EEE543 DCS – LECTURE 2 PART I

Pulse Code
Modulation – Part I

1
OUTLINE
 Sampling
 Quantization

 Pulse Code Modulation (PCM)

 Companding in PCM Systems

2
 SAMPLING
Consider a band-limited signal g t with bandwidth 𝐵 spectrum
defined as

𝐺 𝑓 = 0, 𝑓 > 𝐵
 If we sample this signal at a uniform rate 𝑅 > 2𝐵 then the signal can
be exactly reconstructed (without any error) from the discrete
samples
 The minimum frequency for successful recovery is:
𝑓𝑠 ≥ 2𝐵
 The sampling rate 𝑓𝑠 Hz means we take 𝑓𝑠 uniform samples per
second at an interval 𝑇𝑠 .
1
 The parameter 𝑇𝑠 is called the sampling period and the reciprocal 𝑇
𝑠
denoted 𝑓𝑠 is called the sampling rate or frequency
 We obtain a sequence of samples by multiplying 𝑔(𝑡) by an impulse
train 𝛿𝑇𝑠 𝑡 .
 The result is a sequence {g 𝑛𝑇𝑠 }, (denoted gത(𝑡) in Figure 2.1) where
𝑛 is an integer, which uniquely define g 𝑡 . 3
SAMPLING CONT’D
 The figure below show the original signal and the sampled version
of the signal together with their Fourier transforms

Figure 2.1

4
SAMPLING THEOREM CONT’D
 Let the samples of the signal g 𝑡 be denoted by gത 𝑡 and represented
as follows
∞

gത 𝑡 = g 𝑡 . 𝛿𝑇 𝑠 𝑡 = ෍ g 𝑛𝑇𝑠 𝛿(𝑡 − 𝑛𝑇𝑠 )

𝑛=−∞
 Taking the Fourier Transform of the above signal we have
∞
1 𝑛
𝐺ҧ 𝑓 = 𝐺 𝑓 ⋆ ෍ 𝛿 𝑓−
𝑇𝑠 𝑇𝑠
𝑛=−∞
∞
1
= ෍ 𝐺 𝑓 − 𝑛𝑓𝑠
𝑇𝑠
𝑛=−∞
 Question: Can g 𝑡 be recovered from gത 𝑡 without distortion?
 From figure 2.1 we not that the recovery is possible as long as the
spectrum of 𝐺ҧ 𝑓 does not overlap. This requires the following
condition to hold
1 1
𝑓𝑠 > 2B ⇒ 𝑇𝑠 < ⇒ 𝑇𝑠 <
𝑓𝑠 2𝐵
5
SAMPLING THEOREM
 The minimum sampling rate 𝒇𝒔 = 𝟐𝑩 required to recover 𝐠 𝒕 from
its samples 𝐠ത 𝐭 is called the Nyquist rate for 𝐠 𝒕 and the
𝟏
corresponding sampling interval 𝑻𝒔 = 𝟐𝑩 is called the Nyquist
interval for the low-pass signal 𝐠 𝒕
 The sampling theorem states that if the signal g 𝑡 is sampled at
1
intervals 𝑇𝑠 ≤ , then it is possible to reconstruct the original
2𝐵
signal from the samples.
1
 If 𝑇𝑠 > 2𝐵, then the spectrum of the sampled signal will overlap and
reconstruction of the original signal is not possible – this is called
undersampling
 The effect of undersampling is called aliasing (aliasing refers to
reconstruction of a possible signal which is not the original signal)

 If the spectrum 𝐺 𝑓 has no impulse (or its derivatives) at the

highest frequency B, then the overlap is still zero as long as the
sampling rate is greater than or equal to the Nyquist rate, that is,
6
𝑓𝑠 ≥ 2B
SIGNAL RECONSTRUCTION
 The process of reconstructing a continuous time signal 𝐠 𝒕 from its
samples is also known as interpolation.
 This processes takes place after passing the sampled signal
through a lowpass filter of bandwidth B Hz and gain 𝑇𝑠 .
 The filter response has the transfer function

