EE 422G Notes: Chapter 10

Instructor: Zhang

Chapter 10 The Discrete Fourier Transform and Fast Fourier Transform Algorithms 10-1 Introduction
1. x(t) --- Continuous-time signal X(f) --- Fourier Transform, frequency characteristics Can we find
ft X ( f ) =x (t )e j 2 dt

if we dont have a mathematical equation for x(t) ? No! 2. What can we do? (1) Sample x(t) => x0, x1, , xN-1 over T (for example 1000 seconds) t Sampling period (interval) N (samples) over T => t = T / N Can we have infinite T and N? Impossible! (2) Discrete Fourier Transform (DFT):
j 2kn / N => X k = x n e n =0 N 1

k = 1, 2,..., N
k T

for the line spectrum at frequency k = (2 ) 3. Limited N and T => limited frequency resolution 2
1 T

limited frequency band (from to in Fourier transform to):

j 2kn / N 4. xn = N X k e ---- periodic function (period N) k =0 x(t) --- general function sampling and inverse transform xn --- periodic function

N 1

0 < 2N / T

5. X k

( k = 2

k T

line spectrum) period function (period N)

Xk =

N 1 n =0

xn e j 2kn / N

Page 10-1

EE 422G Notes: Chapter 10

Instructor: Zhang

10-2 Error Sources in the DFT

1. Preparations (1) Ideal sampling waveform

y s (t ) =

m =

(t mTs ) :

y s (t ) f s ( f nf s ) n =
Fourier Transform Pair

( fs =

1 ) Ts

in DFT Ts = t = T / N (2) Rectangular Pulse (window)

t t 0 ) T sin c (Tf )e j 2t0 f T

width of the window

Center of the window

sin x x sin Tf sin c (Tf ) = Tf sin cX =

when t0 = 0

t ( ) T sin c (Tf ) T

(3) x1 (t ) * x2 (t ) X 1 ( f ) X 2 ( f ) (4) x1 (t ) x2 (t ) X 1 ( f ) * X 2 ( f )

2. Illustration of Error Sources Example 10-1: Continuous time signal: two-sided exponential signal
x (t ) = e
|t |

Its Fourier transform

Page 10-2

EE 422G Notes: Chapter 10

X(f ) = 2 1 + (2f ) 2 ( ,+) in
not lim ited as

Instructor: Zhang

(1) If we sample x(t)

in

DFT

with sampling frequency fs :

sampled signal x (t ) = y (t )e s s its Fourier transform:

X s ( f ) = Ys ( f ) * F ( e X s ( f ) = fs
n =

|t |

|t |

) = Ys ( f ) * X ( f )
n =

X ( f nf s ) = 2f s

1 + (2 ( f

1 nf s )) 2

sampling => (1) possible overlapping if f s 2 f h is not held. (2) periodic function, introduce frequencies beyond fs . (2) Limited T (over which x(t) is sampled to collect data for DFT)
t ) T t x sw (t ) = x s (t )( ) T

window

X sw ( f ) = ( 2f s

n =

{1 + 2 ( f

nf s ) 2 }1 ) * (T sin c(Tf ))

Fourier transform given by sampled data in limited window (T) X sw ( f ) is a worse estimate of X(f) than Xs(f) due to the introduction of ( Tsinc( Tf ) ) for convolution! Effect of limited T X ( f ) (3) Dose DFT give s for every f ? No! only discrete frequencies. DFT as an estimate for X(f): even worse than X s ( f ) due to the limited frequency resolution.

Page 10-3

EE 422G Notes: Chapter 10

Instructor: Zhang

3. Effect of sampling frequency (or number of points) on accuracy when T is given: Example for e 4. Effect of T (window size) use Compare
t ( ) 4 t ( ) 4
|t |

and

t ( ) 8

for e

|t |

Page 10-4

EE 422G Notes: Chapter 10

Instructor: Zhang

Page 10-5

EE 422G Notes: Chapter 10

Instructor: Zhang

5. DFT Errors (1) Aliasing

X s ( f ) = fs
n =

X ( f

nf s )

superposition of the analog signal spectrum X(f) and its translates (X(f-nfs) n0)

Caused by sampling Overlapping of X(f) and its translates: aliasing (sampling effect) (2) Leakage Effect limited window size T
( )) ( T t ( f ) = F (( )) T t t X s ( f ) = X s ( f ) * F (( )) T

worse than Xs(f) as approximation of X(f).

