www.jntuworldupdates.

org

8.1 FINITE WORD LENGTH EFFECTS

Finite Wordlength Effects

All the signals and systems are digital in DSP. The digital implementation has finite accuracy. When numbers are
represented in digital form, errors are introduced due to their finite accuracy. These errors generate finite precision
effects or finite wordlength effects.
Let us consider an example if the first order IIR filter to illustrate how errors are encountered in
discretization. Such filter can be described as,
y(n)= α y(n-1)+x(n) .......(1)
The z-transform of above equation gives
z[y(n)]= α z[y(n-1)+x(n)]
X(z)= α z −1 Y(z)+X(z)
Hence the transfer function

1 z
H(z)= −1 =
1 − αz z −α
Here observe that ‘ α ’ is the filter coefficient when this filter is implemented on some DSP processor or
software, ‘ α ’ can have only discrete values. Let the discrete values of α be represented by α .
Hence the actual transfer function which is implemented is given as,

z
H (z) =
z −α
The transfer function given by above equation is slightly different from H(z). Hence the actual frequency
response will be different from desired response.
The input x(n) is obtained by sampling the analog input signal. Since the quantizer takes only fixed(discrete)
values of x(n) ,error is introduced. The actual input can be denoted by x(n) .

x(n) =x(n)+e(n)
Here e(n) is the error introduced during A/D conversion process due to finite wordlength of the quantizer.
Similarly error is introduced in the multiplication of α and y(n-1) in equation(1). This is because the product
α y(n-1) has to be quantized to one of the available discrete values. This introduces error. These errors generate
finite wordlength effects.
Finite Wordlength Effects in IIR Digital Filters
When an IIR filter is implemented in a small system, such as an 8-bit microcomputer, errors arise in representing
the filter coefficients and in performing the arithmetic operations indicated by the difference equation. These
errors degrade the performance of the filter and in extreme cases lead to instability.
Before implementing an IIR filter, it is important to ascertain the extent to which its performance will be
degraded by finite wordlength effects and to find a remedy if the degradation is not acceptable. The effects of
these errors can be reduced to acceptable levels by using more bits but this may be at the expense of increased
cost.
The main errors in digital IIR filters are:
i. ADC Quantization Noise:
This noise is caused by representing the samples of the input data ,by only a small number of bits.

www.specworld.in 1 www.smartzworld.com
 www.jntuworldupdates.org

ii. Coefficient quantization errors:

These errors are caused by representing the IIR filter coefficients by a finite number of bits.
iii. Overflow errors
These errors are caused by the additions or accumulation of partial results in a limited register length.
iv. Product round-off errors
These errors are caused when the output ,and results of internal arithmetic operations are rounded to the permissible
wordlength.
Finite Wordlength Effects in FFT Filters
As in most DSP algorithms, the main errors arising from implementing FFT algorithms using fixed point arithmetic
are
i. Round off errors
These errors are produced when the product W k B is truncated or rounded to the system wordlength.
ii. Overflow errors
These errors result when the output of a butterfly exceeds the permissible wordlength.
iii. Coefficient quantization errors
These errors result from representing the twiddle factors using a limited number of bits.

8.2 LIMIT CYCLES

8.2.1 Overflow oscillations
Limit Cycle: The finite wordlength effects are analyzed using the linear model of the digital systems. But
nonlinearities are introduced because of quantization of arithmetic operations. Because of these nonlinearities,
the stable digital filter under infinite precision may become unstable under finite preision.Because of this instability,
oscillating periodic output is generated.Such output is called limit cycle.The limit cycle occur in IIR filters due to
feedback paths.
Types of Limit Cycles
There are two types of limit cycles.
(1) Granular and
(2) Overflow.
1. Granular Limit Cycles
The granular limit cycles are of low amplitude. These cycles occur in digital filters when the input signal levels
are very low. The granular limit cycles are of two types. They are
i. Inaccessible limit cycles
ii. Accessible limit cycles.
2. Overflow Limit Cycles
Overflow limit cycles occur because of overflow due to addition in digital filters implemented with finite precision.
The amplitudes of overflow limit cycles are very large and it can cover complete dynamic range of the register.
This further leads to overflow causing cumulative effect. Hence overflow limit cycles are more serious than
granular limit cycles.

www.specworld.in 2 www.smartzworld.com
 www.jntuworldupdates.org

Transfer Characteristics and Example

f(v )

-1 0 1 2 v
-1

Figure: Transfer characteristics of adder having overflow limit cycles

Because of overflow limit cycle oscillations the output fluctuates between minimum and maximum values.
The above figure shows the transfer characteristic of an adder that exihibit overflow limit cycle oscillations. Here
f(v) indicates addition operation. Consider the addition of following two numbers in sign magnitude form.

