DFT PDF

Signal Processing Fundamentals – Part I
Spectrum Analysis and Filtering

5. Fourier Transform and Spectrum Analysis
5. Fourier Transform and Spectrum

Analysis
1
Spectrum of Non-periodic
Signals
• Fourier series help us to find the spectrum of
periodic signals
• Most signals are not periodic

• Speech, audio, etc.
• Need another tool to find the spectrum of non-

periodic (aperiodic) signals
⇒ Fourier Transform
2
Fourier Transform of Discrete-

time Signals
• Let x(t) be an aperiodic continuous-time signal, x[n]
is the samples of x(t) such that:
x[n] = x(nTs) Radian frequency
• The spectrum of x[n] is given by: ωˆ = ωTs

∞ ∞
X p (ω ) = ∑ x[n]e − jωnTs
or X p (ωˆ ) = ∑ x[n]e − jωˆ n
n= −∞ n= −∞
3
Aperiodic Signals have Periodic

Spectrum
• It is interesting to note that Xp(ω) is periodic since
∞
X p (ωˆ + 2πk ) = ∑ x[ n ]e − j (ωˆ + 2πk ) n
n= −∞ 1
∞
= ∑ x[n]e − jωˆ n − j 2πnk
e = X p (ωˆ )
n= −∞
4
X p (ωˆ )
-2π 0 2π
ωˆ = ωTs
5
• If x(t) has a spectrum of X(ω) and x[n] = x(nTs) has

a spectrum of Xp(ω), it can be shown that
∞
1
X p (ωˆ ) =
Ts
∑ X (ωˆ + 2πk ) −∞ ≤ω ≤ ∞
k = −∞
1
= K+ X (ωˆ − 2π )
Ts
1
+ X (ωˆ )
Ts
1
+ X (ωˆ + 2π )+ K
Ts
6
Time Domain Frequency Domain
x(t) X(ω)
1
0 t -B 0 B ω
Ideal low
X p (ωˆ ) pass filter
x[n]=x(nTs)
ωˆ = ωTs
0 | |←Ts n -2π 0 2π
7

x(t) X(ω)
1
0 t -B 0 B ω
x[n] X p (ωˆ )
n
0 | |←Ts -2π 0 2π ωˆ = ωTs
x[n] X p (ωˆ )
0 | |←Ts n
-2π 0 2π
8 ωˆ = ωTs

x(t) X(ω)
1
0 t -B 0 B ω
x[n] X p (ωˆ )
0 | |←Ts n
ωˆ = ωTs
x’(t) X’(ω)
0 t -B 0 B ω
9
• If the signal has frequency components beyond |π|,

after sampling, these frequency components will
affect the other replicas in the spectrum
• Even with an ideal low pass filter, the original
signal cannot be reconstructed. This is the so-
called alias effect
• Restate the Shannon Sampling Theorem for
general aperiodic signals
ωˆ ≤ π ⇒ 2πf maxTs ≤ π
⇒ 2πf max ≤ πf s or f s ≥ 2 f max
10
Shannon Sampling Theorem
A continuous-time aperiodic signal x(t) with

frequencies no higher than fmax can be
reconstructed exactly from its samples x[n] =
x(nTs) if the samples are taken at a rate fs =
1/Ts that is greater than 2fmax
Nyquist
NyquistFrequency
Frequency
11
Real Examples
Resulted signal
Ideal low pass filter

12
Resulted signal
13
Resulted signal
14
How to Solve Aliasing

Problems?
1. Increase the sampling rate such that fs ≥ 2fmax
2. Use anti-aliasing filter first
Pre-filter
Pre-filterthe
theinput
inputsignal
signalsuch
suchthat
thatititwill
willnever
never
has
hasfrequency
frequencycomponents
componentsbeyond
beyond|π|
|π|
Pre- A/D Digital Signal D/A Post-

filter Processor filter
15

x(t) X(ω)
1
0 t -B 0 B ω
X(ω)
Anti-aliasing 1
filter
ω
x[n] X p (ωˆ )
n
0 | |←Ts -2π 0 2π ωˆ = ωTs
16
With anti-aliasing filter Without anti-aliasing filter

X p (ωˆ ) X p (ωˆ )
ωˆ = ωTs ωˆ = ωTs
Ideal low Ideal low
pass filter pass filter
X’(ω) X’(ω)
ω ω
Look better
17
Hear the effect!
Original 8kHz 8kHz

44.1kHz sampling sampling
sampling with with anti-
aliasing aliasing filter
18
Discrete Fourier Transform

