Professional Documents
Culture Documents
Applications of Fourier
Representations
Chih-Wei Liu
Outline
Introduction
Fourier Transform of Periodic Signals
Convolution/Multiplication with (Non-)Periodic Signals
Fourier Transform of Discrete-Time Signals
Sampling
Reconstruction of Continuous-Time Signals
Discrete-Time Processing of Continuous-Time Signals
Fourier Series of Finite-Duration Nonperiodic Signals
2
Introduction
Four classes of signals in Fourier representations
Continuous- and discrete-time signals
Periodic and nonperiodc signals
In order to use Fourier methods to analyze a general system
involving a mixing of noperiodic (says, impulse response) and
periodic (input) signals, we must build bridges between Fourier
representations of different classes of signals
FT/DTFT are most commonly used for analysis applications
We must develop FT/DTFT representations of periodic signals
DTFS is the primary representation used for computational
applications (the only one can be evaluated on a computer)
Use DTFS to represent the FT, FS, and DTFT
3
FT Representations of Periodic Signals
Recall the FS representation of a periodic signal x(t): x(t ) X [k ]e 0
jk t
k
Note that 1 FT
2 ( )
Using the frequency-shift property, we have e jk0t
FT
2 k0
Let’s take the FT of x(t):
jk0t
FT x (t ) X [k ]e X [k ]FT e jk0t
2 X [k ] ( k0 )
k k k
x (t ) X [k ]e
k
jk0t
FT
X j 2 X [k ] k
k
0
4
FS and FT Representations of a
Periodic Continuous-Time Signal
0
X ( j )
6
4
2
0 0 0 0
2
5
Example 4.1 FT of a Cosine
Find the FT representation of x(t) = cos(ot)
<Sol.> 1
, k 1
0 X k 2
1. FS of x(t): cos t FS ;0
0, k 1
cos( 0t ) X ( j )
2 2 0 0
0 0
6
Example 4.2 FT of a Unit Impulse Train
Find the FT of the impulse train pt t nT
<Sol.>
n
Pk t e jk0t dt 1 / T
1 T /2
T T / 2
3. FT of p(t): P( j)
2
2
P j k
0 T
T k
4 2 2 4
T T T T
7
Relating the DTFT to the DTFS
Recall the DTFS representation of a periodic DT signal x[n]:
N 1
x n X [k ]e jk 0n
k 0
Note that the inverse DTFT of a frequency shifted impulse, i.e. (-k), is
a complex sinusoid. With one period of e jk0n , we have
1 jk0n DTFT
e ( k0 ) , k0
2
Let’s construct an infinite series of shifted impulses separated by 2, i.e. to
obtain the following 2-periodic function
1 jk0n DTFT
e ( k0 m 2 )
2 m
k 0 k 0 m
8 Since X[k] is N periodic and N0 = 2, we may combine the two sums
Relating the DTFT to the DTFS
N 1
x n X [k ]e jk 0n
DTFT
X e j
2 X [k ] k 0
k 0 k
Weighting factor = 2
DTFS DTFT
4
4
2
The DTFS X[k] and the corresponding
DTFT X(e j) have similar shape.
4 2 4
9 2
Example 4.3 DTFT of a Periodic Signal
Determine the inverse DTFT of the frequency-domain representation depicted in the
following figure, where Ω1 = π /N.
