Discrete Fourier Transformation

© PFH Private Hochschule Göttingen 1

The Fourier transform
The Discrete Time Fourier Transform (DTFT)

Fourier transform

▪ Time continuously  = 2   f
F ( j ) =  f (t )  e
− jt
 dt
−
▪ Frequency continuously

Discrete time Fourier transform f ( k ) = f ( k  Ts )

▪ Time discrete k = k =
Fd ( j ) = 
k =−
f (k  Ts )  e − j k Ts = 
k =−
f ( k )  e − j   k fs = 1
Ts
▪ Frequency continuously

Normalized angular frequency:  =   Ts = 2    f  Ts = 0 2

f =0 fs

© PFH Private Hochschule Göttingen 2

The Fourier transform
The Discrete Fourier Transform (DFT)
f ( k ) = f ( k  Ts )
Discrete time Fourier transform
k = k =
▪ Time discrete Fd ( j ) = 
k =−
f (k  Ts )  e − j k Ts = 
k =−
f ( k )  e − j   k

▪ Frequency continuously Normalized angular frequency:  =   Ts = 2    f  Ts = 0 2

Discrete Fourier transform

▪ Time discrete in N values t = 0, Ts , 2Ts , , ( N − 1) Ts
1 2 N −1
▪ Frequency discrete in N values n = 0,  s ,  s , ,  s
N N N
2  n
1 2 N −1 n =
 n = 0,  2 ,  2 , ,  2 N
N N N
N −1 N −1
Fd (n) = Fd ( jn ) = Fd ( j n ) =  f (k )  e − jk n =  f (k )  e − jk 2 n / N
k =0 k =0

© PFH Private Hochschule Göttingen 3

The Fourier transformation
Fourier transform
▪ Time continuously 
F ( j ) =
 = 2   f
 f (t )  e
− jt
 dt
▪ Frequency continuously −

Discrete time Fourier transform fs = 1

n =
Ts
▪ Time discret Fd ( j ) = 
n =−
f (n  Ts )  e − j nTs  =   Ts
n =
 = 2    f  Ts
▪ Frequency continuously Fd ( j ) = 
n =−
f ( n)  e − j   n
=0 2
f =0 fs
Discrete Fourier transform
▪ Time discrete in N values t = 0, Ts , 2Ts , , ( N − 1) Ts
1 2 N −1 n
▪ Frequency discrete in N values n = 0,  s ,  s , ,  s n =  2
N N N N
N −1
Fd (n) = Fd ( jn ) = Fd ( j n ) =  f (k )  e − j2 nk / N
k =0 © PFH Private Hochschule Göttingen 4
The discrete Fourier transform
Transition to the discrete Fourier transform
Assumptions
▪ A periodic signal is considered x(t)
▪ The sampling frequency is fs
▪ There are exactly N samples per signal period x0 , x1 , x2 , . . . , xN − 1 , xN = x0
▪ The signal period should be an integer multiple of the sampling period
N −1 j  2  n  k N −1
−
Fd ( n ) X d (n) =  xk  e N
=  xk W − k n
k =0 k =0
j  2
W =e N
N-th complex root of unity

Inverse discrete Fourier transform (IDFT)

j  2  n  k
1 N −1 1 N −1
xk =   X d ( n )  e N
=   X d ( n )  W nk
N n =0 N n =0
© PFH Private Hochschule Göttingen 5
The discrete Fourier transform
▪ The continuous Fourier transform can be
calculated by wrapping a period of the signal
around the complex unit circle and then
integrating once around it Imaginary-Part

▪ With the DFT, instead of a complete integral, j  2  n

W =e N
only a few discrete values ​are added at the j  2

regularly distributed points of the multiples of the t = nTs W =e N

Real-Part
root of unity t = N  Ts

▪ If the points are close enough together, i.e. N is

large enough, then discrete summation is
sufficient as a substitute for integration

▪ N sufficiently large corresponds to the fulfillment

of the sampling theorem

© PFH Private Hochschule Göttingen 6

The discrete Fourier transform
Transition to the discrete Fourier transform
Assumptions
▪ A periodic signal is considered x(t)
▪ The sampling frequency is fs
▪ There are exactly N samples per signal period x0 , x1 , x2 , . . . , xN − 1 , xN = x0
▪ The signal period should be an integer multiple of the sampling period
N −1 j  2  n  k N −1
−
Fd ( n ) X d (n) =  xk  e N
=  xk WNnk
k =0 k =0
j  2
−
WN = e N
N-th complex root of unity

