S.

Esakkirajan
T. Veerakumar
Badri N. Subudhi

Digital
Signal
Processing
Illustration Using Python
Digital Signal Processing
S. Esakkirajan • T. Veerakumar • Badri N. Subudhi

Digital Signal Processing

Illustration Using Python
S. Esakkirajan T. Veerakumar
Dept of Instrumentation & Control Eng. Dept of Electronics & Communication Eng.
PSG College of Technology National Institute of Technology Goa
Coimbatore, Tamil Nadu, India Ponda, Goa, India

Badri N. Subudhi
Dept of Electrical Engineering
Indian Institute of Technology Jammu
Jammu, Jammu and Kashmir, India

ISBN 978-981-99-6751-3 ISBN 978-981-99-6752-0 (eBook)

https://doi.org/10.1007/978-981-99-6752-0

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore
Pte Ltd. 2024
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by
similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore

Paper in this product is recyclable.

Preface

Motivation

The objective of this book is to implement signal processing algorithms in Python.

During COVID-19 pandemic, it was a challenge to conduct the signal processing
laboratory session in online mode. It was difficult for the students to install propri-
etary software as it was costly and hence not affordable. This motivated us to turn
our attention toward open-source software. There are many open-source software
packages available to implement signal processing algorithms. The reasons for
choosing Python are (1) it is a general-purpose programming language that can be
used for various tasks beyond scientific computing. (2) Python has an active
community of developers who create and maintain a wide range of libraries and
frameworks. (3) Python has become the language of choice for many machine
learning and deep learning applications with powerful libraries such as TensorFlow,
PyTorch, and Keras. The main aim of signal processing is to extract information
from the signal. After extracting useful information, further processing, like classi-
fication of information, has to be done effectively using machine learning and deep
learning libraries in Python.
In this book, Python is used as a tool to implement signal processing algorithms.
Teaching Python is not the main aim of this book. Python is used as a vehicle to
present concepts related to signal processing. In this book, the signals are generated,
manipulated, transformed, and useful information is extracted using libraries avail-
able in Python. The Python programs used in this book are purposively made simple
and illustrative. The libraries used in this book include (1) Numpy, (2) Scipy,
(3) Matplotlib, etc. These libraries provide a wide range of tools and functions for
performing operations like filtering, resampling, prediction, etc.

v
vi Preface

Target Audience

This book is suited for undergraduate students, postgraduate students, research

scholars, and faculties working in signal processing. The reader is assumed to be
familiar with basic Python programming.

Salient Features of the Book

The salient features of the book are summarized below:

• PreLab questions are included in each chapter. The questions are framed to be
concise, clear, and thought-provoking.
• Numerous examples with Python illustrations are provided in each chapter.
Python codes that implement signal processing algorithms are explained in step
by-step approach. Tasks are given at the end of Python examples. These tasks will
help the reader to vary the parameters in the algorithm and realize its impact.
• Exercises are provided in each chapter. These exercises help the reader to develop
a deeper understanding of the concepts discussed in the chapter.
• Objective questions are given in each chapter. It helps the reader to prepare for
competitive examinations like GATE, IES, etc.

Coimbatore, Tamil Nadu, India S. Esakkirajan

Ponda, Goa, India T. Veerakumar
Jammu, Jammu and Kashmir, India Badri N. Subudhi
Organization of the Book

The book comprises of 12 chapters. Chapter 1 deals with the generation and
visualization of continuous-time signals which include periodic signals,
non-stationary signals, pulse signals, and standard test signals. Chapter 2 focuses
on sampling, quantization, and reconstruction of signals. Both the time domain and
frequency domain view of sampling, the effect of undersampling, uniform and
non-uniform quantization, and different types of reconstruction like zero-order
hold, first-order hold, and sinc interpolation are discussed in this chapter.
Chapter 3 is dedicated to the generation of discrete-time signals and mathematical
operations that are performed on the discrete-time signals. In this chapter, standard
discrete-time signals like unit sample, unit step, unit ramp, exponential, and sinu-
soidal signals are generated, and mathematical operations like folding, shifting, and
scaling are performed on the generated signals. This chapter also discusses two
important signal processing operations: convolution and correlation. Different forms
of representation of discrete-time system, properties of discrete-time systems, and
responses of discrete-time systems are explained with examples in Chap. 4. One of
the important topics is signal processing which is analysis of signals and systems
using transform. Chapter 5 is devoted to transform domain analysis of signals and
systems. Different transforms discussed in this chapter include Z-transform, Fourier
transform, Short-Time Fourier transform, and Wavelet transform. Chapter 6 deals
with the design of a simple filter using pole-zero placement technique. Different
filters discussed in this chapter include moving average filter, digital resonator, notch
filter, comb filter, and all-pass filter. Chapter 7 covers the types of FIR filters and the
design of FIR filters. Three design approaches covered in this chapter include
window-based FIR filter design, frequency sampling-based FIR filter design, and
optimal FIR filter design. Chapter 8 deals with the design of IIR filter, mapping from
the analog domain to the digital domain. The types of IIR filters discussed in this
chapter include Butterworth filter, Chebyshev filter, and Elliptic filter. The mapping
techniques discussed in this chapter include the backward difference, impulse
invariant, and matched Z-transform techniques. The impact of the finite word length
effect in the FIR and IIR filters is discussed in Chap. 9. Concepts like limit cycle

