Chap 03

Digital Communications
Chapter 3: Digital Modulation Schemes

Po-Ning Chen, Professor
Institute of Communications Engineering

National Chiao-Tung University, Taiwan
Digital Communications: Chapter 3
Ver. 2015.03.09
Po-Ning Chen
1 / 160
3.1 Representation of digitally

modulated signals
Ver. 2015.03.09
Po-Ning Chen
2 / 160
Note that the channel symbols are bandpass signals.
Ver. 2015.03.09
Po-Ning Chen
3 / 160
Memoryless modulation: sm` (t), m` {1, 2, . . . , M},

m` =function of Block`
L = Constraint length of mod-
Modulation with memory: sm` (t), ulation (with memory)

m` =function of (Block` , Block`1 , , Block`(L1) )
Ver. 2015.03.09
Po-Ning Chen
4 / 160
Terminology
Signal sm (t), 1 m M, t [0, Ts )
Signaling interval: Ts (For convenience, we will use T
instead sometimes.)
Signaling rate (or symbol rate): Rs = T1s
(Equivalent) Bit interval: Tb =
Ts
log2 M
(Eqiuvalent) Bit rate: Rb = T1b = Rs log2 M

Average signal energy (assume equal-probable in message
m)
Ts
1 M
2
E avg =
sm (t) dt
M m=1 0
(Equivalent) Average bit energy: Ebavg =
Average power: Pavg =
Ver. 2015.03.09
Eavg
Ts
= Rs Eavg =
Eavg
log2 M
Ebavg
Tb
Po-Ning Chen
= Rb Ebavg
5 / 160
3.2 Memoryless modulation

methods
Ver. 2015.03.09
Po-Ning Chen
6 / 160
Example studies of memoryless modulation

Digital pulse amplitude modulated (PAM) signals
(Amplitude-shift keying or ASK)
Digital phase-modulated (PM) signals (Phase shift keying
or PSK)
Quadrature amplitude modulated (QAM) signals
Multidimensional modulated signals
Orthogonal
Bi-orthogonal
Simplex signals
Ver. 2015.03.09
Po-Ning Chen
7 / 160
M-ary pulse amplitude modulation (M-PAM)

PAM bandpass waveform
sm (t) = Re {Am g (t)e 2fc t } = Am g (t) cos (2fc t) , t [0, Ts ),
where Am = (2m 1 M)d, and m = 1, 2, , M
Example 1 (M=4)
The amplitude
s1 (t)
s2 (t)
s3 (t)
s4 (t)
=
=
=
=
3d g (t) cos (2fc t)

d g (t) cos (2fc t)
+d g (t) cos (2fc t)
+3d g (t) cos (2fc t)
difference between two adjacent signals = 2d.
Ver. 2015.03.09
Po-Ning Chen
8 / 160
sm (t) = Re {Am g (t)e 2fc t } = Am g (t) cos (2fc t) , t [0, Ts )

g (t) is the pulse shaping function.
Ts is usually assumed to be a multiple of
Ver. 2015.03.09
Po-Ning Chen
1
fc
in principle.
9 / 160
Vectorization of M-PAM signals (Gram-Schmidt)

g (t)
g (t)
1 (t) =
2 cos (2fc t) =
2 cos (2fc t)
g (t)
Eg
Am
g (t)] , a one-dimensional vector
sm = [
2
By the correct Gram-Schmidt procedure,

1 (t) =
g (t) cos(2fc t)
g (t) cos(2fc t)
g (t) cos(2fc t)
g (t) cos(2fc t)
=
1
g (t) T cos(2fc t)
g
(t)
1/2
s
The idea behind the above derivation is to single out g (t) in the
expression! This justifies the necessity of introducing the lowpass
equivalent signal where the influence of fc has been relaxed.
For a time-limited signal, we can only claim Ex` 2Ex !

Ver. 2015.03.09
Po-Ning Chen
10 / 160
1 (t)
Ts
g 2 (t) cos2 (2fc t) dt

2
g (t) 0
Ts
2
1 + cos (4fc t)
=
g 2 (t) [
] dt
2
2
g (t) 0
Ts
1
=
g 2 (t) dt
2
g (t) 0
Ts
1
g 2 (t) cos (4fc t) dt
+
2
0
g (t)
Ts
1
g 2 (t) dt = 1
2
g (t) 0
=
If g (t) is constant for t [0, Ts ) and Ts is a multiple of

then the above becomes =.
Ver. 2015.03.09
Po-Ning Chen
1
fc ,
11 / 160
Based on the pseudo-vectorization,

Transmission energy of M-PAM signals
Ts
Em =
0
A2 g (t)
sm (t) dt m
2
2
1 2
A Eg
2 m
Error consideration
The most possible error is the erroneous selection of an
adjacent amplitude to the transmitted signal amplitude.
Therefore, the mapping (from bit pattern to channel
symbol) is assigned such that the adjacent signal
amplitudes differ by exactly one bit. (Gray encoding)
In such a way, the most possible bit error pattern
(caused by the noise) is a single bit error.
Ver. 2015.03.09
Po-Ning Chen
12 / 160
Gray code (Signal space diagram : one dimension)
Euclidean distance
Am g (t) An g (t)
2
2
g (t)
(2m 1 M)d (2n 1 M)d
=
2
= d 2 g (t) m n
sm (t) sn (t)
Ver. 2015.03.09
Po-Ning Chen
13 / 160
Single side band (SSB) PAM

1
g (t) is real G (f ) is Hermitian symmetric.
Consequently, the previous PAM is based on the DSB

transmission which requires twice the bandwidth.
Recall
1
[g (t) + g (t)] = g+ (t)
2
where g (t) is the Hilbert transform of g (t).
F 1 {u1 (f )G (f )} =
Ver. 2015.03.09
Po-Ning Chen
14 / 160
sm,SSB (t) = Re {Am g+ (t)e 2fc t }
Re { 2 g+ (t)e 2fc t }
Re {Am g+ (t)e 2fc t }
1,SSB (t)
=
g+ (t)
g+ (t) 1Ts Re {Am e 2fc t }
Am
g+ (t)]
s m,SSB = [
2
Ts
g+ (t)
0
Ts
= 2
0
=
21,SSB (t) dt
2
Re {g+ (t)e 2fc t } dt
1 Ts
2
[g+ (t)e 2fc t + g+ (t)e 2fc t ] dt
2 0
Ts
Ts
g+ (t) dt +
=
0
g+ (t) cos [4fc t + (t)] dt

2
g+ (t) dt = g+ (t)
Ver. 2015.03.09
Po-Ning Chen
15 / 160
Am
[g (t) g (t)] e 2fc t }
2
Am
Am
=
g (t) cos (2fc t)
g (t) sin (2fc t)
2
2
sm,SSB (t) = Re {
Ver. 2015.03.09
Po-Ning Chen
16 / 160
2
1
1
1
2
g+ (t) = g (t) + g (t) = g (t)
2
2
2
2
Recall from Slide 2-22, x+ (t) = 12 (x(t) + x(t)) and Ex = 2Ex+ .
To summarize
g (t)
(t)
=
1(,DSB)
g (t) 2 cos (2fc t)
Am
s m(,DSB) =
g (t)
g+ (t)
2fc t }
1,SSB (t) = Re { g+ (t) 2e
Am
s m,SSB =
g+ (t)
2-level PAM signals are particularly named antipodal signals.

(1 signals)
Ver. 2015.03.09
Po-Ning Chen
17 / 160
Applications of PAM
Ts = T
Ver. 2015.03.09
Po-Ning Chen
18 / 160
Phase-modulation (PM)
Bandpass PM
sm (t) = Re [g (t)e 2(m1)/M e 2fc t ]
= g (t) cos (2fc t + m )
= cos (m ) g (t) cos (2fc t) sin(m ) g (t) sin(2fc t)

1
where m = 2(m 1)/M, m = 1, 2, , M

Example 2 (M=4)
s1 (t)
s2 (t)
s3 (t)
s4 (t)
=
=
=
=
g (t) cos (2fc t)

g (t) cos (2fc t + /2)
g (t) cos (2fc t + )
g (t) cos (2fc t + 3/2)
Ver. 2015.03.09
Po-Ning Chen
19 / 160
Signal space of PM signals
g (t)
1 (t) g (t) 2 cos (2fc t)
g (t)
2 (t) g (t) 2 sin (2fc t)
g (t)
g (t)
s m = [ cos(m ), sin(m )]
2
2
Ver. 2015.03.09
Po-Ning Chen
20 / 160
Transmission energy of PM Signals

