MATCHED FILTER

In the detection probability equations ((106), (109) and (110)) we note

that Pd depends directly on SNR . That is, Pd increases as SNR increases.
Because of this we want to try to be sure that the receiver is designed to
maximize SNR . We do this by including a matched filter in the receiver. As
indicated in Figures 3-1 and 3-2 the matched filter is usually included right
before the signal processor. In fact, in some radars, the matched filter is the
signal processor.
The statements in the above paragraph provide the design requirement
for the matched filter. Specifically, given some signal, s ( t ) , and noise n ( t ) , we
want to find a filter impulse response, h ( t ) , that maximizes SNR at the filter
output. For purposes of this discussion we assume that the signal is not a
random process. Actually, what we are assuming is that the form of the signal
is deterministic (its amplitude can still be random). This is the case in practice,
even for SW2 or SW4 targets.

As indicated in Figure 8 if the input to the matched filter is s ( t ) the

output will be so ( t ) . If the input is n ( t ) the output will be no ( t ) . The output,
instantaneous, normalized, signal power is

Pso ( t ) = so ( t ) .
2
(121)

For purposes of the matched filter design, we define the peak signal power at
the matched filter output as

PS = max Pso ( t ) = Pso ( to ) = so ( to ) .

2
(122)
t

Since no ( t ) is a random process (that we assume is wide-sense

stationary) we must work with its average power. Thus, the average noise
power at the output of the matched filter is

(
Pn = E no ( t )
2
).
1
(123)

With the above we can now define the design criterion for the matched
filter. Specifically, we choose the matched filter so as to maximize the peak
signal power-to-average noise power at the output of the matched filter. In
equation form,
PS
h ( t ) : max (124)
h( t ) Pn

Equation (124) states that we must first write PS and Pn in terms of h ( t ) and
then maximize it with respect to h ( t ) .
1
The choice of symbols for peak signal and average noise power is not coincidental. As will be
shown, these are the same peak signal and average noise power terms included in the radar
range equation.

2005 M. C. Budge, Jr 27
 Figure 8 Matched Filter Block Diagram

If we assume that h ( t ) is linear we can write

so ( t ) = s ( t ) * h ( t ) (125)

and
no ( t ) = n ( t ) * h ( t ) (126)

where * denotes convolution.

We will choose to solve the optimization problem in the frequency domain
through the use of Fourier transforms. To this end we write

H ( f ) =
h ( t)
, (127)

S ( f ) =
s ( t)
and (128)

So ( f ) =
so ( t )
(129)

and
So ( f ) = H ( f ) S ( f ) . (130)

In the above [ x ] denotes the Fourier transform.

Since n ( t ) and no ( t ) are random processes we must deal with them as

such. This means we write

E { n ( t + t ) n* ( t ) }
N ( f ) =
, (131)

E { no ( t + t ) n*o ( t ) }
N o ( f ) =
and (132)

No ( f ) = H ( f ) N ( f ) .
2
(133)

Although not obvious we took advantage of the fact that n ( t ) and no ( t ) are
wide-sense stationary.

We now recognize that N o ( f ) is a power spectral density and thus that

2005 M. C. Budge, Jr 28

N 0 ( f ) df = H ( f ) N ( f ) df .
2
Pn =
- -
(134)

The peak signal power is given by

PS = so ( to ) .
2
(135)

However, we can write

so ( to ) = -1
So ( f )
t =to = S( f ) H ( f ) e

j 2p fto
df . (136)
-

If we combine (136), (135), (134) and (124) we get

2

S( f ) H( f ) e

j 2p fto
df
P
h ( t ) : max S = max -

. (137)
h( t ) P h( t )

H ( f ) N ( f ) df
n 2

At this point we make the assumption that n ( t ) is white with a noise power
spectral density of
N ( f ) = kT0 FnG (138)

and write
2

S( f ) H ( f ) e

j 2p fto
df
h ( t ) : max -

(139)
h( t )
H ( f ) df
2
kT0 FnG
-

In the above equation G is the gain of all receiver components between the
antenna and the input to the matched filter. Thus, N ( f ) is the noise power
spectral density at the input to the matched filter.
We start the maximization process by applying one of the Cauchy-
Schwartz inequalities to the numerator. The particular Cauchy-Schwartz
inequality of interest is
b 2
b b

