Gram-Schmidt Orthogonalization

The principle of Gram-Schmidt Orthogonalization (GSO) states that, any set of M energy signals, {si(t)}, 1 i M can be expressed as linear combinations of N orthonormal basis functions, where N M. If s1(t), s2(t), .., sM(t) are real valued energy signals, each of duration T sec,
N 0 t T si (t ) = sij j (t ) ; j =1 i = 1, 2,...., M N

4.16.1

where, i = 1, 2,......, M 4.16.2 sij = si (t ) j (t )dt ; j = 1, 2,......, N 0 The j(t)-s are the basis functions and sij-s are scalar coefficients. We will consider realvalued basis functions j (t) - s which are orthonormal to each other, i.e.,
T T

(t ).
i 0

1, if i = j (t ) dt = 0, if i j

4.16.3

Note that each basis function has unit energy over the symbol duration T. Now, if the basis functions are known and the scalars are given, we can generate the energy signals, by following Fig. 4.16.1. Or, alternatively, if we know the signals and the basis functions, we know the corresponding scalar coefficients (Fig. 4.16.2).

si 1

1 (t)

si 2

2 (t)

+ +

si (t) 0t T

siN

N (t) Fig. 4.16.1 Pictorial depiction of Equation 4.16.1

Correlator Structure
T

s2(t) s1(t) |T|T| 1 (t)

( )dt
0

si1

si ri (t)

2 (t)

( )dt
0

si2

N (t)

( )dt
0

siN Scalar

Fig. 4.16.2 Pictorial depiction of Equation 4.16.2

Justification for G-S-O procedure

Part I: We show that any given set of energy signals, {si (t)}, 1 i M over 0 t < T, can be completely described by a subset of energy signals whose elements are linearly independent. To start with, let us assume that all si(t) -s are not linearly independent. Then, there must exist a set of coefficients {ai}, 1 < i M, not all of which are zero, such that, a1s1 (t) + a2s2 (t) + + aM sM (t) = 0, 0t<T 4.16.4

Verify that even if two coefficients are not zero, e.g. a1 0 and a3 0, then s1(t) and s3(t) are dependent signals. Let us arbitrarily set, aM 0. Then, 1 sM (t ) = [ a1s1 (t ) + a2 s2 (t ) + ........ + aM 1sM 1 (t )] aM

1 aM

M 1 i =1

a s (t )
i i

4.16.5

Eq.4.16.5 shows that sM(t) could be expressed as a linear combination of other si(t) s, i = 1, 2, .., (M 1). i = 1,2,.., (M 1). Next, we consider a reduced set with (M-1) signals {si(t)}, This set may be either linearly independent or not. If not, there exists a set of {bi}, i = 1,2, (M 1), not all equal to zero such that,
M 1 i =1

b s (t ) = 0, 0 t < T
i i

4.16.6

Again, arbitrarily assuming that bM-1 0, we may express sM-1(t) as: 1 M 2 4.16.7 sM 1 (t ) = bi si (t ) bM 1 i =1 Now, following the above procedure for testing linear independence of the remaining signals, eventually we will end up with a subset of linearly independent signals. Let {si(t)}, i = 1, 2, ., N M denote this subset. Part II : We now show that it is possible to construct a set of N orthonormal basis functions 1(t), 2(t), .., N(t) from {si(t)}, i = 1, 2, .., N. s (t ) Let us choose the first basis function as, 1 (t ) = 1 , where E1 denotes the energy of the E1 first signal s1(t), i.e.,
E1 = s12 (t )dt :
0 T

s1 (t ) = E1 .1 (t ) = s111 (t ) Where, s11 = E1 Now, from Eq. 4.16.2, we can write

4.16.8

s 21 = s2(t) 1(t )dt

0

4.16.9 4.16.10

Let us now define an intermediate function: g2(t) = s2(t) s211(t); 0 t < T Note that,
T

g 2 (t )1 (t )dt = s2 (t )1 (t )dt s21 1 (t )1 (t )dt

0 0 0

= s21 s21 = 0 g 2 (t ) Orthogonal to 1(t); So, we verified that the function g2(t) is orthogonal to the first basis function. This gives us a clue to determine the second basis function. Now, energy of g2(t)
= g 2 2 (t )dt
0 T

