Filter Design Toolbox User`s Guide

Filter Design Toolbox User`s Guide
Filter Design
Toolbox
For Use with MATLAB
®
Computation
Visualization
Programming
User’s Guide
Version 2
How to Contact The MathWorks:
☎
508-647-7000
Phone
508-647-7001
Fax
The MathWorks, Inc.
3 Apple Hill Drive
Natick, MA 01760-2098
Mail
http://www.mathworks.com
Web
Anonymous FTP server
Newsgroup
PHONE
FAX
✉
MAIL
INTERNET
ftp.mathworks.com
comp.soft-sys.matlab
@
E-MAIL
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
Technical support
Product enhancement suggestions
Bug reports
Documentation error reports
Subscribing user registration
Order status, license renewals, passcodes
Sales, pricing, and general information
Filter Design Toolbox User’s Guide
 COPYRIGHT 2000 by The MathWorks, Inc.
The software described in this document is furnished under a license agreement. The software may be used
or copied only under the terms of the license agreement. No part of this manual may be photocopied or reproduced in any form without prior written consent from The MathWorks, Inc.
FEDERAL ACQUISITION: This provision applies to all acquisitions of the Program and Documentation by
or for the federal government of the United States. By accepting delivery of the Program, the government
hereby agrees that this software qualifies as "commercial" computer software within the meaning of FAR
Part 12.212, DFARS Part 227.7202-1, DFARS Part 227.7202-3, DFARS Part 252.227-7013, and DFARS Part
252.227-7014. The terms and conditions of The MathWorks, Inc. Software License Agreement shall pertain
to the government’s use and disclosure of the Program and Documentation, and shall supersede any
conflicting contractual terms or conditions. If this license fails to meet the government’s minimum needs or
is inconsistent in any respect with federal procurement law, the government agrees to return the Program
and Documentation, unused, to MathWorks.
MATLAB, Simulink, Stateflow, Handle Graphics, and Real-Time Workshop are registered trademarks, and
Target Language Compiler is a trademark of The MathWorks, Inc.
Other product or brand names are trademarks or registered trademarks of their respective holders.
Printing History: March 2000
September 2000
New for Version 1.0 (online only)
First Printing Revised for Version 2 (Release 12)
Contents
Preface
What Is Filter Design Toolbox? . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Related Products List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
Using This Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
New Users of This Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
Experienced Users of This Toolbox . . . . . . . . . . . . . . . . . . . . . . . xiv
Organization of This Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Configuration Information . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii
Technical Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii
Typographical Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . xix
Filter Design Toolbox Overview
1
Filter Design Functions in the Toolbox . . . . . . . . . . . . . . . . . 1-4
Quantization Functions in the Toolbox . . . . . . . . . . . . . . . . .
Data Quantizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Quantized Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Quantized Fast Fourier Transforms . . . . . . . . . . . . . . . . . . . . .
1-7
1-8
1-9
1-9
Comparison to the Signal Processing Toolbox . . . . . . . . . . 1-11
Filters in Signal Processing Toolbox . . . . . . . . . . . . . . . . . . . . 1-11
Filters in Filter Design Toolbox . . . . . . . . . . . . . . . . . . . . . . . . 1-13
i
Getting Started with the Toolbox . . . . . . . . . . . . . . . . . . . . . .
Example - Creating a Quantized IIR Filter . . . . . . . . . . . . . . .
Designing the IIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Quantizing the IIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1-15
1-15
1-17
1-20
Selected Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-28
Designing Advanced Filters
2
The Optimal Filter Design Problem . . . . . . . . . . . . . . . . . . . . . 2-2
Optimal Filter Design Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
Optimal Filter Design Solutions . . . . . . . . . . . . . . . . . . . . . . . . . 2-5
Advanced FIR Filter Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
gremez Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8
Advanced IIR Filter Designs . . . . . . . . . . . . . . . . . . . . . . . . . .
iirlpnorm Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iirlpnormc Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iirgrpdelay Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2-37
2-40
2-45
2-51
Robust Filter Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-59
Filter Design Example That Includes Quantization . . . . . . . . 2-62
Selected Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-68
Quantization and Quantized Filtering
3
Binary Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Digital Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Quantized Filter Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Quantized Filter Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ii
Contents
3-3
3-3
3-4
3-4
Data Format for Quantized Filters . . . . . . . . . . . . . . . . . . . . . . . 3-5
Quantized FFTs and Quantized Inverse FFTs . . . . . . . . . . . . . . 3-6
Introductory Quantized Filter Example . . . . . . . . . . . . . . . . . 3-7
Constructing an Eight-Bit Quantized Filter . . . . . . . . . . . . . . . 3-8
Analyzing Poles and Zeros with zplane . . . . . . . . . . . . . . . . . . 3-10
Analyzing the Impulse Response with impz . . . . . . . . . . . . . . . 3-10
Analyzing the Frequency Response with freqz . . . . . . . . . . . . 3-11
Noise Loading Frequency Response Analysis: nlm . . . . . . . . . 3-12
Analyzing Limit Cycles with limitcycle . . . . . . . . . . . . . . . . . . 3-13
Fixed-Point Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Radix Point Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dynamic Range and Precision . . . . . . . . . . . . . . . . . . . . . . . . . .
Overflows and Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3-15
3-16
3-16
3-17
Floating-Point Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Scientific Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The IEEE Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The Fraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The Sign Bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Single-Precision Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Double-Precision Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Custom Floating-point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dynamic Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exceptional Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3-18
3-18
3-19
3-19
3-19
3-19
3-20
3-20
3-21
3-21
3-23
Working with Objects
4
Objects for Quantized Filtering . . . . . . . . . . . . . . . . . . . . . . . . 4-2
Constructing Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
Copying Objects to Inherit Properties . . . . . . . . . . . . . . . . . . . . 4-4
Properties and Property Values . . . . . . . . . . . . . . . . . . . . . . . . 4-5
Setting and Retrieving Property Values . . . . . . . . . . . . . . . . . . . 4-5
iii
Setting Property Values Directly at Construction . . . . . . . . . . .
Setting Property Values with the set Command . . . . . . . . . . . .
Retrieving Properties with the get Command . . . . . . . . . . . . . .
Direct Property Referencing to Set and Get Values . . . . . . . . . .
4-5
4-6
4-8
4-9
Functions Acting on Objects . . . . . . . . . . . . . . . . . . . . . . . . . . 4-10
Using Command Line Help . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-11
Command Line Help For Nonoverloaded Functions . . . . . . . . 4-11
Command Line Help For Overloaded Functions . . . . . . . . . . . 4-11
Using Cell Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-13
Indexing into a Cell Array of Vectors or Matrices . . . . . . . . . . 4-13
Indexing into a Cell Array of Cell Arrays . . . . . . . . . . . . . . . . . 4-14
Working with Quantizers
5
Quantizers and Unit Quantizers . . . . . . . . . . . . . . . . . . . . . . . . 5-2
Constructing Quantizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-3
Constructor for Quantizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-3
Quantizer Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Properties and Property Values . . . . . . . . . . . . . . . . . . . . . . . . .
Settable Quantizer Properties . . . . . . . . . . . . . . . . . . . . . . . . . . .
Setting Quantizer Properties Without Naming Them . . . . . . . .
Read-Only Quantizer Properties . . . . . . . . . . . . . . . . . . . . . . . . .
5-4
5-4
5-4
5-5
5-5
Quantizing Data with Quantizers . . . . . . . . . . . . . . . . . . . . . . . 5-7
Example — Data-Related Quantizer Information . . . . . . . . . . . 5-7
Transformations for Quantized Data . . . . . . . . . . . . . . . . . . . . 5-9
Quantizer Data Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-10
iv
Contents
Working with Quantized Filters
6
Constructing Quantized Filters . . . . . . . . . . . . . . . . . . . . . . . .
Constructor for Quantized Filters . . . . . . . . . . . . . . . . . . . . . . . .
Constructing a Quantized Filter from a Reference . . . . . . . . . .
Copying Filters to Inherit Properties . . . . . . . . . . . . . . . . . . . . .
Changing Filter Property Values After Construction . . . . . . . .
6-3
6-3
6-4
6-5
6-5
Quantized Filter Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
Properties and Property Values . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
Basic Filter Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
Specifying the Filter’s Reference Coefficients . . . . . . . . . . . . . . 6-7
Specifying the Quantized Filter Structure . . . . . . . . . . . . . . . . . 6-8
Specifying the Data Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-9
Specifying All Data Format Properties at Once . . . . . . . . . . . . 6-10
Specifying the Format Parameters with setbits . . . . . . . . . . . . 6-11
Using normalize to Scale Coefficients . . . . . . . . . . . . . . . . . . . . 6-12
Filtering Data with Quantized Filters . . . . . . . . . . . . . . . . . . 6-14
Transformation Functions for
Quantized Filter Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . 6-15
Working with Quantized FFTs
7
Constructing Quantized FFTs . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
Constructor for Quantized FFTs . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
Copying Quantized FFTs to Inherit Properties . . . . . . . . . . . . . 7-4
Quantized FFT Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Properties and Property Values . . . . . . . . . . . . . . . . . . . . . . . . .
Basic Quantized FFT Properties . . . . . . . . . . . . . . . . . . . . . . . . .
Specifying the Data Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Specifying All Data Format Properties at Once . . . . . . . . . . . . .
Specifying the Format Parameters with setbits . . . . . . . . . . . . .
7-6
7-6
7-6
7-7
7-8
7-8
v
Computing a Quantized FFT or Inverse FFT of Data . . . . 7-10
Quantized Filtering Analysis Examples
8
Example — Quantized Filtering of Noisy Speech . . . . . . . . . 8-3
Loading a Speech Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-3
Analyzing the Frequency Content of the Speech . . . . . . . . . . . . 8-4
Adding Noise to the Speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
Creating a Filter to Extract the 3000Hz Noise . . . . . . . . . . . . . 8-5
Quantizing the Filter As a Fixed-Point Filter . . . . . . . . . . . . . . 8-8
Normalizing the Quantized Filter Coefficients . . . . . . . . . . . . . 8-8
Analyzing the Filter Poles and Zeros Using zplane . . . . . . . . . . 8-9
Creating a Filter with Second-Order Sections . . . . . . . . . . . . . 8-11
Quantized Filter Frequency Response Analysis . . . . . . . . . . . 8-12
Filtering with Quantized Filters . . . . . . . . . . . . . . . . . . . . . . . . 8-13
Analyzing the filter Function Logged Results . . . . . . . . . . . . . 8-14
Example — A Quantized Filter Bank . . . . . . . . . . . . . . . . . . . 8-16
Filtering Data with the Filter Bank . . . . . . . . . . . . . . . . . . . . . 8-17
Creating a DFT Polyphase FIR Quantized Filter Bank . . . . . 8-17
Example — Effects of Quantized Arithmetic . . . . . . . . . . . .
Creating a Quantizer for Data . . . . . . . . . . . . . . . . . . . . . . . . .
Creating a Fixed-Point Filter from a Quantized Reference . . .
Creating a Double-Precision Quantized Filter . . . . . . . . . . . . .
Quantizing a Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Filtering the Quantized Data with Both Filters . . . . . . . . . . .
Comparing the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vi
Contents
8-22
8-22
8-22
8-23
8-23
8-23
8-24
Quantization Tool Overview
9
Switching FDATool to Quantization Mode . . . . . . . . . . . . . . . 9-3
Getting Help for FDATool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
Context-Sensitive Help: The What’s This? Option . . . . . . . . . . . 9-5
Additional Help for FDATool . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
Quantizing Filters in the Filter Design and Analysis Tool . 9-6
To Quantize Reference Filters . . . . . . . . . . . . . . . . . . . . . . . . . . 9-10
To Change the Quantization Properties of Quantized Filters . 9-11
Choosing Your Quantized Filter Structure . . . . . . . . . . . . . 9-12
Converting the Structure of a Quantized Filter . . . . . . . . . . . . 9-12
To Change the Structure of a Quantized Filter . . . . . . . . . . . . 9-13
Scaling Transfer Function Coefficients . . . . . . . . . . . . . . . . 9-16
To Scale Transfer Function Coefficients . . . . . . . . . . . . . . . . . . 9-16
Scaling Inputs and Outputs of Quantized Filters . . . . . . . . 9-18
To Enter Scale Values for Quantized Filters . . . . . . . . . . . . . . 9-18
Importing and Exporting Quantized Filters . . . . . . . . . . . . 9-20
To Import Quantized Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-21
To Export Quantized Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-22
Property Reference
10
A Quick Guide to Quantizer Properties . . . . . . . . . . . . . . . . 10-2
Quantizer Properties Reference . . . . . . . . . . . . . . . . . . . . . . .
Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Max . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Min . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10-3
10-3
10-5
10-5
10-5
vii
NOperations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
NOverflows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
NUnderflows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
OverflowMode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
RoundMode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10-6
10-6
10-6
10-7
10-8
A Quick Guide to Quantized Filter Properties . . . . . . . . . 10-10
Quantized Filter Properties Reference . . . . . . . . . . . . . . . .
CoefficientFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
FilterStructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
InputFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
NumberOfSections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MultiplicandFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
OutputFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ProductFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
QuantizedCoefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ReferenceCoefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ScaleValues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
StatesPerSection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
SumFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10-11
10-11
10-12
10-32
10-32
10-32
10-33
10-33
10-33
10-34
10-42
10-44
10-44
A Quick Guide to Quantized FFT Properties . . . . . . . . . . . 10-46
Quantized FFT Properties Reference . . . . . . . . . . . . . . . . .
CoefficientFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
InputFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
NumberOfSections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MultiplicandFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
OutputFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ProductFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Radix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ScaleValues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
SumFormat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
viii Contents
10-47
10-47
10-47
10-48
10-48
10-48
10-49
10-49
10-49
10-49
10-50
Function Reference
11
Filter Design Toolbox Functions . . . . . . . . . . . . . . . . . . . . . . 11-2
Functions Operating on Quantized Filters . . . . . . . . . . . . . 11-8
Functions Operating on Quantizers . . . . . . . . . . . . . . . . . . . 11-10
Functions Operating on Quantized FFTs . . . . . . . . . . . . . . 11-12
Functions for Designing Digital Filters . . . . . . . . . . . . . . . 11-14
Filter Design Toolbox Functions Listed Alphabetically . 11-15
Bibliography
12
ix
Preface
What Is Filter Design Toolbox?
. . . . . . . . . . . . x
Related Products List . . . . . . . . . . . . . . . . . xi
Using This Guide . . . . . . .
New Users of This Toolbox . . . .
Experienced Users of This Toolbox
Organization of This Guide . . .
Configuration Information
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.xii
.xii
xiii
xiv
. . . . . . . . . . . . . xvi
Technical Conventions . . . . . . . . . . . . . . . xvii
Typographical Conventions . . . . . . . . . . . . . xviii
What Is Filter Design Toolbox?
Filter Design Toolbox is a collection of tools built on top of the MATLAB®
computing environment and the Signal Processing Toolbox. The toolbox
includes a number of advanced filter design techniques that support designing,
simulating, and analyzing fixed-point and custom floating-point filters for a
wide range of precisions.
Note A preliminary version of Filter Design Toolbox, Version 2 was released
as Quantized Filtering Toolbox, Version 1.
xi
Preface
Related Products List
The MathWorks provides several products that are especially relevant to the
tasks you perform with Filter Design Toolbox.
For more information about any of these products, refer to:
• The online documentation for that product, if it is installed or if you are
reading the documentation from the documentation CD
• The MathWorks Web site, at http://www.mathworks.com; navigate to the
“products” section
Note The toolboxes listed below include functions that extend MATLAB
capabilities. The blocksets include blocks that extend Simulink® capabilities.
xii
Product
Description
DSP Blockset
Simulink block libraries for the design,
simulation, and prototyping of digital signal
processing systems
Fixed-Point Blockset
Simulink blocks that model, simulate, and
automatically generate pure integer code for
fixed-point applications
Signal Processing
Toolbox
Tool for algorithm development, signal and
linear system analysis, and time-series data
modeling
Simulink
Interactive, graphical environment for
modeling, simulating, and prototyping
dynamic systems
Using This Guide
Using This Guide
All users of the toolbox should read this guide. You should be generally familiar
with basic digital signal processing concepts before you use the toolbox and this
User’s Guide. The quantization portion of this toolbox assumes some
familiarity with fixed-point and floating-point arithmetic in the context of
digital filtering applications.
New Users of This Toolbox
You can use this toolbox to:
• Design filters using advanced design methods
• Convert filters to and from coupled-allpass forms
• Convert filters to second-order section form
• Quantize filters and filter data
• Quantize data
• Compute quantized FFTs and IFFTs
This toolbox relies on object-oriented programming techniques using objects for
quantized filtering and analysis. You do not need to be familiar with these
techniques to use this toolbox. However, you may want to review the concepts
of MATLAB structures and cell arrays, as these are used in the syntax for
several toolbox methods. For more information on MATLAB structures and cell
arrays, refer to “Programming and Data Types” in your MATLAB
documentation.
As a new user of this toolbox, read the entire guide. Of particular interest are:
• Chapter 2, “Designing Advanced Filters” for its background information on
the advanced filter design techniques in this toolbox
• Chapter 3, “Quantization and Quantized Filtering” for its background
information on fixed-point and floating-point filters
• Chapter 4, “Working with Objects” for an introduction to the object-oriented
techniques you need for this toolbox
• Chapter 5, “Working with Quantizers” for information on constructing and
using quantizers
xiii
Preface
• Chapter 6, “Working with Quantized Filters” for information on constructing
and using quantized filters
• Chapter 7, “Working with Quantized FFTs” for information on constructing
and using quantized FFTs
• “Example — Quantized Filtering of Noisy Speech” on page 8-3 for a detailed
example of designing and analyzing a fixed-point filter
• “Example — A Quantized Filter Bank” on page 8-16 for an example of
designing and analyzing a fixed-point polyphase DFT filter bank
• Chapter 9, “Quantization Tool Overview” for information about using Filter
Design and Analysis Tool to quantize filters and investigate the effects of
quantization on filter performance
• “Quantizer Properties Reference” on page 10-3 for a description of the
quantizer properties
• “Quantized Filter Properties Reference” on page 10-11 for a description of the
quantized filter properties
• “Quantized FFT Properties Reference” on page 10-47 for a description of the
quantized FFT properties
• “Filter Design Toolbox Functions” on page 11-2 for a description of every
function in the toolbox
Experienced Users of This Toolbox
As an experienced user of this toolbox, you may find the following sections to
be useful reference guides for the toolbox:
• “Quantizer Properties Reference” on page 10-3
• “Quantized Filter Properties Reference” on page 10-11
• “Quantized FFT Properties Reference” on page 10-47
• “Filter Design Toolbox Functions” on page 11-2
xiv
Using This Guide
Organization of This Guide
This guide is organized as follows.
Chapter Title
Description
“Filter Design Toolbox Overview”
Offers an overview of the toolbox and an example to get
you started using toolbox features and functions
“Designing Advanced Filters”
Provides background information on the advanced filter
design methods in this toolbox
“Quantization and Quantized
Filtering”
Introduces:
• The concepts of quantization and filtering
• An example of using, creating, and analyzing quantized
filters
• Some tutorial information on fixed- and floating-point
arithmetic
“Working with Objects”
Introduces the object-oriented programming techniques
relevant to this toolbox
“Working with Quantizers”
Provides information about constructing and using
quantizers
“Working with Quantized Filters”
Covers quantized-filter specific characteristics and
analysis techniques
“Working with Quantized FFTs”
Introduces constructing and using quantized FFTs
“Quantized Filtering Analysis
Examples”
Presents approaches to solving some applied problems
with this toolbox
“Quantization Tool Overview”
Presents a detailed reference covering the quantization
page of the Filter Design and Analysis Tool
xv
Preface
Chapter Title
Description (Continued)
“Property Reference”
Provides:
• A summary of the quantized filter properties
• A detailed quantized filter property reference, including
descriptions of the filter structures
“Function Reference” (online only)
Provides:
• Tables that include short descriptions of the functions in
this toolbox
• A detailed alphabetical function reference
“Bibliography”
xvi
Lists references for quantized filtering
Configuration Information
Configuration Information
To determine whether Filter Design Toolbox is installed on your system, type
this command at the MATLAB prompt.
ver
When you enter this command, MATLAB displays information about the
version of MATLAB you are running, including a list of all toolboxes installed
on your system and their version numbers.
For information about installing the toolbox, refer to the MATLAB Installation
Guide for your platform.
Note For up-to-date information about system requirements, visit the system
requirements page, available in the products area at the MathWorks Web site
(www.mathworks.com).
xvii
Preface
Technical Conventions
This manual and the functions in Filter Design Toolbox use the following
technical notations.
xviii
Nyquist frequency
One-half the sampling frequency. Some Signal
Processing Toolbox functions normalize this to 1.
x(1)
The first element of a data sequence or filter,
corresponding to zero lag.
w (used in syntax
examples)
Digital frequency in radians per sample.
f (used in syntax
examples)
Digital frequency in hertz.
[x, y)
The interval from x to y, including x but not
including y.
... (used in
syntax examples)
Ellipses in the argument list for a given syntax on a
function reference page. These indicate that all
argument options listed prior to the current syntax
are valid for the function.
Typographical Conventions
Typographical Conventions
This manual uses some or all of these conventions.
Item
Convention to Use
Example
Example code
Monospace font
To assign the value 5 to A,
enter
A = 5
Function names/syntax
Monospace font
The cos function finds the
cosine of each array element.
Syntax line example is
MLGetVar ML_var_name
Keys
Boldface with an initial
Press the Enter key.
capital letter
Literal strings (in syntax
descriptions in Reference
chapters)
Mathematical
expressions
MATLAB output
Monospace bold for
f = freqspace(n,'whole')
literals.
Variables in italics
Functions, operators, and
constants in standard text.
Monospace font
This vector represents the
polynomial
p = x2 + 2x + 3
MATLAB responds with
A =
5
Menu names, menu items, and
controls
Boldface with an initial
capital letter
Choose the File menu.
New terms
Italics
An array is an ordered
collection of information.
String variables (from a finite
list)
Monospace italics
sysc = d2c(sysd, 'method')
xix
Preface
xx
1
Filter Design Toolbox
Overview
Filter Design Functions in the Toolbox . . . . . . . . 1-4
Quantization Functions in the Toolbox
Data Quantizers . . . . . . . . . . .
Quantized Filters . . . . . . . . . .
Quantized Fast Fourier Transforms . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1-7
1-8
1-9
1-9
Comparison to the Signal Processing Toolbox . . . . . 1-11
Filters in Signal Processing Toolbox . . . . . . . . . . . 1-11
Filters in Filter Design Toolbox . . . . . . . . . . . . . 1-13
Getting Started with the Toolbox . .
Example - Creating a Quantized IIR Filter
Designing the IIR Filter . . . . . . . .
Quantizing the IIR Filter . . . . . . .
Selected Bibliography
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1-15
1-15
1-17
1-20
. . . . . . . . . . . . . . . 1-28
1
Filter Design Toolbox Overview
When you install Filter Design Toolbox in your MATLAB® environment, you
can perform digital filter design, fixed- and floating-point filter quantization,
and filter performance analysis on your desktop computer. But what are
filtering and quantization and what benefits do they provide?
Designers use filtering and its variant, digital filtering, for many tasks:
• To separate signals that have been combined, such as a musical recording
and the noise added during the recording process
• To separate signals into their constituent frequencies
• To demodulate signals
• To restore signals that have been degraded by some process, known or
unknown
You can use analog filters to accomplish these tasks, but digital filters offer
greater flexibility and accuracy than analog filters. In addition, digital signal
processing (DSP) depends in large measure on digital filtering to meet the
needs of its users.
Analog filters can be cheaper, faster, and have greater dynamic range; digital
filters outstrip their analog cousins in flexibility. The ability to create filters
that have arbitrary shape frequency response curves, and filters that meet
performance constraints, such as bandpass width and transition region width,
is well beyond that of analog filters.
Quantization is a natural outgrowth of digital filtering and digital signal
processing development. Also, there is a growing need for fixed-point filters
that meet power, cost, and size restrictions. When you convert a filter from
floating-point to fixed-point, you use quantization to perform the conversion.
As filter designers began to use digital filters in applications where power
limitations and size constraints drove the filter design, they moved from
double-precision, floating-point filters to fixed-point filters. When you have
enough power to run a floating-point digital signal processor, such as on your
desktop PC or in your car, fixed-point processing and filtering are unnecessary.
But, when your filter needs to run in a cellular phone, or you want to run a
hearing aid for hours instead of seconds, fixed-point processing can be essential
to ensure long battery life and small size.
Filter Design Toolbox provides the functions you need to develop filters that
meet the needs of fixed-point algorithms and electronics systems. In addition
to offering tools for analyzing the effects of quantization on filter performance
1-2
and signal processing performance, the toolbox offers filter structures for you
to use to develop prototype filter designs. With structures ranging from finite
impulse response (FIR) filters to infinite impulse response (IIR) filters, you can
investigate alternative fixed-point realizations of filters that might meet your
goals.
This section contains the following subsections introducing filter design:
• “Filter Design Functions in the Toolbox” on page 1-4
• “Quantization Functions in the Toolbox” on page 1-7
• “Comparison to the Signal Processing Toolbox” on page 1-11
• “Getting Started with the Toolbox” on page 1-15
• “Selected Bibliography” on page 1-28
1-3
1
Filter Design Toolbox Overview
Filter Design Functions in the Toolbox
In a system that has unlimited power and size, any filter structure that met
your performance specifications would do. You would design a floating-point
filter whose frequency response achieved your aims and implement that filter
in your system.
When you need a fixed-point filter to meet your requirements, the filter
structure you choose can depend very much on how quantization affects the
performance of the filter. Filter Design Toolbox offers both FIR and IIR filter
design tools and structures that let you experiment with multiple filter designs
to see how each responds to quantization effects.
Filter Structures
The following tables detail some of the quantized FIR and IIR filter structures
available in the toolbox. For lists of all the architectures available in the
toolbox, refer to the section “Quantized Filter Properties Reference” on page
10-11 in this guide.
1-4
FIR Filter Structures
Description
'antisymmetricfir’
Antisymmetric finite impulse response (FIR)
'fir'
Finite impulse response (FIR)
'firt'
Transposed finite impulse response (FIR)
'latticema'
Moving average (MA) lattice form
'symmetricfir'
Symmetric FIR
IIR Filter Structures
Description
'df1'
Direct form I
'df1t'
Direct form I transposed
'df2'
Direct form II
'df2t'
Direct form II transposed
Filter Design Functions in the Toolbox
IIR Filter Structures (Continued)
Description
'latticeca'
Coupled allpass lattice
'latticecapc'
Power-complementary output coupled
allpass lattice form
'latticear'
Autoregressive (AR) lattice form
'latticearma'
Autoregressive, moving average (ARMA)
lattice form
'statespace'
Single-input/single-output state-space
Each of the structures supports floating-point or fixed-point realizations, and
you use the same toolbox function, qfilt, to create each one. To review
schematics of the filter structures available in this toolbox, perform the
following steps to run the demo “Quantized Filter Construction” in the Filter
Design folder in MATLAB demos.
To run the filter construction demo.
1 Enter demo at the MATLAB command line prompt.
The MATLAB Demo window opens on the desktop.
2 Double-click the entry Toolboxes in the left pane. The list of available
toolboxes appears in the left pane.
3 Click Filter Design.
1-5
1
Filter Design Toolbox Overview
4 Click Quantized Filter Construction in the list of demos on the lower
right.
5 Click Run Quantized Filter to run the demonstration model.
To access these demos directly from the MATLAB command line, enter
qfiltconstruction at the prompt.
1-6
Quantization Functions in the Toolbox
Quantization Functions in the Toolbox
Designing floating-point filters solves only part of the filter design problem. In
most cases, floating-point filter realizations are not appropriate for digital
signal processing applications. Many real-world DSP systems require that
their filters use minimum power, generate minimum heat, and do not induce
computational overload in their processors. Meeting these constraints often
means using fixed-point filters. Unfortunately, converting a floating-point
filter to fixed-point realization (called quantizing) can result in lost filter
performance and accuracy. To simulate and determine the effects of
quantization, and allow you to investigate how switching from floating-point to
fixed-point arithmetic affects the performance of your filter, the toolbox
includes quantization functions. You use the toolbox quantization functions for
constructing, applying, and analyzing quantizers, quantized filters, and
quantized fast Fourier transforms (FFT).
The following sections introduce the quantization functions in the toolbox. You
can find details about the functions in these sections:
• “Quantizer Properties Reference” on page 10-3
• “Quantized Filter Properties Reference” on page 10-11
• “Quantized FFT Properties Reference” on page 10-47
As you read the sections about the properties, you will see that quantizers,
quantized filters, and quantized FFTs share common properties and methods.
At the lowest level, a quantizer forms the basis of all the quantizers in the
toolbox. Each property of a quantized object is an instantiation of a data
quantizer. The relationship between quantizers, quantized filters, quantized
FFTs, and their underlying quantizer is shown in the following figure.
1-7
1
Filter Design Toolbox Overview
Quantized Objects
Quantizer
Quantized Filter
Quantized FFT
Data Quantizer
Coefficient
Quantizer
Input
Quantizer
Output
Quantizer
Multiplicand
Quantizer
Product
Quantizer
Sum
Quantizer
Figure 1-1: Unified Modeling Language Diagram for Filter Design Toolbox
Objects
Data Quantizers
To determine how quantization affects a signal, you construct quantizers that
you use to quantize a signal or data set in MATLAB. By adjusting the
quantization parameters of your quantizer, you can investigate the output
from various quantization schemes when you apply them to a data set or
signal. In addition to experimenting with data quantization, quantizers
determine how quantized filters and quantized FFTs quantize data to which
they are applied.
Each quantizer you construct has the following properties that you can set
when you construct the quantizer:
• format — determines the quantization format properties
• mode — determines the arithmetic data type
• overflowmode — determines how overflows are handled during arithmetic
operations
• roundmode — determines the rounding method applied to data values
1-8
Quantization Functions in the Toolbox
When you apply a quantizer, five more properties report the results of the
operation:
• max — reports the maximum value encountered while quantizing input
signals
• min — reports the minimum value encountered while quantizing input
signals
• noperations — reports the total number of quantized operations performed
while quantizing input signals
• noverflows — reports the total number of overflows, both negative and
positive, that occurred while quantizing input signals
• nunderflows — reports the number of underflows that occurred while
quantizing input signals
You cannot set these properties; they are read-only, and reflect the results of
all the quantization operations that you perform with a given quantizer. Use
reset to return quantizers to their initial settings.
Quantized Filters
Quantization, or the effect of word length on filter performance, can lead to
erroneous behavior in filter designs. Finite word lengths can change the
frequency response of a filter from its desired performance. To help you
investigate quantization effects that occur during filtering, the toolbox
provides two ways to construct a quantized filter:
• Use the function qfilt to create a default quantized filter.
• Use qfilt and specify a reference filter to quantize as an input argument.
In both techniques, your quantized filters have the same properties as
quantizers.
Quantized Fast Fourier Transforms
In developing digital signal processing (DSP) algorithms, the fast Fourier
transform (FFT) is one of the essential building blocks. It may be the most
common transform for handling data and signals. To implement an FFT on a
fixed-point DSP, you must consider the effects of word length on the output of
the transform, in much the same way that you must consider the quantization
effects in a digital filter. Filter Design Toolbox includes a quantized FFT (qfft)
1-9
1
Filter Design Toolbox Overview
function that you use to construct and apply quantized FFTs to signals and
data in MATLAB. To help you investigate quantization effects that occur
during the FFT, the toolbox provides you two ways to construct quantized
FFTs:
• Use the function qfft to create a default quantized fast Fourier transform.
• Use qfft and specify a reference filter to quantize as an input argument.
In both techniques, your quantized FFTs have the same properties as
quantizers and quantized filters.
Quantized FFTs have other properties as well; some you can set and some are
read-only:
• length — determines the length of the FFT. Must be a power of the radix
• numberofsections — a read-only property reporting the number of sections
in your quantized FFT
• radix — indicates the form of the FFT to use
• scalevalues — specifies the scaling for the input for each stage of the FFT
1-10
Comparison to the Signal Processing Toolbox
Comparison to the Signal Processing Toolbox
You use Signal Processing Toolbox and Filter Design Toolbox to design and
analyze filters. Filter Design Toolbox offers advanced filter design methods for
FIR and IIR filters that enhance the functionality of Signal Processing Toolbox.
Filters in Signal Processing Toolbox
Signal Processing Toolbox is data-oriented. You create separate variables for
each parameter required to characterize a given filter type. For instance, to
specify a state-space realization of a filter, you need four variables: one for each
of the four parameters that characterize a state-space model.
Filters you design in Signal Processing Toolbox are in double-precision. You
cannot design single-precision, custom-precision, or fixed-point filters. The
filter design methods in Signal Processing Toolbox are listed in the following
tables. Each table includes brief descriptions of the methods and functions,
separated into IIR and FIR architectures:
• Table 1-1 — describes IIR filter design methods
• Table 1-2 — describes filter order estimation functions
• Table 1-3 — describes FIR filter design methods
Table 1-1: Available IIR Filter Design Methods in Signal Processing Toolbox
IIR Filter Design—Classical and Direct
besself
Bessel analog filter design
butter
Butterworth analog and digital filter design
cheby1
Chebyshev type I filter design (passband ripple)
cheby2
Chebyshev type II filter design (stopband ripple)
ellip
Elliptic (Cauer) filter design
maxflat
Generalized digital Butterworth filter design
yulewalk
Recursive digital filter design
1-11
1
Filter Design Toolbox Overview
Table 1-2: Filter Order Estimation Functions in Signal Processing Toolbox
IIR Filter Order Estimation
buttord
Calculate the order and cutoff frequency for a
Butterworth filter
cheb1ord
Calculate the order for a Chebyshev type I filter
cheb2ord
Calculate the order for a Chebyshev type II filter
ellipord
Calculate the minimum order for elliptic filters
remezord
Parks-McClellan optimal FIR filter order estimation
Table 1-3: FIR Filter Design Methods in Signal Processing Toolbox
1-12
FIR Filter Design
Description
cremez
Complex and nonlinear-phase equiripple FIR filter
design
fir1
Design a window-based finite impulse response filter
fir2
Design a frequency sampling-based finite impulse
response filter
fircls
Constrained least square FIR filter design for
multiband filters
fircls1
Constrained least square filter design for lowpass and
highpass linear phase FIR filters
firls
Least square linear-phase FIR filter design
firrcos
Raised cosine FIR filter design
intfilt
Interpolation FIR filter design
kaiserord
Estimate parameters for an FIR filter design with
Kaiser window
remez
Compute the Parks-McClellan optimal FIR filter design
Comparison to the Signal Processing Toolbox
Table 1-3: FIR Filter Design Methods in Signal Processing Toolbox (Continued)
FIR Filter Design
Description
remezord
Parks-McClellan optimal FIR filter order estimation
sgolay
Savitzky-Golay filter design
Filters in Filter Design Toolbox
To help you create and analyze quantized filters, Filter Design Toolbox is
object-oriented. You encapsulate the parameters needed to specify your
quantized filter under one variable name in a quantized filter object. To specify
the parameters associated with a quantized filter, you set the property values
for its associated named properties. These properties are assigned to the
quantized filter object that represents your quantized filter.
You can design a wide range of fixed-point and custom floating-point filters in
Filter Design Toolbox. You use the double-precision filters you design in Signal
Processing Toolbox and Filter Design Toolbox as reference filters to create
quantized filters in this toolbox. To develop a quantized filter, use either
toolbox to create a double-precision filter that meets your requirements, then
use the quantization functions in this toolbox to convert the double-precision
filter to a quantized filter.
Refer to Table 1-4 for a list of the filter design methods in this toolbox.
Table 1-4: Filter Design Methods in the Toolbox—FIR and IIR
Filter Function
Filter Description
firlpnorm
Design minimax solution FIR filters using the least-pth
algorithm
gremez
Use the generalized Remez exchange algorithm to
design optimal solution FIR filters with arbitrary
response curves
iirgrpdelay
Design optimal solution IIR filters where you specify
the group delay in the passband frequencies
1-13
1
Filter Design Toolbox Overview
Table 1-4: Filter Design Methods in the Toolbox—FIR and IIR (Continued)
Filter Function
Filter Description
iirlpnorm
Design minimax solution IIR filters using the least-pth
algorithm
iirlpnormc
Design minimax solution IIR filters using the least-pth
algorithm. In addition, restrict the filter poles and
zeros to lie within a fixed radius around the origin of
the z-plane
You can construct these filters as single-precision, double-precision,
custom-precision floating-point, or fixed-point structures.
1-14
Getting Started with the Toolbox
Getting Started with the Toolbox
This section provides an example to get you started using Filter Design
Toolbox. You can run the code in this example from the Help browser (select
the code, right-click the selection, and choose Evaluate Selection from the
context menu) or you can enter the code on the command line. This exercise
also introduces Filter Design and Analysis Tool (FDATool). You use it to design
and analyze filters, and to quantize filters.
As you follow the example, you are introduced to some of the basic tasks of
designing a filter and using FDATool. You will engage some of the quantization
capabilities of the toolbox, and a few of the filter design architectures provided
as well.
Before you begin this example, start MATLAB and verify that you have
installed Signal Processing and Filter Design Toolboxes (type ver at the
command prompt). You should see Filter Design Toolbox, version 2.0 and
Signal Processing Toolbox, version 5.0, among others, in the list of installed
products.
Example - Creating a Quantized IIR Filter
Example Background. Wireless communications technologies, such as cellular
telephones, need to account for the receiver’s motion relative to the transmitter
and for path changes between the stations. To model the channel fading and
frequency shifting that occurs when the receiver is moving, wireless
communications models apply a lowpass filter to the transmitted signal. With
a narrow passband of 0 to 40Hz that modifies the transmitted signal, the
lowpass filter simulates the Doppler shift caused by the motion between the
transmitter and receiver. As the lowpass filter requires a rather peculiar rising
shape across the passband and an extremely sharp transition region, designing
and quantizing the filter presents an interesting study in filter design. In
Figure 1-2, you see the frequency response curve for the RFC filter. Notice the
narrow passband with the rising shape and the sharp cutoff transition. Also
note that the y-axis is a linear scale that dramatizes the shape of the passband.
1-15
1
Filter Design Toolbox Overview
Distinctive passband
shape to match the
Doppler shift effects
40 Hz cutoff frequency
with sharp transition
Figure 1-2: Frequency Response of the Filter Used to Simulate the Rayleigh
Fading Channel Phenomenon
To create a filter with the passband shape in the figure, we define four vectors
that describe the shape.
1-16
Vector
Definition
Frequency
vector
Specifies frequency points along the frequency response
curve. frequency can be in Hz or normalized. In our
example, we are using normalized entries.
Edge vector
Specifies the edges, in Hz or in normalized values, of the
passband and stopband for the filter. In our example, we are
using normalized entries.
Magnitude
vector
Specifies the filter response magnitude at each frequency
specified in the frequency vector. These values produce the
distinctive passband we require.
Weight
vector
Specifies the weighting for each frequency in the frequency
vector.
Getting Started with the Toolbox
Most filter designs do not require you to define four vectors to specify the filter
response. Because the passband of the filter we want is not standard, we are
going to use the Arbitrary Magnitude filter type in FDATool when we design
our filter. This type requires four input vectors to specify the filter. You can also
design filters with more normal passband specifications directly in FDATool.
You can enter the four vectors in FDATool, but long vectors are easier to enter
at the command line. If the vectors exist as files, you can use MATLAB
commands to import the vectors into your MATLAB workspace.
Designing the IIR Filter
Start to design the filter by clearing the MATLAB workspace and defining the
four required vectors:
1 Clear your MATLAB workspace of all variables and close all your open
figure windows. Enter
clear; close all;
2 At the MATLAB prompt, enter the following commands to create the four
vectors that define the desired IIR filter frequency response.
PBfreq = 0:.0005:.0175; % Define the passband frequencies
Now specify the amplitude at each passband frequency. We use the right
array divide operator (./) to perform element-wise division.
PBamp = .4845 ./ (1-(PBfreq ./ 0.0179).^2).^0.25;
Use PBfreq and PBamp to generate the final frequency F and amplitude A
vectors for our IIR filter. While defining these vectors, define edges and W,
the edge and weight vectors.
F = [PBfreq .02 .0215 .025 1];
edges = [0 .0175 .02 .0215 .025 1];
A = [PBamp 0 0 0 0];
W = [ones(1,length(A)-1) 300];
Issuing these commands created four vectors in your MATLAB workspace.
FDATool uses these vectors to create an IIR lowpass filter with a specified
magnitude response curve. Vectors F and A each contain 40 elements, and
vectors W and edges contain 40 and 6 elements. If we were not designing
1-17
1
Filter Design Toolbox Overview
a specific passband shape, you would not have needed to define these
vectors.
3 Open FDATool by typing fdatool at the command prompt.
FDATool opens in Design Filter mode.
4 Under Filter Type, select Arbitrary Magnitude from the list.
Although we want a lowpass filter, Lowpass does not let us specify the
shape of the passband. So we use the Arbitrary Magnitude option to get
precisely the curve we need. You could plot F and A to see that the curve is
similar to the response in Figure 1-2. Use the command plot(F,A) to view
a simple plot of the specified passband shape.
When you select Arbitrary Magnitude from the list, the options under
Filter Specification change to require four vectors: Freq. vector, Freq.
edges, Mag. vector, and Weight vector.
5 Under Filter Specifications, select Normalized (0 to 1) from the
Frequency Units list.
1-18
Getting Started with the Toolbox
6 Under Filter Specifications, enter the variable names that define the four
vectors required to specify the filter response.
Freq. vector, Freq. edges, Mag. vector, and Weight vector: F, edges, A,
and W.
Required Vector
Variable
Freq. vector
F
Freq. edges
edges
Mag. vector
A
Weight vector
W
7 Specify the filter order by entering 8 for the numerator and denominator
orders under Filter Order.
8 Complete the IIR filter design by selecting IIR under Design Method,
choosing Constrained least Pth-norm from the list.
9 Click Design Filter.
FDATool designed the filter and computed the filter response. In the
analysis area, you see the magnitude response of the filter displayed on a
logarithmic scale.
1-19
1
Filter Design Toolbox Overview
In the upper left corner, the plot shows the region of interest for this filter.
Click
on the FDATool toolbar and use the zoom feature to inspect the
filter passband between 0 and 0.05 (as shown in the figure). You see that the
shape of the passband for the IIR filter generally matches the shape in
Figure 1-2 (accounting for the shift from a linear to a logarithmic y-axis).
For now, we have an eighth-order, stable filter based on the direct form II
transposed structure. It consists of one section.
10 To see the poles and zeros for the filter, select Pole/Zero Plot from the
Analysis menu in FDATool.
For this filter, which is stable, the poles lie on or very close to the unit circle,
and close to one another. Generally, when roots are close, they can be
sensitive to coefficient quantization effects. Changes to the positions of the
poles or zeros could cause the filter to become unstable. This is your first hint
that quantizing this double-precision filter might be difficult.
Quantizing the IIR Filter
You used the filter design tools in FDATool to design an IIR filter with a
passband you defined. To demonstrate the effects of quantization on this filter,
we can convert the filter to fixed-point arithmetic and quantize its transfer
1-20
Getting Started with the Toolbox
function coefficients. So, to complete the design process, we need to quantize
the IIR filter, keeping its performance intact through the quantization process.
You use FDATool in quantization mode to accomplish this operation:
1 In FDATool, click Set Quantization Parameters to switch FDATool to
quantization mode.
2 To quantize your IIR filter, select Turn quantization on under
Quantization in FDATool.
You have quantized the current filter using the defaults. Under Current Filter
Information you see the filter is still stable, eighth-order, and consists of one
section. Notice that Source now reads Designed(Quantized). If you import a
filter into FDATool, Source changes to read Imported.
For now the filter uses the default structure — Direct form II transposed.
3 Look at the Magnitude Response in the FDATool analysis area, which now
shows the response curves for both your original IIR filter (Reference) and
the quantized version (Quantized).
1-21
1
Filter Design Toolbox Overview
While the new filter is stable, quantizing the filter coefficients seriously
degraded its response. Truncating some of the coefficients, as you did when
you quantized the filter, caused the coefficients to exceed the limits [-1,1) of
the fixed data type (called overflow). Those coefficients were truncated to fall
within the range -1 to 1. Maybe we can scale the transfer function
coefficients of the reference filter so that quantizing the filter does not do
such damage.
If you select View Filter Coefficients from the Analysis menu in FDATool,
you can review the coefficients of the reference and quantized filters. When
you scroll to the bottom of the display in the analysis area, you see that eight
coefficients overflowed during quantization. In the left column of the
analysis area, the symbols +,-, and 0 appear to indicate which coefficients
overflowed or underflowed, and in which direction (toward ±infinity or
toward zero. The following table summarizes the meaning of the symbols.
1-22
Getting Started with the Toolbox
Symbol
Meaning
+
Coefficients marked with this symbol overflowed toward
positive infinity. FDATool handled the overflow as directed
by the Overflow mode property value for the Coefficient
property. In this case the setting is saturate.
-
Coefficients marked with this symbol overflowed toward
negative infinity. FDATool handled the overflow as directed
by the Overflow mode property value for the Coefficient
property. In this case the setting is saturate.
0
Coefficients marked with this symbol underflowed to zero.
FDATool handled the underflow as directed by the round
mode property value for the Coefficient property. In this
case the setting is round toward nearest.
For example, the ninth numerator coefficient underflowed toward zero, and
eight of the nine denominator coefficients overflowed toward plus or minus
infinity and were saturated to (1-eps) or -1.0. The following code shows the
filter coefficients.
Numerator
QuantizedCoefficients{1}
0 (1)
0.000000000000000
(2)
-0.000030517578125
(3)
0.000091552734375
(4)
-0.000122070312500
(5)
0.000091552734375
(6)
-0.000061035156250
(7)
0.000030517578125
0 (8)
0.000000000000000
0 (9)
0.000000000000000
Denominator
QuantizedCoefficients{2}
+ (1)
0.999969482421875
- (2)
-1.000000000000000
+ (3)
0.999969482421875
- (4)
-1.000000000000000
+ (5)
0.999969482421875
- (6)
-1.000000000000000
ReferenceCoefficients{1}
0.000006805499528066
-0.000037137669916463
0.000087218236138384
-0.000115464066452111
0.000094505093411602
-0.000048910514376539
0.000015426381218102
-0.000002610607681069
0.000000167701400337
ReferenceCoefficients{2}
1.000000000000000000
-7.532602606298016000
24.769238091848504000
-46.426612663362768000
54.235802381593018000
-40.420742464796888000
1-23
1
Filter Design Toolbox Overview
+ (7)
- (8)
(9)
0.999969482421875
-1.000000000000000
0.569671630859375
18.759623438876325000
-4.954375991471868800
0.569669813719955510
Warning: 8 overflows in coefficients.
4 Click Scale transfer-fcn coeffs<=1.
FDATool scales the reference filter coefficients, then quantizes the reference
filter again. This time, the coefficients do not overflow or underflow and the
filter response in the stop band appears to closely match the reference filter
response, as shown in the next figure.
Your quantized filter is now unstable (check FDATool for the Current Filter
Information). When the reference filter poles and zeros are so close to one
another, they can be very sensitive to the effects of quantization. In this case,
quantizing the filter moved some of the poles outside the unit circle. If you
switch to the Pole/Zero plot by selecting Pole/Zero from the Analysis menu in
FDATool, you see the poles and zeros for the quantized filter.
1-24
Getting Started with the Toolbox
We can resolve this problem by converting our filter structure to one that is
more robust to quantization effects. For example, we could change from direct
form II transposed to a lattice structure, or we could use second-order sections
(SOS) to implement our quantized filter. Second-order section form offers a
strong option because when we convert to SOSs, we reduce the order of the
polynomials that define the filter, and thus reduce the filter sensitivity to
quantization.
5 To convert the filter to second-order section form, click Convert structure
under Current Filter Information in FDATool.
On the Convert Structure dialog, you select the filter structure to convert
to, and you choose whether to use second-order sections and scaling. For this
example, under Convert To, select Direct form II transposed because we
are not changing the filter structure.
1-25
1
Filter Design Toolbox Overview
You can keep your filter structure the same and convert to SOS form. Or you
can change your filter structure and adopt SOS form. We want to keep the
transposed direct form II structure, but use second-order sections to implement
the filter.
6 Select Use second-order sections.
When you convert to second-order sections (SOS), you have the option of
treating the error between the reference filter magnitude response and the
quantized filter magnitude response in one of three ways. The Scale option
determines which method FDATool uses:
- None — ignore scaling when determining the SOS coefficients
- L-2 — use Euclidean norm when determining the SOS coefficients
- L-infininty — use L∞ scaling when determining the SOS coefficients
FDATool optimizes the order of the second-order sections according to the
scaling option you choose. (The tf2sos function that performs the
conversion is called with option 'down' for L-2 and 'up' for L-infinity
scaling.)
Our IIR filter does not need to be scaled to meet our needs, so select None
from the Scale list.
1-26
Getting Started with the Toolbox
7 Click OK to close the dialog and convert the filter according to the settings
you selected.
8 Select Magnitude Response from the FDATool Analysis menu.
Our quantized second-order section filter now has the magnitude response
we require, and matches the unquantized filter specifications. In the
following figure showing the magnitude response curves for both filters, you
cannot distinguish between the reference and quantized filter curves.
As you followed this example, you created an arbitrary magnitude IIR filter to
match an ideal filter response. Then you quantized the filter and converted it
to second-order section form. All of this you accomplished using FDATool,
although you could have used the command line to perform the same filter
design and quantization operations.
To save the filter you created in FDATool, either use the Save Session option
on the FDATool File menu to save the session and the FDATool interface
settings, or use the FDATool option Export on the File menu to export the
filter to your MATLAB workspace in transfer function form.
1-27
1
Filter Design Toolbox Overview
Selected Bibliography
For further information about the algorithms and computer models used to
design filters and apply quantization in the toolbox, refer to one or more of the
following references.
Digital Filters
Antoniou, Andreas, Digital Filters, Second Edition, McGraw-Hill, Inc, 1993
Mitra, Sanjit K., Digital Signal Processing: A Computer-Based Approach,
McGraw-Hill, Inc, 1998
Oppenheim, Alan. V., R.W. Schafer, Discrete-Time Signal Processing,
Prentice-Hall, Inc, 1989
Quantization and Signal Processing
Lapsley, Phil, J, Bier, A. Shoham, E.A. Lee, DSP Processor Fundamentals,
IEEE Press, 1997
McClellan, James H., C.S. Burrus, A.V. Oppenheim, T.W. Parks, R.W. Schafer,
H.W. Schuessler, Computer-Based Exercises for Signal Processing Using
MATLAB 5, Prentice-Hall, Inc., 1998
Roberts, Richard A., C.T. Mullis, Digital Signal Processing, Addison-Wesley
Publishing Company, 1987
Van Loan, Charles, Computational Frameworks for the Fast Fourier
Transform, SIAM,1992
1-28
2
Designing Advanced
Filters
The Optimal Filter Design Problem . . . . . . . . . 2-2
Optimal Filter Design Theory . . . . . . . . . . . . . 2-2
Optimal Filter Design Solutions . . . . . . . . . . . . 2-5
Advanced FIR Filter Designs . . . . . . . . . . . . 2-7
gremez Examples . . . . . . . . . . . . . . . . . . 2-8
Advanced IIR Filter Designs
iirlpnorm Examples . . . .
iirlpnormc Examples . . . .
iirgrpdelay Examples . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 2-37
. 2-40
. 2-45
. 2-51
Robust Filter Architectures . . . . . . . . . . . . . 2-59
Filter Design Example That Includes Quantization . . . . . 2-62
Selected Bibliography
. . . . . . . . . . . . . . . 2-68
2
Designing Advanced Filters
The Optimal Filter Design Problem
Filter Design Toolbox provides you with the tools to design optimal filters in
the finite impulse response (FIR) and infinite impulse response (IIR) domains.
Often, filter design techniques and algorithms result in filters that are easy to
apply and put relatively light demands on computational systems. While these
filters are acceptable in many instances, they are not optimal solutions to the
filtering needs of some digital signal processing implementations. Suboptimal
filter designs can meet the performance specifications for the filter, but
generally at the expense of increased filter order. This can result in increased
arithmetic computational load for each input sample and lower operating
speed than may be possible and necessary.
You use the functions firlpnorm, gremez, iirlpnorm, and iirlpnormc to
design optimal filters. The following sections review the optimal filter design
problem and introduce the filter design functions included in the toolbox:
• “Optimal Filter Design Theory” on page 2-2
• “Optimal Filter Design Solutions” on page 2-5
• “Advanced FIR Filter Designs” on page 2-7
• “Examples — Using gremez to Design FIR Filters” on page 2-8
• “Advanced IIR Filter Designs” on page 2-37
• “Examples — Using Filter Design Toolbox Functions to Design IIR Filters”
on page 2-38
Optimal Filter Design Theory
How do you design a filter that meets your performance needs, such as having
the required passbands, stopbands, or transition regions, and is also the
optimal solution? (The optimal solution filter minimizes a measure of the error
between your desired frequency response and the actual filter response.)
2-2
Consider two filter frequency response curves:
• D(ω) — the response of your ideal filter, as defined by your signal processing
needs and specifications
• H(ω) — the frequency response of the filter implementation you select
In the following figure you see the response curves for D(ω) and H(ω), both
lowpass filters.
Comparison of desired and actual frequency responses
1.4
Ideal Filter Response D(ω)
1.2
Actual Filter Response H(ω)
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.05
0.1
0.15
0.2
Normalized Frequency (×π rad/sample)
0.25
0.3
Figure 2-1: Response Curves for Ideal and Actual Lowpass Filters
Optimal filter design theory seeks to make H(ω) match D(ω) as closely as
possible by a given measure of closeness.
2-3
2
Designing Advanced Filters
More precisely, if we define a weighted error
E(ω)= W(ω)[H(ω ) – D(ω)]
where E(ω) is the error between the ideal and actual filter response values and
W(ω) is the weighting factor, the optimal filter design problem is to determine
an H(ω) that minimizes some measure or norm of E(ω) given a particular
weighting function W(ω) and a desired response D(ω).
W(ω), the weighting function, lets you determine which portions of the actual
filter response curve are most important to your filter performance, whether
passband response or attenuation in the stopband.
Usually, developers use the Lp norm to measure the error. This norm is given
by
ò [E[ω]]
p
Ω
and this is the quantity we minimize.
In practice, the two most commonly used norms are L2 and L∞, meaning that
p equals 2 and p equals infinity.
Filter designs that minimize the L∞ are attractive because they lead to
equiripple solutions. Their equiripple characteristics tend to produce the
lowest order filter that satisfies some prescribed specification.
When p goes to infinity, L∞ norm simplifies to
max|E(ω)|
ωεΩ
meaning that when p equals ∞, the optimal design minimizes the maximum
magnitude of the weighted error. Hence, it yields a minimax solution.
Notice that the Lp norm is computed over a region Ω that uses a subset of the
positive Nyquist interval [0,π]. Ω covers the positive Nyquist interval except
for certain frequency bands deemed to be “don’t care” regions or transition
bands that are not included in the optimization.
2-4
Optimal Filter Design Solutions
We have stated that the optimal filter design problem is to find the filter whose
magnitude response, |H(ω)|, minimizes
ò [ W( w )( H( w)
p
– D ( w ) ) ] dw
ω
for a given Ω, p, W(ω) and D(ω). You can use both FIR and IIR filters to meet
this requirement.
For the FIR case, with p equals ∞, and the additional constraint that the filter
must have linear phase, you can use a very efficient design method, based on
the Remez exchange algorithm to determine the optimal solution.
Function gremez in the toolbox implements this method. Additionally, gremez
provides optional calling syntaxes that enable variations and enhancements to
the filter design problem.
To design optimal FIR solutions in the general case where p is not necessarily
equal to infinity, the toolbox includes the function firlpnorm. You may find
this useful in cases where minimax solutions lead to abrupt time-domain
responses. firlpnorm does not use the Remez exchange algorithm and
generally takes longer to design a filter than gremez and other filter design
functions. Moreover, firlpnorm is not constrained to linear phase filters.
Note that Signal Processing Toolbox provides the function firls, an efficient
FIR linear phase solution to the optimal filter design problem in the
least-squares sense, that is, when p equals 2.
IIR solutions to the optimal filter design problem are more involved than their
FIR counterparts. Filter Design Toolbox offers two functions that design IIR
filters that are optimal in the least-p norm sense: iirlpnorm and iirlpnormc.
iirplnorm uses a somewhat faster, unconstrained algorithm, while
iirplnormc uses a constrained algorithm that designs an optimal filter that
meets the specifications while restricting the maximum radius of its poles to a
specified value less than one.
Elliptic filters, such as those you use the function ellip (in Signal Processing
Toolbox) to design, are optimal IIR filters for the case p equals infinity, when
the desired magnitude response is piecewise constant, and the filter numerator
and denominator orders are the same.
2-5
2
Designing Advanced Filters
The Parks-McClellan method, which implements the Remez exchange
algorithm, produces a filter design that just meets your design requirements,
but does not exceed them. In many instances, when you use the window method
to design a filter, the result is a filter that performs too well in the stopband.
This wastes performance and taxes computational power by using more filter
coefficients than necessary. When you use a rectangular window in the window
design method, the resulting filter can be shown to be the optimal, unweighted
least squares solution to the filter design problem. In summary, the optimal
solution is not always a good solution to the filter design problem.
Filters designed using the Parks-McClellan method have equal ripple in their
passbands and stopbands. For this reason, they are often called equiripple
filters. They represent the most efficient filter designs for a given specification,
meeting your frequency response specification with the lowest order filter.
2-6
Advanced FIR Filter Designs
Advanced FIR Filter Designs
Filter Design Toolbox includes a function, gremez, for designing FIR filters that
represent the optimal solutions to filter design requirements. gremez provides
a minimax filter design algorithm that you use to design the following real FIR
filters:
• Types 1 through 4 linear phase
• Minimum phase
• Maximum phase
• Minimum order, even or odd
• Extra-ripple
• Maximal-ripple
• Constrained-ripple
• Single-point band
• Forced gain
• Arbitrarily shaped frequency response
For examples of filters that use gremez design features, refer to “gremez
Examples” on page 2-8.
FIR filters, when implemented nonrecursively, do not use feedback in their
architectures. This limits the filter design so that you include current and past
inputs to the filter, as opposed to including past outputs (feedback) to calculate
the current output of the filter. In this toolbox, you use the function gremez to
design FIR filters. Among other features, gremez lets you:
• Define filters that have arbitrary shape frequency response curves
• Set a range of performance limits for a filter
• Set the weighting for the error between the desired response and the actual
response in each band of interest in a filter
remez and gremez respond the same way to the same input and output
arguments, where the input arguments are valid for both functions. gremez
extends the remez algorithm to support the new filter designs by adding new
input argument options.
