1 Finite Wordlength

Professor A G Constantinides 1
Finite Wordlength Effects

Finite register lengths and A/D converters
cause errors in:-
(i) Input quantisation.
(ii) Coefficient (or multiplier)
quantisation
(iii) Products of multiplication truncated
or rounded due to machine length
Quantisation

Q
Output
Input
2
) (
2
,
Q
k e
Q
o i
s s
) (k e
i
) (k e
o
The pdf for e using rounding

Noise power
or
2
Q
2
Q
Q
1
}
= =
2
2
2 2 2
} { ). (
Q
Q
e E de e p e o
12
2
2
Q
= o
Let input signal be sinusoidal of unity
amplitude. Then total signal power

If b bits used for binary then
so that
Hence

or dB
2
1
= P
b
Q 2 2 =
3 2
2 2 b
= o
b
P
2 2
2 .
2
3
+
= o
b 6 8 . 1 SNR + =
Consider a simple example of finite
precision on the coefficients a,b of second
order system with poles

where

2 1
1
1
) (

+
=
bz az
z H
u
j
e
2 2 1
. . cos 2 1
1
) (

+
=
z z
z H
u
2
= b
u cos 2 = a

bit pattern
000 0 0
001 0.125 0.354
010 0.25 0.5
011 0.375 0.611
100 0.5 0.707
101 0.625 0.791
110 0.75 0.866
111 0.875 0.935
1.0 1.0 1.0
2
, cos 2 u

Finite wordlength computations

+
INPUT
OUTPU
T
+
+
Limit-cycles; "Effective Pole"
Model; Deadband

Observe that for

instability occurs when
i.e. poles are
(i) either on unit circle when complex
(ii) or one real pole is outside unit
circle.
Instability under the "effective pole" model
is considered as follows
) 1 (
1
) (
2
2
1
1

+ +
=
z b z b
z H
1
2
b
In the time domain with

With for instability we have
indistinguishable from
Where is quantisation

) (
) (
) (
z X
z Y
z H =
) 2 ( ) 1 ( ) ( ) (
2 1
= n y b n y b n x n y
1
2
b
| | ) 2 (
2
n y b Q ) 2 ( n y
| | Q
With rounding, therefore we have

are indistinguishable (for integers)
or
Hence

With both positive and negative numbers

5 . 0 ) 2 (
2
n y b ) 2 ( n y
) 2 ( 5 . 0 ) 2 (
2
= n y n y b
2
1
5 . 0
) 2 (
b
n y
=
2
1
5 . 0
) 2 (
b
n y
=
The range of integers

constitutes a set of integers that cannot be
individually distinguished as separate or from the
asymptotic system behaviour.
The band of integers

is known as the "deadband".
In the second order system, under rounding, the
output assumes a cyclic set of values of the
deadband. This is a limit-cycle.

2
1
5 . 0
b
|
|
.
|
\
|
2 2
1
5 . 0
,
1
5 . 0
b b
Consider the transfer function

if poles are complex then impulse response
is given by

) 1 (
1
) (
2
2
1
1

+ +
=
z b z b
z G
2 2 1 1
=
k k k k
y b y b x y
k
h
| | u
u
) 1 ( sin .
sin
+ = k h
k
k
Where
If then the response is sinusiodal
with frequency

Thus product quantisation causes instability
implying an "effective .
2
b =
|
.
|
\
|
=

2
1
1
2
cos
b
b
u
1
2
= b
|
.
|
\
|
=

2
cos
1
1
1
b
T
e
1
2
= b
Consider infinite precision computations for
2 1
9 . 0

+ =
k k k k
y y x y
0 ; 0
10
0
= =
=
k x
x
k
-10 -5 0 5 10
-10
-8
-6
-4
-2
0
2
4
6
8
10
Now the same operation with integer
precision
-10 -5 0 5 10
-10
-8
-6
-4
-2
0
2
4
6
8
10

Notice that with infinite precision the
response converges to the origin

With finite precision the reponse does not
converge to the origin but assumes
cyclically a set of values the Limit Cycle
Assume , .. are not
correlated, random processes etc.

Hence total output noise power

Where and
{ } ) (
1
k e { } ) (
2
k e
=

=0
2 2 2
0
) (
k
i e i
k h o o
12
2
2 Q
e
= o
| |
+
= + =

=
0
2
2
2
2
2
02
2
01
2
0
sin
) 1 ( sin
.
12
2
. 2
k
k
b
k
u
u
o o o
b
Q

= 2
| |
0 ;
sin
) 1 ( sin
. ) ( ) (
2 1
>
+
= = k
k
k h k h
k
u
u

ie
(
+
+
=

u
o
2 cos 2 1
1
.
1
1
6
2
2 4 2
2 2
2
0
b

For FFT
A(n)
B(n)
B(n+1)
B(n+1)
-
A(n)
B(n+1)
B(n)W(n)
B(n)
A(n+1)
) ( ). ( ) ( ) 1 (
) ( ). ( ) ( ) 1 (
n B n W n A n B
n B n W n A n A
= +
+ = +
W(n)
FFT

AVERAGE GROWTH: 1/2 BIT/PASS
2 ) 1 ( ) 1 (
2 2
= + + + n B n A
) ( 2 ) (
) ( 2 ) 1 (
2 2
n A n A
n A n A
=
= +
FFT

PEAK GROWTH: 1.21.. BITS/PASS
IMAG
REAL
1.0
1.0
-1.0
-1.0
.... 414 . 2 ) ( ) ( 0 . 1
) (
) 1 (
) ( ) ( ) ( ) ( ) ( ) 1 (
) ( ) ( ) ( ) ( ) ( ) 1 (
= + <
+
+ < +
+ = +
n S n C
n A
n A
n S n B n C n B n A n A
n S n B n C n B n A n A
x
x
y x x x
y x x x
Linear modelling of product quantisation

Modelled as
x(n) ) (
~
n x
x(n)
q(n)
+
) ( ) ( ) (
~
n q n x n x + =
| | Q
For rounding operations q(n) is uniform
distributed between , and where Q is
the quantisation step (i.e. in a wordlength of
bits with sign magnitude representation or
mod 2, ).
A discrete-time system with quantisation at
the output of each multiplier may be
considered as a multi-input linear system
2
Q
2
Q
b
Q

= 2

Then

where is the impulse response of the
system from the output of the multiplier
to y(n).
h(n)
) ( )... ( )... (
2 1
n q n q n q
p
{ } ) (n x { } ) (n y
=
=
=
p
r r
r n h r q r n h r x n y
1 0 0
) ( ). ( ) ( ). ( ) (

) (n h
For zero input i.e. we can write

where is the maximum of
which is not more than

ie

n n x = , 0 ) (

s
=
=
p
r
r n h q n y
1 0
) ( . ) (
q
r r q , , ) (

2
Q

s
=
=
p
n
r n h
Q
n y
1 0
) ( .
2
) (

However

And hence

ie we can estimate the maximum swing at
the output from the system parameters and
quantisation level

s
= 0 0
) ( ) (
n n
n h n h
s

=0
) ( .
2
) (
n
n h
pQ
n y

1 Finite Wordlength

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 Finite Wordlength

Uploaded by

Copyright:

Available Formats

Professor A G Constantinides 1

Finite Wordlength Effects

You might also like