X s ( f ) : contribution of X s ( f + f ) to X s ( f ) : determined by
( f) weight frequency energy leaks from one frequency to another!

(3) Picket Fence Effect: As an estimation of X(f), does X s ( f ) have picket fence effect? No! DFT: discrete frequencies (not blocked by the fence). 6. Minimization of DFT Error Effects.

Major ways: increase T and fs Problem: DFT for large N.

10-3 Examples Illustrating the computation of the DFT

(Preparation for Mathematical Derivation of FFT)
Page 10-6

EE 422G Notes: Chapter 10

Instructor: Zhang

1. DFT Algorithm
X ( k ) = x ( n )e
n =0 N 1 j 2kn / N

= x ( n ) e j 2 / N
n =0

N 1

nk

Denote WN = e
X (k ) =
N 1 n =0

j 2 / N

, then

x(n)WN nk
WN N = e j 2 = 1

Properties of WN m : (1) WN 0 = (e j 2 / N )0 = e 0 = 1, N +m m = WN (2) W N

WN N + m = (e j 2 / N ) N + m = (e j 2 / N ) N (e j 2 / N ) m = 1 (e j 2 / N ) m = WN m

(3) WN N / 2 = e j 2 /( N / 2) / N = e j = 1

WN N / 4 = e j 2 /( N / 4) / N = e j / 2 = j
WN
3N / 4

= e j 2 /(3 N / 4) / N = e j 3 / 2 = j

2. Examples Example 10-3: Two-Point DFT

nk x(0), x(1): X (k ) = x(n)W2 n =0 1 1 1

k = 0,1

X (0) = x(n)W2 n 0 = x(n) = x(0) + x(1)

n =0
1

n =0
1

X (1) = x (n)W2
n =0 0

n1

= x (n)W2
n =0 1 (1 / 2 ) 2

= x(0)W2 + x(1)W2 = x(0) + x(1)W2 = x(0) x(1) = x(0) + x(1)(1)

Example 10-4: Generalization of derivation in example 10-3 to a four-point DFT x(0), x(1), x(2), x(3)
X (k ) =
n =0

x(n)W4 nk

k = 0,1 ,2,3,

Page 10-7

EE 422G Notes: Chapter 10

X ( 0) =

Instructor: Zhang
3

n =0
3

x(n)W4 n0 = x(n) = x(0) + x(1) + x(2) + x(3)

n =0
n 0 1 2 3

X (1) = x ( n)W4 = x (0)W4 + x (1)W4 + x ( 2)W4 + x (3)W4

n =0

= x (0) jx (1) x ( 2) + jx (3)

X ( 2) = x(n)W4
n =0

2n

= x (0)W4 + x(1)W4 + x( 2)W4 + x(3)W4

2

= x(0) + x(1)(1) + x(2)(1) + x(3)W4 = x(0) x(1) + x( 2) x(3)

X (3) = x( n)W4
n =0 3 3n 0

= x(0)W4 + x(1)W4 + x( 2)W4 + x(3)W4

3 2 1

= x(0) + x(1)W4 + x( 2)(1)W4 + x (3)W4 = x(0) + jx(1) + ( 1) x( 2) + ( j ) x (3) = x(0) + jx(1) x ( 2) jx (3)
X (0) =[ x (0) + x ( 2)] +[ x (1) + x (3)]

X (1) =[ x (0) x ( 2)] +( j )[ x (1) x (3)] X ( 2) =[ x (0) + x ( 2)] [ x (1) + x (3)] X (3) =[ x (0) x ( 2)] + j[ x (1) x (3)]

