f

BASEBAND PULSETRANSMISSION ~
448

of the pulse-amplitude modulator and the decisionomaking device in an M-ary

PAM are more complex than those in a binary PAM system. Intersymbol inter-
t ference, noise, and imperfect synchronization cause errors to appear at the re-
ceive output. The transmit and receive filters are designed to minimize these
errors. Procedures used for the design of these filters are similar to those dis-
cussed in Sections 7.4 and 7.5 for baseband binary PAM systems.

7.8 . TAPPED-DELAY-LINE EQUALIZATION

A communication channel that readily comes to mind for the transmission of

digital data (e.g., computer data) is a telephone channel, which is characterized by
a high signal-to-noise ratio. However, the telephone channel is bandwidth-limited,
as illustrated in Fig. 7.21 for a typical toll connection. Figure 7.21a shows the
insertion loss of the channel plotted versus frequency; insertion loss (in dB) is
def!ned as 10 10giO (Pol P2), where P2 is the power delivered to a load by the
channel, and Po is the power delivered to the same load when it is connected
directly to the source (i.e., the channel is removed). Figure 7.21 b shows the
corresponding plots of the phase response and envelope (group) delay versus
frequency; for the definition of envelope delay, see Section' 2.14. Figure 7.21
']). . tol clearly illustrates the dispersive nature of the telephone channel. ~ efficient
-\-t, 10'" approach to high-speed transmission of digital data over such a channel uses a

~~-:t\~'hcw
1..t
combination
~iscrete
of two basic signal-processIng
PAM, which involves encoding
operatlOps:
the amplitudes of successive pulses in
a periodic pulse train with a discrete set of possible amplitude levels.
/A linear modulation scheme, which offers bandwidth conservation to transmit
the encoded pulse tra~er the telephone channel.

At the receiving end of the system, the received signal is demodulated and syn-
chronously sampled, and then decisions are made as to which particular symbols
were transmitted. As a result of dispersion of the pulse shape by the telephone
channel, we find that the number of detectable amplitude levels is often Ii .
by intersymbol interference rat er t an by additive noise. In principle, if the
channel is known precisely, it is virtually always possible to make the intersymbol
interference at the sampling instants arbitrarily small by using a suitable pair of
transmit and receive filters, so as to control the overall pulse shape in the manner

~ ~ ~ described previously. The transmit filter is placed directly before the modulator,
.1.~X 11 ( whereas the receive filter is placed dIrectly after the demodulator. Thus, insofar
~ k "",,,, ek ".,...J as intersymbol interference is concerned, we may consider the data transmission
~\ U",k."-,,, Ch"'.'-L\over the telephone channel as being baseband.
~Q..C-k.,.;.'-\.;l~ In practice, however, we seldom have prior knowledge of the exact channel
Ov;t"ftkl>~"" characteristics. Also, there is the unavoidable problem of imprecision that arises
in the physical implementation of the transmit and receive filters. The net result
of all these effects is that there will be some residual distortion for ISI to be a
limiting factor on the data rate of the system. To compensate for the intrinsic
residual distortion, ~e may use a process known as equalization. The filter used
to perform such a process is called an equalizer,
A device well-suited for the design of a linear equalizer is the J:.a~E.e2-<:!~L
line filter, as depicted in Fig. 7.22. For symmetry, the total nunfb'(;rv of taps IS
('./\/V"Vvv

tII
7.8 TAPPED-DELAY-LINE EQUALIZATION 449

20

15

cc
:3-
'"'"
0
c: 10
.g
Q;
'"
E

2 3 4 5
Frequency (kHz)

(a)

4 10

I
..-'"
Envelope
..-..-
3 delay
,..-"-\. en
Ideal "
I>
T
""13
>
linear
'"
Q; phase .t:
0
"Q.
"0
2 :.c:
'"
0 "'"
Q;
>
'"
.J::
c
w c..

