SDSU CS 662 Theory of Parallel Algorithms
Numerical Method part 4 FFT

[To Lecture Notes Index]
San Diego State University -- This page last updated May 7, 1996
----------


FFT - Fast Fourier Transform

Who is Fourier?How did he get so Fast?Who cares?

Bases and functions


Bases for 3-dimension points: (x, y, z)

Bases for functions:
Polynomials
Fourier bases
k = 0, 1, ...
k = 0, 1, ...
Then the Fourier transform of f(x) is:

Example
f(x) = | x | for
The Fourier transform of f(x) is:

Complex numbers, X-Y plane and all that

Let represent the unit vector in the y-axis

Represent points in the plane as x + y

So (1/2, 1) is 1/2 +

Thus (a + b)(c + d) = ac+ bd + (b+d)

but , so

(a + b)(c + d) = ac-bd + (bc+ad)
Roots of Unity

Define as:

where

then

The n'th roots of unity are:
Example

Let n = 4 then
Example

Let n = 8 then

Discrete vs. Continuous (DFT vs FT)

Let

Define for j,k = 0, 1, ..., n-1

If n = 4 then

Let x be an n-dimensional vector, the discrete Fourier transform (DFT) of x is the vector:


Then

Note defines We have
Now the Fast Part

Assume n is a power of 2 and compute

Since we have

But this is the 2p'th component of the DFT of the vector:

Similarly we get:

But this is the 2p+1'th component of the DFT of the vector:

This means to compute the DFT of an n vector we can compute the DFT of two n/2 vectors.

Computing DFT

Theorem
The above processes will compute the DFT of an n-dimensional vector in O(log(n)) time using total of O(nlog(n)) operations
The inverse DFT of an n-dimensional vector can be computed in O(log(n)) time using total of O(nlog(n)) operations

Example Use of FFT
Polynomial Multiplication

Let and

Let

where

So can compute in parallel using log(n+m) time and performing O(mn) operations using scan and n*n processors
Using FFT

Let and

Compute and where

Compute

Compute where

Then for k = 0, 1, ..., n+m-1

So can be computed in O( log(n+m) ) time and O( (n+m)log(n+m) ) operations using O(n) processors
----------