7 5
x1 = 0.111 i.e., x2 = 0.101 i.e.,
8 8
2
Then x1 + x2 = 1.010 i.e., −
8
Here overflow has occured in addition due to finite precision and the digit before decimal point makes the
number negative.
Signal Scaling:
Need for Scaling: Limit cycle oscillations can be avoided by using the nonlinear transfer characteristic. But it
introduces distortion in the output. Hence it is better to perform signal scaling such that overflow or underflow
does not occur and hence limit cycle oscillations can be avoided.
Implementation of Signal Scaling

Figure shows the direct form-II structure of IIR filter. Let the input x (n ) be scaled by a factor s 0 before the
summing node to prevent overflow. With the scaling, the transfer function will be,

b0 + b1z −1 + b 2 z −2 B(z )
H(z ) = s 0 = s0
−1
1 + a1z + a 2 z −2
A(z )
w (n )
x (n ) + + y (n )
s0

z -1
-a 1 -b 1
+ +

z -1
-a 2 -b 2

Figure: Direct form-II realization for second order IIR filter

Let us define the transfer function

W (z )
H ’(z ) =
X(z )
From figure we can write above transfer function as,

www.specworld.in 3 www.smartzworld.com
 www.jntuworldupdates.org

H ’(z ) =
s0
1 + a 1z + a 2 z − 2
−1

Since A(z ) = 1 + a1z −1 + a 2 z −2

H ’(z ) =
s0
A(z )
W (z )
Since, H (z ) =
’

X(z )
’

W (z ) s
= 0
X(z ) A(z )
s 0 X(z )
W (z ) =
A(z )
(or)

S(z ) =
1
A(z )
Let ’ , then above equation becomes,

W (z ) = s 0S(z )X(z )
Evaluating z-transform on unit circle, we put z = e jω in above equation,

( ) ( )( )
W e jω = s 0S e jω X e jω
Taking inverse Fourier transform of above equation,

ω(n ) =
1
2π ∫ ( )
W e jω e jωn dω =
1
2π ∫ ( )( )
s 0S e jω X e jω e jωn dω

Hence ω (n ) =
s 02
( )( )
2

∫ S e X e e dω
2 jω jω j ωn

4π 2
Schwartz inequality states that,

∫ x1 (t )x 2 (t )dt ≤∫ x1 (t ) dt ∫ x 2 (t ) dt
2 2 2

Using this relation we can write above equation as,

2 s 02 
4π 
2
( )
ω (n ) ≤ 2 ∫ S e jω dω.∫ X e jω dω ,
2


( )
jωn
since, e =1

1
( ) 
∴ ω2 (n ) ≤ s 02  ∫ S e jω dω. ∫ X e jω dω
2 1 2
( )
 2π 2π 

( ) = ∑ x (n ). Then above equation can be written as,

∞
1
∫
2
Parseval’s theorem states that X e jω 2

2π n =0

1 
( )
∞
ω2 (n ) ≤ s 02  ∫ S e jω dω.∑ x 2 (n )
2

 2π n =0 

www.specworld.in 4 www.smartzworld.com
 www.jntuworldupdates.org

We have put z = e jω . Hence dz = j e jωdω or

dz dz
dω = jω = , since e jω = z
je jz
Putting these values in equation
1 
2 dz ∞
( )
ω2 (n ) ≤ s 02  ∫ S e jω . .∑ x 2 (n )
 2π jz n =0 
∞
≤ s 02 ∑ x 2 (n ). S(z ) .
1 2 dz

n =0 2πj ∫ z
( )
Here S(z ) = S(z ).S z −1 . Then we have,
2

( )
∞
ω2 (n ) ≤ s 02 ∑ x 2 (n ). S(z ).S z −1 z −1dz
1
n =0 2πj ∫
( )
∞

Here the integration is executed over a closed contour i.e. ω (n ) ≤ s 0 ∑ x (n ). S(z ).S z −1 z −1dz
1
∫
2 2 2

n =0 2πj C

(or)
∞
[ ( )
ω2 (n ) ≤ ∑ x 2 (n ) s 02 ∫ S(z ).S z −1 z −1dz
n =0
C
]
Here ω2 (n ) represents instantaneous energy of signal after first summing node. And x 2 (n ) represents
instantaneous energy of input signal. Overflow will not occur if
∞
ω2 (n ) ≤ ∑ x 2 (n )
n =0

