Chap 4

Chapter 4 Baseband Pulse
Transmission
Techniques for the transmission of (originally)

digital data over a baseband channel are the main focus of this chapter.
4.1 Introduction
o Transmission of digital data (bit stream) over a noisy baseband channel typically suffers two channel imperfections
n Intersymbol interference (ISI)
n Background noise (e.g., AWGN)
o These two interferences/noises often occur simultaneously. However, for simplicity, they are often separately considered
Chapter 4-2
1
4.1 ISI
(t) Impulse response
ISI
channel
ISI
ISI
channel
ISI
ISI
channel
Chapter 4-3
4.2 Matched Filter

Matched filter is a device for the optimal detection of a digital pulse. It is so named because the impulse response of the m
System model without ISI
channel
Chapter 4-4
4.2 Design Criterion
o To find h(t) such that the output signal-to-noise ratio SNRO
is maximized.
x(t) g(t) w(t) for 0 t T y(t) [g(t) w(t)]* h(t)
g(t)* h(t) w(t)* h(t)
go (t) n(t)
| g (T ) |2
SNR O o
E[n2 (T )]
Chapter 4-5
4.2 Analysis of Matched Filter
go (t) H ( f )G( f ) exp( j2ft)df

| g (T ) |2 H ( f )G( f ) exp( j2fT )df

o 2
With w(t) being white with PSD N 0 / 2,

N0
SN ( f ) SW ( f ) | H ( f ) |2 | H ( f ) |2
2
2
N0 2
E[n (T )] S ( fN )df
2 | H ( f ) | df
Chapter 4-6
2

G( f )H ( f ) exp( j2fT )df

N0
2
2 | H ( f ) | df
Cauchy-Schwarz inequality
2

| (x) |
2 2
dx|
(x) | dx
(x) 2 (x) dx
1 1 2
*
with equality holding if, and only if, 1(x) k 2(x) for some constant k.
Chapter 4-7

o By Cauchy-Schwarzs inequality,
2

2 2
| H ( f ) | df | G( f )exp( j2 fT ) | df
H ( f )G( f )exp( j2fT )df

2
| H ( f ) | df | G( f ) |2 df
2
2
| G( f ) | df

N0
N0
2
| H ( f ) | df
2
This is a constant bound, independent of the choice of h(t). Hence, the optimal is achieved by:
H( f ) k G*( f )exp( j2fT )

Chapter 4-8
4.2 Analysis of Matched F
Z 1
hopt(t)= k t(
1
Z 1
t(f ) exp(j2f (T t))df =k
1
=kg(T t).
o Hence, under additive white noise, the optimal received filter matches the input signal in the se
4.2 Properties of Matched Filter

o The maximum output signal-to-noise ratio only depends on the energy of the input, and is nothing to do with the pulse sh
n Namely, whether the pulse shape is sinusoidal, rectangular, triangular, etc is irrelevant to the maximum output signal-to-n
2E
max s , where E s| G( f ) |2 df .
0

N
Chapter 4-
Example 4.1 Matched Filter for Rectangular Pulse
Also h(t)
Chapter 4-11
Example 4.1 Matched Filter for Rectangular Pulse

o hopt(t) in this example can be implemented as integrate-and- dump circuit
Chapter 4-12
4.3 Error Rate due to Noise
o In what follows, we analyze the error rate of polar non- return-to-zero (NRZ) signaling in a syste
s(t) I g(t), where I {1,1}.

y(T ) [I g(t)]* h(t) t T
w(t)* h(t) t T

I h( )g(T )d h( )w(T )d
I kg (T )g(T )d kg (T )w(T )d

* *
I k | g( ) | 2d k g (* )w( )d
I kE g kn, where E g | g( ) | d2 and n g (* )w( )d .

For notational convenience, brief y(T ) / k by y.
(The integration can be taken over [0,T ) since g(t) is zero outside this range, as does in text. I however use the entire
real line as the integration range here for convenience.)
Chapter 4-14
By AWGN assumption of w(t), and real g(t) assumption,
n g * ( )w( )d is Gussian distributed with

E[n] g ( )E[w(
*
)]d 0.

2
E[n ] g(s)g(t)E[w(s)w(t)]dsdt

N0
g(s)g(t)2 (s t)dsdt
N 0 g 2 (s)ds N0 E

2
2
g
1 ( y) Normal (E g , E g N 0 / 2), if I 1;
n
yg I E 1 ( y) Normal (E g , E g N 0 / 2), if I 1
Chapter 4-15
Let be the set for which a decision favors 1 is made.

BER Pr[I 1] Pr{guess (1) | I 1} Pr[I 1] Pr{guess (1) | I 1}
Pr[I 1] Pr{y | I 1} Pr[I 1] Pr{y | I 1}
p1 Pr{y | I 1} (1 p) Pr{y | I 1}
p (1 p) Pr{y | I 1} p Pr{y | I 1}
p [(1 p)1( y) p1( y)]dy, where p Pr[I 1].
To minimize BER, the optimal set opt y : (1 p)1( y) p1( y) 0.

Thus, the optimal decision maker is given by :
d ( y) 1, (1 p)1( y) p1( y)
.
1,(1 p)1( y) p1( y)
Chapter 4-16
1 ( y) Normal (E g , E g N 0 / 2), if I 1;

1 ( y) Normal (E g , E g N 0 / 2), if I 1
Let Eg and Eg N0 / 2.
2
1 ( y )2
1 exp 2
(1 p) 1( y) 2 2 2

p 1 ( y) 1 ( y )2
exp
2 2
1
2 2

exp 2y exp 2Eg y exp 4 y
2
E g N 0 / 2 N0
1
Thisy threshold
N 0 log (1 p) on N0; hence, the best decision relies on the accuracy of N0 estimate.
depends

1 4 p
Chapter 4-17
4.3 Error Rate due to Noise under Uniform Input

o In order to free the system dependence on N0 estimate, a uniform I is transmited in which case, p = .
1

o The best decision now becomes y0.
1
1 0 1
BER opt
20 ( y)dy
1
2 ( y)dy
1
1 1 ( y )2
exp dy
0 2 2
2
2

