Digital Comm Notes

Digital Modulation 1
Lecture Notes
Ingmar Land and Bernard H. Fleury
Navigation and Communications (NavCom)
Department of Electronic Systems
Aalborg University, DK
Version: February 5, 2007
i
Contents
I Basic Concepts 1
1 The Basic Constituents of a Digital Communication System 2
1.1 Discrete Information Sources . . . . . . . . . . . . 4
1.2 Considered System Model . . . . . . . . . . . . . . 6
1.3 The Digital Transmitter . . . . . . . . . . . . . . . 10
1.3.1 Waveform Look-up Table . . . . . . . . . . 10
1.3.2 Waveform Synthesis . . . . . . . . . . . . . 12
1.3.3 Canonical Decomposition . . . . . . . . . . 13
1.4 The Additive White Gaussian-Noise Channel . . . . 15
1.5 The Digital Receiver . . . . . . . . . . . . . . . . . 17
1.5.1 Bank of Correlators . . . . . . . . . . . . . 17
1.5.2 Canonical Decomposition . . . . . . . . . . 18
1.5.3 Analysis of the correlator outputs . . . . . . 20
1.5.4 Statistical Properties of the Noise Vector . . 23
1.6 The AWGN Vector Channel . . . . . . . . . . . . . 27
1.7 Vector Representation of a Digital Communication System 28
1.8 Signal-to-Noise Ratio for the AWGN Channel . . . 31
2 Optimum Decoding for the AWGN Channel 33
2.1 Optimality Criterion . . . . . . . . . . . . . . . . . 33
2.2 Decision Rules . . . . . . . . . . . . . . . . . . . . 34
2.3 Decision Regions . . . . . . . . . . . . . . . . . . . 36
Land, Fleury: Digital Modulation 1 NavCom
ii
2.4 Optimum Decision Rules . . . . . . . . . . . . . . 37
2.5 MAP Symbol Decision Rule . . . . . . . . . . . . . 39
2.6 ML Symbol Decision Rule . . . . . . . . . . . . . . 41
2.7 ML Decision for the AWGN Channel . . . . . . . . 42
2.8 Symbol-Error Probability . . . . . . . . . . . . . . 44
2.9 Bit-Error Probability . . . . . . . . . . . . . . . . . 46
2.10 Gray Encoding . . . . . . . . . . . . . . . . . . . . 48
II Digital Communication Techniques 50
3 Pulse Amplitude Modulation (PAM) 51
3.1 BPAM . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1.1 Signaling Waveforms . . . . . . . . . . . . . 51
3.1.2 Decomposition of the Transmitter . . . . . . 52
3.1.3 ML Decoding . . . . . . . . . . . . . . . . . 53
3.1.4 Vector Representation of the Transceiver . . 54
3.1.5 Symbol-Error Probability . . . . . . . . . . 55
3.1.6 Bit-Error Probability . . . . . . . . . . . . . 57
3.2 MPAM . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.3 ML Decoding . . . . . . . . . . . . . . . . . 62
iii
4 Pulse Position Modulation (PPM) 67
4.1 BPPM . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.3 ML Decoding . . . . . . . . . . . . . . . . . 69
4.2 MPPM . . . . . . . . . . . . . . . . . . . . . . . . 74
4.2.3 ML Decoding . . . . . . . . . . . . . . . . . 77
5 Phase-Shift Keying (PSK) 85
5.1 Signaling Waveforms . . . . . . . . . . . . . . . . . 85
5.2 Signal Constellation . . . . . . . . . . . . . . . . . 87
5.3 ML Decoding . . . . . . . . . . . . . . . . . . . . . 90
5.4 Vector Representation of the MPSK Transceiver . . 93
5.5 Symbol-Error Probability . . . . . . . . . . . . . . 94
6 Oset Quadrature Phase-Shift Keying (OQPSK)100
6.1 Signaling Scheme . . . . . . . . . . . . . . . . . . . 100
6.2 Transceiver . . . . . . . . . . . . . . . . . . . . . . 101
6.3 Phase Transitions for QPSK and OQPSK . . . . . 102
iv
7 Frequency-Shift Keying (FSK) 104
7.1 BFSK with Coherent Demodulation . . . . . . . . . 104
7.1.1 Signal Waveforms . . . . . . . . . . . . . . 104
7.1.2 Signal Constellation . . . . . . . . . . . . . 106
7.1.3 ML Decoding . . . . . . . . . . . . . . . . . 107
7.1.5 Error Probabilities . . . . . . . . . . . . . . 109
7.2 MFSK with Coherent Demodulation . . . . . . . . 110
7.2.4 Error Probabilities . . . . . . . . . . . . . . 113
7.3 BFSK with Noncoherent Demodulation . . . . . . . 114
7.3.3 Noncoherent Demodulator . . . . . . . . . . 116
7.3.4 Conditional Probability of Received Vector . 117
7.3.5 ML Decoding . . . . . . . . . . . . . . . . . 119
7.4 MFSK with Noncoherent Demodulation . . . . . . 121
7.4.3 Conditional Probability of Received Vector . 123
7.4.4 ML Decision Rule . . . . . . . . . . . . . . 123
7.4.5 Optimal Noncoherent Receiver for MFSK . . 124
v
8 Minimum-Shift Keying (MSK) 130
8.1 Motivation . . . . . . . . . . . . . . . . . . . . . . 130
8.2 Transmit Signal . . . . . . . . . . . . . . . . . . . 132
8.3 Signal Constellation . . . . . . . . . . . . . . . . . 135
8.4 A Finite-State Machine . . . . . . . . . . . . . . . 138
8.5 Vector Encoder . . . . . . . . . . . . . . . . . . . 140
8.6 Demodulator . . . . . . . . . . . . . . . . . . . . . 144
8.7 Conditional Probability Density of Received Sequence145
8.8 ML Sequence Estimation . . . . . . . . . . . . . . 146
8.9 Canonical Decomposition of MSK . . . . . . . . . . 149
8.10 The Viterbi Algorithm . . . . . . . . . . . . . . . . 150
8.11 Simplied ML Decoder . . . . . . . . . . . . . . . 155
8.13 Dierential Encoding and Dierential Decoding . . 161
8.14 Precoded MSK . . . . . . . . . . . . . . . . . . . . 162
8.15 Gaussian MSK . . . . . . . . . . . . . . . . . . . . 166
9 Quadrature Amplitude Modulation 169
10 BPAM with Bandlimited Waveforms 170
III Appendix 171
vi
A Geometrical Representation of Signals 172
A.1 Sets of Orthonormal Functions . . . . . . . . . . . 172
A.2 Signal Space Spanned by an Orthonormal Set of Functions173
A.3 Vector Representation of Elements in the Vector Space174
A.4 Vector Isomorphism . . . . . . . . . . . . . . . . . 175
A.5 Scalar Product and Isometry . . . . . . . . . . . . 176
A.6 Waveform Synthesis . . . . . . . . . . . . . . . . . 178
A.7 Implementation of the Scalar Product . . . . . . . . 179
A.8 Computation of the Vector Representation . . . . . 181
A.9 Gram-Schmidt Orthonormalization . . . . . . . . . 182
B Orthogonal Transformation of a White Gaussian Noise Vector 185
B.1 The Two-Dimensional Case . . . . . . . . . . . . . 185
B.2 The General Case . . . . . . . . . . . . . . . . . . 186
C The Shannon Limit 188
C.1 Capacity of the Band-limited AWGN Channel . . . 188
C.2 Capacity of the Band-Unlimited AWGN Channel . 191
vii
References
[1] J. G. Proakis and M. Salehi, Communication Systems Engi-
neering, 2nd ed. Prentice-Hall, 2002.
[2] B. Rimoldi, A decomposition approach to CPM, IEEE
Trans. Inform. Theory, vol. 34, no. 2, pp. 260270, Mar. 1988.
1
Part I
Basic Concepts
2
1 The Basic Constituents of a Digi-
tal Communication System
Block diagram of a digital communication system:
PSfrag replacements
Binary Information Source
Binary Information Sink
Digital Transmitter
Digital Receiver
Information
Information
Source
Source
Source
Encoder Encoder
Vector
Vector
Waveform
Waveform
Waveform
Modulator
Channel
Decoder Decoder Demodulator
Sink
U
k
U
k
X(t)
Y (t)
1/T
b
1/T
b
1/T
1/T
Time duration of one binary symbol (bit) U
k
: T
b
Time duration of one waveform X(t): T
Waveform channel may introduce
- distortion,
- interference,
- noise
Goal: Signal of information source and signal at information sink
should be as similar as possible according to some measure.
3
Some properties:
Information source: may be analog or digital
Source encoder: may perform sampling, quantization and
compression; generates one binary symbol (bit) U
k
0, 1
per time interval T
b
Waveform modulator: generates one waveform x(t) per time
interval T
Vector decoder: generates one binary symbol (bit)

U
k

0, 1 (estimates of U
k
) per time interval T
b
Source decoder: reconstructs the original source signal
4
1.1 Discrete Information Sources
A discrete memoryless source (DMS) is a source that generates
a sequence U
1
, U
2
, . . . of independent and identically distributed
(i.i.d.) discrete random variables, called an i.i.d. sequence.
PSfrag replacements
DMS
U
1
, U
2
, . . .
Properties of the output sequence:
discrete U
1
, U
2
, . . . a
1
, a
2
, . . . , a
Q
.
memoryless U
1
, U
2
, . . . are statistically independent.
stationary U
1
, U
2
, . . . are identically distributed.
Notice: The symbol alphabet a
1
, a
2
, . . . , a
Q
has cardinality Q.
The value of Q may be innite, the elements of the alphabet only
have to be countable.
Example: Discrete sources
(i) Output of the PC keyboard, SMS
(usually not memoryless).
(ii) Compressed le
(near memoryless).
(iii) The numbers drawn on a roulette table in a casino
(ought to be memoryless, but may not. . . ).
3
5
A DMS is statistically completely described by the probabilities
p
U
(a
q
) = Pr(U = a
q
)
of the symbols a
q
, q = 1, 2, . . . , Q. Notice that
(i) p
U
(a
q
) 0 for all q = 1, 2, . . . , Q, and
(ii)
Q
q=1
p
U
(a
q
) = 1.
As the sequence is i.i.d., we have Pr(U
i
= a
q
) = Pr(U
j
= a
q
).
For convenience, we may write p
U
(a) shortly as p(a).
Binary Symmetric Source
A binary memoryless source is a DMS with a binary symbol al-
phabet.
Remark:
Commonly used binary symbol alphabets are F
2
:= 0, 1 and
B := 1, +1. In the following, we will use F
2
.
A binary symmetric source (BSS) is a binary memoryless source
with equiprobable symbols,
p
U
(0) = p
U
(1) =
1
2
.
Example: Binary Symmetric Source
All length-K sequences of a BSS have the same probability
p
U
1
U
2
...U
K
(u
1
u
2
. . . u
K
) =
K
i=1
p
U
(u
i
) =
_
1
2
_
K
.
3
6
1.2 Considered System Model
This course focuses on the transmitter and the receiver. Therefore,
we replace the information source by a BSS
PSfrag replacements
BSS
Binary
Information Source
Source Encoder
Decoder
Sink
U
1
, U
2
, . . .
U
1
, U
2
, . . .
and the information sink by a binary (information) sink.
PSfrag replacements
BSS
Binary
Information Source
Encoder
Decoder
Sink
Sink
U
1
, U
2
, . . .

U
1
,

U
2
, . . .
U
1
,

U
2
, . . .
7
Some Remarks
(i) There is no loss of generality resulting from these substitutions.
Indeed it can be demonstrated within Shannons Information The-
ory that an ecient source encoder converts the output of an in-
formation source into a random binary independent and uniformly
distributed (i.u.d.) sequence. Thus, the output of a perfect source
encoder looks like the output of a BSS.
(ii) Our main concern is the design of communication systems for
reliable transmission of the output symbols of a BSS. We will not
address the various methods for ecient source encoding.
8
Digital communication system considered in this course:
PSfrag replacements
BSS
Binary
Sink
Digital
Digital
Transmitter
Receiver
Waveform
Channel
U
U
X(t)
Y (t)
bit rate 1/T
b
rate 1/T
Source: U = [U
1
, . . . , U
K
], U
i
0, 1
Transmitter: X(t) = x(t, u)
Sink:

U = [
U
1
, . . . ,

U
K
],

U
i
0, 1
Bit rate: 1/T
b
Rate (of waveforms): 1/T = 1/(KT
b
)
Bit error probability
P
b
=
1
K
K
k=1
Pr(U
k
,=

U
k
)
Objective: Design of ecient digital communication systems
Eciency means
_
small bit error probability and
high bit rate 1/T
b
.
9
Constraints and limitations:
limited power
limited bandwidth
impairments (distortion, interference, noise) of the channel
Design goals:
good waveforms
low-complexity transmitters
low-complexity receivers
10
1.3 The Digital Transmitter
1.3.1 Waveform Look-up Table
Example: 4PPM (Pulse-Position Modulation)
Set of four dierent waveforms, S = s
1
(t), s
2
(t), s
3
(t), s
4
(t):
PSfrag replacements
0
0
T
A
t
s
1
(t)
PSfrag replacements
0
0
T
A
t
s
1
(t)
s
2
(t)
PSfrag replacements
0
0
T
A
t
s
1
(t)
s
2
(t)
s
3
(t)
PSfrag replacements
0
0
T
A
t
s
1
(t)
s
2
(t)
s
3
(t)
s
4
(t)
Each waveform X(t) S is addressed by vector [U
1
, U
2
] 0, 1
2
:
[U
1
, U
2
] X(t).
The mapping may be implemented by a waveform look-up table.
PSfrag replacements
4PPM
Transmitter
[U
1
, U
2
] X(t)
[00] s
1
(t)
[01] s
2
(t)
[10] s
3
(t)
[11] s
4
(t)
Example: [00011100] is transmitted as
PSfrag replacements
0
0
T 2T 3T 4T
A
t
x(t)
3
11
For an arbitrary digital transmitter, we have the following:
PSfrag replacements
Digital
Transmitter
Waveform
Look-up
Table
[U
1
, U
2
, . . . , U
K
] X(t)
Input Binary vectors of length K from the input set U:
[U
1
, U
2
, . . . , U
K
] U := 0, 1
K
.
The duration of one binary symbol U
k
is T
b
.
Output Waveforms of duration T from the output set S:
x(t) S := s
1
(t), s
2
(t), . . . , s
M
(t).
Waveform duration T means that for m = 1, 2, . . . , M,
s
m
(t) = 0 for t / [0, T].
The look-up table maps each input vector [u
1
, . . . , u
K
] U to one
waveform x(t) S. Thus the digital transmitter may dened by
a mapping U S with
[u
1
, . . . , u
K
] x(t).
The mapping is one-to-one and onto such that
M = 2
K
.
Relation between signaling interval (waveform duration) T and bit
interval (bit duration) T
b
:
T = K T
b
.
12
1.3.2 Waveform Synthesis
The set of waveforms,
S := s
1
(t), s
2
(t), . . . , s
M
(t),
spans a vector space. Applying the Gram-Schmidt orthogonaliza-
tion procedure, we can nd a set of orthonormal functions
S
:=
1
(t),
2
(t), . . . ,
D
(t) with D M
such that the space spanned by S
contains the space spanned

by S.
Hence, each waveform s
m
(t) can be represented by a linear combi-
nation of the orthonormal functions:
s
m
(t) =
D
i=1
s
m,i

i
(t),
m = 1, 2, . . . , M. Each signal s
m
(t) can thus be geometrically
represented by the D-dimensional vector
s
m
= [s
m,1
, s
m,2
, . . . , s
m,D
]
T
R
D
with respect to the set S
.
Further details are given in Appendix A.
13
1.3.3 Canonical Decomposition
The digital transmitter may be represented by a waveform look-up
table:
u = [u
1
, . . . , u
K
] x(t),
where u U and x(t) S.
From the previous section, we know how to synthesize the wave-
form s
m
(t) from s
m
(see also Appendix A) with respect to a set of
basis functions
S
:=
1
(t),
2
(t), . . . ,
D
(t).
Making use of this method, we can split the waveform look-up
table into a vector look-up table and a waveform synthesizer:
u = [u
1
, . . . , u
K
] x = [x
1
, x
2
, . . . , x
D
] x(t),
where
u U = 0, 1
K
,
x X = s
1
, s
2
, . . . , s
D
R
D
,
x(t) S = s
1
(t), s
2
(t), . . . , s
M
(t).
This splitting procedure leads to the sought canonical decomposi-
tion of a digital transmitter:
PSfrag replacements
Vector
Vector
Encoder
Waveform
Modulator
Look-up Table
[U
1
, . . . , U
K
]
X
1
X
D
1
(t)
D
(t)
X(t)
[0 . . . 0] s
1
[1 . . . 1] s
M
14
Example: 4PPM
The four orthonormal basis functions are
PSfrag replacements
0
0
T
2/
T
t
1
(t)
PSfrag replacements
0
0
T
2/
T
t
1
(t)
2
(t)
PSfrag replacements
0
0
T
2/
T
t
1
(t)
2
(t)
3
(t)
PSfrag replacements
0
0
T
2/
T
t
1
(t)
2
(t)
3
(t)
4
(t)
The canonical decomposition of the receiver is the following, where
E
s
= A
T/2.
PSfrag replacements
Vector Encoder
Vector
Encoder
Waveform Modulator
Waveform
Modulator
Vector
Vector
Look-up Table
[U
1
, U
2
]
X
1
X
2
X
3
X
4
1
(t)
2
(t)
3
(t)
4
(t)
X(t)
[00] [
E
s
, 0, 0, 0]
T
[01] [0,
E
s
, 0, 0]
T
[10] [0, 0,
E
s
, 0]
T
[11] [0, 0, 0,
E
s
]
T
Remarks:
- The signal energy of each basis function is equal to one:
_

d
(t)dt = 1 (due to their construction).
- The basis functions are orthonormal (due to their construction).
- The value
E
s
is chosen such that the synthesized signals have
the same energy as the original signals, namely E
s
(compare
Example 3).
3
15
1.4 The Additive White Gaussian-Noise
Channel
PSfrag replacements
X(t) Y (t)
W(t)
The additive white Gaussian noise (AWGN) channel-model is
widely used in communications. The transmitted signal X(t) is
superimposed by a stochastic noise signal W(t), such that the
transmitted signal reads
Y (t) = X(t) + W(t).
The stochastic process W(t) is a stationary process with the
following properties:
(i) W(t) is a Gaussian process, i.e., for each time t, the
samples v = w(t) are Gaussian distributed with zero mean
and variance
2
:
p
V
(v) =
1
2
2
exp
_
v
2
2
2
_
.
(ii) W(t) has a at power spectrum with height N
0
/2:
S
W
(f) =
N
0
2
(Therefore it is called white.)
(iii) W(t) has the autocorrelation function
R
W
() =
N
0
2
().
16
Remarks
1. Autocorrelation function and power spectrum:
R
W
() = E
_
W(t)W(t + )
, S
W
(f) = T
_
R
W
()
_
.
2. For Gaussian processes, weak stationarity implies strong sta-
tionarity.
3. An WGN is an idealized process without physical reality:
The process is so wild that its realizations are not
ordinary functions of time.
Its power is innite.
However, a WGN is a useful approximation of a noise with a
at power spectrum in the bandwidth B used by a commu-
nication system:
PSfrag replacements
S
noise
(f)
N
0
/2
f
f
0
f
0
B B
4. In satellite communications, W(t) is the thermal noise of the
receiver front-end. In this case, N
0
/2 is proportional to the
squared temperature.
17
1.5 The Digital Receiver
Consider a transmission system using the waveforms
S =
_
s
1
(t), s
2
(t), . . . , s
M
(t)
_
with s
m
(t) = 0 for t / [0, T], m = 1, 2, . . . , M, i.e., with dura-
tion T.
Assume transmission over an AWGN channel, such that
y(t) = x(t) + w(t),
x(t) S.
1.5.1 Bank of Correlators
The transmitted waveform may be recovered from the received
waveform y(t) by correlating y(t) with all possible waveforms:
c
m
= y(t), s
m
(t)),
m = 1, 2, . . . , M. Based on the correlations [c
1
, . . . , c
M
], the wave-
form s(t) with the highest correlation is chosen and the correspond-
ing (estimated) source vector u = [ u
1
, . . . , u
K
] is output.
PSfrag replacements
y(t)
s
1
(t)
s
M
(t)
_
T
0
(.)dt
_
T
0
(.)dt
c
1
c
M
estimation
of x(t)
and
table
look-up
[ u
1
, . . . , u
K
]
Disadvantage: High complexity.
18
1.5.2 Canonical Decomposition
Consider a set of orthonormal functions
S
=
_
1
(t),
2
(t), . . . ,
D
(t)
_
obtained by applying the Gram-Schmidt procedure to
s
1
(t), s
2
(t), . . . , s
M
(t). Then,
s
m
(t) =
D
d=1
s
m,d

d
(t),
m = 1, 2, . . . , M. The vector
s
m
= [s
m,1
, s
m,2
, . . . , s
m,D
]
T
entirely determines s
m
(t) with respect to the orthonormal set S
.
The set S
spans the vector space

o
:=
_
s(t) =
D
i=1
s
i
i
(t) : [s
1
, s
2
, . . . , s
D
] R
D
_
.
Basic Idea
The received waveform y(t) may contain components outside
of o
. However, only the components inside o
are relevant,
as x(t) o
.
Determine the vector representation y of the components
of y(t) that are inside o
.
This vector y is sucient for estimating the transmitted
waveform.
19
Canonical decomposition of the optimal receiver for
the AWGN channel
Appendix A describes two ways to compute the vector represen-
tation of a waveform. Accordingly, the demodulator may be im-
plemented in two ways, leading to the following two receiver struc-
tures.
Correlator-based Demodulator and Vector Decoder
PSfrag replacements
Vector
Decoder
Y (t)
1
(t)
D
(t)
_
T
0
(.)dt
_
T
0
(.)dt
Y
1
Y
D
[
U
1
, . . . ,

