Mathematics for Economics and Finance

Answer Key to Final Exam

Instructor: Norman Schürhoff

Date: 16.01.2017

Last Name, First Name:

Instructions:

• Exam paper is provided by the proctors. Fill in the required information on every
sheet. No own paper is allowed. You must turn in all sheets at the end of the exam,
including any scratch paper. All exam questions must be returned.

• The exam includes 4 questions. Tell the proctor immediately if something is missing.
Start your answers to every question at the top of a new page.

• The duration of the exam is 180 minutes. If you arrive late for the exam, you must
finish with all other students. No extension can be granted.

• This is a closed-book exam. You are allowed to use a black or blue ink pen, a
non-programmable calculator, and a dictionary only. The pen must not write
red or green. No pencils are allowed. The dictionary may not have any markings.
You are allowed no other material, including cell phones, computers, class material,
books, cheat sheets.

• Show all derivations and computations. Numerical results without corresponding ana-
lytical derivations receive no points except if stated otherwise. The solution approach
has to be clear to the grader. Write legibly.

• The total number of points is 100.

1
Question I (25 points)
1. Are the following statements true or false? Just state TRUE or FALSE.

(a) The eigenvalues of A are the same as the eigenvalues of kA for any positive
constant k.
(b) If a function f (x) is strictly concave, then its Hessian H(x) is negative definite.
(c) If two events A and B such that P (A) > 0 and P (B) > 0 are disjoint, then they
are also independent.
(d) If a matrix A has at least one eigenvalue equal to 0, then det(A) = 0.
(e) If a function f (x) is continuous at a point x = x0 , then it is also differentiable at
x = x0 .
(f) If two random variables X and Y have the same moments, i.e., E(X n ) = E(Y n )
∀ n ∈ N, then they are equal in distribution.
(g) The Weierstrass Theorem provides sufficient conditions for the existence of solu-
tions to Kuhn-Tucker problems.
(h) If two random variables X and Y are independent, then the expectation of their
product E(XY ) is equal to zero.

2. Compute the following expressions:

(a)
 x1
x2 + 1

lim
x→∞ x+2
(b)
Z e3
√
1 + log x
dx
1 x
(c) ZZ
2
e−y dxdy
D
where
D = (x, y) ∈ R2 : 0 ≤ x ≤ 1, x ≤ y ≤ 1 .


3. Prove the following statement:

∀n ≥ 0 : (1 + a)n ≥ 1 + na.

2
Answers to Question I
1. Answers: (1 point for each)

(a) False. To see this, compare det(A − λI) and det(kA − λI) and notice that in
general the solutions are not the same.
(b) False. The reverse implication holds.
(c) False. If two sets are disjoint, then P(A ∩ B) = 0. If two sets are independent
(and each has positive probability), then P(A ∩ B) = P (A)P (B).
(d) True. For any square matrix with K rows/ columns we have: det(A) = ΠK
i=1 λi .

(e) False. In general the reverse is true. See Proposition 3.28 in the lecture notes.
(f) False. The reverse is true in general. See Proposition 4.8 in the lecture notes.
(g) True. See Theorem 3.16 in the lecture notes.
(h) False. See Definition 4.21 and Proposition 4.21 in the lecture notes.

2. Solutions: (4 points for each item)

(a) We cannot directly apply de l’Hôpital’s theorem because the expression is ”equal”
to ∞0 . Thus we transform the problem:
 2 1  2  1 !!   2  
x +1 x x +1 x x +1
lim = lim exp ln = exp lim ln /x .
x→∞ x+2 x→∞ x+2 x→∞ x+2

Now we can apply l’Hôpital’s theorem:

!
2x 1
x2 +1
− x+2
exp lim = exp(0) = 1.
x→∞ 1

(b)
e3
Z p Z 3 Z 4
1 + log(x) log(x) = y 1 y+1=k 1
dx = = (1+y) dy = 2 = k 2 dk
1 x dx = xdy 0 dy = dk 1

4
2 3 2 2 14
= k2 = 7− = .
3 1 3 3 3
(c) ZZ Z x=1 Z y=1  Z y=1 Z x=y 
−y 2 −y 2 −y 2
e dydx = e dy dx = e dx dy
D x=0 y=x y=0 x=0
1 y 1 1 1
e−1
Z Z Z Z
−y 2 −y 2 −y 2 1 2
= e dxdy = e {x}y0 dy = ye dy = − e−y = .
0 0 0 0 2 0 2e

3
3. For n = 0 the inequality clearly holds, as we have:

(1 + a)0 = 1 = 1 + 0 · a.

