2
Why digital modulation?
If our goal was to design a digital
baseband communication system, we have
done that
Problem is baseband communication wont
takes us far, literally and figuratively
Digital modulation to a square pulse is
what analog modulation was to messages
3
A block diagram
Messsage
source
Source
coder
Line coder
Pulse
shaping
demodulator detector
channel
modulator
decision
1011
GEOMETRIC
REPRESENTATION OF
SIGNALS
5
The idea
We are used to seeing signals expressed
either in time or frequency domain
There is another representation space that
portrays signals in more intuitive format
In this section we develop the idea of
signals as multidimensional vectors
6
Have we seen this before?
Why yes! Remember the beloved e
j2fct
which can be written as
e
j2fct
=cos(2f
c
t)+jsin(2f
c
t)
inphase
quadrature
7
Expressing signals as a
weighted sum
Suppose a signal set consists of M signals
s
i
(t),I=1,,M. Each signal can be
represented by a linear sum of basis
functions
s
i
t ( ) = s
ij

j
t ( )
j =1
N
i =1,..., M
0 s t s T
8
Conditions on basis functions
For the expansion to hold, basis functions
must be orthonormal to each other
Mathematically:
Geometrically:

i
t ( )
j
t ( )
}
dt =
0 i = j
1 i = j

i

j

k
9
Components of the signal
vector
Each signal needs N numbers to be
represented by a vector. These N numbers
are given by projecting each signal onto
the individual basis functions:
s
ij
means projection of s
i
(t)on 
j
(t)
s
ij
= s
i
(t)
j
t ( )
0
T
}
dt
s
ij
s
i

j
10
Signal space dimension
How many basis functions does it take to
express a signal? It depends on the
dimensionality of the signal
Some need just 1 some need an infinite
number.
The number of dimensions is N and is
always less than the number of signals in
the set
N<=M
11
Example: Fourier series
Remember Fouirer series? A signal was
expanded as a linear sum of sines and
cosines of different frequencies. Sounds
familiar?
Sines and cosines are the basis functions
and are in fact orthogonal to each other
cos 2tnf
o
t ( )
T
o
}
cos 2tmf
o
t ( )dt = 0, m = n
f
o
=1/ T
o
12
Example: four signal set
A communication system sends one of 4
possible signals. Expand each signal in
terms of two given basis functions
1 1
1 1 2
1
0.5
2
1
13
Components of s
1
(t)
This is a 2Dsignal space. Therefore, each
signal can be represented by a pair of
numbers. Lets find them
For s
1
(t)
s
11
= s
1
(t)
1
t ( )
0
2
}
dt = 1 ( ) 1 ( )
0
1
}
dt + 0 =1
s
12
= s
1
(t)
2
t ( )
0
2
}
dt = 0 + 0.5 ( ) 1 ( )
0
1
}
dt = 0.5
t
t
1 2
1
1
0.5
s
1
(t)

1
s=(1,0.5)
14
Interpretation
s
1
(t) is now condensed into just two
numbers. We can reconstruct s
1
(t) like
this
s
1
(t)=(1)
1
(t)+(0.5)
2
(t)
Another way of looking at it is this
1
0.5

1

2
15
Signal constellation
Finding individual components of each
signal along the two dimensions gets us
the constellation
s4
s1
s2
s3

1

2
0.5
0.5 0.5
16
Learning from the
constellation
So many signal properties can be inferred
by simple visual inspection or simple math
Orthogonality:
s
1
and s
4
or orthogonal. To show that, simply find
their inner product, < s
1
, s
4
>
< s
1
, s
4
>=s
11
xs
41
+s
12
xs
42
(1)(0.5)+(1)(0.5)=0
17
Finding the energy from the
constellation
This is a simple matter. Remember,
Replace the signal by its expansion
E
i
= s
i
2
(t)dt
0
T
}
E
i
= s
ij

j
(t)
j =1
N
(
(
0
T
}
s
ik

k
(t)
k =1
N
(
(
dt
18
Exploiting the orthogonality
of basis functions
Expanding the summation, all cross
product terms integrate to zero. What
remains are N terms where j=k
E
i
= s
ij
2

j
2
t ( )
j =1
N
(
(
0
T
}
dt = s
ij
2

j
2
t ( )dt
0
T
}
(
(
j=1
N
s
ij
2

j
2
t ( )dt
0
T
}
=1
j =1
N
= s
ij
2
j =1
N
19
Energy in simple language
What we just saw says that the energy of a
signal is simply the square of the length of
its corresponding constellation vector
3
2
E=9+4=13
E
20
Constrained energy signals
Lets say you are under peak energy E
p
constraint in your application. Just make
sure all your signals are inside a circle of
radius sqrt(E
p
)
E
p
21
Correlation of two signals
A very desirable situation in is to have
signals that are mutually orthogonal. How
do we test this? Find the angle between
them
u
s1
s2
cos u
12
( ) =
s
1
T
s
2
s
1
s
2
transpose
22
Find the angle between s
1
and
s
2
Given that s
1
=(1,2)
T
and s
2
=(2,1)
T
, what is
the angle between the two?
s
1
T
s
2
= 1 2
 
