Digital Communications

– Lecture Notes –

Dr. Majid Safari

Email:[Link]@[Link]
Chapter 2
Analogue to Digital Conversion (ADC)

Abstract This chapter discusses sampling, quantisation and pulse code modulation PCM, which are the required steps to
convert a continuous-time input signal to a digital binary signal. We focus on different types of sampling, including ideal,
natural and flat topped sampling. In addition, linear and non-linear quantisation are discussed. In PCM, both uniformly-
distributed and non-uniformly-distributed signals are considered. Detailed references to the individual sections of the
textbook that are relevant to the material presented here are indicated at the title of each section below. For example, H2.3
would refer to section section 3 of chapter 2 of Haykin’s textbook [Haykin(2014)].

2.1 Introduction

Consider a simplified digital communication system as depicted in Fig. 2.1 with an analog input signal. The transmitter
consists of an anti-aliasing filter and an analog-to-digital converter (ADC) and the receiver consists of a digital-to-analog
converter (DAC) and an amplifier. Please note that we will mainly focus on the ADC part, but bear in mind that the
processes at the ADC are reversed at the DAC.

Fig. 2.1 Block diagram of an ADC model discussed in this chapter.

The anti-aliasing filter is a low-pass filter used to block a high-frequency signal to ensure that the input signal satisfies
the ADC requirements. In this course, the ADC is modeled by a concatenation of a sampler, a quantiser and a PCM
encoder. The input signal is first sampled so that we have a discrete-time signal. Recall that a digital signal is both discrete
time and discrete amplitude (i.e., it is made up of a countable alphabet). Hence, a quantizer is needed to round the values
of the sampled-signal to a predefined set of values. This discrete values can be then represented by binary values encoded
by a PCM encoder. Note that this type of communication system is one of the earliest digital systems used for baseband
transmission such as landline telephone systems or generally for encoding and recording a digital audio signal.

21
Digital Communications Dr. Majid Safari

2.2 Sampling H6.2,H6.3

Fig. 2.2 shows a basic operation of a sampler, which can be thought of as a switch that only conducts over short reoccurring
periods. Mathematically speaking, the sampling process is equivalent to multiplying an analog signal (g(t)) with a periodic
pulse train (s(t)). Generally, the pulse train has the pulse width t and the pulse interval or sampling period Ts . The output
signal gs (t) is merely a pulse train with the same sampling period and pulse width but with an amplitude proportional to
the amplitude of the analog signal over the sampling pulse periods.

Fig. 2.2 A first view on a sampling process.

Definition 2.1 (Nyquist sampling rate). Nyquist sampling theorem states that a band-limited signal of finite energy
which has no frequency component on or higher than W Hertz is completely described by the samples of the signal taken
at least at the Nyquist rate of fN = 2W , i.e., fs = 1/Ts fN = 2W .

Considering the sampling period Ts or sampling rate fs = 1/Ts , we may have three sampling cases including oversampled
( fs > 2W ), critically sampled ( fs = 2W ) or undersampled ( fs < 2W ) signals defined by the Nyquist sampling theorem. To
illustrate this in Fig. 2.3, we consider the highest frequency sinusoid component of a signal with at W Hz. If fs = 8W as
shown in Fig. 2.3(a) the signal is oversampled (more than the minimum required number of samples). Critically sampled
case is obtained if fs = 2W as defined by the equality in the Nyquist sampling and undersampling occurs if fs < 2W , see
Figs. 2.3(b) and (c).
Considering Fig. 2.3, if we interpolate the oversampled or critically sampled values to get a sinusoid, the sinusoid with
correct frequency of W Hz should be obtained. However, this is not necessarily the case with the undersampled case where
the samples may be confused to be taken from a lower frequency sinusoid. Such effect is called aliasing. Next, we discuss
three types of sampling models, i.e., ideal (Fig. 2.4), natural (Fig. 2.5) and flat topped samplings (Fig. 2.6).

(a) Oversampled (b) Critically sampled (c) Undersampled

Fig. 2.3 Types of sampling based on the sampling time Ts .

22
Digital Communications Dr. Majid Safari

2.2.1 Ideal sampling

Fig. 2.4 shows the process of ideal sampling in time and frequency. The ideal sampling uses a train of impulses (rather
than pulses with a finite width) to sample the analog signal and this assumption makes the model mathematically tractable
(easily manipulated algebraically). Considering an arbitrary band-limited analog signal g(t) with the Fourier transform
G( f ) and the highest frequency component at fH Hz (see Fig. 2.4(a)), the ideally sampled signal is demonstrated in both
time and frequency in Fig. 2.4(c).

(a) Original signal

(b) Pulse train

(c) Sampled signal

Fig. 2.4 Ideal sampling.

