MFT

Linear Time Invariant (LTI) Systems and Matched Filter
1

Intuitive Guide to Principles of Communications
www.complextoreal.com


Matched filter is a theoretical frame work and not the name of a specific type of filter. It
is an ideal filter which processes a received signal to minimize the effect of noise. Hence,
it maximizes the signal to noise ratio (SNR) of the filtered signal.

Lets start by assuming that for a given transmitted signal shape, such as the two square
pulses shown below, noise will be added while they are traveling through the medium
and also at the receiver, Now we want to filter the received signal to reduce additive
noise. What kind of filter should we use? FIR, IIR, Butterworth, what? This is the
question we will answer.

It so happens that an optimum filter does exist for each signal shape transmitted and it is
a function only of the transmitted pulse shape. Because of its direct relationship to the
transmitted pulse shape, it is called a matched filter. At the receiver it filters out the noise
such that the received signal SNR is maximum possible and the probability of BER is
lowest.

(a) Transmitted signal, square pulses
(b) At the receiver, distorted by a lot of noise.
(c) After the receive filter, looking a good deal more like the transmitted signal.
Fig. 1 Received vs. transmitted signal shapes.

Let the input to the receiver be ( ) ( ) s t n t + where is the transmitted signal and the
additive noise. Now filter this signal so that additive noise is removed before
demodulation. Since there is no way to remove noise without also removing some of the
signal, we want a filter that does this in an optimum way.
( ) s t ( ) n t
( ) n t

Actually what we want is to maximize is the ratio of the received signal to the noise at
some particular time, t
d
. And not to keep you in any suspense, t
d
is at the end of the

Copyright 2003 Charan Langton www.complextoreal.com
Tutorial 20 LTI and Matched Filter

2
symbol. Its like waiting until you have all the information before you make a decision,
which is a good policy for signal processing as well!

The received signal is designated as ( ) ( )
o o
s t n t + . The subscript is there to denote that this
is an output signal. The ratio of the signal power to the noise power, or the SNR at the
precise instant of the decision time t
d
is equal to

2
2
( )
( )
o d
o
s t
SNR
n t
= = (1)

We wish to maximize this number. This is like any other maximization process. We are
trying to increase the mean square value of the numerator which is the signal power, and
minimize the undesired quantity in the denominator, which is the noise power.

All communications text books spend a whole page or two going through the solution to
this maximization, as it is elegant and impressive. The derivation uses Schwartz
inequality and other complex terms, is a lot of work but leads to a very simple
conclusion. And the conclusion is:

The signal to noise ratio is maximized when the impulse response of that filter is
exactly a reversed and time delayed copy of the transmitted signal.

(a) Transmitted signal, (b) the required impulse response of the receive filter.
Fig. 2 The impulse response of the receive filter should be a reversed copy of the transmitted signal
in order to maximize the SNR.

If the signal shape transmitted is 2(a) then the filter that maximizes the SNR has an
impulse response shown in this figure 2(b).

What if the transmitted signal is square pulse, then what? Nothing. The signal is
symmetrical. The impulse response is the same as itself. This is true of all symmetrical
pulses and since nearly all practical pulses are symmetrical, this makes our task easy.

We learned in the BER section that the bit error rate experienced by a signal during
demodulation is a function of the signal to noise ratio. A matched filter which maximizes
SNR will provide the lowest possible BER. So obviously finding the optimum filter is a
worth-while thing to do. The two most common matched filters are Integrate and Dump
and Root-raised Cosine.


3
Linear Time Invariant System

To examine what a matched filter does, we need to visit the concept of a Linear Time
Invariant (LTI) system.

Fig. 3 The LTI transforms an incoming signal based on a fixed rule.

The middle box in Fig. 3 is the LTI box which does something to a signal entering it. In
this case, we show a single pulse entering this box. The LTI box takes this pulse, and
does something to it. All LTIs are different and all do different something depending on
what you need them to do. All signal processing boxes including filters, amplifiers, etc.
can be characterized at LTIs.

The LTI transformation (Eq. 4 and Eq. 5) is the single most important relationship in the
field of communications since it gives us a way to relate input and output signals.

The following is model of an LTI.

