HayAppDv3.

fm Page 509 Thursday, January 22, 2004 9:10 PM

A P P E N D I X D

Matched Filters

D.1 MATCHED-FILTER RECEIVER

Consider a known signal s(t) corrupted by additive white Gaussian noise w(t), resulting
in the received signal

x(t) = s(t) + w(t) 0≤t≤T (D.1)

What is the optimum receiver for detecting the known signal s(t) in the received signal
x(t)? To answer this fundamental question, we first note the following two important
points:
1. The power spectral density of white noise, with sample function w(t), is defined
by

N
S w ( f ) = -----0- for all f in the entire interval – ∞ < f < ∞ (D.2)
2

The power spectral density of white noise is illustrated in Fig. D.1(a). For a sta-
tionary random process, the autocorrelation function is the inverse Fourier trans-
form of the power spectral density. (See Appendix C.) It follows, therefore, that
the autocorrelation function of white noise consists of a Dirac delta function
δ (τ), weighted by N0/2, as shown in Fig. D.1(b). That is,

R w ( τ ) = E [ w ( τ )w ( t – τ ) ]
N (D.3)
= -----0- δ ( τ )
2
where E is the statistical expectation operator. Accordingly, any two different
samples of white noise are uncorrelated, no matter how closely together in time
they are taken. If the white noise w(t) is also Gaussian, then the two samples are
statistically independent. In a sense, white Gaussian noise represents the ultimate
in randomness.

509
HayAppDv3.fm Page 510 Thursday, January 22, 2004 9:10 PM

510 Appendix D Matched Filters

SW(f )

N0 /2

0 f
(a)

RW(f )

N0
d(t)
2

0 t
(b)

FIGURE D.1 (a) Power spectrum of the additive white noise W(t).
(b) Autocorrelation function of W(t).

2. Since the signal s(t) is known and therefore deterministic, it follows that s(t) and
w(t) are as uncorrelated (i.e., dissimilar) as they could ever be.
In light of point 2, we may intuitively state that, for the problem described herein, the
optimum receiver consists of a correlator with two inputs, one being the noisy received
signal x(t) and the other being a locally generated replica of the known signal s(t), as
shown in Fig. D.2. For obvious reasons, this optimum receiver is known as the correla-
tion receiver.
Another way of constructing the optimum receiver is to use a matched filter,
defined as a linear filter whose impulse response h(t) is a time-reversed, delayed ver-
sion of the known signal s(t); that is,


h(t) =  s(T – t) 0≤t≤T (D.4)
 0 otherwise

x (t) T
y
K0 dt

s(t)

FIGURE D.2 Correlation receiver.

HayAppDv3.fm Page 511 Thursday, January 22, 2004 9:10 PM

Section D.2 Probability of Detection 511

Sample
at t ⫽ T
x (t) Matched filter: y
h(t) ⫽ s(T ⫺ t)

FIGURE D.3 Matched-filter receiver.

Figure D.3 shows a matched filter receiver, which consists of a matched filter followed
by a sampler that is activated at the end of the signaling interval t = T. The important
point to note here is that the correlation receiver of Fig. D.2 and the matched filter
receiver of Fig. D.3 are equivalent insofar as their overall output samples are con-
cerned. Specifically, for the same input signal and at the end of a signaling interval, the
resulting output samples produced by these two receivers are identical.

