MIT 6.

02 DRAFT Lecture Notes

(Last update: March 17, 2012)
Comments, questions or bug reports?
Please contact {verghese, hari} at mit.edu

C HAPTER 10
Models for Physical Communication
Channels

To preview what this chapter is about, it will be helpful first to look back briefly at the
territory we have covered. The previous chapters have made the case for a digital (versus
analog) communication paradigm, and have exposed us to communication at the level of
bits or, more generally, at the level of the discrete symbols that encode messages.
We showed, in Chapters 2 and 3, how to obtain compressed or non-redundant repre-
sentations of a discrete set of messages through source coding, which produced codewords
that reflected the inherent information content or entropy of the source. In Chapter 4 we
examined how the source transmitter might map a bit sequence to clocked signals that are
suited for transmission on a physical channel (for example, as voltage levels).
Chapter 5 introduced the binary symmetric channel (BSC) abstraction for bit errors on
a channel, with some associated probability of corrupting an individual bit on passage
through the channel, independently of what happens to every other bit in the sequence.
That chapter, together with Chapters 6, 7, and 8, showed how to re-introduce redundancy,
but in a controlled way using parity bits. This resulted in error-correction codes, or channel
codes, that provide some level of protection against the bit-errors introduced by the BSC.
Chapter 9 considered the challenges of “demapping” back from the received noise-
corrupted signal to the underlying bit stream, assuming that the channel introduced no
deterministic distortion, only additive white noise on the discrete-time samples of the re-
ceived signal. A key idea from Chapter 9 was showing how Gaussian noise experienced
by analog signals led to the BSC bit-flip probability for the discrete version of the channel.
The present chapter begins the process—continued through several subsequent
chapters—of representing, modeling, analyzing, and exploiting the characteristics of the
physical transmission channel. This is the channel seen between the signal transmitted
from the source and the signal captured at the receiver. Referring back to the “single-link
view” in the upper half of Figure 4-8 in Chapter 4, our intent is to study in more detail the
portion of the communication channel represented by the connection between “Mapper +
Xmit samples” at the source side, and “Recv samples + Demapper” at the receiver side.

119
120 Modeling Channel
CHAPTER Behavior
10. MODELS FOR PHYSICAL COMMUNICATION CHANNELS

codeword
bits in generate
101110101 digitized modulate DAC
samples

NOISY & DISTORTING ANALOG CHANNEL

demodulate sample &

ADC & filter threshold
101110101
codeword
bits out

Digitized Samples
Figure 10-1: Elements of a communication channel between the channel coding step at the transmitter and
channel decoding at the receiver.

6.02 Fall 2011 Lecture 10, Slide #2

1 0 0 1 1 1 0 1 0 1

Sample
value

Sample number n
Distorted noise-free signal (discrete-time
y[n] at receiverindex)

Figure 10-2: Digitized samples of the baseband signal.

� 10.1 Getting the Message Across

� 10.1.1 The Baseband Signal
In Figure 10-1 we see an expanded version of what might come between the channel cod-
ing operation at the transmitter and the channel decoding operation at the receiver (as
described in the upper portion of Figure 4-8). At the source, the first stage is to Lecture
6.02 Fall 2011 convert10, Slide #4
the input bit stream to a digitized and discrete-time (DT) signal, represented by samples pro-
duced at a certain sample rate f s samples/s. We denote this signal by x[n], where n is the
integer-valued discrete-time index, ranging in the most general case from −∞ to +∞.
In the simplest case, which we will continue to use for illustration, each bit is repre-
sented by a signal level held for a certain number of samples, for instance a voltage level
of V0 = 0 held for 8 samples to indicate a 0 bit, and a voltage level of V1 = 1 held for 8
samples to indicate a 1 bit, as in Figure 10-2. The sample clock in this example operates
SECTION 10.1. GETTING THE MESSAGE ACROSS 121

at 8 times the rate of the bit clock, so the bit rate is f s /8 bits/s. Such a signal is usually
referred to as the baseband signal.