Noting the properties of an impulse, the sampled signal in time domain constitute a train of impulses with amplitudes
equal to the amplitude of the analog signal in the sampling instances. Recall that the Fourier transform of a train of
impulses separated by Ts is also a train of impulses separated by 1/Ts = fs . This is shown in Fig. 2.4(b). Recall our
discussion on the constant time-bandwidth product for this matter. Moreover, From Table 1.2, we know that the Fourier
transform of the product of two functions is the convolution of their Fourier transforms. Noting the sifting property of
the impulse, the convolution of the impulse train in the frequency domain with the signal results in the replication of
G( f ) every fs Hz as shown in Fig. 2.4(c). Applying an ideal low-pass filter with cut-off frequency fH , we can eliminate
the copies of the signal generated by the sampling process at higher frequency and therefore reconstruct the signal. This
can be done without aliasing distortions if the copies of the signal spectrum in frequency domain do not overlap. From
Fig. 2.4, we can deduce that to avoid such overlap (i.e., aliasing), we should have fs 2 fH which confirms teh Nyquist
sampling theorem.
Let us now analyse the ideal sampling process (graphically illustrated in Fig. 2.4) using mathematical formulation to
get some insight to the reconstruction process. The sampled signal gs (t) in Fig. 2.4(c) can be denotes as:
•
gs (t) = Â g(nTs )d (t nTs ). (2.1)
n= •

Using Tables 1.1 and Table 1.2, the sampled signal in frequency domain can be expressed as:

23
Digital Communications Dr. Majid Safari
•
Gs ( f ) = f s  G( f m fs )
m= •
•
= fs G( f ) + fs  G( f m fs ),
m= •
m6=0

If fs 2 fH , one can show that Gs ( f ) = fs G( f ), | f | < fH and thus no aliasing occurs.

We will now describe how to reconstruct the original signal g(t) from the sampled signal gs (t). Taking the Fourier
transform of (2.1) and using the linearity of the transform, the spectrum of the sampled signal can also be written as:
• • ✓ ◆ ✓ ◆
n jpn f
Gs ( f ) = Â g(nTs ) exp ( j2pn f Ts ) = Â g exp .
n= • n= • 2 fH fH

where the second equality is written assuming that the signal is critically sampled at the Nyquist rate fs = 2 fH . as
indicated before, G( f ) can be easily obtained by means of an ideal low-pass filter (LPF) within | f | < fH , which yields:
8 • ⇣ ⌘ ⇣ ⌘
< 1 Â g n exp jpn f
2 fH 2 fH fH , | f | < fH
G( f ) = n= •
:0, otherwise

Now, taking the inverse Fourier transform of the equation above, we have:

Z• ZfH • ✓ ◆ ✓ ◆
1 n jpn f
g(t) = G( f ) exp( j2p f t)d f =
2 fH Â g 2 fH exp
fH
exp( j2p f t)d f
n= •
• fH

• ✓ ◆ ZfH  ✓ ◆ • ✓ ◆
n 1 n n
= Â g exp j2p f t df = Â g sinc(2 fH t n),
n= • 2 fH 2 fH 2 fH n= • 2 fH
fH

where the last term shows that the signal can be perfectly reconstructed by interpolation of the ideal samples in time using
the sinc function as the interpolation function.

2.2.2 Natural sampling

Unlike, ideal sampling, the natural sampling uses a train of pulses with a finite width to sample the analog signal; that
is the output of the natural sampler would be the product of the pulse train and the analog signal in time. Therefore, the
pulses carry the amplitude of the signal over their width t. For the natural sampling, the Fourier transform pair of the
pulse train is shown in Fig. 2.5(a). Observe that the pulse train in time domain is no longer a fixed-amplitude impulse train
in the frequency domain, but it is an impulse train at a rate of fs with amplitudes dictated by the sinc function tsinc(t f ).
To explain this, we can think of the time domain pulse train with finite width as the convolution of a train of impulses
(at the same rate as the pulse train) and a rectangular pulse of width t. Recalling that convolution in time translate into
multiplication in frequency, the frequency response of the pulse train can be described as the multiplication of an impulse
train (Fourier transform of the impulse train in time) and the sinc function tsinc(t f ).
As shown in Fig. 2.5(b), the sampling of the analog signal using the finite-width pulse train can be described in
frequency domain as the convolution of the variable amplitude impulse train with the spectrum of the analog signal.
Therefore, similar to ideal sampling, the spectrum of the sampled signal with the natural sampling is the replication of
many spectra of the input signal, but it decays according to the sinc function. Therefore, The reconstruction process is
the same as that of the ideal sampling using a low-pass filter and passing through the copy of the spectrum at the zero
frequency.

