Exercises from Linear algebra, signal

processing, and wavelets. A unified

approach.
Python version

Øyvind Ryan

Feb 8, 2017
Contents

1 Sound and Fourier series 3

2 Digital sound and Discrete Fourier analysis 20

3 Operations on digital sound: digital filters 38

4 Symmetric filters and the DCT 53

5 Motivation for wavelets and some simple examples 65

6 The filter representation of wavelets 91

7 Constructing interesting wavelets 105

8 The polyphase representation and wavelets 108

9 Digital images 123

10 Using tensor products to apply wavelets to images 137

A Basic Linear Algebra 143

B Signal processing and linear algebra: a translation guide 144

2
Chapter 1

Sound and Fourier series

Exercise 1.6: The Krakatoa explosion

Compute the loudness of the Krakatoa explosion on the decibel scale, assuming
that the variation in air pressure peaked at 100 000 Pa.

Solution. Setting pref =0.00002 Pa and p=100 000 Pa in the decibel expression
we get

105
     
p 100000
20 log10 = 20 log10 = 20 log10
pref 0.00002 2 × 10−5
 10 
10
= 20 log10 = 20 (10 − log10 2) ≈ 194db.
2

Exercise 1.7: Sum of two pure tones

Consider a sum of two pure tones, f (t) = A1 sin(2πν1 t) + A2 sin(2πν2 t). For
which values of A1 , A2 , ν1 , ν2 is f periodic? What is the period of f when it is
periodic?

Solution. sin(2πν1 t) has period 1/ν1 , while sin(2πν2 t) has period 1/ν2 . The
period is not unique, however. The first one also has period n/ν1 , and the
second also n/ν2 , for any n. The sum is periodic if there exist n1 , n2 so that
n1 /ν1 = n2 /ν2 , i.e. so that there exists a common period between the two. This
common period will also be a period of f . This amounts to that ν1 /ν2 = n1 /n2 ,
i.e. that ν1 /ν2 is a rational number.

Exercise 1.8: Sum of two pure tones

Find two constant a and b so that the function f (t) = a sin(2π440t)+b sin(2π4400t)
resembles the right plot of Figure 1.1 in the book as closely as possible. Generate
the samples of this sound, and listen to it.

3
CHAPTER 1. SOUND AND FOURIER SERIES 4

Solution. The important thing to note here is that there are two oscillations
present in the right part of Figure 1.1 in the book: One slow oscillation with
a higher amplitude, and one faster oscillation, with a lower amplitude. We see
that there are 10 periods of the smaller oscillation within one period of the
larger oscillation, so that we should be able to reconstruct the figure by using
frequencies where one is 10 times the other, such as 440Hz and 4400Hz. Also,
we see from the figure that the amplitude of the larger oscillation is close to 1,
and close to 0.3 for the smaller oscillation. A good choice therefore seems to be
a = 1, b = 0.3. The code can look this:

t = arange(0,3,1/float(fs))
x = sin(2*pi*440*t) + 0.3*sin(2*pi*4400*t)
x /= abs(x).max()
play(x, fs)

Exercise 1.9: Playing with different sample rates

If we provide another samples rate fs to the play functions, the sound card will
assume a different time distance between neighboring samples. Play and listen
to the audio sample file again, but with three different sample rates: 2*fs, fs,
and fs/2, where fs is the sample rate returned by audioread.

Solution. The following code can be used

play(x, fs)

play(x, 2*fs)

play(x, fs/2)

The sample file castanets.wav played at double sampling rate can be found in
the file castanetsdouble.wav, and in castanetshalf.wav for half the sampling rate.

Exercise 1.10: Play sound with added noise

To remove noise from recorded sound can be very challenging, but adding noise
is simple. There are many kinds of noise, but one kind is easily obtained by
adding random numbers to the samples of a sound. For this we can use the
function random.random as follows.

z = x + c*(2*random.random(shape(x))-1)

This adds noise to all channels. The function for returning random numbers
returns numbers between 0 and 1, and above we have adjusted these so that
CHAPTER 1. SOUND AND FOURIER SERIES 5

they are between −1 and 1 instead, as for other sound which can be played by
the computer. c is a constant (usually smaller than 1) that dampens the noise.
Write code which adds noise to the audio sample file, and listen to the result
for damping constants c=0.4 and c=0.1. Remember to scale the sound values
after you have added noise, since they may be outside [−1, 1] then.

Solution. The following code can be used.

z = x + c*(2*random.random(shape(x))-1)
z /= abs(z).max()
play(z, fs)

With c = 0.4 the resulting sound can be found in the file castanetsnoisehigh.wav,
while with c = 0.1 it can be found in the file castanetsnoiselow.wav.

Exercise 1.11: Playing the triangle wave

Repeat what you did in Example 1.4 in the book, but now for the triangle wave
of Example 1.5 in the book. Start by generating the samples for one period of
the triangle wave, then plot five periods, before you generate the sound over a
period of three seconds, and play it. Verify that you generate the same sound as
in Example 1.5 in the book.

Solution. The triangle wave can be plotted as follows:

totransl=1-4*abs(t-T/2)/T
plt.figure()
plt.plot(concatenate([t, t+T, t+2*T, t+3*T, t+4*T]),tile(totransl,5), ’k-’)

The samples for one period are created as follows.

oneperiod = hstack([linspace(-1, 1, samplesperperiod/2), \

linspace(1, -1, samplesperperiod/2)])

Then we repeat one period to obtain a sound with the desired length, and play
it as follows.

x = tile(oneperiod, antsec*f)
play(x, fs)

Exercise 1.15: Shifting the Fourier basis vectors

Show that sin(2πnt/T + a) ∈ VN,T when |n| ≤ N , regardless of the value of a.

Solution. Write sin(2πnt/T + a) = cos a sin(2πnt/T ) + sin a cos(2πnt/T ).

CHAPTER 1. SOUND AND FOURIER SERIES 6

Exercise 1.16: Playing the Fourier series of the triangle

wave
a) Plot the Fourier series of the triangle wave.

Solution. The following code can be used.

t = linspace(0, T, 100)
x = zeros(len(t))
for k in range(1, 20, 2):
x -= (8/(k*pi)**2)*cos(2*pi*k*t/T)
plt.figure()
plt.plot(t, x, ’k-’)

b) Write code so that you can listen to the Fourier series of the triangle wave.
How high must you choose N for the Fourier series to be indistuingishable from
the square/triangle waves themselves?

Solution. The following code can be used.

x = tile(x, antsec/T)
play(x, fs)

Exercise 1.17: Riemann-integrable functions which are not

square-integrable
RT
Find a function f which is Riemann-integrable on [0, T ], and so that 0
f (t)2 dt
is infinite.

Solution. The function f (t) = 1

√
t
= t−1/2 can be used since it has the
properties

Z T Z T h iT
f (t)dt = lim t−1/2 dt = lim 2t1/2
0 x→0+ x x→0+ x
1/2 1/2 1/2
= lim (2T − 2x ) = 2T
x→0+
Z T Z T
T
f (t)2 dt = lim t−1 dt = lim [ln t]x
0 x→0+ x x→0+

= ln T − lim ln x = ∞.
x→0+

Exercise 1.18: When are Fourier spaces included in each

other?
Given the two Fourier spaces VN1 ,T1 , VN2 ,T2 . Find necessary and sufficient
conditions in order for VN1 ,T1 ⊂ VN2 ,T2 .
CHAPTER 1. SOUND AND FOURIER SERIES 7

Solution. The space VN1 ,T1 is spanned by pure tones with frequencies 1/T1 , . . . , N1 /T1 ,
while VN2 ,T2 is spanned by pure tones with frequencies 1/T2 , . . . , N2 /T2 . We
must have that the first set of frequencies is contained in the second. This is
achieved if and only if 1/T1 = k/T2 for some integer k, and also N1 /T1 ≤ N2 /T2 .
In other words, T2 /T1 must be an integer, and T2 /T1 ≤ N2 /N1 .

Exercise 1.19: antisymmetric functions are sine-series

Prove the second part of Theorem 1.10 in the book, i.e. show that if f is
antisymmetric about 0 (i.e. f (−t) = −f (t) for all t), then an = 0, i.e. the Fourier
series is actually a sine-series.

Exercise 1.20: More connections between symmetric-/antisymmetric

functions and sine-/cosine series
Show that
PN
a) Any cosine series a0 + n=1 an cos(2πnt/T ) is a symmetric function.
PN
b) Any sine series n=1 bn sin(2πnt/T ) is an antisymmetric function.
c) Any periodic function can be written as a sum of a symmetric - and an
antisymmetric function by writing f (t) = f (t)+f
2
(−t)
+ f (t)−f
2
(−t)
.
PN
d) If fN (t) = a0 + n=1 (an cos(2πnt/T ) + bn sin(2πnt/T )), then

N
fN (t) + fN (−t) X
= a0 + an cos(2πnt/T )
2 n=1
N
fN (t) − fN (−t) X
= bn sin(2πnt/T ).
2 n=1

Exercise 1.21: Fourier series for low-degree polynomials

Find the Fourier series coefficients of the periodic functions with period T defined
by being f (t) = t, f (t) = t2 , and f (t) = t3 , on [0, T ].

1
RT T
Solution. For f (t) = t we get that a0 = T 0
tdt = 2 . We also get
CHAPTER 1. SOUND AND FOURIER SERIES 8

Z T
2
an = t cos(2πnt/T )dt
T 0
 T Z T !
2 T T
= t sin(2πnt/T ) − sin(2πnt/T )dt = 0
T 2πn 0 2πn 0
Z T
2
bn = t sin(2πnt/T )dt
T 0
 T Z T !
2 T T T
= − t cos(2πnt/T ) + cos(2πnt/T )dt = − .
T 2πn 0 2πn 0 πn

The Fourier series is thus

T X T
− sin(2πnt/T ).
2 πn
n≥1

Note that this is almost a sine series, since it has a constant term, but no other
cosine terms. If we had subtracted T /2 we would have obtained a function which
is antisymmetric, and thus a pure sine series.
RT 2
For f (t) = t2 we get that a0 = T1 0 t2 dt = T3 . We also get

Z T
2
an = t2 cos(2πnt/T )dt
T 0
 T Z T !
2 T 2 T
= t sin(2πnt/T ) − t sin(2πnt/T )dt
T 2πn 0 πn 0
T2
  
T T
= − − = 2 2
πn πn π n
Z T
2
bn = t2 sin(2πnt/T )dt
T 0
 T Z T !
2 T 2 T
= − t cos(2πnt/T ) + t cos(2πnt/T )dt
T 2πn 0 πn 0
T2
=− .
πn
Here we see that we could use the expressions for the Fourier coefficients of
f (t) = t to save some work. The Fourier series is thus

T2 X
 2
T2

T
+ cos(2πnt/T ) − sin(2πnt/T ) .
3 π 2 n2 πn
n≥1

1
RT T3
For f (t) = t3 we get that a0 = T 0
t3 dt = 4 . We also get
CHAPTER 1. SOUND AND FOURIER SERIES 9

Z T
2
an = t3 cos(2πnt/T )dt
T 0
 T Z T !
2 T 3 3T 2
= t sin(2πnt/T ) − t sin(2πnt/T )dt
T 2πn 0 2πn 0
T2 3T 3
  
3T
= − − =
2πn πn 2π 2 n2
Z T
2
bn = t3 sin(2πnt/T )dt
T 0
 T Z T !
2 T 3 3T 2
= − t cos(2πnt/T ) + t cos(2πnt/T )dt
T 2πn 0 2πn 0
T3 3T T 2 T3 3T 3
=− + = − + .
πn 2πn π 2 n2 πn 2π 3 n3
Also here we saved some work, by reusing the expressions for the Fourier
coefficients of f (t) = t2 . The Fourier series is thus

T3 X 3T 3 T3 3T 3
   
+ cos(2πnt/T ) + − + sin(2πnt/T ) .
4 2π 2 n2 πn 2π 3 n3
n≥1

We see that all three Fourier series converge slowly. This is connected to the
fact that none of the functions are continuous at the borders of the periods.

Exercise 1.22: Fourier series for polynomials

Write down difference equations for finding the Fourier coefficients of f (t) = tk+1
from those of f (t) = tk , and write a program which uses this recursion. Use the
program to verify what you computed in Exercise 1.

Solution. Let us define an,k , bn,k as the Fourier coefficients of tk . When k > 0
and n > 0, integration by parts gives us the following difference equations:
CHAPTER 1. SOUND AND FOURIER SERIES 10

Z T
2
an,k = tk cos(2πnt/T )dt
T 0
T !
 Z T
2 T k kT k−1
= t sin(2πnt/T ) − t sin(2πnt/T )dt
T 2πn 0 2πn 0

kT
=− bn,k−1
2πn
Z T
2
bn,k = tk sin(2πnt/T )dt
T 0
 T Z T !
2 T k kT k−1
= − t cos(2πnt/T ) + t cos(2πnt/T )dt
T 2πn 0 2πn 0
Tk kT
=− + an,k−1 .
πn 2πn
When n > 0, these can be used to express an,k , bn,k in terms of an,0 , bn,0 , for
Tk
which we clearly have an,0 = bn,0 = 0. For n = 0 we have that a0,k = k+1 for all
k.
The following program computes an,k , bn,k recursively when n > 0.

def findfouriercoeffs(n, k, T):

ank, bnk = 0, 0
if k > 0:
ankprev, bnkprev = findfouriercoeffs(n,k-1,T)
ank = -k*T*bnkprev/(2*pi*n)
bnk = -T**k/(pi*n) + k*T*ankprev/(2*pi*n)
return ank, bnk

Exercise 1.23: Fourier series of a given polynomial

Use the previous exercise to find the Fourier series for f (x) = − 13 x3 + 12 x2 − 16
3
x+1
on the interval [0, 1]. Plot the 9th order Fourier series for this function. You
should obtain the plots from Figure 1.5 in the book.

Exercise 1.27: Orthonormality of Complex Fourier basis

Show that the complex functions e2πint/T are orthonormal.

Solution. For n1 6= n2 we have that

CHAPTER 1. SOUND AND FOURIER SERIES 11

Z T Z T
1 1
he2πin1 t/T , e2πin2 t/T i = e2πin1 t/T e−2πin2 t/T dt = e2πi(n1 −n2 )t/T dt
T 0 T 0
 T
1 T
= e2πi(n1 −n2 )t/T
T 2πi(n1 − n2 ) 0
1 1
= − = 0.
2πi(n1 − n2 ) 2πi(n1 − n2 )

When n1 = n2 the integrand computes to 1, so that ke2πint/T k = 1.

Exercise 1.28: Complex Fourier series of f (t) = sin2 (2πt/T )

Compute the complex Fourier series of the function f (t) = sin2 (2πt/T ).

Solution. We have that

 2
1 2πit/T
f (t) = sin2 (2πt/T ) = (e − e−2πit/T
2i
1 1 1 1
= − (e2πi2t/T − 2 + e−2πi2t/T ) = − e2πi2t/T + − e−2πi2t/T .
4 4 2 4
This gives the Fourier series of the function (with y2 = y−2 = −1/4, y0 = 1/2).
This could also have been shown by using the trigonometric identity sin2 x =
1 1 T −2πint/T
R
2 (1 − cos(2x)) first, or by computing the integral T 0 f (t)e dt (but this
is rather cumbersome).

Exercise 1.29: Complex Fourier series of polynomials

Repeat Exercise 1, computing the complex Fourier series instead of the real
Fourier series.

Exercise 1.30: Complex Fourier series and Pascals triangle

In this exercise we will find a connection with certain Fourier series and the rows
in Pascal’s triangle.
a) Show that both cosn (t) and sinn (t) are in VN,2π for 1 ≤ n ≤ N .

Solution. We have that

 n
n 1 it −it
cos (t) = (e + e )
2
 n
1 it
sinn (t) = (e − e−it )
2i
CHAPTER 1. SOUND AND FOURIER SERIES 12

If we multiply out here, we get a sum of terms of the form eikt , where −n ≤ k ≤ n.
As long as n ≤ N it is clear that this is in VN,2π .
b) Write down the N ’th order complex Fourier series for f1 (t) = cos t, f2 (t) =
cos2 t, og f3 (t) = cos3 t.

Solution. We have that

1 it
cos(t) = (e + e−it )
2
1 1 1 1
cos2 (t) = (eit + e−it )2 = e2it + + e−2it
4 4 2 4
1 1 3 3 1
cos3 (t) = (eit + e−it )3 = e3it + eit + e−it + e−3it .
8 8 8 8 8
Therefore, for the first function the nonzero Fourier coefficients are y−1 = 1/2,
y1 = 1/2, for the second function y−2 = 1/4, y0 = 1/2, y2 = 1/4, for the third
function y−3 = 1/8, y−1 = 3/8, y1 = 3/8, y3 = 1/8.
c) In b) you should be able to see a connection between the Fourier coefficients
and the three first rows in Pascal’s triangle. Formulate and prove a general
relationship between row n in Pascal’s triangle and the Fourier coefficients of
fn (t) = cosn t.

Solution. In order to find the Fourier coefficients of cosn (t) we have to multiply
out the expression 21n (eit + e−it )n . The coefficients we get after this can alos be
obtained from Pascal’s triangle.

Exercise 1.31: Complex Fourier coefficients of the square

wave
Compute the complex Fourier coefficients of the square wave using Equation
(1.22) in the book, i.e. repeat the calculations from Example 1.12 in the book
for the complex case. Use Theorem 1.15 in the book to verify your result.

Solution. We obtain that

Z T /2 Z T
1 1
yn = e−2πint/T dt − e−2πint/T dt
T 0 T T /2
 T /2  T
1 T −2πint/T T −2πint/T 1
=− e e +
T 2πin 0 2πin T T /2
1
−e−πin + 1 + 1 − e−πin +


=
2πin (
1 −πin

0, if n is even;
= 1−e = .
πin 2/(πin), if n is odd.
CHAPTER 1. SOUND AND FOURIER SERIES 13

Instead using Theorem 1.15 in the book together with the coefficients bn =
2(1−cos(nπ)
nπ we computed in Example 1.12 in the book, we obtain

( (
1 1 0, if n is even; 0, if n is even;
yn = (an − ibn ) = − i =
2 2 4/(nπ), if n is odd. 2/(πin), if n is odd.

when n > 0. The case n < 0 follows similarly.

Exercise 1.32: Complex Fourier coefficients of the triangle

wave
Repeat Exercise 1 for the triangle wave.

Exercise 1.33: Complex Fourier coefficients of low-degree

polynomials
Use Equation (1.22) in the book to compute the complex Fourier coefficients of
the periodic functions with period T defined by, respectively, f (t) = t, f (t) = t2 ,
and f (t) = t3 , on [0, T ]. Use Theorem 1.15 in the book to verify your calculations
from Exercise 1.

Solution. For f (t) = t we get

T Z T !
Z T 
1 −2πint/T 1 T −2πint/T T −2πint/T
yn = te dt = − te + e dt
T 0 T 2πin 0 0 2πin

T T
=− = i.
2πin 2πn
T
From Exercise 1 we had bn = − πn , for which Theorem 1.15 in the book gives
T
yn = 2πn i for n > 0, which coincides with the expression we obtained. The case
n < 0 follows similarly.
For f (t) = t2 we get

T !
Z T  Z T
1 2 −2πint/T 1 T 2 −2πint/T T −2πint/T
yn = t e dt = − t e +2 te dt
T 0 T 2πin 0 0 2πin

T2 T2 T2 T2
=− + 2 2 = + i.
2πin 2π n 2π 2 n2 2πn
2 2
T T
From Exercise 1 we had an2 = π2 n22 and bn = − πn , for which Theorem 1.15 in
the book gives yn = 12 πT2 n2 + i πn
T
for n > 0, which also is seen to coincide
with what we obtained. The case n < 0 follows similarly.
For f (t) = t3 we get
CHAPTER 1. SOUND AND FOURIER SERIES 14

T !
Z T  Z T
1 3 −2πint/T T 3 −2πint/T
1 T 2 −2πint/T
yn = t e −
dt = t e +3 t e dt
T 0 T
2πin 0 0 2πin

T3 T2 T2 T3
 3
T3

T T
=− +3 ( + i) = 3 2 2 + −3 3 3 i=
2πin 2πin 2π 2 n2 2πn 4π n 2πn 4π n
3T 3 T 3 3
From Exercise 1 we had an = 2π 2 n2 and bn = − πn + 2π3T3 n3 for which Theorem 1.15
in the book gives

3T 3 T3 3T 3 3T 3 T3 3T 3
    
1
yn = +i − 3 3 = + − 3 3 i
2 2π 2 n2 πn 2π n 4π 2 n2 2πn 4π n

for n > 0, which also is seen to coincide with what we obtained. The case n < 0
follows similarly.

Exercise 1.34: Complex Fourier coefficients for symmetric

and antisymmetric functions
In this exercise we will prove a version of Theorem 1.10 in the book for complex
Fourier coefficients.
a) If f is symmetric about 0, show that yn is real, and that y−n = yn .

Solution. If f is symmetric about 0 we have that bn = 0. Theorem 1.15 in

the book then gives that yn = 12 an , which is real. The same theorem gives that
y−n = 12 an = yn .
b) If f is antisymmetric about 0, show that the yn are purely imaginary, y0 = 0,
and that y−n = −yn .

Solution. If f is antisymmetric about 0 we have that an = 0. Theorem 1.15

in the book then gives that yn = − 12 bn , which is purely imaginary. The same
theorem gives that y−n = 21 bn = −yn .
PN
c) Show that n=−N yn e2πint/T is symmetric when y−n = yn for all n, and
rewrite it as a cosine-series.

Solution. When yn = y−n we can write

y−n e2πi(−n)t/T + yn e2πint/T = yn (e2πint/T + e−2πint/T ) = 2yn cos(2πnt/T )

PN
This is clearly symmetric, but then also n=−N yn e2πint/T is symmetric since
it is a sum of symmetric functions.
CHAPTER 1. SOUND AND FOURIER SERIES 15

PN
d) Show that n=−N yn e2πint/T is antisymmetric when y0 = 0 and y−n = −yn
for all n, and rewrite it as a sine-series.

Solution. When yn = −y−n we can write

y−n e2πi(−n)t/T + yn e2πint/T = yn (−e2πint/T + e2πint/T ) = 2iyn sin(2πnt/T )

PN
This is clearly antisymmetric, but then also n=−N yn e2πint/T is antisymmetric
since it is a sum of antisymmetric functions, and since y0 = 0.

Exercise 1.36: Fourier series of a delayed square wave

Define the function f with period T on [−T /2, T /2) by
(
1, if −T /4 ≤ t < T /4;
f (t) =
−1, if T /4 ≤ |t| < T /2.
f is just the square wave, delayed with d = −T /4. Compute the Fourier
coefficients of f directly, and use Property 4 in Theorem 1.17 in the book to
verify your result.

Solution. We obtain that

T /4
1 −T /4 −2πint/T 1 T /2 −2πint/T
Z Z Z
1 −2πint/T
yn = e dt − e dt − e dt
T −T /4 T −T /2 T T /4
 T /4  −T /4  T /2
1 −2πint/T 1 −2πint/T 1 −2πint/T
=− e + e + e
2πin −T /4 2πin −T /2 2πin T /4
1  −πin/2 
= −e + eπin/2 + eπin/2 − eπin + e−πin − e−πin/2
2πin
1 2
= (2 sin(πn/2) − sin(πn)) = sin(πn/2).
πn πn
The square wave defined in this exercise can be obtained by delaying our original
square wave with −T /4. Using Property 3 in Theorem 1.17 in the book with
d = −T /4 on the complex Fourier coefficients
(
0, if n is even;
yn =
2/(πin), if n is odd,
which we obtained for the square wave in Exercise 1, we obtain the Fourier
coefficients
CHAPTER 1. SOUND AND FOURIER SERIES 16

( (
2πin(T /4)/T 0, if n is even; 0, if n is even;
e = 2i sin(πn/2)
2/(πin), if n is odd. πin , if n is odd.
(
0, if n is even;
= 2
.
πn sin(πn/2), if n is odd.
This verifies the result.

Exercise 1.37: Find function from its Fourier series

Find a function f which has the complex Fourier series
X 4
e2πint/T .
π(n + 4)
n odd

Hint. Attempt to use one of the properties in Theorem 1.17 in the book on
the Fourier series of the square wave.

Solution. Since the real Fourier series of the square wave is

X 4
sin(2πnt/T ),
πn
n≥1,n odd

Theorem 1.15 in the book gives us that the complex Fourier coefficients are
yn = − 12 i πn
4 2i
= − πn , and y−n = 12 i πn
4 2i
= πn for n > 0. This means that
2i
yn = − πn for all n, so that the complex Fourier series of the square wave is
X 2i
− e2πint/T .
πn
n odd

Using Property 4 in Theorem 1.17 in the book we get that the e−2πi4t/T (i.e. set
2i
d = −4) times the square wave has its n’th Fourier coefficient equal to − π(n+4) .
Using linearity, this means that 2ie−2πi4t/T times the square wave has its n’th
4
Fourier coefficient equal to π(n+4) . We thus have that the function
(
2ie−2πi4t/T , 0 ≤ t < T /2
f (t) =
−2ie−2πi4t/T , T /2 ≤ t < T
has the desired Fourier series.

Exercise 1.38: Relation between complex Fourier coeffi-

cients of f and cosine-coefficients of f˘
Show that the complex Fourier coefficients yn of f , and the cosine-coefficients
an of f˘ are related by a2n = yn + y−n . This result is not enough to obtain the
entire Fourier series of f˘, but at least it gives us half of it.
CHAPTER 1. SOUND AND FOURIER SERIES 17

Solution. The 2nth complex Fourier coefficient of f˘ is

Z 2T
1
f˘(t)e−2πi2nt/(2T ) dt
2T 0
Z T Z 2T
1 −2πint/T 1
= f (t)e dt + f (2T − t)e−2πint/T dt.
2T 0 2T T

Substituting u = 2T − t in the second integral we see that this is

Z T Z 0
1 1
= f (t)e−2πint/T dt − f (u)e2πinu/T du
2T 0 2T T
Z T Z T
1 1
= f (t)e−2πint/T dt + f (t)e2πint/T dt
2T 0 2T 0
1 1
= yn + y−n .
2 2
Therefore we have a2n = yn − y−n .

Exercise 1.39: Filters preserve sine- and cosine-series

An analog filter where λs (ν) = λs (−ν) is also called a symmetric filter.
a) Prove that, if the input to a symmetric filter is a Fourier series which is a
cosine series/sine-series, then the output also is a cosine/sine series.

Solution. We have that

 
1 2πint/T −2πint/T
s(cos(2πnt/T )) = s (e +e )
2
1 1
= λs (n/T )e2πint/T + λs (−n/T )e−2πint/T
2  2 
1 2πint/T −2πint/T
= λs (n/T ) (e +e ) = λs (n/T ) cos(2πnt/T ),
2

so that s preserves cosine-series. A similar computation holds for sine-series.

Ra
b) Show that s(f ) = −a g(s)f (t − s)ds is a symmetric filter whenever g is
symmetric around 0 and supported on [−a, a].

Solution. We have that

Z a Z a
−2πiνs
λs (ν) = g(s)e ds = g(s)e2πiνs ds = λs (−ν),
−a −a

so that s is symmetric.
CHAPTER 1. SOUND AND FOURIER SERIES 18

We saw that the symmetric extension of a function took the form of a cosine-
series, and that this converged faster to the symmetric extension than the Fourier
series did to the function. If a filter preserves cosine-series it will also preserve
symmetric extensions, and therefore also map fast-converging Fourier series to
fast-converging Fourier series.

Exercise 1.40: Approximation in norm with continuous

functions
Show that if f is a function with only a finite number of discontinuities, there
exists a continuous function g so that kf − gk < .

Exercise 1.41: The Dirichlet kernel

The Dirichlet kernel is defined as
N
X N
X
DN (t) = e2πint/T = 1 + 2 cos(2πnt/T ).
n=−N n=1

DN is clearly trigonometric, and of degree N .

sin(π(2N +1)t/T )
a) Show that DN (t) = sin(πt/T ) .
RT
b) Show that fN (t) = T1 0 f (t − u)DN (u)du. Proving that limN →∞ fN (t) =
RT
f (t) is thus equivalent to limN →∞ T1 0 f (t − u)DN (u)du = f (t).

Solution. We have that

N
1 T
X Z
fN (t) = f (s)e−2πins/T ds e2πint/T
T 0
n=−N
Z T N
1 X
= f (s) e−2πins/T e2πint/T ds
T 0 n=−N
Z T N Z T
1 X 1
= f (s) e2πin(t−s)/T ds = f (s)DN (t − s)ds.
T 0 T 0
n=−N

c) Prove that DN (t) satisfies only two of the properties of a summability kernel.
d) Write a function which takes N and T as arguments, and plots DN (t) over
[−T /2, T /2].
CHAPTER 1. SOUND AND FOURIER SERIES 19

Exercise 1.42: The Fejer summability kernel

The Fejer kernel is defined as
N  
X |n|
FN (t) = 1− e2πint/N .
N +1
n=−N

FN is clearly trigonometric, and of degree N .

 2
a) Show that FN (t) = N1+1 sin(π(N +1)t/T )
sin(πt/T ) , and conclude from this that
T2
0 ≤ FN (t) ≤ 4(N +1)t2 .

2
Hint. Use that π |u| ≤ | sin u| when u ∈ [−π/2, π/2].
b) Show that FN (t) satisfies the three properties of a summability kernel.
RT PN
c) Show that T1 0 f (t − u)FN (u)du = N1+1 n=0 fn .

1
PN
Hint. Show that FN (t) = N +1 n=0 Dn (t), and use Exercis 1 b).
d) Write a function which takes N and T as arguments, and plots FN (t) over
[−T /2, T /2].
Chapter 2

Digital sound and Discrete

Fourier analysis

Exercise 2.6: Computing the DFT by hand

Compute F4 x when x = (2, 3, 4, 5).

Solution. As in Example 2.3 in the book we get

    
2 1 1 1 1 2
3 1 1 −i −1 i  3
F4 
4 = 2 1 −1 1 −1 4
   

5 1 i −1 −i 5
   
2+3+4+5 7
1 2 − 3i − 4 + 5i
 = −1 + i .
 
= 
2  2−3+4−5   −1 
2 + 3i − 4 − 5i −1 − i

Exercise 2.7: Exact form of low-order DFT matrix

As in Example 2.3 in the book, state the exact cartesian form of the Fourier
matrix for the cases N = 6, N = 8, and N = 12.

Solution. For N = 6 the entries are on the form √1 e−2πink/6 = √1 e−πink/3 .

6 6
This means that the entries in the Fourier matrix are the numbers √1 e−πi/3 =
√ √ 6
√1 (1/2 − i 3/2), √1 e−2πi/3 = √1 (−1/2 − i 3/2), and so on. The matrix is
6 6 6
thus

20
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS21

 
1 1√ 1 √ 1 1 √ 1√
1
 1/2 − i √3/2 −1/2 − i√3/2 −1 −1/2 + i√3/2 1/2 + i √2/2 

1  1 −1/2 − i 3/2 −1/2 + i 3/2 1 −1/2 − i 3/2 +1/2 − i 3/2
F6 = √  
6 1
 −1 √ 1 √ −1 1 √ −1 √ 

1 −1/2 + √i 3/2 −1/2 − i√3/2 1 −1/2 + i√3/2 −1/2 − √i 3/2
1 1/2 + i 2/2 −1/2 + i 3/2 −1 −1/2 − i 3/2 1/2 − i 3/2

The cases N = 8 and N = 12 follow similarly, but are even more tedious. For
N = 8 the entries are √18 eπink/4 , which can be expressed exactly since we can
express exactly any sines and cosines of a multiple of π/4. For N = 12 we get
the base angle π/6, for which we also have exact values for sines and cosines for
all multiples.

