Fast Fourier Transform (FFT)

Borbala Hunyadi

Delft University of Technology, The Netherlands

Recap: Discrete Fourier Transform

Definition
The Discrete Fourier Transform (DFT) of a sequence x[n] is
N−1
X [k] = ∑ x[n]e −j2π N , for 0 ≤ k ≤ N − 1
kn

n=0

Applications:
● Filtering
● Spectral analysis

2 / 13
Recap: Discrete Fourier Transform

Definition
The Discrete Fourier Transform (DFT) of a sequence x[n] is
N−1
X [k] = ∑ x[n]e −j2π N , for 0 ≤ k ≤ N − 1
kn

n=0

Applications:
● Filtering
● Spectral analysis
Is DFT efficient enough?

2 / 13
Computational complexity of the DFT
Let’s define WN = e −j2π/N ! Then the DFT can be expressed as:
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Steps of the direct computation algorithm:

Stage 1:
Compute and store the values
WNl = e −j2πl/N = cos(2πl/N) − j ⋅ sin(2πl/N)
Stage 2:
for k = 0 ∶ N − 1
X [k] ← x[0]
for n = 1 ∶ N − 1
l = (kn)N
X [k] ← X [k] + x[n]WNl
end
end
3 / 13
Computational complexity of the DFT
Let’s define WN = e −j2π/N ! Then the DFT can be expressed as:
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Steps of the direct computation algorithm:

Stage 1:
Compute and store the values N evaluations of sin
WNl = e −j2πl/N = cos(2πl/N) − j ⋅ sin(2πl/N) and cos functions
Stage 2:
for k = 0 ∶ N − 1
X [k] ← x[0]
for n = 1 ∶ N − 1
l = (kn)N
X [k] ← X [k] + x[n]WNl
end
end
3 / 13
Computational complexity of the DFT
Let’s define WN = e −j2π/N ! Then the DFT can be expressed as:
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Steps of the direct computation algorithm:

Stage 1:
Compute and store the values N evaluations of sin
WNl = e −j2πl/N = cos(2πl/N) − j ⋅ sin(2πl/N) and cos functions
Stage 2:
for k = 0 ∶ N − 1 N 2 complex
X [k] ← x[0] multiplications and
for n = 1 ∶ N − 1 N(N − 1) complex
additions
l = (kn)N
X [k] ← X [k] + x[n]WNl
end
end
3 / 13
Computational complexity of the DFT
Let’s define WN = e −j2π/N ! Then the DFT can be expressed as:
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Steps of the direct computation algorithm:

Stage 1:
Compute and store the values N evaluations of sin
WNl = e −j2πl/N = cos(2πl/N) − j ⋅ sin(2πl/N) and cos functions
Stage 2:
for k = 0 ∶ N − 1 N 2 complex
X [k] ← x[0] multiplications and
for n = 1 ∶ N − 1 N(N − 1) complex
additions
l = (kn)N
X [k] ← X [k] + x[n]WNl
+ overhead:
end addressing, indexing...
end
3 / 13
Computational complexity of the DFT
Let’s define WN = e −j2π/N ! Then the DFT can be expressed as:
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Steps of the direct computation algorithm: O(N 2 ) - very costly

Stage 1:
Compute and store the values N evaluations of sin
WNl = e −j2πl/N = cos(2πl/N) − j ⋅ sin(2πl/N) and cos functions
Stage 2:
for k = 0 ∶ N − 1 N 2 complex
X [k] ← x[0] multiplications and
for n = 1 ∶ N − 1 N(N − 1) complex
additions
l = (kn)N
X [k] ← X [k] + x[n]WNl
+ overhead:
end addressing, indexing...
end
3 / 13
Fast Fourier Transform

● A family of computationally efficient algorithms to compute DFT

● Not a new transform!

Different working principles:

1 Divide and conquer approach
2 DFT as convolution: linear filtering approach

4 / 13
Fast Fourier Transform

● A family of computationally efficient algorithms to compute DFT

● Not a new transform!