T
Em =
sm2 (t) dt
2
Eg
g (t)
[cos2 (m ) + sin2 (m )] =
2
2
Advantages of PM signals : Equal energy for every

channel symbol.
Error consideration
The most possible error is the erroneous selection of an
adjacent phase of the transmitted signal.
Therefore, we assign the mapping from bit pattern to
channel symbol as the adjacent signal phase differs only
by one bit. (Gray encoding)
The most possible bit error pattern caused by the noise
is a single-bit error.
Ver. 2015.03.09
Po-Ning Chen
21 / 160
Signal space diagram of PM with Gray code
Ver. 2015.03.09
Po-Ning Chen
22 / 160
g (t)
g (t)
s m = [ cos(m ), sin(m )]
2
2
Euclidean distance
sm (t) sn (t)
g (t)
2
2
=
cos(m ) cos(n ) + sin(m ) sin(n )
2
= g (t) 1 cos(m n )
SSB?
Note that g (t)e m is not real; hence spectrum is not
symmetric and there is no SSB PM.
Ver. 2015.03.09
Po-Ning Chen
23 / 160
/4-QPSK
A variant of 4-phase PSK (QPSK), named /4-QPSK, is
obtained by introducing an additional /4 phase shift in the
carrier phase in each symbol interval.
Ver. 2015.03.09
Po-Ning Chen
24 / 160
Quadrature amplitude modulation (QAM)

Bandpass QAM
sm (t) = xi (t) cos(2fc t) xq (t) sin(2fc t)
where xi (t) and xq (t) are quadrature components. Let
xi (t) = Ami g (t) and xq (t) = Amq g (t); then bandpass QAM is
sm (t) = Ami g (t) cos(2fc t) Amq g (t) sin(2fc t)
Advantage: Transmit more digital information by using both
quadrature components as information carriers. As a result,
the digital data transfer rate is doubled.
Ver. 2015.03.09
Po-Ning Chen
25 / 160
Vectorization of QAM signals

1
g (t)
1 (t) g (t) 2 cos(2fc t)
g (t)
2 (t) g (t) 2 sin(2fc t)
Amq
Ami
s m = [ g (t) , g (t)]
2
2
Ver. 2015.03.09
Po-Ning Chen
26 / 160
Amq
Ami
s m = [ g (t) , g (t)]
2
2
Transmission energy of QAM signals
T
Em = sm2 (t) dt
0
1
1
2
2
g (t) A2mi + g (t) A2mq
=
2
2
1
2
g (t) (A2mi + A2mq )
=
2
1
=
Eg (A2mi + A2mq )
2
Euclidean Distance

Eg
2
2
sm (t) sn (t) =
Ami Ani + Amq Anq
2
Ver. 2015.03.09
Po-Ning Chen
27 / 160
Signal space diagram for rectangular QAM
Ver. 2015.03.09
Po-Ning Chen
28 / 160
Amq
Ami
s m = [ g (t) , g (t)] ,
2
2
where Ami , Amq {(2m 1 M) m = 1, 2, , M}

Minimum Euclidean distance (of square QAM)
Eg
Ami Ani 2 + Amq Anq 2 = 2Eg
min
mn
2

=4
=0
Average symbol energy (of square QAM)
Eavg
Eg 2M(M 1) M 1
1 Eg M M 2
=
=
Eg
(Ami + A2nq ) =
M 2 m=1 n=1
2M
3
3
Average bit energy (of square QAM)

M 1
Ebavg =
Eg
3 log2 M
Ver. 2015.03.09
Po-Ning Chen
29 / 160
Example of applications of square QAM

CCITT V.22 modem
Serial binary, asynchronous or synchronous, full duplex,
dial-up
2400 bps or 600 baud (symbols/sec)
QAM, 16-point rectangular-type signal constellation
Ver. 2015.03.09
Po-Ning Chen
30 / 160
Alternative viewpoint of QAM

QAM = PM (PSK) + PAM (ASK)
Use both amplitude and phase as digital information
bearers.
sm (t) = Re [Vm1 e m2 g (t)e 2fc t ] = Vm1 g (t) cos (2fc t + m2 )
Compare with the previous viewpoint
= Vm1 g (t) cos (2fc t + m2 )
where Vm1 = A2mi + A2mq and m2 = tan1 (Amq /Ami )

There is a one-to-one correspondence mapping from
(Ami , Amq ) domain to (Vm1 , m2 ) domain
Ver. 2015.03.09
Po-Ning Chen
31 / 160
Signal space for non-rectangular QAM (AM-PSK)
Ver. 2015.03.09
Po-Ning Chen
32 / 160
Multi-dimensional signals
PAM : one-dimensional
PM : two-dimensional
QAM : two-dimensional
How to create three or higher dimensional signal?
1
Subdivision of time
Example. N time slots can be used to form 2N vector
basis elements (each has two quadrature bearers.)
Subdivision of frequency
Example. N frequency subbands can be used to form 2N
vector basis elements (each has two quadrature bearers.)
Subdivision of both time and frequency
Ver. 2015.03.09
Po-Ning Chen
33 / 160
Frequency shift keying or FSK

Subdivision of frequency
Bandpass orthogonal multidimensional signals (Frequency
shift keying or FSK)
2E 2(mf )t 2fc t
sm (t) = Re
e
e
2E
cos (2fc t + 2(mf )t)
=
T
Vectorization of FSK signals under orthogonality
conditions (introduced in next few slides)
1
m (t) = sm (t) and s m = [0, . . . , 0, E , 0, . . . , 0]T
E
m-th
position
Ver. 2015.03.09
Po-Ning Chen
34 / 160
Crosscorrelations of FSK signals
2E 2(mf )t
e
T
sm,` (t) =
mn,` =
and
sm,` (t) =
2E
T
sm,` (t), sn,` (t)
1
= e 2(mn)f t dt
sm,` (t) sn,` (t) T 0
= sinc [T (m n)f ] e T (mn)f

sm (t), sn (t)
sin (T (m n)f )
= Re{mn,` } =
cos (T (m n)f )
sm (t) sn (t)
T (m n)f
= sinc(2T (m n)f )
k
When f = 2T
, Re{mn,` } = 0 for m n. In other words, the
minimum frequency separation between adjacent (bandpass)
1
signals for orthogonality is f = 2T
.
Ver. 2015.03.09
Po-Ning Chen
35 / 160
Transmission energy of FSK signals

T
Em =
sm (t)2 dt = E
Equal transmission power for each channel symbol.

Signal space diagram for FSK
Ver. 2015.03.09
Po-Ning Chen
36 / 160
Euclidean distance between FSK signals
Equal distance between signals
[s 1 s 2
E 0
0
E
s M ] =
0
0
s m s n =
Ver. 2015.03.09
0
0

2E
Po-Ning Chen
37 / 160
Biorthogonal multidimensional FSK signals
Ver. 2015.03.09
Po-Ning Chen
38 / 160
Transmission energy for biorthogonal FSK signals

T
Em =
sm (t)2 dt = E
Still, equal transmission power for each channel symbol.

Cross-correlation of baseband biorthogonal FSK signals
2E 2m(f ) t
sm,` (t) = sgn(m)
e
, m = 1, 2, , M/2
T
1, m = n
mn,` = 1, m = n
0, otherwise
Ver. 2015.03.09
Po-Ning Chen
39 / 160
Euclidean distance between signals
[s 1 s M/2 s 1
E
0
0
E
=
0
0
Hence
s M/2 ]
2E
s m s n =
2 E
Ver. 2015.03.09
E 0
0
E
0
0
if m n
if m = n
Po-Ning Chen
40 / 160
Simplex signals
Given the vector representations of orthogonal and
equal-power channel symbols (such as FSK)
s m = [am1 , am2 , , amk ]
for m = 1, 2, , M, its center is
c = [
1 M
1 M
1 M
am1 ,
am2 , ,
amk ]
M m=1
M m=1
M m=1
Define new channel symbol as

s m = s m c
Then {s 1 , s 2 , , s M } is called the simplex signal.
Ver. 2015.03.09
Po-Ning Chen
41 / 160
Transmission energy of simplex signals

T
Em = sm (t) dt
0
2
= s m c
= s m + c s m , c c, s m
= s m + c
2
= s m +
= (1
(c =
1 M
si )
M i=1
1 M
1 M
s m , s i s i , s m
M i=1
M i=1
1
2
2
2
s m s m
M
M
by orthogonality
(
)
and equal-power
1
2
) s m
M
The transmission energy of a signal is reduced by

simplexing it.
Ver. 2015.03.09
Po-Ning Chen
42 / 160
Crosscorrelation of simplex signals
mn =
s m , s n
s m c, s n c
=
s m s n
(1 M1 ) s m 2
s m , s n s m , c c, s n + c, c
=
(1 M1 ) s m 2
2
1
s m 2 M
s m 2 + M
s m 2
m=n
(1 M
)s m 2
= 2
1
0 M s m 2 + M
s m 2
mn
1
(1 M )s m 2
1
m=n
= {
1
M1 m n
Simplex signals are equally correlated !