U
K
]
Matched-lter based Demodulator and Vector De-
coder
PSfrag replacements
Vector
Decoder
Y (t)
1
(T t)
D
(T t)
MF
MF
Y
1
Y
D
T
[
U
1
, . . . ,

U
K
]
The optimality of the receiver structures is shown in the following
sections.
20
1.5.3 Analysis of the correlator outputs
Assume that the waveform s
m
(t) is transmitted, i.e.,
x(t) = s
m
(t).
The received waveform is thus
y(t) = x(t) + w(t) = s
m
(t) + w(t).
The correlator outputs are
y
d
=
T
_
0
y(t)
d
(t)dt
=
T
_
0
_
s
m
(t) + w(t)
_
d
(t)dt
=
T
_
0
s
m
(t)
d
(t)dt
. .
s
m,d
+
T
_
0
w(t)
d
(t)dt
. .
w
d
= s
m,d
+ w
d
,
d = 1, 2, . . . , D.
Using the notation
y = [y
1
, . . . , y
D
]
T
,
w = [w
1
, . . . , w
D
]
T
, w
d
=
T
_
0
w(t)
d
(t)dt,
we can recast the correlator outputs in the compact form
y = s
m
+ w.
The waveform channel between x(t) = s
m
(t) and y(t) is thus trans-
formed into a vector channel between x = s
m
and y.
21
Remark:
Since the vector decoder sees noisy observations of the vectors
s
m
, these vectors are also called signal points with respect to S
,
and the set of signal points is called the signal constellation.
From Appendix A, we know that s
m
(t) entirely characterizes s
m
(t).
The vector y, however, does not represent the complete wave-
form y(t); it represents only the waveform
y
(t) =
D
d=1
y
d
d
(t)
which is the part of y(t) within the vector space spanned by the
set o
of orthonormal functions, i.e.,

y
(t) o
.
The remaining part of y(t), i.e., the part outside of o
, reads
y
(t) = y(t) y
(t)
= y(t)
D
d=1
y
d
d
(t)
= s
m
(t) + w(t)
D
d=1
[s
m,d
+ w
d
]
d
(t)
= s
m
(t)
D
d=1
s
m,d
d
(t)
. .
= 0
+w(t)
D
d=1
w
d
d
(t)
= w(t)
D
d=1
w
d
d
(t).
22
Observations
The waveform y
(t) contains no information about s

m
(t).
The waveform y
(t) is completely represented by y.

Therefore, the receiver structures do not lead to a loss of informa-
tion, and so they are optimal.
23
1.5.4 Statistical Properties of the Noise Vector
The components of the noise vector
w = [w
1
, . . . , w
D
]
T
are given by
w
d
=
T
_
0
w(t)
d
(t)dt,
d = 1, 2, . . . , D. The random variables W
d
are Gaussian as they
result from a linear transformation of the Gaussian process W(t).
Hence the probability density function of W
d
is of the form
p
W
d
(w
d
) =
1
_
2
2
d
exp
_
(w
d
d
)
2
2
2
d
_
.
where the mean value
d
and the variance
2
d
are given by
d
= E
_
W
d
,
2
d
= E
_
(W
d
d
)
2
.
These values are now determined.
Mean value
Using the denition of W
d
and taking into account that the pro-
cess W(t) has zero mean, we obtain
E
_
W
d
= E
_
T
_
0
W(t)
d
(t)dt
_
=
T
_
0
E
_
W(t)
d
(t)dt
= 0.
Thus, all random variables W
d
have zero mean.
24
Variance
For computing the variance, we use
d
= 0, the denition of W
d
,
and the autocorrelation function of W(t),
R
W
() =
N
0
2
().
We obtain the following chain of equations:
E
_
(W
d
d
)
2
= E
_
W
2
d
= E
_
_
T
_
0
W(t)
d
(t)dt
__
T
_
0
W(t
)
d
(t
)dt
_
_
=
T
_
0
T
_
0
E
_
W(t)W(t
. .
R
W
(t
t)
d
(t)
d
(t
)dtdt
=
N
0
2
T
_
0
_
T
_
0
(t t
)
d
(t
)dt
_
. .
(t)
d
(t) =
d
(t)
d
(t)dt
=
N
0
2
T
_
0
2
d
(t)dt =
N
0
2
.
Distribution of the noise vector components
Thus we have
d
= 0 and
2
d
= N
0
/2, and the distributions read
p
W
d
(w
d
) =
1
N
0
exp
_
w
2
d
N
0
_
.
for d = 1, 2, . . . , D. Notice that the components W
d
are all iden-
tically distributed.
25
As the Gaussian distribution is very important, the shorthand
notation
A ^(,
2
)
is often used. This notation means: The random variable A is
Gaussian distributed with mean value and variance
2
.
Using this notation, we have for the components of the noise vector
W
d
^(0, N
0
/2),
d = 1, 2, . . . , D.
Distribution of the noise vector
Using the same reasoning as before, we can conclude that W is a
Gaussian noise vector, i.e.,
W ^(, ),
where the expectation R
D1
and the covariance matrix
R
DD
are given by
= E
_
W
, = E
_
(W )(W )
T
.
From the previous considerations, we have immediately
= 0.
Using a similar approach as for computing the variances, the sta-
tistical independence of the components of W can be shown. That
is,
E
_
W
i
W
j
=
N
0
2

i,j
,
where
i,j
denotes the Kronecker symbol dened as
i,j
=
_
1 for i = j,
0 otherwise.
26
As a consequence, the covariance matrix of W is
=
N
0
2
I
D
,
where I
D
denotes the D D identity matrix.
To sum up, the noise vector W is distributed as
W ^
_
0,
N
0
2
I
D
_
.
As the components of W are statistically independent, the prob-
ability density function of W can be written as
p
W
(w) =
D
d=1
p
W
d
(w
d
)
=
D
d=1
1
N
0
exp
_
w
2
d
N
0
_
=
_
1
N
0
_
D
exp
_
1
N
0
D
d=1
w
2
d
_
=
_
N
0
_
D/2
exp
_
1
N
0
|w|
2
_
with |w| denoting the Euclidean norm
|w| =
_
D
d=1
w
2
d
_
1/2
of the D-dimensional vector w = [w
1
, . . . , w
D
]
T
.
The independency of the components of W give rise to the AWGN
vector channel.
27
1.6 The AWGN Vector Channel
The additive white Gaussian noise vector channel is dened as a
channel with input vector
X = [X
1
, X
2
, . . . , X
D
]
T
,
output vector
Y = [Y
1
, Y
2
, . . . , Y
D
]
T
,
and the relation
Y = X + W,
where
W = [W
1
, W
2
, . . . , W
D
]
T
is a vector of D independent random variables distributed as
W
d
^(0, N
0
/2).
PSfrag replacements
X Y
W
(As nobody would have been expected dierently. . . )
28
1.7 Vector Representation of a Digital
Communication System
Consider a digital communication system transmitting over
an AWGN channel. Let s
1
(t), s
2
(t), . . . , s
M
(t) , s
m
(t) = 0 for
t / [0, T], m = 1, 2, . . . , M, denote the waveforms. Let further
S
=
_
1
(t),
2
(t), . . . ,
D
(t)
_
denote a set of orthonormal functions for these waveforms.
The canonical decompositions of the digital transmitter and the
digital receiver convert the block formed by the waveform modu-
lator, the AWGN channel, and the waveform demodulator into an
AWGN vector channel.
Modulator, AWGN channel, demodulator
PSfrag replacements
1
(t)
1
(t)
D
(t)
D
(t)
_
T
0
(.)dt
_
T
0
(.)dt
X(t)
X
1
X
D
Y (t)
Y
1
Y
D
W(t)
AWGN vector channel
PSfrag replacements
[X
1
, . . . , X
D
] [Y
1
, . . . , Y
D
]
[W
1
, . . . , W
D
]
29
Using this equivalence, the communication system operating over a
(waveform) AWGN channel can be represented as a communication
system operating over an AWGN vector channel:
PSfrag replacements
BSS
Binary
Sink Vector
Vector
Vector
Encoder
Decoder
AWGN
Channel
U
U
X
Y
W
Source vector: U = [U
1
, . . . , U
K
], U
i
0, 1
Transmit vector: X = [X
1
, . . . , X
D
]
Receive vector: Y = [Y
1
, . . . , Y
D
]
Estimated source vector:

U = [
U
1
, . . . ,

U
K
], U
i
0, 1
AWGN vector: W = [W
1
, . . . , W
D
],
W ^
_
0,
N
0
2
I
D
_
This is the vector representation of a digital communication system
transmitting over an AWGN channel.
30
For estimating the transmitted vector x, we are interested in the
likelihood of x for the given received vector y:
p
Y |X
(y[x) = p
W
(y x)
=
_
N
0
_
D/2
exp
_
1
N
0
|y x|
2
_
.
Symbolically, we may write
Y [X = s
m
^
_
s
m
,
N
0
2
I
D
_
.
Remark:
The likelihood of x is the conditional probability density function
of y given x, and it is regarded as a function of x.
Example: Signal buried in AWGN
Consider D = 3 and x = s
m
. Then the Gaussian cloud
around s
m
may look like this:
PSfrag replacements
3
s
m
3
31
1.8 Signal-to-Noise Ratio for the AWGN
Channel
Consider a digital communication system using the set of wave-
forms
S =
_
s
1
(t), s
2
(t), . . . , s
M
(t)
_
,
s
m
(t) = 0 for t / [0, T], m = 1, 2, . . . , M, for transmission over
an AWGN channel with power spectrum
S
W
(f) =
N
0
2
.
Let E
s
m
denote the energy of waveform s
m
(t), and let E
s
denote
the average waveform energy:
E
s
m
=
T
_
0
_
s
m
(t)
_
2
dt, E
s
=
1
M
M
m=1
E
s
m
Thus, E
s
is the average energy per waveform x(t) and also
the average energy per vector x (due to the one-to-one correspon-
dence.)
The average signal-to-noise ratio (SNR) per waveform is
then dened as
s
=
E
s
N
0
.
32
On the other hand, the average energy per source bit u
k
is of
interest. Since K source bits u
k
are transmitted per waveform, the
average energy per bit is
E
b
=
1
K
E
s
.
Accordingly, the average signal-to-noise ratio (SNR) per
bit is then dened as
b
=
E
b
N
0
.
As K bits are transmitted per waveform, we have the relation
b
=
1
K
s
.
Remark:
The energies and signal-to-noise ratios usually denote received val-
ues.
33
2 Optimum Decoding for the
AWGN Channel
2.1 Optimality Criterion
In the previous chapter, the vector representation of a digital
transceiver transmitting over an AWGN channel was derived.
PSfrag replacements
BSS
Binary
Sink Vector
Vector
Vector
Encoder
Decoder
AWGN
Channel
U
U
X
Y
W
Nomenclature
u = [u
1
, . . . , u
K
] U, x = [x
1
, . . . , x
D
] X,
u = [ u
1
, . . . , u
K
] U, y = [y
1
, . . . , y
D
] R
D
,
U = 0, 1
K
, X = s
1
, . . . , s
M
.
U : set of all binary sequences of length K,
X : set of vector representations of the waveforms.
The set X is called the signal constellation, and its elements
are called signals, signal points, or modulation symbols.
For convenience, we may refer to u
k
as bits and to x as symbols.
(But we have to be careful not to cause ambiguity.)
34
As the vector representation implies no loss of information, re-
covering the transmitted block u = [u
1
, . . . , u
K
] from the re-
ceived waveform y(t) is equivalent to recovering u from the vec-
tor y = [y
1
, . . . , y
D
].
Functionality of the vector encoder
enc : u x = enc(u)
U X
Functionality of the vector decoder
dec
u
: y u = dec
u
(y)
R
D
U
Optimality criterion
A vector decoder is called an optimal vector decoder if it
minimizes the block error probability Pr(U ,=

U).
2.2 Decision Rules
Since the encoder mapping is one-to-one, the functionality of the
vector decoder can be decomposed into the following two opera-
tions:
(i) Modulation-symbol decision:
dec : y x = dec(y)
R
D
X
(ii) Encoder inverse:
enc
1
: x u = enc
1
( x)
X U
35
Thus, we have the following chain of mappings:
y
dec
x
enc
1
u.
Block Diagram
PSfrag replacements
Vector
Encoder
Encoder
Vector Decoder
Inverse
Mod. Symbol
Decision
enc
enc
1
dec
U
U
X
X
Y
W
U
U X
X
R
D
R
D
A decision rule is a function which maps each observation y
R
D
to a modulation symbol x X, i.e., to an element of the signal
constellation X.
Example: Decision rule
PSfrag replacements
R
D
y
(1)
y
(2)
y
(3)
_ _
, , ,
. . . s
1
s
2
s
M = X
3
36
2.3 Decision Regions
A decision rule dec : R
D
X denes a partition of R
D
into deci-
sion regions
T
m
:=
_
y R
D
: dec(y) = s
m
_
,
m = 1, 2, . . . , M.
Example: Decision regions
.
.
.
PSfrag replacements
R
D
T
1
T
2
T
M
_ _
, , ,
. . . s
1
s
2
s
M = X
3
Since the decision regions T
1
, T
2
, . . . , T
M
form a partition, they
have the following properties:
(i) They are disjoint:
T
i
T
j
= for i ,= j.
(ii) They cover the whole observation space R
D
:
T
1
T
2
T
M
= R
D
.
Obviously, any partition of R
D
denes a decision rule as well.
37
2.4 Optimum Decision Rules
Optimality criterion for decision rules
Since the mapping enc : U X of the vector encoder is one-to-
one, we have
U ,= U

X ,= X.
Therefore, the decision rule of an optimal vector decoder minimizes
the modulation-symbol error probability Pr(

X ,= X).
Such a decision rule is called optimal.
Sucient condition for an optimal decision rule
Let dec : R
D
X be a decision rule which minimizes the
a-posteriori probability of making a decision error,
Pr
_
dec(y)
. .
x
,= X
Y = y
_
,
for any observation y. Then dec is optimal.
Remark:
The a-posteriori probability of a decoding error is
the probability of the event dec(y) ,= X conditioned on the
event Y = y:
PSfrag replacements
Mod. Symbol
Decision
dec
X
X = dec(y)
Y = y
W
38
Proof:
Let dec : R
D
X denote any decision rule, and let
dec
opt
: R
D
X denote an optimal decision rule. Then we
have
Pr
_
dec(Y ) ,= X
_
=
_
R
D
Pr
_
dec(Y ) ,= X
Y = y
_
p
Y
(y)dy
=
_
R
D
Pr
_
dec(y) ,= X
Y = y
_
p
Y
(y)dy
()
_
R
D
Pr
_
dec
opt
(y) ,= X
Y = y
_
p
Y
(y)dy
=
_
R
D
Pr
_
dec
opt
(Y ) ,= X
Y = y
_
p
Y
(y)dy
= Pr
_
dec
opt
(Y ) ,= X
_
.
(): by denition of dec
opt
, we have
Pr
_
dec(y) ,= X
Y = y
_
Pr
_
dec
opt
(y) ,= X
Y = y
_
for all x X.
39
2.5 MAP Symbol Decision Rule
Starting with the necessary condition for an optimal decision rule
given above, we derive now a method to make an optimal decision.
The corresponding decision rule is called the maximum a-posteriori
(MAP) decision rule.
Consider an optimal decision rule
dec : y x = dec(y),
where the estimated modulation-symbol is denoted by x.
As the decision rule is assumed to be optimum, the probability
Pr( x ,= X[Y = y)
is minimal.
For any x X, we have
Pr(X ,= x[Y = y) = 1 Pr(X = x[Y = y)
= 1 p
X|Y
(x[y)
= 1
p
Y |X
(y[x) p
X
(x)
p
Y
(y)
,
where Bayes rule has been used in the last line. Notice that
Pr(X = x[Y = y) = p
X|Y
(x[y)
is the a-posteriori probability of X = x given Y = y.
40
Using the above equalities, we can conclude that
Pr(X ,= x[Y = y) is minimal
p
X|Y
(x[y) is maximal
p
Y |X
(y[x) p
X
(x) is maximal,
since p
Y
(y) does not depend on x.
Thus, the estimated modulation-symbol x = dec(y) resulting from
the optimal decision rule satises the two equivalent conditions
p
X|Y
( x[y) p
X|Y
(x[y) for all x X;
p
Y |X
(y[ x) p
X
( x) p
Y |X
(y[x) p
X
(x) for all x X.
These two conditions may be used to dene the optimal decision
rule.
Maximum a-posteriori (MAP) decision rule:
x = dec
MAP
(y) := argmax
xX
_
p
X|Y
(x[y)
_
:= argmax
xX
_
p
Y |X
(y[x) p
X
(x)
_
.
The decision rule is called the MAP decision rule as the selected
modulation-symbol x maximizes the a-posteriori probability den-
sity function p
X|Y
(x[y). Notice that the two formulations are
equivalent.
41
2.6 ML Symbol Decision Rule
If the modulation-symbols at the output of the vector encoder are
equiprobable, i.e., if
Pr(X = s
m
) =
1
M
for m = 1, 2, . . . , M, the MAP decision rule reduces to the follow-
ing decision rule.
Maximum likelihood (ML) decision rule:
x = dec
ML
(y) := argmax
xX
_
p
Y |X
(y[x)
_
.
In this case, the the selected modulation-symbol x maximizes the
likelihood function p
Y |X
(y[x) of x.
42
2.7 ML Decision for the AWGN Channel
Using the canonical decomposition of the transmitter and of the
receiver, we obtain the AWGN vector-channel:
PSfrag replacements
X X
Y R
D
W ^
_
0,
N
0
2
I
D
_
The conditional probability density function of Y given
X = x X is
p
Y |X
(y[x) =
_
N
0
_
D/2
exp
_
1
N
0
|y x|
2
_
.
Since exp(.) is a monotonically increasing function, we obtain the
following formulation of the ML decision rule
x = dec
ML
(y) = argmax
xX
_
p
Y |X
(y[x)
_
= argmax
xX
_
exp
_
1
N
0
|y x|
2
_
_
= argmin
xX
_
|y x|
2
_
.
Thus, the ML decision rule for the AWGN vector channel selects
the vector in the signal constellation X that is the closest one to the
observation y, where the distance measure is the squared Euclidean
distance.
43
Implementation of the ML decision rule
The squared Euclidean distance can be expanded as
|y x|
2
= |y|
2
2y, x) + |x|
2
.
The term |y|
2
is irrelevant for the minimization with respect to x.
Thus we get the following equivalent formulation of the ML decision
rule:
x = dec
ML
(y) = argmin
xX
_
2y, x) + |x|
2
_
= argmax
xX
_
2y, x) |x|
2
_
= argmax
xX
_
y, x)
1
2
|x|
2
_
.
Block diagram of the ML decision rule for the AWGN vector-
channel:
.
.
.
.
.
.
PSfrag replacements
y s
1
s
M
1
2
|s
1
|
2
1
2
|s
M
|
2
x
select
largest
one
Vector correlator:
PSfrag replacements
a R
D
b R
D
a, b) =
D
d=1
a
d
b
d
R
44
Remark:
Some signal constellations X have the property that |x|
2
has the
same value for all x X. In these cases (but only then), the ML
decision rule simplies further to
x = dec
ML
(y) = argmax
xX
_
y, x)
_
.
2.8 Symbol-Error Probability
The error probability for modulation-symbols X, or for short, the
symbol-error probability is dened as
P
s
= Pr(

