Digital Signal Processing

Markus Kuhn

Computer Laboratory

http://www.cl.cam.ac.uk/teaching/0809/DSP/

Lent 2009 – Part II

Signals
→ flow of information

→ measured quantity that varies with time (or position)

→ electrical signal received from a transducer

(microphone, thermometer, accelerometer, antenna, etc.)

→ electrical signal that controls a process

Continuous-time signals: voltage, current, temperature, speed, . . .

Discrete-time signals: daily minimum/maximum temperature,
lap intervals in races, sampled continuous signals, . . .

Electronics (unlike optics) can only deal easily with time-dependent signals, therefore spatial
signals, such as images, are typically first converted into a time signal with a scanning process
(TV, fax, etc.).

2
 Signal processing
Signals may have to be transformed in order to
→ amplify or filter out embedded information
→ detect patterns
→ prepare the signal to survive a transmission channel
→ prevent interference with other signals sharing a medium
→ undo distortions contributed by a transmission channel
→ compensate for sensor deficiencies
→ find information encoded in a different domain
To do so, we also need
→ methods to measure, characterise, model and simulate trans-
mission channels
→ mathematical tools that split common channels and transfor-
mations into easily manipulated building blocks
3

Analog electronics

Passive networks (resistors, capacitors, R

inductances, crystals, SAW filters),
non-linear elements (diodes, . . . ),
Uin L C Uout
(roughly) linear operational amplifiers
Advantages:

• passive networks are highly linear

Uin
over a very large dynamic range
Uin
and large bandwidths
Uout

• analog signal-processing circuits Uout

√
require little or no power 0 1/ LC ω (= 2πf) t

• analog circuits cause little addi- t

Uin − Uout 1 dUout
Z
tional interference = Uout dτ + C
R L −∞ dt

4
 Digital signal processing
Analog/digital and digital/analog converter, CPU, DSP, ASIC, FPGA.
Advantages:
→ noise is easy to control after initial quantization
→ highly linear (within limited dynamic range)
→ complex algorithms fit into a single chip
→ flexibility, parameters can easily be varied in software
→ digital processing is insensitive to component tolerances, aging,
environmental conditions, electromagnetic interference
But:
→ discrete-time processing artifacts (aliasing)
→ can require significantly more power (battery, cooling)
→ digital clock and switching cause interference
5

Typical DSP applications

→ communication systems → astronomy
modulation/demodulation, channel VLBI, speckle interferometry
equalization, echo cancellation

→ consumer electronics → experimental physics

perceptual coding of audio and video sensor-data evaluation
on DVDs, speech synthesis, speech
recognition
→ aviation
→ music
radar, radio navigation

synthetic instruments, audio effects,

noise reduction → security
steganography, digital watermarking,
→ medical diagnostics biometric identification, surveillance
systems, signals intelligence, elec-
magnetic-resonance and ultrasonic
tronic warfare
imaging, computer tomography,
ECG, EEG, MEG, AED, audiology
→ engineering
→ geophysics control systems, feature extraction
seismology, oil exploration for pattern recognition
6
 Syllabus
Signals and systems. Discrete sequences and systems, their types and proper-
ties. Linear time-invariant systems, convolution. Harmonic phasors are the eigen
functions of linear time-invariant systems. Review of complex arithmetic. Some
examples from electronics, optics and acoustics.
MATLAB. Use of MATLAB on PWF machines to perform numerical experiments
and visualise the results in homework exercises.
Fourier transform. Harmonic phasors as orthogonal base functions. Forms of the
Fourier transform, convolution theorem, Dirac’s delta function, impulse combs in
the time and frequency domain.
Discrete sequences and spectra. Periodic sampling of continuous signals, pe-
riodic signals, aliasing, sampling and reconstruction of low-pass and band-pass
signals, spectral inversion.
Discrete Fourier transform. Continuous versus discrete Fourier transform, sym-
metry, linearity, review of the FFT, real-valued FFT.
Spectral estimation. Leakage and scalloping phenomena, windowing, zero padding.

Finite and infinite impulse-response filters. Properties of filters, implementa-

tion forms, window-based FIR design, use of frequency-inversion to obtain high-
pass filters, use of modulation to obtain band-pass filters, FFT-based convolution,
polynomial representation, z-transform, zeros and poles, use of analog IIR design
techniques (Butterworth, Chebyshev I/II, elliptic filters).
Random sequences and noise. Random variables, stationary processes, autocor-
relation, crosscorrelation, deterministic crosscorrelation sequences, filtered random
sequences, white noise, exponential averaging.
Correlation coding. Random vectors, dependence versus correlation, covariance,
decorrelation, matrix diagonalisation, eigen decomposition, Karhunen-Loève trans-
form, principal/independent component analysis. Relation to orthogonal transform
coding using fixed basis vectors, such as DCT.
Lossy versus lossless compression. What information is discarded by human
senses and can be eliminated by encoders? Perceptual scales, masking, spatial
resolution, colour coordinates, some demonstration experiments.
Quantization, image and audio coding standards. A/µ-law coding, delta cod-
ing, JPEG photographic still-image compression, motion compensation, MPEG
video encoding, MPEG audio encoding.
Note: The last three lectures on audio-visual coding were previously part of the course “Informa-
tion Theory and Coding”. A brief introduction to MATLAB was given in “Unix Tools”.

8
 Objectives
By the end of the course, you should be able to
→ apply basic properties of time-invariant linear systems
→ understand sampling, aliasing, convolution, filtering, the pitfalls of
spectral estimation
→ explain the above in time and frequency domain representations
→ use filter-design software
→ visualise and discuss digital filters in the z-domain
→ use the FFT for convolution, deconvolution, filtering
→ implement, apply and evaluate simple DSP applications in MATLAB
→ apply transforms that reduce correlation between several signal sources
→ understand and explain limits in human perception that are ex-
ploited by lossy compression techniques
→ provide a good overview of the principles and characteristics of sev-
eral widely-used compression techniques and standards for audio-
visual signals
9

Textbooks
→ R.G. Lyons: Understanding digital signal processing. Prentice-
Hall, 2004. (£45)
→ A.V. Oppenheim, R.W. Schafer: Discrete-time signal process-
ing. 2nd ed., Prentice-Hall, 1999. (£47)
→ J. Stein: Digital signal processing – a computer science per-
spective. Wiley, 2000. (£74)
→ S.W. Smith: Digital signal processing – a practical guide for
engineers and scientists. Newness, 2003. (£40)
→ K. Steiglitz: A digital signal processing primer – with appli-
cations to digital audio and computer music. Addison-Wesley,
1996. (£40)
→ Sanjit K. Mitra: Digital signal processing – a computer-based
approach. McGraw-Hill, 2002. (£38)
10
 Sequences and systems
A discrete sequence {xn }∞
n=−∞ is a sequence of numbers

. . . , x−2 , x−1 , x0 , x1 , x2 , . . .

where xn denotes the n-th number in the sequence (n ∈ Z). A discrete

sequence maps integer numbers onto real (or complex) numbers.
We normally abbreviate {xn }∞ n=−∞ to {xn }, or to {xn }n if the running index is not obvious.
The notation is not well standardized. Some authors write x[n] instead of xn , others x(n).

Where a discrete sequence {xn } samples a continuous function x(t) as

xn = x(ts · n) = x(n/fs ),

we call ts the sampling period and fs = 1/ts the sampling frequency.

A discrete system T receives as input a sequence {xn } and transforms
it into an output sequence {yn } = T {xn }:
discrete
. . . , x2 , x1 , x0 , x−1 , . . . . . . , y2 , y1 , y0 , y−1 , . . .
system T

11

Properties of sequences
A sequence {xn } is
∞
X
absolutely summable ⇔ |xn | < ∞
n=−∞
X∞
square summable ⇔ |xn |2 < ∞
n=−∞
periodic ⇔ ∃k > 0 : ∀n ∈ Z : xn = xn+k
A square-summable sequence is also called an energy signal, and
∞
X
|xn |2
n=−∞

is its energy. This terminology reflects that if U is a voltageR supplied to a load

resistor R, then P = U I = U 2 /R is the power consumed, and P (t) dt the energy.
So even where we drop physical units (e.g., volts) for simplicity in calculations, it
is still customary to refer to the squared values of a sequence as power and to its
sum or integral over time as energy.
12
A non-square-summable sequence is a power signal if its average power
k
1 X
lim |xn |2
k→∞ 1 + 2k
n=−k

exists.

Special sequences
Unit-step sequence: 
0, n < 0
un =
1, n ≥ 0
Impulse sequence:

1, n = 0
δn =
0, n 6= 0
= un − un−1

13

Types of discrete systems

A causal system cannot look into the future:
yn = f (xn , xn−1 , xn−2 , . . .)
A memory-less system depends only on the current input value:
yn = f (xn )
A delay system shifts a sequence in time:
yn = xn−d
T is a time-invariant system if for any d
{yn } = T {xn } ⇐⇒ {yn−d } = T {xn−d }.
T is a linear system if for any pair of sequences {xn } and {x′n }
T {a · xn + b · x′n } = a · T {xn } + b · T {x′n }.
14
Examples:
The accumulator system
n
X
yn = xk
k=−∞

is a causal, linear, time-invariant system with memory, as are the back-

ward difference system
yn = xn − xn−1 ,
the M-point moving average system
M −1
1 X xn−M +1 + · · · + xn−1 + xn
yn = xn−k =
M k=0 M
and the exponential averaging system
∞
X
yn = α · xn + (1 − α) · yn−1 = α (1 − α)k · xn−k .
k=0

15

Examples for time-invariant non-linear memory-less systems:

yn = x2n , yn = log2 xn , yn = max{min{⌊256xn ⌋, 255}, 0}

Examples for linear but not time-invariant systems:


xn , n ≥ 0
yn = = xn · un
0, n < 0
yn = x⌊n/4⌋
yn = xn · ℜ(eωjn )

Examples for linear time-invariant non-causal systems:

1
yn = (xn−1 + xn+1 )
2
9
X sin(πkω)
yn = xn+k · · [0.5 + 0.5 · cos(πk/10)]
k=−9
πkω

16
 Constant-coefficient difference equations
Of particular practical interest are causal linear time-invariant systems
of the form
xn b0 yn
N
X
yn = b0 · xn − ak · yn−k −a1
z −1
k=1 yn−1

z −1
−a2
Block diagram representation yn−2
of sequence operations:
x′n z −1
−a3
xn xn + x′n yn−3
Addition:
Multiplication xn a axn
by constant:
The ak and bm are
constant coefficients.
xn xn−1
Delay: z −1
17

or
xn xn−1 xn−2 xn−3
M z −1 z −1 z −1
X
yn = bm · xn−m b0 b1 b2 b3
m=0

yn

or the combination of both: xn b0 a−1 yn

0

z −1 z −1
b1 −a1
N
X M
X xn−1 yn−1
ak · yn−k = bm · xn−m
k=0 m=0 z −1 z −1
b2 −a2
xn−2 yn−2

z −1 z −1
b3 −a3
xn−3 yn−3
The MATLAB function filter is an efficient implementation of the last variant.
18
 Convolution
All linear time-invariant (LTI) systems can be represented in the form
∞
X
yn = ak · xn−k
k=−∞

where {ak } is a suitably chosen sequence of coefficients.

This operation over sequences is called convolution and defined as
∞
X
{pn } ∗ {qn } = {rn } ⇐⇒ ∀n ∈ Z : rn = pk · qn−k .
k=−∞

If {yn } = {an } ∗ {xn } is a representation of an LTI system T , with

{yn } = T {xn }, then we call the sequence {an } the impulse response
of T , because {an } = T {δn }.

19

Convolution examples
A B C D

E F A∗B A∗C

C∗A A∗E D∗E A∗F

20
 Properties of convolution
For arbitrary sequences {pn }, {qn }, {rn } and scalars a, b:
→ Convolution is associative
({pn } ∗ {qn }) ∗ {rn } = {pn } ∗ ({qn } ∗ {rn })
→ Convolution is commutative
{pn } ∗ {qn } = {qn } ∗ {pn }
→ Convolution is linear
{pn } ∗ {a · qn + b · rn } = a · ({pn } ∗ {qn }) + b · ({pn } ∗ {rn })
→ The impulse sequence (slide 13) is neutral under convolution
{pn } ∗ {δn } = {δn } ∗ {pn } = {pn }
→ Sequence shifting is equivalent to convolving with a shifted
impulse
{pn−d } = {pn } ∗ {δn−d }
21

Can all LTI systems be represented by convolution?

Any sequence {xn } can be decomposed into a weighted sum of shifted
impulse sequences:
∞
X
{xn } = xk · {δn−k }
k=−∞

Let’s see what happens if we apply a linear(∗) time-invariant(∗∗) system

T to such a decomposed sequence:
∞ ∞
!
X (∗) X
T {xn } = T xk · {δn−k } = xk · T {δn−k }
k=−∞ k=−∞
∞ ∞
!
(∗∗) X X
= xk · {δn−k } ∗ T {δn } = xk · {δn−k } ∗ T {δn }
k=−∞ k=−∞
= {xn } ∗ T {δn } q.e.d.

⇒ The impulse response T {δn } fully characterizes an LTI system.

22
Exercise 1 What type of discrete system (linear/non-linear, time-invariant/
non-time-invariant, causal/non-causal, causal, memory-less, etc.) is:

(a) yn = |xn | 3xn−1 + xn−2

(e) yn =
xn−3
(b) yn = −xn−1 + 2xn − xn+1 (f) yn = xn · en/14
8
(c) yn =
Y
xn−i (g) yn = xn · un
∞
i=0 X
(h) yn = xi · δi−n+2
(d) yn = 21 (x2n + x2n+1 ) i=−∞

Exercise 2
Prove that convolution is (a) commutative and (b) associative.

23

Exercise 3 A finite-length sequence is non-zero only at a finite number of

positions. If m and n are the first and last non-zero positions, respectively,
then we call n − m + 1 the length of that sequence. What maximum length
can the result of convolving two sequences of length k and l have?

Exercise 4 The length-3 sequence a0 = −3, a1 = 2, a2 = 1 is convolved

with a second sequence {bn } of length 5.
(a) Write down this linear operation as a matrix multiplication involving a
matrix A, a vector ~b ∈ R5 , and a result vector ~c.
(b) Use MATLAB to multiply your matrix by the vector ~b = (1, 0, 0, 2, 2)
and compare the result with that of using the conv function.
(c) Use the MATLAB facilities for solving systems of linear equations to
undo the above convolution step.

Exercise 5 (a) Find a pair of sequences {an } and {bn }, where each one
contains at least three different values and where the convolution {an }∗{bn }
results in an all-zero sequence.
(b) Does every LTI system T have an inverse LTI system T −1 such that
{xn } = T −1 T {xn } for all sequences {xn }? Why?
24
 Direct form I and II implementations
xn b0 a−1
0 yn xn a−1
0 b0 yn

z −1 z −1 z −1
b1 −a1 −a1 b1
xn−1 yn−1

z −1 z −1
= z −1
b2 −a2 −a2 b2
xn−2 yn−2

z −1 z −1 z −1
b3 −a3 −a3 b3
xn−3 yn−3

The block diagram representation of the constant-coefficient difference

equation on slide 18 is called the direct form I implementation.
The number of delay elements can be halved by using the commuta-
tivity of convolution to swap the two feedback loops, leading to the
direct form II implementation of the same LTI system.
These two forms are only equivalent with ideal arithmetic (no rounding errors and range limits).
25

Convolution: optics example

If a projective lens is out of focus, the blurred image is equal to the
original image convolved with the aperture shape (e.g., a filled circle):

∗ =

as
Point-spread function h (disk, r = 2f
): image plane focal plane

1
x2 + y 2 ≤ r 2

r2 π
,
h(x, y) =
0, x2 + y 2 > r 2
a
Original image I, blurred image B = I ∗ h, i.e.
ZZ
B(x, y) = I(x − x′ , y − y ′ ) · h(x′ , y ′ ) · dx′ dy ′ s

f
26
 Convolution: electronics example
R Uin
Uin U
√in
2

Uout
Uin C Uout
Uout

0
t 0 1/RC ω (= 2πf)

Any passive network (R, L, C) convolves its input voltage Uin with an
impulse response function h, leading to Uout = Uin ∗ h, that is
Z ∞
Uout (t) = Uin (t − τ ) · h(τ ) · dτ
−∞

In this example:
 −t
Uin − Uout dUout 1
· e RC , t ≥ 0
=C· , h(t) = RC
R dt 0, t<0

27

Why are sine waves useful?

