Analog and Digital Signals and Systems
R.K. Rao Yarlagadda
Analog and Digital Signals and Systems
13
R.K. Rao Yar...

Author:
R. K. Rao Yarlagadda

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Analog and Digital Signals and Systems

R.K. Rao Yarlagadda

Analog and Digital Signals and Systems

13

R.K. Rao Yarlagadda School of Electrical & Computer Engineering Oklahoma State University Stillwater OK 74078-6028 202 Engineering South USA [email protected]

ISBN 978-1-4419-0033-3 e-ISBN 978-1-4419-0034-0 DOI 10.1007/978-1-4419-0034-0 Springer New York Dordrecht Heidelberg London Library of Congress Control Number: 2009929744 # Springer ScienceþBusiness Media, LLC 2010 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer ScienceþBusiness Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed on acid-free paper Springer is part of Springer ScienceþBusiness Media (www.springer.com)

This book is dedicated to my wife Marceil, children, Tammy Bardwell, Ryan Yarlagadda and Travis Yarlagadda and their families

Note to Instructors

The solutions manual can be located on the book’s webpage http://www/ springer.com/engineering/cirucitsþ %26þsystems/bok/978-1-4419-0033-3

vii

Preface

This book presents a systematic, comprehensive treatment of analog and discrete signal analysis and synthesis and an introduction to analog communication theory. This evolved from my 40 years of teaching at Oklahoma State University (OSU). It is based on three courses, Signal Analysis (a second semester junior level course), Active Filters (a first semester senior level course), and Digital signal processing (a second semester senior level course). I have taught these courses a number of times using this material along with existing texts. The references for the books and journals (over 160 references) are listed in the bibliography section. At the undergraduate level, most signal analysis courses do not require probability theory. Only, a very small portion of this topic is included here. I emphasized the basics in the book with simple mathematics and the sophistication is minimal. Theorem-proof type of material is not emphasized. The book uses the following model: 1. Learn basics 2. Check the work using bench marks 3. Use software to see if the results are accurate The book provides detailed examples (over 400) with applications. A threenumber system is used consisting of chapter number – section number – example or problem number, thus allowing the student to quickly identify the related material in the appropriate section of the book. The book includes well over 400 homework problems. Problem numbers are identified using the above three-number system. Hints are provided wherever additional details may be needed and may not have been given in the main part of the text. A detailed solution manual will be available from the publisher for the instructors.

Summary of the Chapters This book starts with an introductory chapter that includes most of the basic material that a junior in electrical engineering had in the beginning classes. For those who have forgotten, or have not seen the material recently, it gives enough

ix

x

background to follow the text. The topics in this chapter include singularity functions, periodic functions, and others. Chapter 2 deals with convolution and correlation of periodic and aperiodic functions. Chapter 3 deals with approximating a function by using a set of basis functions, referred to as the generalized Fourier series expansion. From these concepts, the three basic Fourier series expansions are derived. The discussion includes detailed discussion on the operational properties of the Fourier series and their convergence. Chapter 4 deals with Fourier transform theory derived from the Fourier series. Fourier series and transforms are the bases to this text. Considerable material in the book is based on these topics. Chapter 5 deals with the relatives of the Fourier transforms, including Laplace, cosine and sine, Hartley and Hilbert transforms. Chapter 6 deals with basic systems analysis that includes linear timeinvariant systems, stability concepts, impulse response, transfer functions, linear and nonlinear systems, and very simple filter circuits and concepts. Chapter 7 starts with the Bode plots and later deals with approximations using classical analog Butterworth, Chebyshev, and Bessel filter functions. Design techniques, based on both amplitude and phase based, are discussed. Last part of this chapter deals with analysis and synthesis of active filter circuits. Examples of basic low-pass, high-pass, band-pass, band elimination, and delay line filters are included. Chapter 8 builds a bridge to go from the continuous-time to discrete-time analysis by starting with sampling theory and the Fourier transform of the ideally sampled signals. Bulk of this chapter deals with discrete basis functions, discrete-time Fourier series, discrete-time Fourier transform (DTFT), and the discrete Fourier transform (DFT). Chapter 9 deals with fast implementations of the DFT, discrete convolution, and correlation. Second part of the chapter deals the z-transforms and their use in the design of discrete-data systems. Digital filter designs based on impulse invariance and bilinear transformations are presented. The chapter ends with digital filter realizations. Chapter 10 presents an introduction to analog communication theory, which includes basic material on analog modulation, such as AM and FM, demodulation, and multiplexing. Pulse modulation methods are introduced. Appendix A reviews the basics on matrices; Appendix B gives a brief introduction on MATLAB; and Appendix C gives a list of useful formulae. The book concludes with a list of references and Author and Subject indexes.

Suggested Course Content Instructor is the final judge of what topics will best suit his or her class and in what depth. The suggestions given below are intended to serve as a guide only. The book permits flexibility in teaching analysis, synthesis of continuous-time and discrete-time systems, analog filters, digital signal processing, and an introduction to analog communications. The following table gives suggestions for courses.

Preface

Preface

xi

Topical Title

Related topics in chapters

One semester

(

One semester

Systems and analog filters

Chapters 4, 5*, 6, 7

One semester

( (

Chapters 4*, 6*, 8, 9

Two semesters *Partial coverage

)

Fundamentals of analog signals and systems

Chapters 1–4, 6

)

Introduction to digital signal processing

)

Signals and an introduction to analog communications

Chapters 1–4, 5*, 6, 8*, 10

Acknowledgements

The process of writing this book has taken me several years. I am indebted to all the students who have studied with me and taken classes from me. Education is a two-way street. The teachers learn from the students, as well as the students learn from the teachers. Writing a book is a learning process. Dr. Jack Cartinhour went through the material in the early stages of the text and helped me in completing the solution manual. His suggestions made the text better. I am deeply indebted to him. Dr. George Scheets used an earlier version of this book in his signal analysis and communications theory class. Dr. Martin Hagan has reviewed a chapter. Their comments were incorporated into the manuscript. Beau Lacefield did most of the artwork in the manuscript. Vijay Venkataraman and Wen Fung Leong have gone through some of the chapters and their suggestions have been incorporated. In addition, Vijay and Wen have provided some of the MATLAB programs and artwork. I appreciated Vijay’s help in formatting the final version of the manuscript. An old adage of the uncertainty principle is, no matter how many times the author goes through the text, mistakes will remain. I sincerely appreciate all the support provided by Springer. Thanks to Alex Greene. He believed in me to complete this project. I appreciated the patience and support of Katie Chen. Thanks to Shanty Jaganathan and her associates of Integra-India. They have been helpful and gracious in the editorial process. Dr. Keith Teague, Head, School of Electrical and Computer Engineering at Oklahoma State University has been very supportive of this project and I appreciated his encouragement. Finally, the time spent on this book is the time taken away from my wife Marceil, children Tammy, Ryan and Travis and my grandchildren. Without my family’s understanding, I could not have completed this book. Oklahoma, USA

R.K. Rao Yarlagadda

xiii

Contents

1

Basic Concepts in Signals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Introduction to the Book and Signals . . . . . . . . . . . . . . . . . . . . 1.1.1 Different Ways of Looking at a Signal . . . . . . . . . . . . . . 1.1.2 Continuous-Time and Discrete-Time Signals . . . . . . . . . 1.1.3 Analog Versus Digital Signal Processing . . . . . . . . . . . . 1.1.4 Examples of Simple Functions . . . . . . . . . . . . . . . . . . . . 1.2 Useful Signal Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 Time Shifting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.2 Time Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.3 Time Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.4 Amplitude Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.5 Simple Symmetries: Even and Odd Functions . . . . . . . . 1.2.6 Products of Even and Odd Functions . . . . . . . . . . . . . . . 1.2.7 Signum (or sgn) Function . . . . . . . . . . . . . . . . . . . . . . . . 1.2.8 Sinc and Sinc2 Functions. . . . . . . . . . . . . . . . . . . . . . . . . 1.2.9 Sine Integral Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Derivatives and Integrals of Functions . . . . . . . . . . . . . . . . . . . 1.3.1 Integrals of Functions with Symmetries . . . . . . . . . . . . . 1.3.2 Useful Functions from Unit Step Function . . . . . . . . . . 1.3.3 Leibniz’s Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.4 Interchange of a Derivative and an Integral . . . . . . . . . . 1.3.5 Interchange of Integrals. . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Singularity Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 Unit Impulse as the Limit of a Sequence. . . . . . . . . . . . . 1.4.2 Step Function and the Impulse Function . . . . . . . . . . . . 1.4.3 Functions of Generalized Functions . . . . . . . . . . . . . . . . 1.4.4 Functions of Impulse Functions . . . . . . . . . . . . . . . . . . . 1.4.5 Functions of Step Functions . . . . . . . . . . . . . . . . . . . . . . 1.5 Signal Classification Based on Integrals . . . . . . . . . . . . . . . . . . 1.5.1 Effects of Operations on Signals . . . . . . . . . . . . . . . . . . . 1.5.2 Periodic Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.3 Sum of Two Periodic Functions . . . . . . . . . . . . . . . . . . . 1.6 Complex Numbers, Periodic, and Symmetric Periodic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Complex Numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 1 1 3 5 6 8 8 8 8 8 9 9 10 10 10 11 12 12 13 13 13 14 15 16 17 18 19 19 21 21 23 24 25 xv

xvi

2

3

Contents

1.6.2 Complex Periodic Functions . . . . . . . . . . . . . . . . . . . . . . 1.6.3 Functions of Periodic Functions . . . . . . . . . . . . . . . . . . . 1.6.4 Periodic Functions with Additional Symmetries. . . . . . . 1.7 Examples of Probability Density Functions and their Moments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.8 Generation of Periodic Functions from Aperiodic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9 Decibel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.10 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27 27 28

Convolution and Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 Scalar Product and Norm . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Properties of the Convolution Integral . . . . . . . . . . . . . . 2.2.2 Existence of the Convolution Integral. . . . . . . . . . . . . . . 2.3 Interesting Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Convolution and Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 Repeated Convolution and the Central Limit Theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Deconvolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 Convolution Involving Periodic and Aperiodic Functions . . . . 2.5.1 Convolution of a Periodic Function with an Aperiodic Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.2 Convolution of Two Periodic Functions. . . . . . . . . . . . . 2.6 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.1 Basic Properties of Cross-Correlation Functions . . . . . . 2.6.2 Cross-Correlation and Convolution . . . . . . . . . . . . . . . . 2.6.3 Bounds on the Cross-Correlation Functions . . . . . . . . . 2.6.4 Quantitative Measures of Cross-Correlation . . . . . . . . . 2.7 Autocorrelation Functions of Energy Signals . . . . . . . . . . . . . . 2.8 Cross- and Autocorrelation of Periodic Functions . . . . . . . . . . 2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39 39 40 41 41 44 44 50

Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Orthogonal Basis Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Gram–Schmidt Orthogonalization . . . . . . . . . . . . . . . . . 3.3 Approximation Measures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 Computation of c[k] Based on Partials . . . . . . . . . . . . . . 3.3.2 Computation of c[k] Using the Method of Perfect Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.3 Parseval’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.1 Complex Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.2 Trigonometric Fourier Series . . . . . . . . . . . . . . . . . . . . . 3.4.3 Complex F-series and the Trigonometric F-series Coefficients-Relations . . . . . . . . . . . . . . . . . . . . . . . . . . .

71 71 72 74 75 77

29 31 32 34 35

52 53 54 54 55 56 57 57 58 59 63 65 68 68

77 78 80 80 83 83

Contents

xvii

3.4.4 3.4.5 3.4.6

4

Harmonic Form of Trigonometric Fourier Series. . . . . Parseval’s Theorem Revisited . . . . . . . . . . . . . . . . . . . . Advantages and Disadvantages of the Three Forms of Fourier Series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Fourier Series of Functions with Simple Symmetries. . . . . . . . 3.5.1 Simplification of the Fourier Series Coefficient Integral . 3.6 Operational Properties of Fourier Series . . . . . . . . . . . . . . . . . 3.6.1 Principle of Superposition . . . . . . . . . . . . . . . . . . . . . . . 3.6.2 Time Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.3 Time and Frequency Scaling . . . . . . . . . . . . . . . . . . . . . 3.6.4 Fourier Series Using Derivatives . . . . . . . . . . . . . . . . . . 3.6.5 Bounds and Rates of Fourier Series Convergence by the Derivative Method . . . . . . . . . . . . . . . . . . . . . . 3.6.6 Integral of a Function and Its Fourier Series . . . . . . . . 3.6.7 Modulation in Time. . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.8 Multiplication in Time. . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.9 Frequency Modulation . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.10 Central Ordinate Theorems . . . . . . . . . . . . . . . . . . . . . . 3.6.11 Plancherel’s Relation (or Theorem). . . . . . . . . . . . . . . . 3.6.12 Power Spectral Analysis . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Convergence of the Fourier Series and the Gibbs Phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.1 Fourier’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.2 Gibbs Phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.3 Spectral Window Smoothing. . . . . . . . . . . . . . . . . . . . . 3.8 Fourier Series Expansion of Periodic Functions with Special Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.8.1 Half-Wave Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . 3.8.2 Quarter-Wave Symmetry. . . . . . . . . . . . . . . . . . . . . . . 3.8.3 Even Quarter-Wave Symmetry . . . . . . . . . . . . . . . . . . 3.8.4 Odd Quarter-Wave Symmetry. . . . . . . . . . . . . . . . . . . 3.8.5 Hidden Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.9 Half-Range Series Expansions . . . . . . . . . . . . . . . . . . . . . . . . . 3.10 Fourier Series Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.11 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

83 84

100 100 102 102 102 103 103 104 104 106

Fourier Transform Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Fourier Series to Fourier Integral . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Amplitude and Phase Spectra . . . . . . . . . . . . . . . . . . . . . 4.2.2 Bandwidth-Simplistic Ideas . . . . . . . . . . . . . . . . . . . . . . . 4.3 Fourier Transform Theorems, Part 1. . . . . . . . . . . . . . . . . . . . . 4.3.1 Rayleigh’s Energy Theorem . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Superposition Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.3 Time Delay Theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.4 Scale Change Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.5 Symmetry or Duality Theorem . . . . . . . . . . . . . . . . . . . . 4.3.6 Fourier Central Ordinate Theorems . . . . . . . . . . . . . . . .

109 109 109 112 114 114 114 115 116 116 118 119

85 85 86 87 87 87 88 89 91 93 93 94 95 95 95 95 96 96 97 99

xviii

Contents

4.4

Fourier Transform Theorems, Part 2 . . . . . . . . . . . . . . . . . . . . 4.4.1 Frequency Translation Theorem . . . . . . . . . . . . . . . . . 4.4.2 Modulation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.3 Fourier Transforms of Periodic and Some Special Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.4 Time Differentiation Theorem . . . . . . . . . . . . . . . . . . 4.4.5 Times-t Property: Frequency Differentiation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.6 Initial Value Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.7 Integration Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Convolution and Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 Convolution in Time . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.2 Proof of the Integration Theorem . . . . . . . . . . . . . . . . 4.5.3 Multiplication Theorem (Convolution in Frequency) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.4 Energy Spectral Density . . . . . . . . . . . . . . . . . . . . . . . 4.6 Autocorrelation and Cross-Correlation . . . . . . . . . . . . . . . . . . 4.6.1 Power Spectral Density . . . . . . . . . . . . . . . . . . . . . . . . 4.7 Bandwidth of a Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.1 Measures Based on Areas of the Time and Frequency Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.2 Measures Based on Moments . . . . . . . . . . . . . . . . . . . 4.7.3 Uncertainty Principle in Fourier Analysis. . . . . . . . . . 4.8 Moments and the Fourier Transform . . . . . . . . . . . . . . . . . . . 4.9 Bounds on the Fourier Transform . . . . . . . . . . . . . . . . . . . . . . 4.10 Poisson’s Summation Formula . . . . . . . . . . . . . . . . . . . . . . . . 4.11 Interesting Examples and a Short Fourier Transform Table . . 4.11.1 Raised-Cosine Pulse Function. . . . . . . . . . . . . . . . . . . 4.12 Tables of Fourier Transforms Properties and Pairs. . . . . . . . . 4.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Relatives of Fourier Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Fourier Cosine and Sine Transforms . . . . . . . . . . . . . . . . . . . . . 5.3 Hartley Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.1 Region of Convergence (ROC) . . . . . . . . . . . . . . . . . . . . 5.4.2 Inverse Transform of Two-Sided Laplace Transform. . . 5.4.3 Region of Convergence (ROC) of Rational Functions – Properties. . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Basic Two-Sided Laplace Transform Theorems . . . . . . . . . . . . 5.5.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.2 Time Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.3 Shift in s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.4 Time Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.5 Time Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.6 Differentiation in Time . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.7 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.8 Convolution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

119 120 120 121 124 126 128 128 129 129 132 133 135 136 138 139 139 140 141 143 144 145 145 146 147 147 147 155 155 156 159 161 163 164 165 165 165 165 165 165 166 166 166 166

Contents

xix

5.6

One-Sided Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.1 Properties of the One-Sided Laplace Transform . . . . . 5.6.2 Comments on the Properties (or Theorems) of Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . 5.7 Rational Transform Functions and Inverse Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.7.1 Rational Functions, Poles, and Zeros . . . . . . . . . . . . . 5.7.2 Return to the Initial and Final Value Theorems and Their Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.8 Solutions of Constant Coefficient Differential Equations Using Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.8.1 Inverse Laplace Transforms . . . . . . . . . . . . . . . . . . . . 5.8.2 Partial Fraction Expansions . . . . . . . . . . . . . . . . . . . . 5.9 Relationship Between Laplace Transforms and Other Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.9.1 Laplace Transforms and Fourier Transforms . . . . . . . 5.9.2 Hartley Transforms and Laplace Transforms . . . . . . . 5.10 Hilbert Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.10.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.10.2 Hilbert Transform of Signals with Non-overlapping Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.10.3 Analytic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.11 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6

Systems and Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Linear Systems, an Introduction . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 Two-Terminal Component Equations . . . . . . . . . . . . . . 6.3.2 Kirchhoff’s Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Time-Invariant and Time-Varying Systems . . . . . . . . . . . . . . . . 6.5 Impulse Response. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.1 Eigenfunctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.2 Bounded-Input/Bounded-Output (BIBO) Stability . . . . 6.5.3 Routh–Hurwitz Criterion (R–H criterion) . . . . . . . . . . . 6.5.4 Eigenfunctions in the Fourier Domain . . . . . . . . . . . . . . 6.6 Step Response. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7 Distortionless Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7.1 Group Delay and Phase Delay . . . . . . . . . . . . . . . . . . . . 6.8 System Bandwidth Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.8.1 Bandwidth Measures Using the Impulse Response hðtÞ and Its Transform Hðj!Þ . . . . . . . . . . . . . . . . . . . . . 6.8.2 Half-Power or 3 dB Bandwidth. . . . . . . . . . . . . . . . . . . . 6.8.3 Equivalent Bandwidth or Noise Bandwidth . . . . . . . . . . 6.8.4 Root Mean-Squared (RMS) Bandwidth . . . . . . . . . . . . . 6.9 Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.9.1 Distortion Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . .

166 167 167 174 175 176 178 179 179 183 184 185 186 186 188 189 190 190 193 193 193 194 195 197 198 199 202 202 203 206 208 213 213 216 216 217 217 218 219 220

xx

Contents

6.9.2 Output Fourier-Transform of a Nonlinear System . . . 6.9.3 Linearization of Nonlinear System Functions . . . . . . 6.10 Ideal Filters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.10.1 Low-Pass, High-Pass, Band-Pass, and Band-Elimination Filters . . . . . . . . . . . . . . . . . . . . . . . 6.11 Real and Imaginary Parts of the Fourier Transform of a Causal Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.11.1 Relationship Between Real and Imaginary Parts of the Fourier Transform of a Causal Function Using Hilbert Transform . . . . . . . . . . . . . . . . . . . . . . . 6.11.2 Amplitude Spectrum jHðj!Þj to a Minimum Phase Function HðsÞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12 More on Filters: Source and Load Impedances . . . . . . . . . . . . 6.12.1 Simple Low-Pass Filters . . . . . . . . . . . . . . . . . . . . . . . 6.12.2 Simple High-Pass Filters . . . . . . . . . . . . . . . . . . . . . . . 6.12.3 Simple Band-Pass Filters . . . . . . . . . . . . . . . . . . . . . . . 6.12.4 Simple Band-Elimination or Band-Reject or Notch Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12.5 Maximum Power Transfer. . . . . . . . . . . . . . . . . . . . . . 6.12.6 A Simple Delay Line Circuit . . . . . . . . . . . . . . . . . . . . 6.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Approximations and Filter Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Bode Plots. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Gain and Phase Margins . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 Classical Analog Filter Functions . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 Amplitude-Based Design . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.2 Butterworth Approximations . . . . . . . . . . . . . . . . . . . . . 7.3.3 Chebyshev (Tschebyscheff) Approximations . . . . . . . . . 7.4 Phase-Based Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.1 Maximally Flat Delay Approximation . . . . . . . . . . . . . . 7.4.2 Group Delay of Bessel Functions . . . . . . . . . . . . . . . . . . 7.5 Frequency Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.1 Normalized Low-Pass to High-Pass Transformation . . . 7.5.2 Normalized Low-Pass to Band-Pass Transformation. . . 7.5.3 Normalized Low-Pass to Band-Elimination Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.4 Conversions of Specifications from Low-Pass, High-Pass, Band-Pass, and Band Elimination Filters to Normalized Low-Pass Filters . . . . . . . . . . . . . . . . . . . 7.6 Multi-terminal Components. . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6.1 Two-Port Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6.2 Circuit Analysis Involving Multi-terminal Components and Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6.3 Controlled Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.7 Active Filter Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.7.1 Operational Amplifiers, an Introduction . . . . . . . . . . . .

220 221 221 222 227

228 229 229 231 231 233 235 238 239 239 239 243 243 246 252 254 254 255 257 262 263 264 266 266 268 268

270 273 273 277 278 279 279

Contents

xxi

8

7.7.2 Inverting Operational Amplifier Circuits . . . . . . . . . . . 7.7.3 Non-inverting Operational Amplifier Circuits . . . . . . . 7.7.4 Simple Second-Order Low-Pass and All-Pass Circuits. .. 7.8 Gain Constant Adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.9 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.9.1 Amplitude (or Magnitude) Scaling, RLC Circuits . . . . 7.9.2 Frequency Scaling, RLC Circuits . . . . . . . . . . . . . . . . . 7.9.3 Amplitude and Frequency Scaling in Active Filters . . . 7.9.4 Delay Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.10 RC–CR Transformations: Low-Pass to High-Pass Circuits . . 7.11 Band-Pass, Band-Elimination and Biquad Filters . . . . . . . . . . 7.12 Sensitivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

280 282 284 285 287 287 288 288 290 292 294 298 301 301

Discrete-Time Signals and Their Fourier Transforms . . . . . . . . . . . . . 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Sampling of a Signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1 Ideal Sampling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2 Uniform Low-Pass Sampling or the Nyquist Low-Pass Sampling Theorem . . . . . . . . . . . . . . . . . . . . . 8.2.3 Interpolation Formula and the Generalized Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.4 Problems Associated with Sampling Below the Nyquist Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.5 Flat Top Sampling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.6 Uniform Band-Pass Sampling Theorem . . . . . . . . . . . . . 8.2.7 Equivalent continuous-time and discrete-time systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Basic Discrete-Time (DT) Signals . . . . . . . . . . . . . . . . . . . . . . . 8.3.1 Operations on a Discrete Signal . . . . . . . . . . . . . . . . . . . 8.3.2 Discrete-Time Convolution and Correlation . . . . . . . . . 8.3.3 Finite duration, right-sided, left-sided, two-sided, and causal sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.4 Discrete-Time Energy and Power Signals . . . . . . . . . . . . 8.4 Discrete-Time Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.1 Periodic Convolution of Two Sequences with the Same Period. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.2 Parseval’s Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5 Discrete-Time Fourier Transforms . . . . . . . . . . . . . . . . . . . . . . 8.5.1 Discrete-Time Fourier Transforms (DTFTs) . . . . . . . . . 8.5.2 Discrete-Time Fourier Transforms of Real Signals with Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6 Properties of the Discrete-Time Fourier Transforms. . . . . . . . . 8.6.1 Periodic Nature of the Discrete-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.2 Superposition or Linearity . . . . . . . . . . . . . . . . . . . . . . . 8.6.3 Time Shift or Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.4 Modulation or Frequency Shifting . . . . . . . . . . . . . . . . .

311 311 312 312 314 317 319 322 324 325 325 327 329 330 330 332 334 334 335 335 336 339 339 340 341 341

xxii

Contents

8.6.5 Time Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.6 Differentiation in Frequency . . . . . . . . . . . . . . . . . . . 8.6.7 Differencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.8 Summation or Accumulation . . . . . . . . . . . . . . . . . . 8.6.9 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.10 Multiplication in Time. . . . . . . . . . . . . . . . . . . . . . . . 8.6.11 Parseval’s Identities . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.12 Central Ordinate Theorems . . . . . . . . . . . . . . . . . . . . 8.6.13 Simple Digital Encryption . . . . . . . . . . . . . . . . . . . . . 8.7 Tables of Discrete-Time Fourier Transform (DTFT) Properties and Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.8 Discrete-Time Fourier-transforms from Samples of the Continuous-Time Fourier-Transforms . . . . . . . . . . . . . . . . . . 8.9 Discrete Fourier Transforms (DFTs) . . . . . . . . . . . . . . . . . . . . 8.9.1 Matrix Representations of the DFT and the IDFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.9.2 Requirements for Direct Computation of the DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10 Discrete Fourier Transform Properties . . . . . . . . . . . . . . . . . . 8.10.1 DFTs and IDFTs of Real Sequences. . . . . . . . . . . . . 8.10.2 Linearity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.3 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.4 Time Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.5 Frequency Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.6 Even Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.7 Odd Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.8 Discrete-Time Convolution Theorem . . . . . . . . . . . . 8.10.9 Discrete-Frequency Convolution Theorem . . . . . . . . 8.10.10 Discrete-Time Correlation Theorem . . . . . . . . . . . . . 8.10.11 Parseval’s Identity or Theorem . . . . . . . . . . . . . . . . . 8.10.12 Zero Padding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.13 Signal Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.14 Decimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.11 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Discrete Data Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 Computation of Discrete Fourier Transforms (DFTs) . . . . . . . 9.2.1 Symbolic Diagrams in Discrete-Time Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2.2 Fast Fourier Transforms (FFTs). . . . . . . . . . . . . . . . . . . 9.3 DFT (FFT) Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.1 Hidden Periodicity in a Signal. . . . . . . . . . . . . . . . . . . . . 9.3.2 Convolution of Time-Limited Sequences . . . . . . . . . . . . 9.3.3 Correlation of Discrete Signals . . . . . . . . . . . . . . . . . . . . 9.3.4 Discrete Deconvolution. . . . . . . . . . . . . . . . . . . . . . . . . . 9.4 z-Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.1 Region of Convergence (ROC) . . . . . . . . . . . . . . . . . . . .

341 342 342 344 344 345 346 346 346 347 348 350 352 353 354 354 354 355 355 356 356 356 357 358 359 359 359 360 361 361 361 367 367 368 368 369 372 372 374 377 378 380 381

Contents

xxiii

9.4.2

z-Transform and the Discrete-Time Fourier Transform (DTFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . 384 9.5 Properties of the z-Transform . . . . . . . . . . . . . . . . . . . . . . . . . 384 9.5.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384 9.5.2 Time-Shifted Sequences . . . . . . . . . . . . . . . . . . . . . . . . 385 9.5.3 Time Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385 9.5.4 Multiplication by an Exponential . . . . . . . . . . . . . . . . 385 9.5.5 Multiplication by n . . . . . . . . . . . . . . . . . . . . . . . . . . . 386 9.5.6 Difference and Accumulation . . . . . . . . . . . . . . . . . . . 386 9.5.7 Convolution Theorem and the z-Transform . . . . . . . . 386 9.5.8 Correlation Theorem and the z-Transform . . . . . . . . . 387 9.5.9 Initial Value Theorem in the Discrete Domain . . . . . . 388 9.5.10 Final Value Theorem in the Discrete Domain . . . . . . 388 9.6 Tables of z-Transform Properties and Pairs. . . . . . . . . . . . . . . 389 9.7 Inverse z-Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390 9.7.1 Inversion Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390 9.7.2 Use of Transform Tables (Partial Fraction Expansion Method) . . . . . . . . . . . . . . . . . . . . . . . . . . . 391 9.7.3 Inverse z-Transforms by Power Series Expansion. . . . 394 9.8 The Unilateral or the One-Sided z-Transform . . . . . . . . . . . . . 395 9.8.1 Time-Shifting Property . . . . . . . . . . . . . . . . . . . . . . . . 395 9.9 Discrete-Data Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 9.9.1 Discrete-Time Transfer Functions. . . . . . . . . . . . . . . . 400 9.9.2 Schur–Cohn Stability Test. . . . . . . . . . . . . . . . . . . . . . 401 9.9.3 Bilinear Transformations. . . . . . . . . . . . . . . . . . . . . . . 401 9.10 Designs by the Time and Frequency Domain Criteria. . . . . . . 403 9.10.1 Impulse Invariance Method by Using the Time Domain Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405 9.10.2 Bilinear Transformation Method by Using the Frequency Domain Criterion . . . . . . . . . . . . . . . . . . . . 407 9.11 Finite Impulse Response (FIR) Filter Design . . . . . . . . . . . . . 410 9.11.1 Low-Pass FIR Filter Design . . . . . . . . . . . . . . . . . . . . 411 9.11.2 High-Pass, Band-Pass, and Band-Elimination FIR Filter Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413 9.11.3 Windows in Fourier Design. . . . . . . . . . . . . . . . . . . . . . 1416 .... 9.12 Digital Filter Realizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 9.12.1 Cascade Form of Realization . . . . . . . . . . . . . . . . . . . 422 9.12.2 Parallel Form of Realization . . . . . . . . . . . . . . . . . . . . 422 9.12.3 All-Pass Filter Realization. . . . . . . . . . . . . . . . . . . . . . 423 9.12.4 Digital Filter Transposed Structures . . . . . . . . . . . . . . 423 9.12.5 FIR Filter Realizations . . . . . . . . . . . . . . . . . . . . . . . . 423 9.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425 10

Analog Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

429

10.1 10.2

429 431 432 432

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Limiters and Mixers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2.1 Mixers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3 Linear Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xxiv

Contents

10.4 10.5

10.6

10.7

10.8

10.9 10.10 10.11

10.12

10.13

10.14

10.15

10.16

10.17

10.3.1 Double-Sideband (DSB) Modulation . . . . . . . . . . . 10.3.2 Demodulation of DSB Signals. . . . . . . . . . . . . . . . . Frequency Multipliers and Dividers. . . . . . . . . . . . . . . . . . . . Amplitude Modulation (AM). . . . . . . . . . . . . . . . . . . . . . . . . 10.5.1 Percentage Modulation . . . . . . . . . . . . . . . . . . . . . . 10.5.2 Bandwidth Requirements . . . . . . . . . . . . . . . . . . . . 10.5.3 Power and Efficiency of an Amplitude Modulated Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.5.4 Average Power Contained in an AM Signal . . . . . . Generation of AM Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.1 Square-Law Modulators . . . . . . . . . . . . . . . . . . . . . 10.6.2 Switching Modulators . . . . . . . . . . . . . . . . . . . . . . . 10.6.3 Balanced Modulators. . . . . . . . . . . . . . . . . . . . . . . . Demodulation of AM Signals . . . . . . . . . . . . . . . . . . . . . . . . 10.7.1 Rectifier Detector. . . . . . . . . . . . . . . . . . . . . . . . . . . 10.7.2 Coherent or a Synchronous Detector . . . . . . . . . . . 10.7.3 Square-Law Detector. . . . . . . . . . . . . . . . . . . . . . . . 10.7.4 Envelope Detector . . . . . . . . . . . . . . . . . . . . . . . . . . Asymmetric Sideband Signals . . . . . . . . . . . . . . . . . . . . . . . . 10.8.1 Single-Sideband Signals . . . . . . . . . . . . . . . . . . . . . . 10.8.2 Vestigial Sideband Modulated Signals . . . . . . . . . . 10.8.3 Demodulation of SSB and VSB Signals . . . . . . . . . 10.8.4 Non-coherent Demodulation of SSB. . . . . . . . . . . . 10.8.5 Phase-Shift Modulators and Demodulators . . . . . . Frequency Translation and Mixing . . . . . . . . . . . . . . . . . . . . Superheterodyne AM Receiver. . . . . . . . . . . . . . . . . . . . . . . . Angle Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.11.1 Narrowband (NB) Angle Modulation. . . . . . . . . . . 10.11.2 Generation of Angle Modulated Signals . . . . . . . . . Spectrum of an Angle Modulated Signal . . . . . . . . . . . . . . . . 10.12.1 Properties of Bessel Functions . . . . . . . . . . . . . . . . . 10.12.2 Power Content in an Angle Modulated Signal . . . . Demodulation of Angle Modulated Signals. . . . . . . . . . . . . . 10.13.1 Frequency Discriminators . . . . . . . . . . . . . . . . . . . . 10.13.2 Delay Lines as Differentiators . . . . . . . . . . . . . . . . . FM Receivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.14.1 Distortions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.14.2 Pre-emphasis and De-emphasis . . . . . . . . . . . . . . . . 10.14.3 Distortions Caused by Multipath Effect . . . . . . . . . Frequency-Division Multiplexing (FDM) . . . . . . . . . . . . . . . 10.15.1 Quadrature Amplitude Modulation (QAM) or Quadrature Multiplexing (QM). . . . . . . . . . . . . . 10.15.2 FM Stereo Multiplexing and the FM Radio . . . . . . Pulse Modulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.16.1 Pulse Amplitude Modulation (PAM) . . . . . . . . . . . 10.16.2 Problems with Pulse Modulations . . . . . . . . . . . . . . 10.16.3 Time-Division Multiplexing (TDM) . . . . . . . . . . . . Pulse Code Modulation (PCM) . . . . . . . . . . . . . . . . . . . . . . . 10.17.1 Quantization Process . . . . . . . . . . . . . . . . . . . . . . . . 10.17.2 More on Coding. . . . . . . . . . . . . . . . . . . . . . . . . . . .

432 433 435 437 438 438 439 440 441 441 441 442 443 443 443 444 444 446 446 447 448 449 449 450 453 455 458 459 460 461 463 465 465 467 468 468 469 470 471 472 473 474 475 475 477 478 478 480

Contents

xxv

10.17.3

Tradeoffs Between Channel Bandwidth and Signal-to-Quantization Noise Ratio . . . . . . . . . . . . 10.17.4 Digital Carrier Modulation . . . . . . . . . . . . . . . . . . . 10.18 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

481 482 484 484

Appendix A: Matrix Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 A.1 Matrix Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 A.2 Elements of Matrix Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . 490 A.2.1 Vector Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491 A.3 Solutions of Matrix Equations . . . . . . . . . . . . . . . . . . . . . . . . . 492 A.3.1 Determinants. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492 A.3.2 Cramer’s Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493 A.3.3 Rank of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494 A.4 Inverses of Matrices and Their Use in Determining the Solutions of a Set of Equations . . . . . . . . . . . . . . . . . . . . . . 495 A.5 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . 496 A.6 Singular Value Decomposition (SVD) . . . . . . . . . . . . . . . . . . . 500 A.7 Generalized Inverses of Matrices . . . . . . . . . . . . . . . . . . . . . . . 501 A.8 Over- and Underdetermined System of Equations . . . . . . . . . . 502 A.8.1 Least-Squares Solutions of Overdetermined System of Equations (m > n) . . . . . . . . . . . . . . . . . . . . 502 A.8.2 Least-Squares Solution of Underdetermined System of Equations (m n) . . . . . . . . . . . . . . . . . . . . 504 A.9 Numerical-Based Interpolations: Polynomial and Lagrange Interpolations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505 A.9.1 Polynomial Approximations . . . . . . . . . . . . . . . . . . . . . . . 505 ... A.9.2 Lagrange Interpolation Formula . . . . . . . . . . . . . . . . . . . 506 ... Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506 Appendix B: MATLAB1 for Digital Signal Processing . . . . . . . . . . . . . . . B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.2 Signal Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.3 Signal Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.4 Fast Fourier Transforms (FFTs) . . . . . . . . . . . . . . . . . . . . . . . B.5 Convolution of Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.6 Differentiation Using Numerical Methods . . . . . . . . . . . . . . . B.7 Fourier Series Computation . . . . . . . . . . . . . . . . . . . . . . . . . . B.8 Roots of Polynomials, Partial Fraction Expansions, PoleZero Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.8.1 Partial Fraction Expansions . . . . . . . . . . . . . . . . . . . . B.9 Bode Plots, Impulse and Step Responses . . . . . . . . . . . . . . . . B.9.1 Bode Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.9.2 Impulse and Step Responses . . . . . . . . . . . . . . . . . . . . B.10 Frequency Responses of Digital Filter Transfer Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.11 Introduction to the Construction of Simple MATLAB Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.12 Additional MATLAB Code. . . . . . . . . . . . . . . . . . . . . . . . . . .

509 509 509 511 511 513 515 515 517 518 518 518 518 520 520 521

xxvi

Contents

Appendix C: Mathematical Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.1 Trigonometric Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.2 Logarithms, Exponents and Complex Numbers . . . . . . . . . . . . C.3 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.4 Indefinite Integrals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.5 Definite Integrals and Useful Identities. . . . . . . . . . . . . . . . . . . C.6 Summation Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.7 Series Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.8 Special Constants and Factorials . . . . . . . . . . . . . . . . . . . . . . .

523 523 523 524 524 525 525 526 526

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

527

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

531

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

535

List of Tables

Table 1.4.1 Table 1.9.1 Table 1.9.2 Table 2.4.1 Table 2.6.1 Table 3.4.1 Table 3.10.1 Table 3.10.2 Table 4.12.1 Table 4.12.2 Table 5.6.1 Table 5.6.2 Table 5.8.1 Table 5.9.1 Table 5.10.1 Table 7.1.1 Table 7.4.1

Table 7.5.1 Table 7.7.1 Table 8.1.1 Table 8.2.1 Table 8.2.2 Table 8.3.1 Table 8.7.1 Table 8.7.2 Table 8.10.1 Table 9.1.1

Properties of the impulse function. . . . . . . . . . . . . . . . . . . Sound Power (loudness) Comparison . . . . . . . . . . . . . . . . Power ratios and their corresponding values in dB . . . . . . Properties of aperiodic convolution . . . . . . . . . . . . . . . . . Example 2.6.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Summary of the three Fourier series representations . . . . Symmetries of real periodic functions and their Fourier-series coefficients . . . . . . . . . . . . . . . . . . . . . . . . . Periodic functions and their Trigonometric Fourier Series . . Fourier transform properties. . . . . . . . . . . . . . . . . . . . . . . Fourier Transform Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . One-sided Laplace transform properties . . . . . . . . . . . . . . One-sided Laplace tranform pairs . . . . . . . . . . . . . . . . . . . Typical rational replace transforms and their inverses . . . One sided Laplace transforms and Fourier transforms. . . Hilbert transform pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . Formula for computing sensitivities . . . . . . . . . . . . . . . . . Normalized frequencies, ! ¼ !0 . Time delay and a loss table giving the normalized frequency ! at which the zero frequency delay and loss values deviate by specified amounts for Bessel filter functions . . . . . . . . . . . . . . . . . . Frequency transformations . . . . . . . . . . . . . . . . . . . . . . . . Guidelines for passive components . . . . . . . . . . . . . . . . . . Fourier representations of discrete-time and continuous-time signals . . . . . . . . . . . . . . . . . . . . . . . . . . . Common interpolation functions . . . . . . . . . . . . . . . . . . . Spectral occupancy of Xðjð! n!s ÞÞ; ! ¼ 2f; n ¼ 0; 1; 2; 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Properties of discrete convolution . . . . . . . . . . . . . . . . . . . Discrete-time Fourier transform (DTFT) properties . . . . Discrete-time Fourier transform (DTFT) pairs. . . . . . . . . Discrete Fourier transform (DFT) properties . . . . . . . . . . Discrete-time and continuous-time signals and their transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18 33 33 53 62 84 105 105 148 149 168 175 182 185 190 244

265 269 283 312 319 325 329 347 348 361 367

xxvii

xxviii

Table 9.2.1 Table 9.6.1 Table 9.6.2 Table 9.11.1 Table 9.11.2 Table 10.9.1 Table 10.12.1 Table 10.17.1 Table 10.17.2 Table 10.17.3 Table B.7.1

List of Tables

Properties of the function WN ¼ ejð2=NÞ . . . . . . . . . . . . . Z-transform properties . . . . . . . . . . . . . . . . . . . . . . . . . . . Z-transform pairs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ideal low-pass filter FIR coefficients with c ¼ =4. . . . . FIR Filter Coefficients for the Four Basic Filters. . . . . . . Inputs and outputs of the system in Fig. 10.9.1 . . . . . . . . Bessel function values . . . . . . . . . . . . . . . . . . . . . . . . . . . . Quantization values and codes corresponding to Fig. 10.17.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Binary representation of quantized values . . . . . . . . . . . . Normal binary and Gray code representations for N¼8.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Amplitudes and phase angles of the harmonic Fourier series coefficients (Example B.7.1). . . . . . . . . . . . . . . . . . .

369 389 390 412 415 453 463 479 480 481 516

Chapter 1

Basic Concepts in Signals

1.1 Introduction to the Book and Signals the research and developments of many signal proThe primary goal of this book is to introduce the reader on the basic principles of signals and to provide tools thereby to deal with the analysis of analog and digital signals, either obtained naturally or by sampling analog signals, study the concepts of various transforming techniques, filtering analog and digital signals, and finally introduce the concepts of communicating analog signals using simple modulation techniques. The basic material in this book can be found in several books. See references at the end of the book. A signal is a pattern of some kind used to convey a message. Examples include smoke signals, a set of flags, traffic lights, speech, image, seismic signals, and many others. Smoke signals were used for conveying information that goes back before recorded history. Greeks and Romans used light beacons in the pre-Christian era. England employed a long chain of beacons to warn that Spanish Armada is approaching in the late sixteenth century. Around this time, the word signal came into use perceptible by sight, hearing, etc., conveying information. The present day signaling started with the invention of the Morse code in 1838. Since then, a variety of signals have been studied. These include the following inventions: Facsimile by Alexander Bain in 1843; telephone by Alexander Bell in 1876; wireless telegraph system by Gugliemo Marconi in 1897; transmission of speech signals via radio by Reginald Fessenden in 1905, invention and demonstration of television, the birth of television by Vladimir Zworykin in the 1920 s, and many others. In addition, the development of radar and television systems during World War II, proposition of satellite communication systems, demonstration of a laser in 1955, and

cessing techniques and their use in communication systems. Since the early stages of communications, research has exploded into several areas connected directly, or indirectly, to signal analysis and communications. Signal analysis has taken a significant role in medicine, for example, monitoring the heart beat, blood pressure and temperature of a patient, and vital signs of patients. Others include the study of weather phenomenon, the geological formations below the surface and deep in the ground and under the ocean floors for oil and gas exploration, mapping the underground surface using seismometers, and others. Researchers have concluded that computers are powerful and necessary that they need to be an integral part of any communication system, thus generating significant research in digital signal processing, development of Internet, research on HDTV, mobile and cellular telephone systems, and others. Defense industry has been one of the major organizations in advancing research in signal processing, coding, and transmission of data. Several research areas have surfaced in signals that include processing of speech, image, radar, seismic, medical, and other signals.

1.1.1 Different Ways of Looking at a Signal Consider a signal xðtÞ, a function representing a physical quantity, such as voltage, current, pressure, or any other variable with respect to a second variable t, such as time. The terms of interest are the time t and the signal xðtÞ. One of the main topics of

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_1, Ó Springer ScienceþBusiness Media, LLC 2010

1

2

1 Basic Concepts in Signals

this book is the analysis of signals. Webster’s dictionary defines the analysis as 1. Separation of a thing into the parts or elements of which it is composed. 2. An examination of a thing to determine its parts or elements. 3. A statement showing the results of such an examination. There are other definitions. In the following the three parts are considered using simple examples. Consider the sinusoidal function and its expansion using Euler’s formula: xðtÞ ¼ A0 cosðo0 t þ y0 Þ A0 jy0 jo0 t A0 jy0 jo0 t e e þ e e ¼ 2 2

(1:1:1)

¼ ReðA0 ejo0 t ejy0 Þ: In (1.1.1) A0 is assumed to be positive and real and A0 ejy0 is a complex number carrying the amplitude and phase angle of the sinusoidal function and is by definition the phasor representation of the given sinusoidal function. Some authors refer to this as phasor transform of the sinusoidal signal, as it transforms the time domain sinusoidal function to the complex frequency domain. A brief discussion on complex numbers is included later in Section 1.6. This signal can be described in another domain, i.e., such as the frequency domain. The amplitude is ðA0 =2Þ and the phase angles of y0 corresponding to the frequencies f0 ¼ o0 =2p Hz. In reality, only positive frequencies are available, but Euler’s formula in (1.1.1) dictates that both the positive and negative frequencies need to be identified as illustrated in Fig. 1.1.1a. This description is the twosided amplitude and phase line spectra of xðtÞ. Amplitudes are always positive and are located at f ¼ o0 =2p ¼ f0 Hz, symmetrically located around the zero frequency, i.e., with even symmetry. The phase spectrum consists of two angles y ¼ y0 corresponding to the positive and negative frequencies, respectively, with odd symmetry. Since ðtÞ is real, we can pictorially describe it by one- or two-sided amplitude and phase line spectra as shown in Fig. 1.1.1a,b,c,d. The following example illustrates the three steps.

Fig. 1.1.1 xðtÞ ¼ A0 cosðo0 t þ y0 Þ. (a) Two-sided amplitude spectrum, (b) two-sided phase spectrum, (c) one-sided amplitude spectrum, and (d) one-sided phase spectrum

Example 1.1.1 Express the following function in terms of a sum of cosine functions: xðtÞ ¼ A0 þ A1 cosðo1 t þ y1 Þ

(1:1:2) A2 cosðo2 tÞ A3 sinðo3 t þ y3 Þ; Ai > 0:

Solution: Using trigonometric relations to express each term in (1.1.2) in the form of Ai cosðoi t þ yi Þ results in xðtÞ ¼ A0 cosðð0Þt 180 Þ þ A1 cosðo1 t þ y1 Þ þ A2 cosðo2 t 180 Þ (1:1:3) þ A3 cosðo3 t þ y3 þ 90 Þ:

In the first and the third terms either 1808 or 1808 could be used, as the end result is the same. The two-sided line spectra of the function in (1.1.2) are shown in Fig. 1.1.2. How would one get the functions of the type shown in (1.1.3) for an arbitrary

1.1 Introduction to the Book and Signals

3

Fig. 1.1.2 (a) Two-sided amplitude spectra and (b) twosided phase spectra

Fig. 1.1.3 Speech . . .sho in . . .show ._male 2000 Samples @ 8000 samples per second. Printed with the permission from Hassan et al. (1994)

function? The sign and cosine functions are the building blocks of the Fourier series in Chapter 3 and later the Fourier transforms in Chapter 4. The function xðtÞ has four frequencies:

Example(s) 1.1.2 In this example several specific examples of interest are considered. In the first one, part of the time signal illustrating a male voice of speech in the sentence ‘‘. . .Show the rich lady out’’ is shown in Fig. 1.1.3. The speech signal is sampled at 8000 samples per second. There are three portions of the speech ‘‘/. . ./, /sh/, /o/’’ shown in the figure. The first part of the signal does not have any speech in it and the small amplitudes of the signal represent the noise in the tape recorder and/or in the room where the speech was recorded. It represents a random signal and can be described only by statistical means. Random signal analysis is not discussed in any detail in this book, as it requires knowledge of probability theory. The second part represents the phoneme ‘‘sh’’ that does not show any observable pattern. It is a time signal for a very short time and has finite energy. Power and energy signals are studied in Section 1.5. Third part of the figure represents the vowel ‘‘o,’’ showing a structure of (almost) periodic pulses for a short time. In this book, aperiodic or non-periodic signals with finite energy and periodic signals with finite average power will be studied. One goal is to come up with a model for each portion of a signal that can be transmitted and reconstructed at the receiver. Next three examples are from food industry. Small businesses are sprouting that use signal processing. For example, when we go to a grocery store we may like to buy a watermelon. It may not always be possible to judge the ripeness of the watermelon

f1 ð¼0Þ; f2 ; f3 ; f4 with amplitudes A0 ; A1 ; A2 ; A3 and phases 180o ; y1 ; 180o ; y3 þ 90o : Figure 1.1.2 illustrates pictorially the discrete locations of the frequencies, their amplitudes, and phases. The signal in (1.1.2) can be described by using the time domain function or in terms of frequencies. In the figures, o0 ð¼ 2pf Þ0 s in radians per second could have been used rather than f 0 s in Hz.&

1.1.2 Continuous-Time and Discrete-Time Signals A signal xðtÞ is a continuous-time signal if t is a continuous variable. It can take on any value in the continuous interval ða; bÞ. Continuous-time signal is an analog signal. If a function y½n is defined at discrete times, then it is a discrete-time signal, where n takes integer values. In Chapter 8 discrete-time signals will be studied by sampling the continuous signals at equal sampling intervals of ts seconds and write xðnts Þ; where n an integer. This is expressed by x½n xðnts Þ:

(1:1:4)

4

by outward characteristics such as external color, stem conditions, or just the way it looks. A sure way of looking at the quality is to cut the watermelon open and taste it before we buy it. This implies we break it first, which is destructive testing. Instead, we can use our grandmother’s procedure in selecting a watermelon. She uses her knuckles to send a signal into the watermelon. From the audio response of the watermelon she decides whether it is good or not based on her prior experience. We can simulate this by putting the watermelon on a stand, use a small hammer like device, give a slight tap on the watermelon, and record the response. A simplistic model of this is shown in Fig. 1.1.4. The responses can be categorized by studying the outputs of tasty watermelons. For an interesting research work on this topic, see Stone et al (1996). Image processing can be used to check for burned crusts, topping amount distribution, such as the location of pepperoni pizza slices, and others. For an interesting article on this subject, see Wagner (1983), which has several applications in the food industry. The next two examples are from the surface seismic signal analysis. In the first one, we use a source in the form of dynamite sticks representing a source, dig a small hole, and blow them in the hole. The ground responds to this input and the response is recorded using a seismometer and a tape recorder. The analysis of the recorded waveform can provide information about the underground cavities and pockets of oil and other important measures.

1 Basic Concepts in Signals

Geologists drill holes into the ground and a small slice of the core sample is used to measure the oil content by looking at the percentage of the area with dark spots on the slice, which is image processing. Another example of interest is measuring the distance from a ground station to an airplane. Send a signal with square wave pulses toward the airplane and when the signal hits it, a return signal is received at the ground station. A simple model is shown in Fig. 1.1.5. If we can measure the time between the time the signal left from the ground station and the time it returned, identified as T in the figure, we can determine the distance between the ground station and the target by the formula x ¼ 3ð108 Þ ðm=sÞ Tðsignal round trip time in secondsÞ=2:

(1:1:5)

The constant c ¼ 3ð108 Þ m/s is the speed of light. Radar and sonar signal processing are two important areas of signal processing applications. An exciting field of study is the biomedical area. We are well aware of a healthy heart that beats periodically, which can be seen from a record of an electrocardiogram (ECG). The ECG represents changes in the voltage potential due

(a)

(b)

Fig. 1.1.4 Watermelon responses to a tap

Fig. 1.1.5 (a) Radar range measurement and (b) transmitted and received filtered signals

1.1 Introduction to the Book and Signals

5

to electrochemical processes in the heart cells. Inferences can be made about the health of the heart under observation from the ECG. Another important example is the electroencephalogram (EEG), which measures the electrical activity in & brain. Signal processing is an important area that interests every engineer. Pattern recognition and classification is almost on top of the list. See, for example, O’Shaughnessy (1987) and Tou and Gonzalez (1974). For example, how do we distinguish two phonemes, one is a vowel and the other one is a consonant. A rough measure of frequency of a waveform with zero average value is the number of zero crossings per unit time. We will study in much more detail the frequency content in a signal later in terms of Fourier transforms in Chapter 4. Vowel sounds have lower frequency content than the consonants. A simple procedure to measure frequency in a speech segment is by computing the number of zero crossings in that segment. To differentiate a vowel from a consonant, set a threshold level for the frequency content for vowels and consonants that differentiate between vowels and consonants. If the frequency content is higher than this threshold, then the phoneme is a consonant. Otherwise, it is a vowel. If we like to distinguish one vowel from another we may need more than one measure. Vocal tract can be modeled as an acoustic tube with resonances, called formants. Two formant frequencies can be used to distinguish two vowels, say /u/ and / a/. See Problem 1.1.1. Two formant frequencies may not be enough to distinguish all the phonemes, especially if the signal is corrupted by noise. Consider a simple pattern classification problem with M prototype patterns z1 ; z2 ; . . . ; zM , where zi is a vector representing an ith pattern. For simplicity we assume that each pattern can be represented by a pair of numbers, say zi ¼ ðzi1 ; zi2 Þ; i = 1,2 . . . M and classify an arbitrary pattern x ¼ ðx1 ; x2 Þ to represent one of the prototype patterns. The Euclidean distance between a pattern x and the ith prototype pattern is defined by D i ¼ kx z i k ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ðx1 zi1 Þ2 þðx2 zi2 Þ2 :

(1:1:6)

A simple classifier is a minimum distance classifier that computes the distance from a pattern x of the

signal to be classified to the prototype of each class and assigns the unknown pattern to the class which it is closest to. That is, if Di 5Dj ; for all i 6¼ j, then we make the decision that x belongs to the ith prototype pattern. Ties are rare and if there are, they are resolved arbitrarily. In the above discussion two measures are assumed for each pattern. More measures give a better separation between classes. There are several issues that would interest a biomedical signal processor. These include removal of any noise present in the signals, such as 60-Hz interference picked up by the instruments, interference of the tools or meters that measure a parameter, and other signals that interfere with the desired signal. Finding the important facets in a signal, such as the frequency content, and many others is of interest. &

1.1.3 Analog Versus Digital Signal Processing Most signals are analog signals. Analog signal processing uses analog circuit elements, such as resistors, capacitors, inductors, and active components, such as operational amplifiers and non-linear devices. Since the inductors are made from magnetic material, they have inherent resistance and capacitance. This brings the quality of the components low. They tend to be bulky and their effectiveness is reduced. To alleviate this problem, active RC networks have been popular. Analog processing is a natural way to solve differential equations that describe physical systems, without having to resort to approximate solutions. Solutions are obtained in real time. In Chapter 10 we will see an example of analog encryption of a signal, wherein the analog speech is scrambled by the use of modulation techniques. Digital signal processing makes use of a special purpose computer, which has three basic elements, namely adders, multipliers, and memory for storage. Digital signal processing consists of numerical computations and there is no guarantee that the processing can be done in real time. To encrypt a set of numbers, these need to be converted into another set of numbers in the digital encryption scheme, for example. The complete encrypted signal is needed before it can be decrypted. In addition, if the input and the output signals are analog, then an

6

1 Basic Concepts in Signals

fx½ng ¼ f. . . ; 1; a; a2 ; . . . ; an ; . . .g or fx½ng #

In (1.1.8) reference points are not identified. In (1.1.9), the arrow below 1 is the 0 index term. The first term is assumed to be zero index term if there is no arrow and all the values of the sequence are zero for n50. We will come back to this in Chapter 8. A signal xðtÞ is a real signal if its value at some t is a real number. A complex signal xðtÞ consists of two real signals, x1 ðtÞ and x2 ðtÞ such that x(t) = x_1 pﬃﬃﬃﬃﬃﬃﬃﬃ (t) + jx2 ðtÞ; where j ¼ 1: The symbol j (or i) is used to represent the imaginary part. Interesting functions: a. P Function: The P function is centered at t0 with a width of t s shown in Fig. 1.1.6. It is not defined at t ¼ t0 t=2 and is symbolically expressed by

Π

2

n

fx½ng ¼ f. . . ; 0; 0; 0; 1; a; a ; . . . ; a ; . . .g;

(1:1:8)

t − t0 τ

Fig. 1.1.6 A P function

Yht t0 i 1;

jt t0 j5t=2 (1:1:10) t 0; otherwise ht t i xðtÞ; jt t j5t=2 0 0 xðtÞP ¼ t 0; otherwise

1.1.4 Examples of Simple Functions To begin the study we need to look into the concept of expressing a signal in terms of functions that can be generated in a laboratory. One such function is the sinusoidal function xðtÞ ¼ A0 cosðo0 t þ y0 Þ seen earlier in (1.1.1), where A0 ; o0 ; and y0 are some constants. A digital signal can be defined as a sequence in the forms n a ;n 0 x½n ¼ ; (1:1:7) 0; n50

(1:1:9)

¼ f1; a; a2 ; . . . ; an ; . . .g #

analog-to-digital converter (A/D), a digital processor, and a digital-to-analog converter (D/A) are needed to implement analog processing by digital means. Special purpose processor with A/D and D/A converters can be expensive. Digital approach has distinct advantages over analog approaches. Digital processor can be used to implement different versions of a system by changing the software on the processor. It has flexibility and repeatability. In the analog case, the system has to be redesigned every time the specifications are changed. Design components may not be available and may have to live with the component values within some tolerance. Components suffer from parameter variations due to room temperature, humidity, supply voltages, and many other aspects, such as aging, component failure. In a particular situation, many of the above problems need to be investigated before a complete decision can be made. Future appears to be more and more digital. Many of the digital signal processing filter designs are based on using analog filter designs. Learning both analog and digital signal processing is desired. Deterministic and random signals: Deterministic signals are specified for any given time. They can be modeled by a known function of time. There is no uncertainty with respect to any value at any time. For example, xðtÞ ¼ sinðtÞ is a deterministic signal. A random signal yðtÞ ¼ xðtÞ þ nðtÞ can take random values at any given time, as there is uncertainty about the noise signal nðtÞ. We can only describe such signals through statistical means, and the discussion on this topic will be minimal.

This P function is a deterministic signal. It is even, as P½t ¼ P½t and P½t0 t ¼ P½t t0 . Some use the symbol ‘‘rect’’ and ð1=tÞrectððt t0 Þ=tÞ is a rectangular pulse of width t s centered at t ¼ t0 with a height ð1=tÞ: b. L Function: The triangular function shown in Fig. 1.1.7 is defined by

L

ht t i 0

t

( ¼

0j 5 1 jtt t ; jt t0 j t

0;

otherwise

:

(1:1:11)

1.1 Introduction to the Book and Signals Λ

7

t − t0 τ

x(t ) = X 0 e–t / Tc u (t ), τ > 0

Fig. 1.1.9 Exponential decaying function

Fig. 1.1.7 A L function

Rectangular function defined in (1.1.10) has a width of t seconds, whereas the triangular function defined in (1.1.11) has a width of 2t s. The symbol ‘‘tri’’ is also used for a triangular function and triððt t0 Þ=tÞ describes the function in (1.1.11). c. Unit step function: It is shown in Fig. 1.1.8 and is uð t Þ ¼

1; t40 ðnot defined at t ¼ 0Þ: 0; t50

xðTc Þ ¼ e1 :37 or xðTc Þ ¼ :37X0 : X0

(1:1:14)

xðtÞ decreases to about 37% of its initial value in Tc s and is the time constant. It decreases to 2% in four time constants and Xð4Tc Þ :018X0 : A measure associated with exponential functions is the half life Th defined by

(1:1:12) 1 xðTh Þ ¼ X0 ; eTh =Tc 2 1 ¼ ; Th ¼ Tc loge ð2Þ ﬃ :693Tc : 2

(1:1:15)

Fig. 1.1.8 Unit step function

e. One-sided and two-sided exponentials: These are described by

The unit step function at t ¼ 0 can be defined explicitly as 0 or 1 or ½uð0þ Þ þ uð0 Þ=2=.5

x1 ðtÞ ¼

d. Exponential decaying function: A simple such function is ( X0 et=Tc ; t 0; Tc 40 : (1:1:13) xðtÞ ¼ 0; otherwise See Fig. 1.1.9. It has a special significance, as xðtÞ is the solution of a first-order differential equation. The constants X0 and Tc can take different values and

eat ; t 0 ; a4 0 0; t50 0; t 0 a40; x3 ðtÞ ¼ eajtj ; a40: x2 ðtÞ ¼ eat ; t50 (1:1:16)

x1 ðtÞ is the right-sided exponential, x2 ðtÞ is the leftsided exponential, and x3 ðtÞ is the two-sided exponential. These are sketched in Fig. 1.1.10. Using the unit step function, we have x2 ðtÞ ¼ x3 ðtÞuðtÞ and x1 ðtÞ ¼ x3 ðtÞuðtÞ. x2 (t )

Fig. 1.1.10 Exponential functions (a) x2 (t) = eat u(t),a > 0, (b) x1 ðtÞ ¼ eat uðtÞ; a > 0; and (c) x3 ðtÞ ¼ eajtj ; a > 0

(a)

x1 (t )

(b)

x3 (t )

(c)

8

1 Basic Concepts in Signals

1.2 Useful Signal Operations

1.2.2 Time Scaling

1.2.1 Time Shifting

The compression or expansion of a signal in time is known as time scaling. It is expanded in time if a5 1 and compressed in time if a > 1 in

Consider an arbitrary signal starting at t ¼ 0 shown in Fig. 1.2.1a. It can be shifted to the right as shown in Fig. 1.2.1b. It starts at time t ¼ a > 0, a delayed version of the one in Fig. 1.2.1a. Similarly it can be shifted to the left starting at time a shown in Fig. 1.2.1c. It is an advanced version of the one in Fig. 1.2.1a. We now have three functions: xðtÞ, xðt aÞ; and xðt þ aÞ with a > 0. The delayed and advanced unit step functions are

fðtÞ ¼ xðatÞ; a > 0:

(1:2:3)

Example 1.2.1 Illustrate the rectangular pulse functions P½t; P½2t; and P½t=2. Solution: These are shown in Fig. 1.2.2 and are of widths 1, (1/2), and 2, respectively. The pulse P½2t is a compressed version and the pulse P½t=2 is an & expanded version of the pulse function P½t.

1; t4a ; 0; t5a 1; t4 a uðt þ aÞ ¼ ; ða40Þ: 0; t5 a

uð t aÞ ¼

1.2.3 Time Reversal (1:2:1) If a ¼ 1 in (1.2.3), that is, fðtÞ ¼ xðtÞ, then the signal is time reversed (or folded).

From (1.1.16), the right-sided delayed exponential decaying function is x1 ðt tÞ ¼ eaðttÞ uðt tÞ; a > 0; t > 0:

x(t)

(1:2:2)

x(t + a)

x(t – a)

(a)

(b)

(c)

Fig. 1.2.1 (a) x(t), (b) x(t a), (c) x(t + a), a > 0

Fig. 1.2.2 Pulse functions

Example 1.2.2 Let x1 ðtÞ ¼ eat uðtÞ: Give its timereversed signal. Solution: The time-reversed signal of x1 ðtÞ is & x2 ðtÞ ¼ eat uðtÞ.

1.2.4 Amplitude Shift The amplitude shift of xðtÞ by a constant K is fðtÞ ¼ K þ xðtÞ. Combined operations: Some of the above signal operations can be combined into a general form. The signal yðtÞ ¼ xðat t0 Þ may be described by one of the two ways, namely:

Π[t]

Π[2t]

(a)

(b)

Π[t /2]

(c)

1.2 Useful Signal Operations

9

1. Time shift of t0 followed by time scaling by a. 2. Time scaling by (a) followed by time shift of ðt0 =aÞ.

equations in (1.2.6) and yð0Þ ¼ xð3Þ ¼ 0 and & yðt0 =aÞ ¼ yð3=2Þ ¼ 0.

These can be visualized by the following:

1: xðtÞ

shift scale ! vðtÞ ¼ xðt t0 Þ ! yðtÞ t ! t t0 t ! at (1:2:4)

Continuous-time even and odd functions satisfy

¼ vðatÞ ¼ xðat t0 Þ: scale shift 2: xðtÞ ! yðtÞ ! gðtÞ ¼ xðatÞ t ! at t ! t ðt0 =aÞ (1:2:5) ¼ gðt ðt0 =aÞÞ ¼ xðat t0 Þ:

Notation can be simplified by writing yðtÞ ¼ xðaðt ðt0 =aÞÞÞ. Noting that b ¼ at t0 is a linear equation in terms of two constants a and t0 , it follows: yð0Þ ¼ xðt0 Þ and yðt0 =aÞ ¼ xð0Þ:

1.2.5 Simple Symmetries: Even and Odd Functions

(1:2:6)

These two equations provide checks to verify the end result of the transformation. Following example illustrates some pitfalls in the order of time shifting and time scaling. Example 1.2.3 Derive the expression yðtÞ ¼ xð3t þ 2Þ assuming xðtÞ ¼ P½t=2.

xðtÞ ¼ xðtÞ xe ðtÞ, an even function, x ( t) = xðtÞ x0 ðtÞ, an odd function. (1.2.7) Examples of even and odd functions are shown in Fig. 1.2.3. The function cosðo0 tÞ is an even function and x0 ðtÞ ¼ sinðo0 tÞ is an odd function. An arbitrary real signal, xðtÞ; can be expressed in terms of its even and odd parts by xðtÞ ¼ xe ðtÞ þ x0 ðtÞ; xe ðtÞ ¼ ½ðxðtÞ þ xðtÞÞ=2; x0 ðtÞ

(1:2:8)

¼ ½ðxðtÞ xðtÞÞ=2: xe(t)

x0(t)

(a)

(b)

for

Solution: Using (1.2.4) with a ¼ 3 and t0 ¼ 2, we have tþ2 ; vðtÞ ¼xðt t0 Þ ¼ P 2 3t þ 2 t þ ð2=3Þ yðtÞ ¼vð3tÞ ¼ P ¼P : 2 2=3 Using (1.2.5), we have 3t ; gðtÞ ¼xðatÞ ¼ P 2 t0 2 yðtÞ ¼gðt Þ ¼ gðt þ Þ a 3 3ðt þ ð2=3ÞÞ t þ ð2=3Þ ¼P : ¼P 2 2=3 It is a rectangular pulse of unit amplitude centered at t ¼ ð2=3Þ with width (2/3). We can check the

Fig. 1.2.3 (a) Even function and (b) odd function

1.2.6 Products of Even and Odd Functions Let xe ðtÞ and ye ðtÞ be two even functions and x0 ðtÞ and y0 ðtÞ be two odd functions and arbitrary. Some general comments can be made about their products. xe ðtÞye ðtÞ ¼ xe ðtÞye ðtÞ; even function:

(1:2:9)

xe ðtÞy0 ðtÞ ¼ xe ðtÞy0 ðtÞ; odd function: (1:2:10) x0 ðtÞy0 ðtÞ ¼ ð1Þ2 x0 ðtÞy0 ðtÞ ¼ x0 ðtÞy0 ðtÞ; even function:

(1:2:11)

10

1 Basic Concepts in Signals

Fig. 1.2.4 (a) x1 ðtÞ, (b) xie (t) = x2 (t) even part of x1(t), and (c) x10 ðtÞ = x3 (t) odd part of x1 ðtÞ

x1(t)

x3(t)

x2(t)

(a)

(b)

Note that the functions P[t], L[t] and P[t]cosðo0 tÞ are even functions and P½t sinðo0 tÞ is an odd function. The even and odd parts of the exponential pulse x1 ðtÞ ¼ et uðtÞ are shown in Fig. 1.2.4 and are 1 x1e ðtÞ ¼ ðx1 ðtÞ þ x1 ðtÞÞ; 2 1 x10 ðtÞ ¼ ðx1 ðtÞ x1 ðtÞÞ: 2

sincðplÞ ¼

(c) sinðplÞ sin2 ðplÞ ; sinc2 ðplÞ ¼ pl ðplÞ2

(1:2:15)

Some authors use sincðlÞ for sincðplÞ in (1.2.15). Notation in (1.2.15) is common. Sinc ðplÞ is indeterminate at t ¼ 0: Using the L’Hospital’s rule,

(1:2:12) sinðplÞ ¼ lim l!0 l!0 pl lim

d sinðplÞ dl dðplÞ dl

p cosðplÞ ¼ 1: l!0 p

¼ lim

(1:2:16)

1.2.7 Signum (or sgn) Function

In addition, since sinðplÞ is equal to zero for l ¼ n; n an integer, it follows that

The signum (or sgn) function is an odd function shown in Fig. 1.2.5: sgn(t)

sincðplÞ ¼ 0;

l ¼ n;

n 6¼ 0 and an integer: (1:2:17a)

Interestingly, the function jsincðplÞj is bounded by jð1=plÞ j as jsinðplÞj 1. The side lobes of jsincðplÞj are larger than the side lobes of sinc2 ðplÞ, which follows from the fact that the square of a fraction is less than the fraction we started with. Both the sinc and the sinc2 functions are even. That is,

t

Fig. 1.2.5 Signum function sgnðtÞ

sincð plÞ ¼ sincðplÞ and sinc2 ðplÞ ¼ sinc2 ðplÞ:

sgnðtÞ ¼ uðtÞ uðtÞ ¼ 2uðtÞ 1:

(1:2:13)

(1:2:17b)

sgnðtÞ ¼ lim½eat uðtÞ eat uðtÞ; a > 0:

(1:2:14)

These functions can be evaluated easily by a calculator. For the sketch of a sinc function using MATLAB, see Fig. B.5.2 in Appendix B.

a!0

It is not defined at t ¼ 0 and is chosen as 0.

1.2.8 Sinc and Sinc2 Functions

1.2.9 Sine Integral Function

The sinc and sinc2 functions are defined in terms of an independent variable l by

The sine integral function is an odd function defined by (Spiegel, 1968)

1.3 Derivatives and Integrals of Functions

SiðyÞ ¼

Zy

sinðaÞ da: a

11

(1:2:18a)

0

The values of this function can be computed numerically using the series expression SiðyÞ ¼

y y3 y5 y7 þ þ ... ð1Þ1! ð3Þ3! ð5Þ5! ð7Þ7! (1:2:18b)

@xðt; aÞ xðt þ Dt; aÞ xðt; aÞ ¼ lim ; Dt!0 @t Dt @xðt; aÞ xðt; a þ DaÞ xðt; aÞ ¼ lim : Da!0 @a Da

(1:3:2)

Assuming the second (first) variable is not a function of the first (second) variable, the differential of xðt; aÞ is dx ¼

@x @x dt þ da: @t @a

Some of its important properties are The integral of a function over an interval is the area of the function over that interval.

SiðyÞ ¼ SiðyÞ; Sið0Þ ¼ 0; SiðpÞ ﬃ 2:0123; Sið1Þ ¼ ðp=2Þ :

(1:2:18c)

Si function converges fast and only a few terms in (1.2.18b) are needed for a good approximation.

Example 1.3.1 Compute the value of the integral of xðtÞ shown in Fig. 1.3.1.

1.3 Derivatives and Integrals of Functions It will be assumed that the reader is familiar with some of the basic properties associated with the derivative and integral operations. We should caution that derivatives of discontinuous functions do not exist in the conventional sense. To handle such cases, generalized functions are defined in the next section. The three well-known formulas to approximate a derivative of a function, referred to as forward difference, central difference, and backward difference, are dxðtÞ xðt þ hÞ xðhÞ : ; dt h xðt þ hÞ xðt hÞ ; x0 ðtÞ : 2h xðtÞ xðt hÞ : x0 ðtÞ : h

x0 ðtÞ ¼

(1:3:1)

MATLAB evaluations of the derivatives are given in Appendix B. If we have a function of two variables, then we have the possibility of taking the derivatives one or the other, leading to partial derivatives. Let xðt; aÞ be a function of two variables. The two partial derivatives of xðt; aÞ with respect to t, keeping a constant, and with respect to a, keeping t constant are, respectively, given by

Fig. 1.3.1 Computation of Integral of x(t) using areas

Solution: Divide the area into three parts as identified in the figure. The three parts are in the intervals ð0; aÞ, ða; bÞ, and ðb; cÞ, respectively. The areas of the two triangles are identified by A1 and A2 and the area of the rectangle by A3. They can be individually computed and then add the three areas to get the total area. That is, 1 1 A1 ¼ aB; A2 ¼ ðb aÞC; A3 ¼ ðc bÞB; 2 2 Zc xðtÞdt: A ¼ A1 þ A2 þ A3 ¼

&

0

If the function is arbitrary and cannot be divided into simple functions like in the above example, we can approximate the integral by dividing the area into small rectangular strips and compute the area by adding the areas in each strip.

12

1 Basic Concepts in Signals

If the time interval Dt ¼ ðb aÞ=N is sufficiently small, then the difference between the two formulas in (1.3.3a) and (1.3.3b) would be small and (1.3.3a) & is adequate.

x(t)

1.3.1 Integrals of Functions with Symmetries Fig. 1.3.2 xðtÞ and its approximation using its samples

Example 1.3.2 Consider the function xðtÞ shown in Fig. 1.3.2. Find the integral of this function using the above approximation for a5t5b. Assume the values of the function are known as xðaÞ; xða þ DtÞ; xða þ 2DtÞ; :::; and x((a þ N - 1)D t) Solution: Assuming Dt is small enough that we can approximate the area in terms of rectangular strips using the rectangular integration formula and " # Zb N1 X xðtÞdt xða þ nDtÞ Dt; Dt ¼ ðb aÞ=N: a

n¼0

(1:3:3a) Note that xðaÞDt gives the approximate area in the first strip. If the width of the strips gets smaller and the number of strips increases correspondingly, then the approximation gets better. In the limit, i.e., when Dt ! 0, approximation approaches the value of the integral. In computing the area of the kth rectangular strip, xða þ ðk 1ÞDtÞ was used to approximate the height of the pulse. Some other value of the function in the interval, such as the value of the function in the middle of the strip, could be used. MATLAB evaluation of integrals is discussed in Appendix B. In Chapter 8 appropriate values for Dt will be considered in terms of the frequency content in the signal. Instead of rectangular integration formula, there are other formulas that are useful. One could assume that each strip is a trapezoid and using the trapezoidal integration formula the integral is approximated by Zb

The integrals of functions with even and odd symmetries around a symmetric interval are Za Za Za xe ðtÞdt ¼ 2 xe ðtÞdt and x0 ðtÞdt ¼ 0; a

(1:3:4) where a is an arbitrary positive number. Example 1.3.3 Evaluate the integrals of the functions given below: x1 ðtÞ ¼ P½t=2a;

a

þ 2xða þ ðN 1ÞDtÞ þ xða þ NDtÞðDt=2Þ: (1:3:3b)

x2 ðtÞ ¼ tx1 ðtÞ:

(1:3:5)

Solution: x1 ðtÞ is a rectangular pulse with an even symmetry and x2 ðtÞ is an odd function with an odd symmetry. The integrals are Za Za x1 ðtÞdt ¼ 2 dt ¼ 2a; A1 ¼

A2 ¼

a Za

0

x2 ðtÞdt ¼ 0:

&

a

1.3.2 Useful Functions from Unit Step Function The ramp and the parabolic functions can be obtained by Zt xr ðtÞ ¼ uðtÞdt ¼ tuðtÞ and 0

xp ðtÞ ¼ xðtÞdt ½xðaÞ þ 2xða þ DtÞ þ 2xða þ 2DtÞ þ

a

0

Zt

xr ðtÞdt ¼ðt2 =2ÞuðtÞ:

(1:3:6)

0

Section 1.4 considers the derivatives of the unit step functions.

1.3 Derivatives and Integrals of Functions

13

Now consider Leibniz’s rule, interchange of derivative and integral, and interchange of integrals without proofs. For a summary, see Peebles (2001).

Z1 Z1

jxðt;aÞjdtda51;

1 1 Z1 Z1

½

1.3.3 Leibniz’s Rule

LetgðtÞ ¼

1

ZbðtÞ zðx; tÞdx:

(1:3:7)

aðtÞ

dgðtÞ daðtÞ dbðtÞ ¼ z½bðtÞ; tÞ z½aðtÞ; t dt dt dt bðtÞ Z @zðx; tÞ dx: þ @t

(1:3:8)

aðtÞ

1.3.4 Interchange of a Derivative and an Integral

1

jxðt;aÞdtjda51;

(1:3:10)

1

2 4

1

Z1

3 xðt; aÞdt5da ¼

1

Z1 1

2 4

Z1

3 xðt; aÞdt5da

1

Z1 Z1

¼

xðt; aÞdtda:

(1:3:11)

1 1

Signals generated in a lab are well behaved and they are valid. In Chapter 3 on Fourier series, integrating a product of a simple function, say hðtÞ, with its nth derivative goes to zero; such a polynomial, and the other one is a sinusoidal function gðtÞ, such as sinðo0 tÞ or cosðo0 tÞ or ejo0 t is applicable. The generalized integration by parts formula comes in handy. R

When the limits in (1.3.7) are constants, say aðtÞ ¼ a and bðtÞ ¼ b, then the derivatives of these limits will be zero and (1.3.7) and (1.3.8) reduce to gðtÞ ¼

1

is true, then Fubini’s theorem (see Korn and Korn (1961) for a proof) states that Z1

aðtÞ and bðtÞ are assumed to be real differentiable functions of a real parameter t, and zðx; tÞ and its derivative dzðx; tÞ=dt are both continuous functions of x and t. The derivative of the integral with respect to t is Leibniz’s rule Spiegel (1968) and is

Z1 Z1 ½ jxðt;aÞjdadt51;

hðnÞ ðtÞgðtÞdt ¼ hðn1Þ ðtÞgðtÞ hðn2Þ ðtÞg0 ðtÞ R þhðn3Þ ðtÞg00 ðtÞ ð1Þn hðtÞgðnÞ ðtÞdt; d k gðtÞ g ðtÞ ¼ ; dtk ðkÞ

Zb zðx; tÞdx;

and

: d k hðtÞ h ðtÞ ¼ dtk (1:3:12) ðkÞ

a

dgðtÞ d ¼ dt dt

Zb a

zðx; tÞdt ¼

Zb

@zðx; tÞ dx: @t

(1:3:9)

Using (1.3.12), the following equalities can be seen: Z

a

The derivative and the integral operations may be interchanged.

1.3.5 Interchange of Integrals

Z

Z Z

If any one of the following conditions,

t cosðtÞdt ¼ cosðtÞ þ t sinðtÞ; t sinðtÞdt ¼ sinðtÞ t cosðtÞ

(1:3:13a)

t2 cosðtÞdt ¼ 2t cosðtÞ þ ðt2 2Þ sinðtÞ; t2 sinðtÞdt ¼ 2t sinðtÞ ðt2 2Þ cosðtÞ: (1:3:13b)

14

1 Basic Concepts in Signals

1.4 Singularity Functions

Zt2

The impulse function, or the Dirac delta function, a singularity function, is defined by

t1

dðtÞ ¼

0;

t 6¼ 0

1;

t¼0

Z1 with

dðtÞdt ¼ 1:

(1:4:1)

1

dðtÞ takes the value of infinity at t ¼ 0 and is zero everywhere else. See Fig. 1.4.1b. Impulse function is a continuous function and the area under this function is equal to one. Note that a line has a zero area. Here, a generalized or a distribution function is defined that is nonzero only at one point and has a unit area. A delayed or an advanced impulse function can be defined by dðt t0 Þ, where t0 is assumed to be positive in the expressions. The ideal impulse function cannot be synthesized. It is useful in the limit. For example, 1 hti P ; e!0 e e

dðtÞ ¼ lim

h i 1 t L dðtÞ ¼ lim : e!0 e e (1:4:2)

Figure 1.4.1a illustrates the progression of rectangular pulses of unit area toward the delta function. As e is reduced, the height increases and, in the limit, the function approaches infinity at t ¼ 0 and the area of the rectangle is 1. There are other functions that approximate the impulse function in the limit. A nice definition is given in terms of an integral of a product of an impulse and a test function fðtÞ by Korn and Korn (1961):

fðtÞdðt t0 Þdt

8 0; t0 5t1 or t2 5t0 > > > < ð1=2Þ fðtþ Þ þ fðt Þ; 0 0 ¼ þ > ð1=2Þfðt Þ; t ¼ t1 0 > 0 > : ð1=2Þfðt0 Þ; t0 ¼ t2

t1 5t0 5t2

:

(1:4:3)

fðtÞ is a testing (or a test) function of t and is assumed to be continuous and bounded in the neighborhood of t ¼ t0 and is zero outside a finite interval. That is fð 1Þ ¼ 0. The integral in (1.4.3) is not an ordinary (Riemann) integral. In this sense, dðtÞ is a generalized function. A simpler form of (1.4.3) is adequate. dðtÞ has the property that Z1

fðtÞdðt t0 Þdt ¼ fðt0 Þ;

(1:4:4a)

1

where fðtÞ is a test function that is continuous at t ¼ t0 : As a special case, consider t0 ¼ 0 in (1.4.4a) and fðtÞ ¼ 1. Equation (1.4.4a) can be written as Z1

dðtÞdt ¼

1

Z0þ

dðtÞdt ¼ 1:

(1:4:4b)

0

That is, the area under the impulse function is 1. The integral in (1.4.4a) sifts the value of fðtÞ at t ¼ t0 and dðtÞ is called a sifting function. In summary, the impulse dðt t0 Þ has unit area (or weight) centered at the point t ¼ t0 and zero everywhere else. Since the dðt t0 Þ exists only at t ¼ t0 , and fðtÞ at t ¼ t0 is fðt0 Þ, we have an important result fðtÞdðt t0 Þ ¼ fðt0 Þdðt t0 Þ:

(1:4:5)

There are many limiting forms of impulse functions. Some of these are given below. Notes: In the limit, the following functions can be used to approximate dðtÞ:

(a)

(b)

Fig. 1.4.1 (a) Progression toward an impulse as e ! 0 and (b) symbol for dðtÞ

2 x1 ðtÞ ¼ ejt=tj ; t!0 t 1 1 x2 ðtÞ ¼ sincðt=tÞ; x3 ðtÞ ¼ sinc2 ðt=tÞ t t dðtÞ ¼ lim xi ðtÞ;

1.4 Singularity Functions 2 1 x4 ðtÞ ¼ epðt=tÞ ; t

15

x5 ðtÞ ¼

pðt2

t : þ t2 Þ

(1:4:6a)

x1 ðtÞ is a two-sided exponential function; x2 ðtÞ and x3 ðtÞ are sinc functions. Sinc function does not go to zero in the limit for all t. See the discussion by Papoulis (1962). x4 ðtÞ is a Gaussian function and x5 ðtÞ is a Lorentzian function. To prove these, approximate the impulse function in the limit, using (1.4.4a) and (1.4.4b). Example 1.4.1 Show that the Lorentzian function x5 ðtÞ approaches an impulse function as t ! 0. Show the result by using the equations: a. (1.4.1) and b. (1.4.4a and b). Solution: a. Clearly as t ! 0, x5 ðtÞ ! 0 for t not equal to zero. As t ! 0; x5 ðtÞ ! 1. From tables, 1 p

Z1 t2

t dt ¼ 1; lim x5 ðtÞ ¼ dðtÞ: t!0 þ t2

(1:4:6b)

1

differentiable everywhere any number of times, and, in addition, the function and its derivative decrease at least as rapidly as (1=tn ) as t ! 1 for all n. The derivative of a good function is another good function and the sums and the products of two good functions are good functions. A sequence of good functions fxn ðtÞg fx1 ðtÞ; x2 ðtÞ; . . . ; xn ðtÞg

(1:4:7)

is called regular, if for any good function fðtÞ, the following limit exists: Z1 fxn ðtÞgfðtÞdt: (1:4:8) Vx ðfÞ ¼ lim n!1

1

Example 1.4.2 Find the limit in (1.4.8) of the following sequence: n 2 2o (1:4:9) fxn ðtÞg ¼ eðt =n Þ : Solution: The limit in (1.4.8) is

b. First by (1.4.4a), with t0 ¼ 0 and fðtÞ ¼ 1 in the neighborhood of t ¼ 0 results in

Z1

Vx ðfÞ ¼

fðtÞdt:

(1:4:10) &

1

Z1 lim

t!0

fðtÞx5 ðt t0 Þdt ¼ lim

Z1

t!0

1

x5 ðtÞdt ¼ 1:

1

(1:4:6c)

Two regular sequences of good functions are considered equivalent if the limit in (1.4.8) is the same 4 4 for the two sequences. For example, eðt =n Þ and 2 2 eðt =n Þ are equivalent only in that sense. The function Vx ðfÞ defines a distribution xðtÞ and the limit of the sequence

Using (1.4.4b) and the integral tables, it follows that xðtÞ lim fxn ðtÞg: Z0þ 0

¼

n!1

0þ 1 t t 1 1 t dt ¼ tan p ðt2 þ t2 Þ p t t 0

1 hp p i ¼ 1: p 2 2

(1:4:11)

An impulse function can be defined in terms of a sequence of functions and write & dðtÞ lim fxn ðtÞg if lim n!1

Z1

n!1

fðtÞfxn ðtÞgdt ¼ fð0Þ

1

(1:4:12)

1.4.1 Unit Impulse as the Limit of a Sequence

¼)

Z1

fðtÞdðtÞdt ¼ fð0Þ:

(1:4:13)

1

Another approach to the above is through a sequence. See the work of Lighthill (1958). Also see Baher, (1990). A good function fðtÞ is

Many sequences can be used to approximate an impulse.

16

1 Basic Concepts in Signals

Notes: A constant can be interpreted as a generalized function defined by the regular sequence fxn ðtÞg so that, for any good function fðtÞ Z1 Z1 lim fðtÞdt: (1:4:14) fxn ðtÞgfðtÞdt ¼ K n!1

1

Solution: Using the integration by parts, we have the result below and (1.4.19) follows: Z1 Z1 1 0 x ðtÞfðtÞdt ¼ xðtÞfðtÞj1 xðtÞf0ðtÞdt: 1

1

(1:4:20) &

1

Using the function in (1.4.9), it follows that 2

2

2

fxn ðtÞg Ket =n ;K ¼ lim fKeðt

=n2 Þ

n!1

g:

(1:4:15)

Notes: Let g1 ðtÞ and g2 ðtÞ be two generalized functions and Z1 Z1 fðtÞg1 ðtÞdt ¼ fðtÞg2 ðtÞdt: (1:4:21) 1

1.4.2 Step Function and the Impulse Function Noting the area under the impulse function is one, it follows that Zt dðaÞda ¼ uðtÞ: (1:4:16) 1

Asymmetrical functions ðdþ ðtÞ and d ðtÞÞ and (uþ ðtÞ and u ðtÞÞ can be defined and are uþ ðtÞ ¼

u ðtÞ ¼

Z1 1 Z1 1

1; dþ ðtÞdt ¼ 0; d ðtÞdt ¼

1; 0;

8 t>0 > < 1; uðtÞ ¼ ð1=2Þ; t ¼ 0 > : 0; t 50

(1:4:17)

1

x ðtÞfðtÞdt ¼

xðtÞf0 ðtÞdt; f0 ðtÞ ¼

uðtÞfðtÞdt ¼

1

(1:4:19)

fðtÞdt:

(1:4:22)

0

1

Z1

f0 ðtÞdt ¼ ½fð1Þ fð0Þ ¼ fð0Þ; (1:4:23)

0

Z1

df dt

Z1

Noting that fð1Þ ¼ 0, it follows that Z1 Z1 0 u ðtÞfðtÞdt ¼ uðtÞf0 ðtÞdt

¼

t! 1

Z1

Solution: First Z1

(1:4:18)

Example 1.4.3 Using the property of the test function, lim fðtÞ ¼ 0, show that 0

Example 1.4.4 Show that the derivative of the unit step function is an impulse function using the equivalence property.

1

These step functions differ in how the value of the function at t ¼ 0 is assigned and are only of theoretical interest. Here the step function is assumed to be u ðtÞ and ignore the subscript: Derivative of the unit step function is an impulse function, which can be shown by using the generalized function concept.

Z1

The two generalized functions are equal, i.e., g1 ðtÞ ¼ g2 ðtÞ only in the sense of (1.4.21). It is called the equivalency property.

1

t>0 ; t 0 t0 t 50

1

1

Z1

0

u ðtÞfðtÞdt¼

dðtÞfðtÞdtand u0 ðtÞ

1

¼

duðtÞ ¼ dðtÞ: dt

(1:4:24)

The derivative of a parabolic function results in a ramp function, the derivative of a ramp function results in a unit step function, and the derivative of a unit step function results in an impulse function. All of these are true only in the generalized sense. What is the derivative of an impulse function? That is, d0 ðtÞ ¼

ddðtÞ : dt

(1:4:25)

1.4 Singularity Functions

17

It is defined by the relation Z1

d0 ðtÞfðtÞdt ¼

1

Z1

P½t=e ¼ uðt þ e=2Þ uðt e=2Þ

dP et e e ¼ dðt þ Þ dðt Þ; d0 ðtÞ 2 2 dt h e e i ¼ lim dðt þ Þ dðt Þ : e!0 2 2

dðtÞf0 ðtÞdt ¼ f0ð0Þ:

1

(1:4:26) Generalizing to higher-order derivatives of the impulse function results in Z1

dðnÞ ðtÞfðtÞdt ¼ ð1Þn fðnÞ ð0Þ:

(1:4:27)

1 hti 2 1 lim 2 P ¼ lim ¼ 1: e!0 e e!0 e e

Example 1.4.5 Evaluate the following integrals: a: A ¼

b: B ¼

(1:4:34)

d2 dðtÞ The impulse function is not square integrable. The square of a distribution is not defined Papoulis, dt2 (1962). Note that dðtÞ is an even function and (1:4:28) d0 ðtÞ is an odd function:

ðt2 þ 2t þ 1Þdð2Þ ðt 1Þdt; dð2Þ ðtÞ ¼

1

Z2

(1:4:33)

The derivative of the impulse function results in two impulses illustrated in Fig. 1.4.2. It is an odd function called a doublet. The square of an impulse function is not defined as

1

Z1

(1:4:32)

½ðt 1Þ2 dðt 1Þ þ 5dðt þ 1Þ þ 6tdðtÞdt:

∞

:5

(1:4:29)

t

0

Solution: These follow Z1 ½ða þ 1Þ2 þ 2ða þ 1Þ þ 1dð2Þ ðaÞda a: A ¼

Fig. 1.4.2 Symbolic representation of d0 ðtÞ, a doublet

–∞

1

¼ ð1Þ2

b: B ¼

Z2

d2 ½ða þ 1Þ2 þ 2ða þ 1Þ þ 1ja¼0 ¼ 4 da2

1.4.3 Functions of Generalized Functions Using the equivalence property of the generalized function, the following is true:

½ðt 1Þ2 dðt 1Þ þ 5dðt þ 1Þ

:5

(1:4:30)

gðtÞ ¼ dðat bÞ ¼

When an impulse is outside the integration limits, then the integral is 0. In addition, it is assumed that the limits do not fall at the exact location of the & impulses.

This can be seen from

þ6tdðtÞdt ¼ 0:

Z1 1

1 fðtÞdðat bÞdt¼ j aj

dðtÞ can be approximated using various functions in the limit. In Fig. 1.4.1 it is expressed in terms of a rectangular pulse function. That is, 1 hti dðtÞ ¼ lim x1 ðtÞ; x1 ðtÞ ¼ P : e!0 e e

(1:4:31)

The generalized derivative of a pulse function is as follows:

dðt b=aÞ ; a 6¼ 0: j aj

Z1

(1:4:35)

fðy=aÞdðy bÞdy

1

1 ¼ fðb=aÞ: j aj

(1:4:36)

From the equivalence property, the equality in (1.4.35) now follows. From (1.4.35), it follows that dðoÞ ¼

1 dðfÞ or dðfÞ ¼ 2pdðoÞ: 2p

(1:4:37)

18

1 Basic Concepts in Signals Table 1.4.1 Properties of the impulse function Z1 dðt t0 ÞfðtÞdt ¼ fðt0 Þ: 1

Z1

dðtÞfðt t0 Þdt ¼ fðt0 Þ:

1

Z1

fðtÞdðtÞdt ¼ fð0Þ:

1

Z1 1

Z1

ddðtÞ dfðtÞ : fðtÞdt ¼ dt dt t¼0

dðnÞ ðtÞfðtÞdt ¼ ð1Þn fðnÞ ð0Þ:

1

xðtÞdðtÞ ¼ xð0ÞdðtÞ: Z1

dðtÞdðt0 tÞdt ¼ dðt0 Þ:

1

dðat bÞ ¼

1 b dðt Þ ; a 6¼ 0: a jaj

dðjoÞ ¼ dðoÞ: dðtÞ ¼ dðtÞ:

1.4.4 Functions of Impulse Functions In (1.4.35), dðxðtÞÞ is considered with xðtÞ being a linear function of time. Now consider other cases, where xðtÞ is assumed to have simple zeros, i.e., no multiple zeros. Example 1.4.6 Evaluate the following integral using (1.4.35) and the following cases for the limits. a. x ¼ 0; y ¼ 4 and b. x ¼ 10; y ¼ 5.

A¼

Zy

ðt 1Þðt þ 5Þdð2t þ 5Þdt:

x

Solution: First, changing the variables, a ¼ 2t þ 5, i.e., t ¼ 12ða 5Þ, dt ¼ 12 da, results in a: t ¼ 0¼)a ¼ 5; t ¼ 4¼)a ¼ 13;

Z13 1 7 1 5 1 ¼)A ¼ a a þ dðaÞ da ¼ 0 2 2 2 2 2 5

ða ¼ 0 is outside the range 55a513Þ: b. Similarly, changing the variables, a ¼ 2t þ 5, i.e., t ¼ 12ða 5Þ, dt ¼ 12 da, results in t ¼ 5¼)a ¼ 5; t ¼ 0¼)a ¼ 5; Z5 1 7 1 5 1 ¼)A ¼ a a þ dðaÞ da 2 2 2 2 2 5 1 7 1 5 1 35 a aþ ¼ ja¼0 ¼ : 2 2 2 2 2 8

&

Example 1.4.7 Using the equivalence property of the impulse functions, show that

1.5 Signal Classification Based on Integrals

dðt2 a2 Þ ¼

19

1 ðdðt þ aÞ þ dðt aÞÞ; a 6¼ 0: j2aj (1:4:38)

Solution: Since t2 a2 ¼ ðt aÞðt þ aÞ ¼ 0 ! t ¼ a 6¼ 0 at t ¼ a, it follows that Z1

Example 1.4.8 Give the expression for dðsinðtÞÞ: Solution: Since sinðtÞjt¼np ¼ 0; d sinðtÞ=dt ¼ cosðtÞ; andj cosðnpÞj ¼ 1, it follows that 1 X dðsinðtÞÞ ¼ dðt npÞ: (1:4:41) & n¼1

Z0 dðt a ÞfðtÞdt ¼ dðt2 a2 ÞfðtÞdt 2

2

1

1.4.5 Functions of Step Functions

1

þ

Z1

dðt2 a2 ÞfðtÞdt: (1:4:39)

0

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ With t ¼ ðy þ a2 Þ, Z0

Example 1.4.9 Given xðtÞ ¼ t2 1, sketch the function yðtÞ ¼ uðxðtÞÞ ¼ uðt2 1Þ, where uðtÞ is the unit step function. Solution: Since xðtÞ 0 for 1 t 1, it follows that

dðt2 a2 ÞfðtÞdt

xðtÞ 0 for 1 t 1 ! yðtÞ ¼ 0; 15t51;

1

Za2

¼

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ dðyÞf y þ a2

1

1 : ¼ fðaÞ j2aj

Fig. 1.4.3 shows the sketches for xðtÞ and yðtÞ.

Note that when t ¼ 0; y ¼ a2 and when t ¼ 1; y ¼ 1. In a similar manner, we can evaluate the second integral in (1.4.39). Combining them, it follows that Z1 1

xðtÞ > 0 for 15t5 1 and 5 1 t51 ! yðtÞ ¼ 1; t5 1; t > 1:

! 1 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ dy 2 y þ a2

1.5 Signal Classification Based on Integrals The area of a signal xðtÞ is

1 fðaÞ fðaÞ ½dðt aÞ þ dðt þ aÞfðtÞdt ¼ þ : j2aj j2aj j2aj

Z1

Area½xðtÞ ¼

xðtÞdt:

(1:5:1)

1

This can be generalized. If xðtÞ has simple roots at t ¼ tn , then d½xðtÞ ¼

&

X tn

If a signal xðtÞ is said to be absolutely integrable, then

1 dxðtÞ dðt tn Þ; x0 ðtn Þ ¼ jt¼tn dt jx0 ðtn Þj

Area½jxðtÞj ¼

(1:4:40) &

Z1

jxðtÞjdt51:

(1:5:2)

1

x(t) y(t)

t

Fig. 1.4.3 (a) xðtÞ ¼ t2 1 and (b) yðtÞ ¼ uðt2 1Þ

(a)

t

(b)

20

1 Basic Concepts in Signals

If a signal is square integrable, i.e., a finite energy signal satisfies Z1 2 (1:5:3) Area½jxðtÞj ¼ jxðtÞj2 dt51: Consider a resistor of value R O. Ohm’s law states that the voltage, vðtÞ; across this resistor is equal to R times the current iðtÞ passing through the resistor and vðtÞ ¼ RiðtÞ. The instantaneous power delivered to the resistor is pR ðtÞ ¼ i2 ðtÞR. The total energy delivered to the resistor is ER ¼

Z1

pR ðtÞdt ¼ R

1

1 i ðtÞdt ¼ R 2

1

Z1

v2 ðtÞdt:

1

If we normalize the resistor value to 1O, that is, R ¼ 1, then we can consider the function, xðtÞ as a generic (i.e., either voltage or current) function. The energy in xðtÞ is defined by Ex ¼

Z1

2

jxðtÞj dt:

(1:5:4)

1

The normalized average power of a signal xðtÞ is defined by 1 Px ¼ lim T!1 T

Aeat dt ¼

ZT=2

jxðtÞj2 dt:

(1:5:5)

The signal xðtÞ is an energy signal if 05Ex 51, that is, Ex is finite and Px ¼ 0. The function xðtÞ is a power signal if 05Px 51, i.e., Px is finite and therefore Ex is infinite. If a signal does not satisfy one of these conditions, then it is neither an energy signal nor a power signal. The power and energy signals are mutually exclusive. Example 1.5.1 Show that the function given below is an energy signal. xðtÞ ¼ Aeat uðtÞ; a > 0:

(1:5:6)

Find its area, the absolute area, and its energy assuming A 6¼ 0 is finite. Solution: The area, the absolute area, and the energy are, respectively, given by

Z1

Ex ¼

jxðtÞj2 dt ¼ jAj2

1

Z1 h

2

jeat j uðtÞdt ¼ jAj2

1

¼

i

A 2a

2

Z1

e2at dt

0

! jAj2 : 2a

e2at t¼1 t¼0 ¼

From this it follows that Ex is finite. Clearly, Px ¼ 0 & implying it is an energy signal. Example 1.5.2 Using a ! 0 in the above example show that xðtÞ is a power signal. xðtÞ ¼ AuðtÞ; A 6¼ 0 and finite:

(1:5:7)

Solution: The energy contained in a step function Ex is infinite, whereas 2 3 ZT=2 6 1 7 Px ¼ lim 4 jxðtÞj2 dt5 T!1 T T=2

2

3 ZT=2 6 1 7 ¼ lim 4 jAj2 dt5 T!1 T 0

T=2

A at 1 A e j0 ¼ ; a a

0

Area½jxðtÞj ¼ jAj=a;

1

Z1

Area½xðtÞ ¼

Z1

¼ lim jAj2 T!1

T=2 T

¼

jAj2 2

is finite. It follows that xðtÞ in (1.5.7) is a power signal. In determining whether a signal is a power or an energy signal, we can check either its energy or power. If Ex is finite, then Px ¼ 0. If Px is finite, then Ex is infinite. We do not have to check both of them. If Ex is infinite, then we need to check the average power before making the decision on whether the signal is a power signal or neither a & power nor an energy signal. Example 1.5.3 Show that xðtÞ defined below is neither an energy nor a power signal. xðtÞ ¼ tuðtÞ

(1:5:8)

Solution: The energy and the average power in this signal are, respectively, given by

1.5 Signal Classification Based on Integrals

Z1

1 Ex ¼ t dt ¼ 1; Px ¼ lim T!1 T 0 3 1 T ¼ 1: ¼ lim T1 !1 3 T 2

21

ZT=2

Z1 2

t dt

xðt aÞdt ¼

1

0

Example 1.5.4 Show that xðtÞ ¼ A; a finite constant is a power signal. Solution: The average power contained in xðtÞ is given by 2 3 ZT=2 61 7 Px ¼ lim 4 A2 dt5 ¼ A2 ¼)E ¼ 1: T!1 T T=2

It is a power signal.

1

&

Z1 1

1 xðatÞdt ¼ j aj

Notes: The signals dðtÞ and d ðtÞ are neither nor power signals since the squares of these functions are not defined. There are two classes of power signals. These are periodic and random signals. Random signals require some knowledge of probability & theory, which is beyond the scope here.

1.5.1 Effects of Operations on Signals Example 1.5.5 In signal analysis, scaling and shifting of a function are quite common. a. Show that the functions xðtÞ and xðt aÞ have the same areas and energies. b. Show AreaðxðatÞÞ ¼ ð1=jajÞAreaðxðtÞÞ; a 6¼ 0: jxðt aÞj dt ¼

1

Z1

jxðbÞj2 db:

(1:5:9a) (1:5:9b)

1

Solution: a. Using a change of variable in the integral b ¼ t a results in

T1

1 lim T1 !1 T1

Z2 T

21

N X 1 1 yT ðtÞdt¼ lim N!1 2 N þ 1 T k¼N

ð2Z kþ1ÞT2 ð2 k1ÞT2

Z1 xðbÞdb: 1

Equation (1.5.9b) can be shown in both cases and is & left as an exercise.

1.5.2 Periodic Functions A function xðtÞ is periodic or T-periodic if there is a number T for all time such that xðt þ TÞ ¼ xðtÞ:

0

2

xðbÞdb:

b. Similarly, for at ¼ b results in

Since both the energy and the average power go to & infinity, the result follows.

Z1

Z1

(1:5:10)

It is common to use the actual period, such as T as a subscript on x and write xT ðtÞ ¼ xT ðt þ TÞ:

(1:5:11)

The smallest positive number T that satisfies (1.5.11) is called the fundamental period and it defines the duration of one complete cycle. The reciprocal of the fundamental period is the fundamental frequency. That is, f0 ¼ 1=T Hz and the period T ¼ 1=f0 s:

(1:5:12)

Clearly, if (1.5.11) is satisfied, then for all integers of n, xT ðt þ nTÞ ¼ xT ðtÞ. If there is no T that satisfies (1.5.10), then xðtÞ is called an aperiodic or a nonperiodic signal. Note that the integral over any one period of a periodic function is the same. Earlier we were interested in finding the average power and the average energy in a signal. With periodic signals, we can make some simplifications of the integrals. Consider the normalized integral of a periodic signal with period T1 .

T

2N þ 1 1 yT ðtÞdt¼ lim N!1 2 N þ 1 T

Z2 T2

T

1 yT ðtÞdt ¼ T

Z2 T2

yT ðtÞdt: (1:5:13)

22

1 Basic Concepts in Signals

Therefore, any one period can be used and written in symbolic short hand notation by T1 =2 Z

1 lim T1 !1 T1

yT ðtÞdt ¼

1 T

Z

yT ðtÞdt:

(1:5:14)

T

T1 =2

Fig 1.5.1 Half-rectified sine wave

The terms that are of interest in dealing with periodic functions are the duty cycle of an on–off signal (i.e., the ratio of on-time to the period), average value of the signal xave , the average signal power Px , and the root mean square (rms) value xrms . It is also called the effective value of the periodic function yT ðtÞ. These values are defined by xave

1 ¼ T

Z T

Px ¼

1 T

Z

1 Px ¼ T

x2T ðtÞdt

T

Z

2

¼

A T

jxT j2 dt; xrms ¼

Z

cos2 ðo0 t þ y0 Þdt

T

1 þ cosð2ðo0 t þ y0 ÞÞ A2 : (1:5:17) dt ¼ 2 2

T

The rms value is pﬃﬃﬃ xrms ¼ ðjAj= 2Þ:

xT ðtÞdt; Z

A2 ¼ T

pﬃﬃﬃﬃﬃﬃ Px :

(1:5:15)

T

Since the average power in a periodic signal is finite, the energy is infinite, it follows that all periodic signals are power signals. In Chapter 3 periodic functions will be discussed in detail, where the average value of a periodic function can never exceed the rms value will be shown.

It is best to express sinusoidal functions in terms of Hertz to compute the period. The period is the inverse of the fundamental frequency. Interestingly, if the frequency f0 ¼ o0 =2p is 1 MHz, then the signal completes 1 million cycles every second. & Example 1.5.7 Consider the half-wave sinusoidal periodic function

sinðtÞ; 0 t5p ; x2p ðtÞ ¼ x2p ðt þ 2pÞ 0; p t 5 2p

x2p ðtÞ ¼ Example 1.5.6 Consider the function xT ðtÞ with A; o0 ; and y0 being real constants given below. a. Show that it is a periodic function with period T ¼ ð2p=o0 Þ ¼ ð1=f0 Þ. xT ðtÞ ¼ A cosðo0 t þ y0 Þ:

(1:5:16)

b. Find xave ; Px ; and xrms for the above periodic signal. Solution: a. Using tables it can be seen that xT ðtÞ is a periodic function, i.e., xT ðt þ TÞ ¼ A cosðo0 ðt þ ð2p=o0 ÞÞ þ y0 Þ ¼ A cosðo0 t þ y0 Þ cosð2pÞ A sinðo0 t þ y0 Þ sinð2pÞ ¼ A cosðo0 t þ y0 Þ:

(1:5:19) shown in Fig. 1.5.1. Find its duty cycle, average, average signal power, and its rms value. Show that average value of the function is less than its rms value. Solution: Clearly the duty cycle is (1/2), as the signal is on for half the time. The average, the power, and the corresponding root mean square values of x2p ðtÞ are Zp Zp 1 1 1 x2p ðtÞdt ¼ sinðtÞdt ¼ (1:5:20a) xave ¼ 2p 2p p 0

1 Px ¼ 2p

Zp 0

b. The average value of a sine or a cosine function is zero as their positive areas cancel out with their negative areas. The average power is independent of y0 and

(1:5:18)

¼

1 4p

Zp 0

0

1 sin ðtÞdt ¼ 2p 2

Zp

1 ð1 sinð2tÞÞdt 2

0

ð1 sinð2tÞÞdt ¼

1 4

(1:5:20b)

1.5 Signal Classification Based on Integrals

23

Fig. 1.5.2 xT1 ðtÞ ¼ sinðð2p=4ÞtÞ; xT2 ðtÞ ¼ sinðð2p=6ÞtÞ

xrms ¼

pﬃﬃﬃﬃﬃﬃ Px ¼ 1=2:

(1:5:20c)

Note that the average value is less than the rms & value, as ð1=pÞ ﬃ 0:31835ð1=2Þ ¼ 0:5: Notes: The average value of a periodic function does not exceed the rms value, i.e.,xave xrms . The rms value of the sinusoidal voltage supplied to the outlet of a US home is 120 V with a frequency of 60 Hz. The pﬃﬃﬃ maximum value of the voltage at the & outlet is 2ð120Þ ¼ 169:71 V 170 V.

1.5.3 Sum of Two Periodic Functions

least common multiple of 4 and 6 is 12 and, therefore, xT ðtÞ=xT1 ðtÞ þ xT2 ðtÞ is periodic with period T ¼ 12 s. This can be seen from the fact that in 12 s, xT1 ðtÞ will have three full cycles, xT2 ðtÞ will have two full cycles, and xT ðtÞ will have one full cycle. Figure 1.5.2 gives sketches of xT1 ðtÞ and xT2 ðtÞ. If each of the signals is shifted by different amounts, say if yT ðtÞ ¼ A1 sinð2pð1=4Þtþy1 ÞþA2 sinð2pð1=6Þtþy2 Þ, then yT ðtÞ is still periodic with period T=12 s for any set of constants A1 ; A2 and angles y1 and y2 . This can be generalized and state that for any constants X½0; h½k, and y½k, the function xT ðtÞ ¼ X½0 þ

1 X

h½k cosðko0 t þ y½kÞ

(1:5:22)

k¼1

If xT1 ðtÞ and xT2 ðtÞ are two periodic functions with periods T1 and T2 , respectively, then xðtÞ ¼ xT1 ðtÞ þ xT2 ðtÞ is periodic with period T if T ¼ nT1 ¼ mT2 or ½T1 =T2 ¼ ½m=n:

(1:5:21)

m and n are some integers and ðT1 =T2 Þ is a rational number. The period of xðtÞ is equal to the least common multiple (LCM) of T1 and T2 . The LCM of two integers, m and n, is the smallest integer divisible by both m and n. If T1 =T2 is an irrational number, it cannot be written in terms of a ratio of two integers and xðtÞ is not periodic. Example 1.5.8 Let xT ðtÞ ¼ a1 cosðo0 tÞ and yT ðtÞ ¼ b1 sinðo0 tÞ, with T ¼ ð2p=o0 Þ. Show that zT ðt þ TÞ ¼ zT ðtÞ ¼ xT ðtÞ þ yT ðtÞ. Solution: Since yT ðt þ TÞ ¼ yT ðtÞ and xT ðt þ TÞ ¼ & xT ðtÞ implies zT ðtÞ is periodic. Example 1.5.9 Let xT1 ðtÞ ¼ A1 sinð2pð1=4ÞtÞ and xT2 ðtÞ ¼ A2 sinð2pð1=6ÞtÞ. Show that the period of xT ðtÞ ¼ xT1 þ xT2 is 12. Sketch the function xT ðtÞ. Solution: The period of xT1 ðtÞ is T1 =4 s and the period of xT2 ðtÞ is T2 ¼ 6 s. The ratio, ðT1 =T2 Þ ¼ ð4=6Þ is a rational number, and the

is periodic with period T ¼ o0 =2p. Noting that (o0 T ¼ 2p), we have cosðko0 ðt þ TÞ þ y½kÞ ¼ cosðko0 t þ y½kÞ cosðko0 TÞ sinðko0 t þ y½kÞ sinðko0 TÞ ¼ cosðko0 t þ y½kÞ: The term for k ¼ 1 is called the fundamental and the kth term is called the kth harmonic. The dc term is X½0. The above can be generalized and state that the following function is periodic with period T ¼ 2p=o0 for any constants X½0; A½k, and B½k: xT ðtÞ ¼ X½0 þ

1 X

A½k cosðko0 tÞ

k¼1

þ

1 X

B½k sinðko0 tÞ:

(1:5:23) &

k¼1

Example 1.5.10 Let xT1 ðtÞ ¼ cosð4tÞ, xT2 ðtÞ ¼ cosð2ptÞ, and xðtÞ ¼ xT1 ðtÞ þ xT2 ðtÞ. Show that xðtÞ is not a periodic function. Solution: The period of xT 1 ðtÞ ¼ cosð2pð2=pÞtÞ is T1 ¼ ðp=2Þ and the period of xT 2 ðtÞ is T2 ¼ 1: The ratio ðT1 =T2 Þ ¼ðp=2Þ is an irrational number and & xðtÞ is not periodic.

24

1 Basic Concepts in Signals

For functions such as the one given in the above example, there is no repetition. These types of combination of periodic functions are referred as quasi periodic or almost periodic. For a study of almost periodic functions, see Chuanyi, (2003). Example 1.5.11 Compute the average power in xðtÞ given below for the two cases: xðtÞ ¼ C1 cosðo1 t þ y1 Þ þ C2 cosðo2 t þ y2 Þ a. o1 ¼ no0 6¼ o2 ¼ ko0 , where n and k are integers b. o1 ¼ o2 ¼ o0 . Assume C1 ; C2 ; y1 ; and y2 are arbitrary constants. Solution: Without loosing any generality, assume T ¼ ð2p=o0 Þ. 2p 2p xðt þ Þ ¼ C1 cosðno0 ðt þ Þ þ y1 Þ o0 o0 2p þ C2 cosðko0 ðt þ Þ þ y2 Þ o0 ¼ C1 cosðno0 t þ y1 Þ þ C2 cosðko0 t þ y2 Þ: This indicates that xðtÞ ¼ xT ðtÞ is periodic for both cases with period T ¼ ðo0 =2pÞ. a: P ¼

1 T0

Z

jxT ðtÞj2 dt ¼

1 T0

T0

Z

½C1 cosðo1 t þ y1 Þ

T0 2

þC2 cosðo2 t þ y2 Þ dt Z 1 ½C21 þ C22 þ C21 cosð2ðo1 t þ y1 ÞÞ ¼ 2T0 T0

þC22

cosð2ðo2 t þ y2 Þdt Z 1 ½cosððo1 þ o2 Þt þ ðy1 þ y2 ÞÞ þ C1 C2 T0

xT ðtÞ ¼ C1 cosðo0 t þ y1 Þ þ C2 cosðo0 t þ y1 ðp=2ÞÞ ¼ C1 cosðo0 t þ y1 Þ C2 sinðo0 t þ y1 Þ:

&

Notes: Consider xðtÞ ¼ C cosðo0 t þ yÞ ¼ C cosðyÞ cosðo0 tÞ C sinðyÞ sinðo0 tÞ ¼ a cosðo0 tÞ þ b sinðo0 tÞ; pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ a ¼ C cosðyÞ; b ¼ C sinðyÞ; C ¼ a2 þ b2 ; y ¼ tan1 ðb=aÞ:

(1:5:26)

One should be careful in computing y as tan1 ðb=aÞ 6¼ tan1 ðb= aÞ; tan1 ðb= aÞ 6¼ tan1 ðb=aÞ:

&

Exponentially varying sinusoids: An example of such a function is xðtÞ ¼ Aeat cosðo0 t þ yÞ:

(1:5:27)

It becomes unbounded for at50. Our interest is for only positive t and such functions are referred as causal signals. If xðtÞ is defined for all t, then its causal part is yðtÞ ¼ xðtÞuðtÞ. In the case of xðtÞ in (1.5.27), yðtÞ ¼ xðtÞuðtÞ ¼ Aeat cosðo0 t þ yÞuðtÞ:

(1:5:28)

If a > 0 ða50Þ, xðtÞ in (1.5.28) is an exponentially decaying (increasing) sinusoidal function. These functions can be sketched using the envelopes Aeat and Aeat as constraints and the function cosðo0 t þ yÞ between the envelopes. Notes: Even for temporal signals, the analysis and design of noncausal systems is important. For example, the analysis of prerecorded data is applicable. &

T0

1 þ cosððo2 o1 Þt þ ðy2 y1 ÞÞdt ¼ ½C21 þ C22 2 (1:5:24) 1.6 Complex Numbers, Periodic, and

Symmetric Periodic Functions b. In the case of o2 ¼ o1 , the average power is 1 P ¼ ½C21 þ C22 þ C1 C2 cosðy2 y1 Þ: 2

(1:5:25)

In the above equation C1 C2 cosðy2 y1 Þ is equal to zero only when ðy2 y1 Þ ¼ p=2 and

A complex number ci ¼ ai þ jbi , where ai ¼ Reðci Þ is the real part and bi ¼ Imðci Þ is the imaginary part. Similarly if xðtÞ is a complex function, we can write it as xðtÞ ¼ ReðxðtÞÞ þ jImðxðtÞÞ:

(1:6:1)

1.6 Complex Numbers, Periodic, and Symmetric Periodic Functions

1.6.1 Complex Numbers A complex number can be written in terms of its real and imaginary parts or in terms of its magnitude and phase. Consider the complex number ci ¼ ai þ jbi ¼ ri ðcos yi þ j sin yi Þ ¼ ri ejyi ; ri qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ a2i þ b2i ; yi ¼ arctanðci Þ ¼ tan1 ðbi =ai Þ: (1:6:2) The representation ci ¼ ri ejyi is referred to as the polar form of the complex number, where ri and yi are, respectively, called the amplitude (or the modulus) and the phase angle associated with the complex number. One needs to be careful in using the formula yi ¼ arctanðbi =ai Þ, especially when the real part of ai is negative. Example 1.6.1 Sketch the following complex numbers as vectors on a complex plane. c1 ¼ 1 þ j1; c2 ¼ 1 þ j1; c3 ¼ 1 j1; c4 ¼ 1 j1:

25

Solution: The complex numbers c1 ; c2 ; c3 ; and c4 can be represented on the complex plane as vectors, where the length of the vector is equal to ri , and are illustrated in Fig. 1.6.1. They are located in the first, second, third, and fourth quadrants, respectively. Also, pﬃﬃﬃ p jci j ¼ ri ¼ 1= 2; i ¼ 1; 2; 3; 4; y1 ¼ ¼ 45 ; 4 3p 5p 7p ¼ 135 ; y3 ¼ ¼ 225 ; y4 ¼ ¼ 315 y2 ¼ 4 4 4 Noting that if we use the arctangent function, i.e., yi ¼ arctanðbi =ai Þ, we have y1 ¼ y3 and y2 ¼ y4 . Note the ambiguity in taking the ratio of positive (negative)/negative (positive) quantities. This ambiguity can be reconciled, for example, by noting that when the vector lies in the second quadrant, the angle must satisfy 90 y 180 . The correct angle is 180 y2 ¼ 180 arctanðb2 =a2 Þ ¼ 180 45 ¼ 135 :

&

A general method for obtaining the phase angle of a complex number c ¼ a þ jb is

(a)

(b)

(c)

(d)

Fig. 1.6.1 (a) c1 ¼ 1 þ j1; ðbÞ c2 ¼ 1 þ j1; ðcÞ c3 ¼ 1 j1; ðdÞ c4 ¼ 1 j1

26

1 Basic Concepts in Signals

y¼

if a 0 if a50

arctanðb=aÞ 180 þ arctanðb=aÞ

(1:6:3)

When a50, select the appropriate one so that jyj 1800 or p radians. In terms of power series, in radians (1 rad ¼ p=180 ), 3

5

7

arctanðyÞ ¼ y y 3 þ y5 y7 þ ; jyj51 þ; y 1 : p 1 1 1 arctanðyÞ ¼ 2 y þ 3y2 5y5 þ ; ; y 1 (1:6:4) Convergence of the series is fast for most values of y and very few terms are needed to compute the arctangent function. The worst case is when y ¼ 1. The conjugate, polar representation, sums, differences, multiplications, and divisions of complex numbers are given below, where we assume ci ¼ ai þ jbi . c 1 ¼ ða1 jb1 Þ; ci ¼ ri ejyi ; c1 c2 ¼ ða1 þ a2 Þ jðb1 þ b2 Þ

(1:6:5a)

c1 c2 ¼ ða1 þ jb1 Þða2 þ jb2 Þ ¼ ða1 a2 b1 b2 Þ þ jða1 b2 þ b1 a2 Þ ¼ r1 r2 ejðy1 þy2 Þ

(1:6:5b)

c1 a1 þ jb1 a1 þ jb1 a2 jb2 ¼ ¼ c2 a2 þ jb2 a2 þ jb2 a2 jb2 ða1 a2 þ b1 b2 Þ ða1 b2 b1 a2 Þ ¼ j ; c2 6¼ 0 2 2 a 2 þ b2 a22 þ b22 ðc1 =c2 Þ ¼ ðr1 =r2 Þejðy1 y2 Þ :

r ¼ absðcÞ ;y ¼ angleðcÞ ¼ atan2ðimagðcÞ; realðcÞÞ: (1:6:8) Notes: MATLAB atanðxÞ function computes the arctangent or inverse tangent of x. The function returns an angle in radians between ðp=2Þ and ðp=2Þ. MATLAB atan(x, y) computes the arctangent of ðy=xÞ. It returns an angle in radians between p and p and the signs on both x and y plays a role. & See Appendix B for MATLAB. In the last section we considered a sum of two sinusoids. Euler’s formula can be used to express sinusoids in terms of complex exponentials and vice versa. These are cosðyÞ ¼ ½ejy þ ejy =2;

ejy ¼ cosðyÞ þ j sinðyÞ; ejy ¼ cosðyÞ j sinðyÞ: (1:6:9b) Notes: MATLAB atanðxÞ function computes the arctangent or inverse tangent of x. The function returns an angle in radians between ðp=2Þ and ðp=2Þ. MATLAB atan(x, y) computes the arctangent of ðy=xÞ. It returns an angle in radians between p and p and the signs on both x and y plays a role. See Appendix B for MATLAB. & Example 1.6.2 Express the following functions in terms of single sinusoids. See (1.5.26).

(1:6:5c)

c1 ða21 b21 Þ 2a1 b1 ¼ 2 ; þj 2

2 c1 ða1 þ b1 Þ a1 þ b21 jc j jr j 1 ¼ 1 ¼ 1; y ¼ 2y1 ¼ 2 arctanðb1 =a1 Þ: (1:6:6) c j r 1 j

¼)

1

Consider the polar form of the complex number c and its natural log:

sinðyÞ ¼ ½ejy ejy =2j: (1:6:9a)

a: xðtÞ ¼ cosðo0 tÞ

pﬃﬃﬃ 2 sinðo0 tÞ; b: xðtÞ

¼ cosðo0 tÞ þ sinðo0 tÞ:

(1:6:10)

pﬃﬃﬃ Solution: a: In this case; we have a ¼ 1; b ¼ 2. From (1.6.10), we have qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃ 2ﬃ pﬃﬃﬃ pﬃﬃﬃ C ¼ 1 þ ð 2Þ ¼ 3; y ¼ tan1 ð 2=1Þ ﬃ 54:73 ; pﬃﬃﬃ 3 cosðo0 t þ 54:73 Þ pﬃﬃﬃ ¼ 3 cosð54:73 Þ cosðo0 tÞ pﬃﬃﬃ 3 sinð54:73 Þ sinðo0 tÞ:

xðtÞ ¼ c ¼ rejy ;

lnðcÞ ¼ lnðrejy Þ ¼ lnðrÞ þ jy r ¼ jcj; y ¼ ImðlnðcÞÞ

(1:6:7)

MATLAB can operate in terms of real or complex numbers. MATLAB commands for computing the amplitude and phase are

b. In this case a ¼ 1; b ¼ 1. From (1.6.10), it follows that

1.6 Complex Numbers, Periodic, and Symmetric Periodic Functions

pﬃﬃﬃ 2; y ¼ tan1 ½1= 1 ¼ 135 ! xðtÞ pﬃﬃﬃ ¼ 2 cosðo0 t 135 Þ:

C¼

Case a: a=1; b=1; [theta, C] = cart2pol(a, b); theta_deg = (180/pi)*theta; C, theta_deg! C ¼ 1:7321; y ¼ 54:73560 Case b: a=1; b=1; [theta, C] = cart2pol(a,-b); theta_deg = (180/pi)*theta; C, theta_deg!C ¼ 1:4142; y ¼ 135 .

27

Solution: a. ða1 jb1 oÞ ða2 jb2 oÞ XðjoÞ ¼ ða2 þ jb2 oÞ ða2 þ jb2 Þ ða1 a2 b1 b2 o2 Þ ða1 b2 þ a2 b1 Þo j : ¼ 2 2 2 ða2 þ b2 o Þ ða22 þ b22 o2 Þ b. In this part jXðjoÞj ¼

sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ a21 þ ðb1 oÞ2

; yðoÞ a22 þ ðb2 oÞ2 ða1 b2 þ a2 b1 Þo ¼)jXðjoÞja1 ¼a2 ;b1 ¼b2 ¼ arctan a1 a2 b1 b2 o2 sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ a2 þ b2 o 2 ¼ jXðjoÞj ¼ a2 þ b2 o 2 ¼ 1; yðoÞ ¼ 2 arctanðbo=aÞ:

De Moivre’s theorem defines the power of a complex number c. For p a real number,

&

In Chapter 7, systematic methods for sketching the magnitude and phase representations of complex rational functions of o will be considered. MATLAB plots will be considered in Appendix B.

ðcÞp ¼ ðrejy Þp ¼ ½rðcos y þ j sin yÞp ¼ rp ½cosðpyÞ þ j sinðpyÞ:ðcÞ1=n y þ 2 kp y þ 2 kp 1=n þ j sin ; ¼r cos n n

1.6.2 Complex Periodic Functions

k ¼ 0; 1; 2; . . . ; n 1; n ¼ an integer:

xðtÞ ¼ Aejðo0 tþy0 Þ ¼ Acosðo0 t þ y0 Þ þ jAsinðo0 t þ y0 Þ: (1:6:13)

Approximation of the amplitude and the phase angle of a complex number: The magnitude pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃof ﬃ a complex number c ¼ a þ jb is jcj ¼ ða2 þ b2 Þ ¼ r. Representing c in terms of ða; bÞ is the rectangular coordinate representation and ðr; yÞ is the polar coordinate representation with y ¼ arctanðb=aÞ. Finding the square root of a number is not as simple as a multiplication. To save the computational time, the magnitude jcj is usually approximated. For a simple algorithm see Problem 1.6.1. Example 1.6.3 Consider the complex function below with the real variable o. a . Derive the rectangular and polar form expressions for XðjoÞ: XðjoÞ ¼

a1 jb1 o ; ai > 0; bi > 0; i ¼ 1; 2: (1:6:12) a2 þ jb2 o

b. Simplify the expressions when a1 ¼ a2 ¼ a and b1 ¼ b2 ¼ b:

For A; o0 ; and y0 constants, Euler’s identity is

Note that xðtÞ ¼ xðt þ TÞ ¼ xT ðtÞ with the period T ¼ 2p=o0 and ej2p ¼ 1. xT ðt þ ð2p=o0 ÞÞ ¼ Aeðjo0 ðtþð2p=TÞþy0 Þ ¼ Aejðo0 tþy0 Þ ejo0 ð2p=TÞ ¼ Aejðo0 tþy0 Þ

(1:6:14)

In Chapter 3, periodic functions xT ðtÞ will be approximated by the complex Fourier series expansion: xT ðtÞ ¼

1 X

Xs ½kejno0 t :

(1:6:15)

k¼1

1.6.3 Functions of Periodic Functions If a function xT ðtÞ is periodic, then gðxT ðt þ TÞÞ ¼ gT1 ðxT ðtÞÞ; T1 T:

(1:6:16)

28

1 Basic Concepts in Signals

That is, gðtÞ is also periodic with the same period T or smaller. One problem of interest is given the period of xT ðtÞ the period of the function gðtÞ is to be computed.

Example 1.6.6 Show that the function xT ðtÞ is a half-wave symmetric periodic function. xT ðtÞ ¼

b: g2 ðtÞ ¼ ejx2 ðtÞ ; x2 ðtÞ ¼ sinðo0 tÞ:

(1:6:17)

Solution: a. Noting the absolute value jcosðo0 tÞj, the fundamental period of g1 ðtÞ is T=2 ¼ ðp=o0 Þ. b. In this case, g2 ðtÞ is periodic with the period T ¼ ð2p=o0 Þ since g2 ðt þ TÞ ¼ ej sinðo0 ðtþTÞÞ ¼ ej sinðo0 tÞ :

(1:6:18) &

fa½2 k 1 cosð2 k 1Þo0 t

k¼1

Example 1.6.4 Find the fundamental periods of the following periodic functions: a: g1 ðtÞ ¼ jx1 ðtÞj; x1 ðtÞ ¼ cosðo0 tÞ;

1 X

þ b½2 k 1 sinð2 k 1Þo0 tg:

(1:6:21)

Solution: Noting that xT ðtÞ is an algebraic sum of sine and cosine terms with the same period T ¼ 2p=o0 , it is a periodic function. Furthermore, since o0 ðT=2Þ ¼ p and ð2 k 1Þ is an odd integer, we can use cos½2 k 1p ¼ 1 and sin½2 k 1p ¼ 0. Using these and the following, we have xðt þ ðT=2ÞÞ ¼ xðtÞ. cosð½2 k 1o0 ðt þ ðT=2ÞÞÞ ¼ cos½2 k 1o0 t cosð½2 k 1o0 ðT=2ÞÞ sin½2 k 1o0 t sinð½2 k 1o0 ðT=2ÞÞ

1.6.4 Periodic Functions with Additional Symmetries Even and odd functions apply for both energy and periodic signals. cosðo0 tÞ and sinðo0 tÞ are periodic functions with period ð2p=o0 Þ and have even and odd symmetries, respectively. Interestingly sine and cosine functions have four distinct parts in one period. These are for 0 t5T=4; T=4 t5T=2; T=2 t53T=4; and 3T=4 t5T. Having information for one-fourth of the period of a sine wave provides the information about the other three parts. This gives a clue on the sizes of the transmitting and receiving antennas to transmit and receive the sine wave, which is discussed in Chapter 10. Half-wave symmetric functions: A periodic function xT ðtÞ is half-wave symmetric if xT ðtÞ ¼ xT ðt ðT=2ÞÞ:

þ cos½2 k 1o0 t sinð½2 k 1o0 ðT=2ÞÞ

xT ðtÞ ¼

xT ðtÞ ¼ xT ðtÞ xT ðtÞ ¼ xT ðt þ T2 Þ

;

xT ðt þ TÞ ¼ xT ðtÞ : even quarter wave symmetry: (1:6:22) xT ðtÞ ¼ xT ðtÞ xT ðtÞ ¼ T ; xT ðtÞ ¼ xT ðt þ 2 Þ xT ðt þ TÞ ¼ xT ðtÞ : odd quarter wave symmetry (1:6:23) Example 1.6.7 Show that the function below xT ðtÞ has the even quarter-wave symmetry.

(1:6:20) xT ðtÞ ¼

Solution: Since xT ðtÞ is periodic with period T, we can write T T T & xT ðt Þ ¼ xT ðt þ T Þ ¼ xT ðt þ Þ: 2 2 2

&

Quarter-wave symmetric functions: If a periodic function xT ðtÞ has half-wave symmetry and, in addition, is either even or odd function, then it is said to have even or odd quarter-wave symmetry. That is,

(1:6:19)

Example 1.6.5 Show that if a function is half-wave symmetric, then xT ðtÞ ¼ xT ðt þ ðT=2ÞÞ:

sinð½2 k 1o0 ðt þ ðT=2ÞÞ ¼ sin½2 k 1o0 t cosð½2 k 1o0 ðT=2ÞÞ

1 X

a½2 k 1 cosðð2 k 1Þo0 tÞ:

(1:6:24a)

k¼1

Solution: Since xT ðtÞ is a sum of cosine terms with a zero phase, it is an even symmetric function. It has even quarter-wave symmetry since

1.7 Examples of Probability Density Functions and their Moments

T cosð½2 k 1o0 ðt þ ÞÞ 2 T ¼ cosð½2 k 1o0 tÞ cosð½2 k 1o0 Þ 2 T sinð½2 k 1o0 tÞ sinð½2 k 1o0 Þ 2 ¼ cosð½2 k 1o0 tÞ (1:6:24b) &

29

probability of existence of the noise is always positive. That is pðxÞ 0. Furthermore, the existence of noise is certain and therefore the integral of the probability density function must be 1. In summary, Z1 pðxÞ 0 and pðxÞdt ¼ 1: (1:7:1) 1

In a similar manner, it can be shown that the function below has odd quarter-wave symmetry. This is left as an exercise. yT ðtÞ ¼

1 X

b½2 k 1 sinðð2 k 1Þo0 tÞ:

For a good discussion on probability theory, see Peebles (2001). Any nonnegative function with area 1 can serve as a probability density function. The nth moment of pðxÞ is defined as

(1:6:25)

k¼1

mn ¼

As examples, Fig. 1.6.2a and b has even and odd quarter-wave symmetries, respectively. Hidden symmetries: The symmetries can be hidden within a constant as given by xT ðtÞ ¼ A þ A sinðo0 tÞ:

(1:6:26)

It is neither even nor an odd function. On the other hand ðxT ðtÞ AÞ is an odd function.

Fig. 1.6.2 (a) Even quarterwave symmetric function and (b) odd quarter-wave symmetric function

xn pðxÞdx:

(1:7:2)

1

The zero and the first moments are, respectively, defined by m0¼

Z1

pðxÞdx ¼ 1 and m1 ¼

1

Z1 xpðxÞdx: (1:7:3) 1

The moments about the mean are called the central moments and are defined by

1.7 Examples of Probability Density Functions and their Moments In this section, a brief introduction to the probability density function (PDF) pðxÞ of a random variable xðtÞ, such as the amplitude of a noise signal, is presented. Noise, by its nature, is unpredictable. It is assumed that the amplitude can take any value in a continuous range 15a5x5b51. The level of the noise amplitude can only be described in terms of averages. Since noise is ever present,

Z1

mn ¼

Z1

ðx m1 Þn pðxÞdx:

(1:7:4a)

1

¼)m2 ðxÞ ¼ Variance of xðtÞ ¼ s2x Z1 ¼ ðx m1 Þ2 pðxÞdx:

(1:7:4b)

1

pﬃﬃﬃﬃﬃ The positive square root of the variance sx ¼ þ s2x is called the standard deviation. It gives a measure of the spread of the probability density function. Now

xT (t)

xT (t)

(a)

(b)

30

1 Basic Concepts in Signals

m2 ¼

Z1

ðx m0 Þ2 pðxÞdx ¼

1

Z1

Z1

x2 pðxÞdx 2m0

ðða þ bÞ=2Þ2 ¼

1

Z1

xpðxÞdxþ m21

1

Variance :s2x ¼ m2 m21 ¼ ð1=3Þ½b2 þ ab þ a2

pðxÞdx ¼ m2 2m21 þ m21

1

¼ m2 m21 :

(1:7:5) s2x ¼ m2 m21 :

(1:7:6)

Mean and the variance are basic statistical values in the study of the p probability density functions. In ﬃﬃﬃﬃﬃ addition, sx ¼ þ s2x measures its effective width or duration.

ðb aÞ2 : 12

(1:7:10)

pﬃﬃﬃ Standard deviation ¼ sx ¼ ðb aÞ=2 3: (1:7:11) & The function in Fig. 1.7.1 is the uniform density function, as the variable x is equally likely to take any value in the range ½a; b. It will be used in Chapter 10 to describe the error caused by quantization of samples. Example 1.7.2 Consider the Gaussian probability density function shown in Fig. 1.7.2 and it is given in (1.7.12). Determine m0 ; m1 ; and s2x of this PDF.

Example 1.7.1 Determine m0 ; m1 ; and s2x for the function in Fig. 1.7.1.

p(x)

p(x)

x

x Fig. 1.7.1 Uniform density function

1 x ððb þ aÞ=2Þ P ðuniform PDFÞ: pðxÞ ¼ ba ba (1:7:7) Solution: By inspection, we have the area under the function is 1. That is, m0 ¼ 1: Also m1 ¼

Z1

xpðxÞdx ¼

1

Zb

(1:7:12) Solution: From tables, Z1 Z1 1 1 2 ðtax Þ2 =s2x e dt ¼ pﬃﬃﬃﬃﬃﬃ ey =2 dy ¼ 1: m0 ¼ pﬃﬃﬃﬃﬃﬃ 2psx 2p 1

1

(1:7:13)

a

(1:7:8)

Z1 1

3

1b a 1 ¼ ðb2 þ ab þ a2 Þ: 3 ba 3

That is, the area under the Gaussian function is equal to 1. The mean value is m1 ¼

a

¼

2 1 2 pðxÞ ¼ pﬃﬃﬃﬃﬃﬃ eðxax Þ =sx ðGaussian PDFÞ: 2pðsx Þ

x dx ba

bþa ðaverage value of xÞ: ¼ 2 Zb 2

2 t m2 ¼ Area t xðtÞ ¼ dt ba 3

Fig. 1.7.2 Gaussian density function

(1:7:9)

1 xpð xÞdt ¼ pﬃﬃﬃﬃﬃﬃ 2psx

Z1

xeðxax Þ

2

=2s2x

dx:

1

(1:7:14) Using the change of variable y ¼ ðx ax Þ=sx , we have

1.8 Generation of Periodic Functions from Aperiodic Functions

1 m1 ¼ pﬃﬃﬃﬃﬃﬃ 2p

Z1

ax 2 ðsx y þ ax Þey =2 dy ¼ pﬃﬃﬃﬃﬃﬃ 2p

1 Z1

sx þ pﬃﬃﬃﬃﬃﬃ 2p

yey

2

=2

Z1

ey

2

=2

dy

1

(1:7:15)

dy¼ ax :

1

These follow since the second one above on the right is an integral of an odd function of y over a symmetrical interval and it is zero. The first integral in (1.7.15) reduces to (1.7.13). To derive the variance s2x , start with 1 m0 ¼ pﬃﬃﬃﬃﬃﬃ 2psx Z1

Z1

2

2

eðxax Þ =2sx dt ¼ 1 or

1

2

2

eðxax Þ =sx dt ¼

pﬃﬃﬃﬃﬃﬃ 2psx :

31

Notes: Noise is random and unpredictable. When it is added to the information bearing signal, the message signal is masked or even obliterated. Noise cannot be eliminated. A measure of corruption of the signal by noise is an important measure. It is the ratio of the average signal power to variance of the noise. It is Signaltonoise ratio ¼ SNR Average message signal power ¼ : Variance of the noise;s2x

1.8 Generation of Periodic Functions from Aperiodic Functions Now we like to construct a periodic function from an aperiodic function, say jðtÞ.

1

Taking the derivative with respect to sx results in pﬃﬃﬃﬃﬃﬃ 2p ¼

Z1 1

¼

Z1 1

¼)s2x

1 ¼ pﬃﬃﬃﬃﬃﬃ 2psx

deðxax Þ dsx

2

yT ðtÞ ¼

=2s2x

dx

ðx ax Þ ðtax Þ2 =2s2x e dx: s3x 2

2

ðx ax Þ2 eðxax Þ =2sx dx: (1:7:16)

1

Gaussian PDF (see Peebles (2001)) is one of the most important PDF, as most of the noise processes & observed in practice are Gaussian. Example 1.7.3 Find m0 and s2x for the Laplace PDF defined by pðxÞ ¼ ðb=2Þe

bj xj

; b > 0; 15x51:

(1:7:17)

Solution: From integral tables, the mean and the variance are Z1 Z1 b b jxj e m0 ¼ dx ¼ b ebx dx ¼ ebx 1 x¼0 ¼ 1; 2 s2x ¼

1 Z1 1

0

b 2 jxj=b 2 x e dx ¼ 2 : 2 b

1 X

jðt þ kTÞ:

(1:8:1)

k¼1

2

Z1

&

jðtÞ is the principal segment of the periodic extension. Clearly, yT ðtÞ is a periodic function with period T s and is the periodic extension of jðtÞ. If jðtÞ is not time limited to a T s interval (for example, jðtÞ is nonzero for t > T and t 5 0), then jðtÞ and jðt þ TÞ terms will overlap and jðtÞ cannot be extracted from yT ðtÞ. Example 1.8.1 Using the principal segment jðtÞ ¼ L½t=t, sketch the periodic extensions for the following cases. a: T 2t; b: T52t. Solution: The periodic extension of the triangular function is yT ðtÞ ¼

1 X k¼1

L½

t þ kT : t

(1:8:2)

The function jðtÞ and its periodic extensions are sketched in Fig. 1.8.1a, b, and c. For simplicity, in the sketch for part a, T ¼ 2t is assumed. a. The functions L½ðt þ kTÞ=t and L½ðt þ ðk þ 1Þ TÞ=t do not overlap and the function jðtÞ can be extracted

(1:7:18) & yT ðtÞjk¼0 ¼L½t=t ¼ jðtÞ; jtj t:

(1:8:3)

32

1 Basic Concepts in Signals

yT (t),T = 2τ

(a)

yT (t),T < 2τ

(b)

(c)

Fig. 1.8.1 (a) L tt , (b) yT ðtÞ; T ¼ 2t, (c) yT ðtÞ; T52t

b. For T52t, L½ðt þ kTÞ=t and L½ðt þ ðk þ 1ÞTÞ=t overlap (See Fig. 1.8.1b). Recovery of fðtÞ from yT ðtÞ: is not possible. It is interesting to note the area of one period of the periodic extension of jðtÞ equals to the area of the function jðtÞ. This is a consistency check. See Ambardar (1995). Example 1.8.2 Show that the area under one period of the periodic extension of jðtÞ ¼ et=t uðtÞ is equal to the area under jðtÞ.

ZT

x1 ðtÞdt ¼ ½1 þ eT=t þ e2T=t þ . . .

0

ZT e

t=t

1 dt ¼ 1 eT=t

0

ZT e

t=t

1 eT=t 1 1 ¼ : dt ¼ T=t t 1e t

0

Z1

jðtÞdt ¼

0

Z1

1 et=t uðtÞdt ¼ : t

(1:8:6) &

0

Solution: The periodic extension of jðtÞ with a period of T can be written as yT ðtÞ ¼½eðt=tÞ uðtÞ þ eðtþTÞ=t uðt þ TÞ þ eðtþ2TÞ=t uðt þ 2TÞ þ þ ½eðtTÞ=t uðt TÞ þ eðt2TÞ uðt 2TÞ þ ...

(1:8:4)

x1 ðtÞ ¼ et=t ½uðtÞ þ eT=t uðt þ tÞ þ e2T=t uðt þ 2tÞ þ . . .; t > 0:

(1:8:5)

1.9 Decibel Decibel or dB is a logarithmic unit named after Alexander Graham Bell that is used to express power ratios. Given two powers P 1 and P 2 and their ratio ðP 2 =P 1 Þ, then

1.9 Decibel

33

Table 1.9.1 Sound Power (loudness) Comparison Threshold of audibility 0 dB Whisper 15 dB Average home 45 dB Riveting machine (30’ away) 100 dB Threshold of hearing 120 dB Jet plane 140 dB

Power ratio in dB ¼ 10 log10 ½P2 =P1 :

(1:9:1)

For computing the dB values from the amplitudes using MATLAB, see Section B.12 in Appendix B. The unit of bel is too large. For example, a human ear can detect audio power level differences of onetenth of a bel or 1 dB. The loudness of a few typical activities are shown in table 1.9.1. Also, 1 dB is approximately equal to the attenuation of one mile of a standard telephone cable. Power ratio is expressed by Power ratio ¼ ðP2 =P1 Þ ¼ 10dB=10 :

(1:9:2)

Even though decibels were originally meant to be used with respect to power ratios, they can be used to express absolute values of power interpreted as a ratio of power P2 to P 1 ¼ 1 Watt, referred to as 1 dBW. That is, 1 Watt is a reference and 10 log10 ðP2 =P1 Þ ¼ 10 log10 ðP2 in W=1 WattÞ dB W: (1:9:3) Positive decibels imply A ¼ P 2 =P 1 > 1, zero decibels imply A ¼ 1, and negative decibels imply A51. Decibels are used to express gains or losses in a system; gain is the output divided by input and loss is the input divided output. Instead of using 1 Watt as a reference, 1 mW can be used as a reference to compare small signal power levels, such as powers of radar echoes and the unit is dBm. Now

Table 1.9.2 Power ratios and their corresponding values in dB dB 0 1 2 3 4 5 6 7 8 9 Power ratios 1 1.26 1.6 2 2.5 3.2 4 5 6.3 8

The logarithmic decibel function provides a greater resolution when the power ratio is small, indicating a good way to recognize very small differences in the power levels (Table 1.9.2). The smaller the power ratio, that is less than 1, the larger the number of negative dBs required. As the ratio approaches zero, the negative dB increases without limit. Interestingly when you round of to the nearest whole decibel, the error in the power ratio is at most only 1 part in 7. The dB provides greater resolution when the power ratio is small. Small differences in the power levels are important in spectral analysis, analog and digital filter designs, control system designs, communication system designs, etc. A power ratio of 2 to 1 (1 to 2) is 3 dB (3 dB). Bandwidths of 3 dB play a major role in filter designs. A ratio of 108 to is only 80 dB and 108 is 80 dB. Power levels in radars have a large range. Dealing with large number of digits can be troublesome, as dropping a zero at the end of a large number makes the radar calculations wrong. The dB scale makes the numbers compressed. The power levels associated with seismic signals are low. One nice thing about dB scale is that we will be dealing with additions and subtractions rather than multiplications and divisions. Gain (or loss) is the term used for an increase (or decrease) in power level. For example, for an amplifier, Gain Output signal power coming out of the amplifier ¼ : Input signal power going into the amplifier (1:9:5)

P2 =P1 jP1 ¼1 mW ¼ 10 log10 ðP2 in Watts=1 mWÞ dB m: If the output power is 100 times the input power, (1:9:4) then As examples, a power of 1 Watt is 0 dBW, a power of 3 Watt is 4.77 dBW, and a power of 1 kW is 30 dBW. A power of 1 kilowatt is 0 dBm and a power of 1010 mW is 10 log10 ð1010 mW=1 mWÞ ¼ 100 dBm:

Gain dB ¼ 10 logð100Þ ¼ 20 dB:

(1:9:6)

If we use a wave guide or a cable, we have a loss in the power. That is, Loss ¼

Input power to the device : Output power of the device

(1:9:7)

34

1 Basic Concepts in Signals

As an example, let an amplifier be connected to an antenna by a waveguide and the guide absorbs 20% of the power. If the ratio of the input power to the output power is 10 to 8, i.e., 1.25 indicating a loss of: Loss in dB ¼ 10 logð1:25Þ ﬃ :9 dB ðor; gainisð :9Þ dBÞ::

(1:9:8)

Since logðABÞ ¼ logðAÞ þ logðBÞ, we can simply find the power gain (or loss) in dB by adding or subtracting the numbers in dB. This is illustrated in the example below. Example 1.9.1 Consider that the amplifier considered above is connected to an antenna by a waveguide. See Fig. 1.9.1. Assume the amplifier gain is Gamplifier =100 and the waveguide absorbs about 20% of the power. Determine the total gain.

power gain or loss by adding or subtracting the appropriate quantities. It can be obtained by finding the transfer functions of each block and determine the total transfer function. Since the transfer functions are functions of frequency, the power gains or losses are functions of frequencies. Notes: Filters keep approximately the same amplitude in the filter pass band and provide attenuation in the filter stop band. If HðjoÞ is the transfer function of the filter, then we have attenuation (gain) at a frequency o1 ¼ 2pf1 , if jHðjo1 Þj 1 (jHðjo1 Þj > 1). The corresponding attenuation and gain are as follows: a ¼ 20 logjHðjo1 Þj dB > 0 ðLossÞ; A ¼ 20 logjHðjo1 Þj ðGainÞ:

(1:9:11)

Solving for jHðjoÞj from (1.9.11) results in Loss : jHðjo1 Þj ¼ 1=ð10:05 Þa ; Gain :jHðjo1 Þj ¼ ð10:05 ÞA :

(1:9:12)

Loss in terms of dB and the corresponding decrease in amplitudes are given below: Fig. 1.9.1 Amplifier and a waveguide

1dB¼)approximately 10% decrease in jHj Solution: The total gain from the input to the amplifier to the antenna is GTotal ¼ Gamplifier =Gwave guide ¼ 100=1:25 ¼ 80: (1:9:9) We can determine the gain or loss by simply subtracting gain from the loss. That is,

¼) Power ratio ¼ Pr ¼ 10ð19:03=10Þ ’ 80:

2dB¼)approximately 20% decrease in jHj from 1 to :794: 3dB¼)approximately 30% decrease in jHj from 1 to :708: 6dB¼)approximately 50% decrease in jHj from 1 to :501:

ðGTotal ÞdB ¼ ðGamplifier ÞdB ðGwave guide ÞdB ﬃ 20 :97 ¼ 19:03 dB:

from 1 to :891:

&

(1:9:10) &

Systems are designed part by part, arranged in a cascade, generally drawn symbolically by a block diagram shown in Fig. 1.9.2. We assume that the system identified by block k is not loading the system identified by block (k1). That is, there is no loading effect. Corresponding to this we can compute the

1.10 Summary In this chapter some of the basics on signals are presented that a second semester junior in an electrical engineering program may have gone through. Some of the material, such as complex numbers, periodic functions, integrals, decibels, and others, are included to refresh the reader’s memory. Specific principal topics that were included are

Various types of continuous signals Useful signal operations involving time shifting, Fig. 1.9.2 A cascaded system

scaling, reversal, and amplitude shift

Problems

35

Approximations and simplifications for integrals

with symmetries Singularity functions that include impulse functions, step functions, etc., and functions that can be used to approximate impulses Signal classifications based on power and energy Periodic signals and special classes of periodic functions with symmetries Complex numbers and complex functions Energy signals and their moments Periodic extension of aperiodic functions Decibels

1.1.1 Peterson and Barney (1952) collected average formant frequencies for vowels spoken by adult male and female speakers. For example, first two average formant frequencies in Hertz for the two vowels =i= and =a= are given in the table below. For a particular subject the first two formant frequencies of one of the vowels given were determined and they are F1 ¼ 500 Hz and F2 ¼ 1600 Hz. Using the minimum distance classifier, and the below table, determine if this subject is a male or a female and what is the vowel =i= or =a=?

Fig. P1.2.1

1.2.1 Given the functions in Fig. P1.2.1, sketch a: xi ðt þ 1Þ; b: xi ðt 1Þ; c: xi ð2t 3Þ; d: xi ð2t þ 1Þ; i ¼ 1; 2:

jij

jaj

270 310 2290 2790

730 850 1090 1220

a: x1 ðtÞ ¼ et uðt 1Þ; b: x2 ðtÞ ¼ et ; c: x3 ðtÞ ¼ ½ðt þ 1Þ=ðt 1Þ: d: x4 ðtÞ ¼ cosðtÞ þ sinðtÞ; e: x5 ðtÞ ¼ P½ðt 1Þ=2: 1.2.3 Sketch the functions sincðplÞ and sinc2 ðplÞ and give an approximate value of the magnitude of the first side lobe and their values in dB. Use MATLAB if you have the access. See Appendix B for information on MATLAB. 1.3.1 Approximate the following integral for the intervals shown by using a. the rectangular integration formula and b. the trapezoidal integration formula.

1.1.2 Sketch the following:

Zp

p xðtÞdt; xðtÞ ¼ sinðtÞ; intervals : 0; ; 4 0 p p p 3p 3p ; ; ; ; ;p : 4 2 2 4 4

A¼

x2(t)

1.2.2 Find the even and odd parts of the following functions:

Problems

F1 Male Female F2 Male Female

x1(t)

h i t1 2t t1 t P ; b: x2 ðtÞ ¼ P :L ; Use xðtÞ at t ¼ kðp=4Þ; k ¼ 0; 1; 2; 3 to approximate 2 2 2 2 the area by assuming each strip is a rectangle or a c: x3 ðtÞ ¼ uðt1Þ:uðbtÞ for b ¼ 5;0;1;2 trapezoid. Compare these values to the actual value of the integral. 1.1.3 Response of a first-order system is given by 1.3.2 Evaluate the following integral: yðtÞ ¼ Að1 et=2 ÞuðtÞ. What is the time constant of the system and compute the value of the time at ZT=2 1 t the time yðtÞ reaches 63.21% of the response? What sinðo0 tÞdt: 2 T do we call this time?

a: x1 ðtÞ ¼ P

0

36

1 Basic Concepts in Signals

1.3.3 Derive the expressions for the following partials:

Z5 b:

Z1

@uðs; oÞ @vðs; oÞ @uðs;oÞ @vðs; oÞ ; ; ; and : @s @s o @o

a: x1 ðtÞ ¼ sinðtÞ; b: x2 ðtÞ ¼ cosðtÞ; c: x3 ðtÞ ¼ tanðtÞ; d: x4 ðtÞ ¼ secðtÞ; e: x5 ðtÞ ¼ cotðtÞ; f: x6 ðtÞ ¼ cscðtÞ: 1.4.2 Let xðtÞ ¼ dððt 1Þðt 2ÞÞ. Express the function in terms of a sum of impulses. 1.4.3 Show the following functions are limiting forms of impulse functions: a: x1 ðtÞ ¼ ðT=2ÞejtjT ;

b: x2 ðtÞ ¼ Tð1 ðjtj=TÞÞ; Tsin Tt T sin Tt 2 c: x3 ðtÞ ¼ ; d: x4 ðtÞ ¼ p tT p Tt 1 T 2 2 e: x5 ðtÞ ¼ Tept T ; f: x6 ðtÞ ¼ 2 p t þ T2

1.4.4 Show that xn ðtÞ can be used as an impulse representation

dðtÞ ¼ lim xn ðtÞ; xn ðtÞ ¼ :5 nenjtj n!1 Z 1 use xn ðtÞdt ¼ 1; lim xn ðtÞ ¼ 0 : n!1

1

a: d0 ðtÞ ¼ d0 ðtÞ; b: td0 ðtÞ ¼ dðtÞ; R1 dxðtÞ 0Þ c: xðtÞ ddðtt dt ¼ ; dt dt d:

t¼t0

1 e ¼ ddðtÞ dt ; xe ðtÞ ¼ p t2 þe2 ; e > 0:

1.4.6 Evaluate the following integrals using the properties of the impulse functions: Z5 a: 0

ðt2 þ 2t 1Þdðt 1Þdt;

c:

et uðtÞ

dðt 1Þ dt: dt

1

1.5.1a. Express an arbitrary real-valued signal in terms of its even and odd parts. b. Express the unit step function uðtÞ in terms of its even and odd parts and sketch the even and odd parts. Assume uð0Þ ¼ 1 in sketching the function. 1.5.2 Classify each of the following functions as either an energy signal or a power signal or neither. If the functions are either energy or power signals, give the corresponding energy or power. Otherwise, explain why they are not. t1 a: x1 ðtÞ ¼ L ; b: x2 ðtÞ ¼ eajtj ; a > 0; 2 c: x3 ðtÞ ¼ t et uðtÞ;

d: x4 ðtÞ ¼ sincðptÞ;

e: x5 ðtÞ ¼ 1=½pð1 þ t2 Þ;

2

f: x6 ðtÞ ¼ ept ;

g: x7 ðtÞ ¼ d0 ðtÞ; h: x8 ðtÞ ¼ sin2 ðtÞ: 1.5.3 Show that Z1 1

2

x ðtÞdt ¼

Z1 1

x2e ðtÞdt

þ

Z1

x20 ðtÞdt:

1

1.5.4 Given xðtÞ ¼ cosð2pð1000ÞtÞ, find the period of xðatÞ; a > 0. What can you say in the general case of a periodic function xT ðtÞ and xT ðatÞ; a 6¼ 0? 1.5.5 Find the mean, rms, and the peak values of the function

1.4.5 Show the following:

1 lim dxdte ðtÞ e!0

ddðt 1Þ dt; dt

0

GðjoÞ ¼ 1=ðs þ joÞ ¼ uðoÞ þ jvðoÞ;

and yi ðtÞ ¼ uðxi ðtÞÞ; 1.4.1 Sketch xi ðtÞ i ¼ 1; 2; 3; 4; 5; 6 for p=25t53p=2 given

ðt2 þ 2t 1Þ

xðtÞ ¼ A cosðo0 t þ f1 Þ þ B cosð2o0 t þ f2 Þ: What is the effect of the phase angles on the peak value of this function? 1.6.1 The power series expansion of the square root function is Spiegel (1968) qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ða2 þ b2 Þ ¼ a ð1 þ x2 Þ 1 1 2 1:3 3 ¼a 1þ x x þ x ; 2 2ð4Þ ð2Þð4Þ6Þ b x ¼ ; 1 5 x 1: a

Problems

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Simplest approximation: a2 þ b2 a þ ðb=2Þ. For c ¼ 1 þ j1, find jcj by the approximation and the error between the direct computation and the simplified method. 1.6.2 Solve for the roots of the polynomial 1 þ ð1Þn s2n ¼ 0 in general terms. The roots may be complex. Give the roots on the left half-plane for n ¼ 1; 2; 3; 4; 5. Plot these roots on a complex plane and comment on the magnitudes of these roots. 1.6.3 Give the even and odd parts of the function xðtÞ ¼ t; 05t5T=2 and zero elsewhere. 1.6.4 Give examples of half-wave, even, and odd quarter-wave symmetric functions. 1.6.5 Show that the function given in (1.6.25) has odd quarter-wave symmetry. 1.6.6 Consider the periodic function xT ðtÞ given below. What can you say about the hidden symmetry in this periodic function?

37

a: xL ðtÞ ¼ ðb=2Þebjtaj ; 15t51; b > 0 : Laplace function: 2

b: xR ðtÞ ¼ ð2=bÞðt aÞeðtaÞ =b uðt aÞ; 15t51; b > 0 : Rayleigh function: 1.8.1 Assuming T ¼ t; b: T ¼ t=2, give the expressions for the periodic extension of the function x1 ðtÞ ¼ P½t=t. 1.9.1 Show that if you round off to the nearest whole decibel, the error in the power at most 1 in 7 by noting that the plot of decibel versus power ratio in the interval between 0 and 1 is approximately a straight line (Simpson, Hughes Air Craft Company, 1983). 1.9.2 Convert the following to power ratios and approximate them in dB: pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃﬃﬃ Magnitude ratios : 1= 10; 1=2; 1= 2; 1; 2; 2; pﬃﬃﬃﬃﬃ 10; 5; 10; 100; 1000:

1.9.3 Two radar signals x1 ðtÞ and x2 ðtÞ are assumed xT ðtÞ ¼ ð1 ðt=TÞÞ; 05t5T and xT ðt þ TÞ ¼ xðtÞ: to have an average power of 3 dBm and10 dBm, 1.7.1 Determine the mean and the variances of the respectively. What are the corresponding absolute power levels? following two functions:

Chapter 2

Convolution and Correlation

2.1 Introduction In this chapter we will consider two signal analysis concepts, namely convolution and correlation. Signals under consideration are assumed to be real unless otherwise mentioned. Convolution operation is basic to linear systems analysis and in determining the probability density function of a sum of two independent random variables. Impulse functions were defined in terms of an integral (see (1.4.4a)) using a test function fðtÞ. Z1

fðtÞdðt t0 Þdt ¼ fðt0 Þ:

(2:1:1)

1

This integral is the convolution of two functions, fðtÞ and the impulse function dðtÞ to be discussed shortly. In a later chapter we will see that the response of a linear time-invariant (LTI) system to an impulse input dðtÞ is described by the convolution of the input signal and the impulse response of the system. Convolution operation lends itself to spectral analysis. There are two ways to present the discussion on convolution, first as a basic mathematical operation and second as a mathematical description of a response of a linear time-invariant system depending on the input and the description of the linear system. The later approach requires knowledge of systems along with Fourier series and transforms. This approach will be considered in Chapter 6. Although we will not be discussing random signals in any detail, convolution is applicable in dealing with random variables.

The process of correlation is useful in comparing two deterministic signals and it provides a measure of similarity between the first signal and a timedelayed version of the second signal (or the first signal). A simple way to look at correlation is to consider two signals: x1 ðtÞ and x2 ðtÞ. One of these signals could be a delayed, or an advanced, version of the other. In this case we can write x2 ðtÞ ¼ x1 ðt þ tÞ; 1 < t < 1. Multiplying point by point and adding all the products, x1 ðtÞx1 ðt þ tÞ will give us a large number for t ¼ 0, as the product is the square of the function. On the other hand if t 6¼ 0, then adding all these numbers will result in an equal or a lower value since a positive number times a negative number results in a negative number and the sum will be less than or equal to the peak value. In terms of continuous functions, this information can be obtained by the following integral, called the autocorrelation function of xðtÞ, as a function of t not t. Z1 Rxx ðtÞ ¼ xðtÞxðt þ tÞdt ¼ AC ½xðtÞ Rx ðtÞ: 1

(2:1:2) This gives a comparison of the function xðtÞ with its shifted version xðt þ tÞ. Autocorrelation (AC) provides a nice way to determine the spectral content of a random signal. To compare two different functions, we use the cross-correlation function defined by

Rxh ðtÞ ¼ xðtÞ hðtÞ ¼

Z1

xðtÞhðt þ tÞdt: (2:1:3)

1

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_2, Ó Springer ScienceþBusiness Media, LLC 2010

39

40

2 Convolution and Correlation

Note the symbol (**) for correlation. Correlation isIf xðtÞ and yðtÞ are orthogonal, the energy and power related to the convolution. As in autocorrelation, contained in the energy or power signal the cross-correlation in (2.1.3) is a function of t, the zðtÞ ¼ xðtÞ þ yðtÞ are respectively given by time shift between the function xðtÞ, and the shifted version of the function hðtÞ. Ez ¼ Ex þ Ey or Pz ¼ Px þ Py : (2:1:8)

2.1.1 Scalar Product and Norm

Some of the important properties of the norm are stated as follows:

The scalar valued function hxðtÞ; yðtÞi of two signals xðtÞ and yðtÞ of the same class of signals, i.e., either energy or power signals, is defined by 8 R1 > > xðtÞy ðtÞdt; > > > 1 > > > < energysignals: ð2:1:4aÞ hxðtÞ;yðtÞi¼ T=2 R > > > lim T1 xðtÞy ðtÞdt; > > T!1 > T=2 > > : powersignals: ð2:1:4bÞ Superscript (*) indicates complex conjugation. Our discussion will be limited to a subclass of power signals, namely periodic signals. In that case, assuming that both the time functions have the same period (2.1.4b) can be written in the symbolic form as follows: 1 hxT ðtÞyT ðtÞi ¼ T

Z

xðtÞy ðtÞdt:

T

1: kxðtÞk ¼ 0 if and only if xðtÞ ¼ 0;

(2:1:9a)

2: kxðtÞ þ yðtÞk kxðtÞk þ kyðtÞk; triangular inequality

(2:1:9b)

3: kaxðtÞk ¼ jajkxðtÞk:

(2:1:9c)

In (2.1.9c), a is some constant. One measure of distance, or dissimilarity, between xðtÞ and yðtÞ is kxðtÞ yðtÞk. A useful inequality is the Schwarz’s inequality given by jhxðtÞ; yðtÞij kxðtÞkkyðtÞk:

(2:1:9d)

(2:1:4c) The two sides are equal when xðtÞ or yðtÞ is zero or if yðtÞ ¼ axðtÞ where a is a scalar to be determined. This can be seen by noting that

Even though our interest is in real functions, for generality, we have used complex conjugates in the above equations. The norm of the function is defined by Ex ; energy signals : kxðtÞk ¼ hxðtÞ; xðtÞi1=2 ¼ Px ; power signals (2:1:5)

kxðtÞ þ ayðtÞk2 ¼ hxðtÞ þ ayðtÞ; xðtÞ þ ayðtÞi ¼ hxðtÞxðtÞi þ a hxðtÞ; yðtÞi þ ahxðtÞ; yðtÞi þjaj2 hyðtÞ; yðtÞi ¼ kxðtÞk2 þa hxðtÞ; yðtÞi þ ahxðtÞ; yðtÞi þjaj2 kyðtÞk2 : (2:1:10)

It gives the energy or power in the given energy or the power signal. The two functions, xðtÞ and yðtÞ, are orthogonal if Since a is arbitrary, select hxðtÞ; yðtÞi ¼ 0:

(2:1:6) a ¼ hxðtÞ; yðtÞi=kyðtÞk2 :

(2:1:11)

In that case, kxðtÞ þ yðtÞk2 ¼ kxðtÞk2 þkyðtÞk2 :

(2:1:7)

Substituting this in (2.1.10), the last two terms cancel out, resulting in

2.2

Convolution

41

2 hxðtÞ; yðtÞi

This definition describes a higher algebra and allows us to study the response of a linear timekxðtÞ þ ayðtÞk ¼ kxðtÞk kyðtÞk2 invariant system in terms of a signal and a system response to be discussed in Chapter 6. It should be ) kxðtÞk2 kyðtÞk2 hxðtÞ; yðtÞi2 0: emphasized that the end result of the convolution (2:1:12) operation is a function of time. Coming back to the sifting property of the impulse functions, consider Equality exists in (2.1.9d) only if xðtÞ þ ayðtÞ ¼ 0. the equation given in (2.1.1). Two special cases are Another possibility is the trivial case being either one of interest. of the functions or both are equal to zero. Ziemer Z1 and Tranter (2002) provide important applications fðaÞdðt aÞda fðtÞ dðtÞ ¼ on this important topic. 1 Correlations in terms of time averages: Cross(2:2:2a) Z1 correlation and autocorrelation functions can be expressed in terms of the time average symbols and ¼ fðt bÞdðbÞdb ¼ dðtÞ fðtÞ; 2

Rxh ðtÞ ¼

Z1

2

1

xðtÞhðt þ tÞdt ¼ hxðtÞhðt þ tÞi; (2:1:13a)

1

dðtÞ dðtÞ ¼

Z1

dðaÞdðt aÞda ¼ dðtÞ: (2:2:2b)

1

RT;xh ðtÞ ¼

ZT=2

1 T

xT ðtÞhT ðt þ tÞdt

T=2

¼

1 T

Z

(2:1:13b)

2.2.1 Properties of the Convolution Integral

xT ðtÞhT ðt þ tÞdt ¼hxT ðtÞhT ðt þ tÞi: 1. Convolution of two functions, x1 ðtÞ and x2 ðtÞ, satisfies the commutative property,

T

In the early part of this chapter we will deal with convolution and correlation associated with aperiodic signals. In the later part we will concentrate on convolution and correlation with respect to both periodic and aperiodic signals. Most of the material in this chapter is fairly standard and can be seen in circuits and systems books. For example, see Ambardar (1995), Carlson (1975), Ziemer and Tranter (2002), Simpson and Houts (1971), Peebles (1980), and others.

yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ x2 ðtÞ x1 ðtÞ:

(2:2:3)

This equality can be shown by defining a new variable, b ¼ t a, in the first integral in (2.2.1) and simplifying the equation. 2. Convolution operation satisfies the distributive property, i.e., x1 ðtÞ ½x2 ðtÞ þ x3 ðtÞ ¼ x1 ðtÞ x2 ðtÞ þ x1 ðtÞ x3 ðtÞ:

2.2 Convolution

(2:2:4)

The convolution of two functions, x1 ðtÞ and x2 ðtÞ, is defined by Z1 x1 ðaÞx2 ðt aÞda yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ ¼

Z1 1

1

x2 ðbÞx1 ðt bÞdb ¼ x2 ðtÞ x1 ðtÞ:

(2:2:1)

3. Convolution operation satisfies the associative property, i.e., x1 ðtÞ ðx2 ðtÞ x3 ðtÞÞ ¼ ðx1 ðtÞ x2 ðtÞÞ x3 ðtÞ: (2:2:5) The proofs of the last two properties follow from the definition. 4. The derivative of the convolution operation can be written in a simple form and

42

2 Convolution and Correlation

2 1 3 Z dyðtÞ d y 0 ðtÞ ¼ x1 ðaÞx2 ðt aÞda5 ¼ 4 dt dt

¼

1

¼

Z1

x1 ðaÞ

Z1

dx2 ðt aÞ db ¼ x1 ðtÞ x02 ðtÞ; dt

Z1

¼

1

2

dyðtÞ d ) ¼ ½x1 ðtÞ x2 ðtÞ dt dt dx1 ðtÞ dx2 ðtÞ ¼ x2 ðtÞ ¼ x1 ðtÞ : dt dt

¼4

3 x2 ða bÞda5x1 ðbÞdb

1

2 4

Ztb

3 x2 ðlÞdl5x1 ðbÞdb;

1

3

2

x2 ðlÞdl5x1 ðtÞ¼4

x1 ðbÞdb5x2 ðtÞ:

Example 2.2.1 Find the convolution of a function xðtÞ and the unit step function uðtÞ and show it is a running integral of xðtÞ: Solution: This can be seen from Z1 xðbÞuðt bÞdb xðtÞ uðtÞ ¼

m

d x1 ðtÞ d yðtÞ x2 ðtÞ ¼ m dt dtm

¼

1 Zt

xðbÞdb; 1

ðnÞ x2 ðtÞ

3

1

(2:2:6b) d i xi ðtÞ ðmÞ ðmÞ ; ¼ y ðtÞ Note xi ðtÞ ¼ dti

ðmÞ x1 ðtÞ

Zt

(2:2:8)

Equation (2.2.6a) can be generalized for higher order derivatives. We can then write m

Zt

Zt

1

(2:2:6a)

ðmÞ

4

1

1

x1 ðtÞ x2 ðtÞ ¼

2

1; uðt bÞ ¼ 0;

b5t : b4t

(2:2:9) &

d m x1 ðtÞ dn x2 ðtÞ ¼ dtm dtn (2:2:6c) d mþn yðtÞ ¼ ¼ yðmþnÞ ðtÞ: dtmþn

Since the impulse function is the generalized derivative of the unit step function uðtÞ (see Section 1.4.2.), we have yðtÞ ¼ uðtÞ hðtÞ ) y0 ðtÞ

6. Convolution of two delayed functions x1 ðt t1 Þ and x2 ðt t2 Þ are related to the convolution of x1 ðtÞ and x2 ðtÞ. yðtÞ ¼ x1 ðtÞ x2 ðtÞ ) x1 ðt t1 Þ x2 ðt t2 Þ (2:2:10) ¼ yðt ðt1 þ t2 ÞÞ: This can be seen from

(2:2:7) ¼ u0 ðtÞ hðtÞ ¼ dðtÞ hðtÞ ¼ hðtÞ: 5. Convolution is an integral operation and if we know the convolution of two functions and desire to compute its running integral, we can use Zt

Zt yðaÞda¼

1

¼

1 Zt 1

x1 ðt t1 Þ x2 ðt t2 Þ Z1 ¼ x1 ða t1 Þx2 ðt a t2 Þda

¼

4

Z1 1

x1 ðbÞx2 ð½t ðt1 þ t2 Þ bÞdb

1

½x1 ðaÞ x2 ðaÞda 2

1 Z1

¼ yðt ðt1 þ t2 ÞÞ:

(2:2:11)

3 x1 ðbÞx2 ða bÞdb5da;

Example 2.2.2 Derive the expression yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ dðt t1 Þdðt t2 Þ.

for

2.2

Convolution

43

Solution: Using the integral expression, we have Z1

Z1

x1 ðaÞx2 ðt aÞda¼

1

x1e ðtÞ x20 ðtÞ ¼ y01 ðtÞ; x10 ðtÞ x2e ðtÞ ¼ y02 ðtÞ; odd functions:

dða t1 Þdðt t2 aÞda

(2:2:14c)

1

¼ dðt t2 aÞja¼t1 ¼ dðt t1 t2 Þ: 8. The area of a signal was defined in Chapter 1 (see (1.5.1)) by Z1 Noting dðtÞ dðtÞ ¼ dðtÞ and using (2.2.11), we have A½xi ðtÞ ¼ xi ðaÞda: (2:2:15) & dðt t1 Þ dðt t2 Þ ¼ dðt t1 t2 Þ. 1

7. The time scaling property of the convolution operation is if yðtÞ ¼ x1 ðtÞ x2 ðtÞ, then x1 ðctÞ x2 ðctÞ ¼

Z1

x1 ðcbÞx2 ðcðt bÞÞdb

Area property of the convolution applies if the areas of the individual functions do not change with a shift in time. It is given by A½yðtÞ ¼ A½x1 ðtÞ x2 ðtÞ ¼ A½x1 ðtÞA½x2 ðtÞ: (2:2:16)

1

1 ¼ yðctÞ; c 6¼ 0: jcj

(2:2:12)

This can be proved by A½yðtÞ

¼

Assuming c < 0 and using the change of variables a ¼ cb, and simplifying, we have

x1 ðctÞ x2 ðctÞ ¼

1 c

Z1 1

¼

1 j cj

¼

1 Z1 1

x1 ðaÞx2 ðct aÞdy

Z1

Z1

x1 ðaÞx2 ðct aÞda ¼

1

1 yðctÞ: jcj

¼

yðbÞdb ¼ 2

Z1

4

:

x1 ðtÞ x2 ðtÞ ¼ yðtÞ:

(2:2:13)

This property simplifies the convolution if there are symmetries in the functions. In Chapter 1, even and odd functions were identified by subscripts e for even and 0 for odd (see (1.2.7)). From these xie ðtÞ ¼ xie ðtÞ; an even function; xi0 ðtÞ ¼ xi0 ðtÞ; an odd function

(2:2:14a)

x1e ðtÞ x2e ðtÞ ¼ ye1 ðtÞ; x10 ðtÞ x20 ðtÞ ¼ ye2 ðtÞ; even functions (2:2:14b)

½x1 ðbÞ x2 ðbÞdb 3

x1 ðaÞx2 ðb aÞda5db

x1 ðaÞ

¼ A½x2 ðtÞ A similar argument can be given in the case of c > 0: Scaling property applies only when both functions are scaled by the same constant c 6¼ 0. When c ¼ 1, then

1

1

8 Z1 < 1

Z 1

Z1

Z1 1

9 = x2 ðb aÞdb da ;

x1 ðaÞda ¼ A½x2 ðtÞA½x1 ðtÞ:

1

9. Consider the signals x1 ðtÞ and x2 ðtÞ that are nonzero for the time intervals of tx1 and tx2 , respectively. That is, we have two time-limited signals, x1 ðtÞ and x2 ðtÞ, with time widths tx1 and tx2 . Then, the time width ty of the signal yðtÞ ¼ x1 ðtÞ x2 ðtÞ is the sum of the time widths of the two convolved signals and ty ¼ tx1 þ tx2 . This is referred to as the time duration property of the convolution. We will come back to some intricacies in this property, as there are some exceptions to this property. Example 2.2.3 Derive the expression for the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ, where xi ðtÞ; i ¼ 1; 2 are as follows: x1 ðtÞ ¼ 0:5dðt 1Þ þ 0:5dðt 2Þ; x2 ðtÞ ¼ 0:3dðt þ 1Þ þ 0:7dðt 3Þ:

44

2 Convolution and Correlation

Solution: Convolution of these two functions is Z1 x1 ðaÞx2 ðt aÞda yðtÞ ¼

¼

1 Z1

½0:5dða 1Þ þ 0:5dða 2Þ

1

½0:3dðt a þ 1Þ þ 0:7dðt a 3Þda Z1

¼

ð0:5Þð0:3Þdða 1Þdðt a þ 1Þda

1

þ

Z1

ð0:5Þð0:7Þdða 1Þdðt a 3Þda

1

þ

þ

Z1 1 Z1

ð0:5Þð0:3Þdða 2Þdðt a þ 1Þda

ð0:5Þð0:7Þdða 2Þdðt a 3Þda

1

¼ ð0:15ÞdðtÞ þ 0:35dðt 4Þ þ 0:15dðt 1Þ þ 0:35dðt 5Þ:

properties to simplify the evaluations are illustrated. A few comments are in order before the examples. First, the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ is an integral operation and can use either one of the integrals in (2.2.1). Note that yðtÞ; 1 < t < 1 is a time function. The expression for the convolution, say at t ¼ t0 , will yield a zero value for those values of t0 over which x1 ðbÞ and x2 ðt0 bÞ do not overlap. The area under the product ½x1 ðbÞx2 ðt0 bÞ, i.e., the integral of this product gives the value of the convolution at t ¼ t0 . Sketches of the function x1 ðbÞ and the time reversed and delayed function x2 ðt0 bÞ on the same figure would be helpful in identifying the limits of integration of the product ½x1 ðbÞx2 ðt0 bÞ. As a check, the value of the convolution at end points of each range must match, except in the case of impulses and/or their derivatives in the integrand of the convolution integral. This is referred to as the consistency check. The following steps can be used to compute the convolution of two functions x1 ðtÞ and x2 ðtÞ: New variable

Notes: If an impulse function is in the integrand of the form dðat bÞ, then use (see (1.4.35), which is

Convolution of two functions exists if the convolution integral exists. Existence can be given only in terms of sufficient conditions. These are related to signal energy, area, and one sidedness. It is simple to give examples, where the convolution does not exist. Some of these are a*a, a*u(t), cos(t)*u(t), eat *eat, a > 0. Convolution of energy signals and the samesided signals always exist. In Chapter 4 we will be discussing Fourier transforms and the transforms make it convenient to find the convolution.

Shift

New variable

Multiply the two functions

!x1 ðbÞ !x1 ðbÞx2 ðt bÞ: x1 ðtÞ R1 Integrate ! x1 ðbÞx2 ðt bÞdb ¼ yðtÞ:

dðat bÞ ¼ ð1=jajÞdðt ðb=aÞÞ:

2.2.2 Existence of the Convolution Integral

Reverse

x2 ðtÞ !x2 ðbÞ !x2 ðbÞ!x2 ðt bÞ

1

Example 2.3.1 Derive the expression for the convolution of the two pulse functions shown in Fig. 2.3.1 a,b. These are 1 t ða=2Þ and x1 ðtÞ ¼ P a a 1 t ðb=2Þ x2 ðtÞ ¼ P ; b a > 0: b b

(2:3:1)

Solution: First

yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼

Z1

x1 ðbÞx2 ðt bÞdb: (2:3:2)

1

2.3 Interesting Examples In the following, the basics of the convolution operation, along with using some of the above

Figure 2.3.1c,d,e,f give the functions x1 ðbÞ; x2 ðbÞ; x2 ð bÞ; and x2 ðt bÞ, respectively. Note that the variable t is some value between 1 and 1 on the b axis. Different cases are considered, and

2.3

Interesting Examples

45 x2(t)

x1(t)

(a) x1(β)

(b) x2(–β)

x2(β)

(c)

(d)

(e) t ≤ 0 : x2 (t – β) and x1 (β)

x2(t – β)

(g)

(f) t > 0 : x2 (t – β) and x1 (β)

t > a : x2 (t – β) and x1 (β)

(h)

(i) t – b > a : x2 (t – β) and x1 (β)

t – b < a : x2 (t – β) and x1 (β)

(k)

(j)

b = a : y (t) = x1 (t)∗x2 (t)

b > a : y (t) = x1 (t)∗x2 (t)

(l)

(m)

Fig. 2.3.1 Convolution of two rectangular pulses (b a)

in each case, we keep the first function x1 ðbÞ stationary and move (or shift) the second function x2 ðbÞ, resulting in x2 ðt bÞ. Case 1: t 0: For this case the two functions are sketched in Fig. 2.3.1 g on the same figure. Noting that there is no overlap of these two functions, it follows that

yðtÞ ¼ 0; t 0:

(2:3:3)

Case 2: 05t a: The two functions are sketched for this case in Fig. 2.3.1 h. The two functions overlap and the convolution is yðtÞ ¼

Zt 0

x1 ðbÞx2 ðt bÞdb ¼

1 t; 05t a: (2:3:4) ab

46

2 Convolution and Correlation

Case 3: a5t b: This corresponds to the complete overlap of the two functions and the functions are shown in Fig. 2.3.1i. The convolution integral and the area is

yðtÞ ¼

Za

1 1 db ¼ ; a5t b: ab b

(2:3:5)

integral operation, which is a smoothing operation. Convolution values at end points of each range must match (consistency check) as we do not have any impulse functions or their derivatives in the functions that are convolved. Some of these are discussed below. The areas of the two pulses are each equal to 1 and the area of the trapezoid is given by

0

Case 4: 0 < t b a < b or b < t ðb þ aÞ. This corresponds to a partial overlap of the two functions and is shown in Fig. 2.3.1j. The convolution integral and the area is

yðtÞ ¼

Za

1 ða þ b tÞ db ¼ ; b5t ðb þ aÞ: (2:3:6) ab ab

tb

Case 5: t b > a or t a þ b: The two functions corresponding to this range are sketched in Fig. 2.3.1k and from the sketches we see that the two functions do not overlap and yðtÞ ¼ 0; t ða þ bÞ:

(2:3:7)

Summary: 8 > > > > > > > > > > > >

b > > > > > aþbt > > ; > > ab > > : 0;

Area½yðtÞ ¼ ð1=2Það1=bÞ þ ðb aÞð1=bÞ þ ð1=2Það1=bÞ ¼ 1 ¼ Area½x1 ðtÞArea½x2 ðtÞ:

This shows that the area property is satisfied. Peebles (2001) shows the probability density function of the sum of the two independent random variables is also a probability density function. We should note that the probability density function is nonnegative and the area under this function is 1 (see Section 1.7). From the above discussion, it follows that the convolution of two rectangular pulses (these can be considered as uniform probability density functions) results in a nonnegative function and the area under this function is 1. The function yðtÞ satisfies the conditions of a probability density function. The time duration of yðtÞ, ty is ty ¼ tx1 þ tx2 and tx1 ¼ a; tx2 ¼ b ) ty ¼ tx1 þ tx2 ¼ a þ b:

t0

(2:3:9)

(2:3:10)

05 t 5 a a t 5b

:

(2:3:8)

b t 5a þ b taþb

This function is sketched in Fig. 2.3.1 l and m for the cases of b > a and b ¼ a. There are several interesting aspects in this example that should be noted. First, the two functions we started with have firstorder discontinuous and the convolution is an

A special case is when a ¼ b and the function yðtÞ given in Fig. 2.3.1m, a triangle, is P

ht ai t a=2 t a=2 P ¼L : (2:3:11) a a a

&

Example 2.3.2 Give the expressions for the convolution of the following functions:

t1 x1 ðtÞ ¼ uðtÞ and x2 ðtÞ ¼ sinðptÞP : (2:3:12) 2

2.3

Interesting Examples

47

Solution: The convolution integral

yðtÞ ¼

Z1

x2 ðbÞx1 ðt bÞda ¼

1

Z2

Solution:

sinðpbÞ uðt bÞdb

yðtÞ ¼ xðtÞ hðtÞ ¼

0

8 > Zt < 0; t 0 ¼ sinðpbÞdb ¼ ð1=pÞð1 cosðptÞÞ; 05t52 ; > : 0; t 2 0

yðtÞ ¼

Zt sinðpbÞdb 0

8 > < 0; t 0 ¼ ð1=pÞð1 cosðptÞÞ; 05t52 : > : 0; t 2

(2:3:13)

The time duration of the unit step function is 1 and the time duration of x2 ðtÞ is 2. The duration of the function yðtÞ is 2, which illustrates a pathological case where the time duration property of the convolution is not satisfied. The integral or the area of a sine or a cosine function over one period is equal to zero. The period of the function sinðptÞ is equal to 2 and therefore t1 A½x2 ðtÞ ¼ A sinðptÞ:P 2 Z2 ¼ 0 ) A½yðtÞ ¼ ð1=pÞ ½1 cosðptÞdt

¼

Z1 1 Z1

xðbÞhðt bÞdb

hðaÞxðt aÞda:

(2:3:14b)

1

In computing the convolution, we keep one of the functions at one location and the other function is time reversed and then shifted. In this example, since the function hðtÞ ¼ 0 for t < 0, we have a benchmark to keep track of the movement of the function hðt bÞ as t varies. Therefore, the first integral in (2.3.14b) is simpler to use. The functions xðbÞ; hðbÞ; hðbÞ; and hðt bÞ are shown in Fig. 2.3.2 c, d, e, and f respectively. As before, we will compute the convolution for different intervals of time. Case 1: t T : the two functions, hðt bÞ and xðbÞ, are sketched in Fig. 2.3.2 g. Clearly there is no overlap of the two functions and therefore the integral is zero. That is yðtÞ ¼ 0; t T:

(2:3:15)

Case 2: T < t < T: The two functions hðt bÞ and xðbÞ are sketched in Fig. 2.3.2 h in the same figure for this interval. There is a partial overlap of the two functions in the interval T4t4T. The convolution can be expressed by

0

¼ 1=p

Z2

dt ¼ 2=p:

yðtÞ ¼

0

Noting that A½x1 ðtÞ ¼ A½uðtÞ ¼ 1 and A½yðtÞ ¼ 2=p, we can see that the area property of the convolution is not satisfied. See Ambardar & (1995) for an additional discussion. Example 2.3.3 Derive the expression for the convolution of the following functions shown in Fig. 2.3.2a,b: h t i xðtÞ ¼ P and hðtÞ ¼ eat uðtÞ; a > 0: (2:3:14a) 2T

Z1

xðbÞhðt bÞdb ¼

1

¼ eat

Zt

ð1ÞeaðtbÞ db

T

Zt

(2:3:16) h i 1 eab db¼ 1 eaðtþTÞ ; T5t5T: a

T

Case 3: t > T : From the sketch of the two functions in Fig. 2.3.2 h, the two functions overlap in this range T t T and the convolution integral is yðtÞ ¼

ZT

eaðtbÞ db ¼

1 aT e eaT eat ; a

t4T:

T

(2:3:17)

48

2 Convolution and Correlation

Fig. 2.3.2 Convolution of a rectangular pulse with an exponentially decaying pulse

x (t )

h(t)

(a) x(β)

(b)

h(β)

(c)

(d) h ( t − β ), t ≥ 0

h(−β)

(e)

(f)

h ( t − β ) , t < –T and x ( β )

h ( t − β ) , t < T and x ( β )

(g)

(h) y(t)

(i)

Summary: 8 0; t T > > > i > > > 1 > : eaT eaT eat ; t4T a This function is sketched in Fig. 2.3.2i. Note yðtÞ is smoother than either of the given functions used in the convolution. Computing the area of yðtÞ is not as simple as finding the areas of the two functions, xðtÞ and hðtÞ: Using the area property,

A½yðtÞ ¼ A½xðtÞA½hðtÞ ¼ ð2TÞð1=aÞ: (2:3:19) & Notes: In computing the convolution, one of the sticky points is finding the integral of the product ½xðbÞhðt bÞ in (2.3.14b), which requires finding the region of overlap of the two functions. Sketching both functions on the same figure allows for an easy determination of this overlap. The delay property is quite useful. For example, if yðtÞ ¼ xðtÞ hðtÞ then it implies y1 ðtÞ ¼ xðt TÞ hðtÞ ¼ yðt TÞ. In Example 2.3.3, x(t) ¼ P[t/2T] ¼ u[tþT] u½t T. Therefore

2.3

Interesting Examples

49

Example 2.3.5 Derive the expression yi ðtÞ ¼ hðtÞ xi ðtÞ for the following two cases:

yðtÞ ¼ hðtÞ xðtÞ ¼ hðtÞ ½uðt þ TÞ xðt TÞ ¼ hðtÞ uðt þ TÞ hðtÞ uðt TÞ:

&

a:x1 ðtÞ ¼ uðtÞ; b:x2 ðtÞ ¼ dðtÞ:

Example 2.3.4 Determine the convolution yðtÞ ¼ xðtÞ xðtÞ with xðtÞ ¼ eat uðtÞ, a > 0:

Solution: a. Since uðt aÞ ¼ 0; a > t, we have the running integral

Solution: The convolution is y1 ðtÞ ¼ hðtÞ uðtÞ ¼

yðtÞ ¼ eat uðtÞ eat uðtÞ Z1 ¼ eab eaðtbÞ ½uðbÞuðt bÞdb

¼e

Zt

hðaÞuðt aÞda

1

¼

1 at

Z1

Zt (2:3:22)

hðaÞda: 1

db ¼ teat uðtÞ:

(2:3:20)

0

In evaluating the integral, the following expression is used (see Fig. 2.3.3a): ½uðbÞuðt bÞ ¼

b. Noting that the impulse function is the generalized derivative of the unit step function, we can compute the convolution y2 ðtÞ ¼ hðtÞdðtÞ ¼ hðtÞ

0; 1;

b50 and b4t : 05 b 4 t

(2:3:21)

The functions xðtÞ and yðtÞ are shown in Fig. 2.3.3b,c. Note that the function xðtÞ has a discontinuity at t ¼ 0: The function yðtÞ, obtained by convolving two identically decaying signals, xðtÞ and xðtÞ is smoother than either one of the convolved signals. This is to be expected as the convolution operation is a smoothing operation. &

duðtÞ ¼ y 0 1 ðtÞ ¼ hðtÞ: dt (2:3:23) &

Example 2.3.6 Let hðtÞ ¼ eat uðtÞ; a > 0 a. Determine the running integral of hðtÞ. b. Using (2.3.23), determine y2 ðtÞ: Solution:

a: y1 ðtÞ ¼

Zt

hðbÞdb ¼

1

Zt

eab uðbÞdb

1

1 ¼ ð1 eat ÞuðtÞ; a

(2:3:24)

dy1 ðtÞ 1 d ¼ ð1 eat ÞuðtÞ dt a dt 1 d 1 dð1 eat Þ ¼ ð1 eat Þ uðtÞ þ uðtÞ a dt a dt ¼ ð1=aÞð1 eat ÞdðtÞ þ eat uðtÞ

b: y2 ðtÞ ¼

(a) x(t ) = e

− at

u (t )

x(t ) = e

− at

u (t )

y (t ) = x(t ) * x(t )

¼ ð1=aÞ½dðtÞ dðtÞ þ eat uðtÞ ¼ ð1=aÞeat uðtÞ:

(b) Fig. 2.3.3 Example 2.3.4

(c)

(2:3:25) &

In a later chapter this result will be used in dealing with step and impulse inputs to an RC circuit with an impulse response hðtÞ ¼ eat uðtÞ.

50

2 Convolution and Correlation

Example 2.3.7 Express the following integral in the form of xðtÞ pðtÞ; ðpðtÞ is a pulse function:

b: y2 ðtÞ ¼

Z1

uðaÞuða þ tÞda

1 tþT=2 Z

yðtÞ ¼

xðaÞda:

(2:3:26) ¼

tT=2

8 R1 > > uðaÞda ! 1; t 0 > < t

> R1 > > : uða þ tÞda ! 1; t40

(2:3:30) :

0

Solution:

yðtÞ ¼

tþðT=2Þ Z

xðaÞda

1

It follows that y2 ðtÞ ¼ 1; 15t51. In this case, & convolution does not exist.

tðT=2Þ Z

xðaÞda 1

¼ xðtÞ uðt þ ðT=2ÞÞ xðtÞ uðt ðT=2ÞÞ ¼ xðtÞ ½uðt þ ðT=2ÞÞ uðt T=2ÞÞ hti ¼ xðtÞ P : T

2.4 Convolution and Moments (2:3:27)

The output is the convolution of xðtÞ with a pulse width of T with unit amplitude and the process is a & running average. Example 2.3.8 Find the derivative of the running average of the function in (2.3.27) and express the function xðtÞ in terms of the derivative of yðtÞ. Solution: McGillem and Cooper (1991) give an interesting solution for this problem. duðt þ ðT=2ÞÞ duðt ðT=2ÞÞ dt dt T T xðtÞ d t ¼ xðtÞ d t þ 2 2

y0 ðtÞ ¼ xðtÞ

In the examples considered so far, except in the cases of impulses, convolution is found to be a smoothing operation. We like to quantify and compare the results of the convolution of nonimpulse functions to the Gaussian function. In Section 1.7.1, the moments associated with probability density functions were considered. A useful result can be determined by considering the center of gravity convolution in terms of the centers of gravity of the factors in the convolution. First, the moments mn ðxÞ of a waveform xðtÞ and its center of gravity Z are, respectively, defined as

mn ðxÞ ¼

¼ xðt þ ðT=2ÞÞ xðt ðT=2ÞÞ

Example 2.3.9 Derive the expressions a: y1 ðtÞ ¼ uðtÞ uðtÞ; b: y2 ðtÞ ¼ uðtÞ uðtÞ: Solution:

a: y1 ðtÞ ¼ uðtÞ uðtÞ ¼

¼

0

( ð1Þdt ¼

(2:4:1)

R1 Z ¼ 1 R1

txðtÞdt ¼ xðtÞdt

m1 ðxÞ : m0 ðxÞ

(2:4:2)

1

We note that we can define a term like the variance in Section 1.7.1 by uðaÞuðt aÞda

1

Zt

tn xðtÞdt;

1

) xðtÞ ¼ y0ðt ðT=2ÞÞ þ xðt TÞ: (2:3:28) &

Z1

Z1

0; t40 t; t50

(2:3:29)

) ¼ tuðtÞ;

s2 ðxÞ ¼

m2 ðxÞ Z2 : m0 ðxÞ

(2:4:3)

Now consider the expressions for the convolution yðtÞ ¼ gðtÞ hðtÞ. First,

2.4

Convolution and Moments

m1 ðyÞ ¼

Z1

tyðtÞdt ¼

1

¼

Z1

Z1

2 gðlÞ4

1

51

2 4t

1

Z1

3

Z1 1

Signal-to-noise ratio ¼ gðlÞhðt lÞdl5dt

(2:4:8)

3

Example 2.4.1 Verify the result is true in (2.4.7) using the functions

thðt lÞdt5dl:

1

Defining a new variable x ¼ t l on the right and rewriting the above equation results in 2 m1 ðyÞ ¼ 4

Z1

Z1 gðlÞ

1

¼

ðx þ lÞhðxÞdx5dl

Solution: Using integral tables, it can be shown that Z1 Z1 et dt ¼ 1; m1 ðgÞ ¼ tet dt ¼ 1; m0 ðgÞ ¼ 0

Z1 lgðlÞdl

1

Z1

3

gðtÞ ¼ hðtÞ ¼ et and yðtÞ ¼ gðtÞ hðtÞ:

1

Z1

Average signal power : Noise power; s2n

Z1

hðxÞdx þ

1

xhðxÞdx

m2 ðgÞ ¼

Z1

t2 et dt ¼ 2;

0

1

gðlÞdl¼ m1 ðgÞm0 ðhÞ þ m1 ðhÞm0 ðgÞ:

0

(2:4:4)

1

m1 ðgÞ m2 ðgÞ ¼ 1; s2g ¼ Z2g ¼ 1; m0 ðgÞ m0 ðgÞ

Zg ¼

s2h ¼ 1 ðnote gðtÞ ¼ hðtÞÞ;

From the area property, it follows that m0 ðyÞ ¼ m0 ðgÞm0 ðhÞ. The center of gravity is m1 ðyÞ m1 ðgÞ m1 ðhÞ ¼ þ ) Zy ¼ Zg þ Zh : m0 ðyÞ m0 ðgÞ m0 ðhÞ

(2:4:5)

s2y ¼

m2 ðyÞ m1 ðyÞ m0 ðyÞ m0 ðyÞ

¼

s2g

þ

s2h :

t

te dt ¼1; m1 ðyÞ ¼

m2 ðyÞ ¼

0

t2 et dt ¼ 2;

0

Z1

Zy ¼

Z1

t3 et dt ¼ 6;

m1 ðyÞ m2 ðyÞ ¼ 2; s2y ¼ Z2y ¼ 2 ) m0 ðyÞ m0 ðyÞ

s2y ¼ s2g þ s2h ¼ 1 þ 1 ¼ 2:

2 :

(2:4:6)

Using the expressions for m0(y), m1 ðyÞ and m2 ðyÞ and simplifying the integrals results in s2y

m0 ðyÞ ¼

Z1 0

Consider the expression for the squares of the spread of yðtÞ in terms of the squares of the spreads of gðtÞ and hðtÞ. The derivation is rather long and only results are presented.

yðtÞ ¼ gðtÞ hðtÞ ¼ tet uðtÞðsee Example 2:3:4Þ:

(2:4:7)

That is, the variance of y is equal to the sum of the variances of the two factors. It also verifies that convolution is a broadening operation for pulses. Noting that if gðtÞ and hðtÞ are probability density functions then (2.4.7) is valid. In communications theory we are faced with a signal, say gðtÞ is corrupted by a noise nðtÞ with the variance, s2n . The signal-to-noise ratio (SNR) is given by

As an example, consider that we have signal gðtÞ ¼ A cosðo0 tÞ and is corrupted by a noise with a variance equal to s2n . Then, the signal-to-noise ratio is SNR ¼

A2 =2 : s2n

In Chapter 10, we will make use of this in quantization methods, wherein A and SNR are given and determine s2n . This, in turn, provides the information on the size of the error that can be tolerated. Notes: For readers interested in independent random variables, the probability density function of a sum of two independent random variables is the convolution of the density functions of the two factors of the

52

2 Convolution and Correlation

convolution, and the variance of the sum of the two random variables equals the sum of their variances. For a detailed discussion on this, see Peebles (2001).&

Solution: yðtÞ is a triangular function (see Example 2.3.1) given by 1 ht ai yðtÞ ¼ L : (2:4:12) a a

2.4.1 Repeated Convolution and the Central Limit Theorem

The mean values of the two rectangular pulses are a/2 (see Section 1.7). The mean value of yðtÞ is 2ða=2Þ ¼ a. The variance of each of the rectangular pulses is

Convolution operation is an integral operation, which is a smoothing operation. In Example 2.3.1, we have considered the special case of the convolution of two identical rectangular pulses and the convolution of these two pulses resulted in a triangular pulse (see Fig. 2.3.1m). The discontinuities in the functions being convolved are not there in the convolved signal. As more and more pulse functions convolve, the resultant functions become smoother and smoother. Repeated convolution begins to take on the bellshaped Gaussian function. The generalized version of this phenomenon is called the central limit theorem. It is commonly presented in terms of probability density functions. In simple terms, it states that if we convolve N functions and one function does not dominate the others, then the convolution of the N functions approaches a Gaussian function as N ! 1. In the general form of the central limit theorem, the means and variances of the individual functions that are convolved are related to the mean and the variance of the Gaussian function (see Peebles (2001)). Given xi ðtÞ; i ¼ 1; 2; :::; N; the convolution of these functions is yðtÞ ¼ x1 ðtÞ x2 ðtÞ ::: xN ðtÞ:

(2:4:9)

The function yðtÞ can be approximated using ðm0 ÞN , the sum of the individual means of the functions, and s2N the sum of the individual variances by 2 1 2 yðtÞ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eðtðm0 ÞN Þ =2sN : 2ps2N

(2:4:10)

Example 2.4.2 Illustrate the effects of convolution and compare yðtÞ to a Gaussian function by considering the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ; 1 t a=2 ; i ¼ 1; 2: xi ðtÞ ¼ P a a

s2i ¼ m2 m21 ¼ a2 =12; i ¼ 1; 2:

The variance is given by s2y ¼ s21 þ s22 ¼ a2 =6. The Gaussian approximation is 2 1 2 ðyðtÞÞjN¼2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eðtaÞ =ða =3Þ : pða2 =3Þ

(2:4:13b)

This Gaussian and the triangle functions are symmetric around a. They are sketched in Fig. 2.4.1. Even with N ¼ 2, we have a good approximation. &

Fig. 2.4.1 Triangle function yðtÞ in (2.4.12) and the Gaussian function in (2.4.13b)

Example 2.4.3 In Example 2.4.1 we considered two identically exponentially decaying functions: x1 ðtÞ ¼ et uðtÞ ¼ x2 ðtÞ. The convolution of these two functions is given by y2 ðtÞ ¼ tet uðtÞ. Approximate this function using the Gaussian function. Solution: The Gaussian function approximations of yn ðtÞ, considering n ¼ 2 and for n large, are, respectively, given below. Note that m0 ðyÞ ¼ 2. 2 1 y2 ðtÞ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eððt2Þ =2ð2ÞÞ ; 2pð2Þ 2 1 yn ðtÞ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eððtnÞ =2ðnÞÞ : 2pðnÞ

(2:4:11)

(2:4:13a)

(2:4:14)

For sketches of these functions for various values of & n, see Ambardar (1995).

2.4

Convolution and Moments

53

2.4.2 Deconvolution

xðtÞ ¼ yðtÞ hinv ðtÞ ¼ xðtÞ hðtÞ hinv ðtÞ

In this chapter, we have defined the convolution yðtÞ ¼ hðtÞ xðtÞ as a mathematical operation. If xðtÞ needs to be recovered from yðtÞ, we use a process called the deconvolution defined by

¼ xðtÞ ½hðtÞ hinv ðtÞ; ) hðtÞ hinv ðtÞ ¼ dðtÞ and xðtÞ dðtÞ ¼ xðtÞ: (2:4:15) It is a difficult problem to find hinv ðtÞ, which may not even exist.

Table 2.4.1 Properties of aperiodic convolution Definition: yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼

R1

1

x1 ðaÞx2 ðt aÞda ¼

R1

1

x2 ðaÞx1 ðt aÞda:

Amplitude scaling: ax1 ðtÞ bx2 ðtÞ ¼ abðxðtÞ hðtÞÞ; a and b are constants: Commutative: x1 ðtÞ x2 ðtÞ ¼ x2 ðtÞ x1 ðtÞ: Distributive: x1 ðtÞ ½x2 ðtÞ þ x3 ðtÞ ¼ x1 ðtÞ x2 ðtÞ þ x1 ðtÞ x3 ðtÞ: Associative: x1 ðtÞ ½x2 ðtÞ x3 ðtÞ ¼ ½x1 ðtÞ x2 ðtÞ x3 ðtÞ: Delay: x1 ðt t1 Þ x2 ðt t2 Þ ¼ x1 ðt t2 Þ x2 ðt t1 Þ ¼ yðt ðt1 þ t2 ÞÞ: Impulse response: xðtÞ dðtÞ ¼ xðtÞ: Derivatives: x1 ðtÞ x02 ðtÞ ¼ x01 ðtÞ x2 ðtÞ ¼ y0 ðtÞ;

ðmÞ

ðnÞ

x1 ðtÞ x2 ðtÞ ¼ yðmþnÞ ðtÞ:

Step response: yðtÞ ¼ xðtÞ uðtÞ ¼

Rt

1

xðaÞda;

y0 ðtÞ ¼ xðtÞ dðtÞ ¼ xðtÞ:

Area: A½x1 ðtÞ x2 ðtÞ ¼ A½yðtÞ; where A½xðtÞ ¼

R1

1

xðtÞdt:

Duration: tx 1 þ tx 2 ¼ ty : Symmetry: x1e ðtÞ x2e ðtÞ ¼ ye ðtÞ;

x1e ðtÞ x20 ðtÞ ¼ y0 ðtÞ;

Time scaling: x1 ðctÞ x2 ðctÞ ¼ j1cj yðctÞ; c 6¼ 0:

x10 ðtÞ x20 ðtÞ ¼ ye ðtÞ:

54

2 Convolution and Correlation δ T (t )

2.5 Convolution Involving Periodic and Aperiodic Functions 2.5.1 Convolution of a Periodic Function with an Aperiodic Function

h (t )

(a)

Let hðtÞ be an aperiodic function and xT ðtÞ be a periodic function with a period T. We desire to find the convolution of these two functions. That is, find yðtÞ ¼ xT ðtÞ hðtÞ.

yT ( t )

Example 2.5.1 Derive the expressions for the convolution of the following two functions: dT ðtÞ and hðtÞ assuming T ¼ 1:5 and T ¼ 2 and sketch the results for the two cases. 1 X

dT ðtÞ ¼

(b)

yT ( t )

dðt nTÞ;

hðtÞ ¼ L½t:

(c)

(2:5:1)

k¼1

Derive the expressions for the convolution of these two functions assuming T ¼ 1:5 and T ¼ 2 and sketch the results of the convolution for the two cases.

yðtÞ ¼ hðtÞ dT ðtÞ ¼ hðtÞ

1 X

dðt kTÞ

k¼1

¼

1 X

hðtÞ dðt kTÞ:

(2:5:2)

k¼1

(d) Fig. 2.5.1 (a) Periodic impulse sequence, (b) L½t; (c) yT ðtÞ; T ¼ 2, and (d) yT ðtÞ; T ¼ 2

R1 Solution: yðtÞ ¼ hðtÞ xT ðtÞ ¼ eab cosðo0 ðt bÞþ 0 yÞdb Z1 eab ½cosðo0 t þ yÞ cosðo0 bÞ ¼ 0

Noting that hðtÞ dðt kTÞ ¼ hðt kTÞ, it follows that

þ sinðo0 t þ yÞ sinðo0 bÞdb 2 1 3 Z ¼ 4 eab cosðo0 bÞdb5 cosðo0 t þ yÞ 0

yðtÞ ¼

1 X

2

hðt kTÞ ¼ yT ðtÞ:

(2:5:3)

k¼1

þ4

Z1

3 eab sinðo0 bÞdb5 sinðo0 t þ yÞ:

(2:5:5)

0

Figure 2.5.1a,b gives the sketches of the functions dT ðtÞ and hðtÞ. The sketches for the convolution are shown in Fig. 2.5.1c,d. In the first case, there were no overlaps, whereas in the second case there are & overlaps. Example 2.5.2 Derive an expression for the convolution yðtÞ ¼ hðtÞ xT ðtÞ, xT ðtÞ ¼ cosðo0 t þ yÞ and hðtÞ ¼ eat uðtÞ:

(2:5:4)

Using the identities given below (see (2.5.7 a, b, and c.)), (2.5.5) can be simplified. yðtÞ ¼ a=ða2 þ o20 Þ cosðo0 t þ yÞ þ o0 =ða2 þ o20 Þ sinðo0 t þ yÞ 1 ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ cosðo0 t þ y tan1 ðo0 =aÞÞ a2 þ o20 yT ðtÞ;

(2:5:6)

2.5

Convolution Involving Periodic and Aperiodic Functions

Z1 e 0

ab

The periodic convolution of two periodic functions, xT ðtÞ and hT ðtÞ, is defined by

eab sinðo0 bÞdb ¼ 2 ½a sinðo0 bÞ a þ o20 o0 cosðo0 bÞ1 0 ¼

o0 ; 2 a þ o20 (2:5:7a)

Z1

eab cosðo0 bÞdb ¼

0

eab ½a cosðo0 bÞ þ o20 a ; 2 a þ o20 (2:5:7b)

a cosðo0 t þ yÞ þ b sinðo0 t þ yÞ ¼ c cosðo0 t þ fÞ; qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ c ¼ a2 þ b2 ; f ¼ ½ tan1 ðb=aÞ þ y: (2:5:7c) The functions yðtÞ ¼ yT ðtÞ and xT ðtÞ are sinusoids at the same frequency o0 . The amplitude and the phase & of yðtÞ are different compared to that of xT ðtÞ. The derivation given above can be generalized for a periodic function 1 X xT ðtÞ ¼ Xs ½0 þ c½k cosðko0 t þ y½kÞ; o0 ¼ 2p=T; k¼0

(2:5:8a) yðtÞ ¼ xT ðtÞ hðtÞ ¼ Xs ½0 hðtÞ þ

1 X

c½k½cosðko0 t

k¼0

þ y½kÞ hðtÞ; o0 ¼ 2p=T:

tZ0 þT 1 xT ðaÞhT ðtaÞda yT ðtÞ ¼ xT ðtÞ hT ðtÞ ¼ T t0 Z Z 1 1 ¼ xT ðaÞhT ðtaÞda ¼ xT ðtaÞhT ðaÞda T T T

a2

þ o0 sinðo0 bÞ1 0 ¼

55

hT ðt þ T aÞ ¼ hT ðt aÞ and xT ðt þ T aÞ ¼ xT ðt aÞ:

(2:5:9b)

Also, periodic convolution is commutative. Many of the aperiodic convolution properties discussed earlier are applicable for periodic convolution with some modifications. The expression for the periodic convolution can be obtained by considering aperiodic convolution for one period of each of the two functions. Consider the periodic functions in the form 1 X

xT ðtÞ ¼

xðt nTÞ and hT ðtÞ ¼

n¼1

1 X

hðt nTÞ;

n¼1

(2:5:10a)

xT ðtÞ; t0 t < t0 þ T

0; otherwise hT ðtÞ; t0 t < t0 þ T

xðtÞ ¼ hðtÞ ¼

In Section 1.5 energy and power signals were considered. The energy in a periodic function is infinity and its average power is finite. One period of a periodic function has all its information. In the same vein, the average convolution is a useful measure of periodic convolution. Such averaging process is called periodic or cyclic convolution. The convolution of two periodic functions with different periods is very difficult and is limited here to the convolution of two periodic functions, each with the same period.

(2:5:9a)

Note that the symbol used for the periodic convolution and the constant ðTÞ in the denominator in (2.5.9a) indicates that it is an average periodic convolution. yT ðtÞ is periodic since

(2:5:8b)

2.5.2 Convolution of Two Periodic Functions

T

¼ hT ðtÞ xT ðtÞ:

;

0; otherwise:

(2:5:10b)

Note that the time-limited functions, xðtÞ and hðtÞ, are defined from the periodic functions xT ðtÞ and hT ðtÞ. Using (2.5.10b) the periodic convolution is 1 yT ðtÞ ¼ T

Z

xT ðaÞhT ðt aÞda

T

¼

1 T

Z T

xT ðaÞ

1 X n¼1

hðt a nTÞda

56

2 Convolution and Correlation

¼

1 Z 1 X xðaÞhðt a nTÞda T n¼1

Convolution of almost periodic or random signals, xðtÞ and hðtÞ, is defined by

T

1 1 X xðtÞ hðt nTÞ; ¼ T n¼1

(2:5:11a)

ZT=2

1 yðtÞ ¼ lim T!1 T

xðaÞhðt aÞda:

(2:5:14)

T=2

yT ðtÞ ¼ xT ðtÞ hT ðtÞ 1 1 X yðt nTÞ; yðtÞ ¼ xðtÞ hðtÞ: (2:5:11b) ¼ T n¼1

This reduces to the periodic convolution if xðtÞ and hðtÞ are periodic with the same period.

That is, yT ðtÞ can be determined by considering one period of each of the two functions and finding the aperiodic convolution.

2.6 Correlation

Example 2.5.3 a. Determine and sketch the aperiodic convolution yðtÞ ¼ hðtÞ xðtÞ. 1 t1 1 t 1:5 ; hðtÞ ¼ P : (2:5:12) xðtÞ ¼ P 2 2 3 3

Equation (2.1.3) gives the cross-correlation of xðtÞ and hðtÞ as the integral of the product of two functions, one displaced by the other by t between the interval a < t < b and is given by

Zb b. Determine and sketch the periodic convolution Rxh ðtÞ¼xðtÞhðtÞ¼ xðtÞhðtþtÞdt¼ hxðtÞhðtþti: yT ðtÞ ¼ xT ðtÞ hT ðtÞ for periods T ¼ 6 and 4: a 1 1 X X xðt kTÞ and hT ðtÞ ¼ hðt kTÞ: xT ðtÞ ¼ Cross-correlation function gives the similarity k¼1 k¼1 between the two functions: xðtÞ and hðt þ tÞ. Many (2:5:13) a times the second function hðtÞ may be a corrupted Solution: a. From (2.5.13), the results for the aper- version of xðtÞ, such as hðtÞ ¼ xðtÞ þ nðtÞ, where iodic convolution can be derived. The sketches of nðtÞ is a noise signal. In the case of xðtÞ ¼ hðtÞ, the two functions and the result of the convolution cross-correlation reduces to autocorrelation. In are shown in Fig. 2.5.2a. The periodic convolutions this case, at t ¼ 0, the autocorrelation integral for the two different periods are shown in Fig. gives the highest value at t ¼ 0. Comparison of 2.5.2b,c. There are no overlaps of the functions two functions appears in many identification situafrom one period to the next in Fig. 2.5.2b, whereas tions. For example, to identify an individual based & upon his speech pattern, we can store his speech in Fig. 2.5.2c, the pulses overlap.

(a)

yT (t ) Fig. 2.5.2 Example 2.5.1 (a) Aperiodic convolution; (b) periodic convolution T ¼ 6; (c) periodic convolution, T ¼ 4

(b)

yT (t )

(c)

2.6

Correlation

57

segment in a computer. When he enters, say a secure area, we can request him to speak and compute the cross-correlation between the stored and the recorded. Then decide on the individual’s identification based on the cross-correlation function. Generally, an individual is identified if the peak of the cross-correlation is close to the possible peak autocorrelation value. Allowance is necessary since the speech is a function of the individual’s physical and mental status of the day the test is made. Quantitative measures on the cross-correlation will be considered a bit later. The order of the subscripts on the cross-correlation function Rxh ðtÞ is important and will get to it shortly. In the case of xðtÞ ¼ hðtÞ, we have the autocorrelation and the function is referred to as Rx ðtÞ with a single subscript. The cross- and autocorrelation functions are functions of t and not t. Correlation is applicable to periodic, aperiodic, and random signals. In the case of periodic functions, we assume that both are periodic with the same period.

xðtÞhðt þ tÞdt

(2:6:1e)

For periodic functions, (2.6.1e) reduces to (2.6.1b).

2.6.1 Basic Properties of Cross-Correlation Functions Folding relationship between the two cross-correlation functions is Rxh ðtÞ ¼ Rhx ðtÞ; ) Rxh ðtÞ ¼

¼ Z1

xðtÞhðt þ tÞdt:

T=2

Cross-correlation: Aperiodic: Rxh ðtÞ ¼

ZT=2

1 Ra;xh ðtÞ ¼ lim T!1 T

Z1 1 Z1

(2:6:2)

xðtÞhðt þ tÞdt

xða tÞhðaÞda¼ Rhx ðtÞ: (2:6:3)

1

(2:6:1a)

1

2.6.2 Cross-Correlation and Convolution

Cross-correlation: Periodic: RT; xhðtÞ ¼

1 T

Z

xT ðtÞhT ðt þ tÞdt

(2:6:1b)

T

Autocorrelation: Aperiodic: Rx ðtÞ ¼

Z1

xðtÞxðt þ tÞdt

(2:6:1c)

1

The cross-correlation function is related to the convolution. From (2.6.3) we have Rxh ðtÞ ¼ xðtÞ hðtÞ ¼ xðtÞ hðtÞ;

(2:6:4a)

Rhx ðtÞ ¼ hðtÞ xðtÞ ¼ hðtÞ xðtÞ:

(2:6:4b)

Equation (2.6.4a) can be seen by first rewriting the first integral in (2.6.3) using a new variable t ¼ a, and then simplifying it. That is,

Autocorrelation: Periodic: 1 RT;x ðtÞ ¼ T

Z

xT ðtÞxT ðt þ tÞdt:

(2:6:1d)

T

Rxh ðtÞ ¼

Z1

xðtÞhðt þ tÞdt ¼

1

¼ xðtÞ hðtÞ: Notes: Cross- and autocorrelations of periodic functions and random signals are referred to as average periodic cross- and autocorrelation functions. In the case of random or noise signals, the average cross-correlation function is defined by

Z1

xðaÞhðt aÞda

1

(2:6:4c)

Equation (2.6.4b) can be similarly shown. Noting the explicit relation between correlation and convolution, many of the convolution properties are applicable to the correlation. To compute the cross–

58

2 Convolution and Correlation

correlation, Rxh ðtÞ, one can use either of the integral in (2.6.3) or the integral in (2.6.4c). Rxh ðtÞ is not always equal to Rhx ðtÞ. In case, if one of the functions is symmetric, say xðtÞ ¼ xðtÞ; then Rxh ðtÞ ¼ xðtÞ hðtÞ ¼ xðtÞ hðtÞ:

Rg ð0Þ ¼

Z1

g2 ðtÞdt ¼ Eg ; Rh ð0Þ ¼

1

1

Consider the integral Z1 Z1 2 ½xðtÞ hðt þ tÞ dt ¼ x2 ðtÞdt

þ

Z1

1 Z1

h2 ðt þ tÞdt 2

1

1

jRxh ðtÞj

(2:6:7a)

jRxh ðtÞj ðRx ð0Þ þ Rh ð0ÞÞ=2:

@

32 xðtÞhðt þ tÞdt5

1

Z1 1

10

jxðtÞj [email protected] 2

Z1 1

1 jhðt þ tÞj dtA; 2

1

(2:6:9a) pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Rx ð0ÞRh ð0Þ:

(2:6:9b)

(2:6:9c)

Write the resulting equation in a quadratic form in terms of a. In order for the equation in (2.6.10) to be true, the roots of the quadratic equation have to be real and equal or the roots have to be complex conjugates. The proof is left as a homework problem.

(2:6:8)

h t i ; hðtÞ ¼ eat uðtÞ; a > 0: 2T

(2:6:11a)

Solution: Example 2.3.3 dealt with computing the convolution of these two functions. The cross-correlation functions are as follows: Rhx ðtÞ ¼

Rxh ðtÞ ¼

0

h2 ðtÞdtA

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Rx ð0ÞRh ð0Þ ðRx ð0Þ þ Rh ð0ÞÞ=2:

(2:6:7b)

An interesting bound can be derived using the Schwarz’s inequality. See (2.1.9d).

hxðtÞhðt þ ti ¼ 4

1

Equation (2.6.9b) represents a tighter bound compared to the one in (2.6.7b), as the geometric mean cannot exceed the arithmetic mean. That is,

xðtÞhðt þ tÞdt

This follows since the integrand in (2.6.7a) is nonnegative and

Z1

x2 ðtÞ[email protected]

Z1

¼ Rx ð0ÞRh ð0Þ;

xðtÞ ¼ P

¼ Rx ð0Þ þ Rh ð0Þ 2Rxh ðtÞ 0:

2

10

Example 2.6.1 Determine the cross-correlation of the functions given in Fig. 2.3.2.

1

2

Z1

Another way to prove (2.6.9b) is as follows. Start with the inequality below. Expand the function and 2 h ðtÞdt¼ Eh : identify the auto- and cross-correlation terms. Z 1 ½xðtÞ þ ahðt þ tÞ2 dt 0: (2:6:10) (2:6:6) & 1

2.6.3 Bounds on the Cross-Correlation Functions

1

)jRxh ðtÞj2 @

(2:6:5)

Example 2.6.1 illustrates the use of this property. In particular, the area and duration properties for convolution also apply to the correlation. We should note that the correlations are functions of t and not t, where t is the time shift between xðtÞ and hðt þ tÞ. In the case of energy signals, the energies in the real signals, gðtÞ and hðtÞ, are Z1

0

Z1 1 Z1

hðtÞxðt þ tÞdt ¼ hðtÞ xðtÞ;

xðtÞhðt þ tÞdt ¼ xðtÞ hðtÞ:

1

(2:6:11b) Note that we have xðtÞ ¼ xðtÞ, and therefore the cross-correlation Rxh ðtÞ ¼ xðtÞ hðtÞ is the convolution determined before (see (2.3.18).), except the cross-correlation is a function of t rather than t. It is given below. The two cross-correlation functions are sketched in Fig. 2.6.1a,b. Note Rhx ðtÞ ¼ Rxh ðtÞ

2.6

Correlation

59

Fig. 2.6.1 Crosscorrelations (a)Rxh ðtÞ, (b) Rhx ðtÞðR xh ðTÞ ¼ 1 2aT ¼ Rhx ðTÞ) a 1e

8 0; t T > > > i > > > 1 > : eaT eaT eat ; t > T a (2:6:11c) &

2.6.4 Quantitative Measures of Cross-Correlation

The significance of rxh ðtÞ can be seen by considering some extreme cases. When xðtÞ ¼ ahðtÞ; a > 0, we have the correlation coefficient rxh ðtÞ ¼ 1. In the case of xðtÞ ¼ ahðtÞ; a < 0 and rxh ðtÞ ¼ 1. In communication theory, we will be interested in signals that are corrupted by noise, usually identified by nðtÞ, which can be defined only in statistical terms. In the following, we will consider the analysis without going through statistical analysis. Noise signal nðtÞ is assumed to have a zero average value. That is,

The amplitudes of Rxh ðtÞ ðand Rhx ðtÞÞ vary. It is appropriate to consider the normalized correlation coefficient (or correlation coefficient) of two energy signals defined by Rxh ðtÞ Rxh ðtÞ ﬃ; rxh ðtÞ ¼ sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 1 ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Ex Eh R R 2 2 x ðtÞdt h ðtÞdt 1

1

(2:6:12a) ) jrxh ðtÞj 1:

nðtÞdt ¼ 0:

(2:6:14)

T=2

Cross-correlation function can be used to compare two signals. The signals xðtÞ and hðtÞ are uncorrelated if the average cross-correlation satisfies the relation

(2:6:12b)

Equation (2.6.12b) can be shown as follows. From (2.1.13a) and using the Schwarz’s inequality (see (2.1.9d)), we have

ZT=2

1 lim T!1 T

1 Ra;xh ðtÞ ¼ lim T!1 T 2

ZT=2 T=2

1 6 ¼ 4 lim T!1 T

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Rxh ðtÞ ¼ hxðtÞhðt þ tÞi kxðtÞkkhðt þ tÞk ¼ Ex Eh

xðtÞhðt þ tÞdt

ZT=2 T=2

32 1 76 xðtÞdt54 lim T!1 T

ZT=2

3 7 hðtÞdt5:

T=2

(2:6:15) It should be noted that the case of xðtÞ ¼ hðtÞ, the correlation coefficient reduces to rxx ðtÞ ¼

Rx ðtÞ : Rx ð0Þ

Example 2.6.2 If the signals x(t) and a zero average noise signal n(t) are uncorrelated, then show

(2:6:13)

Correlation measures are very useful in statistical analysis. See Yates and Goodman (1999), Cooper and McGillem (1999) and others.

1 lim T!1 T

ZT=2 T=2

xðtÞnðt tÞdt ¼ 0 for all t:

(2:6:16)

60

2 Convolution and Correlation

Fig. 2.6.2 Correlation detector

Solution: Using (2.6.14) and (2.6.15), we have lim

1

ZT=2

T!1 T

2

xðtÞnðt þ tÞdt

T=2

1 6 ¼ 4 lim T!1 T

ZT=2

32 1 76 xðtÞdt54 lim T!1 T

T=2

ZT=2

3

correlation detector (or receiver) shown in Fig. 2.6.2. The received signals yi ðtÞ are assumed to be of the form in (2.6.19). Decide which signal has been transmitted using the cross-correlation function.

7 nðt þ tÞdt5 ¼ 0:

T=2

(2:6:17) Cross-correlation function can be used to estimate the delay caused by a system. Suppose we know that a finite duration signal xðtÞ is passed through an ideal transmission line resulting in the output function yðtÞ ¼ xðt t0 Þ. The delay t0 caused by the transmission line is unknown and can be estimated using the cross-correlation function Rxy ðtÞ: At t ¼ t0 , Rxy ðt0 Þ gives a maximum value. Then determine t corresponding to the maximum value of & Rxy ðtÞ.

yi ðtÞ ¼ xi ðtÞ þ noise; i ¼ 1 or 2:

Solution: Let the transmitted signal be x1 ðtÞ. Using the top path in Fig. 2.6.2, we have ZTs

½x1 ðtÞ þ nðtÞx1 ðtÞdt ¼ A

0

ZTs

xi ðtÞxj ðtÞdt ¼

Exi ¼ Ex ; i ¼ j 0;

i 6¼ j; i ¼ 1; 2

:

(2:6:18)

0

Ex is the energy contained in each signal. The two signals to be transmitted are assumed to be available at the receiver. A simple receiver is the binary

ZTs

x21 ðtÞdt:

(2:6:20)

0

Using the bottom path, with the transmitted signal equal to x1 ðtÞ, we have ZTs

½x1 ðtÞþnðtÞx2 ðtÞdt¼

0

Example 2.6.3 Consider the transmitted signals x1 ðtÞ and x2 ðtÞ in the interval 0 < t < Ts and zero otherwise. Use the cross-correlation function to determine which signal was transmitted out of the two. They are assumed to be mutually orthogonal (see Section 2.1.1) over the interval and satisfy

(2:6:19)

ZTs

½x1 ðtÞx2 ðtÞþx1 ðtÞnðtÞdt

0

¼

ZTs

x1 ðtÞnðtÞdt¼B:

(2:6:21)

0

Since the noise signal has no relation to x1 ðtÞ, B will be near zero and A 44B, implying x1 ðtÞ was transmitted. If x2 ðtÞ was transmitted, the roles are reversed and B A. The correlation method of detection is based on the following: 1. If A > B ) transmitted signal is x1 ðtÞ. 2. If B > A ) transmitted signal is x2 ðtÞ. 3. If B ¼ A ) no decision can be made as noise & swamped the transmitted signal.

2.6

Correlation

61

Example 2.6.4 Derive the expressions for the crossand Rhx ðtÞ assuming correlation Rxh ðtÞ xðtÞ ¼ et uðtÞ; hðtÞ ¼ e2t uðtÞ:

xðtÞ ¼ P½t :5; hðtÞ ¼ tP ¼

Solution: Using the expression in (2.6.3), we have Rxh ðtÞ ¼

¼

Z1 1 Z1

xða tÞhðaÞda

eðatÞ e2a ½uða tÞuðaÞda: (2:6:22a)

Consider the following and then the corresponding correlations: ½uðaÞuða tÞ ¼ t 1 or t 1:

;

Rxh ðtÞ ¼

1 3a 1 1 2t e uðtÞ; (2:6:22b) t ¼ e 3 3 Z1

(2:6:24)

Case 2: 0 < t 1 or 1 < t 0 : See Fig. 2.6.4e. Using Table 2.6.1

e3a da

t

¼ et

Solution: See Fig. 2.6.4c for hðt þ tÞ for an arbitrary t. The function hðt þ tÞ starts at t ¼ t and ends at t ¼ 2 t. As t varies from 1 to 1, there are five possible regions we need to consider. These are sketched in Fig. 2.6.4 d,e,f,g,h. In each of these cases both the functions are sketched in the same figure, which allows us to find the regions of overlap. The regions of overlap are listed in Table 2.6.1.

Z1

ðt þ tÞdt ¼

t2 þ ttt¼1 t¼t 2

t

¼

ðt þ 1Þ2 ; 1 t < 0: 2

(2:6:25)

1 t 3a 1 et : ee 0 ¼ 3 Case 3: 0 < t 1: See Fig. 2.6.4 f. Using Table ð3Þ 0 2.6.1, we have (2:6:22c) Z1 t2 1 þ 2t Rxh ðtÞ is shown in Fig. 2.6.3. Note that Rxh ðtÞ ¼ ðt þ tÞdt ¼ þ ttt¼1 ; 0 < t 1: t¼0 ¼ 2 2 & Rhx ðtÞ ¼ Rxh ðtÞ. 0 t < 0 : Rxh ðtÞ ¼ et

e3a da ¼

(2:6:26)

Rxh (τ )

Case 4: 1 < t 2: See Fig. 2.6.3 g. Using Table 2.6.1 we have

Rxh ðtÞ ¼

Z2t

ðt þ tÞdt ¼

t2 þ ttt¼2t t¼0 2

0

Fig. 2.6.3 Rxh ðtÞ

Example 2.6.5 Derive the cross–correlation Rxh ðtÞ for the following functions:

4 t2 ; 1 < t 2: ¼ 2

(2:6:27)

Case 5: 2 < t: See Fig. 2.6.4 h. There is no overlap and

62 Fig. 2.6.4 (a) xðtÞ, (b) hðtÞ; (c) hðt þ tÞ, (d) xðtÞ and hðt þ tÞ; t > 1ðor t 1Þ; (e) xðtÞandhðtþtÞ;1 2, (i) Rxh ðtÞ

2 Convolution and Correlation

(a)

(b)

(c)

(d)

(e)

(f)

(h)

(g)

Rxh (τ )

τ (i)

Table 2.6.1 Example 2.6.4 Case t Range of overlap/ integration range 1 t 1 No over lap 2 1 < t 0 t < t < 1 3 0 0 : Rx ðtÞ ¼

¼

½xðtÞ xðt tÞ½xðtÞ xðt tÞdt 0: (2:7:4)

Z1 1 Z1

1;

t>t

0;

otherwise

; (2:7:9)

xðtÞxðt tÞdt

eat uðtÞeaðttÞ uðt tÞdt

1

2

x ðtÞdt þ

1

Z1

2

x ðt tÞdt 2

1

¼ 2½

Z1

Z1

xðtÞxðt tÞdt

Z1

x ðtÞdt

1

Z1

¼

Rx ðtÞ ¼ ð1=2aÞeajtj :

1

xðtÞxðt tÞdt:

(2:7:10)

The energy contained in the exponentially decaying pulse is E ¼ Rx ð0Þ ¼ ð1=2aÞ. The autocorrelation & function is sketched in Fig. 2.7.1. (2:7:5)

Third,

1

eat : 2a

Using the symmetry property of the AC, we have xðtÞxðt tÞdt 0:

x2 ðtÞdt jRx ðtÞj

1 1 Z

Z1

e2 at dt ¼

t

1

) Rx ð0Þ ¼

Z1

1

2

Ex ¼ Rx ð0Þ ¼

¼e

at

x2 ðtÞdt:ðenergy in xðtÞÞ: (2:7:6)

Example 2.7.2 Consider the function xðtÞ ¼ P½t 1=2. Determine its autocorrelation function and its energy using this function. Solution: The AC function for t 0 is Z1 Z1 Rx ðtÞ¼ xðtÞxðttÞdt¼ P½t:5P½tt:5dt: 1

1

64

2 Convolution and Correlation

AC function is much easier to compute using this property. The AC of the pulse function P½t :5 can be computed by ignoring the delay. That is, ACfP½t :5g ¼ ACfP½tg. Interestingly, n h t io hti AC P ¼ TL : T T Fig. 2.7.1 Example 2.7.1

The function P½t 1=2 is a rectangular pulse centered at t ¼ 1=2 with a width of 1, and P½t ðt þ ð1=2ÞÞ is a rectangular pulse centered at ðt þ 0:5Þ with a width of 1. See Fig. (2.7.2a) for the case 0 < t < 1. In the case of t 1, there is no overlap indicating that Rx ðtÞ ¼ 0; t 1.

Rx ðtÞ ¼

Z1

The AC function of a rectangular pulse of width T is a triangular pulse of width 2T and its amplitude at t ¼ 0 is T: We can verify the last part by noting n h t io hti ¼ T: AC P jt¼0 ¼ TL j T T t¼0 Note xðtÞP Tt extracts xðtÞ for the time T=2 < t < T=2. That is,

dt ¼ ð1 tÞ; 0 t < 1:

t

xðtÞP Using the symmetry property, we have Rx ðtÞ ¼ Rx ðtÞ ¼

(2:7:11b)

ð1 jtjÞ; 0 jtj 1 0; Otherwise

¼ L½t: (2:7:11a)

hti T

¼

xðtÞ; T=2 < t < T=2 : 0; otherwise

(2:7:12)

Example 2.7.3 Find the autocorrelation of the function yðtÞ ¼ cosðo0 tÞP½t=T. Solution:

This is sketched in Fig. 2.7.2b indicating that there is correlation for jtj < 1 and no correlation for jtj 1. The peak value of the autocorrelation is when t ¼ 0 and is R x ð0Þ ¼ 1. The energy contained in the unit rectangular pulse is equal to 1 and by using the autocorrelation function, i.e., Rx ð0Þ ¼ 1, the same by both the methods. Noting that the autocorrelation function of a given function and its delayed or advanced version are the same, the

Ry ðtÞ¼

Z1 P

h t i htti P cosðo0 tÞcosðo0 ðttÞÞdt T T

1

Z1

cosðo0 tÞ ¼ 2

P

h t i htti P dt T T

1

þ

1 2

Z1 P

h t i htti P cosð2o0 ttÞdt: T T

1

Rx(τ)

(a) Fig. 2.7.2 Example 2.7.2 Autocorrelation of a rectangular pulse

(b)

2.8

Cross- and Autocorrelation of Periodic Functions

¼

ð1=2ÞTL Tt cosðo0 tÞ þ B; jtj T : 0; jtj > T

2.8 Cross- and Autocorrelation (2:7:13) of Periodic Functions

Now consider the evaluation of B. For t 0, Z1

1 B¼ 2

P

65

h t i ht ti P cosð2o0 t tÞdt T T

The cross- and the autocorrelation functions of periodic functions of xT ðtÞ and hT ðtÞ are Z 1 xT ðtÞhT ðt þ tÞdt ¼ hxT ðtÞhT ðt þ tÞi ; RT; xhðtÞ ¼ T T

1

(2:8:1a) ¼

1 2

ZT=2

cosð2o0 t o0 tÞdt

1 RT;x ðtÞ ¼ T

t

¼

1 ½sinðo0 T o0 tÞ sinðo0 tÞ: 4o0

Z

xT ðtÞxT ðt þ tÞdt ¼ hxT ðtÞxT ðt þ tÞi :

T

(2:7:14)

(2:8:1b)

If o0 is large, Ry ðtÞ in (2.7.13) can be approximated by the first term and

Note that the periods of the functions, xT ðtÞ and hT ðtÞ, are assumed to be the same and the constant (1/T) before the integrals in (2.8.1a and b). If they have different periods, computation of (2.8.1a) is difficult and these cases will not be discussed here. Many of the cross-correlation and AC function properties derived earlier for the aperiodic case apply for the periodic functions with some modifications. Note that

hti 1 RY ðtÞ ’ TL cosðo0 tÞ: 2 T

(2:7:15)

The envelope of the autocorrelation function in (2.7.16) is a triangular function, which follows since the correlation of the two identical rectangular functions is a triangular function. Noting that the cosine function oscillates between 1, the envelope of the autocorrelation function in (2.7.15) is shown & in Fig. 2.7.3. Ry(τ)

Fig. 2.7.3 Sketch of Ry(t)

RT;xh ðtÞ ¼ RT;hx ðtÞ;

RT;x ð0Þ þ RT;h ð0Þ 2RT;xh ðtÞ:

In Section 2.5.1, aperiodic convolution was used to find periodic convolution. The same type of analysis can be used to determine periodic cross-correlations using aperiodic cross-correlations. Furthermore, as discussed before, correlation is related to convolution. First define two finite duration functions, xðtÞ and hðtÞ, over the interval t0 t < t0 þ T. Assume that they are zero outside this interval. Now create two periodic functions:

xT ðtÞ ¼

1 X n¼1

Notes: Conditions for the existence of an aperiodic autocorrelation are similar to those of convolution (see Section 2.2.3). But there are a few exceptions. For example, the autocorrelation of the unit step function does not exist.

(2:8:2)

xðt nTÞ;

hT ðtÞ ¼

1 X

hðt nTÞ:

n¼1

(2:8:3a) The periodic cross-correlation function is defined by

66

2 Convolution and Correlation

1 RT;xh ðtÞ ¼ T ¼

tZ0 þT

t0 1 X

Solution: 1 a: RT;xT;1 ðtÞ ¼ T

xT ðtÞhT ðt þ tÞdt; hT ðt þ tÞ

Z

X2s ½0dt ¼ X2s ½0;

(2:8:5a)

T

hðt þ t nTÞ:

(2:8:3b)

n¼1

The expression for periodic convolution is given in terms of aperiodic convolution and

1 RT;xT;2 ðtÞ ¼ T

Z T

¼

c2 ½k 2T

xT;2 ðtÞxT;2 ðt þ tÞdt Z cosðko0 tÞdt T

1 1 X RT;xh ðtÞ ¼ Rxh ðt nTÞ; T n¼1

Rxh ðt nTÞ ¼

tZ0 þT

xðtÞhðt þ t nTÞ:

c2 ½k þ 2T

Z

cosðko0 ð2t þ tÞ þ 2y½kÞdt

T

(2:8:3c)

t0

¼

c2 ½k cosðko0 tÞ 2T

ZT

dt ¼

c2 ½k cosðko0 tÞ : 2

0

(2:8:5b) The details of the derivation are left as an exercise. Copies of Rxh ðtÞ will overlap if the width of Rxh ðtÞ is wider than T. Example 2.8.1 Give the lower bound on the period T so that there are no overlaps in the cross-correlation of the functions xT ðtÞ and hT ðtÞ given below. See Example 2.6.5.

t1 ; xðtÞ ¼ P½t :5; hðtÞ ¼ tP 2 1 1 X X xðt þ nTÞ; hT ðtÞ ¼ hðt þ nTÞ: xT ðtÞ ¼ n¼1

n¼1

Note that the integral of a cosine function over any integer number of periods is zero. b. The cross-correlation of a constant and a cosine function over one period is zero. Also note that the two functions are orthogonal. That is & hxT1 ðtÞ; xT2 ðtÞi ¼ 0. Example 2.8.3 Find the AC of xT ðtÞ given below with k 6¼ m; kandmare integers. xT ðtÞ ¼ xT;1 ðtÞ þ xT;2 ðtÞ; xT;1 ðtÞ ¼ c½k cosðko0 t þ y½kÞ; xT;2 ðtÞ ¼ c½m cosðmo0 t þ y½mÞ:

Solution: If the period T is larger than 3, then there are no overlaps in the periodic cross-correlation function. In that case, one period of the cross-correlation function can be obtained from the aperiodic cross-correlation in that example and dividing it by the period T. If the period is less than 3, then & there will be overlaps. Example 2.8.2 Consider the periodic functions xT;1 ðtÞ ¼ Xs ½0;

xT;2 ðtÞ ¼ c½k cosðko0 t þ y½kÞ: (2:8:4)

a. Find the AC functions for the functions in (2.8.4). b. Find the cross-correlation of the two functions.

Solution: The periodic autocorrelations are determined as follows: Z 1 ½xT;1 ðtÞ þ xT;2 ðtÞ½xT;1 ðt þ tÞ RT;x ðtÞ ¼ T T Z 1 þ xT;2 ðt þ tÞdt ¼ xT;1 ðtÞxT;1 ðt þ tÞdt T T Z 1 þ xT;2 ðtÞxT;2 ðt þ tÞdt T T Z 1 þ xT;1 ðtÞxT;2 ðt þ tÞdt T T Z 1 þ xT;2 ðtÞxT;1 ðt þ tÞdt: (2:8:6) T T

2.8

Cross- and Autocorrelation of Periodic Functions

67

Note Z

1 T

xT;1 ðtÞxT;2 ðt þ tÞdt

T

¼

P ¼ X2s ½0 þ

1 2T

Z

1 þy½mÞdt þ 2T

Z

Variance ¼ h½kh½m cos½ðk mÞo0 t mo0 tÞ

T

þðy½k y½mÞdt ¼ 0: Similarly the fourth term in (2.8.6) goes to zero. From the last example, RT;x ðtÞ ¼

(2:8:10)

The difference between the total power and the dc power is the variance and is given by

c½kc½m cos½ðk þ mÞo0 t þ mo0 t þ y½k

T

1 1X c2 ½k: 2 k¼1

1 1X c2 ½k: 2 k¼1

(2:8:11) &

Example 2.8.4 Consider the corrupted signal yðtÞ ¼ xðtÞ þ nðtÞ, where nðtÞ is assumed to be noise. Assuming the signal xðtÞ and noise nðtÞ are uncorrelated, derive an expression for the autocorrelation function of yðtÞ.

Solution: c2 ½k c2 ½m cosðko0 tÞ þ cosðmo0 tÞ; k 6¼ m: 2 2 ZT=2 (2:8:7) & 1 Ryy ðtÞ ¼ lim yðtÞyðt þ tÞdt T!1 T T=2

These results can be generalized using the last two examples and the autocorrelation of a periodic function xT ðtÞ is given as follows: xT ðtÞ ¼ Xs ½0 þ

1 X

c½k cosðko0 t þ y½kÞ;

(2:8:8)

1

¼ lim

T!1 T

½xðtÞ þ nðtÞ½xðt þ tÞnðt þ tÞdt;

T=2

ZT=2

¼ lim

T!1

k¼1

) RT;x ðtÞ ¼ X2s ½0 þ

ZT=2

xðtÞxðt þ tÞdt þ lim

T!1

T=2 1 1X c2 ½kcosðko0 tÞ;o0 ¼ 2p=T: 2 k¼1

þ lim

T!1

T=2 R T=2

Notes: The AC function of a constant Xs ½0 is X2s ½0. The AC of the sinusoid c½k cosðko0 t þ y½kÞ is ðc2 ½k=2Þ cosðko0 tÞ. That is, it loses the phase information in the function in the sinusoid. The power contained in the periodic function xT ðtÞ in (2.8.8) can be computed from the autocorrelation function evaluated at t ¼ 0. That is,

xðtÞnðt þ tÞdt

T=2

nðtÞxðt þ tÞdt þ lim

T!1

T=2 R

xðtÞxðt þ tÞdt:

T=2

(2:8:12)

(2:8:9) AC function of a periodic function is also a periodic function with the same period. It is independent of y½k. It does not have the phase information contained in (2.8.8). In the next chapter, (2.8.8) will be derived for an arbitrary periodic function and will be referred to as the harmonic form of Fourier series of a periodic & function xT ðtÞ.

