EE 121: Digital Communications

April 22, 2008

Matched Filters

Introduction
Starting from this lecture, we focus on how to communicate over LTI channels. This lecture
we focus on matched filtering. This strategy aims to mitigate ISI by filtering the received
signal.

The Matched Filter

To motivate this linear processor, let us start with a very simple sequential communication
scheme: we send independent information bits at different time instants, but interleave L − 1
zeros between every information bit. In other words, we send information only once in L
time samples. Of course, such a scheme entails a very low communication rate but it has
the major advantage of ensuring that there is no ISI at the receiver. Denoting the voltages
chosen by the sequential communication scheme by d[n], n ≥ 1, the transmit voltages are
½
d[n] if ` = 0
x[(n − 1)L + ` + 1] = , n ≥ 1. (1)
0 if ` = 1 . . . L − 1
Now the received voltages are
y[(n − 1)L + ` + 1] = h` d[n] + w[(n − 1)L + ` + 1], n ≥ 1, ` = 0, . . . , L − 1. (2)
So, to recover the information bits that generated the nth data voltage d[n] we only need to
focus on L receive voltages:
y[(n − 1)L + ` + 1], ` = 0, . . . , L − 1. (3)
Each of these voltages is simply the (scaled) voltage of interest (d[n]) corrupted by additive
white Gaussian noise. This situation is reminiscent of repetition coding over the AWGN
channel.
We have derived the vector ML rule in an earlier lecture to be a nearest-neighbor rule
where the distance is now a Euclidean one. If d[n] is equal to the k th possible voltage level
(among the 2B choices, if we are sending B information bits in each of the transmit voltages),
say vk , then the squared Euclidean distance between the received voltages and the vector
 
h0 vk
 h1 vk 
 
 ..  (4)
 . 
hL−1 vk
is
L−1
X
(y[(n − 1)L + `] − h` vk )2 (5)
`=0

1
which can be expanded out as
ÃL−1 ! ÃL−1 ! ÃL−1 !
X X X
h2` vk2 + (y[(n − 1)L + `])2 − 2vk h` y[(n − 1)L + `] . (6)
`=0 `=0 `=0

In comparing the Euclidean distances for different choices of k (among the 2B different
choices), we only need to keep track of the sufficient statistic:
L−1
X
h` y[(n − 1)L + `], (7)
`=0

the weighted sum of the received voltages. Indeed, when all the filter coefficients are equal,
this sufficient statistic is simply the sum of the received voltages and featured in the ML rule
for repetition coding. This operation of taking a weighted sum of the received voltages, with
the weights in proportion to the channel coefficients is known as match filtering; the idea is
that the filter weights are matched to those of the channel. The output of the matched filter
L−1
X
def
y MF [n] = h` y[(n − 1)L + `] (8)
`=0
ÃL−1 ! ÃL−1 !
X X
= h2l d[n] + h` w[(n − 1)L + `] . (9)
`=0 `=0

Thus we have a plain AWGN channel between the voltage of interest d[n] and the output
of the matched filter y MF [n]. We could now employ the nearest-neighbor rule on the output
of the matched filter (MF) to detect the information bits that generated the voltage d[n].
The operating SNR of this AWGN channel is
ÃL−1 !
def
X
SNRMF = h2` SNR (10)
`=0

where we have denoted the ratio of average transmit energy (say, E) and the noise variance
σ 2 ) by SNR. The effect of the matched filter is to boost the operating SNR level by carefully
harnessing the delayed copies of the transmit symbols. This SNR level directly impacts √ the
reliability of communication; indeed, with the voltage d[n] constrained to be within ± E
Volts and there are 2B possible levels, the average probability of error is (from an earlier
lecture) µ ¶ µ√ ¶
1 SNRMF
2 − B−1 Q . (11)
2 2B − 1
Note that the operation (8) is called matched filtering because it can actually be rewritten
as a convolution operation followed by sampling every L time instants:
−(L−1)
X
MF
y [n] = h−` y[(n − 1)L − `] = (h̃ ∗ y)[(n − 1)L]
`=0