2 0 1
( y )2
1 exp dy
2
2
2 2 2

Chapter 4-18
1 1 y2
exp dy
BERopt 2 2
2 2
2
1 1 y 2
exp 2 2 dy
2 2

2
1 y2 y
exp 2 2 dy, z
2 2 2 2

1
/ exp zdz
2
2 2
1 1erfc Eg
erfc
2 2 2 2 N0

2 2
where erfc(u) exp(z )dz is the complementary error function.
u
Chapter 4-19
4.3 Error Function

o Error function erf (u) 2 u
2
exp(z )dz

0
2
2
o Complementary error function erfc(u) exp(z )dz

u
1 z2

o Q-function Q(u) exp dz
u 2
2
erf (u) erf (u)
erfc(u) 1 erf (u)
Q(u) 1 erfc u

2
2
Chapter 4-20
10
4.3 Error Function
o Bounds for error function
1 x
2
e1 11 3 1 3 5 !
erfc( x)
x
22 43 6
2x2 x2 x
1 x
2
e1 1 erfc( x) 1e x 2
For x 0, 2
x 2x x
(The bound is good when x is large.)
Chapter 4-21
4.3 Error rate due to noise

o The optimal BER formula is important in communications:
BER 1 erfc Eg 2Eg

opt
2 N0 Q N0

1

o The best decision is y0.
1
Chapter 4-22
11
s(t) I g(t), where I {1,1}.
TT
In this case, 0 0
b E E[s 2 (t)]dt E[I 2 ]g 2 (t)dt E g
Chapter 4-23
4.4 Intersymbol Interference

The channel is usually dispersive in nature.
In this section, we only consider discrete pulse-amplitude modulation (PAM). Consideration of PDM and PPM will be out of
bk {0,1}, ak 2bk 1 and s(t) kak g(t kTb ).

Chapter 4-24
4.4 Intersymbol Interference
Notably, in the previous section, we only consider one interval of input.
s(t) I g(t)
This is justifiable because of no ISI.
However, in this section, we have to consider
s(t) ak g(t kTb ).
since ISI is involved.

o We also assume perfect synchronization to simplify the analysis.
Chapter 4-25

Information of ak is carried at [kTb, (k+1)Tb).
s(t) k ak g(t kTb )

x(t) s(t) * h(t) w(t) k ak [g(t kTb ) * h(t)] w(t)

y(t) x(t) * c(t) k ak [g(t kTb ) * h(t) * c(t)] w(t) * c(t)

b
y(iT ) a [g(t kT )*b h(t)* c(t)]
k
k t iTb
w(t)* c(t) t iTb
Chapter 4-26

g(t kTb ) * h(t) * c(t) G( f ) exp{ j2fkTb} H ( f ) C( f ) exp{ j2ft}dt

G( f ) H ( f ) C( f ) exp{ j2f (t kTb )}dt

p(t kTb )
where p(t) G( f )H ( f )C( f )exp{ j2ft}dt.
y(t) k
ak p(t kTb ) n(t), where n(t) w(t) * c(t)

y(iTb ) ak p(iTb kTb ) n(iTb ) k

Chapter 4-27
4.4 ISI and Noise

o Without ISI,
H ( f ) 1 p(t) G( f )C( f )exp{ j2ft}dt.

* *
With matched filter C( f ) G ( f )exp{ j2 fTb }, or c(t) g (T b t),
Z 1
p(iTb) = c( )g(iTb )d
1
Z 1
= g (Tb )g(iTb )d(Let s = Tb )
1
8<0, if i 6= 1
Z 1
= g (s)g(s + (i 1)Tb )ds = Z 1
2
|g(s)| ds, if i = 1
1 :
Chapter 4-28

y((i 1)Tb ) ak p((i 1 k)Tb ) n((i 1)Tb ) ai p(Tb ) n((i 1)Tb )

k
As a result, without ISI and additive noise,

y((i 1)Tb ) ak p((ik

1 k )Tb ) ai p(Tb )
and {ai} can be completely reconstructed by{y((i 1)Tb )}.
The text sets p(0) = 1 for simplicity but this is a litle confused (See Slide 4-28)! The text is correct whe
during [(i-1)Tb, iTb).
Information of ai is actually carried at [iTb, (i+1)Tb).
So in order to recover ai, correlation (convolution) operation should start at iTb, and end (i.e., is sam
Hence, y((i+1)Tb) is used to reconstruct ai.
Chapter 4-29
4.5 Nyquists Criterion for Distortionless

Baseband Binary Transmission
o Is it possible to completely eliminate ISI (in principle) by selecting a proper g(t) ?
Z 1
Choose g(t) and c(t) such that p(t) =t(f )H(f )C(f ) exp(j2f t)dt
1
( 0, if i 6= 0
satisfies p(iTb) = .
1, if i = 0
(Here, I assume that information of ai is carried at [(i 1)Tb, iTb).)
Chapter 4-30
Let P(f) = G(f)H(f)C(f).
Sample p(t) with sampling period Tb to produce P(f).
From Slide 3-4, we get:
1 n
P ( f ) T bn
P f T
b
o Also from Slide 3-4, we have:

P ( f ) np(nTb ) exp j2nTb f 1

(1, if n = 0 !
Because p(nTb)=
0, if n 6=0 Chapter 4-31

o This concludes that the condition for zero ISI is:

n

n
nP f T b Tb (Indeed, f
P T constant.)
b
n
o This is named the Nyquist criterion.

n The overall system frequency function P(f) suffers no ISI for samples taken at interval Tb if it satisfies the above equation.
n Notably, P(f) represents the overall accumulative effect of transmit filter, channel response, and receive filter.
Chapter 4-32
4.5 Ideal Nyquist Channel
o The simplest P(f) that satisfies Nyquist criterion is the rectangular function:
T ,| f | W 1
b 2Tb 1
P( f )
2Tb and P(W ) P(W ) Tb.
0, | f | W

p(t) sin(2Wt) sinc(2Wt)
2Wt
Chapter 4-33
Chapter 4-34
(i 1)Tb ~ iTb
1
0
1
1
0
1
0
Chapter 4-35
4.5 Infeasibility of Ideal Nyquist Channel

o Rectangular P(f) is infeasible because:
n p(t) extends to negative infinity, which means that each ak
have already been transmited at t = !
n A system response being flat from W to W, and zero elsewhere is physically unrealizable.
n The margin of error is quite small, as a slight shift (error) in sampling time (such as, iTb+) would cause very large ISI.
o Note that p(t) decays to zero at a very slow rate of 1/|t|.
Chapter 4-36
4.5 Infeasibility of Ideal Nyquist Channel
o Examination of timing error margin
n Let t be the sampling time difference between transmiter and receiver.