Exercise 2.8: DFT of a delayed vector

We have a real vector x with length N , and define the vector z by delaying
all elements in x with 5 cyclically, i.e. z5 = x0 , z6 = x1 ,. . . ,zN −1 = xN −6 ,
and z0 = xN −5 ,. . . ,z4 = xN −1 . For a given n, if |(FN x)n | = 2, what is then
|(FN z)n |? Justify the answer.

Solution. z is the vector x delayed with d = 5 samples, and then Property

3 of Theorem 2.7 in the book gives us that (FN z)n = e−2πi5k/N (FN x)n . In
particular |(FN z)n | = |(FN x)n | = 2, since |e−2πi5k/N | = 1.

Exercise 2.9: Using symmetry property

Given a real vector x of length 8 where (F8 (x))2 = 2 − i, what is (F8 (x))6 ?

Solution. By Theorem 2.7 in the book we know that (FN (x))N −n = (FN (x))n
when x is a real vector. If we set N = 8 and n = 2 we get that (F8 (x))6 =
(F8 (x))2 = 2 − i = 2 + i.

Exercise 2.10: DFT of cos2 (2πk/N )

Let x be the vector of length N where xk = cos2 (2πk/N ). What is then FN x?

Solution. The idea is to express x as a linear combination of the Fourier basis

vectors φn , and use that FN φn = en . We have that
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS22

 
1 2πik/N 2
2 −2πikn/N
cos (2πk/N ) = e +e
2
1 1 1 1 1 1
= e2πi2k/N + + e−2πi2k/N = e2πi2k/N + + e2πi(N −2)k/N
4  2 4 4 2 4
√

1 1 1
= N φ2 + φ0 + φN −2 .
4 2 4

We here used the periodicity of e2πikn/N , i.e. that e−2πi2k/N = e2πi(N −2)k/N .
Since FN is linear and FN (φn ) = en , we have that

√ √
 
1 1 1
FN (x) = N e2 + e0 + eN −2 = N (1/2, 0, 1/4, 0, . . . , 0, 1/4, 0) .
4 2 4

Exercise 2.11: DFT of ck x

Let x be the vector with entries xk = ck . Show that the DFT of x is given by
the vector with components

1 − cN
yn =
1 − ce−2πin/N
for n = 0, . . . , N − 1.

Solution. We get

N
X −1 N
X −1
k −2πink/N
yn = c e = (ce−2πin/N )k
k=0 k=0
1 − (ce−2πin/N )N 1 − cN
= = .
1 − ce−2πin/N 1 − ce−2πin/N

Exercise 2.12: Rewrite a complex DFT as real DFT’s

If x is complex, Write the DFT in terms of the DFT on real sequences.

Hint. Split into real and imaginary parts, and use linearity of the DFT.

Exercise 2.13: DFT implementation

Extend the code for the function DFTImpl in Example 2.4 in the book so that

• The function also takes a second parameter called forward. If this is true
the DFT is applied. If it is false, the IDFT is applied. If this parameter is
not present, then the forward transform should be assumed.
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS23

• If the input x is two-dimensional (i.e. a matrix), the DFT/IDFT should be

applied to each column of x. This ensures that, in the case of sound, the
FFT is applied to each channel in the sound when the enrire sound is used
as input, as we are used to when applying different operations to sound.

Also, write documentation for the code.

Solution. The code can look like this:

def DFTImpl(x, forward=True):

"""
Compute the DFT of the vector x using standard matrix
multiplication. To avoid out of memory situations, we do not
allocate the entire DFT matrix, only one row of it at a time.
Note that this function differs from the FFT in that it includes
the normalizing factor 1/sqrt(N). The DFT is computed along axis
0. If there is another axis, the DFT is computed for each element
in this as well.
x: a vector
forward: Whether or not this is forward (i.e. DFT)
or reverse (i.e. IDFT)
"""
y = zeros_like(x).astype(complex)
N = len(x)
sign = -(2*forward - 1)
if ndim(x) == 1:
for n in xrange(N):
D = exp(sign*2*pi*n*1j*arange(float(N))/N)
y[n] = dot(D, x)
else:
for n in range(N):
D = exp(sign*2*pi*n*1j*arange(float(N))/N)
for s2 in xrange(shape(x)[1]):
y[n,s2] = dot(D,x[:, s2])
if sign == 1:
y /= float(N)
return y

Exercise 2.14: Symmetry

Assume that N is even.
a) Show that, if xk+N/2 = xk for all 0 ≤ k < N/2, then yn = 0 when n is odd.

Solution. We have that

CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS24

 
N/2−1 N −1
1  X X
yn = √ xk e−2πikn/N + xk e−2πikn/N 
N k=0 k=N/2
 
N/2−1 N/2−1
1  X X
=√ xk e−2πikn/N + xk e−2πi(k+N/2)n/N 
N k=0 k=0
N/2−1
1 X
=√ xk (e−2πikn/N + (−1)n e−2πikn/N )
N k=0
N/2−1
1 X
= (1 + (−1)n ) √ xk e−2πikn/N
N k=0
If n is odd, we see that yn = 0.
b) Show that, if xk+N/2 = −xk for all 0 ≤ k < N/2, then yn = 0 when n is
even.

Solution. The proof is the same as in a), except for a sign change.
c) Show also the converse statements in a) and b).

Solution. Clearly the set of vectors which satisfies xk+N/2 = ±xk is a vector
space V of dimension N/2. The set of vectors where every second component is
zero is also a vector space of dimension N/2, let us denote this by W . We have
shown that FN (V ) ⊂ W , but since FN is unitary, FN (V ) also has dimension
N/2, so that FN (V ) = W . This shows that when every second yn is 0, we must
have that xk+N/2 = ±xk , and the proof is done.
d) Also show the following:
• xn = 0 for all odd n if and only if yk+N/2 = yk for all 0 ≤ k < N/2.
• xn = 0 for all even n if and only if yk+N/2 = −yk for all 0 ≤ k < N/2.

Solution. In the proofs above, compute the IDFT instead.

Exercise 2.15: DFT on complex and real data

Let x1 , x2 be real vectors, and set x = x1 + ix2 . Use Theorem 2.7 in the book
to show that

1 
(FN (x1 ))k = (FN (x))k + (FN (x))N −k
2
1  
(FN (x2 ))k = (FN (x))k − (FN (x))N −k
2i
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS25

This shows that we can compute two DFT’s on real data from one DFT on
complex data, and 2N extra additions.

Solution. We have that

(FN (x))k = (FN (x1 + ix2 ))k = (FN (x1 ))k + i(FN (x2 ))k
(FN (x))N −k = (FN (x1 ))N −k + i(FN (x2 ))N −k = (FN (x1 ))k + i(FN (x2 ))k ,

where we have used Property 1 of Theorem 2.7 in the book. If we take the
complex conjugate in the last equation, we are left with the two equations

(FN (x))k = (FN (x1 ))k + i(FN (x2 ))k

(FN (x))N −k = (FN (x1 ))k − i(FN (x2 ))k .

If we add these we get

1 
(FN (x1 ))k = (FN (x))k + (FN (x))N −k ,
2
which is the first equation. If we instead subtract the equations we get
1  
(FN (x2 ))k = (FN (x))k − (FN (x))N −k ,
2i
which is the second equation

Exercise 2.19: Comment code

Explain what the code below does, line by line:

x = x[0:2**17]
y = fft.fft(x, axis=0)
y[(2**17/4):(3*2**17/4)] = 0
newx = abs(fft.ifft(y))
newx /= abs(newx).max()
play(newx, fs)

Comment in particular why we adjust the sound samples by dividing with the
maximum value of the sound samples. What changes in the sound do you expect
to hear?

Solution. First a sound file is read. We then restrict to the first 212 sound
samples, perform a DFT, zero out the frequencies which correspond to DFT-
indices between 210 and 212 − 210 − 1, and perform an IDFT. Finally we scale
the sound samples so that these lie between −1 and 1, which is the range we
demand for the sound samples, and play the new sound.
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS26

Exercise 2.20: Which frequency is changed?

In the code from the previous exercise it turns out that fs = 44100Hz, and that
the number of sound samples is N = 292570. Which frequencies in the sound
file will be changed on the line where we zero out some of the DFT coefficients?

Solution. As we have seen, DFT index n corresponds to frequency ν = nfs /N .

Above N = 217 , so that we get the connection ν = nfs /N = n × 44100/217 .
We zeroed the DFT indices above n = 215 , so that frequencies above ν =
215 × 44100/217 = 11025Hz are affected.

Exercise 2.21: Implement interpolant

Implement code where you do the following:
• at the top you define the function f (x) = cos6 (x), and M = 3,
• compute the unique interpolant from VM,T (i.e. by taking N = 2M + 1
samples over one period), as guaranteed by Proposition 2.9 in the book,
• plot the interpolant against f over one period.
Finally run the code also for M = 4, M = 5, and M = 6. Explain why the plots
coincide for M = 6, but not for M < 6. Does increasing M above M = 6 have
any effect on the plots?

Solution. The code can look as follows.

import matplotlib.pyplot as plt

from numpy import *
f = lambda t:cos(t)**6
M = 5
T = 2*pi
N = 2*M + 1
t = linspace(0, T, 100)
x = f(linspace(0, T - T/float(N), N))
y = fft.fft(x, axis=0)/N
s = real(y[0])*ones(len(t))
for k in range(1,(N+1)/2):
s += 2*real(y[k]*exp(2*pi*1j*k*t/float(T)))
plt.plot(t, s, ’r’, t, f(t), ’g’)
plt.legend([’Interpolant from V_{M,T}’,’f’])

Exercise 2.22: Extra results for the FFT when N = N1 N2

When N is composite, there are a couple of results we can state regarding
polyphase components.
a) Assume that N = N1 N2 , and that x ∈ RN satisfies xk+rN1 = xk for all k, r,
i.e. x has period N1 . Show that yn = 0 for all n which are not a multiplum of
N2 .
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS27

Solution. We have that x(p) is a constant vector of length N2 for 0 ≤ p < N1 .

But then the DFT of all the x(p) has zero outside entry zero. Multiplying with
e−2πikn/N does not affect this. The last N2 − 1 rows are thus zero before the
final DFT is applied, so that these rows are zero also after this final DFT. After
assembling the polyphase components again we have that yrN2 are the only
nonzero DFT-coefficients.
b) Assume that N = N1 N2 , and that x(p) = 0 for p = 6 0. Show that the
polyphase components y (p) of y = DFTN x are constant vectors for all p.
But what about the case when N is a prime number? Rader’s algorithm
[?] handles this case by expressing a DFT with N a prime number in terms of
DFT’s of length N − 1 (which is not a prime number). Our previous scenario can
then be followed, but stops quickly again if N − 1 has prime factors of high order.
Since there are some computational penalties in applying Rader’s algorithm, it
may be inefficient some cases. Winograd’s FFT algorithm [?] extends Rader’s
algorithm to work for the case when N = pr . This algorithm tends to reduce
the number of multiplications, at the price of an increased number of additions.
It is difficult to program, and is rarely used in practice.

Exercise 2.23: Extend implementation

Recall that, in Exercise 2, we extended the direct DFT implementation so that it
accepted a second parameter telling us if the forward or reverse transform should
be applied. Extend the general function and the standard kernel in the same
way. Again, the forward transform should be used if the forward parameter is
not present. Assume also that the kernel accepts only one-dimensional data,
and that the general function applies the kernel to each column in the input if
the input is two-dimensional (so that the FFT can be applied to all channels
in a sound with only one call). The signatures for our methods should thus be
changed as follows:

def FFTImpl(x, FFTKernel, forward = True):

def FFTKernelStandard(x, forward):

It should be straightforward to make the modifications for the reverse transform

by consulting the second part of Theorem 2.18 in the book. For simplicity,
let FFTImpl take care of the additional division with N we need to do in case
of the IDFT. In the following we will assume these signatures for the FFT
implementation and the corresponding kernels.

Solution. The functions can be implemented as follows:

def FFTImpl(x, FFTKernel, forward = True):

"""
Compute the FFT or IFFT of the vector x. Note that this function
differs from the DFT in that the normalizing factor 1/sqrt(N) is
not included. The FFT is computed along axis 0. If there is
another axis, the FFT is computed for each element in this as
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS28

well. This function calls a kernel for computing the FFT. The
kernel assumes that the input has been bit-reversed, and contains
only one axis. This function is where the actual bit reversal and
the splitting of the axes take place.
x: a vector
FFTKernel: can be any of FFTKernelStandard, FFTKernelNonrec, and
FFTKernelSplitradix. The kernel assumes that the input has been
bit-reversed, and contains only one axis.
forward: Whether the FFT or the IFFT is applied
"""
if ndim(x) == 1:
bitreverse(x)
FFTKernel(x, forward)
else:
bitreversearr(x)
for s2 in xrange(shape(x)[1]):
FFTKernel(x[:, s2], forward)
if not forward:
x /= len(x)

def FFTKernelStandard(x, forward):

"""
Compute the FFT of x, using a standard FFT algorithm.
x: a bit-reversed version of the input. Should have only one axis
forward: Whether the FFT or the IFFT is applied
"""
N = len(x)
sign = -1
if not forward:
sign = 1
if N > 1:
xe, xo = x[0:(N/2)], x[(N/2):]
FFTKernelStandard(xe, forward)
FFTKernelStandard(xo, forward)
D = exp(sign*2*pi*1j*arange(float(N/2))/N)
xo *= D
x[:] = concatenate([xe + xo, xe - xo])

Exercise 2.24: Compare execution time

In this exercise we will compare execution times for the different methods for
computing the DFT.
a) Write code which compares the execution times for an N -point DFT for the
following three cases: Direct implementation of the DFT (as in Example 2.4
in the book), the FFT implementation used in this chapter, and the built-in
fft-function. Your code should use the sample audio file castanets.wav, apply
the different DFT implementations to the first N = 2r samples of the file for
r = 3 to r = 15, store the execution times in a vector, and plot these. You can
use the function time() in the time module to measure the execution time.
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS29

b) A problem for large N is that there is such a big difference in the execution
times between the two implementations. We can address this by using a loglog-
plot instead. Plot N against execution times using the function loglog. How
should the fact that the number of arithmetic operations are 8N 2 and 5N log2 N
be reflected in the plot?

Solution. The two different curves you see should have a derivative approxi-
mately equal to one and two, respectively.
c) It seems that the built-in FFT is much faster than our own FFT implemen-
tation, even though they may use similar algorithms. Try to explain what can
be the cause of this.

Solution. There may be several reasons for this. One is that Python code
runs slowly when compared to native code, which is much used in the built-in
FFT. Also, the built-in fft has been subject to much more optimization than
we have covered here.

Solution. The code can look as follows.

x0, fs = audioread(’sounds/castanets.wav’)
kvals = arange(3,16)
slowtime = zeros(len(kvals))
fasttime = zeros(len(kvals))
fastesttime = zeros(len(kvals))
N = 2**kvals
for k in kvals:
x = x0[0:2**k].astype(complex)
start = time()
DFTImpl(x)
slowtime[k - kvals[0]] = time() - start
start = time()
FFTImpl(x, FFTKernelStandard)
fasttime[k - kvals[0]] = time() - start
start = time()
fft.fft(x, axis=0)
fastesttime[k - kvals[0]] = time() - start
# a.
plt.plot(kvals, slowtime, ’ro-’, \
kvals, fasttime, ’go-’, \
kvals, fastesttime, ’bo-’)
plt.grid(’on’)
plt.title(’time usage of the DFT methods’)
plt.legend([’DFT’, ’Standard FFT’, ’Built-in FFT’])
plt.xlabel(’log_2 N’)
plt.ylabel(’time used [s]’)
plt.show()
plt.figure()
# b.
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS30

plt.loglog(N, slowtime, ’ro-’, N, fasttime, ’go-’, N, fastesttime, ’bo-’)

plt.axis(’equal’)
plt.legend([’DFT’, ’Standard FFT’, ’Built-in FFT’])

Exercise 2.25: Combine two FFT’s

Let x1 = (1, 3, 5, 7) and x2 = (2, 4, 6, 8). Compute DFT4 x1 and DFT4 x2 . Ex-
plain how you can compute DFT8 (1, 2, 3, 4, 5, 6, 7, 8) based on these computations
(you don’t need to perform the actual computation). What are the benefits of
this approach?

Solution. We get

    
1 1 1 1 1 16
1 −i −1 i  3 = −4 + 4i
   
DFT4 x1 = 
1 −1 1 −1 5  −4 
1 i −1 −i 7 −4 − 4i
    
1 1 1 1 2 20
1 −i −1 i  4 = −4 + 4i
   
DFT4 x2 = 
1 −1 1 −1 6  −4 
1 i −1 −i 8 −4 − 4i

In the FFT-algorithm we split the computation of DFT4 (x) into the computation
of DFT2 (x(e) ) and DFT2 (x(o) ), where x(e) and x(o) are vectors of length 4 with
even-indexed and odd-indexed components, respectively. In this case we have
x(e) = (1, 3, 5, 7) and x(o) = (2, 4, 6, 8). In other words, the FFT-algorithm uses
the FFT-computations we first made, so that we can save computation. The
benefit of using the FFT-algorithm is that we save computations, so that we end
up with O(5N log2 N ) real arithmetic operations.

Exercise 2.26: FFT operation count

When we wrote down the difference equation for the number of multiplications in
the FFT algorithm, you could argue that some multiplications were not counted.
Which multiplications in the FFT algorithm were not counted when writing down
this difference equation? Do you have a suggestion to why these multiplications
were not counted?

Solution. When we compute e−2πin/N , we do some multiplications/divisions

in the exponent. These are not counted because they do not depend on x, and
may therefore be precomputed.
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS31

Exercise 2.27: FFT algorithm adapted to real data

In this exercise we will look at how we can adapt an FFT algorithm to real input.
Since yN −n = yn for real input, there is no additional complexity in computing
the second half of the DFT coefficients, once the first half has been computed.
We will now rewrite Equation (2.13) in the book for indices n and N/2 − n as

yn = (DFTN/2 x(e) )n + e−2πin/N (DFTN/2 x(o) )n

yN/2−n = (DFTN/2 x(e) )N/2−n + e−2πi(N/2−n)/N (DFTN/2 x(o) )N/2−n
= (DFTN/2 x(e) )n − e2πin/N (DFTN/2 x(o) )n
= (DFTN/2 x(e) )n − e−2πin/N (DFTN/2 x(o) )n .

We see here that, if we already have computed DFTN/2 x(e) and DFTN/2 x(o) ,
we need one additional complex multiplication for each yn with 0 ≤ n < N/4
(since e−2πin/N and (DFTN/2 x(o) )n are complex). No further multiplications
are needed in order to compute yN/2−n , since we simply conjugate terms before
adding them. Again yN/2 must be handled explicitly with this approach. For
this we can use the formula

yN/2 = (DFTN/2 x(e) )0 − (DN/2 DFTN/2 x(o) )0

instead.
a) Conclude from this that an FFT algorithm adapted to real data at each step
requires N/4 complex additions and N/2 complex additions. Conclude from this
as before that an algorithm based on real
data requires MN = O(N log2 N ) real
multiplications and AN = O 32 N log2 N real additions (i.e. half the operation
count for complex input).

Solution. The formulas at the top show that we need only compute the
e−2πin/N (DFTN/2 x(o) )n for 0 ≤ n < N/4, i.e. N/4 complex additions. Since
N/2 formulas are listed, N/2 complex additions are needed. This translates to
N real multiplications, and N/2 + N = 3N/2 real additions. These are exactly
half the numbers we obtained for the complex version of the algorithm, for
which we obtained MN = O(2N log2 (N )) and AN = O(3N log2 (N )). It is then
not difficult to see that the real version
will require MN = O(N log2 N ) real
multiplications and AN = O 32 N log2 N real additions.
b) Find an IFFT algorithm adapted to vectors with conjugate symmetry, which
has the same operation count as this FFT algorithm adapted to real data.

Hint. Consider the vectors z, w with entries zn = yn + yN/2−n ∈ RN/2 and

wn = e2πin/N (yn − yN/2−n ) ∈ RN/2 . From the equations above, how can these
be used in an IFFT?
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS32

Solution. Note that

zN/2−n = yN/2−n + yn = yn + yN/2−n = zn

wN/2−n = e2πi(N/2−n)/N (yN/2−n − yn ) = −e−2πin/N (yN/2−n − yn )
= e2πin/N (yn − yN/2−n ) = wn ,

showing that the vectors z and w also have conjugate symmetry. Note also that,
if we add and subtract the equations at the top, we obtain

2(DFTN/2 x(e) )n = yn + yN/2−n = zn

2(DFTN/2 x(o) )n = e2πin/N (yn − yN/2−n ) = wn .

so that

1 1
x(e) = IDFTN/2 z x(o) = IDFTN/2 w.
2 2
The point is that the number of arithmetic operations in computing the vectors
z and w is the same as the FFT in a), so that the overall complexity is the same.

Exercise 2.28: Non-recursive FFT algorithm

Use the factorization in (2.18) in the book to write a kernel function FFTKernelNonrec
for a non-recursive FFT implementation. In your code, perform the matrix
multiplications in Equation (2.18) in the book from right to left in an (outer)
for-loop. For each matrix loop through the different blocks on the diagonal in an
(inner) for-loop. Make sure you have the right number of blocks on the diagonal,
each block being on the form
 
I DN/2k
.
I −DN/2k
It may be a good idea to start by implementing multiplication with such a simple
matrix first as these are the building blocks in the algorithm (also attempt to do
this so that everything is computed in-place). Also compare the execution times
with our original FFT algorithm, as we did in Exercise 2, and try to explain
what you see in this comparison.

Solution. The algorithm for the non-recursive FFT can look as follows

def FFTKernelNonrec(x, forward):

"""
Compute the FFT of x, using a non-recursive FFT algorithm.
x: a bit-reversed version of the input. Should have only one axis
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS33

forward: Whether the FFT or the IFFT is applied

"""
N = len(x)
sign = -1
if not forward:
sign = 1
D = exp(sign*2*pi*1j*arange(float(N/2))/N)
nextN = 1
while nextN < N:
k = 0
while k < N:
xe, xo = x[k:(k + nextN)], x[(k + nextN):(k + 2*nextN)]
xo *= D[0::(N/(2*nextN))]
x[k:(k+2*nextN)] = concatenate([xe + xo, xe - xo])
k += 2*nextN
nextN *= 2

If you add the non-recursive algorithm to the code from Exercise 2, you will see
that the non-recursive algorithm performs much better. There may be several
reasons for this. First of all, there are no recursive function calls. Secondly,
the values in the matrices DN/2 are constructed once and for all with the non-
recursive algorithm. Code which compares execution times for the original FFT
algorithm, our non-recursive implementation, and the split-radix algorithm of
the next exercise, can look as follows:

x0, fs = audioread(’sounds/castanets.wav’)
kvals = arange(3,16)
slowtime = zeros(len(kvals))
fasttime = zeros(len(kvals))
fastesttime = zeros(len(kvals))
N = 2**kvals
for k in kvals:
x = x0[0:2**k].astype(complex)
start = time()
FFTImpl(x, FFTKernelStandard)
slowtime[k - kvals[0]] = time() - start
start = time()
FFTImpl(x, FFTKernelNonrec)
fasttime[k - kvals[0]] = time() - start
start = time()
FFTImpl(x, FFTKernelSplitradix)
fastesttime[k - kvals[0]] = time() - start
plt.plot(kvals, slowtime, ’ro-’, \
kvals, fasttime, ’bo-’, \
kvals, fastesttime, ’go-’)
plt.grid(’on’)
plt.title(’time usage of the DFT methods’)
plt.legend([’Standard FFT algorithm’, \
’Non-recursive FFT’, \
’Split radix FFT’])
plt.xlabel(’log_2 N’)
plt.ylabel(’time used [s]’)
plt.show()
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS34

Exercise 2.29: The Split-radix FFT algorithm

In this exercise we will develop a variant of the FFT algorithm called the split-
radix FFT algorithm, which until recently held the record for the lowest operation
count for any FFT algorithm.
We start by splitting the rightmost DFTN/2 in Equation (2.17) in the book
by using this equation again, to obtain

  
DFTN/4 DN/4 DFTN/4  (e) 
DFTN/2 DN/2 x
DFT N/4 −D N/4 DFT N/4   (oe) 

DFTN x =   x .
 DFTN/4 DN/4 DFTN/4 
DFTN/2 −DN/2 x(oo)
DFTN/4 −DN/4 DFTN/4
(2.1)
The term radix describes how an FFT is split into FFT’s of smaller sizes, i.e. how
the sum in an FFT is split into smaller sums. The FFT algorithm we started
this section with is called a radix 2 algorithm, since it splits an FFT of length
N into FFT’s of length N/2. If an algorithm instead splits into FFT’s of length
N/4, it is called a radix 4 FFT algorithm. The algorithm we go through here is
called the split radix algorithm, since it uses FFT’s of both length N/2 and N/4.
a) Let GN/4 be the 
(N/4)×(N/4) diagonal
 matrix with e−2πin/N on the diagonal.
GN/4 0
Show that DN/2 = .
0 −iGN/4
b) Let HN/4 be the (N/4) × (N/4) diagonal matrix GD/4 DN/4 . Verify the
following rewriting of Equation (2.1):

  
GN/4 DFTN/4 HN/4 DFTN/4  (e) 
DFTN/2 x
−iG N/4 DFT N/4 iH N/4 DFT N/4   (oe) 

DFTN x =   x
 −GN/4 DFTN/4 −HN/4 DFTN/4 
DFTN/2 x(oo)
iGN/4 DFTN/4 −iHN/4 DFTN/4
 
I 0 GN/4 HN/4    (e) 
0 I −iGN/4 iHN/4  DFTN/2 0 0 x
=  0 DFT N/4 0  x(oe) 
I 0 −GN/4 −HN/4 
0 0 DFTN/4 x(oo)
0 I iGN/4 −iHN/4
   
GN/4 HN/4
DFTN/2 x(e)
 
I −iG iH
 N/4 N/4  

= DFTN/4 x(oe) 
 GN/4 HN/4 
I − DFTN/4 x(oo)
−iGN/4 iHN/4
GN/4 DFTN/4 x(oe) + HN/4 DFTN/4 x(oo)

  
(e)
DFTN/2 x + −i GN/4 DFTN/4 x(oe) − HN/4 DFTN/4 x(oo) 
= 
GN/4 DFTN/4 x(oe) + HN/4 DFTN/4 x(oo)


 (e)
DFTN/2 x −
−i GN/4 DFTN/4 x(oe) − HN/4 DFTN/4 x(oo)
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS35

c) Explain from the above expression why, once the three FFT’s above have
been computed, the rest can be computed with N/2 complex multiplications,
and 2 × N/4 + N = 3N/2 complex additions. This is equivalent to 2N real
multiplications and N + 3N = 4N real additions.

Hint. It is important that GN/4 DFTN/4 x(oe) and HN/4 DFTN/4 x(oo) are com-
puted first, and the sum and difference of these two afterwards.
d) Due to what we just showed, our new algorithm leads to real multiplication
and addition counts which satisfy

MN = MN/2 + 2MN/4 + 2N AN = AN/2 + 2AN/4 + 4N

Find the general solutions
to these difference equations

and conclude from these
that MN = O 34 N log2 N , and AN = O 83 N log2 N . The operation count is
thus O (4N log2 N ), which is a reduction of N log2 N from the FFT algorithm.
e) Write an FFT kernel function FFTKernelSplitradix for the split-radix
algorithm (again this should handle both the forward and reverse transforms).
Are there more or less recursive function calls in this function than in the
original FFT algorithm? Also compare the execution times with our original
FFT algorithm, as we did in Exercise 2. Try to explain what you see in this
comparison.

Solution. If you add the split-radix FFT algorithm also to the code from
Exercise 2, you will see that it performs better than the FFT algorithm, but
worse than the non-recursive algorithm. That it performs better than the FFT
algorithm is as expected, since it has a reduced number of arithmetic operations,
and also a smaller number of recursive calls. It is not surprising that the non-
recursive function performs better, since only that function omits recursive calls,
and computes the values in the diagonal matrices once and for all.
By carefully examining the algorithm we have developed, one can reduce the
operation count to 4N log2 N − 6N + 8. This does not reduce the order of the
algorithm, but for small N (which often is the case in applications) this reduces
the number of operations considerably, since 6N is large compared to 4N log2 N
for small N . In addition to having a lower number of operations than the FFT
algorithm of Theorem 2.15 in the book, a bigger percentage of the operations
are additions for our new algorithm: there are now twice as many additions
than multiplications. Since multiplications may be more time-consuming than
additions (depending on how the CPU computes floating-point arithmetic), this
can be a big advantage.

Solution. The code for the split-radix algorithm can look as follows
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS36

def FFTKernelSplitradix(x, forward):

"""
Compute the FFT of x, using the split-radix FFT algorithm.
x: a bit-reversed version of the input. Should have only one axis
forward: Whether the FFT or the IFFT is applied
"""
N = len(x)
sign = -1
if not forward:
sign = 1
if N == 2:
x[:] = [x[0] + x[1], x[0] - x[1]]
elif N > 2:
xe, xo1, xo2 = x[0:(N/2)], x[(N/2):(3*N/4)], x[(3*N/4):N]
FFTKernelSplitradix(xe, forward)
FFTKernelSplitradix(xo1, forward)
FFTKernelSplitradix(xo2, forward)
G = exp(sign*2*pi*1j*arange(float(N/4))/N)
H = G*exp(sign*2*pi*1j*arange(float(N/4))/(N/2))
xo1 *= G
xo2 *= H
xo = concatenate( [xo1 + xo2, -sign*1j*(xo2 - xo1)] )
x[:] = concatenate([xe + xo, xe - xo])

Exercise 2.30: Bit-reversal

In this exercise we will make some considerations which will help us explain the
code for bit-reversal. This is perhaps not a mathematically challenging exercise,
but nevertheless a good exercise in how to think when developing an efficient
algorithm. We will use the notation i for an index, and j for its bit-reverse. If
we bit-reverse k bits, we will write N = 2k for the number of possible indices.
a) Consider the following code

j = 0
for i in range(N-1):
print j
m = N/2
while (m >= 1 and j >= m):
j -= m
m /= 2
j += m

Explain that the code prints all numbers in [0, N − 1] in bit-reversed order (i.e. j).
Verify this by running the program, and writing down the bits for all numbers
for, say N = 16. In particular explain the decrements and increments made to
the variable j. The code above thus produces pairs of numbers (i, j), where j is
the bit-reverse of i. As can be seen, bitreverse applies similar code, and then
swaps the values xi and xj in x, as it should.