( ) ( ) ( )
sin
x t h t
goe LTI did something to it
y t

x(t) goes in, the LTI does h(t) to it and out comes y(t) transformed version of x(t). Each
pulse goes through the same transformation no matter where in time. Collective response
to many pulses spread in time is determined by the Principle of Superposition and forms
the transformation of signal through the LTI.

Lets take a look at what happens to the following signal as it goes through this LTI box.

Fig. 4 Putting a signal composed of many pulses through a LTI


4

The first pulse to go through the box is at t =0 time. Its response is the same as the
impulse response (because it has a positive unit amplitude!) of the LTI box.

Fig 5a Response to the impulse at time 0

Now at the next instant, t =1, the second pulse of magnitude 2 enters the LTI. The
response is the same, just twice the magnitude and shifted in time so it starts at t =1.

Fig 5b Response to the impulse at time 1

And similarly the response to the third pulse of magnitude -1 is same as for 1 but
reversed and of course starts at t =2, as shown in Fig. 5c

Fig 5c Response to the impulse at time 2

The signal consists of just 3 pulses. The impulse response of LTI was 4 pulses long for
each of the 3 input pulses. The complete response is the sum of all these individual
impulse responses. Sadly the output bears no likeness to the input.


5

Fig 5d All responses added up to give the combined response to three pulses.

We can write the output at any one moment in time as

[ ] [0] [ ] [1] [ 1] [2] [ 2]
[3] [ 3] [4] [ 4] [5] [ 5]
y n x h n x h n x h n
x h n x h n x h n
= + +
+ +
+
+

You notice that in this expression, for each nth value of the response, the input signal x[n]
is staying put but the impulse response h[n-k] is moving backwards. We can see this by
writing a couple of terms.

[0] [0] [0] [1] [ 1] [2] [ 2]
[3] [ 3] [4] [ 4]
y x h x h x h
x h x h
= + +
+ [5] [ 5] x h +

Since impulse response of the LTI has no 4
th
or 5
th
terms, the last two products are
crossed out. The next two terms of the output are

[1] [0] [1] [1] [0] [2] [ 1]
[3] [ 2] [4] [ 3] [5] [ 4]
y x h x h x h
x h x h x h
= + +
+ +
+

[2] [0] [2] [1] [1] [2] [0]
[3] [ 1] [4] [ 2] [5] [ 3]
y x h x h x h
x h x h x h
= + +
+ +
+

The sum can be written out as the familiar but dreaded convolution equation.

[ ] [ ] [ ]
k
y n x k h n k
=
=

(2)

(The negative index k in h(n-k) indicates the h(t) shifts in the opposite direction of x(t).)
We interpret the response of the LTI as the summation of the weighted and shifted input
pulses. It is mathematically a fairly simple concept, although convolution is hard concept
to understand intuitively.


6
The LTI transformation equation is given in time domain by

( ) ( )* ( ) y t h t x t = (3)

where h(t) is the impulse response of the LTI and y(t) the output/

In frequency domain, we write

( ) ( ) ( ) Y f H f X f = (4)

The spectral density of the output signal is given by

2
( ) ( ) ( )
y x
S f S f H f = (5)

and another important relationship (both of these given here without proof) about the
power in the received signal after it has come out of the LTI is

2
2
( ) ( ) ( )
x
E y t H f S f df
(6)

Convolution and autocorrelation

The autocorrelation function (ACF) is similar in concept to convolution, but ACF is a lot
easier to grasp.

[ ] [ ] [ ]
k
y n x k x k n
=
=

(7)

The ACF, as convolution also shifts one of the sequences in time. The only difference
between ACF and convolution is that we do not reverse one of the sequences as we do in
convolution. Note that the index k here is positive as opposed to Eq. 3. Otherwise the
two concepts are identical.

Now the math: You can skip this derivation, if you are feeling math-shy. This part is not
essential to understanding what a matched filter is. It is an amazingly simple concept but
is wrapped in the mathematical clothes so thick that we can lose sight of what it means.

Statement
The matched filter maximizes the SNR of the filtered signal and has an impulse response
that is reverse time-shifted version of the input signal.

