From Fourier Series to

Fourier Transform to FFT

 ∞

𝑥𝑇 𝑡 = 𝑐𝑘 𝑒 𝑖𝑘𝜔0𝑡 ,  𝑡 ∈ ℝ, 𝜔0 = 2𝜋
𝑇
𝑘=−∞
𝑇/2
1
𝑐𝑘 = 𝑥(𝑡)𝑒 −𝑖𝑘𝜔𝑜 𝑡 𝑑𝑡,  𝑘 = 0, ±1, ±2, …
𝑇
−𝑇/2
x(t )  lim xT (t )
T 
• What happens to the frequency components of 𝑥𝑇 (𝑡) as 𝑇 → ∞ ?
• For 𝑘 = 0:
𝑐0 = 1/𝑇
• For 𝑘 = ±1, ±2, … :

2 𝑘𝜔0 1 𝑘𝜔0
𝑐𝑘 = sin = sin
𝑘𝜔0 𝑇 2 𝑘𝜋 2

𝜔0 = 2𝜋/𝑇
T  2,5,10
 𝜔
lim 𝑇 𝑐𝑘 = sinc ,  𝜔 ∈ ℝ
𝑇→∞ 2𝜋

sin( 𝜋𝜆)
sinc( 𝜆) ≐
𝜋𝜆
• Given a signal x(t), its Fourier transform 𝑋(𝜔) is defined as
∞

𝑋(𝜔) = 𝑥(𝑡)𝑒 −𝑖𝜔𝑡 𝑑𝑡,  𝜔 ∈ ℝ

−∞

• A signal x(t) is said to have a Fourier transform in the ordinary sense if

the above integral converges
• The integral does converge if
1. the signal x(t) is “well-behaved”
2. and x(t) is absolutely integrable, namely,
∞

|𝑥(𝑡)|𝑑𝑡 < ∞
−∞
• Note: well behaved means that the signal has a finite number of
discontinuities, maxima, and minima within any finite time interval
• Given a signal x(t) with Fourier transform 𝑋(𝜔), x(t) can be recomputed
from 𝑋(𝜔) by applying the inverse Fourier transform given by
∞
1
𝑥(𝑡) = 𝑋(𝜔)𝑒 𝑖𝜔𝑡 𝑑𝜔,  𝑡 ∈ ℝ
2𝜋
−∞

• Transform pair
𝑥(𝑡) ↔ 𝑋(𝜔)
 Properties of the Fourier Transform

x(t )  X ( ) y (t )  Y ( )
• Linearity:
 x(t )   y(t )   X ( )   Y ( )
• Left or Right Shift in Time:

𝑥(𝑡 − 𝑡0 ) ↔ 𝑋(𝜔)𝑒 −𝑖𝜔𝑡0

• Time Scaling:
1  
x(at )  X  
a a
Properties of the Fourier Transform

• Time Reversal:
x(t )  X ( )
• Multiplication by a Power of t:
𝑛
𝑑
𝑡 𝑛 𝑥(𝑡) ↔ (𝑖)𝑛 𝑛
𝑋(𝜔)
𝑑𝜔
• Multiplication by a Complex Exponential:

𝑥(𝑡)𝑒 𝑖𝜔0 𝑡 ↔ 𝑋(𝜔 − 𝜔0 )

Properties of the Fourier Transform

• Multiplication by a Sinusoid (Modulation):

𝑖
𝑥(𝑡) sin( 𝜔0 𝑡) ↔ 𝑋(𝜔 + 𝜔0 ) − 𝑋(𝜔 − 𝜔0 )
2
1
𝑥(𝑡) cos( 𝜔0 𝑡) ↔ 𝑋(𝜔 + 𝜔0 ) + 𝑋(𝜔 − 𝜔0 )
2
• Differentiation in the Time Domain:

𝑑𝑛 𝑛
𝑛
𝑥(𝑡) ↔ (𝑖𝜔) 𝑋(𝜔)
𝑑𝑡
Properties of the Fourier Transform
• Integration in the Time Domain:
𝑡
1
𝑥(𝜏)𝑑𝜏 ↔ 𝑋(𝜔) + 𝜋𝑋(0)𝛿(𝜔)
𝑖𝜔
−∞
• Convolution in the Time Domain:
x(t )  y (t )  X ( )Y ( )
• Multiplication in the Time Domain:
x(t ) y(t )  X ( )  Y ( )
Properties of the Fourier Transform
• Parseval’s Theorem:
1
 x(t ) y(t )dt  2 X ( )Y ( )d


1
if y (t )  x(t )  | x(t ) | dt 
2
 | X ( ) | d
2

2
• Duality:

X (t )  2 x( )
Example: Linearity
x(t )  p4 (t )  p2 (t )

 2   
X ( )  4sinc    2sinc  
    
Example: Time Shift

x(t )  p2 (t  1)