= [ s2 (t ) s211 (t )] dt
2

= s2 2 (t )dt 2.s21 s2 (t )1 (t )dt + s212 12 (t )dt

0 0 0

0 T

= E2 2.s21.s21 + s21 = E2 s
2

2 21

( Say ) So, we now set,

4.16.11

2 (t ) =

g 2 (t )
T

s2 (t ) s211 (t ) E2 s212

4.16.12

g
0

2 2

(t )dt

and Verify that:

E2 = s2 2 (t )dt : Energy of s2(t)

0 T

0 1

2 2

(t )dt = 1 ,

i.e. 2(t) is a time limited energy signal of unit energy.

and

(t ).
0

(t )dt = 0 , i.e. 1(t) and 2(t) are orthonormal to each other.

Proceeding in a similar manner, we can determine the third basis function, 3(t). For i=3,

g3 (t ) = s3 (t ) s3 j J (t ) ;
j 1

0 t <T

= s3 (t ) [ s311 (t ) + s32 2 (t )]
where,
s31 = s3 (t )1 (t )dt and s32 = s3 (t ) 2 (t )dt
0 0 T T

It is now easy to identify that, g3 (t ) 3 (t ) =

T

4.16.13

g
0

2 3

(t )dt
gi (t ) Egi

Indeed, in general,

i (t ) =

gi (t )
T

4.16.14

g
0

2 i

(t )dt
i 1

for i = 1, 2,.., N, where

gi (t ) = si (t ) sij j (t )
j =1

4.16.15

and

sij = si (t ). j (t )dt
0

4.16.16

for i = 1, 2,.., N and j = 1, 2, , M Let us summarize the steps to determine the orthonormal basis functions following the Gram-Schmidt Orthogonalization procedure: If the signal set {sj(t)} is known for j = 1, 2,.., M, 0 t <T, Derive a subset of linearly independent energy signals, {si(t)},i = 1, 2,.., N M. Find the energy of s1(t) as this energy helps in determining the first basis function 1(t), which is a normalized form of the first signal. Note that the choice of this first signal is arbitrary. Find the scalar s21, energy of the second signal (E 2), a special function g2(t) which is orthogonal to the first basis function and then finally the second orthonormal basis function 2(t) Follow the same procedure as that of finding the second basis function to obtain the other basis functions.

Concept of signal space

Let, for a convenient set of {j (t)}, j = 1,2,,N and 0 t <T,
si (t ) = sij j (t ), i= 1, 2,, M and 0 t < T, such that,
j =1 N

sij = si (t ) j (t )dt
0

Now, we can represent a signal si(t) as a column vector whose elements are the scalar coefficients sij, j = 1, 2, .., N : si1 s si = i 2 ; i= 1, 2,, M siN 1 N

4.16.17

These M energy signals or vectors can be viewed as a set of M points in an N dimensional Euclidean space, known as the Signal Space (Fig.4.16.3). Signal Constellation is the collection of M signals points (or messages) on the signal space.

2 E2

E1

s2
0

s1
s11

s12
1 (t )

s3

E3
2D Signal Space

Fig. 4.16.3 Sketch of a 2-dimentional-signal space showing three signal vectors

s1 , s 2 and s3
Now, the length or norm of a vector is denoted as si . The squared norm is the inner product of the vector:
2

si

= si , si = s 2ij
j =1

4.16.18

The cosine of the angle between two vectors is defined as: si , s j cos angle between si & s j = si s j

4.16.19

si & s j are orthogonal to each other if si , s j = 0 . If E i is the energy of the i-th signal vector,

Ei =

T 2 i

s ( t ) dt =
N T

N N sij j ( t ) sik k ( t ) dt j=1 k=1

= sij sik j ( t ) k ( t ) dt as { j ( t )} forms an ortho-normal set

j=1 k=1 N 0 2 = sij = si j=1 2

4.16.20
2 N 2 T 2

For a pair of signals si (t) and sk(t), si -sk = ( sij -s kj ) = si ( t ) -s k ( t ) dt 0

j=1

It may now be guessed intuitively that we should choose si (t) and sk (t) such that the Euclidean distance between them, i.e. si -sk is as much as possible to ensure that their detection is more robust even in presence of noise. For example, if s1 (t) and s2 (t) have same energy E, (i.e. they are equidistance from the origin), then an obvious choice for maximum distance of separation is, s1(t) = -s2(t).