For this equation to be true we get following condition from equation

s 02
1
2πj ∫C
( )
S(z ).S z −1 z −1dz = 1

Earlier we have defined S(z ) =

1
A(z )
. Hence above condition becomes,

1 z −1dz
2πj ∫C A(z ).A z −1
=1
s 02
( )
1
s 02 =
1 z −1dz
2πj ∫C A(z ).A z −1 ( )
(or)

Above equation gives the value of scaling factor s 0 to avoid overflow

www.specworld.in 5 www.smartzworld.com
 www.jntuworldupdates.org

8.3 ROUND OFF NOISE IN IIR DIGITAL FILTERS

Statistical Model for Analysis of Round-off Error Multiplication:
We perform arithmetic operations like addition and multiplication some errors will be occured. Those errors are
called arithmetic errors. The results of arithmetic operations are required to be quantized so that they can occupy
one of the finite set of digital levels. Such operation can be visualized as multiplier (or other arithmetic operation)
with quantizer at its output.

α
u (n ) Q v (n )
v (n )
Figure: Quantization of multiplication or product
The above process can be represented by a statistical model for error analysis. The output υ (n) and error
eα (n) in product quantization process.i.e.,

υ (n) = υ (n) + eα (n)

The statistical model is shown below.
α
u(n) + v(n)
v(n)

e α(n )

Figure: Statistical model for analysis of round-off error multiplication

For the analysis purpose following assumptions are made.

i) { }
The error sequence eα (n ) is the sample sequence of a stationary white noise process.

ii eα (n) is having uniform distribution over the range of quantization error.

iii) The sequence {eα (n)} is uncorrelated with the sequence υ (n) and input sequence x(n).
8.3.1Computational output round off noise
Product Round-off Errors and its Reduction:
The results of product or multiplication operations are quantized to fit into the finite wordlength,when the digital
filters are implemented using fixed point arithmetic. Hence errors generated in such operation are called product
round off errors.
The effect of product Round-off errors can be analyzed using the statistical model of the quantization
process. The noise due to product round-off errors reduces the signal to noise ratio at the output of the filter.
Some times this ratio may be reduces below acceptable levels. Hence it is necessary to reduce the effects of
product round-off errors.
There are two solutions available to reduce product round-off errors.
a) Error feedback structures and
b )State space structure.
The error feedback structures use the difference between the unquantized and quantized signal to reduce
the round-off noise. The difference between unquantized and quantized signal is fed back to the digital filter
structure in such a way that output noise power due to round-off errors is reduced.

www.specworld.in 6 www.smartzworld.com
 www.jntuworldupdates.org

First Order Error-feedback Structure to reduce Round-off Error:

The results of product or multiplication operations are quantized to fit into the finite wordlength, when the digital
filters are implemented using fixed point arithmetic. Hence errors generated in such operation are called product
round off errors.
The effect of product round-off errors can be analyzed using the statistical model of the quantization
process.
Let the quantization error signal be given as the difference between unquantized signal y(n) and quantized signal
υ (n) .i.e.,
e(n) = y(n)- υ (n)

z -1
e(n )
β - + +
K
x (n ) + Q y (n )
v (n )
z -1

α
Figure: First order error feedback structure to reduce product round-off error
This error signal is fed back in the structure such that round-off noise is reduced. Such structure for first order
digital filter is shown in figure.
The incorporation of quantization error feedback as shown in figure helps in reducing the noise power at the
output . This statement can be proved mathematically.
Round-off Errors in FFT Algorithms:
FFT is used in large number of applications. Hence it is necessary to analyze the effects due to finite wordlengths
in FFT algorithms. The most critical error in FFT computation occurs due to arithmetic round-off errors.
The DFT is used in large number of applications such as filtering, correlation, spectrum analysis etc. In
such applications DFT is computed with the help of FFT algorithms. Therefore it is important to study the
quantization errors in FFT algorithms. These quantization effects mainly take place because of round-off errors.
These errors are introduced when multiplications are performed in fixed point arithmetic.
FFT algorithms require less number of multiplications compared to direct computation of DFT. But it
does not mean that quantization errors are also reduced in FFT algorithms.

Let σ x represents the variance of output DFT coefficients i.e., X (k ) .

2

For N-point DFT σ x is given as,

1
σ 2x = ......(1)
3N
For direct computation of DFT, the variance of quantization errors in multiplications is given as,

N 2
σq2 = .∆ ......(2)
3

Here σq is variance of quantization errors and ∆ is step size which is given as,
2

∆ = 2−b .....(3)

www.specworld.in 7 www.smartzworld.com
 www.jntuworldupdates.org

And b is the number of bits to represent one level.