𝜔 −𝑖𝜔
𝑋(𝜔) = 2sinc 𝑒
𝜋
Example: Time Scaling

 
p2 (t ) 2sinc  
 

 
sinc  
p2 (2t )  2 

a  1 time compression  frequency expansion

0  a  1 time expansion  frequency compression
Example: Multiplication by a Sinusoid

x(t )  p (t ) cos(0t ) sinusoidal

burst

1   (   0 )    (   0 )  
X ( )   sinc     sinc  
2  2   2 
 Example: Multiplication by a
Sinusoid – Cont’d

1   (   0 )    (   0 )  
X ( )   sinc     sinc  
2  2   2 

 0  60 rad / sec

  0.5
Generalized Fourier Transform of Sinusoidal
Signals

cos(0t )    (  0 )   (  0 )

sin( 𝜔0 𝑡) ↔ 𝑖𝜋 𝛿(𝜔 + 𝜔0 ) − 𝛿(𝜔 − 𝜔0 )

 Image Processing

Optical
scene System
image

𝑓 𝑔 ℎ ℎ =𝑓∗𝑔
 Convolution

ℎ 𝑥 = 𝑓 𝜏 𝑔 𝑥 − 𝜏 𝑑𝜏 𝑔 =𝑓∗ℎ
0
 Convolution Theorem
Here we derive the most important formula of image processing and other applications of Fourier Transform

Let g  f h
∞
Then 𝐺 𝜔 = 𝑔 𝑥 𝑒 −𝑖2𝜋𝜔𝑥 𝑑𝑥
−∞
∞ ∞
From = 𝑓 𝜏 ℎ 𝑥 − 𝜏 𝑒 −𝑖2𝜋𝜔𝑥 𝑑𝜏𝑑𝑥Add and
definition −∞ −∞ subtract tau
∞ ∞
= 𝑓 𝜏 𝑒 −𝑖2𝜋𝜔𝜏 𝑑𝜏 ℎ 𝑥 − 𝜏 𝑒 −𝑖2𝜋𝜔 𝑥−𝜏 𝑑𝑥
−∞ −∞
∞ ∞
= 𝑓 𝜏 𝑒 −𝑖2𝜋𝜔𝜏 𝑑𝜏 ℎ 𝑥′ 𝑒 −𝑖2𝜋𝜔𝑥′ 𝑑𝑥′ Separate
domains
−∞ −∞
=𝐹 𝜔 𝐻 𝜔
Convolution in spatial domain  Multiplication in frequency domain
Discrete Fourier Transform (DFT)
• The DFT provides uniformly spaced samples of the Discrete-Time
Fourier Transform (DTFT)
• DFT definition:
𝑁−1 𝑁−1
2𝜋𝑛𝑘 1 2𝜋𝑛𝑘
𝑖 𝑁
𝑋[𝑘] = 𝑥[𝑛]𝑒 −𝑖 𝑁 𝑥[𝑛] = 𝑋[𝑘]𝑒
𝑁
𝑛=0 𝑛=0
• Requires N2 complex multiplies and N(N-1) complex additions
 Faster DFT computation?
• Take advantage of the symmetry
2𝑖𝜋
and periodicity of the complex
exponential (let 𝑊𝑁 = 𝑒 𝑁 )
• symmetry:
𝑘[𝑁−𝑛]
𝑊𝑁 = 𝑊𝑁−𝑘𝑛 = (𝑊𝑁𝑘𝑛 )∗
• periodicity:
𝑘[𝑛+𝑁] [𝑘+𝑁]𝑛
𝑊𝑁𝑘𝑛 = 𝑊𝑁 = 𝑊𝑁
• Note that two length N/2 DFTs take less computation than one length N
DFT: 2(N/2)2 < N2
• Algorithms that exploit computational savings are collectively called Fast
Fourier Transforms
 Decimation-in-Time Algorithm
• Consider expressing DFT with even and odd input samples:
𝑁−1

𝑋[𝑘] = 𝑥[𝑛]𝑊𝑁𝑛𝑘
𝑛=0

= 𝑥[𝑛]𝑊𝑁𝑛𝑘 + 𝑥[𝑛]𝑊𝑁𝑛𝑘
𝑛 𝑒𝑣𝑒𝑛 𝑛 𝑜𝑑𝑑
𝑁 𝑁
2 −1 2 −1
= 𝑥[2𝑟](𝑊𝑁2 )𝑟𝑘 + 𝑊𝑁𝑘 𝑥[2𝑟 + 1](𝑊𝑁2 )𝑟𝑘
𝑟=0 𝑟=0
𝑁 𝑁
2 −1 2 −1
𝑟𝑘 𝑟𝑘
= 𝑥[2𝑟]𝑊𝑁/2 + 𝑊𝑁𝑘 𝑥[2𝑟 + 1]𝑊𝑁/2
𝑟=0 𝑟=0