2
1
(
(
= 2 + 2 = 4
s
1
= 1+ 4 = 5
s
2
= 4 +1 = 5
cos u
12
( ) =
4
5 5
=
4
5
u
12
= 36.9
o
23
Distance between two signals
The closer signals are together the more
chances of detection error. Here is how we
can find their separation
d
12
2
= s
1
s
2
2
= s
1j
s
2 j
( )
2
j =1
N
= (1)
2
+ (1)
2
= 2
d
12
= 2
1 2
1
2
24
Constellation building using
correlator banks
We can decompose the signal into its
components as follows
s(t)

1

2

N
dt
0
T
}
dt
0
T
}
dt
0
T
}
s
1
s
2
s
N
N components
25
Detection in the constellation
space
Received signal is put through the filter
bank below and mapped to a point
s(t)

1

2

N
dt
0
T
}
dt
0
T
}
dt
0
T
}
s
1
s
2
s
N
components
mapped to a single point
26
Constellation recovery in
noise
Assume signal is contaminated with noise.
All N components will also be affected.
The original position of s
i
(t) will be
disturbed
27
Actual example
Here is a 16level constellation which is
reconstructed in the presence of noise
2 1.5 1 0.5 0 0.5 1 1.5 2
2
1.5
1
0.5
0
0.5
1
1.5
2
Eb/No=5 dB
28
Detection in signal space
One of the M allowable signals is
transmitted, processed through the bank of
correlators and mapped onto constellation
question is based on what we see , what
was the transmitted signal?
received signal
which of the four did it
come from
29
Minimum distance decision
rule
It can be shown that the optimum decision,
in the sense of lowest BER, is to pick the
signal that is closest to the received
vector. This is called maximum likelihood
decision making
this is the most likely
transmitted signal
received
30
Defining decision regions
An easy detection method, is to compute
decision regions offline. Here are a few
examples
decide s1
decide s2
s1
s2
measurement
decide s1
decide s2
decide s3 decide s4
s1
s2
s3
s4
decide s1
s1
31
More formally...
Partition the decision space into M
decision regions Z
i
, i=1,,M. Let X be the
measurement vector extracted from the
received signal. Then
if XeZ
i
s
i
was transmitted
32
How does detection error
occur?
Detection error occurs when X lands in Z
i
but it wasnt s
i
that was transmitted.
Noise, among others, may be the culprit
departure from transmitted
position due to noise
X
s
i
33
Error probability
we can write an expression for error like
this
P{errors
i
}=P{X does not lie in Z
i
s
i
was
transmitted}
Generally
P
e
= P X eZ
i
 s
i
{ }P{
i =1
M
s
i
}
34
Example: BPSK
(binary phase shift keying)
BPSK is a well known digital modulation
obtained by carrier modulating a polar NRZ
signal. The rule is
1: s
1
=Acos(2f
c
t)
0:s
2
=  Acos(2f
c
t)
1s and 0s are identified by 180 degree
phase reversal at bit transitions
35
Signal space for BPSK
Look at s
1
and s
2
. What is the basis
function for them? Both signals can be
uniquely written as a scalar multiple of a
cosine. So a single cosine is the sole basis
function. We have a 1D constellation
A A
cos(2pif
c
t)
36
Bringing in E
b
We want each bit to have an energy E
b
.
Bits in BPSK are RF pulses of amplitude A
and duration T
b
. Their energy is A
2
T
b
/2 .
Therefore
E
b
= A
2
T
b
/2 >A=sqrt(2E
b
/T
b
)
We can write the two bits as follows
s
1
t ( ) =
2E
b
T
b
cos 2tf
c
t ( )
s
2
t ( ) =
2E
b
T
b
cos 2tf
c
t
( )
37
BPSK basis function
As a 1D signal, there is one basis function.
We also know that basis functions must
have unit energy. Using a normalization
factor
E

=1

1
t ( ) =
2
T
b
cos 2tf
c
t ( )
38
Formulating BER
BPSK constellation looks like this
E
b
E
b
X1=[E
b
+n,n]
transmitted
received
noise
P
e1
= P E
b
+ n < 0  1 is transmitted
{ }
if noise is negative enough, it will push
X to the left of the boundary, deciding 0
instead
39
Finding BER
Lets rewrite BER
But n is gaussian with mean 0 and
variance N
o
/2
P
e1
= P E
b
+ n < 0  1
{ }
= P n < E
b
{ }
sqrt(E
b
)
40
BER for BPSK
Using the trick to find the area under a
gaussian density(after normalization with
respect to variance)
BER=Q[(2E
b
/No)
0.5
]
or
BER=0.5erfc[(E
b
/No)
0.5
]
41
BPSK Example
Data is transmitted at R
b
=10
6
b/s. Noise
PSD is 10
6
and pulses are rectangular with
amplitude 0.2 volt. What is the BER?
First we need energy per bit, E
b
. 1s and 0s
are sent by
2E
b
T
b
cos(2tf
c
t)
2E
b
T
b
= 0.2
42
Solving for E
b
Since bit rate is 10
6
, bit length must be
1/R
b
=10
6
Therefore,
E
b
=20x10
6
=20 wsec
Remember, this is the received energy.
What was transmitted are probably several
orders of magnitude bigger
43
Solving for BER
Noise PSD is N
o
/2 =10
6
. We know for BPSK
BER=0.5erfc[(E
b
/No)
0.5
]
What we have is then
Finish this using erf tables
BER=
1
2
erfc
E
b
N
o

\

.
 =
1
2
erfc
2 10
7
2 10
6

\

.