vii
viii Organization of the Book

oscillation, impact of coefficient quantization, and the nature of coefficient error are
discussed in this chapter. Chapter 10 is devoted to multi-rate signal processing.
Concepts like multi-rate operators, noble identities, polyphase decomposition, filter
bank, and transmultiplexer are covered with detailed examples in this chapter.
Design of optimal and adaptive filters and their applications are discussed in
Chap. 11. This chapter discusses Wiener filter, LMS algorithm and its variants,
RLS algorithm and its applications with necessary examples. Chapter 12 is devoted
to case study which discusses the application of signal processing algorithms in
analyzing speech signal, ECG signal, and power line signal.
Acknowledgments

The authors are always thankful to the Almighty for guiding them in their persever-
ance and blessing them with achievements. The authors wish to thank Shri
L. Gopalakrishnan, Managing Trustee, PSG Institutions; Dr. K. Prakasan, Principal,
PSG College of Technology, Coimbatore; Prof. Gopal Mugeraya, Director, National
Institute of Technology, Goa; and Prof. Manoj Singh Gaur, Director, IIT Jammu for
their wholehearted cooperation and constant encouragement given in this successful
endeavor.
Dr. S. Esakkirajan would like to express his gratitude to his parents,
Mr. G. Sankaralingam, and Mrs. S. Saraswathi, wife Mrs. K. Sornalatha and sons
Azhaku Vignesh and Krishnan, for their love and encouragement. He would like to
thank his students Mr. Senthil Murugan, Mr. Vijay Bhaskar, Ms. B. Keerthiveena,
and Mr. Upendra Vishwanath for their continual support and encouragement.
Dr. T. Veerakumar would like to thank his life guru Dr. S. Esakkirajan for his
guidance, motivation, and constant support in completing this work. He also wants
to thank his parents, Mr. Thangaraj and Mrs. Muniammal, brothers Mr. Tamilselvan
and Mr. Karl Marks, and sister, Mrs. Muniponnu, for their wholehearted support.
Finally, he would like to thank his wife, Banupriya, and daughters, Harini and
Ishani, for tolerating his late coming home and their support in completing this
work on time.
Dr. Badri Narayan Subudhi would like to express his gratitude to his parents,
Mr. Ananda Chandra Subudhi and Ms. Subasini Subudhi, wife Ms. Bandanarani
Subudhi and children: Aaradhya and Anwit for their unflagging love and support
throughout life. He would also like to thank his brother Mr. Rashimi Ranjan Subudhi
and Prof. Sarat Kumar Patra for their encouragement and support during his life.

ix
Contents

1 Generation of Continuous-Time Signals . . . . . . . . . . . . . . . . . . . . . 1

1.1 Continuous-Time Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Continuous-Time Periodic Signal . . . . . . . . . . . . . . . . 3
1.1.2 Exponential Function . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Non-stationary Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3 Non-sinusoidal Waveform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.1 Square Waveform . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.2 Triangle and Sawtooth Waveform . . . . . . . . . . . . . . . . 22
1.3.3 Sinc Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.4 Pulse Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3.5 Gaussian Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2 Sampling and Quantization of Signals . . . . . . . . . . . . . . . . . . . . . . . 35
2.1 Sampling of Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.1.1 Violation of Sampling Theorem . . . . . . . . . . . . . . . . . 39
2.1.2 Quantization of Signal . . . . . . . . . . . . . . . . . . . . . . . . 45
2.1.3 Mid-Rise Quantizer . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2 Non-uniform Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.3 Signal Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3.1 Zero-Order Hold Interpolation . . . . . . . . . . . . . . . . . . . 60
2.3.2 First-Order Hold Interpolation . . . . . . . . . . . . . . . . . . . 62
2.3.3 Ideal or Sinc Interpolation . . . . . . . . . . . . . . . . . . . . . . 64
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3 Generation and Operation on Discrete-Time Sequence . . . . . . . . . . 71
3.1 Generation of Discrete-Time Signals . . . . . . . . . . . . . . . . . . . . 73
3.2 Mathematical Operation on Discrete-Time Signals . . . . . . . . . . 84
3.2.1 Amplitude Modification on DT Signal . . . . . . . . . . . . . 84
3.3 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.4 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
xi
xii Contents