1) Adding together sine waves of equal frequency, but arbitrary ampli-
tude and phase, results in another sine wave of the same frequency:
A1 · sin(ωt + ϕ1 ) + A2 · sin(ωt + ϕ2 ) = A · sin(ωt + ϕ)
with
q
A = A21 + A22 + 2A1 A2 cos(ϕ2 − ϕ1 )

A1 sin ϕ1 + A2 sin ϕ2
tan ϕ =
A1 cos ϕ1 + A2 cos ϕ2 A
A2 A2 · sin(ϕ2 )

ϕ2
Sine waves of any phase can be
A1 A2 · cos(ϕ2 )
formed from sin and cos alone:
ωt ϕ A1 · sin(ϕ1 )
ϕ1
A · sin(ωt + ϕ) = A1 · cos(ϕ1 )

a · sin(ωt) + b · cos(ωt)
√
with a = A · cos(ϕ), b = A · sin(ϕ) and A = a2 + b2 , tan ϕ = ab .
28
Note: Convolution of a discrete sequence {xn } with another sequence
{yn } is nothing but adding together scaled and delayed copies of {xn }.
(Think of {yn } decomposed into a sum of impulses.)
If {xn } is a sampled sine wave of frequency f , so is {xn } ∗ {yn }!
=⇒ Sine-wave sequences form a family of discrete sequences
that is closed under convolution with arbitrary sequences.
The same applies for continuous sine waves and convolution.

2) Sine waves are orthogonal to each other:

Z ∞
sin(ω1 t + ϕ1 ) · sin(ω2 t + ϕ2 ) dt “=” 0
−∞

⇐⇒ ω1 6= ω2 ∨ ϕ1 − ϕ2 = (2k + 1)π/2 (k ∈ Z)
They can be used to form an orthogonal function basis for a transform.
The term “orthogonal” is used here in the context of an (infinitely dimensional) vector space,
where the “vectors” Rare functions of the form f : R → R (or f : R → C) and the scalar product
∞
is defined as f · g = −∞ f (t) · g(t) dt.
29

Why are exponential functions useful?

Adding together two exponential functions with the same base z, but
different scale factor and offset, results in another exponential function
with the same base:
A1 · z t+ϕ1 + A2 · z t+ϕ2 = A1 · z t · z ϕ1 + A2 · z t · z ϕ2
= (A1 · z ϕ1 + A2 · z ϕ2 ) · z t = A · z t

Likewise, if we convolve a sequence {xn } of values

. . . , z −3 , z −2 , z −1 , 1, z, z 2 , z 3 , . . .
xn = z n with an arbitrary sequence {hn }, we get {yn } = {z n } ∗ {hn },
∞
X ∞
X ∞
X
yn = xn−k · hk = z n−k · hk = z n · z −k · hk = z n · H(z)
k=−∞ k=−∞ k=−∞

where H(z) is independent of n.

Exponential sequences are closed under convolution with
arbitrary sequences. The same applies in the continuous case.
30
 Why are complex numbers so useful?
1) They give us all n solutions√(“roots”) of equations involving poly-
nomials up to degree n (the “ −1 = j ” story).
2) They give us the “great unifying theory” that combines sine and
exponential functions:
1 jωt
e + e− jωt


cos(ωt) =
2
1
e jωt − e− jωt


sin(ωt) =
2j
or
1 jωt+ϕ − jωt−ϕ


cos(ωt + ϕ) = e +e
2
or
cos(ωn + ϕ) = ℜ(e jωn+ϕ ) = ℜ[(e jω )n · e jϕ ]
sin(ωn + ϕ) = ℑ(e jωn+ϕ ) = ℑ[(e jω )n · e jϕ ]
Notation: ℜ(a + jb) := a and ℑ(a + jb) := b where j2 = −1 and a, b ∈ R.
31

We can now represent sine waves as projections of a rotating complex

vector. This allows us to represent sine-wave sequences as exponential
sequences with basis e jω .
A phase shift in such a sequence corresponds to a rotation of a complex
vector.
3) Complex multiplication allows us to modify the amplitude and phase
of a complex rotating vector using a single operation and value.
Rotation of a 2D vector in (x, y)-form is notationally slightly messy,
but fortunately j2 = −1 does exactly what is required here:
     
x3 x2 −y2 x1
= · (x3 , y3 )
y3 y2 x2 y1
 
x1 x2 − y1 y2 (−y2 , x2 )
=
x1 y2 + x2 y1 (x2 , y2 )
(x1 , y1 )
z1 = x1 + jy1 , z2 = x2 + jy2
z1 · z2 = x1 x2 − y1 y2 + j(x1 y2 + x2 y1 )

32
 Complex phasors
Amplitude and phase are two distinct characteristics of a sine function
that are inconvenient to keep separate notationally.
Complex functions (and discrete sequences) of the form
A · e j(ωt+ϕ) = A · [cos(ωt + ϕ) + j · sin(ωt + ϕ)]
(where j2 = −1) are able to represent both amplitude and phase in
one single algebraic object.
Thanks to complex multiplication, we can also incorporate in one single
factor both a multiplicative change of amplitude and an additive change
of phase of such a function. This makes discrete sequences of the form
xn = e jωn
eigensequences with respect to an LTI system T , because for each ω,
there is a complex number (eigenvalue) H(ω) such that
T {xn } = H(ω) · {xn }
In the notation of slide 30, where the argument of H is the base, we would write H(e jω ).
33

Recall: Fourier transform

The Fourier integral transform and its inverse are defined as
Z ∞
F{g(t)}(ω) = G(ω) = α g(t) · e∓ jωt dt
−∞
Z ∞
F −1
{G(ω)}(t) = g(t) = β G(ω) · e± jωt dω
−∞

where α and β are constants chosen such that αβ = 1/(2π).

Many equivalent forms of the Fourier transform are used in the literature. There is no strong
consensus on whether the forward transform uses e− jωt and the backwards transform e jωt , or
vice versa. We follow here those authors who set α = 1 and β =√1/(2π), to keep the convolution
theorem free of a constant prefactor; others prefer α = β = 1/ 2π, in the interest of symmetry.

The substitution ω = 2πf leads to a form without prefactors:

Z ∞
F{h(t)}(f ) = H(f ) = h(t) · e∓2π jf t dt
−∞
Z ∞
F −1
{H(f )}(t) = h(t) = H(f )· e±2π jf t df
−∞

34
Another notation is in the continuous case
Z ∞
F{h(t)}(ω) = H(e ) =jω
h(t) · e− jωt dt
−∞
∞
1
Z
F −1 jω
{H(e )}(t) = h(t) = H(e jω ) · e jωt dω
2π −∞

and in the discrete-sequence case

∞
X
F{hn }(ω) = H(e jω ) = hn · e− jωn
n=−∞

1 π
Z
F −1 jω
{H(e )}(t) = hn = H(e jω ) · e jωn dω
2π −π

which treats the Fourier transform as a special case of the z-transform

(to be introduced shortly).

35

Properties of the Fourier transform

If
x(t) •−◦ X(f ) and y(t) •−◦ Y (f )
are pairs of functions that are mapped onto each other by the Fourier
transform, then so are the following pairs.
Linearity:
ax(t) + by(t) •−◦ aX(f ) + bY (f )

Time scaling:  
1 f
x(at) •−◦ X
|a| a
Frequency scaling:
 
1 t
x •−◦ X(af )
|a| a

36
Time shifting:

x(t − ∆t) •−◦ X(f ) · e−2π jf ∆t

Frequency shifting:

x(t) · e2π j∆f t •−◦ X(f − ∆f )

Parseval’s theorem (total power):

Z ∞ Z ∞
2
|x(t)| dt = |X(f )|2 df
−∞ −∞

37

Fourier transform example: rect and sinc

The Fourier transform of the “rectangular function”

1
 1 if |t| < 2

1
rect(t) = 2
if |t| = 12

 0 otherwise

is the “(normalized) sinc function”

1
sin πf
Z
2
F{rect(t)}(f ) = e−2π jf t dt = = sinc(f )
− 21 πf

and vice versa

F{sinc(t)}(f ) = rect(f ).
Some noteworthy properties of these functions:
R∞ R∞
• −∞ sinc(t) dt = 1 = −∞ rect(t) dt
• sinc(0) = 1 = rect(0)
• ∀n ∈ Z \ {0} : sinc(k) = 0
38
 Convolution theorem
Continuous form:

F{(f ∗ g)(t)} = F{f (t)} · F{g(t)}

F{f (t) · g(t)} = F{f (t)} ∗ F{g(t)}

Discrete form:

{xn } ∗ {yn } = {zn } ⇐⇒ X(e jω ) · Y (e jω ) = Z(e jω )

Convolution in the time domain is equivalent to (complex) scalar mul-

tiplication in the frequency domain.
Convolution in the frequency domain corresponds to scalar multiplica-
tion in the time domain.
− jωr dr = − jωr dsdr =
R R R R
Proof: z(r) = s x(s)y(r − s)ds ⇐⇒ r z(r)e r s x(s)y(r − s)e
− jωr drds = − jωs − jω(r−s) drds t:=r−s
R R R R
s x(s) r y(r − s)e s x(s)e r y(r − s)e =
− jωs − jωt dtds = − jωs ds · − jωt dt. (Same for
R R R R P R
s x(s)e t y(t)e s x(s)e t y(t)e instead of .)
39

Dirac’s delta function

The continuous equivalent of the impulse sequence {δn } is known as
Dirac’s delta function δ(x). It is a generalized function, defined such
that

0, x 6= 0 1
δ(x) =
∞, x = 0
Z ∞
δ(x) dx = 1
−∞
0 x
and can be thought of as the limit of function sequences such as

0, |x| ≥ 1/n
δ(x) = lim
n→∞ n/2, |x| < 1/n
or
n 2 2
δ(x) = lim √ e−n x
n→∞ π
The delta function is mathematically speaking not a function, but a distribution, that is an
expression that is only defined when integrated.
40
Some properties of Dirac’s delta function:
Z ∞
f (x)δ(x − a) dx = f (a)
−∞
Z ∞
e±2π jf t df = δ(t)
−∞
∞
1
Z
e± jωt dω = δ(t)
2π −∞

Fourier transform:
Z ∞
F{δ(t)}(ω) = δ(t) · e− jωt dt = e0 = 1
−∞
∞
1
Z
F −1
{1}(t) = 1 · e jωt dω = δ(t)
2π −∞

http://mathworld.wolfram.com/DeltaFunction.html
41

Sine and cosine in the frequency domain

1 2π jf0 t 1 −2π jf0 t 1 2π jf0 t 1 −2π jf0 t
cos(2πf0 t) = e + e e sin(2πf0 t) =
− e
2 2 2j 2j
1 1
F{cos(2πf0 t)}(f ) = δ(f − f0 ) + δ(f + f0 )
2 2
j j
F{sin(2πf0 t)}(f ) = − δ(f − f0 ) + δ(f + f0 )
2 2
ℜ ℜ
1 1
2 2
ℑ ℑ
1 1
2 j 2 j

−f0 f0 f −f0 f0 f

As any real-valued signal x(t) can be represented as a combination of sine and cosine functions,
the spectrum of any real-valued signal will show the symmetry X(e jω ) = [X(e− jω )]∗ , where ∗
denotes the complex conjugate (i.e., negated imaginary part).
42
 Fourier transform symmetries
We call a function x(t)

odd if x(−t) = −x(t)

even if x(−t) = x(t)

and ·∗ is the complex conjugate, such that (a + jb)∗ = (a − jb).

Then
x(t) is real ⇔ X(−f ) = [X(f )]∗
x(t) is imaginary ⇔ X(−f ) = −[X(f )]∗
x(t) is even ⇔ X(f ) is even
x(t) is odd ⇔ X(f ) is odd
x(t) is real and even ⇔ X(f ) is real and even
x(t) is real and odd ⇔ X(f ) is imaginary and odd
x(t) is imaginary and even ⇔ X(f ) is imaginary and even
x(t) is imaginary and odd ⇔ X(f ) is real and odd

43

Example: amplitude modulation

Communication channels usually permit only the use of a given fre-
quency interval, such as 300–3400 Hz for the analog phone network or
590–598 MHz for TV channel 36. Modulation with a carrier frequency
fc shifts the spectrum of a signal x(t) into the desired band.
Amplitude modulation (AM):

y(t) = A · cos(2πtfc ) · x(t)

X(f ) Y (f )
∗ =

−fl 0 fl f −fc fc f −fc 0 fc f

The spectrum of the baseband signal in the interval −fl < f < fl is
shifted by the modulation to the intervals ±fc − fl < f < ±fc + fl .
How can such a signal be demodulated?
44
 Sampling using a Dirac comb
The loss of information in the sampling process that converts a con-
tinuous function x(t) into a discrete sequence {xn } defined by

xn = x(ts · n) = x(n/fs )

can be modelled through multiplying x(t) by a comb of Dirac impulses

∞
X
s(t) = ts · δ(t − ts · n)
n=−∞

to obtain the sampled function

x̂(t) = x(t) · s(t)

The function x̂(t) now contains exactly the same information as the
discrete sequence {xn }, but is still in a form that can be analysed using
the Fourier transform on continuous functions.
45

The Fourier transform of a Dirac comb

∞
X ∞
X
s(t) = ts · δ(t − ts · n) = e2π jnt/ts
n=−∞ n=−∞

is another Dirac comb

( ∞
)
X
S(f ) = F ts · δ(t − ts n) (f ) =
n=−∞
Z∞ X
∞  ∞ 
n
X
ts · δ(t − ts n) e2π jf t dt = δ f− .
n=−∞ n=−∞
ts
−∞

s(t) S(f )

−2ts −ts 0 ts 2ts t −2fs −fs 0 fs 2fs f

46
 Sampling and aliasing
sample
cos(2π tf)
cos(2π t(k⋅ fs± f))

Sampled at frequency fs , the function cos(2πtf ) cannot be distin-

guished from cos[2πt(kfs ± f )] for any k ∈ Z.
47

Frequency-domain view of sampling

x(t) s(t) x̂(t)

· =
... ... ... ...
0 t −1/fs 0 1/fs t −1/fs 0 1/fs t
X(f ) S(f ) X̂(f )

∗ =
... ... ... ...
0 f −fs fs f −fs 0 fs f

Sampling a signal in the time domain corresponds in the frequency

domain to convolving its spectrum with a Dirac comb. The resulting
copies of the original signal spectrum in the spectrum of the sampled
signal are called “images”.
48
 Nyquist limit and anti-aliasing filters
If the (double-sided) bandwidth of a signal to be sampled is larger than
the sampling frequency fs , the images of the signal that emerge during
sampling may overlap with the original spectrum.
Such an overlap will hinder reconstruction of the original continuous
signal by removing the aliasing frequencies with a reconstruction filter.
Therefore, it is advisable to limit the bandwidth of the input signal to
the sampling frequency fs before sampling, using an anti-aliasing filter.
In the common case of a real-valued base-band signal (with frequency
content down to 0 Hz), all frequencies f that occur in the signal with
non-zero power should be limited to the interval −fs /2 < f < fs /2.
The upper limit fs /2 for the single-sided bandwidth of a baseband
signal is known as the “Nyquist limit”.

49

Nyquist limit and anti-aliasing filters

Without anti-aliasing filter With anti-aliasing filter
single-sided Nyquist
X(f ) bandwidth X(f ) limit = fs /2
anti-aliasing filter

0 f −fs 0 fs f
double-sided bandwidth
reconstruction filter
X̂(f ) X̂(f )

−2fs −fs 0 fs 2fs f −2fs −fs 0 fs 2fs f

Anti-aliasing and reconstruction filters both suppress frequencies outside |f | < fs /2.