Hence equation (2) becomes,

N −2 b
σq2 = .2 .....(4)
3
The signal to noise power ratio at the output (i.e., DFT coefficients) can be considered as the measure of
( )
quantization errors. This ratio is the ratio of variance of DFT coefficients σ 2x to the variance of quantization

( )
errors σ q i.e.,
2

 σ 2x 
DFT =  
Signal to noise ratio in direct computation of  σ2 
 q  Direct DFT

From equation (1) and equation (2) we have,

1
 σ 2x 
  = 3N 22b
 σ q2  N −2 b = - ......(5)
  Direct DFT .2 N2
3
When DFT is computed using FFT algorithms, the variance of the signal remains same i.e.,
1
σ 2x = from equation (1) ......(6)
3N
But with algorithms the variance of the quantization errors is given as,
2
σq2 = .2 − 2 b ......(7)
3
Hence signal to noise ration in FFT algorithms is,
1
 σ 2x  22b
  = 3N
 σ2  =
 q  FFT 2 .2 − 2 b 2N
3
In the above expression, the signal to noise ration is inversely proportional to N. whereas in direct DFT
computation the signal to noise ratio is inversely proportional to N 2 as given by equation (5). This means
quantization errors increase fast with increase in ‘N’ in direct computation of DFT. But in FFT algorithms the
quantization errors increase slowly with increase in ‘N’.
Product of Round-off Errors in IIR Digital Filters:
The results of product or multiplication operations are quantized to fit into the finite wordlength, when the digital
filters are implemented using fixed point arithmetic. Hence errors generated in such operation are called product
round off errors.
Product round-off error analysis is an extensive topic.Our presentation here will be brief and aims to
make you aware of the nature of the errors, their effects and how to reduce them if necessary.
The basic operations in IIR filtering are defined by the familiar second- order difference equation:
2 2
y (n) = ∑ bk x(n − k ) − ∑ ak y (n − k )
k =0 k =1

Where x(n-k) and y(n-k) are the input and output data samples,and bk and a k are the filter coefficients. In
practice these variables are often represented as fixed point numbers. Typically , each of the products bk x(n-k)

www.specworld.in 8 www.smartzworld.com
 www.jntuworldupdates.org

and a k y(n-k) would require more bits to represent than any of the operands. For example, the product of a B-bit
data and a B-bit coefficient is 2B bits long.
Truncation or rounding is used to quantize the products back to the permissible wordlength. Quantizing
the products leads to errors,popularly known as round-off errors,in the output data and hence a reduction in the
SNR. These errors can also lead to small-scale oscillations in the output of the digital filter,even when there is no
input to the filter.
x (n ) K 2 B b its B b its
(a ) Q
y (n )

x (n ) K 2 B b its B b its
(b ) Σ
y (n )

c(n )
Figure: Representation of the product quantization error: (a) a block diagram representation of the
quantization process; (b) a linear model of the quantization process
The figure(a) represents a block diagram of the product quantization process,and figure (b) represents a
linear model of the effect of product quantization. The model consists of an ideal multiplier,witk infinite precision,
in series with an adder fed by a noise sample, e(n), representing the error in the quantized product ,where we have
assumed,for simplicity,that x(n),y(n), and K are each represented by B bits. Thus
y(n) = Kx(n) + e(n)
The noise power, due to each product quantization, is given by

q2
σr =
2

12
Where r symbolizes the round-off error and q is the quantization step defined by the wordlength to which
product is quantized. The round-off noise is assumed to be a random variable with zero mean and constant
variance. Although this assumption may not always be valid, it is useful in assesing the performance of the filter.
Product of Round-off Errors on Filter Performance:
The effects of round-off errors on filter performance depend on the type of filter structure used and the point at
which the results are quantized.
The above figure represents the quantization noise model for the direct form building block. It is assumed
in the figure that the input data,x(n),output data,y(n),and the filter coefficients are represented as B-bit numbers
(including the sign bit). The products are quantized back to B bits after multiplication by rounding (or truncation).
e (n )

x(n) B b its 2B B b its B y(n)

Σ Σ y (n )
b0 s1 x (n )
s1 Σ
b0 s1
e1 s1
z -1 z -1

2B B B 2B z -1 z -1
Σ Σ
b1 -a 1
s1
z -1 e2 z -1 b1 -a 1
s1
2B B B 2B z -1
Σ Σ z -1
b2 -a 2
s1
e3 b2 -a 2
s1
Figure: Product quantization noise model for the direct form filter section. All the noise sources in (a) have

www.specworld.in 9 www.smartzworld.com
 www.jntuworldupdates.org

been combined in (b) as they feed to the same point

Since all five noise sources, e1 to e5 in figure(a),feed to the same point (that is into the middle adder), the
total output noise power is the sum of the individual noise powers(figure(b)).