� 10.1.2 Modulation
The DT baseband signal shown in Figure 10-2 is typically not ready to be transmitted on
the physical transmission channel. For one thing, physical channels typically operate in
continuous-time (CT) analog fashion, so at the very least one needs a digital-to-analog
converter (DAC) to produce a continuous-time signal that can be applied to the channel.
The DAC is usually a simple zero-order hold, which maintains or holds the most recent
sample value for a time interval of 1/ f s . With such a DAC conversion, the DT “rectangular-
wave” in Figure 10-2 becomes a CT rectangular wave, each bit now corresponding to a
signal value that is held for 8/ f s seconds.
Conversion to an analog CT signal will not suffice in general, because the physical
channel is usually not well suited to the transmission of rectangular waves of this sort.
For instance, a speech signal from a microphone may, after appropriate coding for digital
transmission, result in 64 kilobits of data per second (a consequence of sampling the micro-
phone waveform at 8 kHz and 8-bit resolution), but a rectangular wave switching between
two levels at this rate is not adapted to direct radio transmission. The reasons include the
fact that efficient projection of wave energy requires antennas of dimension comparable
with the wavelength of the signal, typically a quarter wavelength in the case of a tower an-
tenna. At 32 kHz, corresponding to the waveform associated with alternating 1’s and 0’s in
the coded microphone output, and with the electromagnetic waves propagating at 3 × 108
meters/s (the speed of light), a quarter-wavelength antenna would be a rather unwieldy
3 × 108 /(4 × 32 × 103 ) = 2344 meters long!
Even if we could arrange for such direct transmission of the baseband signal (after
digital-to-analog conversion), there would be issues related to the required transmitter
power, the attenuation caused by the atmosphere at this frequency, interference between
this transmission and everyone else’s, and so on. Regulatory organizations such as the
U.S. Federal Communications Commission (FCC), and equivalent bodies in other coun-
tries, impose constraints on transmissions, which further restrict what sort of signal can be
applied to a physical channel.
In order to match the baseband signal to the physical and regulatory specifications of a
transmission channel, one typically has to go through a modulation process. This process
converts the digitized samples to a form better suited for transmission on the available
channel. Consider, for example, the case of direct transmission of digital information on
an acoustic channel, from the speaker on your computer to the microphone on your com-
puter (or another computer within “hearing” distance). The speaker does not respond
effectively to the piecewise-constant voltages that arise from our baseband signal. It is in-
stead designed to respond to oscillatory voltages at frequencies in the appropriate range,
producing and projecting a wave of oscillatory acoustic pressure. Excitation by a sinu-
soidal wave produces a pure acoustic tone. With a speaker aperture dimension of about 5
cm (0.05 meters), and a sound speed of around 340 meters/s, we anticipate effective pro-
jection of tones with frequencies in the low kilohertz range, which is indeed in (the high
end of) the audible range.
A simple way to accomplish the desired modulation in the acoustic wave exam-
122 CHAPTER 10. MODELS FOR PHYSICAL COMMUNICATION CHANNELS

ple above is to apply—at the output of the digital-to-analog converter, which feeds the
loudspeaker—a voltage V0 cos(2π f c t) for some duration of time to signal a 0 bit, and a
voltage of the form V1 cos(2π f c t) for the same duration of time to signal a 1 bit.1 Here
cos(2π f c t) is referred to as the carrier signal and f c is the carrier frequency, chosen to be ap-
propriately matched to the channel characteristics. This particular way of imprinting the
baseband signal on a carrier by varying its amplitude is referred to as amplitude modulation
(AM), which we will study in more detail in Chapter 14. The choice V0 = 0 and V1 = 1 is
also referred to as on-off keying, with a burst of pure tone (“on”) signaling a 1 bit, and an
interval of silence (“off”) signaling a 0.
One could also choose V0 = −1 and V1 = +1, which would result in a sinusoidal voltage
that switches phase by π/2 each time the bit stream goes from 0 to 1 or from 1 to 0. This
approach may be referred to as polar keying (particularly when it is thought of as an instance
of amplitude modulation), but is more commonly termed binary phase-shift keying (BPSK).
Yet another modulation possibility for this acoustic example is frequency modulation (FM),
where a tone burst of a particular frequency in the neighborhood of f c is used to signal a 0
bit, and a tone burst at another frequency to signal a 1 bit. All these schemes are applicable
to radio frequency (RF) transmissions as well, not just acoustic transmissions, and are in
fact commonly used in practice for RF communication.