24
Digital Communications Dr. Majid Safari

(a) Pulse train

(b) Sampled signal

Fig. 2.5 Natural sampling.

2.2.3 Flat topped sampling

The flat topped sampling is an alternative sampling technique which allows for relatively simpler practical implementation,
for example, by a hold circuit. Roughly speaking, the flat topped sampling is carried out by sampling a value at a sampling
instance and hold it for the finite width of of the pulse. Therefore, unlike the natural sampling, the amplitude of the
sampling pulse is constant rather than following the analog signal over the pulse duration. The model for the flat topped
sampling can be mathematically analysed as an extension of the ideal sampling as shown in Figs. 2.6. Recall the sifting
property of the delta function; that is the convolution of an impulse and a function gives that function centered at the
location of the impulse and multiplied with the amplitude of the impulse. Hence, as shown in Figs. 2.6, if the ideally
sampled signal comprising a train of impulses with different amplitudes is convolved with a rectangular pulse of unit
amplitude and width t, the pulse will appear repeatedly at the sampling points with a constant amplitude equal to the
sampling value, and this is equivalent to flat-topped sampling. Note that the flat topped sampling is also known as a pulse
amplitude modulation (PAM).
In frequency domain, as shown in Figs. 2.6, The spectrum of the ideal sampling (i.e., the repeated copies of the
analog signal spectrum with the same amplitudes) is multiplied with a wide sinc function (i.e., sinc(t f )) that distorts the
amplitude of the spectrum of the copies of the signal as can be seen in Figs. 2.6(c). Unlike the ideal and natural samplings,
the reconstruction filter described by H( f ) = 1/sinc(t f ), 8 fH  f  fH , requires an equalizer to compensate for
the distortion caused by the constant-amplitude pulses while filtering the additional copies of the signal. The frequency
response of such reconstruction filter is shown in Fig. 2.6(d).

2.2.4 Distortion caused by aliasing

Generally, aliasing refers to the phenomenon where a high frequency component of a signal takes the identity of lower
frequency component. As already mentioned, this can happen when the copies of the spectrum of the signal generated
through the sampling process interfere with each other due to undersampling. In practice, the signals are not strictly
band-limited and thus there will be always some level of undersampling which cause aliasing effect. This effect cannot be
removed after sampling using basic reconstruction procedures described before and remains as a distortion. To minimise
such aliasing effect, typically we:

25
Digital Communications Dr. Majid Safari

1. use an anti-aliasing filter (a low-pass filter with cut-off frequency at fH ) before the sampling to attenuate unnecessary
high frequency components, and

2. sampling at a rate slightly higher than Nyquist rate. This helps to simplify the design of reconstruction filter which
must have perfectly sharp tails otherwise.
There is a rule-of-thumb for the latter method, which is

fs 2.2 fH .

Note that although the application of a anti-aliasing filter may pre-distort the signal by attenuating the high frequency
components, it avoids the larger distortions that would occur otherwise due to undersampling. To measure the distortions
caused by Aliasing in the signal, we define signal-to-distortion ratio (SDR) as the ratio of un-aliased to aliased power:
fR
s /2
G( f )d f
0
SDR = R• , (2.2)
G( f )d f
fs /2

(a) Ideally sampled signal

(b) Applying the constant amplitude pulses

(c) Flat topped sampled signal

(d) Frequency response of the reconstruction filter

Fig. 2.6 Flat topped sampling.

26
Digital Communications Dr. Majid Safari

where G( f ) the double-sided PSD. If the effect of a non-ideal reconstruction filter H( f ) is considered, then the SDR can
be approximated as:
R•
G( f ) |H( f )|2 d f
0
SDR = R• . (2.3)
G( f fs ) |H( f )|2 d f
0

This measure can be used to calculate the suitable sampling rate or cut-off frequency for anti-aliasing filter. Alternatively,
the sampling rate can be determined based on the highest frequency with significant spectral amplitude as specified by
some other standards (e.g., 3.4 kHz for voice in Europe).

2.2.5 Pulse modulation and time division multiplexing (TDM)

The sampling theorem allows for the representation of a continuous-time signal with separated samples which can be
transmitted over the channel using separated pulses. Therefore, the idle time between separated pulses can be used for
the transmission of other independent signals. This type of transmission which allows multiple signals share the same
channel at different times is called time-division multiplexing (TDM). Fig. 2.7 shows a representation of the TDM of two
signals where the switch ensures only one signal is sampled at each sampling time. The low-pass filter at the final stage
of TDM system minimises the frequency content of the multiplexed signal to their required transmission bandwidth. Note
that TDM can be done if the bandwidth of the channel can also support multiple signals as multiplexing several signals
in time expands the bandwidth of the multiplex signal accordingly (roughly with the same factor as the number of signals
assuming they have the same bandwidth). In addition, in order for reliable demultiplexing and detection of multiple signals
at the receiver side of a TDM system, robust synchronization is required.