Two point DFT If we denote z(0) = x(0), z(1) = x(2) => Z(0) = z(0) + z(1) = x(0) + x(2) Z(1) = z(0) - z(1) = x(0) - x(2) v(0) = x(1), v(1) = x(3) => V(0) = v(0) + v(1) = x(1) + x(3) V(1) = v(0) - v(1) = x(1) - x(3) Four point DFT Two-point DFT

X(0) = Z(0) + V(0) X(1) = Z(1) + (-j)V(1)

Page 10-8

EE 422G Notes: Chapter 10

Instructor: Zhang

X(2) = Z(0) - V(0) X(3) = Z(1) + jV(1)

One Four point DFT

Two Two point DFT

10-4 Mathematical Derivation of the FFT

10-4A Decimation-in-Time FFT Algorithm x(0), x(1), , x(N-1)
N = 2m

N g ( 0 ), g ( 1 ), , g ( 1) 2 (( x(0), x (2),, x( N 2)) => h(0), h(1),, h( N 1) 2 (( x(1), x(3),, x( N 1))

enen

N po int s 2 ( g ( r ) = x (2r )) N odd po int s 2 (h( r ) = x(2r + 1))

Page 10-9

EE 422G Notes: Chapter 10

X (k ) = x( n)W N
n =0 N 1 kn

Instructor: Zhang

= =
2 kr

N / 2 1 r =0

g (r )WN g (r )WN
r =0

k (2r )

N / 2 1 r =0 k

h(r )W
N / 2 1 r =0

k ( 2 r +1) N

(k = 0,1,..., N 1)

N / 2 1

2 kr

+ WN

h(r )W

2 kr N

WN

= (e j 2 . / N ) 2 kr = (e j 2 /(. N / 2 ) ) kr = W N
2 N / 2 1 r =0

kr

X (k ) =

g (r )W

kr N /2 k

+WN

N / 2 1 r =0

h(r )W

kr N /2

= G ( k ) + W N H (k )

( G(k): N/2 point DFT output (even indexed), H(k) : N/2 point DFT output (odd indexed))
X ( k ) = G ( k ) + WN k H ( k ) G (k ) = H (k ) =
N / 2 1 r =0

k = 0,1,..., N 1
N / 2 1 r =0

g ( r )WN / 2

kr

x ( 2r )WN / 2

kr

N / 2 1 r =0

h ( r )WN / 2 kr =

N / 2 1 r =0

x ( 2r + 1)WN / 2 kr

Question: X(k) needs G(k), H(k), k= N-1 How do we obtain G(k), H(k), for k > N/2-1 ? G(k) = G(N/2+k) k <= N/2-1 H(k) = H(N/2+k) k <= N/2-1

Page 10-10

EE 422G Notes: Chapter 10

Instructor: Zhang

Future Decimation g(0), g(1), , g(N/2-1) h(0), h(1), , h(N/2-1)

N g (0), g ( 2), , g ( 2) 2 N ge(0), ge(1),...ge( 1) 4 N g (1), g (3), , g ( 1) 2 N go(0), go(1),...go( 1) 4

G(k) H(k)
G (k ) = =
N / 4 1 m =0 N / 2 1 r =0

g (r )W
N / 4 1 m =0

kr

( N / 2)

ge(m)W
k

km

( N / 4)

+W( N / 2 )

go(m)W
k

km ( N / 4)

= GE ( k ) +W( N / 2 ) Go( k )

even indexed g (N/4 point)

WN / 2 k = WN 2 k ?

odd indexed g (N/4 point)

Page 10-11

EE 422G Notes: Chapter 10

WN / 2 k = ( e j 2 /( N / 2 ) ) k = ( e j 2 2 / N ) k = ( e j 2 / N ) 2 k = WN 2 k