A ( f ) B ( f ) df A ( f ) df B ( f ) df
2 2

a

a


a
(140)

with equality when

A ( f ) = KB * ( f ) (141)

where K is an arbitrary constant. If we apply (140) to the ratio of (139) with

the associations

2005 M. C. Budge, Jr 29
 A( f ) = H ( f ) (142)

and
B ( f ) = S ( f ) e j 2p fto (143)

we get
2


S( f ) H ( f ) e ( ) S ( f ) df
j 2p fto 2 2
df H f df

-

-

- (144)
H ( f ) df H ( f ) df
2 2
kT0 FnG kT0 FnG
- -

where we have made use of the fact that

S ( f ) e j 2p fto = S ( f ) . (145)

We note that (144) reduces to

2

S( f ) H ( f ) e S ( f )
j 2p fto 2

-
df df

- . (146)
kT0 FnG
H( f )
2
kT0 FnG df
-

Equation (146) tells us that for all H ( f ) the upper bound on the left side is
equal to the right side. That is, we have found the maximum value of PS Pn
over all h ( t ) . That is, we have solved part of the maximization problem. To
find the h ( t ) that yields the maximum PS Pn we invoke the second part of the
Cauchy-Schwartz inequality given in (141). Specifically, we say that

S ( f )
2
df
P (147)
max S = -
h( t ) P kT0 FnG
n

when we choose h ( t ) as

h ( t ) = -1
KS * ( f ) e - j 2p fto
. (148)

Thus, we have found the impulse response of the filter that maximizes peak
signal-to-average noise power at the filter output. Furthermore, we have and
equation for the maximum in the form of (147) and have determined that the
maximum occurs at t = to .
We note from the form of (148) that

H ( f ) = KS ( f ) . (149)

2005 M. C. Budge, Jr 30
In other words, the matched filter frequency response has the same shape as
the frequency spectrum of the signal. They simply differ by a scaling factor K .
This is the reason we term h ( t ) a matched filter.

We now want to look at the specific form of h ( t ) relative to s ( t ) . From

(148) we can write

h( t) = KS * ( f ) e - j 2p fto e j 2p ft df = K
S * ( f ) e - j 2p f ( to -t ) df
- -
* . (150)


= K S ( f ) e j 2p f ( to -t ) df = Ks * ( to - t )
-

Thus, h ( t ) is the conjugate of a scaled (by K ), time reversed (because of the -t

) and shifted (by to ) version of the transmit signal, s ( t ) . This is another reason
we refer to h ( t ) as a matched filter. This operation is illustrated in Figure 9.
The left sketch of this figure is that of s ( t ) while the right figure is a sketch of
s* ( -t ) (we have assumed K = 1 since it is arbitrary). Finally, the right figure is
s* ( to - t ) .

Figure 9 Evolution of

Now that we have established the equation for the maximum value of the
SNR at the output of the matched filter, and have a filter that can provide the
maximum SNR, we want to determine its value. Specifically, we want to relate
the maximum SNR to values of SNR we compute from the radar range equation.

We start by recognizing that s ( t ) is the radar pulse at the input to the

matched filter. It is the rectangular pulse that we have been discussing in our
previous work. Thus, we can write s ( t ) as

2005 M. C. Budge, Jr 31

t -t
s ( t ) = Vrect R (151)
t p

where t R is the range delay to the target and t p is the pulse width. The
representation of (151) is idealized in that it assumes that the receiver
components prior to the matched filter can pass a rectangular pulse without
distortion. Although this is not realistic, for purposes of this development it is a
fairly reasonable assumption. V is, in general, a complex number that may
include a carrier term (the IF frequency), a phase and, possibly, some type of
phase modulation.

When we derived the radar range equation we concluded that the signal
power at the output of the antenna was

PT GT GR l 2s
Ps = . (152)
( 4p )
3
R4L

From above we said that the gain of the receiver from the antenna to the input
to the matched filter is G . If we assume that this same gain is applied to the
signal the power at the input to the matched filter is

