DFT

DFT
1
Course Information
Instructors: Mohamed Khairy & Yasmine Fahmy
TAs: Ahmed AbdelKarim& HazemSoliman
Lectures: Sundays and Wednesdays
Grading: Midterm 22%, Projects: 8%, Final 70%
2
Lec. Topic Source
1 Introduction and Fourier Transforms Chapter 5, Proakis DSP
book
2 DFT:
- Sampling in the frequency domain
- Time-Domain aliasing
Chapter 5, Proakis DSP
book
3 DFT:
- Properties
- Circular convolution and linear convolution using circular convolution
book
4 DFT:
- Frequency resolution and windowing
Wireline Channel
- Properties
- Interference sources
book
Lecture notes
5 Fading:
- Origin of fading
- Doppler frequency
- Classification of fading channels
Chapter 4Rappaport
6 Fading:
- Fast and slow channels
- Flat and frequency selective channels
Chapter 4 Rappaport
7 Fading:
- Delay spread and coherence bandwidth
- Doppler spread and coherence bandwidth
Chapter 4 Rappaport
3
8 Multichannel Modulation (MCM):
- Advantages and how MCM combats ISI
- Block diagram of MCM transceiver
- Basis functions
Haykin, section 6.12
9 MCM:
- The water-filling algorithm
Haykin, section 6.12
10 Discrete Multi-tone DMT:
- Using DFT symmetry properties to generate real baseband MCM signal
- DSL basics
Haykin, section 6.12,
and Cioffis tutorial
11 OFDM:
- Properties of the wireless channel and introduction to multipath fading,
and the delay spread
- Advantages and disadvantages of OFDM systems in wireless channels
- Guard time and cyclic extension
Prasads OFDM book
12 OFDM:
- Block diagram of a digital OFDM transceiver
- Choice of OFDM parameters
Prasads OFDM book
4
Wireless Communications
5
History of Wireless
The Birth of Radio
1897 The Birth of Radio - Marconi awarded patent for wireless telegraph
1897 First Marconi station established on Needles island to communicate
with English coast
1898 Marconi awarded English patent no. 7777 for tuned communication
1898 Wireless telegraphic connection between England and France
established
Transoceanic Communication
1901 Marconi successfully transmits radio signal across Atlantic Ocean from
(first wireless communication across the ocean) Cornwall to Newfoundland
1902 First bidirectional communication across Atlantic
1909 Marconi awarded Nobel prize for physics
http://wireless.ece.ufl.edu/jshea/wireless_history.html
6
History of Wireless (2)
Voice over Radio
1914 First voice over radio transmission
1920s Mobile receivers installed in police cars in Detroit
1930s Mobile transmitters developed; radio equipment occupied most of police car trunk
1935 Frequency modulation (FM) demonstrated by Armstrong
1940s Majority of police systems converted to FM
Birth of Mobile Telephony
1946 First interconnection of mobile users to public switched telephone network (PSTN)
1949 FCC recognizes mobile radio as new class of service
1940s Number of mobile users > 50K
1950s Number of mobile users > 500K
1960s Number of mobile users > 1.4M
1960s Improved Mobile Telephone Service (IMTS) introduced; supports full-duplex, auto
dial, auto trunking
1976 Bell Mobile Phone has 543 pay customers using 12 channels in the New York City
area; waiting list is 3700 people; service is poor due to blocking
7
Cellular Mobile Telephony
1979 NTT/Japan deploys first cellular communication system
1983 Advanced Mobile Phone System (AMPS) deployed in US in 900 MHz band: supports 666
duplex channels
1989 Groupe Spcial Mobile defines European digital cellular standard, GSM
1991 US Digital Cellular phone system introduced
1993 IS-95 code-division multiple-access (CDMA) spread- spectrum digital cellular system
deployed in US
1994 GSM system deployed in US, relabeled ``Global System for Mobile Communications''
Wireless Local Area Networks
1990 Formation of IEEE 802.11 Working Group to define standards for Wireless Local Area
Networks (WLANs)
1997-2003 Releases of IEEE 802.11 WLAN protocol, supporting 1-54 Mbit/s data rates in the
2.4/5 GHz ISM bands based on Orthogonal Frequency Division Multiplexing (OFDM)
2009 Release of IEEE 802.11n WLAN protocol, supporting up to 150 Mbit/s data rates in both the
2.4 GHz and 5 GHz ISM bands.
8
Wireless Metropolitan Area Networks
1999 Formation of IEEE 802.16 Working Group to
define standards for Wireless Metropolitan Area
Networks (WLANs)
2004 release of 802.16d (fixed WiMAX standard)
(OFDM)
2005 release of 802.16e (Mobile WiMAX standard)
2009 Cairo University hosted the WIMAX standard
meeting to discuss development of WiMAX release 2
2012 WiMAX release 2 commercially available
9
3G networks and beyond
2001 UMTS deployment based on WCDMA and
CDMA2000
2007 HSPA often referred to as 3.5G supporting
14Mbps on the downlink
2008 HSPA+ often referred to as 3.75G supporting
42Mbps on the downlink
2010 Number of cellular phones surpassed 4 billion
worldwide and 65 million in Egypt.
2009 first LTE (long term evolution) system
deployment is Sweden supporting 100Mbps on the
downlink. LTE is based on OFDM
10
Fourier transforms
Freq domain Time domain
Fourier transform
Fourier series
Discrete time Fourier
transform
11
Fourier transforms
Fourier transform Cont/aperiodic Aperiodic/cont
Fourier series
transform
12
Fourier transforms
Fourier series Disc/aperiodic Periodic/cont
transform
13
Fourier transforms
Fourier series Disc/aperiodic Periodic/cont
transform
cont/periodic Aperiodic/disc
14
Fourier Transform
2
2
( ) ( )
( ) ( )
j ft
j ft
x f x t e dt
x t x f e df
t
t
=
=
}
}
15
Fourier Series
/ 2
2
/ 2
1
( )
n
T
j f t
n
T
C x t e dt
T
t
=
}
2
( )
n
j f t
n
n
x t C e
t
=
=