 For the ideal reconstruction case, the inverse Fourier Transform of

the H(f) is
ℎ 𝑡 = 2𝐵𝑇𝑠 . 𝑠𝑖𝑛𝑐 2𝜋𝐵𝑡
 Assuming Nyquist sampling rate, 2𝐵𝑇𝑠 = 1, then

ℎ 𝑡 = 𝑠𝑖𝑛𝑐 2𝜋𝐵𝑡
7
SIGNAL RECONSTRUCTION CONT’D
 When the sampled signal gത t is applied to the filter defined above
the output is g t
 The figure below shows the properties of the filter and the
reconstructed signal

8
Figure 2.2
SIGNAL RECONSTRUCTION CONT’D
 The k-th sample of the signal gത t = g 𝑘𝑇𝑠 . 𝛿 𝑡 − 𝑘𝑇𝑠 , the output of
the filter is: g 𝑘𝑇𝑠 . ℎ 𝑡 − 𝑘𝑇𝑠 .
 The filter output signal g 𝑡 in Figure 2.2 c) given by
∞

g t = ෍ g 𝑘𝑇𝑠 . ℎ 𝑡 − 𝑘𝑇𝑠
𝑘=−∞
∞

= ෍ g 𝑘𝑇𝑠 . 𝑠𝑖𝑛𝑐 2𝜋𝐵 𝑡 − 𝑘𝑇𝑠

𝑘=−∞
∞

= ෍ g 𝑘𝑇𝑠 . 𝑠𝑖𝑛𝑐 2𝜋𝐵𝑡 − 𝑘𝜋

𝑘=−∞
 Where we have used 2𝐵𝑇𝑠 = 1

9
EXAMPLE

10
PRACTICAL SIGNAL RECONSTRUCTION
 The idea reconstruction low-pass filter ℎ 𝑡 is practically
unrealizable, This can be equivalently seen from the infinitely long
nature of the sinc reconstruction pulse
 For practical application of signal reconstruction, we need to
implement realizable signal reconstruction systems from the
uniform signal samples.
 A solution would be a reconstruction pulse 𝑝 𝑡 which must be easy
to generate which is shown in the figure below:

11

Figure 2.3
PRACTICAL SIGNAL RECONSTRUCTION
 We must first use the nonideal interpolation pulse p(t) to analyze
the accuracy of the reconstructed signal. First we define the signal
reconstructed using p(t) as

 Its relation to the original analog signal g(t) is as follows:

 The Fourier transform of g෤ 𝑡 is given by

12
PRACTICAL SIGNAL RECONSTRUCTION
 We note that that the reconstructed signal g(t) using pulse p(t)
consists of multiple replicas of G(f) shifted to the frequency center
nfs and filtered by P(f)
 To fully recover g(t), further filtering of g
෤ 𝑡 is required using filters
referred to as equalizers
 Let the equalizer transfer function be defined as E(f) distortionless
reconstruction requires that

 Where

 Then the replicas of 𝐺 𝑓 − 𝑛𝑓𝑠 will be removed for all the

summation except for 𝑛 = 0
13
PRACTICAL SIGNAL RECONSTRUCTION
 Additionally, distortionless reconstruction requires that
𝐸 𝑓 . 𝑃 𝑓 = 𝑇𝑠 𝑓𝑜𝑟 |𝑓| < 𝐵
 The equalizer filter E(f) must be low-pass in nature to stop all

frequency content above 𝑓𝑠 − 𝐵 𝐻𝑧 B Hz, and it should be the

inverse of P(f) within the signal bandwidth of B Hz.
 The diagram below demonstrates a practical signal reconstruction
system utilizing such an equalizer.

