Chapter 6

DETECTION AND ESTIMATION:

Fundamental issues in digital communications are

1. Detection and
2. Estimation

Detection theory: It deals with the design and evaluation of decision making
processor that observes the received signal and guesses which particular symbol was
transmitted according to some set of rules.

Estimation Theory: It deals with the design and evaluation of a processor that uses
information in the received signal to extract estimates of physical parameters or
waveforms of interest.
The results of detection and estimation are always subject to errors

Model of digital communication system

Consider a source that emits one symbol every T seconds, with the symbols belonging to
an alphabet of M symbols which we denote as m1, m2, . . . . . . mM.
We assume that all M symbols of the alphabet are equally likely. Then
pi p (mi emitted )
1
for all i
M
The output of the message source is presented to a vector transmitter producing vector of
real number
S i1
S
i2 Where the dimension N M.
.
S i i 1 , 2 , ..... , M
.
.

S iN

The modulator then constructs a distinct signal si(t) of duration T seconds. The signal
si(t) is necessarily of finite energy.

The Channel is assumed to have two characteristics:

Channel is linear, with a bandwidth that is large enough to accommodate
the transmission of the modulator output si(t) without distortion.
The transmitted signal si(t) is perturbed by an additive, zero-mean,
stationary, white, Gaussian noise process.
such a channel is referred as AWGN ( additive white Gaussian noise ) channel

GRAM SCHMIDT ORTHOGONALIZATION PROCEDURE:

In case of Gram-Schmidt Orthogonalization procedure, any set of M energy signals

{Si(t)} can be represented by a linear combination of N orthonormal basis functions
where NM. That is we may represent the given set of real valued energy signals S 1(t),
S2(t). . . . . . . SM(t) each of duration T seconds in the form

S1 (t ) S111 (t ) S12 2 (t ) . . . . . . . . S1N N (t )

S 2 (t ) S 211 (t ) S 222 (t ) . . . . . . . . S 2 N N (t )
.
.
.
.
S M (t ) S M 11 (t ) S M 22 (t ) . . . . . . . . S MNN (t )

N
0 t T
S i (t ) S ij j ( t )
j 1 i 1, 2 , 3 . . . . . . M ( 6 . 1 )
 Where the Co-efficient of expansion are defined by

T
i 1, 2,3 . . . . . M
S ij (t ) S i (t ) j (t )dt
0 j 1, 2,3 . . . . . . N ( 6 .2 )

The basic functions 1 (t ), 2 (t ) ......... N (t ) are orthonormal by which

T
1 if i j
0 i (t ) j (t ) dt 0 if i j (6.3)
The co-efficient Sij may be viewed as the jth element of the N dimensional
Vector Si
S i1
S
i 2

'

Therefore S i ' i = 1,2,3 . . . . . . M
'

'
S
iN

Let S1 31 (t ) 42 (t ) S 2 1 (t ) 22 (t )

Vector
3 1
S1 S2
4 2
Geometric interpretation of signal:

Using N orthonormal basis functions we can represent M signals as

N
S i (t ) S ij j (t ) 0 t T i 1, 2 , ..... , M (6.4)
j 1
Coefficients are given by

T
S ij S i (t ) j (t ) dt i 1, 2 , ..... , M
0

j 1, 2 , ..... , N (6.5)

Given the set of coefficients {sij}, j= 1, 2, .N operating as input we may use the scheme
as shown in fig(a) to generate the signal si(t) i = 1 to M. It consists of a bank of N
multipliers, with each multiplier supplied with its own basic function, followed by a
summer.

fig(a)

conversely given a set of signals si(t) i = 1 to M operating as input we may use the
scheme shown in fig (b) to calculate the set of coefficients {sij}, j= 1, 2, .N
 fig(b)

Si1 The vector si is called signal vector

S
i2
.
S i i 1, 2 , ..... , M
.
.