/
n
f n T =
16
DTFT
( ) ( )
j n
n
X x n e
e
e

=
=

1
( ) ( )
2
j n
x n X e d
t
e
t
e e
t
+
=
}
17
Discrete Fourier Transform
(DFT)
18
Motivation
We need a transform that is discrete in both
domains, to be able to manipulate signals on
processors.
For example, given a discrete time signal, we
need a DISCRETE frequency domain
representation, unlike the DTFT which is
continuous in the frequency domain.
19
Approach
Analogous to sampling in the time domain, we
will consider sampling the DTFT. Remember
that sampling in the TD causes repetition of
the spectrum in the FD.
Take
20
Sampling in the Freq Domain
Time domain
Frequency domain
(magnitude)
21
0 2pi
n
x[n]
e
|X[e]|
Sampling in the Freq Domain (2)
22
Sampling in the Freq Domain (3)
1
2 /
1
2 /
0
2
( ) ( )
( )
lN N
j kn N
l n lN
N
j kn N
n l
k
X x n e
N
x n lN e
t
t
t +

= =

= =
=
=

We define a new signal
( ) ( )
p
l
x n x n lN
=
=
1
2 /
0
1
2 /
0
2
( ) ( )
1 2
( ) ( )
N
j kn N
p
n
N
j kn N
p
k
k
X x n e
N
k
x n X e
N N
t
t
t
t
=
=
=
23
Fourier Transform pair
The DFT
N>L
N=L
( )
p
x n
24
Time Domain Aliasing

N<L
( ) ( )
p
l
x n x n lN
=
=
25
Example
1 0 n L-1
( )
0 otherwise
x n
s s
Find the DFT, X(k), of

1
( 1) / 2
0
( 1) /
sin( / 2)
( ) ( )
sin( / 2)
sin( / )
(2 / )
sin( / )
L
j n j L
n
j k L N
L
X x n e e
kL N
X k N e
k N
e e
t
e
e
e
t
t
t

=

= =
=
26
n
Example (cntd.)
27
0 2pi
e
|X(e)|
Example (cntd.)
DTFT, N=L, N=2L, 4L
28
e
|X(e)|
Time Domain Sampling
29
time
x(t)
X(f)
Time Domain Reconstruction
30
LPF
time
x(t)
X(f)
Frequency domain sampling
31
freq
time
Window 0N-1

Analogy
32
sampling filter
x(t)
x[n]
x(t)
/ 2 / 2
1 -
( )
0 o.w.
s s
f f f
H f
< <
sampling filter
X(e)
X[k]
X(e)
1 0 1
( )
0 o.w.
sin( / 2)
( )
sin( / 2)
n N
h n
N
H
e
e
e
s s
=
Example
Determine the reconstructed spectrum
33
( ) ( ) 0< <1
n
x n a u n a =
0
1
( )
1
n j n
j
n
X a e
ae
e
e
e

=
= =
0 2pi
x(n)
|X(e)|
e
Example (cntd.)
34
2 /
1
( )
1
j k N
X k
ae
t
=
Take IDFT to get ( ) 0 n N-1