Proof

A signal s(t) is transmitted.. It goes through a filter with an impulse response of h(t), It
picks up noise n(t) along the way. The output signal is


7

( ) ( ) ( )
o
s t s t n t +

Lets say that the transformation occurs inside a LTI which we will call the matched filter
The requirement is that the output signal be as large as possible compared to noise at
decision time t =t
d
. We can write the expression for average SNR as

2
2
( )
( )
o
o
d
s t
SNR
n t
= = (8)

Since the matched filter is characterized by its impulse response, we want to find an
expression for h(t) that will maximize this SNR.

We will do the maximization in three parts, first we will write the expression for the
numerator, the signal, then secondly we will examine the denominator, noise. Thirdly we
will use Schwarz inequality to extract useful information about the SNR and make a
statement about the optimality of the filter.

The signal

We model the matched filter as a linear time-invariant system which has an impulse
response of h(t) and frequency response of H(f). The output out of the LTI is

{ }
1
( ) ( )* ( ) ( ) ( )
o
s t h t s t F H f S f
= = (9)

Lets substitute the expression for the inverse Fourier transform

2
( ) ( )* ( ) ( ) ( )
j ft
o
s t h t s t S f H f e df
= =
(10)

At time t
d
, this equation is written as

2
( ) ( ) ( )
d
j f t
o d
s t S f H f e df
(11)

The output power is proportional to the square of this signal amplitude, so we can write

2
2
2
( ) ( ) ( )
j t
o
S s t H S e dt
= =

(12)

This is the numerator of the expression for SNR that we are trying to maximize.

The Noise


8

Now lets turn our attention to noise, the denominator term that we need to minimize.
The noise is assumed to be Gaussian and has a noise spectral density of N
0
/2. The noise is
zero mean, and has a flat power spectrum wherein all frequencies are represented and
have equal amount of power.

We state that the spectral density of the input noise is equal to

0
( ) /
2
ni
N
S watts Hz = (13)

The factor of 2 is there because this is a two-sided spectral density. The assumption of
N
0
/2 gives us correct intuitive results. Over a bandwidth B, we get total power which is
equal to N
0
B as shown below.

0
0
0
0
( )
2 ( )
2
2
B
n n
B
B
n
B
P S f df
S f df
N
df N B
=
=
= =
(14)

The spectral density of noise-out (no) after the LTI is equal to (using Eq. 5)

2
0
( ) ( ) ( )
n ni
S f S f H f = (15)

Now substitute value for the noise-in (ni) spectral density (Eq. 13), we get

2
0
( ) ( )
2
no
N
S H = f (16)

Employing Eq. 14, the noise power is equal to

2
2
0
( ) ( )
2
o
N
n t H f df
=

(17)

The fact that noise is assumed to be Gaussian is an important thing to recognize. A
matched filter thus derived works as expected only under those conditions. If we have a
fading link or other non-Gaussian phenomena then this filter would not be optimum.

Maximizing the ratio


9
Now we put together both of numerator and denominator is their full glory. Eq. 12 for the
numerator and Eq. 17 for the denominator.

2
2
2
2
2
0
( ) ( )
( )
( )
( )
2
d
o
o
j f t
d
H f S f e df
s t
SNR
N
n t
H f df
= = =

= (18)

Here is where Schwarz inequality comes in handy. The inequality states the equivalence
of H(f) and S(f) by the relationship

2
2 2
1 2 1 2
( ) ( ) ( ) ( ) f x f x dx f x df f x dx

(10)

but this hold true only if and only if

*
1 2
( ) ( ) f x k f x = (20)

where
*
2
( ) f x is complex conjugate of the function
1
( ) f x . This condition is really
important, because we will use it to formulate the final conclusion about the matched
filter.

Now set
1
( ) ( ) f x H f =

And

2
2
( ) ( )
d
j ft
f x S f e

=

we write using Eq, 16,

2
2
2
( ) ( ) ( ) ( )
d
j f t
H f S f e df H f df S f df
(21)

Now substitute the right side of the above expression into the numerator of Eq. 15, we get

2
2
2
2
2
2
0
( )
( ) ( )
( )
( )
( )
2
d
o
o
j f t
d
H f df
H f S f e df
s t
N
n t
H f df
2
2
0
( )
( )
2
S f df
N
H f df
(22)


10

Apply above to Eq. 17, we can say that the SNR must be less than or equal to

2
2
2
0
( )
2
( )
( )
o d
o
s t
S f df
N
n t

(23)

So the maximum SNR possible is equal to

2
2
2
0
( )
2
( )
( )
o d
o Max
s t
S f df
N
n t
=

(24)

Now going back to the condition of Eq. 19, this is true only if the following condition is
true.