Use of Signal Space for Signal Detection in a Receiver

The signal space defined above, is very useful for designing a receiver as well. In a sense, much of the features of a modulation scheme, such as the number of symbols used and the energy carried by the symbols, is embedded in the description of its signal space. So, in absence of any noise, the receiver should detect one of these valid symbols only. However, the received symbols are usually corrupted and once placed in the signal space, they may not match with the valid signal points in some respect or the other. Let us briefly consider the task of a good receiver in such a situation. Let us assume the following: 1. One of the M signals si(t), i=1,2,.,M is transmitted in each time slot of duration T sec. 2. All symbols are equally probable, i.e. the probability of occurrence of si(t) = 1/M, for all i. 3. Additive White Gaussian Noise (AWGN) processes W (t) is assumed with a noise N sample function w (t) having mean = 0 and power spectral density 0 [N0: 2 single sided power spectral density of additive white Gaussian noise. Noise is discussed more in next two lessons] 4. Detection is on a symbol-by-symbol basis. Now, if R (t) denotes the received random process with a sample function r (t), we may write, r ( t ) =si ( t ) +w ( t ) ; 0 t < T and i = 1,2,.,M. The job of the receiver is to make best estimate of the transmitted signal si (t) (or, equivalently, the corresponding message symbol mi) upon receiving r(t). We map the received sample function r (t) on the signal space to include a received vector or

received signal point. This helps us to identify a noise vector, w(t) ,also. The detection problem can now be stated as: Given an observation / received signal vector ( r ), the receiver has to perform a mapping from r to an estimate m for the transmitted symbol mi in a way that would minimize the average probability of symbol error. Maximum Likelihood Detection scheme provides a general solution to this problem when the noise is additive and Gaussian.

Maximum likelihood (ML) detection:

We start with the following assumptions: Number of information-bearing signals (symbols), designed after the G-S-O approach, is M and one of these M signals is received from the AWGN channel in each time slot of T-sec. Let the messages be denoted by mi, i = 1, 2, , M. Each message, as discussed in an earlier lesson, may be represented by a group of bits, e.g. by a group of m bits each such that 2m = M. All symbols are equi- probable. This is not a grave assumption since, if the input message probabilities are different and known, they can be incorporated following Bayesian approach. However, for a bandwidth efficient transmission scheme, as is often needed in wireless systems, the source coding operation should be emphasized to ensure that all the symbols are independent and equally likely. Alternatively, the number of symbols, M, may also be decided appropriately to approach this desirable condition. AWGN process with a mean = 0 and double-sided psd No/2. Let w(t) denote a noise sample function over 0 t < T.

Let R(t) denote the received random process with sample function over a symbol duration denoted as r(t), 0 t T. Now, a received sample function can be expressed in terms of the corresponding transmitted information-bearing symbol, say si(t), and a sample function w(t) of the Gaussian noise process simply as: 4.19.1 r(t) = si(t) + w(t), 0 t < T At the receiver, we do not know which si(t) has been transmitted over the interval 0 t < T. So, the job of an efficient receiver is to make best estimate of transmitted signal [si(t)] upon receiving r(t) and to repeat the same process during all successive symbol intervals. This problem can be explained nicely using the concept of signal space. Depending on the modulation and transmission strategy,the receiver usually has the knowledge about the signal constellation that is in use. This also means that the receiver knows all the nominal basis functions used by the transmitter. For convenience, we will mostly consider a transmission strategy involving two basis functions, 1 and 2 (described now as unit vectors) for explanation though most of the discussion will hold for any number of basis functions. Fig.4.19.1 shows a two-dimensional signal space showing a signal vector si and a received vector r . Note the noise vector . as well.

