TN 303

Digital Communications

Lecture #5
Differential Pulse Code Modulation
& Delta Modulation

Prof. A.N Mvuma UDOM

 Layout of Lecture #5

 Linear Prediction

 Differential Pulse Code Modulation (DPCM)

 Delta Modulation

Prof. A.N Mvuma UDOM

 Introduction

 Disadvantages of pulse code modulation (PCM)

are increased system complexity and increased
channel bandwidth.

 When simplicity of implementation is a basic

system requirement, delta modulation (DM) is
used as an alternative to PCM.

 When channel bandwidth saving is the most

important system requirement, differential pulse
code modulation (DPCM) is used.
Prof. A.N Mvuma UDOM
 Linear Prediction

 Present signal sample m  nTs  can be predicted

from N previous sample values
m  nTs  Ts  , m  nTs  2Ts  ,..., m  nTs  NTs  using a linear
prediction filter of order N .

 The predicted sample mˆ  nTs  is defined by the

convolutional sum as
N
mˆ  n    wi m  n  i  ( with a normalized Ts )
i 1
wi i  1, 2,..., N : filter coefficients
Prof. A.N Mvuma UDOM
 Linear Prediction

m  n  1 m  n  2 mn  N 
m n Delay Delay Delay
Ts Ts Ts

w1 w2 wN

m̂  n 

Prof. A.N Mvuma UDOM

 Linear Prediction

 Prediction error e  n  is the difference between

the true sample value m  n  and its predicted
value m̂  n  i.e.,
N
e  n   m  n   mˆ  n   m  n    wi m  n  i 
i 1

 The mean-square of the prediction error is given

by
  N
 
2

E e  n    E  m  n    wi m  n  i   
2

  i 1  

Prof. A.N Mvuma UDOM

 Linear Prediction
N
E e 2  n    E  m 2  n    2 wi E  m  n  m  n  i  
i 1
N N
  wi w j E  m  n  i  m  n  j  
i 1 j 1

 Assuming m  t  to be a sample function of a

zero-mean WSS random process M  t  and

E  m 2  n     M2

E  m  n  m  n  i    RM  i  , E  m  n  i  m  n  j    RM  i  j 

Prof. A.N Mvuma UDOM

 Linear Prediction

N N N
E e 2  n     M2  2 wi RM  i    wi w j RM  i  j 
i 1 i 1 j 1

 Differentiating E e 2  n   wrt wi

 N
E  e 2  n    2 RM  i   2 w j RM  i  j 
wi j 1

 For optimum w j
N

 j RM  i  j   RM  i  , i  1, 2,..., N (Wiener-Hopf Eqns.)

w opt

j 1

Prof. A.N Mvuma UDOM

 Linear Prediction

 Matrix form expression

R M w opt  rM
 RM  0   RM  N  1 
 
RM       :autocorrelation matrix
 R  N  1  R  0  
 M M 

rM   RM 1 RM  2  ... RM  N   : autocorrelation vector

T

T
w opt
  w
opt
1
opt
w
2 ... w   R -1M rM : optimum coefficient vector
opt
N

Prof. A.N Mvuma UDOM

 Linear Prediction

 Minimum mean-square of the prediction error

E e 2  n     M2  rMT R -1M rM   M2  e  n   m  n 

 Example: WWS processes M 1  t  and M 2  t  have

the following autocorrelation functions:

RM 1  0   1, RM 1 1  0.8; RM 2  0   1, RM 2 1  0.2

 What is optimum coefficient values of a one-tap

predictor and variance of the prediction errors?
Prof. A.N Mvuma UDOM
 Linear Prediction

RM 1 1
 Soln. R M1   RM 1  0   , rM1   RM 1 1   w opt
  0.8
RM 1  0 
1

E e12  n    RM 1  0   RM2 1 1 RM 1  0   0.36

RM 1 1
R M2   RM 2  0   , rM2   RM 2 1   w opt
  0.2
RM 1  0 
1

E e2 2  n    RM 2  0   RM2 2 1 RM 2  0   0.96

Prof. A.N Mvuma UDOM

 Differential Pulse Code
Modulation
 Successive samples m  nTs  of the message
signal m  t  (voice and video) taken at Nyquist
rate or higher do not change rapidly (they are
correlated).
 An approximation m  nTs  of the current sample
m  nTs  can be obtained from N previous samples,
m  nTs  Ts  , m  nTs  2Ts  , ...,m  nTs  NTs  by a N th predictor.