2
3 2 2 3 4
2
<Sol.>
1. We express one period of X(e j) as
X e j
1 1
1 1 ,
2j 2j
from which we infer that 2. The inverse DTFT:
1/ 4 j , k 1 N 1
X k 1/ 4 j , k 1 x[n ] X [k ]e jk0n
k 0
0, otherwise on 1 k N 2
1 j1n j1n
x n
1
2 2 j e e
1
2 sin 1n
10
Outline
Introduction
Fourier Transform of Periodic Signals
Convolution/Multiplication with (Non-)Periodic Signals
Fourier Transform of Discrete-Time Signals
Sampling
Reconstruction of Continuous-Time Signals
Discrete-Time Processing of Continuous-Time Signals
Fourier Series of Finite-Duration Nonperiodic Signals
11
Convolution and Multiplication with
Mixtures of Periodic and Nonperiodic Signals
Example: a periodic signal fed into a stable, nonperiodic impulse response
Periodic input x(t) y(t) = x(t) h(t)
Stable filter
Nonperiodic impulse response h(t)
y t x t * h t Y j 2
FT
X [k ] k H j
k
0
12
y t x t * h t Y j 2
FT
H jk X [k ] k
k
0 0
Convolution with Mixtures of Periodic and
Nonperiodic Signals
y t x t * h t Y j 2
FT
H jk X [k ] k
k
0 0
X ( j ) Magnitude
2 X [0]
2 X [ 2] Y ( j )
adjustment
0 0
2 X [0] H ( j 0)
4 0 2 0 2 0 4 0 2 X [2 ] H ( j 2 0 )
2 X [1] 0 0
H ( j ) 4 0 2 0 2 0 4 0
2 X [1] H ( j 0 )
4 0 2 0 0 0 2 0 4 0
13
Example 4.4
A periodic square wave is applied to a system with impulse response h(t)=(1/(t))sin(t.
Use the convolution property to find the output of this system.
<Sol.>
1. The frequency response of the LTI system is by taking the FT of the impulse response h(t):
1,
h t
FT
H j H ( j )
0,
2. FT of the periodic square wave:
2 sin k / 2
X j
X ( j )
k
k k 2
3. The FT of the system output is Y(j) = H(j)X(j):
Y j 2 2
2 2
2 2
H(j) acts as a low-pass filter, passing harmonics
at /2, 0, and /2, while suppressing all others.
4. Inverse
14 FT of Y(j): y t 1 / 2 2 / cos t / 2
2 2
Convolution with
Mixtures of Periodic and Nonperiodic Signals
Periodic input x[n] y[n] = x[n] h[n]
Stable filter
Nonperiodic impulse response h[n]
For DT signals:
If x[n] is DT periodic signal, then
N 1
x n X [k ]e jk 0n
DTFT
X e j
2 X [k ] k 0
k 0 k
y n x n * b n Y e
DTFT j
2 H e X [k ] k .
jk 0
0
k
The form of Y(e j) indicates that y[n] is also periodic with the same period as x[n].
15
Multiplication of
Periodic and Nonperiodic Signals
Continuous-time case:
Original multiplication property:
1
y t g t x t
FT
Y j G j X ( j )
2
If x(t) is periodic. The FT of x(t) is
x (t ) X [k ]e
k
jk0t
FT
X j 2 X [k ] k
k
0
y t g t x t Y j G ( j ) X [k ] ( k0 )
FT
k
y t g t x t Y j
FT
X [k ]G j( k )
k
0
G ( j )
Y ( j )
G ( j ) G ( j )
X ( j ) 2 0 2 0
0 0
2 0 2 0
17
Example 4.4
Consider a system with output y(t) = g(t)x(t).
Let x(t) be the square wave and g(t) = cos(t /2), Find Y(j) in terms of G(j).
<Sol.>
FS representation of the square wave:
sin k / 2
x t
FS ; / 2
X k
k
And, we have G j 1 / 2 1 / 2
sin k / 2
Y j 1 / 2 k / 2 1 / 2 k / 2
k k
18
Example 4.6 AM Radio
A simplified AM transmitter and receiver are depicted in Fig. 4.13(a). The effect of propagation
and channel noise are ignored in this system. The signal at the receiver antenna, r(t), is
assumed equal to the transmitted signal. The passband of the low-pass filter in the receiver is
equal to the message bandwidth, W < < W. Analyze this system in the frequency domain.