2-7
2
Designing Advanced Filters
Note To provide improved FIR filter design optimization, gremez uses a
generalized Remez algorithm that is not identical to the Remez algorithm
used by remez. Specifically, gremez uses a higher density frequency grid in
filter transition regions, such as at the cutoff points. Thus the frequency grid
is not constant, but changes density across the frequency spectrum, letting the
algorithm more closely optimize filter performance in those areas.
For more straightforward filter designs, remez and gremez generate the same
filter coefficients and the same design. As the filter gets more complex, such as
higher order or more bands or steeper transition regions, the filter designs
may diverge. Generally, gremez provides better optimized filter designs in
these cases.
Using gremez to design filters places the following restrictions on your designs:
• Design must be FIR.
• You can select the number of filter coefficients.
• The frequency response curve must be divided into a series of passbands and
stopbands separated by transition or “don’t care” bands.
• Within each passband and stopband, you specify your desired amplitude
response as a piecewise constant function.
• You cannot constrain the amplitude response in transition bands.
With these considerations in place, gremez designs equiripple, or minimax,
filters to meet your specifications.
gremez Examples
Each of these examples uses one or more features provided in the function
gremez. Review each example to get an overview of the capabilities of the
function.
Examples — Using gremez to Design FIR Filters
gremez provides a wide range of new capabilities for FIR filter design. Because
of the comprehensive nature of the generalized Remez algorithm, the best way
to learn what you can do with the new function is by example. This section
presents a series of examples that investigate the filters you can design
2-8
Advanced FIR Filter Designs
through gremez. You can view these examples as a demonstration program in
MATLAB by opening the MATLAB demos and selecting Filter Design from
Toolboxes. Listed there you see a number of demonstration programs. Select
Minimax FIR Filter Design to see function gremez used to create many filters,
from a lowpass filter to a constrained stopband design to a minimum phase,
lowpass filter with a constrained stopband.
To open the FIR filter design demo.
Follow these steps to open the FIR filter design demo in MATLAB.
1 Start MATLAB.
2 At the MATLAB prompt, enter demos.
The MATLAB Demo Window dialog opens.
3 On the left-hand list, double-click Toolboxes to expand the directory tree.
You see a list of the toolbox demonstration programs available in MATLAB.
4 Select Filter Design.
5 From the right-hand list, select Minimax FIR Filter Design.
A few examples include comparisons to other filter designs and some include
analysis notes. For details about using function gremez, refer to Chapter 11,
“Function Reference.” While this set of examples covers some of the options for
gremez, many options exist that do not appear in these examples. Examples
cover common or interesting gremez options to demonstrate some of the
capabilities.
2-9
2
Designing Advanced Filters
In each of the examples in this section, we use the output argument res to
return the structure res that contains information about the filter.
Structure res
Element
Contents
res.order
Filter order.
res.fgrid
Vector containing the frequency grid used in the filter
design optimization.
res.H
Actual frequency response on the grid in fgrid.
res.error
Error at each point on the frequency grid (desired
response- actual response).
res.des
Desired response at each point on fgrid.
res.wt
Weights at each point on fgrid.
res.iextr
Vector of indices into fgrid of extremal frequencies.
res.fextr
Vector of extremal frequencies.
res.iterations
Number of Remez iterations for the optimization.
res.evals
Number of function evaluations for the optimization.
res.edgeCheck
Results of the transition-region anomaly check.
Computed when the 'check' option is specified. One
element returned per band edge. Returned values can
be:
• 1 = OK
• 0 = Probable transition-region anomaly
• -1 = Edge not checked. In the normalized frequency
domain, the edges at f=0 and f=1 cannot have
anomalies and are not checked.
2-10
Advanced FIR Filter Designs
Example — Designing a Minimax Filter
To use gremez to design an equiripple or minimax filter, we use the following
statement.
[b,err,res] = gremez(22,[0 0.4 0.5 1],[1 1 0 0],[1,5]);
If you use the same statement, replacing gremez with remez, you get the same
filter. You can reproduce any filter that remez generates by replacing remez
with gremez in the statement. gremez retains full compatibility with remez.
Here’s a plot of the magnitude response of the minimax filter as created by
gremez. The following code creates this figure.
[h,w]=freqz(b); plot(w,abs(h))
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Our filter ends up as a 22nd-order filter with magnitude response that has
ripples about 1 in the passband and ripples about 0 in the stopband. Using the
weight vector, we chose to emphasize meeting the stopband performance five
times as much as meeting the passband performance. Hence the reduced ripple
2-11
2
Designing Advanced Filters
in the stopband relative to the passband. In the next figure, we switch the
weighting to emphasize the passband, and see that the passband ripple is
much smaller than the stopband ripple.
[b,err,res] = gremez(22,...,[5,1]);
plot(res.fgrid,abs(res.H))
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Example — Designing a Minimax Filter, Odd-Order, Antisymmetric
In this example, gremez designs a filter that remez cannot. When you evaluate
the following code in MATLAB, the result is a minimax FIR filter, this time
having odd-order and antisymmetric structure, known as type 4. You can see
from the figure that the magnitude response now represents a high pass filter.
In this example, we specify the filter as type 4 ('4' in the statement) to get the
odd-order, antisymmetric design we want.
[b,err,res]=gremez(21,[0 0.4 0.5 1], [0 0 1 1],[2 1],'4');
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
2-12
Advanced FIR Filter Designs
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
We have weighted the stopband more heavily than the passband ([2 1]) in the
function syntax. The 2 and 1 tell gremez that we care about meeting the
stopband specification twice as much as the passband specification. Notice that
the weighting is relative, not absolute. Our weights say that the stopband is
twice as important as the passband. They do not specify the weighting in
absolute terms.
Example — Designing a “Least Squares-Like” Filter
gremez lets you design filters that resemble least squares design. In this
example, we design a 53rd-order filter and use the user-supplied file
taperedresp.m to specify a frequency response weighting function to perform
the error weighting for the design. So you can reproduce this example, the file
taperedresp.m is in the matlabroot\toolbox\filterdesign\filtdesdemos
folder. taperedresp.m contains the following code to specify the weighting.
2-13
2
Designing Advanced Filters
% Example for a user-supplied frequency-response function
% taperedresp.m
function [des,wt] = taperedresp(order, ff, grid, wtx, aa)
nbands = length(ff)/2;
% Create output vectors of the appropriate size
des=grid;
wt=grid;
for i=1:nbands
k = find(grid >= ff(2*i-1) & grid <= ff(2*i));
npoints = length(k); t = 0:npoints-1;
des(k) = linspace(aa(2*i-1), aa(2*i), npoints);
if i == 1
wt(k) = wtx(i) * (1.5 + cos((t)*pi/(npoints-1)));
elseif i == nbands
wt(k) = wtx(i) * (1.5 + cos(pi+(t)*pi/(npoints-1)));
else
wt(k) = wtx(i) * (1.5 - cos((t)*2*pi/(npoints-1)));
end
end
To generate the least-squares-like filter, use the following code.
[b,err,res]=gremez(53, [0 0.3 0.33 0.77 0.8 1],...
{'taperedresp',[0 0 1 1 0 0]}, [2 2 1]);
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
When you issue these statements at the MATLAB prompt, you get the
following plot for the filter magnitude response.
2-14
Advanced FIR Filter Designs
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Example — Designing a Constrained Lowpass Filter
With gremez, you can both apply weighting to the passband and apply a limit
or constraint to the error in the stopband, called constraining. Limiting the
stopband error can be useful in circumstances where your filter must meet a
specified stopband requirement. To create a lowpass filter with a constrained
stopband and weighted passband response, we use gremez with the 'w'
optional input argument to weight the passband. The optional input argument
'c' constrains the filter stopband error not to exceed 0.2. Note that to use the
constraining and weighting options, your filter must have at least one
unconstrained band. That is, cell array c must contain at least one 'w' entry.
In our example, c is {'w' 'c'}.
[b,err,res]=gremez(12,[0 0.4 0.5 1], [1 1 0 0], [1 0.2],...
{'w' 'c'});
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
2-15
2
Designing Advanced Filters
The next figure shows the lowpass filter with the constraints applied.
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
When you use constraining values in your gremez filter design, check to see
that your filter actually touches the constraining value in the stopband. If it
does not, increase the error weighting ('w') for your unconstrained bands. This
change causes the constrained errors to approach the constraint value more
quickly. Notice that the plot shows our filter just touches the desired constraint
of 0.2.
Example — Designing a Constrained Bandstop Filter
Continuing with the concept of using weighting in gremez, we design a
bandstop filter whose passband ripple we constrain not to exceed 0.05 and 0.1.
In this instance, cell array c is {'c' 'w' 'c'} to constrain the passbands and
we use the optional input vector W=[0.05 1 0.1] to constrain the passband
ripple not to exceed 0.05 and 0.1.
2-16
Advanced FIR Filter Designs
[b,err,res]=gremez(22,[0 0.4 0.5 0.7 0.8 1], [1 1 0 0 1 1],...
[0.05 1 0.1], {'c' 'w' 'c'});
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
As expected the magnitude response shows different peak ripple values in the
passbands — 0.05 for the low frequency band and 0.1 for the high frequency
band.
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
2-17
2
Designing Advanced Filters
Example — Designing a Single-Point Band Filter
The following statements
[b,err,res]=gremez(42,[0 0.2 0.25 0.3 0.5 0.55 0.6 1],...
[1 1 0 1 1 0 1 1], {'n' 'n' 's' 'n' 'n' 's' 'n' 'n'});
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
generate an interesting filter that you cannot design when you use functions in
Signal Processing Toolbox: a multiple stopband filter where the stop bands are
defined by single points. In the gremez command in this example, the syntax is
b=gremez(N,F,A,S). The input vectors F, A, and S, each containing eight
values, define the response curve for the filter.
Input Vector
Use
F=[0 0.2 0.25 0.3 0.5 0.55 0.6 1]
Defines the points of interest in the frequency
response. In this case, you are working with
frequencies normalized between 0 and 1.
A=[1 1 0 1 1 0 1 1]
Set the gain at each frequency point.
S={'n' 'n' 's' 'n' 'n' 's' 'n' 'n'}
Specifies whether the frequency points represent
normal or single-point bands. By comparing the
frequency and type vector entries, we see that
F=0.25 and F=0.55 are single point bands (marked
by s), and the gain at those points is 0. The other
bands are normal bands (marked with n) with
gain =1.
From the next figure, you see that the filter has just the response we defined,
with zeros at F = 0.25 and F = 0.55.
2-18
Advanced FIR Filter Designs
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Example — Designing a Filter with a Specified In-band Value
In some filter design tasks, you want a filter whose inband value you determine
exactly. For example, you might want a 60Hz noise rejection filter to have zero
gain at F = 0.06 (F = 60Hz in real frequency). For this example, the sampling
frequency is 2KHz, so 60Hz is F = 0.06 when we normalize the frequency. We
use the following code example to design such a filter.
[b,err,res]=gremez(82,[0 0.055 0.06 0.1 0.15 1], [0 0 0 0 1 1],...
{'n' 'i' 'f' 'n' 'n' 'n'});
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
At F = 0.06, we require the gain of the filter response to be exactly 0.0. So we
force the gain at F = 0.06 to zero by adding the 'f' input option to the S vector.
As shown in the plot, the filter response is zero at F = 0.06, and the resulting
filter rejects 60Hz noise quite effectively.
2-19
2
Designing Advanced Filters
0.3
0.25
Magnitude
0.2
0.15
0.1
0.0 at F=0.06
0.05
0
0
0.02
0.04
0.06
0.08
0.1
0.12
Normalized Frequency (×π rad/sample)
0.14
0.16
0.18
You might have noticed in the gremez statement that the S vector includes an
'i' option. Entries in the S vector have any of the following values.
Vector Symbol
Meaning
n
Represents a normal frequency point
s
Represents a single-point band frequency
f
Forces the gain at this frequency to a fixed value, as
specified in the weighting vector W
i
Represents an indeterminate frequency point. Usually
used when the band should abut the next band
For our noise rejecting filter, the sampling frequency is 2 KHz, so 60 Hz is
f=0.06 in normalized frequency.
2-20
Advanced FIR Filter Designs
Example — Designing Extra-Ripple and Maximal-Ripple Filters
Extra-ripple and maximal-ripple filters have some interesting properties:
• They have locally minimum transition region widths
• They tend to converge very quickly
gremez lets you use multiple independent approximation errors to directly
design extra- and maximal ripple filters. In this example, we use independent
errors to design two filters, then we revisit our 60Hz noise rejection filter to
compare these two different approaches to designing the same filter.
Example of an Extra-Ripple Lowpass Filter
The code to design our extra-ripple filter is
[b,err,res]=gremez(12,[0 0.4 0.5 1], [1 1 0 0], [1 1],...
{'e1' 'e2'});
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
The last entries in the command, [1 1] and {'e1' 'e2'}, are the vectors W and
E that determine the weights and independent approximation errors for filters
with special properties. 'e1' is applied to the passband and 'e2' applied to the
stopband. Where the gremez algorithm usually results in equiripple filters,
using the approximations lets gremez adjust the ripple in each band
separately, as we have done in this design.
2-21
2
Designing Advanced Filters
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Example of an Extra-Ripple Bandstop Filter With Two Independent
Approximation Errors
Now we extend the extra-ripple concept by using two independent error
approximations. The two passbands share the first approximation error 'e1'.
The stopband uses 'e2'. So you can see the effectiveness of this design
approach, also create and plot a single approximation error filter for
comparison.
[b,err,res]=gremez(28,[0 0.4 0.5 0.7 0.8 1], [1 1 0 0 1 1],...
[1 1 2], {'e1' 'e2' 'e1'}); % Extra-rippple filter design
[b2,err2,res2]=gremez(28,[0 0.4 0.5 0.7 0.8 1],...
[1 1 0 0 1 1],[1 1 2]); % Weighted-Chebyshev design
[H,W]=freqz(b,1,1024);[H2,W]=freqz(b2,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot([H H2],W,S);
2-22
Advanced FIR Filter Designs
In the figure, the responses are similar for the two designs, but the extra-ripple
design shows less ripple in the passbands and slightly more in the stopband. If
you evaluate the example code in MATLAB to create the plot, you can select
Zoom in from the Tools menu in the figure window to examine the curves more
closely.
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
For this design, we let gremez use the same error approximation for the
passbands and a different one in the stopband. The result is a filter that has
minimum total error in the passbands, and minimum error in the stopband.
Example — Comparing Two 60Hz Noise Rejection Filters
With the extra-ripple filter design technique available in gremez, we can use
two different design techniques to redo our 60Hz noise rejection filter. We use
three independent error approximations in this design, one for each band, as
shown in the following code.
2-23
2
Designing Advanced Filters
[b,err,res]=gremez(82,[0 0.055 0.06 0.1 0.15 1],[0 0 0 0 1 1],...
{'n' 'i' 'f' 'n' 'n' 'n'},[10 1 1],{'e1' 'e2' 'e3'}); % New filter
[b2,err,res]=gremez(82,[0 0.055 0.06 0.1 0.15 1],...
[0 0 0 0 1 1], {'n' 'i' 'f' 'n' 'n' 'n'}); % Original filter
[H,W]=freqz(b,1,1024);
[H2,W]=freqz(b2,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot([H H2],W,S);
We have included the second gremez statement in this example to reproduce
the earlier noise rejection filter for comparison. We plot them on the same
figure for easy reference. In the stopband, the original design has lower ripple;
the new, independent error design has less ripple in the passband. Also, the
new filter has slightly steeper transition region performance.
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Using independent approximation errors, as we did in this filter when we
specified 'e1', 'e2', and 'e3', can result in better filter performance. The strings
'e1', 'e2', and so on direct gremez to consider the associated band alone, or with
2-24
Advanced FIR Filter Designs
other bands that use the same error approximation. By assigning independent
errors to each band, we let the generalized Remez algorithm used by gremez
minimize the error in each band without considering the error in the other
bands. If we do not use independent errors, the algorithm minimizes the total
error in all bands at once.
At times, you need to use independent approximation errors to get designs that
use forced inband values to converge. Error approximations are needed where
the polynomial used to approximate the filter becomes undetermined when you
try to force the inband values to converge.
Example — Checking for Transition-Region Anomalies
To allow you to check your filter designs for anomalies, gremez provides an
input option called 'check'. With the check option included in the command,
gremez reports anomalies in the response curve for the filter. An anomaly in
gremez is defined as out-of-the-ordinary response behavior in a transition, or
“don’t care,” region of the filter response.
To demonstrate anomaly checking, we use gremez to design a filter with an
anomaly, and include the 'check' optional input argument.
[b,err,res]=gremez(44,[0 0.3 0.4 0.6 0.8 1],...
[1 1 0 0 1 1],'check');
[H,W]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
With the 'check' option, gremez returns the results vector res.edgeCheck in
the structure res. Each zero-valued entry in this vector represents the location
of a probable anomaly in the filter response. Entries that are not checked, such
as the edges at f=1 and f=0, have -1 entries in res.edgeCheck.
To check for anomalies, the following command returns the vector containing
the edge check results.
2-25
2
Designing Advanced Filters
res.edgeCheck
ans =
-1
1
1
0
0
-1
There are anomalies between the f=0.6 and f=0.8 edges, as shown clearly in the
figure. This represents a transition region for our filter. Notice that the edges
at f=0 and f=1 were not checked.
1.8
1.6
1.4
Magnitude
1.2
1
0.8
0.6
0.4
0.2
0
2-26
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Advanced FIR Filter Designs
In our example, the anomalous behavior happened because of the width of the
transition region. When we define a narrower transition band, the anomaly
disappears. Generally, reducing the transition region width eliminates
anomalies in the filter response.
Example — Using Automatic Minimum Filter Order Determination
Rather than entering the filter order N in the gremez command, you can let the
generalized Remez algorithm determine the minimum order for your filter.
You set the specifications for the filter and the generalized Remez algorithm
repeatedly designs the filter until the design just meets your specifications.
You have three options for setting the minimum order option for the filter:
• 'minorder' directs the Remez algorithm to iterate over the filter design
until it finds a design that just fulfills your design specifications and is the
lowest possible order. Using this option directs gremez to use remezord to get
an initial estimate of the filter order.
• 'mineven' directs the Remez algorithm to iterate over the filter design until
it finds a design that just fulfills your design specifications and is the lowest
possible even order.
• 'minodd' directs the Remez algorithm to iterate over the filter design until it
finds a design that just fulfills your design specifications and is the lowest
possible odd order.
Note When you use the minimum order option 'minorder', gremez treats the
weights in the W vector as maximum error values for the associated
frequencies in the frequency vector F. Also, constraints become absolute limits;
gremez designs a filter that does not exceed the constraints.
For this example, we let the Remez algorithm find a minimum order filter that
implements a lowpass filter with a transition band between f=0.4 and f=0.5.
[b,err,res]=gremez('minorder',[0 0.4 0.5 1], [1 1 0 0],...
[0.1 0.02]);
[H,W,S]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
2-27
2
Designing Advanced Filters
Our filter, shown in the figure, demonstrates the desired ripple in the
passbands and stopbands, 0.1 and 0.02; the transition region meets our
specifications; and the filter order (found from res.order) is 22.
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
When you use the minimum order feature, you can specify the initial order
(your best guess) in the gremez statement. When you estimate the order,
gremez does not use remezord to make an estimate of the filter order. This is
important when remezord does not support your desired filter type, such as
differentiators and Hilbert transformers, as well as for filters that use
frequency response functions that you supply. For the following filter example,
we provide an initial estimate of 18 for the filter order, and we specify that we
want our filter to have the minimum even order possible by adding the
'mineven' option.
[b,err,res]=gremez({'mineven',18},[0 0.4 0.5 1], [1 1 0 0],...
[0.1 0.02]);
[H,W,S]=freqz(b,1,1024);
2-28
Advanced FIR Filter Designs
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
Though we provided an initial estimate of 18 for the order, the final order for
our filter is again 22. If we had specified 'minodd', the result would be a
23rd-order filter.
Example — Designing an Interpolation Filter
Now let us design an interpolation filter. These are usually used to upsample
a band-limited signal by an integer factor, for example after the signal has been
decimated by downsampling. Upsampling is often used while designing
multirate filters to reduce the computational load required to use a filter. In
Signal Processing Toolbox, you can use the function intfilt to design an
interpolation filter. While intfilt provides a way to design the filter, it does
not provide the control that gremez offers. Input options for gremez let you
define the filter response and errors in each passband and stopband, and the
weighting of the band responses in the filter design.
[b,err,res]=gremez(30,[0 0.1 0.4 0.6 0.9 1], [4 4 0 0 0 0],...
[1 100 100]);
[H,W]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'db';
freqzplot(H,W,S);
We specify a 30th-order filter with edges at 0.1, 0.4, 0.6, and 0.9, and weight
them as [1 100 100]. The resulting design has stopbands between f=0.4 and
f=0.6, and f=0.9 and f=1.0.
The next figure shows a filter designed by gremez.
2-29
2
Designing Advanced Filters
20
0
−20
Magnitude (dB)
−40
−60
−80
−100
−120
−140
−160
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Example — Comparing Filters Designed by gremez and intfilt
Now, to see that gremez lets you develop a better interpolation filter than
intfilt, we compare filters designed by both functions. We need three sets of
code to display the filters for our comparison — the first set generates the detail
plot of the first stopband, the second set displays the second stopband in detail,
and the third plot focuses on the stopband ripple. To keep the frequency
response displays consistent, we use the MATLAB plot function to ensure that
the axes and labels are the same for both filters. freqzplot does not provide
enough control of the plotting functions.
2-30
Advanced FIR Filter Designs
Code to display the first stopband.
[b,err]=gremez(30,[0 0.1 0.4 0.6 0.9 1], [4 4 0 0 0 0],...
[1 100 100]);
b2=intfilt(4, 4, 0.4);
w=linspace(0.4, 0.6)*pi; h=freqz(b,1,w); h2=freqz(b2,1,w);
plot(w/pi,20*log10(abs([h' h2']))); ylabel('Stopband #1 (dB)');
v=axis; v=[0.4 0.6 -100 v(4)]; axis(v);
Code set to display the second stopband.
[b,err]=gremez(30,[0 0.1 0.4 0.6 0.9 1], [4 4 0 0 0 0],...
[1 100 100]);
b2=intfilt(4, 4, 0.4);
w=linspace(0.9, 1)*pi; h=freqz(b,1,w); h2=freqz(b2,1,w);
plot(w/pi,20*log10(abs([h' h2']))); ylabel('Stopband #2 (dB)');
v=axis; v=[0.9 1 -100 v(4)]; axis(v);
Code set to display the passband ripple.
[b,err]=gremez(30,[0 0.1 0.4 0.6 0.9 1], [4 4 0 0 0 0],...
[1 100 100]);
b2=intfilt(4, 4, 0.4);
w=linspace(0, .1)*pi; h=freqz(b,1,w); h2=freqz(b2,1,w);
plot(w/pi,20*log10(abs([h' h2']))); ylabel('Passband (dB)');
In the next figure, showing the first stopband in detail, you see that using the
weighting function in gremez improved the minimum stopband attenuation by
almost 20dB over the intfilt design.
2-31
2
Designing Advanced Filters
−70
−75
intfilt
Stopband #1 (dB)
−80
igremez
−85
−90
−95
−100
0.4
0.42
0.44
0.46
0.48
0.5
0.52
0.54
0.56
0.58
0.6
If we switch to a plot of the second stopband, shown in the next figure, you see
that the equiripple attenuation throughout the band is about 6dB larger for the
gremez-generated filter than the minimum stopband attenuation of the filter
designed by intfilt.
2-32
Advanced FIR Filter Designs
−80
−82
−84
Stopband #2 (dB)
−86
gremez
−88
intfilt
−90
−92
−94
−96
−98
−100
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
Finally, let’s look at the passbands of the two filters, shown in the next figure.
Here, the ripple in the gremez-designed filter is slightly larger than the
passband ripple for the intfilt design. Still, both are very small, less than
0.014 dB peak-to-peak.
2-33
2
Designing Advanced Filters
12.048
12.046
gremez
12.044
Passband (dB)
intfilt
12.042
12.04
12.038
12.036
12.034
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Example — Designing a Minimum Phase Lowpass Filter with a
Constrained Stopband
With gremez you can determine whether the FIR filter you design is minimum
phase, maximum phase, or linear phase. Through this example we show a
minimum phase filter and look at the roots of the filter transfer function to see
that no roots lie outside the unit circle in the z-plane. First, we create the
minimum phase filter by using gremez with the 'minphase' optional input
argument.
[b,err,res]=gremez(12,[0 0.4 0.5 1], [1 1 0 0],[1 0.1],...
{'w' 'c'},{64},'minphase');
gremez generates a lowpass filter with constrained stopband magnitude equal
to 0.1, and the filter is minimum phase as well. We could have specified a
maximum phase design by replacing the 'minphase' option with 'maxphase'. In
the gremez statement, you might have noticed the cell array {64} entry. The
cell array entries define the grid density for points across the frequency
spectrum.
2-34
Advanced FIR Filter Designs
Now, plot the filter to view the frequency response.
[H,W]=freqz(b,1,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
We have a lowpass filter with stopband ripple not exceeding 0.1, as desired.
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
In the next figure, viewing our filter roots on the z-plane plot shows us that the
roots lie in or on the unit circle. The zeros of a minimum phase delay FIR filter
lie on or inside the unit circle. Maximum phase delay filters have zeros that lie
on or outside the unit circle.
[b,err,res]=gremez(12,[0 0.4 0.5 1], [1 1 0 0],[1 0.1],...
{'w' 'c'},{64},'minphase');
[H,W]=freqz(b,1,1024);
zplane(roots(b));
2-35
2
Designing Advanced Filters
1
0.8
0.6
Imaginary Part
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1
−0.5
0
Real Part
0.5
1
Notice that the filter, with eight zeros on the unit circle, could be very sensitive
to quantization. You could use FDATool to investigate the effects of quantizing
this filter.
2-36
Advanced IIR Filter Designs
Advanced IIR Filter Designs
Many digital filters use both input values and previous output values from the
filter to calculate the current output value. FIR filters can be implemented with
feedback, although this is unusual. Cascaded integrated comb filters are one
example.
For IIR filters, the transfer function is a ratio of polynomials:
• The numerator of the transfer function. When this expression falls to zero,
the value of the transfer function is zero as well. Called a zero of the function.
• The denominator of the transfer function. When this expression goes to zero
(division by zero), the value of the transfer function tends to infinity; called
a pole of the function or filter.
Filter Design Toolbox introduces three functions: iirlpnorm, iirlpnormc, and
iirgrpdelay for designing IIR filters that design optimal solutions to your
filter requirements. With these new filter functions, you can design filters to
meet your specifications that you could not design using the IIR filter design
functions in Signal Processing Toolbox.
Function iirlpnorm uses a least-pth norm unconstrained optimization
algorithm to design IIR filters that have arbitrary shape magnitude response
curves. iirlpnormc uses a least-pth norm optimization algorithm as well, only
this version is constrained to let you restrict the radius of the poles of the IIR
filter.
To let you design allpass IIR filters that meet a prescribed group delay
specification, iirgrpdelay uses a least-pth constrained optimization
algorithm. For basic information about the least-pth algorithms used in the IIR
filter design functions, refer to Digital Filters by Antoniou [1].
This section uses examples to introduce the IIR filter design functions in the
toolbox. As you review these examples, you may notice that the IIR design
functions use the same syntax, input, and output arguments. Because the
design functions use very similar algorithms, common input and output
arguments apply. Arguments are used in the same way, and carry the same
defaults and restrictions. That said, if an example of one IIR function uses a
syntax that does not appear under another IIR design function, chances are
you can use the first syntax in the other design function as well.
2-37
2
Designing Advanced Filters
Examples — Using Filter Design Toolbox Functions to Design IIR Filters
Filter Design Toolbox provides new capabilities for IIR filter design. Because
of the comprehensive nature of the new IIR design functions, the best way to
learn what you can do with them is by example. This section presents a series
of examples that investigate the filters you can implement through IIR filter
design in Filter Design Toolbox. You can view these examples as a
demonstration program in MATLAB by opening the MATLAB demos and
selecting Filter Design from Toolboxes. Listed there you see a number of
demonstration programs. Select one of the following demos to see the IIR filter
design functions being used to design a variety of filters:
• Least P-norm Optimal IIR Filter Design demonstrates IIR filter design
function iirlpnorm. Examples include:
- “Example — Using iirlpnorm to Design a Lowpass Filter” on page 2-40
- “Example — Using iirlpnorm to Design a Low Order Filter” on page 2-41
- “Example — Using iirlpnorm to Design a Bandstop Filter” on page 2-42
- “Example — Using iirlpnorm to Design a Noise-Shaping Filter” on page
2-44
• Constrained Least P-norm IIR Filter Design demonstrates IIR filter
design function iirlpnormc. This set of examples includes:
- “Example — Using iirlpnormc to Design a Lowpass Filter” on page 2-45
- “Example — Using iirlpnormc to Design a Bandstop Filter with a
Constrained Pole Radius” on page 2-47
- “Example — Using iirlpnormc to Design a High-Order Notch Filter” on
page 2-48
- “Example — Using iirlpnormc to Change an Elliptic Filter to a
Constrained Lowpass Filter” on page 2-49
• IIR Filter Design Given a Prescribed Group Delay demonstrates IIR
filter design function iirgrpdelay. These examples include:
- “Example — Using iirgrpdelay to Design a Filter with a User-Specified
Group Delay Contour” on page 2-52
- “Example — Using iirgrpdelay to Design a Lowpass Elliptic Filter with
Equalized Group Delay” on page 2-54
2-38
Advanced IIR Filter Designs
To Open the IIR Filter Design Demos
Follow these steps to open the IIR filter design demos:
1 Start MATLAB.
2 At the MATLAB prompt, enter demos.
The MATLAB Demo Window dialog opens.
3 On the list on the left, double-click Toolboxes to expand the directory tree.
You see a list of the toolbox demonstration programs available in MATLAB.
4 Select Filter Design.
5 From the list on the right, select one of the following demonstration
programs:
- Least P-norm Optimal IIR Filter Design
- Constrained Least P-norm IIR Filter Design
- IIR Filter Design Given a Prescribed Group Delay
A few examples include comparisons to other filter design functions and
analysis notes. For details about using the IIR design functions iirlpnorm,
iirlpnormc, and iirgrpdelay, refer to Chapter 11, “Function Reference.”
While this set of examples covers many of the options for the functions, more
options exist that do not appear in these examples. Examples cover common or
interesting IIR design options to highlight some of the capabilities of the design
functions.
In these examples, you can see that iirlpnorm, iirlpnormc, and iirgrpdelay
use many of the input arguments used by gremez, plus others such as the
denominator order. At the most basic level, each IIR filter design function uses
the input arguments N, D, F, Edges, and A — the filter order for the numerator
and denominator (so you can specify different order numerators and
denominators), the vector containing the filter cutoff frequencies, the band
edge frequencies, and the filter response at each frequency point. F and A must
have matching numbers of elements; they can exceed the number of elements
in Edges. You use this feature to specify a gain contour within a band defined
by the entries in Edges. Every frequency that appears in Edges must also be an
element of F. Also, the first band edge must equal the first frequency and the
last band edge must equal the last frequency in F.
2-39
2
Designing Advanced Filters
iirlpnorm Examples
Each of these examples uses one or more feature provided in the function
iirlpnorm. The examples build on one another, although they can be run
separately. Review each example to get an overview of the capabilities of the
function.
Example — Using iirlpnorm to Design a Lowpass Filter
To design a lowpass filter with maximum gain of 1.6 in the passband, we use
the syntax iirlpnorm(N,D,F,Edges,A,W). To duplicate the filter in the figure,
use this code.
[b,a]=iirlpnorm(3, 11, [0 0.15 0.4 0.5 1], [0 0.4 0.5 1],...
[1 1.6 1 0 0], [1 1 1 100 100]);
[H,W,S]=freqz(b,a,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
hold on; plot([0 0.15 0.4 0.5 1], [1 1.6 1 0 0], 'r'); hold off;
When you look at the magnitude response curve, notice the response reaches
1.6 in the passband.
2-40
Advanced IIR Filter Designs
s
1.6
1.4
1.2
Magnitude
1
Filter response calculated by
iirlpnorm
0.8
0.6
Ideal filter response
defined by F and A
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Example — Using iirlpnorm to Design a Low Order Filter
The curves in the next figure show the results of using iirplnorm to design a
low-order filter with a single band. For this design, we introduce a new
two-element vector P=[Pmin,Pmax] that defines the minimum and maximum
values of P in the least-pth norm algorithm. If you do not specify P, the default
values are [2 128], resulting in the L∞ or Chebyshev norm. Specify Pmin and
Pmax to be even numbers. To view the placement of the poles and zeros for your
filter before the optimization takes place, replace [Pmin Pmax] with the string
'inspect'. With the option 'inspect' in use, the algorithm does not optimize
the filter design.
2-41
2
Designing Advanced Filters
1.4
1.2
1
Magnitude
Desired filter response
0.8
Low-order filter
response from
iirlpnorm
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
We specified a lowpass filter with third-order numerator and denominator, and
used the P vector to limit the optimization range, by using the function syntax
iirlpnorm(N,D,F,Edges,A,W,P).
[b,a]=iirlpnorm(3, 3, [0 .2 .6 .8 1], [0 1], [0 .4 .2 0 1],...
[1 1 1 1 1], [2 64]);
[H,W,S]=freqz(b,a,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
hold on; plot([0 .2 .6 .8 1], [0 .4 .2 0 1], 'r'); hold off;
Setting W=[1 1 1 1 1] is the same as not setting weight values.
Example — Using iirlpnorm to Design a Bandstop Filter
Designing IIR bandstop filters is straightforward. Enter the frequency,
magnitude, edges, and weight vectors using the syntax
iirlpnorm(N,D,F,Edges,A,W) as shown here. To ensure that the stopband
2-42
Advanced IIR Filter Designs
rejects undesired frequencies aggressively, we weight the magnitude response
in the stopband more heavily by entering the weight vector [1 1 5 5 1 1],
telling the optimization algorithm that meeting the inband response
specification is five times as important as meeting the out-of-band response.
[b,a]=iirlpnorm(10, 7, [0 .25 .35 .7 .8 1],...
[0 .25 .35 .7 .8 1], [1 1 0 0 1 1], [1 1 5 5 1 1]);
[H,W,S]=freqz(b,a,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
hold on; plot([0 .25 .35 .7 .8 1], [1 1 0 0 1 1], 'r'); hold off;
As you can see from the following figure, the filter meets our design needs quite
closely.
1.4
1.2
Magnitude
1
Designed Filter
0.8
0.6
Ideal Response
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
2-43
2
Designing Advanced Filters
Example — Using iirlpnorm to Design a Noise-Shaping Filter
In this example, we create a lowpass filter with a rising magnitude in the
passband. Communications designers use the filter when they simulate the
effects of motion between a transmitter and receiver, such as you find in
cellular telephone networks. Here, we use iirlpnorm to design the same filter.
Because of the complex shape of the passband, we define the vectors F, A, W, and
Edges in the workspace, then use the vector names in the iirplnorm
statement.
F = 0:0.01:0.4;
A = 1.0 ./ (1 - (F./0.42).^2).^0.25;
F = [F 0.45 1];
A = [A 0 0];
edges = [0 0.4 0.45 1];
W = ones(1, length(A));
[b,a]=iirlpnorm(4, 6, F, edges, A, W);
[H,W,S]=freqz(b,a,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
hold on; plot(F,A, 'r'); hold off;
When you compare the figure below to the filter design in “Getting Started with
the Toolbox” on page 1-15, you see they match very well.
2-44
Advanced IIR Filter Designs
2
1.8
1.6
1.4
Magnitude
1.2
Actual noise-shaping
filter design
1
0.8
Desired filter response
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
iirlpnormc Examples
Each of these examples uses one or more feature provided in the function
iirlpnormc. Review each example to get an overview of the capabilities of the
function.
Example — Using iirlpnormc to Design a Lowpass Filter
Just as you use iirlpnorm to design lowpass filters, you can use iirlpnormc to
design them as well. iirlpnormc lets you limit the radius of the filter poles
when you specify the filter in the function. By restricting the poles to be less
than a certain distance from the origin of the unit circle in the z-plane, the filter
remains stable, while possibly improving the robustness of the filter to
quantization effects. In this lowpass filter example, we restrict the pole radius
not to exceed 0.95, using the function syntax
iirlpnormc(N,D,F,Edges,A,W,Radius).
2-45
2
Designing Advanced Filters
[b,a]=iirlpnormc(3, 11, [0 0.15 0.4 0.5 1], [0 0.4 0.5 1],...
[1 1.6 1 0 0], [1 1 1 100 100], 0.95);
[H,W,S]=freqz(b,a,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
hold on; plot([0 0.15 0.4 0.5 1], [1 1.6 1 0 0], 'r'); hold off;
Radius takes values between 0 and 1.
1.6
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Compared to the unconstrained iirlpnorm lowpass filter example (refer to
“iirlpnorm Examples” on page 2-40), you see that the filter performance is
about the same, although the ripple in the passband is slightly greater, and the
transition somewhat sharper. The difference between these two designs is the
constraint applied to the poles when you use iirlpnormc with a radius value.
Both filters demonstrate peaks in their passband.
2-46
Advanced IIR Filter Designs
Example — Using iirlpnormc to Design a Bandstop Filter with a
Constrained Pole Radius
Here we use iirlpnormc to design a bandstop filter. Notice that we specify
different orders for the numerator (N=10) and denominator (D=7) and the
frequency and edges vectors are the same. With Radius=.91, none of the 11
filter poles lies farther than 0.91 away from the origin, as you can see in the
zero-pole plot.
f = [0 .25 .35 .7 .8 1];
[b,a]=iirlpnormc(10, 7, f, f, [1 1 0 0 1 1], [1 1 5 5 1 1], .91);
[H,W,S]=freqz(b,a,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
hold on; plot([0 .25 .35 .7 .8 1], [1 1 0 0 1 1], 'r'); hold off;
To generate the zero-pole plot, use zplane(b,a) at the MATLAB prompt.
When we plot the magnitude response curve, the emphasis we placed on
reducing the error in the stopband is clear — note the close match between the
desired and calculated responses. (We weighted the magnitude response
W=[1 1 5 5 1 1] to minimize the error in the vicinity of the stopband
frequency points.)
2-47
2
Designing Advanced Filters
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Example — Using iirlpnormc to Design a High-Order Notch Filter
To create an optimized design for an IIR high-order notch filter, use
iirlpnormc to design the filter. The following code results in the optimal
solution to creating a filter with different numerator and denominator orders,
and with a maximum pole radius of 0.92.
f = [0 0.37 0.399 0.401 0.43 1];
[b,a]=iirlpnormc(2, 17, f, f, [1 1 0 0 1 1], [1 1 2 2 1 1], 0.92);
[H,W,S]=freqz(b,a,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot(H,W,S);
hold on;
plot([0 0.37 0.399 0.401 0.43 1], [1 1 0 0 1 1], 'r'); hold off;
Note the frequency vector entries 0.37, 0.399, 0.401, and 0.43. These
represent the cutoff points for the filter stopband, a fairly narrow filter.
Looking at the filter response plot, you see it is similar to the single-point filter
2-48
Advanced IIR Filter Designs
example we designed with the gremez function (refer to “Example — Designing
a Single-Point Band Filter” on page 2-18).
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
This filter has two pairs of constrained poles.
Example — Using iirlpnormc to Change an Elliptic Filter to a Constrained
Lowpass Filter
Using an elliptic filter design as the initial conditions, with a maximum pole
radius of 0.96, we reduce the pole radius to 0.95 when we use iirlpnormc to
create an optimal filter solution. The result is a filter with the same band edge
frequencies, and a gain in the passband greater than one. The following code
uses the function ellip from Signal Processing Toolbox to create an elliptical
filter. Then we use the function iirlpnormc with the syntax
iirlpnormc(N,D,F,Edges,A,W,Radius,P, Dens,Initnum,Initden). Initnum
and Initden are the initial estimates of the filter numerator and denominator
coefficients. We use be and ae from our elliptic filter as the vectors Initnum and
initden.
2-49
2
Designing Advanced Filters
1.4
1.2
Magnitude
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
[be,ae]=ellip(4,1,20,0.3);
f = [0 0.3 0.323 1];
[b,a]=iirlpnormc(4, 4, f, f, [1 1 0 0], [1 1 1 1], .95,...
[128 128], 20, be, ae);
[H,W,S]=freqz(b,a,1024);
He=freqz(be,ae,1024);
S.plot = 'mag'; S.yunits = 'linear';
freqzplot([H He],W,S);
A few points to think about when you use iirlpnormc. These hints can help you
converge on a good filter design:
• iirlpnormc implements a weighted, least-pth optimization algorithm.
• Check the location and radii of the designed filter poles and zeros.
• If the zeros are on the unit circle and the poles are well inside the circle, try
increasing the numerator order N, or reducing the error weighting (W) in the
stopband.
2-50
Advanced IIR Filter Designs
• If several poles have large radii, and the zeros are well inside the unit circle,
try increasing D, the denominator order, or reducing the error weighting in
the passband.
• As you reduce the pole radius, you may need to increase the denominator
order.
iirgrpdelay Examples
Filter Design Toolbox provides a new filter design function iirgrpdelay for
designing allpass IIR filters that have group delay characteristics that meet
your needs. When you cascade these allpass filters with other IIR filters, they
act as compensating elements. They produce equalized or specified group delay
across the combined filter frequency response while maintaining the IIR filter
pass and stop bands. For more information about group delay in filters, refer
to “Signal Processing Basics” in Signal Processing Toolbox User’s Guide.
Note iirgrpdelay creates allpass filters you use to compensate for the phase
changes caused by other filters. You cannot use iirgrpdelay to create filters
that both filter input signals and compensate for phase changes in output
signals.
In this section, we introduce the function iirgrpdelay through a series of
examples. Each of these examples uses one or more feature provided in the
function. The examples build on one another, although they can be run
separately. By reviewing each example you get an overview of the capabilities
of the design function.
In much the same way that you use other IIR filter design functions to create
filters with arbitrary magnitude response curves, you use iirgrpdelay to
create filters that have arbitrary group delay curves in the filter passband and
stopband. (In most cases, specifying the group delay in the stopband is not
useful; the filter rejects those frequencies by design. Nonetheless, you can
specify the group delay for frequencies that fall within filter stopbands.)
To specify a filter that approximates a given relative group delay, use
iirgrpdelay with the following input argument syntax
iirgrpdelay(N,F,Edges,Gd)
2-51
2
Designing Advanced Filters
where N is the filter order, F is a vector containing frequencies between 0 and
1, Gd is a vector whose elements are the desired group delay at the frequencies
specified in F, and Edges specifies the band edges. Filter order N must be an
even number, and the vectors F and Gd must have the same number of
elements. To let you specify the shape of the group delay within a band or
bands, vectors F and Gd can contain more elements than Edges.
Considering the following ideas can help you design your group delay
compensator:
• After you use iirgrpdelay to design a filter, use freqz, grpdelay, and
zplane to check your design for undesirable features.
• Remember that allpass filters have positive group delay. You cannot develop
allpass filters that have negative group delay characteristics.
• For some difficult filter optimization problems, use the iirgrpdelay syntax
iirgrpdelay(n,d,edgees,a,w,radius,p,dens,initden)
where initden is a vector containing your estimates of the transfer function
coefficients for the denominator. You can use the Pole-Zero editor in Signal
Processing Toolbox to generate values for initden.
• If the poles and zeros of your filter design cluster together, you may need to
increase the filter order or relax the pole radius restriction (if you used one).
Example — Using iirgrpdelay to Design a Filter with a User-Specified
Group Delay Contour
To show the ability to create an arbitrary shape group delay contour in the
passband of an IIR filter, we use iirgrpdelay and specify the group delay we
desire. Notice that we also specify the maximum pole radius of 0.99. We plot
the ideal group delay contour on the figure as well to compare the desired result
to the designed filter.
[b,a,tau] = iirgrpdelay(8, [0 0.1 1], [0 1], [2 3 1],...
[1 1 1], 0.99);
[G,F] = grpdelay(b,a, 0:0.001:1, 2);
plot(F, G); hold on; plot([0 0.1 1], [2 3 1]+tau, 'r'); hold off;
2-52
Advanced IIR Filter Designs
9
8.5
Group Delay
8
7.5
7
6.5
0
0.1
0.2
0.3
0.4
0.5
0.6
Normalized Frequency
0.7
0.8
0.9
1
The straight lines represent the desired group delay contour, the wavy line the
designed contour. The desired group delay, [2 3 1], is relative. Note that the
actual group delay approximates [8 9 7]. If we increase the filter order, to 10
for example, the approximation improves, but the absolute group delay
increases.
One of the output arguments for iirgrpdelay is tau, the resulting group delay
offset. In all cases, filters created by iirgrpdelay have a group delay that
approximates (gd + tau) where gd is the specified relative group delay of the
filter.
When you look at the zero-pole plot for our filter (use the function zplane), you
can see that the poles stay well within the radius constraint. Optimizing the
filter may not result in poles that are near the constraint. Pole constraints
come into play only when needed to limit the optimization. In this example, our
design did not require the constraint to stay within the bounds of the unit
circle.
2-53
2
Designing Advanced Filters
You can verify that this is an allpass filter by plotting the magnitude response
curve for the design. Use freqz(b,a) to plot the curve.
In general, you determine the contour to use for the group delay equalization
of an IIR filter by subtracting the filter group delay from the filter maximum
group delay. In the next example, we use this process to create our lowpass
filter.
Example — Using iirgrpdelay to Design a Lowpass Elliptic Filter with
Equalized Group Delay
The following code designs a pair of filters that together create a lowpass filter
with equalized group delay.
[be,ae] = ellip(4,1,40,0.2); % Lowpass filter
f = 0:0.001:0.2;
g = grpdelay(be,ae,f,2);
g1 = max(g)-g;
[b,a,tau] = iirgrpdelay(8, f, [0 0.2], g1); % Phase compensator
gd = grpdelay(b,a,f,2);
plot(f, g); hold on; plot(f, g+gd, 'r'); hold off;
Cascading the filters is the same as adding the group delay for each filter
frequency-point by frequency-point (g+gd in the plot function input
arguments). In the figure, the lower curve is the group delay for the elliptic
filter. The compensated, or equalized, group delay is the upper curve — an
essentially flat group delay across the passband from 0 to 0.2. Since this
example used the lowpass elliptic filter from our earlier iirlpnorm examples,
you can see that combining these filters results in a lowpass filter with
equalized group delay. Note that the group delay of the combination is twice
the maximum group delay of the reference filter. When you use an allpass filter
to equalize the group delay of a reference filter, the final group delay is the sum
of the group delays of the reference and allpass filters.
2-54
Advanced IIR Filter Designs
40
35
30
Group Delay
25
20
15
10
5
0
0
0.02
0.04
0.06
0.08
0.1
0.12
Normalized Frequency
0.14
0.16
0.18
0.2
To determine the group delay contour necessary to compensate for the phase
effects of our elliptic filter, we use the elliptic filter group delay as a reference.
In the example, we used grpdelay to return vector g containing the group delay
value at many frequencies across the elliptic filter passband. After determining
the maximum group delay in the elliptic filter passband (returned by max(g) in
the example code), we subtract each individual group delay from the maximum
group delay (g1=max(g)-g). The result is vector g1 containing values that
define a curve that is the mirror image of the group delay contour of our elliptic
filter. Then we use g1 as the input group delay values to iirgrpdelay, and the
resulting allpass filter has a group delay contour that equalizes the group delay
of our lowpass elliptic filter, as shown in the figure.
Example — Demonstrating Passband Equalization for a Bandpass
Chebyshev Filter
You can use iirgrpdelay to create filters that compensate for the group delay
of many kinds of filters. In this example, we create an allpass filter that
2-55
2
Designing Advanced Filters
equalizes the group delay of a bandpass filter. In the figure, the lower curve is
the group delay of the bandpass filter and the upper curve is the equalized
group delay for the combination of the bandpass filter and the allpass filter.
Group delay variation across the passband is less than 0.2.
60
Passband
50
Equalized group delay
Group Delay
40
30
Original group delay
20
10
0
−10
0
0.1
0.2
0.3
0.4
0.5
0.6
Normalized Frequency
0.7
0.8
0.9
1
[bc,ac] = cheby1(2,1,[0.3 0.4]); % Bandpass filter design
f = 0.3:0.001:0.4;
g = grpdelay(bc,ac,f,2);
g1 = max(g)-g;
wt = ones(1, length(f));
[b,a,tau] = iirgrpdelay(8, f, [0.3 0.4], g1, wt, 0.95);
f = 0:0.001:1;
g = grpdelay(bc,ac,f,2);
gd = grpdelay(b,a,f,2);
plot(f, g); hold on; plot(f, g+gd, 'r'); hold off;
2-56
Advanced IIR Filter Designs
Example — Demonstrating Passband Equalization for a Bandstop
Chebyshev Filter
Our final example shows how to equalize the group delay in the passband of a
bandstop filter. Since this filter has two passbands, we equalize the group
delay in each band according to the needs of each band. Vectors g1 and g2 in
the example code contain the group delays within each passband of the
bandpass filter. We ignore the stopband group delay for this case. To determine
the group delay contour across both passbands, we concatenate g1 and g2
(using the command g = [g1; g2]), then use the vector g as the basis for the
group delay input argument gx to iirgrpdelay.
[bc,ac] = cheby2(3,1,[0.3 0.8], 'stop'); % Bandstop filter
f1 = 0.0:0.001:0.3;
g1 = grpdelay(bc,ac,f1,2);
f2 = 0.8:0.001:1;
g2 = grpdelay(bc,ac,f2,2);
f = [f1 f2]; g = [g1; g2]; % Concatenate the passband group delays
gx = max(g)-g;
wt = ones(1, length(f));
[b,a,tau] = iirgrpdelay(14, f, [0 0.3 0.8 1], gx, wt, 0.95);
f = 0:0.001:1;
g = grpdelay(bc,ac,f,2);
gd = grpdelay(b,a,f,2);
plot(f, g); hold on; plot(f, g+gd, 'r'); hold off;
The figure shows that our approach works. You see that the group delay in the
passbands is well-equalized (illustrated by the upper curve; the lower curve
presents the nonequalized group delay). The stop band is unaffected, and the
overall equalized group delay variation in the passbands is close to a constant.
2-57
2
Designing Advanced Filters
40
Passband 1 (f1)
35
Passband 2 (f2)
Stopband
30
gd
Group Delay
25
20
15
10
5
0
2-58
g2
g1
0
0.1
0.2
0.3
0.4
0.5
0.6
Normalized Frequency
0.7
0.8
0.9
1
Robust Filter Architectures
Robust Filter Architectures
To this point, we have been considering FIR and IIR filters whose transfer
function is represented by constant coefficients and where the input signals
and coefficients can be any double-precision value from -∞ to +∞. These
systems are in the discrete time domain, with infinite precision values for the
dependent variable, often magnitude. When you represent filters in software,
or in general purpose or special purpose computing hardware, the inputs to the
filters and the filter coefficients can be represented only by discrete values. The
process of converting the infinite precision variables to discrete values is called
quantization and represents a source of error when you implement digital
filters.
Converting to the discrete domain produces three sources of errors:
• Error caused by the discrete representation of infinitely precise information,
such as filter transfer function coefficients or signal amplitude values. Real
systems create error when they quantize amplitude values.
• Analog-to-digital conversion error in the input signal.
• Arithmetic errors resulting from roundoff caused by the limited word length
available to represent the data in the arithmetic process.
Transfer Function Coefficient Quantization Error
To illustrate the effects of converting from continuous to discrete
representations, and to show error sources resulting from quantization,
consider the following first-order IIR filter.
x[n]
y[n]
X
Z-1
v[n]
a
The constant coefficient difference equation that defines this filter is
y [ n ] = αy [ n – 1 ] + x [ n ]
2-59
2
Designing Advanced Filters
where y[n] and x[n] are the output and input signal variables. In transfer
function form, the following equation describes our IIR filter.
1
z
H ( z ) = --------------------- = -----------–
1
z–α
1 – αz
When you implement this filter form in hardware, the filter coefficient
α assumes discrete values that approximate the design value. Therefore, the
actual transfer function that you implement is
z
ˆ ( z ) = ----------H
z – αˆ
ˆ and αˆ are the close approximations to the original H and α in the
where H
filter design. Notice that this transfer function differs from the theoretical
function H(z). As a result, the actual filter response may differ substantially
from the ideal response.
The main effect of transfer function coefficient quantization is to move the
poles and zeros to different locations in the z-plane, away from their desired, or
designed locations (the locations for the ideal, nonquantized coefficient filter).
Moving the poles or zeros can have two effects:
• Changing the frequency response of the quantized filter so it is not the same
as the ideal or designed filter.
• Moving poles from inside to outside the unit circle, causing the quantized IIR
filter to be unstable. Applies only to IIR filters.
Input Sampling Error (A/D Error)
Given the difference equation for our IIR filter, from earlier
y [ n ] = αy [ n – 1 ] + x [ n ]
where x[n] is the sampled output from an analog to digital converter. Sampling
the continuous signal xa(t) results in x[n]. Then the sampled input to the filter
from the A/D convertor, xˆ [ n ] , is
xˆ [ n ] = x [ n ] + e [ n ]
and e[n] is the error in the A/D conversion process. Our discrete input to the
filter no longer matches the continuous signal xa(t). Discrete-time input xa(t)
does not match x[n] because analog-to-digital conversion discretized the input
2-60
Robust Filter Architectures
in time. Similarly, quantized input xˆ [ n ] does not match x[n] because it has
been discretized in amplitude.
Arithmetic Quantization Error
Quantization in arithmetic operations causes another error. For our first-order
filter example, the output from our multiplier v[n] is generated by multiplying
the signal, y[n-1] with the transfer function coefficients, α
v [ n ] = αy [ n – 1 ]
and storing the result. When we quantize the result to fit it into a storage
register, we generate a quantized value vˆ [ n ] that we write as
v [ n ] = v [ n ] + eα [ n ]
where eα[n] is the error sequence resulting from the product quantization
process.
There is another source of errors in digital filter implementation, caused by the
nonlinearity of quantized arithmetic operations. These errors are apparent in
an effect called limit cycling that occurs at the filter output. Limit cycles
usually appear when there is no input to the filter, or the input to the filter is
constant or sinusoidal. For more information on limit cycles and the toolbox
function limitcycle, refer to limitcycle reference page in the online
documentation. To learn more about quantization, refer to Chapter 3,
“Quantization and Quantized Filtering.”
Low Sensitivity Filter Architectures
Quantizing filter coefficients can have serious effects on the performance of
digital filters. As a result of coefficient quantization, the frequency response of
the filter with quantized coefficients can be significantly different from the
desired filter without quantized coefficients. In some cases, the performance of
the quantized filter can make it unsuitable for your application.
Low sensitivity filter architectures, or robust architectures as they are
sometimes called, are interesting because they can reduce the effects of
coefficient quantization. By being inherently less sensitive to coefficient
quantization, these filter architectures withstand the quantization process and
result in filters that retain the performance of the original filter.
2-61
2
Designing Advanced Filters
You can try either of two approaches to designing low sensitivity filters:
• Convert low sensitivity analog filters composed of inductors, capacitors, and
resistors to digital architectures by replacing the analog components and
connections with their digital equivalents so the digital filter approximates
the analog version.
• Develop digital filter implementations that respond directly to the conditions
that create low coefficient sensitivity in a digital filter designs.
Filter Design Toolbox uses the latter approach to provide low sensitivity filter
architectures.
Generally, filter architecture sensitivity ranges from high for direct forms to
very low for coupled allpass forms. For reference, the following list ranks the
filter structures in the toolbox by their sensitivity to coefficient quantization,
from high sensitivity to low:
1 Direct forms
2 Lattice forms
3 Allpass forms
Quantization sensitivity is also a function of the locations of the poles and zeros
for a filter, so use this list for guidance only.
Within the forms, FIR filters tend to be less sensitive than IIR filters. For the
direct forms, second-order section filter implementations are often less
sensitive to coefficient quantization as well.
Filter Design Example That Includes Quantization
To demonstrate the effects of coefficient quantization on the performance of a
filter, this example creates a 5th-order, lowpass elliptic IIR filter. We choose a
cutoff frequency of 0.4π radians (normalized frequency from 0 to 1), passband
ripple less than 0.5 dB, and stopband attenuation of at least 40 dB. In the
figure you see the filter response. We used the Filter Design and Analysis tool
(FDATool) to design the filter. Notice that we used the default filter structure
df2t, or Direct form 2 transposed. When we want to compare the quantized
version of the filter to the floating-point filter, FDATool lets us quantize the
filter and display the filter response curves together.
2-62
Robust Filter Architectures
We could have used the function ellip from Signal Processing Toolbox to
create the filter.
[b,a] = ellip(5,0.5,40,0.4);
The results are identical because FDATool uses the same function to design the
lowpass filter.
We quantize the filter by selecting Turn quantization on. FDATool quantizes
our elliptic filter and displays the magnitude response for both the original (or
reference) filter and the quantized filter. For this quantization process we use
the default coefficient format settings in FDATool. Later in this example we
change the coefficient format to illustrate the effects of changing the word
length used to represent the filter coefficients.
2-63
2
Designing Advanced Filters
Quantizing the coefficients has damaged our filter magnitude response. Our
quantized filter transition band starts much earlier and is much shallower, and
the stopband attenuation has been reduced. When we look at the zero-pole plot
for the unquantized and quantized versions of our filter, we see that
quantization has moved the poles from their designed locations. Coefficient
overflow, rather than sensitivity to quantization, caused the terrible quantized
response in this filter. Coefficient quantization changes filter coefficients by at
most one quantization level. Overflow can change the coefficients by an
arbitrarily large amount. In this case, quantization changed the largest
magnitude coefficient from 2.49 to saturation at 1.0. You can see this from the
coefficient view by selecting Analysis -> View Filter Coefficients. Thus we see
how sensitive this direct form IIR filter is to coefficient quantization.
2-64
Robust Filter Architectures
To continue this example, we look at the effects of changing the coefficient
format from fixed-point, 16-bits to fixed-point, 8-bits. After we make the
desired change, we see the response curves shown in this figure.
2-65
2
Designing Advanced Filters
When you inspect the entries in the Set Quantization Parameters dialog, you
see that we changed the coefficient format to [8 7], meaning we are using
eight-bit wordlength and seven-bit fraction length to represent each filter
coefficient. Changing the coefficient format to 8-bit, fixed point representation
causes the effects shown in the figure — the passband rolls off early, the
transition is less sharp, and the cutoff frequency lies beyond our 0.4
specification.
In FDATool, select Analysis->Pole/Zero Plot to view the poles and zeros for
the 8-bit filter plotted on the unit circle. Or you might select Analysis->View
Filter Coefficients to see the coefficient numerical values for the filter.
One more experiment in this example. We try changing the Direct form II
transposed (df2t) filter structure to use second-order sections, which tend to
be resistant to quantization effects. As we see in the figure, the elliptic filter
2-66
Robust Filter Architectures
that uses second-order sections, even with the 8-bit coefficient format,
performs identically to our reference filter.