1
n
p
N
a
x n
a
= s s
If N is taken large enough, x

p
(n) will approach x(n)
To reconstruct the frequency domain, calculate
1
0
1 1
( ) ( )
1 1
N j N
N
j n
p
N j
n
a e
X x n e
a ae
e
e
e
e

= =

If N is large, a
N
will decrease and the error will decrease
Example (cntd.)
35
Reconstructed
spectrum
e
|X(e)|
DFT Properties
36
1
0
1
0
2 /
( ) ( )
( ) ( )
1
( ) ( )
( ) ( )
( ) ( )
N
kn
N
n
N
kn
N
k
j N
N
x n X k
X k x n W
x n X k W
N
W e
x n N x n
X k N X k
t
=
=
=
+ =
+ =
Symmetry Properties of DFT

If x(n) is real, then
More properties can be obtained for even,
odd, imaginary, signals
37
*
( ) ( ) X k X N k =
Parsevals Theorem
38
1 1
* *
0 0
1 1
2 2
0 0
1
( ) ( ) ( ) ( )
1
( ) ( )
N N
n k
N N
n k
x n y n X k Y k
N
x n X k
N

= =

= =
=
=

Linearity and Shift
39
( ) ( ) ( ) ( )
DFT
N
ax n by n aX k bY k + +
The N-point DFT of a finite duration sequence, x(n) of length L<N
is equivalent to the N-point DFT of a periodic sequence x
p
(n)
of period N
Any shift by m units, will be applied to x
p
(n) such that
( ) ( )
p
l
x n m x n m lN
=
=
Shift
40
4
(( )) ( mod )
N
x n m x n m N =
4
Circular Shift
41
2 4
3
1
x(0)
x(1)
x(2)
x(3)
x(n)
3 1
4
2
y(0)
y(1)
y(2)
y(3)
y(n)= x((n-1))
4
4
4
y(n)
Multiplication of 2 DFTs
In analog systems, using Fourier transform
If
Then
Using the DFT
If
42
( ) ( ) ( )
( ) ( ) ( )
Y f H f x f
y t h x t d t t t
=
=
}
1
0
( ) ( ) ( )
( ) ( ) (( ))
N
N
n
Y k X k H k
y m x n h m n
=
=
=
Note that both signals must

have the same size N
Circular convolution
Example
43
Perform circular convolution between the
following 2 sequences:
x=[2 1 2 1 ] h=[1 2 3 4]
2 2
1
1
x(0)
x(1)
x(2)
x(3)
x(n)
4 1
4
0
( ) ( ) (( ))
n
y m x n h m n
=
=
Example
44
following 2 sequences:
x=[2 1 2 1 ] h=[1 2 3 4]
2 2
1
1
x(0)
x(1)
x(2)
x(3)
x(n)
3 1
2
4
h(0)
h(1)
h(2)
h(3)
h(n)
4 1
4
0
( ) ( ) (( ))
n
y m x n h m n
=
=
Example
45
following 2 sequences: x=[2 1 2 1 ] h=[1 2 3 4]
2 2
1
1
x(0)
x(1)
x(2)
x(3)
x(n)
3 1
2
4
h(0)
h(1)
h(2)
h(3)
h(n)
3 1
4
2
h((-n))
4 1
4
0
( ) ( ) (( ))
n
y m x n h m n
=
=
Example (cntd.)
46
4 1
4
0
(0) ( ) (( )) 14
n
y x n h n
=
= =
2 2
1
1
x(0)
x(1)
x(2)
x(3)
x(n)
3 1
4
2
h((-n))
Example (cntd.)
47
4 1
4
0
(1) ( ) ((1 )) 16
n
y x n h n
=
= =
2 2
1
1
x(0)
x(1)
x(2)
x(3)
x(n)
4 2
1
3
h((1-n))
Example (cntd.)
48 48
4 1
4
0
(2) ( ) ((2 )) 14
n
y x n h n
=
= =
2 2
1
1
x(0)
x(1)
x(2)
x(3)
x(n)
1 3
2
4
h((2-n))
Example (cntd.)
49
4 1
4
0
(3) ( ) ((1 )) 16
n
y x n h n
=
= =
2 2
1
1
x(0)
x(1)
x(2)
x(3)
x(n)
2 4
3
1
h((3-n))
y(n)=[14 16 14 16]
Linear Filtering Based on DFT
50
1
0
( ) ( ) ( )
( ) ( ) (( ))
N
N
n
Y k X k H k
y m x n h m n
=
=
=
To perform circular convolution, x and h must be of the

same size, N, and y is also of size N.
y is NOT equal to the linear convolution of x and h
How can we FORCE the circular convolution to calculate
the linear convolution?
Assume x is of length L and h is of length M
The linear convolution output will be of length
N=L+M-1
In the case of linear convolution,
51
1
0
( ) ( ) ( )
( ) ( ) ( )
L
n
Y X H
y m x n h m n
e e e
=
=
=