Solution. Note that, if the bit representation of i ends with 0 1| .{z

. . 1}, then
n
i + 1 has a bit representation which ends with 1 0| .{z
. . 0}, with the remaining first
n
CHAPTER 2. DIGITAL SOUND AND DISCRETE FOURIER ANALYSIS37

bits unaltered. Clearly the bit-reverse of i then starts with 1| .{z

. . 1} 0 and the
n
bit-reverse of i + 1 starts with 1 0| .{z
. . 0}. We see that the bit reverse of i + 1 can
n
be obtained from the bit-reverse of i by replacing the first consecutive set of ones
by zeros, and the following zero by one. This is performed by the line above
where j is decreased by m: Decreasing j by N/2 when j ≥ N/2 changes the first
bit from 1 to 0, and similarly for the next n bits. The line where j is increased
with m changes bit number n + 1 from 0 to 1.
Since bit-reverse is its own inverse (i.e. P 2 = I), it can be performed by
swapping elements i and j. One way to secure that bit-reverse is done only once,
is to perform it only when j > i. You see that bitreverse includes this check.
b) Explain that N − j − 1 is the bit-reverse of N − i − 1. Due to this, when
i, j < N/2, we have that N − i − 1, N − j − l ≥ N/2, and that bitreversal can
swap them. Moreover, all swaps where i, j ≥ N/2 can be performed immediately
when pairs where i, j < N/2 are encountered. Explain also that j < N/2
if and only if i is even. In the code you can see that the swaps (i, j) and
(N − i − 1, N − j − 1) are performed together when i is even, due to this.

Solution. Clearly N − i − 1 has a bit representation obtained by changing

every bit in i. That N − j − 1 is the bit-reverse of N − i − 1 follows immediately
from this. If i is even, the least significant bit is 0. After bit-reversal, this
becomes the most significant bit, and the most significant bit of j is 0 which is
the case if and only if j < N/2.
c) Assume that i < N/2 is odd. Explain that j ≥ N/2, so that j > i. This says
that when i < N/2 is odd, we can always swap i and j (this is the last swap
performed in the code). All swaps where 0 ≤ j < N/2 and N/2 ≤ j < N can be
performed in this way.

Solution. If i < N/2 is odd, then the least significant bit is 1. This means
that the most significant bit of j is 1, so that j ≥ N/2, so that j > i.
In bitreversal, you can see that the bit-reversal of 2r and 2r +1 are handled
together (i.e. i is increased with 2 in the for-loop). The effect of this is that the
number of if-tests can be reduced, due to the observations from b) and c).
Chapter 3

Operations on digital sound:

digital filters

Exercise 3.4: Finding the filter coefficients and the matrix

Assume that the filter S is defined by the formula
1 1 1 1
zn = xn+1 + xn + xn−1 + xn−2 .
4 4 4 4
Write down the filter coefficients tk , and the matrix for S when N = 8.

Solution. Here we have that t−1 = 1/4, t0 = 1/4, t1 = 1/4, and t2 = 1/4. We
now get that s0 = t0 = 1/4, s1 = t1 = 1/4, and s2 = t2 = 1/4 (first formula),
and sN −1 = s7 = t−1 = 1/4 (second formula). This means that the matrix of S
is
 
1 1 0 0 0 0 1 1
1 1 1 0 0 0 0 1
 
1 1 1 1 0 0 0 0
 
1 0 1 1 1 1 0 0 0
S=   .
4 0 0 1 1 1 1 0 0 
0 0 0 1 1 1 1 0
 
0 0 0 0 1 1 1 1
1 0 0 0 0 1 1 1

Exercise 3.5: Finding the filter coefficients from the matrix

Given the circulant Toeplitz matrix
 
1 2 0 0
0 1 2 0
S= 0
,
0 1 2
2 0 0 1

38
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS39

write down the filter coefficients tk .

Exercise 3.6: Convolution and polynomials

Compute the convolution of {1, 2, 1} with itself. interpret the result in terms of
two polynomials.

Exercise 3.7: Implementation of convolution

Implement code which computes t ∗ x in the two ways described after Equation
(3.3) in the book, i.e. as a double for loop, and as a simple for loop in k, with n
vectorized. As your t, take k randomly generated numbers. Compare execution
times for these two methods and the convolve function, for different values of
k. Present the result as a plot where k runs along the x-axis, and execution
times run along the y-axis. Your result will depend on how Python performs
vectorization.

Solution. The code can look as follows.

from sound import *

from time import *
from numpy import *
import matplotlib.pyplot as plt
x, fs = audioread(’sounds/castanets.wav’)
x = x[:,0]
N= len(x)
kmax=100
vals1 = zeros(kmax/10)
vals2 = zeros(kmax/10)
vals3 = zeros(kmax/10)
ind = 0
for k in range(10, kmax+1,10):
t = random.random(k)
start = time()
convolve(t, x)
vals1[ind] = time() - start
z = zeros(N)
start = time()
for s in range(k):
z[(k-1):N] += t[s]*x[(k-1-s):(N-s)]
vals2[ind] = time() - start
z = zeros(N)
start = time()
for n in range(k-1, N):
for s in range(k):
z[n] += t[s]*x[n-s]
vals3[ind] = time() - start
ind += 1
plt.plot( range(10, kmax+1,10),log(vals1), ’r-’, range(10, kmax+1,10), log(vals2), ’g-’, range(10,
plt.legend([’conv’,’simple for’,’double for’])
plt.show()
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS40

Exercise 3.8: Filters with a different number of coefficients

with positive and negative indices
Assume that S = {t−E , . . . , t0 , . . . , tF }. Formulate a generalization of Proposi-
tion 3.5 in the book for such filters, i.e. to filters where there may be a different
number of filter coefficients with positive and negative indices. You should only
need to make some small changes to the proof of Proposition 3.5 in the book to
achieve this.

Exercise 3.9: Implementing filtering with convolution

Implement a function filterS which uses Proposition 3.5 in the book and the
convolve function Sx when S = {t−L , . . . , t0 , . . . , tL } The function should take
the vectors (t−L , . . . , t0 , . . . , tL ) and x as input.

Solution. The code can look like this:

def filterS(t, x):

L = (len(t) - 1)/2
N = len(x)
y = concatenate([ x[(N - L):], x, x[:L]])
y = convolve(t, y)
y = y[(2*L):(len(y)-2*L)]

Exercise 3.13: Time reversal is not a filter

In Example 1.2 in the book we looked at time reversal as an operation on digital
sound. In RN this can be defined as the linear mapping which sends the vector
ek to eN −1−k for all 0 ≤ k ≤ N − 1.
a) Write down the matrix for the time reversal linear mapping, and explain
from this why time reversal is not a digital filter.

Solution. The matrix for time reversal is the matrix

 
0 0 ··· 0 1
0
 0 · · · 1 0 
 .. .. .. .. .. 
. . . . .
 
0 1 · · · 0 0
1 0 ··· 0 0
This is not a circulant Toeplitz matrix, since all diagonals assume the values 0
and 1, so that they are not constant on each diagonal. Time reversal is thus not
a digital filter.
b) Prove directly that time reversal is not a time-invariant operation.
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS41

Solution. Let S denote time reversal. Clearly Se1 = eN −2 . If S was time-

invariant we would have that Se0 = eN −3 , where we have delayed the input and
output. But this clearly is not the case, since by definition Se0 = eN −1 .

Exercise 3.14: When is a filter symmetric?

Let S be a digital filter. Show that S is symmetric if and only if the frequency
response satisfies λS,n = λS,N −n for all n.

Exercise 3.15: Eigenvectors and eigenvalues

Consider the matrix
 
4 1 3 1
1 4 1 3
S=
3
.
1 4 1
1 3 1 4
a) Compute the eigenvalues and eigenvectors of S using the results of this
section. You should only need to perform one DFT in order to achieve this.

Solution. The eigenvalues of S are 1, 5, 9,√ and are found by computing a DFT
of the first column (and multiplying by N = 2). The eigenvectors are the
Fourier basis vectors. 1 has multiplicity 2.
b) Verify the result from a) by computing the eigenvectors and eigenvalues the
way you taught in your first course in linear algebra. This should be a much
more tedious task.
c) Use a computer to compute the eigenvectors and eigenvalues of S also. For
some reason some of the eigenvectors seem to be different from the Fourier basis
vectors, which you would expect from the theory in this section. Try to find an
explanation for this.

Solution. The computer uses some numeric algorithm to find the eigenvectors.
However, eigenvectors may not be unique, so you have no control on which
eigenvectors it actually selects. In particular, here the eigenspace for λ = 1 has
dimension 2, so that any linear combination of the two eigenvectors from this
eigenspace also is an eigenvector. Here it seems that a linear combination is
chosen which is different from a Fourier basis vector.

Exercise 3.16: Composing filters

Assume that S1 and S2 are two circulant Toeplitz matrices.
a) How can you express the eigenvalues of S1 + S2 in terms of the eigenvalues
of S1 and S2 ?
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS42

Solution. If we write S1 = FNH D1 FN and S2 = FNH D2 FN we get

S1 + S2 = FNH (D1 + D2 )FN S1 S2 = FNH D1 FN FNH D2 FN = FNH D1 D2 FN

This means that the eigenvalues of S1 + S2 are the sum of the eigenvalues of S1
and S2 . The actual eigenvalues which are added are dictated by the index of the
frequency response, i.e. λS1 +S2 ,n = λS1 ,n + λS2 ,n .
b) How can you express the eigenvalues of S1 S2 in terms of the eigenvalues of
S1 and S2 ?

Solution. As above we have that S1 S2 = FNH D1 FN FNH D2 FN = FNH D1 D2 FN ,

and the same reasoning gives that the eigenvalues of S1 S2 are the product of
the eigenvalues of S1 and S2 . The actual eigenvalues which are multiplied are
dictated by the index of the frequency response, i.e. λS1 S2 ,n = λS1 ,n λS2 ,n .
c) If A and B are general matrices, can you find a formula which expresses the
eigenvalues of A + B and AB in terms of those of A and B? If not, can you find
a counterexample to what you found in a) and b)?

Solution. In general there is no reason to believe that there is a formula for

the eigenvalues for the sum or product of two matrices, based on eigenvalues of
the individual matrices. However, the same type of argument as for filters can
be used in all cases where the eigenvectors are equal.

Exercise 3.17: Keeping every second component

Consider the linear mapping S which keeps every second component in RN ,
i.e. S(e2k ) = e2k , and S(e2k−1 ) = 0. Is S a digital filter?

Solution. The matrix for the operation which keeps every second component
is
 
1 0 ··· 0 0
0 0 · · · 0 0
 
 .. .. .. .. ..  ,
. . . . .
 
0 0 · · · 1 0
0 0 ··· 0 0
where 1 and 0 are repeated in alternating order along the main diagonal. Since
the matrix is not constant on the main diagonal, it is not a circulant Toeplitz
matrix, and hence not a filter.
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS43

Exercise 3.20: Plotting a simple frequency response

Let again S be the filter defined by the equation
1 1 1 1
zn =
xn+1 + xn + xn−1 + xn−2 ,
4 4 4 4
as in Exercise 3. Compute and plot (the magnitude of) λS (ω).

Solution. The frequency response is

1 iω eiω (1 − e−4iω ) 1 sin(2ω)

λS (ω) = (e + 1 + e−iω + e−2iω ) = −iω
= e−iω/2 .
4 4(1 − e ) 4 sin(ω/2)

Exercise 3.21: Low-pass and high-pass filters

A filter S is defined by the equation
1
zn = (xn + 3xn−1 + 3xn−2 + xn−3 ).
3
a) Compute and plot the (magnitude of the continuous) frequency response of
the filter, i.e. |λS (ω)|. Is the filter a low-pass filter or a high-pass filter?

Solution. The filter coefficients are t0 = t3 = 1/3, t1 = t2 = 1. We have that

X 1
λS (ω) = tk e−ikω = (1 + 3e−iω + 3e−2iω + e−3iω )
3
k
2 1
= e−3iω/2 (e3iω/2 + 3eiω/2 + 3e−iω/2 + e−3iω/2 )
3 2
2 −3iω/2
= e (cos(3ω/2) + 3 cos(ω/2)).
3
From this expression it is easy to plot the frequency response, but since this is
complex, we have to plot the magnitude, i.e. |λS (ω)| = 23 | cos(3ω/2)+3 cos(ω/2)|.
We also see that λS (0) = 23 , and that λS (π) = 0, so that the filter is a low-pass
filter.
b) Find an expression for the vector frequency response λS,2 . What is Sx when
x is the vector of length N with components e2πi2k/N ?

Solution. If we use the connection between the vector frequency response and
the continuous frequency response we get
2 −6πi/N
λS,2 = λS (2π2/N ) = e (cos(6π/N ) + 3 cos(2π/N )).
3
Alternatively you can here compute that the first column in the circulant Toeplitz
matrix for S is given by s0 = t1 , s2 = t2 , s3 = t3 , and s4 = t4 , and insert this in
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS44

PN −1
the definition of the vector frequency response, λS,2 = k=0 sk e−2πi2k/N . We
know that e2πi2k/N is an eigenvector for S since S is a filter, and that λS,2 is
the corresponding eigenvalue. We therefore get that
2 −6πi/N
Sx = λS,2 x = e (cos(6π/N ) + 3 cos(2π/N ))x.
3

Exercise 3.22: Circulant matrices

A filter S1 is defined by the equation
1
zn = (xn+2 + 4xn+1 + 6xn + 4xn−1 + xn−2 ).
16
a) Write down an 8 × 8 circulant Toeplitz matrix which corresponds to applying
S1 on a periodic signal with period N = 8.

Solution. Since clearly t−2 = t2 = 1/16, t−1 = t1 = 1/4, and t0 = 6/16,

the first column s in the circulant Toeplitz matrix is given by s0 = t0 = 6/16,
s1 = t1 = 4/16, s2 = t2 = 1/16, sN −2 = t−2 = 1/16, sN −1 = t−1 = 4/16. An
8 × 8 circulant Toeplitz matrix which corresponds to applying S1 to a periodic
signal of length N = 8 is therefore
 
6 4 1 0 0 0 1 4
 4 6 4 1 0 0 0 1
 
 1 4 6 4 1 0 0 0
 
1  0 1 4 6 4 1 0 0 .

16 
 0 0 1 4 6 4 1 0 

 0 0 0 1 4 6 4 1
 
 1 0 0 0 1 4 6 4
4 1 0 0 0 1 4 6
b) Compute and plot (the continuous) frequency response of the filter. Is the
filter a low-pass filter or a high-pass filter?

Solution. The frequency response is

 4
1 2iω 1 iω/2
λS1 (ω) = (e +4eiω +6+4e−iω +e−2iω ) = (e + e−iω/2 ) = cos4 (ω/2),
16 2

where we recognized (1, 4, 6, 4, 1) as a row in Pascal’s triangle, so that we could

write the expression as a power. From this expression it is easy to plot the
frequency response, and it is clear that the filter is a low-pass filter, since
λS1 (0) = 1, λS1 (π) = 0.
c) Another filter S2 has (continuous) frequency response λS2 (ω) = (eiω + 2 +
e−iω )/4. Write down the filter coefficients for the filter S1 S2 .
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS45

Solution. We have that

 2
1 iω/2
λS2 (ω) = (eiω + 2 + e−iω )/4 = (e + e−iω/2 ) = cos2 (ω/2).
2
We then get that

λS1 S2 (ω) = λS1 (ω)λS2 (ω) = cos4 (ω/2) cos2 (ω/2) = cos6 (ω/2)
 6
1 iω/2
= (e + e−iω/2 )
2
1 3iω
= (e + 6e2iω + 15eiω + 20 + 15e−iω + 6e−2iω + e−3iω ),
64
where we have used that, since we have a sixth power, the values can be obtained
from fra a row in Pascal’s triangle also here. It is now clear that
1
S1 S2 = {1, 6, 15, 20, 15, 6, 1}.
64
1
You could also argue here by taking the convolution of 16 (1, 4, 6, 4, 1) with
1
4 (1, 2, 1).

Exercise 3.23: Composite filters

Assume that the filters S1 and S2 have the frequency responses λS1 (ω) =
2 + 4 cos(ω), λS2 (ω) = 3 sin(2ω).
a) Compute and plot the frequency response of the filter S1 S2 .
b) Write down the filter coefficients tk and the impulse response s for the filter
S1 S2 .

Exercise 3.24: Maximum and minimum

Compute and plot the continuous frequency response of the filter S = {1/4, 1/2, 1/4}.
Where does the frequency response achieve its maximum and minimum value,
and what are these values?

Solution. We have that λS (ω) = 12 (1 + cos ω). This clearly has the maximum
point (0, 1), and the minimum point (π, 0).

Exercise 3.25: Plotting a simple frequency response

Plot the continuous frequency response of the filter T = {1/4, −1/2, 1/4}. Where
does the frequency response achieve its maximum and minimum value, and what
are these values? Can you write down a connection between this frequency
response and that from Exercise 3?
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS46

Solution. We have that |λT (ω)| = 12 (1 − cos ω). This clearly has the maximum
point (π, 1), and the minimum point (0, 0). The connection between the frequency
responses is that λT (ω) = λS (ω + π).

Exercise 3.26: Continuous- and vector frequency responses

Define the filter S by S = {1, 2, 3, 4, 5, 6}. Write down the matrix for S when
N = 8. Plot (the magnitude of) λS (ω), and indicate the values λS,n for N = 8
in this plot.

Solution. Here we have that s0 = t0 = 3, s1 = t1 = 4, s2 = t2 = 5, and

s3 = t3 = 6 (first formula), and sN −2 = t−2 = 1, sN −1 = t−1 = 2 (second
formula). This means that the matrix of S is
 
3 2 1 0 0 6 5 4
 4 3 2 1 0 0 6 5
 
 5 4 3 2 1 0 0 6
 
 6 5 4 3 2 1 0 0
S=  
 0 6 5 4 3 2 1 0

 0 0 6 5 4 3 2 1
 
 1 0 0 6 5 4 3 2
2 1 0 0 6 5 4 3
The frequency response is

λS (ω) = e2iω + 2eiω + 3 + 4e−iω + 5e−2iω + 6e−3iω .

Exercise 3.27: Starting with circulant matrices

Given the circulant Toeplitz matrix
 
1 1 1 ··· 1
1
 1 1 ··· 0 
0
 1 1 ··· 0 
1 . .. .. .. .. 
.
S = . . . . .
5 0

 0 0 ··· 1 
1
 0 0 ··· 1 
1 1 0 ··· 1
1 1 1 ··· 1
Write down the compact notation for this filter. Compute and plot (the magni-
tude) of λS (ω).

Solution. The filter coefficients are t0 = s0 = 1/5, t1 = s1 = 1/5 (first

formula), and t−1 = sN −1 = 1/5, t−2 = sN −2 = 1/5, t−3 = sN −3 = 1/5 (second
formula). All other tk are zero. This means that the filter can be written as
1
5 {1, 1, 1, 1, 1}, using our compact notation.
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS47

Exercise 3.28: When the filter coefficients are powers

Assume that S = {1, c, c2 , . . . , ck }. Compute and plot λS (ω) when k = 4 and
k = 8. How does the choice of k influence the frequency response? How does
the choice of c influence the frequency response?

Solution. The frequency response is

k
X 1 − ck+1 e−i(k+1)ω
cs e−isω = .
s=0
1 − ce−iω

It is straightforward to compute the limit as ω → 0 as ck (k + 1). This means

that as we increase k or c, this limit also increases. The value of k also dictates
oscillations in the frequency response, since the numerator oscillates fastest.
When c = 1, k dictates how often the frequency response hits 0.

Exercise 3.29: The Hanning window

The Hanning window is defined by wn = 1−cos(2πn/(N −1)). Compute and plot
the window coefficients and the continuous frequency response of this window for
N = 32, and compare with the window coefficients and the frequency responses
for the rectangular- and the Hamming window.

Exercise 3.37: Composing time delay filters

Let Ed1 and Ed2 be two time delay filters. Show that Ed1 Ed2 = Ed1 +d2 (i.e. that
the composition of two time delays again is a time delay) in two different ways:
a) Give a direct argument which uses no computations.
b) By using Property 3 in Theorem 2.7 in the book, i.e. by using a property for
the Discrete Fourier Transform.

Exercise 3.38: Adding echo filters

Consider the two filters S1 = {1, 0, . . . , 0, c} and S2 = {1, 0, . . . , 0, −c}. Both of
these can be interpreted as filters which add an echo. Show that 12 (S1 + S2 ) = I.
What is the interpretation of this relation in terms of echos?

Solution. The sum of two digital filters is again a digital filter, and the first
column in the sum can be obtained by summing the first columns in the two
matrices. This means that the filter coefficients in 12 (S1 + S2 ) can be obtained
by summing the filter coefficients of S1 and S2 , and we obtain
1
({1, 0, . . . , 0, c} + {1, 0, . . . , 0, −c}) = {1}.
2
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS48

This means that 12 (S1 + S2 ) = I, since I is the unique filter with e0 as first
column. The interpretation in terms of echos is that the echo from S2 cancels
that from S1 .

Exercise 3.39: Reducing bass and treble

Write code where you reduce the treble and the bass as described in examples 2.7
in the book and 3.36 in the book, generate the sounds you heard in these
examples, and verify that they are the same. In your code, it will be necessary
to scale the values after reducing the bass, but not the values after reducing
the bass. Explain why this is the case. How high must k be in order for you
to hear difference from the actual sound? How high can you choose k and still
recognize the sound at all? If you solved Exercise 3, you can also use the function
filterS to perform the filtering, rather than sing the convolve function (the
latter disregards circularity).

Solution. The code can look like this, when we reduce the treble. We have
used the convolve function:

t = [1.]
for kval in range(k):
t = convolve(t, [1/2., 1/2.])

z = convolve(t, x[:, 0])

play(z,fs)

If we reduce the bass instead, the code can look like this.

t = [1.]
for kval in range(k):
t = convolve(t, [1/2., -1/2.])

z = convolve(t, x[:, 0])

z /= abs(z).max()
play(z,fs)

Exercise 3.40: Constructing a high-pass filter

Consider again Example 3.32 in the book. Find an expression for a filter so that
only frequencies so that |ω − π| < ωc are kept, i.e. the filter should only keep
angular frequencies close to π (i.e. here we construct a high-pass filter).

Exercise 3.41: Combining low-pass and high-pass filters

In this exercise we will investigate how we can combine low-pass and high-pass
filters to produce other filters
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS49

a) Assume that S1 and S2 are low-pass filters. What kind of filter is S1 S2 ?

What if both S1 and S2 are high-pass filters?
b) Assume that one of S1 , S2 is a high-pass filter, and that the other is a low-pass
filter. What kind of filter S1 S2 in this case?

Exercise 3.42: Composing filters

A filter S1 has the frequency response 12 (1 + cos ω), and another filter has the
frequency response 12 (1 + cos(2ω)).
a) Is S1 S2 a low-pass filter, or a high-pass filter?
b) What does the filter S1 S2 do with angular frequencies close to ω = π/2.
c) Find the filter coefficients of S1 S2 .

Hint. Use Theorem 3.18 in the book to compute the frequency response of
S1 S2 first.
d) Write down the matrix of the filter S1 S2 for N = 8.

Exercise 3.43: Composing filters

An operation describing some transfer of data in a system is defined as the
composition of the following three filters:

• First a time delay filter with delay d1 = 2, due to internal transfer of data
in the system,
• then the treble-reducing filter T = {1/4, 1/2, 1/4},
• finally a time delay filter with delay d2 = 4 due to internal transfer of the
filtered data.

We denote by T2 = Ed2 T Ed1 = E4 T E2 the operation which applies these filters

in succession.
a) Explain why T2 also is a digital filter. What is (the magnitude of) the
frequency response of Ed1 ? What is the connection between (the magnitude of)
the frequency response of T and T2 ?
b) Show that T2 = {0, 0, 0, 0, 0, 1/4, 1/2, 1/4}.

Hint. Use the expressions (Ed1 x)n = xn−d1 , (T x)n = 14 xn+1 + 12 xn + 14 xn−1 ,
(Ed2 x)n = xn−d2 , and compute first (Ed1 x)n , then (T Ed1 x)n , and finally
(T2 x)n = (Ed2 T Ed1 x)n . From the last expression you should be able to read
out the filter coefficients.
c) Assume that N = 8. Write down the 8 × 8-circulant Toeplitz matrix for the
filter T2 .
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS50

Exercise 3.44: Filters in the MP3 standard

In Example 3.34 in the book, we mentioned that the filters used in the MP3-
standard were constructed from a low-pass prototype filter by multiplying the
filter coefficients with a complex exponential. Clearly this means that the new
frequency response is a shift of the old one. The disadvantage is, however, that
the new filter coefficients are complex. It is possible to address this problem as
follows. Assume that tk are the filter coefficients of a filter S1 , and that S2 is
the filter with filter coefficients cos(2πkn/N )tk , where n ∈ N. Show that
1
(λS1 (ω − 2πn/N ) + λS1 (ω + 2πn/N )).
λS2 (ω) =
2
In other words, when we multiply (modulate) the filter coefficients with a cosine,
the new frequency response can be obtained by shifting the old frequency response
with 2πn/N in both directions, and taking the average of the two.

Solution. We have that

X 1 X 2πikn/N
λS2 (ω) = cos(2πkn/N )tk e−ikω = (e + e−2πikn/N )tk e−ikω
2
k k
!
1 X
−ik(ω−2πn/N )
X
−ik(ω+2πn/N )
= tk e + tk e
2
k k
1
= (λS1 (ω − 2πn/N ) + λS1 (ω + 2πn/N )).
2

Exercise 3.45: Explain code

a) Explain what the code below does, line by line.

x, fs = audioread(’sounds/castanets.wav’)
N, nchannels = shape(x)
z = zeros((N, nchannels))
for n in range(1,N-1):
z[n] = 2*x[n+1] + 4*x[n] + 2*x[n-1]
z[0] = 2*x[1] + 4*x[0] + 2*x[N-1]
z[N-1] = 2*x[0] + 4*x[N-1] + 2*x[N-2]
z = z/abs(z).max()
play(z, fs)

Comment in particular on what happens in the three lines directly after the
for-loop, and why we do this. What kind of changes in the sound do you expect
to hear?

Solution. In the code a filter is run on the sound samples from the file
castanets.wav. Finally the new sound is played. In the first two lines after the
for-loop, the first and the last sound samples in the filtered sound are computed,
under the assumption that the sound has been extended to a periodic sound
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS51

with period N. After this, the sound is normalized so that the sound samples lie
in the range between −1 and 1. In this case the filter is a low-pass-filter (as we
show in b)), so that we can expect that that the treble in the sound is reduced.
b) Write down the compact filter notation for the filter which is used in the
code, and write down a 5 × 5 circulant Toeplitz matrix which corresponds to
this filter. Plot the (continuous) frequency response. Is the filter a low-pass- or
a high-pass filter?

Solution. Compact filter notation for the filter which is run is {2, 4, 2}. A
5 × 5 circulant Toeplitz matrix becomes
 
4 2 0 0 2
2 4 2 0 0
 
0 2 4 2 0
 .
0 0 2 4 2
2 0 0 2 4
The frequency response is λS (ω) = 2eiω + 4 + 2e−iω = 4 + 4 cos ω. It is clear
that this gives a low-pass filter.
c) Another filter is given by the circulant Toeplitz matrix
 
4 −2 0 0 −2
−2 4 −2 0 0
 
 0 −2 4 −2 0 
 .
0 0 −2 4 −2
−2 0 0 −2 4
Express a connection between the frequency responses of this filter and the filter
from b). Is the new filter a low-pass- or a high-pass filter?

Solution. The frequency response for the new filter is

−2eiω + 4 − 2e−iω = 4 − 4 cos ω = 4 + 4 cos(ω + π) = λS (ω + π),

where S is the filter from the first part of the exercise. The new filter therefore
becomes a high-pass filter, since to add π to ω corresponds to swapping the
frequencies 0 and π. We could also here refered to Observation 3.22 in the book,
where we stated that adding an alternating sign in the filter coefficients turns a
low-pass filter into a high-pass filter and vice versa.

Exercise 3.47: A concrete IIR filter

A filter is defined by demanding that zn+2 − zn+1 + zn = xn+1 − xn .
a) Compute and plot the frequency response of the filter.
CHAPTER 3. OPERATIONS ON DIGITAL SOUND: DIGITAL FILTERS52

b) Use a computer to compute the output when the input vector is x =

(1, 2, . . . , 10). In order to do this you should write down two 10 × 10-circulant
Toeplitz matrices.

Exercise 3.48: Implementing the factorization

Write a function filterdftimpl, which takes the filter coefficients t and the
value k0 from this section, computes the optimal M , and implements the filter
as here.

Exercise 3.49: Factoring concrete filter

Factor the filter S = {1, 5, 10, 6} into a product of two filters, one with two filter
coefficients, and one with three filter coefficients.
Chapter 4

Symmetric filters and the

DCT

Exercise 4.2: Computing eigenvalues

Consider the matrix
 
2 1 0 0 0 0
1 1 1 0 0 0
 
1 0 1 1 1 0 0
S=  
30 0 1 1 1 0
0 0 0 1 1 1
0 0 0 0 1 2
a) Compute the eigenvalues and eigenvectors of S using the results of this
section. You should only need to perform one DFT or one DCT in order to
achieve this.
b) Use a computer to compute the eigenvectors and eigenvalues of S also. What
are the differences from what you found in a)?
c) Find a filter T so that S = Tr . What kind of filter is T ?

Exercise 4.3: Writing down lower order Sr

Consider the averaging filter S = { 14 , 12 , 14 }. Write down the matrix Sr for the
case when N = 4.

Solution. First we obtain the matrix S as

53
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 54

1 1 1
 
2 4 0 0 0 0 0 4
1 1 1

4 2 4 0 0 0 0 0 
1 1 1
 
 0 4 2 4 0 0 0 0 
1 1 1
 

 0 0 4 2 4 0 0 0 

1 1 1

 0 0 0 4 2 4 0 0 

1 1 1

 0 0 0 0 4 2 4 0 

1 1 1
 0 0 0 0 0 4 2 4

1 1 1
4 0 0 0 0 0 4 2

where we have drawn the boundaries between the blocks S1 , S2 , S3 , S4 . From

this we see that

1 1
  1
 1 
2 4 0 0 0 0 0 4 4 0 0 0
1 1 1
0 0 0 0 0 0 0 0 0
4
S1 =  2
1
4
1 1
 S2 = 
  (S2 )f =  .
0 4 2 4 0 0 0 0 0 0 0 0
1 1 1 1
0 0 4 2 4 0 0 0 0 0 0 4

From this we get

3 1

4 4 0 0
1 1 1
0
Sr = S1 + (S2 )f =  4 2
1
4
1 1 .