50
 Reconstruction of a continuous
band-limited waveform
The ideal anti-aliasing filter for eliminating any frequency content above
fs /2 before sampling with a frequency of fs has the Fourier transform
(
1 if |f | < f2s
H(f ) = fs = rect(ts f ).
0 if |f | > 2
This leads, after an inverse Fourier transform, to the impulse response
 
sin πtfs 1 t
h(t) = fs · = · sinc .
πtfs ts ts
The original band-limited signal can be reconstructed by convolving
this with the sampled signal x̂(t), which eliminates the periodicity of
the frequency domain introduced by the sampling process:
x(t) = h(t) ∗ x̂(t)
Note that sampling h(t) gives the impulse function: h(t) · s(t) = δ(t).
51

Impulse response of ideal low-pass filter with cut-off frequency fs /2:

−3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 3

t⋅ fs
52
 Reconstruction filter example

sampled signal
interpolation result
scaled/shifted sin(x)/x pulses

1 2 3 4 5
53

Reconstruction filters
The mathematically ideal form of a reconstruction filter for suppressing
aliasing frequencies interpolates the sampled signal xn = x(ts · n) back
into the continuous waveform
∞
X sin π(t − ts · n)
x(t) = xn · .
n=−∞
π(t − ts · n)

Choice of sampling frequency

Due to causality and economic constraints, practical analog filters can only approx-
imate such an ideal low-pass filter. Instead of a sharp transition between the “pass
band” (< fs /2) and the “stop band” (> fs /2), they feature a “transition band”
in which their signal attenuation gradually increases.
The sampling frequency is therefore usually chosen somewhat higher than twice
the highest frequency of interest in the continuous signal (e.g., 4×). On the other
hand, the higher the sampling frequency, the higher are CPU, power and memory
requirements. Therefore, the choice of sampling frequency is a tradeoff between
signal quality, analog filter cost and digital subsystem expenses.
54
Exercise 6 Digital-to-analog converters cannot output Dirac pulses. In-
stead, for each sample, they hold the output voltage (approximately) con-
stant, until the next sample arrives. How can this behaviour be modeled
mathematically as a linear time-invariant system, and how does it affect the
spectrum of the output signal?

Exercise 7 Many DSP systems use “oversampling” to lessen the require-

ments on the design of an analog reconstruction filter. They use (a finite
approximation of) the sinc-interpolation formula to multiply the sampling
frequency fs of the initial sampled signal by a factor N before passing it to
the digital-to-analog converter. While this requires more CPU operations
and a faster D/A converter, the requirements on the subsequently applied
analog reconstruction filter are much less stringent. Explain why, and draw
schematic representations of the signal spectrum before and after all the
relevant signal-processing steps.

Exercise 8 Similarly, explain how oversampling can be applied to lessen

the requirements on the design of an analog anti-aliasing filter.

55

Band-pass signal sampling

Sampled signals can also be reconstructed if their spectral components
remain entirely within the interval n · fs /2 < |f | < (n + 1) · fs /2 for
some n ∈ N. (The baseband case discussed so far is just n = 0.)
In this case, the aliasing copies of the positive and the negative frequencies will interleave instead
of overlap, and can therefore be removed again with a reconstruction filter with the impulse
response

„ «
sin πtfs /2 2n + 1 sin πt(n + 1)fs sin πtnfs
h(t) = fs · cos 2πtfs = (n + 1)fs − nfs .
πtfs /2 4 πt(n + 1)fs πtnfs

X(f ) anti-aliasing filter X̂(f ) reconstruction filter

− 45 fs 0 5
4 fs f −fs −fs 0 fs fs f
2 2
n=2

56
Exercise 9 Reconstructing a sampled baseband signal:

• Generate a one second long Gaussian noise sequence {rn } (using

MATLAB function randn) with a sampling rate of 300 Hz.
• Use the fir1(50, 45/150) function to design a finite impulse re-
sponse low-pass filter with a cut-off frequency of 45 Hz. Use the
filtfilt function in order to apply that filter to the generated noise
signal, resulting in the filtered noise signal {xn }.
• Then sample {xn } at 100 Hz by setting all but every third sample
value to zero, resulting in sequence {yn }.
• Generate another low-pass filter with a cut-off frequency of 50 Hz
and apply it to {yn }, in order to interpolate the reconstructed filtered
noise signal {zn }. Multiply the result by three, to compensate the
energy lost during sampling.
• Plot {xn }, {yn }, and {zn }. Finally compare {xn } and {zn }.

Why should the first filter have a lower cut-off frequency than the second?
57

Exercise 10 Reconstructing a sampled band-pass signal:

• Generate a 1 s noise sequence {rn }, as in exercise 9, but this time
use a sampling frequency of 3 kHz.
• Apply to that a band-pass filter that attenuates frequencies outside
the interval 31–44 Hz, which the MATLAB Signal Processing Toolbox
function cheby2(3, 30, [31 44]/1500) will design for you.
• Then sample the resulting signal at 30 Hz by setting all but every
100-th sample value to zero.
• Generate with cheby2(3, 20, [30 45]/1500) another band-pass
filter for the interval 30–45 Hz and apply it to the above 30-Hz-
sampled signal, to reconstruct the original signal. (You’ll have to
multiply it by 100, to compensate the energy lost during sampling.)
• Plot all the produced sequences and compare the original band-pass
signal and that reconstructed after being sampled at 30 Hz.
Why does the reconstructed waveform differ much more from the original
if you reduce the cut-off frequencies of both band-pass filters by 5 Hz?
58
 Spectrum of a periodic signal
A signal x(t) that is periodic with frequency fp can be factored into a
single period ẋ(t) convolved with an impulse comb p(t). This corre-
sponds in the frequency domain to the multiplication of the spectrum
of the single period with a comb of impulses spaced fp apart.
x(t) ẋ(t) p(t)

= ∗
... ... ... ...
−1/fp 0 1/fp t t −1/fp 0 1/fp t
X(f ) Ẋ(f ) P (f )

= ·
... ...
−fp 0 fp f f −fp 0 fp f
59

Spectrum of a sampled signal

A signal x(t) that is sampled with frequency fs has a spectrum that is
periodic with a period of fs .

x(t) s(t) x̂(t)

· =
... ... ... ...
0 t −1/fs 0 1/fs t −1/fs 0 1/fs t
X(f ) S(f ) X̂(f )

∗ =
... ... ... ...
0 f −fs fs f −fs 0 fs f

60
 Continuous vs discrete Fourier transform
• Sampling a continuous signal makes its spectrum periodic
• A periodic signal has a sampled spectrum

We sample a signal x(t) with fs , getting x̂(t). We take n consecutive

samples of x̂(t) and repeat these periodically, getting a new signal ẍ(t)
with period n/fs . Its spectrum Ẍ(f ) is sampled (i.e., has non-zero
value) at frequency intervals fs /n and repeats itself with a period fs .
Now both ẍ(t) and its spectrum Ẍ(f ) are finite vectors of length n.

ẍ(t) Ẍ(f )

... ... ... ...

−n/fs fs−1 0 fs−1 n/fs t −fs −fs /n 0 fs /n fs f
61

Discrete Fourier Transform (DFT)

n−1 n−1
X
−2π j ik 1 X ik
Xk = xi · e n xk = · Xi · e2π j n
i=0
n i=0
The n-point DFT multiplies a vector with an n × n matrix
1 1 1 1 ··· 1
 
1 2 3 n−1

 1 e−2π j n e−2π j n e−2π j n ··· e−2π j n 

2 4 6 2(n−1)

 1 e−2π j n e−2π j n e−2π j n ··· e−2π j n 

Fn = 
 1
3
e−2π j n
6
e−2π j n
9
e−2π j n ··· e−2π j n
3(n−1) 

.. .. .. .. .. ..
 
.
 
 . . . . . 
n−1 2(n−1) 3(n−1) (n−1)(n−1)
1 e−2π j n e−2π j n e−2π j n · · · e−2π j n

       
x0 X0 X0 x0
 x1   X1   X1   x1 
  
 x2   X2 
 1 
∗  X2 
 
x2

Fn ·  = , ·F · =
 
 ..   ..  n n  ..   ..


 .   .   .   . 
xn−1 Xn−1 Xn−1 xn−1

62
 Discrete Fourier Transform visualized
     
x0 X0
x1 X1
     
     
x2 X2
     
     
     


 
· x3  
= X3 



 
  x4  
  X4 



 
  x5  
  X5 

x6 X6
     
     
x7 X7

The n-point DFT of a signal {xi } sampled at frequency fs contains in

the elements X0 to Xn/2 of the resulting frequency-domain vector the
frequency components 0, fs /n, 2fs /n, 3fs /n, . . . , fs /2, and contains
in Xn−1 downto Xn/2 the corresponding negative frequencies. Note
that for a real-valued input vector, both X0 and Xn/2 will be real, too.
Why is there no phase information recovered at fs /2?
63

Inverse DFT visualized

     
X0 x0
X1 x1
     
     
X2 x2
     
     
     
1    X3   x3 
· · = 
8 

 
  X4  
  x4 



 
  X5  
  x5 

X6 x6
     
     
X7 x7

64
 Fast Fourier Transform (FFT)
n−1
X ik
n−1
xi · e−2π j n


Fn {xi }i=0 k
=
i=0
n n
2
−1 2
−1
ik k ik
−2π j n/2 −2π j n/2
X X
= x2i · e + e−2π j n x2i+1 · e
i=0 i=0
  n n
−1 −1
 k
 
2 −2π j n 2 n
 F n2 {x2i }i=0 + e · F n2 {x2i+1 }i=0 , k<

2

k k
= n n
−1 −1
  k
 
2
 F n2 {x2i }i=0 + e−2π j n 2
· F n2 {x2i+1 }i=0 , k≥ n


k− n2 n
k− 2 2

The DFT over n-element vectors can be reduced to two DFTs over
n/2-element vectors plus n multiplications and n additions, leading to
log2 n rounds and n log2 n additions and multiplications overall, com-
pared to n2 for the equivalent matrix multiplication.
A high-performance FFT implementation in C with many processor-specific optimizations and
support for non-power-of-2 sizes is available at http://www.fftw.org/.
65

Efficient real-valued FFT

The symmetry properties of the Fourier transform applied to the discrete
n−1 n−1
Fourier transform {Xi }i=0 = Fn {xi }i=0 have the form
∀i : xi = ℜ(xi ) ⇐⇒ ∀i : Xn−i = Xi∗
∀i : xi = j · ℑ(xi ) ⇐⇒ ∀i : Xn−i = −Xi∗
These two symmetries, combined with the linearity of the DFT, allows us
to calculate two real-valued n-point DFTs
n−1 n−1 n−1 n−1
{Xi′ }i=0 = Fn {x′i }i=0 {Xi′′ }i=0 = Fn {x′′i }i=0
simultaneously in a single complex-valued n-point DFT, by composing its
input as
xi = x′i + j · x′′i
and decomposing its output as
1 1
Xi′ = (Xi + Xn−i ∗
) Xi′′ = (Xi − Xn−i
∗
)
2 2

To optimize the calculation of a single real-valued FFT, use this trick to calculate the two half-size
real-value FFTs that occur in the first round.
66
 Fast complex multiplication
Calculating the product of two complex numbers as

(a + jb) · (c + jd) = (ac − bd) + j(ad + bc)

involves four (real-valued) multiplications and two additions.

The alternative calculation
α = a(c + d)
(a + jb) · (c + jd) = (α − β) + j(α + γ) with β = d(a + b)
γ = c(b − a)

provides the same result with three multiplications and five additions.
The latter may perform faster on CPUs where multiplications take three
or more times longer than additions.
This trick is most helpful on simpler microcontrollers. Specialized signal-processing CPUs (DSPs)
feature 1-clock-cycle multipliers. High-end desktop processors use pipelined multipliers that stall
where operations depend on each other.

67

FFT-based convolution
m−1 n−1
Calculating the convolution of two finite sequences {xi }i=0 and {yi }i=0
of lengths m and n via
min{m−1,i}
X
zi = xj · yi−j , 0≤i<m+n−1
j=max{0,i−(n−1)}

takes mn multiplications.
Can we apply the FFT and the convolution theorem to calculate the
convolution faster, in just O(m log m + n log n) multiplications?
{zi } = F −1 (F{xi } · F{yi })
There is obviously no problem if this condition is fulfilled:
{xi } and {yi } are periodic, with equal period lengths
In this case, the fact that the DFT interprets its input as a single period
of a periodic signal will do exactly what is needed, and the FFT and
inverse FFT can be applied directly as above.
68
In the general case, measures have to be taken to prevent a wrap-over:
−1
A B F [F(A)⋅F(B)]

−1
A’ B’ F [F(A’)⋅F(B’)]

Both sequences are padded with zero values to a length of at least m+n−1.
This ensures that the start and end of the resulting sequence do not overlap.
69

Zero padding is usually applied to extend both sequence lengths to the

next higher power of two (2⌈log2 (m+n−1)⌉ ), which facilitates the FFT.
With a causal sequence, simply append the padding zeros at the end.
With a non-causal sequence, values with a negative index number are
wrapped around the DFT block boundaries and appear at the right
end. In this case, zero-padding is applied in the center of the block,
between the last and first element of the sequence.
Thanks to the periodic nature of the DFT, zero padding at both ends
has the same effect as padding only at one end.
If both sequences can be loaded entirely into RAM, the FFT can be ap-
plied to them in one step. However, one of the sequences might be too
large for that. It could also be a realtime waveform (e.g., a telephone
signal) that cannot be delayed until the end of the transmission.
In such cases, the sequence has to be split into shorter blocks that are
separately convolved and then added together with a suitable overlap.

70
Each block is zero-padded at both ends and then convolved as before:

∗ ∗ ∗

= = =

The regions originally added as zero padding are, after convolution, aligned
to overlap with the unpadded ends of their respective neighbour blocks.
The overlapping parts of the blocks are then added together.
71

Deconvolution
A signal u(t) was distorted by convolution with a known impulse re-
sponse h(t) (e.g., through a transmission channel or a sensor problem).
The “smeared” result s(t) was recorded.
Can we undo the damage and restore (or at least estimate) u(t)?

∗ =

∗ =

72
The convolution theorem turns the problem into one of multiplication:
Z
s(t) = u(t − τ ) · h(τ ) · dτ

s = u∗h
F{s} = F{u} · F{h}
F{u} = F{s}/F{h}
u = F −1 {F{s}/F{h}}
In practice, we also record some noise n(t) (quantization, etc.):
Z
c(t) = s(t) + n(t) = u(t − τ ) · h(τ ) · dτ + n(t)

Problem – At frequencies f where F{h}(f ) approaches zero, the

noise will be amplified (potentially enormously) during deconvolution:
ũ = F −1 {F{c}/F{h}} = u + F −1 {F{n}/F{h}}
73

Typical workarounds:
→ Modify the Fourier transform of the impulse response, such that
|F{h}(f )| > ǫ for some experimentally chosen threshold ǫ.
→ If estimates of the signal spectrum |F{s}(f )| and the noise
spectrum |F{n}(f )| can be obtained, then we can apply the
“Wiener filter” (“optimal filter”)
|F{s}(f )|2
W (f ) =
|F{s}(f )|2 + |F{n}(f )|2
before deconvolution:
ũ = F −1 {W · F{c}/F{h}}
Exercise 11 Use MATLAB to deconvolve the blurred stars from slide 26.
The files stars-blurred.png with the blurred-stars image and stars-psf.png with the impulse
response (point-spread function) are available on the course-material web page. You may find
the MATLAB functions imread, double, imagesc, circshift, fft2, ifft2 of use.
Try different ways to control the noise (see above) and distortions near the margins (window-
ing). [The MATLAB image processing toolbox provides ready-made “professional” functions
deconvwnr, deconvreg, deconvlucy, edgetaper, for such tasks. Do not use these, except per-
haps to compare their outputs with the results of your own attempts.]
74
 Spectral estimation
Sine wave 4×fs/32 Discrete Fourier Transform
1
15

10
0
5

−1 0
0 10 20 30 0 10 20 30
Sine wave 4.61×f /32 Discrete Fourier Transform
s
1
15

10
0
5

−1 0
0 10 20 30 0 10 20 30

75

We introduced the DFT as a special case of the continuous Fourier

transform, where the input is sampled and periodic.
If the input is sampled, but not periodic, the DFT can still be used
to calculate an approximation of the Fourier transform of the original
continuous signal. However, there are two effects to consider. They
are particularly visible when analysing pure sine waves.
Sine waves whose frequency is a multiple of the base frequency (fs /n)
of the DFT are identical to their periodic extension beyond the size
of the DFT. They are, therefore, represented exactly by a single sharp
peak in the DFT. All their energy falls into one single frequency “bin”
in the DFT result.
Sine waves with other frequencies, which do not match exactly one of
the output frequency bins of the DFT, are still represented by a peak
at the output bin that represents the nearest integer multiple of the
DFT’s base frequency. However, such a peak is distorted in two ways:
→ Its amplitude is lower (down to 63.7%).
→ Much signal energy has “leaked” to other frequencies.
76
 35

30

25

20

15

10

0
0 16
5 10 15 15.5
20 25 30 15
input freq.
DFT index

The leakage of energy to other frequency bins not only blurs the estimated spec-
trum. The peak amplitude also changes significantly as the frequency of a tone
changes from that associated with one output bin to the next, a phenomenon
known as scalloping. In the above graphic, an input sine wave gradually changes
from the frequency of bin 15 to that of bin 16 (only positive frequencies shown).
77

Windowing
Sine wave Discrete Fourier Transform
1 300

200
0
100

−1 0
0 200 400 0 200 400
Sine wave multiplied with window function Discrete Fourier Transform
1 100

0 50

−1 0
0 200 400 0 200 400
78
The reason for the leakage and scalloping losses is easy to visualize with the
help of the convolution theorem:
The operation of cutting a sequence of the size of the DFT input vector out
of a longer original signal (the one whose continuous Fourier spectrum we
try to estimate) is equivalent to multiplying this signal with a rectangular
function. This destroys all information and continuity outside the “window”
that is fed into the DFT.
Multiplication with a rectangular window of length T in the time domain is
equivalent to convolution with sin(πf T )/(πf T ) in the frequency domain.
The subsequent interpretation of this window as a periodic sequence by
the DFT leads to sampling of this convolution result (sampling meaning
multiplication with a Dirac comb whose impulses are spaced fs /n apart).
Where the window length was an exact multiple of the original signal period,
sampling of the sin(πf T )/(πf T ) curve leads to a single Dirac pulse, and
the windowing causes no distortion. In all other cases, the effects of the con-
volution become visible in the frequency domain as leakage and scalloping
losses.

