Fourier Series/Transform

Review
 Fourier Review

Fourier and inverse Fourier are the mathematical

tools that are used to switch from time-domain to
frequency-domain and vice versa, in order to view
a signal.
This conversion between time and frequency
domains is very helpful in enhancing the
understanding of signals and systems.
This is called Fourier analysis.
 Fourier Review

Fourier Series and Transforms try to form a signal out

of sinusoids. These sinusoids have a specific
frequency and go on forever. That is, time series
which is represented by points in time will now be
represented by points in frequency.
This is why we use the terms “Fourier domain” and
“frequency domain” interchangeably.
Reminder:

ae ( jt )  a cos(t )  ja sin(t )

 What Transform, When?

Start Discrete or Periodic Transform

Domain Continuous
Time Discrete Yes DTFS
Time Discrete No DTFT
Time Continuous Yes FS
Time Continuous No FT
Frequency Discrete Yes I-DTFS
Frequency Discrete No I-DTFT
Frequency Continuous Yes I-FS
Frequency Continuous No I-FT
 Discrete Time Fourier Series

1
DTFS: X [k ]  n  N  x[n]e  jk0 n

N
1
I-DTFS: x[n]  k  N  X [k ]e jk0 n
N

X[k] and x[n] have period N

w0 = 2π/N
 Discrete Time Fourier Transform

DTFT: X [ e j ]   x[
n  
n ]e  j n


1

j jn
I-DTFT: x[n]  X ( e ) e d
2 

X[k] has period 2π

 Fourier Series

1
FS: X [k ]   x(t )e  jk 0t
dt
T T 

I-FS: x(t )   X [k ]e
k  
jk o t

X(t) has period T

w0 = 2π/T
 Fourier Transform



 jt
FT: X ( j )  x (t ) e dt


1

j t
I-FT: x(t )  X ( j  ) e d
2 

The Fourier Transform is the general transform, it can

handle periodic and non-periodic signals. For a periodic
signal it can be thought of as a transformation of the
Fourier Series

X ( j )  2  X [k ] (  n 0 )
k  
 Fourier Series
1
X [k ]   x(t )e  jk 0t dt
T T 
1 1
X [k ]   x(t ) cos( o kt )dt  j  x(t ) sin( o kt )dt
T T  T T 
               
Ak Bk

 tan 1  
Bk
Ak 
X [k ]  Ak2  Bk2 e      X k e k

    k
Xk e
 Fourier Series – Real Signals

1 1
X [k ]   x(t ) cos( o kt )dt  j  x(t ) sin( o kt )dt
T T  T T 
               
Ak Bk

If x(t) is real valued: Ak = A-k Bk = -B-k

   X [0]    A 
  
x(t )   X [k ]e jk o t
 X [0]   X [k ]e jk o t
 X [k ]e  jk o t
k  jBk  e jk ot   A k  jB k  e  jkot
k   k 1 k 1

      
 
x(t )  X [0]    Ak  jBk  e jk ot   Ak  jBk  e  jkot  X [0]   Ak e jkot  e  jkot  jBk e jk ot  e  jkot
k 1 k 1

 
 
x(t )  X [0]  2  Ak cos(k ot )  Bk sin(k ot )   X [0]  2 Re X [k ]e jkot
k 1 k 1


  X [ 0]  2  X [k ] cos(k t  

x(t )  X [0]  2 Re X [k ] e j k e jk ot 
k 1 0 k )
k 1
 Fourier Series – Real +Even/Odd

 

x(t )  X [0]  2 Re X [k ] e j k e jkot
k 1

x(t )  X [0]  2 Re  Ak  jBk  cos(k ot )  j sin(k ot ) 
k 1

x(t )  X [0]  2  Ak cos(k ot )  Bk sin(k ot ) 
k 1

Even: f(t) = f(-t), therefore Bk = 0; Cosine Series

Odd: f(t) = -f(-t), therefore Ak = 0; Sine Series
 Cosine Fourier Series

1 j0t 1  j 0t
f (t )  cos( 0t )  e  e
2 2
Even Function
FS X [1]  X [1]  1
2
FT = 2π(FS) X ( j )   (   0 )   (   0 )
When is FT the continuous counterpart to 2πFS?
How do the Delta’s move as frequency changes?
 Sine Fourier Transform