D.2 PROBABILITY OF DETECTION

To detect a signal with a correlation receiver or a matched-filter receiver, the output
sample is compared against a threshold and then a decision is made by the receiver,
depending on whether the threshold is exceeded or not. In so doing, the receiver
makes a decision in favor of one of two hypotheses:
Hypothesis H1: The known signal s(t) is present in the received signal x(t), a
decision that is made when the threshold is exceeded.
Hypothesis H0: The received signal x(t) consists solely of noise w(t), a deci-
sion that is made when the threshold is not exceeded.
Clearly, the receiver is subject to errors due to the random behavior of the additive
noise w(t) in the received signal x(t).
To calculate the average probability of error incurred by the receiver, we proceed
by using Eq. (D.1) as the input signal applied to the correlation receiver of Fig. D.2.
The resulting output sample is
T
y = ∫0 x ( t )s ( t ) dt
T
= ∫0 ( s ( t ) + w ( t ) )s ( t ) dt
(D.5)
T 2 T
= ∫0 s ( t ) dt + ∫0 w ( t )s ( t ) dt
T
= E+ ∫0 w ( t )s ( t ) dt
where
T 2
E = ∫0 s ( t ) dt (D.6)

is the energy of the known signal s(t).

HayAppDv3.fm Page 512 Thursday, January 22, 2004 9:10 PM

512 Appendix D Matched Filters

Since, by assumption, the white noise w(t) is the sample function of a Gaussian
process W(t), it follows that the receiver output y is the sample of a Gaussian-distributed
random variable Y. To complete the characterization of the receiver output, we need to
determine its mean and variance.
The mean of the random variable Y is

µY = E [ Y ]
T
= E+E ∫0 W ( t )s ( t ) dt
(D.7)
T
= E+E ∫0 [ W ( t ) ]s ( t ) dt
= E

where we have used two facts: First, the known signal s(t) is deterministic and there-
fore unaffected by the expectation operator E. Second, by assumption, the mean of
the white noise process W(t) is zero.
The variance of the random variable Y is

2 2
σY = E [ ( Y – µY ) ]
T T
= E ∫0 ∫0 W ( t1 )W ( t2 )s ( t1 )s ( t2 ) dt1 dt2 (D.8)
T T
= ∫0 ∫0 E [ W ( t1 )W ( t2 ) ]s ( t1 )s ( t2 ) dt1 dt2
Invoking the use of Eq. (D.2), we may write

N
E [ W ( t 1 )W ( t 2 ) ] = -----0- δ ( t 1 – t 2 ) (D.9)
2

Substituting Eq. (D.9) into (D.8) yields

2 N T T
σ Y = -----0- ∫ ∫ δ ( t 1 – t 2 )s ( t 1 )s ( t 2 ) dt 1 dt 2
2 0 0
N T 2
= -----0- ∫0 s ( t 1 ) dt 1 (D.10)
2
N0 E
= ---------
-
2

where E is the signal energy.

Putting all the pieces together, we can now say that the correlation receiver
output y is the sample value of a Gaussian-distributed random variable Y with mean
HayAppDv3.fm Page 513 Thursday, January 22, 2004 9:10 PM

Section D.2 Probability of Detection 513

µY = E and variance σ Y2 = N 0 E ⁄ 2 . Accordingly, we may express the probability den-

sity function of the random variable Y as

2
1  ( y – µY ) 
f Y ( y ) = ----------------- exp  – ---------------------
-
2 πσ Y  2σ
2 
Y
(D.11)
1  ( y – E ) 2
= ------------------ exp  – -------------------
π N0 E  N0 E 

which is plotted in Fig. D.4.

Let λ denote the threshold against which the correlator output y is compared. As
stated previously, when y > λ, the receiver decides in favor of hypothesis H1; otherwise
it decides in favor of hypothesis H0. Accordingly, the conditional probability of error,
given that the known signal s(t) is present in the receiver input, is defined by
λ
Prob ( say H 0 H 1 is true ) = ∫–∞ fY ( y ) dy (D.12)

which is illustrated graphically in Fig. D.4. Substituting Eq. (D.11) into (D.12) yields

1 λ  ( y – E ) 2
Prob ( say H 0 H 1 is true ) = ------------------
π N0 E ∫ –∞
exp  – ------------------- dy
 N0 E 
(D.13)

To simplify matters, let

y–E
z = -------------- (D.14)
N0 E

which means that

dy
dz = --------------
N0 E

fY (y)

Conditional
probability of
error, given
that the known signal
s(t) is present
(i.e., y ⬎ l)

0 l E y

FIGURE D.4 Probability distribution of the correlation receiver output.

