Technische Universität München WS 2017/18

Lehrstuhl für Informatik V Scientific Computing

Univ.-Prof. Dr. M. Bader 06.11.2017/08.11.2017
M.Sc. S. Seckler

Worksheet 1
Sample Solutions
Notation

In the worksheets you will find different exercises marked by the following symbols:
• (I) - marks interactive exercises, they are solved by students during class hours with the
help of tutors. Solutions of these exercises might not be provided in a document!
• (H) - marks homework exercises, tutors explain and solve these problems during class
hours. Solutions are also provided in a document after the tutorial.
• (H*) - marks optional homework exercises for self-study, they are not explained during
tutorials. However, solutions are provided after the tutorial.

Eigenvalues

Repetition

The eigenvalue theory can be used to characterise (amongst others) linear systems with re-
spect to the amplification, reduction and frequency of the underlying matrix-based operati-
ons. The eigenvalues λi together with the corresponding eigenvectors vi for a matrix A ∈
R N × N are all pairs for which hold: A · vi = λi vi . It can be shown that the eigenvalues are
identical with the roots of the characteristic polynomial det( A − λ · 1) where 1 denotes the
eye-matrix; ‘det’ represents the determinant of the matrix A − λ · 1. Hence, if λi is an eigen-
value, then it holds that det( A − λi · 1) = 0. Similar to the roots of other polynomials, the
eigenvalues λi do not need to necessarily be real values; they might also lie in the space of
complex numbers.
Depending on the properties of the matrix A, one can find out information about its eigenva-
lues. In the following, three examples should be given for such properties:
1 A matrix A ∈ R N × N is called diagonalisable if it can be written as A = PDP−1 with
invertible matrix P ∈ R N × N and diagonal matrix D ∈ diag( N ). In this case, the diagonal
matrix contains the eigenvalues λi on its diagonal and the columns of P represent the
corresponding eigenvectors.

1
 2 A matrix A ∈ R N × N is called symmetric if it holds Aij = A ji for all i, j = 1, ..., N. If a
matrix is symmetric, all eigenvalues are real values.
3 A matrix A ∈ R N × N is called positive definite if x > Ax > 0 for all vectors x ∈ R N \{~0}
and positive semi-definite if x > Ax ≥ 0 for all x ∈ R N \{~0}. Analogous definitions hold
for negative (semi-)definiteness. For symmetric matrices the following equivalence is true
matrix A is positive definite ⇔ all eigenvalues are positive
In the following exercises, we will practise the computation of eigenvalues and the characte-
risation of matrix-based systems in terms of their eigenvalues.

Remarks & Hints

• You may use the following rule to compute the determinant of a 2 × 2 matrix:
 
a11 a12
det = a11 a22 − a12 a21 (1)
a21 a22

• You may use the following rule to compute the determinant of a 3 × 3 matrix:
 
a11 a12 a13
det  a21 a22 a23  = a11 a22 a33 + a12 a23 a31 + a13 a21 a32
(2)
a31 a32 a33
− a31 a22 a13 − a32 a23 a11 − a33 a21 a12

(I) Exercise 1: Direct Computation of Eigenvalues

Consider the matrices

−6 −14 −12
   
1 0 0
A :=  4 9 6  , B : =  −1 2 −1  (3)
1 2 3 0 0 1

(a) Compute the eigenvalues of the matrices.

(b) Compute the eigenvectors of B.
(c) Which properties does the matrix B have?
Solution:
(a) Using the characteristic polynomial and the formula for the determinant from above yields

2
for matrix A:

−6 − λ −14 −12
 

det ( A − λ1) = det  4 9−λ 6 

1 2 3−λ
(4)
= (−6 − λ)(9 − λ)(3 − λ) + (−14) · 6 · 1 + (−12) · 4 · 2
− 1 · (9 − λ)(−12) − 2 · 6 · (−6 − λ) − (3 − λ) · 4 · (−14)
= ... = −λ3 + 6λ2 − 11λ + 6

