Lecture 6: Discrete Fourier Transform

Topics:
– Sampling
– Discrete Fourier Transform (DFT)
– Properties of DFT
– Fast Fourier Transform
 Lecture 6: Discrete Fourier Transform
Topics:
– Sampling
– Discrete Fourier Transform (DFT)
– Properties of DFT
– Fast Fourier Transform
Sampling

Recap: The sampling theorem

F ( )
f (t ) A

t − 2B 2B 
−B 0 B F (Hz)
 T (t )
1 
F () =  F ( − ns )
T n=−

T t F ( ) Lowpass
filter

f (t ) A/T

−s − 2B 2B s 

t
− Fs −B 0 B Fs F (Hz)
Sampling

Recap: The sampling theorem

1 
F () =  F ( − ns )
T n=−

F ( )
A

Fs  2B
 Nyquist - Shannon 
− 2B 2B or, equivalently,
−B 0 B
sampling theorem :  1
F (Hz)
 T 
2B
F ( )
Lowpass
filter
samples
A/T Fs  2B , (2B = spectral width)
second

−s − 2B 2B s 

− Fs −B 0 B Fs F (Hz)
Sampling

The spectral sampling theorem

F ()
f (t)

0  t 0 

 
F () =  f (t)e − j t
dt =  f (t)e− j t dt
− 0


fT0 (t) 2π
fT0 (t) =  Dn e jn0 t , 0 =
n=− T0

1 T0 1 
0 
T0
t Dn =  f (t) e− jn0 t dt =  f (t) e− jn0 t dt
T0 0 T0 0
1
 Dn = F (n0 )
T0
Sampling

The spectral sampling theorem

1
Dn = F (n0 )
T0
F ()
f (t)

0  t 0 

Dn 1
fT0 (t) F ()
T0

0  t 0 
T0
1 F
F0 =
T0
Condition for recovery:
1 samples
T0      R  , ( = signal width)
F0 hertz
 Lecture 6: Discrete Fourier Transform
Topics:
– Sampling
– Discrete Fourier Transform (DFT)
– Properties of DFT
– Fast Fourier Transform
The discrete Fourier transform (DFT)

f (t) F ()

0  t 0 F
(a) (b)

f (t) F ()

(c) 0 t (d) 0 Fs Fs F
T 2
1
F0 =
T0
T

0 t 0 Fs =
1 F
(e) T0 (f) T
The discrete Fourier transform (DFT)

1
F0 =
T0
T

0 t 0 Fs =
1 F
T0
T

Number of samples per period:

T0 F
N0 = N0' = s
T F0
1 1
F0 = Fs =
T0 T

T0 F
N0 = = s = N0'
T F0
The discrete Fourier transform (DFT)
f (t)
T
N0 −1
f (t) =  f (kT )  (t − kT )
k =0
0 t
T0 F ( )
A

 (t − kT )  e− jkT 
− 2B 2B
−B 0 B F (Hz)

 
N0 −1
F () = F f (t) =  f (kT ) e
F ( ) Lowpass
− jkT
filter
A/T
k =0
F () s s
F () = , −   −s − 2B 2B s 
T 2 2
− Fs −B 0 B Fs F (Hz)
N0 −1
s
F () = TF () = T  f (kT ) e− jkT ,   1 
k =0 2 F () =  F ( − ns )
T n=−
Samples of F () at multiples of 0 :
N0 −1 N0 −1
2 2
Fr = F (r0 ) =  T f (kT ) e − jkr 0T
=  fk e− jr 0k , 0 =0T = T= .
k =0 0 k =0 T0 N0
fk
The discrete Fourier transform (DFT)

N0 −1
Fr =  fk e− jr 0k
k =0

Change subindex:
N0 −1
Fr =  fm e− jr 0m
m=0

N0 −1

N0 −1 N0 −1
− jr 0m  jr 0k
N0 −1
N0 −1 j 0(k −m)r 
F e r
jr 0k
=   fm e
r =0  m=0
e =  fm  e 
r =0  m=0  r =0 