Received Vector

si

Signal vector

0
Noise Vector

Fig. 4. 19.1 Signal space showing a signal vector si and a received vector r The job of the receiver can now be formally restated as: Given received signal vectors r , find estimates mi for all valid transmit symbols mi-s once in each symbol duration in a way that would minimize the probability of erroneous decision of a symbol on an average (continuous transmission of symbols is implicit). The principle of Maximum Likelihood (ML) detection provides a general solution to this problem and leads naturally to the structure of an optimum receiver. When the receiver takes a decision that m = mi , the associated probability of symbol decision error may be expressed as: Pe mi , r = probability of decision on receiving r that mi was transmitted = Pr (mi not sent r ) = 1 Pr (mi sent r ). In the above, Pr (mi not sent r ) denotes the probability that mi was not transmitted while r is received. So, an optimum decision rule may heuristically be framed as: Set m = mi if Pr (mi sent r ) Pr (mk sent r ), for all k i 4.19.2

This decision rule is known as maximum a posteriori probability rule. This rule requires the receiver to determine the probability of transmission of a message from the received vector. Now, for practical convenience, we invoke Bayes rule to obtain an equivalent statement of optimum decision rule in terms of a priori probability: r mi r

( )

r r

()

= r r mi r ( mi )
a priori prob. of r given 'mi ' 1 M

4.19.3

a posteriori prob. joint probability of r of 'mi ' given r

r mi r

( )

: A posteriori probability of mi given r

10

r r : Joint pdf of r , defined over the entire set of signals { si(t) }; independent of any specific message mi r r mi : Probability that a specific r will be received if the message mi is transmitted;

()

( )

known as the a priori probability of r given mi r (mi ) : 1 M From Eq. 4.19.3, we see that determination of maximum a posteriori probability is equivalent to determination of maximum a priori probability r r mi . This a priori

( )

probability is also known as the likelihood function. So the decision rule can equivalently be stated as: Set m = mi if r r mi is maximum for k = i

( )
)

Usually, ln p r r|m k , i.e. natural logarithm of the likelihood function is considered. As the likelihood function is non-negative, another equivalent form for the decision rule is: 4.19.4 Set m = mi if ln [ r r mi ] is maximum for k = i

( )

A Maximum Likelihood Detector realizes the above decision rule.Towards this, the signal space is divides in M decision regions, Zi , i = 1, 2, , M such that, vector r lies inside 'Zi 'if, 4.19.5 ln Pr r mk is maximum for k = i

Fig. 4.19.2 indicates two decision zones in a two-dimensional signal space. The received vector r lies inside region Zi if ln p r r|m k is maximum for k = i.

2
s1

Z2 Z1
s2

Fig. 4.19.2 Decision zones in a two-dimensional signal space

11

Now for an AWGN channel, the following statement is equivalent to ML decision: Received vector r lies inside decision region Zi if,

(r s )
j kj j =1

is minimum for k = i

4.19.6

That is, the decision rule simply is to choose the signal point si if the received vector r is closest to si in terms of Euclidean distance. So, it appears that Euclidean distances of a received vector r from all the signal points are to be determined for optimum decisionmaking. This can, however, be simplified. Note that, on expansion we get,

( rj skj ) = rj 2 2 rj .skj + s 2kj

N 2 N N N j =1 j =1 j =1 j =1

4.19.7

It is interesting that, the first term on the R.H.S, i.e.,

r
j =1

2 j

is independent of k and
N

hence need not be computed for our purpose. The second term, 2 rj.skj is the inner
j =1

product of two vectors. The third term, i.e.

s
j =1

2 kj

is the energy of the k-th symbol. If the

modulation format is so chosen that all symbols carry same energy, this term also need not be computed. Many popular digital modulation schemes such as BPSK, QPSK exhibit this property in a linear time invariant channel. So, a convenient observation is: the received vector r lies in decision region Zi if,

N 1 rj skj Ek is maximum for k = i 2 j =1

That is, a convenient form of the ML decision rule is: N 1 Choose m = mi if rj skj Ek is maximum for k = i 2 j =1

4.19.8

A Correlation Receiver, consisting of a Correlation Detector and a Vector Receiver implements the M L decision rule [4.19.8] by, (a) first finding r with a correlation detector and then (b) computing the metric in [4.19.8] and taking decision in a vector receiver. Fig. 4.19.3 shows the structure of a Correlation Detector for determining the received vector r from the received signal r(t). Fig. 4.19.4 highlights the operation of a Vector Receiver.