14
Figure 2.5
PRACTICAL SIGNAL RECONSTRUCTION
 An example of p(t) is the simple interpolating pulse generator that
generates short (zero-order hold) pulses given by:

 This is a gate pulse of unit height with pulse duration 𝑇𝑝

 The reconstruction will first generate

 The transfer function of filter 𝑃(𝑓) is the Fourier transform of

ς 𝑡Τ𝑇𝑝 shifted by 0.5 𝑇𝑝
𝑃 𝑓 = 𝑇𝑝 sinc 𝜋𝑓𝑇𝑝 𝑒 −𝑗𝜋𝑓𝑇𝑝

15
PRACTICAL SIGNAL RECONSTRUCTION
 The equalizer frequency response should satisfy

 In the design of the equalizer we must ascertain that the equalizer

passband response is realizable, we first add another time delay in
the reconstruction such that

 We choose short 𝑇𝑝 so that the passband gain of 𝐸 𝑓 is well defined

this implies that

16
PRACTICAL SIGNAL RECONSTRUCTION
 This means that the equalizer E(f) does not need to achieve infinite
gain. Otherwise the equalizer would become unrealizable – this
requires that
1
𝑇𝑝 <
𝐵

 In practice, 𝑇𝑝 can be chosen very small, to yield the following

equalizer passband response:

 The short rectangular pulses make very little distortions and the
design of the equalizer either unnecessary or very simple
 We can use the first-order-hold filter, which results in a linear
interpolation instead of the staircase interpolation
17
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 If a signal is sampled at the Nyquist rate fs = 2B Hz, the spectrum
𝐺ҧ 𝑓 consists of repetitions of G(f) without any gap between successive
cycles,
 To recover g 𝑡 from gത 𝑡 , we need to pass the sampled signal gത 𝑡
through an ideal low-pass filter (dotted lines)

Figure 2.6
 Such a filter is unrealizable in practice; It can be closely approximated
only with infinite time delay in the response.
 This means that we can recover the signal g(t) from its samples with
infinite time delay. 18
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 Solution to the above problem is to sample the signal at a rate
higher than the Nyquist rate gത 𝑓𝑠 > 2𝐵 𝑜𝑟 𝜔𝑠 > 4𝜋𝐵 and the output
is as follows

Figure 2.7
 In this case a low-pass filter with a gradual cutoff characteristic can
be employed to recover 𝐺 𝑓
 The filter gain is required to be zero beyond the first cycle, also such
a filter is unrealizable it can only be better approximated with a
smaller time delay.
Note: Even if the sampling rate is higher than the Nyquist rate, it is
19
impossible in practice to recover a band-limited signal g(t) exactly
from its samples
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 The premise of the sampling theorem for signal reconstruction from
its samples was that the signal g(t) is band-limited, at most: All
practical signals are time-limited; that is, they are of finite duration
or width
 Practically, this is not the case: A time-limited signal cannot be
band-limited, and vice versa (but a signal can be simultaneously
non-time-limited and non-band-limited).

 All practical signals, which are necessarily time-limited, are non-

band-limited: they have infinite bandwidth, and the spectrum
𝐺ҧ 𝑓 consists of overlapping cycles of 𝐺 𝑓 repeating every 𝑓𝑠 Hz (the
sampling frequency)

 The spectral overlap is unavoidable, regardless of the sampling rate.

 Sampling at a higher rate reduces but does not eliminate

20
overlapping between repeating spectral cycles
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 Illustrations

a) Spectrum of a
practical signal g(t)

b) Spectrum of
sampled g (t)

c) Reconstructed
signal spectrum
21

Figure 2.8
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 Due to the overlapping tails, 𝐺ҧ 𝑓 no longer has complete
information about G(f), recover g(t) exactly from the sampled signal
g(t) is not possible
 After passing the signal thought the low pass filter low-pass filter of
cutoff frequency 𝑓𝑠ൗ2 the output is a distorted version of 𝐺 𝑓 denoted
𝐺𝑎 𝑓
 The distortion is caused by:
 The reappearance of this tail inverted or folded back onto the spectrum
𝑓
 The loss of the tail of G(f) beyond 𝑓𝑠 > 𝑠ൗ2 Hz