Inverse discrete Fourier transform (IDFT)

j  2  k  n
1 N −1 1 N −1
xk =   X d ( n )  e N
=   X d ( n ) WN− k n
N n =0 N n =0
© PFH Private Hochschule Göttingen 7
The discrete Fourier transform
▪ They’re called this because, in complex arithmetic, (W )
k N
N = 1 for all k

▪ They’re vertices of a regular polygon inscribed in the unit circle of the complex plane, with one vertex
at (1,0)
j  2
−
WN = e N

j j

−j −j

Roots of unity for N = 2, N = 4, and N = 8, graphed in the complex plane

© PFH Private Hochschule Göttingen 8

The discrete Fourier transform
Inverse discrete Fourier transform (IDFT)
j  2  n  k
1 N −1
xk =   X d ( n )  e N
N n =0
Relationship to the complex Fourier coefficients
T0 2
1 −j nt
cn =  x (t )  e  dt
T0

T0 0

x ( t ) → x ( kT ) → xk
dt → T
t → kT
T0 → NT

 →
© PFH Private Hochschule Göttingen 9
The discrete Fourier transform
Relationship to the complex Fourier coefficients

x ( t ) → x ( kT ) → xk
dt → T
t → kT
T0 → NT

 →

T0 2
1 −j nt
cn =  x (t )  e  dt
T0

T0 0

2 nk N −1 nk
N −1 N −1 − j  2 
X d (n) =  xk  e
−j nk T 1 − j  2 
x
1
cn =
NT
x
k =0
k e NT
T cn =
N k =0
k e N

k =0
N

© PFH Private Hochschule Göttingen 10

The discrete Fourier transform
Relationship to the complex Fourier coefficients

x ( t ) → x ( kT ) → xk
dt → T
t → kT
T0 → NT

 →

N −1 nk N −1 nk
− j  2 
X d (n) =  xk  e
1 − j  2 
cn =
N
x
k =0
k e N N

k =0

1
cn =  X d (n)
N
© PFH Private Hochschule Göttingen 11
The discrete Fourier transform
Properties of the DFT

DFT a  x1 ( k ) + b  x2 ( k ) = a  X 1 ( n ) + b  X 2 ( n )
▪ Linearity
j 2 n
DFT  x ( n + l ) = X ( n )  e
− l

▪ Time shift
N

▪ Frequency shift IDFT  X ( n + l ) = x ( k )  e j 2 nl / N

▪ Convolution DFT  x1 ( k )  x2 ( k ) = X 1 ( n )  X 2 ( n )

▪ Multiplication DFT  x1 ( k )  x2 ( k ) = X 1 ( n )  X 2 ( n )

1 N −1
▪ Inverse DFT IDFT  X ( n ) = x ( k ) =  X ( n )  e j 2 kn / N
N n =0

© PFH Private Hochschule Göttingen 12

The discrete Fourier transform
Some special cases

➢ Periodicity in the frequency domain

➢ Periodicity in the time domain

➢ Frequency resolution increasing through zero padding

➢ Picket fence effect

➢ Leakage effect

➢ Effect of windowing
© PFH Private Hochschule Göttingen 14
The discrete Fourier transform
Periodicity in the frequency domain

Since the e-function is periodic with j2π, the following holds

e j = e j( + 2 ) = e j( + 4 ) = Fd (n) = Fd (n + N ) = Fd (n + 2 N ) =

time signal x ( k ) spectrum X ( n ) Fd ( n )

N =9 N =9

0 1 2 3 4 5 6 7 8
discrete time k discrete frequencies n
© PFH Private Hochschule Göttingen 15
The discrete Fourier transform
Periodicity in the time domain
Since the 2π-periodic e-function also occurs in a comparable way in the
inverse transformation, the time signal is also periodic in the DFT, in contrast to
the time-discrete Fourier transformation
1 N −1
F  X ( n ) = x ( k ) =  X ( n )  e j 2 kn / N
−1
x(k ) = x(k + N ) = x(k + 2 N ) =
N n =0
spectrum X ( n ) discrete time signal x ( k )
N =9
N =9