e1 e2

x (n ) w (n ) B B b its x (n ) w (n )
Σ B b its Σ
Σ Σ Σ Σ
1 /s s 1b 0 y (n ) 1 /s 1 s1b 0 y (n )
z -1
e 1 (n ) w (n -1 ) e 6 (n )
z -1
B 2B 2B B
Σ Σ w (n -1 )
-a 1 s1b 1
s1b 1
z -1 -a 1
e 3 (n ) e 5 (n )
w (n -2 )
B 2B 2B B z -1
Σ Σ -b 2
-a 2 s 1b 2
w (n -2 )
e 2 (n ) e 4 (n )
-a 2 s1b 2
Figure: Product quantization noise model for the canonic filter section. The noise sources feeding the same
point in (a) have been combined in (b)

5q 2  1 −1 dz
 2 5q 2  ∞ 2  2 5q 2
σ or = ∫ ∑ f (k ) s1
2
  s1
2 2
F ( z ) F ( z )
z  = 12  k =0 = F ( z ) 2 s1
12 2∏ j c 12

1
Where F(z) =
1 + a1 z + a2 z − 2
−1

f(k) = Z −1 [F ( z )] is the inverse z-transform of F(z),which is also the impulse response from each noise source to

2 q2
the filter output, . 2 is the L2 norm squared and is the intrinsic product round-off noise power. The total
12
noise power at the filter output is the sum of the product round-off noise and the ADC quantization noise.

σ 0 = σ 0 A + σ or
2 2 2

q2  ∞ 2  q2
[ ]
∞

= ∑
12  k = 0
h ( k ) + 5 s1 ∑
2

k =0
f 2
( k )  =
 12
H ( z ) 2 + 5s12 F ( z )
2 2
2

For canonic section, figure(a) , the noise model again includes a scale factor as this generates a round-off
error of its own. The noise sources e1 (n) to e3 (n) all feed to the left adder, whilst the noise sources e4 (n) to
e6 (n) feed directly into the filter output. Combination of the noise sources feeding to the same point leads to the
noise model of figure(b). Assuming uncorrelated noise sources, the total noise contribution is simply the sum of
the individual noise contributions:

[ ]
∞
3q 2 3q 2 3q 2
σ or = ∑ f 2 (k ) +
2 2
= F ( z) 2 + 1
12 k =0 12 12

www.specworld.in 10 www.smartzworld.com
 www.jntuworldupdates.org

Where f(k) is the impulse response from the noise source e 1 to the filter output, and F(z) the corresponding
transfer function given by
b0 + b1 z −1 + b2 z −2
F ( z ) = s1
1 + a1 z −1 + a2 z − 2 s1
= H(z)

The total noise (ADC+round-off noises ) at the filter output is given by

σ 02 = σ 02A + σ or2

q2    ∞ 2  q2
{[ ]+ H ( z) }
∞

31 + s1 ∑ h (k ) + ∑ h (k ) =
2 2
3 1 + s12 H ( z )
2 2
=
12   =
k 0  k 0
=  12 2 2

8.4 METHODS TO PREVENT OVERFLOW

Prevent Overflow Limit Cycle Oscillations: The overflow limit cycles occur because of overflow due to addition
in digital filters implemented with finite precision. The amplitudes of overflow limit cycles are very large and it
can cover complete dynamic range of the register.
The specific design of filter coefficients do not assure prevention of overflow limit cycle oscillations. The
transfer characteristic can be modified to avoid overflow limit cycle oscillations.

f(v )

v
-1 0 1
-1

Figure: Prevention of overflow limit cycle oscillations

As shown in figure when an overflow or underflow is sensed, the output of the adder is set to its full scale
value of ± 1. This prevents oscillatory output. This nonlinearity of the characteristic causes very small distortion
in the output because overflow/underflow occurs rarely.
Scaling is also used to prevent overflow limit cycle oscillations. Limit cycle free structures are normally
used to avoid the effects of limit cycles.
Characteristics of a Limit Cycle Oscillation with respect to the System by the following Difference equation
y ( n ) = 0.95 y ( n − 1) + x ( n ) .

Let y r (n ) be the output of the system after the product term 0.95 y(n − 1) is quantized after rounding. i.e.,

y r (n ) = Q r [0.95y(n − 1)]+ x (n )

0.75 for n = 0
Let x (n ) = 
 0 for n ≠ 0

Let b = 4 bits are used to represent the quantized product excluding sign bit.
With n=0

www.specworld.in 11 www.smartzworld.com
 www.jntuworldupdates.org

y r (n ) = Q r [0.95y r (n − 1)]+ x (n )

∴ y r (0) = Q r [0.95y r (− 1)]+ x (0 ) = Q r [0.95 × 0]+ 0.75

Since y r (− 1) = 0 = 0.75

[0.75]10 = [0.11]2
∴ 4-bits rounded value of [0.11]2 will be [0.1100]2 i.e., 0.75 only.