79

Some better window functions

0.8

0.6

0.4

0.2
Rectangular window
Triangular window
0
Hann window
Hamming window

0 0.2 0.4 0.6 0.8 1

All these functions are 0 outside the interval [0,1].
80
 Rectangular window (64−point) Triangular window

Magnitude (dB) 20 20

Magnitude (dB)
0 0

−20 −20

−40 −40

−60 −60
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)
Hann window Hamming window

20 20
Magnitude (dB)

Magnitude (dB)
0 0

−20 −20

−40 −40

−60 −60
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)

81

Numerous alternatives to the rectangular window have been proposed

that reduce leakage and scalloping in spectral estimation. These are
vectors multiplied element-wise with the input vector before applying
the DFT to it. They all force the signal amplitude smoothly down to
zero at the edge of the window, thereby avoiding the introduction of
sharp jumps in the signal when it is extended periodically by the DFT.
n−1
Three examples of such window vectors {wi }i=0 are:
Triangular window (Bartlett window):
 
 i 
wi = 1 −  1 −
 
n/2 
Hann window (raised-cosine window, Hanning window):
 
i
wi = 0.5 − 0.5 × cos 2π
n−1
Hamming window:
 
i
wi = 0.54 − 0.46 × cos 2π
n−1
82
 Zero padding increases DFT resolution
The two figures below show two spectra of the 16-element sequence
si = cos(2π · 3i/16) + cos(2π · 4i/16), i ∈ {0, . . . , 15}.
The left plot shows the DFT of the windowed sequence
xi = si · wi , i ∈ {0, . . . , 15}
and the right plot shows the DFT of the zero-padded windowed sequence

si · wi , i ∈ {0, . . . , 15}
x′i =
0, i ∈ {16, . . . , 63}
where wi = 0.54 − 0.46 × cos (2πi/15) is the Hamming window.

DFT without zero padding DFT with 48 zeros appended to window

4 4

2 2

0 0
0 5 10 15 0 20 40 60
83

Applying the discrete Fourier transform to an n-element long real-

valued sequence leads to a spectrum consisting of only n/2+1 discrete
frequencies.
Since the resulting spectrum has already been distorted by multiplying
the (hypothetically longer) signal with a windowing function that limits
its length to n non-zero values and forces the waveform smoothly down
to zero at the window boundaries, appending further zeros outside the
window will not distort the signal further.
The frequency resolution of the DFT is the sampling frequency divided
by the block size of the DFT. Zero padding can therefore be used to
increase the frequency resolution of the DFT.
Note that zero padding does not add any additional information to the
signal. The spectrum has already been “low-pass filtered” by being
convolved with the spectrum of the windowing function. Zero padding
in the time domain merely samples this spectrum blurred by the win-
dowing step at a higher resolution, thereby making it easier to visually
distinguish spectral lines and to locate their peak more precisely.
84
 Frequency inversion
In order to turn the spectrum X(f ) of a real-valued signal xi sampled at fs
into an inverted spectrum X ′ (f ) = X(fs /2 − f ), we merely have to shift
the periodic spectrum by fs /2:
X ′ (f ) X(f )

= ∗
... ... ... ...
−fs 0 fs f −fs 0 fs f − f2s 0 fs
2
f

This can be accomplished by multiplying the sampled sequence xi with yi =

cos πfs t = cos πi, which is nothing but multiplication with the sequence

. . . , 1, −1, 1, −1, 1, −1, 1, −1, . . .

So in order to design a discrete high-pass filter that attenuates all frequencies

f outside the range fc < |f | < fs /2, we merely have to design a low-pass
filter that attenuates all frequencies outside the range −fc < f < fc , and
then multiply every second value of its impulse response with −1.
85

Window-based design of FIR filters

Recall that the ideal continuous low-pass filter with cut-off frequency
fc has the frequency characteristic
  
1 if |f | < fc f
H(f ) = = rect
0 if |f | > fc 2fc
and the impulse response
sin 2πtfc
h(t) = 2fc = 2fc · sinc(2fc · t).
2πtfc
Sampling this impulse response with the sampling frequency fs of the
signal to be processed will lead to a periodic frequency characteristic,
that matches the periodic spectrum of the sampled signal.
There are two problems though:
→ the impulse response is infinitely long
→ this filter is not causal, that is h(t) 6= 0 for t < 0
86
Solutions:

→ Make the impulse response finite by multiplying the sampled

h(t) with a windowing function
→ Make the impulse response causal by adding a delay of half the
window size

The impulse response of an n-th order low-pass filter is then chosen as

sin[2π(i − n/2)fc /fs ]
hi = 2fc /fs · · wi
2π(i − n/2)fc /fs
where {wi } is a windowing sequence, such as the Hamming window

wi = 0.54 − 0.46 × cos (2πi/n)

with wi = 0 for i < 0 and i > n.

Note that for fc = fs /4, we have hi = 0 for all even values of i. Therefore, this special case
requires only half the number of multiplications during the convolution. Such “half-band” FIR
filters are used, for example, as anti-aliasing filters wherever a sampling rate needs to be halved.
87

FIR low-pass filter design example

Impulse Response
0.3
1
Imaginary Part

0.2
Amplitude

0.5
30
0 0.1
−0.5 0
−1
−0.1
−1 0 1 0 10 20 30
Real Part n (samples)
0 0
Phase (degrees)
Magnitude (dB)

−20 −500

−40 −1000

−60 −1500
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)

order: n = 30, cutoff frequency (−6 dB): fc = 0.25 × fs /2, window: Hamming
88
We truncate the ideal, infinitely-long impulse response by multiplication with a window sequence.
In the frequency domain, this will convolve the rectangular frequency response of the ideal low-
pass filter with the frequency characteristic of the window. The width of the main lobe determines
the width of the transition band, and the side lobes cause ripples in the passband and stopband.

Converting a low-pass into a band-pass filter

To obtain a band-pass filter that attenuates all frequencies f outside
the range fl < f < fh , we first design a low-pass filter with a cut-off
frequency (fh − fl )/2 and multiply its impulse response with a sine
wave of frequency (fh + fl )/2, before applying the usual windowing:

sin[π(i − n/2)(fh − fl )/fs ]

hi = (fh − fl )/fs · · sin[π(fh + fl )] · wi
π(i − n/2)(fh − fl )/fs

H(f )
= ∗

−fh −fl 0 fl fh f − fh −f
2
l fh −fl
2
f − fh +f
2
l
0 fh +fl
2
f

89

Exercise 12 Explain the difference between the DFT, FFT, and FFTW.

Exercise 13 Push-button telephones use a combination of two sine tones

to signal, which button is currently being pressed:
1209 Hz 1336 Hz 1477 Hz 1633 Hz
697 Hz 1 2 3 A
770 Hz 4 5 6 B
852 Hz 7 8 9 C
941 Hz * 0 # D
(a) You receive a digital telephone signal with a sampling frequency of
8 kHz. You cut a 256-sample window out of this sequence, multiply it with a
windowing function and apply a 256-point DFT. What are the indices where
the resulting vector (X0 , X1 , . . . , X255 ) will show the highest amplitude if
button 9 was pushed at the time of the recording?
(b) Use MATLAB to determine, which button sequence was typed in the
touch tones recorded in the file touchtone.wav on the course-material web
page.
90
 Polynomial representation of sequences
We can represent sequences {xn } as polynomials:
∞
X
X(v) = xn v n
n=−∞

Example of polynomial multiplication:

(1 + 2v + 3v 2 ) · (2 + 1v)
2 + 4v + 6v 2
+ 1v + 2v 2 + 3v 3
= 2 + 5v + 8v 2 + 3v 3

Compare this with the convolution of two sequences (in MATLAB):

conv([1 2 3], [2 1]) equals [2 5 8 3]

91

Convolution of sequences is equivalent to polynomial multiplication:

∞
X
{hn } ∗ {xn } = {yn } ⇒ yn = hk · xn−k
k=−∞
↓ ↓
∞
! ∞
!
X X
H(v) · X(v) = hn v n · xn v n
n=−∞ n=−∞
∞
X ∞
X
= hk · xn−k · v n
n=−∞ k=−∞

Note how the Fourier transform of a sequence can be accessed easily

from its polynomial form:
∞
X
− jω
X(e )= xn e− jωn
n=−∞

92
Example of polynomial division:
∞
1 2 2 3 3
X
= 1 + av + a v + a v + · · · = an v n
1 − av n=0

1 + av + a2 v 2 + · · ·
1 − av 1
1 − av
av
av − a2 v 2
a2 v 2
a2 v 2 − a 3 v 3
···

Rational functions (quotients of two polynomials) can provide a con-

venient closed-form representations for infinitely-long exponential se-
quences, in particular the impulse responses of IIR filters.

93

The z-transform
The z-transform of a sequence {xn } is defined as:
∞
X
X(z) = xn z −n
n=−∞

Note that is differs only in the sign of the exponent from the polynomial representation discussed
on the preceeding slides.

Recall that the above X(z) is exactly the factor with which an expo-
nential sequence {z n } is multiplied, if it is convolved with {xn }:

{z n } ∗ {xn } = {yn }
∞
X ∞
X
⇒ yn = z n−k xk = z n · z −k xk = z n · X(z)
k=−∞ k=−∞

94
The z-transform defines for each sequence a continuous complex-valued
surface over the complex plane C. For finite sequences, its value is al-
ways defined across the entire complex plane.
For infinite sequences, it can be shown that the z-transform converges
only for the region
   
 xn+1   xn+1 
lim   < |z| < lim  
n→∞  xn  n→−∞  xn 

The z-transform identifies a sequence unambiguously only in conjunction with a given region of
convergence. In other words, there exist different sequences, that have the same expression as
their z-transform, but that converge for different amplitudes of z.

The z-transform is a generalization of the Fourier transform, which it

contains on the complex unit circle (|z| = 1):
∞
X
F{xn }(ω) = X(e jω ) = xn e− jωn
n=−∞

95

The z-transform of the impulse xn b0 a−1

0 yn
response {hn } of the causal LTI
system defined by z −1 z −1
b1 −a1
xn−1 yn−1
k
X m
X
al · yn−l = bl · xn−l z −1 z −1
··· ···
l=0 l=0 ··· ···
with {yn } = {hn } ∗ {xn } is the z −1 z −1
bm −ak
rational function xn−m yn−k

b0 + b1 z −1 + b2 z −2 + · · · + bm z −m
H(z) =
a0 + a1 z −1 + a2 z −2 + · · · + ak z −k
(bm =
6 0, ak 6= 0) which can also be written as
zk m m−l
P
l=0 bl z
H(z) = .
z m kl=0 al z k−l
P

H(z) has m zeros and k poles at non-zero locations in the z plane,

plus k − m zeros (if k > m) or m − k poles (if m > k) at z = 0.
96
This function can be converted into the form
m
Y Ym
−1
(1 − cl · z ) (z − cl )
b0 l=1 b0 k−m l=1
H(z) = · k = ·z · k
a0 Y a0 Y
−1
(1 − dl · z ) (z − dl )
l=1 l=1

where the cl are the non-zero positions of zeros (H(cl ) = 0) and the dl
are the non-zero positions of the poles (i.e., z → dl ⇒ |H(z)| → ∞)
of H(z). Except for a constant factor, H(z) is entirely characterized
by the position of these zeros and poles.
As with the Fourier transform, convolution in the time domain corre-
sponds to complex multiplication in the z-domain:
{xn } •−◦ X(z), {yn } •−◦ Y (z) ⇒ {xn } ∗ {yn } •−◦ X(z) · Y (z)
Delaying a sequence by one corresponds in the z-domain to multipli-
cation with z −1 :
{xn−∆n } •−◦ X(z) · z −∆n
97

2
1.75
1.5
1.25
|H(z)|

1
0.75
0.5
0.25
0
1
0.5 1
0 0.5
−0.5 0
−0.5
−1 −1
imaginary real

This example is an amplitude plot of xn 0.8 yn

0.8 0.8z z −1
H(z) = = 0.2
1 − 0.2 · z −1 z − 0.2
yn−1
which features a zero at 0 and a pole at 0.2.
98
 z 1
H(z) = z−0.7
= 1−0.7·z −1
Imaginary Part z Plane Impulse Response
1 1

Amplitude
0 0.5

−1 0
−1 0 1 0 10 20 30
Real Part n (samples)
z 1
H(z) = z−0.9
= 1−0.9·z −1

z Plane Impulse Response

Imaginary Part

1 1
Amplitude

0 0.5

−1 0
−1 0 1 0 10 20 30
Real Part n (samples)
99

z 1
H(z) = z−1
= 1−z −1

z Plane Impulse Response

Imaginary Part

1 1
Amplitude

0 0.5

−1 0
−1 0 1 0 10 20 30
Real Part n (samples)
z 1
H(z) = z−1.1
= 1−1.1·z −1

z Plane Impulse Response

Imaginary Part

1 20
Amplitude

0 10

−1 0
−1 0 1 0 10 20 30
Real Part n (samples)
100
 z2 1
H(z) = (z−0.9·e jπ/6 )·(z−0.9·e− jπ/6 )
= 1−1.8 cos(π/6)z −1 +0.92 ·z −2

z Plane Impulse Response

Imaginary Part
1 2

Amplitude
2
0 0

−1 −2
−1 0 1 0 10 20 30
Real Part n (samples)
z2 1
H(z) = (z−e jπ/6 )·(z−e− jπ/6 )
= 1−2 cos(π/6)z −1 +z −2

z Plane Impulse Response

Imaginary Part

1 5
Amplitude

2
0 0

−1 −5
−1 0 1 0 10 20 30
Real Part n (samples)
101

z2 1 1
H(z) = (z−0.9·e jπ/2 )·(z−0.9·e− jπ/2 )
= 1−1.8 cos(π/2)z −1 +0.92 ·z −2
= 1+0.92 ·z −2

z Plane Impulse Response

Imaginary Part

1 1
Amplitude

2
0 0

−1 −1
−1 0 1 0 10 20 30
Real Part n (samples)
z 1
H(z) = z+1
= 1+z −1

z Plane Impulse Response

Imaginary Part

1 1
Amplitude

0 0

−1 −1
−1 0 1 0 10 20 30
Real Part n (samples)
102
 IIR Filter design techniques
The design of a filter starts with specifying the desired parameters:

→ The passband is the frequency range where we want to approx-

imate a gain of one.

→ The stopband is the frequency range where we want to approx-

imate a gain of zero.

→ The order of a filter is the number of poles it uses in the

z-domain, and equivalently the number of delay elements nec-
essary to implement it.

→ Both passband and stopband will in practice not have gains

of exactly one and zero, respectively, but may show several
deviations from these ideal values, and these ripples may have
a specified maximum quotient between the highest and lowest
gain.
103

→ There will in practice not be an abrupt change of gain between

passband and stopband, but a transition band where the fre-
quency response will gradually change from its passband to its
stopband value.

The designer can then trade off conflicting goals such as a small tran-
sition band, a low order, a low ripple amplitude, or even an absence of
ripples.
Design techniques for making these tradeoffs for analog filters (involv-
ing capacitors, resistors, coils) can also be used to design digital IIR
filters:
Butterworth filters
Have no ripples, gain falls monotonically across the pass and transition
p
band. Within the passband, the gain drops slowly down to 1 − 1/2
(−3 dB). Outside the passband, it drops asymptotically by a factor 2N
per octave (N · 20 dB/decade).

104
Chebyshev type I filters
Distribute the gain error uniformly throughout the passband (equirip-
ples) and drop off monotonically outside.
Chebyshev type II filters
Distribute the gain error uniformly throughout the stopband (equirip-
ples) and drop off monotonically in the passband.
Elliptic filters (Cauer filters)
Distribute the gain error as equiripples both in the passband and stop-
band. This type of filter is optimal in terms of the combination of the
passband-gain tolerance, stopband-gain tolerance, and transition-band
width that can be achieved at a given filter order.

All these filter design techniques are implemented in the MATLAB Signal Processing Toolbox in
the functions butter, cheby1, cheby2, and ellip, which output the coefficients an and bn of the
difference equation that describes the filter. These can be applied with filter to a sequence, or
can be visualized with zplane as poles/zeros in the z-domain, with impz as an impulse response,
and with freqz as an amplitude and phase spectrum. The commands sptool and fdatool
provide interactive GUIs to design digital filters.
105

Butterworth filter design example

Impulse Response
1 0.8
Imaginary Part

0.6
Amplitude

0.5
0 0.4
−0.5 0.2
−1 0
−1 0 1 0 10 20 30
Real Part n (samples)
0 0
Phase (degrees)
Magnitude (dB)

−20
−50
−40

−60 −100
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)
order: 1, cutoff frequency (−3 dB): 0.25 × fs /2
106
 Butterworth filter design example
Impulse Response
1 0.3
Imaginary Part

0.2

Amplitude
0.5
0 0.1
−0.5 0
−1 −0.1
−1 0 1 0 10 20 30
Real Part n (samples)
0 0

Phase (degrees)
Magnitude (dB)

−20 −200

−40 −400

−60 −600
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)
order: 5, cutoff frequency (−3 dB): 0.25 × fs /2
107

Chebyshev type I filter design example

Impulse Response
1 0.6
Imaginary Part

0.4
Amplitude

0.5
0 0.2
−0.5 0
−1 −0.2
−1 0 1 0 10 20 30
Real Part n (samples)
0 0
Phase (degrees)
Magnitude (dB)

−20 −200

−40 −400

−60 −600
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)
order: 5, cutoff frequency: 0.5 × fs /2, pass-band ripple: −3 dB
108
 Chebyshev type II filter design example
Impulse Response
1 0.6
Imaginary Part

0.4

Amplitude
0.5
0 0.2
−0.5 0
−1 −0.2
−1 0 1 0 10 20 30
Real Part n (samples)
0 100

Phase (degrees)
Magnitude (dB)

0
−20
−100
−40
−200

−60 −300
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)
order: 5, cutoff frequency: 0.5 × fs /2, stop-band ripple: −20 dB
109

Elliptic filter design example

Impulse Response
1 0.6
Imaginary Part

0.4
Amplitude

0.5
0 0.2
−0.5 0
−1 −0.2
−1 0 1 0 10 20 30
Real Part n (samples)
0 0
Phase (degrees)
Magnitude (dB)

−100
−20
−200
−40
−300

−60 −400
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample) Normalized Frequency (×π rad/sample)
order: 5, cutoff frequency: 0.5 × fs /2, pass-band ripple: −3 dB, stop-band ripple: −20 dB
110
Exercise 14 Draw the direct form II block diagrams of the causal infinite-
impulse response filters described by the following z-transforms and write
down a formula describing their time-domain impulse responses:
1
(a) H(z) =
1 − 21 z −1
1 −4
′ 1− 44
z
(b) H (z) =
1 − 14 z −1
1 1 1
(c) H ′′ (z) = + z −1 + z −2
2 4 2

Exercise 15 (a) Perform the polynomial division of the rational function

given in exercise 14 (a) until you have found the coefficient of z −5 in the
result.
(b) Perform the polynomial division of the rational function given in exercise
14 (b) until you have found the coefficient of z −10 in the result.
(c) Use its z-transform to show that the filter in exercise 14 (b) has actually
a finite impulse response and draw the corresponding block diagram.

111

Exercise 16 Consider the system h : {xn } → {yn } with yn + yn−1 =

xn − xn−4 .
(a) Draw the direct form I block diagram of a digital filter that realises h.
(b) What is the impulse response of h?
(c) What is the step response of h (i.e., h ∗ u)?
(d) Apply the z-transform to (the impulse response of) h to express it as a
rational function H(z).
(e) Can you eliminate a common factor from numerator and denominator?
What does this mean?
(f) For what values z ∈ C is H(z) = 0?
(g) How many poles does H have in the complex plane?
(h) Write H as a fraction using the position of its poles and zeros and draw
their location in relation to the complex unit circle.
(i) If h is applied to a sound file with a sampling frequency of 8000 Hz,
sine waves of what frequency will be eliminated and sine waves of what
frequency will be quadrupled in their amplitude?

112
 Random sequences and noise
A discrete random sequence {xn } is a sequence of numbers
. . . , x−2 , x−1 , x0 , x1 , x2 , . . .
where each value xn is the outcome of a random variable xn in a
corresponding sequence of random variables
. . . , x−2 , x−1 , x0 , x1 , x2 , . . .
Such a collection of random variables is called a random process. Each
individual random variable xn is characterized by its probability distri-
bution function
Pxn (a) = Prob(xn ≤ a)
and the entire random process is characterized completely by all joint
probability distribution functions
Pxn1 ,...,xnk (a1 , . . . , ak ) = Prob(xn1 ≤ a1 ∧ . . . ∧ xnk ≤ ak )
for all possible sets {xn1 , . . . , xnk }.
113

Two random variables xn and xm are called independent if

Pxn ,xm (a, b) = Pxn (a) · Pxm (b)
and a random process is called stationary if
Pxn1 +l ,...,xnk +l (a1 , . . . , ak ) = Pxn1 ,...,xnk (a1 , . . . , ak )
for all l, that is, if the probability distributions are time invariant.
The derivative pxn (a) = Px′ n (a) is called the probability density func-
tion, and helps us to define quantities such as the
→ expected value E(xn) = apxn (a) da
R

→ mean-square value (average power) E(|xn|2) = |a|2pxn (a) da

R

→ variance Var(xn ) = E[|xn − E(xn )|2 ] = E(|xn |2 ) − |E(xn )|2

→ correlation Cor(xn , xm ) = E(xn · x∗m )
Remember that E(·) is linear, that is E(ax) = aE(x) and E(x + y) = E(x) + E(y). Also,
Var(ax) = a2 Var(x) and, if x and y are independent, Var(x + y) = Var(x) + Var(y).
114
A stationary random process {xn } can be characterized by its mean
value
mx = E(xn ),
its variance
σx2 = E(|xn − mx |2 ) = γxx (0)
(σx is also called standard deviation), its autocorrelation sequence
φxx (k) = E(xn+k · x∗n )
and its autocovariance sequence
γxx (k) = E[(xn+k − mx ) · (xn − mx )∗ ] = φxx (k) − |mx |2
A pair of stationary random processes {xn } and {yn } can, in addition,
be characterized by its crosscorrelation sequence
φxy (k) = E(xn+k · yn∗ )
and its crosscovariance sequence
γxy (k) = E[(xn+k − mx ) · (yn − my )∗ ] = φxy (k) − mx m∗y
115

Deterministic crosscorrelation sequence

For deterministic sequences {xn } and {yn }, the crosscorrelation sequence
is
∞
X
cxy (k) = xi+k yi .
i=−∞
After dividing through the overlapping length of the finite sequences involved, cxy (k) can be
used to estimate, from a finite sample of a stationary random sequence, the underlying φxy (k).
MATLAB’s xcorr function does that with option unbiased.
If {xn } is similar to {yn }, but lags l elements behind (xn ≈ yn−l ), then
cxy (l) will be a peak in the crosscorrelation sequence. It is therefore widely
calculated to locate shifted versions of a known sequence in another one.
The deterministic crosscorrelation sequence is a close cousin of the convo-
lution, with just the second input sequence mirrored:
{cxy (n)} = {xn } ∗ {y−n }
It can therefore be calculated equally easily via the Fourier transform:
Cxy (f ) = X(f ) · Y ∗ (f )
Swapping the input sequences mirrors the output sequence: cxy (k) = cyx (−k).
116
Equivalently, we define the deterministic autocorrelation sequence in
the time domain as
∞
X
cxx (k) = xi+k xi .
i=−∞

which corresponds in the frequency domain to

Cxx (f ) = X(f ) · X ∗ (f ) = |X(f )|2 .

In other words, the Fourier transform Cxx (f ) of the autocorrelation

sequence {cxx (n)} of a sequence {xn } is identical to the squared am-
plitudes of the Fourier transform, or power spectrum, of {xn }.
This suggests, that the Fourier transform of the autocorrelation se-
quence of a random process might be a suitable way for defining the
power spectrum of that random process.
What can we say about the phase in the Fourier spectrum of a time-invariant random process?

117

Filtered random sequences

Let {xn } be a random sequence from a stationary random process.
The output
∞
X ∞
X
yn = hk · xn−k = hn−k · xk
k=−∞ k=−∞

of an LTI applied to it will then be another random sequence, charac-

terized by
X∞
my = mx hk
k=−∞

and
∞
X φxx (k) = E(xn+k · x∗n )
φyy (k) = φxx (k−i)chh (i), where P∞
i=−∞ chh (k) = i=−∞ hi+k hi .

118
In other words:
{φyy (n)} = {chh (n)} ∗ {φxx (n)}
{yn } = {hn } ∗ {xn } ⇒
Φyy (f ) = |H(f )|2 · Φxx (f )

Similarly:
{φyx (n)} = {hn } ∗ {φxx (n)}
{yn } = {hn } ∗ {xn } ⇒
Φyx (f ) = H(f ) · Φxx (f )

White noise
A random sequence {xn } is a white noise signal, if mx = 0 and

φxx (k) = σx2 δk .

The power spectrum of a white noise signal is flat:

Φxx (f ) = σx2 .
119

Application example:
Where an LTI {yn } = {hn } ∗ {xn } can be observed to operate on
white noise {xn } with φxx (k) = σx2 δk , the crosscorrelation between
input and output will reveal the impulse response of the system:

φyx (k) = σx2 · hk

where φyx (k) = φxy (−k) = E(yn+k · x∗n ).

120
 DFT averaging

The above diagrams show different types of spectral estimates of a sequence

xi = sin(2πj × 8/64) + sin(2πj × 14.32/64) + ni with φnn (i) = 4δi .
Left is a single 64-element DFT of {xi } (with rectangular window). The
flat spectrum of white noise is only an expected value. In a single discrete
Fourier transform of such a sequence, the significant variance of the noise
spectrum becomes visible. It almost drowns the two peaks from sine waves.
After cutting {xi } into 1000 windows of 64 elements each, calculating their
DFT, and plotting the average of their absolute values, the centre figure
shows an approximation of the expected value of the amplitude spectrum,
with a flat noise floor. Taking the absolute value before spectral averaging
is called incoherent averaging, as the phase information is thrown away.

121

The rightmost figure was generated from the same set of 1000 windows,
but this time the complex values of the DFTs were averaged before the
absolute value was taken. This is called coherent averaging and, because
of the linearity of the DFT, identical to first averaging the 1000 windows
and then applying a single DFT and taking its absolute value. The windows
start 64 samples apart. Only periodic waveforms with a period that divides
64 are not averaged away. This periodic averaging step suppresses both the
noise and the second sine wave.

Periodic averaging
If a zero-mean signal {xi } has a periodic component with period p, the
periodic component can be isolated by periodic averaging :
k
1 X
x̄i = lim xi+pn
k→∞ 2k + 1
n=−k

Periodic averaging
P corresponds in the time domain to convolution with a
Dirac comb n δi−pn . In the frequency domain, this means multiplication
with a Dirac comb that eliminates all frequencies but multiples of 1/p.
122
 Image, video and audio compression
Structure of modern audiovisual communication systems:

- sensor + - perceptual - entropy channel

signal sampling
-
coding
coding coding

?
noise - channel

?
human  perceptual  entropy  channel
senses
 display decoding
decoding decoding

123

Audio-visual lossy coding today typically consists of these steps:

→ A transducer converts the original stimulus into a voltage.
→ This analog signal is then sampled and quantized.
The digitization parameters (sampling frequency, quantization levels) are preferably
chosen generously beyond the ability of human senses or output devices.

→ The digitized sensor-domain signal is then transformed into a

perceptual domain.
This step often mimics some of the first neural processing steps in humans.

→ This signal is quantized again, based on a perceptual model of what

level of quantization-noise humans can still sense.
→ The resulting quantized levels may still be highly statistically de-
pendent. A prediction or decorrelation transform exploits this and
produces a less dependent symbol sequence of lower entropy.
→ An entropy coder turns that into an apparently-random bit string,
whose length approximates the remaining entropy.
The first neural processing steps in humans are in effect often a kind of decorrelation transform;
our eyes and ears were optimized like any other AV communications system. This allows us to
use the same transform for decorrelating and transforming into a perceptually relevant domain.
124
 Outline of the remaining lectures
→ Quick review of entropy coding

→ Transform coding: techniques for converting sequences of highly-

dependent symbols into less-dependent lower-entropy sequences.
• run-length coding
• decorrelation, Karhunen-Loève transform (PCA)
• other orthogonal transforms (especially DCT)

→ Introduction to some characteristics and limits of human senses

• perceptual scales and sensitivity limits
• colour vision
• human hearing limits, critical bands, audio masking

→ Quantization techniques to remove information that is irrelevant to

human senses
125

→ Image and audio coding standards

• A/µ-law coding (digital telephone network)
• JPEG
• MPEG video
• MPEG audio

Literature
→ D. Salomon: A guide to data compression methods.
ISBN 0387952608, 2002.

→ L. Gulick, G. Gescheider, R. Frisina: Hearing. ISBN 0195043073,

1989.

→ H. Schiffman: Sensation and perception. ISBN 0471082082, 1982.

126
 Entropy coding review – Huffman
X 1
Entropy: H = p(α) · log2
1.00 p(α)
α∈A
0 1 = 2.3016 bit

0.40 0.60

0 1 0 1
v w 0.25
0.20 0.20 u
0.35 0 1

x
0.10
Mean codeword length: 2.35 bit 0.15

0 1
Huffman’s algorithm constructs an optimal code-word tree for a set of
symbols with known probability distribution. It iteratively picks the two y z
elements of the set with the smallest probability and combines them into 0.05 0.05
a tree by adding a common root. The resulting tree goes back into the
set, labeled with the sum of the probabilities of the elements it combines.
The algorithm terminates when less than two elements are left.
127

Entropy coding review – arithmetic coding

Partition [0,1] according 0.0 0.35 0.55 0.75 0.9 0.95 1.0
to symbol probabilities: u v w x y z
Encode text wuvw . . . as numeric value (0.58. . . ) in nested intervals:
1.0 0.75 0.62 0.5885 0.5850
z z z z z
y y y y y

x x x x x

w w w w w

v v v v v

u u u u u

0.55
0.0 0.55 0.5745 0.5822

128
 Arithmetic coding
Several advantages:

→ Length of output bitstring can approximate the theoretical in-

formation content of the input to within 1 bit.

→ Performs well with probabilities > 0.5, where the information

per symbol is less than one bit.

→ Interval arithmetic makes it easy to change symbol probabilities

(no need to modify code-word tree) ⇒ convenient for adaptive
coding

Can be implemented efficiently with fixed-length arithmetic by rounding

probabilities and shifting out leading digits as soon as leading zeros
appear in interval size. Usually combined with adaptive probability
estimation.
Huffman coding remains popular because of its simplicity and lack of patent-licence issues.
129

Coding of sources with memory and

correlated symbols
Run-length coding:

↓
5 7 12 3 3

Predictive coding:
encoder decoder
f(t) g(t) g(t) f(t)
− +

predictor predictor
P(f(t−1), f(t−2), ...) P(f(t−1), f(t−2), ...)

Delta coding (DPCM): P (x) = x

n
X
Linear predictive coding: P (x1 , . . . , xn ) = ai x i
i=1
130
 Old (Group 3 MH) fax code
pixels white code black code
• Run-length encoding plus modified Huffman 0 00110101 0000110111
code 1 000111 010
• Fixed code table (from eight sample pages) 2 0111 11
3 1000 10
• separate codes for runs of white and black
4 1011 011
pixels
5 1100 0011
• termination code in the range 0–63 switches 6 1110 0010
between black and white code 7 1111 00011
• makeup code can extend length of a run by 8 10011 000101
a multiple of 64 9 10100 000100
• termination run length 0 needed where run 10 00111 0000100
length is a multiple of 64 11 01000 0000101
12 001000 0000111
• single white column added on left side be- 13 000011 00000100
fore transmission 14 110100 00000111
• makeup codes above 1728 equal for black 15 110101 000011000
and white 16 101010 0000010111
• 12-bit end-of-line marker: 000000000001 ... ... ...
(can be prefixed by up to seven zero-bits 63 00110100 000001100111
to reach next byte boundary) 64 11011 0000001111
128 10010 000011001000
Example: line with 2 w, 4 b, 200 w, 3 b, EOL → 192 010111 000011001001
1000|011|010111|10011|10|000000000001 ... ... ...
1728 010011011 0000001100101
131

Modern (JBIG) fax code

Performs context-sensitive arithmetic coding of binary pixels. Both encoder
and decoder maintain statistics on how the black/white probability of each
pixel depends on these 10 previously transmitted neighbours:

Based on the counted numbers nblack and nwhite of how often each pixel
value has been encountered so far in each of the 1024 contexts, the proba-
bility for the next pixel being black is estimated as
nblack + 1
pblack =
nwhite + nblack + 2

The encoder updates its estimate only after the newly counted pixel has
been encoded, such that the decoder knows the exact same statistics.
Joint Bi-level Expert Group: International Standard ISO 11544, 1993.
Example implementation: http://www.cl.cam.ac.uk/~mgk25/jbigkit/
132
 Statistical dependence
Random variables X, Y are dependent iff ∃x, y:
P (X = x ∧ Y = y) 6= P (X = x) · P (Y = y).

If X, Y are dependent, then

⇒ ∃x, y : P (X = x | Y = y) 6= P (X = x) ∨
P (Y = y | X = x) 6= P (Y = y)
⇒ H(X|Y ) < H(X) ∨ H(Y |X) < H(Y )

Application
Where x is the value of the next symbol to be transmitted and y is
the vector of all symbols transmitted so far, accurate knowledge of the
conditional probability P (X = x | Y = y) will allow a transmitter to
remove all redundancy.
An application example of this approach is JBIG, but there y is limited
to 10 past single-bit pixels and P (X = x | Y = y) is only an estimate.
133

Practical limits of measuring conditional probabilities

The practical estimation of conditional probabilities, in their most gen-
eral form, based on statistical measurements of example signals, quickly
reaches practical limits. JBIG needs an array of only 211 = 2048 count-
ing registers to maintain estimator statistics for its 10-bit context.
If we wanted to encode each 24-bit pixel of a colour image based on
its statistical dependence of the full colour information from just ten
previous neighbour pixels, the required number of
(224 )11 ≈ 3 × 1080
registers for storing each probability will exceed the estimated number
of particles in this universe. (Neither will we encounter enough pixels
to record statistically significant occurrences in all (224 )10 contexts.)
This example is far from excessive. It is easy to show that in colour im-
ages, pixel values show statistical significant dependence across colour
channels, and across locations more than eight pixels apart.
A simpler approximation of dependence is needed: correlation.
134
 Correlation
Two random variables X ∈ R and Y ∈ R are correlated iff

E{[X − E(X)] · [Y − E(Y )]} 6= 0

where E(· · ·) denotes the expected value of a random-variable term.

Correlation implies dependence, but de- Dependent but not correlated:
pendence does not always lead to corre- 1

lation (see example to the right).

However, most dependency in audiovi-
0
sual data is a consequence of correlation,
which is algorithmically much easier to
exploit.
−1
−1 0 1

Positive correlation: higher X ⇔ higher Y , lower X ⇔ lower Y

Negative correlation: lower X ⇔ higher Y , higher X ⇔ lower Y
135

Correlation of neighbour pixels

Values of neighbour pixels at distance 1 Values of neighbour pixels at distance 2
256 256

192 192

128 128

64 64

0 0
0 64 128 192 256 0 64 128 192 256
Values of neighbour pixels at distance 4 Values of neighbour pixels at distance 8
256 256

192 192

128 128

64 64

0 0
0 64 128 192 256 0 64 128 192 256

136
 Covariance and correlation
We define the covariance of two random variables X and Y as

Cov(X, Y ) = E{[X −E(X)]·[Y −E(Y )]} = E(X ·Y )−E(X)·E(Y )

and the variance as Var(X) = Cov(X, X) = E{[X − E(X)]2 }.

The Pearson correlation coefficient

Cov(X, Y )
ρX,Y = p
Var(X) · Var(Y )

is a normalized form of the covariance. It is limited to the range [−1, 1].

If the correlation coefficient has one of the values ρX,Y = ±1, this
implies that X and Y are exactly linearly dependent, i.e. Y = aX + b,
with a = Cov(X, Y )/ Var(X) and b = E(Y ) − E(X).

137

Covariance Matrix
For a random vector X = (X1 , X2 , . . . , Xn ) ∈ Rn we define the co-
variance matrix

Cov(X) = E (X − E(X)) · (X − E(X))T = (Cov(Xi , Xj ))i,j =

 
Cov(X1 , X1 ) Cov(X1 , X2 ) Cov(X1 , X3 ) · · · Cov(X1 , Xn )
 Cov(X2 , X1 ) Cov(X2 , X2 ) Cov(X2 , X3 ) · · · Cov(X2 , Xn ) 
 
 Cov(X3 , X1 ) Cov(X3 , X2 ) Cov(X3 , X3 ) · · · Cov(X3 , Xn ) 
.. .. .. .
 
 . .. .. 
 . . . 
Cov(Xn , X1 ) Cov(Xn , X2 ) Cov(Xn , X3 ) · · · Cov(Xn , Xn )

The elements of a random vector X are uncorrelated if and only if

Cov(X) is a diagonal matrix.
Cov(X, Y ) = Cov(Y, X), so all covariance matrices are symmetric:
Cov(X) = CovT (X).

138
 Decorrelation by coordinate transform
Neighbour−pixel value pairs Decorrelated neighbour−pixel value pairs
256 320

256
192
192

128 128

64
64
0

0 −64
0 64 128 192 256 −64 0 64 128 192 256 320
Probability distribution and entropy

correlated value pair (H = 13.90 bit)

Idea: Take the values of a group of cor-
decorrelated value 1 (H = 7.12 bit) related symbols (e.g., neighbour pixels) as
decorrelated value 2 (H = 4.75 bit)
a random vector. Find a coordinate trans-
form (multiplication with an orthonormal
matrix) that leads to a new random vector
whose covariance matrix is diagonal. The
vector components in this transformed co-
ordinate system will no longer be corre-
lated. This will hopefully reduce the en-
tropy of some of these components.
−64 0 64 128 192 256 320

139

Theorem: Let X ∈ Rn and Y ∈ Rn be random vectors that are

linearly dependent with Y = AX + b, where A ∈ Rn×n and b ∈ Rn
are constants. Then

E(Y) = A · E(X) + b
Cov(Y) = A · Cov(X) · AT

Proof: The first equation follows from the linearity of the expected-
value operator E(·), as does E(A · X · B) = A · E(X) · B for matrices
A, B. With that, we can transform

Cov(Y) = E (Y − E(Y)) · (Y − E(Y))T

T


= E (AX − AE(X)) · (AX − AE(X))
= E A(X − E(X)) · (X − E(X))T AT

= A · E (X − E(X)) · (X − E(X))T · AT

= A · Cov(X) · AT

140
Quick review: eigenvectors and eigenvalues
We are given a square matrix A ∈ Rn×n . The vector x ∈ Rn is an
eigenvector of A if there exists a scalar value λ ∈ R such that

Ax = λx.

The corresponding λ is the eigenvalue of A associated with x.

The length of an eigenvector is irrelevant, as any multiple of it is also
an eigenvector. Eigenvectors are in practice normalized to length 1.
Spectral decomposition
Any real, symmetric matrix A = AT ∈ Rn×n can be diagonalized into
the form
A = U ΛU T ,
where Λ = diag(λ1 , λ2 , . . . , λn ) is the diagonal matrix of the ordered
eigenvalues of A (with λ1 ≥ λ2 ≥ · · · ≥ λn ), and the columns of U
are the n corresponding orthonormal eigenvectors of A.
141

Karhunen-Loève transform (KLT)

We are given a random vector variable X ∈ Rn . The correlation of the
elements of X is described by the covariance matrix Cov(X).
How can we find a transform matrix A that decorrelates X, i.e. that
turns Cov(AX) = A · Cov(X) · AT into a diagonal matrix? A would
provide us the transformed representation Y = AX of our random
vector, in which all elements are mutually uncorrelated.
Note that Cov(X) is symmetric. It therefore has n real eigenvalues
λ1 ≥ λ2 ≥ · · · ≥ λn and a set of associated mutually orthogonal
eigenvectors b1 , b2 , . . . , bn of length 1 with
Cov(X)bi = λi bi .

We convert this set of equations into matrix notation using the matrix
B = (b1 , b2 , . . . , bn ) that has these eigenvectors as columns and the
diagonal matrix D = diag(λ1 , λ2 , . . . , λn ) that consists of the corre-
sponding eigenvalues:
Cov(X)B = BD
142
B is orthonormal, that is BB T = I.
Multiplying the above from the right with B T leads to the spectral
decomposition
Cov(X) = BDB T
of the covariance matrix. Similarly multiplying instead from the left
with B T leads to
B T Cov(X)B = D
and therefore shows with
Cov(B T X) = D
that the eigenvector matrix B T is the wanted transform.
The Karhunen-Loève transform (also known as Hotelling transform
or Principal Component Analysis) is the multiplication of a correlated
random vector X with the orthonormal eigenvector matrix B T from the
spectral decomposition Cov(X) = BDB T of its covariance matrix.
This leads to a decorrelated random vector B T X whose covariance
matrix is diagonal.
143

Karhunen-Loève transform example I

colour image red channel green channel blue channel

The colour image (left) has m = r2 pixels, each We can now define the mean colour vector
of which is an n = 3-dimensional RGB vector 0 1
T m 0.4839
Ix,y = (rx,y , gx,y , bx,y ) 1 X
S̄c = Sc,i , S̄ = @ 0.4456 A
m i=1 0.3411
The three rightmost images show each of these
colour planes separately as a black/white im-
age. and the covariance matrix
We want to apply the KLT on a set of such m
Rn colour vectors. Therefore, we reformat the 1 X
Cc,d = (Sc,i − S̄c )(Sd,i − S̄d )
image I into an n × m matrix of the form m − 1 i=1
0 1 0 1
r1,1 r1,2 r1,3 · · · rr,r 0.0328 0.0256 0.0160
S = @ g1,1 g1,2 g1,3 · · · gr,r A C = @ 0.0256 0.0216 0.0140 A
b1,1 b1,2 b1,3 · · · br,r 0.0160 0.0140 0.0109
144
[When estimating a covariance from a number of samples, the sum is divided by the number of
samples minus one. This takes into account the variance of the mean S̄c , which is not the exact
expected value, but only an estimate of it.]
The resulting covariance matrix has three eigenvalues 0.0622, 0.0025, and 0.0006
0 10 1 0 1
0.0328 0.0256 0.0160 0.7167 0.7167
@ 0.0256 0.0216 0.0140 A @ 0.5833 A = 0.0622 @ 0.5833 A
0.0160 0.0140 0.0109 0.3822 0.3822
0 10 1 0 1
0.0328 0.0256 0.0160 −0.5509 −0.5509
@ 0.0256 0.0216 0.0140 A @ 0.1373 A = 0.0025 @ 0.1373 A
0.0160 0.0140 0.0109 0.8232 0.8232
0 10 1 0 1
0.0328 0.0256 0.0160 −0.4277 −0.4277
@ 0.0256 0.0216 0.0140 A @ 0.8005 A = 0.0006 @ 0.8005 A
0.0160 0.0140 0.0109 −0.4198 −0.4198

and can therefore be diagonalized as

0 1
0.0328 0.0256 0.0160
@ 0.0256 0.0216 0.0140 A = C = U · D · U T =
0.0160 0.0140 0.0109
0 10 10 1
0.7167 −0.5509 −0.4277 0.0622 0 0 0.7167 0.5833 0.3822
@ 0.5833 0.1373 0.8005 A @ 0 0.0025 0 A @ −0.5509 0.1373 0.8232 A
0.3822 0.8232 −0.4198 0 0 0.0006 −0.4277 0.8005 −0.4198

(e.g. using MATLAB’s singular-value decomposition function svd).

145

Karhunen-Loève transform example I

Before KLT: We finally apply the orthogonal 3 × 3 transform
matrix U , which we just used to diagonalize the
covariance matrix, to the entire image:
2 0 13
S̄1 S̄1 · · · S̄1
T = U T · 4S − @ S̄2 S̄2 · · · S̄2 A5
S̄3 S̄3 · · · S̄3
red green blue 0 1
S̄1 S̄1 · · · S̄1
After KLT: + @ S̄2 S̄2 · · · S̄2 A
S̄3 S̄3 · · · S̄3

The resulting transformed image

0 1
u1,1 u1,2 u1,3 · · · ur,r
u v w T = @ v1,1 v1,2 v1,3 · · · vr,r A
Projections on eigenvector subspaces: w1,1 w1,2 w1,3 · · · wr,r

consists of three new “colour” planes whose

pixel values have no longer any correlation to
the pixels at the same coordinates in another
plane. [The bear disappeared from the last of
these (w), which represents mostly some of the
v=w=0 w=0 original green grass in the background.]
146
 Spatial correlation
The previous example used the Karhunen-Loève transform in order to
eliminate correlation between colour planes. While this is of some
relevance for image compression, far more correlation can be found
between neighbour pixels within each colour plane.
In order to exploit such correlation using the KLT, the sample set has
to be extended from individual pixels to entire images. The underlying
calculation is the same as in the preceeding example, but this time
the columns of S are entire (monochrome) images. The rows are the
different images found in the set of test images that we use to examine
typical correlations between neighbour pixels.
In other words, we use the same formulas as in the previous example, but this time n is the number
of pixels per image and m is the number of sample images. The Karhunen-Loève transform is
here no longer a rotation in a 3-dimensional colour space, but it operates now in a much larger
vector space that has as many dimensions as an image has pixels.
To keep things simple, we look in the next experiment only at m = 9000 1-dimensional “images”
with n = 32 pixels each. As a further simplification, we use not real images, but random noise
that was filtered such that its amplitude spectrum is proportional to 1/f , where f is the frequency.
The result would be similar in a sufficiently large collection of real test images.

147

Karhunen-Loève transform example II

Matrix columns of S filled with samples of 1/f filtered noise

...
Covariance matrix C Matrix U with eigenvector columns

148
 Matrix U ′ with normalised KLT Matrix with Discrete Cosine
eigenvector columns Transform base vector columns

Breakthrough: Ahmed/Natarajan/Rao discovered the DCT as an ex-

cellent approximation of the KLT for typical photographic images, but
far more efficient to calculate.
Ahmed, Natarajan, Rao: Discrete Cosine Transform. IEEE Transactions on Computers, Vol. 23,
January 1974, pp. 90–93.
149

Discrete cosine transform (DCT)

The forward and inverse discrete cosine transform
N −1
C(u) X (2x + 1)uπ
S(u) = p s(x) cos
N/2 x=0 2N
N −1
X C(u) (2x + 1)uπ
s(x) = p S(u) cos
u=0 N/2 2N

with
√1

2
u=0
C(u) =
1 u>0
is an orthonormal transform:
N −1
(2x + 1)uπ C(u′ ) (2x + 1)u′ π 1 u = u′

X C(u)
cos ·p cos =
0 u=6 u′
p
x=0 N/2 2N N/2 2N

150
The 2-dimensional variant of the DCT applies the 1-D transform on
both rows and columns of an image:

C(u) C(v)
S(u, v) = p p ·
N/2 N/2
N −1 N −1
X X (2x + 1)uπ (2y + 1)vπ
s(x, y) cos cos
x=0 y=0
2N 2N

s(x, y) =
N −1 N −1
X X C(u) C(v) (2x + 1)uπ (2y + 1)vπ
p p · S(u, v) cos cos
u=0 v=0 N/2 N/2 2N 2N

A range of fast algorithms have been found for calculating 1-D and
2-D DCTs (e.g., Ligtenberg/Vetterli).

151

Whole-image DCT

2D Discrete Cosine Transform (log10) Original image

−1

−2

−3

−4

152
Whole-image DCT, 80% coefficient cutoff

80% truncated 2D DCT (log10) 80% truncated DCT: reconstructed image

−1

−2

−3

−4

153

Whole-image DCT, 90% coefficient cutoff

90% truncated 2D DCT (log10) 90% truncated DCT: reconstructed image

−1

−2

−3

−4

154
Whole-image DCT, 95% coefficient cutoff

95% truncated 2D DCT (log10) 95% truncated DCT: reconstructed image

−1

−2

−3

−4

155

Whole-image DCT, 99% coefficient cutoff

99% truncated 2D DCT (log10) 99% truncated DCT: reconstructed image

−1

−2

−3

−4

156
 Base vectors of 8×8 DCT
v
0 1 2 3 4 5 6 7
0

3
u

157

Discrete Wavelet Transform

158
The n-point Discrete Fourier Transform (DFT) can be viewed as a device that
sends an input signal through a bank of n non-overlapping band-pass filters, each
reducing the bandwidth of the signal to 1/n of its original bandwidth.
According to the sampling theorem, after a reduction of the bandwidth by 1/n,
the number of samples needed to reconstruct the original signal can equally be
reduced by 1/n. The DFT splits a wide-band signal represented by n input signals
into n separate narrow-band samples, each represented by a single sample.
A Discrete Wavelet Transform (DWT) can equally be viewed as such a frequency-
band splitting device. However, with the DWT, the bandwidth of each output signal
is proportional to the highest input frequency that it contains. High-frequency
components are represented in output signals with a high bandwidth, and therefore
a large number of samples. Low-frequency signals end up in output signals with
low bandwidth, and are correspondingly represented with a low number of samples.
As a result, high-frequency information is preserved with higher spatial resolution
than low-frequency information.
Both the DFT and the DWT are linear orthogonal transforms that preserve all
input information in their output without adding anything redundant.
As with the DFT, the 1-dimensional DWT can be extended to 2-D images by trans-
forming both rows and columns (the order of which happens first is not relevant).

159

A DWT is defined by a combination of a low-pass filter, which smoothes

a signal by allowing only the bottom half of all frequencies to pass
through, and a high-pass filter, which preserves only the upper half of
the spectrum. These two filters must be chosen to be “orthogonal”
to each other, in the sense that together they split up the information
content of their input signal without any mutual information in their
outputs.
A widely used 1-D filter pair is DAUB4 (by Ingrid Daubechies). The
low-pass filter convolves a signal with the 4-point sequence c0 , c1 , c2 , c3 ,
and the matching high-pass filter convolves with c3 , −c2 , c1 , −c0 . Writ-
ten as a transformation matrix, DAUB4 has the form
c0 c1 c2 c3
0 1
B c3 −c2 c1 −c0 C
c0 c1 c2 c3
B C
B C
c3 −c2 c1 −c0
B C
B C
B . ..
B C
B .. .. C
B . . C
C
B
B c0 c1 c2 c3 C
C
B
B c3 −c2 c1 −c0 C
C
@ c c3 c0 c1 A
2
c1 −c0 c3 −c2
160
An orthogonal matrix multiplied with itself transposed is the identity
matrix, which is fulfilled for the above one when

c20 + c21 + c22 + c23 = 1

c2 c0 + c3 c1 = 0

To determine four unknown variables we need four equations, there-

fore we demand that the high-pass filter will not pass through any
information about polynomials of degree 1:

c3 − c2 + c1 − c0 = 0
0c3 − 1c2 + 2c1 − 3c0 = 0

This leads to the solution

√ √ √ √
c0 = (1 + 3)/(4 2), c1 = (3 + 3)/(4 2)
√ √ √ √
c2 = (3 − 3)/(4 2), c3 = (1 − 3)/(4 2)

Daubechies tabulated also similar filters with more coefficients.

161

In an n-point DWT, an input vector is convolved separately with a low-

pass and a high-pass filter. The result are two output sequences of n
numbers. But as each sequence has now only half the input bandwidth,
each second value is redundant, can be reconstructed by interpolation
with the same filter, and can therefore be discarded.
The remaining output values of the high-pass filter (“detail”) are part
of the final output of the DWT. The remaining values of the low-pass
filter (“approximation”) are recursively treated the same way, until they
consist – after log2 n steps – of only a single value, namely the average
of the entire input.
Like with the DFT and DCT, for many real-world input signals, the
DWT accumulates most energy into only a fraction of its output values.
A commonly used approach for wavelet-based compression of signals is
to replace all coefficients below an adjustable threshold with zero and
encode only the values and positions of the remaining ones.

162
 Discrete Wavelet Transform compression
80% truncated 2D DAUB8 DWT 90% truncated 2D DAUB8 DWT

95% truncated 2D DAUB8 DWT 99% truncated 2D DAUB8 DWT

163

Psychophysics of perception
Sensation limit (SL) = lowest intensity stimulus that can still be perceived
Difference limit (DL) = smallest perceivable stimulus difference at given
intensity level
Weber’s law
Difference limit ∆φ is proportional to the intensity φ of the stimu-
lus (except for a small correction constant a, to describe deviation of
experimental results near SL):
∆φ = c · (φ + a)

Fechner’s scale
Define a perception intensity scale ψ using the sensation limit φ0 as
the origin and the respective difference limit ∆φ = c · φ as a unit step.
The result is a logarithmic relationship between stimulus intensity and
scale value:
φ
ψ = logc
φ0
164
Fechner’s scale matches older subjective intensity scales that follow
differentiability of stimuli, e.g. the astronomical magnitude numbers
for star brightness introduced by Hipparchos (≈150 BC).

Stevens’ law
A sound that is 20 DL over SL is perceived as more than twice as loud
as one that is 10 DL over SL, i.e. Fechner’s scale does not describe
well perceived intensity. A rational scale attempts to reflect subjective
relations perceived between different values of stimulus intensity φ.
Stevens observed that such rational scales ψ follow a power law:

ψ = k · (φ − φ0 )a

Example coefficients a: temperature 1.6, weight 1.45, loudness 0.6,

brightness 0.33.

165

Decibel
Communications engineers often use logarithmic units:
→ Quantities often vary over many orders of magnitude → difficult
to agree on a common SI prefix
→ Quotient of quantities (amplification/attenuation) usually more
interesting than difference
→ Signal strength usefully expressed as field quantity (voltage,
current, pressure, etc.) or power, but quadratic relationship
between these two (P = U 2 /R = I 2 R) rather inconvenient
→ Weber/Fechner: perception is logarithmic
Plus: Using magic special-purpose units has its own odd attractions (→ typographers, navigators)

Neper (Np) denotes the natural logarithm of the quotient of a field

quantity F and a reference value F0 .
Bel (B) denotes the base-10 logarithm of the quotient of a power P
and a reference power P0 . Common prefix: 10 decibel (dB) = 1 bel.
166
Where P is some power and P0 a 0 dB reference power, or equally
where F is a field quantity and F0 the corresponding reference level:
P F
10 dB · log10 = 20 dB · log10
P0 F0
Common reference values are indicated with an additional letter after
the “dB”:
0 dBW = 1W
0 dBm = 1 mW = −30 dBW
0 dBµV = 1 µV
0 dBSPL = 20 µPa (sound pressure level)
0 dBSL = perception threshold (sensation limit)

3 dB = double power, 6 dB = double pressure/voltage/etc.

10 dB = 10× power, 20 dB = 10× pressure/voltage/etc.
W.H. Martin: Decibel – the new name for the transmission unit. Bell System Technical Journal,
January 1929.
167

RGB video colour coordinates

Hardware interface (VGA): red, green, blue signals with 0–0.7 V
Electron-beam current and photon count of cathode-ray displays are
roughly proportional to (v − v0 )γ , where v is the video-interface or
control-grid voltage and γ is a device parameter that is typically in
the range 1.5–3.0. In broadcast TV, this CRT non-linearity is com-
pensated electronically in TV cameras. A welcome side effect is that
it approximates Stevens’ scale and therefore helps to reduce perceived
noise.
Software interfaces map RGB voltage linearly to {0, 1, . . . , 255} or 0–1.
How numeric RGB values map to colour and luminosity depends at
present still highly on the hardware and sometimes even on the oper-
ating system or device driver.
The new specification “sRGB” aims to standardize the meaning of
an RGB value with the parameter γ = 2.2 and with standard colour
coordinates of the three primary colours.
http://www.w3.org/Graphics/Color/sRGB, IEC 61966
168
 YUV video colour coordinates

The human eye processes colour and luminosity at different resolutions.

To exploit this phenomenon, many image transmission systems use a
colour space with a luminance coordinate
Y = 0.3R + 0.6G + 0.1B
and colour (“chrominance”) components
V = R − Y = 0.7R − 0.6G − 0.1B
U = B − Y = −0.3R − 0.6G + 0.9B
169

YUV transform example

original Y channel U and V channels

The centre image shows only the luminance channel as a black/white
image. In the right image, the luminance channel (Y) was replaced
with a constant, such that only the chrominance information remains.
This example and the next make only sense when viewed in colour. On a black/white printout of
this slide, only the Y channel information will be present.
170
 Y versus UV sensitivity example

original blurred U and V blurred Y channel

In the centre image, the chrominance channels have been severely low-
pass filtered (Gaussian impulse response ). But the human eye
perceives this distortion as far less severe than if the exact same filtering
is applied to the luminance channel (right image).
171

YCrCb video colour coordinates

Since −0.7 ≤ V ≤ 0.7 and −0.9 ≤ U ≤ 0.9, a more convenient
normalized encoding of chrominance is:
Y=0.1 Y=0.3 Y=0.5
1 1 1

0.8 0.8 0.8

U
0.6 0.6 0.6
Cr

Cr

Cr

Cb = + 0.5
0.4 0.4 0.4

0.2 0.2 0.2

2.0 0
0 0.5 1
0
0 0.5 1
0
0 0.5 1
Cb Cb Cb
Y=0.7 Y=0.9 Y=0.99
1 1 1

V 0.8 0.8 0.8

Cr = + 0.5 0.6 0.6 0.6

1.6
Cr

Cr

Cr

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
Cb Cb Cb

Modern image compression techniques operate on Y , Cr, Cb channels

separately, using half the resolution of Y for storing Cr, Cb.
Some digital-television engineering terminology:
If each pixel is represented by its own Y , Cr and Cb byte, this is called a “4:4:4” format. In the
compacter “4:2:2” format, a Cr and Cb value is transmitted only for every second pixel, reducing
the horizontal chrominance resolution by a factor two. The “4:2:0” format transmits in alternat-
ing lines either Cr or Cb for every second pixel, thus halving the chrominance resolution both
horizontally and vertically. The “4:1:1” format reduces the chrominance resolution horizontally
by a quarter and “4:1:0” does so in both directions. [ITU-R BT.601]
172
 The human auditory system

→ frequency range 20–16000 Hz (babies: 20 kHz)

→ sound pressure range 0–140 dBSPL (about 10−5 –102 pascal)

→ mechanical filter bank (cochlea) splits input into frequency

components, physiological equivalent of Fourier transform

→ most signal processing happens in the frequency domain where

phase information is lost

→ some time-domain processing below 500 Hz and for directional

hearing

→ sensitivity and difference limit are frequency dependent

173

Equiloudness curves and the unit “phon”

Each curve represents a loudness level in phon. At 1 kHz, the loudness unit
phon is identical to dBSPL and 0 phon is the sensation limit.
174
Sound waves cause vibration in the eardrum. The three smallest human bones in
the middle ear (malleus, incus, stapes) provide an “impedance match” between air
and liquid and conduct the sound via a second membrane, the oval window, to the
cochlea. Its three chambers are rolled up into a spiral. The basilar membrane that
separates the two main chambers decreases in stiffness along the spiral, such that
the end near the stapes vibrates best at the highest frequencies, whereas for lower
frequencies that amplitude peak moves to the far end.
175

Frequency discrimination and critical bands

A pair of pure tones (sine functions) cannot be distinguished as two
separate frequencies if both are in the same frequency group (“critical
band”). Their loudness adds up, and both are perceived with their
average frequency.
The human ear has about 24 critical bands whose width grows non-
linearly with the center frequency.
Each audible frequency can be expressed on the “Bark scale” with
values in the range 0–24. A good closed-form approximation is
26.81
b≈ − 0.53
1 + 1960f Hz

where f is the frequency and b the corresponding point on the Bark

scale.
Two frequencies are in the same critical band if their distance is below
1 bark.
176
 Masking
→ Louder tones increase the sensation limit for nearby frequencies and
suppress the perception of quieter tones.
→ This increase is not symmetric. It extends about 3 barks to lower
frequencies and 8 barks to higher ones.
→ The sensation limit is increased less for pure tones of nearby fre-
quencies, as these can still be perceived via their beat frequency.
For the study of masking effects, pure tones therefore need to be
distinguished from narrowband noise.
→ Temporal masking: SL rises shortly before and after a masker.

177

Audio demo: loudness and masking

loudness.wav
Two sequences of tones with frequencies 40, 63, 100, 160, 250, 400,
630, 1000, 1600, 2500, 4000, 6300, 10000, and 16000 Hz.
→ Sequence 1: tones have equal amplitude
→ Sequence 2: tones have roughly equal perceived loudness
Amplitude adjusted to IEC 60651 “A” weighting curve for soundlevel meters.

masking.wav
Twelve sequences, each with twelve probe-tone pulses and a 1200 Hz
masking tone during pulses 5 to 8.
Probing tone frequency and relative masking tone amplitude:
10 dB 20 dB 30 dB 40 dB
1300 Hz
1900 Hz
700 Hz
178
 Audio demo: loudness.wav
80 0 dBA curve (SL)
first series
second series
70

60

50
SPL

40
dB

30

20

10

40 63 100 160 250 400 630 1000 1600 2500 4000 6300 10000 16000
Hz
179

Audio demo: masking.wav

80 0 dBA curve (SL)
masking tones
probing tones
70 masking thresholds

60

50
SPL

40
dB

30

20

10

40 63 100 160 250 400 630 1000 1600 2500 4000 6300 10000 16000
Hz
180
 Quantization
Uniform/linear quantization: Non-uniform quantization:
6 6
5 5
4 4
3 3
2 2
1 1
0 0
−1 −1
−2 −2
−3 −3
−4 −4
−5 −5
−6 −6
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6

Quantization is the mapping from a continuous or large set of val-

ues (e.g., analog voltage, floating-point number) to a smaller set of
(typically 28 , 212 or 216 ) values.
This introduces two types of error:
→ the amplitude of quantization noise reaches up to half the max-
imum difference between neighbouring quantization levels
→ clipping occurs where the input amplitude exceeds the value of
the highest (or lowest) quantization level
181

Example of a linear quantizer (resolution R, peak value V ):

   
x 1
y = max −V, min V, R +
R 2
Adding a noise signal that is uniformly distributed on [0, 1] instead of adding 21 helps to spread
the frequency spectrum of the quantization noise more evenly. This is known as dithering.

Variant with even number of output values (no zero):

    
x 1
y = max −V, min V, R +
R 2

Improving the resolution by a factor of two (i.e., adding 1 bit) reduces

the quantization noise by 6 dB.
Linearly quantized signals are easiest to process, but analog input levels
need to be adjusted carefully to achieve a good tradeoff between the
signal-to-quantization-noise ratio and the risk of clipping. Non-uniform
quantization can reduce quantization noise where input values are not
uniformly distributed and can approximate human perception limits.
182
 Logarithmic quantization
Rounding the logarithm of the signal amplitude makes the quantiza-
tion error scale-invariant and is used where the signal level is not very
predictable. Two alternative schemes are widely used to make the
logarithm function odd and linearize it across zero before quantization:
µ-law:

V log(1 + µ|x|/V )
y= sgn(x) for −V ≤ x ≤ V
log(1 + µ)

A-law:
( A|x| V
1+log A
sgn(x) for 0 ≤ |x| ≤ A
y= A|x|
V (1+log V ) V
1+log A
sgn(x) for A
≤ |x| ≤ V

European digital telephone networks use A-law quantization (A = 87.6), North American ones
use µ-law (µ=255), both with 8-bit resolution and 8 kHz sampling frequency (64 kbit/s). [ITU-T
G.711]
183

V µ−law (US)
A−law (Europe)
signal voltage

−V

−128 −96 −64 −32 0 32 64 96 128

byte value

184
Joint Photographic Experts Group – JPEG
Working group “ISO/TC97/SC2/WG8 (Coded representation of picture and audio information)”
was set up in 1982 by the International Organization for Standardization.

Goals:
→ continuous tone gray-scale and colour images
→ recognizable images at 0.083 bit/pixel
→ useful images at 0.25 bit/pixel
→ excellent image quality at 0.75 bit/pixel
→ indistinguishable images at 2.25 bit/pixel
→ feasibility of 64 kbit/s (ISDN fax) compression with late 1980s
hardware (16 MHz Intel 80386).
→ workload equal for compression and decompression
The JPEG standard (ISO 10918) was finally published in 1994.
William B. Pennebaker, Joan L. Mitchell: JPEG still image compression standard. Van Nostrad
Reinhold, New York, ISBN 0442012721, 1993.
Gregory K. Wallace: The JPEG Still Picture Compression Standard. Communications of the
ACM 34(4)30–44, April 1991, http://doi.acm.org/10.1145/103085.103089
185

Summary of the baseline JPEG algorithm

The most widely used lossy method from the JPEG standard:

→ Colour component transform: 8-bit RGB → 8-bit YCrCb

→ Reduce resolution of Cr and Cb by a factor 2

→ For the rest of the algorithm, process Y , Cr and Cb compo-

nents independently (like separate gray-scale images)
The above steps are obviously skipped where the input is a gray-scale image.

→ Split each image component into 8 × 8 pixel blocks

Partial blocks at the right/bottom margin may have to be padded by repeating the
last column/row until a multiple of eight is reached. The decoder will remove these
padding pixels.

→ Apply the 8 × 8 forward DCT on each block

On unsigned 8-bit input, the resulting DCT coefficients will be signed 11-bit integers.

186
 → Quantization: divide each DCT coefficient with the correspond-
ing value from an 8×8 table, then round to the nearest integer:
The two standard quantization-matrix examples for luminance and chrominance are:
16 11 10 16 24 40 51 61 17 18 24 47 99 99 99 99
12 12 14 19 26 58 60 55 18 21 26 66 99 99 99 99
14 13 16 24 40 57 69 56 24 26 56 99 99 99 99 99
14 17 22 29 51 87 80 62 47 66 99 99 99 99 99 99
18 22 37 56 68 109 103 77 99 99 99 99 99 99 99 99
24 35 55 64 81 104 113 92 99 99 99 99 99 99 99 99
49 64 78 87 103 121 120 101 99 99 99 99 99 99 99 99
72 92 95 98 112 100 103 99 99 99 99 99 99 99 99 99

→ apply DPCM coding to quantized DC coefficients from DCT

→ read remaining quantized values from DCT in zigzag pattern
→ locate sequences of zero coefficients (run-length coding)
→ apply Huffman coding on zero run-lengths and magnitude of
AC values
→ add standard header with compression parameters
http://www.jpeg.org/
Example implementation: http://www.ijg.org/
187

Storing DCT coefficients in zigzag order

horizontal frequency

0 1 5 6 14 15 27 28
2 4 7 13 16 26 29 42
vertical frequency

3 8 12 17 25 30 41 43
9 11 18 24 31 40 44 53
10 19 23 32 39 45 52 54
20 22 33 38 46 51 55 60
21 34 37 47 50 56 59 61
35 36 48 49 57 58 62 63
After the 8×8 coefficients produced by the discrete cosine transform
have been quantized, the values are processed in the above zigzag order
by a run-length encoding step.
The idea is to group all higher-frequency coefficients together at the end of the sequence. As many
image blocks contain little high-frequency information, the bottom-right corner of the quantized
DCT matrix is often entirely zero. The zigzag scan helps the run-length coder to make best use
of this observation.
188
 Huffman coding in JPEG
s value range
0 0
1 −1, 1
2 −3, −2, 2, 3
3 −7 . . . − 4, 4 . . . 7
4 −15 . . . − 8, 8 . . . 15
5 −31 . . . − 16, 16 . . . 31
6 −63 . . . − 32, 32 . . . 63
... ...
i −(2i − 1) . . . − 2i−1 , 2i−1 . . . 2i − 1
DCT coefficients have 11-bit resolution and would lead to huge Huffman
tables (up to 2048 code words). JPEG therefore uses a Huffman table only
to encode the magnitude category s = ⌈log2 (|v| + 1)⌉ of a DCT value v. A
sign bit plus the (s − 1)-bit binary value |v| − 2s−1 are appended to each
Huffman code word, to distinguish between the 2s different values within
magnitude category s.
When storing DCT coefficients in zigzag order, the symbols in the Huffman tree are actually
tuples (r, s), where r is the number of zero coefficients preceding the coded value (run-length).
189

Lossless JPEG algorithm

In addition to the DCT-based lossy compression, JPEG also defines a
lossless mode. It offers a selection of seven linear prediction mecha-
nisms based on three previously coded neighbour pixels:

1: x=a
2: x=b
3: x=c
4: x=a+b−c
c b
5: x = a + (b − c)/2
6: x = b + (a − c)/2 a ?
7: x = (a + b)/2

Predictor 1 is used for the top row, predictor 2 for the left-most row.
The predictor used for the rest of the image is chosen in a header. The
difference between the predicted and actual value is fed into either a
Huffman or arithmetic coder.
190
 Advanced JPEG features
Beyond the baseline and lossless modes already discussed, JPEG pro-
vides these additional features:
→ 8 or 12 bits per pixel input resolution for DCT modes
→ 2–16 bits per pixel for lossless mode
→ progressive mode permits the transmission of more-significant
DCT bits or lower-frequency DCT coefficients first, such that
a low-quality version of the image can be displayed early during
a transmission
→ the transmission order of colour components, lines, as well as
DCT coefficients and their bits can be interleaved in many ways
→ the hierarchical mode first transmits a low-resolution image,
followed by a sequence of differential layers that code the dif-
ference to the next higher resolution
Not all of these features are widely used today.
191

JPEG-2000 (JP2)
Processing steps:

→ Preprocessing: If pixel values are unsigned, subtract half of the

maximum value → symmetric value range.

→ Colour transform: In lossy mode, use RGB ↔ YCrCb.

In lossless mode, use RGB ↔ YUV with integer approximation
Y = ⌊(R + 2G + B)/4⌋.

→ Cut each colour plane of the image into tiles (optional), to be

compressed independently, symmetric extension at edges.

→ Apply discrete wavelet transform to each tile, via recursive ap-

plication (typically six steps) of a 2-channel uniformly maximally-
decimated filter bank.
In lossy mode, use a 9-tap/7-tap real-valued filter (Daubechies),
in lossless mode, use a 5-tap/3-tap integer-arithmetic filter.
192
 → Quantization of DWT coefficients (lossy mode only), same
quantization step per subband.
→ Each subband is subdivided into rectangles (code blocks). These
are split into bit planes and encoded with an adaptive arithmetic
encoder (probability estimates based on 9 contexts).
For details of this complex multi-pass process, see D. Taubman: High-performance
scalable image compression with EBCOT. IEEE Trans. Image Processing 9(7)1158–
1170, July 2000. (On http://ieeexplore.ieee.org/)

→ The bit streams for the independently encoded code blocks

are then truncated (lossy mode only), to achieve the required
compression rate.
Features:
→ progressive recovery by fidelity or resolution
→ lower compression for specified region-of-interest
→ CrCb subsampling can be handled via DWT quantization
ISO 15444-1, example implementation: http://www.ece.uvic.ca/~mdadams/jasper/
193

JPEG examples (baseline DCT)

1:5 (1.6 bit/pixel) 1:10 (0.8 bit/pixel)

194
 JPEG2000 examples (DWT)

1:5 (1.6 bit/pixel) 1:10 (0.8 bit/pixel)

195

JPEG examples (baseline DCT)

1:20 (0.4 bit/pixel) 1:50 (0.16 bit/pixel)

Better image quality at a compression ratio 1:50

can be achieved by applying DCT JPEG to a 50%
scaled down version of the image (and then inter-
polate back to full resolution after decompression):

196
 JPEG2000 examples (DWT)

1:20 (0.4 bit/pixel) 1:50 (0.16 bit/pixel)

197

Moving Pictures Experts Group – MPEG

→ MPEG-1: Coding of video and audio optimized for 1.5 Mbit/s
(1× CD-ROM). ISO 11172 (1993).
→ MPEG-2: Adds support for interlaced video scan, optimized
for broadcast TV (2–8 Mbit/s) and HDTV, scalability options.
Used by DVD and DVB. ISO 13818 (1995).
→ MPEG-4: Adds algorithmic or segmented description of audio-
visual objects for very-low bitrate applications. ISO 14496
(2001).
→ System layer multiplexes several audio and video streams, time
stamp synchronization, buffer control.
→ Standard defines decoder semantics.
→ Asymmetric workload: Encoder needs significantly more com-
putational power than decoder (for bit-rate adjustment, motion
estimation, perceptual modeling, etc.)
http://www.chiariglione.org/mpeg/
198
 MPEG video coding
→ Uses YCrCb colour transform, 8×8-pixel DCT, quantization,
zigzag scan, run-length and Huffman encoding, similar to JPEG
→ the zigzag scan pattern is adapted to handle interlaced fields
→ Huffman coding with fixed code tables defined in the standard
MPEG has no arithmetic coder option.

→ adaptive quantization
→ SNR and spatially scalable coding (enables separate transmis-
sion of a moderate-quality video signal and an enhancement
signal to reduce noise or improve resolution)
→ Predictive coding with motion compensation based on 16×16
macro blocks.
J. Mitchell, W. Pennebaker, Ch. Fogg, D. LeGall: MPEG video compression standard.
ISBN 0412087715, 1997. (CL library: I.4.20)
B. Haskell et al.: Digital Video: Introduction to MPEG-2. Kluwer Academic, 1997.
(CL library: I.4.27)
John Watkinson: The MPEG Handbook. Focal Press, 2001. (CL library: I.4.31)
199

MPEG motion compensation

time

backward current picture forward

reference picture reference picture

Each MPEG image is split into 16×16-pixel large macroblocks. The predic-
tor forms a linear combination of the content of one or two other blocks of
the same size in a preceding (and following) reference image. The relative
positions of these reference blocks are encoded along with the differences.
200
 MPEG reordering of reference images
Display order of frames:
time

I B B B P B B B P B B B P

Coding order:
time

I P B B B P B B B P B B B

MPEG distinguishes between I-frames that encode an image independent of any others, P-frames
that encode differences to a previous P- or I-frame, and B-frames that interpolate between the
two neighbouring B- and/or I-frames. A frame has to be transmitted before the first B-frame
that makes a forward reference to it. This requires the coding order to differ from the display
order.
201

MPEG system layer: buffer management

encoder decoder
encoder decoder
buffer fixed−bitrate buffer
channel
buffer content

buffer content
encoder

decoder

time time

MPEG can be used both with variable-bitrate (e.g., file, DVD) and fixed-bitrate (e.g., ISDN)
channels. The bitrate of the compressed data stream varies with the complexity of the input
data and the current quantization values. Buffers match the short-term variability of the encoder
bitrate with the channel bitrate. A control loop continuously adjusts the average bitrate via the
quantization values to prevent under- or overflow of the buffer.
The MPEG system layer can interleave many audio and video streams in a single data stream.
Buffers match the bitrate required by the codecs with the bitrate available in the multiplex and
encoders can dynamically redistribute bitrate among different streams.
MPEG encoders implement a 27 MHz clock counter as a timing reference and add its value as a
system clock reference (SCR) several times per second to the data stream. Decoders synchronize
with a phase-locked loop their own 27 MHz clock with the incoming SCRs.
Each compressed frame is annotated with a presentation time stamp (PTS) that determines when
its samples need to be output. Decoding timestamps specify when data needs to be available to
the decoder.
202
 MPEG audio coding
Three different algorithms are specified, each increasing the processing
power required in the decoder.
Supported sampling frequencies: 32, 44.1 or 48 kHz.

Layer I
→ Waveforms are split into segments of 384 samples each (8 ms at 48 kHz).
→ Each segment is passed through an orthogonal filter bank that splits the
signal into 32 subbands, each 750 Hz wide (for 48 kHz).
This approximates the critical bands of human hearing.

→ Each subband is then sampled at 1.5 kHz (for 48 kHz).

12 samples per window → again 384 samples for all 32 bands

→ This is followed by scaling, bit allocation and uniform quantization.

Each subband gets a 6-bit scale factor (2 dB resolution, 120 dB range, like floating-
point coding). Layer I uses a fixed bitrate without buffering. A bit allocation step
uses the psychoacoustic model to distribute all available resolution bits across the 32
bands (0–15 bits for each sample). With a sufficient bit rate, the quantization noise
will remain below the sensation limit.

→ Encoded frame contains bit allocation, scale factors and sub-band samples.
203

Layer II
Uses better encoding of scale factors and bit allocation information.
Unless there is significant change, only one out of three scale factors is transmitted. Explicit zero
code leads to odd numbers of quantization levels and wastes one codeword. Layer II combines
several quantized values into a granule that is encoded via a lookup table (e.g., 3 × 5 levels: 125
values require 7 instead of 9 bits). Layer II is used in Digital Audio Broadcasting (DAB).

Layer III
→ Modified DCT step decomposes subbands further into 18 or 6 frequencies

→ dynamic switching between MDCT with 36-samples (28 ms, 576 freq.)
and 12-samples (8 ms, 192 freq.)
enables control of pre-echos before sharp percussive sounds (Heisenberg)

→ non-uniform quantization

→ Huffman entropy coding

→ buffer with short-term variable bitrate

→ joint stereo processing

MPEG audio layer III is the widely used “MP3” music compression format.
204
 Psychoacoustic models
MPEG audio encoders use a psychoacoustic model to estimate the spectral
and temporal masking that the human ear will apply. The subband quan-
tization levels are selected such that the quantization noise remains below
the masking threshold in each subband.
The masking model is not standardized and each encoder developer can
chose a different one. The steps typically involved are:
→ Fourier transform for spectral analysis
→ Group the resulting frequencies into “critical bands” within which
masking effects will not vary significantly
→ Distinguish tonal and non-tonal (noise-like) components
→ Apply masking function
→ Calculate threshold per subband
→ Calculate signal-to-mask ratio (SMR) for each subband
Masking is not linear and can be estimated accurately only if the actual sound pressure levels
reaching the ear are known. Encoder operators usually cannot know the sound pressure level
selected by the decoder user. Therefore the model must use worst-case SMRs.
205

Exercise 17 Compare the quantization techniques used in the digital tele-

phone network and in audio compact disks. Which factors to you think led
to the choice of different techniques and parameters here?

Exercise 18 Which steps of the JPEG (DCT baseline) algorithm cause a

loss of information? Distinguish between accidental loss due to rounding
errors and information that is removed for a purpose.

Exercise 19 How can you rotate by multiples of ±90◦ or mirror a DCT-

JPEG compressed image without losing any further information. Why might
the resulting JPEG file not have the exact same file length?

Exercise 20 Decompress this G3-fax encoded line:

1101011011110111101100110100000000000001

Exercise 21 You adjust the volume of your 16-bit linearly quantizing sound-
card, such that you can just about hear a 1 kHz sine wave with a peak
amplitude of 200. What peak amplitude do you expect will a 90 Hz sine
wave need to have, to appear equally loud (assuming ideal headphones)?
206
 Outlook
Further topics that we have not covered in this brief introductory tour
through DSP, but for the understanding of which you should now have
a good theoretical foundation:

→ multirate systems

→ effects of rounding errors

→ adaptive filters

→ DSP hardware architectures

→ modulation and symbol detection techniques

→ sound effects
If you find a typo or mistake in these lecture notes, please notify [email protected].

207

Some final thoughts about redundancy . . .

Aoccdrnig to rsceearh at Cmabrigde Uinervtisy, it
deosn’t mttaer in waht oredr the ltteers in a wrod are,
the olny iprmoetnt tihng is taht the frist and lsat
ltteer be at the rghit pclae. The rset can be a total
mses and you can sitll raed it wouthit porbelm. Tihs is
bcuseae the huamn mnid deos not raed ervey lteter by
istlef, but the wrod as a wlohe.

. . . and perception
Count how many Fs there are in this text:
FINISHED FILES ARE THE RE-
SULT OF YEARS OF SCIENTIF-
IC STUDY COMBINED WITH THE
EXPERIENCE OF YEARS

208