SiN
We may visualize signal vectors as a set of M points in an N dimensional Euclidean
space, which is also called signal space
The squared-length of any vector si is given by inner product or the dot product of si
N
Si ( Si , Si ) Sij2
2

j 1

Where sij are the elements of si

Two vectors are orthogonal if their inner product is zero

The energy of the signal is given by

T
Ei Si2 (t ) dt
0
substituting the value si(t) from equation 6.1
T N N
Ei [ Sij j (t )] [ Sikk (t )] dt
0 j 1 k 1

interchanging the order of summation and integration

N N T
Ei S ij S ik j (t )k (t )dt
j 1 k 1 0
since j (t ) forms an orthonormal set, the above equation reduce to

N
Ei Sij2
j 1
this shows that the energy of the signal si(t) is equal to the squared-length of the signal
vector si
The Euclidean distance between the points represented by the signal vectors si and sk is
N
Si S k ( Sij S kj ) 2
2

j 1
T
[ Si (t ) S k (t )]2 dt
0

Response of bank of correlators to noisy input

Received Signal X(t) is given by

X (t ) S i (t ) W (t ) 0t T
i 1,2,3. . . . . . ., M (6.6)

where W(t) is AWGN with Zero Mean and PSD N0/2

Output of each correlator is a random variable defined by
T
X j X (t ) j (t )dt
o

S ij W j j 1,2,........N (6.7)

The first Component Sij is deterministic quantity contributed by the transmitted signal
Si(t), it is defined by
T
S ij S i (t ) j (t )dt (6.8)
0

The second Component Wj is a random variable due to the presence of the noise at the
input, it is defined by

T
W j W (t ) j (t )dt (6.9)
o
let X ' (t ) is a new random variable defined as

N
X ' (t ) X (t ) Xj 1
j j (t ) (6.10)

substituting the values of X(t) from 6.6 and Xj from 6.7 we get

N
X ' (t ) Si (t ) W (t ) (S
j 1
ij W j ) j (t )

N
W (t ) W (t )
j 1
j j

W ' (t )

which depends only on noise W(t) at the front end of the receiver and not at all on the
transmitted signal si(t). Thus we may express the received random process as
N
X (t ) X (t ) X ' (t )
j 1
j j

N
X (t ) W ' (t )
j 1
j j

Now we may characterize the set of correlator output, {Xj}, j = 1 to N , since the received
random process X(t) is Gaussian , we deduce that each Xj is a Gaussian random variable.
Hence, each Xj is characterized completely by its mean and variance.

Mean and variance:

The noise process W(t) has zero mean, hence the random variable W j extracted from W(t)
also has zero mean. Thus the mean value of he jth correlator output depends only on Sij as

m x j E[ X j ] from eqn 6.7

E[ S i j W j ]
S ij E[W j ] but E[W j ] 0
S ij

variance of Xj is given by

2 x j Var[ X j ]
E[( X j m x j ) 2 ] substituting m x j S ij
E[( X j S ij ) 2 ] from equton 6.7
E[W j2 ]
 substituting the value of Wj from eqn 6.9
T T

2 x j E W (t ) j (t ) dt W (u ) j (u ) du
0 0
T T
E j (t ) j (u )W (t )W (u ) dt du
0 0

T T
2x j
0 0
j (t ) j (u ) E [ W (t )W (u ) ] dt du

T T

0 0
j (t ) j (u ) Rw (t , u ) dt du (6.11)

where

Rw (t , u ) = E [ W (t )W (u ) ] autocorrelation function of the noise process W(t).Science the

noise is stationary, with psd N0/2 ,Rw(t,u) depends only on the time difference (t-u) and
expressed as

N0
Rw (t , u ) (t u ) (6.12)
2
substituting this value in the equation 6.11 we get
T T
N
xj 0 (t ) (u ) (t u ) dt du
2
j j
2 0 0
T
N0

2 0 j2 (t ) dt

Science the j (t ) have unit energy, the above equation reduce to

N0
2x j for all j
2
This shows that all the correlator outputs {Xj}, j = 1 to N have a variance equal to the psd
No/2 of the additive noise process W(t).
Science the j (t ) forms an orthogonal set, then the Xj are mutually uncorrelated, as
shown by
 Cov[ X j X k ] E[( X j mx j )( X k mx k )]
E[( X j S i j )( X k S ik )]
E[W jWk ]
T T

E W (t ) j (t )dt W (u )k (u )du
0 0
T T
(t ) (u ) R
0 0
j k w (t , u )dt du

T T
N0

2 (t ) (u ) (t u )dt du
0 0
j k

T
N
0
2
0
j (t )k (u )dt

0 jk

Since the Xj are Gaussian random variables, from the above equation it is implied that
they are also statistically independent.

