0% found this document useful (0 votes)
59 views19 pages

2 - Detection of Digital Signals

This document discusses detecting digital signals in noise using optimum decision thresholds. It describes how to calculate the net error probability and optimum threshold voltage for binary signals affected by noise. The optimum threshold depends on the signal and noise probability density functions. It provides the formula to calculate the threshold for unipolar signals in Gaussian noise. It also discusses finding the error probability in terms of signal-to-noise ratio and extending the analysis to ternary signals.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
59 views19 pages

2 - Detection of Digital Signals

This document discusses detecting digital signals in noise using optimum decision thresholds. It describes how to calculate the net error probability and optimum threshold voltage for binary signals affected by noise. The optimum threshold depends on the signal and noise probability density functions. It provides the formula to calculate the threshold for unipolar signals in Gaussian noise. It also discusses finding the error probability in terms of signal-to-noise ratio and extending the analysis to ternary signals.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 19

1

Detection of Digital Signals in Noise


Net Error Probability and Optimum Decision Threshold
Binary Signals:
Let x(t) be a binary signal having two waveforms shapes(one for logic"0"
and the other for logic"1"). n(t) is the noise component added due to
channel such that the PDF of n(t) is f(n) with zero mean.

y(Tb) is the detector output at t=Tb( bit duration for both "0" and "1"
signals ). A comparator with threshold voltage Vth is used to decide if the

y(Tb)
x(t)+n(t) Detector
(binary) + data

Vth

received x(t) is "0" or "1" according to the decision rules:

if y(Tb)>Vth then decide the received x(t) is a "1".


if y(Tb)<Vth then decide the received x(t) is a "0".

The problem is to find the best choice of Vth such that the overall net error
probability(average error prob) of "1" & "0" is minimum. This optimum
Vth depends on type of the binary signals and the PDF of n(t).

Unipolar binary signals: First we discuss the simplest case if x(t)=0 volt
for logic"0" and x(t)=+A volt for logic"1" over the bit duration Tb.
2

+A
x(t) 0 0
1 1
0
Tb
p(1T)
p(0T)

Note that for noiseless case (y(t)=x(t)) then the PDF of y(t) is discrete
similar to the PDF of x(t).
When n(t)≠0, the PDF of y(t) will be the same as the PDF of n(t) (f(n)) but
d.c. shifted by 0 volt or +A volts corresponding to x(t). Let f1(y) be the
PDF of y when "1" is transmitted and fo(y) be the PDF of y when "0" is
transmitted.

According to previous decision rules, then:


Vth 
p(0 R / 1T )   f 1( y)dy

and p(1R / 0T )   fo( y)dy
Vth
These two probabilities are those previously given for the binary channel
model.
3

0T 0R
p(0R/1T)

p(1R/0T)
1T
1R

The net error prob will be:


PNet= p(0R,1T) + p(1R,0T)=p(1T)p(0R/1T) + p(0T)p(1R/0T)
Vth 
pNet  p(1T )  f 1( y )dy  p(0T )  fo( y )dy
 Vth
To minimize PNet differentiate PNet with respect to Vth to get:
dPNet
 0  p(1T ) f 1(Vth)  p(0T ) fo(Vth) or:
dVth
p(0T ) f 1(Vth)
 This is a general formula used to find Vth if f(n) and
p(1T ) fo(Vth)
input prob of x(t) are known.

Take for example the case when n(t) is Gaussian, then:


n2
1 
f ( n)  e 2 2
with =0 and 2 =variance of n(t). Using previous
2 
equation, then:
(Vth A ) 2
1 
2 2
e
p (0T ) f 1(Vth) 2 
q  V 2th
p (1T ) fo(Vth) 1 
2 2
e
2 
This gives:
ln q  ( A 2 / 2 2 )
Vth 
(A / 2 ) This is used only for unipolar x(t) and n(t) is
Gaussian.
Special and practical case: if q=1 (p(0T)=p(1T)), then: Vth=A/2 and
PNet= pe=p(0R/1T)=p(1R/0T)=area under Gaussian curve calculated either
from fo(y) or f1(y).
4

  y2
1 
pe   fo( y)dy  
Vth A/ 2 2 
e 2 2
dy . Note that this definite integral can

not be evaluated analytically. The following substitution will help to


evaluate this integral numerically:
Let z=y/, then dy=dz.
For y=A/2, then z=A/(2), and for y, then z. Changing the
variable and the limits of this integral reduces it into:
5
 z2
1  A
pe  A 2
e 2
dz  Q(
2
)
where Q(x) is a function called
2
Marcum function which gives the area under the standard Gaussian
curve(=1, =0) from x=(A/2) to .