� 10.1.3 Demodulation
We shall have more to say about demodulation later, so for now it suffices to think of it
as a process that is inverse to modulation, aimed at extracting the baseband signal from
the received signal. While part of this process could act directly on the received CT analog
signal, the block diagram in Figure 10-1 shows it all happening in DT, following conversion
of the received signal using an analog-to-digital converter (ADC). The block diagram also
indicates that a filtering step may be involved, to separate the channel noise as much as
possible from the signal component of the received signal, as well as to compensate for
deterministic distortion introduced by the channel. These ideas will be explored further in
later chapters.

� 10.1.4 The Baseband Channel

The result of the demodulation step and any associated filtering is a DT signal y[n], com-
prising samples arriving at the rate f s used for transmission at the source. We assume
issues of clock synchronization are taken care of separately. We also neglect the effects of
any signal attenuation, as this can be simply compensated for at the receiver by choosing
an appropriate amplifier gain.
In the ideal case of no distortion, no noise on the channel, and insignificant propagation
delay, y[n] would exactly equal the modulating baseband signal x[n] used at the source, for
all n. If there is a fixed and known propagation delay on the channel, it can be convenient
to simply set the clock at the receiver that much later than the clock at the sender. If this
is done, then again we would find that in the absence of distortion and random noise, we
get y[n] = x[n] for all n.
1
A zero-order-hold DAC will produce only an approximation of a pure sinusoid, but if the sample rate f s
is sufficiently high, the speaker may not sense the difference.
 Transmission over a Channel
SECTION 10.2. LINEAR TIME-INVARIANT (LTI) MODELS 123

Signal x[n] from digitized samples at transmitter

Example of distorted noise-free signal y[n] at receiver

Figure 10-3: Channel distortion example. The distortion is deterministic.

6.02 Fall 2011 Lecture 10, Slide #9

More realistically, the channel does distort the baseband signal, so the output DT signal
may look (in the noise-free case) as the lower waveform in Figure 10-3. Our objective in
what follows is to develop and analyze an important class of models, namely linear and
time-invariant (LTI) models, that are quite effective in accounting for such distortion, in a
vast variety of settings. The models would be used to represent the end-to-end behavior
of what might be called the baseband channel, whose input is x[n] and output is y[n], as in
Figure 10-3.

� 10.2 Linear Time-Invariant (LTI) Models

� 10.2.1 Baseband Channel Model
Our baseband channel model, as represented in the block diagram in Figure 10-4 takes the
DT sequence or signal x[.] as input and produces the sequence or signal y[.] as output.
We will often use the notation x[.]—or even simply just x—to indicate the entire DT signal
or function. Another way to point to the entire signal, though more cumbersome, is by
referring to “x[n] for −∞ < n < ∞”; this often gets abbreviated to just “the signal x[n]”,
at the risk of being misinterpreted as referring to just the value at a single time n.
Figure 10-4 shows x[n] at the input and y[n] at the output, but that is only to indicate
that this is a snapshot of the system at time n, so indeed we see x[n] at the input and y[n] at
 124 System Input and Output CHAPTER 10. MODELS FOR PHYSICAL COMMUNICATION CHANNELS

input response

x[n] S y[n]

Figure 10-4: Input and output of baseband channel.

A discrete-time signal such as x[n] or y[n] is described by
an infinite sequence of values, i.e., the time index n takes
values in ! to +!. The above picture is a snapshot at a
particular time n.