1 j0t 1  j 0t
f (t )  sin( 0t )  e  e
2j 2j
Odd Function
FS X [1]   X [1]  1
2j
FT = 2π(FS) X ( j )  j (   0 )  j (   0 )
The Fourier Transform of an Odd Signal is Odd.
Notice the Fourier Domain graph is in jF(ω). It is imaginary.
 DC Fourier Transform

f ( t )  1  e j0 t at 0  0
DC Function
FS X [0]  1

FT (FS) X ( j )  2  X [k ]   k 
k  
0

FT X ( j )  2 ( )
The FT of a signal with a DC component is separable.
The DC component of a time signal is statistically the MEAN.
 Delta Fourier Transform

f (t )   (0)
Delta Function
FS - No Fourier Series, Not Periodic


FT X ( j )    (t )e  j 2kt dt  e  j 2k ( 0 )  1

The FT is only congruent with the FS for PERIODIC signals.
A delta has an infinitely steep rise time, therefore it has a great
deal of high frequencies
 Pulse Train Fourier Transform


f (t )    (t  nT )
n  
Function with Period T
FS X [k ]  1 for all k
T
  2k 
    
X ( j )     (t  nT )e
 j 2kt
dt   e
 j 2knT
dt  2   
T T
 n    n   k  

What happens in the Frequency Domain when the time

between pulses is shortened? When T  0? When T = 0?
 Time Window Fourier Transform

 
1, t  
f (t )   2
 rect t
0, t   2 
Not Periodic – No FS
2 sin 
FT X ( j ) 


2   Sinc  
2
sin  x 
Sinc x   Sa x  
x
 Ideal Filter Fourier Transform

x(t )  SincWt 
Not Periodic – No FS
 ,   W 
FT X ( j )   W
 0,  W W

 rect 
2W

Why is this called the “ideal filter”?
Notice similarities between this and rectangular time window, and
how W here is a counterpart to τ there in controlling width.
 Triangle Fourier Transform

 t
 
x (t )  1   , t     t

 0, t 
Not Periodic – No FS
FT   
X( j)   Sinc 
2
2

Sinc squared can never be negative. Why are we introducing

these signals? They are the foundation of most analog
communication signals.
 More Complex Example
1

0.5

0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

0.5

0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

0.5

0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

A pulse train with period (T) one second is convolved with

a time windowing function with timing (τ) of 0.5 seconds,
to produce a 50% duty cycle square wave.
 More Complex Example
The spectrum of the pulse train is:
    2k T 

X 1 ( j )  2
T k  
The spectrum of the square-wave is:
X 2 ( j )   Sinc   2

Convolution turns into Multiplication in the Freq Domain

X 1 ( j )  X 2 ( j )  2
T

Sinc  2     2k T 


k  

This turns into a line spectra, and how it changes with

changing the parameters is very informative
 Constant τ  0.5
T=2

T=4

T=8

• Amplitude DECREASES as 1/T

• Line spectra resolution INCREASES as T
• The envelope is INDEPENDENT of T
 Constant T T=2
  0.25

  0.5
 1
• Amplitude INCREASES in proportion to t
• Line spectra resolution is INDEPENDENT of t
• The spectrum SPREADS as the window shortens
!!! TIME RESOLUTION AND FREQUENCY RESOLUTION
ARE INVERSELY RELATED !!!!!!!!
 The Sampling Theorem
One of the fundamental concepts in dealing with the
representation of analog signals in the digital domain is
the Nyquist Rate, or Minimum Time-Bandwidth product.
This law states the minimum sample frequency necessary
to exactly represent an analog signal as a digital signal.
Since one of the main constraints in judging the efficiency
of a communication system is spectral efficiency, the
Nyquist rate forms a large part of the back-bone of system
design.