In the Quantization Parameters options, you may note that the Input/output
scaling changed when we converted our filter to second-order sections.
Although we did not explicitly change the scaling by using the Scale
transfer-fcn <=1 option, converting the filter structure required that the gain
for the new sections be changed to maintain the same overall gain for the filter.
Thus our converted filter, which now has three sections, has unique scale
factors for each section. The vector entries [0.0625 1 2 1] represent the scale
factors applied to each section. 0.0625 is the scale factor applied to the input,
1 and 2 are the factors applied to the inputs of the second and third sections,
and 1 is applied to the output from the third section. The resulting filter has
the same gain as the original filter.
2-67
2
Designing Advanced Filters
Selected Bibliography
[1] Antoniou, A., Digital Filters: Analysis, Design, and Applications, Second
Edition, McGraw-Hill, Inc. 1993, 330–360.
[2] Mitra, S. K., Digital Signal Processing: A Computer-Based Approach,
McGraw-Hill, Inc, 1998, 573–584.
2-68
3
Quantization and
Quantized Filtering
Binary Data Types . . . . . . . . . .
Digital Filters . . . . . . . . . . . . .
Quantized Filter Types . . . . . . . . .
Quantized Filter Structures . . . . . . .
Data Format for Quantized Filters . . . .
Quantized FFTs and Quantized Inverse FFTs
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Introductory Quantized Filter Example . .
Constructing an Eight-Bit Quantized Filter . . .
Analyzing Poles and Zeros with zplane . . . . .
Analyzing the Impulse Response with impz . . .
Analyzing the Frequency Response with freqz . .
Noise Loading Frequency Response Analysis: nlm
Analyzing Limit Cycles with limitcycle . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 3-7
. 3-8
. 3-10
. 3-10
. 3-11
. 3-12
. 3-13
Fixed-Point Arithmetic . .
Radix Point Interpretation . .
Dynamic Range and Precision
Overflows and Scaling . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Floating-Point Arithmetic
Scientific Notation . . . .
The IEEE Format . . . .
The Exponent . . . . . .
The Fraction . . . . . .
The Sign Bit . . . . . . .
Single-Precision Format . .
Double-Precision Format .
Custom Floating-point . .
Dynamic Range . . . . .
Exceptional Arithmetic . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 3-18
. 3-18
. 3-19
. 3-19
. 3-19
. 3-19
. 3-20
. 3-20
. 3-21
. 3-21
. 3-23
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3-3
3-3
3-4
3-4
3-5
3-6
3-15
3-16
3-16
3-17
3
Quantization and Quantized Filtering
In the Filter Design Toolbox you can implement and analyze single-input
single-output filters either as fixed-point filters, or as custom floating-point
filters. Either type of filter is referred to as a quantized filter.
You can create a quantized filter from a reference filter, that is, a filter whose
coefficients and arithmetic operations you want to quantize in some fashion.
You can also implement quantized FFTs and quantized inverse FFTs in this
toolbox.
You can specify quantization parameters for quantized filters and FFTs with
quantizers. Quantizers specify how data is quantized. You can also quantize
any data set with a quantizer.
When you apply a quantized filter to data, not only are the filter coefficients
quantized to your specification, but so are:
• The data that you filter
• The results of any arithmetic operations that occur during filtering
See “Bibliography” on page 12-2 for a list of relevant references on quantized
filtering.
3-2
Binary Data Types
Binary Data Types
Binary data is coded and stored as ones and zeros.
Binary Data for Fixed-Point Arithmetic
Binary data that is coded for fixed-point arithmetic is characterized by word
length (in bits) and the placement of the radix (binary) point. The radix point
placement determines the fraction length of a binary word, and also
determines how the binary words are scaled. You can specify fixed-point words
in this toolbox with word lengths of up to 53 bits. The fraction length can range
from 0 bits (for integers) to one bit less than the word length.
Binary Data for Floating-Point Arithmetic
Binary data that is coded for floating-point arithmetic is characterized by the
lengths of the fraction (mantissa) and the exponent (or equivalently, by the
word length and the exponent length). In addition to having the ability to
specify the standard IEEE single-precision and double-precision formats, you
can specify filters in a custom floating-point format, with word lengths ranging
from 2 to 64 bits, and exponent lengths ranging from 0 to 11 bits.
Different data coding methods and precisions affect the following:
• The numerical range of the result
• The quantization error
You can use the Filter Design Toolbox to analyze quantized filters, quantized
FFTs, or quantizers, and see how all of these factors affect your filter
performance on data sets.
Digital Filters
Digital computers generate coded binary data. Binary data is usually coded in
a fixed-point or floating-point format. You use digital filters to process binary
data. Digital filters are modeled as discrete-time linear systems.
You can use digital filters to:
• Filter out noise in measurements
• Enhance signals
• Represent signals
3-3
3
Quantization and Quantized Filtering
Quantized Filter Types
You can specify any type of filter in this toolbox as a quantized filter:
• Fixed-point filters
Fixed-point filters are useful for modeling fixed-point Digital Signal
Processing (DSP) processors that operate on data using fixed-point
arithmetic.
• Double-precision floating-point filters
• Single-precision floating-point filters
• Custom floating-point filters
You can use custom floating-point filters in this toolbox to model
floating-point DSP processors that operate on data using specific
floating-point formats.
Quantized Filter Structures
The quantized filters you can implement in this toolbox can have any of the
following structures:
• Direct form I
• Direct form I transposed
• Direct form II
• Direct form II transposed
• Direct form finite impulse response (FIR)
• Direct form FIR transposed filters
• Direct form antisymmetric FIR filters
• Direct form symmetric FIR filters
• Lattice allpass
• Lattice coupled-allpass filters
• Lattice moving average (MA) minimum phase filters
• Lattice MA maximum phase filters
• Lattice autoregressive (AR) filters
• Lattice ARMA filters
3-4
Binary Data Types
• Lattice coupled-allpass power complementary filters
• Single-input single-output state-space filters
Data Format for Quantized Filters
You can specify the precision and dynamic range for fixed-point filters with two
fixed-point data format parameters:
• Word length
• Fraction length
The word length is the length in bits of any binary word. The fraction length is
the length in bits of the binary word up to the radix point.
You can specify the precision, dynamic range, and other quantization
parameters for floating-point filters with two floating-point data format
parameters:
• Word length
• Exponent length
You can specify the precision, dynamic range, and other quantization
parameters when you specify the data format properties. You can specify these
properties using quantizers.
Except for when you specify a double- or single-precision quantized filter, you
can specify the precision and dynamic range for each of the following
quantization results individually:
• Inputs to a filtering operation
• Outputs of a filtering operation
• Quantized filter coefficients
• Sums that result from filtering
• Products that result from filtering
• Terms that are multiplied by filter coefficients (multiplicands)
These filter characteristics allow you to specify different quantization
parameters for data and arithmetic instructions.
3-5
3
Quantization and Quantized Filtering
Quantized FFTs and Quantized Inverse FFTs
You can specify any type of radix two or radix four quantized FFT in this
toolbox with the following data formats:
• Fixed-point FFTs
• Double-precision floating-point FFTs
• Single-precision floating-point FFTs
• Custom floating-point FFTs
The data formats for quantized FFTs are identical to those of quantized filters.
3-6
Introductory Quantized Filter Example
Introductory Quantized Filter Example
Follow the example in this section to:
• Construct a quantized filter.
• Plot the filter’s poles and zeros.
• Plot the filter’s impulse response.
• Plot the filter’s frequency response from the quantized filter coefficients. The
method used does not account for other quantization effects on the frequency
response computation.
• Plot the filter’s frequency response using the noise loading method. This
method takes all quantization effects into account.
• Test the filter for limit cycles.
You can construct quantized filters by either:
• Using the quantized filter constructor function qfilt
• Copying a quantized filter from an existing one
Quantized filters have many properties, including the filter structure and the
quantization formats.
When you use the function qfilt to create a quantized filter Hq, you can either:
• Type
Hq = qfilt
at the command line to accept the default filter properties, and change the
property values later.
• Use a modified syntax for qfilt to set property values when you create Hq.
You can construct quantized filters with any of several filter structures.
Once you construct a filter, use the filter command to apply it to data. In
addition, the following analysis functions apply to quantized filters:
• zplane for pole/zero plots
• impz for quantized impulse response plots
• freqz for computing and plotting the linear frequency response from the
quantized filter coefficients
3-7
3
Quantization and Quantized Filtering
• nlm for estimating and plotting the frequency response using the noise
loading method
• limitcycle for limit cycle detection and analysis
The first three of these Filter Design Toolbox functions are overloaded for
quantized filters. They behave similarly to the functions with the same name
in the Signal Processing Toolbox.
The introductory example presented in this section is included to illustrate
some of the features of this toolbox. In this example, you can use the code
presented to construct an eight-bit fixed-point quantized FIR filter, and
analyze it with the response functions listed above.
To learn more about quantized filters, see Chapter 6, “Working with Quantized
Filters” and Chapter 8, “Quantized Filtering Analysis Examples.”
Constructing an Eight-Bit Quantized Filter
1 Use gremez to design an FIR low-pass filter in the frequency domain with a
normalized cutoff frequency of approximately 0.4 radians/sample. Specify:
- 27 filter coefficients
- Four frequency points [0 .4 .6 1]
- Four corresponding gains [1 1 0 0]
b = gremez(27,[0 .4 .6 1],[1 1 0 0]);
The entries in the vector b are the coefficients of the (numerator) polynomial
of the FIR filter. This is your reference filter.
2 Construct a fixed-point quantized FIR filter from your reference filter with
the following characteristics:
- 8-bit word length for all data formats
- 7-bit fraction length for all data formats
- Direct form finite impulse response (FIR) filter structure
- The 'convergent' method used to round quantized numbers to the
nearest allowable quantized value.
You can create quantized filters using the qfilt command. When you create a
quantized filter, you must enter the vector of reference filter coefficients b in
a cell array by enclosing the coefficients in curly braces, {b}.
3-8
Introductory Quantized Filter Example
Hq = qfilt('fir',{b},'Format',{[8,7],'convergent'})
Hq =
Quantized Direct form FIR filter.
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
0
( 1) 0.0000000
0.001722275146612721
0
( 2) 0.0000000
0.003409515867936453
( 3) -0.0078125
-0.004898115162792102
( 4) -0.0078125
-0.006325311495727597
( 5) 0.0078125
0.009418759615173328
( 6) 0.0156250
0.012524352890295399
( 7) -0.0156250
-0.017394015777896423
( 8) -0.0234375
-0.022634462311564768
( 9) 0.0312500
0.030838037214625479
(10) 0.0390625
0.040914907859937441
(11) -0.0546875
-0.057578619191314129
(12) -0.0859375
-0.084552886463529736
(13) 0.1484375
0.147259140040687880
(14) 0.4453125
0.448759881582659110
(15) 0.4453125
0.448759881582659110
(16) 0.1484375
0.147259140040687880
(17) -0.0859375
-0.084552886463529736
(18) -0.0546875
-0.057578619191314129
(19) 0.0390625
0.040914907859937441
(20) 0.0312500
0.030838037214625479
(21) -0.0234375
-0.022634462311564768
(22) -0.0156250
-0.017394015777896423
(23) 0.0156250
0.012524352890295399
(24) 0.0078125
0.009418759615173328
(25) -0.0078125
-0.006325311495727597
(26) -0.0078125
-0.004898115162792102
0
(27) 0.0000000
0.003409515867936453
0
(28) 0.0000000
0.001722275146612721
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
=
=
=
=
=
=
=
=
=
=
fir
[1]
1
[27]
unitquantizer('fixed', 'convergent', 'saturate',
quantizer('fixed', 'convergent', 'saturate', [8
quantizer('fixed', 'convergent', 'saturate', [8
quantizer('fixed', 'convergent', 'saturate', [8
quantizer('fixed', 'convergent', 'saturate', [8
quantizer('fixed', 'convergent', 'saturate', [8
[8 7])
7])
7])
7])
7])
7])
Notice that the display provides information about the filter and its property
values. For this example, we created a filter whose product and sum quantizer
3-9
3
Quantization and Quantized Filtering
formats are the same size as the coefficient format to illustrate the
quantization effects.
Analyzing Poles and Zeros with zplane
To compare poles and zeros of the reference filter to those of the quantized filter
Hq you just constructed, type
zplane(Hq)
2.5
Quantized zeros
Quantized poles
Reference zeros
Reference poles
2
1.5
Imaginary part
1
0.5
27
27
0
−0.5
−1
−1.5
−2
−2.5
−3
−2
−1
0
1
2
Real part
Notice that the quantized zeros are not very close to the reference poles and
zeros on the plot.
Analyzing the Impulse Response with impz
To compare the impulse response plot of the quantized filter Hq you just
constructed to that of its floating-point reference (b), use the impz command.
impz(Hq)
3-10
Introductory Quantized Filter Example
The impulse response computed by impz is the response of the fixed-point
quantized filter Hq to a quantized impulse.
0.6
Quantized response
Reference response
0.5
0.4
0.3
0.2
0.1
0
−0.1
0
5
10
15
20
25
Analyzing the Frequency Response with freqz
To compare the frequency response plot of the quantized filter Hq you just
constructed to that of its floating-point reference (b), use the freqz command.
freqz(Hq)
3-11
3
Quantization and Quantized Filtering
20
Quantized response
Reference response
Magnitude (dB)
0
−20
−40
−60
−80
−100
−120
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
Phase (degrees)
0
−500
−1000
−1500
−2000
The freqz command computes the linear frequency response of the two filters
whose coefficients are, respectively:
• The quantized filter coefficients
• The reference filter coefficients
Noise Loading Frequency Response Analysis: nlm
You can estimate the frequency response of the filter Hq you just created using
the noise loading method computed with nlm. The noise loading method takes
quantization effects into account. This method estimates the quantization
noise figure when it runs a set of Monte Carlo frequency response calculations
by filtering a set of sinusoids with randomly varying phase.
3-12
Introductory Quantized Filter Example
nlm(Hq)
Noise Loading Method. Noise figure = 7.6573 dB
20
Magnitude (dB)
0
−20
−40
−60
−80
−100
−120
Quantized NLM
Quantized FREQZ
Reference FREQZ
Noise Power Spectrum
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
0
Phase (degrees)
−500
−1000
−1500
−2000
−2500
−3000
Difference Between nlm and freqz for Frequency Response Analysis
The frequency response computed by freqz is determined using the true linear
frequency response of the transfer function associated with the quantized filter
coefficients. It does not take any other quantization effects into account, and is
not computed from the filter structure you specify.
The frequency response computed by nlm is an estimate of the frequency
response that accounts for nonlinear quantization effects due to your choice of:
• Filter structure
• Other quantization parameters
Analyzing Limit Cycles with limitcycle
You can analyze limit cycles of the filter Hq with limitcycle. This function
computes a Monte Carlo simulation to detect the presence of limit cycles.
3-13
3
Quantization and Quantized Filtering
limitcycle(Hq)
No limit cycles detected after 20 Monte Carlo trials.
As is guaranteed for FIR filters, no limit cycles are detected for this model.
3-14
Fixed-Point Arithmetic
Fixed-Point Arithmetic
You can specify how numbers are quantized using fixed-point arithmetic in this
toolbox with two quantities:
• Word length in bits
• Fraction length in bits
Note This toolbox does bit-true fixed-point arithmetic for word lengths of 53
bits and fewer. It simulates fixed-point arithmetic for word lengths greater
than 53 bits, such as 64 bits.
Although the 64-bit fixed-point arithmetic is not be bit-true to the last bit, it
properly handles overflows and the results are almost indistinguishable from
bit-true results when the numbers are scaled properly. For example, (small
numbers + small numbers) work correctly and (large numbers + large
numbers) are right as well, but (large numbers + small numbers) will be
dominated by the large number and some precision loss will occur.
Fraction length can be up to one bit less than the word length.
A general representation for a two’s complement binary fixed-point number is
Word length
Fraction length
bw – 1
Sign bit
bw – 2
…
b5
b4
•
b3
Radix point
b2
b1
b0
Least significant bit
3-15
3
Quantization and Quantized Filtering
where:
• bi are the binary digits (bits, zeros or ones).
• The word length in bits is given by w.
• The most significant bit (MSB) is the leftmost bit. It is represented by the
location of bw-1. In Filter Design Toolbox, this number represents the sign
bit; a 1 indicates the number is negative, and a 0 indicates it is nonnegative.
• The least significant bit (LSB) is the rightmost bit, represented by the
location of b0.
• The radix (binary) point is shown four places to the left of the LSB for this
example.
• The fraction length f is the distance from the LSB to the radix point.
Radix Point Interpretation
Where you place the radix point determines how fixed-point numbers are
interpreted in two’s complement arithmetic. For example, the five bit binary
number:
• 10110 represents the integer –24+22+2 = –10.
• 10.110 represents –2+2–1+2–2 = –1.25.
• 1.0110 represents –2–0+2–2+2–3 = –0.625.
Dynamic Range and Precision
A fixed-point quantization scheme determines the dynamic range of the
numbers that can be applied to it. Numbers outside of this range are always
mapped to fixed-point numbers within the range when you quantize them. The
precision is the distance between successive numbers occurring within the
dynamic range in a fixed-point representation. The dynamic range and
precision depend on the word length and the fraction length.
For a signed fixed-point number with word length w and fraction length f, the
range is from –2w–f–1 to 2w–f–1–2–f.
For an unsigned fixed-point number with word length w and fraction length f,
the range is from 0 to 2w–f–2–f.
In either case the precision is 2–f.
3-16
Fixed-Point Arithmetic
Overflows and Scaling
When you quantize a number that is outside of the dynamic range for your
specified precision, overflows occur. Overflows occur more frequently with
fixed-point quantization than with floating-point, because the dynamic range
of fixed-point numbers is much less than that of floating-point numbers with
equivalent word lengths.
Overflows can occur when you create a fixed-point quantized filter from an
arbitrary floating-point design. You can normalize your fixed-point filter
coefficients and introduce a corresponding scaling factor for filtering to avoid
overflows in the coefficients.
In this toolbox you can specify how you want overflows to be handled:
• Saturate on the overflow
• Wrap on the overflow
3-17
3
Quantization and Quantized Filtering
Floating-Point Arithmetic
Fixed-point numbers are limited in that they cannot simultaneously represent
very large or very small numbers using a reasonable word length. This
limitation is overcome by using scientific notation. With scientific notation, you
can dynamically place the radix point at a convenient location and use powers
of the radix to keep track of that location. Thus, a range of very large and very
small numbers can be represented with only a few digits.
Any binary floating-point number can be represented in floating-point using
±E
scientific notation form as ± F × 2
where F is the fraction or mantissa (of
length f), 2 is the radix or base (binary in this case), and E is the exponent of
the radix (of length e). The floating-point word length w is f+e+1. The extra bit
is for the sign bit.
You can specify single-precision and double-precision floating-point quantized
filters with the Filter Design Toolbox. In addition, you can specify custom
floating-point quantized filters with word lengths of up to 64 bits, and exponent
lengths of up to 11 bits.
See http://www.mathworks.com/company/newsletter/pdf/Fall96Cleve.pdf
for more information on floating-point computation.
Scientific Notation
A direct analogy exists between scientific notation and radix point notation.
For example, scientific notation using five decimal digits for the mantissa
would take the form
p
± d.dddd × 10 = ± ddddd.0 × 10
p–4
= ± 0 .ddddd × 10
p+1
where p is an integer of unrestricted range. Radix point notation using 5 bits
for the mantissa is the same except for the number base
q
± b.bbbb × 2 = ± bbbbb.0 × 2
q–4
= ± 0.bbbbb × 2
q+1
where q is an integer of unrestricted range. The previous equation is valid for
both fixed- and floating-point numbers. For both these data types, the mantissa
can be changed at any time by the processor. However, for fixed-point numbers,
the exponent never changes, while for floating-point numbers, the exponent
can be changed any time by the processor.
3-18
Floating-Point Arithmetic
The IEEE Format
The IEEE 754 Standard for binary floating-point arithmetic has been widely
adopted for use on DSP processors.
This standard specifies four floating-point number formats including singleand double-precision. Each format contains three components:
• Exponent
• Fraction
• Sign bit
The Exponent
In the IEEE format, exponent representations are biased. This means a fixed
value (the bias) is subtracted from the field to get the true exponent value. For
example, if the exponent field is 8 bits, then the numbers 0 through 255 are
represented, and there is a bias of 127. Some values of the exponent are
reserved for flagging inf, NaN, and denormalized numbers, so the true
exponent values range from –126 to 127. If the exponent length is e, the bias is
given by 2e–1–1.
The Fraction
In general, floating-point numbers can be represented in many different ways
by shifting the number to the left or right of the radix point and decreasing or
increasing the exponent of the radix by a corresponding amount. To simplify
operations on these numbers, they are normalized in the IEEE format.
A normalized binary number has a fraction with the form 1.F where F has a
fixed size for a given data type. Since the leftmost fraction bit is always a 1, it
is unnecessary to store this bit and is therefore implicit (or hidden). Thus, an
n-bit fraction stores an n+1-bit number. If the exponent length is e and the
word length is w, then the fraction length f = w–e–1. IEEE also supports
denormalized numbers.
The Sign Bit
IEEE floating-point numbers use a sign/magnitude representation where the
sign bit is explicitly included in the word. Using this representation, a sign bit
of 0 represents a positive number and a sign bit of 1 represents a negative
number. Both the fraction and the exponent can be positive or negative, but
3-19
3
Quantization and Quantized Filtering
only the fraction has a sign bit. The sign of the exponent is determined by the
bias.
Single-Precision Format
The IEEE 754 single precision floating-point format is a 32 bit word divided
into a 1-bit sign indicator s, an 8-bit biased exponent E, and a 23-bit fraction F.
A representation of this format is given below.
b31 b30
s
b22
E
b0
F
The relationship between this format and the representation of real numbers
is given below.
Number Characterization
Value
Normalized, 0<E<255
(–1)s(2E–127)(1.F)
Denormalized, E=0; F≠0
(–1)s(2–126)(0.F)
Zero, E=0; F=0
(–1)s(0)
Otherwise
exceptional value
Denormalized values are discussed in “Exceptional Arithmetic” on page 3-23.
Double-Precision Format
The IEEE 754 double precision (64-bit) floating-point format consists of a 1-bit
sign indicator s, an 11-bit biased exponent E, and a 52-bit fraction F.
A representation of this format is given below.
b63 b62
s
3-20
b51
E
b0
F
Floating-Point Arithmetic
The relationship between this format and the representation of real numbers
is given below.
Number Characterization
Value
Normalized, 0<E<2047
(–1)s(2E–1023)(1.F)
Denormalized, E=0; F≠0
(–1)s(2–1022)(0.F)
Zero, E=0; F=0
(–1)s(0)
Otherwise
exceptional value
Denormalized values are discussed in “Exceptional Arithmetic” on page 3-23.
Custom Floating-point
This toolbox supports custom (nonstandard) IEEE-style floating-point data
types. These data types adhere to the definitions and formulas previously given
for IEEE single- and double-precision numbers.
The fraction length and the bias for the exponent are calculated from the word
length and exponent length you supply. You can specify:
• Any exponent length up to 11 bits
• Any word length greater than the exponent length up to 64 bits
When specifying a custom format, keep in mind that the exponent length
largely determines the dynamic range, while the fraction length largely
determines the precision of the result.
Dynamic Range
A floating-point quantization scheme determines the dynamic range of the
numbers that can be applied to it. Numbers outside of this range are always
mapped to ±inf.
3-21
3
Quantization and Quantized Filtering
The range of representable numbers for an IEEE floating-point number with
word length w, exponent length e, fraction length f = w–e–1, and the exponent
bias given by bias = 2e – 1– 1 is described in the following diagram
negative
overflow
negative numbers
negative
underflow
positive numbers
positive
underflow
positive
overflow
where:
• Normalized positive numbers are defined within the range 21–bias to
(2 – 2–f).2bias.
• Normalized negative numbers are defined within the range –21–bias to
–(2 – 2–f).2bias.
• Positive numbers greater than (2 – 2–f).2bias, and negative numbers greater
than –(2 – 2–f).2bias are called overflows.
• Positive numbers less than 21–bias, and negative numbers less than –21–bias
are either underflows or denormalized numbers.
• Zero is specified by a E=0; F=0.
Overflows and underflows result from exceptional arithmetic conditions.
Exceptional arithmetic is discussed “Exceptional Arithmetic” on page 3-23.
Note You can use the MATLAB functions realmin and realmax to determine
the dynamic range of double-precision floating-point values for your computer.
Precision
The precision is the distance between 1.0 and the next largest floating-point
number. The dynamic range and precision depend on the word length and the
exponent length.
The precision for floating-point numbers is 2–f.
3-22
Floating-Point Arithmetic
Note In MATLAB, floating-point relative accuracy is given by the command
eps which returns the distance from 1.0 to the next largest floating-point
number. For computers that support the IEEE standard for floating-point
numbers, eps = 2–52 or 2.2204 10–16.
Floating-Point Data Type Parameters
The maximum and minimum absolute values, exponent bias, and precision for
the floating-point formats supported by this toolbox are given below.
Table 3-1: Floating-Point Data Type Parameters
Floating-Point
Data Type
Normalized
Minimum
Maximum
Exponent
Bias
Precision
Single
2–126≈10–38
(2–2–23)2127≈3(1038)
127
2–23≈10–7
Double
2–1022≈2(10–308)
(2–2–52)21023≈1.7(10308)
1023
2–52≈10–16
Custom
21–bias
(2–2–f)2bias
2e–1–1
2–f
Due to the sign/magnitude representation of floating-point numbers, there are
two representations of zero, one positive and one negative. For both
representations E = 0 and F = 0.
Exceptional Arithmetic
In addition to specifying a floating-point format, the IEEE 754 Standard for
binary floating-point arithmetic specifies practices and procedures so that
predictable results are produced independent of the hardware platform.
Specifically, denormalized numbers, are defined to deal with exceptional
arithmetic (underflow and overflow).
Denormalized Numbers
Denormalized numbers are used to handle cases of exponent underflow. When
the exponent of the result is too small (i.e., a negative exponent with too large
a magnitude), the result is denormalized by right-shifting the fraction and
leaving the exponent at its minimum value. The use of denormalized numbers
is also referred to as gradual underflow. Without denormalized numbers, the
3-23
3
Quantization and Quantized Filtering
gap between the smallest representable nonzero number and zero is much
wider than the gap between the smallest representable nonzero number and
the next larger number. Gradual underflow fills that gap and reduces the
impact of exponent underflow to a level comparable with roundoff among the
normalized numbers. Thus, denormalized numbers provide extended range for
small numbers at the expense of precision.
For more information about denormalized single- and double-precision
numbers, refer to “Single-Precision Format” on page 3-20 and
“Double-Precision Format” on page 3-20.
3-24
4
Working with Objects
Objects for Quantized Filtering . . . . . . . . . . . 4-2
Constructing Objects . . . . . . . . . . . . . . . . . 4-3
Copying Objects to Inherit Properties . . . . . . . . . . 4-4
Properties and Property Values . . . . . .
Setting and Retrieving Property Values . . . .
Setting Property Values Directly at Construction
Setting Property Values with the set Command .
Retrieving Properties with the get Command . .
Direct Property Referencing to Set and Get Values
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4-5
4-5
4-5
4-6
4-7
4-9
Functions Acting on Objects . . . . . . . . . . . . . 4-10
Using Command Line Help . . . . . . . . . . . . . 4-11
Command Line Help For Nonoverloaded Functions . . . . . 4-11
Command Line Help For Overloaded Functions . . . . . . 4-11
Using Cell Arrays . . . . . . . . . . . . . . . . . . 4-13
Indexing into a Cell Array of Vectors or Matrices . . . . . . 4-13
Indexing into a Cell Array of Cell Arrays . . . . . . . . . 4-14
4
Working with Objects
Objects for Quantized Filtering
The Filter Design Toolbox uses objects to create:
• Quantizers
• Quantized filters
• Quantized FFTs
Concepts you need to know about the objects for quantized filtering in this
toolbox are covered in these sections:
• “Constructing Objects”
• “Copying Objects to Inherit Properties”
• “Properties and Property Values”
• “Setting and Retrieving Property Values”
- “Setting Property Values Directly at Construction”
- “Setting Property Values with the set Command”
- “Retrieving Properties with the get Command”
- “Direct Property Referencing to Set and Get Values”
• “Functions Acting on Objects”
• “Using Command Line Help”
• “Using Cell Arrays”
- “Indexing into a Cell Array of Vectors or Matrices”
- “Indexing into a Cell Array of Cell Arrays”
Note Although the examples in this section use quantized filters, the
techniques discussed here apply to quantizers and quantized FFTs. See
“MATLAB Classes and Objects” in your MATLAB documentation for more
details on object-oriented programming in MATLAB.
4-2
Constructing Objects
You use one of the two methods Filter Design Toolbox offers to construct
objects:
• Use the object constructor function
• Copy an existing object
For example, when you create a quantized filter using the qfilt command, you
are creating a Qfilt object. The Qfilt object implementation relies on MATLAB
object-oriented programming capabilities.
Like other MATLAB structures, objects in this toolbox have predefined fields
called object properties.
You specify object property values by either:
• Specifying the property values when you create the object
• Creating an object with default property values, and changing some or all of
these property values later
For examples of setting quantized filter properties, see “Quantized Filter
Properties” on page 6-6.
Example — Constructor for Quantized Filters
The easiest way to create a quantized filter (qfilt object) is to create one with
the default properties. You can create a quantized filter Hq by typing
Hq = qfilt
MATLAB lists the properties of the filter Hq you created along with the
associated default property values.
Quantized Direct form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
+ (1)
0.999969482421875 1.000000000000000000
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
+ (1)
0.999969482421875 1.000000000000000000
FilterStructure = df2t
ScaleValues = [1]
NumberOfSections = 1
4-3
4
Working with Objects
StatesPerSection =
CoefficientFormat =
InputFormat =
OutputFormat =
MultiplicandFormat =
ProductFormat =
SumFormat =
Warning: 2 overflows
[0]
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
in coefficients.
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
15])
15])
15])
30])
30])
The properties of this filter are described in Table 10-3, Quick Guide to
Quantized Filter Properties, on page 10-10, and in more detail in “Quantized
Filter Properties Reference” on page 10-11. All of these properties are set to
default values when you construct them.
For information on quantizer properties, see “A Quick Guide to Quantizer
Properties” on page 10-2 or “Quantizer Properties Reference” on page 10-3 for
more details.
For information on quantized FFT properties, see “A Quick Guide to Quantized
FFT Properties” on page 10-46, or “Quantized FFT Properties Reference” on
page 10-47 for more details.
Copying Objects to Inherit Properties
If you already have an object with the property values set the way you want
them, you can create a new one with the same property values by copying the
first object.
This feature is convenient to use when you want to change a small number of
properties on a set of objects.
Example — Copying Quantized Filters to Inherit Properties
To create a new quantized filter Hq2 with the same property values as an
existing quantized filter Hq, type
Hq2 = copyobj(Hq);
4-4
Properties and Property Values
Properties and Property Values
All objects in this toolbox have properties associated with them. Each property
associated with an object is assigned a value. You can set the values of most
properties. However, some properties have read-only values.
To learn about properties that are specific to quantized filters, see “Quantized
Filter Properties” on page 6-6.
To learn about properties that are specific to quantizers, see “Quantizer
Properties Reference” on page 10-3.
To learn about properties that are specific to quantized FFTs, see “Quantized
FFT Properties Reference” on page 10-47.
Setting and Retrieving Property Values
You can set Filter Design Toolbox object property values:
• Directly when you create the object
• By using the set command with an existing object
You can retrieve quantized filter property values using the get command.
In addition, direct property referencing lets you either set or retrieve property
values.
Setting Property Values Directly at Construction
To set property values directly when you construct an object, simply include the
following in the argument list for the object construction command:
• A string for the property name you want to set followed by a comma
• The associated property value. Sometimes this value is also a string
Include as many property names in the argument list for the object
construction command as there are properties you want to set directly.
4-5
4
Working with Objects
Example — Setting Quantized Filter Property Values at Construction
Suppose you want to set the following filter characteristics when you create a
fixed-point quantized filter:
• The filter structure has a direct form II transposed structure
• The reference filter transfer function has numerator [1 .5] and denominator
[1 .7 .89]
Do this by typing
Hq = qfilt('FilterStructure','df2t','ReferenceCoefficients',...
{[1 .5] [1 .7 .89]});
These properties are described in “Quantized Filter Properties Reference” on
page 10-11.
Note When you set any object property values, the strings for property
names and their values are case-insensitive. In addition, you only need to type
the shortest uniquely identifying string in each case. For example, you could
have typed the above code as
Hq = qfilt('filt','df2t','ref',{[1 .5] [1 .7 .89]});
Setting Property Values with the set Command
Once you construct an object, you can modify its property values using the set
command.
You can use the set command to both:
• Set specific property values
• Display a listing of all property values you can set
Example — Setting Fixed-Point Quantized Filter Property Values Using set
For example, set the following specifications for the fixed-point filter Hq you
just created:
• Set the input quantization format to [24 23]
• Set the filter structure to a direct form I structure
4-6
Properties and Property Values
To do this, type
set(Hq,'inputformat',[24 23],'filterstructure','df1')
Hq.input.format
ans =
24
23
Hq.filt
ans =
df1
Notice how the display reflects the changes in the property values.
To display a listing of all of the properties associated with a quantized filter Hq
that you can set, type
set(Hq)
QuantizedCoefficients: Quantized from reference coefficients.
ReferenceCoefficients: Cell array of coefficients. One cell per section.
{num,den} | {{num1,den1},{num2,den2},...} |
{num} | {{num1},{num2},...} |
{k} | {{k1},{k2},...} |
{k,v} | {{k1,v1},{k2,v2},...} |
{k1,k2,beta} | {{k11,k21,beta1},{k12,k22,beta2},...} |
{A,B,C,D} | {{A1,B1,C1,D1},{A2,B2,C2,D2},...}
FilterStructure: [df1 | df1t | df2 | <df2t> | fir | firt |
symmetricfir | antisymmetricfir |
latticear | latcallpass |
latticema | latcmax | latticearma |
latticeca | latticecapc | statespace]
ScaleValues: Vector of scale values between sections.
CoefficientFormat: quantizer
InputFormat: quantizer
OutputFormat: quantizer
MultiplicandFormat: quantizer
ProductFormat: quantizer
SumFormat: quantizer
4-7
4
Working with Objects
Retrieving Properties with the get Command
You can use the get command to:
• Retrieve property values for an object
• Display a listing of all the properties associated with an object and the
associated property values
Example — Retrieving Quantized Filter Property Values
For example, to retrieve the value of the quantization data format for the input,
type
v = get(Hq,'FilterStructure')
v =
df1
Note When you retrieve properties, the strings for property names and their
values are case-insensitive. In addition, you only need to type the shortest
uniquely identifying string in each case. For example, you could have typed
the above code as
v = get(Hq,'in');
To display a listing of all of the properties of a quantized filter Hq, and their
values, type
get(Hq)
Quantized Direct Form I (df1) filter.
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
(1) 1.000000000000000
1.000000000000000000
(2) 0.500000000000000
0.500000000000000000
Denominator
QuantizedCoefficients{2}
(1) 1.000000000000000
(2) 0.699981689453125
(3) 0.889984130859375
4-8
ReferenceCoefficients{2}
1.000000000000000000
0.699999999999999960
0.890000000000000010
Properties and Property Values
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
=
=
=
=
=
=
=
=
=
=
df1
[1]
1
[3]
unitquantizer('fixed', 'floor', 'saturate', [16 15])
quantizer('fixed', 'floor', 'saturate', [24 23])
quantizer('fixed', 'floor', 'saturate', [16 15])
quantizer('fixed', 'floor', 'saturate', [16 15])
quantizer('fixed', 'floor', 'saturate', [32 30])
quantizer('fixed', 'floor', 'saturate', [32 30])
Direct Property Referencing to Set and Get Values
You can reference directly into a property for setting or retrieving property
values using MATLAB structure-like referencing. You do this by using a period
(full stop) to index into a property by name.
Example — Direct Property Referencing in Quantized Filters
For example:
1 Create a filter with default values.
2 Change its reference filter coefficients.
Hq = qfilt;
Hq.ref = {[1 .5],[1 .7 .89]};
Notice that you don’t have to type the full name of the ReferenceCoefficients
property name, and you can use lower case to refer to the property name.
To retrieve any property values, you can also use direct property referencing.
v = Hq.ref
v =
[1x2 double]
[1x3 double]
Notice that v is a cell array, and you need to index into it in order to retrieve
its values. See “Using Cell Arrays” on page 4-13 for help with indexing into cell
arrays.
4-9
4
Working with Objects
Functions Acting on Objects
Several functions in this toolbox have the same name as functions in the Signal
Processing Toolbox or in MATLAB. These Filter Design Toolbox functions
behave similarly to their original counterparts, but you apply these functions
directly to an object. This concept of having functions with the same name
operate on different types of objects (or on data) is called overloading of
functions.
For example, the filter command is overloaded for quantized filters (Qfilt
objects). Once you specify your quantized filter by assigning values to its
properties, you can apply many of the functions in this toolbox (such as freqz
for frequency response analysis) directly to the variable name you assign to
your quantized filter, without having to specify filter parameters again.
For a complete list of functions that act on quantizers, see “Functions
Operating on Quantizers” on page 11-10.
For a complete list of functions that act on quantized filters, see “Functions
Operating on Quantized FFTs” on page 11-12.
For a complete list of functions that act on quantized FFTs, see “Functions
Operating on Quantized FFTs” on page 11-12.
4-10
Using Command Line Help
Using Command Line Help
How you get command line help on a function depends on whether the function
is overloaded.
Command Line Help For Nonoverloaded Functions
You can use the usual syntax for getting command line help on functions that
are not overloaded.
Type
help FuncName
to get command line help on functions in this toolbox that are not overloaded.
Command Line Help For Overloaded Functions
Because many of the toolbox functions are overloaded, you need to refer to the
object name when you are trying to get command line help for overloaded
functions.
Command Line Help for Overloaded Functions on Quantized Filters
To get command line help for an overloaded function MethodName that operates
on quantized filters (Qfilt objects), type
help qfilt/MethodName
Similarly, for command line help on overloaded methods for quantizers or
quantized FFTs, type
help quantizer/MethodName
help qfft/MethodName
For example, to get help on the zplane function in this toolbox, type
help qfilt/zplane
You can find a list of the overloaded functions for quantized filters in
“Functions Operating on Quantized FFTs” on page 11-12.
You can find a list of the overloaded functions for quantizers, in “Functions
Operating on Quantizers” on page 11-10.
4-11
4
Working with Objects
You can find a list of the overloaded functions for quantized FFTs, in
“Functions Operating on Quantized FFTs” on page 11-12.
Note Many of the toolbox functions are overloaded. MATLAB does not
necessarily display the appropriate help text for a given object command
MethodName when you type
help MethodName
To get the appropriate help for an overloaded function, you may need to
specify the type of object to which you are applying the function. For example,
help qfilt/MethodName
help qfft/MethodName
4-12
Using Cell Arrays
Using Cell Arrays
The syntax for constructing quantized filters requires that you enter the
reference filter coefficients as cell arrays.
Cell arrays can store any type of data: strings, vectors, matrices, cell arrays,
and so forth. You specify a cell array using curly braces ({}). You need to use
these braces to index into a cell array to retrieve its contents.
When you index into a cell array you use one set of braces to index into each
layer of a cell array.
For details on constructing and using quantized filters in this toolbox, see
Chapter 6, “Working with Quantized Filters.” For detailed information on cell
arrays, see Using MATLAB.
The next sections provide guidance and examples of how to index into a cell
array:
• “Indexing into a Cell Array of Vectors or Matrices” on page 4-13
• “Indexing into a Cell Array of Cell Arrays” on page 4-14
Indexing into a Cell Array of Vectors or Matrices
To index into a cell array of matrices (as opposed to a cell array of cell arrays),
you only need one set of braces to index into the cell array.
Here’s an example of accessing cell array information from a quantized filter
with a single section. In this case, the filter coefficient information is stored as
a cell array of vectors.
Example — Accessing Coefficient Information from Filters with One
Section
You can specify a sample quantized filter by typing
Hq = qfilt('ref',{[1 .5],[1 .7 .89]});
Hq.ReferenceCoefficients
ans =
[1x2 double]
[1x3 double]
Notice that the filter reference coefficients are stored in a two-by-one cell array
of vectors, the way you specified them.
4-13
4
Working with Objects
Suppose that you want to retrieve the values stored in this property.
Use curly braces to index into and access the first entry of the cell array
Hq.ReferenceCoefficients. You can use the shorthand for property names
when you index into the properties of Hq.
Hq.ref{1}
ans =
1.0000
0.5000
Similarly,
Hq.ref{2}
ans =
1.0000
0.7000
0.8900
To access the third entry in Hq.ref{2}, index into Hq.ref{2} in the standard
way.
Hq.ref{2}(3)
ans =
0.8900
Indexing into a Cell Array of Cell Arrays
To index into a cell array of cell arrays, you have to use as many sets of braces
as you have layers of cells.
Here’s an example of indexing into the cell arrays of multisection quantized
filters.
Example — Accessing Coefficient Information from Multisection Filters
When you create quantized filters with multiple sections, specify the reference
filter coefficients as a cell array of cell arrays, using one cell array to enter the
numerator and denominator of each section. In this case, use sequences of curly
braces to index into these cell arrays.
For example, suppose you want to quantize and design a sixth-order
Butterworth filter you create using the Signal Processing Toolbox.
[b,a] = butter(6,.5);
4-14
Using Cell Arrays
Filters whose transfer functions are factored into second-order sections are
much more robust against quantization error, so use sos to put your direct
form II filter into a second-order sections form.
Hq = sos(qfilt('df2',{b,a}));
Hq.ReferenceCoefficients
ans =
{1x2 cell}
{1x2 cell}
{1x2 cell}
The reference coefficients are contained in a three-by-one cell array of cells
Hq.ReferenceCoefficients. This cell array is created from the values you set
for the ReferenceCoefficients property. You can index into one of the three
cell arrays of cells by:
1 Creating a cell array c from the cell array Hq.ReferenceCoefficients
2 Indexing into it
c = Hq.ref;
c{2}{1:2}
ans =
0.2500
ans =
1.0000
0.5012
0.2511
0.0000
0.1716
Notice that you can use the colon operator to obtain the contents of both entries
in the cell array contained in the cell array c{2}.
Note You do not have to create another cell array to index into the reference
coefficients data for one section of the filter. You do have to create another cell
array if you want to index into multiple entries of the cell array, as in this
example.
4-15
4
Working with Objects
4-16
5
Working with Quantizers
Quantizers and Unit Quantizers . . . . . . . . . . . 5-2
Constructing Quantizers . . . . . . . . . . . . . . 5-3
Constructor for Quantizers . . . . . . . . . . . . . . . 5-3
Quantizer Properties . . . . . . . . . . . .
Properties and Property Values . . . . . . . . .
Settable Quantizer Properties . . . . . . . . .
Setting Quantizer Properties Without Naming Them
Read-Only Quantizer Properties . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5-4
5-4
5-4
5-5
5-5
Quantizing Data with Quantizers . . . . . . . . . . 5-7
Example — Data-Related Quantizer Information . . . . . 5-7
Transformations for Quantized Data . . . . . . . . . 5-9
Quantizer Data Functions . . . . . . . . . . . . . . 5-10
5
Working with Quantizers
Quantizers and Unit Quantizers
There are two types of quantizers you can construct in this toolbox:
• Quantizers
• Unit quantizers
These two types of quantizers are the same, except that unit quantizers
quantize any number within the quantization level (eps(q)) of 1 to 1, where q
is a quantizer.
You can construct quantizers to specify quantization parameters you want to
use when you quantize data sets. You can also use quantizers for:
• Specifying data formats for quantized filters or FFTs
• Obtaining information about the data sets you quantize
This chapter covers quantizer-specific information:
• Constructing quantizers
• Quantizer properties
• Quantizing data with quantizers
- Accessing data-related quantization information using a quantizer
• Displaying quantized data in binary or hexadecimal format
• Accessing quantizer data
The quantizers you create in this toolbox are objects with properties. Most of
the basic information you need to know about setting and retrieving property
values is found in Chapter 4, “Working with Objects.” See “Quantizer
Properties Reference” on page 10-3 for information on quantizer properties.
5-2
Constructing Quantizers
Constructing Quantizers
You can construct quantizers by either:
• Using either quantizer constructor function:
- quantizer
- unitquantizer
• Copying a quantizer from an existing one using the copyobj function
Note You can also use the constructor unitquantizer to transform an
existing quantizer into a unit quantizer.
All quantizer parameters are stored as properties that you can set or retrieve.
Some of these quantizer parameters include:
• Quantization format
• Data type (signed or unsigned fixed-point, or double-, single-, or
custom-precision floating-point)
• Rounding method used in quantization
• Overflow method used in quantization
Constructor for Quantizers
The easiest way to create a quantizer is to create one with the default
properties. You can create a quantizer q by typing
q = quantizer
A listing of all of the properties of the quantizer q you just created is displayed
along with the associated property values. All property values are set to
defaults when you construct a quantizer this way. See “Example —
Constructor for Quantized Filters” on page 4-3 for more details.
To construct a unit quantizer q with all of the default quantizer properties, type
q = unitquantizer
5-3
5
Working with Quantizers
Quantizer Properties
Since a quantizer is an object, it has properties associated with it. You can set
the values of some quantizer properties. However, some properties have
read-only values. This sections covers both settable and read-only properties:
• “Settable Quantizer Properties” on page 5-4
• “Read-Only Quantizer Properties” on page 5-5
Properties and Property Values
Each property associated with a quantizer is assigned a value. When you
construct a quantizer, you can assign some of the property values.
Most of the basic information you need to know about setting and retrieving
property values is found in Chapter 4, “Working with Objects.”
A complete list of properties of quantized filters is provided in Table 10-3,
Quick Guide to Quantized Filter Properties, on page 10-10. Properties are
described in more detail in “Quantized Filter Properties Reference” on page
10-11.
Settable Quantizer Properties
You can set the following four quantizer properties:
• Mode property — specifying the data type:
- Fixed-point (signed or unsigned)
- Custom floating-point
- Double-precision floating-point
- Single-precision floating-point
• Format property — specifying quantization format parameters
• OverflowMode property — specifying how overflows are handled in
arithmetic operations
• RoundMode property — specifying the rounding method used in quantization
See “Quantizer Properties Reference” on page 10-3 for full details on all
properties.
5-4
Quantizer Properties
For example, create a fixed-point quantizer with:
• The Format property value set to [16,14]
• The OverflowMode property value set to 'saturate'
• The RoundMode property value set to 'ceil'
You can do this with the following command.
q = quantizer('mode','fixed','format',[16,14],'overflowmode',...
'saturate','roundmode','ceil')
Setting Quantizer Properties Without Naming Them
You don’t have to include quantizer property names when you set quantizer
property values.
For example, you can create quantizer q from the previous example by typing
q = quantizer('fixed',[16,14],'saturate','ceil')
Note You do not have to include default property values when you construct
a quantizer. In this example, you could leave out 'fixed' and 'saturate'.
Read-Only Quantizer Properties
Quantizers have five read-only properties:
• Max
• Min
• NOperations
• NOverflows
• NUnderflows
These properties log quantization information each time you use quantize to
quantize data with a quantizer. The associated property values change each
5-5
5
Working with Quantizers
time you use quantize with a given quantizer. You can reset these values to
the default value using reset.
For an example, see “Example — Data-Related Quantizer Information” on
page 5-7.
5-6
Quantizing Data with Quantizers
Quantizing Data with Quantizers
You construct a quantizer to specify the quantization parameters to use when
you quantize data sets. You can use the quantize function to quantize data
according to a quantizer’s specifications.
Once you quantize data with a quantizer, its data-related, read-only property
values may change.
The following example shows:
• How you use quantize to quantize data
• How quantization affects the read-only properties
• How you reset the read-only properties to their default values using reset
Example — Data-Related Quantizer Information
1 Construct an example data set and a quantizer.
randn('state',0);
x = randn(100,4);
q = quantizer([16,14]);
2 Retrieve the values of the Max and Noverflows properties.
q.max
ans =
reset
q.noverflows
ans =
0
3 Quantize the data set according to the quantizer’s specifications.
y = quantize(q,x);
4 Check the quantizer property values.
q.max
5-7
5
Working with Quantizers
ans =
2.3726
q.noverflows
ans =
15
5 Reset the read-only properties and check them.
reset(q)
q.max
ans =
reset
q.noverflows
ans =
0
5-8
Transformations for Quantized Data
Transformations for Quantized Data
You can convert data values from numeric to hexadecimal or binary according
to a quantizer’s specifications.
Use:
• num2bin to convert data to binary
• num2hex to convert data to hexadecimal
• hex2num to convert hexadecimal data to numeric
• bin2num to convert binary data to numeric
For example,
q = quantizer([3 2]);
x = [0.75
-0.25
0.50
-0.50
0.25
-0.75
0
-1
];
b = num2bin(q,x)
b =
011
010
001
000
111
110
101
100
produces all two’s complement fractional representations of three-bit
fixed-point numbers.
5-9
5
Working with Quantizers
Quantizer Data Functions
Filter Design Toolbox provides a number of data functions to retrieve
information about a quantizer. These functions include:
• denormalmax — the largest denormalized quantized number
• denormalmin — the smallest denormalized quantized number
• eps — the quantization level
• exponentbias — the exponent bias of a quantizer
• exponentlength — the exponent length of a floating-point quantizer
• exponentmax — the maximum exponent allowable for a floating-point
quantizer
• fractionlength — the fraction length of a fixed-point quantizer
• range — the numerical range of a quantizer
• realmax — the largest positive number a quantizer can produce
• realmin — the smallest positive normal number a quantizer can produce
• wordlength — the word length of a quantizer
For example, to find the largest positive quantized number the default
quantizer can create, type
format long
q = quantizer;
r = realmax(q)
r =
0.99996948242188
5-10
6
Working with Quantized
Filters
Constructing Quantized Filters . . . . . . .
Constructor for Quantized Filters . . . . . . . .
Constructing a Quantized Filter from a Reference .
Copying Filters to Inherit Properties . . . . . . .
Changing Filter Property Values After Construction
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Quantized Filter Properties . . . . . . .
Properties and Property Values . . . . . . .
Basic Filter Properties . . . . . . . . . .
Specifying the Filter’s Reference Coefficients .
Specifying the Quantized Filter Structure . .
Specifying the Data Formats . . . . . . . .
Specifying All Data Format Properties at Once
Specifying the Format Parameters with setbits
Using normalize to Scale Coefficients . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 6-6
. 6-6
. 6-6
. 6-7
. 6-8
. 6-9
. 6-10
. 6-11
. 6-12
Filtering Data with Quantized Filters
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6-3
6-3
6-4
6-5
6-5
. . . . . . . . 6-14
Transformation Functions for
Quantized Filter Coefficients . . . . . . . . . . 6-15
6
Working with Quantized Filters
This chapter covers what you need to know to construct and use quantized
filters:
• Constructing quantized filters
• Quantized filter properties
• Filtering data with quantized filters
• Transformation Functions for Quantized Filter Coefficients
The quantized filters you create in this toolbox are objects with properties.
Most of the basic information you need to know about setting and retrieving
property values is found in Chapter 4, “Working with Objects.”
6-2
Constructing Quantized Filters
Constructing Quantized Filters
You can construct quantized filters in the Filter Design Toolbox by either:
• Using the quantized filter constructor function qfilt
• Copying an existing one
All filter characteristics are stored as properties that you can set or retrieve.
Some of these quantized filter characteristics include:
• Filter structure.
• Reference filter coefficients.
• Filter topology (single section or cascaded nth-order sections). The syntax
you use to enter the reference filter coefficients determines the topology.
• Quantized filter data format parameters:
- Quantization parameters (precisions).
- Data type (signed or unsigned fixed-point, or, double-, single-, or
custom-precision floating-point).
- Rounding method used in quantization.
- Overflow method used in quantization.
• Scaling factors for each section of a cascade of nth-order sections.
You can specify quantized filter properties by either:
• Specifying all of the filter properties when you create it
• Creating a quantized filter with default property values, and changing some
or all of these property values later
Constructor for Quantized Filters
The most direct way to create a quantized filter (Qfilt object) is to create one
with the default properties. You create a default quantized filter Hq by typing
Hq = qfilt
A listing of all of the properties of the filter Hq you just created is displayed
along with the associated property values. All property values are set to
defaults when you construct a quantized filter this way.
6-3
6
Working with Quantized Filters
To construct a quantized filter with properties other than the default values,
follow the procedure outlined in “Setting Property Values Directly at
Construction” on page 4-5.
For some examples of using the quantized filter constructor to construct a filter
while specifying some properties at construction, see:
• “Constructing an Eight-Bit Quantized Filter” on page 3-8
• “Example — Accessing Coefficient Information from Filters with One
Section” on page 4-13
• “Example — Accessing Coefficient Information from Multisection Filters” on
page 4-14
Constructing a Quantized Filter from a Reference
In general you construct quantized filters from reference filters. You begin with
a set of unquantized (or quantized) reference filter coefficients to implement in
a quantized filter.
Suppose you design a quantized filter from a fourth-order elliptic filter. You can
use the Signal Processing Toolbox filter design functions to help you. First,
design a filter with parameters in transfer function form.
[b,a] = ellip(4,3,20,.6);
Filters designed with a second-order section topology are more robust against
quantization errors than those composed of higher order transfer functions.
Converting a Filter to Second-Order Sections Form
You can construct a quantized filter in second-order sections form as follows:
1 Create a quantized filter using the elliptic filter’s transfer function
parameters as reference coefficients.
Hq = qfilt('df2t',{b,a});
This filter is not in second-order sections form and has coefficient overflow.
2 Use sos to convert the filter to second-order sections form.
Hq = sos(Hq);
6-4
Constructing Quantized Filters
Copying Filters to Inherit Properties
If you already have a quantized filter Hq with the property values set the way
you want them, you can create a new quantized filter Hq2 with the same
property values as Hq by typing
Hq2 = copyobj(Hq)
This function is convenient to use when you are changing a small number of
properties on a set of filters.
For example, create a 16-bit precision filter Hq from an FIR reference filter with
b = fir1(80,0.5,kaiser(81,8)); % Reference filter
Hq = qfilt('fir',{b})
Except for the filter coefficients provided by {b}, Hq inherits the default property
values for a quantized filter.
Changing Filter Property Values After Construction
Now suppose you want to analyze the response of this same reference filter b
when you:
• Change all of the data format property values using setbits
• Change the ScaleValues property value to [0.5 0.5]
You can do this by first copying Hq, and then changing only those properties you
want to change.
Hq2 = copyobj(Hq);
setbits(Hq2,[16,14])
Hq2.ScaleValues = [0.5 0.5];
Hq2.scale
ans =
0.5000
0.5000
For more information on setting filter properties, see “Setting Property Values
with the set Command” on page 4-6 and “Direct Property Referencing to Set
and Get Values” on page 4-9.
6-5
6
Working with Quantized Filters
Quantized Filter Properties
Since a quantized filter is a Qfilt object, it has properties associated with it.
These properties prescribe the most basic filter qualities, such as the data
format for each data path or the rounding methods used for quantization and
filtering. You can set the values of most properties. However, some properties
have read-only values.
Properties and Property Values
Each property associated with a quantized filter is assigned a value. When you
construct a quantized filter, you assign some of the quantized filter property
values to design a quantized filter to your own specifications. You can set or
retrieve quantized filter properties according to the information in “Setting and
Retrieving Property Values” on page 4-5.
A complete list of properties of quantized filters is provided in Table 10-3,
Quick Guide to Quantized Filter Properties, on page 10-10. Properties are
described in more detail in “Quantized Filter Properties Reference” on page
10-11.
Basic Filter Properties
Basic filter properties include:
• The ReferenceCoefficients property — specifying the filter’s reference
coefficients
• The FilterStructure property — specifying the quantized filter structure
• The data format properties for setting quantization parameters for data and
arithmetic operations:
- CoefficientFormat — specifying how the reference filter coefficients are
quantized
- InputFormat — specifying how the inputs are quantized
- MultiplicandFormat — specifying how data is quantized before it is
multiplied by a coefficient
6-6
Quantized Filter Properties
- OutputFormat — specifying how the outputs are quantized
- ProductFormat — specifying how the results of multiplication are
quantized
- SumFormat — specifying how the results of addition are quantized
See “Quantized Filter Properties Reference” on page 10-11 for full details on all
properties.
Specifying the Filter’s Reference Coefficients
The ReferenceCoefficients property value contains the filter parameters for
the reference filter that specifies your quantized filter. “Constructing a
Quantized Filter from a Reference” on page 6-4 uses the
ReferenceCoefficients property in an example of quantized filter
construction.
The syntax you use for assigning reference filter coefficients depends on the
filter structure and topology you want to assign. See “Assigning Reference
Filter Coefficients” on page 10-34 for more information on the required syntax
for each filter structure and topology.
For example, to assign a direct form II transposed filter structure with one
second-order section for the transfer function
–1
1 + 0.5z
H ( z ) = ----------------------------------------------------–1
–2
1 + 0.7z + 0.89z
type
Hq = qfilt('FilterStructure','df2t','ReferenceCoefficients',...
{[1 .5] [1 .7 .89]});
In this example, you use the constructor qfilt to specify the quantized filter.
You set the FilterStructure and the ReferenceCoefficients property values
at the same time that you specify the filter. All other filter properties retain
their default values.
Notice that you enter the numerator and denominator polynomial coefficients
in one cell array for this filter with one second-order section. In general you
supply the reference filter coefficients for each cascaded section in a quantized
filter in its own cell array.
6-7
6
Working with Quantized Filters
Specifying the Quantized Filter Structure
In Filter Design Toolbox, you can create quantized filters with 16 different
filter structures:
• Direct form I
• Direct form I transposed
• Direct form II
• Direct form II transposed
• Direct form Finite Impulse Response (FIR)
• Direct form FIR transposed
• Direct form antisymmetric FIR (odd and even orders)
• Direct form symmetric FIR filters (odd and even orders)
• Lattice allpass
• Lattice coupled-allpass filters
• Lattice coupled allpass power-complementary filters
• Lattice Moving Average (MA) minimum phase filters
• Lattice MA maximum phase filters
• Lattice Autoregressive (AR) filters
• Lattice ARMA filters
• Single-input single-output state-space filters
Filter structures are described in detail in the description for the property
FilterStructure on page 10-12.
You can create filters with two possible filter topologies:
• A single section
• Cascaded nth-order sections
Topology. You set the topology when you specify the reference filter coefficients
for your quantized filter. See “Assigning Reference Filter Coefficients” on page
10-34 for more information. After you create your quantized filter with the
topology you choose, use Filter Design and Analysis Tool (FDATool) in
quantization mode to change the filter topology. For more information about
FDATool, refer to Chapter 9, “Quantization Tool Overview.”