The sequence y(n) can be uniquely
represented in the frequency domain by
N=M+L-1 samples of its spectrum Y(e)
Where X(k) and H(k) are the N-point DFTs of
x(n) and h(n)
52
2 /
2 /
2 /
( ) ( ) 0,1,..., -1
= ( ) ( ) 0,1,..., -1
= ( ) ( ) 0,1,..., -1
k N
k N
k N
Y k Y k N
X H k N
X k H k k N
e t
e t
e t
e
e e
=
=
=
= =
=
=
We PAD x(n) and h(n) with zeros to increase
their duration to N
Padding does not change the spectra X(e)
and H(e)
By increasing the lengths of x[n] and h[n] to
N and then circularly convolving the resulting
sequences, we obtain the same result we
would obtain with linear convolution
53
Example
Perform Linear convolution using Circular
convolution for x(n)=[1 2 2 1], h(n)=[1 2 3 ]
L=3, M=4, use N=4+3-1=6
Do circular convolution between
x(n)=[1 2 2 1 0 0 ], h(n)=[1 2 3 0 0 0]
You can also get the DFT(x) and DFT(h) with N=6,
Calculate Y(k)=X(k)H(K), N=6
Then y(n)=IDFT(Y)
54
Frequency Analysis Using DFT
To compute the spectrum, the signal values for
all time are required, which is non-practical.
What are the implications of using a finite data
record in frequency analysis using the DFT?
Assume a band-limited, analog signal with
BW=B, sampling rate F
s
=1/T>2B, and L total
samples.
55
56
( 1) / 2
1 0 1
( )
0 o.w.
sin( / 2)
( )
sin( / 2)
j L
n N
w n
L
W e
e
e
e
e

s s
=
( ) ( ) ( ) x n x n w n =
Assume x(n) is the signal to be analyzed, limiting the number
of samples to L is equivalent to multiplying x(n) by a
rectangular window w(n) of length L
57
Assume that the input signal, x(n)=cos(w
o
n)
58
| |
( ) 1/ 2 ( ) ( )
o o
X W W e e e e e = + +
0 w
o
2pi-w
o
2pi
| ( ) | X e
59
To compute we use the DFT, by padding
with N-L zeros to compute the N-point
DFT of the L point sequence
Notice that the windowed spectrum is not
localized to a single frequency but the power
leaked out into the entire frequency range
Leakage not only distort the spectrum but
decreases resolution as well
( ) X e
( ) x n
Example
60
0 1 2
0 1 2
( ) cos( ) cos( ) cos( )
0.2 , 0.22 , 0.6
x n n n n e e e
e t e t e t
= + +
= = =
L=25, N=1000
0 0.2pi 0.6pi 1.4pi 1.8pi 2pi
Example (cntd.)
61
0 0.2pi 0.6pi 1.4pi 1.8pi 2pi
L=200,N=1000
Frequency Resolution of DFT
The spectrum of the rectangular window
sequence has its first zero at
If the 2 window functions
and overlap and thus the 2
spectral lines are not distinguishable
2t/L defines the frequency resolution of the
DFT
Increasing N increase the visibility of the
spectrum already defined by L
62
2 / L e t =
1 2
2 / L e e t <
1
( ) W e e
2
( ) W e e
Example
63
, t 0
( )
0 t<0
t
a
e
x t

>
=

Is sampled at the rate Fs=20 samples per

sec, and a block of length 100 samples is
used to estimate the spectrum. Compare
the spectrum of the truncated discrete
signal to the spectrum of the analog signal.
1
( )
1 2
a
X F
j F t
=
+
Example (cntd.)
64
Example (cntd.)
65
/ 20
( ) (0.95) , n 0
s
nT n n
x n e e

= = = >
Example (cntd.)
We would like to use L samples only of the
discrete signal to compute the spectrum of
the signal
66
L=100, N=200
Example (cntd.)
67
L=25, N=200

DFT

Uploaded by

DFT

Uploaded by

DFT

Time Domain Aliasing

Find the DFT, X(k), of

Take IDFT to get ( ) 0 n N-1

If N is taken large enough, x

Symmetry Properties of DFT

Note that both signals must

To perform circular convolution, x and h must be of the

Linear Filtering Based on DFT

Is sampled at the rate Fs=20 samples per

You might also like