A real-valued band-limited signal having no spectral

components above a frequency of B Hz is determined
uniquely by its values at uniform intervals spaced no
greater than 1/2B seconds apart
 Sampling Theorem
Consider a signal f(t) sampled with an impulse train p(t)
f s (t )  f (t ) p (t )

p (t )   e jn 0 t
,  0  2 T
n  

f s (t )  f (t )  e jn 0t , Fourier Transform
n  

Fs ( )   F ( ) (n 0 )
n  

Fs ( )   F (  n 0 )
n  
Sampling Theorem Visual
Band limited signal +
spectrum

Periodic gating
function + spectrum

Size of sampling
window controls
envelope of
spectrum, sample
frequency controls
spacing of original
spectrum replicas
 Nyquist Rate
Since the periodic gating function controls the center of
the replicas and the replicas are 2W (W = 2πB) wide,
then to make sure there is no overlap:
2
 2W
T
1
T
2B
If the signal is sampled at a lower rate there will be
overlap, and in the final spectrum you won’t know if the
overlapped part is from the spectrum that is suppose to
be there or from the “ALIASED” part of the spectrum
Signals and Systems Review
 Energy and Power
Signal Energy

Ef   f (t ) f * (t )dt , UNITS [V 2 s ]


Signal Power T
2
1
Pf  lim
T  T
T f (t ) f * (t )dt , UNITS [V 2 ]

2

An energy signal cannot be a power signal, nor vice-versa

To be an energy signal:
Amplitude  0
As |Time|  inf
 Energy and Power Example
Find Ex
x(t )  A cos( 0 t   )
 2 
A
E x   A 2 cos 2 ( 0 t   )dt 
2  1  cos(2 0t  2 ) dt
 
 
2 2
At A
Ex   sin  2 0 t  2  
2 t  
4 0 t  
0
T /2 2 T /2
1 1 A
Px  lim  ( 0 t   )dt  lim  1  cos(2 0t  2 ) dt
2 2
A cos
T  T T / 2 T  T 2 T / 2

A 2T A2 A2
Px  lim   sin  0T  2   sin  0T  2   
T   2T 4T 0 2
 Parseval’s Theorem
Energy calculated in the Time domain is equal to energy
calculated in the Frequency domain.
 
1
 f ( t ) f ( t ) dt   F ( ) F ( ) d 
* *

 2 

1
F
j t
f (t ) 
* *
(  )e  d
2 
 
 1 

  
j t
f ( t ) f ( t ) dt 
*
f (t )  F * (  )e  d   dt
   2  

1 
  
  F (  )   f ( t )e  j t
f ( t ) f ( t ) dt 
* *
dt  d 
 2   



j t
F ( )  f ( t )e  dt

 
1
 f ( t ) f ( t ) dt   F ( ) F * ( ) d 
*

 2 
 Power Spectral Density
T /2
1 2
Pf  lim
T  2 T  F ( ) d 
T / 2

1
Pf 
2 Sf ( ) d 


 
1 1 1 2

2  S f ( ) d   lim
T  T 2  F ( ) d 
 
 
1 1 1 2
G f ( ) 
2  S f ( u ) du  lim
T  T 2  F ( u ) du
 

  2
F (u )
2  G f ( )   S f ( u ) du   lim
T  T
du
 
2
dG f ( ) F ( )
2  S f ( )  lim
d T  T
 PSD
Sf(ω) is the power spectral
density function, it has units
of power per Hz.

Gf(ω) is the cumulative

spectral power function, it
the amount of energy in the
signal in those components
less then ω.
 Autocorrelation
2
F ( )
S f ( )  lim
T  T

IFT S 
f 
1
2  lim
1
T
F * ( ) F ( ) e j 
d
 T 
 T / 2 T / 2
IFT S   lim
f
1 1
T 2  
*
f (t )e j t
dt  f ( t1 ) e  j t1
dt 1 e j 
d
T    T / 2 T / 2
T / 2
 1
T / 2 

IFT S   lim
f
1
T  f ( t )  f ( t1 )  2
*
e
j  (  t  t 1 )
d   dt 1 dt
T  T / 2 T / 2   
T / 2 T / 2
IFT S   lim
f
1
T  f * (t )  f ( t 1 ) (  t  t 1 ) dt 1 dt
T  T / 2 T / 2
T / 2
IFT S   lim
f
1
T  f * ( t ) f ( t   ) dt  R f ( )
T  T / 2
 Autocorrelation
Rf(τ) should look familiar in a way. It is equivalent to
convolving the function f(t) with f(-t).
T /2
1
R f ( )  lim
T  T
 f * (t ) f (t   )dt
T / 2

f * ( )  f (  )   f (t ) f (t   )dt


The autocorrelation function is often used for signal

detection in a background of random noise. When we get
into random noise it will become very evident why this is
so.
Linear Time Invariant Systems
Fundamental way of describing many components in a
communication system. Models filters, amplifiers and
equalizers very well.