y(iTb t) akk
p((i k)Tb t)
n For simplicity, set i = 0.

y(t) k ak p(t kTb )

sin[2W (t kTb)]
k a

k
2W (t kTb )
Chapter 4-37

sin[2Wt k ]
y(t)
ak 2Wt k
k
(1) k sin[2Wt] 2Wt k

a
k
k
sin[2Wt] sin[2Wt](1)

ka

k
a0
2Wt k 2Wt k

k 0

(1) a k k
There exists {ak} such that 2Wt k for any fixed small t 0.
k
k 0
Question: How to make p(t) decays faster? Answer: Make P(f) smoother.
Chapter 4-38
4.5 Raised Cosine Spectrum
For a nonnegative function p(t),

k k P( f )
ift p(t)dt , then
exists.
f k

Chapter 4-39

o We extend the bandwidth of p(t) from W to 2W, and require
that 1
P( f ) P( f 2W ) P( f 2W ) for | f | W .
2W
n So the price to pay here is a larger bandwidth.
n One of the P(f) that satisfies the above condition is the
raised cosine spectrum.
1
2W , 0 | f | (1 )W
1 (| f | (1 )W )
P( f ) (1 )W | f | (1 )W
4W 1
cos , 2W
| f | (1 )W
0,
Chapter 4-40
20
The text puts BT = W(1+) may not be justifiable!

o The transmission bandwidth of the raised cosine spectrum is equal to:
BT 2W(1 )
where is the rolloff factor, which is the excess bandwidth
over the ideal solution.
Chapter 4-41
cos(2Wt)
p(t) 22 2
sinc(2Wt) 1 16 W t

1
~as | t | large
| t |3
Chapter 4-42
21
p(t) sinc(2Wt)
o 22 2 consists of two terms:
cos(2Wt) 1 16 W t
n The first term ensures the desired zero crossing of p(t).
n The second term provides the necessary tail convergence rate of p(t).
o The special case of = 1 is known as the full-cosine rolloff

characteristic.
sinc(4Wt)
p(t)
1 16W 2 t2
Chapter 4-43

o Useful property of full-cosine spectrum.
1,
i
iT 1 0
p b
2 2, i
0, 1
n We have more zero-crossing at 3Tb/2, 5Tb/2, 7Tb/2, i 2in addition to the desired Tb, 2Tb, 3Tb
n This is useful in synchronization. (Think of when synchronized, the quantity should be small both at 3Tb/2,
5Tb/2, 7Tb/2, and at Tb, 2Tb, 3Tb)
n However, the price to pay for this excessive synchronization
information is to double the bandwidth.
Chapter 4-44
Example 4.2 Bandwidth Requirement of the T1
System
For T1 transmission, a frame consists of 24 PCM-encoded voice channels and 1 framing bit.
n The resultant number of bits in a frame is 24 8 + 1 = 193.
The duration of each frame is 125s.
Hence,
125s 0.647688s
Tb
193
1
W 772kHz.
2Tb WW
BT ,( Baseband) W (1 ) 772(1 )kHz
Chapter 4-45
4.6 Correlative-Level Coding

ISI, when generated in an uncontrolled manner, is an undesirable phenomenon.
However, ISI may become a friend if it is added to the transmited signal in a controlled manner.
n Known fact: A signal of bandwidth W can be distortionlessly transmited using its samples with sampling rate 2W.
n Conversely, in a channel with bandwidth W Hz, the theoretical maximum signal rate is 2W symbols per second.
Chapter 4-46
A channel with bandwidth W Hz
-BB
-W W
-W W
The maximum signal rate is
2W samples per second.
-WW
Chapter 4-47

o Why intentionally adding ISI? Answer: To have beter bandwidth efficiency.
n Ideal Nyquist pulse shaping is efficient; it cannot be realized.
n Raised consine pulse shaping is realizable; it is bandwidth inefficient.
n By adding ISI to the transmited symbols in a controlled manner, we can achieve the Nyqusit rate 2W in a channel bandw
o Correlative-level coding or Partial-response signaling
Chapter 4-48
4.6 One Example of Correlative-Level Coding
o Duobinary signaling (or class I partial response)
HduoB( f )
where {bk} i.i.d.
Chapter 4-49
4.6 Duobinary Signaling

o Let us ignore the effect of HNyquist(f) first in the block diagram in the previous slide. We directly obtain:
HDuoB( f ) 1 exp( j2fTb )

n Note that ck has three levels (2,0,2).
o The transfer function of the overall system is thus:
Chapter 4-50
o HNyqusit(f):
n Only for derivation purpose (do not need it finally)
n As shown in the next slide, the response HI(f) is realizable.
Chapter 4-51

o HI(f)
||
Chapter 4-52
o hI(t):
Text omits this term

hNyquist (t) hNyquist (t Tb ) by saying except for a scaling factor.
. sin(vt/Tb)sin[v(t Tb)/Tb] .
= +vt/Tb x Tb
v(t Tb)/Tb
. sin(vt/Tb)sin[vt/Tb] .
= vt/T v(t T )/T bb x Tb
b
Tbsin(vt/Tb)
= x Tb
(Tb t) vt/Tb
Chapter 4-53

o hI(t):
Chapter 4-54
o Bandwidth efficiency of duobinary signaling
n Example.

The transmited signal k a k g(t kTb ) ak (t kTb ) * g(t)
k

a (t kT )
k
k b
g(t)
The input to this filter may not be WSS!

Then we should use the time-average autocorrelation function.
Chapter 4-55
4.6 Time-Average Autocorrelation Function
X (t) H(f) Y (t)
Time Average Autocorrelation Function

1 T 1 T
RX ( ) lim E[ X (t) X (t )]dt lim R X (t,t )dt
T 2T T 2T
T T
1 T
RY ( ) RY (t,t )dt
T 2T T
lim
1 T

h( )h(12X )R (t 1 ,t )d212d dt
lim T 2T T
(Assume that limit and integration are interchangeable.)
Chapter 4-56