Fig. 2.7 TDM of two independent signals.

The TDM system above is based on PAM where the sampled values are represented as the amplitude of transmitted
pulses and can be naturally generated by flat topped sampling. Besides PAM, the sampled values can be also modulated
as the width or position of pulses resulting in pulse width modulation (PWM) and pulse position modulation (PPM),
respectively (see Fig. 2.8). Note that PAM requires larger signal-to-noise ratio (SNR) as the noise directly affects the
amplitude of the signal where the data is modulated. However, as shown in Fig. 2.9, since the signal in PWM and PPM

27
Digital Communications Dr. Majid Safari

are not modulated on the amplitude, the noise can be mostly compensated by clipping the pulses at the receiver without
changing the width or position of the pulse that carry the data. This is particularly true when the pulses are closer to ideal
rectangular pulses. On the other hand, PAM requires smaller bandwidth than PWM or PPM, which are implemented based
on short pulses occupying a larger bandwidth. Also, PWM has an easy demodulation process through Integration over
symbol periods.

Fig. 2.8 Different pulse modulation techniques.

Fig. 2.9 Effect of noise on pulse width and position.

2.3 Quantisation H6.4

Quantisation is a matter of mapping an uncountable set of values of the sampled signal (v(t)) to a countable one, which
can be then typically represented by binary values. While Nyquist sampling may allow for perfect reconstruction of a
signal, signal distortion is inevitable in the process of quantisation. Therefore, determining a quantisation increment q is
important and it highly relates to the range of the amplitude of the sampled signal. We can be confident that the range is
finite as an electronic device cannot handle an infinite amplitude. As illustrated in Fig. 2.10, the sampled analog signal
is going through a quantiser which is defined by a set of finite values over the dynamic range of signal separated by the

28
Digital Communications Dr. Majid Safari

quantisation increment q. For a linear quantiser as illustrated in the figure, q is constant but this is not true for nonlinear
quantisers that will be discussed later. The quantiser rounds each sampled value to the nearest value of the quantiser and
outputs the quantised signal, v0 (t), which is a similar train of pulses but with quantised amplitudes.

(a) Sampled signal. (b) Quantiser (c) Quantised signal

Fig. 2.10 Quantisation Process.

After quantisation, the statistical pdf of v(t) changes from a continuous one to a discrete one for v0 (t) although the
general form of the pdf does not change significantly as illustrated in Fig. 2.11. Note that the pdf illustrated here is just an
example of the statistics of an arbitrary signal. Let’s define p(v) as a PDF of sampled values in v(t). The PDF of v0 (t) is
denoted by p(v0 ). Then, the probabilities for the discrete values of v0 (t) can be calculated as the area under the curve p(v)
over the interval [kq q/2, kq + q/2) for k 2 Z. In fact, this interval is the range of values which are rounded to kq by the
quantiser.

(a) PDF of v(t) (b) PDF of v0 (t)

Fig. 2.11 Distributions of original v(t) and quantised v0 (t) signals.

In order to design a good quantiser, we must be able to first charactrise the distortion induced by the quantiser. A metric
for distortion due to quantisation is referred to as average signal to quantisation noise ratio (SNq R), and it is defined as
the ratio of the mean square of the signal (signal power) and the mean square of quantisation error (noise power), i.e.,

v02
SNq R , , (2.4)
eq2

where the quantisation error is defined as eq (t) = ga (t) g(t), where g(t) and ga (t) are respectively the input signal
original value and its quantised value respectively as illustrated in Fig. 2.10. Here, we take a generalization of the input
of the quantiser as it can also be any continuous signal, not only the sampled signal discussed before. Note that in the
equation above the power of the quantised signal is sometimes is approximated by the power of the original signal, v2 , as
the quantisation process does not change the the power of the signal significantly.
To derive SNq R, a few assumptions are made for simplicity. We assume that:
1. equal increments between quantisation levels (linear quantisation), i.e., q is fixed,
2. zero mean signal, i.e. symmetrical PDF about 0V, and

29
Digital Communications Dr. Majid Safari

(a) Analogue (ga (t)) and quantised signal (g(t)).

(b) Quantisation error eq (t) = ga (t) g(t)

Fig. 2.12 Distortion in quantisation.