discrete frequency n diskrete time k

© PFH Private Hochschule Göttingen 16
The discrete Fourier transform
Perodicity
If x(k) is real (normal case), then the amplitude response is an even function and the phase response is
an odd function, i.e. both are already completely defined by half of the points, the other half can be
generated by mirroring:
▪ If N is even: (N/2+1) points are needed
▪ If N is odd: (N+1)/2 points are needed WN = e− j2 / N
spectrum X (n) phase X ( n )
N =9 N =9
( N + 1) / 2 = 5 ( N + 1) / 2 = 5

redundant redundant

discrete frequncy n discrete frequency n

© PFH Private Hochschule Göttingen 17
 The discrete Fourier transform
Perodicity

➢ Equivalent representations of the spectrum WN = e− j2 / N

redundant redundant

n: N −N N N
2 N 2 2 2
f: fs fs − fs fs fs
2 2 2 2
:  2 −  

© PFH Private Hochschule Göttingen 18

The discrete Fourier transform
DFT of cosine-shaped signals with complete periods
32
32 T0 =  Ts
T0 = 32  Ts T0 =  Ts 5
2
N = 32 N = 32 N = 32
WN = e− j2 / N

discrete time k discrete time k discrete time k

redundant redundant redundant

1 2 5
f = fs f = fs f = fs
32 32 32

discrete frequency n discrete frequency n discrete frequency n

© PFH Private Hochschule Göttingen 19
The discrete Fourier transform
Calculation scheme
N −1 N −1
DFT  x ( k ) = X ( n ) =  x(k )  e − j2 nk / N =  x(k ) WNnk WN = e− j2 / N
k =0 k =0

 X (0)  1 1 1 1   x(0) 
  1 WN1 WN2 WNN −1   
 X (1)     x(1) 
 X (2)  = 1 WN2 WN4 WN2( N −1)    x(2) 
     
     
 X ( N − 1)   ( N −1)2   x( N − 1) 
  1 WNN −1 WN2( N −1) WN   

The following operations are necessary to carry out this matrix-vector

multiplication:
N 2 complex additions
N 2 complex multiplications
© PFH Private Hochschule Göttingen 20
The discrete Fourier transform
Hints for the calculation
N −1 j  2  n  k N −1
−
Two-point DFT (N=2) X d (n) =  xk  e N
=  xk WNnk
k =0 k =0
j  2
j  2 −
W2 = e
−
2
= e − j = −1 WN = e N

1
X d (n) =  xk  ( −1) = x0  ( −1) + x1  ( −1) = x0 + x1  ( −1)
nk n 0 n1 n

k =0

X d (0) = x0 + x1  ( −1) = x0 + x1
0

X d (1) = x0 + x1  ( −1) = x0 − x1
1

© PFH Private Hochschule Göttingen 21

The discrete Fourier transform
Hints for the calculation
N −1 j  2  n  k N −1
−
Four-point DFT (N=4) X d (n) =  xk  e N
=  xk WNnk
k =0 k =0
j  2 j 
− − j  2
−
W4 = e 4
=e 2
=−j WN = e N

3
X d (n) =  xk  ( − j ) = x0  ( − j ) + x1  ( − j ) + x2  ( − j ) + x3  ( − j )
nk n 0 n1 n2 n3

k =0

3
X d (n) =  xk  ( − j ) = x0 + x1  ( − j ) + x2  ( −1) + x3  j n
nk n n

k =0

X d (0) = x0 + x1 + x2 + x3
X d (1) = x0 − jx1 − x2 + jx3
X d (2) = x0 − x1 + x2 − x3
X d (3) = x0 + jx1 − x2 − jx3
© PFH Private Hochschule Göttingen 22
The discrete Fourier transform
Hints for the calculation

This can also be written as a matrix multiply:

X d (0) = x0 + x1 + x2 + x3
X d (1) = x0 − jx1 − x2 + jx3
X d (2) = x0 − x1 + x2 − x3
X d (3) = x0 + jx1 − x2 − jx3

 X d (0)  1 1 1 1   x0 
 X (1)  1 − j −1 j   x 
 d =  1
 X d (2)  1 −1 1 −1  x2 
     
 d  
X (3) 1 j −1 −1  x3 
© PFH Private Hochschule Göttingen 23
The discrete Fourier transform
Hints for the calculation

To compute A quickly, we can pre-compute common subexpressions:

