Digital Signal Processing - Lecture 1: Introduction
Signal Processing EE2S31
Delft University of Technology, The Netherlands
Course Organization
2 / 41
Course organization - DSP track
● Information
● Website: general overview
● Brightspace: more detailed information, quiz, forum
● Organization
● DSP 1x a week(±) on Monday, Tuesday, or Thursday
● Exam comprises both tracks
3 / 41
Prerequisite: EE2S11 Signals and Systems
● Continuous-time vs discrete-time signals
● Linear time-invariant systems
● Fourier Transform, spectral representation
● Discrete-time Fourier transform
⋮
4 / 41
Study materials - DSP track
● Theory
● Lectures
● Book (Proakis, Manolakis: Digital Signal Processing)
● Collegerama videos
● Important notes:
1. Studying the slides is not sufficient; you need to read the book!
2. Attending lectures is important; we solve exercises during lectures
● Practice
● Brightspace Quiz (easy)
● Exercises from book (more advanced)
● Past exams on website
5 / 41
Lectures: Digital Signal Processing Track
Geethu Joseph
● Lectures 1-5
Mid-term exam (Lectures 1-4)
Bori Hunyadi
● Lectures 6 - 8
● Exercise session
Final exam (Lectures 5-8)
6 / 41
Exam
● The exam is conducted in two parts; both partial exams contain 50% of
questions from each track
● The final grade is the average of the two partial exam results, rounded to half a
digit
● The re-examination is conducted in one part (over all lecture material)
● The exams are closed-book, with one A4-size page (2 sides) of handwritten
notes permitted
7 / 41
Introduction and Applications
8 / 41
Introduction
What is a signal?
9 / 41
Introduction
What is a signal?
Any measurable quantity that conveys information
9 / 41
Introduction
What is a signal?
Any measurable quantity that conveys information
9 / 41
Introduction
What is a signal?
Any measurable quantity that conveys information
Examples
1 electrical: voltage output of amplifier
2 mechanical: acceleration of a car
3 acoustic: air pressure measured by a microphone
4 biological: body temperature
5 image and video: intensities of each pixel
9 / 41
Classification of Signals
1 Continuous-time vs discrete-time
2 Unquantized (continuous-amplitude) vs quantized (discrete amplitude)
10 / 41
Classification of Signals
1 Continuous-time vs discrete-time
2 Unquantized (continuous-amplitude) vs quantized (discrete amplitude)
10 / 41
Introduction
Digital Signal Processing
Processing of analog signals employing discrete-time operations implemented on
digital hardware
11 / 41
Analog vs Digital Signal Processing
⇓ instead
12 / 41
Analog vs Digital Signal Processing
⇓ instead
Pro:
● accuracy
● flexibility
● ease of data storage
12 / 41
Analog vs Digital Signal Processing
⇓ instead
Pro: Cons:
● accuracy ● extra complexity
● flexibility ● limited bandwidth
● ease of data storage ● quantization effects
12 / 41
Digital signal processing
sampling filtering reconstruction
quantization spectral analysis
13 / 41
DSP applications
● Digital communication
● Audio signal processing
● Speech signal processing
● Image Processing
● Medical applications
14 / 41
DSP applications (1)
Mobile communication:
15 / 41
DSP Applications (2)
EEG processing for epileptic seizure detection:
16 / 41
DSP Applications (3):
Seizure detection pipeline:
17 / 41
This Course
● Sampling and reconstruction
● Non-ideal sampling and reconstruction
● Sampling in the frequency domain: DFT
● DFT basics
● Spectral analysis and filtering using DFT
● Efficient implementation of DFT: FFT
● Quantization and effects
● Quantization, coding, sigma-delta
● Round-off effects and filter structures
● Multirate signal processing
18 / 41
Recap: Ideal sampling and reconstruction
Reference: Chapter 6.1 of the textbook
19 / 41
Ideal sampling and reconstruction
Under which conditions can we
reconstruct xa (t)?
20 / 41
Ideal sampling and reconstruction
Under which conditions can we
reconstruct xa (t)?
To answer this question, we will
investigate the form of the digital
signal in the frequency domain.