6-8
Quantized Filter Properties
For example, you can construct a quantized filter with a lattice AR structure
by:
1 Specifying a vector of AR reflection coefficients
k = [.66 .7 .44];
2 Constructing a quantized filter with a lattice AR filter structure
Hq = qfilt('latticear',{k});
Notice that:
• You don’t have to type the 'FilterStructure' property name at
construction
• You specify the reflection reference filter coefficients in a cell array
Specifying the Data Formats
Quantized filters have six data format properties you can set:
• CoefficientFormat
• InputFormat
• MultiplicandFormat
• OutputFormat
• ProductFormat
• SumFormat
Specify the data format property values for quantized filters using quantizers.
For each data format, you can specify:
• Data type
• Quantization format parameters
• Method for handling quantization overflows
• Method for rounding
For example, the quantization format of the CoefficientFormat property for
Hq has the default value of [16,15] (as do all data format properties for this
filter). To change the quantization format for the CoefficientFormat property
value to [16,14], type
6-9
6
Working with Quantized Filters
Hq.CoefficientFormat.Format = [16,14];
Hq.CoefficientFormat.Format
ans =
16
14
Here you are changing the Format property of the quantizer for the
CoefficientFormat property. This syntax leaves all other property values for
the quantizer for the CoefficientFormat property unchanged.
Specifying All Data Format Properties at Once
To implement the quantized lattice filter Hq you just specified using
floating-point calculations, you need to set the Mode property value for each
data format property quantizer for Hq to 'float'. You can do this using the
quantizer syntax for accessing the data format properties. See qfilt for more
information on this syntax.
Hq.quantizer = {'float', [24,8]}
Hq =
Quantized Autoregressive Lattice (latticear) filter.
Lattice
QuantizedCoefficients{1}
ReferenceCoefficients{1}
(1) 0.659988403320313
0.660000000000000030
(2) 0.699996948242188
0.699999999999999960
(3) 0.439994812011719
0.440000000000000000
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
6-10
=
=
=
=
=
=
=
=
=
=
latticear
[1]
1
[3]
quantizer('float',
quantizer('float',
quantizer('float',
quantizer('float',
quantizer('float',
quantizer('float',
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
[24
[24
[24
[24
[24
[24
8])
8])
8])
8])
8])
8])
Quantized Filter Properties
Note The quantizer syntax lets you use one line of code to change the Mode
and Format property values for all data format quantizers. You can also do
this with the following six commands.
Hq.CoefficientFormat = quantizer('float',[24,8])
Hq.InputFormat = quantizer('float',[24,8])
Hq.MultiplicandFormat = quantizer('float',[24,8])
Hq.OutputFormat = quantizer('float',[24,8])
Hq.ProductFormat = quantizer('float',[24,8])
Hq.SumFormat = quantizer('float',[24,8])
Specifying the Format Parameters with setbits
Suppose you want to change all of the arithmetic and quantization data format
parameters for the custom floating-point filter Hq in the previous example to
[24 8]. You can do this in three ways:
• Using the setbits command
• Using the quantizer syntax
• Setting each data format property separately
To do this using the setbits command, type
setbits(Hq,[24,8])
To do this using the quantizer syntax, type
Hq.quantizer = [24,8];
These two commands are equivalent for floating-point filters.
Note The setbits command behaves slightly differently for fixed-point
filters. It doubles the quantization data formats for products and sums.
6-11
6
Working with Quantized Filters
Using normalize to Scale Coefficients
Even though you can specify how overflows are treated, they are not corrected
for automatically. You can use normalize to account for coefficient
quantization overflows for all of the direct form and FIR fixed-point filter
structures. This function normalizes the coefficients and modifies the filter’s
scaling.
For example, if you create an elliptic filter with the Signal Processing Toolbox
and directly quantize it with fixed-point arithmetic, there may be some
coefficient overflows.
[b,a] = ellip(5,2,40,0.4);
Hq = qfilt('df2t',{b,a})
Warning: 5 overflows in coefficients.
A warning is displayed indicating that there are coefficient overflows in this
fixed-point filter. This type of warning is displayed whenever you create a filter
with coefficient overflow and you have MATLAB warning set on.
You can normalize the coefficients and modify the scaling using normalize.
Hq = normalize(Hq)
Hq.ScaleValues
ans =
0.0313
1.0000
Notice that:
• The ScaleValues property value has been modified from its original (default)
value of 1.
• There is no longer any coefficient overflow in Hq.
You can apply normalize to direct form IIR and FIR filters. The
FilterStructure property value must be one of the following:
• 'antisymmetricfir'
• 'df1'
• 'df1t'
• 'df2'
• 'df2t'
6-12
Quantized Filter Properties
• 'fir'
• 'firt'
• 'symmetricfir'
6-13
6
Working with Quantized Filters
Filtering Data with Quantized Filters
You can filter data with quantized filters using the filter function.
Example: Filtering Data with a Quantized Filter
warning on
randn('state',0);
x = randn(100,2);
[b,a] = butter(3,.9,'high');
Hq = sos(qfilt('ReferenceCoefficients',{0.5* b,0.5*a},...
'CoefficientFormat',unitquantizer([26 24])));
y = filter(Hq,x);
Warning: 64 overflows in QFILT/FILTER.
Max
Min
NOverflows
Coefficient
1
-0.7419
0
0.8238
-1
0
Input
2.183
-2.202
64
Output 0.4324
-0.4462
0
Multiplicand
1
-1
0
0.4324
-0.4462
0
Product 0.01276
-0.01224
0
0.4324
-0.4462
0
Sum 0.01276
-0.01221
0
0.2162
-0.2231
0
NUnderflows
0
0
0
0
3
0
0
0
0
0
NOperations
4
6
200
200
1200
1400
1200
1400
600
1000
Notice that a record of the overflows that occurred in filtering is displayed if
you have set warning on.
Use qreport to get this listing when needed as well.
6-14
Transformation Functions for Quantized Filter Coefficients
Transformation Functions for Quantized Filter Coefficients
You can change the display for quantized filter coefficients to:
• Binary, using num2bin
• Hexidecimal, using num2hex
For example, to display the coefficients of the filter Hq you just created as
hexidemimal numbers, type
num2hex(Hq)
Hq.QuantizedCoefficients{1} =
05D8
0655
0B99
0B99
0655
05D8
Hq.QuantizedCoefficients{2} =
7FFF
8000
7FFF
8000
7FFF
CAE8
6-15
6
Working with Quantized Filters
6-16
7
Working with Quantized
FFTs
Constructing Quantized FFTs . . . . . . . . . . . . 7-3
Constructor for Quantized FFTs . . . . . . . . . . . . 7-3
Copying Quantized FFTs to Inherit Properties . . . . . . . 7-4
Quantized FFT Properties . . . . . . .
Properties and Property Values . . . . . . .
Basic Quantized FFT Properties . . . . . .
Specifying the Data Formats . . . . . . . .
Specifying All Data Format Properties at Once
Specifying the Format Parameters with setbits
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7-6
7-6
7-6
7-7
7-8
7-8
Computing a Quantized FFT or Inverse FFT of Data . . 7-10
7
Working with Quantized FFTs
Use quantized fast Fourier transforms (FFTs) to specify quantization
parameters for computing a quantized FFT or inverse FFT.
This chapter covers what you need to know to construct and use quantized
FFTs:
• Constructing quantized FFTs
• Quantized FFT properties
• Computing quantized FFTs and quantized inverse FFTs
The quantized FFTs you create in this toolbox are called QFFT objects. These
objects have properties. Most of the basic information you need to know about
setting and retrieving property values is found in Chapter 4, “Working with
Objects.”
7-2
Constructing Quantized FFTs
You can construct quantized FFTs in the Filter Design Toolbox by either:
• Using the quantized FFT constructor function qfft
• Copying a quantized FFT from an existing one
All quantized FFT characteristics are stored as properties that you can set or
retrieve. Some of these quantized FFT characteristics include:
• The FFT length.
• The radix number. Either 2 or 4.
• The number of sections in the FFT. Computed from the length and radix of
the FFT.
• Quantized FFT data format parameters:
- Quantization parameters (precisions).
- Data type (signed or unsigned fixed-point; or double-, single-, or
custom-precision floating-point).
- Rounding method used in quantization.
- Overflow method used in quantization.
• Scaling factors for each stage of the FFT.
You can specify quantized FFT properties by either:
• Specifying them when you create a quantized FFT
• Creating a quantized FFT with default property values, and changing some
or all of these property values later
Constructor for Quantized FFTs
The easiest way to create a quantized FFT (QFFT object) is to create one with
the default properties. You create a default quantized FFT F by typing
F = qfft
A listing of the properties of the FFT F you just created is displayed along with
the associated property values. All property values are set to defaults when you
create a quantized FFT this way.
7-3
7
Working with Quantized FFTs
To construct a quantized FFT with properties other than the default values,
follow the procedure outlined in “Setting Property Values Directly at
Construction” on page 4-5.
Copying Quantized FFTs to Inherit Properties
If you have a quantized FFT F with the property values set the way you want
them, you can create a new quantized FFT F2 with the same property values
as F by typing
F2 = copyobj(F)
For example, create a length 32, radix 2, FFT F by typing
F = qfft('length',32, 'radix', 2)
F =
Radix = 2
Length = 32
CoefficientFormat = quantizer('fixed',
InputFormat = quantizer('fixed',
OutputFormat = quantizer('fixed',
MultiplicandFormat = quantizer('fixed',
ProductFormat = quantizer('fixed',
SumFormat = quantizer('fixed',
NumberOfSections = 5
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
15])
15])
15])
30])
30])
ScaleValues = [1]
Except for the length and the number of sections, F inherits all of the default
property values for a quantized filter.
Changing Some FFT Property Values After Construction
You can create another quantized FFT F2, which has the same properties as F,
but scales each stage of the FFT differently. To do this:
1 Copy F.
2 Change the ScaleValues property value.
For example, you can do this by typing
F2 = copyobj(F);
F2.ScaleValues = [1 0.5 0.25 0.5 1];
7-4
For more information on setting FFT properties, see “Setting Property Values
with the set Command” on page 4-6 and “Direct Property Referencing to Set
and Get Values” on page 4-9.
7-5
7
Working with Quantized FFTs
Quantized FFT Properties
Since a quantized FFT is a QFFT object, it has properties associated with it.
These properties prescribe the FFT characteristics, such as the FFT length and
the radix number. You can set the values of most properties. However, some
properties have read-only values.
Properties and Property Values
Each property associated with a quantized FFT is assigned a value. When you
construct a quantized FFT, you can assign some of the quantized FFT property
values. You can set or retrieve quantized FFT properties according to the
information in “Setting and Retrieving Property Values” on page 4-5.
A complete list of properties of quantized FFTs is provided in Table 10-6, Quick
Guide to Quantized FFT Properties, on page 10-46. Properties are described in
more detail in “Quantized FFT Properties Reference” on page 10-47.
Basic Quantized FFT Properties
Basic quantized FFT properties include:
• The Radix property — specifying the FFT’s radix number (2 or 4)
• The Length property — specifying the quantized FFT length (a power of the
radix number)
• The data format properties for setting quantization parameters for data and
arithmetic operations:
- CoefficientFormat — specifying how the FFT coefficients (twiddle
factors) are quantized
- InputFormat — specifying how the inputs are quantized
- MultiplicandFormat — specifying how data is quantized before it is
multiplied by a coefficient
- OutputFormat — specifying how the outputs are quantized
- ProductFormat — specifying how the results of multiplication are
quantized
- SumFormat — specifying how the results of addition are quantized
See “Quantized FFT Properties Reference” on page 10-47 for full details on all
properties.
7-6
Quantized FFT Properties
Specifying the Data Formats
Quantized FFTs have six data format properties you can set:
• CoefficientFormat
• InputFormat
• MultiplicandFormat
• OutputFormat
• ProductFormat
• SumFormat
Specify the data format property values for quantized FFTs using quantizers.
For each data format, you can specify:
• Data type
• Quantization format parameters
• Method for handling quantization overflows
• Method for rounding
For example:
1 Create a default quantized FFT F.
2 Change the quantization format parameters for the CoefficientFormat
property value to [16,14].
% Create a default quantized FFT.
F = qfft;
% Display the format of the coefficient quantization.
F.CoefficientFormat.Format
ans =
16
15
% Change the coefficient quantization to [16,14].
F.CoefficientFormat.Format = [16,14];
F.CoefficientFormat.Format
ans =
16
14
7-7
7
Working with Quantized FFTs
Here you are changing the Format property of the quantizer for the quantized
FFT’s CoefficientFormat property. This syntax leaves all other property
values for the quantizer for the CoefficientFormat property unchanged.
Specifying All Data Format Properties at Once
To implement the quantized FFT F you just specified using floating-point
calculations, set the Mode property value for each data format property
quantizer for F to 'float'. You do this using the quantizer syntax for
accessing the data format properties. See qfft for more information on this
syntax.
F.quantizer = {'float', [24,8]}
F =
Radix =
Length =
CoefficientFormat =
InputFormat =
OutputFormat =
MultiplicandFormat =
ProductFormat =
SumFormat =
NumberOfSections
ScaleValues =
2
16
quantizer('float',
quantizer('float',
quantizer('float',
quantizer('float',
quantizer('float',
quantizer('float',
= 4
[1]
'floor',
'floor',
'floor',
'floor',
'floor',
'floor',
[24
[24
[24
[24
[24
[24
8])
8])
8])
8])
8])
8])
Specifying the Format Parameters with setbits
Suppose you want to change all of the arithmetic and quantization data format
parameters for the custom floating-point FFT F in the previous example to [24
4]. You can do this in three ways:
• Using the setbits command
• Using the quantizer syntax
• Setting each data format property separately
To do this using the setbits command, type
setbits(F,[24,4])
To do this using the quantizer syntax, type
7-8
Quantized FFT Properties
F.quantizer = [24,4];
These two commands are equivalent for floating-point FFTs.
Note The setbits command behaves slightly differently for fixed-point FFTs
in that it doubles the quantization data formats for products and sums.
7-9
7
Working with Quantized FFTs
Computing a Quantized FFT or Inverse FFT of Data
To compute a quantized FFT or inverse FFT of a data set:
1 Create a quantized FFT F.
2 Obtain or create the data set.
3 Apply fft to F for a quantized FFT or ifft to F for a quantized inverse FFT.
For example, type
warning on
randn('state',0)
F = qfft;
% Create a quantized FFT.
x = randn(100,3); % Create a sample data set x.
y = fft(F,x);
% Compute a quantized FFT of x.
Warning: 542 overflows in quantized fft.
Max
Coefficient
Min
NOverflows
NUnderflows
NOperations
1
-1
5
4
62
Input
2.309
-2.365
97
0
300
Output
2
-2
71
0
192
Multiplicand
Product
Sum
2
-2
350
0
3840
1
-1
0
0
960
2.414
-2.414
24
0
2400
Notice that a record of the overflows that occurred in filtering is displayed if
you have warnings turned on.
You can also use qreport to get this report.
7-10
8
Quantized Filtering
Analysis Examples
Example — Quantized Filtering of Noisy Speech
Loading a Speech Signal . . . . . . . . . . . . .
Analyzing the Frequency Content of the Speech . . .
Adding Noise to the Speech . . . . . . . . . . .
Creating a Filter to Extract the 3000Hz Noise . . . .
Quantizing the Filter As a Fixed-Point Filter . . . .
Normalizing the Quantized Filter Coefficients . . . .
Analyzing the Filter’s Poles and Zeros Using zplane .
Creating a Filter with Second-Order Sections . . . .
Quantized Filter Frequency Response Analysis . . .
Filtering with Quantized Filters . . . . . . . . .
Analyzing the filter Function Logged Results . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8-3
8-3
8-4
8-4
8-5
8-8
8-8
8-9
8-11
8-12
8-13
8-14
Example — A Quantized Filter Bank . . . . . . . . . 8-16
Filtering Data with the Filter Bank . . . . . . . . . . . 8-17
Creating a DFT Polyphase FIR Quantized Filter Bank . . . 8-17
Example — Effects of Quantized Arithmetic . . .
Creating a Quantizer for Data . . . . . . . . . . .
Creating a Fixed-Point Filter from a Quantized Reference
Creating a Double-Precision Quantized Filter . . . . .
Quantizing a Data Set . . . . . . . . . . . . . .
Filtering the Quantized Data with Both Filters . . . .
Comparing the Results . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8-22
8-22
8-22
8-23
8-23
8-23
8-24
8
Quantized Filtering Analysis Examples
This chapter includes the following examples of how you use the quantized
filtering features of this toolbox:
• “Example — Quantized Filtering of Noisy Speech”
• “Example — A Quantized Filter Bank”
• “Example — Effects of Quantized Arithmetic”
8-2
Example — Quantized Filtering of Noisy Speech
This example covers the following procedure that demonstrates filtering of a
noisy signal:
1 “Loading a Speech Signal” on page 8-3
2 “Analyzing the Frequency Content of the Speech” on page 8-4
3 “Adding Noise to the Speech” on page 8-4
4 “Creating a Filter to Extract the 3000Hz Noise” on page 8-5
5 “Quantizing the Filter As a Fixed-Point Filter” on page 8-8
6 “Normalizing the Quantized Filter Coefficients” on page 8-8
7 “Analyzing the Filter Poles and Zeros Using zplane” on page 8-9
8 “Creating a Filter with Second-Order Sections” on page 8-11
9 “Quantized Filter Frequency Response Analysis” on page 8-12
10 “Filtering with Quantized Filters” on page 8-13
11 “Analyzing the filter Function Logged Results” on page 8-14
Loading a Speech Signal
To load a speech signal contained in a matrix mtlb, along with its associated
sampling frequency Fs, type
load mtlb
If you have speakers and a sound card, you can type
sound(mtlb)
and hear this speech signal.
8-3
8
Quantized Filtering Analysis Examples
Analyzing the Frequency Content of the Speech
Next look at the power spectral density of this signal using the pwelch
command.
n = length(mtlb);
nfft = 128;
pwelch(mtlb,[],[],nfft,Fs)
Welch PSD Estimate
−20
Power Spectral Density (dB/Hz)
−30
−40
−50
−60
−70
−80
0
500
1000
1500
2000
Frequency (Hz)
2500
3000
3500
Adding Noise to the Speech
Now add noise to the speech signal at 3000 hertz (Hz) and 3100 Hz and look at
its power spectral density.
f1 = 3000;
f2 = 3100;
t = (0:n-1)'/Fs;
% Noise frequency in Hz.
% Noise frequency in Hz.
% Time duration of the noise signal.
noise = sin(2*pi*f1*t) + 0.8*sin(2*pi*f2*t);
u = mtlb + noise;
% Add noise to the mtlb signal.
8-4
If you have speakers and a sound card, type
sound(u)
Otherwise, use pwelch to look at the power spectral density for u and compare
it to that of mtlb.
pwelch(u,[],[],nfft,Fs);
Welch PSD Estimate
−25
−30
Power Spectral Density (dB/Hz)
−35
−40
−45
−50
−55
−60
−65
−70
−75
0
500
1000
1500
2000
Frequency (Hz)
2500
3000
3500
Notice the difference between the two power spectral densities in the 3000 to
3100 Hz range.
Creating a Filter to Extract the 3000Hz Noise
Consider this simple notched filter design to remove the 3000Hz noise.
A Notched Filter Design
To design a notched filter in MATLAB to remove noise at a given frequency, for
each frequency you want to remove:
1 Calculate the (normalized) frequency you want to remove in rad/sample.
8-5
8
Quantized Filtering Analysis Examples
2 Place a complex zero on the unit circle at this normalized frequency.
3 Place a stable complex pole close to this zero, but inside the unit circle.
4 Determine the filter numerator and denominator by:
a Specifying factors of the numerator and denominator polynomials using
the pole and zero
b Using conv to multiply the factors by their conjugates
For this example, you want to remove noise at both 3000 Hz and 3100 Hz, so
you can follow these steps for both f1=3000 and f2=3100, and put the two
notched filters together.
Here are the steps for f1=3000. Repeat these for f2=3100 for the final design.
The frequency you want to remove is calculated in rad/sample as
wo = 2*pi*f1/Fs;
A notched filter has a zero on the unit circle at a frequency corresponding to an
angle of wo radians. This removes any noise at this frequency. You can find the
real and imaginary parts (x and y) of the corresponding zero using
rez = cos(wo);
imz = sin(wo);
The next step in the notched filter design is to add a pole close to the zero, but
inside the unit circle. This essentially eliminates the effect of the notched filter
at frequencies other than 3000 Hz, while keeping the filter stable. The closer
the pole is to the zero, the narrower the notch will be.
rez1 = .99*cos(wo);
imz1 = .99*sin(wo);
You can define this portion of the filter’s numerator and denominator
polynomials b and a by introducing the complex conjugate factors and using
conv.
b1 = conv([1 -rez-i*imz],[1 -rez+i*imz]);
a1 = conv([1 -rez1-i*imz1],[1 -rez1+i*imz1]);
8-6
Similarly, you can follow these steps to remove 3100 Hz noise.
b2 = conv([1 -cos(2*pi*f2/Fs)-i*sin(2*pi*f2/Fs)],...
[1 -cos(2*pi*f2/Fs)+i*sin(2*pi*f2/Fs)]);
a2 = conv([1 -0.99*cos(2*pi*f2/Fs)-i*0.99*sin(2*pi*f2/Fs)],...
[1 -0.99*cos(2*pi*f2/Fs)+i*0.99*sin(2*pi*f2/Fs)]);
Finally, put these two filters together and look at the frequency response.
b = conv(b1,b2);
a = conv(a1,a2);
freqz(b,a,512,Fs);
1
Magnitude (dB)
0
−1
−2
−3
−4
−5
−6
0
500
1000
1500
2000
Frequency (Hz)
2500
3000
3500
0
500
1000
1500
2000
Frequency (Hz)
2500
3000
3500
60
Phase (degrees)
40
20
0
−20
−40
−60
8-7
8
Quantized Filtering Analysis Examples
Quantizing the Filter As a Fixed-Point Filter
You can create a direct form II transposed fixed-point quantized filter using the
elliptic filter you just created as a reference. Name the filter Hq1.
Hq1 = qfilt('df2t',{b,a});
Warning: 9 Overflows in coefficients.
Normalizing the Quantized Filter Coefficients
MATLAB displays a warning because the filter you just created has some
coefficient overflow associated with it. You can use the normalize command to
scale the coefficients and account for this overflow.
Hq1 = normalize(Hq1);
In addition to scaling the filter coefficients, the normalize also modifies the
ScaleValues property value to account for the coefficient scaling when you
filter.
Hq1
Hq1 =
Quantized Direct form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
(1)
0.125000000000000 0.125000000000000000
(2)
0.423736572265625 0.423728525076514370
(3)
0.608825683593750 0.608840299517598550
(4)
0.423736572265625 0.423728525076514370
(5)
0.125000000000000 0.125000000000000000
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
(1)
0.125000000000000 0.125000000000000000
(2)
0.419494628906250 0.419491239825749210
(3)
0.596710205078125 0.596724377557198320
(4)
0.411132812500000 0.411143364153216890
(5)
0.120086669921875 0.120074501250000020
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
8-8
=
=
=
=
=
=
df2t
[1 1]
1
[4]
quantizer('fixed', 'round', 'saturate', [16
quantizer('fixed', 'floor', 'saturate', [16
15])
15])
OutputFormat = quantizer('fixed', 'floor', 'saturate', [16 15])
MultiplicandFormat = quantizer('fixed', 'floor', 'saturate', [16 15])
ProductFormat = quantizer('fixed', 'floor', 'saturate', [32 30])
SumFormat = quantizer('fixed', 'floor', 'saturate', [32 30])
Note In this example, the ScaleValues property value is [1 1]. There is
effectively no scaling associated with the sections of this particular filter, even
after it has been normalized. This is because the required scaling for the
numerator and denominator of each filter section is the same.
Analyzing the Filter Poles and Zeros Using zplane
You can apply zplane to a quantized filter to analyze its poles and zeros.
zplane(Hq1)
Quantized zeros
Quantized poles
Reference zeros
Reference poles
1
0.8
0.6
Imaginary part
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1
−0.5
0
Real part
0.5
1
8-9
8
Quantized Filtering Analysis Examples
At first glance, this looks like you’ve done a good job at the fixed point notched
filter design. If you zoom in, you can see that the quantized poles are not really
at the correct angles for the notched filter. This is caused by quantization error.
Quantized zeros
Quantized poles
Reference zeros
Reference poles
0.505
0.5
Imaginary part
0.495
0.49
0.485
0.48
0.475
−0.89
−0.885
−0.88
−0.875
−0.87
Real part
−0.865
−0.86
−0.855
−0.85
Having poles located at incorrect angles is not the only problem in the filter.
There are overflow limit cycles that you detect by
rand('state',0)
limitcycle(Hq1)
resulting in the warning
Overflow limit cycle detected.
To see the destructive behavior of the limit cycles, look at the plot from the
noise loading method nlm.
nlm(Hq1)
The quantized noise loading method is random noise around the filter notches.
Also, zplane(Hq1) shows oscillating behavior for the filter.
8-10
Creating a Filter with Second-Order Sections
Filters whose transfer functions have been factored into second-order sections
are less susceptible to coefficient quantization errors. If you are using a
quantized filter with a transfer function filter structure, you can use sos to
convert the normalized quantized filter to second-order sections form.
Hq2 = sos(Hq1);
Now look at the poles and zeros using zplane.
zplane(Hq2)
Sections 1 − 2
Quantized zeros
Quantized poles
Reference zeros
Reference poles
1
0.8
0.6
Imaginary part
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
1
05
0
05
1
Zoom in, as shown in the next figure, to see that the quantized notched filter
design poles and zeros are lined up the way you designed them.
Also, the overflow limit cycle problem has cleared up. You can verify this with
limitcycle(Hq2)
and
nlm(Hq2)
8-11
8
Quantized Filtering Analysis Examples
By zooming in on the tail of the impulse response, plotted by impz(Hq2), you
see the granular limit cycle, but this is not as big an issue as overflow limit
cycles.
Sections 1 − 2
Quantized zeros
Quantized poles
Reference zeros
Reference poles
0.56
0.54
Imaginary part
0.52
0.5
0.48
0.46
0.44
−0.92
−0.9
−0.88
−0.86
−0.84
−0.82
Real part
−0.8
−0.78
−0.76
Quantized Filter Frequency Response Analysis
You can use freqz to analyze the frequency response of a quantized filter.
[H,F,units,Hr] = freqz(Hq2,512,Fs);
This syntax allows you to compare the frequency response H of the quantized
filter, to that (Hr) of the reference filter.
plot(F,20*log10(abs([H Hr])));
ylabel('Magnitude (dB)')
xlabel('Frequency (Hz)')
legend('Quantized','Reference',3)
8-12
Quantized and Unquantized Notched Filter
5
0
Magnitude (dB)
−5
−10
−15
−20
−25
−30
0
500
1000
1500
2000
Frequency (Hz)
2500
3000
3500
4000
The two responses are almost identical.
Filtering with Quantized Filters
Now that you’ve designed a quantized filter you are happy with, use the filter
command to apply it to the noisy speech signal and see how well it does.
y = filter(Hq2,u/5);
This scaling of the input is to avoid overflows.
You can listen to the filtered speech signal by typing
sound(y)
8-13
8
Quantized Filtering Analysis Examples
Sounds pretty good. The power spectral density also looks like the original.
pwelch(y,[],[],nfft,Fs)
Welch PSD Estimate
−40
−45
Power Spectral Density (dB/Hz)
−50
−55
−60
−65
−70
−75
−80
−85
−90
0
500
1000
1500
2000
Frequency (Hz)
2500
3000
3500
Analyzing the filter Function Logged Results
You can use an alternate syntax for the filter command to monitor the
maximum and minimum values as well as the overflows and underflows that
occur during filtering. Suppose you didn’t realize there would be input
overflows and hadn’t scaled the input.
warning on
y = filter(Hq2,u);
Warning: 1557 overflows in QFILT/FILTER.
Max
Min
Coefficient
0.8612
0.49
0.8699
0.4901
Input
4.127
-3.665
Output
1
-1
Multiplicand
1.81
-1.759
2
-2
Product
1.81
-1.759
2
-2
8-14
NOverflows
0
0
1557
0
2637
1964
0
0
NUnderflows
0
0
0
0
0
0
0
0
NOperations
6
6
4001
4001
32008
28007
32008
28007
Sum
1.095
1.688
-1.248
-1.604
0
0
0
0
20005
20005
A report of all underflows and overflows is displayed when you filter the data.
qreport(Hq2) provides the logged function output as well.
8-15
8
Quantized Filtering Analysis Examples
Example — A Quantized Filter Bank
You can use filter banks to create a set of filters that partition input signals
into separate frequency bands or channels. Discrete Fourier Transform (DFT)
polyphase FIR filter banks [3] provide a computationally efficient way to
implement a filter bank that supports a large number of channels. Some cell
phone base stations use DFT polyphase FIR fixed-point filter banks.
The polyphase DFT FIR filter bank is equivalent to a bank of long FIR filters
operating at a relatively high sample rate.
A model for a polyphase DFT FIR filter bank is shown below. The impulse
response coefficients of the original FIR filter are sampled and partitioned
among the 16 FIR filters Hi(z), i=1, ... , 16. The incoming signal is successively
delayed and downsampled, before it enters any of the FIR filters. The outputs
of the FIR filters are then scaled and sent through an FFT. The 16 outputs of
the FFT represent the 16 channel signals.
X (z)
16
H1(z)
X1 (z)
16
H2(z)
X2 (z)
-1
z
z-1
FFT
z-1
16
H16(z)
X16 (z)
Figure 8-1: Model for a Polyphase DFT FIR Filter Bank
You can follow the example in this section to create a bank of DFT polyphase
FIR fixed-point filters using quantized filters and quantized FFTs.
8-16
Example — A Quantized Filter Bank
Filtering Data with the Filter Bank
To implement the filter bank shown in Figure 8-1, Model for a Polyphase DFT
FIR Filter Bank, on page 8-16:
1 Create a quantized filter bank of 16 FIR filters followed by a quantized FFT.
For linear analysis, adjust the ScaleValues property of the quantized FFT
so that no overflows occur.
2 Successively delay and downsample an incoming data stream so that every
ith signal sample enters the ith FIR filter.
3 Filter the data through the bank of FIR filters using filter on each
quantized filter in the bank.
4 Put the output of the bank of filters through a 16-point FFT using fft on the
quantized FFT.
5 Rescale the output of the FFT to account for the scaling introduced by its
ScaleValues property.
Creating a DFT Polyphase FIR Quantized Filter Bank
This example follows the five steps listed in “Filtering Data with the Filter
Bank” using a set of unit sinusoids at different frequencies for the incoming
data.
This demo takes some time to run and produces the two frequency response
plots shown after the example code. You only see eight channels of filters in the
magnitude response of the filter bank because FFTs produce conjugate signals
for real-valued inputs. The second figure shows all 16 channels, presenting the
channel amplitude for each channel.
%
%
%
M
N
Create a DFT Polyphase FIR Quantized Filter Bank.
Initialize two variables to define the filters and the filter
bank.
= 16; % Number of channels in the filter bank.
= 8;
% Number of taps in each FIR filter.
% Calculate the coefficients b for the prototype lowpass filter,
% and zero-pad so that it has length M*N.
b = fir1(M*N-2,1/M);
8-17
8
Quantized Filtering Analysis Examples
b = [b,zeros(1,M*N-length(b))];
%
%
%
B
Reshape the filter coefficients into a matrix whos rows
represent the individual polyphase filters to be distributed
among the filter bank.
= flipud(reshape(b,M,N));
Hq = cell(M,1);
for k=1:M
Hq{k} = qfilt('fir',{B(k,:)});
end
% Create a quantized FFT F of length M.
% Set the ScaleValues property value according to the
% NumberOfSections property value. Scale each section by 1/2.
F = qfft('length',M,'scale',0.5*ones(1,log2(M)));
% Retain the FFT scaling to weight the FFT correctly.
g = 1/prod(F.ScaleValues);
%
%
%
%
Construct a bank of M quantized filters and an M-point quantized
FFT. Filter a sinusoid that is stepped in frequency from 0 to
pi radians, store the power of the filtered signal, and plot the
results for each channel in the filter bank.
Nfreq = 200; % Number of frequencies to sweep.
w = linspace(0,pi,Nfreq); % Frequency vector from 0 to pi.
P = 100;
% Number of output points from each channel.
t = 1:M*N*P; % Time vector.
HH = zeros(M,length(w)); % Stores output power for each channel.
for j=1:length(w)
disp([num2str(j),' out of ',num2str(length(w))])
x = sin(w(j)*t);
% Signal to filter
%
%
%
X
X
8-18
EXECUTE THE FILTER BANK:
Reshape the input so that it represents parallel channels of
data going into the filter bank.
= [x(:);zeros(M*ceil(length(x)/M)-length(x), 1)];
= reshape(X,M,length(X)/M);
Example — A Quantized Filter Bank
% Make the output the same size as the input.
Y = zeros(size(X));
% FIR filter bank.
for k=1:M
Y(k,:) = filter(Hq{k},X(k,:));
end
% FFT
Y = fft(F,Y);
HH(:,j) = var(Y.')';
end
% Store the output power.
% Compensate for FFT scaling.
s = 1/prod(scalevalues(F));
HH = HH*s^2;
% Plot the results.
figure(1)
plot(w,10*log10(HH))
title('Filter Bank Frequency Response')
xlabel('Frequency (normalized to channel center)')
ylabel('Magnitude Response (dB)')
set(gca,'xtick',(1:M/2)*w(end)/M*2)
set(gca,'xticklabel',(1:M/2))
figure(2)
strips(HH')
set(gca,'yticklabel',1:M)
set(gca,'xtick',(1:M/2)*Nfreq/M*2)
set(gca,'xticklabels',(1:M/2))
grid off
title('Filter Bank Frequency Response')
xlabel('Frequency (normalized to channel center)')
ylabel('Channel, Amplitude in Each Channel')
Look at the next two figures to see the results of the example code.
8-19
8
Quantized Filtering Analysis Examples
Filter Bank Frequency Response
0
−10
−20
Magnitude Response (dB)
−30
−40
−50
−60
−70
−80
−90
−100
8-20
0
0.5
1
1.5
2
2.5
Frequency (normalized to channel center)
3
3.5
Example — A Quantized Filter Bank
Filter Bank Frequency Response
16
15
14
Channel, Amplitude in Each Channel
13
12
11
10
9
8
7
6
5
4
3
2
1
1
2
3
4
5
Frequency (normalized to channel center)
6
7
8
8-21
8
Quantized Filtering Analysis Examples
Example — Effects of Quantized Arithmetic
When you filter data with a fixed-point quantized filter, your results may vary
from those obtained by filtering with a double-precision reference filter. This is
due to a number of factors, including:
• Quantization of the input to the filter
• Quantization of the output from the filter
• Quantization of the filter coefficients
• Quantization occurring during the various arithmetic operations performed
by the filter
You can isolate the effects of fixed-point quantization that result solely from
arithmetic operations by:
1 Creating quantizer q for data.
2 Creating a fixed-point filter Hq from a reference using quantizer q data
formats.
3 Creating a double-precision quantized filter Hd from Hq, with the same
(quantized) coefficients.
4 Quantizing a data set x according to the quantizer specifications.
5 Filtering the quantized data set x with both filters.
6 Comparing the results.
Creating a Quantizer for Data
Create a 16-bit default quantizer.
q = quantizer;
Creating a Fixed-Point Filter from a Quantized
Reference
1 Create an example double-precision reference filter and quantize and scale
the filter coefficients.
8-22
Example — Effects of Quantized Arithmetic
[b,a] = ellip(7,.1,40,.4);
c = quantize(q,{b/8, a/8}); % Coefficients are in a cell array.
2 Create a fixed-point quantized filter from the coefficients c, with data
formats specified by the quantizer q.
Hq = qfilt('df2t',c,'quantizer',q);
Creating a Double-Precision Quantized Filter
You can create a quantized double-precision filter Hd from Hq by changing the
value of the Mode property for each of the quantizers that specify the data
formats of Hq.
Hd = Hq;
Hd.quantizer = 'double';
Quantizing a Data Set
Create a random data set and quantize it.
rand('state',0);
n = 1000;
x = quantize(q,0.5*(2*rand(n,1) - 1));
This data set is scaled to prevent overflows. If you do not prevent overflows, you
cannot isolate the quantization effects of arithmetic.
Filtering the Quantized Data with Both Filters
Filter the quantized data with the double-precision filter and the fixed-point
filter.
yq = filter(Hq,x);
yd = filter(Hd,x);
8-23
8
Quantized Filtering Analysis Examples
Comparing the Results
Analyze the error signal and its histogram.
e = yd - yq;
hist(e,20)
140
120
100
80
60
40
20
0
−0.02
−0.015
−0.01
−0.005
0
0.005
0.01
0.015
0.02
0.025
The error is approximately normally distributed. The nonzero mean is caused
by choosing 'floor' for the rounding method.
8-24
9
Quantization Tool
Overview
Switching FDATool to Quantization Mode
. . . . . . 9-3
Getting Help for FDATool . . . . . . . . . . . . . . 9-5
Context-Sensitive Help: The What’s This? Option . . . . . 9-5
Additional Help for FDATool . . . . . . . . . . . . . . 9-5
Quantizing Filters in the Filter Design and
Analysis Tool . . . . . . . . . . . . . . . . . 9-6
To Quantize Reference Filters . . . . . . . . . . . . . 9-10
To Change the Quantization Properties of Quantized Filters . 9-11
Choosing Your Quantized Filter Structure . . . . . . 9-12
Converting the Structure of a Quantized Filter . . . . . . 9-12
To Change the Structure of a Quantized Filter . . . . . . 9-13
Scaling Transfer Function Coefficients . . . . . . . . 9-16
To Scale Transfer Function Coefficients . . . . . . . . . 9-16
Scaling Inputs and Outputs of Quantized Filters . . . 9-18
To Enter Scale Values for Quantized Filters . . . . . . . . 9-18
Importing and Exporting Quantized Filters . . . . . . 9-20
To Import Quantized Filters . . . . . . . . . . . . . . 9-21
To Export Quantized Filters . . . . . . . . . . . . . . 9-22
9
Quantization Tool Overview
Filter Design Toolbox adds a new dialog and operating mode to the Filter
Design and Analysis Tool (FDATool) provided by the Signal Processing
Toolbox. From the new dialog, titled Set Quantization Parameters, you can:
• View Simulink models of the filter structures available in the toolbox.
• Quantize double-precision filters you design in this GUI using the design
mode.
• Quantize double-precision filters you import into this GUI using the import
mode.
• Perform analysis of quantized filters.
• Scale the transfer function coefficients for a filter to be less than or equal to 1.
• Select the quantization settings for the properties of the quantized filter
displayed by the tool:
- Coefficient
- Input
- Output
- Multiplicand
- Product
- Sum
• Change the input and output scale values for a filter.
After you import a filter in to FDATool, the options on the quantization dialog
let you quantize the filter and investigate the effects of various quantization
settings.
This section presents the following information and procedures for using
FDATool:
• “Switching FDATool to Quantization Mode” on page 9-3
• “Quantizing Filters in the Filter Design and Analysis Tool” on page 9-6
• “Choosing Your Quantized Filter Structure” on page 9-12
• “Scaling Transfer Function Coefficients” on page 9-16
• “Scaling Inputs and Outputs of Quantized Filters” on page 9-18
9-2
Switching FDATool to Quantization Mode
Switching FDATool to Quantization Mode
You use the quantization mode in FDATool to quantize filters. Quantization
represents the third operating mode for FDATool, along with the filter design
and import modes. To switch to quantization mode, open FDATool from the
MATLAB command prompt by entering
fdatool
When FDATool opens, click Set Quantization Parameters. FDATool switches
to quantization mode and you see the following dialog, with the default values
shown. Controls within the dialog let you quantize filters and investigate the
effects of changing quantization settings. To enable the quantization options,
select the Turn quantization on check box.
9-3
9
Quantization Tool Overview
You use the following controls in the dialog to perform tasks related to
quantizing filters in FDATool:
• Turn quantization on check box — quantizes the filter displayed in
Current Filter Information.
• Set Quantization Parameters — changes Filter Design and Analysis Tool
to quantization mode to configure and quantize filters that you design or
import.
• Scale transfer-fcn coeffs <=1 — scales the filter transfer function
coefficients to be less than or equal to one.
• Input/Output scaling — reports the results of scaling the transfer function
coefficients. You can enter a vector containing scale values, or enter the
name of a variable in your workspace that contains the scale values to use
for scaling.
• Show filter structures — opens the quantization demonstration program
that shows the filter structures available in Filter Design Toolbox.
• Apply — applies changes you make to the quantization parameters or input/
output scaling for your filter.
• Quantizer property lists, such as Coefficient and Multiplicand — these
lists let you set values for the properties of the quantizers that constitute
your quantized filter. Under Format, the entries contain [wordlength
fractionlength] for each quantizer property.
9-4
Getting Help for FDATool
Getting Help for FDATool
To find out more about the buttons or options in the FDATool dialogs, use the
What’s This? button to access context-sensitive help.
Context-Sensitive Help: The What’s This? Option
To find information on a particular option or region of the dialog:
1 Click the What’s This? button
Your cursor changes to
.
.
2 Click on the region or option of interest.
For example, click Turn quantization on to find out what the option does.
You can also select What’s this? from the Help menu to launch
context-sensitive help.
Additional Help for FDATool
For help about importing filters into FDATool, or for details about using
FDATool to create and analyze double-precision filters, refer to “Filter Design
and Analysis Tool Overview” in Signal Processing Toolbox User’s Guide.
9-5
9
Quantization Tool Overview
Quantizing Filters in the Filter Design and Analysis Tool
Quantized filters have properties that define how they quantize data you filter.
Use the Set Quantization Parameters dialog in FDATool to set the properties.
Using options in the Set Quantization Parameters dialog, FDATool lets you
perform a number of tasks:
• Create a quantized filter from a reference filter after either importing the
reference filter from your workspace, or using FDATool to design the
reference filter.
• Create a quantized filter that has the default structure (Direct form II
transposed) and other property values you select.
• Change the quantization property values for a quantized filter after you
design the filter or import it from your workspace.
When you click Set Quantization Parameters, the dialog opens in FDATool,
with all options set to default values.
To let you set the properties for the six quantizers that make up a quantized
filter, FDATool lists each quantizer. Table 9-1 lists each component quantizer,
9-6
Quantizing Filters in the Filter Design and Analysis Tool
its full property name, and includes a short description of what the quantizer
does in the filter.
Table 9-1: These Quantizers Define the Behavior of a Quantized Filter
Quantizer
Filter Property Name
Description
Coefficient
CoefficientFormat
Determines how the coefficient quantizer handles
filter coefficients. When you quantize a filter, the
properties of this quantizer govern the
quantization.
Input
InputFormat
Specifies how data input to the filter is quantized.
Output
OutputFormat
Specifies how date output by the filter is quantized.
Multiplicand
MultiplicandFormat
Specfies how filter multiplicands are quantized.
Multiplicands are the inputs to multiply operations.
Product
ProductFormat
Determines how to quantize the results of multiply
operations.
Sum
SumFormat
Determines how to quantize the results of
arithmetic sums in the filter.
Every quantizer has five properties. For each quantizer, such as Coefficient
and Output, you select values for its properties to determine how the filter
performs quantization. The properties that make up each quantizer in
a quantized filter are listed in Table 9-2.
9-7
9
Quantization Tool Overview
Table 9-2: Five Properties Specify Each Quantizer
Quantizer Property
Description
Quantizer type
Specifies how the quantizer quantizes values in the
category, such as inputs, or products. Your choices
are:
• quantizer — quantize all values according to the
settings for the other properties of the quantizer.
quantizer is the default setting.
• unitquantizer — quantize values as above,
except quantize values that lie between eps of the
quantizer and 1 to be equal to 1.
Mode
Selects one of four arithmetic modes for the
quantizer:
• fixed — to specify fixed-point arithmetic. fixed is
the default setting.
• float — to specify floating-point arithmetic
• double — to specify double-precision arithmetic
• single — to specify single-precision arithmetic
9-8
Quantizing Filters in the Filter Design and Analysis Tool
Table 9-2: Five Properties Specify Each Quantizer (Continued)
Quantizer Property
Description
Round mode
Sets the way in which the quantizer handles values
after it quantizes them. You have five options to
choose from:
• ceil — round values to the nearest integer
towards plus infinity.
• convergent — round values to the nearest
integer, except in a tie, then round down if the
next-to-last bit is even, up if odd.
• fix — round values to the nearest integer
towards zero.
• floor — round values to the nearest integer
towards minus infinity. The default setting for all
quantizers except the Coefficient quantizer.
• round — round values to the nearest integer.
Negative numbers that lie halfway between two
values are rounded towards negative infinity.
Positive numbers that lie halfway between two
values are rounded towards positive infinity. Ties
round toward positive infinity. The default setting
for the Coefficient quantizer.
9-9
9
Quantization Tool Overview
Table 9-2: Five Properties Specify Each Quantizer (Continued)
Quantizer Property
Description
Overflow mode
When the result of a quantization operation exceeds
the range that the format can represent, this value
tells the quantizer how to handle the overflow.
Choices are
• saturate — set values that fall outside the
representable range to the minimum or maximum
values in the range. Values greater than the
maximum value are set to the maximum range
value. Values less than the minimum value are
set to the minimum range value. This is the
default setting.
• wrap — map values that fall outside the
representable range of the format back into the
range using modular arithmetic.
Format
Specifies the word length and fraction length for the
Mode value you specified. [16 15] is the default
setting for word length and fraction length. Notice
that the Product and Sum quantizers default to
[2*word length 2*fraction length], or [32 30].
To Quantize Reference Filters
When you are quantizing a reference filter, follow these steps to set the
Coefficient property values that control the quantization process. Before you
begin, verify that Turn quantization on is not selected:
1 Click Set Quantization Parameters to open the Set Quantization
Parameters dialog.
2 Select Turn quantization on.
When you turn quantization on, FDATool quantizes the current filter
according to the Coefficient properties, and changes the information
displayed in the analysis area to show quantized filter data.
9-10
Quantizing Filters in the Filter Design and Analysis Tool
3 Review the settings for the Coefficient properties: Quantizer type, Mode,
Round mode, Overflow mode, and Format.
4 Change the Coefficient properties as required to quantize your filter
correctly.
5 Click Apply.
FDATool quantizes your filter using the new settings.
6 Use the analysis features in FDATool to determine whether the new
quantized filter meets your requirements.
To Change the Quantization Properties of
Quantized Filters
When you are changing the property values for a quantized filter, or after you
import a quantized filter from your MATLAB workspace, follow these steps to
set the property values for the quantized filter:
1 Verify that the current filter is quantized.
2 Click Set Quantization Parameters to open the Set Quantization
Parameters dialog.
3 Review and select property settings for the filter quantizers: Coefficient,
Input, Output, Multiplicand, Product, and Sum. Settings for these
properties determine how your filter quantizes data during filtering
operations.
4 Click Apply to update your current quantized filter to use the new
quantization property settings from Step 2.
5 Use the analysis features in FDATool to determine whether your new
quantized filter meets your requirements.
9-11
9
Quantization Tool Overview
Choosing Your Quantized Filter Structure
FDATool lets you change the structure of any quantized filter. Use the Convert
structure option to change the structure of your filter to one that meets your
needs.
Converting the Structure of a Quantized Filter
You use the Convert structure option to change the structure of filter. When
the current filter source is Designed(Quantized) or Imported(Quantized),
Convert structure lets you recast the filter to one of the following structures:
• “Direct Form II Transposed Filter Structure” on page 10-19
• “Direct Form I Transposed Filter Structure” on page 10-17
• “Direct Form II Filter Structure” on page 10-18
• “Direct Form I Filter Structure” on page 10-16
• “Direct Form Finite Impulse Response (FIR) Filter Structure” on page 10-20
• “Direct Form FIR Transposed Filter Structure” on page 10-21
• “Lattice Autoregressive Moving Average (ARMA) Filter Structure” on page
10-28
• “State-Space Filter Structure” on page 10-29
Starting from any quantized filter, you can convert to one of the following
representation:
• Direct form I
• Direct form II
• Direct form I transposed
• Direct form II transposed
• State space
• Lattice ARMA
Additionally, FDATool lets you do the following conversions:
• Minimum phase FIR filter to Lattice MA minimum phase
• Maximum phase FIR filter to Lattice MA maximum phase
• Allpass filters to Lattice allpass
9-12
Choosing Your Quantized Filter Structure
Refer to “FilterStructure” on page 10-12 for details about each of these
structures.
When you convert the filter structure, you can use the Use second-order
sections option to implement the structure as second-order sections. In
addition, you can specify the type of scaling to be applied to the second-order
sections during the conversion — None (the default setting), L-2 for L2 norm, or
L-infinity for L∞norm. For example, to use L∞ norm scaling, select
L-infinity from the Scale list.
FDATool optimizes the order of the second-order sections according to the
scaling option you choose. To optimize the order of the sections, FDATool calls
tf2sos with the optional 'down' argument when you select L-2 norm scaling,
and with argument 'up' for L-infinity scaling.
To Change the Structure of a Quantized Filter
To change the structure of a quantized filter in FDATool, follow these steps:
1 Import or design a reference filter in FDATool. For details about importing
filters in FDATool, refer to “Importing the Filter Design” in Signal
Processing Toolbox User’s Guide. To learn more about using FDATool to
design filters, refer to “Filter Design and Analysis Graphical User Interface”
in Signal Processing Toolbox User’s Guide.
2 Click Convert structure.
You see the Convert dialog open as shown here.
9-13
9
Quantization Tool Overview
3 Select the structure to convert to under Convert to.
4 To use second-order sections, select Use second-order sections.
Selecting Use second-order sections enables the Scale option, with None as
the default setting. Use the list to change the scaling to one of:
a None
b L-2
c
L-infinity
You can find details about the scaling options in the books listed in “Selected
Bibliography” on page 1-28 in this guide, and in the bibliography included in
the Signal Processing Toolbox User’s Guide.
5 Click OK to convert your filter to the selected structure and close the dialog.
Click Apply to convert to the new structure without closing the dialog.
To View Schematics of Filter Structures in the Toolbox
Often it helps to see the structure of a filter. From the Set Quantization
Parameters dialog in FDATool, you can open a demonstration program that
provides Simulink models of each filter structure included in the toolbox. To
9-14
Choosing Your Quantized Filter Structure
open the demonstration, click Show possible filter structures. Once the demo
Quantized Filter Construction opens, select the filter structure to view from
the Select filter structure list.
9-15
9
Quantization Tool Overview
Scaling Transfer Function Coefficients
All filters in FDATool are in transfer function form. To mitigate the effects of
quantization on the performance of your filter, you can scale the transfer
function coefficients. After you import or design a filter in FDATool (to create
your reference filter), you can scale the filter transfer function coefficients not
to exceed ±1. Scaling the coefficients prevents overflow and underflow
conditions from occurring during quantization.
A few things to note about using scaling:
• When you choose to scale your transfer function coefficients, FDATool does
two things:
- It scales the coefficients as directed.
- It changes the filter gain to keep the filter magnitude response the same
after scaling. If FDATool did not change the gain, the response of the filter
to a given input would change when you scaled the coefficients.
• If you choose to remove the scaling factors, FDATool restores the transfer
function coefficients to their values before scaling. FDATool does not remove
the filter gain it added when you scaled the coefficients. So the resulting
filter may demonstrate changed magnitude response after you remove the
scale factors.
To Scale Transfer Function Coefficients
To scale the transfer function coefficients of a filter in FDATool, follow these
steps:
1 Design a filter, or import a filter into FDATool. This is your reference filter.
Under Current Filter Information, the characteristics of your filter are
structure, source, order, and whether the filter is stable.
2 Click Set Quantization Parameters.
The bottom half of the FDATool window (the quantization region) shows the
options for quantizing a filter, including options for scaling filter transfer
function coefficients and setting the property values for the quantization
properties of the filter.
9-16
Scaling Transfer Function Coefficients
3 Select Turn quantization on to quantize the filter in Current Filter
Information.
You can review the transfer function coefficients for your filter. Select View
Filter Coefficients from the Analysis menu. The analysis area changes to
list the coefficients for the reference and quantized filters. Scroll through the
list to review the coefficients and to check for coefficient overflow or
underflow that can occur during quantization.
Notice that the left column in the analysis area contains symbols. They
indicate whether the quantized coefficient over- or underflowed during
quantization. A minus sign signals that the coefficient on that line
overflowed toward positive infinity. A plus sign indicates an overflow toward
negative infinity. Coefficients marked with zero had reference values that
underflowed to zero.
4 Click Scale transfer-fnc coeffs <=1.
5 Review the scaled coefficients to see that no overflow warning appears at the
end of the list of coefficients
Warning: 1 overflow in coefficients.
and no plus, zero, or minus symbols appear in the left column.
Once you have scaled a filter, you cannot remove the scale factors. You must
recreate the filter from the beginning by redesigning or reimporting the filter.
9-17
9
Quantization Tool Overview
Scaling Inputs and Outputs of Quantized Filters
Although clicking Scale transfer-fcn coeffs <=1 scales both the filter
coefficients and the input/output values for your quantized filter, you can set
input/output scaling without using the Scale transfer-fcn coeffs <=1 option.
For any filter structure, each filter section has two scale values associated with
it, an input and an output. When you click Show possible filter structures...
to look at the filter structures provided by FDATool, you see that each structure
includes two scale values, s(1) and s(2). If the filter has multiple sections, the
number of scale values is (number of sections +1). For example, a filter with
three sections has four scale values: s(1), s(2), s(3), and s(4), because the
output scale value for each section is the input value to the next section. So the
number of scale values you need for your filter depends on the filter structure.
You enter input/output scale values in three ways:
• Enter a scalar. FDATool uses the scalar for the input scale value s(1) in the
structure.
• Enter a vector of scale values. The vector must be of length (number of
sections +1), where each entry is a real number.
• Enter a variable name that represents a vector in your MATLAB workspace.
The length of the vector must be (number of sections +1).
To Enter Scale Values for Quantized Filters
Scale values apply to quantized filters. To specify the scale values for the
current quantized filter in FDATool, follow these steps:
1 Import a quantized filter or design a quantized filter in FDATool.
2 Click Set Quantization Parameters.
3 Check the number of sections in your filter.
The number of scale values you need for your filter depends on the number
of sections used in the filter design. For example, a filter with four sections
requires you to enter either one scale value or 5 (the number of sections +1).
9-18
Scaling Inputs and Outputs of Quantized Filters
4 Enter one of the following into Input/output scaling:
a A scalar. FDATool uses the scalar for the input scale value in the filter.
b A vector of scale values. The vector must be of length
(number of sections +1), where each entry is a real number.
c
A variable name that represents a vector in your MATLAB workspace.
The length of the vector in the workspace must be
(number of sections +1).
5 Click Apply.
9-19
9
Quantization Tool Overview
Importing and Exporting Quantized Filters
When you import a quantized filter into FDATool, or export a quantized filter
from FDATool to your workspace, the import and export functions use objects
and you specify the filter as a variable. This contrasts with importing and
exporting nonquantized filters, where you select the filter structure and enter
the filter numerator and denominator for the filter transfer function.
You have the option of exporting quantized filters to your MATLAB workspace,
exporting them to text files, or exporting them to MAT-files.
This section includes:
• “To Import Quantized Filters”
• “To Export Quantized Filters”
For general information about importing and exporting filters in FDATool,
refer to “Filter Design and Analysis Tool” in your Signal Processing Toolbox
User’s Guide.
FDATool imports quantized filters having the following structures:
• Direct form I
• Direct form II
• Direct form I transposed
• Direct form II transposed
• Direct form symmetric FIR
• Direct form antisymmetric FIR
• Lattice allpass
• Lattice AR
• Lattice MA minimum phase
• Lattice MA maximum phase
• Lattice ARMA
• Lattice coupled-allpass
• Lattice coupled-allpass power complementary
• State-space
9-20
Importing and Exporting Quantized Filters
To Import Quantized Filters
After you design or open a quantized filter in your MATLAB workspace,
FDATool lets you import the filter for analysis. Follow these steps to import
your filter in to FDATool:
1 Open FDATool.
2 Select Filter->Import Filter from the menu bar.
In the lower region of FDATool, the Design Filter tab becomes Import
Filter, and options appear for importing quantized filters, as shown.
3 From the Filter Structure list, select Quantized filter (Qfilt object).
The options for importing filters change to include:
- Quantized filter — Enter the variable name for the quantized filter in
your workspace. You can also enter qfilt to direct FDATool to construct
a quantized filter. When you enter qfilt, FDATool creates a quantized
filter according to the qfilt syntax you use.
- Frequency units — select the frequency units from the Units list, and
specify the sampling frequency value in Fs. Your sampling frequency must
correspond to the units you select. For example, when you select
Normalized (0 to 1), Fs should be one.
9-21
9
Quantization Tool Overview
4 Click Import to import or construct the filter.
FDATool checks your workspace for the specified filter. It imports the filter
if it finds it, displaying the magnitude response for the filter in the analysis
area. If you entered the quantized filter constructor in Quantized filter,
FDATool creates the filter and displays the filter magnitude response.
To Export Quantized Filters
To save your filter design, FDATool lets you export the quantized filter to your
MATLAB workspace (or you can save the current session in FDATool). When
you choose to save the quantized filter by exporting it, you select one of these
options:
• Export to your MATLAB workspace
• Export to a text file
• Export to a MAT-file
Exporting to Your Workspace
To save your quantized filter to your workspace, follow these steps:
1 Select Export from the File menu. The Export dialog opens.
2 Select Workspace under Export to.
9-22
Importing and Exporting Quantized Filters
3 Under Variable Names, assign a variable name for the quantized filter.
To overwrite variables that you have in your workspace, select Overwrite
existing variables. Clear the check box to ensure that FDATool does not
overwrite existing variables when you export the filter.
4 Click OK to export the filter and close the dialog. Click Apply to export the
filter without closing the Export dialog. Clicking Apply lets you export your
quantized filter to more than one name without leaving the Export dialog.
If you try to export the filter to a variable name that exists in your
workspace, and you did not select Overwrite existing variables, FDATool
stops the export operation and returns a warning that the variable you
specified as the quantized filter name already exists in the workspace. To
continue to export the filter to the existing variable, click OK to dismiss the
warning dialog, select the Overwrite existing variables check box and click
OK or Apply.
Getting Filter Coefficients after Exporting
To extract the filter coefficients from your quantized filter after you export the
quantized filter to MATLAB, use the celldisp function in MATLAB. For
example, create a quantized filter in FDATool and export the filter as Hq. To
extract the filter coefficients for Hq, use
celldisp(Hq.referencecoefficients)
which returns the cell array containing the filter reference coefficients, or
celldisp(Hq.quantizedcoefficients)
to return the quantized coefficients.