*
( ) ( ) H f k S f = (25)

Alternately we can say that this means that the magnitude characteristic of the filter
should be such that

( ) ( ) H f S f = (26)

This says it all. The filter should be such that it only filters those frequencies in the
signal. How simple is that!

This can be written in time domain as

*
( ) ( )
d
h t k s t t = (27)

The k is an arbitrary constant and can be used to represent gains through the channel.

This is the final result of the whole thing. The impulse response of the LTI or the
matched filter is equal to a time reversed copy of the input signal. And this particular
value gives the maximum SNR out of the filter.

Properties of Matched Filter

The output of the LTI system which we will now call the Matched Filter output (as
opposed to its impulse response) is the signal s
o
(t
d
).

Note that Rayleighs Theorem tells us that the energy of a signal is equal to


11
2
( ) E S f
df

And the output of the matched filter can be written as

2
0
1
( ) ( )
2
d
s t S d E
=

= (28)

The mean square noise output out of the matched filter is

2 0
0
( )
2
N
n t =

Which gives us

0
2E
SNR
N
= (29)

No matter what the shape of the transmitted signal, the output of the Matched filter is the
energy of the signal at the instant of the decision.

The output of the Matched filter, which is called matched filer response, is a signal
which may not look like the transmitted signal but has a value of E at the moment of
decision.

When we talk about matched filters, be aware if it is the impulse response or the output
response, because they are two different things. The output response may look nothing
like the impulse response as we shall see soon. In the above math, the impulse response is
h(t) and the output response is y(t).

The Correlation receiver

The process of deriving the matched Filter does not help much in developing an
understanding of what the matched Filter is all about. There is n alternate way to look at
the Matched Filter process that is much more intuitive. This method is called the
Correlation receiver.

For matched filter now we know that h(t) is equal to

*
( ) ( )
d
h t ks t t = (30)

The output response of the matched filter is given by the convolution equation


12
( ) ( ) ( ) y t x h t d
(31)

for a single bit period t
d
, we can write this equation again by changing the limits of
integration,

0
( ) ( ) ( )
d
t
y t x h t d =
(32)

now lets take the impulse response h(t) (Eq. 26) and shift it by so we can plug in its
shifted value in the above equation. Correspondingly the right sight shifts by + .
Remember that the time shift in h(t) has to be opposite to the time shift in s(t) for a
matched filter.

*
( ) (
d
h t k s t t ) + = (33)

Substituting into Eq. 29 into Eq. 28, we have

[ ]
0 0
2
( ) ( ) ( ) ( )
d
t
d
k
y t s n s t t d
N
= + +

Now set t =t
d
, and we can do that because it we are only interested in the results of this
operation at the precise moment when t =t
d
or the end of the bit time when we will make
a decision about what was sent.

[ ]
0 0
2
( ) ( ) ( ) ( )
d
t
k
y t s n s d
N
= +
(34)

The above expression is exactly the same as doing cross-correlation between the input
signal which includes noise and the original signal without noise. This leads us to an
alternative way to implement a matched filter, called the correlator.

The correlation receiver is exactly the same thing mathematically as the matched filter
but much easier to comprehend and much easier to implement in hardware.

Examples

In Fig 6, we show the matched filter output for a square pulse shown. Impulse response
of a square pulse is itself. Note what the filter output value at the end of the symbol
period which is 5. The value is maximum at t =5, the end of the symbol period. The
receiver now makes a decision based on this value and declares the bit to be 1. Then the
second half of the output is not needed, so the value (of 5) is discarded from memory, the
registers are flushed and the process starts anew with the next symbol.


13
0
1
2
3
4
5
6
0 5 10
Symbol Matched filter output
Do not calculate
these values
Make decision based on th
max value.

Fig. 6 A square pulse and its matched filter output.