12

1 (t )

dt
0

r1 = si1 + i1
t =T

r (t )

2 (t )

dt
0

r2

Received Vector

0t T

t =T

r1 r r= 2 : rN 1 N

N (t )

dt
0

rN

t =T

Fig. 4.19.3 The structure of a Correlation Detector for determining the received vector r from the received signal r(t)

13

( r, s )
1

Inner product

1 E1 2

s1

Inner product

( r, s )
2

Select

1 E2 2

the Largest

Estimate of m

s2

Inner product

( r, s )
M

1 EM 2

sM

Inner product

( r, s )
i

si

Accumulator

( r, s )
i

SCALAR

si

Fig. 4. 19.4 Block schematic diagram for the Vector Receiver

Features of the received vector

We will now discuss briefly about the statistical features of the received vector r as obtained at the output of the correlation detector [Fig. 4.19.3]. The j-th element of r , which is obtained at the output of the j-th correlator once in T second, can be expressed as:
rj = r (t ) j (t )dt = [ si (t ) + w(t ) ] j (t )dt
0 0 T T

= sij + w j ;

j=1,2,.., N

4.19.9

14

Here wj is a Gaussian distributed random variable with zero mean and sij is a scalar signal component of si . Now, the mean of the correlator out put is, E rj = E sij + w j = E sij = sij = mrj , say. We note that the mean of the correlator out put is independent of the noise process. However, the variances of the correlator outputs are dependent on the strength of accompanying noise:
2 = E ( rj sij ) = E w j 2 j T T = E w(t ) j (t )dt w(u ) j (u )du 0 0 T T = E j (t ) j (u ).w(t ) w(u )dtdu 0 0 Taking the expectation operation inside, we can write

Var r j = r

rj = j (t ) j (u ) E w ( t ) .w ( u ) dtdu
2

T T

= j (t ) j (u ) Rw (t , u )dtdu
0 0

0 0 T T

4.19.10

Here, Rw (t-u) is the auto correlation of the noise process. As we have learnt earlier, additive white Gaussian noise process is a WSS random process and hence the autocorrelation function may be expressed as, Rw ( t , u ) = Rw ( t u ) and further,

N0 ( t u ) , where No is the single-sided noise power spectral density in 2 Watt/Hz. So, the variance of the correlator output now reduces to: T T N0 2 rj = j (t ) j (u ) (t u )dtdu 2 0 0 Rw ( t u ) = N N = 0 j 2 (t )dt = 0 2 0 2
T

4.19.11

It is interesting to note that the variance of the random signals at the out puts of all N correlators are a) same, b) independent of information-bearing signal waveform and c) dependent only on the noise psd. Now, the likelihood function for si(t), as introduced earlier in Eq.4.19.3 and the ML decision rule [4.19.5] , can be expressed in terms of the output of the correlation detector. The likelihood function for mi = r r mi = f r r mi = f r r si (t ) , where, fr(r|mi) is the conditional pdf of r given mi.

( )

( )

15

In our case, fr r mi = fr j ( r j mi ) , i = 1,2,,M

N j =1

4.19.12
2
j

where, fr j ( r j mi ) is the pdf of a Gaussian random variable with mean sij & var. = r =
N0 , i.e., fr j r j mi = 2

1 2 r
2
j

( r j sij )2 2 rj 2

.e

4.19.13

Combining Eq. 4.19.12 and 4.19.13, we finally obtain,

2 1 N rj sij ) , i=1,2,..,M 4.19.14 .exp ( N 0 j =1 This generic expression is of fundamental importance in analyzing error performance of digital modulation schemes .

f r r mi = ( N 0 )

( )

N 2

MATCHED FILTER
Certain structural modification and simplifications of the correlation receiver are possible by observing that, (a) All orthonormal basis functions j s are defined between 0 t Tb and they are zero outside this range . (b) Analog multiplication, which is not always very simple and accurate to implement, of the received signal r(t) with time limited basis functions may be replaced by some filtering operation. Let, h j (t ) represent the impulse response of a linear filter to which r(t) is applied. Then, the filter output y j (t ) may be expressed as:

y j (t ) = r ( ) h j (t )d
Now, let, h j (t ) = j (T t ), a time reversed and time-shifted version of j (t ) .