 x

Note that: Q( x)   p( z )dz   p( z)dz , Q(0)=0.5. This Q(x) is


x 
numerically evaluated and tabulated for x<3, or can be approximated for
large x as:
1 0.7 0.5 x 2
Q(x) (1  )e for x>2
2 x x2
Ex: Equiprobable binary signals are represented by 0 & 2 volts. For
AWGN with 2=0.16 volts2, find the optimum threshold setting and the
net error prob.
Solution: For equiprobable special case with unipolar signals affected by
AWGN, then Vth=A/2=2/2=1 volt, and pe=Q(A/2)=Q(2/2*0.4)=Q(2.5)
( 2=0.16, then =0.4)., then:
1 0.7 0.5*6.25
Pe=Q(2.5)  2 2.5 )e (1  6.2*10 -3
6.25
Ex: Repeat previous example if p(0T)=0.25.
Solution:
Note that this is not the equiprobable special case, then for AWGN, then:
ln q  ( A 2 / 2 2 )
Vth 
(A / 2 ) where q=p(0T)/p(1T)=0.25/0.75=1/3. Putting
q=1/3, A=2, and =0.4, get Vth=0.912 volt and:
6


Vth 0.912
p(1R / 0T )  
Vth
fo( y )dy  Q(

)  Q(
0.4
)  Q(2.28)  0.01125

and
A  Vth 2  0.912
Vth
p(0 R / 1T )  

f 1( y)dy  Q(

)  Q(
0.4
)  Q(2.72)  3.28 *103

and PNet= p(1T)p(0R/1T) + p(0T)p(1R/0T)


=0.75*3.28*10-3+0.25*0.01125=5.2*10-3

Homework: Show that for bipolar signals A over AWGN channel with
variance 2 and q=p(0T)/p(1T), then:

 2 ln q
Vth | optimum  and if q=1, then Vth=0, and pe=Q(A/)
2A

Note: In general, for equiprobable signals and AWGN, then:


7

Vth|op=(A1+A2)/2 and p(0R/1T)=p(1R/0T)=pe=Q[d/(2)], where d=A2-A1

Ex: Nongaussian noise case: if n(t) is not Gaussian, say f(n)=0.75(1-n2),


|n|<1. If this noise affects the unipolar signal x(t): 0 volt & +1 volt, and if

p(0T)=p(1T)=0.5, then Vth=(1+0)/2=0.5 volts(using previous note for


equiprobable binary signals). To find the error prob:

1 0.5
p(1R / 0T )   0.75(1  y )dy  0.156  p(0 R / 1T )   0.75[1  ( y  1) 2 ]dy
2

0.5 0

Note that p(1R/0T)=p(0R/1T) since x(t) has equiprobable signals.


Note: If n(t) is not Gaussian remember that there is no answer in terms
of Q(x) function

Homework: Repeat previous example if p(0T)=0.25.


p(0T ) f 1(Vth)
(Hint: use the equation: p(1 )  fo(Vth) )
T
8

Finding the error prob. in terms of SNR:


Sometimes, it is required to find the error prob in terms of SNR power
ratio( S is the average signal power and N is the noise power). We will
find this for:
1-unipolar equiprobable case affected by AWGN in which we found that:
Pe=Q[A/(2)]. Here S=p(1T)*A2 +p(0T)*(0)2=0.5A2, where the signal
power is normalized to 1ohm resistor and p(1T)=p(0T)=0.5.
Also N=2 and again the noise power is also normalized to 1ohm resistor.
2S SNR
Hence: pe  Q( )  Q( )
2 N 2
2- bipolar equiprobable case affected by AWGN, then pe=Q(A/), where:
S=0.5*A2+0.5(-A)2=A2 and N=2, then pe  Q( SNR)

Ternary signals: Suppose that the ternary voltage levels –A, 0, +A volts
are used to represent logic states 0,1,2 respectively. The decision rules will
be:

If y<Vth2 then decide on "0" state.