∴ y r (0) = 0.75 after 4 bits rounding

With n = 1

y r (1) = Q r [0.95y r (0 )]+ x (1) = Q r [0.95 × 0.75]+ 0 = Q r [0.7125]

[0.7125]10 = [0.1 0 1 1 0 1 1 0 0 1 1 0 0 ⋅ ⋅ ⋅]2

Note that 0.7125 requires infinite binary digits for its representation. Let us round it to 4 bits.

∴ Q r [0.7125]10 = [0.1 0 1 1]2 upto 4 bits

But decimal equivalent of [0.1 0 1 1]2 is 0.6875.

∴ y r (1) = 0.6875
This means the actual value of y r (1) = 0.7125 is changes to 0.6875 due to 4-bits quantization.
With n = 2

y r (2 ) = Q r [0.95 y r (1)]+ x (2 ) = Q r [0.95 × 0.6875]+ 0 = Q r [0.653125]

[0.653125]10 = [0.1 0 1 0 0 1 1 1 0 0 1 1 0 0 1 1 0 0 ⋅ ⋅ ⋅]
∴ Q r [0.653125]10 = [0.1 0 1 0]2 upto 4 bits = [0.625]10

∴ y r (2 ) = 0.625
With n = 3

y r (3) = Q r [0.95y r (2 )]+ x (3) = Q r [0.95 × 0.625]+ 0 = Q r [0.59375]

[0.59375]10 = [0.1 0 1 0 1]2

∴ Q r [0.59375]10 = [0.1 0 1 1]2 upto 4 bits = 0.625

∴ y r (3) = 0.625

Thus y r (2 ) = y r (3) = ........ = 0.625

Thus the system enters into limit oscillation when n = 2 .
To calculate dead band
Consider equation
ä/2
y(n − 1) ≤
1− á

www.specworld.in 12 www.smartzworld.com
 www.jntuworldupdates.org

1 1
Here δ = b
= 4 = 0.0625
2 2

y(n − 1) ≤
0.0625/2
1 − 0.95
≤ 0.625
Dead band = [-0.625, +0.625]
Signal Scaling to Prevent Limit Cycle Oscillations: This is zero input condition. Following table lists the
values of y(n ) before and after quantization. Here the values are rounded to nearest integer value.
n y(n ) before quantization y (n ) after quantization
-1 12 12
0 10.8 11
1 9.72 10
2 8.748 9
3 7.8732 8
4 7.08588 7
5 6.377292 6
6 5.7395628 6
7 5.1656065 5
8 4.6490459 5
Table: Values of y(n ) before and after quantization

From table observe that if y(− 1) ≤ 5 , y(n ) = y(− 1) for n ≥ 0 for zero input. Hence the dead band will be [− 5,5].

Since the values are rounded to nearest integer after quantization, the step size will be δ = 1 . Hence
dead band can also be calculated as follows:

δ/2
y(n − 1) = ,Here α = 0.9 , y(n − 1) =
1/ 2
=5
1− α 1 − 0.9
Thus the dead band is [− 5,5] .
Dynamic Range Scaling to Prevent the Effects of Overflow: The overflow can take place at some internal
nodes when the digital filters are implemented by using fixed point arithmetic. Such nodes can be inputs/outputs
of address or multipliers. This overflow can take place even if the inputs are scaled. Because of such overflow at
intermediate points,produces totally undesired output or oscillations. The overflow can be avoided by scaling the
internal signal levels with the help of scaling multipliers. These scaling multipliers are inserted at the apprppriate
points in the filter structure to avoid possibilities of overflow. Sometimes these scaling multipliers are absorbed
with the existing multipliers in the structure to reduce the total number and complexity.
At which node the overflow will take place is not known in advance. This is because the overflow
depends upon type of input signal samples. Hence whenever overflow takes place at some node, the scaling
should be done dynamically. Hence dynamic range scaling in the filter structure can avoid the effects of overflow.
Let ur (n) be the signal sample at r th node in the structure. Then the scaling should ensure that,

u r (n) ≤ 1 for all r and n.

www.specworld.in 13 www.smartzworld.com
 www.jntuworldupdates.org

8.5 TRADE OFF BETWEEN ROUND-OFF AND OVERFLOW NOISE, MEASUREMENT OF

COEFFICIENT QUANTIZATION EFFECTS THROUGH POLE-ZERO MOVEMENT
Errors in Rounding and Truncation Operations : The computations like multiplication or addition are performed
the result is truncated or rounded to nearest available digital level. This operation introduces an error. Hence the
performance of the system is changed from expected value.
Truncation Error: This error is introduced whenever the number is represented by reduced number of bits.