An educated guess shows that the values λ0 = 1, λ1 = 2, λ2 = 3 are the roots of the polyno-
mial.
Analogously, one obtains the eigenvalues for the matrix B: λ0 = 1, λ1 = 1, λ2 = 2.
(b) The eigenvectors vi of B can be computed from the equation system B · vi = λi vi , i = 0, 1, 2.
We obtain v0,1 = {(w0 , w1 , w2 )> ∈ R3 : w1 = w0 + w2 }, v2 = {(w0 , w1 , w2 )> ∈ R3 : w0 =
w2 = 0 } .
(c)
• B is not symmetric
• Though B’s eigenvalues are all bigger than zero, we can not immediately say if it’s po-
sitive definite, since it is not symmetric (following point 3 from the introduction on
eigenvalues). (However detailed studies show, that x T Bx > 0, ∀ x 6= 0. Therefore B is
positive definite.)
• B is diagonalisable, since P can be found, such that B = PDP−1

Population Models

(I) Exercise 2: Web-stores

Two web-stores were successfully selling one product for a similar price. To increase their
income they decided to change the prices. Complete the following tasks to find out if their
strategies worked fine.
(a) The first web-store decided to enhance the selling of this product by gradually decrea-
sing the price. Each week they would change the price by taking a sum of 50% of their
current price and 40% of their competitors current price. Write down a recursive formu-
la to find the price an+1 at week n + 1, if the price at week n is an and the competitors’
price is fixed to b. What value will the price approach after many weeks? In other words,
what is the stationary solution of the recursive formula?
(b) The second web-store decided to increase the price of the product. They would make a
weekly price adjustment by taking 50% of their current price and 60% of their compe-
titors current price. Find a recursive formula for the price bn+1 , if the current price is bn
and the competitors’ price is fixed to a. What value will the price approach in the long
run?

3
 (c) The web-stores started applying their strategies simultaneously. Find the matrix A in
the recursive relation    
a ( n + 1) a(n)
= A· (5)
b ( n + 1) b(n)

(d) After some time the two web-stores noticed that something was going wrong with the
product prices. Find the eigenvalues of matrix A and explain what could go wrong.
Solution:
(a) The recursive formula of the first web-store price strategy is

an+1 = 0.5an + 0.4b

In the long run we expect the price not to change anymore, an+1 = an = a and from the
recursive relation it follows that a = 0.4b/0.5 = 0.8b.
(b) The recursive formula of the second web-store product price is

bn+1 = 0.5bn + 0.6a

The stationary solution is b = 0.6a/0.5 = 1.2a.

(c) Matrix A can be written directly from the recursive formulas of an+1 and bn+1 :
 
0.5 0.4
A=
0.6 0.5

(d) The eigenvalues of matrix A are given by

√
1± 0.96
λ1,2 = ≤1
2
Initial prices of the products can be written as a linear combination of the two eigenvectors
( a(0), b(0))> = c1 · v1 + c2 · v2 with constants c1 , c2 ∈ R. Then, we can compute the prices in
week n via:    
a(n) n a (0)
= A
b(n) b (0)
n
= A ( c1 v1 + c2 v2 ) (6)
n
= c1 A v1 + c2 A v2 n

= c1 λ1n v1 + c2 λ2n v2
From this equation we see that in the long run the prices will gradually approach 0, because
both eigenvalues λ1,2 are less than one.

(H) Exercise 3: Rabbits and Fibonacci Numbers

Let’s consider the Fibonacci model for the evolution of rabbits. Each pair is assumed to have
one pair of children each year (one male and one female rabbit). In their first year, the young
rabbits do not have children. After the first year, they will also give birth to one pair of rabbits
each year.