Instructor: Zhang

=> G ( k ) = GE ( k ) + WN 2 k Go( k ) Similarly,

H ( k ) = HE ( k ) + WN 2 k Ho(k )

even indexed h (N/4 point) For 8 point

g ( 0) x ( 0) ge( 0) g (1) x ( 2) go (0) g ( 2) x ( 4) ge(1) g ( 3) x ( 6) go(1)

odd indexed h (N/4 point)

h ( 0) x (1) he( 0)

h (1) x ( 3) ho(0)

h ( 2) x (5) he(1)

h ( 3) x ( 7) ho(1)

Page 10-12

EE 422G Notes: Chapter 10

Instructor: Zhang

10-4B Decimation-in-Frequency FFT Algorithm

x(0), x(1), , x(N-1)
X ( k ) = x ( n)W N
n =0 N 1 nk

N = 2m

N / 2 1 n =0

x(n)W N

nk

n =N / 2

x(n)W

N 1

nk N

let m = n-N/2 (n = N/2+m)

n = N/2 => m = N/2-N/2 = 0

Page 10-13

EE 422G Notes: Chapter 10

Instructor: Zhang

n = N-1 => m = N-1-N/2 = N/2-1

X (k ) = =
N / 2 1 n =0
N

N / 2 1 n =0

x ( n )WN nk +
N / 2 1 m =0

N / 2 1 m =0

x ( N / 2 + m )WN ( N / 2 + m ) k
N

x ( n )WN nk +
N

x ( N / 2 + m )WN mkWN 2 k

WN 2 = 1 WN 2 k = ( 1) k X (k ) = =
N / 2 1 n =0 N / 2 1 n =0

x ( n )WN nk +

N / 2 1 m =0

( 1) k x ( N / 2 + m )WN mk

[ x ( n ) + ( 1) k x ( N / 2 + n )]WN nk
N / 2 1 n =0

k : even (k = 2r ) X (k ) = X (2r ) = WN
N/2 point DFT
2 rn

[ x(n) + x( N / 2 + n)]W

2 rn N rn

= (e

j 2 / N

) 2 rn = (e j 2 /( N / 2 ) ) rn = W N / 2
N /2

X ( k ) = X ( 2r ) =
Y (r) =

N / 2 1 n =0

x(n) + x( N / 2 + n)]W [
y (n)

rn

N / 2 1 n =0

y ( n )WN / 2 rn

Z (r )

k : odd k = 2r + 1 X (k ) = X ( 2r + 1) = = = =
N / 2 1 n =0

[ x(n) x( N / 2 + n)]W
n =0 z (n) 2 rn N

n ( 2 r +1) N

N / 2 1

x( n) x( N / 2 + n)]W W [
n N

2 rn N

N / 2 1 n =0

z (n)W z (n)W

N / 2 1

rn N /2

Z (r ) =

n =0 N / 2 1

z (n)W
n =0

rn N /2

N po int 2

DFT

of

z (0), ,

z(

N 1) 2

X(k) : N-point DFT of x(0), , x(N) two N/2 point DFT

Page 10-14

EE 422G Notes: Chapter 10

Instructor: Zhang

One N/2 point DFT => two N/4 point DFT two point DFTs Consider N/2 point DFT y(0), y(1), , y(N/2-1)
Y (k ) = =
N / 4 1 n =0 N / 2 1 n =0

y ( n )WN / 2 rn

[ y ( n ) + ( 1) k y ( N / 4 + n )]WN / 2 nk

k : even k = 2 r Y (k ) = Y (2r ) = Y 1( r ) =
N / 4 1 n =0 N / 4 1 n =0

[ y( (N n )]WN / 2 nk n )+ y /4 +
y1( n )

y1( n )WN / 4 nk N / 4

po int

DFt

Page 10-15

EE 422G Notes: Chapter 10 k : odd k = 2 r + 1 Y ( k ) = Y ( 2 r + 1) = Y 2( r ) =