Now assume that the inequality holds for n = k. Then for n = k + 1 we have:

(1 + a)k+1 = (1 + a)(1 + a)k ≥ (1 + a)(1 + ka)

= 1 + ka + a + ka2 ≥ 1 + (k + 1)a,

where the first inequality comes from the induction hypothesis for n = k, while the
second one holds as ka2 ≥ 0. Thus the inequality also holds for all n ≥ 1.
(5 points)

4
Question II (25 points)
You are a trader at V CB and you have a budget of 1 billion dollars at your disposal that
you can invest in two stocks. You do not have to invest the whole amount, as the rest can
be put into a riskless bond that has expected return of zero. Let x1 and, respectively, x2 be
the dollar amount (in billions) invested in each stock. Based on historical data you conclude
that the expected return and the variance of your investment are functions of the allocation.
They are given by the following expressions:

Expected return (in %): µinv (x1 , x2 ) = 10x1 + 12x2

2
Variance of investment: σinv (x1 , x2 ) = 5000x21 − 6000x1 x2 + 8000x22
Your objective is to maximize the expected return of your investment by choosing the allo-
cation in the two stocks. Given the risk aversion of your client you are constrained to the
risk limit that the variance of your investment cannot be greater than 625.

1. What is the optimal asset allocation without risk limits but with short-sale constraints?

2. Write down the optimal asset allocation problem with risk limits. Explain why it has
a solution.

3. Check for which points the constraint qualifications are satisfied.

4. Write down the Lagrangian and the Kuhn-Tucker conditions. Are the Kuhn-Tucker
conditions sufficient for a global maximum?

5. Find and discuss the solutions to the optimal asset allocation problem with risk limits.

6. How much does the expected value of your investment change as you vary the risk
constraint? Explain.

5
Answers to Question II
1. Asset 2 has the higher expected return, so you invest all your money in asset 2.
(3 points)
2. The optimization problem can be stated in the standard form as follows:
max f (x1 , x2 ) = 10x1 + 12x2
x1 , x2
s.t. g1 (x1 , x2 ) = 1 − x1 − x2 ≥ 0
g2 (x1 , x2 ) = 625 − 5000x21 + 60001 x2 − 8000x22 ≥ 0
The objective function is continuous and the constraint set D is closed (inequalities).
In order to establish that the constraint set is compact, we have to show that it is
also bounded. This is straightforward, given that from g1 (x1 , x2 ) we can immediately
see that x1 ≤ 1 and x2 ≤ 1. Thus, by Weierstrass’ theorem the objective function
possesses a maximum on D.
(3 points)
3. The constraint set D contains two constraint functions g1 (x1 , x2 ) and g2 (x1 , x2 ). We
have to consider two different cases: either only the constraint {g1 } is binding, or two
constraints {g1 , g2 } are binding. We have to check CQ for both cases. The gradients
are:
Dg1 (x1 , x2 ) = (−1 − 1) and Dg2 (x1 , x2 ) = (−10000x1 + 6000x2 6000x1 − 16000x2 ).
Therefore:
{g1 } : rank(Dg1 (x1 , x2 )) = 1 ∀ x1 , x2
 