0 4 2 4
1 3
0 0 4 4

Exercise 4.4: Writing down lower order DCTs

As in Example 4.1 in the book, state the exact cartesian form of the DCT matrix
for the case N = 3.
q q
Solution. We first see that d0,3 = 13 and dk,3 = 23 for k = 1, 2. We also
have that
     
n 1 n 1
cos 2π k+ = cos π k+ ,
2N 2 3 2
so that the DCT matrix can be written as
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 55

 q q q 
1 1 1
q 3 3 3
π1
q
π3
q
π5


DCT3 =  23 cos 2 2



cos cos
 
3 2 3 3 2 3 3 2

q q q

2 2π 1 2 2π 3 2 2π 5



3 cos 3 2 cos
3 3 2 3cos 3 2
 q q q 
1 1 1
q 3 q q 3  3

=  23 cos(π/6) 2
cos(π/2) 2
cos(5π/6)
 
q3 q3

q 
2 2 2
3 cos(π/3) 3 cos(π) 3 cos(5π/3)
 q q q 
1 1 1
q √ 3 3 q √
3 
 2 2
=  3 ( 3/2 + i/2) 0 (− 3/2 + i/2)

q √ q q3 √ 
2
3 (1/2 + 3i/2) − 23 2
3 (1/2 − 3i/2)

Exercise 4.5: DCT-IV

n 
n+ 1
oN −1
Show that the vectors cos 2π 2N2 k + 12 in RN are orthogonal, with
n=0 q
n+ 1
p 

lengths N/2. This means that the matrix with entries N2 cos 2π 2N2 k + 12
is orthogonal. Since this matrix also is symmetric, it is its own inverse. This is
(IV)
the DCT-IV, which we denote by DCTN . Although we will not consider this,
the DCT-IV also has an efficient implementation.

Hint. Compare with the orthogonal vectors dn , used in the DCT.

Solution. We can write

n + 21
     
1 2n + 1 1
cos 2π k+ = cos 2π k+ .
2N 2 4N 2
If we consider these as vectors of length 2N , we recognize these as thep
unit vectors
d
p 2n+1 in the 2N -dimensional DCT, divided
√ by the factor dn = 2/(2N ) =
1/N , so that these vectors have length N . To see that these vectors are
orthogonal when we restrict to the first N elements we use Equation (6.36) in
the book as follows:
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 56

2N −1
n1 + 21 n2 + 12
     
X 1 1
N δn1 ,n2 = cos 2π k+ cos 2π k+
2N 2 2N 2
k=0
N −1
n1 + 12 n2 + 21
     
X 1 1
= cos 2π k+ cos 2π k+
2N 2 2N 2
k=0
2N −1
n1 + 21 n2 + 12
     
X 1 1
+ cos 2π k+ cos 2π k+
2N 2 2N 2
k=N
N −1
n1 + 12 n2 + 21
     
X 1 1
= cos 2π k+ cos 2π k+
2N 2 2N 2
k=0
N −1
n1 + 12 n2 + 12
     
X 1 1
+ cos 2π k+ cos 2π k+
2N 2 2N 2
k=0
N −1
n1 + 21 n2 + 12
     
X 1 1
=2 cos 2π k+ cos 2π k+ .
2N 2 2N 2
k=0

This shows that

N −1  1    1  
X n1 + 2 1 n2 + 2 1 N
cos 2π k+ cos 2π k+ = δn1 ,n2 ,
2N 2 2N 2 2
k=0
p
so that the vectors are orthogonal with lengths N/2.

Exercise 4.6: MDCT

The MDCT is defined as the N ×(2N )-matrix M with elements Mn,k = cos(2π(n+
1/2)(k + 1/2 + N/2)/(2N )). This exercise will take you through the details of
the transformation which corresponds to multiplication with this matrix. The
MDCT is very useful, and is also used in the MP3 standard and in more recent
standards.
a) Show that

r  
N (IV) 0 A
M= DCTN
2 B 0

where A and B are the (N/2) × N -matrices

CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 57

 
··· ··· 0 −1 −1 0 ··· ···
 .. .. .. .. .. .. .. ..   
f
A= . . . . . . . . 
 = −IN/2 −IN/2

0 −1 ··· ··· · · · · · · −1 0 
−1 0 ··· ··· · · · · · · 0 −1
 
1 0 ··· ··· · · · · · · 0 −1
0 1 ··· ··· · · · · · · −1 0  
f

B= . ..  = IN/2 −IN/2 .
 
.. .. .. .. .. ..
 .. . . . . . . . 
··· ··· 0 1 −1 0 · · · · · ·

Due to this expression, any algorihtm for the DCT-IV can be used to compute
the MDCT.

Solution. Clearly, columns 0, . . . , N/2−1 of the MDCT are columns N/2, . . . , N −

1 of the DCT-IV. For the remaining columns, note first that, for 0 ≤ k < N , the
properties

cos(2π(n + 1/2)((2N − 1 − k) + 1/2)/(2N )) = − cos(2π(n + 1/2)(k + 1/2)/(2N ))

cos(2π(n + 1/2)((k + 2N ) + 1/2)/(2N )) = − cos(2π(n + 1/2)(k + 1/2)/(2N ))

are easy to verify. From the first property it follows that columns N/2, . . . , 3N/2−
1 of the MDCT are columns N − 1, N − 2, . . . , 0 of the DCT-IV, with a sign
change (they occur in opposite order). From the second property, it follows
that columns 3N/2, . . . , 2N − 1 of the MDCT are columns 0, . . . , N/2 − 1 of the
DCT-IV, with a sign change. This means that, if y is a vector of length 2N , the
MDCT of y can be written as

 
−y3N/2 − y3N/2−1
−y3N/2+1 − y3N/2−2 
..
 
 
r 
 . 

N (IV)  −y2N −1 − yN 
.
DCTN 
2 
 y0 − yN −1 


 y1 − yN/2+1 

 .. 
 . 
yN/2−1 − yN/2
q q
N
The factor 2 was added since N2 was added in front of the cosine-matrix
(IV)
in order to make
 DCT  N orthogonal. The result now follows by noting that
0 A
we can write y for the vector on the right hand side, with A and B as
B 0
defined in the text of the exercise.
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 58

b) The MDCT is not invertible, since it is not a square matrix. We will show
here that it still can be used in connection with invertible transformations. We
first define the IMDCT as the matrix M T /N . Transposing the matrix expression
we obtained in a) gives

BT
 
1 0 (IV)
√ DCTN
2N AT 0
for the IMDCT, which thus also has an efficient implementation. Show that if

x0 = (x0 , . . . , xN −1 ) x1 = (xN , . . . , x2N −1 ) x2 = (x2N , . . . , x3N −1 )

and

   
x0 x1
y0,1 =M y1,2 = M
x1 x2
(i.e. we compute two MDCT’s where half of the data overlap), then
−1 N −1
x1 = {IMDCT(y0,1 )}2N
k=N + {IMDCT(y1,2 )}k=0 .
Even though the MDCT itself is not invertible, the input can still be recovered
from overlapping MDCT’s.

Solution. Applying the MDCT first, and then the IMDCT, gives us the matrix

BT
    T 
1 0 0 A 1 B B 0
=
2 AT 0 B 0 2 0 AT A
Note that

f
! f
!
T
IN/2 IN/2 T
IN/2 −IN/2
A A= f B B= f .
IN/2 IN/2 −IN/2 IN/2

Inserting this in the above gives

 f

IN/2 −IN/2 0 0
 f f
−I

I 0 0  1 IN − IN
 
1  N/2
 N/2 0
f 
 = f .
2 0
 0 IN/2 IN/2  2 0 IN + IN
f
0 0 IN/2 IN/2
 
x0
Assume now that we have computed the MDCT of = (x0 , . . . , x2N −1 ),
  x1
x1
and of = (xN , . . . , x3N −1 ). Performing the IMDCT on these two thus
x2
gives
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 59

x0 − xrev 1 x1 − xrev
   
1 0 1
and
2 x1 + xrev
1 2 x2 + xrev
2

Adding the second component of the first and the first component of the second
gives x1 , which proves the result.

Exercise 4.7: Component expressions for a symmetric filter

Assume that S = t−L , . . . , t0 , . . . , tL is a symmetric filter. Use Equation (4.7) in
the book to show that zn = (Sx)n in this case can be split into the following
different formulas, depending on n:
a) 0 ≤ n < L:

n
X L
X
zn = t0 xn + tk (xn+k + xn−k ) + tk (xn+k + xn−k+N ). (4.1)
k=1 k=n+1

b) L ≤ n < N − L:
L
X
zn = t0 xn + tk (xn+k + xn−k ). (4.2)
k=1

c) N − L ≤ n < N :

−1−n
NX L
X
zn = t 0 xn + tk (xn+k + xn−k ) + tk (xn+k−N + xn−k ). (4.3)
k=1 k=N −1−n+1

The convolve function may not pick up this reduction in the number of
multiplications, since it does not assume that the filter is symmetric. We will
still use the convolve function in implementations, however, due to its heavy
optimization.

Exercise 4.8: Trick for reducing the number of multiplica-

tions with the DCT
In this exercise we will take a look at a small trick which reduces the number of
additional multiplications we need for DCT algorithm from Theorem 4.22 in the
book. This exercise does not reduce the order of the DCT algorithms, but we
will see in Exercise 4 how the result can be used to achieve this.
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 60

a) Assume that x is a real signal. Equation (4.12) in the book, which said that

 n   n 
yn = cos π <((DFTN x(1) )n ) + sin π =((DFTN x(1) )n )
2N
 n  2N
 n 
yN −n = sin π <((DFTN x(1) )n ) − cos π =((DFTN x(1) )n )
2N 2N
for the n’th and N − n’th coefficient of the DCT. This can also be rewritten as

   n 
yn = <((DFTN x(1) )n ) + =((DFTN x(1) )n ) cos π
 n   n  2N
(1)
− =((DFTN x )n )(cos π − sin π )
 2N 2N  n 
yN −n = − <((DFTN x(1) )n ) + =((DFTN x(1) )n ) cos π
 n   n  2N
(1)
+ <((DFTN x )n )(sin π + cos π ).
2N 2N
Explain that the first two equations require 4 multiplications to compute yn and
yN −n , and that the last two equations require 3 multiplications to compute yn
and yN −n .
b) Explain why the trick in a) reduces the number of additional multiplications
in a DCT, from 2N to 3N/2.
c) Explain why the trick in a) can be used to reduce the number of additional
multiplications in an IDCT with the same number.

Hint. match the expression eπin/(2N ) (yn − iyN −n ) you encountered in the
IDCT with the rewriting you did in b).
d) Show that the penalty of the trick we here have used to reduce the number
of multiplications, is an increase in the number of additional additions from N
to 3N/2. Why can this trick still be useful?

Exercise 4.9: An efficient joint implementation of the DCT

and the FFT
In this exercise we will explain another joint implementation of the DFT and
the DCT, which has the benefit of a low multiplication count, at the expense
of a higher addition count. It also has the benefit that it is specialized to real
vectors, with a very structured implementation (this is not always the case for
the quickest FFT implementations. Not surprisingly, one often sacrifices clarity
of code when one pursues higher computational speed). a) of this exercise can be
skipped, as it is difficult and quite technical. For further details of the algorithm
the reader is refered to [?].
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 61

a) Let y = DFTN x be the N -point DFT of the real vector x. Show that


 <((DFTN/2 x(e) )n ) + (CN/4 z)n 0 ≤ n ≤ N/4 − 1
<(yn ) = <((DFTN/2 x(e) )n ) n = N/4
<((DFTN/2 x(e) )n ) − (CN/4 z)N/2−n N/4 + 1 ≤ n ≤ N/2 − 1

(4.4)

 =((DFTN/2 x(e) )n ) n=0
=(yn ) = =((DFTN/2 x(e) )n ) + (CN/4 w)N/4−n 1 ≤ n ≤ N/4 − 1 (4.5)
=((DFTN/2 x(e) )n ) + (CN/4 w)n−N/4 N/4 ≤ n ≤ N/2 − 1


where x(e) is as defined in Theorem 2.15 in the book, where z, w ∈ RN/4 defined
by

zk = x2k+1 + xN −2k−1 0 ≤ k ≤ N/4 − 1,

k
wk = (−1) (xN −2k−1 − x2k+1 ) 0 ≤ k ≤ N/4 − 1,

Explain from this how you can make an algorithm which reduces an FFT of
length N to an FFT of length N/2 (on x(e) ), and two DCT’s of length N/4 (on
z and w). We will call this algorithm the revised FFT algorithm.

Solution. Taking real and imaginary parts in Equation (2.13) in the book for
the FFT algorithm we obtain

 
<(yn ) = < (DFTN/2 x(e) )n + <((DN/2 DFTN/2 x(o) )n
 
=(yn ) = = (DFTN/2 x(e) )n + =((DN/2 DFTN/2 x(o) )n ,

These equations explain the first terms <((DFTN/2 x(e) )n ) and =((DFTN/2 x(e) )n )
on the right hand sides
in equations (4.4) and (4.5).
It remains to rewrite
< (DN/2 DFTN/2 x(o) )n and = (DN/2 DFTN/2 x(o) )n so that the remaining
terms on the right hand sides can be seen. Let us first consider the equation for
the real part with 0 ≤ n ≤ N/4 − 1. In this case we can write
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 62

<((DN/2 DFTN/2 x(o) )n )

   
N/2−1 N/2−1
1
X X
= < e−2πin/N (x(o) )k e−2πink/(N/2)  = <  (x(o) )k e−2πin(k+ 2 )/(N/2) 
k=0 k=0
N/2−1
n(k + 12 )
X  
= (x(o) )k cos 2π
N/2
k=0
N/4−1
n(k + 12 )
X  
= (x(o) )k cos 2π
N/2
k=0
N/4−1
n(N/2 − 1 − k + 21 )
X  
+ (x(o) )N/2−1−k cos 2π
N/2
k=0
N/4−1 1

!
X n k +
= ((x(o) )k + (x(o) )N/2−1−k ) cos 2π 2
N/2
k=0
N/4−1
!
X n k + 21
= zk cos 2π ,
N/2
k=0

where we have used that cos is periodic with period 2π, that cos is symmetric, and
where z is the vector defined in the text of the theorem. When 0 ≤ n ≤ N/4 − 1
this can also be written as
N/4−1
X
(CN/4 )n,k zk = (CN/4 z)n ,
k=0

This proves the first formula in Equation (4.4).

For N/4 + 1 ≤ n ≤ N/2 − 1, everything above is valid, except for that
cos(2πn(k + 1/2)/(N/2)) are not entries in the matrix CN/4 , since n is outside
the legal range of the indices. However, N/2 − n is now a legal index in CN/4 ,
and using that
N

!
n(k + 12 ) k + 12


2 −n
 
cos 2π = − cos 2π ,
N/2 N/2

we arrive at −(CN/4 z)N/2−n instead, and this proves the third formula in
Equation (4.4). For the case n = N4 all the cosine entries in the sum are zero,
and this completes the proof of Equation (4.4).
For the imaginary part, using that sin is periodic with period 2π, and that
sin is anti-symmetric, analogous calculations as above give
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 63

N/4−1
!
(o)
X
(o) (o) n k + 21
=((DN/2 DFTN/2 x )n ) = ((x )N/2−1−k − (x )k ) sin 2π .
N/2
k=0
(4.6)
Using that

n(k + 12 ) n(k + 21 ) (N/4 − n)(k + 21 )

     
π
sin 2π = cos − 2π = cos 2π − kπ
N/2 2 N/2 N/2
(N/4 − n)(k + 21 )
 
= (−1)k cos 2π ,
N/2
Equation (4.6) can be rewritten as

N/4−1
(N/4 − n)(k + 21 )
X  
((x(o) )N/2−1−k − (x(o) )k )(−1)k cos 2π
N/2
k=0
N/4−1
(N/4 − n)(k + 21 )
X  
= wk cos 2π ,
N/2
k=0

where w is the vector defined as in the text of the theorem. When n = 0 this is 0
since all the cosines entries are zero. When 1 ≤ n ≤ N/4 this is (CN/4 w)N/4−n ,
since cos(2π(N/4 − n)(k + 1/2)/(N/2)) are entries in the matrix CN/4 . This
proves the second formula in Equation (4.5).
For N/4 ≤ n ≤ N/2 − 1 we can use that cos(2π(N/4 − n)(k + 1/2)/(N/2)) =
cos(2π(n − N/4)(k + 1/2)/(N/2)), which is an entry in the matrix DCTN/4
as well, so that we get (CN/4 z)n−N/4 . This also proves the third formula in
Equation (4.5), and the proof is done.
a) says nothing about the coefficients yn for n > N2 . These are obtained in
the same way as before through symmetry. a) also says nothing about yN/2 .
This can be obtained with the same formula as in Theorem 2.15 in the book.
Let us now compute the number of arithmetic operations our revised algorithm
needs. Denote by the number of real multiplications needed by the revised N -
point FFT algorithm
b) Explain from the algorithm in a) that

MN = 2(MN/4 + 3N/8) + MN/2 AN = 2(AN/4 + 3N/8) + AN/2 + 3N/2

(4.7)

Hint. 3N/8 should come from the extra additions/multiplications (see Exer-
cise 4) you need to compute when you run the algorithm from Theorem 4.22 in
the book for CN/4 . Note also that the equations in a) require no extra multipli-
cations, but that there are xix equations involved, each needing N/4 additions,
so that we need 6N/4 = 3N/2 extra additions.
CHAPTER 4. SYMMETRIC FILTERS AND THE DCT 64

c) Explain why xr = M2r is the solution to the difference equation

xr+2 − xr+1 − 2xr = 3 × 2r ,

and that xr = A2r is the solution to

xr+2 − xr+1 − 2xr = 9 × 2r .

and show that the general solution to these are xr = 12 r2r + C2r + D(−1)r for
multiplications, and xr = 32 r2r + C2r + D(−1)r for additions.
d) Explain why, regardless of initial conditions

to the difference equations,
MN = O 12 N log2 N and AN = O 32 N log2 N both for the revised FFT and

the revised DCT. The total number of operations is thus O(2N log2 N ), i.e. half
the operation count of the split-radix algorithm. The orders of these algorithms
are thus the same, since we here have adapted to read data.
e) Explain that, if you had not employed the
trick from Exercise 4, we
would
instead have obtained MN = O 23 log2 N , and AN = O 43 log2 N , which
equal the orders for the number of multiplications/additions for the split-radix
algorithm. In particular, the order of the operation count remains the same, but
the trick from Exercise 4 turned a bigger percentage of the arithmetic operations
into additions.
The algorithm we here have developed thus is constructed from the beginning
to apply for real data only. Another advantage of the new algorithm is that it
can be used to compute both the DCT and the DFT.

Exercise 4.10: Implementation of the IFFT/IDCT

We did not write down corresponding algorithms for the revised IFFT and IDCT
algorithms. We will consider this in this exercise.
a) Using equations (4.4)-(4.5), show that

<(yn ) − <(yN/2−n ) = 2(CN/4 z)n

=(yn ) + =(yN/2−n ) = 2(CN/4 w)N/4−n

for 1 ≤ n ≤ N/4 − 1. Explain how one can compute z and w from this using
two IDCT’s of length N/4.
b) Using equations (4.4)-(4.5), show that

<(yn ) + <(yN/2−n ) = <((DFTN/2 x(e) )n )

=(yn ) − =(yN/2−n ) = =((DFTN/2 x(e) )n ),

and explain how one can compute x(e) from this using an IFFT of length N/2.
Chapter 5

Motivation for wavelets and

some simple examples

Exercise 5.1: The vector of samples is the coordinate vector

Show that the coordinate vector for f ∈ V0 in the basis {φ0,0 , φ0,1 , . . . , φ0,N −1 }
is (f (0), f (1), . . . .f (N − 1)). This shows that, for f ∈ Vm , there is no loss of
information in working with the samples of f rather than f itself.
PN −1
Solution. We have that f (t) = n=0 cn φ0,n , where cn are the coordinates of
f in the basis {φ0,0 , φ0,1 , . . . , φ0,N −1 }. We now get that
N
X −1
f (k) = cn φ0,n (k) = ck ,
n=0

6 k. This shows that (f (0), f (1), . . . .f (N − 1)) are

since φ0,n (k) = 0 when n =
the coordinates of f .

Exercise 5.2: Proposition 5.12 in the book

Prove Proposition 5.12 in the book.

Solution. Since f is constant and equal to f (n) on [n, n + 1/2), and constant
and equal to f (n + 1/2) on [n + 1/2, n + 1), we get that

65
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES66

Z N Z n+1
hf, φ0,n i = f (t)φ0,n (t)dt = f (t)dt
0 n
Z n+1/2 Z n+1
= f (t)dt + f (t)dt
n n+1/2
Z n+1/2 Z n+1
= f (n)dt + f (n + 1/2)dt
n n+1/2

= f (n)/2 + f (n + 1/2)/2 = (f (n) + f (n + 1/2))/2.

The orthogonal decomposition theorem gives that

N −1 N −1
X X f (n) + f (n + 1/2)
projV0 f = hf, φ0,n iφ0,n = φ0,n .
n=0 n=0
2

Since φ0,n is 1 on [n, n + 1) and 0 elsewhere, projV0 f is the piecewise constant

function which is equal to (f (n) + f (n + 1/2))/2 on [n, n + 1).

Exercise 5.3: Computing projections 1

In this exercise we will consider the two projections from V1 onto V0 and W0 .
a) Consider the projection projV0 of V1 onto V0 . Use Lemma 5.11 in the book
to write down the matrix for projV0 relative to the bases φ1 and φ0 .

Solution. From Lemma 5.11 in the book it follows that

√
projV0 (φ1,2n ) = φ0,n / 2
√
projV0 (φ1,2n+1 ) = φ0,n / 2

This means that

√
[projV0 (φ1,2n )]φ0 = en / 2
√
[projV0 (φ1,2n+1 )]φ0 = en / 2.

These are the columns in the matrix for projV0 relative to the bases φ1 and φ0 .
This matrix is thus
 
1 1 0 0 0 ··· 0 0 0
1 0 0 1 1 0 · · ·
 0 0 0
√ . . . . . ..  .

.. .. ..
2  .. .. .. .. .. . . . .
0 0 0 0 0 ··· 0 1 1
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES67

b) Similarly, use Lemma 5.11 in the book to write down the matrix for projW0 :
V1 → W0 relative to the bases φ1 and ψ0 .

Solution. From Lemma 5.11 in the book it follows that

√
projW0 (φ1,2n ) = ψ0,n / 2
√
projW0 (φ1,2n+1 ) = −ψ0,n / 2

This means that

√
[projW0 (φ1,2n )]ψ0 = en / 2
√
[projW0 (φ1,2n+1 )]ψ0 = −en / 2.

These are the columns in the matrix for projW0 relative to the bases φ1 and ψ0 .
This matrix is thus
 
1 −1 0 0 0 ··· 0 0 0
1 0 0 1 −1
 0 ··· 0 0 0 
√ . .

.. .. .. .. .. .. .. ..
2  .. . . . . . . . .
0 0 0 0 0 ··· 0 1 −1

Exercise 5.4: Computing projections 2

Consider again the projection projV0 of V1 onto V0 .
a) Explain why projV0 (φ) = φ and projV0 (ψ) = 0.

Solution. Since φ ∈ V0 we must have that projV0 (φ) = φ. Since ψ is in the

orthogonal complement of V0 in V1 we must have that projV0 (ψ) = 0.
b) Show that the matrix of projV0 relative to (φ0 , ψ0 ) is given by the diagonal
matrix where the first half of the entries on the diagonal are 1, the second half 0.

Solution. The first columns in the matrix of projV0 relative to (φ0 , ψ0 ) are

[projV0 (φ0,0 )](φ0 ,ψ0 ) = [φ0,0 ](φ0 ,ψ0 ) = e0

[projV0 (φ0,1 )](φ0 ,ψ0 ) = [φ0,1 ](φ0 ,ψ0 ) = e1
....
..

The last columns in the matrix of projV0 relative to (φ0 , ψ0 ) are

CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES68

[projV0 (ψ0,0 )](φ0 ,ψ0 ) = [0](φ0 ,ψ0 ) = 0

[projV0 (ψ0,1 )](φ0 ,ψ0 ) = [0](φ0 ,ψ0 ) = 0
....
..

It follows that the matrix of projV0 relative to (φ0 , ψ0 ) is given by the diagonal
matrix where the first half of the entries on the diagonal are 1, the second half 0.
c) Show in a similar way that the projection of V1 onto W0 has a matrix relative
to (φ0 , ψ0 ) given by the diagonal matrix where the first half of the entries on
the diagonal are 0, the second half 1.

Solution. Follows in the same way as b).

Exercise 5.5: Computing projections 3

Show that
N
X −1 Z n+1 
projV0 (f ) = f (t)dt φ0,n (t) (5.1)
n=0 n

for any f . Show also that the first part of Proposition 5.12 in the book follows
from this.

Solution. We have that

−1 −1 Z n+1
N
! N
X Z N X 
projV0 (f ) = f (t)φ0,n (t)dt φ0,n = f (t)dt φ0,n ,
n=0 0 n=0 n

where we have used the orthogonal decomposition formula. Note also that, if
f (t) ∈ V1 , and fn,1 is the value f attains on [n, n + 1/2), and fn,2 is the value f
attains on [n + 1/2, n + 1), we have that

N
X −1 Z n+1 
projV0 (f ) = f (t)dt φ0,n (t)
n=0 n
N −1   N −1
X 1 1 X fn,1 + fn,2
= fn,1 + fn,2 φ0,n (t) = φ0,n (t),
n=0
2 2 n=0
2

which is the function which is (fn,1 + fn,2 )/2 on [n, n + 1). This proves the first
part of Proposition 5.12 in the book.
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES69

Exercise 5.6: Finding the least squares error

Show that

X Z n+1  X Z n+1 2
k f (t)dt φ0,n (t) − f k2 = hf, f i − f (t)dt .
n n n n

This, together with the previous exercise, gives us an expression for the least-
squares error for f from V0 (at least after taking square roots). 2DO: Generalize
to m

Solution. We have that

kf − projV0 (f )k2 = hf − projV0 (f ), f − projV0 (f )i

= hf, f i − 2hf, projV0 (f )i + hprojV0 (f ), projV0 (f )i

Now, note that

N
X −1 Z n+1 2
hprojV0 (f ), projV0 (f )i = f (t)dt
n=0 n

from what we just showed in Exercise 5 (use that the φ0,n are orthonormal).
This means that the above can be written

N
X −1 Z N Z n+1  N
X −1 Z n+1 2
= hf, f i − 2 f (s)ds φ0,n (t)f (t)dt + f (t)dt
n=0 0 n n=0 n
N
X −1 Z n+1 Z n+1  N
X −1 Z n+1 2
= hf, f i − 2 f (s)ds f (t)dt + f (t)dt
n=0 n n n=0 n
N
X −1 Z n+1 2 NX
−1 Z n+1 2
= hf, f i − 2 f (t)dt + f (t)dt
n=0 n n=0 n
N
X −1 Z n+1 2
= hf, f i − f (t)dt .
n=0 n

Exercise 5.7: Projecting on W0

Show that
−1
N
!
X Z n+1/2 Z n+1
projW0 (f ) = f (t)dt − f (t)dt ψ0,n (t) (5.2)
n=0 n n+1/2

for any f . Show also that the second part of Proposition 5.12 in the book follows
from this.
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES70

Solution. The orthogonal decomposition theorem gives that

−1 −1
N N
!
X X Z N
projW0 (f ) = hf, ψ0,n iψ0,n (t) = f (t)ψ0,n (t)dt ψ0,n (t)
n=0 n=0 0
N
X −1 Z n+1 
= f (t)ψ0,n (t)dt ψ0,n (t)
n=0 n
−1
N
!
X Z n+1/2 Z n+1
= f (t)dt − f (t)dt ψ0,n (t),
n=0 n n+1/2

where we used that ψ0,n is nonzero only on [n, n + 1), and is 1 on [n, n + 1/2),
and −1 on [n + 1/2, n + 1). Note also that, if f (t) ∈ V1 , and fn,1 is the value f
attains on [n, n + 1/2), and fn,2 is the value f attains on [n + 1/2, n + 1), we
have that

−1
N
!
X Z n+1/2 Z n+1
projW0 (f ) = f (t)dt − f (t)dt ψ0,n (t)
n=0 n n+1/2
N −1   N −1
X 1 1 X fn,1 − fn,2
= fn,1 − fn,2 ψ0,n (t) = ψ0,n (t),
n=0
2 2 n=0
2

which is the function which is (fn,1 − fn,2 )/2 on [n, n + 1/2), and −(fn,1 − fn,2 )/2
on [n + 1/2, n + 1). This proves the second part of Proposition 5.12 in the book.

Exercise 5.8: When N is odd

When N is odd, the (first stage in a) DWT is defined as the change of coordinates
from (φ1,0 , φ1,1 , . . . , φ1,N −1 ) to

(φ0,0 , ψ0,0 , φ0,1 , ψ0,1 , . . . , φ0,(N −1)/2 , ψ(N −1)/2 , φ0,(N +1)/2 ).
Since all functions are assumed to have period N , we have that

1 1
φ0,(N +1)/2 = √ (φ1,N −1 + φ1,N ) = √ (φ1,0 + φ1,N −1 ).
2 2
From this relation one can find the last column in the change of coordinate
matrix from φ0 to (φ1 , ψ1 ), i.e. the IDWT matrix. In particular, when N is
odd, we see that the last column in the IDWT matrix circulates to the upper
right corner. In terms of coordinates, we thus have that

1 1
c1,0 = √ (c0,0 + w0,0 + c0,(N +1)/2 ) c1,N −1 = √ c0,(N +1)/2 . (5.3)
2 2
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES71

a) If N = 3, the DWT matrix equals

 
1 1 1
1
√ 1 −1 0 ,
2 0 0 1
and the inverse of this is
 
1 1 −1
1 
√ 1 −1 −1 .
2 0 0 2
Explain from this that, when N is odd, the DWT matrix can be constructed by
adding a column on the form √12 (−1, −1, 0, . . . , 0, 2) to the DWT matrices we
had for N even (in the last row zeros are also added). In terms of the coordinates,
we thus have the additional formulas

1
c0,0 = √ (c1,0+c1,1 −c1,N −1 )
2
1
w0,0 = √ (c1,0−c1,1 −c1,N −1 )
2
1
c0,(N +1)/2 = √ 2c1,N −1 . (5.4)
2
b) Explain that the DWT matrix is orthogonal if and only if N is even. Also
explain that it is only the last column which spoils the orthogonality.

Exercise 5.13: Implement IDWT for The Haar wavelet

Write a function idwt_kernel_haar which uses the formulas (5.3) to implement
the IDWT, similarly to how the function dwt_kernel_haar implemented the
DWT using the formulas (5.4).

Solution. The following code can be used:

def idwt_kernel_haar(x, bd_mode):

x /= sqrt(2)
if mod(len(x), 2)==1:
a, b = x[0] + x[1] + x[-1], x[0] - x[1]
x[0], x[1] = a, b
for k in range(2,len(x) - 2, 2):
a, b = x[k] + x[k+1], x[k] - x[k+1]
x[k], x[k+1] = a, b
else:
for k in range(0,len(x) - 1, 2):
a, b = x[k] + x[k+1], x[k] - x[k+1]
x[k], x[k+1] = a, b
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES72

Exercise 5.14: Computing projections 4

Generalize Exercise 5 to the projections from Vm+1 onto Vm and Wm .

Solution. Since φm,n ∈ Vm we must have that T (φm,n ) = φm,n . Since ψm,n
is in the orthogonal complement of Vm in Vm+1 we must have that T (ψm,n ) = 0.
The first half of the columns in the matrix of projVm relative to (φm , ψm ) are

[projVm (φm,0 )](φm ,ψm ) = [φm,0 ](φm ,ψm ) = e0

[projVm (φm,1 )](φm ,ψm ) = [φm,1 ](φm ,ψm ) = e1
....
...