Different working principles:

1 Divide and conquer approach
2 DFT as convolution: linear filtering approach

4 / 13
Divide and conquer FFT
Essential ingredients:

5 / 13
Divide and conquer FFT
Essential ingredients:
● Break down the N-point DFT to a cascade of smaller-size DFTs

5 / 13
Divide and conquer FFT
Essential ingredients:
● Break down the N-point DFT to a cascade of smaller-size DFTs

Guessing game: I am thinking of a random number between 1 and

16. Can you guess which number is it?

5 / 13
Divide and conquer FFT
Essential ingredients:
● Break down the N-point DFT to a cascade of smaller-size DFTs

Guessing game: I am thinking of a random number between 1 and

16. Can you guess which number is it?

● Exploit symmetries

5 / 13
Divide and conquer FFT
Essential ingredients:
● Break down the N-point DFT to a cascade of smaller-size DFTs

Guessing game: I am thinking of a random number between 1 and

16. Can you guess which number is it?

● Exploit symmetries

WNk = WNk+N
WNLk = WN/L
k

5 / 13
Radix-2 FFT

Radix-2 FFT is the most important divide and conquer type FFT
algorithm. It can be used if N = 2r . This can always be achieved using
zero-padding the sequence.

Decimation in time (DIT) solution:

● Divide the N long sequence x[n] to 2 N/2 long sequences
● The N-point DFT of x[n] can be computed by properly combining
the 2 N/2-point DFTs
● Repeat the subdivision until the sequences are 2 samples long
(2-point DFT)

6 / 13
Radix-2 FFT
N-point DFT (N = 2r ) solved by a cascade of r stages:

The following slides show the exact operations performed during these
stages.
7 / 13
2-point DFT: How to compute in a simple way?
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Let us write out the expression for both DFT coefficients:

X [0] = x[0] + x[1]W20 = x[0] + x[1]

X [1] = x[0] + x[1]W21 = x[0] − x[1]

8 / 13
2-point DFT: How to compute in a simple way?
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Let us write out the expression for both DFT coefficients:

X [0] = x[0] + x[1]W20 = x[0] + x[1]

X [1] = x[0] + x[1]W21 = x[0] − x[1]

8 / 13
2-point DFT: How to compute in a simple way?
N−1
X [k] = ∑ x[n]WNkn , for 0 ≤ k ≤ N − 1
n=0
Let us write out the expression for both DFT coefficients:

X [0] = x[0] + x[1]W20 = x[0] + x[1]

X [1] = x[0] + x[1]W21 = x[0] − x[1]

The 2-point DFT coefficients are given by taking the sum and the
difference of the samples. This simple operation is represented by the
so-called butterfly diagram.
8 / 13
Combine two 2-point DFTs into a 4-point DFT

3
X [k] = ∑ x[n]W4kn , for 0 ≤ k ≤ 3
n=0

X [k] = x[0] + x[1]W4k + x[2]W42k + x[3]W43k

9 / 13
Combine two 2-point DFTs into a 4-point DFT

3
X [k] = ∑ x[n]W4kn , for 0 ≤ k ≤ 3
n=0

X [k] = x[0] + x[1]W4k + x[2]W42k + x[3]W43k

= (x[0] + x[2]W42k ) + (x[1]W4k + x[3]W43k )

Decimation in time: divide the

sum to a sum of even and a
sum of odd samples

9 / 13
Combine two 2-point DFTs into a 4-point DFT

3
X [k] = ∑ x[n]W4kn , for 0 ≤ k ≤ 3
n=0

X [k] = x[0] + x[1]W4k + x[2]W42k + x[3]W43k

= (x[0] + x[2]W42k ) + (x[1]W4k + x[3]W43k )

= (x[0] + x[2]W42k ) + W4k (x[1] + x[3]W42k )

9 / 13
Combine two 2-point DFTs into a 4-point DFT

3
X [k] = ∑ x[n]W4kn , for 0 ≤ k ≤ 3
n=0

X [k] = x[0] + x[1]W4k + x[2]W42k + x[3]W43k

= (x[0] + x[2]W42k ) + (x[1]W4k + x[3]W43k )