4
 (a) Let Y denote the number of pairs of “young” rabbits and G the number of pairs of
“grown-up” rabbits. Model the evolution of the rabbits by a recursive scheme and a
respective linear relationship between the rabbits Y (n), Y (n + 1), G (n), G (n + 1) of sub-
sequent years n, n + 1:
   
Y ( n + 1) Y (n)
= A· (7)
G ( n + 1) G (n)
with matrix A ∈ R2×2 .
(b) Which properties does A have? Compute the eigenvalues and eigenvectors of A!
(c) Assume an initial rabbit population (Y (0), G (0))> := (0, 1)> . How can we describe
the evolution of the rabbits in terms of eigenvectors? How can we easily estimate the
population in the year 20 by only considering the eigenvalues and eigenvectors of the
system?
Hint: decompose the initial population (Y (0), G (0))> into its eigenvector contributions
and compute their evolution separately.
(d) Assume that each year p % of the grown-up rabbits and q % of the young rabbits die due
to a big fat wolf in their forest. How can you include this assumption into the recursive
model from above? How do the eigenvalues of the update matrix A change in this case?
Remark: for the sake of simplicity, assume that dying and giving birth is a strictly se-
quential process ;-), i.e. the respective percentage of rabbits dies first and the remaining
rabbits give birth to new pairs of rabbits.
Solution
(a) Each new year, there will be as many young rabbits as grown-up rabbits: Y (n + 1) = G (n).
Besides, the number of grown-up rabbits is arising from the original grown-ups G (n) and the
young rabbits of that year Y (n), G (n + 1) = Y (n) + G (n). We hence obtain:
     
Y ( n + 1) 0 1 Y (n)
= · (8)
G ( n + 1) 1 1 G (n)
 
0 1
(b) The matrix A = is symmetric. The eigenvalues must hence be real-valued.
1 1
The characteristic polynomial√
is given by det( A − λ · 1) = λ2 − λ − 1. Its roots, that is the
eigenvalues, are λ0,1 = 1±2 5 .
(c) First, let’s try to understand the hint from this exercise: assume we can write our initial
condition (Y (0), G (0))> as a linear combination of the two eigenvectors v0 , v1 (that belong
to λ0 and λ1 , respectively), that is (Y (0), G (0))> = c0 · v0 + c1 · v1 with constants c0 , c1 ∈ R.
Then, we can compute the population in the year n via:
   
Y (n) Y (0)
= An
G (n) G (0)
= A n ( c0 v0 + c1 v1 ) (9)
= c0 A n v0 + c1 A n v1
= c0 λ0n v0 + c1 λ1n v1

5
From the last equation, it can be seen that we can track the evolution of the rabbits by con-
sidering the evolution of each eigenvector: depending on its corresponding eigenvalue, we
can just compute λin and see how the eigenvector
√
is either amplified
√
or reduced over time. In
our case, the eigenvalues are λ0 = 1+2 5 ≈ 1.618 and λ1 = 1−2 5 ≈ −0.618. For n → ∞, the
n→∞
first term of the upper equation hence tends to infinity, c0 λ0n v0 → ∞, since the magnitude of
the eigenvalue λ0 is bigger than 1. For the second eigenvalue λ1 , it holds that kλ1 k < 1. The
n→∞
respective contribution c1 λ1n v1 is consequently decreasing over time, c1 λ1n v1 → 0.
In our example, the decomposition of the initial √vector can be determined
√
as (0, 1)> = c0 ·
v0 + c1 · v1 with c0 = √15 , c1 = − √15 , v0 = (1, 1+2 5 )> , v1 = (1, 1−2 5 )> . For n = 20, the va-
lue λ20
1 is already so small that we can completely neglect the contribution of the respective
summand to the overall population (λ201 < 1e − 4). We obtain:

(Y (20), G (20))> ≈ c0 λ20

0 v0
≈ 0.45 · 15100 · (1, 1.62)> (10)
≈ (6800, 11000)>

The “exact” iteration using the matrix A delivers: (Y (20), G (20))> = A20 (Y (0), G (0))> =
(6765, 10946)> .
(d) As p% of the grown-up rabbits die, we will only have 1 − p/100 new young rabbits,
Y (n + 1) = (1 − p/100) G ( N + 1). With q % of the former young rabbits dying, we will have
less grown-ups as well: G (n + 1) = (1 − q/100)Y (n) + (1 − p/100) G (n).
The arising matrix looks as follows:
 p 
0 1− 100
A= q p (11)
1 − 100 1− 100
r  
p p
2 p+q pq
1− 100 ± 1− 100 − 4 −1 + 100 − 10000
The eigenvalues evolve at λ0,1 = .
2