In the diagram above, the sequence of output values y[.] is

the response of system S to the input sequence x[.]
Figure 10-5: A unit step. In the picture on the left the unit step is unshifted, switching from 0 to 1 at index
Question:
(time) 0. OnWhy
the right,didn’t I iswrite:
the unit step shifted forward in time by 3 units (shifting forward in time means
that we use the − sign in the argument because we want the switch to occur with n − 3 = 0).

“In the diagram above, the sequence of output values y[n] is

the output of the system. What the diagram should not be interpreted as indicating is that
the response
the value of theof system
output signal y[.S
] atto the
time input
n depends sequence
exclusively x[n]”
on the value ??
of the input
signal at that same time n. In general, the value of the output y[.] at time n, namely y[n],
6.02 Fall 2011 Lecture 10, Slide #13
could depend on the values of the input x[.] at all times. We are most often interested in
causal models, however, and those are characterized by y[n] only depending on past and
present values of x[.], i.e., x[k] for k ≤ n.

� 10.2.2 Unit Sample Response h[n] and Unit Step Response s[n]
There are two particular signals that will be very useful in our description and study of
LTI channel models. The unit step signal or function u[n] is defined as

u[n] = 1 if n ≥ 0
u[n] = 0 otherwise (10.1)

It takes the value 0 for negative values of its argument, and 1 everywhere else, as shown
in Figure 10-5. Thus u[1 − n], for example, is 0 for n > 1 and 1 elsewhere.
The unit sample signal or function δ [n], also called the unit impulse function, is defined
as

δ [n] = 1 if n = 0
δ [n] = 0 otherwise. (10.2)

It takes the value 1 when its argument is 0, and 0 everywhere else, as shown in Figure 10-6.
SECTION 10.2. LINEAR TIME-INVARIANT (LTI) MODELS 125

Time Invariant Systems

Let y[.] be the response of S to input x[.]
Figure 10-6: A unit sample. In the picture on the left the unit sample is unshifted, with the spike occurring
If 0.for
at index (time) Onall possible
the right, the unit sequences
sample is shifted x[n] and
backward integers
in time by 5 unitsD(shifting backward
in time means that we use the + sign in the argument because we want the switch to occur with n + 5 = 0).

x[n-D] S y[n-D]

Figure 10-7: Time-invariance: if for all possible sequences x[.] and integers D, the relationship between
then
input and output system
is as S is
shown above, said
then to tobe
S is said time invariant
(TI). A time
be “time-invariant” (TI).
shift in the input sequence to S results in an identical
Thus δ [n −time
3] is 1shift
whereof
n =the output
3 and sequence.
0 everywhere else. One can also deduce easily that

In particular, for δa[n]TI=system, − 1]

u[n] − u[n a ,
shifted unit sample (10.3)
function
where addition and subtraction ofatsignals
![n ! D] the input such asgenerates
u[n] and u[n an
− 1]identically
are defined “point-
wise”, i.e., shifted
by addingunit samplethe
or subtracting response h[ntime
values at each ! D] at the
instant. output.
Similarly, the multipli-
cation of a signal by a scalar constant is defined as pointwise multiplication by this scalar,
so for instance
6.02 Fall 2011 2u[n − 3] has the value 0 for n < 3, and the value 2 everywhere else.10, Slide #20
Lecture
The response y[n] of the system in Figure 10-4 when its input x[n] is the unit sample
signal δ [n] is referred to as the unit sample response, or sometimes the unit impulse
response. We denote the output in this special case by h[n]. Similarly, the response to the
unit step signal u[n] is referred to as the unit step response, and denoted by s[n].
A particularly valuable use of the unit step function, as we shall see, is in representing
a rectangular-wave signal as an alternating sum of delayed (and possibly scaled) unit step
functions. An example is shown in Figure 10-9. We shall return to this decomposition later.

� 10.2.3 Time-Invariance
Consider a DT system with input signal x[.] and output signal y[.], so x[n] and y[n] are
the values seen at the input and output at time n. The system is said to be time-invariant if
shifting the input signal x[.] in time by an arbitrary positive or negative integer D to get
a new input signal x D [n] = x[n − D] produces a corresponding output signal y D [n] that is
just y[n − D], i.e., is the result of simply applying the same shift D to the response y[.] that
was obtained with the original unshifted input signal. The shift corresponds to delaying
the signal by D if D > 0, and advancing it by | D| if D < 0. In particular, for a TI system,
a shifted unit sample function at the input generates an identically shifted unit sample
response at the output. Figure 10-7 illustrates time-invariance.
 the response to x2[.]

If the response to linear combinations of these two

126 inputs equals theCHAPTER
same linear combination of the
10. MODELS FOR PHYSICAL COMMUNICATION CHANNELS
individual responses, then system S is said to be linear.

a1 x1[n]+ a2 x2 [n] S a1 y1[n]+ a2 y2 [n]

If the input is the weighted sum of several signals, the

Figure 10-8: Linearity: if the input is the weighted sum of several signals, the response is the corresponding
response is the corresponding superposition (i.e.,
superposition (i.e., weighted sum) of the response to those signals.
weighted sum) of the response to those signals.

The key6.02to
Fallrecognizing
2011 time-invariance in a given system description isSlide
Lecture 10, to ask
#21 whether

the rules or equations by which the input values x[.] are combined, to create the output
y[n], involve knowing the value of n itself (or something equivalent to knowing n), or just
time differences from the time n. If only the time differences from n are needed, the system is
time-invariant. In this case, the same behavior occurs whether the system is run yesterday
or today, in the following sense: if yesterday’s inputs are applied today instead, then the
output today is what we would have obtained yesterday, just occurring a day later.
Another operational way to recognize time-invariance is to ask whether shifting the
pair of signals x[.] and y[.] by the arbitrary but identical amount D results in new signals
x D [.] and y D [.] that still satisfy the equations defining the system. More generally, a set
of signals that jointly satisfies the equations defining a system, such as x[.] and y[.] in our
input-output example, is referred to as a behavior of the system. And what time-invariance
requires is that time-shifting any behavior of the system by an arbitrary amount D results
in a set of signals that is still a behavior of the system.
Consider a few examples. A system defined by the relation

y[n] = 0.5y[n − 1] + 3x[n] + x[n − 1] for all n (10.4)

is time-invariant, because to construct y[.] at any time instant n, we only need values of
y[.] and x[.] at the same time step and one time step back, no matter what n is — so we
don’t need to know n itself. To see this more concretely, note that the above relation holds
for all n, so we can write

y[n − D] = 0.5y[(n − D) − 1] + 3x[n − D] + x[(n − D) − 1] for all n

or
y D [n] = 0.5y D [n − 1] + 3x D [n] + x D [n − 1] for all n .
In other words, the time-shifted input and output signals, x D [.] and y D [.] respectively, also
satisfy the equation that defines the system.
The system defined by
y[n] = n3 x[n] for all n (10.5)
is not time-invariant, because the value of n is crucial to defining the output at time n. A
little more subtle is the system defined by

y[n] = x[0] + x[n] for all n . (10.6)

This again is not time-invariant, because the signal value at the absolute time 0 is needed,
rather than a signal value that is offset from n by an amount that doesn’t depend on n. We
SECTION 10.2. LINEAR TIME-INVARIANT (LTI) MODELS 127

Unit Step
Decompositio
“Rectangular-wave” digita
signaling waveforms, of th
we have been considering
easily decomposed into ti
shifted, scaled unit steps
transition corresponds to
shifted, scaled unit step).

In this example, x[n] is th

transmission of 1001110
samples/bit: x[n]
= u[n]
! u[n !
+ u[

6.02 Fall 2011 Lecture

Figure 10-9: A rectangular-wave signal can be represented as an alternating sum of delayed (and possibly
scaled) unit step functions. In this example, x[n] = u[n] − u[n − 4] + u[n − 12] − u[n − 24].

have y D [n] = x[0] + x D [n] rather than what would be needed for time-invariance, namely
y D [n] = x D [0] + x D [n].

� 10.2.4 Linearity
Before defining the concept of linearity, it is useful to recall two operations on signals
or time-functions that were defined in connection with Equation (10.3) above, namely (i)
addition of signals, and (ii) scalar multiplication of a signal by a constant. These operations
were defined as pointwise (i.e., occurring at each time-step), and were natural definitions.
(They are completely analogous to vector addition and scalar multiplication of vectors,
the only difference being that instead of the finite array of numbers that we think of for
a vector, we have an infinite array, corresponding to a signal that can take values for all
integer n.)
With these operations in hand, one can talk of weighted linear combinations of signals.
Thus, if x1 [.] and x2 [.] are two possible input signals to a system, for instance the signals
associated with experiments numbered 1 and 2, then we can consider an experiment 3 in
which the input x3 [.] a weighted linear combination of the inputs in the other two experi-
ments:
x3 [n] = a1 x1 [n] + a2 x2 [n] for all n ,
128 CHAPTER 10. MODELS FOR PHYSICAL COMMUNICATION CHANNELS

where a1 and a2 are scalar constants.

The system is termed linear if the response to this weighted linear combination of the
two signals is the same weighted combination of the responses to the two signals, for all possible
choices of x1 [.], x2 [.] a1 and a2 , i.e., if

y3 [n] = a1 y1 [n] + a2 y2 [n] for all n ,

where yi [.] denotes the response of the system to input xi [.] for i = 1, 2, 3.
This relationship is shown in Figure 10-8. If this property holds, we say that the results
of any two experiments can be superposed to yield the results of other experiments; a linear
system is said to have the superposition property. (In terms of the notion of behaviors
defined earlier, what linearity requires is that weighted linear combinations, or superposi-
tions, of behaviors are again behaviors of the system.)
We can revisit the examples introduced in Equations (10.4), (10.5), (10.6) to apply this
definition, and recognize that all three systems are linear. The following are examples of
systems that are not linear:
y[n] = x[n] + 3 ;
y[n] = x[n] + x2 [n − 1] ;
� x2 [n] �
y[n] = cos 2 .
x [n] + 1
All three examples here are time-invariant.

� 10.2.5 Linear, Time-Invariant (LTI) Models

Models that are both linear and time-invariant, or LTI models, are hugely important in en-
gineering and other domains. We will mention some of the reasons in the next chapter. We
will develop insights into their behavior and tools for their analysis, and then return to ap-
ply what we have learned, to better understand signal transmission on physical channels.
In the context of audio communication using a computer’s speaker and microphone,
transmissions are done using bursts at the loudspeaker of a computer, and receptions by
detecting the response at a microphone. The input x[n] in this case is a baseband signal
of the form in Figure 10-3, but alternating regularly between high and low values. This
was converted through a modulation process into the tone bursts that you heard. The
signal received at the microphone is then demodulated to reconstruct an estimate y[n] of
the baseband input.
With the microphone in a fixed position, responses have some consistency from one
transition to the next (between tone and no-tone), despite the presence of random fluctua-
tions riding on top of things. The deterministic or repeatable part of the response y[n] does
show distortion, i.e., deviation from x[n], though more “real-world” than what is shown
in the synthetic example in Figure 10-3. However, when the microphone is very close to
the speaker, the distortion is low.
There were features of the system response in this communication system to suggest
that it may not be unreasonable to model the baseband acoustic channel as LTI. Time-
invariance (at least over the time-horizon of the demo!) is suggested by the repeatability
of the transient responses to the various transitions. Linearity is suggested by the fact that
SECTION 10.2. LINEAR TIME-INVARIANT (LTI) MODELS 129

the downward transients caused by negative (i.e., downward) steps at the input look like
reflections of the upward transients caused by positive (i.e., upward) steps of the same
magnitude at the input, and is also suggested by the appropriate scaling of the response
when the input is scaled.