−1 −1
{g1 , g2 } : rank = 2,
−10000x1 + 6000x2 6000x1 − 16000x2
with rank becoming less than 2 if and only if 16x1 = 22x2 . However, this point does
not satisfy the binding constraints: it implies that x1 = 11 x and plugging it into g1
8 2
gives x̃2 = 1/(1 + 11 8
). This value violates g2 (as 8000x̃ 2
2 > 625 ⇒ g2 < 0).
(3 points)
4. The Lagrangian is given by:
L(x1 , x2 , λ) = 0.1x1 + 0.12x2 − λ1 (1 − x1 − x2 ) − λ2 (625 − 5000x21 + 6000x1 x2 − 8000x22 ).
Consequently, the resulting KT conditions are:
∂L
= 10 + 10000x1 λ2 − 6000x2 λ2 + λ1 = 0 (1)
∂x1
∂L
= 12 + 16000x2 λ2 − 6000x1 λ2 + λ1 = 0 (2)
∂x2
g1 (x1 , x2 ) = 1 − x1 − x2 ≥ 0 (3)
g2 (x1 , x2 ) = 625 − 5000x21 + 6000x1 x2 − 8000x22 ≥ 0 (4)
λ1 (1 − x1 − x2 ) = 0 (5)
λ2 (625 − 5000x1 + 6000x1 x2 − 8000x22 )
2
= 0 (6)
λ1 ≥ 0 (7)
λ2 ≥ 0. (8)

6
 The KT conditions are sufficient for global optimum because the objective function as
well as the constrains are concave (the objective function is linear, thus both convex
and concave and the concavity of g2 is easily established by looking at the Hessian).
(4+2 points)

5. Let us consider four cases when:

(a) both the constraints are binding (λ1 > 0, λ2 > 0),
(b) only the second constraint is binding (λ1 = 0),
(c) only the first constraint is binding (λ2 = 0),
(d) both the constraints are not binding (λ1 = 0, λ2 = 0).

We have:

(a) When both constraints are binding, we have to solve the following system of
equations:
(
g1 (x1 , x2 ) = 1 − x1 − x2 = 0
g2 (x1 , x2 ) = 625 − 5000x21 + 6000x1 x2 − 8000x22 = 0

Simple manipulation yields the following quadratic equation:

19000x22 − 16000x2 + 4375 = 0,

which has no real solutions. Therefore this case is not feasible.
(b) When λ1 = 0, KT conditions give:
(
10 + λ2 = 0
12 + λ2 = 0

which has no solution, thus this case is not feasible.

(c) When λ1 = 0, KT conditions give the following system of equations:

10 + 10000x1 λ2 − 6000x2 λ2 = 0

12 + 16000x2 λ2 − 6000x1 λ2 = 0

625 − 5000x21 + 6000x1 x2 − 8000x22 = 0


Dividing the first two expressions by λ2 and equating them gives:

10 0.12
= .
10000x1 − 6000x2 16000x2 − 6000x1
Consequently, x1 ≈ 1.29x2 . Substituting x1 into the third equation gives the final
candidate solution of (x1 , x2 ) ≈ (0.3481, 0.2699), in billions of dollars. Note that
not all of your budget will be allocated.
(d) When none of the constraints is binding (λ1 = λ2 = 0), the first order conditions
∂L ∂L
are not satisfied, as ∂x 1
= 10 = 0 and ∂x2
= 12 = 0 can never hold.

7
 Therefore the global optimum of the problem is given by (x∗1 , x∗2 ) ≈ (0.3481, 0.2699).
(6+1 points)

6. The Lagrange multiplier is the marginal effect on the value function, when the con-
strained is ”relaxed” marginally. In other words, λ1 is indicative of ”how much does
the expected value of the portfolio change as the budget constraint changes” while
λ2 helps answer the question of ”how much does the expected value of the portfolio
change as the risk constraint is altered”. They consitute therefore the shadow prices
of relaxing the constraints.
(3 points)

8
Question III (25 points)
The Swiss alpine town of Zernez is inhabited by 1,200 people. Each citizen makes the weekly
grocery shopping in the only grocery store in town, run by Mr. and Mrs. Zipf. One day a
new, cheaper chain store Micros opens its doors. The Zipf family does not expect that
everybody immediately switches to doing their shopping in the new cheaper store, but they
anticipate that 20% of those shopping at Zipf ’s store each week switch to Micros in the
following week. At the same time, some customers who switched miss the personal service
(and gossip) and they switch back. The Zipfs expect that 10% of those shopping at Micros
each week go back to Zipf ’s in the following week. The state of the town, as far as grocery
shopping is concerned, can be represented as follows:
 
z(t)
x(t) = ,
m(t)
where z(t) and m(t) are the numbers of citizens shopping at Zipf ’s store and, respectively,
at Micros in week t after Micros opens. Assume that z(0) = 1, 200 and m(0) = 0.