=
1
2
erfc( 0.1) =
1
2
erfc(0.316)
44
Binary FSK
(Frequency Shift Keying)
Another method to transmit 1s and 0s is
to use two distinct tones, f
1
and f
2
of the
form below
But what is the requirements on the tones?
Can they be any tones?
s
i
t ( ) =
2E
b
T
b
cos 2tf
i
t ( ), 0 s t s T
b
0
45
Picking the right tones
It is desirable to keep the tones orthogonal
Since tones are sinusoids, it is sufficient
for the tones to be separated by an integer
multiple of inverse duration, i.e.
f
i
=
n
c
+ i
T
b
, i =1, 2
n
c
= some integer
46
Example tones
Lets say we are sending data at the rate
of 1 Mb/sec in BFSK, What are some
typical tones?
Bit length is 10
6
sec. Therefore, possible
tones are (use n
c
=0)
f
1
=1/T
b
=1 MHz
f
2
=2/T
b
=2MHz
47
BFSK dimensionality
What does the constellation of BFSK look
like? We first have to find its dimension
s
1
and s
2
can be represented by two
orthonormal basis functions:
Notice f
1
and f
2
are selected to make them
orthogonal

i
t ( ) =
2
T
b
cos 2tf
i
t ( ), 0 s t s T
b
48
BFKS constellation
There are two dimensions. Find the
components of signals along each
dimension using
s
11
= s
1
t ( )
0
T
b
}

1
t ( )dt = E
b
s
12
= s
1
t ( )
0
T
b
}

2
t ( )dt = 0
s
1
= ( E
b
, 0)
E
b
E
b
49
Decision regions in BFSK
Decisions are made based on distances.
Signals closer to s
1
will be classified as s
1
and vice versa
50
Detection error in BFSK
Let the received signal land where shown.
Assume s
1
is sent. How would a detection
error occur?
x
2
>x
1
puts X in the
s
2
partition
s
1
s
2
X=received
x
1
x
2
P
e1
=P{x
2
>x
1
s
1
was sent}
51
Where do (x
1
,x
2
) come from?
Use the correlator bank to extract signal
components
x=
s
1
(t)+noise

1

2
dt
0
T
b
}
dt
0
T
b
}
x
1
(gaussian)
x
2
(gaussian)
52
Finding BER
We have to answer this question: what is
the probability of one random variable
exceeding another random variable?
To cast P(x
2
>x
1
) into like of P(x>2), rewrite
P(x
2
>x
1
x
1
)
x
1
is now treated as constant. Then,
integrate out x
1
to eliminate it
53
BER for BFSK
Skipping the details of derivation, we get
P
e
= BER=
1
2
erfc
E
b
2N
o

\

.

54
BPSK and BFSK comparison:
energy efficiency
Lets compare their
BERs
P
e
=
1
2
erfc
E
b
2N
o

\

.

, BFSK
P
e
=
1
2
erfc
E
b
N
o

\

.