Exporting as a Text File
To save your quantized filter as a text file, follow these steps:
1 Select Export from the File menu.
2 Select Text-file under Export to.
9-23
9
Quantization Tool Overview
3 Click OK to export the filter and close the dialog. Click Apply to export the
filter without closing the Export dialog. Clicking Apply lets you export your
quantized filter to more than one name without leaving the Export dialog.
The Export Filter Coefficients to Text-file dialog appears. This is the
standard Microsoft Windows save file dialog.
4 Choose or enter a directory and filename for the text file and click OK.
FDATool exports your quantized filter as a text file with the name you
provided, and the MATLAB editor opens, displaying the file for editing.
Exporting as a MAT-File
To save your quantized filter as a MAT-file, follow these steps:
1 Select Export from the File menu.
2 Select MAT-file under Export to.
3 Assign a variable name for the filter.
4 Click OK to export the filter and close the dialog. Click Apply to export the
filter without closing the Export dialog. Clicking Apply lets you export your
quantized filter to more than one name without leaving the Export dialog.
The Export Filter Coefficients to MAT-file dialog appears. This is the
standard Microsoft Windows save file dialog.
5 Choose or enter a directory and filename for the text file and click OK.
FDATool exports your quantized filter as a MAT-file with the specified
name.
9-24
10
Property Reference
A Quick Guide to Quantizer Properties . . . . . . . . 10-2
Quantizer Properties Reference . . . . . . . . . . . 10-3
A Quick Guide to Quantized Filter Properties . . . . 10-10
Quantized Filter Properties Reference . . . . . . . 10-11
A Quick Guide to Quantized FFT Properties
. . . . 10-46
Quantized FFT Properties Reference . . . . . . . . 10-47
10
Property Reference
A Quick Guide to Quantizer Properties
The following table summarizes the quantizer properties and provides a brief
description of each. A table providing a full description of each property follows
in the next section.
Table 10-1: Quick Guide to Quantizer Properties
10-2
Property
Brief Description of What the Property Specifies
Format
Quantization format
Max
Maximum value encountered when the quantizer quantizes data
Min
Minimum value encountered when the quantizer quantizes data
Mode
Type of quantized arithmetic
NOperations
Number of quantization operations performed by a quantizer
NOverflows
Number of overflows encountered when the quantizer quantizes
data
NUnderflows
Number of underflows encountered when the quantizer quantizes
data
OverflowMode
Handling of arithmetic overflows
RoundMode
Rounding method used in quantization
Quantizer Properties Reference
Quantizer Properties Reference
To quantize data using quantize, you need to specify quantization parameters
in a quantizer. When you create a quantizer, you are creating a MATLAB
object. You specify the quantization parameters as values assigned to the
quantizer properties. With these property values, you specify the quantizer:
• Data format
• Arithmetic method
• Rounding method
• Overflow method
For a quick reference to properties, see Table 10-1, Quick Guide to Quantizer
Properties, on page 10-2. Details of all of the properties associated with
quantizers are described in the following sections in alphabetical order.
Format
You can set the data format of a quantizer according to its Format property
value. The interpretation of this property value depends on the value of the
Mode property.
For example, whether you specify the Mode property with fixed- or
floating-point arithmetic affects the interpretation of the data format property.
For some Mode property values, the data format property is read-only.
The following table shows you how to interpret the values for the Format
property value when you specify it, or how it is specified in read-only cases.
10-3
10
Property Reference
Table 10-2: Interpreting Format Property for Different Arithmetic Types (Mode Property Values)
Filter Arithmetic
Mode Property
Value
Interpreting the Format Property Values
Fixed-point
'fixed' or
'ufixed'
You specify the Format property value as a vector. The
number of bits for the quantizer word length is the first
entry of this vector, and the number of bits for the
quantizer fraction length is the second entry.
The word length can range from 2 to 53. The fraction
length can range from 0 to one less than the word length.
Floating-point
'float'
You specify the Format property value as a vector. The
number of bits you want for the quantizer word length is
the first entry of this vector, and the number of bits you
want for the quantizer exponent length is the second
entry.
The word length can range from 2 to 64. The exponent
length can range from 0 to 11.
Floating-point
'double'
The Format property value is specified automatically (is
read-only) when you set the Mode property to 'double'.
The value is [64 11], specifying the word length and
exponent length, respectively.
Floating-point
'single'
The Format property value is specified automatically (is
read-only) when you set the Mode property to 'single'.
The value is [32 8], specifying the word length and
exponent length, respectively.
Default value: 'fixed'
The Format property for quantizers affects the following quantized filter and
quantized FFT data format properties:
• The CoefficientFormat property
• The InputFormat property
• The MultiplicandFormat property
10-4
Quantizer Properties Reference
• The OutputFormat property
• The ProductFormat property
• The SumFormat property
Set each of these data format properties using a quantizer.
Max
The Max property is read-only. The value of the Max property is the maximum
value data has before a quantizer is applied to it, that is, before quantization
using quantize. This value accumulates if you use the same quantizer to
quantize several data sets. You can reset the value using reset.
Default value: reset
Min
The Min property is read-only. The value of the Min property is the minimum
value data has before a quantizer is applied to it, that is, before quantization
using quantize. This value accumulates if you use the same quantizer to
quantize several data sets. You can reset the value using reset.
Default value: reset
Mode
You specify Mode property values as one of the following strings to indicate the
type of arithmetic used in filtering and quantization.
Mode Property Setting
Description
'fixed'
Signed fixed-point calculations
'float'
User-specified floating-point calculations
'double'
Floating-point calculations using
double-precision
'single'
Floating-point calculations using
single-precision
'ufixed'
Unsigned fixed-point calculations
10-5
10
Property Reference
Default value: 'fixed'
Remarks: When you set the Mode property value to 'double' or 'single' the
Format property value becomes read-only.
The Mode property for quantizers affects the following quantized filter and
quantized FFT data format properties:
• The CoefficientFormat property
• The InputFormat property
• The MultiplicandFormat property
• The OutputFormat property
• The ProductFormat property
• The SumFormat property
Set each of these data format properties using a quantizer.
NOperations
The NOperations property is read-only. The value of the NOperations property
is the number of quantization operations that occurred during quantization
when you use a quantizer, quantized filter, or quantized FFT. This value
accumulates when you use the same quantizer, quantized filter, or quantized
FFT to process several data sets. You reset the value using reset.
Default value: 0
NOverflows
The NOverflows property is read-only. The value of the NOverflows property is
the number of overflows that occur during quantization using quantize. This
value accumulates if you use the same quantizer to quantize several data sets.
You can reset the value using reset.
Default value: 0
NUnderflows
The NUnderflows property is read-only. The value of the NUnderflows property
is the number of underflows that occur during quantization using quantize.
10-6
Quantizer Properties Reference
This value accumulates when you use the same quantizer to quantize several
data sets. You can reset the value using reset.
Default value: 0
OverflowMode
The OverflowMode property values are specified as one of the following two
strings indicating how overflows in fixed-point arithmetic are handled:
• 'saturate' — saturate overflows.
When the values of data to be quantized lie outside of the range of the largest
and smallest representable numbers (as specified by the data format
properties), these values are quantized to the value of either the largest or
smallest representable value, depending on which is closest.
• 'wrap' — wrap all overflows to the range of representable values.
When the values of data to be quantized lie outside of the range of the largest
and smallest representable numbers (as specified by the data format
properties), these values are wrapped back into that range using modular
arithmetic relative to the smallest representable number.
Default value: 'saturate'
Note Numbers in floating-point filters that extend beyond the dynamic
range overflow to ±inf.
The OverflowMode property value is set to 'saturate' and becomes a read-only
property when you set the value of the Mode property to either 'float',
'double', or 'single'.
The OverflowMode property for quantizers affects the following quantized filter
and quantized FFT data format properties:
• The CoefficientFormat property
• The InputFormat property
• The MultiplicandFormat property
• The OutputFormat property
10-7
10
Property Reference
• The ProductFormat property
• The SumFormat property
Set each of these data format properties using a quantizer.
RoundMode
The RoundMode property values specify the rounding method used for
quantizing numerical values. Specify the RoundMode property values as one of
the following five strings.
RoundMode String
Description of Rounding Algorithm
'ceil'
Round up to the next allowable quantized value.
'convergent'
Round to the nearest allowable quantized value.
Numbers that are exactly halfway between the
two nearest allowable quantized values are
rounded up only if the least significant bit (after
rounding) would be set to 1.
'fix'
Round negative numbers up and positive
numbers down to the next allowable quantized
value.
'floor'
Round down to the next allowable quantized
value.
'round'
Round to the nearest allowable quantized value.
Numbers that are halfway between the two
nearest allowable quantized values are rounded
up.
Default value: 'floor'
Remarks: The RoundMode property for quantizers affects the following
quantized filter and quantized FFT data format properties:
• The CoefficientFormat property
• The InputFormat property
• The MultiplicandFormat property
10-8
Quantizer Properties Reference
• The OutputFormat property
• The ProductFormat property
• The SumFormat property
Use a quantizer to set each of these data format properties.
10-9
10
Property Reference
A Quick Guide to Quantized Filter Properties
The following table summarizes the quantized filter properties and provides a
brief description of each. A table providing a full description of each property
follows in the next section.
Table 10-3: Quick Guide to Quantized Filter Properties
10-10
Property
Brief Description of What the Property Specifies
CoefficientFormat
Quantization format for filter coefficients
FilterStructure
Filter structure
InputFormat
Quantization format applied to inputs during filtering
NumberOfSections
Number of cascaded sections in the filter
MultiplicandFormat
Quantization format for inputs that are multiplied by
coefficients in filtering operations
OutputFormat
Quantization format applied to outputs during filtering
ProductFormat
Quantization format for results of multiplication in filtering
QuantizedCoefficients
Filter coefficients after quantization
ReferenceCoefficients
Filter coefficients before quantization
ScaleValues
Scaling for the quantized filter
StatesPerSection
Number of states (delays) in each section of the filter
SumFormat
Quantization format for results of addition in filtering
Quantized Filter Properties Reference
Quantized Filter Properties Reference
When you create a quantized filter, you are creating a MATLAB object. The
quantized filter object you create has many properties to which you assign
values. You use these property values to assign the characteristics of the
quantized filters you create, including:
• The filter structure
• The double-precision coefficients that specify the original reference filter
(before quantization)
• The data formats used in quantization and filtering operations
You specify the ReferenceCoefficients property value as a cell array. For
more information, see “Using Cell Arrays” on page 4-13.
For a quick reference to properties, see Table 10-1, Quick Guide to Quantizer
Properties. Details of all of the properties associated with quantized filters are
described in the following sections in alphabetical order.
CoefficientFormat
The CoefficientFormat property values specify how filter coefficients are
quantized. You specify these values with a quantizer. You set them according
to the quantizer property values:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
The value you set for this property is used to calculate the
QuantizedCoefficients property values.
Default value: quantizer('fixed','round','saturate',[16,15])
Note Coefficient overflows that occur due to quantization are not corrected
automatically. You can use normalize with several filter structures to account
for coefficient overflows.
10-11
10
Property Reference
FilterStructure
The FilterStructure property values are specified as one of the following
strings indicating the quantized filter architecture:
'Default value: 'df2t'
FilterStructure
Property Name
Filter Description
'antisymmetricfir'
Antisymmetric finite impulse response (FIR).
Even and odd forms.
'df1'
Direct form I.
'df1t'
Direct form I transposed.
'df2'
Direct form II.
'df2t'
Direct form II transposed. Default filter structure.
'fir'
Direct form FIR .
'firt'
Direct form FIR transposed.
'latcallpass'
Lattice allpass.
'latticeca'
Lattice coupled-allpass.
'latticecapc'
Lattice coupled-allpass power-complementary.
'latticear'
Lattice autoregressive (AR).
'latticema'
Lattice moving average (MA) minimum phase.
'latcmax'
Lattice moving average (MA) maximum phase.
'latticearma'
Lattice ARMA.
'statespace'
Single-input/single-output state-space.
'symmetricfir'
Symmetric FIR. Even and odd forms.
Remarks: The syntax for entering values for the ReferenceCoefficients
property is constrained by the FilterStructure property value. See Table
10-12
Quantized Filter Properties Reference
10-4: Syntax for Assigning Reference Filter Coefficients (Single Section) on
page 10-35, for information on how to enter these coefficients for each filter
architecture.
Quantized Filter Structures
You can choose among several different filter structures when you create a
quantized filter. You can also specify filters with single or multiple cascaded
sections of the same type. Because quantization is a nonlinear process,
different filter structures produce different results.
You specify the filter structure by assigning a value to the FilterStructure
property. See the function reference listings for qfilt and set for information
on setting property values.
The FilterStructure property value constrains the syntax you can use for
specifying the filter reference coefficients. For details on the syntax to use for
specifying a filter with either a single section, or multiple (L) cascaded sections,
see Table 10-4, Syntax for Assigning Reference Filter Coefficients (Single
Section), and Table 10-5, Syntax for Assigning Reference Filter Coefficients (L
Sections).
The figures in the following subsections of this chapter serve as a visual aid to
help you determine how to enter the reference filter coefficients for each filter
structure. Each subsection contains a simple example for constructing a filter
of a given structure.
For clarity, the scale factors appear outside the filter structure, labeled s(1)
and s(2) for the input and output factors.
10-13
10
Property Reference
Direct Form Antisymmetric FIR Filter Structure (Odd Order)
The following figure depicts an direct form antisymmetric FIR filter structure
that directly realizes a fifth-order antisymmetric FIR filter. The filter
coefficients are labeled b(i), i = 1, ..., 6, and the initial and final state values in
filtering are labeled z(i).
antisymmetricfir
(Antisymmetric FIR)
Even number of coefficients, length(b) = 6.
b(i) == − b(end − i + 1)
b(1)
1
x
s(2)
1
y
s(1)
1
z(1)
z
1
z(5)
z
b(2)
1
z(2)
z
1
z(4)
z
b(3)
1
z
z(3)
~2
Use the string 'antisymmetricfir' for the value of the FilterStructure
property to design a quantized filter with this structure.
Example — Specifying an Odd-Order Direct Form Antisymmetric FIR Filter Structure.
Specify a fifth-order direct form antisymmetric FIR filter structure for a
quantized filter Hq with the following code.
10-14
Quantized Filter Properties Reference
b = [-0.008 0.06 -0.44 0.44 -0.06 0.008];
Hq = qfilt('antisymmetricfir',{b});
Antisymmetric FIR Filter Structure (Even Order)
The following figure depicts a direct form antisymmetric FIR filter structure
that directly realizes a fourth-order antisymmetric FIR filter. The filter
coefficients are labeled b(i), i = 1, ..., 5, and the states (used for initial and final
state values in filtering) are labeled z(i).
antisymmetricfir
(Antisymmetric FIR)
Odd number of coefficients, length(b) = 5.
Note that antisymmetry is defined as
b(i) == −b(end − i + 1)
so that the middle coefficient is zero for odd length
b((end+1)/2) = 0
b(1)
1
x
s(2)
1
y
s(1)
1
z(1)
z
1
z(4)
z
b(2)
1
z(2)
z
1
z(3)
z
~2
Use the string 'antisymmetricfir' to specify the value of the
FilterStructure property for a quantized filter with this structure.
10-15
10
Property Reference
Example — Specifying an Even-Order Direct Form Antisymmetric FIR Filter Structure. You
can specify a fourth-order direct form antisymmetric FIR filter structure for a
quantized filter Hq with the following code.
b = [-0.01 0.1 0.0 -0.1 0.01];
Hq = qfilt('antisymmetricfir',{b});
Direct Form I Filter Structure
The following figure depicts a direct form I filter structure that directly realizes
a transfer function with a second-order numerator and denominator. The
numerator coefficients are labeled b(i), the denominator coefficients are labeled
a(i), i = 1, 2, 3, and the states (used for initial and final state values in filtering)
are labeled z(i).
df1
(Direct Form I)
b(1)
1
x
1/a(1)
s(1)
s(2)
1
z
z(1)
b(2)
1
z
1
z
z(3)
1
z
z(4)
1
y
−a(2)
z(2)
b(3)
−a(3)
~2
Use the string 'df1' to specify the value of the FilterStructure property for
a quantized filter with this structure.
Example — Specifying a Direct Form I Filter Structure. You can specify a second-order
direct form I structure for a quantized filter Hq with the following code.
b = [0.3 0.6 0.3];
a = [1 0 0.2];
Hq = qfilt('df1',{b,a});
10-16
Quantized Filter Properties Reference
Direct Form I Transposed Filter Structure
The following figure depicts a direct form I transposed filter structure that
directly realizes a transfer function with a second-order numerator and
denominator. The numerator coefficients are labeled b(i), the denominator
coefficients are labeled a(i), i = 1, 2, 3, and the states (used for initial and final
state values in filtering) are labeled z(i).
df1t
(Transposed Direct Form I)
1/a(1)
1
x
b(1)
s(1)
s(2)
1
z
z(1)
−a(2)
1
z
1
z
z(3)
1
z
z(4)
1
y
b(2)
z(2)
−a(3)
b(3)
~2
Use the string 'df1t' to specify the value of the FilterStructure property for
a quantized filter with this structure.
Example — Specifying a Direct Form I Transposed Filter Structure. You can specify a
second-order direct form I transposed filter structure for a quantized filter Hq
with the following code.
b = [0.3 0.6 0.3];
a = [1 0 0.2];
Hq = qfilt('df1t',{b,a});
10-17
10
Property Reference
Direct Form II Filter Structure
The following figure depicts a direct form II filter structure that directly
realizes a transfer function with a second-order numerator and denominator.
The numerator coefficients are labeled b(i), the denominator coefficients are
labeled a(i), i = 1, 2, 3, and the states (used for initial and final state values in
filtering) are labeled z(i).
df2
(Direct Form II)
1/a(1)
1
x
b(1)
s(1)
s(2)
1
z
1
y
z(1)
−a(2)
b(2)
1
z
−a(3)
z(2)
b(3)
~2
Use the string 'df2' to specify the value of the FilterStructure property for
a quantized filter with this structure.
Example — Specifying a Direct Form II Filter Structure. You can specify a second-order
direct form II filter structure for a quantized filter Hq with the following code.
b = [0.3 0.6 0.3];
a = [1 0 0.2];
Hq = qfilt('df2',{b,a});
10-18
Quantized Filter Properties Reference
Direct Form II Transposed Filter Structure
The following figure depicts a direct form II transposed filter structure that
directly realizes a transfer function with a second-order numerator and
denominator. The numerator coefficients are labeled b(i), the denominator
coefficients are labeled a(i), i = 1, 2, 3, and the states (used for initial and final
state values in filtering) are labeled z(i).
df2t
(Transposed Direct Form II)
b(1)
1
x
1/a(1)
s(1)
s(2)
1
z
1
y
z(1)
−a(2)
b(2)
1
z
z(2)
−a(3)
b(3)
~2
Use the string 'df2t' to specify the value of the FilterStructure property for
a quantized filter with this structure.
Example — Specifying a Direct Form II Transposed Filter Structure. You can specify a
second-order direct form II transposed filter structure for a quantized filter Hq
with the following code.
b = [0.3 0.6 0.3];
a = [1 0 0.2];
Hq = qfilt('df2t',{b,a});
10-19
10
Property Reference
Direct Form Finite Impulse Response (FIR) Filter Structure
The following figure depicts a direct form finite impulse response (FIR) filter
structure that directly realizes a second-order FIR filter. The filter coefficients
are labeled b(i), i = 1, 2, 3, and the states (used for initial and final state values
in filtering) are labeled z(i).
fir
(Direct Form FIR = Tapped delay line)
b(1)
1
x
s(1)
s(2)
1
y
1
z(1)
z
b(2)
1
z(2)
z
b(3)
~2
Use the string 'fir' to specify the value of the FilterStructure property for
a quantized filter with this structure.
Example — Specifying a Direct Form FIR Filter Structure. You can specify a
second-order direct form FIR filter structure for a quantized filter Hq with the
following code.
b = [0.05 0.9 0.05];
Hq = qfilt('fir',{b});
10-20
Quantized Filter Properties Reference
Direct Form FIR Transposed Filter Structure
The following figure depicts a direct form finite impulse response (FIR)
transposed filter structure that directly realizes a second-order FIR filter. The
filter coefficients are labeled b(i), i = 1, 2, 3, and the states (used for initial and
final state values in filtering) are labeled z(i).
firt
(Transposed Direct Form FIR)
b(1)
1
x
s(1)
s(2)
1
z
z(1)
1
z
z(2)
1
y
b(2)
b(3)
~2
Use the string 'firt' to specify the value of the FilterStructure property for
a quantized filter with this structure.
Example — Specifying a Direct Forn FIR Transposed Filter Structure. You can specify a
second-order direct form FIR transposed filter structure for a quantized filter
Hq with the following code.
b = [0.05 0.9 0.05];
Hq = qfilt('firt',{b});
10-21
10
Property Reference
Lattice Allpass Filter Structure
The following figure depicts a lattice allpass filter structure. The pictured
structure directly realizes third-order lattice allpass filters. The filter reflection
coefficients are labeled k1(i), i = 1, 2, 3. The states (used for initial and final
state values in filtering) are labeled z(i).
latcallpass
(Lattice AR All−Pass)
1
x
1
y
s(1)
s(2)
k(3)
k(2)
k(1)
conj(k(3))
conj(k(2))
conj(k(1))
1
z
z(3)
1
z
z(2)
1
z
z(1)
~2
Use the string 'latcallpass' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying a Lattice Allpass Filter Structure. You can specify a third-order
lattice allpass filter structure for a quantized filter Hq with the following code.
k = [.66 .7 .44];
Hq = qfilt('latcallpass',{k});
10-22
Quantized Filter Properties Reference
Lattice Moving Average Maximum Phase Filter Structure
The following figure depicts a lattice moving average maximum phase filter
structure that directly realizes a third-order lattice moving average (MA) filter
with maximum phase. The filter reflection coefficients are labeled k1(i), i = 1,
2, 3. The states (used for initial and final state values in filtering) are labeled
z(i).
latcmax
(Lattice MA Max phase)
1
x
s(1)
1
z
z(1)
k(3)
k(2)
k(1)
conj(k(3))
conj(k(2))
conj(k(1))
1
z
z(2)
1
z
z(3)
s(2)
1
y
~2
Use the string 'latcmax' to specify the value of the FilterStructure property
for a quantized filter with this structure.
Example — Specifying a Lattice Moving Average Maximum Phase Filter Structure. You can
specify a fourth-order lattice MA maximum phase filter structure for a
quantized filter Hq with the following code.
k = [.66 .7 .44 .33];
Hq = qfilt('latcmax',{k});
10-23
10
Property Reference
Lattice Coupled-Allpass Filter Structure
The following figure depicts a lattice coupled-allpass filter structure. The filter
is composed of two third-order allpass lattice filters. The filter reflection
coefficients for the first filter are labeled k1(i), i = 1, 2, 3. The filter reflection
coefficients for the second filter are labeled k2(i), i = 1, 2, 3. The unity gain
complex coupling coefficient is beta. The states (used for initial and final state
values in filtering) are labeled z(i).
latticeca
(Coupled Allpass Lattice)
H1(z)
k1(3)
k1(2)
k1(1)
conj(k1(3))
conj(k1(2))
conj(k1(1))
1
z
z(3)
1
z
z(2)
0.5
1
x
1
z
z(1)
~2
conj(beta)
s(1)
s(2)
1
y
H2(z)
k2(3)
k2(2)
k2(1)
conj(k2(3))
conj(k2(2))
conj(k2(1))
1
z
z(6)
1
z
z(5)
1
z
z(4)
~1
beta
~3
Use the string 'latticeca' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying a Lattice Coupled-Allpass Filter Structure. You can specify a
third-order lattice coupled allpass filter structure for a quantized filter Hq with
the following code.
k1 =
k2 =
beta
Hq =
10-24
[0.9511 + 0.3088i; 0.7511 + 0.1158i]
0.7502 - 0.1218i
= 0.1385 + 0.9904i
qfilt('latticeca',{k1,k2,beta});
Quantized Filter Properties Reference
Lattice Coupled-Allpass Power Complementary Filter Structure
The following figure depicts a lattice coupled-allpass power complementary
filter structure. The filter is composed of two third-order allpass lattice filters.
The filter reflection coefficients for the first filter are labeled k1(i), i = 1, 2, 3.
The filter reflection coefficients for the second filter are labeled k2(i), i = 1, 2, 3.
The unity gain complex coupling coefficient is beta. The states used for initial
and final state values in filtering are labeled z(i). The resulting filter transfer
function is the power-complementary transfer function of the coupled allpass
lattice filter (formed from the same coefficients).
latticecapc
(Coupled Allpass Lattice, Power Complementary output)
H1(z)
k1(3)
k1(2)
k1(1)
conj(k1(3))
conj(k1(2))
conj(k1(1))
1
z
z(3)
1
z
z(2)
j/2
1
x
1
z
z(1)
~2
conj(beta)
s(1)
s(2)
1
y
H2(z)
k2(3)
k2(2)
k2(1)
conj(k2(3))
conj(k2(2))
conj(k2(1))
−1
1
z
z(6)
1
z
z(5)
1
z
z(4)
~1
beta
~3
Use the string 'latticecapc' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying a Lattice Coupled-Allpass Power Complementary Filter Structure.
Specify a third-order lattice coupled-allpass power complementary filter
structure for a quantized filter Hq with the following code.
k1 =
k2 =
beta
Hq =
[0.9511 + 0.3088i; 0.7511 + 0.1158i]
0.7502 - 0.1218i
= 0.1385 + 0.9904i
qfilt('latticecapc',{k1,k2,beta});
10-25
10
Property Reference
Lattice Autoregressive (AR) Filter Structure
The following figure depicts an lattice autoregressive (AR) filter structure that
directly realizes a third-order lattice AR filter. The filter reflection coefficients
are labeled k(i), i = 1, 2, 3, and the states (used for initial and final state values
in filtering) are labeled z(i).
latticear
(Autoregressive Lattice)
1
x
s(1)
k(3)
k(2)
k(1)
conj(k(3))
conj(k(2))
conj(k(1))
1
z
z(3)
1
z
z(2)
s(2)
1
y
1
z
z(1)
~2
Use the string 'latticear' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying an Lattice AR Filter Structure. You can specify a third-order
lattice AR filter structure for a quantized filter Hq with the following code.
k = [.66 .7 .44];
Hq = qfilt('latticear',{k});
10-26
Quantized Filter Properties Reference
Lattice Moving Average (MA) Filter Structure
The following figure depicts a lattice moving average (MA) filter structure that
directly realizes a third-order lattice MA filter. The filter reflection coefficients
are labeled k(i), i = 1, 2, 3, and the states (used for initial and final state values
in filtering) are labeled z(i).
latticema
(Moving Average Lattice)
1
x
s(1)
1
z
z(1)
k(3)
k(2)
k(1)
conj(k(3))
conj(k(2))
conj(k(1))
1
z
z(2)
s(2)
1
y
1
z
z(3)
~2
Use the string 'latticema' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying an Lattice MA Filter Structure. You can specify a third-order
lattice MA filter structure for a quantized filter Hq with the following code.
k = [.66 .7 .44];
Hq = qfilt('latticema',{k});
10-27
10
Property Reference
Lattice Autoregressive Moving Average (ARMA) Filter Structure
The following figure depicts an lattice autoregressive moving average (ARMA)
filter structure that directly realizes a fourth-order lattice ARMA filter. The
filter reflection coefficients are labeled k(i), i = 1, ..., 4, the ladder coefficients
are labeled v(i), i = 1, 2, 3, and the states (used for initial and final state values
in filtering) are labeled z(i).
latticearma
(ARMA Lattice)
1
x
s(1)
k(4)
k(3)
k(2)
k(1)
conj(k(4))
conj(k(3))
conj(k(2))
conj(k(1))
1
z
z(4)
1
z
z(3)
1
z
z(2)
v(3)
1
z
z(1)
v(2)
v(1)
s(2)
1
y
~2
Use the string 'latticearma' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying an Lattice ARMA Filter Structure. You can specify a fourth-order
lattice ARMA filter structure for a quantized filter Hq with the following code.
k = [.66 .7 .44 .66];
v = [1 0 0];
Hq = qfilt('latticearma',{k,v});
10-28
Quantized Filter Properties Reference
State-Space Filter Structure
State-space models with input sequence xk and output sequence yk have the
following form.
z k + 1 = Az k + Bx k
y k = Cz k + Dx k
If the states zk are vectors of length n, then the matrices A, B, C, and D are
n-by-n, n-by-1, 1-by-n, and 1-by-1 respectively.
Statespace
x(k+1) = Ax(k) + Bu(k)
y(k) = Cx(k) + Du(k)
D
DU(k)
Y(k) = CX(k)+DU(k)
s(2)
1
Output
CX(k)
U(k)
C
1
Input
U(k)
B
BU(k)
AX(k)+BU(k)
1
z
X(k)
s(1)
AX(k)
A
~2
Use the string 'statespace' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying a State-Space Filter Structure. You can specify a second-order
state-space filter structure for a quantized filter Hq with the following code.
[A,B,C,D] = butter(2,0.5);
Hq = qfilt('statespace',{A,B,C,D});
10-29
10
Property Reference
Direct Form Symmetric FIR Filter Structure (Odd Order)
The following figure depicts a direct form symmetric FIR filter structure that
directly realizes a fifth-order direct form symmetric FIR filter. The filter
coefficients are labeled b(i), i = 1, ..., 6, and the states (used for initial and final
state values in filtering) are labeled z(i).
symmetricfir
(Symmetric FIR)
Even number of coefficients, length(b) = 6.
b(i) == b(end − i + 1)
b(1)
1
x
s(2)
1
y
s(1)
1
z
z(1)
1
z
z(5)
1
z
z(4)
b(2)
1
z
z(2)
b(3)
1
z
z(3)
~2
Use the string 'symmetricfir' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying an Odd-Order Direct Form Symmetric FIR Filter Structure. You can
specify a fifth-order direct form symmetric FIR filter structure for a quantized
filter Hq with the following code.
b = [-0.008 0.06 0.44 0.44 0.06 -0.008];
Hq = qfilt('symmetricfir',{b});
10-30
Quantized Filter Properties Reference
Direct Form Symmetric FIR Filter Structure (Even Order)
The following figure depicts a direct form symmetric FIR filter structure that
directly realizes a fourth-order direct form symmetric FIR filter. The filter
coefficients are labeled b(i), i = 1, ..., 5, and the states (used for initial and final
state values in filtering) are labeled z(i).
symmetricfir
(Symmetric FIR)
Odd number of coefficients, length(b) = 5.
b(i) == b(end − i + 1)
b(1)
1
x
s(2)
1
y
s(1)
1
z
z(1)
1
z
z(4)
1
z
z(3)
b(2)
1
z
z(2)
b(3)
~2
Use the string 'symmetricfir' to specify the value of the FilterStructure
property for a quantized filter with this structure.
Example — Specifying an Even-Order Direct Form Symmetric FIR Filter Structure. You can
specify a fourth-order direct form symmetric FIR filter structure for
a quantized filter Hq with the following code.
b = [-0.01 0.1 0.8 0.1 -0.01];
Hq = qfilt('symmetricfir',{b});
10-31
10
Property Reference
InputFormat
The InputFormat property values specify how inputs are quantized during the
filtering operation. You specify these values with a quantizer. You set them
according to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[16,15])
NumberOfSections
The value of this read-only property is a scalar that specifies the number of
cascaded sections in your quantized filter. You specify the number of sections
for your filter by the way you specify the ReferenceCoefficients property
value.
Default value: 1
MultiplicandFormat
Products in quantized filters always involve two types of multiplicands:
• Inputs (data)
• Coefficients
MultiplicandFormat property values specify how inputs that are multiplied by
coefficients are quantized during the filtering operation. You specify these
values with a quantizer. You set them according to the property values of a
quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[16,15])
10-32
Quantized Filter Properties Reference
OutputFormat
The OutputFormat property values specify how outputs are quantized during
the filtering operation. You specify these values with a quantizer. You set them
according to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[16,15])
ProductFormat
The ProductFormat property values specify how the results of multiplication
are quantized during the filtering operation. You specify these values with a
quantizer. You set them according to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[32,30])
QuantizedCoefficients
The values for this read-only property are stored in a cell array containing the
quantized filter coefficients calculated from the value of the
ReferenceCoefficients property. The quantization is specified by the value of
the CoefficientFormat property.
Default value: {1 1}
Remarks: If any filter coefficient overflows occur as a result of quantization, a
warning is displayed.
The cell array for the QuantizedCoefficients property value has the same
form as that of the corresponding ReferenceCoefficients property value,
described in “Assigning Reference Filter Coefficients” on page 10-34.
10-33
10
Property Reference
ReferenceCoefficients
The ReferenceCoefficients property values are specified as a cell array that
specifies the original (unquantized) reference filter coefficients. You specify
these in double-precision, using a syntax specific to the value of the
FilterStructure property.
Default value: {1 1}
Assigning Reference Filter Coefficients
To assign the coefficients that specify the filter that serves as the reference for
your quantized filter, specify the value of the ReferenceCoefficients
property. The syntax you use to assign reference filter coefficients for your
quantized filter depends on the value you assign to the FilterStructure
property. These syntaxes are described in the following two tables. The first
table explains the syntax for the ReferenceCoefficients property value when
you want to specify one section for your filter. The next table explains how to
specify the coefficients for filters with L cascaded sections.
10-34
Quantized Filter Properties Reference
Table 10-4: Syntax for Assigning Reference Filter
Coefficients (Single Section)
FilterStructure Property
Value
Syntax for ReferenceCoefficients Property
Value
'antisymmetricfir'
{b}:
• This is a cell array containing one vector.
• b(i) = -b(n-i+1); i = 1, ..., n
• n-1 is the order of the polynomial represented
by b.
• When n is odd, the center coefficient, b((n+1)/
2) should be 0.
If you don’t supply an antisymmetric vector b, it
is converted to be antisymmetric automatically.
'df1'
{b,a}:
• This is a cell array of vectors.
• b is the vector representing the coefficients of
the transfer function numerator polynomial.
• a is the vector representing the coefficients of
the transfer function denominator polynomial.
'df1t'
{b,a}: This is a cell array of vectors.
'df2'
{b,a}: This is a cell array of vectors.
'df2t'
{b,a}: This is a cell array of vectors.
'fir'
{b}: This is a cell array containing one vector.
'firt'
{b}: This is a cell array containing one vector.
10-35
10
Property Reference
Table 10-4: Syntax for Assigning Reference Filter
Coefficients (Single Section) (Continued)
FilterStructure Property
Value
Syntax for ReferenceCoefficients Property
Value
'latticeca'
{k1,k2,beta}:
• This is a cell array of vectors.
• k1 and k2 are the vectors of reflection
coefficients for the two lattice allpass filters in
the coupled allpass structure.
• beta is the unity gain complex scalar coupling
coefficient.
'latticecapc'
{k1,k2,beta}:
• This is a cell array of vectors.
• k1 and k2 are the vectors of reflection
coefficients for the two lattice allpass filters in
the coupled allpass structure.
• beta is the unity gain complex scalar coupling
coefficient.
'latticear'
{k}:
• This is a cell array containing one vector.
• k is the vector of reflection coefficients for an
all-pole (AR) lattice filter.
'latticema'
{k}:
• This is a cell array containing one vector.
• k is the vector of reflection coefficients for an
FIR (MA) lattice filter.
10-36
Quantized Filter Properties Reference
Table 10-4: Syntax for Assigning Reference Filter
Coefficients (Single Section) (Continued)
FilterStructure Property
Value
Syntax for ReferenceCoefficients Property
Value
'latticearma'
{k,v}:
• This is a cell array of vectors.
• k is the vector of reflection coefficients for an
IIR (ARMA) lattice filter.
• v is the vector of ladder coefficients for an IIR
lattice filter.
'statespace'
{A,B,C,D}:
• This is a cell array of matrices.
• A is the n-by-n state transition matrix (n
states).
• B is the n-by-1 input to state transmission
vector.
• C is the 1-by-n state to output transmission
vector.
• D is the input to output transmission scalar.
'symmetricfir'
{b}:
• This is a cell array containing one vector.
• b(i) = b(n-i+1); i = 1, ..., n
• n-1 is the order of the polynomial represented
by b.
• If you don’t supply a symmetric vector b, it is
converted to be symmetric automatically.
You can specify quantized filters with multiple sections for all of the filter
structures.
10-37
10
Property Reference
The following table describes the syntax for entering reference coefficients to
specify a quantized filter with L second-order or arbitrary-order sections.
Table 10-5: Syntax for Assigning Reference Filter
Coefficients (L Sections)
Section
Structure
Syntax for ReferenceCoefficients Property Value
L second-order
sections
{ {b1 a1} {b2 a2} ... {bL aL} }
• This is a 1-by-L cell array of 1-by-2 cell arrays.
• bi is a 1-by-3 row vector for the numerator of the ith
section, i=1, ... , L.
• ai is a 1-by-3 row vector for the denominator of the ith
section, i=1, ... , L.
You can use tf2sos and sos2cell to covert a transfer
function directly into this format (a cell array of cells).
You can also use sos to convert quantized filters with
other topologies directly to a second-order sections form.
L sections,
each of
arbitrary order
(except FIR
filters)
{ {b1 a1} {b2 a2} ... {bL aL} }
• This is a 1-by-L cell array of 1-by-2 cell arrays.
• bi is a 1-by-nbi row vector for the numerator of the ith
section, i=1, ... , L.
• ai is a 1-by-nai row vector for the denominator of the
ith section, i=1, ... , L.
• nbi is the order of the numerator of the ith section.
• nai is the order of the denominator of the ith section.
L sections,
each of
arbitrary order
(only FIR)
{ {b1} {b2} ... {bL} }
• This is a 1-by-L cell array of one-dimensional cell
arrays.
• bi is a 1-by-nbi row vector for the numerator of the ith
FIR section, i=1, ... , L.
• nbi is the order of the ith FIR section.
10-38
Quantized Filter Properties Reference
Table 10-5: Syntax for Assigning Reference Filter
Coefficients (L Sections) (Continued)
Section
Structure
Syntax for ReferenceCoefficients Property Value
L sections of
coupled allpass
lattice filters
{{k11,k21,beta1},...,{k1L,k2L,betaL)}:
• This is a 1-by-L cell array of 1-by-3 cell arrays.
• k1i and k2i are the vectors of reflection coefficients for
the two lattice allpass filters in the ith coupled allpass
structure, i=1, ... , L.
• betai is the unity gain complex scalar coupling
coefficient in the ith coupled allpass structure, i=1, ... ,
L.
L sections of
lattice ARMA
filters
{{k1,v1},...,{kL,vL}}:
• This is a 1-by-L cell array of 1-by-2 cell arrays.
• ki is the vector of reflection coefficients for the ith IIR
lattice (ARMA) filter in the cascade, i=1, ... , L.
• vi is the vector of ladder coefficients for ith IIR lattice
(ARMA) filter in the cascade, i=1, ... , L.
10-39
10
Property Reference
Table 10-5: Syntax for Assigning Reference Filter
Coefficients (L Sections) (Continued)
Section
Structure
Syntax for ReferenceCoefficients Property Value
L sections of
lattice AR or
MA filters
{{k1},...,{kL}}:
• This is a 1-by-L cell array of one-dimensional cell
arrays.
• ki is the vector of reflection coefficients for the ith
lattice AR or MA filter in the cascade, i=1, ... , L.
L sections of
state-space
filters
{{A1,B1,C1,D1},...,{AL,BL,CL,DL}}:
• This is a cell array of 1-by4 cell arrays.
• Ai is the ni-by-ni state transition matrix (ni states) of
the ith state-space filter in the cascade, i=1, ... , L.
• Bi is the ni-by-1 input to state transmission vector of
the ith state-space filter in the cascade, i=1, ... , L.
• Ci is the 1-by-ni state to output transmission vector of
the ith state-space filter in the cascade, i=1, ... , L.
• Di is the input to output transmission scalar of the ith
state-space filter in the cascade, i=1, ... , L.
10-40
Quantized Filter Properties Reference
Conversion functions in this toolbox and in Signal Processing Toolbox let you
convert transfer functions to other filter forms and from filter forms to transfer
functions. Relevant conversion functions include the following functions.
Conversion Function
Description
ca2tf
Converts from a coupled allpass filter to a
transfer function.
cl2tf
Converts from a lattice coupled allpass filter to
a transfer function.
sos
Converts quantized filters to create
second-order sections. This is the
recommended method for converting quantized
filters to second-order sections.
tf2ca
Converts from a transfer function to a coupled
allpass filter.
tf2cl
Converts from a transfer function to a lattice
coupled allpass filter.
tf2latc
Converts from a transfer function to a lattice
filter.
tf2sos
Converts from a transfer function to a
second-order section form.
tf2ss
Converts from a transfer function to
state-space form.
tf2zp
Converts from a rational transfer function to
its factored (single section) form
(zero-pole-gain form).
zp2sos
Converts a zero-pole-gain form to a
second-order section form.
10-41
10
Property Reference
Conversion Function
(Continued)
Description
zp2ss
Conversion of zero-pole-gain form to a
state-space form.
zp2tf
Conversion of zero-pole-gain form to transfer
functions of multiple order sections.
You can specify a filter with L sections of arbitrary order by:
1 Factoring your entire transfer function with tf2zp.
2 Using zp2tf to compose the transfer function for each section from the
selected first-order factors obtained in step 1.
Note You are not required to normalize the leading coefficients of each
section’s denominator polynomial when specifying second-order sections,
though tf2sos does.
ScaleValues
The ScaleValues property values are specified as a scalar (or vector) that
introduces scaling for inputs (and the outputs from cascaded sections in the
vector case) during filtering:
• When you only have a single section in your filter:
- Specify the ScaleValues property value as a scalar if you only want to
scale the input to your filter.
- Specify the ScaleValues property as a vector of length 2 if you want to
specify scaling to the input (scaled with the first entry in the vector) and
the output (scaled with the last entry in the vector).
• When you have L cascaded sections in your filter:
- Specify the ScaleValues property value as a scalar if you only want to
scale the input to your filter.
10-42
Quantized Filter Properties Reference
- Specify the value for the ScaleValues property as a vector of length L+1 if
you want to scale the inputs to every stage in your filter, along with the
output:
-The first entry of your vector specifies the input scaling
- Each successive entry specifies the scaling at the output of the next section
- The final entry specifies the scaling for the filter output.
The interpretation of this property is described below with diagrams in
“Interpreting the ScaleValues Property”.
Default value: 1
Remarks: The value of the ScaleValues property is not quantized. Data
affected by the presence of a scaling factor in the filter is quantized according
to the appropriate data format.
When you apply normalize to a quantized filter, the value for the ScaleValues
property is changed accordingly.
It is good practice to choose values for this property that are either positive or
negative powers of two.
Interpreting the ScaleValues Property
When you specify the values of the ScaleValues property of a quantized filter,
the values are entered as a vector, the length of which is determined by the
number of cascaded sections in your filter:
• When you have only one section, the value of the Scalevalues property can
be a a scalar or a two-element vector.
• When you have L cascaded sections in your filter, the value of the
Scalevalues property can be a scalar or an L+1-element vector.
10-43
10
Property Reference
The following diagram shows how the ScaleValues property values are applied
to a quantized filter with only one section.
Application of ScaleValues
to a Single Section
ScaleValues(1)
−K−
1
Input
ScaleValues(2)
Input
Output
−K−
1
Output
Filter
The following diagram shows how the ScaleValues property values are applied
to a quantized filter with two sections.
Application of ScaleValues
to Multiple Sections
ScaleValues(1)
1
Input
−K−
ScaleValues(2)
Input
Output
−K−
Filter1
ScaleValues(3)
Input
Output
−K−
1
Output
Filter2
StatesPerSection
This read-only property value is an 1-by-L vector that specifies the number of
states (delays) in each section of a quantized filter with L cascaded sections.
Default value: 0
SumFormat
The SumFormat property values specify how the results of addition are
quantized during the filtering operation. You specify these values with a
quantizer. You set them according to the property values of a quantizer’s:
10-44
Quantized Filter Properties Reference
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[32,30])
10-45
10
Property Reference
A Quick Guide to Quantized FFT Properties
The following table summarizes the quantized FFT properties and provides a
brief description of each. These properties are implemented when you use a
quantized FFT in conjunction with an FFT or inverse FFT (IFFT)) algorithm
(fft or ifft). A table providing a full description of each property follows in the
next section.
Table 10-6: Quick Guide to Quantized FFT Properties
10-46
Property
Brief Description of What the Property Specifies
CoefficientFormat
Quantization format for FFT or IFFT coefficients (twiddle factors)
InputFormat
Quantization format applied to inputs to the FFT or IFFT algorithm
Length
Length of the quantized FFT or IFFT
NumberOfSections
Number of sections used in the quantized FFT algorithm
MultiplicandFormat
Quantization format for inputs that are multiplied by coefficients in
the FFT or IFFT algorithm
OutputFormat
Quantization format applied to outputs of the FFT or IFFT algorithm
ProductFormat
Quantization format for results of multiplication within the FFT or
IFFT algorithm
Radix
Radix value for the FFT algorithm
ScaleValues
Scaling for the inputs and stages of the FFT or IFFT algorithm
SumFormat
Quantization format for results of addition within the FFT or IFFT
algorithm
Quantized FFT Properties Reference
Quantized FFT Properties Reference
To implement an FFT or inverse FFT (IFFT) algorithm, you specify a quantized
FFT, along with its property values. When you create a quantized FFT, you are
creating a MATLAB object. You specify the FFT quantization parameters as
values assigned to the quantized FFT’s properties. With these property values,
you specify the quantized FFT’s:
• Data formats
• Length
• Radix number (2 or 4)
• Scaling values for each stage
For a quick reference to properties, see “A Quick Guide to Quantized FFT
Properties” on page 10-46. Details of all of the properties associated with
quantized FFTs are described in the following sections in alphabetical order.
CoefficientFormat
The CoefficientFormat property values specify how FFT coefficients (twiddle
factors) are quantized in the FFT algorithm. You specify these values with a
quantizer. You set them according to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','round','saturate',[16,15])
InputFormat
The InputFormat property values specify how inputs are quantized in the FFT
algorithm. You specify these values with a quantizer. You set them according
to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
10-47
10
Property Reference
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[16,15])
Length
The Length property value is a scalar integer indicating the length of the FFT.
Specify the length as a power of the radix number.
Default value: 16
NumberOfSections
The value of this read-only property is a scalar that specifies the number of
sections (stages) in your FFT algorithm. This number is computed from the
Length and the Radix property values as
log2(Length)/log2(Radix)
Default value: 4
MultiplicandFormat
Products in quantized FFTs always involve two types of multiplicands:
• Inputs (data)
• Coefficients
The MultiplicandFormat property values specify how inputs that are
multiplied by coefficients are quantized in the FFT algorithm. You specify
these values with a quantizer. You set them according to the property values
of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[16,15])
10-48
Quantized FFT Properties Reference
OutputFormat
The OutputFormat property values specify how outputs are quantized in the
FFT algorithm. You specify these values with a quantizer. You set them
according to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[16,15])
ProductFormat
The ProductFormat property values specify how the results of multiplication
are quantized in the FFT algorithm. You specify these values with a quantizer.
You set them according to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[32,30])
Radix
The Radix property indicates the form of the FFT algorithm you want to apply.
The Radix property value can be either:
• 2 (default)
•4
ScaleValues
The ScaleValues property values are specified as a scalar (or vector) that
introduces scaling for inputs (and the outputs from each FFT section in the
vector case) to the FFT algorithm:
10-49
10
Property Reference
• Specify the ScaleValues property value as a scalar if you only want to scale
the input to the FFT algorithm.
• Specify the ScaleValues property as a vector of length L if you have L
sections in your FFT, and you want to scale:
- The input to the first section (with the first entry in the vector you supply)
- The input to each subsequent section (with each corresponding entry in
the vector you supply)
Default value: 1
Remarks: The value of the ScaleValues property is not quantized. Data
affected by the presence of a scaling factor within the FFT algorithm is
quantized according to the appropriate data format.
It is good practice to choose values for this property that are positive or
negative powers of two.
SumFormat
The SumFormat property values specify how the results of addition are
quantized in the FFT algorithm. You specify these values with a quantizer. You
set them according to the property values of a quantizer’s:
• Format (except when the Mode property value is set to 'double' or 'single')
• Mode
• OverflowMode
• RoundMode
Default value: quantizer('fixed','floor','saturate',[32,30])
10-50
11
Function Reference
Filter Design Toolbox Functions . . . . . . . .
Quantized Filter Construction and Property Functions
Quantized Filter Analysis Functions . . . . . . . .
Quantized Filtering Functions . . . . . . . . . .
Second-Order Sections Conversion Functions . . . .
Quantizer Construction and Property Functions . . .
Quantizer Analysis Functions . . . . . . . . . .
Quantized FFT Construction and Property Functions .
Quantized FFT Analysis Functions . . . . . . . .
Filter Design Functions . . . . . . . . . . . . .
Filter Conversion Functions . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 11-2
. 11-2
. 11-2
. 11-3
. 11-4
. 11-4
. 11-4
. 11-5
. 11-6
. 11-6
. 11-7
Functions Operating on Quantized Filters . . . . . . 11-8
Functions Operating on Quantizers
. . . . . . . . 11-10
Functions Operating on Quantized FFTs . . . . . . 11-12
Functions for Designing Digital Filters . . . . . . . 11-14
Filter Design Toolbox Functions Listed Alphabetically 11-15
11
Function Reference
Filter Design Toolbox Functions
You use the Filter Design (FDQ) Toolbox to create, apply, and analyze
quantized filters, quantizers, and quantized Fast Fourier Transforms (FFTs).
This chapter contains detailed descriptions of all FDQ Toolbox functions
grouped by subject area, and continues with the reference entries listed
alphabetically. The following tables list the functions in the FDQ Toolbox,
separated by quantization application—quantized filter, quantizer, or
quantized FFT. In many instances, you can apply a function to more than one
application; those functions are called overloaded functions and they appear in
more than one table.
Table 11-1: Quantized Filter Construction and Property Functions
Function
Description
get
Get properties of a quantized filter.
isreal
Test if filter coefficients are real.
num2bin
Convert a number to two’s-complement binary string.
num2hex
Convert a number to hexadecimal string.
qfilt
Construct a quantized filter (Qfilt object).
qreport
Returns the listing of a quantize filter and its properties.
reset
Reset the properties of a quantized filter to their initial
values.
set
Set properties of a quantized filter.
setbits
Set the data format property values for a quantized filter.
Table 11-2: Quantized Filter Analysis Functions
11-2
Function
Description
freqz
Compute the frequency response for a quantized filter.
impz
Compute the impulse response for a quantized filter.
Table 11-2: Quantized Filter Analysis Functions (Continued)
Function
Description
isallpass
Test quantized filters to determine if they are allpass
structures.
isfir
Test quantized filters to see if they are FIR filters.
islinphase
Test quantized filters to see if they have one or more
linear phase sections.
ismaxphase
Test quantized filters to see if they are maximum phase
filters.
isminphase
Test quantized filters to see if they are minimum phase
filters.
isreal
Test qauntized filters for purely real coefficients.
issos
Test whether quantized filters are composed of
second-order sections.
isstable
Test for stability of quantized filters.
limitcycle
Detect limit cycles in a quantized filter.
nlm
Use the Noise Loading Method to estimate the frequency
response of a quantized filter.
zplane
Compute a pole-zero plot for a quantized filter.
Table 11-3: Quantized Filtering Functions
Function
Description
filter
Filter data with a quantized filter.
normalize
Normalize quantized filter coefficients.
11-3
11
Function Reference
Table 11-4:
Second-Order Sections Conversion Functions
Function
Description
cell2sos
Convert a cell array to a second-order sections matrix.
sos
Convert a quantized filter to second-order sections form,
order, and scale.
sos2cell
Convert a second-order sections matrix to a cell array.
Table 11-5: Quantizer Construction and Property Functions
Function
Description
bin2num
Convert binary string to number.
get
Return the property values for a quantizer
num2bin
Convert a number to two’s-complement binary string.
num2hex
Convert a number to hexadecimal string.
qreport
Returns the listing of a quantizer and its properties.
quantize
Apply a quantizer to a data set.
quantizer
Construct a quantizer object.
reset
Reset the properties of a quantizer to their initial
values.
set
Set and display the property values of a quantizer.
unitquantize
Set numbers between eps(q) and 1 equal to 1.
wordlength
Return the wordlength of a quantizer.
Table 11-6: Quantizer Analysis Functions
11-4
Function
Description
denormalmax
Return the largest denormalized quantized number.
denormalmin
Return the smallest denormalized quantized number.
Table 11-6: Quantizer Analysis Functions
Function
Description
eps
Return the quantized relative accuracy of a quantizer.
exponentbias
Return the exponent bias for a quantizer.
exponentlength
Return the exponent length for a quantizer.
exponentmax
Return the maximum exponent for a quantizer.
exponentmin
Return the minimum exponent for a quantizer
fractionlength
Return the fraction length for a quantizer.
noverflows
Return the number of overflows encountered while
using a quantizer on one or more data sets.
range
Return the numerical range of a quantizer.
realmax
Return the largest positive quantized number.
realmin
Return the smallest positive quantized number
nunderflows
Return the number of underflows encountered while
using a quantizer on one or more data sets.
Table 11-7: Quantized FFT Construction and Property Functions
Function
Description
fft
Apply a quantized FFT to a data set.
get
Return the property values for a quantized FFT.
ifft
Apply an inverse quantized FFT to a data set.
qfft
Construct a quantized FFT.
qreport
Returns the listing of a quantized FFT and its
properties.
quantizer
Return all the quantizers associated with a quantized
FFT.
11-5
11
Function Reference
Table 11-7: Quantized FFT Construction and Property Functions
Function
Description
reset
Reset the properties of a quantized FFT to their initial
values.
set
Set and display the property values of a quantized FFT.
setbits
Set and one or more property values of a quantized
FFT.
Table 11-8: Quantized FFT Analysis Functions
Function
Description
noverflows
Return the number of overflows resulting from the
most recent application of a quantized FFT.
range
Return the numerical range of a quantized FFT.
twiddles
Return the twiddle factors for a quantized FFT.
Table 11-9: Filter Design Functions
11-6
Function
Description
firlpnorm
Design least-pth norm optimal FIR filters.
gremez
Design optimal equiripple FIR (finite impulse
response) digital filters based on the Parks-McClellan
algorithm.
iirgrpdelay
Design least-pth norm IIR filters with given group
delay.
iirlpnorm
Design least-pth norm IIR filters.
iirlpnormc
Design constrained least-pth norm IIR filters.
Table 11-10: Filter Conversion Functions
Function
Description
ca2tf
Convert coupled allpass filters to transfer function
form.
cl2tf
Convert lattice coupled allpass filters to transfer
function form.
iirpowcomp
Compute the power complementary IIR filter.
tf2ca
Convert transfer function form to coupled allpass form.
tf2cl
Convert transfer function form to lattice coupled
allpass form.
11-7
11
Function Reference
Functions Operating on Quantized Filters
The following table lists functions that operate directly on quantized filters.
Some are overloaded and operate on other quantized objects, such as quantized
FFTs as well. Overloaded functions are marked in the table.
Functions
11-8
Functions That Operate
Directly on Quantized Filters
Overloaded Functions
convert
√
copyobj
√
√
disp
√
√
eps
√
√
filter
√
√
freqz
√
√
get
√
√
impz
√
√
isallpass
√
isfir
√
islinphase
√
ismaxphase
√
isminphase
√
isreal
√
issos
√
isstable
√
limitcycle
√
nlm
√
√
Functions Operating on Quantized Filters
Functions
Functions That Operate
Directly on Quantized Filters
Overloaded Functions
noperations
√
√
normalize
√
noverflows
√
num2bin
√
num2hex
√
optimizeunityg
ains
√
√
order
√
qfilt
√
qfilt2tf
√
range
√
√
reset
√
√
set
√
√
setbits
√
√
sos
√
zplane
√
√
To get command line help on an overloaded function FunctionName for
quantized filters, type
help qfilt/FunctionName
11-9
11
Function Reference
Functions Operating on Quantizers
The following table lists functions that operate directly on quantizers. Some
are overloaded and operate on other quantized objects, such as quantized FFTs
as well. Overloaded functions are marked in the table
Functions
11-10
Functions That Operate
Directly on Quantizers
Overloaded Functions
bin2num
√
√
copyobj
√
√
denormalmax
√
denormalmin
√
disp
√
√
eps
√
√
exponentbias
√
exponentlength
√
exponentmax
√
exponentmin
√
fractionlength
√
get
√
hex2num
√
max
√
min
√
noperations
√
√
noverflows
√
√
num2bin
√
√
√
Functions Operating on Quantizers
Functions
Functions That Operate
Directly on Quantizers
Overloaded Functions
num2hex
√
√
nunderflows
√
qreport
√
quantize
√
quantizer
√
randquant
√
range
√
realmax
√
realmin
√
reset
√
√
set
√
√
tostring
√
√
unitquantize
√
unitquantizer
√
wordlength
√
√
√
To get command line help for an overloaded function FunctionName for
quantizers, type
help quantizer/FunctionName
11-11
11
Function Reference
Functions Operating on Quantized FFTs
The following table lists functions that operate directly on quantized FFTs.
Some are overloaded and operate on other quantized objects, such as quantized
filters as well. Overloaded functions are marked in the table
Functions
11-12
Functions That Operate
Directly on Quantized FFTs
Overloaded
Functions
copyobj
√
√
disp
√
√
eps
√
√
fft
√
get
√
ifft
√
noperations
√
√
noverflows
√
√
optimizeunitygains
√
√
qfft
√
qreport
√
√
quantizer
√
√
range
√
√
reset
√
√
set
√
√
setbits
√
√
tostring
√
√
twiddles
√√
√
Functions Operating on Quantized FFTs
To get command line help on an overloaded function FunctionName for
quantized FFTs, type
help qfft/FunctionName
11-13
11
Function Reference
Functions for Designing Digital Filters
The following functions design digital FIR filters:
• firlpnorm
• gremez
The following functions design digital IIR filters:
• iirgrpdelay
• iirlpnorm
• iirlpnormc
The following functions convert the structures of digital filters:
• ca2tf
• cl2tf
• iirpowcomp
• qfilt2tf
• tf2ca
• tf2cl
To get command line help on a design or conversion function such as gremez or
quantizer, type either
• help gremez
• help objecttype/quantizer where objecttype is one of the following
strings that specify the version of help to see:
- qfilt
- qfft
- quant
11-14
Filter Design Toolbox Functions Listed Alphabetically
Filter Design Toolbox Functions Listed Alphabetically
The following reference pages list the functions included in the Filter Design
Toolbox. Each function listing provides a purpose, syntax, description,
algorithm (optional), and examples for the function.
