Fast Fourier Transforms

Quote of the Day

On two occasions I have been asked, “Pray, Mr.
Babbage, if you put into the machine wrong figures,
will the right answers come out?” I am not able
rightly to apprehend the kind of confusion of ideas
that could provoke such a question.

                              Charles Babbage

Content and Figures are from Discrete-Time Signal Processing, 2e by Oppenheim, Shafer, and Buck, ©1999-2000 Prentice Hall Inc.
 Discrete Fourier Transform
• The DFT pair was given as
1 N 1
x[n]   Xk e j 2  / Nkn
N 1
Xk    x[n]e  j 2  / N  kn
N k 0
n0

• Baseline for computational complexity:

– Each DFT coefficient requires
• N complex multiplications
• N-1 complex additions
– All N DFT coefficients require
• N2 complex multiplications
• N(N-1) complex additions
• Complexity in terms of real operations
• 4N2 real multiplications
• 2N(N-1) real additions
• Most fast methods are based on symmetry properties
 j 2  / N  k  N  n 
– Conjugate symmetry e  e j 2 / NkNe  j 2  / Nk  n  e j 2  / Nkn
– Periodicity in n and k e  j 2  / N kn  e  j 2  / N k  n N  e j 2  / N k N n

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 2

 The Goertzel Algorithm
• Makes use of the periodicity
e j 2  / NNk  e j2 k  1
• Multiply DFT equation with this factor
N 1 N 1
Xk   e j 2  / N  kN
 x[r]e  j 2  / N  rn
  x[r]e j 2  / N r  N n 

r 0 r 0

• Define 
yk n   x[r]e j 2  / N  k  n  r 
un  r 
r  
• With this definition and using x[n]=0 for n<0 and n>N-1

Xk   yk n n N
• X[k] can be viewed as the output of a filter to the input x[n]
– Impulse response of filter:

e j 2  / Nknun

– X[k] is the output of the filter at time n=N

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 3

 The Goertzel Filter
• Goertzel Filter
1
Hk  z   2
j k
1e N
z 1
• Computational complexity
– 4N real multiplications
– 2N real additions
– Slightly less efficient than the direct method

• Multiply both numerator and denominator

2 2
j k j k
1e z 1 N
1e z 1 N
Hk  z   
 2
j k  2
j k  2k 1
1  e N z 1 1  e N z 1  1  2 cos z  z 2
   N
  

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 4

 Second Order Goertzel Filter
• Second order Goertzel Filter

2
j k
1e z 1 N
Hk  z  
2k 1
1  2 cos z  z2
N

• Complexity for one DFT coefficient

– Poles: 2N real multiplications and 4N real additions
– Zeros: Need to be implement only once
• 4 real multiplications and 4 real additions
• Complexity for all DFT coefficients
– Each pole is used for two DFT coefficients
• Approximately N2 real multiplications and 2N2 real additions
• Do not need to evaluate all N DFT coefficients
– Goertzel Algorithm is more efficient than FFT if
• less than M DFT coefficients are needed
• M < log2N

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 5

 Decimation-In-Time FFT Algorithms
• Makes use of both symmetry and periodicity
• Consider special case of N an integer power of 2
• Separate x[n] into two sequence of length N/2
– Even indexed samples in the first sequence
– Odd indexed samples in the other sequence
N 1 N 1 N 1
Xk    x[n]e  j 2  / N  kn
  x[n]e  j 2  / N  kn
  x[n]e  j 2  / N  kn

n0 n even n odd

• Substitute variables n=2r for n even and n=2r+1 for odd

N / 2 1 N / 2 1
Xk    x[2r]W 2rk
N   x[2r  1]W N
2r 1  k

r 0 r 0
N / 2 1 N / 2 1
  x[2r]W
r 0
rk
N/2 W k
N  x[2r  1]W
r 0
rk
N/2

 Gk   W Hk  k
N

• G[k] and H[k] are the N/2-point DFT’s of each subsequence

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 6
 Decimation In Time
• 8-point DFT example using
decimation-in-time
• Two N/2-point DFTs
– 2(N/2)2 complex multiplications
– 2(N/2)2 complex additions
• Combining the DFT outputs
– N complex multiplications
– N complex additions
• Total complexity
– N2/2+N complex multiplications
– N2/2+N complex additions
– More efficient than direct DFT
• Repeat same process
– Divide N/2-point DFTs into
– Two N/4-point DFTs
– Combine outputs

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 7

 Decimation In Time Cont’d
• After two steps of decimation in time

• Repeat until we’re left with two-point DFT’s

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 8

 Decimation-In-Time FFT Algorithm
• Final flow graph for 8-point decimation in time

• Complexity:
– Nlog2N complex multiplications and additions

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 9

 Butterfly Computation
• Flow graph constitutes of butterflies

• We can implement each butterfly with one multiplication

• Final complexity for decimation-in-time FFT

– (N/2)log2N complex multiplications and additions

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 10

 In-Place Computation
• Decimation-in-time flow graphs require two sets of registers
– Input and output for each stage
• Note the arrangement of the input indices
– Bit reversed indexing
X 0 0  x0  X 0 000  x000
X 0 1  x 4  X 0 001  x100
X 0 2  x2  X0 010  x010
X 0 3  x6  X0 011  x110
X 0  4  x1  X 0 100  x001
X 0 5  x5  X 0 101  x101
X 0 6  x3  X0 110  x011
X 0 7  x7  X0 111  x111

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 11

 Decimation-In-Frequency FFT Algorithm
• The DFT equation
N 1
Xk    x[n]Wnk
N
n0

• Split the DFT equation into even and odd frequency indexes
N 1 N / 2 1 N 1
X2r    x[n]W n2r
N   x[n]W n2r
N   x[n]Wn2r
N
n0 n0 n N / 2

• Substitute variables to get

N / 2 1 N / 2 1 N / 2 1
X2r    x[n]W n2r
N   x[n  N / 2]W  n N / 2  2r
N    x[n]  x[n  N / 2]W
nr
N/2
n0 n0 n0

• Similarly for odd-numbered frequencies

N / 2 1
X2r  1    x[n]  x[n  N / 2]W  n 2r 1
N/2
n0

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 12

 Decimation-In-Frequency FFT Algorithm
• Final flow graph for 8-point decimation in frequency

Copyright (C) 2005 Güner Arslan 351M Digital Signal Processing 13