FFT(Fast Fourier

Transform)
 FFT
A( x )  a0  a1 x  .....  an 1 x n 1
B( x )  b0  b1 x  .....  bn 1 x n 1
C ( x )  A( x ) B ( x )
= c0  c1 x  c2 x 2  .....  c2 n 3 x 2 n 3  c2 n 2 x 2 n 2
j
c j   ak b j  k
k 0

•Coefficient representation:
How to evaluate A(x0)?

p2.
 Horner’s rule:
A( x0 )  a0  x0 (a1  x0 (a2  .....  x0 (an  2  x0 an 1 )...))
(n)

 Point-value representation:
A point-value representation of a polynomial A(x) of
degree-bound n is a set of n point-value pairs {(x 0 , y0 ),
(x1 , y1 ),.....(x n-1 , yn 1 )}, where y k  A( xk )

A :{(x 0 , y0 ), (x1 , y1 ),.....(x 2n-1 , y2 n 1 )}

B :{(x 0 , y0' ), (x1 , y1' ),.....(x 2n-1 , y2' n 1 )}
C :{(x 0 , y0 y0' ), (x1 , y1 y1' ),.....(x 2n-1 , y2 n 1 y2' n 1 )}

p3.
 Thm1
For any set {(x 0 , y0 ),(x1, y1 ),.....(x n-1, yn 1 )} of n point-value
pairs, there is a unique poly A(x) of degree  n-1, such that
y k  A( xk ) for k=0, 1,...., n-1
Pf:
 1 x0 x02  x0n 1 
 2 n 1

1 x1 x1  x1 
if X =  then
     
 
1 x 2 n 1 
 n 1 xn 1  xn 1 

 a0   y0 
   
a y
X 1  1 
    
   
a y
 n 1   n 1 

p4.
 det( X )   ( xk  x j )  0 if xk 's are distinct
jk

Thus, a j 's can be solved uniquely.

 a0   y0 
   
 a1  X 
1
y1 
    
   y 
a
 n 1   n 1 

 n-point interpolation:
Lagrange’s formula:

 (x  x )
n 1 j

A( x)   y
j k

 (x  x )
k
k 0 k j
jk

p5.
a0 , a1 ,...., an 1 Ordinary Mult c0 , c1 ,...., c2 n  2 Cof.
rep.
b0 , b1 ,...., bn 1 ( n 2 )

Evalation Interpolation
(n log n) (n log n)

A(02 n ), B (02 n ) C (02 n )

A(02 n ), B (02 n ) Pointwise Mult C (02 n ) Point-value
( n) rep.
 ,  
A(22 nn 1 ), B(22 nn 1 ) C (22 nn 1 )

p6.
 Thm2
The product of 2 polynomials of deg-bound n can
be computed in time (nlog n), with both the input
and output representation in coefficient form
Complex roots of unity:
n  1, e 2ik/n for k=0, 1, ..., n-1 e iu  cos u  i sin u
n  e 2i/n , the principal n-th root of unity
0n , 1n , 2n ,..., nn 1
nn =?, 0n =1
nj kn =(n j  k ) mod n
n1  nn 1
p7.
 Lemma 3 (Cancellation Lemma)
dk
n, k, d: non-negative integers, dn  k
n

Pf:
2 i / dn dk 2 i / n k
dk
dn  ( e )  ( e )  k
n

 Cor. 4 n: even positive integer

nn / 2  2  1

p8.
 Lemma 5 (Halving lemma)
n: even positive integer
The squares of the n complex n-th roots of unity are n/2
complex (n/2)th roots of unity.
Pf:
(kn )2  2n k  kn / 2, where k  Z   {0}
(kn  n / 2 ) 2  2n k  n  2n k  kn / 2
 kn and kn  n / 2 have the same squre
 Lemma 6 (Summation lemma)
n-1
n  Z  , k  Z   {0}, n | k ,  n)  0
(
j=0
 k j

Pf:
n-1
(kn ) n  1 (nn ) k  1

j=0
( ) 
k
n
j

 k
 1

 k
 1
0
n n

p9.
 DFT n -1
Evaluate A( x)   a j x j at 0n , 1n ,....., nn-1 ,
j 0

Assume n is a power of 2
n -1
Let a  a0 , a1 ,...., an -1 , and yk  A(kn )   a j kjn
j 0

y  y0 , y1 ,...., yn -1  is the DFT of the coefficient

vector a=<a 0 , a1 ,...., a n-1 ,
y  DFTn (a)

p10.
 Interpolation at the complex roots of unity:

 y0  1 1 1  1   a0 
   n 1  
y  1   2
 n  1 a
 1  n n

 y2  = 1 2 4  2( n 1) 
 a2 
   n n n
 
          
y   n 1 2( n 1) ( n 1)( n 1) 
  an 1 
 n 1   1  n  n  n 
y  Vn a, (Vn ) k , j  nkj
a  Vn1 y

p11.
 Thm 7
1  kj
j, k  0,1,....., n  1, (V ) n j ,k  n /n
Pf:
n 1
V V  I n , (V V ) j , j '   (Vn1 ) j ,k (Vn ) k , j '
n
1
n n
1
n
k 0
n 1
1
=  n kj kjn '
k 0 n