ZT=2

Noting that the signal and the noise are uncorrelated, i.e., Rxn ðtÞ ¼ Rnx ðtÞ ¼ 0, we have Ryy ðtÞ ¼ Rxx ðtÞ þ Rnn ðtÞ:

(2:8:13)

&

The average power contained in the signal and the noise is given by Py ¼ Px þ Pn ¼ Rx ð0Þ þ Rn ð0Þ ¼ Rx ð0Þ þ s2n : (2:8:14) The signal-to-noise ratio (SNR), Px =Pn , can be computed. It is normally identified in terms of decibels. See Section 1.9.

68

2 Convolution and Correlation

2.2.3 Use the area property of convolution to find the integrals of yðtÞ in Problem 2.2.2.

2.9 Summary We have introduced the basics associated with the two important signal analysis concepts: convolution and correlation. Specific principal topics that were included are

Convolution integral: its computations and its properties Moments associated with functions Central limit theorem Periodic convolutions Auto- and cross-correlations Examples of correlations involving noise without going into probability theory

Quantitative measures of cross-correlation functions and the correlation coefficient

Auto- and cross-correlation functions of energy and periodic signals

Signal-to-noise ratios

2.3.1 a. Derive the expression for the convolution of two pulse functions given by xðtÞ ¼ P½t 1 and h½t ¼ P½t 2. Compute this directly first and then verify your result by using the delay property of convolution. b. Verify the time duration property of the convolution using the above problems. 2.3.2 Determine the area of yðtÞ in (2.3.18) using the area property of the convolution. 2.4.1 Approximate the function yðtÞ in Example 2.3.1 using the Gaussian function. 2.4.2 Use the derivative property of the convolution to derive the convolution of the two functions given below using the results in Example 2.5.2. xT ðtÞ ¼ sinðo0 tÞ; hðtÞ ¼ eat uðtÞ; a > 0: 2.4.3 Use the delay property of the convolution to determine

Problems

xðtÞ ¼ eat uðtÞ uðt 1Þ: 2.1.1 Consider the following functions defined over 0 < t < 1. Using (2.1.3), identify the two functions that give the maximum cross-correlation at t ¼ 0. x1 ðtÞ ¼ et ; x2 ðtÞ ¼ sinðtÞ; x3 ðtÞ ¼ ð1=tÞ: 2.2.1 Prove the commutative, distributive, and the associate properties of the convolution. 2.2.2 Find the convolution yðtÞ ¼ hðtÞ xðtÞ for the following functions:

2.5.1 Derive the expressions for the periodic convolution of the two periodic functions 1 1 X X t nT xðtÞ ¼ dðt nTÞ; hðtÞ ¼ P : T=2 n¼1 n¼1 2.6.1 Find the cross-correlation of the functions xðtÞ and hðtÞ given in (2.6.11a) by directly deriving the result and verify the result using the results in Example 2.6.1.

a: xðtÞ ¼ :5dðt 1Þ þ :5dðt 2Þ; hðtÞ ¼ :5dðt 2Þ þ :5dðt 3Þ

2.6.2 Show the bounds given in (2.6.7a and b) and (2.6.9b) are valid. Use (2.6.11a).

b: xðtÞ ¼ ðt 1ÞP½t 1;

2.6.3 Show (2.6.9b) using (2.6.10).

hðtÞ ¼ xðtÞ;

c: xðtÞ ¼ ð1 t2 Þ; 1 t 1; hðtÞ ¼ P½t; d: xðtÞ ¼ eat uðtÞ;

hðtÞ ¼ ebt uðtÞ

for cases : 1:a > 0; b > 0; e: xðtÞ ¼ P½t=2; f: xðtÞ ¼ dðt 1Þ;

2.7.1 Find the autocorrelations of the following functions:

2:a ¼ 0; b > 0

hðtÞ ¼ P½t :5 P½t 1:5 hðtÞ ¼ et uðtÞ

g: xðtÞ ¼ cosðptÞP½t;

hðtÞ ¼ et uðtÞ:

a:x1 ðtÞ ¼ P½t :5 P½t 1:5; b:x2 ðtÞ ¼ uðt :5Þ uðt þ :5Þ;

c:x3 ðtÞ ¼ tP½t:

Compute the energies contained in the functions directly and then verify the results using the autocorrelation functions derived in the first part. 2.7.2 Verify the result in (2.7.3) using the results in Example 2.7.1.

Problems

69

2.7.3 Show the identity AC½xðt t0 Þ ¼ AC½xðtÞ: 2.7.4 Derive the AC function step by step for the function xðtÞ ¼ cosðo0 tÞP½t=T. Use the integral formula by assuming o0 ¼ p and T ¼ 4. Verify the results in Example 2.7.3 using the information provided in this problem. Give the appropriate bounds. 2.7.5 Show that the autocorrelations of the function x2 ðtÞ ¼ eat uðtÞ for a 0 do not exist. 2.8.1 a. Derive the time-average periodic autocorrelation function Rx; T ðtÞ for the following periodic function using the integral formula. xT ðtÞ ¼ A1 cosðo0 t þ y1 Þ þ A2 cosð2o0 t þ y2 Þ: b. Verify the result using (2.8.8) and (2.8.9). c. Compute the average power contained in the function directly and by evaluating the autocorrelation function at t ¼ 0. Sketch the function xðtÞ by assuming the values A1 ¼ 5; A2 ¼ 2; y1 ¼ 200 ; y2 ¼ 1200 . Sketch the autocorrelation function using

these constants. Suppose we are interested in determining the period T from these two sketches, which function is better, the given function or its autocorrelation? Why? 2.8.2 Let yT ðtÞ ¼ A þ xT ðtÞ; A constant. Repeat the last problem, except for the plots. 2.8.3 a. Show that the following functions are orthogonal over a period: xT ðtÞ ¼ cosðo0 t þ yÞ; yðtÞ ¼ A b. Show the functions xðtÞ ¼ P½t; y½t ¼ t are orthogonal. 2.8.4 Consider the signal zðtÞ ¼ xðtÞ þ yðtÞ. Show that the AC of this function is given by Rz ðtÞ ¼ Rx ðtÞ þ Ry ðtÞ þ Rxy ðtÞ þ Ryx ðtÞ: Simplify the expression for Rz ðtÞ by assuming that xðtÞ is orthogonal to yðtÞ for all t. 2.8.5 Complete the details in deriving the periodic cross-correlation function in terms of the aperiodic convolution leading up to Equation (2.8.3c). 2.8.6 Show (2.8.3c) using (2.6.5).

Chapter 3

Fourier Series

3.1 Introduction In this chapter we will consider approximating a function by a linear combination of basis functions, which are simple functions that can be generated in a laboratory. Joseph Fourier (1768–1830) developed the mathematical theory of heat conduction using a set of trigonometric (sine and cosine) series of the form we now call Fourier series (Fourier, J.B.J., 1955 (A. Freeman, translation)). He established that an arbitrary mathematical function can be represented by its Fourier series. This idea was new and startling and met with vigorous opposition from some of the leading mathematicians at the time, see Hawking (2005). Fourier series and the Fourier transform are basics to mathematics and science, especially to the theory of communications. For example, a phoneme in a speech signal is smooth and wavy. A linear combination of a few sinusoidal functions would approximate a segment of speech within some error tolerance. Suppose we like to build a structure that allows us to climb from the first floor to the second floor of a building. We can have a staircase approximating a ramp function using a linear combination of pulse functions. The amplitudes and the width of the pulses can be determined based on the error between the ramp and the staircase. Apart from the staircase problem, this type of analysis is important in electrical engineering, for example, when converting an analog signal to a discrete signal. The term ‘‘well-behaved’’ function, xðtÞ defined in the interval, ðt0 ; t0 þ TÞ is given in terms of the following Dirichlet conditions: 1. The function xðtÞ must be single valued within the given interval of T seconds.

2. The function xðtÞ can have at most a finite number of discontinuities and a finite number of maxima and minima in the time interval. 3. The function xðtÞ must be absolutely integrable on the interval, i.e., tZ0 þT

jxðtÞjdt ¼ finite51:

t0

Fortunately, all signals that we will be interested in satisfy these properties. The functions that do not satisfy the Dirichlet conditions are only of theoretical interest. Dirichlet gave an example that does not satisfy the conditions mentioned above and is x2p ðtÞ ¼

1;

t-rational

0; t-irrational

:

Our goal is to express a well-behaved function xðtÞ by an approximate function xa ðtÞ in terms of an independent set of functions ffk ðtÞg and a set of constants c½k in the form

xa ðtÞ ¼

N X

c½kfk ðtÞ:

(3:1:1)

k¼N

The subscript a on x in (3.1.1) denotes that it is an approximation of the function xðtÞ: Without loosing any generality we can assume that the limits on the sum N ! 1. We will be interested in a finite N that satisfies some constraints on the error signal, i.e., the difference between the given signal and its approximation. The entries in the expansion are assumed to have the following properties:

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_3, Ó Springer ScienceþBusiness Media, LLC 2010

71

72

1. The constants c½k are assumed to be some constants and k is an integer. 2. The set ffk ðtÞ; k ¼ N; ðN 1Þ; . . . 1; 0; 1; . . . ; ðN 1Þ; Ng is a linearly independent set. That is, fk ðtÞ cannot be obtained as a linear combination of the other fn ðtÞ; n 6¼ k. Such a set is called a basis function set and the members of this set are called basis functions. The basis functions can be real or complex. 3. Finally, we like to consider a basis set that is independent of xðtÞ: These properties are based on common sense. The first one allows for a level adjustment. The second property allows for the use of a set of independent basis functions. The third property allows for a general analysis. Later on we will see that some basis functions may be more attractive than others for a particular application. Fourier used the sine and cosine functions as basis functions. The most important aspect of generalized Fourier series expansion is that it allows an arbitrary function, defined over a finite interval, and may have discontinuities to be represented as a sum of basis functions, such as sine and cosine functions instead of using Taylor’s series:

3 Fourier Series

most of the standard circuits and systems text books. For example, see Ambardar (1995), Haykin and Van Veen (1999), Carlson (1998), Hsu (1967), and many others. Also, see Carslaw (1950), Jeffrey, (1956), Tolstov (1962), and Zygmund (1955). In Chapter 8 we will approximate a function by making use of samples of a signal in combination with some interesting interpolation functions. The presentation starts with the generalized Fourier series and later the Fourier series as a member of the generalized class of series. The basis functions are independent. The expansion gets easier if the basis functions are orthogonal.

3.2 Orthogonal Basis Functions The set of basis functions ffk ðtÞg is an orthogonal basis set if the functions satisfy tZ0 þT t0

x00 ðaÞðxaÞ2 xðtÞ¼xðaÞþx0 ðaÞðxaÞþ þ 2! xðnÞ ðaÞðtaÞn d n xðtÞ þ; xðnÞ ðaÞ¼ þ jt¼a n! dtn (3:1:2) This is an approximation of the function xðtÞ based upon the value of the given function at a point t ¼ a and the values of the derivatives of the function at that point. The Taylor series gives a strict prediction of xðtÞ at a finite distance from xðtÞjt¼a , whereas the Fourier series gives information of the function over the entire range t0 t 5t0 þ T. Another striking difference is the coefficients in the Taylor series are based upon the derivatives of the function at t ¼ a and the Fourier series coefficients are obtained by integration. Furthermore we can use (3.1.2) only if we know all the derivatives of the function at t ¼ t0 : If not, we have to resort to other methods, such as approximating them. The material in this chapter is fairly standard and can be found in

fk ðaÞfm ðaÞda ¼

Ek k ¼ m : 0 k 6¼ m

(3:2:1)

The superscript (*) on fm ðtÞ indicates complex conjugation. If fm ðtÞ is real, then fm ðtÞ ¼ fm ðtÞ. The symbol Ek is used to denote the energy in the basis function, fk ðtÞ in the given time interval and Ek is real. That is, Ek > 0ð assuming fk ðtÞ 6¼ 0Þ:

(3:2:2)

When k 6¼ m in (3.2.1), the integral is zero, which is the orthogonality property of the basis functions. If Ek ¼ 1 in (3.2.1), then the basis set is an orthonormal set. Orthonormality is not critical in our expansion, as we can create an orthonormal set by normalizing an orthogonal set, i.e., by replacing pﬃﬃﬃﬃﬃﬃ fk ðtÞ by fk ðtÞ= Ek . Therefore, we will concentrate on using orthogonal basis sets instead of orthonormal basis sets. Example 3.2.1 Show that the set ff1 ðtÞ; f2 ðtÞg given below is an independent set: t1 f1 ðtÞ ¼ P½t 0:5 and f2 ðtÞ ¼ P : (3:2:3) 2

3.2 Orthogonal Basis Functions

φ1(t)

73

φ2(t)

Example 3.2.3 Show that the following set is an orthogonal basis set over the time interval t0 t5t0 þ T and give the values of Ek :

Solution: The members of the set are sketched in Fig. 3.2.1 and the set is an independent set since any one of the members cannot be expressed in terms of the others. Since they overlap, and the time width of the second member is longer than the first member, they are not orthogonal. This can be seen from the integral f1 ðaÞf2 ðaÞda ¼

Z1

0

¼ 1 6¼ 0:

(3:2:4) &

Example 3.2.2 Consider the pulse functions given below and show that they form an orthogonal basis set. Find the value of A that makes the set an orthonormal set. t ðT=6Þ f1 ðtÞ ¼ AP ; T=3 t ðT=2Þ ; f2 ðtÞ ¼ AP T=3 t ð5T=6Þ f3 ðtÞ ¼ AP : (3:2:5) T=3 Solution: The pulse functions are shown in Fig. 3.2.2. Clearly, each pulse function exists in a different interval and therefore they are orthogonal. Since all the pulses are of the same width and the same height, we can write 2

2

A da ¼ ðA TÞ=3:

0

The functions are orthonormal if A ¼ Ei ¼ 1; i ¼ 1; 2; 3: φ1(t)

φ2 (t)

φ3 (t)

(3:2:6) qﬃﬃﬃ 3 T

tZ0 þT

xT ðaÞda ¼

Z

xT ðaÞda:

(3:2:8)

T

The integral on the right is over any period. Using the orthogonality property, we have Z Z k 6¼ m : fk ðaÞfm ðaÞda ¼ ejko0 a ejmo0 e dt T

¼

ZT e

T

jðkmÞo0 a

da ¼ e

jðkmÞo0 a

T 1 jðk mÞo0 0

0

¼

11 ¼ 0: ðk mÞo0 Z k ¼ m : ejko0 a ejmo0 a da

(3:2:9a) &

T

¼

ZT

(3:2:9b) da ¼ T; Ek ¼ T ¼ E:

0

Example 3.2.4 Test the orthogonality over the interval ðt0 ; t1 Þ of the set of functions ffk ðtÞ; k ¼ 0; 1; 2; . . .g ¼ 1; t; t2 ; . . . : (3:2:10)

and &

Solution: First Zt1 t0

Fig. 3.2.2 Pulse functions fi ðtÞ; i ¼ 1; 2; 3

(3:2:7b)

If a function xT ðtÞ ¼ xT ðt þ TÞ; then a short hand notation (see (1.5.14)) is

t0

0

E1 ¼ E2 ¼ E3 ¼

ejko0 ðtþTÞ ¼ ejko0 t ejko0 T ¼ ejko0 t ; o0 T ¼ 2p; f0 ¼ 2p=o0 ¼ 1=T:

P½a :5P½ða 1Þ=2da

ZT=3

(3:2:7a)

Solution: Note that fk ðtÞ ¼ fk ðt þ TÞ with the period T ¼ 2p=o0 since

Fig. 3.2.1 Pulse functions fi ðtÞ, i = 1, 2

Z2

fk ðtÞ ¼ ejko0 t ; k ¼ 0; 1; 2; . . . :

f1 ðaÞf2 ðaÞda ¼

Zt1

1 ada ¼ ½t21 t20 : 2

t0

The functions f1 ðtÞ and f2 ðtÞ are orthogonal if t1 ¼ t0 or if t1 ¼ t0 . Now consider the two functions f0 ðtÞ ¼ 1 and f2 ðtÞ ¼ t2 .With these, we have

74

3 Fourier Series

Zt1

f0 ðaÞf2 ðaÞda ¼

t0

Zt1

Z1

1 a da ¼ ðt31 t30 Þ: 3 2

t0

P1 ðtÞ ¼ f1 ðtÞ 1

The functions f0 ðtÞ and f2 ðtÞ are not orthogonal over any interval, except in the trivial case of t0 ¼ t1 . The set in (3.2.10) is not a good set to represent all & signals.

P0 ðaÞf1 ðaÞda Z1

P20 ðaÞda

1

Z1 P0 ðtÞ ¼ t 1

3.2.1 Gram–Schmidt Orthogonalization

ð1Þada P0 ðtÞ ¼ t:

Z1 da

1

Consider a linear set of independent real functions, f1 ðtÞ; f2 ðtÞ; . . . ; fk ðtÞ; . . ., defined on the interval ½a; b. Now define a new set of functions fj1 ðtÞ; j2 ðtÞ; . . . ; jk ðtÞ; . . .g by j1 ðtÞ ¼ f1 ðtÞ; Zb j1 ðaÞf2 ðaÞda j2 ðtÞ ¼ f2 ðtÞ

Similarly we can determine P2 ðtÞ ¼ t2 ð1=3Þ. We can multiply these polynomials by a constant since multiplying a polynomial in the set by a constant does not change the orthogonality of the polynomials. The above process generates the Legendre polynomials within a constant. The first five Legendre polynomials are listed below:

j1 ðtÞ

a

Zb

L0 ðtÞ ¼ 1;

j21 ðaÞda

L1 ðtÞ ¼ t;

Zb

L3 ðtÞ ¼ ð1=2Þð5t3 3tÞ;

L2 ðtÞ ¼ ð1=2Þð3t2 1Þ;

a

j3 ðtÞ ¼ f3 ðtÞ

j1 ðaÞf3 ðaÞda

L4 ðtÞ ¼ ð1=8Þð35t4 30t2 þ 3Þ:

a

Zb

Note the constant factors between Pi ðtÞ and Li ðtÞ. These polynomials can be generated by Rodrigue’s formula Spiegel (1968):

j21 ðaÞda

a

Zb j1 ðtÞ

j2 ðaÞf3 ðaÞda

a

Zb

Lk ðtÞ ¼ j2 ðtÞ; . . .:

(3:2:11)

j22 ðaÞda

a

This process of generating an orthogonal set of functions starting with an independent set is called the Gram–Schmidt orthogonalization process. Example 3.2.5 Use the Gram–Schmidt process to generate an orthogonal basis set, fPn ðtÞ; n ¼ 0; 1; 2; 3; . . .g; in the interval 1 t 1 using (3.2.11) in Example 3.2.4. Solution: First two are P0 ðtÞ ¼ f0 ðtÞ ¼ 1:

1 dðt2 1Þk ; k ¼ 0; 1; 2; 3; . . . : (3:2:12a) dtk 2k k!

The polynomials generated by this process are referred to as special Legendre polynomials. Note the subscript k is used as an index, which is different from p used in the Lp measures. They satisfy the orthogonality property ( Z1 0; m 6¼ k Lm ðaÞLk ðaÞda ¼ : 2 Ek ¼ ð2 kþ1Þ ;m¼k 1

(3:2:12b) & Example 3.2.6 Show the set of periodic functions given below is an orthogonal basis set over one period and compute the energy in each of the basis functions in one period:

3.3 Approximation Measures

75

f1; cosðo0 tÞ; cosð2o0 tÞ; . . . ; cosðko0 tÞ; . . . ; sinðo0 tÞ; sinð2o0 tÞ; . . . ; sinðko0 tÞ; . . .g:

Z T

ð1Þda ¼T;

Z

(3:2:13)

ð1Þ cosðko0 aÞda ¼ 0;

T

Solution: The members of the set are periodic with period T ¼ 2p=o0 and we need to show (3.2.1) using the members of the given set: Z

ð1Þ sinðko0 aÞda ¼ 0;

k ¼ 1; 2; . . . :

(3:2:14a)

T

Using trigonometric identities, we have 8 R R R cos2 ðko0 aÞda ¼ 12 da þ 12 cosð2 ko0 aÞda ¼ T2 ; k ¼ m > < T T T R : cosðko0 aÞ cosðmo0 aÞda ¼ 1 R 1 > : 2 cosððk þ mÞo0 aÞda þ 2 cosððk mÞo0 aÞda ¼ 0; k 6¼ m

Z T

T

Z

sinðko0 aÞ sinðmo0 aÞda ¼

1 2

Z

T

Z

T

cosðk mÞo0 ada

1 2

T

sinðko0 aÞ cosðmo0 aÞda ¼

1 2

Z

T

Z

T cosðk þ mÞo0 ada ¼

T

sinððk þ mÞo0 aÞda þ

T

1 2

Z

;k ¼ m : 0; k 6¼ m 2

sinððk mÞo0 aÞda ¼ 0; (3:2:14b)

T

for all k and m: These prove that the set in (3.2.13) is an orthogonal set. The energies contained in the members of the basis set in one period are as follows: ðEÞ1 ¼ T; ðEÞsine or a cosine function ¼ T=2:

(3:2:15) &

The set is an orthogonal set and not orthonormal set. There are many other basis sets.

3.3 Approximation Measures We are interested in approximating a given function xðtÞ over an interval (t0 ; t0 þ TÞ by xa ðtÞ using a set of orthogonal basis functions. How do we measure the approximation and then how good is the approximation? It can be measured by the error ½xðtÞ xa ðtÞ. Figure 3.3.1 illustrates an example where xðtÞ is the given function and its approximation is xa ðtÞ. The hatched area represents the error. Since the functions can be complex and a positive error is just as bad as a negative error, and to make it general, we would like to consider the magnitude of

Fig. 3.3.1 xðtÞ Given function, xa ðtÞ Function approximating xðtÞ

the error. In addition, if the error measure is a number, we can compare and evaluate a particular approximation with respect to a number of basis sets. These goals can be achieved by considering the integral of the pth power of the magnitude of the error function, i.e., tZ0 þT

jxðtÞ xa ðtÞjp dt; 1 p:

(3:3:1)

t0

Notes: There is a good deal of interest in the area of inverse problems, such as deconvolution of signals based on Lp ; 1 p measures. For a

76

3 Fourier Series

review see Tarantola (1987) and Hassan et al. (1994). Statisticians have investigated the Lp measures based on probabilistic behavior of signals. Our discussion here does not involve any details of statistical analysis. For readers interested in statistical details of these measures, p is selected based upon the kurtosis defined as the fourth moment normalized by the square of the variance of a probability density function Money (1982). In Section 1.7 we have considered three important density functions, uniform, Gaussian, and Laplacian. These can be used in selecting the value of p. Kurtosis values are k = 1.8 (Uniform, 3 ðGaussianÞ; and 6 ðLaplacianÞ: The constant p is selected by

9 p¼ ; 1 k 1; k2 þ 1 9 8 > = < k > 3:8 use L1 > 2:25k53:8 use L2 : > > ; : k52:2 use L1

(3:3:2)

1 MSE ¼ T

(3:3:3)

Minimizing the error (see e2 in (3.3.a)) corresponds to maximizing the probability, thus

tZ0 þT

jxðtÞ xa ðtÞj2 dt; xa ðtÞ

t0

¼

N X

c½kfk ðtÞ S2 Nþ1 :

(3:3:4a)

k¼N

The interval T will be the same for different approximations in a particular situation and the normalization constant can be omitted and compare the approximations by using an orthogonal basis set ffk ðtÞg and the integral-squared error (ISE), i.e., tZ0 þT

Lp measures are used for speech, seismic, radar, and other signal coding. For example, for vowel sounds, L1 is preferable and for nonvowel sounds, L2 is preferable, see Lansford and Yarlagadda (1988). Since seismic signals have spiky noise, L1 seems to work well, see Yarlagadda et al. (1985). See Schroeder and Yarlagadda (1989) on spectral estimation using L1 norm. An old adage is if you have a fork in the road and have a choice to select either L1 or L2 measure, L2 tells you to go in the middle of the two roads, not one of the two possible paths, whereas L1 suggests taking one or the other paths. L2 measure thinks like a machine, whereas L 1 measure thinks like a human, see Problem 3.2.2 at the end of the chapter. For most applications, L2 the least-squares error measure is adequate and simple to use. In Section 1.7.1, the Gaussian probability density function was introduced, see (1.7.12). Writing it terms of the error e with mean 0 and variance s2e , the density function is 1 2 2 fe ðeÞ ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ee =2se : 2 2pse

providing a mathematical basis for the least& squares approach. The mean-squared error (MSE) is defined by considering ð2 N þ 1Þ terms, we have

ISE¼

jxðtÞxa ðtÞj2 dt

t0

¼

tZ0 þT t0

2 N X c½kfk ðtÞ dt e2Nþ1 : (3:3:4b) xðtÞ k¼N

The subscript on e; 2 N þ 1 corresponds to the number of terms in the approximation and some of the coefficients c½k may be zero and N could go to infinity. It is convenient to consider odd number of terms. Since c½k0s are unknowns, there is no loss in generality. The constants c½k can be determined from the basis set ffk ðtÞ; k ¼ N; . . . ; 2; 1; 0; 1; 2; . . . ; Ng by minimizing the ISE. We will consider two ways of computing the constants that minimize the integral-squared error. The first one is based upon taking the partials of the ISE with respect to c½k, equating the partials to zero, and then solving for them. The second one is based on using perfect squares by rewriting the ISE in terms of two parts. First term is independent of c½k and the second is a sum of perfect square terms involving c½k. Equating the perfect square terms to zero and solving for c½k give the desired result.

3.3 Approximation Measures

77

Using this in (3.3.6) results in

3.3.1 Computation of c[k] Based on Partials First,

tZ0 þT

tZ0 þT

e2 Nþ1 ¼

2 N X c½kfk ðtÞ dt xðtÞ k¼N

t0 tZ0 þT

¼

x ðtÞfk ðtÞdt þ c ½kðEk Þ ¼ 0:

t0

The coefficients c½k are the generalized Fourier series coefficients giving an explicit formula given below to compute c½k given xðtÞ and the orthogonal basis setfjk ðtÞg:

jxðtÞ S2 Nþ1 j2 dt

t0 tZ0 þT "

xðtÞ

¼

t0

#

N X

tZ0 þT

1 c½k ¼ Ek

c½mfm ðtÞ #

N X

x ðtÞ

(3:3:9)

t0

m¼N

"

xðtÞfk ðtÞdt:

c ½kfk ðtÞ dt:

(3:3:5)

k¼N

Note the two different variables k and m in the above summations. This allows us to keep track of the terms in the summation products. Multiplying the product terms and since the integral of a sum is equal to the sum of the integrals, we have

e2Nþ1 ¼

tZ0 þT

xðtÞx ðtÞdt

N X

c ½k

k¼N

þ

tZ0 þT

c½m

m¼N

t0

N X

tZ0 þT

x ðtÞfm ðtÞdt

t0

xðtÞfk ðtÞdt

t0

N N X X

c½mc ½k

k¼N m¼N

tZ0 þT

fm ðtÞfk ðtÞdt:

t0

(3:3:6) Take the partial derivatives of e2 Nþ1 with respect to c½m and equate them to zero: @e2Nþ1 ¼ @c½m

tZ0 þT

½x ðtÞfm ðtÞdt

t0

þ

N X

tZ0 þT

c½k

k¼N

3.3.2 Computation of c[k] Using the Method of Perfect Squares Example 3.3.1 Consider the second-order polynomial yðtÞ ¼ t2 þ bt. Determine the minimum value of yðtÞ by having a part of the expression that is a perfect square. Solution: By adding and subtracting the term ðb=2Þ2 to yðtÞ, we have

2 2 b b 2 2 yðtÞ ¼ t þ bt ¼ t þ bt þ 2 2

2 2 b b ¼ tþ : 2 2 This function takes a minimum value for t ¼ b=2 & and yðtÞ ¼ ðb=2Þ2 : This idea can be used in minimizing the integralsquared error, See Ziemer and Tranter (2002). Add and subtract the following term to e2 Nþ1 in (3.3.6): 2 t þT Z0 N X 1 xðtÞfk ðtÞdt : E k¼N k t0

fm ðtÞfk ðtÞdt ¼ 0: (3:3:7)

t0

Orthogonality of the basis functions implies

eð2Nþ1Þ ¼

tZ0 þT

( 2

jxðtÞj dtþ

k¼N

t0 tZ0 þT t0

fk ðtÞfm ðtÞdt ¼

Ek > 0 and real; k ¼ m 0;

otherwise

: (3:3:8)

N X

N X k¼N

tZ0 þT

c½k t0

tZ0 þT

c½k t0

xðtÞfk ðtÞdt

x ðtÞfk ðtÞdt

78

3 Fourier Series

2 t þT approximation to the function xðtÞ, only in the Z0 1 1 X X sense that the approximation gives a smaller mean 1 þ xðtÞfk ðtÞdt þ Ek ðc½kc ½kÞg square error. In the limit, for any complete set of E k¼1 k¼1 k t0 basis functions defined below: 2 t þT 0 Z 1 X 1 (3:3:13) lim e2 Nþ1 ¼ 0: N!1 xðtÞfk ðtÞdt : (3:3:10) E k¼1 k t 0 From (3.3.12), we have Parseval’s equation, or forThe terms inside the brackets {.} can be expressed as mula, or identity given by a sum of squares and

e2 Nþ1 ¼

tZ0 þT

2

jxðtÞj dt þ

tZ0 þT N X

jxðtÞj2 dt ¼

Ek jc½kj2 :

(3:3:14)

k¼1

t0

k¼N

t0

2 tZ0 þT pﬃﬃﬃﬃﬃﬃ 1 Ek c½k pﬃﬃﬃﬃﬃﬃ xðtÞfk ðtÞdt Ek t0 2 t þT Z0 N X 1 xðtÞfk ðtÞdt : (3:3:11) E k¼N k

Summary: Given a time-limited function xðtÞ; t0 t5t0 þ T and a set of orthogonal basis functions ffk ðtÞ; k ¼ 0; 1; 2; . . .g, the function xðtÞ is approximated by

To show that (3.3.11) reduces to (3.3.10), expand the middle term on the right and then cancel the equal terms that have opposite signs. The parameters c½k are not in the first and the third terms in (3.3.11). It is only included in the middle term, which is a sum of absolute values. Therefore, minimization of e2 Nþ1 is achieved when the middle term is zero and c½k is given by (3.3.9).

1 c½k ¼ Ek

The minimum ISE with ð2 N þ 1Þ terms is (see (3.3.5))

Ek jc½kj2

tZ0 þT

xðtÞfk ðtÞdt

(3:3:15)

t0

and the integral-squared error is equal to zero. The function xa ðtÞ is an approximation of xðtÞ in the identified interval. Only in the sense that the integral-squared error goes to zero, we write the equality of the given function to the approximate function by 1 X

c½kfk ðtÞ:

(3:3:16)

jxðtÞ S2 Nþ1 j2 dt 0:

It should be emphasized that xðtÞ and xa ðtÞ are not equal in the true sense. Differences between these two functions will be considered a bit later. In simple terms, the coefficients c½k of the generalized Fourier series are 2 3 6 6 c½k ¼ 6 4

k¼N

t0

1 ¼ T

N X

tZ0 þT

k¼1

3.3.3 Parseval’s Theorem

jxðtÞj2 dt

c½kfk ðtÞ;

k¼1

xðtÞ ¼ xa ðtÞ ¼

tZ0 þT

1 X

xa ðtÞ ’

t0

e2 Nþ1 ¼

1 X

(3:3:12)

t0

As N increases, the partial the quantity ðe2 Nþ1 Þ can only decrease. Therefore, as N increases, the partial sums of the F-series give a closer and closer

7 tZ0 þT 1 7 7 Ek , Energy in fk ðtÞ 5 t0 in the interval ðt0 ;t0 þ TÞ ðconjugate of the dt: (3:3:17a) xðtÞ basis function, fk ðtÞÞ

The error in (3.3.12) with ð2 N þ 1Þ coefficients in the F-series expansion is equal to

3.3 Approximation Measures

79

E2 Nþ1 ¼ ðEnergy in the given T second interval of the function) " # N X th ðSquared magnitude of the kth coefficient) (Energy in the k basis function) k¼N

¼

tZ0 þT t0

jxT ðtÞj2 dt

N X

jc½kj2 Ek ¼ ISE: (3:3:17b)

k¼N

Later, the convergence of the approximated signal xa ðtÞ to xðtÞ in terms of the number of coefficients in the approximation will be considered. The signal xðtÞ is assumed to satisfy the Dirichlet conditions. The value of the integral in (3.3.17b) is unique and the generalized Fourier series is unique for a given set of basis functions. Let us illustrate the above by a detailed example, see Simpson and Houts (1971). &

Earlier we have shown that fi ðtÞ; i ¼ 1; 2; 3 form an orthogonal basis set with Ek ¼ A2 T=3; k ¼ 1; 2; 3, see Example 3.2.2. Noting that the time interval of the given function is 0 t5T and from (3.3.9), the coefficients are as follows: 1 c½1 ¼ E1

xðtÞf1 ðtÞdt

t0 ¼0

1 ¼ 2 A ðT=3Þ

ZT

3 tf ðtÞdt T 1

0

Example 3.3.2 Find the generalized Fourier series expansion of the function xðtÞ given below in Fig. 3.3.2 using the three orthogonal basis functions in (3.2.5). Then compute the mean-squared error using the direct method and Parseval’s theorem.

ZT=3 1 3 t:Adt ¼ 2 A ðT=3Þ T 0

2 1 3 3 t t ¼ T3 1 A ¼ ¼ 2 2 t ¼ 0 2A A T T

Solution: The expression for the time function xðtÞ and the three basis functions are t T2 3 ; xðtÞ ¼ t P T G t T6 ; f1 ðtÞ ¼ AP T=3 t T2 f2 ðtÞ ¼ AP ; and T=3 t 5T 6 f3 ðtÞ ¼ AP : T=3

t0 Z þT¼T

c½ 2 ¼

1

T

A2

3

2T=3 Z

3 3 5 t Adt ¼ ; c½ 3 ¼ : T 2A 2A

T=3

xa ðtÞ ¼ c½1f1 ðtÞ þ c½2f2 ðtÞ þ c½3f3 ðtÞ 1 t T=6 3 t T=2 ¼ P þ P 2 T=3 2 T=3 5 t 5t=6 : þ P 2 T=3 (3:3:18)

The functions xðtÞ and xa ðtÞ are sketched in Fig. 3.3.2. The error is shown by the hatched marks and it has six equal parts. Clearly, ZT=6 ZT=6

3 1 2 2 ISE ¼ 6 t ½xðtÞ xa ðtÞ dt ¼ 6 dt T 2 0

0

9t2 1 3t þ dt T2 4 T 0 1 3 t2 9 t3 T=6 T þ ¼ ¼6 t 4 T 2 T2 3 0 12

¼6

Fig. 3.3.2 xðtÞ, xa ðtÞ and the error between these two functions

(3:3:19)

ZT=6

(3:3:20a)

This procedure is complicated and unnecessary since some of this work has already been done

80

3 Fourier Series

and should be evident from the context. The equality is only true in the sense that the integralsquared error between the periodic function and the F-series in the given interval is zero. The given function and the corresponding Fourier series 2 # coefficients are identified by the symbolic notation:

in deriving the generalized Fourier series coefficients. Using Parseval’s theorem and (3.3.17b) results in the same value as in (3.3.20a) and is shown below.

ISE ¼

ZT

3 t T

2

" 2 2

T 1 3 5 dtA þ þ 3 2A 2A 2A 2

0

FS;T

xT ðtÞ ! Xs ½k

35 T ¼3T T¼ : 12 12

(3:4:1b)

(3:3:20b) &

3.4 Fourier Series The generalized Fourier series developed earlier can be used to study both the complex and the trigonometric Fourier series of a periodic function xT ðtÞ with period of T seconds. The functions ejno0 t ; sinðno0 tÞ; and cosðno0 tÞ are nice functions in the sense that all the derivatives of these functions exist. Approximation of a given function by the Fourier series (F-series) gives a smooth function even when the function being approximated has discontinuities.

The subscript s on X is used to denote that the coefficients are the complex F-series coefficients. Note the difference in the sign of the exponents in the sum and the integral expressions in (3.4.1a). Fourier coefficients are computed by an integral and the integral has a unique value. That is, F-series expansion is unique. The complex F-series can be used to approximate an aperiodic function in a time interval (t0 ; t0 þ TÞ, where t0 is arbitrary and T is the interval of the function that is under consideration. The approximation, in terms of periodic basis functions, will be valid only in the given time interval and, outside this interval, it is not valid. Complex Fourier series is applicable to both real and complex functions. When the function xT ðtÞ is real, the F-series coefficients Xs ½k and Xs ½k are related. Z 1 xT ðtÞejko0 t dt T T 9 8 = 2:

The F-series in (3.4.13b) contains only sine terms since x2p ðtÞ is an odd function. It is more appealing & since the given function is real.

Since the function is real and even, the above validates Xs ½k ¼ Xs ½k. The F-series coefficients are unique. Using trigonometric identity, it follows that

Example 3.4.4 Derive the complex F-series given the periodic impulse sequence

xT ðtÞ ¼ ð1=2Þ þ 2 cosðo0 tÞ ð1=2Þ sinð2o0 tÞ:

xT ðtÞ ¼

1 X

Adðt kTÞ:

(3:4:14)

k¼1

Solution: The complex F-series coefficients are

&

The complex F-series expansion is applicable for both real and complex periodic functions and the complex F-series leads to Fourier transforms. Most functions, in reality, are real functions and the trigonometric F-series are more desirable.

3.4 Fourier Series

83

3.4.2 Trigonometric Fourier Series

3.4.3 Complex F-series and the Trigonometric F-series The set of basis functions for the trigonometric Coefficients-Relations Fourier series with period T ¼ 2p=o0 is

f1; cosðko0 tÞ; sinðko0 tÞ; k ¼ 1; 2; 3; . . .g: (3:4:16a)

By using Euler’s formula and comparing the results in (3.4.1a) and (3.4.17), we have

Equation (3.2.15) gives the energy contents in one period of these functions and are Xs ½k ¼ ðEÞ1 ¼ T; ðEÞ sine or a cosine function ¼ T=2:

1 X

Z

xðtÞejko0 t dt

T

(3:4:16b)

In the trigonometric F-series, Xs ½0 for the dc term, a½k for the coefficients for the cosine terms, and b½k for the coefficients of the sine terms will be used. With (3.4.16b) and (3.3.17a), the following trigonometric F-series result: xT ðtÞ ¼ Xs ½0 þ

1 T

1 ¼ T

Z

j xðtÞ cosðko0 tÞdt T

T

¼

(3:4:19a)

Since the cosine and sine functions are even and odd functions, we can see that

fa½k cosðko0 tÞ þ b½k sinðko0 tÞg; Xs ½k ¼

Z 1 Xs ½0 ¼ xT ðtÞdt T T Z 2 a½k ¼ xT ðtÞ cosðko0 tÞdt: T > > T > > Z > > 2 > > > b½k ¼ xT ðtÞ sinðko0 tÞdt > : T

xðtÞ sinðko0 tÞdt T

a½k b½k j : 2 2

k¼1

8 > > > > > > > > > >

0:

(3:6:3)

The magnitudes of the F-series coefficients of a periodic signal and its delayed (or advanced) version are the same. The delay (or advance) t appears explicitly in the phase angle ðko0 tÞ for the advance and ðko0 tÞ for the delay. In case of trigonometric F-series, for t > 0 (for t50, replace t by t in the following.), we have

88

3 Fourier Series

xT ðttÞ¼Xs ½0þ

1 X ½a½kcosðko0 ðttÞÞ

Solution: Substituting T ¼ 2p, i.e., o0 ¼ 1 in (3.5.9), we have

k¼1

þb½ksinðko0 ðttÞÞ¼Xs ½0 1 X ½a½kcosðko0 tÞb½ksinðko0 tÞ þ k¼1

cosðko0 tÞ þ

1 X

2 4 x2p ðtÞ ¼ p p cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : 1ð 3Þ 3ð 5Þ 5ð 7Þ (3:6:7) &

½a½ksinðko0 tÞ

k¼1

þb½kcosðko0 tÞsinðko0 tÞ:

(3:6:4)

Example 3.6.2 Find the trigonometric F-series of the signal below using the full-wave rectified signal series: y2p ðtÞ ¼

3.6.3 Time and Frequency Scaling Given a function xðtÞ, its time-scaled version is given by xðatÞ; a > 0. If a51, the signal is expanded and if a > 1, the signal is compressed (see (1.2.3)). The F-series of the scaled signal is obtained by replacing t ! at in the complex F-series and 1 X xT ðatÞ ¼ Xs ½kejko0 ðatÞ : (3:6:5)

sinðtÞ; 0;

05t5p : p5t52p

(3:6:8)

Solution: The full-wave rectified function x2p ðtÞ was shown in Fig. 3.5.1 with period T ¼ 2p. The half-wave rectified signal is shown in Fig. 3.6.1 and y2p ðtÞ ¼ ½x2p ðtÞ þ sinðtÞ=2:

(3:6:9)

k¼1

The frequency locations moved from ko0 to kðao0 Þ and the F-series coefficients are Xs ½k. The time-scaled signal changes the period from T to ðT=aÞ: Harmonics are now located at ko0 a ¼ kð2pÞf0 a ¼ kð2pÞða=TÞ. The more compressed the time function is, farther apart its harmonics are. Harmonics of the compressed signal are located at frequencies kf0 a and a > 1: If a ¼ 1, then a time-reversed or a folded function results and xT ðtÞ ¼

1 X

Xs ½kejko0 t ¼

k¼1 FS;T

1 X Xs ½kejko0 t :

Fig. 3.6.1 y2p ðtÞ Half-wave rectified signal

Using the linearity property of the Fourier series, we have 1 1 2 y2p ðtÞ¼ þ sinðtÞ p 2 p cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : 1ð3Þ 3ð5Þ 5ð7Þ

k¼1

xT ðtÞ ! X ½k; jXs ½kj ¼

jXs ½kj:

(3:6:10) (3:6:6)

Example 3.6.1 Using Example 3.5.2, find the trigonometric F-series of the full-wave rectified signal x2p ðtÞ ¼ jsinðtÞj assuming the period is 2p.

In Chapter 1, we studied the one-sided and the two-sided line spectra by expressing each term in terms of cosines and sines. Noting that cosða 900 Þ ¼ sinðaÞ and cosða 1800 Þ ¼ cosðaÞ, we can write

1 1 2 cosð2t 1800 Þ cosð4t 1800 Þ cosð6t 1800 Þ 0 þ þ þ : y2p ðtÞ ¼ cosðt þ 90 Þ þ p 2 p 1ð3Þ 3ð5Þ 5ð7Þ

(3:6:11)

3.6 Operational Properties of Fourier Series

89

2 4 cosð2t 1800 Þ cosð4t 1800 Þ cosð6t 1800 Þ þ þ þ : x2p ðtÞ ¼ þ p p 1ð3Þ 3ð5Þ 5ð7Þ These are the harmonic forms of the trigonometric Fourier series for the two given functions. The two-sided amplitude line spectra associated with these two functions are sketched in Fig. 3.6.2. The only difference between the two functions x2p ðtÞ and y2p ðtÞ is the component at the frequency o0 ¼ 1 or f0 ¼ 1=2p. If we can remove or filter out this frequency component from y2p ðtÞ, we can obtain the function x2p ðtÞ illustrating one of the remarkable insights into the description of signals provided by the Fourier series. Filter design & will be discussed in later chapters.

(3:6:12)

The dc term in the F-series of the derivative goes to zero and the other coefficients are multiplied by ðjko0 Þ, k 6¼ 0: The spectral components of x0T ðtÞ have significantly higher frequency content compared to the spectral components of xT ðtÞ. Note the F-series coefficients of xT ðtÞ are multiplied by jko0 to obtain the F-series coefficients of x0T ðtÞ. Derivative operation enhances the details in the signal. We can state that d n xT ðtÞ FS;T !ðjko0 Þn Xs ½k; k 6¼ 0: dtn

(3:6:13)

In a similar manner, the derivatives of the trigonometric F-series are given by xT ðtÞ ¼ Xs ½0 þ

1 X

a½k cosðko0 tÞ

k¼0

þ

(a)

1 X

b½k sinðko0 tÞ:

k¼0 1 dxT ðtÞ X ¼ ðko0 Þa½k sinðko0 tÞ dt k¼1

þ

1 X

ðko0 Þb½k cosðko0 tÞ:

k¼1

(b) Fig. 3.6.2 Two-sided amplitude line spectra (a)x2p ðtÞ and (b) y2p ðtÞ

¼

1 X

þ

3.6.4 Fourier Series Using Derivatives

a1 ½k cosðko0 tÞ

k¼1 1 X

b1 ½k sinðko0 tÞ;

a1 ½k ¼ b½kðko0 Þ;

k¼1

b1 ½k ¼ a½kðko0 Þ All the derivatives of the F-series exist since all the derivatives of the functions ejno0 t ; sinðno0 tÞ; and cosðno0 tÞ exist. In that sense, F-series is a nice function. The complex F-series of a periodic function and its derivative are 1 X Xs ½kejko0 t ; xT ðtÞ ¼ k¼1 1 dxT ðtÞ X x0T ðtÞ ¼ ½Xs ½kjðko0 Þejko0 t ¼ dt k¼1 k6¼0

b1 ½k a1 ½k ; b½k ¼ ; k 6¼ 0; ko0 ko0 Z 1 Xs ½0 ¼ xT ðtÞdt: T

a½k ¼

(3:6:14)

T

In (3.6.14), the subscript ‘‘1’’ on a and b indicates the trigonometric F-series are for the first derivative of the periodic function. The dc component needs to be computed directly from the given periodic

90

function. This approach of finding the F-series is the derivative method of finding F-series. The derivative property allows simplifies the computing the F-series coefficients. Most of the signals we deal with are pulses that do not have derivatives in the conventional sense. The derivatives of such functions can only be considered in the sense of generalized functions discussed in Section 1.4, resulting in impulse functions in the derivatives. See, for example, the derivative of the rectangular pulse in (1.4.33). Fourier series coefficients are

3 Fourier Series

determined by use of integrals. Integrals involving impulses are trivial to compute and, therefore, computing the Fourier series coefficients by the derivative method makes it simple. Example 3.6.3 Find the trigonometric F-series of the trapezoidal waveform shown in Fig. 3.6.3a using the derivative method. Solution: The first two derivatives of the wave form are shown in Figs. 3.6.3b and c. The second derivative of the function xT ðtÞ is

(a)

(b)

Fig. 3.6.3 (a) Trapezoidal wave-form xT ðtÞ, (b) xT0 ðtÞ, and (c) xT00 ðtÞ

(c)

3.6 Operational Properties of Fourier Series

x00T ðtÞ ¼

91

d2 xT ðtÞ 1 ¼ dt2 ðd2 d1 Þ

½dðt þ d2 Þ dðt þ d1 Þ dðt d1 Þ þ dðt d2 Þ; d2 d1 ; T=25t5T=2; xT ðt þ TÞ ¼ xT ðtÞ:

considered above decay proportional to (1/k2 Þ. In the case of d2 ¼ d1 ; we have a rectangular pulse waveform. Using L’Hospital’s rule, the coefficients in (3.6.16a) can be simplified. Assuming d2 ¼ d1 þ e,

(3:6:15) Since the second derivative has an even symmetry, we can use some simplifications:

( a½k ¼lim e!0

¼lim

4 Teðko0 Þ2 4

e!0 Tðko

b2 ½k ¼ 0;

2 a2 ½ k ¼ T

ZT=2

x00T ðtÞ cosðko0 tÞdt

T=2

¼

4 ðd2 d1 ÞT

ZT=2

½cosðko0 ðd1 þeÞÞcosðko0 d1 Þ

ðko0 Þsinðko0 ðd1 þeÞÞ 4sinðko0 d1 Þ ¼ 1 kTo0 0Þ 2

1 2d1 X 4 sinðko0 d1 Þ þ cosðko0 tÞ: xT ðtÞ d1 ¼d2 ¼ T ko0 T k¼1 (3:6:16c)

½dðt d1 Þ þ dð1 d2 Þ

0

cosðko0 tÞdt;

a2 ½k ¼

)

4 ½ cosðko0 d1 Þ þ cosðko0 d2 Þ: Tðd2 d1 Þ

Noting that a2 ½k ¼ ðko0 Þ2 a½k, we have the dc term (to be computed directly) and Zd2 1 1 xT ðtÞdt ¼ ðd1 þ d2 Þ: Xs ½0 ¼ T T 4 d2 ½cosðko0 d1 Þ cosðko0 d2 Þ: a½k ¼ Tðd2 d1 Þðko0 Þ2

From (3.6.16b and c), the F-series coefficients of the trapezoidal and the rectangular pulse sequences decay at a rate proportional to ð1=k2 Þ and (1/k), respectively. The derivative of an even (odd) function is an odd (even) function. See Fig. 3.6.3 a,b,c and the F-series to verify the above assertion and the chain rule below: dy dy da dxe ðtÞ ¼ ! yðtÞ ¼ ; dt da dt dt yðtÞ ¼

dxe ðtÞ dxe ðtÞ ¼ ð1Þ ¼ yðtÞ: dt dt (3:6:17) &

The trigonometric F-series are xT ðtÞ ¼

d1 þ d2 4 þ 2 T o0 Tðd2 d1 Þ 1 X 1 ½cosðko0 d1 Þ cosðko0 d2 Þ cosðko0 tÞ: 2 k k¼1

(3:6:16a) xT ðtÞjd1 ¼0 ¼

d2 4 þ T d2 Tðo20 Þ

1 X 1 ½1 cosðko0 d2 Þ cosðko0 tÞ: k2 k¼1

(3:6:16b)

Note that when d1 ¼ 0, we have a triangular pulse wave form. The F-series coefficients for the two cases

3.6.5 Bounds and Rates of Fourier Series Convergence by the Derivative Method The Fourier series coefficients Xs ½k of a periodic signal xT ðtÞ usually decay at a rate inversely proportional to kn , where k is the harmonic index. An exception is the periodic impulse sequence. In Example 3.4.4, we have seen that the periodic impulse sequence has the complex F-series given by Xs ½k ¼ ðA=TÞ, i.e., the coefficients are independent of k. Higher the value of n in kn is, the faster the high-frequency component decays. An estimated value of n can be determined without actually

92

3 Fourier Series

computing the Fourier series coefficient Xs ½k using the derivative properties of xT ðtÞ: In (3.6.13) we have seen that the Fourier series coefficients of the nth derivative of a periodic function are related to the Fourier coefficients of the function multiplied by ðjko0 Þn . If we differentiate an arbitrary periodic function xT ðtÞ n times before the first set of impulses appear, then the F-series coefficients have the property that Xs ½k / ð1=jko0 jÞn . That is, the decay rate of the coefficients is proportional to 1/ðjkjÞn . The decay rate is a good indicator for large k, as some of the early coefficients may be even zero. Since the complex F-series and trigonometric F-series are related, the decay rate of the trigonometric F-series coefficients is tied to the decay rate of the complex F-series coefficients. The derivative property of the F-series provides bounds on the F-series coefficients, referred to as the spectral bounds. For k 6¼ 0, n Z 1 0 T T Z 1 Tjkjn jo0 jn

1 jXs ðkÞj ¼ jko

ðnÞ jko0 t xT ðtÞe dt ðnÞ xT ðtÞ dt:

n ¼ 0 bound : jXs ½kj

Z1=2

1 2ðo0 Þ

0

(3:6:18a)

In deriving the right side of the above equation, the fundamental theorem of calculus is used and Z Z

yðtÞdt jyðtÞjdt note ejko0 t ¼ 1 : For example, the F-series coefficients in Example 3.4.2 show their decay rate is proportional to ð1=kÞ, see (3.4.8). The function xT ðtÞ has discontinuities at t ¼ t=2 in one period of the time function. Correspondingly, the bound on the F-series coefficient is (3:6:18b)

The number of nonzero bounds that can be determined equals the number of times the function can be differentiated before derivatives of impulses occur in the derivatives, see Ambardar (1995), Morrison, (1994), and others. Example 3.6.4 Find all the nonzero spectral bounds for the rectangular pulse xðtÞ ¼ P½t; xT ðtÞ ¼ xT ðt þ TÞ; T ¼ 2:

jxT ðtÞjdt ¼

1 : 2

(3:6:19a)

1=2

n ¼ 1 bound : Z1 Z1 dx2 ðtÞ 1 jXs ½kj dt dt ¼ 2p 2jo0 j1 1 1 1 1 1 dðt þ Þ þ dðt Þ dt ¼ : 2 2 p 1

The bounds above n ¼ 1 are not defined.

(3:6:19b) &

Bounds on the trigonometric Fourier series coefficients: If xT ðtÞ has discontinuities, then its trigonometric F-series coefficients satisfy for large k ja½kj5

T

jXs ½kj5B=k; B a constant:

Solution: The function, its first, and second derivatives are shown in Fig. 3.6.4. Noting that o0 ¼ ð2p=2Þ ¼ p, we have

ðKa1 and Kb1

Ka1 Kb1 and jb½kj5 k k are some constants):

(3:6:20)

If xT ðtÞ is continuous but x0T ðtÞ is discontinuous, then for large k Ka2 Kb2 and jb½kj5 2 ðKa2 and Kb2 2 k k are some constants): (3:6:21)

ja½kj5

Convergence is a function of the continuity of the highest derivative of xT ðtÞ. The convergence rates of the coefficients a½k and b½k may be different. Example 3.6.5 Consider the function and its F-series coefficients given below. Comment on the convergence rates. x2p ðtÞ ¼ et ; p5t5p; x2p ðt þ 2pÞ ¼ x2p ðtÞ: (3:6:22a) " # 1 2 sinhðpÞ 1 X ð1Þk þ x2p ðtÞ ¼ p 2 k¼1 ð1 þ k2 Þ ðcosðktÞ k sinðktÞÞ:

(3:6:22b)

Solution: The sine series coefficients b½k converge like ðKb =kÞ, whereas the cosine series coefficients a½k converge like ðKa =k2 Þ, implying that the Fourier series, as a whole, converges like ðK=kÞ; K 0 s

3.6 Operational Properties of Fourier Series

93

Fig. 3.6.4 (a) Periodic pulse waveform, (b) periodic impulse sequence, and (c) periodic doublet sequence

(a)

(b)

(c) are constants. The dc term and the first few harmonics contain bulk of the power for low-frequency & signals.

3.6.6 Integral of a Function and Its Fourier Series Consider the integral of a periodic function and its F-series. The integral of a periodic function with a dc component cannot be periodic as the integral of a constant is a ramp. In the case of Xs ½0 ¼ 0, we can derive the complex F-series of an integral of a function: # Zt Zt " X 1 jko0 a Xs ½ke da yT ðtÞ¼ xT ðaÞda¼ 0 1 X

¼

Ys ½ke

k¼1 1 X

¼

0 jko0 t

Ys ½k ¼

½Xs ½k=jko0 ; k 6¼ 0 constant; k ¼ 0

:

(3:6:23b)

Note the division by ko0 above indicating the F-series of the integral of a function xT ðtÞ converges faster than the F-series convergence of xT ðtÞ. Integration is a smoothing operation. Therefore, the integrated signal has much smaller high-frequency content than xT ðtÞ. Without the dc term Xs ½0, integration and differentiation can be thought of as inverse operations.

3.6.7 Modulation in Time

k¼1;k6¼0

Consider the F-series xT ðtÞejat ¼

fð1=jko0 ÞXs ½kgejko0 t þconstant:

k¼1;k6¼0

¼ (3:6:23a)

1 X k¼1 1 X k¼1

Xs ½kejko0 t ejat Xs ½kejðko0 aÞt :

(3:6:24)

94

3 Fourier Series

Multiplying a function by ejat shifts the frequencies from ko0 to ko0 a. This is modulation. Multiplying xT ðtÞ by cosðatÞ and using Euler’s formula results in y1 ðtÞ ¼ xT ðtÞ cosðatÞ ¼

1 1 X Xs ½kejðko0 þaÞt 2 k¼1

1 1 X þ Xs ½kejðko0 aÞt : 2 k¼1

(3:6:25)

Example 3.6.6 Consider the periodic signal xT ðtÞ ¼ cosðo0 tÞ modulated by the same function. The resulting function is yT ðtÞ ¼ cos2 ðo0 tÞ. Determine the frequency shifts.

1 X

Ws ½k nejðknÞo0 t : k¼1 " # 1 1 X X jno0 t jðknÞo0 t yT ðtÞ ¼ Xs ½ne Ws ½k ne wT ðtÞ ¼

n¼1

k¼1

¼

1 1 X X

Xs ½nWs ½k nejno0 t ejðknÞo0 t

k¼1 n¼1

¼ ¼

"

1 X

#

1 X

Xs ½nWs ½k n ejko0 t

n¼1

k¼1 1 X

Ys ½kejko0 t :

(3:6:27)

k¼1

Solution: yT ðtÞ ¼ :5ð1 þ cosð2o0 tÞÞ ¼ :25ej2o0 t þ 0ejo0 t þ :5 þ 0ejo0 t þ :25ej2o0 t : Modulation shifted the frequencies from o0 to 2o0 ; o0 ; 0, with one of the frequencies & having zero amplitude.

Equation (3.6.26b) now follows from (3.6.27). The expression Ys ½k in terms of Xs ½k and Ws ½k in (3.6.26b) is a discrete convolution. Periodic time convolution: If xT ðtÞ and wT ðtÞ are two periodic functions and their F-series are (3.6.26a), then (see Section 2.5) 1 T

yT ðtÞ ¼

3.6.8 Multiplication in Time

xT ðtÞ¼ yT ðtÞ¼

n¼1 1 X

Xs ½nejno0 t ; wT ðtÞ¼

1 X

Ws ½mejmo0 t ;

m¼1

Ys ½kejko0 t :

(3:6:26a)

k¼1

Following shows that the complex F-series coefficients of the product yT ðtÞ ¼ xT ðtÞwT ðtÞ are

Ys ½k ¼

1 X n¼1

Xs ½nWs ½k n¼

1 X

xT ðt tÞwT ðtÞdt

T

Let xT ðtÞ and wT ðtÞ be two periodic functions with the same period T. The product yT ðtÞ ¼ xT ðtÞwT ðtÞ is also periodic with period T. The relation between the F-series coefficients is derived below. First, let the F-series expansions of these are as follows: 1 X

Z

1 ¼ T

Z

¼

Ws ½nXs ½k n :

yT ðtÞ ¼ xT ðtÞ wT ðtÞ ¼ wT ðtÞ xT ðtÞ:

(3:6:28a) (3:6:28b)

The Fourier series expansion of the periodic convolution in (3.6.28a) can be determined by using the F-series for the two functions as shown below: Z 1 yT ðtÞ ¼ xT ðt tÞwT ðtÞdt T T

1 ¼ T ¼

Z

1 X k¼1

1 X k¼1

(3:6:26b) ¼ First, use the change of variable m ¼ k n in F-series expansion of wT ðtÞ, substitute this expression in the F-series expansion of yT ðtÞ, and then simplify the expression yT ðtÞ:

Xs ½kWs ½kejko0 t :

k¼1

T

n¼1

xT ðtÞwT ðt tÞdt

T 1 X

1 X

Xs ½kejko0 ðttÞ wT ðtÞdt 2

Xs ½k4

1 T

Z

3 wT ðtÞejko0 t dt5ejko0 t

T

Xs ½kWs ½kejko0 t :

(3:6:29a)

k¼1 FS;T

yT ðtÞ ¼ xT ðtÞ wT ðtÞ ! Xs ½kWs ½k ¼ Ys ½k:

(3:6:29b)

3.6 Operational Properties of Fourier Series

95

3.6.9 Frequency Modulation The dual of time-domain modulation is frequency modulation. Using the superposition and the delay properties of the F-series, we can determine the F-series coefficients of the function yT ðtÞ ¼ xT ðt þ aÞ þ xT ðt aÞ. It follows that yT ðtÞ ¼ ¼

1 X k¼1 1 X

2

1 X

¼

3 xT ðtÞejko0 t dt5

T

Xs ½kYs ½k:

(3:6:33)

k¼1

(3:6:30a)

k¼1

Px ¼

¼)Ys ½k ¼ 2Xs ½k cosðko0 aÞ:

1 X

Z

If yðtÞ ¼ xðtÞ (i.e., Ys ½k ¼ Xs ½k), then the average power in a complex or a real periodic function with period T is

Xs ½k ½ejko0 a þ ejko0 a ejko0 t Ys ½kejko0 t :

1 Ys ½k4 ¼ T k¼1

1 T

Z

jxT ðtÞj2 dt

T

(3:6:30b) ¼

1 X

jXs ½kj2 ðParseval 0 s formulaÞ:

(3:6:34)

k¼1

3.6.10 Central Ordinate Theorems

3.6.12 Power Spectral Analysis

The following results from the F-series at t ¼ 0 and the F-series coefficient at k ¼ 0:

The power density spectrum of the periodic signal xT ðtÞ is defined by ! 1 X 2 jko0 t xT ðtÞ ¼ Xs ½ke Sx ½k ¼ jXs ½kj (3:6:35) :

xT ½0 ¼

1 X

Xs ½k;

Xs ½0 ¼

k¼1

1 T

Z

xT ðtÞdt: (3:6:31)

k¼1

T

3.6.11 Plancherel’s Relation (or Theorem) Let xT ðtÞ and yT ðtÞ be two periodic functions. Then, Plancherel’s relation is 1 T

Z

xT ðtÞyT ðtÞdt ¼

1 X

Xs ½kYs ½k:

(3:6:32)

k¼1

T

Note, for generality, the expression in (3.6.32) is given for complex functions and the superscript (*) corresponds to the conjugation. The above relation can be derived by substituting the F-series coefficients for the two time functions: 1 T

Z T

1 xT ðtÞyT ðtÞdt ¼ T

Z T

" xT ðtÞ

1 X k¼1

The average power contained in xT ðtÞ is then given by Px ¼

dt

jXs ½kj2 ¼

k¼1

1 X

Sx ½k:

(3:6:36)

k¼1

Notes: In Chapter 1, it was pointed that periodic and random signals are power signals. Although we will not be going through any discussion on random signals or processes, as it is beyond our scope, the average power contained in a random process is expressed in terms of power spectral density, which is real, even, and nonnegative function of frequency ð f Þ, identified by Sx ð f Þ. The average power in the process is expressed by

# Ys ½kejko0 t

1 X

Px ¼

Z1 1

1 Sx ðfÞdf ¼ 2p

Z1 1

Sx ðoÞdo:

(3:6:37)

96

3 Fourier Series

It is interesting to tie the average power Px in (3.6.36) and (3.6.37). This is achieved by using impulse functions (note dðfÞ ¼ 2pdðoÞ, see (1.4.37) 1 X

Sx ðoÞ ¼ 2p

jXs ½kj2 dðo ko0 Þ:

(3:6:38)

k¼1

Using this expression results in (3.6.37) 1 2p

Z1

Sx ðoÞdo

1

1 ¼ 2p ¼

Z1 " 2p

1 1 X

Solution: The nonzero F-series coefficients are pro& portional to (1/k). Notes: Before considering the convergence of F-series let us briefly summarize the theoretical constraints on the periodic function xT ðtÞ of interest and its F-series existence. It is assumed that xT ðtÞ is square integrable. That is, Z (3:7:3) jxT ðtÞj2 dt51: T

1 X

# jXs ½kj2 dðoko0 Þ do

k¼1

jXs ðkÞj2 ¼Px :

(3:6:39)

Second, the periodic function is assumed to satisfy the Dirichlet conditions, see Section 3.1. All physically realizable functions satisfy these conditions and & therefore, we will not be dealing with these.

k¼1

3.7.1 Fourier’s Theorem 3.7 Convergence of the Fourier Series and the Gibbs Phenomenon In Chapter 1, the average value, the average power, and the root mean-squared (rms) values were defined (see (1.5.15)). It was pointed out that the average value of a periodic pﬃﬃﬃﬃﬃﬃ function Xs ½0can never exceed the rms value Px . Using (3.4.23), we have Px ¼ jXs ð0Þj2 þ

1 X

jXs ½kj2

Dirichlet proved first that Fourier series approximation converges to xT ðtÞ at every point xT ðtÞ is continuous and to ½xT ðtþ Þ þ xT ðt Þ=2, the halfvalue, wherever the function xT ðtÞ is discontinuous, i.e., the F-series converges to the average value of the function. This result is called Fourier’s theorem. Example 3.7.2 Let xT ðtÞ has a discontinuity at t ¼ t0 as shown in Fig. 3.7.1. The Fourier series approximation of xT ðtÞ with ð2n þ 1Þ terms is assumed to be

k¼1;k6¼0

¼)Px jXs ½0j2 or

pﬃﬃﬃﬃﬃﬃ Px jXs ½0j:

(3:7:1)

Furthermore, from (3.4.24a), the mean-squared value of a periodic function is equal to the sum of the mean-squared values of its dc component and its harmonics.

xT;2nþ1 ðtÞ ¼

n X

Xs ½kejko0 t :

(3:7:4)

k¼n

Example 3.7.1 Consider the periodic function given in (3.5.1) with T ¼ 2p. It is discontinuous at t ¼ kp and its F-series is given below (it follows from (3.5.2)). Identify how fast the F-series coefficients decrease. 4 1 1 x2p ðtÞ ¼ ½sinðtÞ þ sinð3tÞ þ sinð5tÞ þ . . . p 3 5 1 X 4 sinðð2 k 1ÞtÞ: (3:7:2) ¼ ð2 k 1Þp k¼1

Fig. 3.7.1 xT ðtÞ with a discontinuity

The mean-squared error at the discontinuity between xT ðtÞ and its ð2n þ 1Þ term F-series is

3.7 Convergence of the Fourier Series and the Gibbs Phenomenon

97

h i Solution: Consider the approximations by consid2 þ 2 e2 ¼ fxT;2nþ1 ðtÞxT ðt 0 Þg þfxT;2nþ1 ðtÞxT ðt0 Þg : ering the first few terms. Let (3:7:5) 4 4 1 þ The time entries in the above equation t s1 ðtÞ ¼ sinðtÞ; s3 ðtÞ ¼ ½sinðtÞ þ sinð3tÞ; 0 and t0 are p p 3 the values of t before and after the discontinuity at t ¼ t0 . Find the minimum value of the meansquared error, with respect to xTð2nþ1Þ ðtÞ by taking the partial derivative of e2 with respect to xT;2 nþ1 ðtÞ, equating the result to zero, and then solving for xT;2nþ1 ðtÞ at t ¼ t0 . Solution: Taking the partial derivative and equating it to zero at t ¼ t0 , we have @e2 ¼ 2½xT;2nþ1 ðtÞ xT ðt 0 Þ @xT;2nþ1 þ 2½xT;2nþ1 ðtÞ xT ðtþ 0 Þjt¼t0 ¼ 0: (3:7:6) ) xT;2nþ1 ðtÞjt¼t0 ¼ ½xT ðtþ 0 Þ þ xT ðt0 Þ=2:

(3:7:7)

That is, the F-series approximation gives the average value of the function before and after the discontinuity. This value is referred to as the half-value & of xT ðtÞ at the discontinuity t ¼ t0 .

4 1 1 s5 ðtÞ ¼ ½sinðtÞ þ sinð3tÞ þ sinð5tÞ;...: p 3 5

(3:7:8)

The functions s1 ðtÞ; s3 ðtÞ; and sð2 k1Þ ðtÞ; k large are sketched in Fig. 3.7.2 for one period. They are odd periodic functions. Fourier series approximation gives the value of 0 at t ¼ 0, the average value (or half-value) of the function equals to ð1 1Þ=2 ¼ 0. First, consider only the positive values of t; 05t5p. The maximum value of s1 ðtÞ is equal to ð4=pÞ ¼ 1:2732. This function crosses the value of 1 when ðs1 ðtÞ 1Þ ¼ 0 for positive t. The roots of this equation are t ¼ :9033 and t ¼ 2:2383 located symmetrically around the middle t ¼ 1:5708. More number of terms we consider, the better the approximation of the given function is, and in the limit, the integral-squared error goes to zero.

Summary on s2 k1 ðtÞðt > 0Þ :

3.7.2 Gibbs Phenomenon

The function rises rapidly as t goes from 0. It overshoots the value of 1 and oscillates

From Fourier’s theorem and the following discussion we will see that at both sides of a discontinuity the finite F-series approximation exhibits ripples before and after the discontinuity. This behavior is called the Gibbs phenomenon (or effect). Historically, Albert Michalson observed the phenomenon and reported to Josiah Gibbs, a theoretical physicist. Gibbs investigated this behavior of oscillations with overshoots and undershoots before and after the discontinuity associated with Fourier series. Equality of the function to its F-series is only in the sense the integralsquared error between the two goes to zero when infinite number of terms is included in the F-series approximation. Fourier’s theorem points out that at a point of discontinuity the series converges to the average value or the half-value given in (3.7.7).

about the line xðtÞ ¼ 1 with increasing frequency and decreasing amplitudes. Although the magnitude of the peak overshoots and undershoots before and after the discontinuity at t ¼ 0 diminish as k increases, there is a lower bound of 9% on the overshoots or undershoots even as k ! 1: Furthermore, the F-series converges to every point of xT ðtÞ that is continuous with rare exceptions. It is possible that the Fourier series of a continuous function to be divergent at some point. Kolmogoroff Zygmund (1955) has given a function whose Fourier series is everywhere divergent. At the point of discontinuity in xT ðtÞ, the series converges to the half-value of the function, i.e., the average value of the function before and after the discontinuity. Since sk ðtÞ is a periodic odd function, the & sketches follow for p5t50:

Example 3.7.3 Illustrate the convergence of the F-series expansion to the function given in Example 3.7.1 by using the first few terms in the series.

98

3 Fourier Series

Fig. 3.7.2 (a)s1 ðtÞ, (b)s3 ðtÞ, and (c)s2 k1 ðtÞ, k large

s1 (t ) =

4 sin(t ) π

s3 (t ) =

4 1 sin( t) + sin(3 t) π 3

s2 k −1 (t ) =

4 1 sin( t) + sin(3 t) +... π 3 +

Notes: Gibbs published his results on the phenomenon (or effect) in Nature Magazine in 1899. Fourier did not discuss the convergence of F-series in his paper. If N is small, the value of the overshoot may be different. The Gibbs effect occurs only for waveforms with jump discontinuities, see Carslaw & (1950). Example 3.7.4 Find a series expansion for p using the results in Example 3.7.1. Solution: At t ¼ ð1=2Þp the function is continuous and therefore the series converges to the actual value of the function. Substituting t ¼ p=2 in (3.7.2), and simplifying, the expression for p is

Notes: Beckmann (1971), in his book on A History of p, gives many interesting aspects associated with & respect to the constant p. Example 3.7.5 Find a series expansion for p2 using the function in Example 3.7.3. Solution: The average power is 1 P¼ 2p

Zp p

ð1Þ2 dt¼1¼

1 1X b2 ½2k1Þ 2 k¼1

1 16 1 1 ¼ 1þ þ þ : 2 p2 9 25

1 1 ¼)p ¼8 1þ þ þ : 9 25

4 1 1 1 1 þ þ ¼ 1 p 3 5 7 1 1 1 ¼)p ¼ 4 1 þ þ : 3 5 7

1 sin((2k − 1)t ) 2k– 1

2

(3:7:9)

The series converge at a rate of 4ð1=ð2 k 1ÞÞ cor& responding to the (2k1)th term.

(3:7:10) &

How many terms in the F-series are needed to have a‘‘good’’approximation of the given function? The answer can only be given for a particular application. The integral-squared error (ISE) between the periodic function xT ðtÞ and its Fourier series approximation is

3.7 Convergence of the Fourier Series and the Gibbs Phenomenon

" # Z Z 1 X 2 2 2 ISE¼ jxT ðtÞxa ðtÞj dt¼ xT ðtÞdtT jX½kj T

T

k¼1

" # Z 1 TX 2 2 2 2 ¼ xT ðtÞdt TXs ½0þ ða ½kþb ½kÞ ¼0: 2 k¼1 T

(3:7:11) The ISE goes to zero only if an infinite number of terms are used in the expansion, which is impractical. A goal is to approximate the function xT ðtÞ using N trigonometric F-series with a bound on the ISE. First, xT;N ðtÞ ¼ Xs ½0 þ

N X

99

N¼1:

Z 1 ðISEÞ1 ¼ y22p ðtÞdt2p X2s ½0þ a2 ½1þb2 ½1 2 2p

¼ 1:571 2p½ð1=pÞ2 þ ð1=2Þð1=2Þ2 ¼ :149:

(3:7:16)

The percentage error is (.149/1.571) = 9.5%. This implies that N ¼ 1 satisfies the requirements for Part a. Continuing this procedure, we have N¼2: ðISEÞ2 ¼1:5712p½ð1=pÞ2 þð1=2Þð1=2Þ2 þð1=2Þð4=9p2 Þ¼:0075:

a½k cosðko0 tÞ

(3:7:17)

k¼1

þ

N X

b½k sinðko0 tÞ:

(3:7:12)

k¼1

Find the smallest integer value of N that results within certain percentage of ISE. The ISE, keeping only the dc term and N harmonics, is ( ) Z N 1X 2 2 2 2 ða ½kþb ½kÞ : ðISEÞN ¼ xT ðtÞdtT Xs ½0þ 2 k¼1

In this case the percentage integral-squared error is (.0075/1.571)=.5% and N ¼ 2 satisfies the requirement for part b. The coefficients die out like ðK=k2 Þ; where K is a constant. Very few terms are needed to approximate the half-wave rectified signal. Unfortunately, there is no general formula to determine the number of harmonics needed for a & given set of specifications.

T

(3:7:13) Example 3.7.6 Consider the half-wave rectified periodic function in Example 3.6.2 with sinðtÞ; 0 t5p y2p ðtÞ ¼ ; 0; p t52p yðt þ 2pÞ ¼ yðtÞ:

(3:7:14)

The trigonometric F-series of this periodic function was given by (see (3.6.10)) y2p ðtÞ ¼

1 1 þ sinðtÞ p 2 2 cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : p 1ð3Þ 3ð5Þ 5ð7Þ (3:7:15)

Find the smallest N in (3.7.13) for the two cases: a:10% or less b: 2% or less, see Gibson (1993). Solution: First Zp Z 1 p 2 y2p ðtÞdt ¼ ð1 cosð2tÞÞdt ¼ ¼ 1:571: 2 2 2p

0

The ISE can be reduced by increasing N. It goes to 0 when N ! 1. On the other hand, the peak 9% overshoots and undershoots before and after the discontinuities discussed earlier cannot be reduced even if infinite number of terms is included in the F-series expansion. There is another measure, average error, which can be used in judging an approximation, which is not very attractive as the positive errors may cancel out with the negative errors. Squaring the error function accentuates the larger errors. Least-squares error measure gives a convenient and a simple way to calculate the parameters in the approximation. Overshoots and undershoots can be reduced by smoothing.

3.7.3 Spectral Window Smoothing The ripples generated by the Fourier series approximation of a discontinuous function are due to the abrupt change of the function before and after the discontinuity. The use of a taper, instead of a discontinuity at a transition, yields a smoother

100

3 Fourier Series

reconstruction from the basis functions. The windowed signal is defined by the periodic convolution (see Section 2.5) by yT ðtÞ ¼ xT ðtÞ wT ðtÞ Z 1 xT ðaÞwT ðt aÞda: ¼ T

Ws;H ½k¼

:0800;:1256;:2532;:4376;:6424;:8268;:9544;1; :9544;:8268;:6424;:4376;:2532;:1256;:0800

ðy2p ðtÞÞjN¼7 (3:7:18a)

T

:

4 1 ¼ ð:9544Þ sinðtÞ þ ð:6424Þ sinð3tÞ p 3 1 1 þ ð:2532Þsinð5tÞþ ð:0800Þsinð7tÞ 5 7 (3:7:22)

Considering (2N+1) complex F-series coefficients of yT ðtÞ (see (3.6.29a)), we have yT;N ðtÞ ¼

N X

Ws ½kXs ½kejko0 t :

(3:7:18b)

k¼N

The sequence Ws ½k is a window and its weights (or coefficients) typically decrease with increasing jkj. The rectangular and hamming window sequences are

The two functions x2p ðtÞ (identified as xðtÞ on the top figure) and y2p ðtÞ (identified as yðtÞ on the bottom figure) are plotted in Fig. 3.7.3 using MATLAB. Note the overshoots and undershoots before and after the discontinuities in each case. The later case has hardly any ripples. The slope in the transition region is much higher in the case of the rectangular window com& pared to the Hamming window case.

3.8 Fourier Series Expansion of Periodic Functions with Ws;H ½k ¼ :54 þ :46 cosðkp=NÞ; Special Symmetries N k N: (3:7:20)

Ws;R ½k ¼ 1; N k N:

(3:7:19)

The use of special windows reduces or even eliminates the overshoots and undershoots in the approximated signal before and after a discontinuity, see Ambardar (1995). Example 3.7.7 Consider the trigonometric F-series in (3.7.2). Give the expression using the rectangular window with N ¼ 7. Illustrate the window smoothing by first sketching the F-series and then the Hamming windowed series. Solution: The trigonometric F-series approximation is 4 1 1 1 ðx2p ðtÞÞjN¼7 ¼ sinðtÞþ sinð3tÞþ sinð5tÞþ sinð7tÞ : p 3 5 7 (3:7:21)

Note the odd harmonic terms are all zero. The 15 tapered Hamming window coefficients and the Hamming windowed function y2p ðtÞ are, respectively, given by

In Section 1.6, periodic functions with half-wave and quarter-wave symmetries were considered. Computation of the F-series for these cases is considered next.

3.8.1 Half-Wave Symmetry Figure 3.8.1 illustrates a periodic function with period T and with half-wave symmetry (or rotation symmetry). Such functions satisfy (see (1.6.19) and (1.6.20))

T : (3:8:1) xT ðtÞ ¼ xT t 2 Periodic functions with half-wave symmetry have odd harmonics, i.e., Xs ½k ¼ 0;k even, which can be seen from the following. First Xs ½k ¼

1 T

Z T

xT ðtÞejko0 t dt

3.8 Fourier Series Expansion of Periodic Functions with Special Symmetries Fig. 3.7.3 Window smoothing

101 Use of rectangular window

2

x(t)

1 0 –1 –2

0

1

2

0

1

2

3

4 5 t Use of Hamming window

6

7

8

6

7

8

2

y(t)

1 0 –1 –2

3

1 T

4 t

5

ZT=2

T xT ða Þejko0 a ejko0 T=2 da 2 3 2 0 ZT=2 T 7 61 ¼4 xT ða Þejko0 a da5ejkp T 2 0

¼ð1Þ

k1

Z0

1 xT ðtÞejko0 t dtþ T

ZT=2

(3:8:4)

Using (3.8.4) in (3.8.3), we have xT ðtÞejko0 t dt: (3:8:2)

0

T=2

T

T xT ðt Þejko0 t dt: 2

0

Fig. 3.8.1 A half-wave symmetric function

1 ¼ T

ZT

1 Xs ½k ¼ T

ZT=2

T ejko0 t dt: xðtÞ þ ð1Þk x t 2

0

Consider the change of variable from t to a ðT=2Þ in the first integral on the right in (3.8.2), which results in 1 Xs ½k ¼ T

ZT=2

T T xT ða Þejko0 ða 2 Þ da 2

0

1 þ T

ZT=2

xT ðtÞejko0 t dt:

From the half-wave symmetry property in (3.8.1) we see that Xs ½k ¼ 0, k even, thus establishing the halfwave symmetric functions that contain only odd harmonics. The complex F-series of these functions and the corresponding trigonometric F-series are

(3:8:3)

0

The limits follow from and t ¼ T=2¼)a ¼ 0 and

(3:8:5)

t ¼ 0¼)a ¼ T=2

xT ðtÞ¼

1 X k¼1 k6¼0;kodd

Xs ½ke

jko0 t

2 ;Xs ½k¼ T

ZT=2

xT ðtÞejko0 t dt:

0

(3:8:6)

102

3 Fourier Series

xT ðtÞ ¼

1 X

fa½2 k 1 cosðð2 k 1Þo0 tÞ þ b½2 k 1

k¼1

ZT=4

4 ¼ T

0

sinðð2 k 1Þo0 tÞg; 4 a½2 k 1 ¼ T

ZT=2

þ xT ðtÞ cosðð2 k 1Þo0 tÞdt

0

4 b½2 k 1 ¼ T

ZT=2

xT ðtÞ cosðð2 k 1Þo0 tÞdt Z0

4 T

xT ðtÞ cos½ð2 k 1Þo0 tdt:

T=4

:

Trigonometric and complex F-series for the even quarter-wave symmetric function:

xT ðtÞ sinðð2 k 1Þo0 tÞdt

0

(3:8:7)

xT ðtÞ ¼

1 X

a½2 k 1 cos½ð2 k 1Þo0 t;

n¼1

8 a½2 k 1 ¼ T

3.8.2 Quarter-Wave Symmetry

ZT=4

xT ðtÞcosðð2 k 1Þo0 tÞdt:

0

If xT ðtÞ is a periodic function with half-wave symmetry and, in addition, is either even or an odd function, then xðtÞ is said to have even or odd quarter-wavesymmetry, respectively, see Section 1.6.4. They satisfy the following properties:

(3:8:10) 1 X

¼

k¼1

þ

1 X

Even quarter-wave symmetry : xT ðtÞ ¼ xT ðtÞ and xT ðtÞ ¼ xT ðt T=2Þ:

ð1=2Þa½2 k 1ejð2 k1Þo0 t

(3:8:8a)

¼

ð1=2Þa½2 k 1ejð2 k1Þo0 t

k¼1 1 X

0t Xs ½kjko e

k¼1

Odd quarter-wave symmetry : xT ðtÞ ¼ xT ðtÞ and xT ðtÞ ¼ xT ðt T=2Þ:

(3:8:8b)

¼)Xs ½2 k 1 ¼ Xs ½ð2 k 1Þ ¼ a½2 k 1=2:

(3:8:11)

3.8.3 Even Quarter-Wave Symmetry

3.8.4 Odd Quarter-Wave Symmetry

Since the function must be a half-wave symmetric to be a quarter-wave symmetric, it follows that Xs ½0 ¼ 0 and a½2 k ¼ 0: In addition, xðtÞ is even and b½k ¼ 0. Therefore,

The F-series for this case are as follows. Derivation is left as an exercise. 1 X b½2 k 1 sin½ð2 k 1Þo0 t; xT ðtÞ ¼

Xs ½0 ¼ 0; b½k ¼ 0; a½2 k 1 ¼

4 T

ZT=2

xT ðtÞ cosðð2 k 1Þo0 tÞdt; a½2 k ¼ 0:

ZT=4

(3:8:9a) xT ðtÞcosðð2k1Þo0 tÞdt

xT ðtÞ¼

T=4

xT ðtÞ sin½ð2 k 1Þo0 tdt: (3:8:12) 0

1 X

X½2k1ejð2k1Þo0 t

k¼1 1 X

þ

X½ð2k1Þejð2k1Þo0 t :

(3:8:13)

k¼1

0

4 þ T

8 T

b½2 k 1 ¼

0

4 a½2k1¼ T

k¼1

Z

T=2 Z

T=4

xT ðtÞcosðð2k1Þo0 tÞdt: (3:8:9b)

Compare this with the F-series expansion and equate the corresponding coefficients.

3.9 Half-Range Series Expansions

xT ðtÞ ¼

1 X

103

Xs ½nejno0 t

n¼1; n6¼0

¼)X½2 k 1 ¼

1 X

xeT ðtÞ ¼

xe ðt þ kTÞ and

k¼1

b½2 k 1 ; X½ð2 k 1Þ 2j

1 X

x0T ðtÞ ¼

x0 ðt þ kTÞ:

(3:9:2)

k¼1

b½2 k 1 ¼ : 2j

(3:8:14) The trigonometric F-series of the even periodic function has dc and cosine terms and the odd periodic function has only sine terms. The two periodic functions xeT ðtÞ and x0T ðtÞ can be expressed by the following with o0 ¼ 2p=T (see (3.4.18)):

3.8.5 Hidden Symmetry Example 3.8.1 Symmetry of a periodic function can be obscured by a constant. Consider the periodic saw-tooth waveform, x2p ðtÞ ¼ ð1 ðt=2pÞÞ; 05t52p in Fig. 3.8.2. It does not have any obvious symmetry. Find the F-series of the function x2p ðtÞ by noting ðx2p ðtÞ ð1=2ÞÞ is an odd function.

xeT ðtÞ ¼ Xs ½0 þ

1 X

a½k cosðko0 tÞ:

(3:9:3a)

xðtÞ cosðko0 tÞdt:

(3:9:3b)

k¼1

2 Xs ½0 ¼ T

ZT=2 xðtÞdt; 0

2 a½k ¼ ðT=2Þ

ZT=2 0

x0T ðtÞ ¼

1 X

b½k sinðko0 tÞ;

k¼1

Fig. 3.8.2 xT ðtÞ with hidden symmetry, T = 2p

4 b½k ¼ T

ZT=2 xðtÞ sinðko0 tÞdt:

(3:9:4)

0

Solution: From the odd symmetry,

X 1 1 x2p ðtÞ b½k sinðko0 tÞ ¼ 2 k¼1 1 1 X b½k sinðko0 tÞ : ¼)x2p ðtÞ ¼ þ 2 k¼1

(3:8:15) &

3.9 Half-Range Series Expansions Consider an aperiodic function x(t) over the interval ð0; T=2Þ and zero everywhere else. Even and odd functions can be generated in the interval T=25t5T=2 by xe ðtÞ ¼ xðtÞ þ xðtÞ; x0 ðtÞ ¼ xðtÞ xðtÞ: (3:9:1) Even and odd periodic extensions (see Section 1.8.1.) of these are

The functions xeT ðtÞ and x0T ðtÞ in (3.9.3a and b) and (3.9.4) represent the same function in the interval (0, T/2). Outside this interval (3.9.3a) represents an even periodic function and (3.9.4) represents an odd periodic function. These expansions are called the half-range Fourier series expansions of the aperiodic function xðtÞ. Example 3.9.1 Given the aperiodic function xðtÞ ¼ sinðtÞ; 05t5p and 0 otherwise, expand this function in terms of a cosine series expansion and the sine series expansion in the interval 05t5p. Give the even and odd periodic extensions of xðtÞ. Solution: It is simple to see that xeT ðtÞ ¼ jsinðtÞj and the odd periodic extension x0T ðtÞ ¼ sinðtÞ. In the interval 05t5p, the two functions xeT ðtÞ and x0T ðtÞ are equal. Since jsinðtÞj and sinðtÞ are continuous functions, the F-series converges and (see (3.6.7))

104

3 Fourier Series

xeT ðtÞ ¼ jsinðtÞj 2 4 cosð2tÞ cosð4tÞ þ þ ; 05 t 5 p : ¼ p p ð1Þð3Þ ð3Þð5Þ (3:9:5) x0T ðtÞ ¼ sinðtÞ; 05t5p:

(3:9:6) &

Example 3.9.2 Expand the function given below in terms of a cosine series expansion in the interval 05t5p. Give the even periodic extensions of xðtÞ. xðtÞ ¼

0; 05t5p=2 : 1; p=25t5p

Zp

p=4

xeT ðtÞ cosðktÞdt

0

¼

2 p

Zp

cosðktÞdt ¼

sinðktÞdt p=2

¼

2 ½cosðkpÞ cosðð1=2ÞkpÞ: kp

2 1 1 sinðtÞ þ sinð3tÞ þ sinð5tÞ þ x0 ðtÞ ¼ p 3 5 2 1 1 sinð2tÞ þ sinð6tÞ þ sinð10tÞ þ : p 3 5 &

(3:9:7)

2 1 ¼ ½ðp=2Þ ðp=4Þdt ¼ : p 2 2 a½k ¼ p

Zp

2 p

05t5p:

Solution: Cosine series need a symmetric function defined by xe ðtÞ ¼ xðtÞ þ xðtÞ. The F-series can now be obtained by considering the interval p5t5p, i.e., the period T ¼ 2p. Noting that o0 ¼ 2p=T ¼ 1, the trigonometric F-series can be computed. Since it is even, it follows that b½k ¼ 0. The coefficients Xs ½0 and a½k are, respectively, given by Zp Zp=2 1 2 xeT ðtÞdt ¼ dt X½0 ¼ 2p p p

b½k ¼

2 sinðkp=2Þ: kp

p=2

1 2 1 1 xeT ðtÞ ¼ cosðtÞ cosð3tÞ þ cosð5tÞ ; 2 p 3 5 05t5p: This gives an approximation of xðtÞ in terms of cosine series in the time interval 05t5p. In a similar manner, the odd periodic extension of the function can be determined. Since the function is odd, it follows that Xs ½0 ¼ 0 and a½k ¼ 0; k ¼ 0; 1; 2; . . .. The coefficients b½k and the corresponding odd period extension are given by

3.10 Fourier Series Tables Refer tables 3.10.1 and 3.10.2 for Fourier Series.

3.11 Summary In this chapter we have introduced some of the basis functions and their use in approximating a given function in an interval. The important set of basis functions are the sine and the cosine functions and the periodic exponential function leading to the discussion on Fourier series. This chapter dealt with Fourier series, their properties and the computations of the Fourier series coefficients, in general, and in cases of special symmetries in the given function. Convergence of the coefficients and the number of coefficients required for a given set of specifications are discussed. Approximation measures are discussed in terms of basis functions. Specific principal topics that were included are

Various basis functions and error measures of a function and their approximations

Basics of complex and trigonometric Fourier series and the relationships between the trigonometric and the complex F-series Computation and simplification of the F-series coefficients of periodic functions that have simple and special symmetries Operational properties of the Fourier series that include simple methods that allow for simplification in the computation of the F-series

3.11 Summary

105 Table 3.10.1 Symmetries of real periodic functions and their Fourier-series coefficients

Type of symmetry

Constraints Periodic, xT ðtÞ ¼ xT ðt þ TÞ; o0 ¼ 2p=T:

Trignometric Fourier-series

Even

xT ðtÞ ¼ xT ðtÞ

xT ðtÞ ¼ XS ½0 þ

1 P

Fourier series coefficients

a½k cosðko0 tÞ

XS ½0 ¼ T2

k¼1

a½k ¼ T4 xT ðtÞ ¼ xT ðtÞ

Odd

x T ð tÞ ¼

1 P

b½k ¼ T4

b½k sinðko0 tÞ

k¼1

xT ðtÞ ¼ xT ðt þ T=2Þ

Half-wave

xðtÞ ¼

1 P k¼1

þ

b½2 k 1 sinð2 k 1Þo0 t

k¼1

Even quarterwave

xT ðtÞ ¼ xT ðtÞ; xT ðtÞ ¼ xT ðt þ T=2Þ

x T ð tÞ ¼

1 P

a½2 k 1 cosð2 k 1Þo0 t

k¼1

0 T=2 R

0 T=2 R

xT ðtÞdt

xT ðtÞ cosðko0 tÞdt

xT ðtÞ sinðko0 tÞdt

0

a½2 k 1 ¼ T4

a½2 k 1 cos½2 k 1o0 t

1 P

T=2 R

b½2 k 1 ¼ T4 a½2 k 1 ¼ T8

T=2 R 0 T=2 R 0 T=4 R

xT ðtÞcosðð2 k 1Þo0 tÞdt xT ðtÞ sinðð2 k 1Þo0 tÞdt xT ðtÞcosðð2 k 1Þo0 tÞdt

0

T=4 1 R P b½2 k 1 sinð2 k 1Þo0 t Odd xT ðtÞ ¼ xT ðtÞ; x T ð tÞ ¼ b½2 k 1 ¼ T8 xT ðtÞsinðð2 k 1Þo0 tÞdt k¼1 0 quarterxT ðtÞ ¼ xT ðt þ T=2Þ wave There are extensive tables in literature that list Fourier series of functions. See Abramowitz and Stegun (1964), Gradshteyn and Ryzhik (1980), and others. To standardize the tables, we will assume the period is T ¼ 2p. Spiegel (1968) has several other interesting periodic functions and the corresponding Fourier series.

Table 3.10.2 Periodic functions and their Trigonometric Fourier Series Periodic Function x2p ðtÞ ¼ x2p ðt þ 2pÞ Trigonometric Fourier-series h i 5 5 sinð3tÞ sinð5tÞ 1; 0 t p 4 sinðtÞ þ þ þ x2p ðtÞ ¼ p 1 3 5 1; p5t50 x2p ðtÞ ¼ jtj ¼

t; t;

05t5p p5t50

2

x2p ðtÞ ¼ t2 ; p5t5p x2p ðtÞ ¼ jsinðtÞj; p5t5p sinðtÞ; 05t5p 0; p5t52p x2p ðtÞ ¼

cosðtÞ; cosðtÞ;

x2p ðtÞ ¼ et ; p5t5p

p4

05t5p p5t50

h

h

cosðtÞ 12

þ cos3ð23tÞ þ cos5ð25tÞ

i

i sin2ð2tÞ þ sin3ð3tÞ þ h i cosðtÞ cosð2tÞ cosð3tÞ p2 4 þ 2 2 3 1 2 3

x2p ðtÞ ¼ t; p5t5p

x2p ðtÞ ¼

p 2

sinðtÞ 1

h

cosð2tÞ ð1Þð3Þ

ð4tÞ cosð6tÞ þ cos ð3Þð5Þ þ ð5Þð7Þ þ

2 p

p4

1 p

þ 12 sinðtÞ p2 h

8 sinð2tÞ p ð1Þð3Þ

h

cosð2tÞ ð1Þð3Þ

i

ð4tÞ cosð6tÞ þ cos ð3Þð5Þ þ ð5Þð7Þ þ

i ð4tÞ 3 sinð6tÞ þ 2ðsin þ þ 3Þð5Þ ð5Þð7Þ

2 sinh p 1 p 2

þ

1 P k¼1

ð1Þk ðcosðktÞ ð1þk2 Þ

k sinðktÞÞ

i

106

3 Fourier Series

Bounds and convergence of the F-series to the given function Half-range expansions

Problems 3.1.1 The set of functions fi ðtÞ; i ¼ 1; 2; 3; 4 shown in Fig. P3.1.1 are a set of Walsh functions. Show that they are orthogonal in the interval [0, 1]. 3.1.2 Consider the set ff1 ðtÞ; f2 ðtÞg with f1 ðtÞ ¼ 1 and f2 ðtÞ ¼ cð1 2tÞ. Is this an orthogonal set in the interval [0, 1]? If so, compute the value of c that makes the functions f1 ðtÞ and f2 ðtÞ become an orthonormal basis set. 3.1.3 The function xðtÞ ¼ sinðtÞ is approximated by xðtÞ ¼ c1 f1 ðtÞ þ c2 f2 ðtÞ. Use the results in Problem 3.1.2 and find the constants c1 and c2 so that the

1 2

1 4

3 4

t

1

-1

t 1 2

3 4

1

(b) φ3 (t ) 1 0 -1

1 2

1 t 3 4

1 4

(c)

-1

3.2.1 Use the Walsh functions given in Problem 3.1.1 to approximate the following function and find the mean-squared error between the given function and approximation: xðtÞ ¼

t;

05 t 5 1

0;

otherwise

:

The Lp error between the two sides of the above equation is Ep ¼ jð1 aÞjp þjð2 aÞjp . If p ¼ 2; we call that as a least-squares error or L2 error and for p ¼ 1, we call that as the L1 error. Find the value of a that minimizes the L1 and L2 errors. The leastsquares error can be computed by taking the partial of E2 with respect to a, equating it to zero and solving for a, see Section A.8.1. A simple way to solve the L 1 error problem is solve each equation and find the value of a out of the two solutions that gives the minimum L 1 error. For iterative Lp solutions, see Yarlagadda, Bednar and Watt (1986). 3.3.1 Determine the complex Fourier series of the function

φ4 (t ) 1 0

0; 15t50 : 1; 05t51

1 ¼ a: 2 1

φ2 (t ) 1 1 4

xðtÞ ¼

1

(a)

0

3.1.5 Use the first five Legendre polynomials to approximate the following function:

3.2.2 Consider the equations given in matrix form given below. There is no value of a that satisfies the set of equations given below. The system is called an overdetermined system of equations:

φ 1 (t ) 1 0

mean-squared error is minimized between the given function and the approximated function. pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 3.1.4 Show the set fk ðtÞ ¼ ð2=TÞ sinðkp=TÞt; k ¼ 1; 2; 3; . . . is an orthogonal set in the interval (0, T).

1 4

1 2

(d) Fig. P3.1.1 Walsh functions

3 4

1

t

xT ðtÞ ¼ cosð2pf0 t 1Þ þ sinð2pð2f0 Þt 2Þ: 3.3.2 a. Determine the period of the function 1 X 1 xT ðtÞ ¼ ejð3pkt=2Þ : ða þ jpkbÞ k¼1

Problems

107

b. What is the average value of the periodic function xT ðtÞ? c. Determine the amplitude and phase values of the third harmonic.

Derive an expression for p using the F-series. See Example 3.7.4. What is the value of the F-series at t ¼ 0? Compare this to the actual value of the function at that location.

3.3.3 Expand the following periodic functions with xi ðtÞ ¼ xi ðt þ 2pÞ in trigonometric F-series:

3.5.2 Using the Fourier series expansion, determine the sum A identified below

a: x1 ðtÞ ¼ eat ; p5t5p; a is some constant

b: x2 ðtÞ ¼ cosh at; p t p: c: x3 ðtÞ ¼ tðp tÞ; 05t5p; cosðtÞ; 05t5p : d: x4 ðtÞ ¼ cosðtÞ; p5t50

xðtÞ ¼ cosh at ¼ p2 sinhðapÞ 1 P k a 1 þ ð1Þ cosðktÞ ; p t p 2 2 2a k þa k¼1

1 þ A ¼ 2a

1 P k¼1

:

1 ð1Þk k2 þa 2

function

3.5.3 Find the F-series expansion of the function yðtÞ ¼ sinhðatÞ using the above results.

3.3.5 Show that the equation in (3.3.11) reduces to the equation in (3.3.10). 3.4.1 Use the derivative method to determine the trigonometric Fourier series of the periodic function x2p ðtÞ ¼ t; 0 t52p.

3.6.1 Consider the full-wave rectified function xðtÞ ¼ jsinðtÞj; p5t5p. Estimate the rms value of the full-wave rectified signal by using the first four nonzero terms in the Fourier series representation of the function. Calculate the percentage of error in the estimation.

3.4.2 Find the trigonometric F-series of the function shown below using a. derivative method. b. What can you say about the convergence of its F-series coefficients? 2 cos½pt=2; 15t51 xT ðtÞ ¼ : 0; 15jtj5T=2

3.6.2 Consider the triangular wave function given below and derive the trigonometric F-series of this function. ( T5 t0 1 þ 4t T ; 2 xT ðtÞ ¼ ; xT ðt þ TÞ ¼ xT ðtÞ: 4t 1 T ; 0 t5 T2

c. What can you say about the convergence of the trigonometric F-series coefficients?

3.6.3 Show that

3.3.4 Find the F-series of x1 ðtÞ ¼ At2 þ Bt þ C; p5t5p.

the

3.4.3 Use the generalized derivatives of the function given in Problem3.4.2 and see how fast the series converge without actually computing the Fourier series and verify the results obtained in that problem. 3.4.4 Show that jXs ½kj ¼ jXs ½kj for a real periodic function xT ðtÞ. Give a complex periodic function where this is not true. 3.5.1 The function and its trigonometric F-series expansion are given by x2p ðtÞ ¼ jsinðtÞj; 0 t52p; 2 4 jsinðtÞj ¼ p p cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : 3 15 35

1 T 1 T

Z

Z

xT ðtÞxT ðt tÞdt ¼

T

1 xTe ðtÞxTe ðt tÞdt þ T

T

Z

xT0 ðtÞxT0 ðt tÞdt:

T

3.6.4 Show the integral of xT ðtÞ with a nonzero average value is non-periodic. 3.6.5 Show the derivative of an even function is an odd function and the derivative of an odd function is an even function. Verify this property using the trigonometric F-series. 3.7.1 Using the F-series in Table 3.10.2, determine 1 X A¼ ½1=ð2 k 1Þ2 : k¼1

108

3 Fourier Series

3.7.2 Verify the Fourier series of the periodic function x2p ðtÞ ¼ t; p5t5p are

g2p ðtÞ ¼

t ðp=2Þ; t ðp=2Þ;

05 t 5 p ; p5t50

g2p ðtÞ ¼ g2p ðt þ 2pÞ: x2p ðtÞ ¼ 2

1 X

½ð1Þk1 =k sinðktÞ:

k¼1

What value does this function converges to at t ¼ 0; ðp=2Þ; p?

3.8.4 Give the complex F-series of the full-wave rectified signal x2p ðtÞ ¼ jcosðtÞj. Use the Fourier series of the results on the full-wave rectified sine wave in Table 3.10.2.

3.7.3 Give the convergence rate of the F-series coefficients of the periodic functions without actually using the F-series.

3.9.1 Verify the trigonometric Fourier series given in Table 3.10.2 for the following periodic functions in the range p5t5p.

a:x1 ðtÞ ¼ t 1; 05t51; x1 ðt þ TÞ ¼ x1 ðtÞ; sinðtÞ; 05t5p : T ¼ 1; b:x2p ðtÞ ¼ 0; p5t52p

a: x2p ðtÞ ¼ t2 ; b: x2p ðtÞ ¼ et ; c: x2p ðtÞ ¼ cosðatÞ:

3.8.1 Derive the expressions for the coefficients b½k in (3.8.13). 3.8.2 Consider the periodic function x2p ðtÞ below and identify any symmetry this has. x2p ðtÞ ¼

1; 1;

05 t 5 p ; x2p ðtÞ ¼ x2p ðt þ 2pÞ: p5t50

3.9.2 Use the results in Table 3.10.2 and the derivative method to derive the F-series of the periodic function x2p ðtÞ ¼ sinðatÞ; p5t5p. 3.9.3 Derive an expression for the F-series of the function x2p ðt pÞ; p5t5p given

x2p ðtÞ ¼ Xs ½0 þ

1 X

a½k cosðko0 tÞ

k¼1

þ

1 X

b½k sinðko0 tÞ; p5t5p:

k¼1

a. Use that to derive the trigonometric F-series of this function. Give the corresponding complex F-series. b. Use the derivative method to derive the trigonometric F-series. 3.8.3 Consider the periodic triangular wave function below. Does this function have any symmetry? Derive the trigonometric F-series of this function.

3.9.4 Consider the functions xðtÞ ¼ cosðtÞ and yðtÞ ¼ sinðtÞ over the interval 05t5p and 0 otherwise. Expand these functions using the Fourier sine series and cosine series by directly going through the procedure discussed in Section 3.9. Can you think of a simpler method knowing the results given in Example 3.9.1?

Chapter 4

Fourier Transform Analysis

4.1 Introduction In Chapter 3 we have discussed the frequency representation of a periodic signal. Fourier series expansions of periodic signals give us a basic understanding how to deal with signals in general. Since most signals we deal with are aperiodic energy signals, we will study these in terms of their Fourier transforms in this chapter. Fourier transforms can be derived from the Fourier series by considering the period of the periodic function going to infinity. Fourier transform theory is basic in the study of signal analysis, communication theory, and, in general, the design of systems. Fourier transforms are more general than Fourier series in some sense. Even periodic signals can be described using Fourier transforms. Most of the material in this chapter is standard (see Carlson, (1975), Lathi, (1983), Papoulis, (1962), Morrison, (1994), Ziemer and Tranter, (2002), Haykin and Van Veen, (1999), Simpson and Houts, (1971), Baher, (1990), Poulariskas and Seely, (1991), Hsu, (1967, 1993), Roberts, (2007), and others).

The frequency f0 ¼ o0 =2p is the fundamental frequency of the signal, which is the inverse of the period of the signal f0 ¼ ð1=TÞ. The Fourier series coefficients are complex in general. To make the analysis simple we assume the signal under consideration is real and the amplitude of the Fourier coefficients is given by jXs ½kj. Figure 4.2.1b gives the sketch of the amplitude line spectra of the complex Fourier series of a periodic function shown in Fig. 4.2.1a. The frequencies are located at ko0 ¼ 2pðkf0 Þ; k ¼ 0; 1; 2; : : : and the frequency interval between the adjacent line spectra is o0 ¼ 2pf0 ¼ 2p=T: In this example, we assumed t=T ¼ 1=5. As T ¼ 2p=o0 ! 1, o0 goes to zero and the spectral lines merge. To quantify this, let

xT (t)

4.2 Fourier Series to Fourier Integral Consider a periodic signal xT ðtÞ with period T and its complex Fourier series xT ðtÞ ¼

1 X

X s [k ]

Xs ½ke jko0 t ;

k¼1

Xs ½k ¼

1 T

ZT=2

xT ðtÞejko0 t dt;

o0 ¼ 2p=T:

T=2

(4:2:1)

Fig. 4.2.1 (a) xT ðtÞ and (b) jXs ðkÞj

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_4, Ó Springer ScienceþBusiness Media, LLC 2010

109

110

4 Fourier Transform Analysis

us extract one period of the periodic signal by defining xðtÞ ¼

xT ðtÞ;

T2 5t5 T2

0;

Otherwise

:

(4:2:2)

"

1 1 X Xðko0 Þe jðko0 Þt o0 xðtÞ ¼ lim T!1 2p k¼1

1 ¼ 2p

Z1

XðjoÞejot do:

#

(4:2:7)

1

We can consider the function in (4.2.2) as a periodic signal with period equal to 1. In the expression for the Fourier series coefficients in (4.2.1), k and o0 appear as product ðko0 Þ ¼ ðkð2pÞ=TÞ. As T ! 1, the expression for the Fourier series coefficients in (4.2.1) results in a value equal to zero, which does not provide any spectral information of the signal. To avoid this problem, define

We now have the Fourier transform of the time function xðtÞ, XðjoÞ; and its inverse Fourier transform. Some authors use XðjoÞ instead of XðoÞ for the Fourier transform, indicating the transform is a function of complex variable ðjoÞ. The pair of functions xðtÞ and XðjoÞ is referred to as a Fourier transform pair:

XðjoÞ ¼ F½xðtÞ ¼ ZT=2

Xðko0 Þ ¼ TXs ½k ¼

(4:2:3) xðtÞ ﬃ x~ðtÞ ¼ F

Note that k is an integer and it can take any integer value from 1 to 1. Furthermore, as T ! 1, o0 ¼ ð2p=TÞ becomes an incremental value, ko0 becomes a continuous variable o on the frequency axis, and Xðko0 Þ becomes XðoÞ. From this, we have the analysis equation for our single pulse:

T!1

ZT=2 xðtÞe

T!1

jko0 t

dt

T=2

¼

Z1

Xs ½ke jko0 t ;

(4:2:4)

T=25t5T=2; (4:2:5)

k¼1

xT ðtÞ ¼

1

1 ½Xð joÞ ¼ 2p

Z1

XðjoÞe jot do; (4:2:8b)

FT

(4:2:8c)

xðtÞ !Xð joÞ:

The transform and its inverse transform can be written in terms of frequency variable f in Hertz instead of o ¼ 2pf: Xð jf Þ ¼

Z1 xðtÞe

j2pft

dt;

xðtÞ ¼

Z1

Xð jf Þej2pft df:

1

(4:2:8d)

xðtÞejot dt:

Now consider the synthesis equation in terms of the Fourier series in the forms 1 X

(4:2:8a)

1

1

1

xT ðtÞ ¼

xðtÞejot dt;

1

xðtÞejko0 t dt:

T=2

XðjoÞ ¼ lim ðTXs ½kÞ ¼ lim

Z1

1 1 X Xðko0 Þe jko0 t : T k¼1

(4:2:6)

One can see that as T ! 1, ko0 ! o; a continuous variable, and o0 ¼ 2p=T ! do, an incremental value and the summation becomes an integral. These result in

Equation (4.2.8c) shows that the transform and its inverse have the same general form, one has the time function and an exponential term with negative exponent and the other has the transform and an exponential term with a positive exponent in the corresponding integrands. F-transforms are applicable for both real and complex functions. Most practical signals are real signals. The transforms are generally complex. Integrals in the transform and its inverse are with respect to a real variable. The relations between the Laplace transforms, considered in Chapter 5, and the Fourier transforms become evident with this form. The form in (4.2.8d) is adopted by the engineers in the communications area. The transforms are computed by integration and the inverse transforms are determined by using

4.2 Fourier Series to Fourier Integral

111

transform tables. Since the Fourier transform is derived from the Fourier series, we can now say that

xT ðtÞ ¼ xðtÞ; jtj5t=2; xT ðtÞ ¼

1 X

xðt þ nTÞ:

n¼1 FT

Inverse FT

xT ðtÞ ¼ AP½t=t; jtj5T=2; xT ðtÞ ¼ xT ðt þ TÞ; (4:2:10) yT ðtÞ ¼ xT ðt ðt=2ÞÞ:

xðtÞ ! XðjoÞ ! x~ðtÞ ’ xðtÞ:

The inverse transform of XðjoÞ; F1 ½XðjoÞ identified by x~ðtÞ is an approximation of xðtÞ and it may not be the same as the function xðtÞ. We will have Solution: The complex Fourier series coefficients of xT ðtÞ were given in Example 3.4.2. The Fourier more on this shortly. series coefficients and their amplitudes of the two Fourier transform is applicable to signals that functions are given by obey the Dirichlet conditions (see Section 3.1), with the exception now that xðtÞ must be absosinðko0 t=2Þ Xs ½k ¼ Aðt=TÞ ; lutely integrable over all time, which is a sufficient ðko0 t=2Þ but not a necessary condition. Periodic functions (4:2:11) Ys ½k ¼ Xs ½kejko0 ðt=2Þ ; o0 ¼ 2p=T: violate the last condition of absolute integrability over all time and will be considered in a later Atsinðko0 ðt=2ÞÞ Y ½ k ¼ X ½ k ¼ j j j j section. There are many functions that do not s s T ðko t=2Þ : 0 have Fourier transforms. For example, the Fourier transform of the function eat ; a40 is not defined. By using the complex Fourier series in (4.2.4), the The functions that can be generated in a laboratory transforms of the two pulses xðtÞ and yðtÞ are given have Fourier transforms. Existence of the Fourier below: transforms will not be discussed any further. In the yðtÞ ¼ AP½ðt t=2Þ=t; (4:2:12) synthesis equation, the inverse transform given by (4.2.8b) is an approximation of the function xðtÞ. sinðot=2Þ joðt=2Þ lim TYs ½k ¼ At ; (4:2:13a) e The function xðtÞ and its approximation x~ðtÞ are YðjoÞ ¼ T!1 ðot=2Þ equal in the sense that the error eðtÞ ¼ ½xðtÞ x~ðtÞ sinðot=2Þ is not equal to zero for all t and may differ signifi: (4:2:13b) XðjoÞ ¼ lim TXs ½k ¼ At T!1 cantly from zero at a discrete set of points t, but ðot=2Þ Z1

Obviously it is simpler to compute the transform directly. For example, jeðtÞj2 dt ¼ 0:

(4:2:9)

1

In the sense that the integral squared error is zero, the equality xðtÞ ¼ x~ðtÞ between the function and its approximation is valid. Physically realizable signals have F-transforms and when they are inverted, they provide the original function. Physically realizable signals do not have any jump discontinuities. That is x~ðtÞ ¼ xðtÞ. If the function xðtÞ has jump discontinuities, then the reconstructed function x~ðtÞ exhibits Gibbs phenomenon. The reconstructed function converges to the halfpoint at the discontinuity and will have overshoots and undershoots before and after the discontinuity (see Section 3.7.1). Example 4.2.1 Determine FfxðtÞg ¼ FfAP½t=tg and FfyðtÞg ¼ Ffxðt ðt=2ÞÞg using the F-series of

YðjoÞ ¼

Z1

1

xðtÞe

jot

Zt dt ¼ Aejot dt ¼

A ejot t0 ðjoÞ

0

1 ejot ejoðt=2Þ ejoðt=2Þ jot=2 ¼ At e ¼A jo 2jot=2 sinðot=2Þ jot=2 ¼ At : (4:2:13c) e ðot=2Þ The transform of the pulse xðtÞ is XðjoÞ ¼ At

sinðot=2Þ : ðot=2Þ

(4:2:14)

We should note that yðtÞ is a delayed version of xðtÞ and the delay explicitly appears in the phase spectra. See the difference between (4.2.13c) and & (4.2.14).

112

4 Fourier Transform Analysis

Since the F-transform is derived from the F-series, many of the properties for the F-series can be modified to derive the transform properties with some exceptions. Let a time limited function transform FT xðtÞ Xð joÞ be defined over the interval t0 5t5t0 þ T and zero everywhere else. This implies

$

xT ðtÞ ¼

1 X

xðt þ nTÞ ¼

n¼1

1 X

Xs ½ke jko0 t

k¼1

) Xs ½k ¼ ð1=TÞXðjoÞjo¼ko0 :

(4:2:15)

If xðtÞ is not time limited to a T second interval, then the function xðtÞ cannot be extracted from xT ðtÞ and (4.2.15) is not valid.

4.2.1 Amplitude and Phase Spectra

RðoÞ ¼

Z1 xðtÞ cosðotÞdt; 1

IðoÞ ¼

Z1 xðtÞ sinðotÞdt: 1

Since cosðotÞ is even and sinðotÞ is odd, the equalities in (4.2.18a) follow. As a consequence, for any real signal xðtÞ, we have XðjoÞ ¼ RðoÞ þ jIðoÞ ¼ RðoÞ ¼ jIðoÞ ¼ X ð joÞ: (4:2:18b) From (4.2.16b) and (4.2.17), the amplitude spectrum jXðjoÞj of a real signal is even and the phase spectrum yðoÞ is odd. That is, jXðjoÞj ¼ jXðjoÞj; yðoÞ ¼ yðoÞ:

Let the Fourier transform of a real signal xðtÞ be given by XðjoÞ. It is usually complex and can be written as either in terms of the rectangular or the polar form: XðjoÞ ¼ RðoÞ þ jIðoÞ ¼ AðoÞe jyðoÞ :

(4:2:16a)

The functions RðoÞ and IðoÞ are the real and the imaginary parts of the spectrum. In the polar form, the magnitude or the amplitude and the phase spectra are given by qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ AðoÞ ¼ jXðjoÞj ¼ R2 ðoÞ þ I 2 ðoÞ; yðoÞ ¼ tan1 ½IðoÞ=RðoÞ:

FT

xðtÞ ¼ xe ðtÞ þ x0 ðtÞ !RðoÞ þ jIðoÞ;

IðoÞ ¼ IðoÞ:

FT FT

x0 ðtÞ ¼ ½xðtÞ xðtÞ=2 !jIðoÞ:

¼

¼

XðjoÞ ¼

Z1 xðtÞe 1

j

jot

dt ¼

Z1

1

Z1 xðtÞejot dt 1 Z1

½xe ðtÞ þ x0 ðtÞ½cosðotÞ j sinðotÞdt

1 Z1

Z1 xe ðtÞ cosðotÞdt j x0 ðtÞ sinðotÞdt

1

xðtÞ sinðotÞdt ¼ RðoÞ þ jIðoÞ;

1

Z1 Z1 j xe ðtÞ sinðotÞdt þ x0 ðtÞ cosðotÞdt

xðtÞ cosðotÞdt 1

Z1

(4:2:21)

These can be seen from

(4:2:17)

These can be easily verified using Euler’s identity and

(4:2:20)

xe ðtÞ ¼ ½xðtÞ þ xðtÞ=2 !RðoÞ and

XðjoÞ ¼

RðoÞ ¼ RðoÞ and

(4:2:19)

Interesting transform relations in terms of the even and odd parts of a real function: If xðtÞ ¼ xe ðtÞ þ x0 ðtÞ, a real function, then the following is true:

(4:2:16b) When xðtÞ is real, the transform satisfies the properties that RðoÞ and IðoÞ are even and odd functions of o, respectively. That is,

(4:2:18a)

1

1

Z1 Z1 ¼ xe ðtÞ cosðotÞdt j x0 ðtÞ sinðotÞdt 1

¼ RðoÞ þ jIðoÞ

1

(4:2:22a)

4.2 Fourier Series to Fourier Integral

113

sinðoðt=2ÞÞ jot e 0 ¼ tjsincðot=2Þj; jXðjoÞj ¼ t oðt=2Þ

Z1 RðoÞ ¼ xe ðtÞ cosðotÞdt; 1

Z1 IðoÞ ¼ x0 ðtÞ sinðotÞdt:

(4:2:22b)

yðoÞ ¼

(4:2:24b) ðot0 Þ;

sincðot=2Þ40

ðot0 Þ p;

sincðot=2Þ50

:

(4:2:24c)

1

Note that integral of an odd function over a symmetric interval is zero. The F-transform of a real and even function is real and even and the F-transform of a real and odd function is pure imaginary. The transform of a real function xðtÞ can be expressed in terms of a real integral:

1 2p

1 ¼ 2p

1 Z1

jXðjoÞje jyðoÞ e jot do

¼ ðp þ pÞ ¼ 0:

1 Z1

jXðjoÞje jðotþyðoÞÞ do

1

Z1 1 ¼ jXðjoÞj cosðot þ yðoÞÞdo: p

(4:2:25a) The discontinuity in the phase spectrum at o ¼ 2p=t can be seen from 8 þ9 8 9 8 98 t 9 2p > 1;: ; >2p > > : > ; ¼ 2p ¼ p; y> y: : ; t t 2 t

Z1 1 xðtÞ ¼ XðjoÞe jot do 2p ¼

The time function xðtÞ; the magnitude spectrum jXðjoÞj, and the phase spectrum are sketched in Fig. 4.2.2 assuming t0 ¼ t=2. At o ¼ k2p=t, Xð j k2pÞ ¼ t; k ¼ 0 : t 0; k 6¼ 0 and k, an integer

(4:2:23)

We have added p in determining the phase angle in determining yð2pþ =tÞ taking into consideration that the sinc function is negative in the range 05ð2p=tÞ5o52ð2p=tÞ, i.e., in the first side lobe. In sketching the plots, appropriate multiples of (2p)

0

Note jXðjoÞj and sinðot þ yðoÞÞ are even and odd functions, respectively. Example 4.2.2 Rectangular (or a gating pulse) is given by xðtÞ ¼ P½ðt t0 Þ=t. a. Give the expression for the transform using (4.2.13a). b. Compute the amplitude and the phase spectra associated with the gating pulse function. c. Sketch the magnitude and phase spectra of this function assuming t0 ¼ t=2:

(a)

Solution: a. From (4.2.13a), we have sinðot=2Þ jot0 t XðjoÞ ¼ t e ¼ tsincðo Þejot0 2 ðot=2Þ

(b)

t 2

¼ tsincð2pf Þejot0 ; ht t i 0 FT jot0 ¼ t sincðpftÞej2pft0 : !t sincðot=2Þe t (4:2:24a) b. The magnitude and the phase spectra are, respectively, given by

(4:2:25b)

P

(c) Fig. 4.2.2 (a) xðtÞ, (b) jXðjoÞj, and (c) ﬀXðjoÞ

114

have been subtracted to make the phase spectrum compact and the phase angle is bounded between p and p. Noting sinðoðt=2ÞÞ / 1 ; (4:2:25c) jXðjoÞj ¼ t oðt=2Þ joj it can be seen that the envelope of the magnitude spectrum gets smaller for larger frequencies. The exact frequency representation of the square pulse should include all frequencies in the reconstruction of this pulse, which is impractical. Therefore, keep only the frequencies that are significant and the range or the width of those significant frequencies is referred to as the ‘‘bandwidth’’. Keeping the desired frequencies is achieved by using a filter. Filters will be discussed in later chapters. There are several interpretations of bandwidth. For the present, the following explanation of the bandwidth is adequate and will quantify these measures at a later & time.

4.2.2 Bandwidth-Simplistic Ideas 1. The width of the band of positive frequencies passed by a filter of an electrical system. 2. The width of the positive band of frequencies by the central lobe of the spectrum. 3. The band of frequencies that have most of the signal power. 4. The bandwidth includes the positive frequency range lying between two points at which the power is reduced to half that of the maximum. This width is referred to as the half-power bandwidth or the 3 dB bandwidth. Note that only positive frequencies are used in defining the bandwidth. With bandwidth in mind, let us look at jXðjoÞj in Example 4.2.2, where t is assumed to be the width of the pulse. Any signal that is nonzero for a finite period of time is referred to as a time-limited signal and the signal given in Example 4.2.2 is a time-limited signal. The main lobe width of the magnitude spectrum for positive frequencies is ð1=tÞ Hz. The spectrum is not frequency limited, as the spectrum occupies the entire frequency range, except that it is zero at o ¼ kð2p=tÞ; k 6¼ 0; k integer. The amplitude

4 Fourier Transform Analysis

spectrum jXðjoÞj gives the value jXðjo i Þj at the frequency oi ¼ 2pfi . Noting that most of the energy is in the main lobe, a standard assumption of bandwidth is generally assumed to be equal to k times half of the main lobe width ð2p=tÞ, i.e., (kð2p=tÞ) rad/s or ðk=tÞ Hz. An interesting formula that ties the time and frequency widths is (Time width) (Frequency width or bandwidth) ¼ Constant (4:2:26) Clearly as t decreases (increases), the main lobe width increases (decreases) and we say that bandwidth is inversely proportional to the time width. For most applications, k ¼ 1 is assumed. Bandwidth is generally given in terms of Hz rather than rad/s.

4.3 Fourier Transform Theorems, Part 1 We will consider first a set of theorems or properties associated with the energy function xðtÞ and its Ftransform XðjoÞ: Transforms are applicable for both the real and complex functions. In Chapter 3, Parseval’s theorem was given and it can be generalized to include energy signals and is referred to as generalized Parseval’s theorem, Plancheral’s theorem or Rayleigh’s energy theorem.

4.3.1 Rayleigh’s Energy Theorem The energy in a complex or a real signal is Z1

Ex ¼

¼

1 Z1

1 jxðtÞj dt ¼ 2p 2

Z1

jXðjoÞj2 do

1

jXðjfÞj2 df:

(4:3:1)

1

This is proved in general terms first. FT FT Given xðtÞ !XðjoÞ and yðtÞ !YðjoÞ, then

Exy ¼

Z1 1

1 xðtÞy ðtÞdt ¼ 2p

Z1

XðjoÞY ðjoÞdo:

1

(4:3:2)

4.3 Fourier Transform Theorems, Part 1

115

First, Z1

Exy ¼

¼

1 Z1

rectangular formula or the trapezoidal formula dis& cussed in Chapter 1.

1 y ðtÞ½ 2p

1

¼

¼

1 2p 1 2p

Z1

XðjoÞe jot dodt

1 Z1

Z1

1 ¼ 2p

Example 4.3.2 Compute the energy in the main lobe of the sinc function and compare with the total energy in the function using the following:

y ðtÞxðtÞdt

XðjoÞ½ 1 Z1

1 Z1

XðjoÞ½

EMain lobe

y ðtÞe

jot

XðjoÞY ðjoÞdo:

9 81 ð > > 2 > > > > sin ðpaÞ > : sinc2 ðot=2Þdo> da ¼ pp > > > > 2 ; : a

sinc2 ðpaÞda:

(4:3:6)

Solution: To obtain (4.3.6) from (4.3.5), change of variable, pa ¼ ot=2, is used and the limits are from o ¼ 2p=t to a ¼ 1. Using the rectangular method of integration, the energies in the main lobe and in the pulse E1 (see (4.3.4)) are EMain lobe 0:924=t; E1 ¼ ð1=tÞ:

(4:3:7)

The ratio of the energy in the main lobe, EMain lobe , of the spectrum to the total energy in the pulse is 92.4%. That is, the main lobe has over 90% of the total energy in the pulse function. Therefore, a bandwidth of (1/tÞ Hz is a reasonable estimate of & the pulse function.

1

(4:3:3) P

hti t

4.3.2 Superposition Theorem

FT

!t sincðot=2Þ:

Solution: Using the energy theorem, Z1

Z1

1

1

1 E1 ¼ 2p

(4:3:5)

1 1 Z1

Example 4.3.1 Compute the energy in the pulse E1 using Rayleigh’s energy theorem, the transform pair, and the identity Spiegel (1968) given below: 1 2p

1 ¼ t

yðtÞejot dt do

When xðtÞ ¼ yðtÞ, the proof of Rayleigh’s energy theorem in (4.3.1) follows. The transform of the function or the function itself can be used to find the energy.

E1 ¼

sinc2 ðot=2Þdo;

2p=t

dtdo

1

1 ð

Z2p=t

1 ¼ 2p

2 Z1 t 1 sinc ðot=2Þdo ¼ P2 dt t t 2

1

1

2 Zt=2 1 1 1 ðtÞ ¼ ¼ 2 dt: ¼ t t t

The Fourier transform of a linear combination of functions F½xi ðtÞ ¼ Xi ðjoÞ, i ¼ 1; 2; :::; n with constants ai ; i ¼ 1; 2:::; n is " F

n X

# ai xi ðtÞ ¼

i¼1 n X

(4:3:4)

i¼1

n X

ai Xi ðjoÞ;

i¼1 FT

ai xi ðtÞ !

n X

ai Xi ðjoÞ:

(4:3:8)

i¼1

t=2

Bandwidth of a rectangular pulse of width t is usually taken (1/tÞ Hz corresponding to the first zero crossing point of the spectrum. Energy contained in the main lobe of the sinc function can be computed by numerical methods, such as the

Since the integral of a sum is equal to the sum of the integrals, the proof follows. This theorem is useful in computing transforms of a function expressible as a sum of simple functions with known transforms. The F-transform of the function x ðtÞ is related to the transform of xðtÞ. This can be seen from

116

Z1

4 Fourier Transform Analysis

2

x ðtÞe

jot

dt ¼ 4

1

3

Z1 xðtÞe

jot

dt5 ¼ X ðjoÞ;

1 FT

) x ðtÞ !X ðjoÞ:

(4:3:9)

4.3.3 Time Delay Theorem The F-transform of a delayed function is given by F½xðt td Þ ¼ ejotd XðjoÞ:

(4:3:10)

This can be shown directly by using the change of variable a ¼ t td in the transform integral and F½xðt td Þ ¼

Z1

Solution: xðtÞ can be expressed as a sum of two rectangular pulses and is

t þ ðt=2Þ t ðt=2Þ P : xðtÞ ¼ P t t

(4:3:13a)

Using the superposition and delay theorems, we have

t þ ðt=2Þ t ðt=2Þ P XðjoÞ ¼ F½xðtÞ ¼ F P t t h i n h t io ¼ ejot=2 ejot=2 F P t sinðot=2Þ : (4:3:13b) ¼ 2jt sinðot=2Þ ðot=2Þ Note that xðtÞ is an odd function and therefore the & transform is pure imaginary.

xðt td Þejot dt

Notes: In the above example a time-limited function, i.e., xðtÞ ¼ 0 for jtj4t, was used and its trans2 1 3 Z form is not frequency limited, as its spectrum occuxðaÞejoa da5ejotd ¼ XðjoÞejotd pies the entire frequency range. A signal xðtÞ and its ¼4 transform XðjoÞ cannot be both time and frequency 1 & limited. We will come back to this later. ? ? ? ? ? ? ? ? )?F½xðt td Þ?¼?XðjoÞejotd?¼?XðjoÞ?¼?F½xðtÞ?: (4:3:11) 1

Superposition and delay theorems are useful in finding the Fourier transform pairs: FT

xðt tÞ þ xðt þ tÞ !2XðjoÞ cosðotÞ; (4:3:12) FT xðt tÞ xðt þ tÞ ! 2jXðjoÞ sinðotÞ: Example 4.3.3 Using the superposition and the delay theorem, compute the F-transform of the function shown in Fig. 4.3.1.

4.3.4 Scale Change Theorem The scale change theorem states that Z1 1 o xðatÞejot dt ¼ X j ; a 6¼ 0: F½xðatÞ ¼ a j aj 1

(4:3:14) This can be shown by considering the two possibilities, a50 and a40. For a50, by using the change of variable b ¼ at in (4.3.14) in the integral expression, we have

F½xðtÞ ¼

Z1 xðbÞe

jðo=aÞb

1 db a

1

Z1 1 ¼ xðbÞejðo=aÞb db a 1

Fig. 4.3.1 Example 4.3.3

1 ¼ j aj

Z1 1

o

xðbÞejð a Þb db ¼

1 o X j : a j aj

(4:3:15)

4.3 Fourier Transform Theorems, Part 1

117

When a is a negative number, a ¼ jaj. For a40, the proof similarly follows. The scale change theorem states that timescale contraction (expansion) corresponds to the frequency-scale expansion (contraction). Example 4.3.4 Use the scale change theorem to find the F-transforms of the following:

hti t x1 ðtÞ ¼ P and x2 ðtÞ ¼ P : (4:3:16a) ðt=2Þ 2t Solution: Consider the (4.2.24a)) with t0 ¼ 0: xðtÞ ¼ P½t=t

transform

$t sincðot=2Þ ¼ XðjoÞ: FT

Using the result in (4.3.14), we have

pair

(see

(4:3:16b)

x1 ðtÞ ¼ xð2tÞ ¼ P

t sinðoðt=4ÞÞ ! 2 oðt=4Þ

FT

t ¼ sincðoðt=4ÞÞ ¼ X1 ðjoÞ; 2 t h t i FT sinðotÞ x2 ðtÞ ¼ x ð2tÞ ¼P 2 2t ! ot ¼ ð2tÞ sincðotÞ ¼ X2 ðjoÞ:

(4:3:16c)

(4:3:16d)

The two functions and their amplitude spectra are sketched in Figs. 4.3.2a–d. Comparing the magnitude spectra, the main lobe width of jX1 ðjoÞj is twice that of the main lobe width of jXðjoÞj, whereas the main lobe width of jX2 ðjoÞj is half the main lobe width of jXðjoÞj. Consider Figs. 4.2.2 and 4.3.2. The main lobe width times its height in each of the cases are equal and tð2p=tÞ ¼ ðt=2Þð2pð2=tÞÞ ¼ ð2tÞðp=tÞ ¼ 2p. The

(a)

(b)

(c)

Fig. 4.3.2 (a) x1 ðtÞ, (b) jX1 ðjoÞj, (c) x2 ðtÞ, and (d) jX2 ðjoÞj

t t=2

(d)

118

4 Fourier Transform Analysis

pulse amplitudes are all assumed to be equal to 1 for simplicity. For any a io n h io ? ? n h ?F½xðtÞ? ¼ F P t a ¼ F P t : t t & Time reversal theorem: A special case of the scale change theorem is time reversal and F½xðtÞ ¼ XðjoÞ:

2a FT eajtj ! 2 ; a40: a þ o2

The time and frequency functions are not limited in & time and frequency, respectively.

(4:3:17a)

This follows from the scale change theorem by using a ¼ 1 in (4.3.14). We note that

4.3.5 Symmetry or Duality Theorem FT

FT

FT

xðtÞ !XðjoÞ; x1 ðtÞ ¼ xðtÞ !XðjoÞ ¼ X1 ðjoÞ;

jX1 ðjoÞj ¼ jXðjoÞj ¼ jXðjoÞj; ﬀX1 ðoÞ ¼ ﬀXðjoÞ ¼ ﬀXðjoÞ:

(4:3:17b)

Example 4.3.5 Find the F-transform of the following functions: a: x1 ðtÞ ¼ e x3 ðtÞ ¼ e

at

uðtÞ;

ajtj

b: x2 ðtÞ ¼ e uðtÞ;

¼

eðaþjoÞt t¼1 1 : ¼ ða þ joÞ t¼0 ða þ joÞ

(4:3:18c)

c. Noting that x3 ðtÞ ¼ eat uðtÞ þ eat uðtÞ, the Ftransform can be computed using the superposition theorem and the results in the last two parts. That is, X3 ðjoÞ ¼ F½eat uðtÞ þ F½eat uðtÞ 1 1 2a ¼ :(4:3:18d) þ ¼ ða þ joÞ ða joÞ a2 þ ðoÞ2 &

Summary: 1 FT eat uðtÞ ! ; a40; ða þ joÞ 1 FT ; a40; eat uðtÞ ! a jo

1 Z1

XðjoÞejot do ! 2p xðtÞ

XðjoÞejot do:

(4:3:22)

1

Interchanging t and jo in (4.3.22) results in 2p xðjoÞ ¼

Z1

XðtÞejot dt:

(4:3:23)

1

(4:3:18b)

b. Using the time reversal theorem and the last part results in X2 ðjoÞ ¼ X1 ðjoÞ ¼ ½1=ða joÞ:

Z1

2p xðtÞ ¼

¼

0

(4:3:21)

Starting with the expression for 2p xðtÞ and changing the variable from t to t, we have

(4:3:18a)

Solution: a. Using the F-transform integral results in Z1 Z1 X1 ðjoÞ ¼ eat uðtÞejot dt ¼ eðaþjoÞt dt 1

FT

xðtÞ !XðjoÞ ) XðtÞ !2pxðjoÞ:

at

; a40:

(4:3:20)

(4:3:19a) (4:3:19b)

This proves the result in (4.3.21). In terms of f (4.3.21) can be written as follows: FT

FT

xðtÞ !XðjfÞ; XðtÞ !xðjfÞ:

(4:3:24)

A consequence of the symmetry property is if an Ftransform table is available with N entries, then this property allows for doubling the size of the table. Example 4.3.6 Using the duality theorem, show that yðtÞ ¼

1 FT p ajoj e ¼ YðjoÞ: a2 þ t2 ! a

(4:3:25)

Solution: Using (4.3.20) and the duality property of the F-transforms, we have 1 ajtj FT 1 e ; ! 2 2a a þ o2 1 p FT 1 ð2pÞeajjoj ¼ eajoj : a4 0 ! 2 ! 2 a þt 2a a

4.4 Fourier Transform Theorems, Part 2

119

One can appreciate the simplicity of using the duality theorem compared to finding the transform directly in terms of difficult integrals given below: Z1 Z1 1 1 jot e do ¼ cosðotÞdo YðjoÞ ¼ a2 þ t 2 a2 þ t2 1

j

1

Z1

The value of the given function at t ¼ 0 and its transform value at o ¼ 0 are given by Z1 Z1 1 Xð0Þ ¼ xðtÞdt; xð0Þ ¼ XðjoÞdo: 2p 1

1 sinðotÞdo: 2 a þ t2

&

1

Example 4.3.7 Determine the F-transform of xðtÞ using the transform of the rectangular pulse given below and the duality theorem: h t i FT sinðot=2Þ sinðatÞ t ; P : xðtÞ ¼ pt t ! ðot=2Þ Solution: Using the duality theorem and noting that P-function is even, it follows

hoi sinðtt=2Þ FT jo t 2pP ¼ 2pP ; (4:3:26) ! ðtt=2Þ t t sinðatÞ FT h o i (4:3:27) !P 2a : pt Note ðt=2Þ ¼ a in (4.3.26). For later use, let a ¼ 2pB. Using this in (4.3.27) results in y ðt Þ ¼

4.3.6 Fourier Central Ordinate Theorems

sinð2pBtÞ FT 1 o ¼ YðjoÞ: (4:3:28) P ! 2B ð2pBtÞ 2pð2BÞ

Time domain sinc pulses are not time limited but are band limited. The sinc pulse and its transform in (4.3.28) are sketched in Fig. 4.3.3a,b, respectively. &

1

(4:3:29) Equation (4.3.29) points out that if we know the transform of a function, we can compute the integral of this function for all time by evaluating the spectrum at o ¼ 0. In a similar manner the integral of the spectrum for all frequencies is given by ð2pÞxð0Þ. Example 4.3.8 Use the transform pair in (4.3.28) to illustrate the ordinate theorems in (4.3.29) using the identity Spiegel (1968) 8 > Z1 < p; p40 sinðpaÞ (4:3:30) da ¼ 0; p ¼ 0 : > a : 1 p; p50 Solution: The integrals of the sinc function and the area of the pulse are A1 ¼

Z1 1

sinð2pWtÞ 1 dt ¼ ð2pWtÞ 2pW

Z1

sinð2pWtÞ dt t

1

p 1 ¼ : (4:3:31a) ¼ 2pW 2W

1 o 1 A2 ¼ P ) A2 ¼ A1 : jo¼0 ¼ 2W 2pð2WÞ 2W (4:3:31b) In a similar manner, sinð2pWtÞ jt¼0 ¼ 1; ð2pWtÞ Z1 1 1 B2 ¼ YðjoÞdo ¼ 2pð2 WÞ 2p 2pð2 WÞ

B1 ¼

1

(a)

¼ 1 ) B1 ¼ B2 :

(4:3:32) &

4.4 Fourier Transform Theorems, Part 2 (b) Fig. 4.3.3 (a) yðtÞ ¼

sinð2pWtÞ ð2pWtÞ

1 and (b) YðjoÞ=2W P

h

o 2pð2WÞ

i

Impulse functions are used in finding the transforms of periodic functions below.

120

4 Fourier Transform Analysis

Example 4.4.1 Find the Fourier transform of the impulse function in time domain and the inverse transform of the impulse function in the frequency domain.

Example 4.4.2 Show the following: 9 8 j ðo o c Þ> joc t FT 1 > ;: : xðatÞe ! j aj X a

(4:4:7a)

Solution: Using the scale change theorem results in Solution: Clearly 1 Z jo FT 1 jot jot jot0 xðatÞ ! X : (4:4:7b) F½dðt t0 Þ ¼ dðt t0 Þe dt ¼ e : jt¼t0 ¼ e a jaj 1

(4:4:1) That is, an impulse function contains all frequencies with the same amplitude. That is jF½dðt t0 Þj ¼ 1. The inverse transform F

1

1 ½dðo o0 Þ ¼ 2p

Z1

dðo o0 Þe jot do ¼

1 jo0 t e ; 2p

1

) F 1 ½dðoÞ ¼ 1=2p;

F½1 ¼ 2pdðoÞ:

(4:4:2) (4:4:3)

A constant contains only the single frequency at o ¼ 0 (or f ¼ 0Þ. We refer to a constant as a dc signal. Symbolically we can express FT

FT

dðt t0 Þ !ejot0 ; e jo0 t !2pdðo o0 Þ:

(4:4:4)

The result on the right in the above equation follows & by using the duality theorem.

4.4.1 Frequency Translation Theorem Multiplying a time domain function xðtÞ by ejoc t shifts all frequencies in the signal xðtÞ by oc . In general, the following transform pair is true: FT

xðtÞejoc t !Xðjðo oc ÞÞ:

(4:4:5)

Note F

1

Z1 1 ½Xðjðo oc ÞÞ ¼ Xðjðo oc ÞÞe jot do 2p 1 2 3 Z1 1 jat ¼4 XðjaÞe da5 e joc t ¼ xðtÞe joc t : 2p 1

(4:4:6) This provides a way to modify a time function to shift its frequencies. The scale change and the frequency translation theorems can be combined.

Using the frequency translation theorem, i.e., multiplying the function by e joc t causes a shift in the frequency. That is, replace o by o oc and the & result in (4.4.7a) follows.

4.4.2 Modulation Theorem The frequency translation theorem directly leads to the modulation theorem. Given F½xðtÞ ¼ XðjðoÞÞ and yðtÞ ¼ xðtÞ cosðoc t þ yÞ, the modulation theorem results in YðjoÞ ¼ F½xðtÞ cosðoc t þ yÞ

1 1 ¼ F ðxðtÞe jy Þe joc t þ ðxðtÞejy Þejoc t 2 2 1 1 ¼ ejy Xðjðo þ oc ÞÞ þ e jy Xðjðo oc ÞÞ: 2 2 (4:4:8) In simple words, multiplying a signal by a sinusoid translates the spectrum of a signal around o ¼ 0 to the locations around oc and oc . If the spectrum of the signal xðtÞ is frequency (or band) limited to o0 , i.e., jXðjoÞj ¼ 0; joj4o0 , then jYðjoÞj ¼ 0 for joj4joc þ o0 j and joj5joc o0 j: (4:4:9) Figure 4.4.1 gives sketches of the signals and their spectra. The signal xðtÞ is assumed to cross the time axis. There is no real significance in the shape of the spectrum. Since xðtÞ is real, it has even magnitude and odd phase spectrum. The signal is band limited to f0 ¼ o0 =2p Hz. The modulated signal yðtÞ shown in Fig. 4.4.1b assumes y ¼ 0 in (4.4.8). The positive and negative envelopes of the modulated signal are shown by the dotted lines. Note the envelopes cross the axis wherever the

4.4 Fourier Transform Theorems, Part 2

121

(a)

(b) Fig. 4.4.1 (a) xðtÞ and jXðjoÞj, (b) yðtÞ and jYðjoÞj

function xðtÞ ¼ 0. The magnitude and phase spectra of the modulated signal are shown in Fig. 4.4.1b. The bandwidth of the modulated signal is twice the bandwidth of the message signal. Note the factor & (1/2) in both terms in (4.4.8).

The signal yðtÞ is being seen through a rectangular (window) function xðtÞ. Outside this window, no signal is available. The study of windowed signals is an important topic for signal processors and it is humorously called as window carpentry. We will & come back to this topic later.

Example 4.4.3 Find the F½xðtÞ cosðoc tÞ and F½xðtÞ sinðoc tÞ in terms of F½xðtÞ: Solution: Clearly when y ¼ 0 and y ¼ p=2 in (4.4.8), the F-transform pairs are 1 FT 1 xðtÞ cosðoc tÞ ! Xðjðo oc ÞÞ þ Xðjðo þ oc ÞÞ; 2 2 (4:4:10a) 1 FT 1 xðtÞ sinðoc tÞ ! Xðjðo oc ÞÞ Xðjðo þ oc ÞÞ: 2j 2j (4:4:10b) Modulation theorem provides a powerful tool for finding the Fourier transforms of functions that are seen (or windowed) through a function xðtÞ. For example, hti hti xðtÞ ¼ P ! yðtÞ ¼ P cosðoc tÞ t t cosðoc tÞ; jtj52t : ¼ 0; otherwise

4.4.3 Fourier Transforms of Periodic and Some Special Functions Modulation theorem gives a back door way to find the Fourier transforms of periodic functions, such as sine and cosine functions. A sufficient condition for the existence of F½xðtÞ is Z1

jxðtÞjdt51 (absolute integrability condition):

1

Clearly the sine, cosine, unit step, and many other functions violate this condition. Use of the generalized functions allows for the derivation of the Ftransforms of these functions. Example 4.4.4 Use the transform F½1 ¼ 2pdðoÞ and the modulation theorem to find the Fourier transforms of xðtÞ ¼ cosðo0 tÞ and yðtÞ ¼ sinðo0 tÞ.

122

4 Fourier Transform Analysis

Solution: Using (4.4.8), we have the transforms. These are given below in terms of o and f: In the latter case, we have used dðoÞ ¼ dðfÞ=2p:

Notes: A narrowband band-pass signal with a slowly changing envelope RðtÞ and phase fðtÞ has the forms xðtÞ ¼ RðtÞ cosðoc t þ fðtÞÞ; RðtÞ 0;

FT

xðtÞ¼cosðo0 tÞ !XðjoÞ¼pdðoþo0 Þþpdðoo0 Þ; FT

yðtÞ¼sinðo0 tÞ !YðjoÞ¼jpdðoþo0 Þjpdðoo0 Þ; (4:4:11a)

(4:4:12a)

xðtÞ ¼ xi ðtÞ cosðoc tÞ xq ðtÞ sinðoc tÞ; D

xi ðtÞ ¼ RðtÞ cosðfðtÞÞ; xq ðtÞ ¼ RðtÞ sinðfðtÞÞ: (4:4:12b)

FT

cosð2pf0 tÞ ! 12 dðf þ f0 Þ þ 12 dðf f0 Þ; FT

sinð2pf0 tÞ ! 2j dðf þ f0 Þ 2j dðf f0 Þ:

(4:4:11b)

The spectra of the cosine and the sine functions are shown in Figs. 4.4.2. The spectra of these are located at o ¼ o0 ¼ 2pf0 with the same magnitude, but the phases are different. In reality, we do not have any negative frequencies. Euler’s formula illustrates that sinðoc tÞ and cosðoc tÞ are not the same functions, even though they have the same frequencies. Noting that the real part of the transform of a real signal is even and the phase spectra is odd, the negative frequency component does not give any additional information regarding what frequency is present. The average power represented by the negative frequency component simply adds to the average power of the positive frequency component resulting in the total average power at that frequency. In the case of an arbitrary signal resolved into in-phase and quadrature-phase components, the negative frequency terms do contribute additional information. A cosine wave reaches its positive peak 908 before a sine wave does. By convention the cosine wave is called the in-phase or i (or I) component and the sine wave is called the quadra& ture-phase or the q (or Q) component.

Equation (4.4.12a) gives the envelope-and-phase description and (4.4.12b) gives the in-phase and quadrature-carrier description. The components xi ðtÞ and xq ðtÞ are the in-phase and quadraturephase components. Now consider a windowed cosine function and & see the effects of that window. Example 4.4.5 Find the Fourier transform of the cosinusoidal pulse function. Plot the functions XðjoÞ and YðjoÞ and identify the important parameters: yðtÞ ¼ xðtÞ cosðo0 tÞ; xðtÞ ¼ P

Fig. 4.4.2 Transform of the cosine and sine functions

FT

!t sin cðot=2Þ:

Solution: The transform of yðtÞ is YðjoÞ ¼

t sin½ðo o0 Þðt=2Þ t sin½ðo þ o0 Þðt=2Þ þ : 2 ½ðo o0 Þðt=2Þ 2 ½ðo þ o0 Þðt=2Þ (4:4:13b)

The functions xðtÞ; XðjoÞ; yðtÞ; and YðjoÞ are sketched in Fig. 4.4.3a–d, respectively. Noting that XðjoÞ and YðjoÞ are real functions, the main lobe F[sin(ω0t)] π

0

t

(4:4:13a)

F[cos(ω0t)]

– ω0

hti

ω0

jπ

ω

ω

0

– jπ

4.4 Fourier Transform Theorems, Part 2

123

(a)

(b)

Y( jω)

(c)

(d)

Fig. 4.4.3 (a) and (b) Pulse function and its transform; (c) and (d) windowed cosine function and its transform

width of XðjoÞ corresponding to the positive frequencies is ð2p=tÞ. The function YðjoÞ has two main lobes centered at o ¼ o0 ¼ 2pf0 . Again considering only positive frequencies, the main lobe width of YðjoÞ is twice that of XðjoÞ equal to ð4p=tÞ. That is, the process of modulation doubles the bandwidth. As in XðjoÞ, we have side lobes in YðjoÞ that decay as we go away from the center frequency. The peak of the main lobe in XðjoÞ is t, whereas the peaks of the main lobes of YðjoÞ are equal to (t=2). Clearly, if we are interested in finding the frequency o0 ¼ 2pf0 , the steps could include the following: 1. Find the transform. 2. Find the peak value of the spectrum and its location. In a practical problem, we may have several frequencies. Finding the locations of these frequencies and their amplitudes is of interest. This problem is usually referred to as spectral estimation. The spectrum of a cosine function consists of two impulses located at o ¼ o0 . The spectrum of the windowed cosine function contains two sinc functions. We generally assume that o0 ð2p=tÞ and therefore the overlap of the two sinc functions at dc is

assumed to be negligible. Rectangular window modified the impulse spectra of the signal to a spectra consisting of sinc functions. Windowing a func& tion results in spectral leakage.

Fourier transforms of arbitrary periodic functions: In Chapter 3, we derived that if xT ðtÞ is a periodic function with period T and xT ðtÞ can be expressed into its F-series, 1 X xT ðtÞ ¼ Xs ½ke jko0 t ; k¼1

1 Xs ½k ¼ T

Z

xðtÞejko0 t dt;

o0 ¼

2p : T

(4:4:14)

T

where Xs ½k0 s in (4.4.14) are generally complex. The transform can be derived by noting that F½ejko0 t ¼ 2pdðo no0 Þ; F½xT ðtÞ ¼ ¼

1 X k¼1 1 X k¼1

(4:4:15)

Xs ½kF½e jko0 t Xs ½kð2pÞdðo ko0 Þ:

(4:4:16)

124

4 Fourier Transform Analysis

Example 4.4.6 Find the transform F½dT ðtÞ½¼ P F½ 1 k¼1 dðt kTÞ. Solution: The F-series of the function dT ðtÞ is given by (see (3.4.15b)) dT ðtÞ ¼

1 1 X e jko0 t ; o0 ¼ 2p=T: T k¼1

Using the linearity and frequency shift properties of the Fourier transforms, we have dT ðtÞ ¼

1 X

FT 2p dðt nTÞ ! T n¼1 1 X 2p dðo ko0 Þ ¼ do ðoÞ: T 0 k¼1

(4:4:17)

Example 4.4.7 Show the Gaussian pulse transform pair as follows:

e Solution: XðjoÞ ¼

¼

Z1 1 Z1

xðtÞejot dt ¼

Z1

(4:4:18)

XðjoÞ ¼ e

2 o4a

4p1ﬃﬃﬃ a

3

Z1 e

r

2

2

o dr5 ¼ e 4a

pﬃﬃﬃ jo aðt þ Þ; 2a

rﬃﬃﬃ p : a

(4:4:19)

Integral tables are used in (4.4.19). The transform pair in (4.4.18) now follows, that is, the Fourier transform of a Gaussian function is also a Gaussian function. Both time and frequency functions are not & limited in time and in frequency, respectively. The following pairs are valid and can be verified using Fourier transform theorems.

rﬃﬃﬃ 2 FT p o ap ; cos cos at2 ! 4a a

rﬃﬃﬃ 2 2 FT o ap p sin at ; ð4:4:20aÞ ! a sin 4a FT pﬃﬃﬃﬃﬃﬃ (4:4:20b) jtj1=2 ! 2pjoj1=2 :

2

eat ejot dt

eaðyÞ dt; y ¼ t2 þ j

Fig. 4.4.4 Periodic Impulse Sequence and its transform

By the change of variable, we have r ¼ pﬃﬃﬃ dt ¼ dr= a; t ) 1; r ! 1, and

For a catalog of Fourier transform pairs, see Abromowitz and Stegun (1964) and Poularikis (1996).

4.4.4 Time Differentiation Theorem

1

1

1

1

One question should come to our mind, that is, are there other functions and their transforms have the same general form? The answer is yes.

rﬃﬃﬃ p ðoÞ2 ! a e 4a ; a40:

ot ot o2 o2 ¼ t2 þ j þ 2 2 4a a a 4a 2 jo ðoÞ ¼ ðt þ Þ2 þ 2 ) XðjoÞ 4a 2a o2 Z1 o aðt þ j Þ2 2a dt: e ¼ e 4a

y ¼ t2 þ j

2

We have an interesting result: the Fourier transform of a periodic impulse sequence dT ðtÞ with period T is also a periodic impulse sequence ð2p=TÞdo0 ðoÞ with & period o0 . They are sketched in Fig. 4.4.4.

at2 FT

Now add and subtract the term ðo2 =aÞ to the term y in the exponent inside the integral

ot : a

If F½xðtÞ ¼ XðjoÞ and xðtÞ is differentiable for all time and vanishes as t ! 1, then

4.4 Fourier Transform Theorems, Part 2

F

dxðtÞ ¼ F½x0 ðtÞ ¼ ðjoÞ XðjoÞ: dt

125

(4:4:21)

Using integration by parts, we have

using the derivative method. Sketch the transform of the triangular function and compare the transform of the rectangular or P function with the transform of the L function.

Solution: xðtÞ; x0 ðtÞ; and x00 ðtÞ are sketched in Z1 d Fig. 4.4.5 a–c. Clearly, F xðtÞ ¼ x0 ðtÞejot dt ¼xðtÞejot t¼1 t¼1 þ jo dt 1 1 2 1 x00 ðtÞ ¼ dðt þ tÞ dðtÞ þ dðt tÞ: (4:4:24) Z1 t t t xðtÞejot dt ¼ jo XðjoÞ: 1

Differentiation of a function in time corresponds to multiplication of its transform by ðjoÞ, provided that the function xðtÞ ! 0 as t ! 1. If xðtÞ has a finite number of discontinuities, then x0 ðtÞ contains impulses. Then, (4.4.21) can be generalized and n

d xðtÞ F ¼ ðjoÞn XðjoÞ; n ¼ 1; 2; :::: (4:4:22) dtn The above does not provide a proof of the existence of the Fourier transform of the nth derivative of the function. It merely shows that if the transform exists, then it can be computed by the above formula. This theorem is useful if the transforms of derivatives of functions can be found easier than finding the transforms of functions. For example, dxðtÞ FT FT 1 joXðjoÞ ) xðtÞ ! F½x0 ðtÞ: ! dt jo Use of this approach in finding transforms is referred to as the derivative method. Example 4.4.8 Find the Fourier transform of the triangular function ( hti 1 tt ; jtj t xðtÞ ¼ L ¼ (4:4:23) t 0; Otherwise

Using the derivative theorem and solving for XðjoÞ, we have 1 2 1 FT x00 ðtÞ !ðjoÞ2 XðjoÞ ¼ e jot þ ejot ; t t t jot=2

2 V e ejot=2 2 ð4Þ ðjoÞ XðjoÞ ¼ 2j t 4 ¼ sin2 ðot=2Þ; t h t i FT sin2 ðot=2Þ xðtÞ ¼ V l Vt ! t ðot=2Þ2 ¼ Vt sinc2 ðot=2Þ ¼ XðjoÞ:

Equation (4.4.25) gives the spectrum of the triangular function of width ð2tÞ s. The rectangular function of width t s and its transform were given earlier by h t i FT sinðot=2Þ Vt ¼ Vt sincðot=2Þ: (4:4:26) VP t ! ðot=2Þ The time width of the rectangular window function in (4.4.26) is t s, whereas the time width of the triangular window function in (4.4.25) is 2t s. Note the square of the sinc2 function in the spectrum of the triangle function and the sinc function in the spectrum of the rectangular window. Since jsincðot=2Þj2 jsincðot=2Þj;

Fig. 4.4.5 (a) xðtÞ, (b) x0 ðtÞ, and (c) x0 ðtÞx0 ðtÞx00 ðtÞ

(4:4:25) &

126

4 Fourier Transform Analysis

the spectral amplitudes of the triangular function have lower side lobe levels compared to the spectral amplitudes of the rectangular pulses. Since the square of a fraction is smaller than the fraction itself, the side lobes in the transform of the triangular function are much smaller than the side lobes in the transform of the rectangular function. There is less leakage in the side lobes for the triangular (window) pulse function compared to the rectangular (window) function. See Fig. B.4.1 for a sketch of the sinc function. High-frequency decay rate: In the Fourier series discussion, the decay rate of the F-series coefficients Xs ½k was determined using the derivatives of the periodic function (see Section 3.6.5). Similarly, the F-transforms of pulse functions decay rate can be determined without actually finding the transform of the function. Given a pulse function xðtÞ, find the successive derivatives, xðnÞ ðtÞ, of the function until the first set of impulses appear in the nth derivative, then the decay rate is proportional to ð1=on Þ. In Example 4.4.7 the triangular pulse was considered and, in this case, the second derivative exhibits impulses indicating that the high-frequency decay rate of the transform is ð1=o2 Þ (see (4.4.25)). Similarly the first derivative of the rectangular pulse function and the exponential decaying function eat uðtÞ; a40 exhibit impulses indicating that the high-frequency decay rate of these transforms is ð1=jojÞ:

The similarities between the time and frequency differentiation theorems illustrate the duality properties with the F-transform pairs. Example 4.4.9 Show the following relationship using the times-t property: FT

teat uðtÞ !

1 ða þ joÞ2

; a40:

(4:4:28)

Solution: Noting the times-t property given above with xðtÞ ¼ eat uðtÞ, we have F½teat uðtÞ ¼ j

dXðjoÞ dð1=ða þ joÞÞ 1 ¼j ¼ : do do ða þ joÞ2

This can be generalized to obtain the following and the proof is left as an exercise: tn1 at 1 FT e uð t Þ ! ; a40: ðn 1Þ! ða þ joÞn

(4:4:29) &

Example 4.4.10 Noting that 1 FT eðajbÞt uðtÞ ! ; a40; a þ j ð o bÞ

(4:4:30a)

show the following is true: b

FT

xðtÞ ¼ eat sinðbtÞuðtÞ !

ða þ joÞ2 þb2

¼ XðjoÞ; (4:4:30b)

4.4.5 Times-t Property: Frequency Differentiation Theorem If XðjoÞ ¼ F½xðtÞ and if the derivative of the transform exists, then F½ðjtÞxðtÞ ¼

dXðjoÞ : do

(4:4:27)

This can be shown by dXðjoÞ d ¼ do do ¼

Z1 xðtÞe

1 Z1

jot

dt ¼

Z1 xðtÞ

dðejot Þ dt do

1

½ðjtÞxðtÞejot dt ¼ F½jtxðtÞ:

1

ða þ joÞ

FT

yðtÞ ¼ eat cosðbtÞuðtÞ !

ða þ joÞ2 þb2

¼ YðjoÞ: (4:4:30c)

Solution: These can be shown by first expressing the sine and cosine functions by Euler’s formulas, taking the transforms and then combining the & complex–conjugate terms. Example 4.4.11 Using lim eat uðtÞ ¼ uðtÞ; a40, a!0 find F½uðtÞ: 1 : a!0 a þ jo

F½uðtÞ ¼ lim

Solution: Noting that the limiting process is on the complex function, we need to take the limits on the

4.4 Fourier Transform Theorems, Part 2

127

real and the imaginary parts of the complex function separately. That is,

1 a o þ j lim : ¼ lim 2 lim a!0 a þ jo a!0 a þ o2 a!0 a2 þ o2

unit step function is real. These are illustrated in Fig. 4.4.6. Interestingly the spectrum of the delayed unit step uðt 1Þ is

(4:4:31)

1 F½uðt 1Þ ¼ ½pdðoÞ þ ejo ; jo

1 jF½uðt 1Þj ¼ pdðoÞ þ ; ﬀF½uðt 1Þ o o p=2; o40 ¼ : (4:4:34b) o þ p=2; o50

The second term in the above, i.e., the Lorentzian function, approaches an impulse function. That is,

a ¼ pdðoÞ: (4:4:32) lim a!0 a2 þ o2 Using this result in (4.3.30),

1 1 ¼ pdðoÞ þ ; lim a!0 a þ jo jo

(4:4:33)

1 ; (4:4:34a) jo p=2; o40 : jUðjoÞj ¼ pdðoÞ þ j1=oj; ﬀUðjoÞ ¼ p=2; o50 ) UðjoÞ ¼ F½uðtÞ ¼ pdðoÞ þ

Note that the amplitude is an even function and the phase angle function is an odd function, as the

Since delay of a function depends on the phase angle, it follows that jF½uðt 1Þj ¼ jF½uðtÞj. The phase spectrum of the delayed unit step function is sketched in Fig. 4.4.7. The Fourier transform of the unit function has two parts. The first part corresponds to the transform of the average value of the unit step function and the other part is the transform of the signum function. That is, F½uðtÞ ¼ F½ð1=2Þ þ ð1=2Þ sgnðtÞ ¼ F½1=2 þ ð1=2ÞF½sgnðtÞ ¼ pdðoÞ þ ð1=joÞ:

Fig. 4.4.6 (a) Magnitude and (b) phase spectra of the unit step function

Fig. 4.4.7 Phase spectrum of u(t1)

&

128

4 Fourier Transform Analysis

Notes: If we had ignored that we had to take the limit on the real and the imaginary parts of the complex function in (4.4.31) separately and take the limit on the complex function as a whole, the result would be wrong. That is,

FT

ð1=tÞ ! jp sgnðoÞ ¼ jp j2puðoÞ: This can be generalized and 1 FT ðjoÞn1 jp sgnðoÞ: ! ðn 1Þ! tn

1 1 lim ¼ 6¼ F½uðtÞ: a!0 a þ jo jo This indicates the transform is imaginary and the time function must be odd. This cannot be true since the unit step function is not an odd function and & F½uðtÞ 6¼ ð1=joÞ. The sgn (or signum) function is used in communications and control theory and can be expressed in terms of the unit step function. The sgn function and its transform are as follows: 8 > < 1; t40 (4:4:35) sgnðtÞ ¼ 2uðtÞ 1 ¼ 0; t ¼ 0: > : 1; t50 F½2uðtÞ F½1 ¼ 2pdðoÞ þ ð2=joÞ 2pdðoÞ ¼ ð2=joÞ:

(4:4:36)

The times-t property and the transform of the unit step function can be used to determine the Fourier transform of the ramp function and is given as

FT tuðtÞ ! jpd0 ðoÞ 1=o2 :

(4:4:37)

(4:4:39a)

(4:4:39b) &

4.4.6 Initial Value Theorem The initial value theorem is applicable for the rightsided signals, i.e., the functions of the form yðtÞ ¼ xðtÞuðtÞ and is stated below without proof: yð0þ Þ ¼ lim joYðjoÞ: o!1

(4:4:40)

Example 4.4.13 The unit step function is not defined at t ¼ 0, whereas uð0þ Þ ¼ 1; which is well defined. Verify the initial value theorem for the unit step function by noting dðoÞ ¼ 0; o 6¼ 0, and odðoÞ ¼ 0: Solution: uð0þ Þ ¼ lim fjoF½uðtÞg o!1

1 ¼ 1: ¼ lim jo pdðoÞ þ o!1 jo

&

Noting that jtj ¼ 2t uðtÞ t, we have the following transform pair:

4.4.7 Integration Theorem

FT jtj ! 2=o2 :

(4:4:38)

Example 4.4.12 Find the Fourier transform of the function xðtÞ ¼ ð1=tÞ using the duality theorem and the Fourier transform of the signum function. Solution: Using the duality theorem, we have

It states that yð t Þ ¼

Zt

FT XðjoÞ þ pXð0ÞdðoÞ ¼ YðjoÞ: xðaÞda ! jo

1

(4:4:41) FT

Duality theorem

FT

! XðtÞ !2pxðjoÞ; xðtÞ !XðjoÞ 1 FT FT 2 sgnðtÞ ! ; ð1Þp sgnðoÞ ¼ p sgnðoÞ: jo jt ! We can write sgn ðjoÞ ¼ sgnð oÞ ¼ sgnðoÞ. It follows that

This is true only if Xð0Þ; i.e., the area under xðtÞ, is finite. If the area under xðtÞ is zero, then the second term on the right in (4.4.41) disappears. Note that if Xð0Þ ¼ 0, integration and differentiation operations are inverse operations. Integration operation is a smoothing operation. Integral of a function has

4.5 Convolution and Correlation

129

lower frequency content than the function that is integrated. On the other hand, since x 0 ðtÞ ¼ joXðjoÞ, differentiation accentuates the higher frequencies. Integration theorem is not applicable if Xð0Þ is infinity. This theorem will be proved in Section 4.5. Example 4.4.14 Find the Fourier transform of uðtÞ using the integration theorem and Zt FT dðaÞda: dðtÞ !1; uðtÞ ¼ 1

Solution: By the integration theorem 2 t 3 Z 1 dðaÞda5 ¼ þ pdðoÞ: F½uðtÞ ¼ F4 jo 1

&

Example 4.4.15 Use the Fourier transform of the function xðtÞ ¼ cosðo0 tÞ; o0 6¼ 0, and the integration theorem to find the Fourier transform of the function sinðo0 tÞ. Solution: First for o0 6¼ 0, from (4.4.11a), we have FT

xðtÞ ¼ cosðo0 tÞ !pdðo þ o0 Þ þ pdðo o0 Þ ¼ XðjoÞ; Xð0Þ ¼ 0 and yðtÞ ¼

Zt

xðaÞda ¼

1

Zt cosðo0 aÞda (4:4:42a)

See the comment below in regard to the evaluation of the limit at 1 in the above integral. The integration property gives us Zt

(4:4:43a)

The limit does not exist as an ordinary limit and is a generalized limit in the sense of distributions. Using Euler’s formula and the limit in (4.4.43a), computation of the integral in (4.4.42a) follows. Switched functions are very useful in system theory. In computing the derivatives of such functions, one needs to be careful. For example, d½cosðtÞuðtÞ d½cosðtÞ d½uðtÞ ¼ uðtÞ þ cosðtÞ dt dt dt ¼ sinðtÞuðtÞ þ dðtÞ; (4:4:43b) d½sinðtÞuðtÞ ¼ cosðtÞuðtÞ þ sinðtÞdðtÞ ¼ cosðtÞuðtÞ: dt (4:4:43c) To find the transforms of such functions we can make use of modulation theorem. Derivative theorem can be used to find transforms of many functions such as xðtÞ ¼ ea t uðtÞ; a40. We should keep in mind that if the pulse is not time limited, we need to add a frequency domain delta function, whose weight is equal to 2p times the average of the pulse over the entire time axis to the transform result of the successive differentiation. See the discussion on & finding the transform of a unit step function.

4.5 Convolution and Correlation Chapter 2 considered convolution and correlation. Here we will consider the transforms of the signals that are convolved and correlated.

FT

cosðo0 aÞda !

1

o0 ½pdðo þ o0 Þ þ pdðo o0 Þ þ o0 pXð0ÞdðoÞ: jo With Xð0Þ ¼ 0 and dðo o0 Þ=o ¼ dðo o0 Þ=o0 , we have result as in (4.4.11a). Zt

lim ejo t ¼ 0:

t!1

1

¼ ð1=o0 Þ sinðo0 tÞ:

o 0 yð t Þ ¼ o 0

Notes: Papoulis (1962) discusses the concepts of generalized limits. For example

4.5.1 Convolution in Time Convolution of two time functions x1 ðtÞ and x2 ðtÞ is defined by Z1 yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ x1 ðaÞx2 ðt aÞda 1

FT

cosðo0 aÞda ¼ sinðo0 tÞ !jpdðo þ o0 Þ ¼

1

jpdðo o0 Þ:

(4:4:42b) &

Z1 1

x2 ðbÞx1 ðt bÞdb ¼ x2 ðtÞ x1 ðtÞ: (4:5:1)

130

4 Fourier Transform Analysis FT

Assuming that xi ðtÞ !Xi ðjoÞ; i ¼ 1; 2, convolution theorem is given by FT

x1 ðtÞ x2 ðtÞ !X1 ðjoÞX2 ðjoÞ:

the

(4:5:2)

This can be proven by using the transform pair F½x2 ðt aÞ ¼ X2 ðjoÞejoa in (4.5.1) and the resulting integral is the inverse transform of ½X1 ðjoÞX2 ðjoÞ. That is, yðtÞ ¼ x1 ðtÞ x2 ðtÞ 2 3 Z1 Z1 1 x1 ðaÞ4 X2 ðjoÞe joðtaÞ do5da ¼ 2p 1 1 2 1 3 Z1 Z 1 X2 ðjoÞ4 x1 ðaÞejoa da5e jot do ¼ 2p ¼

1 2p

1 Z1

inverse transform of YðjoÞ. Second, in most applications, the function may not be given in an analytical or equation form and may be given in the form of a plot or a set of data and we have to resort to digital means to find the values for yðtÞ. We will consider the discrete Fourier transforms in Chapters 8 and 9. Example 4.5.1 Determine the function yðtÞ ¼ P½t :5 P½t :5 by using the transforms. Solution: Using the transforms of the pulse functions, we have

1 sinðo=2Þ jo=2 F P t ¼ e ; 2 ðo=2Þ " # sin2 ðo=2Þ jo (4:5:4) YðjoÞ ¼ e : ðo=2Þ2

1

½X2 ðjoÞX1 ðjoÞe jot do:j

(4:5:3)

1

Convolution theorem follows from the above equation. It gives a method for computing the convolution of two aperiodic functions via Fourier transforms. This method is the transform method of computing the convolution. The direct method is by the use of the convolution integral and all the operations are in the time domain. The transform method involves the following steps: a. Find F½x1 ðtÞ ¼ X1 ðjoÞ and F½x2 ðtÞ ¼ X2 ðjoÞ. b. Determine YðjoÞ ¼ X1 ðjoÞX2 ðjoÞ. c. Find the inverse transform of the function YðjoÞ to obtain yðtÞ. There are several problems with the transform method of computing the convolution. First, the given function may not have analytical expressions for the transforms. Even if does, we may not be able to find the

Fig. 4.5.1 Convolution of two square pulses

Using (4.4.25) and the time delay theorem, we have a triangle or a tent function given by yðtÞ ¼ L½t 1:

(4:5:5)

The given time functions and the result of the con& volution are shown in Fig. 4.5.1. Example 4.5.2 Consider the two delayed functions x1 ðt t1 Þ and x2 ðt t2 Þ. a. Assuming yðtÞ ¼ x1 ðtÞ x2 ðtÞ is known, show the following is true by using the transform method: zðtÞ ¼ x1 ðt t1 Þ x2 ðt t2 Þ ¼ yðt ðt1 þ t2 ÞÞ: (4:5:6) b. Using the results in (4.5.6), determine the convolution of the two impulse functions yðtÞ ¼ dðt t1 Þ dðt t2 Þ: Solution: a. Using the convolution and time delay theorems, we have

4.5 Convolution and Correlation

131

ZðjoÞ ¼ F½zðtÞ ¼ F½x1 ðt t1 ÞF½x2 ðt t2 Þ ¼ F½x1 ðtÞejot1 F½x2 ðtÞejot2 ¼ X1 ðjoÞX2 ðjoÞejoðt1 þt2 Þ ¼ YðjoÞejoðt1 þt2 Þ :

In (4.5.9b) we have made use of partial fraction expansion and the transforms of the unit step func& tion and the exponential decaying function. Example 4.5.5 Determine the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ in each case below using the transforms. a. The two Gaussian functions and their transforms are given by

The inverse transform of ZðjoÞ is given by yðt ðt1 þ t2 ÞÞ. b. Noting that F½dðt ti Þ ¼ ejoti ; i ¼ 1; 2, we have h i 1 2 2 ðosi Þ2 =2 YðjoÞ ¼ ejoðt1 þt2 Þ ; F1 ejoðt1 þt2 Þ ¼ dðt ðt1 þ t2 ÞÞ: xi ðtÞ ¼ pﬃﬃﬃﬃﬃﬃ et =2si FT ¼ Xi ðjoÞ; i ¼ 1; 2:: !e si 2p (4:5:10a) That is, the inverse transform is a delayed impulse function and b. The two sinc functions and their transforms yðtÞ ¼ dðt t1 Þ dðt t2 Þ ¼ dðt ðt1 þ t2 ÞÞ: (4:5:7) &

Example 4.5.3 Show by using the transform method yðtÞ ¼ xðtÞ dðtÞ ¼ xðtÞ:

are given by

sinðtti =2Þ FT o ¼ Xi ðjoÞ; 2pP ! ðtti =2Þ ti (4:5:10b) i ¼ 1; 2; t1 5t2 :

xi ð t Þ ¼ ti

(4:5:8) c.

Solution: We have F½xðtÞ dðtÞ ¼ F½xðtÞF½dðtÞ ¼ F½xðtÞ and yðtÞ ¼ xðtÞ:

&

FT

xi ðtÞ ¼ 1=jpt !sgnðoÞ; i ¼ 1; 2::

Solution: a. Noting that the transform of a Gaussian pulse is a Gaussian pulse, the product of the two Gaussian pulses is a Gaussian pulse: FT

2=2

2=2

Example 4.5.4 Determine the convolution yðtÞ ¼ eat uðtÞ uðtÞ; a40 by a. the direct method and b. by the transform method.

yðtÞ ¼ x1 ðtÞ x2 ðtÞ !eðos1 Þ eðos2 Þ 2 2 2 ¼ eo ðs1 þs2 Þ=2 ¼ YðjoÞ:

Solution: a. By the direct method,

The inverse transform of this function is again a Gaussian pulse with

yðtÞ ¼

Z1

eaðtbÞ ½uðt bÞuðbÞdb ¼ eat

1

1 1 at ¼ eat eab b¼t ÞuðtÞ: b¼0 ¼ ð1 e a a

Zt

eab db

0

(4:5:9a)

b. By the transform method,

1 1 pdðoÞ þ YðjoÞ ¼ F½e uðtÞF½uðtÞ ¼ ða þ joÞ jo pdðoÞ 1 þ ; ¼ a joða þ joÞ at

1 1 1 1 YðjoÞ ¼ pdðoÞ þ ! a jo a a þ jo 1 yðtÞ ¼ ð1 eat ÞuðtÞ; a40: a

1 2 2 yðtÞ ¼ pﬃﬃﬃﬃﬃﬃ et =2s ; s2 ¼ s21 þ s22 : s 2p

(4:5:10c)

(4:5:10d)

b. The rectangular pulses in the Fourier domain overlap. The product of the two rectangular pulses is a rectangular pulse and its inverse transform is a sinc pulse. The details are left as an exercise. c. The convolution of the two functions and its transform are given by yð t Þ ¼

1 1 FT sgn2 ðoÞ ¼ 1; F1 ½1 ¼ dðtÞ: jpt jpt ! (4:5:10e) &

(4:5:9b)

Notes: The transform method is simpler if the transforms of the individual functions and the

132

4 Fourier Transform Analysis

inverse transform of the convolution are known. In Chapter 2 we discussed the duration property associated with convolution and pointed out that there are exceptions. Part c of the above example illustrates an exception. Convolutions of some functions do not exist. For example, yðtÞ ¼ uðtÞ uðtÞ does not exist since its transform has a term that is a square of an impulse function, which is not & defined.

4.5.2 Proof of the Integration Theorem In Section 4.4.6 the integration theorem is stated (see (4.4.42)) and is yð t Þ ¼

Zt

FT XðjoÞ þ pXð0ÞdðoÞ ¼ YðjoÞ: xðaÞda ! jo

1 1 F1 aþjo bþjo Zt at bt ¼e uðtÞe uðtÞ¼ eaa uðaÞebðtaÞ uðtaÞda

xðtÞ¼F1

0

¼

Zt e

1

xðaÞda¼

Z1

e

dt¼e

bt

eðbaÞt dt

0

Second, by using partial fraction expansion, we have XðjoÞ ¼

1 1 ; a 6¼ b ðb aÞða þ joÞ ðb aÞðb þ joÞ (4:5:12c)

) xðtÞ ¼ F 1 ½XðjoÞ ¼

1 eat uðtÞ ðb aÞ

1 ebt uðtÞ: ðb aÞ

(4:5:12d)

This coincides with the solution in (4.5.12b). b. When a ¼ b; the transform function has a double pole. By the convolution method,

xðaÞuðt aÞda

1 FT

¼ xðtÞ uðtÞ !XðjoÞF½uðtÞ;

1 YðjoÞ ¼ XðjoÞ pdðoÞ þ jo 1 ¼ pXð0ÞdðoÞ þ XðjoÞ: jo

at

xðtÞ ¼ e

uðtÞ e

at

uðtÞ ¼

Zt

eat eaðttÞ dt

0

¼ eat

Zt

dt ¼ teat uðtÞ:

(4:5:12e)

0

Since the function has a double pole, we can find its inverse transform by using times-t property of the Fourier transforms or from tables. Now

This proves the integration theorem. Example 4.5.6 Find the inverse transform of the function XðjoÞ given below for two cases: a. a 6¼ b; a40; b40 and b. a ¼ b40 XðjoÞ ¼ 1=½ða þ joÞðb þ joÞ:

Zt

1 h ðbaÞt it¼t eat ebt e uðtÞ: ¼ ¼ebt t¼0 ðbaÞ ðbaÞ (4:5:12b)

Since uðt aÞ ¼ 0 for a4t, we can write the above running integral as a convolution: Zt

at bðttÞ

0

1

(4:5:11)

1 d 1 1 ¼ ¼ XðjoÞ; ðjÞ do ða þ joÞ ða þ joÞ2

(4:5:12a)

( xðtÞ ¼ F

Solution: a. This can be solved by first noting that convolution in time domain corresponds to the multiplication in the frequency domain. Therefore,

1

1

) ¼ F 1

1 d 1 ðjÞ do ða þ joÞ

ða þ joÞ2 1 1 1 ¼ ðjtÞF ¼ teat uðtÞ : ðjÞ a þ jo

4.5 Convolution and Correlation

133

This coincides with the result obtained in (4.5.12e). & Example 4.5.7 Find yðtÞ for the function below by using a. the derivative theorem and b. the long division: YðjoÞ ¼

jo ; a40: ða þ joÞ

(4:5:13)

Solution: a. Using the derivative theorem, we have d½F1 ð1=ða þ joÞ dðeat uðtÞÞ ¼ dt dt at duðtÞ dðe Þ ¼ eat þ uðtÞ dt dt ¼ eat dðtÞ aeat uðtÞ ¼ dðtÞ aeat uðtÞ: (4:5:14)

yðtÞ ¼

b. Also, dividing jo by ða þ joÞ by long division and using the superposition property of the

Fourier transforms gives the same result as in (4.5.14). That is, jo a FT ¼1 dðtÞ aeat uðtÞ: a þ jo a þ jo !

&

4.5.3 Multiplication Theorem (Convolution in Frequency) The dual to the time convolution theorem is the convolution in frequency theorem. It is given below and can be shown directly by using a proof similar to the time domain convolution theorem. An alternate way of showing is by using the symmetry theorem FT 1 yðtÞ ¼ x1 ðtÞx2 ðtÞ ! X1 ðjoÞ X2 ðjoÞ: 2p

(4:5:15)

Summary: Convolution in time and in frequency: FT

Convolution in time: ½xðtÞ x2 ðtÞ !½X1 ðjoÞX2 ðjoÞ : Multiplication in frequency FT 1 Multiplicaion in time: ½x1 ðtÞx2 ðtÞ ! 2p ½X1 ðjoÞ X2 ðjoÞ : Convolution in frequency Example 4.5.8 Consider the time function and its transform hoi FT xðtÞ !XðjoÞ ¼ P : (4:5:16) W Find the Fourier transform of the function yðtÞ ¼ x2 ðtÞ and its bandwidth by assuming the bandwidth of xðtÞ is ðW=2Þ rad/s.

function L½o=W is (W). Note the duration property of the convolution is satisfied since the width of the triangular pulse is twice that of the rectangular pulse. From the area property of the convolution, we have using the time averages n h o io n h o io A P ¼ W2 : A P W W

Coming back to the bandwidths, if x1 ðtÞ and x2 ðtÞ have bandwidths of B1 and B2 Hz, respectively, then the bandwidth of yðtÞ ¼ x1 ðtÞx2 ðtÞ is equal to ðB1 þ B2 Þ Hz. Multiplication of two time functions hoi hoi h o i increases the bandwidth of the resulting time funcYðjoÞ ¼ XðjoÞ XðjoÞ ¼ P P ¼ WL : tion. The above property is dual to the time width W W W property of the convolution. We have seen some (4:5:17) & pathological cases where the time width property of the convolution does not hold. What about in the Notes: It is instructive to review the properties of frequency domain? Obviously, the same is true in the convolution of the transform functions in the frequency domain for pathological cases. For (4.5.17). The bandwidth of the pulse function practical signals, the above discussion applies. We P½o=W is W=2, whereas the bandwidth of the will come back to this topic at a later time, as it Solution: Example 2.3.1 considered the time domain convolution of two rectangular pulse functions. Using these results, we have

134

4 Fourier Transform Analysis

pertains to the important topic of nonlinear systems and the bandwidth requirements of such systems. Fourier transform computation of windowed periodic functions: Stanley et al. (1984) present a nice approach in finding the transforms of windowed time-limited trigonometric functions using the multiplication theorem, which is presented below. Let gT ðtÞ be a periodic function with period T and F½gT ðtÞ ¼ GðjoÞ. Let pðtÞ be a pulse function with PðjoÞ ¼ F½pðtÞ and a function F½wðtÞ ¼ WðjoÞ is defined by FT

wðtÞ ¼ pðtÞgT ðtÞ !PðjoÞ GðjoÞ:

(4:5:18)

We like to find F½wðtÞ using F½e jko0 t ¼ 2pdðo ko0 Þ and gT ðtÞ ¼

1 X

Gs ½ke jko0 t ;

k¼1

Gs ½k ¼

1 T

Z

gT ðtÞejko0 t dt;

T

o0 ¼ 2p=T; GðjoÞ ¼ F½gT ðtÞ 1 X Gs ½kdðo ko0 Þ: ¼ 2p

gT ðtÞ ¼ :54 þ :46 cosð2pt=TÞ;

(4:5:22a) ) wH ðtÞ ¼ gT ðtÞP

k¼1

¼ 2p ¼ 2p

1 X k¼1 1 X

Gs ½kPðjðo ko0 ÞÞ:

(4:5:20)

Example 4.5.9 Find the Fourier transform of the Hamming window function given below using the above method: :54 þ :46 cosð2pt=TÞ; jtj T=2 : wH ðtÞ ¼ 0; Otherwise (4:5:21) Solution: Define a periodic function using the window function in (4.5.21) by

(4:5:22b)

:

þ :23ð2pÞdðo þ o0 Þ; o0 ¼ 2p=T; (4:5:23) h h t ii WH ðjoÞ ¼ F gT ðtÞP ¼ GðjoÞ T

sinðoT=2Þ T sinðoT=2Þ :54ð2pÞ dðoÞ T ðoT=2Þ ðoT=2Þ

T sinðoT=2Þ þ :23ð2pÞ dðo o0 Þ ðoT=2Þ

T sinðoT=2Þ : þ :23ð2pÞ dðo þ o0 Þ ðoT=2Þ

1 dðo o0 Þ YðjoÞ ¼ 2p

Z1

dða o0 ÞYðjðo aÞÞda

1

1 ¼ Yðjðo o0 ÞÞ; 2p

we have

T sinðoT=2Þ T sinððo o0 ÞT=2Þ þ :23 ðoT=2Þ ððo o0 ÞT=2Þ T sinððo þ o0 ÞT=2Þ : þ :23 ððo þ o0 ÞT=2Þ

WH ðjoÞ ¼ :54

(4:5:24)

Gs ½kfPðjoÞ dðo ko0 Þg

k¼1

T

GðjoÞ ¼:54ð2pÞdðoÞ þ :23ð2pÞdðo o0 Þ

k¼1

The transform of the function wðtÞ can be obtained by the convolution of the two transform functions PðjoÞ ¼ F½pðtÞ and GðjoÞ. That is, " # 1 X WðjoÞ ¼ PðjoÞ 2p Gs ½kdðo ko0 Þ

hti

This function gT ðtÞ contains a constant and a cosine function. Its Fourier transform is

With

(4:5:19)

gT ðt þ TÞ ¼ gT ðtÞ;

Noting sinðoðT=2Þ o0 ðT=2ÞÞ ¼ sinðoðT=2Þ pÞ ¼ sinðoT=2Þ and using this in (4.5.24) results in WH ðjoÞ¼:54T

sinðoT=2Þ ðoT=2Þ

:23TsinðoT=2Þ

1 1 þ : ðoT=2Þp ðoT=2Þþp

In terms of f, we have " # T sinðpfTÞ :54 :08ðfTÞ2 ; o ¼ 2pf: WH ðjoÞ ¼ ðpfTÞ 1 ðfTÞ2 (4:5:25) &

4.5 Convolution and Correlation

135

4.5.4 Energy Spectral Density From Rayleigh’s energy theorem, the energy contained in an energy signal F½xðtÞ ¼ XðjoÞ can be computed either by the time domain function or by the frequency domain function and the energy contained in the signal is E¼

Z1

1 jxðtÞj dt ¼ 2p 2

1

¼

Z1

Z1

c: E:95

ZW :95 1 do ¼ ¼ do 2 2a 2p a þ o2 W 1 1 W tan ! a tanð:95ðp=2ÞÞ ¼ W ; ¼ ap a

W ¼ 2pF; F ð2:022aÞ Hz:

(4:5:27d) &

Example 4.5.11 Consider the pulse function xðtÞ ¼ P½t=t. Find the percentage of energy contained in the frequency range W5o5W; W ¼ 2pfc .

jXðjoÞj2 do

1

jXðjfÞj2 df; GðfÞ ¼ jXðjfÞj2 ¼ jXðjoÞj2 =2p:

1

Solution: The spectrum and the energy spectral densities are, respectively, given by

(4:5:26) Note that GðfÞ ¼ jXðjfÞj2 ¼ jXðjoÞj2 =2p is the energy spectral density. Example 4.5.10 a. Derive the energy spectral density of the function FT

xðtÞ ¼ eat uðtÞ !½1=ða þ joÞ ¼ XðjoÞ; a40 b. Illustrate the validity of Rayleigh’s energy theorem. c. Select the frequency band W5o5W so that 95% of the total energy is in this band. Solution: a. The energy spectral density is given by (4:5:27a) jXðjoÞj2 ¼ 1= a2 þ o2 : b. By using the time domain function, the energy contained in the function is ETotal ¼

Z1

2

jxðtÞj dt ¼

1

Z1

e2 at dt ¼

1 : (4:5:27b) 2a

0

XðjoÞ ¼ t

sinððo=2ÞtÞ ; ððo=2ÞtÞ

1 1 t2 sin2 ððo=2ÞtÞ : jXðjoÞj2 ¼ 2p 2p ððo=2ÞtÞ2

The total energy and the energy contained in the frequency range fc 5f5fc of the pulse are 2

ETotal ¼ ð1Þ t ¼ t; Efc ¼

ETotal ¼

1 2p

Z1

o 1 1 ¼ 1 : jXðjoÞj2 do ¼ tan1 2pa a 1 2a

1

(4:5:27c) The above two equations validate Rayleigh’s energy theorem.

Zfc

t2

fc

sin2 ðpftÞ ðpftÞ2

df: (4:5:29)

Using the change of variable b ¼ ft; df ¼ db=t; and f ¼ fc ! b ¼ fc t, the energy contained in the frequency range fc 5f5fc can be computed and the ratio of this to the total energy contained in the pulse. These follow

Efc ¼ 2t

Zfc t 0

We can make use of the frequency function to determine the energy as well and is

(4:5:28)

Efc ¼2 ETotal

Zfc t 0

sin2 ðpbÞ ðpbÞ2

sin2 ðpbÞ ðpbÞ2

db;

db:

(4:5:30a)

We can only compute this integral numerically. In the case of fc t ¼ 1, we have ðEfc =ETotal Þ :9028:

(4:5:30b)

136

4 Fourier Transform Analysis

That is, approximately 90% of the energy is contained in the spectral main lobe of the signal. If we include the side lobes, more energy will be included and Efc !1 ¼ ETotal . Ninety percent of the energy is reasonably sufficient to represent a rectangular pulse. & An interesting formula can be derived to find the energy of a causal signal xðtÞ, i.e., xðtÞ ¼ 0; t50 in terms of its real and imaginary parts of its transform XðjoÞ ¼ RðoÞ þ jIðoÞ. The energy is given by Papoulis (1977) as follows: Z1

x2 ðtÞdt ¼

2 p

0

Z1

R2 ðoÞdo ¼

2 p

0

Z1

¼

1 Z1

xðtÞxðt þ tÞdt

xðtÞxðt tÞdt¼ Rx ðtÞ:

xðtÞhðt þ tÞdt ¼

1

(4:6:1)

Z1

Rhx ðtÞ ¼

1

Gx ðoÞe jot do ¼F1 ½Gx ðoÞ ;

1 FT

(4:6:4)

Note that the autocorrelation function is the integral of the product of two functions, the function and its shifted version. It is a function of t, which is the shift between the given function and its shifted version. The Fourier transform pair relationship in (4.6.4) is referred to as the Wiener–Khintchine theorem. Also, we should note that a function xðtÞ and its delayed or advanced version xðt t0 Þ have the same autocorrelation function and therefore they have the same energy spectral densities. That is, Ry ðtÞ ¼

Z1 1 Z1

xðt t0 Þxðt t0 þ tÞdt

xðaÞxða þ tÞdt ¼ Rx ðtÞ

(4:6:5)

Gy ðoÞ ¼ Gx ðoÞ:

(4:6:6)

xða tÞhðaÞda;

Correlations were expressed in terms of convolution (see (2.6.4a and b)). Now, Rx ðtÞ ¼ xðtÞ xðtÞ:

(4:6:7)

Using the convolution theorem F½xðtÞ ¼ XðjoÞ ¼ X ðjoÞ, it follows that

and

1

(4:6:2a) Z1

Z1

1 Rx ðtÞ ¼ 2p

1

Note the single subscript in the case of autocorrelation and a double subscript in the case of cross-correlation below. The cross-correlations (see (2.6.3)) are

Rxh ðtÞ ¼

Rx ðtÞejot dt;

1

¼

1

Z1

Z1

Rx ðtÞ !Gx ðoÞ:

In this section we will see that the inverse Fourier transform of the energy spectral density discussed in the last section is the autocorrelation (AC) function defined in Chapter 2 (see (2.7.1)). The AC function of a real function xðtÞ is Rxx ðtÞ ¼

Gx ðoÞ ¼

0

4.6 Autocorrelation and CrossCorrelation

(4:6:3)

Cross-correlation reduces to the autocorrelation when hðtÞ ¼ xðtÞ. The Fourier transform of the AC function is the energy spectral density and is

I2 ðoÞdo: (4:5:31)

This can be shown by noting xðtÞ ¼ 2xe ðtÞuðtÞ ¼ 2x0 ðtÞuðtÞ and then using the transforms of real and imaginary parts. Details are left as an exercise.

Z1

Rhx ðtÞ ¼ Rxh ðtÞ:

hðtÞxðt þ tÞdt ¼

Z1

F½Rx ðtÞ ¼ F½xðtÞ xðtÞ ¼ F½xðtÞF½xðtÞ ¼ jXðjoÞj2 ¼ Gx ðoÞ:

(4:6:8)

hðb tÞxðbÞdb;

1

(4:6:2b)

Example 4.6.1 Show that the energy spectral densities of xðtÞ and xðt t0 Þ are the same and therefore

4.6 Autocorrelation and Cross-Correlation

137

the autocorrelation functions of the two functions are identical (see (4.6.5) and (4.6.6)). FT

Solution: Noting xðt t0 Þ !e jot0 XðjoÞ, we have e jot0 XðjoÞejot0 X ðjoÞ ¼ jXðjoÞj2 ¼ Gx ðoÞ: (4:6:9)

fx ðtÞ ¼

The energy spectral density and its inverse transform are (see (4.3.20)) jXðjoÞj2 ¼

The corresponding autocorrelation function is given by h i Rx ðtÞ ¼ F1 ½Gx ðoÞ ¼ F1 jXðjoÞj2 : (4:6:10) &

This example illustrates that the autocorrelation function Rx ðtÞ does not have the phase information contained in the function xðtÞ. This can be seen from the fact that t0 is not in either of the expressions Rx ðtÞ or Gx ðoÞ. The autocorrelation function is even and its spectrum, the energy spectral density, is real and even and Ex ¼ Rx ðtÞjt¼0 ¼

Z1

Z1

xðtÞxðtþtÞdtjt¼0 ¼

1

1 ajtj e : 2a

1

F

1 1 ¼ ; ða þ joÞða joÞ a2 þ o2

1 1 ¼ eajtj ¼ Rx ðtÞ: a2 þ o2 2a (4:6:12b)

b. By using the AC function and the energy spectral density, the energy in xðtÞ is Ex ¼ Rx ðtÞjt¼0 ¼ 1=2a, 1 p

Z1

1 1 1 1 o 1 tan ¼ ¼ Ex : do ¼ a2 þ o2 ap a 0 2a

0

(4:6:12c) & The cross-correlation signals is

jxðtÞj2 dt:

theorem

for

FT

Rhx ðtÞ !H ðjoÞXðjoÞ:

aperiodic (4:6:13)

1

(4:6:11a) This gives the energy in the signal. Using the Wiener–Khintchine theorem, we have

It can be shown by

Rhx ðtÞ ¼

Z1

hðtÞxðt þ tÞdt

1

Ex ¼ Rx ðtÞjt¼0

1 ¼ 2p

Z1

2

jXðjoÞj do:

1 ¼ 2p

(4:6:11b)

1

1 ¼ 2p Example 4.6.2 Consider the pulse function and its transform given by

¼

FT

xðtÞ ¼ eat uðtÞ !1=ða þ joÞ ¼ XðjoÞ; a40: (4:6:12a) a. Give the expression for the energy spectral density and its inverse transform, the corresponding autocorrelation function. b. Compute the energy in xðtÞ using its AC function and its energy spectral density. Solution: a. From Example 2.7.1 and (2.7.10), we have

1 2p

Z1

Z1 hðtÞ½

1 Z1

XðjoÞe joðtþtÞ dodt

1

Z1 XðjoÞ½

1 Z1

hðtÞe jot dte jot do

1

H ðjoÞXðjoÞe jot do:

1

Comparing these, (4.6.13) follows. For t ¼ 0, Rhx ð0Þ ¼

Z1 1

1 hðtÞxðtÞdt ¼ 2p

Z1

H ðjoÞXðjoÞdo:

1

(4:6:14) Equation (4.6.14) is a generalized version of Parseval’s theorem. The cross-correlation theorem

138

4 Fourier Transform Analysis

reduces to the case of autocorrelation by replacing HðjoÞ by XðjoÞ in (4.6.13).

Example 4.6.3 Consider the harmonic form of Fourier series of a periodic function

4.6.1 Power Spectral Density

xT ðtÞ ¼ Xs ½0 þ

N X

d½k cosðko0 t þ y½kÞ; o0 ¼ 2p=T:

k¼1

Earlier we have studied the power signals that include periodic signals and random signals. We will not be discussing random signals in this book in any detail. The autocorrelation of a periodic signal xT ðtÞ is given by Z 1 xT ðtÞxT ðt þ tÞdt: (4:6:15a) RT;x ðtÞ ¼ T

(4:6:16) Find its auto correlation. Solution: The autocorrelation function is given by (see (2.8.9)) RT;x ðtÞ ¼ X2s ½0 þ

T

Note that the above integral is over any one period. We have seen in Chapter 2 that the autocorrelation function of a periodic function is also a periodic function with the same period. The Fourier transform of the autocorrelation function is called the power spectral density (PSD) and its inverse is the autocorrelation function. The autocorrelation function and the corresponding spectral density function form a Fourier transform pair Sx ðoÞ ¼ F Rx;T ðtÞ ; Rx;T ðtÞ ¼ F1 ½Sx ðoÞ; FT

Rx;T ðtÞ !Sx ðoÞ:

(4:6:15b)

This relation is referred to as the Wiener–Khintchine theorem for periodic signals. The power spectral density of a periodic signal can be determined from Z1 1 Sx ðoÞdo ¼ Rx; T ð0Þ 2p 1 Z 1 xðtÞxðt þ tÞdtjt¼0 : ¼ T

N 1X d2 ½k cosðko0 tÞ: 2 k¼1

(4:6:17)

The phase terms y½ks are not in the autocorrelation function. The PSD is Sx ðoÞ ¼ 2pX2s ½0dðoÞ þ

N pX d2 ½kfdðo ko0 Þ þ dðo þ ko0 Þg: 2 k¼1

(4:6:18) The AC function and the PSD do not have any phase information and the frequencies are located at o ¼ ko0 ; k ¼ 1; 2; :::; N. The average power can be computed from the AC function or from the power spectral density. From (4.6.17), we have the average power Px ¼ RT;x ð0Þ ¼ X2s ½0 þ

N 1X d2 ½k: 2 k¼1

(4:6:19)

Using the power spectral density, we have

P¼

(4:6:15c)

T

Formal proof of the general Wiener–Khintchine theorem is beyond the scope here (see Ziemer and Tranter, 2002 and Peebles, 2001). In the following we will assume that Sx ðoÞ is given by the transform of the periodic autocorrelation function. Notes: Power signals include periodic and random signals. The autocorrelation function of a periodic function is periodic and the power spectral density & contains impulses.

1 Px ¼ 2p

Z1

Sx ðoÞdo

1

2p ¼ 2p

Z1

fX ½0dðoÞ þ 14 X d ½k:fdðo ko Þ N

2 s

2

1

þ dðo þ ko0 Þg

0

k¼1

gdo:

Since the integrand contains only impulses, the integral can be evaluated by inspection and is & RT;x ð0Þ in (4.6.19). Notes: In the case of energy signals the square of the magnitude spectrum gives the energy spectral

4.7 Bandwidth of a Signal

139

density (ESD), and the energy content is obtained by integrating the ESD. In the case of periodic signals, the spectrum contains impulses and the square of an impulse function is not defined. The AC function and the Wiener–Khintchine theorem are used to compute the power spectral density (PSD) of the periodic signal and its integral is the average power contained in the power signal. Most signals are corrupted by noise. Autocorrelation function ‘‘cleans’’ the signal and it provides a better insight into the essential qualities of the signal. Note that the AC function does not have the phase information in the signal. Since the AC of a periodic function is also a periodic with the same period, it can be determined by identifying the peaks in the autocorrelation function, thereby identifying the fundamental frequency. Finding the pitch period of a vowel in a noisy speech signal by using the autocorrelation function is very effective on a short& time basis (see Rabiner and Schafer, 1978).

4.7 Bandwidth of a Signal In Section 4.2.2, bandwidth (BW) of a signal was discussed in simple terms. One definition is the range of positive frequencies in which most of the signal energy or power is contained. This is vague since the word ‘‘most’’ can be interpreted differently. We will consider this here in more detail. A signal is considered time limited if the signal is zero outside an interval. For example, a pulse function P½ðt t0 Þ=t is nonzero for jt t0 j5t=2 and zero outside this range. It is nonzero for t s. The signals considered in this book are real signals and their spectral amplitudes are even and the spectral phase angles are odd. A signal xðtÞ is said to be band limited to B Hz if jXðjoÞj ¼ 0;

o ¼ j2pfj4W ¼ 2pB:

(4:7:1)

Since it is band limited to B Hz, B is defined as the bandwidth of the signal. In this case the signal occupies only low frequencies, i.e., it is a lowfrequency signal or sometimes referred to as a lowpass signal. Note that the bandwidth is defined using only positive frequencies. Band-pass signals are common in communication theory. Band-pass spectrum can be defined as follows:

8 0; joj5o0 ðW=2Þ > > > < H ; jo o j5ðW=2Þ 0 0 : jXðjoÞj ¼ > H0 ; jo þ o0 j5ðW=2Þ > > : 0; joj4o0 þ ðW=2Þ

(4:7:2)

It is an ideal band-pass signal. Most practical signals are not band limited. There are functions that are neither time limited nor band limited. For example, consider the double exponential function given below and its transform derived earlier FT xðtÞ ¼ eajtj !2a= a2 þ o2 ¼ XðjoÞ; a40: (4:7:3) The question we need to answer is, what is a meaningful definition of the time width of an arbitrary nontime-limited signal? How about a meaningful definition of the frequency width of an arbitrary non-bandlimited signal? In Section 4.2.2, we have seen that a shorter time width signal corresponds to a broader spectrum. For example, the Fourier transform pair of a rectangular pulse function is given by xðtÞ ¼ P

hti t

sinðot=2Þ !t ðot=2Þ ¼ XðjoÞ:

FT

(4:7:4)

Most of the energy is contained in the main lobe of the spectrum, which occupies the frequency band between the two zeros of XðjoÞ located at o ¼ 2p=t. The energy content of this pulse is quantified in terms of the frequency content in Example 4.5.12. The side lobes contain a small portion of the energy. The time width of the pulse is obviously equal to t s and the frequency width is approximately 1=t Hz, considering only positive frequencies. Increasing (decreasing) the time width reduces (increases) the frequency width. At least, from this example, we see that the two widths are inversely proportional to each other. In the following we will consider a few standard definitions of time and frequency widths and they give some meaning.

4.7.1 Measures Based on Areas of the Time and Frequency Functions Using the ordinate theorems discussed earlier, we have

140

4 Fourier Transform Analysis

Z1 Xð0Þ¼

xðtÞdt; xð0Þ¼ 1

Z1

1 2p

XðjoÞdo: (4:7:5) 1

The dc value of a signal is zero if it is an odd function. Interestingly, n

xðtÞ ¼ d yðtÞ=dt

1 FT ¼ XðjoÞ; a40: xðtÞ ¼ eat uðtÞ ! a þ jo

n FT

n

!ðjoÞ YðjoÞ ¼ XðjoÞ:

(4:7:6)

It is zero at o ¼ 0 provided YðjoÞ has no poles at the origin that can be canceled by ðjoÞn . If the time function has a discontinuity at the origin, then xð0Þ is obtained from the integral in (4.7.5) and is the average value or the half-value at the discontinuity. Let the time width and the frequency widths, respectively, be defined by R1

R1

xðtÞdt ; tw ¼ 1 xð0Þ

ow ¼ 1

XðjoÞdo : Xð0Þ

1 Ð 0

Z1

1 a: xð0Þ ¼ 2p ¼

1 2p

1 Z1

eat dt ¼ 1a.

XðjoÞe jot dojt¼0 a a2 þ ð o Þ

1

do þ 2

j 2p

Z1 1

ðoÞ a2 þ ðoÞ2

do:

The integrand in the second integral is an odd function and therefore it is zero: 1 xð0Þ ¼ 2p

Z1

a 1 1 tan1 ðo=aÞ1 do ¼ 1 ¼ : a2 þ o2 2p 2

1

(4:7:7)

Then the product tw ow ¼ 2p:

Solution: a. Xð0Þ ¼

(4:7:8)

That is, the product of time and frequency widths is a constant. Some authors use the frequency f in Hz rather than o ¼ 2pf in rad/s in the time–bandwidth product. These simple measures have drawbacks illustrated below. Example 4.7.1 Consider the function given by xðtÞ ¼ P½t :5 P½t þ :5. It is an odd function. If the above definition is used, the time width is zero, even though the actual width of this function & is 2 s. Example 4.7.2 Consider the pair

Note the exponential time-decaying function xðtÞ is discontinuous at t ¼ 0 and the above result verifies that the inverse transform converges to the halfvalue, i.e., the average value of the function before & and after the discontinuity.

4.7.2 Measures Based on Moments The time width Tw of a real non-time-limited function is defined by ðTw Þ2 ¼

Z1

1 kxk

Ex ¼ kxk2 ¼

2

ðt tÞ2 x2 ðtÞdt; kxk2 ¼

1

Z1

The area under the pulse function is 1. The area under the sinc function is 2p and the time–band& width product is 2p. Example 4.7.3 Consider the Fourier transform pair corresponding to the exponential decaying function to find the values of the functions a. Xð0Þ and b. F1 ½XðjoÞjt¼0 :

x2 ðtÞdt;

1

x2 ðtÞdt ¼

1

sinðo=2Þ : P½t ! ðo=2Þ FT

Z1

1 2p

Z1

(4:7:9) jXðjoÞj2 do51:

1

(4:7:10) The center of gravity of the area of the function is 1 t ¼ Ex

Z1

tx2 ðtÞdt:

(4:7:11)

1

Tw is a measure of the signal spread about t and is the signal dispersion in time. Notes: These measures can be seen noting pðtÞ ¼ ½x2 ðtÞ=Ex >0 is a valid probability density

4.7 Bandwidth of a Signal

141

function, as is nonnegative for all t and the area under it is 1. In statistical terminology t is the mean & and ðTW Þ2 is the variance (see Peebles, 2001). Example 4.7.4 Consider the exponential decaying function xðtÞ ¼ eat uðtÞ; a40. Find the center of gravity and the time dispersion Tw . Solution: The energy contained in the pulse is E ¼ ð1=2aÞ. The center of gravity is Ð 1 2 at te dt e2 at ½1 þ 2 at 1 1 ¼ 2a t ¼ 0 : 0 ¼ 2 ð1=2aÞ 2a ð2aÞ

A bound on the time–bandwidth product Tw Ww ¼ Tw ð2pFw Þ is derived using Rayleigh’s energy theorem and Schwarz’s inequality (see Section 2.1.). The inequality is briefly reviewed below. Schwarz’s inequality: The inequality is (see (2.1.9d)) khxðtÞþyðtÞik kxðtÞkkyðtÞk Zb Zb Zb 2 2 ) jxðtÞyðtÞj dt jxðtÞj dt jyðtÞj2 dt: a

a

a

(4:7:15)

(4:7:12) In addition, using the following integral formulas, Tw is as follows:

t2 2t 2 2þ 3 ; b b b

Z Z 1 1 bt bt t ; and ebt dt ¼ ebt ; te dt ¼ e b b2 a

Z

t2 ebt dt ¼ ebt

4.7.3 Uncertainty Principle in Fourier Analysis

The uncertainty principle in spectral analysis states that if the integrals in (4.7.9) and (4.7.14a) are finite and pﬃﬃ lim txðtÞ ¼ 0; (4:7:16) t!1

ðTw Þ2 ¼2a

Z1

then ðtð1=2aÞÞ2 e2at dt

Tw Ww

0

Z ¼ð2aÞ

1

½t2 ð1=aÞtþð1=2aÞ2 e2at dt¼ð1=2aÞ2

0

)Tw ¼ð1=2aÞ:

(4:7:13)

The frequency width, a frequency measure, Ww ¼ 2pFw , can be defined by

W2w

¼

kXk2

1 ¼ Ex ð2pÞ Note : o

k X k2 ¼

Z1

1

Z1

1 1 or Tw Fw

; Ww ¼ 2pFw : 2 2ð2pÞ

(4:7:17) Using the expressions for Tw and Ww from (4.7.9) and (4.7.14a) results in ðTw Ww Þ2 ¼

2 jXðoÞj2 do ðo oÞ

3

Z1

4 ðt tÞ2 x2 ðtÞdt5 kx k 2 kX k2 1 2 1 3 Z o2 jXðjoÞj2 do5;

4 1

1

R1

2

1

2 ojXðjoÞj do ¼ 0 ; (4:7:14a)

(4:7:18)

1 2

kX k ¼

jXðjoÞj2 do ¼ 2pkxk2

Z1

2

2

jXðjoÞj do; kxk ¼

1

Z1

2 x ðtÞdt;

1

1

ðRayleigh’s energy theoremÞ:

(4:7:14b)

This follows since the integrand in the above equation is odd and the integral of an odd function over & a symmetric interval is zero.

kXk2 ¼ 2pkxk2 :

(4:7:19)

Noting the Fourier transform derivative theorem, i.e., F½x0 ðtÞ ¼ ðjoÞXðjoÞ and using Rayleigh’s energy theorem results in

142

4 Fourier Transform Analysis

Z1

1 jx ðtÞj dt ¼ 2p 2

0

1

Z1

2

2

o jXðjoÞj do: (4:7:20)

Z1

1 kxk2 kXk2 1 2

kxk kXk

¼

kxk4

2 ðt tÞ2 x2 ðtÞdt4

1 1 ð

Z1

3 o2 jXðjoÞj2 do5

1

ðt tÞ2 x2 ðtÞdtð2pÞ

2

Z1

1

1

Z1

Z1

2

eat dt¼

1

ðt tÞ2 x2 ðtÞdt

1

½x0 ðtÞ2 dt:

(4:7:21)

1

Canceling the negative signs in (4.7.25b) results in Z1

rﬃﬃﬃﬃﬃ p dt ¼ : 2a

(4:7:26)

From tables,

1

Z1

x2 ðtÞdt:

2

k xk ¼

Z1

1

(4:7:23)

2 at2

t e

t!1

1 x2 ðtÞdt ¼ kxk2 : 2

e

2 at2

Now let a ¼ 2a in (4.7.25c), which results in

Assuming that lim ðt tÞx2 ðtÞ ¼ 0, it follows that Z1

Z1 1

1

1

(4:7:25c)

(4:7:22)

dx 1 1 dt ¼ ðt tÞx2 ðtÞ1 1 dt 2 2

1 2

rﬃﬃﬃ p : a

1 ¼ 2a

1

Considering the right-hand side of the above equation and integrating it by parts, we have Ð Ð udv ¼ uv vdu, with u ¼ ðt tÞ; du=dt ¼ 1; dv=dt ¼ xðtÞx0 ðtÞ; and v ¼ ð1=2Þx2 ðtÞ

ðt tÞxðtÞx0 ðtÞdt ¼

2 at2 dt

t e

1

1

Z1

(4:7:25b)

1

2 1 Z 0 ðt tÞxðtÞx ðtÞdt :

ðt tÞxðtÞ

1

pﬃﬃﬃ dða1=2 Þ pﬃﬃﬃ a1:5 ¼ p ¼ p 2 da rﬃﬃﬃ 1 p : ¼ 2a a

½x0 ðtÞ2 dt

1

Z1

rﬃﬃﬃ Z1 Z1 at2 p de 2 dt ¼ ) ðt2 Þeat dt da a

1

Using Schwarz’s inequality in (4.7.15) results in 2 1 32 1 3 Z Z 4 ðt tÞ2 x2 ðtÞdt54 ½x0 ðtÞ2 dt5

Z1

rﬃﬃﬃ p ðoÞ2=4a ¼ XðjoÞ; a40: (4:7:25a) ! ae

FT

Solution: First, differentiate both sides of the following equation with respect to a:

ðTw Ww Þ2

¼

2

1

Therefore

¼

xðtÞ ¼ eat

1

dt ¼

Z1

2 2 at2 dt

t e

1 ¼ 4a

rﬃﬃﬃﬃﬃ p : 2a

1

Noting that the Gaussian pulse in this example is even and the integrand in (4.7.11) is odd, it follows that t ¼ 0. The time width can be computed from (4.7.9) and

Using this and (4.7.21) in (4.7.20), we have 2

4

1 1 kxk ¼ or Tw Ww

2 4kxk4 4 1 ; Ww ¼ 2pFw : Tw F w

2ð2pÞ

ðTw Ww Þ2

ðTw Þ ¼

or

1 kxk2

Z1

t2 e2 at dt ¼ ð1=

1

(4:7:27)

(4:7:24)

Example 4.7.5 Illustrate the uncertainty principle using the Gaussian transform pair

rﬃﬃﬃﬃﬃ rﬃﬃﬃﬃﬃ p 1 p 1 Þ ¼ : 2a 4a 2a 4a

Noting that kXk2 ¼ 2pkxk2 ¼ 2p that

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ p=2a, it follows

4.8 Moments and the Fourier Transform

ðWw Þ2 ¼

¼

Z1

1 kXk2

143

will derive the transform in terms of mi by using the power series expansion

o2 jXðjoÞj2 do

1

Z1

1 kxk2 ð2pÞ

ejot ¼

p 2ðoÞ2 o2 e 4a do a

1

pﬃﬃﬃﬃﬃ Z1 2a p o2 pﬃﬃﬃð Þ ¼ o2 e 2a do ð2pÞ p a 1 pﬃﬃﬃﬃﬃ 2a p pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃ a pð2aÞ ¼ a ) ðTw Ww Þ2 ¼ ð2pÞ pa 1 1 (4:7:28) ¼ a¼ : 4a 4

1

1

¼

1

The moment theorem relates the derivatives of the transform of a function at o ¼ 0:

m0 ¼

Z1

n ¼ 0; 1; 2; ::: ;

(4:8:2)

Z1

dXðjoÞ jo¼0 ¼ do

xðtÞ

dejot dtjo¼0 do

1 Z1

¼j

n!

xðtÞdt

tn xðtÞdt

1

(4:8:4)

This holds only if the integral of the terms in the above equation is valid. From (4.8.2) XðjoÞ ¼

1 X

ðjÞn mn ½on =n!:

(4:8:5)

n¼0

Although the moment theorem is given in terms of a series expansion, it can be used to compute the transforms of functions (see Papoulis, 1962). Example 4.8.1 Use the moment theorem and the following identity to show that Fourier transform of the 2 Gaussian pulse xðtÞ ¼ eat is also a Gaussian pulse. Z1

at2

rﬃﬃﬃﬃﬃ rﬃﬃﬃ Z1 p 1 p 2 ; a40: dt ¼ t2 eat dt ¼ ) a 2 a3

1

1

(4:8:6a)

xðtÞdt ¼ XðjoÞjo¼0 (Ordinate theorem);

1

Z1 1 X ðjoÞn n¼0

e d n XðjoÞ ðjÞ mn ¼ jo¼0 ; do n n

n!

n¼0

1 n X d XðoÞ on : ¼ j o¼0 n! don n¼0

4.8 Moments and the Fourier Transform The nth moment mn of xðtÞ is defined by (see Section 1.7.) Z1 tn xðtÞdt; n ¼ 0; 1; 2; ::: : (4:8:1) mn ¼

Z1 X 1 ðjoÞn tn

¼

(4:7:29)

This shows the equality in (4.7.24) in the Gaussian & case. See Hsu (1967) for additional examples.

(4:8:3)

Substituting this in the transform and using (4.8.1) and (4.8.2) result in Z1 XðjoÞ ¼ F½xðtÞ ¼ xðtÞejot dt

The time–bandwidth product of a Gaussian pulse is obtained by using Ww ¼ 2pFw and Tw Fw ¼ 1=ð2ð2pÞÞ:

1 X 1 ðjotÞn : n! n¼0

Solution: Equation (4.8.6a) on the right can be generalized and 1 ð 1

txðtÞe

jot

dtjo¼0 ¼ jm1 :

1

Repeating this process and evaluating the derivatives at o ¼ 0 proves the result in (4.8.2). Now we

2

rﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ p a2nþ1 rﬃﬃﬃ 1:3:::ð2n 1Þ p ¼ ¼ m2n : ð2aÞn a

t2n eat dt ¼

1:3:::ð2n 1Þ 2n

(4:8:6b)

This gives even moments and the odd moments are zero since xðtÞ is even:

144

4 Fourier Transform Analysis

XðjoÞ ¼ ¼

1 X n¼0 1 X

ðjoÞ2n

m2n ð2nÞ!

ð1Þn ðoÞ2n

n¼0

1 1:3:::ð2n 1Þ ð2nÞ! ð2aÞn

rﬃﬃﬃ p : a

jXðjoÞj

This can be expressed in a compact form by noting ð1Þð3Þð5Þ:::ð2n 1Þ ð1Þð3Þð5Þ:::ð2n 1Þ ¼ n n ð1=2 Þð2nÞ! ð1=2 Þð1Þð2Þð3Þ:::ð2n 1Þð2nÞ 1 ¼ ð1=2n Þð2Þð4Þ:::ð2nÞ 1 1 ¼ ; ¼ ð1Þð2Þ:::ðnÞ n!

a ¼ XðjoÞ; a40; ½a þ jo

it follows that

ða þ joÞ2

jo¼0

1 Z Z1 jot xðtÞejot dt xðtÞe dt jXðjoÞj ¼ 1

¼

Z1

1

jxðtÞjdt: FT

1 joj

Z1 dxðtÞ dt dt:

1

In a similar manner, we can prove the third bound in (4.9.2) if it exists. Example 4.9.1 Find the first two bounds in (4.9.2) using the Fourier transform pair FT

&

4.9 Bounds on the Fourier Transform In Chapter 3, we have learned that the derivative of a periodic function plays a role on the bounds on its F-series coefficients. We can use the Fourier time differentiation theorem to find the bounds on the transform. First d n xðtÞ FT n !ðjoÞ XðjoÞ: dtn

First bound:

xðtÞ ¼ P½t !½sinðo=2Þ=ðo=2Þ ¼ XðjoÞ:

m0 ¼ Xð0Þ ¼ 1; m1 ¼ j ja

(4:9:2)

1

Solution: Noting

¼j

jxðtÞjdt

Second bound: With x0ðtÞ !joXðjoÞ, we have Z1 dxðtÞ jot ðjoÞXðjoÞ ¼ e dt ! jXðjoÞj dt &

Example 4.8.2 Find the first two moments of xðtÞ ¼ aeat uðtÞ; a40.

dXðjoÞ jo¼0 do a 1 ¼ 2¼ : a a

1 Ð

1 1 Ð dxðtÞ 1 dt dt joj > 1 > > > 1 > Ð d 2 xðtÞ > > : o12 dt2 dt 1

1

rﬃﬃﬃ 1 p X ð1Þ2 o2n XðjoÞ ¼ a n¼0 ð4aÞn n! pﬃﬃﬃﬃﬃﬃﬃﬃ 2 2 FT pﬃﬃﬃﬃﬃﬃﬃﬃ 2 ¼ p=aeo =4a ) eat ! p=aeo =4a :

F½xðtÞ ¼ F½aeat uðtÞ ¼

8 > > > > > > >

s = s + jo-plane, contour of integration, Example 5.4.1

X2;II ðsÞ ¼

Z0

x2 ðtÞe

st

dt ¼

1

transform and can be obtained from the Laplace transform by simply substituting s ¼ jo in the Laplace transform. The example we have consid& ered above exists for positive time. A two-sided signal can be written in two parts – one for positive time (causal) including t ¼ 0 and the other part for negative time (anti-causal). The twosided function xðtÞ can be separated into two parts x1 ðtÞ and x2 ðtÞ, representing the causal and the anti-causal part, respectively. See Fig. 5.4.2, where we have the following: xðtÞ : twosided;

x2 ðaÞest da

0

x2 ðaÞest da

0

) X2;II ðsÞ ¼

Z1

x2 ðtÞest dt:

(5:4:6b)

0

That is, we have expressed the two-sided Laplace transform in terms of two one-sided Laplace transforms. Changing the sign of s in the function X2;II ðsÞ gives X2;II ðsÞ. Notes:

x1 ðtÞ ¼ xðtÞuðtÞ : causal part x2 ðtÞ ¼ xðtÞuðtÞ : anticausalpart; x2 ðtÞ ¼ xðtÞuðtÞ : invertedanticausalpart x1 ðtÞ and x2 ðtÞ are now causal signals. The bilateral or two-sided transform is given by

Fig. 5.4.2 (a) Two-sided signal, (b) causal, (c) anticausal, and (d) inverted anticausal parts

¼

Z1

Z1

1. Express xðtÞ ¼ ½xðtÞuðtÞ þ ½xðtÞuðtÞ ¼ x1 ðtÞ þ x2 ðtÞ. 2. Use the causal signals x1 ðtÞ and x2 ðtÞ and find their one-sided transforms. 3. The two-sided Laplace transform of the signal is & given by XðsÞ ¼ X1 ðsÞ þ X2 ðsÞ.

(a)

(b)

(c)

(d)

5.4 Laplace Transforms

163

Example 5.4.2 Find the two-sided Laplace transform of the function ( xðtÞ ¼

)

ebt ;

b40; t50

eat ;

a40; t 0

(5:4:7)

¼ ½ebt uðtÞ þ ½eat uðtÞ ¼ x2 ðtÞ þ x1 ðtÞ

To see these constraints, we can separate the bilateral Laplace transform into two integrals, one for positive time and the other for negative time. That is, Z 1 xðtÞest dt XII ðsÞ ¼ LII ½xðtÞ ¼ ¼

0

xðtÞe

Solution: From Example 5.4.1, we have 1 ðs þ aÞ ¼ X1 ðsÞ; ReðsÞ ¼ s4 a:

1

Z

st

dt þ

Z

1

LTII

x1 ðtÞ ¼ eat uðtÞ !

(5:4:8a)

1 ðs þ bÞ ¼ X2 ðsÞ; s ¼ ReðsÞ4 b; LTII

xðtÞest dt:

¼

Z

Z

0 bt st

Me e 1 Z 0

dt þ

MeðbsÞt dt þ

1

0 Z 1

1

Meat est dt MeðaþsÞt dt:

0

(5:4:10c)

LTII

x2 ðtÞ ¼ ebt uðtÞ ! X2 ðsÞ 1 ¼ ; s ¼ ReðsÞ5b: sb

(5:4:8b)

Combining the two equations in (5.4.8a) and (5.4.8b), we have LTII

xðtÞ !

(5:4:10b)

0

The integrals in the above equation must be absolutely integrable in order for the transform to exist. Using (5.4.10a), we have XII ðsÞ

x2 ðtÞ ¼ ebt uðtÞ !

1

1 1 þ ; a5s5b: ðs bÞ ðs þ aÞ

Noting that limits on the integration, we can state that the transform exists if ðb sÞ40;

ða þ sÞ40 ! a5s5b:

(5:4:11)

This defines the ROC and is illustrated by the dark area in Fig. 5.4.3.

(5:4:9)

Note the constraints on s. The transform does not exist if ðaÞ4b. This will be tied to the region of convergence shortly. From this example the causal signal results in the transform that has poles on the left half of the s-plane and the anti-causal signal results in the transform that has poles on the right & half of the s-plane.

jω

Poles for

Poles for

x (t )u ( −t )

x (t )u (t ) −α

0

β

σ

5.4.1 Region of Convergence (ROC) We defined the two-sided Laplace transform of a signal xðtÞ by assuming that the function (xðtÞest ) is absolutely convergent. This implies that there exists a pair of constants a and b and M a real positive number, such that Meat ; t40 : (5:4:10a) xðtÞ Mebt ; t50

Fig. 5.4.3 Region of convergence for bilateral Laplace transform

Now we will relate this to a rational Laplace transform function XII ðsÞ in terms of its poles and zeros. We say that XII ðsÞ has a pole at s ¼ pi if XðsÞs¼pi ¼ 1 and has a zero at s ¼ zi

164

5 Relatives of Fourier Transforms

if XðsÞjs¼zi ¼ 0. The poles of the function XðsÞ must lie to the right of the line s ¼ b þ jo for t50, whereas the poles corresponding to the positive time, i.e., xðtÞ; t40 must lie to the left of the line s ¼ a þ jo in the complex s-plane. Example 5.4.3 Find the two-sided Laplace transform of the following function directly: xðtÞ ¼ ea j tj ¼ eat uðtÞ þ eat uðtÞ; a40:

(5:4:12)

Solution: The transform of this function is XðsÞ ¼

Z

0

eðasÞt es t dt þ 1

¼

Z

1

eðaþsÞ t dt

0

1 1 2a þ ¼ 2 ; a5s5a: ða sÞ ðs þ aÞ a s2 (5:4:13) &

The region of convergence is indicated on the right of (5.4.13). The Laplace transform does not converge at the pole locations and in Fig. 5.4.1 we have shown that the region of convergence goes around the poles (see the half moons around the poles). The region of convergence includes the jo-axis and the function evaluated on the jo-axis gives the Fourier transform of the function. That is, XðjoÞ ¼ XII ðsÞs¼jo :

XII ðsÞ ¼

Z

1

eat est dt ¼

0

YII ðsÞ ¼

Z

1 ; sþa

0

eat est dt ¼

1

1 ; ReðsÞ5 a: sþa (5:4:15)

We can see that the transforms are the same, except that they have different regions of convergence. These indicate the ambiguity in identifying the time function that the transform came from if the region of convergence of a two-sided Laplace transform is not given. In this case it is not possible to compute the corresponding time function. However, in practice, the ambiguity can be resolved on the basis of physical considerations, as the time functions increase without limit as t approaches either þ1 or 1. This gives a way to select a time function among & many possibilities. Procedure to find the two-sided inverse LT of a rational function: 1. Expand the given rational transform function by using partial fraction expansion. 2. The terms in the partial fraction expansion that come from the left half-plane poles will result in time functions that exist only for t 0. 3. The terms in the partial fraction expansion that come from the right half-plane poles will result in the time functions that exist only for t50. Example 5.4.5 Find the inverse transform of the XII ðsÞ ¼ 1=½ðs 1Þðs þ 2Þ.

5.4.2 Inverse Transform of Two-Sided Laplace Transform One needs to be careful in finding the inverse transforms of a two-sided Laplace transform, as the positive and negative time portions must be handled separately. If the region of convergence is not specified, then there is some ambiguity. Example 5.4.4 Consider the time functions xðtÞ ¼ e

at

uðtÞ;

at

yðtÞ ¼ e

uðtÞ:

(5:4:14)

Solution: The two-sided Laplace transforms of these functions are given by

Solution: The partial fraction expansion of this function is A B þ ; ðs 1Þ ðs þ 2Þ 1 1 A¼ js¼ 1 ¼ ; ðs þ 2Þ 3 1 1 B¼ js¼ 2 ¼ ðs 1Þ 3

XII ðsÞ ¼

) XII ðsÞ ¼

1=3 1=3 þ ; ðs 1Þ ðs þ 2Þ

) xðtÞ ¼ ð1=3Þet uðtÞ ð1=3Þe2t uðtÞ; 25ReðsÞ51:

(5:4:16)

(5:4:17)

(5:4:18)

5.5 Basic Two-Sided Laplace Transform Theorems

The region of convergence can be obtained by noting that the first term has a pole at s ¼ 1 and the region of convergence for this term is s ¼ ReðsÞ51. The second term has a pole at s ¼ 2 and the region of convergence for this term is s4 2. The intersection of these two regions of convergence is given & by 25ReðsÞ51. The two-sided Laplace transform of a function can be determined by decomposing the two-sided function into two one-sided functions and then transforming each one. In the case of finding the inverse Laplace transform, we first separate the transform into two parts, one with poles on the right half-plane and other with poles in the left halfplane and the imaginary axis. Then determine the two time functions. Most of our discussion on inverse transforms covers rational functions.

5.4.3 Region of Convergence (ROC) of Rational Functions – Properties 1. The ROC does not contain any poles of the function. 2. If xðtÞ ¼ 0, except in a finite interval, then the ROC is the entire s-plane except possibly s ¼ 0 and s ¼ 1. 3. If xðtÞ is right-sided, then the ROC is right-sided, i.e., s ¼ ReðsÞ4 a, where (a) is the real part of the left-most pole. 4. If xðtÞ is left-sided, then the ROC is left-sided, i.e., s ¼ ReðsÞ5b , where b is the real part of the right-most pole. 5. If xðtÞ is two-sided function, i.e., the sum of leftand right-sided functions, then the ROC is either a strip defined by a5ReðsÞ5b or the individual regions of convergence will not overlap and, in that case, the ROC is the null set.

5.5 Basic Two-Sided Laplace Transform Theorems Now consider some of the important two-sided Laplace transform theorems that are given below without proofs for most. Assume the region of convergence of xi ðtÞ is Rxi .

165

5.5.1 Linearity The Laplace transform of a sum is the sum of the Laplace transforms and can be stated as 2 2 X LTII X xðtÞ ¼ ai xi ðtÞ ! ai Xi ðsÞ (5:5:1) i¼1 i¼1 ¼ XII ðsÞ; ROC : at least Rx1 \ Rx2 Example 5.5.1 Let xðtÞ ¼ x1 ðtÞ þ x2 ðtÞ; x1 ðtÞ ¼ et uðtÞ; x2 ðtÞ ¼ et uðtÞ. Find the Laplace transform of the function xðtÞ and identify the region of convergence. Solution: XII ðsÞ ¼

1 1 þ : sþ1 s1

The ROC of x1 ðtÞ is ReðsÞ4 1 and the ROC of x2 ðtÞ is ReðsÞ51: The ROC of xðtÞ is the intersection of the two and is given by & 15ReðsÞ51.

5.5.2 Time Shift LII ½xðt t0 Þ ¼ est0 XII ðsÞ:

(5:5:2)

The region of convergence is the same for both the original and its shifted version.

5.5.3 Shift in s LII ½eat xðtÞ ¼ XII ðs þ aÞ:

(5:5:3)

Since the poles will be shifted to the left by a, the ROC will be shifted to the left by a.

5.5.4 Time Scaling LII ½xðatÞ ¼

1 XII ðs=aÞ; ðROCÞnew ¼ ðROCÞold =a: jaj (5:5:4)

Time scaling makes the ROC scaled as well.

166

5 Relatives of Fourier Transforms

5.5.5 Time Reversal LII ½xðtÞ ¼ XII ðsÞ; ðROCÞnew ¼ ðROCÞold :

(5:5:5)

Note that the right half-plane poles become left half-plane poles and vice versa.

5.5.6 Differentiation in Time LII

dxðtÞ ¼ sXII ðsÞ; ROCnew ROCold : dt

(5:5:6)

The ROC will not change unless there is a pole–zero cancellation in the product ðsXII ðsÞÞ.

5.5.7 Integration Z LII

t

1 xðaÞda ¼ XII ðsÞ: s 1

(5:5:7)

Noting the term (1/s) in the transform and ROCnew ¼ ROCold IfReðsÞ 4 0g.

5.5.8 Convolution In Chapter 2 we defined the convolution of two functions by Z 1

x1 ðaÞx2 ðt aÞda

yðtÞ ¼

¼

1 Z 1

(5:5:8a) x2 ðaÞx1 ðt aÞda:

1

The transform is Z 1 Z 1 x1 ðaÞx2 ðt aÞda est dt YII ðsÞ ¼ 1 1 Z 1 Z 1 st x1 ðaÞ x2 ðt aÞe dt da ¼ 1 Z1 1 ¼ x1 ðaÞXII;2 ðesa Þda 1 Z 1 ¼ XII;2 ðsÞ x1 ðaÞesa da ¼ XII;2 ðsÞXII;1 ðsÞ: 1

(5:5:8b)

The ROC satisfies ROCnew ðROCÞ1 \ ðROCÞ2 . The ROC of the convolution may be larger. When two transforms are multiplied, there is a possibility of pole cancellations.

5.6 One-Sided Laplace Transform So far we have been discussing the bilateral or twosided transform. A special form of the bilateral transform is the one-sided or unilateral or simply Laplace transform, which was defined in (5.4.4). It was pointed out that the bilateral transform can be computed by using the one-sided Laplace transform. In real-life systems, there is no negative time. However, bilateral transforms provide a structure that we can work with, as the bilateral Laplace transforms relate to the Fourier transforms. We can make the discussion simpler by considering the unilateral Laplace transform. The unilateral transform is fundamental in circuits, systems, and control, where we are interested in the response of a system with initial conditions. In a later chapter we will describe a linear timeinvariant system by constant coefficient differential equations. The unilateral Laplace transform provides a powerful tool in the analysis and design of systems. As mentioned earlier we will use the notation XðsÞ ¼ LfxðtÞg and xðtÞ is a causal signal. Furthermore, the ROC is the right half s-plane for the unilateral Laplace transforms. For simplicity, we generally do not explicitly identify the region of convergence. Unless otherwise mentioned, we will assume that the transform functions are unilateral and will not be mentioned explicitly. The unilateral Laplace transform of a signal xðtÞ is defined earlier and is repeated below: XðsÞ ¼ L½xðtÞ ¼

Z

1

xðtÞest dt:

(5:6:1)

0

Symbolic relation: LT

xðtÞ ! XðsÞ:

(5:6:2)

Notes: In defining the transform integral in (5.6.1), we have used the lower limit of 0 . This allows us to include signals such as the unit impulse function dðtÞ. From now on we will use the lower limit on

5.6 One-Sided Laplace Transform

167

the integral as zero except in special cases and assume that the limit is 0 . In cases where there is some ambiguity we will explicitly identify the lower limit on the integral as 0 . Some texts use the limits of integration on the Laplace integral as 0þ to infinity. This implies that origin is excluded. This approach is impractical in the theoretical study of linear systems. The Laplace transform of a function exists if ðxðtÞes t Þ is absolutely integrable. We can select the range of s that ensures the convergence and this is referred to as the region of convergence.This is one of the nice aspects of the Laplace transform. For example, the L-transform of eat uðtÞ; a40 exists only if s4a and we can select such a range. Noting that the Laplace transform is an integral operation, the transform is unique. Example 5.6.1 Find the Laplace transforms of the functions by using the definition a: x1 ðtÞ ¼ dðt t0 Þ; t0 40; b: x2 ðtÞ ¼ uðtÞ; c: x3 ðtÞ ¼ e

a t

(5:6:3)

uðtÞ; a40:

Solution: a. The Laplace transform of the impulse function is given by Z1

X1 ðsÞ ¼

dðt t0 Þest dt ¼ es t0 ; (5:6:4)

0 LT

dðt t0 Þ ! es t0 : b: X2 ðsÞ ¼

Z

1 0

1 uðtÞest dt ¼ est 1 t¼0 s

(5:6:5)

1 LT 1 ! uðtÞ ! ; ¼ s s Z 1 Z 1 c: X3 ðsÞ ¼ eat est dt ¼ eðsþaÞt dt 0

¼

Z1

t1=2 est dt:

0

Using the change of variable a ¼ t1=2 ; da ¼ ð1=2Þt1=2 dt and the integral tables result in 4 X1 ðsÞ ¼ pﬃﬃﬃ p

Z1

2

a2 es a da ¼

1 : s3=2

0

1 b: X2 ðsÞ ¼ L½sinhðtÞ ¼ 2

Z1 h

i eðsaÞt eðsþaÞt dt

0

a : ¼ 2 s a2

(5:6:7)

The two transform pairs in this example are given by rﬃﬃﬃ t LT 1 ! 3=2 ; 2 p s

LT

sinhðatÞ !

s2

a2 : a2

(5:6:8) &

As in the case of F-transforms we can derive the properties of the L-transforms. Most of the proofs are similar in these cases. We will not go through the details but will point out some of the important facets of the properties. In cases where they are different we will go through the proofs. Following is a table of some of the one-sided Laplace transforms theorems. Note that the regions of convergence are not identified.

5.6.1 Properties of the One-Sided Laplace Transform

(5:6:6) &

Example 5.6.2 Find the unilateral Laplace transforms of the following functions: rﬃﬃﬃ t ; b:x2 ðtÞ ¼ sinhðtÞ: p

a:x1 ðtÞ ¼ 2

2 a:X1 ðsÞ ¼ pﬃﬃﬃ p

Unilateral Laplace transforms properties are given in Table 5.6.1.

0

1 1 LT ; eat uðtÞ ! : sþa sþa

Solution:

5.6.2 Comments on the Properties (or Theorems) of Laplace Transforms The proof of the linearity property of the Laplace transforms is straightforward as the integral of a

168

5 Relatives of Fourier Transforms Table 5.6.1 One-sided Laplace transform properties Superposition (linearity): xðtÞ ¼

N P

N P

LT

ai xi ðtÞ !

i¼1

ai Xi ðsÞ:

ð5:6:9Þ

i¼1

Time delay: LT

xðt tÞuðt tÞ ! est XðsÞ; t > 0:

ð5:6:10Þ

Complex frequency shift (times-exponential): LT

eat xðtÞ ! Xðs þ aÞ: Time scaling: LT 1 xðatÞ ! Xðs=aÞ; a 6¼ 0; a is a constant: jaj LT 1 xðatÞes0 t ! Xðs s0 Þ: jaj

ð5:6:11Þ

ð5:6:12Þ ð5:6:13Þ

Convolution in time: LT

xi ðtÞ xj ðtÞ ! Xi ðsÞXj ðsÞ: Multiplication in time LT 1 ½X1 ðsÞ X2 ðsÞ: x1 ðtÞx2 ðtÞ ! 2pj

ð5:6:14Þ

ð5:6:15Þ

Times-t: LT

ðtÞn xðtÞ !ð1Þn

d n XðsÞ : dsn

Times-(1/t): Z 1 xðtÞ LT XðaÞda ! t s

ð5:6:16aÞ

ð5:6:16bÞ

Derivative: d n xðtÞ LT n ! s XðsÞ sn1 xð0 Þ sn2 xð1Þ ð0 Þ ::: xn1 ð0 Þ; dtn di xðtÞ xðiÞ ð0 Þ ¼ jt¼0 : dti Integration: Z 0 Z t XðsÞ LT 1 xðaÞda ! xðaÞda þ : s 1 s 1

ð5:6:17Þ

ð5:6:18Þ

Initial Value: xð0þ Þ ¼ lim ½sXðsÞ; ðXðsÞ is properÞ: s!1

ð5:6:19Þ

Final value: lim xðtÞ ¼ lim ½sXðsÞ; ðPoles of XðsÞ lie in left half s planeÞ

t!1

s!1

ð5:6:20Þ

Switched periodic (xT ðtÞis periodic with period T): LT

xðtÞ ¼ xT ðtÞuðtÞ !

XðsÞ : 1 esT

ð5:6:21Þ

Differentiation with respect to a second independent variable: @xðt; rÞ LT @Xðs; rÞ ! ; ðr is independent of s and tÞ: @r @r Integration with respect to a second independent variable: Z r Z r LT Xðs; rÞdr ! xðt; bÞdb; ðr is independent of s and tÞ: r0

r0

ð5:6:22Þ

ð5:6:23Þ

5.6 One-Sided Laplace Transform

169

2

sum is equal to the sum of the integrals. The time delay property can be shown by using a change of variable in the transform integral. The complex shift property can be shown by the following: Z1

s0 t

½e xðtÞe

s t

xðtÞe

0

ðsþaÞt

Convolution property: The convolution of two functions was defined in Chapter 2. Assuming the functions start at t ¼ 0, we can write

¼

0

Z1

3 x2 ðxÞesx dx5 ¼ X1 ðsÞX2 ðsÞ:

(5:6:25)

Complex frequency shift: This is also referred to as times-exponential property and Lfe

at

xðtÞg ¼

Z1

eat xðtÞest dt

0

¼

Z1

xðtÞeðsþaÞt dt ¼ Xðs þ aÞ: (6:5:26)

0

x1 ðaÞx2 ðt aÞda Times-t property: This follows from the following equation:

0

Z1

x1 ðaÞesa da5

This proves the convolution theorem.

The time-scaling and frequency-shifting properties follow by combining the time-scaling and the complex frequency-shifting properties.

x1 ðtÞ x2 ðtÞ ¼

3

0

dt ¼ Xðs þ aÞ:

0

Z1

2 4

Z1

dt ¼

L½x1 ðtÞ x2 ðtÞ ¼ 4

Z1

x2 ðbÞx1 ðt bÞdb:

(5:6:24) dXðsÞ ¼ ds

0

Z1

dest dt ¼ xðtÞ ds

0

Z1

½txðtÞest dt: (5:6:27)

0

The transform of this function is given by

L½x1 ðtÞ x2 ðtÞ ¼

Z1 0

¼

Z1

2 1 3 Z 4 x1 ðaÞx2 ðt aÞda5est dt 0

2 1 3 Z x1 ðaÞ4 x2 ðt aÞest dt5da:

0

L½x1 ðtÞ x2 ðtÞ ¼

0

2 x1 ðaÞ4

Z1

Example 5.6.3 In Example 5.6.2 we have derived the Laplace transform of the hyperbolic sine function. Use the times-t property to show that the following is true: 2as

LT

0

yðtÞ ¼ t sinhðatÞ !

Using the change of variable x ¼ t a and simplifying the integral, we have Z1

We can generalize this result by repeated derivatives of the transform.

3 x2 ðxÞesðxþaÞ dx5da:

a2 Þ 2

¼ YðsÞ:

(5:6:28)

Solution: Taking the derivative of the transform function in (5.6.8), we have Z1 dXðsÞ d 2as ½ðtÞ sinhðatÞest dt ¼ ¼ ds ds ðs2 a2 Þ 0

a

¼ We are considering only positive time functions, so x2 ðxÞ ¼ 0 for x50, which allows us to change the lower limit on the second integral in the above equation and

ðs2

2as ðs2

a2 Þ 2

:

The result in (5.6.28) follows by identifying the term in the integrand and its transform on the right in the & above equation.

170

5 Relatives of Fourier Transforms

Times-1/t property or complex integration property: This is Z1

xðtÞ provided XðbÞdb ¼ L t

(5:6:29)

s

parts. The convolution and the derivative properties are the most used properties in linear systems theory. Integration property: This property can be seen using the integration by parts.

lim½xðtÞ=t exists: t!0

L This can be shown by Za

2 3 Za Z1 XðbÞdb ¼ 4 ebt xðtÞdt5db

s

s

¼

¼

¼

0

Z1 0 Z1 0 Z1

2

xðtÞ4

8 t 0:

teat uðtÞ; a > 0: eat cosðbtÞuðtÞ; a > 0: eat sinðbtÞuðtÞ; a > 0: uðtÞ cosðbtÞuðtÞ sinðbtÞuðtÞ

Fourier Transform

1 1 ðs þ aÞ

1 1 ðjo þ aÞ

1

1

ðs þ aÞ2

ðjo þ aÞ2

sþa 2

jo þ a b2

ðjo þ aÞ2 þ b2

b

b

ðs þ aÞ2 þ b2

ðjo þ aÞ2 þ b2

ðs þ aÞ þ

1 s s s2 þ o 2 b s2 þ o 2

computation of transforms and their inverses. That is, one algorithm can be used to compute both the forward and inverse Fourier transforms with few modifications. There is no symmetry property of Laplace transforms. 5. Noting that XðsÞ is a function of s ¼ s þ jo, a complex quantity, it can only be plotted as a surface plot. On the other hand, the Fourier transform is a function of jo and therefore it is the cross section of the surface plot along the jo-axis. 6. Noting Item 5 above, the circuits and systems literature use the Laplace transform to compute the frequency characteristics of the function by simply substituting s ¼ jo in the Laplace transform. The complex function is generally written in terms of the magnitude and phase frequency characteristics of the signal.