Ver. 2015.03.09
Po-Ning Chen
43 / 160
Example of simplex signals
[s 1
E 0
0
E
sM] =
0
0
[s 1
(1 1 )E
1
M
M E
1
1
E
(1
)
E
M
M
s M ] =
M1 E
M1 E
Ver. 2015.03.09
Po-Ning Chen
M1 E
M1 E
(1 M1 ) E
44 / 160
Subdivision of time: N time slots

For example: BPSK in each dimension
s m = [cm,0 , cm,1 , , cm,N1 ] ,
1mM
where NTc = T
cm,j = 0 represents g1 (t) is transmitted at time slot j
cm,j = 1 represents g2 (t) is transmitted at time slot j
2Ec
2E
g1 (t) = +
cos(2fc t), g2 (t) =
cos(2fc t),
Tc
T
with t [0, Tc )
2Ec N1
sm (t) =
(1)cm,j cos (2fc (t jTc )) 1{jTc t < (j+1)Tc }
Tc j=0
Ver. 2015.03.09
Po-Ning Chen
45 / 160
Crosscorrelation coefficient of adjacent signals

(i.e., with only one distinct component)
For those identical components
Tc
Tc
g1 (t)2 dt =
g2 (t)2 dt = Ec
For the single distinct component

Tc
g1 (t)g2 (t) dt =
Tc
0
g1 (t)2 dt = Ec
Hence
mn =
s m , s n
(N 1)Ec Ec
2
=
=1
s m s n
NEc
N
Minimum Euclidean distance between adjacent codewords
2
2
min s m s n = min s m + s n s m , s n s n , s m
mn
mn
N 2
=
NEc + NEc 2(NEc )
= 2 Ec
N
Ver. 2015.03.09
Po-Ning Chen
46 / 160
Transmission energy of multidimensional BPSK signals

T
Em =
Tc
sm (t) dt = N g1 (t) = N
g1 (t)2 dt = NEc
Largest number of channel symbols

M 2N
Vectorization of BPSK signals
Ec

E
sm =

Ec
N1
Can we properly choose {s m }M
m=1 such that they are
orthogonal to each other ?
Ver. 2015.03.09
Po-Ning Chen
47 / 160
Orthogonal multidimensional signals: Hadamard

signals
Definition: The Hadamard signals of size M = 2n can be
recursively defined as
H
Hn1
Hn = [ n1
]
Hn1 Hn1
with initial value H0 = [1].
For example,
1 1
H1 = [
] and H2 =
1 -1
Ver. 2015.03.09
1 1 1 1
1 -1 1 -1
1 1 -1 -1
1 -1 -1 1
Po-Ning Chen
48 / 160
Hence, when M = 4, the Hadamard multidimensional

orthogonal (BPSK) signals are
Ec
Ec
Ec
Ec
E E
Ec Ec
[s 1 s 2 s 3 s 4 ] = c c
Ec
E
Ec Ec
c
Ec Ec Ec Ec
Ver. 2015.03.09
Po-Ning Chen
49 / 160
3.3 Signaling schemes with memory
Ver. 2015.03.09
Po-Ning Chen
50 / 160
Memoryless modulation: smi (t), mi {1, 2, . . . , M},

mi =function of Blocki
Modulation with memory: smi (t),
mi =function of (Blocki , Blocki1 , , Blocki(L1) )
Linear modulation: The modulated part of smi (t) is a
linear function of the digital waveform.
Linearity = Principle of superposition
If a1 b1 and a2 b2 , then a1 + a2 b1 + b2 .
Non-linear modulation:
Ver. 2015.03.09
Po-Ning Chen
51 / 160
Why introducing memory into signals?

The signal dependence is introduced for the purpose of
shaping the spectrum of transmitted signal so that it
matches the spectral characteristics of the channel.
Linearity
For example, smi (t) = Re {Ami e 2fc t }.
+1
+3
Re {3e 2fc t }
Re {1e 2fc t }
Re {+1e 2fc t }
Re {+3e 2fc t }
So, if the modulated part of smi (t) cannot be made as a

linear function of the digital waveform, the modulation is
classified as non-linear.
Ver. 2015.03.09
Po-Ning Chen
52 / 160
Ver. 2015.03.09
Po-Ning Chen
53 / 160
Linear modulations with/without memory

NRZ (Non-return-to-zero) =Binary PAM or binary PSK :
memoryless
channel code bit = input bit
NRZI (Non-return-to-zero, inverted) =Differential
encoding : with memory
(channel code bit)k = (input bit)k (channel code bit)k1
(channel code bit)k = (channel code bit)k1 , when (input bit)k = 0
(channel code bit)k = (channel code bit)k1 , when (input bit)k = 1

Ver. 2015.03.09
Po-Ning Chen
54 / 160
Application: DBPSK/DQPSK in Wireless LAN
Ver. 2015.03.09
Po-Ning Chen
55 / 160
Advantage of modulation with memory

Why adding differential encoding before BPSK ?
For PSK modulations, digital information is carried by
absolute phase
Synchronization is often achieved by either adding a
small pilot signal or using some self-synchronization
scheme.
The demodulator needs to detect the phase, which may
have a phase ambiguity due to noise and other
constraints.
Ver. 2015.03.09
Po-Ning Chen
56 / 160
Example of phase ambiguity (frequency shift)

Ideal (noiseless) case
{
ftransmiter = fc receive cos (2fc t + )

freceiver = fc estimate it based on fc
estimate =
Ambiguous case
{
ftransmiter = fc receive cos (2fc t + )

freceiver fc estimate it based on fc
receive cos (2fc t + [2(fc fc )t] + )

estimate it based on fc
estimate = [2(fc fc )t] +
Ver. 2015.03.09
Po-Ning Chen
57 / 160
Advantage of differential encoding
(channel code bit)k = (input bit)k (channel code bit)k1

The phases or signs of the received waveforms are not
important for detection.
What is important is the change in the sign of successive
pulses.
The sign change can be detected even if the
demodulating carrier has a phase ambiguity.
Ver. 2015.03.09
Po-Ning Chen
58 / 160
Advantage of diff encode (Noncoherent demod.)

No need to generate a local carrier (at the receiver side).
Use the received signal itself as a carrier.
A cos(2fc t)
y (t)
-
6
Delay
T0
Lowpass
fileter
z(t)-
y (t T0 )
2 cos2 (2f t) = Ac + 1 cos (4f t) 1 A2 ,
c
c
c
2
2
2 c
if
y (t) = y (t T0 )
z(t) =
2 cos2 (2f t) = Ac 1 cos (4f t) 1 A2 ,
A
c
c
c
2
2
2 c
if
y
(t)
= y (t T0 )
Ver. 2015.03.09
Po-Ning Chen
59 / 160
Nonlinear modulation methods with

memory
Ver. 2015.03.09
Po-Ning Chen
60 / 160
Linear modulation: The modulated part of smi (t) is a linear function of the
digital waveform.
Linearity = Principle of superposition
If a1 b1 and a2 b2 , then a1 + a2 b1 + b2 .
Non-linear modulation: The modulated part of smi (t) cannot be made as a
linear function of the digital waveform.
(Linear) Frequency shift keying or FSK
2E 2(mf )t 2fc t
sm (t) = Re
e
e
where m = 1, 2, , (M 1)
Ver. 2015.03.09
Po-Ning Chen
61 / 160
Motivation: Disadvantages of FSK

Potential obstacles of multidimensional FSK with (M 1)
oscillators for each desired frequency
Abrupt switching from one oscillator to another will
result in relatively large spectral side lobes outside of the
main spectral band of the signal.
Continuous-Phase FSK (CPFSK)

Alternative implementation of multidimensional FSK
A single carrier whose frequency is changed continuously.
This is considered as a modulated signal with memory
(we will explain this point in the next few slides).
Ver. 2015.03.09
Po-Ning Chen
62 / 160
Recall
s(t) = Re {s` (t)e 2fc t } ,
s` (t) = xi (t) + xq (t)
s` (t) is the baseband version of the bandpass signal s(t).

For ideal FSK signals
2E 2(mf )t 2fc t
sm (t) = Re
e
e
2E 2(mf )t
e
sm,` (t) =
T
where f = fd and m = 1, 2, , (M 1).
Ver. 2015.03.09
Po-Ning Chen
63 / 160
Example of ideal (2-OSC) FSK signals

Let T = 0.5 sec, E = 0.25, fd = 0.5, In (= m) {1, 1}, and
fc = 1.5 Hz.
s(t) = Re {s` (t)e 2fc t } = {
Ver. 2015.03.09
cos(4t) In = 1
cos(2t), In = 1
Po-Ning Chen
64 / 160
Discontinuous phase of (2-OSC) FSK
Phase of s` (t) = {
t, In = 1
for t [nT , (n + 1)T )
t, In = 1
Ver. 2015.03.09
Po-Ning Chen
65 / 160
Phase change of (2-OSC) FSK

(Normalized) phase change (for t [nT , (n + 1)T ))
d(t) =
phase of s` (t)
=
4Tfd
t (2In fd t)
4Tfd
In
2T
is the derivative of the phase!