3. uniform signal pdf, i.e., all quantisation levels are equally likely.
The last assumption above asserts that:
M 1 ✓ ◆
1 qk
p(v ) = Â
0
d v0 , k is odd, M is even. (2.5)
k= M+1 M 2

The mean square (signal power) of the signal after quantisation is:
Z•
v02 = v02 p(v0 )dv0
•
2 3
Z Z ✓ ◆ Z ✓ ◆
2 4 02 ⇣ 0
(a) q⌘ 0 3q (M 1)q
= v d v dv + v d v
02 0
dv0 + · · · + v02 d v0 05
dv
M 2 2 2
0 0 0
2 ⇣ q ⌘2 ⇥ 2 ⇤
= 1 + 32 + · · · + (M 1) 2
M 2 
2 ⇣ q ⌘2 M(M 1)(M1 )
=
M 2 6
M2 1 2
= q , (2.6)
12
where the equality (a) is obtained by using (2.5). It remains to calculate the noise power due to quantisation. The PDF
of eq is typically modeled as U ( q/2, q/2) meaning that the rounding error is completely random for different samples
within the range of quantisation error. This is based on experimental observations and can be also observed in Fig. 2.12(b).
Therefore, we have:
(
1/q, q/2  eq  q/2
p(eq ) = . (2.7)
0, otherwise

The mean square quantisation error (noise power) is:

30
Digital Communications Dr. Majid Safari

Zq/2
eq2 = eq2 p(eq )deq
q/2
✓ ◆
1 3 q/2 1 q3 q3
= e = ( 1)
3q q q/2 3q 8 8
q2
= . (2.8)
12
Finally, the SNq R can be obtained by using (2.4), (2.6) and (2.8) as:

q2 M 2 1 12
SNq R = = M2 1. (2.9)
q2 12
Peak signal to quantisation noise ratio is another measure of distortion due to quantisation and would be beneficial partic-
ularly when the statistics of the input signal is not known (since it is only based on the maximum signal power rather than
average signal power). Replacing the average signal power in the numerator of SNq R with the maximum signal power
(i.e, the square of the maximum value of the quantiser’s dynamic range), (SNq R) p is obtained as:

v2peak [(M 1)q/2]2 12q2 (M 1)2

(SNq R) p = = = = 3(M 1)2 . (2.10)
eq2 eq2 4q2

As mentioned, the peak SNq R is independent of the signal statistics and only related to the characteristics of the
quantiser. The average SNq R can be then obtained for any signal with an arbitrary pdf (not necessarily uniform) based on
the known peak SNq R as

v2 v2peak 1 1
SNq R = = = (SNq R) p . (2.11)
eq2 eq2 a a

where a denotes the ratio of the peak and the average of the input signal with the arbitrary pdf as

v2peak
a, . (2.12)
v2
which can be calculated for any pdf. For example, the typical values of a is 3 dB for a sinusoidal signal1 , 12 dB for a
zero-mean clipped Gaussian distributed random signal2 with sg = vpeak /4 and 10 dB for a speech signal. Having a and
using equations 2.11 and 2.10, the average SNq R can be then obtained as:

3(M 1)2
SNq R = .
a
Note that by introducing a, we now obtained the the average SNq R for a signal with any arbitrary distribution (not
necessary uniform distribution assumed earlier).

2.4 Pulse code modulation (PCM) H6.5

A PCM is used to convert an analog signal (e.g., speech signal) to a digital train of pulses that fits into the requirement
of the channel over which it can be transmitted. In PCM, typically after quantisation, the signal is represented in binary
values, which is further referred to as codewords. For example, if the number of quantisation levels is 16, the quantised

1 Note that the time average of the square of a sinusoid with amplitude a is a2 /2.
2 note that since the peak amplitude is assumed to be the quarter of the standard deviation, we have a = v2peak /sg2 = 42 = 16 = 12 dB.

31
Digital Communications Dr. Majid Safari

signal can be sufficiently represented with 4-bit binary values (0000 to 1111). In general, for M = 2n quantisation levels,
each quantised value is represented based on binary codes of length n = log2 (M). Then, the binary values are represented
as a stream of binary pulses based on On-Off keying (similar to traditional systems based on Morse code). The 0 bit value
corresponds to an ‘off’ pulse, while the 1 bit value corresponds to an ‘on’ pulse. The whole process from the input analog
signal to the PCM signal is illustrated as in Fig. 2.13.

(a) Input signal. (b) Flat toped sampled signal.

(c) Quantised signal (PAM). (d) PCM signal

Fig. 2.13 From input signal to output PCM.

It is worth mentioning here that PCM signal in Fig. 2.13 is more immune to noise (more power efficient) compared to
PAM signal (the quantised train of pulses) because the receiver only needs to distinguish between the two binary values.
On the other hand, PCM requires higher bandwidth (less spectrally efficient) compared to PAM as during each sampling
period the data is represented by multiple pulses (rather than one pulse in PAM) leading to faster changes of the signal
from high to low values. We will see this type of power-bandwidth trade-off in our future discussion and it mostly applies
to a general modulation techniques.