The second half of the columns are

[T (ψm,0 )](φm ,ψm ) = [0](φm ,ψm ) = 0

[T (ψm,1 )](φm ,ψm ) = [0](φm ,ψm ) = 0
....
...

Thus, as before, the matrix of projVm relative to (φm , ψm ) is given by the

diagonal matrix where the first half of the diagonal consists of 1’s, and the
second half consists of 0’s. c) follows in the same way.

Exercise 5.15: Scaling a function

Show that f (t) ∈ Vm if and only if g(t) = f (2t) ∈ Vm+1 .

P2m N −1
Solution. If f ∈ Vm we can write f (t) = n=0 cm,n φm,n (t). We now get

2mX
N −1 2mX
N −1
g(t) = f (2t) = cm,n φm,n (2t) = cm,n 2m/2 φ(2m 2t − n)
n=0 n=0
2mX
N −1 2mX
N −1
−1/2 (m+1)/2
= cm,n 2 2 φ(2 m+1
t − n) = cm,n 2−1/2 φm+1,n (t).
n=0 n=0

This shows that g ∈ Vm+1 . To prove the other way, assume that g(t) = f (2t) ∈
P2m+1 N −1
Vm+1 . This means that we can write g(t) = n=0 cm+1,n φm+1,n (t). We
now have
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES73

2m+1
X N −1 2m+1
X N −1
f (t) = g(t/2) = cm+1,n φm+1,n (t/2) = cm+1,n 2(m+1)/2 φ(2m t − n)
n=0 n=0
2mX
N −1 2m+1
X N −1
= cm+1,n 2(m+1)/2 φ(2m t − n) + cm+1,n 2(m+1)/2 φ(2m t − n)
n=0 n=2m N
2mX
N −1 2mX
N −1
= cm+1,n 2(m+1)/2 φ(2m t − n) + cm+1,n+2m N 2(m+1)/2 φ(2m t − n − 2m N )
n=0 n=0
2mX
N −1 2mX
N −1
= cm+1,n 2(m+1)/2 φ(2m t − n) + cm+1,n+2m N 2(m+1)/2 φ(2m t − n)
n=0 n=0
2mX
N −1
= (cm+1,n + cm+1,n+2m N )21/2 2m/2 φ(2m t − n)
n=0
2mX
N −1
= (cm+1,n + cm+1,n+2m N )21/2 φm,n (t) ∈ Vm
n=0

The thing which made this a bit difficult was that the range of the n-indices
here was outside [0, 2m N − 1] (which describe the legal indices in the basis Vm ),
so that we had to use the periodicity of φ.

Exercise 5.16: Direct sums

Let C1 , C2 . . . , Cn be independent vector spaces, and let Ti : Ci → Ci be linear
transformations. The direct sum of T1 , T2 ,. . . ,Tn , written as T1 ⊕ T2 ⊕ . . . ⊕ Tn ,
denotes the linear transformation from C1 ⊕ C2 ⊕ · · · ⊕ Cn to itself defined by

T1 ⊕ T2 ⊕ . . . ⊕ Tn (c1 + c2 + · · · + cn ) = T1 (c1 ) + T2 (c2 ) + · · · + Tn (cn )

when c1 ∈ C1 , c2 ∈ C2 , . . . , cn ∈ Cn . Similarly, when A1 , A2 , . . . , An are square

matrices, A1 ⊕ A2 ⊕ · · · ⊕ An is defined as the block matrix where the blocks
along the diagonal are A1 , A2 , . . . , An , and where all other blocks are 0. Show
that, if Bi is a basis for Ci then

[T1 ⊕ T2 ⊕ . . . ⊕ Tn ](B1 ,B2 ,...,Bn ) = [T1 ]B1 ⊕ [T2 ]B2 ⊕ · · · ⊕ [Tn ]Bn ,

Here two new concepts are used: a direct sum of matrices, and a direct sum of
linear transformations.

Solution. By definition, [T1 ]B1 ⊕ [T2 ]B2 ⊕ · · · ⊕ [Tn ]Bn is a block matrix where
the blocks on the diagonal are the matrices [T1 ]B1 , [T2 ]B2 , and so on. If bi are
the basis vectors in Bi , the columns in [Ti ]Bi are [T (bj )]Bi . This means that
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES74

[T1 ]B1 ⊕ [T2 ]B2 ⊕ · · · ⊕ [Tn ]Bn has [T (bj )]Bi in the j’th block, and 0 elsewhere.
This means that we can write it as

0 ⊕ · · · 0 ⊕ [T (bj )]Bi ⊕ 0 · · · 0.
On the other hand, [T1 ⊕ T2 ⊕ . . . ⊕ Tn ](B1 ,B2 ,...,Bn ) is a matrix of the same
size, and the corresponding column to that of the above is

[(T1 ⊕ T2 ⊕ . . . ⊕ Tn )(0 ⊕ · · · 0 ⊕ bj ⊕ 0 · · · 0)](B1 ,B2 ,...,Bn )

= [0 ⊕ · · · 0 ⊕ T (bj ) ⊕ 0 · · · 0](B1 ,B2 ,...,Bn )
= 0 ⊕ · · · 0 ⊕ [T (bj )]Bi ⊕ 0 · · · 0.

Here bj occurs as the i’th summand. This is clearly the same as what we
computed for the right hand side above.

Exercise 5.17: Eigenvectors of direct sums

Assume that T1 and T2 are matrices, and that the eigenvalues of T1 are equal
to those of T2 . What are the eigenvalues of T1 ⊕ T2 ? Can you express the
eigenvectors of T1 ⊕ T2 in terms of those of T1 and T2 ?

Solution. Assume that λ is an eigenvalue common to both T1 and T2 . Then

there exists a vector v1 so that T1 v1 = λv1 , and a vector v2 so that T2 v2 = λv2 .
We now have that

  
T1 0 v1
(T1 ⊕ T2 )(v1 ⊕ v2 ) =
0 T2 v2
   
T1 v1 λv1
= =
T2 v2 λv2
 
v
= λ 1 = λ(v1 ⊕ v2 ).
v2

This shows that λ is an eigenvalue for λ also, and that v1 ⊕ v2 is a corresponding

eigenvector.

Exercise 5.18: Invertibility of direct sums

Assume that A and B are square matrices which are invertible. Show that A ⊕ B
is invertible, and that (A ⊕ B)−1 = A−1 ⊕ B −1 .

Solution. We have that

CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES75

   −1 
A 0 A 0
(A ⊕ B)(A−1 ⊕ B −1 ) =
0 B 0 B −1
AA−1
   
0 I 0
= = =I
0 BB −1 0 I

where we have multiplied as block matrices. This proves that A ⊕ B is invertible,

and states what the inverse is.

Exercise 5.19: Multiplying direct sums

Let A, B, C, D be square matrices of the same dimensions. Show that (A ⊕
B)(C ⊕ D) = (AC) ⊕ (BD).

Solution. We have that

    
A 0 C 0 AC 0
(A ⊕ B)(C ⊕ D) = = = (AC) ⊕ (BD)
0 B 0 D 0 BD

where we again have multiplied as block matrices.

Exercise 5.20: Finding N

Assume that you run an m-level DWT on a vector of length r. What value of
N does this correspond to? Note that an m-level DWT performs a change of
coordinates from φm to (φ0 , ψ0 , ψ1 , . . . , ψm−2 , ψm−1 ).

Exercise 5.21: Different DWTs for similar vectors

In Figure 5.1 we have plotted the DWT’s of two vectors x1 and x2 . In both
vectors we have 16 ones followed by 16 zeros, and this pattern repeats cyclically
so that the length of both vectors is 256. The only difference is that the second
vector is obtained by delaying the first vector with one element.
You see that the two DWT’s are very different: For the first vector we see
that there is much detail present (the second part of the plot), while for the
second vector there is no detail present. Attempt to explain why this is the case.
Based on your answer, also attempt to explain what can happen if you change
the point of discontinuity for the piecewise constant function in the left part of
Figure 5.11 in the book to something else.

Exercise 5.22: Construct a sound

Attempt to construct a (nonzero) sound where the low resolution approximations
equal the sound itself for m = 1, m = 2.
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES76

1.2 1.2
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 50 100 150 200 250 0.0 50 100 150 200 250
Figure 5.1: 2 vectors x1 and x2 which seem equal, but where the DWT’s are
very different.

Exercise 5.23: Exact computation of wavelet coefficients 1

Compute the wavelet detail coefficients analytically for the functions in Ex-
RN
ample 5.11 in the book, i.e. compute the quantities wm,n = 0 f (t)ψm,n (t)dt
similarly to how this was done in Example 5.12 in the book.

Solution. Note first that, similarly to the computation in Exercise 5, we have

that

(n+1/2)2−m (n+1)2−m
!
Z N Z Z
m/2
f (t)ψm,n (t)dt = 2 f (t)dt − f (t)dt .
0 n2−m (n+1/2)2−m

With f (t) = 1 − 2|1/2 − t/N | we have two possibilities: when n < N 2m−1
we have that [n2−m , (n + 1)2−m ) ⊂ [0, N/2], so that f (t) = 2t/N , and we get

(n+1/2)2−m (n+1)2−m
Z Z !
m/2
wm,n = 2 2t/N dt − 2t/N dt
n2−m (n+1/2)2−m
(n+1/2)2−m (n+1)2−m
= 2m/2 [t2 /N ]n2−m − 2m/2 [t2 /N ](n+1/2)2−m
2−3m/2 2−3m/2−1
2(n + 1/2)2 − n2 − (n + 1)2 = −


= .
N N
RN
When n ≥ N 2m−1 we have that f (t) = 2−2t/N , and using that 0 ψm,n (t)dt = 0
−3m/2−1
we must get that wm,n = 2 N .
For f (t) = 1/2 + cos(2πt/N )/2, note first that this has the same coefficients
RN
as cos(2πt/N )/2, since 0 ψm,n (t)dt = 0. We now get
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES77

(n+1/2)2−m (n+1)2−m
Z Z !
m/2
wm,n = 2 cos(2πt/N )/2dt − cos(2πt/N )/2dt
n2−m (n+1/2)2−m
(n+1/2)2−m (n+1)2−m
= 2m/2 [N sin(2πt/N )/(4π)]n2−m − 2m/2 [N sin(2πt/N )/(4π)](n+1/2)2−m
2m/2−2 N
2 sin(2π(n + 1/2)2−m /N ) − sin(2πn2−m /N ) − sin(2π(n + 1)2−m /N ) .


=
π
There seems to be no more possibilities for simplification here.

Exercise 5.24: Exact compution of wavelet coefficients 2

k
Compute the wavelet detail coefficients analytically for the functions f (t) = Nt ,
R N t
k
i.e. compute the quantities wm,n = 0 N ψm,n (t)dt similarly to how this was
done in Example 5.12 in the book. How do these compare with the coefficients
from the Exercise 5?

Solution. We get

(n+1/2)2−m (n+1)2−m
Z Z !
m/2 k k
wm,n = 2 (t/N ) dt − (t/N ) dt
n2−m (n+1/2)2−m
(n+1/2)2−m (n+1)2−m
= 2m/2 [tk+1 /((k + 1)N k )]n2−m − 2m/2 [tk+1 /((k + 1)N k )](n+1/2)2−m
2−m(k+1/2)
2(n + 1/2)k+1 − nk+1 − (n + 1)k+1 .


= k
(k + 1)N

The leading term nk+1 will here cancel, but the others will not, so there is no
room for further simplification here.

Exercise 5.25: Computing the DWT of a simple vector

Suppose that we have the vector x with length 210 = 1024, defined by xn = 1
for n even, xn = −1 for n odd. What will be the result if you run a 10-level
DWT on x? Use the function DWTImpl to verify what you have found.
√
Hint. We defined ψ by ψ(t) = (φ1,0 (t) √− φ1,1 (t))/ 2. From this connection
√ it
follows that ψ9,n = (φ10,2n − φ10,2n+1 )/ 2, and thus φ10,2n − φ10,2n+1 = 2ψ9,n .
Try to couple this identity with the alternating sign you see in x.

Solution. The vector x is the coordinate vector of the function √ f (t) =

P1023 n
(−1) φ 10,n in the basis φ10 for V 10 . Since φ 10,2n − φ 10,2n+1 = 2ψ9,n ,
n=0 P1023 √
we can write f (t) = n=0 2ψ9,n . Since a 10-level-DWT gives as a result the
coordinate vector of f in
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES78

(φ0 , ψ0 , ψ1 , ψ2 , ψ3 , ψ4 , ψ5 , ψ6 , ψ7 , ψ8 , ψ9 ),
(the DWT is nothing√but the change of coordinates from φ10 to this basis), and
P1023
since f (t) √= n=0 2ψ9,n , it is clear that the coordinate vector of f in this
basis has 2 in the second part (the ψ9 -coordinates), and 0 elsewhere. The
10-level DWT of x therefore
√ gives the vector of length 1024 which is 0 on the
first half, and equal to 2 on the second half. m = 10 is here arbitrarily chosen:
The result would have been the same for m = 1, m = 2, and so on. The following
code verifies the result:

x = tile([1.,-1.], 512)
DWTImpl(x, 10, ’Haar’)
print x

Exercise 5.26: The Haar wavelet when N is odd

Use the results from Exercise 5 to rewrite the implementations dwt_kernel_haar
and idwt_kernel_haar so that they also work in the case when N is odd.

Solution. The following code can be used.

def dwt_kernel_haar(x, bd_mode):

x /= sqrt(2)
if mod(len(x), 2)==1:
a, b = x[0] + x[1] - x[-1], x[0] - x[1] - x[-1]
x[0], x[1] = a, b
x[-1] *= 2
else:
a, b = x[0] + x[1], x[0] - x[1]
x[0], x[1] = a, b
for k in range(2,len(x) - 1,2):
a, b = x[k] + x[k+1], x[k] - x[k+1]
x[k], x[k+1] = a, b

def idwt_kernel_haar(x, bd_mode):

x /= sqrt(2)
if mod(len(x), 2)==1:
a, b = x[0] + x[1] + x[-1], x[0] - x[1]
x[0], x[1] = a, b
for k in range(2,len(x) - 2, 2):
a, b = x[k] + x[k+1], x[k] - x[k+1]
x[k], x[k+1] = a, b
else:
for k in range(0,len(x) - 1, 2):
a, b = x[k] + x[k+1], x[k] - x[k+1]
x[k], x[k+1] = a, b
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES79

Exercise 5.27: in-place DWT

Show that the coordinates in φm after an in-place m-level DWT end up at
indices k2m , k = 0, 1, 2, . . .. Show similarly that the coordinates in ψm after an
in-place m-level DWT end up at indices 2m−1 + k2m , k = 0, 1, 2, . . .. Find these
indices in the code for the function reorganize_coefficients.

Exercise 5.30: The vector of samples is the coordinate vec-

tor 2
Show that, for f ∈ V0 we have that [f ]φ0 = (f (0), f (1), . . . , f (N − 1)). This
shows that, also for the piecewise linear wavelet, there is no loss of information
in working with the samples of f rather than f itself.
PN −1
Solution. Let us write f (t) = n=0 cn φ0,n (t). If k is an integer we have that
N
X −1 N
X −1
f (k) = cn φ0,n (k) = cn φ(k − n).
n=0 n=0

Clearly the only integer for which φ(s) 6= 0 is s = 0 (since φ(0) = 1), so that the
only n which contributes in the sum is n = k. This means that f (k) = ck , so
that [f ]φ0 = (f (0), f (1), . . . , f (N − 1)).

Exercise 5.31: Computing projections

In this exercise we will show how the projection of φ1,1 onto V0 can be computed.
We will see from this that it is nonzero, and that its support is the entire [0, N ].
Let f = projV0 φ1,1 , and let xn = f (n) for 0 ≤ n < N . This means that, on
(n, n + 1), f (t) = xn + (xn+1 − xn )(t − n).
R n+1
a) Show that n f (t)2 dt = (x2n + xn xn+1 + x2n+1 )/3.

Solution. We have that

Z n+1 Z n+1 Z 1
f (t)2 dt = (xn + (xn+1 − xn )(t − n))2 dt = (xn + (xn+1 − xn )t)2 dt
n n 0
Z 1
= (x2n + 2xn (xn+1 − xn )t + (xn+1 − xn )2 t2 )dt
0
1
= x2n t + xn (xn+1 − xn )t2 + (xn+1 − xn )2 t3 /3 0


1
= x2n + xn (xn+1 − xn ) + (xn+1 − xn )2 /3 = (x2n + xn xn+1 + x2n+1 ).
3
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES80

b) Show that

1/2 √
Z  
1 1
(x0 + (x1 − x0 )t)φ1,1 (t)dt = 2 2 x0 + x1
0 12 24
1 √
Z  
1 1
(x0 + (x1 − x0 )t)φ1,1 (t)dt = 2 2 x0 + x1 .
1/2 24 12

Solution. We have that

Z 1/2
(x0 + (x1 − x0 )t)φ1,1 (t)dt
0
Z 1/2 √ √ Z 1/2
= (x0 + (x1 − x0 )t)2 2tdt = 2 2 (x0 t + (x1 − x0 )t2 )dt
0 0
1/2
√ 1 √
  
2 1 3 1 1
= 2 2 x0 t + (x1 − x0 )t =2 2 x0 + (x1 − x0 ))
2 3 0 8 24
√
 
1 1
=2 2 x0 + x1 .
12 24
In the same way

Z 1
(x0 + (x1 − x0 )t)φ1,1 (t)dt
1/2
Z 1 √√ Z 1
= (x0 + (x1 − x0 )t)2 2(1 − t)tdt = 2 2 (x0 + (x1 − 2x0 )t − (x1 − x0 )t2 )tdt
1/2 1/2
1
√ √
  
1 2 1 3 1 3 7
= 2 2 x0 t + (x1 − 2x0 )t − (x1 − x0 )t =2 2 x0 + (x1 − 2x0 ) − (x1 − x0 )
2 3 1/2 2 8 24
√
 
1 1
=2 2 x0 + x1 .
24 12
c) Use the fact that

Z N N
X −1
(φ1,1 (t) − xn φ0,n (t))2 dt
0 n=0
Z 1 Z 1/2 Z 1
= φ1,1 (t)2 dt − 2 (x0 + (x1 − x0 )t)φ1,1 (t)dt − 2 (x0 + (x1 − x0 )t)φ1,1 (t)dt
0 0 1/2
N
X −1 Z n+1
+ (xn + (xn−1 − xn )t)2 dt
n=0 n

PN −1
and a) and b) to find an expression for kφ1,1 (t) − n=0 xn φ0,n (t)k2 .
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES81

Solution. Using a) and b) we see that the above can be written as

N −1
√ √
    
2 X 1 2 1 1 1 1
+ (xn + xn xn+1 + x2n+1 ) − 2 2 2 x0 + x1 − 2 2 x0 + x1
3 n=0 3 12 24 24 12
N −1 N −1 √
2 2 X 2 1 X 2
= + xn + xn xn+1 − (x0 + x1 ).
3 3 n=0 3 n=0 2

d) To find the minimum least squares error, we can set the gradient of the
expression in c) to zero, and thus find the expression for the projection of φ1,1
−1
onto V0 . Show that the values {xn }Nn=0 can be found by solving the equation
1
Sx = b, where S = 3 {1, 4, 1} is an N × N symmetric filter, and b is the vector
√
with components b0 = b1 = 2/2, and bk = 0 for k ≥ 2.

Solution. We see that the partial derivatives of the function in c) are

√
∂f 1 4 1 2
= xN −1 + x0 + x1 −
∂x0 3 3 3 2
√
∂f 1 4 1 2
= x0 + x1 + x1 −
∂x1 3 3 3 2
∂f 1 4 1
= xi−1 + xi + xi+1 2 ≤ i < N − 1
∂xi 3 3 3
∂f 1 4 1
= xN −2 + xN −1 + x0 .
∂xN −1 3 3 3
√
Moving the two terms 22 over to the right hand side, setting the gradient equal
to zero is the same as solving the system Sx = b which we stated.
e) Solve the system in d. for some values of N to verify that the projection of
φ1,1 onto V0 is nonzero, and that its support covers the entire [0, N ].

Solution. The following code can be used

from numpy import *

import matplotlib.pyplot as plt
N = 16
S = zeros((N, N))
S[0,N-1] = 1/3.; S[0,0] = 4/3.; S[0,1] = 1/3.; # First row
for k in range(1,N-1):
S[k,(k-1):(k+2)] = [1/3., 4/3., 1/3.]
S[N-1,N-2] = 1/3.; S[N-1,N-1]=4/3.; S[N-1,0]=1/3.; # Last row
b=zeros(N); b[0]=sqrt(2)/2; b[1]=sqrt(2)/2;
plt.plot(range(0,N),linalg.solve(S,b)) # Plots the projection
plt.show()
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES82

Exercise 5.32: Non-orthogonality for the piecewise linear

wavelet
Show that

2 1
hφ0,n , φ0,n i = hφ0,n , φ0,n±1 i = hφ0,n , φ0,n±k i = 0 for k > 1.
3 6
As a consequence, the {φ0,n }n are neither orthogonal, nor have norm 1.

Solution. We have that

Z n+1
hφ0,n , φ0,n i = (1 − |t − n|)2 dt
n−1
Z n+1
1 − 2|t − n| + (t − n)2 dt


=
n−1
 n+1
1 2
=2−2+ (t − n)3 = .
3 n−1 3

We also have

Z n+1 Z 1
hφ0,n , φ0,n+1 = (1 − (t − n))(1 + (t − n − 1))dt = (1 − u)(1 + u − 1)du
n 0
Z 1
1 1 1
= (t − t2 )dt = − = .
0 2 3 6

Finally, the supports of φ0,n and φ0,n±k are disjoint for k > 1, so that we must
have hφ0,n , φ0,n±k i = 0 in that case.

Exercise 5.33: Implement elementary lifting steps of odd

type
Write functions lifting_odd_symm which takes λ, a vector x, and symm as input,
and applies an elementary lifting matrix of odd type (Equation (5.35) in the
book) to x. Assume that N is even. The parameter symm should do nothing,
we will return to this parameter later. lifting_odd_symm should not perform
matrix multiplication, and apply as few multiplications as possible.

Solution. The code can look like this:

def lifting_odd_symm(lmbda, x, bd_mode):

if (bd_mode.lower() == ’per’) and mod(len(x), 2)!=0:
raise AssertionError()
x[1:-1:2] += lmbda*(x[0:-2:2] + x[2::2])
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES83

if mod(len(x), 2)==0:
if bd_mode.lower() == ’symm’:
x[-1] += 2*lmbda*x[-2] # With symmetric extension
else:
x[-1] += lmbda*(x[0]+x[-2])

Since λ appears twice in each row, note that we have added entries before
multiplication, to avoid extra multiplication.

Exercise 5.34: Wavelets based on polynomials

The convolution of two functions defined on (−∞, ∞) is defined by
Z ∞
(f ∗ g)(x) = f (t)g(x − t)dt.
−∞

Show that we can obtain the piecewise linear φ we have defined as φ = χ[−1/2,1/2) ∗
χ[−1/2,1/2) (recall that χ[−1/2,1/2) is the function which is 1 on [−1/2, 1/2) and
0 elsewhere). This gives us a nice connection between the piecewise constant
scaling function (which is similar to χ[−1/2,1/2) ) and the piecewise linear scaling
function in terms of convolution.

Solution. We have that

Z ∞
χ[−1/2,1/2) ∗ χ[−1/2,1/2) (x) = χ[−1/2,1/2) (t)χ[−1/2,1/2) (x − t)dt.
−∞

The integrand here is 1 when −1/2 < t < 1/2 and −1/2 < x − t < 1/2, or
in other words when max(−1/2, −1/2 + x) < t < min(1/2, 1/2 + x) (else it is
0). When x > 0 this happens when −1/2 + x < t < 1/2, and when x < 0 this
happens when −1/2 < t < 1/2 + x. This means that
(R 1/2
dt = 1 − x , x > 0
χ[−1/2,1/2) ∗ χ[−1/2,1/2) (x) = R−1/2+x
1/2+x
−1/2
dt = 1 + x , x < 0.
But this is by definition φ.

Exercise 5.37: Implement elementary lifting steps of even

type
Write a function lifting_even_symm which takes λ, a vector x, and symm as
input, and applies an elementary lifting matrix of even type (Equation (5.44) in
the book) to x. As before, assume that N is even, and that the parameter symm
does nothing. lifting_even_symm should also here apply as few multiplications
as possible.
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES84

Solution. The code can look like this:

def lifting_even_symm(lmbda, x, bd_mode):

if (bd_mode.lower() == ’per’) and mod(len(x), 2)!=0:
raise AssertionError()
if bd_mode.lower() == ’symm’:
x[0] += 2*lmbda*x[1] # With symmetric extension
else:
x[0] += lmbda*(x[1]+x[-1])
x[2:-1:2] += lmbda*(x[1:-2:2] + x[3::2])
if mod(len(x), 2)==1 and bd_mode.lower() == ’symm’:
x[-1] += 2*lmbda*x[-2] # With symmetric extension

Since also here λ appears twice in each row, we have added entries before
multiplication, to avoid extra multiplication.

Exercise 5.38: Two vanishing moments

In this exercise we will show that there is a unique function on the form given
by Equation (5.38) in the book which has two vanishing moments.
a) Show that, when ψ̂ is defined by Equation (5.38) in the book, we have that



 −αt − α for − 1 ≤ t < 0
(2 + α − β)t − α for 0 ≤ t < 1/2



ψ̂(t) = (α − β − 2)t − α + 2 for 1/2 ≤ t < 1

βt − 2β for 1 ≤ t < 2





0 for all other t

Solution. The function ψ̂ is a sum of the functions ψ = φ1,1 , φ, and φ0,1

(i.e. we have set n = 0 in Equation (5.38) in the book). All these are continuous
and piecewise linear, and we can write


2t
 0 ≤ t < 1/2
φ1,1 (t) = 2 − 2t 1/2 ≤ t < 1

0 elsewhere


1 + t
 −1 ≤ t < 0
φ0,0 (t) = 1 − t 0≤t<1

0 elsewhere


t
 0≤t<1
φ0,1 (t) = 2 − t 1≤t<2 .

0 elsewhere


It follows that ψ̂(t) = φ1,1 (t) − αφ(t) − βφ1,1 is piecewise linear, and linear on
the segments [−1, 0], [0, 1/2], [1/2, 1], [1, 2].
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES85

On the segment [−1, 0] only the function φ is seen to be nonzero, and since
φ(t) = 1 + t here, we have that ψ̂(t) = −α(1 + t) = −α − αt here.
On the segment [0, 1/2] all three functions are nonzero, and

φ1,1 (t) = 2t
φ0,0 (t) = 1 − t
φ0,1 (t) = t

on this interval. This means that ψ̂(t) = 2t − α(1 − t) − βt = (2 + α − β)t − α

on [0, 1/2].
On the segment [0, 1/2] all three functions are nonzero, and

φ1,1 (t) = 2 − 2t
φ0,0 (t) = 1 − t
φ0,1 (t) = t

on this interval. This means that ψ̂(t) = 2−2t−α(1−t)−βt = (α−β −2)t−α+2

on [1/2, 1].
On the segment [1, 2] only the function φ0,1 is seen to be nonzero, and since
φ0,1 (t) = 2 − t here, we have that ψ̂(t) = −β(2 − t) = βt − 2β here. For all other
values of t, ψ̂ is zero. This proves the formulas for ψ̂ on the different intervals.
b) Show that

Z N Z N
1 1
ψ̂(t)dt = − α − β, tψ̂(t)dt = − β.
0 2 0 4

Solution. We can write

CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES86

Z N Z 2 Z 0 Z 1/2 Z 1 Z 2
ψ̂(t)dt = ψ̂(t)dt = ψ̂(t)dt + ψ̂(t)dt + ψ̂(t)dt + ψ̂(t)dt
0 −1 −1 0 1/2 1
Z 0 Z 1/2
= (−α − αt)dt + (2 + α − β)t − α)dt
−1 0
Z 1 Z 2
+ ((α − β − 2)t − α + 2)dt + (βt − 2β)dt
1/2 1
 0  1/2
1 2 1 2
= −αt − αt + (2 + α − β)t − αt
2 −1 2 0
 1  2
1 2 1 2
+ (α − β − 2)t + (2 − α)t + βt − 2βt
2 1/2 2 1
1 1 1 3 1 3
= −α + α + (2 + α − β) − α + (α − β − 2) + (2 − α) + β − 2β
2 8 2 8 2 2
1
= − α − β,
2
RN 1
0
tψ̂(t)dt is computed similarly, so that we in the end arrive at 4 − β.
c) Explain why there is a unique function on the form given by Equation (5.38)
in the book which has two vanishing moments, and that this function is given
by Equation (5.40) in the book.

Solution. The equation system

1
−α−β =0
2
1
−β =0
4
has the unique solution α = β = 41 , which we already have found.

Exercise 5.39: Implement finding ψ with vanishing mo-

ments
In the previous exercise we ended up with a lot of calculations to find α, β in
Equation (5.38) in the book. Let us try to make a program which does this for
us, and which also makes us able to generalize the result.
a) Define

Z 1 Z 2 Z 1
k k
ak = t (1 − |t|)dt, bk = t (1 − |t − 1|)dt, ek = tk (1 − 2|t − 1/2|)dt,
−1 0 0
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES87

for k ≥ 0. Explain why finding α, β so that we have two vanishing moments in

Equation (5.38) in the book is equivalent to solving the following equation:
    
a0 b0 α e
= 0
a1 b1 β e1
Write a program which sets up and solves this system of equations, and use this
program to verify the values for α, β we previously have found.

Hint. you can integrate functions in Python with the function quad in the
package scipy.integrate . As an example, the function φ(t), which is nonzero
only on [−1, 1], can be integrated as follows:

res, err = quad(lambda t: t**k*(1-abs(t)), -1, 1)

Solution.
R RIn order for ψ to have vanishing moments we must have that
ψ̂(t)dt = tψ̂(t)dt = 0 Substituting ψ̂ = ψ − αφ0,0 − βφ0,1 we see that,
for k = 0, 1,
Z Z
t (αφ0,0 + βφ0,1 ) dt = tk ψ(t)dt.
k

The left hand side can here be written

Z Z Z
k k
t (αφ0,0 + βφ0,1 ) dt = α t φ0,0 dt + β tk φ0,1 (t)dt
Z 1 Z 2
k
=α t (1 − |t|)dt + β tk (1 − |t − 1|)dt = αak + βbk .
−1 0

The right hand side is

Z Z Z 1
k k
t ψ(t)dt = t φ1,1 (t)dt = (1 − 2|t − 1/2|)dt = ek .
0
The following program sets up the corresponding equation systems, and solves it
by finding α, β.

