Chapter 7 - Quantization Noise in DSP

Download as pdf or txt
Download as pdf or txt
You are on page 1of 31

7-1

7. Quantization Noise in DSP


Reference: Sections 6.6 to 6.9, 9.7 of Text
In our discussions of the filtering and FFT operations, it was implicitly
assumed that all the computations can be performed with infinite precision.
This is quite true with floating point implementation.
With fixed point implementation,
- the filter parameters will be quantized (finite word length
representation) and
- any intermediate result arising from the multiplication and/or addition
of numbers will be quantized (finite precision computation)
Some effects of quantization are:
- it changes the locations of the poles and zeros of a filter,
- it creates the equivalent of an additive noise term at the filter/FFT
output .
7.1 Number Representation and Quantization
In some DSP chips or special purpose hardware, numbers/signals are
represented and manipulated using fixed-point binary arithmetic.
7-2
Examples of fixed-point binary representations are sign and magnitude,
ones complement, and twos complement.
The twos complement representation is the most common format.
In the twos complement numbering system, a real number
x
within the
range
m m
X x X
is quantized to the number
0
1
2
B
i
m i m B
i
x X b b X x

_
+

,

,
where
, 0,1,...,
i
b i B
are binary (0,1) numbers, with
0
b
being the sign bit.
With this numbering scheme, the smallest difference between numbers is
2
B
m
X


and the fractional part of a number, i.e. the term
0
1
2
B
i
B i
i
x b b

,
can be represented by the binary pattern
0 1 2
...
B B
x b bb b o
,
where o is the binary point.
7-3
The number
m
X must be large enough that the risk of overflow is small,
and small enough that the quantization noise is kept to an acceptable level.
When B ,
x becomes the real number
x
.
The quantization of
x
to
x
can be done through
1. rounding, or
2. truncation.
The figures below illustrate these two operations for the case of 2 B .
Clearly the mapping from
x
to
x
is nonlinear in both cases.
rounding
7-4
truncation
Let
[ ]
B
Q g
denotes a 1 B + bit quantizer and let the quantization error be

[ ] .
B
e x x
Q x x


Then it can be easily deduced from the above figures that for rounding,
/ 2 / 2 e <
,
and for truncation,
0 e <
.
7-5
In studying the effect of quantization, the quantization error is usually
modelled as a uniform random variable. For rounding, the probability
density function (pdf) of this random variable is
1/ / 2 / 2
( ) ,
0 otherwise
e
p e
<

'

and for truncation, the pdf is


1/ 0
( ) .
0 otherwise
e
p e
<

'

The mean square quantization error for rounding is


/ 2 / 2
2
2 2 2
/ 2 / 2
1
( ) ,
12
e
e p e de e de


and for truncation is
0 0
2
2 2 2
1
( ) .
3
e
e p e de e de


Thus rounding is clearly more desirable than truncation.
Consider the multiplication of the quantized numbers
x
and
a during a
DSP operation. The product,
y xa
,
7-6
itself will be quantized to a 1 B + number
[ ]
B
y Q y
.
The quantized product can be written in terms of the unquantized product
xa
and the quantization error
y
e
as
[ ]

B y y
y Q y y e xa e + +
If the quantization error is modelled as a uniform random variable, the
model for a fixed-point multiplier becomes
This linearlized model will be adopted in our study of the effect of
rounding error in IIR filters later on.
7.2 The Effects of Coefficient Quantization
The transfer function of an IIR filter can be written as
[ ]
B
Q g
a
x
y
y
e
y
a
x
7-7
0
1
( )
( ) ,
( )
1
M
k
k
k
N
k
k
k
b z
B z
H z
A z
a z

where
1
( ) 1
N
k
k
k
A z a z

and
0
( )
M
k
k
k
B z b z

.
The Direct form implementation of this filter is shown below. All the filter
coefficients are at their original unquantized values.
We want to address in this section the issue of poles and zeros relocation
when the filter coefficients are quantized.
7-8
Consider the denominator polynomial
( ) A z
. This polynomial can be
written in product form as
( )
1
1 1
( ) 1 1
N N
k
k k
k k
A z a z p z



,
where the
k
p
s are the roots of ( ) A z , or equivalently the poles of ( ) H z . It
is understood that the
k
p
s are nonlinear functions of the
k
a
s. For
example when 2 N , then
1 1 2
a p p +
and
2 1 2
a p p
.
So how would the
k
p
s be affected when the
k
a
s are quantized?
Lets take the partial derivative of ( ) A z with respect to
i
a
. From the
summation form, we obtain
( )
i
i
A z
z
a