1. Find a 2 × 2 matrix A such that x(t + 1) = Ax(t).

2. Does A satisfy any of the following properties: symmetric, orthogonal, idempotent?

3. Using only A and x(t), what is the expression for x(t + n), that is, x given n periods
in the future?

4. Compute the eigenvalues and eigenvectors of the matrix A.

5. How many citizens will shop in each store after t weeks?

Hint: Use matrix decomposition.

6. What fractions of clients do the Zipfs lose in the long run (that is, when t → ∞)?

7. The Zipfs expect that they must close down their store when they have less than 500
customers a week. When does this happen?

9
Answers to Question III
1. Note that the state of the town can be described by the following system of equations:

z(t + 1) = 0.8z(t) + 0.1m(t)

m(t + 1) = 0.2z(t) + 0.9m(t).

Thus matrix A is given by:  

0.8 0.1
A= .
0.2 0.9
(3 points)

2. A is not symmetric: A 6= AT .
A is not orthogonal: AAT 6= I.
A is not idempotent: AA 6= A.
(3 points)

3. We know that x(t + 1) = Ax(t). Consequently, x(t + 2) = Ax(t + 1) = A(A(x(t))) =

A2 x(t). As such, iterating forward gives x(t + n) = An x(t).
(3 points)

4. First we compute the eigenvalues using the characteristic polynomial: det(A − λI) = 0.
 
0.8 0.1
det = 0 ⇐⇒ (0.8 − λ)(0.9 − λ) − 0.02 = 0 ⇒ λ1 = 1, λ2 = 0.7.
0.2 0.9

Then we look for the corresponding eigenvectors:

      
−0.2 0.1 x1 0 1
λ1 = 1 ⇒ = ⇒ vλ1 = c, c ∈ R.
0.2 −0.1 x2 0 2
      
0.2 0.1 x1 0 1
λ2 = 0.7 ⇒ = ⇒ vλ2 = c, c ∈ R.
0.2 0.1 x2 0 −1
(5 points)

5. To answer this question, we compute the spectral decomposition of A: A = CΛC −1

and use the property that At = CΛt C −1 . We have:
     
1 0 1 1 −1 1 1 1
Λ= , C= , C = .
0 0.7 2 −1 3 2 −1

We know that x(t) = At x(0) = CΛt C −1 (1200, 0)T . Further computations yield:
      
z(t) 1 1 + 2(0.7)t 1 − (0.7)t 1200 1 + 2(0.7)t
= = 400 .
m(t) 3 2 − 2(0.7)t 2 + (0.7)t 0 2 − 2(0.7)t
(5 points)

10
6. When t → ∞, then we have (z(∞), m(∞)) = (400, 800). Therefore the Zipfs lose
(1200 − 400)/1200 = 2/3 of their clients.
(3 points)

7. To answer this question, we have to solve the following inequality for t:

log(0.125) −2.079
400(1+2(0.7)t ) < 500 ⇐⇒ (0.7)t < 0.125 ⇐⇒ t < ≈ ≈ 5.8 ⇒ t < 6.
log(0.7) −0.357

(3 points)

11
Question IV (25 points)
Let X1 , . . . , Xn be independent and identically distributed random variables with density
2x
fXi (x) = , 0 ≤ x ≤ θ.
θ2
1. Find E(Xi ) and V ar(Xi ).

2. For W = n1 ni=1 Xi , find E(W ) and V ar(W ).

P

3. For what value of a would the estimator

θ̂W = aW

be unbiased? What is V ar(θ̂W ) for this value of a?

4. Let Y = max(X1 , . . . , Xn ). Find the cdf of Y .

Hint: Use the independence of Xi along with the fact that max(a, b) < K ⇒ (a <
K) ∩ (b < K).