<Sol.>
2C 2C
Y ( j )
R ( j )
C C
20
Multiplication of
Periodic and Nonperiodic Signals
Discrete-time case:
Original multiplication property:
1
y[n ] x[n ]z[n ] Y ( e j )
DTFT
X ( e j ) Z ( e j )
2
If x[n] is periodic. The DTFS of x[n] and its DTFT is
N 1
x n X [k ]e jk 0n
DTFT
X e j
2 X [k ] k 0
k 0 k
1
j
Y (e ) X ( e ) Z ( e ) X [k ] ( k0 ) Z ( e j ( ) )d
j j
2
k
In any 2 interval of , there are exactly N multiples of the form ( 0),
since 0= 2/N.
N 1 N 1
Y ( e ) X [k ] ( k0 ) Z ( e
j j ( )
)d X [k ]Z ( e j ( k0 ) )
k 0 k 0
X [k ]Z e
N 1
y n x n z n Y e
DTFT j j k 0
21
k 0
Example 4.7 Windowing Effect
Windowing (or truncating): one common data-processing applications, which access only
to a portion of a data record.
Consider the signal xn cos
7 9
n cos n
16 16
Using only the 2M+1 values of x[n], n≦ M, evaluate the effect of computing the DTFT.
<Sol.>
N 1
1. Since x n X [k ]e jk 0n
DTFT
X e j
2 X [k ] k 0
k 0 k
Then X e j 9 7 7 9
16 16 16 16
which consists of impulses at ± 7/16 and ± 9/16.
1, n M
2. Define a signal y[n] = x[n]w[n], where wn
0, n M
1
y[n ] x[n ]z[n ] Y ( e j )
DTFT
X ( e j ) Z ( e j )
2
22 Y e W e
j 1
2
j 9 /16
W e
j 7 /16
W e
j 7 /16
W e
j 9 /16
3. The windowing introduces replicas of W(e j ) centered at the frequencies 7/16 and
9/16, instead of the impulses that are present in X(e j ) .
4. We may view this state of affairs as a smearing of broadening of the original impulses
5. The energy in Y(e j ) is now smeared over a band centered on the frequencies of the
cosines
n
Define the continuous time signal x(t) with the Fourier transform X(j):
X j X e j
|
Ts x n e jTs n
n
Taking the inverse FT of X(j), using the FT pair t nTs e jTs n
FT
24
x t x n t nT .
n
s
Relating the FT to the DTFT
x t
n
x n t nTs
FT
X j
n
x n e jTsn ,
2 2
Ts
X d ( j )
2
2 2
25 Ts Ts
Example 4.8
Determine the FT pair associated with the DTFT pair
X e j
1
x n a n u n
DTFT
1 ae j
This pair is derived in Example 3.17 assuming that |a| < 1 so the DTFT converges.
<Sol.>
1. We first define the continuous-time signal x t t nTs
n
a
n 0
2. Using = Ts gives
1
x t
FT
X j
1 ae jTs
26
Relating the FT to the DTFS
Suppose that x[n] is an N-periodic signal, then
N 1
x n X [k ]e jk 0n
DTFT
X e j
2 X [k ] k
0
k 0 k
Use the scaling property of the impulse, (a) = (1/a)(), to rewrite X(j) as
2 k 0
X j T
X k
Ts k s
27
2 k 0
x (t ) x[n ] (t nTs ) X ( j )
FT
X [k ] T
n Ts s
1. Recall that X[k] is N-periodic, which implies that X(j) is periodic with periodic
N0/Ts = 2/Ts
2. Recall that x[n] is N-periodic, which implies that x(t) is periodic with periodic NTs
0
X d ( j )
2 2
Ts 2 Ts
NTs
X[k] and x[n] are N-periodic functions
x(t) and X(j) are periodic impulse trains with period NTs and 2/Ts.
28
Outline
Introduction
Fourier Transform of Periodic Signals
Convolution/Multiplication with (Non-)Periodic Signals
Fourier Transform of Discrete-Time Signals
Sampling
Reconstruction of Continuous-Time Signals
Discrete-Time Processing of Continuous-Time Signals
Fourier Series of Finite-Duration Nonperiodic Signals
29
Sampling
We use FT representation of discrete-time signals to analyze the effects of
uniformly sampling a signal.