11-15
bin2num
Purpose
11bin2num
Convert a two’s complement binary string to a number
Syntax
y = bin2num(q,b)
Description
y = bin2num(q,b) uses the properties of quantizer q to convert binary string
b to numeric array y. When b is a cell array containing binary strings, y will be
a cell array of the same dimension containing numeric arrays. The fixed-point
binary representation is two’s complement. The floating-point binary
representation is in IEEE Standard 754 style.
bin2num and num2bin are inverses of one another. Note that num2bin always
returns columnwise.
Examples
Create a quantizer and an array of numeric strings. Convert the numeric
strings to binary strings, then use bin2num to convert them back to numeric
strings.
q=quantizer([4 3]);
[a,b]=range(q);
x=(b:-eps(q):a)';
b = num2bin(q,x)
b =
0111
0110
0101
0100
0011
0010
0001
0000
1111
1110
1101
1100
1011
1010
1001
1000
11-16
bin2num
bin2num performs the inverse operation of num2bin.
y=bin2num(q,b)
y =
0.8750
0.7500
0.6250
0.5000
0.3750
0.2500
0.1250
0
-0.1250
-0.2500
-0.3750
-0.5000
-0.6250
-0.7500
-0.8750
-1.0000
See Also
num2bin
Convert a numeric string to binary.
11-17
ca2tf
Purpose
11ca2tf
Convert coupled allpass filter form to transfer function forms
Syntax
[b,a] = ca2tf(d1,d2)
[b,a] = ca2tf(d1,d2,beta)
[b,a,bp] = ca2tf(d1,d2)
[b,a,bp] = ca2tf(d1,d2,beta)
Description
[b,a]=ca2tf(d1,d2) returns the vector of coefficients b and the vector of
coefficients a corresponding to the numerator and the denominator of the
transfer function
1
H ( z ) = B ( z ) ⁄ A ( z ) = --- [ H1 ( z ) + H2 ( z ) ]
2
d1 and d2 are real vectors corresponding to the denominators of the allpass
filters H1(z) and H2(z).
[b,a]=ca2tf(d1,d2,beta) where d1, d2 and beta are complex, returns the
vector of coefficients b and the vector of coefficients a corresponding to the
numerator and the denominator of the transfer function
1
H ( z ) = B ( z ) ⁄ A ( z ) = --- [ – ( β ) • H1 ( z ) + β • H2 ( z ) ]
2
[b,a,bp]=ca2tf(d1,d2), where d1 and d2 are real, returns the vector bp of real
coefficients corresponding to the numerator of the power complementary filter
G(z):
1
G ( z ) = Bp ( z ) ⁄ A ( z ) = --- [ H1 ( z ) – H2 ( z ) ]
2
[b,a,bp]=ca2tf(d1,d2,beta), where d1, d2 and beta are complex, returns the
vector of coefficients bp of real or complex coefficients that correspond to the
numerator of the power complementary filter G(z):
1
G ( z ) = Bp ( z ) ⁄ A ( z ) = ----- [ – ( β ) • H1 ( z ) + β • H2 ( z ) ]
2j
11-18
ca2tf
Examples
Create a filter, convert the filter to coupled allpass form, and convert the result
back to the original structure (create the power complementary filter as well).
[b,a]=cheby1(10,.5,.4)
[d1,d2,beta]=tf2ca(b,a)
[num,den,numpc]=ca2tf(d1,d2,beta)
[H,w,s]=freqz(num,den)
Hpc = freqz(numpc,den)
s.plot = 'mag'
s.yunits = 'sq'
freqzplot([H Hpc],w,s)
See Also
cl2tf
iirpowcomp
tf2ca
tf2cl
%tf2ca returns the
denominators of the allpasses.
%Reconstruct the original
filter plus the power
complementary one.
%Plot the mag response of the
original filter and the power
complementary one.
Convert coupled allpass forms to transfer function
forms.
Compute a power complementary filter.
Convert transfer function forms to coupled allpass
forms.
Convert transfer function forms to coupled allpass
lattice forms.
11-19
cell2sos
Purpose
11cell2sos
Convert cell array to second-order-section matrix
Syntax
s = cell2sos(c)
[s,g] = cell2sos(c)
Description
s = cell2sos(c) converts cell array c of the form
c = { {b1,a1}, {b2,a2}, ... {bi,ai} }
where each numerator vector bi and denominator vector ai contains the
coefficients of a linear or quadratic polynomial, to an L-by-6 second-order
section matrix s of the form
s = [b1 a1
b2 a2
...
bi ai]
When cell2sos encounters linear sections, it zero-pads the sections on the
right.
[s,g] = cell2sos(c) when the first element of c is a pair of scalars, forms the
scalar gain g with the first pair and uses the remaining elements of c to build
the s matrix.
Examples
c = {{[0.0181 0.0181],[1.0000 -0.5095]},{[1 2 1],[1 -1.2505
0.5457]}}
c =
{1x2 cell}
{1x2 cell}
s = cell2sos(c)
s =
0.0181
1.0000
See Also
11-20
sos2cell
sos2ss
sos2tf
0.0181
2.0000
0
1.0000
1.0000
1.0000
-0.5095
-1.2505
0
0.5457
Convert second-order section to cell array form.
Convert second-order section to state-space form.
Convert second-order section to transfer function form.
cell2sos
sos2zp
ss2sos
tf2sos
zp2sos
Convert second-order-section to zero-pole-gain form.
Convert state-space to second-order section form.
Convert digital filter transfer function data to
second-order sections form.
Convert zero-pole-gain to second-order section form.
11-21
cl2tf
Purpose
11cl2tf
Convert coupled allpass lattice to transfer function form
Syntax
[b,a] = cl2tf(k1,k2)
[b,a] = cl2tf(k1,k2,beta)
[b,a,bp] = ca2tf(k1,k2)
[b,a,bp] = ca2tf(k1,k2,beta)
Description
[b,a] = cl2tf(k1,k2) returns the numerator and denominator vectors of
coefficients b and a corresponding to the transfer function
1
H ( z ) = B ( z ) ⁄ A ( z ) = --- [ H1 ( z ) + H2 ( z ) ]
2
where H1(z) and H2(z) are the transfer functions of the allpass filters
determined by k1 and k2, and k1 and k2 are real vectors of reflection
coefficients corresponding to allpass lattice structures.
[b,a] = cl2tf(k1,k2,beta) where k1, k2 and beta are complex, returns the
numerator and denominator vectors of coefficients b and a corresponding to the
transfer function
1
H ( z ) = B ( z ) ⁄ A ( z ) = --- [ – ( β ) • H1 ( z ) + β • H2 ( z ) ]
2
[b,a,bp] = ca2tf(k1,k2) where k1 and k2 are real, returns the vector bp of
real coefficients corresponding to the numerator of the power complementary
filter G(z)
1
G ( z ) = Bp ( z ) ⁄ A ( z ) = --- [ H1 ( z ) – H2 ( z ) ]
2
[b,a,bp] = ca2tf(k1,k2,beta) where k1, k2 and beta are complex, returns
the vector of coefficients bp of possibly complex coefficients corresponding to
the numerator of the power complementary filter G(z)
1
G ( z ) = Bp ( z ) ⁄ A ( z ) = ----- [ – ( β ) • H1 ( z ) + β • H2 ( z ) ]
2j
Examples
11-22
[b,a]=cheby1(10,.5,.4);
[k1,k2,beta]=tf2cl(b,a); %TF2CL returns the reflection coeffs
cl2tf
% Reconstruct the original filter
% plus the power complementary one.
[num,den,numpc]=cl2tf(k1,k2,beta);
[h,w,s1]=freqz(num,den);
hpc = freqz(numpc,den);
s.plot = 'mag';
s.yunits = 'sq';
% Plot the mag response of the original filter and the power
% complementary one.
freqzplot([h hpc],w,s1);
See Also
tf2cl
tf2ca
ca2tf
tf2latc
latc2tf
iirpowcomp
Convert transfer function form to coupled allpass
lattice form.
Convert transfer function form to coupled allpass form.
Convert coupled allpass form to transfer function form.
Convert transfer function to lattice filter form.
Convert lattice function form to transfer function form.
Compute the power complementary filter.
11-23
convergent
Purpose
11convergent
Apply convergent rounding
Syntax
convergent(x)
Description
convergent(x) rounds the elements of x to the nearest integer, except in a tie,
then round to the nearest even integer.
Examples
round and convergent differ in the way they treat values whose fractional part
is 0.5. In round, every tie is rounded up in absolute value. convergent rounds
ties to the nearest even integer.
x=[-3.5:3.5]';
[x convergent(x) round(x)]
ans =
-3.5000
-2.5000
-1.5000
-0.5000
0.5000
1.5000
2.5000
3.5000
See Also
11-24
-4.0000
-2.0000
-2.0000
0
0
2.0000
2.0000
4.0000
quantizer/quantize
-4.0000
-3.0000
-2.0000
-1.0000
1.0000
2.0000
3.0000
4.0000
Quantize numeric data.
convert
Purpose
11convert
Convert filter structures of quantized filters
Syntax
hq = convert(hq,newstruct)
Description
hq = convert(hq,newstruct) returns a quantized filter whose structure has
been transformed to the filter structure specified by string newstruct. You can
enter any one of the following quantized filter structures:
• 'antisymmetricfir': Antisymmetric finite impulse response (FIR).
• 'df1': Direct form I.
• 'df1t': Direct form I transposed.
• 'df2': Direct form II.
• 'df2t': Direct form II transposed. Default filter structure.
• 'fir': FIR.
• 'firt': Direct form FIR transposed.
• 'latcallpass': Lattice allpass.
• 'latticeca': Lattice coupled-allpass.
• 'latticecapc': Lattice coupled-allpass power-complementary.
• 'latticear': Lattice autoregressive (AR).
• 'latticema': Lattice moving average (MA) minimum phase.
• 'latcmax': Lattice moving average (MA) maximum phase.
• 'latticearma': Lattice ARMA.
• 'statespace': Single-input/single-output state-space.
• 'symmetricfir': Symmetric FIR. Even and odd forms.
All filters can be converted to the following structures:
• df1
• df1t
• df2
• df2t
• statespace
• latticearma
11-25
convert
For the following filter classes, you can specify other conversions as well:
• Minimum phase FIR filters can be converted to latticema
• Maximum phase FIR filters can be converted to latcmax
• Allpass filters can be converted to latcallpass
convert generates an error if you specify a conversion that is not possible.
Examples
[b,a]=ellip(5,3,40,.7);
Hq = qfilt('df2t',{b,a});
Hq2 = convert(Hq,'statespace')
Hq2 =
Quantized State-space filter
.
.
.
FilterStructure = statespace
ScaleValues = [1]
NumberOfSections = 1
StatesPerSection = [5]
CoefficientFormat = quantizer('fixed',
InputFormat = quantizer('fixed',
OutputFormat = quantizer('fixed',
MultiplicandFormat = quantizer('fixed',
ProductFormat = quantizer('fixed',
SumFormat = quantizer('fixed',
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
Warning: 9 overflows in coefficients.
See Also
11-26
qfilt
Construct a quantized filter.
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
15])
15])
15])
30])
30])
copyobj
Purpose
11copyobj
Make an independent copy of a quantizer, quantized filter, or quantized FFT
Syntax
obj1 = copyobj(obj)
[obj1,obj2,...] = copyobj(obja,objb,...)
Description
obj1 = copyobj(obj) makes a copy of obj and returns it in obj1. obj can be a
quantizer, quantized filter, or quantized FFT.
[obj1,obj2,...] = copyobj(obja,objb,...) copies obja into obj1, objb
into obj2, and so on. All objects can be quantizers, quantized filters, or
quantized FFTs.
Using copyobj to copy a quantizer, quantized filter, or quantized FFT is not the
same as using the command syntax object1 = object to copy a quantized
object. Quantized filters, quantized FFTs, and quantizers have memory (their
read-only properties). When you use copyobj, the resulting copy is
independent of the original item—it does not share the original object’s
memory, such as the values of the properties min, max, noverflows, or
noperations. Using object1 = object creates a new object that is an alias for
the original and shares the original object’s memory, and thus its property
values.
Examples
q = quantizer('CoefficientFormat',[8 7]);
q1 = copyobj(q);
You can combine quantizers and quantized filters in the same copyobj
command. You cannot include quantized FFTs with other quantized objects in
one copyobj command.
hq = qfilt;
q=quantizer;
[hq1,q1] = copyobj(hq,q)
See Also
qfilt
qfft
quantizer
get
set
Create a quantized filter.
Create a quantized FFT.
Create a quantizer.
Return the property values for quantized filters,
quantizers, and quantized FFTs.
Set or display property values for quantized filters,
quantizers, and quantized FFTs.
11-27
denormalmax
Purpose
11denormalmax
Return the largest denormalized quantized number for a quantizer
Syntax
x = denormalmax(q)
Description
x = denormalmax(q) is the largest positive denormalized quantized number
where q is a quantizer. Anything larger than x is a normalized number.
Denormalized numbers apply only to floating-point format. When q represents
fixed-point numbers, this function returns eps(q).
Examples
q = quantizer('float',[6 3]);
x = denormalmax(q)
returns the value x = 0.1875 = 3/16.
Algorithm
When q is a floating-point quantizer,
denormalmax(q) = realmin(q) - denormalmin(q).
When q is a fixed-point quantizer, denormalmax(q) = eps(q).
See Also
11-28
denormalmin
eps
quantizer
Smallest denormalized quantized number.
Return the relative accuracy of a quantizer.
Construct a quantizer.
denormalmin
Purpose
11denormalmin
Return the smallest denormalized quantized number for a quantizer
Syntax
x = denormalmin(q)
Description
x = denormalmin(q) is the smallest positive denormalized quantized number
where q is a quantizer. Anything smaller than x underflows to zero with
respect to the quantizer q. Denormalized numbers apply only to floating-point
format. When q represents a fixed-point number, denormalmin returns eps(q).
Examples
q = quantizer('float',[6 3]);
denormalmin(q)
returns the value 0.0625 = 1/16.
Algorithm
When q is a floating-point quantizer, x = 2
exponent(q).
Emin – f
When q is a fixed-point quantizer, x = eps ( q ) = 2
fractionlength(q).
See Also
denormalmax
eps
quantizer
, where Emin is equal to
–f
, where f is equal to
Largest denormalized quantized number.
Return the relative accuracy of a quantizer.
Construct a quantizer.
11-29
disp
Purpose
11disp
Description
Similar to omitting the closing semicolon from an expression on the command
line, except that disp does not display the variable name. disp lists the
property names and property values for a quantizer, quantized filter, or
quantized FFT.
Display a quantizer, quantized FFT, or quantized filter
The following examples illustrate the default display for a quantized FFT F and
a quantizer q.
F = qfft;
disp(F)
Radix
Length
CoefficientFormat
InputFormat
OutputFormat
OperandFormat
ProductFormat
SumFormat
NumberOfSections
ScaleValues
q = quantizer
q =
Mode
RoundMode
OverflowMode
Format
Max
Min
NOverflows
NUnderflows
See Also
11-30
set
=
=
=
=
=
=
=
=
=
=
2
16
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
4
[1]q = quantizer
=
=
=
=
=
=
=
=
fixed
floor
saturate
[16 15]
reset
reset
0
0
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
15])
15])
15])
30])
30])
Set or display property values for quantized filters,
quantizers, and quantized FFTs.
eps
Purpose
11eps
Return the quantized relative accuracy for quantized filters, quantizers, and
quantized FFTs
Syntax
eps(q)
eps(hq)
eps(f)
Description
eps(q) returns the quantization level of quantizer object q, or the distance
from 1.0 to the next largest floating-point number when q is a floating-point
quantizer object.
eps(hq) returns the quantization level of quantized filter hq, or the distance
from 1.0 to the next largest floating-point number when hq is a floating-point
quantized filter.
eps(f) returns the quantization level of quantized FFT f, or the distance from
1.0 to the next largest floating-point number when f is a floating-point
quantized FFT.
Examples
q = quantizer('float',[6 3]);
eps(q)
returns the value 0.25. The following code
hq = qfilt;
f = qfft;
eps(hq)
eps(f)
returns the values
coefficient
input
output
multiplicand
product
sum
3.051757813e-005
3.051757813e-005
3.051757813e-005
3.051757813e-005
9.313225746e-010
9.313225746e-010
for the quantizers in the quantized filter, and
11-31
eps
coefficient
input
output
multiplicand
product
sum
3.051757813e-005
3.051757813e-005
3.051757813e-005
3.051757813e-005
9.313225746e-010
9.313225746e-010
for the quantizers in the quantized FFT.
–f
Algorithm
For fixed-point or floating-point numbers, e = 2 where e is the relative
accuracy of the quantizer and f is the fraction length of the quantizer.
See Also
eps
exponentbias
exponentlength
exponentmax
exponentmin
11-32
Return the relative accuracy of a quantizer.
Return the exponent bias for a quantizer.
Return the exponent length for a quantizer.
Return the maximum exponent for a quantizer.
Return the minimum exponent for a quantizer.
exponentbias
Purpose
11exponentbias
Return the exponent bias for a quantizer
Syntax
b = exponentbias(q)
Description
b = exponentbias(q) returns the exponent bias of the quantizer q. For
fixed-point quantizers, exponentbias(q) returns 0.
Examples
q = quantizer('double');
b = exponentbias(q)
returns the value b = 1023.
Algorithm
e–1
For floating-point quantizers, b = 2
– 1 , where e = eps(q), and
exponentbias is the same as the exponent maximum.
For fixed-point quantizers, b = 0 by definition.
See Also
eps
exponentlength
exponentmax
exponentmin
Return the relative accuracy of a quantizer.
Return the exponent length for a quantizer.
Return the maximum exponent for a quantizer.
Return the minimum exponent for a quantizer.
11-33
exponentlength
Purpose
11exponentlength
Return the exponent length of a quantizer
Syntax
e = exponentlength(q)
Description
e = exponentlength(q) returns the exponent length of quantizer q. When q is
a fixed-point quantizer, exponentlength(q) returns 0. This is useful because
exponent length is valid whether the quantizer mode is floating-point or
fixed-point.
Examples
q = quantizer('double');
e = exponentlength(q);
returns the value e = 11.
Algorithm
The exponent length is part of the format of a floating-point quantizer [w, e].
For fixed-point quantizers, e = 0 by definition.
See Also
eps
exponentbias
exponentmax
exponentmin
11-34
Return the relative accuracy of a quantizer.
Return the exponent bias for a quantizer.
Return the maximum exponent for a quantizer.
Return the minimum exponent for a quantizer.
exponentmax
Purpose
11exponentmax
Return the maximum exponent for a quantizer
Syntax
exponentmax(q)
Description
exponentmax(q) returns the maximum exponent for quantizer q. When q is a
fixed-point quantizer, it returns 0.
q = quantizer('double');
exponentmax(q)
returns the value ans = 1023.
Algorithm
For floating-point quantizers, E max = 2
e–1
–1 .
For fixed-point quantizers, E max = 0 by definition.
See Also
eps
exponentbias
exponentlength
exponentmin
Return the relative accuracy of a quantizer.
Return the exponent bias for a quantizer.
Return the exponent length for a quantizer.
Return the minimum exponent for a quantizer.
11-35
exponentmin
Purpose
11exponentmin
Return the minimum exponent for a quantizer
Syntax
emin = exponentmin(q)
Description
emin = exponentmin(q) returns the minimum exponent for quantizer q. If q is
a fixed-point quantizer, exponentmin returns 0.
Examples
q = quantizer('double');
emin = exponentmin(q)
returns the value emin = –1022.
Algorithm
For floating-point quantizers, E min = – 2
e–1
+2.
For fixed-point quantizers, E min = 0 .
See Also
11-36
eps
exponentbias
exponentlength
exponentmax
Return the relative accuracy of a quantizer.
Return the exponent bias for a quantizer.
Return the exponent length for a quantizer.
Return the maximum exponent for a quantizer.
fdatool
Purpose
11fdatool
Open the Filter Design and Analysis Tool.
Syntax
fdatool
Description
fdatool opens the Filter Design and Analysis Tool (FDATool). Use this tool to:
• Design filters
• Analyze filters
• Modify existing filter designs
Refer to Chapter 9, “Quantization Tool Overview” for more information about
using the quantization features of FDATool. For general information about
using FDATool, refer to “Filter Design and Analysis Tool” in Signal Processing
Toolbox User’s Guide.
11-37
fdatool
When you open FDATool and you have Filter Design Toolbox installed,
FDATool incorporates features that are provided by Filter Design Toolbox.
With Filter Design Toolbox installed, FDATool lets you design and analyse
quantized filters, as well as convert quantized filters to various filter
structures. Use the Set Quantization Parameters option to configure the
quantization settings for a quantized filter, or to scale the filter coefficients.
Set Quantization Parameters — provides access to the properties of the
quantizers that compose a quantized filter. When you click Set Quantization
Parameters, you see FDATool displaying the quantization parameters at the
bottom of the dialog, as shown in the figure.
Scale transfer-fnc coeffs <=1 — to avoid coefficient overflows when you
quantize a filter, this option normalizes the quantized filter coefficients not to
exceed one. Scaling filter coefficients changes the coefficient values and adjusts
the filter gain to keep the same filter magnitude response after scaling. You
11-38
fdatool
cannot undo the scaling operation or remove the modified gain value. To
remove the applied scaling and gain, quantize the reference filter again.
Input/output scaling — reports the scale values applied to the coefficients of
the current quantized filter.
Remarks
The Filter Design and Analysis Tool provides more design methods than the
SPTool Filter Designer.
See Also
sptool
Open the interactive digital signal processing tool
(SPTool) included in Signal Processing Toolbox.
11-39
fft
Purpose
11fft
Apply a quantized fast Fourier transform to data
Syntax
y = fft(F,x)
y = fft(F,x,dim)
Description
y = fft(F,x) uses quantized FFT (fast Fourier transform) F to compute the
FFT of vector x. The parameters of the quantized FFT are specified in
quantized FFT F. The radix is specified by F.radix. The decimation is specified
by F.decimation. The length of the FFT is specified by F.length. When the
length of x is less than F.length, x is padded with zeros. When x is longer than
F.length, x is truncated. For matrices, the FFT operation is applied to each
column. For N-D arrays, the FFT operation operates on the first nonsingleton
dimension.
y = fft(F,x,dim) applies the quantized FFT operation across the dimension
dim.
Examples
When you quantize a sinusoid, you generate errors as a result of the
quantization process. This example demonstrates this effect. We create a
sinusoid, quantize it, and look at the error between the quantized and
unquantized sinusoids. Then we plot the FFTs for both signals.
n = 128;
t = (0:n-1/n);
x = sin(2*pi*16*t)/16;% Reference sinusoid
q = quantizer([5 4]);
f = qfft('length',n,'inputformat',q);
plot(t,[quantize(q,x);x]);% Plot both signals
plot(t,[quantize(q,x)-x]);% Plot the error
plot(t,[20*log10(abs(fft(f,x)));...
(20*log10(abs(fft(x)))/20)])% Plot the FFTs for both signals
11-40
fft
The following figure presents the results.
Reference Sinusoid & Quantized Sinusoid
0.1
Linear
0.05
0
−0.05
−0.1
0
0.1
0.2
0
0.1
0.2
0.3
0.4
0.5
Error
0.6
0.7
0.8
0.9
1
0.8
0.9
1
0
Linear
−0.02
−0.04
−0.06
−0.08
0.3
0.4
0.5
0.6
0.7
FFTs of the Reference and Quantized Sinusoids
20
Quantized
Reference
dB
0
−20
−40
0
0.1
0.2
0.3
0.4
0.5
0.6
Time in Seconds
0.7
0.8
0.9
1
Looking at the subplot of the error between the reference and quantized
sinusoids, you see that the error is periodic. Because the error is periodic, the
FFT of the quantized sinusoid includes periodic frequency content not in the
reference signal, as seen in the FFTs subplot.
See Also
qfft
Construct a quantized FFT.
11-41
filter
Purpose
11filter
Apply a quantized filter to data and access states and filtering information
Syntax
y = filter(Hq,x)
[y,zf] = filter(Hq,x)
[...] = filter(Hq,x,zi)
[...] = filter(Hq,x,zi,dim)
[y,zf,s,z,v] = filter(Hq,x...)
Description
y = filter(Hq,x) filters a vector of real or complex input data x through a
quantized filter Hq, producing filtered output data y. The vectors x and y have
the same length.
If x is a matrix, y = filter(Hq,x) filters each column of x to produce a matrix
y. If x is a multidimensional array, y = filter(Hq,x) filters x along the first
nonsingleton dimension of x.
[y,zf] = filter(Hq,x) produces an additional output argument zf.
zf contains the final values for the state vector calculated from zero initial
conditions for the state. The form zf takes depends on the data to be filtered
and the number of stages in the filter, as detailed in Table 11-12, Final State
Form Depends on Filtered Data and Filter Structure.
[...] = filter(Hq,x,zi) specifies the initial conditions for the state vector
in zi. The form for specifying zi is described in Table 11-11, Initial State
Format Depends on the Filter Structure. To specify the same initial condition
for all state components, enter zi as a scalar. You can set zi to zero, [], or {}
to specify zero (the default) initial conditions.
The form of the initial and final states associated with a quantized filter Hq
depends on the filter structure and the data to be filtered. The following tables
give the form for either entering the initial states or retrieving the final states
of the quantized filter.
11-42
filter
+The variables in these tables are described as follows:
Table 11-11: Initial State Format Depends on the Filter Structure
Number of Filter
Sections
Format of the Initial State
1
A column vector of length s1
n
A 1-by-n cell array of vectors of length si,
i=1, 2,...,n
Table 11-12: Final State Form Depends on Filtered Data and Filter Structure
Filtered Data
Number of
Filter Sections
Form of the Final State
Vector
1
A column vector of length s1
Vector
n
A 1-by-n cell array of vectors of
length si, i=1, 2,...,n
Multidimensional
array
1
An s1-by-c matrix
Multidimensional
array
n
1-by-n cell array of si-by-c
matrices, i=1, 2,...,n
• si is the number of states in the ith section of the filter.
• c is prod(size(x))/size(x,dim), where dim is the first nonsingleton
dimension into which you are filtering.
To figure out the dimensions of the initial or final conditions, run the filter once
with empty initial conditions. Then the final conditions are the right size for
the initial conditions:
[y,zf] = filter(Hq,x);
Look at the size and data type of zf. The initial conditions, zi, will be the same
size as zf.
Use the StatesPerSection property of the quantized filter Hq to access the
number of states in each section. See “Quantized Filter Properties Reference”
on page 10-11 for more information on filter properties.
11-43
filter
[...] = filter(Hq,x,zi,dim) applies the quantized filter Hq to the input
data located along the specific dimension of x specified by dim.
[y,zf,s,z,v] = filter(Hq,x...) returns s, a MATLAB structure
containing quantization information (refer to qreport for details); z, the filter’s
state sequence; and v, the number of overflows at each time step of the filter.
When you include four or five output arguments, the input argument x must
be a vector. z is a cell array containing the sequence of states at each time step,
having 1 element per filter and 1 column per time step. The initial conditions
of the kth filter section are in the first column of z{k}:zi{k}=z{k}(:,1). The
final conditions of the kth filter section are in the last column of z{k}:zf{k} =
z{k}(:,end). Overflows for the kth section are in v{k}.
Examples
Find the response of a quantized digital filter.
randn('state',0);
x = randn(100,1);
warning on;
[b,a] = butter(3,.9,'high');
Hq = sos(qfilt('referencecoefficients',{b,a}))
Warning: 3 overflows in coefficients.
y = filter(Hq,x);
Warning: 27 overflows in QFILT/FILTER.
Max
Coefficient
Input
Output
Multiplicand
Product
Sum
11-44
1
0.8238
2.183
0.4361
1
0.4361
0.01276
0.4361
0.01278
0.2181
Min
-0.7419
-1
-2.171
-0.45
-1
-0.45
-0.01227
-0.45
-0.01221
-0.225
NOverflows
0
0
27
0
0
0
0
0
0
0
NUnderflows
0
0
0
0
2
0
0
0
0
0
NOperations
4
6
100
100
600
700
600
700
300
500
filter
Hq.filterstructure
ans =
df2t
Notice the warnings returned during filter quantization and application. The
first warning indicates that one of the filter coefficients overflowed during
quantization before converting the filter to second-order section form. Applying
the function sos to the filter removed the coefficient overflows. The second
warning displays the overflow report, listing details about the filtering
operation.
Note Use qreport to display the information logged during a filtering
operation.
Algorithm
The filter command implements fixed- or floating-point arithmetic on the
quantized filter structure you specify. The state vector z associated with the
filter is a vector whose components are derived from the values of each of the
input signals to each delay in the filter. The length of z is the same as the
number of delays in the filter.
The implementation of filter depends on the filter structure. For example,
the operation of filter at sample m for a transposed direct form II filter is
given by the quantized time domain difference equations for y and the states zi
shown below. Square brackets denote the quantization that takes place for the
input data x, the output data y, the coefficients, the products, and the sums.
[ [ [ b(1) ] [ x(m) ] ] + z 1(m – 1) ]
y(m) = ------------------------------------------------------------------------[a(1)]
z 1(m) = [ [ [ b(2) ] [ x ( m ) ] ] + z 2(m – 1) – [ [ a(2) ]y(m) ] ]
M = M
z n – 2(m) = [ [ [ [ b(n – 1) ] [ x(m) ] ] + z n – 1(m – 1) ] –[ [ a(n – 1) ]y(m) ] ]
z n – 1(m) = [ [ [ b(n) ] [ x(m) ] ] – [ [ a(n) ]y(m) ] ]
11-45
filter
Notice that for this df2t filter structure, you divide by a(1). For efficient
computation, choose a(1) to be a power of 2.
Note qfilt/filter does not normalize the filter coefficients automatically.
Function filter supplied by MATLAB does normalize the coefficients.
See Also
impz
qfilt
qreport
References
[1] Oppenheim, A.V., and R.W. Schafer, Discrete-Time Signal Processing, Prentice-Hall, 1989.
11-46
Compute impulse response for quantized filters.
Construct a quantized filter.
Display the contents of the filtering log.
firlpnorm
Purpose
11firlpnorm
Least P-norm optimal FIR filter design
Syntax
[num,den]
[num,den]
[num,den]
[num,den]
[num,den]
Description
[num,den] = firlpnorm(n,d,f,edges,a) returns a filter having a numerator
order n and denominator order d which is the best approximation to the desired
frequency response described by f and a in the least-pth norm sense. The vector
edges specifies the band-edge frequencies for multi-band designs. An
=
=
=
=
=
firlpnorm(n,d,f,edges,a)
firlpnorm(n,d,f,edges,a,w)
firlpnorm(n,d,f,edges,a,w,p)
firlpnorm(n,d,f,edges,a,w,p,dens)
firlpnorm(n,d,f,edges,a,w,p,dens,initnum,initden)
unconstrained quasi-Newton algorithm is employed and any poles or zeros that
lie outside of the unit circle are reflected back inside. n and d should be chosen
so that the zeros and poles are used effectively. See the “Hints” section below.
Always use freqz to check the resulting filter.
[num,den] = firlpnorm(n,d,f,edges,a,w) uses the weights in w to weight
the error. w has one entry per frequency point (the same length as f and a)
which tells firlpnorm how much emphasis to put on minimizing the error in
the vicinity of each frequency point relative to the other points. f and a must
have the same number of elements, which may exceed the number of elements
in edges. This allows for the specification of filters having any gain contour
within each band. The frequencies specified in edges must also appear in the
vector f. For example,
[num,den] = firlpnorm(5,12,[0 .15 .4 .5 1],[0 .4 .5 1],...
[1 1.6 1 0 0],[1 1 1 10 10])
is a lowpass filter with a peak of 1.6 within the passband.
[num,den] = firlpnorm(n,d,f,edges,a,w,p) where p is a two-element
vector [pmin pmax] allows for the specification of the minimum and maximum
values of p used in the least-pth algorithm. Default is [2 128] which essentially
yields the L-infinity, or Chebyshev, norm. Pmin and pmax should be even. If p is
the string 'inspect', no optimization will occur. This can be used to inspect
the initial pole/zero placement.
11-47
firlpnorm
[num,den] = firlpnorm(n,d,f,edges,a,w,p,dens) specifies the grid density
dens used in the optimization. The number of grid points is [dens*(n+d+1)].
The default is 20. dens can be specified as a single-element cell array. The grid
is not equally spaced.
[num,den] = firlpnorm(n,d,f,edges,a,w,p,dens,initnum,initden)
allows for the specification of the initial estimate of the filter numerator and
denominator coefficients in vectors initnum and initden. This may be useful
for difficult optimization problems. The pole-zero editor in the Signal
Processing Toolbox can be used for generating initnum and initden.
Hints
• This is a weighted least-pth optimization.
• Check the radii and locations of the poles and zeros for your filter. If the zeros
are on the unit circle and the poles are well inside the unit circle, try
increasing the order of the numerator or reducing the error weighting in the
stopband.
• Similarly, if several poles have a large radii and the zeros are well inside of
the unit circle, try increasing the order of the denominator or reducing the
error weighting in the passband.
See Also
iirlpnorm
filter
freqz
iirgrpdelay
iirlpnormc
zplane
11-48
Least-p norm optimal IIR filter design.
Filter data with a recursive (IIR) or nonrecursive (FIR)
filter.
Frequency response of a filter.
Optimal IIR filter design with prescribed group-delay.
Constrained least-p norm optimal IIR filter design.
Zero-pole plot.
fractionlength
Purpose
11fractionlength
Return the fraction length for a quantizer
Syntax
f = fractionlength(q)
Description
f = fractionlength(q) returns the fraction length of quantizer q. This is
useful because fraction length is valid whether the quantizer mode is
floating-point or fixed-point.
Examples
For a floating-point quantizer
q = quantizer('float',[32 8]);
f = fractionlength(q);
returns f = 23 = 32 – 8 – 1 .
For a fixed-point quantizer
q = quantizer('fixed',[6 4])
f = fractionlength(q);
returns f = 4.
Algorithm
For floating-point quantizers, f = w – e – 1, where w is the word length and e is
the exponent length.
For fixed-point quantizers, f is part of the format [w f].
See Also
quantizer
Construct a quantizer.
11-49
freqz
Purpose
11freqz
Compute the frequency response of quantized filters
Syntax
[h,w] = freqz(Hq,n)
h = freqz(Hq,w)
[h,w] = freqz(Hq,n,'whole')
[h,w,units,href] = freqz(Hq,...)
[h,f] = freqz(Hq,n,fs)
h = freqz(Hq,f,fs)
[h,f] = freqz(Hq,n,'whole',fs)
[h,f,units,href] = freqz(Hq,...,fs)
freqz(Hq,...)
Description
[h,w] = freqz(Hq,n) returns the frequency response vector h and the
corresponding frequency vector w for the quantized filter Hq. freqz uses the
transfer function associated with the quantized filter to calculate the frequency
response of the filter. The vectors h and w are both of length n. The frequency
vector w has values ranging from 0 to π radians per sample. If you do not specify
the integer n, or you specify it as the empty vector [], the frequency response
is calculated using the default value of 512 samples.
h = freqz(Hq,w) returns the frequency response vector h calculated at the
frequencies (in radians per sample) supplied by the vector w. The vector w can
have any length.
[h,w] = freqz(Hq,n,'whole') uses n sample points around the entire unit
circle to calculate the frequency response. Frequency vector w has length n and
values ranging from 0 to 2π radians per sample.
[h,w,units,href] = freqz(Hq,...) returns the optional string argument
units, specifying the units for the frequency vector w. The string returned in
units is 'rad/sample', denoting radians per sample. The optional output
argument href is the frequency response of the transfer function associated
with the reference filter used to specify the quantized filter Hq.
[h,f] = freqz(Hq,n,fs) returns the frequency response vector h and the
corresponding frequency vector f for the quantized filter Hq. The vectors h and
f are both of length n. The frequency response calculation uses the sampling
11-50
freqz
frequency specified by the scalar fs (in Hz). The frequency vector f has values
ranging from 0 to (fs/2) Hz.
h = freqz(Hq,f,fs) returns the frequency response vector h calculated at the
frequencies (in Hz) supplied in the vector f. Vector f can be any length.
[h,f] = freqz(Hq,n,'whole',fs) uses n points around the entire unit circle
to calculate the frequency response. Frequency vector f has length n and has
values ranging from 0 to fs Hz.
[h,f,units,href] = freqz(Hq,...,fs) returns the optional MATLAB
structure units, that freqzplot uses for plotting. The string returned in units
is 'Hz' for hertz. The optional output argument href is the frequency response
of the transfer function associated with the reference filter used to specify the
quantized filter Hq.
freqz(Hq,...) plots the magnitude and unwrapped phase of the frequency
response of the quantized filter Hq in the current figure window.
Remarks
There are several ways of analyzing the frequency response of quantized
filters. freqz accounts for quantization effects in the filter coefficients, but does
not account for quantization effects in filtering arithmetic. To account for the
quantization effects in filtering arithmetic, refer to function nlm.
Algorithm
freqz calculates the frequency response for a quantized filter from the filter
transfer function Hq(z). The complex-valued frequency response is calculated
by evaluating Hq(ejω) at discrete values of w specified by the syntax you use.
The integer input argument n determines the number of equally-spaced points
around the upper half of the unit circle at which freqz evaluates the frequency
response. The frequency ranges from 0 to π radians per sample when you do not
supply a sampling frequency as an input argument. When you supply the
scalar sampling frequency fs as an input argument to freqz, the frequency
ranges from 0 to fs/2 Hz.
To calculate the transfer function associated with a quantized filter, freqz
uses the values of the QuantizedCoefficients and FilterStructure
properties.
11-51
freqz
When you include the optional output argument href in the command, freqz
uses the value of the ReferenceCoefficients property to calculate the
frequency response of the reference filter transfer function.
Examples
Plot the estimated frequency response of a quantized filter.
b = fir1(80,0.5,kaiser(81,8));
Hq = qfilt('fir',{b});
[h,w,units,href] = freqz(Hq);
plot(w,20 * log10(abs(h)),'-',w,20 * log10(abs(href)),'--')
legend('Quantized filter','Reference filter',3)
xlabel('Frequency in rad/sample')
ylabel('Magnitude in dB')
title('Magnitude of the Frequency Response Compared')
Magnitude of the Frequency Response Compared
20
0
−20
Magnitude in dB
−40
−60
−80
−100
−120
−140
Quantized filter
Reference filter
−160
See Also
11-52
0
qfilt
freqzplot
0.5
1
1.5
2
Frequency in rad/sample
2.5
Construct a quantized filter.
Plot frequency response data.
3
3.5
get
Purpose
11get
Return the property values for quantized filters, quantizers, and quantized
FFTs
Syntax
get(obj,pn,pv)
get(Hq)
struct = get(Hq)
v = get(Hq,'propertyname')
value = get(F, propertyname)
structure = get(F)
value = get(q, propertyname)
structure = get(q)
Description
get(obj,pn,pv) displays a list of the property names and property values
associated with obj, where obj is one of the following:
• A quantizer
• A quantized filter
• A quantized FFT
pn is the name of a property of the object obj, and pv is the value associated
with pn.
get(Hq) displays a list of the property names and property values associated
with the quantized filter Hq.
struct = get(Hq) returns the MATLAB structure struct, a list of the
properties associated with the quantized filter Hq, along with the properties’
associated values. Each field associated with struct is named according to the
corresponding property name.
v = get(Hq,'propertyname') returns the property value v associated with
the property named in the string 'propertyname' for the quantized filter Hq. If
you replace the string 'propertyname' by a cell array of a vector of strings
containing property names, get returns a cell array of a vector of corresponding
values.
value = get(F, propertyname) returns the property value value associated
with the property named in the string 'propertyname' for the quantized FFT
F. If you replace the string 'propertyname' by a cell array of a vector of strings
11-53
get
containing property names, get returns a cell array of a vector of corresponding
values.
structure = get(F) returns a structure containing the properties and states
of quantized FFT F.
value = get(Q, propertyname) returns the property value value associated
with the property named in the string 'propertyname' for the quantizer Q. If
you replace the string 'propertyname' by a cell array of a vector of strings
containing property names, get returns a cell array of a vector of corresponding
values.
structure = get(Q) returns a structure containing the properties and states
of quantizer Q.
Remarks
For more information on the properties associated with quantized filters, see
“A Quick Guide to Quantized Filter Properties” on page 10-10. For more
information on the properties associated with quantized FFTs, see “A Quick
Guide to Quantized FFT Properties” on page 10-46. For more information on
the properties associated with quantizers, see “A Quick Guide to Quantizer
Properties” on page 10-2.
Examples
Use get to list the properties of quantized filter Hq, along with the property
values. Then retrieve the value associated with the OutputFormat property for
this filter in a structure v.
Hq = qfilt;
get(Hq)
Quantized Direct form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
+ (1)
0.999969482421875 1.000000000000000000
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
+ (1)
0.999969482421875 1.000000000000000000
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
11-54
=
=
=
=
=
=
df2t
[1]
1
[0]
quantizer('fixed', 'round', 'saturate', [16
quantizer('fixed', 'floor', 'saturate', [16
15])
15])
get
OutputFormat =
MultiplicandFormat =
ProductFormat =
SumFormat =
Warning: 2 overflows
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
in coefficients.
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[32
[32
15])
15])
30])
30])
v = get(Hq,'OutputFormat')
v =
Mode
RoundMode
OverflowMode
Format
=
=
=
=
fixed
floor
saturate
[16 15]
Max
Min
NOverflows
NUnderflows
NOperations
=
=
=
=
=
reset
reset
0
0
0
q = quantizer('fixed', 'floor', 'saturate', [16 15])
struct
= get(q)
mode
= get(q, 'mode')
format
= get(q, 'format')
noverflows = get(q, 'noverflows')
get also supports the dot syntax for setting and accessing properties.
q = quantizer('fixed', 'floor', 'saturate', [16 15])
struct
= get(q)
mode
= q.mode
format
= q.format
noverflows = q.noverflows
See Also
qfft
qfilt
quantizer
set
Construct a quantized FFT.
Construct a quantized filter.
Construct a quantizer.
Set property values for quantized filters, quantized
FFTs, and quantizers.
11-55
gremez
Purpose
11gremez
Use the Parks-McClellan technique to design digital FIR filters
Syntax
b
b
b
b
b
b
b
b
b
Description
gremez is a minimax filter design algorithm that you can use to design the
following types of real FIR filters:
=
=
=
=
=
=
=
=
=
gremez(n,f,a,w)
gremez(n,f,a,'hilbert')
gremez(n,f,a,'differentiator')
gremez(m,f,a,r)
gremez({m,ni},f,a,r)
gremez(n,f,a,w,c)
gremez(n,f,a,w,e)
gremez(n,f,a,s)
gremez(n,f,a,s,w,e)
• Types 1-4 linear phase:
- Type 1 is even order, symmetric
- Type 2 is odd order, symmetric
- Type 3 is even order, antisymmetric
- Type 4 is odd order, antisymmetric
• Minimum phase
• Maximum phase
• Minimum order (even or odd)
• Extra ripple
• Maximal ripple
• Constrained ripple
• Single-point band (notching and peaking)
• Forced gain
• Arbitrary shape frequency response curve filters
b = gremez(n,f,a,w) returns a length N+1 linear phase FIR filter which has
the best approximation to the desired frequency response described by f and a
in the minimax sense. W is a vector of weights, one per band. When you omit w,
all bands are weighted equally. For more information on the input arguments,
see the remez help entry.
11-56
gremez
b = gremez(n,f,a,'hilbert') and b = gremez(n,f,a,'differentiator')
design FIR Hilbert transformers and differentiators. For more information on
designing these filters, see the remez help entry.
b = gremez(m,f,a,r), where m is one of 'minorder', 'mineven' or 'minodd',
designs filters repeatedly until the minimum order filter, as specified in m, that
meets the specifications is found. r is a vector containing the peak ripple per
frequency band. You must specify r. When you specify 'mineven' or 'minodd', the
minimum even or odd order filter is found.
b = gremez({m,ni},f,a,r) where m is one of 'minorder', 'mineven' or 'minodd',
uses ni as the initial estimate of the filter order. ni is optional for common filter
designs, but it must be specified for designs in which remezord cannot be used,
such as while designing differentiators or Hilbert transformers.
b = gremez(n,f,a,w,c) designs filters having constrained error magnitudes
(ripples). c is a cell array of strings of length w. The entries of c must be either
'c' to indicate that the corresponding element in w is a constraint (the ripple for
that band cannot exceed w) or 'w' indicating that the corresponding entry in w is
a weight. There must be at least one unconstrained band—c must contain at
least one 'w' entry. For example,
b = gremez(12,[0 0.4 0.5 1], [1 1 0 0], [1 0.2], {'w' 'c'}) uses a
weight of one in the passband, and constrains the stopband ripple to 0.2 or less.
A hint about using constrained values: if the resulting filter does not touch the
constraints, increase the error weighting you apply to the unconstrained
bands.
b = gremez(n,f,a,w,e) specifies independent approximation errors for
different bands. Use this syntax to design extra ripple or maximal ripple filters.
These filters have interesting properties such as having the minimum
transition width. e is a cell array of strings specifying the approximation errors
to use. Its length must equal the number of bands. Entries of e must be in the
form 'e#' where # indicates which approximation error to use for the
corresponding band. For example, when e = {'e1','e2','e1'}, the first and
third bands use the same approximation error 'e1' and the second band uses
a different one 'e2'. Note that when all bands use the same approximation
error, such as {'e1','e1','e1',...}, it is equivalent to omitting e, as in
b = gremez(n,f,a,w).
11-57
gremez
b = gremez(n,f,a,s) is used to design filters with special properties at
certain frequency points. s is a cell array of strings and must be the same
length as f and a. Entries of s must be one of:
• 'n' - normal frequency point.
• 's' - single-point band. The frequency “band” is given by a single point. The
corresponding gain at this frequency point must be specified in a.
• 'f' - forced frequency point. Forces the gain at the specified frequency band
to be the value specified.
• 'i' - indeterminate frequency point. Use this argument when adjacent
bands abut one another (no transition region).
For example, the following command designs a bandstop filter with zero-valued
single-point stop bands (notches) at 0.25 and 0.55.
b = gremez(42,[0 0.2 0.25 0.3 0.5 0.55 0.6 1],[1 1 0 1 1 0 1 1],...
{'n' 'n' 's' 'n' 'n' 's' 'n' 'n'})
b = gremez(82,[0 0.055 0.06 0.1 0.15 1],[0 0 0 0 1 1],...
{'n' 'i' 'f' 'n' 'n' 'n'})
designs a highpass filter with the gain at 0.06 forced to be zero. The band edge
at 0.055 is indeterminate since the first two bands actually touch. The other
band edges are normal.
b = gremez(n,f,a,s,w,e) specifies weights and independent approximation
errors for filters with special properties. The weights and properties are
included in vectors w and e. Sometimes, you may need to use independent
approximation errors to get designs with forced values to converge. For
example,
b = gremez(82,[0 0.055 0.06 0.1 0.15 1], [0 0 0 0 1 1],...
{'n' 'i' 'f' 'n' 'n' 'n'}, [10 1 1] ,{'e1' 'e2' 'e3'});
b = gremez(...,'1') designs a type 1 filter (even-order symmetric). You can
specify type 2 (odd-order symmetric), type 3 (even-order antisymmetric), and
type 4 (odd-order antisymmetric) filters as well. Note that restrictions apply to
a at f=0 or f=1 for FIR filter types 2, 3, and 4.
b = gremez(...,'minphase') designs a minimum-phase FIR filter. You can
use the argument 'maxphase' to design a maximum phase FIR filter.
11-58
gremez
b = gremez(..., 'check') returns a warning when there are potential
transition-region anomalies.
b = remez(...,{lgrid}), where {lgrid} is a scalar cell array. The value of
the scalar controls the density of the frequency grid by setting the number of
samples used along the frequency axis.
[b,err] = gremez(...) returns the unweighted approximation error
magnitudes. err contains one element for each independent approximation
error returned by the function.
[b,err,res] = gremez(...) returns the structure res comprising optional
results computed by gremez. res contains the following fields.
Structure Element
Contents
res.fgrid
Vector containing the frequency grid used in
the filter design optimization
res.des
Desired response on fgrid
res.wt
Weights on fgrid
res.h
Actual frequency response on the frequency
grid
res.error
Error at each point (desired response - actual
response) on the frequency grid
res.iextr
Vector of indices into fgrid of extremal
frequencies
res.fextr
Vector of extremal frequencies
res.order
Filter order
11-59
gremez
Structure Element
Contents
res.edgecheck
Transition-region anomaly check. One element
per band edge. Element values have the
following meanings:
1 = OK
0 = probable transition-region anomaly
-1 = edge not checked
Computed when you specify the 'check' input
option in the function syntax.
res.iterations
Number of Remez iterations for the
optimization
res.evals
Number of function evaluations for the
optimization
gremez is also a “function function”, allowing you to write a function that
defines the desired frequency response.
b = gremez(n,f,fresp,w) returns a length N+1 FIR filter which has the best
approximation to the desired frequency response as returned by the
user-defined function fresp. gremez uses the following syntax to call fresp
[dh,dw] = fresp(n,f,gf,w)
where:
• fresp is the string variable that identifies the function that you use to define
your desired filter frequency response.
• n is the filter order.
• f is the vector of frequency band edges which must appear monotonically
between 0 and 1, where 1 is one-half of the sampling frequency. The
frequency bands span f(k) to f(k+1) for k odd. The intervals f(k+1) to
f(k+2) for k odd are “transition bands” or “don't care” regions during
optimization.
11-60
gremez
• gf is a vector of grid points that have been chosen over each specified
frequency band by gremez, and determines the frequencies at which gremez
evaluates the response function.
• w is a vector of real, positive weights, one per band, for use during
optimization. w is optional in the call to gremez. If you do not specify w, it is
set to unity weighting before being passed to fresp.
• dh and dw are the desired frequency response and optimization weight
vectors, evaluated at each frequency in grid gf.
gremez includes a predefined frequency response function named 'remezfrf2'.
You can write your own based on the simpler 'remezfrf'. See the help for
private/remezfrf for more information.
b = gremez(n,f,{fresp,p1,p2,...},w) specifies optional arguments p1,
p2,..., pn to be passed to the response function fresp.
b = gremez(n,f,a,w) is a synonym for
b = gremez(n,f,{'remezfrf2',a},w), where a is a vector containing your
specified response amplitudes at each band edge in f. By default, gremez
designs symmetric (even) FIR filters. 'remezfrf2' is the predefined frequency
response function. If you do not specify your own frequency response function
(the fresp string variable), gremez uses 'remezfrf2'.
b = gremez(...,'h') and b = gremez(...,'d') design antisymmetric (odd)
filters. When you omit the 'h' or 'd' arguments from the gremez command
syntax, each frequency response function fresp can tell gremez to design either
an even or odd filter. Use the command syntax
sym = fresp('defaults',{n,f,[],w,p1,p2,...}). gremez expects fresp to
return sym = 'even' or sym = 'odd'. If fresp does not support this call,
gremez assumes even symmetry.
For more information about the input arguments to gremez, see remez.
See Also
remez
cremez
butter
cheby1
cheby2
Parks-McClellan optimal FIR filter design.
Complex and nonlinear-phase equiripple FIR filter
design
Butterworth analog and digital filter design.
Chebyshev type I filter design (passband ripple).
Chebyshev type II filter design (stopband ripple).
11-61
gremez
ellip
freqz
filter
filter
firls
fircls
11-62
Elliptic (Cauer) filter design.
Frequency response of digital filters.
Filter data with a recursive (IIR) or nonrecursive (FIR)
filter (in Signal Processing Toolbox).
Apply a quantized filter to data and access states and
filtering information (in Filter Design Toolbox).
Least square linear-phase FIR filter design.
Constrained least square FIR filter design for
multiband filters.
hex2num
Purpose
11hex2num
Convert hexadecimal string to a number
Syntax
x = hex2num(q,h)
[x1,x2,...] = hex2num(q,h1,h2,...)
Description
x = hex2num(q,h) converts hexadecimal string h to numeric matrix x. The
attributes of the numbers in x are specified by quantizer q. When h is a cell
array containing hexadecimal strings, hex2num returns x as a cell array of the
same dimension containing numbers. For fixed-point hexadecimal strings,
hex2num uses two’s complement representation. For floating-point strings, the
representation is IEEE Standard 754 style.
When there are fewer hexadecimal digits than needed to represent the number,
the fixed-point conversion zero fills on the left. Floating-point conversion zero
fills on the right.
[x1,x2,...] = hex2num(q,h1,h2,...) converts hexadecimal strings h1, h2,...
to numeric matrices x1, x2,....
hex2num and num2hex are inverses of one another, with the distinction that
num2hex returns the hexadecimal strings in a column.
Examples
To create all of the 4-bit fixed-point two’s complement numbers fractional form,
use the following code.
q
h
x
x
= quantizer([4 3]);
= '7 3 F B';'6 2 E A';'5 1 D 9';'4 0 C 8'];
= hex2num(q,h)
=
0.8750
0.7500
0.6250
0.5000
See Also
num2hex
bin2num
num2bin
0.3750
0.2500
0.1250
0
-0.1250
-0.2500
-0.3750
-0.5000
-0.6250
-0.7500
-0.8750
-1.0000
Convert a number to its IEEE hexadecimal equivalent.
Convert a binary number to its decimal value.
Convert a number to its two’s complement binary
value.
11-63
ifft
Purpose
11ifft
Apply the inverse quantized FFT to data
Syntax
y = ifft(F,x)
y = ifft(F,x,dim)
Description
y = ifft(F,x) is the quantized inverse FFT of x. The parameters of the
quantized FFT are specified in quantized FFT F.
y = ifft(F,x,dim) is the quantized inverse FFT of x across the dimension
dim.
See Also
11-64
fft
qfft
Apply a quantized FFT to data.
Create a quantized FFT.
iirgrpdelay
Purpose
11iirgrpdelay
Optimal IIR filter design with prescribed group-delay
Syntax
[num,den] = iirgrpdelay(n,f,edges,a)
[num,den] = iirgrpdelay(n,f,edges,a,w)
[num,den] = iirgrpdelay(n,f,edges,a,w,radius)
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p)
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p,dens)
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p,dens,initden)
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p,dens,initden,tau)
[num,den,tau] = iirgrpdelay(n,f,edges,a,w)
Description
[num,den] = iirgrpdelay(n,f,edges,a) returns an allpass IIR filter of
order n (n must be even) which is the best approximation to the relative
group-delay response described by f and a in the least-pth sense. f is a vector
of frequencies between 0 and 1 and a is specified in samples. The vector edges
specifies the band-edge frequencies for multi-band designs. iirgrpdelay uses
a constrained Newton-type algorithm. Always check your resulting filter using
grpdelay or freqz.
[num,den] = iirgrpdelay(n,f,edges,a,w) uses the weights in w to weight
the error. w has one entry per frequency point and must be the same length
length as f and a). Entries in w tell iirgrpdelay how much emphasis to put on
minimizing the error in the vicinity of each specified frequency point relative
to the other points.
f and a must have the same number of elements. f and a can contains more
elements than the vector edges contains. This lets you use f and a to specify a
filter that has any group-delay contour within each band.
[num,den] = iirgrpdelay(n,f,edges,a,w,radius) returns a filter having a
maximum pole radius equal to radius, where 0<radius<1. radius defaults to
0.999999. Filters whose pole radius you constrain to be less than 1.0 can better
retain transfer function accuracy after quantization.
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p), where p is a
two-element vector [pmin pmax], lets you determine the minimum and
maximum values of p used in the least-pth algorithm. p defaults to [2 128]
which yields filters very similar to the L-infinity, or Chebyshev, norm. pmin and
11-65
iirgrpdelay
pmax should be even. If p is the string 'inspect', no optimization occurs. You
might use this feature to inspect the initial pole/zero placement.
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p,dens) specifies the
grid density dens used in the optimization process. The number of grid points
is (dens*(n+1)). The default is 20. dens can be specified as a single-element
cell array. The grid is not equally spaced.
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p,dens,initden) allows
you to specify the initial estimate of the denominator coefficients in vector
initden. This can be useful for difficult optimization problems. The pole-zero
editor in the Signal Processing Toolbox can be used for generating initden.
[num,den] = iirgrpdelay(n,f,edges,a,w,radius,p,dens,initden,tau)
allows the initial estimate of the group delay offset to be specified by the value
of tau, in samples.
[num,den,tau] = iirgrpdelay(n,f,edges,a,w) returns the resulting group
delay offset. In all cases, the resulting filter has a group delay that
approximates [a + tau]. Allpass filters can have only positive group delay and
a non-zero value of tau accounts for any additional group delay that is needed
to meet the shape of the contour specified by (f,a). The default for tau is
max(a).
Hint: If the zeros or poles cluster together, your filter order may be too low or
the pole radius may be too small (overly constrained). Try increasing n or
radius.
For group-delay equalization of an IIR filter, compute a by subtracting the
filter's group delay from its maximum group delay. For example,
[be,ae] = ellip(4,1,40,0.2);
f = 0:0.001:0.2;
g = grpdelay(be,ae,f,2);
% Equalize only the passband.
a = max(g)-g;
[num,den]=iirgrpdelay(8, f, [0 0.2], a);
11-66
iirgrpdelay
See Also
freqz
filter
grpdelay
iirlpnorm
iirlpnormc
zplane
Frequency response of digital filters.
Filter data with a recursive (IIR) or nonrecursive (FIR)
filter.
Average filter delay (group delay).
Least p-norm optimal IIR filter design.
Constrained least p-norm optimal IIR filter design.
Zero-pole plot.
11-67
iirlpnorm
Purpose
11iirlpnorm
Least p-norm optimal IIR filter design
Syntax
[num,den]
[num,den]
[num,den]
[num,den]
[num,den]
Description
[num,den] = iirlpnorm(n,d,f,edges,a) returns a filter having a numerator
order n and denominator order d which is the best approximation to the desired
frequency response described by f and a in the least-pth sense. The vector
edges specifies the band-edge frequencies for multi-band designs. An
=
=
=
=
=
iirlpnorm(n,d,f,edges,a)
iirlpnorm(n,d,f,edges,a,w)
iirlpnorm(n,d,f,edges,a,w,p)
iirlpnorm(n,d,f,edges,a,w,p,dens)
iirlpnorm(n,d,f,edges,a,w,p,dens,initnum,initden)
unconstrained quasi-Newton algorithm is employed and any poles or zeros that
lie outside of the unit circle are reflected back inside. n and d should be chosen
so that the zeros and poles are used effectively. See the “Hints” section. Always
user freqz to check the resulting filter.
[num,den] = iirlpnorm(n,d,f,edges,a,w) uses the weights in w to weight
the error. w has one entry per frequency point (the same length as f and a)
which tells iirlpnorm how much emphasis to put on minimizing the error in
the vicinity of each frequency point relative to the other points. f and a must
have the same number of elements, which may exceed the number of elements
in edges. This allows for the specification of filters having any gain contour
within each band. The frequencies specified in edges must also appear in the
vector f. For example,
[num,den] = iirlpnorm(5,12,[0 .15 .4 .5 1],[0 .4 .5 1],...
[1 1.6 1 0 0],[1 1 1 10 10])
is a lowpass filter with a peak of 1.6 within the passband.