 X d (0)  1 1 1 1   x0 
 X (1)  1 − j −1 j   x 
 d =  1
 X d (2)  1 −1 1 −1  x2 
     
 X d (3)  1 j −1 −1  x3 

X d (0) = ( x0 + x2 ) + ( x1 + x3 )
X d (1) = ( x0 − x2 ) − j ( x1 − x3 )
X d (2) = ( x0 + x2 ) − ( x1 + x3 )
X d (3) = ( x0 − x2 ) + j ( x1 − x3 )

© PFH Private Hochschule Göttingen 24

The discrete Fourier transform
Hints for the calculation

▪ This saves a lot of adds

▪ Note that each add and multiply here is a complex (not real) operation
▪ If we use the following diagram for a complex multiply and add

p p + q X d (0) = ( x0 + x2 ) + ( x1 + x3 )
X d (1) = ( x0 − x2 ) − j ( x1 − x3 )
 X d (2) = ( x0 + x2 ) − ( x1 + x3 )
q X d (3) = ( x0 − x2 ) + j ( x1 − x3 )

© PFH Private Hochschule Göttingen 25

The discrete Fourier transform
Hints for the calculation

▪ Then we can diagram the 4-point DFT like this:

x0 x0 + x2 X d ( 0)
1 X d (0) = ( x0 + x2 ) + ( x1 + x3 )
x2 x0 − x2 X d (1) X d (1) = ( x0 − x2 ) − j ( x1 − x3 )
−1 −j
X d (2) = ( x0 + x2 ) − ( x1 + x3 )
x1 x1 + x3 X d ( 2) X d (3) = ( x0 − x2 ) + j ( x1 − x3 )
1
x3 x1 − x3 X d ( 3)
−1 j

© PFH Private Hochschule Göttingen 26

The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

▪ The FFT is a fast algorithm for computing the DFT

▪ If we take the 2-point DFT and 4-point DFT and generalize them to 8-point,
16-point, ..., 2r-point, we get the FFT algorithm
▪ To compute the DFT of an N-point sequence would take N² multiplies and
adds
▪ The FFT algorithm computes the DFT using N log(N) multiplies and adds
▪ There are many variants of the FFT algorithm
▪ One of them is the FFT algorithm for sequences whose length is a power of
two

© PFH Private Hochschule Göttingen 27

The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

▪ The diagram of an 8-point FFT, where

W8 = e
−
j  2
8
=e
−
j 
4
=
(1 − j )
2

© PFH Private Hochschule Göttingen 28

The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

▪ The diagram of an 8-point FFT

x0 X d ( 0)
x4 X d (1)
x2 X d ( 2)
x6 X d ( 3)
x1 X d ( 4)
x5 X d ( 5)
x3 X d ( 6)
x7 X d (7)
© PFH Private Hochschule Göttingen 29
The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

Butterflies and Bit-Reversal

▪ The FFT algorithm decomposes the DFT into log2N stages, each of which
consists of N=2 butterfly computations
▪ Each butterfly takes two complex numbers p and q and computes from them
two other numbers, p + αq and p − αq, where α is a complex number
▪ In the diagram of the 8-point FFT above, note that the inputs aren’t in normal
order: a0 ; a1 ; a2 ; a3 ; a4 ; a5 ; a6 ; a7
▪ They’re in the bizarre order: a0 ; a4 ; a2 ; a6 ; a1 ; a5 ; a3 ; a7

© PFH Private Hochschule Göttingen 30

The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

Butterflies and Bit-Reversal

▪ They’re in the bizarre order: a0 ; a4 ; a2 ; a6 ; a1 ; a5 ; a3 ; a7
▪ Below is a table of i and the index of the i-th input sample, ni

i 0 1 2 3 4 5 6 7
ni 0 4 2 6 1 5 3 7
i base 2 000 001 010 011 100 101 110 111
ni base 2 000 100 010 110 001 101 011 111

