The Fourier Transform

& the DFT

• Fourier transform 
 jt
F ( j )   f (t ).e .dt


• Take N samples of f (t ) from 0 to .(N-1)T

• Can F ( j ) be estimated from these?
• Estimate based on rectangular
approximation of integral
N 1
ˆ
F ( j )   f (nT ).e  jnT
.T
n 0

1 Professor A G Constantinides
 The Fourier Transform
& the DFT
• Estimate based on rectangular
approximation of integral
N 1
Fˆ ( j )   f (nT
 
).e  jnT
.T
n 0
• Now take N samples of Fˆ ( j ) at
multiples of 2
0 
NT
2
• Note that Fˆ ( j ) is of period
T

2 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• Then N 1 j
2
.k .n.T
Fˆ ( jk0 )   T . f ( nT ).e NT
n 0
N 1
 T .  f ( nT ).WN  nk  2 
WN  exp j 
n 0  N 

• i.e. DFT relationship where


d (n)  T . f T .n mod N 
• has a DFT D(k) such that
D(k )  Fˆ ( jk0 )
3 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• How good is the estimate?
• Let period of DFT operation be T0  NT
• Hence T  2
0
0
2
• Similarly N0   0 
T
• There are two approximations involved in
estimating F ( j )
• (i)f (t ) is sampled leading to aliasing
for non-bandlimited signals.
4
• (ii) N samples only areProfessor
retained
A G Constantinides
 The Fourier Transform
 

& the DFT

• Point (i) : Sampling yields a new signal

g (n)  T . f (nT )
• With Fourier Transform

G ( j )   F ( j  jk . 0 )
k  

• i.e. it may be a poor approximation toF ( j )

• Point (ii) : Retaining samples 0, N-1 is
effectively windowing the data by
5
 ( n)  1 0  n  N  1Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• Fourier Transform of window
N 1
j .T sin( NT 2)
W ( j )  e 2 .
sin(T 2)
• Actual signal used in DFT is h(n)  g (n).w(n)
• i.e. this leads to convolution in frequency
domain
T 0
H ( j )   G ( j ).W ( j  j ).d
2 0

6 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• Note that

 jnT
H ( j )   T . f (nT ).w(n).e
n  
N 1
  T . f (nT ).w(n).e  jnT  Fˆ ( j )
n 0
• Ie H ( j ) is the estimate of
0  
ˆ T  
F ( j )     F ( j  jk 0 ).W [ j  j ].d
2 0 k  

7 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• Since W ( j ) is periodic with period  0 its
contributions to above may be taken over
the entire domain as
1 
Fˆ ( j )   F ( j ).T .W [ j (   )d
2 
• i.e. the estimate Fˆ ( j ) is the convolution
between the desired transform and V ( j )
( N 1)
j T sin( NT 2)
V ( j )  T .e 2 .
sin(T 2)
8 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• The main lobe ofV ( j ) 2
2  20
is of width NT

• hence convolution "smears" or "blurs" a

step to the same width.
• For greater bandwidth,  0 must be
increased (or T decreased).
• To increase resolution we must decrease 0
• for a fixed T this implies N must increase.

9 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• Frequency resolution is the Minimum
separation between two sinusoids,
resolvable in frequency

• Because of the convolution two impulses in

frequency will be smeared, so that if they
are to be resolvable they must be separated
by at least one frequency bin ie 2
NT
10 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• If maximum frequency in the signal is max
and the required resolution is  then
2 2
 max   2.
2T NT
4 4max
• thus N 
4T 
• For example (rad/s)
  2  1 max  2  103
• then N  4000
11 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

• In practice Fˆ ( j ) is required within a
prescibed resolution and bandwidth.
•   F ( j ) max
• Let be limited
 to , and the
required resolution to be or better.
•  
• Then  0  2max 20  
• 0  
  max T
12
• Hence 2 T i.e. Professor A G Constantinides
max
 The Fourier Transform
 

& the DFT

• Ie  2  4max
20  2    N
 NT  
• If f (t ) is assumed to be of duration tmax
then again  
0
 max  T
2 max
2
• Set T0   NT
0
T0 T0 NT
• and  tmax    tmax
13 2 2 2 Professor A G Constantinides
 The Fourier Transform
 

& the DFT

2tmax
• i.e. N
T

2tmax .max
• Hence N 


tmax .max
where time – bandwidth product
14
of signal Professor A G Constantinides