1 þ pdðoÞ jo jo þ p½dðo þ bÞ þ dðo bÞ b2 o2 b þ jp½dðo þ bÞ dðo bÞ b2 o2

Example 5.9.1 Find the Hartley transform of xðtÞ using its Laplace transform. 1 1 ! ðs þ aÞ jo þ a 1 1 þ j Im : ¼ Re jo þ a jo þ a LT

xðtÞ ¼ eat uðtÞ !

Solution: The Hartley transform of xðtÞ can be obtained from 1 1 aþo Hart : Im ¼ xðtÞ !Re jo þ a jo þ a o2 þ a2 (5:9:7) & Example 5.9.2 Find the Hartley transform of the function uðtÞ from its Laplace transform. LT

xðtÞ ¼ uðtÞ ! XðsÞ ) Xð joÞ ¼ pdðoÞ ð j=oÞ:

5.9.2 Hartley Transforms and Laplace Transforms Hartley transform is the symmetrical form of the Fourier transform. It can be derived from the one-sided Laplace transforms. A few examples are given below for the Laplace transform functions XðsÞ with poles in the left half-plane only and later with poles in the left half-plane and with poles on the imaginary axis.

(5:9:8) Solution: The Hartley transform of the function is XH ðoÞ ¼ pd½o þ ð1=oÞ:

(5:9:9) &

See the chapter by Olejniczak in Poularikis, ed. (1996) for an extensive discussion on the relationship between the Laplace and Hartley transforms.

186

5 Relatives of Fourier Transforms

5.10 Hilbert Transform

yðtÞ ¼ hðtÞ xðtÞ; hðtÞ ¼ F 1 ½Hð joÞ:

Another transform that is closely related to the Fourier transform is the Hilbert transform. It is used in the theoretical descriptions and implementations of analog and digital Hilbert transformers. A device called the Hilbert transformer is basic and has important applications in single sideband modulation of signals and in digital signal processing. Hilbert transforms can be introduced with Euler’s formula e jot ¼ cosðotÞ þ j sinðotÞ. We will see shortly that the Hilbert transform of cosðotÞ is sinðotÞ. Hilbert transforms became an important area with analytic signals that are complex valued with one-sided spectrum. These have the form xa ðtÞ ¼ xðtÞ þ j^ xðtÞ, where x^ðtÞ is the Hilbert transform of xðtÞ. Analytic signals are considered in Section 5.10.3. Also, the real and imaginary parts of transfer functions of systems are tied together by Hilbert transforms.

(5:10:3)

From Chapter 4, F ½sgnðtÞ ¼ 2=ð joÞ. See (4.4.36). Using the symmetry or the duality property of the Fourier transforms, it follows that hðtÞ ¼

1 : pt

(5:10:4)

Using the convolution integral and (5.10.4) results in Z 1 D yðtÞ ¼ hðtÞ xðtÞ ¼ xðaÞhðt aÞda ¼ x^ðtÞ: 1

(5:10:5) x^ðtÞ is the Hilbert transform of the function xðtÞ. Note the hat in x^ðtÞ. Hilbert transform is a convolution operation and is a function of time. This is symbolically represented by HT

HT

xðtÞ ! x^ðtÞ ¼ H½xðtÞ ¼ yðtÞ ! YðjoÞ ¼ HðjoÞXðjoÞ; HðjoÞ ¼ j sgnðoÞ: (5:10:6)

5.10.1 Basic Definitions There are two ways of introducing the Hilbert transforms. One is by using an integral and the other by using the Fourier transform of the function. It is simpler to view it starting with the transform and derive the integral that defines the Hilbert transform. To start with assume xðtÞ is the input and yðtÞ is the output and F½xðtÞ ¼ Xð joÞ and F½yðtÞ ¼ Yð joÞ. The output transform Yð joÞ is assumed to be related to the input transform by

Hilbert transforms can be computed directly by the convolution in (5.10.5) or by using the transforms in (5.10.6). If x^ðtÞ is known, how do we compute xðtÞ from x^ðtÞ, if x^ðtÞ is not identically zero? It turns out that the Hilbert transform of xðtÞ is equal to ½xðtÞ. This can be seen from ^^ðtÞ ¼ ½xðtÞ hðtÞ hðtÞ FT x ! HðjoÞHðjoÞXðjoÞ ¼ ðj sgnðoÞÞ2 XðjoÞ ¼ XðjoÞ:

YðjoÞ ¼ HðjoÞXðjoÞ:

The function Hð joÞ is called the Hilbert transformer and is defined by ( Hð joÞ ¼ j sgnðoÞ ¼

It implies that if^ xðtÞ 6¼ 0 then we have the inversion formula ^^ðtÞ ¼ XðjoÞ ! x ^^ðtÞ F½x

e jp=2 ;

o 40

e jp=2 ;

o 50

(5:10:7)

(5:10:1)

¼ F1 ½XðjoÞ ¼ xðtÞ:

;

(5:10:8)

(5:10:2)

FT

hðtÞ ! Hð joÞ: Noting that multiplication in the frequency domain corresponds to the time-domain convolution, we have

Example 5.10.1 Find their Hilbert transforms of the following functions: a: x1 ðtÞ ¼ ejo0 t ; a0 40;

b: xðtÞ ¼ cosðo0 tÞ;

c: yðtÞ ¼ sinðo0 tÞ; d:x4 ðtÞ ¼ A; a constant

5.10 Hilbert Transform

187

Solution: a. The Fourier transforms of these functions are given by

This can be simplified and the corresponding Hilbert transform pair is

F½ejo0 t ¼ 2p dðo o0 Þ:

cosðo0 tÞ ! sinðo0 tÞ:

HT

(5:10:9)

Figure 5.10.1a,b gives the Fourier transforms of the functions in (5.10.9). Figure 5.10.1c gives HðjoÞ ¼ j sgnðoÞ. It follows that

Hilbert transform operation is an integral operation and the Hilbert transform of a sum is equal to the sum of the Hilbert transforms. c. We can repeat the above process and show that

½HðjoÞð2pdðo þ o0 ÞÞ

HT

¼ j 2psgnðoÞdðo þ o0 Þ ¼ j2pdðo þ o0 Þ:

sinðo0 tÞ ! sinðo0 t ðp=2ÞÞ ¼ cosðo0 tÞ:

(5:10:10)

¼ jejo0 t ¼ ejðo0 tðp=2ÞÞ :

HT

HT

xðtÞ ! x^ðtÞ; x^ðtÞ ! xðtÞ; x^ðtÞ 6¼ 0:

(5:10:12)

b. The Hilbert transform of the cosine function can be obtained by using Euler’s formula 1 jo0 t 1 jo0 t þ e cosðo0 tÞ ¼ e 2 2 1 HT 1 jðo0 tðp=2ÞÞ ! e þ ejðo0 tðp=2ÞÞ 2 2 ¼ cosðo0 t ðp=2ÞÞ:

From the above we note that the Hilbert transform of a sine or a cosine function can be obtained by adding a phase shift of ðp=2Þ. d. In the case of a constant, the transform is an impulse function. Note that sgnðoÞjo¼0 ¼ 0. That is, HðjoÞjo¼0 ¼ 0 and YðjoÞjo¼0 ¼ 0 in (5.10.1). Since the Fourier transform of a constant is an impulse function, it follows that the Hilbert trans& form of a constant is zero. We can generalize the above results and state that any periodic function with zero average value can be written in terms of Fourier cosine and sine series and therefore the Hilbert transform of such a periodic function xT ðtÞ is

p HT xT ðtÞ ! xT t : (5:10:16) 2

(a)

Fig. 5.10.1 (a)F ½e jo0 t , (b) F [e jo0 t ], (c) HðjoÞ ¼ jsgnðoÞ

(5:10:15)

(5:10:11)

Changing the sign in the exponent is a minor matter and we have h i HT ½ejoo t ! ejðo0 tðp=2ÞÞ

(5:10:14) Note also

From (5.10.10), x^ðtÞ ¼ F1 ½j2pdðo þ o0 Þ

(5:10:13)

(b)

(c)

188

5 Relatives of Fourier Transforms

In the next example we will make use of the integral to compute the Hilbert transform. Example 5.10.2 Find the Hilbert transform of the following functions: a: xðtÞ ¼ P½t=t; b: x1 ðtÞ ¼ xðt t=2Þ; c: yðtÞ ¼ A: (5:10:17) Solution: a. Using the integral expression, the Hilbert transform is given by Z

1

xðaÞ 1 da ¼ x^ðtÞ ¼ pðt aÞ p 1

Z

t=2

da ðt aÞ t=2

1 1 t þ t=2 ¼ ½lnðt t=2Þ lnðt þ t=2Þ ¼ ln p p t t=2

P

(5:10:19)

c. Hilbert transform of a constant is zero. This can be seen from (5.6.18) at the limit t ! 1, as & lnð1Þ ¼ 0, which verifies our earlier result. Example 5.10.3 Show that a. the energy (or the power) in an energy signal (or a power signal) xðtÞ and its Hilbert transform x^ðtÞ are equal and b. the signal and its Hilbert transform are orthogonal. Solution: The results are shown for energy signals and the results for power signals are left as exercises. a. The energies in the two functions are given by Ex^ ¼

(5:10:18) The pulse and its Hilbert transform are shown in Fig. 5.10.2. The pulse we started with is time limited, whereas the time-width of the Hilbert transform pulse is infinite. b. Hilbert transform of the delayed pulse can be obtained by changing the variable of integration in (5.10.18) by b ¼ t t=2 and following the above procedure results in

t ðt=2Þ HT 1 t ! ln : t p tt

¼

¼

1 2p 1 2p 1 2p

Z1 1 Z1 1 Z1

xðtÞj2 do jF½^

jjsgnðoÞj2 jXðjoÞj2 do

jXðjoÞj2 do ¼ Ex :

(5:10:20)

1

b. The two functions are orthogonal by the generalized Parseval’s theorem. Z1 Z1 1 xðtÞ^ xðtÞdt¼ XðjoÞ½F½^ xðtÞ dt 2p 1

¼

1 2p

1 Z1

jsgnðoÞjXðjoÞj2 do:

1

(5:10:21) &

(a)

5.10.2 Hilbert Transform of Signals with Non-overlapping Spectra In Chapter 10 single-sided modulation schemes will be studied, where Hilbert transforms play an important role. Consider the signals xðtÞ and gðtÞ with their spectra defined by

(b) Fig. 5.10.2 (a) Pulse function and (b) Hilbert transform of the pulse

FT

xðtÞ ! XðjoÞ; jXðjoÞj ¼ 0; joj4W FT

gðtÞ ! GðjoÞ;

GðjoÞ ¼ 0; joj5W:

(5:10:22)

5.10 Hilbert Transform

189

That is, xðtÞ is a low-pass signal and gðtÞ is a highpass signal. Such is the case in the single sideband modulation. We state that H½xðtÞgðtÞ ¼ xðtÞH½gðtÞ ¼ xðtÞ^ gðtÞ:

(5:10:23)

This can be proven by using the following steps: 1. Write the time function xðtÞgðtÞ using the transform convolution integral in the form and then using the Hilbert transform, we have Z 1 1 Xðo aÞGðaÞda; (5:10:24) xðtÞgðtÞ ¼ 2p 1 FT

H½xðtÞgðtÞ ! jsgnðoÞ Z 1 1 Xðo aÞGðaÞda : 2p 1

The spectrum of the analytic signal xa ðtÞ is the positive portion of the spectrum of the real signal xðtÞ. This property will be useful in the development of single sideband modulation scheme in Chapter 10. Some authors use the symbol xþ ðtÞð¼ xa ðtÞÞ for analytic signals. A real signal xðtÞ can be written in terms of analytic signals HT

xðtÞ ¼ ½xa ðtÞ þ x a ðtÞ=2 !½xa ðtÞ x a ðtÞ=2: (5:10:29)

Example 5.10.4 Let mðtÞ be a low-pass signal with MðoÞ ¼ 0; joj4W and oc 4W, then

H½mðtÞ sinðoc tÞ ¼ mðtÞ cosðoc tÞ:

1 1 xðtÞ ¼ ½1 þ sgnðoÞXðjoÞ F½xa ðtÞ ¼ F½xðtÞ þ j^ 2 2 ( (5:10:28) XðjoÞ; o40 ¼ : 0; o 50

(5:10:25)

2. Write the time-domain product xðtÞ^ gðtÞ in terms of the Fourier convolution integral. 3. Noting the non-overlapping spectra of the two functions, we can relate the two results and then show (5.10.23). The details are left as an exercise.

H½mðtÞ cosðoc tÞ ¼ mðtÞ sinðoc tÞ;

Some authors do not use the constant (1/2) in the definition of the analytic signal. The real significance of the analytic signal is its spectrum and is

(5:10:26)

Solution: These follow from (5.10.23) and the Hilbert transforms of the sine and cosine functions. We will use results in studying single sideband mod& ulations in Chapter 10.

Use of this gives cosðo0 tÞ ¼ ½ejo0 t þ ejo0 t =2; sinðo0 tÞ ¼ ½ejo0 t ejo0 t =2j:

(5:10:30)

Narrowband noise signals: Although statistical description of noise is beyond our scope here, we will study signals with their spectra centered at a frequency fc with a bandwidth B fc . For example, the output of an amplitude modulated signal is xc ðtÞ ¼ mðtÞ cosðoc tÞ with mðtÞ being a low-pass signal with its bandwidth much and much smaller than the carrier frequency fc . Such signals are narrowband (NB) signals. These are expressed in terms of the envelopeRðtÞ, a slowly varying function and the phase fðtÞ written in the form nðtÞ ¼ RðtÞ cosðoc t þ fðtÞÞ; RðtÞ 0:

(5:10:31)

5.10.3 Analytic Signals

In most cases, RðtÞ and fðtÞ are not transformable.

The Hilbert transform is used to define an analytic signal of the real signal xðtÞ by

Example 5.10.5 Find the envelope and the complex envelope of the NB signal in terms of two NB signals nc ðtÞ and ns ðtÞ given by

1 xðtÞ: xa ðtÞ ¼ ½xðtÞ þ j^ 2

(5:10:27)

nðtÞ ¼ nc ðtÞ cosðo0 tÞ ns ðtÞ sinðo0 tÞ:

(5:10:32)

190

5 Relatives of Fourier Transforms Table 5.10.1 Hilbert transform pairs Sinusoids: HT

HT

sinðo0 tÞ ! cosðo0 tÞ; cosðo0 tÞ ! sinðo0 tÞ Exponential: HT

ejot ! jsgnðoÞejot Rectangular pulse: P

hti t

1 t þ t=t22 ! ln p t t=t22

HT

Impulse: HT

dðtÞ ! 1=pt

Solution: First, qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ nðtÞ ¼ RðtÞ cosðo0 t þ fðtÞÞ; RðtÞ ¼ n2c ðtÞ þ n2s ðtÞ; fðtÞ ¼ tan1 ½ns ðtÞ=nc ðtÞ:

(5:10:33)

Hilbert transforms. Bulk of this chapter deals with the one-sided Laplace transforms. Specific topics are:

Fourier cosine and sine and Hartley transforms Laplace transforms and their inverses; regions of convergence

From this representation, the analytic signal can be obtained and is

Basic properties of Laplace transforms; initial and final value theorems

Partial fraction expansions na ðtÞ ¼ nðtÞ þ j^ nðtÞ ¼ nc ðtÞ cosðo0 tÞ ns ðtÞ sinðo0 tÞ Solutions of constant coefficient differential þ jnc ðtÞ sinðo0 tÞ þ jns ðtÞ cosðo0 tÞ ¼ nc ðtÞ½cosðo0 tÞ þ j sinðo0 tÞ þ jns ðtÞ½cosðo0 tÞ þ j sinðo0 tÞ ¼ ½nc ðtÞ þ jns ðtÞejo0 t : The envelope and the complex envelopes are, respectively, given by jna ðtÞj ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ n2c ðtÞ þ n2s ðtÞ;

(5:10:34) &

equations using Laplace transforms

Relationship between Laplace and Fourier transforms

Hilbert transforms and their inverses Various tables listing some simple time functions and their transforms

Problems

n~ðtÞ ¼ nc ðtÞ þ jns ðtÞ: Ziemer and Tranter (2002) give interesting applications of the Hilbert transforms. Refer table 5.10.1 for a short table of Hilbert transforms, see Hahn Poularikis, ed. (1996).

5.2.1 Derive the following properties using Xc ðoÞ ¼ FCT½xðtÞ: FCT

a: Xc ðtÞ !ðp=2ÞxðoÞ; oþb oþb FCT 1 Xc þ Xc ; b: xðatÞ cosðbtÞ ! 2a a a a40; b40; d2 Xc ðoÞ ; do2

5.11 Summary

c: t2 xðtÞ !

In this chapter Fourier transform integral is used to discuss and derive some of the related transforms. These include cosine, sine, Hartley, Laplace, and

d: x00 ðtÞ ! o2 Xc ðoÞ x0 ð0þ Þ:

FCT

FCT

(Assume xðtÞ and x0 ðtÞ vanish as t ! 1.).

Problems

191

5.2.2 Show

5.3.4 Show the following transforms are true.

1 FCT p ao ! e ; 2 þt 2a 1 FCT p ! sinðaoÞ; b: 2 a t2 2a b b FCT þ ! p sinðaoÞebo ; c: ðt aÞ2 þ b2 ðt þ aÞ2 þ b2 d 2 ½eat uðtÞ FCT a3 d: ! 2 : a þ o2 dt2 a: xðtÞ ¼

a2

5.2.3 Derive the following associated with sine transforms using Xs ðoÞ ¼ FST½xðtÞ: 1 oþb ob a: xðatÞ cosðbtÞ ! Xs þ Xs ; 2a a a

Hart

a: eat sinðo0 tÞuðtÞ !

o0 ða2 þ o20 o2 Þ þ 2ðaoÞ ða2 þ o20 o2 Þ þ 2ðaoÞ2

;

Hart

b: eat cosðo0 tÞuðtÞ ! ða oÞða2 þ o20 o2 Þ þ 2oða þ oÞ ða2 þ o20 o2 Þ þ 2ðaoÞ2 c:

1 X

Hart

dðt nTÞ !

n¼1

;

1 p X dðo ðk=TÞÞ; T k¼1

Hart

d: ejo0 t ! pdðo o0 Þ:

FST

FST 1 b: xðatÞ ! Xs ðo=aÞ; a40 a

5.2.4 Show the following are valid: rﬃﬃﬃﬃﬃﬃ 1 FST p ; a: pﬃﬃ ! 2o t b b FST b: 2 2 ! p cosðaoÞebo ; 2 2 b þ ðt aÞ b þ ðt aÞ b40: 5.3.1 The energy spectral density of a signal FT

xðtÞ ! XðjoÞ can be expressed by n ð1=2pÞjXðjoÞj2 ¼ ð1=2pÞ ½ReðXðjoÞÞ2 o þ½ImðXðjoÞÞ2 : Derive this in terms of the Hartley transform XH ðoÞ. Also derive the expression for the phase angle of the spectrum in terms of the Hartley transform.

5.4.1 Find the two-sided Laplace transforms and their ROCs of the following functions: a: x1 ðtÞ ¼ P½t; b: x2 ðtÞ ¼ teajtj ; a40: The following problems are concerned with onesided Laplace transforms. 5.4.2 Find the Laplace transforms of the following functions: pﬃﬃﬃﬃﬃﬃﬃ a: x1 ðtÞ ¼ 2 t=p: b: x2 ðtÞ ¼ L½t; c: x3 ðtÞ ¼ coshðbtÞ: 5.4.3 Find the Laplace transform of the function xðtÞ ¼ tuðtÞ by the following methods: a: Use the times-tproperty and the transform of uðtÞ to show that L½t uðtÞ ¼ 1=s2 . b. Use the result in part a to show by induction LT

that tn uðtÞ ! n!=ðsnþ1 Þ . The proof by induction uses the following procedure. The result is first shown to be true for the case of n ¼ 1. Then verify that if the result is true for n then it is also true for n þ 1. 5.6.1 Find the L-transform of xðtÞ ¼ jsinðtÞj; t 0.

5.3.2 Derive the Hartley transforms of the following functions using Fourier transforms:

c: x3 ðtÞ ¼ xðtÞ cosðo0 tÞ:

5.6.2 a. Find X1(s) ¼ L{x(t)} by using the transform of the unit step function. 1 t T=2 LT ! X1 ðsÞ: x1 ðtÞ ¼ P T T

5.3.3 Derive an expression for the Hartley transform of the convolution yðtÞ ¼ xðtÞ hðtÞ.

b. Take the limit of the transform as T ! 0 and identify the corresponding transform pair.

a: x1 ðtÞ ¼ xðatÞ; a 6¼ 0; b: x2 ðtÞ ¼ xðtÞ cosðo0 tÞ;

192

5 Relatives of Fourier Transforms

5.6.3 Show the following transform pair is true by a.using the integral of the transform and by b. using the second derivative of the Laplace transform to show the above result.

d 2 xðtÞ xðtÞ ¼ sinhðtÞ; xð0Þ ¼ 0 and x0 ð0Þ ¼ 0: dt2 5.9.1 Find

xðtÞ ¼ L½t 1 8 9 > > < t; 05t51 = LT 1 ¼ ð2 tÞ; 15t52 ! 2 ð1 2es þ e2s Þ: > > s : ; 0; otherwise 5.6.4 a. Use the differentiation with respect to the second variable property of the Laplace transforms to show that L½teat ¼ ½1=ðs þ aÞ2 .

xi ðtÞ ¼ L1 ½Xi ðsÞ with 1 es 1 ; b: X2 ðsÞ ¼ 2 2 ; sðs þ 1Þ s ðs þ 4Þ 1 es c: X3 ðsÞ ¼ : sð1 e2s Þ

a: X1 ðsÞ ¼

5.9.2 Show the residues A i can be determined by

5.6.5 Determine the transform of the Laplace transform of the following function xðtÞ: sinðo0 tÞ ; xðtÞ ¼ t

Zo 0

cosðotÞdo ¼

NðsÞ NðsÞ ¼ DðsÞ ðs þ s1 Þðs þ s2 Þ . . . ðs þ sN Þ N X Ai NðsÞ ¼ ; Ai ¼ js¼si : ðs þ si Þ dDðsÞ=ds i¼1

XðsÞ ¼

sinðo0 tÞ : t

0

5.7.1 Verify the following the transform pairs by a. evaluating the transform directly and by b. using the partial fraction expansion and then identify term by term from tables. 1 1 LT : ð1 cosðatÞÞ ! 2 2 a sðs þ a2 Þ 5.7.2 Verify the following by using the Laplace transforms properties: LT

a: x1 ðtÞ ¼ ð1=aÞebt sinhðatÞ ! 1=½ðs bÞ2 a2

a. Use this result to find L1 fXðsÞg ¼ L f1=½sðs þ 3Þg. b. Now consider YðsÞ ¼ ð1=sÞXðsÞ. Use Part a. to generalize this. 1

5.9.3 Assuming the regions of convergence are a: s41; b: s5 2, find 2s þ 1 1 1 : L fXII ðsÞg ¼ L ðs2 þ s 2Þ 5.9.4 Assuming the following Laplace transforms, find the corresponding Fourier transforms:

¼ X1 ðsÞ;

a: XII ðsÞ ¼

h i LT b: x2 ðtÞ ¼ ð1=2aÞt sinðatÞ ! s= ðs2 þ a2 Þ2 ¼ X2 ðsÞ: c: x3 ðtÞ ¼ t=T; 05t5T; x3 ðt þ nTÞ eas LT 1 : ¼ xðtÞ; n 0 ! 2 sð1 eas Þ as 5.8.1 Find the solution of dy þ 3y ¼ t cosðtÞ; yð0 Þ ¼ 1: dt 5.8.2 Using Laplace transforms to find the solution of the differential equation

¼

1 2

ðs þ aÞ

1 ðs aÞ2

; jsj5a; b: XðsÞ

s þ o0 : s2 þ o20

5.10.1 Show 1 HT t ; ! 2 1þt ð1 þ t2 Þ HT 1 b: dðtÞ ! ; pt K X c: Xs ½0 þ Xs ½k cosðko0 t a:

k¼1 HT

þ y½kÞ !

K X k¼1

Xs ½k sinðko0 t þ y½kÞ:

Chapter 6

Systems and Circuits

6.1 Introduction In this chapter we will consider systems in general, and in particular linear systems. Most systems are inherently nonlinear and time varying. A human being is a good example. He can run fast for a while and then speed comes down. If you plot speed versus time, the plot is not going to be a straight line, i.e., the function speed versus time is not linear. Humans are nonlinear and also time-varying systems. For example, if you want to ask your dad for a new car, you do not ask him when he is not happy. Moods change with time. These considerations are important in, for example, speaker identification. Human beings are not only nonlinear but also time-varying complicated systems. Nonlinear time-varying systems are very hard to deal with. Even though many of the systems may have nonlinear behavior characteristics, they can be approximated to be linear systems and they allow for transform analysis. In addition we are interested in systems that operate in the same manner every time we use them. That is, the systems must be independent of time. Linear time-invariant system analysis and design is the basis of present day system analysis and design. Transfer functions associated with these systems are discussed. In addition the frequency analysis makes it very attractive for the design of systems. Majority of the discussion in this chapter is on linear timeinvariant systems. These allow for transfer function analysis. The study of the amplitude and phase frequency responses of linear time-invariant systems is one of the important topics. When a signal through some media, it is modified by the media. In the frequency domain we can say that some

frequencies are amplified and some are attenuated. In addition different frequencies are delayed differently. Our goal is to filter frequencies with appropriate attenuations of the input frequencies with a constant delay at all frequencies in the frequency band of interest. The delay response is related to the phase response of a system. If the delay is not constant, then delay compensation may be required. One of the topics we will be interested is filter circuits. Toward this goal ideal low-pass, high-pass, band-pass, and band-elimination filter functions are introduced in this chapter. In addition to these simple examples of a differentiator, integrator, and a delay circuit are illustrated. A brief introduction to nonlinear systems is included later. The topic of linear systems is one of the topics every undergraduate student in electrical engineering program goes through. See the books Haykin and Van Veen (1999), Lathi (1998), Oppenheim et al. (1997), Nillsson and Riedel (1996), Poularikas and Seely (1991), Carlson (2000) and others.

6.2 Linear Systems, an Introduction Our study starts with a system that has an input and an output. It is symbolically represented by a block diagram shown in Fig. 6.2.1. The T inside the box is some transformation that converts the input signal xðtÞ into the output signal yðtÞ and yðtÞ ¼ T½xðtÞðT maps xðtÞ into yðtÞÞ:

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_6, Ó Springer ScienceþBusiness Media, LLC 2010

(6:2:1)

193

194

6 Systems and Circuits

below. What can you say about the linearity of theses schemes? Fig. 6.2.1 Block diagram of a system

a: y1 ðtÞ ¼ mðtÞ cosðoc tÞ;

Some use L½xðtÞ to represent a linear system. The system described by (6.2.1) is called a linear system if the transformation given by T½xðtÞ satisfies the following conditions: Principle of additivity: If T½x1 ðtÞ ¼ y1 ðtÞ and T½x2 ðtÞ ¼ y2 ðtÞ; then the superposition property states that

b: y2 ðtÞ ¼ ðA þ mðtÞÞ cosðoc tÞ; A 6¼ 0:

T½x1 ðtÞ þ x2 ðtÞ ¼ T½x1 ðtÞ þ T½x2 ðtÞ ¼ y1 ðtÞ þ y2 ðtÞ (6:2:2) Principle of proportionality: If T½xðtÞ ¼ yðtÞ; then for any constant a

(6:2:5)

Solution: a. For the inputs m1 ðtÞ and m2 ðtÞ, the outputs are, respectively, given by m1 ðtÞ cosðoc tÞ and m2 ðtÞ cosðoc tÞ. For the input a1 m1 ðtÞ þ a2 m2 ðtÞ, the output is the sum of the two individual outputs ½a1 m1 ðtÞ þ a2 m2 ðtÞ cosðoc tÞ. Therefore the system is linear. b. For the inputs x1 ðtÞ and x2 ðtÞ, the outputs are, respectively, given by ðA þ m1 ðtÞÞ cosðoc tÞ and ðA þ m2 ðtÞÞ cosðoc tÞ. For the input ða1 m1 ðtÞþ a2 m2 ðtÞÞ, the output is given by ½A þ a1 m1 ðtÞþ a2 m2 ðtÞ cosðoc tÞ which is not equal to ðA þ m1 ðtÞÞ cosðoc tÞ þ ðA þ m2 ðtÞÞ cosðoc tÞ:

T½axðtÞ ¼ aT½xðtÞ ¼ ayðtÞ:

(6:2:3)

This property is also referred to as the homogeneity property. We can combine the two properties into one and state that a linear system satisfies the following property T½a1 xðtÞ þ a2 x2 ðtÞ ¼ a1 y1 ðtÞ þ a2 y2 ðtÞ

(6:2:4)

Therefore the system is nonlinear. Note that the Fourier transform is a linear operation as F½a1 x1 ðtÞ þ a2 x2 ðtÞ ¼ a1 F½x1 ðtÞ þ a2 F½x2 ðtÞ: (6:2:6) & The systems described by the following equations are linear systems and the reader is encouraged to go through the proofs:

for any pair of constants a1 and a2 . Otherwise, the system is called a nonlinear system.

a: yðtÞ ¼ kxðtÞ

Example 6.2.1 Show that a system described by yðtÞ ¼ x2 ðtÞ is a nonlinear system.

b: yðtÞ ¼ k

Solution: For two inputs x1 ðtÞ and x2 ðtÞ, the corresponding outputs are, respectively, given by y1 ðtÞ ¼ x21 ðtÞ and y2 ðtÞ ¼ x22 ðtÞ. If the input is x1 ðtÞ þ x2 ðtÞ, then the output is ðx1 ðtÞ þ x2 ðtÞÞ2 6¼ x21 ðtÞ þ x22 ðtÞ and therefore the system is nonlinear. We can make a general statement that if the output of a system is a power of the input, and the power is not equal to one, then the system is nonlinear. Other examples of nonlinear systems include y1 ðtÞ ¼ logðxðtÞÞ; y2 ðtÞ ¼ jxðtÞj:

c: yðtÞ ¼

dxðtÞ dt

Z xðbÞdb

d: yðtÞ ¼ xðt tÞ; t 0

ðAmplifierÞ

(6:2:7a)

ðDifferentiatorÞ (6:2:7b)

ðIntegratorÞ

(6:2:7c)

ðDelay deviceÞ: (6:2:7d)

6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws

&

Example 6.2.2 Two of the many modulation schemes that we will be interested in are given

In this section we will consider two-terminal passive and active components and laws that pertain to the interconnection of elements. These are two powerful

6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws

laws, referred to as Kirchhoff’s voltage and current laws first formulated by Kirchhoff (pronounced as kear-koff) in 1847. One gives equations in terms of voltages across components and the other gives equations in terms of currents flowing through the components. Component equations and the Kirchhoff’s laws provide us with circuit analysis tools.

6.3.1 Two-Terminal Component Equations Simple circuits include ideal sources, voltage, and current sources and three types of components resistors, inductors, and capacitors. The symbols for the sources are shown in Fig. 6.3.1. An ideal voltage source is a two-terminal component whose voltage across the two terminals is a constant or a function of time regardless of what the current through the component is. Examples of voltage sources are wall outlets, where we assume that the voltage is vs ðtÞ ¼ Vm cosðom tÞ, and batteries, where the voltage across is a constant. The first source we refer to as an alternating current (AC) source and the second one is a constant voltage source (DC). The positive sign on top of the ideal voltage source indicates the higher potential whenever the source voltage is positive. Most generators are voltage sources. The ideal current source is a two-terminal component whose current is a constant or a function of time, regardless of what the voltage across it is. Transistors and many other electronic devices act more like a current source rather than a voltage source. It is important to notice the voltage signs and the direction of the currents through the sources. This convention shows that the sources provide power. The three basic passive components are the resistor, inductor, and the capacitor. The Lumped parameter models are shown in Fig. 6.3.2. These are passive elements, i.e.,

Fig. 6.3.1 Voltage, current, and a constant voltage source

195

The three basic passive components are the resistor, inductor, and the capacitor. The lumped parameter models are shown in Fig. 6.3.2. These are passive elements, i.e., they do not produce any power. Therefore the notation for the three components is that the current flows from the positive terminal to the negative terminal. The resistance is measured in Ohms, the inductor in Henries, and the capacitor in Farads. The voltage across a resistor is related to the current by the Ohm’s law and is given by vR ðtÞ ¼ RiR ðtÞ:

(6:3:1)

The voltage across an inductor is given by vL ð t Þ ¼ L

diL ðtÞ : dt

(6:3:2)

Since the voltage across an inductor is the derivative of current, it is zero for a constant current. The inductor stores energy in a magnetic field produced by current through a coil of wire. Inductor is an energy storage device and the instantaneous stored energy, measured in Joules, is 1 wL ¼ Li2L ðtÞ: 2

(6:3:3)

The current in the inductor can be computed from (6.3.2) and is 1 iL ðtÞ ¼ iL ðt0 Þ þ L

Zt

vL ðbÞdb:

(6:3:4)

t0

The term iL ðt0 Þ corresponds to the initial conditions on the current through the inductor at time t ¼ t0 . The current through a capacitor is given by

i C ðt Þ ¼ C

dvC ðtÞ dt

(6:3:5)

196

6 Systems and Circuits

Fig. 6.3.2 Resistor, inductor, and the capacitor

iR (t)

iL (t) R

L

C

vR (t)

vL (t)

vC (t)

The gap in the capacitor symbol reflects that when the voltage across the capacitor is constant, then the capacitor acts like an open circuit. Capacitor is an energy storage device and the instantaneous energy, measured in Joules, is

iC ðtÞdt:

(6:3:7)

The voltage vc ðt0 Þ corresponds to the initial voltage across the capacitor at time t ¼ t0 . Initial conditions are necessary when we consider transient analysis. In the design of systems we generally do not consider the initial conditions, as the systems are supposed to work for any initial conditions. We assume that the initial conditions on the capacitors and the inductors are assumed to be zero with t0 is equal to 1. These allow us to use both the Laplace and the Fourier transforms. This is especially true in network synthesis, as the design specifications are given in terms the sinusoidal steady state. Once we have the designs, we can always test the systems with initial conditions and any possible changes in the responses of systems associated with the initial conditions. The component equations for the three two-terminal components discussed above can be expressed in terms of Laplace transforms. To avoid any confusion from the inductor values we will make use of the symbol Lf:g for the Laplace transform. These are VR ðsÞ ¼ LfvR ðtÞg ¼ RLfiR ðtÞg ¼ RIR ðsÞ diL ðtÞ VL ðsÞ ¼ LfvL ðtÞg ¼ L L dt

(6:3:8)

¼ sLIL ðsÞ LiL ð0 Þ or IL ðsÞ

(6:3:9)

iL ð0 Þ 1 þ VL ðsÞ s sL

¼

v C ð 0 Þ 1 þ IC ðsÞ or IC ðsÞ s sC

(6:3:10)

Assuming the initial conditions are zero, we have the component equations in terms of the Laplace transformed variable s for the three components

t0

¼

dvc ðtÞ Vc ðsÞ ¼ LfvC ðtÞg ¼ L dt

(6:3:6)

The current in the capacitor is Zt

¼ sCVC ðsÞ CvC ð0 Þ:

1 wC ¼ Cv2C : 2

1 v C ð t Þ ¼ vC ð t 0 Þ þ C

iC (t)

LfvR ðtÞg ¼ VR ðSÞ; LfiR ðtÞg ¼ IR ðsÞ; LfvL ðtÞg ¼ VL ðsÞ; FfiL ðtÞg ¼ IL ðsÞ Lfvc ðtÞg ¼ Vc ðsÞ; Lfic ðtÞg ¼ Ic ðsÞ VR ðsÞ ¼ RIR ðsÞ; VL ðsÞ ¼ LsiL ðsÞ; Vc ðsÞ ¼ ð1=CsÞIC ðsÞ

(6:3:11)

The component equations in terms of the Fourier transform variable jo are VR ðjoÞ¼F½vR ðtÞ;VL ðjoÞ¼F½vL ðtÞ;Vc ðjoÞ¼F½vc ðtÞ (6:3:12a)

IR ðjoÞ ¼ F½iR ðtÞ; iL ðjoÞ ¼ F½iL ðtÞ; ic ðjoÞs ¼ F½ic ðtÞ (6:3:12b) VR ðjoÞ ¼ RIR ð joÞ; VL ð joÞ ¼ joLIL ð joÞ Vc ð joÞ ¼ ð1=joCÞIc ðjoÞ:

(6:3:12c)

In the case of zero initial conditions either Laplace or Fourier transform variables can be used. One can be obtained from the other. Also, the voltage to the current transform ratio is called as an impedance of the component under consideration and its inverse as the admittance of that component. The impedances of the three components are, respectively, given by R; Ls ðor joLÞ and 1=Cs ðor 1=joCÞ.

6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws

6.3.2 Kirchhoff’s Laws Circuit analysis is based on the Kirchhoff’s current and voltage laws and the component equations. The Kirchhoff’s current law (KCL) states that the sum of the currents going into a junction, a node, is equal to the sum of the currents going out of that junction. In other words, the algebraic sum of the currents going into a node is equal to zero. The dual to the current law is the voltage law and is stated for a loop in a circuit. A loop is any path that goes from one node to another node and returns to the starting node. The Kirchhoff’s voltage law (KVL) states that the sum of the voltage drops around any loop is equal to the sum of the voltage rises. Or, the algebraic sum of voltages around a loop is equal to zero. Examples are given in Fig. 6.3.3.

Using the component equation for the capacitor and the relation in (6.3.14) and relating input and output results in iC ¼ C

LT

LT

(6:3:15b)

FT

FT

yðtÞ ! YðjoÞ

(6:3:16)

RCðsÞYðsÞ þ YðsÞ ¼ XðsÞ or

(6:3:17)

ðRCjo þ 1ÞYðjoÞ ¼ XðjoÞ: ð1=CsÞ R þ ð1=CsÞ ð1=RCÞ XðsÞ; XðsÞ ¼ s þ ð1=RCÞ 1 YðjoÞ ¼ XðjoÞ: 1 þ joRC YðsÞ ¼

(6:3:13)

Assuming that the current flowing through the output node is zero, i.e., the circuit is not loaded, we have

Fig. 6.3.3 Illustration of Kirchhoff’s current and voltage laws (a) i1 ðtÞ þ i2 ðtÞ i3 ðtÞ ¼ 0, (b) v1 ðtÞ v2 ðtÞ v3 ðtÞ ¼ 0

(6:3:15a)

xðtÞ ! XðsÞ; yðtÞ ! YðsÞ; xðtÞ ! XðjoÞ and

Solution: Using the KVL, we have the input voltage xðtÞ is equal to the sum of the voltages across the resistor and the capacitor.

xðtÞ vC ðtÞ xðtÞ yðtÞ ¼ : iR ¼ iC ; and iR ¼ R R (6:3:14)

dvC dy dy xðtÞ yðtÞ ¼C ; C ¼ : dt dt dt R dy RC þ yðtÞ ¼ xðtÞ: dt

It is a linear combination of the output and the derivative of the output related to the input. The system described by this differential equation is a linear system. In terms of the Laplace and the Fourier transforms at each step and simplifying, the output transforms can be expressed as follows:

Example 6.3.1 Consider the simple RC circuit shown in Fig. 6.3.4a. Derive the differential equation relating the input and the out put of the circuit.

xðtÞ ¼ vR ðtÞ þ vC ðtÞ:

197

(6:3:18)

&

A simple integrator and a differentiator: The circuit in Fig. 6.3.4a can be used as an integrator in the lowfrequency range. The integral form of the equation in (6.3.15b) is

(a)

(b)

198

6 Systems and Circuits

YðsÞ ¼ s=½ðs þ ð1=RCÞÞXðsÞ ¼ HðsÞXðsÞ; HðsÞ ¼ s=½ðs þ ð1=RCÞÞ: (6:3:23a)

(a)

(b)

Fig. 6.3.4 RC circuits

RCyðtÞ þ

Zt

Zt

yðaÞda ¼

1

xðaÞda:

(6:3:19)

1

If the time constant RC is large enough that the integral on the left side of the equation in (6.3.19) is dominated by RCyðtÞ, then (6.3.19) can be approximated by an integral, a smoothing operation, and is

RCyðtÞ

Zt

The above two examples provide simple circuits for low-pass and high-pass filters. The amplitude response jHðjoÞj can be approximated for small frequencies and rﬃﬃﬃﬃﬃﬃﬃ 2 1 joj2 2 RCo ; joj jHðjoÞj ¼ ð1=RCÞþo2 RC !HðjoÞjoðRCÞ: (6:3:23b) &

1 xðaÞda or yðtÞ ¼ RC

1

Zt xðaÞda: 1

(6:3:20) Example 6.3.2 Relate the input and the output transforms of the circuit in Fig. 6.3.4b. Solution: Using the Kirchhoff’s voltage law, we have dxðtÞ dt dvc ðtÞ dyðtÞ ic ðtÞ dyðtÞ þ ¼ þ ¼ dt dt c dt dxðtÞ dyðtÞ 1 ¼ þ yðtÞ: dt dt RC

xðtÞ ¼ vc ðtÞ þ yðtÞ !

(6:3:21)

If the time constant ðRCÞ is small enough that the second term dominates the first term on the right in the last equation, we can approximate (6.3.21) by yðtÞ RC

dx : dt

The networks containing ideal resistors (R’s), inductors (L’s), and capacitors (C’s) result in constant coefficient differential equations.

6.4 Time-Invariant and Time-Varying Systems For any system of use, we like the system to respond every time the same way when we switch the system on. That is, if we switch today or tomorrow, the system should respond exactly the same. Such a system is called a time-invariant system. Time-invariant system: If the response of a system is yðtÞ to the input xðtÞ; i:e:; yðtÞ ¼ T½xðtÞ, then the system is called a time invariant or a fixed system if T½xðt t0 Þ ¼ yðt t0 Þ. Otherwise, it is a time-varying system. Linear time-invariant system: A system is linear time invariant(LTI)if it is linear and time invariant. Example 6.4.1 The systems described by constant coefficient differential equations are linear time-invariant systems. RLC networks are linear time-invariant systems. Circuits containing diodes, transistors, and other electronic components are & nonlinear. Example 6.4.2 Consider the model of a carbon microphone shown in Fig. 6.4.1. The resistance R

(6:3:22)

The circuit approximates the derivative operation or it acts like a differentiator. Taking the transform of the equation in (6.3.21) and solving for YðsÞ; we have

Fig. 6.4.1 A time-varying system

6.5 Impulse Response

199

is a function of the pressure generated by sound waves on the carbon granules of the microphone, which is a function of time. The circuit has only one loop and using the Kirchhoff’s voltage law (KVL) and the component equations, we can write xðtÞ ¼ vR þ vL ¼ RðtÞiðtÞ þ L

diðtÞ : dt

(6:4:1)

The resistance is a function of time and the resulting differential equation has coefficients that vary with time and the system is a time-varying & system. Earlier we have indicated that a human being is a nonlinear time-varying system. The speech signal is a time-varying signal, albeit, a slowly time-varying signal. To analyze a slowly timevarying signal, we segment the speech signal by using windows and find the needed information for each segment. Obviously the spectral characteristics of different phonemes are different. In the earlier chapters we defined causal signals that are zero for t50. We can similarly define causal systems. Causal systems: Causal systems do not respond until the input is applied. That is, they do not anticipate the input. For a causal system, if the input

Invertibility: A system is said to be invertible if the input of the system can be recovered from the output of the system. Consider that we have the response of the system given by yðtÞ ¼ T½xðtÞ:

(6:4:4)

The system is invertible if there is a transformation T 1 such that T 1 ½yðtÞ ¼ T 1 ½T½xðtÞ ¼ xðtÞ:

(6:4:5)

That is ðT 1 TÞ ¼ I, the identity operator. Simple examples of non-invertible systems include yi ðtÞ ¼ x2 ðtÞ; y2 ðtÞ ¼ uðxðtÞÞ and many others. In each of these cases we cannot determine the function xðtÞ from yi ðtÞ. Example 6.4.3 Give an expression for the derivative of the current in an inductor. Solution: The current in an inductor is Zt 1 diL : iL ðtÞ ¼ vL ðtÞdt ) vL ðtÞ ¼ L dt L

(6:4:6)

1

Derivative operations.

and

the

integral

are

inverse &

Now consider one of the important concepts in system analysis, i.e., the signal and the system interaction.

xðtÞ ¼ 0 for all t T; then the output yðtÞ ¼ 0 for all t T

(6:4:2)

Memory and memoryless systems: A system is called memoryless if the output of the system at a particular time depends only on the input at that time. The resistor is memoryless since vðtÞ ¼ RiðtÞ and the voltage and the current pertaining to this component are related at each value of t. The capacitor voltage is related to the current by an integral and the inductor voltage is related to the voltage by an integral. Capacitors and inductors have memory and initial conditions can be assigned on these. The relations are

vC ðtÞ ¼

1 C

Zt 1

iC ðaÞda; iL ðtÞ ¼

1 L

Zt

6.5 Impulse Response Consider the block diagram of a LTI system in Fig. 6.5.1 with input xðtÞ and the corresponding output is yðtÞ. We like to find a relationship between the input and the output and the system characteristics. If xðtÞ ¼ dðtÞ; an impulse, then the output, the response of the impulse, is the impulse response of the LTI system identified by

vL ðaÞda:

1

(6:4:3)

Fig. 6.5.1 A linear time-invariant system

200

6 Systems and Circuits

yðtÞ ¼ hðtÞ ¼ T½dðtÞ:

(6:5:1)

Therefore the response for the input ai dðt ti Þ is ai hðt ti Þ, i.e., T½ai dðt ti Þ ¼ ai hðt ti Þ where a i ’s are some constants. If the input is a linear combination of impulses, then the response will be a linear combination of the corresponding impulse responses. That is, T½

N2 X

ai dðt ti Þ ¼

i¼N1

N2 X

ai hðt ti Þ:

(6:5:5) The time instant tn ¼ nDt is at some point on the time axis. As Dt ! 0; nDt approaches a continuous variable b, the sum becomes an integral and Dt becomes a differential and

(6:5:2)

i¼N1

xðtÞ ¼

We can tie this relationship to an arbitrary input using the approximation of an impulse and relate the output to the input in terms of a sum of delayed impulse responses. Impulse functions were represented in the limit by (see Section 1.4.) 1 h t ti i P : D t!0 D t Dt

dðt ti Þ ¼ lim

N2 X n¼N1

1 t nDt ½xðnDtÞDt P : Dt Dt

(6:5:4)

Note the multiplication and division by Dt in (6.5.4). The term xðnDtÞDt approximates the area of the pulse centered at t ¼ nDt. Now

Fig. 6.5.2 (a) xðtÞ(b) Pulse centered at t ¼ nDt

(a)

Z1

xðbÞdðt bÞdb:

(6:5:6)

1

This is valid provided xðtÞ is continuous for all t. It is the convolution of the two functions dðtÞ and xðtÞ. See the equation in (2.2.2a) in Chapter 2. That is, xðtÞ ¼ xðtÞ dðtÞ:

(6:5:3)

Consider an arbitrary signal xðtÞ shown in Fig. 6.5.2a. There is no specific significance for the shape of this function. Now divide the time into intervals of Dt seconds apart as shown in Fig. 6.5.2b. The strip centered at t ¼ nDt with a width of Dt can be approximated by xðnDtÞP½ðt nDtÞ=Dt. If Dt is negligibly small, the pulse can be assumed to be a rectangular pulse and the above approximation is good. The function xðtÞ can now be approximated by

xðtÞ ﬃ

t nDt lim xðnDtÞ ð1=DtÞP ¼ xðtn Þdðt tn Þ: Dt!0 Dt

(6:5:7)

In a similar manner, the output expression can be derived. Since the system is a time-invariant system, the input xðnDtÞdðt nDtÞ produces an output xðnDtÞhðt nDtÞ. Combining all the responses, we can pictorially identify N2 X

Produces

xðnDtÞ½dðt nDtÞDt

n¼N1

ﬃ

N2 X

! yðtÞ

the output

xðnDtÞhðt nDtÞDt:

(6:5:8)

n¼N1

In the limit, i.e., when Dt ! 0, nDt becomes a continuous variable b. The time interval Dt becomes a differential db and the summation becomes an integral. Noting the limits on the sum are arbitrary, the sum can be taken as over all positive and negative integers and the integral correspondingly goes from 1 to þ 1.

(b)

6.5 Impulse Response

yðtÞ ¼

201

Z1

xðbÞhðt bÞdb:

(6:5:9)

1

This integral is a superposition or a convolution integral of two functions, input and the impulse response of the linear time-invariant (LTI) system. The response of the LTI system to any input xðtÞ is yðtÞ. Symbolically, it can be written in the form

yðtÞ ¼ xðtÞ hðtÞ ¼

¼

Z1 1 Z1

xðbÞhðt bÞdb

FT

!jXðjoÞj2 ; fh ðtÞ

fx ðtÞ ¼ xðtÞ xðtÞ ¼ hðtÞ hðtÞ fy ðtÞ ¼ yðtÞ yðtÞ

(6:5:14a)

FT

!jHðjoÞj ! ¼ jYðjoÞj2 ; fy ðtÞ 2

FT

¼ hðtÞ hðtÞ fx ðtÞ:

(6:5:14b)

In a similar manner, the output power spectral density of a periodic (or a random signal) can be expressed in terms of the input power spectral density by FT

FT

Rx ðtÞ ! Sx ðoÞ; Ry ðtÞ ! Sy ðoÞ; Sy ðoÞ ¼ jHðjoÞj2 Sx ðoÞ:

hðaÞxðt aÞda ¼ hðtÞ xðtÞ:

(6:5:15)

1

(6:5:10) ) YðsÞ ¼ HðsÞXðsÞ:

(6:5:11)

The function HðsÞ is the transform of the impulse response, HðsÞ ¼ L½hðtÞ and is called the transfer function of the LTI system in Laplace transform domain or s-domain. It is symbolically represented by the block diagram in Fig. 6.5.3. In the Fourier domain, (6.5.11) is expressed by YðjoÞ ¼ HðjoÞXðjoÞ; HðjoÞ ¼ F½hðtÞ:

(6:5:12)

The input, the output, and their Fourier (and Laplace) transforms are identified on the block diagram along with impulse response hðtÞ and its transform HðjoÞ or HðsÞ. Notes: Input–output energy spectral density relations of a linear system: The output energy spectral density is

These operations are basic to the study of linear systems, as they provide a simple way of expressing how the energy (or power) of an input signal is distributed at the output. From Examples (6.3.1) and (6.3.2), the respective transfer functions and the corresponding impulse responses are given by H ðsÞ ¼

ð1=RCÞ s þ ð1=RCÞ

1 t=RC e ! RC uðtÞ ¼ hðtÞ

LT

(6:5:16a) HðjoÞ ¼

1=RC ðjo þ ð1=RCÞÞ

1 ð1=RCÞt ! RC uðtÞ e

FT

(6:5:16b) H ðsÞ ¼

s ðs þ 1=RCÞ

LT

! dðtÞ et=RC uðtÞ ¼ hðtÞ (6:5:17a)

HðjoÞ ¼

jo ðjo þ 1=RCÞ

FT

! dðtÞ et=RC uðtÞ ¼ hðtÞ:

(6:5:13)

(6:5:17b) &

Correspondingly, the output autocorrelation (AC) can be expressed in terms of the input AC as shown below.

Notes: A single input–single output of a linear time-invariant system is described by its transfer function HðsÞ ¼ YðsÞ=XðsÞ in the Laplace domain

jYðjoÞj2 ¼ jHðjoÞj2 jXðjoÞj2 :

Fig. 6.5.3 Inputs, outputs and transfer functions

202

6 Systems and Circuits

or in the Fourier domain by HðjoÞ. Its impulse response is hðtÞ ¼ L1 ½HðsÞ or hðtÞ ¼ F1 ½HðjoÞ: jHðjoÞj and ﬀHðjoÞ are the amplitude and the phase & responses of the linear system. A causal continuous-time LTI system is memoryless if and only if hðtÞ ¼ cdðtÞ, where c is a constant. Noting that the output of a causal continuous-time LTI system is described in terms of the input xðtÞ and its impulse response hðtÞ, it is expressed by Z1 hðaÞxðt aÞda yðtÞ ¼

¼

1 Z1

cdðaÞxðt aÞda ¼ cxðtÞ:

each frequency, corresponding to a linear timeinvariant (LTI) system, are tied by (6.5.20). The response of a LTI system with a transfer function HðjoÞ to a unit input Hð0Þ:

6.5.2 Bounded-Input/Bounded-Output (BIBO) Stability BIBO stability of a LTI system is tied to its impulse response hðtÞ. Consider the output of the LTI system described by the convolution integral to a bounded input xðtÞ with jxðtÞj M. It can be shown that yðtÞ is bounded provided the impulse response hðtÞ is absolutely integrable. That is,

1

yðtÞ ¼

Z1

xðt bÞhðbÞdb ) jyðtÞj

1

6.5.1 Eigenfunctions

The transfer function of a linear time-invariant system can be expressed in terms of its impulse response hðtÞ with the input of the form xðtÞ ¼ est . Then the system output is Z 1 yðtÞ ¼ T½xðtÞ ¼ hðtÞ xðtÞ ¼ hðaÞxðt aÞda 1

¼

Z1

hðaÞesðtaÞ da ¼ est HðsÞ:

1

(6:5:18)

(6:5:19)

is called an eigenfunction (or a characteristic function) and HðsÞ is the eigenvalue (or the characteristic value). That is, est is the eigenfunction and the eigenvalue is defined as the system function. In terms of Fourier transforms, yðtÞ ¼ HðjoÞejot :

jxðt bÞjjhðbÞjdb M

1

Z1

Z1

jhðbÞjdb:

1

(6:5:21) 5 jhðbÞjdb ¼ K 1 ) jyðtÞj MK ¼ N; (6:5:22:)

1

If (6.5.22) is satisfied, then the output is bounded and the system is BIBO stable. Example 6.5.1 Determine if the system described by its impulse response hðtÞ ¼ eat uðtÞ; a40 is BIBO stable. Solution: The system is BIBO stable since

An equation satisfying Tfest g ¼ HðsÞest :

Z1

(6:5:20)

It is the response of a linear time-invariant system with a transfer function HðjoÞ to an input ejot and the relationship holds for each o. The responses for

Z1 1

jhðbÞjdb ¼

Z1 e

2ab

1 2ab 1 1 db ¼ e 0 ¼ 2a 51: 2a

0

&

BIBO stability requires that the transfer function is strictly proper. That is, the degree of the numerator polynomial of the transfer function is less than the degree of the denominator polynomial. Otherwise, the impulse response will contain the derivatives of the impulse function, which are not absolutely integrable. The ideal differentiator is

6.5 Impulse Response

yðtÞ ¼ dxðtÞdt

LT

! sXðsÞ; HðsÞ ¼ s; hðtÞ ¼ d0 ðtÞ:

The derivative of an impulse function is not absolutely integrable. The transfer function is not strictly proper. Similarly the ideal integrator has a transfer function HðsÞ ¼ 1=s has a simple pole on the jo axis at the origin and is marginally stable. Stability analysis is an important topic in all areas of systems engineering, especially in control systems. The literature is extensive in this area and the discussion here is limited to simple ideas. A linear timeinvariant system is stable if every root of its characteristic equation, i.e., the poles of the transfer function HðsÞ have negative real parts. The natural and forced responses of these systems can be described by seeing the properties of the inverse Laplace transforms of response functions with poles at various locations on the s plane. As mentioned earlier the responses of systems have two parts, one is a natural response that is due to the system itself and the other one is the response of the system when there is input. If a characteristic root is a simple real and negative, then the response corresponding to this pole is exponentially decaying. If a root is multiple and real, then the response is a polynomial intmultiplied by an exponentially decaying response. If we have a simple complex conjugate poles on the imaginary axis, then the corresponding response is oscillatory. A system with simple finite poles on the imaginary axis is called wide sense stable or marginally stable. The ideal integrator has a transfer function HðsÞ ¼ 1=s has a simple pole on the jo axis at the origin. It is marginally stable. If we have a pair of complex conjugate poles on the left half plane, the corresponding response is exponentially decaying oscillatory response. If the poles are multiple complex conjugate, then the time response has the form of a polynomial multiplied by an oscillatory decaying response. The systems are stable if its transfer function has all poles on the left half of the splane. The behavior of the impulse response depends on the poles closest to the imaginary axis. For most systems these are simple complex poles, referred to as dominant poles. Complicated systems with transfer function given by HðsÞ generally have many poles and are approximated by a reduced-order system HR ðsÞ by keeping only the poles near the imaginary axis, referred to as a model reduction.

203

Example 6.5.2 Illustrate the model reduction of the system with the transfer function

1 1 HðsÞ ¼100 ðs þ 1Þ ðs þ 5Þ

LT

! 100et uðtÞ

100e5t uðtÞ ¼ hðtÞ Solution: Noting that the pole at s ¼ 5 is farther away than the pole at s ¼ 1, the transfer function HðsÞ can be approximated by a reduced-order function LT 100 ! 100et uðtÞ ¼ hR;1 ðtÞ: HR;1 ðsÞ ¼ ðs þ 1Þ Another way is ignore the poles away from the imaginary axis. Then 400 ) HR;2 ðsÞ 5ðs þ 1Þð:2s þ 1Þ LT 80 ¼ ! 80et uðtÞ ¼ hR;2 ðtÞ: ðs þ 1 Þ

H ðsÞ ¼

&

6.5.3 Routh–Hurwitz Criterion (R–H criterion) The R–H criterion Kuo (1987) provides a test if the roots of a polynomial given below are on the left half plane without actually factoring the polynomial.

DðsÞ ¼ dn sn þ dn1 sn1 þ þ d1 s þ d0

(6:5:23)

Without loosing any generality the coefficient of sn is assumed to be 1. Furthermore, if d0 ¼ 0, i.e., the polynomial has a root at s ¼ 0, the polynomial can be divided by s and test the resulting polynomial for the stability. The Routh array starts by arranging two rows consisting of the coefficients of the polynomial in the following form: Note that ðn kÞth row starts with the coefficient nk ðs Þ. If n is even (odd), then d0 is the last entry in row n (n1). The next step is construct row (n2) by using rows n and n1.

204

6 Systems and Circuits

Row n : ðsn Þ : Row n 1 : ðsn1 Þ :

dn dn1

dn2 dn3

dn4 dn5

::: ::: (6:5:24)

dn1 dn2 dn dn3 dn1 dn1 dn4 dn dn5 ::: dn1

Row n 2 : ðs n2 Þ :

(6:5:25)

The entries in row n2 can be written in terms of determinants. 1 dn dn2 Row n 2 : dn1 dn1 dn3 1 dn dn4 ::: (6:5:26) dn1 dn1 dn5

Row 4 :ðs4 Þ :

DðsÞ ¼ s4 þ 2s3 þ 3s2 þ 4s þ 5:

Row 3 :ðs Þ :

2 3 4

1 2

¼ 1 2 4 1 ð5=2Þ ¼ 1 Row 1 :ðs1 Þ : 1 1 5=2 1 0 ¼ 5=2 Row 0 :ðs0 Þ : 1

1 2

2

Entries in the first column can be written by ½1; 2; 1; 1; 5=2. There are two sign changes indicating that there are two roots on the right half splane. Using MATLAB, the roots of the polynomial can be computed by using the following: d ¼ ½1 2 3 4 5 : coefficents of the polynomial :2878 j1:4161 r ¼ rootsðdÞ : Gives the roots : 1:288 j:8579 d ¼ polyðrÞ : Gives the coefficeients of the polynomial Routh array gives the information about the number of roots on the right half of the splane and not the actual roots. In the Routh array, there are divisions. If these division

(6:5:27)

Solution: Routh array is given below.

1

3

Row 2 :ðs2 Þ :

Row n3 is computed in a similar manner using rows n1 and n2. The procedure is continued until we reach row 0. The R–H criterion states that all the roots of the polynomial D(s) lie on the left half of the s-plane if all the entries in the left most column of the Routh array are nonzero and have the same sign. The number of sign changes in the leftmost column is equal to the number of roots of D(s) in the right half s-plane. Example 6.5.3 Determine the number of roots that are on the right half s-plane of DðsÞ.

3

5

4 0 5

0

2

¼

5 2

0

0

0

0

0

terms are zero, then a different technique is & needed to overcome this. Special cases: 1. Routh array has a zero in the first column of a row. 2. Routh array has an entire row of zeros. 1. If the first entry in the row ðn iÞ; i 6¼ 0 or 1 is zero, to compute the entries in the row ðn i þ 1Þ, a problem of division by zero arises. To alleviate this problem, e is assigned to 0. e is allowed to approach zero either e ! 0þ or 0 . Example 6.5.4 Consider the polynomial DðsÞ ¼ s4 þ s3 þ 2s2 þ 2s þ 1. Use the Routh array to determine the number of roots on the right half of the s plane.

6.5 Impulse Response

205

Solution: Routh array is given by 4

1

2 1

3

Row 3 : ðs Þ : Row 2 : ðs2 Þ :

1 0ðeÞ

2 0 1 0

Row 1 : ðs1 Þ : Row 0 : ðs0 Þ :

12e e 1

0 0

Row 4 : ðs Þ :

In the next step the entries in the first column are written by First column : ½1;1;e;ð12eÞ=e;1 ! fe¼ 0þ ) ½þ;þ;þ;;þg; e¼ 0 ) ½þ;þ;;þ;þg Using either of the two cases, there are two sign changes. There are two roots on the right half of the s-plane. Using MATLAB, the roots of DðsÞ are 0:1247 j1:3066; 0:6217 j0:4406. The polynomial has two roots in the right half of the s plane.& 2. Next consider the case that an entire row in the Routh array consists of zeros. To illustrate this consider the following possibilities: a. Roots are located on the imaginary axis b. Roots with symmetry about origin c. Roots with quadrant symmetry Each of these implies the following type of factors in the polynomial DðsÞ: a: ðs jbÞ ! ðs2 þ b2 Þ

(6:5:28a)

b: ðs aÞ ! ðs2 a2 Þ

c: ðs a jbÞ ! s4 þ½2ða2 þ b2 Þ 4a2 s2 þ ða2 þ b2 Þ2 (6:5:28b)

Row 4 : ðs4 Þ : Row 3 : ðs3 Þ :

1 1

1 1

2 0

Row 2 : ðs2 Þ : Row 1 : ðs1 Þ :

2 4

2 0

0 0

Row 0 : ðs0 Þ :

2

These roots produce even polynomials resulting in a row of zeros in the Routh array. The row before the row of zeros in the array gives the even polynomial identified here as D2 ðsÞ and is called the auxiliary equation. That is DðsÞ ¼ D1 ðsÞD2 ðsÞ. To complete the Routh array, take the derivative of the auxiliary equation and replace the row of zeros by the row obtained from the coefficients of the derivative of the auxiliary equation. Example 6.5.5 Consider the polynomial DðsÞ ¼ s4 þ s3 s2 þ s 2. Show that the system described by this characteristic polynomial is unstable using the Routh array. Solution: Routh array is given by Row 4 : ðs4 Þ :

1

3

Row 3 : ðs Þ : 1 2 Row 2 : ðs Þ : ð1 ð1ÞÞ ¼ 2 Row 1 : ðs1 Þ :

0

1

2

1 2

0 0

0

Noting that the row 1 has all zeros, the auxiliary equation can be written from row 2 and D2 ðsÞ ¼ ðs2 þ 1Þ ¼ 0. Since the (–) sign is irrelevant for the roots, the sign can be ignored and written as DðsÞ ¼ D1 ðsÞðD2 ðsÞÞ and D1 ðsÞ ¼ s2 þ s 2. The auxiliary polynomial D2 ðsÞ has a pair of imaginary roots at s ¼ j1 : In this simple example, the polynomial D1 ðsÞ can be factored and its roots are located at s ¼ 1; 2 indicating that DðsÞ has one root on the right half of the s-plane. If the number of roots D1 ðsÞ is higher than 2, then the Routh array can be continued in the following manner.

ðAuxiliary polynomial; D2 ðsÞ ¼ ð2s2 þ 2ÞÞ ðD02 ðsÞÞ

Entries of the first column in the above Routh array are ½1; 1; 2; 4; 2 indicating there is one

root inside the right half of the s-plane. As mentioned before, Routh array does not provide the

206

6 Systems and Circuits

roots of the polynomial. It merely identifies the number of roots in the right half s-plane. Routh array is frequently used in feedback control systems to determine the condition of stability of a control system. Note that if all the coefficients of the characteristic polynomial DðsÞ do not have the same sign, the polynomial has some roots on the right half s-plane and the corresponding system is unstable. Example 6.5.6 Using the Routh array determine the range of values for K for which all the roots of the polynomial DðsÞ ¼ s3 þ 3s2 þ 3s þ K are located inside the left half plane. Solution: The Routh array is Row 3 :ðs3 Þ : Row 2 :ðs2 Þ :

1 3 ðK 9Þ Row 1 :ðs1 Þ : 3 Row 0 :ðs0 Þ : K

3 K 0

To have all the roots of the polynomial on the left half plane, the coefficients in the first column in the Routh array must have the same signs. This implies that ð9 KÞ40 and K40. All the roots are on the right half plane if 05K59, which gives the range of values of K to keep the system stable. When K ¼ 9, row 1 has all zeros. Correspondingly, the auxiliary polynomial is D2 ðsÞ ¼ ð3s2 þ 9Þ, indicating polynomial has a pair of roots on the imaginary axis. In the case of K ¼ 0, there is a root at s ¼ 0: In the case of a root at s ¼ 0; it is evident from the polynomial that DðsÞ ¼ sD1 ðsÞ and the Routh & array can be determined starting with D1 ðsÞ. Notes: A polynomial DðsÞ with all its roots on the left half s-plane is called a strictly Hurwitz polynomial. If it has all its roots on the left half s-plane and in addition, it has simple poles on the imaginary axis, & then it is called a pseudo-Hurwitz polynomial.

yðtÞ ¼ T ejot ¼ HðjoÞejot ¼ fjHðjoÞjejfðoÞ gejot ¼ jHðjoÞjej½otþfðoÞ :

(6:5:30)

For a particular value of o ¼ o0 , (6.5.30) reduces to T ejko0 t ¼ Hðjko0 Þejko0 t ¼ jHðjko0 Þjej½ko0 tþfðko0 Þ (6:5:31) Since the system under consideration is a LTI system, the response to several frequencies can be determined by (6.4.31). The system with the frequency response HðjoÞ acts like a gate to allow certain frequencies fully or partially through or attenuated or eliminated. Example 6.5.7 Consider a LTI system with a transfer function HðjoÞ. Use (6.5.30) to find the responses to the real periodic inputs given by a: xT ðtÞ ¼

1 X

Xs ½kejko0 t ;

k¼1

b: xT ðtÞ ¼ Xs ½0 þ

1 X

d½k cosðko0 t þ y½kÞ: (6:5:32)

k¼1

Solution: a. Using (6.5.30), the output of the linear time-invariant (LTI) system is yT ðtÞ ¼ HðjoÞ

1 X

Xs ½kejko0 t

k¼1

¼

1 X k¼1

Hðjko0 ÞXs ½kejko0 t ¼

1 X

Ys ½kejko0 t :

k¼1

(6:5:33) If the input to a LTI system is periodic, then the output is also periodic with the same period and the F-series coefficients of the output and the input are related by

6.5.4 Eigenfunctions in the Fourier Domain

Ys ½k ¼ Hðjko0 ÞXs ½k ¼ jHðjko0 ÞjjXs ð½kjejðfðko0 Þþy½kÞ

In terms of the Fourier domain, we have from (6.5.20) that

jYs ½kj ¼ jXs ½kjjHðjko0 Þj; ﬀYs ½k ¼ ﬀXs ½k þ fðko0 Þ: (6:5:34b)

(6:5:34a)

6.5 Impulse Response

207

b: yT ðtÞ ¼Hð0ÞXs ½0 1 X ðjHðko0 Þjd½kÞ cosðko0 t þ fðko0 Þ þ k¼1

þyðko0 ÞÞ:

(6:5:35) &

Notes: The response given in (6.5.33) is the steadystate response of the linear system to a periodic input. A linear time-invariant system does not produce any new frequencies. The output amplitudes and the phases of the harmonics are different from the amplitudes and the phases of the input signal harmonics and are determined by (6.5.34b), for a & real periodic input. Example 6.5.8 Use the Fourier transforms to derive the output given in (6.5.33). Solution: xT ðtÞ ¼ ¼

1 X

Xs ½kejko0 t

T ¼ 2p; o0 ¼ 1:

(6:5:37)

The transfer function, the amplitude, and phase responses are given by

HðjoÞ ¼

pXs ½kdðo ko0 Þ

1 1 ; jHðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ð1 þ joRCÞ 1 þ ðoRCÞ2

ﬀ tanðoRCÞ: 1 X

(6:5:38)

pXs ½kdðo ko0 Þ

k¼1

¼

4 cosðtÞ cosð3tÞ cosð5tÞ xT ðtÞ ¼ þ ::: ; p 1 3 5

! XðjoÞ

k¼1

1 X

Solution: The Fourier series of the input waveform is given by

FT

k¼1 1 X

YðjoÞ ¼ HðjoÞ

Example 6.5.9 Find the output yT ðtÞ of the RC circuit in Fig. 6.5.4b corresponding to the periodic pulse signal shown in Fig. 6.5.4a with a period equal to T ¼ 2p.

The kth harmonic term and the steady-state output response are, respectively, given by

pfHðjko0 ÞXs ½kgdðo ko0 Þ:

k¼1

Taking the inverse transform of the transform of YðjoÞ in (6.4.35), we have yT ðtÞ ¼

1 X

fHðjko0 ÞXs ½kgejko0 t :

(6:5:36)

&

1 jHðjko0 Þj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ﬀHðjko0 Þ 1 þ ðko0 RCÞ2 ¼ tan1 ðko0 RCÞ:

(6:5:39)

k¼1

Fig. 6.5.4 (a) Periodic pulse waveform and (b) RC circuit

(a)

(b)

208

6 Systems and Circuits

4 cosðt tan1 ð1ÞÞ cosð3t tan1 ð3ÞÞ cosð5t tan1 ð5ÞÞ pﬃﬃﬃ pﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃ yT ðtÞ ¼ þ ::: : p 3 10 5 26 1 2

(6:5:40)

Note the kth harmonic input produces the kth harmonic output illustrated below.

4 4 cosðko0 tÞjo0 ¼1 ! qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ cosðko0 t tan1 ðko0 RCÞÞjo0 ¼1 : p p 1 þ k2 o20 ðR2 C2 Þ

Attenuation is proportional to ð1=kÞ and the phase shift is tan1 ðko0 RCÞ in the kth harmonic term. The RC circuit is a low-pass filter. The low frequencies have smaller attenuations and higher frequencies are significantly & attenuated.

hðtÞ ¼

6.6 Step Response ¼

sðtÞ ¼ hðtÞ uðtÞ ¼

Z1

hðbÞuðt bÞdb ¼

1

Zt hðbÞdb: 1

(6:6:1) In the above equation the variable of integration is b not t and uðt bÞ ¼ 0 for t5b. The step response can be obtained by integrating the impulse response and the impulse response can be obtained by differentiating the step response. That is,

hðtÞ ¼

dsðtÞ : dt

(6:6:2)

(6:6:3a)

The step response is

sðtÞ ¼

The step response of a continuous time LTI system is the response to a step input xðtÞ ¼ uðtÞ. The step response sðtÞ; is related to the impulse response hðtÞ (see (6.5.10)) and

1 t=RC e uðtÞ: RC

(6:5:41)

Zt hðbÞdb 1 Zt

1 ðb=ðRCÞÞ uðbÞdb ¼ ð1 eðt=RCÞ ÞuðtÞ: e RC

1

(6:6:3b) The impulse response from the step response by dsðtÞ dð1 et=RC ÞuðtÞ ¼ dt dt 1 t=RC t=RC Þ þ uðtÞ e ¼ dðtÞð1 e RC

hðtÞ ¼

¼ ð1=RCÞet=RC uðtÞ: Note ð1 et=RC Þ is continuous at t ¼ 0 and dðtÞð1 et=RC Þ is zero, see (1.4.5). The impulse and the step responses are sketched in Fig. 6.6.1. The rise time of the RC circuit is the time required for a unit step response to go from 10 to 90% of its final value. It is given by tr ¼t2 t1 ; sðt1 Þ ¼ ð1 et1 =RC Þ ¼ :1;

Example 6.6.1 Determine the step response of the RC circuit in Example 6.3.1 from the impulse response and vice versa. Solution: From (6.5.16a), the impulse response is

sðt2 Þ ¼ 1 et2 =RC 0:9 ¼ et1 =RC ; and 0:1 ¼ et2 =RC ; tr ¼ ðt2 t1 Þ ¼ RC lnð9Þ ¼ 2:197RC:

(6:6:4)

6.6 Step Response

209

Fig. 6.6.1 (a) Impulse response and (b) step response

(a)

Rise time is a measure of how fast the system responds to an input. It is related to the bandwidth & of the circuit and we will discuss this shortly. Rise time and the 3 dB bandwidth: The output transform and the transfer function of the RC circuit given Fig. 6.3.4a are 1 YðjoÞ ¼ XðjoÞ ¼ HðjoÞXðjoÞ; 1 þ joRC 1 : (6:6:5) HðjoÞ ¼ 1 þ joRC The amplitude and the phase responses of the transfer function are 1 20ogjHðjoÞj ¼ 20 log qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ dB; 1 þ ðoRCÞ2 ﬀHðjoÞ ¼ tan1 ðoRCÞ:

(6:6:6)

The responses are shown in Fig. 6.6.2 for positive frequencies. Note the amplitude response is even and the phase response is odd. The amplitude at o ¼ 0 is 1 and in dB, it p isﬃﬃﬃ0 dB. At o ¼ 1=RC, the magnitude is equal to 1/ 2 and in dB this is 3 dB. The 3 dB frequency (or the half-power) is

Fig. 6.6.2 (a) Amplitude response and (b) phase response

(b)

o3dB ¼ 1=RC; or f3dB ¼ 1=2pRC Hertz

(6:6:7)

The amplitude response jHðjoÞj decreases smoothly for higher frequencies and goes to zero at infinity. The rise time is related to the 3 dB bandwidth and is tr ¼ 2:197=ð2pf3dB Þ ¼ :35=f3dB :

(6:6:8)

In summary, the RC circuit is a simple low-pass filter passing frequencies between 0 and f3 dB with small attenuations and all the higher frequencies are attenuated significantly. The phase response is zero at o ¼ 0: At the 3 dB frequency, it is equal to ðp=4Þ and at the infinite frequency the phase & response reaches ðp=2Þ rad or 90 : Ideal integrator: The transfer function of the ideal integrator is HðsÞ ¼ 1=s. The amplitude and the phase responses are, respectively, given by HðjoÞ ¼ð1=joÞ ¼ ðj=oÞ; jHðjoÞj ¼ 1=joj; ﬀHðjoÞ ¼ p=2; o40: (6:6:9) This function represents an ideal integrator by noting that if the input is a sinusoid, say cosðotÞ, the output of the integrator and its transform are

(a)

(b)

210

6 Systems and Circuits

yðtÞ ¼

Zt

xðtÞdt ¼

1 1 p sinðotÞ ¼ cos ot o o 2

1

(6:6:10) jﬀHðjoÞ

YðjoÞ ¼ HðjoÞXðjoÞ; HðjoÞ ¼ jHðjoÞje jHðjoÞj ¼ 1=joj; ﬀHðjoÞ ¼ j sgnðoÞ:

;

(6:6:11)

The amplitude response is inversely proportional to joj; the phase response for o40 is ðp=2Þ, a constant. Since the amplitude gain of an ideal integrator is ð1=jojÞ, it suppresses the higher frequency components and enhances the low-frequency components. The noise signals contain mostly high-frequency components, and the integrator reduces the size of the high-frequency components. Ideal differentiator: The transfer function of an ideal differentiator is HðsÞ ¼ s; HðjoÞ ¼ jo:

(6:6:12)

The amplitude and phase responses are given by jHðjoÞj ¼ jojand ﬀHðjoÞ ¼

p ; o40: 2

(6:6:13)

Consider that a sinusoidal function xðtÞ ¼ cosðotÞ is passed through a differentiator, then

yðtÞ ¼

dxðtÞ d cosðotÞ ¼ ¼ o sinðotÞ dt dt

p ¼ o cos ot þ : 2

Example 6.6.2 Show the circuit shown in Fig. 6.6.3 can be used as a differentiator.

Fig. 6.6.3 RL circuit

Solution: The output transform is YðjoÞ ¼

joL jojL XðjoÞ; jHðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; R þ joL R2 þ ðoLÞ2 p oL : (6:6:15) ﬀHðoÞ ¼ tan1 2 R

For small frequencies, i.e., joj ¼ ðR=LÞ, the output transform can be approximated. Noting the Ftransform derivative theorem, it follows that the circuit acts like a differentiator. That is,

(6:6:14)

Note the amplitude response increases linearly with frequency o and phase response is constant and is equal to ðp=2Þ rad for positive frequencies. From the amplitude response expression, we see that the high frequencies are enhanced. Most corrupted signals contain noise components that are of high

Fig. 6.6.4 RC circuit response to a pulse input

frequencies. Using a derivative function enhances the noise signal much more than a low-frequency signal. Derivative function is used to sharpen a signal. For example, to sharpen an image at the edges, we use a derivative function. Note the discontinuity in the phase response at o ¼ 0.

FT

YðjoÞ jo½LxðjoÞ ! L

dxðtÞ yðtÞ: (6:6:16) dt

&

Example 6.6.3 Find the response of the RC circuit in Fig. 6.3.4a to the input pulse xðtÞ ¼ AP

t T2 : T

(6:6:17)

6.6 Step Response

211

Solution: The transfer function HðjoÞ, its impulse response hðtÞ, output frequency response and the response yðtÞ using the convolution integral, we have YðjoÞ ¼HðjoÞXðjoÞ

HðjoÞ ¼

1 ð1 þ joRCÞ

yðtÞ ¼

Z1

FT

! hðtÞ xðtÞ ¼ yðtÞ;

FT

!

(6:6:18)

1 t=RC e uð t Þ ¼ hð t Þ RC (6:6:19)

hðt bÞxðbÞdb ; hðt bÞ

Assuming the unit step input response is hu ðtÞ, the delayed step input response is hu ðt TÞ. The response to the pulse input is & Aðhu ðtÞ hu ðt TÞÞ: Simple frequency analysis of the RC circuit in the last example: The Fourier transforms of the input, the transfer function and the output transform are t T2 sinðoT=2Þ joT ¼ At e 2; XðjoÞ ¼F AP T ðoT=2Þ 1 : (6:6:22) HðjoÞ ¼ 1 þ joRC

1

8 < 1 ðtbÞ=RC ; b5t e ¼ RC : : 0; b4t

(6:6:20)

9 8 0; t50 > > > > > > > > Zt > > > > A > > ðtbÞ=RC > > > db; 0 5 t 5 T > = < RC e yðtÞ ¼ 0 > > > > > > ZT > > > > A > > ðtbÞ=RC > > 4 > > db; t T e > > ; : RC 0 8 t50 > < 0; t=RC ¼ Að1 e Þ; 05 t 5 T : > : T=RC ðtTÞ=RC Að1 e Þe ; t 4T (6:6:21) Note that the input pulse is the same as in Example 2.2.4, except the pulse is of width T instead of 2T and the pulse started at t ¼ 0 rather than at t ¼ T. The function in (6.6.21) is sketched in Fig. 6.6.4. The response can be visualized by the following argument. For t50, the input is zero and the output is zero as well. At t ¼ 0 we have a step input and the capacitor voltage cannot charge instantaneously and the voltage across the capacitor starts at 0 and increases exponentially with a time constant RC. At t ¼ T, the input becomes zero and for t4T the charge across the capacitor discharges through the resistor and the capacitor voltage decreases exponentially from the peak value of Að1 eT=RC Þ to zero as t ! 1. Another way to derive (6.6.2) is that the input pulse function is AP½ðt ðT=2ÞÞ=T ¼ AuðtÞ Auðt TÞ.

YðjoÞ ¼

At sincðoT=2ÞejoT=2 : ð1 þ joRCÞ

(6:6:23)

We will sketch the amplitude of the output transform by considering two special cases: a. Pulse width T is very large compared to the time constant t ¼ RC (i.e., T RCÞ b. Time constant is very small compared to the time constant (i.e., T RCÞ. Now jYðjoÞj ¼ jHðjoÞjjXðjoÞj:

(6:6:24)

For the two special cases, the functions jHðjoÞj and jXðjoÞj are sketched in Fig. 6.6.5a,b. In case a, the 3 dB bandwidth is assumed to be much larger than the main lobe width of the response. That is, ð1=ð2pÞf3dB Þ ¼ ð1=RCÞ 1=T or T RC and the function jHðjoÞj is essentially flat in the range joj51=T. In this frequency band the amplitude of the output transform is approximately equal to the magnitude of the input transform and we can approximate and jYðoÞj kjXðoÞj; k a constant. The output pulse will be a good approximation of the input pulse. In case b, ð1=ð2pf3dB ÞÞ 1=T or T RC. From Fig. 6.6.5 we see that jXðjoÞj is essentially flat in the 3 dB frequency range. That is, jYðjoÞj jHðjoÞj in this range. This indicates that the amplitude of the output transform looks more like the magnitude of the system transform in this case. We are interested in the input signal transform, not the transform of the system. When a signal is passed through a system, the bandwidth of the system must be much larger than the

212

6 Systems and Circuits

Fig. 6.6.5 Frequency analysis of an RC circuit with a pulse input(a) T ¼ RC, (b) T ¼ RC

(a)

(b) bandwidth of the input signal in order the output response to have some resemblance of the input. Let us now consider some simple ideas about the response of a sequence of pulses. The detection of the existence of the pulse at the output can be improved by increasing the amplitude of the input pulse or increasing the width of the pulse or both. Increasing the amplitude increases the power requirements on the input. Increasing the pulse width implies that the number of pulses that can be transmitted per unit time has to be reduced. In addition the pulses are not band limited. Since the RC circuit is a LTI system, it is conceivable that if the input consists of a set of pulses, the output response will be a sum of the individual responses of the pulses with appropriate delays in the pulse responses. The sum may become unbounded. Next we will consider the process of removing the effects of the RC circuit. This type of a situation appears in measuring the voltage across a

component, seeing a picture through a lens and many others. In these measurements, the signal is affected by the measuring device or the system we visualize with. If the bandwidth of these devices is much, much larger than the signal bandwidths, then the effect of the measuring devices is minimal. Removing the effects of the transmission system from the received signal is an important problem and this process is called the deconvolution and is discussed next. Deconvolution: Let the output response of a LTI system with the transfer function HðjoÞ, and the corresponding impulse response hðtÞ, is given by yðtÞ ¼ hðtÞ xðtÞ

FT

! HðjoÞXðjoÞ ¼ YðjoÞ: (6:6:25)

To recover the signal, xðtÞ from yðtÞ, consider Fig. 6.6.6. The first block represents a system and

6.7 Distortionless Transmission

213

Fig. 6.6.6 Deconvolution

the second block identified by HR ðoÞ represents a system to recover the original signal. The output transfrom; ZðjoÞ is related to the input transform, assuming no loading effects, is ZðjoÞ ¼ HR ðjoÞHðjoÞXðjoÞ:

(6:6:26)

To recover the signal, it is desired to have jHðjoÞjjHR ðjoÞj k, a gain constant in the frequency range of interest. Since there is an inherent delay in every system, say t seconds, this delay can be incorporated and zðtÞ ¼ kxðt tÞ. This implies that

yðtÞ ¼H0 xðt t0 Þ ðH0 and t0 40 are some constants:Þ: (6:7:1) For simplicity, assume H0 40. We essentially tried to obtain a distortionless signal in using the deconvolution process in Fig. 6.6.6. Taking the transform of yðtÞ, the output transform and the transfer functions are as follows: YðjoÞ ¼ H0 ejot0 XðjoÞ ! HðjoÞ ¼ H0 ejot0 : (6:7:2)

HðjoÞHR ðjoÞ ¼ k ejo t or HR ðjoÞ ¼ k ejo t =HðjoÞ: The amplitude and the phase responses of the dis(6:6:27) tortionless system are A circuit that gives the transfer function HR ðjoÞ in (6.6.27) may not always be possible. For example, if HðjoÞ ¼ 0 at some frequency o ¼ o i , the function HR ðjoÞ goes to infinity at this frequency. Therefore, HR ðjoÞ can only be approximated. In terms of the time domain, in the ideal case, the inverse transform of ZðjoÞ is given by zðtÞ ¼ F1 ½ZðjoÞ ¼ F1 ½HðjoÞHR ðjoÞXðjoÞ ¼ hR ðtÞ hðtÞ xðtÞ

(6:6:28)

There is perfect deconvolution if hðtÞ hR ðtÞ ¼ dðtÞ and zðtÞ ¼ dðtÞ xðtÞ ¼ xðtÞ:

jHðjoÞj ¼ jH0 j; ﬀHðjoÞ ¼ ot0 :

(6:7:3)

These functions are shown in Fig. 6.7.1, where H0 is assumed to be positive. This implies that all frequencies are attenuated (or amplified) by the same amount. It is referred to as an all-pass system. The phase response is linear. The delay associated with an ideal delay line, a LTI system, can be seen by considering a sinusoidal input xðtÞ ¼ cosðotÞ. The output is H0 cosðoðt t0 ÞÞ. The amplitude response is the same for all frequencies. In addition, the output is H0 cosðoðt t0 ÞÞ ¼ H0 cosðot ot0 Þ. That is, the time shift is t0 and the phase shift is ot0 and the phase is linearly proportional to the frequency o with a slope of ðt0 Þ

6.7 Distortionless Transmission 6.7.1 Group Delay and Phase Delay A system is called distortionless if the output is the same as the input except the signal may be attenuated by the same amount for all frequencies along with a delay of t0 seconds. A distortionless system has the output

The phase response in (6.7.3) and the delay are respectively given by yðoÞ ¼ ﬀHðjoÞ ¼ ot0

(6:7:4)

214

6 Systems and Circuits

Fig. 6.7.1 Amplitude and phase responses of a distortionless system

t0 ¼

dﬀHðjoÞ dyðoÞ dðot0 Þ ¼ ¼ : do do do

(6:7:5)

Group delay: The group delay deals with a group of frequencies (usually referred as delay for simplicity). The result in (6.7.5) can be used to find the delay associated with a transfer function. HðjoÞ ¼ jHðjoÞjejyðoÞ :

dyðoÞ d tan1 bo=ðc ao2 Þ : ¼ Tg ðoÞ ¼ do do (6:7:10) ¼

ðc ao2 Þb boð2aoÞ

1 2

ðbo=ðc ao2 ÞÞ þ1 bðc þ ao2 Þ ¼ : ðbo2 Þ þ ðc ao2 Þ2

ðc ao2 Þ2

(6:7:6)

(6:7:11)

The group delay deals with a group of frequencies (usually referred as delay for simplicity). It is a nonlinear function of frequency and is defined by

Figure 6.7.2 illustrates the delay function in (6.7.11) & and is not constant for all o: In a later section, filters will be designed that have transfer functions with nonlinear phase. When signals passed through such filters, different frequencies are delayed differently. This is not critical for speech signals, as the human ear compensates for small delays. It is a problem in transmitting data and compensation is necessary so that the filter delay equalizer combination has approximately a constant delay in the desired frequency range. A second-order delay equalizer has the form

Tg ðoÞ ¼

dyðoÞ : do

(6:7:7)

Example 6.7.1 Find the group delay associated with the transfer function 1 ; a; b; c40: HðjoÞ ¼ 2 ðc ao Þ þ bðjoÞ

d tan1 ðxÞ 1 dx ¼ : : Use the identity dy 1 þ x2 dy (6:7:8) Solution: The amplitude, the phase response, and the group delay are given by 1 jHðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ðc ao2 Þ þ ðboÞ2 yðoÞ ¼ ﬀHðjoÞ ¼ tan

1

bo : c ao2 (6:7:9)

Hci ðjoÞ ¼

ðbi o2 Þ jai o ; ai ; bi 40: ðbi ai o2 Þ þ jai o

(6:7:12)

The amplitude and phase responses are qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ðbi o2 Þ2 þ ðai oÞ2 jHci ðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ 1; ðbi o2 Þ2 þ ðai oÞ2 ai o : ﬀHci ðjoÞ ¼ 2 tan1 bi o 2

(6:7:13)

6.7 Distortionless Transmission

215

Fig. 6.7.2 Group delay characteristics in Example 6.7.1

Tg(ω) b c

ω

0

Noting that the phase angles of a product of transfer functions add, the amplitude and phase response of a cascade of n second-order sections result in n

n

i¼1

i¼1

yT ðtÞ ¼AjHðjo0 Þj cosðo0 t þ y0 Þ ¼AjHðjo0 Þj cos½o0 ðt þ ðy0 =o0 ÞÞ:

(6:7:17)

To have a distortionless transmission, the output must have the form

Hc ðjoÞ ¼ P Hci ðjoÞ; jHc ðjoÞj ¼ P jHci ðjoÞj ¼ 1 ﬀHc ðoÞ ¼

N X

yT ðtÞ ¼ B cos½o0 ðt t0 Þ:

ﬀHci ðoÞ:

(6:7:14)

i¼1

The parameters a0i s and b0i s can be adjusted so that the amplitude response of the filter cascaded by the delay equalizer has the same magnitude as the filter function HF ðjoÞ and the corresponding phase angle has approximately linear phase (i.e., approximately constant delay) in the desired frequency range. That is, HðjoÞ ¼ HF ðjoÞHc ðjoÞ; jHðjoÞj jHF ðjoÞj ﬀHðjoÞ ¼ﬀHF ðjoÞ þ ﬀHc ðjoÞ; d ﬀHðjoÞ constant: do

(6:7:15a) (6:7:15b)

Phase delay: Consider the input xT ðtÞ ¼ A cosðo0 tÞ to a LTI system with a transfer function HðjoÞ ¼ jHðjoÞjejﬀHðjoÞ ; ﬀHðjoÞ ¼ yðoÞ:

(6:7:18)

(6:7:16)

If the input to a LTI system is a sinusoid then the output is also a sinusoid at the same frequency, although the output may have a different amplitude and phase. See Example 6.5.6. Let the output of the system to the input xT ðtÞ is

Comparing this with (6.7.17), the time delay between the input and the output at the frequency f0 ¼ o0 =2p is t0 ¼ y0 =o0 . For a single frequency, phase delay is appropriate. The phase delay Tp ðoÞ in terms of the system phase response yðoÞ ¼ argðHðoÞÞ is defined by Tp ðoÞ ¼ yðoÞ=o:

(6:7:19)

Notes: For rational transfer functions, yðoÞ is a transcendental function, whereas the group delay is a rational function of o2 making it easier for filter & design. Example 6.7.2 Find the phase delay and the group delay of the transfer function HðjoÞ ¼ ½ð1 joÞ=ð1 þ joÞ:

(6:7:20a)

Solution: The phase, the group delay, and the phase delay responses are yðoÞ ¼ p 2 tan1 ðoÞ ! Tg ðoÞ ¼

dyðoÞ do

2 yðoÞ : (6:7:20b) & ; Tp ðoÞ ¼ 2 1þo o Earlier, we have seen that the delay associated with a system is a function of its phase response. Transfer ¼

216

6 Systems and Circuits

functions of stable systems can have the same amplitude response with different phase and delay responses. There are three important systems to consider. These are minimum phase, mixed phase, and maximum phase systems.

B0 ¼

Consider the transfer function of a system k

Pðs þ pm Þ

:

hðtÞdt; 1 Z1

1 Hð0Þ

HðjoÞdf; o ¼ 2pf:

(6:8:1)

1

Pðs þ zk Þ HðsÞ ¼ K

Z1

1 T0 ¼ hð0Þ

(6:7:21)

m

For stability reasons all the poles located at s ¼ pk are located on the left half of the s-plane. If the zeros of the transfer function s ¼ zk are on the negative half of the s-plane, then the system is a minimum phase system. If some zeros are on the right half s-plane and some are on the left half s-plane, the system is a mixed phase and if all the zeros are located on the right half of the splane, then the system is a maximum phase system.

6.8 System Bandwidth Measures In Sections 4.2.2, bandwidth measures of a signal were briefly studied. The concentration was on the time–bandwidth product and illustrated examples, wherein the bandwidth is inversely proportional to time width of the signal. Similar ideas can be used using the duration of the impulse response of a system and the system bandwidth. In Section 6.3, a simple, but a practical measure, the half-power or the 3 dB bandwidth was considered. This width is the range of frequencies overpwhich the magnitude ﬃﬃﬃ of the function exceeds ð1= 2Þ of its maximum. Half-power bandwidth comes from the fact that the square pﬃﬃﬃ of the magnitude is power and 20logð1= 2Þ ¼ 3dB. There are different measures that are used and some of these are considered.

6.8.1 Bandwidth Measures Using the Impulse Response hðtÞ and Its Transform Hðj!Þ The time and the frequency durations of the impulse response hðtÞ are defined by

From the central ordinates theorems of the F-transforms, the time width times the bandwidth is equal to 1. This definition for the bandwidth makes use of the spectrum on both sides. That is, T0 B0 ¼ 1:

(6:8:2)

For the one side case, which is what we mostly use, divide the frequency width by 2. Example 6.8.1 Using the above measures show that T0 ¼ 1 and B0 ¼ 1 for the following: a: x1 ðtÞ ¼ PðtÞ

FT

! sinc

o

¼ X1 ðjoÞ; 2 FT 1 b: x2 ðtÞ ¼ eat uðtÞ ! ¼ X2 ðoÞ; a þ jo rﬃﬃﬃ FT 2 c: x3 ðtÞ ¼ eat ! paeo2 =4a

(6:8:3)

Solution: a. In Chapter 4, the areas of the rectangular and the sinc functions were considered and the results are T0 ¼ 1; B0 ¼ 1. b. Noting that hð0Þ ¼ 1=2 ðhðtÞ is discontinuous at t ¼ 0Þ and Hð0Þ ¼ 1=a , it follows that Z1 2

e 1

at

Z1 2 a 1 do uðtÞdt¼ and a 2p a þ jo 1 Z a2 1 1 ¼ do 2p 1 a2 þ o2 Z a2 1 1 a do ¼ ; ¼ p 0 a2 þ o 2 2

T0 ¼ ð2=aÞ; B0 ¼ ða=2Þ; T0 B0 ¼ 1:

(6:8:4)

c. By making use of integral tables, the time width and the bandwidth of the Gaussian pulses both & come out to be one and the product is one.

6.8 System Bandwidth Measures

Time functions can take both positive and negative values, some authors use the magnitudes or the squares of the time functions hðtÞ in defining the time width in (6.8.2a). Others use moments to define the time and bandwidths. In the following, the bandwidth measures that are simple and practical will be considered.

6.8.2 Half-Power or 3 dB Bandwidth In Section 6.6, an RC circuit was considered. On the amplitude spectrum, the 3 dB frequency was identified (see Fig. (6.6.2a)). The half-power or the 3 dB bandwidth is a practical measure and is widely used in systems and circuit theory, especially in filter designs. In identifying the 3 dB bandwidths, only positive frequencies are considered. Example 6.8.2 Show the 3 dB bandwidth of the Gaussian function is W (Carlson (1975)). HðjoÞ ¼ eðlnð2=2Þðo=WÞ ðnote; Hðj0Þ ¼ 1Þ: Solution: The half-power frequency is equal to W since

217

value of the spectrum. This measure is simple and it does nottake into consideration any ripples in jHðjoÞj function between the two 3 dB frequencies. A more generalized measure that takes into consideration the ripples by making use of integrals in computing the bandwidths. These methods are used in random signal analysis, as the spectrum of noisy signals have many peaks. For a good discussion on this topic, see Peebles (2001). These measures have been developed using signals rather than systems. To make it uniform in our discussion we will use XðjoÞ rather than HðjoÞ. When we consider examples of transfer functions, HðjoÞ will be used.

6.8.3 Equivalent Bandwidth or Noise Bandwidth The equivalent noise bandwidth is obtained by equating the areas contained in the signal energy spectrum with a pulse spectrum of bandwidth Weq ¼ WN rad/s. Figure 6.8.1 illustrates a signal spectrum and noise equivalent is computed as follows: Z1

jXðjoÞj2 do ¼ jXðjoÞj2max 2Weq

1 2 1 jHðjo3dB Þj2 ¼ ¼ ðe2ðlnð2Þ=2Þ ðo3dB =WÞ Þjo3dB ¼W 2 1 (6:8:5) & ) ¼ e lnð2Þ : 2

Although the 3 dB bandwidth is the most common one, we could obviously define 6 dB bandwidth or any other value for the bandwidth measure. In summary the 3 dB bandwidth computes the width by considering the peak valuepﬃﬃof ﬃ the spectrum and a value (or values) of the (1/ 2) below the maximum

Fig. 6.8.1 Noise equivalent bandwidth

R1 ) Weq ¼

jXðjoÞj2 do

0

jXðjoÞj2max

; Beq ¼

Weq : 2p

(6:8:6)

If the system bandwidth of the transfer function HðjoÞ is of interest, replace XðjoÞ by HðjoÞ in (6.8.6). Example 6.8.3 Determine the noise equivalent bandwidth of the filter transfer function

218

6 Systems and Circuits

jHðjoÞj2 ¼

1 ¼ ½1=ð1 þ ðo=WÞ2 Þ; 1 þ ðoRCÞ2 o ¼ 2pf; W ¼ 1=RC:

Solution: Using (6.8.6) and jHðjoÞjmax ¼ 1, we have Z1 1 W2 Weq ¼ do ¼ W tan1 ðo=WÞ1 0 2 2 2p ðW þ o Þ 0

¼Wp=2:

(6:8:7)

) Beq ¼ ½1=4RCHz:

(6:8:8)

The 3 dB frequency of the RC circuit, f3dB ¼ 1=2pRC is related to the equivalent noise bandwidth and Beq ¼ 1:57B3dB . The equivalent noise bandwidth works as well for signals that have spectrum in the middle, such as the band-pass spectrum. In such cases, using the center frequency o0 for the peak in the amplitude response of the band-pass filter, the equivalent bandwidth is R1 Weq ¼

jHðjoÞj2 do

0

jHðjo0 Þj2

Weq Hz: ; Beq ¼ 2p

by the bandwidth. We can define the root mean square (RMS) bandwidth as Z1

W2RMS ¼

o2 jXno ðjoÞj2 do ¼

R1

2 2 1 o jXðjoÞj do : R1 2

jXðjoÞj do

1

1

(6:8:12) Example 6.8.4 Compute the RMS bandwidth (see Peebles (2001).) and compare it with the 3 dB frequency which is given by jXðjoÞj2 ¼

10 ½1 þ ðo=10Þ2 2

(6:8:13)

:

Solution: Using integral tables, we have Z1 Z1 10 do 5 do ¼ 10 ¼ 50p: ½ð1 þ ðo=10Þ2 2 ½100 þ o2 2 1

1

(6:8:14a) Z1

(6:8:9) &

2

2

o jXðjoÞj do ¼ 10

1

5

Z1 1

o2 ½100 þ o2 2

do ¼ 5000p: (6:8:14b)

6.8.4 Root Mean-Squared (RMS) Bandwidth

FRMS

The RMS bandwidth comes from the statistical measures, where the variance is a measure of the spread of a density function. Consider the low-pass energy spectral density shown in Fig. 6.8.1. The area under this function is the energy Z1 1 (6:8:10) E¼ jXðjoÞj2 do: 2p 1

Now define the normalized energy spectral density function by 2

2

jXno ðjoÞj ¼ jXðjoÞj =E:

5000p ¼ 100; oRMS ¼ 10 rad=s; 50p ¼ 1:5915 Hz:

o2RMS ¼

(6:8:11)

It is real, even, and positive and the area under the function is 1. It has the same properties as a probability density function (PDF). In the PDF case, we define the variance as a measure of the spread of the density function. In this case the spread is measured

1 10 jXðjo3dB Þj2 ¼ 12 jHð0Þj2 ¼ 10 ¼ 5 ¼ 2 ½1 þ ðo3dB =10Þ2 2 : ) o3dB ¼ 6:436 or f3dB ¼ 1:243Hz: In this special case, the 6 dB bandwidth comes out to be the same as the RMS bandwidth. The concepts of RMS bandwidth can be easily extended to band-pass spectra. Assuming that most of the spectra is around o0 , the RMS bandwidth is given by W2RMS

¼

4

R1 0

ðo o0 Þ2 jXðjoÞj2 do R1 2 0 jXðjoÞj do

(6:8:15)

RMS bandwidth is more general than the 3 dB bandwidth, as it can handle general spectra with several peaks and valleys in the passband of the & energy spectral density.

6.9 Nonlinear Systems

219

6.9 Nonlinear Systems In this section we will consider simple nonlinear systems and illustrate the difficulties in the spectral analysis of the responses. A nonlinear system is described by a time domain relationship between the input and the output. This can be expressed in the form of a graphical representation or in terms of a general output function yðtÞ ¼ gðxðtÞÞ, where the output function gð:Þ is a complicated closed form expression or in terms of a power series of xðtÞ. A system is nonlinear if it has components that have nonlinear characteristics, such as a diode. In many cases, nonlinear systems are approximated by linear systems, as they are easier to handle, see Ziemer and Tranter (2002). The system described by a polynomial function of the input xðtÞ, such as yðtÞ ¼

n X

ai xi ðtÞ:

(6:9:1)

i¼0

Example 6.9.1 Find the output yðtÞ defined below and sketch the one-sided line spectra of the input xðtÞ ¼ B1 cosðo1 tÞ þ B2 cosðo2 tÞ; B1 ; B2 40 with oi ¼ 2pfi ; i ¼ 1; 2 and f2 4f1 and yðtÞ, see Ziemer and Tranter (2002). yðtÞ ¼a0 þ a1 xðtÞ þ a2 x2 ðtÞ; ai 6¼ 0; xðtÞ ¼B1 cosðo1 tÞ þ B2 cosðo2 tÞ; B1 ; B2 40: (6:9:2)

Solution: The output is yðtÞ ¼ a0 þ a1 B1 cosð2pf1 tÞ þ a1 B2 cosð2pf2 tÞ þ a2 B21 cos2 ðo1 tÞ þ a2 B22 cos2 ðo2 tÞ (6:9:3) 1þ a2 ð2B11 B2 Þ cosðo1 tÞ cosðo2 tÞ DCoffsetterm ¼½a0 þ a2 B21 þ a2 B22 2 2 Linearterms þ½a1 B1 cosðo1 tÞþa1 B2 cosðo2 tÞ 1 þ a2 ½B21 cosð2o1 tÞþB22 cosð2o2 Þt Harmonicterms 2 þa2 B1 B2 ½cosðo1 þo2 Þtþcosðo2 o1 Þt Inter modulationterms:

is linear if all ai 0s are zero except a1 : If any of the other ai 0s are nonzero, then the system is nonlinear. An example is a device that has saturation nonlinearity. It has a voltage to current characteristic that is linear within a range and outside that range, the voltage saturates, see Fig. 6.9.1a. A device that may have this type of a characteristic is a resistor. The Ohm’s law says that v ¼ Ri is valid in a certain range of currents and voltages. Outside this range, the resistor is a nonlinear component. A hard limiter is an important example. The output voltage is 1 if the input voltage is positive and 1 if the input voltage is negative (see Fig. 6.9.1b.).

Fig. 6.9.1 Examples of nonllinear input–output characteristics

(a)

(6:9:4) Figure 6.9.2 gives the input and the output onesided amplitude line spectra. System nonlinearity created a DC term, linear terms (frequencies f1 and f2 ), harmonic distortion terms (frequencies, 2f1 and 2f2 ), and inter modulation terms (sums and differences of the input frequencies, ( f2 f1 Þ and ð f1 þ f2 Þ). Note that the output of a linear time-invariant system has the same frequencies as the input with possible changes in the amplitudes and phases. Nonlinear system generates new & frequencies.

(b)

220

6 Systems and Circuits

Fig. 6.9.2 Example 6.9.2, (a) Input line spectra and (b) Output line spectra

(a)

6.9.1 Distortion Measures Most systems have inherent nonlinear components. It may be desirable to operate them in the linear region, if possible. Amplification is a good example, where the nonlinearities may be small and the distortions will be small enough that they can be tolerated. The next question is how does one measure the distortions due to a nonlinear system? A simple way is compare a nonlinear system time response to a linear system time response. To achieve these measures, start with a single input, a sinusoid, say xðtÞ ¼ cosðo0 tÞ and measure the distortion due to the nonlinearities in the system described xðtÞ by X yðtÞ ¼ ai xi ðtÞ; where ai 0 s some constants: i

(6:9:5) The powers of the cosine functions can be expressed in terms of sine and cosine terms, where the frequencies will be multiples of the input frequency, i.e., we will have harmonics. The output can be written in terms of trigonometric Fourier series yðtÞ ¼ Y½0 þ

1 X k¼1

A½k cosðko0 tÞ þ

1 X

B½k sinðko0 Þt:

(b) Obviously the frequency of interest f0 ¼ o0 =2p should be in the passband of the signal. Manufacturers of stereo systems provide literature that gives these numbers in terms of dBs for their systems. Clearly, if the distortion terms D½k0 s; k 6¼ 1 are negligible, then the nonlinear system comes close to a linear system.

6.9.2 Output Fourier Transform of a Nonlinear System In the following example, a system with a polynomial nonlinearity is considered and illustrates the effect of the nonlinearities in terms of the input and the output frequencies. Example 6.9.2 Let the input xðtÞ and the output yðtÞ in terms of the input are as given below. Noting F½x2 ðtÞ ¼ ð1=2pÞ½XðjoÞ XðjoÞ, sketch the output spectrum assuming xðtÞ

FT

! XðjoÞ ¼

yðtÞ ¼ a0 þ a1 xðtÞ þ a2 x2 ðtÞ

Yh o i ; W

FT

! YðjoÞ:

(6:9:8)

k¼1

(6:9:6) The constants Y½0; A½k0 s; and B½k0 s are functions of the constants a0i s and the powers of the input sinusoid. The kth distortion term is measured by the ratio sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ A2 ½k þ B2 ½k D½k ¼ : (6:9:7) A2 ½1 þ B2 ½1

YðjoÞ ¼ a0 2pdðoÞ þ a1 XðjoÞ þ ða2 =2pÞ½XðjoÞ XðjoÞ:

(6:9:9)

Solution: Convolution of two identical rectangular pulses is a triangular pulse (see Example 2.3.1.) and h i h i YðjoÞ ¼ a0 2pdðoÞ þ a1 P o þ ða2 =2pÞL o : W W (6:9:10)

6.10 Ideal Filters

221

Note that the approximation is valid for small excursions of x from x0 and we can neglect higherorder terms. It is a linear relationship between small changes in both the input and output related by the slope of the function mat x ¼ x0 , see Nise (1992).

6.10 Ideal Filters Fig. 6.9.3 Output transform of a nonlinear system

The three parts are explicitly shown in Fig. 6.9.3. Note that the width of the triangular pulse is 2 W. Therefore, the bandwidth of the output signal is two times that of the input signal. If a system has a second-order nonlinearity, then the frequency width of the output signal will be doubled that of the input signal. A system with a nth order nonlinearity, then the bandwidth of the output signal will increase from B Hz to ðnBÞ Hz: Most signals are not band limited and the transmission systems have a limited band& width and filtering is necessary. Frequency analysis of a nonlinear system is difficult and may not even be possible. If it can be approximated by a linear system, then frequency domain analysis provides useful information. Time domain analysis is simpler for nonlinear systems.

6.9.3 Linearization of Nonlinear System Functions Nonlinear systems are hard to deal with in general terms. A function gðxÞ can be approximated about a point x ¼ x0 using Taylor series expansion dg jx¼x0 ðx x0 Þ gðxÞ ¼ gðx0 Þ þ dx " # 2 d g ðx x0 Þ2 þ þ ::: dx2 2!

dg dgðxÞ ¼ gðxÞ gðx0 Þ dx

ðx x0 Þ ¼ mjx¼x0 dx: x¼x0

(6:9:12)

In this section we will consider the basics of lowpass, high-pass, band-pass, band-elimination filters, and the ideal delay line filters. The filters are specified based on a transfer function HðjoÞ in terms of its amplitude, phase, or delay responses, jHðjoÞj, phase ﬀHðjoÞ or ½dHðjoÞ=do. Finding HðjoÞ from the specifications is the first step. The next step involves the synthesis. We will consider here the ideal filter functions that describe their functions, and simple circuits that can be used as filters. Low-pass filters allow low frequencies to pass through with small attenuation and attenuate or eliminate high frequencies; high-pass filters eliminate or attenuate low frequencies and allow high frequencies go through with possibly small attenuations; band-pass filters allow a band of frequencies to go though with small attenuation and attenuate or eliminate frequencies that are outside this band; band-elimination or bandreject filters let the low and high frequencies pass through and attenuate or eliminate a band of frequencies somewhere in the middle. Delay line filters are primarily used in cascade with filters so that the cascaded filter delay line combination has an approximate linear phase characteristics. Filters are used in every communication system. If the frequencies of the two signals are disjoint, then we can remove the undesired signal by using a band-pass filter that allows the desired signal to go through with a small attenuation and attenuate or eliminate the undesired signal. Tuning to a particular radio station involves eliminating, i.e., filtering out all the other signals from the other stations all available at the front end of the radio or TV receivers. The DC component can be removed by using a highpass filter or a simple bias removal component, a capacitor.

222

6 Systems and Circuits

Example 6.10.1 Illustrate the Bell System TOUCHTONE telephone dialing scheme. Solution: Bell systems TOUCH-TONE telephone dialing scheme uses some of the filters. The discussion follows that of Daryanani (1976). The filters are used in the detection of signals generated by a push button telephone. As we dial a telephone number, i.e., by pushing a button on the telephone, a unique set of two-tone signals are generated and transmitted to the telephone central office, where the signals are processed to identify the number that is transmitted. The buttons and the tone assignments of a TOUCHTONE telephone are shown in Fig. 6.10.1. It has 12 buttons. These correspond to 10 decimal digits, a star button, and a pound button. The letters are also identified on the buttons. For example, on the button identified by 2 has the letters ABC indicating that the number 2 represents A, B, and C as well as 2. Operator button (0) is used for zero. The star (*) button and the pound (#) button are used for other special purposes, such as responding to queries from an answering machine. There are four other buttons that are not shown and are used for special purposes. The signaling code provides 16 distinct signals that use 4 low and high frequencies given by Low : ð697 Hz; 770 Hz; 852 Hz; 941 HzÞ;

amplified first and the two tones are then separated into two groups by the low-pass and the high-pass filters. The low-pass filters pass the low frequencies with very low attenuation and block the high frequencies. Similarly the high-pass filters pass the high frequencies with very little attenuation and block the low frequencies. The separated tones are then converted to square waves of fixed amplitudes by using limiters. Signals are then passed through eight band-pass filters. Each of these passes only one tone and rejects the others. The band-pass filter characteristics are such that there is very little attenuation for the particular frequency and a significant attenuation to block the other frequencies. For a detailed discussion on the amplitude characteristics of low-pass, band-pass, and high-pass filters, see Daryanani (1976). The outputs of the bandpass filters are fed into detectors. The detectors are energized when their input voltage exceeds a set threshold value and the outputs of the detector provides the required dc switching level to connect & the caller to the party being called. Filters can be implemented either in terms of analog or digital domain. Next we will consider each of the filter types in a more formal fashion and discuss the generation of simple transfer functions that allow for the analysis of these filters.

High : ð1209 Hz; 1336 Hz; 1477Hz; 1633 HzÞ Pressing one of the buttons generates a pair of unique frequencies, one lower and the other higher. The fourth high-band frequency, 1633 Hz, is for special services. The block diagram shown in Fig. 6.10.2 illustrates the detection scheme in the telephone office. The received two tones are

6.10.1 Low-Pass, High-Pass, Band-Pass, and Band-Elimination Filters The words low-pass means that when the signal xðtÞ is passed through a low-pass filter, only low

697 Hz →

770 Hz →

Low-band frequencies

852 Hz →

941Hz →

Fig. 6.10.1 Tone assignments for TOUCH-TONE dialing

ABC

DEF

1 2 GHI

JKL

4

5

PRS

TUV

7

3 MNO 6 WXY

8

9

Oper *

# 0

↑

↑

↑

↑

1209 Hz

1336 Hz

1477 Hz

1633 Hz

High-band frequencies

6.10 Ideal Filters

223

Fig. 6.10.2 Block diagram of detection scheme in the telephone office

frequencies, say 0 to fc ¼ oc =2p, are passed and block all frequencies above the cutoff frequency fc : The amplitude of the ideal low-pass filter transfer function is (H0 is assumed to be positive) H0 ; joj oc o ¼ ;oc ¼ 2pfc : jHLP ðjoÞj ¼ H0 P 2oc 0; joj4oc (6:10:1) Every transmission system takes time, i.e., the signal will be delayed. It is ideal to have this delay to be a constant, say t0 s for all frequencies,which may not be possible. In terms of frequency domain, the output transform of such a system is

o jot0 e ; YðjoÞ ¼ HLp ðjoÞXðjoÞ; HLP ðjoÞ ¼ H0 P 2oc HLp ðjoÞ ¼ H0 Pðo=2oc Þ; and ﬀHLp ðjoÞ ¼ ot0 (6:10:2)

Fig. 6.10.3 Amplitude and phase responses of an ideal low-pass filter

The amplitude and the phase response plots are shown with respect to the frequency variable o in Fig. 6.10.3. On the magnitude plot the band of frequencies from 0 to fc as the passband and the band of frequencies from fc to 1 as the stopband are shown. Since the amplitude spectrum of a real signal is even and the phase spectrum is odd, the discussion can be limited to only positive frequencies. The phase response is assumed to be linear, i.e., slope is constant. The group delay is t0 ¼ dﬀHLp ðjoÞ=do:

(6:10:3)

Can we design a real circuit that has the transfer function HLp ðjoÞ? For a physically realizable system, the impulse response hðtÞ ¼ 0 for t50, i.e., the system is causal. For a realizable system, the output cannot exist before the input is applied. The impulse response of the ideal low-pass filter can be

224

6 Systems and Circuits

Z1

determined from the results in Chapter 4 (see (4.3.28).). It is repeated below. sinðaðt0 t0 ÞÞ pðt t0 Þ

FT

!

Yh o i 2a

1

e

jot0

(6:10:4)

Y o jot0 hLp ðtÞ ¼F1 ½HLP ðjoÞ ¼ F1 H0 e 2oc sinðoc ðt t0 ÞÞ : (6:10:5) ¼H0 ð2fc Þ oc ðt t0 Þ In Fig. 6.10.4, the input, an impulse function dðtÞ, applied at t ¼ 0, the block diagram representing the ideal low-pass filter and the impulse response are identified. The impulse response, a sinc function, peaks at t ¼ t0 , giving a value of ð2fc H0 Þ at this time. From the figure we can see that the response is nonzero for t50: That is, there is output before the input is applied. The ideal lowpass filter is not causal and is physically unrealizable. We can also see this from the Payley–Wiener criterion stated below. For a causal system, the impulse response hðtÞ ¼ 0 for t50, as it does not respond before the input is applied. The causality condition can be stated in terms of the transfer function HðjoÞ. It is called the Paley–Wiener criterion Papoulis (1962) and is given in terms of the inequality

(6:10:6)

If jHðjoÞj ¼ 0 over a finite frequency band, then the above integral becomes infinite. jHðjoÞj can be zero at isolated frequencies and still satisfy the criterion. The criterion describes the physical reliability conditions and is not of practical value. That is, if jHðjoÞj ¼ 0 over any band of frequencies, the Payley–Wiener criterion states that the system is physically unrealizable. Ideal low-pass filter violates the condition. We can make a general statement that if the amplitude spectrum is a brick wall type function, the corresponding transfer function is physically unrealizable. Since the ideal low-pass filter function is physically unrealizable, the next best thing is find a function that approximates the ideal filter characteristics. First, consider the simple RC circuit in Fig. 6.3.4a. The transfer function of this circuit is HLp ðjoÞ ¼ 1=ð1 þ joRCÞ: The frequency amplitude characteristic is shown in Fig. 6.6.2a with the cutoff frequency fc ¼ f3dB ¼ ð1=ð2pRCÞÞ. The input and the output transforms are related by YðjoÞ ¼ HLp ðjoÞXðjoÞ. At a particular frequency fi ,

(a)

Fig. 6.10.4 (a) Impulse input, (b) block diagram of a low-pass filter, and (c) impulse response

jlnjHðjoÞjj do51: ð1 þ o2 Þ

(b)

(c)

6.10 Ideal Filters

jYðjoi Þj ¼ HLp ðjoi ÞjXðjoi Þj; oi ¼ 2pfi :

225

(6:10:7)

The output spectral amplitudes at frequencies f ¼ fi are attenuated from the input magnitude spectral amplitudes by the factor of HLp ðjoi Þ. For frequencies between 0 f fc ¼ ð1=2pRCÞ; the output amplitude spectrum is within 3 dB of the input amplitude spectrum, whereas for f4fc , the output amplitude spectrum is significantly reduced or attenuated. The simple RC low-pass filter allows the low frequencies 0 to f3dB go through with a small attenuation and the frequencies from f3dB to 1 are attenuated significantly. The circuit has low-pass filter characteristics. To raise the amplitude characteristics in the passband and, at the same time, lower the amplitude characteristics in the stopband, the following amplitude response function would work: 1 HLp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : 1 þ ðo=oc Þ2n

(6:10:8)

Since o=oc is less than 1 in the passband, by taking the power of this by (2nÞ, we are decreasing the value of the denominator in the passband, thus increasing the amplitude in the passband. On the other hand, in the stopband, i.e., the band above the cutoff frequency o4oc , the denominator in (6.10.8) increases as o increases above the cutoff frequency, and the value of the function reduces in this range. Figure 6.10.5 gives two sketches for n; say n1 and n2 ; n2 4n1 : In the limit, i.e., when n ! 1, the filter characteristics approach the ideal

Fig. 6.10.5 Butterworth amplitude filter response n1 ¼ 2; n2 ¼ 3; e ¼ 1; oc ¼ 5

characteristics. Equation (6.10.8) can be generalized to control the passband attenuation by choosing 1 HLp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ﬃ: 1 þ e2 ðo=oc Þ2n

(6:10:9)

There are two parameters in (6.10.9), e and n. e controls how far the amplitude characteristics will go down to from a value of 1 at o ¼ 0 to when o ¼ oc . The value of n controls how fast the attenuation of the amplitude characteristics in the stop-band region. The amplitude characteristic goes from 1 to ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ p 1= ð1 þ e2 Þ corresponding to the frequencies 0 and fc , respectively. By using the power series, for small e, we can write 1 ½1=

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 þ e2 e2 =2:

(6:10:10)

The function in (6.10.9) is the Butterworth filter function. It has interesting properties. At o ¼ 0, ð2n 1Þ derivatives of the function 1=½1 þ e2 ðo=oc Þ2n are zero, identified as a maximally flat amplitude response. For jo=oc j 1, the high-frequency roll-off of an nth order Butterworth function is 20n dB/decade. The proofs of these are left as exercises. As n ! 1, the filter response has the ideal low-pass characteristics. In the low-pass filter specifications, three bands, namely passband (0 ! oc Þ, transition band ðoc ! or Þ, and stopband ðor ! 1Þ are identified. The transition band is not

226

shown in Fig. 6.10.6, as it depends on required attenuations at the edges of the transition band. Returning to the simple RC low-pass filter, the element values of one of the two components R or C can be determined from the given cutoff frequency, see (6.6.7). Normally the capacitor

6 Systems and Circuits

value is selected, as the number of available capacitor values is much smaller than that of the available resistor values. As a final step, the time response of these filters, for the two simple first-order low-pass RC and RL filters is shown in Figs. 6.3.4a and 6.6.3. The response is determined by the time

(a)

(b)

(c)

Fig. 6.10.6 Amplitude and phase plots of ideal filters: (a) low-pass, (b) high-pass, (c) band-pass, and (d) band elimination

(d)

6.11 Real and Imaginary Parts of the Fourier Transform of a Causal Function

constants, t ¼ RC for the RC circuit and t ¼ L=R for the RL circuit. The output of an LTI system with input xðtÞ and the impulse response hðtÞ is given assuming it zero for t50 by the convolution integral yðtÞ ¼

Z

t

hðaÞxðt aÞda:

(6:10:11)

0

The impulse response hðtÞ is the circuit weighting function. It gives the amount of memory the circuit has. For example, if hðtÞ ¼ dðtÞ, it gives zero weight to the past values of the input function as Z t Z t hðaÞxðt aÞda ¼ dðaÞxðt aÞda ¼ xðtÞ: yðtÞ ¼ 0

0

(6:10:12)

6.11 Real and Imaginary Parts of the Fourier Transform of a Causal Function The real and the imaginary parts of the Fourier transform of a causal function xðtÞ are shown to be related below. Let FT

! XðjoÞ ¼ ReðXðjoÞÞ þ j ImðXðjoÞÞ:

xðtÞ

By noting Re½XðjoÞ is even and Im½XðjoÞ is odd and integral of an odd function over a symmetric interval is zero, we have

Z

t

hðaÞxðt aÞda ¼

0

¼

Z

Z

Z1

1 xðtÞ ¼ 2p

If hðtÞ ¼ uðtÞ, then the circuit has a perfect memory giving equal weights and yðtÞ ¼

227

¼

1 2p

1 Z1

XðjoÞejot do

ðRe½XðjoÞ þ j Im½XðjoÞÞ

1

t

uðaÞxðt aÞda

½cosðotÞ þ j sinðotÞdo

0 t

xðt aÞda:

(6:10:13)

0

Ideal filter frequency responses of the low-pass (Lp), high-pass (Hp), band-pass (Bp), and the band-elimination (Be) filters are given below. The amplitude and the phase response plots of these are given for positive frequencies in Fig. 6.10.6a,b,cd, respectively. Note the phase responses of these ideal filter functions are shown as linear. o jo t0 e (6:10:14a) HLp ðjoÞ ¼ H0 P 2oc o ejo t0 HHp ðjoÞ ¼ H0 1 P (6:10:14b) 2oc h ho o i ho o ii 0 0 HBp ðjoÞ ¼ H0 P þP ejo t0 W W (6:10:14c) h h ho o i ho o iii 0 0 HBe ðjoÞ ¼ H0 1 P þP ejo t0 : W W (6:10:14d) The Lp and Hp filter responses have one passband and one stopband. The Bp filter response has one passband and two stopbands. The Be filter response has two passbands and one stopband.

1 ¼ p

Z1 Re½XðjoÞ cosðotÞdo 0

Z1

1 p

Im½XðjoÞ sinðotÞdo: 0

(6:11:1) Noting that xðtÞ is causal, i.e., xðtÞ ¼ 0; t40, results in xðtÞ ¼

1 p

Z1 Re½XðjoÞ cosðotÞdo 0

1 p

Z1

Im½XðjoÞ sinðotÞdo ¼ 0:

0

Since cosðotÞ and sinðotÞ are defined everywhere, Re½XðjoÞ and Im½XðjoÞ are the real and imaginary parts of the transform of the causal function. That is, 1 p

Z1 Re½XðjoÞ cosðotÞdo 0

1 ¼ p

Z1 0

Im½XðjoÞ sinðotÞdo; t40:

228

6 Systems and Circuits

This implies that a causal signal xðtÞ can be expressed in terms of either Re½XðjoÞ or Im½XðjoÞ and xðtÞ ¼

2 p

Z1

Z1 Im½XðjoÞ sinðotÞdo

(6:11:2)

0

These are true as long as there are no impulses at t ¼ 0 and they imply that Re½XðjoÞ and Im½XðjoÞ cannot be specified independently. Using the transform and solving for real and the imaginary parts of the transform, we have

Re½XðjoÞ ¼

2 p

h0 ðtÞ; t40 h0 ðtÞ; t50

¼ h0 ðtÞsgnðtÞ or

h0 ðtÞ ¼ he ðtÞsgnðtÞ:

0

2 p

he ðtÞ ¼

Re½XðjoÞ cosðotÞdo;

xðtÞ ¼

Noting that hðtÞ ¼ 0; t40, the following interesting relations result:

Z1 Z1 Im½XðjvÞ sinðvtÞ cosðotÞdvdt 0

0

(6:11:6)

Using F½sgnðtÞ ¼ 2=jo and using the Fourier time multiplication theorem, h0 ðtÞsgnðtÞ

FT

! Re½HðjoÞ; he ðtÞsgnðtÞjIm½HðjoÞ: (6:11:7)

The real and the imaginary parts of the function can be related by using the frequency convolution theorem studied in Chapter 4. The frequency convolution theorem corresponding to the two functions is given in (6.11.8) using F½xi ðtÞ ¼ Xi ðjoÞ; i ¼ 1; 2.

(6:11:3a) Im½XðjoÞ ¼

2 p

Re½XðjvÞ cosðvtÞ sinðotÞdvdt: 0

! 2p1 ½X1 ðjoÞ X2 ðjoÞ

FT

x1 ðtÞx2 ðtÞ

Z1 Z1 0

(6:11:3b) That is, for a causal signal, the real and imaginary parts of the transform can be expressed in terms of the other. The results in (6.11.3a and b) are difficult to use. A more elegant way of expressing these relations is by using Hilbert transforms discussed in Chapter 5.

6.11.1 Relationship Between Real and Imaginary Parts of the Fourier Transform of a Causal Function Using Hilbert Transform Consider impulse response of a realizable function and its transform hðtÞ ¼ he ðtÞ þ h0 ðtÞ

FT

! HðjoÞ

¼ Re½HðjoÞ þ j Im½HðjoÞ

(6:11:4)

he ðtÞ ¼ ½hðtÞ þ hðtÞ=2; ho ðtÞ ¼ ½hðtÞ hðtÞ=2: (6:11:5)

1 ¼ 2p

Z1

X1 ðjaÞX2 ðjðo aÞÞda:

(6:11:8)

1

Using these in (6.11.7) the following results: 1 2 j Im½HðjoÞ ; 2p jo 1 2 j Im½HðjoÞ ¼ Re½HðjoÞ : 2p jo Re½HðjoÞ ¼

Z1

1 Im½HðjoÞ ¼ p

1

Re½HðjoÞ ¼

1 p

Z1 1

(6:11:9)

Im½HðjoÞ da; ðo aÞ

ImðHðjoÞÞ da: ðo aÞ

(6:11:10)

Note that the causal signal does not contain an impulse at t ¼ 0 is assumed. If it does, then it adds a constant to its transform. Let hðtÞ ¼ KdðtÞ þ h1 ðtÞ, where h1 ðtÞ does not have an impulse at t ¼ 0: The impulse at t ¼ 0 appears in the transform and K ¼ lim HðjoÞ ¼ Re½Hðj1Þ: o!1

(6:11:11)

6.12 More on Filters: Source and Load Impedances

If there is an impulse at t ¼ 0, the real and the imaginary parts of the transform of the causal signal are related by the following relations in terms of Hilbert transforms: Im½HðjoÞ ¼

1 p

Z1 1

Im½HðjoÞ da; ðo aÞ

1 Re½HðjoÞ ¼Re½Hðj1Þ þ p

Z1 1

ImðHðjoÞÞ da: ðo aÞ (6:11:12)

Notes: From these two equations we note that the real and the imaginary parts of a realizable transfer function HðjoÞ are tied together by the Hilbert transform. This implies that HðjoÞ can be found from its real part alone, referred to as the real-part sufficiency. A physically realizable transfer function can also be found from its magnitude spectrum alone, which is referred to as a minimum phase transfer function and was briefly mentioned in Section 6.7. A linear system with a transfer function HðsÞ has no zeros or poles on the right halfs plane is called a minimum phase system. The relations between amplitude and phase responses of causal functions are referred to as Bode relations. Detailed study of these is beyond the scope here. For a discussion on this topic and other relations, see Bode (1945). Finding an impedance function ZðsÞ from Re½ZðjoÞ has been investigated by many authors, see Weinberg (1962). Here, finding the minimum phase transfer function HðsÞ from the given amplitude spectrum jHðjoÞj is of interest. The following gives a simpler procedure compared to the above & results.

229

Nðo2 Þ or jHðjoÞj2 ¼HðsÞHðsÞs¼jo ¼ K2 Dðo2 Þ Nðo2 Þ HðsÞHðsÞ ¼ K2 j 2 2 : (6:11:13) Dðo2 Þ o ¼s Poles and zeros of the function HðsÞHðsÞ have quadrantal symmetry and mirror symmetry about the jo axis giving a choice in selecting the poles and zeros of HðsÞ. 1. Choose only the left half plane roots of Dðo2 Þjo2 ¼s2 . This gives DðsÞ. 2. To have a minimum phase system, choose only the left half plane roots of NðsÞNðsÞ. Obviously there are other choices and those do not result in minimum phase functions. 3. Select a value K40 to match the value of the amplitude function jHðjoÞj at a desirable frequency. Example 6.11.1 Find the minimum phase stable transfer function HðsÞ given the amplitude-squared spectrum below. jHðjoÞj2 ¼

9ðo2 þ 4Þ : þ 10o2 þ 9Þ

ðo4

(6:11:14)

Solution: Substituting o2 ¼ s2 , and using the above procedure, results in 9ð4 s2 Þ ¼ HðsÞHðsÞ ð9 10s2 þ s4 Þ ðs þ 2Þ ðs þ 2Þ ¼K2 ðs þ 1Þðs þ 3Þ ðs þ 1Þðs þ 3Þ

jHðjoÞj2 jo2 ¼s2 ¼

) HðsÞ ¼ K

ðs þ 2Þ ;K ¼ ðs þ 1Þðs þ 3Þ

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ jHðjoÞj2 jo¼0 ¼ 3: &

6.11.2 Amplitude Spectrum jHðj!Þj to a Minimum Phase Function HðsÞ 6.12 More on Filters: Source and Load Given ðjHðjoÞjÞ find the minimum phase function Impedances 2

HðsÞ ¼ KNðsÞ=DðsÞ that is stable. Starting with ðjHðjoÞjÞ2 , we have 2

jHðjoÞj ¼HðjoÞH ðjoÞ ¼HðjoÞHðjoÞ ¼ K2 Nðo2 Þ=Dðo2 Þ

In this section simple passive analog filters are considered. In the next chapter, the design of various types of filters starting from the specifications to the synthesis using passive and active elements will be

230

considered. In the simple examples considered so far we assumed only one resistor and one inductor (or capacitor) in the filter circuit. The filter problem is illustrated in Fig. 6.12.1a, where we have three boxes, one represents a source, second one represents a filter, and the third one represents the load. Using the Thevenin’s equivalent circuit, we can replace the boxes represented by the source by the source plus the source impedance and the box representing the load by the load impedance. This is shown in Fig. 6.12.1b. The source and load impedances are generally assumed to be resistive in the frequency range of interest. This is a standard assumption in most filter design problems as we are operating in a small range of filter frequencies. In stead of Thevenin’s equivalent circuit we could use the Norton’s equivalent circuit shown in Fig. 6.12.1c. That is,

6 Systems and Circuits

replace the series circuit consisting of source and source resistance in Fig. 6.12.1b by a current source in parallel with the source impedance. This procedure allows the designer to separate the work associated with the filters from any designs associated with the left of the filter, i.e., the source and to the right of the filter, i.e., the load. We might also add that the source box and the load box may include several parts and the filter designer does not have to worry about those parts. In a later chapter when we consider two-port circuit analysis we will come back to this. For now let us consider a simple example illustrating the effect of the load. In the following we will derive the transfer functions in the Laplace transform domain. The frequency responses can be derived by replacing s ¼ jo in the transfer functions.

(a)

(b)

Fig. 6.12.1 (a) Filter with source and load resistors, (b) filter using Thevenin’s source equivalent circuit, and (c) filter using source Norton’s equivalent circuit

(c)

6.12 More on Filters: Source and Load Impedances

231

6.12.1 Simple Low-Pass Filters Example 6.12.1 Consider the RC circuit shown in Fig. 6.12.2 with the source and the load resistors. Derive the transfer function and sketch the amplitude characteristic function for the two cases. a: RL ¼ 1 and b: RL ¼ Rs . Solution: The transfer functions are given by

YðsÞ Z2 ¼HLP ðsÞ¼ ; XðsÞ a Z1 þZ2 RL =Cs RL ¼ ; Z1 ¼Rs (6:12:1a) Z2 ¼ RL þð1=CsÞ 1þRL Cs

YðsÞ RL =Rs RL C ¼ HLP ðsÞ ¼ XðsÞ b s þ ½ðRs þ RL Þ=Rs RL C K Rs þ RL : ; K ¼ 1=Rs C; oc ¼ ¼ Rs RL C s þ oc (6:12:1b) In case a., the load resistance is infinite, i.e., the circuit is not loaded. In case b. Rs ¼ RL . For the two cases the corresponding transfer functions are ð1=Rs CÞ ; ðs þ ð1=Rs CÞÞ ð1=Rs CÞ : b: HLP;Rs ¼RL ðsÞ ¼ s þ ð2=Rs CÞ

pﬃﬃﬃ jYðjoÞjo¼o3dB ¼ ð1= 2ÞjXðjoÞjo¼o3dB :

&

Notes: For simplicity, generic functions x(t) for the input and the output voltage yðtÞ are used. Usually, vi ðtÞ ðor vs ðtÞÞ for the input and v0 ðtÞ for the output & are common.

6.12.2 Simple High-Pass Filters In the ideal low- and high-pass filter cases shown in Fig. 6.10.7a, and b, it can be see that HLp ðoÞ ¼ jHHP ðjoÞjo¼1 ¼ 1 and o¼0 HHp ðjoÞ ¼ HLp ðjoÞo¼1 ¼ 0: o¼0

(6:12:3)

In addition, the amplitudes of these functions transition at the frequency o ¼ oc and the change in the amplitudes are as follows:

a: HLP;RL!1 ðsÞ ¼

(6:12:2)

In both cases, the gain constant is the same. However, the cutoff frequency is increased in the case of a load resistance. Note that the peak value of the amplitude response function in case b. is (1/2). So, the 3 dB frequency corresponds to the valuepof ﬃﬃﬃ the magnitude of the function equal to ð1=2Þð1= 2Þ. At o ¼ 0 the filter circuit is transparent; at o ¼ 1

Fig. 6.12.2 Example 6.12.1

there is no signal transmission; in between these frequencies, the output signal amplitude attenuation is determined by the equation jYðjoÞj ¼ jHðjoÞjjXðjoÞj. At the 3 dB frequency, o3dB

1ðlow-passÞ ! 0 ðhigh-passÞ or 0 ðhigh-passÞ ! 1 ðlow-passÞ: A logical conclusion is that o ! ð1=oÞ ði:e:; s ! ð1=sÞÞ provides a transformation that gives a way to find a high-pass filter function from a low-pass filter function. Noting that the impedance of an inductor is ðjoLÞ and the impedance of a capacitor is ð1=joCÞ, a high-pass filter can be obtained from a

232

6 Systems and Circuits

low-pass filter by replacing a capacitor by an inductor and an inductor by a capacitor. Since the impedance of the resistor R is independent of frequency, no change is necessary in the case of resistors. Using the RC and the RL low-pass circuits studied earlier in the low-pass case, we have two simple high-pass filters shown in Fig. 6.12.3a,b, one is a RL and the other one is a RC circuit. The filter is a low-pass if the inductor is in the series arm and the capacitor in the shunt arm. Similarly the filter acts as a highpass filter if the capacitor is in the series arm and the inductor is in the shunt arm. The transfer functions corresponding to the two circuits in Fig. 6.12.3 are

joj HHp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; b o2 þ ðR=LÞ2 ﬀHHpb ðoÞ ¼ 900 tan1 ðoL=RÞ:

(6:12:5b)

The amplitude and phase responses are shown in Fig. 6.12.4 for o40: The maximum value of the amplitude response is 1 or 0 dB. The 3 dB frequencies can be computed bypequating the amplitude ﬃﬃﬃ response function to ð1= 2Þ and solving for o. That is, 1 1 joca j jocb j pﬃﬃﬃ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; pﬃﬃﬃ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : 2 o2 þ ðR=LÞ2 o2 þ ð1=RCÞ2 2 ca

s s HHpa ðsÞ ¼ ; HHpb ðsÞ ¼ s þ ð1=RCÞ s þ ðR=LÞ jo jo ;HHPb ðjoÞ¼ : HHPa ðjoÞ¼ joþð1=RCÞ joþðR=LÞ

cb

(6:12:6) ) oca ¼ ð1=RCÞ; ocb ¼ ðR=LÞ:

(6:12:7)

(6:12:4)

The amplitude and the phase response characteristics of these are as follows: jo j HHp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; a 2 2 o þ ð1=RCÞ ﬀHHpa ðoÞ ¼ 900 tan1 ðoRCÞ

(6:12:5a)

Fig. 6.12.3 Simple high passive filters (a) RC circuit, (b) RL circuit

Fig. 6.12.4 Simple high-pass filter responses: (a) amplitude and (b) phase

As in the low-pass case, given the cutoff frequency oc , one of the reactive component (inductor or capacitor) values can be solved by selecting the resistor value. The high-pass filter is transparent from the input to the output at infinite frequency and no signal transmission at zero frequency. Note the low and high-frequency behavior of the lowpass and the high-pass filter functions HLp ðsÞ (or HHp ðsÞ) at s ¼ 0 and s ¼ 1:

(a)

(a)

(b)

(b)

6.12 More on Filters: Source and Load Impedances

233

6.12.3 Simple Band-Pass Filters These filters pass a band of frequencies called the passband and attenuate or eliminate the frequencies outside the passband, called the stopband. The simplest band-pass filter has a second-order transfer function. The ideal band-pass filters have two cutoff (or 3 dB) frequencies olow and ohigh . These frequencies are defined as the frequencies for which the magnitude function is equal to pﬃﬃﬃ of the transfer maxð1= 2ÞHBp ðjoÞ. In addition to these, a new frequency referred as the center or the resonant frequency o0 ; is of interest. It is defined as the frequency at which the transfer function of the circuit HBp ðjoÞ is purely real. The center frequency is not in the middle of the passband. It is the geometric center of the pass-band edges. It is related to the 3 dB frequencies by pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ (6:12:8) o0 ¼ olow ohigh : The second parameter of interest is the 3 dB bandwidth given by b ¼ ohigh olow :

(6:12:9)

The third parameter, the quality factor is the ratio of the center frequency to the 3 dB bandwidth. It is given by Q¼

o0 : ðohigh olow Þ

(6:12:10)

A second-order function that has the band-pass characteristics is HBp ðsÞ ¼

s2

H0 as H0 ðo0 =QÞs ¼ 2 : þ as þ b s þ ðo0 =QÞs þ o20 (6:12:11)

Fig. 6.12.5 (a) Amplitude and (b) phase responses of a band-pass filter

(a)

For simplicity the gain constant is assumed to be H0 ¼ 1 in the following. The transfer function has a zero at the origin ðs ¼ 0Þ and at infinity ðs ¼ 1Þ indicating that the function goes to zero at o ¼ 0 and at o ¼ 1. For Q41=2, it has a pair of complex poles given by sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ o0 1 : (6:12:12) jo0 1 s1 ; s2 ¼ 2Q 4Q2 The corresponding transfer function, the amplitude, and phase responses are given by ðo0 =QÞo ðo20 o2 Þ þ jðo0 =QÞo 1

ih

i ¼h o0 o 1 þ jQ o0 o 1 þ jQ oo0 oo0

HBp ðjoÞ ¼

(6:12:13) sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

2ﬃ o o 0 ; 1 þ Q2 o0 o

o o0 1 : ﬀHBp ðjoÞ ¼ tan Q o0 o

HBp ðjoÞ ¼1=

(6:12:14)

The amplitude and the phase responses are sketched in Fig. 6.12.5 for positive values of o. From the amplitude response, the peak of the amplitude appears at the center frequency o0 . The peak magnitude is 1 at o ¼ o0 . The phase angle starts at ðp=2Þ, crosses the frequency axis at o ¼ o0 and it asymptotically reaches ðp=2Þ as o ! 1: Higher the value of Q is, the more peaked the amplitude response is and steeper the phase response is around o ¼ o0 . The 3 dB or half-power bandwidth can be determined by assuming olow 5 o0 and ohigh 4o0 . These frequencies can be computed from

(b)

234

6 Systems and Circuits

2 HBp ðjoÞ2 o¼o ;o ¼ 1!Q2 o o0 ¼1) l u 2 o0 o o0 o 2 1 ;o 2 oo0 o20 ¼0: ðo2 o20 Þ¼ Q Q There are four roots of this equation, two for positive and two for negative frequencies. sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2 o0 1 o0 þ 4o20 ; ou ; ol ¼ 2Q 2 Q s ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ﬃ

o0 1 o0 2 ou ; ol ¼ þ 4o20 : 2Q 2 Q

Fig. 6.12.6 A simple band-pass filter

the expressions for the center frequency, bandwidth, and the quality factor. Solution: The transfer functions are V0 ðsÞ ðR=LÞs ¼ 2 ; Vi ðsÞ s þ ðR=LÞs þ ð1=LCÞ ðR=LÞjo : (6:12:18) HBp ðjoÞ ¼ ½ð1=LCÞ o2 Þ þ joðR=LÞ

HBp ðsÞ ¼ Assuming that Q4ð1=2Þ, the positive roots are given by sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 o0 (6:12:15) ol ; ou ¼ o0 1 þ : 4Q2 2Q The 3 dB bandwidth ðB or bÞ and o0 are respectively given by b ¼ ðohigh olow Þ ¼ o0 =Q; pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ o20 ¼ ol ou ! o0 ¼ ol ou :

(6:12:16)

Clearly from the first equation in (6.12.16) the bandwidth is inversely proportional to the value of Q. That is, the bandwidth decreases as Q increases and vice versa. The filter is assumed to be narrowband if o0 is very large compared to the bandwidth of the filter, i.e., o0 B. As a rough measure, we assume the filter is a narrowband filter if Q ¼ o0 =b 10:

(6:12:17a)

For the narrowband case the edges of the passband and o0 are given below. ol ; ou ¼ o0 ðo0 =2QÞ:

(6:12:17b)

In this case, o0 is in the middle of the 3 dB frequencies and is the center frequency. Example 6.12.2 Consider the circuit shown in Fig. 6.12.6. Find the transfer function and derive

The corresponding amplitude and phase responses are given for positive o by ðR=LÞo HBp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ðð1=LCÞ o2 Þ2 þ ððR=LÞoÞ2 ﬀHBp ðjoÞ ¼90o tan1

ððR=LÞo : (6:12:19) ½ð1=LCÞ o2 Þ

From (6.12.19), by using (6.12.16) results in pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ o0 ¼ 1=LC; b ¼ Bandwidth ¼ R=L and pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ (6:12:20) Q ¼ o0 =B ¼ L=CR2 : There two equations had three unknowns. In the design of a second-order band-pass filter, use the following steps. From the given second-order function, find o0 and Q. Select one of the element values, say C, and then use the other two equations to solve for the element values R and L: These are L ¼ 1=o20 C; R ¼ bL. Note that the gain constant is assumed to be H0 ¼ 1. If the gain is higher than 1, then circuit needs amplification. So far we have not considered of having both source and load impedances in the band-pass case. Also, inductors are never ideal and can be modeled by a resistor in series with an ideal inductor shown in Fig. 6.12.7 resulting in the impedance of the nonideal inductor as R i þ sL. The new transfer function and the amplitude response are

6.12 More on Filters: Source and Load Impedances

235

Fig. 6.12.7 A simple band-pass filter with a nonideal inductor

HBp ðsÞ ¼

ðR=LÞs s2 þ ½ðR þ Ri Þ=Ls þ ð1=LCÞ

(6:12:21)

ðR=LÞjoj HBp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : ðð1=LCÞ o2 Þ2 þ oððR þ Ri ÞLÞ (6:12:22) The center frequency is the same as before. The maximum value of the amplitude response is now R=ðRi þ RÞ. Also the new bandwidth is ðR þ Ri Þ=L. The nonideal inductor reduces the peak value and increases the bandwidth of the filter response. Correspondingly, the Q value is reduced. Nonideal inductors make the amplitude response less peaked with an increase in the bandwidth, i.e., the amplitude response becomes lower and broader. A band-pass filter can be viewed as a cascaded lowpass and a high-filter combination with the cutoff frequency of the low-pass filter greater than the cutoff frequency of the high-pass filter, oc;LP 4 oc;HP . An obvious question is, can we obtain a band-pass filter function from a low-pass filter function? In the two simple low-pass circuits considered earlier, the RL circuit, with the replacement of the inductor by a series LC circuit results in the circuit studied above. By replacing the capacitor in the RC low-pass circuit by a parallel LC circuit results in a band-pass circuit shown in Fig. 6.12.7. The element values in the bandpass case need to be determined using the center

Fig. 6.12.8 (a) Low-pass and (b) band-pass filter circuits

(a)

frequency and the required bandwidths. The figure illustrates only the concepts here. In Chapter 7 we will come back to the frequency transformations that involve changing cutoff frequencies of filters, converting various filter specifications into a proto-type lowpass filter specification, finding the appropriate filter function, the appropriate frequency transformations, synthesizing the filter function and finally scaling the circuit to fit the given specifications. Figure 6.12.8a gives a low-pass circuit. A band-pass circuit can be obtained by replacing the capacitor by an LCcircuit as shown in Fig. 6.12.8b. Inductors tend to be bulky and nonideal. That is, they have a resistive part Rw , thus reducing the quality factor of such coils. Impedance of the inductor needs to be replaced from joL to Rw þ joL. The quality factor of the coil is given by Q ¼ ðoL=Rw Þ, a function of frequency. Designing inductors with high Q values is a difficult process. In addition, since the field of operation associated with coils is the magnetic field, there is coupling between different inductors in a circuit. This can be reduced by either shielding one inductor from another and/or by placing in a manner shown in Fig. 6.12.9 requiring more space on the circuit board.

6.12.4 Simple Band-Elimination or Band-Reject or Notch Filters The amplitude response of a band-reject filter has the shape of a notch and is used to remove a band of frequencies somewhere in the middle of the frequency band and pass the low and high frequencies outside this band. A second-order notch filter has a transfer function of the form HBe ðsÞ ¼

ðs2 þ 2bs þ o20 Þ ; s2 þ ðo0 =QÞs þ o20

b55

o0 2Q (6:12:23)

(b)

236

6 Systems and Circuits

The 3 dB frequencies are obtained by equating pﬃﬃﬃ jHðo3dB Þj ¼ 1= 2. The two frequencies are

Fig. 6.12.9 Placement of two inductors to reduce magnetic coupling

ol ; ou ¼ o0 ðo0 =2QÞ; b ¼ o0 =Q:

(6:12:27)

If b 6¼ 0, then Lim HBe ðsÞ ¼ 1 and Lim HBe ðsÞ ¼ 1: s!0

s!1

(6:12:24)

The second-order band-pass and the notch filter functions have the same denominator, see (6.12.11) and (6.11.23). Notch filter passes low and high frequencies of the input signal without much attenuation and attenuates (or eliminates) a band of frequencies in the middle. This can be seen by first computing the zeros of the transfer function. qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ z1 ; z2 ¼ b j o20 b2 ) z1 ; z2 b jo0 ðIn the case of o0 44bÞ: (6:12:25) A special and an interesting case is when b ¼ 0 and for this case 2 o o2 0 jHBe ðjoÞj ﬃ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ðo20 o2 Þ2 þ ½ðo0 =QÞo2 1 ðo0 =QÞo : (6:11:26) ﬀHBe ðjoÞ ¼ tan ðo20 o2 Þ Note there is no output at the notch frequency o0 as jHBe ðjo0 Þj ¼ 0. The amplitude and phase responses of the notch filter are shown in Fig. 6.12.10 for o40. We can see that jHBe ðjoÞjo¼0 ¼ 1 and lim jHBe ðjoÞj ¼ 1: o!1

jHBe ðjoÞjo¼o0 ¼ Qb=o0 :

The attenuation will be significant at the notch frequency as o0 is usually large. Also, note the phase reversal at o ¼ o0 in the phase response. Notch filters are used wherever a narrowband of frequencies needs to be eliminated from a received signal. In any electronic device, 60 Hz undesired hum, is ever present and a notch filter can be used to remove this. There are many applications in the telephone industry. In a long-distance call, a single frequency is transmitted from the caller to the telephone office until the end of the dialing of the number. After the party answers, the tone signal ceases and billing of the call begins and it continues as long as the signal tone is absent until the call is complete. A different application is toll-free long-distance calls that are not billed. For these, the signal tone is transmitted to the telephone office for the entire period of the call. Since the signal is within the voice frequency band, it must be removed from the voice signal before being transmitted from the telephone office to the listener. A simple second-order notch filter could be used for such an application. Notch filters are used wherever a narrowband of frequencies need to be eliminated from a received signal. In any electronic device, 60 Hz, an undesired hum, is ever present and a notch filter can be used to remove this. There are many applications in the telephone industry.

Fig. 6.12.10 (a) |HBe ðjoÞ| and (b) ﬀHBe ðjoÞ

(a)

(6:12:28)

(b)

6.12 More on Filters: Source and Load Impedances

Example 6.12.3 Band-elimination filters can be derived from low-pass filters by replacing an inductor (capacitor) by a parallel LC ðseries LCÞ circuit. Consider the circuit in Fig. 6.12.11 with a nonideal inductor with the equivalent impedance of the coil equal to ðRw þ joLsÞ. Assuming the circuit is not loaded, the output current is zero. Derive the transfer function and show it corresponds to a band-elimination filter.

Fig. 6.12.11 A band-elimination filter with a nonideal inductor

Solution: The transfer function, its amplitude and phase responses are Rw þ Ls þ ð1=CsÞ ðR þ Rw Þ þ Ls þ ð1=CsÞ ð1 þ Rw s þ LCs2 Þ (6:12:29) ¼ ð1 þ LCs2 Þ þ ðR þ Rw ÞCs ðRw þ joL þ ð1=ðjoCÞÞ HBe ðjoÞ ¼ ½ðR þ Rw Þ þ joL þ ð1=ðjoCÞÞ ð1 o2 LCÞ þ joRw C (6:12:30) ¼ ð1 o2 LCÞ þ joðR þ Rw ÞC qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ð1 LCo2 Þ2 þ ðRw CoÞ2 jHBe ðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ð1 o2 LCÞ2 þ ðoðR þ Rw ÞCÞ2 HBe ðsÞ ¼

ﬀHBe ðjoÞ ¼ arctan½Rw Co=ð1 o2 LCÞÞ

(6:12:31a)

arctan½oðR þ Rw ÞC=ð1 o LCÞ: (6:12:31b) The transfer function has the band-elimination characteristics as lim jHBe ðjoÞj ¼ 1; lim jHBe ðjoÞj ¼ 1; and o!1

jHBe ðjoÞjo¼1=pﬃﬃﬃﬃﬃ LC ¼ Rw =ðR þ Rw Þ:

If the winding resistance Rw is small, then the peak value of the amplitude response is close to 1. Using (6.12.27) pﬃﬃﬃﬃﬃﬃﬃ and (6.12.28), we have 2b ¼ Rw =LC, o0 ¼ 1= LC and ðo0 =QÞ ¼ ðR þ Rw ÞC. The notch bandwidth is B ¼ ½o0 =Q ¼ ½ðR þ Rw ÞC. In the ideal case, a second-order notch filter can be obtained from a second-order band-pass filter by replacing the parallel (series) LC circuit by a series (parallel) LC circuit.& Passive filter designs use resistors, inductors (and transformers), and capacitors. Four ladder forms of low-pass, high-pass, band-pass, and band-elimination filter circuits with source and load resistances are shown in Fig. 6.12.12. Source with source resistance is identified by a circle and the load resistance by a square. Between the source and the load, a lossless circuit is inserted, i.e., lossless coupling between the source and load. Figure 6.12.12a: Low-pass filter: When o ¼ 0, inductors will be short and the capacitors will be open and the source is directly connected to the source and the output is v0 ¼ ½RL =ðRi þ RL Þvi . When o ¼ 1, inductors will be open and the capacitors will be short and the load is disconnected from the source and the output is zero. Figure 6.12.12b: High-pass filter: When o ¼ 0, inductors will be open and the capacitors will be short and the load is disconnected from the source and the output is zero. For o ¼ 1, inductors will be short and the capacitors will be open and the source is directly connected to the load and v0 ¼ ½RL =ðRi þ RL Þvi . Figure 6.12.12c: Band-pass filter: At o ¼ 0, inductors will be short and the capacitors will be open and there is no output. At o ¼ 1, there is no output either. At the center frequency o0 , if (Lsi Csi ¼ 1=o20 ), the series arm is short since Lsei s þ ð1=Csei sÞs2 ¼ð1=Lsei Csei Þ ¼ ðLsei Csi s2 þ 1Þ=Csei ss2 ¼ð1=L C Þ ¼ 0: sei

2

o!0

237

(6:12:31c)

sei

In a similar manner we can show that at the frequency o0 , the shunt arm is open since Lshi s= Lshi Cshi s2 þ 1 s2 ¼ð1=L C Þ ¼ 1: shi

shi

Figure 6.12.12d: In the band-elimination case, we can show that v0 ¼ ½RL =ðRi þ RL Þvi at o ¼ 0 and o ¼ 1. At o ¼ o0 , the output is zero. The four filters have the desired transfer characteristics values at o ¼ 0; 1 (and, in addition o ¼ o0 ; in

238

6 Systems and Circuits

(a)

(b)

(c)

(d) Fig. 6.12.12 Passive filters

the case of band-pass and band-elimination filters. The exact characteristics at other frequencies cannot be determined since the actual filter element values are not known. In the passive filter design, and in general in communication theory, power transfer between the source and the load is important.

6.12.5 Maximum Power Transfer If there is no filter in Fig. 6.12.12, i.e., the source resistance is connected directly to the load, the maximum power is available at the output provided Rs ¼ RL :

(6:12:32a)

6.13 Summary

239

This is the maximum power transfer theorem and can be proven by first expressing the power delivered to the load without the filter. That is, pL ¼ RL i2 ; i ¼

vs RL ¼ v2 : ðRs þ RL Þ ðRs þ RL Þ2 s

The transfer function, the amplitude, phase, and the group delay responses are

(6:12:32b) Taking the derivative of pL with respect to RL , then solving for RL results in (6.12.32a). If the filter is inserted between the source and the load, then there will be less power available at the load, which will vary with frequency. This is defined as insertion loss. The design using these concepts is beyond the scope here, see Weinberg (1962).

6.12.6 A Simple Delay Line Circuit In Section 6.4 we considered the properties of a delay function. In this part of the section we will consider a simple circuit that has a constant amplitude response and a phase response that can be adjusted. Consider the circuit shown in Fig. 6.12.13. Assuming the output current is equal to zero, the output voltage can be expressed by V0 ðsÞ ¼Vc ðsÞ Vx ðsÞ ¼ ¼

ð1=CsÞ 1 Vin ðsÞ Vin ðsÞ ðR þ ð1=CsÞÞ 2

1 RCS Vin ðsÞ: 2ðRCs þ 1Þ

jHðjoÞj ¼ 1=2; ﬀHðjoÞ ¼ 2 tan1 ðo=aÞ; dﬀHðjoÞ a ¼ 2 : (6:12:34b) Tg ðoÞ ¼ do a þ o2

(6:12:33)

1 ða sÞ 1 ða joÞ 1 ; HðjoÞ ¼ ;a ¼ 2 ða þ sÞ 2 ða þ joÞ RC jHðjoÞj ¼ 1=2; ﬀHðjoÞ ¼ 2 tan1 ðo=aÞ; (6:12:35a) HðsÞ ¼

Tg ðoÞ ¼

dﬀHðjoÞ a ¼ 2 : do a þ o2

(6:12:35b)

Since the amplitude response is (1/2), a constant for all frequencies, this function is referred to as an allpass function. All-pass filters are used in cascade with the filters to provide the overall phase of the filter delay line combination and have an approximate linear phase characteristics. Additional phase due to the all-pass circuit adds to the filter delay.

6.13 Summary In this chapter we have started with basics of systems analysis and circuits. The circuits considered are simple. Specific topics that were covered in this chapter are given below.

Linear systems and their properties Two-terminal components: resistors, inductors, capacitors, voltage, and current sources

Classification of systems based on linearity, time-invariance, and other concepts

Impulse response of a linear system and the output in terms of the convolution integral

Transfer functions along with examples of simple circuits

System stability concepts and Routh’s stability test Distortionless systems and distortion measures

Fig. 6.12.13 A simple delay line circuit

for nonlinear systems

The transfer function in the s domain and in the frequency domain, the corresponding magnitude and phase responses and the group delay function are respectively given by HðsÞ ¼

1 ða sÞ 1 ða joÞ 1 ; HðjoÞ ¼ ;a ¼ 2 ða þ sÞ 2 ða þ joÞ RC (6:12:34a)

Group delay and phase delay responses System bandwidth measures similar to signal bandwidth

Relations between real and imaginary parts of a Fourier transform of a causal function

Derivation of the minimum phase transfer function from a given magnitude function

Ideal low-pass, high-pass, band-pass, bandelimination filters along with delay lines

240

6 Systems and Circuits

replacing the capacitor by an inductor in Fig. 6.5.4b. Give the corresponding steady-state response.

Problems 6.2.1 Consider the systems described by the following input–output relations. In each case, determine whether the system satisfies the following: 1. memoryless, 2. causal, 3. stable, 4. linear, and 5. time invariant. a: yðtÞ ¼ xð1 tÞ; b: yðtÞ ¼ xðt=2Þ; c: yðtÞ ¼ sinðxðtÞÞ: 6.3.1 Show that the systems a: yðtÞ ¼ x2 ðtÞ; b: yðtÞ ¼ sgnðtÞ are not invertible. 6.3.2 Determine whether the system yðtÞ ¼ T½xðtÞ ¼ txðtÞ is 1. memoryless, 2. causal, 3. stable, 4. linear, and 5. time invariant. 6.3.3 Consider the amplitude modulated function (discussed in Chapter 10) yðtÞ ¼ Ac ½A þ mðtÞ cosðoc tÞ with cases a: A ¼ 0; b: A 6¼ 0. Determine whether the system described by this equation is 1. memoryless, 2. causal, 3. linear, 4. time invariant, and 5. BIBO stable. 6.3.4 Repeat Problem 6.3.3 assuming the output of the system is yðtÞ ¼ xðt tÞ. 6.3.5 Determine whether the system described by the following is a. linear or nonlinear, b. time invariant or time varying yðtÞ ¼

1 X

xðtÞdðt kts Þ ¼

k¼1

1 X

xðkts Þdðt kts Þ:

k¼1

6.3.6 Consider the system described by yðtÞ ¼ xðbtÞ. Determine for what values of b the system is a. causal, b. linear, and c. time invariant.

6.4.3 Determine the impulse responses of the ideal low-pass, high-pass, band-pass, and bandelimination filters defined in (6.10.14a, b, c, and d) and the ideal delay line function in (6.7.1). 6.4.4 Determine the system responses for the inputs a: x1 ðtÞ ¼ uðtÞ; b: x2 ðtÞ ¼ P½t :5 assuming the system impulse response is hðtÞ ¼ tet uðtÞ: 6.4.5 Determine the stability of the integrator and a differentiator given below. Find their impulse responses and then use BIBO stability condition to see their stability. yðtÞ ¼

Zt xðaÞda;

yðtÞ ¼

dxðtÞ : dt

1

6.4.6 Using very simple functions with the properties as identified to show the responses become unbounded. Consider the transfer functions HðsÞ. a: HðsÞ has a pole on the right half s-pane, b.HðsÞ has multiple poles on the imaginary axis, c. HðsÞ has poles on the imaginary axis and the input function has a pole at this location. Explain why the responses become unbounded with the aid of the inverse Laplace transforms. 6.4.7 Consider the transfer function given by TðsÞ ¼

1 K ; HðsÞ ¼ : 1 þ HðsÞ ð1 þ :1sÞða þ sÞ

6.4.1 In Fig. 6.4.1 we have considered a simple RL time-varying circuit. Consider the RC time-varying circuit shown in Fig. P6.4.1. Assume that the time constant is large enough to justify that the circuit can be used as an integrator in an approximate sense. Identify the approximation used in considering this circuit acts like an integrator.

a. Assuming a ¼ 1, use the Routh array to determine the range of K for which the system is stable. b. Repeat the problem in Part a. assuming K ¼ 1 and determine the range of K for which the system is stable. c. Using the Routh array to determine the range of K for which the system has only poles to the left of s ¼ 1 in the s-plane.

6.4.2 Apply the periodic pulse waveform shown in Fig. 6.5.4a to a simple RL circuit obtained by

6.4.8 Use the Routh array and factor the polynomial DðsÞ ¼ s4 þ s3 þ 2s2 þ s þ 1. 6.4.9 Use the Routh array to find the number of right half s-plane roots of the polynomial DðsÞ ¼ s4 þ s2 þ s þ 1.

Fig. P6.4.1 An RC time-varying circuit

6.5.1 Give an RC circuit that approximates a differentiator and sketch the circuit’s amplitude and phase responses.

Problems

241

6.5.2 The transfer function of a linear time-invariant system is HðsÞ ¼

YðsÞ ðs þ 1Þ : ¼ 2 XðsÞ ðs þ s þ 1Þ

Assuming xðtÞ ¼ cosðo0 tÞ, find the steady-state response yðtÞ of the system. 6.5.3 Find the response of the RL circuit shown in Fig. 6.6.3 for input pulses t ðT=2Þ ; T ¼ 2p; a: xa ðtÞ ¼ P T b: xb ðtÞ ¼ sinðtÞxa ðtÞ: 6.5.4 Consider a circuit that has the response yðtÞ ¼ xðtÞ xðt TÞ with the input xðtÞ: Give the system impulse response and the expressions HðjoÞ and jHðjoÞj. 6.5.5 Consider the following two differential equations with xðtÞ ¼ ejot : d 2 y L dy þ þ yðtÞ ¼ xðtÞ dt R dt 2 d y b: LC 2 þ yðtÞ ¼ xðtÞ: dt

a: LC

Derive the transfer functions HðjoÞ ¼ YðjoÞ=XðjoÞ in each case assuming xðtÞ ¼ ejot . 6.5.6 Determine in each case below if the system described by its impulse response is stable or realizable, or both. Explain your results. a:ha ðtÞ ¼ dðtÞ;

b:hb ðtÞ ¼ et uðtÞ;

c:hc ðtÞ ¼ dðtÞ et uðtÞ: 6.5.7 Consider the impulse response and the corresponding transfer function hLp ðtÞ ¼ et uðtÞ; L½hLp ðtÞ ¼ HLp ðsÞ ¼ 1=ðs þ 1Þ. What can you say about the deconvolution filter HR ðsÞ and hR ðtÞ ¼ L1 fHR ðsÞg? Is this function realizable? 6.6.1 Consider the following functions with DðsÞ ¼ ðs þ 1Þðs þ 4Þðs þ 5Þ. Classify these as minimum or mixed or maxi phase systems. Sketch their amplitude and phase responses. a: H1 ðsÞ ¼ ðs 3Þðs þ 2Þ=DðsÞ b: H2 ðsÞ ¼ ðs þ 3Þðs þ 2Þ=DðsÞ c: H3 ðsÞ ¼ ðs 3Þðs 2Þ=DðsÞ:

6.6.2 Sketch the amplitude and phase responses of pﬃﬃﬃﬃﬃ a: H1 ðsÞ ¼ 1=ðs2 þ 2s þ 1Þ; b: H2 ðsÞ ¼ 1=ðs2 þ 3s þ 3Þ: 6.6.3 What can you say about the group delays associated with an all-pass functions at o ¼ 0 and o ¼ 1 in Problem 6.6.2? Can you draw any general conclusions? 6.6.4 The second-order Butterworth function is given in Problem 6.6.2a. a. Find its impulse response and the corresponding step response. b. Find the 10–90% rise time. c. Find the expression for the group delay of this function. 6.6.5 Assume the following node equations of a 3 nodes plus a reference node is given by ½ðVA V1 ÞY1 þ VA Y3 þ ðVA V0 ÞY2 ¼ 0; ðV0 VA ÞY2 þ ðV0 V1 ÞY4 ¼ 0: Give a circuit that has these node equations. Derive the transfer function V0 =V1 assuming Y1 ¼ 1=sL1 ; Y2 ¼ 1=sL2; Y3 ¼ sC2 ; Y3 ¼ sC3 : 6.6.6 Find the impulse and step responses of the following transfer functions with a40. 1 as ; a40; 1 þ as a ; b: HLP ðsÞ ¼ 2 s þ as þ a s c: HBP ðsÞ ¼ 2 s þsþ1 s2 þ a2 ; d: HBe ðsÞ ¼ 2 s þ bs þ a2 s2 : e: HHP ðsÞ ¼ 2 s þsþ1 a: Hd ðsÞ ¼

Give the amplitude and phase responses of these filters. 6.7.1 Consider the impulse response of a system hðtÞ ¼ et uðtÞ. Derive the expressions for its group and the phase delays. Sketch these functions on the same plot. 6.8.1 Show that the noise bandwidth of a band-pass function with center frequency o0 is WN ¼

Z1 h 0

i jHðjoÞj2 =jHðjo0 Þj2 do:

242

6 Systems and Circuits

6.8.2 Determine the RMS and the equivalent bandwidth of HðjoÞH ðjoÞ ¼ L½o=W. 6.9.1 Sketch the input and the output line spectra of a nonlinear circuit described by its input–output assuming xðtÞ ¼ relationship yðtÞ ¼ x2 ðtÞ 2 cosð2pð60ÞtÞ þ sinð2pð60ÞtÞ. 6.9.2 Assuming yðtÞ ¼ xðtÞ þ x2 ðtÞ, sketch YðjoÞ by assuming XðjoÞ ¼ P½ðo þ o0 Þ=W þ P½ðo o0 Þ=W: 6.9.3 Assume the input is xðtÞ ¼ cosðo0 tÞ to a nonlinear system described by yðtÞ ¼ xðtÞ þ :1x2 ðtÞ determine the second-order distortion term. 6.9.4 Consider the systems described by the following input–output relations. In each case determine whether the system is a. memoryless, b. causal, c. stable, d. linear or non-linear, and e. time-invariant system: a: yðtÞ ¼ xð1 tÞ; b: yðtÞ ¼ xðt=2Þ; c:yðtÞ ¼ sinðxðtÞÞ. 6.10.1a. Show the high-frequency slope of the n th order low-pass Butterworth function jHðjoÞj ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1=ð1 þ e2 o2n Þ is 20 n dB=decade:

b. Show that the first ð2n 1Þ derivatives of jHðjoÞj2 are zero at o ¼ 0. Use long division and then compare that to a power series and identify the

Fig. P6.12.1 Circuits to determine the transfer functions

corresponding derivatives. c. Show (6.10.10). d. Assuming n ¼ 2, determine the corresponding second-order Butterworth transfer function HðsÞ and find its impulse response. 6.10.2 Sketch normalized function qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ jHðjo=oc Þj ¼ 1=ð1 þ e2 ðo=oc Þ2n Þ for n ¼ 3 and 4: 6.11.1 Determine the minimum phase transfer function corresponding to the functions a: jHðjoÞj2 ¼ ½1=ð1 þ o2 Þ; b: jHðjoÞj2 ¼ ½1=ð1 þ o4 Þ: 6.12.1 Consider the circuits shown in Fig. P6.12.1. For each of these cases determine the corresponding transfer function. In the case of band-pass or bandelimination filters, determine the center frequencies. Give the 3 dB cutoff frequencies in each case. Give the expression for the quality factor of the circuits wherever appropriate. In addition, identify the type of the filter in each case. Simplify the expressions by assuming Rw ¼ 0. Sketch the amplitude responses for each of the cases and identify the important values. 6.12.2 Prove the maximum power transfer theorem. Sketch the power delivered to the resistor RL . Assume the source resistance is Rs .

Chapter 7

Approximations and Filter Circuits

7.1 Introduction In the first part of this chapter we will consider a graphical representation of the transfer function in terms of its frequency response HðjoÞ ¼ jHðjoÞjeﬀHðjoÞ . Bode diagrams or plots consist of two separate plots, the amplitude jHðjoÞj and the phase angle ﬀHðjoÞ, with respect to the frequency o on a logarithmic scale. These plots are named after Bode, in recognition of his pioneering work Bode (1945). Bode’s basic work was based upon approximate representation of amplitude and phase response plots of a communication system. Wide range of frequencies of interest in a communication system dictated the use of the logarithmic frequency scale. Bode plots use the asymptotic behavior of the amplitude and the phase responses of simple functions by straight-line segments and are then approximated by smooth plots with ease and accuracy. Bode plots can be created by using computer software, such as MATLAB. The topic is mature and can be found in most circuits, systems, and control books. For example, see Melsa and Schultz (1969), Lathi (1998), Close (1966), Nilsson and Riedel (1966), and many others. Filter approximations will be considered in the second part of this chapter. In Section 6.10, Butterworth approximation of an ideal low-pass filter amplitude response was introduced and the amplitude squared Butterworth function is jHBu ðjoÞj2 ¼

1 1 þ e2 ðo=oc Þ2n

:

(7:1:1)

The value of this function at o ¼ 0 is 1, at o ¼ 1, the function goes to zero, and, in between these

frequencies, the function decays. The low-pass filter passes frequencies between 0 and oc with small attenuation and blocks or attenuates the frequencies above oc , the cut-off frequency. In Section 6.11, we have considered deriving the transfer function HðsÞ from jHðjoÞj2 . In the next stage we are interested in coming up with a circuit that has the given transfer function. The circuit may consist of passive elements, such as resistors, inductors, capacitors, and transformers. Early filter designs were done exclusively with passive networks. Mathematics associated with passive network synthesis is elegant. There is very little leeway in the designs. See, for example, Weinberg (1962), and others for the passive filter limtations. Another problem of passive network synthesis is the use of inductors and transformers, as these are not ideal components in reality. Last part of the chapter deals with active filter synthesis using operational amplifiers, resistors, and capacitors. Active filter synthesis avoids the use of coils. Mathematical sophistication in active filter synthesis is much lower than passive filter synthesis. Active filter synthesis is based upon coming up with circuits with different topologies consisting of operational amplifiers, resistors, and capacitors. The circuit is then analyzed in terms of the R0 s and C0 s, assuming the operational amplifier is ideal. Comparing the derived and the given transfer function and equating the corresponding coefficients of s in the two transfer functions result in a set of equations with more unknowns than equations. As a result, we have infinite number of solutions for the component values. This gives a good deal of leeway for a circuit designer to optimize the circuits. One of the optimization criteria is

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_7, Ó Springer ScienceþBusiness Media, LLC 2010

243

244

7 Approximations and Filter Circuits

minimization of the sensitivity of a network with respect to changes in the component values. Introduction to sensitivity: In introducing sensitivity, Bode was concerned about the changes in the transfer function resulting from large changes in the element values in the transmission systems that included vacuum tubes. Even though we are in the era of integrated circuits, we are still interested in the effect of changes in the component values on the transfer function. This effect may be in the form of a shift in a pole frequency op or change in the quality factor Qor any other system parameter with respect to a component value. Pole sensitivity is defined as the per-unit change in the pole frequency, Dop =op , caused by a per-unit change in the desired component value Dx=x. The sensitivity of the parameter op with respect a component value xis defined by p So x ¼ lim

Dx!0

¼

x @op Dop =op =½Dx=x ¼ op @x

@ lnðop Þ : @ lnðxi Þ

Solution: We note that

Example 7.1.1 The transfer function (TF), HðsÞ ¼ V2 ðsÞ=V1 ðsÞ of the circuit in Fig. 7.1.1 is as follows:

Table 7.1.1 Formulae for computing sensitivities Y2 1 Y2 1 ¼ SY SY x x þ Sx

Y2 1 =Y2 1 SY ¼ SY x x Sx

Y SY xn ¼ ð1=nÞSx

Y SY x ¼ nSx

1 þY2 SY ¼ x

Y1 þY2

V2 ðsÞ 1=L1 C1 ¼ 2 V1 ðsÞ s þ ðR1 =L1 Þs þ ð1=L1 C1 Þ o20 ; ¼ 2 s þ ðo0 =QÞs þ o20 1 o0 L1 1 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ L1 =C1 : o0 ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; Q ¼ R1 R1 L1 C 1 (7:1:4)

(7:1:2)

@ðlnð1=YÞÞ @ð lnðYÞÞ ¼ ¼ SY ¼ x : (7:1:3) @ðln xÞ @ðlnðxÞÞ

Y Y Y1 SX1 þY2 Sx 2

HðsÞ ¼

Derive the sensitivities of the functions 1=2 1=2 o0 ¼ L1 C1 and Q with respect to the element values R1 ; L1 ; and C1 .

The parameter can be any that is important to the circuit’s function. We will assume the function of interest is Yi as a function of x. Formulas for sensitivities of simple functions can be seen by inspection. These are given in Table 7.1.1, where Yi ¼ Yi ðxÞ and xis a variable and c is a constant. From the sensitivity equation (7.1.2), we have S1=Y x

Fig. 7.1.1 Example 7.1.1

n

c YðxÞ

Sx

YðxÞ

¼ Sx

; Sxc ¼ 0

1=2

1=2

L1 1=2 C1

SLo10 ¼ SL1

¼

ðL1 C1 Þ1=2 @ðL1 Þ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ @L1 1= L1 C1 3=2

¼

3=2

ðL1 L1 2

Þ

1 ¼ : 2

(7:1:5)

Similarly, SCo10 ¼ ð1=2Þ; SRo10 ¼ 0; SLQ1 ¼ ð1=2Þ; SRQ1 ¼ 1:

(7:1:6)

0 Note that So R1 ¼ 0 since o0 is not a function of R. From (7.1.6), we have a 1% change in any one of the three component values in the RLC circuit in the example resulting in either (1/2)% or 0% change in o0 and (1/2)% or 1% change in Q.The sign of the values indicates whether the change is increasing or & decreasing. Transfer function of a circuit is a function of a set of parameters that are functions of the circuit components. For example, o0 is a function of L1 and C1 in Example 7.1.1. The change in o0 , Do0 can be approximated by using Taylor’s series written in terms of the variables xi in the form

7.1 Introduction

Do0 ¼

245

@o0 @o0 @o0 þ þ ::: þ @x1 @x2 @xn þ second and higher order terms:

MðsÞ ¼ G1 ðsÞEðsÞ ¼ G1 ðsÞðRðsÞ+BðsÞÞ; CðsÞ ¼ G2 ðsÞMðsÞ;

BðsÞ ¼ HðsÞCðsÞ

(7:1:7) CðsÞ ¼ G2 ðsÞG1 ðsÞ½RðsÞ+BðsÞ

See Tomovic and Vukobratovic (1972). If the change in the element values is assumed to be small, then the second-order and higher-order terms can be ignored. Do0 ﬃ

i¼1

Do0 ¼ o0

n X @o0 xi @xi @xi ¼ o0 : @xi @xi o0 xi i¼1

n X @o0

n X i¼1

0 So xi

Dxi ; xi

0 So xi

@o0 xi : ¼ @xi o0

CðsÞ½1 þ G2 ðsÞG1 ðsÞHðsÞ ¼ G1 ðsÞG2 ðsÞRðsÞ Transfer function:

(7:1:8) TðsÞ ¼

CðsÞ G1 ðsÞG2 ðsÞ : ¼ RðsÞ 1 þ G1 ðsÞG2 ðsÞHðsÞ

(7:1:10)

(7:1:9)

In Example 7.1.1 there are two important parameters, one is o0 and the other one is quality factor Q. Per-unit changes in Q in terms of its sensitivities can be expressed with respect to the parameters as well. In turn, the sensitivities of gain and phase of a transfer function in terms of the frequency o0 and Q can be determined. Block diagrams: In system control, system stability is one of the most important properties to be dealt with and closed-loop feedback control is basic to many systems. A simple feedback loop is shown in Fig. 7.1.2, where we have the Laplace transforms of the input signal with L½rðtÞ ¼ RðsÞ, error or actuating signal L½eðtÞ ¼ EðsÞ, control signal L½mðtÞ ¼ MðsÞ, controlled output L½cðtÞ ¼ CðsÞ, and primary feedback signal L½bðtÞ ¼ BðsÞ: The blocks identified by the transforms G1 ðsÞ; G2 ðsÞ; and HðsÞ represent control elements, plant or process, and feedback elements, respectively. The transfer function of the feedback system can be computed by writing the appropriate equations and solving for the output in terms of the input. These are as follows: LT

¼ G2 ðsÞG1 ðsÞ½RðsÞ+HðsÞCðsÞ

eðtÞ ¼ rðtÞ+bðtÞ ! RðsÞ+BðsÞ ¼ EðsÞ

The product GðsÞ ¼ G1 ðsÞG2 ðsÞ is the direct transfer function, HðsÞ is the feedback transfer function, the product GðsÞHðsÞ is the loop transfer function or the open-loop transfer function, and TðsÞ is the closed-loop transfer function. There are several books (for example, DiStefano et al. (1990).) that cover the block diagram algebra that gives simplifications in deriving the transfer functions. A useful transfer function that can be written in a special form is given by DiStfano et al. and the sensitivity of this function with respect to K is as follows: TðsÞ¼

A1 ðsÞþKA2 ðsÞ A3 ðsÞþKA4 ðsÞ ðK is independent of Ai ðsÞ;i¼1;2;3;4Þ:

) STK ¼

(7:1:11)

K½A2 ðsÞA3 ðsÞ A1 ðsÞA4 ðsÞ : (7:1:12) ½A3 ðsÞ þ KA4 ðsÞ½ðA1 ðsÞ þ KA2 ðsÞ

The transfer function can be expressed as a ratio of two polynomials in the form TðsÞ ¼

NðsÞ : DðsÞ

(7:1:13)

Example 7.1.2 Let a: T1 ðsÞ ¼ GðsÞHðsÞ ¼ ½K=ðs2 þ s þ 1Þ; b: T2 ðsÞ ¼ ½K=ðs2 þ s þ 1 þ KÞ. Determine the sensitivities of these functions to the parameter K.

Fig. 7.1.2 A simple feedback system

Solution: a. Using (7.1.11), we have A1 ðsÞ ¼ 0; A2 ðsÞ ¼ 1; A3 ðsÞ ¼ s2 þ s þ 1; A4 ðsÞ ¼ 0. Using T ðsÞ these in (7.1.12), it follows that SK1 ¼ 1 for all K.

246

7 Approximations and Filter Circuits

b. Again, using (7.1.11) and (7.1.12), we have

Consider M

P ð1 þ ð1=zm ÞsÞ

A1 ðsÞ ¼ 0; A2 ðsÞ ¼ 1;

HðsÞ ¼

A3 ðsÞ ¼ s2 þ s þ 1; A4 ðsÞ ¼ 1; T ðsÞ

SK2

Kðs2 þ s þ 1Þ 2 ðs þ s þ 1 þ KÞK 1 : ¼ 1 þ K=ðs2 þ s þ 1Þ

Ksd m¼1 N

(7:2:1b)

P ð1 þ ð1=pn ÞsÞ

n¼1

HðjoÞ ¼ HðsÞ s¼jo

¼

¼ KðjoÞd

(7:1:14)

ð1 þ jo=z1 Þð1 þ jo=z2 Þ:::ð1 þ jo=zM Þ ð1 þ jo=p1 Þð1 þ jo=p2 Þ:::ð1 þ jo=pN Þ

¼ jHðjoÞjejyðoÞ : T ðsÞ SK1

Sensitivity of the open-loop function is 1 for all values of K and the sensitivity of the closed-loop T ðsÞ function SK2 is a function of K and s. Using s ¼ jo, we can observe that for small values of o and K ¼ 1, T ðsÞ SK2 ﬃ :5. The feedback system is less sensitive than the open-loop system with respect to K. In Section 7.12 amplitude and phase sensitivities will & be considered. One of the main topics in this chapter is active filter synthesis. The first step in the active filter design is the analysis of a circuit with the appropriate topology. In Chapter 6 we considered computing the transfer function of a given circuit with twoterminal components using the Kirchhoff’s current and voltage laws and the component equations. Active filter circuits include multiple terminal components, including operational amplifiers (or op amps). These active devices are represented by controlled or dependent sources in the analysis. Kirchhoff’s laws, two-terminal component equations, and the controlled source representations of active devices provide a way to analyze circuits. A brief discussion on two-port representations of circuits is included by making use of the indefinite admittance matrix (Mitra, 1969). Other topics include scaling, frequency normalization, and adjustment of gain constants of the filter.

7.2 Bode Plots In this section we will study the basic concepts associated with a pictorial representation of a rational function, say HðsÞ or HðjoÞ, HðsÞ ¼ ½NðsÞ=DðsÞ; HðsÞjs¼jo ¼ jHðjoÞjejyðoÞ : (7:2:1a)

(7:2:1c)

Since the transfer is a ratio of real polynomials, the complex poles (or zeros) exist as complex-conjugate pairs and are usually simple. Multiple poles and zeros are possible and d is usually negative. In most cases, only a reasonable estimate of the system behavior and that to only at a very few frequencies is desired. The amplitude and phase responses at oi are given by M

P Zm

; Aðoi Þ ¼ jHðjoi Þj ¼ jKjjoi jd m¼1 N P Pn n¼1

Zm ¼ j1 þ joi =zm j; Pn ¼ j1 þ joi =pn j yðoi Þ ¼ﬀK þ dð90 Þ þ

M X

(7:2:1d)

tan1 ðoi =zm Þ

m¼1

N X

tan1 ðoi =pn Þ:

(7:2:1e)

n¼1

Although this approach is simple to see, finding these values and sketching them is time consuming. An alternate one is to obtain approximate sketches for the amplitude jHðjoÞj and the phase response yðoÞ using Bode plots using the following factors: 1. 2. 3. 4.

Constant term, K. Poles or zeros at the origin, s+k : Real poles or zeros, ðts þ 1Þ+k Complex-conjugate poles or zeros, ðt2 s2 þ 2xts þ 1Þ+k , where x is the damping ratio and 0 5x 5 1:

We need to study only simple poles and zeros, as the extensions to the multiple pole cases are simple. Note that log denotes base 10 and (ln) denotes base e.

7.2 Bode Plots

247

20 logjBðjoÞjN ¼ N logjBðjoÞj and ﬀ½BðjoÞN ¼ Nﬀ½BðjoÞ:

(7:2:2a)

logða þ jbÞ ¼ logja þ jbj þ j argða þ jbÞ

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ log a2 þ b2 þ j tan1 ðb=aÞ:

variable be defined by u ¼ logðoÞ or o ¼ 10u . The frequencies o1 and o2 are separated by an octave if o2 ¼ 2o1 and by a decade if o2 ¼ 10o1 . Note u2 u1 ¼ logðo2 Þ logðo1 Þ ¼ logðo2 =o1 Þ: (7:2:3)

(7:2:2b)

Note that the term t is used instead of the explicit poles or zeros and have combined the complex poles and their conjugates. Substituting s ¼ jo in the transfer function HðsÞ, we have HðjoÞ ¼ jHðjoÞj ﬀHðjoÞ or jHðjoÞjdB ¼ 20 logjHðjoÞj and fðoÞ ¼ ﬀHðjoÞ. We are only interested in sketching for positive frequencies. In addition, we will consider only poles that are on the negative half splane, including the imaginary axis and the zeros can be anywhere. In most applications, the poles and zeros are simple, with the exception that may include multiples at the origin. Since the log of a product is the sum of the corresponding logs of the terms, we can sketch the magnitude function by using the simple functions. The phase responses of the terms in the transfer function can be added to obtain the total phase response. The dB magnitude versus logðoÞ plot, i.e., logarithmic magnitude frequency response plot is called the Bode amplitude plot, and the phase angle versus logðoÞ is called Bode phase phase plot (or Bode diagrams). Logarithmic scale for the o-axis makes the sketches simple and allows sketches over a wider range of frequencies than the linear scale. Let the logarithmic frequency

Octaves ¼ log2 ðo2 =o1 Þ ¼ ½log10 ðo2 =o1 Þ= log10 ð2Þ; Decades ¼ log10 ðo2 =o1 Þ:

(7:2:4)

The amplitude and phase plots of HðjoÞ are considered using the four possible factors of a transfer function. Constant K: The logarithm of a constant is a constant with respect to o. The plot of 20 logðjKjÞ versus log(oÞ is a horizontal line. The phase angle is either 08 or –1808 depending upon whether the K is positive or negative. The factor (jo )N : 20 logjjojN ¼ 20 N logjoj; ﬀðjoÞN ¼ Nðp=2Þ:

(7:2:5)

Noting that log10 ð2Þ ¼ :3013, if o1 ¼ a and o2 ¼ 2a , the amplitude in (7.2.5) has increased by 6NdB/octave or 20NdB/decade. The function 20 N logjjoj plots as a straight line on the Bode plot and has a slope equal to 6 N dB/octave or 20N dB/decade. The slope of the line is positive (negative) depending on whether N is positive (negative). The magnitude and phase plots are shown in Fig. 7.2.1a,b for ð1=joÞ. It is simple to obtain the plots for multiple poles.

Bode Plot: Magnitude Plot of 1/jω

Bode Plot: Phase Plot of 1/jω

20

0

15

–45

5

Phase (deg)

Magnitude (dB)

10

0

–5

–90 –10 –15 –20 –1 10

0

10

Frequency (rad /sec)

Fig. 7.2.1 Amplitude and phase plots of (1/ðjoÞ)

10

1

–135 –1 10

0

10

Frequency (rad /sec)

1

10

248

7 Approximations and Filter Circuits

The factor ½1=ðjot þ 1Þ: Noting that A1 ðoÞ ¼ 20 logjð1 þ jotÞj ¼ 10 log½1 þ ðotÞ2 , we have for small and for large values of o, the amplitudes can be approximated by ðotÞ 1; AðoÞ 20 logð1Þ ¼ 0 dB; ðotÞ 1; AðoÞ 20 logðjotjÞ dB :

(7:2:6)

These are the asymptotes to the true curve corresponding to the very small and very large frequencies. The first asymptote is a horizontal line and the second asymptote is a straight line with a slope of –6 dB/octave or –20 dB/decade. The two asymptotes intersect at the corner frequency or the break frequency o ¼ 1=t. The actual value of the magnitudepfunction at this frequency is ﬃﬃﬃ equal to 20 logð 2ÞdB 3dB. It is simple to draw the asymptotic and the actual curves using the following guidelines: 1. The constant t is the break point (or the corner frequency) in the asymptotic plot. 2. From the break point, draw the two asymptotes, one with a zero slope toward the o small and the other one with a –6 dB/octave slope extending toward o ! 1. 3. At the break point, the true response is displaced by –3 dB. In addition, an octave below and above the break point, the true curve is separated by –1 dB. A sketch of the amplitude response using the table and the above guidelines is shown in Fig. 7.2.2a. Note the frequency is

plotted using the log scale. The phase angle of the term ½1=ð1 þ jotÞ is equal to f1 ðoÞ ¼ tan1 ðotÞ radians or ½57:3 tan1 ðotÞ degrees: It can be approximated using the power series expansion (Spiegel, 1966): tan1 ðotÞ ¼ ðotÞ ð1=3ÞðotÞ3 þ ð1=5ÞðotÞ5 :::; jotj41;

(7:2:7a)

tan1 ðotÞ ¼ +ðp=2Þ ½1=ðotÞ ð1=3Þð1=otÞ3 þ ð1=5Þð1=otÞ5 :::; þ if ot 1; if ot 1: (7:2:7b) tan1 ðotÞ ¼ p=4;

ot ¼ 1:

(7:2:7c)

Figure 7.2.2 gives the Bode amplitude and phase plots. The phase angle plot approaches 08 as ðotÞ ! 0 and – 908as ðotÞ ! 1. Noting (7.2.7c), we can see that the phase angle is –458 at the break frequency o ¼ 1=t. These two asymptotes can be connected by drawing a line from the 08 asymptote starting at one decade below the break frequency ð:1=tÞ with 08 phase and draw a line with a slope of 458/decade passing through – 458at the break frequency and continuing to –908 one decade above the break frequency ð10=tÞ. For the zeros, the amplitude and phase response sketches can be similarly drawn since only the signs need to be altered. Quadratic factors: The Bode plots corresponding to a pair of complex poles are usually given in terms of the damping factor x 1 by

Bode Plot: Magnitude Plot of 1/(jω + 1)

Bode Plot: Phase Plot of 1/(jω + 1)

0

0

–5

–15

–30 Phase (deg)

Magnitude (dB)

–10

–20 –25

–60

–30 –35 –40 –45 10–2

10–1

100 Frequency (rad /sec)

101

102

–90 10–2

10–1

100 Frequency (rad /sec)

Fig. 7.2.2 Bode amplitude 1=j1 þ joj and phase ﬀ1=ð1 þ joÞ plots, break frequency ¼ 1

101

102

7.2 Bode Plots

249

1 ½1 þ 2xts þ t2 s2 1 ; 0 x 1: (7:2:8) ¼ ½1 þ ðQp =op Þs þ ð1=o2p Þs2

H2 ðsÞ ¼

jH2 ðjoÞj2 ¼ 1=½ð1 o2 t2 Þ2 þ 4ðxtoÞ2 ; A2 ðoÞ ¼ 20 logðjH2 ðjoÞjÞ

(7:2:9a)

f2 ðoÞ ¼ tan1 2xot=ð1 o2 t2 Þ

(7:2:9b)

It is common in the filter designs to use the quality factor Qp and op in (7.2.8). The peak of the magnitude squared function can be found by taking the derivative of the denominator in (7.2.9a) with respect to o and equating to zero. That is,

The high frequency asymptote is a straight line with a slope of –12 dB/octave (–40 dB/ decade). 2. The break frequency, i.e., the intersection of the low-frequency and the high-frequency asymptotes is located at the frequency o ¼ ð1=tÞ, which can seen from the fact that 40 logðotÞ ¼ 0; o ¼ 1=t for all x. At the break frequency, we have 20 logjH2 ðjoÞjo¼1=t ¼ 20 logð2xÞ dB:

As an example at x ¼ :2, we have the value –7.958 dB and for x ¼ 1=2, the above equation reduces to 0 dB. A few values are given below for (7.2.13). x –20log(2x)dB

d½1=jH2 ðjoÞj2 ¼ 2ð1 o2 t2 Þð2ot2 Þ þ 2ð4x2 t2 oÞ do pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 2x2 : (7:2:10) ¼0!o¼ t pﬃﬃﬃ 1 2x2 40 or x5ð1= 2Þ ¼ :707 ðo is realÞ: (7:2:11) Asymptotic approximations – second-order case: H2 ðjoÞ ¼ 1=½1 þ j2xto t2 o2 1. For low and high frequencies we can write: ðotÞ 1; 20logjH2 ðjoÞj ¼ 0 dBðotÞ 1; 20logjH2 ðjoÞj 40 logðotÞ dB:

(7:2:12)

0 1

.05 20

.1 14

.2 8

.3 4.5

.4 2

.5 0

.6 –1.5

.707 –3

1 –6

Now consider the phase asymptotic plots of the second-order function from the f2 ðoÞ given in (7.2.9b). At the break frequency, f2 ðoÞ o¼ð1=tÞ ¼ tan1 ð1Þ ¼ 90 for all x:

(7:2:14)

The phase starts at 0 at low frequencies. At o ¼ ð:1=tÞ, it starts to decrease at a rate of –908/ decade. At o ¼ ð1=tÞ, it is –908. It continues to decrease and reaches–1808 as o ! 1. The amplitude and phase responses are plotted for a few values of x in Fig. 7.2.3. The phase response will

2

2

Bode Diagram: Magnitude Plot of 1/(1+ j2ξω-ω )

Bode Diagram: Phase Plot of 1/(1+ j2ξω-ω )

30

0 ξ = 0.05 ξ = 0.2 ξ = 0.4

ξ = 0.05

20

ξ = 0.2 10

ξ = 0.4

ξ = 0.6

–45

ξ=1

ξ = 0.6

0

Phase (deg)

Magnitude (dB)

(7:2:13)

ξ=1

–10

1 –20

–90

–135

–30 –40 –50 –1 10

–180 10

0

1

10

–1

10

Frequency (rad/sec)

Fig. 7.2.3 Bode plots jH2 ðjoÞj ¼ 1=½1 þ j2xto t2 o2 ; ﬀH2 ðjoÞ; 0 x 1

0

10

Frequency (rad/sec)

1

10

250

have a discontinuity going from 0 to –1808 at the break point corresponding to x ¼ 0. The above discussion is given in terms of poles. For the zeros of a transfer function, multiply the dB and the phase angle values by –1. Bode plots of the transfer function can be constructed by summing the log magnitudes and phase angle contributions of each pole and zero (or pairs of complex and their conjugates of poles and zeros). DiStefano III et al. (1990), Nise (1992), and others give systematic procedures for sketching the Bode plots with examples to illustrate the construction process. Thaler and Brown (1960) give a tabulation of typical control system transfer functions, with associated polar plots, Bode plots, and root locus plots. There are a few ways of computing the amplitude versus frequency on the Bode plot. Construct each factor separately and at selected values of o add the amplitudes. Then, sketch the amplitude function through these points. It is simpler to use the asymptotes. For this purpose we need to have the transfer function in factored form. Next arrange the transfer function with increased values of poles and zeros. Many of the control system transfer functions have poles at the origin with a multiplicity of l; l 0. Such a system is called a type l system. The system amplitude plot has a slope at low frequencies of 10 l dB/decade or (–6l dB/octave). This slope is maintained until the first corner frequency is reached. At the first corner frequency the slope is changed by +10 dB=decade for a first-order pole or zero. If it is a second-order function, then the slope is changed by +20 dB=decade. This procedure is used to sketch the asymptotic plot. We can construct the composite asymptote provided the exact location of the lowest frequency segment can be located. For l ¼ 0, the lowest-frequency asymptote with a constant gain K is 20 logðKÞ dB. For l ¼ 1, locate the point o ¼ K on the 0 dB axis. The lowest frequency asymptote passes through this point with a slopepof ﬃﬃﬃﬃ –10 dB/decade. For l ¼ 2, locate the point o ¼ K on the 0 dB axis. This frequency asymptote passes through this point with a slope of –20 dB/ decade. It is very rare to have more than a double pole at the origin. Noting that for oT 1, all the factors of the form ðjoT þ 1Þ reduce to 1 in this region. Then, we have

7 Approximations and Filter Circuits

20 logjHðjoÞj ﬃ 20 logðKÞ 20 log ðjol Þ : (7:2:15a) For example, for l ¼ 2, the above equation represents a straight line with a slope of –20 dB/decade and the corresponding intercept is determined by pﬃﬃﬃﬃ 20 logðKÞ 20 log ðjo2 Þ ¼ 0 ! o ¼ K: (7:2:15b) Example 7.2.1 Sketch the Bode plots for the following transfer function: HðsÞ¼ HðjoÞ¼

10ð1þsÞ s2 ð1þðs=4Þþðs=4Þ2 Þ

10ð1þjoÞ 2

ðjoÞ ð1þð1=4Þjoðð1=4ÞÞ:25oÞ2 Þ

; : (7:2:16)

Solution: The corresponding amplitude in terms of dB and the phase responses are 20 logjHðjoÞj ¼ 20 logð10Þ þ 20logjð1 þ joÞj 40logjjoj 20 log 1=½ð1 þ jðo=4Þ ðo=4Þ2 ; (7:2:17)

ﬀHðjoÞ ¼ﬀð1 þ joÞ þ ﬀð1=joÞ2 þ ﬀ½1=ð1 þ jo=4 ðo=4Þ2 Þ:

(7:2:18)

The asymptotic amplitude plot is obtained by adding the asymptotic plots of each of the terms. The first term on the right in (7.2.17) is equal to 20 dB for all values of o. The third term corresponds to a double pole at the origin and the asymptote goes through the corner frequency o ¼ 1 with a slope of –40 dB/decade. The high-frequency asymptote of the second term in (7.2.17) starts at the corner frequency o ¼ 1 and has a slope of 20 dB/decade. The fourth term corresponds to a pair of complex poles. Noting t ¼ 1=4 and 2xt ¼ :25 (x ¼ .5, damping factor). The high-frequency asymptote of the complex pair of poles starts at the corner frequency o ¼ 4 with a slope of –40 dB/decade. All of these are sketched in Fig. 7.2.4 using MATLAB software. Before we obtain the composite amplitude asymptotic Bode plot of the transfer function, we need to locate the lowest frequency asymptote. Using (7.2.15b), the corresponding pﬃﬃﬃﬃ pﬃﬃﬃﬃﬃ asymptote is a line through the point K ¼ 10 at a slope of –40 dB/decade on the 0 dB axis. We are now ready to sketch the asymptotic Bode magnitude plot of the

7.2 Bode Plots

251

Fig. 7.2.4 (a) Bode amplitude and (b) Bode phase plots of individual factors

Bode Diagram

Magnitude (dB)

100

50

0

–50

10 1+s 1/s2 2

1/(1 + 0.25s – (0.25s) )

(a)

–100 90

Phase (deg)

45 0 –45 –90 –135 –180 10–2

(b) 10–1

100

101

102

Frequency (rad/sec)

transfer function by adding the asymptotic plots of each of the four terms. Start at a low frequency say o ¼ :1, follow the low-frequency asymptote that pﬃﬃﬃﬃﬃ passes through o ¼ 10 on the 0 dB axis to the first corner frequency o ¼ 1 corresponding to the (jo þ 1Þ factor. Since this is a numerator factor, the next asymptote changes from –40 dB/ decade to –20 dB/decade and continues to the next corner frequency located at o ¼ 4. The next asymptote takes into consideration of the second-order factor in the denominator. The last high-frequency asymptote starts at o ¼ 4 with a change in slope to –60 dB/decade. The individual

amplitude and phase plots of the transfer function are shown in Fig. 7.2.4a and b. We can obtain the actual amplitude plot by correcting for the errors in the asymptotic plots and the true functions. An easy way to do is find the amplitudes at the corner frequencies and then sketch the function through the computed values. Bode plots are only sketches. If accurate plots are desired, software packages, such as MATLAB, need to be used to obtain the desired results. Phase approximations can be used for the phase angle asymptotic plots. To get a sketch, it is easier to make use of arctangent function to obtain the

MATLAB code for Example 7.2.1 %Plot individual terms (Figure 7.2.4 Example 7.2.1) sys1=tf (10,1 ); sys2=tf ([1 1],1); sys3=tf(1,[1 0 0]); sys4=tf(1,[-0.25 0.25 1]); bode(sys1,’k’,sys2,’k– –’,sys3,’k-.’,sys4,’k:’,{0.01,100}) legend(’10’,’1+s’,’1/s^2’,’1/(1+0.25s-(0.25s)^2)’) %Plot the whole Bode plot (Figure 7.2.5, Example 7.2.1) num=10*[1 1]; den=[-(1/4) (1/4) 1 0 0]; w=logspace(-2,2,100); bode(num,den,w) grid

252

7 Approximations and Filter Circuits

Fig. 7.2.5 Bode amplitude and phase plots of the composite function

Bode Diagram 150

Magnitude (dB)

100

50

0

–50

Phase (deg)

–100 –90

–135

–180 –2 10

10

phase angles at some important frequencies, such as at corner frequencies, and sketch the function using these values. The amplitude and phase plots for the composite are given in Fig. 7.2.5a,b using the MATLAB code given below.

–1

0

10 Frequency (rad/sec)

10

1

10

2

FPM ¼ 1800 þ ﬀH0 ðjoc Þ with Hðjoc Þ ¼ 1; oc ¼ gain crossover frequency:

(7:2:20)

They are measures on how closely H0 ðjoÞ approaches a magnitude of unity and a phase of – MATLAB code for Example 7.2.2

7.2.1 Gain and Phase Margins We would like to consider two important topics that are used in the stability analysis of feedback control systems. Our discussion will be brief. The characteristic polynomial of a feedback control system is DðsÞ ¼ 1 þ H0 ðsÞ ¼ 0 with H0 ðsÞ ¼ G1 ðsÞG2 ðsÞ HðsÞ (see 7.1.10). Practicing engineers use the gain margin (GM ) and the phase margin (FM ), see, for example, Melsa and Schultz (1969). Graphical analysis is more appealing to engineers than analytical analysis. Gain and phase margins are measures of relative stability of the feedback control system. These are defined at the phase and margin crossover frequencies op and oc . 1 with ﬀHðjop Þ ¼ 1800 ; jH0 ðjop Þj op ¼ phase crossover frequency;

GM ¼

(7:2:19)

n=5 ;d=[1 3 4 2]; w=logspace(–2,2,100), [mag, phase]=bode (n,d,w);margin(mag, phase,w)

1808 quantifying the relative stability of the system. They can be read using the Bode plots. Negative phase margin implies instability. Most engineers use the criteria that a phase margin of 308 and a gain margin of 6 dB are safe margins. Analytical computation of gain and phase margins may not be possible since it requires factoring polynomials. Example 7.2.2 Using the following methods obtain the gain and phase margins: a. Analytical methods and b. MATLAB for the following function: 5 ) HðjoÞ H0 ðsÞ ¼ 3 2 s þ 3s þ 4s þ 2 5 ¼ : (7:2:21) ð2 3o2 Þ þ joð4 o2 Þ

7.2 Bode Plots

253

Solution: a. Equate the imaginary part to zero, i.e., ImðH0 ðjoÞÞ ¼ 0. Solving for op and then evaluating the real part at this frequency, we have ImðH0 ðjoÞÞ ¼ 0; op ¼ 2 ! ReðH0 ðjop ÞÞ ¼ 1=2:

We can use MATLAB command roots ([1,0,1,0,4,0,–21]) and obtain the real positive root of the polynomial given by oc ¼ 1:4315 resulting in

(7:2:22)

We can increase the gain by 2 before the real part becomes –1. The gain margin in dB is GM ¼ 20 logð2Þ ﬃ 6 dB:

ﬀH0 ðjoÞjo¼oc ¼1:4315 ¼ ﬀ 5=ð2 3o2 Þ þ joð4 o2 Þ o¼o ¼1:4315 1460 : c

(7:2:23)

We could also solve for op by noting that the characteristic polynomial

The phase margin, the difference between this angle and –1808, is

DðsÞ ¼ 1 þ aH0 ðsÞ ) 0 ! s3 þ 3s2 þ 4s þ 12 ¼ ðs2 þ 4Þðs þ 3Þ ¼ 0

FM 180 146 ¼ 34 :

has imaginary roots given by s ¼ + j2. See the discussion on Routh table Chapter 6. For phase margin, we need to equate jH0 ðjoÞj ¼ 0 and solve for o ¼ oc . This requires software, such as MATLAB. These result in 25 ¼ð2 3o2 Þ2 þ o2 ð4 o2 Þ2 ! o6 þ o4 þ 4o2 21 ¼ 0:

(7:2:24)

b. Analytical computation may not be possible and computational tools, such as MATAB, are good to use. For this example the code is given below. MATLAB Bode plots are given in Fig. 7.2.6. The gain and the phase margins are & shown.

Bode Diagram Gm = 6.03 dB (at 2 rad /sec), Pm = 34.1 deg (at 1.43 rad /sec) 20

Magnitude (dB)

0 –20 –40 –60 –80 –100 –120 0

Phase (deg)

–45 –90 –135 –180 –225

Fig. 7.2.6 Illustration of gain and phase margins, Example 7.2.2

–270 10–2

10–1

100 Frequency (rad/sec)

101

102

254

7 Approximations and Filter Circuits

7.3 Classical Analog Filter Functions 7.3.1 Amplitude-Based Design

The ideal low-pass filter function was defined in Chapter 6 and is HLp ðjoÞ ¼ P½o=2oc ejot0 ;

Noting jHðjoÞj is even, start with

oc ¼ 2pfc ; cut - off frequency:

jHðjoÞj2 ¼ HðjoÞHðjoÞ M P

¼

al o2 l

l¼0

1þ

N P

; N M:

(7:3:1)

bl o2 l

l¼1

The goal is to determine the coefficients al and bl so that (7.3.1) satisfies a given set of specifications on the amplitude response. Letting o ! s=j, we have the product HðsÞHðsÞ. The minimum phase transfer function is obtained by assigning the left halfplane poles and zeros of HðsÞHðsÞ to HðsÞ. In this section we will consider Butterworth, Chebyshev I, and Chebyshev II filter functions.

The ideal low-pass filter response in (7.3.2) is not physically realizable, see Section 6.10.1. The next best thing is approximate an ideal filter response. We have seen that a simple RC circuit approximates a low-pass filter. There are four types of filters identified by 1. 2. 3. 4.

HLp ðjoÞ, low-pass function HHp ðjoÞ, High-pass function HBp ðjoÞ, Band-pass function HBe ðjoÞ, Band-elimination function

Low-pass filter specifications: 9 8 < Pass band : 0 joj oc ; 1 2 HLp ðjoÞ 2 1 or XdB 20 log HLp ðjoÞ 0dB = 1þe : : Stop band : joj o ; H ðjoÞ 2 ð1=A2 Þ; or 20 log H ðjoÞ YdB; o 5o ; r Lp Lp c r

High-pass filter specifications: 8 < Pass band : oc joj 1; : Stop band : 0 joj o ; r

(7:3:2)

(7:3:3a)

9 2 HHp ðjoÞ 1 or X dB 20 log HHp ðjoÞ 0 dB = : (7:3:3b) HHp ðjoÞ 2 ð1=A2 Þ; or 20 log HHp ðjoÞ Y dB; or 5oc : ;

1 1þe2

Band-pass filter specifications: 9 8 1 HBp ðjoÞ 2 1 or X dB 20 log HBp ðjoÞ 0 dB > > o o ; Pass band0 :o j j l h 2 > > 1þe = < 2 2 Stop bands :0 joj o1 and o2 joj 1; HBp ðjoÞ ð1=A Þ; > > > > ; : or 20 log HBp ðjoÞ Y dB; o1 5ol ; oh 5o2 : Band elimination filter specifications: 9 8 2 2 > > Pass bands : 0 o o ; o o 1; ð1=ð1 þ e ÞÞ H ðjoÞ 1 j j j j l h Be > > > > > > = < or X dB 20 logjHBe ðjoÞj 0 dB : 2 > > > > Stopband : o1 joj o2 ; 0 jHBe ðjoÞj ð1=ð1 þ e2 ÞÞ; > > > > ; : or YdB 20 logjHBe ðjoÞj;ol o1 ; oh o2 :

(7:3:3c)

(7:3:3d)

7.3 Classical Analog Filter Functions

255

Fig. 7.3.1 Analog filter amplitude specifications: (a) low pass, (b) high pass, (c) band pass, and (d) band elimination

(a)

(b)

(c)

(d)

These are illustrated in Fig. 7.3.1a,b,c,d in terms of the squares of amplitudes and in dB scale on all the figures. For simplicity the responses are assumed to be smooth.

7.3.2 Butterworth Approximations A simple RC circuit was considered in Chapter 6 with a transfer function (see 6.5.16b) HðjoÞ ¼

1 ; jHðjoÞj ¼ HðsÞ s¼jo joRC þ 1

1 ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; oc ¼ 1=RC: ð1 þ ðo=oc Þ2

(7:3:4)

It has a single parameter oc controling the amplitude response. Assuming the input and the output transforms as XðjoÞ and YðjoÞ, we have YðjoÞ ¼ HðjoÞXðjoÞ. This implies jYðjoÞj ¼ jHðjoÞjjXðjoÞj. The filter acts like a gate in the sense that low frequencies are passed with very little attenuation and the high frequencies are attenuated significantly. The shape of the amplitude response function jHðjoÞj controls what frequencies are allowed through and what frequencies are attenuated or eliminated. To be more specific, in the low-pass filter design, we assume we have three bands defined by Pass band : 0 o oc ; transition band: oc 5o5or ; stop band : or o 1:

(7:3:5)

The frequencies 0 o oc in the input will be allowed to pass through the low-pass filter without much attenuation and therefore we call this band of frequencies as the pass band. The band of frequencies or o 1 in the input signal will be attenuated by the low-pass filter significantly and this band is called the stop band. In between these two bands the filter amplitude response will have to be tapered or gradual and the input frequencies are attenuated gradually. A popular function that can act as a low-pass filter that satisfies the above criterion is a Butterworth function defined by 1 jHBu ðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : 1 þ e2 ðo=oc Þ2n

(7:3:6)

The subscript Bu on H is usually not shown and seen from the context. The e term controls how far the filter amplitude characteristic will go down to when o ¼ oc from 1 at o ¼ 0 and the value of n controls how fast the magnitude characteristic attenuates in the stop-band region. We have shown that the ripple amplitude in the band ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ppass having the value for small e is 1 ð1= ð1 þ e2 Þ e2 =2 (see 6.10.10). The pass-band and stop-band specifications for the low-pass filter are given in (7.3.3a) and the specifications are identified in Fig. 7.3.1a. It is common to specify these two in terms of the dB scale, as identified in this figure. Using the edges of the frequency bands oc and or , we can write

256

7 Approximations and Filter Circuits

Since js2nþ1 j ¼ ðoc =e1=n Þ, poles of the Butterworth function are equally spaced on a circle of radius (oc =ðe1=n ÞÞ on the splane. Selecting the poles on the left half of the s plane, the transfer function is

jHBu ðjoÞj2o¼oc ¼ 10 logð1=ð1 þ e2 ÞÞ ¼ X dB ) e pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ 10ðx=10Þ 1:

(7:3:7) HBu ðsÞ ¼

jHBu ðjoÞj2o¼or ¼10logð1=A2 Þ ¼ YdB:10log

ð Note HBu ð0Þ ¼ 1Þ:

1 1 þ e2 ðor =oc Þ2n

¼ YdB ) ðnÞinteger

ðnÞinteger

ﬃ

log½ð10:1Y 1Þ=ð10:1X 1Þ : (7:3:8) 2logðor =oc Þ log

(7:3:14)

Example 7.3.1 Compute the values of e and n and derive Butterworth transfer function for the specifications given in Fig. 7.3.2 with X ¼ 2 dB and Y ¼ 15dB.

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ðA2 1Þ logðeÞ logðoocr Þ

lnðA=eÞ : ðor =oc Þ 1

(7:3:9)

Now compute the transfer function HBu ðsÞ from (7.3.6). That is, jHBu ðjoÞj2 ¼ HBu ðsÞHBu ðsÞjo2 ¼s2 :

(7:3:10)

Since the amplitude response function has one in the numerator, we need to compute only the poles of the transfer function. They can be found by solving 2 n 2 o 2n 2 ðs Þ 1 þ e ð Þ jo2 ¼s2 ¼ 1 þ e ¼0 o2n oc c o p c ) s2nþ1 ¼ 1=n ejð2nþ1þnÞ2n ; e n ¼ 0; 1; 2; . . . ; 2n 1:

Solution: From (7.3.7) and in (7.3.8), e ¼ :7648 and n 3:76. Since n has to be an integer, it follows that n ¼ 4: From (7.3.13b), the left-half s-plane poles and the transfer function are s1;7 ¼ ðoc =0:93516Þð0:38268+j0:92388Þ;

(7:3:11) HBu ðsÞ ¼

(7:3:12)

p ¼ oc =e1=n ejð2nþ1þnÞ2n ;

n ¼ 0; 1; . . . ; n 1:

Fig. 7.3.2 Specifications in Example 7.3.1

s3;5 ¼ ðoc =0:93516Þð0:92388+j0:38268Þ

The left half-plane poles of HBu ðsÞ in the exponential and trignometric forms are s2nþ1

ðs1 Þðs3 Þ . . . ðs2n1 Þ : ðs s1 Þðs s3 Þ . . . ðs s2n1 Þ

ðs1 Þðs3 Þðs5 Þðs7 Þ : (7:3:15)& ðs s1 Þðs s3 Þðs s5 Þðs s7 Þ

Maximally flat amplitude property Butterworth function: Consider jHBu ðjoÞj2 ¼

1 1þ

e2 ðo=oc Þ2n

of

:

Even function of ðo=oc Þ2 ¼ l: :

(7:3:16)

(7:3:13a)

2n þ 1 s2nþ1 ¼ oc =e1=n ½ sinð pÞ 2n 2n þ 1 pÞ; n ¼ 0; 1; . . . ; n 1: (7:3:13b) þ j cosð 2n

the

That is, jHBu ðjlÞj2 ¼ jHBu ðjoÞj2 ðo=oc Þ2 ¼l ¼

1 1 þ e 2 ln

7.3 Classical Analog Filter Functions

257

Expanding this function in power series in the neighborhood of l ¼ 0, we have 1 ¼1 þ ð0Þl þ ::: þ ð0Þln1 þ ðe2 Þln þ ::: 1 þ e 2 ln (7:3:17) A simple way to see this is divide the numerator (1) by the denominator ð1 þ e2 ln Þ. The coefficients for the terms lk ; k ¼ 1; 2; :::; n 1 are identically zero. That is, ðn 1Þ derivatives of the Butterworth function are equal to zero. This is called the maximally flat property. Butterworth approximation starts with 1 at o ¼ 0 and monotonically goes to zero as o ! 1. It has the maximally flat response in both pass and stop bands. In the pass band p the magniﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ tude of the function goes from 1 to (1= ð1 þ e2 Þ). Zero error at o ¼ 0 and maximum error at the cutoff frequency oc is approximately equal to e2 =2. It may be of interest to distribute the error throughout the pass band and the Chebyshev approximation achieves that.

Cn ðaÞ ¼ cosðn cos1 aÞ cosðn cos1 ðaÞÞ; ¼ coshðn cosh1 ðaÞÞ;

j aj 1 : j aj 4 1

(7:3:18)

It can be expressed as a polynomial. To show this, let nf ¼ cos1 ðaÞ and Cn ðaÞ ¼ cosðnfÞ: ðC0 ðaÞ ¼ 1 and C1 ðaÞ ¼ aÞ:

(7:3:19a)

Using the trigonometric identities, we can write cnþ1 ðaÞ ¼ cos½ðn+1Þf ¼ cosðnfÞ cosðfÞ+ sinðnfÞ sinðfÞ: Cnþ1 ðaÞ ¼ 2aCn ðaÞ Cn1 ðaÞ; C0 ðaÞ ¼ 1; C1 ðaÞ ¼ a: C2n ðaÞ ¼ :5½C2n ðaÞ þ 1:

(7:3:19b)

Using (7.3.19b) the Chebyshev polynomials can be derived. First few of these are C0 ðaÞ ¼ 1; C2 ðaÞ ¼ 2a2 1;

C1 ðaÞ ¼ a C3 ðoÞ ¼ 4a3 3a

C4 ðaÞ ¼ 8a4 8a2 þ 1; C5 ðaÞ ¼ 16a5 20a3 þ 5a

7.3.3 Chebyshev (Tschebyscheff) Approximations

(7:3:19c)

The nth (n40Þ order Chebyshev polynomial is defined by the transcendental function

These are sketched for n ¼ 1; 2; 3; 4 in Fig. 7.3.3 in the range 0 a. Since the Chebyshev polynomials Cn ðaÞ have even (odd) powers of a only for neven

3.5 3

n=1 n=2

2.5

n=3

2

n=4

Cn(α)

1.5 1 0.5 0 –0.5 –1

Fig. 7.3.3 Chebyshev polynomials, cn ðaÞ; n ¼ 1; 2; 3; 4

–1.5

0

0.1

0.2

0.3

0.4

0.5

0.6

α

0.7

0.8

0.9

1

1.1

258

7 Approximations and Filter Circuits

(odd), sketching the polynomials for negative a is straightforward. For large values of a, Cn ðaÞ 2n1 an ;

a 1:

(7:3:19d)

Properties of the Chebyshev polynomials (see Scheid (1968)): 1. Since coshðaÞ is never zero for real o, it follows that Cn ðaÞ ¼ 0 only for jaj 1: The roots of Cn ðaÞ=0, ak ; k ¼ 1; 2; ::; n are real and jak j51: 2. Since Cn ðaÞ ¼ cosðn cos1 ðaÞÞ; jaj 1; it follows that jCn ðaÞj 1 for jaj 1. 3. For jaj41, Cn ðaÞ increases monotonically consistent with the degree n. 4. Cn ðaÞ is an odd (even) polynomial if n odd (even). 5. The polynomial Cn ðaÞ oscillates with an equiripple character varying between a maximum of +1 and a minimum of –1 for joj 1. " # 6. neven:Cn ð0Þ¼ð1Þn=2 ;Cn ð1Þ¼1 : nodd:Cn ð0Þ¼0;Cn ð+1Þ¼+1ðrespectivelyÞ Cn ðaÞ ¼ 0; a ¼ cosðð2 k þ 1Þp=2nÞ; k ¼ 0; 1; 2; :::; n 1; 1 a 1:

(7:3:20a)

Cn ðaÞ ¼ ð1Þk ; a ¼ cosðkp=nÞ; k ¼ 0; 1; 2; :::; n; 1 a 1:

(7:3:20b)

7. Slope

dCn ðaÞ ja¼1 ¼ n2 : da

an even function. Chebyshev polynomial gives the best approximation in the sense that it minimizes the maximum magnitude of the error for a given value of n. Chebyshev 1 approximation: Noting the characteristics of the Chebyshev polynomials in the range 1 a 1, e2 C2n ðaÞ varies between 0 and e2 in the interval jaj 1 and increases rapidly for jaj41 consistent with n. With these properties in mind, a low-pass amplitude response function can be defined, so that the response swings between 1 and pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1= ð1 þ e2 Þ, in the pass band and monotonically decreasing property in the stop band. Such a function is the Chebyshev 1 function given with a subscript ðc1Þ with the argument ðo=oc ) is given below, see, for example, in Figs. 7.3.4 and 7.3.5. jHc1 ðjoÞj2 ¼

1 : 1 þ e2 C2n ðo=oc Þ

For jo=oc j 1; jHc1 ðjoÞj oscillates between 1 and 1=ð1 þ e2 Þ with equal ripple character. 2

jHc1 ð0Þj ¼

ð1=ð1 þ e2 Þ;

n even;

1;

n odd;

(7:3:20c)

Equations (7.3.20a) and (7.3.20b) follow from

Fig.7.3.4 jHc1 ðjoÞj; n ¼ 1; 2; 3

0 ¼ cosðn cos1 ðaÞÞ ) n cos1 ðaÞ ¼ kp=2; k-odd;

(7:3:22a)

(7:3:21a)

+1 ¼ cosðn cos1 ðaÞÞ ) cos1 ðaÞ ¼ kp=n: (7:3:21b) The Chebyshev polynomial has n roots and they are located in the range 1 a 1. Outside this range, Cn ðaÞ is monotonically increasing (or decreasing in the case of negative a) function. Since Cn ðaÞ is either an even or an odd function, it follows that C2n ðoÞ is Fig. 7.3.5 jHc1 ðjoÞj; n ¼ 4; 5

(7:3:22b)

7.3 Classical Analog Filter Functions

259

jHc1 ðjoc Þj2 ¼ ð1=ð1 þ e2 Þ; n even and n odd: (7:3:22c) The number of peaks ðjHc1 ðjoÞj ¼ 1Þ plus the number of valleys ðjHc1 ðjoc Þj) in the positive frequency range of the pass band is equal to n. This is referred to as the equal-ripple property. Fig. 7.3.5 illustrates this for n ¼ 4; 5. For joj4joc j, jHc1 ðjoÞj decreases rapidly consistent with the value of n. For e small, the width of the ripple in the pass band can be approximated and is e2 =2 (see 6.10.10). From the filter specifications, e gives the permissible range of amplitudes of the Chebyshev 1 response in the pass band and the stop-band attenuation constant A gives a measure of acceptable attenuation in the stop band. The range of frequencies between oc and or is the transition band. Chebyshev 1 transfer function can be computed from Hc1 ðsÞHc1 ðsÞ ¼ Hc1 ðjoÞHc1 ðjoÞ o¼s=j 1 o¼s=j : ¼ o 1 þ e2 c2n ðoc Þ

(7:3:23)

Solving for the roots of the equation Cn ðo=oc Þ o¼s=j ¼ +j=e and selecting the left-halfplane roots results in the following poles of the transfer function (see Weinberg, 1962):

1 1 2i þ 1 sin p si ¼ oc sinh sinh1 n e 2n

1 2i þ 1 1 1 cos p ; þ j cosh sinh n e 2n i ¼ 0; 1; 2; :::; n 1:

(7:3:24)

Design parameters: e controls the ripple width in the pass band and n controls the attenuation in the stop band. That is, 1 1 qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ jor ¼oc ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; or 2 2 1 þ e2 1 þ e Cn ðoc Þ 1 1 ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; or 4oc : A 1 þ e2 C2 ðor Þ

Noting Cn ðaÞ ¼ coshðn cosh1 ðaÞÞ; jaj41, the integer n must satisfy pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ cosh1 ðð A2 1Þ=eÞ n

cosh1 ðor =oc Þ lnð2A=eÞ : (7:3:27) ½ð2=oc Þðor oc Þ1=2 The approximation follows from cosh1 ðxÞ ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2 lnðx þ x 1Þ lnð2xÞ; x 1. If the constraints are given in terms of dB, then

n

cosh1 ½ð10:1Y 1Þ=ð10:1X 1Þ : cosh1 ðor =oc Þ

Hc1 ðsÞ ¼

> > > > > :

Solution: Noting C2n ð1Þ ¼ 1, " # 1 10 log jo¼oc ¼ 2 dB 1 þ e2 C2n ðooc Þ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ) e ¼ ð102 1Þ ¼ 0:7648:

ðs1 Þðs2 Þ:::ðsn Þ ; ðs s1 Þðs s2 Þ:::ðs sn Þ (

Hc1 ðsÞjs¼0 ¼

1=

0;

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 þ e2 ;

n even n odd:

n odd

> > > > > ;

(7:3:28)

Example 7.3.2 Find the Chebyshev 1 transfer function that has X ¼2 dB ripple in the pass band and a minimum attenuation in the stop band of Y ¼15 dB.

Chebyshev-1 low-pass transfer functions: 8 9 1 ðs1 Þðs2 Þ. ..ðsn Þ > > > > p ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ﬃ ; n even > > > > > > < 1 þ e2 ðs s1 Þðs s2 Þ. ..ðs sn Þ =

(7:3:26)

n oc

(7:3:29)

This is the same as in Example 7.3.1. The value of n is determined from ;

(7:3:25)

"

# 1 10 log jo¼or ¼1:69196oc ¼ 15 dB 1 þ e2 C2n ðooc Þ ) Cn ð1:69196Þ ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ð101: 5 1Þ=ð10:2 1Þ

¼ 7:2358:

(7:3:30)

260

7 Approximations and Filter Circuits

Cn ð1:69196Þ ¼ coshðn cosh1 ð1:69196ÞÞ ¼ 7:2358; n

cosh1 ð7:2358Þ ¼ 2:387 cosh1 ð1:69196Þ

transformation that takes the zero frequency to 1 and vice versa, which is referred to as low-pass : to high-pass transformation. Consider the Chebyshev 1 function with oc ¼ 1 in the form jHc1 ðjuÞj2 ¼

It follows that n ¼ 3 since n must be an integer. The maximum attenuation in the stop band and the transfer function can be determined. These are

1 ¼ 20:81dB 10 log 1 þ e2 C3 ð1:69196Þ

(7:3:31)

Hc1 ðsÞ ¼ s0;2

s1 ¼ oc ð0:36891Þ:

(7:3:32)

Figure 7.3.6 shows the specifications and the & derived amplitude response.

1þ

:

(7:3:33a)

The low-pass to high-pass transformation n ! ð1=nÞ translates the ripples in the pass band to the stop band in the region n41 and monotonic response in the region of n 1. jHc1 ðj=nÞj2 ¼

ðs0 Þðs1 Þðs2 Þ ; ðs s0 Þðs s1 Þðs s2 Þ ¼ oc ð0:184445+j0:92078Þ;

1 e2 C2n ðnÞ

1 : 1 þ e2 C2n ð1=nÞ

(7:3:33b)

The transformation translates the ripples in the pass band to the stop band in the region n41 and gives monotonic response in the region of n 1. The lowpass function is jHC2 ðjnÞj2 ¼ 1 jHC1 ðj=nÞj2 ¼ 1 ¼

1 1 þ e2 C2n ð1=nÞ

e2 C2n ð1=nÞ 1 þ e2 C2n ð1=nÞ

(7:3:34)

For n ¼ 2 and 3, C2 ðaÞ ¼ 2a2 1 and C3 ðaÞ ¼ 4a3 3a: With a ¼ 1=n, we have C2 ð1=nÞ ¼ 2ð1=nÞ2 1 ¼ ð2 n2 Þ=n2 ; Fig. 7.3.6 Example 7.3.2

Chebyshev 1 response has equal ripple property and is monotonic in the stop band. It has a steep transition compared to the Butterworth approximation. Also, the phase response of the Butterworth filter is better with good delay properties. The idea is to find a maximally flat pass-band response to improve the delay performance and retain the steep transition like the Chebyshev 1. Such a case is Chebyshev 2 approximation and is discussed below. Chebyshev 2 or inverse Chebyshev approximation: The Chebyshev 2 approximation function can be derived from Chebyshev 1 function by first considering the normalized Chebyshev 1 function with the cut-off frequency of one and a

C3 ð1=nÞ ¼ 4ð1=nÞ3 3ð1=nÞ ¼ ð4 3n2 Þ=n3 : Therefore,

2 2 n2 e n2 Hc2;2 ðjnÞ 2 ¼

2 2 n2 1þ e n2 2 e ð4 4n2 þ n4 Þ ¼ 2 ½e ð4 4n2 þ n4 Þ þ n4 Hc2;3 ðjnÞ 2 ¼

e2 ð16 24n2 þ 9n4 Þ : ½e2 ð16 24n2 þ 9n4 Þ þ n6

(7:3:35a)

(7:3:35b)

The responses have the maximally flat property at n ¼ 0. They are monotonic in the range 0 n 1

7.3 Classical Analog Filter Functions

261

Fig. 7.3.7 Chebyshev 2 low-pass amplitude response (a) n ¼ 2, (b) n ¼ 3

(a) and have ripples in the range 15n51. The two functions are sketched in Fig. 7.3.7a,b. e can be computed from jHc2 ðjnÞj2n¼1 ¼

1. n

e2 C2n ð1=nÞ n!1 1 þ e2 C2 ð1=nÞ n

¼

0; n odd 2 : e =ð1 þ e2 Þ ; n even ð7:3:36Þ

The Chebyshev 2 normalized transfer function can be computed from Hc2 ð^ sÞHc2 ð^ sÞ ¼ jHc2 ðjnÞj2 n¼s=j :

(7:3:38a)

2. eI ¼ ½1=eCn ðor =oc Þ

3. ^sn ¼ sinh 1 sinh1 1 sin 2n þ 1p n eI 2n

1 1 2n þ1 þ j cosh sinh1 cos p; n eI 2n n ¼ 0; 1; :::; n 1

lim jHc2 ðjnÞj ¼ lim

the stop-band and the pass-band frequencies or and oc in terms of e computed from the passband edge specification. cosh1 ½ð10:1Y 1Þ=ð10:1X 1Þ cosh1 ðor =oc Þ (Same as in Chebyshev 1:Þ

e2 1 : ¼ 1 þ e2 1 þ ð1=e2 Þ

The term ð1=e2 Þ is related to the ripple width in the stop band similar to the term e2 used in determining the ripple width in the pass band in the Chebyshev 1. Note jHc2 ð0Þj ¼ 1 for all values of n, C2n ð0Þ ¼ 0 for n–odd and C2n ð0Þ=1 for n–even. Therefore,

n!1

(b)

4. Poles: sn ¼ 1=^ sn

(7:3:38c)

(7:3:38d)

5. Zeros: z^m ¼ j secðð2m þ 1Þp=2nÞ; m ¼ 0; 1; :::; ðn 1Þ=2; n odd : m ¼ 0; 1; :::; ðn=2Þ 1; n even (7:3:38e)

(7:3:37)

For the derivation of the Chebyshev 2 function, see Weinberg (1962). A summary is given below. Chebyshev 2 function has a maximally flat response in the pass band as in the Butterworth approximation. Note Hc2 ðjoÞ o¼0 ¼ 1. The left half-plane poles sn and the zeros zm of the Chebyshev 2 function normalized to the frequency 1 are determined using the constant eI obtained from the edges of

The function Cn2 is obtained by substituting s^ ¼ s=or in the normalized function. Example 7.3.3 Find the Chebyshev 2 transfer function that has attenuation of X ¼ 2 dB at the edge of the pass band and the minimum attenuation in the stop band of Y ¼15 dB. Note the pass-band and stop-band specifications are the same as in Example 7.3.2.

262

7 Approximations and Filter Circuits

Solution: From Example 7.3.2, e ¼ :76478 and n ¼ 3. From (7.3.38e), the zeros of the transfer function are z^1 ; z^ 1 ¼ +jð1:1547Þ: The other zero is at infinity. The constant eI is eI ¼ ½1=eC3 ðor =oc Þ ¼ ð1=:76478Þð1=14:2998Þ ﬃ :09144:

7.4 Phase-Based Design A system is distortionless if its output is the same as the input except it is attenuated by the same amount for all frequencies with a constant delay (see Section 6.11). The transfer function of a linear time-invariant (LTI) system is given by

Using (7.3.38d), the poles are ^ s0;2 ¼ ð:6103+ j1:3665Þ; s^1 ¼ ð1:2206Þ. The normalized transfer function is ðs z1 Þðs z 1 Þ ; sÞ ¼ K Hc2 ð^ ðs s0 Þðs s1 Þðs s2 Þ ðs0 Þðs1 Þðs2 Þ ðnote Hc2 ð0Þ ¼ 1:Þ: K¼ ðz1 Þðz2 Þ The denormalized transfer function is obtained from

HðjoÞ ¼ jHðjoÞjejyðoÞ : If yðoÞ is linear, then yðoÞ ¼ ot:

(7:4:2)

Since linear phase analog filters are not realizable, they are approximated. The group delay and the phase delays were defined by (6.7.7) and (6.7.19).

Hc2 ðsÞ ¼ Hc2 ðsÞ s¼s=or : Tg ðoÞ ¼ The amplitude response function is sketched in & Fig. 7.3.8.

(7:4:1)

dyðoÞ ; do

Tp ðoÞ ¼

yðoÞ : o

(7:4:3)

Linear phase implies that the group delay in (7.4.3) is a constant. Since we are more interested in the phase angle, we can write (7.4.1) in the form below and solve for yðoÞ. ln½HðjoÞ ¼ lnjHðjoÞj þ jyðoÞ ¼ ð1=2Þ lnjHðjoÞj2 þjyðoÞ ¼ ð1=2Þ ln½HðjoÞHðjoÞþjyðoÞ: (7:4:4)

yðoÞ ¼ ð1=jÞ ln½HðjoÞ ð1=2jÞ ln½HðjoÞ Fig. 7.3.8 Amplitude response of the Chebyshev 2 transfer function in Example 7.3.3

Elliptic filter approximations: Elliptic filter functions have equal ripple in both bands. Elliptic functions are beyond the scope here (see Storer, 1957). For a given set of filter amplitude response specifications, the order of the filter for the Butterworth (nBu ), Chebyshev 1 and 2 (nC1 and nC2 ), and elliptic filters (nE ) satisfy (see Storer, (1957).): nBu nC1 ¼ nC2 nE :

(7:3:39)

ð1=2jÞ ln½HðjoÞ ¼ ð1=2jÞ ln½HðjoÞ ð1=2jÞ ln½HðjoÞ HðjoÞ : (7:4:5a) ¼ ðj=2Þ ln HðjoÞ The generalized phase function is defined by yðsÞ ¼ :5 ln½HðsÞ=HðsÞ: From (7.4.3), and using the chain rule given below, the group delay is given by

7.4 Phase-Based Design

263

expressed in terms of the transform variable s given below, which is useful in computing the delay associated with a transfer function and prime (0 ) denotes differentiation and

dyðoÞ do 1 dðln½HðjoÞÞ 1 dðln½HðjoÞÞ þ ¼ 2j do 2j do

Tg ðoÞ ¼

1 d ln½HðjoÞ 1 d ln½HðjoÞ þ : ¼ 2 dðjoÞ 2 dðjoÞ

Tg ðsÞ ¼

(7:4:5b)

Chain rule :

d ln½HðjoÞ d ln½HðjoÞ dðjoÞ d ln½HðjoÞ ¼ : ¼j do dðjoÞ do dðjoÞ

d ln½HðsÞ s¼jo : ds

ðTg ðsÞ ¼ Tg ðsÞÞ:

(7:4:9)

This results in the group delay that is real and an : even function of o.

Noting the complex-conjugate terms inside the brackets {.} in (7.4.5b), the group delay is d ln½HðjoÞ Tg ðoÞ ¼ Re dðjoÞ ¼ Ev

1 P0 ðsÞ P0 ðsÞ Q0 ðsÞ Q0 ðsÞ þ ; 2 PðsÞ PðsÞ QðsÞ QðsÞ

(7:4:6)

Notes: The symbol EvfXðsÞg is the even part of XðsÞ and is 1 EvfXðsÞg ¼ ½XðsÞ þ XðsÞ 2 Z1 Z1 1 1 st ¼ xðtÞe dt þ xðtÞest dt; 2 2 1 1 EvfXðsÞg s¼jo ¼½XðjoÞ þ X ðjoÞ=2 ¼ RefXðjoÞg: (7:4:7) & We should note that the group delay function is an even function. Assuming that the transfer function HðsÞ ¼ PðsÞ=QðsÞ is a ratio of two polynomials, we have d ln½HðsÞ d ln½PðsÞ d ln½QðsÞ P0 ðsÞ Q0 ðsÞ ¼ ¼ : ds ds ds PðsÞ QðsÞ The generalized phase and the group delay can be defined in terms of the variable s by 1 HðsÞ ; yðsÞ ¼ ln 2 HðsÞ dyðsÞ yðoÞ ¼ jyðsÞ s¼jo ! Tg ðsÞ ¼ ; ds Tg ðoÞ ¼ Tg ðsÞ s¼jo : (7:4:8) Using the property that ln½HðsÞ=HðsÞ ¼ ln½HðsÞ ln½HðsÞ the group delay can be

Example 7.4.1 Compute the generalized phase and the group delay functions for the transfer function HðsÞ ¼ 1=½as2 þ bs þ c. Solution: From (7.4.8), the generalized phase and group delays are as follows: Phase : yðsÞ ¼ ð1=2Þ ln½HðsÞ=HðsÞ; yðoÞ ¼ jyðsÞ s¼jo Group delay: PðsÞ ¼ 0; P 0 ðsÞ ¼ 0; P 0 ðsÞ ¼ 0; QðsÞ ¼ as2 þ bs þ c; QðsÞ ¼ as2 bs þ c d Q0 ðsÞ ¼ 2as þ b; Q0 ðsÞ ¼ QðsÞ ds d 2 ¼ ½as bs þ c ds ¼ ½2as b ¼ 2as þ b 1 2as þ b 2as þ b ) Tg ðsÞ ¼ 0 þ 0 2 2 2 as þ bs þ c as bs þ c ¼

bc abs2 ðas2 þ cÞ2 b2 s2

Tg ðoÞ ¼ Tg ðsÞ s¼jo ¼

:

(7:4:10a)

bc þ abo2 ðc ao2 Þ2 þ b2 o2

(function is real and even):

(7:4:10b)&

7.4.1 Maximally Flat Delay Approximation If we assume in (7.4.1) that jHðjoÞj ¼ 1, then HðjoÞ ¼ ejot0 ; jHðjoÞj ¼ 1; yðoÞ ¼ ﬀHðjoÞ ¼ ot0 ; tðoÞ ¼

dﬀHðoÞ ¼ t0 : do

(7:4:11)

264

7 Approximations and Filter Circuits

It has flat amplitude, linear phase, and a constant group delay with respect to o. The system described by (7.4.11) is distortionless. Using the analytic continuation(see Balbanian et al., 1969), the transfer function can be written in the Laplace transform domain by replacing jo by s. Let us define a normalized transfer function using HðsÞ ¼ HðsÞ s¼st0 ¼ est0 s¼st0 ¼ es :

(7:4:12)

1 s¼st 0 coshðsÞ þ sinhðsÞ 1 : (7:4:13) ¼ ½ðcoshðsÞ= sinhðsÞÞ þ 1 sinhðsÞ ¼

Storch (1954) approximates es by an nth order rational function of the form in (7.4.13) below using the power series approximation of the hyperbolic sine and cosine functions.

HðsÞ Hn ðsÞ ¼

b0 b0 ¼ ; Bn ðsÞ bn sn þ bn1 sn1 þ ::: þ b0

bn ¼ 1; Hn ð0Þ ¼ 1:

(7:4:14)

Bn ðsÞ are the Bessel polynomials and they can be derived using (see Spiegel, 1968.) Bn ðsÞ ¼ ð2n 1ÞBn1 ðsÞ þ s2 Bn2 ðsÞ; B0 ðsÞ ¼ 1; B1 ðsÞ ¼ 1 þ s:

(7:4:15)

For n ¼ 0; 1; 2; 3; 4, these are B0 ðsÞ ¼ 1;

B2 ðsÞ ¼ 3 þ 3s þ s2 ; B3 ðsÞ ¼ 15 þ 15s þ 6s2 þ s3 ; (7:4:16)

The roots of the polynomials can only be determined numerically. The transfer function Hn ðjoÞ

Hn ðjoÞ ¼

Hn ðjoÞ ¼ jHn ðoÞjejyn ðoÞ ; yn ðoÞ ¼ ﬀHn ðoÞ; tn ðoÞ ¼ dyn ðoÞ=do:

(7:4:17)

Example 7.4.2 Show the maximally flat property of the group delays of Hn ðsÞ; n ¼ 1; 2 . H1 ðsÞ ¼

1 3 ; H2 ðsÞ ¼ 2 : sþ1 s þ 3s þ 3

(7:4:18)

Solution: The phase and the delay responses are given by y1 ðoÞ ¼ tan1 ðoÞ; t1 ðoÞ ¼

dy1 dð tan1 ðoÞÞ 1 ¼ ¼ ; (7:4:19a) do do 1 þ o2

y2 ðoÞ ¼ tan1 t2 ðoÞ ¼

3o ; 3 o2

dy2 ð9 þ 3o2 Þ ¼ do ð9 þ 3o2 Þ þ o4

(7:4:19b)

Expressing these in terms of Maclaurin power series in the neighborhood of o ¼ 0 , we can show that the first ð2n 1Þ derivatives of the group delay function vanish at the zero frequency and the maximally flat property follows. This is valid & for all n.