Prof. A.N Mvuma UDOM

 Differential Pulse Code
Modulation
 In differential pulse code modulation (DPCM)
system the prediction error e  nTs   m  nTs   m  nTs  ,
which has much less average power (variance)
than m  nTs  is quantized and encoded.

 Quantizing e  nTs 
results in much less
quantization error qe  nTs  than qm  nTs  due to
quantizing m  nTs  yielding improved SNRq
compared to conventional PCM.
Prof. A.N Mvuma UDOM
 Differential Pulse Code
Modulation

e  nTs 
m  nTs  + eˆ  nTs 
 Quantizer Encoder
Bit
- stream
m  nTs  +

Predictor 

mˆ  nTs 

Prof. A.N Mvuma UDOM

 Differential Pulse Code
Modulation

Bit
stream eˆ  nTs  mˆ  nTs 
Decoder 

m  nTs 

Predictor
e  nTs   m  nTs   m  nTs 
eˆ  nTs   e  nTs   qe  nTs 
mˆ  nTs   m  nTs   eˆ  nTs   m  nTs   e  nTs   eˆ  nTs   m  nTs   qe  nTs 

Prof. A.N Mvuma UDOM

 Differential Pulse Code
Modulation
 The output signal-to-noise ratio SNRO of the
DPCM system is

 M2  M2  E2 3   M2 2 R
SNRO  2  2 2  G p SNRq  G p 2 2
Q  E Q mmax

 M2 : Variance of the message signal

 Q2 : Variance of the quantization error
 E2 : Variance of the prediction error
G p : Processing gain

Prof. A.N Mvuma UDOM

 Differential Pulse Code
Modulation
 The processing gain G p is maximized by proper
selection of predictor taps that minimize  E2 .

 For a constant SNRQ, DPCM uses fewer bits per

sample hence less transmission bandwidth.

 For the same number of bits per sample, DPCM

results in an improved output signal-to-
quantization noise ratio. (4-11 dB in voice
signals).

Prof. A.N Mvuma UDOM

 Differential Pulse Code
Modulation
 Example 2: Determine processing gain G p when
the one-tap prediction filter with optimum
coefficient realizes a DPCM system for
processing of signals in Example 1.

RM1  0  1
For m1  t  , G p    2.78 or 4.44 dB.
RE1  0  0.36
RM 2  0  1
For m2  t  , G p    1.042 or 0.18 dB.
RE2  0  0.96

Prof. A.N Mvuma UDOM

 Delta Modulation

 Samples m  nTs  of a baseband message signal m  t 

obtained at a sampling rate f s much higher than the
Nyquist rate are highly correlated.

 Average power of prediction error e  nTs  , for a one-

tap prediction with unity weight is smaller
compared to that of the message signal m  nTs  .

 This justifies the use of the simplest quantizer with

only two levels. The rate of information
transmission is equal to the sampling rate.
Prof. A.N Mvuma UDOM
 Delta Modulation

e  nTs 
m  nTs  +  eˆ  nTs 
 Encoder
 Bit
- stream
m  nTs  mˆ  nTs  +
Delay

Ts +

mˆ  nTs 

Prof. A.N Mvuma UDOM

 Delta Modulation

Bit
stream eˆ  nTs  mˆ  nTs  Low- m̂  t 
Decoder  pass
Filter
m  nTs 
Delay
Ts

e  nTs   m  nTs   m  nTs 

 m  nTs   mˆ  nTs  Ts 
eˆ  nTs    sgn e  nTs  
n
mˆ  nTs   mˆ  nTs  Ts   eˆ  nTs    eˆ  nTs 
i 1

Prof. A.N Mvuma UDOM

 Delta Modulation

 Slope-overload distortion occurs when the step

size   2 is too small for a staircase
approximation m̂  t  to follow a steep segment of
the input signal m  t .

 To avoid slope-overload distortion requires that

 dm  t 
 max
Ts dt

Prof. A.N Mvuma UDOM

 Delta Modulation

 Granular noise (quantization noise) power

increases with increase in  and is given by

2
 q2 
3
m t 
m̂  t 

t
Slope-overload Granular noise
distortion

Prof. A.N Mvuma UDOM