2.4.1 Distortion of PCM in a noise-free channel

This subsection discusses the effect of the quantisation distortion to PCM signal in terms of SNq R. Noting the relationship
between number of quantisation levels M and the length PCM codes, n, as:

n = log2 (M). (2.13)

The peak SNq R in (2.10) can be then rewritten in terms of the length PCM codes, n, as:

(SNq R) p = 3(M 1)2 ⇡ 3M 2 = 3(22n ). (2.14)

Then, the average SNq R of PCM signal for a signal with an arbitrary distribution can be expressed in terms of n (in dB)
as:

32
Digital Communications Dr. Majid Safari
✓ ◆
3(22n )
(SNq R)dB = 10 log10 = 10 log10 (3) + 20n log10 (2) 10 log10 (a) ⇡ 4.8 + 6n adB . (2.15)
a

Hence, the SNq R of speech signals (i.e., a ⇡ 10dB) can be estimated based on the following rule of thumb:

SNq R ⇡ 6(n 1). (2.16)

Exercise 2.1. A PCM system uses a uniform quantizer followed by a 7-bit binary encoder. The bit rate of the system is
equal to 50 ⇥ 106 bits/s. (a) What is the maximum message bandwidth for which the system operates satisfactorily. (b)
Determine the output signal-to-(quantization) noise ratio when a full-load sinusoidal wave of frequency 1 MHz is applied
to the input.

Answer (a): Let the message bandwidth be W . Then, sampling the message signal at its Nyquist rate to maintain
satisfactory operation by avoiding aliasing, and using an n-bit code to represent each sample of the message signal, we
find that the bit duration is
Ts 1
Tb = =
n 2W n

The bit rate is

1
= 2W n
Tb

The maximum value of message bandwidth is therefore

50 ⇥ 106
Wmax = = 3.57 ⇥ 106 Hz
2⇥7
Answer (b): The output signal-to-quantizing noise ratio for a sinusoid with a = 3dB is given by

SNq R = 4.8 + 6n adB = 1.8 + 6n = 1.8 + 6 ⇥ 7 = 43.8dB

2.4.2 PCM with noisy channel

In this subsection, we assume transmission of the PCM signal over an AWGN channel and thus in addition to the quan-
tisation error, there is also distortions caused by the channel’s induced noise. We assume that the PCM encoder, assigns
binary codes to the quantisation levels based on their natural order. In such case, the 0th quantisation level is represented
by ‘000’, the 1st is represented by ‘001’ and so on. Therefore, if an error caused by the noise in the channel flips the least
significant bit (LSB) in the PCM code it would translates into a single quantisation level error after decoding the binary
codewords. However, If an error flips a more significant bit, it would correspond to a larger number of quantisation level
error after decoding.
Next, we intend to derive the SNR performance of a PCM system in the presence of noise and quantisation error. To
simplify the analysis, we assume that the system is reliably operating with a low error rate meaning that the probability
of more than one error happening in each codeword is very low; that is, only one error may occur in each n-bit codeword.
Consequently, all bits in the codeword have the same error probability Pe . To index a codeword, a subscript 1, 2, · · · , n is
used, where the subscript 1 is used for the LSB. The possible errors in the decoded signal are:

33
Digital Communications Dr. Majid Safari

e1 = q
e2 = 2q
···
en = 2n 1 q,

where ei denotes the error after decoding if the noise in the channel flips the ith significant bit of the codeword and q
represents the quantisation increment (e.g. if the maximum allowed input voltage of the ADC device is 5 V and M = 8 (3
bits) then q = 5/8 = 0.625 V). Then, the error power is:

e1 = q2
e2 = (2q)2
···
en = (2n 1 q)2 .

The mean square error in the decoded signal due to the channel noise can be obtained as:
n ⇣ 2
⌘ 4n 1 4n 1
2 =
ede  ek2 Pe = q2 + (2q)2 + · · · + 2n 1 q = Pe q2 40 + 41 + · · · + 4n 1 (a)
= Pe q2
4 1
= Pe q2
3
, (2.17)
k=1

where (a) uses the sum of a geometric sequence.3 The SNR that takes into account distortions caused by both quantisation
and channel noise after decoding is defined as:

v2
SNR = . (2.18)
2
eq2 + ede

Then Considering a signal with arbitrary distribution, the average signal power can be written in terms of the peak
power as

v2peak (M 1)2 q2
v2 = = , (2.19)
a 4a
where vpeak = (M 1)q/2 is the maximum quantised value of the quantiser. Now, inserting the signal power calculated
above and the distortion power due to noise and quantisation obtained in (2.17) and (2.8), respectively in to the definition
of SNR above, we have:

v2 3(M 1)2
SNR = n = . (2.20)
q2 /12 + Pe q2 4 3 1 (1 + 4Pe (M 2 1))a

Repeating the same steps, a similar expression can be obtained for a signal with uniform distribution as:

M2 1
SNR = , (2.21)
1 + 4Pe (M 2 1)

We quickly mention here that for most practical systems, a PCM system is quantisation-limited for a large received SNR
and noise-limited for a smaller received SNR. The quantisation-limited case refers to the case where eq2 >> ede 2 (e is
q
dominant). For the noise-limited case, however, the channel noise is dominant. This is illustrated in Fig. 2.14 where the
channel noise decreases (received SNR increases) along the X axis. Therefore, at small values of X axis which corresponds
to high channel noise, the system is still noise-limited (quantisation error is small compared to noise) while when we move
along the X axis the noise becomes gradually negligible and therefore the system becomes quantisation-limited with a
fixed output SNR which is only depends on the quantiser specifications.

n
3 sn = a + ar + ar2 + · · · + ar n 1 = a rr 1
1 .

34
Digital Communications Dr. Majid Safari

Fig. 2.14 SNR for decoded PCM.

2.4.3 PCM based on nonlinear quantisation

This subsection specifically targets the practical PCM systems designed based on nonlinear quantisation. For a general
signal with a non-uniform statistical distribution a linear quantiser with a constant quantisation increment, q, does not lead
to maximum average SNq R. In fact, to achieve higher average SNq R, a more resolution of the quantisation (i.e., smaller
quantisation increment) can be given to the more probable lower amplitudes of the input signal, while less resolution
(larger quantisation increment) is given to the less probable higher amplitudes of the input signal. This also provide a more
balanced quality of the quantised signal whether the input signal amplitude is high or low. Hence, a better performance
can be gained this way as the quantisation noise will be reduced not only for the weaker signals but also on average. A
method to implement this is by using a non-linear quantiser where the quantisation increment changes along the dynamic
range of the quantiser. Comparisons between non-linear and linear quantisers are given in Fig. 2.19. For example, the
nonlinear quantiser shown here would provide higher SNq R for an input signal with a pdf similar to Fig. 2.11, where the
input signal around zero amplitude is more probable.

(a) Non-linear quantiser. (b) Linear quantiser.

Fig. 2.15 Non-linear and linear quantisers.

A practical approach to implement a nonlinear quantiser is using a linear quantiser along with a companding technique.
Companding includes a pre-processing block to compress the dynamic range of signal before the linear quantiser so that
the distribution of the resulting signal after nonlinear compressing function is nearly uniform. Then a post-precessing

35
Digital Communications Dr. Majid Safari

block is used after linear quantiser that expands the dynamic range of signal to its original state. Hence, the expansion
block is based on the inverse function of the compression function. The overall system is referred to as companding
(compression and expansion) method and it is equivalent to a nonlinear quantiser, where the resolution of quantisation at
different signal levels is defined based on the compression function. Typical curves for a companding method is shown
in Fig. 2.16. As in the non-linear quantiser, more sensitivity level is given to the signal around zero amplitude. Since
at the receiver, the compression effect is uniquely inversed by an expansion curve, we can focus on the design of the
compression curve.

(a) Compression characteristic. (b) Expansion characteristic.

Fig. 2.16 Compression and expansion (companding) curves.

2.4.4 Companded PCM for signals with unknown statistics

In some scenarios such as in voice transmission or recording (e.g., in phone conversation), the exact statistics of the
speech signal is not known for the specific users and may also dynamically change over the conversation. In such cases,
optimizing the average SNq R is not feasible. Moreover, in a phone conversation scenario, we require that the voice signal
from both users show the same quality even though the level of the signal in two directions may be different due to
different quality of channels or the presence of parties with different voices volumes. Therefore, the design strategy of the
nonlinear compression function is based on preserving SNq R and thus signal quality for all signal amplitudes. Therefore,
in the following, we seek a compression function that allows for nearly equal performance across different input signal
levels.

Fig. 2.17 Logarithmic com-

pression characteristic.