X ,= X) = Pr(dec(Y ) ,= X).
The symbol-error probability coincides with the block error prob-
ability
P
s
= Pr(

U ,= U) = Pr
_
[
U
1
, . . . ,

U
K
] ,= [U
1
, . . . , U
K
]
_
.
We compute P
s
in two steps:
(i) Compute the conditional symbol-error probability given
X = x for any x X:
Pr(

X ,= X[X = x).
(ii) Average over the signal constellation:
Pr(

X ,= X) =
xX
Pr(

X ,= X[X = x) Pr(X = x).
45
Remember: Decision regions for the symbol decision rule:
.
.
.
PSfrag replacements
R
D
T
1
T
2
T
M
_ _
, , ,
. . . s
1
s
2
s
M = X
Assume that x = s
m
is transmitted. Then,
Pr(

X ,= X[X = s
m
) = Pr(

X ,= s
m
[X = s
m
)
= Pr(Y / T
m
[X = s
m
)
= 1 Pr(Y T
m
[X = s
m
).
For the AWGN vector channel, the conditional probability density
function of Y given X = s
m
is
p
Y |X
(y[s
m
) =
_
N
0
_
D/2
exp
_
1
N
0
|y s
m
|
2
_
.
Hence,
Pr(Y T
m
[X = s
m
) =
_
D
m
p
Y |X
(y[s
m
)dy =
=
_
N
0
_
D/2
_
D
m
exp
_
1
N
0
|y s
m
|
2
_
dy.
The symbol-error probability is then obtained by averaging
over the signal constellation:
Pr(

X ,= X) = 1
M
m=1
Pr(Y T
m
[X = s
m
) Pr(X = s
m
).
46
If the modulation symbols are equiprobable, the formula simplies
to
Pr(

X ,= X) = 1
1
M
M
m=1
Pr(Y T
m
[X = s
m
).
2.9 Bit-Error Probability
The error probability for the source bits U
k
, or for short, the bit-
error probability is dened as
P
b
=
1
K
K
k=1
Pr(
U
k
,= U
k
),
i.e., as the average number of erroneous bits in the block
U = [
U
1
, . . . ,

U
K
].
Relation between the bit-error probability and the
symbol-error probability
Clearly, the relation between P
b
and P
s
depends on the encoding
function
enc : u = [u
1
, . . . , u
k
] x = enc(u).
The analytical derivation of the relation between P
b
and P
s
is
usually cumbersome. However, we can derive an upper bound
and a lower bound for P
b
when P
s
is given (and also vice versa).
These bounds are valid for any encoding function.
The lower bound provides a guideline for the design of a good
encoding function.
47
Lower bound:
If a symbol is erroneous, there is at least one erroneous bit in the
block of K bits. Therefore
P
b
1
K
P
s
.
Upper bound:
If a symbol is erroneous, there are at most K erroneous bits in the
block of K bits. Therefore
P
b
P
s
.
Combination:
Combining the lower bound and the upper bound yields
1
K
P
s
P
b
P
s
.
An ecient encoder should achieve the lower bound for the bit-
error probability, i.e.,
P
b
1
K
P
s
.
Remark:
The bounds can also be obtained by considering the union bound
of the events
U
1
,= U
1
,
U
2
,= U
2
, . . . ,
U
K
,= U
K
,
i.e, by relating the probability
Pr
_
U
1
,= U
1

U
2
,= U
2
. . .
U
K
,= U
K
_
to the probabilities
Pr
_
U
1
,= U
1
_
, Pr
_
U
2
,= U
2
_
, . . . , Pr
_
U
K
,= U
K
_
.
48
2.10 Gray Encoding
In Gray encoding, blocks of bits, u = [u
1
, . . . , u
K
], which dier in
only one bit u
k
are assigned to neighboring vectors (modulation
symbols) x in the signal constellation X.
Example: 4PAM (Pulse Amplitude Modulation)
Signal constellation: X = [3], [1], [+1], [+3].
PSfrag replacements
00 01 10 11
s
1
s
2
s
3
s
4
2 2
3
Motivation
When a decision error occurs, it is very likely that the transmitted
symbol and the detected symbol are neighbors. When the bit
blocks of neighboring symbols dier only in one bit, there will be
only one bit error per symbol error.
Eect
Gray encoding achieves
P
b
1
K
P
s
for high signal-to-noise ratios (SNR).
49
Sketch of the proof:
We consider the case K = 2, i.e., U = [U
1
, U
2
].
P
s
=Pr(

X ,= X)
=Pr
_
[
U
1
,

U
2
] ,= [U
1
, U
2
]
_
=Pr
_
U
1
,= U
1

U
2
,= U
2
_
=Pr(
U
1
,= U
1
) + Pr(
U
2
,= U
2
)
. .
2 P
b
Pr
_
U
1
,= U
1

U
2
,= U
2
_
. .
()
.
The term () is small compared to the other two terms if the SNR
is high. Hence
P
s
2 P
b
.
50
Part II
Digital Communication
Techniques
51
3 Pulse Amplitude Modulation
(PAM)
In pulse amplitude modulation (PAM), the information is conveyed
by the amplitude of the transmitted waveforms.
3.1 BPAM
PAM with only two waveforms is called binary PAM (BPAM).
3.1.1 Signaling Waveforms
Set of two rectangular waveforms with amplitudes A and A over
t [0, T], S :=
_
s
1
(t), s
2
(t)
_
:
PSfrag replacements
0 0
T T
A A
A A
t t
s
1
(t) s
2
(t)
As [S[ = 2, one bit is sucient to address the waveforms (K = 1).
Thus, the BPAM transmitter performs the mapping
U x(t)
with
U U = 0, 1 and x(t) S.
Remark: In the general case, the basic waveform may also have
another shape.
52
3.1.2 Decomposition of the Transmitter
Gram-Schmidt procedure
1
(t) =
1
E
1
s
1
(t)
with
E
1
=
T
_
0
_
s
1
(t)
_
2
dt =
T
_
0
A
2
dt = A
2
T.
Thus
1
(t) =
_
1
T
for t [0, T] ,
0 otherwise.
Both s
1
(t) and s
2
(t) are a linear combination of (t):
s
1
(t) = +
_
E
s

1
(t) = +A
T
1
(t),
s
2
(t) =
_
E
s

1
(t) = A
T
1
(t),
where
E
s
= A
T and E
s
= E
1
= E
2
. (The two waveforms
have the same energy.)
Signal constellation
Vector representation of the two waveforms:
s
1
(t) s
1
= s
1
= +
_
E
s
s
2
(t) s
2
= s
2
=
_
E
s
Thus the signal constellation is X =
_
E
s
, +
E
s
_
.
PSfrag replacements
0
E
s
E
s
1
s
1
s
2
53
Vector Encoder
The vector encoder is usually dened as
x = enc(u) =
_
+
E
s
for u = 0,
E
s
for u = 1.
The signal constellation comprises only two waveforms, and thus
per symbol and per bit, the average transmit energies and the
SNRs are the same:
E
b
= E
s
,
s
=
b
.
3.1.3 ML Decoding
ML symbol decision rule
dec
ML
(y) = argmin
x{
E
s
,+
E
s
}
|y x|
2
=
_
+
E
s
for y 0,
E
s
for y < 0.
Decision regions
T
2
= y R : y < 0, T
1
= y R : y 0.
PSfrag replacements
0
1
y
s
1
= +
E
s
s
2
=
E
s
T
1
T
2
Remark: The decision rule
dec
ML
(y) =
_
+
E
s
for y > 0,
E
s
for y 0.
is also a ML decision rule.
54
The mapping of the encoder inverse reads
u = enc
-1
( x) =
_
0 for x = +
E
s
,
1 for x =
E
s
.
Combining the (modulation) symbol decision rule and the encoder
inverse yields the ML decoder for BPAM:
u =
_
0 for y 0,
1 for y < 0.
3.1.4 Vector Representation of the Transceiver
BPAM transceiver:
PSfrag replacements
Encoder
Decoder
BSS
Sink
1
(t)
1
T
_
T
0
(.)dt
U
U
X(t) X
Y (t) Y
W(t)
0 +
E
s
1
E
s
T
1
0
T
2
1
Vector representation:
PSfrag replacements
Encoder
Decoder
BSS
Sink
1
T
_
T
0
(.)dt
U
U
X
Y
W ^(0, N
0
/2)
0 +
E
s
1
E
s
T
1
0
T
2
1
55
3.1.5 Symbol-Error Probability
Conditional probability densities
p
Y |X
(y[ +
_
E
s
) =
1
N
0
exp
_
1
N
0
[y
_
E
s
[
2
_
p
Y |X
(y[
_
E
s
) =
1
N
0
exp
_
1
N
0
[y +
_
E
s
[
2
_
Conditional symbol-error probability
Assume that X = x = +
E
s
is transmitted.
Pr(

X ,= X[X = +
_
E
s
) = Pr(Y T
2
[X = +
_
E
s
)
= Pr(Y < 0[X = +
_
E
s
)
=
1
N
0
0
_
exp
_
1
N
0
[y
_
E
s
[
2
_
dy
=
1
2E
s
/N
0
_
exp
_
1
2
z
2
_
dz,
where the last equality results from substituting
z = (y
_
E
s
)/
_
N
0
/2.
The Q-function is dened as
Q(v) :=
_
v
q(z)dz, q(z) =
1
2
exp
_
1
2
z
2
_
.
56
The function q(z) denotes a Gaussian pdf with zero mean and vari-
ance 1; i.e., it may be interpreted as the pdf of a random variable
Z ^(0, 1).
Using this denition, the conditional symbol-error probabilities
may be written as
Pr(

X ,= X[X = +
_
E
s
) = Q
_
_
2E
s
N
0
_
,
Pr(

X ,= X[X =
_
E
s
) = Q
_
_
2E
s
N
0
_
.
The second equation follows from symmetry.
Symbol-error probability
The symbol-error probability results from averaging over the con-
ditional symbol-error probabilities:
P
s
= Pr(

X ,= X)
= Pr(X = +
_
E
s
) Pr(

X ,= X[X = +
_
E
s
)
+ Pr(X =
_
E
s
) Pr(

X ,= X[X =
_
E
s
)
= Q
_
_
2E
s
N
0
_
= Q(
_
2
s
)
as Pr(X = +
E
s
) = Pr(X =
E
s
) = 1/2.
Alternative method to derive the conditional symbol-
error probability
Assume again that X =
E
s
is transmitted. Instead of consider-
ing the conditional pdf of Y , we directly use the pdf of the white
Gaussian noise W.
57
Pr(

X ,= X[X = +
_
E
s
) = Pr(Y < 0[X = +
_
E
s
)
= Pr(
_
E
s
+ W < 0[X = +
_
E
s
)
= Pr(W <
_
E
s
[X = +
_
E
s
)
= Pr(W <
_
E
s
)
= Pr(W
<
_
2E
s
/N
0
)
= Q(
_
2E
s
/N
0
),
where we substituted W
= W/
_
N
0
/2; notice that
W
^(0, 1). A plot of the symbol-error probability can found

in [1, p. 412].
The initial event (here Y < 0) is transformed into an equivalent
event (here W
<
_
2E
s
/N
0
), of which the probability can
easily be calculated. This method is frequently applied to deter-
mine symbol-error probabilities.
3.1.6 Bit-Error Probability
In BPAM, each waveform conveys one source bit. Accordingly,
each symbol-error leads to exactly one bit-error, and thus
P
b
= Pr(
U ,= U) = Pr(

X ,= X) = P
s
.
Due to the same reason, the transmit energy per symbol is equal
to the transmit energy per bit, and so also the SNR per symbol is
equal to the SNR per bit:
E
b
= E
s
,
s
=
b
.
Therefore, the bit-error probability (BEP) of BPAM results as
P
b
= Q(
_
2
b
).
Notice: P
b
10
5
is achieved for
b
9.6 dB.
58
3.2 MPAM
In the general case of M-ary pulse amplitude modulation (MPAM)
with M = 2
K
, input blocks of K bits modulate the amplitude of
a unique pulse.
Waveform encoder:
u = [u
1
, . . . , u
K
] x(t)
U = 0, 1
K
S =
_
s
1
(t), . . . , s
M
(t)
_
,
where
s
m
(t) =
_
(2mM 1) Ag(t) for t [0, T],
0 otherwise,
A R, and g(t) is a predened pulse of duration T,
g(t) = 0 for t / [0, T].
Values of the factors (2mM 1)A that are weighting g(t):
(M 1)A, (M 3)A, . . . , A, A, . . . , (M 3)A, (M 1)A.
Example: 4PAM with rectangular pulse.
Using the rectangular pulse
g(t) =
_
1 for t [0, T],
0 otherwise,
the resulting waveforms are
S =
_
3Ag(t), Ag(t), Ag(t), 3Ag(t)
_
.
3
59
Gram-Schmidt procedure:
Orthonormal function
(t) =
1
E
1
s
1
(t) =
1
_
E
g
g(t),
E
1
=
_
T
0
_
s
1
(t)
_
2
dt, E
g
=
_
T
0
_
g(t)
_
2
dt.
Every waveform in S is a linear combination of (only) (t):
s
m
(t) = (2mM 1) A
_
E
g
. .
d
(t)
= (2mM 1)d (t).
s
1
(t) s
1
= (M 1)d
s
2
(t) s
2
= (M 3)d

s
m
(t) s
m
= (2mM 1)d

s
M1
(t) s
M1
= (M 3)d
s
M
(t) s
M
= (M 1)d
X =
_
s
1
, s
2
, . . . , s
M
_
=
_
(M 1)d, (M 3)d, . . . , d, d, . . .
. . . , (M 3)d, (M 1)d
_
.
60
Example: M = 2, 4, 8
PSfrag replacements
0 1
00 01 10 11
000 001 010 011 100 101 110 111
s
1
s
1
s
1
s
2
s
2
s
2
s
3
s
3
s
4
s
4
s
5
s
6
s
7
s
8
3
Vector Encoder
The vector encoder
enc : u = [u
1
, . . . , u
K
] x = enc(u)
U = 0, 1
K
X =
_
(M 1)d, . . . , (M 1)d
_
is any Gray encoding function.
Energy and SNR
The source bits are generated by a BSS, and so the symbols of the
signal constellation X are equiprobable:
Pr(X = s
m
) =
1
M
for all m = 1, 2, . . . , M.
61
The average symbol energy results as
E
s
= E
_
X
2
=
1
M
M
m=1
(s
m
)
2
=
d
2
M
M
m=1
(2mM 1)
2
=
d
2
M
1
3
M(M
2
1)
Thus,
E
s
=
M
2
1
3
d
2
d =
3E
s
M
2
1
As M = 2
K
, the average energy per transmitted source bit results
as
E
b
=
E
s
K
=
M
2
1
3 log
2
M
d
2
d =
_
3 log
2
M
M
2
1
E
b
.
The SNR per symbol and the SNR per bit are related as
b
=
1
K
s
=
1
log
2
M
s
.
62
3.2.3 ML Decoding
dec
ML
(y) = argmin
xX
|y x|
2
Decision regions
Example: Decision regions for M = 4.
PSfrag replacements
T
1
T
2
T
3
T
4
s
1
s
2
s
3
s
4
0 2d 2d
y
3
Starting from the special cases M = 2, 4, 8, the decision regions
can be inferred:
T
1
= y R : y (M 2)d,
T
2
= y R : (M 2)d < y (M 4)d,

T
m
= y R : (2mM 2)d < y (2mM)d,

T
M1
= y R : (M 4)d < y (M 2)d,
T
M
= y R : (M 2)d < y.
Accordingly, the ML symbol decision rule may also be written as
dec
ML
(y) = s
m
for y T
m
.
ML decoder for MPAM
u = enc
-1
(dec
ML
(y))
63
BPAM transceiver:
PSfrag replacements
Gray
Encoder
Decoder
BSS
Sink
(t)
1
T
_
T
0
(.)dt
U
U
X(t) X
Y (t) Y
W(t)
u
m
s
m
T
m
u
m
Vector representation:
PSfrag replacements
Gray
Encoder
Decoder
BSS
Sink
1
T
_
T
0
(.)dt
U
U
X
Y
W ^(0, N
0
/2)
u
m
s
m
T
m
u
m
Notice: u
m
denotes the block of source bits corresponding to s
m
,
i.e.,
s
m
= enc(u
m
), u
m
= enc
-1
(s
m
).
64
Two cases must be considered:
1. The transmitted symbol is an edge symbol (a symbol at the
edge of the signal constellation):
X s
1
, s
M
.
2. The transmitted symbol is not an edge symbol:
X s
2
, s
3
, . . . , s
M1
.
Case 1: X s
1
, s
M
Assume rst X = s
1
. Then
Pr(

X ,= X[X = s
1
) = Pr(Y / T
1
[X = s
1
)
= Pr(Y > s
1
+ d[X = s
1
)
= Pr(W > d)
= Pr
_
W
_
N
0
/2
>
d
_
N
0
/2
_
= Q
_
d
_
N
0
/2
_
.
Notice that W/
_
N
0
/2 ^(0, 1).
By symmetry,
Pr(

X ,= X[X = s
M
) = Q
_
d
_
N
0
/2
_
.
65
Case 2: X s
2
, . . . , s
M1
Assume X = s
m
. Then
Pr(

X ,= X[X = s
m
) = Pr(Y / T
m
[X = s
m
)
= Pr(W < d or W > d)
= Pr(W < d) + Pr(W > d)
= Pr
_
W
_
N
0
/2
<
d
_
N
0
/2
_
+ Pr
_
W
_
N
0
/2
>
d
_
N
0
/2
_
= Q
_
d
_
N
0
/2
_
+ Q
_
d
_
N
0
/2
_
= 2 Q
_
d
_
N
0
/2
_
.
Averaging over the symbols, one obtains
P
s
=
1
M
M
m=1
Pr(