= (x[0] + x[2]W42k ) + W4k (x[1] + x[3]W42k )

= (x[0] + x[2]W2k ) + W4k (x[1] + x[3]W2k )

using the property

WNLK = WN/Lk
N = 4 and
L=2
9 / 13
Combine two 2-point DFTs into a 4-point DFT

3
X [k] = ∑ x[n]W4kn , for 0 ≤ k ≤ 3
n=0

X [k] = x[0] + x[1]W4k + x[2]W42k + x[3]W43k

= (x[0] + x[2]W42k ) + (x[1]W4k + x[3]W43k )

= (x[0] + x[2]W42k ) + W4k (x[1] + x[3]W42k )

= (x[0] + x[2]W2k ) + W4k (x[1] + x[3]W2k ) = G [k] + W4k H[k]

G [k] ≡ x[0] + x[2]W2k is the H[k] ≡ x[1] + x[3]W2k is the

2-point DFT of even samples 2-point DFT of odd samples

9 / 13
4-point DFT from 2-point DFTs

X [k] = G [k] + W4k H[k]

X [0] = G [0] + H[0]

X [1] = G [1] + W4 H[1]
X [2] = G [2] + W42 H[2]
X [3] = G [3] + W43 H[3]

10 / 13
4-point DFT from 2-point DFTs

X [k] = G [k] + W4k H[k]

X [0] = G [0] + H[0]

X [1] = G [1] + W4 H[1]
X [2] = G [2] + W42 H[2] = G [0] + W42 H[0]
X [3] = G [3] + W43 H[3] = G [1] + W43 H[1]

G [k] and H[k] are 2-point

DFTs, hence, 2-periodic

10 / 13
4-point DFT from 2-point DFTs

X [k] = G [k] + W4k H[k]

X [0] = G [0] + H[0]

X [1] = G [1] + W4 H[1]
X [2] = G [2] + W42 H[2] = G [0] + W42 H[0]
X [3] = G [3] + W43 H[3] = G [1] + W43 H[1]

10 / 13
General case: N-point DFT from N/2-point DFTs

N−1
X [k] = ∑ x[n]WNkn
n=0

11 / 13
General case: N-point DFT from N/2-point DFTs

N−1
X [k] = ∑ x[n]WNkn
n=0

N/2−1 N/2−1
= ∑ x[2r ]WN2kr + WNk ∑ x[2r + 1]WN2kr
r =0 r =0

Decimation in time: divide the

sum to a sum of even and a
sum of odd samples

11 / 13
General case: N-point DFT from N/2-point DFTs

N−1
X [k] = ∑ x[n]WNkn
n=0

N/2−1 N/2−1
= ∑ x[2r ]WN2kr + WNk ∑ x[2r + 1]WN2kr
r =0 r =0

N/2−1 N/2−1
kr
= ∑ x[2r ]WN/2 + WNk ∑ x[2r + 1]WN/2
kr
r =0 r =0

using the property

WNLK = WN/L
k

11 / 13
General case: N-point DFT from N/2-point DFTs

N−1
X [k] = ∑ x[n]WNkn
n=0

N/2−1 N/2−1
= ∑ x[2r ]WN2kr + WNk ∑ x[2r + 1]WN2kr
r =0 r =0

N/2−1 N/2−1
kr
= ∑ x[2r ]WN/2 + WNk ∑ x[2r + 1]WN/2
kr
= G [k] + WNk H[k]
r =0 r =0

G [k] is the N/2-point DFT H[k] is the N/2-point DFT

of even samples, hence N/2- of odd samples, hence N/2-
periodic periodic

11 / 13
Example: 8-point FFT

Start with 2-point DFTs of samples arranged in bit-reversed order and

combine the results in each stage! Note that the butterflies can be
k+N/2
further simplfied with WN = −WNk

12 / 13
Computational complexity of Radix-2 FFT
● v = log2 N stages
● per stage, there are N/2 butterflies
● per butterly, 1 complex multiplication and 2 complex additions
Total: log2 N ⋅ N/2 complex multiplications and log2 N ⋅ N complex
additions, i.e. O(Nlog2 N).

13 / 13