(1)
From the product form we obtain
( )
1 1
1 1
( )
1
N N
n
k
n k i i
k n
p A z
p z z
a a




' ;




(2)
Suppose we want to determine
/
j i
p a
. What we can do is to evaluate both
(1) and (2) at
j
z p
and equate them. The end result is
7-9
( )
1
N i
j j
N
i
j k
k
k j
p p
a
p p

Example: When 3 N , ( ) A z can be written as


1 2 3
1 2 3
( ) 1 A z a z a z a z


or
( )( )( )
1 1 1
1 2 3
( ) 1 1 1 A z p z p z p z


.
The partial derivative of the first expression with respect to
1
a
yields
1
1
( ) A z
z
a

.
On the other hand, the partial derivative of the second expression with
respect to
1
a
yields
( ) ( ) ( ) ( )
( ) ( )
1 1 1 1 1 1 3 2
1 2 1 3
1 1 1
1 1 1 1
2 3
1
( )
1 1 1 1
1 1
p A z p
p z p z z p z z p z
a a a
p
z p z p z
a


_ _
+ +


, ,
_

,
If we evaluate the two expressions at
3
z p and equate them, we obtain
( ) ( )
1 1 1 1
3
3 1 3 2 3 3
1
1 1
p
p p p p p p
a

_

,
7-10
which can be simplified to
( )( )
2
3 3
1 3 1 3 2
p p
a p p p p


Let
, 1,2,...,
k
a k N
, be the quantization errors in the
k
a
s when the IIR
filter is implemented using the Direct form computational structure with
fixed-point arithmetic. Then the poles will be shifted by the amounts
1
; 1,2,...,
N
j
j i
i
i
p
p a j N
a

Because of the term


( )
1
N
j k
k
k j
p p

in the denominator of
/
j i
p a
, we can deduce that if the poles are clustered
together, i.e. when
1
j k
p p <<
,
there could be big changes to the poles locations. Consequently, the direct
form implementation structure is quite sensitive to quantization errors in
the filter coefficients.
As shown in Section 6.4, the cascade form of an IIR filter is made up of a
serial concatenation of second order direct form subsystems. The poles
7-11
(zeros) of all the subsystems together form the poles (zeros) of the IIR
filter.
Quantization of the filter coefficients in a subsystem will only affect the
two poles (and 2 zeros) of that subsystem. In other word, quantization
errors are localized.
Thus the cascade form is generally much less sensitive to coefficient
quantization than the direct form.
Let us focus on a 2
nd
order IIR filter of the form
( )( )
1 1
1 2
1
( )
1 1
H z
p z p z


,
where
1
j
p re

and
2
j
p re

are the conjugate pole-pair of the filter. The two poles are located on a
circle with radius r , with one of them at phase angle of and the other at
an angle of .
The denominator polynomial can be written as
( )( )
( )
( )
1 1
1 2
1 2
1 2 1 2
1 2 2
( ) 1 1
1
1 2 cos
A z p z p z
p p z p p z
r z r z




+ +
+
7-12
Consequently the coefficients in the feedback portion of the IIR filter are
( )
1
2 cos a r
and
2
2
a r
The implementation structure of this filter is as shown below
If we assume 4-bit quantization of
1
a
and
2
a
in the interval [-1,+1]. Then
2
r and 2 cos r are numbers from the set:
7 3 5 3 3 5 3 7 1 1 1 1 1 1
8 4 8 2 8 4 8 8 4 8 2 8 4 8
1, , , , , , , , 0, , , , , , ,
.
This means after quantization, the poles can only take on values from the
set shown in diagram in the next page (only the first quadrant in the z-plane
are shown).
From the diagram, we can deduce that poles that are originally around
0
or

are more affected by quantization than those around
/ 2
.
7-13
Exercise: Determine
1
1
p
a