HayAppDv3.fm Page 514 Thursday, January 22, 2004 9:10 PM

514 Appendix D Matched Filters

Hence, we may rewrite Eq. (D.13) as

1 ( λ – E ) ⁄ N0 E 2
Prob ( say H 0 H 1 is true ) = -------
π ∫ –∞ exp ( – z ) dz
(D.15)
1 ∞ 2
= -------
π ∫( E – λ ) ⁄ N0 E
exp ( – z ) dz

At this point in the discussion, we digress briefly to introduce a function that is closely
related to the Gaussian distribution: the error function, defined by

2 u 2
erf ( u ) = -------
π ∫0 exp ( –z ) dz (D.16)

Table E.1 of Appendix E gives values of the error function erf(u) for the argument u
in the interval 0 ≤ u ≤ 3.3. The error function has two useful properties:
1. Symmetry property, described by

erf ( – u ) = – erf ( u ) (D.17)

2. Asymptote property, which, for the argument u approaching infinity, is described
by
2 ∞ 2
erf ( ∞ ) = -------
π ∫0 exp ( –z ) dz
(D.18)
= 1
Another function, the complementary error function, is defined by

2 ∞ 2
erfc ( u ) = -------
π ∫u exp ( –z ) dz (D.19)

which is related to the error function by the formula

erfc ( u ) = 1 – erf ( u ) (D.20)

We may now reformulate the conditional probability of error of Eq. (D.15) in terms of
the complementary error function by writing

1  E – λ
Prob ( say H 0 H 1 is true ) = --- erfc  -------------- (D.21)
2  N E 0

From Eq. (D.21), the following points are noteworthy:

• The signal energy E and noise spectral density N0 have different physical inter-
pretations, in that E is measured in joules whereas N0 is measured in watts/hertz;
yet these two units are in fact equal.
HayAppDv3.fm Page 515 Thursday, January 22, 2004 9:10 PM

Section D.4 Matched Filtering for Complex Signals 515

• Insofar as the signal component is concerned, the probability of error is indepen-

dent of the waveform of the known signal s(t), and the only parameter that mat-
ters is the signal energy E.
• The threshold λ is measured in joules.

D.3 ANOTHER PROPERTY OF THE MATCHED FILTER

Equation (D.21) sums up one important property of the matched-filter receiver in the
combined presence of signal and noise at the filter input. For another important prop-
erty of the matched filter, consider the case of a noiseless input. Then, with the input
x(t) = s(t) and the impulse response h(t) = s(T − t), the resulting filter output is defined
by the convolution integral

y ( t ) = Rs ( T – t )
(D.22)
= Rs ( t – T ) if w ( t ) = 0

The integral of Eq. (D.22) is recognized as the deterministic autocorrelation function

of the signal component s(t) for a lag of T − t, namely, Rs(T − t). Accordingly, we may
write
y ( t ) = Rs ( T – t )
(D.23)
= Rs ( T – t ) wt = 0

where, in the second line, we have used the fact that the autocorrelation function of
a signal of finite energy is an even function of the lag (see Appendix A). In words,
Eq. (D.23) states that the output of a filter matched to an input signal is equal to the
autocorrelation function of that signal, delayed by an amount equal to the duration of
the signal.

D.4 MATCHED FILTERING FOR COMPLEX SIGNALS

The material presented thus far on matched filtering applies to real-valued signals.
When dealing with complex-valued signals, we make a simple modification to Eq. (D.4).
Specifically, the impulse response of a filter matched to a complex-valued signal s(t) is
defined by


h ( t ) =  s* ( T – t ) 0≤t≤T (D.24)
0 otherwise

where the asterisk denotes complex conjuction. Except for this minor modification,
everything else presented in the Appendix remains intact.