Detection of known signals in noise

Assume that in each time slot of duration T seconds, one of the M possible signals S1(t),
S2(t). . . . . . . SM(t) is transmitted with equal probability of 1/M. Then for an AWGN
channel a possible realization of sample function x(t), of the received random process
X(t) is given by

x(t ) Si (t ) w(t ) 0t T
i 1,2,3,.............., M

where w(t) is sample function of the white Gaussian noise process W(t), with zero mean
and PSD N0/2. The receiver has to observe the signal x(t) and make a best estimate of
the transmitted signal si(t) or equivalently symbol mi
The transmitted signal si(t), i= 1to M , is applied to a bank
of correlators, with a common input and supplied with an appropriate set of N
orthonormal basic functions, the resulting correlator outputs define the signal vector Si.
knowing Si is as good as knowing the transmitted signal Si(t) itself, and vice versa. We
may represents si(t) by a point in a Euclidean space of dimensions N M. . Such a point
is referred as transmitted signal point or message point. The collection of M message
points in the N Euclidean space is called a signal constellation.
When the received signal x(t) is applied to the bank o N correlators , the output of the
correlator define a new vector x called observation vector. this vector x differs from the
signal vector si by a random noise vector w
x Si w i 1,2,3,........., M
The vectors x and w are sampled values of the random vectors X and W respectively. the
noise vector w represents that portion of the noise w(t) which will interfere with the
detected process.
Based on the observation vector x, we represent the received signal s(t)by a point in the
same Euclidean space, we refer this point as received signal point. The relation between
them is as shown in the fig

Fig: Illustrating the effect of noise perturbation on location of the received signal point

In the detection problem , the observation vector x is given, we have to perform a

mapping from x to an estimate of the transmitted symbol, in away that would minimize
the average probability of symbol error in the decision. The maximum likelihood
detector provides solution to this problem.

Optimum transmitter & receiver

Probability of error depends on signal to noise ratio
As the SNR increases the probability of error decreases
An optimum transmitter and receiver is one which maximize the SNR and
minimize the probability of error.
Correlative receiver

Observation
Vector x

For an AWGN channel and for the case when the transmitted signals are equally likely,
the optimum receiver consists of two subsystems
1) .Receiver consists of a bank of M product-integrator or correlators
1(t) ,2(t) .M(t) orthonormal function
The bank of correlator operate on the received signal x(t) to produce observation vector x
2). Implemented in the form of maximum likelihood detector that
operates on observation vector x to produce an estimate of the transmitted symbol
mi i = 1 to M, in a way that would minimize the average probability of symbol error.
The N elements of the observation vector x are first multiplied by the corresponding N
elements of each of the M signal vectors s1, s2 sM , and the resulting products are
successively summed in accumulator to form the corresponding set of
Inner products {(x, sk)} k= 1, 2 ..M. The inner products are corrected for the fact that the
transmitted signal energies may be unequal. Finally,
the largest in the resulting set of numbers is selected and a corresponding decision on the
transmitted message made.
The optimum receiver is commonly referred as a correlation receiver

MATCHED FILTER
Science each of t he orthonormal basic functions are 1(t) ,2(t) .M(t) is assumed
to be zero outside the interval 0 t T . we can design a linear filter with impulse
response hj(t), with the received signal x(t) the fitter output is given by the convolution
integral

y j (t ) x( )h (t )d

j

Suppose the impulse response of the system is

h j (t ) j (T t )

Then the filter output is

y j (t ) x( ) j (T t )d

sampling this output at time t = T, we get

y j (T ) x( ) j ( )d

j(t) is zero outside the interval 0 t T ,we get

T
y j (T ) x( ) j ( )d
0

yj(t) = xj
where xj is the j th correlator output produced by the received signal x(t).

A filter whose impulse response is time-reversed and delayed version of the input signal
j (t ) is said to be matched to j (t ) . correspondingly , the optimum receiver based on
this is referred as the matched filter receiver.

For a matched filter operating in real time to be physically realizable, it must be causal.
For causal system

h j (t ) 0 t 0

causality condition is satisfied provided that the signal j (t ) is zero outside the interval
0t T

Maximization of output SNR in matched filter

Let
x(t) = input signal to the matched filter
h(t) = impulse response of the matched filter
w(t) =white noise with power spectral density No/2
(t ) = known signal
Input to the matched filter is given by
x(t ) (t ) w(t ) 0t T
science the filter is linear , the resulting output y(t) is given by
y (t ) 0 (t ) n(t )
where 0 (t ) and n(t) are produced by the signal and noise components of the input x(t).
 The signal to noise ratio at the output of the matched filter at t = T is
0 (T )
2

( SNR ) 0 (6.13)
E[n 2 (t )]

aim is to find the condition which maximize the SNR

let
0 (t ) 0 ( f )
h (t ) H ( f )
are the Fourier transform pairs, hence the output signal 0 (t ) is given by