,
1
2
p
a

,
2
1
p
a

, and
2
2
p
a

for the above second order IIR filter. Assuming the original poles are on a
circle of radius
0.96 r
with phases of /36 t . Where are the
locations of the new poles after 4-bit quantization?
7-14
7.3 Zero Input Limit Cycle
For a stable digital filter, if the input is set to zero at some point in time, the
output should decay to zero.
For finite precision implementations, the output may decay to a non-zero
amplitude and then oscillate. This phenomenon is known as the zero-input
limit cycle; see for example the following results for a first order IIR filter
with 4-bit quantization between 1 and +1.
7-15
Zero-input limit cycle can be very annoying for applications such as
speech, as tones will be generated during the silence periods.
Zero-input limit cycle is unique to IIR filters because of the feedback
mechanism in these filters. No need to worry about this phenomenon for
FIR filters.
As an example, consider a first order IIR filter
[ ] [ 1] [ ]; 1 y n ay n x n a + <
With ( 1) B + -bit fixed-point implementation, the output becomes
[ ]
[ ] [ 1] [ ]; 1
B
y n Q ay n x n a + <
Here, we assume rounding with a quantization step size of .
Suppose the input becomes zero when
o
n n
. Then the unquantized output
at time
o
n
is
[ 1]
o
ay n
. Assuming that both
[ 1]
o
y n and the filter parameter
a are positive, then
[ 1]
o
ay n
will be quantized back to
[ 1]
o
y n if
[ 1] [ 1] / 2
o o
ay n y n
or when
( )
[ 1] ; ( [ 1] 0, 0)
2 1
o o
y n y n a
a

> >

7-16
In general, it can be shown that for a 1
st
order IIR filter, the zero-input limit
cycle exsists when
( )
[ 1]
2 1
o
y n
a

This is known as the dead band of the first order IIR filter.
Limit cycles can also be caused by overflow. In this case, the oscillation is
between large limits.
Limit cycles can be eliminated by adopting computation structures that do
not support them. But these structures are usually more computationally
intensive than the cascade form we discussed.
More bits in the quantization process will reduce the chance of limit cycles.
7.4 Effects of Round-off Noise in IIR Filters
In Section 7.1, we showed that a fixed-point multiplier can be replaced by
an unconstrained (real) multiplier in concatenation with an additive,
uniform noise source that models the round-off error.
PDF of round-off noise in a B+1 bit quantizer
7-17
We want to analyze in this section the effect of the round-off noise on the
output of an IIR filter. With infinite precision implementation, the input
output relationship of such a filter is given by
1 0
[ ] [ ] [ ]
N M
k k
k k
y n a y n k b x n k

+
.
With rounding, this becomes
[ ] [ ]
1 0
[ ] [ ] [ ]
N M
B k B k
k k
y n Q a y n k Q b x n k

+
,
where we assume that the input and the filter coefficients are already
quantized.
The figures below show the additive noise model for the fixed point
implementation of a 2
nd
-order IIR filter using the Direct Form I structure.
(a) Direct Form I (floating point) implementation of a 2
nd
order IIR Filter
7-18
(b) Fixed point implementation of a 2
nd
order IIR filter using the
Direct Form I structure
(c) Additive noise model for the fixed-point 2
nd
order IIR filter in (b).
The terms
0 1 2 3 4
[ ], [ ], [ ], [ ], and [ ] e n e n e n e n e n
in Diagram (c) represent the
round-off noises in the five fixed-point multipliers. Each of these noise
terms has zero mean and a variance of
7-19
2 2 2
2
2
12 12
B
m
X


Furthermore, we assume that each
[ ]
i
e n
is white (i.e. [ ] [ ] [ ] 0
i i
E e n e n m + )
and statistically independent of one another.
Since all the nodes in Diagram (c) represent adders, and since the output of
the first stage is the input to the second stage, we can lump all the noise
sources together into a single noise source
0 1 2 3 4
[ ] [ ] [ ] [ ] [ ] [ ] e n e n e n e n e n e n + + + +
and place it at the input to the second stage; see diagram below.
Like the individual
[ ]
i
e n
s, the combined noise term [ ] e n has zero mean. Its
variance, however, is
2 2
5
e

.
Since each
[ ]
i
e n
is white, so [ ] e n is also white. Consequently, the power
spectral density of [ ] e n is ( )
2
5
j
ee
e

.
7-20
The output
[ ] y n of the linearized model has two components, the desired
(i.e. real value) output [ ] y n and the noise term [ ] f n . The output noise term
is obtained by disabling the input line and feeding the combined noise term
to the feedback filter; see diagram below.
Since the frequency response of the feedback filter is
( )
2
1 2
1
1
j
ef
j j
H e
a e a e