Model an LTI system with the impulse response, h(t), of

the system, the response of an impulse input to the
system. The Fourier Transform of the impulse response is
the frequency transfer function.
y (t )  h(t )  x(t )
x(t) h(t) y(t) 
  h( ) x(t   )d

 Time Operators
f(t)  g(t) 

f(t-a)  g(2t) 

f(t+b)  g(t/2) 

What happens to in the Fourier domain to each of these?

 Invertibility
LTI systems are invertible
If you can determine the input given the output then a
system is called Invertible
Given input x and it’s output is y:
y(t) = 2 x(t)
Is inverted by z:
z(t) = ½ y(t) = x(t)

Not invertible:
y(t) = floor{x(t)}

!!! A non-invertible system usually maps multiple

points from the input space to the same point in the
output space.
 LTI Systems
x(t) h(t) y(t) y (t )  h(t )  x(t )

X(ω) H(ω) Y(ω) Y ( )  H ( ) X ( )

In the frequency domain the convolution integral
becomes a multiplication, and vice-versa. By assessing
the frequency domain magnitude and phase we can see
how H can effect specific frequencies differently:
j y ( )
Y ( ) e  H ( ) e j h ( )
X ( ) e j x ( )

Y ( )  H ( ) X ( ) !!! This is the beginning

of the filtering
 y ( )   h ( )   x ( ) interpretation
 LTI Systems
The Law of Superposition:
Given inputs a and b to system x, a linear system:
x(a)+x(b) = x(a+b)
Given input a and some scalar constant to system x,
x(c a) = c x(a)
The Law of Time Invariance:
Given some input function g(t) and is input to a system X produces
an output f(t)
X{g(t)} = f(t)
If g(t) is shifted in time by T0 then the output has the same shift
X{g(t-T0)} = f(t-T0)
The Law of Commutation:
Given some function g(t) and f(t)
g(t) * f(t) = f(t) * g(t)
 Ideal Filter Introduction
Frequency Response Impulse Response
1
0.08
Low Pass Filter 0.5
0.06
0.04

(LPF) 0
0.02
0
-10 -5 0 5 10 10 20 30 40 50 60

High Pass Filter 0.5

0.05
0

(HPF) 0
-0.05

-10 -5 0 5 10 10 20 30 40 50 60

BandPass Filter 0.5

0.1
0

(BPF) 0
-0.1

-10 -5 0 5 10 10 20 30 40 50 60
1
0.6

BandStop Filter 0.5

0.4
0.2

(BSF) 0
-10 -5 0 5 10
0

10 20 30 40 50 60
 Real Filters
In reality one cannot make the Brick Wall type ideal filters.
This is due to the fundamental tradeoff between time and
frequency resolution. If you have a jump in the frequency
response that is infinitesimally resolved, you’d need infinite
time to represent that.

One deals with filter specifications such as bandwidth, roll-

off, implementation complexity, passband ripple and so on
for most of this course, and for many future courses.

It is of great practical importance to understand the tradeoffs

implicit in the time-frequency bandwidth tradeoff.
 Filters cont’d
Most filters bandwidths are defined by the 3 dB point, or
where the frequency transfer response is 1/2 less then
the maximum point.
 Filter Truncation - Time
One can never implement an ideal filter because the
infinite frequency resolution requires infinite time. What
happens when you just get rid of some of the time
window?
1
0.4

0.2 0.5 W = 100

0
0
-100 -50 0 50 100 0 50 100 150 200

Ringing = Gibbs
0.4

0.2 0.5

-50 0 50
0
0 20 40 60 80 100
50 effect
1
0.4

10
0.2

Longer Time
0.5

0
0

Window, steeper
-10 -5 0 5 10 0 5 10 15 20

1
0.4

0.2

0
0.5 5 frequency roll-off
0
-6 -4 -2 0 2 4 0 2 4 6 8 10