Continue from the previous example with T = 0.5.
d(t) = (+1)[u1 (t) u1 (t T )] (t T )

+ (1)[u1 (t T ) u1 (t 2T )]
+ (1)[u1 (t 2T ) u1 (t 3T )] + 3(t 3T )
+ (+1)[u1 (t 3T ) u1 (t 4T )] +
Ver. 2015.03.09
Po-Ning Chen
66 / 160
d(t) = I0 [u1 (t
) u1 (t T )] + 1{I0 I1 } I1 1 (t T )
+ I1 [u1 (t T ) u1 (t 2T )] + 1{I1 I2 } I2 2 (t 2T )
+ I2 [u1 (t 2T ) u1 (t 3T )] + 1{I2 I3 } I3 3 (t 3T )
+ I3 [u1 (t 3T ) u1 (t 4T )] + 1{I3 I4 } I4 4 (t 4T )
+
Phase change is the derivative of the phase!
Phase is the integration of phase change!
2E 4Tfd t d( )d
e
s` (t) =
T
Those () functions result in discontinuity in
integration! Hence, let us remove them to force
continuity in phase.
Ver. 2015.03.09
Po-Ning Chen
67 / 160
Continuous phase FSK (CPFSK)
s` (t) =
2E 4Tfd t d( )d
e
T
where
d(t) = In g (t nT ) and g (t) =

n=
1
[u1 (t) u1 (t T )] .
2T
In {1, 3, 5, } is the PAM information sequence.

g (t) is the phase shaping function.
It is now chosen as a rectangular pulse of height 1/(2T )
and duration [0, T ) (hence, the area is 1/2.)
T is the symbol duration.

Ver. 2015.03.09
Po-Ning Chen
68 / 160
Re-express s` (t) as
s` (t) =
2E (t;I )
e
T
where
(t; I )
t
= 4Tfd
= 4Tfd
d( ) d
[ In g ( nT )] d
n=
n1
= 4fd T [ Ik (T
k=
n1
1
t nT
) + In
]
2T
2T
= 2fd T Ik + 2fd (t nT )In
for t [nT , (n + 1)T )
for t [nT , (n + 1)T )
k=
Ver. 2015.03.09
Po-Ning Chen
69 / 160

(Cont.)For t [nT , (n + 1)T ), s` (t) =
2E (t;I )
T e
with
n1
(t; I ) = 2fd T Ik + 2fd (t nT )In

k=
= n + 2h In q(t nT ),
where
h = 2fd T (modulation index)
n1
=
h
Ik (accumulation of history/memory)
n
k=
0 t <0
q(t) = 2T 0 t < T
(integration of g (t))
t T
2
Ver. 2015.03.09
Po-Ning Chen
70 / 160
Generalization of CPFSK: CPM

We can further generalize (t; I ) to
n
(t; I ) = 2 hk Ik q(t kT )
k=
for nT t < (n + 1)T

where
1
I = {Ik }
k= is the sequence of PAM symbols in
{1, 3, . . . , (M 1)}.
2
hk is the modulation index.
If hk varies with k, it is called multi-h CPM.
t
3
q(t) = 0 g ( ) d.
If g (t) = 0 for t T (and t < 0), s` (t) is called full-response CPM;
otherwise it is called partial-response CPM.
Ver. 2015.03.09
Po-Ning Chen
71 / 160
Examples of CPMs
Ver. 2015.03.09
Po-Ning Chen
72 / 160
Examples of CPMs
Ver. 2015.03.09
Po-Ning Chen
73 / 160
Some commonly used CPM pulse shapes
LREC (Rectangular): LREC with L = 1 is CPFSK

g (t) =
1
(u1 (t) u1 (t LT ))
2LT
LRC (Raised cosine)

g (t) =
1
2t
(u1 (t) u1 (t LT )) (1 cos (
))
2LT
LT
Ver. 2015.03.09
Po-Ning Chen
74 / 160
Some commonly used CPM pulse shapes

GMSK (Gaussian minimum shift keying)
g (t) = Q (2B (t
where Q(t) = t
T
T
) / ln 2)Q (2B (t + ) / ln 2)
2
2
2
1 e x /2 dx,
2
and B is 3dB Bandwidth
GMSK with BT = 0.3 is used in the European digital

cellular communication system, called GSM (2G).
At BT = 0.3, the GMSK pulse may be truncated at
t = 1.5T with a relatively small error incurred for
t > 1.5T .
Ver. 2015.03.09
Po-Ning Chen
75 / 160
Representations of continuous-phase
Phase trajectory or phase tree

Phase trellis
Ver. 2015.03.09
Po-Ning Chen
76 / 160
Phase trajectory or phase tree

Binary CPFSK (i.e., In = 1 and g (t) is full response
rectangular function)
n1
(t; I ) = h Ik + 2hIn q(t nT )

k=
Ver. 2015.03.09
Po-Ning Chen
77 / 160
Example 3
Quaternary CPFSK (See the next page) with
In {3, 1, +1, +3}.
We observe that the phase trees for CPFSK are piecewise
linear as a consequence of the fact that the pulse g (t) is
rectangular.
Smoother phase trajectories and phase trees are obtained
by using pulses that do not contain discontinuities.
Ver. 2015.03.09
Po-Ning Chen
78 / 160
Ver. 2015.03.09
Po-Ning Chen
79 / 160
If g (t) is continuous (especially at boundaries), phase

trajectory becomes smooth.
Example 4
g (t) =
1
2t
(1 cos (
)) = raised cosine of length 3T
6T
3T
with (I2 , I1 , I0 , I1 , I2 , ) = (+1, +1, +1, 1, 1, 1, +1, +1, 1, +1, )
Ver. 2015.03.09
Po-Ning Chen
80 / 160
Phase trellis
Phase trellis = Phase trajectory is plotted with modulo 2
Example 5
Binary CPFSK with h = 1/2 and g (t) is a full response
rectangular function.
Thus CPM can be decoded by Viterbi trellis decoding.

Ver. 2015.03.09
Po-Ning Chen
81 / 160
Minimum shift keying (MSK)

Recall for nT t < (n + 1)T , CPM has
n
(t; I ) = 2 hk Ik q(t kT ).
k=
CPFSK is a special case of CPM with

g (t) =
1
2T
for 0 t < T
MSK is a special case of binary CPFSK with

hk = 12 , g (t) =
1
2T
for 0 t < T and In {1}
Thus for MSK, we have for nT t < (n + 1)T ,

(t; I ) =
n1
1
t nT
)
Ik + In q(t nT ) = n + In (
2 k=
2
T
Ver. 2015.03.09
Po-Ning Chen
82 / 160
1
t nT
In
nIn
(t; I ) = n + In (
) = 2 ( ) t
+ n
2
T
4T
2
The corresponding modulated carrier wave is
sMSK (t) = A cos (2fc t + (t; I ))
nIn
In
)t
+ n ]
= A cos [2 (fc +
4T
2
Since In {1}, sMSK (t) has two frequency components:

1
4T
1
= fc +
4T
f1 = fc
f2
Ver. 2015.03.09
Po-Ning Chen
83 / 160
Minimum shift keying (MSK)
1
MSK is so named because f2 f1 = 2T
= the minimum
(frequency) shift that makes the two frequency components
orthogonal.
k
, Re{mn,` } = 0 for m n. In
[See Slide 35] When f = 2T
other words, the minimum frequency separation between
1
adjacent signals for orthogonality is f = 2T
.
MSK is sometimes regarded as a kind of OQPSK (Offset

QPSK). Why?
Ver. 2015.03.09
Po-Ning Chen
84 / 160
Offset QPSK
The original QPSK
There could be 180 degree of (sudden) phase change (so, not

continuous phase), e.g., from (+1, +1) to (1, 1).
Ver. 2015.03.09
Po-Ning Chen
85 / 160
sQPSK (t) =
I2n g (t 2nT ) cos(2fc )

n=
I2n+1 g (t 2nT ) sin (2fc t)

n=
(I0 , I1 ) = (+1, +1), (I2 , I3 ) = (1, 1) and (I4 , I5 ) = (1, +1).

g (t) rectangular pulse of unit height and during 2T .
Ver. 2015.03.09
Po-Ning Chen
86 / 160
Offset QPSK (OQPSK)