2
where
h̃` = h−` ,
i.e. the channel in reverse.
Throughout this discussion, we have not had to worry about ISI: the transmission in-
terleaved with zeros so that there is no interference between consecutive transmit voltages.
Now to use sequential communication at all each time sample (and not interleave with L − 1
zeros) while maintaining the simplicity of the matched filter receiver, the key conceptual
idea is the following: even if there were interference but it was negligible compared to the
additive Gaussian noise then for practical purposes it is as if it does not exist. In this situa-
tion, the MF would be quite a sensible strategy to pursue at the receiver. With sequential
communication, the average energy in the interference is
ÃL−1 !
X
E h2` (12)
`=1

where we have denoted the average transmit energy by E. Whenever this is small compared
to the noise energy σ 2 , i.e., the operating SNR
1
SNR ¿ PL−1 , (13)
`=1 h2`

it is reasonable to suspect that the interference can be ignored without much loss of perfor-
mance. In such instances, the matched filter is quite likely to yield near-optimal performance.
Concretely, suppose the transmit voltages are generated based on sequential communi-
cation (no interleaving with zeros this time). Now the matched filter operates at every time
sample and the corresponding output is
L−1
X
MF def
y [m] = h` y[m + `] (14)
`=0
L−1
XX L−1
= ˜
h` h`˜x[m + ` − `] (15)
˜
`=0 `=0
ÃL−1 !   ÃL−1 !
X L−1
X X
= h2` x[m] +  ˜ +
h` h`˜x[m + ` − `] h` w[m + `] . (16)
`=0 ˜
`,`=0:`6 =`˜ `=0

We observe that the channel between the transmitted voltage x[m] and the received voltage
y[m] is not quite AWGN: the main difference is the presence of the interference
L−1
X
def ˜
I[m] = h` h`˜x[m + ` − `], (17)
˜
`,`=0:`6 =`˜

which is discrete valued and not Gaussian. So the ML rule for detecting x[m] from y[m]
does not directly follow from our analysis of the AWGN channel. While in certain instances

3
(which are explored in the homework), we see that the ML rule still comes down to the
simple nearest-neighbor rule, this is not necessarily always the case.
As we have seen earlier in this course, it is useful for the communication engineer to have
a simple rule of thumb to quickly measure the reliability of communication. In the AWGN
channel, the operating SNR serves as a natural criterion; indeed, it is directly related to the
reliability of communication. While the SNR is not so directly related to the reliability of
communication in additive but non-Gaussian channels, it is very easy to measure and still
serves as a surrogate to measure the reliability of communication. In the context of the
additive noise plus interference channel in Equation (16), the signal energy is
ÃL−1 !2
X
h2` E, (18)
`=0

the interference energy is  2

L
X L−1
X
 h` h`˜ E, (19)
`6ˆ=0, `=−L
ˆ ˜
`,`=0:`− ˜ `ˆ
`=

and the noise energy is ÃL−1 !

X
h2` σ2. (20)
`=0

So the signal to interference plus noise ratio (SINR) at the output of the matched filter is
³P ´2
L−1 2
def
h
`=0 ` E
SINRMF = P ³P ´2 ³P ´ (21)
L L−1 L−1 2 2
`6ˆ=0,`=−L
ˆ ˜
`,`=0:`−
h h
˜ `ˆ ` `
`= ˜ E + h
`=0 ` σ
c1 SNR
= , (22)
c2 SNR + 1
where
E
SNR = σ2
(23)
L−1
X
def
c1 = h2` (24)
`=0
 2
L
X L−1
X
def 1 
c2 = h` h`˜ (25)
c1
`6ˆ=0, `=−L
ˆ ˜
`,`=0:`− ˜ `ˆ
`=

From Equation (22) we see clearly how the SINR at the output of the matched filter depends
on the operating SNR. This function is schematically illustated in Figure 1.
Two different regimes are of keen interest:
• Low SNR: For small values of SNR, the SINR is close to linear in SNR. In other
words, a doubling of the operating SNR also doubles the SINRMF . Mathematically,

4
 MF
SINR
6
º

¢
¢
¢
¢
¢ -
¢ SNR

Figure 1: SINR at the output of the matched filter plotted as a function of the operating
SNR.

this regime happens when the interference c2 SNR ¿ 1. In this regime, the interference
is much smaller than the background noise and it makes fine sense to just ignore the
interference. The channel is almost like an AWGN one and hence the linear relationship
between SNR and SINRMF .

• High SNR: For large values of SNR the SINRMF is almost constant and hardly changes.
Mathematically, this regime kicks in when the interference c2 SNR À 1. In this regime,
the interference is much larger than noise and we pay a steep price by just ignoring its
presence. The interference level is directly proportional to the transmit signal energy
and hence the SINRMF is not sensitive to increases in SNR in this regime.

To ameliorate the deleterious presence of interference in the high SNR regime, one needs
to explicitly deal with this by jointly detecting the sequence of symbols instead of detecting
symbol-by-symbol. This will be the topic of the next lecture.