 Folding Frequency
1
 The frequency 𝑓𝑠 = 2𝑇 (figure 2.8 c)) is the folding frequency.
 At this frequency the lost tail folds back onto itself and its components
appear at the reconstructed signal.
𝑓𝑠
 the components of frequencies above 2
reappear as components of
𝑓𝑠
frequencies below
2
 This is known as spectral folding or aliasing (shaded region in figure
2.8 b, and c) 22
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 In aliasing we lose the frequency components above the folding
frequency also these components reappear as lower frequency
components.
 Aliasing also destroys the integrity of the frequency components
𝑓
below the folding frequency 2𝑠
 To avoid aliasing, we need to suppress the higher frequency
𝑓
components beyond the folding frequency 2𝑠 before sampling the
signal g(t)
 The signal is passed through a low-pass filter with cut-off frequency
𝑓𝑠
called anti-aliasing filter as shown below
2

Figure 2.9 23
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 An antialiasing filter essentially band-limits the signal g(t) to 𝑓2𝑠 Hz
so that we lose only the frequency components beyond the folding
frequency.
 The suppressed components cannot reappear and corrupt the
components of frequencies below the folding frequency 𝑓2𝑠
𝑓𝑠
 The reconstructed signal spectrum 𝐺𝑎𝑎 𝑓 = 𝐺 𝑓 for 𝑓 < 2
 Emphasis: The antialiasing operation must be performed before the
signal is sampled.

24

Figure 2.10
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 An antialiasing filter also helps to reduce noise
 Noise, generally, has a wideband spectrum, it cause the noise
components outside the desired signal band to appear in the signal
band
 Antialiasing suppresses the entire noise spectrum beyond frequency
𝑓𝑠
.
2
 The antialiasing filter, being an ideal filter, is unrealizable. In
practice we use a steep-cutoff filter, which leaves a sharply
𝑓
attenuated residual spectrum beyond the folding frequency 2𝑠.

25
PRACTICAL ISSUES WITH SIGNAL SAMPLING & RECONSTRUCTION
 Sampling Forces Non-Band-Limited Signals to Appear Band-
Limited
 Figure 2.8b shows the spectrum of a signal g ത 𝑡 consists of
overlapping cycles of 𝐺 𝑓 . This means that gത t are sub-Nyquist
samples of g t
 We may also view the spectrum in Fig. 2.8b as the spectrum 𝐺𝑎 𝑓
(Fig. 2.8c ), repeating periodically every 𝑓𝑠 Hz without overlap
 The spectrum 𝐺𝑎 𝑓 is band-limited to 𝑓𝑠ൗ2 Hz
 These (sub-Nyquist) samples of g t are actually the Nyquist
samples for signal g a t
 In conclusion, sampling a non-band-limited signal g(t) at a rate 𝑓𝑠 Hz
makes the samples appear to be the Nyquist samples of some signal
g a t , band-limited to 𝑓𝑠ൗ2 Hz.
 In other words, sampling makes a non-band-limited signal appear to
be a band-limited signal g a t , with bandwidth 𝑓𝑠ൗ2 Hz

26
MAXIMUM INFORMATION RATE
 A knowledge of the maximum rate at which information can be
transmitted over a channel of bandwidth B Hz is of fundamental
importance in digital communication
 Theorem: A maximum of 2B independent pieces of information per
second can be transmitted, error free, over a noiseless channel of
bandwidth B Hz.
 The sampling theorem shows that a low-pass signal of bandwidth B
Hz can be fully recovered from samples uniformly taken at the rate
of 2B samples per second
 We need to show that any sequence of independent data at the rate
of 2B Hz can come from uniform samples of a low-pass signal with
bandwidth B
 For an independent data samples denoted as {g 𝑛 } with rate 2B
samples per second there always exists a (not necessarily band-
limited) signal g(t) such that
1
g 𝑡 = g 𝑛𝑇𝑠 𝑇𝑠 =
2𝐵
27
MAXIMUM INFORMATION RATE
 Previously, we have seen that sub-Nyquist sampling of a signal g(t)
generates samples that can be equally well obtained by Nyquist
sampling of a band-limited signal ga(t)
 Sampling g(t) and ga(t) at the rate of 2B Hz will generate the same
independent information sequence {gn}: that is
1
g 𝑛 = g 𝑛𝑇𝑠 = g 𝑎 𝑛𝑇𝑠 𝑇𝑠 =
2𝐵
 the sampling theorem also states that a low-pass signal g 𝑎 𝑡 with
bandwidth B can be reconstructed from its uniform samples