[num,den] = iirlpnorm(n,d,f,edges,a,w,p) where p is a two-element
vector [pmin pmax] allows for the specification of the minimum and maximum
values of p used in the least-pth algorithm. Default is [2 128] which essentially
yields the L-infinity, or Chebyshev, norm. Pmin and pmax should be even. If p is
the string 'inspect', no optimization will occur. This can be used to inspect
the initial pole/zero placement.
11-68
iirlpnorm
[num,den] = iirlpnorm(n,d,f,edges,a,w,p,dens) specifies the grid density
dens used in the optimization. The number of grid points is (dens*(n+d+1)).
The default is 20. dens can be specified as a single-element cell array. The grid
is not equally spaced.
[num,den] = iirlpnorm(n,d,f,edges,a,w,p,dens,initnum,initden)
allows for the specification of the initial estimate of the filter numerator and
denominator coefficients in vectors initnum and initden. This may be useful
for difficult optimization problems. The pole-zero editor in the Signal
Processing Toolbox can be used for generating initnum and initden.
Hints
• This is a weighted least-pth optimization.
• Check the radii and locations of the poles and zeros for your filter. If the zeros
are on the unit circle and the poles are well inside the unit circle, try
increasing the order of the numerator or reducing the error weighting in the
stopband.
• Similarly, if several poles have a large radii and the zeros are well inside of
the unit circle, try increasing the order of the denominator or reducing the
error weighting in the passband.
See Also
iirlpnormc
filter
freqz
iirgrpdelay
zplane
Constrained least-p norm optimal IIR filter design.
Filter data with a recursive (IIR) or nonrecursive (FIR)
filter.
Frequency response of a filter.
Optimal IIR filter design with prescribed group-delay.
Zero-pole plot.
11-69
iirlpnormc
Purpose
11iirlpnormc
Constrained least P-norm optimal IIR filter design
Syntax
[num,den] = iirlpnormc(n,d,f,edges,a)
[num,den] = iirlpnormc(n,d,f,edges,a,w)
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius)
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius,p)
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius,p,dens)
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius,p,...
dens,initnum,initden)
Description
[num,den] = iirlpnormc(n,d,f,edges,a) returns a filter having a
numerator order n and denominator order d which is the best approximation to
the desired frequency response described by f and a in the least-pth sense. The
vector edges specifies the band-edge frequencies for multi-band designs. A
constrained Newton-type algorithm is employed. n and d should be chosen so
that the zeros and poles are used effectively. See the “Hints” section. Always
check the resulting filter using freqz.
[num,den] = iirlpnormc(n,d,f,edges,a,w) uses the weights in w to weight
the error. w has one entry per frequency point (the same length as f and a)
which tells iirlpnormc how much emphasis to put on minimizing the error in
the vicinity of each frequency point relative to the other points. f and a must
have the same number of elements, which can exceed the number of elements
in edges. This allows for the specification of filters having any gain contour
within each band. The frequencies specified in edges must also appear in the
vector f. For example,
[num,den] = iirlpnormc(5,12,[0 .15 .4 .5 1],[0 .4 .5 1],...
[1 1.6 1 0 0],[1 1 1 10 10])
designs a lowpass filter with a peak of 1.6 within the passband.
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius) returns a filter having
a maximum pole radius of radius where 0<radius<1. radius defaults to
0.999999. Filters having a reduced pole radius may retain better transfer
function accuracy after you quantize them.
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius,p) where p is a
two-element vector [pmin pmax] allows for the specification of the minimum
11-70
iirlpnormc
and maximum values of p used in the least-pth algorithm. Default is [2 128]
which essentially yields the L-infinity, or Chebyshev, norm. pmin and pmax
should be even. If p is the string 'inspect', no optimization will occur. This can
be used to inspect the initial pole/zero placement.
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius,p,dens) specifies the
grid density dens used in the optimization. The number of grid points is
(dens*(n+d+1)). The default is 20. dens can be specified as a single-element
cell array. The grid is not equally spaced.
[num,den] = iirlpnormc(n,d,f,edges,a,w,radius,p,dens,...
initnum,initden) allows for the specification of the initial estimate of the
filter numerator and denominator coefficients in vectors initnum and initden.
This may be useful for difficult optimization problems. The pole-zero editor in
the Signal Processing Toolbox can be used for generating initnum and initden.
Hints
• This is a weighted least-pth optimization.
• Check the radii and location of the resulting poles and zeros.
• If the zeros are all on the unit circle and the poles are well inside of the unit
circle, try increasing the order of the numerator or reducing the error
weighting in the stopband.
• Similarly, if several poles have a large radius and the zeros are well inside of
the unit circle, try increasing the order of the denominator or reducing the
error weight in the passband.
• If you reduce the pole radius, it may be necessary to increase the order of the
denominator.
See Also
freqz
filter
iirgrpdelay
iirlpnorm
zplane
Frequency response of a filter.
Filter data with a recursive (IIR) or nonrecursive (FIR)
filter
IIR filter design with prescribed group-delay.
Least p-norm optimal filter design.
Pole-zero plot.
11-71
iirpowcomp
Purpose
11iirpowcomp
Compute power complementary filter.
Syntax
[bp,ap] = iirpowcomp(b,a)
[bp,ap,c] = iirpowcomp(b,a)
Description
[bp,ap] = iirpowcomp(b,a) returns the coefficients of the power
complementary IIR filter g(z) = bp(z)/ap(z) in vectors bp and ap, given the
coefficients of the IIR filter h(z) = b(z)/a(z) in vectors b and a. b must be
symmetric (Hermitian) or antisymmetric (antihermitian) and of the same
length as a. The two power complementary filters satisfy the relation
|H(w)|2 + |G(w)|2 = 1.
[bp,ap,c] = iirpowcomp(b,a) where c is a complex scalar of magnitude =1,
forces bp to satisfy the generalized hermitian property
conj(bp(end:-1:1)) = c*bp.
When c is omitted, it is chosen as follows:
• When b is real, chooses C as 1 or -1, whichever yields bp real
• When b is complex, C defaults to 1
ap is always equal to a.
Examples
See Also
[b,a]=cheby1(10,.5,.4);
[bp,ap]=iirpowcomp(b,a);
[h,w,s]=freqz(b,a); [h1,w,s]=freqz(bp,ap);
s.plot='mag'; s.yunits='sq';freqzplot([h h1],w,s)
tf2ca
tf2cl
ca2tf
cl2tf
11-72
Convert transfer function form to coupled allpass
Convert transfer function form to coupled allpass
lattice
Convert coupled allpass form to transfer function form
Convert coupled lattice form to transfer function form
impz
Purpose
11impz
Compute the impulse response of quantized filters
Syntax
[h,t] = impz(Hq)
[h,t] = impz(Hq,n)
[h,t] = impz(Hq,n,Fs)
[h,t,ref] = impz(Hq,...)
impz(Hq,...)
Description
[h,t] = impz(Hq) computes the response of the quantized filter Hq to an
impulse. impz returns the computed impulse response in the column vector h
and the corresponding sample times in the column vector t (where
t = [0:n-1]' and n = length(t) is computed automatically).
[h,t] = impz(Hq,n) computes n samples of the quantized impulse response
for any positive integer n. In this case, t = [0:n-1]'. When n is a vector of
integers, impz computes the impulse response at those integer locations,
starting the response computation from 0 (and t=n or t=[0 n]). If, instead of n,
you include the empty vector [] as the second argument, the number of
samples is computed automatically.
[h,t] = impz(Hq,n,Fs) computes n samples and produces a vector t of
length n so that the samples are spaced 1/Fs units apart.
[h,t,ref] = impz(Hq,...) returns the impulse response of the quantized
filter Hq in the column vector h, and returns the impulse response of the
reference filter in the vector ref.
impz(Hq,...) with no output arguments plots the impulse response of the
reference filter associated with Hq, and the quantized impulse response of
quantized filter Hq in a new figure window. impz uses stem for plotting the
impulse responses.
Note impz works for both real and complex quantized filters. When you omit
the output arguments, only the real part of the impulse response is plotted.
11-73
impz
Examples
Create a quantized filter for a fourth-order, low-pass elliptic filter with a cutoff
frequency of 0.4 times the Nyquist frequency. Use a second-order sections
structure to resist quantization errors. Plot the first 50 samples of the
quantized impulse response, along with the reference impulse response.
% Specify transfer function parameters for the reference filter.
[b,a] = ellip(4,3,20,.6);
% Create a quantized filter from the reference filter. Convert the
quantized filter to second-order section form, order, and scale.
Hq = sos(qfilt('ref',{b,a}));
Warning: 1 overflow in coefficients.
impz(Hq,50)
0.5
Quantized response
Reference response
0.4
0.3
0.2
0.1
0
−0.1
−0.2
−0.3
11-74
0
5
10
15
20
25
30
35
40
45
impz
Algorithm
impz applied to the quantized filter Hq applies the filter command twice to a
length n impulse sequence:
• Once for a quantized filter whose coefficients are determined by the
ReferenceCoefficients property value for Hq
• Once for a quantized filter whose coefficients are determined by the
QuantizedCoefficients property value for Hq
The resulting plots use stem.
Warnings that occur with impz are a result of the filter command. In
particular, you get an input overflow warning with impz when the InputFormat
property value for the quantized filter Hq is fixed-point and has only one bit to
the left of the radix point.
For example, when your InputFormat property is set to {'fixed',[16,15]},
you get an input overflow warning when you implement impz.
See Also
filter
Filter data through a quantized filter.
11-75
isallpass
Purpose
11isallpass
Test quantized filters to determine if they are allpass structures
Syntax
flag = isallpass(f)
flag = isallpass(f,k)
Description
flag = isallpass(f) determines if the filter object f has allpass filter
sections, returning 1 if true and 0 if false.
flag = isallpass(f,k) determines if the k-th section of the filter object f is
an allpass section and returns 1 if true and 0 if false.
Since lattice coupled allpass filters always have allpass sections, this function
always returns one for these filters.
See Also
isfir
islinphase
ismaxphase
isminphase
isreal
issos
isstable
11-76
Test quantized filters to see if they are FIR filters.
Test quantized filters to see if they are linear phase.
Test quantized filters to see if they are maximum phase
filters.
Test quantized filters to see if they are minimum
phase.
Test quantized filters for purely real coefficients.
Test whether quantized filters are in second-order
section form.
Test for stability of a quantized filter.
isfir
Purpose
11isfir
Test quantized filters to see if they are FIR filters
Syntax
flag = isfir(hq)
flag = isfir(hq,k)
Description
flag = isfir(hq) determines whether quantized filter hq is an FIR filter,
returning flag equal to one when the quantized filter is an FIR filter, and zero
when it is IIR.
flag = isfir(hq,k) determines if the k-th section of the quantized filter hq is
an FIR filter, returning flag equal to 1when the k-th section is an FIR filter
and zero when it is IIR.
isfir(hq) looks at filter hq and determines whether the filter, in transfer
function form, has a scalar for its denominator. If it does, it is an FIR filter.
Examples
hq = qfilt;
isfir(hq)
returns 1 for the status of filter hq; the filter is an FIR structure with
denominator reference coefficient equal to one.
See Also
isallpass
islinphase
ismaxphase
isminphase
isreal
issos
isstable
Test quantized filters to determine if they are allpass
structures.
Test quantized filters to see if they are linear phase.
Test quantized filters to see if they are maximum phase
filters.
Test quantized filters to see if they are minimum
phase.
Test quantized filters for purely real coefficients.
Test whether quantized filters are in second-order
section form.
Test for stability of a quantized filter.
11-77
islinphase
Purpose
11islinphase
Test quantized filters to see if they have linear phase sections (one or more)
Syntax
flag = islinphase(hq)
flag = islinphase(hq,k)
Description
flag = islinphase(hq) determines if the quantized filter hq has linear phase
sections, and returns 1 if true and 0 if false.
flag = islinphase(hq,k) determines if the k-th section of the filter hq is a
linear phase section and returns 1 if true and 0 if false.
The determination is based on the reference coefficients. A filter has linear
phase if it is FIR and its transfer function coefficients are are symmetric or
antisymmetric, or if it is IIR and it has poles on or outside the unit circle and
both numerator and denominator are symmetric or antisymmetric.
Examples
This filter has linear phase.
num=[1 0 0 0 0 -1];
den=[1 -1];
hq = qfilt('df2t',{num,den});
islinphase(hq)
See Also
isallpass
isfir
ismaxphase
isminphase
isreal
issos
isstable
11-78
Test quantized filters to determine if they are allpass
structures.
Test quantized filters to see if they are FIR filters.
Test quantized filters to see if they are maximum phase
filters.
Test quantized filters to see if they are minimum
phase.
Test quantized filters for purely real coefficients.
Test whether quantized filters are in second-order
section form.
Test for stability of a quantized filter.
ismaxphase
Purpose
11ismaxphase
Test quantized filters to see if they are maximum phase filters
Syntax
flag = ismaxphase(hq)
flag = ismaxphase(hq,k)
Description
flag = ismaxphase(hq) determines if the filter hq has maximum phase
sections, and returns 1 if true and 0 if false.
flag = ismaxphase(hq,k) determines if the k-th section of the filter hq is a
maximum phase section and returns 1 if true and 0 if false.
The determination is based on the reference coefficients. A filter is maximum
phase when all the zeros of its transfer function are on or outside the unit
circle, or when the numerator is a scalar.
Examples
hq = qfilt;
ismaxphase(hq)
returns 1 so this is a maximum phase quantized filter.
See Also
isallpass
isfir
islinphase
isminphase
isreal
issos
isstable
Test quantized filters to determine if they are allpass
structures.
Test quantized filters to see if they are FIR filters.
Test quantized filters to see if they are linear phase.
Test quantized filters to see if they are minimum
phase.
Test quantized filters for purely real coefficients.
Test whether quantized filters are in second-order
section form.
Test for stability of a quantized filter.
11-79
isminphase
Purpose
11isminphase
Test quantized filters to see if they are minimum phase
Syntax
flag = isminphase(hq)
flag = isminphase(hq,k)
Description
flag = isminphase(hq) determines if the filter hq has minimum phase
sections, and returns 1 if true and 0 if false.
flag = isminphase(hq,k) determines if the k-th section of the filter hq is a
minimum phase section and returns 1 if true and 0 if false.
The determination is based on the reference coefficients. A filter is minimum
phase when all the zeros of its transfer function are on or inside the unit circle,
or the numerator is a scalar.
Examples
This example create a minimum phase quantized filter.
hq = qfilt;
isminphase(hq)
See Also
isallpass
isfir
islinphase
ismaxphase
isreal
issos
isstable
11-80
Test quantized filters to determine if they are allpass
structures.
Test quantized filters to see if they are FIR filters.
Test quantized filters to see if they are linear phase.
Test quantized filters to see if they are maximum phase
filters.
Test quantized filters for purely real coefficients.
Test whether quantized filters are in second-order
section form.
Test for stability of a quantized filter.
isreal
Purpose
11isreal
Test quantized filters for purely real coefficients
Syntax
r = isreal(hq)
Description
r = isreal(hq) returns the value 1 (equal to true) if all reference filter
coefficients for the quantized filter hq are real, and returns the value 0 (or false)
otherwise.
isreal(hq) returns 1 if all filter coefficients in quantized filter hq have zero
imaginary part. Otherwise, isreal(hq) returns a 0 indicating that the filter
is complex. That is, it detects complex filters, meaning quantized filters that
have one or more coefficients with nonzero imaginary parts.
Note Quantizing a filter cannot make a real filter into a complex filter.
Examples
% Create a reference filter.
[b,a] = ellip(2,0.5,20,0.4);
% Create a quantized filter from the reference filter.
hq = qfilt('df2t',{b,a});
% Test if all filter coefficients are real.
r = isreal(hq)
r =
1
11-81
isreal
See Also
isfir
islinphase
ismaxphase
isminphase
issos
isstable
isallpass
11-82
Test quantized filters to see if they are FIR filters.
Test quantized filters to see if they are linear phase.
Test quantized filters to see if they are maximum phase
filters.
Test quantized filters to see if they are minimum
phase.
Test whether quantized filters are in second-order
section form.
Test for stability of a quantized filter.
Test quantized filters to determine if they are allpass
structures.
issos
Purpose
11issos
Test whether quantized filters are composed of second-order sections
Syntax
flag = issos(hq)
Description
flag = issos(hq) determines whether quantized filter hq consists of
second-order sections. Returns 1 if all sections of quantized filter hq have order
less than or equal to two, and 0 otherwise.
Examples
warning off
[b,a] = butter(5,.5);
hq = sos(qfilt('ref',{b,a}));
v = issos(hq)
v =
1
hq.statespersection
ans =
1
2
2
Quantized filter hq is in second-order section form.
See Also
isallpass
isfir
islinphase
ismaxphase
isminphase
isreal
isstable
Test quantized filters to determine if they are allpass
structures.
Test quantized filters to see if they are FIR filters.
Test quantized filters to see if they are linear phase.
Test quantized filters to see if they are maximum phase
filters.
Test quantized filters to see if they are minimum
phase.
Test quantized filters for purely real coefficients.
Test for stability of a quantized filter.
11-83
isstable
Purpose
11isstable
Test for stability of a quantized filter
Syntax
isstable(hq)
isstable(hq,k)
Description
isstable(hq) tests quantized filter hq to determine whether its poles are
inside the unit circle. If the poles lie on or outside the circle, isstable returns
0. If the poles are inside the circle, isstable returns a 1.
isstable(hq,k) returns the stability of the k-th section of a multiple section
quantized filter. Based on the locations of the poles of the specified section,
isstable returns 1 if the filter section is stable, and 0 otherwise.
To determine the filter stability, isstable checks the quantized filter
coefficients.
Examples
See Also
Hq = qfilt;
isstable(Hq)
ans =
1
isallpass
isfir
islinphase
ismaxphase
isminphase
isreal
issos
zplane
11-84
Test quantized filters to determine if they are allpass
structures.
Test quantized filters to see if they are FIR filters.
Test quantized filters to see if they are linear phase.
Test quantized filters to see if they are maximum phase
filters.
Test quantized filters to see if they are minimum
phase.
Test quantized filters for purely real coefficients.
Test whether quantized filters are in second-order
section form.
Pole-zero plot.
length
Purpose
11length
Return the length of a quantized FFT
Syntax
length(f)
Description
length(f) returns the value of the length property of quantized FFT f. The
value of the length property must be a positive integer that is also a power of
the radix of the quantized FFT (f.radix). The length of the FFT is the length
of the data vector that the FFT operates on.
Examples
F = qfft;
length(F)
returns the default 16 for the length of the FFT.
See Also
qfft
get
set
Construct a quantized FFT.
Return the properties of a quantized filter, quantizer,
or quantized FFT.
Set or display the properties of a quantized filter,
quantizer, or quantized FFT.
11-85
limitcycle
Purpose
11limitcycle
Detect limit cycles in a quantized filter
Syntax
limitcycle(Hq)
limitcycle(Hq, Ntrials, InputLength, StopCriterion, DisplayType)
[LimitcycleType, Zi, StatePeriod, StateSequence, OverflowsPerStep,
Trial, Section] = limitcycle(Hq, ...)
Description
limitcycle(Hq) runs 20 Monte Carlo trials with quantized filter Hq. Each trial
uses a new set of initial states (determined randomly) and zero input vector of
length 100. Monte Carlo processing stops if a zero-input limit cycle is detected
in quantized filter Hq. At completion, limitcycle returns one of the following
strings:
• 'granular' indicating that a granular overflow occurred
• 'overflow' indicating that an overflow limitcycle occurred
• 'none' indicating that no limit cycles were detected during the Monte Carlo
trials
limitcycle(Hq, Ntrials, InputLength, StopCriterion, DisplayType) lets
you set the following arguments:
• Ntrials, the number of Monte Carlo trials (default is 20).
• InputLength, the length of the zero vector used as input to the filter (default
is 100).
• StopCriterion, the criterion for stopping the Monte Carlo trials processing.
StopCriterion can be set to 'either' (the default), 'granular',
'overflow', or 'none'. If StopCriterion is:
- 'either' — the Monte Carlo trials will stop when either a granular or
overflow limit cycle is detected.
- 'granular' — the Monte Carlo trials stop when a granular limit cycle was
detected.
- 'overflow' — the Monte Carlo trials stop when an overflow limit cycle
was detected.
- 'none' — the Monte Carlo trials do not stop until all trials have been run.
• DisplayType, the display type. When DisplayType is nonzero, limitcycle
displays messages about the progress of the Monte Carlo trials.
11-86
limitcycle
[LimitcycleType, Zi, StatePeriod, StateSequence, OverflowsPerStep,
Trial, Section] = limitcycle(Hq,...) also returns
• LimitcycleType — one of 'granular' to indicate that a granular overflow
occurred; 'overflow' to indicate that an overflow limitcycle occurred; or
'none' to indicate that no limit cycles were detected during the Monte Carlo
trials.
• Zi — the initial condition that caused the limit cycle.
• StatePeriod — an integer indicating the repeat period of the limit cycle (-1
if the filter converged and the last state is zero, 0 if the last state is not zero
and no limit cycle was detected).
• StateSequence — a matrix containing the sequence of states at every time
step (one column per time step). The final conditions are in the last column
of StateSequence Zf = StateSequence(:,end). The initial conditions of the
section are in the first column of StateSequence Zi = StateSequence(:,1).
• OverflowsPerStep — a cell array that contains one vector of integers for
each section of the filter that indicates the total number of overflows that
occurred during each time step. The overflows from the Kth section are found
in OverflowsPerStep{K}.
• Trial — the number of the trial on which Monte Carlo processing stopped.
• Section — the number of the section in which the limitcycle was detected.
Only the parameters of the last limit cycle are returned. If Monte Carlo
processing does not detect any limit cycles, the parameters of the last Monte
Carlo trial are returned.
Examples
In this example, there is a region of initial conditions in which no limit cycles
occur, and a region where they do. If no limit cycles are detected before the
Monte Carlo trials are over, the state sequence spirals to zero. When a limit
cycle is found, the states do not end at zero. Each time you run this example, it
uses a different sequence of random initial conditions, so the plot you get may
differ from the one displayed in the following figure.
A = [-1 -1; 0.5 0];
B = [0; 1];
C = [1 0];
D = 0;
Hq = qfilt('statespace',{A,B,C,D},'OverFlowMode','Wrap');
[LimitCycleType, Zi, StatePeriod, StateSequence] = limitcycle(Hq);
plot(StateSequence(1,:), StateSequence(2,:),'-o')
11-87
limitcycle
xlabel('State
ylabel('State
axis([-2 2 -2
title(['Limit
See Also
11-88
freqz
nlm
1');
2');
2]); axis square; grid
cycle type:',LimitCycleType])
Compute the frequency response of quantized filters.
Use the noise loading method to estimate the frequency
response of a quantized filter.
max
Purpose
11max
Return the maximum value of a quantizer object before quantization
Syntax
max(q)
Description
max(q) is the maximum value before quantization during a call to
quantize(q,...) for quantizer q. This value is the maximum value
encountered over successive calls to quantize and is reset with reset(q).
max(q) is equivalent to get(q,'max') and q.max.
Examples
q = quantizer;
warning on
y = quantize(q,-20:10);
max(q)
returns the value 10 and a warning for 29 overflows.
See Also
min
Return the minimum value of a quantizer before
quantization.
11-89
min
11min
Purpose
Return the minimum value of a quantizer object before quantization
Syntax
max(q)
Description
max(q) is the minimum value before quantization during a call to
quantize(q,...) for quantizer q. This value is the minimum value
encountered over successive calls to quantize and is reset with reset(q).
min(q) is equivalent to get(q,'min') and q.min.
Examples
q = quantizer;
warning on
y = quantize(q,-20:10);
min(q)
returns the value -20 and a warning for 29 overflows.
See Also
11-90
max
Return the maximum value of a quantizer before
quantization.
nlm
Purpose
11nlm
Use the noise loading method to estimate the frequency response of a quantized
filter
Syntax
[H,W,Pnn,Nf] = nlm(Hq,N,L)
[H,W,Pnn,Nf] = nlm(Hq,n,l,'whole')
[H,F,...] = nlm(Hq,N,L,Fs)
[H,F,...] = nlm(Hq,N,L,'whole',Fs)
nlm(Hq,...)
Description
[H,W,Pnn,Nf] = nlm(Hq,N,L) uses the noise loading method to estimate the
complex frequency response of quantized filter Hq. nlm returns complex
frequency response H, frequency vector W, in radians/sample, power spectral
density Pnn, and noise figure Nf, for the quantized filter Hq, at N equally-spaced
points around the upper half of the unit circle. Noise figure Nf and power
spectral density Pnn are given in dB. nlm averages over L Monte Carlo trials.
The Monte Carlo trials result in a noise-like signal that contains complete
frequency content across the spectrum. When N or L is omitted from the
command, or empty, N defaults to 512 and L defaults to 10.
[H,W,Pnn,Nf] = nlm(Hq,n,l,'whole') uses N points around the entire unit
circle, rather than the upper half.
[H,F,...] = nlm(Hq,N,L,Fs) and [H,F,...] = nlm(Hq,N,L,'whole',Fs)
return frequency vector F, in Hz, where Fs is the sampling frequency in Hz.
nlm(Hq,...) without output arguments plots the magnitude and unwrapped
phase of Hq, comparing the estimated response to the theoretical frequency
response calculated by [h,w] = freqz(Hq,n) in the current figure window.
Examples
Use the noise loading method to determine the frequency response of a
quantized filter Hq.
[b,a] = butter(6, 0.5);
Hq = qfilt('df2t',{b,a});
nlm(Hq,1024,20)
11-91
nlm
Noise Loading Method. Noise figure = 3.6634 dB
100
0
−100
−200
Quantized NLM
Quantized FREQZ
Reference FREQZ
Noise Power Spectrum
−300
−400
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Normalized Frequency (×π rad/sample)
0.8
0.9
1
0
−500
−1000
−1500
−2000
See Also
freqz
qfilt
References
11-92
Compute the theoretical frequency response of a
quantized filter.
Construct a quantized filter.
McClellan, et al., Computer-Based Exercises for Signal Processing Using
MATLAB 5, Prentice-Hall, 1998, 243.
noperations
Purpose
11noperations
Number of quantization operations performed by a quantizer, quantized filter,
or quantized FFT
Syntax
noperations(q)
noperations(hq)
noperations(f)
Description
noperations(q) is the number of product and sum quantization operations
during a call to quantize(q,...) for quantizer q. This value accumulates over
successive calls to quantize. You reset the value of noperations by issuing the
command reset(q).
noperations(hq) is the number of quantization operations during a call to
filter(hq,...) for quantized filter hq.
noperations(f) is the number of quantization operations during a call to
fft(f,...) or ifft(f,...) for quantized FFT f.
Examples
Noperations returns the number of quantizations it performs. You call it as a
quantizer, or as part of designing a quantized filter or quantized FFT.
noperations reflects the number of sum and product quantizations for
quantized filters and quantized FFTs. For quantizers, noperations reflects all
the quantization operations, not just the sum and product quantizations.
The following code does not perform any adds or multiplies; it quantizes the
specified data according to the properties of quantizer q:
warning on
q=quantizer;
y = quantize(q,-20:10);
noperations(q)
returns 31 and a warning for 29 overflows. Notice that the next example
returns an operations count (NOperations) that includes only quantizations
performed during multiply and add operations.
[b,a] = ellip(4,3,20,.6);
hq = qfilt('df2',{b,a},'roundmode','fix');
y=filter(hq,randn(100,1));
noperations(hq)
11-93
noperations
Warning: 42 overflows in QFILT/FILTER.
Max
Min
NOverflows
NUnderflows
Coefficient
1.398
0.2259
2
0
Input
2.112
-2.202
37
0
Output
0.8863
-0.9275
0
0
Multiplicand
2
-1.956
383
0
Product
1
-1
0
0
Sum
2.389
-1.956
3
0
See Also
get
qfft
qfilt
quantizer
11-94
NOperations
10
100
100
1200
1200
1000
Return the property values for a quantizer, quantized
filter, or quantized FFT.
Construct a quantized FFT.
Construct a quantized filter.
Construct a quantizer.
normalize
Purpose
11normalize
Normalize quantized filter coefficients
Syntax
h = normalize(Hq)
Description
h = normalize(Hq) accounts for quantized filter coefficient overflow by
normalizing the quantized filter coefficients in the quantized filter Hq. The new
quantized filter h contains the normalized coefficients. All quantized filter
coefficients for h stored in the QuantizedCoefficients property value are
modified to have magnitude less than or equal to one. The result also modifies
the ReferenceCoefficients property value for h accordingly. normalize also
modifies the ScaleValues property value for h from that of Hq, so that input
data to each section of h are scaled to compensate for the normalized filter
coefficients. The scaling factors used in normalize are powers of two. There
may be a different scaling factor for each section of the quantized filter. You can
apply normalize to direct form IIR and FIR filters only. To apply normalize to
a quantized filter, its property Hq.FilterStructure must be one of the
following strings:
• 'df1'
• 'df1t'
• 'df2'
• 'df2t'
• 'fir'
• 'firt'
• ‘antisymmetricfir’
• 'symmetricfir'
Examples
Create a direct form II transposed quantized filter and use normalize to
account for overflow.
% Create a low pass reference filter in the Signal Proc. Toolbox.
[b,a] = ellip(5,2,40,0.4);
% Create the quantized filter from the reference.
hq = qfilt('df2t',{b,a});
Warning: 5 overflows in coefficients.
11-95
normalize
You are warned that some of the coefficients have overflowed. To account for
this overflow, use normalize to modify the ReferenceCoefficients,
QuantizedCoefficients, and ScaleValues property values for Hq.
hq = normalize(hq)
hq =
Quantized Direct form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
(1)
0.365295410156250 0.365289835338219130
(2)
0.395721435546875 0.395708380608267300
(3)
0.724884033203125 0.724891008581378560
(4)
0.724884033203125 0.724891008581378120
(5)
0.395721435546875 0.395708380608267240
(6)
0.365295410156250 0.365289835338218350
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
(1)
0.250000000000000
0.250000000000000000
(2)
-0.541015625000000 -0.541012429707579350
(3)
0.790557861328125
0.790542752251058410
(4)
-0.668945312500000 -0.668930473694134720
(5)
0.365966796875000
0.365965902328318770
(6)
-0.103698730468750 -0.103697674644671510
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
=
=
=
=
=
=
=
=
=
=
df2t
[0.03125 1]
1
[5]
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
Notice that none of the coefficients overflow, and that the ScaleValues
property value has changed.
See Also
11-96
get
set
Retrieve property values for quantized filters.
Set property values for quantized filters.
15])
15])
15])
15])
30])
30])
noverflows
Purpose
11noverflows
Return the number of overflows from the most recent FFT or IFFT operation
Syntax
noverflows(F)
noverflows(Hq)
noverflows(Hq,'sum')
noverflows (Q)
Description
noverflows(F) returns the number of overflows resulting from the most recent
fft or ifft operation that used quantized fft (F).
noverflows(Hq) returns the number of overflows resulting from the most
recent filter operation that used quantized filter (Hq).
noverflows(Hq,'sum') returns the number of overflows that resulted from the
most recent qfilt operation. When the quantized filter has one section, this
returns a scalar. When the filter uses two or more sections, noverflows returns
a vector containing one element for each filter section.
noverflows(Q) returns the number of overflows resulting from the most recent
quantize operation that used quantizer (Q).
Examples
Create a quantized fft f and apply it to a data set. Check the number of
overflows that result when you use f. Then apply f to a second data set and
check the overflows again.
warning on
n=128;
t = (1:n)/n;
x = sin(2*pi*10*t)/10;
f = qfft('length',n);
plot(t,abs([fft(f,x);fft(x)]))
noverflows(f)
returns 24 for the number of overflows and a warning of 24 overflows.
Now, apply f to another data set.
x = sin(2*pi*10*t)/5;
plot(t,abs([fft(f,x);fft(x)]))
noverflows(f)
Now you see 58 overflows.
11-97
noverflows
See Also
get
max
range
reset
11-98
Return the property values for quantized filters,
quantized FFTs, and quantizers.
Return the maximum value of a quantizer object before
quantization.
Return the numerical range of the quantizer in a
quantized filter, quantized FFT, or quantizer.
Reset one or more quantizers to their initial conditions.
num2bin
Purpose
11num2bin
Convert a number to a binary string
Syntax
num2bin(Hq)
c = num2bin(Hq)
y = num2bin(q,x)
Description
num2bin(Hq) with no left-hand-side argument displays the quantized
coefficients in quantized filter Hq as binary strings.
c = num2bin(Hq) with a left-hand-side argument c returns a cell array of
quantized coefficients as binary strings. Cell array c inherits the configuration
of cell array Hq.QuantizedCoefficients.
When the mode of Hq is float, double, or single, the coefficients are converted to
IEEE Standard 754 style binary strings.
If the mode of Hq is fixed, the coefficients are converted to two’s complement
binary strings.
y = num2bin(q,x) converts numeric array x into binary strings returned in y.
When x is a cell array, each numeric element of x is converted to binary. If x is
a structure, each numeric field of x is converted to binary.
num2bin and bin2num are inverses of one another, differing in that num2bin
returns the binary strings in a column.
Examples
x=magic(3)/9
x =
0.8889
0.3333
0.4444
0.1111
0.5556
1.0000
0.6667
0.7778
0.2222
q=quantizer([4 3]);
y = num2bin(q,x)
y =
0111
0010
0011
0000
11-99
num2bin
0100
0111
0101
0110
0001
Algorithm
Numeric values in the input data are quantized first by quantizer q, then
converted to their binary equivalents. When Hq has coefficients exactly equal
to 1, or when the input data set x includes values equal to 1 and 1 is outside the
quantizer’s range, 1 is quantized according to the property values set for q
because no binary representation for 1 exists. Beware of this behavior when
q.overflowmode = 'wrap', because the value 1 in the input data or quantized
filter coefficients gets converted and wrapped to –1 (1000 binary).
For example,
q = quantizer([4 3],'wrap');
range (q)
ans =
-1.0000
0.8750
num2bin(q,1)
returns the binary 10002 = –110 because 1 lies outside the maximum value
(0.8750) for q.
Errors
When one or more quantized coefficients have real or imaginary parts that
equal 1, and the number format does not include 1 in its range, those
coefficients are saturated to 1 – ε (where ε is the epsilon of the coefficient
quantizer) and the operation returns a warning message.
See Also
bin2num
hex2num
num2hex
11-100
Convert a two’s complement binary string to a number.
Convert an IEEE hexadecimal number to a
double-precision number.
Convert a number to its hexadecimal equivalent.
num2hex
Purpose
11num2hex
Convert a number to its hexadecimal equivalent
Syntax
num2hex(Hq)
c = num2hex(Hq)
y = num2hex(q,x)
Description
num2hex(Hq) with no left-hand-side argument displays the quantized
coefficients in quantized filter Hq as hexadecimal strings.
c = num2hex(Hq) with a left-hand-side argument c returns a cell array of
quantized coefficients as hexadecimal strings. Cell array c inherits the
configuration of cell array Hq.QuantizedCoefficients.
When the mode of Hq is 'float', 'double', or 'single', the coefficients are
converted to IEEE Standard 754 style hexadecimal strings.
If the mode of Hq is fixed, the coefficients are converted to two’s complement
hexadecimal strings.
y = num2hex(q,x) converts numeric array x into hexadecimal strings returned
in y. When x is a cell array, each numeric element of x is converted to
hexadecimal. If x is a structure, each numeric field of x is converted to
hexadecimal.
For fixed-point quantizers, the representation is two’s complement. For
floating-point quantizers, the representation is IEEE Standard 754 style.
For example, for q = quantizer('double')
num2hex(q,nan)
ans =
fff8000000000000
The leading fraction bit is 1, all other fraction bits are 0. Sign bit is 1, exponent
bits are all 1.
num2hex(q,inf)
ans =
7ff0000000000000
11-101
num2hex
Sign bit is 0, exponent bits are all 1, all fraction bits are 0.
num2hex(q,-inf)
ans =
fff0000000000000
Sign bit is 1, exponent bits are all 1, all fraction bits are 0.
num2hex and hex2num are inverses of each other, except that num2hex returns
the hexadecimal strings in a column.
Examples
This is a floating-point example using a quantizer q that has 6-bit word length
and 3-bit exponent length.
x=magic(3)
x =
8
1
3
5
4
9
6
7
2
q=quantizer('float',[6 3]);
y = num2hex(q,x)
y =
0
8
0
8
8
8
0
8
0
Algorithm
11-102
Call the num2hex method of the coefficient’s quantizer. The numeric values are
quantized first by q; if you have coefficients that are exactly equal to 1, and 1
is not representable in the arithmetic format, no binary representation for 1
num2hex
will exist, and 1 is quantized according to q. Beware of this when
q.overflowmode = 'wrap', because 1 will be quantized to –1.
For example,
q = quantizer([4 3],'wrap');
num2hex(q, 1)
returns the hexadecimal 816 = –110.
Errors
If one or more quantized coefficients has a real or imaginary part that is exactly
equal to 1, and 1 is outside the range for the quantizer, those coefficients are
saturated to 1 – ε (where ε is the epsilon of the coefficient quantizer) and the
operation returns a warning message.
See Also
bin2num
hex2num
num2bin
Convert a two’s complement binary string to a number.
Convert an IEEE hexadecimal number to a
double-precision number.
Convert a number to a binary string.
11-103
num2int
Purpose
11num2int
Convert number to signed integer
Syntax
y = num2int(q,x)
y = num2int(hq)
y = num2int(q,c)
[y1,y2,…] = num2int(q,x1,x2,…)
Description
y = num2int(q,x) uses q.format to convert numeric x to an integer.
y = num2int(hq) uses q.coefficientformat to convert the coefficients of
quantized filter hq to integers. This function is equivalent to
y = num2int(hq.coefficientquantizer,hq.quantizedcoefficients)
y = num2int(q,{c}) uses q.format to convert the entries in cell array c to
integers, returned in cell array y.
[y1,y2,…] = num2int(q,x1,x2,…) uses q.format to convert numeric values
x1, x2, … to integers y1,y2,….
Examples
All of the four-bit, two’s complement, fixed-point numbers in fractional form
are given by
x = [0.875
0.750
0.625
0.500
0.375 -0.125 -0.625
0.250 -0.250 -0.750
0.125 -0.375 -0.875
0 -0.500 -1.000];
q=quantizer([4 3]);
y = num2int(q,x)
11-104
num2int
y =
7
6
5
4
3
2
1
0
-1
-2
-3
-4
-5
-6
-7
-8
For a quantized filter hq
[b,a] = butter(3,.9,'high')
b =
0.0029
-0.0087
0.0087
-0.0029
1.0000
2.3741
1.9294
0.5321
a =
hq = sos(qfilt('referencecoefficients',{b,a}))
hq.format = [4 3]
Warning: 1 overflow in coefficients.
hq =
Quantized Direct form II transposed filter
------- Section 1 ------Numerator
QuantizedCoefficients{1}{1}
ReferenceCoefficients{1}{1}
(1)
0.750
0.741915184087109990
(2)
-0.750 -0.741922736650797670
Denominator
QuantizedCoefficients{1}{2}
ReferenceCoefficients{1}{2}
+ (1)
0.875 0.999969482421875000
(2)
0.750 0.726520355687005010
------- Section 2 ------Numerator
QuantizedCoefficients{2}{1}
ReferenceCoefficients{2}{1}
(1)
0.500
0.500000000000000000
(2)
-1.000 -0.999994910089555320
(3)
0.500
0.499994910141368990
Denominator
QuantizedCoefficients{2}{2}
ReferenceCoefficients{2}{2}
(1)
0.500 0.500000000000000000
11-105
num2int
(2)
(3)
0.875
0.375
0.823776107851993070
0.366169458636399440
FilterStructure = df2t
ScaleValues = [0.00390625
NumberOfSections = 2
StatesPerSection = [1 2]
1
CoefficientFormat = quantizer('fixed',
InputFormat = quantizer('fixed',
OutputFormat = quantizer('fixed',
MultiplicandFormat = quantizer('fixed',
ProductFormat = quantizer('fixed',
SumFormat = quantizer('fixed',
Warning: 1 overflow in coefficients.
1]
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[4
[4
[4
[4
[4
[4
3])
3])
3])
3])
3])
3])
num2int(hq)
hq.QuantizedCoefficients{1}{1} =
6
-6
hq.QuantizedCoefficients{1}{2} =
7
6
hq.QuantizedCoefficients{2}{1} =
4
-8
4
hq.QuantizedCoefficients{2}{2} =
4
Algorithm
7
3
When q is a fixed-point quantizer with f is equal to fractionlength(q), and x
is numeric
y=x*2f.
When q is a floating-point quantizer, y = x. num2int is only meaningful for
fixed-point quantizers.
11-106
num2int
See Also
bin2num
hex2num
num2bin
num2hex
Convert a two’s complement binary string to a number.
Convert a hexadecimal string to a number.
Convert a number to a binary string.
Convert a number to its hexadecimal equivalent.
11-107
numberofsections
Purpose
11numberofsections
Return the number of sections in a quantized filter
Syntax
numberofsections(hq)
Description
numberofsections(hq) returns the number of sections in a quantized filter.
The filter reference coefficients determine the number of sections.
Examples
Create a double-precision filter using the Butterworth method. Convert the
filter to a quantized filter in second-order section form, then use the function
numberofsections to determine the number of sections that make up the filter.
[b,a] = butter(7,.5);
Hq = sos(qfilt('df2t',{b,a}));
numberofsections(Hq)
See Also
get
qfilt
set
sos
11-108
Return the properties for a quantized filter, quantizer,
or quantized FFT.
Construct a quantized filter.
Set or display the properties for a quantized filter,
quantizer, or quantized FFT.
Convert a quantized filter to second-order section form,
order, and scale.
nunderflows
Purpose
11nunderflows
Return the number of underflows from the most recent quantizer operation
Syntax
nunderflows(q)
Description
nunderflows(q) is the number of underflows during a call to quantize(q,...)
for quantizer object q. An underflow is defined as a number that is nonzero
before it is quantized, and zero after it is quantized. The number of underflows
accumulates over successive calls to quantize. Use the function reset(q) to
return nunderflows to zero.
Examples
q = quantizer('fixed','floor',[4 3]);
x = (0:eps(q)/4:2*eps(q))';
y = quantize(q,x);
nunderflows(q)
ans =
3
By looking at x and y, you can see which ones went to zero.
[x,y]
ans =
0
0.0313
0.0625
0.0938
0.1250
0.1563
0.1875
0.2188
0.2500
0
0
0
0
0.1250
0.1250
0.1250
0.1250
0.2500
11-109
nunderflows
See Also
denormalmin
eps
quantize
quantizer
reset
11-110
Return the smallest denormalized quantized number
for a quantizer.
Return the quantized relative accuracy.
Apply a quantizer to data.
Construct a quantizer.
Reset one or more quantizers to their initial values.
optimizeunitygains
Purpose
11optimizeunitygains
Optimize unity gains for a quantized filter
Syntax
optimizeunitygains(hq)
Description
optimizeunitygains(hq) returns the value of the optimizeunitygains
property of quantized filter object hq. The value of the property can be one of
these two strings:
• on — optimize for coefficients whose real or imaginary part is exactly equal
to 1. Even if 1 cannot be represented by the number format specified by the
CoefficientFormat property, skip multiplications by a real or imaginary
part of a coefficient that is equal to 1.
• off — do not optimize for coefficients whose real or imaginary part is exactly
equal to 1. If 1 cannot be represented by the number format specified by the
CoefficientFormat property, then quantize real or imaginary parts of
coefficients that are equal to 1 to the next lower quantization level.
When optimizeunitygains is on, quantizer(hq,'coefficient') returns a
unitquantizer. If optimizeunitygains is off, quantizer(hq,'coefficient')
returns a quantizer.
Examplse
Hq = qfilt;
optimizeunitygains(Hq)
returns the default 'off'.
See Also
qfilt
qfilt/get
quantizer
unitquantizer
Construct a quantized filter.
Return the property values for quantized filters.
Construct a quantizer.
Construct a unitquantizer.
11-111
order
Purpose
11order
Return the filter order of a quantized filter
Syntax
n=order(hq)
n=order(hq,k)
Description
n = order(hq) returns the order n of the quantized filter hq. When hq is a
single-section filter, n is the number of delays required for a minimum
realization of the filter.
When hq has more than one section, n is the number of delays required for a
minimum realization of the overall filter.
n=order(hq,k) returns the order n of the k-th section of quantized filter hq.
Examples
Create a reference filter. Quantize the filter and convert to second-order section
form. Then use order to check the filter order of the second section and the
overall filter.
[b,a] = ellip(4,3,20,.6); % Create the reference filter.
% Quantize the filter and convert to second-order sections.
Hq = sos(qfilt('df2',{b,a},'roundmode','fix'))
n=order(Hq) % Check the order of the overall filter.
n = 4
n=order(Hq,2) % Check the order of the second section, k=2.
n = 2
11-112
qfft
Purpose
Syntax
11qfft
Construct a quantized FFT
f
f
f
f
=
=
=
=
qfft
qfft('propertyname1',propertyvalue1, ...)
qfft(a)
qfft(pn,pv)
f = qfft('quantize’,[14 13])
Description
f = qfft creates a quantized FFT with default property values.
f = qfft('propertyname1',propertyvalue1,...) uses property name/
property value pairs to set the properties of the quantized FFT.
f = qfft(a), where a is a structure whose field names are quantized FFT
property names, sets the properties named in each field name to the values
contained in the structure.
f = qfft(pn,pv) sets the quantized FFT properties specified in the cell array
of strings pn to the corresponding property values in cell array pv.
f = qfft('quantize’,[14 13]) sets all data format properties for the
quantized FFT to the same word length and fraction length.
Refer to “A Quick Guide to Quantized FFT Properties” on page 10-46 for a list
of quantized FFT properties.
Examples
Create a quantized FFT f and apply it to a data set. Plot the result.
warning on
n=128;
t = (1:n)/n;
x = sin(2*pi*10*t)/10;
f = qfft('length',n);
plot(t,abs([fft(f,x);fft(x)]))
11-113
qfft
See Also
fft
ifft
get
set
11-114
Apply a quantized FFT to data.
Apply the inverse FFT to data.
Return the property values for a quantizer, quantized
filter, or quantized FFT.
List or set the property values for a quantizer,
quantized filter, or quantized FFT.
qfilt
Purpose
11qfilt
Construct a quantized filter
Syntax
Hq
Hq
Hq
Hq
Hq
Description
Hq = qfilt creates a quantized filter Hq with default property settings. The
default settings for Hq imply Hq is a fixed-point quantized filter with a
=
=
=
=
=
qfilt
qfilt('Structure',{Coef})
qfilt('prop1',value1,'prop2',value2,...)
qfilt('Structure',{Coef},'prop1',value1,'prop2',value2,...)
qfilt(‘quantizer’,[13, 14])
transposed direct form II filter structure. All of the filter properties, along with
their default values are listed in “Quantized Filter Properties Reference” on
page 10-11.
Hq = qfilt('Structure',{Coef}) creates a quantized filter Hq with all
properties set to default values, except that the filter structure is specified by
the string 'Structure', and the reference filter parameters (the
ReferenceCoefficients property values) are specified in the cell array
{Coef}. The syntax for entering reference coefficients is specified in “Specifying
the Filter’s Reference Coefficients” on page 6-7. 'Structure' can be one of the
strings for the FilterStructure property values listed in the following table.
Table 11-13: Filter Structure Properties
Property Value String
Description
'df1'
Direct form I
'df1t'
Direct form I transposed
'df2'
Direct form II
'df2t'
Direct form II transposed
'fir'
Finite impulse response (FIR)
'firt’
Finite impulse response transposed
'antisymmetricfir’
Direct form antisymmetric FIR, available odd or
even
11-115
qfilt
Table 11-13: Filter Structure Properties (Continued)
Property Value String
Description
'symmetricfir'
Direct form symmetric FIR, available odd or even
'latticear'
Lattice autoregressive (AR)
'latticema'
Lattice moving average (MA)
'latticearma'
Lattice ARMA
'latticeca'
Lattice coupled-allpass
'latticecapc'
Lattice coupled-allpass power complementary
'statespace'
Single-input, single-output state-space
Hq = qfilt('prop1',value1,'prop2',value2,...) creates a quantized
filter Hq with all properties set to the default values, except for those you
specify with the input string arguments 'prop1', 'prop2',..., along with the
corresponding values in value1, value2,.... Filter properties you can set, with
their default values, are listed in “Quantized Filter Properties Reference” on
page 10-11. Any properties that you do not explicitly set when you create the
quantized filter are assigned default values.
You can also use the shortcut
Hq = qfilt('Structure',{Coef},'prop1',value1,'prop2',value2,...)
by first specifying the FilterStructure property value as 'Structure' and the
reference filter parameters (the ReferenceCoefficients property values) in
the cell array {Coef}.
Hq = qfilt('quantizer',[13 14] sets all the data format properties for
quantized filter Hq to the same word length and fraction length.
Examples
Example 1: Quantized Filter with Two Second-Order Sections
From a reference filter, create a fixed-point quantized filter Hq that has two
second-order sections, setting the rounding mode to 'fix' and displaying the
results.
% Create the reference filter transfer function.
11-116
qfilt
[b,a] = ellip(4,3,20,.6);
% Create a quantized filter with 2 second-order sections
% and display the results.
hq = sos(qfilt('df2',{b,a},'roundmode','fix'))
hq =
Quantized Direct form II transposed filter
------- Section 1 ------Numerator
QuantizedCoefficients{1}{1}
ReferenceCoefficients{1}{1}
(1)
0.551605224609375 0.551616219027048720
(2)
0.776458740234375
(3)
0.551605224609375
Denominator
QuantizedCoefficients{1}{2}
(1)
0.999969482421875
(2)
-0.054809570312500
(3)
0.473083496093750
------- Section 2 ------Numerator
QuantizedCoefficients{2}{1}
(1)
0.499969482421875
(2)
0.359802246093750
(3)
0.499969482421875
Denominator
QuantizedCoefficients{2}{2}
(1)
0.999969482421875
(2)
0.588378906250000
(3)
0.957336425781250
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
=
=
=
=
=
=
=
=
=
=
0.776489000631472080
0.551616219027047940
ReferenceCoefficients{1}{2}
0.999969482421875000
-0.054810658312267876
0.473108096805785360
ReferenceCoefficients{2}{1}
0.499984741210937500
0.359832079066733920
0.499984741210938170
ReferenceCoefficients{2}{2}
0.999969482421875000
0.588389482549356520
0.957363508666007170
df2t
[0.5 2 1]
2
[2 2]
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
'fix',
'fix',
'fix',
'fix',
'fix',
'fix',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
15])
15])
15])
30])
30])
Example 2: Quantized Filter from Table of Filter Coefficients
11-117
qfilt
In this example, you create a sixth-order quantized filter from filter coefficients
in a reference table.
Enter the filter coefficients from a table of coefficients. The following
coefficients represent a 6-pole Chebyshev high pass filter, with 0.5% ripple in
the passband and cutoff at 0.25 in normalized frequency.
Numerator:
b=[.0143445 -0.08606701 .2151675 -.28689 -.2151675 0.08606701 0.0143445]
Denominator:
a=[1.0 1.076051 1.662847 1.191062 0.7403085 0.2752156 0.0572225]
Create a quantized filter using the reference coefficients b and a.
hq = qfilt('ref',{b,a})
hq =
Quantized Direct form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
(1)
0.014343261718750
0.014344500000000000
(2)
-0.086059570312500 -0.086067009999999999
(3)
0.215179443359375
0.215167500000000010
(4)
-0.286895751953125 -0.286889999999999980
(5)
0.215179443359375
0.215167500000000010
(6)
-0.086059570312500 -0.086067009999999999
(7)
0.014343261718750
0.014344500000000000
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
+ (1)
0.999969482421875 1.000000000000000000
+ (2)
0.999969482421875 1.076051000000000100
+ (3)
0.999969482421875 1.662847000000000000
+ (4)
0.999969482421875 1.191062000000000100
(5)
0.740295410156250 0.740308500000000040
(6)
0.275207519531250 0.275215600000000000
(7)
0.057220458984375 0.057222500000000003
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
11-118
=
=
=
=
=
=
=
df2t
[1]
1
[6]
quantizer('fixed', 'round', 'saturate', [16
quantizer('fixed', 'floor', 'saturate', [16
quantizer('fixed', 'floor', 'saturate', [16
15])
15])
15])
qfilt
MultiplicandFormat =
ProductFormat =
SumFormat =
Warning: 4 overflows
quantizer('fixed', 'floor', 'saturate', [16
quantizer('fixed', 'floor', 'saturate', [32
quantizer('fixed', 'floor', 'saturate', [32
in coefficients.
15])
30])
30])
Eliminate the overflows by normalizing the coefficients.
hq2 = normalize(hq)
You have a sixth-order, high pass filter with no overflowing coefficients.
Some things to think about when you use coefficients from a table.
• Take care to assign the numerator and denominator values correctly. In your
table, know which coefficients are for the numerator, which for the
denominator.
• Verify that the sign of the denominator coefficients is correct for MATLAB.
• Note whether all coefficients are provided. Some tables omit the first
coefficient for the denominator. If omitted, set the first denominator
coefficient equal to 1.0.
See Also
get
set
setbits
Retrieve property values for quantized filters.
Set property values for quantized filters.
Set data format properties for quantized filters.
11-119
qfilt2tf
Purpose
11qfilt2tf
Convert quantized filters to transfer function form
Syntax
[Bq,Aq,Br,Ar] = qfilt2tf(Hq)
[Cq,Cr] = qfilt2tf(Hq,'sections')
Description
[Bq,Aq,Br,Ar] = qfilt2tf(Hq) converts the quantized filter coefficients
from quantized filter Hq into transfer function form with numerator Bq and
denominator Aq, and the reference coefficients into transfer-function form with
numerator Br and denominator Ar. When quantized filter Hq has more than
one section, all the numerator polynomials are are convolved into the
numerator polynomial of a single transfer function. Similarly, the
denominator polynomials are convolved into a denominator polynomial of a
single transfer function.
[Cq,Cr] = qfilt2tf(Hq,'sections') returns one cell array per section,
where Cq is the transfer function form of the quantized coefficients and Cr is
the transfer function form of the reference coefficients.
Cq = {{Bq1,Aq1},{Bq2,Aq2},...}
Cr = {{Br1,Ar1},{Br2,Ar2},...}
Examples
To demonstrate the conversion, use butter to a create a reference filter in
statespace form. Make a statespace quantized filter from the reference filter
and convert the quantized filter to transfer function form.
[A,B,C,D]=butter(3,.2);
Hq=qfilt('statespace',{A,B,C,D},'mode','double');
[bq,aq]=qfilt2tf(Hq)
bq =
0.0181
0.0543
0.0543
0.0181
-1.7600
1.1829
-0.2781
aq =
1.0000
See Also
11-120
qfilt
Construct a quantized filter.
qreport
Purpose
Syntax
11qreport
Display the results of applying a quantizer, quantized FFT or quantized filter
to data
qreport(obj)
s = qreport (obj)
where obj is one of the following objects:
• Quantizer
• Quantized filter
• Quantized FFT
Description
qreport(obj) displays the minimum (Min), maximum (Max), number of
overflows (NOver), and underflows (NUnder) of the most recent application of
obj to a data set, where obj is a quantized filter or a quantized FFT. Each
section of quantized filter Hq or stage of quantized FFT F is represented by one
line of information in the report.
Setting warning to ON displays this report when a quantized filter or quantized
FFT overflows.
s = qreport(obj) returns a MATLAB structure containing the information.
Also, qreport(s) displays the report for the structure s.
Examples
Display the results of filtering a data set with a quantized filter Hq.
[b,a] = butter(6,.5);
Hq = sos(qfilt('ReferenceCoefficients',{b,a}));
Y = filter(Hq,rand(50,1));
qreport(Hq)
Max
Coefficient
Input
Output
Multiplicand
Product
Min
-5.169e-016
1
-1.11e-016
1
-8.326e-017
0.9501
0.009861
0.9555
0.02808
0.9501
0.0006161
0.394
0.02808
0.9556
0.02808
0.394
-0.001708
0.6424
-0.05511
0.9556
-0.5626
1
NOverflows NUnderflows NOperations
0
1
6
0
1
6
0
1
6
0
0
50
0
0
50
0
0
400
0
0
350
0
0
350
0
0
400
0
0
350
0
0
350
11-121
qreport
Sum
0.09852
0.3212
0.9555
-0.0007188
-0.003827
-0.2523
0
0
0
0
0
0
250
250
250
Display the results of running qfft F on a set of random data.
F = qfft('length',64,'scale',1/64);
Y = fft(F,rand(64,1));
qreport(F)
Max
Coefficient
Input
Output
Multiplicand
Product
Sum
See Also
11-122
disp
get
1
0.9883
0.5364
0.9883
0.2902
0.5364
Min
-1
0.01176
-0.06312
-0.03622
-0.02877
-0.06312
NOverflows
6
0
0
0
0
0
NUnderflows NOperations
5
126
0
64
0
128
0
1536
0
768
0
1920
Display quantized object property values.
Retrieve property values for quantized filters.
quantize
Purpose
11quantize
Apply a quantizer to data
Syntax
y = quantize(q, x)
[y1,y2,...] = quantize(q,x1,x2,...)
Description
y = quantize(q, x) uses the quantizer q to quantize x. When x is a numeric
array, each element of x is quantized. When x is a cell array, each numeric
element of the cell array is quantized. When x is a structure, each numeric field
of x is quantized. Nonnumeric elements or fields of x are left unchanged and
quantize does not issue warnings for nonnumeric values.
[y1,y2,...] = quantize(q,x1,x2,...) is equivalent to
y1 = quantize(q,x1), y2 = quantize(q,x2),...
The quantizer states
'max'
'min'
'noverflows'
'nunderflows'
'noperations'
-
Maximum value before quantizing
Minimum value before quantizing
Number of overflows
Number of underflows
Number of quantization operations
are updated during the call to quantize, and running totals are kept until a call
to reset is made.
Examples
The following examples demonstrate using quantize to quantize data.
Example 1 - Custom Precision Floating-Point
The code listed here produces the plot shown in the following figure.
u=linspace(-15,15,1000);
q=quantizer([6 3],'float');
range(q)
ans =
-14
14
y=quantize(q,u);
plot(u,y);title(tostring(q))
11-123
quantize
quantizer(’float’, ’floor’, [6 3])
15
10
5
0
−5
−10
−15
−15
−10
−5
0
5
10
Example 2 - Fixed-Point
The code listed here produces the plot shown in the following figure.
u=linspace(-15,15,1000);
q=quantizer([6 2],'wrap');
range(q)
ans =
-8.0000
7.7500
y=quantize(q,u);
plot(u,y);title(tostring(q))
11-124
15
quantize
quantizer(’fixed’, ’floor’, ’wrap’, [6 2])
8
6
4
2
0
−2
−4
−6
−8
−15
See Also
quantizer
set
−10
−5
0
5
10
15
Construct a quantizer.