Let Qt (x) be the value after truncation ,then truncation error will be,

ε r = Qt ( x) − ( x)
Here x is the original value of the number.
Rounding Error : This error is introduced whenever the number is rounded off to the nearest digital level.The
number of bits used to represent the rounded number are generally less than the number of bits required for actual
number.
Let Qr (x) be the value after rounding.Then rounding error will be,

ε r = Qr ( x) − x
Here x is the original value of a number.
Tradeoff between roundoff and overflow noise:
Scaling operation
Scaling is a process of readjusting certain internal gain parameters in order to constrain internal signals to a range
appropriate to the hardware with the constraint that the transfer function from input to output should not be
changes.
The filter in figure with unscaled node x has the transfer function
H(z) = D(z)+F(z)G(z) ......(1)
To scale the node x, we divide F(z) by some number β and multiply G(z) by the same number as in
figure. Although the transfer function does not change by this operation, the signal level at node x has been
changes. The scaling parameter β can be chosen to meet any specific scaling rule such as
∞
l1 scaling :β = ∑ f (i ) ......(2)
i =0

∞
l2 scaling :β = δ ∑ f (i )
i =0
2
…..(3)

Where f(i) is the unit-sample response from input to the node x and the parameter δ can be interpreted
to represent the number of standard deviations
D (z)

IN OUT
F (z) x G (z)
(a)

www.specworld.in 14 www.smartzworld.com
 www.jntuworldupdates.org

D (z)

IN OUT
F (z)/β x1 βG (z)

(b )
Figure: A filter with unscaled node x and (b) A filter with scaled node x’
Representable in the register at node x if the input is unit-variance white noise. If the input is bound by
u (n ) ≤ 1 , then,

∞ ∞
x (n ) = ∑ f (i)u (n − i) ≤ ∑ f (i ) ......(4)
i =0 i =0

Equation represents the true bound on the range of x and overflow is completely avoided by l1 scaling in (2),
which is the most stringent scaling policy.
In many cases, input can be assumed to be white noise. Although we cannot compute the variance at node
x. for unit-variance white noise input,

[ ]
∞
E x 2 (n ) = ∑ f 2 (i ) .....(5)
i =0

Since most input signals can be assumed to be white noise, l 2 scaling is commonly used. In addition, (5) can be
easily computed. Since (5) is the variance (not a strict bound), there is a possibility of overflow, which can be
reduced by increasing δ in (3). For large values of δ , the internal variables are scaled conservatively so that no
overflow occurs. However, there is a trade –off between overflow and roundoff noise, since increasing δ deteriorates
the output SNR (signal to noise ratio).
a

8 b its
1 5 b its

u (n ) + x (n )
D
u (n ) + x (n )
D
8 b its
(ro u n d o ff erro r)
Figure: Model of roundoff error
Roundoff Noise: If two W-bit fixed point fraction numbers are multiplies together, the product is (2W-1) bit
long. This product must eventually be quantized to W-bits by rounding or truncation. For example, consider the
1st –order IIR filter shown in figure. Assume that the input wordlength is W=8bits. If the multiplier coefficient
wordlength is also the same, then to maintain full precision in the output we need to increase the output wordlength
by 8 bits per iterations. This is clearly infeasible. The alternative is to roundoff (or truncate) the output to its
nearest 8-bit representation.

www.specworld.in 15 www.smartzworld.com
 www.jntuworldupdates.org

P e (X )

1
∆

x
∆
−∆ 2
2
Figure: Error probability distribution
The result of such quantization introduces roundoff noise e(n). For mathematical ease a system with
roundoff can be modeled as an infinite precision system with an external error input. For example in the previous
case (shown in figure) we round off the output of the multiply add operation and an equivalent model is shown in
figure.
Although rounding is not a linear operation, its effect at the output can be analyzed using linear system
theory with the following assumptions about e(n):
1. E(n) is uniformly distributed white noise.
2. E(n) is a wide –sense stationary random process, i.e., mean and covariance of e(n) are independent of the
time index n.
3. E(n) is uncorrelated to all other signals such as input and other noise signals.
Let the wordlength of the output be W-bits, then the roundoff error e(n) can be given by
− 2 − (w −1) 2− (w −1)
≤ e(n ) ≤ .....(6)
2 2
Since the error is assumed to be uniformly distributed over the interval given in (6), the corresponding
probability distribution is shown in figure, where ∆ is the length of the interval (i.e., 2 − (w −1) ).