, BPSK
What does it take to
have the same BER?
E
b
in BFSK must be
twice as big as BPSK
Conclusion: energy per
bit must be twice as
large in BFSK to
achieve the same BER
55
Comparison in the
constellation space
Distances determine BERs. Lets compare
Both have the same E
b
, but BPSKs are
farther apart, hence lower BER
E
b
E
b
2 E
b
E
b
E
b
1.4 E
b
56
Differential PSK
Concept of differential encoding is very
powerful
Take the the bit sequence 11001001
Differentially encoding of this stream
means that we start we a reference bit and
then record changes
57
Differential encoding example
Data to be encoded
1 0 0 1 0 0 1 1
Set the reference bit to 1, then use the
following rule
Generate a 1 if no change
Generate a 0 if change
1 0 0 1 0 0 1 1
1 1 0 1 1 0 1 1 1
58
Detection logic
Detecting a differentially encoded signal is
based on the comparison of two adjacent
bits
If two coded bits are the same, that means
data bit must have been a 1, otherwise 0
? ? ? ? ? ? ? ?
1 1 0 1 1 0 1 1 1
Encoded received
bits
unknown transmitted
bits
59
DPSK: generation
Once data is differentially encoded, carrier
modulation can be carried out by a straight
BPSK encoding
Digit 1:phase 0
Digit 0:phase 180
1 1 0 1 1 0 1 1 1
0 0 0 0 0 0 0
Differentially encoded data
Phase encoded(BPSK)
60
DPSK detection
Data is detected by a phase comparison of
two adjacent pulses
No phase change: data bit is 1
Phase change: data bit is 0
0 0 0 0 0 0 0
1 0 0 1 0 0 1 1
Detected data
61
Bit errors in DPSK
Bit errors happen in an interesting way
Since detection is done by comparing
adjacent bits, errors have the potential of
propagating
Allow a single detection error in DPSK
0 0 0 0 0 0
1 0 1 0 0 0 1 1
1 0 0 1 0 0 1 1
Back on track:no errors
Transmitted bits
Incoming phases
Detected bits
2 errors
62
Conclusion
In DPSK, if the phase of the RF pulse is
detected in error, error propagates
However, error propagation stops quickly.
Only two bit errors are misdetected. The
rest are correctly recovered
63
Why DPSK?
Detecting regular BPSK needs a coherent
detector, requiring a phase reference
DPSK needs no such thing. The only
reference is the previous bit which is
readily available
64
Mary signaling
Binary communications sends one of only 2
levels; 0 or 1
There is another way: combine several bits
into symbols
1 0 1 1 0 1 1 0 1 1 1 0 0 1 1
Combining two bits at a time gives rise to
4 symbols; a 4ary signaling
65
8level PAM
Here is an example of 8level signaling
0 1 0 1 0 0 0 0 0 0 0 1 1 1 0 1 0 0 1 1 1
binary
7
5
3
2
1
1
3
5
7
66
A few definitions
We used to work with bit length T
b
. Now
we have a new parameter which we call
symbol length,T
1 1
0
T
T
b
67
Bit lengthsymbol length
relationship
When we combine n bits into one symbol;
the following relationships hold
T=nT
b
 symbol length
n=logM bits/symbol
T=T
b
xlogM symbol length
All logarithms are base 2
68
Example
If 8 bits are combined into one symbol, the
resulting symbol is 8 times wider
Using n=8, we have M=2
8
=256 symbols to
pick from
Symbol length T=nT
b
=8T
b
69
Defining baud
When we combine n bits into one symbol,
numerical data rate goes down by a factor
of n
We define baud as the number of
symbols/sec
Symbol rate is a fraction of bit rate
R=symbol rate=R
b
/n=R
b
/logM
For 8level signaling, baud rate is 1/3 of bit
rate
70
Why Mary?
Remember Nyquist bandwidth? It takes a
minimum of R/2 Hz to transmit R
pulses/sec.
If we can reduce the pulse rate, required
bandwidth goes down too
Mary does just that. It takes R
b
bits/sec
and turns it into R
b
/logM pulses sec.
71
Issues in transmitting 9600
bits/sec
Want to transmit 9600 bits/sec. Options:
Nyquists minimum bandwidth:9600/2=4800 Hz
Full roll off raised cosine:9600 Hz
None of them fit inside the 4 KHz wide
phone lines
Go to a 16  level signaling, M=16. Pulse
rate is reduced to
R=R
b
/logM=9600/4=2400 Hz
72
Using 16level signaling
Go to a 16level signaling, M=16. Pulse rate
is then cut down to
R=R
b
/logM=9600/4=2400 pulses/sec
To accommodate 2400 pulses /sec, we
have several options. Using sinc we need
only 1200 Hz. Full rolloff needs 2400Hz
Both fit within the 4 KHz phone line
bandwidth
73
Bandwidth efficiency
Bandwidth efficiency is defined as the
number of bits that can be transmitted
within 1 Hz of bandwidth
q=R
b
/B
T
bits/sec/Hz
In binary communication using sincs,
B
T
=R
b
/2> q=2 bits/sec/Hz
74
Mary bandwidth efficiency
In Mary signaling , pulse rate is given by
R=R
b
/logM. Full rolloff raised cosine
bandwidth is B
T
=R= R
b
/logM.
Bandwidth efficiency is then given by
q=R
b
/B
T
=logM bits/sec/Hz
For M=2, binary we have 1 bit/sec/Hz. For
M=16, we have 4 bits/sec/Hz
75
Mary bandwidth
Summarizing, Mary and binary bandwidth
are related by
B
Mary
=B
binary
/logM
Clearly , Mary bandwidth is reduced by a
factor of logM compared to the binary
bandwidth
76
8ary bandwidth
Let the bit rate be 9600 bits/sec. Binary
bandwidth is nominally equal to the bit
rate, 9600 Hz
We then go to 8level modulation (3
bits/symbol) Mary bandwidth is given by
B
Mary
=B
binary
/logM=9600/log8=3200 Hz
77
Bandwidth efficiency
numbers
Here are some numbers
n(bits/symbol) M(levels) q(bits/sec/Hz)
1 2 1
2 4 2
3 8 3
4 16 4
8 256 8
78
Symbol energy vs. bit energy
Each symbol is made up of n bits. It is not
therefore surprising for a symbol to have n
times the energy of a bit
E(symbol)=nE
b
E
b
E
79
QPSK
quadrature phase shift keying
This is a 4 level modulation.
Every two bits is combined and mapped to
one of 4 phases of an RF signal
These phases are 45
o
,135
o
,225
o
,315
o
s
i
(t) =
2E
T
cos 2tf
c
t +(2i 1)
t
4
(
, i =1, 2,3, 4
0
, 0 s t s T
Symbol energy
Symbol width
80
QPSK constellation
45
o
00
01
11
10
E