X ,= X[X = s
m
)
=
2
M
Q
_
d
_
N
0
/2
_
+
2(M 2)
M
Q
_
d
_
N
0
/2
_
= 2
M 1
M
Q
_
d
_
N
0
/2
_
.
Expressing the symbol-error probability as a function of the SNR
per symbol,
s
, or the SNR per bit,
b
, yields
P
s
= 2
M 1
M
Q
_
_
6
M
2
1
s
_
= 2
M 1
M
Q
_
_
6 log
2
M
M
2
1
b
_
.
66
Plots of the symbol-error probabilities can be found in [1, p. 412].
Remark: When comparing BPAM and 4PAM for the same P
s
,
the SNRs per bit,
b
, dier by 10 log
10
(5/2) 4 dB. (Proof?)
When Gray encoding is used, we have the relation
P
b
P
s
log
2
M
= 2
M 1
M log
2
M
Q
_
_
6 log
2
M
M
2
1
b
_
for high SNR
b
= E
b
/N
0
.
67
4 Pulse Position Modulation (PPM)
In pulse position modulation (PPM), the information is conveyed
by the position of the transmitted waveforms.
4.1 BPPM
PPM with only two waveforms is called binary PPM (BPPM).
Set of two rectangular waveforms with amplitude A and dura-
tion T/2 over t [0, T], S :=
_
s
1
(t), s
2
(t)
_
:
PSfrag replacements
0 0
T T
T
2
T
2
A A
A A
t t
s
1
(t) s
2
(t)
As the cardinality of the set is [S[ = 2, one bit is sucient to
address each waveforms, i.e., K = 1. Thus, the BPPM transmitter
performs the mapping
U x(t)
with
U U = 0, 1 and x(t) S.
Remark: In the general case, the basic waveform may also have
another shape.
68
The waveforms s
1
(t) and s
2
(t) are orthogonal and have the same
energy
E
s
=
T
_
0
s
2
1
(t)dt =
T
_
0
s
2
2
(t)dt = A
2
T
2
.
Notice that
E
s
= A
_
T/2.
The two orthonormal functions and the corresponding waveform
representations are
1
(t) =
1
E
s
s
1
(t), s
1
(t) =
_
E
s

1
(t) + 0
2
(t),
2
(t) =
1
E
s
s
2
(t), s
2
(t) = 0
1
(t) +
_
E
s

2
(t).
Vector representation of the two waveforms:
s
1
(t) s
1
= (
_
E
s
, 0)
T
,
s
2
(t) s
2
= (0,
_
E
s
)
T
.
Thus the signal constellation is X =
_
(
E
s
, 0)
T
, (0,
E
s
)
T
_
.
PSfrag replacements
0
0
E
s
E
s
2E
s
2
s
1
s
2
69
Vector Encoder
The vector encoder is dened as
x = enc(u) =
_
(
E
s
, 0)
T
for u = 0,
(0,
E
s
)
T
for u = 1.
The signal constellation comprises only two waveforms, and thus
per (modulation) symbol and per (source) bit, the average transmit
energies and the SNRs are the same:
E
b
= E
s
,
s
=
b
.
4.1.3 ML Decoding
dec
ML
(y) = argmin
x{s
1
,s
2
}
|y x|
2
=
_
(
E
s
, 0)
T
for y
1
y
2
,
(0,
E
s
)
T
for y
1
< y
2
.
Notice: |y s
1
|
2
|y s
2
|
2
y
1
y
2
.
Decision regions
T
1
= y R
2
: y
1
y
2
, T
2
= y R
2
: y
1
< y
2
.
PSfrag replacements
0
0
E
s
E
s
2E
s
y
1
y
2
s
1
s
2
T
1
T
2
70
Combining the (modulation) symbol decision rule and the encoder
inverse yields the ML decoder for BPPM:
u = enc
-1
(dec
ML
(y)) =
_
0 for y
1
y
2
,
1 for y
1
< y
2
.
The ML decoder may be implemented by
PSfrag replacements
U
Y
1
Y
2
Z
z 0 0
z < 0 1
Notice that Z = Y
1
Y
2
is then the decision variable.
71
BPPM transceiver for the waveform AWGN channel
PSfrag replacements
Vector
Vector
Encoder
Decoder
BSS
Sink
0 (
E
s
, 0)
T
1 (0,
E
s
)
T
1
(t)
2
(t)
2
T
T/2
0
(.)dt
2
T
T
T/2
(.)dt
T
1
0
T
2
1
U
U
X(t)
X
1
X
2
Y (t)
Y
1
Y
2
W(t)
Vector representation including a AWGN vector channel
PSfrag replacements
Vector
Vector
Encoder
Decoder
BSS
Sink
0 (
E
s
, 0)
T
1 (0,
E
s
)
T
1
(t)
2
(t)
2
T
T/2
0
(.)dt
_
2
T
_
T
T/2
(.)dt
T
1
0
T
2
1
U
U
X
1
X
2
Y
1
Y
2
W
1
^(0, N
0
/2)
W
2
^(0, N
0
/2)
72
Assume that X = s
1
= (
E
s
, 0)
T
is transmitted.
Pr(

X ,= X[X = s
1
) = Pr(Y T
2
[X = s
1
)
= Pr(s
1
+ W T
2
[X = s
1
)
= Pr(s
1
+ W T
2
)
The noise vector W may be represented in the coordinate system
(
1
,
2
) or in the rotated coordinate system (
1
,
2
). The latter is
more suited to computing the above probability, as only the noise
component along the
2
-axis is relevant.
PSfrag replacements
0
E
s
E
s
_
E
s
/2
2
= /4
s
1
s
2 s
1
+ w
w
1
w
2
w
1
w
2
T
1
T
2
The elements of the noise vector in the rotated coordinate system
have the same statistical properties as in the non-rotated coordi-
nate system (see Appendix B), i.e.,
W
1
^(0, N
0
/2), W
2
^(0, N
0
/2).
73
Thus, we obtain
Pr(

X ,= X[X = s
1
) = Pr(W
2
>
_
E
s
/2)
= Pr(
W
2
_
N
0
/2
>
_
E
s
N
0
)
= Q(
_
E
s
/N
0
) = Q(
s
).
By symmetry,
Pr(

X ,= X[X = s
2
) = Q(
s
).
Averaging over the conditional symbol-error probabilities, we ob-
tain
P
s
= Pr(

X ,= X) = Q(
s
).
For BPPM, E
b
= E
s
,
s
=
b
, P
b
= P
s
(see above), and so the
bit-error probability (BEP) results as
P
b
= Pr(
U ,= U) = Q(
b
).
A plot of the BEP versus
b
may be found in [1, p. 426]
Remark
Due to their signal constellations, BPPM is called orthogonal sig-
naling and BPAM is called antipodal signaling. Their BEPs are
given by
P
b,BPPM
= Q(
b
), P
b,BPAM
= Q(
_
2
b
).
When comparing the SNR for a certain BEP, BPPM requires 3 dB
more than BPAM. For a plot, see [1, p. 409]
74
4.2 MPPM
In the general case of M-ary pulse position modulation (MPPM)
with M = 2
K
, input blocks of K bits determine the position of a
unique pulse.
Waveform encoder
u = [u
1
, . . . , u
K
] x(t)
U = 0, 1
K
S =
_
s
1
(t), . . . , s
M
(t)
_
,
where
s
m
(t) = g
_
t
(m1)T
M
_
,
m = 1, 2, . . . , M, and g(t) is a predened pulse of duration T/M,
g(t) = 0 for t / [0, T/M].
The waveforms are orthogonal and have the same energy as g(t),
E
s
= E
g
=
T
_
0
g
2
(t)dt.
Example: (M=4)
Consider a rectangular pulse g(t) with amplitude A and du-
ration T/4. Then, we have the set of four waveforms, S =
s
1
(t), s
2
(t), s
3
(t), s
4
(t):
PSfrag replacements
0
0
T
A
t
s
1
(t)
PSfrag replacements
0
0
T
A
t
s
1
(t)
s
2
(t)
PSfrag replacements
0
0
T
A
t
s
1
(t)
s
2
(t)
s
3
(t)
PSfrag replacements
0
0
T
A
t
s
1
(t)
s
2
(t)
s
3
(t)
s
4
(t)
3
75
As all waveforms are orthogonal, we have M orthonormal func-
tions:
m
(t) =
1
E
s
s
m
(t),
m = 1, 2, . . . , M.
The waveforms and the orthonormal functions are related as
s
m
(t) =
_
E
s

m
(t),
m = 1, 2, . . . , M.
Waveforms and their vector representations:
s
1
(t) s
1
= (
M
..
_
E
s
, 0, 0, . . . , 0)
T
s
2
(t) s
2
= (0,
_
E
s
, 0, . . . , 0)
T
.
.
.
s
M
(t) s
M
= (0, 0, 0, . . . ,
_
E
s
)
T
Signal constellation:
X =
_
s
1
, s
2
, . . . , s
M
_
R
M
.
The modulation symbols x X are orthogonal, and they are
placed on corners of an M-dimensional hypercube.
76
Example: M = 3
The signal constellation
X =
_
(
_
E
s
, 0, 0)
T
, (0,
_
E
s
, 0)
T
, (0,
_
E
s
, 0)
T
_
R
3
may be visualized using a cube with side-length
E
s
. 3
Vector Encoder
enc : u = [u
1
, . . . , u
K
] x = enc(u)
U = 0, 1
K
X =
_
s
1
, s
2
, . . . , s
M
_
Energy and SNR
The symbol energy is E
s
, and as M = 2
K
, the average energy per
transmitted source bit results as
E
b
=
E
s
K
The SNR per symbol and the SNR per bit are related as
b
=
1
K
s
=
1
log
2
M
s
.
77
4.2.3 ML Decoding
dec
ML
(y) = argmin
xX
|y x|
2
= argmin
xX
|y|
2
2y, x) + |x|
2
..
E
s
= argmax
xX
y, x)
ML decoder for MPAM
u = enc
-1
(dec
ML
(y))
Implementation of the ML detector using vector correlators:
.
.
.
PSfrag replacements
enc
-1
y
s
1
s
M
y, s
1
)
y, s
M
)
x u
select
max.
Notice that u = [ u
1
, . . . , u
K
].
78
Assume rst that X = s
1
is transmitted. Then
Pr(

X ,= X[X = s
1
) = 1 Pr(

X = X[X = s
1
)
= 1 Pr
_
y, s
1
) y, s
m
) for all m ,= 1[X = s
1
_
.
For X = s
1
= (
E
s
, 0, . . . , 0), the received vector is
Y = s
1
+ W = (
_
E
s
, 0, . . . , 0) + (W
1
, . . . , W
M
),
so that
y, s
m
) =
_
E
s
+
E
s
W
1
for m = 1,
E
s
W
m
for m = 2, . . . , M.
Therefore, the symbol-error probability may be written as follows:
Pr(

X ,= X[X = s
1
) =
= 1 Pr
_
E
s
+
_
E
s
W
1
_
E
s
W
m
for all m ,= 1
_
= 1 Pr
_
E
s
N
0
/2
+
W
1
_
N
0
/2
. .
W
W
m
_
N
0
/2
. .
W
m
for all m ,= 1
_
= 1 Pr
_
W
_
2E
s
N
0
+ W
1
for all m ,= 1
_
with
W
= (W
1
, . . . , W
M
)
T
^(0, I),
i.e., the components W
m
are independent and Gaussian distributed
with zero mean and variance 1.
79
The probability is now written using the pdf of W
1
:
Pr
_
W
_
2E
s
N
0
+ W
1
for all m ,= 1
_
=
=
+
_
Pr
_
W
_
2E
s
N
0
+ W
1
for all m ,= 1
1
= v
_
. .
()
p
W
1
(v)dv.
The events corresponding to dierent values of m are independent,
and thus
() =
M
m=2
Pr
_
W
_
2E
s
N
0
+ W
1
= v
_
. .
Pr
_
W
_
2E
s
N
0
+ v
_
=
M
m=2
Q
_
_
2E
s
N
0
+ v
_
= Q
M1
(
_
2
s
+ v).
Using this result and inserting p
W
1
(v), we obtain the conditional
symbol-error probability
Pr(

X ,= X[X = s
1
) =
=
+
_
_
1 Q
M1
(
_
2
s
+ v)
_
1
2
exp
_
v
2
2
_
dv
The above expression is independent of s
1
and thus valid for all
x X. Therefore, the average symbol-error probability results as
P
s
=
+
_
_
1 Q
M1
(
_
2
s
+ v)
_
1
2
exp
_
v
2
2
_
dv.
80
Hamming distance
The Hamming distance d
H
(u, u
) between two length-K vectors

u and u
is the number of positions in which they dier:

d
H
(u, u
) =
K
k=1
(u
k
, u
k
),
where the indicator function is dened as
(u, u
) =
_
1 for u = u
,
0 for u ,= u
.
Example: 4PPM
Let the blocks of source bits and the modulation symbols be asso-
ciated as
u
1
= enc
-1
(s
1
) = [0, 0]
u
2
= enc
-1
(s
2
) = [0, 1]
u
3
= enc
-1
(s
3
) = [1, 0]
u
4
= enc
-1
(s
4
) = [1, 1].
Then we have for example
d
H
(u
1
, u
1
) = 0, d
H
(u
1
, u
2
) = 1,
d
H
(u
2
, u
3
) = 2, d
H
(u
2
, u
4
) = 1.
3
81
Conditional bit-error probability
Assume that X = s
1
, and let
u
m
= [u
m,1
, . . . , u
m,K
] = enc
-1
(s
m
).
The conditional bit-error probability is dened as
Pr(U ,=

U[X = s
1
) =
1
K
K
k=1
Pr(U
k
,=

U
k
[X = s
1
).
The argument of the sum may be expanded as
Pr(U
k
,=

U
k
[X = s
1
) =
=
M
m=1
Pr(U
k
,=

U
k
[
X = s
m
, X = s
1
) Pr(

X = s
m
[X = s
1
).
Since
X = s
1
U = u
1
U
k
= u
1,k
,
X = s
m

U = u
m

U
k
= u
m,k
,
we have
Pr(U
k
,=

U
k
[
X = s
m
, X = s
1
) =
= Pr(u
1,k
,= u
m,k
[
X = s
m
, X = s
1
)
= (u
1,k
, u
m,k
).
Using this result in the expression for the conditional bit-error
probability, we obtain
Pr(U ,=

U[X = s
1
)
=
1
K
K
k=1
M
m=1
(u
1,k
, u
m,k
) Pr(

X = s
m
[X = s
1
)
=
1
K
M
m=1
K
k=1
(u
1,k
, u
m,k
)
. .
d
H
(u
1
, u
m
)
Pr(

X = s
m
[X = s
1
)
82
Due to the symmetry of the signal constellation, the probabilities
of making a specic decision error are identical, i.e.,
Pr(

X = s
m
[X = s
m
) =
P
s
M 1
=
P
s
2
K
1
for all m ,= m
. Using this in the above sum yields

Pr(U ,=

U[X = s
1
) =
1
K

P
s
2
K
1
M
m=1
d
H
(u
1
, u
m
).
Without loss of generality, we assume u
1
= 0. The sum over the
Hamming distances becomes then the overall number of ones in all
bit blocks of length K.
Example:
Consider K = 3 and write the binary vectors of length 3 in the
columns of a matrix:
_
u
T
1
, . . . , u
T
8
=
_
_
0 0 0 0 1 1 1 1
0 0 1 1 0 0 1 1
0 1 0 1 0 1 0 1
_
_
.
In every row, half of the entries are ones. The overall number of
ones is thus 3 2
31
. 3
Generalizing the example, the above sum can shown to be
M
m=1
d
H
(0, u
m
) = K 2
K1
.
83
Applying this result, we obtain the conditional bit-error probability
Pr(U ,=

U[X = s
1
) =
2
K1
2
K
1
P
s
,
which is independent of s
1
and thus the same for all transmitted
symbols X = s
m
, m = 1, . . . , M.
Average bit-error probability
From above, we obtain
Pr(U ,=

U) =
2
K1
2
K
1
P
s
.
Plots of the BEPs may be found in [1, p. 426]
Notice that for large K,
Pr(U ,=

U)
1
2
P
s
.
Due to the signal constellation, Gray mapping is not possible.
84
Remarks
(i) Provided that the SNR
b
= E
b
/N
0
is larger than ln 2 (cor-
responding to 1.6 dB), the BEP of MPPM can be
made arbitrarily small by selecting M large enough.
The price to pay is an increasing decoding delay and an in-
creasing required bandwidth.
Let the transmission rate be denoted by
R =
K
T
=
log
2
M
T
[bit/sec].
Decoding delay:
T =
1
R
log
2
M.
Hence, for a xed rate R, T for M .
Required bandwidth:
B = B
g

1
T/M
=
M
T
= R
M
log
2
M
,
where B
g
denotes the bandwidth of the pulse g(t).
Hence, for a xed rate R, B for M .
(ii) The power of the MPPM signal is
P =
E
b
K
T
= E
b
R.
Hence, the condition
b
> ln 2 is equivalent to the condition
R <
1
ln 2
P
N
0
= C
.
The value C
denotes the capacity of the (innite band-

width) AWGN channel (see Appendix C).
85
5 Phase-Shift Keying (PSK)
In phase-shift keying (PSK), the waveforms are sines with the same
frequency, and the information is conveyed by the phase of the
waveforms.
5.1 Signaling Waveforms
Set of sines s
m
(t) with the same frequency f
0
, duration T, and
dierent phases
m
:
s
m
(t) =
_
2P cos
_
2f
0
t + 2
m1
M
. .
m
_
for t [0, T],
0 elsewhere,
m = 1, 2, . . . , M. The phase can take the values
_
0, 2
1
M
, 2
2
M
, . . . , 2
M1
M
_
Hence, the digital information is embedded in the phase of a carrier.
The frequency f
0
is called the carrier frequency.
For M = 2, this modulation scheme is called binary phase-
shift keying (BPSK), and for M = 4, it is called quadrature
phase-shift keying. (QPSK).
All waveforms have the same energy
E
s
=
s
2
m
(t)dt = PT.
(Notice: cos
2
x = 1 + cos 2x.)
86
Usually, f
0
1/T. For simplifying the theoretical analysis, we
assume that f
0
= L/T for some large L N.
BPSK
For M = 2, we obtain
m
0, and thus the two waveforms
s
1
(t) =
2P cos
_
2f
0
t
_
,
s
2
(t) =
2P cos
_
2f
0
t +
_
= s
1
(t)
for t [0, T].
QPSK
m
0,

2
, ,
3
2
and thus the waveforms
are
s
1
(t) = +
2P cos
_
2f
0
t
_
,
s
2
(t) =
2P sin
_
2f
0
t
_
,
s
3
(t) =
2P cos
_
2f
0
t
_
,
s
4
(t) = +
2P sin
_
2f
0
t
_
for t [0, T].
87
5.2 Signal Constellation
BPSK
Orthonormal functions:
1
(t) =
1
E
s
s
1
(t) =
_
2
T
cos
_
2f
0
t
_
for t [0, T].
Geometrical representation of the waveforms:
s
1
(t) = +
PT
1
(t) s
1
= +
PT = +
_
E
s
s
2
(t) =
PT
1
(t) s
2
=
PT =
_
E
s
PSfrag replacements
0
E
s
E
s
1
s
1
s
2
Vector encoder:
x = enc(u) =
_
s
1
for u = 1,
s
2
for u = 0.
88
QPSK
1
(t) =
1
E
s
s
1
(t) = +
_
2
T
cos
_
2f
0
t
_
2
(t) =
1
E
s
s
2
(t) =
_
2
T
sin
_
2f
0
t
_
for t [0, T].
s
1
(t) = +
PT
1
(t) s
1
= (
_
E
s
, 0)
T
s
2
(t) =
_
E
s
2
(t) s
2
= (0,
_
E
s
)
T
s
3
(t) =
PT
1
(t) s
3
= (
_
E
s
, 0)
T
s
4
(t) =
_
E
s
2
(t) s
4
= (0,
_
E
s
)
T
with E
s
= PT.
PSfrag replacements
2
s
1
s
2
s
3
s
4
Vector encoder: any Gray encoder
89
General Case
The expression for s
m
(t) can be expanded as follows:
s
m
(t) =
_
E
s
cos
_
2f
0
t +
m
_
=
_
E
s
cos
m
cos(2f
0
t)
_
E
s
sin
m
sin(2f
0
t),
where
m
= 2
m1
M
, m = 1, 2, . . . , M.
(Notice: cos(x + y) = cos x cos y sin x sin y.)
1
(t) =
_
2
T
cos(2f
0
t)
2
(t) =
_
2
T
sin
_
2f
0
t
_
for t [0, T].
s
m
(t) =
_
E
s
cos
m