-5

10
2 3 4 5

Frequency (kHz)

(b)

Figure 7.21 (a) Amplitude response of typical toll connec-

tion. (b) Envelope delay and phase response of typical toll
connection. (Bellamy, 1982.)

chosen to be (2N + 1), with the weights denoted by W-N>"" W_-!, WO,W!,...,
WN' The impulse response of the tapped-delay-line equalizer is therefore
N
k(t) = 2: Wh 8(t - kT) (7.85)
h= -N

where 8(t) is the Dirac delta function, and the delay T is chosen equal to the
symbol duration.
Suppose that the tapped-delay-line-equalizer is connected in cascade with a
linear system whose impulse response is c(t), as depicted in Fig. 7.23. Let pet)

III
 ...:
0)
~
;;::
0)
c
>.
.!!!
0)
"0
h' .:,
H " 0)
';;::, 0.
0.
nJ
I-
N
"'!
r0-
O)
...
::I
CI
u::

450

110
 ...

,
7.8 TAPPED-DELAY-LiNE EQUALIZATION 451

Linear T apped-delay-I ine

system, equalizer,
c(l) h(t)

Impulse response, p(t)

Figure 7.23 Cascade connection of linear

system and tapped-delay-line equalizer.

denote the impulse response of the equalized system. Then P(t) is equal to the
convolution of c(t) and h(t), as shown by

P(t) = c(t) * h(t)

N
= c(t) * k=L-N Wk 8 (t - kT)

Interchanging the order of summation and convolution:

N
P(t) = k= -N
L WkC(t) * 8(t - kT)
N . (7.86)

k= -N
L WkC(t - kT)

where we have made use of the sifting property of the delta function. Evaluating
Eq. (7.86) at the sampling times t = nT, we get the discrete convolution sum
N
p(nT) = L
k= -N
wk c((n - k)T) (7.87)

Note that the sequence {p(nT)} is longer than {c(nT)}.

To eliminate intersymbol interference completely, we must satisfy the
Nyquist criterion for distortionless transmission described in Eq. (7.49), with T
used in place of Tb' It is assumed that the P( t) is defined in such a way that the
normalized condition P(O) = 1 is satisfied in accordance with Eq. (7.46). Thus,
for no intersymbol interference we require that

I, n=
p(nT) =
{ 0, n=/=O

But from Eq. (7.87) we note that there are only (2N + I) adjustable coefficients
at our disposal. Hence, this ideal condition can only be satisfied approximately
as follows:

I, n =
p(nT) = (7.88)
{., n = ::tl, :t2,'..., :tN I
I
I
L
 -,-- = ..,,- ..

452 BASEBAND PULSE TRANSMISSION

To simplify the notation, we let the nth sample of the impulse response c(t) be
wri tten as

Cn = c(nT) (7.89)

Then, imposing the condition of Eq. (7.88) on the discrete convolution sum of
Eq. (7.87), we obtain a set of (2N + 1) simultaneous equations:

I, n = 0
f
k= -N Wk Cn-k = { 0, n = ::!:1, j:2, . . ., ::!:N
(7.90)

Equivalently, in matrix form we may write

(Q ... C-N+I C-N C-N-I . .. C-2N W-N 0

CN-I . .. (Q C-l C-2 ... C-N-I W-I 0

CN
... CI (Q C-I . .. C-N Wo
= 1 I (7.91)

CN+I . .. l".z Cl (Q ... C-N+I WI 0

... 0
G.2N ... CN+l CN CN-I (Q WN

A tapped-delay-line equalizer described by Eq. (7.90) or, equivalently Eq. <7.9.1),

is referred to as a zf;:j1:.!j!!}hl1r.!lJ!::~ Such an equalizer is optimum in the sense
that it minimizes the peak distortion (intersymbol interference). It also has the
nice feature of being relatively simple to implement. In theory, the longer we
make the equalizer (i.e., permit N to approach infinity), the more closely will
the equalized system approach the ideal condition specified by the Nyquist cri-
terion for distortlonless transmission.