If Vth1>y>Vthe2 then decide on "1" state.
If y>Vth1 then decide on "2" state.
To find Vth1 and Vth2, first we assume that the extreme voltages –A and
+A are not interfered and these is no error between state "0" & "2", then:
p (1 ) f 2(Vth1)
Vth1 can be found from p (2 )  f 1(Vth1)
T

p (0 ) f 1(Vth 2)
Vth2 can be found from p (1 )  fo(Vth 2)
T

T
9

Special case: if the ternary signals are equiprobable, the: Vth1=A/2,


Vth2=-A/2, pe=Q[A/(2)]
Note: This case of ternary signaling was already given as a nonsymmetric
and practical channel shown, where x1=0T, x2=1T, x3=2T and y1=0R, y2=1R,
y3=2R.

1-pe

X1 Y1
X2 pe
e
pe
1-2pe
X2 Y2
X3
pe

pe

X3 Y3
1-pe

Ex: Equiprobable ternary signals -2A, 0, +3A volts are affected by zero
mean AWGN with variance 2=0.16A2. Find optimum threshold settings
and the net error prob.
Solution: Since equiprobable, then: Vth1=[0-2A]/2=-A volts,
Vth2=[0+3A]/2=1.5A. Also =0.4A, then:

pe1=p(0R/1T)=p(1R/0T)=Q[d1/(2)]=Q[2A/(2*0.4*A)]=Q(2.5)=6.2*10-3
pe2=p(2R/1T)=p(1R/2T)=Q[d2/(2)]=Q[3A/(2*0.4*A)]=Q(3.75)=8.9*10-5
and the net error prob=(1/3)*pe1+(1/3)*pe1+(1/3)*pe2+(1/3)*pe2
=(2/3)*pe1+(2/3)*pe24.19*10-3
11

Homework: Repeat previous example if p(0T)=p(1T)=0.125.


(Hint: Vth1 as before is (-2A+0)/2=-A since "0" and "1" are equiprobable,
but Vth2 must be found from the formula for unipolar case under AWGN
with A being set to +3A, i.e.:
ln q  (9 A2 / 2 2 )
Vth2  with q=p(1T)/p(2T).
(3 A /  2 )

Block Error Probability:


If pe is the error prob of a channel (binary symmetric), and a block of m
bits is transmitted, then the prob that a block is received with n errors,
(m  n  0), is given by:
p(n erroneousbits in a block of m bits)  Cnm pen (1  pe)mn
m!
Where C n 
m

n! (m  n)! which gives the number of ways that a block of


m bits is received with n errors.
m

errors

Note also that the total prob is: 


C i
m
pei (1  pe) m  i  1
i 0
Ex: Bipolar equiprobable signals are transmitted through AWGN channel
with SNR of 6dB in blocks of 5bits. Find:[1]-prob that a block of 2 errors
is received. [2]-prob that a block with errors at 1st and last positions is
received.
Solution: Fisrt, we find pe from SNR. We recall that for bipolar equiprob
6

case under AWGN, then : pe  Q SNR  Q 10  Q (2)  0.02


10

5*4
[1]p(a block with 2 errors)= C25 pe2 (1  pe) 3  * (0.02) 2 (0.98) 3  0.0037
2
11

[2]p(2 errors at certain locations)=p2e (1-pe)3=(0.02)2(0.98)3=0.00037.