Sampling the (continuous-time) signal is often performed in order to
manipulate the signal on a computer.
CT signals DT signals
Sampling
x[n ] x (nTs )
Let x(t) be a CT signal and x[n], a DT signal, is the “samples” of x(t) at
integer multiples of a sampling interval.
p t t nTs .
n
x t xn t nT
s
30
= n
Sampling Process
We note that
x (t ) x[n] (t nT ) x(nT ) (t nT )
n
s
n
s s
x(t ) (t nT ) x(t ) (t nT ) x(t ) p(t )
n
s
n
s
Then 1
x (t ) x (t ) p(t )
FT
X ( j ) X ( j ) P( j )
2
Example 4.2
1 2
X ( j )
2
X ( j )
Ts
( k )
k
s s = 2/Ts
1
Ts
X ( j( k ))
k
s
31
Sampling Theorem
1
x (t ) x[n ] (t nTs ) X ( j )
FT
X ( j( k )) s
n Ts k
The shifted version of X(j) may overlap with each other if s is not large
enough, compared with the bandwidth of X(j), i.e. W.
X ( j )
X d ( j)
2s s s 2s
sW < W
X d ( j ) sW > W X d ( j)
aliasing
2 s s ss
2 s 3s 2s s s 2s 3s
s W
As Ts increases or s decreases, the shifted replicas of X(j) moves closer together, finally
32
overlapping one another when s < 2W.
Example 4.9
Consider the effect of sampling the sinusoidal signal x t cos t
Determine the FT of the sampled signal for Ts = 1 / 4.
1
<Sol.>
Recall that
x (t )
n
x[n ] (t nTs )
FT
X ( j )
T
X ( j( k )) s
s k
1. First find that X ( j ) ( ) ( )
2. Then, we have X j
Ts
k k
k
s s
Ts = ¼.
33
X d ( j )
2
Ts = 1
4 2 2 4
for Ts = 3/2
X d ( j )
2 / 3
3 2
2 3
3 3
34
Down-Sampling:
Sampling Discrete-Time Signals
1
Recall x (t )
n
x[ n ] ( t nTs ) X FT
( j )
Ts
X ( j( k ))
k
s
qTs
k
X ( j (
q
s ))
n
k m
Let l , 0 m q 1 , then
q q
1
k 1 q 1 1
m
Y ( j ) X ( j ( s ))
q m 0 Ts
X ( j ( ls s )
qTs k q l q
1 q 1 m
X ( j ( s ))
q m 0 q
35
Down-Sampling:
Sampling Discrete-Time Signals
Since y[n ]
DTFT
Y e j Y ( j ) Y ( j ) , therefore
Ts ' qTs
1 q 1 m 1 q 1 m
Y ( e ) X ( j ( s ))
j
X ( j( s ))
q m 0 q
q m 0 qTs q
qTs
1 q 1 m 2 1 q1 j m
X ( j( )) X ( ( 2 ))
q m 0 qTs q Ts q m 0 Ts q q
Use X e X
j
( j ) , we have 1 q 1 1
j ( 2 m )
Ts Y ( e j
)
q m 0
X (e q
)
Summary
1 k
y (t ) x[nq] (t nqTs ) Y ( j )
FT
X ( j ( s ))
n qTs k q
1
1 q 1
y[n ] x[ qn ] Y e X ( e q
j ( 2 m )
DTFT j
)
36
q m 0
Down-Sampling by q
X q (e j (( q1)2 ) )
q
4 2 2 4
2 q 2 q
X d (e j (2 ) )
q
2
Figure 4.29
Effect of subsampling on the DTFT. (a) Original signal spectrum. (b) m = 0 term in Eq. (4.27)
(c) m = 1 term in Eq. (4.27). (d) m = q – 1 term in Eq. (4.27). (e) Y(ej), assuming that W < /q.
(f) Y(ej), assuming that W > /q.
