Digital Signal Processing

(EEE301)
Dr. P. Geethanjali
Fast Fourier Transform Algorithm
• Consider DTFT

• Basic idea is to split the sum into 2 subsequences of length N/2 and continue all
the way down until you have N/2 subsequences of length 2

Log2(8)

N
Decimation in Time
In this method we split x (n) into the even indexed x (2m) and the odd indexed x
(2m + 1) each N/2 long. DFT can be expressed as:

N / 2 1 N / 2 1
2 mk k ( 2 m 1)
X (k )  
m0
x ( 2 m )W N   x ( 2 m  1)W
m0
N

N / 2 1 N / 2 1
X (k )   x ( 2 m )W mk
W k
 x ( 2 m  1)W km (1)
N /2 N N /2
m 0 m0

K = 0,1,2,3,4,……………..,N-1

We have used the fact

WN2 mk  e  j 2 mk 2 / N  e  j 2mk /( N / 2 )  WNmk/ 2

Decimation in Time
Define new functions as
G (K) is a DFT for the even samples of x (n), or
N / 2 1
mk
G (k )   x(2m)W
m0
N /2 k  0,1,2,...., N / 2  1

H (K) is a DFT for the odd samples of x(n), or

N / 2 1
km
H (k )   x(2m  1)W
m0
N /2 k  0,1,2,3,..., N / 2  1

Then equation (1) for the first N/2 points can be expressed as

X (k )  G (k )  WNk H (k ) k  0,1,2,...., N / 2  1
 Decimation in Time
But
WN( N / 2 k )  WNk
G (k )  G ( N / 2  k ), k  0,1,2,..., N / 2  1
H (k )  H ( k  N / 2) k  0,1,2,...., N / 2  1
Then

X ( k  N / 2 )  G ( k )  W Nk H ( k ) k  N / 2, N / 2  1,...., N

This can be expressed in the following diagram

Example
Let’s take a simple example where only two points are given n=0, n=1; N=2
N / 2 1 N / 2 1
mk k mk
X [k ]  WN / 2 x[2m] WN (
m 0
WN / 2 x[2m  1])
m0
2 mk mk
WN  WN / 2
m N / 2 m N /2 m
WN / 2  WN / 2 WN / 2  WN / 2
N
WN  e 2j  cos( 2 )  j sin( 2 )  1
N /2
WN  1

0
0.0 0
0
0.0
Same result
X [ k  0 ]  W 1 x[0] W ( W
1 1 x[1])  x[0]  x[1]
m 0 m0
0 0
0.1 1 0.1 1
X [k  1]  W x[0] W ( W1 x[1])  x[0]  W1 x[1]  x[0]  x[1]
1 1
m0 m 0
FFT Algorithms - Four point FFT
First find even and odd parts and then combine them:
Decimation in Time and Frequency

Index mapping for Fast Fourier Transform

Input Data index Index Bits Reversal Bits Output data

n index
k
0 000 000 0
1 001 100 4
2 010 010 2
3 011 110 6
4 100 001 1
5 101 101 5
6 110 011 3
7 111 111 7
8-Point DIT
Decimation in Time
G(0)
x(0) X(0)
G(1)
x(2) X(1)
4-point
DFT G(2)
x(4) X(2)
G(3)
x(6) X(3)
H(0) W 80
x(1) X(4)
-1
H(1) W81
x(3) X(5)
4-point
-1
DFT H(2) W82
x(5) X(6)
-1
H(3) W 83
x(7) X(7)
-1
Decimation in Time
x(0) 2-point X(0)
DFT
x(4) X(1)
x(2) W80
2-point X(2)
-1
DFT W82
x(6) X(3)
-1
x(1) W 80
2-point X(4)
-1
DFT W81
x(5) X(5)
-1
x(3) W80 W82
2-point X(6)
-1 -1
DFT W82 W 83
x(7) X(7)
-1 -1
Decimation in Time
x(0) X(0)
W80
x(4) X(1)
-1
W80
x(2) X(2)
-1
W80 W82
x(6) X(3)
-1 -1
W 80
x(1) X(4)
-1
W80 W81
x(5) X(5)
-1 -1
W80 W82
x(3) X(6)
-1 -1
W80 W82 W 83
x(7) X(7)
-1 -1 -1
Decimation in Time
Example Given a sequence x(n) where x(0) = 1, x(1) = 2, x(2) = 3,
x(3) = 4 and x(n) = 0 elsewhere ,find DFT for the first four points
Solution

x(0)=1 4
X(0 )= 10
W40  1 -2
x(2)=3 X(1)= -2+2j
-1
6 W40  1
x(1)=2 X(2)= -2
-1
1
W40  1 -2 W4   j
x(3)=4 X(3)= -2-2j
-1 -1
 Inverse Fourier Transform
The inverse discrete Fourier can be calculated using the same method
but after changing the variable WN and multiplying the result by 1/N

Example Given a sequence X(n) given in the previous example.