2
N0 −1 N0 −1 (k −m)r N0 when m = k,
 = e
j
j 0 (k −m)r
e N0
=
r =0 r =0  0 when m  k.

N0 −1
1 N0 −1 jr 0k 2
Fre
r =0
jr 0k
= N0 f k  fk = Fr e
N0 r =0
, 0 =0T =
N0
.
The discrete Fourier transform (DFT)

Sample f (t) (and scale by T ): fk = T f (kT )

2
Sample F (): Fr = F (r0 ), 0 =
T0

N0 −1
2
Fr =  fk e− jr 0k , 0 =0T = DFT r = 0, 1,  , N0 −1
k =0 N0

1 N0 −1 jr 0k fk  Fr
fk =  Fr e IDFT k = 0, 1,  , N0 −1
N0 r =0

1
F0 =
T0
T

0 t 0 Fs =
1 F
T0
T
 Lecture 6: Discrete Fourier Transform
Topics:
– Sampling
– Discrete Fourier Transform (DFT)
– Properties of DFT
– Fast Fourier Transform
The discrete Fourier transform (DFT)

Properties of the DFT

1. Linearity: if fk  Fr and gk  Gr then,

a1 fk + a2 gk  a1Fr + a2Gr

2. Conjugate symmetry: for a real signal fk ,

since Fr is N0-periodic  F−r = F−r + N0 = FN0 −r ,

Hence, we only need to compute ½ of the DFT for real signals fk .

The discrete Fourier transform (DFT)

Properties of the DFT

3. Time-shifting (circular shifting): if fk  Fr then,

fk −n  Fr e− jr0n

4. Frequency-shifting: if fk  Fr then,

fk −n e jk0m  Fr−m

5. Circular convolution:
N0 −1 N0 −1
fk . gk =  fn gk −n =  gn fk −n
n=0 n=0
The discrete Fourier transform (DFT)

Properties of the DFT

5. Circular convolution:
N0 −1 N0 −1
fk . gk =  fn gk −n =  gn fk −n
n=0 n=0

f1 f3 f1 f3
fn f0 f2 fn f0 f2

0 1 2 3 n 0 1 2 3 n

g3 g3 g3 g3
g2 g2 g2 g2
g0−n g1 g1 g1−n g1 g1
g0 g0 g0 g0
−3 − 2 −1 0 1 2 3 4 n − 2 −1 0 1 2 3 4 5 n

 fk . gk  k =0 = f0 g0 + f1 g3 + f2 g2 + f3 g1  fk . gk  k =1 = f0 g1 + f1 g0 + f2 g3 + f3 g2
The discrete Fourier transform (DFT)

Properties of the DFT

5. Circular convolution:
N0 −1 N0 −1
fk . gk =  fn gk −n =  gn fk −n
n=0 n=0

g1 g2
f3 f3
g2 f 2 f 0 g0 g3 f 2 f0 g1
f1 f1
g3 k =0 g0 k =1

 fk . gk  k =0 = f0 g0 + f1 g3 + f2 g2 + f3 g1  fk . gk  k =1 = f0 g1 + f1 g0 + f2 g3 + f3 g2
The discrete Fourier transform (DFT)

Properties of the DFT

5. Circular convolution:
N0 −1 N0 −1
fk . gk =  fn gk −n =  gn fk −n
n=0 n=0

fk . gk  Fr Gr

1
f k gk  Fr . Gr
N0
 Lecture 6: Discrete Fourier Transform
Topics:
– Sampling
– Discrete Fourier Transform (DFT)
– Properties of DFT
– Fast Fourier Transform
The fast Fourier transform (FFT)

Recall: DFT

N0 −1
2
Fr =  fk e − jr 0k
, 0 =0T = , r = 0, 1,  , N0 −1
k =0 N0

N0 complex multiplications  N0

N0 −1 complex additions  N0

Thus, to compute the DFT we require:

2
N0 complex multiplications
N0 ( N0 −1) complex additions
The fast Fourier transform (FFT)
2
−j
Define: WN0 = e N0
= e− j 0

2 2 2 2
−j −j 2  −j 
W N0 = e =e = e  = WN0
( N0 / 2) N0 N0 2
Note that:
 
2  
2 N0
N   N0  −j 
r + 0   
=W =e WN0 = e− j WN0 = −WN0
r N0 2 r r r
W N0  
2
N0  
2 WN0