Let us compute the mean E[e(n)] and variance E e 2 (n ) of this error function. [ ]
∆
∆
 1 x2  2
E[e(n )] = ∫ 2
∆ xPe (x )dx =   = 0 .....(7)
−
2  ∆ 2 −∆
2

Note that since mean is zero, variance is simply E e 2 (n ) [ ]

∆
∆
 1 x3  2
[ ]
E e 2 (n ) = ∫ 2
∆ x 2 Pe (x )dx =   =
∆2 2 − 2 w
=
 ∆ 3  − ∆ 12
− 3 …...(8)
2
2

b CT
u (n ) x (n + 1 ) z -1 x (n ) y (n )

e(n )

www.specworld.in 16 www.smartzworld.com
 www.jntuworldupdates.org

Figure: Signal flow graph

In other words (8) can be rewritten as

2 −2 w
σe2 = ......(9)
3
Where σe2 is the variance of the roundoff error in a finite precision, W-bit wordlength system. Since the
variance is proportional to 2 −2 w , increase in wordlength by 1 bit decreases the error by a factor of 4.
The purpose of analyzing roundoff noise is to determine its effect at the output signal. If the noise
variance at the output is not negligible in comparison to the output signal level, the worlength should be increase
or some low noise structures should be used. Therefore, we need to compute SNR at the output, not just the noise
gain to the output. In the noise analysis, we use a double length accumulator model, which means rounding is
performed after two (2w-1)-bit products are added. Also, notice that multipliers are the sources for roundoff
noise.
Effects of Coefficient Quantization in FIR filters: Let us consider the effects of coefficient quantization in FIR
filters. Consider the transfer function of the FIR filter of length M,
M −1

H(z) = ∑ h( n) z
n =0
−n

The quantization of h(n) takes place during implementation of filter. Let the quantized coefficeints be
denoted by h(n) and e(n) be the error in quantization. Then we can write,

h(n) = h(n)+e(n)
And the new filter transfer function becomes,
M −1 M −1 M −1 M −1
H ( z ) = ∑ h(n) z − n =
n =0
∑ [h(n) + e(n)]z −n = ∑ h(n) z −n + ∑ e(n) z −n
n=0 n=0 n=0
= H(z)+E(z)

M −1

Where, E(z)= ∑ e( n ) z
n =0
−n

H (z )

H (z )

E (z)

Figure: Model of FIR filter with quantizer coefficients

Here observe that the FIR filter with quantized coefficients can be modelled as parallel connection of two
FIR filters H(z) and E(z).
In the figure, H(z) is the FIR filter with unquantized coefficients, and E(z) is the FIR filter representing
coefficient quantization error.
The frequency response of FIR filter with quantized coefficients as,

www.specworld.in 17 www.smartzworld.com
 www.jntuworldupdates.org

H (ω ) = H (ω ) + E (ω )
Here E (ω ) is the error in the desired frequency response which is given as,
M −1

E (ω ) = ∑ e( n)e
− j ωn

n=0

Consider the magnitude of error i.e.,

M −1 M −1 M −1
E (ω ) = ∑ e(n)e − jωn ⇒ E (ω ) ≤ ∑ e(n) e − jωn ⇒ ≤ ∑ e(n)
n =0 n =0 n =0
(∴ e
− jωn
= 1)

The upper bound is reached if all the errors have same sign and have the maximum value in the range. If
we consider e(n) to be statistically independent random variables, then more realistic bound is given by standard
derivation of E( ω ) i.e.;

σ E (ω ) is the standard derivation of error in frequency response i.e., E( ω ).

8.6DEADBAND EFFECTS
Deadband and Deadband of First Order Filter: Dead band is the range of output amplitudes over which limit
cycle oscillations take place
Dead band of first order filter
Consider the first order filter,

y(n ) = αy(n − 1) + x (n )
Here α y(n − 1) is the product term. After rounding it to ‘b’ bits we get,

y(n ) = Q r [αy(n − 1)]+ x (n )

When limit cycle oscillations take place,

Q r [αy(n − 1)] = ± y(n − 1) …..(1)

δ
The error due to rounding is less than . Hence,
2

δ
Q[αy(n − 1) − αy(n − 1)] ≤
2
From equation (1) above equation can be written as,
δ
± y(n − 1) − αy(n − 1) ≤
2
δ
∴ y(n − 1)[1 − α ] ≤
2
δ/2
∴ y(n − 1) ≤ 1 − α

www.specworld.in 18 www.smartzworld.com