How to reduce the 180o phase change to only 90o ?
Simple solution: Do not let the two bits I2n and I2n+1 change
at the same time!
I1 = 1
I3 = 1
I5 = 1
3T
5T
I0 = 1
I2 = 1
I4 = 1
2T
4T
7T
6T
sOQPSK (t) =
I2n g (t 2nT ) cos(2fc t)

n=
I2n+1 g (t (2n + 1)T ) sin (2fc t)

n=
Ver. 2015.03.09
Po-Ning Chen
87 / 160
To synchronize with the textbook, we reverse {I2n+1 } to obtain
sOQPSK (t) =
I2n g (t 2nT ) cos(2fc t)

n=
+ I2n+1 g (t (2n + 1)T ) sin (2fc t)

n=
Ver. 2015.03.09
Po-Ning Chen
88 / 160
OQPSK vs. MSK

MSK can be regarded as a kind of (memoryless) OQPSK.
Why?
Proof: Suppose without loss of generality,
0 =
3
1
.
Ik =
2 k=
2
Then for nT t < (n + 1)T ,

sMSK,` (t) = e (t;I )
= e 2( 4T )t e 2 In e 2 k=0 Ik e 0
n1
t
t
n
= [cos (
) + In sin (
)] (In ) ( (Ik )) ( )
2T
2T
k=0
In
n1
= Inn+1 ( Ik ) sin (
k=0
n1
n1
t
(t T )
) + Inn ( Ik ) sin (
)
2T
2T
k=0
Ver. 2015.03.09
Po-Ning Chen
89 / 160
n
0
1
2
3
4
5
6
Inn+1 (n1
k=0 Ik )
J0 = I0 = J20/2
J0 = I0 = J21/2
J2 = I0 I1 I2 = J22/2
J2 = I0 I1 I2 = J23/2
J4 = I0 I1 I2 I3 I4 = J24/2
J4 = I0 I1 I2 I3 I4 = J25/2
J6 = I0 I1 I2 I3 I4 I5 I6 = J26/2
Inn (n1
k=0 Ik )
J1 = I0 I1 = J2(11)/2+1
J1 = I0 I1 = J2(21)/2+1
J3 = I0 I1 I2 I3 = J2(31)/2+1
J3 = I0 I1 I2 I3 = J2(41)/2+1
J5 = I0 I1 I2 I3 I4 I5 = J2(51)/2+1
J5 = I0 I1 I2 I3 I4 I5 = J2(61)/2+1
For nT t < (n + 1)T ,

(t 2n/2T )
)
2T

sMSK,` (t) = J2n/2 (1)n/2 sin (
g (t2n/2T )
(t 2(n 1)/2T T )
)
2T

J2(n1)/2+1 (1)(n1)/2+1 sin (
g (t2(n1)/2T T )
Ver. 2015.03.09
Po-Ning Chen
90 / 160
For 2mT t < (2m + 1)T ,

sMSK,` (t) = J2m (1)m g (t 2mT ) J2m1 (1)m g (t (2m 1)T )

(1)(2m1)/2
For (2m + 1)T t < (2m + 2)T ,

sMSK,` (t) = J2m (1)m g (t 2mT ) J2m+1(1)m+1g (t (2m + 1)T )

(1)(2m+1)/2
For (2m + 2)T t < (2m + 3)T ,

sMSK,` (t) = J2(m+1) (1)m+1 g (t 2(m + 1)T )
J2m+1(1)m+1g (t (2m + 1)T )

(1)(2m+1)/2
t
) [u1 (t) u1 (t 2T )] .
with g (t) = sin ( 2T
Ver. 2015.03.09
Po-Ning Chen
91 / 160
MSK can be regarded as a memoryless OQPSK by setting
sMSK (t) = [ I2n g (t 2nT )] cos(2fc t)

n=
+ [ I2n+1 g (t (2n + 1)T )] sin (2fc t)

n=
with
In = (1)n/2 Jn = (1)n/2 Ik .
k=0
MSK can be composed using memoryless circuits

with with-memory information sequence I .
Please be noted that the textbook abuses the notation by
using g (t) to denote both the amplitude and phase pulse
shaping functions for CPM signals!
Ver. 2015.03.09
Po-Ning Chen
92 / 160
A linear representation of CPM

The key of OQPSK representation of MSK is that phase can
be pulled down as a multiplicative adjustment in amplitude
when In {1, +1}!
For example,
e 2( 4T )t = cos (
In
t
t
) + In sin (
)
2T
2T
(1986 Laurent)
CPM can also be represented as a linear superposition of
AM signal waveforms (if In {1}).
Such a representation provides an alternative method for
generating CPM signal at the transmitter and for
demodulating the signal at the receiver.
Ver. 2015.03.09
Po-Ning Chen
93 / 160
An important and useful fact

For I {1, +1},
e AI =
sin(B A)
sin(A)
+ e BI
.
sin(B)
sin(B)
Proof:
sin(B)e AI
= sin(B)[cos(A) + I sin(A)]
= sin(B) cos(A) + sin(B I ) sin(A)
= sin(B A) + cos(B) sin(A) + sin(B I ) sin(A)
= sin(B A) + sin(A)[cos(B I ) + sin(B I )]
= sin(B A) + sin(A)e BI
Ver. 2015.03.09
Po-Ning Chen
94 / 160
For general h and g () function of duration L and of integral

1/2 (but each In {1}), we have for nT t < (n + 1)T (for a
binary CPM signal),
sb-CPM,` (t) = e (t;I )
= e (h k= Ik +2h k=nL+1 Ik q(tkT ))
nL
L1
= e h k= Ik e 2hInk q(t(nk )T )
nL
k =0
L1
= e h k= Ik (
nL
k =0
+e BInk
(n k = k)
sin(B 2hq(t (n k )T ))
sin(B)
sin(2hq(t (n k )T ))
),
sin(B)
where B = h.
Ver. 2015.03.09
Po-Ning Chen
95 / 160
Define
sin(2hq(t))
sin(B)
))
s0 (t) = sin(B2hq(tLT
sin(B)
0 t < LT
LT t < 2LT
otherwise
Since q(0) = 0 and q(LT ) = 1/2, s0 (t) is continuous for t R.
Ver. 2015.03.09
Po-Ning Chen
96 / 160
Continue the derivation:

sb-CPM,` (t)
L1
= e h k= Ik (
nL
k =0
sin(B 2hq(t (n k )T + LT LT ))
sin(B)
+e BInk
sin(2hq(t (n k )T ))
)
sin(B)
L1
= e h k= Ik (s0 (t (n k )T + LT )
nL
k =0
+e BInk s0 (t (n k )T ))
nT t < (n + 1)T and 0 k L 1 imply that

0 t (n k )T < LT and LT t (n k )T + LT < 2LT .
Ver. 2015.03.09
Po-Ning Chen
97 / 160
L1
BInk
s0 (t (n k )T ))
(s0 (t (n k )T + LT ) + e
k =0
= ( s0 (t nT + 0 T + LT ) +e BIn0 s0 (t nT + 0 T ) )

ai,0 =0 (k =0)
ai,0 =1 (k =0)
( s0 (t nT + 1 T + LT ) +e BIn1 s0 (t nT + 1 T ) )

ai,1 =1 (k =1)
ai,1 =0 (k =1)
( s0 (t nT + (L 1) T + LT ) +e BIn(L1) s0 (t nT + (L 1) T ) )

ai,L1 =1 (k =L1)
2L 1
ai,L1 =0 (k =L1)
L1
(1ai,k )Ink
e k =0
s0 (t nT + k T + ai,k LT )
L1
i=0
k =0
where (ai,0 , ai,1 , . . . , ai,L1 ) is the binary representation of i with

ai,0 being the most significant bit.
Ver. 2015.03.09
Po-Ning Chen
98 / 160
Continue the derivation:

sb-CPM,` (t)
= e
B nL
k= Ik
2L 1
L1
B
(1ai,k )Ink
e k =0
s0 (t nT + k T + ai,k LT )
L1
i=0
k =0
2L 1
hAi,n
ci (t nT )
e

i=0
complex
amplitude
pulse shaping
function
where
n
L1
L1
Ai,n = Ik ai,k Ink

k=
k =0
and ci (t) = s0 (t+k T +ai,k LT ).

k =0
Binary CPM can be expressed as a weighted sum of 2L real-valued

pulses {ci (t)} where the complex amplitudes depends on the
information sequence.
Ver. 2015.03.09
Po-Ning Chen
99 / 160
Property of ci (t)
Duration: ci (t) = 0 if any of s0 (t + k T + ai,k LT ) = 0.
Hence, ci (t) 0 only possible in
[(k T ai,k LT ) + 2LT ]
max (k T ai,k LT ) t < min
0k <L
0k <L

min
k
T
t
<
LT
0k L
and a
=0
i,k
max
k
1k <L
and ai,k =1
1 for the case

of ai,k = 0 k
L for the case

of ai,k = 1 k
where we define ai,L = 0 and ai,1 = 1. So, the duration is

equal to:
(L + (
min
0k L and ai,k =0
k ) (
max
1k <L and ai,k =1
k ) )T .

kmin0
Ver. 2015.03.09
kmax1
Po-Ning Chen
100 / 160
i
0
1
2
3
4
5
6
7
ai,0 ai,1 ai,2

000
001
010
011
100
101
110
111
kmin0
0
0
0
0
1
1
2
3
L=3
kmax1 kmin0 kmax1
1
1
2
2
1
1
2
2
0
1
2
1
1
1
2
1
L + kmin0 kmax1
4
1
2
1
4
2
4
4
It can be shown that kmin0 kmax1 1, and the upper bound can
always be achieved by i = 0.
Ver. 2015.03.09
Po-Ning Chen
101 / 160

0
t <0
Example. h = 1/2 and q(t) = t/(6T ) 0 t < 3T . Then
otherwise
1/2
t) 0 t < 6T
sin ( 6T
s0 (t) =
otherwise
0
n
Ai,n = Ik ai,k Ink

k=
k =0
Ver. 2015.03.09
and ci (t) = s0 (t+k T +ai,k LT ).

k =0
Po-Ning Chen
102 / 160
ai,0 ai,1 ai,2
duration
0000
1001
2010
3011
4100
5101
6110
7111
[0,4T)
[0,T)
[0,2T)
[0,T)
[-T,3T)
[-T,T)
[-2T,2T)
[-3T,T)
ci (t)
s0 (t)s0 (t + T )s0 (t + 2T )
s0 (t)s0 (t + T )s0 (t + 5T )
s0 (t)s0 (t + 4T )s0 (t + 2T )
s0 (t)s0 (t + 4T )s0 (t + 5T )
s0 (t + 3T )s0 (t + T )s0 (t + 2T )
s0 (t + 3T )s0 (t + T )s0 (t + 5T )
s0 (t + 3T )s0 (t + 4T )s0 (t + 2T )
s0 (t + 3T )s0 (t + 4T )s0 (t + 5T )
Note that
c4 (t) = c0 (t + T )
hA
4,n = e hA0,n1
c6 (t) = c0 (t + 2T )
hA
6,n = e hA0,n2
e hAi,n
e n+1
e (n2 +hIn +hIn1 )
e (n1 +hIn )
e (n2 +hIn )
e n
e (n2 +hIn1 )
e n1
e n2
c7 (t) = c0 (t + 3T )
hA
7,n = e hA0,n3
and
c5 (t) = c2 (t + T )
hA
5,n = e hA2,n1
e
Ver. 2015.03.09
Po-Ning Chen
103 / 160
For nT t < (n + 1)T ,

7
sb-CPM,` (t) = e (t;I ) = e hAi,n ci (t nT )

i=0
= e hA0,n c0 (t nT ) + e hA1,n c1 (t nT ) + e hA2,n c2 (t nT )

+ e hA3,n c3 (t nT ) + e hA4,n c4 (t nT ) + e hA5,n c5 (t nT )
+ e hA6,n c6 (t nT ) + e hA7,n c7 (t nT )
= e hA0,n c0 (t nT ) + e hA1,n c1 (t nT ) + e hA2,n c2 (t nT )
+ e hA3,n c3 (t nT ) + e hA0,n1 c0 (t (n 1)T )
+ e hA2,n1 c2 (t (n 1)T ) + e hA0,n2 c0 (t (n 2)T )
+ e hA0,n3 c0 (t (n 3)T )
hA0,m
c0 (t mT ) + e hA1,m c1 (t mT )
[e
m=
(for m = n 3, n 2, n 1, n)
+e hA2,m c2 (t mT ) + e hA3,m c3 (t mT )]
31

2 1 hAi,m
=
e
c
(t
mT
)

i
m= i=0
Ver. 2015.03.09
Po-Ning Chen
104 / 160
So, we notice that when ai,0 = 1, ci (t) is always a shift-version

of some cj (t) with 0 j 2L1 1.
This concludes to that:
Theorem 1 (Laurent 86)
For nT t < (n + 1)T ,
L1

2 1
sb-CPM,` (t) = e hAi,m ci (t mT )
m= i=0
where
L1
Ai,n = Ik ai,k Ink

k=
and
k =1
L1
ci (t) = s0 (t) s0 (t + k T + ai,k LT )

k =1
with duration 0 t < (L kmax1 )T .

Ver. 2015.03.09
Po-Ning Chen
105 / 160
3.4 Power spectrum of digital

modulated signals
Ver. 2015.03.09
Po-Ning Chen
106 / 160
Why studying spectral characteristics?

Bandwidth limitation in a real channel.
Random process Power spectral density

PAM
CPM
Ver. 2015.03.09
Po-Ning Chen
107 / 160
Power spectra of modulated signals
The modulated waveform s(t) is deterministic given the

information sequence I , so only the information sequence
I = (. . . , I2 , I1 , I0 , I1 , I2 , . . .) is random!
For convenience, we denote the waveform at
nT t < (n + 1)T as s(t nT ; In ) if the modulation is
memoryless, and as s(t nT ; I n ) if the modulation is
with memory, where I n = (. . . , In2 , In1 , In ).
Ver. 2015.03.09
Po-Ning Chen
108 / 160
Hence, the modulated lowpass equivalent signal can be

expressed as
v ` (t) = s(t nT ; I n ).
n=
Note that v ` (t) is usually not a (wide-sense) stationary

process but a cyclostationary process.
Its spectral characteristics is then determined by the
time-averaged autocorrelation function rather than the usual
authocorrelation function for a WSS proess.
Ver. 2015.03.09
Po-Ning Chen
109 / 160
2.7.2 Cyclostationary processes
Ver. 2015.03.09
Po-Ning Chen
110 / 160
How to model a waveform source that carries digital

information?
For example,
X (t) = an g (t nT )
n=
where {an }
n= is a discrete-time random sequence, and
g (t) is a deterministic pulse shaping function.
Ver. 2015.03.09
Po-Ning Chen
111 / 160
Cyclostationary processes
Given that {an }
n= is WSS, what is the statistical property
of X (t)?
X (t) is not necessarily (strictly) stationary. Its mean
becomes periodic with period T :
n=
n=
E[X (t)] = E [ an g (t nT )] = a g (tnT ) = E [X (t+KT )]

Autocorrelation function becomes periodic with period T
RX (t1 , t2 ) = E [X (t1 )X (t2 )]
E[an am ]g (t1 nT )g (t2 mT )
n= m=
Ra (n m)g (t1 nT )g (t2 mT )
n= m=
= RX (t1 + KT , t2 + KT )
Ver. 2015.03.09
Po-Ning Chen
112 / 160
Definition 1 (Cyclostationary process)

A random process is said to be cyclostationary or periodically
stationary in the wide sense if its mean and autocorrelation
function are both periodic.
Time-average autocorrelation function

R X ( ) =
T
1
RX (t + , t) dt
T 0
Average power spectral density

S X (f ) = F {R X ( )}
Ver. 2015.03.09
Po-Ning Chen
113 / 160
3.4-1 Power spectral density of a

digitally modulated signal with
memory
Ver. 2015.03.09
Po-Ning Chen
114 / 160
E [v ` (t)] = E [In ] g (t nT ) = I g (t nT ) = E [v ` (t + T )]
n=
n=
and
Rv ` (t1 , t2 ) = E [v ` (t1 )v ` (t2 )]
E [In Im ] g (t1 nT )g (t2 mT ) = Rv ` (t1 + T , t2 + T )
n= m=
implies that v ` (t) is cyclostationary.

R v ` ( )
T
1
=
RI (n m)g (t + nT )g (t mT )dt
T 0 n= m=
T
1
=
RI (k)g (t + kT mT )g (t mT )dt
T 0 k= m=
(k = n m)
Ver. 2015.03.09
Po-Ning Chen
115 / 160
T
1
RI (k) g (t + kT mT )g (t mT )dt
T k=
m= 0
(m1)T
1
g (u + kT )g (u)du
RI (k)
T k=
m= mT
RI (k) g (u + kT )g (u)du
T k=
1
gk ( kT )
T k=
u=tmT
where
gm ( ) = RI (m)
Ver. 2015.03.09
g (u + )g (u)du.
Po-Ning Chen
116 / 160
Gm (f )
=
v =u+
gm ( )e 2f d
(RI (m)
RI (m)
g (u + )g (u)du) e 2f d
g (u + )e 2f d ) g (u)du
g (v )e 2f (v u) dv ) g (u)du
RI (m)
RI (m) (
RI (m)G (f )2
g (v )e 2fv dv ) (
S v ` (f ) = F {R v ` ( )} =
1
2 2kfT
RI (k)G (f ) e
T k=
1
SI (f )G (f )2
T
Ver. 2015.03.09
g (u)e 2fu du)
1
F {gk ( kT )}
T k=
where SI (f ) = RI (k)e 2kfT .

k=
Po-Ning Chen
117 / 160
Theorem 2
S v ` (f ) =
1
2
SI (f ) G (f )
T
The average power spectrum density of PAM signals is

determined by the pulse shape, as well as the input
information.
Ver. 2015.03.09
Po-Ning Chen
118 / 160
Example
Input information is real and mutually uncorrelated
RI (k) = {
I2 + 2I , k = 0
2I ,
k 0
Hence
SI (f ) = I2 + 2I e 2fkT = I2 +
k=
2I
k
(f )
T k=
T
and
S v ` (f ) =
2
I2
k
2
2
G (f ) + I2 (f ) G (f )
T
T k=
T
Ver. 2015.03.09
Po-Ning Chen
119 / 160
S v ` (f ) =
2
I2
k
2
2
G (f ) + I2 (f ) G (f )
T
T k=
T

continuous
discrete
Observation 1: Discrete spectrum vanishes when the

input information has zero mean, which is often desirable
for digital modulation techniques under consideration.
Observation 2: With a zero-mean input information, the
average power spectrum density is determined by G (f ).
Ver. 2015.03.09
Po-Ning Chen
120 / 160
Example 6
The average power spectrum density for rectangular pulses
g (t) = A [u1 (t) u1 (t T )]
It shows
G (f ) = AT sinc(fT )e fT G (f )2 = A2 T 2 sinc2 (fT )
Hence
S v ` (f ) =
I2
2
k
G (f )2 + I2 (f ) G (f )2
T
T k=
T
= I2 A2 T sinc2 (fT ) + 2I A2 (f )
Ver. 2015.03.09
Po-Ning Chen
121 / 160
Example 7
The average power spectrum density for raised cosine pulse
g (t) =
A
2
T
[1 + cos ( (t ))] (u1 (t) u1 (t T )) .
2
T
2
It gives
G (f ) =
1
AT
sinc(fT )
e fT
2
1 f 2T 2
Hence
S v ` (f ) =
=
I2
2
k
G (f )2 + I2 (f ) G (f )2
T
T k=
T
I2 A2 T sinc2 (fT ) 2I A2
2I A2
2I A2
1
1
+
(f
)
+
(f
)
+
(f + )
2
2
2
4(1 f T )
4
16
T
16
T
Note: limx1
Ver. 2015.03.09
Po-Ning Chen
sinc2 (x)
(1x 2 )2
1
4
122 / 160
Ver. 2015.03.09
Po-Ning Chen
123 / 160
Comparison of the previous two examples
Broader side lobe

Faster decay in tail (f 6 < f 2 )
Ver. 2015.03.09
Po-Ning Chen
124 / 160
Assume A = T = I2 = 1 and I = 0
S v ` (f )
S v ` (0)
Ver. 2015.03.09
Po-Ning Chen
125 / 160
S v ` (f )
S v ` (0)
Ver. 2015.03.09
Po-Ning Chen
126 / 160
The smoother (meaning, continuity of derivatives) the

pulse shape, the greater the bandwidth efficiency (lower
bandwidth occupancy).
The raised cosine pulse shape will result in higher
bandwidth efficiency than the rectangular pulse shape.
Ver. 2015.03.09
Po-Ning Chen
127 / 160
What if I correlated?
Example 8
In = bn + bn1
where {bn } mutually uncorrelated with zero mean and unit
variance.
Then,
2 k =0
RI (k) = 1 k = 1
0 otherwise
SI (f ) = 2 + e 2fT + e 2fT = 2 (1 + cos(2fT )) = 4 cos2 (fT )
1
4
2
2
S v ` (f ) =
G (f ) SI (f ) = G (f ) cos2 (fT )
T
T
Ver. 2015.03.09
Po-Ning Chen
128 / 160
Rectangular pulse shape with A = T = 1
S v ` (f )
S v ` (0)
Ver. 2015.03.09
Po-Ning Chen
129 / 160
Rectangular pulse shape with A = T = 1

Dependence in transmitted information (not the original
information) can improve the bandwidth efficiency.
S v ` (f )
S v ` (0)
Ver. 2015.03.09
Po-Ning Chen
130 / 160
Power spectra of CPFSK and CPM

CPM: Assume I i.i.d.
v ` (t) = e (t;I )
where
(t; I ) = 2h Ik q(t kT )
k=
Rv ` (t1 , t2 )
= E [v ` (t1 )v ` (t2 )]
= E [e (t1 ,I ) e (t2 ,I ) ]
= E [exp ( 2h Ik [q(t1 kT ) q(t2 kT )])]

k=
Ver. 2015.03.09
Po-Ning Chen
131 / 160
Rv ` (t1 , t2 )
= E [ exp ( 2hIk [q(t1 kT ) q(t2 kT )])]
k=
= E [exp ( 2hIk [q(t1 kT ) q(t2 kT )])]

k=
= [ Pn exp ( 2hn [q(t1 kT ) q(t2 kT )])]

k= nS
Rv ` ( )
T
1
= Rv ` (t + , t) dt
T 0
T
1
=
[ Pn e 2hn[q(t+ kT )q(tkT )] ] dt
T 0 k= nS
Ver. 2015.03.09
Po-Ning Chen
132 / 160
Assume 0. For mT = + mT < (m + 1)T and

0 t < T (i.e., the range of integration)
t + (m +1)T = t + T and t + (m +1L)T = t + (1L)T .
R v ` ( )
max{m+1,0}
T
1
=
[ Pn e 2hn[q(t+ kT )q(tkT )] ] dt
T 0 k=min{m+1L,1L} nS
Ver. 2015.03.09
Po-Ning Chen
133 / 160
Hermitian symmetry of R v ` ( )
It suffices to derive R v ` ( ) for 0 because R v ` ( ) = R v ` ( ).

Proof:
R v ` ( ) =
=
=
=
=
T
1
2h
k= Ik [q(t+ kT )q(tkT )] ] dt
E [e
T 0
T
1
2h
k= Ik [q(tkT )q(t+ kT )] ] dt
E [e
T 0
T +
1
E [e 2h k= Ik [q(v kT )q(v kT )] ] dv
T
(v = t + )
T
1
2h
k= Ik [q(v kT )q(v kT )] ] dv
E [e
T 0
R v ` ( )
Ver. 2015.03.09
Po-Ning Chen
134 / 160
Average PSD of CPM
S v ` (f ) = R v ` ( )e 2f d
= R v ` ( )e 2f d +
=
0
=
0
R v ` ( )e 2f d +
0
R v ` ( )e 2f d
R v ` ( )e 2f d
[R v ` ( )e 2f ] d +
= 2Re [
R v ` ( )e
Ver. 2015.03.09
2f
R v ` ( )e 2f d
d ]
Po-Ning Chen
135 / 160
R v ` ( )e 2f d
LT
=
0
LT
=
0
R v ` ( )e 2f d + R v ` ( )e 2f d
LT
R v ` ( )e 2f d +
m=L
(m+1)T
mT
R v ` ( )e 2f d
For m L, the two regions below are non-overlapping!
Ver. 2015.03.09
Po-Ning Chen
136 / 160
R v ` ( ) =
=
T m+1
1
2hn[q(t+ kT )q(tkT )]
] dt
[ Pn e
T 0 k=1L nS
0
T
1
]
( [ Pn e
T 0
k=1L nS
mL
]
[ Pn e
k=1 nS
m+1
[ Pn e 2hn[q(t+ kT )q(tkT )] ]) dt
k=m+1L nS
0
T
T 0
( [ Pn e 2hn[1/2q(tkT )] ]
k=1L nS
mL
2hn[1/20]
]
[ Pn e
k=1 nS
m+1
[ Pn e 2hn[q(t+ kT )0] ]) dt
k=m+1L nS
Ver. 2015.03.09
Po-Ning Chen
137 / 160
R v ` ( ) =
0
T
1
2hn[1/2q(tkT )]
] [I (h)]mL
( [ Pn e
T 0
k=1L nS
1
2hn[q(t+ k
[ Pn e
k =1L
nS
mL
= [I (h)]
T mT )]
]) dt
(k = k m)
( mT )
where I (h) = Pn e hn and

nS
() =
0
T
1
2hn[1/2q(tkT )]
(
]
[ Pn e
T 0
k=1L nS
1
2hn[q(t+k
[ Pn e
T )]
k =1L nS
Ver. 2015.03.09
Po-Ning Chen
]) dt
138 / 160
m=L
(m+1)T
mT
m=L
m=L
R v ` ( )e 2f d
(m+1)T
mT
T
0
[I (h)]mL ( mT )e 2f d
[I (h)]mL ()e 2f (+mT ) d
= ( [I (h)]mL e 2fmT ) (
( = mT )
()e 2f d)
m=L
2fLT
T
( 1e (h)e 2fT ) (0 ()e 2f d) if I (h) < 1
2T (f /T )m
= (e 2fLT
) (0T ()e 2f d)
m =0 e
if I (h) = e 2 = 1
2fLT
T
( 1e (h)e 2fT ) (0 ()e 2f d) if I (h) < 1
1
+m
= e 2fLT ( 21 + 2T
m = (f T ) cot (T (f T )))
T
2f
d)
if I (h) = e 2 = 1
(0 ()e
Ver. 2015.03.09
Po-Ning Chen
139 / 160
We finally obtain a numerically computable formula for the average

PSD of CPM. For example, if I (h) < 1,
LT
S v ` (f ) = 2 Re [
+(
R v ` ( )e 2f d
T
1
)
(
()e 2f (+LT ) d)]
1 I (h)e 2fT
0
where for 0 = + mT < LT ,