 Assuming no noise, this signal can be transmitted over a

distortionless channel of bandwidth B Hz, error free.
 In conclusion the maximum information rate is 2B samples per Hz

 Note: In practice, channel noise is unavoidable, and consequently, 28

this rate will cause some detection errors

NONIDEAL PRACTICAL SAMPLING ANALYSIS
 So far, we have considered an ideal uniform impulse sampling pulse
train to precisely extract the signal value g 𝑘𝑇𝑠 at the precise
instant of 𝑡 = 𝑘𝑇𝑠 . Not possible in practice
 Practical samplers take a snapshot of duration 𝑇𝑝 at time 𝑡 = 𝑘𝑇𝑠 of
the signal g(𝑡)
 The sampler would then generate a sample value at 𝑡 = 𝑘𝑇𝑠 by
averaging the values of signal g(t) over the window 𝑇𝑝

 This averaging may be weighted by a device-dependent averaging

function q(t) such that

29
NONIDEAL PRACTICAL SAMPLING ANALYSIS
 In the practical sampler the uniform weighing is defined as

 Generating pulse train of q(t) we have:

 The product of g(𝑡) and 𝑞𝑇𝑠 (𝑡) is

gො 𝑡 = g 𝑡 . 𝑞𝑇𝑠 𝑡 = ෍ g 𝑛𝑇𝑠 . 𝑞 𝑡 − 𝑛𝑇𝑠
𝑛
 The signal gො 𝑡 is applied to an averaging filter defined as ℎ𝑎 𝑡 with
transfer function 𝐻𝑎 𝑓

30
NONIDEAL PRACTICAL SAMPLING ANALYSIS
 The output of the averaging filter is given by
g1 𝑡 = gො 𝑡 ∗ ℎ𝑎 𝑡
 The practical sampler generate a sampled signal g
ු (t) by sampling
the averaging filter output g1 𝑘𝑇𝑠
gු t = ෍ g1 𝑘𝑇𝑠 . 𝛿 𝑡 − 𝑘𝑇𝑠
𝑘
 The figure in the next slide illustrate the above described practical
sampler .

31
NONIDEAL PRACTICAL SAMPLING ANALYSIS
 Practical Sampler

32

Figure 2.11
NONIDEAL PRACTICAL SAMPLING ANALYSIS
 We examine the frequency domain relationships to analyze the
distortion generated by practical samplers.
 Note that 𝑞𝑇𝑠 (𝑡) is periodic with Fourier series

 The averaging filter output signal is

 The sampled average filter output is given by

33
NONIDEAL PRACTICAL SAMPLING ANALYSIS
 We can define frequency responses

 Then we can write

 For the low-pass signal 𝐺 𝑓 with bandwidth B Hz, applying an ideal

low-pass (interpolation) filter will generate a distorted signal:

𝐹𝑜 𝑓 𝐺 𝑓
Where

 The use of a practical reconstruction pulse p(t) will generate

additional distortion.
34
NONIDEAL PRACTICAL SAMPLING ANALYSIS
 If we reconstruct g (t) by using the practical samples to generate:

 the relationship between the spectra of the reconstruction and the

original message G(f) as
𝐺ෘ 𝑓 = 𝑃 𝑓 ෍ 𝐹𝑛 𝑓 . 𝐺 𝑓 + 𝑛𝑓𝑠

 Since G(f) has bandwidth B Hz, we will need to design a new

equalizer with transfer function E(f) such that the reconstruction is
distortionless within the bandwidth B

 This single equalizer can compensate the non-ideal sampling effect

in 𝐹𝑛 𝑓 and non-ideal reconstruction effect in 𝑃 𝑓
 The equalizer design is made practically possible because both
35
distortions are known in advance.