,
the PSD of the output noise is
( ) ( ) ( )
2
2
2
2
1 2
1
5
1
j j j
ff ee ef
j j
e e H e
a e a e




In general, for an IIR filter with N feedback taps and 1 M + feedforward
taps, the variance of the combined noise term
[ ] e n
in a Direct Form I fixed-
point implementation is
7-21
( )
( )
2 2
2 2
1 2
1
12
B
m
e
N M X
N M

+ +
+ +
,
and its power spectral density is
( )
( )
2 2
2
1 2
12
B
m j
ee e
N M X
e

+ +

.
Since this combined noise term is injected into the input of the feedback
filter, the PSD of the output quantization noise, [ ] f n , is
( ) ( ) ( )
( )
2
2
2
1
2 2
2
1
1
( 1)
1
2 1
1 ,
12
1
j j j
ff ee ef
N
j k
k
k
B
m
N
j k
k
k
e e H e
N M
a e
X
N M
a e

_
+ +

,

and the output noise power is


( ) ( )
2 2
2
2
1
2 1 1
1
2 12 2
1
B
j m
f ff
N
j k
k
k
X d
e d N M
a e

_
+ +

,

While the Cascade Form is preferred over the Direct Forms in terms of
poles/zeros relocation caused by quantization, there are NO similar rules
for output quantization error. A lot actually depends on the coefficients of
the feedforward and feedback filters.
7-22
The diagrams below show the noise model for a fixed-point
implementation of a 2
nd
order IIR filter using the Direct Form II structure.
It is clear that round-off error affects Direct Forms I and II differently.
(a) Individual quantization noise sources in 2
nd
order Direct Form
II IIR filter
(b) Combined quantization noise sources in 2
nd
order Direct Form
II IIR filter
7-23
7.5 Quantization Noise in FFT
The DFT coefficients of a N-sample long signal [ ] x n are:
1
0
[ ] [ ] ; 0,1,..., 1
N
kn
N
n
X k x n W k N

,
where
2 / j N
N
W e

.
The DFT coefficients are actually uniformly spaced samples of ( )
j
H e

, the
spectrum of [ ] x n , between 0 2 < .
The DFT coefficients can be computed efficiently using a FFT algorithm.
The signal flow graph for the the decimation in time FFT algorithm, with
8 N , is shown below.
7-24
For 8 N , the computation of each DFT coefficient involves 7 butterfly
computational structures of the form shown below,
where [ ], [ ]
m m
X p X q represent the p-th and q-th intermediate output
coefficients of the m-th stage. For example in the case of [0] X we have
7-25
and for [2] X we have
In general, the computation of a DFT coefficient always involves 1 N
butterfly structures.
With fixed-point implementation, the complex multiplication of
1
[ ]
m
X q

by
r
N
W
in each butterfly introduces a complex quantization noise term
[ , ] e m q
;
see diagram below.
7-26
The expected value of
2
[ , ] e m q represents the average noise power due to
quantization and it can be determined as follows. Let
1 1 1
c a jb +
and
2 2 2
c a jb +
be two complex numbers with real components
1
a ,
2
a and imaginary
components
1 2
, b b . Their product, with infinite precision implementation, is
1 2 1 2 1 2 2 1
( ) ( ) c a a b b j a b a b + +
With fixed-point implementation though, the product of
1
c
and
2
c
becomes
( ) ( ) { } ( ) ( ) { }
1 2 1 2 1 2 2 1

B B B B
c Q a a Q b b j Q a b Q a b + +
The quantization error is thus
( ) ( ) { }
( ) ( )
{ }
( ) ( )
1 2 1 2 1 2 1 2
1 2 1 2 2 1 2 1
1 2 3 4



B B
B B
c c Q a a a a Q b b b b
j Q a b a b Q a b a b
e e j e e
+ 1 1
] ]
+ 1 1
] ]
+ +
,
7-27
where
( )
( )
( )
( )
1 1 2 1 2
2 1 2 1 2
3 1 2 1 2
4 2 1 2 1
,
,
,
B
B
B
B
e Q a a a a
e Q bb b b
e Q a b a b
e Q a b a b




are the individual real-value quantization errors. It is assumed that these
individual error terms are statistically independent. Consequently, the mean
magnitude square of the complex quantization erro is
( ) ( )
( ) ( )
2 2 2
1 2 3 4
2 2 2 2
1 2 1 2 3 4 3 4
2 2 2 2
1 2 3 4
2