Whats interesting is that the output of the matched filter is maximum at t =5, the last
sample of the symbol. Each point in between amounts to an integration of the symbol up
to that point. So although we did convolution, what we got is an integration, exactly the
same thing we would get if we do cross-correlation between the transmitted symbol( the
square pulse) and the saved reversed replica of this symbol (also a square pulse).

Fig. 7 Transmitted and received pulses

Fig. 7 shows a square pulse signal and its noisy received counterpart. Fig. 8 shows the
timing pulse and the output of the filter. This result gives us the integrate and dump filter.
We integrate the received symbol up to the last sample, dump this value to the slicer and
the reset to start anew with the next symbol. Depending on the polarity of the next
symbol, the next integration may be negative or positive as shown below. A slicer then
samples the signal after the dump and makes a bit decision based on a simple logic.


14

Fig 8 Integrate and Dump output
Integrate received amplitudes over the symbol width which is 12 samples in this case. At the end of 12
samples, a trigger tells the integrator to stop, dump the values to the slicer and start again. So what we see
are little triangles, which is the integration of the square pulse. The perturbation comes from the statistical
variation of the added noise value.

Situations where Integrate and Dump filter in not optimum

A triangular pulse shape

0
2
4
6
8
10
12
14
16
18
20
0 5 10

Fig. 9 A triangular pulse and its matched filter output

In this case, we have a triangular pulse. Note once again, that the out put of the matched
filter is the integration of the triangular pulse with maximum occurring at t =5, a value of
19 which is the area under the pulse. Now once again, we could integrate and make a
decision based on the max value but it turns out the integrate and dump filter is not
optimum for this case.


15

Fig. 10 The effect of noise on the average signal to noise ratio is different for these two pulses.

For a square pulse the signal to noise ratio is constant whereas for the second pulse, the
SNR is low at the edges and is maximum in the center for the very obvious reason that
the signal is maximum at the center. The integrate and dump does a fine job for the
square pulse but it would exaggerate the low SNR region over the high SNR, so is not an
optimum filter for anything other than square pulse.

What is optimum for this type of pulse? Since the pulse is symmetrical, the impulse
response is the same as the pulse, the output is just the correlation of the two and the best
value occurs at the center, hence for this type of signal, we just do a center slice and make
a decision based on the center value.

For most pulses other than square, a center sampling is optimum for decoding purposes,
because the SNR is maximum at this point.

Implementing Matched filter as a correlator

Lets take the pulse of Fig. 11. The transmitted pulse and the received pulse with noise
are shown.

1
2 2 2
0
Time
0 1 2 3 4 5
\

Fig 11 Transmitted and received pulse

Transmitted pulse 1 2 2 2 0
Received pulse 0.8 2.2 2.2 1.8 0.3

x x x x x 2 2 2 1 0
0.3
1.8
0.3
2.2 1.8
0.3
0.3
2.2 1.8 2.2
0.3 1.8 2.2 2.2 0.8
Expected value =13
Received value =12

Fig 12 Matched filter in process (note how similar it looks to a correlator,)


16

In the receiver we have stored the expected pulse shape in reverse as shown in the boxes
next to the multipliers. Now come the samples of the received signal. This is a little
confusing because they come in a reverse order, that is 0.8, 2.2, 2.2, 1.8, 0.3., with .8 at
first, then 2.2, then 2.2 and so on.

The samples are being fed into the registers. Nothing is happening until all five registers
are full. Now perhaps you can see the need for reversing the stored impulse response in
the receiver. This is so that the received samples line up properly and we will get a
maximum response. Now we do a multiplication of the expected with the received. The
expected value is 13. the received value in this case is 13.2.

.8 x 1 +2 x 2.2 +2 x 2.2 +1.8 x 2 +0 x .3 =13.2

Based on the rule that if this number is greater than 0, then decide in favor of pulse 1 if
less than 1 then zero decide in favor of other pulse.

The output of a matched filer is the energy of the signal

An optimum receiver for AWGN channel does two functions. One is the matched filter
which is often the correlator and the other a detector which makes decision about the
pulse. The correlator correlates the received signal with two stored pulses, s
0
and s
1

(assuming this is a binary signal). The result is two different output values and these two
values are then fed to the detector which makes a decision about what was sent. The
output values at time t are given by the expressions

0 0
0
1 1
0
( ) ( ) ( )
( ) ( ) ( )
t
t
y t y s d
y t y s d

=
=

(35)

Example.