7.9 ADAPTIVE EQUALIZATION

The zero-forcing strategy described abrnce works well in the 1aboratory, where we
have access to the system to be equalized, in which case we know the system
coefficients L N' . . . , L l' (Q, Cl' . . . , CN that are needed for the solution of Eq.
(7.91). In a telecommunications environment, however, the channel is usually
time varying. For example, in a switched telephone network, we find that two
factors contribute to the distribution of pulse distortion on different link
/.I connections:
~ ~'lDifferences in the transmission characteristics of the individual links that ,may
be switched together.
.../ Differences in the number of links in a connection.

The result is that the telephone channel is random in the sense of being one of
van ensemble of possible physical realizations. Consequently, the use of a fixed
equalizer designed on the basis of average channel characteristics may not ade-
 ... Jm !IF.!: - .

7.9 ADAPTIVE EQUALIZATION 453

quately reduce intersyrnbol interference. To realize the full transmission capa-

bility of a telephone channel, there is need for ~n. 5 The process ~L
of equalization is said to be adaptivp when tlie equalizer adjusts itself continu- =r--
<?usly and automatically by operating on the input signal.
Among the philosophies for adaptive equalization of data transmission sys-
tems, we have prechannel equalization at the transmitter and postchannel equalization
at the receiver. Because the first approach requires a feedback channel, we con-
sider only adaptive equalization at the receiving end of the system. This~i- ~ ""
zation can be achieved, prior to data transmission, by training the filter with the R 7w:.~
guidance of a suitable ~ce transmitted through the channel so as to r
adjust the filter parameters to E!imum values. The typical telephone channel
changes little during an average data call, so that precall equalization with a
training sequence is sufficient in most cases encountered in practice. As men-
tioned previously, the equalizer is positioned after the receive filter in the
receiver.
In this section we study an adaptive equalizer based on the tapped-delay-line
filter, which is synchronous in the sense that the tap spacing of the equalizer is
the same as the symbol duration T of the transmitted signal (i.e., the reciprocal
of the signaling rate). This equalizer is not only simple to implement but is also
capable of realizing a satisfactory performance.

Least-Mean-Square Algorithm
Consider a tapped-delay-line equalizer, whose tap-weights are adjustable as in-
dicated in Fig. 7.24. The input sequence Ix( nT) I applied to this equalizer is
produced by the transmission of a binary sequence through an unknown channel
that is both dispersive and noisy. I.t is assumed that some form of pulse shaping
is included in the design of the transmission system. The requirement is to co~-
rect for the combined effec!S- oLresidual dis.!Qrtion and noise in the system
t.hrough the use of an ada2tive equalizer. - it, oft)
To simplifY notational matters, we let C\tYIS~''''c\;CM

Xn = x(nT} (7.92)
Yn = y(nT} (7.93)

Then, the output Yn of the tapped-delay-line equalizer in response to the input

sequence IXnl is defined by the discrete convolution sum (see Fig. 7.24)
N
Yn = 2: WkXn-k (7.94)
k= -N

where Wk is the weight at the kth tap, a!1d 2N + 1 is the t~t~l nu_mber of t!lps.
The tap-weights constitute the adaptive filter coefficients. We assume that the
input sequence IXnl pas finite ene!lO'.
The adaptation may be achieved by observing the error between the desired
pulse shape and the actual pulse shape at the filter output, measured at the
sampling instants, and then using this error to estimate the direction in which
the tap-weights of the filter should be changed so as to approach an optimum
set of values. Eor the adaptation, '!!...e I1!..a'y~e a pea!!:...!!isjortion criterion that I!!ini-
ipizes the peak distortion, defined as the worst-case intersymbol interference at