4.20.1

Now, y j (t ) =

r ( ). j [T (t )]d

r ( ). j (T + t )d

4.20.2

If we sample this output at t = T,

y j (T ) =

Let us recall that j (t ) is zero outside the interval 0 t T . Using this, the above equation may be expressed as, y j (T ) = r ( ) j ( )d
0 T

( ). j ( )d

4.20.3

16

From our discursion on correlation receiver, we recognize that, r j = r ( ) j ( )d = y j ( )

0 T

4.20.4

The important expression of (Eq.4.20.4) tells us that the j th correlation output can equivalently be obtained by using a filter with h j (t ) = j (T t ) and sampling its output at t = T. The filter is said to be matched to the orthonormal basis function j (t ) and the alternation receiver structure is known as a matched filter receiver. The detector part of the matched filter receiver is shown in [Fig.4.20.1].

0 (t) 0
2T

0 (t) h (t)

t
t=T Filter h1 (t ) = 1 (T t )

r1

h2 (t ) = 2 (T t )

r2

Observation vector r

r (t ) rN

hN (t ) = N (T t )

Sample at t = T Matched filters Fig. 4.20.1: The block diagram of a matched filter bank that is equivalent to a Correlation Detector
A physically realizable matched filter is to be causal and h j (t ) = 0 for t < 0. Note that if j (t ) is zero outside 0 t T , h j (t ) = j (T t ) is a causal impulse response.

Properties of a Matched Filter

We note that a filter which is matched to a known signal (t ), 0 t T , is characterized by an impulse response h(t) which is a time reversed and delayed version of (t ) i.e. 4.20.5 h(t ) = (T t )
17

In the frequency domain, the matched filter is characterized (without mach explanation at this point), by a transfer function, which is, except for a delay factor, the complex conjugate of the F.T. of (t ) , i.e. ( f ) = ( f ) exp( j 2 fT ) 4.20.6
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 ) denote the F.T. of the filter of output 0 (t ) . Then, 0 ( f ) = ( f ) ( f )

= ( f )( f ) exp( j 2 fT )
= ( f )
2

exp( j 2 fT )

4.20.7

Energy spectral density of ( t )

Property (2): The output signal of a matched filter is proportional to a shifted version of the autocorrelation function of the in the input signal to which the filter is matched. This property follows from Property (1). As the auto-correlation function and the energy spectral density form F.T. pair, by taking IFT of (Eq.4.20.7), we may write, 4.20.8 0 (t ) = R (t T )

Where R ( ) is the act of (t ) for ' lag ' . Note that at t = T,

R (0) = 0 (t ) = Energy of (t ).

4.20.9

Property (3): The output SNR of a matched filter depends only on the ratio of the signal energy to the psd of the white noise at the filter input.

Let us consider a filter matched to the input signal (t ) . From property (2), we see that the maximum value of 0 (t ) at t = T is 0 (t T ) = E . Now, it may be shown that the average noise power at the output of the matched N N0 2 2 filter is given by, E[n (t )] = 4.20.10 ( f ) df = 20 .E 2 The maximum signal power = 0 (T ) = E 2 .
E2 2E ( SNR ) max = = N0 E N0 2 Note that SNR in the above expression is a dimensionless quantity.
2

Hence,

4.20.11

This is a very significant result as we see that the SNRmax depends on E and N0 but not on the shape of (t ) . This means a freedom to the designer to select specific pulse shape to

18

optimize other design requirement (the most usual requirement being the spectrum or, equivalently, the transmission bandwidth) while ensuring same SNR.
Property (4): The matched-filtering operation may be separated into two matching condition: namely, spectral phase matching that produces the desired output peak at t = T and spectral amplitude matching that gives the peak value its optimum SNR. ( f ) = ( f ) exp[ j ( f )] 4.20.12

The filter is said to be matched to the signal (t ) in spectral phase if the transfer function of the filter follows: ( f ) = ( f ) exp[ j ( f ) j 2 fT ] 4.20.13 Here ( f ) is real non-negative and T is a positive constant. The output of such a filter is,