Not so easy !! The samples of a signal do not always uniquely determine the
corresponding continuous-time signal
X ( j )
X ( j )
m m s m m s
(a)
Ideal low-pass filter
H r ( j )
X ( j ) X ( j ) H r ( j )
Ideal reconstruction:
Ts , s 2
H r ( j ) s / 2 s / 2
0 otherwise Non-causal !!
39 s: Nyquist frequency the perfect-reconstruction cannot be
implemented in any practical system
X ( j ) X ( j ) H r ( j ) x(t ) x (t ) hr (t )
H r ( j )
s t
Ts sin( ) t
2
hr (t ) sin c( )
s / 2 s / 2
t Ts
H r ( j )
hr (t ) FT
x (t ) x (t ) hr (t ) x[n ] (t nTs ) hr (t )
n
t nTs
x[n]hr (t nTs )
n
n
x[n ] sin c(
Ts
)
X (t )
40
A Practical Reconstruction –
Zero-Order Hold Filter
The zero-order hold filter, which holds the input value for Ts seconds.
m m
s s 2 s
X ( j ) arg H O ( j )
s s 2 s
s m m s
X O ( j )
s m m s 2 s
Ts seconds holding for x[n] Time shift of Ts/2 in xo(t) for distortion-1
Stair-step approximation main reason for distortion-2 & 3
1. Using Equalizer for non-flattened mag. in passband
2. Using anti-image filter
Compensation filter
Both distortions 1 and 2 are reduced by increasing s or, equivalently, decreasing Ts,
42 e.g. oversampling !!
Compensation Filter
The frequency response: Equalizer
H c ( j )
Anti-imaging
m s m m s m
Ts
2sin(T / 2) , m
H c ( j ) s
0, s m
43
Example 4.13
In this example, we explore the benefits of oversampling in reconstructing a continuous-time
audio signal using an audio compact disc player. Assume that the maximum signal frequency is
fm = 20 kHz. Consider two cases:
(a) reconstruction using the standard digital audio rate of 1/Ts1 = 44.1 kHz, and
(b) reconstruction using eight-times oversampling, for an effective sampling rate of 1/Ts2 =
352.8 kHz.
In each case, determine the constraints on the magnitude response of an anti-imaging filter so
that the overall magnitude response of the zero-order hold reconstruction system is between
0.99 and 1.01 in the signal passband and the images of the original signal’s spectrum centered
at multiples of the sampling frequency are attenuated by a factor of 10 3 or more.
<Sol.>
0.99 1.01
H 'c ( jf ) , 20 kHz f 20 kHz
H 'o ( jf ) H 'o ( jf )
44
45
Outline
Introduction
Fourier Transform of Periodic Signals
Convolution/Multiplication with (Non-)Periodic Signals
Fourier Transform of Discrete-Time Signals
Sampling
Reconstruction of Continuous-Time Signals
Discrete-Time Processing of Continuous-Time Signals
Fourier Series of Finite-Duration Nonperiodic Signals
46
Discrete-Time Signal Processing
Several advantages for DSP algorithm
Signal manipulations are more easily performed by using arithmetic
operations of a computer than through the use of analog components
DSP systems are easily modified in real-time
Direct dependence of the dynamic range and signal-to-noise ratio on the
number of bits used to represent the discrete-time signal
Easily implement the decimation and the interpolation
H a ( j ) H O ( j ) H c ( j )
G ( j )
47
System Response Analysis
Assume that the discrete-time processing operation is represented by a
DT system with frequency response H(e j) = Ts , Ts sampling interval
H a ( j ) H O ( j ) H c ( j )
X a ( j ) H a ( j ) X ( j )
X j k
1
X ( j ) a s
Ts k
H j k X j k
1
a s s
Ts k
Y ( j ) H e H a j ks X j ks
1 jTs
Ts k
Y ( j ) H o ( j ) H c ( j ) H e H a j ks X j ks
1 jTs
48 Ts k
The anti-imaging filter Hc(j) eliminates frequency components above s/2, hence
eliminating all the terms in the infinite sum except for the k = 0 term.