1 T
RY ( ) h( 1 )h( 2 ) RX (t 1 ,t 2 )dt d1d 2
T 2T T

lim
h( 1 )h( 2 )RX ( 2 1 )d 1d
2
SY ( f ) RY ( ) exp(j2f )d
h( 1 )h( 2 ) R X (

2 1 ) exp(j2f )dd 1d 2
h(

1)h( 2 ) RX ( ' ) exp(j2f [ ' 2 1])d ' d 1d 2

h( )e j 2f d2 h(1)e j 2f d1 X
j 2f '
2 1
R ( ' )e d '
2
H ( f )H *( f )S X( f ), if h( ) is real
| H ( f ) |2 SX ( f )
Chapter 4-57

o Now back to the example.
X (t) ak (t kTb ) Y (t) k ak g(t kTb )

g(t)
k (to channel)
1 T
RX ( ) lim E k a (t kT b)a
j (t jT )dtb
2T
T T
k

j

1 T
lim E[ajak ] (t jTb ) (t kTb )dt
T 2T T k j
(1, j = k
Assume E[ajak]= .
0, j 6= k Chapter 4-58
1 T
T 2T (t kTb ) (t kTb )dt
RX ( ) lim
T k

1 T
( T)lim

T k
(t kT b )dt

2T
1 1 2
( ) SY ( f ) | G( f ) |
Tb Tb
Approximately 2T of them
Tb
T 0 T
Chapter 4-59
X (t) ak (t kTb ) Y (t)

g(t) hDuoB (t)
k (to channel)
1
SY ( f ) | G( f ) |2| H ( f ) |2
DuoB
b T
Chapter 4-60
30
HDouB( f ) 2cos(fTb ) exp( jfTb )

Assume g(t) 0 t Tb
0,otherwise . | G f | b sinc fTb
22 2
1,
T
S (f) sinc2 ( fT ),b No Signal ISI
Y
With Signal ISI
2 (fT )sinc
Y S (0)cos b 2 ( fT ), b
sinc
No 2 ( fTb ),
Signal ISI
With2 Signal ISI
sinc (2 fTb ),

Chapter 4-61
1.2
0.8 sinc2 ( fTb ), without Signal ISI

0.6
0.4 sinc2 (2 fTb ), with Signal ISI

0.2
0
-2-1.5-1-0.500.511.52
fTb
Chapter 4-62
31
o Conclusions
n By adding ISI to the transmited signal in a controlled (and reversible) manner, we can reduce the requirement of b
n Hence, in the previous example, {ck} can be transmited in every
Tb/2 seconds!
o Doubling the transmission capacity without introducing additional requirement in bandwidth!
n Duobinary signaling : Duo means doubling the transmission capacity of a straight binary system.
n A larger SNR is required to yield the same error rate because of an increase in the number of signal levels (from 1
Detailed discussion on error rate impact is omited here!
Chapter 4-63

Conclusions (cont.)
n The duobinary signaling is also named class I partial response.
Full response: The transmission wave at each time instance is fully determined by a single information symbol.
Partial response: The transmission wave at each time instance is only partially determined by one information symbol (i.e.,
Chapter 4-64
4.6 Decision Feedback for Correlative-Level Coding
o Recovering of {ak} from {ck}
ak ck ak HduoB( f )
1
n It requires the previous decision to determine the current

symbol.
n So the system should feedback the previous decision.
n Error therefore may propagate!
n How to avoid error propagation? Answer: Precoding.
Chapter 4-65
4.6 Precoding of Correlative Coding

Without precoding
ck ak ak 1
{bk {0,1} i.i.d.} ak 2bk
With precoding 1
{bk {0,1} i.i.d.} ~ ~ ~

bk bk bk 1 ak 2bk 1 ck ak
ak 1
ck ak ak 1 ~ ~
~ ~bk bk 1 bk ck
bk (2 1) bk 1 0 0 0 2
~ ~
2b(21)
2b 2 0 1 1 0
~ k ~ k 1 11 0 1 0
b b 1 0 2
k k bk 1
Po-Ning Chen@ece.nctu Chapter 4-66

o Final notes
n The precode must not change the duo- of the transmission capacity of a straight binary system
n Hence, {b!
k
} must have the same distribution ask{b }
and hence must be i.i.d.
Chapter 4-67