(H*) Exercise 4: Rates and Calculation of Interest

Congratulations! You are just about to open a new bank account. To open it, you initially
invest K (n = 0) euros, that is you start with K (n = 0) euros in the year n = 0. After each year,
you first obtain an interest rate of p % onto your current savings. Besides, you are obliged to
add another J euros each year onto your current account.
(a) Try to find a model for your bank account which–based on a recursive formula–computes
your savings K (n + 1) in the (n + 1)-th year from the savings K (n).
(b) Which value can be considered to be an eigenvalue in our recursive expression? Which
quantities affect the eigenvalue and what happens to your savings when you modify
them?
(c) Convert the recursive relation from (a) into a non-recursive expression.

6
 (d) The people from the bank cheated on you. Though they first announced that the bank
account is for free, you suddenly need to pay n euros in the year n, starting in the very
first year (hence, only the first year was free)! Include the arising costs into the recursive
expression from (a).
(e) How do the costs of n euros in year n enter the non-recursive formula for your savings?
(f) What can you buy from your saved money in 10 years, assuming an interest rate p =
0.05, an initial payment of K (n = 0) = 50 euros and annual investments of J = 50? A
new notebook (∼ 1000 Euros), the latest iPad (∼ 800 Euros), or the latest iPhone (∼ 650
Euros)?
Solution
(a) K (n + 1) = (1 + p) · K (n) + J
(b) The value 1 + p can be considered to be a characteristic value and indeed is the eigenvalue
of our recursive scheme for J = 0. In this case, if p < 0, it implies that we have to pay − p of
our savings to the bank in one year. For n → ∞, our savings tend to zero, K (n → ∞) → 0.
For p > 0, we obtain more and more money, i.e. K (n → ∞) → ∞. For p = 0, our savings stay
constant: K (n → ∞) = K (0).
(c) We can first of all look at the first three years of our savings development:
K (1) = (1 + p ) K (0) + J
K (2) = (1 + p )2 K (0) + (1 + p ) J + J (12)
K (3) = (1 + p )3 K (0) + (1 + p )2 J + (1 + p ) J + J
We can hence write our problem as:
n
K ( n + 1 ) = (1 + p ) n +1 K (0 ) + J ∑ (1 + p ) k (13)
k =0
n
q n +1 − 1
The equality ∑ qn = q−1 –applied to the last term of Eq. (??) yields:
k =0

J 
K ( n + 1 ) = (1 + p ) n +1 K (0 ) + (1 + p ) n +1 − 1 (14)
p
(d) The new formula looks as follows: K (n + 1) = (1 + p)K (n) + J − n
(e) For the non-recursive formula, we can again consider the first three steps
K (1) = (1 + p ) K (0) + J − 0
K (2) = (1 + p )2 K (0) + (1 + p ) J + J − 0 · (1 + p ) − 1 (15)
K (3) = (1 + p)3 K (0) + (1 + p)2 J + (1 + p) J + J − 0 · (1 + p)2 − 1 · (1 + p) − 2,
and–via induction–we obtain the formula
n n
K ( n + 1 ) = ( 1 + p ) n +1 K (0 ) + J ∑ (1 + p ) k − ∑ k (1 + p ) n − k (16)
k =0 k =0
n
n qn+2 −(n+1)qn+1 +q
The equality ∑ k · qk = ( q −1)2
can be used to re-write the term
k =0
n n  k
1
∑ k (1 + p ) n−k
= (1 + p ) n
∑k 1+ p
(17)
k =0 k =0

7
with q = 1/(1 + p). Re-writing the term finally yields:

J  n − (n + 1)(1 + p) + (1 + p)n+1
K ( n + 1 ) = (1 + p ) n +1 K (0 ) + (1 + p ) n +1 − 1 − (18)
p p2

(f) Inserting p = 0.05, J = 50 and K (0) = 50 into the formula from above yields: K (10) =
658.78. Hence, you can enjoy a new iPhone :-)