© PFH Private Hochschule Göttingen 31

The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

FFT Explained Using Matrix Factorization W8 = e

−
j  2
8
=e
−
j 
4
=
(1 − j )
2
 X d ( 0 )  W 0 W0 W0 W0 W0 W0 W0 W 0   x0 
   0   
 X d (1)  W W1 W2 W3 W4 W5 W6 W 7   x1 
 X d ( 2 )  W 0 W2 W4 W6 W0 W2 W4 W 6   x2 
   0   
(
 d  = W
X 3 ) W3 W6 W1 W4 W7 W2 W 5   x3 
 X ( 4 )  W 0 
W4 W0 W4 W0 W4 W0 W 4   x4 
 d   0   
 X d ( 5 )  W W5 W2 W7 W4 W1 W6 W 3   x5 
 X ( 6 )  W 0 W6 W4 W2 W0 W6 W4 W 2   x6 
 d     
 X d ( 7 )  W 0 W7 W6 W5 W4 W3 W2 W   x7 
1

© PFH Private Hochschule Göttingen 32

The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

Rearranging so that the input array a is bit-reversed and factoring the 8×

8 matrix:
 X d ( 0 )  W 0 W0 W0 W0 W0 W0 W0 W 0   x0 
   0   
 X d ( 1)  W W4 W2 W6 W1 W5 W3 W 7   x4 
 X d ( 2 )  W 0 W0 W4 W4 W2 W2 W6 W 6   x2 
   0   
(
 d  = W
X 3 ) W4 W6 W2 W3 W7 W1 W 5   x6 
 X ( 4 )  W 0 
W0 W0 W0 W4 W4 W4 W 4   x1 
 d   0   
 X d ( 5 )  W W4 W2 W6 W5 W1 W7 W 3   x5 
 X ( 6 )  W 0 W0 W4 W4 W6 W6 W2 W 2   x3 
 d     
 X d ( 7 )  W 0 W4 W6 W2 W7 W3 W5 W   x7 
1

© PFH Private Hochschule Göttingen 33

 The discrete Fourier transform
The Fast Fourier Transform (FFT) Algorithm

Rearranging so that the input array a is bit-reversed and factoring the 8×

8 matrix:

 X d ( 0 )  1 0 0 0 W0 0 0 0  1 0 W 0 0 0 0 0 0  1 W 0 0 0 0 0 0 0   x0 
         
 X d (1)  0 1 0 0 0 W1 0 0  0 1 0 W 2 0 0 0 0  1 W 4 0 0 0 0 0 0   x4 
 X d ( 2 )  0 0 1 0 0 0 W 2 0  1 0 W 4 0 0 0 0 0  0 0 1 W0 0 0 0 0   x2 
         
 X d ( 3)  =  0 0 0 1 0 0 0 W 3  0 1 0 W 6

0 0 0 0  0 0

1 W4 0 0 0 0   x6 

 X ( 4 )  1 0 0 0 W4 0 0 0  0 0 0 0 1 0 W 0 0  0 0 0 0 1 W0 0 0   x1 
 d         
 X d ( 5 )  0 0  0 0 0 1 0 W 2  0 0 0 0   x5 
5
1 0 0 0 W 0 0 0 0 0 1 W4
 X ( 6 )  0 0 1 0 0 0 W 6
0   0 0 0 0 1 0 W 4 0  0 0 0 0 0 0 1 W 0   x3 
 d         
 X d ( 7 )  0 0 0 1 0 0 0 W 7  0 0 0 0 0 1 0 W 6  0 0 0 0 0 0 1 W 4   x7 

© PFH Private Hochschule Göttingen 34

The discrete Fourier transform
Choosing the number of data points N and zero padding
N = 20 N = 32 N = 40

discrete time k discrete time k discrete time k

discrete frequency n discrete frequency n discrete frequency n

© PFH Private Hochschule Göttingen 35
The discrete Fourier transform
Choosing the number of data points N and zero padding

Observation:
▪ Doubling the number of values ​N = 20 → 40 doubles the frequency resolution.
▪ The DFT with N = 20 also delivers identical results to a DFT with N = 40 at the identical
frequency points (every 2nd point).

Remarks:
N = 20
▪ The phase of X(n) must not be ignored.
N = 40
The phase can also be interesting at times.
However, we will concentrate here on the
analysis of the amplitude response, as this is the
more important variable in most applications. The values of the DFT for N = 20
and N = 40 are identical!