1
t ( ) =
2
T
cos2tf
c
t

2
t ( ) =
2
T
sin 2tf
c
t
Basis functions
S=[0.7 E, 0.7 E]
81
QPSK decision regions
00
01
11
10
Decision regions re colorcoded
82
QPSK error rate
Symbol error rate for QPSK is given by
This brings up the distinction between
symbol error and bit error. They are not the
same!
P
e
= erfc(
E
2N
o
)
83
Symbol error
Symbol error occurs when received vector
is assigned to the wrong partition in the
constellation
When s
1
is mistaken for s
2
, 00 is mistaken
for 11
00
11
s1
s2
84
Symbol error vs. bit error
When a symbol error occurs, we might
suffer more than one bit error such as
mistaking 00 for 11.
It is however unlikely to have more than
one bit error when a symbol error occurs
10 10 11 10 00
11 10 11 10 00
10 symbols = 20 bits
Sym.error=1/10
Bit error=1/20
85
Interpreting symbol error
Numerically, symbol error is larger than bit
error but in fact they are describing the
same situation; 1 error in 20 bits
In general, if P
e
is symbol error
P
e
logM
s BER s P
e
86
Symbol error and bit error for
QPSK
We saw that symbol error for QPSK was
Assuming no more than 1 bit error for each
symbol error, BER is half of symbol error
Remember symbol energy E=2E
b
P
e
= erfc(
E
2N
o
)
BER =
1
2
erfc(
E
2N
o
)
87
QPSK vs. BPSK
Lets compare the two based on BER and
bandwidth
BER Bandwidth
BPSK QPSK BPSK QPSK
1
2
erfc
E
b
N
o

\

.

1
2
erfc
E
b
N
o

\

.

R
b
R
b
/2
EQUAL
88
Mphase PSK (MPSK)
If you combine 3 bits into one symbol, we
have to realize 2
3
=8 states. We can
accomplish this with a single RF pulse
taking 8 different phases 45
o
apart
s
i
(t) =
2E
T
cos 2tf
c
t +(i 1)
t
4
(
, i =1,...,8
0
, 0 s t s T
89
8PSK constellation
Distribute 8 phasors uniformly around a
circle of radius E
45
o
Decision region
90
Symbol error for MPSK
We can have M phases around the circle
separated by 2/M radians.
It can be shown that symbol error
probability is approximately given by
P
e
~ erfc
E
N
o
sin
t
M

\

.

\

.
 , M > 4
91
Quadrature Amplitude
Modulation (QAM)
MPSK was a phase modulation scheme. All
amplitudes are the same
QAM is described by a constellation
consisting of combination of phase and
amplitudes
The rule governing bitstosymbols are the
same, i.e. n bits are mapped to M=2
n
symbols
92
16QAM constellation using
Gray coding
16QAM has the following constellation
Note gray coding
where adjacent symbols
differ by only 1 bit
0010 0011 0001 0000
1010
1110
0110
1011
1111
0111
1001
1101
0101
1000
1100
0100
93
Vector representation
of 16QAM
There are 16 vectors, each defined by a
pair of coordinates. The following 4x4
matrix describes the 16QAM constellation
[a
i
, b
i
] =
3, 3 ( ) 1, 3 ( ) 1, 3 ( ) 3, 3 ( )
3,1 ( ) 1,1 ( ) 1,1 ( ) 3,1 ( )
3, 1 ( ) 1, 1 ( ) 1, 1 ( ) 3, 1 ( )
3, 3 ( ) 1, 3 ( ) 1, 3 ( ) 3, 3 ( )
(
(
(
(
94
What is energy per symbol in
QAM?
We had no trouble defining energy per
symbol E for MPSK. For QAM, there is no
single symbol energy. There are many
We therefore need to define average
symbol energy E
avg
E
avg
=
1
M
a
i
2
+ b
i
2
( )
i =1
M
95
E
avg
for 16QAM
Using the [a
i
,b
i
] matrix and using
E=a
i
^2+b
i
^2 we get one energy per signal
E =
18 10 10 18
10 2 2 10
10 2 2 10
18 10 10 18
(
(
(
(
E
avg
=10
96
Symbol error for Mary QAM
With the definition of energy in mind,
symbol error is approximated by
P
e
~ 2 1
1
M

\

.
erfc
2E
avg
2 M 1 ( )N
o

\

.

97
Familiar constellations
Here are a few golden oldies
V.22
600 baud
1200 bps
V.22 bis
600 baud
2400 bps
V.32 bis
2400 baud
9600 bps
98
Mary FSK
Using M tones, instead of M
phases/amplitudes is a fundamentally
different way of Mary modulation
The idea is to use M RF pulses. The
frequencies chosen must be orthogonal
s
i
t ( ) =
2E
T
cos 2tf
i
t ( ), 0 s t s T
i =1,..., M
99
MFSK constellation:
3dimensions
MFSK is different from MPSK in that each
signal sits on an orthogonal axis(basis)
s
1
s
2
s
3

1

2

3

i
t ( ) =
2
T
cos 2tf
i
t
( )
,
0 s t s T
i =1,..., M
s
1
=[E ,0, 0]
s
2
=[0,E, 0]
s
3
=[0,0,E]
E
E
E
100
Orthogonal signals:
How many dimensions, how many
signals?
We just saw that in a 3 dimensional space,
we can have no more than 3 orthogonal
signals
Equivalently, 3 orthogonal signals dont
need more than 3 dimensions because
each can sit on one dimension
Therefore, number of dimensions is always
less than or equal to number of signals
101
How to pick the tones?
Orthogonal FSK requires tones that are
orthogonal.
Two carrier frequencies separated by
integer multiples of period are orthogonal
102
Example
Take two tones one at f
1
the other at f
2
. T
must cover one or more periods for the
integral to be zero
2cos 2tf
1
t ( )cos 2tf
2
t ( )dt = cos2t f
1
+ f
2
( )dt
0
T
}
averages to zero
0
T
}
+ cos2t f
1
f
2
( )dt
0
T
}
averages to zero if T =i/(f1 f2)
; i=integer
Take f
1
=1000 and T=1/1000. Then
if f
2
=2000 , the two are orthogonal
so will f
2
=3000,4000 etc
103
MFSK symbol error
Here is the error expression with the usual
notations
P
e
s
1
2
M 1 ( )erfc
E
2N
o

\

.

104
Spectrum of Mary signals
So far E
b
/N
o
, i.e. power, has been our main
concern. The flip side of the coin is
bandwidth.
Frequently the two move in opposite
directions
Lets first look at binary modulation
bandwidth
105
BPSK bandwidth
Remember BPSK was obtained from a
polar signal by carrier modulation
We know the bandwidth of polar NRZ using
square pulses was B
T
=R
b
.
It doesnt take much to realize that carrier
modulation doubles this bandwidth
106
Illustrating BPSK bandwidth
The expression for baseband BPSK (polar)
bandwidth is
S
B
(f)=2E
b
sinc
2
(T
b
f)
B
T
=2R
b
f 1/T
b
BPSK
f
c
+/T
b
f
c
/T
b
f
c
2/T
b
=2R
b
107
BFSK as a sum of two RF
streams
BFSK can be thought of superposition of
two unipolar signals, one at f
1
and the
other at f
2
0 1000 2000 3000 4000 5000 6000 7000 8000
1
0.5
0
0.5
1
0 1000 2000 3000 4000 5000 6000 7000 8000
1
0.5
0
0.5
1
0 1000 2000 3000 4000 5000 6000 7000 8000
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
BFSK for 1 0 0 1 0 1 1
+
108
Modeling of BFSK bandwidth
Each stream is just a carrier modulated
unipolar signal. Each has a sinc spectrum
f
1
f
2
1/T
b
=R
b
f
c
f
c
=(f
1
+f
2
)/2
Af
B
T
=2 Af+2R
b
Af= (f
2
f
1
)/2
109
Example: 1200 bps bandwidth
The old 1200 bps standard used BFSK
modulation using 1200 Hz for mark and
2200 Hz for space. What is the bandwidth?
Use
B
T
=2Af+2R
b
Af=(f
2
f
1
)/2=(22001200)/2=500 Hz
B
T
=2x500+2x1200=3400 Hz
This is more than BPSK of 2R
b
=2400 Hz
110
Sundes FSK
We might have to pick tones f
1
and f
2
that
are not orthogonal. In such a case there
will be a finite correlation between the
tones
=
2
T
b
cos(2tf
1
t)
0
T
b
}
cos(2tf
2
t)dt
1 2 3
2(f
2
f
1
)T
b
Good points,zero correlation
111
Picking the 2nd zero crossing:
Sundes FSK
If we pick the second zc term (the first
term puts the tones too close) we get
2(f
2
f
1
)T
b
=2> Af=1/2T
b
=R
b
/2
remember Af is (f
2
f
1
)/2
Sundes FSK bandwidth is then given by
B
T
=2Af+2R
b
=R
b
+2R
b
=3R
b
The practical bandwidth is a lot smaller
112
Sundes FSK bandwidth
Due to sidelobe cancellation, practical
bandwidth is just B
T
=2Af=R
b
f
1
f
2
1/T
b
=R
b
f
c
f
c
=(f
1
+f
2
)/2
Af
B
T
=2 Af+2R
b
Af= (f
2
f
1
)/2
Af
113
B FSK example
A BFSK system operates at the 3rd zero
crossing of T
b
plane. If the bit rate is 1
Mbps, what is the frequency separation of
the tones?
The 3rd zc is for 2(f
2
f
1
)T
b
=3. Recalling that
Af=(f
2
f
1
)/2 then Af =0.75/T
b
Then Af =0.75/T
b
=0.75x10
6
=750 KHz
And B
T
=2(Af +R
b
)=2(0.75+1)10
6
=3.5 MHz
114
Point to remember
FSK is not a particularly bandwidthfriendly
modulation. In this example, to transmit 1
Mbps, we needed 3.5 MHz.
Of course, it is working at the 3rd zero
crossing that is responsible
Original Sundes FSK requires B
T
=R
b
=1 MHz
Bandwidth of MPSK
modulation
116
MPSK bandwidth review
In MPSK we used pulses that are log
2
M
times wider tan binary hence bandwidth
goes down by the same factor.
T=symbol width=T
b
log
2
M
For example, in a 16phase modulation,
M=16, T=4T
b
.
B
qpsk
=B
bpsk
/log
2
M= B
bpsk
/4
117
MPSK bandwidth
MPSK spectrum is given by
S
B
(f)=(2E
b
log
2
M)sinc
2
(T
b
flog
2
M)
f/R
b
Notice normalized frequency
1/logM
Set to 1 for zero crossing BW
T
b
flog
2
M=1
>f=1/ T
b
flog
2
M
=R
b
/log
2
M
B
T
= R
b
/log
2
M
118
Bandwidth after carrier
modulation
What we just saw is MPSK bandwidth in
baseband
A true MPSK is carrier modulated. This will
only double the bandwidth. Therefore,
Bmpsk=2R
b
/log
2
M
119
QPSK bandwidth
QPSK is a special case of MPSK with M=4
phases. Its baseband spectrum is given by
S
B
(f)=2Esinc
2
(2T
b
f)
f/R
b
0.5
B=0.5R
b
>
half of BPSK
1
After modulation:
B
qpsk
=R
b
120
Some numbers
Take a 9600 bits/sec data stream
Using BPSK: B=2Rb=19,200 Hz (too much
for 4KHz analog phone lines)
QPSK: B=19200/log
2
4=9600Hz, still high
Use 8PSK:B= 19200/log
2
8=6400Hz
Use 16PSK:B=19200/ log
2
16=4800 Hz. This
may barely fit
121
MPSK vs.BPSK
Lets say we fix BER at some level. How do
bandwidth and power levels compare?
M Bmary/Bbinary (Avg.power)M/(Avg.power)bin
4 0.5 0.34 dB
8 1/3 3.91 dB
16 1/4 8.52 dB
32 1/5 13.52 dB
Lesson: By going to multiphase modulation, we save
bandwidth but have to pay in increased power, But why?
122
Powerbandwidth tradeoff
The goal is to keep BER fixed as we
increase M. Consider an 8PSK set.
What happens if you go to 16PSK? Signals
get closer hence higher BER
Solution: go to a larger circle>higher
energy
123
Additional comparisons
Take a 28.8 Kb/sec data rate and lets
compare the required bandwidths
BPSK: B
T
=2(R
b
)=57.6 KHz
BFSK: B
T
= R
b
=28.8 KHz ...Sundes FSK
QPSK: B
T
=half of BPSK=28.8 KHz
16PSK: B
T
=quarter of BPSK=14.4 KHz
64PSK: B
T
=1/6 of BPSK=9.6 KHz
124
Powerlimited systems
Modulations that are powerlimited achieve
their goals with minimum expenditure of
power at the expense of bandwidth.
Examples are MFSK and other orthogonal
signaling
125
Bandwidthlimited systems
Modulations that achieve error rates at a
minimum expenditure of bandwidth but
possibly at the expense of too high a
power are bandwidthlimited
Examples are variations of MPSK and many
QAM
Check BER rate curves for BFSK and
BPSK/QAM cases
126
Bandwidth efficiency index
A while back we defined the following ratio
as a bandwidth efficiency measure in
bits/sec/HZ
=R
b
/B
T
bits/sec/Hz
Every digital modulation has its own
127
for MPSK
At a bit rate of R
b
, BPSK bandwidth is 2R
b
When we go to MPSK, bandwidth goes
down by a factor of log
2
M
B
T
=2Rb/ log
2
M
Then
=R
b
/B
T
= log
2
M/2 bits/sec/Hz
128
Some numbers
Lets evaluate vs. M for MPSK
M 2 4 8 16 32 64
.5 1 1.5 2 2.5 3
Notice that bits/sec/Hz goes up by a factor
of 6 from M=2 and M=64
The price we pay is that if power level is
fixed (constellation radius fixed) BER will
go up. We need more power to keep BER
the same
129
Defining MFSK:
In MFSK we transmit one of M frequencies
for every symbol duration T
These frequencies must be orthogonal.
One way to do that is to space them 1/2T
apart. They could also be spaced 1/T apart.
Following The textbook we choose the
former (this corresponds to using the first
zero crossing of correlation curve)
130
MFSK bandwidth
Symbol duration in MFSK is M times longer
than binary
T=T
b
log
2
M symbol length
Each pair of tones are separated by 1/2T. If
there are M of them,
B
T
=M/2T=M/2T
b
log
2
M
>B
T
=MR
b
/2log
2
M
131
Contrast with MPSK
Variation of bandwidth with M differs
drastically compared to MPSK
MPSK MFSK
B
T
=2R
b
/log
2
M B
T
=MR
b
/2log
2
M
As M goes up, MFSK eats up more
bandwidth but MPSK save bandwidth
132
MFSK bandwidth efficiency
Lets compute s for MFSK
=R
b
/M=2log
2
M/M bits/sec/HzMFSK
M 2 4 8 16 32 64
1 1 .75 .5 .3 .18
Notice bandwidth efficiency drop. We are
sending fewer and fewer bits per 1 Hz of
bandwidth
COMPARISON OF DIGITAL
MODULATIONS*
*B. Sklar, Defining, Designing and Evaluating Digital Communication Systems,
IEEE Communication Magazine, vol. 31, no.11, November 1993, pp. 92101
134
Notations
M = 2
m
# of symbols
m = log
2
M bits/symbol
R =
m
T
s
=
log
2
M
T
s
bits/ sec
T
s
= symbol duration
R
s
= symbol rate
T
b
=
1
R
=
T
s
m
=
1
mR
s
bit length
Bandwidth efficiency
measure
R
W
=
log
2
M
WT
s
=
1
WT
b
135
Bandwidthlimited Systems
There are situations where bandwidth is at
a premium, therefore, we need
modulations with large R/W.
Hence we need standards with large time
bandwidth product
The GSM standard uses Gaussian minimum
shift keying(GMSK) with WT
b
=0.3
136
Case of MPSK
In MPSK, symbols are m times as wide as
binary.
Nyquist bandwidth is W=R
s
/2=1/2Ts.
However, the bandpass bandwidth is twice
that, W=1/Ts
Then
R
W
=
log
2
M
WT
s
= log
2
M bits/sec/Hz
137
Cost of Bandwidth Efficiency
As M increases, modulation becomes more
bandwidth efficient.
Lets fix BER. To maintain this BER while
increasing M requires an increase in E
b
/N
o
.
138
PowerLimited Systems
There are cases that bandwidth is
available but power is limited
In these cases as M goes up, the
bandwidth increases but required power
levels to meet a specified BER remains
stable
139
Case of MFSK
MFSK is an orthogonal modulation scheme.
Nyquist bandwidth is Mtimes the binary
case because of using M orthogonal
frequencies, W=M/Ts=MR
s
Then
R
W
=
log
2
M
WT
s
=
log
2
M
M
bits/sec/Hz
140
Select an Appropriate
Modulation
We have a channel of 4KHz with an
available S/No=53 dBHz
Required data rate R=9600 bits/sec.
Required BER=10
5
.
Choose a modulation scheme to meet
these requirements
141
Minimum Number of Phases
To conserve power, we should pick the
minimum number of phases that still meets
the 4KHz bandwidth
A 9600 bits/sec if encoded as 8PSK results
in 3200 symbols/sec needing 3200Hz
So, M=8
142
What is the required E
b
/No?
S
N
o
=
E
b
R
N
o
=
E
b
N
o
R
E
b
N
o
(dB) =
S
N
o
(dB Hz) R(dBbits/ sec
=13.2dB
143
Is BER met? Yes
The symbol error probability in 8PSK is
Solve for Es/No
Solve for PE
P
E
M ( ) = 2Q
2E
s
N
o
sin
t
M