5
Note that the term C 2 is neglected here since the locations of the two
errors are specified.

Matched filter
Once we studied how to choose the best threshold to minimize the net
error prob, next, we consider the detection problem, i.e. the detection of
the signal x(t) embedded in AWGN noise n(t) (only AWGN case is
considered here). The output of this detector is y(Tb) which was the signal
used at the decision block. The device that carries out the function of
detection is called a Matched Filter . This matched filter is simply a linear
system with impulse response h(t) (or frequency response H()) that acts
| y (Tb ) | 2
as a filter. The job of this filter is to maximize the ratio n 2 (t ) . This
o

ratio is in fact the ratio of the signal power to noise power at the output of
this filter at the instants of sampling at Tb(bit duration). The input

Matched filter
x(t)+ni(t) H() y(t)+no(t)
h(t)

is the signal x(t) embedded in the noise ni(t) which is assumed to be


AWGN. This filter is called Matched filter since for a certain signal x(t),
there exists a filter with impulse response h(t) matched to it(maximizes the
above ratio).
To find the relation between x(t) and h(t), the following derivation is
given:
We know that: Y()=X() H(), then:

1
y(t )  
2 
X ( ) H ( )e jwt d and at t=Tb, then:

1
y (Tb)   X ( ) H ( )e d
jwTb

2 

1
And no2(t)=o/p noise power= 2  Gn ( ) | H ( ) | d where Gn() is the
2


input noise power spectral density which is flat for AWGN and is equal to
o/2 (two-sided spectrum), then:
12

1
no (t)= 2  (o / 2) | H ( ) | d
2
2

The ratio to be maximized will be:

1
| y (Tb ) | 2
| 
2 
H ( ) X ( )e jwTbd |2
 
n o2 (t ) 1 then using Shwartz inequality
2 
( o / 2) | H ( ) | 2
d
  

 f ( x) f ( x)dx |  (  | f1 ( x) | dx ) (  | f 2 ( x) | 2 dx ) where
2 2
that states that: | 1 2
  
*
equal sign holds if f1(x)=k f2 (x) (*=conjugate), k is any constant that
may represent the gain, say k=1.
Now if f1(x)=|H()|, f2(x)=X()ejwTb, then applying this inequality on the
numerator of above ratio get:
 
1
(  | H ( ) | d ) (  | X ( )e jwTb |2 d )
2

| y (Tb ) |2 4
2
 


n o2 (t ) (o / 2)
2 
| H ( ) | 2
d


| y (Tb ) |2 1
   d
2
| X ( ) |
o  
jwTb 2
And since |e | =1, then: n 2 (t ) and since:
o

 | y (Tb ) |2 E

| X () | d  E =energy of x(t), then: n 2o (t ) o or:
2

| y (Tb ) |2 E
[ 2 ] max 
n o (t ) o note that both E and o have the same units
Watts/(rad/sec). But if expressed in Watt/Hz, then:
| y (Tb ) |2 E
[ 2 ] max 
n o (t ) (o / 2) =Max SNR at matched filter output.
o=one-sided AWGN spectral density(Watt/Hz).
To find the condition for maximization, recall that we used equal sign in
Schwartz inequality if f1(x)=f*2(x) , or H()=(X()ejwTb)*=X*()e-jwTb, or
in time domain:
h(t)=x(Tb-t).
13

Hence the impulse response of a matched filter matched to the signal x(t)
is the negative time of x(t) shifted by Tb.
Ex: Find the impulse response and the max SNR at the o/p of a matched
filter used to detect the signal so(t) shown embedded in AWGN with one-
sided spectral density of o.
h(t) so(t)

V V

Tb t Tb t
t

Solution: so(t)=(V/Tb) t for Tb>t>0


Then h(t)=so(Tb-t)=(V/Tb)(Tb-t)=V-(V/Tb)t for Tb>t>0.
And the max SNR =E/(o/2), where E can also be found from the time
Tb Tb
V2 2 V 2Tb
domain as: E   s o (t )dt   Tb 2 t dt  3 , then:
2

0 0

2V 2Tb
(SNR)max= 3 .
o
Practical Implementation of the Matched filter:
We know that h(t)=x(Tb-t), and since y(t)=h(t)x(t) (=convolution)
t Tb
y (t )   x( )h(t   )d and at t=Tb y (Tb )   x( )h(Tb   )d
0 0
Tb Tb
         (t )dt =Energy of x(t) over Tb
2
Or y (Tb ) x ( ) x (Tb Tb ) d x
0 0

Above equation indicates that if the i/p x(t) is noise free then y(Tb) is E,
but in general if the i/p is x(t)+n(t), then y(Tb) will be a Gaussian random
variable (n(t) is Gaussian) with mean E and variance that is linearly
proportional to o. The practical implementation can be deduced as the
product of the input x(t)+n(t) and x(t) then integrate the result over t=0 to
t=Tb. The figure below shows the practical implementation of the matched
14

filter.