A signal y(t) is received. It contains the transmitted signal and noise component. Now we
do the two correlations.

0
( ) ( ) ( ) 0
d
y t s t n t t t = +

0 0
0
0 0 0
0 0
0
( ) ( )
( ) ( ) ( ) ( )
d
d d
t
t t
y y t s t dt
s t s t dt n t s t dt
E n
=
= +
= +

(36)

Recognize that the first term is the energy of the signal.


17
2
d
E A t = (37)

which is kind of a autocorrelation of the signal at time t
d
. The same thing with the other
stored pulse.

1 1
0
0 1
0
( ) ( )
( ) ( )
d
d
t
t
y y t s t dt
s t s t dt
=
=
1
0
0
( ) ( )
d
t
n t s t dt
n
+
=

(38)

where noise components are given by

0 0
0
1 1
0
( ) ( )
( ) ( )
d
d
t
t
n n t s t
n n t s t
=
=

dt
dt
(39)

The first term is zero because the signals are assumed to be orthogonal. So we get two
results, y
1
and y
0
, one is the energy plus noise and the other just noise, so assuming that
noise power is less than signal energy, we would make the decision in favor of the higher
level.

A root raised cosine-like pulse

Symbol
-5
-4
-3
-2
-1
0
1
2
3
0 5 10 15
Symbol
-30
-20
-10
0
10
20
30
40
0 10 20 30

Fig. 13 A RRC-like symbol and its matched filter output.

I have made up a pulse to approximate the root raised cosine pulse. As before the output
of the matched filter is maximum in the middle but fluctuates all over.

Figure 12 below shows a raised cosine pulse shaped signal. The noise is added to this
signal and the received signal is then filtered using a raised cosine filter, which happens


18
to be a matched filter to itself. The resulting pulse is then sampled in the middle of the
symbol to decode the bit.

(a) Transmitted raised cosine pulses

(b) Received signal with noise

c) filtered with a root-raised cosine filter

(d) Timing pulses at symbol rate the sample the signal at the center of the symbol

(e) Center sample, if it is positive decide a 1 bit, if negative decide a 0 bit.
Fig. 14 Root Raised cosine as a matched-pair filter set.


19
A root raised cosine pair by virtue of its symmetrical shape is a matched filter pair but in
addition it also suppresses intersymbol interference so we kill two birds with one stone
and that is why this is such a popular filter set.

The optimum filtering again for this type of pulse is the center value as it is for all other
non square pulses.

Other non-symmetric pulse

Symbol
0
2
4
6
8
10
0 5 10 15
Symbol
0
20
40
60
80
100
120
140
160
180
200
0 10 20 30

Fig. 15 An unsymmetrical received signal and its matched filter output.

In the above case, I have a unsymmetrical pulse. Its maximum also occurs at the end of
the symbol period, so recognizing that the SNR is not constant throughout the symbol, we
would use center sample to decide the fate of the bit.

Summary

1. A matched filter is kind of an averaging filter.
2. It maximizes the signal to noise ratio of a received signal.
3. A matched filter minimizes the BER of the received signal. The BER equations,
assume that a matched filter is used to decode the signal.
4. Its time domain shape it the time-reversed replica of the transmitted shape, For
symmetrical shapes and most are, the same filter can be used to filter the received
signal.
5. Two symmetrical filters make a matched pair, one to transmit, the other to
receive.
6. Root raised cosine is a special such pair, because it also reduces inter-symbol
interference.
7. Integrate and dump filters is a matched filter only for square pulses, usually used
in baseband systems that have no rf carrier.
8. Other channel filtering can make a matched filter non-ideal but is still optimum
due to need to build practical system.
9. A matched filter provides us a framework for creating equalized filters ( to be
covered in a subsequent tutorial.)


20

Charan Langton
I can be contacted at
[email protected]
Other tutorials at
www.complextoreal.com


MFT

Uploaded by

Copyright:

Available Formats

MFT

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MFT

Uploaded by

Copyright:

Available Formats

Linear Time Invariant (LTI) Systems and Matched Filter

You might also like