7.4.2 Group Delay of Bessel Functions

B1 ðsÞ ¼ 1 þ s;

B4 ðsÞ ¼ 105 þ 105s þ 45s2 þ 10s3 þ s4

has maximally flat delay characteristics. The transfer function, the phase, and the group delay responses are given for the Bessel transfer function (in terms of frequency o) by

Baher (1990) gives a relationship between an all pole rational function and its group delay and is summarized below in terms of a Bessel transfer function Hn ðsÞ . First, we can write the transfer function (see (7.4.13)) in the form

b0 b0 ¼ : ðb0 b2 o2 þ b4 o4 :::Þ þ jðb1 o b3 o3 þ b5 o5 :::Þ En ðoÞ þ jOn ðoÞ

(7:4:20)

7.4 Phase-Based Design

265

The amplitude, phase, and the corresponding group delay responses are b20 ; E2n ðoÞ þ O2n ðoÞ On ðoÞ yn ðoÞ ¼ tan1 : En ðoÞ

"

#

n 2 o tn ðoÞ ¼ 1 þ ::: b0

jHn ðjoÞj2 ¼

" ﬃ 1 (7:4:21a)

dyn ðoÞ tn ðoÞ ¼ do 2 3 dOn ðoÞ dEn ðoÞ En ðoÞ do On ðoÞ do 5: ¼ 4 E2n ðoÞ þ O2n ðoÞ ) tn ðoÞ ¼ 1 o2n jHn ðjoÞj2 ð1=b20 Þ:

(7:4:21b)

Solution: These can be shown by n ¼ 1 : 1 o2 jH1 ðjoÞj2 ð1=b20 Þ o2 1 ¼ ¼ t1 ðoÞ 2 1þo 1 þ o2

(7:4:23a)

The amplitude response of a Bessel filter function is Gaussian. The attenuation for a filter of order n43, the attenuation and the 3 dB frequency can be approximated by

Example 7.4.3 Verify the results in (7.4.19a and b) using (7.4.21b).

¼1

! # ð2n n!Þ2 2n ðoÞ : ð2nÞ!

(7:4:22a)

20 logjHn ðjoÞj ﬃ 4:3429o2 =ð2n 1ÞÞ

& See Problem 7.4.5 for its use of this. Example 7.4.4 Determine a. the 3 dB frequency and b.the frequency at which the group delay deviates by 1% for a second-order Bessel function. 3 ; H2 ðsÞ ¼ 2 s þ 3s þ 3 9 : (7:4:24a) jH2 ðjoÞj2 ¼ 2 ½ð3 o Þ2 þ 9o2

Solution: a. It follows that jH2 ðjo3dB Þj2 ¼

n ¼ 2 : 1 o4 jH2 ðjoÞj2 ð1=b20 Þ o4 ð9 þ 3o2 Þ ¼ 2 4 ð9 þ 3o Þ þ o ð9 þ 3o2 Þ þ o4 ¼ t2 ðoÞ: ð7:4:22bÞ

1 9 ¼ 2 ½ð3 o23dB Þ2 þ 9o23dB

) o3dB ¼ 1:36:

¼1

Notes: Note Hn ð0Þ ¼ 1 and the group delay has the maximally flat response with tn ð0Þ ¼ 1: The design involves finding the n for a set of specifications including maximum attenuation in the pass band in dB and a constant delay within a prescribed tolerance in the pass band. The group delay can be approximated by using the first two terms in the series and the approximation is good for n43 (see Temes and Mitra, 1973). Assuming the frequency is normalized by t0 , that is o ¼ ot0 , we have

(7:4:23b)

(7:4:24b)

b. The frequency at which the group delay deviates is computed using (7.4.22b) o4:99 ðt2 ðoÞÞ:99 ¼ 1 9 þ 3o2:99 þ o4:99 ¼ :99 ) o:99 ¼ :56:

(7:4:24c)

&

In this example, the 3 dB frequency and the frequency at which certain percent deviation in tn ðoÞ from 1 can be analytically computed. For an arbitrary n, these can be computed either by (7.4.23) or by tables (see Weinberg, 1962.). Table 7.4.1 gives the

Table 7.4.1 Normalized frequencies, o ¼ ot0 . Time delay and a loss table giving the normalized frequency o at which the zero frequency delay and loss values deviate by specified amounts for Bessel filter functions n 1 2 3 4 5 6 7 8 9 10 11 o3dB o1%deviation o10%deviation

1 0.1 0.34

1.36 0.56 1.09

1.75 1.21 1.94

2.13 1.93 2.84

2.42 2.71 3.76

2.70 3.52 4.69

2.95 4.36 5.64

3.17 5.22 6.59

3.39 6.08 7.55

3.58 6.96 8.52

3.77 7.85 9.48

266

7 Approximations and Filter Circuits

Fig. 7.4.2 Example 7.4.5: (a) amplitude and (b) group delay response specifications

(a) normalized 3 dB frequency and the frequencies at which the tðoÞ deviates 1 and 10% from 1. Compare the results in (7.4.24b and c) to the table. Example 7.4.5 Find n for the Bessel filter specifications in Fig. 7.4.2a (for the amplitude) and Fig. 7.4.2b (for the delay) with 1. A delay of t0 =.25 ms up to 1 MHz within 1% deviation 2. A loss of less than 3 dB up to 1 MHz. Solution: From the specifications, the pass-band edge of the normalized frequency is o3dB ¼ o3dB t0 o¼2pð106 Þ ¼ 2pð106 Þð:25ð106 ÞÞ ﬃ 1:57: (7:4:25) From the first condition, using Table 7.4.1, we have n 4. To satisfy the second condition, again using Table 7.4.1, n must be at least equal to 3, as 1.36 < 2; n ¼ 1 xðnts Þ ¼

Fig. 8.2.5 Interpolation using three sinc functions

1; n ¼ 2 : > : 0; otherwise

(8:2:20)

(8:2:19)

Evaluate the function yðtÞ at t ¼ :5ts ; ts ; 1:5ts ; 2ts using the interpolation formula in (8.2.17) and the sampled values of the function xðtÞ in (8.2.20). Solution: By using the interpolation formula and noting that sincðpfs tÞ is even, we have yðtÞ ¼ 2sincðpðfs t 1ÞÞ sincðpðfs t 2ÞÞ; (8:2:21) yðts Þ ¼ 2sincðpðfs ts 1ÞÞ sincðpðfs ts 2ÞÞ ¼ 2sincð0Þ sincðpÞ ¼ 2 ¼ xðts Þ; yð2ts Þ ¼ 2sincðð2ts fs 1ÞpÞ sincððfs ð2ts Þ 2ÞpÞ ¼ 1 ¼ xð2ts Þ;

8.2 Sampling of a Signal

317

Fig. 8.2.6 Example 8.2.2

The constant kn ¼ ð1=fs Þ is the energy contained in each of the sinc functions. To show this, consider the two functions and their transforms given by 1 o joðnts Þ FT e x1 ðtÞ ¼ sincðpfs ðt nts ÞÞ ! P fs 2pfs

yðts =2Þ ¼ 2sincðpðfs ðts =2Þ 1ÞÞ sincðpðfs ðts =2Þ 2ÞÞ ¼ 2sincðp=2Þ sincð3p=2Þ ¼ 1:2732 ð:212207Þ ﬃ 1:4854 yð1:5ts Þ ¼ 2sincðp=2Þ sincðp=2Þ ﬃ :637: See figure 8.2.6

¼ X1 ðjoÞ; &

Notes: The interpolated function yðtÞ ¼ 0 for t ¼ kts , k is an integer and k 6¼ 1; 2. The interpolating function yðtÞ has oscillating tails that die out and is shown in Fig. 8.2.6. xðtÞ is band limited to B0 ¼ fs =2 and therefore it cannot be time limited as the product of the spectral width and the time duration of the function cannot be less than a certain minimum value. See the uncertainty principle in & Fourier analysis in Section 4.7.3.

x2 ðtÞ ¼ sincðpfs ðt mts ÞÞ ¼ X2 ðjoÞ:

(8:2:24b)

Using generalized Parseval’s theorem assuming n 6¼ m with os ts ¼ 2p and using the transforms of the sinc functions, we have 1 1 ð ð 1 x1 ðtÞx2 ðtÞdt ¼ X1 ðjoÞX2 ðjoÞdo 2p 1

8.2.3 Interpolation Formula and the Generalized Fourier Series The interpolation formula is the generalized Fourier series expansion (see Section 3.3) with the orthogonal basis function set consisting of sinc functions. fsincðpðfs tnÞÞ; n¼2;1;0;1;2;...g:;15t51:

(8:2:24a) 1 o joðmts Þ FT e ! P fs 2pfs

1

2 ð1 1 o ¼ P ejoðnmÞts do; 2pfs 2pðfs Þ2 1 ð os =2 1 ejoðnmÞts do ¼ 2 2pðfs Þ os =2 joðnmÞts o¼os =2 1 e ¼ ; 2 ðn mÞt s o¼os =2 2pðfs Þ h i 1 jðnmÞp jðnmÞp e e ¼ 0: ¼ 2pðfs Þ2 ðn mÞts (8:2:25)

(8:2:22) For n ¼ m;

In the first step, the set in (8.2.22) is shown to be an orthogonal basis set over the interval 15t51. That is, ð1 sincðpðfs t nÞÞsincðpðfs t mÞÞdt 1 kn ¼ ð1=fs Þ; n ¼ m : (8:2:23) ¼ 0; n 6¼ m

ð1 1

2 ð1 o P ejoðnmÞts do 2pfs ð2pfs Þ2 1 ð os =2 1 1 ¼ do ¼ : (8:2:26) fs 2pðfs Þ2 os =2

x1 ðtÞx2 ðtÞdt ¼

1

From (8.2.25) and (8.2.26), it follows that the set in (8.2.22) is an orthogonal basis set. Therefore, the generalized Fourier series expansion of yðtÞ is

318

8 Discrete-Time Signals and Their Fourier Transforms

yðtÞ ¼

1 X

Ys ½ksincðpfs ðt kts ÞÞ:

(8:2:27)

k¼1

The generalized Fourier series coefficients can be determined from ð1 Ys ½k ¼ fs yðtÞsincðpfs ðt kts ÞÞdt: (8:2:28) 1

Noting the transform of the sinc pulse is a rectangular pulse (see (4.3.28)) and the generalized Parseval’s theorem (see (8.2.25)) and F½yðtÞ ¼ F½xðtÞ ¼ XðjoÞ results in the following: sincðpfs ðtnts ÞÞ

( sinðpfs ðtnts ÞÞ FT f1s enots ; joj5 o2s ; ¼ ! pfs ðtnts Þ 0; otherwise ð1 )Ys ½k¼fs xðtÞsincðpðfs tnÞÞdt ð os =2

(8:2:29)

1

1 XðjoÞejonts do; 2p os =2 ð1 fs XðjoÞejot dojt¼kts ¼xðtÞjt¼kts ¼xðkts Þ: ¼ 2pðfs Þ 1 (8:2:30)

¼

Since the transform of the function xðtÞ is band limited to os =2, the limits on the transform integral can be changed from ððos =2Þ; ðos =2ÞÞ to ð1; 1Þ in (8.2.30). Example 8.2.3 Let xðtÞ ¼ sinð2pð1ÞtÞ shown in Fig. 8.2.7. It is sampled at the Nyquist rate of fN ¼ 2ð1Þ ¼ 2 samples per second and sampled at t ¼ 0; :5; 1; 1:5; . . .. The sampled values are equal to zero indicating that the signal cannot be recovered from the samples. Nyquist theorem does not identify & where to sample. The sampling rate has to be larger than the Nyquist rate. Its selection is signal dependent and cost-

effectiveness, as the analog-to-digital (A/D) converters are expensive at both the low and the high sampling rate. Sampling a function at much higher than the Nyquist rate does not help. Recovering an analog signal from the samples requires the computation using more samples than necessary and the errors in computation nullifies any advantage used in high sampling. As a guide, the sampling rate is more than the Nyquist rate, about 2.5–10 times the highest frequency in the signal. For seismic signals, the frequencies of interest are in few hundred Hertz range. In these cases, higher sampling rates are used. For speech, the frequency range of interest is from a few Hertz to 3.5 kHz. The sampling rate is taken as 8 kHz or 10 kHz. For CDs, the frequency range of the input signal is from a low frequency of few Hertz to 20 kHz. The sampling rate is taken as 44.1 kHz and the standard sampling rate for studio quality audio is 48 kHz. The compact disc recording system samples each of the two stereo signals with a 16-bit A/D converter at 44.1 kHz (Haykin and Van Veen (2003)). Example 8.2.4 Consider signal xðtÞ band limited to ð2pBÞ rad/s. Determine the Nyquist rates for the functions: a: y1 ðtÞ ¼ xð2tÞ; b: y2 ðtÞ ¼ xðtÞ cosðo0 tÞ Solution: a. Note y1 ðtÞ is formed from xðtÞ by compressing the time axis by a factor of 2. From the Fourier scale change theorem (see Section 4.3.4), we have the following: FT 1 y1 ðtÞ ¼ xðatÞ$ Xðjo=aÞ; a 6¼ 0 j aj FT 1 Xðjo=2Þ ¼ Y1 ðjoÞ: ) xð2tÞ$ 2

Time compression by a factor of 2 results in expansion in frequency by a factor of 2. The Nyquist rate is given by os1 ¼ 2pð2ð2BÞÞ. It is like playing an audio tape fast. b. The signal is a modulated signal with a center frequency o0 with a bandwidth of 2B Hz and F½y2 ðtÞ ¼ F½xðtÞ cosðo0 tÞ ¼ 0:5Xðjðo o0 ÞÞ þ 0:5Xðjðo þ o0 ÞÞ: The highest frequency in the modulated signal is o0 þ 2pB ¼ ðo0 þ os Þ=2. The Nyquist rate is & os2 ¼ os þ 2o0 .

Fig. 8.2.7 xðtÞ ¼ sin 2pð1Þt, Sampled two times per second

The sinc interpolation function is not the best way to approximate the function from its sample

8.2 Sampling of a Signal

319

values, as it decays only at a rate of ð1=tÞ. There are other better functions. The function xðtÞ is known at t ¼ nts . The interpolation formula can be expressed in terms of a function hi ðtÞ that is 0 at all the sampling instants, except at t ¼ 0, where it is 1. In addition, it is absolutely integrable. Interpolation formula is given by yi ðtÞ ¼

1 X

xðnts Þhi ðt nts Þ; hi ðkts nts Þ

n¼1

¼

1; k ¼ n 0; k 6¼ n

:

Since yi ðtÞjt¼kts ¼ xðkts Þ, i.e., the interpolation formula gives the same values at the sampling instants and, at other times, yi ðtÞ is an approximation of xðtÞ. Most commonly interpolating functions are step, linear, sinc, and raised cosine functions. These are given below in table 8.2.1. See Ambardar (1999) for additional discussion on the interpolation functions. Step interpolation (zero-order-hold) uses a rectangular interpolation function and xðnts Þ to produce a stepwise or a staircase approximation of xðtÞ. This is simple and does not depend on the future values of the signal. It is widely used. The reconstructed signal is (more on this in Section 8.2.5.)

It cannot be implemented online since a future value is required. Sinc interpolation was considered earlier. Raised cosine interpolation function (see (4.11.9a) for the function and its transform in (4.11.9b)) uses the roll-off factor b. It reduces to the sinc interpolation function when b ¼ 0. The raised cosine function’s decaying rate is proportional to ð1=t3 Þ. Faster decaying results in improved reconstruction, if the samples are not at exactly at the sampling instants (i.e., jitter). It requires fewer past values are needed in the reconstruction. Polynomial-based interpolation methods are discussed in Appendix A.9.

8.2.4 Problems Associated with Sampling Below the Nyquist Rate Consider the functions x1 ðtÞ and x2 ðtÞ in Fig. 8.2.8. They are sampled at a rate shown. Both provide the same sample values. The function x1 ðtÞ cannot be reconstructed from the sample values. From the figure, x1 ðtÞ has a higher frequency content than x2 ðtÞ. By sampling the functions at the locations shown, some of

yc ðtÞ ¼ xðnts Þ; nts t5xððn þ 1Þts Þ: Linear interpolation (first-order hold) uses a linear approximation and the reconstructed signal is yl ðtÞ ¼ xðnts Þ þ

xððn þ 1Þts Þ xðnts Þ ts

ðt nts Þ; nts t5ðn þ 1Þts :

Fig. 8.2.8 Two signals sampled at the same locations

Table 8.2.1 Common interpolation functions 1 P xðnts ÞP½ðt nts Þ=ts : constant or step interpolation a. yc ðtÞ ¼ n¼1

b. yl ðtÞ ¼

1 P

xðnts ÞL½ðt nts Þ=ts : linear interpolation

n¼1

c. ys ðtÞ ¼

1 P

xðnts Þsinc½pðt nts Þ=ts : sinc interpolation

n¼1

d. yrc ðtÞ ¼

1 X n¼1

xðnts Þ

cosðpbðt nts Þ=ts Þ ð1 ½2bðt nts Þ=ts 2 Þ

sincðpðt nts Þ=ts Þ: raised cosine interpolation, b roll-off factor; 0 b 1.

320

8 Discrete-Time Signals and Their Fourier Transforms

the peaks and valleys of x1 ðtÞ are missed indicating that x1 ðtÞ needs to be sampled at a higher rate than x2 ðtÞ: A spectrum XðjoÞ, its ideally sampled signal spectra by assuming the sampling rates of os1 2pð2BÞ and os2 52pð2BÞ are shown in Fig. 8.2.9a,b,c. In the case of os1 , the message signal can be recovered. In the second case of os2 , the sampling rate is lower than the Nyquist rate. The resultant spectra of the ideally sampled signal will have overlaps of the adjoining spectra and the spectral components are added around half the sampling frequency and the message signal cannot be recovered by low-pass filtering the ideally sampled signal and the filtered signal will be distorted. The distortion caused by sampling below the Nyquist rate is called aliasing. Most signals are not band limited. Therefore, a band limiter is necessary before sampling to minimize the aliasing errors. Example 8.2.5 Consider the amplitude spectrum of a function xðtÞ given by 2

jXðjoÞj ¼

2oc 2

ðoÞ þ o2c

:

(8:2:31)

The signal is sampled at the sampling frequency os and an ideal low-pass filter is used to recover it from

the sampled signal. Use MATLAB to quantify the effect of loosing the spectral energy outside of half the sampling frequency. For MATLAB, see Appendix B. Solution: The energy contained in the signal is ð ð 1 1 2oc 1 1 do E¼ jXðjoÞj2 do ¼ 2p 1 o2 þ o2c 2p 1 2oc 1 tan1 ðo=oc Þ1 ¼ 1 ¼ p ¼ 1: 2poc p (8:2:32) The signal is a low-frequency signal, as most of the spectral energy is concentrated around f ¼ 0. It is not band limited. If it is filtered using an ideal low-pass filter with a cut-off frequency equal to half the sampling frequency ðos =2Þ, then some information is lost and the loss can be measured using the spectral energy contained in the frequency range joj4os =2 and 1 Error ¼ p

ð1

1 jXðjoÞj do ¼ p os =2 2

os=2

2oc ðoÞ2 þ o2c

do

2 ¼ 1 tan1 ðos =2oc Þ: p (8:2:33)

(a)

(b)

Fig. 8.2.9 (a) XðjoÞ, (b) Xs ðjoÞ; os1 > 2pð2BÞ (sampling rate higher than the Nyquist rate), and (c) Xs ðjoÞ; os2 52pð2BÞ (sampling rate lower than the Nyquist rate)

ð1

(c)

8.2 Sampling of a Signal

321

Fig. 8.2.10 Example 8.2.6

Mean-square error 1 0.9

mean-square error

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

0

0.5

If the sampling rate goes to infinity, i.e., os ! 1, then the error goes to zero as the area under the integral goes to zero. For the case of os ¼ 2oc the error is 0.5 or 50%. The error slowly goes down as we increase the value of the sampling rate. A simple MATLAB routine and a sketch of the mean squared error as a function of the ratio of sampling frequency divided by 2fc is given in Fig. 8.2.10. Note that 2fc is not the Nyquist rate since the spectrum is not band limited to fc Hz. For the two cases fs =2fc ¼ 2 and 4, the errors can be calculated and are 0.2952 and 0.1560, respectively. In the above example we need to have a high enough sampling rate to reduce the mean squared error. We use a pre-sampling filter that band limits the signal allowing for a decrease in sampling rate. Such a filter passes frequency components that are below the frequency os =2 and attenuates significantly or even suppress some of the frequency components above os =2: The bandlimiting filter is referred to as an anti-aliasing filter. Even if the signal is band limited to os =2, an antialiasing filter is generally used to avoid aliasing that may result from noise that is ever present in almost all signals. The anti-aliasing filter may not be shown explicitly and is assumed to be included in the system. The bandwidth of the pre-sampling or anti-

1

1.5

2 2.5 3 3.5 sampling frequency/2fc

4

4.5

5

aliasing filter is signal dependent. In simple words, we state that most of the signal energy is contained within the bandwidth B Hz and the energy contained outside this band is negligible. See Section & 4.7 for bandwidth measures. Most of the practical are low-pass signals that have decaying frequency response. One way to look at the aliasing error is to put a limit on the maximum aliasing error at half the sampling frequency which depends on the bandwidth of the signal. This works out nicely for signals that have decaying frequency response. The maximum error occurs at half the sampling frequency. See Spilker (1977), Ambardar (1995) and others. Another simple method is select the essential bandwidth which is taken as the frequency where the spectrum of the signal xðtÞ given by XðjoÞ reduces to say 1% of its peak value. MATLAB Code for Fig. 8.2.10 x=0:.01:5; y=1-(2/pi)*atan(x); plot(x,y) Title (‘Mean-square error’) xlabel (‘sampling frequency/2fc’) ylabel (‘mean-square error’)

322

8 Discrete-Time Signals and Their Fourier Transforms FT

Example 8.2.6 Let xðtÞ ¼ ea t uðtÞ ! 1=ða þ joÞ ¼ XðjoÞ; and

XðjoÞ ¼

1 X k¼1

1 jXðjoÞj ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; Xð0Þ ¼ 1; o ¼ 2pf: 2 a þ o2 Noting that the maximum aliasing error occurs at o ¼ os =2, find the sampling frequency fs using the following methods: a. a ¼ 1. Maximum aliased magnitude is less than (1) 5% and (2) 1% of the peak value of the function jXðjoÞj. b. a ¼ 2. Use the bandwidth of XðjoÞ as the frequency at which the amplitude reduces to 1% of the peak value. Solution: a. (1). From the statement we have jXðos =2Þj :05jXð0Þj ¼ :05. Now 1 1 qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ! ½1 þ ðos =2Þ2 5400 20 2 1 þ ðos =2Þ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ) os ¼ 4ð399Þ ) fs 6:36 Hz:

(8:2:34)

For a proof and for additional discussion, see Mitra (2006). Some of the digital finite impulse response (FIR) filters are based upon frequency sampling. Discrete-time signal bandwidth: The spectrum of an ideally sampled waveform xs ðtÞ (see 8.2.2) is periodic with period os and the measures of BW used for the continuous signals with nonperiodic spectrum cannot be used here. The ideally sampled signal is uniquely specified for frequencies in the range 0 to fs =2. The bandwidth of xs ðtÞ is the range of positive frequencies within the range 0fs =2, for which the amplitude spectrum is greater than or equal to a times its maximum value, where a is a constant less than 1. The common one is pﬃﬃﬃ a ¼ 1= 2 corresponding to the 3 dB bandwidth.

Flat top sampling uses a sample and hold device illustrated in Fig. 8.2.11 with the input is assumed to be 1 X FT xs ðtÞ ¼ xðnts Þdðt nts Þ ! Xs ðjoÞ: (8:2:35) n¼1

&

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ b: jXðjoÞj ¼ 1= o2 þ 4 ) 1% of jXðj 0Þj ¼ ð1=2Þ ð:01Þ ¼ :005. For o 2, jXðjoÞj 1=o and jXðjoÞj 1=2pB ¼ :005 ) B ¼ ð100=pÞHz ) fs 2B ¼ 200=p 64 Hz:

jkp sinðoTN kpÞ : TN ðoTN kpÞ

8.2.5 Flat Top Sampling

(2). In this case, we have jXðos =2Þj :01 jXð0Þj ¼ :01. Correspondingly, 1 1 qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ! ½1 þ ðos =2Þ2 > 1000 100 2 1 þ ðos =2Þ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ) os ¼ ð4Þ999 ) fs 10:06 Hz

X

The output of the summer is xðtÞ xðt ts Þ. Let the input is dðtÞ. The output of the first block is ½dðtÞ dðt ts Þ. The transfer functions of the first block and the integrator are H1 ðjoÞ ¼ F½dðtÞ dðt ts Þ ¼ ½1 ejots ; H2 ðjoÞ ¼ 1=jo:

&

Notes: For most signals, the actual spectrum may not be available and experimental methods may be needed to determine the bandwidth of & these. Frequency sampling theorem: Since the Fourier transform and its inverse are related so closely, this theorem follows naturally the timesampling theorem. Consider a time-limited function such that xðtÞ ¼ 0; jtj > TN . It has a Fourier transform that can be uniquely determined from samples at frequency intervals of np=TN and

Assuming the input to the [dðtÞ dðt ts Þ], the output is

hðtÞ ¼

ðt

integrator

is

½dðbÞ dðb ts Þdb

1

¼

(8:2:36)

1; 05t5ts 0; otherwise

¼P

t ðts =2Þ : (8:2:37) ts

The system response to an impulse input is a rectangular pulse of width ts s. This operation is the

8.2 Sampling of a Signal

323

Fig. 8.2.11 Zero-order-hold (a) Use of a delay component, a summer and the process of integration, (b) representation of these using block diagrams

(a)

(b) zero-order-hold (ZoH) referred to in the last section. The function yðtÞ is approximated in the interval nts t5ðn þ 1Þts using the first term in the Taylor series. 1 yðtÞ ¼ yðnts Þ þ y0 ðnts Þðt nts Þ þ y00 2!

(8:2:38)

jH0 ðjoÞj ¼ ts jsincðots =2Þj; ﬀH0 ðjoÞ ðots =2Þ; sincðos t=2Þ > 0 ¼ ðots =2Þ p; sincðos t=2Þ50 dﬀH0 ðjoÞ ts ; ¼ Group delay ¼ 2XðjoÞ do YðjoÞ ¼ H0 ðjoÞXs ðjoÞ: (8:2:42)

2

ðnts Þðt nts Þ þ ; nts t5ðn þ 1Þts ; ZoH introduces three modifications: yðtÞ yðnts Þ; nts t5ðn þ 1Þts :

(8:2:39)

It is called the zero-order-hold since the function is approximated by the constant term in the Taylor series. It can be approximated by the first two terms in the series. That is, yðtÞ yðnts Þ þ y0 ðnts Þðt nts Þ; nts t5ðn þ 1Þts ; y0 ðnts Þ ¼

dyðtÞ jt¼nts : dt

(8:2:40)

Most systems use ZoH. The transfer function of the cascaded blocks in Fig. 8.2.11 is H0 ðjoÞ ¼ H1 ðjoÞH2 ðjoÞ ¼ ½1 ejots =jo;

¼ ts

ejots =2 ejots =2 1 ejots =2 2j ots =2

¼ ts sincðots =2Þejots =2 :

(8:2:41)

The amplitude, phase, and the delay frequency responses are

1. A linear phase shift corresponding to a time delay of ðts =2Þs. 2. The input transform is band limited to om . The output transform YðjoÞ is a distorted version of XðjoÞ affected by the curvature of the main lobe of H0 ðjoÞ. 3. Transform of the ideally sampled signal is periodic. The envelope of the sinc function is inversely proportional to the frequency. The output transform contains distorted and attenuated versions of the images of XðjoÞ, centered at nonzero multiples of os . The first one follows since each sample is held constant for ts seconds. The last two effects are caused by constant or step interpolation resulting in high frequency components. The effects of the first two items can be reduced by increasing the sampling rate. If the effects of the last two items are not acceptable, then a continuous-time compensation filter (an anti-imaging filter) in cascade with the zero-order hold is needed. The transfer function of such is given in (8.2.43). See Haykin and Van Veen (2003). 8 ots os < ; joj5om 5 2 : (8:2:43) Hc ðjoÞ 2 sinðots =2Þ : 0; joj4os om

324

8 Discrete-Time Signals and Their Fourier Transforms

Notes: The output of an ideal sampler and a ‘‘flat top’’ sampler are xs ðtÞ ¼

1 X

xðnts Þdðt nts Þ;

(8:2:44a)

n¼1

t ts =2 nts : (8:2:44b) xðnts ÞP xflat top ðtÞ ¼ ts n¼1 1 X

An ideal low-pass filter is needed to recover xðtÞ from xs ðtÞ. On the other hand, an anti-imaging filter (see (8.2.43)) is necessary to recover xðtÞ from xflat top ðtÞ. In both cases, we assumed that fs > 2B; B ¼ bandwidth of the signal in Hertz. The digital-to-analog (D/A) converter with a ZoH takes the sequence x½n and creates a signal xflat top ðtÞ. So far, the discussion was centered on the lowpass signals. Can a sampling rate less than the twice the highest frequency of a band-pass signal and still recover the input signal? The answer is yes and is illustrated below. See Ziemer and Tranter (2002).

8.2.6 Uniform Band-Pass Sampling Theorem Given a signal xðtÞ

FT

!

XðjoÞ with jfjfl : jXðjoÞj¼0;o¼2pf; jfjfu ½bandwidth¼B¼ðfu fl ÞHz:

Fig. 8.2.12 (a) Band-pass spectra, (b) spectra of the ideally sampled signal

(8:2:45)

The signal xðtÞ can be recovered from the sampled signal if the sampling rate is fs ¼ ð2fu =mÞ, where m the largest integer that is not exceeding ðfu =BÞ. All higher sampling rates are not necessarily usable unless they exceed the Nyquist rate of 2fu . Example 8.2.7 Consider the band-pass signal spectrum XðjoÞ shown in Fig. 8.2.12a. with fl ¼ 4 kHz and fu ¼ 5 kHz(bandwidth is 1 kHzÞ: Using the band pass sampling theorem, sketch the ideally sampled signal spectrum assuming a sampling rate that allows for the recovery of the original signal from the sampled signal. Solution: Note fs ¼ 2fu =m; m ¼ Integer part ofðfu =BÞ ¼ 5; fs ¼ 2ð5Þ=5 ¼ 2 kHz;

Xs ðjoÞ ¼

1 os X Xðjðo kos ÞÞ: 2p k¼1

(8:2:46a)

(8:2:46b)

The band-pass signal can be recovered by noting that none of the spectra Xðjðo kos ÞÞ; k 6¼ 0 over laps the spectrum of the continuous signal xðtÞ. The spectra in (8.2.46b), Xðjðo kos ÞÞ are shifted to the right by kos for k positive and to the left by kos for k negative. In Table 8.2.2, the frequency ranges of the terms in (8.2.46b) for n ¼ 0; 1; 2; 3 and the center frequencies of the corresponding spectra are given. For example Xðjðo os ÞÞ is centered at 2 kHz and occupies the frequency range ð3 kHz5f5 2 kHz;

8.3 Basic Discrete-Time (DT) Signals

325

Table 8.2.2 Spectral occupancy of Xðjðo nos ÞÞ; o ¼ 2pf; n ¼ 0; 1; 2; 3 Spectra Frequency ranges (f, kHz) (nfs, kHz) X(jo) X(j(o – os)) X(j(o + os)) X(j(o – 2os)) X(j(o +2os)) X(j(o – 3os)) Xj((o +3os))

–5 < f M:

(8:5:3)

Using (8.5.3), the second equation in (8.5.2) can be expressed in terms of x½n as

&

Xds ½k ¼

8.5 Discrete-Time Fourier Transforms Computation of the continuous F- and the inverse F-transforms involves integrals and analytical computation is possible only in a few cases. Also, the signal may be available in the form of a waveform instead of an analytical expression or in terms of a sequence. The discrete-time Fourier transform of a sequence is derived from the discrete F-series by taking the number of sample points in the discrete F-series to infinity. Presentation is intuitive and follows that of Haykin and Van Veen (2003).

8.5.1 Discrete-Time Fourier Transforms (DTFTs)

O0 ¼

That is, one period of the periodic sequence is extracted and then it is padded with zeros outside of the period. In (8.5.1), as M increases, the periodic replicates of x½n move further and further away from the origin. The discrete-time F-series representation of the periodic signal and the DFS coefficients are as follows: xs ½n ¼

M X

Xds ½kejkO0 n ;

2p : 2M þ 1

XðejO Þ ¼

1 X

x½nejnO ;

n¼1

¼

Xds ½k ¼

1 2p : xs ½nejO0 nk ; O0 ¼ 2 M þ 1 n¼M 2M þ 1 (8:5:2)

1 XðejO ÞjO¼kO0 2M þ 1

1 X 1 x½nejnkO0 : 2 M þ 1 n¼1

(8:5:5)

For real sequences, with O0 ¼ 2p=ð2 M þ 1Þ, we have M 1 X XðejkO0 ÞejkO0 n O0 : 2p k¼M

(8:5:6)

As M increases, the spacing between the harmonics in the discrete Fourier series decreases (see (8.5.4)). In the limit, as M ! 1, dO ¼ O0 and O ¼ kO0 is some value on the frequency axis. The summation becomes an integral and 1 x½n ¼ 2p

ðp

XðejO ÞejnO dO;

(8:5:7)

2pð ÞM ¼ p: 2M þ 1

(8:5:8)

p

lim ð ÞMO0 ¼ lim

M!1

M!1

The discrete-time Fourier transform (DTFT) and its inverse along with their symbolic relations are as follows:

k¼M M X

(8:5:4)

Now define a continuous periodic function of frequency with period equal to 2p, XðejO Þ; so that the scaled samples of this function are the discrete-time Fourier series.

xs ½n ¼ Let x½n be a non-periodic sequence obtained from a single period of the periodic sequence centered at the origin. xs ½n; jnj M x½n ¼ : (8:5:1) 0; j nj 4 M

1 X 1 x½nejO0 nk ; 2 M þ 1 n¼1

XðejO Þ ¼ 1 2p

ðp p

1 X

x½nejnO ;

n¼1

XðejO ÞejnO dO ¼ x½n;

(8:5:9)

336

8 Discrete-Time Signals and Their Fourier Transforms

x½n

DTFT

! XðejO Þ ¼ XðejO ÞejyðOÞ ; Ffx½ng ¼ XðejO Þ:

XðejO Þ ¼ (8:5:10) ¼

jO

The transform Xðe Þ of a non-periodic sequence is the discrete-time Fourier spectrum or the spectrum of the sequence x½n. Some authors use XðOÞ or XðjOÞ instead of XðejO Þ. XðejO Þ is the preferred notation, as it explicitly shows that the spectrum is periodic and it will be used here. As in the continuous case, is in general complex. the DTFT The quantities XðejO Þ and yðOÞ ¼ ﬀXðejO Þ are the amplitude (or magnitude) and phase (or angle) spectra of the sequence x½n. The DTFT is valid for both real and complex sequences and has interesting properties, similar to the properties of the continuous F-transform. For a real sequence x½n, its amplitude spectrum is even and the phase spectrum is odd, which follows directly from the definition. A sufficient condition for the existence of XðejO Þ is the sequence x½n is absolutely summable. That is, 1 X

jx½nj51:

(8:5:11)

1 X n¼1 1 X

xe ½nejnO þ

n¼1

" ¼ x½0 þ 2

1 X

x0 ½nejnO ; (8:5:13)

n¼1 1 X

# xe ½n cosðnOÞ

n¼1

" þ ð1Þj2

1 X

# x0 ½n sinðnOÞ ¼ ReðXðejO ÞÞ

n¼1 jO

þ jImðXðe ÞÞ: Example 8.5.1 Find the DTFT of the following sequences xi ½n; i ¼ 1; 2. Sketch the responses for Part a. and identify the important values for a ¼ :8. a: x1 ½n ¼ an u½n; jaj51ðright-side sequenceÞ b: x2 ½n ¼ an u½n 1ðleft-side sequenceÞ: (8:5:14) Solution: Noting that u½n ¼ 0; n50 and u½n ¼ 1; n 0, and using (8.5.9), we have

n¼1

a: X1 ðejO Þ ¼ Note the unit step sequence does not satisfy (8.5.11).

x½nejnO

1 X

x1 ½nejnO ¼

n¼1

1 X

ðaejO Þn ¼

n¼0

1 ¼ : ð1acosðOÞÞþjasinðOÞ

8.5.2 Discrete-Time Fourier Transforms of Real Signals with Symmetries A real sequence x½n can be written in terms of its even and odd parts xe ½n and x0 ½n (see (8.3.9c)). Making use of the even and odd sequence properties, the DTFT of these can be written as xe ½n

DTFT

! xe ½0 þ 2

1 X

xe ½n cosðnOÞ;

(8:5:12a)

x0 ½n sinðnOÞ:

(8:5:12b)

n¼1

x0 ½n

DTFT

! j2

1 X n¼1

That is, if a real sequence is even, then its DTFT is real and even and if it odd, then its DTFT is pure imaginary and odd. The DTFT of an arbitrary real sequence x½n is

1 1aejO (8:5:15a)

The magnitude and the phase responses are periodic with period 2p and X1 ðejO Þ ¼

1

; ½1 þ a2 2a cosðOÞ1=2 a sinðOÞ jO 1 ﬀX1 ðe Þ ¼ tan : 1 a cosðOÞ

(8:5:15b)

Figure 8.5.1 gives these plots for one period, namely for 0 O52p. We could plot these for p O5p as they are periodic. The amplitude response is even and phase response is odd. The maximum and the minimum values of the amplitude response for a ¼ 0:8 can be seen by noting jcosðOÞj 1. The maximum and minimum values are located at O ¼ 2 kp and O ¼ ð2 k þ 1Þp. The values are given by 1=ð1 aÞ ¼ 5 and 1=ð1 þ aÞ ¼ 0:5555, respectively. The responses are plotted in Fig. 8.5.1 for 0 O52p. The phase response is a bit more complicated, as the arctangent functions is involved. First

8.5 Discrete-Time Fourier Transforms

337

Fig. 8.5.1 (a) X1 ðejO Þ, (b) ﬀX1 ðejO Þ

The group delay tðOÞ of a sequence x½n can be defined using its transform as follows:

ﬀX1 ðe jkp Þ ¼ 0; k ¼ 0; 1; 2; :::: A simple way to find the peak phase angle is by MATLAB. For a ¼ :8; the peak amplitudes and the phases are equal to 5 and 0.932 rad/s, respectively. b: X2 ðejO Þ ¼ a1 ejO a2 ej2O ¼

1 X

x2 ½nejnO

n¼1

¼ a1 ejO ½1 þ a1 ejO þ a2 ej2O þ a1 ejO 1 jO ; a e 51 ! X2 ðejO Þ ¼ ð1 a1 ejO Þ 1 ¼ ; jaj41: (8:5:15c) 1 aejO The transforms have the same forms, except constraints on the constant 0 a0 are different. find the time sequence from the transform, i.e. inverse transform, we need to know whether sequence is a right-side or a left-side sequence.

Fig. 8.5.2 Four sequences, (a) Type 1, (b) Type 2, (c) Type 3, (d) Type 4

the To the the &

x½n

DTFT

! XðejO Þ ¼ XðejO ÞejfðOÞ ; tðOÞ ¼ dfðOÞ=dO:

(8:5:16)

Since fðOÞ is periodic with period 2p, so is the group delay tðOÞ. Time limited sequences: Linear phase is an important property in the digital filter design. In the following we will consider the expressions for the DTFT of a time limited real sequences h½n ¼ 0; n4N; n50 that have the four conditions stated below. We further assume that the samples h½n have a. an even symmetry about the mid point of the sequence and b. an odd & symmetry about the mid point of the sequence. Sequences of interest (see Fig. 8.5.2): Type 1 sequence: N–odd: Sequence with an even symmetry over its mid point.

338

8 Discrete-Time Signals and Their Fourier Transforms

Type 2 sequence: N–even: Sequence with an even symmetry over its mid point. Type 3 sequence: N–odd: Sequence with an odd symmetry over its mid point. Type 4 sequence: N–even: Sequence with an odd symmetry over its mid point. 1. Sequence has an even symmetry if h½n ¼ h½N 1 n

implying the value of the sequence in the middle is zero. These result in h½n ¼ h½N 1 n and h½ðN 1Þ=2 ¼ 0: (8:5:20) Splitting the sequence h½n into fh½0; h½1; :::; h½ðN 3Þ=2g, h½ðN 1Þ=2 ¼ 0 and fh½ðNþ1Þ=2Þ; :::h½N 1g, the transform can be expressed as

(8:5:17)

2. Sequence has an odd symmetry if h½n ¼ h½N 1 n

HðejO Þ ¼

N 1 X

ðN3Þ=2 X

n¼0

(8:5:18)

We will consider Type 2 and 3 sequences and the other two are left as exercises.

h½nejOn :

HðejO Þ¼

h½nejnO þ

n¼0

¼

N1 X

ðN3Þ=2 X

h½nejOn ¼

h½nejnO þ

h½nejnO

n¼0

ðN=2Þ1 X

ðN3Þ=2 X

¼

h½nejOðN1nÞ :

n¼0

h½N1mejOðN1mÞ : Using this result in (8.5.21), we have

m¼0

ðN=2Þ1 X

n

o h½n ejnO þejnO ejðN1ÞO ;

n¼0

¼

h½mejOðN1mÞ

m¼0

n¼N=2

N=21 X

¼

N1 X

(8:5:21)

n¼ðNþ1Þ=2

n¼ðNþ1Þ=2 N=21 X

h½nejnO

n¼0 N 1 X

þ

Type 2 sequence: The even symmetry h½n ¼ h½N 1 n allows us to write

h½nejnO ¼

"ðN=2Þ1 X

HðejO Þ ¼

# 2h½ncosðððN1Þ=2ÞnÞO e

A2 ðe Þe

j2pððN1Þ=2Þ

:

h½nejnO

n¼0 jOðN1Þ=2

:

n¼0

jO

ðN3Þ=2 X

¼ 2j ¼ 2j

We have made use of the following in simplifying the above expression: n o ejnO þ ejnO ejðN1ÞO n o ¼ ejðN1ÞO=2 ej½ðN1Þ=2nO þ ej½ðN1Þ=2nO ¼ 2 cos½ðN 1Þ=2 nO: Note A2 ðejO Þ is real and the phase angle is fðN 1Þ=2gO, which is linear with respect to O. Also, ðN 1Þ=2 is not an integer since N is even. Type 3 sequences: The DTFT of this sequence can be determined by noting that N is odd and the sequence has an odd symmetry over its mid point

h½nejOðN1nÞ

n¼0

ðN3Þ=2 X n¼0

(8:5:19)

ðN3Þ=2 X

ðN3Þ=2 X

jOn e ejOðN1nÞ ; h½n 2j h½nejðN1Þ=2

n¼0 jO½ððN1Þ=2Þn

ejO½ððN1Þ=2Þn ; 2j " ðN3Þ=2 # X N1 ¼j 2 nÞO ejOðN1Þ=2 sin½ð 2 n¼0 e

¼ A3 ðejO ÞðjejOðN1Þ=2 Þ:

(8:5:22)

The phase angle ﬀ jejOðN1Þ=2 is linear. Also, A 3 ðejO Þ is real and odd symmetric about O ¼ 0 and O ¼ p and

jO

A3 ðe Þ ¼

ðN3Þ=2 X n¼0

sin

N1 n O : 2

(8:5:23)

8.6 Properties of the Discrete-Time Fourier Transforms

339

in terms of A i ðejO Þ and h i ½n can be determined jO & ( ) given Ai ðe Þ. ðN3Þ=2 X N1 N1 2h½ncos HðejO Þ¼ h þ n O 2 2 n¼0 Summary: Type 1 sequence:

ejOðN1Þ=2 ¼A1 ðejO ÞejOðN1Þ=2 :

(8:5:24)

A1 ðejO Þ : Even symmetric about O ¼ 0 and O ¼ p; Hðej0 Þ ¼ h½ðN 1Þ=2 þ

ðN3Þ=2 X

2 h½n:

8.6 Properties of the Discrete-Time Fourier Transforms The DTFT of a sequence and its inverse were given before and are (see (8.5.9)):

n¼0

XðejO Þ ¼ Type 2 sequence: "N=21 # X Hðe Þ¼ 2h½ncos½nOððN1Þ=2ÞO ejOðN1Þ=2 jO

n¼0

¼A2 ðejO ÞejOðN1Þ=2 ; Hðej 0 Þ ¼

ðN=2Þ1 X

Type 3 sequence: HðejO Þ ¼ j 2

# h½n sin½ððN 1Þ=2Þ nÞO

n¼0

ejOðN1Þ=2 ¼ A3 ðejO Þejððp=2ÞðN1ÞO=2Þ ;

(8:5:26)

A3 ðejO Þ : Odd symmetric about O ¼ 0 and O ¼ p ! Hðej0 Þ ¼ 0 and Hðejp Þ ¼ 0: Type 4 sequence: " jO

Hðe Þ ¼ 2

ðN=2Þ1 X

1 2p

XðejO ÞejOn dO;

x½n

DTFT

# h½n sin½nOððN 1Þ=2ÞO

n¼0

jejOðN1Þ=2 ¼ A4 ðejO Þejðp=2ðN1ÞO=2Þ :

! XðejO Þ:

p

(8:6:1)

8.6.1 Periodic Nature of the Discrete-Time Fourier Transform

2 h½n:

A2 ðejO Þ: Even symmetric about O ¼ 0 and is odd symmetric about O ¼ p ! Hðejp Þ ¼ 0.

ðN3Þ=2 X

n¼1 ðp

x½nejnO ; x½n

(8:5:25)

n¼0

"

¼

1 X

(8:5:27)

A4 ðejO Þ : Odd symmetric about O ¼ 0 and even & symmetric about O ¼ p ! Hðej0 Þ ¼ 0: These finite length sequences will be useful in studying finite impulse response (FIR) filters in Chapter 9 are considered. Specifications are given

The F-transform of the discrete-time signal x½n is periodic with period 2p. That is, XðejðOþ2pÞ Þ ¼ XðejO Þ:

(8:6:2)

The continuous-time transform is defined in terms of o in radians/second over the entire range 15o51. When the analog signals are sampled at a sampling frequency of fs Hertz, the spectrum of the digitized signal is periodic with period os ¼ 2pfs ¼ ð2p=ts Þ. The normalized frequency O ¼ ð2pf=fs Þ defines the digital frequency. Notes: In the continuous-time domain, periodic signals are expressed in terms of discrete F-series coefficients. In the discrete-time domain, the samples xðnts Þ are located at discrete times and the DTFT is continuous and periodic with period 2p. The interest is in the digital frequency b and jOj p & or in the range 0 O52p. Example 8.6.1 Find the DTFT of the sequence x½n ¼ 1; 0 n5N. Give the expressions for the magnitude and the phase characteristics of the

340

8 Discrete-Time Signals and Their Fourier Transforms

transform. Sketch the magnitude and phase responses for 05O52p assuming N ¼ 21: Solution: The transform, its amplitude and phase responses are as follows: XðejO Þ ¼

N 1 X

h½n and 20logHðejO Þ, 0 O p are shown in Fig. 8.6.1 for N ¼ 21. The side lobes of the amplitude response become smaller as O goes away from p on both of its sides. The peak of the first side lobe appears near the middle of the first side lobe and is & approximately equal to –13.29 dB.

1 1 ejOn X ; jx½nj ¼ N51; jO 1e n¼1 n¼0 Notes: The amplitude spectrum of a typical win(8:6:3) dow is shown in Fig. 8.6.2. It is even and 2p periodic and the frequency interval of interest is 0 O p. eON=2 ðejON=2 ejON=2 Þ=2j The windows of interest have linear phase. The high ¼ jO=2 e ðejO=2 ejO=2 Þ=2j frequency decay rate of the envelope of the spectrum side lobes tells how fast the spectrum envelope sinðON=2Þ ; (8:6:4) decays after the first zero crossing. ¼ ejOðN1Þ=2 sinðO=2Þ Window parameters: (See Fig. 8.6.2):

ejOn ¼

XðejO Þ ¼ jsinðON=2Þj ; ﬀXðejO Þ jsinðO=2Þj sinðON=2Þ : (8:6:5) ¼ ðN 1ÞO=2 þ arg sinðO=2Þ The amplitude XðejO Þ is even and the phase ﬀXðejO Þ is odd. Both are periodic with period 2p. At O ¼ 0, the function is indeterminate and lim

O!0

sinðON=2Þ ¼ N: sinðO=2Þ

GP = Peak gain of main lobe ¼ N; Gp =N ¼ 1 ¼ 0 dB Gs = Peak side lobe gain, Gs =Gp 0:2172 ¼ 13:3 dB OM = Half-width of main lobe = 2p=N (8.6.7) O3 ¼ 3 dB = half-width, W3 =WM ¼ 0:44 O6 ¼ 6 dB = half-width, W6 =WM ¼ 0:6 Os = Half-width of main lobe to reach Ps ; Ws =WM ¼ 0:81

(8:6:6)

Note XðejO Þ ¼ 0 when O ¼ 2 kp=N; k 6¼ 0. The spacing between zero crossings is (2p=NÞ. The phase angle corresponding to the main lobe is ðN 1ÞO=2 resulting in a value of ðN 1Þp=N at O ¼ ½2p=N . At O ¼ ð2p=NÞþ , the phase angle jumps by p rad reaching a value of p=N since sinðON=2Þ= sinðO=2Þ is positive in the main lobe and negative in the first side lobe. This process is repeated and at O ¼ 2p, the phase angle takes the value of 0 completing one period. The sequence

High-frequency attenuation = 20 dB=decade

8.6.2 Superposition or Linearity Assuming DTFT½xi ½n ¼ Xi ðejO Þ and a0i s are constants, the linearity property is M X

ai xi ½n

!

i¼1

ai Xi ðejO Þ:

(8:6:8)

Magnitude spectrum in dB

0

1

M X i¼1

Rectangular window: N = 21

X: 0.4428 Y: -13.29

–10 –20 Magnitude (dB)

0.8 Amplitude

DTFT

0.6 0.4

–30 –40 –50 –60

0.2 –70 0

Fig. 8.6.1 Discrete rectangular window function and its amplitude spectrum

–80 –10 –8 –6 –4 –2

0 2 Index n

4

6

8 10

0

0.5

1

1.5

2

Frequency (Omega)

2.5

3

8.6 Properties of the Discrete-Time Fourier Transforms Fig. 8.6.2 Amplitude spectrum of typical windows

341

DTFT magnitude spectrum of a typical window

P 0.707P 0.5P PSL

High-frequency decay π

0 Ω3

Ω6

8.6.3 Time Shift or Delay

ΩS ΩM

x1 ½n ¼ 1

This property states

DTFT

! 2pdðOÞ:

DTFT

(8:6:9)

x2 ½n ¼cosðO0 nÞ ¼ :5ðeþjO0 n þ ejO0 n Þ

DTFT

! pdðO O0 Þ

jO

þ pdðO þ O0 Þ ¼ X2 ðe Þ:

This follows from

¼

1 X

x½n n0 ejnO

n¼1 1 X

x½mejðmþn0 ÞO

m¼1

"

¼

(8:6:12)

b. Using the Euler’ formula and (8.6.10) results in

x½n n0 ! ejOn0 XðejO Þ:

Ffx½n n0 g ¼

Ω

1 X

# x½mejmO en0 O :

m¼1

Example 8.6.2 Show that the following relationship is true: x½n ¼ ejO0 n XðejO Þ ¼ 2pdðO O0 Þ; jO0 j p: (8:6:10) Solution: This can be shown using the sifting property of the impulse function. 1 ð 1 x½n ¼ 2pdðO O0 ÞejnO dO ¼ ejnO jO¼O0 ¼ejnO0 : & 2p

(8:6:13) &

8.6.4 Modulation or Frequency Shifting The dual of time shifting is the frequency shifting and is given below ejnO0 x½n

DTFT

! XðejðOO0 Þ Þ:

(8:6:14a)

An extension of this is the modulation in time and the corresponding transform pair is jnO0 e þ ejnO0 DTFT ! x½n cosðnO0 Þ ¼ x½n 2 i 1h XðejðOO0 Þ Þ þ XðejðOþO0 Þ Þ : 2 (8:6:14b)

1

Example 8.6.3 Find the DTFTs of the following functions using the pair in (8.6.10) and the shift property. a: x1 ½n ¼ 1; b: x2 ½n ¼ cosðO0 nÞ; jO0 j p: (8:6:11) Solution: a. Using O0 ¼ 0 in (8.6.10), we have the result as follows:

8.6.5 Time Scaling Time scaling deals with the DTFT of x½cn, where ‘‘C’’ is an integer. For example, consider y½n ¼ x½2n, then y½n has only the even samples of x½n. This is decimation (see Section 8.3.1). To simplify the notation define the following sequence assuming m as an integer:

342

8 Discrete-Time Signals and Their Fourier Transforms

xðmÞ ½n ¼

x½n=m ¼ x½k; 0;

The time scaling property is xðmÞ ½n

n ¼ km; n and k are integers n 6¼ km: an u½n

DTFT

! XðejmO Þ:

(8:6:16)

DTFT

!

1 ; jaj51; a 6¼ 0: 1 aejO

ae ! j dð1=ð1dO

DTFT

y½n ¼ nan u½n þ an u½n 1 X

1 X

xðmÞ ½nejnO ¼

n¼1

ð8:6:19Þ

Solution: From the differentiation property,

This can be seen from F½xðmÞ ½n ¼

(8:6:15)

x½kmejkðmOÞ

þ

k¼1

jO

Þ

1 1 ¼ ; 1 aejO ð1 aejO Þ2

¼ XðejmO Þ ðperiodic with period 2p=mÞ: It illustrates the inverse relationship between time and frequency. The signal spreads in time ðm > 1Þ corresponds to its transform being compressed. Time reversal: A special class of time scaling is time reversal and it results in reversal in frequency. That is, 1 X x½nejnO F½x½n ¼ ¼

n¼1 1 X

x½mejmðOÞ ¼ XðejO Þ:

(8:6:17)

m¼1

Note that jF½x½nj ¼ XðejO Þ ¼ XðejO Þ ¼ jF½x½nj:

) y½n ¼ ðn þ 1Þan u½n

jO

! j dXðe dO

DTFT

Þ

(8:6:18)

:

This is shown by dXðejO Þ d ¼ dO dO ¼

"

1 X

# x½ne

1 ; ð1 aejO Þ2 jaj51; a 6¼ 0:

Example 8.6.5 Find the DTFT of x½n ¼ ajnj ; jaj51; a 6¼ 0. Solution: x½n can be expressed as a sum of the right-and left-side sequences in the form x½n ¼ an u½n þ an u½n d½n. The transforms of each of these are 1 DTFT ; (8:6:21a) an u½n ! 1 aejO an u½n

DTFT

!

1 ; 1 aejO DTFT

! 1:

(8:6:21b) (8:6:21c)

Note the time reversal property of the DTFT was used to find the DTFT of an u½n. With these and making use of the linearity property of the DTFT, we have the DTFT pair

This property is nx½n

!

(8:6:20) &

d½n

8.6.6 Differentiation in Frequency

DTFT

x½n ¼ ajnj þ

DTFT

!

1 1 aejO

1 1; jaj51; a 6¼ 0: (8:6:21d) & 1 aejO

jnO

n¼1

1 X

½ðjnÞx½nejnO :

8.6.7 Differencing

n¼1

Example 8.6.4 Derive the DTFT of the function y½n ¼ ðn þ 1Þan u½n; jaj51 using

The differencing property stated below can be shown using the linearity and the time-shifting properties of the DTFT.

8.6 Properties of the Discrete-Time Fourier Transforms

x½n x½n 1

DTFT

!ð1 ejO ÞXðejO Þ:

(8:6:22)

Example 8.6.6 Find the DTFTs of the sequences a: x1 ½n ¼ d½n ðby the direct methodÞ; b: x2 ½n ¼ u½n c: x3 ½n ¼ u½n 1; d: x4 ½n ¼ sgn½n: Solution: a. Ffd½ng ¼

1 X

d½nejnO ¼ 1:

(8:6:23a)

343

pdðOÞ

x½n ¼

b. Let Uðe Þ ¼ DTFTfu½ng. The unit step sequence is a limiting form of the sequence an u½n with a ! 1. Since u½n is not absolutely summable, the transform of the unit step sequence cannot be obtained by taking the limit of the transform as a ! 1 in (8.6.19). Noting that d½n ¼ u½n u½n 1, and defining F½u½n ¼ UðejO Þ, we have F½d½n ¼ 1 ¼ UðejO Þ ejO UðejO Þ ¼ ð1 ejO ÞUðejO Þ:

(8:6:23b)

Since ð1 ejO ÞjO¼0 ¼ 0, it follows that the transform of the unit step sequence will have an impulse, in addition to the transform in (8.6.21a) with a ¼ 1, resulting in u½n

DTFT

1 1 ejO 1 ¼ pdðOÞ þ ; jOj p: 1 ejO (8:6:24)

1 1 ejO

Note the transform is given by the difference between a complex function and its conjugate illustrating the transform of an odd function and is & imaginary. Inverse discrete-time Fourier transform (IDTFT): Finding the inverse transform

n¼1

jO

1 1 ejO 1 1 ¼ : 1 ejO 1 ejO (8:6:26)

d: F½sgn ¼ F½u½n u½n ¼ pdðOÞ þ

1 2p

ðp

XðejO ÞejOn dO:

p

is difficult, as it involves complex integration. Simple cases are illustrated in Example 8.6.7. Alternate methods suggested below are preferable. 1. Since XðejO Þ is periodic, use the F-series of this function (i.e., in the frequency domain O) and then find the corresponding Fourier series coefficients in the time domain. These methods are useful in designing filters and are discussed in Chapter 9. 2. z-Transforms, discussed in the next chapter, can be used to find the IDTFTs. Example 8.6.7 Find the inverse DTFTs of the periodic functions with period 2p

! UðejO Þ ¼ AdðOÞ þ

The average value of the unit step sequence is (1/2) and its transform is pdðOÞ. Example 4.4.10 illustrated the continuous F-transform of a unit step function. c. The DTFT of x3 ½n can be determined by 1 u½n 1 ¼ u½n 1 !pdðOÞ þ 1 ejO 1 2pdðOÞ ¼ pdðOÞ þ : (8:6:25) 1 ejO DTFT

a: XðejO Þ ¼ 2pdðO O0 Þ; 1; jOj W : b: XðejO Þ ¼ 0; W5jOj p

ð8:6:27Þ

c: Use Part b: to find the DTFTs of the sequences x1 ½n ¼ cosðO0 nÞ and x2 ½n ¼ sinðO0 nÞ: Solution: a. The inverse transform is

x½n ¼ ¼

1 2p ðp

ðp

p

XðejO ÞejOn dO

p

dðO O0 ÞejOn dO ¼ ejnO0 ;

344

8 Discrete-Time Signals and Their Fourier Transforms DTFT

) x½n ¼ ejnO0 ! 2pdðO O0 Þ; p5O p;

b:x½n¼

ðp

1 2p

u½n ¼

n X

! pdðOÞ

m¼1

þ XðejO ÞejnO dO¼

DTFT

d½m

1 2p

1 ¼ UðejO Þ; jOj p: (8:6:33) & ð1 ejO Þ

p

8.6.9 Convolution

W ð

e

jnO

sinðWnÞ ; dO¼ pn

Discrete-time convolution property of x1 ½n and x2 ½n is as follows:

W

sinðWnÞ ) x½n ¼ pn

DTFT

!

1; jOj W : 0; W5jOj p

(8:6:29)

1 X

y½n ¼ x1 ½n x2 ½n ¼

x1 ½kx2 ½n k

k¼1

c. Using Euler’s formula and the results in Part a, the DTFTs of the periodic signals are as follows with period 2p: x1 ½n ¼ cosðO0 nÞ ¼ :5ðejO0 n þ ejO0 n Þ

DTFT

!

¼ X1 ðe Þ; p5O0 p;

x1 ½n kx2 ½k ¼ x1 ½n x2 ½n; 1 X

x1 ½kx2 ½n k

k¼1 DTFT

! X1 ðejO ÞX2 ðejO Þ:

(8:6:35)

(8:6:30)

x2 ½n ¼ sinðO0 nÞ ¼ ð1=2jÞðejO0 n ejO0 n Þ

DTFT

!

jp½dðO O0 Þ dðO þ O0 Þ (8:6:31) &

As in the continuous case the time-domain convolution results in the multiplication in the frequency domain and this property plays an important role in discrete-time linear systems. Using the definition of the transform, we have (

8.6.8 Summation or Accumulation

)

n X

jO

Yðe Þ ¼ F

x1 ½kx2 ½n k

k¼1

The accumulation property is the discrete-time counterpart of the integration in the continuous domain. The summation property is shown later in Section (8.6.11) and is n X

x½m

DTFT

j0

!

pXðe ÞdðOÞ

¼

¼

m¼1

þ

1 XðejO Þ; jOj p: 1 ejO

(8:6:34)

k¼1

¼

jO

1 X

y½n ¼ x1 ½n x2 ½n ¼

p½dðO O0 Þ þ dðO þ O0 Þ

¼ X2 ðejO Þ; p5O0 p:

¼

(8:6:32)

¼

!

1 X

n X

n¼1

k¼1

1 X

x1 ½k

k¼1 1 X

x1 ½kx2 ½n k ejOn ;

1 X

! jOn

x2 ½n ke

n¼1

x1 ½kX2 ðejO ÞejOk ;

k¼1

Example 8.6.8 Find the DTFT of u½n using the accumulation property. n P d½m (see (8.3.8b)), Solution: Noting that u½n ¼ d½n

DTFT

! 1, we have

m¼1

" jO

¼ X2 ðe Þ ¼

1 X

# x1 ½ke

k¼1 jO X2 ðe ÞX1 ðejO Þ:

jOk

(8:6:36)

8.6 Properties of the Discrete-Time Fourier Transforms

Convolution in discrete-time corresponds to multiplication in frequency. The accumulation property given in (8.6.32) can now be shown using the convolution theorem and F½u½n ¼ pdðOÞþ jO ½1=ð1 e Þ. That is,

345

Finding the partial fraction expansion in terms of p ¼ ejO would make it a bit easier. Using the DTFT pair in (8.6.21a), the time sequence is

y½n ¼ x½n u½n ¼ ¼

1 X k¼1 n X

x½ku½n k x½k

(8:6:42)

! XðejO ÞUðejO Þ

¼ XðejO Þ pdðOÞ þ

1 ð1 ejO Þ

b. From (8.6.20), we have y½n ¼ ðn þ 1Þan u½n; & jaj51.

1 XðejO Þ; jOj p: ð1 ejO Þ (8:6:37)

Example 8.6.9 Using the DTFTs of the sequences given below (see (8.6.21a)), find the convolution y½n ¼ x1 ½n x2 ½n using the convolution property. x1 ½n ¼ an u½n; x2 ½n ¼ bn u½n; 05jaj; jbj51; a: a 6¼ b; b: a ¼ b: (8:6:38) Solution: a. The DTFTs of the two sequences are given by 1 X1 ðe Þ ¼ Ffx1 ½ng ¼ ; X2 ðejO Þ 1 aejO 1 ¼ Ffx2 ½ng ¼ ; 1 bejO jO

YðejO Þ ¼ X1 ðejO ÞX2 ðejO Þ 1 : ¼ ð1 aejO Þð1 bejO Þ

(8:6:40)

Using the partial fraction expansion, we have

YðejO Þ ¼

anþ1 bnþ1 u½n; a 6¼ b; jaj; jbj51: ab

DTFT

k¼1

¼ pXðej0 Þ þ

¼

a n b a u½n þ bn u½n ba ba

a b þ : ðb aÞð1 aejO Þ ðb aÞð1 bejO Þ (8:6:41)

8.6.10 Multiplication in Time Dual to the convolution in time is multiplication in time and y½n ¼ x1 ½nx2 ½n

DTFT

!

¼ YðejO Þ:

1 ½X1 ðejO Þ X2 ðejO Þ 2p (8:6:43)

Periodic convolution: 1 X1 ðejO Þ X2 ðejO Þ 2p ð 1 X1 ðeja ÞX2 ðejðOaÞ Þda: ¼ 2p

(8:6:44)

2p

The transform of the product of two sequences is the periodic convolution of the two transforms. It can be seen that 1 X

F½y½n ¼ YðejO Þ ¼

¼

1 X

2

x1 ½nx2 ½nejnO

n¼1

41 2p n¼1

ð

X1 ðeja Þejan dax2 ½nejOn :

2p

Interchanging the order of summation and integration results in

346

8 Discrete-Time Signals and Their Fourier Transforms

1 Yðe Þ ¼ 2p jO

"

ð

ja

X1 ðe Þ

1 ¼ 2p

" ja

X1 ðe Þ

da

#

1 X

x2 ½ne

jðOaÞn

da;

Solution: From the F-transform pair, E¼

ð ð

¼ ja

X1 ðe ÞX2 ðe

jðOaÞ

1 X

jx½nj2 ¼

n¼1

X1 ðeja ÞX2 ðejðOaÞ Þda

2p

1 ¼ 2p

Example 8.6.10 Use the Parseval’s identity and the DTFT pair in (8.6.29) to determine the energy contained in the discrete-time signal x½n ¼ sinðWnÞ=pn:

n¼1

2p

1 ¼ 2p

x2 ½nejðOaÞn

n¼1

2p

ð

#

1 X

Þda:

1 2p

W ð

1 X sin2 ðWnÞ

ðpnÞ2

n¼1

ð1Þ2 dO ¼

W : p

(8:6:46) &

W

2p

8.6.12 Central Ordinate Theorems 8.6.11 Parseval’s Identities The discrete versions of the Parseval’s identities are as follows: 1 X

x1 ½nx2 ½n ¼

n¼1

2p

ð

1 2p

X1 ðejO ÞX2 ðejO ÞdO;

2p

ðGeneralized Parseval ’s identityÞ;

1 X

jx½nj2 ¼

n¼1

1 2p

ð

(8:6:45a)

XðejO Þ2 dO;

2p

ðParseval ’s identityÞ;

(8:6:45b)

YðejO ÞjO¼0 ¼

ð

x1 ½nx2 ½nejnO jO¼0 ja

X1 ðe ÞX2 ðe

jðOaÞ

ÞdajO¼0 ;

2p

1 ¼ 2p

ð

1 2p

ð

DTFT

! XðejO Þ,

find the

Ffð1Þn x½ng ¼ ¼

1 X n¼1 1 X

ð1Þn x½nejnO x½nejnðOþpÞ ¼ XðejðOþpÞ Þ:

n¼1

(8:6:50) ja

X1 ðe ÞX2 ðe

ja

Þda

2p

¼

For an introduction to data encryption, see Hershey and Yarlagadda (1986). It is a vast area and most of these techniques are based on manipulating the data in time domain. A simple spectral based encryption can be seen by using the DTFT illustrated below.

Solution:

n¼1

1 ¼ 2p

8.6.13 Simple Digital Encryption

Example 8.6.11 Using x½n DTFT of ð1Þn x½n.

These can be seen by 1 X

From the DTFT and the IDFT, it follows that ð 1 X 1 j0 Xðe Þ ¼ x½n; x½0 ¼ XðejO ÞdO: (8:6:47) 2p n¼1

X1 ðejO ÞX2 ðejO ÞdO:

2p

Note that if x2 ½n ¼ x1 ½n ¼ x½n, then X2 ðejO Þ ¼ X1 ðejO Þ ¼ X ðejO Þ.

Multiplying the time sequence by ð1Þn simply changes the sign of the data with odd indexes. Since the DTFT spectrum is periodic with period equal to 2p, this operation corresponds to the spectral inversion in the frequency band 0 O p. In Chapter 10, Example 10.4.1, we will consider the & analog frequency band inversion.

8.7 Tables of Discrete-Time Fourier Transform (DTFT) Properties and Pairs

8.7 Tables of Discrete-Time Fourier Transform (DTFT) Properties and Pairs Table 8.7.1 Discrete-time Fourier transform (DTFT) properties xi ½n

DTFT

! Xi ðejO Þ

Linearity: x½n ¼

M P

ai xi ½n

M DTFT P

!

i¼1

ai Xi ðejO Þ:

i¼1

Time shift or delay: DTFT

! ejOn XðejO Þ;

x½n n0

n0 is an integer:

0

Frequency shift and modulation: ejO0 n x½n

DTFT

! XðejðOO Þ Þ: 0

x½n cosðnO0 Þ

DTFT

! 12

XðejðOO0 Þ Þ þ XðejðOþO0 Þ Þ :

Conjugation: x ½n

DTFT

! X ðejO Þ:

Time reversal: x½n

DTFT

! XðejO Þ:

Time scaling: x½n=m; n ¼ km xðmÞ ½n ¼ 0; n 6¼ km Times n property: nx½n

DTFT

DTFT

! XðejmO Þ:

jO

Þ ! j dXðe dO :

First difference: x½n x½n 1

DTFT

!ð1 ejO ÞXðejO Þ:

Summation or accumulation: n DTFT P x½k ! pXðej0 ÞdðOÞ þ 1e1jO XðejO Þ; jOjp: k¼1

Time convolution: x1 ½n x2 ½n

DTFT

! X1 ðejO ÞX2 ðejO Þ:

Multiplication in time: DTFT

1 x1 ½nx2 ½n ! 2p X1 ðejO Þ X2 ðejO Þ ; periodic convolution: Even and odd parts of a real function: x½n ¼ xe ½n þ xo ½n xe ½n

DTFT

! ReðXðejO ÞÞ þ jImðXðejO ÞÞ:

DTFT

! ReðXðejO ÞÞ;

xo ½n

DTFT

! jImðXðejO ÞÞ:

Parseval’s theorem: 1 R P 1 x1 ½nx2 ½n ¼ 2p X1 ðejO ÞX2 ðejO ÞdO: n¼1

2p

1 P n¼1

1 jx½nj2 ¼ 2p

R XðejO Þ2 dO:

2p

Central ordinate theorems: Xðej0 Þ ¼

1 P n¼1

1 x½n; x½0 ¼ 2p

R 2p

XðejO ÞdO:

347

348

8 Discrete-Time Signals and Their Fourier Transforms Table 8.7.2 Discrete-time Fourier transform (DTFT) pairs Unit sample function: d½n n0

DTFT

! ejOn

0

Constant: x½n ¼ A

DTFT

! A2pdðOÞ;

jOj p:

Periodic functions: DTFT

! 2pdðO O0 Þ; jOj; jO0 j p: DTFT cosðO0 nÞ ! p½dðO O0 Þ þ dðO þ O0 Þ; jOj; jO0 j p: DTFT sinðO0 nÞ ! jp½dðO O0 Þ dðO þ O0 Þ; jOj; jO0 j p: 1 1 DTFT P P d½n kN ! O0 dðO kO0 Þ; O0 ¼ 2p=N:

ejO0 n

k¼1

k¼1

Unit pulse sequences: u½n

! pdðOÞ þ 1 1ejO ;

DTFT

jOj p:

! pdðOÞ þ 1 1ejO ; jOj p:

DTFT

u½n 1

Exponential sequences: an u½n

1 ! 1 ae ; jO

DTFT

1 ! 1 ae ; jO

DTFT

an u½n 1 ajnj ; a51

! 1 2a cosðOÞ þ a2 1

DTFT

!

ð1 aejO Þ2

Sinc functions:

x½n ¼

jaj > 1:

1 a2

DTFT

ðn þ 1Þan u½n

x½n ¼

jaj51:

sinðWnÞ pn

DTFT

1; 0nN 1 0; otherwise

!

; jaj51:

1; jOj p 0; WjOj p

! ej OðN1Þ=2 sinðON=2Þ sinðO=2Þ

DTFT

8.8 Discrete-Time Fourier-transforms from Samples of the ContinuousTime Fourier-Transforms In Section 8.2, xðtÞ is sampled signal at a sampling rate of fs ¼ ð1=ts Þ > 2ðBÞ; B ¼ Bandwith of xðtÞ: The continuous-time F-transform of xðtÞ and its inverse are

XðjoÞ ¼

1 ð

xðtÞejot dt; xðtÞ

1

1 ¼ 2p

1 ð

XðjoÞejot do; xðtÞ

FT

! XðjoÞ:

1

(8:8:1)

8.8 Discrete-Time Fourier-transforms from Samples of the Continuous-Time Fourier-Transforms

The transform is then approximated using the rectangular integration formula with a sampling interval of ts s. That is, XðjoÞ ﬃ

1 X

ts ½xðnts Þejonts Xos ðjoÞ; os ¼ 2pfs :

n¼1

(8:8:2) Note that Xos ðjoÞ is a periodic function with period os ¼ ð2p=ts Þ and is an approximation of XðjoÞ in the frequency range joj os =2, as Xos ðjðo þ ð2p=ts ÞÞ ¼ ts ¼ ts

1 X n¼1 1 X

xðnts Þejðoþð2p=ts ÞÞnts

period of the DTFT gives its complete information. 4. Most of the continuous-time functions are time limited to, say T ¼ Nts seconds. In computing the transform, two variables need to be selected, the sampling interval ts and the number of sample points N. Note that if xðtÞ has discontinuities, taking its Fourier transform XðjoÞ and then the inverse transform, F1 ½XðjoÞ, gives half-values of the function at the discontinuities. Therefore, the sampled values at these locations are taken as half-values. For example, if xðtÞ ¼ et uðtÞ, then x½0 ¼ :5: 5. The spectrum of the sampled signal in the interval 0 o5os ¼ ð2p=ts Þ is

xðnts Þejonts ¼ Xos ðjoÞ:

n¼1

(8:8:3)

349

Xos ðjoÞ ﬃ ts

N1 X

xðnts Þejðots Þn ; 0 o5os ¼ 2p=ts :

n¼0

(8:8:6) The transform XðjoÞ is arbitrary and Xos ðjoÞ is periodic. If xðtÞ is band limited to half the sampling rate, the signal xðtÞ can be recovered from the sampled signal and the signal spectrum is XðjoÞ ¼

Xos ðjoÞ; joj os =2 : 0; joj > os =2

(8:8:4)

The discrete-time Fourier transform (DTFT) was defined earlier assuming x½n ¼ xðnts Þ and XðejO Þ ¼

1 X

x½nejnO :

(8:8:5)

n¼1

From this review, the following conclusions can be made: 1. The DTFT is periodic, whereas the continuous Fourier transform is not periodic. 2. The sampling interval ts is not included in (8.8.5). Approximation to the continuous F-transform can be obtained from the DTFT of the sampled signal by multiplying it by ts . 3. The DTFT is defined in terms of the normalized frequency, O ¼ ots ¼ o=fs where fs is the sampling frequency. The normalized frequency O is referred to as the digital frequency and is measured in radians/sample or in radians/cycle. Noting that the DTFT is periodic with period 2p, one

6. Finally, considering Item 4 above in approximating the continuous Fourier transform using the DTFT, the number of sample points, the sampling interval, and the sample values need to be considered. Example 8.8.1 In this example, some of the important facets associated with computing the transform xðtÞ ¼ e2t uðtÞ using DTFT are discussed. What should be the interval T before sampling the signal and the number of samples to be used? Use Example 8.2.6 Part b to find the sampling interval. Solution: First xðtÞ is not time limited. For discrete computations, only a finite interval of time needs to be considered, say T. Find T such that in the interval 0 t5T,xðTÞ551.For T ¼ 4 and 5, we have xð4Þ ¼ :00033 and xð5Þ ¼ :000045. Both are small enough, either one would be adequate and let T ¼ 4: From Example 8.2.6, the sampling interval is ts ¼ ð1=fs Þ p=200. The number of samples is T=ts 255: Discrete computations of transforms are the most efficient if the number of sample points N is a power of 2. Select N ¼ 256: The next step is to identify the sample values. Noting that xðtÞ has a jump discontinuity at t ¼ 0, the first sample value is ½xð0 Þþ xð0þ Þ=2 ¼ :5 ¼ xð0Þ. The sample values are

350

8 Discrete-Time Signals and Their Fourier Transforms

fxðnts Þg ¼ f:5; e2nts ; n ¼ 1; . . . ; 255g: The discrete-time Fourier transform is approximated at the frequency sampled values using (8.8.6). That is, Xos ðjðk=NÞos Þ ﬃ ts

N1 X

intervals of ð2p=Nts Þ in one period, then we have N values Xos ðok Þ; ok ¼ 2pk=N. That is, Xos ðj2pk=NÞ ¼ ts

N1 X

xðnts Þejð2pk=NÞn ;

n¼0

k ¼ 0; 1; 2; . . . ; N 1: xs ½nej2pðkn=NÞ ;

n¼0

k ¼ 0; 1; . . . ; N 1: This is a periodic sequence with period N ¼ 256. The samples in the frequency domain are spaced apart by ð1=NÞfs ¼ ð1=256Þ64 ¼ :25 Hz. Notes: The DTFT is a periodic, continuous function of O and the sampled transform values computed from the DTFT are discrete and periodic. The spectrum of the analog signal is continuous. Increasing the product ðNts Þ implies a longer signal and the discrete transform has more frequency values. Since the sampling frequency is not changed, the effect of increasing ðNts Þ introduces more frequency values. The frequency interval between the spectral components is reduced. The sampling interval ts (ts ¼ 1=fs ) controls the accuracy in the approximation obtained by the DTFT compared to the actual evaluation of & the continuous Fourier transform. As an example, consider that a signal of a 10-s interval that is band limited to 4 kHz. We are interested in estimating the spectrum of the segment using the above procedure with a resolution of, say 0.1 Hz in the spectral spacing. From the Nyquist sampling theorem, a sampling rate of 10 kHz would be adequate. ) Number of samples ¼ fs =frequency resolution ¼ 10 kHz=:1 Hz ¼ 100; 000: Transform algorithms are most efficient if the number of sample points, N is a power of 2. The next highest number that is a power of 2 is 217 ¼ 131072. The length of the corresponding segment is equal to T ¼ Nts ¼ N=fs ¼ 131; 072=10; 000 13:1 s:

8.9 Discrete Fourier Transforms (DFTs) In the last section, the spectrum of the sampled signal Xos ðoÞ in the interval 0 o5os ¼ ð2p=ts Þ was given in (8.8.6). If we sample this function at

(8:9:1)

There are N sample values in time xðnts Þ and N sample values of the spectrum in (8.9.1). These results can be applied to digital data by starting with x½n ¼ xðnts Þ; n ¼ 0; 1; 2; . . . ; N 1 and defining the discrete Fourier transform (DFT) by X½k DFT½x½n ¼

N1 X

x½nejð2p=NÞkn ;

n¼0

k ¼ 0; 1; 2; . . . ; N 1:

(8:9:2)

Note that multiplying X½k by ts results in the spectral values in (8.9.1). The next question is how can the data x½n be obtained from the discrete Fourier transform coefficients X½k? It turns out that these can be determined by x½n ¼

N1 1X X½kejð2p=NÞkn ; N k¼0

n ¼ 0; 1; 2; . . . ; ðN 1Þ:

(8:9:3)

The following shows that (8.9.3) is valid. Substituting the DFT values (see (8.9.2)) in (8.9.3) results in N1 1X X½kejð2p=NÞnk N k¼0 N1 X N1 1X ½ x½mejð2p=NÞmk ejð2p=NÞkn N k¼0 m¼0 N1 N1 X 1X ¼ x½m ½ejð2p=NÞðnmÞk N m¼0 k¼0 " # N1 N1 X X 1 jð2p=NÞðnmÞ k (8:9:4) ¼ x½m ½e : N m¼0 k¼0

¼

Using the summation formula for the finite geometric series in (C.6.1a) results in N 1 X

ðe

k¼0

jð2p=NÞðnkÞ k

Þ ¼

0; n 6¼ m : N: n ¼ m

(8:9:5)

8.9 Discrete Fourier Transforms (DFTs)

" # N1 N1 X 1X jð2p=NÞðnmÞ k ) x½m ½e ¼ x½n: N m¼0 k¼0

351

Therefore, (8:9:6) le jð2p=NÞkn ¼ e jð2p=NÞ½mþlN ¼ e jð2p=NÞm e jð2p=NÞðlNÞ

The data x½n is now tied to the discrete Fourier transform coefficients, or DFTs identified as X½ks. We can summarize the results in terms of DFTs, inverse DFTs, and the symbol for the transform pair as follows:

¼ e jð2p=NÞm ; 0 m N 1:

Noting this, only N terms in (8.9.9) are needed to compute the DFT. The DFT and IDFT have implied periodicity with period N. That is,

X½k DFT½x½n ¼

N 1 X

X½k þ N ¼ x½ne

jð2p=NÞkn

¼

(8:9:7a) N1 1X X½kejð2p=NÞkn ; N k¼0

n ¼ 0; 1; 2; . . . ; ðN 1Þ; DFT

x½n ! X½k:

Notes: There are N equations each to determine the DFTs and the IDFTs. The exponential function e jð2p=NÞ is periodic with period N: Indices in the DFT and IDFT variables are restricted to the principal range 0 n; k N 1. The multiplicative terms in the forward and the inverse transforms e jð2p=NÞkn take one of the values in the set n

o 1; e jð2p=NÞ ; e j2ð2p=NÞ ; . . . ; e jðN1Þð2p=NÞ :

(8:9:9)

This follows from the fact that for 0 k; n N 1; the product ðknÞ can be written as kn ¼lN þ m; 0 k; n; m N 1; k; n; l; m and N; integers:

(8:9:10)

In compact form we can use modulo (mod) N arithmetic. That is, kn ¼ lN þ m m modðNÞ ½mmodðNÞ ¼ ½mðNÞ ; 0 m N 1:

N1 X

(8:9:11)

N 1 X

x½nejð2p=NÞnk ejð2p=NÞnN ;

x½nejð2p=NÞnk ¼ X½k:

(8:9:13a)

n¼0

(8:9:7b)

The sequences x½n; 0 n N 1 and X½k; 0 k N 1 form a discrete Fourier transform (DFT) pair.

x½nejð2p=NÞnðkþNÞ

n¼0

¼

(8:9:8)

N1 X n¼0

; k ¼ 0; 1; 2; . . . ; N 1;

n¼0

x½n ¼ IDFT½X½k ¼

(8:9:12)

x½n þ N ¼ ¼

N1 1X X½kej2pðnþNÞk=N N n¼0 N1 1X X½kej2pðnÞk=N ¼ x½n: N n¼0

(8:9:13b)

At a later time, time and frequency shifts will be considered, such as x½n n0 and X½k k0 , where k0 and n0 are some integers. Since the integers ½n n0 and ½k k0 may fall outside of the range ½0; N 1, these integers need to be converted to a number in the principal range using mod N arithmetic. For example, ½1mod 16 ¼ 15; ½17mod 16 ¼ 1; x½½k þ 1mod 16 ¼ x½ðk þ 1 16Þ ¼ x½k 15: These will not be identified explicitly and implied & from the context. Notes: Modular arithmetic was introduced by Carl Friedrich Gauss in his work Disquisitions Arithmetica, see Hawking (2005). Gauss is considered to be one of the great mathematicians who ever lived. His work laid the foundation for number theory. Many of the digital coding and encryption algorithms are based on number theory. See Gilbert and Hatcher (2000), & Hershey and Yarlagadda (1986), and others.

352

8 Discrete-Time Signals and Their Fourier Transforms

Interestingly, the same algorithm can be used to compute both the forward and the inverse DFT transforms, which can be seen by rewriting (8.9.7b) as

x½n ¼

N1 1X ½X ½kejð2p=NÞnk : N k¼0

8.9.1 Matrix Representations of the DFT and the IDFT The discrete Fourier transform (DFT) can be written in a matrix form and in compact form (see Appendix 1 for a brief review of matrices) using the equations N1 1X DFT x½n ¼ X½kejnð2p=nÞk ! X½k N n¼0

(8:9:14a)

Pictorially (8.9.8) can be described by X½k ! X ½k ! DFTfX ½kg 1 ! ðDFTfX ½kgÞ ¼ x½n: N 2 6 6 6 6 6 6 6 6 4

X½0 X½1 : : :

3

2

1 7 61 7 6 7 6 7 6: 7¼6 7 6: 7 6 7 6 5 4:

X½N 1

¼

e

X ¼ ADFT x:

(8:9:14b)

(8:9:15)

These are

ejð2p=NÞ

: : : :

: :

: :

: : : :

: :

:

: :

:

jð2p=NÞðN1Þ

: :

The vectors X and x are N-dimensional column vectors and the matrix ADFT is a N N complex matrix and is referred to as a discrete Fourier transform (or DFT) matrix. A typical entry in ADFT , say ðk; nÞ entry, is ðk1Þðn1Þ

32

1

6 ejð2p=NÞðN1Þ 7 76 76 76 : 76 76 : 76 76 54 :

: e

jð2p=NÞðN1Þ2

x½0 x½1 : : :

3 7 7 7 7 7; 7 7 7 5

(8:9:16a)

x½N 1

The constant WN is an N th root of unity, as WN N ¼ 1. From the second row or column in the ADFT matrix we have the roots of unity. The exponent t in the entries WtN ¼ ejð2p=NÞt is called the twiddle factor or rotation factor, see Rabiner and Gold (1975). All the entries in ADFT can be simplified to one of the values in the following set (see 8.9.12):

(8:9:16b)

ADFT ðk; nÞ ¼ ejð2p=NÞðk1Þðn1Þ ¼ WN

x½nejkð2p=NÞn

k¼0

1

1

N 1 X

n

; WN

o 1; ejð2p=NÞ ; ejð2p=NÞ2 ; . . . ; ejð2p=NÞðN1Þ : (8:9:18)

¼ ejð2p=NÞ ; 1 k N; 1 n N: The IDFT in (8.9.21b), in a matrix form and its compact from are

(8:9:17)

2 6 6 6 6 6 6 6 6 4

x½0 x½1 : : : x½N 1

3

2

1

7 61 7 6 7 6 7 16: 7¼ 6 7 N6 : 7 6 7 6 5 4:

32

1

:

: :

1

ejð2p=NÞ :

: :

: : : :

ejð2p=NÞðN1Þ :

: :

: :

: : : :

: :

:

: : ejð2p=NÞðN1Þ

1 ejð2p=NÞðN1Þ

2

76 76 76 76 76 76 76 76 54

X½0 X½1 : : : X½N 1

3 7 7 7 7 7; 7 7 7 5

(8:9:19a)

8.9 Discrete Fourier Transforms (DFTs)

x¼

353

1 X: A N DFT

pﬃﬃﬃﬃ pﬃﬃﬃﬃ )ðADFT Þð1= NÞð1= NÞðADFT Þ

(8:9:19b)

(8:9:22) ¼ IN or A1 DFT ¼ ð1=NÞADFT : pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ That is, ð ð1=NÞÞADFT is a unitary matrix.

A typical entry, say the ðn; kÞ entry in ADFT is ADFT ðn; kÞ ¼ ejð2p=NÞðn1Þðk1Þ ðn1Þðk1Þ

¼ WN

; WN ¼ ejð2p=NÞ : (8:9:20)

These matrices are known as Vandermonde matrices (see Hohn (1958)). Note that the matrices ADFT and ðð1=NÞADFT Þ are symmetric matrices. That is, ADFT ðk; nÞ ¼ ADFT ðn; kÞ and ADFT ðk; nÞ ¼ ADFT ðn; kÞ; 2

1

(8:9:21) 3

6 1 7 6 7 x¼6 7; ADFT 4 1 5

2

1 61 6 ¼6 41

2

1

1 j ðjÞ

2

ðjÞ3

ðjÞ

1 j

2

4

ðjÞ6

32 3 2 3 1 1 1 1 1 3 6 1 j 1 j 76 1 7 6 j3 7 6 76 7 6 7 X¼6 76 7¼6 7: (8:9:24) 4 1 1 1 1 54 1 5 4 1 5 j

Solution: For N ¼ 4; we have ðejð2p=NÞ Þkn ¼ ðejð2p=4Þ Þkn ¼ ðejðp=2Þ Þkn ¼ ðjÞkn . Now, the data vector, the DFT matrix and DFT sequence (or coefficients) are

1 ðjÞ2

2

1

Example 8.9.1 Given the data x½0 ¼ 1; x½1 ¼ 1; x½2 ¼ 1; x½3 ¼ 2, find the discrete Fourier transform (DFT) of this sequence using the DFT matrix. Using the DFT coefficients and the ADFT matrix, find the corresponding data sequence.

j3

The DFT matrix is symmetric and it contains only N ¼ 4 distinct elements 1, 1; j; j2 ¼ 1; and j3 ¼ j corresponding to the 4-point DFT. If we can save 1 and j, we can generate the others by changing the sign of these. The data sequence can be computed from the DFT vector by

ðjÞ9

1

32 3 2 3 1 1 1 1 3 1 76 7 6 7 1 16 6 1 j 1 j 76 j3 7 6 1 7 x ¼ ADFT X ¼ 6 76 7¼6 7: 4 44 1 1 1 1 54 1 5 4 1 5 1 j 1 j j3 2

X½k ¼

N 1 X

1

1

j 1

1 1

j 7 7 7 1 5

j

1

j

(8:9:23)

½Refx½ng

n¼1

kn ; þ jImfx½ng Re Wkn N þ jIm WN ¼

N1 X

N1 X Refx½ngRe Wkn Imfx½ngIm Wkn N N

n¼0

"

þj

8.9.2 Requirements for Direct Computation of the DFT

3

1

values in frequency( see (8.9.16a)). It involves the multiplication of a N N complex matrix by a Ndimensional vector. The direct computation of DFT requires N2 complex multiplications and NðN 1Þ additions. Fast Fourier transform (FFT) algorithms, considered in the next chapter, reduces these numbers significantly. FFT algorithms are most effective when the number of data points N is a power of 2. DFT is applicable for real and complex data and can be implemented using real multiplications. First, let x½n ¼ Re½x½n þ jIm½x½n and

2

(8:9:25) &

3 2 1 1 61 ðjÞ3 7 7 6 7¼6 ðjÞ6 5 4 1

n¼0 N1 X

Refx½ngIm Wkn N

n¼0

þ

N 1 X

# kn Imfx½ngRe WN ; k ¼ 0; 1:::; N 1:

n¼0

DFT is a transformation that takes a set of N complex (or real) values in time to N complex (or real)

¼ Re½X½k þ j Im½X½k; k ¼ 0; 1 . . . ; N 1: (8:9:26)

354

8 Discrete-Time Signals and Their Fourier Transforms

Computation of X½k requires 4 N2 real multiplications. For fixed or floating point operations, multiplications are computationally more expensive compared to the additions or transfer of data. We will come back to this topic in Section 9.3. Interestingly, x ½n

DFT

! X ½kModðNÞ :

(8:9:27)

N1 X

¼

N 1 X

X½k ¼ X ½k ¼ X½N k:

X ½k ¼ " x ½nej2pnk=N ¼

N1 X

# x½nej2pnk=N

;

N1 X

¼

¼ X ½kmodðNÞ :

X½N k ¼

modðNÞ

DFT

¼ DFTfx i ½ng; i ¼ 1; 2; . . .

N1 X

x½nejð2p=NÞnk ¼ X ½k: X½k

is

real,

then

x½n ¼x1 ½n þ jx2 ½n

DFT

! X½k; x½n

DFT

! X1 ½k; x2 ½n

DFT

! X½k

DFT

! X2 ½k

1 X1 ½k ¼ fX½k þ X ½N kg; X2 ½k 2 1 ¼ jfX½k X ½N kg: 2

xi ½nejð2p=NÞnk ; xi ½n

k¼0

(8:10:1)

Notes: In the following one variable may be replaced by another variable. The variable n and k will be used to identify the time and the frequency sequences. Most of the proofs follow by using the basic definitions of the DFT (or IDFT). Both x½n and X½k are assumed to have implied periodicity with period N, i.e. x½n þ N ¼ x½n and X½k þ N ¼ X½k; x½n ¼ x½nmodN x½nN and X½k ¼ X½kmodN X½kN :

x½nejð2p=NÞnðNkÞ

Notes: The DFT of two real valued sequences x1 ½n and x2 ½n can be determined from the DFT of the complex sequence x½n ¼ x1 ½n þ jx2 ½n as follows:

x1 ½n

IDFTfXi ½kg ¼ xi ½n ! X i ½k

N1 1X Xi ½kejð2p=NÞnk : N k¼0

N1 X

We can show that if x½n ¼ x ½n ¼ x½N n.

DFT properties are similar to the continuous case. In proving these properties, we will assume that the discrete-time sequences are given by xdi ½n; 0 n N 1, and their discrete Fourier transforms are given by Xdi ½k; 0 k N 1. That is,

¼

x½nej2pnðkÞ=N ¼ X½k;

k¼0

8.10 Discrete Fourier Transform Properties

Xi ½k ¼

x½ne

k¼0

¼

N1 X

# j2pnk=N

n¼0

#

n¼0

N1 X n¼0

n¼0

x½nej2pnðkÞ=N

(8:10:2)

These can be shown as follows:

n¼0

"

The DFT of a real sequence xd ½n has conjugate symmetry. That is,

"

This follows from DFTfx ½ng ¼

8.10.1 DFTs and IDFTs of Real Sequences

8.10.2 Linearity The linearity property follows directly from the definition: M X i¼1

ai xi ½n

DFT

!

M X

ai Xi ½k:

(8:10:3)

i¼1

Example 8.10.1 Find the Nð¼ 8Þ point DFT of the sequence {x½n} and sketch the coefficients. x½n ¼ cosð2pn=NÞ; n ¼ 0; 1; 2; . . . ; N 1:

(8:10:4)

8.10 Discrete Fourier Transform Properties

355

Solution: Using Euler’s identity and the linearity property, the DFT coefficients are

variable other than n and k, say l, and then replace l by –l. That is,

X½k ¼ DFTfx½ng ¼ DFTfcosð2pn=NÞÞg

x½n ¼

¼ :5DFTfejð2pn=NÞ g þ :5 DFTfejð2pn=NÞ g; " # N1 N1 X 1 X jð2pn=NÞð1kÞ jð2pn=NÞð1þkÞ ¼ e þ e (8:10:5) 2 n¼0 n¼0 N N ) X½k ¼ d½k 1 þ d½k þ 1 2 2 N N ¼ d½k 1 þ d½k N þ 1: 2 2

! x½l ¼

N1 1X X½kejð2p=NÞl k ; N k¼1

) x½l ¼ (8:10:6a)

The closed form expression for X½k is obtained by using (C.6.1b). Also, note that d½k þ 1 is outside of the interval 0 n5N, which can be resolved by noting the DFT coefficients have implicit periodicity with period N and write d½k þ 1 ¼ d½k N þ 1. Thus there are two discrete-time impulses in the interval 0 k5N. Here, Xd ½k reduces to X½k ¼ 4d½k 1 þ 4d½k 7:

N1 1X X½kejð2p=NÞnk N k¼1

(8:10:6b)

The transform sequence is shown in Fig. 8.10.1. Note the DFT sequence is real and has even symmetry. Furthermore, we can see that X ½k ¼ X½k ¼ & X½N k.

N1 1X X½rejð2p=NÞl r : N r¼0

Now let l ¼ k; r ¼ n; and taking the ð1=NÞ inside the summation results in the proof of the duality property as follows: x½k ¼

N1 X

ðX½n=NÞejð2p=NÞnk ;

(8:10:8)

n¼0

) X½n=N ¼ IDFT of x½k or the DFTfX½n=Ng ¼ x½k: Example 8.10.2 Use the function in Example 8.10.1 to verify the duality principle. Solution: We have DFT

!ðN=2Þd½k 1

x½n ¼ cosð2pn=NÞn

þ ðN=2Þd½k ðN 1Þ ¼ XðkÞ: Now consider y½n

(8:10:9)

DFT

! Y½k with

8.10.3 Duality y½n ¼ :5d½n 1 The duality property is x½n

1 ! X½k ! X½n Change to N

DFT

DFT

! x½k:

N1 X

:5fd½n 1

þ d½n ðN 1Þge

To show this, start with the IDFT in terms of X½k and rewrite the function in terms of a different

) Y½k ¼ :5½ejð2p=NÞk þ ejð2p=NÞk ¼ cosðð2p=NÞkÞ ¼ cosð2pk=NÞ Note y½k ¼ x½k:

8.10.4 Time Shift Fig. 8.10.1 Example 8.10.1

!

n¼0 j2pnk=N

(8:10:7)

[X]k

DFT

þ :5d½n ðN 1Þ

The time shift property is

(8:10:10) &

356

8 Discrete-Time Signals and Their Fourier Transforms

x½n mmodðNÞ

DFT

! X½kejð2p=NÞkm or

x½n mmodðNÞ ejð2p=NÞkm

DFT

! X½k:

(8:10:11)

We know that x½n and X½k can be considered as periodic sequences and X½k represents the DFT coefficients for one period of x½n. By using a new variable l ¼ n m, and simplifying, we have DFTfx½n mg ¼

N1 X

x½n mejð2p=NÞnk

¼

x½lejð2p=NÞðlþmÞk

l¼0

¼

( N1 X

) x½le

jð2p=NÞlk

ejð2p=NÞmk

l¼0

¼ X½kejð2p=NÞmk :

2p 2p xe ½m sin Nk cos ðmkÞ ¼ N N m¼0 2p 2p ðNk Þ sin mk ; þ cos N N N 1 X 2p 2p ¼ Nk cos ðmkÞ xe ½m sin N N m¼0 2p 2p þ cos ðNkÞ sin mk ; N N N 1 X 2p ¼ xe ½m sin km N m¼0 N 1 X 2p nk ¼0: xe ½n sin ¼ N n¼0 N 1 X

n¼0 N1 X

The second term on the right is N N1 1 X X 2p 2p nk ¼ nk xe ½n sin xe ½n sin N N n¼0 n¼0 N 1 X 2p xe ½N n sin nk ; ¼ N n¼0 N1 X 2p xe ½m sin ðN mÞk ; ¼ N m¼0

(8:10:12)

8.10.5 Frequency Shift The dual to time shift property is the frequency shift property given below and can be shown taking the IDFT of the coefficients on the right. (8:10:13)

A number is equal to its negative only when it is zero. The coefficients are real and follow from (8.10.16) and (8.10.17). Noting the periodicity of the DFT coefficients, we have

If a function is even, i.e., x½n ¼ x½n xe ½n, then the DFT is real and even. It can be written as N1 X 2p DFT nk : (8:10:14) xe ½n ! Xe ½k ¼ xe ½n cos N n¼0

2p Xe ½k ¼ Xe ½N k ¼ nðN kÞ ; xe ½n cos N n¼0 N1 X 2p 2p kN cos ðknÞ xe ½n cos ¼ N N n¼0 2p 2p þ sin kN sin kn ; N N N1 X 2p ¼ nk ¼ Xe ½k ð ) Xe ½k xe ½n cos N n¼0

x½ne

jð2p=NÞmn

DFT

! X½k m modðNÞ :

8.10.6 Even Sequences

N1 X

This can be verified using xe ½n ¼ xe ½N n ¼ xe ½n; ) DFTfxe ½ng ¼

N1 X

xe ½ne

(8:10:15)

¼ Xe ½k; even sequenceÞ:

(8:10:18)

jð2p=NÞkn

n¼0

8.10.7 Odd Sequences 2p nk N n¼0 If a function is odd, i.e., x½n ¼ x½n x0 ½n, then N1 X 2p xe ½n sin nk : (8:10:16) the DFT is real and even. The DFT coefficients of j N n¼0 an odd function are odd and imaginary and

¼

N1 X

xe ½n cos

8.10 Discrete Fourier Transform Properties

2p DFT½x0 ½n ¼ j x0 ½n sin nk : N k¼0 N1 X

357

(8:10:19)

The proof of this is very similar to the last property. As a final comment on this topic, we have seen that a sequence can be expressed in terms of its even and odd parts. The above two properties allow for the computation of the DFT of an arbitrary periodic sequence in terms of the sum of DFTs of even and odd sequences. That is, DFT

! XðejO Þ

x½n ¼ xe ½n þ x0 ½n

¼ RefXðejO Þg þ jImfXðejO Þg ; xe ½n x0 ½n

(8:10:20)

DFT

! RefXðejO Þg;

DFT

jO

! ImfXðe Þg:

¼

¼

N1 X

N1 N1 X 1X X½kH½mejð2p=NÞmn N k¼0 m¼0 " # N1 1X jð2p=NÞik jð2p=NÞim e e ; N i¼0

x½ih½n i ¼

i¼0

(8:10:24)

X 1 N1 X½kH½kejð2p=NÞkn : (8:10:25) N k¼0

Example 8.10.3 Write the periodic convolution of the following two periodic sequences with period N ¼ 3. Compute these using a. the time sequence and b. the DFT. fx½ng ¼ fx½0; x½1; x½2g ¼ f1; 2; 3g; fh½ng ¼ fh½0; h½1; h½2g ¼ f1; 1; 1g;

(8:10:21)

It follows that if x½n is real and even, then XðejO Þ is real and even and if x½n is real and odd then XðejO Þ is real and imaginary.

y½n ¼

N1 X

(8:10:26)

x½ih½n i;y½n

i¼0

¼ x½0h½n þ x½1h½n 1 þ x½2h½n 2: (8:10:27)

8.10.8 Discrete-Time Convolution Theorem

Solution: a. Using h½n ¼ h½N n, the periodic convolution values are as follows: y½0 ¼ x½0h½0 þ x½1h½1 þ x½2h½2 ¼ x½0h½0 þ x½1h½2 þ x½2h½1;

In Section 8.4.1 the periodic (or cyclic) convolution of two functions x½n and h½n with the same period N was defined by N 1 X y½n ¼ x½n h½n ¼ x½mh½n m modðNÞ

y½1 ¼ x½0h½1 þ x½1h½0 þ x½2h½2; y½2 ¼ x½0h½2 þ x½1h½1 þ x½2h½0;

m¼0

¼

N1 X

x½n m modðNÞ h½m:

m¼0

(8:10:22) The time convolution theorem stated below can be proven by starting with the left side of the above equation and rearranging the terms and then simplifying it. That is, DFT

! X½kH½k (8:10:23) " # N 1 N1 N1 X X 1 X x½ih½n i ¼ X½kejð2p=NÞik y½n ¼ N k¼0 i¼0 i¼0 " # N1 1X jð2p=NÞmðniÞ H½me ; N m¼0 x½n h½n modðNÞ

32 3 3 2 h½0 h½2 h½1 x½0 y½0 76 7 6 7 6 Matrix form : 4 y½1 5 ¼ 4 h½1 h½0 h½2 54 x½1 5: 2

h½2 h½1 h½0

y½2

x½2 (8:10:28)

Note the structure of the coefficient matrix on the right in (8.10.28) has a pattern, which can be written in general terms after this example. Noting that x½0 ¼ 1; x½1 ¼ 2; x½2 ¼ 3 and h½0 ¼ 1; h½1 ¼ 1; h½2 ¼ 1, the convolution values are 2

3 2 y½0 1 6 7 6 4 y½1 5 ¼ 4 1

1 1

1

1

y½2

32 3 2 3 1 1 0 76 7 6 7 1 54 2 5 ¼ 4 4 5: 1

3

2

(8:10:29)

358

8 Discrete-Time Signals and Their Fourier Transforms

b. In matrix form, the transform values are 2

3 2 1 X½0 6 7 6 4 X½1 5 ¼ 4 1

H½0

3

2

e

j2ð2p=3Þ

1

3 3 2 6 x½0 7 76 7 6 ej2ð2p=3Þ 54 x½1 5 ﬃ 4 1:5 þ j:8660 5 ; 32

1

ejð2p=3Þ

1

X½2 2

1

1

6 7 6 4 H½1 5 ¼ 4 1 ejð2p=3Þ H½2 1 ej2ð2p=3Þ

e

j4ð2p=3Þ

1:5 j:8660

x½2 32

1

h½0

(8:10:30)

3

2

3

1

76 7 6 7 ej2ð2p=3Þ 54 h½1 5 ﬃ 4 1 þ j1:7321 5: e

j4ð2p=3Þ

(8:10:31)

1 j1:7321

h½2

The product of the transform coefficients in matrix form are given by 2

Y½0

3

2

X½0H½0

3

2

3

1ð6Þ

2

6

3

6 7 6 7 6 7 6 7 4 Y½1 5 ¼ 4 X½1H½1 5 ¼ 4 ð1 þ j1:7321Þð1:5 þ j:8660Þ 5 ¼ 4 3 j1:7321 5: Y½2 X½2H½2 ð1 j1:7321Þð1:5 j:8660Þ 3 þ j1:7321 These involve complex arithmetic resulting in rounded values. The IDFT of the vector in (8.10.32) gives approximations of the results in (8.10.29). &

y½n ¼ x½n h½n ¼

6 6 6 6 6 6 6 6 6 6 6 4

y½0 y½1 y½2 : : : y½N 1

3

2

7 6 7 6 7 6 7 6 7 6 7 6 7¼6 7 6 7 6 7 6 7 6 5 4

¼

N 1 X

h½N 1 h½0

h½2

h½1

h½0

:

:

: :

: :

: :

: :

: :

: : h½N 3 :

: :

y ¼ Hx:

h½N 2 : h½N 1 :

(8:10:34b)

The vectors y and x are N-dimensional column vectors and H is a N N circulant matrix having N distinct elements with a pattern. First, h½n; n ¼ 0; 1; 2; . . . ; N 1 is placed in column 1 in H. Column 2 is obtained by circularly shifting column 1 down by 1. Similarly, column 3 is obtained by circularly shifting the column 2 down by 1 and

h½ix½n imodðNÞ :

(8:10:33)

i¼0

h½0 h½1

: : h½N 1 h½N 2

x½ih½n imodðNÞ

i¼0

The periodic convolution can be written in general matrix and symbolic forms of two periodic sequences x½0; x½1; :::; x½N 1 and h½0; h½1; :::; h½N 1 as follows: 2

N 1 X

(8:10:32)

32 3 x½0 h½1 6 7 h½2 7 76 x½1 7 76 7 6 7 : h½3 7 76 x½2 7 76 7 : : 76 : 7; 76 7 7 6 7 : : 76 : 7 76 7 : : 54 : 5 : h½0 x½N 1

: : : :

(8:10:34a)

so on. The diagonal entries are the same and the entries in each sub diagonal are the same.

8.10.9 Discrete-Frequency Convolution Theorem The discrete-frequency convolution theorem is a dual to the time convolution theorem and is given

8.10 Discrete Fourier Transform Properties

359

below. The proof of this is very similar to the time convolution theorem. DFT 1 x½nh½n ! ½X½k H½k modðNÞ N N1 1X ¼ X½iH½k i modðNÞ : (8:10:35) N i¼0

Note the bracketed term in (8.10.38a) is equal to 1 if m ¼ k and zero otherwise. As in the periodic convolution, DFTs can be used to compute the cross correlation by first computing the DFTs of the two sequences and then take the IDFT of Xd ½kHd ½k.

8.10.11 Parseval’s Identity or Theorem 8.10.10 Discrete-Time Correlation Theorem

It states that if x½n is real, then

In Section 8.3.2 we briefly discussed the discrete cross correlation and the convolution (see (8.3.20a and b)). The discrete cross correlation of two N point sequences is rxh ½n ¼

N1 X

! DFTfrxh ½ng: (8:10:36)

Note that we are using the variable n for the cross correlation function, as we are using the variable k for the DFT function. The discrete correlation theorem is stated by xðiÞhðn þ iÞmodðNÞ

DFT

! X ½kH½k:

(8:10:37)

i¼0

This can be proven using the following steps: " # N1 N1 1 X X X 1N jð2p=NÞik x½ih½n þ imodðNÞ ¼ X½ke N k¼0 i¼0 i¼0 " # 1 X 1N jð2p=NÞðnþiÞm H½me ; N m¼0 ¼

i¼0

n¼0

N k¼0

¼

(8:10:39)

Example 8.10.4 Verify the Parseval’s theorem using hd ½n in Example 8.10.3. Solution: h½0 ¼ 1; h½1 ¼ 1; h½2 ¼ 1 ) H½0 ¼ 1; H½1 ¼ 1 þ jð1:7321Þ; H½2 ¼ 1 j1:7321; 2 X

h2 ½n ¼ 3;

X ½ke

jð2p=NÞik

# N1 1X jð2p=NÞðnþiÞm ; H½me N m¼0

ð1=3Þ

2 X n¼0

2 X

h2 ½n ¼

2 1X jH½kj2 ¼ 3 3 k¼0

(8:10:40)

jH½kj2 ¼ ð1 þ j1 þ jð1:7321Þj2

k¼0

þ j1 jð1:7321Þj2 Þ=3 ¼ ð1 þ 2ð4:0001ÞÞ=3 ﬃ 3:

&

# "

N1 X N1 1X X ½kH½mejð2p=NÞmn N k¼0 m¼0 " # N1 1X jð2p=NÞik jð2p=NÞim e e ; N i¼0

¼

N1 1X jX½kj2 : N k¼0

This can be shown using (8.10.36) with x½n ¼ h½n and is left as an exercise.

n¼0

" N1 N1 X 1X

x2 ½n ¼

DFT

x½ih½i þ nmodðNÞ

i¼0

N1 X

N1 X

N1 1X X ½kH½kejð2p=NÞnk : N k¼0

(8:10:38a)

(8:10:38b)

8.10.12 Zero Padding As mentioned earlier, computational complexity is significantly lower in the computation of DFT when N, the number of sample points in the data, is a power of 2, i.e., with the use of fast Fourier transform (FFT) algorithms discussed in the next chapter. This brings up the interesting question, what is the effect of adding zeros to the end of a sequence? Let the sequence have N1 sample points and let N2

360

8 Discrete-Time Signals and Their Fourier Transforms

zeros be added resulting in N ¼ N1 þ N2 sample points. Noting that the DFT spectrum is periodic with period 2p, the sample points are now spaced ð2p=ðN1 þ N2 ÞÞ instead of ð2p=N1 Þ apart. That is, as more zeros are added, DFT provides closer spaced samples of the transform of the original sequence. We should note that we do not have any more frequency information content than before. It gives a better display. Also, by appropriately padding a required number of zeros ðN2 Þ so that N ¼ N1 þ N2 is a power of two, fast DFT algorithms can be used.

8.10.13 Signal Interpolation In Chapter 1 and in an earlier part of this chapter we have made use of different interpolation functions. In this chapter we have discussed using the sinc and other functions to find interpolated values of the sampled signal. We can make use of the idea of zero padding in the frequency domain using the DFT, which is the dual of improving the spectral resolution by zero padding in the time domain discussed in the last section. Since the sampling frequency is fs ¼ 1=ts , increasing the sampling rate reduces the sampling interval, which, in turn, increases the number of samples in the interval. Let fs1 be the sampling rate used to determine N sampled values. Increasing the sampling rate from fs1 to Mfs1 would introduce interpolated values between samples. Procedure: The sample sequence with Nsample points with even and odd cases by 1 x½n : x½0; x½1; x½2; x½ ðN 1Þ; 2 :::; x½N 1; N odd ; 1 x½n : xd ½0; x½1; x½2; x½ N; 2 :::; x½N 1; N even :

N-odd: Form the MN point DFT Y½k as 1 Y½k : X½0; X½1; X½2; :::; X½ ðN 1Þ; 2 1 ððMN NÞ zerosÞ;X½ ðN þ 1Þ; . . . ; X½N 1: 2 (8:10:41c) N-even:Form the MN point DFT Y½k as 1 1 Y½k : X½0; X½1; X½2; :::; X½ N; 2 2 ððMN N 1Þ zerosÞ; 1 1 1 1 X½ N; X½ N þ 1; . . . ; X½N 1: 2 2 2 2

3. Determine IDFT [Y½k] to obtain the MN point sequence y½n, which may be complex. Since x½n is a real sequence, use only Re fy½ng and multiply by M. Example 8.10.5 Use the above method to interpolate the two sequences given below using the factor M ¼ 1 in the above procedure assuming the cases N ¼ 3 and 4. a: x½0 ¼ 0; x½1 ¼ 1; x½2 ¼ 2; b: x½0 ¼ 0; x½1 ¼ 1; x½2; x½3 ¼ 3: Solution: With the steps given above, the following results:

a. N =3: x½n : 0; 1; 2;

X½k : 3; 1:5 þ j:866; 1:5 j:866

Y½k : 3; 1:5 þ j:866; 0; 0; 0; 1:5 j:866 ! xint ½n : 0; 0; 1; 2; 2; 1 b. N = 4: x½n : 0; 1; 2; 3;

(8:10:41a)

(8:10:41d)

X½k : 6; 2 þ j2; 2; 2 j2

Y½k : 6; 2 þ j2; 12ð2Þ; 0; 0; 0; 12ð2Þ; 2 j2 xint ½n : 0; :0858; 1; 1:5; 2; 2:9142; 3; 1:5

(8:10:41b)

1. Take the DFT of the given sequence. DFTfx½ng ¼ X½k. 2. Insert zeros in the middle of the DFT sequence to create a MN point DFT. The cases for N even and odd are handled differently.

Note that x½k ¼ xint ½2 k; k ¼ 0; 1; 2; . . . ; N 1. The interpolated values are the values in between. Note that in the second case x½0 ¼ 0; x½3 ¼ 3; and x½4 ¼ 0 indicating that the interpolated value at xd;int ½7 will be the average value between 0 and 3, which is equal to 1.5. Similar arguments can be & given for the odd case.

8.10 Discrete Fourier Transform Properties

361

Notes: If a band-limited signal is sampled at a rate higher than the Nyquist rate, then the interpolated sequence will be exact at the sampling

intervals and the values between the samples will be interpolated values. In the case of periodic band-limited signals, the interpolation is exact.

Table 8.10.1 Discrete Fourier transform (DFT) properties Linearity: x½n ¼

M P

ai xi ½n

M DTFT P

!

i¼1

i¼1

ai Xi ½k ¼ X½k; ai 0 s are constants:

Time shift or delay: x½n i modðNÞ

DTFT

! X½kejð2p=NÞik :

Frequency shift: x½nejð2p=NÞni

DTFT

! X½k i modðNÞ :

Time reversal: x½n modðNÞ

DTFT

! X½k modðNÞ :

Alternate inversion formula: N1 P X ½kejð2p=NÞ : x½n ¼ N1 k¼0

Conjugation: x ½n

DTFT

! X ½k modðNÞ :

Duality: X½n

DTFT

! Nx½kmodðNÞ :

Circular convolution and correlation: N1 P

x½nh½n i modðNÞ ¼ x½n h½n modðNÞ

DTFT

! X½kH½k:

i¼0 N1 P i¼0 N1 P i¼0

x½ih½n þ imodðNÞ x½ix½n þ i modðNÞ

DTFT

! X ½kH½k:

DTFT

!jX½kj2 :

Multiplication: x½nh½n

DTFT

! N1 ½X½k H½kmodðNÞ ¼ N1

Real sequences: x½n ¼ xe ½n þ x0 ½n xe ½n

DTFT

! A½k;

x0 ½n

DTFT

! A½k þ jB½k:

DTFT

! Bd ½k:

Parseval’s theorem: N1 P n¼0

jx½nj2 ¼ N1

N1 P k¼0

jX½kj2 :

N1 P i¼0

x½iH½k i modðNÞ :

362

In other cases the interpolation can be poor. If the signals are not band limited, then the interpolation will be obviously poor. For x½n real, the discrete transform coefficients satisfy the conjugate symmetry property, X½N k ¼ X ½k: If the procedure for insertion of the zeros discussed earlier is followed for the interpolation, the conjugate symmetry will be preserved in Y½k. That is, Y½MN k ¼ Yd ½k: IDFT of Y½k will result in a & real sequence, see Ambardar (1995).

8.10.14 Decimation Decimation is an inverse operation of interpolation. It reduces the number of samples by discarding M 1 samples and retaining every M th sample. Note that the corresponding new sampling rate must be above the Nyquist sampling rate to avoid aliasing. This to be of any value, the original signal is assumed to be oversampled.

8.11 Summary

8 Discrete-Time Signals and Their Fourier Transforms

Zero-padding, interpolation, and decimation associated with discrete-time signals

Tables of properties associated with discrete Fourier transforms.

Problems 8.2.1 Consider the function xðtÞ ¼ cosðo0 tÞ. Illustrate the aliasing phenomenon by decreasing os , or equivalently, increasing the sampling interval. Use a low-pass filter of bandwidth equal to ðos =2Þ. In your solution use the following steps. Work out the solution using os > 2o0 and show that the cosine function is recoverable. Now reduce the sampling frequency such that os 52o0 . Sketch the spectrum of the ideally sampled signal and show that the signal exists in the frequency range 05ðos o0 Þ5os =2. 8.2.2 Given xðtÞ is band limited to os =2, determine the Nyquist rates for the functions. a: ya ðtÞ ¼ dxðtÞ dt ; b: yb ðtÞ ¼ xÐt2 ðtÞ; xðaÞda; c: yc ðtÞ ¼ 1

This chapter started with analog signals that are sampled to obtain discrete-time signals. Fourier analysis of discrete time limited signals is discussed in terms of discrete-time and discrete Fourier transforms. The following gives a list of some of the specific topics:

Ideal sampling of a continuous signal Continuous Fourier transforms of the sampled signals

Low-pass and band-pass sampling theorems Basic discrete-time signals and operations, including decimation and interpolation

Basic concepts of discrete-time convolution and correlation

Discrete-time periodic signals and the corresponding discrete Fourier series and their properties

Derivation of the discrete-time Fourier transform Properties of the discrete-time Fourier transform Discrete Fourier transforms and the inverse discrete Fourier transforms

Periodic convolution and correlation and their computations directly and through DFT

8.2.3 Consider the function xðtÞ ¼ cosðo0 t þ yÞ; f0 ¼ o0 =2p ¼ 200 Hz. From the low-pass sampling theorem we know that there will not be any aliasing if os > 2o0 . Now consider that xðtÞ is sampled at two different frequencies one below and one above the Nyquist frequency given by a: fs ¼ 600 Hz; b: fs ¼ 160 Hz. In the first case, we know that there will not be any aliasing. In the second case, the signal xðtÞ sampled at fs ¼ 160 Hz describes a cosine function that is not the given function, but a sampled version of some other cosine function. Give the corresponding function xa ðtÞ ¼ A cosð2pfa t þ yÞ. That is, find fa . Sketch the two functions xðtÞ and xa ðtÞ on the same figure and identify the points where the two functions coincide. (xa ðtÞ ¼ Aliased version of xðtÞ). 8.2.4 The acoustic pulse received by a receiver is represented by xðtÞ ¼ Asinc2 ðo0 tÞ. Noting the transform of this function is a triangular function, give the minimum sampling rate, the expression for the spectrum of the ideally sampled signal, and the minimum band width of the ideally low-pass filter required to reconstruct xðtÞ from the sampled signal.

Problems

363

8.2.5 Find the minimum sampling rate that can be used to determine the samples that completely specify the following signals by assuming ideal sampling:

8.3.2 Find the even and odd parts of the functions. a: xa ½n ¼ u½n; b: xb ½n ¼ ð1=2Þn u½n. 8.3.3 Let x½n ¼ xe ½n þ x0 ½n. Show

a. x1 ðtÞ ¼ ½sinð2pð100ÞtÞ=ð2pð100ÞtÞ; b. b: x2 ðtÞ ¼ cosð2pð100Þt þðp=3ÞÞ þ sinð2pð200ÞtÞ 8.2.6 A signal xðtÞ is band limited to the range f0 5f5500 Hz. Find the minimum sampling rate for xðtÞ without aliasing assuming a. f0 ¼ 0; b. f0 ¼ 100 Hz 8.2.7 The signal xðtÞ ¼ A cosð2pð100ÞtÞ is sampled at 150 Hz. Describe the corresponding signal after the sampled signal is passed through the following filters: a. An ideal low-pass filter with a cut-off frequency of 20 Hz b. An ideal band-pass filter with a pass band between 60 Hz and 120 Hz 8.2.8 Consider the sampled sequence xð0Þ ¼ 1; xðts Þ ¼ 0; xð2ts Þ ¼ 1; xðnts Þ ¼ 0; n 6¼ 0; 1; 2. Sketch the interpolated functions using a. step, b. linear, and c. sinc interpolations. 8.2.9 Let F½xðtÞ ¼ XðjoÞ with XðjoÞ ¼ 0; joj2pB. Using the results in Section 8.2.3, show 1 ð 1

1 X

1 x ðnts Þ; ts ¼ ; fs ts ¼ 1: jxðtÞj dt ¼ ts 2B n¼1 2

2

8.2.10 Use the band-pass sampling theorem to determine the possible sampling rates so that the following signal can be recovered from the sampled signal:

o þ oc o oc þP ; xðtÞ ! XðjoÞ ¼ P 2pð2BÞ 2pð2BÞ FT

B ¼ 8 kHz; oc ¼ 2pfc ¼ fc ¼ 64 kHz: Assuming the sampling rates of a. fsa ¼ 200 kHz; b. fsb ¼ 20 kHz; and c. fsc ¼ 16 kHz, illustrate how the signal can be recovered from the sampled signals if possible. 8.3.1 Sketch the following sequences assuming x½n ¼ ð1 nÞfu½n u½n 3g: a: ya ½n ¼ x½2n 1; b: yb ½n ¼ x½n2 1; c: yc ½n ¼ x½1 n:

1 X

E¼

x2 ½n ¼

n¼1

1 X

x2e ½n þ

n¼1

1 X

x20 ½n:

n¼1

8.3.4 Derive the following identities and then simplify the results when N ! 1: N1 X

a: S ¼

an ¼

n¼0

b: c:

N 1 X

nan ¼

n¼0 1 X

1 aN ; jaj51; 1a

ðN 1ÞaNþ1 NaN þ a ð1 aÞ2

eajnj ¼

n¼1

;

1 þ ea : 1 ea

8.3.5 Find the closed form expression for y½n ¼ an u½n u½n; jaj51. 8.3.6 Find the cross correlation of the two sequences given by x½n ¼ u½n u½n nx and h½n ¼ u½n u½n nh for the cases: a: nx ¼ nh ¼ 2; b: nx ¼ 2; nh ¼ 3. 8.4.1 Determine the DTFS of the following sequences by using Euler’s theorem and then by identifying the discrete Fourier series coefficients. Identify the periods. a: xs a ½n¼1 þ sinðpn=2 þ yÞ; b: xsb ½n¼ cosðnp=20Þþ sinðnp=40Þ; c. xsc ½n ¼ cos2 ½np=8 8.4.2 Find the DTFS coefficients of the N-periodic discrete-time functions 1 P

a. xsa ½n ¼

d½n lN;

l ¼1

b. xb ½n ¼

1; 0 jnj M . 0; M5n5N M

8.4.3 Determine the time-domain sequences with period N ¼ 7 with the DTFS coefficients a. Xsa ½k ¼ ð1=2Þ; b. Xsb ½k ¼ cosð2 kp=NÞ:

364

8 Discrete-Time Signals and Their Fourier Transforms

8.4.4 a. Show that Xs ½k ¼ Xs ½N k for the following sequence: 1; 0 n5ðN 1Þ=2 xs ½n ¼ 0; ðN 1Þ=2 þ 1 n N 1; xs ½n ¼ xs ½n þ N: b. Given the periodic sequences xs ½n ¼ f0; 0; 1; 2g; hs ½n ¼ f1; 2; 0; 0g;

8.5.7 Verify the results given in Section 8.5.2 for Type 1 and 4 sequences. 8.6.1 Prove the time reversal property in (8.6.17). 8.6.2 a. Determine the DTFT of the function x½n ¼ ð1=3Þu½n: b. Use the time reversal property to determine the DTFT of ð3Þn u½n. 8.6.3 Find the DTFT of the function

xs ½n ¼ xs ½n þ 4 and hs ½n ¼ hs ½n þ 4: find the DTFS of the function ys ½n ¼ xs ½nhs ½n. Illustrate the generalized Parseval’s identity by using the DTFS of the functions xs ½n; hs ½n and ys ½n.

y½n ¼ ðn þ 1Þ2 x½n: 8.6.4 Determine the convolutions x1 ½n x2 ½n for the following cases:

8.4.5 Use the sequences in Problem 8.4.4b to determine ys ½n ¼ xs ½n hs ½n.

a: x1 ½n ¼ u½n; x2 ½n ¼ u½n;

8.4.6 Give an example of two sinusoidal sequences that are equal. Hint: Assume cosðO1 pk þ yÞ ¼ cosðO2 pk þ yÞ with O1 6¼ O2 and show the two functions are equal.

8.6.5 Determine a:

8.5.1 Show that x ½n

b: x1 ½n ¼ u½n; x2 ½n ¼ :5n u½n:

1 X n¼1

! X ðejO Þ and x ½n

! X ðejO Þ:

DTFT

DTFT

8.5.2 Derive an expression for the convolution y½n ¼ x½n x½n; x½n ¼ an u½n.

ð1=2Þjnj ; b:

1 X

nð1=2Þn :

n¼0

by using the central ordinate theorems. 8.8.1 Find the DFTs of the sequences a: fx½ng ¼ ½1; 1; 1; 1; b: fx½ng ¼ ½1; 1; 1; 1

8.5.3 Show that DFTf:5d½n þ :25d½n 2 þ :25d½n þ 2g ¼ cos2 ðOÞ: 8.5.4 Find the inverse transform of XðejO Þ ¼ 1; jOj Oc ; XðejO Þ ¼ 0; Oc 5jOj p: 8.5.5 Consider the two-sided sequence jnj 5 x½n ¼ a ; jaj 1. Write this expression in terms of the right-side and left-side sequences. Then, derive the expression for the DTFT of this sequence using the time reversal property. Be careful about the sample point at n ¼ 0: 8.5.6 Use the central ordinate theorems to evaluate the sums. a:

1 X n¼0

nan ; b:

1 X n¼1

ajnj ; c:

1 X n¼1

sin

WnÞ ðpnÞ

8.8.2 Compute the DFTs of the following N- point sequences. For Part c., use Euler’s formula for the cosine function in determining the DFT assuming k0 is an integer. a: x½n ¼ an ; 0 n5N; b: x½n ¼ u½n u½n n0 ; 05n0 5N; c: x½n ¼ cosðno0 Þ; o0 ¼ 2pk0 =N; 0 n5N 1; k0 is an integer. 8.8.3 Determine the 8-point DFT sequence of x½n ¼ d½n þ 2d½n 3. 8.8.4 Consider a sequence x½n; 0 n N 1 with X½k ¼ DFTfx½ng. Find the DFTs the two sequences given below in terms of X½k:

Problems

365

fy½ng ¼ ¼

bandwidth of xðtÞ as the frequency where jXðjoÞj is 10% of its maximum.

x½n=2; n even ; fy½ng 0; n odd x½n; n ¼ 0; 1; 2; . . . ; N 1 0; n ¼ N; N þ 1; . . . ; ð2 N 1Þ

:

8.8.5 Compute the DFT of x½n 3 mod ðNÞ directly and then using the time-shift theorem. 8.8.6 Determine y½n ¼ x½n h½n directly and then using the DFT for the sequences x½n ¼ ð1=3Þn ; h½n ¼ sinððp=2ÞnÞ; n ¼ 0; 1; 2; 3: 8.8.7 Show that the DFT of a real sequence x½n satisfies the relation X½N k ¼ X ½k: (*) denotes conjugation. 8.8.8 The DFT sequence of a real time signal is given by fX½kg ¼ f4; j; 0; Xg, where X is the missing value. Use the symmetry property of DFT to determine the missing value. Find the corresponding time sequence. 8.8.9 Derive an expression DFT½ð1Þn x½n in terms of X½k. 8.9.1 Show that for N ¼ 4, (an identity matrix)

DFT

½y½n ¼

ð1=NÞADFT ADFT

¼ IN

ð1=NÞA2DFT

8.9.2 Derive the matrix with N ¼ 4. What can you say about this matrix? 8.9.3 Given xðtÞ ¼ LðtÞ, estimate the sampling frequency and sampling interval by choosing the

8.10.1 Use Example 8.10.3 to compare the number of multiplications required to compute the convolution directly and by using the DFT. 8.10.2 Write the sequence r½n in matrix form r½n ¼

N1 X

x½ix½n þ iModðNÞ :

i¼0

8.10.3 Consider the two discrete N-point real sequences xd1 ½n and xd2 ½n and x½n ¼ x1 ½nþ jx2 ½n; n ¼ 0; 1; . . . ; N 1 with F½xi ½n ¼ Xi ½k; i ¼ 1; 2 and ðxd1 ÞT ¼ ½ c 0123; ðxd2 ÞT ¼ ½ c 2345; xd ¼ xd1 þ jxd2 : a. First show the following in general terms and then b. verify this using the sequences: X1 ½k ¼ :5fXd ½k þ X ½N kg; X2 ½k ¼ :5jfX½k X ½N kg: 8.10.4 Find the N point DFT of the sequences x1 ½n ¼ ejO0 n for two cases: a: O0 ¼ 2pk0 =N; b: O0 6¼ 2pk0 =N:ðk0 is an integerÞ: 8.10.5 Consider the sequence xd ½n ¼ f0; 1; 0; 1g. Compute its DFT and then use the interpolation technique discussed in Section 8.10 assuming M ¼ 2 and 4.

Chapter 9

Discrete Data Systems

9.1 Introduction In the last chapter we have discussed the concepts of discrete Fourier transforms (DFTs). In this chapter we will briefly review these and discuss its fast implementations. There are several algorithms that come under the topic-fast Fourier transforms (FFTs). The first FFTmethod of computing the DFT was developed by Cooley and Tukey (1965). These are innovative and useful in the signal processing area. Continuous Fourier transforms (CFTs) in the analog and the discrete Fourier transforms (DFTs) in the discrete domains are the corner stones of signal analysis. In the continuous domain we studied the Laplace transforms, which are related to the continuous Fourier transforms. The discrete counter part of the Laplace transforms is z-transforms related to the discrete-time Fourier transforms (DTFTs). Table 9.1.1 summarizes the variables in the continuous-time Fourier transforms, the Laplace transforms, the discrete-time Fourier transforms, and the z-transforms. In this chapter we will study some of the basics associated with the z-transforms and its applications. Digital filters have been popular in recent years and will continue to be in the future. In its simplest

Table 9.1.1 Discrete-time and continuous-time signals and their transforms Continuous-time transform/ Discrete-time transform/ variable variable Continuous Fourier transform/ o or f Laplace transform/s

Discrete time Fourier transform/O z-transform/z

form, a digital filter is a computer program that takes a set of data and converts into another set of data. Discrete data systems may correspond to filtering or some other operation. In the analog case we have to worry about component value tolerances and the responses can change in time. The responses of analog systems cannot be duplicated, as the component values may be different from one batch to another. The responses of the filters can change if the operating conditions of the filter change. On the other hand, in the digital case, every time we process a set of data the output will be the same. Digital filters are more flexible and can be altered by simply changing the computer code. At low frequencies, analog components are bulky. We may have to deal with magnetic coupling if inductors or transforms are used as components in the analog system. Analog filters may have to be redesigned and the circuit implementations may be different if the frequencies change. On the other hand, modifying digital filters may represent a change of computer code. Digital technology is modern and powerful signal processing algorithms can be designed. Digital filters can be time shared and process several signals simultaneously. Digital integrated circuits design is much simpler compared to analog integrated circuit technology. They require lower power consumption and the digital circuitry can be fabricated in smaller packages. Digital storing is much cheaper. Searching and selecting digital information is simple and processing the data is straightforward. Digital reproduction is much more reliable and the cost of digital hardware continues to come down every year. Most source signals and the recipients are analog in nature. To replace an analog filter by a digital filter, the analog signal

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_9, Springer ScienceþBusiness Media, LLC 2010

367

368

9 Discrete Data Systems

needs to be converted to a digital signal by using an analog-to-digital (A/D) converter. The digital signal is then passed through a digital filter and the output of the filter needs to be converted back to analog data using a digital-to-analog (D/A) converter. Digital signal processing (DSP) area has been popular during the past 30+ years. It will continue to be of interest in many areas, including seismic signal processing, speech processing, image processing, radar signal processing, and others. Telephone industry has taken the lead in the signal processing area. There are excellent texts available in the general area of signal processing. Some of these include Ambardar (2007), Strum and Kirk (1988), Mitra (1998), Oppenheim and Schafer (1975), Rabiner and Gold (1975), Cartinhour (2000), Ludeman (1986), and many others. For an excellent review on the spectral analysis, see Otnes and Enochson (1972), Marple, (1989), Press et al. (1989), and many others. For a historical survey on the spectral estimation, see Robinson (1982). MATLAB provides digital analysis and design software, see Ingle and Proakis (2007). Also, see Ramirez (1975), Smith III (2007), Smith, (2002) on FFT and its applications.

1 jXð0Þj2 ; N2 1 Pðos =2Þ ¼ 2 jXðN=2Þj2 ; N os ¼ 2pfs ; fs ¼ 1=ts i 1 h Pðok Þ ¼ 2 jX½kj2 þjX½N kj2 ; N 2pk N ; k ¼ 1; 2; :::; 1: ok ¼ Nts 2 Pð0Þ ¼

(9:2:1b)

From Chapter 4 we note that a rectangular window spectrum has a great deal of leakage into the side lobes. A tapered window w½n, such as a Hamming window to be discussed later, can be used in estimating the spectrum to reduce the spectral leakage. A windowed signal y½n ¼ x½nw½n is to be used in the estimation. Another popular method of spectral estimation is the Blackman–Tukey method, see Press et al. (1990). In its simplest form, it involves the computation of the data autocorrelation and then determining the spectrum using DFT. The spectrum of the autocorrelation is the power spectral density.

9.2.1 Symbolic Diagrams in Discrete-Time Representations 9.2 Computation of Discrete Fourier Transforms (DFTs) Power spectrum: Most signals in practice are analog signals. Spectral analysis and estimation of these signals is basic. A simple method of power spectrum estimation of an analog signal xðtÞ involves N values of xðtÞ sampled every ts s (or fs ¼ 1=ts samples/s) resulting in x½n ¼ xðnts Þ; n ¼ 0; 1; 2; . . . ; N 1. The DFT of the signal x½n is (see Section 8.9) X½k ¼

N1 X

Symbolic diagrams or signal flow graphsare a network of directed branches connected at nodes is a pictorial representation of an algorithm. Figure 9.2.1 gives the flow graph symbols that are common in two different forms. Source nodes do not have any incoming braches and are used for input x[n]

+

x[n]

y[n]

×

ax[n]

x[n] z–1

x[n −1]

a

(a)

(b) x[n] + y[n]

x[n]

x½nejð2p=NÞnk ;

x[n] + y[n]

x[n]

a

(c) ax[n]

x[n]

z −1 x [n−1]

n¼0

k ¼ 0; 1; 2; . . . ; N 1:

(9:2:1a)

y[n]

(d) The power spectrum estimate is defined at ðN=2Þ þ 1 frequencies by

(e)

(f)

Fig. 9.2.1 Two flow graph representations: (a) and (d), summers; (b) and (e), multipliers; (c) and (f), delays

9.2 Computation of Discrete Fourier Transforms (DFTs) Fig. 9.2.2 Example 9.2.1

w[n]

+

x

x[n]

369

+

w[n] y[n]

x[n]

a

−1

z

a

x

y[n] −1

z

x c

b

b

(b)

9.2.2 Fast Fourier Transforms (FFTs)

sequences. Sink node has only one entering branch and is used for the output sequence. In addition, summers and multiplier symbols are shown and are self-explanatory. The symbols z1 are used to identify delay components. In Section 9.4 z-transforms will be studied. Some authors use the multiplier constant above the line and others use it below the line. If the multiplier constant a is not shown, then it is 1.

First a brief review of the discrete Fourier transform (DFT) is given below. The discrete Fourier transform is a transformation that takes a set of N values in time to N values in frequency. First, the transform vector is given by X ¼ ADFT x (see (8.9.16a and b)), where the matrix ADFT is a N N matrix with its ðk; nÞ entry being (see (8.9.17))

Example 9.2.1 Using the symbol representations in Fig. 9.2.1, write difference equations relating the variables w½n and y½n in terms of x½n and w½n 1 in Fig. 9.2.2a,b.

ADFT ðk; nÞ ¼ ejð2p=NÞðk1Þðn1Þ ðk1Þðn1Þ

WN

;

1 k; n N; WN ¼ ej2p=N : (9:2:2a)

Solution: The two diagrams in Fig. 9.2.2 result in the same equations and are given as follows: w½n ¼ ax½n þ bw½n 1;

(a)

c

The vectors X and x are N-dimensional column vectors. In matrix form the DFT coefficients can be expressed in terms of WNn ¼ ðej2p=N Þn by

y½n ¼ w½n þ cw½n 1: &

2 6 6 6 6 6 6 6 6 4

2

3

X½0 X½1

1

1

7 6 1 W1N 7 6 7 6 X½2 7 6 1 W2N 7¼6 6 7 : : : 7 6 7 6 5 6 : : : 4 N1 X½N 1 1 WN

1

:

:

W2N

:

:

W4N :

: :

: :

: :

: :

: :

Note WNn takes one of the values in the set jð2p=NÞ ; ejð2p=NÞ2 ; :::; ejð2p=NÞðN1Þ for any N, 1; e see (8.9.18).

32 3 x½0 7 N1 WN 76 x½1 7 7 76 7 6 2ðN1Þ 76 WN 76 x½2 7 7: 76 7 76 : : 7 76 7 74 5 : : 5 ðN1Þ2 x½N 1 WN 1

The properties in Table 9.2.1 allow for the derivation of a fast Fourier transform (FFT) algorithm. We will consider an Nð¼ 2n Þ-point decimation-in-

Table 9.2.1 Properties of the function WN ¼ ejð2p=NÞ 1. WNnþN ¼ ejð2p=NÞðnþNÞ ¼ ejð2p=NÞn ¼ WNn 2.

nþN=2 WN

3.

WNkN

¼ e

¼e

jð2p=NÞ

jð2pÞk

¼ 1;

¼

WNn

2 A ¼ 1; e

;e

(9.2.3a) (9.2.3b) (9.2.3c)

k is an integer

W2Nk ¼ ej 2ð2p=NÞk ¼ ej;ð2p=ðN=2ÞÞk ¼ WkN=2 jð2p=NÞ jð2p=NÞ2 jð2p=NÞn jð2p=NÞðN1Þ

4. 5. e

(9:2:2b)

; :::; e

(9.2.3d) (9.2.3e)

370

9 Discrete Data Systems

frequency FFT algorithm. It is based on expressing one N-point DFT algorithm by two N/2-point DFTs, then four (N/4)-point DFTs, and so on. The algorithm at the end reduces to ðN=2Þ-2-point transforms. The 1-point transform is trivial as X½0 ¼ x½0: The DFT of a 2-point sequence is determined by noting W12 ¼ ej2p=2 ¼ 1: The DFT values in scalar and matrix form are as follows: X½0 ¼ x½0W02 þ x½1W02 ¼ x½0 þ x½1 ; X½1 ¼ x½0W02 þ x½1W12 ¼ x½0 x½1

X½0 1 ¼ X½1 1

1 1

x½0 : x½1

(9:2:4a)

N1 X

x½nejð2p=NÞnk

n¼0

¼

N1 X

x½nWNnk

(9:2:5)

n¼0

¼

N=21 X

x½nWNnk þ

n¼0

N1 X

x½nWNkn ¼

n¼N=2

N=21 X m¼0

n¼N=2

¼

N=21 X m¼0

X½k ¼

x

N=21 X n¼0

N kðmþNÞ þ m WN 2 2

N kðN=2Þ þ m WNmk WN ; (9:2:6) x 2

N x½n þ ð1Þk x n þ WNkn ; 2

k ¼ 0; 1; 2; :::; N 1:

N nð2 kÞ WN x½n þ ð1Þ x n þ X½2 k ¼ 2 n¼0 N=21 X N Wnk ¼ x½n þ x n þ (9:2:8a) N=2 ; 2 n¼0 N=21 X N nð2kþ1Þ X½2k þ 1 ¼ x½n þ ð1Þð2kþ1Þ x n þ WN 2 n¼0 2k

N nk WN=2 x½n x n þ WNn ; 2 n¼0 (9:2:8b) N k ¼ 0; 1; 2; . . . ; 1: 2 ¼

N=21 X

One N-point DFT is reduced to two ðN=2Þ-point DFTs. One N-point DFT requires N2 multiplications and ððN 1Þ additions), see (9.2.2b). Two ðN=2Þ-point DFTs require only 2ðN=2Þ2 multiplications and 2ðN 1Þ additions. Example 9.2.2 Assuming N ¼ 4, show that the use of (9.2.8a and b) successively results in the DFT values. Illustrate the algorithm using the flow graph representation. Solution: a. From (9.2.8a) and (9.2.8b), we have (note W40 ¼ 1; W14 ¼ j; and W42 ¼ 1)

x½nWNnk :

Note the variable n is used for the time variable and k is used for the frequency variable Using m ¼ n ðN=2Þ in the second summation in (9.2.5) kðN=2Þ ¼ ejð2p=NÞkðN=2Þ ¼ ð1Þk result and noting WN in N1 X

N=21 X

(9:2:4b)

Decimation-in-frequency FFT algorithm: Starting with the DFT of a set of data x½n; n ¼ 0; 1; ::; N 1 and N ¼ 2n , the DFT coefficients (see 9.2.1a.), X½k, are obtained in terms of WN ¼ ej2p=N as follows: X½k ¼

Now separate the coefficients into X½2 k and X½2 k þ 1 and use Table 9.2.1:

X½2 k ¼

1 X

fx½n þ x½n þ 2gW2nk 4 ;

n¼0

X½2 k þ 1 ¼

1 X

fx½n x½n þ 2gWn4 W2nk 4 ; k ¼ 0; 1:

n¼0

(9:2:9) First, at stage 0, i.e., to start with, define x0 ½n ¼ x½n; n ¼ 0; 1; 2; 3: At stage i, identify the variables as xi ½n. Algorithm has two stages corresponding to N ¼ 4 ¼ 2n ; n ¼ 2. From (9.2.9), we have the following. Direct: X½0 ¼ fx0 ½0 þ x0 ½2g þ fx0 ½1 þ x0 ½3g

(9:2:7)

¼ fx1 ½0g þ fx1 ½1g ¼ x2 ½0 ¼ X½0 ; (9:2:10a)

9.2 Computation of Discrete Fourier Transforms (DFTs)

X½2 ¼ fx0 ½0 þ x0 ½2g fx0 ½1 þ x½3g ¼ fx1 ½0g fx1 ½1g ¼ x2 ½1 ¼ X½2 ; (9:2:10b) X½1 ¼ fx0 ½0 x0 ½2g þ fW 40 ðx0 ½1 x½3Þg ¼ fx1 ½2g þ fx1 ½3g ¼ x2 ½2 ¼ X½1 ;

(9:2:10c)

X½3 ¼ fx0 ½0 x0 ½2g fW14 ðx0 ½1 x0 ½3Þg ¼ fx1 ½2g fx1 ½3g ¼ x2 ½3 ¼ X½3 : (9:2:10d) Individual identifications at each stage from (9.2.10): Stage 0: x0 ½0 ¼ x½0; x0 ½1 ¼ x½1; x0 ½2 ¼ x½2:

(9:2:11a)

Stage 1: x1 ½0 ¼ fx0 ½0 þ x0 ½2g ; x1 ½1 ¼ fx0 ½1 þ x0 ½3g ; x1 ½2 ¼ fx0 ½0 x0 ½2g ; x1 ½3 ¼ fW04 ðx0 ½1 x½3Þg : (9:2:11b) Stage 2: x2 ½0 ¼ fx1 ½0g þ fx1 ½1g ; x2 ½1 ¼ fx1 ½0g fx1 ½1g ; x2 ½2 ¼ fx1 ½2g þ fx1 ½3g ; x2 ½3 ¼ fx1 ½2g fx1 ½3g : (9:2:11c) End results: X½0 ¼ x2 ½0; X½2 ¼ x2 ½1; X½1 ¼ x2 ½2; X½3 ¼ x2 ½3 : (9:2:11d)

Fig. 9.2.3 Flow graph representations for N ¼ 4 using the decimation-infrequency FFT algorithm

371

These equations can be used to draw the flow graph using the symbols in Fig. 9.2.3. For clarity, the multipliers are shown under the lines rather than above. Interestingly, the above equations can be seen from the flow graph in Fig. 9.2.3. Interestingly, if the variables arein binary form, the argument k in X½k ¼ X½ðk1 k0 Þ2 is related to the argument n in x2 ½ðn1 n0 Þ2 by the relation k ¼ ðk1 ¼ n0 ; k0 ¼ n1 Þ: For additional information & on this, see Oppenheim and Schafer (1999). The above results can be extended for any N ¼ 2n with n stages. Figure 9.2.4 gives the flow graph for N ¼ 8 ¼ 23 . Note the multipliers are identified above and below the lines for clarity. For a general derivation of the decimation-infrequency algorithm and other algorithms, see Oppenheim and Schafer (1999), and others. Notes: The decimation refers to the process of reducing the number of operations for an N ¼ 2n point DFT, expressing the N-point DFT in terms of 2 ðN=2Þ ¼ 2n1 -point DFTs and successively expressing them in n stages with the input sequence in natural order. Number of computations in an FFT algorithm: In Section 8.9.2 the computational aspects of discrete Fourier transforms were considered. These results are compared with FFT computational requirements. In the N-point FFTalgorithm with N ¼ 2n , we have n ¼ log2 ðNÞ stages. FFT computation requires ðN=2Þn ¼ ðN=2Þ log2 ðNÞ complex multiplications and nN ¼ N log2 ðNÞ complex additions. Computers use real arithmetic and each complex multiplication requires four real multiplications and three real additions. The amount of effort to do multiplication is much larger than additions. We

372

9 Discrete Data Systems

Fig. 9.2.4 Flow graph representations for N ¼ 8 using the decimation-infrequency FFT algorithm

can compare the number of multiplications by the direct method versus FFT by the ratio R¼

N2 N N ¼ : :5 N log2 ðNÞ :5n n

(9:2:12)

For a large N ¼ 2n , the difference in the number of computations by FFT is significantly lower. Note that :5 N log2 ðNÞ is nearly linear, whereas N2 is quadratic. For small N, the difference in the number of computations in computing the DFT and FFTis not that significant. As an example, consider N ¼ 210 , the ratio in (9.2.12) is R 204. The DFT requires N2 values of Wkn N ; k; n ¼ 0; 1; 2; . . . ; N 1, whereas FFT requires at most N such values at each stage. Earlier, we have seen that ejð2p=NÞnk ¼ ejð2p=NÞm ; 0 m N 1. The logical way of course is to compute WNk once, k ¼ 0; 1; 2; . . . ; N 1, store them, and use them again and again in each stage. Only about ð3=4ÞN of these WkN are distinct in the FFT algorithm, see Ambardar (2007). The FFT approach N ¼ 2n is computationally efficient compared to the direct method only for n45 ðN432Þ. See Wilf (1986) for an interesting discussion of algorithms and their complexity. MATLAB function for computing the DFT of a signal is the fft function. It can be used for any N. For example, to compute the DFT of a sequence x of N values, MATLAB routine is X ¼ fftðxÞ to get the spectral values and the routine x ¼ ifftðXÞ gives the data from the spectral values.

Just like in the continuous case, other discrete transforms related to discrete Fourier transforms can be considered, including discrete cosine, sine, Hartley, and Hilbert transforms. These are beyond the scope here. See the handbook by Poularikas (1996).

9.3 DFT (FFT) Applications In Section 9.2, spectral analysis based on DFT was considered. Computing DFT via FFT is a tool to reduce the number of computations. FFT is applicable wherever DFT can be used. See, for example, Marple (1987), O’Shaughnessy (1987), Otnes and Enochson (1972), Poularikas (1996), Rabiner and Schafer (1979), Shenoi (1995), and others for FFT applications.

9.3.1 Hidden Periodicity in a Signal Although, nothing is forever, some signals can be considered as periodic at least on a short-time basis. For example, vowel speech sounds can be considered as periodic on a short-time basis. Investors in the stock market would like to know if the price of a stock has a periodic part in the signal that is hidden. If so, the investor can sell when the stock is high and buy when it is low. For a good presentation on

9.3 DFT (FFT) Applications

373

applying spectral analysis to various physical signals, see Marple (1987). Example 9.3.1 Consider the sinusoid xðtÞ ¼ cosð2pð100ÞtÞ that is sampled at twice the Nyquist rate for three full periods. Find the corresponding DFT values. Solution: The frequency of the sinusoid is 100 Hz. The period of the signal is T ¼ ð1=100Þ s. The Nyquist rate is 200 Hz. The corresponding sampling rate and the sampling interval are 400 Hz and ts ¼ ð1=400Þ s. Since three periods are used, we have four samples per period and have N ¼ 12 samples to find the DFT. Note cosð2pf0 tÞjt¼nts ¼ cosð2pnðf0 =fs ÞÞ; n ¼ 0; 1; 2; . . . ; N 1 ¼ 11

(9:3:1)

The sampling frequency is divided into N ¼ 12 intervals and the frequency interval is F ¼ fs =N ¼ fs =12 ¼ 400=12 ¼ 100=3 referred to as the digital frequency in Section 8.6.1, where O ¼ 2pðf=fs Þ ¼ 2pF was used. The DFT frequencies are kF ¼ kfs =N ¼ kðfs =12Þ; k ¼ 0; 1; 2; :::; N 1 ¼ 11:

(9:3:2)

Now, assume f0 is one of these frequencies and kF ¼ kðfs =NÞ ¼ kðfs =12Þ ¼ f0 for some k. That is, f0 k k ¼ ¼ ; a rational number: fs N 12

(9:3:3)

For f0 ¼ 100 Hz; fs ¼ 400 Hz, and k ¼ 3, X½3 gives the appropriate spectral value. This can be verified by computing DFT of the sequence x½n ¼ cosð2pðf0 =fs ÞnÞ ¼ cosð2pðknÞ=NÞ ¼ cosð2pðknÞ=12Þ; n ¼ 0; 1; 2; . . . ; 11; k ¼ 3 ) fx½ng ¼ f1; 0; 1; 0; 1; 0; 1; 0; 1; 0; 1; 0g: (9:3:4) DFTfcosð2pðnkÞ=NÞg ¼ 12 DFT ej2pnm=N þ ej2pnm=N ¼ 12

N1 P n¼0

ejð2p=NÞnðmkÞ þ 12

N1 P n¼0

: ejð2p=NÞnðmþkÞ ¼ X½k

Using the summation formula for the geometric series results in the DFT values

X½k ¼

8 > < > :

N=2; k ¼ m N=2; k ¼ N m 0; otherwise

9 > = > ;

) X½k : f0; 0; 0; 6; 0; 0; 0; 0; 0; 6; 0; 0g: (9:3:5) Noting N ¼ 12 and the amplitude of the sinusoid is & 1, we have X½3 ¼ X½9 ¼ 6. Notes: These results can be extended to a periodic function xðtÞ ¼ cosð2pðk0 =NÞ þ yÞ. The DFT of xðtÞ has only two nonzero discrete frequency values and are X½kjk¼k0 ¼ðN=2Þejy and X½N kjk¼k0 ¼ ðN=2Þejy : The frequency spacing F ¼ f0 =fs needs to be a rational function so that the discrete frequency falls on the input signal frequency. FFT algorithm can be used with this in mind. Also, leakage results in DFT if a periodic signal is not sampled for an integer number of periods. This results in nonzero spectral components at frequencies other than the harmonic frequencies of the signal. Many times it is not easy to find the period of a signal up front. One solution to this is to use large enough number of samples. Larger is the time interval, the more closely is the spectrum sampled. That is, the spectral spacing F ¼ fs =N is reduced, thus giving a more accurate estimate of spectrum of the given signal. Most practical signals are noise corrupted and are known only for a short time. That is, the signal is a windowed signal. As we have seen in Chapter 4, the spectrum of the windowed sinusoid is not a spike, but a sinc function indicating leakage into the side lobes. Tapered windows need to be used in any spectral analysis. If a signal contains several frequencies, then the spectrum of the windowed signal is the sum of the spectra of each sinusoid and it may not have distinct peaks as the main lobes of the sinusoids might merge. This happens if the two frequencies in the input signal are located close enough, then the frequency peaks may merge. Instead of two separate peaks, there may be only one single peak. Choosing larger

374

9 Discrete Data Systems

DFT lengths and good windows improves the accuracy of the spectral estimates, see Marple (1987). &

9.3.2 Convolution of Time-Limited Sequences The convolution of the sequences x½n and h½n was defined by (see (8.3.16)) y½n ¼

1 X

x½kh½n k ¼

k¼1

1 X

h½mx½n m:

Consider x[n] = 0 for n < 0 and n L, h[n] = 0 for n < 0 and n M. Equation (9.3.6) can be expressed as follows: L1 X

x½kh½n k ¼

k¼0

(9:3:10) Solution: 2 3 y½n ¼ 0; n5 0 6 7 6 y½0 ¼ x½0h½0 þ x½1h½1 ¼ x½0h½0 ¼ 1 7 6 7 6 7 6 y½1 ¼ x½0h½1 þ x½1h½0 ¼ 1 þ 1 ¼ 0 7: 6 7 6 7 6 y½2 ¼ x½0h½2 þ x½1h½1 ¼ x½1h½1 ¼ 1 7 4 5 y½n ¼ 0; n 3 (9:3:11)

m¼1

(9:3:6)

y½n ¼

x½0 ¼ 1; x½1 ¼ 1; x½n ¼ 0 for n50 and n41:

M 1 X

h½mx½n m: (9:3:7)

m¼0

The equations in (9.3.11) for y½n 6¼ 0 can be written in two equivalent matrix forms and 2

y½0

3

2

7 6 6 y ¼ 4 y½1 5 ¼ 4 h½1 0 y½2 2

x½0

6 ¼ Hx ¼ 4 x½1 0

Sequence:

h½0

0

3

7 xð0Þ h½0 5 xð1Þ h½1

3 7 h½0 x½0 5 ¼ Xh: h½1 x½1 0

(9:3:12)

The coefficient matrices X and H are ð3 2Þ size matrices. The column vectors y; h; and x are of the dimensions 3 1; 2 1; and 2 1, respectively. The entries in the coefficient matrices X and H have a special structure. For example, the first and the second columns of the coefficient matrix H are, respectively, given by

y½n ¼ 0; n50 y½0 ¼ h½0x½0 y½1 ¼ h½1x½0 þ h½0x½1 y½2 ¼ h½2x½0 þ h½1x½1 þ h½0x½2 ::: y½M 1 ¼ h½M 1x½0þ ::: þ h½0x½M 1:::

2

y½M þ L 2 ¼ h½M 1x½L 1 y½n ¼ 0; n M þ L 1: (9:3:8)

h½0

3

0

3

7 6 Colðh½0; h½1; 0Þ ¼ 4 h½1 5; 0 2

Example 9.3.2 Find the sequence y½n corresponding to the convolution of the following sequences and express y½n in matrix form:

7 6 Colð0; h½0; h½1Þ ¼ 4 h½0 5: h½1

h½0 ¼ 1; h½1 ¼ 1; h½n ¼ 0 for n50 and n 4 1;

The second column is obtained from the first by rotating the first entry to the second, second entry & to the third, and the third entry to the first.

(9:3:9)

(9:3:13)

9.3 DFT (FFT) Applications

375

Equations in (9.3.7) can be written in matrix and symbolic forms for y½n 6¼ 0 as follows: 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4

3

y½0

2

0

h½0

0

:

0

:

3

7 7 6 7 7 6 h½1 h½0 0 : : 0 7 7 6 7 7 6 h½2 h½1 h½0 : : 0 7 7 6 72 7 6 3 7 7 6 : x½0 : h½2 h½1 : : 7 7 6 76 x½1 7 7 6 : : : : : : 76 7 7 6 76 7 7 6 76 7 7 6 : : : : : : : 76 7; y ¼ Hx: 7¼6 6 7 6 7 7 : y½M 1 7 6 h½M 1 h½M 2 h½M 3 : : : 76 7 76 7 7 6 74 5 7 6 2 : : : 0 h½M 1 h½M y½M : 7 7 6 6 7 7 : 0 h½M 1 : : : : 7 x½L 1 7 6 7 7 6 7 7 6 : : : : : : : 7 7 6 6 7 : : : : : h½M 2 7 : 5 5 4 0 0 0 0 0 h½M 1 y½M þ L 2 y½1 y½2

Note that x is an L1 column matrix with the entries x½0; x½1; . . . ; x½L 1. The vector y is a ðM þ L 1Þ 1 column matrix and H is a ðM þ L 1Þ L matrix. The entries in the matrix H have the following pattern. The j th column in H is given by 9 8 ðj 1Þ zeros > > > > > = < N coefficients written in the order > ; 1 j L: > > h½0; h½1; . . . ; h½M 1 > > > > ; : ðL jÞ zeros (9:3:14b) If one column or one row of H is known, the entire matrix can be constructed. Noting that the convolution is commutative, i.e., y½n ¼ h½n x½n ¼ x½n h½n, equations similar to (9.3.14a and b) can be written by replacing h½n by x½n and vice versa. Computation of convolution via DFT: In Section 8.10, it was shown that a periodic convolution can be implemented by using the DFT. This idea can be used here as well and is illustrated by a simple example. Equation in (9.3.12) can be written as 2

y½0

3

2

h½0

7 6 6 y ¼ 4 y½1 5 ¼ 4 h½1 0 y½2

0

h½1

32

x½0

3

7 76 h½0 0 54 x½1 5 (9:3:15) 0 h½1 h½0

(9:3:14a)

This set of equations gives the same results as the set in (9.3.12). Interestingly, (9.3.15) is the same as the one in (8.10.28), with h½2 ¼ 0 corresponding to a periodic convolution. Equation (9.3.15) can be modified by adding the fourth column in the coefficient matrix and appending the data by two zeros resulting in (9.3.16). 2

3 2 32 3 y½0 h½0 0 0 h½1 x½0 6 y½1 7 6 h½1 h½0 0 6 7 0 7 6 7 6 76 x½1 7 ya ¼ 6 7¼6 76 7 ¼ Ha xa : 4 y½2 5 4 0 h½1 h½0 0 54 0 5 0

0

0

h½1 h½0

0 (9:3:16)

The reason for using N ¼ 4 for the extended sequences is that FFT can be used to find the convolution. In summary, given h½n; n ¼ 0; 1; . . . ; M 1 and x½n; n ¼ 0; 1; 2; . . . ; L 1, we can convolve h½n with x½n using the appended sequences as follows. Pad the sequences h½n and x½n with zeros so that they are of length N L þ M 1 resulting in the appended sequences ha ½n and xa ½n of length N, a power of 2. Convolution of the two extended sequences results in ya ½n ¼ ha ½n xa ½n. ya ½n is an appended sequence of y½n of length 2 N 1. To

376

9 Discrete Data Systems

determine the convolution via DFT (or FFT), do the following steps: 1. Determine Xa ½k ¼ DFTfxa ½ng and Ha ½k ¼ DFT fha ½ng . 2. Multiply the DFTs to form the products Ya ½k ¼ Ha ½kXa ½k. 3. Find the inverse DFT of Ya ½k. Discard the last ðN ðM þ L 1ÞÞ data points out of ya ½n to obtain y½n; n ¼ 0; 1; :::; N þ L 2: There are two methods for computing the discrete convolution, one by the convolution formula and the other by using DFTs. It may appear that the effort in computing the convolution via DFT is more computationally intensive than the direct convolution. However, using FFT, the number of computations is fewer, roughly for N432. Even though the number of computations is fewer for large N, there are difficulties with the use of DFT. The data sequence x½n may be long, say M points of data. The sequence h½n is of reasonable size, say L M. The computation of the DFTof x½n may not be possible due to computer storage constraints involving large amount of computations resulting in a significant delay. There are two methods, overlap-add and overlap-save, which can be used for large set of data. Both section long input sequence into shorter sections. They are suitable for online implementation if the process can tolerate slight delays. The overlap-add method is discussed below; see Ambardar (2007) for the overlap-save method. Overlap-add method: The sequence h½n of length Nis assumed to start at n ¼ 0. The sequence x½n is a much longer sequence of length M, also starting n ¼ 0. Partition x½n into k segments each of length N (zero padding the last segment if needed). The data can be expressed in a mathematical form using a rectangular window wR ½n of length N by

x½n ¼

k1 X

1;

n ¼ 0; 1; . . . ; N 1

0; otherwise

:

fx½ng ¼ ffx0 ½ng; fx2 ½ng; . . . ; fxk1 ½ngg; xi ½n x½n Ni; i ¼ 0; 1; . . . ; k 1 ¼ : 0; elsewhere (9:3:17b) We can now write y½n ¼ h½n x½n ¼ h½n

k1 X

xi ½n

i¼0

¼

k1 X i¼0

h½n xi ½n ¼

k1 X

yi ½n:

i¼0

It follows that the total convolution is the sum of the individual convolutions resulting in y½n ¼ y0 ½n þ y1 ½n N þ ::: þ yk1 ½n ðk 1ÞN:

(9:3:17a)

(9:3:18)

The ith segment of the output begins at n ¼ iN, as do the input segment xi ½n. However, each yi ½n segment has a length equal to ð2 N 1Þ and therefore yi ½n s ‘‘overlap’’ each other. We can think of each yi ½n as having the same length of 2 N 1 points, where each yi ½n includes zero padding before and/ or after as appropriate, such that the positions of the sequences are in correct location. Example 9.3.3 Consider the longer and shorter sequences given by x½n ¼ f1; 2; 3; 4g and h½n ¼ f1; 1g. a. Use the overlap-add method to determine the convolution of the sequences. b. Verify the results using direct convolution. Solution: a. Here M ¼ 2 and L ¼ 4. Section the sequence x½n into two sequences x0 ½n ¼ f1; 2g and x1 ½n ¼ f3; 4g. Then determine y0 ½n ¼ x0 ½n h½n and y1 ½n ¼ x1 ½n h½n. By the direct convolution, the sequences are as follows: 2

xi ½n; xi ½n ¼ x½nw½n iN;

i¼0

w½n ¼

That is,

3 2 3 2 3 2 3 1 0 1 3 3 0 6 7 1 6 7 6 7 1 6 7 ¼ 4 1 5; 4 4 3 5 ¼4 1 5 42 15 1 1 0 2 2 0 4 4 ) y0 ½n ¼ f1; 1; 2g; y1 ½n ¼ f3; 1; 4g:

9.3 DFT (FFT) Applications

377

Since the length of the sequence y½n is ð4 þ 2 1Þ ¼ 5, the sequence y0 ½n needs to be padded by two zeros at the end. Also, pad two zeros before the sequence y1 ½n. The result is obtained by overlapping the data and adding them at appropriate locations given below: y½n ¼ y0 ½n þ y1 ½n N

rhx ½k ¼ h½k x½k ¼

1 X

h½nx½n þ k

n¼1

¼ h½k x½k:

(9:3:19b)

The cross-correlation of two causal sequences x½k and h½k with M and L sample points, respectively, are

¼ f1; 1; 2; 0; 0g þ f0; 0; 3; 1 4g ¼ f1; 1; 1; 1; 4g:

rxh ½k ¼ x½k h½k ¼

M1 X

x½nh½n þ k;

n¼0

b. The result can be verified by directly using the direct convolution given below and the result is the same in both cases. See the equations in (9.3.19a) and (9.3.19b):

rhx ½k ¼ h½k x½k ¼

L1 X

h½nx½n þ k;

(9:3:19c)

n¼0

rhx ½k ¼ rxh ½k: 2

y½ 0

3

2

x½0

6 7 6 6 7 6 6 y½1 7 6 x½1 6 7 6 6 7 6 6 y½0 7 ¼ 6 x½2 6 7 6 6 7 6 6 7 6 6 y½3 7 6 x½3 4 5 4 y½ 4 0 2

1 0

6 6 62 6 6 ¼6 63 6 6 64 4 0

0

3 The integer k represents the shift of the second sequence with respect to the first.

7 7 x½0 7" 7 h½0 # 7 x½ 1 7 7 7 h½0 7 x½2 7 5 x½3 3

2

Example 9.3.4 Consider the data sequences given earlier in Example 9.3.2. Give the correlations of these two sequences and write them in a matrix form.

1

Solution: Using (9.3.20c), we have rhx ½k ¼ 0; k 2 and rhx ½k ¼ 0; k 2:

3

7 7 6 7 7 6 1 7" 6 17 # 7 1 6 7 7 7 6 7 6 ¼6 1 7 27 7: 7 1 7 6 7 7 6 37 6 17 5 5 4 4 4

rhx ½1 ¼ h½0x½1 þ h½1x½0 ¼ h½1x½0; rhx ½0 ¼ h½0x½0 þ h½1x½1; rhx ½1 ¼ h½0x½1 þ h½1x½2 ¼ h½0x½1; rhx ½2 ¼ h½0x½2 þ h½1x½1 ¼ 0: &

9.3.3 Correlation of Discrete Signals Discrete cross-correlations of two sequences (see (8.3.20a and b)) were defined as follows: rxh ½k ¼ x½k h½k ¼

1 X

(9:3:19d)

x½nh½n þ k

2

rhx ½1

6 6 In matrix form ) 6 rhx ½0 4 rhx ½1

3 2

h½ 1

0

3

" # 7 6 7 x½0 7 6 7 7¼ 6 h½0 h½1 7 5 4 5 x½0 0 h½ 0

2

3 2 3 1 1 0 6 7 1 6 7 ¼ 4 1 1 5 ¼4 0 5 1 0 1 1

n¼1

¼ x½k h½k;

(9:3:19a)

&

378

9 Discrete Data Systems

The matrix equation can be generalized. The nonzero cross-correlations can be written in the following matrix and symbolic forms: 3 2 h½M 1 0 rhx ½ðM 1Þ 6 r ½ðM 2Þ 7 6 h½M 2 h½M 1 7 6 6 hx 7 6 6 7 6 6 : : : 7 6 6 7 6 6 : : : 7 6 6 7 6 6 7 6 6 : : : 7 6 6 7 6 6 h½1 rhx ½0 7 ¼ 6 h½0 6 7 6 6 7 6 6 0 h½0 rhx ½1 7 6 6 7 6 6 : 0 : 7 6 6 7 6 6 7 6 6 : : : 7 6 6 7 6 6 5 4 : : 4 : 2

rhx ½L 1

0

0

: :

: : : :

: :

: : : :

:

: :

: :

: : : :

: :

: : : :

:

: :

:

: :

3

0 0

7 7 7 72 : 3 7 x½0 h½M 1 7 76 76 x½1 7 7 h½M 2 7 7 76 7 6 : 76 7 ) y ¼ Hcorr x; : 76 7 76 : 7 76 : 7 74 5 : 7 : 7 7 x½L 1 7 : 7 7 5 :

(9:3:20)

h½0

where y is a column vector of dimension ðMþL1Þ, x is a column matrix of dimension L, and Hcorr is a rectangular matrix of dimensions ðM þ L 1Þ L. If xðkÞ ¼ hðkÞ then the cross-correlation coefficients are the autocorrelation coefficients.

Power spectral density: The autocorrelation (AC) sequence of x½n plays a major role in spectral estimation, as its power spectral density is jX½kj 2 ¼ Sx ½k. It is

Notes: In comparing (9.3.14a) and (9.3.20), for convolution, the first column of H has h½n in the normal order. For correlation, the first column of Hcorr has h½n in reverse order. In both cases the other columns can be determined from the first column.

Sx ½k ¼

Computation of the cross-correlation using DFT: Given h½n; n ¼ 0; 1; . . . ; M 1 and x½n; n ¼ 0; 1; . . . ; L 1 determine the cross-correlation function X rhx ½n ¼ h½kx½k þ n: (9:3:21)

9.3.4 Discrete Deconvolution

k

Considering the equations for the convolution (see (9.3.14a)) and the cross-correlation (see (9.3.20)), we see that both have the same general form and the same computational procedure can be used for both cases. The following step-by-step procedure can be used. 1. Zero-pad both sequences to length N Lþ M 1. To use FFT, use N a power of 2. 2. Find the DFTs of h½n and x½n. 3. Rhx ½k ¼ H ½kX½k; k ¼ 0; 1; 2; :::; N 1. 4. Find the inverse DFT of Rhx ½k.

N1 X

rx ½nejð2p=NÞnk ¼

þ

rx ½nejð2p=NÞnk

n¼0

n¼ðN1Þ N 1 X

N1 X

rx ½nejð2p=NÞnk rx ½0:

(9:3:22)

n¼0

We have seen in the analog domain when a signal goes through a linear time-invariant system, then the signal is modified by the impulse response of the system. The same is true in the digital domain. The convolution of two sequences that are of finite width was defined earlier and y½n ¼ x½n h½n ¼

n X

h½kx½n k:

(9:3:23)

k¼0

There are three functions x½n; h½n; and y½n. In finding the convolution, x½n and h½n are known and y½n is determined by (9.3.23). In the deconvolution problem, the output sequence y½n and the input data x½n are known and h½n is to be determined. There are four ways to achieve this

9.3 DFT (FFT) Applications

379

goal. These are as follows: 1. recursion, 2. polynomial division, 3. using DFT, and 4. L p deconvolution. Deconvolution by recursion: From (9.3.23) h½0 ¼ y½0=x½0. Now separate the term h½n in (9.3.23) and write in the following form and determine successively the values of h½n for n40: y½n ¼

n X

h½kx½n k ¼ h½nx½0 þ

k¼0

n1 X

h½kx½n k:

k¼0

(9:3:24) Example 9.3.5 In Example 9.3.3 the convolution sequence y½0 ¼ 1; y½1 ¼ 0; y½2 ¼ 1 was computed using the sequences x½0 ¼ 1; x½1 ¼ 1 and h½0 ¼ 1; h½1 ¼ 1. Verify the sequence h½n using the recursion method. Solution:

Minimization of this error in terms of the unknowns h½n is a difficult problem for an arbitrary p. The general solution can only be determined by iterative means, see the articles by Byrd and Payne (1979) and Yarlagadda et al. (1985). There is a simple solution when p ¼ 2, which is used if the noise sequence is from a Gaussian distribution. These problems can be described under the general problem of solving a set of equations that are overdetermined and underdetermined system of equations. In Section A.6 we consider the solutions of overdetermined and underdetermined system of equations. Consider the system of equations in the symbolic matrix form Ah ¼ y:

(9:3:28a)

h½0 ¼ y½0=x½0 ¼ 1;

The least-squares solution to the overdetermined system in (9.3.28a) is (see (A.8.16b))

h½1 ¼ ð1=x½0Þfy½1 h½0x½1g ¼ 0 1 ¼ 1: (9:3:25) &

y ¼ Ah ) ðAT AÞh ¼ AT y ) h ¼ ðAT AÞ1 AT y: (9:3:28b)

This method is not practical in the presence of noise. Deconvolution using polynomial division in terms of z-transforms will be considered in Section 9.8.1. The DFT method makes use of DFTsof the sequences with Y½k ¼ H½kX½k. Then, H½k ¼ Y½k=X½k and its inverse DFT gives h½n. This procedure is similar to the one in the analog domain. It has at least two disadvantages. One of them is X½ks may be zero resulting in division by zero. Also, it is sensitive to noise in the input. Fourth method is based on minimizing the Lp errors discussed in Section 3.3. Deconvolution by Lp methods: The output is assumed to be the convolution of two sequences, say an input sequence x½n, a linear discrete system response sequence given by h½n, and an additive noise sequence e½n. The output is y½n ¼ h½n x½n þ e½n; n ¼ 0; 1; 2; . ..; N 1: (9:3:26) The noise signal can only be described by statistical measures. An interesting error measure is the Lp ; 1 p 1 measure defined by jejp ¼

N 1 X n¼0

ðy½n ðh½n x½nÞÞp :

(9:3:27)

The matrix ½ðAT AÞ1 A] is a pseudo-inverse of the matrix A. The MATLAB routine to compute this inverse is pinvðAÞ ¼ ðAT AÞ1 AT :

(9:3:29)

The inverses of the matrix ðAT AÞ may not exist. In such cases, a diagonal matrix dI, where d is a small positive number, is added to the matrix ðAT AÞ. This is called diagonal loading. An approximate solution of (9.3.28) is then given by h ﬃ ðAT A þ dIÞ1 Ay:

(9:3:30)

Example 9.3.6 Solve the following set of equations using the least-squares solution: 2

3 2 3 1 0 e 6 7 1 6 7 Ah ¼ 4 2 1 5 þ 4 e 5 1 0 2 e 2 3 2 3 2 3 e e 1 6 7 6 7 6 7 ¼ 4 1 5 þ 4 e 5 ¼ y þ 4 e 5: 2

e

e

(9:3:31)

380

9 Discrete Data Systems

Solution: The pseudo-inverse of A and the solution vector are, respectively, given as follows: 1 5 2 1 2 0 ðAT AÞ1 AT ¼ 21 2 5 0 1 2 1 5 8 4 ; ¼ 21 2 1 10 82 3 2 39 1 e > > = < 5 8 4 1 6 7 6 7 h¼ 4 1 5 þ 4 e 5 > 21 2 1 10 > ; : 2 e 21 1 :3333e 1 1 7e ¼ þ ¼ : 21 21 21 7e 1 þ :3333e Clearly if e ¼ 0, the solution coincides with the & vector h we started with. Notes: Implementation of discrete algorithms generally requires multiplications, which are expensive compared to additions. The following table gives a rough comparison of how expensive the additions, multiplications, and data transfers are for fixed and floating point machines by assuming one unit of expense corresponding to an addition compared to other operations. This gives a comparison of the computational expense and not the individual machine comparison, see Stine (2003) and Swartzlander Jr. (2001).

Fixed point Floating point

Multiplication

Addition

Transfer

10 2

1 1

0.5 0.5

From this table one can appreciate how much FFT algorithms are cost-effective in implementing the discrete Fourier transform when the number of & data points N is large. In the following, z-transforms, the discrete-time counterpart of the L-transforms, will be presented. Theory behind z-transforms is rather sophisticated and our presentation will be simple. See Oppenheim and Schafer (1999) for a detailed discussion on this topic.

9.4 z-Transforms

1 X

jx½nj51:

(9:4:1)

n¼1

The DTFT pair is jO

Xðe Þ ¼

1 X

x½ne

1 ! 2p

jnO DTFT

n¼1

ðp

XðejO ÞejnO dO

p

¼ x½n:

(9:4:2)

The DTFT of x½nens and the corresponding DTFT are as follows: DTFT½x½nens ¼ ¼

1 X

½x½nens ejnO

n¼1 1 X

x½nejðsþjOÞn ;

(9:4:3)

n¼1

x½nejnO

DTFT

! XðejðsþjOÞ Þ:

(9:4:4)

The convergence of the sequence x½n ejnO can now be defined in terms of ens , which is similar to the convergence of L-transforms, see Section 5.4. It is desirable to use the notation z ¼ esþjO ¼ es ejO ¼ rejO and lnðzÞ ¼ s þ jO and ð1=zÞdz ¼ jdO:

(9:4:5)

Using these in (9.4.1), the time sequence and the corresponding z-transform are 1 x½n ¼ 2pj ¼

I

1 X

XðzÞzn1 dz;

XðzÞ

x½nzn :

(9:4:6)

n¼1

The z-transform of a discrete-time sequence x½n is defined in terms of a complex variable z by Zfx½ng ¼ XðzÞ ¼

1 X

x½nzn :

(9:4:7a)

n¼1

The DTFT of the sequence x½n, XðejO Þ exists provided that x½n is absolutely summable (see (8.5.11), which is repeated below in (9.4.1)). This is sufficient but not necessary:

The range of values of the complex variable z for which the summation converges is called the region of convergence (ROC). The inverse z-transform and

9.4 z-Transforms

381

Fig. 9.4.1 Contour of integration on the z-plane.

Fig. 9.4.2 Example 9.4.1: region of convergence ða40Þ

Im(z)

Im(z)

r

0

Re(z)

Re(z )

the symbolic relationship between x½n and XðzÞ are, respectively, given by I 1 1 n1 dz x½n ¼ Z ½XðzÞ ¼ C XðzÞz 2pj

I z contour integral ; x½n $ XðzÞ: C (9:4:7b) The contour integral is around a circle of radius r in the counterclockwise direction enclosing the origin on the z-plane, see Fig. 9.4.1. The complex domain integration requires knowledge of complex variables, which is beyond our scope. The z-transform exists when the sum in (9.4.7a) converges. A necessary condition for convergence is absolute summability of x½nzn . Let z ¼ rejO . The absolute summability of x½nzn is

ROC 1 is the range of values of z for which az 51 or jzj4jaj. The transform is represented by a rational function of the complex variable z. As in the Laplace transforms, we can describe a rational function XðzÞ in terms of its poles (the roots of the denominator) and zeros (the roots of the numerator) on the complex z-plane. There is a pole at z ¼ a and a zero at z ¼ 0 and these are shown in Fig. 9.4.2. The ROC is outside the circle of radius a. The boundary of the ROC is jzj ¼ jaj. The ROC does not contain any poles and is outside & the circle of radius jzj ¼ a. Example 9.4.2 Consider the left-side sequence x2 ½n ¼ bn u½n 1; b 6¼ 0. Solution: The z-transform is 1 1 X X x2 ½nzn ¼ bn u½n 1zn : X2 ðzÞ ¼ n¼1

1 X

jx½nrn j51:

n¼1

(9:4:10)

(9:4:8)

n¼1

The range of r for which the sum converges is the region of convergence (ROC).

u½n; u½n; and u½n 1 ¼ u½ðn þ 1Þ are sketched in Fig. 9.4.3a,b,c and X2 ðzÞ ¼

1 X

ðb=zÞn :

(9:4:11)

n¼1

9.4.1 Region of Convergence (ROC)

Using the change of variable m ¼ n in (9.4.12), we have

Example 9.4.1 Determine the z-transform of the right-sided sequence x1 ½n ¼ an u½n.

n¼1

¼

n¼0

ðaz1 Þn ¼

1 X m¼1

Solution: The z-transform of x1 ½n is 1 1 X X x1 ½nzn ¼ an u½nzn X1 ðzÞ ¼ 1 X

X2 ðzÞ ¼

ðz=bÞm ¼ 1

1 X

ðz=bÞn

m¼0

1 z ¼1 ¼ ; jzj5jbj: 1 ðz=bÞ z b

(9:4:12)

n¼1

1 z ¼ ; jzj4jaj: 1 1 az za (9:4:9)

See the pole–zero plot and the region of convergence in Fig. 9.4.4. In the last two examples, a right-side and a leftside sequences were considered. If a ¼ b, the two

382

9 Discrete Data Systems

Fig. 9.4.3 (a) u½n, (b) u½n, and (c) u½ðn 1Þ

(a)

(b)

transforms are identical except the ROC is different. The z-transform of the sequence and the ROC need to be known before the sequence can be identified. Example 9.4.3 Determine the z-transformof the two-sided sequence y½n ¼ an u½n þ bn u½n 1 and the ROC. Solution: The z-transform of the sum is obtained by adding the two transforms and 1 X ðan u½n bn u½n 1Þzn YðzÞ ¼ n¼1

¼

1 X

an zn 1 þ

n¼0

1 X

ðz=bÞn :

(9:4:13)

m¼0

First and the second sums on the right-hand side have the ROCs jzj4jaj and jzj5jbj, respectively. The sum of the two functions and the corresponding ROC are given by z z YðzÞ ¼ þ ; za zb ROC : fjzj4jajg \ fjzj5jbjg: (9:4:14)

x½n ¼

(c)

6¼ 0; N1 n N2 0; Otherwise

z

! XðzÞ ¼

N2 X

x½nzn :

n¼N1

(9:4:15) Solution: For z 6¼ 0 or 1, each term will be finite and the function XðzÞ converges. If N1 50 and N2 40, the sum includes both negative and positive powers of z. As jzj ! 0, the terms with negative powers of z become unbounded. As jzj ! 1, the terms with positive powers of z become unbounded. Therefore, the ROC of the function XðzÞ of a finite sequence is the entire z-plane except for z ¼ 0 and z ¼ 1. If N1 0; the ROC includes z ¼ 1 and if N2 0, the ROC & includes z ¼ 0. Example 9.4.5 Find XðzÞ for x½n in Fig. 9.4.6 and make a pole–zero plot. Solution: First, XðzÞ ¼

2 X

x½nzn ¼ 1:zð2Þ þ ð6Þ:zð1Þ

n¼2

The ROC in (9.4.15) exists only if there is an overlap of the regions identified by the regions fjzj4jajg and fjzj5jbjg. If jbj4jaj, then the transform converges in the annular region shown in Fig. 9.4.5a identified by jaj5jzj5jbj. If jbj5jaj, there is no region of overlap and therefore the ROC is the & null set.

þ 9:z0 þ 4:z1 þ ð12Þ:z2 ¼ z2 6z þ 9 þ 4z1 12z2 z4 6z3 þ 9z2 þ 4z 12 z2 2 ðz 2Þ ðz 3Þðz þ 1Þ ¼ : z2 ¼

Example 9.4.4 Give the ROC of the sequence.

Im(z)

Im(z)

Im(z) α

β

β Re(z)

β

β Re(z)

Fig. 9.4.4 Pole–zero plot and ROC of X2 ðzÞ.

α+β 2

α Re(z) α+β 2

Fig. 9.4.5 (a) ROC of YðzÞ and (b) no region of convergence when jaj4jbj

9.4 z-Transforms

383

Fig. 9.4.6 (a) x½n and (b) pole–zero plot

The poles of XðzÞ are at the origin and have zeros at z ¼ 1; 2; and 3. See the pole–zero plot in Fig. 9.4.6b. The ROC of this function spans the entire z-plane corresponding to the region enclosed between & the poles at zero and those located at infinity. Example 9.4.6 Derive the z-transform of the sequence for the following two cases: a being arbitrary and a ¼ 1: n a ; 0nN1 : (9:4:16) x½n ¼ 0; otherwise Solution: XðzÞ ¼

N1 X

an zn ¼

n¼0

N1 X

ðaz1 Þ

n¼0

1 aN zN ; ROC : jzj40: ¼ 1 az1 a ¼ 1 ) x½n ¼

1; 0 n N 1

0; otherwise

(9:4:17)

z

!

1 zN 1 z1

¼ XðzÞ; ROC : jzj 6¼ 0:

(9:4:18)

See Fig. 9.4.7 for the pole-zero plot assuming & N¼11 Notes on the ROC of a rational function XðzÞ: The ROC depends on the poles of the Im(z)

function XðzÞ at z ¼ ri ejyi ; ri 40. The maximum and minimum magnitudes of these poles are identified by rmax ¼ maxjri j and rmin ¼ minjri j. If the degree of the denominator of XðzÞ is smaller than the degree of the numerator, then XðzÞ has at least one pole at 1. 1. The ROC does not contain any poles. 2. If the sequence is a finite sequence, then the ROC of XðzÞ is the entire z-plane except possibly z ¼ 0 or z ¼ 1. 3. If x½n is a right-side sequence, i.e., x½n ¼ 0; n5N1 51, and XðzÞ converges for some values of z, the ROC is rmax 5jzj 1 with a possible exception of z ¼ 1. 4. If x½n is a left-sided sequence, i.e., x½n ¼ 0; 15N2 5n and XðzÞ converges for some values of z, then the ROC is 0 jzj5rmin with a possible exception of z ¼ 0 5. If x½n is a two-sided sequence and the region of convergence of the right- and left-sided sequences are, respectively, given by r1 5jzj and jzj5r2 and XðzÞ converges for some values of z, then the ROC takes the form r1 5jzj5r2 , where r1 and r2 are the magnitudes of the poles of XðzÞ. Example 9.4.7 Find the z-transforms of the following sequences and their ROCs: a: x1 ½n ¼ d½n; b: x2 ½n ¼ u½n;

11th order pole

c: x3 ½n ¼ u½n 1;

6 a Re(z)

d: x6 ½n ¼ ajnj ; jaj51: Solution: a: X1 ðzÞ ¼

1 X

d½nzn ¼ z0 ¼ 1;

n¼1 z

Fig. 9.4.7 Example 9.4.6: pole–zero plot ðN ¼ 11Þ

d½n ! 1; ROC : all z

(9:4:19)

384

b: X2 ðzÞ ¼

9 Discrete Data Systems 1 X

zn ¼

n¼0

1 z ; ROC : jzj41 ¼ 1 z1 z 1 (9:4:20)

c: X3 ðzÞ ¼ 1 jnj

1 z ¼ ; ROC : jzj51 1 1z z1 (9:4:21) n

Example 9.4.8 Use the z-transforms to determine the DTFT of the discrete-time function z

x½n ¼ u½n u½n N; N40 $ XðzÞ ¼

n

d: x6 ½n ¼ a ¼ a u½n þ a u½n 1 z z z ! z a z ð1=aÞ ða2 1Þ z ; ¼ a ðz aÞðz ð1=aÞÞ 1 ROC : jaj5jzj5 ; jaj51: a

response. The amplitude and the phase responses are periodic with period 2p.

N 1 X

zn ; ROC : jzj40:

(9:4:24)

n¼0

Solution: Since ROC of XðzÞ includes the unit circle, the DTFT XðejO Þ exists. It is periodic with period 2p and phase response is linear: (9:4:22) & XðejO Þ ¼ XðzÞjz¼ejO ¼

N 1 X n¼0

9.4.2 z-Transform and the Discrete-Time Fourier Transform (DTFT) Note XðzÞ is a function of the complex variable z. The point z ¼ rejO is located at a distance r from the origin at an angle O from the positive real axis. If x½n is absolutely summable, then the DTFT can be obtained from the z-transform by setting z ¼ rejO jr¼1 :

X ejO ¼ XðzÞjz¼e jO (9:4:23) The equation jzj ¼ ejO ¼ 1 describes a circle of unit radius ðr ¼ 1Þ centered at the origin in the z-plane. The frequency O in the discrete-time Fourier transform corresponds to the point on the unit circle at an angle O in radians with respect to the positive real axis. As the discrete-time frequency varies in the range p to p, in the z-plane, it corresponds to one time around the unit circle. In words, (9.4.23) states that the DTFT of a discrete-time signal x½n can be obtained from the z-transform XðzÞ by evaluating it on the unit circle. The z-transform of the sequence is assumed to exist and the DTFT of the sequence exists provided that the region of convergence of XðzÞ includes the unit circle. The DTFT function jO XðejO Þ ejyðOÞ , where is represented by Xðe Þ ¼ XðejO Þ is called the amplitude (or magnitude) response and yðOÞ is called the phase (or angle)

zn ¼

1 zN j jO 1 z1 z¼e

1 ejNO sinðNO=2Þ : ¼ ejOðN1Þ=2 ¼ 1 ejO sinðO=2Þ (9:4:25)

The amplitude and the phase responses are, respectively, given by jO sinðNO=2Þ

jO X e ¼ sinðO=2Þ ;ﬀX e ¼OðN1Þ=2: (9:4:26) &

9.5 Properties of the z-Transform Let the ROC of xi ½n is R xi and R 0 is the ROC after the appropriate operation. The ROC is stated in terms of set theory. The proofs are simple for many of these and are omitted.

9.5.1 Linearity z

Let xi ½n ! Xi ðzÞ, then z

x½n ¼ a1 x1 ½n þ a2 x2 ½n $ a1 X1 ðzÞ þ a2 X2 ðzÞ ¼ XðzÞ; ROC :R0 Rx1 \ Rx2 ; (9:5:1)

9.5 Properties of the z-Transform

385

where R 0 is the ROC of x½n, which is the proper subset of the ROCs of x1 and x2 . Note the intersection of the subsets represented by Rx1 \ Rx2 in (9.5.1). Expansion of the ROC takes into consideration pole cancellations with zeros in XðzÞ.

1 X

Zfx½ng ¼

x½nzn ¼

n¼1

XðzÞ ¼

m¼1

1 X

x½nzn ¼

n¼1

Zfx½nn0 g¼

1 X

x½nn0 zn

1 X

x½mzðmþn0 Þ ¼zn0

m¼1

1 X

¼

0 X

ðz1 Þn ¼

n¼1

x½mzm ¼zn0 XðzÞ;

m¼1

z

x½nn0 ! zn0 XðzÞ;ROC:R0 R\f05jzj51g: (9:5:2) Noting the multiplication factor zn0 , additional poles are introduced when n0 40 and, at the same time, some of the poles at 1 are deleted. In a similar manner if n0 50 then additional zeros are introduced at z ¼ 0 and some of the poles at 1 are deleted. This implies that the points z ¼ 0 and z ¼ 1 are either added or deleted from the ROC by time shifting the function. The special cases include the unit delay and advance operations:

1 ; ROC : R0 ¼ jzj51: 1z

Solution: Noting that u½n !ð1=ð1 z1 ÞÞ and using the transformation z ! 1=z results in the above equation verifying the time reversal theorem. We note that the closed-form expression for the sum is valid if jzj51. Also, the ROC of the unit step sequence is jzj41. Reversing the sequence results & in the ROC from jzj51.

9.5.4 Multiplication by an Exponential If a is a complex number, then Zfan x½ng ¼

1 X

an x½nzn ¼

n¼1

1 X

x½nðz=aÞn

n¼1

¼ Xðz=aÞ; an x½n ! Xðz=aÞ; ROC : R0 ¼ jajR:

1

x½n 1 ! z XðzÞ; ROC : R0 ¼ Rx \ f05jzjg;

(9:5:3a)

z

x½n þ 1 ! zXðzÞ; ROC : R0 ¼ R \ fjzj51g:

(9:5:3b)

(9:5:5)

Change in the argument of Xðz=aÞ resulted in the multiplication of ROC boundaries by jaj. ROC expands or contracts by the factor of jaj. In the special case of a ¼ ejO0 n : z

ejO0 n x½n ! XðejO0 zÞ; ROC :R0 ¼ R:

(9:5:6)

Example 9.5.2 Determine the z-transform of the real sequence y½n ¼ rn cosðO0 nÞu½n.

9.5.3 Time Reversal z

zn

n¼0

z

z

1 X

z

n¼1

¼

x½mzm ¼ Xð1=zÞ:

Example 9.5.1 Find z½x½n ¼ z½u½n directly and then verify using the result in (9.5.4):

9.5.2 Time-Shifted Sequences The z-transform of the time-shifted sequence and the corresponding ROC are

1 X

z

If x½n ! XðzÞ; ROC ¼ R, then x½n ! Xð1=zÞ; ð9:5:4Þ ROC :R0 ¼ 1=R: Reversing in time results in the transformation z ) ð1=zÞ in the transform and, the points in the ROC R0 corresponds to the inverses of the points in R. This can be shown by

z

Solution: Noting rn u½n ! z=ðz rÞ; ROC : jzj4jrj we can write 1 1 z z 1 y½n ¼ rn ejO0 n x½n þ rn ejO0 n u½n ! 2 2 2 z rejO0 1 z þ ¼ YðzÞ ; 2 z rejO0

386

9 Discrete Data Systems

9.5.6 Difference and Accumulation

z

y½n ¼ rn cosðO0 nÞu½n ! YðzÞ ¼

z2 r cosðO0 Þz ; ROC : jzj4jrj: z2 2r cosðO0 Þz þ r2

The z-transforms of these are

(9:5:7a) In a similar manner, r sinðO0 Þz ; r sinðO0 nÞu½n ! 2 z 2r cosðO0 Þz þ r2 (9:5:7b) ROC : jzj4r:

z

y1 ½n ¼ x½n x½n 1 ! XðzÞ½1 z1 ¼ Y1 ðzÞ; R0 R \ fjzj40g; (9:5:10a) n X 1 z XðzÞ y2 ½n ¼ x½k ! ð1 z1 Þ k¼1 ¼ Y2 ðzÞ; R0 R \ fjzj41g :

z

n

&

9.5.5 Multiplication by n

(9:5:10b)

9.5.7 Convolution Theorem and the z-Transform Convolution theorem states that the convolution in the time domain corresponds to the multiplication in the z-domain and z

y½n ¼ x½n h½n ! XðzÞHðzÞ The z-transform of nx½n is

¼ YðzÞ; ROC : Ry ðRx \ Rh Þ:

z

y½n ¼ nx½n ! z½dXðzÞ=dz ¼ YðzÞ; ROC : R0 ¼ R:

(9:5:8)

This can be shown using the expression for the convolution of two sequences x½n and h½n and then by using the transform pairs as shown below: z

This can be seen by differentiating both sides with respect to zof the following equation:

z

x½n ! XðzÞ and h½n k ! zk HðzÞ; y½n ¼

1 X

x½kh½n k ¼

k¼1

XðzÞ ¼

1 X

x½nzn !

n¼1

(9:5:11)

1 X

h½kx½n k;

k¼1

1 X " # dXðzÞ 1 1 1 ðnÞx½nzn1 : ¼ X X X n dz y½nz ¼ x½kh½n k zn YðzÞ ¼ n¼1 n¼1

Multiplying both sides by (z) and identifying the appropriate time and transform terms, (9.5.8) follows. The region of convergence is the same for XðzÞ and YðzÞ. Example 9.5.3 Determine the z-transform of the right-side sequence x½n ¼ nan u½n; a40 using the multiplication by n property.

¼

1 X

n¼1

" x½k

1 X

#

h½n kzn

n¼1

k¼1

¼

1 X

k¼1

x½k HðzÞzk

k¼1

¼ HðzÞ

1 X

x½kzk ¼ HðzÞXðzÞ

k¼1

z

Solution: Noting that an u½n !ðz=ðz aÞÞ; ROC : jzj4jaj, by using (9.5.8), we have

z

nan u½n ! z

dðz=ðz aÞÞ az ; ¼ dz ðz aÞ2

(9:5:9)

ROC : jzj4jaj: &

The ROC of YðzÞ contains the intersection of the ROC of XðzÞ and YðzÞ: If a zero of one of the transforms cancels with a pole of the other, then the ROCof YðzÞ will be larger than the intersection of R x and R h . Example 9.5.4 Verify the result in (9.5.10b) by using (9.5.11).

9.5 Properties of the z-Transform

387

Solution: Noting u½n k ¼ 0; k4n, y½n can be expressed by n 1 X X x½k ¼ x½ku½n k y½n ¼ k¼1

We have a pole outside and one inside the unit circle resulting in a transform that has the annular region of convergence given in (9.5.14d). See Fig. 9.5.1 for & pole-zero plots and ROC.

k¼1 z

¼ x½n u½n ! XðzÞzfu½ng z ¼ XðzÞ ; ROC :R0 ðR \ fjzj41gÞ: ðz 1Þ (9:5:12) & Example 9.5.5 Find the z-transform of the following sequence: y½n ¼ x½n h½n; x½n ¼ nan u½n; h½n ¼ ðbÞn u½n; a ¼ b ¼ 1=2:

(9:5:13)

Solution: Using the multiplication by n property, we have az z x½n ¼ nan u½n ! ¼ XðzÞ; ðz aÞ2 ROC : jzj4jaj; (9:5:14a) z h½n ¼ bn u½n ! ; ROC : jzj4jbj: (9:5:14b) ðz bÞ z

9.5.8 Correlation Theorem and the z-Transform In Section 8.3 the cross-correlation of two sequences x½n and h½n was defined by rxh ½k ¼ x½k h½k ¼ ¼

1 X

x½nh½n þ k

n¼1 1 X

x½m kh½m: (9:5:15)

m¼1

Correlation theorem: rxh ½k¼

1 X

z

x½nh½nþk ! XðzÞHð1=zÞ

n¼1

¼Rxh ðzÞ;ROC:Rxh ðRx \Rh Þ: Using (9.5.14a and b), the time reversal property, and the convolution theorem, we have z

x½n ¼ nð1=2Þn u½n !

ð1=2Þz ðz þ ð1=2ÞÞ2

¼ XðzÞ; ROC :jzj4ð1=2Þ ; z

h½n ¼ ð1=2Þn u½n !

1=z ð1=zÞ ð1=2Þ

(9:5:14c)

2 ¼ HðzÞ; ROC :jzj52 ; z2 z YðzÞ ¼ HðzÞXðzÞ ¼ ; ðz 2Þðz þ ð1=2ÞÞ2 1 5 jzj 5 2: (9:5:14d) ROC : 2

(9:5:16)

This can be seen by first noting that rxh ½k ¼ x½k h½k. Using the convolution and the time reversal properties, we can see the result in (9.5.16). Again there is a possibility of pole–zero cancellations and therefore Rxh ðRx \ Rh Þ. In the case of autocorrelation, we have h½n ¼ x½n and the autocorrelation (AC) theorem in (9.5.16) reduces to

¼

Im(z)

−

Fig. 9.5.1 Example 9.5.5: pole–zero plots and the ROCS: (a) XðzÞ; (b) HðzÞ; and (c) YðzÞ

1 2

rxx ½k ¼ rx ½k ¼

1 X n¼1

¼ Rx ðzÞ; ROC : Rxh ðRx \ Rh Þ:

1 2 Re(z)

(9:5:17)

Im(z)

Im(z)

1

z

x½nx½n þ k ! XðzÞXð1=zÞ

Re(z)

−

1 2

1 2 Re(z)

388

9 Discrete Data Systems

Example 9.5.6 Using a. the direct and b. the transform methods, find the AC of x½n ¼ an u½n; jaj51:

(9:5:18)

Solution: a. Since the AC function is even, we need to find rx ½k; k 0 and then use rx ½k ¼ rx ½k. For k 0, since jaj51, we have

9.5.10 Final Value Theorem in the Discrete Domain The final value theorem applies only to causal sequences x½n and if all the poles of XðzÞ lie within the unit circle, with the exception that it can have one pole at z ¼ 1, then lim x½n ¼ limð1 z1 ÞXðzÞ if x½1exists: (9:5:21)

n!1

rx ½k ¼

1 X

1 X

an anþk ¼ ak

n¼0

ða2 Þn ¼

n¼0

1 ak ; k 0: 1 a2

Autocorrelation is even ) rx ½k jkj

2

z!1

This can be seen by Zfx½n x½n 1g ¼ ð1 z1 ÞXðzÞ

(9:5:19a)

¼ a =ð1 a Þ; 15k51:

N!1

b. By the autocorrelation theorem,

lim lim

N X

z!1 N!1

Rx ðzÞ ¼ XðzÞXðz1 Þ ¼ ¼

1 1 1 az1 1 az

1 z ; a ðz aÞðz ð1=aÞÞ

N X

¼ lim

fx½n x½n 1gzn ;

(9:5:22)

n¼0

fx½n x½n 1gzn

n¼0

¼ lim lim

N!1 z!1

N X

fx½n x½n 1gzn

n¼0

(9:5:19b)

lim ½x½0 x½1 þ x½1 x½0 þ x½2 x½1 þ :::

N!1

1 ROC : jaj5jzj5 : jaj

¼ lim x½N: N!1

z

) ajkj !

2

a 1 z : a ðz aÞðz ð1=aÞÞ

(9:5:19c)

The region of convergence is an annular ring and the AC sequence is a two-sided sequence. Note the & poles of the z-transform in (9.5.19c).

Notes: The final value x½1 is equal to zero if all the poles of XðzÞ lie within the unit circle. This follows from the fact that the corresponding time function contains exponentially damped terms. It is a constant if XðzÞ has a single pole at z ¼ 1. If the poles are outside of the unit circle, then the final value theorem gives incorrect results. Also, x½1 is indeterminate if & there are complex poles on the unit circle.

9.5.9 Initial Value Theorem in the Discrete Example 9.5.7 Find the initial and final values of Domain x½n for the function If the sequence x½n is causal, i.e., x½n ¼ 0; n50; as z ! 1; zn ! 0 for n40, we have 1 X

x½0 ¼ lim XðzÞ ¼ lim z!1

z!1

n¼0

1

¼ lim ½x½0 þ x½1z z!1

x½nz

n

þ x½2z

(9:5:20a) 2

þ ;

XðzÞ ¼

Solution: The initial and final values are, respectively, given by ðzð1=3ÞÞ ¼0 x½0¼ lim XðzÞ¼ lim z!1 z!1 ðz1Þðzð1=2ÞÞ lim x½n¼limð1z1 ÞXðzÞ

n!1

x½0 ¼ 0 ) x½1 ¼ lim zXðzÞ: z!1

ðz ð1=3ÞÞ ; jzj41: ðz 1Þðz ð1=2ÞÞ

(9:5:20b)

z!1

ðz1Þðzð1=3ÞÞ 4 ¼ : ¼lim z!1 zðz1Þðzð1=2ÞÞ 3

&

9.6 Tables of z-Transform Properties and Pairs

389

y½n ¼ x½nu½n ¼ f0; 1; 0; 1; 0; 1; 0; 1; :::g: #

Switched periodic sequences and their z-transforms: Consider the periodic sequence x½n with the property x½n ¼ x½n þ N. Now define a causal sequence y½n ¼ x½nu½n. Let the z-transform of the first N-point sequence is N1 X X1 ðzÞ ¼ x½nzn :

Solution: The period of the sequence sinðnp=2Þ is 4. Therefore z

n¼0

x1 ½n ¼ f0; 1; 0; 1g ! z1 z3 ¼ X1 ðzÞ #

The z-transform of the function y½n is given by YðzÞ ¼ X1 ðzÞ þ z

N

X1 ðzÞ þ z

2 N

) XðzÞ ¼

X1 ðzÞ þ :::

X1 ðzÞ z1 z3 z : ¼ ¼ 2 1 z4 1 z4 z þ1

¼ X1 ðzÞ½1 þ zN þ z2 N þ ::: &

zN X1 ðzÞ; jzj41: ¼ N z 1 Example 9.5.8 Find the z-transform of the sequence y½n ¼ x½nu½n; x½n ¼ sinðnp=2Þ:

9.6 Tables of z-Transform Properties and Pairs

Table 9.6.1 Z-transform properties z

z

Two sided signals x½n ! XðzÞ; h½n ! HðzÞ Superposition: z ax½n þ bh½n ! aXðzÞ þ bHðzÞ

ð9:6:1Þ

Time shift: z x½n n0 ! zn0 XðzÞ

ð9:6:2Þ

Scaling: z an x½n ! Xðz=aÞ

ð9:6:3Þ

Multiplication byejnO0 : z

ejnO0 x½n ! XðejO0 zÞ

ð9:6:4Þ

Time reversal: z x½n ! Xð1=zÞ

ð9:6:5Þ

Multiplication by n: z

nx½n ! z dXðzÞ dz

ð9:6:6Þ

Accumulation: n P z z x½k ! z1 XðzÞ

ð9:6:7Þ

k¼1

Difference: z x½n x½n 1 !ð1 z1 ÞXðzÞ Convolution: z x½n h½n ! XðzÞHðzÞ Cross correlation: z x½k h½k ! XðzÞHð1=zÞ Autocorrelation: z x½n x½n ! XðzÞXð1=zÞ

ð9:6:8Þ ð9:6:9Þ ð9:6:10aÞ ð9:6:10bÞ

The following properties hold for causal sequences x½n ¼ 0; n50: Initial value theorem: x½0 ¼ lim XðzÞ

ð9:6:11aÞ

Final value theorem: lim x½n ¼ limðz 1ÞXðzÞ

ð9:6:11bÞ

z!1

n!1

z!1

Switched periodic functions: x½n ¼ x½n þ N; X1 ðzÞ ¼

N1 P n¼0

z

N

x½nzn ; x½nu½n ! zNz1 X1 ðzÞ

ð9:6:12Þ

390

9 Discrete Data Systems Table 9.6.2 Z-transform pairs Unit sample: z

d½n ! 1; ROC : all z: z

ð9:6:13aÞ

k

ð9:6:13bÞ

d½n k ! z ; k > 0; ROC : jzj > 0: z

d½n þ k ! zk ; k > 0; ROC : jzj51:

ð9:6:13cÞ

Unit step: z ; ROC : jzj > 1: z1 z z u½n 1 ! ; ROC : jzj51: z1 z

u½n !

ð9:6:14aÞ ð9:6:14bÞ

Exponential: z ; ROC : jzj > jaj: za z z bn u½n 1 ! ; ROC : jzj5jbj: zb General type: az z nan u½n ! ; ROC : jzj > jaj: ðz aÞ2 z

an u½n !

ð9:6:15aÞ ð9:6:15bÞ

ð9:6:16aÞ

nðn 1Þ:::ðn ðk 2ÞÞankþ1 u½n z z ! ; ROC :jzj > jaj: ðk 1Þ! ðz aÞk az z nan u½n 1 ! ; ROC : jzj5jaj: ðz aÞ2 z2

z

ðn þ 1Þan u½n !

ðz aÞ2

ð9:6:16bÞ ð9:6:16cÞ ð9:6:16dÞ

; ROC : jzj > a:

Sequences involving sinusoids: z2 cosðO0 Þz ; ROC : jzj > 1: z2 ð2 cosðO0 ÞÞz þ 1

ð9:6:17aÞ

sinðO0 Þz ; ROC : jzj > 1: z2 ð2 cosðO0 ÞÞz þ 1

ð9:6:17bÞ

z

cosðO0 nÞu½n ! z

sinðO0 nÞu½n !

z2 r cosðO0 Þz ; ROC : jzj > r: ð2r cosðO0 ÞÞz þ r2

ð9:6:17cÞ

r sinðO0 Þz ; ROC : jzj > r: z2 ð2r cosðO0 ÞÞz þ r2

ð9:6:17dÞ

z

rn cosðO0 nÞu½n ! z

rn sinðO0 nÞu½n !

z2

Finite sequence: n N N z 1a z a ;0 n N1 ; ROC : jzj > 0: ! 0; otherwise 1 az1

ð9:6:18Þ

9.7 Inverse z-Transforms

9.7.1 Inversion Formula

In this section we will consider determining x½n ¼ Z1 fXðzÞg by the following methods: (1) inversion formula, (2) use of z-transform tables, and (3) power series expansion.

The inverse z-transform is x½n ¼

1 2pj

I c

XðzÞzn1 dz:

(9:7:1)

9.7 Inverse z-Transforms

391

It is a contour integral over a closed path C encircling the origin in a counterclockwise direction that lies within the region of convergence of XðzÞ in the zplane. The proof requires the knowledge of complex variables, which is beyond the scope here. See Churchill (1948), Poularikas (1996), and others. Let fak g be the set of poles of XðzÞzn1 inside the contour Cand fbk g be the set of poles of XðzÞzn1 outside the contour C in a finite region of the z-plane. Now x½n ¼ Z1 ½XðzÞ 8P n1 > > < k ResfXðzÞz ; ak Þ; n 0; ¼ P > > : ResfXðzÞzn1 ; bk g; n50:

Re( z )

ð9:7:2aÞ ð9:7:2bÞ

The residue at the multiple and the simple poles at z ¼ p0 of order k are, respectively, given by 1 d k1 ½ðz p0 Þk XðzÞzn1 z!p0 ðk 1Þ! dzk1

ResfXðzÞzn1 g ¼ lim

(9:7:2c)

ResfXðzÞzn1 ; p0 g ¼ XðzÞzn1 ðz p0 Þ z¼p0 ðSimple poleÞ:

1 2

1 2

Fig. 9.7.1 Example 9.7.1: Poles, zeros, and the ROC

k

ðMultiple poleÞ;

Im( z )

(9:7:2d)

9.7.2 Use of Transform Tables (Partial Fraction Expansion Method) This method is based upon expressing XðzÞ as a sum of simple functions Xi ðzÞ by using partial fraction expansion (see Section 5.9.2.), where each one of these functions have inverse transforms that are readily available in a table. This method is limited to rational functions and will need a table of z-transform pairs that provides the appropriate z transform pairs xi ½n ! Xi ðzÞ. The inverse transform is given by ( ) M X 1 1 Xi ðzÞ x½n ¼ Z fXðzÞg ¼ Z i¼1

Example 9.7.1 Find the inverse z-transform of the following function using the residues z ; ðz :5Þðz 2Þ ROC : :55jzj52: See Fig 9:7:1

¼

M X i¼1

Z1 fXi ðzÞg ¼

M X

xi ½n:

(9:7:4)

i¼1

XðzÞ ¼

(9:7:3a)

Solution: The function has a single pole inside and a single pole outside the unit circle. Therefore, it is a two-sided sequence. From (9.7.2c), we have x½n ¼ ResfXðzÞzn1 ; :5g ¼

zðz :5Þzn1 ð:5Þn ; n 0; jz¼:5 ¼ ðz :5Þðz 2Þ 1:5 (9:7:3b)

x½n ¼ ResfXðzÞzn1 ; 2g zðz 2Þzn1 2n ; n50: ¼ jz¼2 ¼ ðz :5Þðz 2Þ 1:5 (9:7:3c) &

We considered partial fraction expansions when we studied Laplace transforms and the procedure here is the same with a slight modification. This approach provides closed-form solutions. From Table 9.6.2 we see that z appears in the numerator of the z- transform functions. Therefore, find XðzÞ=z, and then use the partial fraction expansion discussed in Section 5.4. The expansion of a rational function XðzÞ can be obtained by multiplying each term in the expansion by z. The function XðzÞ=z takes the form XðzÞ a0 þ a1 z þ þ aM zM : ¼ b0 þ b1 z þ þ b N z N z

(9:7:5)

If M4N, then divide the numerator polynomial by the denominator polynomial and

392

9 Discrete Data Systems

½XðzÞ=z ¼ RðzÞ þ X0 ðzÞ; RðzÞ ¼ ½cMN zMN þ cMN1 zMN1 þ þ c1 z þ c0 d0 þ d1 z þ þ dN1 zN1 : X0 ðzÞ ¼ b0 þ b1 z þ þ bN zN

(9:7:6)

Note that the numerator polynomial in the rational function X0 ðzÞ in (9.7.6) is of degree (N1) or less. The denominator of this rational function is then factored and expanded using partial fraction expansion. The inverse z-transform of XðzÞ can now be computed:

Z1 ½XðzÞ ¼ Z1 ½zRðzÞ þ Z1 ½zfpartial fraction expansion of X0 ðzÞg ¼ x1 ½n þ x2 ½n;

(9:7:7)

lZ1 ½c0 z ¼ c0 dðn þ 1Þ; Z1 ½c1 z2 ¼ c1 d½n þ 2; :::;

Multiple pole case: See Section 5.8 on the partial fraction expansion with multiple poles. Example 9.7.3 Find x½n ¼ Z1 fXðzÞg for the cases: a: jzj41 and b: jzj5ð1=2Þ :

XðzÞ ¼

3z3 ð5=2Þz2 ðz ð1=2ÞÞ2 ðz 1Þ

Solution: a. The sequence is a right-side sequence since the ROC is jzj41. We have a double pole at z ¼ ð1=2Þ and a single pole at z ¼ 1: XðzÞ ð3z2 ð5=2ÞzÞ ¼ z ðzð1=2ÞÞ2 ðz1Þ A12 A11 A3 þ ; þ ¼ 2 ðzð1=2ÞÞ ðz1Þ ðzð1=2ÞÞ

X0 ðzÞ¼

M X

cMm d½n þ ðM m þ 1Þ:

(9:7:8)

m¼N

In the following we will concentrate on the second part in (9.7.7) by considering all simple poles and later we will consider a single multiple pole plus simple poles.

A11 ¼

¼

d 3z2 ð5=2Þz z¼1=2 dz ðz 1Þ " # ðz 1Þð6z ð5=2Þ ð3z2 ð5=2ÞzÞ ðz 1Þ2

Example 9.7.2 Find the inverse z-transform of the function

XðzÞ ¼

½z2

z½2z ð4=3Þ ; ROC : jzj41: (9:7:9) ð4=3Þz þ ð1=3Þ

Solution: From the ROC the sequence is a rightside sequence. Now XðzÞ 1 1 ¼ þ z z ð1=3Þ z 1 z z þ ; ROC : jzj41: ) XðzÞ ¼ z ð1=3Þ z 1 XðzÞ z

1 1 z z ¼ zð1=3Þ þ z1 ) XðzÞ ¼ zð1=3Þ þ z1 ;

ROC : jzj41 ) x½n ¼ ð1=3Þn u½n þ u½n

:

(9:7:11)

A12 ¼ ðz ð1=2ÞÞ2 X0 ðzÞ z¼1=2 3z2 ð5=2Þz ¼ z¼1=2 ¼ 1; ðz 1Þ

Z1 ½ck zkþ1 ¼ ck d½n þ k þ 1; :::; x1 ½n ¼ Z1 ½zRðzÞ ¼

; ROC : jzj41: (9:7:10)

A3 ¼

X0 ðzÞ ¼

3z2 ð5=2Þz ðz ð1=2ÞÞ2

z¼1=2

¼ 1;

jz¼1 ¼ 2;

ð3z2 ð5=2ÞzÞ

ðz ð1=2ÞÞ2 ðz 1Þ 1 1 2 þ (9:7:12) þ ¼ 2 ðz ð1=2ÞÞ z 1 ðz ð1=2ÞÞ

) XðzÞ ¼

z ðz ð1=2ÞÞ

2

þ

z 2z þ : z ð1=2Þ z 1 (9:7:13)

&

The last step involves determining the inverse transforms. In doing so, the ROC of the z-domain function should be kept in mind. From Table 9.6.2,

9.7 Inverse z-Transforms

393

z z z z ! an u½n; ! nan1 u½n; 2 ðz aÞ ðz aÞ z

z

ðz aÞ

!

3

Solution: Using the partial fraction expansion, we can write

1 nðn 1Þan2 u½n; . . . : 2!

(9:7:14)

X0 ðzÞ ¼

XðzÞ ðz þ 1Þ A1 ¼ z ðz þ 0:5Þðz 2Þðz 0:75Þ ðz þ 0:5Þ þ

The ROC is outside of the unit circle and the inverse transform is x½n ¼ ½nð1=2Þ

n1

n

þ ð1=2Þ þ 2u½n;

(9:7:15)

x½0 ¼ 3; x½1 ¼ 1 þ ð1=2Þ þ 2 ¼ 7=2; X½2 ¼ 1 þ ð1=4Þ þ 2 ¼ 13=4; . . .

(9:7:16)

b. Now use the transform pairs corresponding to the left-side sequences given below: z ; jzj51; z1 z z an u½n 1 ! ; za az

z

nan u½n 1 !

) XðzÞ ¼

z ðz ð1=2ÞÞ

2

ðz aÞ2

þ

(9:7:18)

z 2z þ : z ð1=2Þ z 1 (9:7:19)

u½n 1 2u½n 1:

(9:7:20a)

x½0 ¼ 0; x½1

(9:7:20b)

x½n ¼ f. . . ; 38; 10; 0; 0; . . .g:

(9:7:21)

# &

Notes: It is uncommon to come across multiple poles of order more than 2. It is simple to use the repeated application of simple pole case, see & Example 5.8.2. Example 9.7.4 Find the sequence x½n ¼ z1 fXðzÞg: XðzÞ ¼

A2 ¼

ðz þ 1Þ ¼ 24 ; and ðz þ 0:5Þðz 0:75Þ z¼2 25

ðz þ 1Þ 28 A3 ¼ ¼ ðz þ 0:5Þðz 2Þ z¼0:75 25 4 z 24 z þ 25 ðz þ 0:5Þ 25 ðz 2Þ

28 z : 25 ðz 0:75Þ

Since the ROC is not specified, the sequence cannot be uniquely determined from the XðzÞ alone. Therefore we will identify all possible ROCs corresponding to this function and find the sequences associated with each of them. To find the various possible ROCs, we first make a pole–zero plot as shown in Fig. 9.7.2a. Using the properties of the ROC, the different possible ROCs are shown in Fig. 9.7.2b–e. 1. ROC : jzj5:5: ROC extends inward to include the origin and x½n is a left-sided sequence:

¼ 4 2 2 ¼ 0; x½2 ¼ 16 4 2 ¼ 10; x½3 ¼ 6ð8Þ 8 2 ¼ 38; ::: ;

ðz þ 1Þ 4 ¼ ; ðz 2Þðz 0:75Þ z¼0:5 25

(9:7:17)

; jzj5jaj:

x½n ¼ ð2Þðnð1=2Þn Þu½n 1 ð1=2Þn

A1 ¼

) XðzÞ ¼

z

u½n 1 !

A2 A3 þ ; ðz 2Þ ðz 0:75Þ

zðz þ 1Þ : ðz þ 0:5Þðz 2Þðz 0:75Þ

x½n ¼

n 4 1 24 u½n 1 ð2Þn 25 2 25

n 28 3 u½n 1 þ u½n 1: 25 4

2. ROC : ð1=2Þ5jzj5ð3=4Þ: ROC is a ring and x½n is two sided:

n 4 1 24 u½n ð2Þn x½n ¼ 25 2 25

n 28 3 u½n 1 þ u½n 1: 25 4

394

9 Discrete Data Systems

Im(z)

Im(z)

Im(z)

Unit Circle

1

1 − 2

Unit Circle

2 Re(z)

1

2

1 − 2

3 4

ROC: z

2

(d)

(e)

Fig. 9.7.2 (a) Pole–zero plot, (b) ROC : jzj5:5, (c) ROC : ð1=2Þ5jzj5ð3=4Þ, and (d) ROC : ð3=4Þ5jzj52,ROC : jzj42

3. ROC : ð3=4Þ5jzj52: ROC is a ring and x½n sequence is two sided:

n 4 1 24 u½n þ ð2Þn x½n ¼ 25 2 25

n 28 3 u½n þ u½n 1: 25 4 4. ROC : jzj42: x½n is a right-sided sequence:

n 4 1 24 x½n ¼ u½n þ ð2Þn 25 2 25

n 28 3 u½n u½n: 25 4

&

9.7.3 Inverse z-Transforms by Power Series Expansion From the definition of the z-transform of a sequence x½n, we can write 1 X XðzÞ ¼ x½nzn ¼ þ x½2z2 þ x½1z1 n¼1

þ x½0 þ x½1z1 þ x½2z2 þ :

(9:7:22)

If XðzÞ can be expanded in a power series, x½n can be determined for positive n(negative n) by identifying the coefficients for the negative powers of z (positive powers of z). Example 9.7.5 Find the inverse z-transform using power series for the function in (9.7.10) assuming two cases of ROC identified by a: jzj41; b: jzj5ð1=2Þ. Solution: a. Since the ROC is jzj41, the sequence is a right-side sequence and therefore the power series of the function contain negative powers of z. Divide the numerator by the denominator in (9.6.23) by division. It can be written by ð7=2Þz2 ð15=4Þz þ ð3=4Þ z3 2z2 þ ð5=4Þz ð1=4Þ 7 ð13=4Þz ð29=8Þ þ ð7=8Þz1 ¼ 3 þ z1 þ 3 z 2z2 þ ð5=4Þz ð1=4Þ 2

XðzÞ ¼ 3 þ

) x½0 ¼ 3; x½1 ¼ 7=2; x½2 ¼ 13=4; ::: (9:7:23) b. Since the ROCis inside the circle of radius (1/2), the sequence is a left-side sequence. It can be

9.8 The Unilateral or the One-Sided z-Transform

395

obtained by expanding the function in terms of positive powers. This is achieved by first writing the polynomials in the numerator and the denominator in reverse order and then making use of long division to obtain series in terms of positive powers of z: ð5=2Þz2 þ 3z3 ð1=4Þ þ ð5=4Þz 2z2 þ z3 ð38=4Þz3 þ 20z4 10z3 ¼ 10z2 þ ð1=4Þ þ ð5=4Þz 2z2 þ z3 ) x½1 ¼ 0; x½2 ¼ 10; x½3 ¼ 38; . . . :

XðzÞ ¼

Notes: If the ROC is in an annular region, we can separate the z-transform corresponding to the rightside sequence and the left-side sequence and follow & the above procedure. Example 9.7.6 Find the inverse z-transform of the following function:

An important property of this transform is its ROC and is outside of a circle in the z-plane.

Consider the transform pair x½n ! XI ðzÞ. The transforms of the delayed and advanced sequences are given below and can be shown by starting with the definition of the unilateral transform of these sequences and reducing them into the appropriate forms: z

a: x½n m ! zm XI ðzÞ þ zmþ1 x½1 þ zmþ2 x½2 þ þ x½m; m40;

( ) x½n ¼

ðaz1 Þn

ð1=nÞa ; n 1 0; n 0

zm1 x½1 zx½m 1; m40:

(9:8:3)

a. First consider (9.8.2). The one-sided transform of the delayed sequence is given by

ZI fx½n mg ¼

1 X n¼0

x½n mzn ¼

1 X

x½kzðmþkÞ :

k¼m

(9:7:26)

n

(9:8:2)

z

b: x½n þ m ! zm XI ðzÞ zm x½0

(9:7:25)

Solution: Using the power series expansion Spiegel (1968), we have n n¼1

(9:8:1)

n¼0

1 X 1

x½nzn :

z

&

XðzÞ ¼ logð1 az1 Þ ¼

1 X

9.8.1 Time-Shifting Property

(9:7:24)

1 ; jzj4jaj: XðzÞ ¼ log 1 az1

XI ðzÞ ¼ ZI fx½ng ¼

(9:8:4)

:

The ROC is outside the circle of radius jaj and x½n is & a right-side sequence.

Separating (9.8.4) into two parts x½n; n50 and x½n for n 0, we have 1 X k¼m

x½kzðmþkÞ

¼ fzmþ1 x½1 þ zmþ2 x½2 þ þ x½mg þ zm

9.8 The Unilateral or the One-Sided z-Transform The unilateral transform is useful since most of our sequences are right sided. The causal part of an arbitrary sequence y½n is y½nu½n. The unilateral or one-sided transform is

1 X

x½kzk

k¼0

¼ zm fx½1z þ x½2z2 þ ::: þ x½mzm g þ zm XI ðzÞ; m40:

(9:8:5)

b. Now consider the one-sided transform of the advanced sequence

396

9 Discrete Data Systems

ZI fx½n þ mg ¼

1 X

x½n þ mzn

n¼0

¼

1 X k¼0

x½kzðkmÞ

m1 X

x½kzðkmÞ

k¼0

(9:8:6a)

Note y½n at location n is the sum of the values at the two previous locations. The sequence results in Fibonacci numbers, see Hershey and Yarlagadda (1986). b. Taking the one-sided z-transform of the equation in (9.8.7) results in YI ½z ¼ Zfy½n 1g þ Zfy½n 2g ¼ z1 YI ðzÞ þ y½1 þ z2 YI ðzÞ

¼ fzm x½0 þ x½1zm1 þ þ x½m 1z1 g þ zm XI ½z:

þ z1 y½1 þ y½2

(9:8:6b)

The relations in (9.8.2) and (9.8.3) provide a method to solve constant coefficient difference equations. This generally involves two discrete functions, an output y½n and an input x½n. The procedure parallels that of solving constant coefficient differential equations and the Laplace transform. In the following it is assumed that one-sided transforms are in use and the subscript (I) will not be shown on X explicitly. The procedure involves first finding the z-transform of the difference equation in terms of the two transforms YðzÞ and XðzÞ. In determining YðzÞ, the initial conditions on y½n need to be known. The input x½nand thereforeXðzÞ is assumed to be known. Solve for YðzÞ and then take the inverse transform of this function resulting in y½n. Two simple examples are considered below, one with zero input, but with initial conditions, and the second one has both input and initial conditions. Example 9.8.1 Consider the second-order difference equation given by

¼ z1 YI ðzÞ þ 1 þ z2 YI ðzÞ þ z1 1

(9:8:8)

z1 z ; ¼ 2 ) YI ðzÞ ¼ 1 z2 1 z z pﬃﬃﬃ pzﬃﬃ ﬃ 1 Poles : p1 ¼ ð1=2Þ þ ð 5=2Þ; p2 ¼ ð1=2Þ ð 5=2Þ: (9:8:9) pﬃﬃﬃ The region of convergence is jzj4 12 1 j 5 . The partial fraction expansion is YI ðzÞ 1 A B ¼ þ ¼ 2 z z z 1 z z 1 z z2 A B pﬃﬃﬃ pﬃﬃﬃ þ ; ¼ z ðð1 5Þ=2Þ z ðð1 þ 5Þ=2Þ 1 1 A ¼ pﬃﬃﬃ ; B ¼ pﬃﬃﬃ 5 5 pﬃﬃﬃ pﬃﬃﬃ ð1= 5Þz ð1= 5Þz pﬃﬃﬃ pﬃﬃﬃ ) YI ðzÞ ¼ þ z ðð1 þ 5Þ=2Þ z ðð1 5Þ=2Þ ( pﬃﬃﬃn pﬃﬃﬃn ) 1 1þ 5 1 1 5 pﬃﬃﬃ u½n; ) y½n ¼ pﬃﬃﬃ 2 2 5 5 (9:8:10)

y½n ¼ y½n 1 þ y½n 2; y½2 ¼ 1; y½1 ¼ 1:

y½n :4472ð1:618Þn ; n 1;

(9:8:7)

Solution: a. By the direct method, we have

1 y½0 ¼ pﬃﬃﬃ ½1 1 ¼ 0; y½1 5 pﬃﬃﬃ

pﬃﬃﬃ

5 5 1 1 1 þ ¼ 1: ¼ pﬃﬃﬃ 2 2 2 5 2

Example 9.8.2 Determine y½n ¼ x½n h½n; x½n ¼ f1; 1; 1g; and h½n ¼ f1; 1; 1g. #

y½0 ¼ 0; y½1 ¼ 1; y½2 ¼ 1;

&

#

Determine y½n for several values of n by a. using the equation in (9.8.7) directly and then b. verify this result using the one-sided z-transform. Cadzow (1973) uses (9.8.7) to generate a model for rabbit population.

(9:8:11)

y½3 ¼ 2; y½4 ¼ 3; y½5 ¼ 5;

Solution: The z-transforms of the two finite length sequences x½n and h½n are

y½6 ¼ 8; y½7 ¼ 13; . . . :

XðzÞ ¼ 1 þ z1 þ z2 ;HðzÞ ¼ 1 z1 þ z2 ;

(9:8:12a)

9.9 Discrete-Data Systems

397

YðzÞ ¼ HðzÞXðzÞ ¼ 1 þ z2 þ z4 ) y½n ¼ f1; 0; 1; 0; 1g:

(9:8:12b) &

#

In the convolution of two sequences x½n and h½n are known and y½n ¼ x½n h½n needs to be found. In Section 9.3.4, the deconvolution problem was identified and three methods were discussed. In the deconvolution problem, y½n and x½n are assumed to be known and h½n is to be determined. Such a problem has practical importance, as it is a system identification problem. That is, determine the unit sample response of a system h½n. In the method of deconvolution using polynomial long division, HðzÞ is obtained YðzÞ=XðzÞ. This is illustrated in the following example. Example 9.8.3 Determine h½n using (9.8.12b). Solution: YðzÞ 1 þ z2 þ z4 z1 þ z4 ¼ ¼ 1 þ XðzÞ 1 z1 þ z2 1 þ z1 þ z2 2 3 4 z þz þz ¼ 1 z1 þ ¼ 1 z1 þ z2 1 þ z1 þ z2 ) h½n ¼ f1; 1; 1g &

H ð zÞ ¼

Principles of additivity and proportionality: A system is said to be additive if Tfx1 ½n þ x2 ½ng ¼ Tfx1 ½ng þ Tfx2 ½ng. This is sometimes referred to as the superposition property. A system is homogeneous if it satisfies the principle of proportionality, y½n ¼ Tfax½ng ¼ afx½ng for a constant a and for any input sequence. Linear systems: A system that is both additive and homogeneous is called a linear system. A system is linear if for any inputs xi ½n and for any constants ai ; i ¼ 1; 2, Tfa1 x1 ½n þ a2 x2 ½ng ¼ a1 Tfx1 ½ng þ a2 Tfx2 ½ng:

(9:9:2)

Example 9.9.1 Consider the systems described by the following transformations. In each case, determine whether the corresponding system is linear or nonlinear: a: y1 ½n ¼ Ax½n þ B; B 6¼ 0; b: y2 ½n ¼ x2 ½n; c: y3 ½n ¼ nx½n:

#

Solution: Using (9.9.2), it follows that a1 Tfx1 ½ng ¼ a1 Ax1 ½n þ a1 B ;

9.9 Discrete-Data Systems

a2 Tfx2 ½ng ¼ a2 Ax2 ½n þ a2 B;

In this section, basic concepts associated with discrete-time systems will be discussed. Presentation will be very similar to the continuous-time systems studied in Chapter 6. Our discussion will be brief. A discrete-time system is represented by a block diagram shown in Fig. 9.9.1 mapping x½n into y½n. The T inside the block diagram is some transformation that converts the input data into output data. We can characterize a discrete-time data system by putting constraints on the transformation T: y½n ¼ Tfx½ng:

(9:9:1)

Tfa1 x1 ½n þ a2 x2 ½ng ¼ a1 Ax1 ½n þ a2 Ax2 ½n þ B 6¼ a1 Tfx1 ½ng þ a2 Tfx2 ½ng: a. This indicates that the system is nonlinear. It is a linear system if B ¼ 0: b. It is easy to see that the system is nonlinear. Any time, if the transformation has a power of the input other than one, the system is nonlinear. c. The outputs corresponding to the inputs x1 ðtÞ; x2 ðtÞ and a1 x1 ðtÞ þ a2 x2 ðtÞ are ai Tfxi ½ng ¼ nxi ½n; i ¼ 1; 2 and Tfa1 x1 ðnÞ

x½n ! Tf:g ! y½n: Fig. 9.9.1 A discrete-data system

þ a2 x2 ðnÞg ¼ nfa1 x1 ½n þ a2 x2 ½ng ¼ a1 Tfx1 ½ng þ a2 Tfx2 ½ng ) System is linear:

&

398

9 Discrete Data Systems

Shift (or time) invariance: A discrete-time system is shift invariant if and only if y½n ¼ Tfx½ng implies that for every input signal x½n and every time shift k, Tfx½n kg ¼ y½n k

(9:9:3)

Example 9.9.2 Are the following systems shift invariant? a: y½n ¼ x2 ½n;

system. We will use the impulse response here. If h½n is the response of a linear time-invariant system to the input unit sample d½n, then the response to the input d½n k is h½n k. Using the linearity property, we can write the response of a linear system to an input x½ngiven in (9.9.4). The output is " y½n ¼ Tfx½ng ¼ T

# x½kd½n k

k¼1

¼

b: y½n ¼ nx½n;

1 X

1 X

x½kTfd½n kg ¼

k¼1

c: y½n ¼ x½k0 n; k0 a positive integer: Solution: a. The response of this system corresponding to the input x½n n0 is y1 ½n ¼ Tfx½n n0 g ¼ x2 ½n n0 ¼ y½n n0 . Therefore, the system is shift invariant. b. The response of this system corresponding to the input x½n n0 is y2 ½n ¼ nx½n n0 6¼ ðn n0 Þy ½n n0 . Therefore, the system is shift variant. c. The response of this system corresponding to the input x½n n0 is y3 ½n ¼ x½k0 n n0 6¼ x½k0 ðn n0 Þ. Therefore, the system is a shift-variant system and is a compressor. If k0 ¼ 1 then the sys& tem is shift invariant.

Linear shift-invariant systems: A linear shiftinvariant system (LSI) is linear and also shift invariant. From Chapter 8, we can write that a discretetime signal x½n is 1 X x½kd½n k: (9:9:4) x½n ¼ k¼1

Modeling a discrete-data system is an important topic of study. In the case of analog systems we have defined the impulse response and we called the transform of the impulse response as the transfer function of the system. In a similar manner we can define the unit sample (discrete impulse) response or simply impulse response. To stick with the analog impulse response notation, many authors use impulse response rather than a unit sample response. The unit sample response or the impulse response provides a complete description of a linear shift-invariant

1 X

x½kh½n k

k¼1

¼ x½n h½k:

(9:9:5)

The output of an LSI system is the convolution of the input sequence with the unit impulse response h½n, which gives a complete characterization of the LSI system. Causality: A causal signal x½n is zero for n50. A system is causal if, for any time n, the response of the system y½n depends only on the present and the past inputs x½n; x½n 1; :::; and does not depend on the future inputs x½n þ 1; x½n þ 2; :::. The response of a causal system satisfies the following, where ffg is an arbitrary function: y½n ¼ ffx½n; x½n 1; :::g:

(9:9:6)

If a system does not satisfy this constraint, then the system is non-causal. In real-time processing applications, we cannot predict the future values and therefore non-causal systems are not physically realizable. In the case of off-line processing, i.e., if we have all the values of the signal, then it is possible to design a non-causal system to process the data. Such a situation is common in data processing. Example 9.9.3 Classify each of the following systems are causal or not; a: y½n ¼ x½n x½n 1 þ x½n 2; b: y½n ¼ x½n þ x½n þ 1; c: y½n ¼ x½2n : Solution: The system in part a is causal, whereas the systems in parts b and care non-causal as they & require the knowledge of future values.

9.9 Discrete-Data Systems

399

The systems described by constant coefficient difference equations given below are linear shift-invariant systems: y½n ¼

N X

ak y½n k þ

k¼1

L X

bk x½n k:

(9:9:7)

k¼0

A system described by (9.9.7) is a recursive linear system if at least one of the coefficients ak is not zero and the output depends on previous values of the output as well as the input. For a non-recursive linear system, a½k ¼ 0 and is described by the difference equation, y½n ¼

L X

1: jlj51; lim ln ! 0; n!1

bk x½n k:

(9:9:8)

k¼0

(9:9:9a) For a system to be BIBO stable, i.e., to have a bounded output jy½nj51 to a bounded input, the unit sample response of a shift-invariant system must be absolutely summable: jh½kj51:

2: jlj41; lim ln ! 1; n!1 n

3: jlj ¼ 1; jlj ! 1 for all n:

Stability: The stability of a discrete system can be defined in terms of the input–output behavior as in the analog case. A system is called BIBO stable if every bounded input produces a bounded output. For linear systems BIBO stability requires that the sample response h½n must be absolutely summable. We can show this by starting with a bounded input such that jx½nj M for all n and the convolution sum in (9.9.5). That is, 1 1 X X h½kx½n k M jy½nj jh½kj: k¼1 k¼1

1 X

be asymptotically stable, if and only if the natural response goes to zero as n ! 1, and is unstable if it grows without bound. We can see this very clearly by making use of the transforms and by expanding the z-domain function into its partial fraction expansion and then taking the inverse terms. The natural response depends upon the characteristic modes of the system. The roots of the characteristic polynomial of the system are the modes. Let a typical root be given by z ¼ l ¼ jljejb . Noting that ln ¼ jljn ejbn , we can summarize the results using the three cases:

(9:9:9b)

k¼1

Example 9.9.4 Show the system described by y½n ¼ njx½nj; x½n ¼ Au½n; A40 and finite is not BIBO stable. Solution: The output y½n ! 1 as n ! 1 and the & system is not BIBO stable. Notes: The response of a discrete-time linear timeinvariant system y½n consists of two parts, one due to the natural response (due to initial conditions) and the second due to the source. A system is said to

(9:9:10)

The linear discrete-time system is asymptotically stable if and only if the characteristic roots, i.e., the poles of the transfer function of the system, are inside the unit circle. It is unstable if there is at least one root outside the unit circle and/or if there are multiple roots on the unit circle. It is marginally stable if and only if there are no roots outside the unit circle and only simple roots on the unit circle. A marginally stable system is not BIBO stable. Discussion on general stability analysis is beyond the scope here. Example 9.9.5 Show that the system described by the following equation is stable: y½n þ 2 þ y½n þ 1 þ 2y½n ¼ x½n þ 1 þ x½n: Solution: The characteristic polynomial can be obtained as follows: zfy½n þ 2 þ y½n þ 1 þ 2y½ng ¼ z2 þ z þ 2; ¼ ðz þ z1 Þðz þ z2 Þ ¼ 0: pﬃﬃﬃ ) z1 ¼ ð1=2Þ þ j 7=2; z2 ¼ z 1 ; z1 z2 ¼ 2. Since jzi j41

the system is unstable.

&

Classification of LSI systems based on the duration of the impulse response: We can classify the linear shift-invariant (LSI) systems based upon the duration of their (discrete impulse or simply impulse) responses. Without losing any generality, we will

400

9 Discrete Data Systems

consider causal systems. The systems that have finite-duration impulse response (FIR) are called FIR systems. On the other hand, the systems with infinite-duration impulse response are called IIR systems. The system described by (9.9.7) is an IIR system if at least one ak 6¼ 0, whereas the system described by (9.9.8) is an FIRsystem. As in the analog systems the transform analysis is basic to filter designs. Since the systems need to work with any set of initial conditions, the designs are based on zero initial conditions.

impulse response h½n or the discrete-time transfer z function HðzÞ ! h½n. Since the impulse response and the corresponding transfer function are related, we can discuss the stability of a discrete linear timeinvariant system in terms of the poles of the transfer function. The transfer function in (9.9.11b) is L P

bk z k¼0 N P

HðzÞ ¼

k

ak zk

1þ

(9:9:12a)

k¼1

¼K

ðz z1 Þðz z2 Þ ðz zL Þ : ðz p1 Þðz p2 Þ ðz pN Þ

9.9.1 Discrete-Time Transfer Functions Consider the difference equation in (9.9.7). Assuming the initial conditions are all equal to zero and taking the z-transform of this equation result in N L X X YðzÞ ¼ ak zk YðzÞ þ bk zk XðzÞ; (9:9:11a) k¼1

k¼0 L P

) YðzÞ ¼

bk z k¼0 N P

k

XðzÞ ¼ HðzÞXðzÞ;

ak zk

1þ

k¼1 L P

YðzÞ HðzÞ ¼ ¼ XðzÞ

bk z k¼0 N P

1þ

k

ak

:

(9:9:11b)

zk

k¼1

As in the analog case HðzÞ is called the transfer function and h½n is the impulse response or the unit sample response. A discrete-time linear time-invariant system can be described by its difference equation, its transfer function, or by its poles and zeros. From our earlier discussion on the z-transforms, we can write the expressions for the system input–output relations in the time domain or in terms of the transform domain. That is,

Notes: he poles of the transfer function are called the natural frequencies or natural modes and they determine the time domain behavior of the system response. For example, if we have poles outside the unit circle, the response grows exponentially and the system is unstable. If we have multiple poles on the unit circle, then the response has polynomial growth. If a system has a simple pole on the unit circle, then the system is referred to as marginally stable. If we require that the function is a minimum phase function, then all the zeros and poles must be & inside the unit circle. Special cases of the general model are useful in defining digital filters. These are 1. the autoregressive moving average filter (ARMA); moving average filter (MA); and the autoregressive filter (AR). These are explicitly expressed by L P

bk z k¼0 N P

HARMA ðzÞ ¼

1þ

ak

1

(9:9:11c)

; zk

k¼1

YðzÞ ¼ HARMA ðzÞXðzÞ; y½n ¼

N X

ak y½n k þ

k¼1

L X

bk x½n k;

k¼0

(9:9:12b)

z

y½n ¼ h½n x½n $ YðzÞ ¼ HðzÞXðzÞ;

k

HMA ðzÞ ¼

HðzÞ ¼ zfh½ng; h½n ¼ z fHðzÞg:

L X

bk zk ;

k¼0

YðzÞ ¼ HMA ðzÞXðzÞ; Linear time-invariant discrete-time systems are described by either constant coefficient difference equations relating the output to the input or the

y½n ¼

L X k¼0

bk x½n k

(9:9:12c)

9.9 Discrete-Data Systems

HAR ðzÞ ¼ 1þ

1 N P

401

;YðzÞ ¼ HAR ðzÞXðzÞ; ak zk

The polynomial AðzÞ given in (9.9.13) has all its roots inside the unit circle if and only if the coefficients jKm j51; m ¼ 1; 2; :::; N.

k¼1

y½n ¼ x½n

N X

ak y½n k:

(9:9:12d)

k¼0

All three models are used in different applications. They are related, at least in the limit, see Marple (1987). The AR models are used extensively in spectral estimation. AR and ARMA models are used in digital

R.K. Rao Yarlagadda

Analog and Digital Signals and Systems

13

R.K. Rao Yarlagadda School of Electrical & Computer Engineering Oklahoma State University Stillwater OK 74078-6028 202 Engineering South USA [email protected]

ISBN 978-1-4419-0033-3 e-ISBN 978-1-4419-0034-0 DOI 10.1007/978-1-4419-0034-0 Springer New York Dordrecht Heidelberg London Library of Congress Control Number: 2009929744 # Springer ScienceþBusiness Media, LLC 2010 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer ScienceþBusiness Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed on acid-free paper Springer is part of Springer ScienceþBusiness Media (www.springer.com)

This book is dedicated to my wife Marceil, children, Tammy Bardwell, Ryan Yarlagadda and Travis Yarlagadda and their families

Note to Instructors

The solutions manual can be located on the book’s webpage http://www/ springer.com/engineering/cirucitsþ %26þsystems/bok/978-1-4419-0033-3

vii

Preface

This book presents a systematic, comprehensive treatment of analog and discrete signal analysis and synthesis and an introduction to analog communication theory. This evolved from my 40 years of teaching at Oklahoma State University (OSU). It is based on three courses, Signal Analysis (a second semester junior level course), Active Filters (a first semester senior level course), and Digital signal processing (a second semester senior level course). I have taught these courses a number of times using this material along with existing texts. The references for the books and journals (over 160 references) are listed in the bibliography section. At the undergraduate level, most signal analysis courses do not require probability theory. Only, a very small portion of this topic is included here. I emphasized the basics in the book with simple mathematics and the sophistication is minimal. Theorem-proof type of material is not emphasized. The book uses the following model: 1. Learn basics 2. Check the work using bench marks 3. Use software to see if the results are accurate The book provides detailed examples (over 400) with applications. A threenumber system is used consisting of chapter number – section number – example or problem number, thus allowing the student to quickly identify the related material in the appropriate section of the book. The book includes well over 400 homework problems. Problem numbers are identified using the above three-number system. Hints are provided wherever additional details may be needed and may not have been given in the main part of the text. A detailed solution manual will be available from the publisher for the instructors.

