6.

02 Fall 2012

Lecture #14

Spectral content via the DTFT

6.02 Fall 2012 Lecture 14 Slide #1

 Demo: Deconvolving Output of

Channel with Echo

x[n] y[n] z[n]

Channel, Receiver
h1[.] filter, h2[.]

Suppose channel is LTI with

h 1[n]=[n]+0.8[n-1]

H1() = ?? = 1
jm
h [m]e
m
= 1+ 0.8ej = 1 + 0.8cos() j0.8sin()
So:
|H1()| = [1.64 + 1.6cos()]1/2 EVEN function of ;

<H1() = arctan [(0.8sin()/[1 + 0.8cos()] ODD .

6.02 Fall 2012 Lecture 14 Slide #2
A Frequency-Domain view of Deconvolution

x[n] y[n] z[n]

Channel, Receiver
H 1() filter, H2()

Noise w[n]
Given H1(), what should H2() be, to get z[n]=x[n]?

H2()=1/H1() Inverse filter

= (1/|H1()|). exp{j<H1()}

Inverse filter at receiver does very badly in the presence of noise

that adds to y[n]:

filter has high gain for noise precisely at frequencies where
channel gain|H 1()| is low (and channel output is weak)!
6.02 Fall 2012 Lecture 14 Slide #3
 DT Fourier Transform (DTFT) for

Spectral Representation of General x[n]

If we can write

1
h[n] =
2
H ()e jn d where H () = h[m]e jm
<2 > Any contiguous
m
interval of length

then we can write 2

1
x[n] =
2
X()e jn
d where X() = x[m]e jm
<2 > m

This Fourier representation expresses x[n] as

a weighted combination of e jn for all in [,].

X()d is the spectral content of x[n]

6.02 Fall 2012 in the frequency interval [, + d ] Lecture 14 Slide #4
The spectrum of the exponential signal (0.5)nu[n] is shown over the

frequency range = 2f in [-4,4], The angle has units of degrees.

Courtesy of Don Johnson. Used with permission; available under a CC-BY license.

6.02 Fall 2012

http://cnx.org/content/m0524/latest/ Lecture 14 Slide #5

 )

x[n] and X(

6.02 Fall 2012 Lecture 14 Slide #6

 Input/Output Behavior of

LTI System in Frequency Domain

1 1
x[n] =
2
X()e jn
d y[n] =
2
H ()X()e jn d
<2 > <2 >
H()
1
y[n] =
2
Y ()e jn d
<2 >

Y () = H ()X()
Compare with y[n]=(h*x)[n]

Again, convolution in time

has mapped to
multiplication in frequency
6.02 Fall 2012 Lecture 14 Slide #7
 Magnitude and Angle

Y () = H ()X()

| Y () |= |H () |. | X() |
and

< Y () = < H ()+ < X()

6.02 Fall 2012 Lecture 14 Slide #8
 Core of the Story
1. A huge class of DT and CT signals

can be written --- using Fourier transforms --- as a

weighted sums of sinusoids (ranging from very slow to very fast)

or (equivalently, but more compactly) complex exponentials.

The sums can be discrete or continuous (or both).

2. LTI systems act very simply on sums of sinusoids:

superposition of responses to each sinusoid, with the

frequency response determining the frequency-dependent

scaling of magnitude, shifting in phase.

6.02 Fall 2012 Lecture 14 Slide #9

 Loudspeaker Bandpass Frequency Response

SPL versus Frequency

(Speaker Sensitivity = 85dB)

100

97

94

91

88
SPL (dB)

85

82

79

76 -3dB @ 56.5Hz -3dB @ 12.5k Hz

73

70
10 100 1,000 10,000 100,000

Frequency (Hz)

Image by MIT OpenCourseWare.

F ll 2012
6.02 Fall Lecture 14 Slide #10
 PC Magazine. All rights reserved. This content is excluded from our Creative
Commons license. For more information, see http://ocw.mit.edu/fairuse.
6.02 Fall 2012 Lecture 14 Slide #11
 Spectral Content of Various Sounds

Human Voice
Cymbal Crash
Snare Drum
Bass Drum
Guitar
Bass Guitar
Synthesizer
Piano

13.75 Hz- 27.5 Hz- 55 Hz- 110 Hz- 220 Hz- 440 Hz- 880 Hz- 1,760 Hz- 3,520 Hz- 7,040 Hz- 14,080 Hz-
27.5 Hz 55 Hz 110 Hz 220 Hz 440 Hz 880 Hz 1,760 Hz 3,520 Hz 7,040 Hz 14,080 Hz 28,160 Hz

Image by MIT OpenCourseWare.

6.02 Fall 2012 Lecture 14 Slide #12

 Connection between CT and DT

The continuous-time (CT) signal

x(t) = cos( t) = cos(2ft)

sampled every T seconds, i.e., at a sampling

frequency of fs = 1/T, gives rise to the discrete-time
(DT) signal

x[n] = x(nT) = cos(nT) = cos(n)

So = 

and = corresponds to = /T or f = 1/(2T) = fs/2

6.02 Fall 2012 Lecture 14 Slide #13
 Signal x[n] that has its frequency content

uniformly distributed in [c , c]

1
x[n] =
2
X()e jn
d
<2 >

C
1
=
2
1 e jn d
C

sin(C n)
= , n0
n DT sinc function
= C / , n=0 (extends to in time, 
falls off only as 1/n)
6.02 Fall 2012 Lecture 14 Slide #14
 )

x[n] and X(

6.02 Fall 2012 Lecture 14 Slide #15

 X() and x[n]

6.02 Fall 2012 Lecture 14 Slide #16

 Fast Fourier Transform (FFT) to compute

samples of the DTFT for

signals of finite duration

P1 (P/2)1
1
X( k ) = x[m]e
jk m
, x[n] = X( k )e jk n

m=0 P k=P/2
where k = k(2/P), P is some integer (preferably a power of 2)
such that P is longer than the time interval [0,L-1] over which
x[n] is nonzero, and k ranges from P/2 to (P/2)1 (for even P).

Computing these series involves O(P2) operations when P gets

large, the computations get very s l o w.

Happily, in 1965 Cooley and Tukey published a fast method for

computing the Fourier transform (aka FFT, IFFT), rediscovering
a technique known to Gauss. This method takes O(P log P)
operations.
6.02 Fall 2012
P = 1024, P2 = 1,048,576, P logP 10,240
Lecture 14 Slide #17
 Where do the k live?

e.g., for P=6 (even)

3
2
1 0 1 2  3

 0 

exp(j2) j exp(j1)

exp(j3) 1 . exp(j0)
1
= exp(j
3)

exp(j
2)
j exp(j
1)

6.02 Fall 2012 Lecture 14 Slide #18

 Spectrum of Digital Transmissions

(scaled version of DTFT samples)

6.02 Fall 2012 Lecture 14 Slide #19

 Spectrum of Digital Transmissions

6.02 Fall 2012 Lecture 14 Slide #20

 Observations on previous figure

The waveform x[n] cannot vary faster than the step change every 7
samples, so we expect the highest frequency components in the
waveform to have a period around 14 samples. (The is rough and
qualitative, as x[n] is not sinusoidal.)

A period of 14 corresponds to a frequency of 2/14 = /7, which

is 1/7 of the way from 0 to the positive end of the frequency axis
at  (so k approximately 100/7 or 14 in the figure). And that
indeed is the neighborhood of where the Fourier coefficients drop
off significantly in magnitude.

There are also lower-frequency components corresponding to the

fact that the 1 or 0 level may be held for several bit slots.

And there are higher-frequency components that result from the

transitions between voltage levels being sudden, not gradual.

6.02 Fall 2012 Lecture 14 Slide #21

 Effect of Low-Pass Channel

6.02
6 02 Fall 2012 Lecture 14 Slide #22

 How Low Can We Go?

7 samples/bit  14 samples/period  k=(N/14)=(196/14)=14

6.02 Fall 2012 Lecture 14 Slide #23

MIT OpenCourseWare
http://ocw.mit.edu

6.02 Introduction to EECS II: Digital Communication Systems

Fall 2012

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.