Tb t=Tb

x(t)+n(t)
 ( )dt
0 Y(Tb)

x(t)

Error probability of binary signal detection using matched filter:


Assume that the matched filter is used to detect the binary signal
waveforms so(t) and s1(t) embedded in AWGN. Here, we need two
matched filters, one matched to so(t) with impulse response ho(t)=so(Tb-t)
and the other matched to s1(t) with impulse response h1(t)=s1(Tb-t)

Matched
filter yo(Tb)
ho(t)

s(t)+n(t)
data
Matched
filter y1(Tb)
h1(t)

Tb

Note that E o   s o (t )dt =mean of yo(Tb)=energy of so(t)


2

=o/p of ho(t) due to so(t).


Also:
Tb
E1   s12 (t )dt =mean of y1(Tb)=energy of s1(t)
0

=o/p of h1(t) due to s1(t).


Tb

And: Eo1   so (t ) s1 (t )dt =cross energy between so(t) and s1(t)


0

=o/p of ho(t) due to s1(t)


= o/p of h1(t) due to so(t)
15

Recall that for equiprobable case p(0T)=p(1T)=0.5, and for Gaussian case,
then:
Tb

 [s o (t )  s1 (t )]2 dt
Eo  E1  2 E01
pe  Q( 0
) = Q( )
2 o 2 o
Ex: Matched filter is used to detect the two equiprobable signals shown in
AWGN. Find the error prob in terms of average signal power, bit rate, and
noise spectral density
o.
s1(t)
so(t) Tb
V
t

-V
Tb t

Solution: s1(t)=(V/Tb) t, so(t)=-(V/Tb) t over bit duration 0 to Tb.


(V ) 2 2
Tb
Eo  E1  2 E01 V2
pe= Q( ) where : E o   2 t dt  Tb
2 o 0 T b 3
Tb
(V ) 2 2 V2
and E1   T 2b t dt  3 Tb
0

(V )(V ) 2
Tb
V2
Eo1   2
t dt   Tb
0
T b 3

And S=average signal power=[0.5Eo+0.5E1]/Tb=V2/3


(V 2 / 3)Tb  (V 2 / 3)Tb  (2V 2 / 3)Tb
pe  Q{ } , if Rb=bit rate=1/Tb
2 o
2V 2 2S
pe  Q{ }  Q( )
3 o Rb  o Rb

Homework: Repeat previous example for unipolar 0,+A


signals(orthogonal signals) and bipolar A signals(antipodal signals).
S 2S
(Ans: for unipolar pe  Q(  o Rb
) , for bipolar pe  Q(
 o Rb
)
16

Error prob for digital carrier systems:

[A]-coherent detection with matched filter:


1- ASK (OOK) (ON-OFF) keying: Here the two signals are so(t)=0,
s1(t)=Acost over bit duration Tb. We use the general equation:
Eo  E1  2 E01
pe= Q( 2 o
) , where for OOK, Eo=E01=0, and:

E1=(A2Tb)/2 (s1(t) is a sinusoid whose normalized power=A2/2), then:


A 2Tb
pe  Q( )
4 o and to express in terms of average signal power, then:
S=[0.5Eo+0.5E1]/Tb=A2/4, then:
S S E
pe  Q( ) . For simplicity, let:    b  energy/ bit ,
 o Rb  o Rb  o
then: pe  Q(  ) .

2-BPSK signals(Binary Phase Shift Keying): Here the two signals are
so(t)=Acost, s1(t)=-Acost over bit duration Tb. Again, using the same
general equation, then:
Eo=E1=[A2Tb]/2 since both so(t) and s1(t) are two sinusoids, also:
Tb

E01=   A cos  t dt =-[A2Tb]/2.


2 2

0
Also S=[0.5Eo+0.5E1]/Tb=A2/2=average signal power, then:

[0.5 A2Tb  0.5 A2Tb  A2Tb A2Tb


pe  Q ( )  Q( )
2 o o
2S
pe  Q( )  Q( 2 ) which is better than OOK.
o Rb
3-FSK (frequency shift keying): Here the two signals are so(t)=Acos1t,
s1(t)=Acos2t over bit duration Tb. Note that pe depends also on
=2-1
17

f1 fc f2
fd=f

We write so(t)=A cos[(c-0.5d) t] and s1(t)=A cos[(c+0.5d) t], then:


Eo=E1=[A2Tb]/2, since the deviation d does not affect the power of a
sinusoid. In fact, the effect of d appears on the cross energy Eo1given by:
Tb
E01   A 2 cos[( c  0.5 d )t ] cos[( c  0.5 d )t ]dt
0

A2 sin d Tb
Tb
A2
2 0
E01  [cos  t  cos 2 t ] dt 
d
d c
2
A 2Tb sin  d Tb
Or E o1  2  d Tb which is a sinc function as shown:

This has nulls at dTb=, 2, 3,……and [Eo1]min=-0.217(A2Tb/2) at


dTb=4.49rad or at fdTb=0.715. Next to find pe, then:
sin d Tb
[0.5 A2Tb  0.5 A2Tb  A2Tb ( )]
d Tb
pe  Q( )
2 o
18

sin d Tb sin d Tb
[ A2Tb (1  ( )] [ S (1  ( )]
d Tb d Tb
pe  Q( )  Q( )
2 o  o Rb
S
Where S=[0.5Eo+0.5E1]/Tb=0.5A , and if    Rb , then:
2
o

sin  d Tb
pe  Q(  (1  ) which is a general formula used for matched
 d Tb
filter detection of FSK signals.

Special cases in FSK:


1-if dTb=, 2, 3,.., or fd=0.5Rb, Rb, 1.5Rb,... this gives Eo1=0
(orthogonal FSK), and pe  Q(  ) .
sin  d Tb
2-if dTb=4.49rad, or fd=0.715Rb, then:  Tb  0.217
d

And this gives: pe  Q( 1.217 ) which is the best


(optimum)performance of FSK.

Noncoherent detection:
Previous equations for error prob of digital carrier systems using matched
filter under AWGN can not be used for noncoherent detection(not
matched filter). In fact, exact derivation for noncoherent case is
mathematically complicated. Approximate results are given without
derivations:
1-noncoherent ASK(OOK):
pe=0.5[e-/2+Q()]  0.5 e-/2 for >>1, since here e-/2 >>Q()
2-noncoherent FSK:
pe  0.5 e-/2 for >>1(the same as OOK)
3-noncoherent PSK=DPSK:
pe  0.5 e- for >>1 (better than OOK or FSK)
19

Ex: Matched filter detection is used to detect BPSK signals at a rate of


600bps. If transmitted power is 5KW over an HF channel having
estimated path losses of 150dB, find the error prob if the noise at detector
input has one sided spectral density of 10-15 Watt/Hz.
Solution:
PT=trans power=5000 W, then, PR=S=PT*10-0.1*losses in dB=5000*10-0.1*150
S=5*10-12 W=average signal power.
S 5 *1012
  15  8.333 , then
 o Rb 10 * 600
pe  Q( 2 )  Q( 2 * 8.3333)  Q(4.08)  2 *105
Ex: Pepeat previous example for FSK signals with f1=700Hz, f2=2000Hz.
Solution:
Here d=2(2000-700)=2600, and d Tb=2600/600=4.3333, then and
for the same =8.333:

sin d Tb sin 4.333


pe  Q(  (1  )  Q( 8.333(1  )  Q(2.84)  2.3 * 103
d Tb 4.333

Which is worst than BPSK for the same .( remember that both BPSK and
FSK have the same average power of A2 /2, but the performance of BPSK
is better at the expense of the requirement of carrier recovery at the
receiver).

You might also like