PT GT GR l 2s G
PM = PsG = . (153)
( 4p )
3
R4L

If we assume a normalized impedance of 1 ohm, we get

PT GT GR l 2s G
V = PM = . (154)
( 4p )
3
R4L

From (147) we have

S ( f )
2
df
. (155)
SNRMAX = -
kT0 FnG

Since s ( t ) has finite energy and power we can invoke Parsevals Theorem for
the numerator and write

s ( t )
2
dt
. (156)
SNRMAX = -
kT0 FnG
Substituting (151) yields

2005 M. C. Budge, Jr 32

2 t -t

V rect R
t
dt 2

SNRMAX = - p = V tp (157)
kT0 FnG kT0 FnG
or with (154)
PT GT GR l 2s Gt p PT GT GRl 2st p
SNRMAX = = . (158)
( 4p ) ( 4p )
3 3
R 4 LkT0 FnG R 4kT0 Fn L
We recognize (158) as the SNR that we found when we derived the radar range
equation. This tells us that the peak value of SNR at the output of the matched
filter is the SNR we obtain from the radar range equation. In essence, it says
that the matched filter ekes out the maximum possible SNR from the signal and
noise that the radar must deal with. There is no other filter that will give a
larger value of SNR for the transmitted signal and the case where the
interference is due to white noise at the input to the matched. If the
interference is other than white noise, e.g. clutter, there are other filters that
will provide larger values of signal-to-interference power ratio (SIR) than the
filter defined by (150).

MATCHED FILTER RESPONSE FOR AN UNMODULATED PULSE

We want to derive an equation for the matched filter response for a
signal, s ( t ) , and then use it to derive the matched filter response for the case
where s ( t ) is an unmodulated pulse. The results can also be extended to
finding the matched filter response when s ( t ) is an LFM pulse.

From (150) we have

h ( t ) = Ks* ( t0 - t ) (159)

where K is an arbitrary (complex) constant and t0 is the value of t at which the

matched filter response to s ( t ) will reach its peak.

Since we can pick K and t0 to be anything we want, we can let, without

loss of generality, K = 1 and t0 = 0 . The latter statement says that the output of
the matched filter will reach its peak at a relative time of zero. With this we get
h ( t ) = s * ( -t ) . (160)

The response of h ( t ) to s ( t ) is given by

so ( t ) = h ( t ) * s ( t ) = s ( a ) h ( t - a ) da .
(161)
-

* *
( *
)
But h ( t ) = s ( -t ) so h ( t - a ) = s - ( t - a ) = s ( a - t ) and

2005 M. C. Budge, Jr 33

so ( t ) = s ( a ) s ( a - t ) da .

*
(162)
-

For purposes of the rest of the analyses we will treat s ( t ) and s ( t ) as separate
*

functions. Let

t -t 2

s ( t ) = Ae jq rect p (163)
tp
and

t -t 2

s* ( t ) = Ae - jq rect p . (164)
tp

A plot of s ( t ) is shown in Figure 10. The plot of s ( t ) would look the same
*

except that the height would be Ae - jq rather than Ae jq .

Figure 10 Unmodulated pulse

In the so ( t ) integral we note that t is the separation between s ( a ) and

s* ( a - t ) as show in Figure 11. Figure 11 corresponds to the case where t 0 .

Figure 11 Plot of and for

It should be clear that if t t p or t -t p , or t t p , then s ( a ) and s ( a - t )

*

will not overlap, s ( a ) s ( a - t ) = 0 and so ( t ) = 0 . Thus

*

so ( t ) = 0 t t p . (165)

2005 M. C. Budge, Jr 34
 `For 0 t < t p the overlap region of s ( a ) and s ( a - t ) is t a < t p .
*

Further, over this region s ( a ) s ( a - t ) = Ae Ae = A and thus

* jq - jq 2

tp

so ( t ) = A da = A ( t

2 2
p - t) . (166)
t

We note that since t 0 , t = t . Thus we replace t with t to get

so ( t ) = A2 ( t - t ) 0 t < t p . (167)

The arrangement of s ( a ) and s ( a - t ) for -t p < t < 0 is shown in Figure

*

12.

Figure 12 Plot of and for

In this case the overlap region is 0 < a < t + t p and

t +t p

so ( t ) = A da = A ( t + t) .
2 2
p (168)
0

Here we note that since t < 0 , t = - t . Thus we replace t with - t to get

so ( t ) = A2 ( t p - t ) t p < t < 0 . (169)

Since this is the same form as for 0 t < t p , we can combine these to get

so ( t ) = A2 ( t p - t ) t <tp. (170)

Finally, if we combine this with the result for t t p we get

0 t t p
so ( t ) = 2
A ( t p - t ) t <tp
. (171)
t
= A2 ( t p - t ) rect
2t p

2005 M. C. Budge, Jr 35
A plot of so ( t ) is shown in Figure 13.

Figure 13 Plot of Matched Filter Output

2005 M. C. Budge, Jr 36