Matched Filter Receiver [2, 3]

Since we assume symbol-by-symbol detection, therefore, for the following discussion,

only the signal in a symbol interval (say in the first symbol interval) is considered. The
corresponding transmitted signal is
𝑠(𝑡) = 𝑎0 𝑔(𝑡), 0≤𝑡≤𝑇 (1)
For the following discussion, 𝑎0 is not considered as it is acting as amplitude scaling
(Take 𝑎0 = 1).
Received signal is
𝑟(𝑡) = 𝑠(𝑡) + 𝑤(𝑡), 0≤𝑡≤𝑇
Putting (1), we have
𝑟(𝑡) = 𝑔(𝑡) + 𝑤(𝑡), 0≤𝑡≤𝑇 (2)
𝑁0
Where, 𝑤(𝑡) is a zero-mean white Gaussian noise process with PSD = .
2
Output of the receive filter is
𝑦(𝑡) = 𝑟(𝑡) ∗ ℎ(𝑡) (3)
𝑡

𝑦(𝑡) = ∫ 𝑟(𝜏) ℎ(𝑡 − 𝜏)𝑑𝜏

0
Putting (2) in (3), we have
𝑡 𝑡
𝑦(𝑡) = ∫0 𝑔(𝜏) ℎ(𝑡 − 𝜏)𝑑𝜏 + ∫0 𝑤(𝜏) ℎ(𝑡 − 𝜏)𝑑𝜏
= 𝑔0 (𝑡) + 𝑛(𝑡) (4)
Where 𝑔0 (𝑡) and 𝑛(𝑡) are produced by the signal and noise components of the input
r(t), respectively.
Let the output of the receive filter be sampled at 𝑡 = 𝑡0 , then
𝑡0 𝑡0

𝑦(𝑡0 ) = ∫ 𝑔(𝜏) ℎ(𝑡0 − 𝜏)𝑑𝜏 + ∫ 𝑤(𝜏) ℎ(𝑡0 − 𝜏)𝑑𝜏

0 0
= 𝑔0 (𝑡0 ) + 𝑛(𝑡0 ) (5)
 The goal is to select the receive filter impulse response ℎ(𝑡) such that SNR at the
receive filter output is maximize (or equivalently probability error in detection is
minimized)
Instantaneous Output Signal Power 𝑔𝑜2 (𝑡0 )
(SNR)𝑜 = = (6)
Average Output Noise Power 𝐸[𝑛2 (𝑡0 )]
As 𝑤(𝑡) is zero mean, therefore
𝐸[𝑛(𝑡0 )] = 0
𝑡 2
𝐸[𝑛2 (𝑡0 )] = 𝐸 [(∫0 0 𝑤(𝜏) ℎ(𝑡0 − 𝜏)𝑑𝜏) ]
𝑡 𝑡
= 𝐸 {∫0 0 𝜔(𝜏) ℎ(𝑡0 − 𝜏)𝑑𝜏 ∫0 0 𝜔(𝑡) ℎ(𝑡0 − 𝑡)𝑑𝑡}
𝑡 𝑡
= ∫0 0 ∫0 0 𝐸[𝑤(𝜏) 𝑤(𝑡)] ℎ(𝑡0 − 𝜏)ℎ(𝑡0 − 𝑡) 𝑑𝜏 𝑑𝑡
𝑁0 𝑡0 0 𝑡
= ∫0 ∫0 𝛿(𝜏 − 𝑡) ℎ(𝑡0 − 𝜏)ℎ(𝑡0 − 𝑡) 𝑑𝜏 𝑑𝑡
2
 𝑁0 𝑡
= 0 2
⏟0 ℎ (𝑡0 − 𝑡) 𝑑𝑡
∫ (7)
2
Energy in the IR ℎ(𝑡)
Putting (7) in (6), we have
𝑡 2 𝑡 2
[∫0 0 𝑔(𝑡) ℎ(𝑡0 −𝑡)𝑑𝑡] [∫0 0 𝑔(𝑡0 −𝑡) ℎ(𝑡)𝑑𝑡]
(SNR)𝑜 = 𝑁0 𝑡0 2 = 𝑁0 𝑡0 2 (8)
∫ ℎ (𝑡0 −𝑡) 𝑑𝑡
2 0 ⏟2 ∫0 ℎ (𝑡0 −𝑡) 𝑑𝑡
Constant

Keeping denominator to be constant, (SNR)𝑜 is maximized by maximizing the

numerator.
Using Cauchy-Schwarz inequality
∞ 2 ∞ ∞

[ ∫ 𝑔1 (𝑡) 𝑔2 (𝑡) 𝑑𝑡] ≤ ∫ 𝑔12 (𝑡) 𝑑𝑡 ∫ 𝑔22 (𝑡) 𝑑𝑡

−∞ −∞ −∞
Where the equality holds when 𝑔1 (𝑡) = 𝐶𝑔2 (𝑡) for any arbitrary constant C. Using
Cauchy-Schwarz inequality for the denominator of (8), we have
𝑡 2 𝑡 𝑡
[∫0 0 𝑔(𝑡0 − 𝑡) ℎ(𝑡)𝑑𝑡] ≤ ∫0 0 ℎ2 (𝑡)𝑑𝑡 ∫0 0 𝑔2 (𝑡0 − 𝑡)𝑑𝑡 (9)
Eqn. (9) is maximum when
ℎ(𝑡) = 𝐶𝑔(𝑡0 − 𝑡) (10)
i.e., impulse response of the optimum receive filter is matched to the signal input.
Putting (10) in (8), we have
𝑡 2
𝐶 2 [∫0 0 𝑔2 (𝑡0 −𝑡)𝑑𝑡] 𝐸2 2𝐸
(SNR)𝑜 = 𝑁0 𝑡 = 𝑁0 = (11)
2
𝐶 2 ∫0 0 𝑔2 (𝑡)𝑑𝑡 2
𝐸 𝑁0