▪ It is also usual to omit the upper (redundant)

half of the spectrum!

discrete frequency n

© PFH Private Hochschule Göttingen 36

The discrete Fourier transform
Increasing the sampling frequency

discrete frequency n discrete frequency n discrete frequency n

© PFH Private Hochschule Göttingen 37

 N = 64
The discrete Fourier transform
If the time signal x(k) with k = 0, 1, ... N-1 contains exactly M T0 T0 T0
periods of a sine or cosine signal of time duration T 0, then it holds

M  T0 = N  Ts

It follows that the frequency f0 automatically occurs exactly in the N  Ts = 64  Ts

frequency division: M  T0 = 3  T0
discrete time k
N M f
T0 =  Ts → f0 =  fs = M  s
M N N

Due to the periodicity of the complex e-function, the signal in a

DFT analysis is placed together infinitely often

Because the oscillations always span a full number of periods,

these signals match without discontinuities
discrete time k
© PFH Private Hochschule Göttingen 38
The discrete Fourier transform
DFT of Sinusoidal Signals with Complete Periods (Superposition)

▪ If the time signal contains exactly an integer multiple of periods of the oscillation, then the DFT also
reproduces the spectrum exactly
▪ All of the energy is concentrated in one spectral line
▪ This spectral line represents the exact right frequency
▪ Due to the linearity property of the DFT, this also applies to any number of additively superimposed
oscillations
x(k ) = 2   cos(2  f 2  k ) + cos(2  f3  k ) 
2
f2 = fs
32 redundant

5
f3 = fs
32

discrete time k discrete time k

© PFH Private Hochschule Göttingen 39
The discrete Fourier transform
DFT of sinusoidal signals with incomplete periods

▪ However, it is often not possible to choose N in such a way that all oscillations only occur with
complete periods, because:
▪ The periods of the oscillations of interest are not exactly known
▪ There are many vibrations of different frequencies to be analyzed, and it is therefore difficult to find a
suitable N that satisfies the condition for all vibration-frequencies of interest at the same time
▪ The result when the condition is violated

discrete time k discrete frequency n

© PFH Private Hochschule Göttingen 40
The discrete Fourier transform
Peaky fence and leakage effect
The frequency f0 of the periodic signal no longer occurs in the frequency division!
Therefore, the amplitude associated with 2.5/32*fs must be divided between the existing neighboring
discrete frequencies 2/32*fs and 3/32*fs.
In addition, however, part of the amplitude “smeared” over the entire frequency range. This is a
consequence of the discontinuities that arise during the periodic continuation of the time signal and
induce interfering frequencies.

discrete frequency n
© PFH Private Hochschule Göttingen 41
discrete time k
The Fourier transform
Example: Fourier transform of a non periodic signal
f (t )
A  A for − T  t  T
f (t ) =  2 2
0 otherwise

− 𝑇ൗ2 𝑇ൗ
2
t

T
 2
F ( j ) =  f ( t ) e − jt
 dt = A   e − jt  dt
− −T
2
T 2

 ( e jT 2 − e − jT 2 )
A − jt A
=− e =
j −T 2
j

F ( j ) = A 
j  2  sin T ( )
2 = A (
sin T
2 )
j  
2

F ( j ) = A  T 
sin T ( 2 )
T
2
© PFH Private Hochschule Göttingen 42
The Fourier transform
Example: Fourier transform of a non periodic signal

  /2

t  - / 2 F ( ) =  f ( t ) e − jt  dt =  1e
− j 2  f t
 dt
0 für
 − 
− /2
f (t ) = 1 für - / 2  t   / 2  /2
0 1 − jt 2  e j /2 − e − j /2 
 für t  / 2 =− e =  
j − /2
  2j 

F ( ) =  
(
sin   
2 )
▪ The narrower the rectangle is in the time   2
domain, the wider the sinc-function (sine-
cardinal) is in the frequency domain (i.e. sin (  f  )
the further apart the zeros are) F ( f ) =  sinc-function
  f 
© PFH Private Hochschule Göttingen 43
The Fourier transform
Example: Fourier transform of a non periodic signal

0 t  -

für
sin ( 2  f  )
f (t ) =  A für -  t   F ( f ) = 2   A 
0 2  f 
 für t 

0 t  -2

für
sin ( 4    f  )
f (t ) =  A für -2  t  2 F ( f ) = 4   A 
0 4    f 
 für t  2

0 t  -4

für
sin ( 8    f  )
f (t ) =  A für -4  t  4 F ( f ) = 8   A 
0 8    f 
 für t  4
© PFH Private Hochschule Göttingen 44
The Fourier transform
Leackage effect and windowing