Therefore, we have
H o ( j ) H c ( j ) H e jTs H a ( j ) X ( j )
1
Y ( j )
Ts
The overall system is equivalent to a continuous-time LTI system having the response:
H o ( j ) H c ( j ) H e jTs H a ( j )
1
G ( j )
Ts
49
Oversampling A high sampling rate leads to high computation cost
Sampling !!
4 2 2 4
x2[n] g[n]
4 2 2 4
4 2 2 4
51
The maximum frequency component of X2(e j ) satisfies WTs2 < /q
Decimation Filter
Decimation filter hd[n] : a low-pass filter prevents from aliasing problem
when downsampling by q.
j 1, Ts
H d (e ) q
0, otherwise
52
Interpolation
Interpolation increases the sampling rate and requires that we somehow
produce values between two samples of the signal
2 2
q-fold squeeze
2
4
2
WTs1 WTs1 2 4 2
q q q q q q
53 Image occurs !!
Interpolation Filter
Decimation filter hi[n] : a low-pass filter prevents from image problem when
upsampling by q.
Discrete-time
Upsampling xi[n]
x[n] by q
low pass
Hi(ej)
1, T
H i ( e j ) s
q
0, otherwise
2 2
q q
2 2
54
Outline
Introduction
Fourier Transform of Periodic Signals
Convolution/Multiplication with (Non-)Periodic Signals
Fourier Transform of Discrete-Time Signals
Sampling
Reconstruction of Continuous-Time Signals
Discrete-Time Processing of Continuous-Time Signals
Fourier Series of Finite-Duration Nonperiodic Signals
55
Why DTFS for Finite Nonperiodic Signals?
The primary motivation is for the numerical computation of Fourier
transform
DTFS is only Fourier representation that can be evaluated numerically !!
Let x[n] be a finite-duration signal of length M
Introduce a periodic DT signal x[n ] with period NM
zero padding
M 1
j
X (e ) x n e
n 0
j n
56
2 2
Relating the DTFS to the DTFT
Consider a periodic DT signal x[n ] with period NM, then
N 1 M 1
1 1
~ ~ jk0 n ~ jk 0n
X [k ] x [ n ]e Apply x[n], we have X [k ] x[ n ]e
N n 0 = 2/N
M 1 o
N n 0
j
Recall that X ( e ) x n e
n 0
j n
, we conclude that
~ 1
X [k ] X ( e j )
N k 0
The DTFS coefficients of x[n ] are samples of the DTFT of x[n], divided by N and
evaluated at intervals of 2/N.
58
e j ( ko 3 / 8)16 e j ( ko 3 / 8)16 e j ( ko 3 / 8)16
X k j ( k 3 / 8)2 j ( k 3 / 8)2
j ( k o 3 / 8) e
1 o
e o
2 Ne 2
e j ( ko 3 / 8)16 e j ( ko 3 / 8)16 e j ( ko 3 / 8)16
j ( k o 3 / 8) / 2 j ( k 0 3 / 8)2 j ( k o 3 / 8)2
2 Ne e e
e j 31( ko 3 / 8) / 2 sin 16 k o 3 8
2N sin k o 3 8 2
e j 31( ko 3 / 8) / 2 sin 16 k o 3 8
2N sin k o 3 8 2
59
Relating the FS to the FT
The relationship between the FS and the FT of a finite-duration non-
periodic continuous-time signal is analogous to that of discrete-time case
Let x(t) have duration T0, so that x (t ) 0, t 0 or t T0
Construct a periodic signal x (t ) x (t mT ) with T To
m
Consider the FS of x (t )
1 T 11 TTo0
jkot jkjko0tt
X [k ] x (t )e
dt
x
x (
( tt))ee dt
dt
T 0 TT 0 0
Recall that
X j x (t )e
T
x (t )
DT jt
dt x (t )e jt dt
0
1
X [k ] X ( j )
T ko