Consider a compression function y = F(x) and let qy = q and qx = q̂(x) be linear (output) and nonlinear (input)
quantisation increments, respectively as shown in Fig. 2.17. Note that since the output of compression function y goes
through a linear quantiser qy is constant. Then, for small increments, we can write:

dy qy q
= = .
dx qx q̂(x)

The SNq R at different signal levels of x can be then defined as:

36
Digital Communications Dr. Majid Safari

x2 12x2
SNq R = = ,
q2x /12 q̂2 (x)

As discussed above, the objective of our design for the compression function is to have nearly equal SNq R for all signal
levels. Therefore, here, we assume SNq R to be constant at different values of x as:
x
= c,
q̂(x)

where c is a constant. Using the above equations, we have:

x x dy dy cq
= c =) = c =) = =) F(x) = y = cq log(x).
q̂(x) q dx dx x

Therefore, the compression function should be logarithmic to keep SNq R constant at different signal levels.
Given that we know y µ log(x) is not enough as it can be anything as in Fig. 2.18(a). As we assume that the PDF of
the input signal is symmetric at 0, then the y should be able to accommodate the negative values as well; hence, we have
an odd function constraint as in Fig. 2.18(b). In addition, the dynamic range of the input signal is typically continuous;
therefore, we also have the continuity constraint as in Fig. 2.18(c). This also ensures that we have a unique inverse function
at the receiver as the compression function is one-to-one mapping.

(a) y µ log(x) (b) Odd function constraint. (c) Continuity constraint.

Fig. 2.18 Constraints of the logarithmic function for the compression curve.

In order to meet the constraints discussed above, a standardized compression function is designed and termed as A-law
compression, which is mostly implemented in Europe. The curve for A-law compression is expressed as:
(
sgn(x) 1+ln(A|x|)
1+ln(A) , 1/A  |x|  1
F(x) = A|x| , (2.22)
sgn(x) 1+ln(A) , 0 < |x| < 1/A

where A defined the curvature law with A = 1 giving the linear response as shown in Fig. 2.19(a) and:

v
|x| = .
vpeak

In addition sgn(x) is the sign function defined as

(
1, x 0
sgn(x) = . (2.23)
1, x < 0

Note that A = 87.6 is widely used as shown in Fig. 2.19(a). Considering this value, 24 dB improvement in SNq R over
linear PCM is achieved for small signal levels (i.e., |x|  1/A). This gain can be calculated by noting that the ratio of the
quantisation noise for linear and nonlinear cases can be described as:

37
Digital Communications Dr. Majid Safari

(a) A-law compression curves for different values of A. (b) A-law segmented implementation.

Fig. 2.19 A-law Companding in Europe.

✓ ◆2 ✓ ◆2
SNq Rnonlinear q2 qy dF(x)
linear
= 2 = =
SNq R q̂ (x) qx dx

Another important characteristic of these companding techniques are the achievable dynamic range of constant SNq R
(i.e., |x| > 1/A) which can be calculated as 20 log10 (1/(1/A)) ⇡ 39dB for A = 87.6.
In the US and Japan, the µ-law compression is used, and it is defined as:

ln(1 + µ |x|)
F(x) = sgn(x) , 0  |x|  1.
ln(1 + µ)

where µ is the curvature law here. In the µ-law standard adopted by ITU-T, µ = 255. Compared to A-law, µ-law provides
slightly larger dynamic range of signals with equal SNq R at the cost of less SNq R improvement for small signals. The
µ-law and A-law adopted by ITU-T are Implemented segmented piecewise linear curves. See Fig. 2.19(b) for the 16-
segment A-law example noting the there are 8 positive and 8 negative segments. For an 8-bit A-law PCM, 1 bit is used to
define the polarity, 3 bits are used to define the particular segment of the curve (e.g., positive segments 1 to 8 shown in
Fig. 2.19(b)), and 4 bits are used to define the position of the signal in that segment of the curve.

Exercise 2.2. An 8-bit A-law companded PCM system is to be designed with piecewise linear approximation as shown in
Fig. 2.19(b). 16 segments are employed (with 4 co-linear near the origin) and the segments join at 1/2, 1/4, etc. of the
full scale value, as shown in Fig. 2.19(b). Calculate the approximate SNR gain for full scale and small signal values.
dF(x)
Answer: For small signals (in segments 7 and 8) the average gradient, dx ,
is given by:
!
1 1
dF(x) F 64 1 + ln 87.6 ⇤ 64 1
|small signals = 1
= /
dx 64
1 + ln 87.6 64
0.24
=
1/64
= 15.36

Thus there will be 15.36 times more levels in segments 7 and 8 than there would be for linear quantisation. This represents
an improvement in SNq R of 20 log10 (15.36) = 23.7 dB
For large signals (segement 1) graident is given by:

38
Digital Communications Dr. Majid Safari

dF(x) F(1) F(1/2)

|large signals =
dx 1/2
✓ ◆
1+ln 87.6⇤1/2
1 1+ln 87.6
=
1/2
1 0.873
=
1/2
= 0.253

Thus there will be 0.253 times less levels in segment 1 than there would be for linear quantisation representing a SNq R
degradation of 20 log10 (0.253) = 11.9 dB. ⌅

39
Digital Communications Dr. Majid Safari

References

[Haykin(2014)] Haykin S (2014) Digital Communication Systems. Wiley

40