II
 IIIJ!I!!I

7.9 ADAPTIVE EQUALIZATION 455

.the output of the equalize!:;. The development of an adaptive equalizer using

(q..",.;~
such a cnterion builds on the zero-forcing cp(lcept described in the preceding
"",,-
~~~t section. However, the equalizer is optimum only when the peak distortion at its II
input is less than 100 percent (i.e., the intersymbol interference is not too se-
vere). ~~~~!)S ~e may use a mean-square error criterion, which is more
generafin application; also an adaptive equalizer based on the mean-square error
criterion appears to be less sensitive to timing perturbationsJhan one based on
th.e peak distortion criterion. Accordingly, in what follows we will use the mean-
square error criterion for the development of the adaptive equalizer.
Let an denote the desired response defined as the polar representation of the
nth transmitted binary symbol. Let en denote the error signal defined as the dif-
ference between the desired response an and the actual response Yn of the equal-
Izer, as shown by

en = an - Yn (7.95)

Then, a criterion commonly used in practice (because of its mathematical tract-

ability) is the mean-square error, defined by the cost function

M~ 'i8 = E[e;] (7.96)

(J,\\Qsv\1ffl
~
where E is the statistical expectation operator. Using Eqs. (7.94) to (7.96), the
gradient of the mean-square error 'i8with respect to the kth tap-weight Wkmay
be expressed as

a'i8 aen
aWk 2 E [ enaw~]
aYn
(7.97)
= -2 E [ enaWk]
= - 2 E[enxn-k]

The expectation on the right-hand side of Eq. (7.97) is the ensemble-averaged

cross-correlation between the error signal en and the input signal Xn for a lag of
k samples; that is,

. Rex(k) = E[enxn-k] (7.98)

We may thus simplifY Eq. (7.97) to

I
I
a'i8
- 2 Rex(k) (7.99)
aWk

The optimality condition for minimum mean-square error may now be. ex-
presse.d simply as

./ a'i8= 0 for k = 0, ::!::1,..., ::!::N (7.100)

aWk

.
 ~ , ~ .
- . ;- -. - - .. '. 0lIl' '-.'. .:;;;;;;;jiI -

456 BASEBAND PULSE TRANSMISSION

In light ofEq. (7.99), this condition is equivaJent to the requirement that

Rex(k) = for k = 0, :t 1,..., -tN (7.101)

That is, for minimum mean-square error, the cross-correlation between the output error
sequence{enI and the znput sequence{xnl must have zerosfor the (2N +1) c~mpo;'ents
with integer lags corresponding -to the Tndix values of the available tap-weights of thejilter.
This important result is known as the principle of orthogonality.
Substituting Eqs. (7.94) and (7.95) in (7.96) and expanding terms, we find .1

that the mean-square error <&is precisely a second-order function of the tap-
weights W-N' . . . , Ui-I' WO,WI'. . . , WN'The mean-sq\,lareerror performanceof
the equalizer may therefore be visualized as a multidimensional bowl-shared
surface that is a parabolic function of the tap-weights. The adaptive process,
through successive adjustments of the tap-weights, has the task of continually
seeking t.he bottom of the bowl; at this unique point, the mean-square error <&attains
its mtnimum value <&min'It is therefore intuitively reasonable that successive
adjustments to the tap-weights be in the direction of steepest descent of the
error surface (i.e., in a direction opposite to the vector of gradients aW,/aWk'
- N:so k .:5N), which should lead to the minimum mean-square error <&min' This
.
is the basic idea of the steepestdescentalgoritlJ!!t,described by the recursive fortllilla
~

1 a<&
k = 0, :t "1,. . ., :t N (7.102)
wk(n + 1) = wk(n) - '2J-LaWk'

I where J-Lis a small positive constant called the step-size parameter, and the factor
" 1/2 has been introduced to cancel the factor 2 in the defining equation for
a'f: faWk' The index n is the iteration number. Thus the use of Eq. (7.99) in
(7.102) yields

Wk(n + 1) = wk(n) + J-LRex(k), k = O,:t 1,...,:tN (7.103)

The use of the steepest-descent algorithm requires knowledge of the cross-

correlation function Rex (k). However, this knowledge is not available when oper-
ating in an unknown environment. We may overcome this difficulty by using an
instantaneous estimate for the cross-correlation function Rex (k). Specifically, on