• Spectrum of aperiodic discrete-time signals is
periodic and continuous
• Difficult to be handled by computer
• Since the spectrum is periodic, there’s no point to
keep all periods – one period is enough
• Computer cannot handle continuous data, we can
only keep some samples of the spectrum
• Interesting enough, such requirements lead to a
very simple way to find the spectrum of signals
⇒ Discrete Fourier Transform
19
• Recall the Fourier transform of an aperiodic

discrete sequence
∞
X p (ωˆ ) = ∑ x[ n ]e − jωˆ n
n= −∞
• Assume x[n] is an aperiodic sequence with N values,

i.e. {x[n] : n = 0, 1, ..., N-1}
N −1
X p (ωˆ ) = ∑ x[ n ]e − jωˆ n
n=0
20
• If we are now interested only in N equally spaced

frequencies of 1 period of the Fourier spectrum, i.e.
 k .2π 
X [k ] = X p   k = 0,1, K N − 1
 N 
X p (ωˆ ) IfIfNN==13
13
k=0 k=4
2π/N ωˆ = ω T s
-2π 0 2π
21
• Now if we want to compute the value of these N

frequencies,
 k .2π 
N −1 − j   nTs
X [k ] = ∑ x[n]e  Ts N  W Nnk = e − j 2πnk / N
n=0
N −1 N −1
= ∑ x[ n ]e − jk .2πn / N
= ∑ x[ n ]W nk
N
n=0 n=0
for k = 0,1,…, N-1
Discrete Fourier Transform

22
• Discrete Fourier Transform (DFT) is exactly the

output of the Fourier Transform of an aperiodic
sequence at some particular frequencies
N −1
• Inherently
periodic since
X [k ] = ∑ x[ n ]e − j 2πnk / N
n=0
X[k+N] = X[k], N −1
although we X [k + N ] = ∑ x[ n ]e − j 2πn( k + N ) / N
always only n=0

1
consider one N −1
period of X[k] = ∑ x[ n ]e − j 2πnk / N − j 2πn
e = X [k ]
n=0
23
• If we know X[k], we can reconstruct back the signal

x[n] via the inverse discrete Fourier transform
N −1
1
x[n] =
N
∑ X [ k ]e j 2πnk / N
n=0
N −1
1
=
N
∑ X [ k ]
W − nk
N for n = 0,1,…, N-1
n=0
Inverse Discrete Fourier Transform

24
• It can be proven as follows:
N −1 N −1 N −1
1 1
N
∑ X [ k ]e jk 2πn / N
=
N
∑ ∑ x[ m ]e − jk 2πm / N jk 2πn / N
e
k =0 k =0 m =0
N −1 N −1
1
=
N
∑ ∑ x[ m ]e jk .2π ( n − m ) / N
k =0 m =0
N −1 N −1 jk .2π ( n− m ) / N 
1
0

if n ≠ m =
N
∑ x[m ] ∑ e 

N otherwise m =0  k =0
= x[n] for n = 0,1,..., N − 1
25
• Although DFT gives exact frequency response of a

signal, sometimes it may not give the desired
spectrum
Fourier
• Example One period of
Transform
X p (ωˆ )
x[n]
n 10
X[k] if N = 10
0 9 So different from
NN==10
10 X p (ωˆ )
k
DFT
26
• Need improved resolution

• Achieve by padding zero to the end of x[n] to make
N bigger
x[n]
Pad
Pad40
40zeros
zeros
n
0 9 50
x[n] = {1 1 1 1 1 1 1 1 1 1 0 0 0 0 … 0}
40 zeros
27
N
N == 50
50
28
N
N == 100
100
29
Exercise
Given that x[n] is defined in the following figure,
determine its spectrum using DFT with N = 4
x[n]
2 2
1 1
n
0 3
30

DFT PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

DFT PDF

Uploaded by

Copyright:

Available Formats

Signal Processing Fundamentals – Part I

Spectrum Analysis and Filtering

5. Fourier Transform and Spectrum

• Most signals are not periodic

• Need another tool to find the spectrum of non-

Fourier Transform of Discrete-

• The spectrum of x[n] is given by: ωˆ = ωTs

Aperiodic Signals have Periodic

• If x(t) has a spectrum of X(ω) and x[n] = x(nTs) has

Time Domain Frequency Domain

Time Domain Frequency Domain

Time Domain Frequency Domain

• If the signal has frequency components beyond |π|,

Shannon Sampling Theorem

A continuous-time aperiodic signal x(t) with

Ideal low pass filter

Time Domain Frequency Domain

Ideal low pass filter

Time Domain Frequency Domain

Ideal low pass filter

How to Solve Aliasing

Pre- A/D Digital Signal D/A Post-

Time Domain Frequency Domain

With anti-aliasing filter Without anti-aliasing filter

Hear the effect!

Original 8kHz 8kHz

Discrete Fourier Transform

• Recall the Fourier transform of an aperiodic

• Assume x[n] is an aperiodic sequence with N values,

• If we are now interested only in N equally spaced

• Now if we want to compute the value of these N

Discrete Fourier Transform

• Discrete Fourier Transform (DFT) is exactly the

always only n=0

• If we know X[k], we can reconstruct back the signal

Inverse Discrete Fourier Transform

• It can be proven as follows:

• Although DFT gives exact frequency response of a

• Need improved resolution

You might also like