1 n 1 k ( j ' j )
=  n -------(*)
n k 0
if j=j', then (*)=1, if j  j', then by lemma 6, (*)=0
 -(n-1) < j'-j <n-1, and n | (j'-j)
1 n 1 n 1
a j   yk n k j , yk   a j kn j
n k 0 j 0

p12.
 FFT
A[0] ( x )  a0  a2 x  a4 x 2  .....  an  2 x n / 21
A[1] ( x )  a1  a3 x  a5 x 2  .....  an 1x n / 21
(*) A( x )  A[0] ( x 2 )  xA[1] ( x 2 )
Thus evaluating A(x) at 0n , 1n ,....., nn 1 reduce to
1. evaluating A[0] ( x) and A[1] ( x) at
(0n ) 2 , (1n ) 2 ,....., (nn 1 )2
2. combining the results according to (*)

p13.
  yk[0]  A[0] (kn / 2 )
Let  [1]
 yk  A (n / 2 )
[1] k

yk  A(kn )  A[0] (2n k )  kn A[1] (2n k )

= A[0] (kn / 2 )  kn A[1] (kn / 2 )
= yk[0]  nk yk[1]

yk  n / 2  A(nk  n / 2 )  A[0] (2n k  n )  kn  n / 2 A[1] (2n k  n )

= A[0] (kn / 2 )  kn  n / 2 A[1] (kn / 2 )
= yk[0]  kn  n / 2 yk[1]  yk[ 0]  kn yk[1]

p14.
Recursive-FFT(a)
{ n=length[a]; /* n: power of 2 */
if n=1 the return a;
n  e 2 i / n ;
=1
a [0]  (a0 , a2 ,....., an  2 );
a [1]  (a1 , a3 ,....., an 1 );
y[0]  Recursive-FFT(a [0] );
y[1]  Recursive-FFT(a [1] );
for k=0 to (n/2 - 1) do
{
y k  y[0]   y[1]
k ;
T (n)  2T (n / 2)  ( n)
k

y k+n/2  y  y ;
[0]
k
[1]
k

=n ; } =(nlog n)
}
 Thm 8 (Convolution thm)
n: power of 2
a, b: vectors of length n
a  b=DFT2-1n (DFT2 n (a)  DFT2 n (b))
Componentwise product

a  b  a0b0 , a0b1  a1b0 , a0b2  a1b1  a2b0 ,...... 

j
j-th elt:  ak b j  k
k=0

p16.
 Efficient FFT implement
yk[0]  y [0]
k  y k
n
[1]
k

kn
yk[1] y [0]
 y k [1]
k n k

(a0 , a1 , a2 , a3 , a4 , a5 , a6 , a7 )
(a0 , a2 , a4 , a6 ) (a1 , a3 , a5 , a7 )
(a0 , a4 ) (a2 , a6 ) (a1 , a5 ) (a3 , a7 )
a0 a4 a2 a6 a1 a5 a3 a7
000 100 010 110 001 101 011 111
Idea :
for s=1 to (lg n) do
for k=0 to n-1 by 2s
combine the two 2 s-1  element DFT's in y[1]
y[0] A[k..k+2 s-1  1] and A[k+2s-1..k+2s-1  1]
into one 2s  element DFT in A[k..k+2s  1]
FFT-Base(a)
{ n = length[a];
for s=1 to (lg n) do
{ m=2s ;
m  e 2 i / m ;
for k=0 to n-1 by m do
{ =1;
for j=0 to (m/2 - 1) do
{ t=A[k  j  m / 2];
u=A[k+j];
A[k+j]=u+t;
A[k+j+m/2]=u-t;
  m ; }
} } }
Iterative-FFT(a)
{ Bit-Reverse-Copy(a, A);
n=length[a];
Bit-Reverse-Copy(a,A)
for s=1 to (lg n) do { n=length[a];
{ m=2s ; m  e 2 i / m ;   1; for k=0 to n-1 do
for j=0 to (m/2 -1) do A[rev(k)] = a k
{ for k=j to n-1 by m do }
{ t = A[k  m / 2]; eg
u = A[k]; rev(011)=110, rev(001)=100

A[k]=u+t;
A[k+m/2]=u-t; }
  m
}
}
}
p19.
FFT circuit
a0 y0
 0
2
a1 y1
04
a2 y2
02 14
a3 y3
 0
8
a4 y4
02 18
a5 y5
 0 82
4
a6 y6
02 14 83
a7 y7

S=1 S=2 S=3

n8
In general, depth: (lg n)
size : (nlg n)