Set and list the properties of a quantizer, quantized
filter, or quantized FFT.
11-125
quantizer
Purpose
11quantizer
Construct a quantizer
Syntax
q = quantizer
q = quantizer('PropertyName1',PropertyValue1, ... )
q = quantizer(PropertyValue1, PropertyValue2, ... )
q = quantizer(a)
q = quantizer(pn,pv)
[qcoefficient,qinput,qoutput,qmultiplicand,qproduct,...
qsum] = quantizer(F)
[q1, q2, ...] = quantizer(F, format1, format2, ...)
Description
q = quantizer creates a quantizer with properties set to their default values.
q = quantizer('PropertyName1',PropertyValue1,...) uses property
name/ property value pairs.
q = quantizer(PropertyValue1,PropertyValue2,...) creates a quantizer
with the listed property values. When two values conflict, quantizer sets the
last property value in the list. Property values are unique; you can set the
property names by specifying just the property values in the command.
q = quantizer(a) where a is a structure whose field names are property
names, sets the properties named in each field name with the values contained
in the structure.
q = quantizer(pn,pv) sets the named properties specified in the cell array of
strings pn to the corresponding values in the cell array pv.
These are the quantizer property values, sorted by associated property name:
Property Name
Property Value
Description
Mode
'double'
Double-precision mode. Override
all other parameters.
'float'
Custom-precision floating-point
mode.
'fixed'
Signed fixed-point mode.
11-126
quantizer
Property Name
(Continued)
Roundmode
Overflowmode
(fixed-point only)
Format
Property Value
Description
'single'
Single-precision mode. Override
all other parameters.
'ufixed'
Unsigned fixed-point mode.
'ceil'
Round towards negative infinity.
'convergent'
Convergent rounding.
'fix'
Round towards zero.
'floor'
Round towards positive infinity.
'round'
Round towards nearest.
'saturate'
Saturate at max value on
overflow.
'wrap'
Wrap on overflow.
[wordlength
exponentlength]
The format for fixed or ufixed
mode.
[wordlength
exponentlength]
The format for float mode.
The default property values for a quantizer are
mode = 'fixed';
roundmode = 'floor';
overflowmode = 'saturate';
format = [16 15];
Along with the preceding properties, quantizers have read-only properties:
'max', 'min', 'noverflows', 'nunderflows', and 'noperations'. They can
be accessed through quantizer/get or q.max, q.min, q.noverflows,
q.nunderflows, and q.noperations, but they cannot be set. They are updated
during the quantizer/quantize method, and are reset by the quantizer/
reset method.
11-127
quantizer
The following table lists the read-only quantizer properties:
Property Name
Description
'max'
Maximum value before quantizing
'min'
Minimum value before quantizing
'noverflows'
Number of overflows
'nunderflows'
Number of underflows.
'noperations'
Number of data points quantized
[qcoefficient,qinput,qoutput,qmultilplicand,qproduct,qsum] =…
quantizer(F) returns property values associated with the quantized FFT F for
the twiddle factors, input, output, product, and sum quantizers.
[q1, q2,...] = quantizer(F, formatName1, formatName2,...) returns
quantizers q1, q2,..., associated with formatName1,formatName2,..., where
format k is a string that can be one of 'twiddle', 'input', 'output',
'multiplicand', 'product', or 'sum'.
Examples
The following example operations are equivalent.
Setting quantizer properties by listing property values only in the command.
q = quantizer('fixed', 'ceil', 'saturate', [5 4])
Using a structure a to set quantizer properties.
a.mode = 'fixed';
a.roundmode = 'ceil';
a.overflowmode = 'saturate';
a.format = [5 4];
q = quantizer(a);
Using property name and property value cell arrays pn and pv to set quantizer
properties.
pn = {'mode', 'roundmode', 'overflowmode', 'format'};
pv = {'fixed', 'ceil', 'saturate', [5 4]};
q = quantizer(pn, pv)
Using property name/property value pairs to configure a quantizer.
11-128
quantizer
q = quantizer( 'mode', fixed','roundmode','ceil',...
'overflowmode', 'saturate', 'format', [5 4]);
See Also
quantize
set
Apply a quantizer to data.
Set and list the properties of a quantizer, quantized
filter, or quantized FFT.
11-129
radix
Purpose
11radix
Return the radix of a quantized FFT
Syntax
radix(f)
Description
radix(f) returns the radix of quantized FFT f.
Examples
After you create a default quantized FFT, the radix function returns 2 as the
value of the radix, as shown in this example.
F = qfft;
radix(F)
returns the default 2.
See Also
qfft
qfft/get
qfft/set
11-130
Construct a quantized FFT.
Return the property values for quantized filters,
quantizers, and quantized FFTs.
Set or display property values for quantized filters,
quantizers, or quantized FFTs.
randquant
Purpose
Syntax
Description
11randquant
Generate a uniformly distributed, quantized random number
randquant(q,n)
randquant(q,m,n)
randquant(q,m,n,p,...)
randquant(q,[m,n])
randquant(q,[m,n,p,...])
randquant(q,n) uses quantizer q to generate An n by n matrix with random
entries whose values cover the range of q when q is a fixed-point quantizer.
When q is a floating-point quantizer, randquant populates the n by n array
with values covering the range -[square root of realmax(q)] to [square root of
realmax(q)].
randquant(q,m,n) uses quantizer q to generate an m by n matrix with random
entries whose values cover the range of q when q is a fixed-point quantizer.
When q is a floating-point quantizer, randquant populates the m by n array
with values covering the range -[square root of realmax(q)] to [square root of
realmax(q)].
randquant(q,m,n,p,...) uses quantizer q to generate an m by n by p by …
matrix with random entries whose values cover the range of q when q is
fixed-point quantizer. When q is a floating-point quantizer, randquant
populates the matrix with values covering the range -[square root of
realmax(q)] to [square root of realmax(q)].
randquant(q,[m,n]) uses quantizer q to generate an m by n matrix with
random entries whose values cover the range of q when q is a fixed-point
quantizer. When q is a floating-point quantizer, randquant populates the
m by n array with values covering the range -[square root of realmax(q)] to
[square root of realmax(q)].
randquant(q,[m,n,p,...]) uses quantizer q to generate p m by n matrices
containing random entries whose values cover the range of q when q is a
fixed-point quantizer. When q is a floating-point quantizer, randquant
populates the m by n arrays with values covering the range -[square root of
realmax(q)] to [square root of realmax(q)].
11-131
randquant
randquant produces pseudorandom numbers. The number sequence
randquant generates during each call is determined by the state of the
generator. Since MATLAB resets the random number generator state at
start-up, the sequence of random numbers generated by the function remains
the same unless you change the state.
randquant works like rand in most respects, including the generator used, but
it does not support the 'state' and 'seed' options available in rand.
Examples
q=quantizer([4 3]);
rand('state',0)
randquant(q,3)
ans =
0.7500
-0.6250
0.1250
See Also
-0.1250
0.6250
0.3750
quantizer
quantizer/range
quantizer/realmax
rand
11-132
-0.2500
-1.0000
0.5000
Construct a quantizer.
Return the numerical range of quantizers in a
quantized FFT, and the range of a quantizer.
Return the largest positive real number for a
quantizer.
Generate random numbers.
range
Purpose
Syntax
Description
11range
Return the numerical range of quantizers in a quantized FFT, or the range of
a quantizer
range(F)
rtwiddle = range(F)
[rtwiddle, rinput, routput, rproduct, rsum] = range(F)
[r1, r2, ...] = range(F, formattype1, formattype2, ...)
r = range(q)
[a, b] = range(q)
range(F) displays the numerical range of all the quantizers in quantized
FFT F.
rtwiddle = range(F) returns the numerical range of the twiddle factor
quantizer (although twiddle factors always have magnitudes less than 1).
[rtwiddle, rinput, routput, rproduct, rsum] = range(F) returns the
range of each of the quantizers.
[r1, r2,...] = range(F, formattype1, formattype2,...) returns the
range of the quantizers specified by strings formattype i, which may take on
the values 'twiddle', 'input', 'output', 'product', 'sum'.
r = range(q) returns the two-element row vector r = [a b] such that for all
real x, y = quantize(q,x) returns y in the range a ≤ y ≤ b.
[a, b] = range(q) returns the minimum and maximum values of the range
in separate output variables.
Examples
q = quantizer('float',[6 3]);
r = range(q)
returns r = [–14, 14].
q = quantizer('fixed',[4 2],'floor');
[a,b] = range(q)
returns a = –2, b = 1.75 = 2 –eps(q).
Algorithm
If q is a floating-point quantizer, a = -realmax(q), b = realmax(q).
If q is a signed fixed-point quantizer (mode = 'fixed'),
11-133
range
w–1
–2
a = – real max ( q ) – eps ( q ) = -----------------f
2
w–1
–1
2
b = real max ( q ) = -----------------------f
2
If q is an unsigned fixed-point quantizer (mode = 'ufixed'),
a=0
w
2 –1
b = real max ( q ) = -------------f
2
See realmax for more information.
Errors
If you use more than two output arguments, MATLAB returns the error
message Too many output arguments and aborts the function.
See Also
realmax
realmin
exponentmin
fractionlength
11-134
Return the largest positive floating-point number.
Return the smallest positive floating-point number.
Return the minimum biased exponent for quantizer q.
Return the fraction length of quantizer object q.
realmax
Purpose
11realmax
Return the largest positive quantized number
Syntax
x = realmax(q)
Description
x = realmax(q) is the largest quantized number representable where q is a
quantizer. Anything larger overflows.
Examples
q = quantizer('float',[6 3]);
x = realmax(q)
returns x = 14.
Algorithm
If q is a floating-point quantizer, the largest positive number, x, is
x=2
E ma x
⋅ ( 2 – eps ( q ) )
If q is a signed fixed-point quantizer, the largest positive number, x, is
w–1
–1
2
x = --------------------f
2
If q is an unsigned fixed-point quantizer (mode = 'ufixed'), the largest
positive number, x, is
w
2 –1
x = ---------------f
2
See Also
quantizer
realmin
exponentmin
fractionlength
Construct a quantizer.
Return the smallest positive normal quantized
number.
Return the minimum exponent for a quantizer.
Return the fraction length for a quantizer.
11-135
realmin
Purpose
11realmin
Return the smallest positive normal quantized number
Syntax
x = realmin(q)
Description
x = realmin(q) is the smallest positive normal quantized number where q is
a quantizer. Anything smaller than x underflows or is an IEEE “denormal”
number.
Examples
q = quantizer('float',[6 3]);
realmin(q)
returns the value 0.25.
Algorithm
If q is a floating-point quantizer, x = 2
the minimum exponent.
E min
where E min = exponentmin ( q ) is
If q is a signed or unsigned fixed-point quantizer, x = 2
fraction length.
See Also
11-136
exponentmin
fractionlength
–f
= ε where f is the
Return the minimum exponent for a quantizer.
Return the fraction length for a quantizer.
reset
Purpose
11reset
Reset one or more quantizers, quantized filters, or quantized FFTs to their
initial conditions
Syntax
reset(q)
reset(q1, q2, ...)
reset(hq)
reset(hq1,hq2,...)
reset(f)
reset(f1,f2,...)
Description
reset(q) resets quantizer q to its initial conditions. Works for quantized filters
and quantized FFTs as well by replacing the quantizer with a quantized filter
or quantized FFT in the command syntax.
reset(q1, q2,...) resets the states of the quantizers q1, q2,.... to the states
they were in when you created them — their initial conditions.
The states of a quantizer are
'max'
'min'
'noverflows'
'nunderflows'
'noperations'
-
Maximum value before quantizing.
Minimum value before quantizing.
Number of overflows.
Number of underflows.
Number of quantization operations performed.
reset(hq1,hq2,...) resets the states of quantized filters hq1, hq2,... to the
states you set when you created them — their initial conditions.
The states of a quantized filter are
'FilterStructure' - Structure of the filter
'ScaleValues' - Scale values between filter sections
'NumberOfSections' - Number of filter sections
'StatesPerSection' - Number of states in each filter section
'CoefficientFormat' - quantizer
'InputFormat' - quantizer
'OutputFormat' - quantizer
'MultiplicandFormat’ - quantizer
'ProductFormat' - quantizer
'SumFormat' - quantizer
11-137
reset
reset(f1,f2,...) resets the states of quantized FFTs f1, f2,... to the states
you set when you created them — their initial conditions.
The states of a quantized FFT are
'Radix' - Either 2 or 4
'Length' - Scalar integer, length of the FFT
'CoefficientFormat' - quantizer
'InputFormat' - quantizer
'OutputFormat' - quantizer
'MultiplicandFormat' - quantizer
'ProductFormat' - quantizer
'SumFormat' - quantizer
'NumberOfSections' - 4
'ScaleValues' - Vector of the scale values between FFT
sections
Examples
See Also
11-138
q1 = quantizer('fixed','ceil','saturate',[4 3])
q2 = quantizer('double')
y1 = quantize(q1, -1.2:.1:1.2 )
y2 = quantize(q2, -1.2:.1:1.2 )
q1, q2
reset(q1, q2)
q1, q2
quantizer
set
Construct a quantizer.
Set and list the properties of a quantizer, quantized
filter, or quantized FFT.
scalevalues
Purpose
11scalevalues
Return the scalevalues property of a quantized filter
Syntax
s = scalevalues(hq)
Description
s = scalevalues(hq) returns the scale values of the quantized filter hq. The
scale values for the filter scale the input to each filter section. The value of the
scalevalues property must be a scalar, or a vector of length
numberofsections(hq). For efficient computation, set the scale values to be
powers of 2.
If s is a scalar, the input to the first section of the quantized filter is scaled by s.
When s is a vector, the input to the k-th section of the filter is scaled by s(k),
the k-th element of vector s.
Examples
Hq = qfilt;
scalevalues(Hq)
ans =
1
See Also
qfilt
get
set
Construct a quantized filter.
Return the properties for a quantized filter, quantizer,
or quantized FFT.
Set or display the properties for a quantized filter,
quantizer, or quantized FFT.
11-139
set
Purpose
11set
Set or display property values for quantized filters, quantizers, and quantized
FFTs
Syntax
set(Hq)
set(Hq,'prop',value)
set(Hq,'prop1',value1, 'prop2',value2,...)
s = set(Hq)
set(Hq,struct)
set(Hq,{'prop1','prop2',...},{value1,value2,...})
set(q, PropertyValue1, PropertyValue2, ... )
set(q,a)
set(q,pn,pv)
set(q,’PropertyName1’,PropertyValue1,’PropertyName2’,
PropertyValue2,...)
q.PropertyName = Value
set(q)
s = set(q)
set(F,'PropertyName',PropertyValue)
set(F,'PropertyName1',PropertyValue1,'PropertyName2',
PropertyValue2,...)
set(F,a)
set(F,pn,pv)
F.PropertyName = Value
set(F)
s = set(F)
Description
set(Hq) displays all of the property names and their possible values for a given
quantized filter Hq. The display indicates the default values for properties in
braces. When the default values for a property cannot be represented by a finite
list, set(Hq) does not display the property’s default values.
set(Hq,'prop',value) sets the values for the property 'prop' of a quantized
filter Hq. You specify the property name by the string 'prop', and the
associated value in value. 'prop' can be any of the properties listed in
Table 10-3, Quick Guide to Quantized Filter Properties, on page 10-10. value
can be a string, a numerical value, or a cell array containing numerical values.
The possible values for each property are described in detail in “Quantized
Filter Properties Reference” on page 10-11.
11-140
set
set(Hq,'prop1',value1,'prop2',value2,...) lets you set multiple
properties in one command.
s = set(Hq) returns all property names and their possible values for a
quantized filter Hq. s is a MATLAB structure whose field names are the
property names of Hq and whose values are cell arrays of possible property
values, except when the possible values for the property cannot be described
with a finite list. In this case the values are empty cell arrays.
set(Hq,struct) sets the properties of the quantized filter Hq according to the
values associated with the field names of the MATLAB structure struct. All
field names for the structure s must be valid quantized filter properties. See
Table 10-3, Quick Guide to Quantized Filter Properties, on page 10-10 for a list
of all property names.
set(Hq,{'prop1','prop2',...},{value1,value2,...}) sets the listed
properties specified in the cell array of a vector of strings
{'prop1','prop2',...} to the corresponding values listed in the cell array
{value1,value2,...} for quantized filter object Hq. The two cell array input
arguments must be the same size, and the values must be valid for the
corresponding properties.
set(q, PropertyValue1, PropertyValue2,...) sets the properties of
quantizer q. If two property values conflict, the last value in the list is the one
that is set.
set(q,a) where a is a structure whose field names are object property names,
sets the properties named in each field name with the values contained in the
structure.
set(q,pn,pv) sets the named properties specified in the cell array of strings pn
to the corresponding values in the cell array pv.
set(q,’PropertyName1’,PropertyValue1,’PropertyName2’,
PropertyValue2,...) sets multiple property values with a single statement.
Note that you can use property name/property value string pairs, structures,
and property name/property value cell array pairs in the same call to set.
q.PropertyName = Value uses the dot notation to set property PropertyName
to Value.
set(q) displays the possible values for all properties of quantizer q.
11-141
set
s = set(q) returns a structure containing the possible values for the
properties of quantizer q.
The states are cleared when you set any value other than WarnIfOverflow.
For a quantizer, these are the possible property values, sorted by property
name.
Property Name
Property Value
Description
Mode
'double'
Double-precision mode. Override
all other parameters.
'float'
Custom-precision floating-point
mode.
'fixed'
Signed fixed-point mode.
'single'
Single-precision mode. Override
all other parameters.
'ufixed'
Unsigned fixed-point mode.
'ceil'
Round towards negative infinity.
'convergent'
Convergent rounding.
'fix'
Round towards zero.
'floor'
Round towards positive infinity.
'round'
Round towards nearest.
Roundmode
Overflowmode
(fixed-point only)
Format
11-142
'saturate'
Saturate at max value on
overflow.
'wrap'
Wrap on overflow.
[wordlength
exponentlength]
The format for fixed or ufixed
mode.
[wordlength
exponentlength]
The format for float mode.
set
Property Name
(Continued)
Property Value
Description
Max
Maximum value before quantize.
Min
Minimum value before quantize.
NOverflows
Number of overflows.
NUnderflows
Number of underflows.
set(F,'PropertyName',PropertyValue) sets the value of the specified
property for the quantized FFT F.
set(F,'PropertyName1',PropertyValue1,'PropertyName2',PropertyValue
2,...) sets multiple property values with a single statement. Note that you
can use property name/property value string pairs, structures, and property
name/property value cell array pairs in the same call to set.
set(F,a) where a is a structure whose field names are object property names,
sets the properties named in each field name with the values contained in the
structure.
set(F,pn,pv) sets the named properties specified in the cell array of strings pn
to the corresponding values in the cell array pv for all objects specified in H.
F.PropertyName = Value uses the dot notation to set property PropertyName
to Value.
set(F) displays the possible values.
s = set(F) returns a structure with the possible values.
Remarks
• Property names are not case sensitive.
• You can abbreviate property names to the shortest uniquely identifying
string.
• You can use direct property referencing to set properties with a
structure-like syntax. The following two statements are equivalent:
- set(Hq,'roundm','convergent');
- Hq.round = 'convergent';
11-143
set
Examples
Create a quantized filter and change the values for the
ReferenceCoefficients and InputFormat properties.
Hq = qfilt;
set(Hq,'ref',{[1 .5] [1 .7 .89]},'inp',[16,14])
Hq
Hq =
Quantized direct-form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
+ (1)
0.999969482421875 1.000000000000000000
(2)
0.500000000000000 0.500000000000000000
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
+ (1)
0.999969482421875 1.000000000000000000
(2)
0.700012207031250 0.699999999999999960
(3)
0.890014648437500 0.890000000000000010
FilterStructure =
ScaleValues =
NumberOfSections =
StatesPerSection =
CoefficientFormat =
InputFormat =
OutputFormat =
MultiplicandFormat =
ProductFormat =
SumFormat =
Warning: 2 overflows
df2t
[1]
1
[2]
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
in coefficients.
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
14])
15])
15])
30])
30])
You can create a structure to assign the same data format property values to a
set of filters.
s.InputFor = [16,14];
s.Coefficient = [16,14];
s.SumF = [17,15];
s.Prod = [16,15];
s.output = [24,23];
Now assign those property values to the filter in the previous example.
set(Hq,s)
Hq
Hq =
11-144
set
Quantized Direct-form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
(1)
1.00000000000000 1.000000000000000000
(2)
0.50000000000000 0.500000000000000000
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
(1)
1.00000000000000 1.000000000000000000
(2)
0.70001220703125 0.699999999999999960
(3)
0.89001464843750 0.890000000000000010
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
=
=
=
=
=
=
=
=
=
=
df2t
[1]
1
[2]
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[24
[16
[16
[17
14])
14])
23])
15])
15])
15])
Notice that you do not have to provide complete property names when you
created the structure fields.
You can set property values by using the property value only.
q = quantizer;
set(q, 'fixed', 'ceil', 'wrap', [24 22]);
Using property name/property value pairs to set quantizer properties.
q = quantizer;
set(q, 'mode','fixed', ...
'roundmode','ceil', ...
'overflowmode','wrap', ...
'format',[24 22]);
Using dot notation to enter
q = quantizer;
q.mode = 'fixed';
q.roundmode = 'ceil';
q.overflowmode = 'wrap';
q.format = [24 22];
11-145
set
With no output arguments and one input argument, set displays the defaults
for the quantizer, quantized filter, or quantized FFT.
q = quantizer;
set(q)
Mode: [double | float | {fixed} | single | ufixed]
RoundMode: [ceil | convergent | fix | {floor} | round]
OverflowMode: [{saturate} | wrap]
Format: [wordlength fractionlength] - In 'fixed', 'ufixed' mode.
[wordlength exponentlength] - In 'float' mode.
[16 15] = default.
Max: Maximum value before quantize.
Min: Minimum value before quantize.
NOverflows: Number of overflows.
NUnderflows: Number of underflows.
With one output argument and one input argument, set returns a structure.
q = quantizer;
s = set(q)
returns
s =
Mode: {'double' 'float' 'fixed' 'single' 'ufixed'}
RoundMode: {'ceil' 'convergent' 'fix' 'floor' 'round'}
OverflowMode: {'saturate' 'wrap'}
Format: {}
Max: {}
Min: {}
Overflows: {}
NUnderflows: {}
See Also
11-146
get
qfilt
setbits
sos2cell
sos
Get property values for quantized filters.
Construct a quantized filter.
Set all data formats for quantized filters.
Convert a second-order section matrix to cell arrays.
Convert a quantized filter to second-order form, order
and scale.
setbits
Purpose
Syntax
11setbits
Set all data format property values for quantized filters and quantized FFTs
setbits(Hq,format)
setbits(F,fmt)
Description
When Hq is a floating-point quantized filter, setbits(Hq,format) sets all data
format properties for the quantized filter Hq to the values specified by format.
In this case, format is a two-element vector of integers whose entries are
described as follows:
• The first entry in format sets the word length in bits.
• The second entry in format sets the exponent length in bits.
When Hq is a fixed-point quantized filter, setbits(Hq,format) sets the
properties CoefficientFormat, InputFormat, and OutputFormat to the value
specified by format, whereas the property values SumFormat and
ProductFormat are specified by 2*format. In this case, format is a two-element
vector of integers whose entries are described as follows:
• The first entry in format sets the word length in bits.
• The second entry in format sets the fraction length (the number of bits after
the radix point).
Note When 2*format exceeds the maximum values for the SumFormat and
ProductFormat properties, their maximum values are used instead.
setbits(F,fmt) sets all data format property values for quantized FFT F.
When F is a fixed-point quantized FFT, fmt = [w, f] where w is the word
length and f is the fraction length. The twiddle, input, and output formats are
set to [w, f]. The sum and product formats are set to [2w, 2f].
When F is a floating-point quantized FFT, fmt = [w, e] where w is the word
length and e is the exponent length. All formats are set to [w, e].
If the specified formats exceed the maximum allowed, they are set to the
maximum.
11-147
setbits
Examples
Create a quantized filter with default data format property values. Set the
CoefficientFormat, InputFormat, and OutputFormat property values for a
24-bit word length, and a 23-bit fraction length, while setting the SumFormat
and ProductFormat property values to a 48-bit word length and a 46-bit
fraction length.
Hq = qfilt;
setbits(Hq,[24 23])
get(Hq)
Quantized Direct form II transposed filter
Numerator
QuantizedCoefficients{1}
ReferenceCoefficients{1}
+ (1)
0.9999998807907105 1.000000000000000000
(2)
0.5000000000000000 0.500000000000000000
Denominator
QuantizedCoefficients{2}
ReferenceCoefficients{2}
+ (1)
0.9999998807907105 1.000000000000000000
(2)
0.7000000476837158 0.699999999999999960
(3)
0.8899999856948853 0.890000000000000010
FilterStructure
ScaleValues
NumberOfSections
StatesPerSection
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
See Also
11-148
get
qfilt
set
=
=
=
=
=
=
=
=
=
=
df2t
[1]
1
[2]
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
Get property values for quantized filters.
Construct a quantized filter.
Set property values for quantized filters.
[24
[24
[24
[24
[48
[48
23])
23])
23])
23])
46])
46])
sos
Purpose
11sos
Convert a quantized filter to second-order section form, order, and scale.
Syntax
Hq2 = sos(Hq)
Hq2 = sos(Hq, order)
Hq2 = sos(Hq, order, scale)
Description
Hq2 = sos(Hq) returns a quantized filter Hq2 that has second-order sections
and the dft2 structure. You can use the same optional arguments used in
tf2sos.
Hq2 = sos(Hq, order) specifies the order of the sections in Hq2, where order
is either of the following strings:
• 'down' — to order the sections so the first section of Hq2 contains the poles
closest to the unit circle (L∞ norm scaling)
• 'up' — to order the sections so the first section of Hq2 contains the poles
farthest from the unit circle (L2 norm scaling and the default)
Hq2 = sos(Hq, order, scale) also specifies the desired scaling of the gain
and numerator coefficients of all second-order sections, where scale is one of
the following strings:
• 'none' — to apply no scaling (default)
• 'inf' — to apply infinity-norm scaling
• 'two' — to apply 2-norm scaling
Use infinity-norm scaling in conjunction with up-ordering to minimize the
probability of overflow in the filter realization. Consider using 2-norm scaling
in conjunction with down-ordering to minimize the peak round-off noise.
When Hq is a fixed-point filter, the filter coefficients are normalized so that the
magnitude of the maximum coefficient in each section is 1. The gain of the filter
is applied to the first scale value of Hq2.
sos uses the direct form II transposed (dft2) structure to implement secondorder section filters.
11-149
tf2ca
Purpose
11tf2ca
Transfer function to coupled allpass conversion
Syntax
[d1,d2] = tf2ca(b,a)
[d1,d2] = tf2ca(b,a)
[d1,d2,beta] = tf2ca(b,a)
Description
[d1,d2] = tf2ca(b,a) where b is a real, symmetric vector of numerator
coefficients and a is a real vector of denominator coefficients, corresponding to
a stable digital filter, returns real vectors d1 and d2 containing the
denominator coefficients of the allpass filters H1(z) and H2(z) such that
B(z)
1
H ( z ) = ------------ = -----------------------------------------------A( z )
2 [ H1 ( z ) + H2 ( z ) ]
representing a coupled allpass decomposition.
[d1,d2] = tf2ca(b,a) where b is a real, antisymmetric vector of numerator
coefficients and a is a real vector of denominator coefficients, corresponding to
a stable digital filter, returns real vectors d1 and d2 containing the
denominator coefficients of the allpass filters H1(z) and H2(z) such that
1
B(z)
H ( z ) = ------------ = -- - [ H1 ( z ) – H2 ( z ) ]
è 2ø
A( z )
In some cases, the decomposition is not possible with real H1(z) and H2(z). In
those cases a generalized coupled allpass decomposition may be possible,
whose syntax is
[d1,d2,beta] = tf2ca(b,a)
to return complex vectors d1 and d2 containing the denominator coefficients of
the allpass filters H1(z) and H2(z), and a complex scalar beta, satisfying
|beta| = 1, such that
B(z)
1
H ( z ) = ------------ = -- - [ β • H1 ( z ) + β • H2 ( z ) ]
è 2ø
A( z )
representing the generalized allpass decomposition.
11-150
tf2ca
In the above equations, H1(z) and H2(z) are real or complex allpass IIR filters
given by
fliplr ( ( D1 ( z ) ) )
fliplr ( ( D2 ( 1 ) ( z ) ) )
H1 ( z ) = ------------------------------------------ , H2 ( 1 ) ( z ) = -------------------------------------------------D1 ( z )
D2 ( 1 ) ( z )
where D1(z) and D2(z) are polynomials whose coefficients are given by d1 and
d2.
Note A coupled allpass decomposition is not always possible. Nevertheless,
Butterworth, Chebyshev, and Elliptic IIR filters, among others, can be
factored in this manner. For details, refer to Signal Processing Toolbox User's
Guide.
Examples
See Also
[b,a]=cheby1(9,.5,.4);
[d1,d2]=tf2ca(b,a); % TF2CA returns denominators of the allpass.
num = 0.5*conv(fliplr(d1),d2)+0.5*conv(fliplr(d2),d1);
den = conv(d1,d2); % Reconstruct numerator and denonimator.
max([max(b-num),max(a-den)]) % Compare original and reconstructed
% numerator and denominators.
ca2tf
cl2tf
iirpowcomp
latc2tf
tf2ca
tf2latc
Convert coupled allpass filter form to transfer function
forms.
Convert coupled allpass lattice filter to transfer
function forms.
Compute power complementary filter.
Convert lattice filter to transfer function forms.
Convert transfer function form to coupled allpass
forms.
Convert transfer function form to lattice filter forms.
11-151
tf2cl
Purpose
Syntax
11tf2cl
Transfer function to coupled allpass lattice conversion
[k1,k2] = tf2cl(b,a)
[k1,k2] = tf2cl(b,a)
Description
[k1,k2] = tf2cl(b,a) where b is a real, symmetric vector of numerator
coefficients and a is a real vector of denominator coefficients, corresponding to
a stable digital filter, will perform the coupled allpass decomposition
B(z)
1
H ( z ) = ------------ = -----------------------------------------------A( z )
2 [ H1 ( z ) + H2 ( z ) ]
of a stable IIR filter H(z) and convert the allpass transfer functions H1(z) and
H2(z) to a coupled lattice allpass structure with coefficients given in vectors k1
and k2.
[k1,k2] = tf2cl(b,a) where b is a real, antisymmetric vector of numerator
coefficients and a is a real vector of denominator coefficients, corresponding to
a stable digital filter, performs the coupled allpass decomposition
1
B(z)
H ( z ) = ------------ = -- - [ H1 ( z ) – H2 ( z ) ]
è 2ø
A( z )
of a stable IIR filter H(z) and converts the allpass transfer functions H1(z) and
H2(z) to a coupled lattice allpass structure with coefficients given in vectors k1
and k2.
In some cases, the decomposition is not possible with real H1(z) and H2(z). In
those cases, a generalized coupled allpass decomposition may be possible, using
the command syntax
[k1,k2,beta] = tf2cl(b,a)
to perform the generalized allpass decomposition of a stable IIR filter H(z) and
convert the complex allpass transfer functions H1(z) and H2(z) to
corresponding lattice allpass filters
1
B(z)
H ( z ) = ------------ = -- - [ β • H1 ( z ) + β • H2 ( z ) ]
è 2ø
A( z )
where beta is a complex scalar of magnitude equal to 1.
11-152
tf2cl
Note Coupled allpass decomposition is not always possible. Nevertheless,
Butterworth, Chebyshev, and Elliptic IIR filters, among others, can be
factored in this manner. For details, refer to Signal Processing Toolbox User's
Guide.
Examples
See Also
[b,a]=cheby1(9,.5,.4);
[k1,k2]=tf2cl(b,a); % Get the reflection coeffs. for the lattices.
[num1,den1]=latc2tf(k1,'allpass'); % Convert each allpass lattice
[num2,den2]=latc2tf(k2,'allpass'); % back to transfer function.
num = 0.5*conv(num1,den2)+0.5*conv(num2,den1);
den = conv(den1,den2); % Reconstruct numerator and denonimator.
max([max(b-num),max(a-den)]) % Compare original and reconstructed
% numerator and denominators.
ca2tf
cl2tf
iirpowcomp
latc2tf
tf2ca
tf2latc
Convert coupled allpass filter form to transfer function
forms.
Convert coupled allpass lattice filter to transfer
function forms.
Compute power complementary filter.
Convert lattice filter to transfer function forms.
Convert transfer function form to coupled allpass
forms.
Convert transfer function form to lattice filter forms.
11-153
tostring
Purpose
11tostring
Convert a quantizer, unitquantizer, or quantized FFT to a string
Syntax
s = tostring(q)
s = tostring(q)
s = tostring(f)
Description
s = tostring(q) converts quantizer q to a string s. After converting q to a
string, the function eval(s) can use s to create a quantizer with the same
properties as q.
s = tostring(q) converts unitquantizer q to a string s. After converting q to
a string, the function eval(s) can use s to create a quantizer with the same
properties as q.
s = tostring(q) converts quantized FFT f to a string s. After converting f to
a string, the function eval(s) can use f to create a quantized FFT with the
same properties as f.
Examples
When you use tostring with a quantizer or unitquantizer, you see the
following response.
q = quantizer
q =
Mode = fixed
RoundMode = floor
OverflowMode = saturate
Format = [16 15]
Max
Min
NOverflows
NUnderflows
NOperations
=
=
=
=
=
reset
reset
0
0
0
s = tostring(q)
s =
quantizer('fixed', 'floor', 'saturate', [16
eval(s)
11-154
15])
tostring
ans =
Mode
RoundMode
OverflowMode
Format
=
=
=
=
fixed
floor
saturate
[16 15]
Max
Min
NOverflows
NUnderflows
NOperations
=
=
=
=
=
reset
reset
0
0
0
and s is the same as q.
For a quantized FFT, the result is the same.
f = qfft
f =
Radix = 2
Length
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
NumberOfSections
ScaleValues
s=tostring(f)
=
=
=
=
=
=
=
=
=
16
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
4
[1]
=
=
=
=
=
=
=
=
2
16
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
quantizer('fixed',
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
15])
15])
15])
30])
30])
'round',
'floor',
'floor',
'floor',
'floor',
'floor',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
'saturate',
[16
[16
[16
[16
[32
[32
15])
15])
15])
15])
30])
30])
eval(s)
ans =
Radix
Length
CoefficientFormat
InputFormat
OutputFormat
MultiplicandFormat
ProductFormat
SumFormat
11-155
tostring
NumberOfSections = 4
ScaleValues = [1]
See Also
11-156
quantizer
qfft
unitquantizer
Construct a quantizer.
Construct a quantized FFT.
Construct a unitquantizer.
twiddles
Purpose
11twiddles
Return the quantized twiddle factors for quantized FFTs
Syntax
w = twiddles(F)
Description
w = twiddles(F) returns a vector of the quantized FFT coefficients specified
by quantized FFT F. FFT coefficients are also called twiddle factors.
Examples
f = qfft;
w = twiddles(f)
Warning: 4 overflows.
w =
1.0000
1.0000
0
1.0000
0.7071
0
-0.7071
1.0000
0.9239
0.7071
0.3827
0
-0.3827
-0.7071
-0.9239
See Also
qfft
- 1.0000i
- 0.7071i
- 1.0000i
- 0.7071i
-
0.3827i
0.7071i
0.9239i
1.0000i
0.9239i
0.7071i
0.3827i
Construct a quantized FFT.
11-157
unitquantize
Purpose
11unitquantize
Quantize all numbers in a data set except numbers within eps of 1
Syntax
unitquantize(q,...)
Description
unitquantize(q,...) works the same as quantize except that numbers
within eps(q) of 1 are made exactly equal to 1 (see quantize for a description
of the parameters).
This function is especially useful for quantizing fixed-point coefficients.
Examples
[b,a] = ellip(4,3,20,.6);
m = tf2sos(b,a)
m =
0.2758
1.0000
0.3883
0.7197
0.2758
1.0000
1.0000
1.0000
-0.0548
0.5884
0.4731
0.9574
m==1
ans =
0
1
0
0
0
0
1
1
0
0
0
0
0
1
1
1
0
0
0
0
q=quantizer;
m > realmax(q)
ans =
0
1
0
0
It appears that there are four elements that are exactly equal to 1. In fact, there
are only three. Element m(2,3) is greater than realmax(q), but less than 1.
Ordinarily, m(2,3) would be counted as an overflow and be set to realmax(q).
However, the desired behavior would be to force m(2,3) to be equal to 1 without
recording an overflow. This is what unitquantize does, as shown in the
following example.
11-158
unitquantize
m = unitquantize(q,m)
m =
0.2758
1.0000
m==1
0.3882
0.7197
0.2758
1.0000
1.0000
1.0000
-0.0548
0.5884
0.4731
0.9574
ans =
0
1
0
0
0
1
1
1
0
0
0
0
By forcing values between eps and 1 to be equal to 1, signal processing
algorithms can avoid multiplication operations that involve these numbers,
saving processing steps and time.
See Also
qfft
quantize
Construct a quantized FFT.
Construct a quantizer.
11-159
unitquantizer
Purpose
11unitquantizer
Construct a unit quantizer
For help on this function, enter help unitquantizer at the MATLAB prompt.
Syntax
q = unitquantizer(...)
q = unitquantizer(...) constructs a unitquantizer, which is identical to a
quantizer in all respects except that its quantize method quantizes numbers
within eps(q) of 1 to be equal to 1. Refer to quantizer for arguments and
parameters for the unitquantizer function.
Examples
u = unitquantizer([4 3]);
quantize(u,1)
ans =
1
q = quantizer([4 3]);
quantize(q,1)
Warning: 1 overflow.
ans =
0.8750
See Also
11-160
quantizer
unitquantize
Construct a quantizer.
Construct a unitquantizer.
wordlength
Purpose
11wordlength
Return the word length for a quantizer
Syntax
wordlength(q)
Description
wordlength(q) returns the word length of quantizer q.
Examples
q = quantizer([16 15]);
wordlength(q)
returns 16.
Even though the word length can be read in two stages,
q = quantizer([16 15]);
fmt = q.format;
w = fmt(1);
it is handy to have it available for use in equations. For example, the algorithm
for realmax(q) when q is a signed fixed-point quantizer (q.mode = 'fixed')
is
r=2
w–f–1
–ε
which can be coded as
q = quantizer('fixed',[8 4]);
r = pow2(wordlength(q) - fractionlength(q) - 1) - eps(q)
See Also
fractionlength
exponentlength
Return the exponent length of a quantizer.
Return the fraction length for a quantizer.
11-161
zplane
Purpose
11zplane
Compute a zero-pole plot for quantized filters
Syntax
zplane(Hq)
zplane(Hq,'plotoption')
zplane(Hq,'plotoption','plotoption2')
[zq,pq,kq] = zplane(Hq)
[zq,pq,kq,zr,pr,kr] = zplane(Hq)
Description
This function displays the poles and zeros of quantized filters, as well as the
poles and zeros of the associated unquantized reference filter.
zplane(Hq) plots the zeros and poles of a quantized filter Hq in the current
figure window. The poles and zeros of the quantized and unquantized filters
are plotted by default. The symbol o represents a zero of the unquantized
reference filter, and the symbol x represents a pole of that filter. The symbols
and + are used to plot the zeros and poles of the quantized filter Hq. The plot
includes the unit circle for reference.
zplane(Hq,'plotoption') plots the poles and zeros associated with the
quantized filter Hq according to one specified plot option. The string
'plotoption' can be either of the following reference filter display options:
• 'on' to display the poles and zeros of both the quantized filter and the
associated reference filter (default)
• 'off' to display the poles and zeros of only the quantized filter
zplane(Hq,'plotoption','plotoption2') plots the poles and zeros
associated with the quantized filter Hq according to two specified plot options.
The string 'plotoption' can be selected from the reference filter display
options listed in the previous syntax. The string 'plotoption2' can be selected
from the section-by-section plotting style options described below:
• 'individual' to display the poles and zeros of each section of the filter in a
separate figure window
• 'overlay' to display the poles and zeros of all sections of the filter on the
same plot
• 'tile' to display the poles and zeros of each section of the filter in a separate
plot in the same figure window
11-162
zplane
[zq,pq,kq] = zplane(Hq) returns the vectors of zeros zq, poles pq, and gains
kq. If Hq has n sections, zq, pq, and kq are returned as 1-by-n cell arrays. If
there are no zeros (or no poles), zq (or pq) is set to the empty matrix [].
[zq,pq,kq,zr,pr,kr] = zplane(Hq) returns the vectors of zeros zr, poles pr,
and gains kr of the reference filter associated with the quantized filter Hq, and
returns the vectors of zeros zq, poles pq, and gains kq for the quantized filter Hq.
Examples
Create a quantized filter Hq from a fourth-order digital filter with cutoff
frequency of 0.6. Scale the transfer function parameters to avoid overflows due
to coefficient quantization. Plot the quantized and unquantized poles and zeros
associated with this quantized filter.
[b,a] = ellip(4,.5,20,.6);
Hq = qfilt('df2',{b/2 a/2});
zplane(Hq);
Quantized zeros
Quantized poles
Reference zeros
Reference poles
1
0.8
0.6
Imaginary part
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1
See Also
freqz
impz
−0.5
0
Real part
0.5
1
Estimate the frequency response of quantized filters.
Compute the impulse response for quantized filters.
11-163
zplane
11-164
12
Bibliography
12
Bibliography
[1] Antoniou, A., Digital Filters: Analysis, Design, and Applications, Second
Edition, McGraw-Hill, Inc. 1993.
[2] Chirlian, P.M., Signals and Filters, Van Nostrand Reinhold, 1994.
[3] Fliege, N.J., Mulitrate Digital Signal Processing, John Wiley and Sons,
1994.
[4] Jackson, L., Digital Filtering and Signal Processing with MATLAB
Exercises, Third edition, Kluwer Academic Publishers, 1996.
[5] Lapsley, P., J.Bier, A. Sholam, and E.A. Lee, DSP Processor Fundamentals:
Architectures and Features, IEEE Press, 1997.
[6] McClellan, J.H., C.S. Burrus, A.V. Oppenheim, T.W. Parks, R.W. Shafer,
and H.W. Schuessler, Computer-Based Exercises for Signal Processing Using
MATLAB 5, Prentice-Hall, 1998.
[7] Moler, C., “Floating points: IEEE Standard unifies arithmetic model,”
Cleve’s Corner, The MathWorks, Inc., 1996. See http://www.mathworks.com/
company/newsletter/pdf/Fall96Cleve.pdf.
[8] Oppenheim, A.V., and R.W. Schafer, Discrete-Time Signal Processing,
Prentice-Hall, 1989.
12-2
Index
A
abbreviating property names 4-6
accessing properties 4-5
addition, format for
quantized FFTs 10-50
quantized filters 10-44
advanced FIR filter design 2-7
advanced IIR filter design 2-37
antisymmetricfir 10-14
arithmetic
quantized filtering, effects on 8-22
B
basic filter properties
quantized filters 6-6, 7-6
quantizers 5-4
bias 3-19
bibliography 12-2
binary
coding 3-3
data types 3-3
binary point 3-16
bits
definition 3-16
setting, quantized FFTs 7-8
setting, quantized filters 6-11
brackets, indicating closed interval xviii
C
cell arrays
indexing into cell arrays of cell arrays 4-14
indexing into cell arrays of matrices 4-13
cell arrays, quantized filter coefficients 10-11
changing quantized filter properties in FDATool
9-11
coefficient overflow indicator 9-17
coefficient underflow indicator 9-17
CoefficientFormat property
setting 10-4
command line help 4-11
constructing objects 4-3
context-sensitive help 9-5
controls, FDATool 9-4
conventions in our documentation (table) xix
convert structure dialog 9-14
converting filter structures in FDATool 9-12
copying objects 4-4
custom floating-point 3-21
D
data formats
operands, quantized FFTs 10-48
operands, quantized filters 10-32
outputs, quantized FFTs 10-49
outputs, quantized filters 10-33
properties 11-147
quantized FFTs, setting all 7-8
quantized filters 3-5, 6-11, 7-8
quantized filters, setting all 6-10
setting 10-4
data types
binary 3-3
quantized filters 6-9, 7-7
denormalized numbers 3-23
designing advanced FIR filters 2-7
designing advanced IIR filters 2-37
df1 10-16
df1t 10-17
df2 10-18
df2t 10-19
I-1
Index
digital filters
fixed-point 3-3
floating-point 3-4
digital frequency xviii
direct form I 10-16
transposed 10-17
direct form II 10-18
transposed 10-19
direct property referencing 11-143
dot notation 11-143
double-precision 3-20
DSP processors 3-4
dynamic range
fixed-point 3-16
floating-point 3-21
E
ellipses, in syntax xviii
entering transfer function scale values in
FDATool 9-18
envelope delay. See group delay
equiripple filters 2-6
error, Lp norm 2-4
errors, quantization 2-59
exceptional arithmetic 3-23
exponents 3-3
length 3-18, 10-4
export dialog in FDATool 9-22
export filter dialog options 9-22
export to 9-22
overwrite existing variables 9-22
variable names 9-22
exporting filters 9-22
exporting quantized filters in FDATool 9-22
I-2
F
FDATool
about importing and exporting filters 9-20
apply option 9-4
changing quantized filter properties 9-11
convert structure dialog 9-14
convert structure option 9-12
converting filter structures 9-12
entering transfer function scale values 9-18
export dialog 9-22
exporting quantized filters 9-22
getting help 9-5
import filter dialog 9-21
importable filter structures 9-20
importing filters 9-21
input/output scaling 9-4
quantized filter properties 9-6
quantizer property lists 9-4
quantizing filters 9-6
quantizing reference filters 9-10
scale trans-fcn coeffs <=1 option 9-4
scaling transfer function coefficients 9-16
scaling transfer function coefficients manually
9-18
set quantization mode 9-4
set quantization parameters dialog 9-6
setting properties 9-6
show filter structures option 9-4
turn quantization on option 9-4
user options 9-4
using input/output scaling 9-18
viewing filter structure schematics 9-14
fdatool
about 9-2
about quantization mode 9-3
context-sensitive help 9-5
switching to quantization mode 9-3
Index
FFTs
computing quantized 7-10
filter 11-42
filter banks
quantized 8-16
filter conversions 10-42
Filter Design and Analysis Tool. See fdatool
filter design GUI
context-sensitive help 9-5
help on 9-5
filter design methods
firlpnorm 2-5
gremez 2-7
gremez design examples 2-8
iirgrpdelay 2-37
iirlpnorm 2-37
iirlpnorm design examples 2-40
iirlpnormc 2-37
iirlpnormc design examples 2-45
filter design, advanced FIR 2-7
filter design, advanced IIR 2-37
filter design, minimax 2-4
filter design, optimal 2-2
filter sections
specifying 10-42
filter structures 6-8
direct form FIR 10-20
direct form I 10-16
direct form I transposed 10-17
direct form II 10-18
direct form II transposed 10-19
direct form symmetric FIR 10-30
FIR transposed 10-21
lattice allpass 10-22
lattice AR 10-26
lattice ARMA 10-28
lattice coupled-allpass 10-36
lattice coupled-allpass power complementary
10-36
lattice MA minimum phase 10-27
lattice moving average maximum phase 10-23
state-space 10-29
filtering data
function for 11-42
logs of overflows 11-45
logs of underflows 11-45
obtaining states 11-45
filters
about equiripple 2-6
direct form 10-12
exporting as MAT-file 9-24
exporting as text file 9-23
exporting from FDATool 9-22
FIR 10-12
getting filter coefficients after exporting 9-23
importing and exporting 9-20
lattice 10-12
state-space 10-12
testing for allpass structure 11-3
testing for FIR structure 11-3
testing for linear phase sections 11-3
testing for maximum phase design 11-3
testing for minimum phase design 11-3
testing for purely real coefficients 11-3
testing for second-order sections 11-3
testing for stability 11-3
filters, exporting to workspace 9-22
filters, importing into FDATool 9-21
filters, low-sensitivity 2-59
filters, robust 2-59
FilterStructure property 10-12
finite impulse response
antisymmetric 10-14
symmetric 10-30
I-3
Index
I-4
fir 10-20
freqz 11-50
FIR filters 10-12
firlpnorm design method 2-5
firt 10-21
fixed-point 3-15
data formats 3-5
filters 3-4
fraction length 3-3
ranges 3-3
sign bit 3-15
word length 3-3
fixed-point numbers
scaling 3-17
floating-point 3-18
bias 3-19
custom 3-21
double precision 3-20
dynamic range 3-21
exponents 3-19
filters 3-4
fractions 3-19
IEEE format 3-19
mantissa 3-3
precision 3-22
ranges 3-3
sign bits 3-19
single precision 3-20
word length 3-18
fractions 3-19
determining length 3-3
limitations on length 10-4
frequency
digital xviii
Nyquist xviii
frequency response 11-50
noise loading method 3-12
frequency response plots 3-11
function for opening FDATool 9-3
functions, overloading 4-10
G
get 11-53
getting filter coefficients after exporting 9-23
getting properties 4-8
command for 11-53
getting started 1-15
getting started example 1-15
gremez 2-7
gremez design examples 2-8
group delay, about 2-51
group delay, prescribed 2-37
H
help
command line 4-11
I
IEEE
format 3-19
nonstandard format 3-21
iirgrpdelay 2-37
iirgrpdelay design examples 2-51
iirlpnorm 2-37
iirlpnorm design examples 2-38
iirlpnormc 2-37
iirlpnormc design examples 2-38
import filter dialog in FDATool 9-21
import filter dialog options 9-21
frequency units 9-21
quantized filter 9-21
Index
import/export filters in FDATool 9-20
importing filters 9-21
importing quantized filters in FDATool 9-21
impulse response 11-73
impulse response plots 3-10
impz 11-73
indexing
cell arrays of cell arrays 4-14
cell arrays of matrices 4-13
vectors xviii
indicator, overflow 9-17
indicator, underflow 9-17
InputFormat property 10-4
interval notation xviii
inverse FFTs
computing quantized 7-10
isallpass 11-3
isfir 11-3
islinphase 11-3
ismaxphase 11-3
isminphase 11-3
isreal 11-2, 11-81
issos 11-3
isstable 11-3
L
moving average maximum phase 10-23
moving average minimum phase 10-27
latticear 10-26
latticearma 10-28
latticeca 10-22, 10-23, 10-24
latticecapc 10-25
latticema 10-27
least significant bit 3-16
limit cycles in quantized filters 3-13
low-sensitivity filters 2-59
Lp norm 2-4
LSB 3-16
M
mantissa 3-3
minimax filter designs 2-4
Mode property 10-5
most significant bit 3-16
MSB 3-16
multiple sections
specifying 10-42
MultiplicandFormat property
quantized FFTs 10-48
quantized filter 10-32
latcallpass 10-22
N
latcmax 10-23
new users, tips for xiii
nlm 3-12
noise loading method 3-12
nonstandard IEEE format 3-21
NOperations property 10-6
normalize 3-17
normalize 11-95
normalizing quantized filters 6-12
NumberOfSections property
lattice filters
allpass 10-22
AR 10-26
ARMA 10-28
autoregressive 10-26
coupled-allpass 10-24
coupled-allpass power complementary 10-25
MA 10-27
I-5
Index
quantized filters 10-32
NumberOfStages property
quantized FFTs 10-48
NUnderflows property 10-6
Nyquist frequency xviii
O
object properties 1-13
objects
constructing 4-3
copying 4-4
objects in this toolbox 1-13
opening FDATool, function for 9-3
optimal filter design
problem statement 2-2
solutions 2-5
theory 2-2
options, FDATool 9-4
OutputFormat property
quantized FFTs 10-49
quantized filters 10-32, 10-33
setting 10-4
overflow 3-22
overflow indicator 9-17
overflow mode property
saturate 9-10
wrap 9-10
overflow, checking for 9-17
OverflowMode property 10-6
overflows
addressing, function for 11-95
overloading 4-10
P
parentheses, indicating open interval xviii
I-6
Parks-McClellan method 2-6
plots
frequency response 3-11
impulse response 3-10
impulse response, command for 11-73
noise loading method 3-12
pole/zero 3-10
zero-pole, command for 11-162
pole/zero plots 3-10
pole-zero plots 11-162
precision
fixed-point 3-16
floating-point 3-22
prescribed group delay 2-37
properties 1-13
abbreviating names 4-6
accessing, command for 11-53
data formats
quantized filters 6-11, 7-8
setting 11-147
FilterStructure 10-12
Mode 10-5
MultiplicandFormat, quantized FFT 10-48
MultiplicandFormat, quantized filter 10-32
NumberOfSections, quantized filters 10-32
NumberOfStages, quantized FFTs 10-48
OutputFormat, quantized FFTs 10-49
OutputFormat, quantized filters 10-33
OverflowMode 10-6
QuantizedCoefficients 10-33
Radix 10-49
ReferenceCoefficients 6-7, 10-34
referencing directly 4-9
retrieving 4-5
function for 4-8
retrieving by direct property referencing 4-9
RoundMode 10-8
Index
ScaleValues 10-42
setting 4-5
setting, function for 11-140
StatesPerSection 10-44
SumFormat, quantized FFTs 10-50
SumFormat, quantized filters 10-44
property values
abbreviating 4-8
quantized FFTs 7-6
quantized filters 6-6
quantizers 5-4
Q
QFFT objects 7-2
qfilt 11-115
Qfilt objects 1-13
See also quantized filters
quantization
precision, quantized FFTs 7-8
precision, quantized filters 6-11
quantization errors 2-59
quantization level 3-23
quantization mode in FDATool 9-3
quantization, errors during 2-59
quantized FFT properties
CoefficientFormat 10-47
InputFormat 10-47
MultiplicandFormat 10-48
NumberOfStages 10-48
OutputFormat 10-49
ProductFormat 10-49
Radix 10-49
ScaleValues 10-49
SumFormat 10-50
quantized FFTs 7-2
addition 10-50
basic properties 7-6
computing 7-10
constructing 7-3
data formats 7-7
input formats 10-47
multiplicand formats 10-48
output formats 10-49
product formats 10-49
properties 10-47
property values 7-6
scaling 10-49
stages, number of 10-48
quantized filter formats
inputs 10-32
operands 10-32
outputs 10-33
products 10-33
sums 10-44, 10-50
quantized filter properties
CoefficientFormat 10-11
FilterStructure 10-12
InputFormat 10-32
NumberOfSections 10-32
OperandFormat 10-32
OutputFormat 10-33
ProductFormat 10-33
QuantizedCoefficients 10-33
ReferenceCoefficients 10-34
setting 4-9
setting, command for 11-140
quantized filter properties, changing in FDATool
9-11
quantized filters
accessing properties 11-53
addition 10-44
analysis with 8-1
applications 8-1
I-7
Index
architecture 10-12
arithmetic effects 8-22
basic properties 6-6
cascaded sections 10-38
coefficients, accessing for multiple sections
4-14
coefficients, accessing for single section 4-13
coefficients, overflows 11-95
coefficients, quantized 10-33
coefficients, reference 10-34
constructing 6-3
function for 11-115
data formats 6-9, 6-11, 7-8
data formats, setting all 6-10, 7-8
defining 4-3
direct form FIR 10-20
direct form FIR transposed 10-21
direct form symmetric FIR 10-30
examples 4-13
exponent length 10-4
filter banks 8-16
filter types 3-7
filtering data 6-14, 11-42
finite impulse response 10-20, 10-21
floating point 10-4
fraction length 10-4
frequency response 11-50
noise loading method 3-12
getting properties 4-8
impulse response 11-73
lattice allpass 10-22
lattice AR 10-26
lattice ARMA 10-28
lattice coupled-allpass 10-22, 10-24
lattice coupled-allpass power complementary
10-25
lattice MA maximum phase 10-23
I-8
lattice MA minimum phase 10-27
limit cycles 3-13
multiple sections, specifying coefficients 10-42
table 10-38
normalizing 11-95
objects 4-3
overflow handling 10-6
overflows, logging 6-14
precision, setting 11-147
property values 6-6
Qfilt objects 1-13
real coefficients 11-81
reference coefficients 6-7
reference filter 10-34
rounding, property for 10-8
scaling 10-42
second-order sections 6-4
sections, number of 10-32
setting data formats 11-147
specifying 10-34
state vectors 11-45
states 10-44
structures 10-12
symmetric FIR 10-14
topology 6-8
word length 10-4
zero-pole plots 11-162
quantized filters properties
getting 4-9
ScaleValues 10-42
specifying, command for 11-115
StatesPerSection 10-44
SumFormat 10-44
quantized inverse FFTs
computing 7-10
QuantizedCoefficients property 10-33
quantizers
Index
constructing 5-3
construction
shortcuts 5-5
data types
property for 10-5
properties
Format 10-3
Max 10-5
Min 10-5
Mode 10-5
NOperations 10-6
NOverflows 10-6
NUnderflows 10-6
OverflowMode 10-7
property names, leaving out 5-5
RoundMode 10-8
settable 5-4
property values 5-4
unit 5-2
unity 5-3
quantizing filters in FDATool 9-10
R
Radix 10-49
radix point 3-3
interpretation 3-16
range
fixed-point 3-16
floating-point 3-21
range notation xviii
reference coefficients
specifying 10-34
reference filters
quantized filters, specifying from 6-4
specifying 6-7
ReferenceCoefficients property 10-34
Remez exchange algorithm 2-6
robust filters 2-59
rounding
property for 10-8
RoundMode property 10-8
S
saturate property value 9-10
ScaleValues property 10-42
interpreting 10-43
scaling
2 norm 2-4
implementing for quantized filters 10-43
infinity norm 2-4
Lp norm 2-4
quantized filters 10-42
scientific notation 3-18
second-order sections
normalizing 10-42
set 11-140
set quantization parameters dialog 9-6
setbits 11-147
setting filter properties in FDATool 9-6
setting properties 4-5
set function 11-140
dot notation 11-143
sign bits 3-19
single precision 3-20
solution, minimax 2-4
starting FDATool 9-3
state vectors 11-45
state-space filters 10-29
StatesPerSection property 10-44
structure-like referencing 4-9
SumFormat property
quantized FFTs 10-50
I-9
Index
quantized filters 10-44
sums, data format for
quantized FFTs 10-50
quantized filters 10-44
symmetricfir 10-30
syntax, ellipses (...) xviii
T
toolbox
getting started 1-15
topology 6-8
twiddle factors 10-47
two’s complement arithmetic 3-15
U
underflow indicator 9-17
underflow, checking for 9-17
underflows 3-22
unit quantizers 5-2
unity quantizers 5-3
V
vectors, indexing of xviii
W
word length
fixed-point 3-3
floating-point 3-18
limitations 10-4
setting 10-4
all formats 11-147
wrap property value 9-10
I-10
Z
zero-pole plots 11-162
zplane 11-162
plotting options 11-162
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement