Digital Modulation Methods

in AWGN Channels (2)

Sina Vafi
College of Engineering, IT and Environment
Matched-filter type demodulator

Binary antipodal signals:

r(t)=smψ (t) + n(t) 0 ≤t < Tb

Assume the received signal r(t) passes through a linear time-invariant filter

h(t)=ψ(Tb-t) 0≤ t < Tb

• The filter output: y(t)= ℎ −

• We sample the output of the filter at t=Tb, y(Tb)= ℎ −
• We know h(Tb-τ)=ψ (τ). Therefore,

[ + ( )] = ( ) + ( ) = +

Conclusion: The output of filter at t=Tb is the same as the output obtained from the cross
correlator

A filter with impulse response h(t)=s(T-t) (0 ≤t <T) is named matched filter to the signal s(t).
y(t)= − +
Binary Orthogonal signals
• Binary Orthogonal signals are two-dimensional signals
• In place of correlators, two linear time invariant filter is applied

r(t)=sm(t)+n(t) 0≤t <Tb m=1,2

• The impulse response of two filters matched to ψ1(t) and ψ2(t) are defined as:

h1(t)=ψ1(Tb-t) 0 ≤ t <Tb and h2(t)=ψ2(Tb-t) 0 ≤ t <Tb

• The received signal is passed through two filters. Output of filters are:
= ℎ − = 1,2
• Sample ym(t) at t=Tb :

= = ℎ − = m=1,2

The output obtained from matched filter is the same as the output from cross-correlator.
Properties of Matched filter
Received signal: r(t)=s(t) +n(t)
n(t): AWGN with zero mean and a power spectral density Sn(f)=N0/2 W/Hz

• r(t) is passed through a filter with the impulse response h(t), 0 ≤ t <T

• The filter output:

= ℎ −

= ℎ − + ℎ −

• Samples of y(t) at time t=T

= ( )ℎ − + ( )ℎ − = + !

ys(T): Signal component

yn(T): noise component

We select a filter so as to maximise the output Signal to Noise Ratio (SNR)

" ( )
=
# $ %( ! ( )
 % ! = % ℎ − ℎ −
&'
= ( − ℎ − ℎ −
&'
= ℎ −

• Variance depends on the power spectral density of noise and energy in impulse response h(t)

" ℎ − ℎ −
= =
# # #
$
2 ℎ − 2 ℎ −

• (S/N)o is maximised when the denominator is constant and nominator is in a maximum value

How to maximise nominator?

+ + +
Apply Cauchy-Schwartz inequality: )
,+ *
) ≤ )
,+ * ,+
)

• Equality holds when )* = =) , where = is a constant.

• Take )* = ℎ and ) (t)=s(T-t). (S/N)o is maximised when h(t)=Cs(T-t). This means h(t) is
matched to the signal s(t).
 " 2 2%
= =
# $
# #$
Es: Energy of the signal s(t).

• Frequency domain implementation of matched filter

We have h(t)=s(T-t)
• Fourier transform of h(t):

− C ,D EF = ( )C D EFG C ,D EF = " ∗ (B)C ,D EF

A B =

Conclusion:
1. |H(f)|=|S(f)|
2. Phase of H(f) is the negative of phase of S(f).
3. The output of waveform:
+ +
= I B CD EF
B= "(B) C ,D EF
CD EF
B
,+ ,+
By sampling ys(t) at time t=T:
+
= "(B) B= =%
,+
4. The noise of the output of matched filter has a power spectral density of:
#
" B = A(B)
2
5. The total noise power at the output of the matched filter:
+ + +
# # %#
J! = " B B= A(B) B= "(B) B=
,+ ,+ 2 2 ,+ 2

The signal power: J = ( )

K LM OMP OM
Then, & $
= LN
= Q = &'
OM '
P
Performance of optimum detector:

Binary antipodal signals

• The output of demodulator is: y=sm+n =± %
• The input to the detector is scalar. This means, the detector compares y with a threshold α.
• If y>α: signal s1(t) was transmitted
• If y<α: signal s2(t) was transmitted

The average probability of error (in terms of α):

T +
J S = J( * ) B * + J( ) B( | )
,+ T
• P(s1) and P(s2) are the a-priori probabilities of two possible transmitted signals.

What value of α minimizes the average probability of error?

VLP (T)
• We need to find VT
=0.
F(T| W) L( P)
• P(s1)f(α|s1)-P(s2)f(α|s2)=0. Hence, =
F(T| P ) L( W)
* P
,(X, Y ) /&'
• We have B( | )= C m=1,2
E&'
 P P
T, O [ T\ O [ J(" )
C &' C &' =
J("* )
Then,
]T O /&'
J(" ) J(" ) #
C = → S= `
J("* ) 4 % J("* )

• If P(s1) > P(S2) : α <0

• If P(s1) < P(S2) : α >0
• If P(s1)=P(s2), then α=0

Average probability of error:

P
1 1 1 X, O [
J = B * + B * = B * = C &'
2 ,+ 2 ,+ ,+ a# ,+

Note:
&'
Let x=(y- % )/ b 1 ,c
1 +
,iP / P/
h d = C j= C ,i j
*
Pe
bQ ,c P / O 2a ,+ 2a c
J = E ,+
'C d= Q( &'
)
Binary Orthogonal signals
• y1 = % + * , =
• n1 and n2: statistically independent with zero
mean Gaussian variables and k! = &'b
• Probability of error depends on the difference
between y1 and y2
• If y1-y2<0, transmission of s1(t) is decided.
• If y1-y2>0, transmission of s2(t) is decided.
• Let z=y1-y2= % + * −
• z is Gaussian random variable with mean % and variance kl = #
* P
Probability density function of z: B(m)= C ,(l, O ) / &'
E&'
The average probability of error :
O
, &
1 ' P %
J =J m<0 = B m m= C ,c ⁄ d=h
,+ 2a ,+ #
Remember:
O
• For binary antipodal signals: J = h
&'
• For the same probability of error, antipodal signals require 3 dB less signal energy than
orthogonal signals.
Expression of probability of error
+
• Based on Euclidean distance between two signals: * = ,+ * − ( )
For antipodal signals:
• Signals are separated by distance d12: * = 2 %
P
qWP Pb qWP
Hence, % = ] and J = h .
&'

qWP Pb
For orthogonal signals: * = 2% , % =
M-Ary digital modulation
• Binary information sequence is subdivided into blocks of k bits (symbols)
• Each block (symbol) is represented by one of M=2k signals (each of duration T)
• Signalling rate Rs=1/T
• Each signal carries k=log2M bits of information.
• Bit rate is Rb=kRs=k/T
• Bit interval: Tb=1/Rb=T/k
Note:
• M-ary signals may be one dimensional or multidimensional.
Optimum receiver for M-ary signals
• Each of M=2k symbols is expressed by: sm(t), m=1,2,…,M and 0 ≤ t <T
• Received signal (after passing through the channel) is given by r(t)=sm(t)+n(t)
• n(t): AWGN with power spectral density Sn(f)=N0/2 W/Hz
Signal demodulator:
• Converts the received waveform r(t) into an N-dimensional vector y=(y1,y2,…,yN)
Signal detector:
• Decides which of M possible signals were transmitted.
• M-ary signal waveforms: = ∑& st* s s ( ) 0 ≤t <T m=1,2,…,M
Note:
sm=(sm1,sm2,…,smN)
s : N orthogonal basis waveforms that span the N-dimensional space.
Correlation-type demodulator Matched filter type demodulator
Correlation-type demodulator
• The correlator outputs at the end of the signal interval are:

s = + s

yk=smk+nk k=1,2,…,N

=
s s and s = s
sm: depends on which of M signals was transmitted.

& & &

=u +u + v( )=u + v( )
s s s s s s
st* st* st*

v = − ∑&st* s s (n’(t): zero mean Gaussian noise)

Mean value of noise components:

% s = % s =0
Covariance of noise components :
&'
% s = % s = ( − s
&' &'
= s = ( s
PDF of noise components
P
N•
* , ∑Q
B(n)=∏&
~t* B ~ = Q/P C •€WQ' (*)
E&'

E( s ) = % s+ s = s
kX = k! = # ⁄2

• Noise components {nk} are uncorrelated Gaussian random variables (statistically independent).
• The correlator outputs ({yk}) are also statistically independent
Conditional probability density function of y=(y1,y2,…,yN)
&

B •‚ = ƒ B( s| s)
st*
* X„ , Y„ P /&'
B s s = C, k=1,2,…,N (**)
E&'

From (*) and (**):

&
1
B •‚ = Cd… − u s − s /#
a# &/
st*
*
= E&' Q/P
exp − • − ‚ /# = 1,2, … , ‡
 Is n’(t) uncorrelated with N correlator outputs ({yk})?
v v v v
% s =% s +% s =% s
=% − ∑&
Dt* D D( ) s
&' &'
= % s − ∑&
Dt* %( D s) D( ) = s − s =0
We know:
# #
% s = ( − s = s( )
2 2

As discrete signals (similar to the continuous one represented by the above integral:
&
#
u% D s D = s
2
Dt*

Conclusions:
1. For detection of one of yks, no additional relevant information is extracted from n’(t).
2. n‘(t) and rk (or {yk}) are uncorrelated (reason: n’(t) is zero mean)
3. n’(t) does not contain any information that is relevant to the decision as to which signal
waveform was transmitted.
4. n’(t) can be ignored.