Summary of the Chapters This book starts with an introductory chapter that includes most of the basic material that a junior in electrical engineering had in the beginning classes. For those who have forgotten, or have not seen the material recently, it gives enough

ix

x

background to follow the text. The topics in this chapter include singularity functions, periodic functions, and others. Chapter 2 deals with convolution and correlation of periodic and aperiodic functions. Chapter 3 deals with approximating a function by using a set of basis functions, referred to as the generalized Fourier series expansion. From these concepts, the three basic Fourier series expansions are derived. The discussion includes detailed discussion on the operational properties of the Fourier series and their convergence. Chapter 4 deals with Fourier transform theory derived from the Fourier series. Fourier series and transforms are the bases to this text. Considerable material in the book is based on these topics. Chapter 5 deals with the relatives of the Fourier transforms, including Laplace, cosine and sine, Hartley and Hilbert transforms. Chapter 6 deals with basic systems analysis that includes linear timeinvariant systems, stability concepts, impulse response, transfer functions, linear and nonlinear systems, and very simple filter circuits and concepts. Chapter 7 starts with the Bode plots and later deals with approximations using classical analog Butterworth, Chebyshev, and Bessel filter functions. Design techniques, based on both amplitude and phase based, are discussed. Last part of this chapter deals with analysis and synthesis of active filter circuits. Examples of basic low-pass, high-pass, band-pass, band elimination, and delay line filters are included. Chapter 8 builds a bridge to go from the continuous-time to discrete-time analysis by starting with sampling theory and the Fourier transform of the ideally sampled signals. Bulk of this chapter deals with discrete basis functions, discrete-time Fourier series, discrete-time Fourier transform (DTFT), and the discrete Fourier transform (DFT). Chapter 9 deals with fast implementations of the DFT, discrete convolution, and correlation. Second part of the chapter deals the z-transforms and their use in the design of discrete-data systems. Digital filter designs based on impulse invariance and bilinear transformations are presented. The chapter ends with digital filter realizations. Chapter 10 presents an introduction to analog communication theory, which includes basic material on analog modulation, such as AM and FM, demodulation, and multiplexing. Pulse modulation methods are introduced. Appendix A reviews the basics on matrices; Appendix B gives a brief introduction on MATLAB; and Appendix C gives a list of useful formulae. The book concludes with a list of references and Author and Subject indexes.

Suggested Course Content Instructor is the final judge of what topics will best suit his or her class and in what depth. The suggestions given below are intended to serve as a guide only. The book permits flexibility in teaching analysis, synthesis of continuous-time and discrete-time systems, analog filters, digital signal processing, and an introduction to analog communications. The following table gives suggestions for courses.

Preface

Preface

xi

Topical Title

Related topics in chapters

One semester

(

One semester

Systems and analog filters

Chapters 4, 5*, 6, 7

One semester

( (

Chapters 4*, 6*, 8, 9

Two semesters *Partial coverage

)

Fundamentals of analog signals and systems

Chapters 1–4, 6

)

Introduction to digital signal processing

)

Signals and an introduction to analog communications

Chapters 1–4, 5*, 6, 8*, 10

Acknowledgements

The process of writing this book has taken me several years. I am indebted to all the students who have studied with me and taken classes from me. Education is a two-way street. The teachers learn from the students, as well as the students learn from the teachers. Writing a book is a learning process. Dr. Jack Cartinhour went through the material in the early stages of the text and helped me in completing the solution manual. His suggestions made the text better. I am deeply indebted to him. Dr. George Scheets used an earlier version of this book in his signal analysis and communications theory class. Dr. Martin Hagan has reviewed a chapter. Their comments were incorporated into the manuscript. Beau Lacefield did most of the artwork in the manuscript. Vijay Venkataraman and Wen Fung Leong have gone through some of the chapters and their suggestions have been incorporated. In addition, Vijay and Wen have provided some of the MATLAB programs and artwork. I appreciated Vijay’s help in formatting the final version of the manuscript. An old adage of the uncertainty principle is, no matter how many times the author goes through the text, mistakes will remain. I sincerely appreciate all the support provided by Springer. Thanks to Alex Greene. He believed in me to complete this project. I appreciated the patience and support of Katie Chen. Thanks to Shanty Jaganathan and her associates of Integra-India. They have been helpful and gracious in the editorial process. Dr. Keith Teague, Head, School of Electrical and Computer Engineering at Oklahoma State University has been very supportive of this project and I appreciated his encouragement. Finally, the time spent on this book is the time taken away from my wife Marceil, children Tammy, Ryan and Travis and my grandchildren. Without my family’s understanding, I could not have completed this book. Oklahoma, USA

R.K. Rao Yarlagadda

xiii

Contents

1

Basic Concepts in Signals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Introduction to the Book and Signals . . . . . . . . . . . . . . . . . . . . 1.1.1 Different Ways of Looking at a Signal . . . . . . . . . . . . . . 1.1.2 Continuous-Time and Discrete-Time Signals . . . . . . . . . 1.1.3 Analog Versus Digital Signal Processing . . . . . . . . . . . . 1.1.4 Examples of Simple Functions . . . . . . . . . . . . . . . . . . . . 1.2 Useful Signal Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 Time Shifting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.2 Time Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.3 Time Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.4 Amplitude Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.5 Simple Symmetries: Even and Odd Functions . . . . . . . . 1.2.6 Products of Even and Odd Functions . . . . . . . . . . . . . . . 1.2.7 Signum (or sgn) Function . . . . . . . . . . . . . . . . . . . . . . . . 1.2.8 Sinc and Sinc2 Functions. . . . . . . . . . . . . . . . . . . . . . . . . 1.2.9 Sine Integral Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Derivatives and Integrals of Functions . . . . . . . . . . . . . . . . . . . 1.3.1 Integrals of Functions with Symmetries . . . . . . . . . . . . . 1.3.2 Useful Functions from Unit Step Function . . . . . . . . . . 1.3.3 Leibniz’s Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.4 Interchange of a Derivative and an Integral . . . . . . . . . . 1.3.5 Interchange of Integrals. . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Singularity Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 Unit Impulse as the Limit of a Sequence. . . . . . . . . . . . . 1.4.2 Step Function and the Impulse Function . . . . . . . . . . . . 1.4.3 Functions of Generalized Functions . . . . . . . . . . . . . . . . 1.4.4 Functions of Impulse Functions . . . . . . . . . . . . . . . . . . . 1.4.5 Functions of Step Functions . . . . . . . . . . . . . . . . . . . . . . 1.5 Signal Classification Based on Integrals . . . . . . . . . . . . . . . . . . 1.5.1 Effects of Operations on Signals . . . . . . . . . . . . . . . . . . . 1.5.2 Periodic Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.3 Sum of Two Periodic Functions . . . . . . . . . . . . . . . . . . . 1.6 Complex Numbers, Periodic, and Symmetric Periodic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Complex Numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 1 1 3 5 6 8 8 8 8 8 9 9 10 10 10 11 12 12 13 13 13 14 15 16 17 18 19 19 21 21 23 24 25 xv

xvi

2

3

Contents

1.6.2 Complex Periodic Functions . . . . . . . . . . . . . . . . . . . . . . 1.6.3 Functions of Periodic Functions . . . . . . . . . . . . . . . . . . . 1.6.4 Periodic Functions with Additional Symmetries. . . . . . . 1.7 Examples of Probability Density Functions and their Moments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.8 Generation of Periodic Functions from Aperiodic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9 Decibel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.10 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27 27 28

Convolution and Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 Scalar Product and Norm . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Properties of the Convolution Integral . . . . . . . . . . . . . . 2.2.2 Existence of the Convolution Integral. . . . . . . . . . . . . . . 2.3 Interesting Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Convolution and Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 Repeated Convolution and the Central Limit Theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Deconvolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 Convolution Involving Periodic and Aperiodic Functions . . . . 2.5.1 Convolution of a Periodic Function with an Aperiodic Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.2 Convolution of Two Periodic Functions. . . . . . . . . . . . . 2.6 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.1 Basic Properties of Cross-Correlation Functions . . . . . . 2.6.2 Cross-Correlation and Convolution . . . . . . . . . . . . . . . . 2.6.3 Bounds on the Cross-Correlation Functions . . . . . . . . . 2.6.4 Quantitative Measures of Cross-Correlation . . . . . . . . . 2.7 Autocorrelation Functions of Energy Signals . . . . . . . . . . . . . . 2.8 Cross- and Autocorrelation of Periodic Functions . . . . . . . . . . 2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39 39 40 41 41 44 44 50

Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Orthogonal Basis Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Gram–Schmidt Orthogonalization . . . . . . . . . . . . . . . . . 3.3 Approximation Measures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 Computation of c[k] Based on Partials . . . . . . . . . . . . . . 3.3.2 Computation of c[k] Using the Method of Perfect Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.3 Parseval’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.1 Complex Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.2 Trigonometric Fourier Series . . . . . . . . . . . . . . . . . . . . . 3.4.3 Complex F-series and the Trigonometric F-series Coefficients-Relations . . . . . . . . . . . . . . . . . . . . . . . . . . .

71 71 72 74 75 77

29 31 32 34 35

52 53 54 54 55 56 57 57 58 59 63 65 68 68

77 78 80 80 83 83

Contents

xvii

3.4.4 3.4.5 3.4.6

4

Harmonic Form of Trigonometric Fourier Series. . . . . Parseval’s Theorem Revisited . . . . . . . . . . . . . . . . . . . . Advantages and Disadvantages of the Three Forms of Fourier Series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Fourier Series of Functions with Simple Symmetries. . . . . . . . 3.5.1 Simplification of the Fourier Series Coefficient Integral . 3.6 Operational Properties of Fourier Series . . . . . . . . . . . . . . . . . 3.6.1 Principle of Superposition . . . . . . . . . . . . . . . . . . . . . . . 3.6.2 Time Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.3 Time and Frequency Scaling . . . . . . . . . . . . . . . . . . . . . 3.6.4 Fourier Series Using Derivatives . . . . . . . . . . . . . . . . . . 3.6.5 Bounds and Rates of Fourier Series Convergence by the Derivative Method . . . . . . . . . . . . . . . . . . . . . . 3.6.6 Integral of a Function and Its Fourier Series . . . . . . . . 3.6.7 Modulation in Time. . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.8 Multiplication in Time. . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.9 Frequency Modulation . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.10 Central Ordinate Theorems . . . . . . . . . . . . . . . . . . . . . . 3.6.11 Plancherel’s Relation (or Theorem). . . . . . . . . . . . . . . . 3.6.12 Power Spectral Analysis . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Convergence of the Fourier Series and the Gibbs Phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.1 Fourier’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.2 Gibbs Phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.3 Spectral Window Smoothing. . . . . . . . . . . . . . . . . . . . . 3.8 Fourier Series Expansion of Periodic Functions with Special Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.8.1 Half-Wave Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . 3.8.2 Quarter-Wave Symmetry. . . . . . . . . . . . . . . . . . . . . . . 3.8.3 Even Quarter-Wave Symmetry . . . . . . . . . . . . . . . . . . 3.8.4 Odd Quarter-Wave Symmetry. . . . . . . . . . . . . . . . . . . 3.8.5 Hidden Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.9 Half-Range Series Expansions . . . . . . . . . . . . . . . . . . . . . . . . . 3.10 Fourier Series Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.11 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

83 84

100 100 102 102 102 103 103 104 104 106

Fourier Transform Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Fourier Series to Fourier Integral . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Amplitude and Phase Spectra . . . . . . . . . . . . . . . . . . . . . 4.2.2 Bandwidth-Simplistic Ideas . . . . . . . . . . . . . . . . . . . . . . . 4.3 Fourier Transform Theorems, Part 1. . . . . . . . . . . . . . . . . . . . . 4.3.1 Rayleigh’s Energy Theorem . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Superposition Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.3 Time Delay Theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.4 Scale Change Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.5 Symmetry or Duality Theorem . . . . . . . . . . . . . . . . . . . . 4.3.6 Fourier Central Ordinate Theorems . . . . . . . . . . . . . . . .

109 109 109 112 114 114 114 115 116 116 118 119

85 85 86 87 87 87 88 89 91 93 93 94 95 95 95 95 96 96 97 99

xviii

Contents

4.4

Fourier Transform Theorems, Part 2 . . . . . . . . . . . . . . . . . . . . 4.4.1 Frequency Translation Theorem . . . . . . . . . . . . . . . . . 4.4.2 Modulation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.3 Fourier Transforms of Periodic and Some Special Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.4 Time Differentiation Theorem . . . . . . . . . . . . . . . . . . 4.4.5 Times-t Property: Frequency Differentiation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.6 Initial Value Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.7 Integration Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Convolution and Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 Convolution in Time . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.2 Proof of the Integration Theorem . . . . . . . . . . . . . . . . 4.5.3 Multiplication Theorem (Convolution in Frequency) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.4 Energy Spectral Density . . . . . . . . . . . . . . . . . . . . . . . 4.6 Autocorrelation and Cross-Correlation . . . . . . . . . . . . . . . . . . 4.6.1 Power Spectral Density . . . . . . . . . . . . . . . . . . . . . . . . 4.7 Bandwidth of a Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.1 Measures Based on Areas of the Time and Frequency Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.2 Measures Based on Moments . . . . . . . . . . . . . . . . . . . 4.7.3 Uncertainty Principle in Fourier Analysis. . . . . . . . . . 4.8 Moments and the Fourier Transform . . . . . . . . . . . . . . . . . . . 4.9 Bounds on the Fourier Transform . . . . . . . . . . . . . . . . . . . . . . 4.10 Poisson’s Summation Formula . . . . . . . . . . . . . . . . . . . . . . . . 4.11 Interesting Examples and a Short Fourier Transform Table . . 4.11.1 Raised-Cosine Pulse Function. . . . . . . . . . . . . . . . . . . 4.12 Tables of Fourier Transforms Properties and Pairs. . . . . . . . . 4.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Relatives of Fourier Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Fourier Cosine and Sine Transforms . . . . . . . . . . . . . . . . . . . . . 5.3 Hartley Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.1 Region of Convergence (ROC) . . . . . . . . . . . . . . . . . . . . 5.4.2 Inverse Transform of Two-Sided Laplace Transform. . . 5.4.3 Region of Convergence (ROC) of Rational Functions – Properties. . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Basic Two-Sided Laplace Transform Theorems . . . . . . . . . . . . 5.5.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.2 Time Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.3 Shift in s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.4 Time Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.5 Time Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.6 Differentiation in Time . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.7 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.8 Convolution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

119 120 120 121 124 126 128 128 129 129 132 133 135 136 138 139 139 140 141 143 144 145 145 146 147 147 147 155 155 156 159 161 163 164 165 165 165 165 165 165 166 166 166 166

Contents

xix

5.6

One-Sided Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.1 Properties of the One-Sided Laplace Transform . . . . . 5.6.2 Comments on the Properties (or Theorems) of Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . 5.7 Rational Transform Functions and Inverse Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.7.1 Rational Functions, Poles, and Zeros . . . . . . . . . . . . . 5.7.2 Return to the Initial and Final Value Theorems and Their Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.8 Solutions of Constant Coefficient Differential Equations Using Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.8.1 Inverse Laplace Transforms . . . . . . . . . . . . . . . . . . . . 5.8.2 Partial Fraction Expansions . . . . . . . . . . . . . . . . . . . . 5.9 Relationship Between Laplace Transforms and Other Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.9.1 Laplace Transforms and Fourier Transforms . . . . . . . 5.9.2 Hartley Transforms and Laplace Transforms . . . . . . . 5.10 Hilbert Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.10.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.10.2 Hilbert Transform of Signals with Non-overlapping Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.10.3 Analytic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.11 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6

Systems and Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Linear Systems, an Introduction . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 Two-Terminal Component Equations . . . . . . . . . . . . . . 6.3.2 Kirchhoff’s Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Time-Invariant and Time-Varying Systems . . . . . . . . . . . . . . . . 6.5 Impulse Response. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.1 Eigenfunctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.2 Bounded-Input/Bounded-Output (BIBO) Stability . . . . 6.5.3 Routh–Hurwitz Criterion (R–H criterion) . . . . . . . . . . . 6.5.4 Eigenfunctions in the Fourier Domain . . . . . . . . . . . . . . 6.6 Step Response. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7 Distortionless Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7.1 Group Delay and Phase Delay . . . . . . . . . . . . . . . . . . . . 6.8 System Bandwidth Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.8.1 Bandwidth Measures Using the Impulse Response hðtÞ and Its Transform Hðj!Þ . . . . . . . . . . . . . . . . . . . . . 6.8.2 Half-Power or 3 dB Bandwidth. . . . . . . . . . . . . . . . . . . . 6.8.3 Equivalent Bandwidth or Noise Bandwidth . . . . . . . . . . 6.8.4 Root Mean-Squared (RMS) Bandwidth . . . . . . . . . . . . . 6.9 Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.9.1 Distortion Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . .

166 167 167 174 175 176 178 179 179 183 184 185 186 186 188 189 190 190 193 193 193 194 195 197 198 199 202 202 203 206 208 213 213 216 216 217 217 218 219 220

xx

Contents

6.9.2 Output Fourier-Transform of a Nonlinear System . . . 6.9.3 Linearization of Nonlinear System Functions . . . . . . 6.10 Ideal Filters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.10.1 Low-Pass, High-Pass, Band-Pass, and Band-Elimination Filters . . . . . . . . . . . . . . . . . . . . . . . 6.11 Real and Imaginary Parts of the Fourier Transform of a Causal Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.11.1 Relationship Between Real and Imaginary Parts of the Fourier Transform of a Causal Function Using Hilbert Transform . . . . . . . . . . . . . . . . . . . . . . . 6.11.2 Amplitude Spectrum jHðj!Þj to a Minimum Phase Function HðsÞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12 More on Filters: Source and Load Impedances . . . . . . . . . . . . 6.12.1 Simple Low-Pass Filters . . . . . . . . . . . . . . . . . . . . . . . 6.12.2 Simple High-Pass Filters . . . . . . . . . . . . . . . . . . . . . . . 6.12.3 Simple Band-Pass Filters . . . . . . . . . . . . . . . . . . . . . . . 6.12.4 Simple Band-Elimination or Band-Reject or Notch Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12.5 Maximum Power Transfer. . . . . . . . . . . . . . . . . . . . . . 6.12.6 A Simple Delay Line Circuit . . . . . . . . . . . . . . . . . . . . 6.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Approximations and Filter Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Bode Plots. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Gain and Phase Margins . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 Classical Analog Filter Functions . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 Amplitude-Based Design . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.2 Butterworth Approximations . . . . . . . . . . . . . . . . . . . . . 7.3.3 Chebyshev (Tschebyscheff) Approximations . . . . . . . . . 7.4 Phase-Based Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.1 Maximally Flat Delay Approximation . . . . . . . . . . . . . . 7.4.2 Group Delay of Bessel Functions . . . . . . . . . . . . . . . . . . 7.5 Frequency Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.1 Normalized Low-Pass to High-Pass Transformation . . . 7.5.2 Normalized Low-Pass to Band-Pass Transformation. . . 7.5.3 Normalized Low-Pass to Band-Elimination Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.4 Conversions of Specifications from Low-Pass, High-Pass, Band-Pass, and Band Elimination Filters to Normalized Low-Pass Filters . . . . . . . . . . . . . . . . . . . 7.6 Multi-terminal Components. . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6.1 Two-Port Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6.2 Circuit Analysis Involving Multi-terminal Components and Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6.3 Controlled Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.7 Active Filter Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.7.1 Operational Amplifiers, an Introduction . . . . . . . . . . . .

220 221 221 222 227

228 229 229 231 231 233 235 238 239 239 239 243 243 246 252 254 254 255 257 262 263 264 266 266 268 268

270 273 273 277 278 279 279

Contents

xxi

8

7.7.2 Inverting Operational Amplifier Circuits . . . . . . . . . . . 7.7.3 Non-inverting Operational Amplifier Circuits . . . . . . . 7.7.4 Simple Second-Order Low-Pass and All-Pass Circuits. .. 7.8 Gain Constant Adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.9 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.9.1 Amplitude (or Magnitude) Scaling, RLC Circuits . . . . 7.9.2 Frequency Scaling, RLC Circuits . . . . . . . . . . . . . . . . . 7.9.3 Amplitude and Frequency Scaling in Active Filters . . . 7.9.4 Delay Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.10 RC–CR Transformations: Low-Pass to High-Pass Circuits . . 7.11 Band-Pass, Band-Elimination and Biquad Filters . . . . . . . . . . 7.12 Sensitivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

280 282 284 285 287 287 288 288 290 292 294 298 301 301

Discrete-Time Signals and Their Fourier Transforms . . . . . . . . . . . . . 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Sampling of a Signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1 Ideal Sampling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2 Uniform Low-Pass Sampling or the Nyquist Low-Pass Sampling Theorem . . . . . . . . . . . . . . . . . . . . . 8.2.3 Interpolation Formula and the Generalized Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.4 Problems Associated with Sampling Below the Nyquist Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.5 Flat Top Sampling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.6 Uniform Band-Pass Sampling Theorem . . . . . . . . . . . . . 8.2.7 Equivalent continuous-time and discrete-time systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Basic Discrete-Time (DT) Signals . . . . . . . . . . . . . . . . . . . . . . . 8.3.1 Operations on a Discrete Signal . . . . . . . . . . . . . . . . . . . 8.3.2 Discrete-Time Convolution and Correlation . . . . . . . . . 8.3.3 Finite duration, right-sided, left-sided, two-sided, and causal sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.4 Discrete-Time Energy and Power Signals . . . . . . . . . . . . 8.4 Discrete-Time Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.1 Periodic Convolution of Two Sequences with the Same Period. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.2 Parseval’s Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5 Discrete-Time Fourier Transforms . . . . . . . . . . . . . . . . . . . . . . 8.5.1 Discrete-Time Fourier Transforms (DTFTs) . . . . . . . . . 8.5.2 Discrete-Time Fourier Transforms of Real Signals with Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6 Properties of the Discrete-Time Fourier Transforms. . . . . . . . . 8.6.1 Periodic Nature of the Discrete-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.2 Superposition or Linearity . . . . . . . . . . . . . . . . . . . . . . . 8.6.3 Time Shift or Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.4 Modulation or Frequency Shifting . . . . . . . . . . . . . . . . .

311 311 312 312 314 317 319 322 324 325 325 327 329 330 330 332 334 334 335 335 336 339 339 340 341 341

xxii

Contents

8.6.5 Time Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.6 Differentiation in Frequency . . . . . . . . . . . . . . . . . . . 8.6.7 Differencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.8 Summation or Accumulation . . . . . . . . . . . . . . . . . . 8.6.9 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.10 Multiplication in Time. . . . . . . . . . . . . . . . . . . . . . . . 8.6.11 Parseval’s Identities . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6.12 Central Ordinate Theorems . . . . . . . . . . . . . . . . . . . . 8.6.13 Simple Digital Encryption . . . . . . . . . . . . . . . . . . . . . 8.7 Tables of Discrete-Time Fourier Transform (DTFT) Properties and Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.8 Discrete-Time Fourier-transforms from Samples of the Continuous-Time Fourier-Transforms . . . . . . . . . . . . . . . . . . 8.9 Discrete Fourier Transforms (DFTs) . . . . . . . . . . . . . . . . . . . . 8.9.1 Matrix Representations of the DFT and the IDFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.9.2 Requirements for Direct Computation of the DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10 Discrete Fourier Transform Properties . . . . . . . . . . . . . . . . . . 8.10.1 DFTs and IDFTs of Real Sequences. . . . . . . . . . . . . 8.10.2 Linearity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.3 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.4 Time Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.5 Frequency Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.6 Even Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.7 Odd Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.8 Discrete-Time Convolution Theorem . . . . . . . . . . . . 8.10.9 Discrete-Frequency Convolution Theorem . . . . . . . . 8.10.10 Discrete-Time Correlation Theorem . . . . . . . . . . . . . 8.10.11 Parseval’s Identity or Theorem . . . . . . . . . . . . . . . . . 8.10.12 Zero Padding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.13 Signal Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10.14 Decimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.11 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Discrete Data Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 Computation of Discrete Fourier Transforms (DFTs) . . . . . . . 9.2.1 Symbolic Diagrams in Discrete-Time Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2.2 Fast Fourier Transforms (FFTs). . . . . . . . . . . . . . . . . . . 9.3 DFT (FFT) Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.1 Hidden Periodicity in a Signal. . . . . . . . . . . . . . . . . . . . . 9.3.2 Convolution of Time-Limited Sequences . . . . . . . . . . . . 9.3.3 Correlation of Discrete Signals . . . . . . . . . . . . . . . . . . . . 9.3.4 Discrete Deconvolution. . . . . . . . . . . . . . . . . . . . . . . . . . 9.4 z-Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.1 Region of Convergence (ROC) . . . . . . . . . . . . . . . . . . . .

341 342 342 344 344 345 346 346 346 347 348 350 352 353 354 354 354 355 355 356 356 356 357 358 359 359 359 360 361 361 361 367 367 368 368 369 372 372 374 377 378 380 381

Contents

xxiii

9.4.2

z-Transform and the Discrete-Time Fourier Transform (DTFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . 384 9.5 Properties of the z-Transform . . . . . . . . . . . . . . . . . . . . . . . . . 384 9.5.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384 9.5.2 Time-Shifted Sequences . . . . . . . . . . . . . . . . . . . . . . . . 385 9.5.3 Time Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385 9.5.4 Multiplication by an Exponential . . . . . . . . . . . . . . . . 385 9.5.5 Multiplication by n . . . . . . . . . . . . . . . . . . . . . . . . . . . 386 9.5.6 Difference and Accumulation . . . . . . . . . . . . . . . . . . . 386 9.5.7 Convolution Theorem and the z-Transform . . . . . . . . 386 9.5.8 Correlation Theorem and the z-Transform . . . . . . . . . 387 9.5.9 Initial Value Theorem in the Discrete Domain . . . . . . 388 9.5.10 Final Value Theorem in the Discrete Domain . . . . . . 388 9.6 Tables of z-Transform Properties and Pairs. . . . . . . . . . . . . . . 389 9.7 Inverse z-Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390 9.7.1 Inversion Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390 9.7.2 Use of Transform Tables (Partial Fraction Expansion Method) . . . . . . . . . . . . . . . . . . . . . . . . . . . 391 9.7.3 Inverse z-Transforms by Power Series Expansion. . . . 394 9.8 The Unilateral or the One-Sided z-Transform . . . . . . . . . . . . . 395 9.8.1 Time-Shifting Property . . . . . . . . . . . . . . . . . . . . . . . . 395 9.9 Discrete-Data Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 9.9.1 Discrete-Time Transfer Functions. . . . . . . . . . . . . . . . 400 9.9.2 Schur–Cohn Stability Test. . . . . . . . . . . . . . . . . . . . . . 401 9.9.3 Bilinear Transformations. . . . . . . . . . . . . . . . . . . . . . . 401 9.10 Designs by the Time and Frequency Domain Criteria. . . . . . . 403 9.10.1 Impulse Invariance Method by Using the Time Domain Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405 9.10.2 Bilinear Transformation Method by Using the Frequency Domain Criterion . . . . . . . . . . . . . . . . . . . . 407 9.11 Finite Impulse Response (FIR) Filter Design . . . . . . . . . . . . . 410 9.11.1 Low-Pass FIR Filter Design . . . . . . . . . . . . . . . . . . . . 411 9.11.2 High-Pass, Band-Pass, and Band-Elimination FIR Filter Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413 9.11.3 Windows in Fourier Design. . . . . . . . . . . . . . . . . . . . . . 1416 .... 9.12 Digital Filter Realizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 9.12.1 Cascade Form of Realization . . . . . . . . . . . . . . . . . . . 422 9.12.2 Parallel Form of Realization . . . . . . . . . . . . . . . . . . . . 422 9.12.3 All-Pass Filter Realization. . . . . . . . . . . . . . . . . . . . . . 423 9.12.4 Digital Filter Transposed Structures . . . . . . . . . . . . . . 423 9.12.5 FIR Filter Realizations . . . . . . . . . . . . . . . . . . . . . . . . 423 9.13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425 10

Analog Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

429

10.1 10.2

429 431 432 432

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Limiters and Mixers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2.1 Mixers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3 Linear Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xxiv

Contents

10.4 10.5

10.6

10.7

10.8

10.9 10.10 10.11

10.12

10.13

10.14

10.15

10.16

10.17

10.3.1 Double-Sideband (DSB) Modulation . . . . . . . . . . . 10.3.2 Demodulation of DSB Signals. . . . . . . . . . . . . . . . . Frequency Multipliers and Dividers. . . . . . . . . . . . . . . . . . . . Amplitude Modulation (AM). . . . . . . . . . . . . . . . . . . . . . . . . 10.5.1 Percentage Modulation . . . . . . . . . . . . . . . . . . . . . . 10.5.2 Bandwidth Requirements . . . . . . . . . . . . . . . . . . . . 10.5.3 Power and Efficiency of an Amplitude Modulated Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.5.4 Average Power Contained in an AM Signal . . . . . . Generation of AM Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6.1 Square-Law Modulators . . . . . . . . . . . . . . . . . . . . . 10.6.2 Switching Modulators . . . . . . . . . . . . . . . . . . . . . . . 10.6.3 Balanced Modulators. . . . . . . . . . . . . . . . . . . . . . . . Demodulation of AM Signals . . . . . . . . . . . . . . . . . . . . . . . . 10.7.1 Rectifier Detector. . . . . . . . . . . . . . . . . . . . . . . . . . . 10.7.2 Coherent or a Synchronous Detector . . . . . . . . . . . 10.7.3 Square-Law Detector. . . . . . . . . . . . . . . . . . . . . . . . 10.7.4 Envelope Detector . . . . . . . . . . . . . . . . . . . . . . . . . . Asymmetric Sideband Signals . . . . . . . . . . . . . . . . . . . . . . . . 10.8.1 Single-Sideband Signals . . . . . . . . . . . . . . . . . . . . . . 10.8.2 Vestigial Sideband Modulated Signals . . . . . . . . . . 10.8.3 Demodulation of SSB and VSB Signals . . . . . . . . . 10.8.4 Non-coherent Demodulation of SSB. . . . . . . . . . . . 10.8.5 Phase-Shift Modulators and Demodulators . . . . . . Frequency Translation and Mixing . . . . . . . . . . . . . . . . . . . . Superheterodyne AM Receiver. . . . . . . . . . . . . . . . . . . . . . . . Angle Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.11.1 Narrowband (NB) Angle Modulation. . . . . . . . . . . 10.11.2 Generation of Angle Modulated Signals . . . . . . . . . Spectrum of an Angle Modulated Signal . . . . . . . . . . . . . . . . 10.12.1 Properties of Bessel Functions . . . . . . . . . . . . . . . . . 10.12.2 Power Content in an Angle Modulated Signal . . . . Demodulation of Angle Modulated Signals. . . . . . . . . . . . . . 10.13.1 Frequency Discriminators . . . . . . . . . . . . . . . . . . . . 10.13.2 Delay Lines as Differentiators . . . . . . . . . . . . . . . . . FM Receivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.14.1 Distortions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.14.2 Pre-emphasis and De-emphasis . . . . . . . . . . . . . . . . 10.14.3 Distortions Caused by Multipath Effect . . . . . . . . . Frequency-Division Multiplexing (FDM) . . . . . . . . . . . . . . . 10.15.1 Quadrature Amplitude Modulation (QAM) or Quadrature Multiplexing (QM). . . . . . . . . . . . . . 10.15.2 FM Stereo Multiplexing and the FM Radio . . . . . . Pulse Modulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.16.1 Pulse Amplitude Modulation (PAM) . . . . . . . . . . . 10.16.2 Problems with Pulse Modulations . . . . . . . . . . . . . . 10.16.3 Time-Division Multiplexing (TDM) . . . . . . . . . . . . Pulse Code Modulation (PCM) . . . . . . . . . . . . . . . . . . . . . . . 10.17.1 Quantization Process . . . . . . . . . . . . . . . . . . . . . . . . 10.17.2 More on Coding. . . . . . . . . . . . . . . . . . . . . . . . . . . .

432 433 435 437 438 438 439 440 441 441 441 442 443 443 443 444 444 446 446 447 448 449 449 450 453 455 458 459 460 461 463 465 465 467 468 468 469 470 471 472 473 474 475 475 477 478 478 480

Contents

xxv

10.17.3

Tradeoffs Between Channel Bandwidth and Signal-to-Quantization Noise Ratio . . . . . . . . . . . . 10.17.4 Digital Carrier Modulation . . . . . . . . . . . . . . . . . . . 10.18 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

481 482 484 484

Appendix A: Matrix Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 A.1 Matrix Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 A.2 Elements of Matrix Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . 490 A.2.1 Vector Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491 A.3 Solutions of Matrix Equations . . . . . . . . . . . . . . . . . . . . . . . . . 492 A.3.1 Determinants. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492 A.3.2 Cramer’s Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493 A.3.3 Rank of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494 A.4 Inverses of Matrices and Their Use in Determining the Solutions of a Set of Equations . . . . . . . . . . . . . . . . . . . . . . 495 A.5 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . 496 A.6 Singular Value Decomposition (SVD) . . . . . . . . . . . . . . . . . . . 500 A.7 Generalized Inverses of Matrices . . . . . . . . . . . . . . . . . . . . . . . 501 A.8 Over- and Underdetermined System of Equations . . . . . . . . . . 502 A.8.1 Least-Squares Solutions of Overdetermined System of Equations (m > n) . . . . . . . . . . . . . . . . . . . . 502 A.8.2 Least-Squares Solution of Underdetermined System of Equations (m n) . . . . . . . . . . . . . . . . . . . . 504 A.9 Numerical-Based Interpolations: Polynomial and Lagrange Interpolations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505 A.9.1 Polynomial Approximations . . . . . . . . . . . . . . . . . . . . . . . 505 ... A.9.2 Lagrange Interpolation Formula . . . . . . . . . . . . . . . . . . . 506 ... Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506 Appendix B: MATLAB1 for Digital Signal Processing . . . . . . . . . . . . . . . B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.2 Signal Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.3 Signal Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.4 Fast Fourier Transforms (FFTs) . . . . . . . . . . . . . . . . . . . . . . . B.5 Convolution of Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.6 Differentiation Using Numerical Methods . . . . . . . . . . . . . . . B.7 Fourier Series Computation . . . . . . . . . . . . . . . . . . . . . . . . . . B.8 Roots of Polynomials, Partial Fraction Expansions, PoleZero Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.8.1 Partial Fraction Expansions . . . . . . . . . . . . . . . . . . . . B.9 Bode Plots, Impulse and Step Responses . . . . . . . . . . . . . . . . B.9.1 Bode Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.9.2 Impulse and Step Responses . . . . . . . . . . . . . . . . . . . . B.10 Frequency Responses of Digital Filter Transfer Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.11 Introduction to the Construction of Simple MATLAB Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.12 Additional MATLAB Code. . . . . . . . . . . . . . . . . . . . . . . . . . .

509 509 509 511 511 513 515 515 517 518 518 518 518 520 520 521

xxvi

Contents

Appendix C: Mathematical Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.1 Trigonometric Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.2 Logarithms, Exponents and Complex Numbers . . . . . . . . . . . . C.3 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.4 Indefinite Integrals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.5 Definite Integrals and Useful Identities. . . . . . . . . . . . . . . . . . . C.6 Summation Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.7 Series Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.8 Special Constants and Factorials . . . . . . . . . . . . . . . . . . . . . . .

523 523 523 524 524 525 525 526 526

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

527

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

531

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

535

List of Tables

Table 1.4.1 Table 1.9.1 Table 1.9.2 Table 2.4.1 Table 2.6.1 Table 3.4.1 Table 3.10.1 Table 3.10.2 Table 4.12.1 Table 4.12.2 Table 5.6.1 Table 5.6.2 Table 5.8.1 Table 5.9.1 Table 5.10.1 Table 7.1.1 Table 7.4.1

Table 7.5.1 Table 7.7.1 Table 8.1.1 Table 8.2.1 Table 8.2.2 Table 8.3.1 Table 8.7.1 Table 8.7.2 Table 8.10.1 Table 9.1.1

Properties of the impulse function. . . . . . . . . . . . . . . . . . . Sound Power (loudness) Comparison . . . . . . . . . . . . . . . . Power ratios and their corresponding values in dB . . . . . . Properties of aperiodic convolution . . . . . . . . . . . . . . . . . Example 2.6.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Summary of the three Fourier series representations . . . . Symmetries of real periodic functions and their Fourier-series coefficients . . . . . . . . . . . . . . . . . . . . . . . . . Periodic functions and their Trigonometric Fourier Series . . Fourier transform properties. . . . . . . . . . . . . . . . . . . . . . . Fourier Transform Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . One-sided Laplace transform properties . . . . . . . . . . . . . . One-sided Laplace tranform pairs . . . . . . . . . . . . . . . . . . . Typical rational replace transforms and their inverses . . . One sided Laplace transforms and Fourier transforms. . . Hilbert transform pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . Formula for computing sensitivities . . . . . . . . . . . . . . . . . Normalized frequencies, ! ¼ !0 . Time delay and a loss table giving the normalized frequency ! at which the zero frequency delay and loss values deviate by specified amounts for Bessel filter functions . . . . . . . . . . . . . . . . . . Frequency transformations . . . . . . . . . . . . . . . . . . . . . . . . Guidelines for passive components . . . . . . . . . . . . . . . . . . Fourier representations of discrete-time and continuous-time signals . . . . . . . . . . . . . . . . . . . . . . . . . . . Common interpolation functions . . . . . . . . . . . . . . . . . . . Spectral occupancy of Xðjð! n!s ÞÞ; ! ¼ 2f; n ¼ 0; 1; 2; 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Properties of discrete convolution . . . . . . . . . . . . . . . . . . . Discrete-time Fourier transform (DTFT) properties . . . . Discrete-time Fourier transform (DTFT) pairs. . . . . . . . . Discrete Fourier transform (DFT) properties . . . . . . . . . . Discrete-time and continuous-time signals and their transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18 33 33 53 62 84 105 105 148 149 168 175 182 185 190 244

265 269 283 312 319 325 329 347 348 361 367

xxvii

xxviii

Table 9.2.1 Table 9.6.1 Table 9.6.2 Table 9.11.1 Table 9.11.2 Table 10.9.1 Table 10.12.1 Table 10.17.1 Table 10.17.2 Table 10.17.3 Table B.7.1

List of Tables

Properties of the function WN ¼ ejð2=NÞ . . . . . . . . . . . . . Z-transform properties . . . . . . . . . . . . . . . . . . . . . . . . . . . Z-transform pairs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ideal low-pass filter FIR coefficients with c ¼ =4. . . . . FIR Filter Coefficients for the Four Basic Filters. . . . . . . Inputs and outputs of the system in Fig. 10.9.1 . . . . . . . . Bessel function values . . . . . . . . . . . . . . . . . . . . . . . . . . . . Quantization values and codes corresponding to Fig. 10.17.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Binary representation of quantized values . . . . . . . . . . . . Normal binary and Gray code representations for N¼8.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Amplitudes and phase angles of the harmonic Fourier series coefficients (Example B.7.1). . . . . . . . . . . . . . . . . . .

369 389 390 412 415 453 463 479 480 481 516

Chapter 1

Basic Concepts in Signals

1.1 Introduction to the Book and Signals the research and developments of many signal proThe primary goal of this book is to introduce the reader on the basic principles of signals and to provide tools thereby to deal with the analysis of analog and digital signals, either obtained naturally or by sampling analog signals, study the concepts of various transforming techniques, filtering analog and digital signals, and finally introduce the concepts of communicating analog signals using simple modulation techniques. The basic material in this book can be found in several books. See references at the end of the book. A signal is a pattern of some kind used to convey a message. Examples include smoke signals, a set of flags, traffic lights, speech, image, seismic signals, and many others. Smoke signals were used for conveying information that goes back before recorded history. Greeks and Romans used light beacons in the pre-Christian era. England employed a long chain of beacons to warn that Spanish Armada is approaching in the late sixteenth century. Around this time, the word signal came into use perceptible by sight, hearing, etc., conveying information. The present day signaling started with the invention of the Morse code in 1838. Since then, a variety of signals have been studied. These include the following inventions: Facsimile by Alexander Bain in 1843; telephone by Alexander Bell in 1876; wireless telegraph system by Gugliemo Marconi in 1897; transmission of speech signals via radio by Reginald Fessenden in 1905, invention and demonstration of television, the birth of television by Vladimir Zworykin in the 1920 s, and many others. In addition, the development of radar and television systems during World War II, proposition of satellite communication systems, demonstration of a laser in 1955, and

cessing techniques and their use in communication systems. Since the early stages of communications, research has exploded into several areas connected directly, or indirectly, to signal analysis and communications. Signal analysis has taken a significant role in medicine, for example, monitoring the heart beat, blood pressure and temperature of a patient, and vital signs of patients. Others include the study of weather phenomenon, the geological formations below the surface and deep in the ground and under the ocean floors for oil and gas exploration, mapping the underground surface using seismometers, and others. Researchers have concluded that computers are powerful and necessary that they need to be an integral part of any communication system, thus generating significant research in digital signal processing, development of Internet, research on HDTV, mobile and cellular telephone systems, and others. Defense industry has been one of the major organizations in advancing research in signal processing, coding, and transmission of data. Several research areas have surfaced in signals that include processing of speech, image, radar, seismic, medical, and other signals.

1.1.1 Different Ways of Looking at a Signal Consider a signal xðtÞ, a function representing a physical quantity, such as voltage, current, pressure, or any other variable with respect to a second variable t, such as time. The terms of interest are the time t and the signal xðtÞ. One of the main topics of

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_1, Ó Springer ScienceþBusiness Media, LLC 2010

1

2

1 Basic Concepts in Signals

this book is the analysis of signals. Webster’s dictionary defines the analysis as 1. Separation of a thing into the parts or elements of which it is composed. 2. An examination of a thing to determine its parts or elements. 3. A statement showing the results of such an examination. There are other definitions. In the following the three parts are considered using simple examples. Consider the sinusoidal function and its expansion using Euler’s formula: xðtÞ ¼ A0 cosðo0 t þ y0 Þ A0 jy0 jo0 t A0 jy0 jo0 t e e þ e e ¼ 2 2

(1:1:1)

¼ ReðA0 ejo0 t ejy0 Þ: In (1.1.1) A0 is assumed to be positive and real and A0 ejy0 is a complex number carrying the amplitude and phase angle of the sinusoidal function and is by definition the phasor representation of the given sinusoidal function. Some authors refer to this as phasor transform of the sinusoidal signal, as it transforms the time domain sinusoidal function to the complex frequency domain. A brief discussion on complex numbers is included later in Section 1.6. This signal can be described in another domain, i.e., such as the frequency domain. The amplitude is ðA0 =2Þ and the phase angles of y0 corresponding to the frequencies f0 ¼ o0 =2p Hz. In reality, only positive frequencies are available, but Euler’s formula in (1.1.1) dictates that both the positive and negative frequencies need to be identified as illustrated in Fig. 1.1.1a. This description is the twosided amplitude and phase line spectra of xðtÞ. Amplitudes are always positive and are located at f ¼ o0 =2p ¼ f0 Hz, symmetrically located around the zero frequency, i.e., with even symmetry. The phase spectrum consists of two angles y ¼ y0 corresponding to the positive and negative frequencies, respectively, with odd symmetry. Since ðtÞ is real, we can pictorially describe it by one- or two-sided amplitude and phase line spectra as shown in Fig. 1.1.1a,b,c,d. The following example illustrates the three steps.

Fig. 1.1.1 xðtÞ ¼ A0 cosðo0 t þ y0 Þ. (a) Two-sided amplitude spectrum, (b) two-sided phase spectrum, (c) one-sided amplitude spectrum, and (d) one-sided phase spectrum

Example 1.1.1 Express the following function in terms of a sum of cosine functions: xðtÞ ¼ A0 þ A1 cosðo1 t þ y1 Þ

(1:1:2) A2 cosðo2 tÞ A3 sinðo3 t þ y3 Þ; Ai > 0:

Solution: Using trigonometric relations to express each term in (1.1.2) in the form of Ai cosðoi t þ yi Þ results in xðtÞ ¼ A0 cosðð0Þt 180 Þ þ A1 cosðo1 t þ y1 Þ þ A2 cosðo2 t 180 Þ (1:1:3) þ A3 cosðo3 t þ y3 þ 90 Þ:

In the first and the third terms either 1808 or 1808 could be used, as the end result is the same. The two-sided line spectra of the function in (1.1.2) are shown in Fig. 1.1.2. How would one get the functions of the type shown in (1.1.3) for an arbitrary

1.1 Introduction to the Book and Signals

3

Fig. 1.1.2 (a) Two-sided amplitude spectra and (b) twosided phase spectra

Fig. 1.1.3 Speech . . .sho in . . .show ._male 2000 Samples @ 8000 samples per second. Printed with the permission from Hassan et al. (1994)

function? The sign and cosine functions are the building blocks of the Fourier series in Chapter 3 and later the Fourier transforms in Chapter 4. The function xðtÞ has four frequencies:

Example(s) 1.1.2 In this example several specific examples of interest are considered. In the first one, part of the time signal illustrating a male voice of speech in the sentence ‘‘. . .Show the rich lady out’’ is shown in Fig. 1.1.3. The speech signal is sampled at 8000 samples per second. There are three portions of the speech ‘‘/. . ./, /sh/, /o/’’ shown in the figure. The first part of the signal does not have any speech in it and the small amplitudes of the signal represent the noise in the tape recorder and/or in the room where the speech was recorded. It represents a random signal and can be described only by statistical means. Random signal analysis is not discussed in any detail in this book, as it requires knowledge of probability theory. The second part represents the phoneme ‘‘sh’’ that does not show any observable pattern. It is a time signal for a very short time and has finite energy. Power and energy signals are studied in Section 1.5. Third part of the figure represents the vowel ‘‘o,’’ showing a structure of (almost) periodic pulses for a short time. In this book, aperiodic or non-periodic signals with finite energy and periodic signals with finite average power will be studied. One goal is to come up with a model for each portion of a signal that can be transmitted and reconstructed at the receiver. Next three examples are from food industry. Small businesses are sprouting that use signal processing. For example, when we go to a grocery store we may like to buy a watermelon. It may not always be possible to judge the ripeness of the watermelon

f1 ð¼0Þ; f2 ; f3 ; f4 with amplitudes A0 ; A1 ; A2 ; A3 and phases 180o ; y1 ; 180o ; y3 þ 90o : Figure 1.1.2 illustrates pictorially the discrete locations of the frequencies, their amplitudes, and phases. The signal in (1.1.2) can be described by using the time domain function or in terms of frequencies. In the figures, o0 ð¼ 2pf Þ0 s in radians per second could have been used rather than f 0 s in Hz.&

1.1.2 Continuous-Time and Discrete-Time Signals A signal xðtÞ is a continuous-time signal if t is a continuous variable. It can take on any value in the continuous interval ða; bÞ. Continuous-time signal is an analog signal. If a function y½n is defined at discrete times, then it is a discrete-time signal, where n takes integer values. In Chapter 8 discrete-time signals will be studied by sampling the continuous signals at equal sampling intervals of ts seconds and write xðnts Þ; where n an integer. This is expressed by x½n xðnts Þ:

(1:1:4)

4

by outward characteristics such as external color, stem conditions, or just the way it looks. A sure way of looking at the quality is to cut the watermelon open and taste it before we buy it. This implies we break it first, which is destructive testing. Instead, we can use our grandmother’s procedure in selecting a watermelon. She uses her knuckles to send a signal into the watermelon. From the audio response of the watermelon she decides whether it is good or not based on her prior experience. We can simulate this by putting the watermelon on a stand, use a small hammer like device, give a slight tap on the watermelon, and record the response. A simplistic model of this is shown in Fig. 1.1.4. The responses can be categorized by studying the outputs of tasty watermelons. For an interesting research work on this topic, see Stone et al (1996). Image processing can be used to check for burned crusts, topping amount distribution, such as the location of pepperoni pizza slices, and others. For an interesting article on this subject, see Wagner (1983), which has several applications in the food industry. The next two examples are from the surface seismic signal analysis. In the first one, we use a source in the form of dynamite sticks representing a source, dig a small hole, and blow them in the hole. The ground responds to this input and the response is recorded using a seismometer and a tape recorder. The analysis of the recorded waveform can provide information about the underground cavities and pockets of oil and other important measures.

1 Basic Concepts in Signals

Geologists drill holes into the ground and a small slice of the core sample is used to measure the oil content by looking at the percentage of the area with dark spots on the slice, which is image processing. Another example of interest is measuring the distance from a ground station to an airplane. Send a signal with square wave pulses toward the airplane and when the signal hits it, a return signal is received at the ground station. A simple model is shown in Fig. 1.1.5. If we can measure the time between the time the signal left from the ground station and the time it returned, identified as T in the figure, we can determine the distance between the ground station and the target by the formula x ¼ 3ð108 Þ ðm=sÞ Tðsignal round trip time in secondsÞ=2:

(1:1:5)

The constant c ¼ 3ð108 Þ m/s is the speed of light. Radar and sonar signal processing are two important areas of signal processing applications. An exciting field of study is the biomedical area. We are well aware of a healthy heart that beats periodically, which can be seen from a record of an electrocardiogram (ECG). The ECG represents changes in the voltage potential due

(a)

(b)

Fig. 1.1.4 Watermelon responses to a tap

Fig. 1.1.5 (a) Radar range measurement and (b) transmitted and received filtered signals

1.1 Introduction to the Book and Signals

5

to electrochemical processes in the heart cells. Inferences can be made about the health of the heart under observation from the ECG. Another important example is the electroencephalogram (EEG), which measures the electrical activity in & brain. Signal processing is an important area that interests every engineer. Pattern recognition and classification is almost on top of the list. See, for example, O’Shaughnessy (1987) and Tou and Gonzalez (1974). For example, how do we distinguish two phonemes, one is a vowel and the other one is a consonant. A rough measure of frequency of a waveform with zero average value is the number of zero crossings per unit time. We will study in much more detail the frequency content in a signal later in terms of Fourier transforms in Chapter 4. Vowel sounds have lower frequency content than the consonants. A simple procedure to measure frequency in a speech segment is by computing the number of zero crossings in that segment. To differentiate a vowel from a consonant, set a threshold level for the frequency content for vowels and consonants that differentiate between vowels and consonants. If the frequency content is higher than this threshold, then the phoneme is a consonant. Otherwise, it is a vowel. If we like to distinguish one vowel from another we may need more than one measure. Vocal tract can be modeled as an acoustic tube with resonances, called formants. Two formant frequencies can be used to distinguish two vowels, say /u/ and / a/. See Problem 1.1.1. Two formant frequencies may not be enough to distinguish all the phonemes, especially if the signal is corrupted by noise. Consider a simple pattern classification problem with M prototype patterns z1 ; z2 ; . . . ; zM , where zi is a vector representing an ith pattern. For simplicity we assume that each pattern can be represented by a pair of numbers, say zi ¼ ðzi1 ; zi2 Þ; i = 1,2 . . . M and classify an arbitrary pattern x ¼ ðx1 ; x2 Þ to represent one of the prototype patterns. The Euclidean distance between a pattern x and the ith prototype pattern is defined by D i ¼ kx z i k ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ðx1 zi1 Þ2 þðx2 zi2 Þ2 :

(1:1:6)

A simple classifier is a minimum distance classifier that computes the distance from a pattern x of the

signal to be classified to the prototype of each class and assigns the unknown pattern to the class which it is closest to. That is, if Di 5Dj ; for all i 6¼ j, then we make the decision that x belongs to the ith prototype pattern. Ties are rare and if there are, they are resolved arbitrarily. In the above discussion two measures are assumed for each pattern. More measures give a better separation between classes. There are several issues that would interest a biomedical signal processor. These include removal of any noise present in the signals, such as 60-Hz interference picked up by the instruments, interference of the tools or meters that measure a parameter, and other signals that interfere with the desired signal. Finding the important facets in a signal, such as the frequency content, and many others is of interest. &

1.1.3 Analog Versus Digital Signal Processing Most signals are analog signals. Analog signal processing uses analog circuit elements, such as resistors, capacitors, inductors, and active components, such as operational amplifiers and non-linear devices. Since the inductors are made from magnetic material, they have inherent resistance and capacitance. This brings the quality of the components low. They tend to be bulky and their effectiveness is reduced. To alleviate this problem, active RC networks have been popular. Analog processing is a natural way to solve differential equations that describe physical systems, without having to resort to approximate solutions. Solutions are obtained in real time. In Chapter 10 we will see an example of analog encryption of a signal, wherein the analog speech is scrambled by the use of modulation techniques. Digital signal processing makes use of a special purpose computer, which has three basic elements, namely adders, multipliers, and memory for storage. Digital signal processing consists of numerical computations and there is no guarantee that the processing can be done in real time. To encrypt a set of numbers, these need to be converted into another set of numbers in the digital encryption scheme, for example. The complete encrypted signal is needed before it can be decrypted. In addition, if the input and the output signals are analog, then an

6

1 Basic Concepts in Signals

fx½ng ¼ f. . . ; 1; a; a2 ; . . . ; an ; . . .g or fx½ng #

In (1.1.8) reference points are not identified. In (1.1.9), the arrow below 1 is the 0 index term. The first term is assumed to be zero index term if there is no arrow and all the values of the sequence are zero for n50. We will come back to this in Chapter 8. A signal xðtÞ is a real signal if its value at some t is a real number. A complex signal xðtÞ consists of two real signals, x1 ðtÞ and x2 ðtÞ such that x(t) = x_1 pﬃﬃﬃﬃﬃﬃﬃﬃ (t) + jx2 ðtÞ; where j ¼ 1: The symbol j (or i) is used to represent the imaginary part. Interesting functions: a. P Function: The P function is centered at t0 with a width of t s shown in Fig. 1.1.6. It is not defined at t ¼ t0 t=2 and is symbolically expressed by

Π

2

n

fx½ng ¼ f. . . ; 0; 0; 0; 1; a; a ; . . . ; a ; . . .g;

(1:1:8)

t − t0 τ

Fig. 1.1.6 A P function

Yht t0 i 1;

jt t0 j5t=2 (1:1:10) t 0; otherwise ht t i xðtÞ; jt t j5t=2 0 0 xðtÞP ¼ t 0; otherwise

1.1.4 Examples of Simple Functions To begin the study we need to look into the concept of expressing a signal in terms of functions that can be generated in a laboratory. One such function is the sinusoidal function xðtÞ ¼ A0 cosðo0 t þ y0 Þ seen earlier in (1.1.1), where A0 ; o0 ; and y0 are some constants. A digital signal can be defined as a sequence in the forms n a ;n 0 x½n ¼ ; (1:1:7) 0; n50

(1:1:9)

¼ f1; a; a2 ; . . . ; an ; . . .g #

analog-to-digital converter (A/D), a digital processor, and a digital-to-analog converter (D/A) are needed to implement analog processing by digital means. Special purpose processor with A/D and D/A converters can be expensive. Digital approach has distinct advantages over analog approaches. Digital processor can be used to implement different versions of a system by changing the software on the processor. It has flexibility and repeatability. In the analog case, the system has to be redesigned every time the specifications are changed. Design components may not be available and may have to live with the component values within some tolerance. Components suffer from parameter variations due to room temperature, humidity, supply voltages, and many other aspects, such as aging, component failure. In a particular situation, many of the above problems need to be investigated before a complete decision can be made. Future appears to be more and more digital. Many of the digital signal processing filter designs are based on using analog filter designs. Learning both analog and digital signal processing is desired. Deterministic and random signals: Deterministic signals are specified for any given time. They can be modeled by a known function of time. There is no uncertainty with respect to any value at any time. For example, xðtÞ ¼ sinðtÞ is a deterministic signal. A random signal yðtÞ ¼ xðtÞ þ nðtÞ can take random values at any given time, as there is uncertainty about the noise signal nðtÞ. We can only describe such signals through statistical means, and the discussion on this topic will be minimal.

This P function is a deterministic signal. It is even, as P½t ¼ P½t and P½t0 t ¼ P½t t0 . Some use the symbol ‘‘rect’’ and ð1=tÞrectððt t0 Þ=tÞ is a rectangular pulse of width t s centered at t ¼ t0 with a height ð1=tÞ: b. L Function: The triangular function shown in Fig. 1.1.7 is defined by

L

ht t i 0

t

( ¼

0j 5 1 jtt t ; jt t0 j t

0;

otherwise

:

(1:1:11)

1.1 Introduction to the Book and Signals Λ

7

t − t0 τ

x(t ) = X 0 e–t / Tc u (t ), τ > 0

Fig. 1.1.9 Exponential decaying function

Fig. 1.1.7 A L function

Rectangular function defined in (1.1.10) has a width of t seconds, whereas the triangular function defined in (1.1.11) has a width of 2t s. The symbol ‘‘tri’’ is also used for a triangular function and triððt t0 Þ=tÞ describes the function in (1.1.11). c. Unit step function: It is shown in Fig. 1.1.8 and is uð t Þ ¼

1; t40 ðnot defined at t ¼ 0Þ: 0; t50

xðTc Þ ¼ e1 :37 or xðTc Þ ¼ :37X0 : X0

(1:1:14)

xðtÞ decreases to about 37% of its initial value in Tc s and is the time constant. It decreases to 2% in four time constants and Xð4Tc Þ :018X0 : A measure associated with exponential functions is the half life Th defined by

(1:1:12) 1 xðTh Þ ¼ X0 ; eTh =Tc 2 1 ¼ ; Th ¼ Tc loge ð2Þ ﬃ :693Tc : 2

(1:1:15)

Fig. 1.1.8 Unit step function

e. One-sided and two-sided exponentials: These are described by

The unit step function at t ¼ 0 can be defined explicitly as 0 or 1 or ½uð0þ Þ þ uð0 Þ=2=.5

x1 ðtÞ ¼

d. Exponential decaying function: A simple such function is ( X0 et=Tc ; t 0; Tc 40 : (1:1:13) xðtÞ ¼ 0; otherwise See Fig. 1.1.9. It has a special significance, as xðtÞ is the solution of a first-order differential equation. The constants X0 and Tc can take different values and

eat ; t 0 ; a4 0 0; t50 0; t 0 a40; x3 ðtÞ ¼ eajtj ; a40: x2 ðtÞ ¼ eat ; t50 (1:1:16)

x1 ðtÞ is the right-sided exponential, x2 ðtÞ is the leftsided exponential, and x3 ðtÞ is the two-sided exponential. These are sketched in Fig. 1.1.10. Using the unit step function, we have x2 ðtÞ ¼ x3 ðtÞuðtÞ and x1 ðtÞ ¼ x3 ðtÞuðtÞ. x2 (t )

Fig. 1.1.10 Exponential functions (a) x2 (t) = eat u(t),a > 0, (b) x1 ðtÞ ¼ eat uðtÞ; a > 0; and (c) x3 ðtÞ ¼ eajtj ; a > 0

(a)

x1 (t )

(b)

x3 (t )

(c)

8

1 Basic Concepts in Signals

1.2 Useful Signal Operations

1.2.2 Time Scaling

1.2.1 Time Shifting

The compression or expansion of a signal in time is known as time scaling. It is expanded in time if a5 1 and compressed in time if a > 1 in

Consider an arbitrary signal starting at t ¼ 0 shown in Fig. 1.2.1a. It can be shifted to the right as shown in Fig. 1.2.1b. It starts at time t ¼ a > 0, a delayed version of the one in Fig. 1.2.1a. Similarly it can be shifted to the left starting at time a shown in Fig. 1.2.1c. It is an advanced version of the one in Fig. 1.2.1a. We now have three functions: xðtÞ, xðt aÞ; and xðt þ aÞ with a > 0. The delayed and advanced unit step functions are

fðtÞ ¼ xðatÞ; a > 0:

(1:2:3)

Example 1.2.1 Illustrate the rectangular pulse functions P½t; P½2t; and P½t=2. Solution: These are shown in Fig. 1.2.2 and are of widths 1, (1/2), and 2, respectively. The pulse P½2t is a compressed version and the pulse P½t=2 is an & expanded version of the pulse function P½t.

1; t4a ; 0; t5a 1; t4 a uðt þ aÞ ¼ ; ða40Þ: 0; t5 a

uð t aÞ ¼

1.2.3 Time Reversal (1:2:1) If a ¼ 1 in (1.2.3), that is, fðtÞ ¼ xðtÞ, then the signal is time reversed (or folded).

From (1.1.16), the right-sided delayed exponential decaying function is x1 ðt tÞ ¼ eaðttÞ uðt tÞ; a > 0; t > 0:

x(t)

(1:2:2)

x(t + a)

x(t – a)

(a)

(b)

(c)

Fig. 1.2.1 (a) x(t), (b) x(t a), (c) x(t + a), a > 0

Fig. 1.2.2 Pulse functions

Example 1.2.2 Let x1 ðtÞ ¼ eat uðtÞ: Give its timereversed signal. Solution: The time-reversed signal of x1 ðtÞ is & x2 ðtÞ ¼ eat uðtÞ.

1.2.4 Amplitude Shift The amplitude shift of xðtÞ by a constant K is fðtÞ ¼ K þ xðtÞ. Combined operations: Some of the above signal operations can be combined into a general form. The signal yðtÞ ¼ xðat t0 Þ may be described by one of the two ways, namely:

Π[t]

Π[2t]

(a)

(b)

Π[t /2]

(c)

1.2 Useful Signal Operations

9

1. Time shift of t0 followed by time scaling by a. 2. Time scaling by (a) followed by time shift of ðt0 =aÞ.

equations in (1.2.6) and yð0Þ ¼ xð3Þ ¼ 0 and & yðt0 =aÞ ¼ yð3=2Þ ¼ 0.

These can be visualized by the following:

1: xðtÞ

shift scale ! vðtÞ ¼ xðt t0 Þ ! yðtÞ t ! t t0 t ! at (1:2:4)

Continuous-time even and odd functions satisfy

¼ vðatÞ ¼ xðat t0 Þ: scale shift 2: xðtÞ ! yðtÞ ! gðtÞ ¼ xðatÞ t ! at t ! t ðt0 =aÞ (1:2:5) ¼ gðt ðt0 =aÞÞ ¼ xðat t0 Þ:

Notation can be simplified by writing yðtÞ ¼ xðaðt ðt0 =aÞÞÞ. Noting that b ¼ at t0 is a linear equation in terms of two constants a and t0 , it follows: yð0Þ ¼ xðt0 Þ and yðt0 =aÞ ¼ xð0Þ:

1.2.5 Simple Symmetries: Even and Odd Functions

(1:2:6)

These two equations provide checks to verify the end result of the transformation. Following example illustrates some pitfalls in the order of time shifting and time scaling. Example 1.2.3 Derive the expression yðtÞ ¼ xð3t þ 2Þ assuming xðtÞ ¼ P½t=2.

xðtÞ ¼ xðtÞ xe ðtÞ, an even function, x ( t) = xðtÞ x0 ðtÞ, an odd function. (1.2.7) Examples of even and odd functions are shown in Fig. 1.2.3. The function cosðo0 tÞ is an even function and x0 ðtÞ ¼ sinðo0 tÞ is an odd function. An arbitrary real signal, xðtÞ; can be expressed in terms of its even and odd parts by xðtÞ ¼ xe ðtÞ þ x0 ðtÞ; xe ðtÞ ¼ ½ðxðtÞ þ xðtÞÞ=2; x0 ðtÞ

(1:2:8)

¼ ½ðxðtÞ xðtÞÞ=2: xe(t)

x0(t)

(a)

(b)

for

Solution: Using (1.2.4) with a ¼ 3 and t0 ¼ 2, we have tþ2 ; vðtÞ ¼xðt t0 Þ ¼ P 2 3t þ 2 t þ ð2=3Þ yðtÞ ¼vð3tÞ ¼ P ¼P : 2 2=3 Using (1.2.5), we have 3t ; gðtÞ ¼xðatÞ ¼ P 2 t0 2 yðtÞ ¼gðt Þ ¼ gðt þ Þ a 3 3ðt þ ð2=3ÞÞ t þ ð2=3Þ ¼P : ¼P 2 2=3 It is a rectangular pulse of unit amplitude centered at t ¼ ð2=3Þ with width (2/3). We can check the

Fig. 1.2.3 (a) Even function and (b) odd function

1.2.6 Products of Even and Odd Functions Let xe ðtÞ and ye ðtÞ be two even functions and x0 ðtÞ and y0 ðtÞ be two odd functions and arbitrary. Some general comments can be made about their products. xe ðtÞye ðtÞ ¼ xe ðtÞye ðtÞ; even function:

(1:2:9)

xe ðtÞy0 ðtÞ ¼ xe ðtÞy0 ðtÞ; odd function: (1:2:10) x0 ðtÞy0 ðtÞ ¼ ð1Þ2 x0 ðtÞy0 ðtÞ ¼ x0 ðtÞy0 ðtÞ; even function:

(1:2:11)

10

1 Basic Concepts in Signals

Fig. 1.2.4 (a) x1 ðtÞ, (b) xie (t) = x2 (t) even part of x1(t), and (c) x10 ðtÞ = x3 (t) odd part of x1 ðtÞ

x1(t)

x3(t)

x2(t)

(a)

(b)

Note that the functions P[t], L[t] and P[t]cosðo0 tÞ are even functions and P½t sinðo0 tÞ is an odd function. The even and odd parts of the exponential pulse x1 ðtÞ ¼ et uðtÞ are shown in Fig. 1.2.4 and are 1 x1e ðtÞ ¼ ðx1 ðtÞ þ x1 ðtÞÞ; 2 1 x10 ðtÞ ¼ ðx1 ðtÞ x1 ðtÞÞ: 2

sincðplÞ ¼

(c) sinðplÞ sin2 ðplÞ ; sinc2 ðplÞ ¼ pl ðplÞ2

(1:2:15)

Some authors use sincðlÞ for sincðplÞ in (1.2.15). Notation in (1.2.15) is common. Sinc ðplÞ is indeterminate at t ¼ 0: Using the L’Hospital’s rule,

(1:2:12) sinðplÞ ¼ lim l!0 l!0 pl lim

d sinðplÞ dl dðplÞ dl

p cosðplÞ ¼ 1: l!0 p

¼ lim

(1:2:16)

1.2.7 Signum (or sgn) Function

In addition, since sinðplÞ is equal to zero for l ¼ n; n an integer, it follows that

The signum (or sgn) function is an odd function shown in Fig. 1.2.5: sgn(t)

sincðplÞ ¼ 0;

l ¼ n;

n 6¼ 0 and an integer: (1:2:17a)

Interestingly, the function jsincðplÞj is bounded by jð1=plÞ j as jsinðplÞj 1. The side lobes of jsincðplÞj are larger than the side lobes of sinc2 ðplÞ, which follows from the fact that the square of a fraction is less than the fraction we started with. Both the sinc and the sinc2 functions are even. That is,

t

Fig. 1.2.5 Signum function sgnðtÞ

sincð plÞ ¼ sincðplÞ and sinc2 ðplÞ ¼ sinc2 ðplÞ:

sgnðtÞ ¼ uðtÞ uðtÞ ¼ 2uðtÞ 1:

(1:2:13)

(1:2:17b)

sgnðtÞ ¼ lim½eat uðtÞ eat uðtÞ; a > 0:

(1:2:14)

These functions can be evaluated easily by a calculator. For the sketch of a sinc function using MATLAB, see Fig. B.5.2 in Appendix B.

a!0

It is not defined at t ¼ 0 and is chosen as 0.

1.2.8 Sinc and Sinc2 Functions

1.2.9 Sine Integral Function

The sinc and sinc2 functions are defined in terms of an independent variable l by

The sine integral function is an odd function defined by (Spiegel, 1968)

1.3 Derivatives and Integrals of Functions

SiðyÞ ¼

Zy

sinðaÞ da: a

11

(1:2:18a)

0

The values of this function can be computed numerically using the series expression SiðyÞ ¼

y y3 y5 y7 þ þ ... ð1Þ1! ð3Þ3! ð5Þ5! ð7Þ7! (1:2:18b)

@xðt; aÞ xðt þ Dt; aÞ xðt; aÞ ¼ lim ; Dt!0 @t Dt @xðt; aÞ xðt; a þ DaÞ xðt; aÞ ¼ lim : Da!0 @a Da

(1:3:2)

Assuming the second (first) variable is not a function of the first (second) variable, the differential of xðt; aÞ is dx ¼

@x @x dt þ da: @t @a

Some of its important properties are The integral of a function over an interval is the area of the function over that interval.

SiðyÞ ¼ SiðyÞ; Sið0Þ ¼ 0; SiðpÞ ﬃ 2:0123; Sið1Þ ¼ ðp=2Þ :

(1:2:18c)

Si function converges fast and only a few terms in (1.2.18b) are needed for a good approximation.

Example 1.3.1 Compute the value of the integral of xðtÞ shown in Fig. 1.3.1.

1.3 Derivatives and Integrals of Functions It will be assumed that the reader is familiar with some of the basic properties associated with the derivative and integral operations. We should caution that derivatives of discontinuous functions do not exist in the conventional sense. To handle such cases, generalized functions are defined in the next section. The three well-known formulas to approximate a derivative of a function, referred to as forward difference, central difference, and backward difference, are dxðtÞ xðt þ hÞ xðhÞ : ; dt h xðt þ hÞ xðt hÞ ; x0 ðtÞ : 2h xðtÞ xðt hÞ : x0 ðtÞ : h

x0 ðtÞ ¼

(1:3:1)

MATLAB evaluations of the derivatives are given in Appendix B. If we have a function of two variables, then we have the possibility of taking the derivatives one or the other, leading to partial derivatives. Let xðt; aÞ be a function of two variables. The two partial derivatives of xðt; aÞ with respect to t, keeping a constant, and with respect to a, keeping t constant are, respectively, given by

Fig. 1.3.1 Computation of Integral of x(t) using areas

Solution: Divide the area into three parts as identified in the figure. The three parts are in the intervals ð0; aÞ, ða; bÞ, and ðb; cÞ, respectively. The areas of the two triangles are identified by A1 and A2 and the area of the rectangle by A3. They can be individually computed and then add the three areas to get the total area. That is, 1 1 A1 ¼ aB; A2 ¼ ðb aÞC; A3 ¼ ðc bÞB; 2 2 Zc xðtÞdt: A ¼ A1 þ A2 þ A3 ¼

&

0

If the function is arbitrary and cannot be divided into simple functions like in the above example, we can approximate the integral by dividing the area into small rectangular strips and compute the area by adding the areas in each strip.

12

1 Basic Concepts in Signals

If the time interval Dt ¼ ðb aÞ=N is sufficiently small, then the difference between the two formulas in (1.3.3a) and (1.3.3b) would be small and (1.3.3a) & is adequate.

x(t)

1.3.1 Integrals of Functions with Symmetries Fig. 1.3.2 xðtÞ and its approximation using its samples

Example 1.3.2 Consider the function xðtÞ shown in Fig. 1.3.2. Find the integral of this function using the above approximation for a5t5b. Assume the values of the function are known as xðaÞ; xða þ DtÞ; xða þ 2DtÞ; :::; and x((a þ N - 1)D t) Solution: Assuming Dt is small enough that we can approximate the area in terms of rectangular strips using the rectangular integration formula and " # Zb N1 X xðtÞdt xða þ nDtÞ Dt; Dt ¼ ðb aÞ=N: a

n¼0

(1:3:3a) Note that xðaÞDt gives the approximate area in the first strip. If the width of the strips gets smaller and the number of strips increases correspondingly, then the approximation gets better. In the limit, i.e., when Dt ! 0, approximation approaches the value of the integral. In computing the area of the kth rectangular strip, xða þ ðk 1ÞDtÞ was used to approximate the height of the pulse. Some other value of the function in the interval, such as the value of the function in the middle of the strip, could be used. MATLAB evaluation of integrals is discussed in Appendix B. In Chapter 8 appropriate values for Dt will be considered in terms of the frequency content in the signal. Instead of rectangular integration formula, there are other formulas that are useful. One could assume that each strip is a trapezoid and using the trapezoidal integration formula the integral is approximated by Zb

The integrals of functions with even and odd symmetries around a symmetric interval are Za Za Za xe ðtÞdt ¼ 2 xe ðtÞdt and x0 ðtÞdt ¼ 0; a

(1:3:4) where a is an arbitrary positive number. Example 1.3.3 Evaluate the integrals of the functions given below: x1 ðtÞ ¼ P½t=2a;

a

þ 2xða þ ðN 1ÞDtÞ þ xða þ NDtÞðDt=2Þ: (1:3:3b)

x2 ðtÞ ¼ tx1 ðtÞ:

(1:3:5)

Solution: x1 ðtÞ is a rectangular pulse with an even symmetry and x2 ðtÞ is an odd function with an odd symmetry. The integrals are Za Za x1 ðtÞdt ¼ 2 dt ¼ 2a; A1 ¼

A2 ¼

a Za

0

x2 ðtÞdt ¼ 0:

&

a

1.3.2 Useful Functions from Unit Step Function The ramp and the parabolic functions can be obtained by Zt xr ðtÞ ¼ uðtÞdt ¼ tuðtÞ and 0

xp ðtÞ ¼ xðtÞdt ½xðaÞ þ 2xða þ DtÞ þ 2xða þ 2DtÞ þ

a

0

Zt

xr ðtÞdt ¼ðt2 =2ÞuðtÞ:

(1:3:6)

0

Section 1.4 considers the derivatives of the unit step functions.

1.3 Derivatives and Integrals of Functions

13

Now consider Leibniz’s rule, interchange of derivative and integral, and interchange of integrals without proofs. For a summary, see Peebles (2001).

Z1 Z1

jxðt;aÞjdtda51;

1 1 Z1 Z1

½

1.3.3 Leibniz’s Rule

LetgðtÞ ¼

1

ZbðtÞ zðx; tÞdx:

(1:3:7)

aðtÞ

dgðtÞ daðtÞ dbðtÞ ¼ z½bðtÞ; tÞ z½aðtÞ; t dt dt dt bðtÞ Z @zðx; tÞ dx: þ @t

(1:3:8)

aðtÞ

1.3.4 Interchange of a Derivative and an Integral

1

jxðt;aÞdtjda51;

(1:3:10)

1

2 4

1

Z1

3 xðt; aÞdt5da ¼

1

Z1 1

2 4

Z1

3 xðt; aÞdt5da

1

Z1 Z1

¼

xðt; aÞdtda:

(1:3:11)

1 1

Signals generated in a lab are well behaved and they are valid. In Chapter 3 on Fourier series, integrating a product of a simple function, say hðtÞ, with its nth derivative goes to zero; such a polynomial, and the other one is a sinusoidal function gðtÞ, such as sinðo0 tÞ or cosðo0 tÞ or ejo0 t is applicable. The generalized integration by parts formula comes in handy. R

When the limits in (1.3.7) are constants, say aðtÞ ¼ a and bðtÞ ¼ b, then the derivatives of these limits will be zero and (1.3.7) and (1.3.8) reduce to gðtÞ ¼

1

is true, then Fubini’s theorem (see Korn and Korn (1961) for a proof) states that Z1

aðtÞ and bðtÞ are assumed to be real differentiable functions of a real parameter t, and zðx; tÞ and its derivative dzðx; tÞ=dt are both continuous functions of x and t. The derivative of the integral with respect to t is Leibniz’s rule Spiegel (1968) and is

Z1 Z1 ½ jxðt;aÞjdadt51;

hðnÞ ðtÞgðtÞdt ¼ hðn1Þ ðtÞgðtÞ hðn2Þ ðtÞg0 ðtÞ R þhðn3Þ ðtÞg00 ðtÞ ð1Þn hðtÞgðnÞ ðtÞdt; d k gðtÞ g ðtÞ ¼ ; dtk ðkÞ

Zb zðx; tÞdx;

and

: d k hðtÞ h ðtÞ ¼ dtk (1:3:12) ðkÞ

a

dgðtÞ d ¼ dt dt

Zb a

zðx; tÞdt ¼

Zb

@zðx; tÞ dx: @t

(1:3:9)

Using (1.3.12), the following equalities can be seen: Z

a

The derivative and the integral operations may be interchanged.

1.3.5 Interchange of Integrals

Z

Z Z

If any one of the following conditions,

t cosðtÞdt ¼ cosðtÞ þ t sinðtÞ; t sinðtÞdt ¼ sinðtÞ t cosðtÞ

(1:3:13a)

t2 cosðtÞdt ¼ 2t cosðtÞ þ ðt2 2Þ sinðtÞ; t2 sinðtÞdt ¼ 2t sinðtÞ ðt2 2Þ cosðtÞ: (1:3:13b)

14

1 Basic Concepts in Signals

1.4 Singularity Functions

Zt2

The impulse function, or the Dirac delta function, a singularity function, is defined by

t1

dðtÞ ¼

0;

t 6¼ 0

1;

t¼0

Z1 with

dðtÞdt ¼ 1:

(1:4:1)

1

dðtÞ takes the value of infinity at t ¼ 0 and is zero everywhere else. See Fig. 1.4.1b. Impulse function is a continuous function and the area under this function is equal to one. Note that a line has a zero area. Here, a generalized or a distribution function is defined that is nonzero only at one point and has a unit area. A delayed or an advanced impulse function can be defined by dðt t0 Þ, where t0 is assumed to be positive in the expressions. The ideal impulse function cannot be synthesized. It is useful in the limit. For example, 1 hti P ; e!0 e e

dðtÞ ¼ lim

h i 1 t L dðtÞ ¼ lim : e!0 e e (1:4:2)

Figure 1.4.1a illustrates the progression of rectangular pulses of unit area toward the delta function. As e is reduced, the height increases and, in the limit, the function approaches infinity at t ¼ 0 and the area of the rectangle is 1. There are other functions that approximate the impulse function in the limit. A nice definition is given in terms of an integral of a product of an impulse and a test function fðtÞ by Korn and Korn (1961):

fðtÞdðt t0 Þdt

8 0; t0 5t1 or t2 5t0 > > > < ð1=2Þ fðtþ Þ þ fðt Þ; 0 0 ¼ þ > ð1=2Þfðt Þ; t ¼ t1 0 > 0 > : ð1=2Þfðt0 Þ; t0 ¼ t2

t1 5t0 5t2

:

(1:4:3)

fðtÞ is a testing (or a test) function of t and is assumed to be continuous and bounded in the neighborhood of t ¼ t0 and is zero outside a finite interval. That is fð 1Þ ¼ 0. The integral in (1.4.3) is not an ordinary (Riemann) integral. In this sense, dðtÞ is a generalized function. A simpler form of (1.4.3) is adequate. dðtÞ has the property that Z1

fðtÞdðt t0 Þdt ¼ fðt0 Þ;

(1:4:4a)

1

where fðtÞ is a test function that is continuous at t ¼ t0 : As a special case, consider t0 ¼ 0 in (1.4.4a) and fðtÞ ¼ 1. Equation (1.4.4a) can be written as Z1

dðtÞdt ¼

1

Z0þ

dðtÞdt ¼ 1:

(1:4:4b)

0

That is, the area under the impulse function is 1. The integral in (1.4.4a) sifts the value of fðtÞ at t ¼ t0 and dðtÞ is called a sifting function. In summary, the impulse dðt t0 Þ has unit area (or weight) centered at the point t ¼ t0 and zero everywhere else. Since the dðt t0 Þ exists only at t ¼ t0 , and fðtÞ at t ¼ t0 is fðt0 Þ, we have an important result fðtÞdðt t0 Þ ¼ fðt0 Þdðt t0 Þ:

(1:4:5)

There are many limiting forms of impulse functions. Some of these are given below. Notes: In the limit, the following functions can be used to approximate dðtÞ:

(a)

(b)

Fig. 1.4.1 (a) Progression toward an impulse as e ! 0 and (b) symbol for dðtÞ

2 x1 ðtÞ ¼ ejt=tj ; t!0 t 1 1 x2 ðtÞ ¼ sincðt=tÞ; x3 ðtÞ ¼ sinc2 ðt=tÞ t t dðtÞ ¼ lim xi ðtÞ;

1.4 Singularity Functions 2 1 x4 ðtÞ ¼ epðt=tÞ ; t

15

x5 ðtÞ ¼

pðt2

t : þ t2 Þ

(1:4:6a)

x1 ðtÞ is a two-sided exponential function; x2 ðtÞ and x3 ðtÞ are sinc functions. Sinc function does not go to zero in the limit for all t. See the discussion by Papoulis (1962). x4 ðtÞ is a Gaussian function and x5 ðtÞ is a Lorentzian function. To prove these, approximate the impulse function in the limit, using (1.4.4a) and (1.4.4b). Example 1.4.1 Show that the Lorentzian function x5 ðtÞ approaches an impulse function as t ! 0. Show the result by using the equations: a. (1.4.1) and b. (1.4.4a and b). Solution: a. Clearly as t ! 0, x5 ðtÞ ! 0 for t not equal to zero. As t ! 0; x5 ðtÞ ! 1. From tables, 1 p

Z1 t2

t dt ¼ 1; lim x5 ðtÞ ¼ dðtÞ: t!0 þ t2

(1:4:6b)

1

differentiable everywhere any number of times, and, in addition, the function and its derivative decrease at least as rapidly as (1=tn ) as t ! 1 for all n. The derivative of a good function is another good function and the sums and the products of two good functions are good functions. A sequence of good functions fxn ðtÞg fx1 ðtÞ; x2 ðtÞ; . . . ; xn ðtÞg

(1:4:7)

is called regular, if for any good function fðtÞ, the following limit exists: Z1 fxn ðtÞgfðtÞdt: (1:4:8) Vx ðfÞ ¼ lim n!1

1

Example 1.4.2 Find the limit in (1.4.8) of the following sequence: n 2 2o (1:4:9) fxn ðtÞg ¼ eðt =n Þ : Solution: The limit in (1.4.8) is

b. First by (1.4.4a), with t0 ¼ 0 and fðtÞ ¼ 1 in the neighborhood of t ¼ 0 results in

Z1

Vx ðfÞ ¼

fðtÞdt:

(1:4:10) &

1

Z1 lim

t!0

fðtÞx5 ðt t0 Þdt ¼ lim

Z1

t!0

1

x5 ðtÞdt ¼ 1:

1

(1:4:6c)

Two regular sequences of good functions are considered equivalent if the limit in (1.4.8) is the same 4 4 for the two sequences. For example, eðt =n Þ and 2 2 eðt =n Þ are equivalent only in that sense. The function Vx ðfÞ defines a distribution xðtÞ and the limit of the sequence

Using (1.4.4b) and the integral tables, it follows that xðtÞ lim fxn ðtÞg: Z0þ 0

¼

n!1

0þ 1 t t 1 1 t dt ¼ tan p ðt2 þ t2 Þ p t t 0

1 hp p i ¼ 1: p 2 2

(1:4:11)

An impulse function can be defined in terms of a sequence of functions and write & dðtÞ lim fxn ðtÞg if lim n!1

Z1

n!1

fðtÞfxn ðtÞgdt ¼ fð0Þ

1

(1:4:12)

1.4.1 Unit Impulse as the Limit of a Sequence

¼)

Z1

fðtÞdðtÞdt ¼ fð0Þ:

(1:4:13)

1

Another approach to the above is through a sequence. See the work of Lighthill (1958). Also see Baher, (1990). A good function fðtÞ is

Many sequences can be used to approximate an impulse.

16

1 Basic Concepts in Signals

Notes: A constant can be interpreted as a generalized function defined by the regular sequence fxn ðtÞg so that, for any good function fðtÞ Z1 Z1 lim fðtÞdt: (1:4:14) fxn ðtÞgfðtÞdt ¼ K n!1

1

Solution: Using the integration by parts, we have the result below and (1.4.19) follows: Z1 Z1 1 0 x ðtÞfðtÞdt ¼ xðtÞfðtÞj1 xðtÞf0ðtÞdt: 1

1

(1:4:20) &

1

Using the function in (1.4.9), it follows that 2

2

2

fxn ðtÞg Ket =n ;K ¼ lim fKeðt

=n2 Þ

n!1

g:

(1:4:15)

Notes: Let g1 ðtÞ and g2 ðtÞ be two generalized functions and Z1 Z1 fðtÞg1 ðtÞdt ¼ fðtÞg2 ðtÞdt: (1:4:21) 1

1.4.2 Step Function and the Impulse Function Noting the area under the impulse function is one, it follows that Zt dðaÞda ¼ uðtÞ: (1:4:16) 1

Asymmetrical functions ðdþ ðtÞ and d ðtÞÞ and (uþ ðtÞ and u ðtÞÞ can be defined and are uþ ðtÞ ¼

u ðtÞ ¼

Z1 1 Z1 1

1; dþ ðtÞdt ¼ 0; d ðtÞdt ¼

1; 0;

8 t>0 > < 1; uðtÞ ¼ ð1=2Þ; t ¼ 0 > : 0; t 50

(1:4:17)

1

x ðtÞfðtÞdt ¼

xðtÞf0 ðtÞdt; f0 ðtÞ ¼

uðtÞfðtÞdt ¼

1

(1:4:19)

fðtÞdt:

(1:4:22)

0

1

Z1

f0 ðtÞdt ¼ ½fð1Þ fð0Þ ¼ fð0Þ; (1:4:23)

0

Z1

df dt

Z1

Noting that fð1Þ ¼ 0, it follows that Z1 Z1 0 u ðtÞfðtÞdt ¼ uðtÞf0 ðtÞdt

¼

t! 1

Z1

Solution: First Z1

(1:4:18)

Example 1.4.3 Using the property of the test function, lim fðtÞ ¼ 0, show that 0

Example 1.4.4 Show that the derivative of the unit step function is an impulse function using the equivalence property.

1

These step functions differ in how the value of the function at t ¼ 0 is assigned and are only of theoretical interest. Here the step function is assumed to be u ðtÞ and ignore the subscript: Derivative of the unit step function is an impulse function, which can be shown by using the generalized function concept.

Z1

The two generalized functions are equal, i.e., g1 ðtÞ ¼ g2 ðtÞ only in the sense of (1.4.21). It is called the equivalency property.

1

t>0 ; t 0 t0 t 50

1

1

Z1

0

u ðtÞfðtÞdt¼

dðtÞfðtÞdtand u0 ðtÞ

1

¼

duðtÞ ¼ dðtÞ: dt

(1:4:24)

The derivative of a parabolic function results in a ramp function, the derivative of a ramp function results in a unit step function, and the derivative of a unit step function results in an impulse function. All of these are true only in the generalized sense. What is the derivative of an impulse function? That is, d0 ðtÞ ¼

ddðtÞ : dt

(1:4:25)

1.4 Singularity Functions

17

It is defined by the relation Z1

d0 ðtÞfðtÞdt ¼

1

Z1

P½t=e ¼ uðt þ e=2Þ uðt e=2Þ

dP et e e ¼ dðt þ Þ dðt Þ; d0 ðtÞ 2 2 dt h e e i ¼ lim dðt þ Þ dðt Þ : e!0 2 2

dðtÞf0 ðtÞdt ¼ f0ð0Þ:

1

(1:4:26) Generalizing to higher-order derivatives of the impulse function results in Z1

dðnÞ ðtÞfðtÞdt ¼ ð1Þn fðnÞ ð0Þ:

(1:4:27)

1 hti 2 1 lim 2 P ¼ lim ¼ 1: e!0 e e!0 e e

Example 1.4.5 Evaluate the following integrals: a: A ¼

b: B ¼

(1:4:34)

d2 dðtÞ The impulse function is not square integrable. The square of a distribution is not defined Papoulis, dt2 (1962). Note that dðtÞ is an even function and (1:4:28) d0 ðtÞ is an odd function:

ðt2 þ 2t þ 1Þdð2Þ ðt 1Þdt; dð2Þ ðtÞ ¼

1

Z2

(1:4:33)

The derivative of the impulse function results in two impulses illustrated in Fig. 1.4.2. It is an odd function called a doublet. The square of an impulse function is not defined as

1

Z1

(1:4:32)

½ðt 1Þ2 dðt 1Þ þ 5dðt þ 1Þ þ 6tdðtÞdt:

∞

:5

(1:4:29)

t

0

Solution: These follow Z1 ½ða þ 1Þ2 þ 2ða þ 1Þ þ 1dð2Þ ðaÞda a: A ¼

Fig. 1.4.2 Symbolic representation of d0 ðtÞ, a doublet

–∞

1

¼ ð1Þ2

b: B ¼

Z2

d2 ½ða þ 1Þ2 þ 2ða þ 1Þ þ 1ja¼0 ¼ 4 da2

1.4.3 Functions of Generalized Functions Using the equivalence property of the generalized function, the following is true:

½ðt 1Þ2 dðt 1Þ þ 5dðt þ 1Þ

:5

(1:4:30)

gðtÞ ¼ dðat bÞ ¼

When an impulse is outside the integration limits, then the integral is 0. In addition, it is assumed that the limits do not fall at the exact location of the & impulses.

This can be seen from

þ6tdðtÞdt ¼ 0:

Z1 1

1 fðtÞdðat bÞdt¼ j aj

dðtÞ can be approximated using various functions in the limit. In Fig. 1.4.1 it is expressed in terms of a rectangular pulse function. That is, 1 hti dðtÞ ¼ lim x1 ðtÞ; x1 ðtÞ ¼ P : e!0 e e

(1:4:31)

The generalized derivative of a pulse function is as follows:

dðt b=aÞ ; a 6¼ 0: j aj

Z1

(1:4:35)

fðy=aÞdðy bÞdy

1

1 ¼ fðb=aÞ: j aj

(1:4:36)

From the equivalence property, the equality in (1.4.35) now follows. From (1.4.35), it follows that dðoÞ ¼

1 dðfÞ or dðfÞ ¼ 2pdðoÞ: 2p

(1:4:37)

18

1 Basic Concepts in Signals Table 1.4.1 Properties of the impulse function Z1 dðt t0 ÞfðtÞdt ¼ fðt0 Þ: 1

Z1

dðtÞfðt t0 Þdt ¼ fðt0 Þ:

1

Z1

fðtÞdðtÞdt ¼ fð0Þ:

1

Z1 1

Z1

ddðtÞ dfðtÞ : fðtÞdt ¼ dt dt t¼0

dðnÞ ðtÞfðtÞdt ¼ ð1Þn fðnÞ ð0Þ:

1

xðtÞdðtÞ ¼ xð0ÞdðtÞ: Z1

dðtÞdðt0 tÞdt ¼ dðt0 Þ:

1

dðat bÞ ¼

1 b dðt Þ ; a 6¼ 0: a jaj

dðjoÞ ¼ dðoÞ: dðtÞ ¼ dðtÞ:

1.4.4 Functions of Impulse Functions In (1.4.35), dðxðtÞÞ is considered with xðtÞ being a linear function of time. Now consider other cases, where xðtÞ is assumed to have simple zeros, i.e., no multiple zeros. Example 1.4.6 Evaluate the following integral using (1.4.35) and the following cases for the limits. a. x ¼ 0; y ¼ 4 and b. x ¼ 10; y ¼ 5.

A¼

Zy

ðt 1Þðt þ 5Þdð2t þ 5Þdt:

x

Solution: First, changing the variables, a ¼ 2t þ 5, i.e., t ¼ 12ða 5Þ, dt ¼ 12 da, results in a: t ¼ 0¼)a ¼ 5; t ¼ 4¼)a ¼ 13;

Z13 1 7 1 5 1 ¼)A ¼ a a þ dðaÞ da ¼ 0 2 2 2 2 2 5

ða ¼ 0 is outside the range 55a513Þ: b. Similarly, changing the variables, a ¼ 2t þ 5, i.e., t ¼ 12ða 5Þ, dt ¼ 12 da, results in t ¼ 5¼)a ¼ 5; t ¼ 0¼)a ¼ 5; Z5 1 7 1 5 1 ¼)A ¼ a a þ dðaÞ da 2 2 2 2 2 5 1 7 1 5 1 35 a aþ ¼ ja¼0 ¼ : 2 2 2 2 2 8

&

Example 1.4.7 Using the equivalence property of the impulse functions, show that

1.5 Signal Classification Based on Integrals

dðt2 a2 Þ ¼

19

1 ðdðt þ aÞ þ dðt aÞÞ; a 6¼ 0: j2aj (1:4:38)

Solution: Since t2 a2 ¼ ðt aÞðt þ aÞ ¼ 0 ! t ¼ a 6¼ 0 at t ¼ a, it follows that Z1

Example 1.4.8 Give the expression for dðsinðtÞÞ: Solution: Since sinðtÞjt¼np ¼ 0; d sinðtÞ=dt ¼ cosðtÞ; andj cosðnpÞj ¼ 1, it follows that 1 X dðsinðtÞÞ ¼ dðt npÞ: (1:4:41) & n¼1

Z0 dðt a ÞfðtÞdt ¼ dðt2 a2 ÞfðtÞdt 2

2

1

1.4.5 Functions of Step Functions

1

þ

Z1

dðt2 a2 ÞfðtÞdt: (1:4:39)

0

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ With t ¼ ðy þ a2 Þ, Z0

Example 1.4.9 Given xðtÞ ¼ t2 1, sketch the function yðtÞ ¼ uðxðtÞÞ ¼ uðt2 1Þ, where uðtÞ is the unit step function. Solution: Since xðtÞ 0 for 1 t 1, it follows that

dðt2 a2 ÞfðtÞdt

xðtÞ 0 for 1 t 1 ! yðtÞ ¼ 0; 15t51;

1

Za2

¼

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ dðyÞf y þ a2

1

1 : ¼ fðaÞ j2aj

Fig. 1.4.3 shows the sketches for xðtÞ and yðtÞ.

Note that when t ¼ 0; y ¼ a2 and when t ¼ 1; y ¼ 1. In a similar manner, we can evaluate the second integral in (1.4.39). Combining them, it follows that Z1 1

xðtÞ > 0 for 15t5 1 and 5 1 t51 ! yðtÞ ¼ 1; t5 1; t > 1:

! 1 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ dy 2 y þ a2

1.5 Signal Classification Based on Integrals The area of a signal xðtÞ is

1 fðaÞ fðaÞ ½dðt aÞ þ dðt þ aÞfðtÞdt ¼ þ : j2aj j2aj j2aj

Z1

Area½xðtÞ ¼

xðtÞdt:

(1:5:1)

1

This can be generalized. If xðtÞ has simple roots at t ¼ tn , then d½xðtÞ ¼

&

X tn

If a signal xðtÞ is said to be absolutely integrable, then

1 dxðtÞ dðt tn Þ; x0 ðtn Þ ¼ jt¼tn dt jx0 ðtn Þj

Area½jxðtÞj ¼

(1:4:40) &

Z1

jxðtÞjdt51:

(1:5:2)

1

x(t) y(t)

t

Fig. 1.4.3 (a) xðtÞ ¼ t2 1 and (b) yðtÞ ¼ uðt2 1Þ

(a)

t

(b)

20

1 Basic Concepts in Signals

If a signal is square integrable, i.e., a finite energy signal satisfies Z1 2 (1:5:3) Area½jxðtÞj ¼ jxðtÞj2 dt51: Consider a resistor of value R O. Ohm’s law states that the voltage, vðtÞ; across this resistor is equal to R times the current iðtÞ passing through the resistor and vðtÞ ¼ RiðtÞ. The instantaneous power delivered to the resistor is pR ðtÞ ¼ i2 ðtÞR. The total energy delivered to the resistor is ER ¼

Z1

pR ðtÞdt ¼ R

1

1 i ðtÞdt ¼ R 2

1

Z1

v2 ðtÞdt:

1

If we normalize the resistor value to 1O, that is, R ¼ 1, then we can consider the function, xðtÞ as a generic (i.e., either voltage or current) function. The energy in xðtÞ is defined by Ex ¼

Z1

2

jxðtÞj dt:

(1:5:4)

1

The normalized average power of a signal xðtÞ is defined by 1 Px ¼ lim T!1 T

Aeat dt ¼

ZT=2

jxðtÞj2 dt:

(1:5:5)

The signal xðtÞ is an energy signal if 05Ex 51, that is, Ex is finite and Px ¼ 0. The function xðtÞ is a power signal if 05Px 51, i.e., Px is finite and therefore Ex is infinite. If a signal does not satisfy one of these conditions, then it is neither an energy signal nor a power signal. The power and energy signals are mutually exclusive. Example 1.5.1 Show that the function given below is an energy signal. xðtÞ ¼ Aeat uðtÞ; a > 0:

(1:5:6)

Find its area, the absolute area, and its energy assuming A 6¼ 0 is finite. Solution: The area, the absolute area, and the energy are, respectively, given by

Z1

Ex ¼

jxðtÞj2 dt ¼ jAj2

1

Z1 h

2

jeat j uðtÞdt ¼ jAj2

1

¼

i

A 2a

2

Z1

e2at dt

0

! jAj2 : 2a

e2at t¼1 t¼0 ¼

From this it follows that Ex is finite. Clearly, Px ¼ 0 & implying it is an energy signal. Example 1.5.2 Using a ! 0 in the above example show that xðtÞ is a power signal. xðtÞ ¼ AuðtÞ; A 6¼ 0 and finite:

(1:5:7)

Solution: The energy contained in a step function Ex is infinite, whereas 2 3 ZT=2 6 1 7 Px ¼ lim 4 jxðtÞj2 dt5 T!1 T T=2

2

3 ZT=2 6 1 7 ¼ lim 4 jAj2 dt5 T!1 T 0

T=2

A at 1 A e j0 ¼ ; a a

0

Area½jxðtÞj ¼ jAj=a;

1

Z1

Area½xðtÞ ¼

Z1

¼ lim jAj2 T!1

T=2 T

¼

jAj2 2

is finite. It follows that xðtÞ in (1.5.7) is a power signal. In determining whether a signal is a power or an energy signal, we can check either its energy or power. If Ex is finite, then Px ¼ 0. If Px is finite, then Ex is infinite. We do not have to check both of them. If Ex is infinite, then we need to check the average power before making the decision on whether the signal is a power signal or neither a & power nor an energy signal. Example 1.5.3 Show that xðtÞ defined below is neither an energy nor a power signal. xðtÞ ¼ tuðtÞ

(1:5:8)

Solution: The energy and the average power in this signal are, respectively, given by

1.5 Signal Classification Based on Integrals

Z1

1 Ex ¼ t dt ¼ 1; Px ¼ lim T!1 T 0 3 1 T ¼ 1: ¼ lim T1 !1 3 T 2

21

ZT=2

Z1 2

t dt

xðt aÞdt ¼

1

0

Example 1.5.4 Show that xðtÞ ¼ A; a finite constant is a power signal. Solution: The average power contained in xðtÞ is given by 2 3 ZT=2 61 7 Px ¼ lim 4 A2 dt5 ¼ A2 ¼)E ¼ 1: T!1 T T=2

It is a power signal.

1

&

Z1 1

1 xðatÞdt ¼ j aj

Notes: The signals dðtÞ and d ðtÞ are neither nor power signals since the squares of these functions are not defined. There are two classes of power signals. These are periodic and random signals. Random signals require some knowledge of probability & theory, which is beyond the scope here.

1.5.1 Effects of Operations on Signals Example 1.5.5 In signal analysis, scaling and shifting of a function are quite common. a. Show that the functions xðtÞ and xðt aÞ have the same areas and energies. b. Show AreaðxðatÞÞ ¼ ð1=jajÞAreaðxðtÞÞ; a 6¼ 0: jxðt aÞj dt ¼

1

Z1

jxðbÞj2 db:

(1:5:9a) (1:5:9b)

1

Solution: a. Using a change of variable in the integral b ¼ t a results in

T1

1 lim T1 !1 T1

Z2 T

21

N X 1 1 yT ðtÞdt¼ lim N!1 2 N þ 1 T k¼N

ð2Z kþ1ÞT2 ð2 k1ÞT2

Z1 xðbÞdb: 1

Equation (1.5.9b) can be shown in both cases and is & left as an exercise.

1.5.2 Periodic Functions A function xðtÞ is periodic or T-periodic if there is a number T for all time such that xðt þ TÞ ¼ xðtÞ:

0

2

xðbÞdb:

b. Similarly, for at ¼ b results in

Since both the energy and the average power go to & infinity, the result follows.

Z1

Z1

(1:5:10)

It is common to use the actual period, such as T as a subscript on x and write xT ðtÞ ¼ xT ðt þ TÞ:

(1:5:11)

The smallest positive number T that satisfies (1.5.11) is called the fundamental period and it defines the duration of one complete cycle. The reciprocal of the fundamental period is the fundamental frequency. That is, f0 ¼ 1=T Hz and the period T ¼ 1=f0 s:

(1:5:12)

Clearly, if (1.5.11) is satisfied, then for all integers of n, xT ðt þ nTÞ ¼ xT ðtÞ. If there is no T that satisfies (1.5.10), then xðtÞ is called an aperiodic or a nonperiodic signal. Note that the integral over any one period of a periodic function is the same. Earlier we were interested in finding the average power and the average energy in a signal. With periodic signals, we can make some simplifications of the integrals. Consider the normalized integral of a periodic signal with period T1 .

T

2N þ 1 1 yT ðtÞdt¼ lim N!1 2 N þ 1 T

Z2 T2

T

1 yT ðtÞdt ¼ T

Z2 T2

yT ðtÞdt: (1:5:13)

22

1 Basic Concepts in Signals

Therefore, any one period can be used and written in symbolic short hand notation by T1 =2 Z

1 lim T1 !1 T1

yT ðtÞdt ¼

1 T

Z

yT ðtÞdt:

(1:5:14)

T

T1 =2

Fig 1.5.1 Half-rectified sine wave

The terms that are of interest in dealing with periodic functions are the duty cycle of an on–off signal (i.e., the ratio of on-time to the period), average value of the signal xave , the average signal power Px , and the root mean square (rms) value xrms . It is also called the effective value of the periodic function yT ðtÞ. These values are defined by xave

1 ¼ T

Z T

Px ¼

1 T

Z

1 Px ¼ T

x2T ðtÞdt

T

Z

2

¼

A T

jxT j2 dt; xrms ¼

Z

cos2 ðo0 t þ y0 Þdt

T

1 þ cosð2ðo0 t þ y0 ÞÞ A2 : (1:5:17) dt ¼ 2 2

T

The rms value is pﬃﬃﬃ xrms ¼ ðjAj= 2Þ:

xT ðtÞdt; Z

A2 ¼ T

pﬃﬃﬃﬃﬃﬃ Px :

(1:5:15)

T

Since the average power in a periodic signal is finite, the energy is infinite, it follows that all periodic signals are power signals. In Chapter 3 periodic functions will be discussed in detail, where the average value of a periodic function can never exceed the rms value will be shown.

It is best to express sinusoidal functions in terms of Hertz to compute the period. The period is the inverse of the fundamental frequency. Interestingly, if the frequency f0 ¼ o0 =2p is 1 MHz, then the signal completes 1 million cycles every second. & Example 1.5.7 Consider the half-wave sinusoidal periodic function

sinðtÞ; 0 t5p ; x2p ðtÞ ¼ x2p ðt þ 2pÞ 0; p t 5 2p

x2p ðtÞ ¼ Example 1.5.6 Consider the function xT ðtÞ with A; o0 ; and y0 being real constants given below. a. Show that it is a periodic function with period T ¼ ð2p=o0 Þ ¼ ð1=f0 Þ. xT ðtÞ ¼ A cosðo0 t þ y0 Þ:

(1:5:16)

b. Find xave ; Px ; and xrms for the above periodic signal. Solution: a. Using tables it can be seen that xT ðtÞ is a periodic function, i.e., xT ðt þ TÞ ¼ A cosðo0 ðt þ ð2p=o0 ÞÞ þ y0 Þ ¼ A cosðo0 t þ y0 Þ cosð2pÞ A sinðo0 t þ y0 Þ sinð2pÞ ¼ A cosðo0 t þ y0 Þ:

(1:5:19) shown in Fig. 1.5.1. Find its duty cycle, average, average signal power, and its rms value. Show that average value of the function is less than its rms value. Solution: Clearly the duty cycle is (1/2), as the signal is on for half the time. The average, the power, and the corresponding root mean square values of x2p ðtÞ are Zp Zp 1 1 1 x2p ðtÞdt ¼ sinðtÞdt ¼ (1:5:20a) xave ¼ 2p 2p p 0

1 Px ¼ 2p

Zp 0

b. The average value of a sine or a cosine function is zero as their positive areas cancel out with their negative areas. The average power is independent of y0 and

(1:5:18)

¼

1 4p

Zp 0

0

1 sin ðtÞdt ¼ 2p 2

Zp

1 ð1 sinð2tÞÞdt 2

0

ð1 sinð2tÞÞdt ¼

1 4

(1:5:20b)

1.5 Signal Classification Based on Integrals

23

Fig. 1.5.2 xT1 ðtÞ ¼ sinðð2p=4ÞtÞ; xT2 ðtÞ ¼ sinðð2p=6ÞtÞ

xrms ¼

pﬃﬃﬃﬃﬃﬃ Px ¼ 1=2:

(1:5:20c)

Note that the average value is less than the rms & value, as ð1=pÞ ﬃ 0:31835ð1=2Þ ¼ 0:5: Notes: The average value of a periodic function does not exceed the rms value, i.e.,xave xrms . The rms value of the sinusoidal voltage supplied to the outlet of a US home is 120 V with a frequency of 60 Hz. The pﬃﬃﬃ maximum value of the voltage at the & outlet is 2ð120Þ ¼ 169:71 V 170 V.

1.5.3 Sum of Two Periodic Functions

least common multiple of 4 and 6 is 12 and, therefore, xT ðtÞ=xT1 ðtÞ þ xT2 ðtÞ is periodic with period T ¼ 12 s. This can be seen from the fact that in 12 s, xT1 ðtÞ will have three full cycles, xT2 ðtÞ will have two full cycles, and xT ðtÞ will have one full cycle. Figure 1.5.2 gives sketches of xT1 ðtÞ and xT2 ðtÞ. If each of the signals is shifted by different amounts, say if yT ðtÞ ¼ A1 sinð2pð1=4Þtþy1 ÞþA2 sinð2pð1=6Þtþy2 Þ, then yT ðtÞ is still periodic with period T=12 s for any set of constants A1 ; A2 and angles y1 and y2 . This can be generalized and state that for any constants X½0; h½k, and y½k, the function xT ðtÞ ¼ X½0 þ

1 X

h½k cosðko0 t þ y½kÞ

(1:5:22)

k¼1

If xT1 ðtÞ and xT2 ðtÞ are two periodic functions with periods T1 and T2 , respectively, then xðtÞ ¼ xT1 ðtÞ þ xT2 ðtÞ is periodic with period T if T ¼ nT1 ¼ mT2 or ½T1 =T2 ¼ ½m=n:

(1:5:21)

m and n are some integers and ðT1 =T2 Þ is a rational number. The period of xðtÞ is equal to the least common multiple (LCM) of T1 and T2 . The LCM of two integers, m and n, is the smallest integer divisible by both m and n. If T1 =T2 is an irrational number, it cannot be written in terms of a ratio of two integers and xðtÞ is not periodic. Example 1.5.8 Let xT ðtÞ ¼ a1 cosðo0 tÞ and yT ðtÞ ¼ b1 sinðo0 tÞ, with T ¼ ð2p=o0 Þ. Show that zT ðt þ TÞ ¼ zT ðtÞ ¼ xT ðtÞ þ yT ðtÞ. Solution: Since yT ðt þ TÞ ¼ yT ðtÞ and xT ðt þ TÞ ¼ & xT ðtÞ implies zT ðtÞ is periodic. Example 1.5.9 Let xT1 ðtÞ ¼ A1 sinð2pð1=4ÞtÞ and xT2 ðtÞ ¼ A2 sinð2pð1=6ÞtÞ. Show that the period of xT ðtÞ ¼ xT1 þ xT2 is 12. Sketch the function xT ðtÞ. Solution: The period of xT1 ðtÞ is T1 =4 s and the period of xT2 ðtÞ is T2 ¼ 6 s. The ratio, ðT1 =T2 Þ ¼ ð4=6Þ is a rational number, and the

is periodic with period T ¼ o0 =2p. Noting that (o0 T ¼ 2p), we have cosðko0 ðt þ TÞ þ y½kÞ ¼ cosðko0 t þ y½kÞ cosðko0 TÞ sinðko0 t þ y½kÞ sinðko0 TÞ ¼ cosðko0 t þ y½kÞ: The term for k ¼ 1 is called the fundamental and the kth term is called the kth harmonic. The dc term is X½0. The above can be generalized and state that the following function is periodic with period T ¼ 2p=o0 for any constants X½0; A½k, and B½k: xT ðtÞ ¼ X½0 þ

1 X

A½k cosðko0 tÞ

k¼1

þ

1 X

B½k sinðko0 tÞ:

(1:5:23) &

k¼1

Example 1.5.10 Let xT1 ðtÞ ¼ cosð4tÞ, xT2 ðtÞ ¼ cosð2ptÞ, and xðtÞ ¼ xT1 ðtÞ þ xT2 ðtÞ. Show that xðtÞ is not a periodic function. Solution: The period of xT 1 ðtÞ ¼ cosð2pð2=pÞtÞ is T1 ¼ ðp=2Þ and the period of xT 2 ðtÞ is T2 ¼ 1: The ratio ðT1 =T2 Þ ¼ðp=2Þ is an irrational number and & xðtÞ is not periodic.

24

1 Basic Concepts in Signals

For functions such as the one given in the above example, there is no repetition. These types of combination of periodic functions are referred as quasi periodic or almost periodic. For a study of almost periodic functions, see Chuanyi, (2003). Example 1.5.11 Compute the average power in xðtÞ given below for the two cases: xðtÞ ¼ C1 cosðo1 t þ y1 Þ þ C2 cosðo2 t þ y2 Þ a. o1 ¼ no0 6¼ o2 ¼ ko0 , where n and k are integers b. o1 ¼ o2 ¼ o0 . Assume C1 ; C2 ; y1 ; and y2 are arbitrary constants. Solution: Without loosing any generality, assume T ¼ ð2p=o0 Þ. 2p 2p xðt þ Þ ¼ C1 cosðno0 ðt þ Þ þ y1 Þ o0 o0 2p þ C2 cosðko0 ðt þ Þ þ y2 Þ o0 ¼ C1 cosðno0 t þ y1 Þ þ C2 cosðko0 t þ y2 Þ: This indicates that xðtÞ ¼ xT ðtÞ is periodic for both cases with period T ¼ ðo0 =2pÞ. a: P ¼

1 T0

Z

jxT ðtÞj2 dt ¼

1 T0

T0

Z

½C1 cosðo1 t þ y1 Þ

T0 2

þC2 cosðo2 t þ y2 Þ dt Z 1 ½C21 þ C22 þ C21 cosð2ðo1 t þ y1 ÞÞ ¼ 2T0 T0

þC22

cosð2ðo2 t þ y2 Þdt Z 1 ½cosððo1 þ o2 Þt þ ðy1 þ y2 ÞÞ þ C1 C2 T0

xT ðtÞ ¼ C1 cosðo0 t þ y1 Þ þ C2 cosðo0 t þ y1 ðp=2ÞÞ ¼ C1 cosðo0 t þ y1 Þ C2 sinðo0 t þ y1 Þ:

&

Notes: Consider xðtÞ ¼ C cosðo0 t þ yÞ ¼ C cosðyÞ cosðo0 tÞ C sinðyÞ sinðo0 tÞ ¼ a cosðo0 tÞ þ b sinðo0 tÞ; pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ a ¼ C cosðyÞ; b ¼ C sinðyÞ; C ¼ a2 þ b2 ; y ¼ tan1 ðb=aÞ:

(1:5:26)

One should be careful in computing y as tan1 ðb=aÞ 6¼ tan1 ðb= aÞ; tan1 ðb= aÞ 6¼ tan1 ðb=aÞ:

&

Exponentially varying sinusoids: An example of such a function is xðtÞ ¼ Aeat cosðo0 t þ yÞ:

(1:5:27)

It becomes unbounded for at50. Our interest is for only positive t and such functions are referred as causal signals. If xðtÞ is defined for all t, then its causal part is yðtÞ ¼ xðtÞuðtÞ. In the case of xðtÞ in (1.5.27), yðtÞ ¼ xðtÞuðtÞ ¼ Aeat cosðo0 t þ yÞuðtÞ:

(1:5:28)

If a > 0 ða50Þ, xðtÞ in (1.5.28) is an exponentially decaying (increasing) sinusoidal function. These functions can be sketched using the envelopes Aeat and Aeat as constraints and the function cosðo0 t þ yÞ between the envelopes. Notes: Even for temporal signals, the analysis and design of noncausal systems is important. For example, the analysis of prerecorded data is applicable. &

T0

1 þ cosððo2 o1 Þt þ ðy2 y1 ÞÞdt ¼ ½C21 þ C22 2 (1:5:24) 1.6 Complex Numbers, Periodic, and

Symmetric Periodic Functions b. In the case of o2 ¼ o1 , the average power is 1 P ¼ ½C21 þ C22 þ C1 C2 cosðy2 y1 Þ: 2

(1:5:25)

In the above equation C1 C2 cosðy2 y1 Þ is equal to zero only when ðy2 y1 Þ ¼ p=2 and

A complex number ci ¼ ai þ jbi , where ai ¼ Reðci Þ is the real part and bi ¼ Imðci Þ is the imaginary part. Similarly if xðtÞ is a complex function, we can write it as xðtÞ ¼ ReðxðtÞÞ þ jImðxðtÞÞ:

(1:6:1)

1.6 Complex Numbers, Periodic, and Symmetric Periodic Functions

1.6.1 Complex Numbers A complex number can be written in terms of its real and imaginary parts or in terms of its magnitude and phase. Consider the complex number ci ¼ ai þ jbi ¼ ri ðcos yi þ j sin yi Þ ¼ ri ejyi ; ri qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ a2i þ b2i ; yi ¼ arctanðci Þ ¼ tan1 ðbi =ai Þ: (1:6:2) The representation ci ¼ ri ejyi is referred to as the polar form of the complex number, where ri and yi are, respectively, called the amplitude (or the modulus) and the phase angle associated with the complex number. One needs to be careful in using the formula yi ¼ arctanðbi =ai Þ, especially when the real part of ai is negative. Example 1.6.1 Sketch the following complex numbers as vectors on a complex plane. c1 ¼ 1 þ j1; c2 ¼ 1 þ j1; c3 ¼ 1 j1; c4 ¼ 1 j1:

25

Solution: The complex numbers c1 ; c2 ; c3 ; and c4 can be represented on the complex plane as vectors, where the length of the vector is equal to ri , and are illustrated in Fig. 1.6.1. They are located in the first, second, third, and fourth quadrants, respectively. Also, pﬃﬃﬃ p jci j ¼ ri ¼ 1= 2; i ¼ 1; 2; 3; 4; y1 ¼ ¼ 45 ; 4 3p 5p 7p ¼ 135 ; y3 ¼ ¼ 225 ; y4 ¼ ¼ 315 y2 ¼ 4 4 4 Noting that if we use the arctangent function, i.e., yi ¼ arctanðbi =ai Þ, we have y1 ¼ y3 and y2 ¼ y4 . Note the ambiguity in taking the ratio of positive (negative)/negative (positive) quantities. This ambiguity can be reconciled, for example, by noting that when the vector lies in the second quadrant, the angle must satisfy 90 y 180 . The correct angle is 180 y2 ¼ 180 arctanðb2 =a2 Þ ¼ 180 45 ¼ 135 :

&

A general method for obtaining the phase angle of a complex number c ¼ a þ jb is

(a)

(b)

(c)

(d)

Fig. 1.6.1 (a) c1 ¼ 1 þ j1; ðbÞ c2 ¼ 1 þ j1; ðcÞ c3 ¼ 1 j1; ðdÞ c4 ¼ 1 j1

26

1 Basic Concepts in Signals

y¼

if a 0 if a50

arctanðb=aÞ 180 þ arctanðb=aÞ

(1:6:3)

When a50, select the appropriate one so that jyj 1800 or p radians. In terms of power series, in radians (1 rad ¼ p=180 ), 3

5

7

arctanðyÞ ¼ y y 3 þ y5 y7 þ ; jyj51 þ; y 1 : p 1 1 1 arctanðyÞ ¼ 2 y þ 3y2 5y5 þ ; ; y 1 (1:6:4) Convergence of the series is fast for most values of y and very few terms are needed to compute the arctangent function. The worst case is when y ¼ 1. The conjugate, polar representation, sums, differences, multiplications, and divisions of complex numbers are given below, where we assume ci ¼ ai þ jbi . c 1 ¼ ða1 jb1 Þ; ci ¼ ri ejyi ; c1 c2 ¼ ða1 þ a2 Þ jðb1 þ b2 Þ

(1:6:5a)

c1 c2 ¼ ða1 þ jb1 Þða2 þ jb2 Þ ¼ ða1 a2 b1 b2 Þ þ jða1 b2 þ b1 a2 Þ ¼ r1 r2 ejðy1 þy2 Þ

(1:6:5b)

c1 a1 þ jb1 a1 þ jb1 a2 jb2 ¼ ¼ c2 a2 þ jb2 a2 þ jb2 a2 jb2 ða1 a2 þ b1 b2 Þ ða1 b2 b1 a2 Þ ¼ j ; c2 6¼ 0 2 2 a 2 þ b2 a22 þ b22 ðc1 =c2 Þ ¼ ðr1 =r2 Þejðy1 y2 Þ :

r ¼ absðcÞ ;y ¼ angleðcÞ ¼ atan2ðimagðcÞ; realðcÞÞ: (1:6:8) Notes: MATLAB atanðxÞ function computes the arctangent or inverse tangent of x. The function returns an angle in radians between ðp=2Þ and ðp=2Þ. MATLAB atan(x, y) computes the arctangent of ðy=xÞ. It returns an angle in radians between p and p and the signs on both x and y plays a role. & See Appendix B for MATLAB. In the last section we considered a sum of two sinusoids. Euler’s formula can be used to express sinusoids in terms of complex exponentials and vice versa. These are cosðyÞ ¼ ½ejy þ ejy =2;

ejy ¼ cosðyÞ þ j sinðyÞ; ejy ¼ cosðyÞ j sinðyÞ: (1:6:9b) Notes: MATLAB atanðxÞ function computes the arctangent or inverse tangent of x. The function returns an angle in radians between ðp=2Þ and ðp=2Þ. MATLAB atan(x, y) computes the arctangent of ðy=xÞ. It returns an angle in radians between p and p and the signs on both x and y plays a role. See Appendix B for MATLAB. & Example 1.6.2 Express the following functions in terms of single sinusoids. See (1.5.26).

(1:6:5c)

c1 ða21 b21 Þ 2a1 b1 ¼ 2 ; þj 2

2 c1 ða1 þ b1 Þ a1 þ b21 jc j jr j 1 ¼ 1 ¼ 1; y ¼ 2y1 ¼ 2 arctanðb1 =a1 Þ: (1:6:6) c j r 1 j

¼)

1

Consider the polar form of the complex number c and its natural log:

sinðyÞ ¼ ½ejy ejy =2j: (1:6:9a)

a: xðtÞ ¼ cosðo0 tÞ

pﬃﬃﬃ 2 sinðo0 tÞ; b: xðtÞ

¼ cosðo0 tÞ þ sinðo0 tÞ:

(1:6:10)

pﬃﬃﬃ Solution: a: In this case; we have a ¼ 1; b ¼ 2. From (1.6.10), we have qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃ 2ﬃ pﬃﬃﬃ pﬃﬃﬃ C ¼ 1 þ ð 2Þ ¼ 3; y ¼ tan1 ð 2=1Þ ﬃ 54:73 ; pﬃﬃﬃ 3 cosðo0 t þ 54:73 Þ pﬃﬃﬃ ¼ 3 cosð54:73 Þ cosðo0 tÞ pﬃﬃﬃ 3 sinð54:73 Þ sinðo0 tÞ:

xðtÞ ¼ c ¼ rejy ;

lnðcÞ ¼ lnðrejy Þ ¼ lnðrÞ þ jy r ¼ jcj; y ¼ ImðlnðcÞÞ

(1:6:7)

MATLAB can operate in terms of real or complex numbers. MATLAB commands for computing the amplitude and phase are

b. In this case a ¼ 1; b ¼ 1. From (1.6.10), it follows that

1.6 Complex Numbers, Periodic, and Symmetric Periodic Functions

pﬃﬃﬃ 2; y ¼ tan1 ½1= 1 ¼ 135 ! xðtÞ pﬃﬃﬃ ¼ 2 cosðo0 t 135 Þ:

C¼

Case a: a=1; b=1; [theta, C] = cart2pol(a, b); theta_deg = (180/pi)*theta; C, theta_deg! C ¼ 1:7321; y ¼ 54:73560 Case b: a=1; b=1; [theta, C] = cart2pol(a,-b); theta_deg = (180/pi)*theta; C, theta_deg!C ¼ 1:4142; y ¼ 135 .

27

Solution: a. ða1 jb1 oÞ ða2 jb2 oÞ XðjoÞ ¼ ða2 þ jb2 oÞ ða2 þ jb2 Þ ða1 a2 b1 b2 o2 Þ ða1 b2 þ a2 b1 Þo j : ¼ 2 2 2 ða2 þ b2 o Þ ða22 þ b22 o2 Þ b. In this part jXðjoÞj ¼

sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ a21 þ ðb1 oÞ2

; yðoÞ a22 þ ðb2 oÞ2 ða1 b2 þ a2 b1 Þo ¼)jXðjoÞja1 ¼a2 ;b1 ¼b2 ¼ arctan a1 a2 b1 b2 o2 sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ a2 þ b2 o 2 ¼ jXðjoÞj ¼ a2 þ b2 o 2 ¼ 1; yðoÞ ¼ 2 arctanðbo=aÞ:

De Moivre’s theorem defines the power of a complex number c. For p a real number,

&

In Chapter 7, systematic methods for sketching the magnitude and phase representations of complex rational functions of o will be considered. MATLAB plots will be considered in Appendix B.

ðcÞp ¼ ðrejy Þp ¼ ½rðcos y þ j sin yÞp ¼ rp ½cosðpyÞ þ j sinðpyÞ:ðcÞ1=n y þ 2 kp y þ 2 kp 1=n þ j sin ; ¼r cos n n

1.6.2 Complex Periodic Functions

k ¼ 0; 1; 2; . . . ; n 1; n ¼ an integer:

xðtÞ ¼ Aejðo0 tþy0 Þ ¼ Acosðo0 t þ y0 Þ þ jAsinðo0 t þ y0 Þ: (1:6:13)

Approximation of the amplitude and the phase angle of a complex number: The magnitude pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃof ﬃ a complex number c ¼ a þ jb is jcj ¼ ða2 þ b2 Þ ¼ r. Representing c in terms of ða; bÞ is the rectangular coordinate representation and ðr; yÞ is the polar coordinate representation with y ¼ arctanðb=aÞ. Finding the square root of a number is not as simple as a multiplication. To save the computational time, the magnitude jcj is usually approximated. For a simple algorithm see Problem 1.6.1. Example 1.6.3 Consider the complex function below with the real variable o. a . Derive the rectangular and polar form expressions for XðjoÞ: XðjoÞ ¼

a1 jb1 o ; ai > 0; bi > 0; i ¼ 1; 2: (1:6:12) a2 þ jb2 o

b. Simplify the expressions when a1 ¼ a2 ¼ a and b1 ¼ b2 ¼ b:

For A; o0 ; and y0 constants, Euler’s identity is

Note that xðtÞ ¼ xðt þ TÞ ¼ xT ðtÞ with the period T ¼ 2p=o0 and ej2p ¼ 1. xT ðt þ ð2p=o0 ÞÞ ¼ Aeðjo0 ðtþð2p=TÞþy0 Þ ¼ Aejðo0 tþy0 Þ ejo0 ð2p=TÞ ¼ Aejðo0 tþy0 Þ

(1:6:14)

In Chapter 3, periodic functions xT ðtÞ will be approximated by the complex Fourier series expansion: xT ðtÞ ¼

1 X

Xs ½kejno0 t :

(1:6:15)

k¼1

1.6.3 Functions of Periodic Functions If a function xT ðtÞ is periodic, then gðxT ðt þ TÞÞ ¼ gT1 ðxT ðtÞÞ; T1 T:

(1:6:16)

28

1 Basic Concepts in Signals

That is, gðtÞ is also periodic with the same period T or smaller. One problem of interest is given the period of xT ðtÞ the period of the function gðtÞ is to be computed.

Example 1.6.6 Show that the function xT ðtÞ is a half-wave symmetric periodic function. xT ðtÞ ¼

b: g2 ðtÞ ¼ ejx2 ðtÞ ; x2 ðtÞ ¼ sinðo0 tÞ:

(1:6:17)

Solution: a. Noting the absolute value jcosðo0 tÞj, the fundamental period of g1 ðtÞ is T=2 ¼ ðp=o0 Þ. b. In this case, g2 ðtÞ is periodic with the period T ¼ ð2p=o0 Þ since g2 ðt þ TÞ ¼ ej sinðo0 ðtþTÞÞ ¼ ej sinðo0 tÞ :

(1:6:18) &

fa½2 k 1 cosð2 k 1Þo0 t

k¼1

Example 1.6.4 Find the fundamental periods of the following periodic functions: a: g1 ðtÞ ¼ jx1 ðtÞj; x1 ðtÞ ¼ cosðo0 tÞ;

1 X

þ b½2 k 1 sinð2 k 1Þo0 tg:

(1:6:21)

Solution: Noting that xT ðtÞ is an algebraic sum of sine and cosine terms with the same period T ¼ 2p=o0 , it is a periodic function. Furthermore, since o0 ðT=2Þ ¼ p and ð2 k 1Þ is an odd integer, we can use cos½2 k 1p ¼ 1 and sin½2 k 1p ¼ 0. Using these and the following, we have xðt þ ðT=2ÞÞ ¼ xðtÞ. cosð½2 k 1o0 ðt þ ðT=2ÞÞÞ ¼ cos½2 k 1o0 t cosð½2 k 1o0 ðT=2ÞÞ sin½2 k 1o0 t sinð½2 k 1o0 ðT=2ÞÞ

1.6.4 Periodic Functions with Additional Symmetries Even and odd functions apply for both energy and periodic signals. cosðo0 tÞ and sinðo0 tÞ are periodic functions with period ð2p=o0 Þ and have even and odd symmetries, respectively. Interestingly sine and cosine functions have four distinct parts in one period. These are for 0 t5T=4; T=4 t5T=2; T=2 t53T=4; and 3T=4 t5T. Having information for one-fourth of the period of a sine wave provides the information about the other three parts. This gives a clue on the sizes of the transmitting and receiving antennas to transmit and receive the sine wave, which is discussed in Chapter 10. Half-wave symmetric functions: A periodic function xT ðtÞ is half-wave symmetric if xT ðtÞ ¼ xT ðt ðT=2ÞÞ:

þ cos½2 k 1o0 t sinð½2 k 1o0 ðT=2ÞÞ

xT ðtÞ ¼

xT ðtÞ ¼ xT ðtÞ xT ðtÞ ¼ xT ðt þ T2 Þ

;

xT ðt þ TÞ ¼ xT ðtÞ : even quarter wave symmetry: (1:6:22) xT ðtÞ ¼ xT ðtÞ xT ðtÞ ¼ T ; xT ðtÞ ¼ xT ðt þ 2 Þ xT ðt þ TÞ ¼ xT ðtÞ : odd quarter wave symmetry (1:6:23) Example 1.6.7 Show that the function below xT ðtÞ has the even quarter-wave symmetry.

(1:6:20) xT ðtÞ ¼

Solution: Since xT ðtÞ is periodic with period T, we can write T T T & xT ðt Þ ¼ xT ðt þ T Þ ¼ xT ðt þ Þ: 2 2 2

&

Quarter-wave symmetric functions: If a periodic function xT ðtÞ has half-wave symmetry and, in addition, is either even or odd function, then it is said to have even or odd quarter-wave symmetry. That is,

(1:6:19)

Example 1.6.5 Show that if a function is half-wave symmetric, then xT ðtÞ ¼ xT ðt þ ðT=2ÞÞ:

sinð½2 k 1o0 ðt þ ðT=2ÞÞ ¼ sin½2 k 1o0 t cosð½2 k 1o0 ðT=2ÞÞ

1 X

a½2 k 1 cosðð2 k 1Þo0 tÞ:

(1:6:24a)

k¼1

Solution: Since xT ðtÞ is a sum of cosine terms with a zero phase, it is an even symmetric function. It has even quarter-wave symmetry since

1.7 Examples of Probability Density Functions and their Moments

T cosð½2 k 1o0 ðt þ ÞÞ 2 T ¼ cosð½2 k 1o0 tÞ cosð½2 k 1o0 Þ 2 T sinð½2 k 1o0 tÞ sinð½2 k 1o0 Þ 2 ¼ cosð½2 k 1o0 tÞ (1:6:24b) &

29

probability of existence of the noise is always positive. That is pðxÞ 0. Furthermore, the existence of noise is certain and therefore the integral of the probability density function must be 1. In summary, Z1 pðxÞ 0 and pðxÞdt ¼ 1: (1:7:1) 1

In a similar manner, it can be shown that the function below has odd quarter-wave symmetry. This is left as an exercise. yT ðtÞ ¼

1 X

b½2 k 1 sinðð2 k 1Þo0 tÞ:

For a good discussion on probability theory, see Peebles (2001). Any nonnegative function with area 1 can serve as a probability density function. The nth moment of pðxÞ is defined as

(1:6:25)

k¼1

mn ¼

As examples, Fig. 1.6.2a and b has even and odd quarter-wave symmetries, respectively. Hidden symmetries: The symmetries can be hidden within a constant as given by xT ðtÞ ¼ A þ A sinðo0 tÞ:

(1:6:26)

It is neither even nor an odd function. On the other hand ðxT ðtÞ AÞ is an odd function.

Fig. 1.6.2 (a) Even quarterwave symmetric function and (b) odd quarter-wave symmetric function

xn pðxÞdx:

(1:7:2)

1

The zero and the first moments are, respectively, defined by m0¼

Z1

pðxÞdx ¼ 1 and m1 ¼

1

Z1 xpðxÞdx: (1:7:3) 1

The moments about the mean are called the central moments and are defined by

1.7 Examples of Probability Density Functions and their Moments In this section, a brief introduction to the probability density function (PDF) pðxÞ of a random variable xðtÞ, such as the amplitude of a noise signal, is presented. Noise, by its nature, is unpredictable. It is assumed that the amplitude can take any value in a continuous range 15a5x5b51. The level of the noise amplitude can only be described in terms of averages. Since noise is ever present,

Z1

mn ¼

Z1

ðx m1 Þn pðxÞdx:

(1:7:4a)

1

¼)m2 ðxÞ ¼ Variance of xðtÞ ¼ s2x Z1 ¼ ðx m1 Þ2 pðxÞdx:

(1:7:4b)

1

pﬃﬃﬃﬃﬃ The positive square root of the variance sx ¼ þ s2x is called the standard deviation. It gives a measure of the spread of the probability density function. Now

xT (t)

xT (t)

(a)

(b)

30

1 Basic Concepts in Signals

m2 ¼

Z1

ðx m0 Þ2 pðxÞdx ¼

1

Z1

Z1

x2 pðxÞdx 2m0

ðða þ bÞ=2Þ2 ¼

1

Z1

xpðxÞdxþ m21

1

Variance :s2x ¼ m2 m21 ¼ ð1=3Þ½b2 þ ab þ a2

pðxÞdx ¼ m2 2m21 þ m21

1

¼ m2 m21 :

(1:7:5) s2x ¼ m2 m21 :

(1:7:6)

Mean and the variance are basic statistical values in the study of the p probability density functions. In ﬃﬃﬃﬃﬃ addition, sx ¼ þ s2x measures its effective width or duration.

ðb aÞ2 : 12

(1:7:10)

pﬃﬃﬃ Standard deviation ¼ sx ¼ ðb aÞ=2 3: (1:7:11) & The function in Fig. 1.7.1 is the uniform density function, as the variable x is equally likely to take any value in the range ½a; b. It will be used in Chapter 10 to describe the error caused by quantization of samples. Example 1.7.2 Consider the Gaussian probability density function shown in Fig. 1.7.2 and it is given in (1.7.12). Determine m0 ; m1 ; and s2x of this PDF.

Example 1.7.1 Determine m0 ; m1 ; and s2x for the function in Fig. 1.7.1.

p(x)

p(x)

x

x Fig. 1.7.1 Uniform density function

1 x ððb þ aÞ=2Þ P ðuniform PDFÞ: pðxÞ ¼ ba ba (1:7:7) Solution: By inspection, we have the area under the function is 1. That is, m0 ¼ 1: Also m1 ¼

Z1

xpðxÞdx ¼

1

Zb

(1:7:12) Solution: From tables, Z1 Z1 1 1 2 ðtax Þ2 =s2x e dt ¼ pﬃﬃﬃﬃﬃﬃ ey =2 dy ¼ 1: m0 ¼ pﬃﬃﬃﬃﬃﬃ 2psx 2p 1

1

(1:7:13)

a

(1:7:8)

Z1 1

3

1b a 1 ¼ ðb2 þ ab þ a2 Þ: 3 ba 3

That is, the area under the Gaussian function is equal to 1. The mean value is m1 ¼

a

¼

2 1 2 pðxÞ ¼ pﬃﬃﬃﬃﬃﬃ eðxax Þ =sx ðGaussian PDFÞ: 2pðsx Þ

x dx ba

bþa ðaverage value of xÞ: ¼ 2 Zb 2

2 t m2 ¼ Area t xðtÞ ¼ dt ba 3

Fig. 1.7.2 Gaussian density function

(1:7:9)

1 xpð xÞdt ¼ pﬃﬃﬃﬃﬃﬃ 2psx

Z1

xeðxax Þ

2

=2s2x

dx:

1

(1:7:14) Using the change of variable y ¼ ðx ax Þ=sx , we have

1.8 Generation of Periodic Functions from Aperiodic Functions

1 m1 ¼ pﬃﬃﬃﬃﬃﬃ 2p

Z1

ax 2 ðsx y þ ax Þey =2 dy ¼ pﬃﬃﬃﬃﬃﬃ 2p

1 Z1

sx þ pﬃﬃﬃﬃﬃﬃ 2p

yey

2

=2

Z1

ey

2

=2

dy

1

(1:7:15)

dy¼ ax :

1

These follow since the second one above on the right is an integral of an odd function of y over a symmetrical interval and it is zero. The first integral in (1.7.15) reduces to (1.7.13). To derive the variance s2x , start with 1 m0 ¼ pﬃﬃﬃﬃﬃﬃ 2psx Z1

Z1

2

2

eðxax Þ =2sx dt ¼ 1 or

1

2

2

eðxax Þ =sx dt ¼

pﬃﬃﬃﬃﬃﬃ 2psx :

31

Notes: Noise is random and unpredictable. When it is added to the information bearing signal, the message signal is masked or even obliterated. Noise cannot be eliminated. A measure of corruption of the signal by noise is an important measure. It is the ratio of the average signal power to variance of the noise. It is Signaltonoise ratio ¼ SNR Average message signal power ¼ : Variance of the noise;s2x

1.8 Generation of Periodic Functions from Aperiodic Functions Now we like to construct a periodic function from an aperiodic function, say jðtÞ.

1

Taking the derivative with respect to sx results in pﬃﬃﬃﬃﬃﬃ 2p ¼

Z1 1

¼

Z1 1

¼)s2x

1 ¼ pﬃﬃﬃﬃﬃﬃ 2psx

deðxax Þ dsx

2

yT ðtÞ ¼

=2s2x

dx

ðx ax Þ ðtax Þ2 =2s2x e dx: s3x 2

2

ðx ax Þ2 eðxax Þ =2sx dx: (1:7:16)

1

Gaussian PDF (see Peebles (2001)) is one of the most important PDF, as most of the noise processes & observed in practice are Gaussian. Example 1.7.3 Find m0 and s2x for the Laplace PDF defined by pðxÞ ¼ ðb=2Þe

bj xj

; b > 0; 15x51:

(1:7:17)

Solution: From integral tables, the mean and the variance are Z1 Z1 b b jxj e m0 ¼ dx ¼ b ebx dx ¼ ebx 1 x¼0 ¼ 1; 2 s2x ¼

1 Z1 1

0

b 2 jxj=b 2 x e dx ¼ 2 : 2 b

1 X

jðt þ kTÞ:

(1:8:1)

k¼1

2

Z1

&

jðtÞ is the principal segment of the periodic extension. Clearly, yT ðtÞ is a periodic function with period T s and is the periodic extension of jðtÞ. If jðtÞ is not time limited to a T s interval (for example, jðtÞ is nonzero for t > T and t 5 0), then jðtÞ and jðt þ TÞ terms will overlap and jðtÞ cannot be extracted from yT ðtÞ. Example 1.8.1 Using the principal segment jðtÞ ¼ L½t=t, sketch the periodic extensions for the following cases. a: T 2t; b: T52t. Solution: The periodic extension of the triangular function is yT ðtÞ ¼

1 X k¼1

L½

t þ kT : t

(1:8:2)

The function jðtÞ and its periodic extensions are sketched in Fig. 1.8.1a, b, and c. For simplicity, in the sketch for part a, T ¼ 2t is assumed. a. The functions L½ðt þ kTÞ=t and L½ðt þ ðk þ 1Þ TÞ=t do not overlap and the function jðtÞ can be extracted

(1:7:18) & yT ðtÞjk¼0 ¼L½t=t ¼ jðtÞ; jtj t:

(1:8:3)

32

1 Basic Concepts in Signals

yT (t),T = 2τ

(a)

yT (t),T < 2τ

(b)

(c)

Fig. 1.8.1 (a) L tt , (b) yT ðtÞ; T ¼ 2t, (c) yT ðtÞ; T52t

b. For T52t, L½ðt þ kTÞ=t and L½ðt þ ðk þ 1ÞTÞ=t overlap (See Fig. 1.8.1b). Recovery of fðtÞ from yT ðtÞ: is not possible. It is interesting to note the area of one period of the periodic extension of jðtÞ equals to the area of the function jðtÞ. This is a consistency check. See Ambardar (1995). Example 1.8.2 Show that the area under one period of the periodic extension of jðtÞ ¼ et=t uðtÞ is equal to the area under jðtÞ.

ZT

x1 ðtÞdt ¼ ½1 þ eT=t þ e2T=t þ . . .

0

ZT e

t=t

1 dt ¼ 1 eT=t

0

ZT e

t=t

1 eT=t 1 1 ¼ : dt ¼ T=t t 1e t

0

Z1

jðtÞdt ¼

0

Z1

1 et=t uðtÞdt ¼ : t

(1:8:6) &

0

Solution: The periodic extension of jðtÞ with a period of T can be written as yT ðtÞ ¼½eðt=tÞ uðtÞ þ eðtþTÞ=t uðt þ TÞ þ eðtþ2TÞ=t uðt þ 2TÞ þ þ ½eðtTÞ=t uðt TÞ þ eðt2TÞ uðt 2TÞ þ ...

(1:8:4)

x1 ðtÞ ¼ et=t ½uðtÞ þ eT=t uðt þ tÞ þ e2T=t uðt þ 2tÞ þ . . .; t > 0:

(1:8:5)

1.9 Decibel Decibel or dB is a logarithmic unit named after Alexander Graham Bell that is used to express power ratios. Given two powers P 1 and P 2 and their ratio ðP 2 =P 1 Þ, then

1.9 Decibel

33

Table 1.9.1 Sound Power (loudness) Comparison Threshold of audibility 0 dB Whisper 15 dB Average home 45 dB Riveting machine (30’ away) 100 dB Threshold of hearing 120 dB Jet plane 140 dB

Power ratio in dB ¼ 10 log10 ½P2 =P1 :

(1:9:1)

For computing the dB values from the amplitudes using MATLAB, see Section B.12 in Appendix B. The unit of bel is too large. For example, a human ear can detect audio power level differences of onetenth of a bel or 1 dB. The loudness of a few typical activities are shown in table 1.9.1. Also, 1 dB is approximately equal to the attenuation of one mile of a standard telephone cable. Power ratio is expressed by Power ratio ¼ ðP2 =P1 Þ ¼ 10dB=10 :

(1:9:2)

Even though decibels were originally meant to be used with respect to power ratios, they can be used to express absolute values of power interpreted as a ratio of power P2 to P 1 ¼ 1 Watt, referred to as 1 dBW. That is, 1 Watt is a reference and 10 log10 ðP2 =P1 Þ ¼ 10 log10 ðP2 in W=1 WattÞ dB W: (1:9:3) Positive decibels imply A ¼ P 2 =P 1 > 1, zero decibels imply A ¼ 1, and negative decibels imply A51. Decibels are used to express gains or losses in a system; gain is the output divided by input and loss is the input divided output. Instead of using 1 Watt as a reference, 1 mW can be used as a reference to compare small signal power levels, such as powers of radar echoes and the unit is dBm. Now

Table 1.9.2 Power ratios and their corresponding values in dB dB 0 1 2 3 4 5 6 7 8 9 Power ratios 1 1.26 1.6 2 2.5 3.2 4 5 6.3 8

The logarithmic decibel function provides a greater resolution when the power ratio is small, indicating a good way to recognize very small differences in the power levels (Table 1.9.2). The smaller the power ratio, that is less than 1, the larger the number of negative dBs required. As the ratio approaches zero, the negative dB increases without limit. Interestingly when you round of to the nearest whole decibel, the error in the power ratio is at most only 1 part in 7. The dB provides greater resolution when the power ratio is small. Small differences in the power levels are important in spectral analysis, analog and digital filter designs, control system designs, communication system designs, etc. A power ratio of 2 to 1 (1 to 2) is 3 dB (3 dB). Bandwidths of 3 dB play a major role in filter designs. A ratio of 108 to is only 80 dB and 108 is 80 dB. Power levels in radars have a large range. Dealing with large number of digits can be troublesome, as dropping a zero at the end of a large number makes the radar calculations wrong. The dB scale makes the numbers compressed. The power levels associated with seismic signals are low. One nice thing about dB scale is that we will be dealing with additions and subtractions rather than multiplications and divisions. Gain (or loss) is the term used for an increase (or decrease) in power level. For example, for an amplifier, Gain Output signal power coming out of the amplifier ¼ : Input signal power going into the amplifier (1:9:5)

P2 =P1 jP1 ¼1 mW ¼ 10 log10 ðP2 in Watts=1 mWÞ dB m: If the output power is 100 times the input power, (1:9:4) then As examples, a power of 1 Watt is 0 dBW, a power of 3 Watt is 4.77 dBW, and a power of 1 kW is 30 dBW. A power of 1 kilowatt is 0 dBm and a power of 1010 mW is 10 log10 ð1010 mW=1 mWÞ ¼ 100 dBm:

Gain dB ¼ 10 logð100Þ ¼ 20 dB:

(1:9:6)

If we use a wave guide or a cable, we have a loss in the power. That is, Loss ¼

Input power to the device : Output power of the device

(1:9:7)

34

1 Basic Concepts in Signals

As an example, let an amplifier be connected to an antenna by a waveguide and the guide absorbs 20% of the power. If the ratio of the input power to the output power is 10 to 8, i.e., 1.25 indicating a loss of: Loss in dB ¼ 10 logð1:25Þ ﬃ :9 dB ðor; gainisð :9Þ dBÞ::

(1:9:8)

Since logðABÞ ¼ logðAÞ þ logðBÞ, we can simply find the power gain (or loss) in dB by adding or subtracting the numbers in dB. This is illustrated in the example below. Example 1.9.1 Consider that the amplifier considered above is connected to an antenna by a waveguide. See Fig. 1.9.1. Assume the amplifier gain is Gamplifier =100 and the waveguide absorbs about 20% of the power. Determine the total gain.

power gain or loss by adding or subtracting the appropriate quantities. It can be obtained by finding the transfer functions of each block and determine the total transfer function. Since the transfer functions are functions of frequency, the power gains or losses are functions of frequencies. Notes: Filters keep approximately the same amplitude in the filter pass band and provide attenuation in the filter stop band. If HðjoÞ is the transfer function of the filter, then we have attenuation (gain) at a frequency o1 ¼ 2pf1 , if jHðjo1 Þj 1 (jHðjo1 Þj > 1). The corresponding attenuation and gain are as follows: a ¼ 20 logjHðjo1 Þj dB > 0 ðLossÞ; A ¼ 20 logjHðjo1 Þj ðGainÞ:

(1:9:11)

Solving for jHðjoÞj from (1.9.11) results in Loss : jHðjo1 Þj ¼ 1=ð10:05 Þa ; Gain :jHðjo1 Þj ¼ ð10:05 ÞA :

(1:9:12)

Loss in terms of dB and the corresponding decrease in amplitudes are given below: Fig. 1.9.1 Amplifier and a waveguide

1dB¼)approximately 10% decrease in jHj Solution: The total gain from the input to the amplifier to the antenna is GTotal ¼ Gamplifier =Gwave guide ¼ 100=1:25 ¼ 80: (1:9:9) We can determine the gain or loss by simply subtracting gain from the loss. That is,

¼) Power ratio ¼ Pr ¼ 10ð19:03=10Þ ’ 80:

2dB¼)approximately 20% decrease in jHj from 1 to :794: 3dB¼)approximately 30% decrease in jHj from 1 to :708: 6dB¼)approximately 50% decrease in jHj from 1 to :501:

ðGTotal ÞdB ¼ ðGamplifier ÞdB ðGwave guide ÞdB ﬃ 20 :97 ¼ 19:03 dB:

from 1 to :891:

&

(1:9:10) &

Systems are designed part by part, arranged in a cascade, generally drawn symbolically by a block diagram shown in Fig. 1.9.2. We assume that the system identified by block k is not loading the system identified by block (k1). That is, there is no loading effect. Corresponding to this we can compute the

1.10 Summary In this chapter some of the basics on signals are presented that a second semester junior in an electrical engineering program may have gone through. Some of the material, such as complex numbers, periodic functions, integrals, decibels, and others, are included to refresh the reader’s memory. Specific principal topics that were included are

Various types of continuous signals Useful signal operations involving time shifting, Fig. 1.9.2 A cascaded system

scaling, reversal, and amplitude shift

Problems

35

Approximations and simplifications for integrals

with symmetries Singularity functions that include impulse functions, step functions, etc., and functions that can be used to approximate impulses Signal classifications based on power and energy Periodic signals and special classes of periodic functions with symmetries Complex numbers and complex functions Energy signals and their moments Periodic extension of aperiodic functions Decibels

1.1.1 Peterson and Barney (1952) collected average formant frequencies for vowels spoken by adult male and female speakers. For example, first two average formant frequencies in Hertz for the two vowels =i= and =a= are given in the table below. For a particular subject the first two formant frequencies of one of the vowels given were determined and they are F1 ¼ 500 Hz and F2 ¼ 1600 Hz. Using the minimum distance classifier, and the below table, determine if this subject is a male or a female and what is the vowel =i= or =a=?

Fig. P1.2.1

1.2.1 Given the functions in Fig. P1.2.1, sketch a: xi ðt þ 1Þ; b: xi ðt 1Þ; c: xi ð2t 3Þ; d: xi ð2t þ 1Þ; i ¼ 1; 2:

jij

jaj

270 310 2290 2790

730 850 1090 1220

a: x1 ðtÞ ¼ et uðt 1Þ; b: x2 ðtÞ ¼ et ; c: x3 ðtÞ ¼ ½ðt þ 1Þ=ðt 1Þ: d: x4 ðtÞ ¼ cosðtÞ þ sinðtÞ; e: x5 ðtÞ ¼ P½ðt 1Þ=2: 1.2.3 Sketch the functions sincðplÞ and sinc2 ðplÞ and give an approximate value of the magnitude of the first side lobe and their values in dB. Use MATLAB if you have the access. See Appendix B for information on MATLAB. 1.3.1 Approximate the following integral for the intervals shown by using a. the rectangular integration formula and b. the trapezoidal integration formula.

1.1.2 Sketch the following:

Zp

p xðtÞdt; xðtÞ ¼ sinðtÞ; intervals : 0; ; 4 0 p p p 3p 3p ; ; ; ; ;p : 4 2 2 4 4

A¼

x2(t)

1.2.2 Find the even and odd parts of the following functions:

Problems

F1 Male Female F2 Male Female

x1(t)

h i t1 2t t1 t P ; b: x2 ðtÞ ¼ P :L ; Use xðtÞ at t ¼ kðp=4Þ; k ¼ 0; 1; 2; 3 to approximate 2 2 2 2 the area by assuming each strip is a rectangle or a c: x3 ðtÞ ¼ uðt1Þ:uðbtÞ for b ¼ 5;0;1;2 trapezoid. Compare these values to the actual value of the integral. 1.1.3 Response of a first-order system is given by 1.3.2 Evaluate the following integral: yðtÞ ¼ Að1 et=2 ÞuðtÞ. What is the time constant of the system and compute the value of the time at ZT=2 1 t the time yðtÞ reaches 63.21% of the response? What sinðo0 tÞdt: 2 T do we call this time?

a: x1 ðtÞ ¼ P

0

36

1 Basic Concepts in Signals

1.3.3 Derive the expressions for the following partials:

Z5 b:

Z1

@uðs; oÞ @vðs; oÞ @uðs;oÞ @vðs; oÞ ; ; ; and : @s @s o @o

a: x1 ðtÞ ¼ sinðtÞ; b: x2 ðtÞ ¼ cosðtÞ; c: x3 ðtÞ ¼ tanðtÞ; d: x4 ðtÞ ¼ secðtÞ; e: x5 ðtÞ ¼ cotðtÞ; f: x6 ðtÞ ¼ cscðtÞ: 1.4.2 Let xðtÞ ¼ dððt 1Þðt 2ÞÞ. Express the function in terms of a sum of impulses. 1.4.3 Show the following functions are limiting forms of impulse functions: a: x1 ðtÞ ¼ ðT=2ÞejtjT ;

b: x2 ðtÞ ¼ Tð1 ðjtj=TÞÞ; Tsin Tt T sin Tt 2 c: x3 ðtÞ ¼ ; d: x4 ðtÞ ¼ p tT p Tt 1 T 2 2 e: x5 ðtÞ ¼ Tept T ; f: x6 ðtÞ ¼ 2 p t þ T2

1.4.4 Show that xn ðtÞ can be used as an impulse representation

dðtÞ ¼ lim xn ðtÞ; xn ðtÞ ¼ :5 nenjtj n!1 Z 1 use xn ðtÞdt ¼ 1; lim xn ðtÞ ¼ 0 : n!1

1

a: d0 ðtÞ ¼ d0 ðtÞ; b: td0 ðtÞ ¼ dðtÞ; R1 dxðtÞ 0Þ c: xðtÞ ddðtt dt ¼ ; dt dt d:

t¼t0

1 e ¼ ddðtÞ dt ; xe ðtÞ ¼ p t2 þe2 ; e > 0:

1.4.6 Evaluate the following integrals using the properties of the impulse functions: Z5 a: 0

ðt2 þ 2t 1Þdðt 1Þdt;

c:

et uðtÞ

dðt 1Þ dt: dt

1

1.5.1a. Express an arbitrary real-valued signal in terms of its even and odd parts. b. Express the unit step function uðtÞ in terms of its even and odd parts and sketch the even and odd parts. Assume uð0Þ ¼ 1 in sketching the function. 1.5.2 Classify each of the following functions as either an energy signal or a power signal or neither. If the functions are either energy or power signals, give the corresponding energy or power. Otherwise, explain why they are not. t1 a: x1 ðtÞ ¼ L ; b: x2 ðtÞ ¼ eajtj ; a > 0; 2 c: x3 ðtÞ ¼ t et uðtÞ;

d: x4 ðtÞ ¼ sincðptÞ;

e: x5 ðtÞ ¼ 1=½pð1 þ t2 Þ;

2

f: x6 ðtÞ ¼ ept ;

g: x7 ðtÞ ¼ d0 ðtÞ; h: x8 ðtÞ ¼ sin2 ðtÞ: 1.5.3 Show that Z1 1

2

x ðtÞdt ¼

Z1 1

x2e ðtÞdt

þ

Z1

x20 ðtÞdt:

1

1.5.4 Given xðtÞ ¼ cosð2pð1000ÞtÞ, find the period of xðatÞ; a > 0. What can you say in the general case of a periodic function xT ðtÞ and xT ðatÞ; a 6¼ 0? 1.5.5 Find the mean, rms, and the peak values of the function

1.4.5 Show the following:

1 lim dxdte ðtÞ e!0

ddðt 1Þ dt; dt

0

GðjoÞ ¼ 1=ðs þ joÞ ¼ uðoÞ þ jvðoÞ;

and yi ðtÞ ¼ uðxi ðtÞÞ; 1.4.1 Sketch xi ðtÞ i ¼ 1; 2; 3; 4; 5; 6 for p=25t53p=2 given

ðt2 þ 2t 1Þ

xðtÞ ¼ A cosðo0 t þ f1 Þ þ B cosð2o0 t þ f2 Þ: What is the effect of the phase angles on the peak value of this function? 1.6.1 The power series expansion of the square root function is Spiegel (1968) qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ða2 þ b2 Þ ¼ a ð1 þ x2 Þ 1 1 2 1:3 3 ¼a 1þ x x þ x ; 2 2ð4Þ ð2Þð4Þ6Þ b x ¼ ; 1 5 x 1: a

Problems

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Simplest approximation: a2 þ b2 a þ ðb=2Þ. For c ¼ 1 þ j1, find jcj by the approximation and the error between the direct computation and the simplified method. 1.6.2 Solve for the roots of the polynomial 1 þ ð1Þn s2n ¼ 0 in general terms. The roots may be complex. Give the roots on the left half-plane for n ¼ 1; 2; 3; 4; 5. Plot these roots on a complex plane and comment on the magnitudes of these roots. 1.6.3 Give the even and odd parts of the function xðtÞ ¼ t; 05t5T=2 and zero elsewhere. 1.6.4 Give examples of half-wave, even, and odd quarter-wave symmetric functions. 1.6.5 Show that the function given in (1.6.25) has odd quarter-wave symmetry. 1.6.6 Consider the periodic function xT ðtÞ given below. What can you say about the hidden symmetry in this periodic function?

37

a: xL ðtÞ ¼ ðb=2Þebjtaj ; 15t51; b > 0 : Laplace function: 2

b: xR ðtÞ ¼ ð2=bÞðt aÞeðtaÞ =b uðt aÞ; 15t51; b > 0 : Rayleigh function: 1.8.1 Assuming T ¼ t; b: T ¼ t=2, give the expressions for the periodic extension of the function x1 ðtÞ ¼ P½t=t. 1.9.1 Show that if you round off to the nearest whole decibel, the error in the power at most 1 in 7 by noting that the plot of decibel versus power ratio in the interval between 0 and 1 is approximately a straight line (Simpson, Hughes Air Craft Company, 1983). 1.9.2 Convert the following to power ratios and approximate them in dB: pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃﬃﬃ Magnitude ratios : 1= 10; 1=2; 1= 2; 1; 2; 2; pﬃﬃﬃﬃﬃ 10; 5; 10; 100; 1000:

1.9.3 Two radar signals x1 ðtÞ and x2 ðtÞ are assumed xT ðtÞ ¼ ð1 ðt=TÞÞ; 05t5T and xT ðt þ TÞ ¼ xðtÞ: to have an average power of 3 dBm and10 dBm, 1.7.1 Determine the mean and the variances of the respectively. What are the corresponding absolute power levels? following two functions:

Chapter 2

Convolution and Correlation

2.1 Introduction In this chapter we will consider two signal analysis concepts, namely convolution and correlation. Signals under consideration are assumed to be real unless otherwise mentioned. Convolution operation is basic to linear systems analysis and in determining the probability density function of a sum of two independent random variables. Impulse functions were defined in terms of an integral (see (1.4.4a)) using a test function fðtÞ. Z1

fðtÞdðt t0 Þdt ¼ fðt0 Þ:

(2:1:1)

1

This integral is the convolution of two functions, fðtÞ and the impulse function dðtÞ to be discussed shortly. In a later chapter we will see that the response of a linear time-invariant (LTI) system to an impulse input dðtÞ is described by the convolution of the input signal and the impulse response of the system. Convolution operation lends itself to spectral analysis. There are two ways to present the discussion on convolution, first as a basic mathematical operation and second as a mathematical description of a response of a linear time-invariant system depending on the input and the description of the linear system. The later approach requires knowledge of systems along with Fourier series and transforms. This approach will be considered in Chapter 6. Although we will not be discussing random signals in any detail, convolution is applicable in dealing with random variables.

The process of correlation is useful in comparing two deterministic signals and it provides a measure of similarity between the first signal and a timedelayed version of the second signal (or the first signal). A simple way to look at correlation is to consider two signals: x1 ðtÞ and x2 ðtÞ. One of these signals could be a delayed, or an advanced, version of the other. In this case we can write x2 ðtÞ ¼ x1 ðt þ tÞ; 1 < t < 1. Multiplying point by point and adding all the products, x1 ðtÞx1 ðt þ tÞ will give us a large number for t ¼ 0, as the product is the square of the function. On the other hand if t 6¼ 0, then adding all these numbers will result in an equal or a lower value since a positive number times a negative number results in a negative number and the sum will be less than or equal to the peak value. In terms of continuous functions, this information can be obtained by the following integral, called the autocorrelation function of xðtÞ, as a function of t not t. Z1 Rxx ðtÞ ¼ xðtÞxðt þ tÞdt ¼ AC ½xðtÞ Rx ðtÞ: 1

(2:1:2) This gives a comparison of the function xðtÞ with its shifted version xðt þ tÞ. Autocorrelation (AC) provides a nice way to determine the spectral content of a random signal. To compare two different functions, we use the cross-correlation function defined by

Rxh ðtÞ ¼ xðtÞ hðtÞ ¼

Z1

xðtÞhðt þ tÞdt: (2:1:3)

1

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_2, Ó Springer ScienceþBusiness Media, LLC 2010

39

40

2 Convolution and Correlation

Note the symbol (**) for correlation. Correlation isIf xðtÞ and yðtÞ are orthogonal, the energy and power related to the convolution. As in autocorrelation, contained in the energy or power signal the cross-correlation in (2.1.3) is a function of t, the zðtÞ ¼ xðtÞ þ yðtÞ are respectively given by time shift between the function xðtÞ, and the shifted version of the function hðtÞ. Ez ¼ Ex þ Ey or Pz ¼ Px þ Py : (2:1:8)

2.1.1 Scalar Product and Norm

Some of the important properties of the norm are stated as follows:

The scalar valued function hxðtÞ; yðtÞi of two signals xðtÞ and yðtÞ of the same class of signals, i.e., either energy or power signals, is defined by 8 R1 > > xðtÞy ðtÞdt; > > > 1 > > > < energysignals: ð2:1:4aÞ hxðtÞ;yðtÞi¼ T=2 R > > > lim T1 xðtÞy ðtÞdt; > > T!1 > T=2 > > : powersignals: ð2:1:4bÞ Superscript (*) indicates complex conjugation. Our discussion will be limited to a subclass of power signals, namely periodic signals. In that case, assuming that both the time functions have the same period (2.1.4b) can be written in the symbolic form as follows: 1 hxT ðtÞyT ðtÞi ¼ T

Z

xðtÞy ðtÞdt:

T

1: kxðtÞk ¼ 0 if and only if xðtÞ ¼ 0;

(2:1:9a)

2: kxðtÞ þ yðtÞk kxðtÞk þ kyðtÞk; triangular inequality

(2:1:9b)

3: kaxðtÞk ¼ jajkxðtÞk:

(2:1:9c)

In (2.1.9c), a is some constant. One measure of distance, or dissimilarity, between xðtÞ and yðtÞ is kxðtÞ yðtÞk. A useful inequality is the Schwarz’s inequality given by jhxðtÞ; yðtÞij kxðtÞkkyðtÞk:

(2:1:9d)

(2:1:4c) The two sides are equal when xðtÞ or yðtÞ is zero or if yðtÞ ¼ axðtÞ where a is a scalar to be determined. This can be seen by noting that

Even though our interest is in real functions, for generality, we have used complex conjugates in the above equations. The norm of the function is defined by Ex ; energy signals : kxðtÞk ¼ hxðtÞ; xðtÞi1=2 ¼ Px ; power signals (2:1:5)

kxðtÞ þ ayðtÞk2 ¼ hxðtÞ þ ayðtÞ; xðtÞ þ ayðtÞi ¼ hxðtÞxðtÞi þ a hxðtÞ; yðtÞi þ ahxðtÞ; yðtÞi þjaj2 hyðtÞ; yðtÞi ¼ kxðtÞk2 þa hxðtÞ; yðtÞi þ ahxðtÞ; yðtÞi þjaj2 kyðtÞk2 : (2:1:10)

It gives the energy or power in the given energy or the power signal. The two functions, xðtÞ and yðtÞ, are orthogonal if Since a is arbitrary, select hxðtÞ; yðtÞi ¼ 0:

(2:1:6) a ¼ hxðtÞ; yðtÞi=kyðtÞk2 :

(2:1:11)

In that case, kxðtÞ þ yðtÞk2 ¼ kxðtÞk2 þkyðtÞk2 :

(2:1:7)

Substituting this in (2.1.10), the last two terms cancel out, resulting in

2.2

Convolution

41

2 hxðtÞ; yðtÞi

This definition describes a higher algebra and allows us to study the response of a linear timekxðtÞ þ ayðtÞk ¼ kxðtÞk kyðtÞk2 invariant system in terms of a signal and a system response to be discussed in Chapter 6. It should be ) kxðtÞk2 kyðtÞk2 hxðtÞ; yðtÞi2 0: emphasized that the end result of the convolution (2:1:12) operation is a function of time. Coming back to the sifting property of the impulse functions, consider Equality exists in (2.1.9d) only if xðtÞ þ ayðtÞ ¼ 0. the equation given in (2.1.1). Two special cases are Another possibility is the trivial case being either one of interest. of the functions or both are equal to zero. Ziemer Z1 and Tranter (2002) provide important applications fðaÞdðt aÞda fðtÞ dðtÞ ¼ on this important topic. 1 Correlations in terms of time averages: Cross(2:2:2a) Z1 correlation and autocorrelation functions can be expressed in terms of the time average symbols and ¼ fðt bÞdðbÞdb ¼ dðtÞ fðtÞ; 2

Rxh ðtÞ ¼

Z1

2

1

xðtÞhðt þ tÞdt ¼ hxðtÞhðt þ tÞi; (2:1:13a)

1

dðtÞ dðtÞ ¼

Z1

dðaÞdðt aÞda ¼ dðtÞ: (2:2:2b)

1

RT;xh ðtÞ ¼

ZT=2

1 T

xT ðtÞhT ðt þ tÞdt

T=2

¼

1 T

Z

(2:1:13b)

2.2.1 Properties of the Convolution Integral

xT ðtÞhT ðt þ tÞdt ¼hxT ðtÞhT ðt þ tÞi: 1. Convolution of two functions, x1 ðtÞ and x2 ðtÞ, satisfies the commutative property,

T

In the early part of this chapter we will deal with convolution and correlation associated with aperiodic signals. In the later part we will concentrate on convolution and correlation with respect to both periodic and aperiodic signals. Most of the material in this chapter is fairly standard and can be seen in circuits and systems books. For example, see Ambardar (1995), Carlson (1975), Ziemer and Tranter (2002), Simpson and Houts (1971), Peebles (1980), and others.

yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ x2 ðtÞ x1 ðtÞ:

(2:2:3)

This equality can be shown by defining a new variable, b ¼ t a, in the first integral in (2.2.1) and simplifying the equation. 2. Convolution operation satisfies the distributive property, i.e., x1 ðtÞ ½x2 ðtÞ þ x3 ðtÞ ¼ x1 ðtÞ x2 ðtÞ þ x1 ðtÞ x3 ðtÞ:

2.2 Convolution

(2:2:4)

The convolution of two functions, x1 ðtÞ and x2 ðtÞ, is defined by Z1 x1 ðaÞx2 ðt aÞda yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ ¼

Z1 1

1

x2 ðbÞx1 ðt bÞdb ¼ x2 ðtÞ x1 ðtÞ:

(2:2:1)

3. Convolution operation satisfies the associative property, i.e., x1 ðtÞ ðx2 ðtÞ x3 ðtÞÞ ¼ ðx1 ðtÞ x2 ðtÞÞ x3 ðtÞ: (2:2:5) The proofs of the last two properties follow from the definition. 4. The derivative of the convolution operation can be written in a simple form and

42

2 Convolution and Correlation

2 1 3 Z dyðtÞ d y 0 ðtÞ ¼ x1 ðaÞx2 ðt aÞda5 ¼ 4 dt dt

¼

1

¼

Z1

x1 ðaÞ

Z1

dx2 ðt aÞ db ¼ x1 ðtÞ x02 ðtÞ; dt

Z1

¼

1

2

dyðtÞ d ) ¼ ½x1 ðtÞ x2 ðtÞ dt dt dx1 ðtÞ dx2 ðtÞ ¼ x2 ðtÞ ¼ x1 ðtÞ : dt dt

¼4

3 x2 ða bÞda5x1 ðbÞdb

1

2 4

Ztb

3 x2 ðlÞdl5x1 ðbÞdb;

1

3

2

x2 ðlÞdl5x1 ðtÞ¼4

x1 ðbÞdb5x2 ðtÞ:

Example 2.2.1 Find the convolution of a function xðtÞ and the unit step function uðtÞ and show it is a running integral of xðtÞ: Solution: This can be seen from Z1 xðbÞuðt bÞdb xðtÞ uðtÞ ¼

m

d x1 ðtÞ d yðtÞ x2 ðtÞ ¼ m dt dtm

¼

1 Zt

xðbÞdb; 1

ðnÞ x2 ðtÞ

3

1

(2:2:6b) d i xi ðtÞ ðmÞ ðmÞ ; ¼ y ðtÞ Note xi ðtÞ ¼ dti

ðmÞ x1 ðtÞ

Zt

(2:2:8)

Equation (2.2.6a) can be generalized for higher order derivatives. We can then write m

Zt

Zt

1

(2:2:6a)

ðmÞ

4

1

1

x1 ðtÞ x2 ðtÞ ¼

2

1; uðt bÞ ¼ 0;

b5t : b4t

(2:2:9) &

d m x1 ðtÞ dn x2 ðtÞ ¼ dtm dtn (2:2:6c) d mþn yðtÞ ¼ ¼ yðmþnÞ ðtÞ: dtmþn

Since the impulse function is the generalized derivative of the unit step function uðtÞ (see Section 1.4.2.), we have yðtÞ ¼ uðtÞ hðtÞ ) y0 ðtÞ

6. Convolution of two delayed functions x1 ðt t1 Þ and x2 ðt t2 Þ are related to the convolution of x1 ðtÞ and x2 ðtÞ. yðtÞ ¼ x1 ðtÞ x2 ðtÞ ) x1 ðt t1 Þ x2 ðt t2 Þ (2:2:10) ¼ yðt ðt1 þ t2 ÞÞ: This can be seen from

(2:2:7) ¼ u0 ðtÞ hðtÞ ¼ dðtÞ hðtÞ ¼ hðtÞ: 5. Convolution is an integral operation and if we know the convolution of two functions and desire to compute its running integral, we can use Zt

Zt yðaÞda¼

1

¼

1 Zt 1

x1 ðt t1 Þ x2 ðt t2 Þ Z1 ¼ x1 ða t1 Þx2 ðt a t2 Þda

¼

4

Z1 1

x1 ðbÞx2 ð½t ðt1 þ t2 Þ bÞdb

1

½x1 ðaÞ x2 ðaÞda 2

1 Z1

¼ yðt ðt1 þ t2 ÞÞ:

(2:2:11)

3 x1 ðbÞx2 ða bÞdb5da;

Example 2.2.2 Derive the expression yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ dðt t1 Þdðt t2 Þ.

for

2.2

Convolution

43

Solution: Using the integral expression, we have Z1

Z1

x1 ðaÞx2 ðt aÞda¼

1

x1e ðtÞ x20 ðtÞ ¼ y01 ðtÞ; x10 ðtÞ x2e ðtÞ ¼ y02 ðtÞ; odd functions:

dða t1 Þdðt t2 aÞda

(2:2:14c)

1

¼ dðt t2 aÞja¼t1 ¼ dðt t1 t2 Þ: 8. The area of a signal was defined in Chapter 1 (see (1.5.1)) by Z1 Noting dðtÞ dðtÞ ¼ dðtÞ and using (2.2.11), we have A½xi ðtÞ ¼ xi ðaÞda: (2:2:15) & dðt t1 Þ dðt t2 Þ ¼ dðt t1 t2 Þ. 1

7. The time scaling property of the convolution operation is if yðtÞ ¼ x1 ðtÞ x2 ðtÞ, then x1 ðctÞ x2 ðctÞ ¼

Z1

x1 ðcbÞx2 ðcðt bÞÞdb

Area property of the convolution applies if the areas of the individual functions do not change with a shift in time. It is given by A½yðtÞ ¼ A½x1 ðtÞ x2 ðtÞ ¼ A½x1 ðtÞA½x2 ðtÞ: (2:2:16)

1

1 ¼ yðctÞ; c 6¼ 0: jcj

(2:2:12)

This can be proved by A½yðtÞ

¼

Assuming c < 0 and using the change of variables a ¼ cb, and simplifying, we have

x1 ðctÞ x2 ðctÞ ¼

1 c

Z1 1

¼

1 j cj

¼

1 Z1 1

x1 ðaÞx2 ðct aÞdy

Z1

Z1

x1 ðaÞx2 ðct aÞda ¼

1

1 yðctÞ: jcj

¼

yðbÞdb ¼ 2

Z1

4

:

x1 ðtÞ x2 ðtÞ ¼ yðtÞ:

(2:2:13)

This property simplifies the convolution if there are symmetries in the functions. In Chapter 1, even and odd functions were identified by subscripts e for even and 0 for odd (see (1.2.7)). From these xie ðtÞ ¼ xie ðtÞ; an even function; xi0 ðtÞ ¼ xi0 ðtÞ; an odd function

(2:2:14a)

x1e ðtÞ x2e ðtÞ ¼ ye1 ðtÞ; x10 ðtÞ x20 ðtÞ ¼ ye2 ðtÞ; even functions (2:2:14b)

½x1 ðbÞ x2 ðbÞdb 3

x1 ðaÞx2 ðb aÞda5db

x1 ðaÞ

¼ A½x2 ðtÞ A similar argument can be given in the case of c > 0: Scaling property applies only when both functions are scaled by the same constant c 6¼ 0. When c ¼ 1, then

1

1

8 Z1 < 1

Z 1

Z1

Z1 1

9 = x2 ðb aÞdb da ;

x1 ðaÞda ¼ A½x2 ðtÞA½x1 ðtÞ:

1

9. Consider the signals x1 ðtÞ and x2 ðtÞ that are nonzero for the time intervals of tx1 and tx2 , respectively. That is, we have two time-limited signals, x1 ðtÞ and x2 ðtÞ, with time widths tx1 and tx2 . Then, the time width ty of the signal yðtÞ ¼ x1 ðtÞ x2 ðtÞ is the sum of the time widths of the two convolved signals and ty ¼ tx1 þ tx2 . This is referred to as the time duration property of the convolution. We will come back to some intricacies in this property, as there are some exceptions to this property. Example 2.2.3 Derive the expression for the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ, where xi ðtÞ; i ¼ 1; 2 are as follows: x1 ðtÞ ¼ 0:5dðt 1Þ þ 0:5dðt 2Þ; x2 ðtÞ ¼ 0:3dðt þ 1Þ þ 0:7dðt 3Þ:

44

2 Convolution and Correlation

Solution: Convolution of these two functions is Z1 x1 ðaÞx2 ðt aÞda yðtÞ ¼

¼

1 Z1

½0:5dða 1Þ þ 0:5dða 2Þ

1

½0:3dðt a þ 1Þ þ 0:7dðt a 3Þda Z1

¼

ð0:5Þð0:3Þdða 1Þdðt a þ 1Þda

1

þ

Z1

ð0:5Þð0:7Þdða 1Þdðt a 3Þda

1

þ

þ

Z1 1 Z1

ð0:5Þð0:3Þdða 2Þdðt a þ 1Þda

ð0:5Þð0:7Þdða 2Þdðt a 3Þda

1

¼ ð0:15ÞdðtÞ þ 0:35dðt 4Þ þ 0:15dðt 1Þ þ 0:35dðt 5Þ:

properties to simplify the evaluations are illustrated. A few comments are in order before the examples. First, the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ is an integral operation and can use either one of the integrals in (2.2.1). Note that yðtÞ; 1 < t < 1 is a time function. The expression for the convolution, say at t ¼ t0 , will yield a zero value for those values of t0 over which x1 ðbÞ and x2 ðt0 bÞ do not overlap. The area under the product ½x1 ðbÞx2 ðt0 bÞ, i.e., the integral of this product gives the value of the convolution at t ¼ t0 . Sketches of the function x1 ðbÞ and the time reversed and delayed function x2 ðt0 bÞ on the same figure would be helpful in identifying the limits of integration of the product ½x1 ðbÞx2 ðt0 bÞ. As a check, the value of the convolution at end points of each range must match, except in the case of impulses and/or their derivatives in the integrand of the convolution integral. This is referred to as the consistency check. The following steps can be used to compute the convolution of two functions x1 ðtÞ and x2 ðtÞ: New variable

Notes: If an impulse function is in the integrand of the form dðat bÞ, then use (see (1.4.35), which is

Convolution of two functions exists if the convolution integral exists. Existence can be given only in terms of sufficient conditions. These are related to signal energy, area, and one sidedness. It is simple to give examples, where the convolution does not exist. Some of these are a*a, a*u(t), cos(t)*u(t), eat *eat, a > 0. Convolution of energy signals and the samesided signals always exist. In Chapter 4 we will be discussing Fourier transforms and the transforms make it convenient to find the convolution.

Shift

New variable

Multiply the two functions

!x1 ðbÞ !x1 ðbÞx2 ðt bÞ: x1 ðtÞ R1 Integrate ! x1 ðbÞx2 ðt bÞdb ¼ yðtÞ:

dðat bÞ ¼ ð1=jajÞdðt ðb=aÞÞ:

2.2.2 Existence of the Convolution Integral

Reverse

x2 ðtÞ !x2 ðbÞ !x2 ðbÞ!x2 ðt bÞ

1

Example 2.3.1 Derive the expression for the convolution of the two pulse functions shown in Fig. 2.3.1 a,b. These are 1 t ða=2Þ and x1 ðtÞ ¼ P a a 1 t ðb=2Þ x2 ðtÞ ¼ P ; b a > 0: b b

(2:3:1)

Solution: First

yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼

Z1

x1 ðbÞx2 ðt bÞdb: (2:3:2)

1

2.3 Interesting Examples In the following, the basics of the convolution operation, along with using some of the above

Figure 2.3.1c,d,e,f give the functions x1 ðbÞ; x2 ðbÞ; x2 ð bÞ; and x2 ðt bÞ, respectively. Note that the variable t is some value between 1 and 1 on the b axis. Different cases are considered, and

2.3

Interesting Examples

45 x2(t)

x1(t)

(a) x1(β)

(b) x2(–β)

x2(β)

(c)

(d)

(e) t ≤ 0 : x2 (t – β) and x1 (β)

x2(t – β)

(g)

(f) t > 0 : x2 (t – β) and x1 (β)

t > a : x2 (t – β) and x1 (β)

(h)

(i) t – b > a : x2 (t – β) and x1 (β)

t – b < a : x2 (t – β) and x1 (β)

(k)

(j)

b = a : y (t) = x1 (t)∗x2 (t)

b > a : y (t) = x1 (t)∗x2 (t)

(l)

(m)

Fig. 2.3.1 Convolution of two rectangular pulses (b a)

in each case, we keep the first function x1 ðbÞ stationary and move (or shift) the second function x2 ðbÞ, resulting in x2 ðt bÞ. Case 1: t 0: For this case the two functions are sketched in Fig. 2.3.1 g on the same figure. Noting that there is no overlap of these two functions, it follows that

yðtÞ ¼ 0; t 0:

(2:3:3)

Case 2: 05t a: The two functions are sketched for this case in Fig. 2.3.1 h. The two functions overlap and the convolution is yðtÞ ¼

Zt 0

x1 ðbÞx2 ðt bÞdb ¼

1 t; 05t a: (2:3:4) ab

46

2 Convolution and Correlation

Case 3: a5t b: This corresponds to the complete overlap of the two functions and the functions are shown in Fig. 2.3.1i. The convolution integral and the area is

yðtÞ ¼

Za

1 1 db ¼ ; a5t b: ab b

(2:3:5)

integral operation, which is a smoothing operation. Convolution values at end points of each range must match (consistency check) as we do not have any impulse functions or their derivatives in the functions that are convolved. Some of these are discussed below. The areas of the two pulses are each equal to 1 and the area of the trapezoid is given by

0

Case 4: 0 < t b a < b or b < t ðb þ aÞ. This corresponds to a partial overlap of the two functions and is shown in Fig. 2.3.1j. The convolution integral and the area is

yðtÞ ¼

Za

1 ða þ b tÞ db ¼ ; b5t ðb þ aÞ: (2:3:6) ab ab

tb

Case 5: t b > a or t a þ b: The two functions corresponding to this range are sketched in Fig. 2.3.1k and from the sketches we see that the two functions do not overlap and yðtÞ ¼ 0; t ða þ bÞ:

(2:3:7)

Summary: 8 > > > > > > > > > > > >

b > > > > > aþbt > > ; > > ab > > : 0;

Area½yðtÞ ¼ ð1=2Það1=bÞ þ ðb aÞð1=bÞ þ ð1=2Það1=bÞ ¼ 1 ¼ Area½x1 ðtÞArea½x2 ðtÞ:

This shows that the area property is satisfied. Peebles (2001) shows the probability density function of the sum of the two independent random variables is also a probability density function. We should note that the probability density function is nonnegative and the area under this function is 1 (see Section 1.7). From the above discussion, it follows that the convolution of two rectangular pulses (these can be considered as uniform probability density functions) results in a nonnegative function and the area under this function is 1. The function yðtÞ satisfies the conditions of a probability density function. The time duration of yðtÞ, ty is ty ¼ tx1 þ tx2 and tx1 ¼ a; tx2 ¼ b ) ty ¼ tx1 þ tx2 ¼ a þ b:

t0

(2:3:9)

(2:3:10)

05 t 5 a a t 5b

:

(2:3:8)

b t 5a þ b taþb

This function is sketched in Fig. 2.3.1 l and m for the cases of b > a and b ¼ a. There are several interesting aspects in this example that should be noted. First, the two functions we started with have firstorder discontinuous and the convolution is an

A special case is when a ¼ b and the function yðtÞ given in Fig. 2.3.1m, a triangle, is P

ht ai t a=2 t a=2 P ¼L : (2:3:11) a a a

&

Example 2.3.2 Give the expressions for the convolution of the following functions:

t1 x1 ðtÞ ¼ uðtÞ and x2 ðtÞ ¼ sinðptÞP : (2:3:12) 2

2.3

Interesting Examples

47

Solution: The convolution integral

yðtÞ ¼

Z1

x2 ðbÞx1 ðt bÞda ¼

1

Z2

Solution:

sinðpbÞ uðt bÞdb

yðtÞ ¼ xðtÞ hðtÞ ¼

0

8 > Zt < 0; t 0 ¼ sinðpbÞdb ¼ ð1=pÞð1 cosðptÞÞ; 05t52 ; > : 0; t 2 0

yðtÞ ¼

Zt sinðpbÞdb 0

8 > < 0; t 0 ¼ ð1=pÞð1 cosðptÞÞ; 05t52 : > : 0; t 2

(2:3:13)

The time duration of the unit step function is 1 and the time duration of x2 ðtÞ is 2. The duration of the function yðtÞ is 2, which illustrates a pathological case where the time duration property of the convolution is not satisfied. The integral or the area of a sine or a cosine function over one period is equal to zero. The period of the function sinðptÞ is equal to 2 and therefore t1 A½x2 ðtÞ ¼ A sinðptÞ:P 2 Z2 ¼ 0 ) A½yðtÞ ¼ ð1=pÞ ½1 cosðptÞdt

¼

Z1 1 Z1

xðbÞhðt bÞdb

hðaÞxðt aÞda:

(2:3:14b)

1

In computing the convolution, we keep one of the functions at one location and the other function is time reversed and then shifted. In this example, since the function hðtÞ ¼ 0 for t < 0, we have a benchmark to keep track of the movement of the function hðt bÞ as t varies. Therefore, the first integral in (2.3.14b) is simpler to use. The functions xðbÞ; hðbÞ; hðbÞ; and hðt bÞ are shown in Fig. 2.3.2 c, d, e, and f respectively. As before, we will compute the convolution for different intervals of time. Case 1: t T : the two functions, hðt bÞ and xðbÞ, are sketched in Fig. 2.3.2 g. Clearly there is no overlap of the two functions and therefore the integral is zero. That is yðtÞ ¼ 0; t T:

(2:3:15)

Case 2: T < t < T: The two functions hðt bÞ and xðbÞ are sketched in Fig. 2.3.2 h in the same figure for this interval. There is a partial overlap of the two functions in the interval T4t4T. The convolution can be expressed by

0

¼ 1=p

Z2

dt ¼ 2=p:

yðtÞ ¼

0

Noting that A½x1 ðtÞ ¼ A½uðtÞ ¼ 1 and A½yðtÞ ¼ 2=p, we can see that the area property of the convolution is not satisfied. See Ambardar & (1995) for an additional discussion. Example 2.3.3 Derive the expression for the convolution of the following functions shown in Fig. 2.3.2a,b: h t i xðtÞ ¼ P and hðtÞ ¼ eat uðtÞ; a > 0: (2:3:14a) 2T

Z1

xðbÞhðt bÞdb ¼

1

¼ eat

Zt

ð1ÞeaðtbÞ db

T

Zt

(2:3:16) h i 1 eab db¼ 1 eaðtþTÞ ; T5t5T: a

T

Case 3: t > T : From the sketch of the two functions in Fig. 2.3.2 h, the two functions overlap in this range T t T and the convolution integral is yðtÞ ¼

ZT

eaðtbÞ db ¼

1 aT e eaT eat ; a

t4T:

T

(2:3:17)

48

2 Convolution and Correlation

Fig. 2.3.2 Convolution of a rectangular pulse with an exponentially decaying pulse

x (t )

h(t)

(a) x(β)

(b)

h(β)

(c)

(d) h ( t − β ), t ≥ 0

h(−β)

(e)

(f)

h ( t − β ) , t < –T and x ( β )

h ( t − β ) , t < T and x ( β )

(g)

(h) y(t)

(i)

Summary: 8 0; t T > > > i > > > 1 > : eaT eaT eat ; t4T a This function is sketched in Fig. 2.3.2i. Note yðtÞ is smoother than either of the given functions used in the convolution. Computing the area of yðtÞ is not as simple as finding the areas of the two functions, xðtÞ and hðtÞ: Using the area property,

A½yðtÞ ¼ A½xðtÞA½hðtÞ ¼ ð2TÞð1=aÞ: (2:3:19) & Notes: In computing the convolution, one of the sticky points is finding the integral of the product ½xðbÞhðt bÞ in (2.3.14b), which requires finding the region of overlap of the two functions. Sketching both functions on the same figure allows for an easy determination of this overlap. The delay property is quite useful. For example, if yðtÞ ¼ xðtÞ hðtÞ then it implies y1 ðtÞ ¼ xðt TÞ hðtÞ ¼ yðt TÞ. In Example 2.3.3, x(t) ¼ P[t/2T] ¼ u[tþT] u½t T. Therefore

2.3

Interesting Examples

49

Example 2.3.5 Derive the expression yi ðtÞ ¼ hðtÞ xi ðtÞ for the following two cases:

yðtÞ ¼ hðtÞ xðtÞ ¼ hðtÞ ½uðt þ TÞ xðt TÞ ¼ hðtÞ uðt þ TÞ hðtÞ uðt TÞ:

&

a:x1 ðtÞ ¼ uðtÞ; b:x2 ðtÞ ¼ dðtÞ:

Example 2.3.4 Determine the convolution yðtÞ ¼ xðtÞ xðtÞ with xðtÞ ¼ eat uðtÞ, a > 0:

Solution: a. Since uðt aÞ ¼ 0; a > t, we have the running integral

Solution: The convolution is y1 ðtÞ ¼ hðtÞ uðtÞ ¼

yðtÞ ¼ eat uðtÞ eat uðtÞ Z1 ¼ eab eaðtbÞ ½uðbÞuðt bÞdb

¼e

Zt

hðaÞuðt aÞda

1

¼

1 at

Z1

Zt (2:3:22)

hðaÞda: 1

db ¼ teat uðtÞ:

(2:3:20)

0

In evaluating the integral, the following expression is used (see Fig. 2.3.3a): ½uðbÞuðt bÞ ¼

b. Noting that the impulse function is the generalized derivative of the unit step function, we can compute the convolution y2 ðtÞ ¼ hðtÞdðtÞ ¼ hðtÞ

0; 1;

b50 and b4t : 05 b 4 t

(2:3:21)

The functions xðtÞ and yðtÞ are shown in Fig. 2.3.3b,c. Note that the function xðtÞ has a discontinuity at t ¼ 0: The function yðtÞ, obtained by convolving two identically decaying signals, xðtÞ and xðtÞ is smoother than either one of the convolved signals. This is to be expected as the convolution operation is a smoothing operation. &

duðtÞ ¼ y 0 1 ðtÞ ¼ hðtÞ: dt (2:3:23) &

Example 2.3.6 Let hðtÞ ¼ eat uðtÞ; a > 0 a. Determine the running integral of hðtÞ. b. Using (2.3.23), determine y2 ðtÞ: Solution:

a: y1 ðtÞ ¼

Zt

hðbÞdb ¼

1

Zt

eab uðbÞdb

1

1 ¼ ð1 eat ÞuðtÞ; a

(2:3:24)

dy1 ðtÞ 1 d ¼ ð1 eat ÞuðtÞ dt a dt 1 d 1 dð1 eat Þ ¼ ð1 eat Þ uðtÞ þ uðtÞ a dt a dt ¼ ð1=aÞð1 eat ÞdðtÞ þ eat uðtÞ

b: y2 ðtÞ ¼

(a) x(t ) = e

− at

u (t )

x(t ) = e

− at

u (t )

y (t ) = x(t ) * x(t )

¼ ð1=aÞ½dðtÞ dðtÞ þ eat uðtÞ ¼ ð1=aÞeat uðtÞ:

(b) Fig. 2.3.3 Example 2.3.4

(c)

(2:3:25) &

In a later chapter this result will be used in dealing with step and impulse inputs to an RC circuit with an impulse response hðtÞ ¼ eat uðtÞ.

50

2 Convolution and Correlation

Example 2.3.7 Express the following integral in the form of xðtÞ pðtÞ; ðpðtÞ is a pulse function:

b: y2 ðtÞ ¼

Z1

uðaÞuða þ tÞda

1 tþT=2 Z

yðtÞ ¼

xðaÞda:

(2:3:26) ¼

tT=2

8 R1 > > uðaÞda ! 1; t 0 > < t

> R1 > > : uða þ tÞda ! 1; t40

(2:3:30) :

0

Solution:

yðtÞ ¼

tþðT=2Þ Z

xðaÞda

1

It follows that y2 ðtÞ ¼ 1; 15t51. In this case, & convolution does not exist.

tðT=2Þ Z

xðaÞda 1

¼ xðtÞ uðt þ ðT=2ÞÞ xðtÞ uðt ðT=2ÞÞ ¼ xðtÞ ½uðt þ ðT=2ÞÞ uðt T=2ÞÞ hti ¼ xðtÞ P : T

2.4 Convolution and Moments (2:3:27)

The output is the convolution of xðtÞ with a pulse width of T with unit amplitude and the process is a & running average. Example 2.3.8 Find the derivative of the running average of the function in (2.3.27) and express the function xðtÞ in terms of the derivative of yðtÞ. Solution: McGillem and Cooper (1991) give an interesting solution for this problem. duðt þ ðT=2ÞÞ duðt ðT=2ÞÞ dt dt T T xðtÞ d t ¼ xðtÞ d t þ 2 2

y0 ðtÞ ¼ xðtÞ

In the examples considered so far, except in the cases of impulses, convolution is found to be a smoothing operation. We like to quantify and compare the results of the convolution of nonimpulse functions to the Gaussian function. In Section 1.7.1, the moments associated with probability density functions were considered. A useful result can be determined by considering the center of gravity convolution in terms of the centers of gravity of the factors in the convolution. First, the moments mn ðxÞ of a waveform xðtÞ and its center of gravity Z are, respectively, defined as

mn ðxÞ ¼

¼ xðt þ ðT=2ÞÞ xðt ðT=2ÞÞ

Example 2.3.9 Derive the expressions a: y1 ðtÞ ¼ uðtÞ uðtÞ; b: y2 ðtÞ ¼ uðtÞ uðtÞ: Solution:

a: y1 ðtÞ ¼ uðtÞ uðtÞ ¼

¼

0

( ð1Þdt ¼

(2:4:1)

R1 Z ¼ 1 R1

txðtÞdt ¼ xðtÞdt

m1 ðxÞ : m0 ðxÞ

(2:4:2)

1

We note that we can define a term like the variance in Section 1.7.1 by uðaÞuðt aÞda

1

Zt

tn xðtÞdt;

1

) xðtÞ ¼ y0ðt ðT=2ÞÞ þ xðt TÞ: (2:3:28) &

Z1

Z1

0; t40 t; t50

(2:3:29)

) ¼ tuðtÞ;

s2 ðxÞ ¼

m2 ðxÞ Z2 : m0 ðxÞ

(2:4:3)

Now consider the expressions for the convolution yðtÞ ¼ gðtÞ hðtÞ. First,

2.4

Convolution and Moments

m1 ðyÞ ¼

Z1

tyðtÞdt ¼

1

¼

Z1

Z1

2 gðlÞ4

1

51

2 4t

1

Z1

3

Z1 1

Signal-to-noise ratio ¼ gðlÞhðt lÞdl5dt

(2:4:8)

3

Example 2.4.1 Verify the result is true in (2.4.7) using the functions

thðt lÞdt5dl:

1

Defining a new variable x ¼ t l on the right and rewriting the above equation results in 2 m1 ðyÞ ¼ 4

Z1

Z1 gðlÞ

1

¼

ðx þ lÞhðxÞdx5dl

Solution: Using integral tables, it can be shown that Z1 Z1 et dt ¼ 1; m1 ðgÞ ¼ tet dt ¼ 1; m0 ðgÞ ¼ 0

Z1 lgðlÞdl

1

Z1

3

gðtÞ ¼ hðtÞ ¼ et and yðtÞ ¼ gðtÞ hðtÞ:

1

Z1

Average signal power : Noise power; s2n

Z1

hðxÞdx þ

1

xhðxÞdx

m2 ðgÞ ¼

Z1

t2 et dt ¼ 2;

0

1

gðlÞdl¼ m1 ðgÞm0 ðhÞ þ m1 ðhÞm0 ðgÞ:

0

(2:4:4)

1

m1 ðgÞ m2 ðgÞ ¼ 1; s2g ¼ Z2g ¼ 1; m0 ðgÞ m0 ðgÞ

Zg ¼

s2h ¼ 1 ðnote gðtÞ ¼ hðtÞÞ;

From the area property, it follows that m0 ðyÞ ¼ m0 ðgÞm0 ðhÞ. The center of gravity is m1 ðyÞ m1 ðgÞ m1 ðhÞ ¼ þ ) Zy ¼ Zg þ Zh : m0 ðyÞ m0 ðgÞ m0 ðhÞ

(2:4:5)

s2y ¼

m2 ðyÞ m1 ðyÞ m0 ðyÞ m0 ðyÞ

¼

s2g

þ

s2h :

t

te dt ¼1; m1 ðyÞ ¼

m2 ðyÞ ¼

0

t2 et dt ¼ 2;

0

Z1

Zy ¼

Z1

t3 et dt ¼ 6;

m1 ðyÞ m2 ðyÞ ¼ 2; s2y ¼ Z2y ¼ 2 ) m0 ðyÞ m0 ðyÞ

s2y ¼ s2g þ s2h ¼ 1 þ 1 ¼ 2:

2 :

(2:4:6)

Using the expressions for m0(y), m1 ðyÞ and m2 ðyÞ and simplifying the integrals results in s2y

m0 ðyÞ ¼

Z1 0

Consider the expression for the squares of the spread of yðtÞ in terms of the squares of the spreads of gðtÞ and hðtÞ. The derivation is rather long and only results are presented.

yðtÞ ¼ gðtÞ hðtÞ ¼ tet uðtÞðsee Example 2:3:4Þ:

(2:4:7)

That is, the variance of y is equal to the sum of the variances of the two factors. It also verifies that convolution is a broadening operation for pulses. Noting that if gðtÞ and hðtÞ are probability density functions then (2.4.7) is valid. In communications theory we are faced with a signal, say gðtÞ is corrupted by a noise nðtÞ with the variance, s2n . The signal-to-noise ratio (SNR) is given by

As an example, consider that we have signal gðtÞ ¼ A cosðo0 tÞ and is corrupted by a noise with a variance equal to s2n . Then, the signal-to-noise ratio is SNR ¼

A2 =2 : s2n

In Chapter 10, we will make use of this in quantization methods, wherein A and SNR are given and determine s2n . This, in turn, provides the information on the size of the error that can be tolerated. Notes: For readers interested in independent random variables, the probability density function of a sum of two independent random variables is the convolution of the density functions of the two factors of the

52

2 Convolution and Correlation

convolution, and the variance of the sum of the two random variables equals the sum of their variances. For a detailed discussion on this, see Peebles (2001).&

Solution: yðtÞ is a triangular function (see Example 2.3.1) given by 1 ht ai yðtÞ ¼ L : (2:4:12) a a

2.4.1 Repeated Convolution and the Central Limit Theorem

The mean values of the two rectangular pulses are a/2 (see Section 1.7). The mean value of yðtÞ is 2ða=2Þ ¼ a. The variance of each of the rectangular pulses is

Convolution operation is an integral operation, which is a smoothing operation. In Example 2.3.1, we have considered the special case of the convolution of two identical rectangular pulses and the convolution of these two pulses resulted in a triangular pulse (see Fig. 2.3.1m). The discontinuities in the functions being convolved are not there in the convolved signal. As more and more pulse functions convolve, the resultant functions become smoother and smoother. Repeated convolution begins to take on the bellshaped Gaussian function. The generalized version of this phenomenon is called the central limit theorem. It is commonly presented in terms of probability density functions. In simple terms, it states that if we convolve N functions and one function does not dominate the others, then the convolution of the N functions approaches a Gaussian function as N ! 1. In the general form of the central limit theorem, the means and variances of the individual functions that are convolved are related to the mean and the variance of the Gaussian function (see Peebles (2001)). Given xi ðtÞ; i ¼ 1; 2; :::; N; the convolution of these functions is yðtÞ ¼ x1 ðtÞ x2 ðtÞ ::: xN ðtÞ:

(2:4:9)

The function yðtÞ can be approximated using ðm0 ÞN , the sum of the individual means of the functions, and s2N the sum of the individual variances by 2 1 2 yðtÞ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eðtðm0 ÞN Þ =2sN : 2ps2N

(2:4:10)

Example 2.4.2 Illustrate the effects of convolution and compare yðtÞ to a Gaussian function by considering the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ; 1 t a=2 ; i ¼ 1; 2: xi ðtÞ ¼ P a a

s2i ¼ m2 m21 ¼ a2 =12; i ¼ 1; 2:

The variance is given by s2y ¼ s21 þ s22 ¼ a2 =6. The Gaussian approximation is 2 1 2 ðyðtÞÞjN¼2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eðtaÞ =ða =3Þ : pða2 =3Þ

(2:4:13b)

This Gaussian and the triangle functions are symmetric around a. They are sketched in Fig. 2.4.1. Even with N ¼ 2, we have a good approximation. &

Fig. 2.4.1 Triangle function yðtÞ in (2.4.12) and the Gaussian function in (2.4.13b)

Example 2.4.3 In Example 2.4.1 we considered two identically exponentially decaying functions: x1 ðtÞ ¼ et uðtÞ ¼ x2 ðtÞ. The convolution of these two functions is given by y2 ðtÞ ¼ tet uðtÞ. Approximate this function using the Gaussian function. Solution: The Gaussian function approximations of yn ðtÞ, considering n ¼ 2 and for n large, are, respectively, given below. Note that m0 ðyÞ ¼ 2. 2 1 y2 ðtÞ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eððt2Þ =2ð2ÞÞ ; 2pð2Þ 2 1 yn ðtÞ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ eððtnÞ =2ðnÞÞ : 2pðnÞ

(2:4:11)

(2:4:13a)

(2:4:14)

For sketches of these functions for various values of & n, see Ambardar (1995).

2.4

Convolution and Moments

53

2.4.2 Deconvolution

xðtÞ ¼ yðtÞ hinv ðtÞ ¼ xðtÞ hðtÞ hinv ðtÞ

In this chapter, we have defined the convolution yðtÞ ¼ hðtÞ xðtÞ as a mathematical operation. If xðtÞ needs to be recovered from yðtÞ, we use a process called the deconvolution defined by

¼ xðtÞ ½hðtÞ hinv ðtÞ; ) hðtÞ hinv ðtÞ ¼ dðtÞ and xðtÞ dðtÞ ¼ xðtÞ: (2:4:15) It is a difficult problem to find hinv ðtÞ, which may not even exist.

Table 2.4.1 Properties of aperiodic convolution Definition: yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼

R1

1

x1 ðaÞx2 ðt aÞda ¼

R1

1

x2 ðaÞx1 ðt aÞda:

Amplitude scaling: ax1 ðtÞ bx2 ðtÞ ¼ abðxðtÞ hðtÞÞ; a and b are constants: Commutative: x1 ðtÞ x2 ðtÞ ¼ x2 ðtÞ x1 ðtÞ: Distributive: x1 ðtÞ ½x2 ðtÞ þ x3 ðtÞ ¼ x1 ðtÞ x2 ðtÞ þ x1 ðtÞ x3 ðtÞ: Associative: x1 ðtÞ ½x2 ðtÞ x3 ðtÞ ¼ ½x1 ðtÞ x2 ðtÞ x3 ðtÞ: Delay: x1 ðt t1 Þ x2 ðt t2 Þ ¼ x1 ðt t2 Þ x2 ðt t1 Þ ¼ yðt ðt1 þ t2 ÞÞ: Impulse response: xðtÞ dðtÞ ¼ xðtÞ: Derivatives: x1 ðtÞ x02 ðtÞ ¼ x01 ðtÞ x2 ðtÞ ¼ y0 ðtÞ;

ðmÞ

ðnÞ

x1 ðtÞ x2 ðtÞ ¼ yðmþnÞ ðtÞ:

Step response: yðtÞ ¼ xðtÞ uðtÞ ¼

Rt

1

xðaÞda;

y0 ðtÞ ¼ xðtÞ dðtÞ ¼ xðtÞ:

Area: A½x1 ðtÞ x2 ðtÞ ¼ A½yðtÞ; where A½xðtÞ ¼

R1

1

xðtÞdt:

Duration: tx 1 þ tx 2 ¼ ty : Symmetry: x1e ðtÞ x2e ðtÞ ¼ ye ðtÞ;

x1e ðtÞ x20 ðtÞ ¼ y0 ðtÞ;

Time scaling: x1 ðctÞ x2 ðctÞ ¼ j1cj yðctÞ; c 6¼ 0:

x10 ðtÞ x20 ðtÞ ¼ ye ðtÞ:

54

2 Convolution and Correlation δ T (t )

2.5 Convolution Involving Periodic and Aperiodic Functions 2.5.1 Convolution of a Periodic Function with an Aperiodic Function

h (t )

(a)

Let hðtÞ be an aperiodic function and xT ðtÞ be a periodic function with a period T. We desire to find the convolution of these two functions. That is, find yðtÞ ¼ xT ðtÞ hðtÞ.

yT ( t )

Example 2.5.1 Derive the expressions for the convolution of the following two functions: dT ðtÞ and hðtÞ assuming T ¼ 1:5 and T ¼ 2 and sketch the results for the two cases. 1 X

dT ðtÞ ¼

(b)

yT ( t )

dðt nTÞ;

hðtÞ ¼ L½t:

(c)

(2:5:1)

k¼1

Derive the expressions for the convolution of these two functions assuming T ¼ 1:5 and T ¼ 2 and sketch the results of the convolution for the two cases.

yðtÞ ¼ hðtÞ dT ðtÞ ¼ hðtÞ

1 X

dðt kTÞ

k¼1

¼

1 X

hðtÞ dðt kTÞ:

(2:5:2)

k¼1

(d) Fig. 2.5.1 (a) Periodic impulse sequence, (b) L½t; (c) yT ðtÞ; T ¼ 2, and (d) yT ðtÞ; T ¼ 2

R1 Solution: yðtÞ ¼ hðtÞ xT ðtÞ ¼ eab cosðo0 ðt bÞþ 0 yÞdb Z1 eab ½cosðo0 t þ yÞ cosðo0 bÞ ¼ 0

Noting that hðtÞ dðt kTÞ ¼ hðt kTÞ, it follows that

þ sinðo0 t þ yÞ sinðo0 bÞdb 2 1 3 Z ¼ 4 eab cosðo0 bÞdb5 cosðo0 t þ yÞ 0

yðtÞ ¼

1 X

2

hðt kTÞ ¼ yT ðtÞ:

(2:5:3)

k¼1

þ4

Z1

3 eab sinðo0 bÞdb5 sinðo0 t þ yÞ:

(2:5:5)

0

Figure 2.5.1a,b gives the sketches of the functions dT ðtÞ and hðtÞ. The sketches for the convolution are shown in Fig. 2.5.1c,d. In the first case, there were no overlaps, whereas in the second case there are & overlaps. Example 2.5.2 Derive an expression for the convolution yðtÞ ¼ hðtÞ xT ðtÞ, xT ðtÞ ¼ cosðo0 t þ yÞ and hðtÞ ¼ eat uðtÞ:

(2:5:4)

Using the identities given below (see (2.5.7 a, b, and c.)), (2.5.5) can be simplified. yðtÞ ¼ a=ða2 þ o20 Þ cosðo0 t þ yÞ þ o0 =ða2 þ o20 Þ sinðo0 t þ yÞ 1 ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ cosðo0 t þ y tan1 ðo0 =aÞÞ a2 þ o20 yT ðtÞ;

(2:5:6)

2.5

Convolution Involving Periodic and Aperiodic Functions

Z1 e 0

ab

The periodic convolution of two periodic functions, xT ðtÞ and hT ðtÞ, is defined by

eab sinðo0 bÞdb ¼ 2 ½a sinðo0 bÞ a þ o20 o0 cosðo0 bÞ1 0 ¼

o0 ; 2 a þ o20 (2:5:7a)

Z1

eab cosðo0 bÞdb ¼

0

eab ½a cosðo0 bÞ þ o20 a ; 2 a þ o20 (2:5:7b)

a cosðo0 t þ yÞ þ b sinðo0 t þ yÞ ¼ c cosðo0 t þ fÞ; qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ c ¼ a2 þ b2 ; f ¼ ½ tan1 ðb=aÞ þ y: (2:5:7c) The functions yðtÞ ¼ yT ðtÞ and xT ðtÞ are sinusoids at the same frequency o0 . The amplitude and the phase & of yðtÞ are different compared to that of xT ðtÞ. The derivation given above can be generalized for a periodic function 1 X xT ðtÞ ¼ Xs ½0 þ c½k cosðko0 t þ y½kÞ; o0 ¼ 2p=T; k¼0

(2:5:8a) yðtÞ ¼ xT ðtÞ hðtÞ ¼ Xs ½0 hðtÞ þ

1 X

c½k½cosðko0 t

k¼0

þ y½kÞ hðtÞ; o0 ¼ 2p=T:

tZ0 þT 1 xT ðaÞhT ðtaÞda yT ðtÞ ¼ xT ðtÞ hT ðtÞ ¼ T t0 Z Z 1 1 ¼ xT ðaÞhT ðtaÞda ¼ xT ðtaÞhT ðaÞda T T T

a2

þ o0 sinðo0 bÞ1 0 ¼

55

hT ðt þ T aÞ ¼ hT ðt aÞ and xT ðt þ T aÞ ¼ xT ðt aÞ:

(2:5:9b)

Also, periodic convolution is commutative. Many of the aperiodic convolution properties discussed earlier are applicable for periodic convolution with some modifications. The expression for the periodic convolution can be obtained by considering aperiodic convolution for one period of each of the two functions. Consider the periodic functions in the form 1 X

xT ðtÞ ¼

xðt nTÞ and hT ðtÞ ¼

n¼1

1 X

hðt nTÞ;

n¼1

(2:5:10a)

xT ðtÞ; t0 t < t0 þ T

0; otherwise hT ðtÞ; t0 t < t0 þ T

xðtÞ ¼ hðtÞ ¼

In Section 1.5 energy and power signals were considered. The energy in a periodic function is infinity and its average power is finite. One period of a periodic function has all its information. In the same vein, the average convolution is a useful measure of periodic convolution. Such averaging process is called periodic or cyclic convolution. The convolution of two periodic functions with different periods is very difficult and is limited here to the convolution of two periodic functions, each with the same period.

(2:5:9a)

Note that the symbol used for the periodic convolution and the constant ðTÞ in the denominator in (2.5.9a) indicates that it is an average periodic convolution. yT ðtÞ is periodic since

(2:5:8b)

2.5.2 Convolution of Two Periodic Functions

T

¼ hT ðtÞ xT ðtÞ:

;

0; otherwise:

(2:5:10b)

Note that the time-limited functions, xðtÞ and hðtÞ, are defined from the periodic functions xT ðtÞ and hT ðtÞ. Using (2.5.10b) the periodic convolution is 1 yT ðtÞ ¼ T

Z

xT ðaÞhT ðt aÞda

T

¼

1 T

Z T

xT ðaÞ

1 X n¼1

hðt a nTÞda

56

2 Convolution and Correlation

¼

1 Z 1 X xðaÞhðt a nTÞda T n¼1

Convolution of almost periodic or random signals, xðtÞ and hðtÞ, is defined by

T

1 1 X xðtÞ hðt nTÞ; ¼ T n¼1

(2:5:11a)

ZT=2

1 yðtÞ ¼ lim T!1 T

xðaÞhðt aÞda:

(2:5:14)

T=2

yT ðtÞ ¼ xT ðtÞ hT ðtÞ 1 1 X yðt nTÞ; yðtÞ ¼ xðtÞ hðtÞ: (2:5:11b) ¼ T n¼1

This reduces to the periodic convolution if xðtÞ and hðtÞ are periodic with the same period.

That is, yT ðtÞ can be determined by considering one period of each of the two functions and finding the aperiodic convolution.

2.6 Correlation

Example 2.5.3 a. Determine and sketch the aperiodic convolution yðtÞ ¼ hðtÞ xðtÞ. 1 t1 1 t 1:5 ; hðtÞ ¼ P : (2:5:12) xðtÞ ¼ P 2 2 3 3

Equation (2.1.3) gives the cross-correlation of xðtÞ and hðtÞ as the integral of the product of two functions, one displaced by the other by t between the interval a < t < b and is given by

Zb b. Determine and sketch the periodic convolution Rxh ðtÞ¼xðtÞhðtÞ¼ xðtÞhðtþtÞdt¼ hxðtÞhðtþti: yT ðtÞ ¼ xT ðtÞ hT ðtÞ for periods T ¼ 6 and 4: a 1 1 X X xðt kTÞ and hT ðtÞ ¼ hðt kTÞ: xT ðtÞ ¼ Cross-correlation function gives the similarity k¼1 k¼1 between the two functions: xðtÞ and hðt þ tÞ. Many (2:5:13) a times the second function hðtÞ may be a corrupted Solution: a. From (2.5.13), the results for the aper- version of xðtÞ, such as hðtÞ ¼ xðtÞ þ nðtÞ, where iodic convolution can be derived. The sketches of nðtÞ is a noise signal. In the case of xðtÞ ¼ hðtÞ, the two functions and the result of the convolution cross-correlation reduces to autocorrelation. In are shown in Fig. 2.5.2a. The periodic convolutions this case, at t ¼ 0, the autocorrelation integral for the two different periods are shown in Fig. gives the highest value at t ¼ 0. Comparison of 2.5.2b,c. There are no overlaps of the functions two functions appears in many identification situafrom one period to the next in Fig. 2.5.2b, whereas tions. For example, to identify an individual based & upon his speech pattern, we can store his speech in Fig. 2.5.2c, the pulses overlap.

(a)

yT (t ) Fig. 2.5.2 Example 2.5.1 (a) Aperiodic convolution; (b) periodic convolution T ¼ 6; (c) periodic convolution, T ¼ 4

(b)

yT (t )

(c)

2.6

Correlation

57

segment in a computer. When he enters, say a secure area, we can request him to speak and compute the cross-correlation between the stored and the recorded. Then decide on the individual’s identification based on the cross-correlation function. Generally, an individual is identified if the peak of the cross-correlation is close to the possible peak autocorrelation value. Allowance is necessary since the speech is a function of the individual’s physical and mental status of the day the test is made. Quantitative measures on the cross-correlation will be considered a bit later. The order of the subscripts on the cross-correlation function Rxh ðtÞ is important and will get to it shortly. In the case of xðtÞ ¼ hðtÞ, we have the autocorrelation and the function is referred to as Rx ðtÞ with a single subscript. The cross- and autocorrelation functions are functions of t and not t. Correlation is applicable to periodic, aperiodic, and random signals. In the case of periodic functions, we assume that both are periodic with the same period.

xðtÞhðt þ tÞdt

(2:6:1e)

For periodic functions, (2.6.1e) reduces to (2.6.1b).

2.6.1 Basic Properties of Cross-Correlation Functions Folding relationship between the two cross-correlation functions is Rxh ðtÞ ¼ Rhx ðtÞ; ) Rxh ðtÞ ¼

¼ Z1

xðtÞhðt þ tÞdt:

T=2

Cross-correlation: Aperiodic: Rxh ðtÞ ¼

ZT=2

1 Ra;xh ðtÞ ¼ lim T!1 T

Z1 1 Z1

(2:6:2)

xðtÞhðt þ tÞdt

xða tÞhðaÞda¼ Rhx ðtÞ: (2:6:3)

1

(2:6:1a)

1

2.6.2 Cross-Correlation and Convolution

Cross-correlation: Periodic: RT; xhðtÞ ¼

1 T

Z

xT ðtÞhT ðt þ tÞdt

(2:6:1b)

T

Autocorrelation: Aperiodic: Rx ðtÞ ¼

Z1

xðtÞxðt þ tÞdt

(2:6:1c)

1

The cross-correlation function is related to the convolution. From (2.6.3) we have Rxh ðtÞ ¼ xðtÞ hðtÞ ¼ xðtÞ hðtÞ;

(2:6:4a)

Rhx ðtÞ ¼ hðtÞ xðtÞ ¼ hðtÞ xðtÞ:

(2:6:4b)

Equation (2.6.4a) can be seen by first rewriting the first integral in (2.6.3) using a new variable t ¼ a, and then simplifying it. That is,

Autocorrelation: Periodic: 1 RT;x ðtÞ ¼ T

Z

xT ðtÞxT ðt þ tÞdt:

(2:6:1d)

T

Rxh ðtÞ ¼

Z1

xðtÞhðt þ tÞdt ¼

1

¼ xðtÞ hðtÞ: Notes: Cross- and autocorrelations of periodic functions and random signals are referred to as average periodic cross- and autocorrelation functions. In the case of random or noise signals, the average cross-correlation function is defined by

Z1

xðaÞhðt aÞda

1

(2:6:4c)

Equation (2.6.4b) can be similarly shown. Noting the explicit relation between correlation and convolution, many of the convolution properties are applicable to the correlation. To compute the cross–

58

2 Convolution and Correlation

correlation, Rxh ðtÞ, one can use either of the integral in (2.6.3) or the integral in (2.6.4c). Rxh ðtÞ is not always equal to Rhx ðtÞ. In case, if one of the functions is symmetric, say xðtÞ ¼ xðtÞ; then Rxh ðtÞ ¼ xðtÞ hðtÞ ¼ xðtÞ hðtÞ:

Rg ð0Þ ¼

Z1

g2 ðtÞdt ¼ Eg ; Rh ð0Þ ¼

1

1

Consider the integral Z1 Z1 2 ½xðtÞ hðt þ tÞ dt ¼ x2 ðtÞdt

þ

Z1

1 Z1

h2 ðt þ tÞdt 2

1

1

jRxh ðtÞj

(2:6:7a)

jRxh ðtÞj ðRx ð0Þ þ Rh ð0ÞÞ=2:

@

32 xðtÞhðt þ tÞdt5

1

Z1 1

10

jxðtÞj [email protected] 2

Z1 1

1 jhðt þ tÞj dtA; 2

1

(2:6:9a) pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Rx ð0ÞRh ð0Þ:

(2:6:9b)

(2:6:9c)

Write the resulting equation in a quadratic form in terms of a. In order for the equation in (2.6.10) to be true, the roots of the quadratic equation have to be real and equal or the roots have to be complex conjugates. The proof is left as a homework problem.

(2:6:8)

h t i ; hðtÞ ¼ eat uðtÞ; a > 0: 2T

(2:6:11a)

Solution: Example 2.3.3 dealt with computing the convolution of these two functions. The cross-correlation functions are as follows: Rhx ðtÞ ¼

Rxh ðtÞ ¼

0

h2 ðtÞdtA

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Rx ð0ÞRh ð0Þ ðRx ð0Þ þ Rh ð0ÞÞ=2:

(2:6:7b)

An interesting bound can be derived using the Schwarz’s inequality. See (2.1.9d).

hxðtÞhðt þ ti ¼ 4

1

Equation (2.6.9b) represents a tighter bound compared to the one in (2.6.7b), as the geometric mean cannot exceed the arithmetic mean. That is,

xðtÞhðt þ tÞdt

This follows since the integrand in (2.6.7a) is nonnegative and

Z1

x2 ðtÞ[email protected]

Z1

¼ Rx ð0ÞRh ð0Þ;

xðtÞ ¼ P

¼ Rx ð0Þ þ Rh ð0Þ 2Rxh ðtÞ 0:

2

10

Example 2.6.1 Determine the cross-correlation of the functions given in Fig. 2.3.2.

1

2

Z1

Another way to prove (2.6.9b) is as follows. Start with the inequality below. Expand the function and 2 h ðtÞdt¼ Eh : identify the auto- and cross-correlation terms. Z 1 ½xðtÞ þ ahðt þ tÞ2 dt 0: (2:6:10) (2:6:6) & 1

2.6.3 Bounds on the Cross-Correlation Functions

1

)jRxh ðtÞj2 @

(2:6:5)

Example 2.6.1 illustrates the use of this property. In particular, the area and duration properties for convolution also apply to the correlation. We should note that the correlations are functions of t and not t, where t is the time shift between xðtÞ and hðt þ tÞ. In the case of energy signals, the energies in the real signals, gðtÞ and hðtÞ, are Z1

0

Z1 1 Z1

hðtÞxðt þ tÞdt ¼ hðtÞ xðtÞ;

xðtÞhðt þ tÞdt ¼ xðtÞ hðtÞ:

1

(2:6:11b) Note that we have xðtÞ ¼ xðtÞ, and therefore the cross-correlation Rxh ðtÞ ¼ xðtÞ hðtÞ is the convolution determined before (see (2.3.18).), except the cross-correlation is a function of t rather than t. It is given below. The two cross-correlation functions are sketched in Fig. 2.6.1a,b. Note Rhx ðtÞ ¼ Rxh ðtÞ

2.6

Correlation

59

Fig. 2.6.1 Crosscorrelations (a)Rxh ðtÞ, (b) Rhx ðtÞðR xh ðTÞ ¼ 1 2aT ¼ Rhx ðTÞ) a 1e

8 0; t T > > > i > > > 1 > : eaT eaT eat ; t > T a (2:6:11c) &

2.6.4 Quantitative Measures of Cross-Correlation

The significance of rxh ðtÞ can be seen by considering some extreme cases. When xðtÞ ¼ ahðtÞ; a > 0, we have the correlation coefficient rxh ðtÞ ¼ 1. In the case of xðtÞ ¼ ahðtÞ; a < 0 and rxh ðtÞ ¼ 1. In communication theory, we will be interested in signals that are corrupted by noise, usually identified by nðtÞ, which can be defined only in statistical terms. In the following, we will consider the analysis without going through statistical analysis. Noise signal nðtÞ is assumed to have a zero average value. That is,

The amplitudes of Rxh ðtÞ ðand Rhx ðtÞÞ vary. It is appropriate to consider the normalized correlation coefficient (or correlation coefficient) of two energy signals defined by Rxh ðtÞ Rxh ðtÞ ﬃ; rxh ðtÞ ¼ sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 1 ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Ex Eh R R 2 2 x ðtÞdt h ðtÞdt 1

1

(2:6:12a) ) jrxh ðtÞj 1:

nðtÞdt ¼ 0:

(2:6:14)

T=2

Cross-correlation function can be used to compare two signals. The signals xðtÞ and hðtÞ are uncorrelated if the average cross-correlation satisfies the relation

(2:6:12b)

Equation (2.6.12b) can be shown as follows. From (2.1.13a) and using the Schwarz’s inequality (see (2.1.9d)), we have

ZT=2

1 lim T!1 T

1 Ra;xh ðtÞ ¼ lim T!1 T 2

ZT=2 T=2

1 6 ¼ 4 lim T!1 T

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ Rxh ðtÞ ¼ hxðtÞhðt þ tÞi kxðtÞkkhðt þ tÞk ¼ Ex Eh

xðtÞhðt þ tÞdt

ZT=2 T=2

32 1 76 xðtÞdt54 lim T!1 T

ZT=2

3 7 hðtÞdt5:

T=2

(2:6:15) It should be noted that the case of xðtÞ ¼ hðtÞ, the correlation coefficient reduces to rxx ðtÞ ¼

Rx ðtÞ : Rx ð0Þ

Example 2.6.2 If the signals x(t) and a zero average noise signal n(t) are uncorrelated, then show

(2:6:13)

Correlation measures are very useful in statistical analysis. See Yates and Goodman (1999), Cooper and McGillem (1999) and others.

1 lim T!1 T

ZT=2 T=2

xðtÞnðt tÞdt ¼ 0 for all t:

(2:6:16)

60

2 Convolution and Correlation

Fig. 2.6.2 Correlation detector

Solution: Using (2.6.14) and (2.6.15), we have lim

1

ZT=2

T!1 T

2

xðtÞnðt þ tÞdt

T=2

1 6 ¼ 4 lim T!1 T

ZT=2

32 1 76 xðtÞdt54 lim T!1 T

T=2

ZT=2

3

correlation detector (or receiver) shown in Fig. 2.6.2. The received signals yi ðtÞ are assumed to be of the form in (2.6.19). Decide which signal has been transmitted using the cross-correlation function.

7 nðt þ tÞdt5 ¼ 0:

T=2

(2:6:17) Cross-correlation function can be used to estimate the delay caused by a system. Suppose we know that a finite duration signal xðtÞ is passed through an ideal transmission line resulting in the output function yðtÞ ¼ xðt t0 Þ. The delay t0 caused by the transmission line is unknown and can be estimated using the cross-correlation function Rxy ðtÞ: At t ¼ t0 , Rxy ðt0 Þ gives a maximum value. Then determine t corresponding to the maximum value of & Rxy ðtÞ.

yi ðtÞ ¼ xi ðtÞ þ noise; i ¼ 1 or 2:

Solution: Let the transmitted signal be x1 ðtÞ. Using the top path in Fig. 2.6.2, we have ZTs

½x1 ðtÞ þ nðtÞx1 ðtÞdt ¼ A

0

ZTs

xi ðtÞxj ðtÞdt ¼

Exi ¼ Ex ; i ¼ j 0;

i 6¼ j; i ¼ 1; 2

:

(2:6:18)

0

Ex is the energy contained in each signal. The two signals to be transmitted are assumed to be available at the receiver. A simple receiver is the binary

ZTs

x21 ðtÞdt:

(2:6:20)

0

Using the bottom path, with the transmitted signal equal to x1 ðtÞ, we have ZTs

½x1 ðtÞþnðtÞx2 ðtÞdt¼

0

Example 2.6.3 Consider the transmitted signals x1 ðtÞ and x2 ðtÞ in the interval 0 < t < Ts and zero otherwise. Use the cross-correlation function to determine which signal was transmitted out of the two. They are assumed to be mutually orthogonal (see Section 2.1.1) over the interval and satisfy

(2:6:19)

ZTs

½x1 ðtÞx2 ðtÞþx1 ðtÞnðtÞdt

0

¼

ZTs

x1 ðtÞnðtÞdt¼B:

(2:6:21)

0

Since the noise signal has no relation to x1 ðtÞ, B will be near zero and A 44B, implying x1 ðtÞ was transmitted. If x2 ðtÞ was transmitted, the roles are reversed and B A. The correlation method of detection is based on the following: 1. If A > B ) transmitted signal is x1 ðtÞ. 2. If B > A ) transmitted signal is x2 ðtÞ. 3. If B ¼ A ) no decision can be made as noise & swamped the transmitted signal.

2.6

Correlation

61

Example 2.6.4 Derive the expressions for the crossand Rhx ðtÞ assuming correlation Rxh ðtÞ xðtÞ ¼ et uðtÞ; hðtÞ ¼ e2t uðtÞ:

xðtÞ ¼ P½t :5; hðtÞ ¼ tP ¼

Solution: Using the expression in (2.6.3), we have Rxh ðtÞ ¼

¼

Z1 1 Z1

xða tÞhðaÞda

eðatÞ e2a ½uða tÞuðaÞda: (2:6:22a)

Consider the following and then the corresponding correlations: ½uðaÞuða tÞ ¼ t 1 or t 1:

;

Rxh ðtÞ ¼

1 3a 1 1 2t e uðtÞ; (2:6:22b) t ¼ e 3 3 Z1

(2:6:24)

Case 2: 0 < t 1 or 1 < t 0 : See Fig. 2.6.4e. Using Table 2.6.1

e3a da

t

¼ et

Solution: See Fig. 2.6.4c for hðt þ tÞ for an arbitrary t. The function hðt þ tÞ starts at t ¼ t and ends at t ¼ 2 t. As t varies from 1 to 1, there are five possible regions we need to consider. These are sketched in Fig. 2.6.4 d,e,f,g,h. In each of these cases both the functions are sketched in the same figure, which allows us to find the regions of overlap. The regions of overlap are listed in Table 2.6.1.

Z1

ðt þ tÞdt ¼

t2 þ ttt¼1 t¼t 2

t

¼

ðt þ 1Þ2 ; 1 t < 0: 2

(2:6:25)

1 t 3a 1 et : ee 0 ¼ 3 Case 3: 0 < t 1: See Fig. 2.6.4 f. Using Table ð3Þ 0 2.6.1, we have (2:6:22c) Z1 t2 1 þ 2t Rxh ðtÞ is shown in Fig. 2.6.3. Note that Rxh ðtÞ ¼ ðt þ tÞdt ¼ þ ttt¼1 ; 0 < t 1: t¼0 ¼ 2 2 & Rhx ðtÞ ¼ Rxh ðtÞ. 0 t < 0 : Rxh ðtÞ ¼ et

e3a da ¼

(2:6:26)

Rxh (τ )

Case 4: 1 < t 2: See Fig. 2.6.3 g. Using Table 2.6.1 we have

Rxh ðtÞ ¼

Z2t

ðt þ tÞdt ¼

t2 þ ttt¼2t t¼0 2

0

Fig. 2.6.3 Rxh ðtÞ

Example 2.6.5 Derive the cross–correlation Rxh ðtÞ for the following functions:

4 t2 ; 1 < t 2: ¼ 2

(2:6:27)

Case 5: 2 < t: See Fig. 2.6.4 h. There is no overlap and

62 Fig. 2.6.4 (a) xðtÞ, (b) hðtÞ; (c) hðt þ tÞ, (d) xðtÞ and hðt þ tÞ; t > 1ðor t 1Þ; (e) xðtÞandhðtþtÞ;1 2, (i) Rxh ðtÞ

2 Convolution and Correlation

(a)

(b)

(c)

(d)

(e)

(f)

(h)

(g)

Rxh (τ )

τ (i)

Table 2.6.1 Example 2.6.4 Case t Range of overlap/ integration range 1 t 1 No over lap 2 1 < t 0 t < t < 1 3 0 0 : Rx ðtÞ ¼

¼

½xðtÞ xðt tÞ½xðtÞ xðt tÞdt 0: (2:7:4)

Z1 1 Z1

1;

t>t

0;

otherwise

; (2:7:9)

xðtÞxðt tÞdt

eat uðtÞeaðttÞ uðt tÞdt

1

2

x ðtÞdt þ

1

Z1

2

x ðt tÞdt 2

1

¼ 2½

Z1

Z1

xðtÞxðt tÞdt

Z1

x ðtÞdt

1

Z1

¼

Rx ðtÞ ¼ ð1=2aÞeajtj :

1

xðtÞxðt tÞdt:

(2:7:10)

The energy contained in the exponentially decaying pulse is E ¼ Rx ð0Þ ¼ ð1=2aÞ. The autocorrelation & function is sketched in Fig. 2.7.1. (2:7:5)

Third,

1

eat : 2a

Using the symmetry property of the AC, we have xðtÞxðt tÞdt 0:

x2 ðtÞdt jRx ðtÞj

1 1 Z

Z1

e2 at dt ¼

t

1

) Rx ð0Þ ¼

Z1

1

2

Ex ¼ Rx ð0Þ ¼

¼e

at

x2 ðtÞdt:ðenergy in xðtÞÞ: (2:7:6)

Example 2.7.2 Consider the function xðtÞ ¼ P½t 1=2. Determine its autocorrelation function and its energy using this function. Solution: The AC function for t 0 is Z1 Z1 Rx ðtÞ¼ xðtÞxðttÞdt¼ P½t:5P½tt:5dt: 1

1

64

2 Convolution and Correlation

AC function is much easier to compute using this property. The AC of the pulse function P½t :5 can be computed by ignoring the delay. That is, ACfP½t :5g ¼ ACfP½tg. Interestingly, n h t io hti AC P ¼ TL : T T Fig. 2.7.1 Example 2.7.1

The function P½t 1=2 is a rectangular pulse centered at t ¼ 1=2 with a width of 1, and P½t ðt þ ð1=2ÞÞ is a rectangular pulse centered at ðt þ 0:5Þ with a width of 1. See Fig. (2.7.2a) for the case 0 < t < 1. In the case of t 1, there is no overlap indicating that Rx ðtÞ ¼ 0; t 1.

Rx ðtÞ ¼

Z1

The AC function of a rectangular pulse of width T is a triangular pulse of width 2T and its amplitude at t ¼ 0 is T: We can verify the last part by noting n h t io hti ¼ T: AC P jt¼0 ¼ TL j T T t¼0 Note xðtÞP Tt extracts xðtÞ for the time T=2 < t < T=2. That is,

dt ¼ ð1 tÞ; 0 t < 1:

t

xðtÞP Using the symmetry property, we have Rx ðtÞ ¼ Rx ðtÞ ¼

(2:7:11b)

ð1 jtjÞ; 0 jtj 1 0; Otherwise

¼ L½t: (2:7:11a)

hti T

¼

xðtÞ; T=2 < t < T=2 : 0; otherwise

(2:7:12)

Example 2.7.3 Find the autocorrelation of the function yðtÞ ¼ cosðo0 tÞP½t=T. Solution:

This is sketched in Fig. 2.7.2b indicating that there is correlation for jtj < 1 and no correlation for jtj 1. The peak value of the autocorrelation is when t ¼ 0 and is R x ð0Þ ¼ 1. The energy contained in the unit rectangular pulse is equal to 1 and by using the autocorrelation function, i.e., Rx ð0Þ ¼ 1, the same by both the methods. Noting that the autocorrelation function of a given function and its delayed or advanced version are the same, the

Ry ðtÞ¼

Z1 P

h t i htti P cosðo0 tÞcosðo0 ðttÞÞdt T T

1

Z1

cosðo0 tÞ ¼ 2

P

h t i htti P dt T T

1

þ

1 2

Z1 P

h t i htti P cosð2o0 ttÞdt: T T

1

Rx(τ)

(a) Fig. 2.7.2 Example 2.7.2 Autocorrelation of a rectangular pulse

(b)

2.8

Cross- and Autocorrelation of Periodic Functions

¼

ð1=2ÞTL Tt cosðo0 tÞ þ B; jtj T : 0; jtj > T

2.8 Cross- and Autocorrelation (2:7:13) of Periodic Functions

Now consider the evaluation of B. For t 0, Z1

1 B¼ 2

P

65

h t i ht ti P cosð2o0 t tÞdt T T

The cross- and the autocorrelation functions of periodic functions of xT ðtÞ and hT ðtÞ are Z 1 xT ðtÞhT ðt þ tÞdt ¼ hxT ðtÞhT ðt þ tÞi ; RT; xhðtÞ ¼ T T

1

(2:8:1a) ¼

1 2

ZT=2

cosð2o0 t o0 tÞdt

1 RT;x ðtÞ ¼ T

t

¼

1 ½sinðo0 T o0 tÞ sinðo0 tÞ: 4o0

Z

xT ðtÞxT ðt þ tÞdt ¼ hxT ðtÞxT ðt þ tÞi :

T

(2:7:14)

(2:8:1b)

If o0 is large, Ry ðtÞ in (2.7.13) can be approximated by the first term and

Note that the periods of the functions, xT ðtÞ and hT ðtÞ, are assumed to be the same and the constant (1/T) before the integrals in (2.8.1a and b). If they have different periods, computation of (2.8.1a) is difficult and these cases will not be discussed here. Many of the cross-correlation and AC function properties derived earlier for the aperiodic case apply for the periodic functions with some modifications. Note that

hti 1 RY ðtÞ ’ TL cosðo0 tÞ: 2 T

(2:7:15)

The envelope of the autocorrelation function in (2.7.16) is a triangular function, which follows since the correlation of the two identical rectangular functions is a triangular function. Noting that the cosine function oscillates between 1, the envelope of the autocorrelation function in (2.7.15) is shown & in Fig. 2.7.3. Ry(τ)

Fig. 2.7.3 Sketch of Ry(t)

RT;xh ðtÞ ¼ RT;hx ðtÞ;

RT;x ð0Þ þ RT;h ð0Þ 2RT;xh ðtÞ:

In Section 2.5.1, aperiodic convolution was used to find periodic convolution. The same type of analysis can be used to determine periodic cross-correlations using aperiodic cross-correlations. Furthermore, as discussed before, correlation is related to convolution. First define two finite duration functions, xðtÞ and hðtÞ, over the interval t0 t < t0 þ T. Assume that they are zero outside this interval. Now create two periodic functions:

xT ðtÞ ¼

1 X n¼1

Notes: Conditions for the existence of an aperiodic autocorrelation are similar to those of convolution (see Section 2.2.3). But there are a few exceptions. For example, the autocorrelation of the unit step function does not exist.

(2:8:2)

xðt nTÞ;

hT ðtÞ ¼

1 X

hðt nTÞ:

n¼1

(2:8:3a) The periodic cross-correlation function is defined by

66

2 Convolution and Correlation

1 RT;xh ðtÞ ¼ T ¼

tZ0 þT

t0 1 X

Solution: 1 a: RT;xT;1 ðtÞ ¼ T

xT ðtÞhT ðt þ tÞdt; hT ðt þ tÞ

Z

X2s ½0dt ¼ X2s ½0;

(2:8:5a)

T

hðt þ t nTÞ:

(2:8:3b)

n¼1

The expression for periodic convolution is given in terms of aperiodic convolution and

1 RT;xT;2 ðtÞ ¼ T

Z T

¼

c2 ½k 2T

xT;2 ðtÞxT;2 ðt þ tÞdt Z cosðko0 tÞdt T

1 1 X RT;xh ðtÞ ¼ Rxh ðt nTÞ; T n¼1

Rxh ðt nTÞ ¼

tZ0 þT

xðtÞhðt þ t nTÞ:

c2 ½k þ 2T

Z

cosðko0 ð2t þ tÞ þ 2y½kÞdt

T

(2:8:3c)

t0

¼

c2 ½k cosðko0 tÞ 2T

ZT

dt ¼

c2 ½k cosðko0 tÞ : 2

0

(2:8:5b) The details of the derivation are left as an exercise. Copies of Rxh ðtÞ will overlap if the width of Rxh ðtÞ is wider than T. Example 2.8.1 Give the lower bound on the period T so that there are no overlaps in the cross-correlation of the functions xT ðtÞ and hT ðtÞ given below. See Example 2.6.5.

t1 ; xðtÞ ¼ P½t :5; hðtÞ ¼ tP 2 1 1 X X xðt þ nTÞ; hT ðtÞ ¼ hðt þ nTÞ: xT ðtÞ ¼ n¼1

n¼1

Note that the integral of a cosine function over any integer number of periods is zero. b. The cross-correlation of a constant and a cosine function over one period is zero. Also note that the two functions are orthogonal. That is & hxT1 ðtÞ; xT2 ðtÞi ¼ 0. Example 2.8.3 Find the AC of xT ðtÞ given below with k 6¼ m; kandmare integers. xT ðtÞ ¼ xT;1 ðtÞ þ xT;2 ðtÞ; xT;1 ðtÞ ¼ c½k cosðko0 t þ y½kÞ; xT;2 ðtÞ ¼ c½m cosðmo0 t þ y½mÞ:

Solution: If the period T is larger than 3, then there are no overlaps in the periodic cross-correlation function. In that case, one period of the cross-correlation function can be obtained from the aperiodic cross-correlation in that example and dividing it by the period T. If the period is less than 3, then & there will be overlaps. Example 2.8.2 Consider the periodic functions xT;1 ðtÞ ¼ Xs ½0;

xT;2 ðtÞ ¼ c½k cosðko0 t þ y½kÞ: (2:8:4)

a. Find the AC functions for the functions in (2.8.4). b. Find the cross-correlation of the two functions.

Solution: The periodic autocorrelations are determined as follows: Z 1 ½xT;1 ðtÞ þ xT;2 ðtÞ½xT;1 ðt þ tÞ RT;x ðtÞ ¼ T T Z 1 þ xT;2 ðt þ tÞdt ¼ xT;1 ðtÞxT;1 ðt þ tÞdt T T Z 1 þ xT;2 ðtÞxT;2 ðt þ tÞdt T T Z 1 þ xT;1 ðtÞxT;2 ðt þ tÞdt T T Z 1 þ xT;2 ðtÞxT;1 ðt þ tÞdt: (2:8:6) T T

2.8

Cross- and Autocorrelation of Periodic Functions

67

Note Z

1 T

xT;1 ðtÞxT;2 ðt þ tÞdt

T

¼

P ¼ X2s ½0 þ

1 2T

Z

1 þy½mÞdt þ 2T

Z

Variance ¼ h½kh½m cos½ðk mÞo0 t mo0 tÞ

T

þðy½k y½mÞdt ¼ 0: Similarly the fourth term in (2.8.6) goes to zero. From the last example, RT;x ðtÞ ¼

(2:8:10)

The difference between the total power and the dc power is the variance and is given by

c½kc½m cos½ðk þ mÞo0 t þ mo0 t þ y½k

T

1 1X c2 ½k: 2 k¼1

1 1X c2 ½k: 2 k¼1

(2:8:11) &

Example 2.8.4 Consider the corrupted signal yðtÞ ¼ xðtÞ þ nðtÞ, where nðtÞ is assumed to be noise. Assuming the signal xðtÞ and noise nðtÞ are uncorrelated, derive an expression for the autocorrelation function of yðtÞ.

Solution: c2 ½k c2 ½m cosðko0 tÞ þ cosðmo0 tÞ; k 6¼ m: 2 2 ZT=2 (2:8:7) & 1 Ryy ðtÞ ¼ lim yðtÞyðt þ tÞdt T!1 T T=2

These results can be generalized using the last two examples and the autocorrelation of a periodic function xT ðtÞ is given as follows: xT ðtÞ ¼ Xs ½0 þ

1 X

c½k cosðko0 t þ y½kÞ;

(2:8:8)

1

¼ lim

T!1 T

½xðtÞ þ nðtÞ½xðt þ tÞnðt þ tÞdt;

T=2

ZT=2

¼ lim

T!1

k¼1

) RT;x ðtÞ ¼ X2s ½0 þ

ZT=2

xðtÞxðt þ tÞdt þ lim

T!1

T=2 1 1X c2 ½kcosðko0 tÞ;o0 ¼ 2p=T: 2 k¼1

þ lim

T!1

T=2 R T=2

Notes: The AC function of a constant Xs ½0 is X2s ½0. The AC of the sinusoid c½k cosðko0 t þ y½kÞ is ðc2 ½k=2Þ cosðko0 tÞ. That is, it loses the phase information in the function in the sinusoid. The power contained in the periodic function xT ðtÞ in (2.8.8) can be computed from the autocorrelation function evaluated at t ¼ 0. That is,

xðtÞnðt þ tÞdt

T=2

nðtÞxðt þ tÞdt þ lim

T!1

T=2 R

xðtÞxðt þ tÞdt:

T=2

(2:8:12)

(2:8:9) AC function of a periodic function is also a periodic function with the same period. It is independent of y½k. It does not have the phase information contained in (2.8.8). In the next chapter, (2.8.8) will be derived for an arbitrary periodic function and will be referred to as the harmonic form of Fourier series of a periodic & function xT ðtÞ.

ZT=2

Noting that the signal and the noise are uncorrelated, i.e., Rxn ðtÞ ¼ Rnx ðtÞ ¼ 0, we have Ryy ðtÞ ¼ Rxx ðtÞ þ Rnn ðtÞ:

(2:8:13)

&

The average power contained in the signal and the noise is given by Py ¼ Px þ Pn ¼ Rx ð0Þ þ Rn ð0Þ ¼ Rx ð0Þ þ s2n : (2:8:14) The signal-to-noise ratio (SNR), Px =Pn , can be computed. It is normally identified in terms of decibels. See Section 1.9.

68

2 Convolution and Correlation

2.2.3 Use the area property of convolution to find the integrals of yðtÞ in Problem 2.2.2.

2.9 Summary We have introduced the basics associated with the two important signal analysis concepts: convolution and correlation. Specific principal topics that were included are

Convolution integral: its computations and its properties Moments associated with functions Central limit theorem Periodic convolutions Auto- and cross-correlations Examples of correlations involving noise without going into probability theory

Quantitative measures of cross-correlation functions and the correlation coefficient

Auto- and cross-correlation functions of energy and periodic signals

Signal-to-noise ratios

2.3.1 a. Derive the expression for the convolution of two pulse functions given by xðtÞ ¼ P½t 1 and h½t ¼ P½t 2. Compute this directly first and then verify your result by using the delay property of convolution. b. Verify the time duration property of the convolution using the above problems. 2.3.2 Determine the area of yðtÞ in (2.3.18) using the area property of the convolution. 2.4.1 Approximate the function yðtÞ in Example 2.3.1 using the Gaussian function. 2.4.2 Use the derivative property of the convolution to derive the convolution of the two functions given below using the results in Example 2.5.2. xT ðtÞ ¼ sinðo0 tÞ; hðtÞ ¼ eat uðtÞ; a > 0: 2.4.3 Use the delay property of the convolution to determine

Problems

xðtÞ ¼ eat uðtÞ uðt 1Þ: 2.1.1 Consider the following functions defined over 0 < t < 1. Using (2.1.3), identify the two functions that give the maximum cross-correlation at t ¼ 0. x1 ðtÞ ¼ et ; x2 ðtÞ ¼ sinðtÞ; x3 ðtÞ ¼ ð1=tÞ: 2.2.1 Prove the commutative, distributive, and the associate properties of the convolution. 2.2.2 Find the convolution yðtÞ ¼ hðtÞ xðtÞ for the following functions:

2.5.1 Derive the expressions for the periodic convolution of the two periodic functions 1 1 X X t nT xðtÞ ¼ dðt nTÞ; hðtÞ ¼ P : T=2 n¼1 n¼1 2.6.1 Find the cross-correlation of the functions xðtÞ and hðtÞ given in (2.6.11a) by directly deriving the result and verify the result using the results in Example 2.6.1.

a: xðtÞ ¼ :5dðt 1Þ þ :5dðt 2Þ; hðtÞ ¼ :5dðt 2Þ þ :5dðt 3Þ

2.6.2 Show the bounds given in (2.6.7a and b) and (2.6.9b) are valid. Use (2.6.11a).

b: xðtÞ ¼ ðt 1ÞP½t 1;

2.6.3 Show (2.6.9b) using (2.6.10).

hðtÞ ¼ xðtÞ;

c: xðtÞ ¼ ð1 t2 Þ; 1 t 1; hðtÞ ¼ P½t; d: xðtÞ ¼ eat uðtÞ;

hðtÞ ¼ ebt uðtÞ

for cases : 1:a > 0; b > 0; e: xðtÞ ¼ P½t=2; f: xðtÞ ¼ dðt 1Þ;

2.7.1 Find the autocorrelations of the following functions:

2:a ¼ 0; b > 0

hðtÞ ¼ P½t :5 P½t 1:5 hðtÞ ¼ et uðtÞ

g: xðtÞ ¼ cosðptÞP½t;

hðtÞ ¼ et uðtÞ:

a:x1 ðtÞ ¼ P½t :5 P½t 1:5; b:x2 ðtÞ ¼ uðt :5Þ uðt þ :5Þ;

c:x3 ðtÞ ¼ tP½t:

Compute the energies contained in the functions directly and then verify the results using the autocorrelation functions derived in the first part. 2.7.2 Verify the result in (2.7.3) using the results in Example 2.7.1.

Problems

69

2.7.3 Show the identity AC½xðt t0 Þ ¼ AC½xðtÞ: 2.7.4 Derive the AC function step by step for the function xðtÞ ¼ cosðo0 tÞP½t=T. Use the integral formula by assuming o0 ¼ p and T ¼ 4. Verify the results in Example 2.7.3 using the information provided in this problem. Give the appropriate bounds. 2.7.5 Show that the autocorrelations of the function x2 ðtÞ ¼ eat uðtÞ for a 0 do not exist. 2.8.1 a. Derive the time-average periodic autocorrelation function Rx; T ðtÞ for the following periodic function using the integral formula. xT ðtÞ ¼ A1 cosðo0 t þ y1 Þ þ A2 cosð2o0 t þ y2 Þ: b. Verify the result using (2.8.8) and (2.8.9). c. Compute the average power contained in the function directly and by evaluating the autocorrelation function at t ¼ 0. Sketch the function xðtÞ by assuming the values A1 ¼ 5; A2 ¼ 2; y1 ¼ 200 ; y2 ¼ 1200 . Sketch the autocorrelation function using

these constants. Suppose we are interested in determining the period T from these two sketches, which function is better, the given function or its autocorrelation? Why? 2.8.2 Let yT ðtÞ ¼ A þ xT ðtÞ; A constant. Repeat the last problem, except for the plots. 2.8.3 a. Show that the following functions are orthogonal over a period: xT ðtÞ ¼ cosðo0 t þ yÞ; yðtÞ ¼ A b. Show the functions xðtÞ ¼ P½t; y½t ¼ t are orthogonal. 2.8.4 Consider the signal zðtÞ ¼ xðtÞ þ yðtÞ. Show that the AC of this function is given by Rz ðtÞ ¼ Rx ðtÞ þ Ry ðtÞ þ Rxy ðtÞ þ Ryx ðtÞ: Simplify the expression for Rz ðtÞ by assuming that xðtÞ is orthogonal to yðtÞ for all t. 2.8.5 Complete the details in deriving the periodic cross-correlation function in terms of the aperiodic convolution leading up to Equation (2.8.3c). 2.8.6 Show (2.8.3c) using (2.6.5).

Chapter 3

Fourier Series

3.1 Introduction In this chapter we will consider approximating a function by a linear combination of basis functions, which are simple functions that can be generated in a laboratory. Joseph Fourier (1768–1830) developed the mathematical theory of heat conduction using a set of trigonometric (sine and cosine) series of the form we now call Fourier series (Fourier, J.B.J., 1955 (A. Freeman, translation)). He established that an arbitrary mathematical function can be represented by its Fourier series. This idea was new and startling and met with vigorous opposition from some of the leading mathematicians at the time, see Hawking (2005). Fourier series and the Fourier transform are basics to mathematics and science, especially to the theory of communications. For example, a phoneme in a speech signal is smooth and wavy. A linear combination of a few sinusoidal functions would approximate a segment of speech within some error tolerance. Suppose we like to build a structure that allows us to climb from the first floor to the second floor of a building. We can have a staircase approximating a ramp function using a linear combination of pulse functions. The amplitudes and the width of the pulses can be determined based on the error between the ramp and the staircase. Apart from the staircase problem, this type of analysis is important in electrical engineering, for example, when converting an analog signal to a discrete signal. The term ‘‘well-behaved’’ function, xðtÞ defined in the interval, ðt0 ; t0 þ TÞ is given in terms of the following Dirichlet conditions: 1. The function xðtÞ must be single valued within the given interval of T seconds.

2. The function xðtÞ can have at most a finite number of discontinuities and a finite number of maxima and minima in the time interval. 3. The function xðtÞ must be absolutely integrable on the interval, i.e., tZ0 þT

jxðtÞjdt ¼ finite51:

t0

Fortunately, all signals that we will be interested in satisfy these properties. The functions that do not satisfy the Dirichlet conditions are only of theoretical interest. Dirichlet gave an example that does not satisfy the conditions mentioned above and is x2p ðtÞ ¼

1;

t-rational

0; t-irrational

:

Our goal is to express a well-behaved function xðtÞ by an approximate function xa ðtÞ in terms of an independent set of functions ffk ðtÞg and a set of constants c½k in the form

xa ðtÞ ¼

N X

c½kfk ðtÞ:

(3:1:1)

k¼N

The subscript a on x in (3.1.1) denotes that it is an approximation of the function xðtÞ: Without loosing any generality we can assume that the limits on the sum N ! 1. We will be interested in a finite N that satisfies some constraints on the error signal, i.e., the difference between the given signal and its approximation. The entries in the expansion are assumed to have the following properties:

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_3, Ó Springer ScienceþBusiness Media, LLC 2010

71

72

1. The constants c½k are assumed to be some constants and k is an integer. 2. The set ffk ðtÞ; k ¼ N; ðN 1Þ; . . . 1; 0; 1; . . . ; ðN 1Þ; Ng is a linearly independent set. That is, fk ðtÞ cannot be obtained as a linear combination of the other fn ðtÞ; n 6¼ k. Such a set is called a basis function set and the members of this set are called basis functions. The basis functions can be real or complex. 3. Finally, we like to consider a basis set that is independent of xðtÞ: These properties are based on common sense. The first one allows for a level adjustment. The second property allows for the use of a set of independent basis functions. The third property allows for a general analysis. Later on we will see that some basis functions may be more attractive than others for a particular application. Fourier used the sine and cosine functions as basis functions. The most important aspect of generalized Fourier series expansion is that it allows an arbitrary function, defined over a finite interval, and may have discontinuities to be represented as a sum of basis functions, such as sine and cosine functions instead of using Taylor’s series:

3 Fourier Series

most of the standard circuits and systems text books. For example, see Ambardar (1995), Haykin and Van Veen (1999), Carlson (1998), Hsu (1967), and many others. Also, see Carslaw (1950), Jeffrey, (1956), Tolstov (1962), and Zygmund (1955). In Chapter 8 we will approximate a function by making use of samples of a signal in combination with some interesting interpolation functions. The presentation starts with the generalized Fourier series and later the Fourier series as a member of the generalized class of series. The basis functions are independent. The expansion gets easier if the basis functions are orthogonal.

3.2 Orthogonal Basis Functions The set of basis functions ffk ðtÞg is an orthogonal basis set if the functions satisfy tZ0 þT t0

x00 ðaÞðxaÞ2 xðtÞ¼xðaÞþx0 ðaÞðxaÞþ þ 2! xðnÞ ðaÞðtaÞn d n xðtÞ þ; xðnÞ ðaÞ¼ þ jt¼a n! dtn (3:1:2) This is an approximation of the function xðtÞ based upon the value of the given function at a point t ¼ a and the values of the derivatives of the function at that point. The Taylor series gives a strict prediction of xðtÞ at a finite distance from xðtÞjt¼a , whereas the Fourier series gives information of the function over the entire range t0 t 5t0 þ T. Another striking difference is the coefficients in the Taylor series are based upon the derivatives of the function at t ¼ a and the Fourier series coefficients are obtained by integration. Furthermore we can use (3.1.2) only if we know all the derivatives of the function at t ¼ t0 : If not, we have to resort to other methods, such as approximating them. The material in this chapter is fairly standard and can be found in

fk ðaÞfm ðaÞda ¼

Ek k ¼ m : 0 k 6¼ m

(3:2:1)

The superscript (*) on fm ðtÞ indicates complex conjugation. If fm ðtÞ is real, then fm ðtÞ ¼ fm ðtÞ. The symbol Ek is used to denote the energy in the basis function, fk ðtÞ in the given time interval and Ek is real. That is, Ek > 0ð assuming fk ðtÞ 6¼ 0Þ:

(3:2:2)

When k 6¼ m in (3.2.1), the integral is zero, which is the orthogonality property of the basis functions. If Ek ¼ 1 in (3.2.1), then the basis set is an orthonormal set. Orthonormality is not critical in our expansion, as we can create an orthonormal set by normalizing an orthogonal set, i.e., by replacing pﬃﬃﬃﬃﬃﬃ fk ðtÞ by fk ðtÞ= Ek . Therefore, we will concentrate on using orthogonal basis sets instead of orthonormal basis sets. Example 3.2.1 Show that the set ff1 ðtÞ; f2 ðtÞg given below is an independent set: t1 f1 ðtÞ ¼ P½t 0:5 and f2 ðtÞ ¼ P : (3:2:3) 2

3.2 Orthogonal Basis Functions

φ1(t)

73

φ2(t)

Example 3.2.3 Show that the following set is an orthogonal basis set over the time interval t0 t5t0 þ T and give the values of Ek :

Solution: The members of the set are sketched in Fig. 3.2.1 and the set is an independent set since any one of the members cannot be expressed in terms of the others. Since they overlap, and the time width of the second member is longer than the first member, they are not orthogonal. This can be seen from the integral f1 ðaÞf2 ðaÞda ¼

Z1

0

¼ 1 6¼ 0:

(3:2:4) &

Example 3.2.2 Consider the pulse functions given below and show that they form an orthogonal basis set. Find the value of A that makes the set an orthonormal set. t ðT=6Þ f1 ðtÞ ¼ AP ; T=3 t ðT=2Þ ; f2 ðtÞ ¼ AP T=3 t ð5T=6Þ f3 ðtÞ ¼ AP : (3:2:5) T=3 Solution: The pulse functions are shown in Fig. 3.2.2. Clearly, each pulse function exists in a different interval and therefore they are orthogonal. Since all the pulses are of the same width and the same height, we can write 2

2

A da ¼ ðA TÞ=3:

0

The functions are orthonormal if A ¼ Ei ¼ 1; i ¼ 1; 2; 3: φ1(t)

φ2 (t)

φ3 (t)

(3:2:6) qﬃﬃﬃ 3 T

tZ0 þT

xT ðaÞda ¼

Z

xT ðaÞda:

(3:2:8)

T

The integral on the right is over any period. Using the orthogonality property, we have Z Z k 6¼ m : fk ðaÞfm ðaÞda ¼ ejko0 a ejmo0 e dt T

¼

ZT e

T

jðkmÞo0 a

da ¼ e

jðkmÞo0 a

T 1 jðk mÞo0 0

0

¼

11 ¼ 0: ðk mÞo0 Z k ¼ m : ejko0 a ejmo0 a da

(3:2:9a) &

T

¼

ZT

(3:2:9b) da ¼ T; Ek ¼ T ¼ E:

0

Example 3.2.4 Test the orthogonality over the interval ðt0 ; t1 Þ of the set of functions ffk ðtÞ; k ¼ 0; 1; 2; . . .g ¼ 1; t; t2 ; . . . : (3:2:10)

and &

Solution: First Zt1 t0

Fig. 3.2.2 Pulse functions fi ðtÞ; i ¼ 1; 2; 3

(3:2:7b)

If a function xT ðtÞ ¼ xT ðt þ TÞ; then a short hand notation (see (1.5.14)) is

t0

0

E1 ¼ E2 ¼ E3 ¼

ejko0 ðtþTÞ ¼ ejko0 t ejko0 T ¼ ejko0 t ; o0 T ¼ 2p; f0 ¼ 2p=o0 ¼ 1=T:

P½a :5P½ða 1Þ=2da

ZT=3

(3:2:7a)

Solution: Note that fk ðtÞ ¼ fk ðt þ TÞ with the period T ¼ 2p=o0 since

Fig. 3.2.1 Pulse functions fi ðtÞ, i = 1, 2

Z2

fk ðtÞ ¼ ejko0 t ; k ¼ 0; 1; 2; . . . :

f1 ðaÞf2 ðaÞda ¼

Zt1

1 ada ¼ ½t21 t20 : 2

t0

The functions f1 ðtÞ and f2 ðtÞ are orthogonal if t1 ¼ t0 or if t1 ¼ t0 . Now consider the two functions f0 ðtÞ ¼ 1 and f2 ðtÞ ¼ t2 .With these, we have

74

3 Fourier Series

Zt1

f0 ðaÞf2 ðaÞda ¼

t0

Zt1

Z1

1 a da ¼ ðt31 t30 Þ: 3 2

t0

P1 ðtÞ ¼ f1 ðtÞ 1

The functions f0 ðtÞ and f2 ðtÞ are not orthogonal over any interval, except in the trivial case of t0 ¼ t1 . The set in (3.2.10) is not a good set to represent all & signals.

P0 ðaÞf1 ðaÞda Z1

P20 ðaÞda

1

Z1 P0 ðtÞ ¼ t 1

3.2.1 Gram–Schmidt Orthogonalization

ð1Þada P0 ðtÞ ¼ t:

Z1 da

1

Consider a linear set of independent real functions, f1 ðtÞ; f2 ðtÞ; . . . ; fk ðtÞ; . . ., defined on the interval ½a; b. Now define a new set of functions fj1 ðtÞ; j2 ðtÞ; . . . ; jk ðtÞ; . . .g by j1 ðtÞ ¼ f1 ðtÞ; Zb j1 ðaÞf2 ðaÞda j2 ðtÞ ¼ f2 ðtÞ

Similarly we can determine P2 ðtÞ ¼ t2 ð1=3Þ. We can multiply these polynomials by a constant since multiplying a polynomial in the set by a constant does not change the orthogonality of the polynomials. The above process generates the Legendre polynomials within a constant. The first five Legendre polynomials are listed below:

j1 ðtÞ

a

Zb

L0 ðtÞ ¼ 1;

j21 ðaÞda

L1 ðtÞ ¼ t;

Zb

L3 ðtÞ ¼ ð1=2Þð5t3 3tÞ;

L2 ðtÞ ¼ ð1=2Þð3t2 1Þ;

a

j3 ðtÞ ¼ f3 ðtÞ

j1 ðaÞf3 ðaÞda

L4 ðtÞ ¼ ð1=8Þð35t4 30t2 þ 3Þ:

a

Zb

Note the constant factors between Pi ðtÞ and Li ðtÞ. These polynomials can be generated by Rodrigue’s formula Spiegel (1968):

j21 ðaÞda

a

Zb j1 ðtÞ

j2 ðaÞf3 ðaÞda

a

Zb

Lk ðtÞ ¼ j2 ðtÞ; . . .:

(3:2:11)

j22 ðaÞda

a

This process of generating an orthogonal set of functions starting with an independent set is called the Gram–Schmidt orthogonalization process. Example 3.2.5 Use the Gram–Schmidt process to generate an orthogonal basis set, fPn ðtÞ; n ¼ 0; 1; 2; 3; . . .g; in the interval 1 t 1 using (3.2.11) in Example 3.2.4. Solution: First two are P0 ðtÞ ¼ f0 ðtÞ ¼ 1:

1 dðt2 1Þk ; k ¼ 0; 1; 2; 3; . . . : (3:2:12a) dtk 2k k!

The polynomials generated by this process are referred to as special Legendre polynomials. Note the subscript k is used as an index, which is different from p used in the Lp measures. They satisfy the orthogonality property ( Z1 0; m 6¼ k Lm ðaÞLk ðaÞda ¼ : 2 Ek ¼ ð2 kþ1Þ ;m¼k 1

(3:2:12b) & Example 3.2.6 Show the set of periodic functions given below is an orthogonal basis set over one period and compute the energy in each of the basis functions in one period:

3.3 Approximation Measures

75

f1; cosðo0 tÞ; cosð2o0 tÞ; . . . ; cosðko0 tÞ; . . . ; sinðo0 tÞ; sinð2o0 tÞ; . . . ; sinðko0 tÞ; . . .g:

Z T

ð1Þda ¼T;

Z

(3:2:13)

ð1Þ cosðko0 aÞda ¼ 0;

T

Solution: The members of the set are periodic with period T ¼ 2p=o0 and we need to show (3.2.1) using the members of the given set: Z

ð1Þ sinðko0 aÞda ¼ 0;

k ¼ 1; 2; . . . :

(3:2:14a)

T

Using trigonometric identities, we have 8 R R R cos2 ðko0 aÞda ¼ 12 da þ 12 cosð2 ko0 aÞda ¼ T2 ; k ¼ m > < T T T R : cosðko0 aÞ cosðmo0 aÞda ¼ 1 R 1 > : 2 cosððk þ mÞo0 aÞda þ 2 cosððk mÞo0 aÞda ¼ 0; k 6¼ m

Z T

T

Z

sinðko0 aÞ sinðmo0 aÞda ¼

1 2

Z

T

Z

T

cosðk mÞo0 ada

1 2

T

sinðko0 aÞ cosðmo0 aÞda ¼

1 2

Z

T

Z

T cosðk þ mÞo0 ada ¼

T

sinððk þ mÞo0 aÞda þ

T

1 2

Z

;k ¼ m : 0; k 6¼ m 2

sinððk mÞo0 aÞda ¼ 0; (3:2:14b)

T

for all k and m: These prove that the set in (3.2.13) is an orthogonal set. The energies contained in the members of the basis set in one period are as follows: ðEÞ1 ¼ T; ðEÞsine or a cosine function ¼ T=2:

(3:2:15) &

The set is an orthogonal set and not orthonormal set. There are many other basis sets.

3.3 Approximation Measures We are interested in approximating a given function xðtÞ over an interval (t0 ; t0 þ TÞ by xa ðtÞ using a set of orthogonal basis functions. How do we measure the approximation and then how good is the approximation? It can be measured by the error ½xðtÞ xa ðtÞ. Figure 3.3.1 illustrates an example where xðtÞ is the given function and its approximation is xa ðtÞ. The hatched area represents the error. Since the functions can be complex and a positive error is just as bad as a negative error, and to make it general, we would like to consider the magnitude of

Fig. 3.3.1 xðtÞ Given function, xa ðtÞ Function approximating xðtÞ

the error. In addition, if the error measure is a number, we can compare and evaluate a particular approximation with respect to a number of basis sets. These goals can be achieved by considering the integral of the pth power of the magnitude of the error function, i.e., tZ0 þT

jxðtÞ xa ðtÞjp dt; 1 p:

(3:3:1)

t0

Notes: There is a good deal of interest in the area of inverse problems, such as deconvolution of signals based on Lp ; 1 p measures. For a

76

3 Fourier Series

review see Tarantola (1987) and Hassan et al. (1994). Statisticians have investigated the Lp measures based on probabilistic behavior of signals. Our discussion here does not involve any details of statistical analysis. For readers interested in statistical details of these measures, p is selected based upon the kurtosis defined as the fourth moment normalized by the square of the variance of a probability density function Money (1982). In Section 1.7 we have considered three important density functions, uniform, Gaussian, and Laplacian. These can be used in selecting the value of p. Kurtosis values are k = 1.8 (Uniform, 3 ðGaussianÞ; and 6 ðLaplacianÞ: The constant p is selected by

9 p¼ ; 1 k 1; k2 þ 1 9 8 > = < k > 3:8 use L1 > 2:25k53:8 use L2 : > > ; : k52:2 use L1

(3:3:2)

1 MSE ¼ T

(3:3:3)

Minimizing the error (see e2 in (3.3.a)) corresponds to maximizing the probability, thus

tZ0 þT

jxðtÞ xa ðtÞj2 dt; xa ðtÞ

t0

¼

N X

c½kfk ðtÞ S2 Nþ1 :

(3:3:4a)

k¼N

The interval T will be the same for different approximations in a particular situation and the normalization constant can be omitted and compare the approximations by using an orthogonal basis set ffk ðtÞg and the integral-squared error (ISE), i.e., tZ0 þT

Lp measures are used for speech, seismic, radar, and other signal coding. For example, for vowel sounds, L1 is preferable and for nonvowel sounds, L2 is preferable, see Lansford and Yarlagadda (1988). Since seismic signals have spiky noise, L1 seems to work well, see Yarlagadda et al. (1985). See Schroeder and Yarlagadda (1989) on spectral estimation using L1 norm. An old adage is if you have a fork in the road and have a choice to select either L1 or L2 measure, L2 tells you to go in the middle of the two roads, not one of the two possible paths, whereas L1 suggests taking one or the other paths. L2 measure thinks like a machine, whereas L 1 measure thinks like a human, see Problem 3.2.2 at the end of the chapter. For most applications, L2 the least-squares error measure is adequate and simple to use. In Section 1.7.1, the Gaussian probability density function was introduced, see (1.7.12). Writing it terms of the error e with mean 0 and variance s2e , the density function is 1 2 2 fe ðeÞ ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ee =2se : 2 2pse

providing a mathematical basis for the least& squares approach. The mean-squared error (MSE) is defined by considering ð2 N þ 1Þ terms, we have

ISE¼

jxðtÞxa ðtÞj2 dt

t0

¼

tZ0 þT t0

2 N X c½kfk ðtÞ dt e2Nþ1 : (3:3:4b) xðtÞ k¼N

The subscript on e; 2 N þ 1 corresponds to the number of terms in the approximation and some of the coefficients c½k may be zero and N could go to infinity. It is convenient to consider odd number of terms. Since c½k0s are unknowns, there is no loss in generality. The constants c½k can be determined from the basis set ffk ðtÞ; k ¼ N; . . . ; 2; 1; 0; 1; 2; . . . ; Ng by minimizing the ISE. We will consider two ways of computing the constants that minimize the integral-squared error. The first one is based upon taking the partials of the ISE with respect to c½k, equating the partials to zero, and then solving for them. The second one is based on using perfect squares by rewriting the ISE in terms of two parts. First term is independent of c½k and the second is a sum of perfect square terms involving c½k. Equating the perfect square terms to zero and solving for c½k give the desired result.

3.3 Approximation Measures

77

Using this in (3.3.6) results in

3.3.1 Computation of c[k] Based on Partials First,

tZ0 þT

tZ0 þT

e2 Nþ1 ¼

2 N X c½kfk ðtÞ dt xðtÞ k¼N

t0 tZ0 þT

¼

x ðtÞfk ðtÞdt þ c ½kðEk Þ ¼ 0:

t0

The coefficients c½k are the generalized Fourier series coefficients giving an explicit formula given below to compute c½k given xðtÞ and the orthogonal basis setfjk ðtÞg:

jxðtÞ S2 Nþ1 j2 dt

t0 tZ0 þT "

xðtÞ

¼

t0

#

N X

tZ0 þT

1 c½k ¼ Ek

c½mfm ðtÞ #

N X

x ðtÞ

(3:3:9)

t0

m¼N

"

xðtÞfk ðtÞdt:

c ½kfk ðtÞ dt:

(3:3:5)

k¼N

Note the two different variables k and m in the above summations. This allows us to keep track of the terms in the summation products. Multiplying the product terms and since the integral of a sum is equal to the sum of the integrals, we have

e2Nþ1 ¼

tZ0 þT

xðtÞx ðtÞdt

N X

c ½k

k¼N

þ

tZ0 þT

c½m

m¼N

t0

N X

tZ0 þT

x ðtÞfm ðtÞdt

t0

xðtÞfk ðtÞdt

t0

N N X X

c½mc ½k

k¼N m¼N

tZ0 þT

fm ðtÞfk ðtÞdt:

t0

(3:3:6) Take the partial derivatives of e2 Nþ1 with respect to c½m and equate them to zero: @e2Nþ1 ¼ @c½m

tZ0 þT

½x ðtÞfm ðtÞdt

t0

þ

N X

tZ0 þT

c½k

k¼N

3.3.2 Computation of c[k] Using the Method of Perfect Squares Example 3.3.1 Consider the second-order polynomial yðtÞ ¼ t2 þ bt. Determine the minimum value of yðtÞ by having a part of the expression that is a perfect square. Solution: By adding and subtracting the term ðb=2Þ2 to yðtÞ, we have

2 2 b b 2 2 yðtÞ ¼ t þ bt ¼ t þ bt þ 2 2

2 2 b b ¼ tþ : 2 2 This function takes a minimum value for t ¼ b=2 & and yðtÞ ¼ ðb=2Þ2 : This idea can be used in minimizing the integralsquared error, See Ziemer and Tranter (2002). Add and subtract the following term to e2 Nþ1 in (3.3.6): 2 t þT Z0 N X 1 xðtÞfk ðtÞdt : E k¼N k t0

fm ðtÞfk ðtÞdt ¼ 0: (3:3:7)

t0

Orthogonality of the basis functions implies

eð2Nþ1Þ ¼

tZ0 þT

( 2

jxðtÞj dtþ

k¼N

t0 tZ0 þT t0

fk ðtÞfm ðtÞdt ¼

Ek > 0 and real; k ¼ m 0;

otherwise

: (3:3:8)

N X

N X k¼N

tZ0 þT

c½k t0

tZ0 þT

c½k t0

xðtÞfk ðtÞdt

x ðtÞfk ðtÞdt

78

3 Fourier Series

2 t þT approximation to the function xðtÞ, only in the Z0 1 1 X X sense that the approximation gives a smaller mean 1 þ xðtÞfk ðtÞdt þ Ek ðc½kc ½kÞg square error. In the limit, for any complete set of E k¼1 k¼1 k t0 basis functions defined below: 2 t þT 0 Z 1 X 1 (3:3:13) lim e2 Nþ1 ¼ 0: N!1 xðtÞfk ðtÞdt : (3:3:10) E k¼1 k t 0 From (3.3.12), we have Parseval’s equation, or forThe terms inside the brackets {.} can be expressed as mula, or identity given by a sum of squares and

e2 Nþ1 ¼

tZ0 þT

2

jxðtÞj dt þ

tZ0 þT N X

jxðtÞj2 dt ¼

Ek jc½kj2 :

(3:3:14)

k¼1

t0

k¼N

t0

2 tZ0 þT pﬃﬃﬃﬃﬃﬃ 1 Ek c½k pﬃﬃﬃﬃﬃﬃ xðtÞfk ðtÞdt Ek t0 2 t þT Z0 N X 1 xðtÞfk ðtÞdt : (3:3:11) E k¼N k

Summary: Given a time-limited function xðtÞ; t0 t5t0 þ T and a set of orthogonal basis functions ffk ðtÞ; k ¼ 0; 1; 2; . . .g, the function xðtÞ is approximated by

To show that (3.3.11) reduces to (3.3.10), expand the middle term on the right and then cancel the equal terms that have opposite signs. The parameters c½k are not in the first and the third terms in (3.3.11). It is only included in the middle term, which is a sum of absolute values. Therefore, minimization of e2 Nþ1 is achieved when the middle term is zero and c½k is given by (3.3.9).

1 c½k ¼ Ek

The minimum ISE with ð2 N þ 1Þ terms is (see (3.3.5))

Ek jc½kj2

tZ0 þT

xðtÞfk ðtÞdt

(3:3:15)

t0

and the integral-squared error is equal to zero. The function xa ðtÞ is an approximation of xðtÞ in the identified interval. Only in the sense that the integral-squared error goes to zero, we write the equality of the given function to the approximate function by 1 X

c½kfk ðtÞ:

(3:3:16)

jxðtÞ S2 Nþ1 j2 dt 0:

It should be emphasized that xðtÞ and xa ðtÞ are not equal in the true sense. Differences between these two functions will be considered a bit later. In simple terms, the coefficients c½k of the generalized Fourier series are 2 3 6 6 c½k ¼ 6 4

k¼N

t0

1 ¼ T

N X

tZ0 þT

k¼1

3.3.3 Parseval’s Theorem

jxðtÞj2 dt

c½kfk ðtÞ;

k¼1

xðtÞ ¼ xa ðtÞ ¼

tZ0 þT

1 X

xa ðtÞ ’

t0

e2 Nþ1 ¼

1 X

(3:3:12)

t0

As N increases, the partial the quantity ðe2 Nþ1 Þ can only decrease. Therefore, as N increases, the partial sums of the F-series give a closer and closer

7 tZ0 þT 1 7 7 Ek , Energy in fk ðtÞ 5 t0 in the interval ðt0 ;t0 þ TÞ ðconjugate of the dt: (3:3:17a) xðtÞ basis function, fk ðtÞÞ

The error in (3.3.12) with ð2 N þ 1Þ coefficients in the F-series expansion is equal to

3.3 Approximation Measures

79

E2 Nþ1 ¼ ðEnergy in the given T second interval of the function) " # N X th ðSquared magnitude of the kth coefficient) (Energy in the k basis function) k¼N

¼

tZ0 þT t0

jxT ðtÞj2 dt

N X

jc½kj2 Ek ¼ ISE: (3:3:17b)

k¼N

Later, the convergence of the approximated signal xa ðtÞ to xðtÞ in terms of the number of coefficients in the approximation will be considered. The signal xðtÞ is assumed to satisfy the Dirichlet conditions. The value of the integral in (3.3.17b) is unique and the generalized Fourier series is unique for a given set of basis functions. Let us illustrate the above by a detailed example, see Simpson and Houts (1971). &

Earlier we have shown that fi ðtÞ; i ¼ 1; 2; 3 form an orthogonal basis set with Ek ¼ A2 T=3; k ¼ 1; 2; 3, see Example 3.2.2. Noting that the time interval of the given function is 0 t5T and from (3.3.9), the coefficients are as follows: 1 c½1 ¼ E1

xðtÞf1 ðtÞdt

t0 ¼0

1 ¼ 2 A ðT=3Þ

ZT

3 tf ðtÞdt T 1

0

Example 3.3.2 Find the generalized Fourier series expansion of the function xðtÞ given below in Fig. 3.3.2 using the three orthogonal basis functions in (3.2.5). Then compute the mean-squared error using the direct method and Parseval’s theorem.

ZT=3 1 3 t:Adt ¼ 2 A ðT=3Þ T 0

2 1 3 3 t t ¼ T3 1 A ¼ ¼ 2 2 t ¼ 0 2A A T T

Solution: The expression for the time function xðtÞ and the three basis functions are t T2 3 ; xðtÞ ¼ t P T G t T6 ; f1 ðtÞ ¼ AP T=3 t T2 f2 ðtÞ ¼ AP ; and T=3 t 5T 6 f3 ðtÞ ¼ AP : T=3

t0 Z þT¼T

c½ 2 ¼

1

T

A2

3

2T=3 Z

3 3 5 t Adt ¼ ; c½ 3 ¼ : T 2A 2A

T=3

xa ðtÞ ¼ c½1f1 ðtÞ þ c½2f2 ðtÞ þ c½3f3 ðtÞ 1 t T=6 3 t T=2 ¼ P þ P 2 T=3 2 T=3 5 t 5t=6 : þ P 2 T=3 (3:3:18)

The functions xðtÞ and xa ðtÞ are sketched in Fig. 3.3.2. The error is shown by the hatched marks and it has six equal parts. Clearly, ZT=6 ZT=6

3 1 2 2 ISE ¼ 6 t ½xðtÞ xa ðtÞ dt ¼ 6 dt T 2 0

0

9t2 1 3t þ dt T2 4 T 0 1 3 t2 9 t3 T=6 T þ ¼ ¼6 t 4 T 2 T2 3 0 12

¼6

Fig. 3.3.2 xðtÞ, xa ðtÞ and the error between these two functions

(3:3:19)

ZT=6

(3:3:20a)

This procedure is complicated and unnecessary since some of this work has already been done

80

3 Fourier Series

and should be evident from the context. The equality is only true in the sense that the integralsquared error between the periodic function and the F-series in the given interval is zero. The given function and the corresponding Fourier series 2 # coefficients are identified by the symbolic notation:

in deriving the generalized Fourier series coefficients. Using Parseval’s theorem and (3.3.17b) results in the same value as in (3.3.20a) and is shown below.

ISE ¼

ZT

3 t T

2

" 2 2

T 1 3 5 dtA þ þ 3 2A 2A 2A 2

0

FS;T

xT ðtÞ ! Xs ½k

35 T ¼3T T¼ : 12 12

(3:4:1b)

(3:3:20b) &

3.4 Fourier Series The generalized Fourier series developed earlier can be used to study both the complex and the trigonometric Fourier series of a periodic function xT ðtÞ with period of T seconds. The functions ejno0 t ; sinðno0 tÞ; and cosðno0 tÞ are nice functions in the sense that all the derivatives of these functions exist. Approximation of a given function by the Fourier series (F-series) gives a smooth function even when the function being approximated has discontinuities.

The subscript s on X is used to denote that the coefficients are the complex F-series coefficients. Note the difference in the sign of the exponents in the sum and the integral expressions in (3.4.1a). Fourier coefficients are computed by an integral and the integral has a unique value. That is, F-series expansion is unique. The complex F-series can be used to approximate an aperiodic function in a time interval (t0 ; t0 þ TÞ, where t0 is arbitrary and T is the interval of the function that is under consideration. The approximation, in terms of periodic basis functions, will be valid only in the given time interval and, outside this interval, it is not valid. Complex Fourier series is applicable to both real and complex functions. When the function xT ðtÞ is real, the F-series coefficients Xs ½k and Xs ½k are related. Z 1 xT ðtÞejko0 t dt T T 9 8 = 2:

The F-series in (3.4.13b) contains only sine terms since x2p ðtÞ is an odd function. It is more appealing & since the given function is real.

Since the function is real and even, the above validates Xs ½k ¼ Xs ½k. The F-series coefficients are unique. Using trigonometric identity, it follows that

Example 3.4.4 Derive the complex F-series given the periodic impulse sequence

xT ðtÞ ¼ ð1=2Þ þ 2 cosðo0 tÞ ð1=2Þ sinð2o0 tÞ:

xT ðtÞ ¼

1 X

Adðt kTÞ:

(3:4:14)

k¼1

Solution: The complex F-series coefficients are

&

The complex F-series expansion is applicable for both real and complex periodic functions and the complex F-series leads to Fourier transforms. Most functions, in reality, are real functions and the trigonometric F-series are more desirable.

3.4 Fourier Series

83

3.4.2 Trigonometric Fourier Series

3.4.3 Complex F-series and the Trigonometric F-series The set of basis functions for the trigonometric Coefficients-Relations Fourier series with period T ¼ 2p=o0 is

f1; cosðko0 tÞ; sinðko0 tÞ; k ¼ 1; 2; 3; . . .g: (3:4:16a)

By using Euler’s formula and comparing the results in (3.4.1a) and (3.4.17), we have

Equation (3.2.15) gives the energy contents in one period of these functions and are Xs ½k ¼ ðEÞ1 ¼ T; ðEÞ sine or a cosine function ¼ T=2:

1 X

Z

xðtÞejko0 t dt

T

(3:4:16b)

In the trigonometric F-series, Xs ½0 for the dc term, a½k for the coefficients for the cosine terms, and b½k for the coefficients of the sine terms will be used. With (3.4.16b) and (3.3.17a), the following trigonometric F-series result: xT ðtÞ ¼ Xs ½0 þ

1 T

1 ¼ T

Z

j xðtÞ cosðko0 tÞdt T

T

¼

(3:4:19a)

Since the cosine and sine functions are even and odd functions, we can see that

fa½k cosðko0 tÞ þ b½k sinðko0 tÞg; Xs ½k ¼

Z 1 Xs ½0 ¼ xT ðtÞdt T T Z 2 a½k ¼ xT ðtÞ cosðko0 tÞdt: T > > T > > Z > > 2 > > > b½k ¼ xT ðtÞ sinðko0 tÞdt > : T

xðtÞ sinðko0 tÞdt T

a½k b½k j : 2 2

k¼1

8 > > > > > > > > > >

0:

(3:6:3)

The magnitudes of the F-series coefficients of a periodic signal and its delayed (or advanced) version are the same. The delay (or advance) t appears explicitly in the phase angle ðko0 tÞ for the advance and ðko0 tÞ for the delay. In case of trigonometric F-series, for t > 0 (for t50, replace t by t in the following.), we have

88

3 Fourier Series

xT ðttÞ¼Xs ½0þ

1 X ½a½kcosðko0 ðttÞÞ

Solution: Substituting T ¼ 2p, i.e., o0 ¼ 1 in (3.5.9), we have

k¼1

þb½ksinðko0 ðttÞÞ¼Xs ½0 1 X ½a½kcosðko0 tÞb½ksinðko0 tÞ þ k¼1

cosðko0 tÞ þ

1 X

2 4 x2p ðtÞ ¼ p p cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : 1ð 3Þ 3ð 5Þ 5ð 7Þ (3:6:7) &

½a½ksinðko0 tÞ

k¼1

þb½kcosðko0 tÞsinðko0 tÞ:

(3:6:4)

Example 3.6.2 Find the trigonometric F-series of the signal below using the full-wave rectified signal series: y2p ðtÞ ¼

3.6.3 Time and Frequency Scaling Given a function xðtÞ, its time-scaled version is given by xðatÞ; a > 0. If a51, the signal is expanded and if a > 1, the signal is compressed (see (1.2.3)). The F-series of the scaled signal is obtained by replacing t ! at in the complex F-series and 1 X xT ðatÞ ¼ Xs ½kejko0 ðatÞ : (3:6:5)

sinðtÞ; 0;

05t5p : p5t52p

(3:6:8)

Solution: The full-wave rectified function x2p ðtÞ was shown in Fig. 3.5.1 with period T ¼ 2p. The half-wave rectified signal is shown in Fig. 3.6.1 and y2p ðtÞ ¼ ½x2p ðtÞ þ sinðtÞ=2:

(3:6:9)

k¼1

The frequency locations moved from ko0 to kðao0 Þ and the F-series coefficients are Xs ½k. The time-scaled signal changes the period from T to ðT=aÞ: Harmonics are now located at ko0 a ¼ kð2pÞf0 a ¼ kð2pÞða=TÞ. The more compressed the time function is, farther apart its harmonics are. Harmonics of the compressed signal are located at frequencies kf0 a and a > 1: If a ¼ 1, then a time-reversed or a folded function results and xT ðtÞ ¼

1 X

Xs ½kejko0 t ¼

k¼1 FS;T

1 X Xs ½kejko0 t :

Fig. 3.6.1 y2p ðtÞ Half-wave rectified signal

Using the linearity property of the Fourier series, we have 1 1 2 y2p ðtÞ¼ þ sinðtÞ p 2 p cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : 1ð3Þ 3ð5Þ 5ð7Þ

k¼1

xT ðtÞ ! X ½k; jXs ½kj ¼

jXs ½kj:

(3:6:10) (3:6:6)

Example 3.6.1 Using Example 3.5.2, find the trigonometric F-series of the full-wave rectified signal x2p ðtÞ ¼ jsinðtÞj assuming the period is 2p.

In Chapter 1, we studied the one-sided and the two-sided line spectra by expressing each term in terms of cosines and sines. Noting that cosða 900 Þ ¼ sinðaÞ and cosða 1800 Þ ¼ cosðaÞ, we can write

1 1 2 cosð2t 1800 Þ cosð4t 1800 Þ cosð6t 1800 Þ 0 þ þ þ : y2p ðtÞ ¼ cosðt þ 90 Þ þ p 2 p 1ð3Þ 3ð5Þ 5ð7Þ

(3:6:11)

3.6 Operational Properties of Fourier Series

89

2 4 cosð2t 1800 Þ cosð4t 1800 Þ cosð6t 1800 Þ þ þ þ : x2p ðtÞ ¼ þ p p 1ð3Þ 3ð5Þ 5ð7Þ These are the harmonic forms of the trigonometric Fourier series for the two given functions. The two-sided amplitude line spectra associated with these two functions are sketched in Fig. 3.6.2. The only difference between the two functions x2p ðtÞ and y2p ðtÞ is the component at the frequency o0 ¼ 1 or f0 ¼ 1=2p. If we can remove or filter out this frequency component from y2p ðtÞ, we can obtain the function x2p ðtÞ illustrating one of the remarkable insights into the description of signals provided by the Fourier series. Filter design & will be discussed in later chapters.

(3:6:12)

The dc term in the F-series of the derivative goes to zero and the other coefficients are multiplied by ðjko0 Þ, k 6¼ 0: The spectral components of x0T ðtÞ have significantly higher frequency content compared to the spectral components of xT ðtÞ. Note the F-series coefficients of xT ðtÞ are multiplied by jko0 to obtain the F-series coefficients of x0T ðtÞ. Derivative operation enhances the details in the signal. We can state that d n xT ðtÞ FS;T !ðjko0 Þn Xs ½k; k 6¼ 0: dtn

(3:6:13)

In a similar manner, the derivatives of the trigonometric F-series are given by xT ðtÞ ¼ Xs ½0 þ

1 X

a½k cosðko0 tÞ

k¼0

þ

(a)

1 X

b½k sinðko0 tÞ:

k¼0 1 dxT ðtÞ X ¼ ðko0 Þa½k sinðko0 tÞ dt k¼1

þ

1 X

ðko0 Þb½k cosðko0 tÞ:

k¼1

(b) Fig. 3.6.2 Two-sided amplitude line spectra (a)x2p ðtÞ and (b) y2p ðtÞ

¼

1 X

þ

3.6.4 Fourier Series Using Derivatives

a1 ½k cosðko0 tÞ

k¼1 1 X

b1 ½k sinðko0 tÞ;

a1 ½k ¼ b½kðko0 Þ;

k¼1

b1 ½k ¼ a½kðko0 Þ All the derivatives of the F-series exist since all the derivatives of the functions ejno0 t ; sinðno0 tÞ; and cosðno0 tÞ exist. In that sense, F-series is a nice function. The complex F-series of a periodic function and its derivative are 1 X Xs ½kejko0 t ; xT ðtÞ ¼ k¼1 1 dxT ðtÞ X x0T ðtÞ ¼ ½Xs ½kjðko0 Þejko0 t ¼ dt k¼1 k6¼0

b1 ½k a1 ½k ; b½k ¼ ; k 6¼ 0; ko0 ko0 Z 1 Xs ½0 ¼ xT ðtÞdt: T

a½k ¼

(3:6:14)

T

In (3.6.14), the subscript ‘‘1’’ on a and b indicates the trigonometric F-series are for the first derivative of the periodic function. The dc component needs to be computed directly from the given periodic

90

function. This approach of finding the F-series is the derivative method of finding F-series. The derivative property allows simplifies the computing the F-series coefficients. Most of the signals we deal with are pulses that do not have derivatives in the conventional sense. The derivatives of such functions can only be considered in the sense of generalized functions discussed in Section 1.4, resulting in impulse functions in the derivatives. See, for example, the derivative of the rectangular pulse in (1.4.33). Fourier series coefficients are

3 Fourier Series

determined by use of integrals. Integrals involving impulses are trivial to compute and, therefore, computing the Fourier series coefficients by the derivative method makes it simple. Example 3.6.3 Find the trigonometric F-series of the trapezoidal waveform shown in Fig. 3.6.3a using the derivative method. Solution: The first two derivatives of the wave form are shown in Figs. 3.6.3b and c. The second derivative of the function xT ðtÞ is

(a)

(b)

Fig. 3.6.3 (a) Trapezoidal wave-form xT ðtÞ, (b) xT0 ðtÞ, and (c) xT00 ðtÞ

(c)

3.6 Operational Properties of Fourier Series

x00T ðtÞ ¼

91

d2 xT ðtÞ 1 ¼ dt2 ðd2 d1 Þ

½dðt þ d2 Þ dðt þ d1 Þ dðt d1 Þ þ dðt d2 Þ; d2 d1 ; T=25t5T=2; xT ðt þ TÞ ¼ xT ðtÞ:

considered above decay proportional to (1/k2 Þ. In the case of d2 ¼ d1 ; we have a rectangular pulse waveform. Using L’Hospital’s rule, the coefficients in (3.6.16a) can be simplified. Assuming d2 ¼ d1 þ e,

(3:6:15) Since the second derivative has an even symmetry, we can use some simplifications:

( a½k ¼lim e!0

¼lim

4 Teðko0 Þ2 4

e!0 Tðko

b2 ½k ¼ 0;

2 a2 ½ k ¼ T

ZT=2

x00T ðtÞ cosðko0 tÞdt

T=2

¼

4 ðd2 d1 ÞT

ZT=2

½cosðko0 ðd1 þeÞÞcosðko0 d1 Þ

ðko0 Þsinðko0 ðd1 þeÞÞ 4sinðko0 d1 Þ ¼ 1 kTo0 0Þ 2

1 2d1 X 4 sinðko0 d1 Þ þ cosðko0 tÞ: xT ðtÞ d1 ¼d2 ¼ T ko0 T k¼1 (3:6:16c)

½dðt d1 Þ þ dð1 d2 Þ

0

cosðko0 tÞdt;

a2 ½k ¼

)

4 ½ cosðko0 d1 Þ þ cosðko0 d2 Þ: Tðd2 d1 Þ

Noting that a2 ½k ¼ ðko0 Þ2 a½k, we have the dc term (to be computed directly) and Zd2 1 1 xT ðtÞdt ¼ ðd1 þ d2 Þ: Xs ½0 ¼ T T 4 d2 ½cosðko0 d1 Þ cosðko0 d2 Þ: a½k ¼ Tðd2 d1 Þðko0 Þ2

From (3.6.16b and c), the F-series coefficients of the trapezoidal and the rectangular pulse sequences decay at a rate proportional to ð1=k2 Þ and (1/k), respectively. The derivative of an even (odd) function is an odd (even) function. See Fig. 3.6.3 a,b,c and the F-series to verify the above assertion and the chain rule below: dy dy da dxe ðtÞ ¼ ! yðtÞ ¼ ; dt da dt dt yðtÞ ¼

dxe ðtÞ dxe ðtÞ ¼ ð1Þ ¼ yðtÞ: dt dt (3:6:17) &

The trigonometric F-series are xT ðtÞ ¼

d1 þ d2 4 þ 2 T o0 Tðd2 d1 Þ 1 X 1 ½cosðko0 d1 Þ cosðko0 d2 Þ cosðko0 tÞ: 2 k k¼1

(3:6:16a) xT ðtÞjd1 ¼0 ¼

d2 4 þ T d2 Tðo20 Þ

1 X 1 ½1 cosðko0 d2 Þ cosðko0 tÞ: k2 k¼1

(3:6:16b)

Note that when d1 ¼ 0, we have a triangular pulse wave form. The F-series coefficients for the two cases

3.6.5 Bounds and Rates of Fourier Series Convergence by the Derivative Method The Fourier series coefficients Xs ½k of a periodic signal xT ðtÞ usually decay at a rate inversely proportional to kn , where k is the harmonic index. An exception is the periodic impulse sequence. In Example 3.4.4, we have seen that the periodic impulse sequence has the complex F-series given by Xs ½k ¼ ðA=TÞ, i.e., the coefficients are independent of k. Higher the value of n in kn is, the faster the high-frequency component decays. An estimated value of n can be determined without actually

92

3 Fourier Series

computing the Fourier series coefficient Xs ½k using the derivative properties of xT ðtÞ: In (3.6.13) we have seen that the Fourier series coefficients of the nth derivative of a periodic function are related to the Fourier coefficients of the function multiplied by ðjko0 Þn . If we differentiate an arbitrary periodic function xT ðtÞ n times before the first set of impulses appear, then the F-series coefficients have the property that Xs ½k / ð1=jko0 jÞn . That is, the decay rate of the coefficients is proportional to 1/ðjkjÞn . The decay rate is a good indicator for large k, as some of the early coefficients may be even zero. Since the complex F-series and trigonometric F-series are related, the decay rate of the trigonometric F-series coefficients is tied to the decay rate of the complex F-series coefficients. The derivative property of the F-series provides bounds on the F-series coefficients, referred to as the spectral bounds. For k 6¼ 0, n Z 1 0 T T Z 1 Tjkjn jo0 jn

1 jXs ðkÞj ¼ jko

ðnÞ jko0 t xT ðtÞe dt ðnÞ xT ðtÞ dt:

n ¼ 0 bound : jXs ½kj

Z1=2

1 2ðo0 Þ

0

(3:6:18a)

In deriving the right side of the above equation, the fundamental theorem of calculus is used and Z Z

yðtÞdt jyðtÞjdt note ejko0 t ¼ 1 : For example, the F-series coefficients in Example 3.4.2 show their decay rate is proportional to ð1=kÞ, see (3.4.8). The function xT ðtÞ has discontinuities at t ¼ t=2 in one period of the time function. Correspondingly, the bound on the F-series coefficient is (3:6:18b)

The number of nonzero bounds that can be determined equals the number of times the function can be differentiated before derivatives of impulses occur in the derivatives, see Ambardar (1995), Morrison, (1994), and others. Example 3.6.4 Find all the nonzero spectral bounds for the rectangular pulse xðtÞ ¼ P½t; xT ðtÞ ¼ xT ðt þ TÞ; T ¼ 2:

jxT ðtÞjdt ¼

1 : 2

(3:6:19a)

1=2

n ¼ 1 bound : Z1 Z1 dx2 ðtÞ 1 jXs ½kj dt dt ¼ 2p 2jo0 j1 1 1 1 1 1 dðt þ Þ þ dðt Þ dt ¼ : 2 2 p 1

The bounds above n ¼ 1 are not defined.

(3:6:19b) &

Bounds on the trigonometric Fourier series coefficients: If xT ðtÞ has discontinuities, then its trigonometric F-series coefficients satisfy for large k ja½kj5

T

jXs ½kj5B=k; B a constant:

Solution: The function, its first, and second derivatives are shown in Fig. 3.6.4. Noting that o0 ¼ ð2p=2Þ ¼ p, we have

ðKa1 and Kb1

Ka1 Kb1 and jb½kj5 k k are some constants):

(3:6:20)

If xT ðtÞ is continuous but x0T ðtÞ is discontinuous, then for large k Ka2 Kb2 and jb½kj5 2 ðKa2 and Kb2 2 k k are some constants): (3:6:21)

ja½kj5

Convergence is a function of the continuity of the highest derivative of xT ðtÞ. The convergence rates of the coefficients a½k and b½k may be different. Example 3.6.5 Consider the function and its F-series coefficients given below. Comment on the convergence rates. x2p ðtÞ ¼ et ; p5t5p; x2p ðt þ 2pÞ ¼ x2p ðtÞ: (3:6:22a) " # 1 2 sinhðpÞ 1 X ð1Þk þ x2p ðtÞ ¼ p 2 k¼1 ð1 þ k2 Þ ðcosðktÞ k sinðktÞÞ:

(3:6:22b)

Solution: The sine series coefficients b½k converge like ðKb =kÞ, whereas the cosine series coefficients a½k converge like ðKa =k2 Þ, implying that the Fourier series, as a whole, converges like ðK=kÞ; K 0 s

3.6 Operational Properties of Fourier Series

93

Fig. 3.6.4 (a) Periodic pulse waveform, (b) periodic impulse sequence, and (c) periodic doublet sequence

(a)

(b)

(c) are constants. The dc term and the first few harmonics contain bulk of the power for low-frequency & signals.

3.6.6 Integral of a Function and Its Fourier Series Consider the integral of a periodic function and its F-series. The integral of a periodic function with a dc component cannot be periodic as the integral of a constant is a ramp. In the case of Xs ½0 ¼ 0, we can derive the complex F-series of an integral of a function: # Zt Zt " X 1 jko0 a Xs ½ke da yT ðtÞ¼ xT ðaÞda¼ 0 1 X

¼

Ys ½ke

k¼1 1 X

¼

0 jko0 t

Ys ½k ¼

½Xs ½k=jko0 ; k 6¼ 0 constant; k ¼ 0

:

(3:6:23b)

Note the division by ko0 above indicating the F-series of the integral of a function xT ðtÞ converges faster than the F-series convergence of xT ðtÞ. Integration is a smoothing operation. Therefore, the integrated signal has much smaller high-frequency content than xT ðtÞ. Without the dc term Xs ½0, integration and differentiation can be thought of as inverse operations.

3.6.7 Modulation in Time

k¼1;k6¼0

Consider the F-series xT ðtÞejat ¼

fð1=jko0 ÞXs ½kgejko0 t þconstant:

k¼1;k6¼0

¼ (3:6:23a)

1 X k¼1 1 X k¼1

Xs ½kejko0 t ejat Xs ½kejðko0 aÞt :

(3:6:24)

94

3 Fourier Series

Multiplying a function by ejat shifts the frequencies from ko0 to ko0 a. This is modulation. Multiplying xT ðtÞ by cosðatÞ and using Euler’s formula results in y1 ðtÞ ¼ xT ðtÞ cosðatÞ ¼

1 1 X Xs ½kejðko0 þaÞt 2 k¼1

1 1 X þ Xs ½kejðko0 aÞt : 2 k¼1

(3:6:25)

Example 3.6.6 Consider the periodic signal xT ðtÞ ¼ cosðo0 tÞ modulated by the same function. The resulting function is yT ðtÞ ¼ cos2 ðo0 tÞ. Determine the frequency shifts.

1 X

Ws ½k nejðknÞo0 t : k¼1 " # 1 1 X X jno0 t jðknÞo0 t yT ðtÞ ¼ Xs ½ne Ws ½k ne wT ðtÞ ¼

n¼1

k¼1

¼

1 1 X X

Xs ½nWs ½k nejno0 t ejðknÞo0 t

k¼1 n¼1

¼ ¼

"

1 X

#

1 X

Xs ½nWs ½k n ejko0 t

n¼1

k¼1 1 X

Ys ½kejko0 t :

(3:6:27)

k¼1

Solution: yT ðtÞ ¼ :5ð1 þ cosð2o0 tÞÞ ¼ :25ej2o0 t þ 0ejo0 t þ :5 þ 0ejo0 t þ :25ej2o0 t : Modulation shifted the frequencies from o0 to 2o0 ; o0 ; 0, with one of the frequencies & having zero amplitude.

Equation (3.6.26b) now follows from (3.6.27). The expression Ys ½k in terms of Xs ½k and Ws ½k in (3.6.26b) is a discrete convolution. Periodic time convolution: If xT ðtÞ and wT ðtÞ are two periodic functions and their F-series are (3.6.26a), then (see Section 2.5) 1 T

yT ðtÞ ¼

3.6.8 Multiplication in Time

xT ðtÞ¼ yT ðtÞ¼

n¼1 1 X

Xs ½nejno0 t ; wT ðtÞ¼

1 X

Ws ½mejmo0 t ;

m¼1

Ys ½kejko0 t :

(3:6:26a)

k¼1

Following shows that the complex F-series coefficients of the product yT ðtÞ ¼ xT ðtÞwT ðtÞ are

Ys ½k ¼

1 X n¼1

Xs ½nWs ½k n¼

1 X

xT ðt tÞwT ðtÞdt

T

Let xT ðtÞ and wT ðtÞ be two periodic functions with the same period T. The product yT ðtÞ ¼ xT ðtÞwT ðtÞ is also periodic with period T. The relation between the F-series coefficients is derived below. First, let the F-series expansions of these are as follows: 1 X

Z

1 ¼ T

Z

¼

Ws ½nXs ½k n :

yT ðtÞ ¼ xT ðtÞ wT ðtÞ ¼ wT ðtÞ xT ðtÞ:

(3:6:28a) (3:6:28b)

The Fourier series expansion of the periodic convolution in (3.6.28a) can be determined by using the F-series for the two functions as shown below: Z 1 yT ðtÞ ¼ xT ðt tÞwT ðtÞdt T T

1 ¼ T ¼

Z

1 X k¼1

1 X k¼1

(3:6:26b) ¼ First, use the change of variable m ¼ k n in F-series expansion of wT ðtÞ, substitute this expression in the F-series expansion of yT ðtÞ, and then simplify the expression yT ðtÞ:

Xs ½kWs ½kejko0 t :

k¼1

T

n¼1

xT ðtÞwT ðt tÞdt

T 1 X

1 X

Xs ½kejko0 ðttÞ wT ðtÞdt 2

Xs ½k4

1 T

Z

3 wT ðtÞejko0 t dt5ejko0 t

T

Xs ½kWs ½kejko0 t :

(3:6:29a)

k¼1 FS;T

yT ðtÞ ¼ xT ðtÞ wT ðtÞ ! Xs ½kWs ½k ¼ Ys ½k:

(3:6:29b)

3.6 Operational Properties of Fourier Series

95

3.6.9 Frequency Modulation The dual of time-domain modulation is frequency modulation. Using the superposition and the delay properties of the F-series, we can determine the F-series coefficients of the function yT ðtÞ ¼ xT ðt þ aÞ þ xT ðt aÞ. It follows that yT ðtÞ ¼ ¼

1 X k¼1 1 X

2

1 X

¼

3 xT ðtÞejko0 t dt5

T

Xs ½kYs ½k:

(3:6:33)

k¼1

(3:6:30a)

k¼1

Px ¼

¼)Ys ½k ¼ 2Xs ½k cosðko0 aÞ:

1 X

Z

If yðtÞ ¼ xðtÞ (i.e., Ys ½k ¼ Xs ½k), then the average power in a complex or a real periodic function with period T is

Xs ½k ½ejko0 a þ ejko0 a ejko0 t Ys ½kejko0 t :

1 Ys ½k4 ¼ T k¼1

1 T

Z

jxT ðtÞj2 dt

T

(3:6:30b) ¼

1 X

jXs ½kj2 ðParseval 0 s formulaÞ:

(3:6:34)

k¼1

3.6.10 Central Ordinate Theorems

3.6.12 Power Spectral Analysis

The following results from the F-series at t ¼ 0 and the F-series coefficient at k ¼ 0:

The power density spectrum of the periodic signal xT ðtÞ is defined by ! 1 X 2 jko0 t xT ðtÞ ¼ Xs ½ke Sx ½k ¼ jXs ½kj (3:6:35) :

xT ½0 ¼

1 X

Xs ½k;

Xs ½0 ¼

k¼1

1 T

Z

xT ðtÞdt: (3:6:31)

k¼1

T

3.6.11 Plancherel’s Relation (or Theorem) Let xT ðtÞ and yT ðtÞ be two periodic functions. Then, Plancherel’s relation is 1 T

Z

xT ðtÞyT ðtÞdt ¼

1 X

Xs ½kYs ½k:

(3:6:32)

k¼1

T

Note, for generality, the expression in (3.6.32) is given for complex functions and the superscript (*) corresponds to the conjugation. The above relation can be derived by substituting the F-series coefficients for the two time functions: 1 T

Z T

1 xT ðtÞyT ðtÞdt ¼ T

Z T

" xT ðtÞ

1 X k¼1

The average power contained in xT ðtÞ is then given by Px ¼

dt

jXs ½kj2 ¼

k¼1

1 X

Sx ½k:

(3:6:36)

k¼1

Notes: In Chapter 1, it was pointed that periodic and random signals are power signals. Although we will not be going through any discussion on random signals or processes, as it is beyond our scope, the average power contained in a random process is expressed in terms of power spectral density, which is real, even, and nonnegative function of frequency ð f Þ, identified by Sx ð f Þ. The average power in the process is expressed by

# Ys ½kejko0 t

1 X

Px ¼

Z1 1

1 Sx ðfÞdf ¼ 2p

Z1 1

Sx ðoÞdo:

(3:6:37)

96

3 Fourier Series

It is interesting to tie the average power Px in (3.6.36) and (3.6.37). This is achieved by using impulse functions (note dðfÞ ¼ 2pdðoÞ, see (1.4.37) 1 X

Sx ðoÞ ¼ 2p

jXs ½kj2 dðo ko0 Þ:

(3:6:38)

k¼1

Using this expression results in (3.6.37) 1 2p

Z1

Sx ðoÞdo

1

1 ¼ 2p ¼

Z1 " 2p

1 1 X

Solution: The nonzero F-series coefficients are pro& portional to (1/k). Notes: Before considering the convergence of F-series let us briefly summarize the theoretical constraints on the periodic function xT ðtÞ of interest and its F-series existence. It is assumed that xT ðtÞ is square integrable. That is, Z (3:7:3) jxT ðtÞj2 dt51: T

1 X

# jXs ½kj2 dðoko0 Þ do

k¼1

jXs ðkÞj2 ¼Px :

(3:6:39)

Second, the periodic function is assumed to satisfy the Dirichlet conditions, see Section 3.1. All physically realizable functions satisfy these conditions and & therefore, we will not be dealing with these.

k¼1

3.7.1 Fourier’s Theorem 3.7 Convergence of the Fourier Series and the Gibbs Phenomenon In Chapter 1, the average value, the average power, and the root mean-squared (rms) values were defined (see (1.5.15)). It was pointed out that the average value of a periodic pﬃﬃﬃﬃﬃﬃ function Xs ½0can never exceed the rms value Px . Using (3.4.23), we have Px ¼ jXs ð0Þj2 þ

1 X

jXs ½kj2

Dirichlet proved first that Fourier series approximation converges to xT ðtÞ at every point xT ðtÞ is continuous and to ½xT ðtþ Þ þ xT ðt Þ=2, the halfvalue, wherever the function xT ðtÞ is discontinuous, i.e., the F-series converges to the average value of the function. This result is called Fourier’s theorem. Example 3.7.2 Let xT ðtÞ has a discontinuity at t ¼ t0 as shown in Fig. 3.7.1. The Fourier series approximation of xT ðtÞ with ð2n þ 1Þ terms is assumed to be

k¼1;k6¼0

¼)Px jXs ½0j2 or

pﬃﬃﬃﬃﬃﬃ Px jXs ½0j:

(3:7:1)

Furthermore, from (3.4.24a), the mean-squared value of a periodic function is equal to the sum of the mean-squared values of its dc component and its harmonics.

xT;2nþ1 ðtÞ ¼

n X

Xs ½kejko0 t :

(3:7:4)

k¼n

Example 3.7.1 Consider the periodic function given in (3.5.1) with T ¼ 2p. It is discontinuous at t ¼ kp and its F-series is given below (it follows from (3.5.2)). Identify how fast the F-series coefficients decrease. 4 1 1 x2p ðtÞ ¼ ½sinðtÞ þ sinð3tÞ þ sinð5tÞ þ . . . p 3 5 1 X 4 sinðð2 k 1ÞtÞ: (3:7:2) ¼ ð2 k 1Þp k¼1

Fig. 3.7.1 xT ðtÞ with a discontinuity

The mean-squared error at the discontinuity between xT ðtÞ and its ð2n þ 1Þ term F-series is

3.7 Convergence of the Fourier Series and the Gibbs Phenomenon

97

h i Solution: Consider the approximations by consid2 þ 2 e2 ¼ fxT;2nþ1 ðtÞxT ðt 0 Þg þfxT;2nþ1 ðtÞxT ðt0 Þg : ering the first few terms. Let (3:7:5) 4 4 1 þ The time entries in the above equation t s1 ðtÞ ¼ sinðtÞ; s3 ðtÞ ¼ ½sinðtÞ þ sinð3tÞ; 0 and t0 are p p 3 the values of t before and after the discontinuity at t ¼ t0 . Find the minimum value of the meansquared error, with respect to xTð2nþ1Þ ðtÞ by taking the partial derivative of e2 with respect to xT;2 nþ1 ðtÞ, equating the result to zero, and then solving for xT;2nþ1 ðtÞ at t ¼ t0 . Solution: Taking the partial derivative and equating it to zero at t ¼ t0 , we have @e2 ¼ 2½xT;2nþ1 ðtÞ xT ðt 0 Þ @xT;2nþ1 þ 2½xT;2nþ1 ðtÞ xT ðtþ 0 Þjt¼t0 ¼ 0: (3:7:6) ) xT;2nþ1 ðtÞjt¼t0 ¼ ½xT ðtþ 0 Þ þ xT ðt0 Þ=2:

(3:7:7)

That is, the F-series approximation gives the average value of the function before and after the discontinuity. This value is referred to as the half-value & of xT ðtÞ at the discontinuity t ¼ t0 .

4 1 1 s5 ðtÞ ¼ ½sinðtÞ þ sinð3tÞ þ sinð5tÞ;...: p 3 5

(3:7:8)

The functions s1 ðtÞ; s3 ðtÞ; and sð2 k1Þ ðtÞ; k large are sketched in Fig. 3.7.2 for one period. They are odd periodic functions. Fourier series approximation gives the value of 0 at t ¼ 0, the average value (or half-value) of the function equals to ð1 1Þ=2 ¼ 0. First, consider only the positive values of t; 05t5p. The maximum value of s1 ðtÞ is equal to ð4=pÞ ¼ 1:2732. This function crosses the value of 1 when ðs1 ðtÞ 1Þ ¼ 0 for positive t. The roots of this equation are t ¼ :9033 and t ¼ 2:2383 located symmetrically around the middle t ¼ 1:5708. More number of terms we consider, the better the approximation of the given function is, and in the limit, the integral-squared error goes to zero.

Summary on s2 k1 ðtÞðt > 0Þ :

3.7.2 Gibbs Phenomenon

The function rises rapidly as t goes from 0. It overshoots the value of 1 and oscillates

From Fourier’s theorem and the following discussion we will see that at both sides of a discontinuity the finite F-series approximation exhibits ripples before and after the discontinuity. This behavior is called the Gibbs phenomenon (or effect). Historically, Albert Michalson observed the phenomenon and reported to Josiah Gibbs, a theoretical physicist. Gibbs investigated this behavior of oscillations with overshoots and undershoots before and after the discontinuity associated with Fourier series. Equality of the function to its F-series is only in the sense the integralsquared error between the two goes to zero when infinite number of terms is included in the F-series approximation. Fourier’s theorem points out that at a point of discontinuity the series converges to the average value or the half-value given in (3.7.7).

about the line xðtÞ ¼ 1 with increasing frequency and decreasing amplitudes. Although the magnitude of the peak overshoots and undershoots before and after the discontinuity at t ¼ 0 diminish as k increases, there is a lower bound of 9% on the overshoots or undershoots even as k ! 1: Furthermore, the F-series converges to every point of xT ðtÞ that is continuous with rare exceptions. It is possible that the Fourier series of a continuous function to be divergent at some point. Kolmogoroff Zygmund (1955) has given a function whose Fourier series is everywhere divergent. At the point of discontinuity in xT ðtÞ, the series converges to the half-value of the function, i.e., the average value of the function before and after the discontinuity. Since sk ðtÞ is a periodic odd function, the & sketches follow for p5t50:

Example 3.7.3 Illustrate the convergence of the F-series expansion to the function given in Example 3.7.1 by using the first few terms in the series.

98

3 Fourier Series

Fig. 3.7.2 (a)s1 ðtÞ, (b)s3 ðtÞ, and (c)s2 k1 ðtÞ, k large

s1 (t ) =

4 sin(t ) π

s3 (t ) =

4 1 sin( t) + sin(3 t) π 3

s2 k −1 (t ) =

4 1 sin( t) + sin(3 t) +... π 3 +

Notes: Gibbs published his results on the phenomenon (or effect) in Nature Magazine in 1899. Fourier did not discuss the convergence of F-series in his paper. If N is small, the value of the overshoot may be different. The Gibbs effect occurs only for waveforms with jump discontinuities, see Carslaw & (1950). Example 3.7.4 Find a series expansion for p using the results in Example 3.7.1. Solution: At t ¼ ð1=2Þp the function is continuous and therefore the series converges to the actual value of the function. Substituting t ¼ p=2 in (3.7.2), and simplifying, the expression for p is

Notes: Beckmann (1971), in his book on A History of p, gives many interesting aspects associated with & respect to the constant p. Example 3.7.5 Find a series expansion for p2 using the function in Example 3.7.3. Solution: The average power is 1 P¼ 2p

Zp p

ð1Þ2 dt¼1¼

1 1X b2 ½2k1Þ 2 k¼1

1 16 1 1 ¼ 1þ þ þ : 2 p2 9 25

1 1 ¼)p ¼8 1þ þ þ : 9 25

4 1 1 1 1 þ þ ¼ 1 p 3 5 7 1 1 1 ¼)p ¼ 4 1 þ þ : 3 5 7

1 sin((2k − 1)t ) 2k– 1

2

(3:7:9)

The series converge at a rate of 4ð1=ð2 k 1ÞÞ cor& responding to the (2k1)th term.

(3:7:10) &

How many terms in the F-series are needed to have a‘‘good’’approximation of the given function? The answer can only be given for a particular application. The integral-squared error (ISE) between the periodic function xT ðtÞ and its Fourier series approximation is

3.7 Convergence of the Fourier Series and the Gibbs Phenomenon

" # Z Z 1 X 2 2 2 ISE¼ jxT ðtÞxa ðtÞj dt¼ xT ðtÞdtT jX½kj T

T

k¼1

" # Z 1 TX 2 2 2 2 ¼ xT ðtÞdt TXs ½0þ ða ½kþb ½kÞ ¼0: 2 k¼1 T

(3:7:11) The ISE goes to zero only if an infinite number of terms are used in the expansion, which is impractical. A goal is to approximate the function xT ðtÞ using N trigonometric F-series with a bound on the ISE. First, xT;N ðtÞ ¼ Xs ½0 þ

N X

99

N¼1:

Z 1 ðISEÞ1 ¼ y22p ðtÞdt2p X2s ½0þ a2 ½1þb2 ½1 2 2p

¼ 1:571 2p½ð1=pÞ2 þ ð1=2Þð1=2Þ2 ¼ :149:

(3:7:16)

The percentage error is (.149/1.571) = 9.5%. This implies that N ¼ 1 satisfies the requirements for Part a. Continuing this procedure, we have N¼2: ðISEÞ2 ¼1:5712p½ð1=pÞ2 þð1=2Þð1=2Þ2 þð1=2Þð4=9p2 Þ¼:0075:

a½k cosðko0 tÞ

(3:7:17)

k¼1

þ

N X

b½k sinðko0 tÞ:

(3:7:12)

k¼1

Find the smallest integer value of N that results within certain percentage of ISE. The ISE, keeping only the dc term and N harmonics, is ( ) Z N 1X 2 2 2 2 ða ½kþb ½kÞ : ðISEÞN ¼ xT ðtÞdtT Xs ½0þ 2 k¼1

In this case the percentage integral-squared error is (.0075/1.571)=.5% and N ¼ 2 satisfies the requirement for part b. The coefficients die out like ðK=k2 Þ; where K is a constant. Very few terms are needed to approximate the half-wave rectified signal. Unfortunately, there is no general formula to determine the number of harmonics needed for a & given set of specifications.

T

(3:7:13) Example 3.7.6 Consider the half-wave rectified periodic function in Example 3.6.2 with sinðtÞ; 0 t5p y2p ðtÞ ¼ ; 0; p t52p yðt þ 2pÞ ¼ yðtÞ:

(3:7:14)

The trigonometric F-series of this periodic function was given by (see (3.6.10)) y2p ðtÞ ¼

1 1 þ sinðtÞ p 2 2 cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : p 1ð3Þ 3ð5Þ 5ð7Þ (3:7:15)

Find the smallest N in (3.7.13) for the two cases: a:10% or less b: 2% or less, see Gibson (1993). Solution: First Zp Z 1 p 2 y2p ðtÞdt ¼ ð1 cosð2tÞÞdt ¼ ¼ 1:571: 2 2 2p

0

The ISE can be reduced by increasing N. It goes to 0 when N ! 1. On the other hand, the peak 9% overshoots and undershoots before and after the discontinuities discussed earlier cannot be reduced even if infinite number of terms is included in the F-series expansion. There is another measure, average error, which can be used in judging an approximation, which is not very attractive as the positive errors may cancel out with the negative errors. Squaring the error function accentuates the larger errors. Least-squares error measure gives a convenient and a simple way to calculate the parameters in the approximation. Overshoots and undershoots can be reduced by smoothing.

3.7.3 Spectral Window Smoothing The ripples generated by the Fourier series approximation of a discontinuous function are due to the abrupt change of the function before and after the discontinuity. The use of a taper, instead of a discontinuity at a transition, yields a smoother

100

3 Fourier Series

reconstruction from the basis functions. The windowed signal is defined by the periodic convolution (see Section 2.5) by yT ðtÞ ¼ xT ðtÞ wT ðtÞ Z 1 xT ðaÞwT ðt aÞda: ¼ T

Ws;H ½k¼

:0800;:1256;:2532;:4376;:6424;:8268;:9544;1; :9544;:8268;:6424;:4376;:2532;:1256;:0800

ðy2p ðtÞÞjN¼7 (3:7:18a)

T

:

4 1 ¼ ð:9544Þ sinðtÞ þ ð:6424Þ sinð3tÞ p 3 1 1 þ ð:2532Þsinð5tÞþ ð:0800Þsinð7tÞ 5 7 (3:7:22)

Considering (2N+1) complex F-series coefficients of yT ðtÞ (see (3.6.29a)), we have yT;N ðtÞ ¼

N X

Ws ½kXs ½kejko0 t :

(3:7:18b)

k¼N

The sequence Ws ½k is a window and its weights (or coefficients) typically decrease with increasing jkj. The rectangular and hamming window sequences are

The two functions x2p ðtÞ (identified as xðtÞ on the top figure) and y2p ðtÞ (identified as yðtÞ on the bottom figure) are plotted in Fig. 3.7.3 using MATLAB. Note the overshoots and undershoots before and after the discontinuities in each case. The later case has hardly any ripples. The slope in the transition region is much higher in the case of the rectangular window com& pared to the Hamming window case.

3.8 Fourier Series Expansion of Periodic Functions with Ws;H ½k ¼ :54 þ :46 cosðkp=NÞ; Special Symmetries N k N: (3:7:20)

Ws;R ½k ¼ 1; N k N:

(3:7:19)

The use of special windows reduces or even eliminates the overshoots and undershoots in the approximated signal before and after a discontinuity, see Ambardar (1995). Example 3.7.7 Consider the trigonometric F-series in (3.7.2). Give the expression using the rectangular window with N ¼ 7. Illustrate the window smoothing by first sketching the F-series and then the Hamming windowed series. Solution: The trigonometric F-series approximation is 4 1 1 1 ðx2p ðtÞÞjN¼7 ¼ sinðtÞþ sinð3tÞþ sinð5tÞþ sinð7tÞ : p 3 5 7 (3:7:21)

Note the odd harmonic terms are all zero. The 15 tapered Hamming window coefficients and the Hamming windowed function y2p ðtÞ are, respectively, given by

In Section 1.6, periodic functions with half-wave and quarter-wave symmetries were considered. Computation of the F-series for these cases is considered next.

3.8.1 Half-Wave Symmetry Figure 3.8.1 illustrates a periodic function with period T and with half-wave symmetry (or rotation symmetry). Such functions satisfy (see (1.6.19) and (1.6.20))

T : (3:8:1) xT ðtÞ ¼ xT t 2 Periodic functions with half-wave symmetry have odd harmonics, i.e., Xs ½k ¼ 0;k even, which can be seen from the following. First Xs ½k ¼

1 T

Z T

xT ðtÞejko0 t dt

3.8 Fourier Series Expansion of Periodic Functions with Special Symmetries Fig. 3.7.3 Window smoothing

101 Use of rectangular window

2

x(t)

1 0 –1 –2

0

1

2

0

1

2

3

4 5 t Use of Hamming window

6

7

8

6

7

8

2

y(t)

1 0 –1 –2

3

1 T

4 t

5

ZT=2

T xT ða Þejko0 a ejko0 T=2 da 2 3 2 0 ZT=2 T 7 61 ¼4 xT ða Þejko0 a da5ejkp T 2 0

¼ð1Þ

k1

Z0

1 xT ðtÞejko0 t dtþ T

ZT=2

(3:8:4)

Using (3.8.4) in (3.8.3), we have xT ðtÞejko0 t dt: (3:8:2)

0

T=2

T

T xT ðt Þejko0 t dt: 2

0

Fig. 3.8.1 A half-wave symmetric function

1 ¼ T

ZT

1 Xs ½k ¼ T

ZT=2

T ejko0 t dt: xðtÞ þ ð1Þk x t 2

0

Consider the change of variable from t to a ðT=2Þ in the first integral on the right in (3.8.2), which results in 1 Xs ½k ¼ T

ZT=2

T T xT ða Þejko0 ða 2 Þ da 2

0

1 þ T

ZT=2

xT ðtÞejko0 t dt:

From the half-wave symmetry property in (3.8.1) we see that Xs ½k ¼ 0, k even, thus establishing the halfwave symmetric functions that contain only odd harmonics. The complex F-series of these functions and the corresponding trigonometric F-series are

(3:8:3)

0

The limits follow from and t ¼ T=2¼)a ¼ 0 and

(3:8:5)

t ¼ 0¼)a ¼ T=2

xT ðtÞ¼

1 X k¼1 k6¼0;kodd

Xs ½ke

jko0 t

2 ;Xs ½k¼ T

ZT=2

xT ðtÞejko0 t dt:

0

(3:8:6)

102

3 Fourier Series

xT ðtÞ ¼

1 X

fa½2 k 1 cosðð2 k 1Þo0 tÞ þ b½2 k 1

k¼1

ZT=4

4 ¼ T

0

sinðð2 k 1Þo0 tÞg; 4 a½2 k 1 ¼ T

ZT=2

þ xT ðtÞ cosðð2 k 1Þo0 tÞdt

0

4 b½2 k 1 ¼ T

ZT=2

xT ðtÞ cosðð2 k 1Þo0 tÞdt Z0

4 T

xT ðtÞ cos½ð2 k 1Þo0 tdt:

T=4

:

Trigonometric and complex F-series for the even quarter-wave symmetric function:

xT ðtÞ sinðð2 k 1Þo0 tÞdt

0

(3:8:7)

xT ðtÞ ¼

1 X

a½2 k 1 cos½ð2 k 1Þo0 t;

n¼1

8 a½2 k 1 ¼ T

3.8.2 Quarter-Wave Symmetry

ZT=4

xT ðtÞcosðð2 k 1Þo0 tÞdt:

0

If xT ðtÞ is a periodic function with half-wave symmetry and, in addition, is either even or an odd function, then xðtÞ is said to have even or odd quarter-wavesymmetry, respectively, see Section 1.6.4. They satisfy the following properties:

(3:8:10) 1 X

¼

k¼1

þ

1 X

Even quarter-wave symmetry : xT ðtÞ ¼ xT ðtÞ and xT ðtÞ ¼ xT ðt T=2Þ:

ð1=2Þa½2 k 1ejð2 k1Þo0 t

(3:8:8a)

¼

ð1=2Þa½2 k 1ejð2 k1Þo0 t

k¼1 1 X

0t Xs ½kjko e

k¼1

Odd quarter-wave symmetry : xT ðtÞ ¼ xT ðtÞ and xT ðtÞ ¼ xT ðt T=2Þ:

(3:8:8b)

¼)Xs ½2 k 1 ¼ Xs ½ð2 k 1Þ ¼ a½2 k 1=2:

(3:8:11)

3.8.3 Even Quarter-Wave Symmetry

3.8.4 Odd Quarter-Wave Symmetry

Since the function must be a half-wave symmetric to be a quarter-wave symmetric, it follows that Xs ½0 ¼ 0 and a½2 k ¼ 0: In addition, xðtÞ is even and b½k ¼ 0. Therefore,

The F-series for this case are as follows. Derivation is left as an exercise. 1 X b½2 k 1 sin½ð2 k 1Þo0 t; xT ðtÞ ¼

Xs ½0 ¼ 0; b½k ¼ 0; a½2 k 1 ¼

4 T

ZT=2

xT ðtÞ cosðð2 k 1Þo0 tÞdt; a½2 k ¼ 0:

ZT=4

(3:8:9a) xT ðtÞcosðð2k1Þo0 tÞdt

xT ðtÞ¼

T=4

xT ðtÞ sin½ð2 k 1Þo0 tdt: (3:8:12) 0

1 X

X½2k1ejð2k1Þo0 t

k¼1 1 X

þ

X½ð2k1Þejð2k1Þo0 t :

(3:8:13)

k¼1

0

4 þ T

8 T

b½2 k 1 ¼

0

4 a½2k1¼ T

k¼1

Z

T=2 Z

T=4

xT ðtÞcosðð2k1Þo0 tÞdt: (3:8:9b)

Compare this with the F-series expansion and equate the corresponding coefficients.

3.9 Half-Range Series Expansions

xT ðtÞ ¼

1 X

103

Xs ½nejno0 t

n¼1; n6¼0

¼)X½2 k 1 ¼

1 X

xeT ðtÞ ¼

xe ðt þ kTÞ and

k¼1

b½2 k 1 ; X½ð2 k 1Þ 2j

1 X

x0T ðtÞ ¼

x0 ðt þ kTÞ:

(3:9:2)

k¼1

b½2 k 1 ¼ : 2j

(3:8:14) The trigonometric F-series of the even periodic function has dc and cosine terms and the odd periodic function has only sine terms. The two periodic functions xeT ðtÞ and x0T ðtÞ can be expressed by the following with o0 ¼ 2p=T (see (3.4.18)):

3.8.5 Hidden Symmetry Example 3.8.1 Symmetry of a periodic function can be obscured by a constant. Consider the periodic saw-tooth waveform, x2p ðtÞ ¼ ð1 ðt=2pÞÞ; 05t52p in Fig. 3.8.2. It does not have any obvious symmetry. Find the F-series of the function x2p ðtÞ by noting ðx2p ðtÞ ð1=2ÞÞ is an odd function.

xeT ðtÞ ¼ Xs ½0 þ

1 X

a½k cosðko0 tÞ:

(3:9:3a)

xðtÞ cosðko0 tÞdt:

(3:9:3b)

k¼1

2 Xs ½0 ¼ T

ZT=2 xðtÞdt; 0

2 a½k ¼ ðT=2Þ

ZT=2 0

x0T ðtÞ ¼

1 X

b½k sinðko0 tÞ;

k¼1

Fig. 3.8.2 xT ðtÞ with hidden symmetry, T = 2p

4 b½k ¼ T

ZT=2 xðtÞ sinðko0 tÞdt:

(3:9:4)

0

Solution: From the odd symmetry,

X 1 1 x2p ðtÞ b½k sinðko0 tÞ ¼ 2 k¼1 1 1 X b½k sinðko0 tÞ : ¼)x2p ðtÞ ¼ þ 2 k¼1

(3:8:15) &

3.9 Half-Range Series Expansions Consider an aperiodic function x(t) over the interval ð0; T=2Þ and zero everywhere else. Even and odd functions can be generated in the interval T=25t5T=2 by xe ðtÞ ¼ xðtÞ þ xðtÞ; x0 ðtÞ ¼ xðtÞ xðtÞ: (3:9:1) Even and odd periodic extensions (see Section 1.8.1.) of these are

The functions xeT ðtÞ and x0T ðtÞ in (3.9.3a and b) and (3.9.4) represent the same function in the interval (0, T/2). Outside this interval (3.9.3a) represents an even periodic function and (3.9.4) represents an odd periodic function. These expansions are called the half-range Fourier series expansions of the aperiodic function xðtÞ. Example 3.9.1 Given the aperiodic function xðtÞ ¼ sinðtÞ; 05t5p and 0 otherwise, expand this function in terms of a cosine series expansion and the sine series expansion in the interval 05t5p. Give the even and odd periodic extensions of xðtÞ. Solution: It is simple to see that xeT ðtÞ ¼ jsinðtÞj and the odd periodic extension x0T ðtÞ ¼ sinðtÞ. In the interval 05t5p, the two functions xeT ðtÞ and x0T ðtÞ are equal. Since jsinðtÞj and sinðtÞ are continuous functions, the F-series converges and (see (3.6.7))

104

3 Fourier Series

xeT ðtÞ ¼ jsinðtÞj 2 4 cosð2tÞ cosð4tÞ þ þ ; 05 t 5 p : ¼ p p ð1Þð3Þ ð3Þð5Þ (3:9:5) x0T ðtÞ ¼ sinðtÞ; 05t5p:

(3:9:6) &

Example 3.9.2 Expand the function given below in terms of a cosine series expansion in the interval 05t5p. Give the even periodic extensions of xðtÞ. xðtÞ ¼

0; 05t5p=2 : 1; p=25t5p

Zp

p=4

xeT ðtÞ cosðktÞdt

0

¼

2 p

Zp

cosðktÞdt ¼

sinðktÞdt p=2

¼

2 ½cosðkpÞ cosðð1=2ÞkpÞ: kp

2 1 1 sinðtÞ þ sinð3tÞ þ sinð5tÞ þ x0 ðtÞ ¼ p 3 5 2 1 1 sinð2tÞ þ sinð6tÞ þ sinð10tÞ þ : p 3 5 &

(3:9:7)

2 1 ¼ ½ðp=2Þ ðp=4Þdt ¼ : p 2 2 a½k ¼ p

Zp

2 p

05t5p:

Solution: Cosine series need a symmetric function defined by xe ðtÞ ¼ xðtÞ þ xðtÞ. The F-series can now be obtained by considering the interval p5t5p, i.e., the period T ¼ 2p. Noting that o0 ¼ 2p=T ¼ 1, the trigonometric F-series can be computed. Since it is even, it follows that b½k ¼ 0. The coefficients Xs ½0 and a½k are, respectively, given by Zp Zp=2 1 2 xeT ðtÞdt ¼ dt X½0 ¼ 2p p p

b½k ¼

2 sinðkp=2Þ: kp

p=2

1 2 1 1 xeT ðtÞ ¼ cosðtÞ cosð3tÞ þ cosð5tÞ ; 2 p 3 5 05t5p: This gives an approximation of xðtÞ in terms of cosine series in the time interval 05t5p. In a similar manner, the odd periodic extension of the function can be determined. Since the function is odd, it follows that Xs ½0 ¼ 0 and a½k ¼ 0; k ¼ 0; 1; 2; . . .. The coefficients b½k and the corresponding odd period extension are given by

3.10 Fourier Series Tables Refer tables 3.10.1 and 3.10.2 for Fourier Series.

3.11 Summary In this chapter we have introduced some of the basis functions and their use in approximating a given function in an interval. The important set of basis functions are the sine and the cosine functions and the periodic exponential function leading to the discussion on Fourier series. This chapter dealt with Fourier series, their properties and the computations of the Fourier series coefficients, in general, and in cases of special symmetries in the given function. Convergence of the coefficients and the number of coefficients required for a given set of specifications are discussed. Approximation measures are discussed in terms of basis functions. Specific principal topics that were included are

Various basis functions and error measures of a function and their approximations

Basics of complex and trigonometric Fourier series and the relationships between the trigonometric and the complex F-series Computation and simplification of the F-series coefficients of periodic functions that have simple and special symmetries Operational properties of the Fourier series that include simple methods that allow for simplification in the computation of the F-series

3.11 Summary

105 Table 3.10.1 Symmetries of real periodic functions and their Fourier-series coefficients

Type of symmetry

Constraints Periodic, xT ðtÞ ¼ xT ðt þ TÞ; o0 ¼ 2p=T:

Trignometric Fourier-series

Even

xT ðtÞ ¼ xT ðtÞ

xT ðtÞ ¼ XS ½0 þ

1 P

Fourier series coefficients

a½k cosðko0 tÞ

XS ½0 ¼ T2

k¼1

a½k ¼ T4 xT ðtÞ ¼ xT ðtÞ

Odd

x T ð tÞ ¼

1 P

b½k ¼ T4

b½k sinðko0 tÞ

k¼1

xT ðtÞ ¼ xT ðt þ T=2Þ

Half-wave

xðtÞ ¼

1 P k¼1

þ

b½2 k 1 sinð2 k 1Þo0 t

k¼1

Even quarterwave

xT ðtÞ ¼ xT ðtÞ; xT ðtÞ ¼ xT ðt þ T=2Þ

x T ð tÞ ¼

1 P

a½2 k 1 cosð2 k 1Þo0 t

k¼1

0 T=2 R

0 T=2 R

xT ðtÞdt

xT ðtÞ cosðko0 tÞdt

xT ðtÞ sinðko0 tÞdt

0

a½2 k 1 ¼ T4

a½2 k 1 cos½2 k 1o0 t

1 P

T=2 R

b½2 k 1 ¼ T4 a½2 k 1 ¼ T8

T=2 R 0 T=2 R 0 T=4 R

xT ðtÞcosðð2 k 1Þo0 tÞdt xT ðtÞ sinðð2 k 1Þo0 tÞdt xT ðtÞcosðð2 k 1Þo0 tÞdt

0

T=4 1 R P b½2 k 1 sinð2 k 1Þo0 t Odd xT ðtÞ ¼ xT ðtÞ; x T ð tÞ ¼ b½2 k 1 ¼ T8 xT ðtÞsinðð2 k 1Þo0 tÞdt k¼1 0 quarterxT ðtÞ ¼ xT ðt þ T=2Þ wave There are extensive tables in literature that list Fourier series of functions. See Abramowitz and Stegun (1964), Gradshteyn and Ryzhik (1980), and others. To standardize the tables, we will assume the period is T ¼ 2p. Spiegel (1968) has several other interesting periodic functions and the corresponding Fourier series.

Table 3.10.2 Periodic functions and their Trigonometric Fourier Series Periodic Function x2p ðtÞ ¼ x2p ðt þ 2pÞ Trigonometric Fourier-series h i 5 5 sinð3tÞ sinð5tÞ 1; 0 t p 4 sinðtÞ þ þ þ x2p ðtÞ ¼ p 1 3 5 1; p5t50 x2p ðtÞ ¼ jtj ¼

t; t;

05t5p p5t50

2

x2p ðtÞ ¼ t2 ; p5t5p x2p ðtÞ ¼ jsinðtÞj; p5t5p sinðtÞ; 05t5p 0; p5t52p x2p ðtÞ ¼

cosðtÞ; cosðtÞ;

x2p ðtÞ ¼ et ; p5t5p

p4

05t5p p5t50

h

h

cosðtÞ 12

þ cos3ð23tÞ þ cos5ð25tÞ

i

i sin2ð2tÞ þ sin3ð3tÞ þ h i cosðtÞ cosð2tÞ cosð3tÞ p2 4 þ 2 2 3 1 2 3

x2p ðtÞ ¼ t; p5t5p

x2p ðtÞ ¼

p 2

sinðtÞ 1

h

cosð2tÞ ð1Þð3Þ

ð4tÞ cosð6tÞ þ cos ð3Þð5Þ þ ð5Þð7Þ þ

2 p

p4

1 p

þ 12 sinðtÞ p2 h

8 sinð2tÞ p ð1Þð3Þ

h

cosð2tÞ ð1Þð3Þ

i

ð4tÞ cosð6tÞ þ cos ð3Þð5Þ þ ð5Þð7Þ þ

i ð4tÞ 3 sinð6tÞ þ 2ðsin þ þ 3Þð5Þ ð5Þð7Þ

2 sinh p 1 p 2

þ

1 P k¼1

ð1Þk ðcosðktÞ ð1þk2 Þ

k sinðktÞÞ

i

106

3 Fourier Series

Bounds and convergence of the F-series to the given function Half-range expansions

Problems 3.1.1 The set of functions fi ðtÞ; i ¼ 1; 2; 3; 4 shown in Fig. P3.1.1 are a set of Walsh functions. Show that they are orthogonal in the interval [0, 1]. 3.1.2 Consider the set ff1 ðtÞ; f2 ðtÞg with f1 ðtÞ ¼ 1 and f2 ðtÞ ¼ cð1 2tÞ. Is this an orthogonal set in the interval [0, 1]? If so, compute the value of c that makes the functions f1 ðtÞ and f2 ðtÞ become an orthonormal basis set. 3.1.3 The function xðtÞ ¼ sinðtÞ is approximated by xðtÞ ¼ c1 f1 ðtÞ þ c2 f2 ðtÞ. Use the results in Problem 3.1.2 and find the constants c1 and c2 so that the

1 2

1 4

3 4

t

1

-1

t 1 2

3 4

1

(b) φ3 (t ) 1 0 -1

1 2

1 t 3 4

1 4

(c)

-1

3.2.1 Use the Walsh functions given in Problem 3.1.1 to approximate the following function and find the mean-squared error between the given function and approximation: xðtÞ ¼

t;

05 t 5 1

0;

otherwise

:

The Lp error between the two sides of the above equation is Ep ¼ jð1 aÞjp þjð2 aÞjp . If p ¼ 2; we call that as a least-squares error or L2 error and for p ¼ 1, we call that as the L1 error. Find the value of a that minimizes the L1 and L2 errors. The leastsquares error can be computed by taking the partial of E2 with respect to a, equating it to zero and solving for a, see Section A.8.1. A simple way to solve the L 1 error problem is solve each equation and find the value of a out of the two solutions that gives the minimum L 1 error. For iterative Lp solutions, see Yarlagadda, Bednar and Watt (1986). 3.3.1 Determine the complex Fourier series of the function

φ4 (t ) 1 0

0; 15t50 : 1; 05t51

1 ¼ a: 2 1

φ2 (t ) 1 1 4

xðtÞ ¼

1

(a)

0

3.1.5 Use the first five Legendre polynomials to approximate the following function:

3.2.2 Consider the equations given in matrix form given below. There is no value of a that satisfies the set of equations given below. The system is called an overdetermined system of equations:

φ 1 (t ) 1 0

mean-squared error is minimized between the given function and the approximated function. pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 3.1.4 Show the set fk ðtÞ ¼ ð2=TÞ sinðkp=TÞt; k ¼ 1; 2; 3; . . . is an orthogonal set in the interval (0, T).

1 4

1 2

(d) Fig. P3.1.1 Walsh functions

3 4

1

t

xT ðtÞ ¼ cosð2pf0 t 1Þ þ sinð2pð2f0 Þt 2Þ: 3.3.2 a. Determine the period of the function 1 X 1 xT ðtÞ ¼ ejð3pkt=2Þ : ða þ jpkbÞ k¼1

Problems

107

b. What is the average value of the periodic function xT ðtÞ? c. Determine the amplitude and phase values of the third harmonic.

Derive an expression for p using the F-series. See Example 3.7.4. What is the value of the F-series at t ¼ 0? Compare this to the actual value of the function at that location.

3.3.3 Expand the following periodic functions with xi ðtÞ ¼ xi ðt þ 2pÞ in trigonometric F-series:

3.5.2 Using the Fourier series expansion, determine the sum A identified below

a: x1 ðtÞ ¼ eat ; p5t5p; a is some constant

b: x2 ðtÞ ¼ cosh at; p t p: c: x3 ðtÞ ¼ tðp tÞ; 05t5p; cosðtÞ; 05t5p : d: x4 ðtÞ ¼ cosðtÞ; p5t50

xðtÞ ¼ cosh at ¼ p2 sinhðapÞ 1 P k a 1 þ ð1Þ cosðktÞ ; p t p 2 2 2a k þa k¼1

1 þ A ¼ 2a

1 P k¼1

:

1 ð1Þk k2 þa 2

function

3.5.3 Find the F-series expansion of the function yðtÞ ¼ sinhðatÞ using the above results.

3.3.5 Show that the equation in (3.3.11) reduces to the equation in (3.3.10). 3.4.1 Use the derivative method to determine the trigonometric Fourier series of the periodic function x2p ðtÞ ¼ t; 0 t52p.

3.6.1 Consider the full-wave rectified function xðtÞ ¼ jsinðtÞj; p5t5p. Estimate the rms value of the full-wave rectified signal by using the first four nonzero terms in the Fourier series representation of the function. Calculate the percentage of error in the estimation.

3.4.2 Find the trigonometric F-series of the function shown below using a. derivative method. b. What can you say about the convergence of its F-series coefficients? 2 cos½pt=2; 15t51 xT ðtÞ ¼ : 0; 15jtj5T=2

3.6.2 Consider the triangular wave function given below and derive the trigonometric F-series of this function. ( T5 t0 1 þ 4t T ; 2 xT ðtÞ ¼ ; xT ðt þ TÞ ¼ xT ðtÞ: 4t 1 T ; 0 t5 T2

c. What can you say about the convergence of the trigonometric F-series coefficients?

3.6.3 Show that

3.3.4 Find the F-series of x1 ðtÞ ¼ At2 þ Bt þ C; p5t5p.

the

3.4.3 Use the generalized derivatives of the function given in Problem3.4.2 and see how fast the series converge without actually computing the Fourier series and verify the results obtained in that problem. 3.4.4 Show that jXs ½kj ¼ jXs ½kj for a real periodic function xT ðtÞ. Give a complex periodic function where this is not true. 3.5.1 The function and its trigonometric F-series expansion are given by x2p ðtÞ ¼ jsinðtÞj; 0 t52p; 2 4 jsinðtÞj ¼ p p cosð2tÞ cosð4tÞ cosð6tÞ þ þ þ : 3 15 35

1 T 1 T

Z

Z

xT ðtÞxT ðt tÞdt ¼

T

1 xTe ðtÞxTe ðt tÞdt þ T

T

Z

xT0 ðtÞxT0 ðt tÞdt:

T

3.6.4 Show the integral of xT ðtÞ with a nonzero average value is non-periodic. 3.6.5 Show the derivative of an even function is an odd function and the derivative of an odd function is an even function. Verify this property using the trigonometric F-series. 3.7.1 Using the F-series in Table 3.10.2, determine 1 X A¼ ½1=ð2 k 1Þ2 : k¼1

108

3 Fourier Series

3.7.2 Verify the Fourier series of the periodic function x2p ðtÞ ¼ t; p5t5p are

g2p ðtÞ ¼

t ðp=2Þ; t ðp=2Þ;

05 t 5 p ; p5t50

g2p ðtÞ ¼ g2p ðt þ 2pÞ: x2p ðtÞ ¼ 2

1 X

½ð1Þk1 =k sinðktÞ:

k¼1

What value does this function converges to at t ¼ 0; ðp=2Þ; p?

3.8.4 Give the complex F-series of the full-wave rectified signal x2p ðtÞ ¼ jcosðtÞj. Use the Fourier series of the results on the full-wave rectified sine wave in Table 3.10.2.

3.7.3 Give the convergence rate of the F-series coefficients of the periodic functions without actually using the F-series.

3.9.1 Verify the trigonometric Fourier series given in Table 3.10.2 for the following periodic functions in the range p5t5p.

a:x1 ðtÞ ¼ t 1; 05t51; x1 ðt þ TÞ ¼ x1 ðtÞ; sinðtÞ; 05t5p : T ¼ 1; b:x2p ðtÞ ¼ 0; p5t52p

a: x2p ðtÞ ¼ t2 ; b: x2p ðtÞ ¼ et ; c: x2p ðtÞ ¼ cosðatÞ:

3.8.1 Derive the expressions for the coefficients b½k in (3.8.13). 3.8.2 Consider the periodic function x2p ðtÞ below and identify any symmetry this has. x2p ðtÞ ¼

1; 1;

05 t 5 p ; x2p ðtÞ ¼ x2p ðt þ 2pÞ: p5t50

3.9.2 Use the results in Table 3.10.2 and the derivative method to derive the F-series of the periodic function x2p ðtÞ ¼ sinðatÞ; p5t5p. 3.9.3 Derive an expression for the F-series of the function x2p ðt pÞ; p5t5p given

x2p ðtÞ ¼ Xs ½0 þ

1 X

a½k cosðko0 tÞ

k¼1

þ

1 X

b½k sinðko0 tÞ; p5t5p:

k¼1

a. Use that to derive the trigonometric F-series of this function. Give the corresponding complex F-series. b. Use the derivative method to derive the trigonometric F-series. 3.8.3 Consider the periodic triangular wave function below. Does this function have any symmetry? Derive the trigonometric F-series of this function.

3.9.4 Consider the functions xðtÞ ¼ cosðtÞ and yðtÞ ¼ sinðtÞ over the interval 05t5p and 0 otherwise. Expand these functions using the Fourier sine series and cosine series by directly going through the procedure discussed in Section 3.9. Can you think of a simpler method knowing the results given in Example 3.9.1?

Chapter 4

Fourier Transform Analysis

4.1 Introduction In Chapter 3 we have discussed the frequency representation of a periodic signal. Fourier series expansions of periodic signals give us a basic understanding how to deal with signals in general. Since most signals we deal with are aperiodic energy signals, we will study these in terms of their Fourier transforms in this chapter. Fourier transforms can be derived from the Fourier series by considering the period of the periodic function going to infinity. Fourier transform theory is basic in the study of signal analysis, communication theory, and, in general, the design of systems. Fourier transforms are more general than Fourier series in some sense. Even periodic signals can be described using Fourier transforms. Most of the material in this chapter is standard (see Carlson, (1975), Lathi, (1983), Papoulis, (1962), Morrison, (1994), Ziemer and Tranter, (2002), Haykin and Van Veen, (1999), Simpson and Houts, (1971), Baher, (1990), Poulariskas and Seely, (1991), Hsu, (1967, 1993), Roberts, (2007), and others).

The frequency f0 ¼ o0 =2p is the fundamental frequency of the signal, which is the inverse of the period of the signal f0 ¼ ð1=TÞ. The Fourier series coefficients are complex in general. To make the analysis simple we assume the signal under consideration is real and the amplitude of the Fourier coefficients is given by jXs ½kj. Figure 4.2.1b gives the sketch of the amplitude line spectra of the complex Fourier series of a periodic function shown in Fig. 4.2.1a. The frequencies are located at ko0 ¼ 2pðkf0 Þ; k ¼ 0; 1; 2; : : : and the frequency interval between the adjacent line spectra is o0 ¼ 2pf0 ¼ 2p=T: In this example, we assumed t=T ¼ 1=5. As T ¼ 2p=o0 ! 1, o0 goes to zero and the spectral lines merge. To quantify this, let

xT (t)

4.2 Fourier Series to Fourier Integral Consider a periodic signal xT ðtÞ with period T and its complex Fourier series xT ðtÞ ¼

1 X

X s [k ]

Xs ½ke jko0 t ;

k¼1

Xs ½k ¼

1 T

ZT=2

xT ðtÞejko0 t dt;

o0 ¼ 2p=T:

T=2

(4:2:1)

Fig. 4.2.1 (a) xT ðtÞ and (b) jXs ðkÞj

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_4, Ó Springer ScienceþBusiness Media, LLC 2010

109

110

4 Fourier Transform Analysis

us extract one period of the periodic signal by defining xðtÞ ¼

xT ðtÞ;

T2 5t5 T2

0;

Otherwise

:

(4:2:2)

"

1 1 X Xðko0 Þe jðko0 Þt o0 xðtÞ ¼ lim T!1 2p k¼1

1 ¼ 2p

Z1

XðjoÞejot do:

#

(4:2:7)

1

We can consider the function in (4.2.2) as a periodic signal with period equal to 1. In the expression for the Fourier series coefficients in (4.2.1), k and o0 appear as product ðko0 Þ ¼ ðkð2pÞ=TÞ. As T ! 1, the expression for the Fourier series coefficients in (4.2.1) results in a value equal to zero, which does not provide any spectral information of the signal. To avoid this problem, define

We now have the Fourier transform of the time function xðtÞ, XðjoÞ; and its inverse Fourier transform. Some authors use XðjoÞ instead of XðoÞ for the Fourier transform, indicating the transform is a function of complex variable ðjoÞ. The pair of functions xðtÞ and XðjoÞ is referred to as a Fourier transform pair:

XðjoÞ ¼ F½xðtÞ ¼ ZT=2

Xðko0 Þ ¼ TXs ½k ¼

(4:2:3) xðtÞ ﬃ x~ðtÞ ¼ F

Note that k is an integer and it can take any integer value from 1 to 1. Furthermore, as T ! 1, o0 ¼ ð2p=TÞ becomes an incremental value, ko0 becomes a continuous variable o on the frequency axis, and Xðko0 Þ becomes XðoÞ. From this, we have the analysis equation for our single pulse:

T!1

ZT=2 xðtÞe

T!1

jko0 t

dt

T=2

¼

Z1

Xs ½ke jko0 t ;

(4:2:4)

T=25t5T=2; (4:2:5)

k¼1

xT ðtÞ ¼

1

1 ½Xð joÞ ¼ 2p

Z1

XðjoÞe jot do; (4:2:8b)

FT

(4:2:8c)

xðtÞ !Xð joÞ:

The transform and its inverse transform can be written in terms of frequency variable f in Hertz instead of o ¼ 2pf: Xð jf Þ ¼

Z1 xðtÞe

j2pft

dt;

xðtÞ ¼

Z1

Xð jf Þej2pft df:

1

(4:2:8d)

xðtÞejot dt:

Now consider the synthesis equation in terms of the Fourier series in the forms 1 X

(4:2:8a)

1

1

1

xT ðtÞ ¼

xðtÞejot dt;

1

xðtÞejko0 t dt:

T=2

XðjoÞ ¼ lim ðTXs ½kÞ ¼ lim

Z1

1 1 X Xðko0 Þe jko0 t : T k¼1

(4:2:6)

One can see that as T ! 1, ko0 ! o; a continuous variable, and o0 ¼ 2p=T ! do, an incremental value and the summation becomes an integral. These result in

Equation (4.2.8c) shows that the transform and its inverse have the same general form, one has the time function and an exponential term with negative exponent and the other has the transform and an exponential term with a positive exponent in the corresponding integrands. F-transforms are applicable for both real and complex functions. Most practical signals are real signals. The transforms are generally complex. Integrals in the transform and its inverse are with respect to a real variable. The relations between the Laplace transforms, considered in Chapter 5, and the Fourier transforms become evident with this form. The form in (4.2.8d) is adopted by the engineers in the communications area. The transforms are computed by integration and the inverse transforms are determined by using

4.2 Fourier Series to Fourier Integral

111

transform tables. Since the Fourier transform is derived from the Fourier series, we can now say that

xT ðtÞ ¼ xðtÞ; jtj5t=2; xT ðtÞ ¼

1 X

xðt þ nTÞ:

n¼1 FT

Inverse FT

xT ðtÞ ¼ AP½t=t; jtj5T=2; xT ðtÞ ¼ xT ðt þ TÞ; (4:2:10) yT ðtÞ ¼ xT ðt ðt=2ÞÞ:

xðtÞ ! XðjoÞ ! x~ðtÞ ’ xðtÞ:

The inverse transform of XðjoÞ; F1 ½XðjoÞ identified by x~ðtÞ is an approximation of xðtÞ and it may not be the same as the function xðtÞ. We will have Solution: The complex Fourier series coefficients of xT ðtÞ were given in Example 3.4.2. The Fourier more on this shortly. series coefficients and their amplitudes of the two Fourier transform is applicable to signals that functions are given by obey the Dirichlet conditions (see Section 3.1), with the exception now that xðtÞ must be absosinðko0 t=2Þ Xs ½k ¼ Aðt=TÞ ; lutely integrable over all time, which is a sufficient ðko0 t=2Þ but not a necessary condition. Periodic functions (4:2:11) Ys ½k ¼ Xs ½kejko0 ðt=2Þ ; o0 ¼ 2p=T: violate the last condition of absolute integrability over all time and will be considered in a later Atsinðko0 ðt=2ÞÞ Y ½ k ¼ X ½ k ¼ j j j j section. There are many functions that do not s s T ðko t=2Þ : 0 have Fourier transforms. For example, the Fourier transform of the function eat ; a40 is not defined. By using the complex Fourier series in (4.2.4), the The functions that can be generated in a laboratory transforms of the two pulses xðtÞ and yðtÞ are given have Fourier transforms. Existence of the Fourier below: transforms will not be discussed any further. In the yðtÞ ¼ AP½ðt t=2Þ=t; (4:2:12) synthesis equation, the inverse transform given by (4.2.8b) is an approximation of the function xðtÞ. sinðot=2Þ joðt=2Þ lim TYs ½k ¼ At ; (4:2:13a) e The function xðtÞ and its approximation x~ðtÞ are YðjoÞ ¼ T!1 ðot=2Þ equal in the sense that the error eðtÞ ¼ ½xðtÞ x~ðtÞ sinðot=2Þ is not equal to zero for all t and may differ signifi: (4:2:13b) XðjoÞ ¼ lim TXs ½k ¼ At T!1 cantly from zero at a discrete set of points t, but ðot=2Þ Z1

Obviously it is simpler to compute the transform directly. For example, jeðtÞj2 dt ¼ 0:

(4:2:9)

1

In the sense that the integral squared error is zero, the equality xðtÞ ¼ x~ðtÞ between the function and its approximation is valid. Physically realizable signals have F-transforms and when they are inverted, they provide the original function. Physically realizable signals do not have any jump discontinuities. That is x~ðtÞ ¼ xðtÞ. If the function xðtÞ has jump discontinuities, then the reconstructed function x~ðtÞ exhibits Gibbs phenomenon. The reconstructed function converges to the halfpoint at the discontinuity and will have overshoots and undershoots before and after the discontinuity (see Section 3.7.1). Example 4.2.1 Determine FfxðtÞg ¼ FfAP½t=tg and FfyðtÞg ¼ Ffxðt ðt=2ÞÞg using the F-series of

YðjoÞ ¼

Z1

1

xðtÞe

jot

Zt dt ¼ Aejot dt ¼

A ejot t0 ðjoÞ

0

1 ejot ejoðt=2Þ ejoðt=2Þ jot=2 ¼ At e ¼A jo 2jot=2 sinðot=2Þ jot=2 ¼ At : (4:2:13c) e ðot=2Þ The transform of the pulse xðtÞ is XðjoÞ ¼ At

sinðot=2Þ : ðot=2Þ

(4:2:14)

We should note that yðtÞ is a delayed version of xðtÞ and the delay explicitly appears in the phase spectra. See the difference between (4.2.13c) and & (4.2.14).

112

4 Fourier Transform Analysis

Since the F-transform is derived from the F-series, many of the properties for the F-series can be modified to derive the transform properties with some exceptions. Let a time limited function transform FT xðtÞ Xð joÞ be defined over the interval t0 5t5t0 þ T and zero everywhere else. This implies

$

xT ðtÞ ¼

1 X

xðt þ nTÞ ¼

n¼1

1 X

Xs ½ke jko0 t

k¼1

) Xs ½k ¼ ð1=TÞXðjoÞjo¼ko0 :

(4:2:15)

If xðtÞ is not time limited to a T second interval, then the function xðtÞ cannot be extracted from xT ðtÞ and (4.2.15) is not valid.

4.2.1 Amplitude and Phase Spectra

RðoÞ ¼

Z1 xðtÞ cosðotÞdt; 1

IðoÞ ¼

Z1 xðtÞ sinðotÞdt: 1

Since cosðotÞ is even and sinðotÞ is odd, the equalities in (4.2.18a) follow. As a consequence, for any real signal xðtÞ, we have XðjoÞ ¼ RðoÞ þ jIðoÞ ¼ RðoÞ ¼ jIðoÞ ¼ X ð joÞ: (4:2:18b) From (4.2.16b) and (4.2.17), the amplitude spectrum jXðjoÞj of a real signal is even and the phase spectrum yðoÞ is odd. That is, jXðjoÞj ¼ jXðjoÞj; yðoÞ ¼ yðoÞ:

Let the Fourier transform of a real signal xðtÞ be given by XðjoÞ. It is usually complex and can be written as either in terms of the rectangular or the polar form: XðjoÞ ¼ RðoÞ þ jIðoÞ ¼ AðoÞe jyðoÞ :

(4:2:16a)

The functions RðoÞ and IðoÞ are the real and the imaginary parts of the spectrum. In the polar form, the magnitude or the amplitude and the phase spectra are given by qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ AðoÞ ¼ jXðjoÞj ¼ R2 ðoÞ þ I 2 ðoÞ; yðoÞ ¼ tan1 ½IðoÞ=RðoÞ:

FT

xðtÞ ¼ xe ðtÞ þ x0 ðtÞ !RðoÞ þ jIðoÞ;

IðoÞ ¼ IðoÞ:

FT FT

x0 ðtÞ ¼ ½xðtÞ xðtÞ=2 !jIðoÞ:

¼

¼

XðjoÞ ¼

Z1 xðtÞe 1

j

jot

dt ¼

Z1

1

Z1 xðtÞejot dt 1 Z1

½xe ðtÞ þ x0 ðtÞ½cosðotÞ j sinðotÞdt

1 Z1

Z1 xe ðtÞ cosðotÞdt j x0 ðtÞ sinðotÞdt

1

xðtÞ sinðotÞdt ¼ RðoÞ þ jIðoÞ;

1

Z1 Z1 j xe ðtÞ sinðotÞdt þ x0 ðtÞ cosðotÞdt

xðtÞ cosðotÞdt 1

Z1

(4:2:21)

These can be seen from

(4:2:17)

These can be easily verified using Euler’s identity and

(4:2:20)

xe ðtÞ ¼ ½xðtÞ þ xðtÞ=2 !RðoÞ and

XðjoÞ ¼

RðoÞ ¼ RðoÞ and

(4:2:19)

Interesting transform relations in terms of the even and odd parts of a real function: If xðtÞ ¼ xe ðtÞ þ x0 ðtÞ, a real function, then the following is true:

(4:2:16b) When xðtÞ is real, the transform satisfies the properties that RðoÞ and IðoÞ are even and odd functions of o, respectively. That is,

(4:2:18a)

1

1

Z1 Z1 ¼ xe ðtÞ cosðotÞdt j x0 ðtÞ sinðotÞdt 1

¼ RðoÞ þ jIðoÞ

1

(4:2:22a)

4.2 Fourier Series to Fourier Integral

113

sinðoðt=2ÞÞ jot e 0 ¼ tjsincðot=2Þj; jXðjoÞj ¼ t oðt=2Þ

Z1 RðoÞ ¼ xe ðtÞ cosðotÞdt; 1

Z1 IðoÞ ¼ x0 ðtÞ sinðotÞdt:

(4:2:22b)

yðoÞ ¼

(4:2:24b) ðot0 Þ;

sincðot=2Þ40

ðot0 Þ p;

sincðot=2Þ50

:

(4:2:24c)

1

Note that integral of an odd function over a symmetric interval is zero. The F-transform of a real and even function is real and even and the F-transform of a real and odd function is pure imaginary. The transform of a real function xðtÞ can be expressed in terms of a real integral:

1 2p

1 ¼ 2p

1 Z1

jXðjoÞje jyðoÞ e jot do

¼ ðp þ pÞ ¼ 0:

1 Z1

jXðjoÞje jðotþyðoÞÞ do

1

Z1 1 ¼ jXðjoÞj cosðot þ yðoÞÞdo: p

(4:2:25a) The discontinuity in the phase spectrum at o ¼ 2p=t can be seen from 8 þ9 8 9 8 98 t 9 2p > 1;: ; >2p > > : > ; ¼ 2p ¼ p; y> y: : ; t t 2 t

Z1 1 xðtÞ ¼ XðjoÞe jot do 2p ¼

The time function xðtÞ; the magnitude spectrum jXðjoÞj, and the phase spectrum are sketched in Fig. 4.2.2 assuming t0 ¼ t=2. At o ¼ k2p=t, Xð j k2pÞ ¼ t; k ¼ 0 : t 0; k 6¼ 0 and k, an integer

(4:2:23)

We have added p in determining the phase angle in determining yð2pþ =tÞ taking into consideration that the sinc function is negative in the range 05ð2p=tÞ5o52ð2p=tÞ, i.e., in the first side lobe. In sketching the plots, appropriate multiples of (2p)

0

Note jXðjoÞj and sinðot þ yðoÞÞ are even and odd functions, respectively. Example 4.2.2 Rectangular (or a gating pulse) is given by xðtÞ ¼ P½ðt t0 Þ=t. a. Give the expression for the transform using (4.2.13a). b. Compute the amplitude and the phase spectra associated with the gating pulse function. c. Sketch the magnitude and phase spectra of this function assuming t0 ¼ t=2:

(a)

Solution: a. From (4.2.13a), we have sinðot=2Þ jot0 t XðjoÞ ¼ t e ¼ tsincðo Þejot0 2 ðot=2Þ

(b)

t 2

¼ tsincð2pf Þejot0 ; ht t i 0 FT jot0 ¼ t sincðpftÞej2pft0 : !t sincðot=2Þe t (4:2:24a) b. The magnitude and the phase spectra are, respectively, given by

(4:2:25b)

P

(c) Fig. 4.2.2 (a) xðtÞ, (b) jXðjoÞj, and (c) ﬀXðjoÞ

114

have been subtracted to make the phase spectrum compact and the phase angle is bounded between p and p. Noting sinðoðt=2ÞÞ / 1 ; (4:2:25c) jXðjoÞj ¼ t oðt=2Þ joj it can be seen that the envelope of the magnitude spectrum gets smaller for larger frequencies. The exact frequency representation of the square pulse should include all frequencies in the reconstruction of this pulse, which is impractical. Therefore, keep only the frequencies that are significant and the range or the width of those significant frequencies is referred to as the ‘‘bandwidth’’. Keeping the desired frequencies is achieved by using a filter. Filters will be discussed in later chapters. There are several interpretations of bandwidth. For the present, the following explanation of the bandwidth is adequate and will quantify these measures at a later & time.

4.2.2 Bandwidth-Simplistic Ideas 1. The width of the band of positive frequencies passed by a filter of an electrical system. 2. The width of the positive band of frequencies by the central lobe of the spectrum. 3. The band of frequencies that have most of the signal power. 4. The bandwidth includes the positive frequency range lying between two points at which the power is reduced to half that of the maximum. This width is referred to as the half-power bandwidth or the 3 dB bandwidth. Note that only positive frequencies are used in defining the bandwidth. With bandwidth in mind, let us look at jXðjoÞj in Example 4.2.2, where t is assumed to be the width of the pulse. Any signal that is nonzero for a finite period of time is referred to as a time-limited signal and the signal given in Example 4.2.2 is a time-limited signal. The main lobe width of the magnitude spectrum for positive frequencies is ð1=tÞ Hz. The spectrum is not frequency limited, as the spectrum occupies the entire frequency range, except that it is zero at o ¼ kð2p=tÞ; k 6¼ 0; k integer. The amplitude

4 Fourier Transform Analysis

spectrum jXðjoÞj gives the value jXðjo i Þj at the frequency oi ¼ 2pfi . Noting that most of the energy is in the main lobe, a standard assumption of bandwidth is generally assumed to be equal to k times half of the main lobe width ð2p=tÞ, i.e., (kð2p=tÞ) rad/s or ðk=tÞ Hz. An interesting formula that ties the time and frequency widths is (Time width) (Frequency width or bandwidth) ¼ Constant (4:2:26) Clearly as t decreases (increases), the main lobe width increases (decreases) and we say that bandwidth is inversely proportional to the time width. For most applications, k ¼ 1 is assumed. Bandwidth is generally given in terms of Hz rather than rad/s.

4.3 Fourier Transform Theorems, Part 1 We will consider first a set of theorems or properties associated with the energy function xðtÞ and its Ftransform XðjoÞ: Transforms are applicable for both the real and complex functions. In Chapter 3, Parseval’s theorem was given and it can be generalized to include energy signals and is referred to as generalized Parseval’s theorem, Plancheral’s theorem or Rayleigh’s energy theorem.

4.3.1 Rayleigh’s Energy Theorem The energy in a complex or a real signal is Z1

Ex ¼

¼

1 Z1

1 jxðtÞj dt ¼ 2p 2

Z1

jXðjoÞj2 do

1

jXðjfÞj2 df:

(4:3:1)

1

This is proved in general terms first. FT FT Given xðtÞ !XðjoÞ and yðtÞ !YðjoÞ, then

Exy ¼

Z1 1

1 xðtÞy ðtÞdt ¼ 2p

Z1

XðjoÞY ðjoÞdo:

1

(4:3:2)

4.3 Fourier Transform Theorems, Part 1

115

First, Z1

Exy ¼

¼

1 Z1

rectangular formula or the trapezoidal formula dis& cussed in Chapter 1.

1 y ðtÞ½ 2p

1

¼

¼

1 2p 1 2p

Z1

XðjoÞe jot dodt

1 Z1

Z1

1 ¼ 2p

Example 4.3.2 Compute the energy in the main lobe of the sinc function and compare with the total energy in the function using the following:

y ðtÞxðtÞdt

XðjoÞ½ 1 Z1

1 Z1

XðjoÞ½

EMain lobe

y ðtÞe

jot

XðjoÞY ðjoÞdo:

9 81 ð > > 2 > > > > sin ðpaÞ > : sinc2 ðot=2Þdo> da ¼ pp > > > > 2 ; : a

sinc2 ðpaÞda:

(4:3:6)

Solution: To obtain (4.3.6) from (4.3.5), change of variable, pa ¼ ot=2, is used and the limits are from o ¼ 2p=t to a ¼ 1. Using the rectangular method of integration, the energies in the main lobe and in the pulse E1 (see (4.3.4)) are EMain lobe 0:924=t; E1 ¼ ð1=tÞ:

(4:3:7)

The ratio of the energy in the main lobe, EMain lobe , of the spectrum to the total energy in the pulse is 92.4%. That is, the main lobe has over 90% of the total energy in the pulse function. Therefore, a bandwidth of (1/tÞ Hz is a reasonable estimate of & the pulse function.

1

(4:3:3) P

hti t

4.3.2 Superposition Theorem

FT

!t sincðot=2Þ:

Solution: Using the energy theorem, Z1

Z1

1

1

1 E1 ¼ 2p

(4:3:5)

1 1 Z1

Example 4.3.1 Compute the energy in the pulse E1 using Rayleigh’s energy theorem, the transform pair, and the identity Spiegel (1968) given below: 1 2p

1 ¼ t

yðtÞejot dt do

When xðtÞ ¼ yðtÞ, the proof of Rayleigh’s energy theorem in (4.3.1) follows. The transform of the function or the function itself can be used to find the energy.

E1 ¼

sinc2 ðot=2Þdo;

2p=t

dtdo

1

1 ð

Z2p=t

1 ¼ 2p

2 Z1 t 1 sinc ðot=2Þdo ¼ P2 dt t t 2

1

1

2 Zt=2 1 1 1 ðtÞ ¼ ¼ 2 dt: ¼ t t t

The Fourier transform of a linear combination of functions F½xi ðtÞ ¼ Xi ðjoÞ, i ¼ 1; 2; :::; n with constants ai ; i ¼ 1; 2:::; n is " F

n X

# ai xi ðtÞ ¼

i¼1 n X

(4:3:4)

i¼1

n X

ai Xi ðjoÞ;

i¼1 FT

ai xi ðtÞ !

n X

ai Xi ðjoÞ:

(4:3:8)

i¼1

t=2

Bandwidth of a rectangular pulse of width t is usually taken (1/tÞ Hz corresponding to the first zero crossing point of the spectrum. Energy contained in the main lobe of the sinc function can be computed by numerical methods, such as the

Since the integral of a sum is equal to the sum of the integrals, the proof follows. This theorem is useful in computing transforms of a function expressible as a sum of simple functions with known transforms. The F-transform of the function x ðtÞ is related to the transform of xðtÞ. This can be seen from

116

Z1

4 Fourier Transform Analysis

2

x ðtÞe

jot

dt ¼ 4

1

3

Z1 xðtÞe

jot

dt5 ¼ X ðjoÞ;

1 FT

) x ðtÞ !X ðjoÞ:

(4:3:9)

4.3.3 Time Delay Theorem The F-transform of a delayed function is given by F½xðt td Þ ¼ ejotd XðjoÞ:

(4:3:10)

This can be shown directly by using the change of variable a ¼ t td in the transform integral and F½xðt td Þ ¼

Z1

Solution: xðtÞ can be expressed as a sum of two rectangular pulses and is

t þ ðt=2Þ t ðt=2Þ P : xðtÞ ¼ P t t

(4:3:13a)

Using the superposition and delay theorems, we have

t þ ðt=2Þ t ðt=2Þ P XðjoÞ ¼ F½xðtÞ ¼ F P t t h i n h t io ¼ ejot=2 ejot=2 F P t sinðot=2Þ : (4:3:13b) ¼ 2jt sinðot=2Þ ðot=2Þ Note that xðtÞ is an odd function and therefore the & transform is pure imaginary.

xðt td Þejot dt

Notes: In the above example a time-limited function, i.e., xðtÞ ¼ 0 for jtj4t, was used and its trans2 1 3 Z form is not frequency limited, as its spectrum occuxðaÞejoa da5ejotd ¼ XðjoÞejotd pies the entire frequency range. A signal xðtÞ and its ¼4 transform XðjoÞ cannot be both time and frequency 1 & limited. We will come back to this later. ? ? ? ? ? ? ? ? )?F½xðt td Þ?¼?XðjoÞejotd?¼?XðjoÞ?¼?F½xðtÞ?: (4:3:11) 1

Superposition and delay theorems are useful in finding the Fourier transform pairs: FT

xðt tÞ þ xðt þ tÞ !2XðjoÞ cosðotÞ; (4:3:12) FT xðt tÞ xðt þ tÞ ! 2jXðjoÞ sinðotÞ: Example 4.3.3 Using the superposition and the delay theorem, compute the F-transform of the function shown in Fig. 4.3.1.

4.3.4 Scale Change Theorem The scale change theorem states that Z1 1 o xðatÞejot dt ¼ X j ; a 6¼ 0: F½xðatÞ ¼ a j aj 1

(4:3:14) This can be shown by considering the two possibilities, a50 and a40. For a50, by using the change of variable b ¼ at in (4.3.14) in the integral expression, we have

F½xðtÞ ¼

Z1 xðbÞe

jðo=aÞb

1 db a

1

Z1 1 ¼ xðbÞejðo=aÞb db a 1

Fig. 4.3.1 Example 4.3.3

1 ¼ j aj

Z1 1

o

xðbÞejð a Þb db ¼

1 o X j : a j aj

(4:3:15)

4.3 Fourier Transform Theorems, Part 1

117

When a is a negative number, a ¼ jaj. For a40, the proof similarly follows. The scale change theorem states that timescale contraction (expansion) corresponds to the frequency-scale expansion (contraction). Example 4.3.4 Use the scale change theorem to find the F-transforms of the following:

hti t x1 ðtÞ ¼ P and x2 ðtÞ ¼ P : (4:3:16a) ðt=2Þ 2t Solution: Consider the (4.2.24a)) with t0 ¼ 0: xðtÞ ¼ P½t=t

transform

$t sincðot=2Þ ¼ XðjoÞ: FT

Using the result in (4.3.14), we have

pair

(see

(4:3:16b)

x1 ðtÞ ¼ xð2tÞ ¼ P

t sinðoðt=4ÞÞ ! 2 oðt=4Þ

FT

t ¼ sincðoðt=4ÞÞ ¼ X1 ðjoÞ; 2 t h t i FT sinðotÞ x2 ðtÞ ¼ x ð2tÞ ¼P 2 2t ! ot ¼ ð2tÞ sincðotÞ ¼ X2 ðjoÞ:

(4:3:16c)

(4:3:16d)

The two functions and their amplitude spectra are sketched in Figs. 4.3.2a–d. Comparing the magnitude spectra, the main lobe width of jX1 ðjoÞj is twice that of the main lobe width of jXðjoÞj, whereas the main lobe width of jX2 ðjoÞj is half the main lobe width of jXðjoÞj. Consider Figs. 4.2.2 and 4.3.2. The main lobe width times its height in each of the cases are equal and tð2p=tÞ ¼ ðt=2Þð2pð2=tÞÞ ¼ ð2tÞðp=tÞ ¼ 2p. The

(a)

(b)

(c)

Fig. 4.3.2 (a) x1 ðtÞ, (b) jX1 ðjoÞj, (c) x2 ðtÞ, and (d) jX2 ðjoÞj

t t=2

(d)

118

4 Fourier Transform Analysis

pulse amplitudes are all assumed to be equal to 1 for simplicity. For any a io n h io ? ? n h ?F½xðtÞ? ¼ F P t a ¼ F P t : t t & Time reversal theorem: A special case of the scale change theorem is time reversal and F½xðtÞ ¼ XðjoÞ:

2a FT eajtj ! 2 ; a40: a þ o2

The time and frequency functions are not limited in & time and frequency, respectively.

(4:3:17a)

This follows from the scale change theorem by using a ¼ 1 in (4.3.14). We note that

4.3.5 Symmetry or Duality Theorem FT

FT

FT

xðtÞ !XðjoÞ; x1 ðtÞ ¼ xðtÞ !XðjoÞ ¼ X1 ðjoÞ;

jX1 ðjoÞj ¼ jXðjoÞj ¼ jXðjoÞj; ﬀX1 ðoÞ ¼ ﬀXðjoÞ ¼ ﬀXðjoÞ:

(4:3:17b)

Example 4.3.5 Find the F-transform of the following functions: a: x1 ðtÞ ¼ e x3 ðtÞ ¼ e

at

uðtÞ;

ajtj

b: x2 ðtÞ ¼ e uðtÞ;

¼

eðaþjoÞt t¼1 1 : ¼ ða þ joÞ t¼0 ða þ joÞ

(4:3:18c)

c. Noting that x3 ðtÞ ¼ eat uðtÞ þ eat uðtÞ, the Ftransform can be computed using the superposition theorem and the results in the last two parts. That is, X3 ðjoÞ ¼ F½eat uðtÞ þ F½eat uðtÞ 1 1 2a ¼ :(4:3:18d) þ ¼ ða þ joÞ ða joÞ a2 þ ðoÞ2 &

Summary: 1 FT eat uðtÞ ! ; a40; ða þ joÞ 1 FT ; a40; eat uðtÞ ! a jo

1 Z1

XðjoÞejot do ! 2p xðtÞ

XðjoÞejot do:

(4:3:22)

1

Interchanging t and jo in (4.3.22) results in 2p xðjoÞ ¼

Z1

XðtÞejot dt:

(4:3:23)

1

(4:3:18b)

b. Using the time reversal theorem and the last part results in X2 ðjoÞ ¼ X1 ðjoÞ ¼ ½1=ða joÞ:

Z1

2p xðtÞ ¼

¼

0

(4:3:21)

Starting with the expression for 2p xðtÞ and changing the variable from t to t, we have

(4:3:18a)

Solution: a. Using the F-transform integral results in Z1 Z1 X1 ðjoÞ ¼ eat uðtÞejot dt ¼ eðaþjoÞt dt 1

FT

xðtÞ !XðjoÞ ) XðtÞ !2pxðjoÞ:

at

; a40:

(4:3:20)

(4:3:19a) (4:3:19b)

This proves the result in (4.3.21). In terms of f (4.3.21) can be written as follows: FT

FT

xðtÞ !XðjfÞ; XðtÞ !xðjfÞ:

(4:3:24)

A consequence of the symmetry property is if an Ftransform table is available with N entries, then this property allows for doubling the size of the table. Example 4.3.6 Using the duality theorem, show that yðtÞ ¼

1 FT p ajoj e ¼ YðjoÞ: a2 þ t2 ! a

(4:3:25)

Solution: Using (4.3.20) and the duality property of the F-transforms, we have 1 ajtj FT 1 e ; ! 2 2a a þ o2 1 p FT 1 ð2pÞeajjoj ¼ eajoj : a4 0 ! 2 ! 2 a þt 2a a

4.4 Fourier Transform Theorems, Part 2

119

One can appreciate the simplicity of using the duality theorem compared to finding the transform directly in terms of difficult integrals given below: Z1 Z1 1 1 jot e do ¼ cosðotÞdo YðjoÞ ¼ a2 þ t 2 a2 þ t2 1

j

1

Z1

The value of the given function at t ¼ 0 and its transform value at o ¼ 0 are given by Z1 Z1 1 Xð0Þ ¼ xðtÞdt; xð0Þ ¼ XðjoÞdo: 2p 1

1 sinðotÞdo: 2 a þ t2

&

1

Example 4.3.7 Determine the F-transform of xðtÞ using the transform of the rectangular pulse given below and the duality theorem: h t i FT sinðot=2Þ sinðatÞ t ; P : xðtÞ ¼ pt t ! ðot=2Þ Solution: Using the duality theorem and noting that P-function is even, it follows

hoi sinðtt=2Þ FT jo t 2pP ¼ 2pP ; (4:3:26) ! ðtt=2Þ t t sinðatÞ FT h o i (4:3:27) !P 2a : pt Note ðt=2Þ ¼ a in (4.3.26). For later use, let a ¼ 2pB. Using this in (4.3.27) results in y ðt Þ ¼

4.3.6 Fourier Central Ordinate Theorems

sinð2pBtÞ FT 1 o ¼ YðjoÞ: (4:3:28) P ! 2B ð2pBtÞ 2pð2BÞ

Time domain sinc pulses are not time limited but are band limited. The sinc pulse and its transform in (4.3.28) are sketched in Fig. 4.3.3a,b, respectively. &

1

(4:3:29) Equation (4.3.29) points out that if we know the transform of a function, we can compute the integral of this function for all time by evaluating the spectrum at o ¼ 0. In a similar manner the integral of the spectrum for all frequencies is given by ð2pÞxð0Þ. Example 4.3.8 Use the transform pair in (4.3.28) to illustrate the ordinate theorems in (4.3.29) using the identity Spiegel (1968) 8 > Z1 < p; p40 sinðpaÞ (4:3:30) da ¼ 0; p ¼ 0 : > a : 1 p; p50 Solution: The integrals of the sinc function and the area of the pulse are A1 ¼

Z1 1

sinð2pWtÞ 1 dt ¼ ð2pWtÞ 2pW

Z1

sinð2pWtÞ dt t

1

p 1 ¼ : (4:3:31a) ¼ 2pW 2W

1 o 1 A2 ¼ P ) A2 ¼ A1 : jo¼0 ¼ 2W 2pð2WÞ 2W (4:3:31b) In a similar manner, sinð2pWtÞ jt¼0 ¼ 1; ð2pWtÞ Z1 1 1 B2 ¼ YðjoÞdo ¼ 2pð2 WÞ 2p 2pð2 WÞ

B1 ¼

1

(a)

¼ 1 ) B1 ¼ B2 :

(4:3:32) &

4.4 Fourier Transform Theorems, Part 2 (b) Fig. 4.3.3 (a) yðtÞ ¼

sinð2pWtÞ ð2pWtÞ

1 and (b) YðjoÞ=2W P

h

o 2pð2WÞ

i

Impulse functions are used in finding the transforms of periodic functions below.

120

4 Fourier Transform Analysis

Example 4.4.1 Find the Fourier transform of the impulse function in time domain and the inverse transform of the impulse function in the frequency domain.

Example 4.4.2 Show the following: 9 8 j ðo o c Þ> joc t FT 1 > ;: : xðatÞe ! j aj X a

(4:4:7a)

Solution: Using the scale change theorem results in Solution: Clearly 1 Z jo FT 1 jot jot jot0 xðatÞ ! X : (4:4:7b) F½dðt t0 Þ ¼ dðt t0 Þe dt ¼ e : jt¼t0 ¼ e a jaj 1

(4:4:1) That is, an impulse function contains all frequencies with the same amplitude. That is jF½dðt t0 Þj ¼ 1. The inverse transform F

1

1 ½dðo o0 Þ ¼ 2p

Z1

dðo o0 Þe jot do ¼

1 jo0 t e ; 2p

1

) F 1 ½dðoÞ ¼ 1=2p;

F½1 ¼ 2pdðoÞ:

(4:4:2) (4:4:3)

A constant contains only the single frequency at o ¼ 0 (or f ¼ 0Þ. We refer to a constant as a dc signal. Symbolically we can express FT

FT

dðt t0 Þ !ejot0 ; e jo0 t !2pdðo o0 Þ:

(4:4:4)

The result on the right in the above equation follows & by using the duality theorem.

4.4.1 Frequency Translation Theorem Multiplying a time domain function xðtÞ by ejoc t shifts all frequencies in the signal xðtÞ by oc . In general, the following transform pair is true: FT

xðtÞejoc t !Xðjðo oc ÞÞ:

(4:4:5)

Note F

1

Z1 1 ½Xðjðo oc ÞÞ ¼ Xðjðo oc ÞÞe jot do 2p 1 2 3 Z1 1 jat ¼4 XðjaÞe da5 e joc t ¼ xðtÞe joc t : 2p 1

(4:4:6) This provides a way to modify a time function to shift its frequencies. The scale change and the frequency translation theorems can be combined.

Using the frequency translation theorem, i.e., multiplying the function by e joc t causes a shift in the frequency. That is, replace o by o oc and the & result in (4.4.7a) follows.

4.4.2 Modulation Theorem The frequency translation theorem directly leads to the modulation theorem. Given F½xðtÞ ¼ XðjðoÞÞ and yðtÞ ¼ xðtÞ cosðoc t þ yÞ, the modulation theorem results in YðjoÞ ¼ F½xðtÞ cosðoc t þ yÞ

1 1 ¼ F ðxðtÞe jy Þe joc t þ ðxðtÞejy Þejoc t 2 2 1 1 ¼ ejy Xðjðo þ oc ÞÞ þ e jy Xðjðo oc ÞÞ: 2 2 (4:4:8) In simple words, multiplying a signal by a sinusoid translates the spectrum of a signal around o ¼ 0 to the locations around oc and oc . If the spectrum of the signal xðtÞ is frequency (or band) limited to o0 , i.e., jXðjoÞj ¼ 0; joj4o0 , then jYðjoÞj ¼ 0 for joj4joc þ o0 j and joj5joc o0 j: (4:4:9) Figure 4.4.1 gives sketches of the signals and their spectra. The signal xðtÞ is assumed to cross the time axis. There is no real significance in the shape of the spectrum. Since xðtÞ is real, it has even magnitude and odd phase spectrum. The signal is band limited to f0 ¼ o0 =2p Hz. The modulated signal yðtÞ shown in Fig. 4.4.1b assumes y ¼ 0 in (4.4.8). The positive and negative envelopes of the modulated signal are shown by the dotted lines. Note the envelopes cross the axis wherever the

4.4 Fourier Transform Theorems, Part 2

121

(a)

(b) Fig. 4.4.1 (a) xðtÞ and jXðjoÞj, (b) yðtÞ and jYðjoÞj

function xðtÞ ¼ 0. The magnitude and phase spectra of the modulated signal are shown in Fig. 4.4.1b. The bandwidth of the modulated signal is twice the bandwidth of the message signal. Note the factor & (1/2) in both terms in (4.4.8).

The signal yðtÞ is being seen through a rectangular (window) function xðtÞ. Outside this window, no signal is available. The study of windowed signals is an important topic for signal processors and it is humorously called as window carpentry. We will & come back to this topic later.

Example 4.4.3 Find the F½xðtÞ cosðoc tÞ and F½xðtÞ sinðoc tÞ in terms of F½xðtÞ: Solution: Clearly when y ¼ 0 and y ¼ p=2 in (4.4.8), the F-transform pairs are 1 FT 1 xðtÞ cosðoc tÞ ! Xðjðo oc ÞÞ þ Xðjðo þ oc ÞÞ; 2 2 (4:4:10a) 1 FT 1 xðtÞ sinðoc tÞ ! Xðjðo oc ÞÞ Xðjðo þ oc ÞÞ: 2j 2j (4:4:10b) Modulation theorem provides a powerful tool for finding the Fourier transforms of functions that are seen (or windowed) through a function xðtÞ. For example, hti hti xðtÞ ¼ P ! yðtÞ ¼ P cosðoc tÞ t t cosðoc tÞ; jtj52t : ¼ 0; otherwise

4.4.3 Fourier Transforms of Periodic and Some Special Functions Modulation theorem gives a back door way to find the Fourier transforms of periodic functions, such as sine and cosine functions. A sufficient condition for the existence of F½xðtÞ is Z1

jxðtÞjdt51 (absolute integrability condition):

1

Clearly the sine, cosine, unit step, and many other functions violate this condition. Use of the generalized functions allows for the derivation of the Ftransforms of these functions. Example 4.4.4 Use the transform F½1 ¼ 2pdðoÞ and the modulation theorem to find the Fourier transforms of xðtÞ ¼ cosðo0 tÞ and yðtÞ ¼ sinðo0 tÞ.

122

4 Fourier Transform Analysis

Solution: Using (4.4.8), we have the transforms. These are given below in terms of o and f: In the latter case, we have used dðoÞ ¼ dðfÞ=2p:

Notes: A narrowband band-pass signal with a slowly changing envelope RðtÞ and phase fðtÞ has the forms xðtÞ ¼ RðtÞ cosðoc t þ fðtÞÞ; RðtÞ 0;

FT

xðtÞ¼cosðo0 tÞ !XðjoÞ¼pdðoþo0 Þþpdðoo0 Þ; FT

yðtÞ¼sinðo0 tÞ !YðjoÞ¼jpdðoþo0 Þjpdðoo0 Þ; (4:4:11a)

(4:4:12a)

xðtÞ ¼ xi ðtÞ cosðoc tÞ xq ðtÞ sinðoc tÞ; D

xi ðtÞ ¼ RðtÞ cosðfðtÞÞ; xq ðtÞ ¼ RðtÞ sinðfðtÞÞ: (4:4:12b)

FT

cosð2pf0 tÞ ! 12 dðf þ f0 Þ þ 12 dðf f0 Þ; FT

sinð2pf0 tÞ ! 2j dðf þ f0 Þ 2j dðf f0 Þ:

(4:4:11b)

The spectra of the cosine and the sine functions are shown in Figs. 4.4.2. The spectra of these are located at o ¼ o0 ¼ 2pf0 with the same magnitude, but the phases are different. In reality, we do not have any negative frequencies. Euler’s formula illustrates that sinðoc tÞ and cosðoc tÞ are not the same functions, even though they have the same frequencies. Noting that the real part of the transform of a real signal is even and the phase spectra is odd, the negative frequency component does not give any additional information regarding what frequency is present. The average power represented by the negative frequency component simply adds to the average power of the positive frequency component resulting in the total average power at that frequency. In the case of an arbitrary signal resolved into in-phase and quadrature-phase components, the negative frequency terms do contribute additional information. A cosine wave reaches its positive peak 908 before a sine wave does. By convention the cosine wave is called the in-phase or i (or I) component and the sine wave is called the quadra& ture-phase or the q (or Q) component.

Equation (4.4.12a) gives the envelope-and-phase description and (4.4.12b) gives the in-phase and quadrature-carrier description. The components xi ðtÞ and xq ðtÞ are the in-phase and quadraturephase components. Now consider a windowed cosine function and & see the effects of that window. Example 4.4.5 Find the Fourier transform of the cosinusoidal pulse function. Plot the functions XðjoÞ and YðjoÞ and identify the important parameters: yðtÞ ¼ xðtÞ cosðo0 tÞ; xðtÞ ¼ P

Fig. 4.4.2 Transform of the cosine and sine functions

FT

!t sin cðot=2Þ:

Solution: The transform of yðtÞ is YðjoÞ ¼

t sin½ðo o0 Þðt=2Þ t sin½ðo þ o0 Þðt=2Þ þ : 2 ½ðo o0 Þðt=2Þ 2 ½ðo þ o0 Þðt=2Þ (4:4:13b)

The functions xðtÞ; XðjoÞ; yðtÞ; and YðjoÞ are sketched in Fig. 4.4.3a–d, respectively. Noting that XðjoÞ and YðjoÞ are real functions, the main lobe F[sin(ω0t)] π

0

t

(4:4:13a)

F[cos(ω0t)]

– ω0

hti

ω0

jπ

ω

ω

0

– jπ

4.4 Fourier Transform Theorems, Part 2

123

(a)

(b)

Y( jω)

(c)

(d)

Fig. 4.4.3 (a) and (b) Pulse function and its transform; (c) and (d) windowed cosine function and its transform

width of XðjoÞ corresponding to the positive frequencies is ð2p=tÞ. The function YðjoÞ has two main lobes centered at o ¼ o0 ¼ 2pf0 . Again considering only positive frequencies, the main lobe width of YðjoÞ is twice that of XðjoÞ equal to ð4p=tÞ. That is, the process of modulation doubles the bandwidth. As in XðjoÞ, we have side lobes in YðjoÞ that decay as we go away from the center frequency. The peak of the main lobe in XðjoÞ is t, whereas the peaks of the main lobes of YðjoÞ are equal to (t=2). Clearly, if we are interested in finding the frequency o0 ¼ 2pf0 , the steps could include the following: 1. Find the transform. 2. Find the peak value of the spectrum and its location. In a practical problem, we may have several frequencies. Finding the locations of these frequencies and their amplitudes is of interest. This problem is usually referred to as spectral estimation. The spectrum of a cosine function consists of two impulses located at o ¼ o0 . The spectrum of the windowed cosine function contains two sinc functions. We generally assume that o0 ð2p=tÞ and therefore the overlap of the two sinc functions at dc is

assumed to be negligible. Rectangular window modified the impulse spectra of the signal to a spectra consisting of sinc functions. Windowing a func& tion results in spectral leakage.

Fourier transforms of arbitrary periodic functions: In Chapter 3, we derived that if xT ðtÞ is a periodic function with period T and xT ðtÞ can be expressed into its F-series, 1 X xT ðtÞ ¼ Xs ½ke jko0 t ; k¼1

1 Xs ½k ¼ T

Z

xðtÞejko0 t dt;

o0 ¼

2p : T

(4:4:14)

T

where Xs ½k0 s in (4.4.14) are generally complex. The transform can be derived by noting that F½ejko0 t ¼ 2pdðo no0 Þ; F½xT ðtÞ ¼ ¼

1 X k¼1 1 X k¼1

(4:4:15)

Xs ½kF½e jko0 t Xs ½kð2pÞdðo ko0 Þ:

(4:4:16)

124

4 Fourier Transform Analysis

Example 4.4.6 Find the transform F½dT ðtÞ½¼ P F½ 1 k¼1 dðt kTÞ. Solution: The F-series of the function dT ðtÞ is given by (see (3.4.15b)) dT ðtÞ ¼

1 1 X e jko0 t ; o0 ¼ 2p=T: T k¼1

Using the linearity and frequency shift properties of the Fourier transforms, we have dT ðtÞ ¼

1 X

FT 2p dðt nTÞ ! T n¼1 1 X 2p dðo ko0 Þ ¼ do ðoÞ: T 0 k¼1

(4:4:17)

Example 4.4.7 Show the Gaussian pulse transform pair as follows:

e Solution: XðjoÞ ¼

¼

Z1 1 Z1

xðtÞejot dt ¼

Z1

(4:4:18)

XðjoÞ ¼ e

2 o4a

4p1ﬃﬃﬃ a

3

Z1 e

r

2

2

o dr5 ¼ e 4a

pﬃﬃﬃ jo aðt þ Þ; 2a

rﬃﬃﬃ p : a

(4:4:19)

Integral tables are used in (4.4.19). The transform pair in (4.4.18) now follows, that is, the Fourier transform of a Gaussian function is also a Gaussian function. Both time and frequency functions are not & limited in time and in frequency, respectively. The following pairs are valid and can be verified using Fourier transform theorems.

rﬃﬃﬃ 2 FT p o ap ; cos cos at2 ! 4a a

rﬃﬃﬃ 2 2 FT o ap p sin at ; ð4:4:20aÞ ! a sin 4a FT pﬃﬃﬃﬃﬃﬃ (4:4:20b) jtj1=2 ! 2pjoj1=2 :

2

eat ejot dt

eaðyÞ dt; y ¼ t2 þ j

Fig. 4.4.4 Periodic Impulse Sequence and its transform

By the change of variable, we have r ¼ pﬃﬃﬃ dt ¼ dr= a; t ) 1; r ! 1, and

For a catalog of Fourier transform pairs, see Abromowitz and Stegun (1964) and Poularikis (1996).

4.4.4 Time Differentiation Theorem

1

1

1

1

One question should come to our mind, that is, are there other functions and their transforms have the same general form? The answer is yes.

rﬃﬃﬃ p ðoÞ2 ! a e 4a ; a40:

ot ot o2 o2 ¼ t2 þ j þ 2 2 4a a a 4a 2 jo ðoÞ ¼ ðt þ Þ2 þ 2 ) XðjoÞ 4a 2a o2 Z1 o aðt þ j Þ2 2a dt: e ¼ e 4a

y ¼ t2 þ j

2

We have an interesting result: the Fourier transform of a periodic impulse sequence dT ðtÞ with period T is also a periodic impulse sequence ð2p=TÞdo0 ðoÞ with & period o0 . They are sketched in Fig. 4.4.4.

at2 FT

Now add and subtract the term ðo2 =aÞ to the term y in the exponent inside the integral

ot : a

If F½xðtÞ ¼ XðjoÞ and xðtÞ is differentiable for all time and vanishes as t ! 1, then

4.4 Fourier Transform Theorems, Part 2

F

dxðtÞ ¼ F½x0 ðtÞ ¼ ðjoÞ XðjoÞ: dt

125

(4:4:21)

Using integration by parts, we have

using the derivative method. Sketch the transform of the triangular function and compare the transform of the rectangular or P function with the transform of the L function.

Solution: xðtÞ; x0 ðtÞ; and x00 ðtÞ are sketched in Z1 d Fig. 4.4.5 a–c. Clearly, F xðtÞ ¼ x0 ðtÞejot dt ¼xðtÞejot t¼1 t¼1 þ jo dt 1 1 2 1 x00 ðtÞ ¼ dðt þ tÞ dðtÞ þ dðt tÞ: (4:4:24) Z1 t t t xðtÞejot dt ¼ jo XðjoÞ: 1

Differentiation of a function in time corresponds to multiplication of its transform by ðjoÞ, provided that the function xðtÞ ! 0 as t ! 1. If xðtÞ has a finite number of discontinuities, then x0 ðtÞ contains impulses. Then, (4.4.21) can be generalized and n

d xðtÞ F ¼ ðjoÞn XðjoÞ; n ¼ 1; 2; :::: (4:4:22) dtn The above does not provide a proof of the existence of the Fourier transform of the nth derivative of the function. It merely shows that if the transform exists, then it can be computed by the above formula. This theorem is useful if the transforms of derivatives of functions can be found easier than finding the transforms of functions. For example, dxðtÞ FT FT 1 joXðjoÞ ) xðtÞ ! F½x0 ðtÞ: ! dt jo Use of this approach in finding transforms is referred to as the derivative method. Example 4.4.8 Find the Fourier transform of the triangular function ( hti 1 tt ; jtj t xðtÞ ¼ L ¼ (4:4:23) t 0; Otherwise

Using the derivative theorem and solving for XðjoÞ, we have 1 2 1 FT x00 ðtÞ !ðjoÞ2 XðjoÞ ¼ e jot þ ejot ; t t t jot=2

2 V e ejot=2 2 ð4Þ ðjoÞ XðjoÞ ¼ 2j t 4 ¼ sin2 ðot=2Þ; t h t i FT sin2 ðot=2Þ xðtÞ ¼ V l Vt ! t ðot=2Þ2 ¼ Vt sinc2 ðot=2Þ ¼ XðjoÞ:

Equation (4.4.25) gives the spectrum of the triangular function of width ð2tÞ s. The rectangular function of width t s and its transform were given earlier by h t i FT sinðot=2Þ Vt ¼ Vt sincðot=2Þ: (4:4:26) VP t ! ðot=2Þ The time width of the rectangular window function in (4.4.26) is t s, whereas the time width of the triangular window function in (4.4.25) is 2t s. Note the square of the sinc2 function in the spectrum of the triangle function and the sinc function in the spectrum of the rectangular window. Since jsincðot=2Þj2 jsincðot=2Þj;

Fig. 4.4.5 (a) xðtÞ, (b) x0 ðtÞ, and (c) x0 ðtÞx0 ðtÞx00 ðtÞ

(4:4:25) &

126

4 Fourier Transform Analysis

the spectral amplitudes of the triangular function have lower side lobe levels compared to the spectral amplitudes of the rectangular pulses. Since the square of a fraction is smaller than the fraction itself, the side lobes in the transform of the triangular function are much smaller than the side lobes in the transform of the rectangular function. There is less leakage in the side lobes for the triangular (window) pulse function compared to the rectangular (window) function. See Fig. B.4.1 for a sketch of the sinc function. High-frequency decay rate: In the Fourier series discussion, the decay rate of the F-series coefficients Xs ½k was determined using the derivatives of the periodic function (see Section 3.6.5). Similarly, the F-transforms of pulse functions decay rate can be determined without actually finding the transform of the function. Given a pulse function xðtÞ, find the successive derivatives, xðnÞ ðtÞ, of the function until the first set of impulses appear in the nth derivative, then the decay rate is proportional to ð1=on Þ. In Example 4.4.7 the triangular pulse was considered and, in this case, the second derivative exhibits impulses indicating that the high-frequency decay rate of the transform is ð1=o2 Þ (see (4.4.25)). Similarly the first derivative of the rectangular pulse function and the exponential decaying function eat uðtÞ; a40 exhibit impulses indicating that the high-frequency decay rate of these transforms is ð1=jojÞ:

The similarities between the time and frequency differentiation theorems illustrate the duality properties with the F-transform pairs. Example 4.4.9 Show the following relationship using the times-t property: FT

teat uðtÞ !

1 ða þ joÞ2

; a40:

(4:4:28)

Solution: Noting the times-t property given above with xðtÞ ¼ eat uðtÞ, we have F½teat uðtÞ ¼ j

dXðjoÞ dð1=ða þ joÞÞ 1 ¼j ¼ : do do ða þ joÞ2

This can be generalized to obtain the following and the proof is left as an exercise: tn1 at 1 FT e uð t Þ ! ; a40: ðn 1Þ! ða þ joÞn

(4:4:29) &

Example 4.4.10 Noting that 1 FT eðajbÞt uðtÞ ! ; a40; a þ j ð o bÞ

(4:4:30a)

show the following is true: b

FT

xðtÞ ¼ eat sinðbtÞuðtÞ !

ða þ joÞ2 þb2

¼ XðjoÞ; (4:4:30b)

4.4.5 Times-t Property: Frequency Differentiation Theorem If XðjoÞ ¼ F½xðtÞ and if the derivative of the transform exists, then F½ðjtÞxðtÞ ¼

dXðjoÞ : do

(4:4:27)

This can be shown by dXðjoÞ d ¼ do do ¼

Z1 xðtÞe

1 Z1

jot

dt ¼

Z1 xðtÞ

dðejot Þ dt do

1

½ðjtÞxðtÞejot dt ¼ F½jtxðtÞ:

1

ða þ joÞ

FT

yðtÞ ¼ eat cosðbtÞuðtÞ !

ða þ joÞ2 þb2

¼ YðjoÞ: (4:4:30c)

Solution: These can be shown by first expressing the sine and cosine functions by Euler’s formulas, taking the transforms and then combining the & complex–conjugate terms. Example 4.4.11 Using lim eat uðtÞ ¼ uðtÞ; a40, a!0 find F½uðtÞ: 1 : a!0 a þ jo

F½uðtÞ ¼ lim

Solution: Noting that the limiting process is on the complex function, we need to take the limits on the

4.4 Fourier Transform Theorems, Part 2

127

real and the imaginary parts of the complex function separately. That is,

1 a o þ j lim : ¼ lim 2 lim a!0 a þ jo a!0 a þ o2 a!0 a2 þ o2

unit step function is real. These are illustrated in Fig. 4.4.6. Interestingly the spectrum of the delayed unit step uðt 1Þ is

(4:4:31)

1 F½uðt 1Þ ¼ ½pdðoÞ þ ejo ; jo

1 jF½uðt 1Þj ¼ pdðoÞ þ ; ﬀF½uðt 1Þ o o p=2; o40 ¼ : (4:4:34b) o þ p=2; o50

The second term in the above, i.e., the Lorentzian function, approaches an impulse function. That is,

a ¼ pdðoÞ: (4:4:32) lim a!0 a2 þ o2 Using this result in (4.3.30),

1 1 ¼ pdðoÞ þ ; lim a!0 a þ jo jo

(4:4:33)

1 ; (4:4:34a) jo p=2; o40 : jUðjoÞj ¼ pdðoÞ þ j1=oj; ﬀUðjoÞ ¼ p=2; o50 ) UðjoÞ ¼ F½uðtÞ ¼ pdðoÞ þ

Note that the amplitude is an even function and the phase angle function is an odd function, as the

Since delay of a function depends on the phase angle, it follows that jF½uðt 1Þj ¼ jF½uðtÞj. The phase spectrum of the delayed unit step function is sketched in Fig. 4.4.7. The Fourier transform of the unit function has two parts. The first part corresponds to the transform of the average value of the unit step function and the other part is the transform of the signum function. That is, F½uðtÞ ¼ F½ð1=2Þ þ ð1=2Þ sgnðtÞ ¼ F½1=2 þ ð1=2ÞF½sgnðtÞ ¼ pdðoÞ þ ð1=joÞ:

Fig. 4.4.6 (a) Magnitude and (b) phase spectra of the unit step function

Fig. 4.4.7 Phase spectrum of u(t1)

&

128

4 Fourier Transform Analysis

Notes: If we had ignored that we had to take the limit on the real and the imaginary parts of the complex function in (4.4.31) separately and take the limit on the complex function as a whole, the result would be wrong. That is,

FT

ð1=tÞ ! jp sgnðoÞ ¼ jp j2puðoÞ: This can be generalized and 1 FT ðjoÞn1 jp sgnðoÞ: ! ðn 1Þ! tn

1 1 lim ¼ 6¼ F½uðtÞ: a!0 a þ jo jo This indicates the transform is imaginary and the time function must be odd. This cannot be true since the unit step function is not an odd function and & F½uðtÞ 6¼ ð1=joÞ. The sgn (or signum) function is used in communications and control theory and can be expressed in terms of the unit step function. The sgn function and its transform are as follows: 8 > < 1; t40 (4:4:35) sgnðtÞ ¼ 2uðtÞ 1 ¼ 0; t ¼ 0: > : 1; t50 F½2uðtÞ F½1 ¼ 2pdðoÞ þ ð2=joÞ 2pdðoÞ ¼ ð2=joÞ:

(4:4:36)

The times-t property and the transform of the unit step function can be used to determine the Fourier transform of the ramp function and is given as

FT tuðtÞ ! jpd0 ðoÞ 1=o2 :

(4:4:37)

(4:4:39a)

(4:4:39b) &

4.4.6 Initial Value Theorem The initial value theorem is applicable for the rightsided signals, i.e., the functions of the form yðtÞ ¼ xðtÞuðtÞ and is stated below without proof: yð0þ Þ ¼ lim joYðjoÞ: o!1

(4:4:40)

Example 4.4.13 The unit step function is not defined at t ¼ 0, whereas uð0þ Þ ¼ 1; which is well defined. Verify the initial value theorem for the unit step function by noting dðoÞ ¼ 0; o 6¼ 0, and odðoÞ ¼ 0: Solution: uð0þ Þ ¼ lim fjoF½uðtÞg o!1

1 ¼ 1: ¼ lim jo pdðoÞ þ o!1 jo

&

Noting that jtj ¼ 2t uðtÞ t, we have the following transform pair:

4.4.7 Integration Theorem

FT jtj ! 2=o2 :

(4:4:38)

Example 4.4.12 Find the Fourier transform of the function xðtÞ ¼ ð1=tÞ using the duality theorem and the Fourier transform of the signum function. Solution: Using the duality theorem, we have

It states that yð t Þ ¼

Zt

FT XðjoÞ þ pXð0ÞdðoÞ ¼ YðjoÞ: xðaÞda ! jo

1

(4:4:41) FT

Duality theorem

FT

! XðtÞ !2pxðjoÞ; xðtÞ !XðjoÞ 1 FT FT 2 sgnðtÞ ! ; ð1Þp sgnðoÞ ¼ p sgnðoÞ: jo jt ! We can write sgn ðjoÞ ¼ sgnð oÞ ¼ sgnðoÞ. It follows that

This is true only if Xð0Þ; i.e., the area under xðtÞ, is finite. If the area under xðtÞ is zero, then the second term on the right in (4.4.41) disappears. Note that if Xð0Þ ¼ 0, integration and differentiation operations are inverse operations. Integration operation is a smoothing operation. Integral of a function has

4.5 Convolution and Correlation

129

lower frequency content than the function that is integrated. On the other hand, since x 0 ðtÞ ¼ joXðjoÞ, differentiation accentuates the higher frequencies. Integration theorem is not applicable if Xð0Þ is infinity. This theorem will be proved in Section 4.5. Example 4.4.14 Find the Fourier transform of uðtÞ using the integration theorem and Zt FT dðaÞda: dðtÞ !1; uðtÞ ¼ 1

Solution: By the integration theorem 2 t 3 Z 1 dðaÞda5 ¼ þ pdðoÞ: F½uðtÞ ¼ F4 jo 1

&

Example 4.4.15 Use the Fourier transform of the function xðtÞ ¼ cosðo0 tÞ; o0 6¼ 0, and the integration theorem to find the Fourier transform of the function sinðo0 tÞ. Solution: First for o0 6¼ 0, from (4.4.11a), we have FT

xðtÞ ¼ cosðo0 tÞ !pdðo þ o0 Þ þ pdðo o0 Þ ¼ XðjoÞ; Xð0Þ ¼ 0 and yðtÞ ¼

Zt

xðaÞda ¼

1

Zt cosðo0 aÞda (4:4:42a)

See the comment below in regard to the evaluation of the limit at 1 in the above integral. The integration property gives us Zt

(4:4:43a)

The limit does not exist as an ordinary limit and is a generalized limit in the sense of distributions. Using Euler’s formula and the limit in (4.4.43a), computation of the integral in (4.4.42a) follows. Switched functions are very useful in system theory. In computing the derivatives of such functions, one needs to be careful. For example, d½cosðtÞuðtÞ d½cosðtÞ d½uðtÞ ¼ uðtÞ þ cosðtÞ dt dt dt ¼ sinðtÞuðtÞ þ dðtÞ; (4:4:43b) d½sinðtÞuðtÞ ¼ cosðtÞuðtÞ þ sinðtÞdðtÞ ¼ cosðtÞuðtÞ: dt (4:4:43c) To find the transforms of such functions we can make use of modulation theorem. Derivative theorem can be used to find transforms of many functions such as xðtÞ ¼ ea t uðtÞ; a40. We should keep in mind that if the pulse is not time limited, we need to add a frequency domain delta function, whose weight is equal to 2p times the average of the pulse over the entire time axis to the transform result of the successive differentiation. See the discussion on & finding the transform of a unit step function.

4.5 Convolution and Correlation Chapter 2 considered convolution and correlation. Here we will consider the transforms of the signals that are convolved and correlated.

FT

cosðo0 aÞda !

1

o0 ½pdðo þ o0 Þ þ pdðo o0 Þ þ o0 pXð0ÞdðoÞ: jo With Xð0Þ ¼ 0 and dðo o0 Þ=o ¼ dðo o0 Þ=o0 , we have result as in (4.4.11a). Zt

lim ejo t ¼ 0:

t!1

1

¼ ð1=o0 Þ sinðo0 tÞ:

o 0 yð t Þ ¼ o 0

Notes: Papoulis (1962) discusses the concepts of generalized limits. For example

4.5.1 Convolution in Time Convolution of two time functions x1 ðtÞ and x2 ðtÞ is defined by Z1 yðtÞ ¼ x1 ðtÞ x2 ðtÞ ¼ x1 ðaÞx2 ðt aÞda 1

FT

cosðo0 aÞda ¼ sinðo0 tÞ !jpdðo þ o0 Þ ¼

1

jpdðo o0 Þ:

(4:4:42b) &

Z1 1

x2 ðbÞx1 ðt bÞdb ¼ x2 ðtÞ x1 ðtÞ: (4:5:1)

130

4 Fourier Transform Analysis FT

Assuming that xi ðtÞ !Xi ðjoÞ; i ¼ 1; 2, convolution theorem is given by FT

x1 ðtÞ x2 ðtÞ !X1 ðjoÞX2 ðjoÞ:

the

(4:5:2)

This can be proven by using the transform pair F½x2 ðt aÞ ¼ X2 ðjoÞejoa in (4.5.1) and the resulting integral is the inverse transform of ½X1 ðjoÞX2 ðjoÞ. That is, yðtÞ ¼ x1 ðtÞ x2 ðtÞ 2 3 Z1 Z1 1 x1 ðaÞ4 X2 ðjoÞe joðtaÞ do5da ¼ 2p 1 1 2 1 3 Z1 Z 1 X2 ðjoÞ4 x1 ðaÞejoa da5e jot do ¼ 2p ¼

1 2p

1 Z1

inverse transform of YðjoÞ. Second, in most applications, the function may not be given in an analytical or equation form and may be given in the form of a plot or a set of data and we have to resort to digital means to find the values for yðtÞ. We will consider the discrete Fourier transforms in Chapters 8 and 9. Example 4.5.1 Determine the function yðtÞ ¼ P½t :5 P½t :5 by using the transforms. Solution: Using the transforms of the pulse functions, we have

1 sinðo=2Þ jo=2 F P t ¼ e ; 2 ðo=2Þ " # sin2 ðo=2Þ jo (4:5:4) YðjoÞ ¼ e : ðo=2Þ2

1

½X2 ðjoÞX1 ðjoÞe jot do:j

(4:5:3)

1

Convolution theorem follows from the above equation. It gives a method for computing the convolution of two aperiodic functions via Fourier transforms. This method is the transform method of computing the convolution. The direct method is by the use of the convolution integral and all the operations are in the time domain. The transform method involves the following steps: a. Find F½x1 ðtÞ ¼ X1 ðjoÞ and F½x2 ðtÞ ¼ X2 ðjoÞ. b. Determine YðjoÞ ¼ X1 ðjoÞX2 ðjoÞ. c. Find the inverse transform of the function YðjoÞ to obtain yðtÞ. There are several problems with the transform method of computing the convolution. First, the given function may not have analytical expressions for the transforms. Even if does, we may not be able to find the

Fig. 4.5.1 Convolution of two square pulses

Using (4.4.25) and the time delay theorem, we have a triangle or a tent function given by yðtÞ ¼ L½t 1:

(4:5:5)

The given time functions and the result of the con& volution are shown in Fig. 4.5.1. Example 4.5.2 Consider the two delayed functions x1 ðt t1 Þ and x2 ðt t2 Þ. a. Assuming yðtÞ ¼ x1 ðtÞ x2 ðtÞ is known, show the following is true by using the transform method: zðtÞ ¼ x1 ðt t1 Þ x2 ðt t2 Þ ¼ yðt ðt1 þ t2 ÞÞ: (4:5:6) b. Using the results in (4.5.6), determine the convolution of the two impulse functions yðtÞ ¼ dðt t1 Þ dðt t2 Þ: Solution: a. Using the convolution and time delay theorems, we have

4.5 Convolution and Correlation

131

ZðjoÞ ¼ F½zðtÞ ¼ F½x1 ðt t1 ÞF½x2 ðt t2 Þ ¼ F½x1 ðtÞejot1 F½x2 ðtÞejot2 ¼ X1 ðjoÞX2 ðjoÞejoðt1 þt2 Þ ¼ YðjoÞejoðt1 þt2 Þ :

In (4.5.9b) we have made use of partial fraction expansion and the transforms of the unit step func& tion and the exponential decaying function. Example 4.5.5 Determine the convolution yðtÞ ¼ x1 ðtÞ x2 ðtÞ in each case below using the transforms. a. The two Gaussian functions and their transforms are given by

The inverse transform of ZðjoÞ is given by yðt ðt1 þ t2 ÞÞ. b. Noting that F½dðt ti Þ ¼ ejoti ; i ¼ 1; 2, we have h i 1 2 2 ðosi Þ2 =2 YðjoÞ ¼ ejoðt1 þt2 Þ ; F1 ejoðt1 þt2 Þ ¼ dðt ðt1 þ t2 ÞÞ: xi ðtÞ ¼ pﬃﬃﬃﬃﬃﬃ et =2si FT ¼ Xi ðjoÞ; i ¼ 1; 2:: !e si 2p (4:5:10a) That is, the inverse transform is a delayed impulse function and b. The two sinc functions and their transforms yðtÞ ¼ dðt t1 Þ dðt t2 Þ ¼ dðt ðt1 þ t2 ÞÞ: (4:5:7) &

Example 4.5.3 Show by using the transform method yðtÞ ¼ xðtÞ dðtÞ ¼ xðtÞ:

are given by

sinðtti =2Þ FT o ¼ Xi ðjoÞ; 2pP ! ðtti =2Þ ti (4:5:10b) i ¼ 1; 2; t1 5t2 :

xi ð t Þ ¼ ti

(4:5:8) c.

Solution: We have F½xðtÞ dðtÞ ¼ F½xðtÞF½dðtÞ ¼ F½xðtÞ and yðtÞ ¼ xðtÞ:

&

FT

xi ðtÞ ¼ 1=jpt !sgnðoÞ; i ¼ 1; 2::

Solution: a. Noting that the transform of a Gaussian pulse is a Gaussian pulse, the product of the two Gaussian pulses is a Gaussian pulse: FT

2=2

2=2

Example 4.5.4 Determine the convolution yðtÞ ¼ eat uðtÞ uðtÞ; a40 by a. the direct method and b. by the transform method.

yðtÞ ¼ x1 ðtÞ x2 ðtÞ !eðos1 Þ eðos2 Þ 2 2 2 ¼ eo ðs1 þs2 Þ=2 ¼ YðjoÞ:

Solution: a. By the direct method,

The inverse transform of this function is again a Gaussian pulse with

yðtÞ ¼

Z1

eaðtbÞ ½uðt bÞuðbÞdb ¼ eat

1

1 1 at ¼ eat eab b¼t ÞuðtÞ: b¼0 ¼ ð1 e a a

Zt

eab db

0

(4:5:9a)

b. By the transform method,

1 1 pdðoÞ þ YðjoÞ ¼ F½e uðtÞF½uðtÞ ¼ ða þ joÞ jo pdðoÞ 1 þ ; ¼ a joða þ joÞ at

1 1 1 1 YðjoÞ ¼ pdðoÞ þ ! a jo a a þ jo 1 yðtÞ ¼ ð1 eat ÞuðtÞ; a40: a

1 2 2 yðtÞ ¼ pﬃﬃﬃﬃﬃﬃ et =2s ; s2 ¼ s21 þ s22 : s 2p

(4:5:10c)

(4:5:10d)

b. The rectangular pulses in the Fourier domain overlap. The product of the two rectangular pulses is a rectangular pulse and its inverse transform is a sinc pulse. The details are left as an exercise. c. The convolution of the two functions and its transform are given by yð t Þ ¼

1 1 FT sgn2 ðoÞ ¼ 1; F1 ½1 ¼ dðtÞ: jpt jpt ! (4:5:10e) &

(4:5:9b)

Notes: The transform method is simpler if the transforms of the individual functions and the

132

4 Fourier Transform Analysis

inverse transform of the convolution are known. In Chapter 2 we discussed the duration property associated with convolution and pointed out that there are exceptions. Part c of the above example illustrates an exception. Convolutions of some functions do not exist. For example, yðtÞ ¼ uðtÞ uðtÞ does not exist since its transform has a term that is a square of an impulse function, which is not & defined.

4.5.2 Proof of the Integration Theorem In Section 4.4.6 the integration theorem is stated (see (4.4.42)) and is yð t Þ ¼

Zt

FT XðjoÞ þ pXð0ÞdðoÞ ¼ YðjoÞ: xðaÞda ! jo

1 1 F1 aþjo bþjo Zt at bt ¼e uðtÞe uðtÞ¼ eaa uðaÞebðtaÞ uðtaÞda

xðtÞ¼F1

0

¼

Zt e

1

xðaÞda¼

Z1

e

dt¼e

bt

eðbaÞt dt

0

Second, by using partial fraction expansion, we have XðjoÞ ¼

1 1 ; a 6¼ b ðb aÞða þ joÞ ðb aÞðb þ joÞ (4:5:12c)

) xðtÞ ¼ F 1 ½XðjoÞ ¼

1 eat uðtÞ ðb aÞ

1 ebt uðtÞ: ðb aÞ

(4:5:12d)

This coincides with the solution in (4.5.12b). b. When a ¼ b; the transform function has a double pole. By the convolution method,

xðaÞuðt aÞda

1 FT

¼ xðtÞ uðtÞ !XðjoÞF½uðtÞ;

1 YðjoÞ ¼ XðjoÞ pdðoÞ þ jo 1 ¼ pXð0ÞdðoÞ þ XðjoÞ: jo

at

xðtÞ ¼ e

uðtÞ e

at

uðtÞ ¼

Zt

eat eaðttÞ dt

0

¼ eat

Zt

dt ¼ teat uðtÞ:

(4:5:12e)

0

Since the function has a double pole, we can find its inverse transform by using times-t property of the Fourier transforms or from tables. Now

This proves the integration theorem. Example 4.5.6 Find the inverse transform of the function XðjoÞ given below for two cases: a. a 6¼ b; a40; b40 and b. a ¼ b40 XðjoÞ ¼ 1=½ða þ joÞðb þ joÞ:

Zt

1 h ðbaÞt it¼t eat ebt e uðtÞ: ¼ ¼ebt t¼0 ðbaÞ ðbaÞ (4:5:12b)

Since uðt aÞ ¼ 0 for a4t, we can write the above running integral as a convolution: Zt

at bðttÞ

0

1

(4:5:11)

1 d 1 1 ¼ ¼ XðjoÞ; ðjÞ do ða þ joÞ ða þ joÞ2

(4:5:12a)

( xðtÞ ¼ F

Solution: a. This can be solved by first noting that convolution in time domain corresponds to the multiplication in the frequency domain. Therefore,

1

1

) ¼ F 1

1 d 1 ðjÞ do ða þ joÞ

ða þ joÞ2 1 1 1 ¼ ðjtÞF ¼ teat uðtÞ : ðjÞ a þ jo

4.5 Convolution and Correlation

133

This coincides with the result obtained in (4.5.12e). & Example 4.5.7 Find yðtÞ for the function below by using a. the derivative theorem and b. the long division: YðjoÞ ¼

jo ; a40: ða þ joÞ

(4:5:13)

Solution: a. Using the derivative theorem, we have d½F1 ð1=ða þ joÞ dðeat uðtÞÞ ¼ dt dt at duðtÞ dðe Þ ¼ eat þ uðtÞ dt dt ¼ eat dðtÞ aeat uðtÞ ¼ dðtÞ aeat uðtÞ: (4:5:14)

yðtÞ ¼

b. Also, dividing jo by ða þ joÞ by long division and using the superposition property of the

Fourier transforms gives the same result as in (4.5.14). That is, jo a FT ¼1 dðtÞ aeat uðtÞ: a þ jo a þ jo !

&

4.5.3 Multiplication Theorem (Convolution in Frequency) The dual to the time convolution theorem is the convolution in frequency theorem. It is given below and can be shown directly by using a proof similar to the time domain convolution theorem. An alternate way of showing is by using the symmetry theorem FT 1 yðtÞ ¼ x1 ðtÞx2 ðtÞ ! X1 ðjoÞ X2 ðjoÞ: 2p

(4:5:15)

Summary: Convolution in time and in frequency: FT

Convolution in time: ½xðtÞ x2 ðtÞ !½X1 ðjoÞX2 ðjoÞ : Multiplication in frequency FT 1 Multiplicaion in time: ½x1 ðtÞx2 ðtÞ ! 2p ½X1 ðjoÞ X2 ðjoÞ : Convolution in frequency Example 4.5.8 Consider the time function and its transform hoi FT xðtÞ !XðjoÞ ¼ P : (4:5:16) W Find the Fourier transform of the function yðtÞ ¼ x2 ðtÞ and its bandwidth by assuming the bandwidth of xðtÞ is ðW=2Þ rad/s.

function L½o=W is (W). Note the duration property of the convolution is satisfied since the width of the triangular pulse is twice that of the rectangular pulse. From the area property of the convolution, we have using the time averages n h o io n h o io A P ¼ W2 : A P W W

Coming back to the bandwidths, if x1 ðtÞ and x2 ðtÞ have bandwidths of B1 and B2 Hz, respectively, then the bandwidth of yðtÞ ¼ x1 ðtÞx2 ðtÞ is equal to ðB1 þ B2 Þ Hz. Multiplication of two time functions hoi hoi h o i increases the bandwidth of the resulting time funcYðjoÞ ¼ XðjoÞ XðjoÞ ¼ P P ¼ WL : tion. The above property is dual to the time width W W W property of the convolution. We have seen some (4:5:17) & pathological cases where the time width property of the convolution does not hold. What about in the Notes: It is instructive to review the properties of frequency domain? Obviously, the same is true in the convolution of the transform functions in the frequency domain for pathological cases. For (4.5.17). The bandwidth of the pulse function practical signals, the above discussion applies. We P½o=W is W=2, whereas the bandwidth of the will come back to this topic at a later time, as it Solution: Example 2.3.1 considered the time domain convolution of two rectangular pulse functions. Using these results, we have

134

4 Fourier Transform Analysis

pertains to the important topic of nonlinear systems and the bandwidth requirements of such systems. Fourier transform computation of windowed periodic functions: Stanley et al. (1984) present a nice approach in finding the transforms of windowed time-limited trigonometric functions using the multiplication theorem, which is presented below. Let gT ðtÞ be a periodic function with period T and F½gT ðtÞ ¼ GðjoÞ. Let pðtÞ be a pulse function with PðjoÞ ¼ F½pðtÞ and a function F½wðtÞ ¼ WðjoÞ is defined by FT

wðtÞ ¼ pðtÞgT ðtÞ !PðjoÞ GðjoÞ:

(4:5:18)

We like to find F½wðtÞ using F½e jko0 t ¼ 2pdðo ko0 Þ and gT ðtÞ ¼

1 X

Gs ½ke jko0 t ;

k¼1

Gs ½k ¼

1 T

Z

gT ðtÞejko0 t dt;

T

o0 ¼ 2p=T; GðjoÞ ¼ F½gT ðtÞ 1 X Gs ½kdðo ko0 Þ: ¼ 2p

gT ðtÞ ¼ :54 þ :46 cosð2pt=TÞ;

(4:5:22a) ) wH ðtÞ ¼ gT ðtÞP

k¼1

¼ 2p ¼ 2p

1 X k¼1 1 X

Gs ½kPðjðo ko0 ÞÞ:

(4:5:20)

Example 4.5.9 Find the Fourier transform of the Hamming window function given below using the above method: :54 þ :46 cosð2pt=TÞ; jtj T=2 : wH ðtÞ ¼ 0; Otherwise (4:5:21) Solution: Define a periodic function using the window function in (4.5.21) by

(4:5:22b)

:

þ :23ð2pÞdðo þ o0 Þ; o0 ¼ 2p=T; (4:5:23) h h t ii WH ðjoÞ ¼ F gT ðtÞP ¼ GðjoÞ T

sinðoT=2Þ T sinðoT=2Þ :54ð2pÞ dðoÞ T ðoT=2Þ ðoT=2Þ

T sinðoT=2Þ þ :23ð2pÞ dðo o0 Þ ðoT=2Þ

T sinðoT=2Þ : þ :23ð2pÞ dðo þ o0 Þ ðoT=2Þ

1 dðo o0 Þ YðjoÞ ¼ 2p

Z1

dða o0 ÞYðjðo aÞÞda

1

1 ¼ Yðjðo o0 ÞÞ; 2p

we have

T sinðoT=2Þ T sinððo o0 ÞT=2Þ þ :23 ðoT=2Þ ððo o0 ÞT=2Þ T sinððo þ o0 ÞT=2Þ : þ :23 ððo þ o0 ÞT=2Þ

WH ðjoÞ ¼ :54

(4:5:24)

Gs ½kfPðjoÞ dðo ko0 Þg

k¼1

T

GðjoÞ ¼:54ð2pÞdðoÞ þ :23ð2pÞdðo o0 Þ

k¼1

The transform of the function wðtÞ can be obtained by the convolution of the two transform functions PðjoÞ ¼ F½pðtÞ and GðjoÞ. That is, " # 1 X WðjoÞ ¼ PðjoÞ 2p Gs ½kdðo ko0 Þ

hti

This function gT ðtÞ contains a constant and a cosine function. Its Fourier transform is

With

(4:5:19)

gT ðt þ TÞ ¼ gT ðtÞ;

Noting sinðoðT=2Þ o0 ðT=2ÞÞ ¼ sinðoðT=2Þ pÞ ¼ sinðoT=2Þ and using this in (4.5.24) results in WH ðjoÞ¼:54T

sinðoT=2Þ ðoT=2Þ

:23TsinðoT=2Þ

1 1 þ : ðoT=2Þp ðoT=2Þþp

In terms of f, we have " # T sinðpfTÞ :54 :08ðfTÞ2 ; o ¼ 2pf: WH ðjoÞ ¼ ðpfTÞ 1 ðfTÞ2 (4:5:25) &

4.5 Convolution and Correlation

135

4.5.4 Energy Spectral Density From Rayleigh’s energy theorem, the energy contained in an energy signal F½xðtÞ ¼ XðjoÞ can be computed either by the time domain function or by the frequency domain function and the energy contained in the signal is E¼

Z1

1 jxðtÞj dt ¼ 2p 2

1

¼

Z1

Z1

c: E:95

ZW :95 1 do ¼ ¼ do 2 2a 2p a þ o2 W 1 1 W tan ! a tanð:95ðp=2ÞÞ ¼ W ; ¼ ap a

W ¼ 2pF; F ð2:022aÞ Hz:

(4:5:27d) &

Example 4.5.11 Consider the pulse function xðtÞ ¼ P½t=t. Find the percentage of energy contained in the frequency range W5o5W; W ¼ 2pfc .

jXðjoÞj2 do

1

jXðjfÞj2 df; GðfÞ ¼ jXðjfÞj2 ¼ jXðjoÞj2 =2p:

1

Solution: The spectrum and the energy spectral densities are, respectively, given by

(4:5:26) Note that GðfÞ ¼ jXðjfÞj2 ¼ jXðjoÞj2 =2p is the energy spectral density. Example 4.5.10 a. Derive the energy spectral density of the function FT

xðtÞ ¼ eat uðtÞ !½1=ða þ joÞ ¼ XðjoÞ; a40 b. Illustrate the validity of Rayleigh’s energy theorem. c. Select the frequency band W5o5W so that 95% of the total energy is in this band. Solution: a. The energy spectral density is given by (4:5:27a) jXðjoÞj2 ¼ 1= a2 þ o2 : b. By using the time domain function, the energy contained in the function is ETotal ¼

Z1

2

jxðtÞj dt ¼

1

Z1

e2 at dt ¼

1 : (4:5:27b) 2a

0

XðjoÞ ¼ t

sinððo=2ÞtÞ ; ððo=2ÞtÞ

1 1 t2 sin2 ððo=2ÞtÞ : jXðjoÞj2 ¼ 2p 2p ððo=2ÞtÞ2

The total energy and the energy contained in the frequency range fc 5f5fc of the pulse are 2

ETotal ¼ ð1Þ t ¼ t; Efc ¼

ETotal ¼

1 2p

Z1

o 1 1 ¼ 1 : jXðjoÞj2 do ¼ tan1 2pa a 1 2a

1

(4:5:27c) The above two equations validate Rayleigh’s energy theorem.

Zfc

t2

fc

sin2 ðpftÞ ðpftÞ2

df: (4:5:29)

Using the change of variable b ¼ ft; df ¼ db=t; and f ¼ fc ! b ¼ fc t, the energy contained in the frequency range fc 5f5fc can be computed and the ratio of this to the total energy contained in the pulse. These follow

Efc ¼ 2t

Zfc t 0

We can make use of the frequency function to determine the energy as well and is

(4:5:28)

Efc ¼2 ETotal

Zfc t 0

sin2 ðpbÞ ðpbÞ2

sin2 ðpbÞ ðpbÞ2

db;

db:

(4:5:30a)

We can only compute this integral numerically. In the case of fc t ¼ 1, we have ðEfc =ETotal Þ :9028:

(4:5:30b)

136

4 Fourier Transform Analysis

That is, approximately 90% of the energy is contained in the spectral main lobe of the signal. If we include the side lobes, more energy will be included and Efc !1 ¼ ETotal . Ninety percent of the energy is reasonably sufficient to represent a rectangular pulse. & An interesting formula can be derived to find the energy of a causal signal xðtÞ, i.e., xðtÞ ¼ 0; t50 in terms of its real and imaginary parts of its transform XðjoÞ ¼ RðoÞ þ jIðoÞ. The energy is given by Papoulis (1977) as follows: Z1

x2 ðtÞdt ¼

2 p

0

Z1

R2 ðoÞdo ¼

2 p

0

Z1

¼

1 Z1

xðtÞxðt þ tÞdt

xðtÞxðt tÞdt¼ Rx ðtÞ:

xðtÞhðt þ tÞdt ¼

1

(4:6:1)

Z1

Rhx ðtÞ ¼

1

Gx ðoÞe jot do ¼F1 ½Gx ðoÞ ;

1 FT

(4:6:4)

Note that the autocorrelation function is the integral of the product of two functions, the function and its shifted version. It is a function of t, which is the shift between the given function and its shifted version. The Fourier transform pair relationship in (4.6.4) is referred to as the Wiener–Khintchine theorem. Also, we should note that a function xðtÞ and its delayed or advanced version xðt t0 Þ have the same autocorrelation function and therefore they have the same energy spectral densities. That is, Ry ðtÞ ¼

Z1 1 Z1

xðt t0 Þxðt t0 þ tÞdt

xðaÞxða þ tÞdt ¼ Rx ðtÞ

(4:6:5)

Gy ðoÞ ¼ Gx ðoÞ:

(4:6:6)

xða tÞhðaÞda;

Correlations were expressed in terms of convolution (see (2.6.4a and b)). Now, Rx ðtÞ ¼ xðtÞ xðtÞ:

(4:6:7)

Using the convolution theorem F½xðtÞ ¼ XðjoÞ ¼ X ðjoÞ, it follows that

and

1

(4:6:2a) Z1

Z1

1 Rx ðtÞ ¼ 2p

1

Note the single subscript in the case of autocorrelation and a double subscript in the case of cross-correlation below. The cross-correlations (see (2.6.3)) are

Rxh ðtÞ ¼

Rx ðtÞejot dt;

1

¼

1

Z1

Z1

Rx ðtÞ !Gx ðoÞ:

In this section we will see that the inverse Fourier transform of the energy spectral density discussed in the last section is the autocorrelation (AC) function defined in Chapter 2 (see (2.7.1)). The AC function of a real function xðtÞ is Rxx ðtÞ ¼

Gx ðoÞ ¼

0

4.6 Autocorrelation and CrossCorrelation

(4:6:3)

Cross-correlation reduces to the autocorrelation when hðtÞ ¼ xðtÞ. The Fourier transform of the AC function is the energy spectral density and is

I2 ðoÞdo: (4:5:31)

This can be shown by noting xðtÞ ¼ 2xe ðtÞuðtÞ ¼ 2x0 ðtÞuðtÞ and then using the transforms of real and imaginary parts. Details are left as an exercise.

Z1

Rhx ðtÞ ¼ Rxh ðtÞ:

hðtÞxðt þ tÞdt ¼

Z1

F½Rx ðtÞ ¼ F½xðtÞ xðtÞ ¼ F½xðtÞF½xðtÞ ¼ jXðjoÞj2 ¼ Gx ðoÞ:

(4:6:8)

hðb tÞxðbÞdb;

1

(4:6:2b)

Example 4.6.1 Show that the energy spectral densities of xðtÞ and xðt t0 Þ are the same and therefore

4.6 Autocorrelation and Cross-Correlation

137

the autocorrelation functions of the two functions are identical (see (4.6.5) and (4.6.6)). FT

Solution: Noting xðt t0 Þ !e jot0 XðjoÞ, we have e jot0 XðjoÞejot0 X ðjoÞ ¼ jXðjoÞj2 ¼ Gx ðoÞ: (4:6:9)

fx ðtÞ ¼

The energy spectral density and its inverse transform are (see (4.3.20)) jXðjoÞj2 ¼

The corresponding autocorrelation function is given by h i Rx ðtÞ ¼ F1 ½Gx ðoÞ ¼ F1 jXðjoÞj2 : (4:6:10) &

This example illustrates that the autocorrelation function Rx ðtÞ does not have the phase information contained in the function xðtÞ. This can be seen from the fact that t0 is not in either of the expressions Rx ðtÞ or Gx ðoÞ. The autocorrelation function is even and its spectrum, the energy spectral density, is real and even and Ex ¼ Rx ðtÞjt¼0 ¼

Z1

Z1

xðtÞxðtþtÞdtjt¼0 ¼

1

1 ajtj e : 2a

1

F

1 1 ¼ ; ða þ joÞða joÞ a2 þ o2

1 1 ¼ eajtj ¼ Rx ðtÞ: a2 þ o2 2a (4:6:12b)

b. By using the AC function and the energy spectral density, the energy in xðtÞ is Ex ¼ Rx ðtÞjt¼0 ¼ 1=2a, 1 p

Z1

1 1 1 1 o 1 tan ¼ ¼ Ex : do ¼ a2 þ o2 ap a 0 2a

0

(4:6:12c) & The cross-correlation signals is

jxðtÞj2 dt:

theorem

for

FT

Rhx ðtÞ !H ðjoÞXðjoÞ:

aperiodic (4:6:13)

1

(4:6:11a) This gives the energy in the signal. Using the Wiener–Khintchine theorem, we have

It can be shown by

Rhx ðtÞ ¼

Z1

hðtÞxðt þ tÞdt

1

Ex ¼ Rx ðtÞjt¼0

1 ¼ 2p

Z1

2

jXðjoÞj do:

1 ¼ 2p

(4:6:11b)

1

1 ¼ 2p Example 4.6.2 Consider the pulse function and its transform given by

¼

FT

xðtÞ ¼ eat uðtÞ !1=ða þ joÞ ¼ XðjoÞ; a40: (4:6:12a) a. Give the expression for the energy spectral density and its inverse transform, the corresponding autocorrelation function. b. Compute the energy in xðtÞ using its AC function and its energy spectral density. Solution: a. From Example 2.7.1 and (2.7.10), we have

1 2p

Z1

Z1 hðtÞ½

1 Z1

XðjoÞe joðtþtÞ dodt

1

Z1 XðjoÞ½

1 Z1

hðtÞe jot dte jot do

1

H ðjoÞXðjoÞe jot do:

1

Comparing these, (4.6.13) follows. For t ¼ 0, Rhx ð0Þ ¼

Z1 1

1 hðtÞxðtÞdt ¼ 2p

Z1

H ðjoÞXðjoÞdo:

1

(4:6:14) Equation (4.6.14) is a generalized version of Parseval’s theorem. The cross-correlation theorem

138

4 Fourier Transform Analysis

reduces to the case of autocorrelation by replacing HðjoÞ by XðjoÞ in (4.6.13).

Example 4.6.3 Consider the harmonic form of Fourier series of a periodic function

4.6.1 Power Spectral Density

xT ðtÞ ¼ Xs ½0 þ

N X

d½k cosðko0 t þ y½kÞ; o0 ¼ 2p=T:

k¼1

Earlier we have studied the power signals that include periodic signals and random signals. We will not be discussing random signals in this book in any detail. The autocorrelation of a periodic signal xT ðtÞ is given by Z 1 xT ðtÞxT ðt þ tÞdt: (4:6:15a) RT;x ðtÞ ¼ T

(4:6:16) Find its auto correlation. Solution: The autocorrelation function is given by (see (2.8.9)) RT;x ðtÞ ¼ X2s ½0 þ

T

Note that the above integral is over any one period. We have seen in Chapter 2 that the autocorrelation function of a periodic function is also a periodic function with the same period. The Fourier transform of the autocorrelation function is called the power spectral density (PSD) and its inverse is the autocorrelation function. The autocorrelation function and the corresponding spectral density function form a Fourier transform pair Sx ðoÞ ¼ F Rx;T ðtÞ ; Rx;T ðtÞ ¼ F1 ½Sx ðoÞ; FT

Rx;T ðtÞ !Sx ðoÞ:

(4:6:15b)

This relation is referred to as the Wiener–Khintchine theorem for periodic signals. The power spectral density of a periodic signal can be determined from Z1 1 Sx ðoÞdo ¼ Rx; T ð0Þ 2p 1 Z 1 xðtÞxðt þ tÞdtjt¼0 : ¼ T

N 1X d2 ½k cosðko0 tÞ: 2 k¼1

(4:6:17)

The phase terms y½ks are not in the autocorrelation function. The PSD is Sx ðoÞ ¼ 2pX2s ½0dðoÞ þ

N pX d2 ½kfdðo ko0 Þ þ dðo þ ko0 Þg: 2 k¼1

(4:6:18) The AC function and the PSD do not have any phase information and the frequencies are located at o ¼ ko0 ; k ¼ 1; 2; :::; N. The average power can be computed from the AC function or from the power spectral density. From (4.6.17), we have the average power Px ¼ RT;x ð0Þ ¼ X2s ½0 þ

N 1X d2 ½k: 2 k¼1

(4:6:19)

Using the power spectral density, we have

P¼

(4:6:15c)

T

Formal proof of the general Wiener–Khintchine theorem is beyond the scope here (see Ziemer and Tranter, 2002 and Peebles, 2001). In the following we will assume that Sx ðoÞ is given by the transform of the periodic autocorrelation function. Notes: Power signals include periodic and random signals. The autocorrelation function of a periodic function is periodic and the power spectral density & contains impulses.

1 Px ¼ 2p

Z1

Sx ðoÞdo

1

2p ¼ 2p

Z1

fX ½0dðoÞ þ 14 X d ½k:fdðo ko Þ N

2 s

2

1

þ dðo þ ko0 Þg

0

k¼1

gdo:

Since the integrand contains only impulses, the integral can be evaluated by inspection and is & RT;x ð0Þ in (4.6.19). Notes: In the case of energy signals the square of the magnitude spectrum gives the energy spectral

4.7 Bandwidth of a Signal

139

density (ESD), and the energy content is obtained by integrating the ESD. In the case of periodic signals, the spectrum contains impulses and the square of an impulse function is not defined. The AC function and the Wiener–Khintchine theorem are used to compute the power spectral density (PSD) of the periodic signal and its integral is the average power contained in the power signal. Most signals are corrupted by noise. Autocorrelation function ‘‘cleans’’ the signal and it provides a better insight into the essential qualities of the signal. Note that the AC function does not have the phase information in the signal. Since the AC of a periodic function is also a periodic with the same period, it can be determined by identifying the peaks in the autocorrelation function, thereby identifying the fundamental frequency. Finding the pitch period of a vowel in a noisy speech signal by using the autocorrelation function is very effective on a short& time basis (see Rabiner and Schafer, 1978).

4.7 Bandwidth of a Signal In Section 4.2.2, bandwidth (BW) of a signal was discussed in simple terms. One definition is the range of positive frequencies in which most of the signal energy or power is contained. This is vague since the word ‘‘most’’ can be interpreted differently. We will consider this here in more detail. A signal is considered time limited if the signal is zero outside an interval. For example, a pulse function P½ðt t0 Þ=t is nonzero for jt t0 j5t=2 and zero outside this range. It is nonzero for t s. The signals considered in this book are real signals and their spectral amplitudes are even and the spectral phase angles are odd. A signal xðtÞ is said to be band limited to B Hz if jXðjoÞj ¼ 0;

o ¼ j2pfj4W ¼ 2pB:

(4:7:1)

Since it is band limited to B Hz, B is defined as the bandwidth of the signal. In this case the signal occupies only low frequencies, i.e., it is a lowfrequency signal or sometimes referred to as a lowpass signal. Note that the bandwidth is defined using only positive frequencies. Band-pass signals are common in communication theory. Band-pass spectrum can be defined as follows:

8 0; joj5o0 ðW=2Þ > > > < H ; jo o j5ðW=2Þ 0 0 : jXðjoÞj ¼ > H0 ; jo þ o0 j5ðW=2Þ > > : 0; joj4o0 þ ðW=2Þ

(4:7:2)

It is an ideal band-pass signal. Most practical signals are not band limited. There are functions that are neither time limited nor band limited. For example, consider the double exponential function given below and its transform derived earlier FT xðtÞ ¼ eajtj !2a= a2 þ o2 ¼ XðjoÞ; a40: (4:7:3) The question we need to answer is, what is a meaningful definition of the time width of an arbitrary nontime-limited signal? How about a meaningful definition of the frequency width of an arbitrary non-bandlimited signal? In Section 4.2.2, we have seen that a shorter time width signal corresponds to a broader spectrum. For example, the Fourier transform pair of a rectangular pulse function is given by xðtÞ ¼ P

hti t

sinðot=2Þ !t ðot=2Þ ¼ XðjoÞ:

FT

(4:7:4)

Most of the energy is contained in the main lobe of the spectrum, which occupies the frequency band between the two zeros of XðjoÞ located at o ¼ 2p=t. The energy content of this pulse is quantified in terms of the frequency content in Example 4.5.12. The side lobes contain a small portion of the energy. The time width of the pulse is obviously equal to t s and the frequency width is approximately 1=t Hz, considering only positive frequencies. Increasing (decreasing) the time width reduces (increases) the frequency width. At least, from this example, we see that the two widths are inversely proportional to each other. In the following we will consider a few standard definitions of time and frequency widths and they give some meaning.

4.7.1 Measures Based on Areas of the Time and Frequency Functions Using the ordinate theorems discussed earlier, we have

140

4 Fourier Transform Analysis

Z1 Xð0Þ¼

xðtÞdt; xð0Þ¼ 1

Z1

1 2p

XðjoÞdo: (4:7:5) 1

The dc value of a signal is zero if it is an odd function. Interestingly, n

xðtÞ ¼ d yðtÞ=dt

1 FT ¼ XðjoÞ; a40: xðtÞ ¼ eat uðtÞ ! a þ jo

n FT

n

!ðjoÞ YðjoÞ ¼ XðjoÞ:

(4:7:6)

It is zero at o ¼ 0 provided YðjoÞ has no poles at the origin that can be canceled by ðjoÞn . If the time function has a discontinuity at the origin, then xð0Þ is obtained from the integral in (4.7.5) and is the average value or the half-value at the discontinuity. Let the time width and the frequency widths, respectively, be defined by R1

R1

xðtÞdt ; tw ¼ 1 xð0Þ

ow ¼ 1

XðjoÞdo : Xð0Þ

1 Ð 0

Z1

1 a: xð0Þ ¼ 2p ¼

1 2p

1 Z1

eat dt ¼ 1a.

XðjoÞe jot dojt¼0 a a2 þ ð o Þ

1

do þ 2

j 2p

Z1 1

ðoÞ a2 þ ðoÞ2

do:

The integrand in the second integral is an odd function and therefore it is zero: 1 xð0Þ ¼ 2p

Z1

a 1 1 tan1 ðo=aÞ1 do ¼ 1 ¼ : a2 þ o2 2p 2

1

(4:7:7)

Then the product tw ow ¼ 2p:

Solution: a. Xð0Þ ¼

(4:7:8)

That is, the product of time and frequency widths is a constant. Some authors use the frequency f in Hz rather than o ¼ 2pf in rad/s in the time–bandwidth product. These simple measures have drawbacks illustrated below. Example 4.7.1 Consider the function given by xðtÞ ¼ P½t :5 P½t þ :5. It is an odd function. If the above definition is used, the time width is zero, even though the actual width of this function & is 2 s. Example 4.7.2 Consider the pair

Note the exponential time-decaying function xðtÞ is discontinuous at t ¼ 0 and the above result verifies that the inverse transform converges to the halfvalue, i.e., the average value of the function before & and after the discontinuity.

4.7.2 Measures Based on Moments The time width Tw of a real non-time-limited function is defined by ðTw Þ2 ¼

Z1

1 kxk

Ex ¼ kxk2 ¼

2

ðt tÞ2 x2 ðtÞdt; kxk2 ¼

1

Z1

The area under the pulse function is 1. The area under the sinc function is 2p and the time–band& width product is 2p. Example 4.7.3 Consider the Fourier transform pair corresponding to the exponential decaying function to find the values of the functions a. Xð0Þ and b. F1 ½XðjoÞjt¼0 :

x2 ðtÞdt;

1

x2 ðtÞdt ¼

1

sinðo=2Þ : P½t ! ðo=2Þ FT

Z1

1 2p

Z1

(4:7:9) jXðjoÞj2 do51:

1

(4:7:10) The center of gravity of the area of the function is 1 t ¼ Ex

Z1

tx2 ðtÞdt:

(4:7:11)

1

Tw is a measure of the signal spread about t and is the signal dispersion in time. Notes: These measures can be seen noting pðtÞ ¼ ½x2 ðtÞ=Ex >0 is a valid probability density

4.7 Bandwidth of a Signal

141

function, as is nonnegative for all t and the area under it is 1. In statistical terminology t is the mean & and ðTW Þ2 is the variance (see Peebles, 2001). Example 4.7.4 Consider the exponential decaying function xðtÞ ¼ eat uðtÞ; a40. Find the center of gravity and the time dispersion Tw . Solution: The energy contained in the pulse is E ¼ ð1=2aÞ. The center of gravity is Ð 1 2 at te dt e2 at ½1 þ 2 at 1 1 ¼ 2a t ¼ 0 : 0 ¼ 2 ð1=2aÞ 2a ð2aÞ

A bound on the time–bandwidth product Tw Ww ¼ Tw ð2pFw Þ is derived using Rayleigh’s energy theorem and Schwarz’s inequality (see Section 2.1.). The inequality is briefly reviewed below. Schwarz’s inequality: The inequality is (see (2.1.9d)) khxðtÞþyðtÞik kxðtÞkkyðtÞk Zb Zb Zb 2 2 ) jxðtÞyðtÞj dt jxðtÞj dt jyðtÞj2 dt: a

a

a

(4:7:15)

(4:7:12) In addition, using the following integral formulas, Tw is as follows:

t2 2t 2 2þ 3 ; b b b

Z Z 1 1 bt bt t ; and ebt dt ¼ ebt ; te dt ¼ e b b2 a

Z

t2 ebt dt ¼ ebt

4.7.3 Uncertainty Principle in Fourier Analysis

The uncertainty principle in spectral analysis states that if the integrals in (4.7.9) and (4.7.14a) are finite and pﬃﬃ lim txðtÞ ¼ 0; (4:7:16) t!1

ðTw Þ2 ¼2a

Z1

then ðtð1=2aÞÞ2 e2at dt

Tw Ww

0

Z ¼ð2aÞ

1

½t2 ð1=aÞtþð1=2aÞ2 e2at dt¼ð1=2aÞ2

0

)Tw ¼ð1=2aÞ:

(4:7:13)

The frequency width, a frequency measure, Ww ¼ 2pFw , can be defined by

W2w

¼

kXk2

1 ¼ Ex ð2pÞ Note : o

k X k2 ¼

Z1

1

Z1

1 1 or Tw Fw

; Ww ¼ 2pFw : 2 2ð2pÞ

(4:7:17) Using the expressions for Tw and Ww from (4.7.9) and (4.7.14a) results in ðTw Ww Þ2 ¼

2 jXðoÞj2 do ðo oÞ

3

Z1

4 ðt tÞ2 x2 ðtÞdt5 kx k 2 kX k2 1 2 1 3 Z o2 jXðjoÞj2 do5;

4 1

1

R1

2

1

2 ojXðjoÞj do ¼ 0 ; (4:7:14a)

(4:7:18)

1 2

kX k ¼

jXðjoÞj2 do ¼ 2pkxk2

Z1

2

2

jXðjoÞj do; kxk ¼

1

Z1

2 x ðtÞdt;

1

1

ðRayleigh’s energy theoremÞ:

(4:7:14b)

This follows since the integrand in the above equation is odd and the integral of an odd function over & a symmetric interval is zero.

kXk2 ¼ 2pkxk2 :

(4:7:19)

Noting the Fourier transform derivative theorem, i.e., F½x0 ðtÞ ¼ ðjoÞXðjoÞ and using Rayleigh’s energy theorem results in

142

4 Fourier Transform Analysis

Z1

1 jx ðtÞj dt ¼ 2p 2

0

1

Z1

2

2

o jXðjoÞj do: (4:7:20)

Z1

1 kxk2 kXk2 1 2

kxk kXk

¼

kxk4

2 ðt tÞ2 x2 ðtÞdt4

1 1 ð

Z1

3 o2 jXðjoÞj2 do5

1

ðt tÞ2 x2 ðtÞdtð2pÞ

2

Z1

1

1

Z1

Z1

2

eat dt¼

1

ðt tÞ2 x2 ðtÞdt

1

½x0 ðtÞ2 dt:

(4:7:21)

1

Canceling the negative signs in (4.7.25b) results in Z1

rﬃﬃﬃﬃﬃ p dt ¼ : 2a

(4:7:26)

From tables,

1

Z1

x2 ðtÞdt:

2

k xk ¼

Z1

1

(4:7:23)

2 at2

t e

t!1

1 x2 ðtÞdt ¼ kxk2 : 2

e

2 at2

Now let a ¼ 2a in (4.7.25c), which results in

Assuming that lim ðt tÞx2 ðtÞ ¼ 0, it follows that Z1

Z1 1

1

1

(4:7:25c)

(4:7:22)

dx 1 1 dt ¼ ðt tÞx2 ðtÞ1 1 dt 2 2

1 2

rﬃﬃﬃ p : a

1 ¼ 2a

1

Considering the right-hand side of the above equation and integrating it by parts, we have Ð Ð udv ¼ uv vdu, with u ¼ ðt tÞ; du=dt ¼ 1; dv=dt ¼ xðtÞx0 ðtÞ; and v ¼ ð1=2Þx2 ðtÞ

ðt tÞxðtÞx0 ðtÞdt ¼

2 at2 dt

t e

1

1

Z1

(4:7:25b)

1

2 1 Z 0 ðt tÞxðtÞx ðtÞdt :

ðt tÞxðtÞ

1

pﬃﬃﬃ dða1=2 Þ pﬃﬃﬃ a1:5 ¼ p ¼ p 2 da rﬃﬃﬃ 1 p : ¼ 2a a

½x0 ðtÞ2 dt

1

Z1

rﬃﬃﬃ Z1 Z1 at2 p de 2 dt ¼ ) ðt2 Þeat dt da a

1

Using Schwarz’s inequality in (4.7.15) results in 2 1 32 1 3 Z Z 4 ðt tÞ2 x2 ðtÞdt54 ½x0 ðtÞ2 dt5

Z1

rﬃﬃﬃ p ðoÞ2=4a ¼ XðjoÞ; a40: (4:7:25a) ! ae

FT

Solution: First, differentiate both sides of the following equation with respect to a:

ðTw Ww Þ2

¼

2

1

Therefore

¼

xðtÞ ¼ eat

1

dt ¼

Z1

2 2 at2 dt

t e

1 ¼ 4a

rﬃﬃﬃﬃﬃ p : 2a

1

Noting that the Gaussian pulse in this example is even and the integrand in (4.7.11) is odd, it follows that t ¼ 0. The time width can be computed from (4.7.9) and

Using this and (4.7.21) in (4.7.20), we have 2

4

1 1 kxk ¼ or Tw Ww

2 4kxk4 4 1 ; Ww ¼ 2pFw : Tw F w

2ð2pÞ

ðTw Ww Þ2

ðTw Þ ¼

or

1 kxk2

Z1

t2 e2 at dt ¼ ð1=

1

(4:7:27)

(4:7:24)

Example 4.7.5 Illustrate the uncertainty principle using the Gaussian transform pair

rﬃﬃﬃﬃﬃ rﬃﬃﬃﬃﬃ p 1 p 1 Þ ¼ : 2a 4a 2a 4a

Noting that kXk2 ¼ 2pkxk2 ¼ 2p that

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ p=2a, it follows

4.8 Moments and the Fourier Transform

ðWw Þ2 ¼

¼

Z1

1 kXk2

143

will derive the transform in terms of mi by using the power series expansion

o2 jXðjoÞj2 do

1

Z1

1 kxk2 ð2pÞ

ejot ¼

p 2ðoÞ2 o2 e 4a do a

1

pﬃﬃﬃﬃﬃ Z1 2a p o2 pﬃﬃﬃð Þ ¼ o2 e 2a do ð2pÞ p a 1 pﬃﬃﬃﬃﬃ 2a p pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃ a pð2aÞ ¼ a ) ðTw Ww Þ2 ¼ ð2pÞ pa 1 1 (4:7:28) ¼ a¼ : 4a 4

1

1

¼

1

The moment theorem relates the derivatives of the transform of a function at o ¼ 0:

m0 ¼

Z1

n ¼ 0; 1; 2; ::: ;

(4:8:2)

Z1

dXðjoÞ jo¼0 ¼ do

xðtÞ

dejot dtjo¼0 do

1 Z1

¼j

n!

xðtÞdt

tn xðtÞdt

1

(4:8:4)

This holds only if the integral of the terms in the above equation is valid. From (4.8.2) XðjoÞ ¼

1 X

ðjÞn mn ½on =n!:

(4:8:5)

n¼0

Although the moment theorem is given in terms of a series expansion, it can be used to compute the transforms of functions (see Papoulis, 1962). Example 4.8.1 Use the moment theorem and the following identity to show that Fourier transform of the 2 Gaussian pulse xðtÞ ¼ eat is also a Gaussian pulse. Z1

at2

rﬃﬃﬃﬃﬃ rﬃﬃﬃ Z1 p 1 p 2 ; a40: dt ¼ t2 eat dt ¼ ) a 2 a3

1

1

(4:8:6a)

xðtÞdt ¼ XðjoÞjo¼0 (Ordinate theorem);

1

Z1 1 X ðjoÞn n¼0

e d n XðjoÞ ðjÞ mn ¼ jo¼0 ; do n n

n!

n¼0

1 n X d XðoÞ on : ¼ j o¼0 n! don n¼0

4.8 Moments and the Fourier Transform The nth moment mn of xðtÞ is defined by (see Section 1.7.) Z1 tn xðtÞdt; n ¼ 0; 1; 2; ::: : (4:8:1) mn ¼

Z1 X 1 ðjoÞn tn

¼

(4:7:29)

This shows the equality in (4.7.24) in the Gaussian & case. See Hsu (1967) for additional examples.

(4:8:3)

Substituting this in the transform and using (4.8.1) and (4.8.2) result in Z1 XðjoÞ ¼ F½xðtÞ ¼ xðtÞejot dt

The time–bandwidth product of a Gaussian pulse is obtained by using Ww ¼ 2pFw and Tw Fw ¼ 1=ð2ð2pÞÞ:

1 X 1 ðjotÞn : n! n¼0

Solution: Equation (4.8.6a) on the right can be generalized and 1 ð 1

txðtÞe

jot

dtjo¼0 ¼ jm1 :

1

Repeating this process and evaluating the derivatives at o ¼ 0 proves the result in (4.8.2). Now we

2

rﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ p a2nþ1 rﬃﬃﬃ 1:3:::ð2n 1Þ p ¼ ¼ m2n : ð2aÞn a

t2n eat dt ¼

1:3:::ð2n 1Þ 2n

(4:8:6b)

This gives even moments and the odd moments are zero since xðtÞ is even:

144

4 Fourier Transform Analysis

XðjoÞ ¼ ¼

1 X n¼0 1 X

ðjoÞ2n

m2n ð2nÞ!

ð1Þn ðoÞ2n

n¼0

1 1:3:::ð2n 1Þ ð2nÞ! ð2aÞn

rﬃﬃﬃ p : a

jXðjoÞj

This can be expressed in a compact form by noting ð1Þð3Þð5Þ:::ð2n 1Þ ð1Þð3Þð5Þ:::ð2n 1Þ ¼ n n ð1=2 Þð2nÞ! ð1=2 Þð1Þð2Þð3Þ:::ð2n 1Þð2nÞ 1 ¼ ð1=2n Þð2Þð4Þ:::ð2nÞ 1 1 ¼ ; ¼ ð1Þð2Þ:::ðnÞ n!

a ¼ XðjoÞ; a40; ½a þ jo

it follows that

ða þ joÞ2

jo¼0

1 Z Z1 jot xðtÞejot dt xðtÞe dt jXðjoÞj ¼ 1

¼

Z1

1

jxðtÞjdt: FT

1 joj

Z1 dxðtÞ dt dt:

1

In a similar manner, we can prove the third bound in (4.9.2) if it exists. Example 4.9.1 Find the first two bounds in (4.9.2) using the Fourier transform pair FT

&

4.9 Bounds on the Fourier Transform In Chapter 3, we have learned that the derivative of a periodic function plays a role on the bounds on its F-series coefficients. We can use the Fourier time differentiation theorem to find the bounds on the transform. First d n xðtÞ FT n !ðjoÞ XðjoÞ: dtn

First bound:

xðtÞ ¼ P½t !½sinðo=2Þ=ðo=2Þ ¼ XðjoÞ:

m0 ¼ Xð0Þ ¼ 1; m1 ¼ j ja

(4:9:2)

1

Solution: Noting

¼j

jxðtÞjdt

Second bound: With x0ðtÞ !joXðjoÞ, we have Z1 dxðtÞ jot ðjoÞXðjoÞ ¼ e dt ! jXðjoÞj dt &

Example 4.8.2 Find the first two moments of xðtÞ ¼ aeat uðtÞ; a40.

dXðjoÞ jo¼0 do a 1 ¼ 2¼ : a a

1 Ð

1 1 Ð dxðtÞ 1 dt dt joj > 1 > > > 1 > Ð d 2 xðtÞ > > : o12 dt2 dt 1

1

rﬃﬃﬃ 1 p X ð1Þ2 o2n XðjoÞ ¼ a n¼0 ð4aÞn n! pﬃﬃﬃﬃﬃﬃﬃﬃ 2 2 FT pﬃﬃﬃﬃﬃﬃﬃﬃ 2 ¼ p=aeo =4a ) eat ! p=aeo =4a :

F½xðtÞ ¼ F½aeat uðtÞ ¼

8 > > > > > > >

s = s + jo-plane, contour of integration, Example 5.4.1

X2;II ðsÞ ¼

Z0

x2 ðtÞe

st

dt ¼

1

transform and can be obtained from the Laplace transform by simply substituting s ¼ jo in the Laplace transform. The example we have consid& ered above exists for positive time. A two-sided signal can be written in two parts – one for positive time (causal) including t ¼ 0 and the other part for negative time (anti-causal). The twosided function xðtÞ can be separated into two parts x1 ðtÞ and x2 ðtÞ, representing the causal and the anti-causal part, respectively. See Fig. 5.4.2, where we have the following: xðtÞ : twosided;

x2 ðaÞest da

0

x2 ðaÞest da

0

) X2;II ðsÞ ¼

Z1

x2 ðtÞest dt:

(5:4:6b)

0

That is, we have expressed the two-sided Laplace transform in terms of two one-sided Laplace transforms. Changing the sign of s in the function X2;II ðsÞ gives X2;II ðsÞ. Notes:

x1 ðtÞ ¼ xðtÞuðtÞ : causal part x2 ðtÞ ¼ xðtÞuðtÞ : anticausalpart; x2 ðtÞ ¼ xðtÞuðtÞ : invertedanticausalpart x1 ðtÞ and x2 ðtÞ are now causal signals. The bilateral or two-sided transform is given by

Fig. 5.4.2 (a) Two-sided signal, (b) causal, (c) anticausal, and (d) inverted anticausal parts

¼

Z1

Z1

1. Express xðtÞ ¼ ½xðtÞuðtÞ þ ½xðtÞuðtÞ ¼ x1 ðtÞ þ x2 ðtÞ. 2. Use the causal signals x1 ðtÞ and x2 ðtÞ and find their one-sided transforms. 3. The two-sided Laplace transform of the signal is & given by XðsÞ ¼ X1 ðsÞ þ X2 ðsÞ.

(a)

(b)

(c)

(d)

5.4 Laplace Transforms

163

Example 5.4.2 Find the two-sided Laplace transform of the function ( xðtÞ ¼

)

ebt ;

b40; t50

eat ;

a40; t 0

(5:4:7)

¼ ½ebt uðtÞ þ ½eat uðtÞ ¼ x2 ðtÞ þ x1 ðtÞ

To see these constraints, we can separate the bilateral Laplace transform into two integrals, one for positive time and the other for negative time. That is, Z 1 xðtÞest dt XII ðsÞ ¼ LII ½xðtÞ ¼ ¼

0

xðtÞe

Solution: From Example 5.4.1, we have 1 ðs þ aÞ ¼ X1 ðsÞ; ReðsÞ ¼ s4 a:

1

Z

st

dt þ

Z

1

LTII

x1 ðtÞ ¼ eat uðtÞ !

(5:4:8a)

1 ðs þ bÞ ¼ X2 ðsÞ; s ¼ ReðsÞ4 b; LTII

xðtÞest dt:

¼

Z

Z

0 bt st

Me e 1 Z 0

dt þ

MeðbsÞt dt þ

1

0 Z 1

1

Meat est dt MeðaþsÞt dt:

0

(5:4:10c)

LTII

x2 ðtÞ ¼ ebt uðtÞ ! X2 ðsÞ 1 ¼ ; s ¼ ReðsÞ5b: sb

(5:4:8b)

Combining the two equations in (5.4.8a) and (5.4.8b), we have LTII

xðtÞ !

(5:4:10b)

0

The integrals in the above equation must be absolutely integrable in order for the transform to exist. Using (5.4.10a), we have XII ðsÞ

x2 ðtÞ ¼ ebt uðtÞ !

1

1 1 þ ; a5s5b: ðs bÞ ðs þ aÞ

Noting that limits on the integration, we can state that the transform exists if ðb sÞ40;

ða þ sÞ40 ! a5s5b:

(5:4:11)

This defines the ROC and is illustrated by the dark area in Fig. 5.4.3.

(5:4:9)

Note the constraints on s. The transform does not exist if ðaÞ4b. This will be tied to the region of convergence shortly. From this example the causal signal results in the transform that has poles on the left half of the s-plane and the anti-causal signal results in the transform that has poles on the right & half of the s-plane.

jω

Poles for

Poles for

x (t )u ( −t )

x (t )u (t ) −α

0

β

σ

5.4.1 Region of Convergence (ROC) We defined the two-sided Laplace transform of a signal xðtÞ by assuming that the function (xðtÞest ) is absolutely convergent. This implies that there exists a pair of constants a and b and M a real positive number, such that Meat ; t40 : (5:4:10a) xðtÞ Mebt ; t50

Fig. 5.4.3 Region of convergence for bilateral Laplace transform

Now we will relate this to a rational Laplace transform function XII ðsÞ in terms of its poles and zeros. We say that XII ðsÞ has a pole at s ¼ pi if XðsÞs¼pi ¼ 1 and has a zero at s ¼ zi

164

5 Relatives of Fourier Transforms

if XðsÞjs¼zi ¼ 0. The poles of the function XðsÞ must lie to the right of the line s ¼ b þ jo for t50, whereas the poles corresponding to the positive time, i.e., xðtÞ; t40 must lie to the left of the line s ¼ a þ jo in the complex s-plane. Example 5.4.3 Find the two-sided Laplace transform of the following function directly: xðtÞ ¼ ea j tj ¼ eat uðtÞ þ eat uðtÞ; a40:

(5:4:12)

Solution: The transform of this function is XðsÞ ¼

Z

0

eðasÞt es t dt þ 1

¼

Z

1

eðaþsÞ t dt

0

1 1 2a þ ¼ 2 ; a5s5a: ða sÞ ðs þ aÞ a s2 (5:4:13) &

The region of convergence is indicated on the right of (5.4.13). The Laplace transform does not converge at the pole locations and in Fig. 5.4.1 we have shown that the region of convergence goes around the poles (see the half moons around the poles). The region of convergence includes the jo-axis and the function evaluated on the jo-axis gives the Fourier transform of the function. That is, XðjoÞ ¼ XII ðsÞs¼jo :

XII ðsÞ ¼

Z

1

eat est dt ¼

0

YII ðsÞ ¼

Z

1 ; sþa

0

eat est dt ¼

1

1 ; ReðsÞ5 a: sþa (5:4:15)

We can see that the transforms are the same, except that they have different regions of convergence. These indicate the ambiguity in identifying the time function that the transform came from if the region of convergence of a two-sided Laplace transform is not given. In this case it is not possible to compute the corresponding time function. However, in practice, the ambiguity can be resolved on the basis of physical considerations, as the time functions increase without limit as t approaches either þ1 or 1. This gives a way to select a time function among & many possibilities. Procedure to find the two-sided inverse LT of a rational function: 1. Expand the given rational transform function by using partial fraction expansion. 2. The terms in the partial fraction expansion that come from the left half-plane poles will result in time functions that exist only for t 0. 3. The terms in the partial fraction expansion that come from the right half-plane poles will result in the time functions that exist only for t50. Example 5.4.5 Find the inverse transform of the XII ðsÞ ¼ 1=½ðs 1Þðs þ 2Þ.

5.4.2 Inverse Transform of Two-Sided Laplace Transform One needs to be careful in finding the inverse transforms of a two-sided Laplace transform, as the positive and negative time portions must be handled separately. If the region of convergence is not specified, then there is some ambiguity. Example 5.4.4 Consider the time functions xðtÞ ¼ e

at

uðtÞ;

at

yðtÞ ¼ e

uðtÞ:

(5:4:14)

Solution: The two-sided Laplace transforms of these functions are given by

Solution: The partial fraction expansion of this function is A B þ ; ðs 1Þ ðs þ 2Þ 1 1 A¼ js¼ 1 ¼ ; ðs þ 2Þ 3 1 1 B¼ js¼ 2 ¼ ðs 1Þ 3

XII ðsÞ ¼

) XII ðsÞ ¼

1=3 1=3 þ ; ðs 1Þ ðs þ 2Þ

) xðtÞ ¼ ð1=3Þet uðtÞ ð1=3Þe2t uðtÞ; 25ReðsÞ51:

(5:4:16)

(5:4:17)

(5:4:18)

5.5 Basic Two-Sided Laplace Transform Theorems

The region of convergence can be obtained by noting that the first term has a pole at s ¼ 1 and the region of convergence for this term is s ¼ ReðsÞ51. The second term has a pole at s ¼ 2 and the region of convergence for this term is s4 2. The intersection of these two regions of convergence is given & by 25ReðsÞ51. The two-sided Laplace transform of a function can be determined by decomposing the two-sided function into two one-sided functions and then transforming each one. In the case of finding the inverse Laplace transform, we first separate the transform into two parts, one with poles on the right half-plane and other with poles in the left halfplane and the imaginary axis. Then determine the two time functions. Most of our discussion on inverse transforms covers rational functions.

5.4.3 Region of Convergence (ROC) of Rational Functions – Properties 1. The ROC does not contain any poles of the function. 2. If xðtÞ ¼ 0, except in a finite interval, then the ROC is the entire s-plane except possibly s ¼ 0 and s ¼ 1. 3. If xðtÞ is right-sided, then the ROC is right-sided, i.e., s ¼ ReðsÞ4 a, where (a) is the real part of the left-most pole. 4. If xðtÞ is left-sided, then the ROC is left-sided, i.e., s ¼ ReðsÞ5b , where b is the real part of the right-most pole. 5. If xðtÞ is two-sided function, i.e., the sum of leftand right-sided functions, then the ROC is either a strip defined by a5ReðsÞ5b or the individual regions of convergence will not overlap and, in that case, the ROC is the null set.

5.5 Basic Two-Sided Laplace Transform Theorems Now consider some of the important two-sided Laplace transform theorems that are given below without proofs for most. Assume the region of convergence of xi ðtÞ is Rxi .

165

5.5.1 Linearity The Laplace transform of a sum is the sum of the Laplace transforms and can be stated as 2 2 X LTII X xðtÞ ¼ ai xi ðtÞ ! ai Xi ðsÞ (5:5:1) i¼1 i¼1 ¼ XII ðsÞ; ROC : at least Rx1 \ Rx2 Example 5.5.1 Let xðtÞ ¼ x1 ðtÞ þ x2 ðtÞ; x1 ðtÞ ¼ et uðtÞ; x2 ðtÞ ¼ et uðtÞ. Find the Laplace transform of the function xðtÞ and identify the region of convergence. Solution: XII ðsÞ ¼

1 1 þ : sþ1 s1

The ROC of x1 ðtÞ is ReðsÞ4 1 and the ROC of x2 ðtÞ is ReðsÞ51: The ROC of xðtÞ is the intersection of the two and is given by & 15ReðsÞ51.

5.5.2 Time Shift LII ½xðt t0 Þ ¼ est0 XII ðsÞ:

(5:5:2)

The region of convergence is the same for both the original and its shifted version.

5.5.3 Shift in s LII ½eat xðtÞ ¼ XII ðs þ aÞ:

(5:5:3)

Since the poles will be shifted to the left by a, the ROC will be shifted to the left by a.

5.5.4 Time Scaling LII ½xðatÞ ¼

1 XII ðs=aÞ; ðROCÞnew ¼ ðROCÞold =a: jaj (5:5:4)

Time scaling makes the ROC scaled as well.

166

5 Relatives of Fourier Transforms

5.5.5 Time Reversal LII ½xðtÞ ¼ XII ðsÞ; ðROCÞnew ¼ ðROCÞold :

(5:5:5)

Note that the right half-plane poles become left half-plane poles and vice versa.

5.5.6 Differentiation in Time LII

dxðtÞ ¼ sXII ðsÞ; ROCnew ROCold : dt

(5:5:6)

The ROC will not change unless there is a pole–zero cancellation in the product ðsXII ðsÞÞ.

5.5.7 Integration Z LII

t

1 xðaÞda ¼ XII ðsÞ: s 1

(5:5:7)

Noting the term (1/s) in the transform and ROCnew ¼ ROCold IfReðsÞ 4 0g.

5.5.8 Convolution In Chapter 2 we defined the convolution of two functions by Z 1

x1 ðaÞx2 ðt aÞda

yðtÞ ¼

¼

1 Z 1

(5:5:8a) x2 ðaÞx1 ðt aÞda:

1

The transform is Z 1 Z 1 x1 ðaÞx2 ðt aÞda est dt YII ðsÞ ¼ 1 1 Z 1 Z 1 st x1 ðaÞ x2 ðt aÞe dt da ¼ 1 Z1 1 ¼ x1 ðaÞXII;2 ðesa Þda 1 Z 1 ¼ XII;2 ðsÞ x1 ðaÞesa da ¼ XII;2 ðsÞXII;1 ðsÞ: 1

(5:5:8b)

The ROC satisfies ROCnew ðROCÞ1 \ ðROCÞ2 . The ROC of the convolution may be larger. When two transforms are multiplied, there is a possibility of pole cancellations.

5.6 One-Sided Laplace Transform So far we have been discussing the bilateral or twosided transform. A special form of the bilateral transform is the one-sided or unilateral or simply Laplace transform, which was defined in (5.4.4). It was pointed out that the bilateral transform can be computed by using the one-sided Laplace transform. In real-life systems, there is no negative time. However, bilateral transforms provide a structure that we can work with, as the bilateral Laplace transforms relate to the Fourier transforms. We can make the discussion simpler by considering the unilateral Laplace transform. The unilateral transform is fundamental in circuits, systems, and control, where we are interested in the response of a system with initial conditions. In a later chapter we will describe a linear timeinvariant system by constant coefficient differential equations. The unilateral Laplace transform provides a powerful tool in the analysis and design of systems. As mentioned earlier we will use the notation XðsÞ ¼ LfxðtÞg and xðtÞ is a causal signal. Furthermore, the ROC is the right half s-plane for the unilateral Laplace transforms. For simplicity, we generally do not explicitly identify the region of convergence. Unless otherwise mentioned, we will assume that the transform functions are unilateral and will not be mentioned explicitly. The unilateral Laplace transform of a signal xðtÞ is defined earlier and is repeated below: XðsÞ ¼ L½xðtÞ ¼

Z

1

xðtÞest dt:

(5:6:1)

0

Symbolic relation: LT

xðtÞ ! XðsÞ:

(5:6:2)

Notes: In defining the transform integral in (5.6.1), we have used the lower limit of 0 . This allows us to include signals such as the unit impulse function dðtÞ. From now on we will use the lower limit on

5.6 One-Sided Laplace Transform

167

the integral as zero except in special cases and assume that the limit is 0 . In cases where there is some ambiguity we will explicitly identify the lower limit on the integral as 0 . Some texts use the limits of integration on the Laplace integral as 0þ to infinity. This implies that origin is excluded. This approach is impractical in the theoretical study of linear systems. The Laplace transform of a function exists if ðxðtÞes t Þ is absolutely integrable. We can select the range of s that ensures the convergence and this is referred to as the region of convergence.This is one of the nice aspects of the Laplace transform. For example, the L-transform of eat uðtÞ; a40 exists only if s4a and we can select such a range. Noting that the Laplace transform is an integral operation, the transform is unique. Example 5.6.1 Find the Laplace transforms of the functions by using the definition a: x1 ðtÞ ¼ dðt t0 Þ; t0 40; b: x2 ðtÞ ¼ uðtÞ; c: x3 ðtÞ ¼ e

a t

(5:6:3)

uðtÞ; a40:

Solution: a. The Laplace transform of the impulse function is given by Z1

X1 ðsÞ ¼

dðt t0 Þest dt ¼ es t0 ; (5:6:4)

0 LT

dðt t0 Þ ! es t0 : b: X2 ðsÞ ¼

Z

1 0

1 uðtÞest dt ¼ est 1 t¼0 s

(5:6:5)

1 LT 1 ! uðtÞ ! ; ¼ s s Z 1 Z 1 c: X3 ðsÞ ¼ eat est dt ¼ eðsþaÞt dt 0

¼

Z1

t1=2 est dt:

0

Using the change of variable a ¼ t1=2 ; da ¼ ð1=2Þt1=2 dt and the integral tables result in 4 X1 ðsÞ ¼ pﬃﬃﬃ p

Z1

2

a2 es a da ¼

1 : s3=2

0

1 b: X2 ðsÞ ¼ L½sinhðtÞ ¼ 2

Z1 h

i eðsaÞt eðsþaÞt dt

0

a : ¼ 2 s a2

(5:6:7)

The two transform pairs in this example are given by rﬃﬃﬃ t LT 1 ! 3=2 ; 2 p s

LT

sinhðatÞ !

s2

a2 : a2

(5:6:8) &

As in the case of F-transforms we can derive the properties of the L-transforms. Most of the proofs are similar in these cases. We will not go through the details but will point out some of the important facets of the properties. In cases where they are different we will go through the proofs. Following is a table of some of the one-sided Laplace transforms theorems. Note that the regions of convergence are not identified.

5.6.1 Properties of the One-Sided Laplace Transform

(5:6:6) &

Example 5.6.2 Find the unilateral Laplace transforms of the following functions: rﬃﬃﬃ t ; b:x2 ðtÞ ¼ sinhðtÞ: p

a:x1 ðtÞ ¼ 2

2 a:X1 ðsÞ ¼ pﬃﬃﬃ p

Unilateral Laplace transforms properties are given in Table 5.6.1.

0

1 1 LT ; eat uðtÞ ! : sþa sþa

Solution:

5.6.2 Comments on the Properties (or Theorems) of Laplace Transforms The proof of the linearity property of the Laplace transforms is straightforward as the integral of a

168

5 Relatives of Fourier Transforms Table 5.6.1 One-sided Laplace transform properties Superposition (linearity): xðtÞ ¼

N P

N P

LT

ai xi ðtÞ !

i¼1

ai Xi ðsÞ:

ð5:6:9Þ

i¼1

Time delay: LT

xðt tÞuðt tÞ ! est XðsÞ; t > 0:

ð5:6:10Þ

Complex frequency shift (times-exponential): LT

eat xðtÞ ! Xðs þ aÞ: Time scaling: LT 1 xðatÞ ! Xðs=aÞ; a 6¼ 0; a is a constant: jaj LT 1 xðatÞes0 t ! Xðs s0 Þ: jaj

ð5:6:11Þ

ð5:6:12Þ ð5:6:13Þ

Convolution in time: LT

xi ðtÞ xj ðtÞ ! Xi ðsÞXj ðsÞ: Multiplication in time LT 1 ½X1 ðsÞ X2 ðsÞ: x1 ðtÞx2 ðtÞ ! 2pj

ð5:6:14Þ

ð5:6:15Þ

Times-t: LT

ðtÞn xðtÞ !ð1Þn

d n XðsÞ : dsn

Times-(1/t): Z 1 xðtÞ LT XðaÞda ! t s

ð5:6:16aÞ

ð5:6:16bÞ

Derivative: d n xðtÞ LT n ! s XðsÞ sn1 xð0 Þ sn2 xð1Þ ð0 Þ ::: xn1 ð0 Þ; dtn di xðtÞ xðiÞ ð0 Þ ¼ jt¼0 : dti Integration: Z 0 Z t XðsÞ LT 1 xðaÞda ! xðaÞda þ : s 1 s 1

ð5:6:17Þ

ð5:6:18Þ

Initial Value: xð0þ Þ ¼ lim ½sXðsÞ; ðXðsÞ is properÞ: s!1

ð5:6:19Þ

Final value: lim xðtÞ ¼ lim ½sXðsÞ; ðPoles of XðsÞ lie in left half s planeÞ

t!1

s!1

ð5:6:20Þ

Switched periodic (xT ðtÞis periodic with period T): LT

xðtÞ ¼ xT ðtÞuðtÞ !

XðsÞ : 1 esT

ð5:6:21Þ

Differentiation with respect to a second independent variable: @xðt; rÞ LT @Xðs; rÞ ! ; ðr is independent of s and tÞ: @r @r Integration with respect to a second independent variable: Z r Z r LT Xðs; rÞdr ! xðt; bÞdb; ðr is independent of s and tÞ: r0

r0

ð5:6:22Þ

ð5:6:23Þ

5.6 One-Sided Laplace Transform

169

2

sum is equal to the sum of the integrals. The time delay property can be shown by using a change of variable in the transform integral. The complex shift property can be shown by the following: Z1

s0 t

½e xðtÞe

s t

xðtÞe

0

ðsþaÞt

Convolution property: The convolution of two functions was defined in Chapter 2. Assuming the functions start at t ¼ 0, we can write

¼

0

Z1

3 x2 ðxÞesx dx5 ¼ X1 ðsÞX2 ðsÞ:

(5:6:25)

Complex frequency shift: This is also referred to as times-exponential property and Lfe

at

xðtÞg ¼

Z1

eat xðtÞest dt

0

¼

Z1

xðtÞeðsþaÞt dt ¼ Xðs þ aÞ: (6:5:26)

0

x1 ðaÞx2 ðt aÞda Times-t property: This follows from the following equation:

0

Z1

x1 ðaÞesa da5

This proves the convolution theorem.

The time-scaling and frequency-shifting properties follow by combining the time-scaling and the complex frequency-shifting properties.

x1 ðtÞ x2 ðtÞ ¼

3

0

dt ¼ Xðs þ aÞ:

0

Z1

2 4

Z1

dt ¼

L½x1 ðtÞ x2 ðtÞ ¼ 4

Z1

x2 ðbÞx1 ðt bÞdb:

(5:6:24) dXðsÞ ¼ ds

0

Z1

dest dt ¼ xðtÞ ds

0

Z1

½txðtÞest dt: (5:6:27)

0

The transform of this function is given by

L½x1 ðtÞ x2 ðtÞ ¼

Z1 0

¼

Z1

2 1 3 Z 4 x1 ðaÞx2 ðt aÞda5est dt 0

2 1 3 Z x1 ðaÞ4 x2 ðt aÞest dt5da:

0

L½x1 ðtÞ x2 ðtÞ ¼

0

2 x1 ðaÞ4

Z1

Example 5.6.3 In Example 5.6.2 we have derived the Laplace transform of the hyperbolic sine function. Use the times-t property to show that the following is true: 2as

LT

0

yðtÞ ¼ t sinhðatÞ !

Using the change of variable x ¼ t a and simplifying the integral, we have Z1

We can generalize this result by repeated derivatives of the transform.

3 x2 ðxÞesðxþaÞ dx5da:

a2 Þ 2

¼ YðsÞ:

(5:6:28)

Solution: Taking the derivative of the transform function in (5.6.8), we have Z1 dXðsÞ d 2as ½ðtÞ sinhðatÞest dt ¼ ¼ ds ds ðs2 a2 Þ 0

a

¼ We are considering only positive time functions, so x2 ðxÞ ¼ 0 for x50, which allows us to change the lower limit on the second integral in the above equation and

ðs2

2as ðs2

a2 Þ 2

:

The result in (5.6.28) follows by identifying the term in the integrand and its transform on the right in the & above equation.

170

5 Relatives of Fourier Transforms

Times-1/t property or complex integration property: This is Z1

xðtÞ provided XðbÞdb ¼ L t

(5:6:29)

s

parts. The convolution and the derivative properties are the most used properties in linear systems theory. Integration property: This property can be seen using the integration by parts.

lim½xðtÞ=t exists: t!0

L This can be shown by Za

2 3 Za Z1 XðbÞdb ¼ 4 ebt xðtÞdt5db

s

s

¼

¼

¼

0

Z1 0 Z1 0 Z1

2

xðtÞ4

8 t 0:

teat uðtÞ; a > 0: eat cosðbtÞuðtÞ; a > 0: eat sinðbtÞuðtÞ; a > 0: uðtÞ cosðbtÞuðtÞ sinðbtÞuðtÞ

Fourier Transform

1 1 ðs þ aÞ

1 1 ðjo þ aÞ

1

1

ðs þ aÞ2

ðjo þ aÞ2

sþa 2

jo þ a b2

ðjo þ aÞ2 þ b2

b

b

ðs þ aÞ2 þ b2

ðjo þ aÞ2 þ b2

ðs þ aÞ þ

1 s s s2 þ o 2 b s2 þ o 2

computation of transforms and their inverses. That is, one algorithm can be used to compute both the forward and inverse Fourier transforms with few modifications. There is no symmetry property of Laplace transforms. 5. Noting that XðsÞ is a function of s ¼ s þ jo, a complex quantity, it can only be plotted as a surface plot. On the other hand, the Fourier transform is a function of jo and therefore it is the cross section of the surface plot along the jo-axis. 6. Noting Item 5 above, the circuits and systems literature use the Laplace transform to compute the frequency characteristics of the function by simply substituting s ¼ jo in the Laplace transform. The complex function is generally written in terms of the magnitude and phase frequency characteristics of the signal.

1 þ pdðoÞ jo jo þ p½dðo þ bÞ þ dðo bÞ b2 o2 b þ jp½dðo þ bÞ dðo bÞ b2 o2

Example 5.9.1 Find the Hartley transform of xðtÞ using its Laplace transform. 1 1 ! ðs þ aÞ jo þ a 1 1 þ j Im : ¼ Re jo þ a jo þ a LT

xðtÞ ¼ eat uðtÞ !

Solution: The Hartley transform of xðtÞ can be obtained from 1 1 aþo Hart : Im ¼ xðtÞ !Re jo þ a jo þ a o2 þ a2 (5:9:7) & Example 5.9.2 Find the Hartley transform of the function uðtÞ from its Laplace transform. LT

xðtÞ ¼ uðtÞ ! XðsÞ ) Xð joÞ ¼ pdðoÞ ð j=oÞ:

5.9.2 Hartley Transforms and Laplace Transforms Hartley transform is the symmetrical form of the Fourier transform. It can be derived from the one-sided Laplace transforms. A few examples are given below for the Laplace transform functions XðsÞ with poles in the left half-plane only and later with poles in the left half-plane and with poles on the imaginary axis.

(5:9:8) Solution: The Hartley transform of the function is XH ðoÞ ¼ pd½o þ ð1=oÞ:

(5:9:9) &

See the chapter by Olejniczak in Poularikis, ed. (1996) for an extensive discussion on the relationship between the Laplace and Hartley transforms.

186

5 Relatives of Fourier Transforms

5.10 Hilbert Transform

yðtÞ ¼ hðtÞ xðtÞ; hðtÞ ¼ F 1 ½Hð joÞ:

Another transform that is closely related to the Fourier transform is the Hilbert transform. It is used in the theoretical descriptions and implementations of analog and digital Hilbert transformers. A device called the Hilbert transformer is basic and has important applications in single sideband modulation of signals and in digital signal processing. Hilbert transforms can be introduced with Euler’s formula e jot ¼ cosðotÞ þ j sinðotÞ. We will see shortly that the Hilbert transform of cosðotÞ is sinðotÞ. Hilbert transforms became an important area with analytic signals that are complex valued with one-sided spectrum. These have the form xa ðtÞ ¼ xðtÞ þ j^ xðtÞ, where x^ðtÞ is the Hilbert transform of xðtÞ. Analytic signals are considered in Section 5.10.3. Also, the real and imaginary parts of transfer functions of systems are tied together by Hilbert transforms.

(5:10:3)

From Chapter 4, F ½sgnðtÞ ¼ 2=ð joÞ. See (4.4.36). Using the symmetry or the duality property of the Fourier transforms, it follows that hðtÞ ¼

1 : pt

(5:10:4)

Using the convolution integral and (5.10.4) results in Z 1 D yðtÞ ¼ hðtÞ xðtÞ ¼ xðaÞhðt aÞda ¼ x^ðtÞ: 1

(5:10:5) x^ðtÞ is the Hilbert transform of the function xðtÞ. Note the hat in x^ðtÞ. Hilbert transform is a convolution operation and is a function of time. This is symbolically represented by HT

HT

xðtÞ ! x^ðtÞ ¼ H½xðtÞ ¼ yðtÞ ! YðjoÞ ¼ HðjoÞXðjoÞ; HðjoÞ ¼ j sgnðoÞ: (5:10:6)

5.10.1 Basic Definitions There are two ways of introducing the Hilbert transforms. One is by using an integral and the other by using the Fourier transform of the function. It is simpler to view it starting with the transform and derive the integral that defines the Hilbert transform. To start with assume xðtÞ is the input and yðtÞ is the output and F½xðtÞ ¼ Xð joÞ and F½yðtÞ ¼ Yð joÞ. The output transform Yð joÞ is assumed to be related to the input transform by

Hilbert transforms can be computed directly by the convolution in (5.10.5) or by using the transforms in (5.10.6). If x^ðtÞ is known, how do we compute xðtÞ from x^ðtÞ, if x^ðtÞ is not identically zero? It turns out that the Hilbert transform of xðtÞ is equal to ½xðtÞ. This can be seen from ^^ðtÞ ¼ ½xðtÞ hðtÞ hðtÞ FT x ! HðjoÞHðjoÞXðjoÞ ¼ ðj sgnðoÞÞ2 XðjoÞ ¼ XðjoÞ:

YðjoÞ ¼ HðjoÞXðjoÞ:

The function Hð joÞ is called the Hilbert transformer and is defined by ( Hð joÞ ¼ j sgnðoÞ ¼

It implies that if^ xðtÞ 6¼ 0 then we have the inversion formula ^^ðtÞ ¼ XðjoÞ ! x ^^ðtÞ F½x

e jp=2 ;

o 40

e jp=2 ;

o 50

(5:10:7)

(5:10:1)

¼ F1 ½XðjoÞ ¼ xðtÞ:

;

(5:10:8)

(5:10:2)

FT

hðtÞ ! Hð joÞ: Noting that multiplication in the frequency domain corresponds to the time-domain convolution, we have

Example 5.10.1 Find their Hilbert transforms of the following functions: a: x1 ðtÞ ¼ ejo0 t ; a0 40;

b: xðtÞ ¼ cosðo0 tÞ;

c: yðtÞ ¼ sinðo0 tÞ; d:x4 ðtÞ ¼ A; a constant

5.10 Hilbert Transform

187

Solution: a. The Fourier transforms of these functions are given by

This can be simplified and the corresponding Hilbert transform pair is

F½ejo0 t ¼ 2p dðo o0 Þ:

cosðo0 tÞ ! sinðo0 tÞ:

HT

(5:10:9)

Figure 5.10.1a,b gives the Fourier transforms of the functions in (5.10.9). Figure 5.10.1c gives HðjoÞ ¼ j sgnðoÞ. It follows that

Hilbert transform operation is an integral operation and the Hilbert transform of a sum is equal to the sum of the Hilbert transforms. c. We can repeat the above process and show that

½HðjoÞð2pdðo þ o0 ÞÞ

HT

¼ j 2psgnðoÞdðo þ o0 Þ ¼ j2pdðo þ o0 Þ:

sinðo0 tÞ ! sinðo0 t ðp=2ÞÞ ¼ cosðo0 tÞ:

(5:10:10)

¼ jejo0 t ¼ ejðo0 tðp=2ÞÞ :

HT

HT

xðtÞ ! x^ðtÞ; x^ðtÞ ! xðtÞ; x^ðtÞ 6¼ 0:

(5:10:12)

b. The Hilbert transform of the cosine function can be obtained by using Euler’s formula 1 jo0 t 1 jo0 t þ e cosðo0 tÞ ¼ e 2 2 1 HT 1 jðo0 tðp=2ÞÞ ! e þ ejðo0 tðp=2ÞÞ 2 2 ¼ cosðo0 t ðp=2ÞÞ:

From the above we note that the Hilbert transform of a sine or a cosine function can be obtained by adding a phase shift of ðp=2Þ. d. In the case of a constant, the transform is an impulse function. Note that sgnðoÞjo¼0 ¼ 0. That is, HðjoÞjo¼0 ¼ 0 and YðjoÞjo¼0 ¼ 0 in (5.10.1). Since the Fourier transform of a constant is an impulse function, it follows that the Hilbert trans& form of a constant is zero. We can generalize the above results and state that any periodic function with zero average value can be written in terms of Fourier cosine and sine series and therefore the Hilbert transform of such a periodic function xT ðtÞ is

p HT xT ðtÞ ! xT t : (5:10:16) 2

(a)

Fig. 5.10.1 (a)F ½e jo0 t , (b) F [e jo0 t ], (c) HðjoÞ ¼ jsgnðoÞ

(5:10:15)

(5:10:11)

Changing the sign in the exponent is a minor matter and we have h i HT ½ejoo t ! ejðo0 tðp=2ÞÞ

(5:10:14) Note also

From (5.10.10), x^ðtÞ ¼ F1 ½j2pdðo þ o0 Þ

(5:10:13)

(b)

(c)

188

5 Relatives of Fourier Transforms

In the next example we will make use of the integral to compute the Hilbert transform. Example 5.10.2 Find the Hilbert transform of the following functions: a: xðtÞ ¼ P½t=t; b: x1 ðtÞ ¼ xðt t=2Þ; c: yðtÞ ¼ A: (5:10:17) Solution: a. Using the integral expression, the Hilbert transform is given by Z

1

xðaÞ 1 da ¼ x^ðtÞ ¼ pðt aÞ p 1

Z

t=2

da ðt aÞ t=2

1 1 t þ t=2 ¼ ½lnðt t=2Þ lnðt þ t=2Þ ¼ ln p p t t=2

P

(5:10:19)

c. Hilbert transform of a constant is zero. This can be seen from (5.6.18) at the limit t ! 1, as & lnð1Þ ¼ 0, which verifies our earlier result. Example 5.10.3 Show that a. the energy (or the power) in an energy signal (or a power signal) xðtÞ and its Hilbert transform x^ðtÞ are equal and b. the signal and its Hilbert transform are orthogonal. Solution: The results are shown for energy signals and the results for power signals are left as exercises. a. The energies in the two functions are given by Ex^ ¼

(5:10:18) The pulse and its Hilbert transform are shown in Fig. 5.10.2. The pulse we started with is time limited, whereas the time-width of the Hilbert transform pulse is infinite. b. Hilbert transform of the delayed pulse can be obtained by changing the variable of integration in (5.10.18) by b ¼ t t=2 and following the above procedure results in

t ðt=2Þ HT 1 t ! ln : t p tt

¼

¼

1 2p 1 2p 1 2p

Z1 1 Z1 1 Z1

xðtÞj2 do jF½^

jjsgnðoÞj2 jXðjoÞj2 do

jXðjoÞj2 do ¼ Ex :

(5:10:20)

1

b. The two functions are orthogonal by the generalized Parseval’s theorem. Z1 Z1 1 xðtÞ^ xðtÞdt¼ XðjoÞ½F½^ xðtÞ dt 2p 1

¼

1 2p

1 Z1

jsgnðoÞjXðjoÞj2 do:

1

(5:10:21) &

(a)

5.10.2 Hilbert Transform of Signals with Non-overlapping Spectra In Chapter 10 single-sided modulation schemes will be studied, where Hilbert transforms play an important role. Consider the signals xðtÞ and gðtÞ with their spectra defined by

(b) Fig. 5.10.2 (a) Pulse function and (b) Hilbert transform of the pulse

FT

xðtÞ ! XðjoÞ; jXðjoÞj ¼ 0; joj4W FT

gðtÞ ! GðjoÞ;

GðjoÞ ¼ 0; joj5W:

(5:10:22)

5.10 Hilbert Transform

189

That is, xðtÞ is a low-pass signal and gðtÞ is a highpass signal. Such is the case in the single sideband modulation. We state that H½xðtÞgðtÞ ¼ xðtÞH½gðtÞ ¼ xðtÞ^ gðtÞ:

(5:10:23)

This can be proven by using the following steps: 1. Write the time function xðtÞgðtÞ using the transform convolution integral in the form and then using the Hilbert transform, we have Z 1 1 Xðo aÞGðaÞda; (5:10:24) xðtÞgðtÞ ¼ 2p 1 FT

H½xðtÞgðtÞ ! jsgnðoÞ Z 1 1 Xðo aÞGðaÞda : 2p 1

The spectrum of the analytic signal xa ðtÞ is the positive portion of the spectrum of the real signal xðtÞ. This property will be useful in the development of single sideband modulation scheme in Chapter 10. Some authors use the symbol xþ ðtÞð¼ xa ðtÞÞ for analytic signals. A real signal xðtÞ can be written in terms of analytic signals HT

xðtÞ ¼ ½xa ðtÞ þ x a ðtÞ=2 !½xa ðtÞ x a ðtÞ=2: (5:10:29)

Example 5.10.4 Let mðtÞ be a low-pass signal with MðoÞ ¼ 0; joj4W and oc 4W, then

H½mðtÞ sinðoc tÞ ¼ mðtÞ cosðoc tÞ:

1 1 xðtÞ ¼ ½1 þ sgnðoÞXðjoÞ F½xa ðtÞ ¼ F½xðtÞ þ j^ 2 2 ( (5:10:28) XðjoÞ; o40 ¼ : 0; o 50

(5:10:25)

2. Write the time-domain product xðtÞ^ gðtÞ in terms of the Fourier convolution integral. 3. Noting the non-overlapping spectra of the two functions, we can relate the two results and then show (5.10.23). The details are left as an exercise.

H½mðtÞ cosðoc tÞ ¼ mðtÞ sinðoc tÞ;

Some authors do not use the constant (1/2) in the definition of the analytic signal. The real significance of the analytic signal is its spectrum and is

(5:10:26)

Solution: These follow from (5.10.23) and the Hilbert transforms of the sine and cosine functions. We will use results in studying single sideband mod& ulations in Chapter 10.

Use of this gives cosðo0 tÞ ¼ ½ejo0 t þ ejo0 t =2; sinðo0 tÞ ¼ ½ejo0 t ejo0 t =2j:

(5:10:30)

Narrowband noise signals: Although statistical description of noise is beyond our scope here, we will study signals with their spectra centered at a frequency fc with a bandwidth B fc . For example, the output of an amplitude modulated signal is xc ðtÞ ¼ mðtÞ cosðoc tÞ with mðtÞ being a low-pass signal with its bandwidth much and much smaller than the carrier frequency fc . Such signals are narrowband (NB) signals. These are expressed in terms of the envelopeRðtÞ, a slowly varying function and the phase fðtÞ written in the form nðtÞ ¼ RðtÞ cosðoc t þ fðtÞÞ; RðtÞ 0:

(5:10:31)

5.10.3 Analytic Signals

In most cases, RðtÞ and fðtÞ are not transformable.

The Hilbert transform is used to define an analytic signal of the real signal xðtÞ by

Example 5.10.5 Find the envelope and the complex envelope of the NB signal in terms of two NB signals nc ðtÞ and ns ðtÞ given by

1 xðtÞ: xa ðtÞ ¼ ½xðtÞ þ j^ 2

(5:10:27)

nðtÞ ¼ nc ðtÞ cosðo0 tÞ ns ðtÞ sinðo0 tÞ:

(5:10:32)

190

5 Relatives of Fourier Transforms Table 5.10.1 Hilbert transform pairs Sinusoids: HT

HT

sinðo0 tÞ ! cosðo0 tÞ; cosðo0 tÞ ! sinðo0 tÞ Exponential: HT

ejot ! jsgnðoÞejot Rectangular pulse: P

hti t

1 t þ t=t22 ! ln p t t=t22

HT

Impulse: HT

dðtÞ ! 1=pt

Solution: First, qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ nðtÞ ¼ RðtÞ cosðo0 t þ fðtÞÞ; RðtÞ ¼ n2c ðtÞ þ n2s ðtÞ; fðtÞ ¼ tan1 ½ns ðtÞ=nc ðtÞ:

(5:10:33)

Hilbert transforms. Bulk of this chapter deals with the one-sided Laplace transforms. Specific topics are:

Fourier cosine and sine and Hartley transforms Laplace transforms and their inverses; regions of convergence

From this representation, the analytic signal can be obtained and is

Basic properties of Laplace transforms; initial and final value theorems

Partial fraction expansions na ðtÞ ¼ nðtÞ þ j^ nðtÞ ¼ nc ðtÞ cosðo0 tÞ ns ðtÞ sinðo0 tÞ Solutions of constant coefficient differential þ jnc ðtÞ sinðo0 tÞ þ jns ðtÞ cosðo0 tÞ ¼ nc ðtÞ½cosðo0 tÞ þ j sinðo0 tÞ þ jns ðtÞ½cosðo0 tÞ þ j sinðo0 tÞ ¼ ½nc ðtÞ þ jns ðtÞejo0 t : The envelope and the complex envelopes are, respectively, given by jna ðtÞj ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ n2c ðtÞ þ n2s ðtÞ;

(5:10:34) &

equations using Laplace transforms

Relationship between Laplace and Fourier transforms

Hilbert transforms and their inverses Various tables listing some simple time functions and their transforms

Problems

n~ðtÞ ¼ nc ðtÞ þ jns ðtÞ: Ziemer and Tranter (2002) give interesting applications of the Hilbert transforms. Refer table 5.10.1 for a short table of Hilbert transforms, see Hahn Poularikis, ed. (1996).

5.2.1 Derive the following properties using Xc ðoÞ ¼ FCT½xðtÞ: FCT

a: Xc ðtÞ !ðp=2ÞxðoÞ; oþb oþb FCT 1 Xc þ Xc ; b: xðatÞ cosðbtÞ ! 2a a a a40; b40; d2 Xc ðoÞ ; do2

5.11 Summary

c: t2 xðtÞ !

In this chapter Fourier transform integral is used to discuss and derive some of the related transforms. These include cosine, sine, Hartley, Laplace, and

d: x00 ðtÞ ! o2 Xc ðoÞ x0 ð0þ Þ:

FCT

FCT

(Assume xðtÞ and x0 ðtÞ vanish as t ! 1.).

Problems

191

5.2.2 Show

5.3.4 Show the following transforms are true.

1 FCT p ao ! e ; 2 þt 2a 1 FCT p ! sinðaoÞ; b: 2 a t2 2a b b FCT þ ! p sinðaoÞebo ; c: ðt aÞ2 þ b2 ðt þ aÞ2 þ b2 d 2 ½eat uðtÞ FCT a3 d: ! 2 : a þ o2 dt2 a: xðtÞ ¼

a2

5.2.3 Derive the following associated with sine transforms using Xs ðoÞ ¼ FST½xðtÞ: 1 oþb ob a: xðatÞ cosðbtÞ ! Xs þ Xs ; 2a a a

Hart

a: eat sinðo0 tÞuðtÞ !

o0 ða2 þ o20 o2 Þ þ 2ðaoÞ ða2 þ o20 o2 Þ þ 2ðaoÞ2

;

Hart

b: eat cosðo0 tÞuðtÞ ! ða oÞða2 þ o20 o2 Þ þ 2oða þ oÞ ða2 þ o20 o2 Þ þ 2ðaoÞ2 c:

1 X

Hart

dðt nTÞ !

n¼1

;

1 p X dðo ðk=TÞÞ; T k¼1

Hart

d: ejo0 t ! pdðo o0 Þ:

FST

FST 1 b: xðatÞ ! Xs ðo=aÞ; a40 a

5.2.4 Show the following are valid: rﬃﬃﬃﬃﬃﬃ 1 FST p ; a: pﬃﬃ ! 2o t b b FST b: 2 2 ! p cosðaoÞebo ; 2 2 b þ ðt aÞ b þ ðt aÞ b40: 5.3.1 The energy spectral density of a signal FT

xðtÞ ! XðjoÞ can be expressed by n ð1=2pÞjXðjoÞj2 ¼ ð1=2pÞ ½ReðXðjoÞÞ2 o þ½ImðXðjoÞÞ2 : Derive this in terms of the Hartley transform XH ðoÞ. Also derive the expression for the phase angle of the spectrum in terms of the Hartley transform.

5.4.1 Find the two-sided Laplace transforms and their ROCs of the following functions: a: x1 ðtÞ ¼ P½t; b: x2 ðtÞ ¼ teajtj ; a40: The following problems are concerned with onesided Laplace transforms. 5.4.2 Find the Laplace transforms of the following functions: pﬃﬃﬃﬃﬃﬃﬃ a: x1 ðtÞ ¼ 2 t=p: b: x2 ðtÞ ¼ L½t; c: x3 ðtÞ ¼ coshðbtÞ: 5.4.3 Find the Laplace transform of the function xðtÞ ¼ tuðtÞ by the following methods: a: Use the times-tproperty and the transform of uðtÞ to show that L½t uðtÞ ¼ 1=s2 . b. Use the result in part a to show by induction LT

that tn uðtÞ ! n!=ðsnþ1 Þ . The proof by induction uses the following procedure. The result is first shown to be true for the case of n ¼ 1. Then verify that if the result is true for n then it is also true for n þ 1. 5.6.1 Find the L-transform of xðtÞ ¼ jsinðtÞj; t 0.

5.3.2 Derive the Hartley transforms of the following functions using Fourier transforms:

c: x3 ðtÞ ¼ xðtÞ cosðo0 tÞ:

5.6.2 a. Find X1(s) ¼ L{x(t)} by using the transform of the unit step function. 1 t T=2 LT ! X1 ðsÞ: x1 ðtÞ ¼ P T T

5.3.3 Derive an expression for the Hartley transform of the convolution yðtÞ ¼ xðtÞ hðtÞ.

b. Take the limit of the transform as T ! 0 and identify the corresponding transform pair.

a: x1 ðtÞ ¼ xðatÞ; a 6¼ 0; b: x2 ðtÞ ¼ xðtÞ cosðo0 tÞ;

192

5 Relatives of Fourier Transforms

5.6.3 Show the following transform pair is true by a.using the integral of the transform and by b. using the second derivative of the Laplace transform to show the above result.

d 2 xðtÞ xðtÞ ¼ sinhðtÞ; xð0Þ ¼ 0 and x0 ð0Þ ¼ 0: dt2 5.9.1 Find

xðtÞ ¼ L½t 1 8 9 > > < t; 05t51 = LT 1 ¼ ð2 tÞ; 15t52 ! 2 ð1 2es þ e2s Þ: > > s : ; 0; otherwise 5.6.4 a. Use the differentiation with respect to the second variable property of the Laplace transforms to show that L½teat ¼ ½1=ðs þ aÞ2 .

xi ðtÞ ¼ L1 ½Xi ðsÞ with 1 es 1 ; b: X2 ðsÞ ¼ 2 2 ; sðs þ 1Þ s ðs þ 4Þ 1 es c: X3 ðsÞ ¼ : sð1 e2s Þ

a: X1 ðsÞ ¼

5.9.2 Show the residues A i can be determined by

5.6.5 Determine the transform of the Laplace transform of the following function xðtÞ: sinðo0 tÞ ; xðtÞ ¼ t

Zo 0

cosðotÞdo ¼

NðsÞ NðsÞ ¼ DðsÞ ðs þ s1 Þðs þ s2 Þ . . . ðs þ sN Þ N X Ai NðsÞ ¼ ; Ai ¼ js¼si : ðs þ si Þ dDðsÞ=ds i¼1

XðsÞ ¼

sinðo0 tÞ : t

0

5.7.1 Verify the following the transform pairs by a. evaluating the transform directly and by b. using the partial fraction expansion and then identify term by term from tables. 1 1 LT : ð1 cosðatÞÞ ! 2 2 a sðs þ a2 Þ 5.7.2 Verify the following by using the Laplace transforms properties: LT

a: x1 ðtÞ ¼ ð1=aÞebt sinhðatÞ ! 1=½ðs bÞ2 a2

a. Use this result to find L1 fXðsÞg ¼ L f1=½sðs þ 3Þg. b. Now consider YðsÞ ¼ ð1=sÞXðsÞ. Use Part a. to generalize this. 1

5.9.3 Assuming the regions of convergence are a: s41; b: s5 2, find 2s þ 1 1 1 : L fXII ðsÞg ¼ L ðs2 þ s 2Þ 5.9.4 Assuming the following Laplace transforms, find the corresponding Fourier transforms:

¼ X1 ðsÞ;

a: XII ðsÞ ¼

h i LT b: x2 ðtÞ ¼ ð1=2aÞt sinðatÞ ! s= ðs2 þ a2 Þ2 ¼ X2 ðsÞ: c: x3 ðtÞ ¼ t=T; 05t5T; x3 ðt þ nTÞ eas LT 1 : ¼ xðtÞ; n 0 ! 2 sð1 eas Þ as 5.8.1 Find the solution of dy þ 3y ¼ t cosðtÞ; yð0 Þ ¼ 1: dt 5.8.2 Using Laplace transforms to find the solution of the differential equation

¼

1 2

ðs þ aÞ

1 ðs aÞ2

; jsj5a; b: XðsÞ

s þ o0 : s2 þ o20

5.10.1 Show 1 HT t ; ! 2 1þt ð1 þ t2 Þ HT 1 b: dðtÞ ! ; pt K X c: Xs ½0 þ Xs ½k cosðko0 t a:

k¼1 HT

þ y½kÞ !

K X k¼1

Xs ½k sinðko0 t þ y½kÞ:

Chapter 6

Systems and Circuits

6.1 Introduction In this chapter we will consider systems in general, and in particular linear systems. Most systems are inherently nonlinear and time varying. A human being is a good example. He can run fast for a while and then speed comes down. If you plot speed versus time, the plot is not going to be a straight line, i.e., the function speed versus time is not linear. Humans are nonlinear and also time-varying systems. For example, if you want to ask your dad for a new car, you do not ask him when he is not happy. Moods change with time. These considerations are important in, for example, speaker identification. Human beings are not only nonlinear but also time-varying complicated systems. Nonlinear time-varying systems are very hard to deal with. Even though many of the systems may have nonlinear behavior characteristics, they can be approximated to be linear systems and they allow for transform analysis. In addition we are interested in systems that operate in the same manner every time we use them. That is, the systems must be independent of time. Linear time-invariant system analysis and design is the basis of present day system analysis and design. Transfer functions associated with these systems are discussed. In addition the frequency analysis makes it very attractive for the design of systems. Majority of the discussion in this chapter is on linear timeinvariant systems. These allow for transfer function analysis. The study of the amplitude and phase frequency responses of linear time-invariant systems is one of the important topics. When a signal through some media, it is modified by the media. In the frequency domain we can say that some

frequencies are amplified and some are attenuated. In addition different frequencies are delayed differently. Our goal is to filter frequencies with appropriate attenuations of the input frequencies with a constant delay at all frequencies in the frequency band of interest. The delay response is related to the phase response of a system. If the delay is not constant, then delay compensation may be required. One of the topics we will be interested is filter circuits. Toward this goal ideal low-pass, high-pass, band-pass, and band-elimination filter functions are introduced in this chapter. In addition to these simple examples of a differentiator, integrator, and a delay circuit are illustrated. A brief introduction to nonlinear systems is included later. The topic of linear systems is one of the topics every undergraduate student in electrical engineering program goes through. See the books Haykin and Van Veen (1999), Lathi (1998), Oppenheim et al. (1997), Nillsson and Riedel (1996), Poularikas and Seely (1991), Carlson (2000) and others.

6.2 Linear Systems, an Introduction Our study starts with a system that has an input and an output. It is symbolically represented by a block diagram shown in Fig. 6.2.1. The T inside the box is some transformation that converts the input signal xðtÞ into the output signal yðtÞ and yðtÞ ¼ T½xðtÞðT maps xðtÞ into yðtÞÞ:

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_6, Ó Springer ScienceþBusiness Media, LLC 2010

(6:2:1)

193

194

6 Systems and Circuits

below. What can you say about the linearity of theses schemes? Fig. 6.2.1 Block diagram of a system

a: y1 ðtÞ ¼ mðtÞ cosðoc tÞ;

Some use L½xðtÞ to represent a linear system. The system described by (6.2.1) is called a linear system if the transformation given by T½xðtÞ satisfies the following conditions: Principle of additivity: If T½x1 ðtÞ ¼ y1 ðtÞ and T½x2 ðtÞ ¼ y2 ðtÞ; then the superposition property states that

b: y2 ðtÞ ¼ ðA þ mðtÞÞ cosðoc tÞ; A 6¼ 0:

T½x1 ðtÞ þ x2 ðtÞ ¼ T½x1 ðtÞ þ T½x2 ðtÞ ¼ y1 ðtÞ þ y2 ðtÞ (6:2:2) Principle of proportionality: If T½xðtÞ ¼ yðtÞ; then for any constant a

(6:2:5)

Solution: a. For the inputs m1 ðtÞ and m2 ðtÞ, the outputs are, respectively, given by m1 ðtÞ cosðoc tÞ and m2 ðtÞ cosðoc tÞ. For the input a1 m1 ðtÞ þ a2 m2 ðtÞ, the output is the sum of the two individual outputs ½a1 m1 ðtÞ þ a2 m2 ðtÞ cosðoc tÞ. Therefore the system is linear. b. For the inputs x1 ðtÞ and x2 ðtÞ, the outputs are, respectively, given by ðA þ m1 ðtÞÞ cosðoc tÞ and ðA þ m2 ðtÞÞ cosðoc tÞ. For the input ða1 m1 ðtÞþ a2 m2 ðtÞÞ, the output is given by ½A þ a1 m1 ðtÞþ a2 m2 ðtÞ cosðoc tÞ which is not equal to ðA þ m1 ðtÞÞ cosðoc tÞ þ ðA þ m2 ðtÞÞ cosðoc tÞ:

T½axðtÞ ¼ aT½xðtÞ ¼ ayðtÞ:

(6:2:3)

This property is also referred to as the homogeneity property. We can combine the two properties into one and state that a linear system satisfies the following property T½a1 xðtÞ þ a2 x2 ðtÞ ¼ a1 y1 ðtÞ þ a2 y2 ðtÞ

(6:2:4)

Therefore the system is nonlinear. Note that the Fourier transform is a linear operation as F½a1 x1 ðtÞ þ a2 x2 ðtÞ ¼ a1 F½x1 ðtÞ þ a2 F½x2 ðtÞ: (6:2:6) & The systems described by the following equations are linear systems and the reader is encouraged to go through the proofs:

for any pair of constants a1 and a2 . Otherwise, the system is called a nonlinear system.

a: yðtÞ ¼ kxðtÞ

Example 6.2.1 Show that a system described by yðtÞ ¼ x2 ðtÞ is a nonlinear system.

b: yðtÞ ¼ k

Solution: For two inputs x1 ðtÞ and x2 ðtÞ, the corresponding outputs are, respectively, given by y1 ðtÞ ¼ x21 ðtÞ and y2 ðtÞ ¼ x22 ðtÞ. If the input is x1 ðtÞ þ x2 ðtÞ, then the output is ðx1 ðtÞ þ x2 ðtÞÞ2 6¼ x21 ðtÞ þ x22 ðtÞ and therefore the system is nonlinear. We can make a general statement that if the output of a system is a power of the input, and the power is not equal to one, then the system is nonlinear. Other examples of nonlinear systems include y1 ðtÞ ¼ logðxðtÞÞ; y2 ðtÞ ¼ jxðtÞj:

c: yðtÞ ¼

dxðtÞ dt

Z xðbÞdb

d: yðtÞ ¼ xðt tÞ; t 0

ðAmplifierÞ

(6:2:7a)

ðDifferentiatorÞ (6:2:7b)

ðIntegratorÞ

(6:2:7c)

ðDelay deviceÞ: (6:2:7d)

6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws

&

Example 6.2.2 Two of the many modulation schemes that we will be interested in are given

In this section we will consider two-terminal passive and active components and laws that pertain to the interconnection of elements. These are two powerful

6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws

laws, referred to as Kirchhoff’s voltage and current laws first formulated by Kirchhoff (pronounced as kear-koff) in 1847. One gives equations in terms of voltages across components and the other gives equations in terms of currents flowing through the components. Component equations and the Kirchhoff’s laws provide us with circuit analysis tools.

6.3.1 Two-Terminal Component Equations Simple circuits include ideal sources, voltage, and current sources and three types of components resistors, inductors, and capacitors. The symbols for the sources are shown in Fig. 6.3.1. An ideal voltage source is a two-terminal component whose voltage across the two terminals is a constant or a function of time regardless of what the current through the component is. Examples of voltage sources are wall outlets, where we assume that the voltage is vs ðtÞ ¼ Vm cosðom tÞ, and batteries, where the voltage across is a constant. The first source we refer to as an alternating current (AC) source and the second one is a constant voltage source (DC). The positive sign on top of the ideal voltage source indicates the higher potential whenever the source voltage is positive. Most generators are voltage sources. The ideal current source is a two-terminal component whose current is a constant or a function of time, regardless of what the voltage across it is. Transistors and many other electronic devices act more like a current source rather than a voltage source. It is important to notice the voltage signs and the direction of the currents through the sources. This convention shows that the sources provide power. The three basic passive components are the resistor, inductor, and the capacitor. The Lumped parameter models are shown in Fig. 6.3.2. These are passive elements, i.e.,

Fig. 6.3.1 Voltage, current, and a constant voltage source

195

The three basic passive components are the resistor, inductor, and the capacitor. The lumped parameter models are shown in Fig. 6.3.2. These are passive elements, i.e., they do not produce any power. Therefore the notation for the three components is that the current flows from the positive terminal to the negative terminal. The resistance is measured in Ohms, the inductor in Henries, and the capacitor in Farads. The voltage across a resistor is related to the current by the Ohm’s law and is given by vR ðtÞ ¼ RiR ðtÞ:

(6:3:1)

The voltage across an inductor is given by vL ð t Þ ¼ L

diL ðtÞ : dt

(6:3:2)

Since the voltage across an inductor is the derivative of current, it is zero for a constant current. The inductor stores energy in a magnetic field produced by current through a coil of wire. Inductor is an energy storage device and the instantaneous stored energy, measured in Joules, is 1 wL ¼ Li2L ðtÞ: 2

(6:3:3)

The current in the inductor can be computed from (6.3.2) and is 1 iL ðtÞ ¼ iL ðt0 Þ þ L

Zt

vL ðbÞdb:

(6:3:4)

t0

The term iL ðt0 Þ corresponds to the initial conditions on the current through the inductor at time t ¼ t0 . The current through a capacitor is given by

i C ðt Þ ¼ C

dvC ðtÞ dt

(6:3:5)

196

6 Systems and Circuits

Fig. 6.3.2 Resistor, inductor, and the capacitor

iR (t)

iL (t) R

L

C

vR (t)

vL (t)

vC (t)

The gap in the capacitor symbol reflects that when the voltage across the capacitor is constant, then the capacitor acts like an open circuit. Capacitor is an energy storage device and the instantaneous energy, measured in Joules, is

iC ðtÞdt:

(6:3:7)

The voltage vc ðt0 Þ corresponds to the initial voltage across the capacitor at time t ¼ t0 . Initial conditions are necessary when we consider transient analysis. In the design of systems we generally do not consider the initial conditions, as the systems are supposed to work for any initial conditions. We assume that the initial conditions on the capacitors and the inductors are assumed to be zero with t0 is equal to 1. These allow us to use both the Laplace and the Fourier transforms. This is especially true in network synthesis, as the design specifications are given in terms the sinusoidal steady state. Once we have the designs, we can always test the systems with initial conditions and any possible changes in the responses of systems associated with the initial conditions. The component equations for the three two-terminal components discussed above can be expressed in terms of Laplace transforms. To avoid any confusion from the inductor values we will make use of the symbol Lf:g for the Laplace transform. These are VR ðsÞ ¼ LfvR ðtÞg ¼ RLfiR ðtÞg ¼ RIR ðsÞ diL ðtÞ VL ðsÞ ¼ LfvL ðtÞg ¼ L L dt

(6:3:8)

¼ sLIL ðsÞ LiL ð0 Þ or IL ðsÞ

(6:3:9)

iL ð0 Þ 1 þ VL ðsÞ s sL

¼

v C ð 0 Þ 1 þ IC ðsÞ or IC ðsÞ s sC

(6:3:10)

Assuming the initial conditions are zero, we have the component equations in terms of the Laplace transformed variable s for the three components

t0

¼

dvc ðtÞ Vc ðsÞ ¼ LfvC ðtÞg ¼ L dt

(6:3:6)

The current in the capacitor is Zt

¼ sCVC ðsÞ CvC ð0 Þ:

1 wC ¼ Cv2C : 2

1 v C ð t Þ ¼ vC ð t 0 Þ þ C

iC (t)

LfvR ðtÞg ¼ VR ðSÞ; LfiR ðtÞg ¼ IR ðsÞ; LfvL ðtÞg ¼ VL ðsÞ; FfiL ðtÞg ¼ IL ðsÞ Lfvc ðtÞg ¼ Vc ðsÞ; Lfic ðtÞg ¼ Ic ðsÞ VR ðsÞ ¼ RIR ðsÞ; VL ðsÞ ¼ LsiL ðsÞ; Vc ðsÞ ¼ ð1=CsÞIC ðsÞ

(6:3:11)

The component equations in terms of the Fourier transform variable jo are VR ðjoÞ¼F½vR ðtÞ;VL ðjoÞ¼F½vL ðtÞ;Vc ðjoÞ¼F½vc ðtÞ (6:3:12a)

IR ðjoÞ ¼ F½iR ðtÞ; iL ðjoÞ ¼ F½iL ðtÞ; ic ðjoÞs ¼ F½ic ðtÞ (6:3:12b) VR ðjoÞ ¼ RIR ð joÞ; VL ð joÞ ¼ joLIL ð joÞ Vc ð joÞ ¼ ð1=joCÞIc ðjoÞ:

(6:3:12c)

In the case of zero initial conditions either Laplace or Fourier transform variables can be used. One can be obtained from the other. Also, the voltage to the current transform ratio is called as an impedance of the component under consideration and its inverse as the admittance of that component. The impedances of the three components are, respectively, given by R; Ls ðor joLÞ and 1=Cs ðor 1=joCÞ.

6.3 Ideal Two-Terminal Circuit Components and Kirchhoff’s Laws

6.3.2 Kirchhoff’s Laws Circuit analysis is based on the Kirchhoff’s current and voltage laws and the component equations. The Kirchhoff’s current law (KCL) states that the sum of the currents going into a junction, a node, is equal to the sum of the currents going out of that junction. In other words, the algebraic sum of the currents going into a node is equal to zero. The dual to the current law is the voltage law and is stated for a loop in a circuit. A loop is any path that goes from one node to another node and returns to the starting node. The Kirchhoff’s voltage law (KVL) states that the sum of the voltage drops around any loop is equal to the sum of the voltage rises. Or, the algebraic sum of voltages around a loop is equal to zero. Examples are given in Fig. 6.3.3.

Using the component equation for the capacitor and the relation in (6.3.14) and relating input and output results in iC ¼ C

LT

LT

(6:3:15b)

FT

FT

yðtÞ ! YðjoÞ

(6:3:16)

RCðsÞYðsÞ þ YðsÞ ¼ XðsÞ or

(6:3:17)

ðRCjo þ 1ÞYðjoÞ ¼ XðjoÞ: ð1=CsÞ R þ ð1=CsÞ ð1=RCÞ XðsÞ; XðsÞ ¼ s þ ð1=RCÞ 1 YðjoÞ ¼ XðjoÞ: 1 þ joRC YðsÞ ¼

(6:3:13)

Assuming that the current flowing through the output node is zero, i.e., the circuit is not loaded, we have

Fig. 6.3.3 Illustration of Kirchhoff’s current and voltage laws (a) i1 ðtÞ þ i2 ðtÞ i3 ðtÞ ¼ 0, (b) v1 ðtÞ v2 ðtÞ v3 ðtÞ ¼ 0

(6:3:15a)

xðtÞ ! XðsÞ; yðtÞ ! YðsÞ; xðtÞ ! XðjoÞ and

Solution: Using the KVL, we have the input voltage xðtÞ is equal to the sum of the voltages across the resistor and the capacitor.

xðtÞ vC ðtÞ xðtÞ yðtÞ ¼ : iR ¼ iC ; and iR ¼ R R (6:3:14)

dvC dy dy xðtÞ yðtÞ ¼C ; C ¼ : dt dt dt R dy RC þ yðtÞ ¼ xðtÞ: dt

It is a linear combination of the output and the derivative of the output related to the input. The system described by this differential equation is a linear system. In terms of the Laplace and the Fourier transforms at each step and simplifying, the output transforms can be expressed as follows:

Example 6.3.1 Consider the simple RC circuit shown in Fig. 6.3.4a. Derive the differential equation relating the input and the out put of the circuit.

xðtÞ ¼ vR ðtÞ þ vC ðtÞ:

197

(6:3:18)

&

A simple integrator and a differentiator: The circuit in Fig. 6.3.4a can be used as an integrator in the lowfrequency range. The integral form of the equation in (6.3.15b) is

(a)

(b)

198

6 Systems and Circuits

YðsÞ ¼ s=½ðs þ ð1=RCÞÞXðsÞ ¼ HðsÞXðsÞ; HðsÞ ¼ s=½ðs þ ð1=RCÞÞ: (6:3:23a)

(a)

(b)

Fig. 6.3.4 RC circuits

RCyðtÞ þ

Zt

Zt

yðaÞda ¼

1

xðaÞda:

(6:3:19)

1

If the time constant RC is large enough that the integral on the left side of the equation in (6.3.19) is dominated by RCyðtÞ, then (6.3.19) can be approximated by an integral, a smoothing operation, and is

RCyðtÞ

Zt

The above two examples provide simple circuits for low-pass and high-pass filters. The amplitude response jHðjoÞj can be approximated for small frequencies and rﬃﬃﬃﬃﬃﬃﬃ 2 1 joj2 2 RCo ; joj jHðjoÞj ¼ ð1=RCÞþo2 RC !HðjoÞjoðRCÞ: (6:3:23b) &

1 xðaÞda or yðtÞ ¼ RC

1

Zt xðaÞda: 1

(6:3:20) Example 6.3.2 Relate the input and the output transforms of the circuit in Fig. 6.3.4b. Solution: Using the Kirchhoff’s voltage law, we have dxðtÞ dt dvc ðtÞ dyðtÞ ic ðtÞ dyðtÞ þ ¼ þ ¼ dt dt c dt dxðtÞ dyðtÞ 1 ¼ þ yðtÞ: dt dt RC

xðtÞ ¼ vc ðtÞ þ yðtÞ !

(6:3:21)

If the time constant ðRCÞ is small enough that the second term dominates the first term on the right in the last equation, we can approximate (6.3.21) by yðtÞ RC

dx : dt

The networks containing ideal resistors (R’s), inductors (L’s), and capacitors (C’s) result in constant coefficient differential equations.

6.4 Time-Invariant and Time-Varying Systems For any system of use, we like the system to respond every time the same way when we switch the system on. That is, if we switch today or tomorrow, the system should respond exactly the same. Such a system is called a time-invariant system. Time-invariant system: If the response of a system is yðtÞ to the input xðtÞ; i:e:; yðtÞ ¼ T½xðtÞ, then the system is called a time invariant or a fixed system if T½xðt t0 Þ ¼ yðt t0 Þ. Otherwise, it is a time-varying system. Linear time-invariant system: A system is linear time invariant(LTI)if it is linear and time invariant. Example 6.4.1 The systems described by constant coefficient differential equations are linear time-invariant systems. RLC networks are linear time-invariant systems. Circuits containing diodes, transistors, and other electronic components are & nonlinear. Example 6.4.2 Consider the model of a carbon microphone shown in Fig. 6.4.1. The resistance R

(6:3:22)

The circuit approximates the derivative operation or it acts like a differentiator. Taking the transform of the equation in (6.3.21) and solving for YðsÞ; we have

Fig. 6.4.1 A time-varying system

6.5 Impulse Response

199

is a function of the pressure generated by sound waves on the carbon granules of the microphone, which is a function of time. The circuit has only one loop and using the Kirchhoff’s voltage law (KVL) and the component equations, we can write xðtÞ ¼ vR þ vL ¼ RðtÞiðtÞ þ L

diðtÞ : dt

(6:4:1)

The resistance is a function of time and the resulting differential equation has coefficients that vary with time and the system is a time-varying & system. Earlier we have indicated that a human being is a nonlinear time-varying system. The speech signal is a time-varying signal, albeit, a slowly time-varying signal. To analyze a slowly timevarying signal, we segment the speech signal by using windows and find the needed information for each segment. Obviously the spectral characteristics of different phonemes are different. In the earlier chapters we defined causal signals that are zero for t50. We can similarly define causal systems. Causal systems: Causal systems do not respond until the input is applied. That is, they do not anticipate the input. For a causal system, if the input

Invertibility: A system is said to be invertible if the input of the system can be recovered from the output of the system. Consider that we have the response of the system given by yðtÞ ¼ T½xðtÞ:

(6:4:4)

The system is invertible if there is a transformation T 1 such that T 1 ½yðtÞ ¼ T 1 ½T½xðtÞ ¼ xðtÞ:

(6:4:5)

That is ðT 1 TÞ ¼ I, the identity operator. Simple examples of non-invertible systems include yi ðtÞ ¼ x2 ðtÞ; y2 ðtÞ ¼ uðxðtÞÞ and many others. In each of these cases we cannot determine the function xðtÞ from yi ðtÞ. Example 6.4.3 Give an expression for the derivative of the current in an inductor. Solution: The current in an inductor is Zt 1 diL : iL ðtÞ ¼ vL ðtÞdt ) vL ðtÞ ¼ L dt L

(6:4:6)

1

Derivative operations.

and

the

integral

are

inverse &

Now consider one of the important concepts in system analysis, i.e., the signal and the system interaction.

xðtÞ ¼ 0 for all t T; then the output yðtÞ ¼ 0 for all t T

(6:4:2)

Memory and memoryless systems: A system is called memoryless if the output of the system at a particular time depends only on the input at that time. The resistor is memoryless since vðtÞ ¼ RiðtÞ and the voltage and the current pertaining to this component are related at each value of t. The capacitor voltage is related to the current by an integral and the inductor voltage is related to the voltage by an integral. Capacitors and inductors have memory and initial conditions can be assigned on these. The relations are

vC ðtÞ ¼

1 C

Zt 1

iC ðaÞda; iL ðtÞ ¼

1 L

Zt

6.5 Impulse Response Consider the block diagram of a LTI system in Fig. 6.5.1 with input xðtÞ and the corresponding output is yðtÞ. We like to find a relationship between the input and the output and the system characteristics. If xðtÞ ¼ dðtÞ; an impulse, then the output, the response of the impulse, is the impulse response of the LTI system identified by

vL ðaÞda:

1

(6:4:3)

Fig. 6.5.1 A linear time-invariant system

200

6 Systems and Circuits

yðtÞ ¼ hðtÞ ¼ T½dðtÞ:

(6:5:1)

Therefore the response for the input ai dðt ti Þ is ai hðt ti Þ, i.e., T½ai dðt ti Þ ¼ ai hðt ti Þ where a i ’s are some constants. If the input is a linear combination of impulses, then the response will be a linear combination of the corresponding impulse responses. That is, T½

N2 X

ai dðt ti Þ ¼

i¼N1

N2 X

ai hðt ti Þ:

(6:5:5) The time instant tn ¼ nDt is at some point on the time axis. As Dt ! 0; nDt approaches a continuous variable b, the sum becomes an integral and Dt becomes a differential and

(6:5:2)

i¼N1

xðtÞ ¼

We can tie this relationship to an arbitrary input using the approximation of an impulse and relate the output to the input in terms of a sum of delayed impulse responses. Impulse functions were represented in the limit by (see Section 1.4.) 1 h t ti i P : D t!0 D t Dt

dðt ti Þ ¼ lim

N2 X n¼N1

1 t nDt ½xðnDtÞDt P : Dt Dt

(6:5:4)

Note the multiplication and division by Dt in (6.5.4). The term xðnDtÞDt approximates the area of the pulse centered at t ¼ nDt. Now

Fig. 6.5.2 (a) xðtÞ(b) Pulse centered at t ¼ nDt

(a)

Z1

xðbÞdðt bÞdb:

(6:5:6)

1

This is valid provided xðtÞ is continuous for all t. It is the convolution of the two functions dðtÞ and xðtÞ. See the equation in (2.2.2a) in Chapter 2. That is, xðtÞ ¼ xðtÞ dðtÞ:

(6:5:3)

Consider an arbitrary signal xðtÞ shown in Fig. 6.5.2a. There is no specific significance for the shape of this function. Now divide the time into intervals of Dt seconds apart as shown in Fig. 6.5.2b. The strip centered at t ¼ nDt with a width of Dt can be approximated by xðnDtÞP½ðt nDtÞ=Dt. If Dt is negligibly small, the pulse can be assumed to be a rectangular pulse and the above approximation is good. The function xðtÞ can now be approximated by

xðtÞ ﬃ

t nDt lim xðnDtÞ ð1=DtÞP ¼ xðtn Þdðt tn Þ: Dt!0 Dt

(6:5:7)

In a similar manner, the output expression can be derived. Since the system is a time-invariant system, the input xðnDtÞdðt nDtÞ produces an output xðnDtÞhðt nDtÞ. Combining all the responses, we can pictorially identify N2 X

Produces

xðnDtÞ½dðt nDtÞDt

n¼N1

ﬃ

N2 X

! yðtÞ

the output

xðnDtÞhðt nDtÞDt:

(6:5:8)

n¼N1

In the limit, i.e., when Dt ! 0, nDt becomes a continuous variable b. The time interval Dt becomes a differential db and the summation becomes an integral. Noting the limits on the sum are arbitrary, the sum can be taken as over all positive and negative integers and the integral correspondingly goes from 1 to þ 1.

(b)

6.5 Impulse Response

yðtÞ ¼

201

Z1

xðbÞhðt bÞdb:

(6:5:9)

1

This integral is a superposition or a convolution integral of two functions, input and the impulse response of the linear time-invariant (LTI) system. The response of the LTI system to any input xðtÞ is yðtÞ. Symbolically, it can be written in the form

yðtÞ ¼ xðtÞ hðtÞ ¼

¼

Z1 1 Z1

xðbÞhðt bÞdb

FT

!jXðjoÞj2 ; fh ðtÞ

fx ðtÞ ¼ xðtÞ xðtÞ ¼ hðtÞ hðtÞ fy ðtÞ ¼ yðtÞ yðtÞ

(6:5:14a)

FT

!jHðjoÞj ! ¼ jYðjoÞj2 ; fy ðtÞ 2

FT

¼ hðtÞ hðtÞ fx ðtÞ:

(6:5:14b)

In a similar manner, the output power spectral density of a periodic (or a random signal) can be expressed in terms of the input power spectral density by FT

FT

Rx ðtÞ ! Sx ðoÞ; Ry ðtÞ ! Sy ðoÞ; Sy ðoÞ ¼ jHðjoÞj2 Sx ðoÞ:

hðaÞxðt aÞda ¼ hðtÞ xðtÞ:

(6:5:15)

1

(6:5:10) ) YðsÞ ¼ HðsÞXðsÞ:

(6:5:11)

The function HðsÞ is the transform of the impulse response, HðsÞ ¼ L½hðtÞ and is called the transfer function of the LTI system in Laplace transform domain or s-domain. It is symbolically represented by the block diagram in Fig. 6.5.3. In the Fourier domain, (6.5.11) is expressed by YðjoÞ ¼ HðjoÞXðjoÞ; HðjoÞ ¼ F½hðtÞ:

(6:5:12)

The input, the output, and their Fourier (and Laplace) transforms are identified on the block diagram along with impulse response hðtÞ and its transform HðjoÞ or HðsÞ. Notes: Input–output energy spectral density relations of a linear system: The output energy spectral density is

These operations are basic to the study of linear systems, as they provide a simple way of expressing how the energy (or power) of an input signal is distributed at the output. From Examples (6.3.1) and (6.3.2), the respective transfer functions and the corresponding impulse responses are given by H ðsÞ ¼

ð1=RCÞ s þ ð1=RCÞ

1 t=RC e ! RC uðtÞ ¼ hðtÞ

LT

(6:5:16a) HðjoÞ ¼

1=RC ðjo þ ð1=RCÞÞ

1 ð1=RCÞt ! RC uðtÞ e

FT

(6:5:16b) H ðsÞ ¼

s ðs þ 1=RCÞ

LT

! dðtÞ et=RC uðtÞ ¼ hðtÞ (6:5:17a)

HðjoÞ ¼

jo ðjo þ 1=RCÞ

FT

! dðtÞ et=RC uðtÞ ¼ hðtÞ:

(6:5:13)

(6:5:17b) &

Correspondingly, the output autocorrelation (AC) can be expressed in terms of the input AC as shown below.

Notes: A single input–single output of a linear time-invariant system is described by its transfer function HðsÞ ¼ YðsÞ=XðsÞ in the Laplace domain

jYðjoÞj2 ¼ jHðjoÞj2 jXðjoÞj2 :

Fig. 6.5.3 Inputs, outputs and transfer functions

202

6 Systems and Circuits

or in the Fourier domain by HðjoÞ. Its impulse response is hðtÞ ¼ L1 ½HðsÞ or hðtÞ ¼ F1 ½HðjoÞ: jHðjoÞj and ﬀHðjoÞ are the amplitude and the phase & responses of the linear system. A causal continuous-time LTI system is memoryless if and only if hðtÞ ¼ cdðtÞ, where c is a constant. Noting that the output of a causal continuous-time LTI system is described in terms of the input xðtÞ and its impulse response hðtÞ, it is expressed by Z1 hðaÞxðt aÞda yðtÞ ¼

¼

1 Z1

cdðaÞxðt aÞda ¼ cxðtÞ:

each frequency, corresponding to a linear timeinvariant (LTI) system, are tied by (6.5.20). The response of a LTI system with a transfer function HðjoÞ to a unit input Hð0Þ:

6.5.2 Bounded-Input/Bounded-Output (BIBO) Stability BIBO stability of a LTI system is tied to its impulse response hðtÞ. Consider the output of the LTI system described by the convolution integral to a bounded input xðtÞ with jxðtÞj M. It can be shown that yðtÞ is bounded provided the impulse response hðtÞ is absolutely integrable. That is,

1

yðtÞ ¼

Z1

xðt bÞhðbÞdb ) jyðtÞj

1

6.5.1 Eigenfunctions

The transfer function of a linear time-invariant system can be expressed in terms of its impulse response hðtÞ with the input of the form xðtÞ ¼ est . Then the system output is Z 1 yðtÞ ¼ T½xðtÞ ¼ hðtÞ xðtÞ ¼ hðaÞxðt aÞda 1

¼

Z1

hðaÞesðtaÞ da ¼ est HðsÞ:

1

(6:5:18)

(6:5:19)

is called an eigenfunction (or a characteristic function) and HðsÞ is the eigenvalue (or the characteristic value). That is, est is the eigenfunction and the eigenvalue is defined as the system function. In terms of Fourier transforms, yðtÞ ¼ HðjoÞejot :

jxðt bÞjjhðbÞjdb M

1

Z1

Z1

jhðbÞjdb:

1

(6:5:21) 5 jhðbÞjdb ¼ K 1 ) jyðtÞj MK ¼ N; (6:5:22:)

1

If (6.5.22) is satisfied, then the output is bounded and the system is BIBO stable. Example 6.5.1 Determine if the system described by its impulse response hðtÞ ¼ eat uðtÞ; a40 is BIBO stable. Solution: The system is BIBO stable since

An equation satisfying Tfest g ¼ HðsÞest :

Z1

(6:5:20)

It is the response of a linear time-invariant system with a transfer function HðjoÞ to an input ejot and the relationship holds for each o. The responses for

Z1 1

jhðbÞjdb ¼

Z1 e

2ab

1 2ab 1 1 db ¼ e 0 ¼ 2a 51: 2a

0

&

BIBO stability requires that the transfer function is strictly proper. That is, the degree of the numerator polynomial of the transfer function is less than the degree of the denominator polynomial. Otherwise, the impulse response will contain the derivatives of the impulse function, which are not absolutely integrable. The ideal differentiator is

6.5 Impulse Response

yðtÞ ¼ dxðtÞdt

LT

! sXðsÞ; HðsÞ ¼ s; hðtÞ ¼ d0 ðtÞ:

The derivative of an impulse function is not absolutely integrable. The transfer function is not strictly proper. Similarly the ideal integrator has a transfer function HðsÞ ¼ 1=s has a simple pole on the jo axis at the origin and is marginally stable. Stability analysis is an important topic in all areas of systems engineering, especially in control systems. The literature is extensive in this area and the discussion here is limited to simple ideas. A linear timeinvariant system is stable if every root of its characteristic equation, i.e., the poles of the transfer function HðsÞ have negative real parts. The natural and forced responses of these systems can be described by seeing the properties of the inverse Laplace transforms of response functions with poles at various locations on the s plane. As mentioned earlier the responses of systems have two parts, one is a natural response that is due to the system itself and the other one is the response of the system when there is input. If a characteristic root is a simple real and negative, then the response corresponding to this pole is exponentially decaying. If a root is multiple and real, then the response is a polynomial intmultiplied by an exponentially decaying response. If we have a simple complex conjugate poles on the imaginary axis, then the corresponding response is oscillatory. A system with simple finite poles on the imaginary axis is called wide sense stable or marginally stable. The ideal integrator has a transfer function HðsÞ ¼ 1=s has a simple pole on the jo axis at the origin. It is marginally stable. If we have a pair of complex conjugate poles on the left half plane, the corresponding response is exponentially decaying oscillatory response. If the poles are multiple complex conjugate, then the time response has the form of a polynomial multiplied by an oscillatory decaying response. The systems are stable if its transfer function has all poles on the left half of the splane. The behavior of the impulse response depends on the poles closest to the imaginary axis. For most systems these are simple complex poles, referred to as dominant poles. Complicated systems with transfer function given by HðsÞ generally have many poles and are approximated by a reduced-order system HR ðsÞ by keeping only the poles near the imaginary axis, referred to as a model reduction.

203

Example 6.5.2 Illustrate the model reduction of the system with the transfer function

1 1 HðsÞ ¼100 ðs þ 1Þ ðs þ 5Þ

LT

! 100et uðtÞ

100e5t uðtÞ ¼ hðtÞ Solution: Noting that the pole at s ¼ 5 is farther away than the pole at s ¼ 1, the transfer function HðsÞ can be approximated by a reduced-order function LT 100 ! 100et uðtÞ ¼ hR;1 ðtÞ: HR;1 ðsÞ ¼ ðs þ 1Þ Another way is ignore the poles away from the imaginary axis. Then 400 ) HR;2 ðsÞ 5ðs þ 1Þð:2s þ 1Þ LT 80 ¼ ! 80et uðtÞ ¼ hR;2 ðtÞ: ðs þ 1 Þ

H ðsÞ ¼

&

6.5.3 Routh–Hurwitz Criterion (R–H criterion) The R–H criterion Kuo (1987) provides a test if the roots of a polynomial given below are on the left half plane without actually factoring the polynomial.

DðsÞ ¼ dn sn þ dn1 sn1 þ þ d1 s þ d0

(6:5:23)

Without loosing any generality the coefficient of sn is assumed to be 1. Furthermore, if d0 ¼ 0, i.e., the polynomial has a root at s ¼ 0, the polynomial can be divided by s and test the resulting polynomial for the stability. The Routh array starts by arranging two rows consisting of the coefficients of the polynomial in the following form: Note that ðn kÞth row starts with the coefficient nk ðs Þ. If n is even (odd), then d0 is the last entry in row n (n1). The next step is construct row (n2) by using rows n and n1.

204

6 Systems and Circuits

Row n : ðsn Þ : Row n 1 : ðsn1 Þ :

dn dn1

dn2 dn3

dn4 dn5

::: ::: (6:5:24)

dn1 dn2 dn dn3 dn1 dn1 dn4 dn dn5 ::: dn1

Row n 2 : ðs n2 Þ :

(6:5:25)

The entries in row n2 can be written in terms of determinants. 1 dn dn2 Row n 2 : dn1 dn1 dn3 1 dn dn4 ::: (6:5:26) dn1 dn1 dn5

Row 4 :ðs4 Þ :

DðsÞ ¼ s4 þ 2s3 þ 3s2 þ 4s þ 5:

Row 3 :ðs Þ :

2 3 4

1 2

¼ 1 2 4 1 ð5=2Þ ¼ 1 Row 1 :ðs1 Þ : 1 1 5=2 1 0 ¼ 5=2 Row 0 :ðs0 Þ : 1

1 2

2

Entries in the first column can be written by ½1; 2; 1; 1; 5=2. There are two sign changes indicating that there are two roots on the right half splane. Using MATLAB, the roots of the polynomial can be computed by using the following: d ¼ ½1 2 3 4 5 : coefficents of the polynomial :2878 j1:4161 r ¼ rootsðdÞ : Gives the roots : 1:288 j:8579 d ¼ polyðrÞ : Gives the coefficeients of the polynomial Routh array gives the information about the number of roots on the right half of the splane and not the actual roots. In the Routh array, there are divisions. If these division

(6:5:27)

Solution: Routh array is given below.

1

3

Row 2 :ðs2 Þ :

Row n3 is computed in a similar manner using rows n1 and n2. The procedure is continued until we reach row 0. The R–H criterion states that all the roots of the polynomial D(s) lie on the left half of the s-plane if all the entries in the left most column of the Routh array are nonzero and have the same sign. The number of sign changes in the leftmost column is equal to the number of roots of D(s) in the right half s-plane. Example 6.5.3 Determine the number of roots that are on the right half s-plane of DðsÞ.

3

5

4 0 5

0

2

¼

5 2

0

0

0

0

0

terms are zero, then a different technique is & needed to overcome this. Special cases: 1. Routh array has a zero in the first column of a row. 2. Routh array has an entire row of zeros. 1. If the first entry in the row ðn iÞ; i 6¼ 0 or 1 is zero, to compute the entries in the row ðn i þ 1Þ, a problem of division by zero arises. To alleviate this problem, e is assigned to 0. e is allowed to approach zero either e ! 0þ or 0 . Example 6.5.4 Consider the polynomial DðsÞ ¼ s4 þ s3 þ 2s2 þ 2s þ 1. Use the Routh array to determine the number of roots on the right half of the s plane.

6.5 Impulse Response

205

Solution: Routh array is given by 4

1

2 1

3

Row 3 : ðs Þ : Row 2 : ðs2 Þ :

1 0ðeÞ

2 0 1 0

Row 1 : ðs1 Þ : Row 0 : ðs0 Þ :

12e e 1

0 0

Row 4 : ðs Þ :

In the next step the entries in the first column are written by First column : ½1;1;e;ð12eÞ=e;1 ! fe¼ 0þ ) ½þ;þ;þ;;þg; e¼ 0 ) ½þ;þ;;þ;þg Using either of the two cases, there are two sign changes. There are two roots on the right half of the s-plane. Using MATLAB, the roots of DðsÞ are 0:1247 j1:3066; 0:6217 j0:4406. The polynomial has two roots in the right half of the s plane.& 2. Next consider the case that an entire row in the Routh array consists of zeros. To illustrate this consider the following possibilities: a. Roots are located on the imaginary axis b. Roots with symmetry about origin c. Roots with quadrant symmetry Each of these implies the following type of factors in the polynomial DðsÞ: a: ðs jbÞ ! ðs2 þ b2 Þ

(6:5:28a)

b: ðs aÞ ! ðs2 a2 Þ

c: ðs a jbÞ ! s4 þ½2ða2 þ b2 Þ 4a2 s2 þ ða2 þ b2 Þ2 (6:5:28b)

Row 4 : ðs4 Þ : Row 3 : ðs3 Þ :

1 1

1 1

2 0

Row 2 : ðs2 Þ : Row 1 : ðs1 Þ :

2 4

2 0

0 0

Row 0 : ðs0 Þ :

2

These roots produce even polynomials resulting in a row of zeros in the Routh array. The row before the row of zeros in the array gives the even polynomial identified here as D2 ðsÞ and is called the auxiliary equation. That is DðsÞ ¼ D1 ðsÞD2 ðsÞ. To complete the Routh array, take the derivative of the auxiliary equation and replace the row of zeros by the row obtained from the coefficients of the derivative of the auxiliary equation. Example 6.5.5 Consider the polynomial DðsÞ ¼ s4 þ s3 s2 þ s 2. Show that the system described by this characteristic polynomial is unstable using the Routh array. Solution: Routh array is given by Row 4 : ðs4 Þ :

1

3

Row 3 : ðs Þ : 1 2 Row 2 : ðs Þ : ð1 ð1ÞÞ ¼ 2 Row 1 : ðs1 Þ :

0

1

2

1 2

0 0

0

Noting that the row 1 has all zeros, the auxiliary equation can be written from row 2 and D2 ðsÞ ¼ ðs2 þ 1Þ ¼ 0. Since the (–) sign is irrelevant for the roots, the sign can be ignored and written as DðsÞ ¼ D1 ðsÞðD2 ðsÞÞ and D1 ðsÞ ¼ s2 þ s 2. The auxiliary polynomial D2 ðsÞ has a pair of imaginary roots at s ¼ j1 : In this simple example, the polynomial D1 ðsÞ can be factored and its roots are located at s ¼ 1; 2 indicating that DðsÞ has one root on the right half of the s-plane. If the number of roots D1 ðsÞ is higher than 2, then the Routh array can be continued in the following manner.

ðAuxiliary polynomial; D2 ðsÞ ¼ ð2s2 þ 2ÞÞ ðD02 ðsÞÞ

Entries of the first column in the above Routh array are ½1; 1; 2; 4; 2 indicating there is one

root inside the right half of the s-plane. As mentioned before, Routh array does not provide the

206

6 Systems and Circuits

roots of the polynomial. It merely identifies the number of roots in the right half s-plane. Routh array is frequently used in feedback control systems to determine the condition of stability of a control system. Note that if all the coefficients of the characteristic polynomial DðsÞ do not have the same sign, the polynomial has some roots on the right half s-plane and the corresponding system is unstable. Example 6.5.6 Using the Routh array determine the range of values for K for which all the roots of the polynomial DðsÞ ¼ s3 þ 3s2 þ 3s þ K are located inside the left half plane. Solution: The Routh array is Row 3 :ðs3 Þ : Row 2 :ðs2 Þ :

1 3 ðK 9Þ Row 1 :ðs1 Þ : 3 Row 0 :ðs0 Þ : K

3 K 0

To have all the roots of the polynomial on the left half plane, the coefficients in the first column in the Routh array must have the same signs. This implies that ð9 KÞ40 and K40. All the roots are on the right half plane if 05K59, which gives the range of values of K to keep the system stable. When K ¼ 9, row 1 has all zeros. Correspondingly, the auxiliary polynomial is D2 ðsÞ ¼ ð3s2 þ 9Þ, indicating polynomial has a pair of roots on the imaginary axis. In the case of K ¼ 0, there is a root at s ¼ 0: In the case of a root at s ¼ 0; it is evident from the polynomial that DðsÞ ¼ sD1 ðsÞ and the Routh & array can be determined starting with D1 ðsÞ. Notes: A polynomial DðsÞ with all its roots on the left half s-plane is called a strictly Hurwitz polynomial. If it has all its roots on the left half s-plane and in addition, it has simple poles on the imaginary axis, & then it is called a pseudo-Hurwitz polynomial.

yðtÞ ¼ T ejot ¼ HðjoÞejot ¼ fjHðjoÞjejfðoÞ gejot ¼ jHðjoÞjej½otþfðoÞ :

(6:5:30)

For a particular value of o ¼ o0 , (6.5.30) reduces to T ejko0 t ¼ Hðjko0 Þejko0 t ¼ jHðjko0 Þjej½ko0 tþfðko0 Þ (6:5:31) Since the system under consideration is a LTI system, the response to several frequencies can be determined by (6.4.31). The system with the frequency response HðjoÞ acts like a gate to allow certain frequencies fully or partially through or attenuated or eliminated. Example 6.5.7 Consider a LTI system with a transfer function HðjoÞ. Use (6.5.30) to find the responses to the real periodic inputs given by a: xT ðtÞ ¼

1 X

Xs ½kejko0 t ;

k¼1

b: xT ðtÞ ¼ Xs ½0 þ

1 X

d½k cosðko0 t þ y½kÞ: (6:5:32)

k¼1

Solution: a. Using (6.5.30), the output of the linear time-invariant (LTI) system is yT ðtÞ ¼ HðjoÞ

1 X

Xs ½kejko0 t

k¼1

¼

1 X k¼1

Hðjko0 ÞXs ½kejko0 t ¼

1 X

Ys ½kejko0 t :

k¼1

(6:5:33) If the input to a LTI system is periodic, then the output is also periodic with the same period and the F-series coefficients of the output and the input are related by

6.5.4 Eigenfunctions in the Fourier Domain

Ys ½k ¼ Hðjko0 ÞXs ½k ¼ jHðjko0 ÞjjXs ð½kjejðfðko0 Þþy½kÞ

In terms of the Fourier domain, we have from (6.5.20) that

jYs ½kj ¼ jXs ½kjjHðjko0 Þj; ﬀYs ½k ¼ ﬀXs ½k þ fðko0 Þ: (6:5:34b)

(6:5:34a)

6.5 Impulse Response

207

b: yT ðtÞ ¼Hð0ÞXs ½0 1 X ðjHðko0 Þjd½kÞ cosðko0 t þ fðko0 Þ þ k¼1

þyðko0 ÞÞ:

(6:5:35) &

Notes: The response given in (6.5.33) is the steadystate response of the linear system to a periodic input. A linear time-invariant system does not produce any new frequencies. The output amplitudes and the phases of the harmonics are different from the amplitudes and the phases of the input signal harmonics and are determined by (6.5.34b), for a & real periodic input. Example 6.5.8 Use the Fourier transforms to derive the output given in (6.5.33). Solution: xT ðtÞ ¼ ¼

1 X

Xs ½kejko0 t

T ¼ 2p; o0 ¼ 1:

(6:5:37)

The transfer function, the amplitude, and phase responses are given by

HðjoÞ ¼

pXs ½kdðo ko0 Þ

1 1 ; jHðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ð1 þ joRCÞ 1 þ ðoRCÞ2

ﬀ tanðoRCÞ: 1 X

(6:5:38)

pXs ½kdðo ko0 Þ

k¼1

¼

4 cosðtÞ cosð3tÞ cosð5tÞ xT ðtÞ ¼ þ ::: ; p 1 3 5

! XðjoÞ

k¼1

1 X

Solution: The Fourier series of the input waveform is given by

FT

k¼1 1 X

YðjoÞ ¼ HðjoÞ

Example 6.5.9 Find the output yT ðtÞ of the RC circuit in Fig. 6.5.4b corresponding to the periodic pulse signal shown in Fig. 6.5.4a with a period equal to T ¼ 2p.

The kth harmonic term and the steady-state output response are, respectively, given by

pfHðjko0 ÞXs ½kgdðo ko0 Þ:

k¼1

Taking the inverse transform of the transform of YðjoÞ in (6.4.35), we have yT ðtÞ ¼

1 X

fHðjko0 ÞXs ½kgejko0 t :

(6:5:36)

&

1 jHðjko0 Þj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ﬀHðjko0 Þ 1 þ ðko0 RCÞ2 ¼ tan1 ðko0 RCÞ:

(6:5:39)

k¼1

Fig. 6.5.4 (a) Periodic pulse waveform and (b) RC circuit

(a)

(b)

208

6 Systems and Circuits

4 cosðt tan1 ð1ÞÞ cosð3t tan1 ð3ÞÞ cosð5t tan1 ð5ÞÞ pﬃﬃﬃ pﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃ yT ðtÞ ¼ þ ::: : p 3 10 5 26 1 2

(6:5:40)

Note the kth harmonic input produces the kth harmonic output illustrated below.

4 4 cosðko0 tÞjo0 ¼1 ! qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ cosðko0 t tan1 ðko0 RCÞÞjo0 ¼1 : p p 1 þ k2 o20 ðR2 C2 Þ

Attenuation is proportional to ð1=kÞ and the phase shift is tan1 ðko0 RCÞ in the kth harmonic term. The RC circuit is a low-pass filter. The low frequencies have smaller attenuations and higher frequencies are significantly & attenuated.

hðtÞ ¼

6.6 Step Response ¼

sðtÞ ¼ hðtÞ uðtÞ ¼

Z1

hðbÞuðt bÞdb ¼

1

Zt hðbÞdb: 1

(6:6:1) In the above equation the variable of integration is b not t and uðt bÞ ¼ 0 for t5b. The step response can be obtained by integrating the impulse response and the impulse response can be obtained by differentiating the step response. That is,

hðtÞ ¼

dsðtÞ : dt

(6:6:2)

(6:6:3a)

The step response is

sðtÞ ¼

The step response of a continuous time LTI system is the response to a step input xðtÞ ¼ uðtÞ. The step response sðtÞ; is related to the impulse response hðtÞ (see (6.5.10)) and

1 t=RC e uðtÞ: RC

(6:5:41)

Zt hðbÞdb 1 Zt

1 ðb=ðRCÞÞ uðbÞdb ¼ ð1 eðt=RCÞ ÞuðtÞ: e RC

1

(6:6:3b) The impulse response from the step response by dsðtÞ dð1 et=RC ÞuðtÞ ¼ dt dt 1 t=RC t=RC Þ þ uðtÞ e ¼ dðtÞð1 e RC

hðtÞ ¼

¼ ð1=RCÞet=RC uðtÞ: Note ð1 et=RC Þ is continuous at t ¼ 0 and dðtÞð1 et=RC Þ is zero, see (1.4.5). The impulse and the step responses are sketched in Fig. 6.6.1. The rise time of the RC circuit is the time required for a unit step response to go from 10 to 90% of its final value. It is given by tr ¼t2 t1 ; sðt1 Þ ¼ ð1 et1 =RC Þ ¼ :1;

Example 6.6.1 Determine the step response of the RC circuit in Example 6.3.1 from the impulse response and vice versa. Solution: From (6.5.16a), the impulse response is

sðt2 Þ ¼ 1 et2 =RC 0:9 ¼ et1 =RC ; and 0:1 ¼ et2 =RC ; tr ¼ ðt2 t1 Þ ¼ RC lnð9Þ ¼ 2:197RC:

(6:6:4)

6.6 Step Response

209

Fig. 6.6.1 (a) Impulse response and (b) step response

(a)

Rise time is a measure of how fast the system responds to an input. It is related to the bandwidth & of the circuit and we will discuss this shortly. Rise time and the 3 dB bandwidth: The output transform and the transfer function of the RC circuit given Fig. 6.3.4a are 1 YðjoÞ ¼ XðjoÞ ¼ HðjoÞXðjoÞ; 1 þ joRC 1 : (6:6:5) HðjoÞ ¼ 1 þ joRC The amplitude and the phase responses of the transfer function are 1 20ogjHðjoÞj ¼ 20 log qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ dB; 1 þ ðoRCÞ2 ﬀHðjoÞ ¼ tan1 ðoRCÞ:

(6:6:6)

The responses are shown in Fig. 6.6.2 for positive frequencies. Note the amplitude response is even and the phase response is odd. The amplitude at o ¼ 0 is 1 and in dB, it p isﬃﬃﬃ0 dB. At o ¼ 1=RC, the magnitude is equal to 1/ 2 and in dB this is 3 dB. The 3 dB frequency (or the half-power) is

Fig. 6.6.2 (a) Amplitude response and (b) phase response

(b)

o3dB ¼ 1=RC; or f3dB ¼ 1=2pRC Hertz

(6:6:7)

The amplitude response jHðjoÞj decreases smoothly for higher frequencies and goes to zero at infinity. The rise time is related to the 3 dB bandwidth and is tr ¼ 2:197=ð2pf3dB Þ ¼ :35=f3dB :

(6:6:8)

In summary, the RC circuit is a simple low-pass filter passing frequencies between 0 and f3 dB with small attenuations and all the higher frequencies are attenuated significantly. The phase response is zero at o ¼ 0: At the 3 dB frequency, it is equal to ðp=4Þ and at the infinite frequency the phase & response reaches ðp=2Þ rad or 90 : Ideal integrator: The transfer function of the ideal integrator is HðsÞ ¼ 1=s. The amplitude and the phase responses are, respectively, given by HðjoÞ ¼ð1=joÞ ¼ ðj=oÞ; jHðjoÞj ¼ 1=joj; ﬀHðjoÞ ¼ p=2; o40: (6:6:9) This function represents an ideal integrator by noting that if the input is a sinusoid, say cosðotÞ, the output of the integrator and its transform are

(a)

(b)

210

6 Systems and Circuits

yðtÞ ¼

Zt

xðtÞdt ¼

1 1 p sinðotÞ ¼ cos ot o o 2

1

(6:6:10) jﬀHðjoÞ

YðjoÞ ¼ HðjoÞXðjoÞ; HðjoÞ ¼ jHðjoÞje jHðjoÞj ¼ 1=joj; ﬀHðjoÞ ¼ j sgnðoÞ:

;

(6:6:11)

The amplitude response is inversely proportional to joj; the phase response for o40 is ðp=2Þ, a constant. Since the amplitude gain of an ideal integrator is ð1=jojÞ, it suppresses the higher frequency components and enhances the low-frequency components. The noise signals contain mostly high-frequency components, and the integrator reduces the size of the high-frequency components. Ideal differentiator: The transfer function of an ideal differentiator is HðsÞ ¼ s; HðjoÞ ¼ jo:

(6:6:12)

The amplitude and phase responses are given by jHðjoÞj ¼ jojand ﬀHðjoÞ ¼

p ; o40: 2

(6:6:13)

Consider that a sinusoidal function xðtÞ ¼ cosðotÞ is passed through a differentiator, then

yðtÞ ¼

dxðtÞ d cosðotÞ ¼ ¼ o sinðotÞ dt dt

p ¼ o cos ot þ : 2

Example 6.6.2 Show the circuit shown in Fig. 6.6.3 can be used as a differentiator.

Fig. 6.6.3 RL circuit

Solution: The output transform is YðjoÞ ¼

joL jojL XðjoÞ; jHðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; R þ joL R2 þ ðoLÞ2 p oL : (6:6:15) ﬀHðoÞ ¼ tan1 2 R

For small frequencies, i.e., joj ¼ ðR=LÞ, the output transform can be approximated. Noting the Ftransform derivative theorem, it follows that the circuit acts like a differentiator. That is,

(6:6:14)

Note the amplitude response increases linearly with frequency o and phase response is constant and is equal to ðp=2Þ rad for positive frequencies. From the amplitude response expression, we see that the high frequencies are enhanced. Most corrupted signals contain noise components that are of high

Fig. 6.6.4 RC circuit response to a pulse input

frequencies. Using a derivative function enhances the noise signal much more than a low-frequency signal. Derivative function is used to sharpen a signal. For example, to sharpen an image at the edges, we use a derivative function. Note the discontinuity in the phase response at o ¼ 0.

FT

YðjoÞ jo½LxðjoÞ ! L

dxðtÞ yðtÞ: (6:6:16) dt

&

Example 6.6.3 Find the response of the RC circuit in Fig. 6.3.4a to the input pulse xðtÞ ¼ AP

t T2 : T

(6:6:17)

6.6 Step Response

211

Solution: The transfer function HðjoÞ, its impulse response hðtÞ, output frequency response and the response yðtÞ using the convolution integral, we have YðjoÞ ¼HðjoÞXðjoÞ

HðjoÞ ¼

1 ð1 þ joRCÞ

yðtÞ ¼

Z1

FT

! hðtÞ xðtÞ ¼ yðtÞ;

FT

!

(6:6:18)

1 t=RC e uð t Þ ¼ hð t Þ RC (6:6:19)

hðt bÞxðbÞdb ; hðt bÞ

Assuming the unit step input response is hu ðtÞ, the delayed step input response is hu ðt TÞ. The response to the pulse input is & Aðhu ðtÞ hu ðt TÞÞ: Simple frequency analysis of the RC circuit in the last example: The Fourier transforms of the input, the transfer function and the output transform are t T2 sinðoT=2Þ joT ¼ At e 2; XðjoÞ ¼F AP T ðoT=2Þ 1 : (6:6:22) HðjoÞ ¼ 1 þ joRC

1

8 < 1 ðtbÞ=RC ; b5t e ¼ RC : : 0; b4t

(6:6:20)

9 8 0; t50 > > > > > > > > Zt > > > > A > > ðtbÞ=RC > > > db; 0 5 t 5 T > = < RC e yðtÞ ¼ 0 > > > > > > ZT > > > > A > > ðtbÞ=RC > > 4 > > db; t T e > > ; : RC 0 8 t50 > < 0; t=RC ¼ Að1 e Þ; 05 t 5 T : > : T=RC ðtTÞ=RC Að1 e Þe ; t 4T (6:6:21) Note that the input pulse is the same as in Example 2.2.4, except the pulse is of width T instead of 2T and the pulse started at t ¼ 0 rather than at t ¼ T. The function in (6.6.21) is sketched in Fig. 6.6.4. The response can be visualized by the following argument. For t50, the input is zero and the output is zero as well. At t ¼ 0 we have a step input and the capacitor voltage cannot charge instantaneously and the voltage across the capacitor starts at 0 and increases exponentially with a time constant RC. At t ¼ T, the input becomes zero and for t4T the charge across the capacitor discharges through the resistor and the capacitor voltage decreases exponentially from the peak value of Að1 eT=RC Þ to zero as t ! 1. Another way to derive (6.6.2) is that the input pulse function is AP½ðt ðT=2ÞÞ=T ¼ AuðtÞ Auðt TÞ.

YðjoÞ ¼

At sincðoT=2ÞejoT=2 : ð1 þ joRCÞ

(6:6:23)

We will sketch the amplitude of the output transform by considering two special cases: a. Pulse width T is very large compared to the time constant t ¼ RC (i.e., T RCÞ b. Time constant is very small compared to the time constant (i.e., T RCÞ. Now jYðjoÞj ¼ jHðjoÞjjXðjoÞj:

(6:6:24)

For the two special cases, the functions jHðjoÞj and jXðjoÞj are sketched in Fig. 6.6.5a,b. In case a, the 3 dB bandwidth is assumed to be much larger than the main lobe width of the response. That is, ð1=ð2pÞf3dB Þ ¼ ð1=RCÞ 1=T or T RC and the function jHðjoÞj is essentially flat in the range joj51=T. In this frequency band the amplitude of the output transform is approximately equal to the magnitude of the input transform and we can approximate and jYðoÞj kjXðoÞj; k a constant. The output pulse will be a good approximation of the input pulse. In case b, ð1=ð2pf3dB ÞÞ 1=T or T RC. From Fig. 6.6.5 we see that jXðjoÞj is essentially flat in the 3 dB frequency range. That is, jYðjoÞj jHðjoÞj in this range. This indicates that the amplitude of the output transform looks more like the magnitude of the system transform in this case. We are interested in the input signal transform, not the transform of the system. When a signal is passed through a system, the bandwidth of the system must be much larger than the

212

6 Systems and Circuits

Fig. 6.6.5 Frequency analysis of an RC circuit with a pulse input(a) T ¼ RC, (b) T ¼ RC

(a)

(b) bandwidth of the input signal in order the output response to have some resemblance of the input. Let us now consider some simple ideas about the response of a sequence of pulses. The detection of the existence of the pulse at the output can be improved by increasing the amplitude of the input pulse or increasing the width of the pulse or both. Increasing the amplitude increases the power requirements on the input. Increasing the pulse width implies that the number of pulses that can be transmitted per unit time has to be reduced. In addition the pulses are not band limited. Since the RC circuit is a LTI system, it is conceivable that if the input consists of a set of pulses, the output response will be a sum of the individual responses of the pulses with appropriate delays in the pulse responses. The sum may become unbounded. Next we will consider the process of removing the effects of the RC circuit. This type of a situation appears in measuring the voltage across a

component, seeing a picture through a lens and many others. In these measurements, the signal is affected by the measuring device or the system we visualize with. If the bandwidth of these devices is much, much larger than the signal bandwidths, then the effect of the measuring devices is minimal. Removing the effects of the transmission system from the received signal is an important problem and this process is called the deconvolution and is discussed next. Deconvolution: Let the output response of a LTI system with the transfer function HðjoÞ, and the corresponding impulse response hðtÞ, is given by yðtÞ ¼ hðtÞ xðtÞ

FT

! HðjoÞXðjoÞ ¼ YðjoÞ: (6:6:25)

To recover the signal, xðtÞ from yðtÞ, consider Fig. 6.6.6. The first block represents a system and

6.7 Distortionless Transmission

213

Fig. 6.6.6 Deconvolution

the second block identified by HR ðoÞ represents a system to recover the original signal. The output transfrom; ZðjoÞ is related to the input transform, assuming no loading effects, is ZðjoÞ ¼ HR ðjoÞHðjoÞXðjoÞ:

(6:6:26)

To recover the signal, it is desired to have jHðjoÞjjHR ðjoÞj k, a gain constant in the frequency range of interest. Since there is an inherent delay in every system, say t seconds, this delay can be incorporated and zðtÞ ¼ kxðt tÞ. This implies that

yðtÞ ¼H0 xðt t0 Þ ðH0 and t0 40 are some constants:Þ: (6:7:1) For simplicity, assume H0 40. We essentially tried to obtain a distortionless signal in using the deconvolution process in Fig. 6.6.6. Taking the transform of yðtÞ, the output transform and the transfer functions are as follows: YðjoÞ ¼ H0 ejot0 XðjoÞ ! HðjoÞ ¼ H0 ejot0 : (6:7:2)

HðjoÞHR ðjoÞ ¼ k ejo t or HR ðjoÞ ¼ k ejo t =HðjoÞ: The amplitude and the phase responses of the dis(6:6:27) tortionless system are A circuit that gives the transfer function HR ðjoÞ in (6.6.27) may not always be possible. For example, if HðjoÞ ¼ 0 at some frequency o ¼ o i , the function HR ðjoÞ goes to infinity at this frequency. Therefore, HR ðjoÞ can only be approximated. In terms of the time domain, in the ideal case, the inverse transform of ZðjoÞ is given by zðtÞ ¼ F1 ½ZðjoÞ ¼ F1 ½HðjoÞHR ðjoÞXðjoÞ ¼ hR ðtÞ hðtÞ xðtÞ

(6:6:28)

There is perfect deconvolution if hðtÞ hR ðtÞ ¼ dðtÞ and zðtÞ ¼ dðtÞ xðtÞ ¼ xðtÞ:

jHðjoÞj ¼ jH0 j; ﬀHðjoÞ ¼ ot0 :

(6:7:3)

These functions are shown in Fig. 6.7.1, where H0 is assumed to be positive. This implies that all frequencies are attenuated (or amplified) by the same amount. It is referred to as an all-pass system. The phase response is linear. The delay associated with an ideal delay line, a LTI system, can be seen by considering a sinusoidal input xðtÞ ¼ cosðotÞ. The output is H0 cosðoðt t0 ÞÞ. The amplitude response is the same for all frequencies. In addition, the output is H0 cosðoðt t0 ÞÞ ¼ H0 cosðot ot0 Þ. That is, the time shift is t0 and the phase shift is ot0 and the phase is linearly proportional to the frequency o with a slope of ðt0 Þ

6.7 Distortionless Transmission 6.7.1 Group Delay and Phase Delay A system is called distortionless if the output is the same as the input except the signal may be attenuated by the same amount for all frequencies along with a delay of t0 seconds. A distortionless system has the output

The phase response in (6.7.3) and the delay are respectively given by yðoÞ ¼ ﬀHðjoÞ ¼ ot0

(6:7:4)

214

6 Systems and Circuits

Fig. 6.7.1 Amplitude and phase responses of a distortionless system

t0 ¼

dﬀHðjoÞ dyðoÞ dðot0 Þ ¼ ¼ : do do do

(6:7:5)

Group delay: The group delay deals with a group of frequencies (usually referred as delay for simplicity). The result in (6.7.5) can be used to find the delay associated with a transfer function. HðjoÞ ¼ jHðjoÞjejyðoÞ :

dyðoÞ d tan1 bo=ðc ao2 Þ : ¼ Tg ðoÞ ¼ do do (6:7:10) ¼

ðc ao2 Þb boð2aoÞ

1 2

ðbo=ðc ao2 ÞÞ þ1 bðc þ ao2 Þ ¼ : ðbo2 Þ þ ðc ao2 Þ2

ðc ao2 Þ2

(6:7:6)

(6:7:11)

The group delay deals with a group of frequencies (usually referred as delay for simplicity). It is a nonlinear function of frequency and is defined by

Figure 6.7.2 illustrates the delay function in (6.7.11) & and is not constant for all o: In a later section, filters will be designed that have transfer functions with nonlinear phase. When signals passed through such filters, different frequencies are delayed differently. This is not critical for speech signals, as the human ear compensates for small delays. It is a problem in transmitting data and compensation is necessary so that the filter delay equalizer combination has approximately a constant delay in the desired frequency range. A second-order delay equalizer has the form

Tg ðoÞ ¼

dyðoÞ : do

(6:7:7)

Example 6.7.1 Find the group delay associated with the transfer function 1 ; a; b; c40: HðjoÞ ¼ 2 ðc ao Þ þ bðjoÞ

d tan1 ðxÞ 1 dx ¼ : : Use the identity dy 1 þ x2 dy (6:7:8) Solution: The amplitude, the phase response, and the group delay are given by 1 jHðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ðc ao2 Þ þ ðboÞ2 yðoÞ ¼ ﬀHðjoÞ ¼ tan

1

bo : c ao2 (6:7:9)

Hci ðjoÞ ¼

ðbi o2 Þ jai o ; ai ; bi 40: ðbi ai o2 Þ þ jai o

(6:7:12)

The amplitude and phase responses are qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ðbi o2 Þ2 þ ðai oÞ2 jHci ðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ 1; ðbi o2 Þ2 þ ðai oÞ2 ai o : ﬀHci ðjoÞ ¼ 2 tan1 bi o 2

(6:7:13)

6.7 Distortionless Transmission

215

Fig. 6.7.2 Group delay characteristics in Example 6.7.1

Tg(ω) b c

ω

0

Noting that the phase angles of a product of transfer functions add, the amplitude and phase response of a cascade of n second-order sections result in n

n

i¼1

i¼1

yT ðtÞ ¼AjHðjo0 Þj cosðo0 t þ y0 Þ ¼AjHðjo0 Þj cos½o0 ðt þ ðy0 =o0 ÞÞ:

(6:7:17)

To have a distortionless transmission, the output must have the form

Hc ðjoÞ ¼ P Hci ðjoÞ; jHc ðjoÞj ¼ P jHci ðjoÞj ¼ 1 ﬀHc ðoÞ ¼

N X

yT ðtÞ ¼ B cos½o0 ðt t0 Þ:

ﬀHci ðoÞ:

(6:7:14)

i¼1

The parameters a0i s and b0i s can be adjusted so that the amplitude response of the filter cascaded by the delay equalizer has the same magnitude as the filter function HF ðjoÞ and the corresponding phase angle has approximately linear phase (i.e., approximately constant delay) in the desired frequency range. That is, HðjoÞ ¼ HF ðjoÞHc ðjoÞ; jHðjoÞj jHF ðjoÞj ﬀHðjoÞ ¼ﬀHF ðjoÞ þ ﬀHc ðjoÞ; d ﬀHðjoÞ constant: do

(6:7:15a) (6:7:15b)

Phase delay: Consider the input xT ðtÞ ¼ A cosðo0 tÞ to a LTI system with a transfer function HðjoÞ ¼ jHðjoÞjejﬀHðjoÞ ; ﬀHðjoÞ ¼ yðoÞ:

(6:7:18)

(6:7:16)

If the input to a LTI system is a sinusoid then the output is also a sinusoid at the same frequency, although the output may have a different amplitude and phase. See Example 6.5.6. Let the output of the system to the input xT ðtÞ is

Comparing this with (6.7.17), the time delay between the input and the output at the frequency f0 ¼ o0 =2p is t0 ¼ y0 =o0 . For a single frequency, phase delay is appropriate. The phase delay Tp ðoÞ in terms of the system phase response yðoÞ ¼ argðHðoÞÞ is defined by Tp ðoÞ ¼ yðoÞ=o:

(6:7:19)

Notes: For rational transfer functions, yðoÞ is a transcendental function, whereas the group delay is a rational function of o2 making it easier for filter & design. Example 6.7.2 Find the phase delay and the group delay of the transfer function HðjoÞ ¼ ½ð1 joÞ=ð1 þ joÞ:

(6:7:20a)

Solution: The phase, the group delay, and the phase delay responses are yðoÞ ¼ p 2 tan1 ðoÞ ! Tg ðoÞ ¼

dyðoÞ do

2 yðoÞ : (6:7:20b) & ; Tp ðoÞ ¼ 2 1þo o Earlier, we have seen that the delay associated with a system is a function of its phase response. Transfer ¼

216

6 Systems and Circuits

functions of stable systems can have the same amplitude response with different phase and delay responses. There are three important systems to consider. These are minimum phase, mixed phase, and maximum phase systems.

B0 ¼

Consider the transfer function of a system k

Pðs þ pm Þ

:

hðtÞdt; 1 Z1

1 Hð0Þ

HðjoÞdf; o ¼ 2pf:

(6:8:1)

1

Pðs þ zk Þ HðsÞ ¼ K

Z1

1 T0 ¼ hð0Þ

(6:7:21)

m

For stability reasons all the poles located at s ¼ pk are located on the left half of the s-plane. If the zeros of the transfer function s ¼ zk are on the negative half of the s-plane, then the system is a minimum phase system. If some zeros are on the right half s-plane and some are on the left half s-plane, the system is a mixed phase and if all the zeros are located on the right half of the splane, then the system is a maximum phase system.

6.8 System Bandwidth Measures In Sections 4.2.2, bandwidth measures of a signal were briefly studied. The concentration was on the time–bandwidth product and illustrated examples, wherein the bandwidth is inversely proportional to time width of the signal. Similar ideas can be used using the duration of the impulse response of a system and the system bandwidth. In Section 6.3, a simple, but a practical measure, the half-power or the 3 dB bandwidth was considered. This width is the range of frequencies overpwhich the magnitude ﬃﬃﬃ of the function exceeds ð1= 2Þ of its maximum. Half-power bandwidth comes from the fact that the square pﬃﬃﬃ of the magnitude is power and 20logð1= 2Þ ¼ 3dB. There are different measures that are used and some of these are considered.

6.8.1 Bandwidth Measures Using the Impulse Response hðtÞ and Its Transform Hðj!Þ The time and the frequency durations of the impulse response hðtÞ are defined by

From the central ordinates theorems of the F-transforms, the time width times the bandwidth is equal to 1. This definition for the bandwidth makes use of the spectrum on both sides. That is, T0 B0 ¼ 1:

(6:8:2)

For the one side case, which is what we mostly use, divide the frequency width by 2. Example 6.8.1 Using the above measures show that T0 ¼ 1 and B0 ¼ 1 for the following: a: x1 ðtÞ ¼ PðtÞ

FT

! sinc

o

¼ X1 ðjoÞ; 2 FT 1 b: x2 ðtÞ ¼ eat uðtÞ ! ¼ X2 ðoÞ; a þ jo rﬃﬃﬃ FT 2 c: x3 ðtÞ ¼ eat ! paeo2 =4a

(6:8:3)

Solution: a. In Chapter 4, the areas of the rectangular and the sinc functions were considered and the results are T0 ¼ 1; B0 ¼ 1. b. Noting that hð0Þ ¼ 1=2 ðhðtÞ is discontinuous at t ¼ 0Þ and Hð0Þ ¼ 1=a , it follows that Z1 2

e 1

at

Z1 2 a 1 do uðtÞdt¼ and a 2p a þ jo 1 Z a2 1 1 ¼ do 2p 1 a2 þ o2 Z a2 1 1 a do ¼ ; ¼ p 0 a2 þ o 2 2

T0 ¼ ð2=aÞ; B0 ¼ ða=2Þ; T0 B0 ¼ 1:

(6:8:4)

c. By making use of integral tables, the time width and the bandwidth of the Gaussian pulses both & come out to be one and the product is one.

6.8 System Bandwidth Measures

Time functions can take both positive and negative values, some authors use the magnitudes or the squares of the time functions hðtÞ in defining the time width in (6.8.2a). Others use moments to define the time and bandwidths. In the following, the bandwidth measures that are simple and practical will be considered.

6.8.2 Half-Power or 3 dB Bandwidth In Section 6.6, an RC circuit was considered. On the amplitude spectrum, the 3 dB frequency was identified (see Fig. (6.6.2a)). The half-power or the 3 dB bandwidth is a practical measure and is widely used in systems and circuit theory, especially in filter designs. In identifying the 3 dB bandwidths, only positive frequencies are considered. Example 6.8.2 Show the 3 dB bandwidth of the Gaussian function is W (Carlson (1975)). HðjoÞ ¼ eðlnð2=2Þðo=WÞ ðnote; Hðj0Þ ¼ 1Þ: Solution: The half-power frequency is equal to W since

217

value of the spectrum. This measure is simple and it does nottake into consideration any ripples in jHðjoÞj function between the two 3 dB frequencies. A more generalized measure that takes into consideration the ripples by making use of integrals in computing the bandwidths. These methods are used in random signal analysis, as the spectrum of noisy signals have many peaks. For a good discussion on this topic, see Peebles (2001). These measures have been developed using signals rather than systems. To make it uniform in our discussion we will use XðjoÞ rather than HðjoÞ. When we consider examples of transfer functions, HðjoÞ will be used.

6.8.3 Equivalent Bandwidth or Noise Bandwidth The equivalent noise bandwidth is obtained by equating the areas contained in the signal energy spectrum with a pulse spectrum of bandwidth Weq ¼ WN rad/s. Figure 6.8.1 illustrates a signal spectrum and noise equivalent is computed as follows: Z1

jXðjoÞj2 do ¼ jXðjoÞj2max 2Weq

1 2 1 jHðjo3dB Þj2 ¼ ¼ ðe2ðlnð2Þ=2Þ ðo3dB =WÞ Þjo3dB ¼W 2 1 (6:8:5) & ) ¼ e lnð2Þ : 2

Although the 3 dB bandwidth is the most common one, we could obviously define 6 dB bandwidth or any other value for the bandwidth measure. In summary the 3 dB bandwidth computes the width by considering the peak valuepﬃﬃof ﬃ the spectrum and a value (or values) of the (1/ 2) below the maximum

Fig. 6.8.1 Noise equivalent bandwidth

R1 ) Weq ¼

jXðjoÞj2 do

0

jXðjoÞj2max

; Beq ¼

Weq : 2p

(6:8:6)

If the system bandwidth of the transfer function HðjoÞ is of interest, replace XðjoÞ by HðjoÞ in (6.8.6). Example 6.8.3 Determine the noise equivalent bandwidth of the filter transfer function

218

6 Systems and Circuits

jHðjoÞj2 ¼

1 ¼ ½1=ð1 þ ðo=WÞ2 Þ; 1 þ ðoRCÞ2 o ¼ 2pf; W ¼ 1=RC:

Solution: Using (6.8.6) and jHðjoÞjmax ¼ 1, we have Z1 1 W2 Weq ¼ do ¼ W tan1 ðo=WÞ1 0 2 2 2p ðW þ o Þ 0

¼Wp=2:

(6:8:7)

) Beq ¼ ½1=4RCHz:

(6:8:8)

The 3 dB frequency of the RC circuit, f3dB ¼ 1=2pRC is related to the equivalent noise bandwidth and Beq ¼ 1:57B3dB . The equivalent noise bandwidth works as well for signals that have spectrum in the middle, such as the band-pass spectrum. In such cases, using the center frequency o0 for the peak in the amplitude response of the band-pass filter, the equivalent bandwidth is R1 Weq ¼

jHðjoÞj2 do

0

jHðjo0 Þj2

Weq Hz: ; Beq ¼ 2p

by the bandwidth. We can define the root mean square (RMS) bandwidth as Z1

W2RMS ¼

o2 jXno ðjoÞj2 do ¼

R1

2 2 1 o jXðjoÞj do : R1 2

jXðjoÞj do

1

1

(6:8:12) Example 6.8.4 Compute the RMS bandwidth (see Peebles (2001).) and compare it with the 3 dB frequency which is given by jXðjoÞj2 ¼

10 ½1 þ ðo=10Þ2 2

(6:8:13)

:

Solution: Using integral tables, we have Z1 Z1 10 do 5 do ¼ 10 ¼ 50p: ½ð1 þ ðo=10Þ2 2 ½100 þ o2 2 1

1

(6:8:14a) Z1

(6:8:9) &

2

2

o jXðjoÞj do ¼ 10

1

5

Z1 1

o2 ½100 þ o2 2

do ¼ 5000p: (6:8:14b)

6.8.4 Root Mean-Squared (RMS) Bandwidth

FRMS

The RMS bandwidth comes from the statistical measures, where the variance is a measure of the spread of a density function. Consider the low-pass energy spectral density shown in Fig. 6.8.1. The area under this function is the energy Z1 1 (6:8:10) E¼ jXðjoÞj2 do: 2p 1

Now define the normalized energy spectral density function by 2

2

jXno ðjoÞj ¼ jXðjoÞj =E:

5000p ¼ 100; oRMS ¼ 10 rad=s; 50p ¼ 1:5915 Hz:

o2RMS ¼

(6:8:11)

It is real, even, and positive and the area under the function is 1. It has the same properties as a probability density function (PDF). In the PDF case, we define the variance as a measure of the spread of the density function. In this case the spread is measured

1 10 jXðjo3dB Þj2 ¼ 12 jHð0Þj2 ¼ 10 ¼ 5 ¼ 2 ½1 þ ðo3dB =10Þ2 2 : ) o3dB ¼ 6:436 or f3dB ¼ 1:243Hz: In this special case, the 6 dB bandwidth comes out to be the same as the RMS bandwidth. The concepts of RMS bandwidth can be easily extended to band-pass spectra. Assuming that most of the spectra is around o0 , the RMS bandwidth is given by W2RMS

¼

4

R1 0

ðo o0 Þ2 jXðjoÞj2 do R1 2 0 jXðjoÞj do

(6:8:15)

RMS bandwidth is more general than the 3 dB bandwidth, as it can handle general spectra with several peaks and valleys in the passband of the & energy spectral density.

6.9 Nonlinear Systems

219

6.9 Nonlinear Systems In this section we will consider simple nonlinear systems and illustrate the difficulties in the spectral analysis of the responses. A nonlinear system is described by a time domain relationship between the input and the output. This can be expressed in the form of a graphical representation or in terms of a general output function yðtÞ ¼ gðxðtÞÞ, where the output function gð:Þ is a complicated closed form expression or in terms of a power series of xðtÞ. A system is nonlinear if it has components that have nonlinear characteristics, such as a diode. In many cases, nonlinear systems are approximated by linear systems, as they are easier to handle, see Ziemer and Tranter (2002). The system described by a polynomial function of the input xðtÞ, such as yðtÞ ¼

n X

ai xi ðtÞ:

(6:9:1)

i¼0

Example 6.9.1 Find the output yðtÞ defined below and sketch the one-sided line spectra of the input xðtÞ ¼ B1 cosðo1 tÞ þ B2 cosðo2 tÞ; B1 ; B2 40 with oi ¼ 2pfi ; i ¼ 1; 2 and f2 4f1 and yðtÞ, see Ziemer and Tranter (2002). yðtÞ ¼a0 þ a1 xðtÞ þ a2 x2 ðtÞ; ai 6¼ 0; xðtÞ ¼B1 cosðo1 tÞ þ B2 cosðo2 tÞ; B1 ; B2 40: (6:9:2)

Solution: The output is yðtÞ ¼ a0 þ a1 B1 cosð2pf1 tÞ þ a1 B2 cosð2pf2 tÞ þ a2 B21 cos2 ðo1 tÞ þ a2 B22 cos2 ðo2 tÞ (6:9:3) 1þ a2 ð2B11 B2 Þ cosðo1 tÞ cosðo2 tÞ DCoffsetterm ¼½a0 þ a2 B21 þ a2 B22 2 2 Linearterms þ½a1 B1 cosðo1 tÞþa1 B2 cosðo2 tÞ 1 þ a2 ½B21 cosð2o1 tÞþB22 cosð2o2 Þt Harmonicterms 2 þa2 B1 B2 ½cosðo1 þo2 Þtþcosðo2 o1 Þt Inter modulationterms:

is linear if all ai 0s are zero except a1 : If any of the other ai 0s are nonzero, then the system is nonlinear. An example is a device that has saturation nonlinearity. It has a voltage to current characteristic that is linear within a range and outside that range, the voltage saturates, see Fig. 6.9.1a. A device that may have this type of a characteristic is a resistor. The Ohm’s law says that v ¼ Ri is valid in a certain range of currents and voltages. Outside this range, the resistor is a nonlinear component. A hard limiter is an important example. The output voltage is 1 if the input voltage is positive and 1 if the input voltage is negative (see Fig. 6.9.1b.).

Fig. 6.9.1 Examples of nonllinear input–output characteristics

(a)

(6:9:4) Figure 6.9.2 gives the input and the output onesided amplitude line spectra. System nonlinearity created a DC term, linear terms (frequencies f1 and f2 ), harmonic distortion terms (frequencies, 2f1 and 2f2 ), and inter modulation terms (sums and differences of the input frequencies, ( f2 f1 Þ and ð f1 þ f2 Þ). Note that the output of a linear time-invariant system has the same frequencies as the input with possible changes in the amplitudes and phases. Nonlinear system generates new & frequencies.

(b)

220

6 Systems and Circuits

Fig. 6.9.2 Example 6.9.2, (a) Input line spectra and (b) Output line spectra

(a)

6.9.1 Distortion Measures Most systems have inherent nonlinear components. It may be desirable to operate them in the linear region, if possible. Amplification is a good example, where the nonlinearities may be small and the distortions will be small enough that they can be tolerated. The next question is how does one measure the distortions due to a nonlinear system? A simple way is compare a nonlinear system time response to a linear system time response. To achieve these measures, start with a single input, a sinusoid, say xðtÞ ¼ cosðo0 tÞ and measure the distortion due to the nonlinearities in the system described xðtÞ by X yðtÞ ¼ ai xi ðtÞ; where ai 0 s some constants: i

(6:9:5) The powers of the cosine functions can be expressed in terms of sine and cosine terms, where the frequencies will be multiples of the input frequency, i.e., we will have harmonics. The output can be written in terms of trigonometric Fourier series yðtÞ ¼ Y½0 þ

1 X k¼1

A½k cosðko0 tÞ þ

1 X

B½k sinðko0 Þt:

(b) Obviously the frequency of interest f0 ¼ o0 =2p should be in the passband of the signal. Manufacturers of stereo systems provide literature that gives these numbers in terms of dBs for their systems. Clearly, if the distortion terms D½k0 s; k 6¼ 1 are negligible, then the nonlinear system comes close to a linear system.

6.9.2 Output Fourier Transform of a Nonlinear System In the following example, a system with a polynomial nonlinearity is considered and illustrates the effect of the nonlinearities in terms of the input and the output frequencies. Example 6.9.2 Let the input xðtÞ and the output yðtÞ in terms of the input are as given below. Noting F½x2 ðtÞ ¼ ð1=2pÞ½XðjoÞ XðjoÞ, sketch the output spectrum assuming xðtÞ

FT

! XðjoÞ ¼

yðtÞ ¼ a0 þ a1 xðtÞ þ a2 x2 ðtÞ

Yh o i ; W

FT

! YðjoÞ:

(6:9:8)

k¼1

(6:9:6) The constants Y½0; A½k0 s; and B½k0 s are functions of the constants a0i s and the powers of the input sinusoid. The kth distortion term is measured by the ratio sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ A2 ½k þ B2 ½k D½k ¼ : (6:9:7) A2 ½1 þ B2 ½1

YðjoÞ ¼ a0 2pdðoÞ þ a1 XðjoÞ þ ða2 =2pÞ½XðjoÞ XðjoÞ:

(6:9:9)

Solution: Convolution of two identical rectangular pulses is a triangular pulse (see Example 2.3.1.) and h i h i YðjoÞ ¼ a0 2pdðoÞ þ a1 P o þ ða2 =2pÞL o : W W (6:9:10)

6.10 Ideal Filters

221

Note that the approximation is valid for small excursions of x from x0 and we can neglect higherorder terms. It is a linear relationship between small changes in both the input and output related by the slope of the function mat x ¼ x0 , see Nise (1992).

6.10 Ideal Filters Fig. 6.9.3 Output transform of a nonlinear system

The three parts are explicitly shown in Fig. 6.9.3. Note that the width of the triangular pulse is 2 W. Therefore, the bandwidth of the output signal is two times that of the input signal. If a system has a second-order nonlinearity, then the frequency width of the output signal will be doubled that of the input signal. A system with a nth order nonlinearity, then the bandwidth of the output signal will increase from B Hz to ðnBÞ Hz: Most signals are not band limited and the transmission systems have a limited band& width and filtering is necessary. Frequency analysis of a nonlinear system is difficult and may not even be possible. If it can be approximated by a linear system, then frequency domain analysis provides useful information. Time domain analysis is simpler for nonlinear systems.

6.9.3 Linearization of Nonlinear System Functions Nonlinear systems are hard to deal with in general terms. A function gðxÞ can be approximated about a point x ¼ x0 using Taylor series expansion dg jx¼x0 ðx x0 Þ gðxÞ ¼ gðx0 Þ þ dx " # 2 d g ðx x0 Þ2 þ þ ::: dx2 2!

dg dgðxÞ ¼ gðxÞ gðx0 Þ dx

ðx x0 Þ ¼ mjx¼x0 dx: x¼x0

(6:9:12)

In this section we will consider the basics of lowpass, high-pass, band-pass, band-elimination filters, and the ideal delay line filters. The filters are specified based on a transfer function HðjoÞ in terms of its amplitude, phase, or delay responses, jHðjoÞj, phase ﬀHðjoÞ or ½dHðjoÞ=do. Finding HðjoÞ from the specifications is the first step. The next step involves the synthesis. We will consider here the ideal filter functions that describe their functions, and simple circuits that can be used as filters. Low-pass filters allow low frequencies to pass through with small attenuation and attenuate or eliminate high frequencies; high-pass filters eliminate or attenuate low frequencies and allow high frequencies go through with possibly small attenuations; band-pass filters allow a band of frequencies to go though with small attenuation and attenuate or eliminate frequencies that are outside this band; band-elimination or bandreject filters let the low and high frequencies pass through and attenuate or eliminate a band of frequencies somewhere in the middle. Delay line filters are primarily used in cascade with filters so that the cascaded filter delay line combination has an approximate linear phase characteristics. Filters are used in every communication system. If the frequencies of the two signals are disjoint, then we can remove the undesired signal by using a band-pass filter that allows the desired signal to go through with a small attenuation and attenuate or eliminate the undesired signal. Tuning to a particular radio station involves eliminating, i.e., filtering out all the other signals from the other stations all available at the front end of the radio or TV receivers. The DC component can be removed by using a highpass filter or a simple bias removal component, a capacitor.

222

6 Systems and Circuits

Example 6.10.1 Illustrate the Bell System TOUCHTONE telephone dialing scheme. Solution: Bell systems TOUCH-TONE telephone dialing scheme uses some of the filters. The discussion follows that of Daryanani (1976). The filters are used in the detection of signals generated by a push button telephone. As we dial a telephone number, i.e., by pushing a button on the telephone, a unique set of two-tone signals are generated and transmitted to the telephone central office, where the signals are processed to identify the number that is transmitted. The buttons and the tone assignments of a TOUCHTONE telephone are shown in Fig. 6.10.1. It has 12 buttons. These correspond to 10 decimal digits, a star button, and a pound button. The letters are also identified on the buttons. For example, on the button identified by 2 has the letters ABC indicating that the number 2 represents A, B, and C as well as 2. Operator button (0) is used for zero. The star (*) button and the pound (#) button are used for other special purposes, such as responding to queries from an answering machine. There are four other buttons that are not shown and are used for special purposes. The signaling code provides 16 distinct signals that use 4 low and high frequencies given by Low : ð697 Hz; 770 Hz; 852 Hz; 941 HzÞ;

amplified first and the two tones are then separated into two groups by the low-pass and the high-pass filters. The low-pass filters pass the low frequencies with very low attenuation and block the high frequencies. Similarly the high-pass filters pass the high frequencies with very little attenuation and block the low frequencies. The separated tones are then converted to square waves of fixed amplitudes by using limiters. Signals are then passed through eight band-pass filters. Each of these passes only one tone and rejects the others. The band-pass filter characteristics are such that there is very little attenuation for the particular frequency and a significant attenuation to block the other frequencies. For a detailed discussion on the amplitude characteristics of low-pass, band-pass, and high-pass filters, see Daryanani (1976). The outputs of the bandpass filters are fed into detectors. The detectors are energized when their input voltage exceeds a set threshold value and the outputs of the detector provides the required dc switching level to connect & the caller to the party being called. Filters can be implemented either in terms of analog or digital domain. Next we will consider each of the filter types in a more formal fashion and discuss the generation of simple transfer functions that allow for the analysis of these filters.

High : ð1209 Hz; 1336 Hz; 1477Hz; 1633 HzÞ Pressing one of the buttons generates a pair of unique frequencies, one lower and the other higher. The fourth high-band frequency, 1633 Hz, is for special services. The block diagram shown in Fig. 6.10.2 illustrates the detection scheme in the telephone office. The received two tones are

6.10.1 Low-Pass, High-Pass, Band-Pass, and Band-Elimination Filters The words low-pass means that when the signal xðtÞ is passed through a low-pass filter, only low

697 Hz →

770 Hz →

Low-band frequencies

852 Hz →

941Hz →

Fig. 6.10.1 Tone assignments for TOUCH-TONE dialing

ABC

DEF

1 2 GHI

JKL

4

5

PRS

TUV

7

3 MNO 6 WXY

8

9

Oper *

# 0

↑

↑

↑

↑

1209 Hz

1336 Hz

1477 Hz

1633 Hz

High-band frequencies

6.10 Ideal Filters

223

Fig. 6.10.2 Block diagram of detection scheme in the telephone office

frequencies, say 0 to fc ¼ oc =2p, are passed and block all frequencies above the cutoff frequency fc : The amplitude of the ideal low-pass filter transfer function is (H0 is assumed to be positive) H0 ; joj oc o ¼ ;oc ¼ 2pfc : jHLP ðjoÞj ¼ H0 P 2oc 0; joj4oc (6:10:1) Every transmission system takes time, i.e., the signal will be delayed. It is ideal to have this delay to be a constant, say t0 s for all frequencies,which may not be possible. In terms of frequency domain, the output transform of such a system is

o jot0 e ; YðjoÞ ¼ HLp ðjoÞXðjoÞ; HLP ðjoÞ ¼ H0 P 2oc HLp ðjoÞ ¼ H0 Pðo=2oc Þ; and ﬀHLp ðjoÞ ¼ ot0 (6:10:2)

Fig. 6.10.3 Amplitude and phase responses of an ideal low-pass filter

The amplitude and the phase response plots are shown with respect to the frequency variable o in Fig. 6.10.3. On the magnitude plot the band of frequencies from 0 to fc as the passband and the band of frequencies from fc to 1 as the stopband are shown. Since the amplitude spectrum of a real signal is even and the phase spectrum is odd, the discussion can be limited to only positive frequencies. The phase response is assumed to be linear, i.e., slope is constant. The group delay is t0 ¼ dﬀHLp ðjoÞ=do:

(6:10:3)

Can we design a real circuit that has the transfer function HLp ðjoÞ? For a physically realizable system, the impulse response hðtÞ ¼ 0 for t50, i.e., the system is causal. For a realizable system, the output cannot exist before the input is applied. The impulse response of the ideal low-pass filter can be

224

6 Systems and Circuits

Z1

determined from the results in Chapter 4 (see (4.3.28).). It is repeated below. sinðaðt0 t0 ÞÞ pðt t0 Þ

FT

!

Yh o i 2a

1

e

jot0

(6:10:4)

Y o jot0 hLp ðtÞ ¼F1 ½HLP ðjoÞ ¼ F1 H0 e 2oc sinðoc ðt t0 ÞÞ : (6:10:5) ¼H0 ð2fc Þ oc ðt t0 Þ In Fig. 6.10.4, the input, an impulse function dðtÞ, applied at t ¼ 0, the block diagram representing the ideal low-pass filter and the impulse response are identified. The impulse response, a sinc function, peaks at t ¼ t0 , giving a value of ð2fc H0 Þ at this time. From the figure we can see that the response is nonzero for t50: That is, there is output before the input is applied. The ideal lowpass filter is not causal and is physically unrealizable. We can also see this from the Payley–Wiener criterion stated below. For a causal system, the impulse response hðtÞ ¼ 0 for t50, as it does not respond before the input is applied. The causality condition can be stated in terms of the transfer function HðjoÞ. It is called the Paley–Wiener criterion Papoulis (1962) and is given in terms of the inequality

(6:10:6)

If jHðjoÞj ¼ 0 over a finite frequency band, then the above integral becomes infinite. jHðjoÞj can be zero at isolated frequencies and still satisfy the criterion. The criterion describes the physical reliability conditions and is not of practical value. That is, if jHðjoÞj ¼ 0 over any band of frequencies, the Payley–Wiener criterion states that the system is physically unrealizable. Ideal low-pass filter violates the condition. We can make a general statement that if the amplitude spectrum is a brick wall type function, the corresponding transfer function is physically unrealizable. Since the ideal low-pass filter function is physically unrealizable, the next best thing is find a function that approximates the ideal filter characteristics. First, consider the simple RC circuit in Fig. 6.3.4a. The transfer function of this circuit is HLp ðjoÞ ¼ 1=ð1 þ joRCÞ: The frequency amplitude characteristic is shown in Fig. 6.6.2a with the cutoff frequency fc ¼ f3dB ¼ ð1=ð2pRCÞÞ. The input and the output transforms are related by YðjoÞ ¼ HLp ðjoÞXðjoÞ. At a particular frequency fi ,

(a)

Fig. 6.10.4 (a) Impulse input, (b) block diagram of a low-pass filter, and (c) impulse response

jlnjHðjoÞjj do51: ð1 þ o2 Þ

(b)

(c)

6.10 Ideal Filters

jYðjoi Þj ¼ HLp ðjoi ÞjXðjoi Þj; oi ¼ 2pfi :

225

(6:10:7)

The output spectral amplitudes at frequencies f ¼ fi are attenuated from the input magnitude spectral amplitudes by the factor of HLp ðjoi Þ. For frequencies between 0 f fc ¼ ð1=2pRCÞ; the output amplitude spectrum is within 3 dB of the input amplitude spectrum, whereas for f4fc , the output amplitude spectrum is significantly reduced or attenuated. The simple RC low-pass filter allows the low frequencies 0 to f3dB go through with a small attenuation and the frequencies from f3dB to 1 are attenuated significantly. The circuit has low-pass filter characteristics. To raise the amplitude characteristics in the passband and, at the same time, lower the amplitude characteristics in the stopband, the following amplitude response function would work: 1 HLp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : 1 þ ðo=oc Þ2n

(6:10:8)

Since o=oc is less than 1 in the passband, by taking the power of this by (2nÞ, we are decreasing the value of the denominator in the passband, thus increasing the amplitude in the passband. On the other hand, in the stopband, i.e., the band above the cutoff frequency o4oc , the denominator in (6.10.8) increases as o increases above the cutoff frequency, and the value of the function reduces in this range. Figure 6.10.5 gives two sketches for n; say n1 and n2 ; n2 4n1 : In the limit, i.e., when n ! 1, the filter characteristics approach the ideal

Fig. 6.10.5 Butterworth amplitude filter response n1 ¼ 2; n2 ¼ 3; e ¼ 1; oc ¼ 5

characteristics. Equation (6.10.8) can be generalized to control the passband attenuation by choosing 1 HLp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ﬃ: 1 þ e2 ðo=oc Þ2n

(6:10:9)

There are two parameters in (6.10.9), e and n. e controls how far the amplitude characteristics will go down to from a value of 1 at o ¼ 0 to when o ¼ oc . The value of n controls how fast the attenuation of the amplitude characteristics in the stop-band region. The amplitude characteristic goes from 1 to ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ p 1= ð1 þ e2 Þ corresponding to the frequencies 0 and fc , respectively. By using the power series, for small e, we can write 1 ½1=

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 þ e2 e2 =2:

(6:10:10)

The function in (6.10.9) is the Butterworth filter function. It has interesting properties. At o ¼ 0, ð2n 1Þ derivatives of the function 1=½1 þ e2 ðo=oc Þ2n are zero, identified as a maximally flat amplitude response. For jo=oc j 1, the high-frequency roll-off of an nth order Butterworth function is 20n dB/decade. The proofs of these are left as exercises. As n ! 1, the filter response has the ideal low-pass characteristics. In the low-pass filter specifications, three bands, namely passband (0 ! oc Þ, transition band ðoc ! or Þ, and stopband ðor ! 1Þ are identified. The transition band is not

226

shown in Fig. 6.10.6, as it depends on required attenuations at the edges of the transition band. Returning to the simple RC low-pass filter, the element values of one of the two components R or C can be determined from the given cutoff frequency, see (6.6.7). Normally the capacitor

6 Systems and Circuits

value is selected, as the number of available capacitor values is much smaller than that of the available resistor values. As a final step, the time response of these filters, for the two simple first-order low-pass RC and RL filters is shown in Figs. 6.3.4a and 6.6.3. The response is determined by the time

(a)

(b)

(c)

Fig. 6.10.6 Amplitude and phase plots of ideal filters: (a) low-pass, (b) high-pass, (c) band-pass, and (d) band elimination

(d)

6.11 Real and Imaginary Parts of the Fourier Transform of a Causal Function

constants, t ¼ RC for the RC circuit and t ¼ L=R for the RL circuit. The output of an LTI system with input xðtÞ and the impulse response hðtÞ is given assuming it zero for t50 by the convolution integral yðtÞ ¼

Z

t

hðaÞxðt aÞda:

(6:10:11)

0

The impulse response hðtÞ is the circuit weighting function. It gives the amount of memory the circuit has. For example, if hðtÞ ¼ dðtÞ, it gives zero weight to the past values of the input function as Z t Z t hðaÞxðt aÞda ¼ dðaÞxðt aÞda ¼ xðtÞ: yðtÞ ¼ 0

0

(6:10:12)

6.11 Real and Imaginary Parts of the Fourier Transform of a Causal Function The real and the imaginary parts of the Fourier transform of a causal function xðtÞ are shown to be related below. Let FT

! XðjoÞ ¼ ReðXðjoÞÞ þ j ImðXðjoÞÞ:

xðtÞ

By noting Re½XðjoÞ is even and Im½XðjoÞ is odd and integral of an odd function over a symmetric interval is zero, we have

Z

t

hðaÞxðt aÞda ¼

0

¼

Z

Z

Z1

1 xðtÞ ¼ 2p

If hðtÞ ¼ uðtÞ, then the circuit has a perfect memory giving equal weights and yðtÞ ¼

227

¼

1 2p

1 Z1

XðjoÞejot do

ðRe½XðjoÞ þ j Im½XðjoÞÞ

1

t

uðaÞxðt aÞda

½cosðotÞ þ j sinðotÞdo

0 t

xðt aÞda:

(6:10:13)

0

Ideal filter frequency responses of the low-pass (Lp), high-pass (Hp), band-pass (Bp), and the band-elimination (Be) filters are given below. The amplitude and the phase response plots of these are given for positive frequencies in Fig. 6.10.6a,b,cd, respectively. Note the phase responses of these ideal filter functions are shown as linear. o jo t0 e (6:10:14a) HLp ðjoÞ ¼ H0 P 2oc o ejo t0 HHp ðjoÞ ¼ H0 1 P (6:10:14b) 2oc h ho o i ho o ii 0 0 HBp ðjoÞ ¼ H0 P þP ejo t0 W W (6:10:14c) h h ho o i ho o iii 0 0 HBe ðjoÞ ¼ H0 1 P þP ejo t0 : W W (6:10:14d) The Lp and Hp filter responses have one passband and one stopband. The Bp filter response has one passband and two stopbands. The Be filter response has two passbands and one stopband.

1 ¼ p

Z1 Re½XðjoÞ cosðotÞdo 0

Z1

1 p

Im½XðjoÞ sinðotÞdo: 0

(6:11:1) Noting that xðtÞ is causal, i.e., xðtÞ ¼ 0; t40, results in xðtÞ ¼

1 p

Z1 Re½XðjoÞ cosðotÞdo 0

1 p

Z1

Im½XðjoÞ sinðotÞdo ¼ 0:

0

Since cosðotÞ and sinðotÞ are defined everywhere, Re½XðjoÞ and Im½XðjoÞ are the real and imaginary parts of the transform of the causal function. That is, 1 p

Z1 Re½XðjoÞ cosðotÞdo 0

1 ¼ p

Z1 0

Im½XðjoÞ sinðotÞdo; t40:

228

6 Systems and Circuits

This implies that a causal signal xðtÞ can be expressed in terms of either Re½XðjoÞ or Im½XðjoÞ and xðtÞ ¼

2 p

Z1

Z1 Im½XðjoÞ sinðotÞdo

(6:11:2)

0

These are true as long as there are no impulses at t ¼ 0 and they imply that Re½XðjoÞ and Im½XðjoÞ cannot be specified independently. Using the transform and solving for real and the imaginary parts of the transform, we have

Re½XðjoÞ ¼

2 p

h0 ðtÞ; t40 h0 ðtÞ; t50

¼ h0 ðtÞsgnðtÞ or

h0 ðtÞ ¼ he ðtÞsgnðtÞ:

0

2 p

he ðtÞ ¼

Re½XðjoÞ cosðotÞdo;

xðtÞ ¼

Noting that hðtÞ ¼ 0; t40, the following interesting relations result:

Z1 Z1 Im½XðjvÞ sinðvtÞ cosðotÞdvdt 0

0

(6:11:6)

Using F½sgnðtÞ ¼ 2=jo and using the Fourier time multiplication theorem, h0 ðtÞsgnðtÞ

FT

! Re½HðjoÞ; he ðtÞsgnðtÞjIm½HðjoÞ: (6:11:7)

The real and the imaginary parts of the function can be related by using the frequency convolution theorem studied in Chapter 4. The frequency convolution theorem corresponding to the two functions is given in (6.11.8) using F½xi ðtÞ ¼ Xi ðjoÞ; i ¼ 1; 2.

(6:11:3a) Im½XðjoÞ ¼

2 p

Re½XðjvÞ cosðvtÞ sinðotÞdvdt: 0

! 2p1 ½X1 ðjoÞ X2 ðjoÞ

FT

x1 ðtÞx2 ðtÞ

Z1 Z1 0

(6:11:3b) That is, for a causal signal, the real and imaginary parts of the transform can be expressed in terms of the other. The results in (6.11.3a and b) are difficult to use. A more elegant way of expressing these relations is by using Hilbert transforms discussed in Chapter 5.

6.11.1 Relationship Between Real and Imaginary Parts of the Fourier Transform of a Causal Function Using Hilbert Transform Consider impulse response of a realizable function and its transform hðtÞ ¼ he ðtÞ þ h0 ðtÞ

FT

! HðjoÞ

¼ Re½HðjoÞ þ j Im½HðjoÞ

(6:11:4)

he ðtÞ ¼ ½hðtÞ þ hðtÞ=2; ho ðtÞ ¼ ½hðtÞ hðtÞ=2: (6:11:5)

1 ¼ 2p

Z1

X1 ðjaÞX2 ðjðo aÞÞda:

(6:11:8)

1

Using these in (6.11.7) the following results: 1 2 j Im½HðjoÞ ; 2p jo 1 2 j Im½HðjoÞ ¼ Re½HðjoÞ : 2p jo Re½HðjoÞ ¼

Z1

1 Im½HðjoÞ ¼ p

1

Re½HðjoÞ ¼

1 p

Z1 1

(6:11:9)

Im½HðjoÞ da; ðo aÞ

ImðHðjoÞÞ da: ðo aÞ

(6:11:10)

Note that the causal signal does not contain an impulse at t ¼ 0 is assumed. If it does, then it adds a constant to its transform. Let hðtÞ ¼ KdðtÞ þ h1 ðtÞ, where h1 ðtÞ does not have an impulse at t ¼ 0: The impulse at t ¼ 0 appears in the transform and K ¼ lim HðjoÞ ¼ Re½Hðj1Þ: o!1

(6:11:11)

6.12 More on Filters: Source and Load Impedances

If there is an impulse at t ¼ 0, the real and the imaginary parts of the transform of the causal signal are related by the following relations in terms of Hilbert transforms: Im½HðjoÞ ¼

1 p

Z1 1

Im½HðjoÞ da; ðo aÞ

1 Re½HðjoÞ ¼Re½Hðj1Þ þ p

Z1 1

ImðHðjoÞÞ da: ðo aÞ (6:11:12)

Notes: From these two equations we note that the real and the imaginary parts of a realizable transfer function HðjoÞ are tied together by the Hilbert transform. This implies that HðjoÞ can be found from its real part alone, referred to as the real-part sufficiency. A physically realizable transfer function can also be found from its magnitude spectrum alone, which is referred to as a minimum phase transfer function and was briefly mentioned in Section 6.7. A linear system with a transfer function HðsÞ has no zeros or poles on the right halfs plane is called a minimum phase system. The relations between amplitude and phase responses of causal functions are referred to as Bode relations. Detailed study of these is beyond the scope here. For a discussion on this topic and other relations, see Bode (1945). Finding an impedance function ZðsÞ from Re½ZðjoÞ has been investigated by many authors, see Weinberg (1962). Here, finding the minimum phase transfer function HðsÞ from the given amplitude spectrum jHðjoÞj is of interest. The following gives a simpler procedure compared to the above & results.

229

Nðo2 Þ or jHðjoÞj2 ¼HðsÞHðsÞs¼jo ¼ K2 Dðo2 Þ Nðo2 Þ HðsÞHðsÞ ¼ K2 j 2 2 : (6:11:13) Dðo2 Þ o ¼s Poles and zeros of the function HðsÞHðsÞ have quadrantal symmetry and mirror symmetry about the jo axis giving a choice in selecting the poles and zeros of HðsÞ. 1. Choose only the left half plane roots of Dðo2 Þjo2 ¼s2 . This gives DðsÞ. 2. To have a minimum phase system, choose only the left half plane roots of NðsÞNðsÞ. Obviously there are other choices and those do not result in minimum phase functions. 3. Select a value K40 to match the value of the amplitude function jHðjoÞj at a desirable frequency. Example 6.11.1 Find the minimum phase stable transfer function HðsÞ given the amplitude-squared spectrum below. jHðjoÞj2 ¼

9ðo2 þ 4Þ : þ 10o2 þ 9Þ

ðo4

(6:11:14)

Solution: Substituting o2 ¼ s2 , and using the above procedure, results in 9ð4 s2 Þ ¼ HðsÞHðsÞ ð9 10s2 þ s4 Þ ðs þ 2Þ ðs þ 2Þ ¼K2 ðs þ 1Þðs þ 3Þ ðs þ 1Þðs þ 3Þ

jHðjoÞj2 jo2 ¼s2 ¼

) HðsÞ ¼ K

ðs þ 2Þ ;K ¼ ðs þ 1Þðs þ 3Þ

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ jHðjoÞj2 jo¼0 ¼ 3: &

6.11.2 Amplitude Spectrum jHðj!Þj to a Minimum Phase Function HðsÞ 6.12 More on Filters: Source and Load Given ðjHðjoÞjÞ find the minimum phase function Impedances 2

HðsÞ ¼ KNðsÞ=DðsÞ that is stable. Starting with ðjHðjoÞjÞ2 , we have 2

jHðjoÞj ¼HðjoÞH ðjoÞ ¼HðjoÞHðjoÞ ¼ K2 Nðo2 Þ=Dðo2 Þ

In this section simple passive analog filters are considered. In the next chapter, the design of various types of filters starting from the specifications to the synthesis using passive and active elements will be

230

considered. In the simple examples considered so far we assumed only one resistor and one inductor (or capacitor) in the filter circuit. The filter problem is illustrated in Fig. 6.12.1a, where we have three boxes, one represents a source, second one represents a filter, and the third one represents the load. Using the Thevenin’s equivalent circuit, we can replace the boxes represented by the source by the source plus the source impedance and the box representing the load by the load impedance. This is shown in Fig. 6.12.1b. The source and load impedances are generally assumed to be resistive in the frequency range of interest. This is a standard assumption in most filter design problems as we are operating in a small range of filter frequencies. In stead of Thevenin’s equivalent circuit we could use the Norton’s equivalent circuit shown in Fig. 6.12.1c. That is,

6 Systems and Circuits

replace the series circuit consisting of source and source resistance in Fig. 6.12.1b by a current source in parallel with the source impedance. This procedure allows the designer to separate the work associated with the filters from any designs associated with the left of the filter, i.e., the source and to the right of the filter, i.e., the load. We might also add that the source box and the load box may include several parts and the filter designer does not have to worry about those parts. In a later chapter when we consider two-port circuit analysis we will come back to this. For now let us consider a simple example illustrating the effect of the load. In the following we will derive the transfer functions in the Laplace transform domain. The frequency responses can be derived by replacing s ¼ jo in the transfer functions.

(a)

(b)

Fig. 6.12.1 (a) Filter with source and load resistors, (b) filter using Thevenin’s source equivalent circuit, and (c) filter using source Norton’s equivalent circuit

(c)

6.12 More on Filters: Source and Load Impedances

231

6.12.1 Simple Low-Pass Filters Example 6.12.1 Consider the RC circuit shown in Fig. 6.12.2 with the source and the load resistors. Derive the transfer function and sketch the amplitude characteristic function for the two cases. a: RL ¼ 1 and b: RL ¼ Rs . Solution: The transfer functions are given by

YðsÞ Z2 ¼HLP ðsÞ¼ ; XðsÞ a Z1 þZ2 RL =Cs RL ¼ ; Z1 ¼Rs (6:12:1a) Z2 ¼ RL þð1=CsÞ 1þRL Cs

YðsÞ RL =Rs RL C ¼ HLP ðsÞ ¼ XðsÞ b s þ ½ðRs þ RL Þ=Rs RL C K Rs þ RL : ; K ¼ 1=Rs C; oc ¼ ¼ Rs RL C s þ oc (6:12:1b) In case a., the load resistance is infinite, i.e., the circuit is not loaded. In case b. Rs ¼ RL . For the two cases the corresponding transfer functions are ð1=Rs CÞ ; ðs þ ð1=Rs CÞÞ ð1=Rs CÞ : b: HLP;Rs ¼RL ðsÞ ¼ s þ ð2=Rs CÞ

pﬃﬃﬃ jYðjoÞjo¼o3dB ¼ ð1= 2ÞjXðjoÞjo¼o3dB :

&

Notes: For simplicity, generic functions x(t) for the input and the output voltage yðtÞ are used. Usually, vi ðtÞ ðor vs ðtÞÞ for the input and v0 ðtÞ for the output & are common.

6.12.2 Simple High-Pass Filters In the ideal low- and high-pass filter cases shown in Fig. 6.10.7a, and b, it can be see that HLp ðoÞ ¼ jHHP ðjoÞjo¼1 ¼ 1 and o¼0 HHp ðjoÞ ¼ HLp ðjoÞo¼1 ¼ 0: o¼0

(6:12:3)

In addition, the amplitudes of these functions transition at the frequency o ¼ oc and the change in the amplitudes are as follows:

a: HLP;RL!1 ðsÞ ¼

(6:12:2)

In both cases, the gain constant is the same. However, the cutoff frequency is increased in the case of a load resistance. Note that the peak value of the amplitude response function in case b. is (1/2). So, the 3 dB frequency corresponds to the valuepof ﬃﬃﬃ the magnitude of the function equal to ð1=2Þð1= 2Þ. At o ¼ 0 the filter circuit is transparent; at o ¼ 1

Fig. 6.12.2 Example 6.12.1

there is no signal transmission; in between these frequencies, the output signal amplitude attenuation is determined by the equation jYðjoÞj ¼ jHðjoÞjjXðjoÞj. At the 3 dB frequency, o3dB

1ðlow-passÞ ! 0 ðhigh-passÞ or 0 ðhigh-passÞ ! 1 ðlow-passÞ: A logical conclusion is that o ! ð1=oÞ ði:e:; s ! ð1=sÞÞ provides a transformation that gives a way to find a high-pass filter function from a low-pass filter function. Noting that the impedance of an inductor is ðjoLÞ and the impedance of a capacitor is ð1=joCÞ, a high-pass filter can be obtained from a

232

6 Systems and Circuits

low-pass filter by replacing a capacitor by an inductor and an inductor by a capacitor. Since the impedance of the resistor R is independent of frequency, no change is necessary in the case of resistors. Using the RC and the RL low-pass circuits studied earlier in the low-pass case, we have two simple high-pass filters shown in Fig. 6.12.3a,b, one is a RL and the other one is a RC circuit. The filter is a low-pass if the inductor is in the series arm and the capacitor in the shunt arm. Similarly the filter acts as a highpass filter if the capacitor is in the series arm and the inductor is in the shunt arm. The transfer functions corresponding to the two circuits in Fig. 6.12.3 are

joj HHp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; b o2 þ ðR=LÞ2 ﬀHHpb ðoÞ ¼ 900 tan1 ðoL=RÞ:

(6:12:5b)

The amplitude and phase responses are shown in Fig. 6.12.4 for o40: The maximum value of the amplitude response is 1 or 0 dB. The 3 dB frequencies can be computed bypequating the amplitude ﬃﬃﬃ response function to ð1= 2Þ and solving for o. That is, 1 1 joca j jocb j pﬃﬃﬃ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; pﬃﬃﬃ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : 2 o2 þ ðR=LÞ2 o2 þ ð1=RCÞ2 2 ca

s s HHpa ðsÞ ¼ ; HHpb ðsÞ ¼ s þ ð1=RCÞ s þ ðR=LÞ jo jo ;HHPb ðjoÞ¼ : HHPa ðjoÞ¼ joþð1=RCÞ joþðR=LÞ

cb

(6:12:6) ) oca ¼ ð1=RCÞ; ocb ¼ ðR=LÞ:

(6:12:7)

(6:12:4)

The amplitude and the phase response characteristics of these are as follows: jo j HHp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; a 2 2 o þ ð1=RCÞ ﬀHHpa ðoÞ ¼ 900 tan1 ðoRCÞ

(6:12:5a)

Fig. 6.12.3 Simple high passive filters (a) RC circuit, (b) RL circuit

Fig. 6.12.4 Simple high-pass filter responses: (a) amplitude and (b) phase

As in the low-pass case, given the cutoff frequency oc , one of the reactive component (inductor or capacitor) values can be solved by selecting the resistor value. The high-pass filter is transparent from the input to the output at infinite frequency and no signal transmission at zero frequency. Note the low and high-frequency behavior of the lowpass and the high-pass filter functions HLp ðsÞ (or HHp ðsÞ) at s ¼ 0 and s ¼ 1:

(a)

(a)

(b)

(b)

6.12 More on Filters: Source and Load Impedances

233

6.12.3 Simple Band-Pass Filters These filters pass a band of frequencies called the passband and attenuate or eliminate the frequencies outside the passband, called the stopband. The simplest band-pass filter has a second-order transfer function. The ideal band-pass filters have two cutoff (or 3 dB) frequencies olow and ohigh . These frequencies are defined as the frequencies for which the magnitude function is equal to pﬃﬃﬃ of the transfer maxð1= 2ÞHBp ðjoÞ. In addition to these, a new frequency referred as the center or the resonant frequency o0 ; is of interest. It is defined as the frequency at which the transfer function of the circuit HBp ðjoÞ is purely real. The center frequency is not in the middle of the passband. It is the geometric center of the pass-band edges. It is related to the 3 dB frequencies by pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ (6:12:8) o0 ¼ olow ohigh : The second parameter of interest is the 3 dB bandwidth given by b ¼ ohigh olow :

(6:12:9)

The third parameter, the quality factor is the ratio of the center frequency to the 3 dB bandwidth. It is given by Q¼

o0 : ðohigh olow Þ

(6:12:10)

A second-order function that has the band-pass characteristics is HBp ðsÞ ¼

s2

H0 as H0 ðo0 =QÞs ¼ 2 : þ as þ b s þ ðo0 =QÞs þ o20 (6:12:11)

Fig. 6.12.5 (a) Amplitude and (b) phase responses of a band-pass filter

(a)

For simplicity the gain constant is assumed to be H0 ¼ 1 in the following. The transfer function has a zero at the origin ðs ¼ 0Þ and at infinity ðs ¼ 1Þ indicating that the function goes to zero at o ¼ 0 and at o ¼ 1. For Q41=2, it has a pair of complex poles given by sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ o0 1 : (6:12:12) jo0 1 s1 ; s2 ¼ 2Q 4Q2 The corresponding transfer function, the amplitude, and phase responses are given by ðo0 =QÞo ðo20 o2 Þ þ jðo0 =QÞo 1

ih

i ¼h o0 o 1 þ jQ o0 o 1 þ jQ oo0 oo0

HBp ðjoÞ ¼

(6:12:13) sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

2ﬃ o o 0 ; 1 þ Q2 o0 o

o o0 1 : ﬀHBp ðjoÞ ¼ tan Q o0 o

HBp ðjoÞ ¼1=

(6:12:14)

The amplitude and the phase responses are sketched in Fig. 6.12.5 for positive values of o. From the amplitude response, the peak of the amplitude appears at the center frequency o0 . The peak magnitude is 1 at o ¼ o0 . The phase angle starts at ðp=2Þ, crosses the frequency axis at o ¼ o0 and it asymptotically reaches ðp=2Þ as o ! 1: Higher the value of Q is, the more peaked the amplitude response is and steeper the phase response is around o ¼ o0 . The 3 dB or half-power bandwidth can be determined by assuming olow 5 o0 and ohigh 4o0 . These frequencies can be computed from

(b)

234

6 Systems and Circuits

2 HBp ðjoÞ2 o¼o ;o ¼ 1!Q2 o o0 ¼1) l u 2 o0 o o0 o 2 1 ;o 2 oo0 o20 ¼0: ðo2 o20 Þ¼ Q Q There are four roots of this equation, two for positive and two for negative frequencies. sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2 o0 1 o0 þ 4o20 ; ou ; ol ¼ 2Q 2 Q s ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ﬃ

o0 1 o0 2 ou ; ol ¼ þ 4o20 : 2Q 2 Q

Fig. 6.12.6 A simple band-pass filter

the expressions for the center frequency, bandwidth, and the quality factor. Solution: The transfer functions are V0 ðsÞ ðR=LÞs ¼ 2 ; Vi ðsÞ s þ ðR=LÞs þ ð1=LCÞ ðR=LÞjo : (6:12:18) HBp ðjoÞ ¼ ½ð1=LCÞ o2 Þ þ joðR=LÞ

HBp ðsÞ ¼ Assuming that Q4ð1=2Þ, the positive roots are given by sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 o0 (6:12:15) ol ; ou ¼ o0 1 þ : 4Q2 2Q The 3 dB bandwidth ðB or bÞ and o0 are respectively given by b ¼ ðohigh olow Þ ¼ o0 =Q; pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ o20 ¼ ol ou ! o0 ¼ ol ou :

(6:12:16)

Clearly from the first equation in (6.12.16) the bandwidth is inversely proportional to the value of Q. That is, the bandwidth decreases as Q increases and vice versa. The filter is assumed to be narrowband if o0 is very large compared to the bandwidth of the filter, i.e., o0 B. As a rough measure, we assume the filter is a narrowband filter if Q ¼ o0 =b 10:

(6:12:17a)

For the narrowband case the edges of the passband and o0 are given below. ol ; ou ¼ o0 ðo0 =2QÞ:

(6:12:17b)

In this case, o0 is in the middle of the 3 dB frequencies and is the center frequency. Example 6.12.2 Consider the circuit shown in Fig. 6.12.6. Find the transfer function and derive

The corresponding amplitude and phase responses are given for positive o by ðR=LÞo HBp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ðð1=LCÞ o2 Þ2 þ ððR=LÞoÞ2 ﬀHBp ðjoÞ ¼90o tan1

ððR=LÞo : (6:12:19) ½ð1=LCÞ o2 Þ

From (6.12.19), by using (6.12.16) results in pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ o0 ¼ 1=LC; b ¼ Bandwidth ¼ R=L and pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ (6:12:20) Q ¼ o0 =B ¼ L=CR2 : There two equations had three unknowns. In the design of a second-order band-pass filter, use the following steps. From the given second-order function, find o0 and Q. Select one of the element values, say C, and then use the other two equations to solve for the element values R and L: These are L ¼ 1=o20 C; R ¼ bL. Note that the gain constant is assumed to be H0 ¼ 1. If the gain is higher than 1, then circuit needs amplification. So far we have not considered of having both source and load impedances in the band-pass case. Also, inductors are never ideal and can be modeled by a resistor in series with an ideal inductor shown in Fig. 6.12.7 resulting in the impedance of the nonideal inductor as R i þ sL. The new transfer function and the amplitude response are

6.12 More on Filters: Source and Load Impedances

235

Fig. 6.12.7 A simple band-pass filter with a nonideal inductor

HBp ðsÞ ¼

ðR=LÞs s2 þ ½ðR þ Ri Þ=Ls þ ð1=LCÞ

(6:12:21)

ðR=LÞjoj HBp ðjoÞ ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : ðð1=LCÞ o2 Þ2 þ oððR þ Ri ÞLÞ (6:12:22) The center frequency is the same as before. The maximum value of the amplitude response is now R=ðRi þ RÞ. Also the new bandwidth is ðR þ Ri Þ=L. The nonideal inductor reduces the peak value and increases the bandwidth of the filter response. Correspondingly, the Q value is reduced. Nonideal inductors make the amplitude response less peaked with an increase in the bandwidth, i.e., the amplitude response becomes lower and broader. A band-pass filter can be viewed as a cascaded lowpass and a high-filter combination with the cutoff frequency of the low-pass filter greater than the cutoff frequency of the high-pass filter, oc;LP 4 oc;HP . An obvious question is, can we obtain a band-pass filter function from a low-pass filter function? In the two simple low-pass circuits considered earlier, the RL circuit, with the replacement of the inductor by a series LC circuit results in the circuit studied above. By replacing the capacitor in the RC low-pass circuit by a parallel LC circuit results in a band-pass circuit shown in Fig. 6.12.7. The element values in the bandpass case need to be determined using the center

Fig. 6.12.8 (a) Low-pass and (b) band-pass filter circuits

(a)

frequency and the required bandwidths. The figure illustrates only the concepts here. In Chapter 7 we will come back to the frequency transformations that involve changing cutoff frequencies of filters, converting various filter specifications into a proto-type lowpass filter specification, finding the appropriate filter function, the appropriate frequency transformations, synthesizing the filter function and finally scaling the circuit to fit the given specifications. Figure 6.12.8a gives a low-pass circuit. A band-pass circuit can be obtained by replacing the capacitor by an LCcircuit as shown in Fig. 6.12.8b. Inductors tend to be bulky and nonideal. That is, they have a resistive part Rw , thus reducing the quality factor of such coils. Impedance of the inductor needs to be replaced from joL to Rw þ joL. The quality factor of the coil is given by Q ¼ ðoL=Rw Þ, a function of frequency. Designing inductors with high Q values is a difficult process. In addition, since the field of operation associated with coils is the magnetic field, there is coupling between different inductors in a circuit. This can be reduced by either shielding one inductor from another and/or by placing in a manner shown in Fig. 6.12.9 requiring more space on the circuit board.

6.12.4 Simple Band-Elimination or Band-Reject or Notch Filters The amplitude response of a band-reject filter has the shape of a notch and is used to remove a band of frequencies somewhere in the middle of the frequency band and pass the low and high frequencies outside this band. A second-order notch filter has a transfer function of the form HBe ðsÞ ¼

ðs2 þ 2bs þ o20 Þ ; s2 þ ðo0 =QÞs þ o20

b55

o0 2Q (6:12:23)

(b)

236

6 Systems and Circuits

The 3 dB frequencies are obtained by equating pﬃﬃﬃ jHðo3dB Þj ¼ 1= 2. The two frequencies are

Fig. 6.12.9 Placement of two inductors to reduce magnetic coupling

ol ; ou ¼ o0 ðo0 =2QÞ; b ¼ o0 =Q:

(6:12:27)

If b 6¼ 0, then Lim HBe ðsÞ ¼ 1 and Lim HBe ðsÞ ¼ 1: s!0

s!1

(6:12:24)

The second-order band-pass and the notch filter functions have the same denominator, see (6.12.11) and (6.11.23). Notch filter passes low and high frequencies of the input signal without much attenuation and attenuates (or eliminates) a band of frequencies in the middle. This can be seen by first computing the zeros of the transfer function. qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ z1 ; z2 ¼ b j o20 b2 ) z1 ; z2 b jo0 ðIn the case of o0 44bÞ: (6:12:25) A special and an interesting case is when b ¼ 0 and for this case 2 o o2 0 jHBe ðjoÞj ﬃ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; ðo20 o2 Þ2 þ ½ðo0 =QÞo2 1 ðo0 =QÞo : (6:11:26) ﬀHBe ðjoÞ ¼ tan ðo20 o2 Þ Note there is no output at the notch frequency o0 as jHBe ðjo0 Þj ¼ 0. The amplitude and phase responses of the notch filter are shown in Fig. 6.12.10 for o40. We can see that jHBe ðjoÞjo¼0 ¼ 1 and lim jHBe ðjoÞj ¼ 1: o!1

jHBe ðjoÞjo¼o0 ¼ Qb=o0 :

The attenuation will be significant at the notch frequency as o0 is usually large. Also, note the phase reversal at o ¼ o0 in the phase response. Notch filters are used wherever a narrowband of frequencies needs to be eliminated from a received signal. In any electronic device, 60 Hz undesired hum, is ever present and a notch filter can be used to remove this. There are many applications in the telephone industry. In a long-distance call, a single frequency is transmitted from the caller to the telephone office until the end of the dialing of the number. After the party answers, the tone signal ceases and billing of the call begins and it continues as long as the signal tone is absent until the call is complete. A different application is toll-free long-distance calls that are not billed. For these, the signal tone is transmitted to the telephone office for the entire period of the call. Since the signal is within the voice frequency band, it must be removed from the voice signal before being transmitted from the telephone office to the listener. A simple second-order notch filter could be used for such an application. Notch filters are used wherever a narrowband of frequencies need to be eliminated from a received signal. In any electronic device, 60 Hz, an undesired hum, is ever present and a notch filter can be used to remove this. There are many applications in the telephone industry.

Fig. 6.12.10 (a) |HBe ðjoÞ| and (b) ﬀHBe ðjoÞ

(a)

(6:12:28)

(b)

6.12 More on Filters: Source and Load Impedances

Example 6.12.3 Band-elimination filters can be derived from low-pass filters by replacing an inductor (capacitor) by a parallel LC ðseries LCÞ circuit. Consider the circuit in Fig. 6.12.11 with a nonideal inductor with the equivalent impedance of the coil equal to ðRw þ joLsÞ. Assuming the circuit is not loaded, the output current is zero. Derive the transfer function and show it corresponds to a band-elimination filter.

Fig. 6.12.11 A band-elimination filter with a nonideal inductor

Solution: The transfer function, its amplitude and phase responses are Rw þ Ls þ ð1=CsÞ ðR þ Rw Þ þ Ls þ ð1=CsÞ ð1 þ Rw s þ LCs2 Þ (6:12:29) ¼ ð1 þ LCs2 Þ þ ðR þ Rw ÞCs ðRw þ joL þ ð1=ðjoCÞÞ HBe ðjoÞ ¼ ½ðR þ Rw Þ þ joL þ ð1=ðjoCÞÞ ð1 o2 LCÞ þ joRw C (6:12:30) ¼ ð1 o2 LCÞ þ joðR þ Rw ÞC qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ð1 LCo2 Þ2 þ ðRw CoÞ2 jHBe ðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ð1 o2 LCÞ2 þ ðoðR þ Rw ÞCÞ2 HBe ðsÞ ¼

ﬀHBe ðjoÞ ¼ arctan½Rw Co=ð1 o2 LCÞÞ

(6:12:31a)

arctan½oðR þ Rw ÞC=ð1 o LCÞ: (6:12:31b) The transfer function has the band-elimination characteristics as lim jHBe ðjoÞj ¼ 1; lim jHBe ðjoÞj ¼ 1; and o!1

jHBe ðjoÞjo¼1=pﬃﬃﬃﬃﬃ LC ¼ Rw =ðR þ Rw Þ:

If the winding resistance Rw is small, then the peak value of the amplitude response is close to 1. Using (6.12.27) pﬃﬃﬃﬃﬃﬃﬃ and (6.12.28), we have 2b ¼ Rw =LC, o0 ¼ 1= LC and ðo0 =QÞ ¼ ðR þ Rw ÞC. The notch bandwidth is B ¼ ½o0 =Q ¼ ½ðR þ Rw ÞC. In the ideal case, a second-order notch filter can be obtained from a second-order band-pass filter by replacing the parallel (series) LC circuit by a series (parallel) LC circuit.& Passive filter designs use resistors, inductors (and transformers), and capacitors. Four ladder forms of low-pass, high-pass, band-pass, and band-elimination filter circuits with source and load resistances are shown in Fig. 6.12.12. Source with source resistance is identified by a circle and the load resistance by a square. Between the source and the load, a lossless circuit is inserted, i.e., lossless coupling between the source and load. Figure 6.12.12a: Low-pass filter: When o ¼ 0, inductors will be short and the capacitors will be open and the source is directly connected to the source and the output is v0 ¼ ½RL =ðRi þ RL Þvi . When o ¼ 1, inductors will be open and the capacitors will be short and the load is disconnected from the source and the output is zero. Figure 6.12.12b: High-pass filter: When o ¼ 0, inductors will be open and the capacitors will be short and the load is disconnected from the source and the output is zero. For o ¼ 1, inductors will be short and the capacitors will be open and the source is directly connected to the load and v0 ¼ ½RL =ðRi þ RL Þvi . Figure 6.12.12c: Band-pass filter: At o ¼ 0, inductors will be short and the capacitors will be open and there is no output. At o ¼ 1, there is no output either. At the center frequency o0 , if (Lsi Csi ¼ 1=o20 ), the series arm is short since Lsei s þ ð1=Csei sÞs2 ¼ð1=Lsei Csei Þ ¼ ðLsei Csi s2 þ 1Þ=Csei ss2 ¼ð1=L C Þ ¼ 0: sei

2

o!0

237

(6:12:31c)

sei

In a similar manner we can show that at the frequency o0 , the shunt arm is open since Lshi s= Lshi Cshi s2 þ 1 s2 ¼ð1=L C Þ ¼ 1: shi

shi

Figure 6.12.12d: In the band-elimination case, we can show that v0 ¼ ½RL =ðRi þ RL Þvi at o ¼ 0 and o ¼ 1. At o ¼ o0 , the output is zero. The four filters have the desired transfer characteristics values at o ¼ 0; 1 (and, in addition o ¼ o0 ; in

238

6 Systems and Circuits

(a)

(b)

(c)

(d) Fig. 6.12.12 Passive filters

the case of band-pass and band-elimination filters. The exact characteristics at other frequencies cannot be determined since the actual filter element values are not known. In the passive filter design, and in general in communication theory, power transfer between the source and the load is important.

6.12.5 Maximum Power Transfer If there is no filter in Fig. 6.12.12, i.e., the source resistance is connected directly to the load, the maximum power is available at the output provided Rs ¼ RL :

(6:12:32a)

6.13 Summary

239

This is the maximum power transfer theorem and can be proven by first expressing the power delivered to the load without the filter. That is, pL ¼ RL i2 ; i ¼

vs RL ¼ v2 : ðRs þ RL Þ ðRs þ RL Þ2 s

The transfer function, the amplitude, phase, and the group delay responses are

(6:12:32b) Taking the derivative of pL with respect to RL , then solving for RL results in (6.12.32a). If the filter is inserted between the source and the load, then there will be less power available at the load, which will vary with frequency. This is defined as insertion loss. The design using these concepts is beyond the scope here, see Weinberg (1962).

6.12.6 A Simple Delay Line Circuit In Section 6.4 we considered the properties of a delay function. In this part of the section we will consider a simple circuit that has a constant amplitude response and a phase response that can be adjusted. Consider the circuit shown in Fig. 6.12.13. Assuming the output current is equal to zero, the output voltage can be expressed by V0 ðsÞ ¼Vc ðsÞ Vx ðsÞ ¼ ¼

ð1=CsÞ 1 Vin ðsÞ Vin ðsÞ ðR þ ð1=CsÞÞ 2

1 RCS Vin ðsÞ: 2ðRCs þ 1Þ

jHðjoÞj ¼ 1=2; ﬀHðjoÞ ¼ 2 tan1 ðo=aÞ; dﬀHðjoÞ a ¼ 2 : (6:12:34b) Tg ðoÞ ¼ do a þ o2

(6:12:33)

1 ða sÞ 1 ða joÞ 1 ; HðjoÞ ¼ ;a ¼ 2 ða þ sÞ 2 ða þ joÞ RC jHðjoÞj ¼ 1=2; ﬀHðjoÞ ¼ 2 tan1 ðo=aÞ; (6:12:35a) HðsÞ ¼

Tg ðoÞ ¼

dﬀHðjoÞ a ¼ 2 : do a þ o2

(6:12:35b)

Since the amplitude response is (1/2), a constant for all frequencies, this function is referred to as an allpass function. All-pass filters are used in cascade with the filters to provide the overall phase of the filter delay line combination and have an approximate linear phase characteristics. Additional phase due to the all-pass circuit adds to the filter delay.

6.13 Summary In this chapter we have started with basics of systems analysis and circuits. The circuits considered are simple. Specific topics that were covered in this chapter are given below.

Linear systems and their properties Two-terminal components: resistors, inductors, capacitors, voltage, and current sources

Classification of systems based on linearity, time-invariance, and other concepts

Impulse response of a linear system and the output in terms of the convolution integral

Transfer functions along with examples of simple circuits

System stability concepts and Routh’s stability test Distortionless systems and distortion measures

Fig. 6.12.13 A simple delay line circuit

for nonlinear systems

The transfer function in the s domain and in the frequency domain, the corresponding magnitude and phase responses and the group delay function are respectively given by HðsÞ ¼

1 ða sÞ 1 ða joÞ 1 ; HðjoÞ ¼ ;a ¼ 2 ða þ sÞ 2 ða þ joÞ RC (6:12:34a)

Group delay and phase delay responses System bandwidth measures similar to signal bandwidth

Relations between real and imaginary parts of a Fourier transform of a causal function

Derivation of the minimum phase transfer function from a given magnitude function

Ideal low-pass, high-pass, band-pass, bandelimination filters along with delay lines

240

6 Systems and Circuits

replacing the capacitor by an inductor in Fig. 6.5.4b. Give the corresponding steady-state response.

Problems 6.2.1 Consider the systems described by the following input–output relations. In each case, determine whether the system satisfies the following: 1. memoryless, 2. causal, 3. stable, 4. linear, and 5. time invariant. a: yðtÞ ¼ xð1 tÞ; b: yðtÞ ¼ xðt=2Þ; c: yðtÞ ¼ sinðxðtÞÞ: 6.3.1 Show that the systems a: yðtÞ ¼ x2 ðtÞ; b: yðtÞ ¼ sgnðtÞ are not invertible. 6.3.2 Determine whether the system yðtÞ ¼ T½xðtÞ ¼ txðtÞ is 1. memoryless, 2. causal, 3. stable, 4. linear, and 5. time invariant. 6.3.3 Consider the amplitude modulated function (discussed in Chapter 10) yðtÞ ¼ Ac ½A þ mðtÞ cosðoc tÞ with cases a: A ¼ 0; b: A 6¼ 0. Determine whether the system described by this equation is 1. memoryless, 2. causal, 3. linear, 4. time invariant, and 5. BIBO stable. 6.3.4 Repeat Problem 6.3.3 assuming the output of the system is yðtÞ ¼ xðt tÞ. 6.3.5 Determine whether the system described by the following is a. linear or nonlinear, b. time invariant or time varying yðtÞ ¼

1 X

xðtÞdðt kts Þ ¼

k¼1

1 X

xðkts Þdðt kts Þ:

k¼1

6.3.6 Consider the system described by yðtÞ ¼ xðbtÞ. Determine for what values of b the system is a. causal, b. linear, and c. time invariant.

6.4.3 Determine the impulse responses of the ideal low-pass, high-pass, band-pass, and bandelimination filters defined in (6.10.14a, b, c, and d) and the ideal delay line function in (6.7.1). 6.4.4 Determine the system responses for the inputs a: x1 ðtÞ ¼ uðtÞ; b: x2 ðtÞ ¼ P½t :5 assuming the system impulse response is hðtÞ ¼ tet uðtÞ: 6.4.5 Determine the stability of the integrator and a differentiator given below. Find their impulse responses and then use BIBO stability condition to see their stability. yðtÞ ¼

Zt xðaÞda;

yðtÞ ¼

dxðtÞ : dt

1

6.4.6 Using very simple functions with the properties as identified to show the responses become unbounded. Consider the transfer functions HðsÞ. a: HðsÞ has a pole on the right half s-pane, b.HðsÞ has multiple poles on the imaginary axis, c. HðsÞ has poles on the imaginary axis and the input function has a pole at this location. Explain why the responses become unbounded with the aid of the inverse Laplace transforms. 6.4.7 Consider the transfer function given by TðsÞ ¼

1 K ; HðsÞ ¼ : 1 þ HðsÞ ð1 þ :1sÞða þ sÞ

6.4.1 In Fig. 6.4.1 we have considered a simple RL time-varying circuit. Consider the RC time-varying circuit shown in Fig. P6.4.1. Assume that the time constant is large enough to justify that the circuit can be used as an integrator in an approximate sense. Identify the approximation used in considering this circuit acts like an integrator.

a. Assuming a ¼ 1, use the Routh array to determine the range of K for which the system is stable. b. Repeat the problem in Part a. assuming K ¼ 1 and determine the range of K for which the system is stable. c. Using the Routh array to determine the range of K for which the system has only poles to the left of s ¼ 1 in the s-plane.

6.4.2 Apply the periodic pulse waveform shown in Fig. 6.5.4a to a simple RL circuit obtained by

6.4.8 Use the Routh array and factor the polynomial DðsÞ ¼ s4 þ s3 þ 2s2 þ s þ 1. 6.4.9 Use the Routh array to find the number of right half s-plane roots of the polynomial DðsÞ ¼ s4 þ s2 þ s þ 1.

Fig. P6.4.1 An RC time-varying circuit

6.5.1 Give an RC circuit that approximates a differentiator and sketch the circuit’s amplitude and phase responses.

Problems

241

6.5.2 The transfer function of a linear time-invariant system is HðsÞ ¼

YðsÞ ðs þ 1Þ : ¼ 2 XðsÞ ðs þ s þ 1Þ

Assuming xðtÞ ¼ cosðo0 tÞ, find the steady-state response yðtÞ of the system. 6.5.3 Find the response of the RL circuit shown in Fig. 6.6.3 for input pulses t ðT=2Þ ; T ¼ 2p; a: xa ðtÞ ¼ P T b: xb ðtÞ ¼ sinðtÞxa ðtÞ: 6.5.4 Consider a circuit that has the response yðtÞ ¼ xðtÞ xðt TÞ with the input xðtÞ: Give the system impulse response and the expressions HðjoÞ and jHðjoÞj. 6.5.5 Consider the following two differential equations with xðtÞ ¼ ejot : d 2 y L dy þ þ yðtÞ ¼ xðtÞ dt R dt 2 d y b: LC 2 þ yðtÞ ¼ xðtÞ: dt

a: LC

Derive the transfer functions HðjoÞ ¼ YðjoÞ=XðjoÞ in each case assuming xðtÞ ¼ ejot . 6.5.6 Determine in each case below if the system described by its impulse response is stable or realizable, or both. Explain your results. a:ha ðtÞ ¼ dðtÞ;

b:hb ðtÞ ¼ et uðtÞ;

c:hc ðtÞ ¼ dðtÞ et uðtÞ: 6.5.7 Consider the impulse response and the corresponding transfer function hLp ðtÞ ¼ et uðtÞ; L½hLp ðtÞ ¼ HLp ðsÞ ¼ 1=ðs þ 1Þ. What can you say about the deconvolution filter HR ðsÞ and hR ðtÞ ¼ L1 fHR ðsÞg? Is this function realizable? 6.6.1 Consider the following functions with DðsÞ ¼ ðs þ 1Þðs þ 4Þðs þ 5Þ. Classify these as minimum or mixed or maxi phase systems. Sketch their amplitude and phase responses. a: H1 ðsÞ ¼ ðs 3Þðs þ 2Þ=DðsÞ b: H2 ðsÞ ¼ ðs þ 3Þðs þ 2Þ=DðsÞ c: H3 ðsÞ ¼ ðs 3Þðs 2Þ=DðsÞ:

6.6.2 Sketch the amplitude and phase responses of pﬃﬃﬃﬃﬃ a: H1 ðsÞ ¼ 1=ðs2 þ 2s þ 1Þ; b: H2 ðsÞ ¼ 1=ðs2 þ 3s þ 3Þ: 6.6.3 What can you say about the group delays associated with an all-pass functions at o ¼ 0 and o ¼ 1 in Problem 6.6.2? Can you draw any general conclusions? 6.6.4 The second-order Butterworth function is given in Problem 6.6.2a. a. Find its impulse response and the corresponding step response. b. Find the 10–90% rise time. c. Find the expression for the group delay of this function. 6.6.5 Assume the following node equations of a 3 nodes plus a reference node is given by ½ðVA V1 ÞY1 þ VA Y3 þ ðVA V0 ÞY2 ¼ 0; ðV0 VA ÞY2 þ ðV0 V1 ÞY4 ¼ 0: Give a circuit that has these node equations. Derive the transfer function V0 =V1 assuming Y1 ¼ 1=sL1 ; Y2 ¼ 1=sL2; Y3 ¼ sC2 ; Y3 ¼ sC3 : 6.6.6 Find the impulse and step responses of the following transfer functions with a40. 1 as ; a40; 1 þ as a ; b: HLP ðsÞ ¼ 2 s þ as þ a s c: HBP ðsÞ ¼ 2 s þsþ1 s2 þ a2 ; d: HBe ðsÞ ¼ 2 s þ bs þ a2 s2 : e: HHP ðsÞ ¼ 2 s þsþ1 a: Hd ðsÞ ¼

Give the amplitude and phase responses of these filters. 6.7.1 Consider the impulse response of a system hðtÞ ¼ et uðtÞ. Derive the expressions for its group and the phase delays. Sketch these functions on the same plot. 6.8.1 Show that the noise bandwidth of a band-pass function with center frequency o0 is WN ¼

Z1 h 0

i jHðjoÞj2 =jHðjo0 Þj2 do:

242

6 Systems and Circuits

6.8.2 Determine the RMS and the equivalent bandwidth of HðjoÞH ðjoÞ ¼ L½o=W. 6.9.1 Sketch the input and the output line spectra of a nonlinear circuit described by its input–output assuming xðtÞ ¼ relationship yðtÞ ¼ x2 ðtÞ 2 cosð2pð60ÞtÞ þ sinð2pð60ÞtÞ. 6.9.2 Assuming yðtÞ ¼ xðtÞ þ x2 ðtÞ, sketch YðjoÞ by assuming XðjoÞ ¼ P½ðo þ o0 Þ=W þ P½ðo o0 Þ=W: 6.9.3 Assume the input is xðtÞ ¼ cosðo0 tÞ to a nonlinear system described by yðtÞ ¼ xðtÞ þ :1x2 ðtÞ determine the second-order distortion term. 6.9.4 Consider the systems described by the following input–output relations. In each case determine whether the system is a. memoryless, b. causal, c. stable, d. linear or non-linear, and e. time-invariant system: a: yðtÞ ¼ xð1 tÞ; b: yðtÞ ¼ xðt=2Þ; c:yðtÞ ¼ sinðxðtÞÞ. 6.10.1a. Show the high-frequency slope of the n th order low-pass Butterworth function jHðjoÞj ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1=ð1 þ e2 o2n Þ is 20 n dB=decade:

b. Show that the first ð2n 1Þ derivatives of jHðjoÞj2 are zero at o ¼ 0. Use long division and then compare that to a power series and identify the

Fig. P6.12.1 Circuits to determine the transfer functions

corresponding derivatives. c. Show (6.10.10). d. Assuming n ¼ 2, determine the corresponding second-order Butterworth transfer function HðsÞ and find its impulse response. 6.10.2 Sketch normalized function qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ jHðjo=oc Þj ¼ 1=ð1 þ e2 ðo=oc Þ2n Þ for n ¼ 3 and 4: 6.11.1 Determine the minimum phase transfer function corresponding to the functions a: jHðjoÞj2 ¼ ½1=ð1 þ o2 Þ; b: jHðjoÞj2 ¼ ½1=ð1 þ o4 Þ: 6.12.1 Consider the circuits shown in Fig. P6.12.1. For each of these cases determine the corresponding transfer function. In the case of band-pass or bandelimination filters, determine the center frequencies. Give the 3 dB cutoff frequencies in each case. Give the expression for the quality factor of the circuits wherever appropriate. In addition, identify the type of the filter in each case. Simplify the expressions by assuming Rw ¼ 0. Sketch the amplitude responses for each of the cases and identify the important values. 6.12.2 Prove the maximum power transfer theorem. Sketch the power delivered to the resistor RL . Assume the source resistance is Rs .

Chapter 7

Approximations and Filter Circuits

7.1 Introduction In the first part of this chapter we will consider a graphical representation of the transfer function in terms of its frequency response HðjoÞ ¼ jHðjoÞjeﬀHðjoÞ . Bode diagrams or plots consist of two separate plots, the amplitude jHðjoÞj and the phase angle ﬀHðjoÞ, with respect to the frequency o on a logarithmic scale. These plots are named after Bode, in recognition of his pioneering work Bode (1945). Bode’s basic work was based upon approximate representation of amplitude and phase response plots of a communication system. Wide range of frequencies of interest in a communication system dictated the use of the logarithmic frequency scale. Bode plots use the asymptotic behavior of the amplitude and the phase responses of simple functions by straight-line segments and are then approximated by smooth plots with ease and accuracy. Bode plots can be created by using computer software, such as MATLAB. The topic is mature and can be found in most circuits, systems, and control books. For example, see Melsa and Schultz (1969), Lathi (1998), Close (1966), Nilsson and Riedel (1966), and many others. Filter approximations will be considered in the second part of this chapter. In Section 6.10, Butterworth approximation of an ideal low-pass filter amplitude response was introduced and the amplitude squared Butterworth function is jHBu ðjoÞj2 ¼

1 1 þ e2 ðo=oc Þ2n

:

(7:1:1)

The value of this function at o ¼ 0 is 1, at o ¼ 1, the function goes to zero, and, in between these

frequencies, the function decays. The low-pass filter passes frequencies between 0 and oc with small attenuation and blocks or attenuates the frequencies above oc , the cut-off frequency. In Section 6.11, we have considered deriving the transfer function HðsÞ from jHðjoÞj2 . In the next stage we are interested in coming up with a circuit that has the given transfer function. The circuit may consist of passive elements, such as resistors, inductors, capacitors, and transformers. Early filter designs were done exclusively with passive networks. Mathematics associated with passive network synthesis is elegant. There is very little leeway in the designs. See, for example, Weinberg (1962), and others for the passive filter limtations. Another problem of passive network synthesis is the use of inductors and transformers, as these are not ideal components in reality. Last part of the chapter deals with active filter synthesis using operational amplifiers, resistors, and capacitors. Active filter synthesis avoids the use of coils. Mathematical sophistication in active filter synthesis is much lower than passive filter synthesis. Active filter synthesis is based upon coming up with circuits with different topologies consisting of operational amplifiers, resistors, and capacitors. The circuit is then analyzed in terms of the R0 s and C0 s, assuming the operational amplifier is ideal. Comparing the derived and the given transfer function and equating the corresponding coefficients of s in the two transfer functions result in a set of equations with more unknowns than equations. As a result, we have infinite number of solutions for the component values. This gives a good deal of leeway for a circuit designer to optimize the circuits. One of the optimization criteria is

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_7, Ó Springer ScienceþBusiness Media, LLC 2010

243

244

7 Approximations and Filter Circuits

minimization of the sensitivity of a network with respect to changes in the component values. Introduction to sensitivity: In introducing sensitivity, Bode was concerned about the changes in the transfer function resulting from large changes in the element values in the transmission systems that included vacuum tubes. Even though we are in the era of integrated circuits, we are still interested in the effect of changes in the component values on the transfer function. This effect may be in the form of a shift in a pole frequency op or change in the quality factor Qor any other system parameter with respect to a component value. Pole sensitivity is defined as the per-unit change in the pole frequency, Dop =op , caused by a per-unit change in the desired component value Dx=x. The sensitivity of the parameter op with respect a component value xis defined by p So x ¼ lim

Dx!0

¼

x @op Dop =op =½Dx=x ¼ op @x

@ lnðop Þ : @ lnðxi Þ

Solution: We note that

Example 7.1.1 The transfer function (TF), HðsÞ ¼ V2 ðsÞ=V1 ðsÞ of the circuit in Fig. 7.1.1 is as follows:

Table 7.1.1 Formulae for computing sensitivities Y2 1 Y2 1 ¼ SY SY x x þ Sx

Y2 1 =Y2 1 SY ¼ SY x x Sx

Y SY xn ¼ ð1=nÞSx

Y SY x ¼ nSx

1 þY2 SY ¼ x

Y1 þY2

V2 ðsÞ 1=L1 C1 ¼ 2 V1 ðsÞ s þ ðR1 =L1 Þs þ ð1=L1 C1 Þ o20 ; ¼ 2 s þ ðo0 =QÞs þ o20 1 o0 L1 1 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ L1 =C1 : o0 ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; Q ¼ R1 R1 L1 C 1 (7:1:4)

(7:1:2)

@ðlnð1=YÞÞ @ð lnðYÞÞ ¼ ¼ SY ¼ x : (7:1:3) @ðln xÞ @ðlnðxÞÞ

Y Y Y1 SX1 þY2 Sx 2

HðsÞ ¼

Derive the sensitivities of the functions 1=2 1=2 o0 ¼ L1 C1 and Q with respect to the element values R1 ; L1 ; and C1 .

The parameter can be any that is important to the circuit’s function. We will assume the function of interest is Yi as a function of x. Formulas for sensitivities of simple functions can be seen by inspection. These are given in Table 7.1.1, where Yi ¼ Yi ðxÞ and xis a variable and c is a constant. From the sensitivity equation (7.1.2), we have S1=Y x

Fig. 7.1.1 Example 7.1.1

n

c YðxÞ

Sx

YðxÞ

¼ Sx

; Sxc ¼ 0

1=2

1=2

L1 1=2 C1

SLo10 ¼ SL1

¼

ðL1 C1 Þ1=2 @ðL1 Þ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ @L1 1= L1 C1 3=2

¼

3=2

ðL1 L1 2

Þ

1 ¼ : 2

(7:1:5)

Similarly, SCo10 ¼ ð1=2Þ; SRo10 ¼ 0; SLQ1 ¼ ð1=2Þ; SRQ1 ¼ 1:

(7:1:6)

0 Note that So R1 ¼ 0 since o0 is not a function of R. From (7.1.6), we have a 1% change in any one of the three component values in the RLC circuit in the example resulting in either (1/2)% or 0% change in o0 and (1/2)% or 1% change in Q.The sign of the values indicates whether the change is increasing or & decreasing. Transfer function of a circuit is a function of a set of parameters that are functions of the circuit components. For example, o0 is a function of L1 and C1 in Example 7.1.1. The change in o0 , Do0 can be approximated by using Taylor’s series written in terms of the variables xi in the form

7.1 Introduction

Do0 ¼

245

@o0 @o0 @o0 þ þ ::: þ @x1 @x2 @xn þ second and higher order terms:

MðsÞ ¼ G1 ðsÞEðsÞ ¼ G1 ðsÞðRðsÞ+BðsÞÞ; CðsÞ ¼ G2 ðsÞMðsÞ;

BðsÞ ¼ HðsÞCðsÞ

(7:1:7) CðsÞ ¼ G2 ðsÞG1 ðsÞ½RðsÞ+BðsÞ

See Tomovic and Vukobratovic (1972). If the change in the element values is assumed to be small, then the second-order and higher-order terms can be ignored. Do0 ﬃ

i¼1

Do0 ¼ o0

n X @o0 xi @xi @xi ¼ o0 : @xi @xi o0 xi i¼1

n X @o0

n X i¼1

0 So xi

Dxi ; xi

0 So xi

@o0 xi : ¼ @xi o0

CðsÞ½1 þ G2 ðsÞG1 ðsÞHðsÞ ¼ G1 ðsÞG2 ðsÞRðsÞ Transfer function:

(7:1:8) TðsÞ ¼

CðsÞ G1 ðsÞG2 ðsÞ : ¼ RðsÞ 1 þ G1 ðsÞG2 ðsÞHðsÞ

(7:1:10)

(7:1:9)

In Example 7.1.1 there are two important parameters, one is o0 and the other one is quality factor Q. Per-unit changes in Q in terms of its sensitivities can be expressed with respect to the parameters as well. In turn, the sensitivities of gain and phase of a transfer function in terms of the frequency o0 and Q can be determined. Block diagrams: In system control, system stability is one of the most important properties to be dealt with and closed-loop feedback control is basic to many systems. A simple feedback loop is shown in Fig. 7.1.2, where we have the Laplace transforms of the input signal with L½rðtÞ ¼ RðsÞ, error or actuating signal L½eðtÞ ¼ EðsÞ, control signal L½mðtÞ ¼ MðsÞ, controlled output L½cðtÞ ¼ CðsÞ, and primary feedback signal L½bðtÞ ¼ BðsÞ: The blocks identified by the transforms G1 ðsÞ; G2 ðsÞ; and HðsÞ represent control elements, plant or process, and feedback elements, respectively. The transfer function of the feedback system can be computed by writing the appropriate equations and solving for the output in terms of the input. These are as follows: LT

¼ G2 ðsÞG1 ðsÞ½RðsÞ+HðsÞCðsÞ

eðtÞ ¼ rðtÞ+bðtÞ ! RðsÞ+BðsÞ ¼ EðsÞ

The product GðsÞ ¼ G1 ðsÞG2 ðsÞ is the direct transfer function, HðsÞ is the feedback transfer function, the product GðsÞHðsÞ is the loop transfer function or the open-loop transfer function, and TðsÞ is the closed-loop transfer function. There are several books (for example, DiStefano et al. (1990).) that cover the block diagram algebra that gives simplifications in deriving the transfer functions. A useful transfer function that can be written in a special form is given by DiStfano et al. and the sensitivity of this function with respect to K is as follows: TðsÞ¼

A1 ðsÞþKA2 ðsÞ A3 ðsÞþKA4 ðsÞ ðK is independent of Ai ðsÞ;i¼1;2;3;4Þ:

) STK ¼

(7:1:11)

K½A2 ðsÞA3 ðsÞ A1 ðsÞA4 ðsÞ : (7:1:12) ½A3 ðsÞ þ KA4 ðsÞ½ðA1 ðsÞ þ KA2 ðsÞ

The transfer function can be expressed as a ratio of two polynomials in the form TðsÞ ¼

NðsÞ : DðsÞ

(7:1:13)

Example 7.1.2 Let a: T1 ðsÞ ¼ GðsÞHðsÞ ¼ ½K=ðs2 þ s þ 1Þ; b: T2 ðsÞ ¼ ½K=ðs2 þ s þ 1 þ KÞ. Determine the sensitivities of these functions to the parameter K.

Fig. 7.1.2 A simple feedback system

Solution: a. Using (7.1.11), we have A1 ðsÞ ¼ 0; A2 ðsÞ ¼ 1; A3 ðsÞ ¼ s2 þ s þ 1; A4 ðsÞ ¼ 0. Using T ðsÞ these in (7.1.12), it follows that SK1 ¼ 1 for all K.

246

7 Approximations and Filter Circuits

b. Again, using (7.1.11) and (7.1.12), we have

Consider M

P ð1 þ ð1=zm ÞsÞ

A1 ðsÞ ¼ 0; A2 ðsÞ ¼ 1;

HðsÞ ¼

A3 ðsÞ ¼ s2 þ s þ 1; A4 ðsÞ ¼ 1; T ðsÞ

SK2

Kðs2 þ s þ 1Þ 2 ðs þ s þ 1 þ KÞK 1 : ¼ 1 þ K=ðs2 þ s þ 1Þ

Ksd m¼1 N

(7:2:1b)

P ð1 þ ð1=pn ÞsÞ

n¼1

HðjoÞ ¼ HðsÞ s¼jo

¼

¼ KðjoÞd

(7:1:14)

ð1 þ jo=z1 Þð1 þ jo=z2 Þ:::ð1 þ jo=zM Þ ð1 þ jo=p1 Þð1 þ jo=p2 Þ:::ð1 þ jo=pN Þ

¼ jHðjoÞjejyðoÞ : T ðsÞ SK1

Sensitivity of the open-loop function is 1 for all values of K and the sensitivity of the closed-loop T ðsÞ function SK2 is a function of K and s. Using s ¼ jo, we can observe that for small values of o and K ¼ 1, T ðsÞ SK2 ﬃ :5. The feedback system is less sensitive than the open-loop system with respect to K. In Section 7.12 amplitude and phase sensitivities will & be considered. One of the main topics in this chapter is active filter synthesis. The first step in the active filter design is the analysis of a circuit with the appropriate topology. In Chapter 6 we considered computing the transfer function of a given circuit with twoterminal components using the Kirchhoff’s current and voltage laws and the component equations. Active filter circuits include multiple terminal components, including operational amplifiers (or op amps). These active devices are represented by controlled or dependent sources in the analysis. Kirchhoff’s laws, two-terminal component equations, and the controlled source representations of active devices provide a way to analyze circuits. A brief discussion on two-port representations of circuits is included by making use of the indefinite admittance matrix (Mitra, 1969). Other topics include scaling, frequency normalization, and adjustment of gain constants of the filter.

7.2 Bode Plots In this section we will study the basic concepts associated with a pictorial representation of a rational function, say HðsÞ or HðjoÞ, HðsÞ ¼ ½NðsÞ=DðsÞ; HðsÞjs¼jo ¼ jHðjoÞjejyðoÞ : (7:2:1a)

(7:2:1c)

Since the transfer is a ratio of real polynomials, the complex poles (or zeros) exist as complex-conjugate pairs and are usually simple. Multiple poles and zeros are possible and d is usually negative. In most cases, only a reasonable estimate of the system behavior and that to only at a very few frequencies is desired. The amplitude and phase responses at oi are given by M

P Zm

; Aðoi Þ ¼ jHðjoi Þj ¼ jKjjoi jd m¼1 N P Pn n¼1

Zm ¼ j1 þ joi =zm j; Pn ¼ j1 þ joi =pn j yðoi Þ ¼ﬀK þ dð90 Þ þ

M X

(7:2:1d)

tan1 ðoi =zm Þ

m¼1

N X

tan1 ðoi =pn Þ:

(7:2:1e)

n¼1

Although this approach is simple to see, finding these values and sketching them is time consuming. An alternate one is to obtain approximate sketches for the amplitude jHðjoÞj and the phase response yðoÞ using Bode plots using the following factors: 1. 2. 3. 4.

Constant term, K. Poles or zeros at the origin, s+k : Real poles or zeros, ðts þ 1Þ+k Complex-conjugate poles or zeros, ðt2 s2 þ 2xts þ 1Þ+k , where x is the damping ratio and 0 5x 5 1:

We need to study only simple poles and zeros, as the extensions to the multiple pole cases are simple. Note that log denotes base 10 and (ln) denotes base e.

7.2 Bode Plots

247

20 logjBðjoÞjN ¼ N logjBðjoÞj and ﬀ½BðjoÞN ¼ Nﬀ½BðjoÞ:

(7:2:2a)

logða þ jbÞ ¼ logja þ jbj þ j argða þ jbÞ

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ log a2 þ b2 þ j tan1 ðb=aÞ:

variable be defined by u ¼ logðoÞ or o ¼ 10u . The frequencies o1 and o2 are separated by an octave if o2 ¼ 2o1 and by a decade if o2 ¼ 10o1 . Note u2 u1 ¼ logðo2 Þ logðo1 Þ ¼ logðo2 =o1 Þ: (7:2:3)

(7:2:2b)

Note that the term t is used instead of the explicit poles or zeros and have combined the complex poles and their conjugates. Substituting s ¼ jo in the transfer function HðsÞ, we have HðjoÞ ¼ jHðjoÞj ﬀHðjoÞ or jHðjoÞjdB ¼ 20 logjHðjoÞj and fðoÞ ¼ ﬀHðjoÞ. We are only interested in sketching for positive frequencies. In addition, we will consider only poles that are on the negative half splane, including the imaginary axis and the zeros can be anywhere. In most applications, the poles and zeros are simple, with the exception that may include multiples at the origin. Since the log of a product is the sum of the corresponding logs of the terms, we can sketch the magnitude function by using the simple functions. The phase responses of the terms in the transfer function can be added to obtain the total phase response. The dB magnitude versus logðoÞ plot, i.e., logarithmic magnitude frequency response plot is called the Bode amplitude plot, and the phase angle versus logðoÞ is called Bode phase phase plot (or Bode diagrams). Logarithmic scale for the o-axis makes the sketches simple and allows sketches over a wider range of frequencies than the linear scale. Let the logarithmic frequency

Octaves ¼ log2 ðo2 =o1 Þ ¼ ½log10 ðo2 =o1 Þ= log10 ð2Þ; Decades ¼ log10 ðo2 =o1 Þ:

(7:2:4)

The amplitude and phase plots of HðjoÞ are considered using the four possible factors of a transfer function. Constant K: The logarithm of a constant is a constant with respect to o. The plot of 20 logðjKjÞ versus log(oÞ is a horizontal line. The phase angle is either 08 or –1808 depending upon whether the K is positive or negative. The factor (jo )N : 20 logjjojN ¼ 20 N logjoj; ﬀðjoÞN ¼ Nðp=2Þ:

(7:2:5)

Noting that log10 ð2Þ ¼ :3013, if o1 ¼ a and o2 ¼ 2a , the amplitude in (7.2.5) has increased by 6NdB/octave or 20NdB/decade. The function 20 N logjjoj plots as a straight line on the Bode plot and has a slope equal to 6 N dB/octave or 20N dB/decade. The slope of the line is positive (negative) depending on whether N is positive (negative). The magnitude and phase plots are shown in Fig. 7.2.1a,b for ð1=joÞ. It is simple to obtain the plots for multiple poles.

Bode Plot: Magnitude Plot of 1/jω

Bode Plot: Phase Plot of 1/jω

20

0

15

–45

5

Phase (deg)

Magnitude (dB)

10

0

–5

–90 –10 –15 –20 –1 10

0

10

Frequency (rad /sec)

Fig. 7.2.1 Amplitude and phase plots of (1/ðjoÞ)

10

1

–135 –1 10

0

10

Frequency (rad /sec)

1

10

248

7 Approximations and Filter Circuits

The factor ½1=ðjot þ 1Þ: Noting that A1 ðoÞ ¼ 20 logjð1 þ jotÞj ¼ 10 log½1 þ ðotÞ2 , we have for small and for large values of o, the amplitudes can be approximated by ðotÞ 1; AðoÞ 20 logð1Þ ¼ 0 dB; ðotÞ 1; AðoÞ 20 logðjotjÞ dB :

(7:2:6)

These are the asymptotes to the true curve corresponding to the very small and very large frequencies. The first asymptote is a horizontal line and the second asymptote is a straight line with a slope of –6 dB/octave or –20 dB/decade. The two asymptotes intersect at the corner frequency or the break frequency o ¼ 1=t. The actual value of the magnitudepfunction at this frequency is ﬃﬃﬃ equal to 20 logð 2ÞdB 3dB. It is simple to draw the asymptotic and the actual curves using the following guidelines: 1. The constant t is the break point (or the corner frequency) in the asymptotic plot. 2. From the break point, draw the two asymptotes, one with a zero slope toward the o small and the other one with a –6 dB/octave slope extending toward o ! 1. 3. At the break point, the true response is displaced by –3 dB. In addition, an octave below and above the break point, the true curve is separated by –1 dB. A sketch of the amplitude response using the table and the above guidelines is shown in Fig. 7.2.2a. Note the frequency is

plotted using the log scale. The phase angle of the term ½1=ð1 þ jotÞ is equal to f1 ðoÞ ¼ tan1 ðotÞ radians or ½57:3 tan1 ðotÞ degrees: It can be approximated using the power series expansion (Spiegel, 1966): tan1 ðotÞ ¼ ðotÞ ð1=3ÞðotÞ3 þ ð1=5ÞðotÞ5 :::; jotj41;

(7:2:7a)

tan1 ðotÞ ¼ +ðp=2Þ ½1=ðotÞ ð1=3Þð1=otÞ3 þ ð1=5Þð1=otÞ5 :::; þ if ot 1; if ot 1: (7:2:7b) tan1 ðotÞ ¼ p=4;

ot ¼ 1:

(7:2:7c)

Figure 7.2.2 gives the Bode amplitude and phase plots. The phase angle plot approaches 08 as ðotÞ ! 0 and – 908as ðotÞ ! 1. Noting (7.2.7c), we can see that the phase angle is –458 at the break frequency o ¼ 1=t. These two asymptotes can be connected by drawing a line from the 08 asymptote starting at one decade below the break frequency ð:1=tÞ with 08 phase and draw a line with a slope of 458/decade passing through – 458at the break frequency and continuing to –908 one decade above the break frequency ð10=tÞ. For the zeros, the amplitude and phase response sketches can be similarly drawn since only the signs need to be altered. Quadratic factors: The Bode plots corresponding to a pair of complex poles are usually given in terms of the damping factor x 1 by

Bode Plot: Magnitude Plot of 1/(jω + 1)

Bode Plot: Phase Plot of 1/(jω + 1)

0

0

–5

–15

–30 Phase (deg)

Magnitude (dB)

–10

–20 –25

–60

–30 –35 –40 –45 10–2

10–1

100 Frequency (rad /sec)

101

102

–90 10–2

10–1

100 Frequency (rad /sec)

Fig. 7.2.2 Bode amplitude 1=j1 þ joj and phase ﬀ1=ð1 þ joÞ plots, break frequency ¼ 1

101

102

7.2 Bode Plots

249

1 ½1 þ 2xts þ t2 s2 1 ; 0 x 1: (7:2:8) ¼ ½1 þ ðQp =op Þs þ ð1=o2p Þs2

H2 ðsÞ ¼

jH2 ðjoÞj2 ¼ 1=½ð1 o2 t2 Þ2 þ 4ðxtoÞ2 ; A2 ðoÞ ¼ 20 logðjH2 ðjoÞjÞ

(7:2:9a)

f2 ðoÞ ¼ tan1 2xot=ð1 o2 t2 Þ

(7:2:9b)

It is common in the filter designs to use the quality factor Qp and op in (7.2.8). The peak of the magnitude squared function can be found by taking the derivative of the denominator in (7.2.9a) with respect to o and equating to zero. That is,

The high frequency asymptote is a straight line with a slope of –12 dB/octave (–40 dB/ decade). 2. The break frequency, i.e., the intersection of the low-frequency and the high-frequency asymptotes is located at the frequency o ¼ ð1=tÞ, which can seen from the fact that 40 logðotÞ ¼ 0; o ¼ 1=t for all x. At the break frequency, we have 20 logjH2 ðjoÞjo¼1=t ¼ 20 logð2xÞ dB:

As an example at x ¼ :2, we have the value –7.958 dB and for x ¼ 1=2, the above equation reduces to 0 dB. A few values are given below for (7.2.13). x –20log(2x)dB

d½1=jH2 ðjoÞj2 ¼ 2ð1 o2 t2 Þð2ot2 Þ þ 2ð4x2 t2 oÞ do pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 2x2 : (7:2:10) ¼0!o¼ t pﬃﬃﬃ 1 2x2 40 or x5ð1= 2Þ ¼ :707 ðo is realÞ: (7:2:11) Asymptotic approximations – second-order case: H2 ðjoÞ ¼ 1=½1 þ j2xto t2 o2 1. For low and high frequencies we can write: ðotÞ 1; 20logjH2 ðjoÞj ¼ 0 dBðotÞ 1; 20logjH2 ðjoÞj 40 logðotÞ dB:

(7:2:12)

0 1

.05 20

.1 14

.2 8

.3 4.5

.4 2

.5 0

.6 –1.5

.707 –3

1 –6

Now consider the phase asymptotic plots of the second-order function from the f2 ðoÞ given in (7.2.9b). At the break frequency, f2 ðoÞ o¼ð1=tÞ ¼ tan1 ð1Þ ¼ 90 for all x:

(7:2:14)

The phase starts at 0 at low frequencies. At o ¼ ð:1=tÞ, it starts to decrease at a rate of –908/ decade. At o ¼ ð1=tÞ, it is –908. It continues to decrease and reaches–1808 as o ! 1. The amplitude and phase responses are plotted for a few values of x in Fig. 7.2.3. The phase response will

2

2

Bode Diagram: Magnitude Plot of 1/(1+ j2ξω-ω )

Bode Diagram: Phase Plot of 1/(1+ j2ξω-ω )

30

0 ξ = 0.05 ξ = 0.2 ξ = 0.4

ξ = 0.05

20

ξ = 0.2 10

ξ = 0.4

ξ = 0.6

–45

ξ=1

ξ = 0.6

0

Phase (deg)

Magnitude (dB)

(7:2:13)

ξ=1

–10

1 –20

–90

–135

–30 –40 –50 –1 10

–180 10

0

1

10

–1

10

Frequency (rad/sec)

Fig. 7.2.3 Bode plots jH2 ðjoÞj ¼ 1=½1 þ j2xto t2 o2 ; ﬀH2 ðjoÞ; 0 x 1

0

10

Frequency (rad/sec)

1

10

250

have a discontinuity going from 0 to –1808 at the break point corresponding to x ¼ 0. The above discussion is given in terms of poles. For the zeros of a transfer function, multiply the dB and the phase angle values by –1. Bode plots of the transfer function can be constructed by summing the log magnitudes and phase angle contributions of each pole and zero (or pairs of complex and their conjugates of poles and zeros). DiStefano III et al. (1990), Nise (1992), and others give systematic procedures for sketching the Bode plots with examples to illustrate the construction process. Thaler and Brown (1960) give a tabulation of typical control system transfer functions, with associated polar plots, Bode plots, and root locus plots. There are a few ways of computing the amplitude versus frequency on the Bode plot. Construct each factor separately and at selected values of o add the amplitudes. Then, sketch the amplitude function through these points. It is simpler to use the asymptotes. For this purpose we need to have the transfer function in factored form. Next arrange the transfer function with increased values of poles and zeros. Many of the control system transfer functions have poles at the origin with a multiplicity of l; l 0. Such a system is called a type l system. The system amplitude plot has a slope at low frequencies of 10 l dB/decade or (–6l dB/octave). This slope is maintained until the first corner frequency is reached. At the first corner frequency the slope is changed by +10 dB=decade for a first-order pole or zero. If it is a second-order function, then the slope is changed by +20 dB=decade. This procedure is used to sketch the asymptotic plot. We can construct the composite asymptote provided the exact location of the lowest frequency segment can be located. For l ¼ 0, the lowest-frequency asymptote with a constant gain K is 20 logðKÞ dB. For l ¼ 1, locate the point o ¼ K on the 0 dB axis. The lowest frequency asymptote passes through this point with a slopepof ﬃﬃﬃﬃ –10 dB/decade. For l ¼ 2, locate the point o ¼ K on the 0 dB axis. This frequency asymptote passes through this point with a slope of –20 dB/ decade. It is very rare to have more than a double pole at the origin. Noting that for oT 1, all the factors of the form ðjoT þ 1Þ reduce to 1 in this region. Then, we have

7 Approximations and Filter Circuits

20 logjHðjoÞj ﬃ 20 logðKÞ 20 log ðjol Þ : (7:2:15a) For example, for l ¼ 2, the above equation represents a straight line with a slope of –20 dB/decade and the corresponding intercept is determined by pﬃﬃﬃﬃ 20 logðKÞ 20 log ðjo2 Þ ¼ 0 ! o ¼ K: (7:2:15b) Example 7.2.1 Sketch the Bode plots for the following transfer function: HðsÞ¼ HðjoÞ¼

10ð1þsÞ s2 ð1þðs=4Þþðs=4Þ2 Þ

10ð1þjoÞ 2

ðjoÞ ð1þð1=4Þjoðð1=4ÞÞ:25oÞ2 Þ

; : (7:2:16)

Solution: The corresponding amplitude in terms of dB and the phase responses are 20 logjHðjoÞj ¼ 20 logð10Þ þ 20logjð1 þ joÞj 40logjjoj 20 log 1=½ð1 þ jðo=4Þ ðo=4Þ2 ; (7:2:17)

ﬀHðjoÞ ¼ﬀð1 þ joÞ þ ﬀð1=joÞ2 þ ﬀ½1=ð1 þ jo=4 ðo=4Þ2 Þ:

(7:2:18)

The asymptotic amplitude plot is obtained by adding the asymptotic plots of each of the terms. The first term on the right in (7.2.17) is equal to 20 dB for all values of o. The third term corresponds to a double pole at the origin and the asymptote goes through the corner frequency o ¼ 1 with a slope of –40 dB/decade. The high-frequency asymptote of the second term in (7.2.17) starts at the corner frequency o ¼ 1 and has a slope of 20 dB/decade. The fourth term corresponds to a pair of complex poles. Noting t ¼ 1=4 and 2xt ¼ :25 (x ¼ .5, damping factor). The high-frequency asymptote of the complex pair of poles starts at the corner frequency o ¼ 4 with a slope of –40 dB/decade. All of these are sketched in Fig. 7.2.4 using MATLAB software. Before we obtain the composite amplitude asymptotic Bode plot of the transfer function, we need to locate the lowest frequency asymptote. Using (7.2.15b), the corresponding pﬃﬃﬃﬃ pﬃﬃﬃﬃﬃ asymptote is a line through the point K ¼ 10 at a slope of –40 dB/decade on the 0 dB axis. We are now ready to sketch the asymptotic Bode magnitude plot of the

7.2 Bode Plots

251

Fig. 7.2.4 (a) Bode amplitude and (b) Bode phase plots of individual factors

Bode Diagram

Magnitude (dB)

100

50

0

–50

10 1+s 1/s2 2

1/(1 + 0.25s – (0.25s) )

(a)

–100 90

Phase (deg)

45 0 –45 –90 –135 –180 10–2

(b) 10–1

100

101

102

Frequency (rad/sec)

transfer function by adding the asymptotic plots of each of the four terms. Start at a low frequency say o ¼ :1, follow the low-frequency asymptote that pﬃﬃﬃﬃﬃ passes through o ¼ 10 on the 0 dB axis to the first corner frequency o ¼ 1 corresponding to the (jo þ 1Þ factor. Since this is a numerator factor, the next asymptote changes from –40 dB/ decade to –20 dB/decade and continues to the next corner frequency located at o ¼ 4. The next asymptote takes into consideration of the second-order factor in the denominator. The last high-frequency asymptote starts at o ¼ 4 with a change in slope to –60 dB/decade. The individual

amplitude and phase plots of the transfer function are shown in Fig. 7.2.4a and b. We can obtain the actual amplitude plot by correcting for the errors in the asymptotic plots and the true functions. An easy way to do is find the amplitudes at the corner frequencies and then sketch the function through the computed values. Bode plots are only sketches. If accurate plots are desired, software packages, such as MATLAB, need to be used to obtain the desired results. Phase approximations can be used for the phase angle asymptotic plots. To get a sketch, it is easier to make use of arctangent function to obtain the

MATLAB code for Example 7.2.1 %Plot individual terms (Figure 7.2.4 Example 7.2.1) sys1=tf (10,1 ); sys2=tf ([1 1],1); sys3=tf(1,[1 0 0]); sys4=tf(1,[-0.25 0.25 1]); bode(sys1,’k’,sys2,’k– –’,sys3,’k-.’,sys4,’k:’,{0.01,100}) legend(’10’,’1+s’,’1/s^2’,’1/(1+0.25s-(0.25s)^2)’) %Plot the whole Bode plot (Figure 7.2.5, Example 7.2.1) num=10*[1 1]; den=[-(1/4) (1/4) 1 0 0]; w=logspace(-2,2,100); bode(num,den,w) grid

252

7 Approximations and Filter Circuits

Fig. 7.2.5 Bode amplitude and phase plots of the composite function

Bode Diagram 150

Magnitude (dB)

100

50

0

–50

Phase (deg)

–100 –90

–135

–180 –2 10

10

phase angles at some important frequencies, such as at corner frequencies, and sketch the function using these values. The amplitude and phase plots for the composite are given in Fig. 7.2.5a,b using the MATLAB code given below.

–1

0

10 Frequency (rad/sec)

10

1

10

2

FPM ¼ 1800 þ ﬀH0 ðjoc Þ with Hðjoc Þ ¼ 1; oc ¼ gain crossover frequency:

(7:2:20)

They are measures on how closely H0 ðjoÞ approaches a magnitude of unity and a phase of – MATLAB code for Example 7.2.2

7.2.1 Gain and Phase Margins We would like to consider two important topics that are used in the stability analysis of feedback control systems. Our discussion will be brief. The characteristic polynomial of a feedback control system is DðsÞ ¼ 1 þ H0 ðsÞ ¼ 0 with H0 ðsÞ ¼ G1 ðsÞG2 ðsÞ HðsÞ (see 7.1.10). Practicing engineers use the gain margin (GM ) and the phase margin (FM ), see, for example, Melsa and Schultz (1969). Graphical analysis is more appealing to engineers than analytical analysis. Gain and phase margins are measures of relative stability of the feedback control system. These are defined at the phase and margin crossover frequencies op and oc . 1 with ﬀHðjop Þ ¼ 1800 ; jH0 ðjop Þj op ¼ phase crossover frequency;

GM ¼

(7:2:19)

n=5 ;d=[1 3 4 2]; w=logspace(–2,2,100), [mag, phase]=bode (n,d,w);margin(mag, phase,w)

1808 quantifying the relative stability of the system. They can be read using the Bode plots. Negative phase margin implies instability. Most engineers use the criteria that a phase margin of 308 and a gain margin of 6 dB are safe margins. Analytical computation of gain and phase margins may not be possible since it requires factoring polynomials. Example 7.2.2 Using the following methods obtain the gain and phase margins: a. Analytical methods and b. MATLAB for the following function: 5 ) HðjoÞ H0 ðsÞ ¼ 3 2 s þ 3s þ 4s þ 2 5 ¼ : (7:2:21) ð2 3o2 Þ þ joð4 o2 Þ

7.2 Bode Plots

253

Solution: a. Equate the imaginary part to zero, i.e., ImðH0 ðjoÞÞ ¼ 0. Solving for op and then evaluating the real part at this frequency, we have ImðH0 ðjoÞÞ ¼ 0; op ¼ 2 ! ReðH0 ðjop ÞÞ ¼ 1=2:

We can use MATLAB command roots ([1,0,1,0,4,0,–21]) and obtain the real positive root of the polynomial given by oc ¼ 1:4315 resulting in

(7:2:22)

We can increase the gain by 2 before the real part becomes –1. The gain margin in dB is GM ¼ 20 logð2Þ ﬃ 6 dB:

ﬀH0 ðjoÞjo¼oc ¼1:4315 ¼ ﬀ 5=ð2 3o2 Þ þ joð4 o2 Þ o¼o ¼1:4315 1460 : c

(7:2:23)

We could also solve for op by noting that the characteristic polynomial

The phase margin, the difference between this angle and –1808, is

DðsÞ ¼ 1 þ aH0 ðsÞ ) 0 ! s3 þ 3s2 þ 4s þ 12 ¼ ðs2 þ 4Þðs þ 3Þ ¼ 0

FM 180 146 ¼ 34 :

has imaginary roots given by s ¼ + j2. See the discussion on Routh table Chapter 6. For phase margin, we need to equate jH0 ðjoÞj ¼ 0 and solve for o ¼ oc . This requires software, such as MATLAB. These result in 25 ¼ð2 3o2 Þ2 þ o2 ð4 o2 Þ2 ! o6 þ o4 þ 4o2 21 ¼ 0:

(7:2:24)

b. Analytical computation may not be possible and computational tools, such as MATAB, are good to use. For this example the code is given below. MATLAB Bode plots are given in Fig. 7.2.6. The gain and the phase margins are & shown.

Bode Diagram Gm = 6.03 dB (at 2 rad /sec), Pm = 34.1 deg (at 1.43 rad /sec) 20

Magnitude (dB)

0 –20 –40 –60 –80 –100 –120 0

Phase (deg)

–45 –90 –135 –180 –225

Fig. 7.2.6 Illustration of gain and phase margins, Example 7.2.2

–270 10–2

10–1

100 Frequency (rad/sec)

101

102

254

7 Approximations and Filter Circuits

7.3 Classical Analog Filter Functions 7.3.1 Amplitude-Based Design

The ideal low-pass filter function was defined in Chapter 6 and is HLp ðjoÞ ¼ P½o=2oc ejot0 ;

Noting jHðjoÞj is even, start with

oc ¼ 2pfc ; cut - off frequency:

jHðjoÞj2 ¼ HðjoÞHðjoÞ M P

¼

al o2 l

l¼0

1þ

N P

; N M:

(7:3:1)

bl o2 l

l¼1

The goal is to determine the coefficients al and bl so that (7.3.1) satisfies a given set of specifications on the amplitude response. Letting o ! s=j, we have the product HðsÞHðsÞ. The minimum phase transfer function is obtained by assigning the left halfplane poles and zeros of HðsÞHðsÞ to HðsÞ. In this section we will consider Butterworth, Chebyshev I, and Chebyshev II filter functions.

The ideal low-pass filter response in (7.3.2) is not physically realizable, see Section 6.10.1. The next best thing is approximate an ideal filter response. We have seen that a simple RC circuit approximates a low-pass filter. There are four types of filters identified by 1. 2. 3. 4.

HLp ðjoÞ, low-pass function HHp ðjoÞ, High-pass function HBp ðjoÞ, Band-pass function HBe ðjoÞ, Band-elimination function

Low-pass filter specifications: 9 8 < Pass band : 0 joj oc ; 1 2 HLp ðjoÞ 2 1 or XdB 20 log HLp ðjoÞ 0dB = 1þe : : Stop band : joj o ; H ðjoÞ 2 ð1=A2 Þ; or 20 log H ðjoÞ YdB; o 5o ; r Lp Lp c r

High-pass filter specifications: 8 < Pass band : oc joj 1; : Stop band : 0 joj o ; r

(7:3:2)

(7:3:3a)

9 2 HHp ðjoÞ 1 or X dB 20 log HHp ðjoÞ 0 dB = : (7:3:3b) HHp ðjoÞ 2 ð1=A2 Þ; or 20 log HHp ðjoÞ Y dB; or 5oc : ;

1 1þe2

Band-pass filter specifications: 9 8 1 HBp ðjoÞ 2 1 or X dB 20 log HBp ðjoÞ 0 dB > > o o ; Pass band0 :o j j l h 2 > > 1þe = < 2 2 Stop bands :0 joj o1 and o2 joj 1; HBp ðjoÞ ð1=A Þ; > > > > ; : or 20 log HBp ðjoÞ Y dB; o1 5ol ; oh 5o2 : Band elimination filter specifications: 9 8 2 2 > > Pass bands : 0 o o ; o o 1; ð1=ð1 þ e ÞÞ H ðjoÞ 1 j j j j l h Be > > > > > > = < or X dB 20 logjHBe ðjoÞj 0 dB : 2 > > > > Stopband : o1 joj o2 ; 0 jHBe ðjoÞj ð1=ð1 þ e2 ÞÞ; > > > > ; : or YdB 20 logjHBe ðjoÞj;ol o1 ; oh o2 :

(7:3:3c)

(7:3:3d)

7.3 Classical Analog Filter Functions

255

Fig. 7.3.1 Analog filter amplitude specifications: (a) low pass, (b) high pass, (c) band pass, and (d) band elimination

(a)

(b)

(c)

(d)

These are illustrated in Fig. 7.3.1a,b,c,d in terms of the squares of amplitudes and in dB scale on all the figures. For simplicity the responses are assumed to be smooth.

7.3.2 Butterworth Approximations A simple RC circuit was considered in Chapter 6 with a transfer function (see 6.5.16b) HðjoÞ ¼

1 ; jHðjoÞj ¼ HðsÞ s¼jo joRC þ 1

1 ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; oc ¼ 1=RC: ð1 þ ðo=oc Þ2

(7:3:4)

It has a single parameter oc controling the amplitude response. Assuming the input and the output transforms as XðjoÞ and YðjoÞ, we have YðjoÞ ¼ HðjoÞXðjoÞ. This implies jYðjoÞj ¼ jHðjoÞjjXðjoÞj. The filter acts like a gate in the sense that low frequencies are passed with very little attenuation and the high frequencies are attenuated significantly. The shape of the amplitude response function jHðjoÞj controls what frequencies are allowed through and what frequencies are attenuated or eliminated. To be more specific, in the low-pass filter design, we assume we have three bands defined by Pass band : 0 o oc ; transition band: oc 5o5or ; stop band : or o 1:

(7:3:5)

The frequencies 0 o oc in the input will be allowed to pass through the low-pass filter without much attenuation and therefore we call this band of frequencies as the pass band. The band of frequencies or o 1 in the input signal will be attenuated by the low-pass filter significantly and this band is called the stop band. In between these two bands the filter amplitude response will have to be tapered or gradual and the input frequencies are attenuated gradually. A popular function that can act as a low-pass filter that satisfies the above criterion is a Butterworth function defined by 1 jHBu ðjoÞj ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ : 1 þ e2 ðo=oc Þ2n

(7:3:6)

The subscript Bu on H is usually not shown and seen from the context. The e term controls how far the filter amplitude characteristic will go down to when o ¼ oc from 1 at o ¼ 0 and the value of n controls how fast the magnitude characteristic attenuates in the stop-band region. We have shown that the ripple amplitude in the band ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ppass having the value for small e is 1 ð1= ð1 þ e2 Þ e2 =2 (see 6.10.10). The pass-band and stop-band specifications for the low-pass filter are given in (7.3.3a) and the specifications are identified in Fig. 7.3.1a. It is common to specify these two in terms of the dB scale, as identified in this figure. Using the edges of the frequency bands oc and or , we can write

256

7 Approximations and Filter Circuits

Since js2nþ1 j ¼ ðoc =e1=n Þ, poles of the Butterworth function are equally spaced on a circle of radius (oc =ðe1=n ÞÞ on the splane. Selecting the poles on the left half of the s plane, the transfer function is

jHBu ðjoÞj2o¼oc ¼ 10 logð1=ð1 þ e2 ÞÞ ¼ X dB ) e pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¼ 10ðx=10Þ 1:

(7:3:7) HBu ðsÞ ¼

jHBu ðjoÞj2o¼or ¼10logð1=A2 Þ ¼ YdB:10log

ð Note HBu ð0Þ ¼ 1Þ:

1 1 þ e2 ðor =oc Þ2n

¼ YdB ) ðnÞinteger

ðnÞinteger

ﬃ

log½ð10:1Y 1Þ=ð10:1X 1Þ : (7:3:8) 2logðor =oc Þ log

(7:3:14)

Example 7.3.1 Compute the values of e and n and derive Butterworth transfer function for the specifications given in Fig. 7.3.2 with X ¼ 2 dB and Y ¼ 15dB.

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ðA2 1Þ logðeÞ logðoocr Þ

lnðA=eÞ : ðor =oc Þ 1

(7:3:9)

Now compute the transfer function HBu ðsÞ from (7.3.6). That is, jHBu ðjoÞj2 ¼ HBu ðsÞHBu ðsÞjo2 ¼s2 :

(7:3:10)

Since the amplitude response function has one in the numerator, we need to compute only the poles of the transfer function. They can be found by solving 2 n 2 o 2n 2 ðs Þ 1 þ e ð Þ jo2 ¼s2 ¼ 1 þ e ¼0 o2n oc c o p c ) s2nþ1 ¼ 1=n ejð2nþ1þnÞ2n ; e n ¼ 0; 1; 2; . . . ; 2n 1:

Solution: From (7.3.7) and in (7.3.8), e ¼ :7648 and n 3:76. Since n has to be an integer, it follows that n ¼ 4: From (7.3.13b), the left-half s-plane poles and the transfer function are s1;7 ¼ ðoc =0:93516Þð0:38268+j0:92388Þ;

(7:3:11) HBu ðsÞ ¼

(7:3:12)

p ¼ oc =e1=n ejð2nþ1þnÞ2n ;

n ¼ 0; 1; . . . ; n 1:

Fig. 7.3.2 Specifications in Example 7.3.1

s3;5 ¼ ðoc =0:93516Þð0:92388+j0:38268Þ

The left half-plane poles of HBu ðsÞ in the exponential and trignometric forms are s2nþ1

ðs1 Þðs3 Þ . . . ðs2n1 Þ : ðs s1 Þðs s3 Þ . . . ðs s2n1 Þ

ðs1 Þðs3 Þðs5 Þðs7 Þ : (7:3:15)& ðs s1 Þðs s3 Þðs s5 Þðs s7 Þ

Maximally flat amplitude property Butterworth function: Consider jHBu ðjoÞj2 ¼

1 1þ

e2 ðo=oc Þ2n

of

:

Even function of ðo=oc Þ2 ¼ l: :

(7:3:16)

(7:3:13a)

2n þ 1 s2nþ1 ¼ oc =e1=n ½ sinð pÞ 2n 2n þ 1 pÞ; n ¼ 0; 1; . . . ; n 1: (7:3:13b) þ j cosð 2n

the

That is, jHBu ðjlÞj2 ¼ jHBu ðjoÞj2 ðo=oc Þ2 ¼l ¼

1 1 þ e 2 ln

7.3 Classical Analog Filter Functions

257

Expanding this function in power series in the neighborhood of l ¼ 0, we have 1 ¼1 þ ð0Þl þ ::: þ ð0Þln1 þ ðe2 Þln þ ::: 1 þ e 2 ln (7:3:17) A simple way to see this is divide the numerator (1) by the denominator ð1 þ e2 ln Þ. The coefficients for the terms lk ; k ¼ 1; 2; :::; n 1 are identically zero. That is, ðn 1Þ derivatives of the Butterworth function are equal to zero. This is called the maximally flat property. Butterworth approximation starts with 1 at o ¼ 0 and monotonically goes to zero as o ! 1. It has the maximally flat response in both pass and stop bands. In the pass band p the magniﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ tude of the function goes from 1 to (1= ð1 þ e2 Þ). Zero error at o ¼ 0 and maximum error at the cutoff frequency oc is approximately equal to e2 =2. It may be of interest to distribute the error throughout the pass band and the Chebyshev approximation achieves that.

Cn ðaÞ ¼ cosðn cos1 aÞ cosðn cos1 ðaÞÞ; ¼ coshðn cosh1 ðaÞÞ;

j aj 1 : j aj 4 1

(7:3:18)

It can be expressed as a polynomial. To show this, let nf ¼ cos1 ðaÞ and Cn ðaÞ ¼ cosðnfÞ: ðC0 ðaÞ ¼ 1 and C1 ðaÞ ¼ aÞ:

(7:3:19a)

Using the trigonometric identities, we can write cnþ1 ðaÞ ¼ cos½ðn+1Þf ¼ cosðnfÞ cosðfÞ+ sinðnfÞ sinðfÞ: Cnþ1 ðaÞ ¼ 2aCn ðaÞ Cn1 ðaÞ; C0 ðaÞ ¼ 1; C1 ðaÞ ¼ a: C2n ðaÞ ¼ :5½C2n ðaÞ þ 1:

(7:3:19b)

Using (7.3.19b) the Chebyshev polynomials can be derived. First few of these are C0 ðaÞ ¼ 1; C2 ðaÞ ¼ 2a2 1;

C1 ðaÞ ¼ a C3 ðoÞ ¼ 4a3 3a

C4 ðaÞ ¼ 8a4 8a2 þ 1; C5 ðaÞ ¼ 16a5 20a3 þ 5a

7.3.3 Chebyshev (Tschebyscheff) Approximations

(7:3:19c)

The nth (n40Þ order Chebyshev polynomial is defined by the transcendental function

These are sketched for n ¼ 1; 2; 3; 4 in Fig. 7.3.3 in the range 0 a. Since the Chebyshev polynomials Cn ðaÞ have even (odd) powers of a only for neven

3.5 3

n=1 n=2

2.5

n=3

2

n=4

Cn(α)

1.5 1 0.5 0 –0.5 –1

Fig. 7.3.3 Chebyshev polynomials, cn ðaÞ; n ¼ 1; 2; 3; 4

–1.5

0

0.1

0.2

0.3

0.4

0.5

0.6

α

0.7

0.8

0.9

1

1.1

258

7 Approximations and Filter Circuits

(odd), sketching the polynomials for negative a is straightforward. For large values of a, Cn ðaÞ 2n1 an ;

a 1:

(7:3:19d)

Properties of the Chebyshev polynomials (see Scheid (1968)): 1. Since coshðaÞ is never zero for real o, it follows that Cn ðaÞ ¼ 0 only for jaj 1: The roots of Cn ðaÞ=0, ak ; k ¼ 1; 2; ::; n are real and jak j51: 2. Since Cn ðaÞ ¼ cosðn cos1 ðaÞÞ; jaj 1; it follows that jCn ðaÞj 1 for jaj 1. 3. For jaj41, Cn ðaÞ increases monotonically consistent with the degree n. 4. Cn ðaÞ is an odd (even) polynomial if n odd (even). 5. The polynomial Cn ðaÞ oscillates with an equiripple character varying between a maximum of +1 and a minimum of –1 for joj 1. " # 6. neven:Cn ð0Þ¼ð1Þn=2 ;Cn ð1Þ¼1 : nodd:Cn ð0Þ¼0;Cn ð+1Þ¼+1ðrespectivelyÞ Cn ðaÞ ¼ 0; a ¼ cosðð2 k þ 1Þp=2nÞ; k ¼ 0; 1; 2; :::; n 1; 1 a 1:

(7:3:20a)

Cn ðaÞ ¼ ð1Þk ; a ¼ cosðkp=nÞ; k ¼ 0; 1; 2; :::; n; 1 a 1:

(7:3:20b)

7. Slope

dCn ðaÞ ja¼1 ¼ n2 : da

an even function. Chebyshev polynomial gives the best approximation in the sense that it minimizes the maximum magnitude of the error for a given value of n. Chebyshev 1 approximation: Noting the characteristics of the Chebyshev polynomials in the range 1 a 1, e2 C2n ðaÞ varies between 0 and e2 in the interval jaj 1 and increases rapidly for jaj41 consistent with n. With these properties in mind, a low-pass amplitude response function can be defined, so that the response swings between 1 and pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1= ð1 þ e2 Þ, in the pass band and monotonically decreasing property in the stop band. Such a function is the Chebyshev 1 function given with a subscript ðc1Þ with the argument ðo=oc ) is given below, see, for example, in Figs. 7.3.4 and 7.3.5. jHc1 ðjoÞj2 ¼

1 : 1 þ e2 C2n ðo=oc Þ

For jo=oc j 1; jHc1 ðjoÞj oscillates between 1 and 1=ð1 þ e2 Þ with equal ripple character. 2

jHc1 ð0Þj ¼

ð1=ð1 þ e2 Þ;

n even;

1;

n odd;

(7:3:20c)

Equations (7.3.20a) and (7.3.20b) follow from

Fig.7.3.4 jHc1 ðjoÞj; n ¼ 1; 2; 3

0 ¼ cosðn cos1 ðaÞÞ ) n cos1 ðaÞ ¼ kp=2; k-odd;

(7:3:22a)

(7:3:21a)

+1 ¼ cosðn cos1 ðaÞÞ ) cos1 ðaÞ ¼ kp=n: (7:3:21b) The Chebyshev polynomial has n roots and they are located in the range 1 a 1. Outside this range, Cn ðaÞ is monotonically increasing (or decreasing in the case of negative a) function. Since Cn ðaÞ is either an even or an odd function, it follows that C2n ðoÞ is Fig. 7.3.5 jHc1 ðjoÞj; n ¼ 4; 5

(7:3:22b)

7.3 Classical Analog Filter Functions

259

jHc1 ðjoc Þj2 ¼ ð1=ð1 þ e2 Þ; n even and n odd: (7:3:22c) The number of peaks ðjHc1 ðjoÞj ¼ 1Þ plus the number of valleys ðjHc1 ðjoc Þj) in the positive frequency range of the pass band is equal to n. This is referred to as the equal-ripple property. Fig. 7.3.5 illustrates this for n ¼ 4; 5. For joj4joc j, jHc1 ðjoÞj decreases rapidly consistent with the value of n. For e small, the width of the ripple in the pass band can be approximated and is e2 =2 (see 6.10.10). From the filter specifications, e gives the permissible range of amplitudes of the Chebyshev 1 response in the pass band and the stop-band attenuation constant A gives a measure of acceptable attenuation in the stop band. The range of frequencies between oc and or is the transition band. Chebyshev 1 transfer function can be computed from Hc1 ðsÞHc1 ðsÞ ¼ Hc1 ðjoÞHc1 ðjoÞ o¼s=j 1 o¼s=j : ¼ o 1 þ e2 c2n ðoc Þ

(7:3:23)

Solving for the roots of the equation Cn ðo=oc Þ o¼s=j ¼ +j=e and selecting the left-halfplane roots results in the following poles of the transfer function (see Weinberg, 1962):

1 1 2i þ 1 sin p si ¼ oc sinh sinh1 n e 2n

1 2i þ 1 1 1 cos p ; þ j cosh sinh n e 2n i ¼ 0; 1; 2; :::; n 1:

(7:3:24)

Design parameters: e controls the ripple width in the pass band and n controls the attenuation in the stop band. That is, 1 1 qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ jor ¼oc ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; or 2 2 1 þ e2 1 þ e Cn ðoc Þ 1 1 ¼ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; or 4oc : A 1 þ e2 C2 ðor Þ

Noting Cn ðaÞ ¼ coshðn cosh1 ðaÞÞ; jaj41, the integer n must satisfy pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ cosh1 ðð A2 1Þ=eÞ n

cosh1 ðor =oc Þ lnð2A=eÞ : (7:3:27) ½ð2=oc Þðor oc Þ1=2 The approximation follows from cosh1 ðxÞ ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2 lnðx þ x 1Þ lnð2xÞ; x 1. If the constraints are given in terms of dB, then

n

cosh1 ½ð10:1Y 1Þ=ð10:1X 1Þ : cosh1 ðor =oc Þ

Hc1 ðsÞ ¼

> > > > > :

Solution: Noting C2n ð1Þ ¼ 1, " # 1 10 log jo¼oc ¼ 2 dB 1 þ e2 C2n ðooc Þ qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ) e ¼ ð102 1Þ ¼ 0:7648:

ðs1 Þðs2 Þ:::ðsn Þ ; ðs s1 Þðs s2 Þ:::ðs sn Þ (

Hc1 ðsÞjs¼0 ¼

1=

0;

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 1 þ e2 ;

n even n odd:

n odd

> > > > > ;

(7:3:28)

Example 7.3.2 Find the Chebyshev 1 transfer function that has X ¼2 dB ripple in the pass band and a minimum attenuation in the stop band of Y ¼15 dB.

Chebyshev-1 low-pass transfer functions: 8 9 1 ðs1 Þðs2 Þ. ..ðsn Þ > > > > p ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ﬃ ; n even > > > > > > < 1 þ e2 ðs s1 Þðs s2 Þ. ..ðs sn Þ =

(7:3:26)

n oc

(7:3:29)

This is the same as in Example 7.3.1. The value of n is determined from ;

(7:3:25)

"

# 1 10 log jo¼or ¼1:69196oc ¼ 15 dB 1 þ e2 C2n ðooc Þ ) Cn ð1:69196Þ ¼

qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ð101: 5 1Þ=ð10:2 1Þ

¼ 7:2358:

(7:3:30)

260

7 Approximations and Filter Circuits

Cn ð1:69196Þ ¼ coshðn cosh1 ð1:69196ÞÞ ¼ 7:2358; n

cosh1 ð7:2358Þ ¼ 2:387 cosh1 ð1:69196Þ

transformation that takes the zero frequency to 1 and vice versa, which is referred to as low-pass : to high-pass transformation. Consider the Chebyshev 1 function with oc ¼ 1 in the form jHc1 ðjuÞj2 ¼

It follows that n ¼ 3 since n must be an integer. The maximum attenuation in the stop band and the transfer function can be determined. These are

1 ¼ 20:81dB 10 log 1 þ e2 C3 ð1:69196Þ

(7:3:31)

Hc1 ðsÞ ¼ s0;2

s1 ¼ oc ð0:36891Þ:

(7:3:32)

Figure 7.3.6 shows the specifications and the & derived amplitude response.

1þ

:

(7:3:33a)

The low-pass to high-pass transformation n ! ð1=nÞ translates the ripples in the pass band to the stop band in the region n41 and monotonic response in the region of n 1. jHc1 ðj=nÞj2 ¼

ðs0 Þðs1 Þðs2 Þ ; ðs s0 Þðs s1 Þðs s2 Þ ¼ oc ð0:184445+j0:92078Þ;

1 e2 C2n ðnÞ

1 : 1 þ e2 C2n ð1=nÞ

(7:3:33b)

The transformation translates the ripples in the pass band to the stop band in the region n41 and gives monotonic response in the region of n 1. The lowpass function is jHC2 ðjnÞj2 ¼ 1 jHC1 ðj=nÞj2 ¼ 1 ¼

1 1 þ e2 C2n ð1=nÞ

e2 C2n ð1=nÞ 1 þ e2 C2n ð1=nÞ

(7:3:34)

For n ¼ 2 and 3, C2 ðaÞ ¼ 2a2 1 and C3 ðaÞ ¼ 4a3 3a: With a ¼ 1=n, we have C2 ð1=nÞ ¼ 2ð1=nÞ2 1 ¼ ð2 n2 Þ=n2 ; Fig. 7.3.6 Example 7.3.2

Chebyshev 1 response has equal ripple property and is monotonic in the stop band. It has a steep transition compared to the Butterworth approximation. Also, the phase response of the Butterworth filter is better with good delay properties. The idea is to find a maximally flat pass-band response to improve the delay performance and retain the steep transition like the Chebyshev 1. Such a case is Chebyshev 2 approximation and is discussed below. Chebyshev 2 or inverse Chebyshev approximation: The Chebyshev 2 approximation function can be derived from Chebyshev 1 function by first considering the normalized Chebyshev 1 function with the cut-off frequency of one and a

C3 ð1=nÞ ¼ 4ð1=nÞ3 3ð1=nÞ ¼ ð4 3n2 Þ=n3 : Therefore,

2 2 n2 e n2 Hc2;2 ðjnÞ 2 ¼

2 2 n2 1þ e n2 2 e ð4 4n2 þ n4 Þ ¼ 2 ½e ð4 4n2 þ n4 Þ þ n4 Hc2;3 ðjnÞ 2 ¼

e2 ð16 24n2 þ 9n4 Þ : ½e2 ð16 24n2 þ 9n4 Þ þ n6

(7:3:35a)

(7:3:35b)

The responses have the maximally flat property at n ¼ 0. They are monotonic in the range 0 n 1

7.3 Classical Analog Filter Functions

261

Fig. 7.3.7 Chebyshev 2 low-pass amplitude response (a) n ¼ 2, (b) n ¼ 3

(a) and have ripples in the range 15n51. The two functions are sketched in Fig. 7.3.7a,b. e can be computed from jHc2 ðjnÞj2n¼1 ¼

1. n

e2 C2n ð1=nÞ n!1 1 þ e2 C2 ð1=nÞ n

¼

0; n odd 2 : e =ð1 þ e2 Þ ; n even ð7:3:36Þ

The Chebyshev 2 normalized transfer function can be computed from Hc2 ð^ sÞHc2 ð^ sÞ ¼ jHc2 ðjnÞj2 n¼s=j :

(7:3:38a)

2. eI ¼ ½1=eCn ðor =oc Þ

3. ^sn ¼ sinh 1 sinh1 1 sin 2n þ 1p n eI 2n

1 1 2n þ1 þ j cosh sinh1 cos p; n eI 2n n ¼ 0; 1; :::; n 1

lim jHc2 ðjnÞj ¼ lim

the stop-band and the pass-band frequencies or and oc in terms of e computed from the passband edge specification. cosh1 ½ð10:1Y 1Þ=ð10:1X 1Þ cosh1 ðor =oc Þ (Same as in Chebyshev 1:Þ

e2 1 : ¼ 1 þ e2 1 þ ð1=e2 Þ

The term ð1=e2 Þ is related to the ripple width in the stop band similar to the term e2 used in determining the ripple width in the pass band in the Chebyshev 1. Note jHc2 ð0Þj ¼ 1 for all values of n, C2n ð0Þ ¼ 0 for n–odd and C2n ð0Þ=1 for n–even. Therefore,

n!1

(b)

4. Poles: sn ¼ 1=^ sn

(7:3:38c)

(7:3:38d)

5. Zeros: z^m ¼ j secðð2m þ 1Þp=2nÞ; m ¼ 0; 1; :::; ðn 1Þ=2; n odd : m ¼ 0; 1; :::; ðn=2Þ 1; n even (7:3:38e)

(7:3:37)

For the derivation of the Chebyshev 2 function, see Weinberg (1962). A summary is given below. Chebyshev 2 function has a maximally flat response in the pass band as in the Butterworth approximation. Note Hc2 ðjoÞ o¼0 ¼ 1. The left half-plane poles sn and the zeros zm of the Chebyshev 2 function normalized to the frequency 1 are determined using the constant eI obtained from the edges of

The function Cn2 is obtained by substituting s^ ¼ s=or in the normalized function. Example 7.3.3 Find the Chebyshev 2 transfer function that has attenuation of X ¼ 2 dB at the edge of the pass band and the minimum attenuation in the stop band of Y ¼15 dB. Note the pass-band and stop-band specifications are the same as in Example 7.3.2.

262

7 Approximations and Filter Circuits

Solution: From Example 7.3.2, e ¼ :76478 and n ¼ 3. From (7.3.38e), the zeros of the transfer function are z^1 ; z^ 1 ¼ +jð1:1547Þ: The other zero is at infinity. The constant eI is eI ¼ ½1=eC3 ðor =oc Þ ¼ ð1=:76478Þð1=14:2998Þ ﬃ :09144:

7.4 Phase-Based Design A system is distortionless if its output is the same as the input except it is attenuated by the same amount for all frequencies with a constant delay (see Section 6.11). The transfer function of a linear time-invariant (LTI) system is given by

Using (7.3.38d), the poles are ^ s0;2 ¼ ð:6103+ j1:3665Þ; s^1 ¼ ð1:2206Þ. The normalized transfer function is ðs z1 Þðs z 1 Þ ; sÞ ¼ K Hc2 ð^ ðs s0 Þðs s1 Þðs s2 Þ ðs0 Þðs1 Þðs2 Þ ðnote Hc2 ð0Þ ¼ 1:Þ: K¼ ðz1 Þðz2 Þ The denormalized transfer function is obtained from

HðjoÞ ¼ jHðjoÞjejyðoÞ : If yðoÞ is linear, then yðoÞ ¼ ot:

(7:4:2)

Since linear phase analog filters are not realizable, they are approximated. The group delay and the phase delays were defined by (6.7.7) and (6.7.19).

Hc2 ðsÞ ¼ Hc2 ðsÞ s¼s=or : Tg ðoÞ ¼ The amplitude response function is sketched in & Fig. 7.3.8.

(7:4:1)

dyðoÞ ; do

Tp ðoÞ ¼

yðoÞ : o

(7:4:3)

Linear phase implies that the group delay in (7.4.3) is a constant. Since we are more interested in the phase angle, we can write (7.4.1) in the form below and solve for yðoÞ. ln½HðjoÞ ¼ lnjHðjoÞj þ jyðoÞ ¼ ð1=2Þ lnjHðjoÞj2 þjyðoÞ ¼ ð1=2Þ ln½HðjoÞHðjoÞþjyðoÞ: (7:4:4)

yðoÞ ¼ ð1=jÞ ln½HðjoÞ ð1=2jÞ ln½HðjoÞ Fig. 7.3.8 Amplitude response of the Chebyshev 2 transfer function in Example 7.3.3

Elliptic filter approximations: Elliptic filter functions have equal ripple in both bands. Elliptic functions are beyond the scope here (see Storer, 1957). For a given set of filter amplitude response specifications, the order of the filter for the Butterworth (nBu ), Chebyshev 1 and 2 (nC1 and nC2 ), and elliptic filters (nE ) satisfy (see Storer, (1957).): nBu nC1 ¼ nC2 nE :

(7:3:39)

ð1=2jÞ ln½HðjoÞ ¼ ð1=2jÞ ln½HðjoÞ ð1=2jÞ ln½HðjoÞ HðjoÞ : (7:4:5a) ¼ ðj=2Þ ln HðjoÞ The generalized phase function is defined by yðsÞ ¼ :5 ln½HðsÞ=HðsÞ: From (7.4.3), and using the chain rule given below, the group delay is given by

7.4 Phase-Based Design

263

expressed in terms of the transform variable s given below, which is useful in computing the delay associated with a transfer function and prime (0 ) denotes differentiation and

dyðoÞ do 1 dðln½HðjoÞÞ 1 dðln½HðjoÞÞ þ ¼ 2j do 2j do

Tg ðoÞ ¼

1 d ln½HðjoÞ 1 d ln½HðjoÞ þ : ¼ 2 dðjoÞ 2 dðjoÞ

Tg ðsÞ ¼

(7:4:5b)

Chain rule :

d ln½HðjoÞ d ln½HðjoÞ dðjoÞ d ln½HðjoÞ ¼ : ¼j do dðjoÞ do dðjoÞ

d ln½HðsÞ s¼jo : ds

ðTg ðsÞ ¼ Tg ðsÞÞ:

(7:4:9)

This results in the group delay that is real and an : even function of o.

Noting the complex-conjugate terms inside the brackets {.} in (7.4.5b), the group delay is d ln½HðjoÞ Tg ðoÞ ¼ Re dðjoÞ ¼ Ev

1 P0 ðsÞ P0 ðsÞ Q0 ðsÞ Q0 ðsÞ þ ; 2 PðsÞ PðsÞ QðsÞ QðsÞ

(7:4:6)

Notes: The symbol EvfXðsÞg is the even part of XðsÞ and is 1 EvfXðsÞg ¼ ½XðsÞ þ XðsÞ 2 Z1 Z1 1 1 st ¼ xðtÞe dt þ xðtÞest dt; 2 2 1 1 EvfXðsÞg s¼jo ¼½XðjoÞ þ X ðjoÞ=2 ¼ RefXðjoÞg: (7:4:7) & We should note that the group delay function is an even function. Assuming that the transfer function HðsÞ ¼ PðsÞ=QðsÞ is a ratio of two polynomials, we have d ln½HðsÞ d ln½PðsÞ d ln½QðsÞ P0 ðsÞ Q0 ðsÞ ¼ ¼ : ds ds ds PðsÞ QðsÞ The generalized phase and the group delay can be defined in terms of the variable s by 1 HðsÞ ; yðsÞ ¼ ln 2 HðsÞ dyðsÞ yðoÞ ¼ jyðsÞ s¼jo ! Tg ðsÞ ¼ ; ds Tg ðoÞ ¼ Tg ðsÞ s¼jo : (7:4:8) Using the property that ln½HðsÞ=HðsÞ ¼ ln½HðsÞ ln½HðsÞ the group delay can be

Example 7.4.1 Compute the generalized phase and the group delay functions for the transfer function HðsÞ ¼ 1=½as2 þ bs þ c. Solution: From (7.4.8), the generalized phase and group delays are as follows: Phase : yðsÞ ¼ ð1=2Þ ln½HðsÞ=HðsÞ; yðoÞ ¼ jyðsÞ s¼jo Group delay: PðsÞ ¼ 0; P 0 ðsÞ ¼ 0; P 0 ðsÞ ¼ 0; QðsÞ ¼ as2 þ bs þ c; QðsÞ ¼ as2 bs þ c d Q0 ðsÞ ¼ 2as þ b; Q0 ðsÞ ¼ QðsÞ ds d 2 ¼ ½as bs þ c ds ¼ ½2as b ¼ 2as þ b 1 2as þ b 2as þ b ) Tg ðsÞ ¼ 0 þ 0 2 2 2 as þ bs þ c as bs þ c ¼

bc abs2 ðas2 þ cÞ2 b2 s2

Tg ðoÞ ¼ Tg ðsÞ s¼jo ¼

:

(7:4:10a)

bc þ abo2 ðc ao2 Þ2 þ b2 o2

(function is real and even):

(7:4:10b)&

7.4.1 Maximally Flat Delay Approximation If we assume in (7.4.1) that jHðjoÞj ¼ 1, then HðjoÞ ¼ ejot0 ; jHðjoÞj ¼ 1; yðoÞ ¼ ﬀHðjoÞ ¼ ot0 ; tðoÞ ¼

dﬀHðoÞ ¼ t0 : do

(7:4:11)

264

7 Approximations and Filter Circuits

It has flat amplitude, linear phase, and a constant group delay with respect to o. The system described by (7.4.11) is distortionless. Using the analytic continuation(see Balbanian et al., 1969), the transfer function can be written in the Laplace transform domain by replacing jo by s. Let us define a normalized transfer function using HðsÞ ¼ HðsÞ s¼st0 ¼ est0 s¼st0 ¼ es :

(7:4:12)

1 s¼st 0 coshðsÞ þ sinhðsÞ 1 : (7:4:13) ¼ ½ðcoshðsÞ= sinhðsÞÞ þ 1 sinhðsÞ ¼

Storch (1954) approximates es by an nth order rational function of the form in (7.4.13) below using the power series approximation of the hyperbolic sine and cosine functions.

HðsÞ Hn ðsÞ ¼

b0 b0 ¼ ; Bn ðsÞ bn sn þ bn1 sn1 þ ::: þ b0

bn ¼ 1; Hn ð0Þ ¼ 1:

(7:4:14)

Bn ðsÞ are the Bessel polynomials and they can be derived using (see Spiegel, 1968.) Bn ðsÞ ¼ ð2n 1ÞBn1 ðsÞ þ s2 Bn2 ðsÞ; B0 ðsÞ ¼ 1; B1 ðsÞ ¼ 1 þ s:

(7:4:15)

For n ¼ 0; 1; 2; 3; 4, these are B0 ðsÞ ¼ 1;

B2 ðsÞ ¼ 3 þ 3s þ s2 ; B3 ðsÞ ¼ 15 þ 15s þ 6s2 þ s3 ; (7:4:16)

The roots of the polynomials can only be determined numerically. The transfer function Hn ðjoÞ

Hn ðjoÞ ¼

Hn ðjoÞ ¼ jHn ðoÞjejyn ðoÞ ; yn ðoÞ ¼ ﬀHn ðoÞ; tn ðoÞ ¼ dyn ðoÞ=do:

(7:4:17)

Example 7.4.2 Show the maximally flat property of the group delays of Hn ðsÞ; n ¼ 1; 2 . H1 ðsÞ ¼

1 3 ; H2 ðsÞ ¼ 2 : sþ1 s þ 3s þ 3

(7:4:18)

Solution: The phase and the delay responses are given by y1 ðoÞ ¼ tan1 ðoÞ; t1 ðoÞ ¼

dy1 dð tan1 ðoÞÞ 1 ¼ ¼ ; (7:4:19a) do do 1 þ o2

y2 ðoÞ ¼ tan1 t2 ðoÞ ¼

3o ; 3 o2

dy2 ð9 þ 3o2 Þ ¼ do ð9 þ 3o2 Þ þ o4

(7:4:19b)

Expressing these in terms of Maclaurin power series in the neighborhood of o ¼ 0 , we can show that the first ð2n 1Þ derivatives of the group delay function vanish at the zero frequency and the maximally flat property follows. This is valid & for all n.

7.4.2 Group Delay of Bessel Functions

B1 ðsÞ ¼ 1 þ s;

B4 ðsÞ ¼ 105 þ 105s þ 45s2 þ 10s3 þ s4

has maximally flat delay characteristics. The transfer function, the phase, and the group delay responses are given for the Bessel transfer function (in terms of frequency o) by

Baher (1990) gives a relationship between an all pole rational function and its group delay and is summarized below in terms of a Bessel transfer function Hn ðsÞ . First, we can write the transfer function (see (7.4.13)) in the form

b0 b0 ¼ : ðb0 b2 o2 þ b4 o4 :::Þ þ jðb1 o b3 o3 þ b5 o5 :::Þ En ðoÞ þ jOn ðoÞ

(7:4:20)

7.4 Phase-Based Design

265

The amplitude, phase, and the corresponding group delay responses are b20 ; E2n ðoÞ þ O2n ðoÞ On ðoÞ yn ðoÞ ¼ tan1 : En ðoÞ

"

#

n 2 o tn ðoÞ ¼ 1 þ ::: b0

jHn ðjoÞj2 ¼

" ﬃ 1 (7:4:21a)

dyn ðoÞ tn ðoÞ ¼ do 2 3 dOn ðoÞ dEn ðoÞ En ðoÞ do On ðoÞ do 5: ¼ 4 E2n ðoÞ þ O2n ðoÞ ) tn ðoÞ ¼ 1 o2n jHn ðjoÞj2 ð1=b20 Þ:

(7:4:21b)

Solution: These can be shown by n ¼ 1 : 1 o2 jH1 ðjoÞj2 ð1=b20 Þ o2 1 ¼ ¼ t1 ðoÞ 2 1þo 1 þ o2

(7:4:23a)

The amplitude response of a Bessel filter function is Gaussian. The attenuation for a filter of order n43, the attenuation and the 3 dB frequency can be approximated by

Example 7.4.3 Verify the results in (7.4.19a and b) using (7.4.21b).

¼1

! # ð2n n!Þ2 2n ðoÞ : ð2nÞ!

(7:4:22a)

20 logjHn ðjoÞj ﬃ 4:3429o2 =ð2n 1ÞÞ

& See Problem 7.4.5 for its use of this. Example 7.4.4 Determine a. the 3 dB frequency and b.the frequency at which the group delay deviates by 1% for a second-order Bessel function. 3 ; H2 ðsÞ ¼ 2 s þ 3s þ 3 9 : (7:4:24a) jH2 ðjoÞj2 ¼ 2 ½ð3 o Þ2 þ 9o2

Solution: a. It follows that jH2 ðjo3dB Þj2 ¼

n ¼ 2 : 1 o4 jH2 ðjoÞj2 ð1=b20 Þ o4 ð9 þ 3o2 Þ ¼ 2 4 ð9 þ 3o Þ þ o ð9 þ 3o2 Þ þ o4 ¼ t2 ðoÞ: ð7:4:22bÞ

1 9 ¼ 2 ½ð3 o23dB Þ2 þ 9o23dB

) o3dB ¼ 1:36:

¼1

Notes: Note Hn ð0Þ ¼ 1 and the group delay has the maximally flat response with tn ð0Þ ¼ 1: The design involves finding the n for a set of specifications including maximum attenuation in the pass band in dB and a constant delay within a prescribed tolerance in the pass band. The group delay can be approximated by using the first two terms in the series and the approximation is good for n43 (see Temes and Mitra, 1973). Assuming the frequency is normalized by t0 , that is o ¼ ot0 , we have

(7:4:23b)

(7:4:24b)

b. The frequency at which the group delay deviates is computed using (7.4.22b) o4:99 ðt2 ðoÞÞ:99 ¼ 1 9 þ 3o2:99 þ o4:99 ¼ :99 ) o:99 ¼ :56:

(7:4:24c)

&

In this example, the 3 dB frequency and the frequency at which certain percent deviation in tn ðoÞ from 1 can be analytically computed. For an arbitrary n, these can be computed either by (7.4.23) or by tables (see Weinberg, 1962.). Table 7.4.1 gives the

Table 7.4.1 Normalized frequencies, o ¼ ot0 . Time delay and a loss table giving the normalized frequency o at which the zero frequency delay and loss values deviate by specified amounts for Bessel filter functions n 1 2 3 4 5 6 7 8 9 10 11 o3dB o1%deviation o10%deviation

1 0.1 0.34

1.36 0.56 1.09

1.75 1.21 1.94

2.13 1.93 2.84

2.42 2.71 3.76

2.70 3.52 4.69

2.95 4.36 5.64

3.17 5.22 6.59

3.39 6.08 7.55

3.58 6.96 8.52

3.77 7.85 9.48

266

7 Approximations and Filter Circuits

Fig. 7.4.2 Example 7.4.5: (a) amplitude and (b) group delay response specifications

(a) normalized 3 dB frequency and the frequencies at which the tðoÞ deviates 1 and 10% from 1. Compare the results in (7.4.24b and c) to the table. Example 7.4.5 Find n for the Bessel filter specifications in Fig. 7.4.2a (for the amplitude) and Fig. 7.4.2b (for the delay) with 1. A delay of t0 =.25 ms up to 1 MHz within 1% deviation 2. A loss of less than 3 dB up to 1 MHz. Solution: From the specifications, the pass-band edge of the normalized frequency is o3dB ¼ o3dB t0 o¼2pð106 Þ ¼ 2pð106 Þð:25ð106 ÞÞ ﬃ 1:57: (7:4:25) From the first condition, using Table 7.4.1, we have n 4. To satisfy the second condition, again using Table 7.4.1, n must be at least equal to 3, as 1.36 < 2; n ¼ 1 xðnts Þ ¼

Fig. 8.2.5 Interpolation using three sinc functions

1; n ¼ 2 : > : 0; otherwise

(8:2:20)

(8:2:19)

Evaluate the function yðtÞ at t ¼ :5ts ; ts ; 1:5ts ; 2ts using the interpolation formula in (8.2.17) and the sampled values of the function xðtÞ in (8.2.20). Solution: By using the interpolation formula and noting that sincðpfs tÞ is even, we have yðtÞ ¼ 2sincðpðfs t 1ÞÞ sincðpðfs t 2ÞÞ; (8:2:21) yðts Þ ¼ 2sincðpðfs ts 1ÞÞ sincðpðfs ts 2ÞÞ ¼ 2sincð0Þ sincðpÞ ¼ 2 ¼ xðts Þ; yð2ts Þ ¼ 2sincðð2ts fs 1ÞpÞ sincððfs ð2ts Þ 2ÞpÞ ¼ 1 ¼ xð2ts Þ;

8.2 Sampling of a Signal

317

Fig. 8.2.6 Example 8.2.2

The constant kn ¼ ð1=fs Þ is the energy contained in each of the sinc functions. To show this, consider the two functions and their transforms given by 1 o joðnts Þ FT e x1 ðtÞ ¼ sincðpfs ðt nts ÞÞ ! P fs 2pfs

yðts =2Þ ¼ 2sincðpðfs ðts =2Þ 1ÞÞ sincðpðfs ðts =2Þ 2ÞÞ ¼ 2sincðp=2Þ sincð3p=2Þ ¼ 1:2732 ð:212207Þ ﬃ 1:4854 yð1:5ts Þ ¼ 2sincðp=2Þ sincðp=2Þ ﬃ :637: See figure 8.2.6

¼ X1 ðjoÞ; &

Notes: The interpolated function yðtÞ ¼ 0 for t ¼ kts , k is an integer and k 6¼ 1; 2. The interpolating function yðtÞ has oscillating tails that die out and is shown in Fig. 8.2.6. xðtÞ is band limited to B0 ¼ fs =2 and therefore it cannot be time limited as the product of the spectral width and the time duration of the function cannot be less than a certain minimum value. See the uncertainty principle in & Fourier analysis in Section 4.7.3.

x2 ðtÞ ¼ sincðpfs ðt mts ÞÞ ¼ X2 ðjoÞ:

(8:2:24b)

Using generalized Parseval’s theorem assuming n 6¼ m with os ts ¼ 2p and using the transforms of the sinc functions, we have 1 1 ð ð 1 x1 ðtÞx2 ðtÞdt ¼ X1 ðjoÞX2 ðjoÞdo 2p 1

8.2.3 Interpolation Formula and the Generalized Fourier Series The interpolation formula is the generalized Fourier series expansion (see Section 3.3) with the orthogonal basis function set consisting of sinc functions. fsincðpðfs tnÞÞ; n¼2;1;0;1;2;...g:;15t51:

(8:2:24a) 1 o joðmts Þ FT e ! P fs 2pfs

1

2 ð1 1 o ¼ P ejoðnmÞts do; 2pfs 2pðfs Þ2 1 ð os =2 1 ejoðnmÞts do ¼ 2 2pðfs Þ os =2 joðnmÞts o¼os =2 1 e ¼ ; 2 ðn mÞt s o¼os =2 2pðfs Þ h i 1 jðnmÞp jðnmÞp e e ¼ 0: ¼ 2pðfs Þ2 ðn mÞts (8:2:25)

(8:2:22) For n ¼ m;

In the first step, the set in (8.2.22) is shown to be an orthogonal basis set over the interval 15t51. That is, ð1 sincðpðfs t nÞÞsincðpðfs t mÞÞdt 1 kn ¼ ð1=fs Þ; n ¼ m : (8:2:23) ¼ 0; n 6¼ m

ð1 1

2 ð1 o P ejoðnmÞts do 2pfs ð2pfs Þ2 1 ð os =2 1 1 ¼ do ¼ : (8:2:26) fs 2pðfs Þ2 os =2

x1 ðtÞx2 ðtÞdt ¼

1

From (8.2.25) and (8.2.26), it follows that the set in (8.2.22) is an orthogonal basis set. Therefore, the generalized Fourier series expansion of yðtÞ is

318

8 Discrete-Time Signals and Their Fourier Transforms

yðtÞ ¼

1 X

Ys ½ksincðpfs ðt kts ÞÞ:

(8:2:27)

k¼1

The generalized Fourier series coefficients can be determined from ð1 Ys ½k ¼ fs yðtÞsincðpfs ðt kts ÞÞdt: (8:2:28) 1

Noting the transform of the sinc pulse is a rectangular pulse (see (4.3.28)) and the generalized Parseval’s theorem (see (8.2.25)) and F½yðtÞ ¼ F½xðtÞ ¼ XðjoÞ results in the following: sincðpfs ðtnts ÞÞ

( sinðpfs ðtnts ÞÞ FT f1s enots ; joj5 o2s ; ¼ ! pfs ðtnts Þ 0; otherwise ð1 )Ys ½k¼fs xðtÞsincðpðfs tnÞÞdt ð os =2

(8:2:29)

1

1 XðjoÞejonts do; 2p os =2 ð1 fs XðjoÞejot dojt¼kts ¼xðtÞjt¼kts ¼xðkts Þ: ¼ 2pðfs Þ 1 (8:2:30)

¼

Since the transform of the function xðtÞ is band limited to os =2, the limits on the transform integral can be changed from ððos =2Þ; ðos =2ÞÞ to ð1; 1Þ in (8.2.30). Example 8.2.3 Let xðtÞ ¼ sinð2pð1ÞtÞ shown in Fig. 8.2.7. It is sampled at the Nyquist rate of fN ¼ 2ð1Þ ¼ 2 samples per second and sampled at t ¼ 0; :5; 1; 1:5; . . .. The sampled values are equal to zero indicating that the signal cannot be recovered from the samples. Nyquist theorem does not identify & where to sample. The sampling rate has to be larger than the Nyquist rate. Its selection is signal dependent and cost-

effectiveness, as the analog-to-digital (A/D) converters are expensive at both the low and the high sampling rate. Sampling a function at much higher than the Nyquist rate does not help. Recovering an analog signal from the samples requires the computation using more samples than necessary and the errors in computation nullifies any advantage used in high sampling. As a guide, the sampling rate is more than the Nyquist rate, about 2.5–10 times the highest frequency in the signal. For seismic signals, the frequencies of interest are in few hundred Hertz range. In these cases, higher sampling rates are used. For speech, the frequency range of interest is from a few Hertz to 3.5 kHz. The sampling rate is taken as 8 kHz or 10 kHz. For CDs, the frequency range of the input signal is from a low frequency of few Hertz to 20 kHz. The sampling rate is taken as 44.1 kHz and the standard sampling rate for studio quality audio is 48 kHz. The compact disc recording system samples each of the two stereo signals with a 16-bit A/D converter at 44.1 kHz (Haykin and Van Veen (2003)). Example 8.2.4 Consider signal xðtÞ band limited to ð2pBÞ rad/s. Determine the Nyquist rates for the functions: a: y1 ðtÞ ¼ xð2tÞ; b: y2 ðtÞ ¼ xðtÞ cosðo0 tÞ Solution: a. Note y1 ðtÞ is formed from xðtÞ by compressing the time axis by a factor of 2. From the Fourier scale change theorem (see Section 4.3.4), we have the following: FT 1 y1 ðtÞ ¼ xðatÞ$ Xðjo=aÞ; a 6¼ 0 j aj FT 1 Xðjo=2Þ ¼ Y1 ðjoÞ: ) xð2tÞ$ 2

Time compression by a factor of 2 results in expansion in frequency by a factor of 2. The Nyquist rate is given by os1 ¼ 2pð2ð2BÞÞ. It is like playing an audio tape fast. b. The signal is a modulated signal with a center frequency o0 with a bandwidth of 2B Hz and F½y2 ðtÞ ¼ F½xðtÞ cosðo0 tÞ ¼ 0:5Xðjðo o0 ÞÞ þ 0:5Xðjðo þ o0 ÞÞ: The highest frequency in the modulated signal is o0 þ 2pB ¼ ðo0 þ os Þ=2. The Nyquist rate is & os2 ¼ os þ 2o0 .

Fig. 8.2.7 xðtÞ ¼ sin 2pð1Þt, Sampled two times per second

The sinc interpolation function is not the best way to approximate the function from its sample

8.2 Sampling of a Signal

319

values, as it decays only at a rate of ð1=tÞ. There are other better functions. The function xðtÞ is known at t ¼ nts . The interpolation formula can be expressed in terms of a function hi ðtÞ that is 0 at all the sampling instants, except at t ¼ 0, where it is 1. In addition, it is absolutely integrable. Interpolation formula is given by yi ðtÞ ¼

1 X

xðnts Þhi ðt nts Þ; hi ðkts nts Þ

n¼1

¼

1; k ¼ n 0; k 6¼ n

:

Since yi ðtÞjt¼kts ¼ xðkts Þ, i.e., the interpolation formula gives the same values at the sampling instants and, at other times, yi ðtÞ is an approximation of xðtÞ. Most commonly interpolating functions are step, linear, sinc, and raised cosine functions. These are given below in table 8.2.1. See Ambardar (1999) for additional discussion on the interpolation functions. Step interpolation (zero-order-hold) uses a rectangular interpolation function and xðnts Þ to produce a stepwise or a staircase approximation of xðtÞ. This is simple and does not depend on the future values of the signal. It is widely used. The reconstructed signal is (more on this in Section 8.2.5.)

It cannot be implemented online since a future value is required. Sinc interpolation was considered earlier. Raised cosine interpolation function (see (4.11.9a) for the function and its transform in (4.11.9b)) uses the roll-off factor b. It reduces to the sinc interpolation function when b ¼ 0. The raised cosine function’s decaying rate is proportional to ð1=t3 Þ. Faster decaying results in improved reconstruction, if the samples are not at exactly at the sampling instants (i.e., jitter). It requires fewer past values are needed in the reconstruction. Polynomial-based interpolation methods are discussed in Appendix A.9.

8.2.4 Problems Associated with Sampling Below the Nyquist Rate Consider the functions x1 ðtÞ and x2 ðtÞ in Fig. 8.2.8. They are sampled at a rate shown. Both provide the same sample values. The function x1 ðtÞ cannot be reconstructed from the sample values. From the figure, x1 ðtÞ has a higher frequency content than x2 ðtÞ. By sampling the functions at the locations shown, some of

yc ðtÞ ¼ xðnts Þ; nts t5xððn þ 1Þts Þ: Linear interpolation (first-order hold) uses a linear approximation and the reconstructed signal is yl ðtÞ ¼ xðnts Þ þ

xððn þ 1Þts Þ xðnts Þ ts

ðt nts Þ; nts t5ðn þ 1Þts :

Fig. 8.2.8 Two signals sampled at the same locations

Table 8.2.1 Common interpolation functions 1 P xðnts ÞP½ðt nts Þ=ts : constant or step interpolation a. yc ðtÞ ¼ n¼1

b. yl ðtÞ ¼

1 P

xðnts ÞL½ðt nts Þ=ts : linear interpolation

n¼1

c. ys ðtÞ ¼

1 P

xðnts Þsinc½pðt nts Þ=ts : sinc interpolation

n¼1

d. yrc ðtÞ ¼

1 X n¼1

xðnts Þ

cosðpbðt nts Þ=ts Þ ð1 ½2bðt nts Þ=ts 2 Þ

sincðpðt nts Þ=ts Þ: raised cosine interpolation, b roll-off factor; 0 b 1.

320

8 Discrete-Time Signals and Their Fourier Transforms

the peaks and valleys of x1 ðtÞ are missed indicating that x1 ðtÞ needs to be sampled at a higher rate than x2 ðtÞ: A spectrum XðjoÞ, its ideally sampled signal spectra by assuming the sampling rates of os1 2pð2BÞ and os2 52pð2BÞ are shown in Fig. 8.2.9a,b,c. In the case of os1 , the message signal can be recovered. In the second case of os2 , the sampling rate is lower than the Nyquist rate. The resultant spectra of the ideally sampled signal will have overlaps of the adjoining spectra and the spectral components are added around half the sampling frequency and the message signal cannot be recovered by low-pass filtering the ideally sampled signal and the filtered signal will be distorted. The distortion caused by sampling below the Nyquist rate is called aliasing. Most signals are not band limited. Therefore, a band limiter is necessary before sampling to minimize the aliasing errors. Example 8.2.5 Consider the amplitude spectrum of a function xðtÞ given by 2

jXðjoÞj ¼

2oc 2

ðoÞ þ o2c

:

(8:2:31)

The signal is sampled at the sampling frequency os and an ideal low-pass filter is used to recover it from

the sampled signal. Use MATLAB to quantify the effect of loosing the spectral energy outside of half the sampling frequency. For MATLAB, see Appendix B. Solution: The energy contained in the signal is ð ð 1 1 2oc 1 1 do E¼ jXðjoÞj2 do ¼ 2p 1 o2 þ o2c 2p 1 2oc 1 tan1 ðo=oc Þ1 ¼ 1 ¼ p ¼ 1: 2poc p (8:2:32) The signal is a low-frequency signal, as most of the spectral energy is concentrated around f ¼ 0. It is not band limited. If it is filtered using an ideal low-pass filter with a cut-off frequency equal to half the sampling frequency ðos =2Þ, then some information is lost and the loss can be measured using the spectral energy contained in the frequency range joj4os =2 and 1 Error ¼ p

ð1

1 jXðjoÞj do ¼ p os =2 2

os=2

2oc ðoÞ2 þ o2c

do

2 ¼ 1 tan1 ðos =2oc Þ: p (8:2:33)

(a)

(b)

Fig. 8.2.9 (a) XðjoÞ, (b) Xs ðjoÞ; os1 > 2pð2BÞ (sampling rate higher than the Nyquist rate), and (c) Xs ðjoÞ; os2 52pð2BÞ (sampling rate lower than the Nyquist rate)

ð1

(c)

8.2 Sampling of a Signal

321

Fig. 8.2.10 Example 8.2.6

Mean-square error 1 0.9

mean-square error

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

0

0.5

If the sampling rate goes to infinity, i.e., os ! 1, then the error goes to zero as the area under the integral goes to zero. For the case of os ¼ 2oc the error is 0.5 or 50%. The error slowly goes down as we increase the value of the sampling rate. A simple MATLAB routine and a sketch of the mean squared error as a function of the ratio of sampling frequency divided by 2fc is given in Fig. 8.2.10. Note that 2fc is not the Nyquist rate since the spectrum is not band limited to fc Hz. For the two cases fs =2fc ¼ 2 and 4, the errors can be calculated and are 0.2952 and 0.1560, respectively. In the above example we need to have a high enough sampling rate to reduce the mean squared error. We use a pre-sampling filter that band limits the signal allowing for a decrease in sampling rate. Such a filter passes frequency components that are below the frequency os =2 and attenuates significantly or even suppress some of the frequency components above os =2: The bandlimiting filter is referred to as an anti-aliasing filter. Even if the signal is band limited to os =2, an antialiasing filter is generally used to avoid aliasing that may result from noise that is ever present in almost all signals. The anti-aliasing filter may not be shown explicitly and is assumed to be included in the system. The bandwidth of the pre-sampling or anti-

1

1.5

2 2.5 3 3.5 sampling frequency/2fc

4

4.5

5

aliasing filter is signal dependent. In simple words, we state that most of the signal energy is contained within the bandwidth B Hz and the energy contained outside this band is negligible. See Section & 4.7 for bandwidth measures. Most of the practical are low-pass signals that have decaying frequency response. One way to look at the aliasing error is to put a limit on the maximum aliasing error at half the sampling frequency which depends on the bandwidth of the signal. This works out nicely for signals that have decaying frequency response. The maximum error occurs at half the sampling frequency. See Spilker (1977), Ambardar (1995) and others. Another simple method is select the essential bandwidth which is taken as the frequency where the spectrum of the signal xðtÞ given by XðjoÞ reduces to say 1% of its peak value. MATLAB Code for Fig. 8.2.10 x=0:.01:5; y=1-(2/pi)*atan(x); plot(x,y) Title (‘Mean-square error’) xlabel (‘sampling frequency/2fc’) ylabel (‘mean-square error’)

322

8 Discrete-Time Signals and Their Fourier Transforms FT

Example 8.2.6 Let xðtÞ ¼ ea t uðtÞ ! 1=ða þ joÞ ¼ XðjoÞ; and

XðjoÞ ¼

1 X k¼1

1 jXðjoÞj ¼ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ; Xð0Þ ¼ 1; o ¼ 2pf: 2 a þ o2 Noting that the maximum aliasing error occurs at o ¼ os =2, find the sampling frequency fs using the following methods: a. a ¼ 1. Maximum aliased magnitude is less than (1) 5% and (2) 1% of the peak value of the function jXðjoÞj. b. a ¼ 2. Use the bandwidth of XðjoÞ as the frequency at which the amplitude reduces to 1% of the peak value. Solution: a. (1). From the statement we have jXðos =2Þj :05jXð0Þj ¼ :05. Now 1 1 qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ! ½1 þ ðos =2Þ2 5400 20 2 1 þ ðos =2Þ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ) os ¼ 4ð399Þ ) fs 6:36 Hz:

(8:2:34)

For a proof and for additional discussion, see Mitra (2006). Some of the digital finite impulse response (FIR) filters are based upon frequency sampling. Discrete-time signal bandwidth: The spectrum of an ideally sampled waveform xs ðtÞ (see 8.2.2) is periodic with period os and the measures of BW used for the continuous signals with nonperiodic spectrum cannot be used here. The ideally sampled signal is uniquely specified for frequencies in the range 0 to fs =2. The bandwidth of xs ðtÞ is the range of positive frequencies within the range 0fs =2, for which the amplitude spectrum is greater than or equal to a times its maximum value, where a is a constant less than 1. The common one is pﬃﬃﬃ a ¼ 1= 2 corresponding to the 3 dB bandwidth.

Flat top sampling uses a sample and hold device illustrated in Fig. 8.2.11 with the input is assumed to be 1 X FT xs ðtÞ ¼ xðnts Þdðt nts Þ ! Xs ðjoÞ: (8:2:35) n¼1

&

pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ b: jXðjoÞj ¼ 1= o2 þ 4 ) 1% of jXðj 0Þj ¼ ð1=2Þ ð:01Þ ¼ :005. For o 2, jXðjoÞj 1=o and jXðjoÞj 1=2pB ¼ :005 ) B ¼ ð100=pÞHz ) fs 2B ¼ 200=p 64 Hz:

jkp sinðoTN kpÞ : TN ðoTN kpÞ

8.2.5 Flat Top Sampling

(2). In this case, we have jXðos =2Þj :01 jXð0Þj ¼ :01. Correspondingly, 1 1 qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ! ½1 þ ðos =2Þ2 > 1000 100 2 1 þ ðos =2Þ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ) os ¼ ð4Þ999 ) fs 10:06 Hz

X

The output of the summer is xðtÞ xðt ts Þ. Let the input is dðtÞ. The output of the first block is ½dðtÞ dðt ts Þ. The transfer functions of the first block and the integrator are H1 ðjoÞ ¼ F½dðtÞ dðt ts Þ ¼ ½1 ejots ; H2 ðjoÞ ¼ 1=jo:

&

Notes: For most signals, the actual spectrum may not be available and experimental methods may be needed to determine the bandwidth of & these. Frequency sampling theorem: Since the Fourier transform and its inverse are related so closely, this theorem follows naturally the timesampling theorem. Consider a time-limited function such that xðtÞ ¼ 0; jtj > TN . It has a Fourier transform that can be uniquely determined from samples at frequency intervals of np=TN and

Assuming the input to the [dðtÞ dðt ts Þ], the output is

hðtÞ ¼

ðt

integrator

is

½dðbÞ dðb ts Þdb

1

¼

(8:2:36)

1; 05t5ts 0; otherwise

¼P

t ðts =2Þ : (8:2:37) ts

The system response to an impulse input is a rectangular pulse of width ts s. This operation is the

8.2 Sampling of a Signal

323

Fig. 8.2.11 Zero-order-hold (a) Use of a delay component, a summer and the process of integration, (b) representation of these using block diagrams

(a)

(b) zero-order-hold (ZoH) referred to in the last section. The function yðtÞ is approximated in the interval nts t5ðn þ 1Þts using the first term in the Taylor series. 1 yðtÞ ¼ yðnts Þ þ y0 ðnts Þðt nts Þ þ y00 2!

(8:2:38)

jH0 ðjoÞj ¼ ts jsincðots =2Þj; ﬀH0 ðjoÞ ðots =2Þ; sincðos t=2Þ > 0 ¼ ðots =2Þ p; sincðos t=2Þ50 dﬀH0 ðjoÞ ts ; ¼ Group delay ¼ 2XðjoÞ do YðjoÞ ¼ H0 ðjoÞXs ðjoÞ: (8:2:42)

2

ðnts Þðt nts Þ þ ; nts t5ðn þ 1Þts ; ZoH introduces three modifications: yðtÞ yðnts Þ; nts t5ðn þ 1Þts :

(8:2:39)

It is called the zero-order-hold since the function is approximated by the constant term in the Taylor series. It can be approximated by the first two terms in the series. That is, yðtÞ yðnts Þ þ y0 ðnts Þðt nts Þ; nts t5ðn þ 1Þts ; y0 ðnts Þ ¼

dyðtÞ jt¼nts : dt

(8:2:40)

Most systems use ZoH. The transfer function of the cascaded blocks in Fig. 8.2.11 is H0 ðjoÞ ¼ H1 ðjoÞH2 ðjoÞ ¼ ½1 ejots =jo;

¼ ts

ejots =2 ejots =2 1 ejots =2 2j ots =2

¼ ts sincðots =2Þejots =2 :

(8:2:41)

The amplitude, phase, and the delay frequency responses are

1. A linear phase shift corresponding to a time delay of ðts =2Þs. 2. The input transform is band limited to om . The output transform YðjoÞ is a distorted version of XðjoÞ affected by the curvature of the main lobe of H0 ðjoÞ. 3. Transform of the ideally sampled signal is periodic. The envelope of the sinc function is inversely proportional to the frequency. The output transform contains distorted and attenuated versions of the images of XðjoÞ, centered at nonzero multiples of os . The first one follows since each sample is held constant for ts seconds. The last two effects are caused by constant or step interpolation resulting in high frequency components. The effects of the first two items can be reduced by increasing the sampling rate. If the effects of the last two items are not acceptable, then a continuous-time compensation filter (an anti-imaging filter) in cascade with the zero-order hold is needed. The transfer function of such is given in (8.2.43). See Haykin and Van Veen (2003). 8 ots os < ; joj5om 5 2 : (8:2:43) Hc ðjoÞ 2 sinðots =2Þ : 0; joj4os om

324

8 Discrete-Time Signals and Their Fourier Transforms

Notes: The output of an ideal sampler and a ‘‘flat top’’ sampler are xs ðtÞ ¼

1 X

xðnts Þdðt nts Þ;

(8:2:44a)

n¼1

t ts =2 nts : (8:2:44b) xðnts ÞP xflat top ðtÞ ¼ ts n¼1 1 X

An ideal low-pass filter is needed to recover xðtÞ from xs ðtÞ. On the other hand, an anti-imaging filter (see (8.2.43)) is necessary to recover xðtÞ from xflat top ðtÞ. In both cases, we assumed that fs > 2B; B ¼ bandwidth of the signal in Hertz. The digital-to-analog (D/A) converter with a ZoH takes the sequence x½n and creates a signal xflat top ðtÞ. So far, the discussion was centered on the lowpass signals. Can a sampling rate less than the twice the highest frequency of a band-pass signal and still recover the input signal? The answer is yes and is illustrated below. See Ziemer and Tranter (2002).

8.2.6 Uniform Band-Pass Sampling Theorem Given a signal xðtÞ

FT

!

XðjoÞ with jfjfl : jXðjoÞj¼0;o¼2pf; jfjfu ½bandwidth¼B¼ðfu fl ÞHz:

Fig. 8.2.12 (a) Band-pass spectra, (b) spectra of the ideally sampled signal

(8:2:45)

The signal xðtÞ can be recovered from the sampled signal if the sampling rate is fs ¼ ð2fu =mÞ, where m the largest integer that is not exceeding ðfu =BÞ. All higher sampling rates are not necessarily usable unless they exceed the Nyquist rate of 2fu . Example 8.2.7 Consider the band-pass signal spectrum XðjoÞ shown in Fig. 8.2.12a. with fl ¼ 4 kHz and fu ¼ 5 kHz(bandwidth is 1 kHzÞ: Using the band pass sampling theorem, sketch the ideally sampled signal spectrum assuming a sampling rate that allows for the recovery of the original signal from the sampled signal. Solution: Note fs ¼ 2fu =m; m ¼ Integer part ofðfu =BÞ ¼ 5; fs ¼ 2ð5Þ=5 ¼ 2 kHz;

Xs ðjoÞ ¼

1 os X Xðjðo kos ÞÞ: 2p k¼1

(8:2:46a)

(8:2:46b)

The band-pass signal can be recovered by noting that none of the spectra Xðjðo kos ÞÞ; k 6¼ 0 over laps the spectrum of the continuous signal xðtÞ. The spectra in (8.2.46b), Xðjðo kos ÞÞ are shifted to the right by kos for k positive and to the left by kos for k negative. In Table 8.2.2, the frequency ranges of the terms in (8.2.46b) for n ¼ 0; 1; 2; 3 and the center frequencies of the corresponding spectra are given. For example Xðjðo os ÞÞ is centered at 2 kHz and occupies the frequency range ð3 kHz5f5 2 kHz;

8.3 Basic Discrete-Time (DT) Signals

325

Table 8.2.2 Spectral occupancy of Xðjðo nos ÞÞ; o ¼ 2pf; n ¼ 0; 1; 2; 3 Spectra Frequency ranges (f, kHz) (nfs, kHz) X(jo) X(j(o – os)) X(j(o + os)) X(j(o – 2os)) X(j(o +2os)) X(j(o – 3os)) Xj((o +3os))

–5 < f M:

(8:5:3)

Using (8.5.3), the second equation in (8.5.2) can be expressed in terms of x½n as

&

Xds ½k ¼

8.5 Discrete-Time Fourier Transforms Computation of the continuous F- and the inverse F-transforms involves integrals and analytical computation is possible only in a few cases. Also, the signal may be available in the form of a waveform instead of an analytical expression or in terms of a sequence. The discrete-time Fourier transform of a sequence is derived from the discrete F-series by taking the number of sample points in the discrete F-series to infinity. Presentation is intuitive and follows that of Haykin and Van Veen (2003).

8.5.1 Discrete-Time Fourier Transforms (DTFTs)

O0 ¼

That is, one period of the periodic sequence is extracted and then it is padded with zeros outside of the period. In (8.5.1), as M increases, the periodic replicates of x½n move further and further away from the origin. The discrete-time F-series representation of the periodic signal and the DFS coefficients are as follows: xs ½n ¼

M X

Xds ½kejkO0 n ;

2p : 2M þ 1

XðejO Þ ¼

1 X

x½nejnO ;

n¼1

¼

Xds ½k ¼

1 2p : xs ½nejO0 nk ; O0 ¼ 2 M þ 1 n¼M 2M þ 1 (8:5:2)

1 XðejO ÞjO¼kO0 2M þ 1

1 X 1 x½nejnkO0 : 2 M þ 1 n¼1

(8:5:5)

For real sequences, with O0 ¼ 2p=ð2 M þ 1Þ, we have M 1 X XðejkO0 ÞejkO0 n O0 : 2p k¼M

(8:5:6)

As M increases, the spacing between the harmonics in the discrete Fourier series decreases (see (8.5.4)). In the limit, as M ! 1, dO ¼ O0 and O ¼ kO0 is some value on the frequency axis. The summation becomes an integral and 1 x½n ¼ 2p

ðp

XðejO ÞejnO dO;

(8:5:7)

2pð ÞM ¼ p: 2M þ 1

(8:5:8)

p

lim ð ÞMO0 ¼ lim

M!1

M!1

The discrete-time Fourier transform (DTFT) and its inverse along with their symbolic relations are as follows:

k¼M M X

(8:5:4)

Now define a continuous periodic function of frequency with period equal to 2p, XðejO Þ; so that the scaled samples of this function are the discrete-time Fourier series.

xs ½n ¼ Let x½n be a non-periodic sequence obtained from a single period of the periodic sequence centered at the origin. xs ½n; jnj M x½n ¼ : (8:5:1) 0; j nj 4 M

1 X 1 x½nejO0 nk ; 2 M þ 1 n¼1

XðejO Þ ¼ 1 2p

ðp p

1 X

x½nejnO ;

n¼1

XðejO ÞejnO dO ¼ x½n;

(8:5:9)

336

8 Discrete-Time Signals and Their Fourier Transforms

x½n

DTFT

! XðejO Þ ¼ XðejO ÞejyðOÞ ; Ffx½ng ¼ XðejO Þ:

XðejO Þ ¼ (8:5:10) ¼

jO

The transform Xðe Þ of a non-periodic sequence is the discrete-time Fourier spectrum or the spectrum of the sequence x½n. Some authors use XðOÞ or XðjOÞ instead of XðejO Þ. XðejO Þ is the preferred notation, as it explicitly shows that the spectrum is periodic and it will be used here. As in the continuous case, is in general complex. the DTFT The quantities XðejO Þ and yðOÞ ¼ ﬀXðejO Þ are the amplitude (or magnitude) and phase (or angle) spectra of the sequence x½n. The DTFT is valid for both real and complex sequences and has interesting properties, similar to the properties of the continuous F-transform. For a real sequence x½n, its amplitude spectrum is even and the phase spectrum is odd, which follows directly from the definition. A sufficient condition for the existence of XðejO Þ is the sequence x½n is absolutely summable. That is, 1 X

jx½nj51:

(8:5:11)

1 X n¼1 1 X

xe ½nejnO þ

n¼1

" ¼ x½0 þ 2

1 X

x0 ½nejnO ; (8:5:13)

n¼1 1 X

# xe ½n cosðnOÞ

n¼1

" þ ð1Þj2

1 X

# x0 ½n sinðnOÞ ¼ ReðXðejO ÞÞ

n¼1 jO

þ jImðXðe ÞÞ: Example 8.5.1 Find the DTFT of the following sequences xi ½n; i ¼ 1; 2. Sketch the responses for Part a. and identify the important values for a ¼ :8. a: x1 ½n ¼ an u½n; jaj51ðright-side sequenceÞ b: x2 ½n ¼ an u½n 1ðleft-side sequenceÞ: (8:5:14) Solution: Noting that u½n ¼ 0; n50 and u½n ¼ 1; n 0, and using (8.5.9), we have

n¼1

a: X1 ðejO Þ ¼ Note the unit step sequence does not satisfy (8.5.11).

x½nejnO

1 X

x1 ½nejnO ¼

n¼1

1 X

ðaejO Þn ¼

n¼0

1 ¼ : ð1acosðOÞÞþjasinðOÞ

8.5.2 Discrete-Time Fourier Transforms of Real Signals with Symmetries A real sequence x½n can be written in terms of its even and odd parts xe ½n and x0 ½n (see (8.3.9c)). Making use of the even and odd sequence properties, the DTFT of these can be written as xe ½n

DTFT

! xe ½0 þ 2

1 X

xe ½n cosðnOÞ;

(8:5:12a)

x0 ½n sinðnOÞ:

(8:5:12b)

n¼1

x0 ½n

DTFT

! j2

1 X n¼1

That is, if a real sequence is even, then its DTFT is real and even and if it odd, then its DTFT is pure imaginary and odd. The DTFT of an arbitrary real sequence x½n is

1 1aejO (8:5:15a)

The magnitude and the phase responses are periodic with period 2p and X1 ðejO Þ ¼

1

; ½1 þ a2 2a cosðOÞ1=2 a sinðOÞ jO 1 ﬀX1 ðe Þ ¼ tan : 1 a cosðOÞ

(8:5:15b)

Figure 8.5.1 gives these plots for one period, namely for 0 O52p. We could plot these for p O5p as they are periodic. The amplitude response is even and phase response is odd. The maximum and the minimum values of the amplitude response for a ¼ 0:8 can be seen by noting jcosðOÞj 1. The maximum and minimum values are located at O ¼ 2 kp and O ¼ ð2 k þ 1Þp. The values are given by 1=ð1 aÞ ¼ 5 and 1=ð1 þ aÞ ¼ 0:5555, respectively. The responses are plotted in Fig. 8.5.1 for 0 O52p. The phase response is a bit more complicated, as the arctangent functions is involved. First

8.5 Discrete-Time Fourier Transforms

337

Fig. 8.5.1 (a) X1 ðejO Þ, (b) ﬀX1 ðejO Þ

The group delay tðOÞ of a sequence x½n can be defined using its transform as follows:

ﬀX1 ðe jkp Þ ¼ 0; k ¼ 0; 1; 2; :::: A simple way to find the peak phase angle is by MATLAB. For a ¼ :8; the peak amplitudes and the phases are equal to 5 and 0.932 rad/s, respectively. b: X2 ðejO Þ ¼ a1 ejO a2 ej2O ¼

1 X

x2 ½nejnO

n¼1

¼ a1 ejO ½1 þ a1 ejO þ a2 ej2O þ a1 ejO 1 jO ; a e 51 ! X2 ðejO Þ ¼ ð1 a1 ejO Þ 1 ¼ ; jaj41: (8:5:15c) 1 aejO The transforms have the same forms, except constraints on the constant 0 a0 are different. find the time sequence from the transform, i.e. inverse transform, we need to know whether sequence is a right-side or a left-side sequence.

Fig. 8.5.2 Four sequences, (a) Type 1, (b) Type 2, (c) Type 3, (d) Type 4

the To the the &

x½n

DTFT

! XðejO Þ ¼ XðejO ÞejfðOÞ ; tðOÞ ¼ dfðOÞ=dO:

(8:5:16)

Since fðOÞ is periodic with period 2p, so is the group delay tðOÞ. Time limited sequences: Linear phase is an important property in the digital filter design. In the following we will consider the expressions for the DTFT of a time limited real sequences h½n ¼ 0; n4N; n50 that have the four conditions stated below. We further assume that the samples h½n have a. an even symmetry about the mid point of the sequence and b. an odd & symmetry about the mid point of the sequence. Sequences of interest (see Fig. 8.5.2): Type 1 sequence: N–odd: Sequence with an even symmetry over its mid point.

338

8 Discrete-Time Signals and Their Fourier Transforms

Type 2 sequence: N–even: Sequence with an even symmetry over its mid point. Type 3 sequence: N–odd: Sequence with an odd symmetry over its mid point. Type 4 sequence: N–even: Sequence with an odd symmetry over its mid point. 1. Sequence has an even symmetry if h½n ¼ h½N 1 n

implying the value of the sequence in the middle is zero. These result in h½n ¼ h½N 1 n and h½ðN 1Þ=2 ¼ 0: (8:5:20) Splitting the sequence h½n into fh½0; h½1; :::; h½ðN 3Þ=2g, h½ðN 1Þ=2 ¼ 0 and fh½ðNþ1Þ=2Þ; :::h½N 1g, the transform can be expressed as

(8:5:17)

2. Sequence has an odd symmetry if h½n ¼ h½N 1 n

HðejO Þ ¼

N 1 X

ðN3Þ=2 X

n¼0

(8:5:18)

We will consider Type 2 and 3 sequences and the other two are left as exercises.

h½nejOn :

HðejO Þ¼

h½nejnO þ

n¼0

¼

N1 X

ðN3Þ=2 X

h½nejOn ¼

h½nejnO þ

h½nejnO

n¼0

ðN=2Þ1 X

ðN3Þ=2 X

¼

h½nejOðN1nÞ :

n¼0

h½N1mejOðN1mÞ : Using this result in (8.5.21), we have

m¼0

ðN=2Þ1 X

n

o h½n ejnO þejnO ejðN1ÞO ;

n¼0

¼

h½mejOðN1mÞ

m¼0

n¼N=2

N=21 X

¼

N1 X

(8:5:21)

n¼ðNþ1Þ=2

n¼ðNþ1Þ=2 N=21 X

h½nejnO

n¼0 N 1 X

þ

Type 2 sequence: The even symmetry h½n ¼ h½N 1 n allows us to write

h½nejnO ¼

"ðN=2Þ1 X

HðejO Þ ¼

# 2h½ncosðððN1Þ=2ÞnÞO e

A2 ðe Þe

j2pððN1Þ=2Þ

:

h½nejnO

n¼0 jOðN1Þ=2

:

n¼0

jO

ðN3Þ=2 X

¼ 2j ¼ 2j

We have made use of the following in simplifying the above expression: n o ejnO þ ejnO ejðN1ÞO n o ¼ ejðN1ÞO=2 ej½ðN1Þ=2nO þ ej½ðN1Þ=2nO ¼ 2 cos½ðN 1Þ=2 nO: Note A2 ðejO Þ is real and the phase angle is fðN 1Þ=2gO, which is linear with respect to O. Also, ðN 1Þ=2 is not an integer since N is even. Type 3 sequences: The DTFT of this sequence can be determined by noting that N is odd and the sequence has an odd symmetry over its mid point

h½nejOðN1nÞ

n¼0

ðN3Þ=2 X n¼0

(8:5:19)

ðN3Þ=2 X

ðN3Þ=2 X

jOn e ejOðN1nÞ ; h½n 2j h½nejðN1Þ=2

n¼0 jO½ððN1Þ=2Þn

ejO½ððN1Þ=2Þn ; 2j " ðN3Þ=2 # X N1 ¼j 2 nÞO ejOðN1Þ=2 sin½ð 2 n¼0 e

¼ A3 ðejO ÞðjejOðN1Þ=2 Þ:

(8:5:22)

The phase angle ﬀ jejOðN1Þ=2 is linear. Also, A 3 ðejO Þ is real and odd symmetric about O ¼ 0 and O ¼ p and

jO

A3 ðe Þ ¼

ðN3Þ=2 X n¼0

sin

N1 n O : 2

(8:5:23)

8.6 Properties of the Discrete-Time Fourier Transforms

339

in terms of A i ðejO Þ and h i ½n can be determined jO & ( ) given Ai ðe Þ. ðN3Þ=2 X N1 N1 2h½ncos HðejO Þ¼ h þ n O 2 2 n¼0 Summary: Type 1 sequence:

ejOðN1Þ=2 ¼A1 ðejO ÞejOðN1Þ=2 :

(8:5:24)

A1 ðejO Þ : Even symmetric about O ¼ 0 and O ¼ p; Hðej0 Þ ¼ h½ðN 1Þ=2 þ

ðN3Þ=2 X

2 h½n:

8.6 Properties of the Discrete-Time Fourier Transforms The DTFT of a sequence and its inverse were given before and are (see (8.5.9)):

n¼0

XðejO Þ ¼ Type 2 sequence: "N=21 # X Hðe Þ¼ 2h½ncos½nOððN1Þ=2ÞO ejOðN1Þ=2 jO

n¼0

¼A2 ðejO ÞejOðN1Þ=2 ; Hðej 0 Þ ¼

ðN=2Þ1 X

Type 3 sequence: HðejO Þ ¼ j 2

# h½n sin½ððN 1Þ=2Þ nÞO

n¼0

ejOðN1Þ=2 ¼ A3 ðejO Þejððp=2ÞðN1ÞO=2Þ ;

(8:5:26)

A3 ðejO Þ : Odd symmetric about O ¼ 0 and O ¼ p ! Hðej0 Þ ¼ 0 and Hðejp Þ ¼ 0: Type 4 sequence: " jO

Hðe Þ ¼ 2

ðN=2Þ1 X

1 2p

XðejO ÞejOn dO;

x½n

DTFT

# h½n sin½nOððN 1Þ=2ÞO

n¼0

jejOðN1Þ=2 ¼ A4 ðejO Þejðp=2ðN1ÞO=2Þ :

! XðejO Þ:

p

(8:6:1)

8.6.1 Periodic Nature of the Discrete-Time Fourier Transform

2 h½n:

A2 ðejO Þ: Even symmetric about O ¼ 0 and is odd symmetric about O ¼ p ! Hðejp Þ ¼ 0.

ðN3Þ=2 X

n¼1 ðp

x½nejnO ; x½n

(8:5:25)

n¼0

"

¼

1 X

(8:5:27)

A4 ðejO Þ : Odd symmetric about O ¼ 0 and even & symmetric about O ¼ p ! Hðej0 Þ ¼ 0: These finite length sequences will be useful in studying finite impulse response (FIR) filters in Chapter 9 are considered. Specifications are given

The F-transform of the discrete-time signal x½n is periodic with period 2p. That is, XðejðOþ2pÞ Þ ¼ XðejO Þ:

(8:6:2)

The continuous-time transform is defined in terms of o in radians/second over the entire range 15o51. When the analog signals are sampled at a sampling frequency of fs Hertz, the spectrum of the digitized signal is periodic with period os ¼ 2pfs ¼ ð2p=ts Þ. The normalized frequency O ¼ ð2pf=fs Þ defines the digital frequency. Notes: In the continuous-time domain, periodic signals are expressed in terms of discrete F-series coefficients. In the discrete-time domain, the samples xðnts Þ are located at discrete times and the DTFT is continuous and periodic with period 2p. The interest is in the digital frequency b and jOj p & or in the range 0 O52p. Example 8.6.1 Find the DTFT of the sequence x½n ¼ 1; 0 n5N. Give the expressions for the magnitude and the phase characteristics of the

340

8 Discrete-Time Signals and Their Fourier Transforms

transform. Sketch the magnitude and phase responses for 05O52p assuming N ¼ 21: Solution: The transform, its amplitude and phase responses are as follows: XðejO Þ ¼

N 1 X

h½n and 20logHðejO Þ, 0 O p are shown in Fig. 8.6.1 for N ¼ 21. The side lobes of the amplitude response become smaller as O goes away from p on both of its sides. The peak of the first side lobe appears near the middle of the first side lobe and is & approximately equal to –13.29 dB.

1 1 ejOn X ; jx½nj ¼ N51; jO 1e n¼1 n¼0 Notes: The amplitude spectrum of a typical win(8:6:3) dow is shown in Fig. 8.6.2. It is even and 2p periodic and the frequency interval of interest is 0 O p. eON=2 ðejON=2 ejON=2 Þ=2j The windows of interest have linear phase. The high ¼ jO=2 e ðejO=2 ejO=2 Þ=2j frequency decay rate of the envelope of the spectrum side lobes tells how fast the spectrum envelope sinðON=2Þ ; (8:6:4) decays after the first zero crossing. ¼ ejOðN1Þ=2 sinðO=2Þ Window parameters: (See Fig. 8.6.2):

ejOn ¼

XðejO Þ ¼ jsinðON=2Þj ; ﬀXðejO Þ jsinðO=2Þj sinðON=2Þ : (8:6:5) ¼ ðN 1ÞO=2 þ arg sinðO=2Þ The amplitude XðejO Þ is even and the phase ﬀXðejO Þ is odd. Both are periodic with period 2p. At O ¼ 0, the function is indeterminate and lim

O!0

sinðON=2Þ ¼ N: sinðO=2Þ

GP = Peak gain of main lobe ¼ N; Gp =N ¼ 1 ¼ 0 dB Gs = Peak side lobe gain, Gs =Gp 0:2172 ¼ 13:3 dB OM = Half-width of main lobe = 2p=N (8.6.7) O3 ¼ 3 dB = half-width, W3 =WM ¼ 0:44 O6 ¼ 6 dB = half-width, W6 =WM ¼ 0:6 Os = Half-width of main lobe to reach Ps ; Ws =WM ¼ 0:81

(8:6:6)

Note XðejO Þ ¼ 0 when O ¼ 2 kp=N; k 6¼ 0. The spacing between zero crossings is (2p=NÞ. The phase angle corresponding to the main lobe is ðN 1ÞO=2 resulting in a value of ðN 1Þp=N at O ¼ ½2p=N . At O ¼ ð2p=NÞþ , the phase angle jumps by p rad reaching a value of p=N since sinðON=2Þ= sinðO=2Þ is positive in the main lobe and negative in the first side lobe. This process is repeated and at O ¼ 2p, the phase angle takes the value of 0 completing one period. The sequence

High-frequency attenuation = 20 dB=decade

8.6.2 Superposition or Linearity Assuming DTFT½xi ½n ¼ Xi ðejO Þ and a0i s are constants, the linearity property is M X

ai xi ½n

!

i¼1

ai Xi ðejO Þ:

(8:6:8)

Magnitude spectrum in dB

0

1

M X i¼1

Rectangular window: N = 21

X: 0.4428 Y: -13.29

–10 –20 Magnitude (dB)

0.8 Amplitude

DTFT

0.6 0.4

–30 –40 –50 –60

0.2 –70 0

Fig. 8.6.1 Discrete rectangular window function and its amplitude spectrum

–80 –10 –8 –6 –4 –2

0 2 Index n

4

6

8 10

0

0.5

1

1.5

2

Frequency (Omega)

2.5

3

8.6 Properties of the Discrete-Time Fourier Transforms Fig. 8.6.2 Amplitude spectrum of typical windows

341

DTFT magnitude spectrum of a typical window

P 0.707P 0.5P PSL

High-frequency decay π

0 Ω3

Ω6

8.6.3 Time Shift or Delay

ΩS ΩM

x1 ½n ¼ 1

This property states

DTFT

! 2pdðOÞ:

DTFT

(8:6:9)

x2 ½n ¼cosðO0 nÞ ¼ :5ðeþjO0 n þ ejO0 n Þ

DTFT

! pdðO O0 Þ

jO

þ pdðO þ O0 Þ ¼ X2 ðe Þ:

This follows from

¼

1 X

x½n n0 ejnO

n¼1 1 X

x½mejðmþn0 ÞO

m¼1

"

¼

(8:6:12)

b. Using the Euler’ formula and (8.6.10) results in

x½n n0 ! ejOn0 XðejO Þ:

Ffx½n n0 g ¼

Ω

1 X

# x½mejmO en0 O :

m¼1

Example 8.6.2 Show that the following relationship is true: x½n ¼ ejO0 n XðejO Þ ¼ 2pdðO O0 Þ; jO0 j p: (8:6:10) Solution: This can be shown using the sifting property of the impulse function. 1 ð 1 x½n ¼ 2pdðO O0 ÞejnO dO ¼ ejnO jO¼O0 ¼ejnO0 : & 2p

(8:6:13) &

8.6.4 Modulation or Frequency Shifting The dual of time shifting is the frequency shifting and is given below ejnO0 x½n

DTFT

! XðejðOO0 Þ Þ:

(8:6:14a)

An extension of this is the modulation in time and the corresponding transform pair is jnO0 e þ ejnO0 DTFT ! x½n cosðnO0 Þ ¼ x½n 2 i 1h XðejðOO0 Þ Þ þ XðejðOþO0 Þ Þ : 2 (8:6:14b)

1

Example 8.6.3 Find the DTFTs of the following functions using the pair in (8.6.10) and the shift property. a: x1 ½n ¼ 1; b: x2 ½n ¼ cosðO0 nÞ; jO0 j p: (8:6:11) Solution: a. Using O0 ¼ 0 in (8.6.10), we have the result as follows:

8.6.5 Time Scaling Time scaling deals with the DTFT of x½cn, where ‘‘C’’ is an integer. For example, consider y½n ¼ x½2n, then y½n has only the even samples of x½n. This is decimation (see Section 8.3.1). To simplify the notation define the following sequence assuming m as an integer:

342

8 Discrete-Time Signals and Their Fourier Transforms

xðmÞ ½n ¼

x½n=m ¼ x½k; 0;

The time scaling property is xðmÞ ½n

n ¼ km; n and k are integers n 6¼ km: an u½n

DTFT

! XðejmO Þ:

(8:6:16)

DTFT

!

1 ; jaj51; a 6¼ 0: 1 aejO

ae ! j dð1=ð1dO

DTFT

y½n ¼ nan u½n þ an u½n 1 X

1 X

xðmÞ ½nejnO ¼

n¼1

ð8:6:19Þ

Solution: From the differentiation property,

This can be seen from F½xðmÞ ½n ¼

(8:6:15)

x½kmejkðmOÞ

þ

k¼1

jO

Þ

1 1 ¼ ; 1 aejO ð1 aejO Þ2

¼ XðejmO Þ ðperiodic with period 2p=mÞ: It illustrates the inverse relationship between time and frequency. The signal spreads in time ðm > 1Þ corresponds to its transform being compressed. Time reversal: A special class of time scaling is time reversal and it results in reversal in frequency. That is, 1 X x½nejnO F½x½n ¼ ¼

n¼1 1 X

x½mejmðOÞ ¼ XðejO Þ:

(8:6:17)

m¼1

Note that jF½x½nj ¼ XðejO Þ ¼ XðejO Þ ¼ jF½x½nj:

) y½n ¼ ðn þ 1Þan u½n

jO

! j dXðe dO

DTFT

Þ

(8:6:18)

:

This is shown by dXðejO Þ d ¼ dO dO ¼

"

1 X

# x½ne

1 ; ð1 aejO Þ2 jaj51; a 6¼ 0:

Example 8.6.5 Find the DTFT of x½n ¼ ajnj ; jaj51; a 6¼ 0. Solution: x½n can be expressed as a sum of the right-and left-side sequences in the form x½n ¼ an u½n þ an u½n d½n. The transforms of each of these are 1 DTFT ; (8:6:21a) an u½n ! 1 aejO an u½n

DTFT

!

1 ; 1 aejO DTFT

! 1:

(8:6:21b) (8:6:21c)

Note the time reversal property of the DTFT was used to find the DTFT of an u½n. With these and making use of the linearity property of the DTFT, we have the DTFT pair

This property is nx½n

!

(8:6:20) &

d½n

8.6.6 Differentiation in Frequency

DTFT

x½n ¼ ajnj þ

DTFT

!

1 1 aejO

1 1; jaj51; a 6¼ 0: (8:6:21d) & 1 aejO

jnO

n¼1

1 X

½ðjnÞx½nejnO :

8.6.7 Differencing

n¼1

Example 8.6.4 Derive the DTFT of the function y½n ¼ ðn þ 1Þan u½n; jaj51 using

The differencing property stated below can be shown using the linearity and the time-shifting properties of the DTFT.

8.6 Properties of the Discrete-Time Fourier Transforms

x½n x½n 1

DTFT

!ð1 ejO ÞXðejO Þ:

(8:6:22)

Example 8.6.6 Find the DTFTs of the sequences a: x1 ½n ¼ d½n ðby the direct methodÞ; b: x2 ½n ¼ u½n c: x3 ½n ¼ u½n 1; d: x4 ½n ¼ sgn½n: Solution: a. Ffd½ng ¼

1 X

d½nejnO ¼ 1:

(8:6:23a)

343

pdðOÞ

x½n ¼

b. Let Uðe Þ ¼ DTFTfu½ng. The unit step sequence is a limiting form of the sequence an u½n with a ! 1. Since u½n is not absolutely summable, the transform of the unit step sequence cannot be obtained by taking the limit of the transform as a ! 1 in (8.6.19). Noting that d½n ¼ u½n u½n 1, and defining F½u½n ¼ UðejO Þ, we have F½d½n ¼ 1 ¼ UðejO Þ ejO UðejO Þ ¼ ð1 ejO ÞUðejO Þ:

(8:6:23b)

Since ð1 ejO ÞjO¼0 ¼ 0, it follows that the transform of the unit step sequence will have an impulse, in addition to the transform in (8.6.21a) with a ¼ 1, resulting in u½n

DTFT

1 1 ejO 1 ¼ pdðOÞ þ ; jOj p: 1 ejO (8:6:24)

1 1 ejO

Note the transform is given by the difference between a complex function and its conjugate illustrating the transform of an odd function and is & imaginary. Inverse discrete-time Fourier transform (IDTFT): Finding the inverse transform

n¼1

jO

1 1 ejO 1 1 ¼ : 1 ejO 1 ejO (8:6:26)

d: F½sgn ¼ F½u½n u½n ¼ pdðOÞ þ

1 2p

ðp

XðejO ÞejOn dO:

p

is difficult, as it involves complex integration. Simple cases are illustrated in Example 8.6.7. Alternate methods suggested below are preferable. 1. Since XðejO Þ is periodic, use the F-series of this function (i.e., in the frequency domain O) and then find the corresponding Fourier series coefficients in the time domain. These methods are useful in designing filters and are discussed in Chapter 9. 2. z-Transforms, discussed in the next chapter, can be used to find the IDTFTs. Example 8.6.7 Find the inverse DTFTs of the periodic functions with period 2p

! UðejO Þ ¼ AdðOÞ þ

The average value of the unit step sequence is (1/2) and its transform is pdðOÞ. Example 4.4.10 illustrated the continuous F-transform of a unit step function. c. The DTFT of x3 ½n can be determined by 1 u½n 1 ¼ u½n 1 !pdðOÞ þ 1 ejO 1 2pdðOÞ ¼ pdðOÞ þ : (8:6:25) 1 ejO DTFT

a: XðejO Þ ¼ 2pdðO O0 Þ; 1; jOj W : b: XðejO Þ ¼ 0; W5jOj p

ð8:6:27Þ

c: Use Part b: to find the DTFTs of the sequences x1 ½n ¼ cosðO0 nÞ and x2 ½n ¼ sinðO0 nÞ: Solution: a. The inverse transform is

x½n ¼ ¼

1 2p ðp

ðp

p

XðejO ÞejOn dO

p

dðO O0 ÞejOn dO ¼ ejnO0 ;

344

8 Discrete-Time Signals and Their Fourier Transforms DTFT

) x½n ¼ ejnO0 ! 2pdðO O0 Þ; p5O p;

b:x½n¼

ðp

1 2p

u½n ¼

n X

! pdðOÞ

m¼1

þ XðejO ÞejnO dO¼

DTFT

d½m

1 2p

1 ¼ UðejO Þ; jOj p: (8:6:33) & ð1 ejO Þ

p

8.6.9 Convolution

W ð

e

jnO

sinðWnÞ ; dO¼ pn

Discrete-time convolution property of x1 ½n and x2 ½n is as follows:

W

sinðWnÞ ) x½n ¼ pn

DTFT

!

1; jOj W : 0; W5jOj p

(8:6:29)

1 X

y½n ¼ x1 ½n x2 ½n ¼

x1 ½kx2 ½n k

k¼1

c. Using Euler’s formula and the results in Part a, the DTFTs of the periodic signals are as follows with period 2p: x1 ½n ¼ cosðO0 nÞ ¼ :5ðejO0 n þ ejO0 n Þ

DTFT

!

¼ X1 ðe Þ; p5O0 p;

x1 ½n kx2 ½k ¼ x1 ½n x2 ½n; 1 X

x1 ½kx2 ½n k

k¼1 DTFT

! X1 ðejO ÞX2 ðejO Þ:

(8:6:35)

(8:6:30)

x2 ½n ¼ sinðO0 nÞ ¼ ð1=2jÞðejO0 n ejO0 n Þ

DTFT

!

jp½dðO O0 Þ dðO þ O0 Þ (8:6:31) &

As in the continuous case the time-domain convolution results in the multiplication in the frequency domain and this property plays an important role in discrete-time linear systems. Using the definition of the transform, we have (

8.6.8 Summation or Accumulation

)

n X

jO

Yðe Þ ¼ F

x1 ½kx2 ½n k

k¼1

The accumulation property is the discrete-time counterpart of the integration in the continuous domain. The summation property is shown later in Section (8.6.11) and is n X

x½m

DTFT

j0

!

pXðe ÞdðOÞ

¼

¼

m¼1

þ

1 XðejO Þ; jOj p: 1 ejO

(8:6:34)

k¼1

¼

jO

1 X

y½n ¼ x1 ½n x2 ½n ¼

p½dðO O0 Þ þ dðO þ O0 Þ

¼ X2 ðejO Þ; p5O0 p:

¼

(8:6:32)

¼

!

1 X

n X

n¼1

k¼1

1 X

x1 ½k

k¼1 1 X

x1 ½kx2 ½n k ejOn ;

1 X

! jOn

x2 ½n ke

n¼1

x1 ½kX2 ðejO ÞejOk ;

k¼1

Example 8.6.8 Find the DTFT of u½n using the accumulation property. n P d½m (see (8.3.8b)), Solution: Noting that u½n ¼ d½n

DTFT

! 1, we have

m¼1

" jO

¼ X2 ðe Þ ¼

1 X

# x1 ½ke

k¼1 jO X2 ðe ÞX1 ðejO Þ:

jOk

(8:6:36)

8.6 Properties of the Discrete-Time Fourier Transforms

Convolution in discrete-time corresponds to multiplication in frequency. The accumulation property given in (8.6.32) can now be shown using the convolution theorem and F½u½n ¼ pdðOÞþ jO ½1=ð1 e Þ. That is,

345

Finding the partial fraction expansion in terms of p ¼ ejO would make it a bit easier. Using the DTFT pair in (8.6.21a), the time sequence is

y½n ¼ x½n u½n ¼ ¼

1 X k¼1 n X

x½ku½n k x½k

(8:6:42)

! XðejO ÞUðejO Þ

¼ XðejO Þ pdðOÞ þ

1 ð1 ejO Þ

b. From (8.6.20), we have y½n ¼ ðn þ 1Þan u½n; & jaj51.

1 XðejO Þ; jOj p: ð1 ejO Þ (8:6:37)

Example 8.6.9 Using the DTFTs of the sequences given below (see (8.6.21a)), find the convolution y½n ¼ x1 ½n x2 ½n using the convolution property. x1 ½n ¼ an u½n; x2 ½n ¼ bn u½n; 05jaj; jbj51; a: a 6¼ b; b: a ¼ b: (8:6:38) Solution: a. The DTFTs of the two sequences are given by 1 X1 ðe Þ ¼ Ffx1 ½ng ¼ ; X2 ðejO Þ 1 aejO 1 ¼ Ffx2 ½ng ¼ ; 1 bejO jO

YðejO Þ ¼ X1 ðejO ÞX2 ðejO Þ 1 : ¼ ð1 aejO Þð1 bejO Þ

(8:6:40)

Using the partial fraction expansion, we have

YðejO Þ ¼

anþ1 bnþ1 u½n; a 6¼ b; jaj; jbj51: ab

DTFT

k¼1

¼ pXðej0 Þ þ

¼

a n b a u½n þ bn u½n ba ba

a b þ : ðb aÞð1 aejO Þ ðb aÞð1 bejO Þ (8:6:41)

8.6.10 Multiplication in Time Dual to the convolution in time is multiplication in time and y½n ¼ x1 ½nx2 ½n

DTFT

!

¼ YðejO Þ:

1 ½X1 ðejO Þ X2 ðejO Þ 2p (8:6:43)

Periodic convolution: 1 X1 ðejO Þ X2 ðejO Þ 2p ð 1 X1 ðeja ÞX2 ðejðOaÞ Þda: ¼ 2p

(8:6:44)

2p

The transform of the product of two sequences is the periodic convolution of the two transforms. It can be seen that 1 X

F½y½n ¼ YðejO Þ ¼

¼

1 X

2

x1 ½nx2 ½nejnO

n¼1

41 2p n¼1

ð

X1 ðeja Þejan dax2 ½nejOn :

2p

Interchanging the order of summation and integration results in

346

8 Discrete-Time Signals and Their Fourier Transforms

1 Yðe Þ ¼ 2p jO

"

ð

ja

X1 ðe Þ

1 ¼ 2p

" ja

X1 ðe Þ

da

#

1 X

x2 ½ne

jðOaÞn

da;

Solution: From the F-transform pair, E¼

ð ð

¼ ja

X1 ðe ÞX2 ðe

jðOaÞ

1 X

jx½nj2 ¼

n¼1

X1 ðeja ÞX2 ðejðOaÞ Þda

2p

1 ¼ 2p

Example 8.6.10 Use the Parseval’s identity and the DTFT pair in (8.6.29) to determine the energy contained in the discrete-time signal x½n ¼ sinðWnÞ=pn:

n¼1

2p

1 ¼ 2p

x2 ½nejðOaÞn

n¼1

2p

ð

#

1 X

Þda:

1 2p

W ð

1 X sin2 ðWnÞ

ðpnÞ2

n¼1

ð1Þ2 dO ¼

W : p

(8:6:46) &

W

2p

8.6.12 Central Ordinate Theorems 8.6.11 Parseval’s Identities The discrete versions of the Parseval’s identities are as follows: 1 X

x1 ½nx2 ½n ¼

n¼1

2p

ð

1 2p

X1 ðejO ÞX2 ðejO ÞdO;

2p

ðGeneralized Parseval ’s identityÞ;

1 X

jx½nj2 ¼

n¼1

1 2p

ð

(8:6:45a)

XðejO Þ2 dO;

2p

ðParseval ’s identityÞ;

(8:6:45b)

YðejO ÞjO¼0 ¼

ð

x1 ½nx2 ½nejnO jO¼0 ja

X1 ðe ÞX2 ðe

jðOaÞ

ÞdajO¼0 ;

2p

1 ¼ 2p

ð

1 2p

ð

DTFT

! XðejO Þ,

find the

Ffð1Þn x½ng ¼ ¼

1 X n¼1 1 X

ð1Þn x½nejnO x½nejnðOþpÞ ¼ XðejðOþpÞ Þ:

n¼1

(8:6:50) ja

X1 ðe ÞX2 ðe

ja

Þda

2p

¼

For an introduction to data encryption, see Hershey and Yarlagadda (1986). It is a vast area and most of these techniques are based on manipulating the data in time domain. A simple spectral based encryption can be seen by using the DTFT illustrated below.

Solution:

n¼1

1 ¼ 2p

8.6.13 Simple Digital Encryption

Example 8.6.11 Using x½n DTFT of ð1Þn x½n.

These can be seen by 1 X

From the DTFT and the IDFT, it follows that ð 1 X 1 j0 Xðe Þ ¼ x½n; x½0 ¼ XðejO ÞdO: (8:6:47) 2p n¼1

X1 ðejO ÞX2 ðejO ÞdO:

2p

Note that if x2 ½n ¼ x1 ½n ¼ x½n, then X2 ðejO Þ ¼ X1 ðejO Þ ¼ X ðejO Þ.

Multiplying the time sequence by ð1Þn simply changes the sign of the data with odd indexes. Since the DTFT spectrum is periodic with period equal to 2p, this operation corresponds to the spectral inversion in the frequency band 0 O p. In Chapter 10, Example 10.4.1, we will consider the & analog frequency band inversion.

8.7 Tables of Discrete-Time Fourier Transform (DTFT) Properties and Pairs

8.7 Tables of Discrete-Time Fourier Transform (DTFT) Properties and Pairs Table 8.7.1 Discrete-time Fourier transform (DTFT) properties xi ½n

DTFT

! Xi ðejO Þ

Linearity: x½n ¼

M P

ai xi ½n

M DTFT P

!

i¼1

ai Xi ðejO Þ:

i¼1

Time shift or delay: DTFT

! ejOn XðejO Þ;

x½n n0

n0 is an integer:

0

Frequency shift and modulation: ejO0 n x½n

DTFT

! XðejðOO Þ Þ: 0

x½n cosðnO0 Þ

DTFT

! 12

XðejðOO0 Þ Þ þ XðejðOþO0 Þ Þ :

Conjugation: x ½n

DTFT

! X ðejO Þ:

Time reversal: x½n

DTFT

! XðejO Þ:

Time scaling: x½n=m; n ¼ km xðmÞ ½n ¼ 0; n 6¼ km Times n property: nx½n

DTFT

DTFT

! XðejmO Þ:

jO

Þ ! j dXðe dO :

First difference: x½n x½n 1

DTFT

!ð1 ejO ÞXðejO Þ:

Summation or accumulation: n DTFT P x½k ! pXðej0 ÞdðOÞ þ 1e1jO XðejO Þ; jOjp: k¼1

Time convolution: x1 ½n x2 ½n

DTFT

! X1 ðejO ÞX2 ðejO Þ:

Multiplication in time: DTFT

1 x1 ½nx2 ½n ! 2p X1 ðejO Þ X2 ðejO Þ ; periodic convolution: Even and odd parts of a real function: x½n ¼ xe ½n þ xo ½n xe ½n

DTFT

! ReðXðejO ÞÞ þ jImðXðejO ÞÞ:

DTFT

! ReðXðejO ÞÞ;

xo ½n

DTFT

! jImðXðejO ÞÞ:

Parseval’s theorem: 1 R P 1 x1 ½nx2 ½n ¼ 2p X1 ðejO ÞX2 ðejO ÞdO: n¼1

2p

1 P n¼1

1 jx½nj2 ¼ 2p

R XðejO Þ2 dO:

2p

Central ordinate theorems: Xðej0 Þ ¼

1 P n¼1

1 x½n; x½0 ¼ 2p

R 2p

XðejO ÞdO:

347

348

8 Discrete-Time Signals and Their Fourier Transforms Table 8.7.2 Discrete-time Fourier transform (DTFT) pairs Unit sample function: d½n n0

DTFT

! ejOn

0

Constant: x½n ¼ A

DTFT

! A2pdðOÞ;

jOj p:

Periodic functions: DTFT

! 2pdðO O0 Þ; jOj; jO0 j p: DTFT cosðO0 nÞ ! p½dðO O0 Þ þ dðO þ O0 Þ; jOj; jO0 j p: DTFT sinðO0 nÞ ! jp½dðO O0 Þ dðO þ O0 Þ; jOj; jO0 j p: 1 1 DTFT P P d½n kN ! O0 dðO kO0 Þ; O0 ¼ 2p=N:

ejO0 n

k¼1

k¼1

Unit pulse sequences: u½n

! pdðOÞ þ 1 1ejO ;

DTFT

jOj p:

! pdðOÞ þ 1 1ejO ; jOj p:

DTFT

u½n 1

Exponential sequences: an u½n

1 ! 1 ae ; jO

DTFT

1 ! 1 ae ; jO

DTFT

an u½n 1 ajnj ; a51

! 1 2a cosðOÞ þ a2 1

DTFT

!

ð1 aejO Þ2

Sinc functions:

x½n ¼

jaj > 1:

1 a2

DTFT

ðn þ 1Þan u½n

x½n ¼

jaj51:

sinðWnÞ pn

DTFT

1; 0nN 1 0; otherwise

!

; jaj51:

1; jOj p 0; WjOj p

! ej OðN1Þ=2 sinðON=2Þ sinðO=2Þ

DTFT

8.8 Discrete-Time Fourier-transforms from Samples of the ContinuousTime Fourier-Transforms In Section 8.2, xðtÞ is sampled signal at a sampling rate of fs ¼ ð1=ts Þ > 2ðBÞ; B ¼ Bandwith of xðtÞ: The continuous-time F-transform of xðtÞ and its inverse are

XðjoÞ ¼

1 ð

xðtÞejot dt; xðtÞ

1

1 ¼ 2p

1 ð

XðjoÞejot do; xðtÞ

FT

! XðjoÞ:

1

(8:8:1)

8.8 Discrete-Time Fourier-transforms from Samples of the Continuous-Time Fourier-Transforms

The transform is then approximated using the rectangular integration formula with a sampling interval of ts s. That is, XðjoÞ ﬃ

1 X

ts ½xðnts Þejonts Xos ðjoÞ; os ¼ 2pfs :

n¼1

(8:8:2) Note that Xos ðjoÞ is a periodic function with period os ¼ ð2p=ts Þ and is an approximation of XðjoÞ in the frequency range joj os =2, as Xos ðjðo þ ð2p=ts ÞÞ ¼ ts ¼ ts

1 X n¼1 1 X

xðnts Þejðoþð2p=ts ÞÞnts

period of the DTFT gives its complete information. 4. Most of the continuous-time functions are time limited to, say T ¼ Nts seconds. In computing the transform, two variables need to be selected, the sampling interval ts and the number of sample points N. Note that if xðtÞ has discontinuities, taking its Fourier transform XðjoÞ and then the inverse transform, F1 ½XðjoÞ, gives half-values of the function at the discontinuities. Therefore, the sampled values at these locations are taken as half-values. For example, if xðtÞ ¼ et uðtÞ, then x½0 ¼ :5: 5. The spectrum of the sampled signal in the interval 0 o5os ¼ ð2p=ts Þ is

xðnts Þejonts ¼ Xos ðjoÞ:

n¼1

(8:8:3)

349

Xos ðjoÞ ﬃ ts

N1 X

xðnts Þejðots Þn ; 0 o5os ¼ 2p=ts :

n¼0

(8:8:6) The transform XðjoÞ is arbitrary and Xos ðjoÞ is periodic. If xðtÞ is band limited to half the sampling rate, the signal xðtÞ can be recovered from the sampled signal and the signal spectrum is XðjoÞ ¼

Xos ðjoÞ; joj os =2 : 0; joj > os =2

(8:8:4)

The discrete-time Fourier transform (DTFT) was defined earlier assuming x½n ¼ xðnts Þ and XðejO Þ ¼

1 X

x½nejnO :

(8:8:5)

n¼1

From this review, the following conclusions can be made: 1. The DTFT is periodic, whereas the continuous Fourier transform is not periodic. 2. The sampling interval ts is not included in (8.8.5). Approximation to the continuous F-transform can be obtained from the DTFT of the sampled signal by multiplying it by ts . 3. The DTFT is defined in terms of the normalized frequency, O ¼ ots ¼ o=fs where fs is the sampling frequency. The normalized frequency O is referred to as the digital frequency and is measured in radians/sample or in radians/cycle. Noting that the DTFT is periodic with period 2p, one

6. Finally, considering Item 4 above in approximating the continuous Fourier transform using the DTFT, the number of sample points, the sampling interval, and the sample values need to be considered. Example 8.8.1 In this example, some of the important facets associated with computing the transform xðtÞ ¼ e2t uðtÞ using DTFT are discussed. What should be the interval T before sampling the signal and the number of samples to be used? Use Example 8.2.6 Part b to find the sampling interval. Solution: First xðtÞ is not time limited. For discrete computations, only a finite interval of time needs to be considered, say T. Find T such that in the interval 0 t5T,xðTÞ551.For T ¼ 4 and 5, we have xð4Þ ¼ :00033 and xð5Þ ¼ :000045. Both are small enough, either one would be adequate and let T ¼ 4: From Example 8.2.6, the sampling interval is ts ¼ ð1=fs Þ p=200. The number of samples is T=ts 255: Discrete computations of transforms are the most efficient if the number of sample points N is a power of 2. Select N ¼ 256: The next step is to identify the sample values. Noting that xðtÞ has a jump discontinuity at t ¼ 0, the first sample value is ½xð0 Þþ xð0þ Þ=2 ¼ :5 ¼ xð0Þ. The sample values are

350

8 Discrete-Time Signals and Their Fourier Transforms

fxðnts Þg ¼ f:5; e2nts ; n ¼ 1; . . . ; 255g: The discrete-time Fourier transform is approximated at the frequency sampled values using (8.8.6). That is, Xos ðjðk=NÞos Þ ﬃ ts

N1 X

intervals of ð2p=Nts Þ in one period, then we have N values Xos ðok Þ; ok ¼ 2pk=N. That is, Xos ðj2pk=NÞ ¼ ts

N1 X

xðnts Þejð2pk=NÞn ;

n¼0

k ¼ 0; 1; 2; . . . ; N 1: xs ½nej2pðkn=NÞ ;

n¼0

k ¼ 0; 1; . . . ; N 1: This is a periodic sequence with period N ¼ 256. The samples in the frequency domain are spaced apart by ð1=NÞfs ¼ ð1=256Þ64 ¼ :25 Hz. Notes: The DTFT is a periodic, continuous function of O and the sampled transform values computed from the DTFT are discrete and periodic. The spectrum of the analog signal is continuous. Increasing the product ðNts Þ implies a longer signal and the discrete transform has more frequency values. Since the sampling frequency is not changed, the effect of increasing ðNts Þ introduces more frequency values. The frequency interval between the spectral components is reduced. The sampling interval ts (ts ¼ 1=fs ) controls the accuracy in the approximation obtained by the DTFT compared to the actual evaluation of & the continuous Fourier transform. As an example, consider that a signal of a 10-s interval that is band limited to 4 kHz. We are interested in estimating the spectrum of the segment using the above procedure with a resolution of, say 0.1 Hz in the spectral spacing. From the Nyquist sampling theorem, a sampling rate of 10 kHz would be adequate. ) Number of samples ¼ fs =frequency resolution ¼ 10 kHz=:1 Hz ¼ 100; 000: Transform algorithms are most efficient if the number of sample points, N is a power of 2. The next highest number that is a power of 2 is 217 ¼ 131072. The length of the corresponding segment is equal to T ¼ Nts ¼ N=fs ¼ 131; 072=10; 000 13:1 s:

8.9 Discrete Fourier Transforms (DFTs) In the last section, the spectrum of the sampled signal Xos ðoÞ in the interval 0 o5os ¼ ð2p=ts Þ was given in (8.8.6). If we sample this function at

(8:9:1)

There are N sample values in time xðnts Þ and N sample values of the spectrum in (8.9.1). These results can be applied to digital data by starting with x½n ¼ xðnts Þ; n ¼ 0; 1; 2; . . . ; N 1 and defining the discrete Fourier transform (DFT) by X½k DFT½x½n ¼

N1 X

x½nejð2p=NÞkn ;

n¼0

k ¼ 0; 1; 2; . . . ; N 1:

(8:9:2)

Note that multiplying X½k by ts results in the spectral values in (8.9.1). The next question is how can the data x½n be obtained from the discrete Fourier transform coefficients X½k? It turns out that these can be determined by x½n ¼

N1 1X X½kejð2p=NÞkn ; N k¼0

n ¼ 0; 1; 2; . . . ; ðN 1Þ:

(8:9:3)

The following shows that (8.9.3) is valid. Substituting the DFT values (see (8.9.2)) in (8.9.3) results in N1 1X X½kejð2p=NÞnk N k¼0 N1 X N1 1X ½ x½mejð2p=NÞmk ejð2p=NÞkn N k¼0 m¼0 N1 N1 X 1X ¼ x½m ½ejð2p=NÞðnmÞk N m¼0 k¼0 " # N1 N1 X X 1 jð2p=NÞðnmÞ k (8:9:4) ¼ x½m ½e : N m¼0 k¼0

¼

Using the summation formula for the finite geometric series in (C.6.1a) results in N 1 X

ðe

k¼0

jð2p=NÞðnkÞ k

Þ ¼

0; n 6¼ m : N: n ¼ m

(8:9:5)

8.9 Discrete Fourier Transforms (DFTs)

" # N1 N1 X 1X jð2p=NÞðnmÞ k ) x½m ½e ¼ x½n: N m¼0 k¼0

351

Therefore, (8:9:6) le jð2p=NÞkn ¼ e jð2p=NÞ½mþlN ¼ e jð2p=NÞm e jð2p=NÞðlNÞ

The data x½n is now tied to the discrete Fourier transform coefficients, or DFTs identified as X½ks. We can summarize the results in terms of DFTs, inverse DFTs, and the symbol for the transform pair as follows:

¼ e jð2p=NÞm ; 0 m N 1:

Noting this, only N terms in (8.9.9) are needed to compute the DFT. The DFT and IDFT have implied periodicity with period N. That is,

X½k DFT½x½n ¼

N 1 X

X½k þ N ¼ x½ne

jð2p=NÞkn

¼

(8:9:7a) N1 1X X½kejð2p=NÞkn ; N k¼0

n ¼ 0; 1; 2; . . . ; ðN 1Þ; DFT

x½n ! X½k:

Notes: There are N equations each to determine the DFTs and the IDFTs. The exponential function e jð2p=NÞ is periodic with period N: Indices in the DFT and IDFT variables are restricted to the principal range 0 n; k N 1. The multiplicative terms in the forward and the inverse transforms e jð2p=NÞkn take one of the values in the set n

o 1; e jð2p=NÞ ; e j2ð2p=NÞ ; . . . ; e jðN1Þð2p=NÞ :

(8:9:9)

This follows from the fact that for 0 k; n N 1; the product ðknÞ can be written as kn ¼lN þ m; 0 k; n; m N 1; k; n; l; m and N; integers:

(8:9:10)

In compact form we can use modulo (mod) N arithmetic. That is, kn ¼ lN þ m m modðNÞ ½mmodðNÞ ¼ ½mðNÞ ; 0 m N 1:

N1 X

(8:9:11)

N 1 X

x½nejð2p=NÞnk ejð2p=NÞnN ;

x½nejð2p=NÞnk ¼ X½k:

(8:9:13a)

n¼0

(8:9:7b)

The sequences x½n; 0 n N 1 and X½k; 0 k N 1 form a discrete Fourier transform (DFT) pair.

x½nejð2p=NÞnðkþNÞ

n¼0

¼

(8:9:8)

N1 X n¼0

; k ¼ 0; 1; 2; . . . ; N 1;

n¼0

x½n ¼ IDFT½X½k ¼

(8:9:12)

x½n þ N ¼ ¼

N1 1X X½kej2pðnþNÞk=N N n¼0 N1 1X X½kej2pðnÞk=N ¼ x½n: N n¼0

(8:9:13b)

At a later time, time and frequency shifts will be considered, such as x½n n0 and X½k k0 , where k0 and n0 are some integers. Since the integers ½n n0 and ½k k0 may fall outside of the range ½0; N 1, these integers need to be converted to a number in the principal range using mod N arithmetic. For example, ½1mod 16 ¼ 15; ½17mod 16 ¼ 1; x½½k þ 1mod 16 ¼ x½ðk þ 1 16Þ ¼ x½k 15: These will not be identified explicitly and implied & from the context. Notes: Modular arithmetic was introduced by Carl Friedrich Gauss in his work Disquisitions Arithmetica, see Hawking (2005). Gauss is considered to be one of the great mathematicians who ever lived. His work laid the foundation for number theory. Many of the digital coding and encryption algorithms are based on number theory. See Gilbert and Hatcher (2000), & Hershey and Yarlagadda (1986), and others.

352

8 Discrete-Time Signals and Their Fourier Transforms

Interestingly, the same algorithm can be used to compute both the forward and the inverse DFT transforms, which can be seen by rewriting (8.9.7b) as

x½n ¼

N1 1X ½X ½kejð2p=NÞnk : N k¼0

8.9.1 Matrix Representations of the DFT and the IDFT The discrete Fourier transform (DFT) can be written in a matrix form and in compact form (see Appendix 1 for a brief review of matrices) using the equations N1 1X DFT x½n ¼ X½kejnð2p=nÞk ! X½k N n¼0

(8:9:14a)

Pictorially (8.9.8) can be described by X½k ! X ½k ! DFTfX ½kg 1 ! ðDFTfX ½kgÞ ¼ x½n: N 2 6 6 6 6 6 6 6 6 4

X½0 X½1 : : :

3

2

1 7 61 7 6 7 6 7 6: 7¼6 7 6: 7 6 7 6 5 4:

X½N 1

¼

e

X ¼ ADFT x:

(8:9:14b)

(8:9:15)

These are

ejð2p=NÞ

: : : :

: :

: :

: : : :

: :

:

: :

:

jð2p=NÞðN1Þ

: :

The vectors X and x are N-dimensional column vectors and the matrix ADFT is a N N complex matrix and is referred to as a discrete Fourier transform (or DFT) matrix. A typical entry in ADFT , say ðk; nÞ entry, is ðk1Þðn1Þ

32

1

6 ejð2p=NÞðN1Þ 7 76 76 76 : 76 76 : 76 76 54 :

: e

jð2p=NÞðN1Þ2

x½0 x½1 : : :

3 7 7 7 7 7; 7 7 7 5

(8:9:16a)

x½N 1

The constant WN is an N th root of unity, as WN N ¼ 1. From the second row or column in the ADFT matrix we have the roots of unity. The exponent t in the entries WtN ¼ ejð2p=NÞt is called the twiddle factor or rotation factor, see Rabiner and Gold (1975). All the entries in ADFT can be simplified to one of the values in the following set (see 8.9.12):

(8:9:16b)

ADFT ðk; nÞ ¼ ejð2p=NÞðk1Þðn1Þ ¼ WN

x½nejkð2p=NÞn

k¼0

1

1

N 1 X

n

; WN

o 1; ejð2p=NÞ ; ejð2p=NÞ2 ; . . . ; ejð2p=NÞðN1Þ : (8:9:18)

¼ ejð2p=NÞ ; 1 k N; 1 n N: The IDFT in (8.9.21b), in a matrix form and its compact from are

(8:9:17)

2 6 6 6 6 6 6 6 6 4

x½0 x½1 : : : x½N 1

3

2

1

7 61 7 6 7 6 7 16: 7¼ 6 7 N6 : 7 6 7 6 5 4:

32

1

:

: :

1

ejð2p=NÞ :

: :

: : : :

ejð2p=NÞðN1Þ :

: :

: :

: : : :

: :

:

: : ejð2p=NÞðN1Þ

1 ejð2p=NÞðN1Þ

2

76 76 76 76 76 76 76 76 54

X½0 X½1 : : : X½N 1

3 7 7 7 7 7; 7 7 7 5

(8:9:19a)

8.9 Discrete Fourier Transforms (DFTs)

x¼

353

1 X: A N DFT

pﬃﬃﬃﬃ pﬃﬃﬃﬃ )ðADFT Þð1= NÞð1= NÞðADFT Þ

(8:9:19b)

(8:9:22) ¼ IN or A1 DFT ¼ ð1=NÞADFT : pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ That is, ð ð1=NÞÞADFT is a unitary matrix.

A typical entry, say the ðn; kÞ entry in ADFT is ADFT ðn; kÞ ¼ ejð2p=NÞðn1Þðk1Þ ðn1Þðk1Þ

¼ WN

; WN ¼ ejð2p=NÞ : (8:9:20)

These matrices are known as Vandermonde matrices (see Hohn (1958)). Note that the matrices ADFT and ðð1=NÞADFT Þ are symmetric matrices. That is, ADFT ðk; nÞ ¼ ADFT ðn; kÞ and ADFT ðk; nÞ ¼ ADFT ðn; kÞ; 2

1

(8:9:21) 3

6 1 7 6 7 x¼6 7; ADFT 4 1 5

2

1 61 6 ¼6 41

2

1

1 j ðjÞ

2

ðjÞ3

ðjÞ

1 j

2

4

ðjÞ6

32 3 2 3 1 1 1 1 1 3 6 1 j 1 j 76 1 7 6 j3 7 6 76 7 6 7 X¼6 76 7¼6 7: (8:9:24) 4 1 1 1 1 54 1 5 4 1 5 j

Solution: For N ¼ 4; we have ðejð2p=NÞ Þkn ¼ ðejð2p=4Þ Þkn ¼ ðejðp=2Þ Þkn ¼ ðjÞkn . Now, the data vector, the DFT matrix and DFT sequence (or coefficients) are

1 ðjÞ2

2

1

Example 8.9.1 Given the data x½0 ¼ 1; x½1 ¼ 1; x½2 ¼ 1; x½3 ¼ 2, find the discrete Fourier transform (DFT) of this sequence using the DFT matrix. Using the DFT coefficients and the ADFT matrix, find the corresponding data sequence.

j3

The DFT matrix is symmetric and it contains only N ¼ 4 distinct elements 1, 1; j; j2 ¼ 1; and j3 ¼ j corresponding to the 4-point DFT. If we can save 1 and j, we can generate the others by changing the sign of these. The data sequence can be computed from the DFT vector by

ðjÞ9

1

32 3 2 3 1 1 1 1 3 1 76 7 6 7 1 16 6 1 j 1 j 76 j3 7 6 1 7 x ¼ ADFT X ¼ 6 76 7¼6 7: 4 44 1 1 1 1 54 1 5 4 1 5 1 j 1 j j3 2

X½k ¼

N 1 X

1

1

j 1

1 1

j 7 7 7 1 5

j

1

j

(8:9:23)

½Refx½ng

n¼1

kn ; þ jImfx½ng Re Wkn N þ jIm WN ¼

N1 X

N1 X Refx½ngRe Wkn Imfx½ngIm Wkn N N

n¼0

"

þj

8.9.2 Requirements for Direct Computation of the DFT

3

1

values in frequency( see (8.9.16a)). It involves the multiplication of a N N complex matrix by a Ndimensional vector. The direct computation of DFT requires N2 complex multiplications and NðN 1Þ additions. Fast Fourier transform (FFT) algorithms, considered in the next chapter, reduces these numbers significantly. FFT algorithms are most effective when the number of data points N is a power of 2. DFT is applicable for real and complex data and can be implemented using real multiplications. First, let x½n ¼ Re½x½n þ jIm½x½n and

2

(8:9:25) &

3 2 1 1 61 ðjÞ3 7 7 6 7¼6 ðjÞ6 5 4 1

n¼0 N1 X

Refx½ngIm Wkn N

n¼0

þ

N 1 X

# kn Imfx½ngRe WN ; k ¼ 0; 1:::; N 1:

n¼0

DFT is a transformation that takes a set of N complex (or real) values in time to N complex (or real)

¼ Re½X½k þ j Im½X½k; k ¼ 0; 1 . . . ; N 1: (8:9:26)

354

8 Discrete-Time Signals and Their Fourier Transforms

Computation of X½k requires 4 N2 real multiplications. For fixed or floating point operations, multiplications are computationally more expensive compared to the additions or transfer of data. We will come back to this topic in Section 9.3. Interestingly, x ½n

DFT

! X ½kModðNÞ :

(8:9:27)

N1 X

¼

N 1 X

X½k ¼ X ½k ¼ X½N k:

X ½k ¼ " x ½nej2pnk=N ¼

N1 X

# x½nej2pnk=N

;

N1 X

¼

¼ X ½kmodðNÞ :

X½N k ¼

modðNÞ

DFT

¼ DFTfx i ½ng; i ¼ 1; 2; . . .

N1 X

x½nejð2p=NÞnk ¼ X ½k: X½k

is

real,

then

x½n ¼x1 ½n þ jx2 ½n

DFT

! X½k; x½n

DFT

! X1 ½k; x2 ½n

DFT

! X½k

DFT

! X2 ½k

1 X1 ½k ¼ fX½k þ X ½N kg; X2 ½k 2 1 ¼ jfX½k X ½N kg: 2

xi ½nejð2p=NÞnk ; xi ½n

k¼0

(8:10:1)

Notes: In the following one variable may be replaced by another variable. The variable n and k will be used to identify the time and the frequency sequences. Most of the proofs follow by using the basic definitions of the DFT (or IDFT). Both x½n and X½k are assumed to have implied periodicity with period N, i.e. x½n þ N ¼ x½n and X½k þ N ¼ X½k; x½n ¼ x½nmodN x½nN and X½k ¼ X½kmodN X½kN :

x½nejð2p=NÞnðNkÞ

Notes: The DFT of two real valued sequences x1 ½n and x2 ½n can be determined from the DFT of the complex sequence x½n ¼ x1 ½n þ jx2 ½n as follows:

x1 ½n

IDFTfXi ½kg ¼ xi ½n ! X i ½k

N1 1X Xi ½kejð2p=NÞnk : N k¼0

N1 X

We can show that if x½n ¼ x ½n ¼ x½N n.

DFT properties are similar to the continuous case. In proving these properties, we will assume that the discrete-time sequences are given by xdi ½n; 0 n N 1, and their discrete Fourier transforms are given by Xdi ½k; 0 k N 1. That is,

¼

x½nej2pnðkÞ=N ¼ X½k;

k¼0

8.10 Discrete Fourier Transform Properties

Xi ½k ¼

x½ne

k¼0

¼

N1 X

# j2pnk=N

n¼0

#

n¼0

N1 X n¼0

n¼0

x½nej2pnðkÞ=N

(8:10:2)

These can be shown as follows:

n¼0

"

The DFT of a real sequence xd ½n has conjugate symmetry. That is,

"

This follows from DFTfx ½ng ¼

8.10.1 DFTs and IDFTs of Real Sequences

8.10.2 Linearity The linearity property follows directly from the definition: M X i¼1

ai xi ½n

DFT

!

M X

ai Xi ½k:

(8:10:3)

i¼1

Example 8.10.1 Find the Nð¼ 8Þ point DFT of the sequence {x½n} and sketch the coefficients. x½n ¼ cosð2pn=NÞ; n ¼ 0; 1; 2; . . . ; N 1:

(8:10:4)

8.10 Discrete Fourier Transform Properties

355

Solution: Using Euler’s identity and the linearity property, the DFT coefficients are

variable other than n and k, say l, and then replace l by –l. That is,

X½k ¼ DFTfx½ng ¼ DFTfcosð2pn=NÞÞg

x½n ¼

¼ :5DFTfejð2pn=NÞ g þ :5 DFTfejð2pn=NÞ g; " # N1 N1 X 1 X jð2pn=NÞð1kÞ jð2pn=NÞð1þkÞ ¼ e þ e (8:10:5) 2 n¼0 n¼0 N N ) X½k ¼ d½k 1 þ d½k þ 1 2 2 N N ¼ d½k 1 þ d½k N þ 1: 2 2

! x½l ¼

N1 1X X½kejð2p=NÞl k ; N k¼1

) x½l ¼ (8:10:6a)

The closed form expression for X½k is obtained by using (C.6.1b). Also, note that d½k þ 1 is outside of the interval 0 n5N, which can be resolved by noting the DFT coefficients have implicit periodicity with period N and write d½k þ 1 ¼ d½k N þ 1. Thus there are two discrete-time impulses in the interval 0 k5N. Here, Xd ½k reduces to X½k ¼ 4d½k 1 þ 4d½k 7:

N1 1X X½kejð2p=NÞnk N k¼1

(8:10:6b)

The transform sequence is shown in Fig. 8.10.1. Note the DFT sequence is real and has even symmetry. Furthermore, we can see that X ½k ¼ X½k ¼ & X½N k.

N1 1X X½rejð2p=NÞl r : N r¼0

Now let l ¼ k; r ¼ n; and taking the ð1=NÞ inside the summation results in the proof of the duality property as follows: x½k ¼

N1 X

ðX½n=NÞejð2p=NÞnk ;

(8:10:8)

n¼0

) X½n=N ¼ IDFT of x½k or the DFTfX½n=Ng ¼ x½k: Example 8.10.2 Use the function in Example 8.10.1 to verify the duality principle. Solution: We have DFT

!ðN=2Þd½k 1

x½n ¼ cosð2pn=NÞn

þ ðN=2Þd½k ðN 1Þ ¼ XðkÞ: Now consider y½n

(8:10:9)

DFT

! Y½k with

8.10.3 Duality y½n ¼ :5d½n 1 The duality property is x½n

1 ! X½k ! X½n Change to N

DFT

DFT

! x½k:

N1 X

:5fd½n 1

þ d½n ðN 1Þge

To show this, start with the IDFT in terms of X½k and rewrite the function in terms of a different

) Y½k ¼ :5½ejð2p=NÞk þ ejð2p=NÞk ¼ cosðð2p=NÞkÞ ¼ cosð2pk=NÞ Note y½k ¼ x½k:

8.10.4 Time Shift Fig. 8.10.1 Example 8.10.1

!

n¼0 j2pnk=N

(8:10:7)

[X]k

DFT

þ :5d½n ðN 1Þ

The time shift property is

(8:10:10) &

356

8 Discrete-Time Signals and Their Fourier Transforms

x½n mmodðNÞ

DFT

! X½kejð2p=NÞkm or

x½n mmodðNÞ ejð2p=NÞkm

DFT

! X½k:

(8:10:11)

We know that x½n and X½k can be considered as periodic sequences and X½k represents the DFT coefficients for one period of x½n. By using a new variable l ¼ n m, and simplifying, we have DFTfx½n mg ¼

N1 X

x½n mejð2p=NÞnk

¼

x½lejð2p=NÞðlþmÞk

l¼0

¼

( N1 X

) x½le

jð2p=NÞlk

ejð2p=NÞmk

l¼0

¼ X½kejð2p=NÞmk :

2p 2p xe ½m sin Nk cos ðmkÞ ¼ N N m¼0 2p 2p ðNk Þ sin mk ; þ cos N N N 1 X 2p 2p ¼ Nk cos ðmkÞ xe ½m sin N N m¼0 2p 2p þ cos ðNkÞ sin mk ; N N N 1 X 2p ¼ xe ½m sin km N m¼0 N 1 X 2p nk ¼0: xe ½n sin ¼ N n¼0 N 1 X

n¼0 N1 X

The second term on the right is N N1 1 X X 2p 2p nk ¼ nk xe ½n sin xe ½n sin N N n¼0 n¼0 N 1 X 2p xe ½N n sin nk ; ¼ N n¼0 N1 X 2p xe ½m sin ðN mÞk ; ¼ N m¼0

(8:10:12)

8.10.5 Frequency Shift The dual to time shift property is the frequency shift property given below and can be shown taking the IDFT of the coefficients on the right. (8:10:13)

A number is equal to its negative only when it is zero. The coefficients are real and follow from (8.10.16) and (8.10.17). Noting the periodicity of the DFT coefficients, we have

If a function is even, i.e., x½n ¼ x½n xe ½n, then the DFT is real and even. It can be written as N1 X 2p DFT nk : (8:10:14) xe ½n ! Xe ½k ¼ xe ½n cos N n¼0

2p Xe ½k ¼ Xe ½N k ¼ nðN kÞ ; xe ½n cos N n¼0 N1 X 2p 2p kN cos ðknÞ xe ½n cos ¼ N N n¼0 2p 2p þ sin kN sin kn ; N N N1 X 2p ¼ nk ¼ Xe ½k ð ) Xe ½k xe ½n cos N n¼0

x½ne

jð2p=NÞmn

DFT

! X½k m modðNÞ :

8.10.6 Even Sequences

N1 X

This can be verified using xe ½n ¼ xe ½N n ¼ xe ½n; ) DFTfxe ½ng ¼

N1 X

xe ½ne

(8:10:15)

¼ Xe ½k; even sequenceÞ:

(8:10:18)

jð2p=NÞkn

n¼0

8.10.7 Odd Sequences 2p nk N n¼0 If a function is odd, i.e., x½n ¼ x½n x0 ½n, then N1 X 2p xe ½n sin nk : (8:10:16) the DFT is real and even. The DFT coefficients of j N n¼0 an odd function are odd and imaginary and

¼

N1 X

xe ½n cos

8.10 Discrete Fourier Transform Properties

2p DFT½x0 ½n ¼ j x0 ½n sin nk : N k¼0 N1 X

357

(8:10:19)

The proof of this is very similar to the last property. As a final comment on this topic, we have seen that a sequence can be expressed in terms of its even and odd parts. The above two properties allow for the computation of the DFT of an arbitrary periodic sequence in terms of the sum of DFTs of even and odd sequences. That is, DFT

! XðejO Þ

x½n ¼ xe ½n þ x0 ½n

¼ RefXðejO Þg þ jImfXðejO Þg ; xe ½n x0 ½n

(8:10:20)

DFT

! RefXðejO Þg;

DFT

jO

! ImfXðe Þg:

¼

¼

N1 X

N1 N1 X 1X X½kH½mejð2p=NÞmn N k¼0 m¼0 " # N1 1X jð2p=NÞik jð2p=NÞim e e ; N i¼0

x½ih½n i ¼

i¼0

(8:10:24)

X 1 N1 X½kH½kejð2p=NÞkn : (8:10:25) N k¼0

Example 8.10.3 Write the periodic convolution of the following two periodic sequences with period N ¼ 3. Compute these using a. the time sequence and b. the DFT. fx½ng ¼ fx½0; x½1; x½2g ¼ f1; 2; 3g; fh½ng ¼ fh½0; h½1; h½2g ¼ f1; 1; 1g;

(8:10:21)

It follows that if x½n is real and even, then XðejO Þ is real and even and if x½n is real and odd then XðejO Þ is real and imaginary.

y½n ¼

N1 X

(8:10:26)

x½ih½n i;y½n

i¼0

¼ x½0h½n þ x½1h½n 1 þ x½2h½n 2: (8:10:27)

8.10.8 Discrete-Time Convolution Theorem

Solution: a. Using h½n ¼ h½N n, the periodic convolution values are as follows: y½0 ¼ x½0h½0 þ x½1h½1 þ x½2h½2 ¼ x½0h½0 þ x½1h½2 þ x½2h½1;

In Section 8.4.1 the periodic (or cyclic) convolution of two functions x½n and h½n with the same period N was defined by N 1 X y½n ¼ x½n h½n ¼ x½mh½n m modðNÞ

y½1 ¼ x½0h½1 þ x½1h½0 þ x½2h½2; y½2 ¼ x½0h½2 þ x½1h½1 þ x½2h½0;

m¼0

¼

N1 X

x½n m modðNÞ h½m:

m¼0

(8:10:22) The time convolution theorem stated below can be proven by starting with the left side of the above equation and rearranging the terms and then simplifying it. That is, DFT

! X½kH½k (8:10:23) " # N 1 N1 N1 X X 1 X x½ih½n i ¼ X½kejð2p=NÞik y½n ¼ N k¼0 i¼0 i¼0 " # N1 1X jð2p=NÞmðniÞ H½me ; N m¼0 x½n h½n modðNÞ

32 3 3 2 h½0 h½2 h½1 x½0 y½0 76 7 6 7 6 Matrix form : 4 y½1 5 ¼ 4 h½1 h½0 h½2 54 x½1 5: 2

h½2 h½1 h½0

y½2

x½2 (8:10:28)

Note the structure of the coefficient matrix on the right in (8.10.28) has a pattern, which can be written in general terms after this example. Noting that x½0 ¼ 1; x½1 ¼ 2; x½2 ¼ 3 and h½0 ¼ 1; h½1 ¼ 1; h½2 ¼ 1, the convolution values are 2

3 2 y½0 1 6 7 6 4 y½1 5 ¼ 4 1

1 1

1

1

y½2

32 3 2 3 1 1 0 76 7 6 7 1 54 2 5 ¼ 4 4 5: 1

3

2

(8:10:29)

358

8 Discrete-Time Signals and Their Fourier Transforms

b. In matrix form, the transform values are 2

3 2 1 X½0 6 7 6 4 X½1 5 ¼ 4 1

H½0

3

2

e

j2ð2p=3Þ

1

3 3 2 6 x½0 7 76 7 6 ej2ð2p=3Þ 54 x½1 5 ﬃ 4 1:5 þ j:8660 5 ; 32

1

ejð2p=3Þ

1

X½2 2

1

1

6 7 6 4 H½1 5 ¼ 4 1 ejð2p=3Þ H½2 1 ej2ð2p=3Þ

e

j4ð2p=3Þ

1:5 j:8660

x½2 32

1

h½0

(8:10:30)

3

2

3

1

76 7 6 7 ej2ð2p=3Þ 54 h½1 5 ﬃ 4 1 þ j1:7321 5: e

j4ð2p=3Þ

(8:10:31)

1 j1:7321

h½2

The product of the transform coefficients in matrix form are given by 2

Y½0

3

2

X½0H½0

3

2

3

1ð6Þ

2

6

3

6 7 6 7 6 7 6 7 4 Y½1 5 ¼ 4 X½1H½1 5 ¼ 4 ð1 þ j1:7321Þð1:5 þ j:8660Þ 5 ¼ 4 3 j1:7321 5: Y½2 X½2H½2 ð1 j1:7321Þð1:5 j:8660Þ 3 þ j1:7321 These involve complex arithmetic resulting in rounded values. The IDFT of the vector in (8.10.32) gives approximations of the results in (8.10.29). &

y½n ¼ x½n h½n ¼

6 6 6 6 6 6 6 6 6 6 6 4

y½0 y½1 y½2 : : : y½N 1

3

2

7 6 7 6 7 6 7 6 7 6 7 6 7¼6 7 6 7 6 7 6 7 6 5 4

¼

N 1 X

h½N 1 h½0

h½2

h½1

h½0

:

:

: :

: :

: :

: :

: :

: : h½N 3 :

: :

y ¼ Hx:

h½N 2 : h½N 1 :

(8:10:34b)

The vectors y and x are N-dimensional column vectors and H is a N N circulant matrix having N distinct elements with a pattern. First, h½n; n ¼ 0; 1; 2; . . . ; N 1 is placed in column 1 in H. Column 2 is obtained by circularly shifting column 1 down by 1. Similarly, column 3 is obtained by circularly shifting the column 2 down by 1 and

h½ix½n imodðNÞ :

(8:10:33)

i¼0

h½0 h½1

: : h½N 1 h½N 2

x½ih½n imodðNÞ

i¼0

The periodic convolution can be written in general matrix and symbolic forms of two periodic sequences x½0; x½1; :::; x½N 1 and h½0; h½1; :::; h½N 1 as follows: 2

N 1 X

(8:10:32)

32 3 x½0 h½1 6 7 h½2 7 76 x½1 7 76 7 6 7 : h½3 7 76 x½2 7 76 7 : : 76 : 7; 76 7 7 6 7 : : 76 : 7 76 7 : : 54 : 5 : h½0 x½N 1

: : : :

(8:10:34a)

so on. The diagonal entries are the same and the entries in each sub diagonal are the same.

8.10.9 Discrete-Frequency Convolution Theorem The discrete-frequency convolution theorem is a dual to the time convolution theorem and is given

8.10 Discrete Fourier Transform Properties

359

below. The proof of this is very similar to the time convolution theorem. DFT 1 x½nh½n ! ½X½k H½k modðNÞ N N1 1X ¼ X½iH½k i modðNÞ : (8:10:35) N i¼0

Note the bracketed term in (8.10.38a) is equal to 1 if m ¼ k and zero otherwise. As in the periodic convolution, DFTs can be used to compute the cross correlation by first computing the DFTs of the two sequences and then take the IDFT of Xd ½kHd ½k.

8.10.11 Parseval’s Identity or Theorem 8.10.10 Discrete-Time Correlation Theorem

It states that if x½n is real, then

In Section 8.3.2 we briefly discussed the discrete cross correlation and the convolution (see (8.3.20a and b)). The discrete cross correlation of two N point sequences is rxh ½n ¼

N1 X

! DFTfrxh ½ng: (8:10:36)

Note that we are using the variable n for the cross correlation function, as we are using the variable k for the DFT function. The discrete correlation theorem is stated by xðiÞhðn þ iÞmodðNÞ

DFT

! X ½kH½k:

(8:10:37)

i¼0

This can be proven using the following steps: " # N1 N1 1 X X X 1N jð2p=NÞik x½ih½n þ imodðNÞ ¼ X½ke N k¼0 i¼0 i¼0 " # 1 X 1N jð2p=NÞðnþiÞm H½me ; N m¼0 ¼

i¼0

n¼0

N k¼0

¼

(8:10:39)

Example 8.10.4 Verify the Parseval’s theorem using hd ½n in Example 8.10.3. Solution: h½0 ¼ 1; h½1 ¼ 1; h½2 ¼ 1 ) H½0 ¼ 1; H½1 ¼ 1 þ jð1:7321Þ; H½2 ¼ 1 j1:7321; 2 X

h2 ½n ¼ 3;

X ½ke

jð2p=NÞik

# N1 1X jð2p=NÞðnþiÞm ; H½me N m¼0

ð1=3Þ

2 X n¼0

2 X

h2 ½n ¼

2 1X jH½kj2 ¼ 3 3 k¼0

(8:10:40)

jH½kj2 ¼ ð1 þ j1 þ jð1:7321Þj2

k¼0

þ j1 jð1:7321Þj2 Þ=3 ¼ ð1 þ 2ð4:0001ÞÞ=3 ﬃ 3:

&

# "

N1 X N1 1X X ½kH½mejð2p=NÞmn N k¼0 m¼0 " # N1 1X jð2p=NÞik jð2p=NÞim e e ; N i¼0

¼

N1 1X jX½kj2 : N k¼0

This can be shown using (8.10.36) with x½n ¼ h½n and is left as an exercise.

n¼0

" N1 N1 X 1X

x2 ½n ¼

DFT

x½ih½i þ nmodðNÞ

i¼0

N1 X

N1 X

N1 1X X ½kH½kejð2p=NÞnk : N k¼0

(8:10:38a)

(8:10:38b)

8.10.12 Zero Padding As mentioned earlier, computational complexity is significantly lower in the computation of DFT when N, the number of sample points in the data, is a power of 2, i.e., with the use of fast Fourier transform (FFT) algorithms discussed in the next chapter. This brings up the interesting question, what is the effect of adding zeros to the end of a sequence? Let the sequence have N1 sample points and let N2

360

8 Discrete-Time Signals and Their Fourier Transforms

zeros be added resulting in N ¼ N1 þ N2 sample points. Noting that the DFT spectrum is periodic with period 2p, the sample points are now spaced ð2p=ðN1 þ N2 ÞÞ instead of ð2p=N1 Þ apart. That is, as more zeros are added, DFT provides closer spaced samples of the transform of the original sequence. We should note that we do not have any more frequency information content than before. It gives a better display. Also, by appropriately padding a required number of zeros ðN2 Þ so that N ¼ N1 þ N2 is a power of two, fast DFT algorithms can be used.

8.10.13 Signal Interpolation In Chapter 1 and in an earlier part of this chapter we have made use of different interpolation functions. In this chapter we have discussed using the sinc and other functions to find interpolated values of the sampled signal. We can make use of the idea of zero padding in the frequency domain using the DFT, which is the dual of improving the spectral resolution by zero padding in the time domain discussed in the last section. Since the sampling frequency is fs ¼ 1=ts , increasing the sampling rate reduces the sampling interval, which, in turn, increases the number of samples in the interval. Let fs1 be the sampling rate used to determine N sampled values. Increasing the sampling rate from fs1 to Mfs1 would introduce interpolated values between samples. Procedure: The sample sequence with Nsample points with even and odd cases by 1 x½n : x½0; x½1; x½2; x½ ðN 1Þ; 2 :::; x½N 1; N odd ; 1 x½n : xd ½0; x½1; x½2; x½ N; 2 :::; x½N 1; N even :

N-odd: Form the MN point DFT Y½k as 1 Y½k : X½0; X½1; X½2; :::; X½ ðN 1Þ; 2 1 ððMN NÞ zerosÞ;X½ ðN þ 1Þ; . . . ; X½N 1: 2 (8:10:41c) N-even:Form the MN point DFT Y½k as 1 1 Y½k : X½0; X½1; X½2; :::; X½ N; 2 2 ððMN N 1Þ zerosÞ; 1 1 1 1 X½ N; X½ N þ 1; . . . ; X½N 1: 2 2 2 2

3. Determine IDFT [Y½k] to obtain the MN point sequence y½n, which may be complex. Since x½n is a real sequence, use only Re fy½ng and multiply by M. Example 8.10.5 Use the above method to interpolate the two sequences given below using the factor M ¼ 1 in the above procedure assuming the cases N ¼ 3 and 4. a: x½0 ¼ 0; x½1 ¼ 1; x½2 ¼ 2; b: x½0 ¼ 0; x½1 ¼ 1; x½2; x½3 ¼ 3: Solution: With the steps given above, the following results:

a. N =3: x½n : 0; 1; 2;

X½k : 3; 1:5 þ j:866; 1:5 j:866

Y½k : 3; 1:5 þ j:866; 0; 0; 0; 1:5 j:866 ! xint ½n : 0; 0; 1; 2; 2; 1 b. N = 4: x½n : 0; 1; 2; 3;

(8:10:41a)

(8:10:41d)

X½k : 6; 2 þ j2; 2; 2 j2

Y½k : 6; 2 þ j2; 12ð2Þ; 0; 0; 0; 12ð2Þ; 2 j2 xint ½n : 0; :0858; 1; 1:5; 2; 2:9142; 3; 1:5

(8:10:41b)

1. Take the DFT of the given sequence. DFTfx½ng ¼ X½k. 2. Insert zeros in the middle of the DFT sequence to create a MN point DFT. The cases for N even and odd are handled differently.

Note that x½k ¼ xint ½2 k; k ¼ 0; 1; 2; . . . ; N 1. The interpolated values are the values in between. Note that in the second case x½0 ¼ 0; x½3 ¼ 3; and x½4 ¼ 0 indicating that the interpolated value at xd;int ½7 will be the average value between 0 and 3, which is equal to 1.5. Similar arguments can be & given for the odd case.

8.10 Discrete Fourier Transform Properties

361

Notes: If a band-limited signal is sampled at a rate higher than the Nyquist rate, then the interpolated sequence will be exact at the sampling

intervals and the values between the samples will be interpolated values. In the case of periodic band-limited signals, the interpolation is exact.

Table 8.10.1 Discrete Fourier transform (DFT) properties Linearity: x½n ¼

M P

ai xi ½n

M DTFT P

!

i¼1

i¼1

ai Xi ½k ¼ X½k; ai 0 s are constants:

Time shift or delay: x½n i modðNÞ

DTFT

! X½kejð2p=NÞik :

Frequency shift: x½nejð2p=NÞni

DTFT

! X½k i modðNÞ :

Time reversal: x½n modðNÞ

DTFT

! X½k modðNÞ :

Alternate inversion formula: N1 P X ½kejð2p=NÞ : x½n ¼ N1 k¼0

Conjugation: x ½n

DTFT

! X ½k modðNÞ :

Duality: X½n

DTFT

! Nx½kmodðNÞ :

Circular convolution and correlation: N1 P

x½nh½n i modðNÞ ¼ x½n h½n modðNÞ

DTFT

! X½kH½k:

i¼0 N1 P i¼0 N1 P i¼0

x½ih½n þ imodðNÞ x½ix½n þ i modðNÞ

DTFT

! X ½kH½k:

DTFT

!jX½kj2 :

Multiplication: x½nh½n

DTFT

! N1 ½X½k H½kmodðNÞ ¼ N1

Real sequences: x½n ¼ xe ½n þ x0 ½n xe ½n

DTFT

! A½k;

x0 ½n

DTFT

! A½k þ jB½k:

DTFT

! Bd ½k:

Parseval’s theorem: N1 P n¼0

jx½nj2 ¼ N1

N1 P k¼0

jX½kj2 :

N1 P i¼0

x½iH½k i modðNÞ :

362

In other cases the interpolation can be poor. If the signals are not band limited, then the interpolation will be obviously poor. For x½n real, the discrete transform coefficients satisfy the conjugate symmetry property, X½N k ¼ X ½k: If the procedure for insertion of the zeros discussed earlier is followed for the interpolation, the conjugate symmetry will be preserved in Y½k. That is, Y½MN k ¼ Yd ½k: IDFT of Y½k will result in a & real sequence, see Ambardar (1995).

8.10.14 Decimation Decimation is an inverse operation of interpolation. It reduces the number of samples by discarding M 1 samples and retaining every M th sample. Note that the corresponding new sampling rate must be above the Nyquist sampling rate to avoid aliasing. This to be of any value, the original signal is assumed to be oversampled.

8.11 Summary

8 Discrete-Time Signals and Their Fourier Transforms

Zero-padding, interpolation, and decimation associated with discrete-time signals

Tables of properties associated with discrete Fourier transforms.

Problems 8.2.1 Consider the function xðtÞ ¼ cosðo0 tÞ. Illustrate the aliasing phenomenon by decreasing os , or equivalently, increasing the sampling interval. Use a low-pass filter of bandwidth equal to ðos =2Þ. In your solution use the following steps. Work out the solution using os > 2o0 and show that the cosine function is recoverable. Now reduce the sampling frequency such that os 52o0 . Sketch the spectrum of the ideally sampled signal and show that the signal exists in the frequency range 05ðos o0 Þ5os =2. 8.2.2 Given xðtÞ is band limited to os =2, determine the Nyquist rates for the functions. a: ya ðtÞ ¼ dxðtÞ dt ; b: yb ðtÞ ¼ xÐt2 ðtÞ; xðaÞda; c: yc ðtÞ ¼ 1

This chapter started with analog signals that are sampled to obtain discrete-time signals. Fourier analysis of discrete time limited signals is discussed in terms of discrete-time and discrete Fourier transforms. The following gives a list of some of the specific topics:

Ideal sampling of a continuous signal Continuous Fourier transforms of the sampled signals

Low-pass and band-pass sampling theorems Basic discrete-time signals and operations, including decimation and interpolation

Basic concepts of discrete-time convolution and correlation

Discrete-time periodic signals and the corresponding discrete Fourier series and their properties

Derivation of the discrete-time Fourier transform Properties of the discrete-time Fourier transform Discrete Fourier transforms and the inverse discrete Fourier transforms

Periodic convolution and correlation and their computations directly and through DFT

8.2.3 Consider the function xðtÞ ¼ cosðo0 t þ yÞ; f0 ¼ o0 =2p ¼ 200 Hz. From the low-pass sampling theorem we know that there will not be any aliasing if os > 2o0 . Now consider that xðtÞ is sampled at two different frequencies one below and one above the Nyquist frequency given by a: fs ¼ 600 Hz; b: fs ¼ 160 Hz. In the first case, we know that there will not be any aliasing. In the second case, the signal xðtÞ sampled at fs ¼ 160 Hz describes a cosine function that is not the given function, but a sampled version of some other cosine function. Give the corresponding function xa ðtÞ ¼ A cosð2pfa t þ yÞ. That is, find fa . Sketch the two functions xðtÞ and xa ðtÞ on the same figure and identify the points where the two functions coincide. (xa ðtÞ ¼ Aliased version of xðtÞ). 8.2.4 The acoustic pulse received by a receiver is represented by xðtÞ ¼ Asinc2 ðo0 tÞ. Noting the transform of this function is a triangular function, give the minimum sampling rate, the expression for the spectrum of the ideally sampled signal, and the minimum band width of the ideally low-pass filter required to reconstruct xðtÞ from the sampled signal.

Problems

363

8.2.5 Find the minimum sampling rate that can be used to determine the samples that completely specify the following signals by assuming ideal sampling:

8.3.2 Find the even and odd parts of the functions. a: xa ½n ¼ u½n; b: xb ½n ¼ ð1=2Þn u½n. 8.3.3 Let x½n ¼ xe ½n þ x0 ½n. Show

a. x1 ðtÞ ¼ ½sinð2pð100ÞtÞ=ð2pð100ÞtÞ; b. b: x2 ðtÞ ¼ cosð2pð100Þt þðp=3ÞÞ þ sinð2pð200ÞtÞ 8.2.6 A signal xðtÞ is band limited to the range f0 5f5500 Hz. Find the minimum sampling rate for xðtÞ without aliasing assuming a. f0 ¼ 0; b. f0 ¼ 100 Hz 8.2.7 The signal xðtÞ ¼ A cosð2pð100ÞtÞ is sampled at 150 Hz. Describe the corresponding signal after the sampled signal is passed through the following filters: a. An ideal low-pass filter with a cut-off frequency of 20 Hz b. An ideal band-pass filter with a pass band between 60 Hz and 120 Hz 8.2.8 Consider the sampled sequence xð0Þ ¼ 1; xðts Þ ¼ 0; xð2ts Þ ¼ 1; xðnts Þ ¼ 0; n 6¼ 0; 1; 2. Sketch the interpolated functions using a. step, b. linear, and c. sinc interpolations. 8.2.9 Let F½xðtÞ ¼ XðjoÞ with XðjoÞ ¼ 0; joj2pB. Using the results in Section 8.2.3, show 1 ð 1

1 X

1 x ðnts Þ; ts ¼ ; fs ts ¼ 1: jxðtÞj dt ¼ ts 2B n¼1 2

2

8.2.10 Use the band-pass sampling theorem to determine the possible sampling rates so that the following signal can be recovered from the sampled signal:

o þ oc o oc þP ; xðtÞ ! XðjoÞ ¼ P 2pð2BÞ 2pð2BÞ FT

B ¼ 8 kHz; oc ¼ 2pfc ¼ fc ¼ 64 kHz: Assuming the sampling rates of a. fsa ¼ 200 kHz; b. fsb ¼ 20 kHz; and c. fsc ¼ 16 kHz, illustrate how the signal can be recovered from the sampled signals if possible. 8.3.1 Sketch the following sequences assuming x½n ¼ ð1 nÞfu½n u½n 3g: a: ya ½n ¼ x½2n 1; b: yb ½n ¼ x½n2 1; c: yc ½n ¼ x½1 n:

1 X

E¼

x2 ½n ¼

n¼1

1 X

x2e ½n þ

n¼1

1 X

x20 ½n:

n¼1

8.3.4 Derive the following identities and then simplify the results when N ! 1: N1 X

a: S ¼

an ¼

n¼0

b: c:

N 1 X

nan ¼

n¼0 1 X

1 aN ; jaj51; 1a

ðN 1ÞaNþ1 NaN þ a ð1 aÞ2

eajnj ¼

n¼1

;

1 þ ea : 1 ea

8.3.5 Find the closed form expression for y½n ¼ an u½n u½n; jaj51. 8.3.6 Find the cross correlation of the two sequences given by x½n ¼ u½n u½n nx and h½n ¼ u½n u½n nh for the cases: a: nx ¼ nh ¼ 2; b: nx ¼ 2; nh ¼ 3. 8.4.1 Determine the DTFS of the following sequences by using Euler’s theorem and then by identifying the discrete Fourier series coefficients. Identify the periods. a: xs a ½n¼1 þ sinðpn=2 þ yÞ; b: xsb ½n¼ cosðnp=20Þþ sinðnp=40Þ; c. xsc ½n ¼ cos2 ½np=8 8.4.2 Find the DTFS coefficients of the N-periodic discrete-time functions 1 P

a. xsa ½n ¼

d½n lN;

l ¼1

b. xb ½n ¼

1; 0 jnj M . 0; M5n5N M

8.4.3 Determine the time-domain sequences with period N ¼ 7 with the DTFS coefficients a. Xsa ½k ¼ ð1=2Þ; b. Xsb ½k ¼ cosð2 kp=NÞ:

364

8 Discrete-Time Signals and Their Fourier Transforms

8.4.4 a. Show that Xs ½k ¼ Xs ½N k for the following sequence: 1; 0 n5ðN 1Þ=2 xs ½n ¼ 0; ðN 1Þ=2 þ 1 n N 1; xs ½n ¼ xs ½n þ N: b. Given the periodic sequences xs ½n ¼ f0; 0; 1; 2g; hs ½n ¼ f1; 2; 0; 0g;

8.5.7 Verify the results given in Section 8.5.2 for Type 1 and 4 sequences. 8.6.1 Prove the time reversal property in (8.6.17). 8.6.2 a. Determine the DTFT of the function x½n ¼ ð1=3Þu½n: b. Use the time reversal property to determine the DTFT of ð3Þn u½n. 8.6.3 Find the DTFT of the function

xs ½n ¼ xs ½n þ 4 and hs ½n ¼ hs ½n þ 4: find the DTFS of the function ys ½n ¼ xs ½nhs ½n. Illustrate the generalized Parseval’s identity by using the DTFS of the functions xs ½n; hs ½n and ys ½n.

y½n ¼ ðn þ 1Þ2 x½n: 8.6.4 Determine the convolutions x1 ½n x2 ½n for the following cases:

8.4.5 Use the sequences in Problem 8.4.4b to determine ys ½n ¼ xs ½n hs ½n.

a: x1 ½n ¼ u½n; x2 ½n ¼ u½n;

8.4.6 Give an example of two sinusoidal sequences that are equal. Hint: Assume cosðO1 pk þ yÞ ¼ cosðO2 pk þ yÞ with O1 6¼ O2 and show the two functions are equal.

8.6.5 Determine a:

8.5.1 Show that x ½n

b: x1 ½n ¼ u½n; x2 ½n ¼ :5n u½n:

1 X n¼1

! X ðejO Þ and x ½n

! X ðejO Þ:

DTFT

DTFT

8.5.2 Derive an expression for the convolution y½n ¼ x½n x½n; x½n ¼ an u½n.

ð1=2Þjnj ; b:

1 X

nð1=2Þn :

n¼0

by using the central ordinate theorems. 8.8.1 Find the DFTs of the sequences a: fx½ng ¼ ½1; 1; 1; 1; b: fx½ng ¼ ½1; 1; 1; 1

8.5.3 Show that DFTf:5d½n þ :25d½n 2 þ :25d½n þ 2g ¼ cos2 ðOÞ: 8.5.4 Find the inverse transform of XðejO Þ ¼ 1; jOj Oc ; XðejO Þ ¼ 0; Oc 5jOj p: 8.5.5 Consider the two-sided sequence jnj 5 x½n ¼ a ; jaj 1. Write this expression in terms of the right-side and left-side sequences. Then, derive the expression for the DTFT of this sequence using the time reversal property. Be careful about the sample point at n ¼ 0: 8.5.6 Use the central ordinate theorems to evaluate the sums. a:

1 X n¼0

nan ; b:

1 X n¼1

ajnj ; c:

1 X n¼1

sin

WnÞ ðpnÞ

8.8.2 Compute the DFTs of the following N- point sequences. For Part c., use Euler’s formula for the cosine function in determining the DFT assuming k0 is an integer. a: x½n ¼ an ; 0 n5N; b: x½n ¼ u½n u½n n0 ; 05n0 5N; c: x½n ¼ cosðno0 Þ; o0 ¼ 2pk0 =N; 0 n5N 1; k0 is an integer. 8.8.3 Determine the 8-point DFT sequence of x½n ¼ d½n þ 2d½n 3. 8.8.4 Consider a sequence x½n; 0 n N 1 with X½k ¼ DFTfx½ng. Find the DFTs the two sequences given below in terms of X½k:

Problems

365

fy½ng ¼ ¼

bandwidth of xðtÞ as the frequency where jXðjoÞj is 10% of its maximum.

x½n=2; n even ; fy½ng 0; n odd x½n; n ¼ 0; 1; 2; . . . ; N 1 0; n ¼ N; N þ 1; . . . ; ð2 N 1Þ

:

8.8.5 Compute the DFT of x½n 3 mod ðNÞ directly and then using the time-shift theorem. 8.8.6 Determine y½n ¼ x½n h½n directly and then using the DFT for the sequences x½n ¼ ð1=3Þn ; h½n ¼ sinððp=2ÞnÞ; n ¼ 0; 1; 2; 3: 8.8.7 Show that the DFT of a real sequence x½n satisfies the relation X½N k ¼ X ½k: (*) denotes conjugation. 8.8.8 The DFT sequence of a real time signal is given by fX½kg ¼ f4; j; 0; Xg, where X is the missing value. Use the symmetry property of DFT to determine the missing value. Find the corresponding time sequence. 8.8.9 Derive an expression DFT½ð1Þn x½n in terms of X½k. 8.9.1 Show that for N ¼ 4, (an identity matrix)

DFT

½y½n ¼

ð1=NÞADFT ADFT

¼ IN

ð1=NÞA2DFT

8.9.2 Derive the matrix with N ¼ 4. What can you say about this matrix? 8.9.3 Given xðtÞ ¼ LðtÞ, estimate the sampling frequency and sampling interval by choosing the

8.10.1 Use Example 8.10.3 to compare the number of multiplications required to compute the convolution directly and by using the DFT. 8.10.2 Write the sequence r½n in matrix form r½n ¼

N1 X

x½ix½n þ iModðNÞ :

i¼0

8.10.3 Consider the two discrete N-point real sequences xd1 ½n and xd2 ½n and x½n ¼ x1 ½nþ jx2 ½n; n ¼ 0; 1; . . . ; N 1 with F½xi ½n ¼ Xi ½k; i ¼ 1; 2 and ðxd1 ÞT ¼ ½ c 0123; ðxd2 ÞT ¼ ½ c 2345; xd ¼ xd1 þ jxd2 : a. First show the following in general terms and then b. verify this using the sequences: X1 ½k ¼ :5fXd ½k þ X ½N kg; X2 ½k ¼ :5jfX½k X ½N kg: 8.10.4 Find the N point DFT of the sequences x1 ½n ¼ ejO0 n for two cases: a: O0 ¼ 2pk0 =N; b: O0 6¼ 2pk0 =N:ðk0 is an integerÞ: 8.10.5 Consider the sequence xd ½n ¼ f0; 1; 0; 1g. Compute its DFT and then use the interpolation technique discussed in Section 8.10 assuming M ¼ 2 and 4.

Chapter 9

Discrete Data Systems

9.1 Introduction In the last chapter we have discussed the concepts of discrete Fourier transforms (DFTs). In this chapter we will briefly review these and discuss its fast implementations. There are several algorithms that come under the topic-fast Fourier transforms (FFTs). The first FFTmethod of computing the DFT was developed by Cooley and Tukey (1965). These are innovative and useful in the signal processing area. Continuous Fourier transforms (CFTs) in the analog and the discrete Fourier transforms (DFTs) in the discrete domains are the corner stones of signal analysis. In the continuous domain we studied the Laplace transforms, which are related to the continuous Fourier transforms. The discrete counter part of the Laplace transforms is z-transforms related to the discrete-time Fourier transforms (DTFTs). Table 9.1.1 summarizes the variables in the continuous-time Fourier transforms, the Laplace transforms, the discrete-time Fourier transforms, and the z-transforms. In this chapter we will study some of the basics associated with the z-transforms and its applications. Digital filters have been popular in recent years and will continue to be in the future. In its simplest

Table 9.1.1 Discrete-time and continuous-time signals and their transforms Continuous-time transform/ Discrete-time transform/ variable variable Continuous Fourier transform/ o or f Laplace transform/s

Discrete time Fourier transform/O z-transform/z

form, a digital filter is a computer program that takes a set of data and converts into another set of data. Discrete data systems may correspond to filtering or some other operation. In the analog case we have to worry about component value tolerances and the responses can change in time. The responses of analog systems cannot be duplicated, as the component values may be different from one batch to another. The responses of the filters can change if the operating conditions of the filter change. On the other hand, in the digital case, every time we process a set of data the output will be the same. Digital filters are more flexible and can be altered by simply changing the computer code. At low frequencies, analog components are bulky. We may have to deal with magnetic coupling if inductors or transforms are used as components in the analog system. Analog filters may have to be redesigned and the circuit implementations may be different if the frequencies change. On the other hand, modifying digital filters may represent a change of computer code. Digital technology is modern and powerful signal processing algorithms can be designed. Digital filters can be time shared and process several signals simultaneously. Digital integrated circuits design is much simpler compared to analog integrated circuit technology. They require lower power consumption and the digital circuitry can be fabricated in smaller packages. Digital storing is much cheaper. Searching and selecting digital information is simple and processing the data is straightforward. Digital reproduction is much more reliable and the cost of digital hardware continues to come down every year. Most source signals and the recipients are analog in nature. To replace an analog filter by a digital filter, the analog signal

R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_9, Springer ScienceþBusiness Media, LLC 2010

367

368

9 Discrete Data Systems

needs to be converted to a digital signal by using an analog-to-digital (A/D) converter. The digital signal is then passed through a digital filter and the output of the filter needs to be converted back to analog data using a digital-to-analog (D/A) converter. Digital signal processing (DSP) area has been popular during the past 30+ years. It will continue to be of interest in many areas, including seismic signal processing, speech processing, image processing, radar signal processing, and others. Telephone industry has taken the lead in the signal processing area. There are excellent texts available in the general area of signal processing. Some of these include Ambardar (2007), Strum and Kirk (1988), Mitra (1998), Oppenheim and Schafer (1975), Rabiner and Gold (1975), Cartinhour (2000), Ludeman (1986), and many others. For an excellent review on the spectral analysis, see Otnes and Enochson (1972), Marple, (1989), Press et al. (1989), and many others. For a historical survey on the spectral estimation, see Robinson (1982). MATLAB provides digital analysis and design software, see Ingle and Proakis (2007). Also, see Ramirez (1975), Smith III (2007), Smith, (2002) on FFT and its applications.

1 jXð0Þj2 ; N2 1 Pðos =2Þ ¼ 2 jXðN=2Þj2 ; N os ¼ 2pfs ; fs ¼ 1=ts i 1 h Pðok Þ ¼ 2 jX½kj2 þjX½N kj2 ; N 2pk N ; k ¼ 1; 2; :::; 1: ok ¼ Nts 2 Pð0Þ ¼

(9:2:1b)

From Chapter 4 we note that a rectangular window spectrum has a great deal of leakage into the side lobes. A tapered window w½n, such as a Hamming window to be discussed later, can be used in estimating the spectrum to reduce the spectral leakage. A windowed signal y½n ¼ x½nw½n is to be used in the estimation. Another popular method of spectral estimation is the Blackman–Tukey method, see Press et al. (1990). In its simplest form, it involves the computation of the data autocorrelation and then determining the spectrum using DFT. The spectrum of the autocorrelation is the power spectral density.

9.2.1 Symbolic Diagrams in Discrete-Time Representations 9.2 Computation of Discrete Fourier Transforms (DFTs) Power spectrum: Most signals in practice are analog signals. Spectral analysis and estimation of these signals is basic. A simple method of power spectrum estimation of an analog signal xðtÞ involves N values of xðtÞ sampled every ts s (or fs ¼ 1=ts samples/s) resulting in x½n ¼ xðnts Þ; n ¼ 0; 1; 2; . . . ; N 1. The DFT of the signal x½n is (see Section 8.9) X½k ¼

N1 X

Symbolic diagrams or signal flow graphsare a network of directed branches connected at nodes is a pictorial representation of an algorithm. Figure 9.2.1 gives the flow graph symbols that are common in two different forms. Source nodes do not have any incoming braches and are used for input x[n]

+

x[n]

y[n]

×

ax[n]

x[n] z–1

x[n −1]

a

(a)

(b) x[n] + y[n]

x[n]

x½nejð2p=NÞnk ;

x[n] + y[n]

x[n]

a

(c) ax[n]

x[n]

z −1 x [n−1]

n¼0

k ¼ 0; 1; 2; . . . ; N 1:

(9:2:1a)

y[n]

(d) The power spectrum estimate is defined at ðN=2Þ þ 1 frequencies by

(e)

(f)

Fig. 9.2.1 Two flow graph representations: (a) and (d), summers; (b) and (e), multipliers; (c) and (f), delays

9.2 Computation of Discrete Fourier Transforms (DFTs) Fig. 9.2.2 Example 9.2.1

w[n]

+

x

x[n]

369

+

w[n] y[n]

x[n]

a

−1

z

a

x

y[n] −1

z

x c

b

b

(b)

9.2.2 Fast Fourier Transforms (FFTs)

sequences. Sink node has only one entering branch and is used for the output sequence. In addition, summers and multiplier symbols are shown and are self-explanatory. The symbols z1 are used to identify delay components. In Section 9.4 z-transforms will be studied. Some authors use the multiplier constant above the line and others use it below the line. If the multiplier constant a is not shown, then it is 1.

First a brief review of the discrete Fourier transform (DFT) is given below. The discrete Fourier transform is a transformation that takes a set of N values in time to N values in frequency. First, the transform vector is given by X ¼ ADFT x (see (8.9.16a and b)), where the matrix ADFT is a N N matrix with its ðk; nÞ entry being (see (8.9.17))

Example 9.2.1 Using the symbol representations in Fig. 9.2.1, write difference equations relating the variables w½n and y½n in terms of x½n and w½n 1 in Fig. 9.2.2a,b.

ADFT ðk; nÞ ¼ ejð2p=NÞðk1Þðn1Þ ðk1Þðn1Þ

WN

;

1 k; n N; WN ¼ ej2p=N : (9:2:2a)

Solution: The two diagrams in Fig. 9.2.2 result in the same equations and are given as follows: w½n ¼ ax½n þ bw½n 1;

(a)

c

The vectors X and x are N-dimensional column vectors. In matrix form the DFT coefficients can be expressed in terms of WNn ¼ ðej2p=N Þn by

y½n ¼ w½n þ cw½n 1: &

2 6 6 6 6 6 6 6 6 4

2

3

X½0 X½1

1

1

7 6 1 W1N 7 6 7 6 X½2 7 6 1 W2N 7¼6 6 7 : : : 7 6 7 6 5 6 : : : 4 N1 X½N 1 1 WN

1

:

:

W2N

:

:

W4N :

: :

: :

: :

: :

: :

Note WNn takes one of the values in the set jð2p=NÞ ; ejð2p=NÞ2 ; :::; ejð2p=NÞðN1Þ for any N, 1; e see (8.9.18).

32 3 x½0 7 N1 WN 76 x½1 7 7 76 7 6 2ðN1Þ 76 WN 76 x½2 7 7: 76 7 76 : : 7 76 7 74 5 : : 5 ðN1Þ2 x½N 1 WN 1

The properties in Table 9.2.1 allow for the derivation of a fast Fourier transform (FFT) algorithm. We will consider an Nð¼ 2n Þ-point decimation-in-

Table 9.2.1 Properties of the function WN ¼ ejð2p=NÞ 1. WNnþN ¼ ejð2p=NÞðnþNÞ ¼ ejð2p=NÞn ¼ WNn 2.

nþN=2 WN

3.

WNkN

¼ e

¼e

jð2p=NÞ

jð2pÞk

¼ 1;

¼

WNn

2 A ¼ 1; e

;e

(9.2.3a) (9.2.3b) (9.2.3c)

k is an integer

W2Nk ¼ ej 2ð2p=NÞk ¼ ej;ð2p=ðN=2ÞÞk ¼ WkN=2 jð2p=NÞ jð2p=NÞ2 jð2p=NÞn jð2p=NÞðN1Þ

4. 5. e

(9:2:2b)

; :::; e

(9.2.3d) (9.2.3e)

370

9 Discrete Data Systems

frequency FFT algorithm. It is based on expressing one N-point DFT algorithm by two N/2-point DFTs, then four (N/4)-point DFTs, and so on. The algorithm at the end reduces to ðN=2Þ-2-point transforms. The 1-point transform is trivial as X½0 ¼ x½0: The DFT of a 2-point sequence is determined by noting W12 ¼ ej2p=2 ¼ 1: The DFT values in scalar and matrix form are as follows: X½0 ¼ x½0W02 þ x½1W02 ¼ x½0 þ x½1 ; X½1 ¼ x½0W02 þ x½1W12 ¼ x½0 x½1

X½0 1 ¼ X½1 1

1 1

x½0 : x½1

(9:2:4a)

N1 X

x½nejð2p=NÞnk

n¼0

¼

N1 X

x½nWNnk

(9:2:5)

n¼0

¼

N=21 X

x½nWNnk þ

n¼0

N1 X

x½nWNkn ¼

n¼N=2

N=21 X m¼0

n¼N=2

¼

N=21 X m¼0

X½k ¼

x

N=21 X n¼0

N kðmþNÞ þ m WN 2 2

N kðN=2Þ þ m WNmk WN ; (9:2:6) x 2

N x½n þ ð1Þk x n þ WNkn ; 2

k ¼ 0; 1; 2; :::; N 1:

N nð2 kÞ WN x½n þ ð1Þ x n þ X½2 k ¼ 2 n¼0 N=21 X N Wnk ¼ x½n þ x n þ (9:2:8a) N=2 ; 2 n¼0 N=21 X N nð2kþ1Þ X½2k þ 1 ¼ x½n þ ð1Þð2kþ1Þ x n þ WN 2 n¼0 2k

N nk WN=2 x½n x n þ WNn ; 2 n¼0 (9:2:8b) N k ¼ 0; 1; 2; . . . ; 1: 2 ¼

N=21 X

One N-point DFT is reduced to two ðN=2Þ-point DFTs. One N-point DFT requires N2 multiplications and ððN 1Þ additions), see (9.2.2b). Two ðN=2Þ-point DFTs require only 2ðN=2Þ2 multiplications and 2ðN 1Þ additions. Example 9.2.2 Assuming N ¼ 4, show that the use of (9.2.8a and b) successively results in the DFT values. Illustrate the algorithm using the flow graph representation. Solution: a. From (9.2.8a) and (9.2.8b), we have (note W40 ¼ 1; W14 ¼ j; and W42 ¼ 1)

x½nWNnk :

Note the variable n is used for the time variable and k is used for the frequency variable Using m ¼ n ðN=2Þ in the second summation in (9.2.5) kðN=2Þ ¼ ejð2p=NÞkðN=2Þ ¼ ð1Þk result and noting WN in N1 X

N=21 X

(9:2:4b)

Decimation-in-frequency FFT algorithm: Starting with the DFT of a set of data x½n; n ¼ 0; 1; ::; N 1 and N ¼ 2n , the DFT coefficients (see 9.2.1a.), X½k, are obtained in terms of WN ¼ ej2p=N as follows: X½k ¼

Now separate the coefficients into X½2 k and X½2 k þ 1 and use Table 9.2.1:

X½2 k ¼

1 X

fx½n þ x½n þ 2gW2nk 4 ;

n¼0

X½2 k þ 1 ¼

1 X

fx½n x½n þ 2gWn4 W2nk 4 ; k ¼ 0; 1:

n¼0

(9:2:9) First, at stage 0, i.e., to start with, define x0 ½n ¼ x½n; n ¼ 0; 1; 2; 3: At stage i, identify the variables as xi ½n. Algorithm has two stages corresponding to N ¼ 4 ¼ 2n ; n ¼ 2. From (9.2.9), we have the following. Direct: X½0 ¼ fx0 ½0 þ x0 ½2g þ fx0 ½1 þ x0 ½3g

(9:2:7)

¼ fx1 ½0g þ fx1 ½1g ¼ x2 ½0 ¼ X½0 ; (9:2:10a)

9.2 Computation of Discrete Fourier Transforms (DFTs)

X½2 ¼ fx0 ½0 þ x0 ½2g fx0 ½1 þ x½3g ¼ fx1 ½0g fx1 ½1g ¼ x2 ½1 ¼ X½2 ; (9:2:10b) X½1 ¼ fx0 ½0 x0 ½2g þ fW 40 ðx0 ½1 x½3Þg ¼ fx1 ½2g þ fx1 ½3g ¼ x2 ½2 ¼ X½1 ;

(9:2:10c)

X½3 ¼ fx0 ½0 x0 ½2g fW14 ðx0 ½1 x0 ½3Þg ¼ fx1 ½2g fx1 ½3g ¼ x2 ½3 ¼ X½3 : (9:2:10d) Individual identifications at each stage from (9.2.10): Stage 0: x0 ½0 ¼ x½0; x0 ½1 ¼ x½1; x0 ½2 ¼ x½2:

(9:2:11a)

Stage 1: x1 ½0 ¼ fx0 ½0 þ x0 ½2g ; x1 ½1 ¼ fx0 ½1 þ x0 ½3g ; x1 ½2 ¼ fx0 ½0 x0 ½2g ; x1 ½3 ¼ fW04 ðx0 ½1 x½3Þg : (9:2:11b) Stage 2: x2 ½0 ¼ fx1 ½0g þ fx1 ½1g ; x2 ½1 ¼ fx1 ½0g fx1 ½1g ; x2 ½2 ¼ fx1 ½2g þ fx1 ½3g ; x2 ½3 ¼ fx1 ½2g fx1 ½3g : (9:2:11c) End results: X½0 ¼ x2 ½0; X½2 ¼ x2 ½1; X½1 ¼ x2 ½2; X½3 ¼ x2 ½3 : (9:2:11d)

Fig. 9.2.3 Flow graph representations for N ¼ 4 using the decimation-infrequency FFT algorithm

371

These equations can be used to draw the flow graph using the symbols in Fig. 9.2.3. For clarity, the multipliers are shown under the lines rather than above. Interestingly, the above equations can be seen from the flow graph in Fig. 9.2.3. Interestingly, if the variables arein binary form, the argument k in X½k ¼ X½ðk1 k0 Þ2 is related to the argument n in x2 ½ðn1 n0 Þ2 by the relation k ¼ ðk1 ¼ n0 ; k0 ¼ n1 Þ: For additional information & on this, see Oppenheim and Schafer (1999). The above results can be extended for any N ¼ 2n with n stages. Figure 9.2.4 gives the flow graph for N ¼ 8 ¼ 23 . Note the multipliers are identified above and below the lines for clarity. For a general derivation of the decimation-infrequency algorithm and other algorithms, see Oppenheim and Schafer (1999), and others. Notes: The decimation refers to the process of reducing the number of operations for an N ¼ 2n point DFT, expressing the N-point DFT in terms of 2 ðN=2Þ ¼ 2n1 -point DFTs and successively expressing them in n stages with the input sequence in natural order. Number of computations in an FFT algorithm: In Section 8.9.2 the computational aspects of discrete Fourier transforms were considered. These results are compared with FFT computational requirements. In the N-point FFTalgorithm with N ¼ 2n , we have n ¼ log2 ðNÞ stages. FFT computation requires ðN=2Þn ¼ ðN=2Þ log2 ðNÞ complex multiplications and nN ¼ N log2 ðNÞ complex additions. Computers use real arithmetic and each complex multiplication requires four real multiplications and three real additions. The amount of effort to do multiplication is much larger than additions. We

372

9 Discrete Data Systems

Fig. 9.2.4 Flow graph representations for N ¼ 8 using the decimation-infrequency FFT algorithm

can compare the number of multiplications by the direct method versus FFT by the ratio R¼

N2 N N ¼ : :5 N log2 ðNÞ :5n n

(9:2:12)

For a large N ¼ 2n , the difference in the number of computations by FFT is significantly lower. Note that :5 N log2 ðNÞ is nearly linear, whereas N2 is quadratic. For small N, the difference in the number of computations in computing the DFT and FFTis not that significant. As an example, consider N ¼ 210 , the ratio in (9.2.12) is R 204. The DFT requires N2 values of Wkn N ; k; n ¼ 0; 1; 2; . . . ; N 1, whereas FFT requires at most N such values at each stage. Earlier, we have seen that ejð2p=NÞnk ¼ ejð2p=NÞm ; 0 m N 1. The logical way of course is to compute WNk once, k ¼ 0; 1; 2; . . . ; N 1, store them, and use them again and again in each stage. Only about ð3=4ÞN of these WkN are distinct in the FFT algorithm, see Ambardar (2007). The FFT approach N ¼ 2n is computationally efficient compared to the direct method only for n45 ðN432Þ. See Wilf (1986) for an interesting discussion of algorithms and their complexity. MATLAB function for computing the DFT of a signal is the fft function. It can be used for any N. For example, to compute the DFT of a sequence x of N values, MATLAB routine is X ¼ fftðxÞ to get the spectral values and the routine x ¼ ifftðXÞ gives the data from the spectral values.

Just like in the continuous case, other discrete transforms related to discrete Fourier transforms can be considered, including discrete cosine, sine, Hartley, and Hilbert transforms. These are beyond the scope here. See the handbook by Poularikas (1996).

9.3 DFT (FFT) Applications In Section 9.2, spectral analysis based on DFT was considered. Computing DFT via FFT is a tool to reduce the number of computations. FFT is applicable wherever DFT can be used. See, for example, Marple (1987), O’Shaughnessy (1987), Otnes and Enochson (1972), Poularikas (1996), Rabiner and Schafer (1979), Shenoi (1995), and others for FFT applications.

9.3.1 Hidden Periodicity in a Signal Although, nothing is forever, some signals can be considered as periodic at least on a short-time basis. For example, vowel speech sounds can be considered as periodic on a short-time basis. Investors in the stock market would like to know if the price of a stock has a periodic part in the signal that is hidden. If so, the investor can sell when the stock is high and buy when it is low. For a good presentation on

9.3 DFT (FFT) Applications

373

applying spectral analysis to various physical signals, see Marple (1987). Example 9.3.1 Consider the sinusoid xðtÞ ¼ cosð2pð100ÞtÞ that is sampled at twice the Nyquist rate for three full periods. Find the corresponding DFT values. Solution: The frequency of the sinusoid is 100 Hz. The period of the signal is T ¼ ð1=100Þ s. The Nyquist rate is 200 Hz. The corresponding sampling rate and the sampling interval are 400 Hz and ts ¼ ð1=400Þ s. Since three periods are used, we have four samples per period and have N ¼ 12 samples to find the DFT. Note cosð2pf0 tÞjt¼nts ¼ cosð2pnðf0 =fs ÞÞ; n ¼ 0; 1; 2; . . . ; N 1 ¼ 11

(9:3:1)

The sampling frequency is divided into N ¼ 12 intervals and the frequency interval is F ¼ fs =N ¼ fs =12 ¼ 400=12 ¼ 100=3 referred to as the digital frequency in Section 8.6.1, where O ¼ 2pðf=fs Þ ¼ 2pF was used. The DFT frequencies are kF ¼ kfs =N ¼ kðfs =12Þ; k ¼ 0; 1; 2; :::; N 1 ¼ 11:

(9:3:2)

Now, assume f0 is one of these frequencies and kF ¼ kðfs =NÞ ¼ kðfs =12Þ ¼ f0 for some k. That is, f0 k k ¼ ¼ ; a rational number: fs N 12

(9:3:3)

For f0 ¼ 100 Hz; fs ¼ 400 Hz, and k ¼ 3, X½3 gives the appropriate spectral value. This can be verified by computing DFT of the sequence x½n ¼ cosð2pðf0 =fs ÞnÞ ¼ cosð2pðknÞ=NÞ ¼ cosð2pðknÞ=12Þ; n ¼ 0; 1; 2; . . . ; 11; k ¼ 3 ) fx½ng ¼ f1; 0; 1; 0; 1; 0; 1; 0; 1; 0; 1; 0g: (9:3:4) DFTfcosð2pðnkÞ=NÞg ¼ 12 DFT ej2pnm=N þ ej2pnm=N ¼ 12

N1 P n¼0

ejð2p=NÞnðmkÞ þ 12

N1 P n¼0

: ejð2p=NÞnðmþkÞ ¼ X½k

Using the summation formula for the geometric series results in the DFT values

X½k ¼

8 > < > :

N=2; k ¼ m N=2; k ¼ N m 0; otherwise

9 > = > ;

) X½k : f0; 0; 0; 6; 0; 0; 0; 0; 0; 6; 0; 0g: (9:3:5) Noting N ¼ 12 and the amplitude of the sinusoid is & 1, we have X½3 ¼ X½9 ¼ 6. Notes: These results can be extended to a periodic function xðtÞ ¼ cosð2pðk0 =NÞ þ yÞ. The DFT of xðtÞ has only two nonzero discrete frequency values and are X½kjk¼k0 ¼ðN=2Þejy and X½N kjk¼k0 ¼ ðN=2Þejy : The frequency spacing F ¼ f0 =fs needs to be a rational function so that the discrete frequency falls on the input signal frequency. FFT algorithm can be used with this in mind. Also, leakage results in DFT if a periodic signal is not sampled for an integer number of periods. This results in nonzero spectral components at frequencies other than the harmonic frequencies of the signal. Many times it is not easy to find the period of a signal up front. One solution to this is to use large enough number of samples. Larger is the time interval, the more closely is the spectrum sampled. That is, the spectral spacing F ¼ fs =N is reduced, thus giving a more accurate estimate of spectrum of the given signal. Most practical signals are noise corrupted and are known only for a short time. That is, the signal is a windowed signal. As we have seen in Chapter 4, the spectrum of the windowed sinusoid is not a spike, but a sinc function indicating leakage into the side lobes. Tapered windows need to be used in any spectral analysis. If a signal contains several frequencies, then the spectrum of the windowed signal is the sum of the spectra of each sinusoid and it may not have distinct peaks as the main lobes of the sinusoids might merge. This happens if the two frequencies in the input signal are located close enough, then the frequency peaks may merge. Instead of two separate peaks, there may be only one single peak. Choosing larger

374

9 Discrete Data Systems

DFT lengths and good windows improves the accuracy of the spectral estimates, see Marple (1987). &

9.3.2 Convolution of Time-Limited Sequences The convolution of the sequences x½n and h½n was defined by (see (8.3.16)) y½n ¼

1 X

x½kh½n k ¼

k¼1

1 X

h½mx½n m:

Consider x[n] = 0 for n < 0 and n L, h[n] = 0 for n < 0 and n M. Equation (9.3.6) can be expressed as follows: L1 X

x½kh½n k ¼

k¼0

(9:3:10) Solution: 2 3 y½n ¼ 0; n5 0 6 7 6 y½0 ¼ x½0h½0 þ x½1h½1 ¼ x½0h½0 ¼ 1 7 6 7 6 7 6 y½1 ¼ x½0h½1 þ x½1h½0 ¼ 1 þ 1 ¼ 0 7: 6 7 6 7 6 y½2 ¼ x½0h½2 þ x½1h½1 ¼ x½1h½1 ¼ 1 7 4 5 y½n ¼ 0; n 3 (9:3:11)

m¼1

(9:3:6)

y½n ¼

x½0 ¼ 1; x½1 ¼ 1; x½n ¼ 0 for n50 and n41:

M 1 X

h½mx½n m: (9:3:7)

m¼0

The equations in (9.3.11) for y½n 6¼ 0 can be written in two equivalent matrix forms and 2

y½0

3

2

7 6 6 y ¼ 4 y½1 5 ¼ 4 h½1 0 y½2 2

x½0

6 ¼ Hx ¼ 4 x½1 0

Sequence:

h½0

0

3

7 xð0Þ h½0 5 xð1Þ h½1

3 7 h½0 x½0 5 ¼ Xh: h½1 x½1 0

(9:3:12)

The coefficient matrices X and H are ð3 2Þ size matrices. The column vectors y; h; and x are of the dimensions 3 1; 2 1; and 2 1, respectively. The entries in the coefficient matrices X and H have a special structure. For example, the first and the second columns of the coefficient matrix H are, respectively, given by

y½n ¼ 0; n50 y½0 ¼ h½0x½0 y½1 ¼ h½1x½0 þ h½0x½1 y½2 ¼ h½2x½0 þ h½1x½1 þ h½0x½2 ::: y½M 1 ¼ h½M 1x½0þ ::: þ h½0x½M 1:::

2

y½M þ L 2 ¼ h½M 1x½L 1 y½n ¼ 0; n M þ L 1: (9:3:8)

h½0

3

0

3

7 6 Colðh½0; h½1; 0Þ ¼ 4 h½1 5; 0 2

Example 9.3.2 Find the sequence y½n corresponding to the convolution of the following sequences and express y½n in matrix form:

7 6 Colð0; h½0; h½1Þ ¼ 4 h½0 5: h½1

h½0 ¼ 1; h½1 ¼ 1; h½n ¼ 0 for n50 and n 4 1;

The second column is obtained from the first by rotating the first entry to the second, second entry & to the third, and the third entry to the first.

(9:3:9)

(9:3:13)

9.3 DFT (FFT) Applications

375

Equations in (9.3.7) can be written in matrix and symbolic forms for y½n 6¼ 0 as follows: 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4

3

y½0

2

0

h½0

0

:

0

:

3

7 7 6 7 7 6 h½1 h½0 0 : : 0 7 7 6 7 7 6 h½2 h½1 h½0 : : 0 7 7 6 72 7 6 3 7 7 6 : x½0 : h½2 h½1 : : 7 7 6 76 x½1 7 7 6 : : : : : : 76 7 7 6 76 7 7 6 76 7 7 6 : : : : : : : 76 7; y ¼ Hx: 7¼6 6 7 6 7 7 : y½M 1 7 6 h½M 1 h½M 2 h½M 3 : : : 76 7 76 7 7 6 74 5 7 6 2 : : : 0 h½M 1 h½M y½M : 7 7 6 6 7 7 : 0 h½M 1 : : : : 7 x½L 1 7 6 7 7 6 7 7 6 : : : : : : : 7 7 6 6 7 : : : : : h½M 2 7 : 5 5 4 0 0 0 0 0 h½M 1 y½M þ L 2 y½1 y½2

Note that x is an L1 column matrix with the entries x½0; x½1; . . . ; x½L 1. The vector y is a ðM þ L 1Þ 1 column matrix and H is a ðM þ L 1Þ L matrix. The entries in the matrix H have the following pattern. The j th column in H is given by 9 8 ðj 1Þ zeros > > > > > = < N coefficients written in the order > ; 1 j L: > > h½0; h½1; . . . ; h½M 1 > > > > ; : ðL jÞ zeros (9:3:14b) If one column or one row of H is known, the entire matrix can be constructed. Noting that the convolution is commutative, i.e., y½n ¼ h½n x½n ¼ x½n h½n, equations similar to (9.3.14a and b) can be written by replacing h½n by x½n and vice versa. Computation of convolution via DFT: In Section 8.10, it was shown that a periodic convolution can be implemented by using the DFT. This idea can be used here as well and is illustrated by a simple example. Equation in (9.3.12) can be written as 2

y½0

3

2

h½0

7 6 6 y ¼ 4 y½1 5 ¼ 4 h½1 0 y½2

0

h½1

32

x½0

3

7 76 h½0 0 54 x½1 5 (9:3:15) 0 h½1 h½0

(9:3:14a)

This set of equations gives the same results as the set in (9.3.12). Interestingly, (9.3.15) is the same as the one in (8.10.28), with h½2 ¼ 0 corresponding to a periodic convolution. Equation (9.3.15) can be modified by adding the fourth column in the coefficient matrix and appending the data by two zeros resulting in (9.3.16). 2

3 2 32 3 y½0 h½0 0 0 h½1 x½0 6 y½1 7 6 h½1 h½0 0 6 7 0 7 6 7 6 76 x½1 7 ya ¼ 6 7¼6 76 7 ¼ Ha xa : 4 y½2 5 4 0 h½1 h½0 0 54 0 5 0

0

0

h½1 h½0

0 (9:3:16)

The reason for using N ¼ 4 for the extended sequences is that FFT can be used to find the convolution. In summary, given h½n; n ¼ 0; 1; . . . ; M 1 and x½n; n ¼ 0; 1; 2; . . . ; L 1, we can convolve h½n with x½n using the appended sequences as follows. Pad the sequences h½n and x½n with zeros so that they are of length N L þ M 1 resulting in the appended sequences ha ½n and xa ½n of length N, a power of 2. Convolution of the two extended sequences results in ya ½n ¼ ha ½n xa ½n. ya ½n is an appended sequence of y½n of length 2 N 1. To

376

9 Discrete Data Systems

determine the convolution via DFT (or FFT), do the following steps: 1. Determine Xa ½k ¼ DFTfxa ½ng and Ha ½k ¼ DFT fha ½ng . 2. Multiply the DFTs to form the products Ya ½k ¼ Ha ½kXa ½k. 3. Find the inverse DFT of Ya ½k. Discard the last ðN ðM þ L 1ÞÞ data points out of ya ½n to obtain y½n; n ¼ 0; 1; :::; N þ L 2: There are two methods for computing the discrete convolution, one by the convolution formula and the other by using DFTs. It may appear that the effort in computing the convolution via DFT is more computationally intensive than the direct convolution. However, using FFT, the number of computations is fewer, roughly for N432. Even though the number of computations is fewer for large N, there are difficulties with the use of DFT. The data sequence x½n may be long, say M points of data. The sequence h½n is of reasonable size, say L M. The computation of the DFTof x½n may not be possible due to computer storage constraints involving large amount of computations resulting in a significant delay. There are two methods, overlap-add and overlap-save, which can be used for large set of data. Both section long input sequence into shorter sections. They are suitable for online implementation if the process can tolerate slight delays. The overlap-add method is discussed below; see Ambardar (2007) for the overlap-save method. Overlap-add method: The sequence h½n of length Nis assumed to start at n ¼ 0. The sequence x½n is a much longer sequence of length M, also starting n ¼ 0. Partition x½n into k segments each of length N (zero padding the last segment if needed). The data can be expressed in a mathematical form using a rectangular window wR ½n of length N by

x½n ¼

k1 X

1;

n ¼ 0; 1; . . . ; N 1

0; otherwise

:

fx½ng ¼ ffx0 ½ng; fx2 ½ng; . . . ; fxk1 ½ngg; xi ½n x½n Ni; i ¼ 0; 1; . . . ; k 1 ¼ : 0; elsewhere (9:3:17b) We can now write y½n ¼ h½n x½n ¼ h½n

k1 X

xi ½n

i¼0

¼

k1 X i¼0

h½n xi ½n ¼

k1 X

yi ½n:

i¼0

It follows that the total convolution is the sum of the individual convolutions resulting in y½n ¼ y0 ½n þ y1 ½n N þ ::: þ yk1 ½n ðk 1ÞN:

(9:3:17a)

(9:3:18)

The ith segment of the output begins at n ¼ iN, as do the input segment xi ½n. However, each yi ½n segment has a length equal to ð2 N 1Þ and therefore yi ½n s ‘‘overlap’’ each other. We can think of each yi ½n as having the same length of 2 N 1 points, where each yi ½n includes zero padding before and/ or after as appropriate, such that the positions of the sequences are in correct location. Example 9.3.3 Consider the longer and shorter sequences given by x½n ¼ f1; 2; 3; 4g and h½n ¼ f1; 1g. a. Use the overlap-add method to determine the convolution of the sequences. b. Verify the results using direct convolution. Solution: a. Here M ¼ 2 and L ¼ 4. Section the sequence x½n into two sequences x0 ½n ¼ f1; 2g and x1 ½n ¼ f3; 4g. Then determine y0 ½n ¼ x0 ½n h½n and y1 ½n ¼ x1 ½n h½n. By the direct convolution, the sequences are as follows: 2

xi ½n; xi ½n ¼ x½nw½n iN;

i¼0

w½n ¼

That is,

3 2 3 2 3 2 3 1 0 1 3 3 0 6 7 1 6 7 6 7 1 6 7 ¼ 4 1 5; 4 4 3 5 ¼4 1 5 42 15 1 1 0 2 2 0 4 4 ) y0 ½n ¼ f1; 1; 2g; y1 ½n ¼ f3; 1; 4g:

9.3 DFT (FFT) Applications

377

Since the length of the sequence y½n is ð4 þ 2 1Þ ¼ 5, the sequence y0 ½n needs to be padded by two zeros at the end. Also, pad two zeros before the sequence y1 ½n. The result is obtained by overlapping the data and adding them at appropriate locations given below: y½n ¼ y0 ½n þ y1 ½n N

rhx ½k ¼ h½k x½k ¼

1 X

h½nx½n þ k

n¼1

¼ h½k x½k:

(9:3:19b)

The cross-correlation of two causal sequences x½k and h½k with M and L sample points, respectively, are

¼ f1; 1; 2; 0; 0g þ f0; 0; 3; 1 4g ¼ f1; 1; 1; 1; 4g:

rxh ½k ¼ x½k h½k ¼

M1 X

x½nh½n þ k;

n¼0

b. The result can be verified by directly using the direct convolution given below and the result is the same in both cases. See the equations in (9.3.19a) and (9.3.19b):

rhx ½k ¼ h½k x½k ¼

L1 X

h½nx½n þ k;

(9:3:19c)

n¼0

rhx ½k ¼ rxh ½k: 2

y½ 0

3

2

x½0

6 7 6 6 7 6 6 y½1 7 6 x½1 6 7 6 6 7 6 6 y½0 7 ¼ 6 x½2 6 7 6 6 7 6 6 7 6 6 y½3 7 6 x½3 4 5 4 y½ 4 0 2

1 0

6 6 62 6 6 ¼6 63 6 6 64 4 0

0

3 The integer k represents the shift of the second sequence with respect to the first.

7 7 x½0 7" 7 h½0 # 7 x½ 1 7 7 7 h½0 7 x½2 7 5 x½3 3

2

Example 9.3.4 Consider the data sequences given earlier in Example 9.3.2. Give the correlations of these two sequences and write them in a matrix form.

1

Solution: Using (9.3.20c), we have rhx ½k ¼ 0; k 2 and rhx ½k ¼ 0; k 2:

3

7 7 6 7 7 6 1 7" 6 17 # 7 1 6 7 7 7 6 7 6 ¼6 1 7 27 7: 7 1 7 6 7 7 6 37 6 17 5 5 4 4 4

rhx ½1 ¼ h½0x½1 þ h½1x½0 ¼ h½1x½0; rhx ½0 ¼ h½0x½0 þ h½1x½1; rhx ½1 ¼ h½0x½1 þ h½1x½2 ¼ h½0x½1; rhx ½2 ¼ h½0x½2 þ h½1x½1 ¼ 0: &

9.3.3 Correlation of Discrete Signals Discrete cross-correlations of two sequences (see (8.3.20a and b)) were defined as follows: rxh ½k ¼ x½k h½k ¼

1 X

(9:3:19d)

x½nh½n þ k

2

rhx ½1

6 6 In matrix form ) 6 rhx ½0 4 rhx ½1

3 2

h½ 1

0

3

" # 7 6 7 x½0 7 6 7 7¼ 6 h½0 h½1 7 5 4 5 x½0 0 h½ 0

2

3 2 3 1 1 0 6 7 1 6 7 ¼ 4 1 1 5 ¼4 0 5 1 0 1 1

n¼1

¼ x½k h½k;

(9:3:19a)

&

378

9 Discrete Data Systems

The matrix equation can be generalized. The nonzero cross-correlations can be written in the following matrix and symbolic forms: 3 2 h½M 1 0 rhx ½ðM 1Þ 6 r ½ðM 2Þ 7 6 h½M 2 h½M 1 7 6 6 hx 7 6 6 7 6 6 : : : 7 6 6 7 6 6 : : : 7 6 6 7 6 6 7 6 6 : : : 7 6 6 7 6 6 h½1 rhx ½0 7 ¼ 6 h½0 6 7 6 6 7 6 6 0 h½0 rhx ½1 7 6 6 7 6 6 : 0 : 7 6 6 7 6 6 7 6 6 : : : 7 6 6 7 6 6 5 4 : : 4 : 2

rhx ½L 1

0

0

: :

: : : :

: :

: : : :

:

: :

: :

: : : :

: :

: : : :

:

: :

:

: :

3

0 0

7 7 7 72 : 3 7 x½0 h½M 1 7 76 76 x½1 7 7 h½M 2 7 7 76 7 6 : 76 7 ) y ¼ Hcorr x; : 76 7 76 : 7 76 : 7 74 5 : 7 : 7 7 x½L 1 7 : 7 7 5 :

(9:3:20)

h½0

where y is a column vector of dimension ðMþL1Þ, x is a column matrix of dimension L, and Hcorr is a rectangular matrix of dimensions ðM þ L 1Þ L. If xðkÞ ¼ hðkÞ then the cross-correlation coefficients are the autocorrelation coefficients.

Power spectral density: The autocorrelation (AC) sequence of x½n plays a major role in spectral estimation, as its power spectral density is jX½kj 2 ¼ Sx ½k. It is

Notes: In comparing (9.3.14a) and (9.3.20), for convolution, the first column of H has h½n in the normal order. For correlation, the first column of Hcorr has h½n in reverse order. In both cases the other columns can be determined from the first column.

Sx ½k ¼

Computation of the cross-correlation using DFT: Given h½n; n ¼ 0; 1; . . . ; M 1 and x½n; n ¼ 0; 1; . . . ; L 1 determine the cross-correlation function X rhx ½n ¼ h½kx½k þ n: (9:3:21)

9.3.4 Discrete Deconvolution

k

Considering the equations for the convolution (see (9.3.14a)) and the cross-correlation (see (9.3.20)), we see that both have the same general form and the same computational procedure can be used for both cases. The following step-by-step procedure can be used. 1. Zero-pad both sequences to length N Lþ M 1. To use FFT, use N a power of 2. 2. Find the DFTs of h½n and x½n. 3. Rhx ½k ¼ H ½kX½k; k ¼ 0; 1; 2; :::; N 1. 4. Find the inverse DFT of Rhx ½k.

N1 X

rx ½nejð2p=NÞnk ¼

þ

rx ½nejð2p=NÞnk

n¼0

n¼ðN1Þ N 1 X

N1 X

rx ½nejð2p=NÞnk rx ½0:

(9:3:22)

n¼0

We have seen in the analog domain when a signal goes through a linear time-invariant system, then the signal is modified by the impulse response of the system. The same is true in the digital domain. The convolution of two sequences that are of finite width was defined earlier and y½n ¼ x½n h½n ¼

n X

h½kx½n k:

(9:3:23)

k¼0

There are three functions x½n; h½n; and y½n. In finding the convolution, x½n and h½n are known and y½n is determined by (9.3.23). In the deconvolution problem, the output sequence y½n and the input data x½n are known and h½n is to be determined. There are four ways to achieve this

9.3 DFT (FFT) Applications

379

goal. These are as follows: 1. recursion, 2. polynomial division, 3. using DFT, and 4. L p deconvolution. Deconvolution by recursion: From (9.3.23) h½0 ¼ y½0=x½0. Now separate the term h½n in (9.3.23) and write in the following form and determine successively the values of h½n for n40: y½n ¼

n X

h½kx½n k ¼ h½nx½0 þ

k¼0

n1 X

h½kx½n k:

k¼0

(9:3:24) Example 9.3.5 In Example 9.3.3 the convolution sequence y½0 ¼ 1; y½1 ¼ 0; y½2 ¼ 1 was computed using the sequences x½0 ¼ 1; x½1 ¼ 1 and h½0 ¼ 1; h½1 ¼ 1. Verify the sequence h½n using the recursion method. Solution:

Minimization of this error in terms of the unknowns h½n is a difficult problem for an arbitrary p. The general solution can only be determined by iterative means, see the articles by Byrd and Payne (1979) and Yarlagadda et al. (1985). There is a simple solution when p ¼ 2, which is used if the noise sequence is from a Gaussian distribution. These problems can be described under the general problem of solving a set of equations that are overdetermined and underdetermined system of equations. In Section A.6 we consider the solutions of overdetermined and underdetermined system of equations. Consider the system of equations in the symbolic matrix form Ah ¼ y:

(9:3:28a)

h½0 ¼ y½0=x½0 ¼ 1;

The least-squares solution to the overdetermined system in (9.3.28a) is (see (A.8.16b))

h½1 ¼ ð1=x½0Þfy½1 h½0x½1g ¼ 0 1 ¼ 1: (9:3:25) &

y ¼ Ah ) ðAT AÞh ¼ AT y ) h ¼ ðAT AÞ1 AT y: (9:3:28b)

This method is not practical in the presence of noise. Deconvolution using polynomial division in terms of z-transforms will be considered in Section 9.8.1. The DFT method makes use of DFTsof the sequences with Y½k ¼ H½kX½k. Then, H½k ¼ Y½k=X½k and its inverse DFT gives h½n. This procedure is similar to the one in the analog domain. It has at least two disadvantages. One of them is X½ks may be zero resulting in division by zero. Also, it is sensitive to noise in the input. Fourth method is based on minimizing the Lp errors discussed in Section 3.3. Deconvolution by Lp methods: The output is assumed to be the convolution of two sequences, say an input sequence x½n, a linear discrete system response sequence given by h½n, and an additive noise sequence e½n. The output is y½n ¼ h½n x½n þ e½n; n ¼ 0; 1; 2; . ..; N 1: (9:3:26) The noise signal can only be described by statistical measures. An interesting error measure is the Lp ; 1 p 1 measure defined by jejp ¼

N 1 X n¼0

ðy½n ðh½n x½nÞÞp :

(9:3:27)

The matrix ½ðAT AÞ1 A] is a pseudo-inverse of the matrix A. The MATLAB routine to compute this inverse is pinvðAÞ ¼ ðAT AÞ1 AT :

(9:3:29)

The inverses of the matrix ðAT AÞ may not exist. In such cases, a diagonal matrix dI, where d is a small positive number, is added to the matrix ðAT AÞ. This is called diagonal loading. An approximate solution of (9.3.28) is then given by h ﬃ ðAT A þ dIÞ1 Ay:

(9:3:30)

Example 9.3.6 Solve the following set of equations using the least-squares solution: 2

3 2 3 1 0 e 6 7 1 6 7 Ah ¼ 4 2 1 5 þ 4 e 5 1 0 2 e 2 3 2 3 2 3 e e 1 6 7 6 7 6 7 ¼ 4 1 5 þ 4 e 5 ¼ y þ 4 e 5: 2

e

e

(9:3:31)

380

9 Discrete Data Systems

Solution: The pseudo-inverse of A and the solution vector are, respectively, given as follows: 1 5 2 1 2 0 ðAT AÞ1 AT ¼ 21 2 5 0 1 2 1 5 8 4 ; ¼ 21 2 1 10 82 3 2 39 1 e > > = < 5 8 4 1 6 7 6 7 h¼ 4 1 5 þ 4 e 5 > 21 2 1 10 > ; : 2 e 21 1 :3333e 1 1 7e ¼ þ ¼ : 21 21 21 7e 1 þ :3333e Clearly if e ¼ 0, the solution coincides with the & vector h we started with. Notes: Implementation of discrete algorithms generally requires multiplications, which are expensive compared to additions. The following table gives a rough comparison of how expensive the additions, multiplications, and data transfers are for fixed and floating point machines by assuming one unit of expense corresponding to an addition compared to other operations. This gives a comparison of the computational expense and not the individual machine comparison, see Stine (2003) and Swartzlander Jr. (2001).

Fixed point Floating point

Multiplication

Addition

Transfer

10 2

1 1

0.5 0.5

From this table one can appreciate how much FFT algorithms are cost-effective in implementing the discrete Fourier transform when the number of & data points N is large. In the following, z-transforms, the discrete-time counterpart of the L-transforms, will be presented. Theory behind z-transforms is rather sophisticated and our presentation will be simple. See Oppenheim and Schafer (1999) for a detailed discussion on this topic.

9.4 z-Transforms

1 X

jx½nj51:

(9:4:1)

n¼1

The DTFT pair is jO

Xðe Þ ¼

1 X

x½ne

1 ! 2p

jnO DTFT

n¼1

ðp

XðejO ÞejnO dO

p

¼ x½n:

(9:4:2)

The DTFT of x½nens and the corresponding DTFT are as follows: DTFT½x½nens ¼ ¼

1 X

½x½nens ejnO

n¼1 1 X

x½nejðsþjOÞn ;

(9:4:3)

n¼1

x½nejnO

DTFT

! XðejðsþjOÞ Þ:

(9:4:4)

The convergence of the sequence x½n ejnO can now be defined in terms of ens , which is similar to the convergence of L-transforms, see Section 5.4. It is desirable to use the notation z ¼ esþjO ¼ es ejO ¼ rejO and lnðzÞ ¼ s þ jO and ð1=zÞdz ¼ jdO:

(9:4:5)

Using these in (9.4.1), the time sequence and the corresponding z-transform are 1 x½n ¼ 2pj ¼

I

1 X

XðzÞzn1 dz;

XðzÞ

x½nzn :

(9:4:6)

n¼1

The z-transform of a discrete-time sequence x½n is defined in terms of a complex variable z by Zfx½ng ¼ XðzÞ ¼

1 X

x½nzn :

(9:4:7a)

n¼1

The DTFT of the sequence x½n, XðejO Þ exists provided that x½n is absolutely summable (see (8.5.11), which is repeated below in (9.4.1)). This is sufficient but not necessary:

The range of values of the complex variable z for which the summation converges is called the region of convergence (ROC). The inverse z-transform and

9.4 z-Transforms

381

Fig. 9.4.1 Contour of integration on the z-plane.

Fig. 9.4.2 Example 9.4.1: region of convergence ða40Þ

Im(z)

Im(z)

r

0

Re(z)

Re(z )

the symbolic relationship between x½n and XðzÞ are, respectively, given by I 1 1 n1 dz x½n ¼ Z ½XðzÞ ¼ C XðzÞz 2pj

I z contour integral ; x½n $ XðzÞ: C (9:4:7b) The contour integral is around a circle of radius r in the counterclockwise direction enclosing the origin on the z-plane, see Fig. 9.4.1. The complex domain integration requires knowledge of complex variables, which is beyond our scope. The z-transform exists when the sum in (9.4.7a) converges. A necessary condition for convergence is absolute summability of x½nzn . Let z ¼ rejO . The absolute summability of x½nzn is

ROC 1 is the range of values of z for which az 51 or jzj4jaj. The transform is represented by a rational function of the complex variable z. As in the Laplace transforms, we can describe a rational function XðzÞ in terms of its poles (the roots of the denominator) and zeros (the roots of the numerator) on the complex z-plane. There is a pole at z ¼ a and a zero at z ¼ 0 and these are shown in Fig. 9.4.2. The ROC is outside the circle of radius a. The boundary of the ROC is jzj ¼ jaj. The ROC does not contain any poles and is outside & the circle of radius jzj ¼ a. Example 9.4.2 Consider the left-side sequence x2 ½n ¼ bn u½n 1; b 6¼ 0. Solution: The z-transform is 1 1 X X x2 ½nzn ¼ bn u½n 1zn : X2 ðzÞ ¼ n¼1

1 X

jx½nrn j51:

n¼1

(9:4:10)

(9:4:8)

n¼1

The range of r for which the sum converges is the region of convergence (ROC).

u½n; u½n; and u½n 1 ¼ u½ðn þ 1Þ are sketched in Fig. 9.4.3a,b,c and X2 ðzÞ ¼

1 X

ðb=zÞn :

(9:4:11)

n¼1

9.4.1 Region of Convergence (ROC)

Using the change of variable m ¼ n in (9.4.12), we have

Example 9.4.1 Determine the z-transform of the right-sided sequence x1 ½n ¼ an u½n.

n¼1

¼

n¼0

ðaz1 Þn ¼

1 X m¼1

Solution: The z-transform of x1 ½n is 1 1 X X x1 ½nzn ¼ an u½nzn X1 ðzÞ ¼ 1 X

X2 ðzÞ ¼

ðz=bÞm ¼ 1

1 X

ðz=bÞn

m¼0

1 z ¼1 ¼ ; jzj5jbj: 1 ðz=bÞ z b

(9:4:12)

n¼1

1 z ¼ ; jzj4jaj: 1 1 az za (9:4:9)

See the pole–zero plot and the region of convergence in Fig. 9.4.4. In the last two examples, a right-side and a leftside sequences were considered. If a ¼ b, the two

382

9 Discrete Data Systems

Fig. 9.4.3 (a) u½n, (b) u½n, and (c) u½ðn 1Þ

(a)

(b)

transforms are identical except the ROC is different. The z-transform of the sequence and the ROC need to be known before the sequence can be identified. Example 9.4.3 Determine the z-transformof the two-sided sequence y½n ¼ an u½n þ bn u½n 1 and the ROC. Solution: The z-transform of the sum is obtained by adding the two transforms and 1 X ðan u½n bn u½n 1Þzn YðzÞ ¼ n¼1

¼

1 X

an zn 1 þ

n¼0

1 X

ðz=bÞn :

(9:4:13)

m¼0

First and the second sums on the right-hand side have the ROCs jzj4jaj and jzj5jbj, respectively. The sum of the two functions and the corresponding ROC are given by z z YðzÞ ¼ þ ; za zb ROC : fjzj4jajg \ fjzj5jbjg: (9:4:14)

x½n ¼

(c)

6¼ 0; N1 n N2 0; Otherwise

z

! XðzÞ ¼

N2 X

x½nzn :

n¼N1

(9:4:15) Solution: For z 6¼ 0 or 1, each term will be finite and the function XðzÞ converges. If N1 50 and N2 40, the sum includes both negative and positive powers of z. As jzj ! 0, the terms with negative powers of z become unbounded. As jzj ! 1, the terms with positive powers of z become unbounded. Therefore, the ROC of the function XðzÞ of a finite sequence is the entire z-plane except for z ¼ 0 and z ¼ 1. If N1 0; the ROC includes z ¼ 1 and if N2 0, the ROC & includes z ¼ 0. Example 9.4.5 Find XðzÞ for x½n in Fig. 9.4.6 and make a pole–zero plot. Solution: First, XðzÞ ¼

2 X

x½nzn ¼ 1:zð2Þ þ ð6Þ:zð1Þ

n¼2

The ROC in (9.4.15) exists only if there is an overlap of the regions identified by the regions fjzj4jajg and fjzj5jbjg. If jbj4jaj, then the transform converges in the annular region shown in Fig. 9.4.5a identified by jaj5jzj5jbj. If jbj5jaj, there is no region of overlap and therefore the ROC is the & null set.

þ 9:z0 þ 4:z1 þ ð12Þ:z2 ¼ z2 6z þ 9 þ 4z1 12z2 z4 6z3 þ 9z2 þ 4z 12 z2 2 ðz 2Þ ðz 3Þðz þ 1Þ ¼ : z2 ¼

Example 9.4.4 Give the ROC of the sequence.

Im(z)

Im(z)

Im(z) α

β

β Re(z)

β

β Re(z)

Fig. 9.4.4 Pole–zero plot and ROC of X2 ðzÞ.

α+β 2

α Re(z) α+β 2

Fig. 9.4.5 (a) ROC of YðzÞ and (b) no region of convergence when jaj4jbj

9.4 z-Transforms

383

Fig. 9.4.6 (a) x½n and (b) pole–zero plot

The poles of XðzÞ are at the origin and have zeros at z ¼ 1; 2; and 3. See the pole–zero plot in Fig. 9.4.6b. The ROC of this function spans the entire z-plane corresponding to the region enclosed between & the poles at zero and those located at infinity. Example 9.4.6 Derive the z-transform of the sequence for the following two cases: a being arbitrary and a ¼ 1: n a ; 0nN1 : (9:4:16) x½n ¼ 0; otherwise Solution: XðzÞ ¼

N1 X

an zn ¼

n¼0

N1 X

ðaz1 Þ

n¼0

1 aN zN ; ROC : jzj40: ¼ 1 az1 a ¼ 1 ) x½n ¼

1; 0 n N 1

0; otherwise

(9:4:17)

z

!

1 zN 1 z1

¼ XðzÞ; ROC : jzj 6¼ 0:

(9:4:18)

See Fig. 9.4.7 for the pole-zero plot assuming & N¼11 Notes on the ROC of a rational function XðzÞ: The ROC depends on the poles of the Im(z)

function XðzÞ at z ¼ ri ejyi ; ri 40. The maximum and minimum magnitudes of these poles are identified by rmax ¼ maxjri j and rmin ¼ minjri j. If the degree of the denominator of XðzÞ is smaller than the degree of the numerator, then XðzÞ has at least one pole at 1. 1. The ROC does not contain any poles. 2. If the sequence is a finite sequence, then the ROC of XðzÞ is the entire z-plane except possibly z ¼ 0 or z ¼ 1. 3. If x½n is a right-side sequence, i.e., x½n ¼ 0; n5N1 51, and XðzÞ converges for some values of z, the ROC is rmax 5jzj 1 with a possible exception of z ¼ 1. 4. If x½n is a left-sided sequence, i.e., x½n ¼ 0; 15N2 5n and XðzÞ converges for some values of z, then the ROC is 0 jzj5rmin with a possible exception of z ¼ 0 5. If x½n is a two-sided sequence and the region of convergence of the right- and left-sided sequences are, respectively, given by r1 5jzj and jzj5r2 and XðzÞ converges for some values of z, then the ROC takes the form r1 5jzj5r2 , where r1 and r2 are the magnitudes of the poles of XðzÞ. Example 9.4.7 Find the z-transforms of the following sequences and their ROCs: a: x1 ½n ¼ d½n; b: x2 ½n ¼ u½n;

11th order pole

c: x3 ½n ¼ u½n 1;

6 a Re(z)

d: x6 ½n ¼ ajnj ; jaj51: Solution: a: X1 ðzÞ ¼

1 X

d½nzn ¼ z0 ¼ 1;

n¼1 z

Fig. 9.4.7 Example 9.4.6: pole–zero plot ðN ¼ 11Þ

d½n ! 1; ROC : all z

(9:4:19)

384

b: X2 ðzÞ ¼

9 Discrete Data Systems 1 X

zn ¼

n¼0

1 z ; ROC : jzj41 ¼ 1 z1 z 1 (9:4:20)

c: X3 ðzÞ ¼ 1 jnj

1 z ¼ ; ROC : jzj51 1 1z z1 (9:4:21) n

Example 9.4.8 Use the z-transforms to determine the DTFT of the discrete-time function z

x½n ¼ u½n u½n N; N40 $ XðzÞ ¼

n

d: x6 ½n ¼ a ¼ a u½n þ a u½n 1 z z z ! z a z ð1=aÞ ða2 1Þ z ; ¼ a ðz aÞðz ð1=aÞÞ 1 ROC : jaj5jzj5 ; jaj51: a

response. The amplitude and the phase responses are periodic with period 2p.

N 1 X

zn ; ROC : jzj40:

(9:4:24)

n¼0

Solution: Since ROC of XðzÞ includes the unit circle, the DTFT XðejO Þ exists. It is periodic with period 2p and phase response is linear: (9:4:22) & XðejO Þ ¼ XðzÞjz¼ejO ¼

N 1 X n¼0

9.4.2 z-Transform and the Discrete-Time Fourier Transform (DTFT) Note XðzÞ is a function of the complex variable z. The point z ¼ rejO is located at a distance r from the origin at an angle O from the positive real axis. If x½n is absolutely summable, then the DTFT can be obtained from the z-transform by setting z ¼ rejO jr¼1 :

X ejO ¼ XðzÞjz¼e jO (9:4:23) The equation jzj ¼ ejO ¼ 1 describes a circle of unit radius ðr ¼ 1Þ centered at the origin in the z-plane. The frequency O in the discrete-time Fourier transform corresponds to the point on the unit circle at an angle O in radians with respect to the positive real axis. As the discrete-time frequency varies in the range p to p, in the z-plane, it corresponds to one time around the unit circle. In words, (9.4.23) states that the DTFT of a discrete-time signal x½n can be obtained from the z-transform XðzÞ by evaluating it on the unit circle. The z-transform of the sequence is assumed to exist and the DTFT of the sequence exists provided that the region of convergence of XðzÞ includes the unit circle. The DTFT function jO XðejO Þ ejyðOÞ , where is represented by Xðe Þ ¼ XðejO Þ is called the amplitude (or magnitude) response and yðOÞ is called the phase (or angle)

zn ¼

1 zN j jO 1 z1 z¼e

1 ejNO sinðNO=2Þ : ¼ ejOðN1Þ=2 ¼ 1 ejO sinðO=2Þ (9:4:25)

The amplitude and the phase responses are, respectively, given by jO sinðNO=2Þ

jO X e ¼ sinðO=2Þ ;ﬀX e ¼OðN1Þ=2: (9:4:26) &

9.5 Properties of the z-Transform Let the ROC of xi ½n is R xi and R 0 is the ROC after the appropriate operation. The ROC is stated in terms of set theory. The proofs are simple for many of these and are omitted.

9.5.1 Linearity z

Let xi ½n ! Xi ðzÞ, then z

x½n ¼ a1 x1 ½n þ a2 x2 ½n $ a1 X1 ðzÞ þ a2 X2 ðzÞ ¼ XðzÞ; ROC :R0 Rx1 \ Rx2 ; (9:5:1)

9.5 Properties of the z-Transform

385

where R 0 is the ROC of x½n, which is the proper subset of the ROCs of x1 and x2 . Note the intersection of the subsets represented by Rx1 \ Rx2 in (9.5.1). Expansion of the ROC takes into consideration pole cancellations with zeros in XðzÞ.

1 X

Zfx½ng ¼

x½nzn ¼

n¼1

XðzÞ ¼

m¼1

1 X

x½nzn ¼

n¼1

Zfx½nn0 g¼

1 X

x½nn0 zn

1 X

x½mzðmþn0 Þ ¼zn0

m¼1

1 X

¼

0 X

ðz1 Þn ¼

n¼1

x½mzm ¼zn0 XðzÞ;

m¼1

z

x½nn0 ! zn0 XðzÞ;ROC:R0 R\f05jzj51g: (9:5:2) Noting the multiplication factor zn0 , additional poles are introduced when n0 40 and, at the same time, some of the poles at 1 are deleted. In a similar manner if n0 50 then additional zeros are introduced at z ¼ 0 and some of the poles at 1 are deleted. This implies that the points z ¼ 0 and z ¼ 1 are either added or deleted from the ROC by time shifting the function. The special cases include the unit delay and advance operations:

1 ; ROC : R0 ¼ jzj51: 1z

Solution: Noting that u½n !ð1=ð1 z1 ÞÞ and using the transformation z ! 1=z results in the above equation verifying the time reversal theorem. We note that the closed-form expression for the sum is valid if jzj51. Also, the ROC of the unit step sequence is jzj41. Reversing the sequence results & in the ROC from jzj51.

9.5.4 Multiplication by an Exponential If a is a complex number, then Zfan x½ng ¼

1 X

an x½nzn ¼

n¼1

1 X

x½nðz=aÞn

n¼1

¼ Xðz=aÞ; an x½n ! Xðz=aÞ; ROC : R0 ¼ jajR:

1

x½n 1 ! z XðzÞ; ROC : R0 ¼ Rx \ f05jzjg;

(9:5:3a)

z

x½n þ 1 ! zXðzÞ; ROC : R0 ¼ R \ fjzj51g:

(9:5:3b)

(9:5:5)

Change in the argument of Xðz=aÞ resulted in the multiplication of ROC boundaries by jaj. ROC expands or contracts by the factor of jaj. In the special case of a ¼ ejO0 n : z

ejO0 n x½n ! XðejO0 zÞ; ROC :R0 ¼ R:

(9:5:6)

Example 9.5.2 Determine the z-transform of the real sequence y½n ¼ rn cosðO0 nÞu½n.

9.5.3 Time Reversal z

zn

n¼0

z

z

1 X

z

n¼1

¼

x½mzm ¼ Xð1=zÞ:

Example 9.5.1 Find z½x½n ¼ z½u½n directly and then verify using the result in (9.5.4):

9.5.2 Time-Shifted Sequences The z-transform of the time-shifted sequence and the corresponding ROC are

1 X

z

If x½n ! XðzÞ; ROC ¼ R, then x½n ! Xð1=zÞ; ð9:5:4Þ ROC :R0 ¼ 1=R: Reversing in time results in the transformation z ) ð1=zÞ in the transform and, the points in the ROC R0 corresponds to the inverses of the points in R. This can be shown by

z

Solution: Noting rn u½n ! z=ðz rÞ; ROC : jzj4jrj we can write 1 1 z z 1 y½n ¼ rn ejO0 n x½n þ rn ejO0 n u½n ! 2 2 2 z rejO0 1 z þ ¼ YðzÞ ; 2 z rejO0

386

9 Discrete Data Systems

9.5.6 Difference and Accumulation

z

y½n ¼ rn cosðO0 nÞu½n ! YðzÞ ¼

z2 r cosðO0 Þz ; ROC : jzj4jrj: z2 2r cosðO0 Þz þ r2

The z-transforms of these are

(9:5:7a) In a similar manner, r sinðO0 Þz ; r sinðO0 nÞu½n ! 2 z 2r cosðO0 Þz þ r2 (9:5:7b) ROC : jzj4r:

z

y1 ½n ¼ x½n x½n 1 ! XðzÞ½1 z1 ¼ Y1 ðzÞ; R0 R \ fjzj40g; (9:5:10a) n X 1 z XðzÞ y2 ½n ¼ x½k ! ð1 z1 Þ k¼1 ¼ Y2 ðzÞ; R0 R \ fjzj41g :

z

n

&

9.5.5 Multiplication by n

(9:5:10b)

9.5.7 Convolution Theorem and the z-Transform Convolution theorem states that the convolution in the time domain corresponds to the multiplication in the z-domain and z

y½n ¼ x½n h½n ! XðzÞHðzÞ The z-transform of nx½n is

¼ YðzÞ; ROC : Ry ðRx \ Rh Þ:

z

y½n ¼ nx½n ! z½dXðzÞ=dz ¼ YðzÞ; ROC : R0 ¼ R:

(9:5:8)

This can be shown using the expression for the convolution of two sequences x½n and h½n and then by using the transform pairs as shown below: z

This can be seen by differentiating both sides with respect to zof the following equation:

z

x½n ! XðzÞ and h½n k ! zk HðzÞ; y½n ¼

1 X

x½kh½n k ¼

k¼1

XðzÞ ¼

1 X

x½nzn !

n¼1

(9:5:11)

1 X

h½kx½n k;

k¼1

1 X " # dXðzÞ 1 1 1 ðnÞx½nzn1 : ¼ X X X n dz y½nz ¼ x½kh½n k zn YðzÞ ¼ n¼1 n¼1

Multiplying both sides by (z) and identifying the appropriate time and transform terms, (9.5.8) follows. The region of convergence is the same for XðzÞ and YðzÞ. Example 9.5.3 Determine the z-transform of the right-side sequence x½n ¼ nan u½n; a40 using the multiplication by n property.

¼

1 X

n¼1

" x½k

1 X

#

h½n kzn

n¼1

k¼1

¼

1 X

k¼1

x½k HðzÞzk

k¼1

¼ HðzÞ

1 X

x½kzk ¼ HðzÞXðzÞ

k¼1

z

Solution: Noting that an u½n !ðz=ðz aÞÞ; ROC : jzj4jaj, by using (9.5.8), we have

z

nan u½n ! z

dðz=ðz aÞÞ az ; ¼ dz ðz aÞ2

(9:5:9)

ROC : jzj4jaj: &

The ROC of YðzÞ contains the intersection of the ROC of XðzÞ and YðzÞ: If a zero of one of the transforms cancels with a pole of the other, then the ROCof YðzÞ will be larger than the intersection of R x and R h . Example 9.5.4 Verify the result in (9.5.10b) by using (9.5.11).

9.5 Properties of the z-Transform

387

Solution: Noting u½n k ¼ 0; k4n, y½n can be expressed by n 1 X X x½k ¼ x½ku½n k y½n ¼ k¼1

We have a pole outside and one inside the unit circle resulting in a transform that has the annular region of convergence given in (9.5.14d). See Fig. 9.5.1 for & pole-zero plots and ROC.

k¼1 z

¼ x½n u½n ! XðzÞzfu½ng z ¼ XðzÞ ; ROC :R0 ðR \ fjzj41gÞ: ðz 1Þ (9:5:12) & Example 9.5.5 Find the z-transform of the following sequence: y½n ¼ x½n h½n; x½n ¼ nan u½n; h½n ¼ ðbÞn u½n; a ¼ b ¼ 1=2:

(9:5:13)

Solution: Using the multiplication by n property, we have az z x½n ¼ nan u½n ! ¼ XðzÞ; ðz aÞ2 ROC : jzj4jaj; (9:5:14a) z h½n ¼ bn u½n ! ; ROC : jzj4jbj: (9:5:14b) ðz bÞ z

9.5.8 Correlation Theorem and the z-Transform In Section 8.3 the cross-correlation of two sequences x½n and h½n was defined by rxh ½k ¼ x½k h½k ¼ ¼

1 X

x½nh½n þ k

n¼1 1 X

x½m kh½m: (9:5:15)

m¼1

Correlation theorem: rxh ½k¼

1 X

z

x½nh½nþk ! XðzÞHð1=zÞ

n¼1

¼Rxh ðzÞ;ROC:Rxh ðRx \Rh Þ: Using (9.5.14a and b), the time reversal property, and the convolution theorem, we have z

x½n ¼ nð1=2Þn u½n !

ð1=2Þz ðz þ ð1=2ÞÞ2

¼ XðzÞ; ROC :jzj4ð1=2Þ ; z

h½n ¼ ð1=2Þn u½n !

1=z ð1=zÞ ð1=2Þ

(9:5:14c)

2 ¼ HðzÞ; ROC :jzj52 ; z2 z YðzÞ ¼ HðzÞXðzÞ ¼ ; ðz 2Þðz þ ð1=2ÞÞ2 1 5 jzj 5 2: (9:5:14d) ROC : 2

(9:5:16)

This can be seen by first noting that rxh ½k ¼ x½k h½k. Using the convolution and the time reversal properties, we can see the result in (9.5.16). Again there is a possibility of pole–zero cancellations and therefore Rxh ðRx \ Rh Þ. In the case of autocorrelation, we have h½n ¼ x½n and the autocorrelation (AC) theorem in (9.5.16) reduces to

¼

Im(z)

−

Fig. 9.5.1 Example 9.5.5: pole–zero plots and the ROCS: (a) XðzÞ; (b) HðzÞ; and (c) YðzÞ

1 2

rxx ½k ¼ rx ½k ¼

1 X n¼1

¼ Rx ðzÞ; ROC : Rxh ðRx \ Rh Þ:

1 2 Re(z)

(9:5:17)

Im(z)

Im(z)

1

z

x½nx½n þ k ! XðzÞXð1=zÞ

Re(z)

−

1 2

1 2 Re(z)

388

9 Discrete Data Systems

Example 9.5.6 Using a. the direct and b. the transform methods, find the AC of x½n ¼ an u½n; jaj51:

(9:5:18)

Solution: a. Since the AC function is even, we need to find rx ½k; k 0 and then use rx ½k ¼ rx ½k. For k 0, since jaj51, we have

9.5.10 Final Value Theorem in the Discrete Domain The final value theorem applies only to causal sequences x½n and if all the poles of XðzÞ lie within the unit circle, with the exception that it can have one pole at z ¼ 1, then lim x½n ¼ limð1 z1 ÞXðzÞ if x½1exists: (9:5:21)

n!1

rx ½k ¼

1 X

1 X

an anþk ¼ ak

n¼0

ða2 Þn ¼

n¼0

1 ak ; k 0: 1 a2

Autocorrelation is even ) rx ½k jkj

2

z!1

This can be seen by Zfx½n x½n 1g ¼ ð1 z1 ÞXðzÞ

(9:5:19a)

¼ a =ð1 a Þ; 15k51:

N!1

b. By the autocorrelation theorem,

lim lim

N X

z!1 N!1

Rx ðzÞ ¼ XðzÞXðz1 Þ ¼ ¼

1 1 1 az1 1 az

1 z ; a ðz aÞðz ð1=aÞÞ

N X

¼ lim

fx½n x½n 1gzn ;

(9:5:22)

n¼0

fx½n x½n 1gzn

n¼0

¼ lim lim

N!1 z!1

N X

fx½n x½n 1gzn

n¼0

(9:5:19b)

lim ½x½0 x½1 þ x½1 x½0 þ x½2 x½1 þ :::

N!1

1 ROC : jaj5jzj5 : jaj

¼ lim x½N: N!1

z

) ajkj !

2

a 1 z : a ðz aÞðz ð1=aÞÞ

(9:5:19c)

The region of convergence is an annular ring and the AC sequence is a two-sided sequence. Note the & poles of the z-transform in (9.5.19c).

Notes: The final value x½1 is equal to zero if all the poles of XðzÞ lie within the unit circle. This follows from the fact that the corresponding time function contains exponentially damped terms. It is a constant if XðzÞ has a single pole at z ¼ 1. If the poles are outside of the unit circle, then the final value theorem gives incorrect results. Also, x½1 is indeterminate if & there are complex poles on the unit circle.

9.5.9 Initial Value Theorem in the Discrete Example 9.5.7 Find the initial and final values of Domain x½n for the function If the sequence x½n is causal, i.e., x½n ¼ 0; n50; as z ! 1; zn ! 0 for n40, we have 1 X

x½0 ¼ lim XðzÞ ¼ lim z!1

z!1

n¼0

1

¼ lim ½x½0 þ x½1z z!1

x½nz

n

þ x½2z

(9:5:20a) 2

þ ;

XðzÞ ¼

Solution: The initial and final values are, respectively, given by ðzð1=3ÞÞ ¼0 x½0¼ lim XðzÞ¼ lim z!1 z!1 ðz1Þðzð1=2ÞÞ lim x½n¼limð1z1 ÞXðzÞ

n!1

x½0 ¼ 0 ) x½1 ¼ lim zXðzÞ: z!1

ðz ð1=3ÞÞ ; jzj41: ðz 1Þðz ð1=2ÞÞ

(9:5:20b)

z!1

ðz1Þðzð1=3ÞÞ 4 ¼ : ¼lim z!1 zðz1Þðzð1=2ÞÞ 3

&

9.6 Tables of z-Transform Properties and Pairs

389

y½n ¼ x½nu½n ¼ f0; 1; 0; 1; 0; 1; 0; 1; :::g: #

Switched periodic sequences and their z-transforms: Consider the periodic sequence x½n with the property x½n ¼ x½n þ N. Now define a causal sequence y½n ¼ x½nu½n. Let the z-transform of the first N-point sequence is N1 X X1 ðzÞ ¼ x½nzn :

Solution: The period of the sequence sinðnp=2Þ is 4. Therefore z

n¼0

x1 ½n ¼ f0; 1; 0; 1g ! z1 z3 ¼ X1 ðzÞ #

The z-transform of the function y½n is given by YðzÞ ¼ X1 ðzÞ þ z

N

X1 ðzÞ þ z

2 N

) XðzÞ ¼

X1 ðzÞ þ :::

X1 ðzÞ z1 z3 z : ¼ ¼ 2 1 z4 1 z4 z þ1

¼ X1 ðzÞ½1 þ zN þ z2 N þ ::: &

zN X1 ðzÞ; jzj41: ¼ N z 1 Example 9.5.8 Find the z-transform of the sequence y½n ¼ x½nu½n; x½n ¼ sinðnp=2Þ:

9.6 Tables of z-Transform Properties and Pairs

Table 9.6.1 Z-transform properties z

z

Two sided signals x½n ! XðzÞ; h½n ! HðzÞ Superposition: z ax½n þ bh½n ! aXðzÞ þ bHðzÞ

ð9:6:1Þ

Time shift: z x½n n0 ! zn0 XðzÞ

ð9:6:2Þ

Scaling: z an x½n ! Xðz=aÞ

ð9:6:3Þ

Multiplication byejnO0 : z

ejnO0 x½n ! XðejO0 zÞ

ð9:6:4Þ

Time reversal: z x½n ! Xð1=zÞ

ð9:6:5Þ

Multiplication by n: z

nx½n ! z dXðzÞ dz

ð9:6:6Þ

Accumulation: n P z z x½k ! z1 XðzÞ

ð9:6:7Þ

k¼1

Difference: z x½n x½n 1 !ð1 z1 ÞXðzÞ Convolution: z x½n h½n ! XðzÞHðzÞ Cross correlation: z x½k h½k ! XðzÞHð1=zÞ Autocorrelation: z x½n x½n ! XðzÞXð1=zÞ

ð9:6:8Þ ð9:6:9Þ ð9:6:10aÞ ð9:6:10bÞ

The following properties hold for causal sequences x½n ¼ 0; n50: Initial value theorem: x½0 ¼ lim XðzÞ

ð9:6:11aÞ

Final value theorem: lim x½n ¼ limðz 1ÞXðzÞ

ð9:6:11bÞ

z!1

n!1

z!1

Switched periodic functions: x½n ¼ x½n þ N; X1 ðzÞ ¼

N1 P n¼0

z

N

x½nzn ; x½nu½n ! zNz1 X1 ðzÞ

ð9:6:12Þ

390

9 Discrete Data Systems Table 9.6.2 Z-transform pairs Unit sample: z

d½n ! 1; ROC : all z: z

ð9:6:13aÞ

k

ð9:6:13bÞ

d½n k ! z ; k > 0; ROC : jzj > 0: z

d½n þ k ! zk ; k > 0; ROC : jzj51:

ð9:6:13cÞ

Unit step: z ; ROC : jzj > 1: z1 z z u½n 1 ! ; ROC : jzj51: z1 z

u½n !

ð9:6:14aÞ ð9:6:14bÞ

Exponential: z ; ROC : jzj > jaj: za z z bn u½n 1 ! ; ROC : jzj5jbj: zb General type: az z nan u½n ! ; ROC : jzj > jaj: ðz aÞ2 z

an u½n !

ð9:6:15aÞ ð9:6:15bÞ

ð9:6:16aÞ

nðn 1Þ:::ðn ðk 2ÞÞankþ1 u½n z z ! ; ROC :jzj > jaj: ðk 1Þ! ðz aÞk az z nan u½n 1 ! ; ROC : jzj5jaj: ðz aÞ2 z2

z

ðn þ 1Þan u½n !

ðz aÞ2

ð9:6:16bÞ ð9:6:16cÞ ð9:6:16dÞ

; ROC : jzj > a:

Sequences involving sinusoids: z2 cosðO0 Þz ; ROC : jzj > 1: z2 ð2 cosðO0 ÞÞz þ 1

ð9:6:17aÞ

sinðO0 Þz ; ROC : jzj > 1: z2 ð2 cosðO0 ÞÞz þ 1

ð9:6:17bÞ

z

cosðO0 nÞu½n ! z

sinðO0 nÞu½n !

z2 r cosðO0 Þz ; ROC : jzj > r: ð2r cosðO0 ÞÞz þ r2

ð9:6:17cÞ

r sinðO0 Þz ; ROC : jzj > r: z2 ð2r cosðO0 ÞÞz þ r2

ð9:6:17dÞ

z

rn cosðO0 nÞu½n ! z

rn sinðO0 nÞu½n !

z2

Finite sequence: n N N z 1a z a ;0 n N1 ; ROC : jzj > 0: ! 0; otherwise 1 az1

ð9:6:18Þ

9.7 Inverse z-Transforms

9.7.1 Inversion Formula

In this section we will consider determining x½n ¼ Z1 fXðzÞg by the following methods: (1) inversion formula, (2) use of z-transform tables, and (3) power series expansion.

The inverse z-transform is x½n ¼

1 2pj

I c

XðzÞzn1 dz:

(9:7:1)

9.7 Inverse z-Transforms

391

It is a contour integral over a closed path C encircling the origin in a counterclockwise direction that lies within the region of convergence of XðzÞ in the zplane. The proof requires the knowledge of complex variables, which is beyond the scope here. See Churchill (1948), Poularikas (1996), and others. Let fak g be the set of poles of XðzÞzn1 inside the contour Cand fbk g be the set of poles of XðzÞzn1 outside the contour C in a finite region of the z-plane. Now x½n ¼ Z1 ½XðzÞ 8P n1 > > < k ResfXðzÞz ; ak Þ; n 0; ¼ P > > : ResfXðzÞzn1 ; bk g; n50:

Re( z )

ð9:7:2aÞ ð9:7:2bÞ

The residue at the multiple and the simple poles at z ¼ p0 of order k are, respectively, given by 1 d k1 ½ðz p0 Þk XðzÞzn1 z!p0 ðk 1Þ! dzk1

ResfXðzÞzn1 g ¼ lim

(9:7:2c)

ResfXðzÞzn1 ; p0 g ¼ XðzÞzn1 ðz p0 Þ z¼p0 ðSimple poleÞ:

1 2

1 2

Fig. 9.7.1 Example 9.7.1: Poles, zeros, and the ROC

k

ðMultiple poleÞ;

Im( z )

(9:7:2d)

9.7.2 Use of Transform Tables (Partial Fraction Expansion Method) This method is based upon expressing XðzÞ as a sum of simple functions Xi ðzÞ by using partial fraction expansion (see Section 5.9.2.), where each one of these functions have inverse transforms that are readily available in a table. This method is limited to rational functions and will need a table of z-transform pairs that provides the appropriate z transform pairs xi ½n ! Xi ðzÞ. The inverse transform is given by ( ) M X 1 1 Xi ðzÞ x½n ¼ Z fXðzÞg ¼ Z i¼1

Example 9.7.1 Find the inverse z-transform of the following function using the residues z ; ðz :5Þðz 2Þ ROC : :55jzj52: See Fig 9:7:1

¼

M X i¼1

Z1 fXi ðzÞg ¼

M X

xi ½n:

(9:7:4)

i¼1

XðzÞ ¼

(9:7:3a)

Solution: The function has a single pole inside and a single pole outside the unit circle. Therefore, it is a two-sided sequence. From (9.7.2c), we have x½n ¼ ResfXðzÞzn1 ; :5g ¼

zðz :5Þzn1 ð:5Þn ; n 0; jz¼:5 ¼ ðz :5Þðz 2Þ 1:5 (9:7:3b)

x½n ¼ ResfXðzÞzn1 ; 2g zðz 2Þzn1 2n ; n50: ¼ jz¼2 ¼ ðz :5Þðz 2Þ 1:5 (9:7:3c) &

We considered partial fraction expansions when we studied Laplace transforms and the procedure here is the same with a slight modification. This approach provides closed-form solutions. From Table 9.6.2 we see that z appears in the numerator of the z- transform functions. Therefore, find XðzÞ=z, and then use the partial fraction expansion discussed in Section 5.4. The expansion of a rational function XðzÞ can be obtained by multiplying each term in the expansion by z. The function XðzÞ=z takes the form XðzÞ a0 þ a1 z þ þ aM zM : ¼ b0 þ b1 z þ þ b N z N z

(9:7:5)

If M4N, then divide the numerator polynomial by the denominator polynomial and

392

9 Discrete Data Systems

½XðzÞ=z ¼ RðzÞ þ X0 ðzÞ; RðzÞ ¼ ½cMN zMN þ cMN1 zMN1 þ þ c1 z þ c0 d0 þ d1 z þ þ dN1 zN1 : X0 ðzÞ ¼ b0 þ b1 z þ þ bN zN

(9:7:6)

Note that the numerator polynomial in the rational function X0 ðzÞ in (9.7.6) is of degree (N1) or less. The denominator of this rational function is then factored and expanded using partial fraction expansion. The inverse z-transform of XðzÞ can now be computed:

Z1 ½XðzÞ ¼ Z1 ½zRðzÞ þ Z1 ½zfpartial fraction expansion of X0 ðzÞg ¼ x1 ½n þ x2 ½n;

(9:7:7)

lZ1 ½c0 z ¼ c0 dðn þ 1Þ; Z1 ½c1 z2 ¼ c1 d½n þ 2; :::;

Multiple pole case: See Section 5.8 on the partial fraction expansion with multiple poles. Example 9.7.3 Find x½n ¼ Z1 fXðzÞg for the cases: a: jzj41 and b: jzj5ð1=2Þ :

XðzÞ ¼

3z3 ð5=2Þz2 ðz ð1=2ÞÞ2 ðz 1Þ

Solution: a. The sequence is a right-side sequence since the ROC is jzj41. We have a double pole at z ¼ ð1=2Þ and a single pole at z ¼ 1: XðzÞ ð3z2 ð5=2ÞzÞ ¼ z ðzð1=2ÞÞ2 ðz1Þ A12 A11 A3 þ ; þ ¼ 2 ðzð1=2ÞÞ ðz1Þ ðzð1=2ÞÞ

X0 ðzÞ¼

M X

cMm d½n þ ðM m þ 1Þ:

(9:7:8)

m¼N

In the following we will concentrate on the second part in (9.7.7) by considering all simple poles and later we will consider a single multiple pole plus simple poles.

A11 ¼

¼

d 3z2 ð5=2Þz z¼1=2 dz ðz 1Þ " # ðz 1Þð6z ð5=2Þ ð3z2 ð5=2ÞzÞ ðz 1Þ2

Example 9.7.2 Find the inverse z-transform of the function

XðzÞ ¼

½z2

z½2z ð4=3Þ ; ROC : jzj41: (9:7:9) ð4=3Þz þ ð1=3Þ

Solution: From the ROC the sequence is a rightside sequence. Now XðzÞ 1 1 ¼ þ z z ð1=3Þ z 1 z z þ ; ROC : jzj41: ) XðzÞ ¼ z ð1=3Þ z 1 XðzÞ z

1 1 z z ¼ zð1=3Þ þ z1 ) XðzÞ ¼ zð1=3Þ þ z1 ;

ROC : jzj41 ) x½n ¼ ð1=3Þn u½n þ u½n

:

(9:7:11)

A12 ¼ ðz ð1=2ÞÞ2 X0 ðzÞ z¼1=2 3z2 ð5=2Þz ¼ z¼1=2 ¼ 1; ðz 1Þ

Z1 ½ck zkþ1 ¼ ck d½n þ k þ 1; :::; x1 ½n ¼ Z1 ½zRðzÞ ¼

; ROC : jzj41: (9:7:10)

A3 ¼

X0 ðzÞ ¼

3z2 ð5=2Þz ðz ð1=2ÞÞ2

z¼1=2

¼ 1;

jz¼1 ¼ 2;

ð3z2 ð5=2ÞzÞ

ðz ð1=2ÞÞ2 ðz 1Þ 1 1 2 þ (9:7:12) þ ¼ 2 ðz ð1=2ÞÞ z 1 ðz ð1=2ÞÞ

) XðzÞ ¼

z ðz ð1=2ÞÞ

2

þ

z 2z þ : z ð1=2Þ z 1 (9:7:13)

&

The last step involves determining the inverse transforms. In doing so, the ROC of the z-domain function should be kept in mind. From Table 9.6.2,

9.7 Inverse z-Transforms

393

z z z z ! an u½n; ! nan1 u½n; 2 ðz aÞ ðz aÞ z

z

ðz aÞ

!

3

Solution: Using the partial fraction expansion, we can write

1 nðn 1Þan2 u½n; . . . : 2!

(9:7:14)

X0 ðzÞ ¼

XðzÞ ðz þ 1Þ A1 ¼ z ðz þ 0:5Þðz 2Þðz 0:75Þ ðz þ 0:5Þ þ

The ROC is outside of the unit circle and the inverse transform is x½n ¼ ½nð1=2Þ

n1

n

þ ð1=2Þ þ 2u½n;

(9:7:15)

x½0 ¼ 3; x½1 ¼ 1 þ ð1=2Þ þ 2 ¼ 7=2; X½2 ¼ 1 þ ð1=4Þ þ 2 ¼ 13=4; . . .

(9:7:16)

b. Now use the transform pairs corresponding to the left-side sequences given below: z ; jzj51; z1 z z an u½n 1 ! ; za az

z

nan u½n 1 !

) XðzÞ ¼

z ðz ð1=2ÞÞ

2

ðz aÞ2

þ

(9:7:18)

z 2z þ : z ð1=2Þ z 1 (9:7:19)

u½n 1 2u½n 1:

(9:7:20a)

x½0 ¼ 0; x½1

(9:7:20b)

x½n ¼ f. . . ; 38; 10; 0; 0; . . .g:

(9:7:21)

# &

Notes: It is uncommon to come across multiple poles of order more than 2. It is simple to use the repeated application of simple pole case, see & Example 5.8.2. Example 9.7.4 Find the sequence x½n ¼ z1 fXðzÞg: XðzÞ ¼

A2 ¼

ðz þ 1Þ ¼ 24 ; and ðz þ 0:5Þðz 0:75Þ z¼2 25

ðz þ 1Þ 28 A3 ¼ ¼ ðz þ 0:5Þðz 2Þ z¼0:75 25 4 z 24 z þ 25 ðz þ 0:5Þ 25 ðz 2Þ

28 z : 25 ðz 0:75Þ

Since the ROC is not specified, the sequence cannot be uniquely determined from the XðzÞ alone. Therefore we will identify all possible ROCs corresponding to this function and find the sequences associated with each of them. To find the various possible ROCs, we first make a pole–zero plot as shown in Fig. 9.7.2a. Using the properties of the ROC, the different possible ROCs are shown in Fig. 9.7.2b–e. 1. ROC : jzj5:5: ROC extends inward to include the origin and x½n is a left-sided sequence:

¼ 4 2 2 ¼ 0; x½2 ¼ 16 4 2 ¼ 10; x½3 ¼ 6ð8Þ 8 2 ¼ 38; ::: ;

ðz þ 1Þ 4 ¼ ; ðz 2Þðz 0:75Þ z¼0:5 25

(9:7:17)

; jzj5jaj:

x½n ¼ ð2Þðnð1=2Þn Þu½n 1 ð1=2Þn

A1 ¼

) XðzÞ ¼

z

u½n 1 !

A2 A3 þ ; ðz 2Þ ðz 0:75Þ

zðz þ 1Þ : ðz þ 0:5Þðz 2Þðz 0:75Þ

x½n ¼

n 4 1 24 u½n 1 ð2Þn 25 2 25

n 28 3 u½n 1 þ u½n 1: 25 4

2. ROC : ð1=2Þ5jzj5ð3=4Þ: ROC is a ring and x½n is two sided:

n 4 1 24 u½n ð2Þn x½n ¼ 25 2 25

n 28 3 u½n 1 þ u½n 1: 25 4

394

9 Discrete Data Systems

Im(z)

Im(z)

Im(z)

Unit Circle

1

1 − 2

Unit Circle

2 Re(z)

1

2

1 − 2

3 4

ROC: z

2

(d)

(e)

Fig. 9.7.2 (a) Pole–zero plot, (b) ROC : jzj5:5, (c) ROC : ð1=2Þ5jzj5ð3=4Þ, and (d) ROC : ð3=4Þ5jzj52,ROC : jzj42

3. ROC : ð3=4Þ5jzj52: ROC is a ring and x½n sequence is two sided:

n 4 1 24 u½n þ ð2Þn x½n ¼ 25 2 25

n 28 3 u½n þ u½n 1: 25 4 4. ROC : jzj42: x½n is a right-sided sequence:

n 4 1 24 x½n ¼ u½n þ ð2Þn 25 2 25

n 28 3 u½n u½n: 25 4

&

9.7.3 Inverse z-Transforms by Power Series Expansion From the definition of the z-transform of a sequence x½n, we can write 1 X XðzÞ ¼ x½nzn ¼ þ x½2z2 þ x½1z1 n¼1

þ x½0 þ x½1z1 þ x½2z2 þ :

(9:7:22)

If XðzÞ can be expanded in a power series, x½n can be determined for positive n(negative n) by identifying the coefficients for the negative powers of z (positive powers of z). Example 9.7.5 Find the inverse z-transform using power series for the function in (9.7.10) assuming two cases of ROC identified by a: jzj41; b: jzj5ð1=2Þ. Solution: a. Since the ROC is jzj41, the sequence is a right-side sequence and therefore the power series of the function contain negative powers of z. Divide the numerator by the denominator in (9.6.23) by division. It can be written by ð7=2Þz2 ð15=4Þz þ ð3=4Þ z3 2z2 þ ð5=4Þz ð1=4Þ 7 ð13=4Þz ð29=8Þ þ ð7=8Þz1 ¼ 3 þ z1 þ 3 z 2z2 þ ð5=4Þz ð1=4Þ 2

XðzÞ ¼ 3 þ

) x½0 ¼ 3; x½1 ¼ 7=2; x½2 ¼ 13=4; ::: (9:7:23) b. Since the ROCis inside the circle of radius (1/2), the sequence is a left-side sequence. It can be

9.8 The Unilateral or the One-Sided z-Transform

395

obtained by expanding the function in terms of positive powers. This is achieved by first writing the polynomials in the numerator and the denominator in reverse order and then making use of long division to obtain series in terms of positive powers of z: ð5=2Þz2 þ 3z3 ð1=4Þ þ ð5=4Þz 2z2 þ z3 ð38=4Þz3 þ 20z4 10z3 ¼ 10z2 þ ð1=4Þ þ ð5=4Þz 2z2 þ z3 ) x½1 ¼ 0; x½2 ¼ 10; x½3 ¼ 38; . . . :

XðzÞ ¼

Notes: If the ROC is in an annular region, we can separate the z-transform corresponding to the rightside sequence and the left-side sequence and follow & the above procedure. Example 9.7.6 Find the inverse z-transform of the following function:

An important property of this transform is its ROC and is outside of a circle in the z-plane.

Consider the transform pair x½n ! XI ðzÞ. The transforms of the delayed and advanced sequences are given below and can be shown by starting with the definition of the unilateral transform of these sequences and reducing them into the appropriate forms: z

a: x½n m ! zm XI ðzÞ þ zmþ1 x½1 þ zmþ2 x½2 þ þ x½m; m40;

( ) x½n ¼

ðaz1 Þn

ð1=nÞa ; n 1 0; n 0

zm1 x½1 zx½m 1; m40:

(9:8:3)

a. First consider (9.8.2). The one-sided transform of the delayed sequence is given by

ZI fx½n mg ¼

1 X n¼0

x½n mzn ¼

1 X

x½kzðmþkÞ :

k¼m

(9:7:26)

n

(9:8:2)

z

b: x½n þ m ! zm XI ðzÞ zm x½0

(9:7:25)

Solution: Using the power series expansion Spiegel (1968), we have n n¼1

(9:8:1)

n¼0

1 X 1

x½nzn :

z

&

XðzÞ ¼ logð1 az1 Þ ¼

1 X

9.8.1 Time-Shifting Property

(9:7:24)

1 ; jzj4jaj: XðzÞ ¼ log 1 az1

XI ðzÞ ¼ ZI fx½ng ¼

(9:8:4)

:

The ROC is outside the circle of radius jaj and x½n is & a right-side sequence.

Separating (9.8.4) into two parts x½n; n50 and x½n for n 0, we have 1 X k¼m

x½kzðmþkÞ

¼ fzmþ1 x½1 þ zmþ2 x½2 þ þ x½mg þ zm

9.8 The Unilateral or the One-Sided z-Transform The unilateral transform is useful since most of our sequences are right sided. The causal part of an arbitrary sequence y½n is y½nu½n. The unilateral or one-sided transform is

1 X

x½kzk

k¼0

¼ zm fx½1z þ x½2z2 þ ::: þ x½mzm g þ zm XI ðzÞ; m40:

(9:8:5)

b. Now consider the one-sided transform of the advanced sequence

396

9 Discrete Data Systems

ZI fx½n þ mg ¼

1 X

x½n þ mzn

n¼0

¼

1 X k¼0

x½kzðkmÞ

m1 X

x½kzðkmÞ

k¼0

(9:8:6a)

Note y½n at location n is the sum of the values at the two previous locations. The sequence results in Fibonacci numbers, see Hershey and Yarlagadda (1986). b. Taking the one-sided z-transform of the equation in (9.8.7) results in YI ½z ¼ Zfy½n 1g þ Zfy½n 2g ¼ z1 YI ðzÞ þ y½1 þ z2 YI ðzÞ

¼ fzm x½0 þ x½1zm1 þ þ x½m 1z1 g þ zm XI ½z:

þ z1 y½1 þ y½2

(9:8:6b)

The relations in (9.8.2) and (9.8.3) provide a method to solve constant coefficient difference equations. This generally involves two discrete functions, an output y½n and an input x½n. The procedure parallels that of solving constant coefficient differential equations and the Laplace transform. In the following it is assumed that one-sided transforms are in use and the subscript (I) will not be shown on X explicitly. The procedure involves first finding the z-transform of the difference equation in terms of the two transforms YðzÞ and XðzÞ. In determining YðzÞ, the initial conditions on y½n need to be known. The input x½nand thereforeXðzÞ is assumed to be known. Solve for YðzÞ and then take the inverse transform of this function resulting in y½n. Two simple examples are considered below, one with zero input, but with initial conditions, and the second one has both input and initial conditions. Example 9.8.1 Consider the second-order difference equation given by

¼ z1 YI ðzÞ þ 1 þ z2 YI ðzÞ þ z1 1

(9:8:8)

z1 z ; ¼ 2 ) YI ðzÞ ¼ 1 z2 1 z z pﬃﬃﬃ pzﬃﬃ ﬃ 1 Poles : p1 ¼ ð1=2Þ þ ð 5=2Þ; p2 ¼ ð1=2Þ ð 5=2Þ: (9:8:9) pﬃﬃﬃ The region of convergence is jzj4 12 1 j 5 . The partial fraction expansion is YI ðzÞ 1 A B ¼ þ ¼ 2 z z z 1 z z 1 z z2 A B pﬃﬃﬃ pﬃﬃﬃ þ ; ¼ z ðð1 5Þ=2Þ z ðð1 þ 5Þ=2Þ 1 1 A ¼ pﬃﬃﬃ ; B ¼ pﬃﬃﬃ 5 5 pﬃﬃﬃ pﬃﬃﬃ ð1= 5Þz ð1= 5Þz pﬃﬃﬃ pﬃﬃﬃ ) YI ðzÞ ¼ þ z ðð1 þ 5Þ=2Þ z ðð1 5Þ=2Þ ( pﬃﬃﬃn pﬃﬃﬃn ) 1 1þ 5 1 1 5 pﬃﬃﬃ u½n; ) y½n ¼ pﬃﬃﬃ 2 2 5 5 (9:8:10)

y½n ¼ y½n 1 þ y½n 2; y½2 ¼ 1; y½1 ¼ 1:

y½n :4472ð1:618Þn ; n 1;

(9:8:7)

Solution: a. By the direct method, we have

1 y½0 ¼ pﬃﬃﬃ ½1 1 ¼ 0; y½1 5 pﬃﬃﬃ

pﬃﬃﬃ

5 5 1 1 1 þ ¼ 1: ¼ pﬃﬃﬃ 2 2 2 5 2

Example 9.8.2 Determine y½n ¼ x½n h½n; x½n ¼ f1; 1; 1g; and h½n ¼ f1; 1; 1g. #

y½0 ¼ 0; y½1 ¼ 1; y½2 ¼ 1;

&

#

Determine y½n for several values of n by a. using the equation in (9.8.7) directly and then b. verify this result using the one-sided z-transform. Cadzow (1973) uses (9.8.7) to generate a model for rabbit population.

(9:8:11)

y½3 ¼ 2; y½4 ¼ 3; y½5 ¼ 5;

Solution: The z-transforms of the two finite length sequences x½n and h½n are

y½6 ¼ 8; y½7 ¼ 13; . . . :

XðzÞ ¼ 1 þ z1 þ z2 ;HðzÞ ¼ 1 z1 þ z2 ;

(9:8:12a)

9.9 Discrete-Data Systems

397

YðzÞ ¼ HðzÞXðzÞ ¼ 1 þ z2 þ z4 ) y½n ¼ f1; 0; 1; 0; 1g:

(9:8:12b) &

#

In the convolution of two sequences x½n and h½n are known and y½n ¼ x½n h½n needs to be found. In Section 9.3.4, the deconvolution problem was identified and three methods were discussed. In the deconvolution problem, y½n and x½n are assumed to be known and h½n is to be determined. Such a problem has practical importance, as it is a system identification problem. That is, determine the unit sample response of a system h½n. In the method of deconvolution using polynomial long division, HðzÞ is obtained YðzÞ=XðzÞ. This is illustrated in the following example. Example 9.8.3 Determine h½n using (9.8.12b). Solution: YðzÞ 1 þ z2 þ z4 z1 þ z4 ¼ ¼ 1 þ XðzÞ 1 z1 þ z2 1 þ z1 þ z2 2 3 4 z þz þz ¼ 1 z1 þ ¼ 1 z1 þ z2 1 þ z1 þ z2 ) h½n ¼ f1; 1; 1g &

H ð zÞ ¼

Principles of additivity and proportionality: A system is said to be additive if Tfx1 ½n þ x2 ½ng ¼ Tfx1 ½ng þ Tfx2 ½ng. This is sometimes referred to as the superposition property. A system is homogeneous if it satisfies the principle of proportionality, y½n ¼ Tfax½ng ¼ afx½ng for a constant a and for any input sequence. Linear systems: A system that is both additive and homogeneous is called a linear system. A system is linear if for any inputs xi ½n and for any constants ai ; i ¼ 1; 2, Tfa1 x1 ½n þ a2 x2 ½ng ¼ a1 Tfx1 ½ng þ a2 Tfx2 ½ng:

(9:9:2)

Example 9.9.1 Consider the systems described by the following transformations. In each case, determine whether the corresponding system is linear or nonlinear: a: y1 ½n ¼ Ax½n þ B; B 6¼ 0; b: y2 ½n ¼ x2 ½n; c: y3 ½n ¼ nx½n:

#

Solution: Using (9.9.2), it follows that a1 Tfx1 ½ng ¼ a1 Ax1 ½n þ a1 B ;

9.9 Discrete-Data Systems

a2 Tfx2 ½ng ¼ a2 Ax2 ½n þ a2 B;

In this section, basic concepts associated with discrete-time systems will be discussed. Presentation will be very similar to the continuous-time systems studied in Chapter 6. Our discussion will be brief. A discrete-time system is represented by a block diagram shown in Fig. 9.9.1 mapping x½n into y½n. The T inside the block diagram is some transformation that converts the input data into output data. We can characterize a discrete-time data system by putting constraints on the transformation T: y½n ¼ Tfx½ng:

(9:9:1)

Tfa1 x1 ½n þ a2 x2 ½ng ¼ a1 Ax1 ½n þ a2 Ax2 ½n þ B 6¼ a1 Tfx1 ½ng þ a2 Tfx2 ½ng: a. This indicates that the system is nonlinear. It is a linear system if B ¼ 0: b. It is easy to see that the system is nonlinear. Any time, if the transformation has a power of the input other than one, the system is nonlinear. c. The outputs corresponding to the inputs x1 ðtÞ; x2 ðtÞ and a1 x1 ðtÞ þ a2 x2 ðtÞ are ai Tfxi ½ng ¼ nxi ½n; i ¼ 1; 2 and Tfa1 x1 ðnÞ

x½n ! Tf:g ! y½n: Fig. 9.9.1 A discrete-data system

þ a2 x2 ðnÞg ¼ nfa1 x1 ½n þ a2 x2 ½ng ¼ a1 Tfx1 ½ng þ a2 Tfx2 ½ng ) System is linear:

&

398

9 Discrete Data Systems

Shift (or time) invariance: A discrete-time system is shift invariant if and only if y½n ¼ Tfx½ng implies that for every input signal x½n and every time shift k, Tfx½n kg ¼ y½n k

(9:9:3)

Example 9.9.2 Are the following systems shift invariant? a: y½n ¼ x2 ½n;

system. We will use the impulse response here. If h½n is the response of a linear time-invariant system to the input unit sample d½n, then the response to the input d½n k is h½n k. Using the linearity property, we can write the response of a linear system to an input x½ngiven in (9.9.4). The output is " y½n ¼ Tfx½ng ¼ T

# x½kd½n k

k¼1

¼

b: y½n ¼ nx½n;

1 X

1 X

x½kTfd½n kg ¼

k¼1

c: y½n ¼ x½k0 n; k0 a positive integer: Solution: a. The response of this system corresponding to the input x½n n0 is y1 ½n ¼ Tfx½n n0 g ¼ x2 ½n n0 ¼ y½n n0 . Therefore, the system is shift invariant. b. The response of this system corresponding to the input x½n n0 is y2 ½n ¼ nx½n n0 6¼ ðn n0 Þy ½n n0 . Therefore, the system is shift variant. c. The response of this system corresponding to the input x½n n0 is y3 ½n ¼ x½k0 n n0 6¼ x½k0 ðn n0 Þ. Therefore, the system is a shift-variant system and is a compressor. If k0 ¼ 1 then the sys& tem is shift invariant.

Linear shift-invariant systems: A linear shiftinvariant system (LSI) is linear and also shift invariant. From Chapter 8, we can write that a discretetime signal x½n is 1 X x½kd½n k: (9:9:4) x½n ¼ k¼1

Modeling a discrete-data system is an important topic of study. In the case of analog systems we have defined the impulse response and we called the transform of the impulse response as the transfer function of the system. In a similar manner we can define the unit sample (discrete impulse) response or simply impulse response. To stick with the analog impulse response notation, many authors use impulse response rather than a unit sample response. The unit sample response or the impulse response provides a complete description of a linear shift-invariant

1 X

x½kh½n k

k¼1

¼ x½n h½k:

(9:9:5)

The output of an LSI system is the convolution of the input sequence with the unit impulse response h½n, which gives a complete characterization of the LSI system. Causality: A causal signal x½n is zero for n50. A system is causal if, for any time n, the response of the system y½n depends only on the present and the past inputs x½n; x½n 1; :::; and does not depend on the future inputs x½n þ 1; x½n þ 2; :::. The response of a causal system satisfies the following, where ffg is an arbitrary function: y½n ¼ ffx½n; x½n 1; :::g:

(9:9:6)

If a system does not satisfy this constraint, then the system is non-causal. In real-time processing applications, we cannot predict the future values and therefore non-causal systems are not physically realizable. In the case of off-line processing, i.e., if we have all the values of the signal, then it is possible to design a non-causal system to process the data. Such a situation is common in data processing. Example 9.9.3 Classify each of the following systems are causal or not; a: y½n ¼ x½n x½n 1 þ x½n 2; b: y½n ¼ x½n þ x½n þ 1; c: y½n ¼ x½2n : Solution: The system in part a is causal, whereas the systems in parts b and care non-causal as they & require the knowledge of future values.

9.9 Discrete-Data Systems

399

The systems described by constant coefficient difference equations given below are linear shift-invariant systems: y½n ¼

N X

ak y½n k þ

k¼1

L X

bk x½n k:

(9:9:7)

k¼0

A system described by (9.9.7) is a recursive linear system if at least one of the coefficients ak is not zero and the output depends on previous values of the output as well as the input. For a non-recursive linear system, a½k ¼ 0 and is described by the difference equation, y½n ¼

L X

1: jlj51; lim ln ! 0; n!1

bk x½n k:

(9:9:8)

k¼0

(9:9:9a) For a system to be BIBO stable, i.e., to have a bounded output jy½nj51 to a bounded input, the unit sample response of a shift-invariant system must be absolutely summable: jh½kj51:

2: jlj41; lim ln ! 1; n!1 n

3: jlj ¼ 1; jlj ! 1 for all n:

Stability: The stability of a discrete system can be defined in terms of the input–output behavior as in the analog case. A system is called BIBO stable if every bounded input produces a bounded output. For linear systems BIBO stability requires that the sample response h½n must be absolutely summable. We can show this by starting with a bounded input such that jx½nj M for all n and the convolution sum in (9.9.5). That is, 1 1 X X h½kx½n k M jy½nj jh½kj: k¼1 k¼1

1 X

be asymptotically stable, if and only if the natural response goes to zero as n ! 1, and is unstable if it grows without bound. We can see this very clearly by making use of the transforms and by expanding the z-domain function into its partial fraction expansion and then taking the inverse terms. The natural response depends upon the characteristic modes of the system. The roots of the characteristic polynomial of the system are the modes. Let a typical root be given by z ¼ l ¼ jljejb . Noting that ln ¼ jljn ejbn , we can summarize the results using the three cases:

(9:9:9b)

k¼1

Example 9.9.4 Show the system described by y½n ¼ njx½nj; x½n ¼ Au½n; A40 and finite is not BIBO stable. Solution: The output y½n ! 1 as n ! 1 and the & system is not BIBO stable. Notes: The response of a discrete-time linear timeinvariant system y½n consists of two parts, one due to the natural response (due to initial conditions) and the second due to the source. A system is said to

(9:9:10)

The linear discrete-time system is asymptotically stable if and only if the characteristic roots, i.e., the poles of the transfer function of the system, are inside the unit circle. It is unstable if there is at least one root outside the unit circle and/or if there are multiple roots on the unit circle. It is marginally stable if and only if there are no roots outside the unit circle and only simple roots on the unit circle. A marginally stable system is not BIBO stable. Discussion on general stability analysis is beyond the scope here. Example 9.9.5 Show that the system described by the following equation is stable: y½n þ 2 þ y½n þ 1 þ 2y½n ¼ x½n þ 1 þ x½n: Solution: The characteristic polynomial can be obtained as follows: zfy½n þ 2 þ y½n þ 1 þ 2y½ng ¼ z2 þ z þ 2; ¼ ðz þ z1 Þðz þ z2 Þ ¼ 0: pﬃﬃﬃ ) z1 ¼ ð1=2Þ þ j 7=2; z2 ¼ z 1 ; z1 z2 ¼ 2. Since jzi j41

the system is unstable.

&

Classification of LSI systems based on the duration of the impulse response: We can classify the linear shift-invariant (LSI) systems based upon the duration of their (discrete impulse or simply impulse) responses. Without losing any generality, we will

400

9 Discrete Data Systems

consider causal systems. The systems that have finite-duration impulse response (FIR) are called FIR systems. On the other hand, the systems with infinite-duration impulse response are called IIR systems. The system described by (9.9.7) is an IIR system if at least one ak 6¼ 0, whereas the system described by (9.9.8) is an FIRsystem. As in the analog systems the transform analysis is basic to filter designs. Since the systems need to work with any set of initial conditions, the designs are based on zero initial conditions.

impulse response h½n or the discrete-time transfer z function HðzÞ ! h½n. Since the impulse response and the corresponding transfer function are related, we can discuss the stability of a discrete linear timeinvariant system in terms of the poles of the transfer function. The transfer function in (9.9.11b) is L P

bk z k¼0 N P

HðzÞ ¼

k

ak zk

1þ

(9:9:12a)

k¼1

¼K

ðz z1 Þðz z2 Þ ðz zL Þ : ðz p1 Þðz p2 Þ ðz pN Þ

9.9.1 Discrete-Time Transfer Functions Consider the difference equation in (9.9.7). Assuming the initial conditions are all equal to zero and taking the z-transform of this equation result in N L X X YðzÞ ¼ ak zk YðzÞ þ bk zk XðzÞ; (9:9:11a) k¼1

k¼0 L P

) YðzÞ ¼

bk z k¼0 N P

k

XðzÞ ¼ HðzÞXðzÞ;

ak zk

1þ

k¼1 L P

YðzÞ HðzÞ ¼ ¼ XðzÞ

bk z k¼0 N P

1þ

k

ak

:

(9:9:11b)

zk

k¼1

As in the analog case HðzÞ is called the transfer function and h½n is the impulse response or the unit sample response. A discrete-time linear time-invariant system can be described by its difference equation, its transfer function, or by its poles and zeros. From our earlier discussion on the z-transforms, we can write the expressions for the system input–output relations in the time domain or in terms of the transform domain. That is,

Notes: he poles of the transfer function are called the natural frequencies or natural modes and they determine the time domain behavior of the system response. For example, if we have poles outside the unit circle, the response grows exponentially and the system is unstable. If we have multiple poles on the unit circle, then the response has polynomial growth. If a system has a simple pole on the unit circle, then the system is referred to as marginally stable. If we require that the function is a minimum phase function, then all the zeros and poles must be & inside the unit circle. Special cases of the general model are useful in defining digital filters. These are 1. the autoregressive moving average filter (ARMA); moving average filter (MA); and the autoregressive filter (AR). These are explicitly expressed by L P

bk z k¼0 N P

HARMA ðzÞ ¼

1þ

ak

1

(9:9:11c)

; zk

k¼1

YðzÞ ¼ HARMA ðzÞXðzÞ; y½n ¼

N X

ak y½n k þ

k¼1

L X

bk x½n k;

k¼0

(9:9:12b)

z

y½n ¼ h½n x½n $ YðzÞ ¼ HðzÞXðzÞ;

k

HMA ðzÞ ¼

HðzÞ ¼ zfh½ng; h½n ¼ z fHðzÞg:

L X

bk zk ;

k¼0

YðzÞ ¼ HMA ðzÞXðzÞ; Linear time-invariant discrete-time systems are described by either constant coefficient difference equations relating the output to the input or the

y½n ¼

L X k¼0

bk x½n k

(9:9:12c)

9.9 Discrete-Data Systems

HAR ðzÞ ¼ 1þ

1 N P

401

;YðzÞ ¼ HAR ðzÞXðzÞ; ak zk

The polynomial AðzÞ given in (9.9.13) has all its roots inside the unit circle if and only if the coefficients jKm j51; m ¼ 1; 2; :::; N.

k¼1

y½n ¼ x½n

N X

ak y½n k:

(9:9:12d)

k¼0

All three models are used in different applications. They are related, at least in the limit, see Marple (1987). The AR models are used extensively in spectral estimation. AR and ARMA models are used in digital