2 2

4 ,
E c c E e e e e
E e e e e e e e e
E e E e E e E e

1 1
+ +
] ]
1
+ + + +
]
1 1 1 1 + + +
] ] ] ]

where
2 2
/12
is the mean-square quantization error of a 1 B + -bit real multiplier with a
step size of . From this analysis, we can conclude that the mean
magnitude square of
[ , ] e m q
is
2 2
2
2
[ , ] 4
12 3
B
E e m q

1

]
It should be pointed out that when the term
r
N
W in the butterfly is either
1, 1, , or j j + , then there is actually no quantization error.
7-28
For simplicity in our analysis though, we assume that each butterfly
introduces a complex quantization noise term whose average power is
2
B
.
From the signal flow graphs associated with the computation of [0] X and
[2] X in the 8 N case, we see that a quantization noise term introduced in
an intermediate butterfly will be multiplied by a sequence of complex
exponential terms of the form
r
N
W . This operation will not change the
average power of that noise source.
Consequently, an intermediate noise source can be replaced by an
equivalent noise source with the same power at the output node.
So for the 8 N example, there will be altogether 7 equivalent noise
sources, all with a power of
2
B
, being added to the output node.
It is reasonable to assume that all the 7 equivalent noise sources are
statistically independent. Consequently, they together form a single
combined noise source with a power of
2
7
B
.
In general, quantization in a N-point FFT algorithm produces the
equivalence of a noise term of power
2
( 1)
B
N in each of the DFT
coefficient. In other word, we can express the quantized DFT coefficinet

[ ] X k as

[ ] [ ] [ ] X k X k F k +
,
where [ ] F k is the effective quantization noise and
( )
2
2 2
[ ] 1
B B
E F k N N
1

]
7-29
Recall that
1
0
[ ] [ ] ; 0,1,..., 1
N
kn
N
n
X k x n W k N

Suppose the signal range for our fixed-point implementation is between 1


and +1. So to avoid overflow,
[ ] 1 X k <
.
This can be achieved by scaling the time-domain signal [ ] x n in such a way
that
1
[ ] x n
N
<
.
This stems from the fact that
1 1
0 0
[ ] [ ] [ ]
N N
kn
N
n n
X k x n W x n




.
Assuming that each
[ ] x n
is a uniform random variable in
[0,1/ ] N
. Then
1/
2
2
2
0
1
[ ]
3
N
E x n N y dy
N
1

]

.
Furthermore, if we assume that all the [ ] x n s are statistically independent,
then
7-30
2
1
2
0
1 1
*
0 0
1 1 1
2
*
0 0 0
[ ] [ ]
[ ] [ ]
[ ] [ ] [ ]

N
kn
N
n
N N
kn km
N N
n m
N N N
kn km
N N
n n m
m n
E X k E x n W
E x n W x m W
E x n E x m x n W W

1
1

1
]
1
]
1
_ _

1
, ,
]
1
1
1
+
1
1
]
1
]



2
1

3
1

3
N
N
N

The signal-to-noise (SNR) ratio in

[ ] X k is thus
( )
2
2
2 2 2 2 2 2 2
[ ]
1/ 3
1 1 2
3
[ ]
B
B B
E X k
N
N N N N
E F k


1
]

1
]
,
where 2
B
is the quantization step size of a (B+1)-bit quantizer when
the signal range is between 1 and +1.
The result indicates that if N is doubled (equivalent to adding 1 more FFT
stage), then we must use 1 more bit in the quantization process in order to
maintain the same SNR.
It is possible to reduce the above 1 extra bit per extra stage requirement to
bit per stage if the butterfly structure below is adopted instead.
7-31
Here, the signal [ ] x n (i.e. the input to the FFT) is scaled in such a way that
[ ] 1 x n <
.
(instead of
[ ] 1/ x n N <
). To avoid overflow, the input to any intermediate
butterfly is first scaled by before being multiplied and added. The
effective scaling factor for [ ] x n is thus 1/ 2 1/
v
N , where
2
log v N is the
total number of FFT stages.
A noise source in the th m stage, i.e. [ , ] e m p or [ , ] e m q , will be scaled in
magnitude by the factor
( 1)
2
v m
. This is in contrast to the original butterfly
structure where all noise sources introduced by quantization has a scaling
factor of 1 in magnitude.
This exponential weighting function on noise sources further back in the
overall computational structure leads to the improvement in SNR.

You might also like