▪ A limited time signal xL(k) of length L can be generated from

a periodic signal of greater length N by multiplication with a
rectangular window w(k) of length L
xL ( k ) = w ( k )  x ( k )

▪ This multiplication in the time domain corresponds to a

convolution in the frequency domain

X L (n) = W (n)  X (n)

▪ Here W(n) is the Fourier transform of the rectangular

window w(k)

sin ( Ln / N ) − j ( L −1)n / N
W (n) = e
sin ( n N )

sin ( Ln / N )
W (n) =
sin ( n N )
© PFH Private Hochschule Göttingen 45
The Fourier transform
Leackage effect and windowing
Time signal Spectrum of the time function

Time window

Signal, that have to be analysized

© PFH Private Hochschule Göttingen 46

The Fourier transform
Leakage effect and windowing

▪ So every finite time signal results (in thought) from a multiplication of an infinite
periodically continued signal with a rectangular window
▪ The rectangular shape leads to discontinuities, i.e. a sudden signal change that induces
high frequencies
▪ The errors caused by rectangular windowing (leakage effect) are therefore particularly
large
▪ Instead of an implicitly existing rectangular window that drops off abruptly, a window
with a smoother transition can be selected
▪ Many alternatives have been proposed as windows
▪ The following windows are common: Hanning, Hamming, Blackman, Kaiser, Bartlett,
Gauss
▪ All of these windows are relatively similar and all reduce the leakage effect. However,
due to the gentle fall, they distort the values ​of the time signal at the edge (beginning
and end)
© PFH Private Hochschule Göttingen 47
Leakage effect and windowing

Rectangular

© PFH Private Hochschule Göttingen 48

Leakage effect and windowing

Rectangular

w(n) = 1 (
w ( n ) = 0,5 − 0,5  cos 2 n
N ) (
w ( n ) = 0,54 − 0, 46  cos 2 n
N )

 2n
 N
w(n) = 
0n N
2 (
w ( n ) = 0, 42 − 0,5  cos 2 n
N ) + 0, 08  cos ( 4 n N )
2 − n N N nN
2
© PFH Private Hochschule Göttingen 49
Application of the FFT

© PFH Private Hochschule Göttingen 50

Application of the FFT
Which information can be calculated form this time signal

© PFH Private Hochschule Göttingen 51

Application of the FFT
FFT calculation

© PFH Private Hochschule Göttingen 52

Application of the FFT
FFT calculation and representation in a waterfall diagram

© PFH Private Hochschule Göttingen 53

Application of the FFT
Spectrum of the trigger signal from the drum roller

© PFH Private Hochschule Göttingen 54

Application of the FFT
Spectrum of the trigger signal from the winder roll

© PFH Private Hochschule Göttingen 55

Application of the FFT (MP3 / JPEG)
Signals are manipulated in frequency space in order to compress them
▪ MP3 format: Human hearing in the range 20 Hz - 18 kHz !
Frequencies outside this range are filtered out
▪ JPEG format: Image is divided into 8x8 pixel blocks, which are then Fourier
transformed

© PFH Private Hochschule Göttingen 56

Application of the FFT
Signalanalysis

▪ High-pass / low-pass / band-pass filter of a signal by manipulating the

frequency spectrum e.g. image processing in frequency domain:

▪ Blur a picture using low pass filter

▪ Edge detection using high-pass filter

© PFH Private Hochschule Göttingen 57

Parseval's theorem for energy signals
Linking of the signal energies in the time and frequency domain

  
 1 

W= − x 2
( t ) dt = − x ( t )  x ( t ) dt = − x (t )   2  X ( j )  e
jt
d   dt
− 

1

  1

W=   X ( j )   x ( t )  e jt dt  d  =   X ( j )  X * ( j )  d 
2 −  −  2 −


1
  X ( j )  d 
2
W=
2 −

▪ Parseval's theorem for energy signals

 
1 Parseval ' s theorem for time discrete energy signals :
 x 2 ( t )  dt =   X ( j )  d 
2
W=
2 − N −1
1 N −1

 x2 ( k ) =  X (n)
2
−

k =0 N n =0

 X ( j ) 2  = V = V  s = V  As  V = Ws 1
2 2

  Hz 2 Hz Hz A Hz
© PFH Private Hochschule Göttingen 58