5. Find the pdf of Y , E(Y ), and V ar(Y ).

6. For what value of b would the estimator

θ̂Y = bY

be unbiased? What is V ar(θ̂Y ) for this value of b?

7. Which of the estimators θ̂W and θ̂Y is more precise? What about their relative effi-
ciency?
Hint: Look at what happens to the ratio of the variances when n → ∞.

12
Answers to Question IV
1. θ
Z ∞ Z θ 
2 2 2 3 2
E(Xi ) = xfXi (x)dx = 2
x dx = 2
x = θ.
−∞ 0 θ 3θ 0 3
Z ∞ Z θ  θ
2 3 2 4 1
E(Xi2 ) = x2 fXi (x)dx = 2
x dx = 2
x = θ2 .
−∞ 0 θ 4θ 0 2
 2
1 2 θ2
V ar(Xi ) = E(Xi2 ) − [E(Xi )]2 = θ2 − θ = .
2 3 18
(5 points)
2. By linearity of the expectations operator, we know that for n independent and identi-
cally distributed random variables having Xi ∼ (µ, σ 2 ), the sample mean has distribu-
2
tion X ∼ (µ, σn ). Thus:
2 θ2
 
W ∼ θ, .
3 18n
(5 points)

(Note that direct computation also gives full mark.)

3. An estimator is unbiased if E(θ̂) = θ. We have:
2 3
E(θ̂W )E(aW ) = aE(W ) = a θ ⇒ a = .
3 2
For this value of a, the variance of the estimator is:
9 θ2 θ2
V ar(θ̂W ) = V ar(aW ) = a2 V ar(W ) = = .
4 18n 8n
(3 points)
4. We can compute the CDF of Y using its definition:
FY (y) = P(Y < y) = P(max(X1 , . . . , Xn ) < y) = P(∩ni=1 {Xi ≤ y})
= Πni=1 P(Xi ≤ y) (by independence)
= [P(Xi ≤ y)]n (by identical distribution)
Z y n
= fXi (x)dx (substituting the CDF)
−∞
 Z y Z θ n
2 2
= I(0,θ) (y) 2
dx + I(θ,∞) (y) 2
dx
−∞ θ −∞ θ
 2 n
y
= I(0,θ) (y) 2 + I(θ,∞) (y)
θ
2n
y
= 2n I(0,θ) (y) + I(θ,∞) (y).
θ
(4 points)

(Note that ignoring the indicator functions also gives full mark.)

13
5. The pdf of Y is given by:

d y 2n 2ny 2n−1
 
d
fY (y) = FY y = I (0,θ) (y) + I(θ,∞) (y) = I(0,θ) (y).
dy dy θ2n θ2n

The expected value and variance are given by:

θ θ
2ny 2n−1 2n y 2n+1
Z 
2n
E(Y ) = y dy = = θ.
0 θ2n 2n + 1 θ2n 0 2n + 1

2n 2 n
E(Y 2 ) = θ ⇒ V ar(Y ) = 2
θ2 .
2n + 2 (n + 1)(2n + 1)
(4 points)

6. The estimator is unbiased if and only if

2n 2n + 1
E(θ̂Y ) = E(bY ) = bθ = θ thus b = .
2n + 1 2n
Consequently,

(2n + 1)2 2n θ2
V ar(θ̂Y ) = b2 V ar(Y ) = θ 2
= .
4n2 (2n + 2)(2n + 1)2 4n(n + 1)

(3 points)

7. We can consider the ratio

θ2
V ar(θ̂W ) 8n n+1
= θ2
= .
V ar(θ̂Y ) 4n(n+1)
2

Therefore the variance of θ̂W is greater than the variance of θ̂Y for all n > 1 and the
”relative efficiency” of θ̂Y to θ̂W (which the ratio measures) grows in proportion to n
with increasing sample size. Note that when n = 1, then W = Y and all the above
formulas are in perfect agreement with each other.
(3 points)

(Note that you can get max(x, 25) points for the exercise, with x ∈ [0, 27].)

14