{~ }
o Uniform I.i.d. of bk
~~~ ~

n For i.i.d., it suffices to show Pr b |kk b,b
1k 2 ,... bk
~ ~ ~~~ ~~
bk bk bk 1 Pr b | b,b
kk 1k 2 Pr
,... Prb | b
kk 1
Pr( ~
bk 1
~ ~ bk ~0 | 0) Pr(b

k
0)
Pr(bk 01/|2bk 1 1) Pr(bk 1) ~~~ ~
~
1/ 2~
Pr( bk b1k 1| 0) Pr(b
k
1)

Pr b |kk b,b
1k 2 ,...
bk
~ ~
Pr

Pr(b 1/
k 1 |2bk 1 1) Pr(bk 0)
1/ 2
Chapter 4-68
n For uniformity,
Pr(~ ~ ~ ~
bk 1 bk 1
bk 0) 0) bk 0 |
~ ~ ~
Pr(
Pr(bk 1 1)Pr(
Pr(bk 0)
0 | bk 1
1)~ 1 ~ 1
Pr(bk 1 0) Pr(bk 1
2 2
1)
1

2
~ ~ ~ ~
Pr(bk 1) Pr(bk 1 0) Pr(bk 1 | bk
1 0) ~ ~ ~
b
Pr( k 1
b 1) b 1 k 1
k
~ 1 ~ 11)
Pr(bk 1 0)Pr( 2 Pr(bk |1
1) 2
1
2 Q.E.D.
Chapter 4-69
Example 4.3 Duobinary Coding with Precoding

o Table 4.1 in text
{b~k}
{ }
bk
{ak}
{ck}
{b
}
{bk {0,1} i.i.d.} ~ ~ ~
ak 1
Chapter 4-70
4.6 Modified Duobinary Signaling
The PSD of the signal is nonzero at the origin.
This is considered to be an undesirable feature in some applications, since many communication
Solution: Class IV partial response or modified duobinary technique. 1.2
1
0.8
0.6
0.4
0.2
0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
Ch
{bk {0,1} i.i.d.} ~ ~ ~

ak 2
MDuoB (f)
HMDuoB ( f ) 1 exp( j4fTb )
Chapter 4-72
HMDuoB ( f ) 1 exp( j4fTb )

[exp( j2fTb ) exp( j2fTb )]exp( j2fTb )
2 j sin(2fTb ) exp( j2fTb )
Assume g(t) 0 t Tb
. | G f | b T sinc fT
22 2
1,
0,otherwise
b
2

S ( f ) / T sinc (2 bfT
bY Duobinary (See Slide 4-61)
), 2 2
SY ( f ) /(4Tb ) sin (2fTb )sinc ( fTb ),Modified
Duobinary
Chapter 4-73

1.2
sinc2(2 fT ), Duobinary b
1
0.8 b sin2 (2fTb )sinc2 ( fT ),

Modified duobinary
0.6
0.4
0.2
0
-2-1.5-1-0.500.511.52
fTb
Chapter 4-74
o Precoding is added to eliminate error propagation in decision system.
ck ak ak 2 ~ ~
bk bk 2 bk ck
~ ~ 1)
(2bk 1) (2bk 2 0 0 0
~ ~ 0 1 1 2
2b k 2 0 1 2
~ ~ k 2 1
bk bbk bk 2 1 1 0 0
~ ~ ~
2
{bk {0,1} bk bk bk 2 ak bk 1 ck ak ak 2
i.i.d.}
Po-Ning Chen@ece.nctu Chapter 4-75
4.6 Generalized
Form of Correlative Level Coding (CLC) or Partial Response Signaling
1
w z ! w N 1 zN 1,
H CLC ( f ) w 01
where z exp( j2fTb ).
Chapter 4-76
4.6 Generalized Form of Correlative-Level Coding
or Partial-Response Signaling
Type of Class N w0 w1 w2 w3 w4 Comments
I Duobinary coding
II
III
IV Modified duobinary coding
V
4 cos (fTb )
2
I

16cos (fT )
4
2 II
| G( f ) |
S (f) 4 cos2 (fT ) 8sin2b(2fT ) III
Y
Tb b b
4sin (2fT )
2
IV
b
16sin (2fTb )
4
V
Chapter 4-77
Assume g(t) 0 t Tb
. | G f | bTsinc fTb
22 2
1,
0, otherwise
2
III
1.5
1
IV I
0.5 V II
0
-1.5 -1 -0.5 0 0.5 1 1.5
Chapter 4-78
4.7 Baseband M-ary PAM Transmission
Gray code
Any dibit differs from an adjacent dibit in a single bit position.
Chapter 4-79

For M-ary PAM transmission, there are M possible symbols with symbol duration T.
n 1/T is referred to as the signaling rate or symbol rate or
symbols per second or baud.
Some equivalences
n Each symbol can be equivalently identified with
Baud = log 2M bits.
the number of times a signal changes state per second
n So the baud rate 1/T can be equivalently transformed to bps as:

T Tb log2 (M )
Chapter 4-80
40
o Some equivalences
n Virtually fix the symbol error, namely, fix the level distance (to be 2). For example, (, ) for M
E[S 2 ] M
1 [(M
T
(M 2 1)

3Tb log2 (
E[S 2 ] 1 (M 2 1)
T T 3log (M )
b 2
For fixed Rb = 1/Tb (bps) and level distance = 2, the transmited power of an M-ary PAM transmission signal is
increased by a factor M2/log2M.
Chapter 4-82
41
4.8 Digital Subscriber Lines (DSL)
o A DSL operates over a local loop (often less than 1.5km) that provides a direct connection between a use
n Since it is a direct connection, no dialup is necessary.
n The information-bearing signal is kept in the digital domain all the way from the user terminal to an Inte
SO
4.8 Digital Subscriber Lines

DSL is intended to provide high data-rate, full-duplex, digital transmission capability using local cost configuration (such as t
One of two possible modes can be used to achieve the full- duplex goal.
n Time compression multiplexing (TCM) mode
n Echo-cancellation (EC) mode
Chapter 4-84
Digital Subscriber Lines
n Time-compression multiplexing (TCM) mode
A guard time is often inserted between bursts in the two opposite directions of data.
So the required line rate is slightly greater than twice the data rate.
Transmiter Transmiter
Receiver Receiver
Chapter 4-85

n Echo-cancellation (EC) mode
Support the simultaneous flow of data along the common line in both directions.
In this mode, the line rate is the same as the data rate.
Chapter 4-86
o Hybrid transformer for DSL
n Two-to-four-wire conversion
Zref zl
Chapter 4-87

o Comparison between TCM mode and EC mode
n EC offers a much beter data transmission performance at the expense of increase complexity.
n However, with the recent advance in VLSI, complexity is no longer a main system concern. So in North America, the EC m
Chapter 4-
o Other impairments to DSL
n ISI and Crosstalk
o The transfer function of a twisted pair line can be approximated by
2

| Htwist pair ( f ) | exp f
l
where k, k is a physical constant of the twisted pair, and
l0
l0 and l are respectively the reference length and actual length of the twisted pair.
Chapter 4-89

n ISI
1
Htwist pair( f ) | f | exp j2 f 0
4.5 exp 2
4 4
htwist pair( )
3.5 )2
3 0
2.5 116 2 (
2
1.5 (t) (t 1) htwist pair ( ) 4
1
1 16 2 ( 0.1)2
0.5
z0 = 0.1
0