0 (t ) H ( f )( f ) exp( j 2ft )df

output at t = T is
2

0 (T ) H ( f )( f ) exp( j 2fT )df

2
(6.14)

For the receiver input noise with psd No/2 the receiver output noise psd is given by
N0
S N( f )
2
H ( f ) (6.15)
2
and the noise power is given by

E[n 2 (t )] S

N ( f )df

N
0 H( f )
2
df (6.16)
2

substituting the values of eqns 6.14 & 6.15 in 6.13 we get

2

H ( f )( f ) exp( j 2fT )df

( SNR ) 0
(6.17)
N0
H( f )
2
df
2
 using Schwarzs inequality
2

X 1 ( f ) X 2 ( f )df X
2 2
X 1 ( f ) df 2 ( f ) df (6.18)

*
Eqn 6.16 is equal when X1(f) = kX2 (f)
let X1(f) = H(f)
& X2(f) = ( f ) exp( j 2fT )
under equality condition

H(f) = K * ( f ) exp( j 2fT ) (6.19)

Thus substituting in 6.16 we get the value

H ( f )( f ) exp( j 2fT )df ( f )

2 2
H ( f ) df df

substituting in eqn 6,17 and simplifying

2
( SNR ) 0 ( f )
2
df
N0

Using Rayleighs energy theorem

(t ) dt ( f ) df E , energy of the signal

2 2

2E
( SNR ) 0, max (6.20)
N0
Under maximum SNR condition, the transfer function is given by ( k=1), eqn 6.19

H opt ( f ) * ( f ) exp( j 2fT )

The impulse response in time domain is given by

hopt (t )

( f ) exp[ j 2f (T t )] exp( j 2ft )df
(T t )
Thus the impulse response is folded and shifted version of the input signal (t )
MATCHED FILTER

(t) = input signal

h(t) = impulse response
W(t) =white noise

The impulse response of the matched filter is time-reversed and delayed version of the
input signal
h (t ) (T t )
For causal system

h j (t ) 0 t 0

Matched filter properties

PROPERTY 1
The spectrum of the output signal of a matched filter with the matched signal as
input is, except for a time delay factor, proportional to the energy spectral density of
the input signal.
let 0 ( f ) denotes the Fourier transform of the filter output 0 (t ) , hence

0 ( f ) H opt ( f ) ( f ) sustituting from 6.19

* ( f ) ( f ) exp( j 2fT )
( f ) exp(2 jfT ) (6.21)
2
 PROPERTY 2

The output signal of a Matched Filter is proportional to a shifted version of the

autocorrelation function of the input signal to which the filter is matched.
The autocorrelation function and energy spectral density of a signal forms the Fourier
transform pair, thus taking inverse Fourier transform for eqn 6.21

0 (t ) R (t T )
At time t = T

0 (T ) R (0) E
where E is energy of the signal

PROPERTY 3

The output Signal to Noise Ratio of a Matched filter depends only on the ratio of the
signal energy to the power spectral density of the white noise at the filter input.

SNR at the output of matched filter is eqn 6.13

0 (T )
2

( SNR ) 0 (6.22)
E[n 2 (t )]

output of matched filter is

0 (t ) H ( f )( f ) exp( j 2ft )df

signal power at t = T

0 (T ) H ( f )( f ) exp( j 2fT )df

2

noise psd at the output of receiver is

N0
S N( f )
2
H( f )
2
 noise power is

N
E[n (t )] S N ( f )df 0 H( f )
2 2
df

2

substituting the values in 6.22 we get

2

H ( f )( f ) exp( j 2fT )df

( SNR ) 0 (6.23)
N0

2
H ( f ) df
2
using Schwarzs inequality
2

X 1 ( f ) X 2 ( f )df X
2 2
X 1 ( f ) df 2 ( f ) df (6.24)

*
Eqn 6.24 is equal when X1(f) = kX2 (f)
let X1(f) = H(f)
& X2(f) = ( f ) exp( j 2fT )
under equality condition

H(f) = K * ( f ) exp( j 2fT ) (6.25)

Thus substituting in 6.24 we get the value

H ( f )( f ) exp( j 2fT )df ( f )

2 2
H ( f ) df df

substituting in eqn 6,23 and simplifying

2
( SNR ) 0 ( f )
2
df
N0

Using Rayleighs energy theorem

(t ) dt ( f ) df E , energy of the signal

2 2

2E
( SNR ) 0,max (6.26)
N0
 PROPERTY 4

The Matched Filtering operation may be separated into two matching conditions;
namely spectral phase matching that produces the desired output peak at time T,
and the spectral amplitude matching that gives this peak value its optimum signal to
noise density ratio.