1
(t) +
_
E
s
sin
m

2
(t)
s
m
=
_
_
E
s
cos
m
,
_
E
s
sin
m
_
T
,
m = 1, 2, . . . , M. Remember: E
s
= PT.
X =
_
_
_
E
s
cos 2
m1
M
,
_
E
s
sin 2
m1
M
_
T
: m = 1, 2, . . . , M
_
Example: 8PSK
m
=
m1
4
, m = 1, 2, . . . , 8. (Figure?) 3
90
5.3 ML Decoding
dec
ML
(y) = argmin
xX
|y x|
2
ML decoder for MPSK
u = [ u
1
, u
2
, . . . , u
K
] = enc
-1
(dec
ML
(y))
BPSK
BPSK has the same signal constellation as BPAM. Thus, the ML
detectors for these two modulation schemes are the same.
PSfrag replacements
0
1
y
s
1
(u = 1)
s
2
(u = 0)
T
1
T
2
Implementation of the ML detector:
PSfrag replacements
U
Y
y < 0 0
y 0 1
91
QPSK
Decision regions for the symbols s
m
and corresponding bit blocks
u = [u
1
, u
2
] (Gray encoding):
PSfrag replacements
y
1
y
2
s
1
s
2
s
3
s
4
00
01
10
11
T
1
T
2
T
3
T
4
Decision rule for the data bits:
u
1
=
_
0 for y T
1
T
2
y
1
+ y
2
> 0
1 for y T
3
T
4
y
1
+ y
2
< 0
u
2
=
_
0 for y T
1
T
4
y
2
y
1
< 0
1 for y T
2
T
3
y
2
y
1
> 0
Implementation of the ML decoder:
PSfrag replacements
Y
1
Y
2
Z
1
Z
2
U
1
U
2
z
1
> 0 0
z
1
< 0 1
z
2
< 0 0
z
2
> 0 1
92
General Case
The decision regions for the general case can be determined in a
straight-forward way:
PSfrag replacements
y
1
y
2
s
1
s
2
s
M
T
1
/M
2/M
93
5.4 Vector Representation of the MPSK
Transceiver
PSfrag replacements
Vector
Vector
Encoder
Decoder
AWGN Vector Channel
U X(t)
X
1
X
2
W(t)
Y (t)
Y
1
Y
2
U
T
T
2
T
2
T
cos 2f
0
t
cos 2f
0
t
sin 2f
0
t
sin 2f
0
t
2
T
T
0
(.)dt
2
T
T
0
(.)dt
AWGN Vector Channel:
PSfrag replacements
X
1
X
2
W
1
^(0, N
0
/2)
W
2
^(0, N
0
/2)
Y
1
Y
2
94
5.5 Symbol-Error Probability
BPSK
Since BPSK has the same signal constellation as BPAM, the two
modulation schemes have the same symbol-error probability:
P
s
= Q
__
2E
s
N
0
_
= Q(
_
2
s
).
QPSK
Assume that X = s
1
is transmitted. Then, the conditional
symbol-error probability is given by
Pr(

X ,= X[X = s
1
) = Pr(y / T
1
[X = s
1
)
= 1 Pr(y T
1
[X = s
1
)
= 1 Pr(s
1
+ W T
1
).
For computing this probability, the noise vector is projected on the
rotated coordinate system (
1
,
2
) (compare BPPM):
PSfrag replacements
0
E
s
_
E
s
/2
y
1
y
2
2
/4
s
1
s
2
s
1
+ w
w
1
w
2
w
1
w
2
T
1
T
2
T
3
T
4
d
e
c
i
s
i
o
n
b
o
u
n
d
a
r
y
d
e
c
i
s
i
o
n
b
o
u
n
d
a
r
y
95
The probability may then be written as
Pr(s
1
+ W T
1
) = Pr
_
W
1
>
_
E
s
2
and W
2
<
_
E
s
2
_
= Pr
_
W
1
>
_
E
s
2
_
Pr
_
W
2
<
_
E
s
2
_
= Pr
_
W
N
0
/2
>
_
E
s
N
0
_
Pr
_
W
N
0
/2
<
_
E
s
N
0
_
=
_
1 Q(
s
)
__
1 Q(
s
)
_
=
_
1 Q(
s
)
_
2
.
Hence,
Pr(

X ,= X[X = s
1
) = 1
_
1 Q(
s
)
_
2
= 2 Q(
s
) Q
2
(
s
).
By symmetry, the four conditional symbol-error probabilities are
identical, and thus the average symbol-error probability results as
P
s
= 2 Q(
s
) Q
2
(
s
).
96
General Case
A closed-form expression for P
s
does not exist. It can be upper-
bounded using an union-bound technique, or it can be numerically
computed using an integral expression. (See literature.) A plot of
P
s
versus the SNR can be found in [1, p. 416].
Remark: Union Bound
For two random events A and B, we have
Pr(A B) = Pr(A) + Pr(B) Pr(A B)
. .
0
Pr(A) + Pr(B).
The expression on the right-hand side is called the union bound for
the probability on the left-hand side. In a similar way, we obtain
the union bound for multiple events:
Pr(A
1
A
2
A
M
) Pr(A
1
) + Pr(A
2
) + + Pr(A
M
).
Each of these events may denote the case that the received vector
is within a certain decision region.
97
BPSK
Since E
b
= E
s
,
s
=
b
, P
b
= P
s
, the bit-error probability is given
by
P
b
= Q(
_
2
b
).
QPSK
For Gray encoding, P
b
can directly be computed.
Assume rst that X = s
1
is transmitted, which corresponds to
[U
1
, U
2
] = [0, 0]. Then, the conditional bit-error probability is
given by
Pr(
U ,= U[X = s
1
) =
=
1
2
_
Pr(
U
1
,= U
1
[X = s
1
) + Pr(
U
2
,= U
2
[X = s
1
)
_
=
1
2
_
Pr(
U
1
,= 0[X = s
1
) + Pr(
U
2
,= 0[X = s
1
)
_
=
1
2
_
Pr(
U
1
= 1[X = s
1
) + Pr(
U
2
= 1[X = s
1
)
_
.
From the direct decision rule for data bits, we have
U
1
= 1 Y T
3
T
4
,
U
2
= 1 Y T
2
T
3
.
98
Thus,
Pr(
U ,= U[X = s
1
) =
=
1
2
_
Pr(Y T
3
T
4
) + Pr(Y T
2
T
3
)
_
=
1
2
_
Pr
_
W
1
<
_
E
s
2
_
+ Pr
_
W
2
>
_
E
s
2
_
_
=
1
2
_
Pr
_
W
N
0
/2
<
_
E
s
N
0
_
+ Pr
_
W
N
0
/2
>
_
E
s
N
0
_
_
= Q(
s
).
By symmetry, all four conditional bit-error probabilities are iden-
tical. Since two data bits are transmitted per modulation sym-
bol (waveform) in QPSK, we have the relations E
s
= 2E
b
and
s
= 2
b
. Therefore,
P
b
= Q(
_
2
b
).
(Remember: P
b
10
5
for
b
9.6 dB.)
99
Comparison of QPSK and BPSK
(i) QPSK has exactly the same bit-error probability as BPSK.
(ii) Relation between symbol duration T and (average) bit dura-
tion T
b
= T/K:
BPSK: T = T
b
,
QPSK: T = 2T
b
.
Hence, for a xed symbol duration T, the bit rate of QPSK
is twice as fast as that of BPSK.
General Case
Since Gray encoding is used, we have
P
b
1
log
2
M
P
s
for high SNR.
The symbol-error probability may be determined using one of the
methods mentioned above.
100
6 Oset Quadrature Phase-Shift
Keying (OQPSK)
Oset QPSK is a variant of QPSK. It has the same performance
as QPSK, but the requirements on the transmitter and receiver
hardware are lower.
6.1 Signaling Scheme
Consider QPSK with the following tilted signal constella-
tion and Gray encoding:
PSfrag replacements
E
s
_
E
s
/2
2
00
01
10 11
The phase of QPSK can change by maximal . This is the case
for instance, when two consecutive 2-bit blocks are [00], [11] or
[01], [10]. These phase changes put high requirements on the dy-
namic behavior of the RF-part of the QPSK transmitter and re-
ceiver.
Phase changes of can be avoided by staggering the quadrature
branch of QPSK by T/2 = T
b
. The resulting modulation scheme
is called staggered QPSK or oset QPSK (OQPSK).
101
Without signicant loss of generality, we assume
f
0
=
2L
T
for some L 1, L N.
Remark
The
1
component of the modulation symbol is called the in-phase
component (I), and the
2
component is called the quadrature
component (Q).
6.2 Transceiver
We consider communication over an AWGN channel.
PSfrag replacements
Vector
Vector
Encoder
Decoder
AWGN Vector Channel
[U
1
, U
2
]
X(t)
X
1
X
2
W(t)
Y (t)
Y
1
Y
2
[
U
1
,

U
2
]
T
T
2
3T
2
2
T
2
T
cos 2f
0
t
cos 2f
0
t
sin 2f
0
t
sin 2f
0
t
2
T
T
0
(.)dt
2
T
3T/2
T/2
(.)dt
102
6.3 Phase Transitions for QPSK and
OQPSK
Example: Transmission of three symbols
For the transmission of three symbols, we consider the data bits
u = [u
1
, u
2
], the in-phase component (I) and the quadrature com-
ponent (Q) of the modulation symbol, and the phase change .
For convenience, let a :=
_
E
s
/2.
QPSK
PSfrag replacements
u 00 01
10
11
I
Q
0 T
T
2
= T
b
2T 3T
t
+a +a
+a a
a
a
+
2
OQPSK
PSfrag replacements
u 00 01
10
11
I
Q
0
0 T
T
2
= T
b
2T 3T
t
+a +a
+a
+a +a
+a
a
a a
a
a a
+
2
+
2
3
103
QPSK exhibits phase changes
2
, at times that are mul-
tiples of the signaling interval T. In contrast to this, OQPSK
exhibits phase changes
2
at times that are multiples of the bit
duration T
b
= T/2.
Remarks
(i) OQPSK has the same bit-error probability as QPSK.
(ii) Staggering the quadrature component of QPSK by T/2 does
not prevent phase changes of if the signaling constellation
of QPSK is not tilted. (Why?)
104
7 Frequency-Shift Keying (FSK)
In FSK, the waveforms are sines, and the information is conveyed
by the frequencies of the waveforms. FSK with two waveforms is
called binary FSK (BFSK), and FSK with M waveforms is called
M-ary FSK (MFSK).
7.1 BFSK with Coherent Demodulation
7.1.1 Signal Waveforms
Set S of two sinus waveforms with dierent frequencies f
1
, f
2
and
(possibly) dierent but known phases
1
,
2
[0, 2):
s
1
(t) =
2P cos
_
2f
1
t +
1
_
,
s
2
(t) =
2P cos
_
2f
2
t +
2
_
.
Both values are limited to the time interval t [0, T]. Thus we
have
S =
_
s
1
(t), s
2
(t)
_
.
Usually, f
1
, f
2
1/T. To simplify the theoretical analysis, we
assume that
f
1
=
k
1
T
, f
2
=
k
2
T
,
where k
1
, k
2
N and k
1
, k
2
1.
Both waveforms have the same energy
E
s
=
s
2
1
(t)dt =
s
2
2
(t)dt = PT.
105
We assume the following BFSK transmitter:
0, 1 S
u x(t)
with the mapping
0 s
1
(t)
1 s
2
(t).
The frequency separation
f = [f
1
f
2
[
is selected such that s
1
(t) and s
2
(t) are orthogonal, i.e.,
s
1
(t), s
2
(t)) =
T
_
0
s
1
(t)s
2
(t)dt = 0.
The necessary condition for orthogonality is
f =
k
T
with k N. Notice that the minimum frequency separation is
1/T.
The phases are assumed to be known to the receiver. Demodula-
tion with phase information (known phases or estimated phases)
is called coherent demodulation.
106
7.1.2 Signal Constellation
1
(t) =
1
E
s
s
1
(t) =
_
2
T
cos
_
2f
1
t +
1
_
2
(t) =
1
E
s
s
2
(t) =
_
2
T
cos
_
2f
2
t +
2
_
for t [0, T].
s
1
(t) =
_
E
s

1
(t) + 0
2
(t) s
1
= (
_
E
s
, 0)
T
s
2
(t) = 0
1
(t) +
_
E
s

2
(t) s
1
= (0,
_
E
s
)
T
.
PSfrag replacements
0
0
E
s
E
s
2E
s
2
s
1
s
2
T
1
T
2
BFSK with coherent demodulation has the same signal constella-
tion as BPPM.
Accordingly, these two modulation methods are equivalent.
107
7.1.3 ML Decoding
dec
ML
(y) = argmin
x{s
1
,s
2
}
|y x|
2
=
_
(
E
s
, 0)
T
for y
1
y
2
,
(0,
E
s
)
T
for y
1
< y
2
.
Decision regions
T
1
= y R
2
: y
1
y
2
, T
2
= y R
2
: y
1
< y
2
.
PSfrag replacements
0
0
E
s
E
s
2E
s
y
1
y
2
s
1
s
2
T
1
T
2
(u = 0)
(u = 1)
ML decoder for BFSK
u = enc
-1
(dec
ML
(y)) =
_
0 for y
1
y
2
,
1 for y
1
< y
2
.
Implementation of the ML decoder
PSfrag replacements
U
Y
1
Y
2
Z
z 0 0
z < 0 1
108
We assume BFSK over the AWGN channel and coherent demod-
ulation.
PSfrag replacements
Vector
Vector
Encoder
Decoder
AWGN Vector Channel
U X(t)
X
1
X
2
W(t)
Y (t)
Y
1
Y
2
U
T
T
2
T
2
T
cos(2f
1
t +
1
)
cos(2f
1
t +
1
)
cos(2f
2
t +
2
)
cos(2f
2
t +
2
)
2
T
T
0
(.)dt
2
T
T
0
(.)dt
The dashed box forms a 2-dimensional AWGN vector channel with
independent noise components
W
1
, W
2
^(0, N
0
/2).
Remark: Notice the similarities to BPSK and BPPM.
109
7.1.5 Error Probabilities
As the signal constellation of coherently demodulated BFSK is the
same as that of BPPM, the error probabilities are the same:
P
s
= Q(
s
),
P
b
= Q(
b
).
For the derivation, see analysis of BPPM.
110
7.2 MFSK with Coherent Demodulation
The signal waveforms are
s
m
(t) =
_
2P cos
_
2f
m
t +
m
_
for t [0, T],
0 elsewhere,
m = 1, 2, . . . , M.
The parameters of the waveforms are as follows:
(a) frequencies f
m
= k
m
/T for k
m
N and k
m
1;
(b) frequency separations f
m,n
= [f
m
f
n
[ = k
m,n
/T, k
m,n

N;
(c) phases
m
[0, 2);
m, n = 1, 2, . . . , M.
The phases are assumed to be known to the receiver. Thus, we
consider coherent demodulation.
As in the binary case, all waveforms have the same energy
E
s
=
s
2
m
(t)dt = PT.
111
m
(t) =
1
E
s
s
m
(t) =
_
2
T
cos
_
2f
m
t +
m
_
for t [0, T], m = 1, 2, . . . , M.
s
1
(t) =
_
E
s

1
(t) s
1
= (
M elements
..
_
E
s
, 0, . . . , 0)
T
.
.
.
s
M
(t) =
_
E
s

M
(t) s
1
= (0, . . . , 0,
_
E
s
)
T
.
MFSK with coherent demodulation has the same signal constella-
tion as MPPM.
Accordingly, these two modulation methods are equivalent and
have the same performance.
Vector Encoder
enc : u = [u
1
, . . . , u
K
] x = enc(u)
U = 0, 1
K
X =
_
s
1
, s
2
, . . . , s
M
_
with K = log
2
M.
112
We assume MFSK over the AWGN channel and coherent demod-
ulation.
PSfrag replacements
Vector
Vector
Encoder
Decoder
AWGN Vector Channel
U X(t)
X
1
X
M
W(t)
Y (t)
Y
1
Y
M
U
T
T
2
T
2
T
cos(2f
1
t +
1
)
cos(2f
1
t +
1
)
cos(2f
M
t +
M
)
cos(2f
M
t +
M
)
2
T
T
0
(.)dt
2
T
T
0
(.)dt
The dashed box forms an M-dimensional AWGN vector channel
with mutually independent components
W
m
^(0, N
0
/2),
m = 1, 2, . . . , M.
113
Remarks
The M waveforms s
m
(t) S are orthogonal. The receiver simply
correlates the received waveform y(t) to each of the possibly trans-
mitted waveforms s
m
(t) S and selects the one with the highest
correlation as the estimate.
7.2.4 Error Probabilities
The signal constellation of coherently demodulated MFSK is the
same as that of MPPM. Therefor, the symbol-error probability
and the bit-error probability are the same. For the analysis, see
MPPM.
114
7.3 BFSK with Noncoherent Demodulation
s
1
(t) =
2P cos
_
2f
1
t +
1
_
,
s
2
(t) =
2P cos
_
2f
2
t +
2
_
.
for t [0, T) with
(i) frequencies f
m
= k
m
/T for k
m
N and k
m
1, m = 1, 2,
and
(ii) frequency separation f = [f
1
f
2
[ = k/T for some k N.
As opposed to the case of coherent demodulation, the phases
1
and
2
are not known to the receiver. Instead, they are assumed
as random variables that are uniformly distributed, i.e.,
p
1
(
1
) =
1
2
, p
2
(
2
) =
1
2
for
1
,
2
[0, 2).
Demodulation without use of phase information is called nonco-
herent demodulation. As the phases need not be estimated,
the noncoherent demodulator has a lower complexity than the co-
herent demodulator.
115
Using the same method as for the canonical decomposition of PSK,
the orthonormal functions are obtained as
1
(t) =
_
2
T
cos
_
2f
1
t
_
2
(t) =
_
2
T
sin
_
2f
1
t
_
3
(t) =
_
2
T
cos
_
2f
2
t
_
4
(t) =
_
2
T
sin
_
2f
2
t
_
for t [0, T].
(Remember: cos(x + y) = cos x cos y sin x sin y.)
Geometrical representations of the waveforms:
s
1
(t) =
_
E
s
cos
1

1
(t) +
_
E
s
sin
1

2
(t)
s
1
=
_
_
E
s
cos
1
,
_
E
s
sin
1
, 0, 0
_
T
,
s
2
(t) =
_
E
s
cos
2

3
(t) +
_
E
s
sin
2

4
(t)
s
2
=
_
0, 0,
_
E
s
cos
2
,
_
E
s
sin
2
_
T
.
The signal constellation is thus
X =
_
s
1
, s
2
_
.
The transmitted (modulation) symbols are the four-
dimensional(!) vectors
X =
_
X
1
, X
2
. .
for s
1
, X
3
, X
4
. .
for s
2
_
T
X.
116
7.3.3 Noncoherent Demodulator
As there are four orthonormal basis functions, the demodulator
has four branches:
PSfrag replacements
Y (t)
Y
1
Y
2
Y
3
Y
4
2
T
T
0
(.)dt
2
T
T
0
(.)dt
2
T
T
0
(.)dt
2
T
T
0
(.)dt
cos 2f
1
t
sin 2f
1
t
cos 2f
2
t
sin 2f
2
t
Notice the similarity to PSK.
We introduce the vectors
Y
1
=
_
Y
1
Y
2
_
, Y
2
=
_
Y
3
Y
4
_
and
Y =
_
Y
T
1
, Y
T
2
_
T
=
_
Y
1
, Y
2
, Y
3
, Y
4
_
T
.
(Loosly speaking, Y
1
corresponds to s
1
and Y
2
corresponds to s
2
.)
117
7.3.4 Conditional Probability of Received Vector
Assume that the transmitted vector is
X = s
1
=
_
_
E
s
cos
1
,
_
E
s
sin
1
, 0, 0
_
T
.
Then, the received vector is
Y =
_
_
E
s
cos
1
+ W
1
,
_
E
s
sin
1
+ W
2
. .
Y
T
1
, W
3
, W
4
. .
Y
T
2
_
T
=
_
Y
T
1
, Y
T
2
_
T
.
The conditional probability density of Y given X = s
1
and
1
is
p
Y |X,
1
(y[s
1
,
1
) =
=
_
1
N
0
_
2
exp
_
1
N
0
|y s
1
|
2
_
=
_
1
N
0
_
2
exp
_
1
N
0
_
_
y
1
_
E
s
cos
1
_
2
+
+
_
y
2
_
E
s
sin
1
_
2
+ y
2
3
+ y
2
4
__
=
_
1
N
0
_
2
exp
_
E
s
N
0
_
exp
_
1
N
0
_
|y
1
|
2
+ |y
2
|
2
__
exp
_
2
E
s
N
0
_
y
1
cos
1
+ y
2
sin
1
_
_
With
1
= tan
1
(y
2
/y
1
), we may write
y
1
cos
1
+ y
2
sin
1
= |y
1
|
_
cos
1
cos
1
+ sin
1
sin
1
_
= |y
1
| cos(
1
1
).
118
The phase can be removed from the condition by averaging over
it. The phase is uniformly distributed in [0, 2), and thus
p
Y |X
(y[s
1
) =
=
2
_
0
p
Y |X,
1
(y[s
1
,
1
) p
1
(
1
) d
1
=
1
2
2
_
0
p
Y |X,
1
(y[s
1
,
1
) d
1
=
1
2
2
_
0
_
1
N
0
_
2
exp
_
E
s
N
0
_
exp
_
1
N
0
_
|y
1
|
2
+ |y
2
|
2
__
exp
_
2
E
s
N
0
|y
1
| cos(
1
1
)
_
d
1
=
_
1
N
0
_
2
exp
_
E
s
N
0
_
exp
_
1
N
0
_
|y
1
|
2
+ |y
2
|
2
__
1
2
2
_
0
exp
_
2
E
s
N
0
|y
1
| cos(
1
1
)
_
d
1
. .
I
0
_
2
E
s
N
0
|y
1
|
_
.
The above function is the modied Bessel function of order 0, and
it is dened as
I
0
(v) :=
1
2
2
_
0
exp
_
v cos( )
_
d
for [0, 2). It is monotonically increasing. A plot can be found
in [1, p. 403].
119
Combining all the above results yields
p
Y |X
(y[s
1
) =
_
1
N
0
_
2
exp
_
E
s
N
0
_
exp
_
1
N
0
_
|y
1
|
2
+ |y
2
|
2
__
I
0
_
2
E
s
N
0
|y
1
|
_
.
Using the same reasoning, we obtain
p
Y |X
(y[s
2
) =
_
1
N
0
_
2
exp
_
E
s
N
0
_
exp
_
1
N
0
_
|y
1
|
2
+ |y
2
|
2
__
I
0
_
2
E
s
N
0
|y
2
|
_
.
7.3.5 ML Decoding
Remember the denitions
Y =
_
Y
1
, Y
2
. .
Y
1
, Y
3
, Y
4
. .
Y
2
_
T
=
_
Y
T
1
, Y
T
2
_
T
.
x = dec
ML
(y) = argmax
xX
p
Y |X
(y[x)
Let x = s
m
. Then, the index of the ML vector is
m = argmax
m{1,2}
I
0
_
2
E
s
N
0
|y
m
|
_
= argmax
m{1,2}
|y
m
|
= argmax
m{1,2}
|y
m
|
2
.
Notice that selecting x and selecting m are equivalent.
120
ML decoder
u =
_
0 for x = s
1
|y
1
| |y
2
|,
1 for x = s
2
|y
1
| < |y
2
|.
Optimal noncoherent receiver for BFSK
PSfrag replacements
Y (t)
Y
1
Y
2
Y
3
Y
4
|Y
1
|
2
|Y
2
|
2
Z

U
2
T
T
0
(.)dt
2
T
T
0
(.)dt
2
T
T
0
(.)dt
2
T
T
0
(.)dt
cos 2f
1
t
sin 2f
1
t
cos 2f
2
t
sin 2f
2
t
(.)
2
(.)
2
(.)
2
(.)
2
z 0 0
z < 0 1
Notice:
|Y
1
|
2
=
_
Y
2
1
+ Y
2
2
, Z = |Y
1
|
2
|Y
2
|
2
,
|Y
2
|
2
=
_
Y
2
3
+ Y
2
4
.
121
7.4 MFSK with Noncoherent Demodula-
tion
s
m
(t) =
2P cos
_
2f
m
t +
m
_
for t [0, T), m = 1, 2, . . . , M, with
(i) f
m
= k
m
/T for k
m
N and k
m
1, m = 1, 2, . . . , M,
(ii) fm, n = [f
m
f
n
[ = k/T for some k
m,n
N, m, n =
1, 2, . . . , M.
The phases
1
,
2
, . . . ,
M
are not known to the receiver. They
are assumed to be uniformly distributed in [0, 2).
122
1
(t) =
_
2
T
cos
_
2f
1
t
_
2
(t) =
_
2
T
sin
_
2f
1
t
_
.
.
.
2m1
(t) =
_
2
T
cos
_
2f
m
t
_
2m
(t) =
_
2
T
sin
_
2f
m
t
_
.
.
.
2M1
(t) =
_
2
T
cos
_
2f
M
t
_
2M
(t) =
_
2
T
sin
_
2f
M
t
_
for t [0, T].
Signal vectors (of length 2M):
s
1
=
_
_
E
s
cos
1
,
_
E
s
sin
1
, 0, . . . , 0
_
T
s
2
=
_
0, 0,
_
E
s
cos
2
,
_
E
s
sin
2
, 0, . . . , 0
_
T
.
.
.
s
m
=
_
0, . . . , 0,
_
E
s
cos
m
,
_
E
s
sin
m
. .
positions 2m1 and 2m
, 0, . . . , 0
_
T
.
.
.
s
M
=
_
0, . . . , 0,
_
E
s
cos
M
,
_
E
s
sin
M
_
T
X =
_
s
1
, s
2
, . . . , s
M
_
.
123
7.4.3 Conditional Probability of Received Vector
Using the same reasoning as for BFSK, we can show that
p
Y |X
(y[s
m
) =
_
1
N
0
_
M
exp
_
E
s
N
0
_
exp
_
1
N
0
M
m=1
|y
m
|
2
_
I
0
_
2
E
s
N
0
|y
m
|
_
with the denition
Y =
_
Y
1
, Y
2
. .
Y
1
, Y
3
, Y
4
. .
Y
2
, . . . , Y
2M1
, Y
2M
. .
Y
M
,
_
T
=
_
Y
T
1
, Y
T
2
, . . . , Y
T
M
_
T
.
7.4.4 ML Decision Rule
Original ML decision rule:
x = dec
ML
(y) = argmax
xX
p
Y |X
(y[x)
Let x = s
m
. Then, the original decision rule may equivalently be
expressed as
m = argmax
m{1,...,M}
|y
m
|
2
.
124
7.4.5 Optimal Noncoherent Receiver for MFSK
Remember: noncoherent demodulation means that no phase in-
formation is used for demodulation.
The optimal noncoherent receiver for MFSK over the AWGN chan-
nel is as follows:
PSfrag replacements
Y (t)
Y
1
Y
2
Y
2M1
Y
2M
|Y
1
|
2
|Y
M
|
2
2
T
T
0
(.)dt
2
T
T
0
(.)dt
2
T
T
0
(.)dt
2
T
T
0
(.)dt
cos 2f
1
t
sin 2f
1
t
cos 2f
M
t
sin 2f
M
t
(.)
2
(.)
2
(.)
2
(.)
2
select
maximum
and
compute
encoder
inverse
Notice that

U = [
U
1
, . . . ,

U
K
] with K = log
2
M.
125
Assume that X = s
1
is transmitted. Then, the probability of a
correct decision is
Pr(

X = X[X = s
1
) =
= Pr
_
|Y
m
|
2
|Y
1
|
2
for all m ,= 1[X = s
1
_
.
Provided that X = s
1
is transmitted, we have
Y
1
=
_
_
E
s
cos
1
+ W
1
,
_
E
s
sin
1
+ W
2
_
T
Y
2
=
_
W
3
, W
4
_
T
= W
2
.
.
.
Y
m
=
_
W
2m1
, W
2m
_
T
= W
m
.
.
.
Y
M
=
_
W
2M1
, W
2M
_
T
= W
M
.
Consequently,
Pr(

X = X[X = s
1
) =
= Pr
_
|W
m
|
2
|Y
1
|
2
1
_
= Pr
_
1
N
0
/2
|W
m
|
2
. .
R
m
N
0
/2
|Y
1
|
2
. .
R
1
1
_
= Pr
_
R
m
R
1
1
_
=
_
0
Pr
_
R
m
R
1
1
, R
1
= r
1
_
p
R
1
(r
1
) dr
1
.
126
The probability density function of R
1
is denoted by p
R
1
(r
1
).
The probability of a correct decision can further be written as
Pr(

X = X[X = s
1
) =
=
_
0
Pr
_
R
m
r
1
1
, R
1
= r
1
_
p
R
1
(r
1
) dr
1
=
_
0
Pr
_
R
m
r
1
for all m ,= 1
_
p
R
1
(r
1
) dr
1
The noise vectors W
2
, . . . , W
M
are mutually statistically indepen-
dent, and thus also the values R
2
, . . . , R
M
are mutually statisti-
cally independent. Therefore,
Pr(

X = X[X = s
1
) =
_
0
M
m=2
Pr(R
m
r
1
) p
R
1
(r
1
) dr
1
=
_
0
_
Pr(R
2
r
1
)
_
M1
p
R
1
(r
1
) dr
1
.
Notice that
Pr(R
m
r
1
) = Pr(R
2
r
1
)
for m = 2, 3, . . . , M, because R
2
, . . . , R
M
are identically dis-
tributed.
127
The conditional probability of a symbol-error can thus be written
as
Pr(

X ,= X[X = s
1
) = 1 Pr(

X = X[X = s
1
)
= 1
_
0
_
Pr(R
2
r
1
)
_
M1
p
R
1
(r
1
) dr
1
.
By symmetry, all conditional symbol-error probabilities are the
same. Thus,
P
s
= 1
_
0
_
Pr(R
2
r
1
)
_
M1
p
R
1
(r
1
) dr
1
.
In the above expressions, we have the following distributions (see
literature):
(i) R
2
(and also R
3
, . . . , R
M
) is Rayleigh distributed:
p
R
2
(r
2
) = r
2
e
r
2
2
/2
.
Accordingly, its cumulative distribution function results as
Pr(R
2
r
1
) = 1 e
r
2
1
/2
(ii) R
1
is Rice distributed:
p
R
1
(r
1
) = r
1
exp
_
1
2
(r
1
+ 2
s
)
2
_
I
0
_
_
2
s
r
1
_
Remark
The Rayleigh distribution and the Ricd distribution are often used
to model mobile radio channels.
128
Using the rst relation, we can expand the rst term as
_
Pr(R
2
r
1
)
_
M1
=
_
1 e
r
2
1
/2
_
M1
=
M1
m=0
(1)
m
_
M 1
m
_
e
mr
2
1
/2
After integrating over r
1
, we obtain the following closed-form so-
lution for the symbol-error probability of MFSK with
noncoherent demodulation:
P
s
=
M1
m=1
(1)
m+1
_
M 1
m
_
1
m+ 1
exp
_
m
m+ 1

s
_
.
For M = 2, this sum simplies to
P
s
=
1
2
e
s
/2
.
129
BFSK
We have
b
=
s
and P
b
= P
s
.
Therefore, the bit-error probability is given by
P
b
=
1
2
e
b
/2
.
MFSK
For M > 2, we have
(i)
b
=
1
K
s
,
(ii) P
b
=
2
K1
2
K
1
P
s
with K = log
2
M (see MPPM).
Therefore the bit-error probability results as
P
b
=
2
K1
2
K
1
M1
m=1
(1)
m+1
_
M 1
m
_
1
m+ 1
exp
_
mK
m+ 1

s
_
.
Plots of the error probability can be found in [1, p. 433].
130
8 Minimum-Shift Keying (MSK)
MSK is a binary modulation scheme with coherent detection, and
it is similar to BFSK with coherent detection. The waveforms
are sines, the information is conveyed by the frequencies of the
waveforms, and the phases of the waveforms are chosen such that
there are no phase discontinuities. Thus, the power spectrum of
MSK is more compact than that of BFSK. The description of MSK
follows [2].
8.1 Motivation
Consider BFSK. The signal waveforms of BFSK are
U = 0 s
1
(t) =
2P cos
_
2f
1
t
_
,
U = 1 s
2
(t) =
2P cos
_
2f
2
t
_
,
for t [0, T]. The minimum frequency separation such that s
1
(t)
and s
2
(t) are orthogonal (assuming coherent detection) is
f = [f
1
f
2
[ =
1
2T
.
In the following, we assume minimum frequency separation.
Without loss of generality, we assume that
f
1
=
k
T
, f
2
=
k
T
+
1
2T
=
k + 1/2
T
,
for k N and k 1, i.e., k is a large integer. Thus, we have the
following properties:
waveform number of periods in t [0, T)
s
1
(t) k
s
2
(t) k + 1/2
131
Example: Output of BFSK Transmitter
Consider BFSK with f
1
= 1/T and f
2
= 3/2T such that the fre-
quency separation = 1/T is minimum. Assume the bit sequence
U
[0,4]
= [U
0
, U
1
, U
2
, U
3
, U
4
] = [0, 1, 1, 0, 0].
The resulting output signal x
(t) of the BFSK transmitter shows

discontinuities. 3
The discontinuities produce high side-lobes in the power spectrum
of X(t), which is either inecient or problematic. This can be
avoided by slightly modifying the modulation scheme.
We rewrite the signal waveforms as follows:
s
1
(t) =
2P cos
_
2f
1
t
_
,
s
2
(t) =
2P cos
_
2f
1
t + 2
t
2T
_
,
for t [0, T]. (Remember that f
2
= f
1
+ 1/2T.) Thus, we can
express the output signal as
x(t) =
2P cos
_
2f
1
t +
(t)
_
,
t [0, T), with the phase function
(t) =
_
0 for s
1
(t),
2
t
2T
for s
2
(t).
Notice that
(t) depends directly on the transmitted source bit U.

132
Example: Phase function of BFSK
The phase function
(t) of the previous example shows disconti-

nuities. 3
The discontinuities of the phase and thus of x(t) can be avoided by
making the phase continuous. This can be achieved by disabling
the reset to zero of
(t), yielding the new phase function (t).

Notice that this introduces memory in the transmitter.
Example: Modied BFSK Transmitter
(We continue the previous example.) The phase function (t) of
the modied transmitter is continuous. Accordingly, also the sig-
nal x(t) generated by the modied transmitter is continuous. 3
FSK with phase made continuous as described above is called
continuous-phase FSK (CPFSK). Binary CPFSK with minimum
frequency separation is termed minimum-shift keying (MSK).
8.2 Transmit Signal
The continuous-phase output signal of an MSK transmitter con-
sists of the following waveforms (the expressions on the right-hand
side are only used as abbreviations in the following):
s
1
(t) = +
2P cos
_
2f
1
t
_
f
1
s
2
(t) =
2P cos
_
2f
1
t
_
f
1
s
3
(t) = +
2P cos
_
2f
2
t
_
f
2
s
4
(t) =
2P cos
_
2f
2
t
_
f
2
t [0, T].
133
The input bit U
j
= 0 generates either s
1
(t) or s
2
(t); the input
bit U
j
= 1 generates either s
3
(t) or s
4
(t), which means a phase
dierence of between the beginning and the end of the waveform.
Consider a bit sequence of length L,
u
[0,L1]
= [u
0
, u
1
, . . . , u
L1
].
According to the above considerations (see also examples), the
output signal of the MSK transmitter can be written as
follows:
x(t) =
2P cos
_
2f
1
t +
L1
l=0
u
l
2 b(t lT)
. .
(t)
_
,
t [0, LT), with the phase pulse
b() =
_
_
0 for < 0,
/2T for 0 < T,
1/2 for T .
Remark
The BFSK signal x(t) can be generated with the same expression
as above, when the phase pulse
b
() =
_
_
0 for < 0,
/2T for 0 < T,
0 for T
is used.
134
Plotting all possible trajectories of (t), we obtain the (tilted)
phase tree of MSK.
PSfrag replacements
t
(t)
f
1
f
1
f
1
f
1
f
1
f
1
f
1
f
1
f
1
f
2
f
2
f
2
f
2
f
2
f
2
f
2
f
2
f
2
f
1
f
1
f
1
f
1
f
1
f
1
f
2
f
2
f
2
f
2
f
2
f
2
0
0 T 2T 3T 4T
5T
2
3
4
The physical phase is obtained by taking (t) modulo 2. Plotting
all possible trajectories, we obtain the (tilted) physical-phase
trellis of MSK.
PSfrag replacements
t
(t) mod 2
f
1
f
1
f
1
f
1
f
1
f
2
f
2
f
2
f
2
f
2
f
1
f
1
f
1
f
1
f
2
f
2
f
2
f
2
0
0 T 2T 3T 4T
5T
2
135
8.3 Signal Constellation
Signal waveforms:
s
1
(t) = +
2P cos
_
2f
1
t
_
,
s
2
(t) =
2P cos
_
2f
1
t
_
,
s
3
(t) = +
2P cos
_
2f
2
t
_
,
s
4
(t) =
2P cos
_
2f
2
t
_
,
t [0, T].
1
(t) =
_
2/T cos
_
2f
1
t
_
,
2
(t) =
_
2/T cos
_
2f
2
t
_
,
t [0, T].
Vector representation of the signal waveforms:
s
1
(t) = +
E
s

1
(t) s
1
= (+
E
s
, 0)
T
s
2
(t) =
E
s

1
(t) s
2
= (
E
s
, 0)
T
= s
1
s
3
(t) = +
E
s

2
(t) s
3
= (0, +
E
s
)
T
s
4
(t) =
E
s

2
(t) s
4
= (0,
E
s
)
T
= s
3
with E
s
= PT.
136
Remark:
The modulation symbol at discrete time j is written as
X
j
= (X
j,1
, X
j,2
)
T
.
For example, X
j
= s
1
means X
j,1
=

E
s
and X
j,2
= 0.
PSfrag replacements
2
s
1
s
2
s
3
s
4
Remark:
MSK and QPSK have the same signal constellation. However, the
two modulation schemes are not equivalent because their vector
encoders are dierent. This is discussed in the following.
137
Modulator:
The symbol X
j
= (X
j,1
, X
j,2
)
T
is transmitted in the time interval
t [jT, (j + 1)T).
This time interval can also be written as
t = jT + , [0, T).
The modulation for = t jT [0, T), j = 0, 1, 2, . . . , is then
given by the mapping
X
j
X
j
(),
and it can be implemented by
PSfrag replacements
X
j,1
X
j,2
X
j
()
T
T
2
T
2
T
cos(2f
1
)
cos(2f
2
)
Remark:
The index j is not necessary for describing the modulator. It is
only introduced to have the same notation as being used for the
vector encoder.
138
8.4 A Finite-State Machine
Consider the following device (which is a simple nite-state ma-
chine)
PSfrag replacements
U
j
V
j
V
j+1
T
It consists of the following components:
(i) A modulo-2 adder:
PSfrag replacements
U
j
V
j
V
j+1
It has the functionality
V
j+1
= U
j
V
j
= U
j
+ V
j
mod 2.
(ii) A shift register clocked at times T (every T seconds):
PSfrag replacements
U
j
V
j
V
j+1
T
(iii) The sequence U
j
is a binary sequence with elements
in 0, 1, and so is the sequence V
j
.
The output value V
j
describes the state of this device.
139
As V
j
0, 1, this device can be in two dierent states. All possi-
ble transitions can be depicted in a state transition diagram:
PSfrag replacements
0
1
0
0
0
1
1 1
The values in the circles denote the states v, and the labels of the
arrows denote the inputs u.
The state transitions and the associated input values can also be
depicted in a trellis diagram. The initial state is assumed to be
zero.
PSfrag replacements
0 0 0 0 0 0
1 1 1 1 1 1
0 0
0
0
0
0
0
0
0
1 1
1
1
1
1
1
1
1
0 1 2 3 4 5
The time indices are written below the states.
The emphasized path corresponds to the input sequence
[U
0
, U
1
, U
2
, U
3
, U
4
] = [0, 1, 1, 0, 0].
140
8.5 Vector Encoder
According to the trellis of the physical phase, the initial phase at
the beginning of each symbol interval can take the two values 0
and . These values determine the state of the encoder. We use
the following association:
phase 0 state V = 0,
phase state V = 1.
The functionality of the encoder can be described by the following
state transition diagram:
PSfrag replacements
0
0/s
1
0/s
2
1/s
3
1/s
4
The transitions are labeled by u/x. The value u denotes the binary
source symbol at the encoder input, and x denotes the modulation
symbol at the encoder output.
Remember: s
2
= s
1
and s
4
= s
3
.
Remark
The vector encoder is a nite-state machine as considered in the
previous section.
141
The temporal behavior of the encoder can be described by the
trellis diagram:
PSfrag replacements
0 0 0 0 0 0

0/s
1
0/s
1
0/s
1
0/s
1
0/s
1
0/s
2
0/s
2
0/s
2
0/s
2
1
/
s
3
1
/
s
3
1
/
s
3
1
/
s
3
1
/
s
3
1
/
s
4
1
/
s
4
1
/
s
4
1
/
s
4
0 1 2 3 4 5
The emphasized path corresponds to the input sequence
[U
0
, U
1
, U
2
, U
3
, U
4
] = [0, 1, 1, 0, 0].
General Conventions
(i) The input sequence of length L is written as
u = [u
0
, u
1
, . . . , u
L1
].
(ii) The initial state is V
0
= 0.
(iii) The nal state is V
L
= 0.
The last bit U
L1
is selected such that the encoder is driven back
to state 0:
V
L1
= 0 U
L1
= 0,
V
L1
= 1 U
L1
= 1.
Therefore, U
L1
does not carry information.
142
Generation of the Encoder Outputs
Consider one trellis section:
PSfrag replacements
0 0

0/s
1
0/s
2
1
/
s
3
1
/
s
4
j j + 1
Remember: s
2
= s
1
and s
4
= s
3
.
Observations:
(i) Input symbols output vectors
U
j
= 0 X
j
s
1
, s
2
= s
1
, s
1
,
U
j
= 1 X
j
s
3
, s
4
= s
3
, s
3
.
(ii) Encoder state output vectors
V
j
= 0 (0) X
j
s
1
, s
3
V
j
= 1 () X
j
s
2
, s
4
= s
1
, s
3
.
Using these observations, the above device can be supplemented
to obtain an implementation of the vector encoder for MSK.
143
Block Diagram of the Vector Encoder for MSK
PSfrag replacements
U
j
U
j
V
j
V
j+1
X
j,1
X
j,2
T
U
j
V
j
s
1
s
2
s
3
s
4
00
01
10
11
At time index j:
input symbol U
j
,
state V
j
,
output symbol X
j
= (X
j,1
, X
j,2
)
T
,
subsequent state V
j+1
.
Notice:
(i) U
j
controls the frequency of the transmitted waveform.
(ii) V
j
controls the polarity of the transmitted waveform.
144
8.6 Demodulator
The received symbol Y
j
= (Y
j,1
, Y
j,2
)
T
is determined in the time
interval
t [jT, (j + 1)T).
This time interval can also be written as
t = jT + , [0, T).
The demodulation for = t jT [0, T), j = 0, 1, 2, . . . , can
then be implemented as follows:
PSfrag replacements
Y (t)
Y
j,1
Y
j,2
T
2
T
cos(2f
1
)
cos(2f
2
)
2
T
T
0
(.)d
2
T
T
0
(.)d
145
8.7 Conditional Probability Density of Re-
ceived Sequence
Sequence of (two-dimensional) received symbols
[Y
0
, . . . , Y
L1
] = [X
0
+ W
0
, . . . , X
L1
+ W
L1
].
The vector [W
0
, . . . , W
L1
] is a white Gaussian noise sequence
with the properties:
(i) W
0
, . . . , W
L1
are independent ;
(ii) each (two-dimensional) noise vector is distributed as
W
j
=
_
W
j,1
W
j,2
_
^
_
_
0
0
_
,
_
N
0
/2 0
0 N
0
/2
_
_
Illustration of the AWGN vector channel:
PSfrag replacements
X
j
X
j,1
X
j,2
W
j
W
j,1
W
j,2
Y
j
Y
j,1
Y
j,2
146
Conditional probability of y
[0,L1]
= [y
0
, . . . , y
L1
] given that
x
[0,L1]
= [x
0
, . . . , x
L1
] was transmitted:
p
Y
[0,L1]
|X
[0,L1]
_
y
0
, . . . , y
L1
[x
0
, . . . , x
L1
_
=
=
L1
j=0
p
Y
j
|X
j
_
y
j
[x
j
_
=
L1
j=0
_
1
N
0
exp
_
1
N
0
|y
j
x
j
|
2
_
_
=
_
1
N
0
_
L
exp
_
1
N
0
L1
j=0
|y
j
x
j
|
2
_
.
Notice that this is the likelihood function of x
[0,L1]
for a given
received sequence y
[0,L1]
.
8.8 ML Sequence Estimation
The trellis of the vector encoder has the following two properties:
(i) Each input sequence u
[0,L1]
= [u
0
, . . . , u
L1
] uniquely de-
termines a path of length L in the trellis, and vice versa.
(ii) Each path in the trellis uniquely determines an output se-
quence x
[0,L1]
= [x
0
, . . . , x
L1
]
A sequence x
[0,L1]
of length L generated by the vector encoder is
called an admissible sequence if the initial and the nal state of
the encoder are zero, i.e., if V
0
= V
L
= 0. The set of admissible
sequences is denoted by A.
147
Using these denitions, the ML decoding rule is the following:
Search the trellis for an admissible sequence that max-
imizes the likelihood function.
Or in mathematical terms:
x
[0,L1]
= argmax
x
[0,L1]
A
p(y
[0,L1]
[x
[0,L1]
).
Using the relation
|y
j
x
j
|
2
= |y
j
|
2
. .
indep. of x
j
2y
j
, x
j
) + |x
j
|
2
. .
= E
s
and the expression for the likelihood function, the rule for ML
sequence estimation (MLSE) can equivalently be formulated
as follows:
x
[0,L1]
= argmax
x
[0,L1]
A
exp
_
1
N
0
L1
j=0
|y
j
x
j
|
2
_
= argmin
x
[0,L1]
A
L1
j=0
|y
j
x
j
|
2
= argmax
x
[0,L1]
A
L1
j=0
y
j
, x
j
).
Hence, the ML estimate of the transmitted sequence is a sequence
in A that maximizes the sum
L1
j=0
y
j
, x
j
).
148
149
8.9 Canonical Decomposition of MSK
AWGN Vector Channel
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
U
j
U
j
V
j
V
j
+
1
X
j
,
1
X
j
,
2
T
U
j
V
j
s
1
s
2
s
3
s
4
0
0
0
1
1
0
1
1
X
(
t
)
W
(
t
)
A
W
G
N
V
e
c
t
o
r
C
h
a
n
n
e
l
Y
(
t
)
Y
j
,
1
Y
j
,
2
T
T T
2 T
2 T
c
o
s
(
2
f
1
)
c
o
s
(
2
f
1
)
c
o
s
(
2
f
2
)
c
o
s
(
2
f
2
2 T
T
0
(
.
)
d
2 T
T
0
(
.
)
d
M
a
x
i
m
u
m
-
L
i
k
e
l
i
h
o
o
d
S
e
q
u
e
n
c
e
D
e
c
o
d
i
n
g
U
j
150
8.10 The Viterbi Algorithm
Given Problem:
(i) The admissible sequences
x
[0,L1]
= [x
0
, . . . , x
L1
] A
can be represented in a trellis.
(ii) The value to be maximized (optimization criterion) can be
written as the sum
L1
j=0
y
j
, x
j
).
(Remember: y
j
, x
j
) = y
j,1
x
j,1
+ y
j,2
x
j,2
.)
The Viterbi Algorithm solves such an optimization problem in
a very eective way.
Branch Metrics and Path Metrics
The branches of the trellis are re-labeled as follows:
PSfrag replacements
v
j
v
j
v
j+1
v
j+1
u
j
/
x
j
y
j
,
x
j
)
j j j + 1 j + 1
The new labels are called branch metrics. They may be inter-
preted as a gain or a prot made when going along this branch.
151
When labeling all branches with their branch metrics, we obtain
the following trellis diagram:
PSfrag replacements
0 0 0 0 0 0

y
0
, s
1
y
0
, s
2
y
0
,
s
3
y
0
, s
4
y
1
, s
1
y
1
, s
2
y
1
,
s
3
y
1 ,
s
4
y
2
, s
1
y
2
, s
2
y
2
,
s
3
y
2 ,
s
4
y
3
, s
1
y
3
, s
2
y
3
,
s
3
y
3 ,
s
4
y
4
, s
1
y
4
, s
2
y
4
,
s
3
y
4 ,
s
4
0 1 2 3 4 5
The total gain or prot achieved by going along a path in the
trellis is called the path metric. The path corresponding to the
sequence
x
[0,L1]
= [x
0
, . . . , x
L1
]
has the path metric
L1
j=0
y
j
, x
j
).
Notice: The branch metric and the path metric are not necessarily
a mathematical metric.
Hence, the ML estimate x
[0,L1]
of X
[0,L1]
is an admissible se-
quence ( A) that maximizes the path metric. The corresponding
path is called an optimum path.
There are 2 4
L1
paths of length L. The Viterbi algorithm is
a sequential method that determines an optimum path without
computing the path metrics of all 2 4
L1
paths.
152
Survivors
The Viterbi algorithm relies on the following property of optimum
paths:
A path of length L is optimum if and only if its sub-
paths at any depth n = 1, 2, . . . , L 1 are also opti-
mum.
Example: Optimality of Subpaths
Consider the following trellis. Assume that the thick path (solid
line) is optimum.
PSfrag replacements
0 0 0 0 0 0

0 1 2 3 4 5
The dashed path ending at depth 3 cannot have a metric that is
larger than that of the subpath obtained by pruning the optimum
path at depth 3. 3
For each state in each depth n, there are several subpaths ending
in this state. The one with the largest metric (the best path) is
called the survivor at this state in this depth.
153
From the above considerations, we have the following important
results:
All subpaths of an optimum path are survivors.
Therefore, only survivors need to be considered for determin-
ing an optimum path.
Accordingly, non-survivor paths can be discarded without loss of
optimality. (Therefore the name survivor.)
Viterbi Algorithm
1. Start at the beginning of the trellis (depth 0 and state 0 by
convention).
2. Increase the depth by 1. Extend all previous survivors to the
new depth and determine the new survivor for each state.
3. If the end of the trellis is reached (depth L and state 0 by
convention), the survivor is an optimum path. The associated
sequence is the ML sequence.
Remarks
(i) The complexity of computing the path metrics of all paths is
exponential in the sequence length. (Exhaustive search.)
(ii) The complexity of the Viterbi algorithm is linear in the se-
quence length.
(iii) The complexity of the Viterbi algorithm is exponential in the
number of shift registers in the vector encoder.
154
Example: Viterbi Algorithm
Consider the case L = 5. This corresponds to the following trellis.
PSfrag replacements
0 0 0 0 0 0

0 1 2 3 4 5
The Viterbi algorithm may be illustrated for a randomly drawn re-
ceived sequence y
[0,L1]
. (The receiver must be capable of handling
every received sequence.) 3
155
8.11 Simplied ML Decoder
The MSK trellis has a special structure, and this leads to a special
behavior of the Viterbi algorithm. This can be exploited to derive
a simpler optimal decoder.
Observation
At depth n = 2, the two survivors (solid line and dashed line) will
look as depicted in (a) or (b), but never as in (c) or (d).
PSfrag replacements
0 0
0 0
1 1
1 1
2 2
2 2
(a)
(b)
(c)
(d)
Therefore the nal estimate of the state

V
1
is already decided at
time j = 1. This is proved shortly afterwards.
The general result for the given MSK transmitter follows by induc-
tion:
The two survivors at any depth n are identical for j < n
and dier only in the current state.
Therefore, nal decisions on any state

V
n
(value of the vector-
encoder state at time n) can already be made at time n.
156
We proof now that the Viterbi algorithm makes the nal decision
about state

V
1
(state at time j = 1) already at time 1.
Proof
The inner product is dened as
y
j
, s
m
) = y
j,1
s
m,1
+ y
j,2
s
m,2
.
Due to the signal constellation,
s
2
= s
1
, s
4
= s
3
.
Therefore,
y
j
, s
2
) = y
j
, s
1
) = y
j
, s
1
),
y
j
, s
4
) = y
j
, s
3
) = y
j
, s
3
).
Using these relations, the branches can equivalently be labeled as
PSfrag replacements
0 0 0 0 0 0

y
0
, s
1
y
0
, s
1
y
0
,
s
3
y
0
, s
3
y
1
, s
1
y
1
, s
1
y
1
,
s
3
y
1 ,
s
3
y
2
, s
1
y
2
, s
1
y
2
,
s
3
y
2 ,
s
3
y
3
, s
1
y
3
, s
1
y
3
,
s
3
y
3 ,
s
3
y
4
, s
1
y
4
, s
1
y
4
,
s
3
y
4 ,
s
3
0 1 2 3 4 5
157
Consider the conditions for selecting the survivors, while making
use of these relations:
Survivor at state v
2
= 0 (time j = 1):
PSfrag replacements
0 0 0

y
0
, s
1
)
y
0
, s
1
)
y
0
,
s
3
)
y
0
, s
3
)
y
1
, s
1
)
y
1
, s
1
)
y
1
, s
3
)
y
1
,
s
3
)
0 1 2
y
0
, s
3
) y
1
, s
3
) > y
0
, s
1
) + y
1
, s
1
) upper path
y
0
, s
3
) y
1
, s
3
) < y
0
, s
1
) + y
1
, s
1
) lower path
Survivor at state v
2
= 1 () (time j = 1):
PSfrag replacements
0 0 0

y
0
, s
1
)
y
0
, s
1
)
y
0
,
s
3
)
y
0
, s
3
)
y
1
, s
1
)
y
1
, s
1
)
y
1
,
s
3
)
y
1
, s
3
)
0 1 2
y
0
, s
3
) y
1
, s
1
) > y
0
, s
1
) + y
1
, s
3
) upper path
y
0
, s
3
) y
1
, s
1
) < y
0
, s
1
) + y
1
, s
3
) lower path
Obviously, either the two upper paths or the two lower paths are
the survivors. Therefore, the two survivors go through the same
state

V
1
.
158
ML State Sequence
By induction, we obtain the following optimal decision rule for
state V
j
based on y
j1
and y
j
:
V
j
=
_
0 if y
j1
, s
3
) y
j
, s
3
) < y
j1
, s
1
) + y
j
, s
1
),
1 if y
j1
, s
3
) y
j
, s
3
) > y
j1
, s
1
) + y
j
, s
1
),
j = 1, 2, . . . , L 1. (Remember: V
L
= 0 by convention.)
This decision rule can be simplied. The modulation symbols are
s
1
= (+
E
s
, 0)
T
, s
3
= (0, +
E
s
)
T
.
Thus, the inner products can be written as
y
j
, s
1
) = y
j,1
E
s
, y
j
, s
3
) = y
j,2
E
s
.
Substituting this in the decision rule and dividing by

E
s
, we ob-
tain the following simplied (and still optimal) decision rule:
V
j
=
_
0 if y
j1,2
y
j,2
< y
j1,1
+ y
j,1
,
1 if y
j1,2
y
j,2
> y
j1,1
+ y
j,1
,
j = 1, 2, . . . , L 1.
159
The following device implements this state decision rule:
PSfrag replacements
Y
j1,1
Y
j,1
Y
j1,2
Y
j,2
Z
j,1
Z
j,2
Z
j

V
j
T
T
z
j
> 0 0
z
j
< 0 1
Vector-Encoder Inverse
Given the sequence of state estimates
V
j
, the sequence of bit
estimates
U
j
can computed according to
U
j
=

V
j

V
j+1
,
j = 0, 1, . . . , L 1. (Compare vector encoder.)
The following device implements this operation: PSfrag replacements
V
j

V
j1

U
j1
T
The time indices are j = 0, 1, . . . , L1, and

V
L
= 0 as V
L
= 0 by
convention. Remember that U
L1
does not convey information.
160
The BER performance of MSK is about 2 dB worse than that
of BPSK and QPSK. This means, when comparing the SNR in
E
b
/N
0
required to achieve a certain bit-error probability (BEP),
MSK needs about 2 dB more than BPSK and QPSK. This can be
improved by slightly modifying the vector encoder.
The trellis of the vector encoder has the following properties:
(i) Two dierent paths dier in at least one state.
(ii) Two dierent paths dier in at least two bits.
(See trellis of vector encoder in Section 8.5.)
The most likely error-event for high SNR is that the transmitted
path and the estimated path dier in one state. This error-event
leads to two bit errors.
Idea for improvement:
The state sequence should be equal to the bit sequence, such
that the most likely error-event leads only to one bit error.
This can be achieved by using a pre-coder before the vector
encoder.
The resulting modulation scheme is called precoded MSK
161
8.13 Dierential Encoding and Dierential
Decoding
The two terms dierential encoder and dierential decoder
are mainly used due to historical reasons. Both may be used as
encoders.
In the following, b
j
0, 1 denote the bits before encoding, and
c
j
0, 1 denote the bits after encoding.
Dierential Encoder (DE)
PSfrag replacements
b
j
c
j+1
c
j
T
Operation:
c
j+1
= b
j
c
j
,
j = 0, 1, 2, . . . . Initialization: c
0
= 0.
Dierential Decoder (DD)
PSfrag replacements
b
j
b
j1
c
j
T
Operation:
c
j
= b
j
b
j1
,
j = 0, 1, 2, . . . . Initialization: b
1
= 0.
162
8.14 Precoded MSK
The DD(!) is used as a pre-coder and is placed before the vector
encoder. The DE(!) is placed behind the vector-encoder inverse.
PSfrag replacements
U
j
X(t)
W(t)
Y (t)

U
j1
DD MSK Tx
MSK Rx / DE
Vector Representation of Precoded MSK
PSfrag replacements
U
j
U
j
V
j
= U
j1
X
j,1
X
j,2
T
s
1
s
2
s
3
s
4
00
01
10
11
U
j
V
j
Initialization: V
0
= 0. Notice: U
j
= U
j
V
j
= U
j
U
j1
.
PSfrag replacements
Y
j1,1
Y
j,1
Y
j1,2
Y
j,2
Z
j,1
Z
j,2
Z
j