-1 0 1 2 3 4 5
Chapter 4-90
n Crosstalk
o Capacitive coupling that exists between adjacent twisted pairs in a cable
n Near-end crosstalk (NEXT) and Far-end crosstalk (FEXT)
Near-end crosstalk Far-end crosstalk

Chapter 4-91

n Crosstalk (cont.)
FEXT suffers the same line loss as the signal, whereas NEXT does not.
n This is close to the phenomenon of near-far effect
of wireless channel.
Accordingly, NEXT will be a more serious problem than FEXT. So we can ignore the effect of FEXT, and add NEXT filter to the
Chap
3/ 2
HNEXT( f ) f
Interference (input of HNEXT(f)) often assumes to have the same PSD as the transmited signal, but is Gaussian distribu
Chapter 4-

o Other features of DSL channel
n The PSD of the transmited signal should be zero at zero frequency because no DC transmission through a hybrid transfor
n The PSD of the transmited signal should be low at high frequencies because
o transmission atenuation in a twisted pair is most severe at high frequency;
o crosstalk due to capacitive coupling between adjacent twisted pairs increases dramatically at high frequency (recall that t
Chapte
o Possible candidates for line codes that are suitable for DSL
n Manchester code
o Zero DC component but large spectrum at high frequency so it is vulnerable to NEXT and ISI.
n Bipolar return to zero (BRZ) or Alternate mark inversion (AMI) code
Successive 1s are represented alternately by positive and negative but equal levels, and 0 is repre
Zero DC component. Its NEXT and ISI performance is slightly inferior to the modified duobinary co
Chapter 4-95

n Modified duobinary code
o Of no DC component and moderately spectrally efficient. However, its robustness against NEXT and ISI is about 2 to 3 dB poorer than
n 2B1Q code
Two binary bits encoded into one quaternary symbol (four-level PAM signal).
Zero DC component, and offers the best performance among all the codes introduced. So it is adopted as the standard as the North Am
Chapter 4-96
n 2B1Q code (cont.)
Chapter 4-97

n 2B1Q code (cont.)
o With 2B1Q line coding, adaptive equalizer and echo cancellation, it is possible to achieve BER = 10 7 operating full duplex
Chapter 4-98
4.8 Asymmetric Digital Subscriber Lines
o ADSL is targeted to simultaneously support three services at a single twisted-wire pair
n Data transmission downpstream at 9 Mbps
n Data transmission upstream at 1Mpbs
n Plain old telephone service (POTS)
o Some notes
n It is named asymmetric because the downstream bit rate is much higher than the upstream bit
n The actually achievable bit rates depend on the length of the twisted pair used to do the transm
Chapter 4-99

Frequency-division multiplexing (FDM) technique is used to combine analog voice and DSL data.
Upstream and downstream data transmission are placed in different frequency band to avoid crosstalk.
spliter
Chapter 4-100
50
Various applications can be applied to asymmetric transmissions, such as video-on-demand (VoD
n For example
Downstream = 1.544 Mbps (DS1) for video data
Upstream = 160 kbps for real-time control commands.
Chapter 4-101
4.9 Optimum Linear Receiver

o Zero-forcing equalizer
n A receiver design is to use a zero-forcing equalizer
followed by a decision-making device.
n The design objective of a zero-forcing equalizer is to force the ISI to zero at all sampling instances t = kTb for k 0, provide
Chapter 4-102
51
o Zero-forcing equalizer (cont.)
n This reduces to the Nyquist criterion.

n 1, n 0
nP f T b

Tb or p(nT b ) 0,

n0
where P(f) = G(f)H(f)C(f).
Chapter 4-103

o Zero-forcing equalizer (cont.)
n A serious consequence of the ignorance of w(t) in the design of a zero-forcing equalizer is the performance degradation d
Chapter 4-104
o Example of noise enhancement.
n Suppose that the receiver filter is a tapped-delay-line equalizer, which is of the form
c(t) ck (t kTb )
k 0
n Assume ideally that G(f) = 1.

n Hence, the Nyquist criterion becomes:
) 1, n 0
p(nTb
0, n 0
where P(f) = H(f)C(f).
Chapter 4-105
p(t) h( )c(t )d

h( )ck (t kTb )d
k 0

ck h( ) (t kTb )d
k 0

k
0
ck h(t kTb )

n0
1,
pn p(nTb ) ck h((nk 0knk0)T0b ) ck hnk 0,

Chapter 4-106
It is reasonable to assume that hn 0 for n 0, and h0 1.
1 1
0 0 " 0c0
0 h1 1 0 " 0c
h1 1 " 1
0 h2 0 ! for arbitrary N 0.
! ! ! ! ! c
! hN 1 hN 2 " N 1
0 hN 1 c N
1 1 | | /(2T
b ),0 2Tb
Suppose h( )
0,otherwise
2Tb
h 1, h 1 , and h 0 for n 0,1.
0 1 n
2
n n
cn (1) 2for (N ) n 0, and zero, otherwise.
Chapter 4-107
The above c(t) can successfully remove ISI, provided w(t) = 0.

Now add the additive white Gaussian noise w(t), which also passes the filter c(t).
At any time instance nTb, the sampled noise becomes
=w(z
w(z )c(nTb z )dz

) . ck6(nTb kTb z )dz
k=0
=. ckw(z )6(nTb kTb z )dz = . c
kw(nTb kT b )= . ckwnk
k=0 k=0 k=0
The sampled noise variance then becomes :

Varc w

4 2
ckVar wnk
2 2
2
k nk w
2 2k w
k 0 k 0
w
k 0 3
Chapter 4-108
o An easier way to interpret the noise enhancement phenomenon.
n The Nyquist criterion requires that:

n
n n
n
P f T
b
n
H f T C f T Tb
b b
n A sufficient condition for the Nyquist criterion is that:

H ( f )C( f ) Raised CosineSpectrum
n When H(f) is very small (or zero) at some frequency range,

C(f) has to be very large (or infinity) at the same frequency range in order to equalize the spectr
n Thus, the noise spectrum SW(f)|C(f)|2 after passing through

