The Discrete
Fourier Transform
The Fourier Transform
• “The Fourier transform is a mathematical operation
with many applications in physics and engineering
that expresses a mathematical function of time as a
function of frequency, known as its frequency
spectrum.”
– from [Link]
The Fourier Transform
• “For instance, the transform of a musical chord made
up of pure notes (without overtones) expressed as
amplitude as a function of time, is a mathematical
representation of the amplitudes and phases of the
individual notes that make it up.”
– from [Link]
Amplitude & phase
• f(x) = sin( x + ) +
– is the amplitude
– is the frequency
– is the phase
– is the DC offset
More generally
• f(x) = 1 sin( 1 x + 1 ) + 2 sin( 2 x + 2 ) +
The Fourier Transform
• “The function of time is often called the time domain
representation, and the frequency spectrum the
frequency domain representation.”
– from [Link]
Applications
• differential equations
• geology
• image and signal processing
• optics
• quantum mechanics
• spectroscopy
REVIEW OF COMPLEX NUMBERS
Complex numbers
Complex numbers . . .
• extend the 1D number line to the 2D plane
• are numbers that can be put into the
rectangular form, a+bi where i2 = -1, and a and
b are real numbers.
Complex numbers
(rectangular form)
z x yi
Complex numbers
Complex numbers . . .
• a is the real part; b is the imaginary part
• If a is 0, then a+bi is purely imaginary; if b is 0,
then a+bi is a real number.
• originally called “fictitious” by Girolamo
Cardano in the 16th century
Complex arithmetic
• add/subtract
– add/subtract the real and imaginary parts
separately
Complex arithmetic
• complex conjugate
– often denoted as z * or z
– negate only the imaginary part
Complex arithmetic
• inverse
1 z z
2
z zz a b 2
where
z is a complex number
z bar is the length or magnitude of z
a is the real part
b is the imaginary part
Complex arithmetic
• multiplication (FOIL)
a bi c di ac bd bc ad i
Complex arithmetic
• division
a bi ac bd bc ad
2
2 2 2
i
c di c d c d
a bi c di
c di c di
complex conjugate of denominator
Complex numbers z x yi
(polar form) r e i
r cos i r sin
r ei ( 2 n )
where
r z x2 y2
y
arg z tan
1
x
n is an integer and
x r cos
y r sin
exponential vs. trigonometric
e cos x i sin x
ix
(phasor form)
ix
e e
ix
sin x
2i
ix
e e
ix
cos x
2 Leonhard Euler 1707-1783
DFT
(DISCRETE FOURIER TRANSFORM)
DFT
• Say we have a sequence of N (possibly
complex) numbers, x0 … xN-1.
• The DFT produces a sequence of N (typically
complex) numbers, X0 … XN-1, via the
following:
N 1
X k xn e
i 2 k
N
n
n 0
DFT & IDFT
• The DFT (Discrete Fourier Transform) produces
a sequence of N (typically complex) numbers,
X0 … XN-1, via the following:
N 1
X k xn e
i 2 k
N
n
n 0
• The IDFT (Inverse DFT) is defined as follows:
N 1
1
X
i 2 k
xk
n
k e N
N k 0
Calculating the DFT
N 1
x e
i 2 k
So how can we actually calculate X k
n
n
N
?
n 0
Calculating the DFT
N 1
X k xn e
i 2 k n
• So how can we calculate N
?
n 0
• Let’s use this relationship: e cos x i sin x
ix
• Then
N 1
X k xn cos 2 k
N n i sin 2 k
N n
n 0
• So what does this mean?
Interpretation of DFT
N 1
X k xn cos 2 k
N n xni sin 2 k
N n
n 0
Back to the polar form:
– r/N is the amplitude and is the phase of a
sinusoid with frequency k/N into which xn is
decomposed
CALCULATING THE DFT USING
EXCEL
x_n X_0 X_ 1 X_ 2
N x b sin(b x) b2 sin(b2 x) 0.5*sin(b x) + 0.5*sin(b2 x) real imag real imag real imag
500 0 0.05 0.00 0.1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
1 0.05 0.10 0.07 0.07 0.00 0.07 0.00 0.07 0.00
2 0.10 0.20 0.15 0.15 0.00 0.15 0.00 0.15 -0.01
3 0.15 0.30 0.22 0.22 0.00 0.22 -0.01 0.22 -0.02
4 0.20 0.39 0.29 0.29 0.00 0.29 -0.01 0.29 -0.03
5 0.25 0.48 0.36 0.36 0.00 0.36 -0.02 0.36 -0.05
6 0.30 0.56 0.43 0.43 0.00 0.43 -0.03 0.43 -0.06