20 / 41
Recap: Fourier Transform in continuous and discrete time
FT
∞
Xa (F ) = ∫ xa (t)e −j2πFt dt
−∞
Inverse FT
∞
xa (t) = ∫ Xa (F )e j2πFt dF
−∞
F [Hz]: frequency
Ω [radians/s]: angular frequency
Ω = 2πF
21 / 41
Recap: Fourier Transform in continuous and discrete time
FT DTFT
∞ ∞
Xa (F ) = ∫ xa (t)e −j2πFt dt X (f ) = ∑ x[n]e −j2πfn
−∞
n=−∞
Inverse FT Inverse DTFT
∞
1/2
xa (t) = ∫ Xa (F )e j2πFt dF x[n] = ∫ X (f )e j2πfn df
−∞
−1/2
F [Hz]: frequency f [cycles/sample]: normalized frequency
Ω [radians/s]: angular frequency ω [rad/sample]: normalized angular frequency
Ω = 2πF ω = 2πf
21 / 41
Recap: Fourier Transform in continuous and discrete time
FT DTFT
∞ ∞
Xa (F ) = ∫ xa (t)e −j2πFt dt X (f ) = ∑ x[n]e −j2πfn
−∞
n=−∞
Inverse FT Inverse DTFT
∞
1/2
xa (t) = ∫ Xa (F )e j2πFt dF x[n] = ∫ X (f )e j2πfn df
−∞
−1/2
F [Hz]: frequency f [cycles/sample]: normalized frequency
Ω [radians/s]: angular frequency ω [rad/sample]: normalized angular frequency
Ω = 2πF ω = 2πf
Ω = ω/T F = f ⋅ Fs
21 / 41
Ideal sampling and reconstruction
Can we express the DTFT of the sampled signal using the FT of the analog signal?
22 / 41
DTFT of the sampled signal Vs the FT of the analog signal
● Recall the relation between the sampled and analog signals
x[n] = xa (nT )
23 / 41
DTFT of the sampled signal Vs the FT of the analog signal
● Recall the relation between the sampled and analog signals
x[n] = xa (nT )
● Expressing them in using inverse (DT)FT,
1
2
∞
∫ X (f )e j2πfn df = ∫ Xa (F )e j2πFt dF ∣t=nT
− 21 −∞
∞
=∫ Xa (F )e j2πFTn dF
−∞
23 / 41
DTFT of the sampled signal Vs the FT of the analog signal
● Recall the relation between the sampled and analog signals
x[n] = xa (nT )
● Expressing them in using inverse (DT)FT,
1
2
∞
∫ X (f )e j2πfn df = ∫ Xa (F )e j2πFt dF ∣t=nT
− 21 −∞
∞
=∫ Xa (F )e j2πFTn dF
−∞
● We try to find a function g rewrite
∞ 1
2
∫ Xa (F )e j2πFTn dF = ∫ g (Xa (f ))e j2πfn df
−∞ − 21
Ô⇒ X (f ) = g (Xa (f ))
23 / 41
DTFT of digital vs FT of analog signal
Our goal
∞ 1/2
∫ Xa (F )e j2πFTn dF = ∫ g (Xa (f ))e j2πfn df Ô⇒ X (f ) = g (Xa (f ))
−∞ −1/2
24 / 41
DTFT of digital vs FT of analog signal
Our goal
∞ 1/2
∫ Xa (F )e j2πFTn dF = ∫ g (Xa (f ))e j2πfn df Ô⇒ X (f ) = g (Xa (f ))
−∞ −1/2