V
j
=

U
j1
T
T
z
j
> 0 0
z
j
< 0 1
Notice: The vector-encoder inverse and the DE compensate each
other and can be removed.
163
Trellis of Precoded MSK
PSfrag replacements
0 0 0 0 0 0
1 1 1 1 1 1
0/s
1
0/s
1
0/s
1
0/s
1
0/s
1
1/s
1
1/s
1
1/s
1
1/s
1
1
/
s
3
1
/
s
3
1
/
s
3
1
/
s
3
1
/
s
3
0
/
s
3
0
/
s
3
0
/
s
3
0
/
s
3
0 1 2 3 4 5
The thick path corresponds to the bit sequence
U
[0,4]
= [0, 1, 1, 0, 0].
Bit-Error Probability of Precoded MSK
The bit-error probability of precoded MSK is
P
b
= Q(
_
2
b
),
b
= E
b
/N
0
.
Sketch of the Proof:
The bit-error probability is given as
P
b
= lim
L
1
L
L1
j=0
Pr(
U
j
,= U
j
).
All terms in this sum can shown to be identical, and thus
P
b
= Pr(
U
0
,= U
0
).
Due to the precoding, we have U
0
= V
1
, and therefore
P
b
= Pr(
V
1
,= V
1
).
164
This expression can be written using conditional probabilities:
P
b
=
1
v=0
1
u=0
Pr(
V
1
,= V
1
[V
1
= v, U
1
= u) Pr(V
1
= v, U
1
= u).
(Usual trick.) These conditional probabilities can shown to be all
identical, and thus
P
b
= Pr(
V
1
,= V
1
[V
1
= 0, U
1
= 0).
As V
1
= U
0
(property of precoded MSK), we obtain
P
b
= Pr(
V
1
= 1[U
0
= 0, U
1
= 0).
The ML decision rule is
V
1
=
_
0 if y
0,2
y
1,2
< y
0,1
+ y
1,1
,
1 if y
0,2
y
1,2
> y
0,1
+ y
1,1
.
Given the hypothesis U
0
= U
1
= 0, we have
X
0
= X
1
= s
1
= (
E
s
, 0)
T
,
and so
Y
0,1
=
E
s
+ W
0,1
, Y
0,2
= W
0,2
,
Y
1,1
=
E
s
+ W
1,1
, Y
1,2
= W
1,2
.
Remember:
Y
0
= (Y
0,1
, Y
0,2
)
T
, Y
1
= (Y
1,1
, Y
1,2
)
T
.
165
Therefore, the bit-error probability can be written as
P
b
= Pr(Y
0,2
Y
1,2
> Y
0,1
+ Y
1,1
[U
0
= 0, U
1
= 0)
= Pr(W
0,2
W
1,2
> 2
E
s
+ W
0,1
+ W
1,1
)
= Pr(W
0,2
W
1,2
W
0,1
W
1,1
. .
W
^(0, 2N
0
)
> 2
E
s
)
= Pr
_
W
/
_
2N
0
>
_
2E
s
/N
0
_
= Q
_
_
2E
s
/N
0
_
= Q
_
_
2
b
_
.
As (precoded) MSK is a binary modulation scheme, we have E
s
=
E
b
and
s
=
b
.
Thus, the bit-error probability of precoded MSK is
P
b
= Q(
_
2
b
),
which is the same as that of BPSK and QPSK.
166
8.15 Gaussian MSK
Gaussian MSK is a generalization of MSK, and it is used in the
GSM mobile phone system.
Motivation
The phase function of MSK is continuous, but not its dervative.
These edges in the transmit signal lead to some high-frequency
components. By avoiding these edges, i.e., by smoothing the
phase function, the spectrum of the transmit signal can be made
more compact.
Concept
The output signal of the transmitter is generated similar to that
in MSK:
x(t) =
2P cos
_
2f
1
t +
L1
l=0
u
l
2 b(t lT)
. .
(t)
_
,
t [0, LT), with the phase pulse
b() =
_
_
0 for < 0,
monotonically increasing for 0 < JT,
1/2 for JT ,
for some integer J 1.
167
Notice that there are two new degrees of freedom:
(i) The phase pulse may stretch over more than one symbol
duration T.
(ii) The phase pulse may be be selected such that its derivative
is zero for t = and t = JT.
Phase Pulse and Frequency Pulse
In GMSK, the phase pulse b() is constructed via its deriva-
tive c(), called the frequency pulse:
(i) The frequency pulse c() is the convolution of a rectan-
gular pulse (as in MSK!) with a Gaussian pulse:
c() = rectangular pulse Gaussian pulse,
where c() = 0 for / [0, JT), and
_
c()d = 1/2. The
resulting pulse is
c() = Q
_
2B
ln 2
_
T/2
_
Q
_
2B
ln 2
_
+ T/2
_
_
.
The value B corresponds to a bandwidth and is a design
parameter.
(ii) The phase pulse is the integral over the frequency pulse:
b() =
_
0
c(
)d
.
Due to the denition of the frequency pulse, we have b() =
1/2 for JT.
The resulting phase pulse fullls the conditions given above.
168
In GSM, the parameter B is chosen such that BT 0.3. The
length of the phase pulse is then L 4.
169
9 Quadrature Amplitude Modula-
tion
170
10 BPAM with Bandlimited Wave-
forms
171
Part III
Appendix
172
A Geometrical Representation of
Signals
A.1 Sets of Orthonormal Functions
The D real-valued functions
1
(t),
2
(t), . . . ,
D
(t)
form an orthonormal set if
+
_
k
(t)
l
(t) dt =
_
0 for k ,= l (orthogonality),
1 for k = l (normalization).
Example: D = 3
PSfrag replacements
t
t
t
1
(t)
2
(t)
3
(t)
T
T
T
1/
T
1/
T
1/
T
1/
T
1/
T
1/
T
3
173
A.2 Signal Space Spanned by an Orthonor-
mal Set of Functions
Let
_
1
(t),
2
(t), . . . ,
D
(t)
_
be a set of orthonormal functions.
We denote by o
the set containing all linear combinations of these

functions:
o
:=
_
s(t) : s(t) =
D
i=1
s
i
i
(t); s
1
, s
2
, . . . , s
D
R
_
.
Example: (continued)
The following functions are elements of o
:
s
1
(t) =
1
(t) +
2
(t), s
2
(t) =
1
2
1
(t)
2
(t) +
1
2
3
(t)
PSfrag replacements
t t
s
1
(t) s
2
(t)
T T
2/
T 2/
T
2/
T 2/
T
3
The set o
is a vector space with respect to the following opera-

tions:
Addition
+ :
_
s
1
(t), s
2
(t)
_
s
1
(t) + s
2
(t)
Multiplication with a scalar
:
_
a, s(t)
_
a s(t)
174
A.3 Vector Representation of Elements in
the Vector Space
With each element s(t) o
,
s(t) =
D
i=1
s
i
i
(t),
we associate the D-dimensional real vector
s = [s
1
, s
2
, . . . , s
D
]
T
R
D
.
s
1
(t) =
1
(t) +
2
(t) s
1
= [1, 1, 0]
T
,
s
2
(t) =
1
2
1
(t)
2
(t) +
1
2
3
(t) s
2
= [
1
2
, 1,
1
2
]
T
.
3
Vectors in R, R
2
, R
3
, can be represented geometrically by points
on the line, in the plane, an in the 3-dimensional Euclidean space,
respectively.
PSfrag replacements
3
1
1
1
0.5
0.5
s
1
s
2
3
175
A.4 Vector Isomorphism
The mapping o
R
D
,
s(t) =
D
i=1
s
i
i
(t) s = [s
1
, s
2
, . . . , s
D
]
T
,
is a (vector) isomorphism, i.e.:
(i) The mapping is bijective.
(ii) The mapping preserves the addition and the multiplication
operators:
s
1
(t) + s
2
(t) s
1
+ s
2
,
as(t) as.
Since the mapping o
R
D
is an isomorphism, o
and R
D
are
the same or equivalent from an algebraic point of view.
176
A.5 Scalar Product and Isometry
Scalar product in o
:
s
1
(t), s
2
(t)) :=
+
_
s
1
(t)s
2
(t) dt
for s
1
(t), s
2
(t) o
.
Scalar product in R
D
:
s
1
, s
2
) :=
D
i=1
s
1,i
s
2,i
for s
1
, s
2
R
D
.
Relation of scalar products:
Each element in o
is a linear combination of
1
(t),
2
(t), . . . ,
D
(t):
s
1
(t) =
D
k=1
s
1,k
k
(t), s
2
(t) =
D
l=1
s
2,l
l
(t).
Inserting the two sums into the denition of the scalar product, we
obtain
s
1
(t), s
2
(t)) =
_
_
D
k=1
s
1,k
k
(t)
__
D
l=1
s
2,l
l
(t)
_
dt =
=
D
k=1
D
l=1
s
1,k
s
2,l
_

k
(t)
l
(t)dt
. .
orthonormality
=
D
k=1
s
1,k
s
2,k
= s
1
, s
2
)
Hence the mapping o
R
D
preserves the scalar product, i.e., it
is an isometry.
177
Norms in o
|s(t)|
2
:= s(t), s(t)) =
_
s
2
(t) dt = E
s
The value E
s
is the energy of s(t).
Norms in R
D
:
|s|
2
:= s, s) =
D
i=1
s
2
i
From the isometry and the denitions of the norms, we obtain the
Parseval relation
|s(t)|
2
= |s|
2
.
178
A.6 Waveform Synthesis
Let
_
1
(t),
2
(t), . . . ,
D
(t)
_
be a set of orthonormal functions
spanning the vector space o
. Every element s(t) o
can be
synthesized from its vector representation s = [s
1
, s
2
, . . . , s
D
]
T
by
one of the following devices.
Multiplicative-based Synthesizer
PSfrag replacements
s
s
1
s
2
s
D
1
(t)
2
(t)
D
(t)
(t)
s(t) =
D
k=1
s
k
k
(t)
Filter-bank based Synthesizer
Filter
Filter
Filter
PSfrag replacements
s
s
1
s
2
s
D
1
(t)
2
(t)
D
(t)
(t)
s(t)
s(t) =
D
k=1
s
k
(t)
k
(t) with Dirac function (t)
179
A.7 Implementation of the Scalar Product
Let s
1
(t) and s
2
(t) be time-limited to the interval [0, T]. Then
s
1
(t), s
2
(t)) :=
T
_
0
s
1
(t)s
2
(t) dt.
The scalar product can be computed by the following two devices.
Correlator-based Implementation
PSfrag replacements
s
1
(t)
s
2
(t)
_
T
0
(.)dt s
1
(t), s
2
(t))
Matched-lter based Implementation
We seek a linear lter that outputs the value s
1
(t), s
2
(t)) at a
certain time, say t
0
.
PSfrag replacements
s
1
(t)
g(t)
linear lter
sample at
time t
0
s
1
(t), s
2
(t))
We want
s
1
(t), s
2
(t)) =
T
_
0
s
1
()s
2
() d
!
=
+
_
s
1
()g(t
) d
=t
0
The second equality holds for
s
2
() = g(t
0
) g(t) = s
2
(t
0
t)
180
Example:
PSfrag replacements
t t
t
0
s
2
(t) s
2
(t
0
t)
T T
3
The lter is causal and thus realizable for a function time-limited
to the interval [0, T] if
t
0
T.
The value t
0
is the time at which the value s
1
(t), s
2
(t)) appears
at the lter output. Choosing the minimum value, i.e., t
0
= T, we
obtain the following (realizable) implementation.
PSfrag replacements
s
1
(t)
s
2
(T t)
matched lter
sample at
time T
s
1
(t), s
2
(t))
The lter is matched to s
2
(t).
181
A.8 Computation of the Vector Represen-
tation
Let
_
1
(t),
2
(t), . . . ,
D
(t)
_
be a set of orthonormal functions
time-limited to [0, T] and spanning the vector space o
.
We seek a device that computes the vector representation s R
D
for any function s(t) o
, i.e.,
s = [s
1
, s
2
, . . . , s
D
]
T
such that s(t) =
D
k=1
s
k
k
(t).
The components of s can be computed as
s
k
= s(t),
k
(t)), k = 1, 2, . . . , D.
Correlator-based Implementation
PSfrag replacements
s(t)

1
(t)
D
(t)
_
T
0
(.)dt
_
T
0
(.)dt
s
1
s
D
Matched-lter based Implementation
PSfrag replacements
s(t)
1
(T t)
D
(T t)
MF
MF
s
1
s
D
sample at
time T
182
A.9 Gram-Schmidt Orthonormalization
Assume a set of M signal waveforms
_
s
1
(t), s
2
(t), . . . , s
M
(t)
_
spanning a D-dimensional vector space. From these waveforms,
we can construct a set of D M orthonormal waveforms
_
1
(t),
2
(t), . . . ,
D
(t)
_
by applying the following procedure.
Remember:
s
1
(t), s
2
(t)) =
_
s
1
(t)s
2
(t) dt
|s(t)|
2
:= s(t), s(t)) = E
s
Gram-Schmidt orthonormalization procedure
1. Take the rst waveform s
1
(t) and normalize it to unit energy:
1
(t) :=
s
1
(t)
_
|s
1
(t)|
2
.
2. Take the second waveform s
2
(t). Determine the projection
of s
2
(t) onto
1
(t), and subtract this projection from s
2
(t):
c
2,1
= s
2
(t),
1
(t)),
s
2
(t) = s
2
(t) c
2,1
1
(t).
Normalize s
2
(t) to unit energy:
2
(t) :=
s
2
(t)
_
|s
2
(t)|
2
.
183
3. Take the third waveform s
3
(t). Determine the projections
of s
3
(t) onto
1
(t) and onto
2
(t), and subtract these pro-
jections from s
3
(t):
c
3,1
= s
3
(t),
1
(t)),
c
3,2
= s
3
(t),
2
(t)),
s
3
(t) = s
3
(t) c
3,1
1
(t) c
3,2
2
(t).
Normalize s
3
(t) to unit energy:
3
(t) :=
s
3
(t)
_
|s
3
(t)|
2
.
4. Proceed in this way for all other waveforms s
k
(t) until k = D.
Notice that s
k
(t) = 0 for k > D. (Why?)
The general rule is the following. For k = 1, 2, . . . , D,
c
k,i
= s
k
(t),
i
(t)), i = 1, 2, . . . , k 1
s
k
(t) = s
k
(t)
k1
i=1
c
k,i
i
(t),
k
(t) :=
s
k
(t)
_
|s
k
(t)|
2
.
Due to this procedure, the resulting waveforms
k
(t) are orthog-
onal and normalized, as desired, and they form an orthonormal
basis of the D-dimensional vector space of the waveforms s
i
(t).
Remark: The set of orthonormal waveforms is not unique.
184
Example: Orthonormalization
Original signal waveforms:
PSfrag replacements
t
t t
t
s
1
(t)
s
2
(t)
s
3
(t)
s
4
(t)
0
0 0
0
1
1 1
1
1
1
1
1
2 2
2 2
3
3 3
3
1
1 1
1
1/
2
1/
2
Set 1 of orthonormal waveforms:
PSfrag replacements
t
t
t
1
(t)
2
(t)
3
(t)
0
0
0
1
1
1
1
2
2
2
3
3
3
1
1/
2
1/
2
1/
2
1/
2
Set 2 of orthonormal waveforms:
PSfrag replacements
t
t
t
1
(t)
2
(t)
3
(t)
0
0
0
1
1
1
1
1
1
2
2
2
3
3
3
1
1
1
1/
2
1/
2
3
185
B Orthogonal Transformation of a
White Gaussian Noise Vector
The statistical properties of a white Gaussian noise vector
(WGNV) are invariant with respect to orthonormal transforma-
tions. This is rst shown for a vector with two dimensions, and
then for a vector with an arbitrary number of dimensions.
B.1 The Two-Dimensional Case
Let W denote a 2-dimensional WGNV, i.e.,
W =
_
W
1
W
2
_
^
_
_
0
0
_
,
_
2
0
0
2
_
_
.
Consider the original coordinate system (
1
,
2
) and the rotated
coordinate system (
1
,
2
).
PSfrag replacements
2
= /4
w
w
1
w
2
w
1
w
2
Let [W
1
, W
2
]
T
denote the components of W expressed with respect
to the rotated coordinate system (
1
,
2
):
W
1
= +W
1
cos + W
2
sin
W
2
= W
1
sin + W
2
cos .
186
Writing this in matrix notation as
_
W
1
W
2
_
=
_
cos sin
sin cos
_ _
W
1
W
2
_
,
the vector W
can seen to be the orthonormal transformation of

the vector W.
The resulting vector W
= [W
1
, W
2
]
T
is also a two-dimensional
WGNV with the same covariance matrix as W = [W
1
, W
2
]
T
:
W
=
_
W
1
W
2
_
^
_
_
0
0
_
,
_
2
0
0
2
_
_
.
Proof: See problem.
B.2 The General Case
Let W = [W
1
, . . . , W
D
]
T
denote a D-dimensional WGNV with
zero mean and covariance matrix
2
I, i.e.,
W ^
_
0,
2
I
_
.
(0 is a column vector of length D, and I is the D D identity
matrix.) Let further V denote an orthonormal transformation
R
D
R
D
, i.e., a linear transformation with the property
V V
T
= V
T
V = I.
Then the transformed vector
W
= V W
is also a D-dimensional WGNV with the same covariance matrix
as W:
W
^
_
0,
2
I
_
.
187
Proof
(i) A linear transformation of a Gaussian random vector is again
a Gaussian random vector; thus W
is Gaussian.
(ii) The mean value is
E
_
W
= E
_
V W
= V E
_
W
= 0.
(iii) The covariance matrix is
E
_
W
W
T
= E
_
V WW
T
V
T
= V E
_
WW
T
. .
2
I
V
T
=
2
V V
T
. .
I
=
2
I.
Comment
In the case D = 2, V takes the form
V =
_
cos sin
sin cos
_
for some [0, 2).
188
C The Shannon Limit
C.1 Capacity of the Band-limited AWGN
Channel
Consider an AWGN channel with bandwidth W [1/sec]. The ca-
pacity of this channel is given by
C
W
= W log
2
_
1 +
S
WN
0
_
[bit/sec].
Consider further a digital transmission scheme operating at trans-
mission rate R [bit/sec] with limited transmit power S.
Channel Coding Theorem (C.E. Shannon, 1948):
Consider a binary random sequence generated by a BSS
that is transmitted over an AWGN channel with band-
width W using some transmission scheme.
Transmission with an arbitrarily small probability of er-
ror (using an appropriate coding/modulation scheme)
is possible provided that the transmission rate is less
than the channel capacity, i.e.,
P
b
0 is possible if R < C
W
.
Conversely, error-free transmission is impossible if the
rate is larger than the channel capacity, i.e.,
P
b
0 is impossible if R > C
W
.
189
A coding/modulation scheme can be characterized by its spectral
eciency and its power eciency. The spectral eciency is
the ratio R/W, and it has the unit
bit
sec Hz
. The power eciency
is the SNR
b
= E
b
/N
0
necessary to achieve a certain error prob-
ability (e.g., 10
5
).
Example: MPAM
The rate is given by R =
1
T
log
2
M. The bandwidth required to
transmit pulses of duration T is about W = 1/T. The spectral
eciency is thus
R
W
= log
2
M
bit
sec Hz
.
Error probabilities can be found in [1, p. 412, 435]. 3
Example: MPPM
The rate is given by R =
1
T
log
2
M. The bandwidth required
to transmit pulses of duration T/M is about W = M/T. The
spectral eciency is thus
R
W
=
log
2
M
M
bit
sec Hz
.
Error probabilities can be found in [1, p. 426, 435]. 3
190
A theoretical limit for the relation between the spectral eciency
and the power eciency (for error-free transmission) can be estab-
lished using the channel coding theorem.
The transmit power is given by S = RE
b
. Combining this relation,
the channel capacity, and the condition R < C
W
, we obtain
R
W
< log
2
_
1 +
RE
b
WN
0
_
= log
2
_
1 +
R
W
b
_
.
This can be rewritten as
b
>
2
R/W
1
R/W
.
Using the maximal rate for error-free transmission, R = C, we
obtain the theoretical bound
b
=
2
C/W
1
C/W
.
The plot can be found in [1, pp. 587].
The lower limit of
b
for R/W 0 is called the Shannon limit
for the AWGN channel:
b
> lim
R/W0
2
R/W
1
R/W
= ln 2.
Notice that 10 log
10
(ln 2) 1.6 dB. Thus, error-free transmis-
sion over the AWGN channel is impossible if the SNR per bit is
lower than 1.6 dB.
191
C.2 Capacity of the Band-Unlimited
AWGN Channel
For the AWGN channel with unlimited bandwidth (but still
limited transmit power), the capacity is given by
C
= lim
W
W log
2
_
1 +
S
WN
0
_
=
1
ln 2
S
N
0
.
The condition R < C
becomes then equivalent with the condition
b
> ln 2. (Remember: S = RE
b
.)

Digital Comm Notes

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

Digital Comm Notes

Uploaded by

Copyright:

Available Formats

Digital Modulation 1

contains the space spanned

spans the vector space

. However, only the components inside o

of orthonormal functions, i.e.,

(t) contains no information about s

(t) is completely represented by y.

^(0, 1). A plot of the symbol-error probability can found

) between two length-K vectors

is the number of positions in which they dier:

. Using this in the above sum yields

denotes the capacity of the (innite band-

(t) of the BFSK transmitter shows

(t) depends directly on the transmitted source bit U.

(t) of the previous example shows disconti-

(t), yielding the new phase function (t).

the set containing all linear combinations of these

is a vector space with respect to the following opera-

. Every element s(t) o

can seen to be the orthonormal transformation of

becomes then equivalent with the condition

You might also like