 Note that the output SNR from the matched filter depends on the energy E of the
waveform 𝑔(𝑡) but not on the detailed characteristics (shape etc.) of 𝑔(𝑡). This is
another interesting property of the matched filter.
o For a given E, shape of 𝑔(𝑡) is chosen which is best suited for transmission.
The output of the optimum receive filter depends on
 Pulse shape 𝑔(𝑡)
𝑁0
 Noise PSD
2
 The Sampling instant 𝑡0
Moreover, it is independent of the gain at the receiver C, since the same gain would
apply to both the signal and noise without affecting the (SNR)𝑜 . Therefore, impulse
response of the receive filter is
𝒉(𝒕) = 𝒈(𝒕𝟎 − 𝒕) (12)
i.e., response 𝑔(𝑡0 − 𝑡) is the transmit pulse time reversed 𝑔(−𝑡) and delayed by 𝑡0 .
 What is the best sampling instant 𝑡0 ?
 𝑔(𝑡)

𝑔(−𝑡)

𝑔(𝑡𝑜 − 𝑡)

𝑡𝑜 < 𝑇

(a) 𝑡𝑜
𝑔(𝑡𝑜 − 𝑡)

𝑡𝑜 = 𝑇

(b)
𝑡𝑜
𝑔(𝑡𝑜 − 𝑡)
𝑡𝑜 > 𝑇

(c) 𝑡𝑜

Fig. 3.1 Optimum choice for sampling instant

Three cases are possible as shown above:
Case (a): 𝑡𝑜 < 𝑇, results in a noncausal impulse response, which is unrealizable.
Case (b) and (c) results in physically realizable system. However, case (c) delays the
decision making instant 𝑡𝑜 unnecessarily.
Case (b): 𝑡𝑜 = 𝑇 gives the minimum delay for decision making using a realizable filter
and hence it is taken as the best sampling instant and the corresponding impulse
response of the matched filter is
 𝒉(𝒕) = 𝒈(𝑻 − 𝒕)

Both g(t) and h(t) have duration = T seconds. From (4), the signal component at the
receive filter output is the convolution of g(t) and h(t) and given by
𝑡

𝑔𝑜 (𝑡) = ∫ 𝑔(𝜏) ℎ(𝑡 − 𝜏)𝑑𝜏

0
The duration of the convolution will be 2T. The maximum value of 𝑔𝑜 (𝑡) will occur at
t = T. This also justifies the use of 𝑡𝑜 = 𝑇 value as the best sampling instant.

𝑔(𝑡)

ℎ(𝑡)

ℎ(−𝑡)

ℎ(𝑇 − 𝑡)
Example:
Frequency Domain Interpretation of the Matched Filter [1, 4.2]
ℎ(𝑡) = 𝑔(𝑇 − 𝑡)
Taking the Fourier transform of ℎ(𝑡), we get
𝐻(𝑓) = 𝐺(−𝑓) 𝑒 −𝑗2𝜋𝑓𝑇 (13)
And the signal component at the receive filter output is
𝑡
𝑔𝑜 (𝑡) = ∫0 𝑔(𝜏) ℎ(𝑡 − 𝜏)𝑑𝜏 = 𝑔(𝑡) ∗ ℎ(𝑡) (14)
Fourier transform of (14) is
𝐺𝑜 (𝑓) = 𝐺(𝑓) 𝐻(𝑓) (15)
Putting (13) in (15), we have
𝐺𝑜 (𝑓) = |𝐺(𝑓)|2 𝑒 −𝑗2𝜋𝑓𝑇 (16)
The inverse Fourier transform of (16) is
∞
𝑔𝑜 (𝑡) = ∫−∞ 𝐺𝑜 (𝑓) 𝑒 𝑗2𝜋𝑓𝑡 𝑑𝑓
∞
= ∫−∞ |𝐺(𝑓)|2 𝑒 −𝑗2𝜋𝑓𝑇 𝑒 𝑗2𝜋𝑓𝑡 𝑑𝑓 (17)
Sampling the Filter output at t = T, we get
∞ 𝑇
𝑔𝑜 (𝑇) = ∫−∞ |𝐺(𝑓)|2 𝑑𝑓 = ∫0 𝑔(𝑡)2 𝑑𝑡 = 𝐸 , from Parseval's relation (18)
The noise at the output of the matched filter has a power-spectral density
𝑁0
𝑆𝑛 (𝑓) = |𝐻(𝑓)|2 (19)
2

Total noise power at the output of the matched filter is

∞ 𝑁0 ∞ 𝑁0 ∞ 𝑁0
𝑁 = ∫−∞ 𝑆𝑛 (𝑓) 𝑑𝑓 =
2
∫−∞ |𝐻(𝑓)|2 𝑑𝑓 = 2
∫−∞ |𝐺(𝑓)|2 𝑑𝑓 = 2
𝐸 (20)

The signal power is 𝑃 = [𝑔𝑜 (𝑇)]2 = 𝐸 2 (21)

The output SNR is
𝑃 𝐸2 2𝐸
(𝑆𝑁𝑅)𝑜 = = 𝑁0 = (22)
𝑁 𝐸 𝑁0
2

which agrees with the result given by Eqn. (11).

1. Simon Haykin, Communication Systems, 4th edition, John Wiley & Sons, 2004.
2. B P Lathi and Z Ding, Modem Digital and Analog Communication Systems, 4th edition, Oxford
University Press, 2010. Section 10.1.22.
3. J G Proakis and M Salehi, Communication Systems Engineering, 2nd edition., Pearson Education, 2006,
pp 376-379.