1 Divide the infinite interval to Fs = 1/T long intervals
kFs + 2s
F
∞ ∞ ∞
j2πF /Fs n
∫ Xa (F )e j2πFTn
dF = ∫ Xa (F )e dF = ∑ ∫ Xa (F )e j2πF /Fs n dF
−∞ −∞ k=−∞
kFs − 2s
F
24 / 41
DTFT of digital vs FT of analog signal
Our goal
∞ 1/2
∫ Xa (F )e j2πFTn dF = ∫ g (Xa (f ))e j2πfn df Ô⇒ X (f ) = g (Xa (f ))
−∞ −1/2
1 Divide the infinite interval to Fs = 1/T long intervals
kFs + 2s
F
∞ ∞ ∞
j2πF /Fs n
∫ Xa (F )e j2πFTn
dF = ∫ Xa (F )e dF = ∑ ∫ Xa (F )e j2πF /Fs n dF
−∞ −∞ k=−∞
kFs − 2s
F
2 Change of variables to match the limits of integrals f → F /Fs − k
1/2
∞ ∞
∫ Xa (F )e j2πFTn dF = ∑ ∫ Xa (Fsf + kFs)e j2π(f +k)n Fs df
−∞
−1/2
k=−∞
24 / 41
DTFT of digital vs FT of analog signal
Our goal
∞ 1/2
∫ Xa (F )e j2πFTn dF = ∫ g (Xa (f ))e j2πfn df Ô⇒ X (f ) = g (Xa (f ))
−∞ −1/2
1 Divide the infinite interval to Fs = 1/T long intervals
kFs + 2s
F
∞ ∞ ∞
j2πF /Fs n
∫ Xa (F )e j2πFTn
dF = ∫ Xa (F )e dF = ∑ ∫ Xa (F )e j2πF /Fs n dF
−∞ −∞ k=−∞
kFs − 2s
F
2 Change of variables to match the limits of integrals f → F /Fs − k
1/2
∞ ∞
∫ Xa (F )e j2πFTn dF = ∑ ∫ Xa (Fsf + kFs)e j2π(f +k)n Fs df
−∞
−1/2
k=−∞
3 Exchange sum and integration and note that e j2π(f +k)n = e j2πfn is periodic
1/2
∞ ∞
∫ Xa (F )e j2πFTn
dF = ∫ [Fs ∑ Xa ((f − k)Fs)] e j2πfn df
−∞
−1/2
k=−∞
24 / 41
DTFT of digital vs FT of analog signal
Our goal
∞ 1/2
∫ Xa (F )e j2πFTn dF = ∫ g (Xa (f ))e j2πfn df Ô⇒ X (f ) = g (Xa (f ))
−∞ −1/2
We proved that
1/2
∞ ∞
∫ Xa (F )e j2πFTn dF = ∫ [Fs ∑ Xa ((f − k)Fs)] e j2πfn df
−∞
−1/2
k=−∞
∞
1
Ô⇒ X (f ) = Fs ∑ Xa ((f − k)Fs) = X (F − kFs )
k=−∞ Fs
as F = f Fs
25 / 41
Ideal sampling and reconstruction
26 / 41
Ideal sampling and reconstruction
26 / 41
Spectrum of the sampled signal
∞
X (f ) = Fs ∑ Xa ((f − k)Fs)
k=−∞
27 / 41
Spectrum of the sampled signal
⎧
⎪
⎪ 1 X (F = f Fs ) if∣F ∣ < Fs /2
Xa (F ) = ⎨ Fs
⎪
⎪
⎩0 otherwise
∞
X (f ) = Fs ∑ Xa ((f − k)Fs)
k=−∞
27 / 41
Spectrum of the sampled signal
⎧
⎪
⎪ 1 X (F = f Fs ) if∣F ∣ < Fs /2
Xa (F ) = ⎨ Fs
⎪
⎪
⎩0 otherwise
∞
X (f ) = Fs ∑ Xa ((f − k)Fs)
k=−∞
Sampling theorem
If the signal is bandlimited, it is possible to reconstruct the original signal from the
samples, provided that the sampling rate is at least twice the highest frequency
contained in the signal (i.e., the Nyquist rate).