\

.
(
(
BER=
P
E
log
2
M
=
2.2 10
5
3
= 7.3 10
6
E
s
N
o
= log
2
M ( )
E
b
N
0
= 3 20.89 = 62.67
144
Powerlimited uncoded
system
Same bit rate and BER
Available bandwidth W=45 KHz
Available S/No=48dBHz
Choose a modulation scheme that yields
the required performance
145
Binary vs. Mary Model
Mary Modulator R bits/s
R
s
=
R
log
2
M
symbols/ s
Mary demodulator
S
N
o
=
E
b
N
o
R =
E
s
N
o
R
s
146
Choice of Modulation
With R=9600 bits/sec and W=45 KHz, the
channel is not bandwidth limited
Lets find the available Eb/No
E
b
N
o
(dB) =
S
N
o
dB Hz ( ) R(dB bit / s) =
E
b
N
o
(dB) = 48dB Hz
= (10log9600)dB bits / s
= 8.2dB
147
Choose MFSK
We have a lot of bandwidth but little power
>orthogonal modulation(MFSK)
The larger the M, the more power
efficiency but more bandwidth is needed
Pick the largest M without going beyond
the 45 KHz bandwidth.
148
MFSK Parameters
From Table 1, M=16 for an MFSK
modulation requires a bandwidth of 38.4
KHz for 9600 bits/sec data rate
We also wanted to have a BER<10^5.
Question is if this is met for a 16FSK
modulation.
149
16FSK
Again from Table 1, to achieve BER of 10^
5 we need E
b
/No of 8.1dB.
We solved for the available E
b
/No and that
came to 8.2dB
150
Symbol error for MFSK
For noncoherent orthogonal MFSK, symbol
error probability is
P
E
M ( ) s
M 1
2
exp
E
s
2N
o

\

.

E
s
= E
b
log
2
M
151
BER for MFSK
We found out that Eb/No=8.2dB or 6.61
Relating Es/No and Eb/No
BER and symbol error are related by
E
s
N
o
= log
2
M ( )
E
b
N
o
P
B
=
2
m1
2
m
1
P
E
152
Example
Lets look at the 16FSK case. With 16
levels, we are talking about m=4 bits per
symbol. Therefore,
With Es/No=26.44, symbol error prob.
PE=1.4x10^5>PB=7.3x10^6
P
B
=
2
3
2
4
1
P
E
=
8
15
P
E
153
Summary
Given:
R=9600 bits/s
BER=10^5
Channel bandwith=45
KHz
E
b
/No=8.2dB
Solution
16FSK
required bw=38.4khz
required E
b
/No=8.1dB