C(f) will be enhanced.
Chapter 4-109

To alleviate noise enhancement phenomenon, it is beter to simultaneously consider the ISI and channel noise.
An approach of this kind is to use the mean-square error criterion, and find a balanced solution to the problem of reducing
Chapter 4-1

y(t) c(t)* x(t) c( )x(t )d

x(t) ak q(t kTb ) w(t)
k
q(t) g(t)* h(t)

y(iTb ) ak
k
c( )q(iTb kTb )d c( )w(iTb )d i ni
For perfect receiver, y(iTb ) ai.

So, the error ei (i ni ) ai .
Chapter 4-111
The mean squared error criterion then wishes to minimize :

2 2
Ji E[ei ] E[{( i n )
ii a } ].
2 2 2
E[i ] E[ni ] E[a i ] 2E[ ini ] 2E[n iai ] 2E[ iai ]
1st term
For i.i.d. {ak} where ak = 1,

2
E[ ] i
kl
k l E[aa ]c(
)c( )q(iT kT )q(iT lT )d d
12bb1bb212
c( 1 )c( 2 )Rq ( 1 , 2 ;i)d 1d 2

where Rq ( 1, 2 ;i) q(iTb kTb 1 )q(iTb kTb 2 )
k
Chapter 4-112
Observe that Rq (1, 2;i) q(iTb kTb 1 )q(iTb kTb 2 ) only
k
depends on the difference between 1 and 2, and is invariant with respect to i. We can then

E[i 2] = c(z1)c(z2)Rq(z1 z2)dz1z2

. .
= c(z1)c(z2) Sq (f )e2uf (z1 z2 ) df dz1z2

. . . .
=Sq(f ) c(z2 )e2uf z2 dz2 df
c(z1 )e2u(f )z1 dz1

= Sq(f )C(f )C(f )df

Sq(f )|C(f )|2df.
=
Chapter 4-113
2nd term Assume white w(t) with PSD N0/2.
2
E[n i ] c( )c(
12 )E[w(iT b1 )w(iT )]d
b212 d
c( )c( ) N0 ( N0 2

12
2
1212
) d d
2
c ( ) d
11
N0 . . . .
= C(f1 )e2uf1 z df1 C(f2 )e2uf2 z df2 dz

N0 2 . .
= C(f1)C(f2) df1df2
2
e2u[(f1 +f2 )]z dz
N0
=
2 C(f1)C(f2)6(f1 f2)df1df2
N0
= 2 C(f1)C(f1)df1

N0
= 2 |C(f )|2df.
Chapter 4-114
3rd term
2
For i.i.d. {ak} where a k = 1, E[a i ]
1.
4th and 5th term
By independence of {ak} and w(t), and zero mean of ni,
E[ini ] E[i ]E[ni ] 0 and E[niai ] E[ni ]E[ai ] 0.
6th term

E[iai ] k E[ak ai ] c( )q(iTb kTb )d c( )q( )d

. . . .
= C(f1 )e2uf1 z df1 Q(f2 )e2uf2 (z ) df2 dz

. .
= C(f1)Q(f2) df1df2

e2u(f2 f1 )z dz
Chapter 4-115

E[iai] = C(f1)Q(f2)6(f2 f1)df1df2

= C(f )Q(f )df

= [Cr(f )Qr(f ) Ci(f )Qi(f )] df
where the last step follows from the observation that E[iai] must be a real number, and Cr(f ) and Ci(f ) are respectively the real and imaginary part
Substitute all six terms into Ji.

..
N0 .2 .
Ji = Sq(f )2+ |C(f ) | 2Qrr(f )C (f ) + 2Q (f ii)C (f ) df + 1

s x
A(f )
Chapter 4-116
. N0 . 2
A(f ) = S q (f )2+ |C(f ) | 2Q rr(f )C (f ) + 2Q (f )C
ii (f )
.
= N0 . 2
S q(f )+ rC (frr ) 2Q (f )C (f )
2
.
N . 2
q
+ 0S (f )+ iiiC (f ) + 2Q (f )C (f )
2
.
N
=S0 q..(f ) + Q2(fQ)r(f ).2 r
Cr(f )
2 (Sq(f ) + N0/2)(Sq(f ) + N0/2)
N0 .. Q2(f ) . Qi(f ).2
i
+
Ci(f ) + S q(f ) +
2 (Sq(f ) + N0/2)(Sq(f ) + N0/2)
Q*( f )
Sq ( f ) N0 / 2 for
C(MMSE
f ) equalizer.
An equalizer that is so designed is referred to as the minimum- mean square error (MMSE) equal
Chapter 4-117
4.9 MMSE Equalizer

o Summary
n The MMSE equalizer can be viewed as the concatenation of two filters:
o a matched filter Q*(f) to Q(f) = G(f)H(f)
o an equalizer whose frequency response is the inverse of Sq(f) + N0/2.
Chapter 4-118
4.9 MMSE Equalizer
o Property of Sq(f)
2
1 Q f k
n The text wrote that Sq ( f )
Tb k

Tb
, which
is periodic with period 1/Tb. This implies that Rq()

consists of a series of pulse train with width Tb, which is
not entirely true.
Rq ( 1 2 ) q(kTb 1 )q(kTb 2 )
k
Chapter 4-119

q(kT )q(kT ) exp( j2f )d
q S ( f ) R (q ) exp( j2f )d

b
k
b
k q(kTb ) q(kTb ) exp( j2f )d

k
q(kTb ) q(v) exp( j2f (kTb v))dv

k
q(kTb ) exp( j2fkTb ) q(v) exp( j2fv)dv
*
b
Q ( f )q(kT ) exp( j2 fkTb )
k
*

Q (f) q(t) (t kTb ) exp( j2 ft)dt
k
* 1 k
Q ( f ) Tb kQ fTb
Chapter 4-120
60
4.9 Implementation of MMSE Equalizer
o One can approximate 1/[Sq(f) + N0/2] by a periodic function with:
2
Sq ( f ) Q * ( f ) 1 Q f k1 k ~(f)
Q f Sq
Tb k TT
b b
k

T
b

~
o Since q(f) = 1/[Sq ( f ) N0 / 2] is now periodic with period
1/Tb, we obtain by Fourier series that

q ( f ) kck exp( j2kfTb )

1/(2Tb )
where ck T q( f ) exp( j2 kfTb)df .
b
1/(2Tb )
Chapter 4-121

o We can approximate q(f) by its main 2N+1 terms as:
N N
q ( f ) k ck exp( j2kfTb )
N
q ( ) k ck (t kTb )
N
One can therefore approximate 1/[Sq(f) + N0/2] by a transversal tapped-delay-line equalizer

Chapter 4-122
61
o Final notes
n In a real-life telecommunication environment, the channel is usually time-varying.
n Therefore, an adaptive receiver that provides for the adaptive implementation of both the matc
4.10 Adaptive Equalization

o The equalizer is adjusted under the guidance of a training sequence transmited through the channel.
Chapter 4-124
o Least-mean-square (LMS) algorithm
N
k 0
e[n] d[n] y[n] d[n] wk x[n k ]
o Design objective
n To find the filter coefficients w0, w1, , wN so as to minimize index of performance J:
J e2[n]
Chapter 4-125

o To minimize J, we should update wi toward the botom of the J-bowel.
J
gi
wi
n So when gi > 0, wi should be decreased.
n On the contrary, wi should be increased if gi < 0.
n Hence, we may define the update rule as:
1
w w
i ,nexti ,current
2 g i
where is a chosen constant step size, and is included only for convenience of analysis.
Chapter 4-126
2 N
J d[n] w x[n k

] k
k 0

N
N N
d [n] 2w kd[n]x[n k ] w kjw x[n k ]x[n j]

2
k 0 k 0 j 0
g J 2d[n]x[n i] 2w x[n k ]x[n i]

N
i
wi k
k 0
N
2x[n i] d[n] w x[n k ]

k
k 0

2x[n i]e[n]
Chapter 4-127
8 N
><e[n] =>X
>d[n]
> wk,currentx[n k]
) Repeat k=0
>>For
> 0 i N, wi,next = wi,current + x[n i]e[n]