R v ` ( ) =
T m+1
1
] dt
[ Pn e
T 0 k=1L nS
However, if I (h) = 1, the average PSD of CPM signals has
for integer m .
impulses at fm = +m
T
Ver. 2015.03.09
Po-Ning Chen
140 / 160
Numerically plotted average PSD of the equivalent lowpass

CPFSK signal (M = 2, T = 0.5 and Pn uniform over S = {1})
S v ` (f )
Ver. 2015.03.09
Po-Ning Chen
141 / 160

S v ` (f )
Ver. 2015.03.09
Po-Ning Chen
142 / 160

S v ` (f )
Ver. 2015.03.09
Po-Ning Chen
143 / 160

S v ` (f )
Ver. 2015.03.09
Po-Ning Chen
144 / 160
Observation 1
For h < 1
Its average PSD is relatively smooth and well confined.
Almost all power is confined within
fT < 0.6 or f <
0.6
T
where T is the width of the channel symbols.
Ver. 2015.03.09
Po-Ning Chen
145 / 160
Observation 2
For h > 1
Its average PSD becomes broader and hence the
bandwidth is approximately
fT < 1.2 or f <
1.2
T
This is the main reason why in communication systems,

where CPFSK is used, the modulation index h is usually
taken to be < 1.
Ver. 2015.03.09
Po-Ning Chen
146 / 160
Example: Bluetooth RF specification (Version 1.0)

GFSK (Gaussian FSK) with BT = 0.5
B = Bandwidth (for baseband symbol) = 0.5 MHz,
T = 1 sec
1 = positive frequency deviation
0 = negative frequency deviation
Modulation index 0.28 0.35

Modulation index = 2Tfd , where fd is the peak
frequency deviation
0.28 < h < 0.35 140KHz < fd < 175KHz
Symbol timing shall be better than 20 ppm.
Ver. 2015.03.09
Po-Ning Chen
147 / 160
Observation3
By letting h 1
we can observe M impulses in the average PSD of the
equivalent lowpass CPFSK signal.
Ver. 2015.03.09
Po-Ning Chen
148 / 160
Numerically plotted average PSD of the equivalent lowpass CPFSK

signal (M = 4 and
Pn uniform over S =
{1, 3})
4 impulses appear
when h 1.
The bandwidth
becomes broader
than almost twice
of that of M = 2.
Ver. 2015.03.09
Po-Ning Chen
149 / 160
8 impulses are observed when h 1

Bandwidth becomes broader than almost four times of
that of M = 2
Ver. 2015.03.09
Po-Ning Chen
150 / 160
Re-visit MSK versus OQPSK
S v ` (f )
S v ` (0)
(dB)
Ver. 2015.03.09
Po-Ning Chen
151 / 160
Observations
Main Lobe: MSK is 50% wider than rectangular OQPSK,
i.e., MSK = 1.5 rectangular OQPSK
Side Lobe:
Compare the bandwidth that contains 99% of the total
power: MSK = 1.2/T and rectangular OQPSK = 8/T .
MSK decreases faster than OQPSK.
MSK is significantly more bandwidth efficient than
rectangular OQPSK.
By further decreasing the modulation index h (i.e.,
making h < 1/2), the bandwidth efficiency of MSKs can
be increased. However, in such case MSK signals are no
longer orthogonal. Therefore, its error probability will be
increased as well. fd = 1/(4T ) h = 2fd T = 1/2
Ver. 2015.03.09
Po-Ning Chen
152 / 160
Appendix: Fractional out-of-band power

Fractional in-band power
PIn-band (W ) =
PTotal W
where
PTotal =
S v ` (f )df
S v ` (f )df .
Fractional out-of-band power

POut-of-band (W ) = 1 PIn-band (W )
This quantity is often used to measure the bandwidth

efficiency of a modulation scheme. For example, finding
the bandwidth W under some acceptable condition, say
fractional-out-of-band power is no greater than 0.01.
Ver. 2015.03.09
Po-Ning Chen
153 / 160
Ver. 2015.03.09
Po-Ning Chen
154 / 160
Summary of spectral characteristics of CPFSK

signals
Modulation Index h
In general, the lower the modulation index h, the higher
the bandwidth efficiency.
Pulse shape g (t)
The smoother (meaning, e.g., continuity of the
derivatives) the g (t), the greater the bandwidth
efficiency.
For example, the raised cosine g (t) will result in higher
bandwidth efficiency than the rectangular g (t).
For example, LRC (raised cosine g (t) with duration LT )
with larger L (i.e., smoother) will result in greater
bandwidth efficiency.
Ver. 2015.03.09
Po-Ning Chen
155 / 160
Ver. 2015.03.09
Po-Ning Chen
156 / 160
Ver. 2015.03.09
Po-Ning Chen
157 / 160
What you learn from Chapter 3
(Pseudo-)Vectorization of standard ASK, PSK and QAM

signals
Computation of average energy based on signal space
vector points
Euclidean distance based on signal space vector points
Gray code mapping from binary pattern to the signal
space vector points (in terms of their Euclidean
distances)
(Good to know) QPSK versus /4-QPSK
Ver. 2015.03.09
Po-Ning Chen
158 / 160
Vectorization of standard orthogonal (FSK or

multi-dimensional) and bi-orthogonal signals
Computation of average energy based on signal space
vector points
Euclidean distance based on signal space vector points
No Gray code mapping issue for orthogonal signals?
Why?
(Important) Cross-correlation of FSK bandpass and
lowpass signals (Minimum shift keying)
(Good to know) Simplex signals (from orthogonal signals)
Ver. 2015.03.09
Po-Ning Chen
159 / 160
(Important) Why cyclo-stationarity for digitally

modulated signals and its power spectrum
Modulation with memory CPM signals
Its basic formation
t
(t; I ) = 4Tfd
d( )d
based on phase change d(t) =

n= In g (t nT )
(Good to know) Full response and partial response
MSK versus OQPSK
Linear representation of CPM
(Important) Time-average autocorrelation and power
spectrum (of cyclostationary PAM and MSK)
Ver. 2015.03.09
Po-Ning Chen
160 / 160

Chap 03

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 03

Uploaded by

Copyright:

Available Formats

Digital Communications

Chapter 3: Digital Modulation Schemes

Institute of Communications Engineering

Digital Communications: Chapter 3

3.1 Representation of digitally

Digital Communications: Chapter 3

Note that the channel symbols are bandpass signals.

Digital Communications: Chapter 3

Memoryless modulation: sm` (t), m` {1, 2, . . . , M},

Modulation with memory: sm` (t), ulation (with memory)

(Eqiuvalent) Bit rate: Rb = T1b = Rs log2 M

3.2 Memoryless modulation

Digital Communications: Chapter 3

Example studies of memoryless modulation

Digital Communications: Chapter 3

M-ary pulse amplitude modulation (M-PAM)

3d g (t) cos (2fc t)

difference between two adjacent signals = 2d.

Digital Communications: Chapter 3

sm (t) = Re {Am g (t)e 2fc t } = Am g (t) cos (2fc t) , t [0, Ts )

Digital Communications: Chapter 3

Vectorization of M-PAM signals (Gram-Schmidt)

By the correct Gram-Schmidt procedure,

For a time-limited signal, we can only claim Ex` 2Ex !

g 2 (t) cos2 (2fc t) dt

If g (t) is constant for t [0, Ts ) and Ts is a multiple of

Based on the pseudo-vectorization,

Digital Communications: Chapter 3

Gray code (Signal space diagram : one dimension)

Digital Communications: Chapter 3

Single side band (SSB) PAM

g (t) is real G (f ) is Hermitian symmetric.

Consequently, the previous PAM is based on the DSB

Digital Communications: Chapter 3

sm,SSB (t) = Re {Am g+ (t)e 2fc t }

Re {g+ (t)e 2fc t } dt

g+ (t) cos [4fc t + (t)] dt

Digital Communications: Chapter 3

Digital Communications: Chapter 3

Recall from Slide 2-22, x+ (t) = 12 (x(t) + x(t)) and Ex = 2Ex+ .

g (t) 2 cos (2fc t)

1,SSB (t) = Re { g+ (t) 2e

2-level PAM signals are particularly named antipodal signals.

where m = 2(m 1)/M, m = 1, 2, , M

Digital Communications: Chapter 3

g (t) cos (2fc t)

Signal space of PM signals

1 (t) g (t) 2 cos (2fc t)

2 (t) g (t) 2 sin (2fc t)

Digital Communications: Chapter 3

Transmission energy of PM Signals

Advantages of PM signals : Equal energy for every

Signal space diagram of PM with Gray code

Digital Communications: Chapter 3

Digital Communications: Chapter 3

Digital Communications: Chapter 3

Quadrature amplitude modulation (QAM)

Digital Communications: Chapter 3

Vectorization of QAM signals

sm (t) = Ami g (t) cos(2fc t) Amq g (t) sin(2fc t)

1 (t) g (t) 2 cos(2fc t)

2 (t) g (t) 2 sin(2fc t)

Digital Communications: Chapter 3