N0 −1 N0 −1
Note that: Fr =  fk e − jr 0k
=  fk WN0
kr
0  r  N0 −1
k =0 k =0
N0 N0
−1 −1
2 2
=  f 2k WN0 +  f2k +1 WN0
2k r (2k +1) r

k =0 k =0

f0, f2, f4,  f N0 −2, f1, f3, f5,  f N0 −1

  
gk hk
N0 N0
−1 −1
2 2
Fr = g
k =0
k W N0
kr
+ WN0
r
hW
k =0
k N0
kr
= Gr + WN0 H r ,
r
0  r  N0 −1
2 2
The fast Fourier transform (FFT)

Fr = Gr + WN0 H r , 0  r  N0 −1
r

N 
r + 0 
F N 
=G N 
+W N0  
2 H N 
r + 0  r + 0  r + 0 
 2   2   2 

N 
r + 0 
= −WN0
r
W N0  
2

N 
Since Gr and H r are  0 -periodic: G N 
= Gr and H N 
= H r.
 2  r + 0 
 2 
r + 0 
 2 

= Gr − WN0 H r
r
F  N0 
r + 
 2 

N0
Fr = Gr + WN0 H r , 0  r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F  N0 
r +  2
 2 
The fast Fourier transform (FFT)

N0
Fr = Gr + WN0 H r , 0  r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F  N0 
r +  2
 2 

Butterfly structure:

Gr Fr
WNr0
Hr Fr +( N0 / 2)
−WNr0
The fast Fourier transform (FFT)
N0
Fr = Gr + WN0 H r , 0 r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F N 
r + 0  2
 2 

For example, G0
f0 F0
for N0 =8: G1
f2 F1
N0 = 4 W8
DFT
G2
f4 F2
W82
G3
f6 F3
3
W 8

f1 F4
H0 −1

f3 F5
N0 = 4 H1 −W8
DFT
f5 F6
H2 −W82

f7 F7
H3 −W83
The fast Fourier transform (FFT)
N0
Fr = Gr + WN0 H r , 0 r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F N 
r + 0  2
 2 

For example, A0
f0 G0
N0 = 2
for N0 =8: A1
f4 DFT
G1
B0 W 8
2
(W4 =W82 )
f2 G2
N0 = 2 −1
DFT B1
f6 G3
−W82

f1 H0
N0 = 2 C0
f5 DFT
H1
C1 W 8
2

f3 H2
N0 = 2 D0 −1
DFT
f7 H3
D1 −W82
The fast Fourier transform (FFT)
N0
Fr = Gr + WN0 H r , 0 r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F N 
r + 0  2
 2 

For example, f0 A0
for N0 =8: when N0 = 2:
f4 A1
−1 N0 −1
Fr =  fk e− jr 0k
f2 B0 k =0
1
f6
−1 B1 =  fk e− jr k
k =0
f1 C0
F0 = f0 + f1
f5 C1
−1
F1 = f0 − f1
f3 D0

f7 −1 D1
The fast Fourier transform (FFT)

N0
Fr = Gr + WN0 H r , 0 r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F  N0 
r +  2
 2 

N0
First stage: complex multiplications
2
The fast Fourier transform (FFT)

N0
Fr = Gr + WN0 H r , 0 r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F  N0 
r +  2
 2 

N0
First stage: complex multiplications
2
N0 complex additions
The fast Fourier transform (FFT)

N0
Fr = Gr + WN0 H r , 0 r  −1
r

2
N
= Gr − WN0 H r , 0  r  0 −1
r
F  N0 
r +  2
 2 

N0
complex multiplications
First stage: 2
N0 complex additions

N0 N0
To compute Gr : complex multiplications To compute H r : complex multiplications
4 4
N N
and 0 complex additions and 0 complex additions
2 2

N0
complex multiplications
Second stage: 2
N0 complex additions
The fast Fourier transform (FFT)

N0
complex multiplications
Each stage: 2
N0 complex additions

Number of stages: log 2 N0

Thus, to compute the FFT we require of the order of:

N0
log 2 N0 complex multiplications
2
N0 log 2 N0 complex additions
The fast Fourier transform (FFT)

DFT
16,384

N 02

N0 log 2 N0 FFT
896
0 128 N0