>:For 0 i N, wi,current = wi,next
o Some notes on LMS algorithm

n There is no guarantee that the algorithm converges to a local minimum (could converge to a saddle point).
n There is even no guarantee that the algorithm converges.
Chapter 4-128
o Some notes on LMS algorithm (cont.)
n If is too large, high excess mean-square error may occur.
n If is too small, a slow rate of convergence may arise.
Chapter 4-129
4.10 Operation of the Equalizer

o Two modes of operations for adaptive equalizer
n Training mode (position 1)
n Decision-directed mode (position 2)
Chapter 4-130
4.10 Decision-Directed Mode
In normal operation, the decisions made by the receiver are correct with high probability.
Under such premise, we can use the previous decisions to
calibrate or track the tap coefficients.
In this mode,
n if is too large, high excess mean-square error may occur.
n if is too small, a too-slow tracking may arise.
o We can further extend the idea of decision-directed or
decision-feedback to the decision-feedback equalizer
(DFE).
Chapter 4-131
4.10 Decision-Feedback Equalizer

T
e annn nc v .
Subtract out ISI
wn (1) xn
Let cn w ( 2) and vn , where n sample time at nT
n a n .
T
Denote e annn nc v , where a nis the nth transmited symbol.
Chapter 4-132
4.10 Decision-Feedback Equalizer
o Then DFE gives:
(1)
wn1 wn(1) 1en xn
( 2) .
w( 2)n1 w
n
e2 nan

As anticipated, DFE suffers from error propagation due to
incorrect decisions.
However, error propagation will not persist indefinitely; rather, it tends to occur in bursts.
n E.g., if the number of taps in the feedback section is L, then the influence of one decision error
4.11 Computer Experiments: Eye Paterns

o Eye patern: The synchronized superposition of all possible realizations of the signal of interest viewed within a particular
1
-1 0
12 34567 t / Tb
Chapter 4-134
o Eye patern for pulse shaping function p(t) is half-cycle sine wave with duration Tb, and error-fre
1
-1 0 t / Tb
0.5 1 1.5 2
Chapter 4-135

o Eye patern for pulse shaping function p(t) is half-cycle sine wave with duration 2Tb, and error-free BPSK transmission.
-1
0 0.5 1 1.5 2 t / Tb
Chapter 4-136
Interpretation of Eye Patern
Chapter 4-137
-1
0 0.5 1 1.5 2
Chapter 4-138
Experiment 1: Effect of
channel noise
(Raise-cosine pulse-shaping with roll-off factor = 0.5, W = 0.5 Hz, M = 4)
Eye diagram for noiseless quaternary system.
Eye diagram for quaternary system with SNR 20 dB.
Eye diagram for quaternary system with SNR 10 dB.
Chapter 4-139
Experiment 2: Effect of
bandwidth limitation (Raise-cosine pulse-shaping with roll-off factor = 0.5, W
= 0.5 Hz, M = 4)
(a) Eye diagram for noiseless band-limited quaternary system: cutoff frequency f o 0.975 Hz (b) Eye diagram
for
noiseless band-limited quaternary system: cutoff frequency fo 0.5 Hz
(The channel is now modeled by
a low-pass Buterworth filter with
1
| H ( f ) |2
1 ( f / f 0 )50
Chapter 4-140
70

Chap 4

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chap 4

Uploaded by

Copyright:

Available Formats

Chapter 4 Baseband Pulse

Techniques for the transmission of (originally)

4.2 Matched Filter

4.2 Analysis of Matched Filter

go (t) H ( f )G( f ) exp( j2ft)df

| g (T ) |2 H ( f )G( f ) exp( j2fT )df

With w(t) being white with PSD N 0 / 2,

G( f )H ( f ) exp( j2fT )df

4.2 Analysis of Matched Filter

H( f ) k G*( f )exp( j2fT )

4.2 Properties of Matched Filter

Example 4.1 Matched Filter for Rectangular Pulse

s(t) I g(t), where I {1,1}.

Let be the set for which a decision favors 1 is made.

p [(1 p)1( y) p1( y)]dy, where p Pr[I 1].

To minimize BER, the optimal set opt y : (1 p)1( y) p1( y) 0.

4.3 Error Rate due to Noise under Uniform Input

4.3 Error Function

(The bound is good when x is large.)

4.3 Error rate due to noise

BER 1 erfc Eg 2Eg

4.4 Intersymbol Interference

bk {0,1}, ak 2bk 1 and s(t) kak g(t kTb ).

s(t) ak g(t kTb ).

since ISI is involved.

x(t) s(t) * h(t) w(t) k ak [g(t kTb ) * h(t)] w(t)

y(t) x(t) * c(t) k ak [g(t kTb ) * h(t) * c(t)] w(t) * c(t)

g(t kTb ) * h(t) * c(t) G( f ) exp{ j2fkTb} H ( f ) C( f ) exp{ j2ft}dt

y(iTb ) ak p(iTb kTb ) n(iTb ) k

4.4 ISI and Noise

H ( f ) 1 p(t) G( f )C( f )exp{ j2ft}dt.

y((i 1)Tb ) ak p((i 1 k)Tb ) n((i 1)Tb ) ai p(Tb ) n((i 1)Tb )

As a result, without ISI and additive noise,

y((i 1)Tb ) ak p((ik

4.5 Nyquists Criterion for Distortionless

o Also from Slide 3-4, we have:

P ( f ) np(nTb ) exp j2nTb f 1

4.5 Nyquists Criterion for Distortionless

o This is named the Nyquist criterion.

4.5 Infeasibility of Ideal Nyquist Channel

y(t) k ak p(t kTb )

4.5 Raised Cosine Spectrum

4.5 Raised Cosine Spectrum

o The special case of = 1 is known as the full-cosine rolloff

4.5 Raised Cosine Spectrum

BT ,( Baseband) W (1 ) 772(1 )kHz

4.6 Correlative-Level Coding

4.6 Correlative-Level Coding

where {bk} i.i.d.

4.6 Duobinary Signaling

HDuoB( f ) 1 exp( j2fTb )

n As shown in the next slide, the response HI(f) is realizable.

4.6 Duobinary Signaling

Text omits this term

4.6 Duobinary Signaling

The input to this filter may not be WSS!

4.6 Time-Average Autocorrelation Function

X (t) H(f) Y (t)

Time Average Autocorrelation Function

4.6 Duobinary Signaling

X (t) ak (t kTb ) Y (t) k ak g(t kTb )

4.6 Duobinary Signaling

X (t) ak (t kTb ) Y (t)