A = zeros((2, 2))
b = zeros((2, 1))
for k in range(2):
res1, err1 = quad(lambda t: t**k*(1-abs(t)), -1, 1)
res2, err2 = quad(lambda t: t**k*(1-abs(t-1)), 0, 2)
res3, err3 = quad(lambda t: t**k*(1-2*abs(t-1/2.)), 0, 1)
A[k,:] = [res1, res2]
b[k] = res3
linalg.solve(A,b)
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES88

b) The procedure where we set up a matrix equation in a) allows for generaliza-

tion to more vanishing moments. Define

ψ̂ = ψ0,0 − αφ0,0 − βφ0,1 − γφ0,−1 − δφ0,2 . (5.5)

We would like to choose α, β, γ, δ so that we have 4 vanishing moments. Define
also

Z 0 Z 3
gk = tk (1 − |t + 1|)dt, dk = tk (1 − |t − 2|)dt
−2 1

for k ≥ 0. Show that α, β, γ, δ must solve the equation

    
a0 b0 g0 d0 α e0
a1 b1 g1 d1  β  e1 
   =  ,
a2 b2 g2 d2   γ  e2 
a3 b3 g3 d3 δ e3
and solve this with your computer.

Solution. Similarly to a), Equation (5.5) gives that

Z Z
t (αφ0,0 + βφ0,1 + γφ0,−1 + δφ0,2 ) dt = tk ψ(t)dt.
k

The corresponding equation system is deduced exactly as in a). The following

program sets up the corresponding equation systems, and solves it by finding
α, β, γ, δ.

A=zeros((4, 4))
b=zeros((4, 1))
for k in range(4):
res1, err1 = quad(lambda t: t**k*(1-abs(t)), -1, 1)
res2, err2 = quad(lambda t: t**k*(1-abs(t-1)), 0, 2)
res3, err3 = quad(lambda t: t**k*(1-abs(t+1)), -2, 0)
res4, err4 = quad(lambda t: t**k*(1-abs(t-2)), 1, 3)
res5, err5 = quad(lambda t: t**k*(1-2*abs(t-1/2.)), 0, 1)
A[k,:] = [res1, res2, res3, res4]
b[k] = res5
coeffs = linalg.solve(A,b)

c) Plot the function defined by (5.5), which you found in b).

Hint. If t is the vector of t-values, and you write

(t >= 0)*(t <= 1)*(1-2*abs(t-0.5))

you get the points φ1,1 (t).

CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES89

Solution. The function ψ̂ now is supported on [−2, 3], and can be plotted as
follows:

t=linspace(-2,3,100)
plt.plot(t, (t >= 0)*(t <= 1)*(1-2*abs(t - 0.5)) \
-coeffs[0]*(t >= -1)*(t <= 1)*(1 - abs(t)) \
-coeffs[1]*(t >= 0)*(t <= 2)*(1 - abs(t - 1)) \
-coeffs[2]*(t >= -2)*(t <= 0)*(1 - abs(t + 1)) \
-coeffs[3]*(t >= 1)*(t <= 3)*(1 - abs(t - 2)))

d) Explain why the coordinate vector of ψ̂ in the basis (φ0 , ψ0 ) is

[ψ̂](φ0 ,ψ0 ) = (−α, −β, −δ, 0, . . . , 0 − γ) ⊕ (1, 0, . . . , 0).

Hint. The placement of −γ may seem a bit strange here, and has to with
−1
that φ0,−1 is not one of the basis functions {φ0,n }N
n=0 . However, we have that
φ0,−1 = φ0,N −1 , i.e. φ(t + 1) = φ(t − N + 1), since we always assume that the
functions we work with have period N .
e) Sketch a more general procedure than the one you found in b)., which can
be used to find wavelet bases where we have even more vanishing moments.

Solution. If we define
K
X
ψ̂ = ψ0,0 − (αk φ0,−k − βk φ0,k+1 ) ,
k=0

we have 2k unknowns. These can be determined if we require 2k vanishing

moments.

Exercise 5.40: ψ for the Haar wavelet with two vanishing

moments
Let φ(t) be the function we used when we defined the Haar-wavelet.
a) Compute projV0 (f (t)), where f (t) = t2 , and where f is defined on [0, N ).

b) Find constants α, β so that ψ̂(t) = ψ(t)−αφ0,0 (t)−βφ0,1 (t) has two vanishing
moments, i.e. so that hψ̂, 1i = 0, and hψ̂, ti = 0. Plot also the function ψ̂.
R R R R
Hint.
R Start with R computing the integrals ψ(t)dt, tψ(t)dt, φ0,0 (t)dt, φ0,1 (t)dt,
and tφ0,0 (t)dt, tφ0,1 (t)dt.

c) Express φ and ψ̂ with the help of functions from φ1 , and use this to write
down the change of coordinate matrix from (φ0 , ψ̂0 ) to φ1 .
CHAPTER 5. MOTIVATION FOR WAVELETS AND SOME SIMPLE EXAMPLES90

Exercise 5.41: More vanishing moments for the Haar wavelet

It is also possible to add more vanishing moments to the Haar wavelet. Define

ψ̂ = ψ0,0 − a0 φ0,0 − · · · − ak−1 φ0,k−1 .

R l+1 r R1
Define also cr,l = l t dt, and er = 0 tr ψ(t)dt.

a) Show that ψ̂ has k vanishing moments if and only if a0 , . . . , ak−1 solves the
equation
    
c0,0 c0,1 ··· c0,k−1 a0 e0
 c1,0 c1,1 ··· c1,k−1 
  a1   e1 
   
= (5.6)

 .. .. .. .. . . 
  ..   .. 
   
 . . . .
ck−1,0 ck−1,1 · · · ck−1,k−1 ak−1 ek−1
b) Write a function vanishingmomshaar which takes k as input, solves Equation
(5.6), and returns the vector a = (a0 , a1 , . . . , ak−1 ).

Exercise 5.42: Listening experiments

Run the function forw_comp_rev_DWT for different m for the Haar wavelet, the
piecewise linear wavelet, and the alternative piecewise linear wavelet, but listen
to the detail components W0 ⊕ W1 ⊕ · · · ⊕ Wm−1 instead. Describe the sounds
you hear for different m, and try to explain why the sound seems to get louder
when you increase m.

Solution. The following code can be used:

m = 1
x, fs = forw_comp_rev_DWT(m, ’Haar’)
play(x, fs)

m = 1
x, fs = forw_comp_rev_DWT(m, ’pwl0’)
play(x, fs)

m = 1
x, fs = forw_comp_rev_DWT(m, ’pwl2’)
play(x, fs)
Chapter 6

The filter representation of

wavelets

Exercise 6.5: Implement the dual filter bank transforms

a) Show that ATλ = Bλ and BλT = Aλ , i.e. that the transpose of an elementary
lifting matrix of even/odd type is an elementary lifting matrix of odd/even type.
b) Let H be the kernel of the DWT, and assume that we have a factorization
of it in terms of elementary lifting matrices. Use a) how that the dual DWT is
obtained from the DWT by replacing each Aλ with B−λ , and Bλ with A−λ in
this factorization.

Solution. The kernel of the dual DWT is GT = (H −1 )T . Transposing and

inverting means that the terms are reversed twice, so that the order is the same.
Transposing transforms odd liftings to even liftings and vice versa (according to
a)), and inversion adds the minus sign. The result follows.
c) Previously we expressed the DWT and the IDWT of the piecewise linear
wavelets in terms of elementary liftings. Use b) to write down the dual DWT and
IDWT of these two wavelets in terms of lifting matrices. Verify your answer by
going through the code in the functions dwt_kernel_pwl0, idwt_kernel_pwl0,
dwt_kernel_pwl2, and idwt_kernel_pwl2 where the dual parameter is set to
true..

Exercise 6.6: Implementing the transpose transforms

Since the dual of a wavelet is constructed by transposing filters, one may suspect
that taking the dual is the same as taking the transpose. However, show that
the transposes of the DWT, the dual DWT, the IDWT, and the dual IDWT,
can be computed as follows:

91
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 92

IDWTImpl(x, m, wave_name, True, True); % Transpose of the DWT

IDWTImpl(x, m, wave_name, True, False); % Transpose of the dual DWT
DWTImpl(x, m, wave_name, True, True); % Transpose of the IDWT
DWTImpl(x, m, wave_name, True, False); % Transpose of the dual IDWT

Solution. Assume that the kernel transformations of the DWT and the IDWT
are H and G, respectively. For the transpose of the DWT the point is that, while
its kernel transformation is H T , we compose the kernel with a permutation matrix
when we compute the DWT. When we transpose, the order of the kernel and
the permutation changes, so the transpose must use an IDWT implementation
instead. The remaining code follows similarly, since the kernel of the transpose
of the dual DWT is, G that of the transpose of the IDWT is GT , and that of
the transpose of the dual IDWT is H.

Exercise 6.7: Reduced matrices for elementary lifting

Show that the reduced matrices for elementary lifting are

 
1 2λ 0 0 ··· 0 0 0
0 1 0 0
 ··· 0 0 0 
0 λ 1 λ ··· 0 0 0
(Aλ )r =  . (6.1)
 
.. .. .. .. .. .. .. 
 .. . . . . . . .
 
0 0 0 0 · · · λ 1 λ
0 0 0 0 ··· 0 0 1
 
1 0 0 0 ··· 0 0 0
λ 1 λ 0
 · · · 0 0 0 
0 0 1 0 · · · 0 0 0
(Bλ )r =  . . . . ..  . (6.2)
 
.. .. ..
 .. .. .. .. . . . .
 
0 0 0 0 · · · 0 1 0
0 0 0 0 · · · 0 2λ 1

Also, change the implementations of liftingstevensymm and liftingstoddsymm

so that these expressions are used (rather than Aλ , Bλ ) when the parmeter symm
is set to True.

Solution. Any elementary lifting matrix can be viewed as wavelet transforms

where the filters are {λ, 1, λ} and {1}. Since these are symmetric, Theorem 6.11
in the book gives that they preserve symmetric extensions.

Exercise 6.8: Prove expression for Sr

Show that, with
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 93

 
S1 S2
S= ∈ R2N −2 × R2N −2
S3 S4
a symmetric filter, with S1 ∈ RN × RN , S2 ∈ RN × RN −2 , we have that

(S2 )f


Sr = S1 + 0 0 .
Use the proof of Theorem 4.9 in the book as a guide.

Solution. We compute

 
x0
 .. 
 .         
x0 xN −2 x0 x1
 

xN −2 
 
 .   .   .  f  . 
Sr x = S1 S2  xN −1  = S1  ..  + S2  ..  = S1  ..  + (S2 )  .. 

xN −2 
  xN −1 x1 xN −1 xN −2
 . 
 .. 
x1
   
x0 x0
= S1  ...  + 0 (S2 )f 0  ...  = (S1 + 0 (S2 )f 0 )x,
 
 

xN −1 xN −1

so that

(S2 )f


Sr = S1 + 0 0 .

Exercise 6.9: Orthonormal basis for the symmetric exten-

sions
In this exercise we will establish an orthonormal basis for the symmetric exten-
sions, as defined by Definition 6.10 in the book. This parallels Theorem 4.6 in
the book.
a) Explain why, if x ∈ R2N −2 is a symmetric extension (according to Defini-
x)n = zn e−πin , where z is a real vectors which
tion 4.1 in the book), then (b
satisfies zn = z2N −2−n

Solution. Using Theorem 4.3 in the book with d = N − 1 and with 2N − 2

for N , we obtain that

x)n = zn e−2πidn/(2N −2) = zn e−2πi(N −1)n/(2N −2) = zn e−πin ,

(b
where z is a real vectors which satisfies zn = z2N −2−n .
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 94

b) Show that
(  N −2 )
1
e0 , √ (ei + e2N −2−i ) , eN −1 (6.3)
2 n=1

b with x ∈ R2N −2 a
is an orthonormal basis for the vectors on the form x
symmetric extension.

Solution. Clearly these vectors are an orthonormal basis for the set of vectors
where zn = z2N −2−n . The vectors from a) are obtained by multiplying these
with e−πin . But the orthonormality of these vectors are not affected when we
multiply with e−πin , so we may skip this.
c) Show that

 
1 0
√ cos 2π k
2N − 2 2N − 2
  N −2
1 n
√ cos 2π k
N −1 2N − 2 n=1
 
1 N −1
√ cos 2π k (6.4)
2N − 2 2N − 2

is an orthonormal basis for the symmetric extensions in R2N −2 .

Solution. We compute the IDFT for all vectors in b). Since the IDFT is
unitary, this will give us an orthonormal basis for the symmetric vectors in
R2N −2 . Since (FN )H φn = en we get that

 
H 1 0
(FN ) e0 = φ0 = √ cos 2π k
2N − 2 2N − 2
 
1 1
(FN )H √ (en + e2N −2−n ) = √ (φn + φ2N −2−n )
2 2
1 1  
=√ √ e2πikn/(2N −2) + e−2πikn/(2N −2)
2 2N − 2
 
1 n
=√ cos 2π k
N −1 2N − 2
 
1 N −1
(FN )H eN −1 = φN −1 = √ cos 2π k .
2N − 2 2N − 2
These coincide with the vectors listed in the exercise.
d) Assume that S is symmetric. Show that the vectors listed in (6.4) are
eigenvectors for Sr , when the vectors are viewed as vectors in RN , and that they
are linearly independent. This shows that Sr is diagonalizable.
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 95

Solution. Since S is symmetric, it preserves vectors which are symmetric

around N − 1. In the frequency domain, applying S to a vector listed in (6.4)
corresponds to multiplying the vectors listed in Equation (6.3) with the frequency
response. Since this does not introduce any more components, it is clear that the
new vector must be a multiplum of the same vector, so that these vectors indeed
are eigenvectors. But then the vectors restricted to RN are also eigenvectors
for Sr , since this is simply S when viewed on the first N elements. Since the
vectors in R2N −2 are linearly independent, it is imemdiate that the corresponding
vectors in RN also are linearly independent, since the second part of the vectors
mirror the first part.

Exercise 6.10: Diagonalizing Sr

Let us explain how the matrix Sr can be diagonalized, similarly to how we
previously diagonalized using the DCT. In Exercise 6 we showed that the vectors

  N −1
n
cos 2π k (6.5)
2N − 2 n=0

in RN is a basis of eigenvectors for Sr when S is symmetric. Sr itself is not

symmetric, however, so that this basis can not possibly be orthogonal (S is
symmetric if and only if it is orthogonally digonalizable). However, when the
vectors are viewed in R2N −2 we showed in Exercise 6c) an orthogonality statement
which can be written as


2N −3     2
 if n1 = n2 ∈ {0, N − 1}
X n1 n2
cos 2π k cos 2π k = (N − 1) × 1 if n1 = n2 6∈ {0, N − 1} .
2N − 2 2N − 2 
k=0 
0 if n1 6= n2
(6.6)

a) Show that


1
 if n1 = n2 ∈ {0, N − 1}
1
(N − 1) × if n1 = n2 6∈ {0, N − 1}
2
0 if n1 6= n2

   
1 n1 1 n2
= √ cos 2π · 0 √ cos 2π ·0
2 2N − 2 2 2N − 2
N −2    
X n1 n2
+ cos 2π k cos 2π k
2N − 2 2N − 2
k=1
   
1 n1 1 n2
+ √ cos 2π (N − 1) √ cos 2π (N − 1) .
2 2N − 2 2 2N − 2
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 96

Hint. Use that cos x = cos(2π − x) to pair the summands k and 2N − 2 − k.

Solution. Using that cos x = cos(2π − x) we can here pair the summands k
and 2N − 2 − k to obtain

2N −3    
X n1 n2
cos 2π k cos 2π k
2N − 2 2N − 2
k=0
   
n1 n2
= cos 2π · 0 cos 2π ·0
2N − 2 2N − 2
N −2    
X n1 n2
+2 cos 2π k cos 2π k
2N − 2 2N − 2
k=1
   
n1 n2
+ cos 2π (N − 1) cos 2π (N − 1) .
2N − 2 2N − 2

If we divide by 2 and combine these equations we get the result.

(I)
Now, define the vector dn as

    N −2  !
1 n n 1 n
dn,N √ cos 2π · 0 , cos 2π k , √ cos 2π (N − 1) ,
2 2N − 2 2N − 2 k=1 2 2N − 2

(I) (I) √ (I) p

and define d0,N = dN −1,N = 1/ N − 1, and dn,N = 2/(N − 1) when n > 1.
(I)
The orthogonal N × N matrix where the rows are dn is called the DCT-I,
(I)
and we will denote it by DN . DCT-I is also much used, just as the DCT-II of
Chapter 4 in the book. The main difference from the previous cosine vectors is
that 2N has been replaced by 2N − 2.
(I)
b) Explain that the vectors dn are orthonormal, and that the matrix

 √   √ 
1/ 2 0 ··· 0 0 1/ 2 0 ··· 0 0
r  0 1 ··· 0 0   0
 1 ··· 0 0
2 
 .. .. .. ..

..

n  . .. .. ..

..
 cos 2π 2N −2 k  ..
 
N −1 . . . . . . . . .
 
  
 0 0 ··· 1 0√   0 0 ··· 1 0√ 
0 0 ··· 0 1/ 2 0 0 ··· 0 1/ 2

is orthogonal.
  −1
c) Explain from b) that cos 2π 2Nn−2 k can be written as
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 97

   
1/2 0 ··· 0 0 1/2 0 ··· 0 0
 0 1 ··· 0 0   0 1 ··· 0 0 
2 
 .. .. .. .. ..
   
 . .. .. .. ..

 cos 2π 2Nn−2 k  ..
 
N −1 . . . . . . . . .
 
  
 0 0 ··· 1 0   0 0 ··· 1 0 
0 0 ··· 0 1/2 0 0 ··· 0 1/2

With the expression we found in c) Sr can now be diagonalized as

     −1
cos 2π 2Nn−2 k D cos 2π 2Nn−2 k .

Exercise 6.11: Compute filters and frequency responses 1

Write down the corresponding filters G0 og G1 for Exercise 5. Plot their frequency
responses, and characterize the filters as low-pass- or high-pass filters.

Exercise 6.12: Symmetry of MRA matrices vs. symmetry

of filters 1
Find two symmetric filters, so that the corresponding MRA-matrix, constructed
with alternating rows from these two filters, is not a symmetric matrix.

Solution. You can set for instance H0 = {1/4, 1/2, 1/4}, and H1 = {1} (when
you write down the corresponding matrix you will see that A0,1 = 1/2, A1,0 = 0,
so that the matrix is not symmetric)

Exercise 6.13: Symmetry of MRA matrices vs. symmetry

of filters 2
Assume that an MRA-matrix is symmetric. Are the corresponding filters H0 ,
H1 , G0 , G1 also symmetric? If not, find a counterexample.

Solution. The Haar wavelet is a counterexample.

Exercise 6.14: Finding H0 , H1 from H

Assume that one stage in a DWT is given by the MRA-matrix

 
1/5 1/5 1/5 0 0 0 ··· 0 1/5 1/5
−1/3 1/3 −1/3 0 0 0 ··· 0 0 0 
 
H =  1/5
 1/5 1/5 1/5 1/5 0 ··· 0 0 0 
 0 0 −1/3 1/3 −1/3 0 ···0 0 0 
.. .. .. .. .. .. .. .. .. ..
 
. . . . . . . . . .
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 98

Write down the compact form for the corresponding filters H0 , H1 , and compute
and plot the frequency responses. Are the filters symmetric?

1 1
Solution. We have that H0 = 5 {1, 1, 1, 1, 1}, and H1 = 3 {−1, 1, −1}. The
frequency responses are

1 2iω 1 iω 1 1 −iω 1 −2iω

λH0 (ω) = e + e + + e + e
5 5 5 5 5
2 2 1
= cos(2ω) + cos ω +
5 5 5
1 iω 1 1 −iω 2 1
λH1 (ω) = − e + − e = − cos ω + .
3 3 3 3 3
Both filters are symmetric.

Exercise 6.15: Finding G0 ,G1 from G

Assume that one stage in the IDWT is given by the MRA-matrix
 
1/2 −1/4 0 0 ···
1/4 3/8
 1/4 1/16 · · · 
 0
 −1/4 1/2 −1/4 · · · 
 0
 1/16 1/4 3/8 · · · 
 0 0 0 −1/4 · · ·
G= 0
 
 0 0 1/16 · · · 
 0
 0 0 0 · · ·

 .. .. .. .. .. 
 . . . . . 
 
 0 0 0 0 · · ·
1/4 1/16 0 0 ···
Write down the compact form for the filters G0 , G1 , and compute and plot the
frequency responses. Are the filters symmetric?

Solution. We have that G0 = {1/4, 1/2, 1/4}, and G1 = {1/16, −1/4, 3/8, −1/4, 1/16}.
The frequency responses are

1 iω 1 1 −iω
λG0 (ω) = e + + e
4 2 4
1 1
= cos(ω) +
2 2
1 2iω 1 iω 3 1 −iω 1 −2iω
λG1 (ω) = e − e + − e e
16 4 8 4 16
1 1 3
= cos(2ω) − cos ω + .
8 2 8
Both filters are symmetric.
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 99

Exercise 6.16: Finding H from H0 , H1

Assume that H0 = {1/16, 1/4, 3/8, 1/4, 1/16}, and H1 = {−1/4, 1/2, −1/4}.
Plot the frequency responses of H0 and H1 , and verify that H0 is a lowpass
filter, and that H1 is a highpass filter. Also write down the change of coordinate
matrix PC1 ←φ1 for the wavelet corresponding to these filters.

Solution. The frequency responses are

1 2iω 1 iω 3 1 −iω 1 −2iω

λH0 (ω) = e + e + + e e
16 4 8 4 16
1 1 3
= cos(2ω) + cos ω +
8 2 8
1 iω 1 1 −iω
λH1 (ω) = − e + − e
4 2 4
1 1
= − cos(ω) + .
2 2
The two first rows in PC1 ←φ1 are
 
3/8 1/4 1/16 0 ··· 1/16 1/4
−1/4 1/2 −1/4 0 ··· 0 0
The remaining rows are obtained by translating these in alternating order.

Exercise 6.17: Finding G from G0 , G1

Assume that G0 = 13 {1, 1, 1}, and G1 = 15 {1, −1, 1, −1, 1}. Plot the frequency
responses of G0 and G1 , and verify that G0 is a lowpass filter, and that G1 is a
highpass filter. Also write down the change of coordinate matrix Pφ1 ←C1 for the
wavelet corresponding to these filters.

Solution. The frequency responses are

1 iω 1 1 −iω 2 1
λG0 (ω) = e + + e = cos ω +
3 3 3 3 3
1 2iω 1 iω 1 1 −iω 1 −2iω
λG1 (ω) = e − e + − e e
5 5 5 5 5
2 2 1
= cos(2ω) − cos ω +
5 5 5
The two first columns in Pφ1 ←C1 are
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 100

 
1/3 −1/5
1/3 1/5 
 
 0
 −1/5
 0 1/5 
 
 0 0 
 
 .. .. 
 . . 
 
 0 0 
1/3 1/5
The remaining columns are obtained by translating these in alternating order.

Exercise 6.18: Computing by hand

In Exercise 5 we computed the DWT of two very simple vectors x1 and x2 ,
using the Haar wavelet.
a) Compute H0 x1 , H1 x1 , H0 x2 , and H1 x2 , where H0 and H1 are the filters
used by the Haar wavelet.
b) Compare the odd-indexed elements in H1 x1 with the odd-indexed elements
in H1 x2 . From this comparison, attempt to find an explanation to why the two
vectors have very different detail components.

Exercise 6.19: Comment code

Suppose that we run the following algorithm on the sound represented by the
vector x:

c = (x[0::2] + x[1::2])/sqrt(2)
w = (x[0::2] - x[1::2])/sqrt(2)
newx = concatenate([c, w])
newx /= abs(newx).max()
play(newx,44100)

a) Comment the code and explain what happens. Which wavelet is used? What
do the vectors c and w represent? Describe the sound you believe you will hear.

Solution. c and w represent the coordinates in the wavelet bases φ0 and ψ0 .

The code runs a Haar wavelet transform. The sound is normalized so that the
sound samples lie in the range between −1 and 1, and the resulting sound is
played. The sound is split into two parts, and c represents a low-resolution
version of the sound (with half the number of samples), so that we first will
hear the sound played at double pace. After this we will hear the detail w in
the sound, also played at double pace. We should also be able to recognize the
sound from this detail.
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 101

b) Assume that we add lines in the code above which sets the elements in the
vector w to 0 before we compute the inverse operation. What will you hear if
you play the new sound you then get?

Solution. This corresponds to reconstructing a low-resolution approximation

of the sound.

Exercise 6.20: Computing filters and frequency responses

Let us return to the piecewise linear wavelet from Exercise 5.
a) With ψ̂ as defined as in b) in Exercise 5, compute the coordinates of ψ̂ in
the basis φ1 (i.e. [ψ̂]φ1 ) with N = 8, i.e. compute the IDWT of

[ψ̂](φ0 ,ψ0 ) = (−α, −β, −δ, 0, 0, 0, 0, −γ) ⊕ (1, 0, 0, 0, 0, 0, 0, 0),

which is the coordinate vector you computed in d) in Exercise 5. For this, you
should use the function IDWTImpl, with the kernel of the piecewise linear wavelet
without symmetric extension as input. Explain that this gives you the filter
coefficients of G1 .

Solution. The code which can be used looks like this:

g1 = array([-coeffs[0],-coeffs[1],-coeffs[3],0,0,0,0,-coeffs[2],\
1, 0, 0, 0, 0, 0, 0, 0])
IDWTImpl(g1, 1, ’pwl0’, ’per’)
g1 = hstack([g1[13:16],g1[0:6]]) # Compact filter notation

Note that we have used a kernel which does not make symmetric extensions.
b) Plot the frequency response of G1 .

Solution. The code can look as follows:

omega = linspace(0,2*pi,100)
plt.plot(omega, g1[4] + g1[5]*2*cos(omega) + g1[6]*2*cos(2*omega) \
+ g1[7]*2*cos(3*omega) + g1[8]*2*cos(4*omega))

Exercise 6.21: Computing filters and frequency responses

2
Repeat the previous exercise for the Haar wavelet as in Exercise 5, and plot the
corresponding frequency responses for k = 2, 4, 6.

Exercise 6.22: Implementing with symmetric extension

In Exercise 3 we implemented a symmetric filter applied to a vector, i.e. when a
periodic extension is assumed. The corresponding function was called filterS(t,
x), and used the function numpy.convolve.
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 102

a) Reimplement the function filterS so that it also takes a third parameter

symm. If symm is false a periodic extension of x should be performed (i.e. filtering
as we have defined it, and as the previous version of filterS performs it). If
symm is true, symmetric extensions should be used (as given by Definition 6.10
in the book).

Solution. The code can look like this:

def filterS(t, x, symm):

tlen = len(t)
N0 = (tlen - 1)/2
N = shape(x)[0]
if symm:
y = concatenate([ x[N0:0:(-1)], x, x[(N-2):(N - N0 - 2):(-1)] ])
else:
y = concatenate([ x[(N - N0):], x, x[:N0]])
if ndim(x) == 1:
res = convolve(t, y)
x[:] = res[(2*N0):(len(res)-2*N0)]
else:
n = shape(x)[1]
for k in range(n):
res = convolve(t, y[:, k])
x[:, k] = res[(2*N0):(len(res)-2*N0)]

b) Implement functions DWTKernelFilters(H0, H1, G0, G1, x, symm, dual)

and IDWTKernelFilters(H0, H1, G0, G1, x, symm, dual) which compute
the DWT and IDWT kernels using theorems 6.10 in the book and 6.5 in the
book, respectively. This function thus bases itself on that the filters of the
wavelet are known. The functions should call the function filterS from a).
Recall also the definition of the parameter dual from this section.

Solution. The code can look like this:

def DWTKernelFilters(H0, H1, G0, G1, x, bd_mode, dual):

f0, f1 = H0, H1
if dual:
f0, f1 = G0, G1
N = len(x)
x0 = x.copy()
x1 = x.copy()
filterS(f0, x0, bd_mode)
filterS(f1, x1, bd_mode)
x[::2] = x0[::2]
x[1::2] = x1[1::2]

def IDWTKernelFilters(H0, H1, G0, G1, x, bd_mode, dual):

f0, f1 = G0, G1
if dual:
f0, f1 = H0, H1
N = len(x)
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 103

x0 = x.copy(); x0[1::2] = 0
x1 = x.copy(); x1[::2] = 0
filterS(f0, x0, bd_mode)
filterS(f1, x1, bd_mode)
x[:] = x0 + x1

With the functions defined in b) you can now define standard DWT and
IDWT kernels in the following way, once the filters are known.

f = lambda x, symm, dual: DWTKernelFilters(H0,H1,G0,G1,x,symm,dual)

invf = lambda x, symm, dual: IDWTKernelFilters(H0,H1,G0,G1,x,symm,dual)

Exercise 6.23: Finding FIR filters

Show that it is impossible to find a non-trivial FIR-filter which satisfies Equation
(6.32) in the book.

Exercise 6.24: The Haar wavelet as an alternative QMF

filter bank
Show that the Haar wavelet satisfies λH1 (ω) = −λH0 (ω + π), and G0 = (H0 )T ,
G1 = (H1 )T . The Haar wavelet can thus be considered as an alternative QMF
filter bank.

Exercise 6.25: Plotting frequency responses

The values Cq , Dq can be found by calling the functions mp3ctable, mp3dtable
which can be found on the book’s webpage.
a) Use your computer to verify the connection we stated between the tables C
and D, i.e. that Di = 32Ci for all i.
b) Plot the frequency responses of the corresponding prototype filters, and verify
that they both are lowpass filters. Use the connection from Theorem (6.24) in
the book to find the prototype filter coefficients from the Cq .

Exercise 6.26: Implementing forward and reverse filter bank

transforms
It is not too difficult to make implementations of the forward and reverse steps
as explained in the MP3 standard. In this exercise we will experiment with this.
In your code you can for simplicity assume that the input and output vectors to
your methods all have lengths which are multiples of 32. Also, use the functions
mp3ctable, mp3dtable mentioned in the previous exercise.
a) Write a function mp3forwardfbt which implements the steps in the forward
direction of the MP3 standard.
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 104

b) Write also a function mp3reversefbt which implements the steps in the

reverse direction.

Solution. The code can look as follows:

def mp3forwardfbt(x):
N = len(x)
z = mat(zeros((N,1)))
C = mp3ctable() # The analysis window;
x = x[(N-1)::(-1)]
x = concatenate([x, zeros(512 - 32)])
# The 32x64 matrix M
yvec = arange(0, 32, 1, float)
yvec = yvec.reshape((32, 1))
xvec = arange(-16, 48, 1, float)
xvec = xvec.reshape((1, 64))
M = cos((2*yvec+1)*xvec*pi/64)
start = len(x) - 512;
for n in range(1, N/32 + 1):
X = x[start:(start + 512)]
Z = C*X # Pointwise multiplication
Y = zeros(64)
for j in range(8):
Y += Z[(64*j):(64*(j+1))]
Y = Y.reshape((64, 1))
z[((n-1)*32):(n*32), 0] = mat(M)*mat(Y)
start -= 32
z = array(z).flatten()
return z

def mp3reversefbt(z):
N = len(z)
z = z. reshape((N,1))
z = mat(z)
Ns = N/32
x = zeros(32*Ns)
D = mp3dtable() # The reconstruction window.
V = mat(zeros((1024,1)))
# The 64x32 matrix N
yvec = arange(16, 80, 1, float)
yvec = yvec.reshape((64, 1))
xvec = arange(0, 32, 1, float)
xvec = xvec.reshape((1, 32))
Nmatr = mat(cos(yvec*(2*xvec + 1)*pi/64))
U = zeros(512)
for n in range(1, Ns+1):
V[64:1024, 0] = V[0:(1024-64), 0]
V[0:64, 0] = Nmatr*z[((n-1)*32):(n*32), 0]
for i in range(8):
U[(i*64):(i*64 + 32)] = \
array(V[(i*128):(i*128 + 32), 0]).flatten()
U[(i*64 + 32):((i + 1)*64)] = \
array(V[(i*128 + 96):((i+1)*128), 0]).flatten()
CHAPTER 6. THE FILTER REPRESENTATION OF WAVELETS 105

W = U*D
for i in range(16):
x[((n-1)*32):(n*32)] += W[32*i:(32*(i + 1))]
return x
Chapter 7

Constructing interesting
wavelets

Exercise 7.1: Implementation of the cascade algorithm

a) In the code above, we turned off symmetric extensions (the symm-argument is
0). Attempt to use symmetric extensions instead, and observe the new plots you
obtain. Can you explain why these new plots do not show the correct functions,
while the previous plots are correct?

Exercise 7.2: Using the cascade algorithm

In Exercise 6 we constructed a new mother wavelet ψ̂ for piecewise linear functions
by finding constants α, β, γ, δ so that

ψ̂ = ψ − αφ0,0 − βφ0,1 − δφ0,2 − γφ0,N −1 .

Use the cascade algorithm to plot ψ̂. Do this by using the wavelet kernel for
the piecewise linear wavelet (do not use the code above, since we have not
implemented kernels for this wavelet yet).

Solution. Assuming that the vector coeffs has been set as in Exercise 6, the
code can look as follows

m = 10
t = linspace(-2, 6, 8*2**m)
coordsvm = hstack([[-coeffs[0], -coeffs[1], -coeffs[3], 0, 0, 0, 0,\
-coeffs[2], 1, 0, 0, 0, 0, 0, 0, 0], \
zeros(8*2**m-16)])
IDWTImpl(coordsvm, m, ’pwl0’, ’per’)
coordsvm *= 2**(m/2.)
plt.plot(t,hstack([coordsvm[(6*2**m):(8*2**m+1)], coordsvm[0:(6*2**m)]]))

106
CHAPTER 7. CONSTRUCTING INTERESTING WAVELETS 107

Exercise 7.3: Compute filters

Compute the filters H0 , G0 in Theorem 7.10 in the book when N = N1 = N2 = 4,
and Q1 = Q(4) , Q2 = 1. Compute also filters H1 , G1 so that we have perfect
reconstruction (note that these are not unique).

Solution. We have that

 N1 /2    2  
1 1 1 1
λH0 (ω) = (1 + cos ω) (1 − cos ω) =
Q1 (1 + cos ω) Q(4) (1 − cos ω)
2 2 2 2
 N2 /2    2
1 1 1
λG0 (ω) = (1 + cos ω) Q2 (1 − cos ω) = (1 + cos ω)
2 2 2
 2
1 1 1 1 2iω
= 1 + eiω + e−iω = (e + 4eiω + 6 + 4e−iω + e−2iω ).
4 2 2 16
1
Therefore G0 = 16 {1, 4, 6, 4, 1}. We do not recommend to compute H0 by hand.
With the package sympy in Python you can do as follows to compute H0 .

x = Symbol(’x’)
z = expand( ((1+x/2+1/(2*x))/2)**2* \
(2+8*((1-x/2-1/(2*x))/2)+20*((1-x/2-1/(2*x))/2)**2\
+40*((1-x/2-1/(2*x))/2)**3) )

Here we have substituted x for eiω , 1/x for e−iω . The first part represents
1

2 (4) 2 3
2 (1 + cos ω) , the second part represents

Q (u) = 2 + 8u + 20u + 40u with
1 1 1 iω 1 −iω
u = 2 (1 − cos ω) = 2 1 − 2 e − 2 e . This gives
1
H0 = {−5, 20, −1, −96, 70, 280, 70, −96, −1, 20, −5}.
128
Using Theorem 6.15 in the book with α = 1, d = 0, we get

1
H1 = {1, −4, 6, −4, 1}
16
1
G1 = {5, 20, 1, −96, −70, 280, −70, −96, 1, 20, 5}
128

Exercise 7.4: Viewing the frequency response

In this exercise we will see how we can view the frequency responses, scaling
functions and mother wavelets for any spline wavelet.
a) Write a function which takes N1 and N2 as input, computes the filter coeffi-
cients of H0 and G0 using equation (7.29) in the book, and plots the frequency
responses of G0 and H0 . Recall that the frequency response can be obtained
from the filter coefficients by taking a DFT. You will have use for the conv
function here, and that the frequency response (1 + cos ω)/2 corresponds to the
filter with coefficients {1/4, 1/2, 1/4}.
CHAPTER 7. CONSTRUCTING INTERESTING WAVELETS 108

b) Recall that in Exercise 6 we implemented DWT and IDWT kernels, which

worked for any set of symmetric filters. Combine these kernels with your computa-
tion of the filter coefficients from a), and plot the corresponding scaling functions
and mother wavelets for different N1 and N2 using the cascade algorithm.

Exercise 7.5: Wavelets based on higher degree polynomials

Show that Br (t) = ∗rk=1 χ[−1/2,1/2) (t) is r − 2 times differentiable, and equals a
polynomial of degree r − 1 on subintervals of the form [n, n + 1]. Explain why
these functions can be used as basis for the spaces Vj of functions which are
piecewise polynomials of degree r − 1 on intervals of the form [n2−m , (n + 1)2−m ],
and r − 2 times differentiable. Br is also called the B-spline of order r.
Chapter 8

The polyphase
representation and wavelets

Exercise 8.1: The frequency responses of the polyphase

components
Let H and G be MRA-matrices for a DWT/IDWT, with corresponding filters
H0 , H1 , G0 , G1 , and polyphase components H (i,j) , G(i,j) .
a) Show that

λH0 (ω) = λH (0,0) (2ω) + eiω λH (0,1) (2ω)

λH1 (ω) = λH (1,1) (2ω) + e−iω λH (1,0) (2ω)
λG0 (ω) = λG(0,0) (2ω) + e−iω λG(1,0) (2ω)
λG1 (ω) = λG(1,1) (2ω) + eiω λG(0,1) (2ω).

Solution. G(0,0) , G(1,1) are the even-indexed filter coefficients of G0 , G1 , re-

spectively, so that λG(0,0) (2ω), λG(1,1) (2ω) represents the half of λG0 (ω), λG1 (ω),
respectively, from the even filter coefficients. G(1,0) are the odd-indexed filter
coefficients of G0 . Since coefficient 0 in G(1,0) equals coefficient 1 in G0 , it is clear
that e−iω λG(1,0) (2ω) represents the half of λG0 (ω) from the odd filter coefficients.
This proves the first formula. The second formula follows from the same kind of
reasoning.
If we transpose H (also in polyphase form), we get an MRA-matrix where the
columns are given by the filters (H0 )T , (H1 )T . Inserting these in the formulas
we just proved we get that

λ(H0 )T (ω) = λ(H (0,0) )T (2ω) + e−iω λ(H (0,1) )T (2ω)

λ(H1 )T (ω) = λ(H (1,1) )T (2ω) + eiω λ(H (1,0) )T (2ω).

109
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS110

If we conjugate these expressions we get

λH0 (ω) = λH (0,0) (2ω) + eiω λH (0,1) (2ω)

λH1 (ω) = λH (1,1) (2ω) + e−iω λH (1,0) (2ω),

and the proof is done.

b) In the proof of the last part of Theorem 6.16 in the book, we defered the
last part, namely that equations 6.16 in the book-(8.3) in the book follow from

G(0,0) G(0,1) αE−d H (1,1) −αE−d H (0,1)

   
= .
G(1,0) G(1,1) −αE−d H (1,0) αE−d H (0,0)

Prove this based on the result from a).

Solution. The first column in the matrix on the left hand side gives the filter
G0 . On the right hand side, a) states that the even-indexed columns are taken
from the filter with frequency response

λαE−d H (1,1) (2ω) + e−iω λ−αE−d H (1,0) (2ω)

= αλE−d (2ω) λH (1,1) (2ω) − e−iω λH (1,0) (2ω)


 
= αe2idω λH (1,1) (2(ω + π)) + e−i(ω+π) λH (1,0) (2(ω + π)) = αe2idω λH1 (ω + π).

This shows that λG0 (ω) = αe2idω λH1 (ω + π). We obtain Equation (8.2) in the
book easily from this. Now, the second column in the matrix on the left hand
side gives the filter coefficients of G1 . On the right hand side, a) states that the
odd-indexed columns are taken from the filter with frequency response

λαE−d H (0,0) (2ω) + eiω λ−αE−d H (0,1) (2ω)

= αλE−d (2ω) λH (0,0) (2ω) − eiω λH (0,1) (2ω)


 
= αe2idω λH (0,0) (2(ω + π)) + ei(ω+π) λH (0,1) (2(ω + π)) = αe2idω λH0 (ω + π).

This shows that λG1 (ω) = αe2idω λH0 (ω + π), which is Equation (8.3) in the
book.

Exercise 8.2: Finding new filters

Let S be a filter. Show that
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS111

a)  
I 0
G
S I
is an MRA matrix with filters G̃0 , G1 , where

λG˜0 (ω) = λG0 (ω) + λS (2ω)e−iω λG1 (ω),

Solution. We have that

 (0,0)
G(0,1)
 (0,0)
+ SG(0,1) G(0,1)
  
G I 0 G
= .
G(1,0) G(1,1) S I G (1,0)
+ SG(1,1) G(1,1)
Using Exercise 8a), the even-indexed columns in this matrix are taken from the
filter with frequency response

λG(0,0) +SG(0,1) (2ω) + e−iω λG(1,0) +SG(1,1) (2ω)

= λG(0,0) (2ω) + e−iω λG(1,0) (2ω) + λS (2ω) λG(0,1) (2ω) + e−iω λG(1,1) (2ω)

= λG0 (ω) + λS (2ω)e−iω λG(1,1) (2ω) + eiω λG(0,1) (2ω)

= λG0 (ω) + λS (2ω)e−iω λG1 (ω).

b)  
I S
G
0 I
is an MRA matrix with filters G0 , G̃1 , where

λG˜1 (ω) = λG1 (ω) + λS (2ω)eiω λG0 (ω),

Solution. We have that

 (0,0)
G(0,1)
 (0,0)
SG(0,0) + G(0,1)
  
G I S G
= ,
G(1,0) G(1,1) 0 I G(1,0) SG(1,0) + G(1,1)
so that the odd-indexed columns in this matrix are taken from the filter with
frequency response

λSG(1,0) +G(1,1) (2ω) + eiω λSG(0,0) +G(0,1) (2ω)

= λG(1,1) (2ω) + eiω λG(0,1) (2ω) + λS (2ω) λG(1,0) (2ω) + eiω λG(0,0) (2ω)

= λG1 (ω) + λS (2ω)eiω λG(0,0) (2ω) + e−iω λG(1,0) (2ω)

= λG1 (ω) + λS (2ω)eiω λG0 (ω).

CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS112

c)  
I 0
H
S I
is an MRA-matrix with filters H0 , H̃1 , where

λH̃1 (ω) = λH1 (ω) + λS (2ω)e−iω λH0 (ω).

Solution. We transpose the expression to obtain

I ST
 
HT
0 I
Since H T has filters (H0 )T and (H1 )T in the columns, from b) it follows that

I ST
 
HT
0 I
has columns given by (H0 )T and the filter with frequency response

λ(H1 )T (ω) + λS T (2ω)eiω λ(H0 )T (ω) = λH1 (ω) + λS (2ω)e−iω λH0 (ω),

so that
 
I 0
H
S I
has row filters H0 and a filter H̃1 with frequency response

λH̃1 (ω) = λH1 (ω) + λS (2ω)e−iω λH0 (ω).

d)  
I S
H
0 I
is an MRA-matrix with filters H̃0 , H1 , where

λH̃0 (ω) = λH0 (ω) + λS (2ω)eiω λH1 (ω).

Solution. We transpose the expression to obtain

 
I 0
HT .
ST I
Since H T has filters (H0 )T and (H1 )T in the columns, using a) we see that
 
I 0
HT
ST I
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS113

has columns given by the filter with frequency response

λ(H0 )T (ω) + λS T (2ω)e−iω λ(H1 )T (ω) = λH0 (ω) + λS (2ω)eiω λH1 (ω),

and (H1 )T , so that

 
I S
H
0 I
has a row filter H̃0 with frequency response

λH̃0 (ω) = λH0 (ω) + λS (2ω)eiω λH1 (ω),

and H1 .
In summary, this exercise shows that one can think of the steps in the lifting
factorization as altering one of the filters of an MRA-matrix in alternating order.

Exercise 8.3: Relating to the polyphase components

−1
Show that S is a filter of length kM if and only if the entries {S i,j }M
i,j=0 in the
polyphase representation of S satisfy S (i+r) mod M,(j+r) mod M = Si,j . In other
words, S is a filter if and only if the polyphase representation of S is a “block-
circulant Toeplitz matrix”. This implies a fact that we will use: GH is a filter
(and thus provides alias cancellation) if blocks in the polyphase representations
repeat cyclically as in a Toeplitz matrix (in particular when the matrix is
block-diagonal with the same block repeating on the diagonal).

Solution. If S is a filter we have that S (i+r+s1 M ) mod kM,(j+r+s2 M )kM =

S (i+s1 M ) mod M,(j+s2 M ) mod M , 0 ≤ i, j < M . But since S (i+s1 M ) mod kM,(j+s2 M )kM =
(i,j) ((i+r) mod M,(j+r) mod M ) (i,j)
Ss1 ,s2 , it follows that Ss1 ,s2 = Ss1 ,s2 , so that S (i+r) mod M,(j+r) mod M =
Si,j .

Exercise 8.4: QMF filter banks

Recall from Definition 6.18 in the book that we defined a classical QMF filter
bank as one where M = 2, G0 = H0 , G1 = H1 , and λH1 (ω) = λH0 (ω + π). Show
that the forward and reverse filter bank transforms of a classical QMF filter
bank take the form

 
A −B
H=G=
B A
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS114

Exercise 8.5: Alternative QMF filter banks

Recall from Definition 6.19 in the book that we defined an alternative QMF filter
bank as one where M = 2, G0 = (H0 )T , G1 = (H1 )T , and λH1 (ω) = λH0 (ω + π).
Show that the forward and reverse filter bank transforms of an alternative QMF
filter bank take the form.

T
AT BT −B T AT BT
    
A
H= G= = .
−B A B AT −B A

Exercise 8.6: Alternative QMF filter banks with additional

sign
Consider alternative QMF filter banks where we take in an additional sign, so
that λH1 (ω) = −λH0 (ω + π) (the Haar wavelet was an example of such a filter
bank). Show that the forward and reverse filter bank transforms now take the
form

 T T
BT BT AT BT
   
A A
H= G= = .
B −A B −AT B −A

It is straightforward to check that also these satisfy the alias cancellation con-
dition, and that the perfect reconstruction condition also here takes the form
|λH0 (ω)|2 + |λH0 (ω + π)|2 = 2.

Exercise 8.8: Polyphase components for symmetric filters

Assume that the filters H0 , H1 of a wavelet are symmetric, and denote by S (i,j)
the polyphase components of the corresponding MRA-matrix H. Show that
S (0,0) and S (1,1) are symmetric filters, that the filter coefficients of S (1,0) has
symmetry about −1/2, and that S (0,1) has symmetry about 1/2. Also show a
similar statement for the MRA-matrix G of the inverse DWT.

Exercise 8.9: Implementing kernels transformations using

lifting
Up to now in this chapter we have obtained lifting factorizations for four different
wavelets where the filters are symmetric. Let us now implement the kernel
transformations for these wavelets. Your functions should call the functions
from Exercise 5 and Exercise 5. in order to compute the individual lifting steps.
Recall that the kernel transformations should take the input vector x, symm
(i.e. whether symmetric extension should be applied), and dual (i.e. whether the
dual wavelet transform should be applied) as input. You will need equations 6.19
in the book-(8.12) in the book here, in order to complete the kernels for bot the
transformations and the dual transformations.
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS115

a) Write the DWT and IDWT kernel transformations for the Spline 5/3 wavelet,
using the lifting factorization obtained in Example 8.2.2 in the book. Call your
functions dwt_kernel_53 and idwt_kernel_53.

Solution. The code can look like this:

def dwt_kernel_53(x, bd_mode):
x[0::2] *= 2
x[1::2] *= 0.5
lifting_odd_symm(-0.125, x, bd_mode)
lifting_even_symm(1, x, bd_mode)

def idwt_kernel_53(x, bd_mode):

lifting_even_symm(-1, x, bd_mode)
lifting_odd_symm(0.125, x, bd_mode)
x[0::2] *= 0.5
x[1::2] *= 2

b) Write the DWT and IDWT kernel transformations for the CDF 9/7 wavelet,
using the lifting factorization obtained in Example 8.2.3 in the book. Call your
functions DWTKernel97 and IDWTKernel97.

Solution. The code can look like this:

def dwt_kernel_97(x, bd_mode):
lambda1=-0.586134342059950
lambda2=-0.668067171029734
lambda3=0.070018009414994
lambda4=1.200171016244178
alpha=-1.149604398860250
beta=-0.869864451624777
x[0::2] *= alpha
x[1::2] *= beta
lifting_odd_symm(-lambda4, x, bd_mode)
lifting_even_symm(-lambda3, x, bd_mode)
lifting_odd_symm(-lambda2, x, bd_mode)
lifting_even_symm(-lambda1, x, bd_mode)

def idwt_kernel_97(x, bd_mode):

lambda1=-0.586134342059950
lambda2=-0.668067171029734
lambda3=0.070018009414994
lambda4=1.200171016244178
alpha=-1.149604398860250
beta=-0.869864451624777
lifting_even_symm(lambda1, x, bd_mode)
lifting_odd_symm(lambda2, x, bd_mode)
lifting_even_symm(lambda3, x, bd_mode)
lifting_odd_symm(lambda4, x, bd_mode)
x[0::2] /= alpha
x[1::2] /= beta
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS116

c) In Chapter 5 in the book, we listened to the low-resolution approximations

and detail components in sound for three different wavelets. Repeat these
experiments with the Spline 5/3 and the CDF 9/7 wavelet, using the new kernels
we have implemented in this exercise.

Solution. The following code can be used for listening to the low-resolution
approximations for a given value of m.

m = 1
x, fs = forw_comp_rev_DWT(m, ’cdf53’)
play(x, fs)

m = 1
x, fs = forw_comp_rev_DWT(m, ’cdf97’)
play(x, fs)

d) Plot all scaling functions and mother wavelets for the Spline 5/3 and the CDF
9/7 wavelets, using the cascade algorithm and the kernels you have implemented.

Solution. The code can look as follows.

plotwaveletfunctions(IDWTKernel53, -4, 4)
plotwaveletfunctions(IDWTKernel97, -4, 4)

In the plot for the CDF 9/7 wavelet, it is seen that the functions and their dual
counterparts are close to being equal. This reflects the fact that this wavelet is
close to being orthogonal.

Exercise 8.10: Lifting orthonormal wavelets

In this exercise we will implement the kernel transformations for orthonormal
wavelets.
a) Write functions lifting_even and lifting_odd which take λ1 , λ2 and a
vector x as input, and apply the elementary lifting matrices (8.16) in the book,
respectively, to x. Assume that N is even.

Solution. The code can look like this:

def lifting_even(lmbda1, lmbda2, x):

if mod(len(x), 2)!=0:
raise AssertionError()
x[0] += lmbda1*x[1] + lmbda2*x[-1]
x[2:-1:2] += lmbda1*x[3::2] + lmbda2*x[1:-2:2]
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS117

def lifting_odd(lmbda1, lmbda2, x):

if mod(len(x), 2)!=0:
raise AssertionError()
x[1:-2:2] += lmbda1*x[2:-1:2] + lmbda2*x[0:-3:2]
x[-1] += lmbda1*x[0] + lmbda2*x[-2]

b) Write functions dwt_kernel_ortho and idwt_kernel_ortho which take a

vector x as input, and apply the DWT and IDWT kernel transformations for
orthonormal wavelets to x. You should call the functions lifting_even and
lifting_odd. As mentioned, assume that global variables lambdas, alpha, and
beta have been set, so that the lifting factorization (8.8) in the book holds,
where lambdas is a n × 2-matrix so that the filter coefficients {λ1 , λ2 } or {λ1 , λ2 }
in the i’th lifting step is found in row i of lambdas. Recall that the last lifting
step was even.

Solution. The code can look like this:

def dwt_kernel_ortho( x, filters, bd_mode):

x[0::2] *= filters[’alpha’]
x[1::2] *= filters[’beta’]
for stepnr in range(filters[’lambdas’].shape[0] - 1, 0, -2):
lifting_even(-filters[’lambdas’][stepnr, 0], -filters[’lambdas’][stepnr, 1], x)
lifting_odd(-filters[’lambdas’][stepnr - 1, 0], -filters[’lambdas’][stepnr - 1, 1], x)
if mod(filters[’lambdas’].shape[0], 2)==1:
lifting_even(-filters[’lambdas’][0, 0], -filters[’lambdas’][0, 1], x)

def idwt_kernel_ortho( x, filters, bd_mode):

stepnr = 0
if mod(filters[’lambdas’].shape[0],2) == 1: # Start with an even step
lifting_even(filters[’lambdas’][stepnr, 0], filters[’lambdas’][stepnr, 1], x)
stepnr += 1
while stepnr < filters[’lambdas’].shape[0]:
lifting_odd(filters[’lambdas’][stepnr, 0], filters[’lambdas’][stepnr, 1], x)
lifting_even(filters[’lambdas’][stepnr + 1, 0], filters[’lambdas’][stepnr + 1, 1], x)
stepnr += 2
x[0::2] /= filters[’alpha’]
x[1::2] /=filters[’beta’]

c) Listen to the low-resolution approximations and detail components in sound

for orthonormal wavelets for N = 1, 2, 3, 4. You need to call the function
liftingfactortho in order to set the kernel for the different values of N .

Solution. The following code can be used for listening to the low-resolution
approximations for a given value of m.

x, fs = forw_comp_rev_DWT(m, ’db2’)
play(x, fs)
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS118

x, fs = forw_comp_rev_DWT(m, ’db3’)
play(x, fs)

x, fs = forw_comp_rev_DWT(m, ’db4’)
play(x, fs)

d) Plot all scaling functions and mother wavelets for the orthonormal wavelets for
N = 1, 2, 3, 4, using the cascade algorithm. Since the wavelets are orthonormal,
we should have that φ = φ̃, and ψ = ψ̃. In other words, you should see that the
bottom plots equal the upper plots.

Solution. The code can look as follows.

liftingfactortho(2)
plotwaveletfunctions(IDWTKernelOrtho, -4, 4)
liftingfactortho(3)
plotwaveletfunctions(IDWTKernelOrtho, -4, 4)
liftingfactortho(4)
plotwaveletfunctions(IDWTKernelOrtho, -4, 4)

Exercise 8.11: 4 vanishing moments

In Exercise 5 we found constants α, β, γ, δ which give the coordinates of ψ̂ in
(φ1 , ψ̂1 ), where ψ̂ had four vanishing moments, and where we worked with the
multiresolution analysis of piecewise constant functions.
a) Show that the polyphase representation of G when ψ̂ is used as mother
wavelet can be factored as
  
1 I 0 I {−γ, −α, −β, −δ}
√ . (8.1)
2 {1/2, 1/2} I 0 I
You here need to reconstruct what you did in the lifting factorization for the
alternative piecewise linear wavelet, i.e. write

PD1 ←(φ1 ,ψ̂1 ) = PD1 ←(φ1 ,ψ1 ) P(φ1 ,ψ1 )←(φ1 ,ψ̂1 ) .
By inversion, find also a lifting factorization of H.

Solution. We have found constants α, βγδ so that

[ψ̂](φ0 ,ψ0 ) = (−α, −β, −δ, 0, 0, 0, 0, −γ) ⊕ (1, 0, 0, 0, 0, 0, 0, 0),

From this it is clear that
 
I S2
P(φ1 ,ψ1 )←(φ1 ,ψ̂1 ) =
0 I
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS119

where S2 = {−γ, −α, −β, −δ} This gives as before the lifting factorization
  
1 I 0 I {−γ, −α, −β, −δ}
PD1 ←(φ1 ,ψ̂1 ) = √ . (8.2)
2 S1 I 0 I
where S1 = {1/2, 1/2} as before.

Exercise 8.12: Wavelet based on piecewise quadratic scaling

function
In Exercise 7 you should have found the filters

1
H0 = {−5, 20, −1, −96, 70, 280, 70, −96, −1, 20, −5}
128
1
H1 = {1, −4, 6, −4, 1}
16
1
G0 = {1, 4, 6, 4, 1}
16
1
G1 = {5, 20, 1, −96, −70, 280, −70, −96, 1, 20, 5}.
128
a) Show that

1
I − 14 {1, 1}
    1 
I − 128 {5, −29, −29, 5} I 0 0
G= 4 .
0 I −{1, 1} I 0 I 0 4
From this we can easily derive the lifting factorization of G.

Solution. The polyphase factorization of the IDWT is

1 1

16 {1, 6, 1} 128 {5, 1, −70, −70, 1, 5} .
1 1
16 {4, 4} 128 {20, −96, 280, −96, 20}
We can first apply an even lifting step:

 1 1 1 1
I − 14 {1, 1}
   
16 {1, 6, 1} 128 {5, 1, −70, −70, 1, 5} = 16 {4} 128 {20, −116, −116, 20} .
1 1 1 1
0 I 16 {4, 4} 128 {20, −96, 280, −96, 20} 16 {4, 4} 128 {20, −96, 280, −96, 20}

We can now apply an odd lifting step

1 1
   1 1

I 0 16 {4} 128 {20, −116, −116, 20} = 4 128 {20, −116, −116, 20}
1 1
−{1, 1} I 16 {4, 4} 128 {20, −96, 280, −96, 20} 0 4
Since

1
  1 1
 1 
I − 512 {20, −116, −116, 20} 4 128 {20, −116, −116, 20} 4 0
= ,
0 I 0 4 0 4
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS120

it follows that

1
I − 14 {1, 1}
    1 
I − 128 {5, −29, −29, 5} I 0 0
G= 4 .
0 I −{1, 1} I 0 I 0 4

b) Implement the kernels of the wavelet of this exercise using what you did in
Exercise 6.

H0 = array([-5, 20, -1, -96, 70, 280, 70, -96, -1, 20, -5])/128.
H1 = array([1, -4, 6, -4, 1])/16.
G0 = array([1, 4, 6, 4, 1])/16.
G1 = array([5, 20, 1, -96, -70, 280, -70, -96, 1, 20, 5])/128.
f =lambda x, symm, dual: DWTKernelFilters(H0,H1,G0,G1,x,symm,dual)
invf=lambda x, symm, dual: IDWTKernelFilters(H0,H1,G0,G1,x,symm,dual)

Solution.
c) Listen to the low-resolution approximations and detail components in sound
for this wavelet.

Solution. The following code can be used for listening to the low-resolution
approximations for a given value of m.

playDWT(m, f, invf, True)

d) Plot all scaling functions and mother wavelets for this wavelet, using the
cascade algorithm.

Solution. The code can look as follows.

plotwaveletfunctions(invf, -4, 4)

Exercise 8.13: Run forward and reverse transform

Run the forward and then the reverse transform from Exercise 6 on the vector
(1, 2, 3, . . . , 8192). Verify that there seems to be a delay on 481 elements, as
promised by Therorem 8.14 in the book. Do you get the exact same result back?

Solution. The following code can be used:

x = arange(1,8193)
mp3reversefbt(mp3forwardfbt(x))
plt.plot(x)

There are some small errors from the original vector in the resulting vector, when
one compensates for the delay of 481 elements.
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS121

Exercise 8.14: Verify statement of filters

Use your computer to verify the symmetries we have stated for the symmetries
in the prototype filters, i.e. that
(
−C512−i i 6= 64, 128, . . . , 448
Ci =
C512−i i = 64, 128, . . . , 448.
Explain also that this implies that hi = h512−i for i = 1, . . . , 511. In other words,
the prototype filter has symmetry around (511 + 1)/2 = 256, so that it has linear
phase.

Exercise 8.15: Lifting

We mentioned that we could use the lifting factorization to construct filters on
the form stated in Equation (8.20) in the book, so that the matrices on the form
given by Equation (8.24) in the book, i.e.
 (32−i)
V (i)

V
,
−V (64−i) V (32+i)
are invertible. Let us see what kind of lifting steps produce such matrices.
a) Show that the lifting steps
   
I λE2 I 0
and
0 I λI I
applied in alternating order to a matrix on the form given by Equation (8.24) in
the book, where the filters are on the from given by Equation (8.20) in the book,
again produces matrices and filters on these forms. This explains how we can
parametrize a larger number of such matrices with the help of lifting steps.It
also explain why the inverse matrix is on the form stated in Equation (8.24) in
the book with filters on the same form, since the inverse lifting steps are of the
same type.
b) Explain that 16 numbers {λi }16 i=1 are needed (together with what we start
with on the diagonal in the lifting construction), in order to construct filters so
that the prototype filter has 512 coefficients. Since there are 15 submatrices,
this gives 240 optimization variables.
Lifting gives the following strategy for finding a corresponding synthesis
prototype filter which gives perfect reconstruction: First compute matrices V, W
which are inverses of oneanother using lifting (using the lifting steps of this
exercise ensures that all filters will be on the form stated in Equation (8.20) in
the book), and write
CHAPTER 8. THE POLYPHASE REPRESENTATION AND WAVELETS122

T
V (1) V (2) (W (1) )T (W (2) )T
  (1)
−W (3)
  (1)
V (2)
 
W V
VW = =
−V (3) V (4) W (2) W (4) −V (3) V (4) −(W (3) )T (W (4) )T
T
V (1) V (2) E15 (W (1) )T E15 (W (2) )T
     
E15 0
= = I.
−V (3) V (4) −E15 (W (3) )T E15 (W (4) )T 0 E15

Now, the matrices U (i) = E15 (W (i) )T are on the form stated in Equation (8.20)
in the book, and we have that
 (1)
V (2)
  (1)
U (2)
  
V U E−15 0
=
−V (3) V (4) −U (3) U (4) 0 E−15
We can now conclude from Theorem 8.13 in the book that if we define the
synthesis prototype filter as therein, and set c = 1, d = −15, we have that
GH = 16E481−32·15 = 16E1 .
Chapter 9

Digital images

Exercise 9.6: Generate black and white images

Black and white images can be generated from greyscale images (with values
between 0 and 255) by replacing each pixel value with the one of 0 and 255
which is closest. Use this strategy to generate the black and white image shown
in the right part of Figure 9.2 in the book.

Solution. The following code can be used

X2 = 255*(X1 >= 128)

imwhow(uint8(X2))

Exercise 9.7: Adjust contrast in images

a) Write a function contrastadjust0 which instead uses the function from
Equation (9.1) in the book to increase the contrast. n should be a parameter to
the function.

Solution. The code could look as follows:

def contrastadjust0(X,n):
"""
Assumes that the values are in [0,255]
"""
X /= 255.
X -= 1/2.
X *= n
arctan(X, X)
X /= (2*arctan(n/2.))
X += 1/2.0
X *= 255 # Maps the values back to [0,255]

b) Generate the left and right images in Figure 9.9 in the book on your own by
writing code which uses the two functions contrastadjust0 and contrastadjust.

123
CHAPTER 9. DIGITAL IMAGES 124

Solution. The following code can be used.

X1 = img.copy()
contrastadjust0(X1, 10)
imshow(uint8(X1))

X2 = img.copy()
contrastadjust(X2, 0.01)
imshow(uint8(X2))

Exercise 9.8: Adjust contrast with another function

In this exercise we will look at another function for increasing the contrast of a
picture.
a) Show that the function f : R → R given by

fn (x) = xn ,

for all n maps the interval [0, 1] → [0, 1], and that f 0 (1) → ∞ as n → ∞.
b) The color image secret.jpg,shown in Figure 9.1, contains some information
that is nearly invisible to the naked eye on most computer monitors. Use the
function f (x), to reveal the secret message.

Figure 9.1: Secret message.

Hint. You will first need to convert the image to a greyscale image. You can
then use the function contrastadjust as a starting point for your own program.
CHAPTER 9. DIGITAL IMAGES 125

Solution. The following code can be used

n = 100
X = double(imread(’images/secret.jpg’, ’jpg’))
X = (X[:,:,0] + X[:,:,1]+ X[:,:,2])/3.
mapto01(X)
X **= n
X *= 255
X = uint8(X)
imshow(X)

The resulting images is shown in Figure 9.2.

Figure 9.2: Secret message revealed!

Exercise 9.13: Implement a tensor product

Implement a function tensor_impl which takes a matrix X, and functions S1
and S2 as parameters, and applies S1 to the columns of X, and S2 to the rows of
X.

Solution. The following code can be used:

def tensor_impl(X, S1, S2):

M, N = shape(X)[0:2]
for n in range(N):
S1(X[:,n])
for m in range(M):
S2(X[m, :])
CHAPTER 9. DIGITAL IMAGES 126

Exercise 9.14: Generate images

Write code which calls the function tensor_impl with appropriate filters and
which generate the following images:
a) The right image in Figure 9.12 in the book.

Solution. The following code can be used

X3 = excerpt.copy()
def longmolecule(x):
filterS(S2, x, True)
tensor_impl(X3, longmolecule, longmolecule)
imshow(uint8(X3))

b) The right image in Figure 9.14 in the book.

Solution. The following code can be used

X3 = excerpt.copy()
tensor_impl(X3, donothing, diffxmolecule)
mapto01(X3)
X3 *= 255
contrastadjust0(X3,50)
imshow(uint8(X3))

c) The images in Figure 9.16 in the book.

Solution. The following code can be used

excerpt=createexcerpt()
X1 = excerpt.copy()
tensor_impl(X1, donothing, diffxmolecule)
X2 = excerpt.copy()
tensor_impl(X2, diffymolecule, donothing)
X1 = X1**2 + X2**2
imshow(uint8(X1))

X2[:] = X1
mapto01(X2)
X2 *= 255
imshow(uint8(X2))

X3[:] = X2
contrastadjust0(X3, 50)
imshow(uint8(X3))

d) The images in Figure 9.15 in the book.

CHAPTER 9. DIGITAL IMAGES 127

Solution. The following code can be used

excerpt=createexcerpt()
X1 = excerpt.copy()
tensor_impl(X1, donothing, diffxmolecule)
mapto01(X1)
X1 *= 255
contrastadjust0(X1, 50)
imshow(uint8(X1))

X2 = excerpt.copy()
tensor_impl(X2, diffymolecule, donothing)
mapto01(X2)
X2 *= 255
contrastadjust0(X2, 50)
imshow(uint8(X2))

e) The images in Figure 9.17 in the book.

Solution. The following code can be used

excerpt=createexcerpt()
X1 = excerpt.copy()
tensor_impl(X1, donothing, diffxmolecule)
tensor_impl(X1, donothing, diffxmolecule)
mapto01(X1)
X1 *= 255
contrastadjust0(X1, 100)
imshow(uint8(X1))

X2 = excerpt.copy()
tensor_impl(X2, donothing, diffxmolecule)
tensor_impl(X2, diffymolecule, donothing)
mapto01(X2)
X2 *= 255
contrastadjust0(X2, 100)
imshow(uint8(X2))

X3 = excerpt.copy()
tensor_impl(X3, diffymolecule, donothing)
tensor_impl(X3, diffymolecule, donothing)
mapto01(X3)
X3 *= 255
contrastadjust0(X3, 100)
imshow(uint8(X3))

Exercise 9.15: Interpret tensor products

Let the filter S be defined by S = {−1, 1}.
a) Let X be a matrix which represents the pixel values in an image. What can
you say about how the new images (S ⊗ I)X og (I ⊗ S)X look? What are the
interpretations of these operations?
CHAPTER 9. DIGITAL IMAGES 128

b) Write down the 4 ⊗ 4-matrix X = (1, 1, 1, 1) ⊗ (0, 0, 1, 1). Compute (S ⊗ I)X

by applying the filters to the corresponding rows/columns of X as we have
learned, and interpret the result. Do the same for (I ⊗ S)X.

Exercise 9.16: Computational molecule of moving average

filter
1
Let S be the moving average filter of length 2L+1, i.e. T = L {1,
|
··· , 1, 1, 1, · · · , 1}.
{z }
2L+1 times
What is the computational molecule of S ⊗ S?

Exercise 9.17: Bilinearity of the tensor product

Show that the mapping F (x, y) = x ⊗ y is bi-linear, i.e. that F (αx1 + βx2 , y) =
αF (x1 , y) + βF (x2 , y), and F (x, αy1 + βy2 ) = αF (x, y1 ) + βF (x, y2 ).

Solution. We have that

F (αx1 + βx2 , y) = (αx1 + βx2 ) ⊗ y = (αx1 + βx2 )y T

= αx1 y T + βx2 y T = α(x1 ⊗ y) + β(x ⊗ y)
= αF (x1 , y) + βF (x1 , y).

The second statement follows similarly.

Exercise 9.18: Attempt to write as tensor product

Attempt to find matrices S1 : RM → RM and S2 : RN → RN so that the
following mappings from LM,N (R) to LM,N (R) can be written on the form
X → S1 X(S2 )T = (S1 ⊗ S2 )X. In all the cases, it may be that no such S1 , S2
can be found. If this is the case, prove it.
a) The mapping which reverses the order of the rows in a matrix.

Solution. Multiplicaton with the matrix

 
0 0 0 ··· 0 0 1
0 0 0 · · · 0 1 0
 
S =  ... ... ... .. .. .. .. 

 . . . .
0 1 0 · · · 0 0 0
1 0 0 ··· 0 0 0
reverses the elements in a vector. This means that

((S ⊗ I)(x ⊗ y))i,j = ((Sx) ⊗ y)i,j = (Sx)i yj = xM −1−i yj = (x ⊗ y)M −1−i,j .

CHAPTER 9. DIGITAL IMAGES 129

This means that also ((S ⊗ I)X)i,j = XM −1−i,j for all X, so that S ⊗ I reverses
rows, and thus is a solution to a).
b) The mapping which reverses the order of the columns in a matrix.

Solution. Similarly one shows that I ⊗ S reverses columns, and is thus a

solution to b).
c) The mapping which transposes a matrix.

Solution. It turns out that it is impossible to find S1 and S2 so that transposing

a matrix X corresponds to computing (S1 ⊗ S2 )X. To see why, S1 and S2 would
need to fulfill

(S1 ⊗ S2 )(ei ⊗ ej ) = (S1 ei ) ⊗ (S2 ej ) = ej ⊗ ei ,

since ej ⊗ ei is the transpose of ei ⊗ ej . This would require that S1 ei = ej for
all i, j, which is impossible.

Exercise 9.19: Computational molecules

Let the filter S be defined by S = {1, 2, 1}.
a) Write down the computational molecule of S ⊗ S.

Solution. The computational molecule of S ⊗ S is

   
1
1 2 1
(1, 2, 1) ⊗ (1, 2, 1) = (1, 2, 1) ⊗ (1, 2, 1) = 2 1 2 1 = 2 4 2 .
1 1 2 1

b) Let us define x = (1, 2, 3), y = (3, 2, 1), z = (2, 2, 2), and w = (1, 4, 2).
Compute the matrix A = x ⊗ y + z ⊗ w.

Solution. We get that

   
1
2

A = 2 3 2 1 + 2 1 4 2
3 2
     
3 2 1 2 8 4 5 10 5
= 6 4 2 + 2 8 4 =  8 12 6 .
9 6 3 2 8 4 11 14 7

c) Compute (S ⊗ S)A by applying the filter S to every row and column in the
matrix the way we have learned. If the matrix A was more generally an image,
what can you say about how the new image will look?
CHAPTER 9. DIGITAL IMAGES 130

Solution. We need to compute (S ⊗ S)A = SAS T , which corresponds to first

applying S to every column in the image, and then applying S to every row in
the resulting image.  S to every column in the image we first get
 If we apply
29&46&23
the matrix SA = 32&48&24. If we apply the filter to the rows here we get
35&50&25
 
127&144&121
SAS T = 136&152&128. Since the filter which is applied is a low-pass filter,
145&160&135
the new image should look a bit more smooth than the original image.

Exercise 9.20: Computational molecules 2

Let S = 14 {1, 2, 1} be a filter.
a) What is the effect of applying the tensor products S ⊗ I, I ⊗ S, and S ⊗ S
on an image represented by the matrix X?

Solution. Note forst that the filter is a smoothing filter (a low-pass filter). We
know that S ⊗ I corresponds to applying S to the columns of the matrix, so that
we get the result by applying the smoothing filter to the columns of the matrix.
The result of this is that horizontal edges are smoothed. Similarly, the tensor
product I ⊗ S corresponds to applying S to the rows of the matrix, so that
vertical edges are smoothed. Finally, S ⊗ S corresponds to applying S first to the
columns of the matrix, then to the rows. The result is that both horizontal and
vertical edges are smoothed. You could also have computed the computational
molecules for S ⊗ I, I ⊗ S, and S ⊗ S, by taking the tensor product of the filter
coefficients 14 {1, 2, 1} with itself. From these molecules it is also clear that they
either work on the columns, the rows, or on both rows and columns.
b) Compute (S ⊗ S)(x ⊗ y), where x = (4, 8, 8, 4), y = (8, 4, 8, 4) (i.e. both x
and y are column vectors).

Solution. A 4 × 4 circulant Toeplitz matrix for S is

 
2 1 0 1
1
 1 2 1 0 .

4 0 1 2 1

1 0 1 2
From this we can quickly compute that
CHAPTER 9. DIGITAL IMAGES 131

      
2 1 0 1 4 2+2+1 5
1 1 2 1 0 8 4 + 2 + 1 7
Sx =    =  = 
4 0 1 2 1 8 4 + 2 + 1 7
1 0 1 2 4 2+2+1 5
      
2 1 0 1 8 4+1+1 6
1 1 2 1 0 4 2 + 2 + 2 6
Sy =    =   =  .
4 0 1 2 1 8 4 + 1 + 1 6
1 0 1 2 4 2+2+2 6
From this it is clear that

   
5 30 30 30 30
7
42 42 42 42
(S ⊗ S)(x ⊗ y) = (Sx)(Sy)T = 

7 6 6 6 6 =
 .
42 42 42 42
5 30 30 30 30

Exercise 9.21: Comment on code

Suppose that we have an image given by the M × N -matrix X, and consider the
following code:

for n in range(N):
X[0, n] = 0.25*X[N-1, n] + 0.5*X[0, n] + 0.25*X[1, n]
X[1:(N-1), n] = 0.25*X[0:(N-2), n] + 0.5*X[1:(N-1), n] \
+ 0.25*X[2:N, n]
X[N-1, n] = 0.25*X[N-2, n] + 0.5*X[N-1, n] + 0.25*X[0, n]
for m in range(m):
X[m, 0] = 0.25*X[m, M-1] + 0.5*X[m, 0] + 0.25*X[m, 1]
X[m, 1:(M-1)] = 0.25*X[m, 0:(M-2)] + 0.5*X[m, 1:(M-1)] \
+ 0.25*X[m, 2:M]
X[m, M-1] = 0.25*X[m, M-2] + 0.5*X[m, M-1] + 0.25*X[m, 0]

Which tensor product is applied to the image? Comment what the code does, in
particular the first and third line in the inner for-loop. What effect does the
code have on the image?

Solution. In the code the filter S = {1/4, 1/2, 1/4} is applied to the clumns
and the rows in the image. We have learned that this corresponds to applying
the tensor product S ⊗ S to the image. k=1 in the outer for-loop corresponds
to applying S on the columns, k=2 corresponds to applying S on the rows. The
first and last lines in the inner for-loop are necessary since we apply S to the
periodic extension of the image. Since S is a smoothing filter, the effect will be
that the image is smoothed vertically and horizontally.

Exercise 9.22: Eigenvectors of tensor products

Let vA be an eigenvector of A with eigenvalue λA , and vB an eigenvector of
B with eigenvalue λB . Show that vA ⊗ vB is an eigenvector of A ⊗ B with
CHAPTER 9. DIGITAL IMAGES 132

eigenvalue λA λB . Explain from this why kA ⊗ Bk = kAkkBk, where k · k denotes

the operator norm of a matrix.

Exercise 9.23: The Kronecker product

The Kronecker tensor product of two matrices A and B, written A ⊗k B, is
defined as
 
a1,1 B a1,2 B · · · a1,M B
a2,1 B a2,2 B · · · a2,M B 
A ⊗k B =  . ..  ,
 
.. ..
 .. . . . 
ap,1 B ap,2 B · · · ap,M B
where the entries of A are ai,j . The tensor product of a p × M -matrix, and a
q × N -matrix is thus a (pq) × (M N )-matrix. Note that this tensor product in
particular gives meaning for vectors: if x ∈ RM , y ∈ RN are column vectors,
then x ⊗k y ∈ RM N is also a column vector. In this exercise we will investigate
how the Kronecker tensor product is related to tensor products as we have
defined them in this section.
a) Explain that, if x ∈ RM , y ∈ RN are column vectors, then x ⊗k y is the
column vector where the rows of x ⊗ y have first been stacked into one large
row vector, and this vector transposed. The linear extension of the operation
defined by

x ⊗ y ∈ RM,N → x ⊗k y ∈ RM N
thus stacks the rows of the input matrix into one large row vector, and transposes
the result.
b) Show that (A ⊗k B)(x ⊗k y) = (Ax) ⊗k (By). We can thus use any of
the defined tensor products ⊗, ⊗k to produce the same result, i.e. we have the
commutative diagram shown in Figure 9.3, where the vertical arrows represent
stacking the rows in the matrix, and transposing, and the horizontal arrows
represent the two tensor product linear transformations we have defined. In
particular, we can compute the tensor product in terms of vectors, or in terms
of matrices, and it is clear that the Kronecker tensor product gives the matrix
of tensor product operations.

A⊗B
x⊗y / (Ax) ⊗ (By)

 
A⊗k B
x ⊗k y / (Ax) ⊗k (By),

Figure 9.3: Tensor products

CHAPTER 9. DIGITAL IMAGES 133

Solution. We have that

(A ⊗k B)(x ⊗k y)
    
a11 B a12 B · · · a1M B x1 y (a11 x1 + . . . + a1M xm )By
a21 B a22 B · · · a2M B   x2 y  (a21 x1 + . . . + a2M xm )By 
= . ..   ..  = 
    
.. .. ..
 ..

. . .   .   . 
ap1 B ap2 B · · · apM B xM y (ap1 x1 + . . . + apM xm )By
 
(Ax)1 By
(Ax)2 By 
=  = (Ax) ⊗k (By).
 
..
 . 
(Ax)p By

c) Using the Euclidean inner product on L(M, N ) = RM N , i.e.

M
X −1 N
X −1
hX, Y i = Xi,j Yi,j .
i=0 j=0

and the correspondence in a) we can define the inner product of x1 ⊗ y1 and

x2 ⊗ y2 by

hx1 ⊗ y1 , x2 ⊗ y2 i = hx1 ⊗k y1 , x2 ⊗k y2 i.
Show that

hx1 ⊗ y1 , x2 ⊗ y2 i = hx1 , x2 ihy1 , y2 i.

Clearly this extends linearly to an inner product on LM,N .

Solution. We have that

   
* (x1 )0 y1 (x2 )0 y2 + M −1
hx1 ⊗ y1 , x2 ⊗ y2 i = .. .
..
X
(x1 )i (x2 )i hy1 , y2 i
,  =
   
 .
(x1 )M −1 y1 (x2 )M −1 y2 i=0

M
X −1
= hy1 , y2 i (x1 )i (x2 )i = hx1 , x2 ihy1 , y2 i.
i=0

d) Show that the FFT factorization can be written as

   
FN/2 DN/2 FN/2 I DN/2
= N/2 (I2 ⊗k FN/2 ).
FN/2 −DN/2 FN/2 IN/2 −DN/2
Also rewrite the sparse matrix factorization for the FFT from Equation (2.18)
in the book in terms of tensor products.
CHAPTER 9. DIGITAL IMAGES 134

Exercise 9.26: Implement DFT and DCT on blocks

In this section we have used functions which apply the DCT and the DFT either
to subblocks of size 8 × 8, or to the full image. Implement functions DFTImpl8,
IDFTImpl8, DCTImpl8, and IDCTImpl8 which applies the DFT, IDFT, DCT, and
IDCT, to consecutive segments of length 8.

Solution. The following code can be used.

def DFTImpl8(x):
N = shape(x)[0]
for n in range(0,N,8):
x[n:(n+8), :] = fft.fft(x[n:(n+8), :], axis=0)
def IDFTImpl8(x):
N = shape(x)[0]
for n in range(0, N, 8):
x[n:(n+8), :] = fft.ifft(x[n:(n+8), :], axis=0)

def DCTImpl8(x):
N = shape(x)[0]
for n in range(0,N,8):
x[n:(n+8), :] = dct(x[n:(n+8), :], norm=’ortho’, axis=0)
def IDCTImpl8(x):
N = shape(x)[0]
for n in range(0, N, 8):
x[n:(n+8), :] = idct(x[n:(n+8), :], norm=’ortho’, axis=0)

Exercise 9.27: Implement two-dimensional FFT and DCT

Implement functions DFTImplFull, IDFTImplFull, DCTImplFull, and IDCTImplFull
which applies the DFT, IDFT, DCT, and IDCT, to the entire vector, and use
these to implement. FFT2, IFFT2, DCT2, and IDCT2 on an image, with the
help of the function tensor_impl.

Solution. The following code can be used.

def DFTImplFull(x):
x[:] = fft.fft(x, axis=0)
def IDFTImplFull(x):
x[:] = fft.ifft(x, axis=0)

def DCTImplFull(x):
x[:] = dct(x, norm=’ortho’, axis=0)
def IDCTImplFull(x):
x[:] = idct(x, norm=’ortho’, axis=0)
CHAPTER 9. DIGITAL IMAGES 135

tensor_impl(X, DFTImplFull, DFTImplFull)

tensor_impl(X, IDFTImplFull, IDFTImplFull)
tensor_impl(X, DCTImplFull, DCTImplFull)
tensor_impl(X, IDCTImplFull, IDCTImplFull)

Exercise 9.28: Zeroing out DCT coefficients

The function forw_comp_rev_DCT2 in the module forw_comp_rev applies the
DCT to a part of our sample image, and sets DCT coefficients below a certain
threshold to be zero. This is very similar to what the JPEG standard does.
Run forw_comp_rev_DCT2 for different threshold parameters, and with the
functions DCTImpl8, IDCTImpl8, DCTImplFull, and IDCTImplFull as parame-
ters. Check that this reproduces the DCT test images of this section, and that
the correct numbers of values which have been neglected (i.e. which are below
the threshold) are printed on screen.

Solution. Once

X1 = forw_comp_rev_DCT2(DCTImpl8, IDCTImpl8, 30)

imshow(uint8(X1))

X2 = forw_comp_rev_DCT2(DCTImpl8, IDCTImpl8, 50)

imshow(uint8(X2))

X3 = forw_comp_rev_DCT2(DCTImpl8, IDCTImpl8, 100)

imshow(uint8(X3))

X1 = forw_comp_rev_DCT2(DCTImplFull, IDCTImplFull, 30)

imshow(uint8(X1))

X2 = forw_comp_rev_DCT2(DCTImplFull, IDCTImplFull, 50)

imshow(uint8(X2))

X3 = forw_comp_rev_DCT2(DCTImplFull, IDCTImplFull, 100)

imshow(uint8(X3))

Exercise 9.29: Comment code

Suppose that we have given an image by the matrix X. Consider the following
code:
CHAPTER 9. DIGITAL IMAGES 136

threshold = 30
[M, N] = shape(X)[0:2]
for n in range(N):
FFTImpl(X[:, n], FFTKernelStandard)
for m in range(M):
FFTImpl(X[m, :], FFTKernelStandard)
X = X.*(abs(X) >= threshold)
for n in range(N):
FFTImpl(X[:, n], FFTKernelStandard, 0)
for m in range(M):
FFTImpl((X[m, :], FFTKernelStandard, 0)

Comment what the code does. Comment in particular on the meaning of the
parameter threshold, and what effect this has on the image.

Solution. In the first part of the code one makes a change of coordinates with
the DFT. More precisely, this is a change of coordinates on a tensor product,
as we have defined it. In the last part the change of coordinates is performed
the opposite way. Both these change of coordinates is performed is performed
the way we have described them, first on the rows in the matrix, then on the
columns. The parameter threshold is used to neglect DFT-coefficients which
are below a certain value. We have seen that this can give various visual artefacts
in the image, even though the main contents of the image still may be visible. If
we increase threshold, these artefacts will be more dominating since we then
neglect many DFT-coefficients.
Chapter 10

Using tensor products to

apply wavelets to images

Exercise 10.7: Implement two-dimensional DWT

Implement functions DW2TImpl and IDW2TImpl which perform the m-level DWT2
and the IDWT2, respectively, on an image. The functions should take the
same input as DWTImpl and IDWTImpl, with the input vector replaced with a
two-dimensional object. The functions should at each stage call DWTImpl and
IDWTImpl with m = 1, and each call to these functions should alter the appropri-
ate upper left submatrix in the coordinate matrix. If the image has several color
components, the functions should be applied to each color component. There
are three color components in the test image ’lena.png’.

Solution. The following code can be used:

def DWT2Impl_internal(X, nres, f, bd_mode):

M, N = shape(X)[0:2]
for res in range(nres):
for n in range(0,N,2**res):
f(X[0::2**res, n], bd_mode)
for m in range(0,M,2**res):
f(X[m, 0::2**res], bd_mode)
reorganize_coeffs2_forward(X, nres)

def IDWT2Impl_internal(X, nres, f, bd_mode):

reorganize_coeffs2_reverse(X, nres)
M, N = shape(X)[0:2]
for res in range(nres - 1, -1, -1):
for n in range(0, N, 2**res):
f(X[0::2**res, n], bd_mode)
for m in range(0, M, 2**res):
f(X[m, 0::2**res], bd_mode)

137
CHAPTER 10. USING TENSOR PRODUCTS TO APPLY WAVELETS TO IMAGES138

Exercise 10.8: Comment code

Assume that we have an image represented by the M × N -matrix X, and consider
the following code:

for n in range(N):
c = (X[0:M:2, n] + X[1:M:2, n])/sqrt(2)
w = (X[0:M:2, n] - X[1:M:2, n])/sqrt(2)
X[:, n] = concatenate([c, w])
for m in range(M):
c = (X[m, 0:N:2] + X[m, 1:N:2])/sqrt(2)
w = (X[m, 0:N:2] - X[m, 1:N:2])/sqrt(2)
X[m, :] = concatenate([c,w])

a) Comment what the code does, and explain what you will see if you display X
as an image after the code has run.

Solution. The code runs a DWT over one level, and the Haar wavelet is
used. Inside the for-loops the DWT is applied to every row and column in the
image. k=1 i the for-loop corresponds to applying the DWT to the columns,
k=2 corresponds to applying the DWT to the rows. In the upper left corner we
will see a low-resolution version of the image. In the other three corners you will
see different tyoes of detail: In the upper right corner you will see detail which
corresponds to quick vertical changes, in the lower left corner you will see detail
which corresponds to quick horizontal changes, and in the lower right corner
you will see points where quick changes both vertically and horizontally occur
simultaneously.
b) The code above has an inverse transformation, which reproduce the original
image from the transformed values which we obtained. Assume that you zero
out the values in the lower left and the upper right corner of the matrix X after
the code above has run, and that you then reproduce the image by applying this
inverse transformation. What changes can you then expect in the image?

Solution. By zeroing out the two corners you remove detail which correpond
to quick horizontal and vertical changes. But since we keep the lower right
corner, we keep detail which corresponds to simultaneous changes vertically
and horizontally. The result after the inverse transformation is that most edges
have been smoothed, but we see no smoothing effect in points where quick
changes occur both horizontally and vertically. In Example 10.3 in the book,
this corresponds to that we emphasize the gridpoints in the chess pattern, mut
that we smooth out the horizontal and vertical edges in the chess pattern.

Exercise 10.9: Comment code

In this exercise we will use the filters G0 = {1, 1}, G1 = {1, −1}.
CHAPTER 10. USING TENSOR PRODUCTS TO APPLY WAVELETS TO IMAGES139

a) Let X be a matrix which represents the pixel values in an image. Define

x = (1, 0, 1, 0) and y = (0, 1, 0, 1). Compute (G0 ⊗ G0 )(x ⊗ y).
b) For a general image X, describe how the images (G0 ⊗ G0 )X, (G0 ⊗ G1 )X,
(G1 ⊗ G0 )X, and (G1 ⊗ G1 )X may look.
c) Assume that we run the following code on an image represented by the matrix
X:
M, N = shape(X)
for n in range(N):
c = X[0:M:2, n] + X[1:M:2, n]
w = X[0:M:2, n] - X[1:M:2, n]
X[:, n] = vstack([c,w])
for m in range(M):
c = X[m, 0:N:2] + X[m, 1:N:2]
w = X[m, 0:N:2] - X[m, 1:N:2]
X[m, :] = hstack([c,w])

Comment the code. Describe what will be shown in the upper left corner of
X after the code has run. Do the same for the lower left corner of the matrix.
What is the connection with the images (G0 ⊗ G0 )X, (G0 ⊗ G1 )X, (G1 ⊗ G0 )X,
and (G1 ⊗ G1 )X?

Exercise 10.10: Experiments on a test image

Figure 10.1: A simple image before and after one level of the DWT2. The Haar
wavelet was used.
In Figure 10.1 we have applied the DWT2 with the Haar wavelet to an image
very similar to the one you see in Figure 10.6 in the book. You see here, however,
that there seems to be no detail components, which is very different from what
you saw in Example 10.3 in the book, even though the images are very similar.
Attempt to explain what causes this to happen.
CHAPTER 10. USING TENSOR PRODUCTS TO APPLY WAVELETS TO IMAGES140

Hint. Compare with Exercise 5.

Solution. In Example 10.3 in the book, the borders in the chess pattern was
chosen so that they occur at odd numbers. This means that the image can not
be represented exactly in Vm−1 ⊗ Vm−1 , so that there is detail present in the
image at all the borders in the chess pattern. In the new image, the borders in
the chess pattern was chosen so that they occur at even numbers. This means
that the image can be represented exactly in Vm−1 ⊗ Vm−1 , so that there is no
detail components present in the image.

Exercise 10.11: Implement the fingerprint compression scheme

Write code which generates the images shown in figures 10.3 in the book, 10.3 in
the book, and 10.23 in the book. Use the functions DW2TImpl and IDW2TImpl
with the CDF 9/7 wavelet kernel functions as input.

Solution. The following code generates the different parts of Figure 10.3 in
the book.

img=double(imread(’images/fingerprint_500x500.png’))
imgnew=255*ones((512,512,3))
imgnew[0:500,0:500,:]=img
img=imgnew
l1, l2, l3 = 512, 512, 3
X0 = img.copy()
DWT2Impl(X0, 1, ’cdf97’)
X1 = X0.copy()
mapto01(X0)
X0 *= 255
imshow(uint8(X0))

DWT2Impl(X1[:(l1/2), :(l2/2), :], 1, ’cdf97’)

DWT2Impl(X1[:(l1/2), (l2/2):, :], 1, ’cdf97’)
DWT2Impl(X1[(l1/2):, :(l2/2), :],1,’cdf97’)
DWT2Impl(X1[(l1/2):, (l2/2):512, :],1,’cdf97’)
X2 = X1.copy()
mapto01(X1)
X1 *= 255
imshow(uint8(X1))

DWT2Impl(X2[:(l1/4), :(l2/4), :],1,’cdf97’)

DWT2Impl(X2[:(l1/4), (l2/4):(l2/2), :],1,’cdf97’)
DWT2Impl(X2[(l1/4):(l1/2), :(l2/4), :],1,’cdf97’)
X3 = X2.copy()
mapto01(X2)
X2 *= 255
imshow(uint8(X2))
CHAPTER 10. USING TENSOR PRODUCTS TO APPLY WAVELETS TO IMAGES141

DWT2Impl(X3[:64, :64, :], 1, ’cdf97’)

DWT2Impl(X3[:64, 64:128, :], 1, ’cdf97’)
DWT2Impl(X3[:64, 128:192, :], 1, ’cdf97’)
DWT2Impl(X3[:64, 192:256, :], 1, ’cdf97’)
DWT2Impl(X3[64:128, :64, :], 1, ’cdf97’)
DWT2Impl(X3[64:128, 64:128, :], 1, ’cdf97’)
DWT2Impl(X3[64:128, 128:192, :], 1, ’cdf97’)
DWT2Impl(X3[64:128, 192:256, :], 1, ’cdf97’)
DWT2Impl(X3[128:192, :64, :], 1, ’cdf97’)
DWT2Impl(X3[128:192, 64:128, :], 1, ’cdf97’)
DWT2Impl(X3[192:256, :64, :],1,’cdf97’)
DWT2Impl(X3[192:256, 64:128, :],1,’cdf97’)
Z = X3.copy()
mapto01(X3)
X3 *= 255
imshow(uint8(X3))

The following code generates Figure 10.3 in the book.

DWT2Impl(Z[:32,:32,:],1,’cdf97’)
Y0 = Z.copy()
mapto01(Z)
Z *= 255
Z = uint8(Z)
imshow(Z)

The following code generates the parts of Figure 10.23 in the book.

Y0[16:, :, :] = 0
Y0[:16, 16:, :] = 0
IDWT2Impl(Y0, 5, ’cdf97’)
Y1 = Y0.copy()
mapto01(Y0)
Y0 *= 255
imshow(uint8(Y0))

Y1 = img.copy()
DWT2Impl(Y1, 5, ’cdf97’)
Y1[:16, :16, :] = 0
IDWT2Impl(Y1, 5, ’cdf97’)
mapto01(Y1)
Y1 *= 255
imshow(uint8(Y1))
Appendix A

Basic Linear Algebra

142
Appendix B

Signal processing and linear

algebra: a translation guide

143