the basis of the defining equation (7.98), we may use the following estimate:

Rex (k) = enxn-, k' k = O,:t 1,...,:tN (7.104)

In a corresponding fashion, we use the estimate Wk(n) in place of the tap-weight

wk(n). Naturally, the use of these estimates in Eq. (7.103) results in an approxi-
mation to the steepest-descent algorithm. We may express the new recursive for-
mula for updating the tap-weights of the equalizer as follows:

Wk(n + 1) = wk(n) + J-Lenxn-k' k = 0, :t 1, , . . , :t N (7.105)

This algorithm is known as the least-mean-square (LMS) algorithm.6 Viewing n as

index for the previous IteratIon, wk(n) is the "old value" ot the kth tap-weight.
and J-Lenxn-k is the "correction" applied to it to compute the "updated value"
uJk(n + 1).
 -- - t
II

f
7.9 ADAPTIVE EQUALIZATION 457

The LMS algorithm is an example of a feedback system, as illustrated in the

block diagram of Fig. 7.25. It is therefore possible for the algorithm to diverge
(i.e., for the adaptive equalizer to become unstable). Unfortunately, the conver-
gence behavior of the LMS algorithm is difficult to analyze. Nevertheless,
e- pro-
vided that the step-size parameter JLis assigned a small value, we find that after
alarge number of iterations the benavlOr ot the LMS algorithm is roughly similar
to that of the steepest-descent algorithm, which uses the actual gradient rather
than a noisy estimate for the computation of the tap-weights.
We may simplify the formulation of the LMS algorithm using matrix nota-
tion. Let the (2N + 1)-by-l vector Xn denote the tap-inputs of the equalizer:

Xn = [Xn+N"'" xn+I' Xn, Xn-I"'" Xn-N]T (7.106)

where the superscript T denotes matrix transposition. Correspondingly, let the

(2N + 1)-by-l vector wn denote the tap-weights of the equalizer:

Wn = [w_~n),..., w-I(n), wo(n), wI(n),..., wN(n)]T (7.107)

We may then use matrix"notation to recast the convolution sum ofEq. (7.94) in
the compact form
- T.A
Yn - XnWn (7.108)

where x~ wn is referred to as the inner product of the vectors xn and wn. We may
now summarize the LMS algorithm as follows:
1. Initialize the algorithm by setting WI = 0 (i.e., set all the tap-weights of the
equalizer to zero at n = 1, which corresponds to time t = T).
2. For n = 1, 2, . . . , compute
- T.A
Yn - xnwn

en = an - Yn

Wn+1 = wn + JLe~n
where JL is the step-size parameter.
3. Continue the computation until steady-state conditions are reached.

Correction
lIen Xn - k

+
Old value +
Updated value
(;,k(n)
L (;,k(n + 1)

Unit delay
T

Figure7.25 Signal-flow graph representa-

tion of the LMSalgorithm.
 so iiiI
JI
458 BASEBAND PULSE TRANSMISSION

Operation of the Equalizer

There are two modes of operation for an adaptive equalizer, namely, the training
mode and decision-directed mode, as shown in Fig. 7.26. During the training
mode, a known sequence is transmitted and a synchronized version of:this signal
is generated in the receiver, where (after a time shift equal to the transmission
delay) it is applied to the adaptive equalizer as the desired response; the tap-
weights of the equalizer are thereby adjusted in accordance with the LMS algo-
rithm. A training sequence commonly used in practice is the so-called pS!3ldo-
noise (PN) sequence, which consists of a deterministic sequence with noiselike
~haractenstlcs; a tull discussion of this sequence is presented in Chapter 9.
When the training process is completed, the adaptive equalizer is switched
to its second mode of operation: the decision-directedmode. In this mode of op-
eration, the error signal is defined by ~

en = an - Yn (7.109)

where Yn is the equalizer output at time t = nT, and an is the final (not neces-
sarily) correct estimate of the transmitted symbol an, Now, in normal oper.ation
the decisions m;tde by the receiver are correct with high probability. This means
that the .error estimates are correct most of the time, thereby permitting the
adaptive equalizer to operate satisfactorily. Furthermore, an adaptive equalizer
operating in a decision-directed rr..ode is able to track relatively slow variations in
channel characteristics.
It turns out that the larger the step-size parameter JL, the faster the tracking
capability of the adaptive equalizer. However, a large step-size parameter f.Lmay
result in an unacceptably high excessmean-square error,defined as that part ofthe
.~ mean-square value of the error signal in excess of the minimum attainable value

~t ~min (which results when the tap-weights are at their optimum settings). We
therefore find that in practice the choice of a suitable value for the step-size
parameter JLinvolves making a compromise between fast tracking and reducing
the excess mean-square error.

Implementation Approaches

An important advantage of the LMS algorithm is that it is simple to implement.

The methods of implementing adaptive equalizers may be divided into three

Adaptive Decision
equalizer device
A
an
I
S
an Training
Xn
{wk) 1 ,
Yn
)1 1 seq uence
generator

en
L +

Figure 7.26 Illustrating the two modes of operation of an adaptive equalizer.

r
I 7.9 ADAPTIVE EQUALIZATION 459

broad categories: analog, hardwired digital, and programmabledigital, as described

here: .
1. The analog approach is primarily based on the use of charge-coupleddevice ,
I

(CCD) technology. The basic circuit realization 'of the CCD is a row of field- f
i
effect transistors with drains and sources connected in series, and the drains
capacitively coupled to g<,\tes.The set of adjustable tap-weights are stored in
digital memory locations, and the multiplications of the analog sample values
by the digitized tap-weights take place in analog fashion. This approach has 1!1,

significant potential in applications where the symbol rate is too high for
digital implementation.
2. In hardwired digital implementation of an adaptive equalizer, the equalizer
input is first sampled and then quantized into a form suitable for storage in
shift registers. The set of adjustable tap-weights is also stored in shift registers.
Logic circuits are used to perform the required digital arithmetic (e.g., mul-
tiply and accumulate). In this second approach, the circuitry is hardwired
for the sole purpose of performing equalization. It is the most widely used ~11

method of building adaptive equalizers and lends itself to implementation

in very-large-scale integrated (VLSI) circuit form.
3. The use of a programmable digital processor in the form of a microprocessor,
for example, offers flexibility in that the adaptive equalization is performed
as a series of steps or instructions in the microprocessor. An important ad-
vantage of this approach is that the same hardware may be time shared to
perform a multiplicity of signal-processing functions such as filtering, mod-
ulation, and demodulation in a modem (modulator-demodulator) used to
transmit digital data over a telephone channel.

Decision-Feedback Equalization

To develop further insight into adaptive equalization, consider a baseband chan-

nel with impulse response denoted in its sampled form by the sequence {hnl
where hn = h(nT). The response of this channel to an input sequence {xnl, in
the absence of noise, is given by the discrete convolution sum

Yn = 2:k hk Xn-k
(7.11 0)
- !toxn + k<O
2: hkxn- k + k>O
2: hkxn- k

The first term of Eq. (7.110) represents the desired data symbol. The second
term is due to the precursors of the channel impulse response that occur before
the main sample !to associated with the desired data symbol. The third term is
due to the postcursors of the channel impulse response that occur after the main
sample !to. The 'precursors and postcursors of a channel impulse response are
illustrated in Fig. 7.27. The idea of decisionjeedback equalization7 is to use data
decisions made on the basis of precursors of the channel impulse response to
take care of the postcursors; for the idea to work, however, the decisi9ns would
obviously have to be correct. Provided that this condition is satisfier., a decision-
~
.; -" -
460 BASEBAND PULSETRANSMISSION

hO

Precursors Postcursors

Figure 7.27 Impulse response of a discrete channel.

feedback equalizer is able to provide an improvement over the performance of

the tapped-delay-line equalizer.
A decisionjeedback equalizer consists of a feedforward section, a feedback sec-
tion, and a decision device connected together as shown in Fig. 7.28. The feed-
forward section consists of a tapped-delay-line filter whose taps are spaced at the
reciprocal of the signaling rate. The data sequence to be equalized is applied to
this section. The feedback section consists of another tapped-delay-line filter
whose taps are also spaced at the reciprocal of the signaling rate. The input
applied to the feedback section consists of the decisions made on previously
detected symbols of the input sequence. The function of the feedback section is
to subtract out that portion of the intersymbol interference produced by previ-
ously detected symbols from the estimates of future samples.
Note that the inclusion of the decision device in the feedback loop makes
the equalizer intrinsically nonlinear and therefore more difficult to analyze than
an ordinary tapped-delay-line equalizer. Nevertheless, the mean-square error cri-
terion can be used to obtain a mathematically tractable optimization of a deci-
sion-feedback equalizer. Indeed, the LMS algorithm can be used to jointly adapt
both the feedforward tap-weights and the feedback tap-weights based on a com-
mon error signal. To be specific, let the augmented vector cn denote the com-
I bination of the feedforward and feedback tap-weights, as shown by

w(l)
Cn -- n (7.111)
[ w(2)n ]

A
I
Xn Feedforward Decision an
, section,
device
A (1)
Wn

Feedback
I section,
A (2)
W"
I
I:
I
I Figure 7.28 Block diagram of decision-feedback equalizer.
,
I

L
 . "'"

7.10 EYE PATTERN 461 -)

where the vector w~l) denotes the tap-weights of the feedforward section, and
1
w~2)denotes the tap-weights of the feedback section. Let the augmented vector I
vn denote the combination of input samples for both sections:
I
(7.112)
Vn = [::] 1

I
I
where Xn is the vector of tap-inputs in the feedforward section, and an is the I
vector of tap-inputs (i.e., present and past decisions) in the feedback section.
The common error signal is defined by

en = an - c~v n (7.113)

where the superscript T denotes matrix transposition, and an is the polar rep-
resentation of the nth transmitted binary symbol. The LMS algorithm for the
decision-feedback equalizer is described by the update equations:

w~lll = w~l) + J.Llenxn (7.114)

w~2]. 1 = w~2) + J.L2eiln (7.115)

where J.Lland J.L2are the step-size parameters for the feedforward and feedback
"
sections, respectively. "
A decision-feedback equalizer yields good performance in the presence of
moderate to severe intersymbol interference as experienced in a fading radio
~ for example. . ~

7.10 EYE PATTERN

In previous sections of this chapter we have discussed various techniques for

dealing with the effects of receiver noise and intersymbol interference on the
performance of a baseband pulse-transmission system. In the final analysis, what
really matters is how to evaluate the combined effect of these impairments on
overall system performance in an operational environment.~n experim~tal
tool for such an evaluation in an insightful manner is the so':called eye pattern,
which is defined as the synchronized superposition of all ossible realizations of
t e signa of interest (e.g., received signal, receiver output) viewed within a par-
tKular signaling interval. The eye pattern derives its name from the fact that it
II
resemoles the human eye fur binary waves. The interior region of the eye pattern
is called the eye opening.
r. ,J
An eye pattern provides a great deal of useful information about the per-
formance of a data transmission system, as described in Fig. 7.29. Specifically, II,
I
we may make the following statements:

. The width of the eye opening defines the time interval over which the received
signal can be sampled without errorfrom intersymbol interference; it is apparent that
the preferred time for sampling is the instant of time at which the eye is open
the widest. '

- -- -------
"---I