In polar form the spectrum of the signal (t ) being matched may be expressed as

( f ) ( f ) exp j ( f )

where ( f ) is magnitude spectrum and ( f ) is phase spectrum of the signal.

The filter is said to be spectral phase matched to the signal (t ) if the transfer function of
the filter is defined by

H ( f ) H ( f ) exp j ( f ) j 2fT

The output of such a filter is

0' (t ) H ( f ) ( f ) exp( j 2ft )df


H ( f ) ( f ) exp[ j 2f (t T )]df

The product H ( f ) ( f ) is real and non negative.

The spectral phase matching ensures that all the spectral components of the output add
constructively at t = T , there by causing the output to attain its maximum value.

0' (t ) 0' (T ) ( f ) H ( f ) df

For spectral amplitude matching

H ( f ) ( f )
Problem-1:
Consider the four signals s1(t), s2(t), s3(t) and s4(t) as shown in the fig-P1.1 .
Use Gram-Schmidt Orthogonalization Procedure to find the orthonormal
basis for this set of signals. Also express the signals in terms of the basis functions.

Fig-P1.1: Signals for the problem -1.

Solution:

Given set is not linearly independent because s4(t) = s1(t) + s3(t)

Step-1: Energy of the signal s1(t)

T
T
E1 s12 (t ) dt
0
3
s1 (t )
First basis function 1 (t ) 3 for 0 t T
E1 T 3
Step-2: Coefficient s21
T
s 21 0
s 2 (t ) 1 (t ) dt T
3
T
2T
Energy of s2(t) E 2 s 22 (t ) dt
0
3
 s2 (t ) s21 1 (t )
Second Basis function 2 (t ) 3 for T t 2T
E2 s 2 T 3 3
21

T
Step-3: Coefficient s31: s31 0
s3 (t ) 1 (t ) dt 0

T
Coefficient s32 s32 0
s3 (t ) 2 (t ) dt T
3
Intermediate function
g3(t) = s3(t) - s311(t) - s32 2(t)

g3(t) = 1 for 2T/3 < t < T/3

Third Basis function

g 3 (t )
3 (t ) 3 for 2T t T
T T 3

2
g (t ) dt
3
0
The corresponding orthonormal functions are shown in the figure-P1.2.

Fig-P1.2: Orthonormal functions for the Problem-1

Representation of the signals

S1 (t ) T 1 (t )
3

S 2 (t ) T 1 (t ) T 2 (t )
3 3

S 3 (t ) T 2 (t ) T 3 (t )
3 3

S 4 (t ) T 1 (t ) T 2 (t ) T 3 (t )
3 3 3
PROBLEM-2:

Consider the THREE signals s1(t), s2(t) and s3(t) as shown in the fig P2.1. Use
Gram-Schmidt Orthogonalization Procedure to find the orthonormal basis for this
set of signals. Also express the signals in terms of the basis functions.

Fig-P2.1: Signals for the problem -2.

Solution: The basis functions are shown in fig-P2.2.

Fig-P2.2: Orthonormal functions for the Problem-2

Correspondingly the representation of the signals are:

S1 (t ) 2 1 (t )

S2 (t ) 4 1 (t ) 4 2 (t )

S3 (t ) 3 1 (t ) 3 2 (t ) 3 3 (t )
PROBLEM-3:

Consider the signal s(t) in fig-P3.1

a) Determine the impulse
response of a filter matched
to this signal and sketch it as
a function of time.
b) Plot the matched filter output
as a function of time.
c) What is Peak value of the
output?

Fig P3.1
Solution:

The impulse response of the matched filter is time-reversed and delayed version
of the input signal, h(t) = s(T-t) and the output of the filter, y(t) = x(t) * h(t).

Given s(t) = +1 for 0 < t < 0.5

-1 for 0.5 < t < 1.
(a) With T = 1, the impulse
response h(t) is

h(t) = -1 for 0 < t < 0.5

+1 for 0.5 < t < 1.
 Fig. P3.2

(b) The output of the filter y(t) is obtained by convolving the input s(t) and the
impulse response h(t). The corresponding output is shown in the fig. P3.3.

(c) The peak value of the output is 1.0 unit.

Fig. P3.3

Assignment Problem:

Specify a matched filter for the signal S1(t) shown in Fig.-P4.1 Sketch the output of
the filter matched to the signal S1(t) is applied to the filter input.

Fig P4.1