27 / 41
Ideal reconstruction in frequency domain
Define an ideal low-pass filter G (f ):
⎧
⎪ −1
⎪Fs , if ∣F ∣ ≤ F2s
G (F ) = ⎨
⎪
⎪
⎩0, otherwise
Apply it G (F ) to X (F ):
Xa (F ) = G (F )X (F )
28 / 41
Ideal reconstruction in time domain
● In the frequency domain, we have
Xa (F ) = G (F )X (F ),
⎧
⎪ −1
⎪Fs , if ∣F ∣ ≤ 2s
F
where G (F ) = ⎨
⎪
⎪
⎩0, otherwise
● In the time domain, we have
∞
xa (t) = g (t) ∗ x(t) = ∑ x[n]g (t − nT ),
n=−∞
where the interpolator is
sin(πt/T )
g (t) = inverse DFTF(G (F )) =
πt/T
29 / 41
The ideal interpolator
Sampled signal
x[n]
Terms in the convolution operation
30 / 41
The ideal interpolator
Sampled signal
x[n]
Terms in the convolution operation
sin(π/T (t − 1T ))
x[1]
(π/T )(t − 1T )
30 / 41
The ideal interpolator
Sampled signal
x[n]
Terms in the convolution operation
sin(π/T (t − 2T ))
x[2]
(π/T (t − 2T )
30 / 41
The ideal interpolator
Sampled signal
x[n]
Terms in the convolution operation
sin(π/T (t − 3T ))
x[3]
(π/T )(t − 3T )
30 / 41
The ideal interpolator
Sampled signal
⎧
⎪ −1
⎪Fs , if ∣F ∣ ≤= F2s
G (f ) = ⎨
⎪
⎪
x[n] ⎩0, otherwise
Terms in the convolution operation
sin(π/T (t − 3T ))
x[3]
(π/T )(t − 3T )
30 / 41
The ideal interpolator
Sampled signal
⎧
⎪ −1
⎪Fs , if ∣F ∣ ≤= F2s
G (f ) = ⎨
⎪
⎪
x[n] ⎩0, otherwise
Terms in the convolution operation
sin(π/T (t − 3T ))
x[3]
(π/T )(t − 3T )
30 / 41
Ideal sampling and reconstruction
31 / 41
Ideal sampling and reconstruction
31 / 41
Quiz
Go to www.kahoot.it
32 / 41
Quiz - question 1
The spectrum of a continuous-time signal is depicted above. Which one of the
figures below represents the spectrum of the sampled version of the signal?
33 / 41
Quiz - question 1
The spectrum of a continuous-time signal is depicted above. Which one of the
figures below represents the spectrum of the sampled version of the signal?
Answer c
33 / 41
Quiz - question 2
Aliasing occurs when we
a oversample a signal, i.e. with a sampling rate Fs ≫ 2FH
b sample an aperiodic signal
c sample below the Nyquist rate
34 / 41
Quiz - question 2
Aliasing occurs when we
a oversample a signal, i.e. with a sampling rate Fs ≫ 2FH
b sample an aperiodic signal
c sample below the Nyquist rate
Answer c
34 / 41
Quiz - question 3
The reconstruction of an analog signal from its samples can happen using
a Highpass filter in the frequency domain
b Convolution with a sinc function in the time domain
c The inverse Fourier transform
35 / 41
Quiz - question 3
The reconstruction of an analog signal from its samples can happen using
a Highpass filter in the frequency domain
b Convolution with a sinc function in the time domain
c The inverse Fourier transform
Answer b
35 / 41
Quiz - question 4
What is the Nyquist rate for the analog signal
xa (t) = 3 cos(50πt) + 10sin(300πt) + cos(100πt)?
a 300 Hz
b 600 Hz
c 100 Hz
36 / 41
Quiz - question 4
What is the Nyquist rate for the analog signal
xa (t) = 3 cos(50πt) + 10sin(300πt) + cos(100πt)?
a 300 Hz
b 600 Hz
c 100 Hz
Answer a
36 / 41
Quiz - question 5
If a signal has a maximum frequency between 1000 Hz and 4000 Hz, which of the
below is the most appropriate sampling rate?
a 10000 Hz
b 2000 Hz
c 9000 Hz
37 / 41
Quiz - question 5
If a signal has a maximum frequency between 1000 Hz and 4000 Hz, which of the
below is the most appropriate sampling rate?
a 10000 Hz
b 2000 Hz
c 9000 Hz
Answer c
37 / 41
Summary So Far
Nyquist Rate = Twice the maximum frequency
38 / 41
Non-ideal sampling and reconstruction
1
3
39 / 41
Sampling and reconstruction in practice
1 Delta pulse train for sampling: non-zero duration in practice
2 Signals often are non-low pass, non-bandlimited
● How to sample non-bandlimited signals?
● How to sample bandpass signals?
3 Sinc interpolation in practice is not possible: infinite length
40 / 41
Next Lecture: Non-ideal Cases
1 Delta pulse train for sampling: non-zero duration in practice
2 Signals often are non-low pass, non-bandlimited
● How to sample non-bandlimited signals?
● How to sample bandpass signals?
3 Sinc interpolation in practice is not possible: infinite length
Solve the following exercises from the book: 6.1, 6.2, 6.3, 6.4, 6.5 (solutions
available on BrightSpace)
41 / 41