N / 4 1 n =0 N / 4 1 n =0

Instructor: Zhang

[ y ( n ) x ( N / 4 + n )]WN / 2 n WN / 2 2 rn
y 2( n )

y 2( n )WN / 4 rn N / 4

po int

DFT

10-4 C Computation
N point DFT : 4N(N-1) real multiplications 4N(N-1) real additions N point FFT : 2Nlog2N real multiplications (N = 2m) 3Nlog2N real additions Computation ration

Page 10-16

EE 422G Notes: Chapter 10

FFT ' s DFT ' s computations 5 log2 N = computations 8( N 1) 5 12 = 0.18% 8 4095

Instructor: Zhang

N =212 =4096

10-5 Properties of the DFT

Assumptions (1)
x(n ) X (k ) and y(n) Y (k ) ( n = 0,..., N 1) ( k = 0,..., N 1)

(2) A, B: arbitrary constants (3) Subscript e: Subscript o:

xe(n) : even about

N1 if N even 2 N if N odd 2

Page 10-17

EE 422G Notes: Chapter 10

Instructor: Zhang

xo(n) : odd about

N1 if N even 2 N if N odd 2

N = 10, xe(n)

N = 9, xe(n)

x(4) = x(5) x(3) = x(6) N 1 = 4.5 x(2) = x(7) 2 x(1) = x(8) x(0) = x(9) x(4) = x(5) x(3) = x(6) N = 4.5 x(2) = x(7) 2 x(1) = x(8) x(0) = x(9)

(4) Any real sequence can be expressed in terms of its even and odd parts according to
x ( n ) = xe ( n ) + xo ( n ) 1 1 = [ x ( n ) + x ( N n )] + [ x ( n ) x ( N n )] 2 2

even odd Question 1: x(n) = 1/2[ ] +1/2 [ ] ? Question 2: x(n) + x(N-n) even ? x(n) - x(N-n) odd ?

Yes!

Example: N = 9 => N/2 = 4.5 Consider n = 2 x(2) + x(9-2) = x(2) + x(7)

Page 10-18

EE 422G Notes: Chapter 10

Instructor: Zhang

is

x(2) + x(7) = x(4.5+(4.5-2)) +x(4.5-(7-4.5))?

4.5 + (4.5-2) = 9-2 = 7 4.5 - (7-4.5) = 9-7 = 2 x(2) + x(7) = x(7) + x(2) ? Yes! => x(n) + x(N-n) even Is x(2) - x(7) = - [x(7) + x(2)] ? Yes! => x(n) - x(N-n) odd (5) subscript r : xr(n) a real sequence subscript i : xi(n) Imaginary part of a complex sequence x ( n ) X ( k ) (6) left right side: side: DFT sequence (7) sequences are assumed periodically repeated if necessary Properties 1. Linearity : 2. Time Shift:
Ax ( n ) + By ( n ) AX ( k ) + BX ( k )

x ( n m) X ( k )e j 2km / N = X ( k )WN k m
x ( n )e j 2km / N X ( k m )
N 1 m =0 N 1 n =0

3. Frequency Shift: 4. Duality : why?

N 1 x ( n) X ( k )

X (k ) =

x (m)e j 2mk / N

DFT ( X ( n )) =

X (n )e j 2nk / N

DFT of x(m)

Page 10-19

EE 422G Notes: Chapter 10

x ( n ) = x ( N n ) = e 1 N
N 1 k =0

Instructor: Zhang

X (k )e j 2k ( N n ) / N
= e j 2kN / N e j 2kn / N 1 N
N 1 k =0

j 2k ( N n ) / N

x ( n ) =

X (k )e j 2kn / N
1 N
N 1 n =0

DFT ( N 1 X ( n )) =

X (n )e j 2nk / N

= x ( k )

5. Circular convolution
N 1 m =0

x(m) y (n m) = x(n)y (n) X (k )Y (k )

N 1 m =0

circular convolution

6. Multiplication

new sequence z ( n) = x( n) y ( n)

x ( n) y ( n) N

X (m)Y (k m) = N 1 X (k )Y (k )

7. Parsevals Theorem

| x(n) |2 = N 1 | X (k ) | 2
n =0 k =0

N 1

N 1

8. Transforms of even real functions:

x er ( n) X er ( k )

(the DFT of an even real sequence is even and real ) 9. Transform of odd real functions:
x or ( n) jX oi (k )

(the DFT of an odd real sequence is odd and imaginary ) 10. z(n) = x(n) + jy(n) z(n) Z(k) = X(k) + jY(k) Example 10-7

z (n) = x( n) + jy (n) = e jn / 2 n = 0,1,2,3

= cos( n / 2) + j sin( n / 2)

Four point DFT for x(0), x(1), x(2), x(3):

Page 10-20

EE 422G Notes: Chapter 10

Instructor: Zhang

X(0) = [x(0) + x(2)] + [x(1) + x(3)] X(1) = [x(0) - x(2)] + (-j)[x(1) - x(3)] X(2) = [x(0) + x(2)] - [x(1) + x(3)] X(3) = [x(0) - x(2)] + j[x(1) - x(3)] For x(n) = cos(n / 2) =>
x(0) = 1 x (1) = 0 x (2) = 1 x (3) = 0

X (0) = 1

X (1) = 2

X (2) = 0

X (3) = 2

For y (n) = sin( n / 2) =>

y (0) = 0 y (1) = 1 y ( 2) = 0 y (3) = 1

Y (0) = 1 Y (1) = j 2
Z ( 0) = 0 Z (1) = 2 j 2 Z ( 2) = 0 Z (3) = 2 + j 2

Y ( 2) = 0

Y (3) = j 2

Example 10-8 DFT of x(n) = (n) :

X (k ) = ( n)W N
n =0 N 1 nk

= 1 k = 0,1,..., N 1

Time-shift property
DFT [ x(n n0 )] = (n n0 )W N
n=0 n k N 1 nk

= WN 0 = e j 2kn0 / N Example 10-9: Circular Convolution

x1 (n) = 1 x2 (n) = 1 0 n N 1

Define x3c (n) = x1 (n)x2 (n) = x1 (m) x2 (n m)

m =0

N 1

Page 10-21

EE 422G Notes: Chapter 10

Instructor: Zhang

X 1 (k ) = X 2 (k ) = =
N1 n= 0

N1 n= 0

x1(n)WN = x2 (n)WN nk
nk n= 0

N1

WN

nk

N = 0

k= 0 k 0 k= 0 k 0

N2 X 3 (k ) = X 1 (k ) X 2 (k ) = 0
x3c ( n) = = 1 N
N 1 k =0 N 1 k =0

1 N

N 1 k =0

x3 (k )e j 2nk / N

N 2 (k )e j 2nk / N

= N e j 2nk / N = Ne j 2nk / N = N x3c ( n) = N

10-6 Applications of FFT

1. Filtering x(0), , x(N-1) FFT (DFT) => X(0), , X(1), , X(N-1) X(k): Line spectrum at k = Inverse DFT:

2kt 2k = T N

(1 =

2 T

t =

T ) N

(Over T: x(0), , x(N-1) are sampled.)

x(n) = X (k )e j 2nk / N
k =0 N0

N 1

x(n) = X (k )e j 2nk / N
k =0

Frequencies with > Example 10-10

2N 0 have been filtered! N

Page 10-22

EE 422G Notes: Chapter 10

x( n) = cos

Instructor: Zhang

n 4

+ cos
2 =(

1 =

2 2 ( ) 8 N

2 4 )2= = N 8 2

n 2

0n7

x(0), x(1), , x(7)

X (0), 0 X (1), non zero X ( 2), non zero 0 X (7 ) 0

How to filter frequency higher than

2. Spectrum Analyzers Analog oscilloscopes => time-domain display Spectrum Analyzers: Data Storage, FFT 3. Energy Spectral Density x(0), , x(N-1): its energy definition
E=
N 1 n =0

| x(n) |2

Parsevals Theorem
E=
N 1

| x ( k ) |2 N k =0

Page 10-23