Find the IFFT using decimation in time method
Solution

8 4 1/4
X(0) =10 x(0) = 1
~
W40  1 12 8 1/4
X(2) =-2 x(1) = 2
-1 ~
-4 W40  1 12 1/4
X(1) = -2+2j x(2) = 3
~ -1
W40  1 j4 W~ 1  j
X(3) = -2-2j 4 16 1/4 x(3) = 4
-1 -1
Decimation in Frequency

The discrete Fourier transform can be found using

N 1
kn
X (k )   x ( n )W
n0
N k  0 ,1, 2,....., N  1

Where N = 2, 4, 8, 16,… and

WN  e  j 2 / N
X (k ) can be expressed as
N / 2 1 N 1
kn kn
X (k )  
n0
x ( n )W N   x ( n )W
n N / 2
N

N / 2 1 N / 2 1
kn k (n N / 2)
X (k )  n0
x ( n )W N   x ( n  N / 2 )W
n0
N

N / 2 1 N / 2 1
kn kN / 2 kn
X (k )  n0
x ( n )W N W N  x
n0
( n  N / 2 )W N
Decimation in Frequency

But

W N( N / 2 ) k  e 2 k ( N / 2 ) / N  e k  (  1) k

Then
N / 2 1 N / 2 1
kn k kn
X (k )  
n0
x(n)W N  (1)  x(n  N / 2)W
n0
N

If k = 2m or an even number

N / 2 1
X (2m)   x ( n )  x ( n  N / 2 ) W
n0
N
2 mn

N / 2 1
mn
X ( 2 m)   a
n0
( n )W N /2
Decimation in Frequency
Noting That
WN2 mn  e  j 2 mn2 / N  e  j 2mn /( N / 2 )  WNmn/ 2
a ( n)  x ( n )  x ( n  N / 2)
Then X(2m) is N/2-point DFT for a(n)
If k = 2m+1 (odd number) and using the same method
N / 2 1
X ( 2 m  1)   x ( n )  x ( n  N / 2) W
n 0
N
n
W N2 mn
N / 2 1
X (2m  1)   b(n)W W
n0
n
N
mn
N /2

b( n )  x ( n)  x ( n  N / 2)

X(2m+1) is N/2-point DFT for b(n)WNn

Decimation in Frequency

a(0)
x(0) X(0)
a(1)
x(1) X(2)
N/2-point
a(2) DFT
x(2) X(4)
a(3)
x(3) X(6)
b(0) W80
x(4) X(1)
-1
b(1) W81
x(5) X(3)
-1 N/2-point
b(2) W 82 DFT
x(6) X(5)
-1
b(3) W 83
x(7) X(7)
-1
Decimation in Frequency

x(0) N/4 point X(0)

DFT
x(1) X(4)
W40  W80
x(2) N/4 point X(2)
-1
1
W4  W82
DFT
x(3) X(6)
-1
W80
x(4) N/4 point X(1)
-1
W81 DFT
x(5) X(5)
-1
W82 W40  W80
x(6) X(3)
-1 -1 N/4 point
W 83 W41  W82 DFT
x(7) X(7)
-1 -1
Decimation in Frequency

x(0) X(0)
W80
x(1) X(4)
-1
W80
x(2) X(2)
-1
W82 W80
x(3) X(6)
-1 -1
W80
x(4) X(1)
-1
W81 W80
x(5) X(5)
-1 -1
2 0
W W
x(6) 8 8
X(3)
-1 -1
W 83 W82 W80
X(7)
x(7)
-1 -1 -1
Decimation in Frequency

Using the previous algorithm , the complex multiplications needed

is only 12. While using the normal DFT would require 64 complex
multiplications

In general

Complex multiplication of DFT is: N 2

Complex multiplication of FFT is (N/2) log2(N)

If N = 1024
Complex multiplication of DFT is: 1,048,576
Complex multiplication of FFT is: 5,120
Decimation in Frequency

Example Given a sequence x(n) where x(0) = 1, x(1) = 2, x(2) = 3, x(3) = 4

and x(n) = 0 elsewhere ,find DFT for the first four points
solution

4 10
x(0) =1 X(0)
6 W40  1 -2
x(1) =2 X(2)
-1
-2 W40  1 -2+2j
x(2) =3 X(1)
-1
-2 W41   j W40  1 -2-2j
x(3) =4 X(3)
-1 -1
Inverse Fourier Transform
The inverse discrete Fourier transform can be found using

N 1
1  kn
x(n) 
N
 X ( k )W
k 0
N n  0 ,1, 2,....., N  1

Which can be expressed as

N 1 ~
1 kn
x(n) 
N
 X (k ) W
k 0
N n  0 ,1, 2,....., N  1

~
where W Nkn  W N kn is called the twiddled factor

We can see that the difference between the inverse discrete Fourier and
forward Fourier transform is the twiddled factor and the division by 1/N
Inverse Fourier Transform
X(0) 1/N x(0)
~
W80 1/N x(4)
X(1)
~
-1
W80 1/N
X(2) x(2)
-1 ~ ~

X(3) W82 W80 1/N x(6)

~
-1 -1
X(4)
W80 1/N x(1)
-1 ~ ~

X(5)
W81 W80 1/N x(5)
-1 ~ ~ -1
X(6) W 8
2
W 8
0
1/N x(3)
-1 -1 ~ ~
~

X(7) W 8
3
W 8
2
W80 1/N x(7)
-1 -1 -1
Inverse Fourier Transform

Example Given a sequence X(n) given in the previous example. Find the
IFFT using decimation in frequency method
solution

8 4 1/4
X(0) =10 x(0) = 1
-4 ~
W40  1 12 1/4
X(1) = -2+2j x(2) = 3
-1
12 W~ 0  1 8
X(2) =-2 4 1/4
x(1) = 2
-1 ~
j4 ~
1
W40  1 16
X(3) = -2-2j W  j
4 1/4
x(3) = 4
-1 -1