Finite Precision Numerical Effects

Finite Precision Numerical Effects
Quote of the Day

When you look at yourself from a universal
standpoint, something inside always reminds or
informs you that there are bigger and better things
to worry about.

Albert Einstein
Coefficient Quantization

2
3
Effects of Coefficient Quantization in IIR Systems
When the parameters of a rational system are quantized
The poles and zeros of the system function move
If the system structure of the system is sensitive to
perturbation of coefficients
The resulting system may no longer be stable
The resulting system may no longer meet the original specs
We need to do a detailed sensitivity analysis
Quantize the coefficients and analyze frequency response
Compare frequency response to original response
We would like to have a general sense of the effect of
quantization
4
Effects on Roots

Each root is affected by quantization errors in ALL coefficient
Tightly clustered roots can be significantly effected
Narrow-bandwidth lowpass or bandpass filters can be very
sensitive to quantization noise
The larger the number of roots in a cluster the more sensitive
it becomes
This is the reason why second order cascade structures are
less sensitive to quantization error than higher order system
Each second order system is independent from each other

( )
=
N
1 k
k
k
M
0 k
k
k
z a 1
z b
z H ( )
=
N
1 k
k
k
M
0 k
k
k
z a 1
z b
z H
Quantization
5
MATLAB Demo
6
Poles of Quantized Second-Order Sections
Consider a 2nd order system with complex-conjugate pole pair

The pole locations after quantization will be on the grid point

3-bits
7-bits
7
Coupled-Form Implementation of Complex-Conjugate Pair
Equivalent implementation of
the second order system

But the quantization grid this
time is
8
Effects of Coefficient Quantization in FIR Systems
No poles to worry about only zeros
Direct form is commonly used for FIR systems

Suppose the coefficients are quantized

Quantized system is linearly related to the quantization error

Again quantization noise is higher for clustered zeros
However, most FIR filters have spread zeros
( ) | |
=
M
0 n
n
z n h z H
( ) | | ( ) ( ) z H z H z n h
z H
M
0 n
n
A + = =

=
( ) | |
A = A
M
0 n
n
z n h z H
Matlab Demo
9
Round-Off Noise

10
11
Round-Off Noise in Digital Filters
Difference equations
implemented with
finite-precision
arithmetic are non-
linear systems
Second order direct
form I system
Model with
quantization effect
Density function
error terms for
rounding
12
Analysis of Quantization Error
Combine all error terms to single location to get

The variance of e[n] in the general case is

The contribution of e[n] to the output is

The variance of the output error term f[n] is

| | | | | |
| | | | | | n e n e n e
n e n e n e
4 3 2
1 0
+ + +
+ =
( )
12
2
N 1 M
B 2
2
e
+ + = o
| | | | | |
=
+ =
N
1 k
k
n e k n f a n f
( ) | |
+ + = o
n
2
ef
B 2
2
f
n h
12
2
N 1 M
( ) ( ) z A / 1 z H
ef
=
13
Round-Off Noise in a First-Order System
Suppose we want to implement the following stable system

The quantization error noise variance is

Noise variance increases as |a| gets closer to the unit circle
As |a| gets closer to 1 we have to use more bits to
compensate for the increasing error

( ) 1 a
az 1
b
z H
1
<
=

( ) | |
|
|
.
|
\
|
= = + + = o

2
B 2
0 n
n 2
B 2
n
2
ef
B 2
2
f
a 1
1
12
2
2 a
12
2
2 n h
12
2
N 1 M
Zero-Input Limit Cycles

14
15
Zero-Input Limit Cycles in Fixed-Point Realization of IIR Filters
For stable IIR systems the output will decay to zero when the
input becomes zero
A finite-precision implementation, however, may continue to
oscillate indefinitely
Nonlinear behaviour very difficult to analyze so we sill study
by example
Example: Limite Cycle Behavior in First-Order Systems

Assume x[n] and y[n-1]
are implemented by 4 bit
registers

| | | | | | 1 a n x 1 n ay n y < + =
16
Example Contd

Assume that a=1/2=0.100b and the input is

If we calculate the output for values of n

A finite input caused an oscilation with period 1
| | | | ( ) | | n b 111 . 0 n
8
7
n x o = o =
n y[n] Q(y[n])
0 7/8=0.111b 7/8=0.111b
1 7/16=0.011100b 1/2=0.100b
2 1/4=0.010000b 1/4=0.010b
3 1/8=0.001000b 1/8=0.001b
4 1/16=0.00010b 1/8=0.001b
| | | | | | 1 a n x 1 n ay n y < + =
MATLAB Demo
17
Limite Cycles due to Overflow

18
19
Example: Limite Cycles due to Overflow
Consider a second-order system realized by

Where Q() represents twos complement rounding
Word length is chosen to be 4 bits
Assume a
1
=3/4=0.110b and a
2
=-3/4=1.010b
Also assume

The output at sample n=0 is

After rounding up we get

Binary carry overflows into the sign bit changing the sign
When repeated for n=1

| | | | | | ( ) | | ( ) 2 n y a Q 1 n y a Q n x n y
2 1
+ + =
| | | | b 010 . 1 4 / 3 2 y and b 110 . 0 4 / 3 1 y = = = =
| |
0.100100b 0.100100b
1.010b b 010 . 1 0.110b b 110 . 0 0 y
+ =
+ =
| | -3/4 1.010b 0.101b 0.101b 0 y = = + =
| | 4 / 3 110 . 0 1.010b 1.010b 0 y = = + =
20
Avoiding Limite Cycles
Desirable to get zero output for zero input: Avoid limit-cycles
Generally adding more bits would avoid overflow
Using double-length accumulators at addition points would
decrease likelihood of limit cycles
Trade-off between limit-cycle avoidance and complexity
FIR systems cannot support zero-input limit cycles

Finite Precision Numerical Effects

Uploaded by

Copyright:

Available Formats

Finite Precision Numerical Effects

Uploaded by

Document Information

Original Description:

Original Title

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

Finite Precision Numerical Effects

Uploaded by

Copyright:

Available Formats

Finite Precision Numerical Effects

Quote of the Day

You might also like