7 0.34 0.64 0.49 0.49 0.00 0.49 -0.04 0.49 -0.09
8 0.39 0.72 0.55 0.55 0.00 0.55 -0.06 0.54 -0.11
9 0.43 0.78 0.61 0.61 0.00 0.61 -0.07 0.59 -0.14
10 0.48 0.84 0.66 0.66 0.00 0.66 -0.08 0.64 -0.16
11 0.52 0.89 0.71 0.71 0.00 0.70 -0.10 0.68 -0.19
12 0.56 0.93 0.75 0.75 0.00 0.74 -0.11 0.71 -0.22
13 0.61 0.96 0.78 0.78 0.00 0.77 -0.13 0.74 -0.25
14 0.64 0.99 0.81 0.81 0.00 0.80 -0.14 0.76 -0.28
15 0.68 1.00 0.84 0.84 0.00 0.82 -0.16 0.78 -0.31
16 0.72 1.00 0.86 0.86 0.00 0.84 -0.17 0.79 -0.34
17 0.75 0.99 0.87 0.87 0.00 0.85 -0.18 0.79 -0.36
18 0.78 0.97 0.88 0.88 0.00 0.86 -0.20 0.79 -0.38
19 0.81 0.95 0.88 0.88 0.00 0.85 -0.21 0.78 -0.40
20 0.84 0.91 0.88 0.88 0.00 0.85 -0.22 0.77 -0.42
21 0.87 0.86 0.87 0.87 0.00 0.84 -0.23 0.75 -0.44
22 0.89 0.81 0.85 0.85 0.00 0.82 -0.23 0.72 -0.45
23 0.91 0.75 0.83 0.83 0.00 0.79 -0.24 0.69 -0.45
24 0.93 0.68 0.80 0.80 0.00 0.77 -0.24 0.66 -0.46
25 0.95 0.60 0.77 0.77 0.00 0.74 -0.24 0.63 -0.45
26 0.96 0.52 0.74 0.74 0.00 0.70 -0.24 0.59 -0.45
27 0.98 0.43 0.70 0.70 0.00 0.66 -0.23 0.55 -0.44
28 0.99 0.33 0.66 0.66 0.00 0.62 -0.23 0.50 -0.43
29 0.99 0.24 0.62 0.62 0.00 0.58 -0.22 0.46 -0.41
30 1.00 0.14 0.57 0.57 0.00 0.53 -0.21 0.42 -0.39
31 1.00 0.04 0.52 0.52 0.00 0.48 -0.20 0.37 -0.37
Check w/ matlab/octave
% see X0real = xn .* cos( -2*pi*n*0/N );
[Link] X0imag = xn .* sin ( -2*pi*n*0/N );
b/ref/[Link]
X1real = xn .* cos( -2*pi*n*1/N );
N = 256; % # of samples X1imag = xn .* sin ( -2*pi*n*1/N );
n = (0:N-1); % subscripts
b1 = 0.5; % freq 1 X2real = xn .* cos( -2*pi*n*2/N );
b2 = 2.5; % freq 2 X2imag = xn .* sin ( -2*pi*n*2/N );
xn = 0.5 * sin( b1*n ) + 0.2 * sin( b2*n ); X3real = xn .* cos( -2*pi*n*3/N );
X3imag = xn .* sin ( -2*pi*n*3/N );
plot( xn ); .
Xn = fft( xn ); .
plot( abs(Xn(1:N/2)) ); . Note: .* is element-wise (rather
than matrix) multiplication in
matlab.
Add random noise.
% see X0real = xn .* cos( -2*pi*n*0/N );
[Link] X0imag = xn .* sin ( -2*pi*n*0/N );
b/ref/[Link]
X1real = xn .* cos( -2*pi*n*1/N );
N = 256; % # of samples X1imag = xn .* sin ( -2*pi*n*1/N );
n = (0:N-1); % subscripts
b1 = 0.5; % freq 1 X2real = xn .* cos( -2*pi*n*2/N );
b2 = 2.5; % freq 2 X2imag = xn .* sin ( -2*pi*n*2/N );
r = randn( 1, N ); % noise
xn = 0.5 * sin( b1*n ) + 0.2 * sin( b2*n ) + X3real = xn .* cos( -2*pi*n*3/N );
0.5 * r;
X3imag = xn .* sin ( -2*pi*n*3/N );
plot( xn );
.
Xn = fft( xn );
.
plot( abs(Xn(1:N/2)) );
.
Signal without and with noise.
Signal with noise.
FFT of noisy signal (two major
components are still apparent).
Example of differences in phase.
xn = 0.5 * sin( b1*n ) + 0.2 * sin( b2*n )
xn = 0.5 * sin( b1*n – 0.5 ) + 0.2 * sin( b2*n )
Computational complexity:
DFT vs. FFT
• The DFT is O(N2) complex multiplications.
• In 1965, Cooley (IBM) and Tukey (Princeton)
described the FFT, a fast way (O(N log2 N)) to
compute the FT using digital computers.
– It was later discovered that Gauss described this
algorithm in 1805, and others had “discovered” it
as well before Cooley and Tukey.
– “With N = 106, for example, it is the difference
between, roughly, 30 seconds of CPU time and 2
weeks of CPU time on a microsecond cycle time
computer.” – from Numerical Recipes in C
Extending the DFT to 2D
(and higher)
• Let f(x,y) be a 2D set of sampled points. Then
the DFT of f is the following:
M 1 N 1
F u , v f x, y e
1 j 2 ux / M vy / N
MN x 0 y 0
M 1 N 1
f x, y F u , v e j 2 ux / M vy / N
u 0 v 0
• (Note that engineers often use i for amps
(current) so they use j for -1 instead.)
Extending the DFT to 2D
(and higher)
• In fact, the 2D DFT is separable so it can be
decomposed into a sequence of 1D DFTs.
M 1 N 1
F u, v f x, y e
1 1
e j 2ux / N
N
j 2vy / N
M x 0 y 0
• And this can be generalized to higher and
higher dimensions as well.
The classical “Gibbs phenomenon”
• Visit
[Link]
• Hear it at
[Link]
uvY.