0 (t ) =
'

( f ).( f ), exp( j 2 ft )df

( f ) ( f ) .exp[ j 2 f (t T )df

Note that, (t ) (t ) is real and non-negative. Spectral phase matching ensures that all spectral components of 0 ' (t ) add constructively at t = T and thus cause maximum value of the output:

0 (T ) =
'

( f ) ( f ) df . 0 ' (t )

4.20.14

For spectral amplitude matching, we choose the amplitude response ( f ) of the filter to shape the output for best SNR at t = T by using ( f ) = ( f ) . The standard matched filter achieves both these features.

Maximization of output Signal to-Noise Raito:

Let, h(t) be the impulse response of a linear filter and x(t ) = (t ) + (t ), 0 t T : is the input to the filter where (t ) is a known signal and (t ) is an additive white noise sample function with zero mean and psd of (N0/2) Watt/Hz. Let, (t ) be one of the orthonormal basis functions. As the filter is linear, its output can be expressed as, y(t) = 0(t) + n(t), where 0(t) is the output due to the signal component (t) and n(t) is the output due to the noise component (t).[Fig. 4.20.2].

19

 (t )

+ +

(t )
White noise

h(t)

y(t)

Sample at t = T

Fig. 4.20.2: A matched filter is fed with a noisy basis function to which it is matched We can now re-frame the requirement of minimum probability of error (or maximum likelihood detection) as: The filter should make power of 0 (t ) considerably greater (in fact, as large as possible) compared to the power of n(t) at t = T. That is, the filter should maximize the output signal-to-noise power ratio [(SNR)0] 2 0 (T ) max E[n 2 (t )] The following discussion shows that the SNR is indeed maximized when h(t) is matched to the known input signal (t ) .

Let,

( f ) : F.T. of known signal (t ) ( f ) : Transfer function of the linear filter. 0 ( f ) = ( f ) ( f )

and

0 (t ) =

( f )( f ) exp( j 2 ft )df
2

4.20.15

The filter output is sampled at t = T. Now,

0 (T ) =

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

4.20.16

Let, SN(f) : Power spectral density of noise at the output of the linear filter. So, N 2 S N ( f ) = 0 . ( f ) 2 Now, the average noise power at the output of the filter

4.20.17

= E[n 2 (t )] =

S N ( f )df
4.20.18

N = 0 2

( f ) df

Form Eq. 4.20.16 and 4.20.18, we can write an expression of the output SNR as:

20

( SNR )0 =

2 (T )

E[n 2 (t )]

( f ). ( f ) exp( j 2 fT )df
N0 2

4.20.19 ( f ) df
2

Our aim now is to find a suitable form of H(f) such that (SNR)0 is maximized. We use Schwarzs inequality for the purpose.

Schwarzs Inequality
Let x(t ) and y (t ) denote any pair of complex-valued signals with finite energy, i.e.

x(t ) dt < &

y (t ) dt < . Schwarzs Inequality states that,

x(t ) y (t )dt

x(t ) dt. y (t ) dt.

4.20.20

The equality holds if and only if y (t ) = k .x (t ), where k is a scalar constant. This implies, y (t ) x(t ) = k .x(t ) x (t ) a real quantity. Now, applying Schwarzs inequality on the numerator of (Eq.4.20.19), we may write,

2

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

( f ) df

( f ) df

4.20.21

Using inequality (4.20.21), equation (4.20.19) may be expressed as,

2 ( SNR )0 N0

( f ) df

4.20.22

Now, from Schwarzs inequality, the SNR is maximum i.e. the equality holds, when opt ( f ) = ( f ).exp( j 2 fT ). [Assuming k = 1, a scalar)

We see, hopt (t ) =

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

4.20.23

Now, (t ) is a real valued signal and hence,

( f ) = ( f ) Using Eq. 4.20.24 we see,

4.20.24

hopt (t ) =

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

= (T t )

hopt (t ) = (T t )

This relation is the same as we obtained previously for a matched filter receiver. So, we can infer that, SNR maximization is an operation, which is equivalent to minimization of average symbol error (Pe) for an AWGN Channel.

21