4 Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

4.1 Discrete-Time System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.2 Representation of DT Systems . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.2.1 Difference Equation Representation of Discrete-Time
Linear Time-Invariant System . . . . . . . . . . . . . . . . . . . 125
4.2.2 State-Space Model of a Discrete-Time System . . . . . . . 130
4.2.3 Impulse Response and Step Response of Discrete-Time
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.2.4 Pole-Zero Plot of Discrete-Time System . . . . . . . . . . . 139
4.3 Responses of Discrete-Time System . . . . . . . . . . . . . . . . . . . . . 142
4.4 Different Representations and Response of Unit Delay
Discrete-Time System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.5 Properties of Discrete-Time System . . . . . . . . . . . . . . . . . . . . . 149
4.5.1 Linearity Property . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.5.2 Time-Invariant and Time-Variant System . . . . . . . . . . . 151
4.5.3 Causal and Non-causal System . . . . . . . . . . . . . . . . . . 153
4.5.4 Stability of Discrete-Time System . . . . . . . . . . . . . . . . 154
4.5.5 Invertibility of Discrete-Time System . . . . . . . . . . . . . 158
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5 Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.1 Introduction to Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.2 Z-Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.2.1 Z-Transform of Standard Test Sequences . . . . . . . . . . . 170
5.3 Inverse Z-Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
5.4 Family of Fourier Series and Transforms . . . . . . . . . . . . . . . . . 174
5.4.1 Continuous-Time Fourier Transform (CTFT) . . . . . . . . 175
5.4.2 Fourier Transform of Standard Test Signals . . . . . . . . . 179
5.4.3 Discrete-Time Fourier Transform (DTFT) . . . . . . . . . . 191
5.4.4 Analysis of Discrete-Time LTI System Using DTFT . . . 193
5.4.5 Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . 196
5.4.6 Properties of DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.4.7 Limitations of Fourier Transform . . . . . . . . . . . . . . . . 204
5.5 Discrete Cosine Transform (DCT) . . . . . . . . . . . . . . . . . . . . . . 206
5.6 Short-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . 210
5.6.1 Shortcoming of STFT . . . . . . . . . . . . . . . . . . . . . . . . . 219
5.7 Continuous Wavelet Transform (CWT) . . . . . . . . . . . . . . . . . . 219
5.7.1 Continuous Wavelets Family . . . . . . . . . . . . . . . . . . . . 220
5.7.2 Drawback of CWT . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
5.8 Discrete Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
6 Filter Design Using Pole-Zero Placement Method . . . . . . . . . . . . . . 231
6.1 First-Order IIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
6.2 Moving Average filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Contents xiii

6.3 M-Point Exponentially Weighted Moving Average Filter

(EWMA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
6.4 Digital Resonator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
6.5 Notch Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
6.6 All-Pass Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
6.7 Comb Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.7.1 Location of Poles and Zeros of Comb Filter . . . . . . . . . 256
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
7 FIR Filter Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.1 FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
7.2 Classification of FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
7.3 Design of FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
7.3.1 Steps in Window-Based FIR Filter Design . . . . . . . . . . 272
7.3.2 Window-Based FIR Lowpass Filter . . . . . . . . . . . . . . . 273
7.3.3 Window-Based FIR Highpass Filter . . . . . . . . . . . . . . . 276
7.3.4 Window-Based FIR Bandpass Filter . . . . . . . . . . . . . . 278
7.3.5 Window-Based FIR Band Reject Filter . . . . . . . . . . . . 280
7.3.6 Design of FIR Filter Using Built-In Function . . . . . . . . 282
7.3.7 Window Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
7.4 Frequency Sampling-Based FIR Filter Design . . . . . . . . . . . . . . 291
7.5 Design of Optimal FIR filter . . . . . . . . . . . . . . . . . . . . . . . . . . 293
7.6 Applications of FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
8 Infinite Impulse Response Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
8.1 IIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
8.2 Mapping Techniques in the Design of IIR Filter . . . . . . . . . . . . 306
8.2.1 Backward Difference Method . . . . . . . . . . . . . . . . . . . 307
8.2.2 Impulse Invariant Technique . . . . . . . . . . . . . . . . . . . . 310
8.2.3 Bilinear Transformation Technique (BLT) . . . . . . . . . . 320
8.2.4 Matched Z-Transform Technique . . . . . . . . . . . . . . . . . 323
8.3 Analog Frequency Transformation . . . . . . . . . . . . . . . . . . . . . . 326
8.4 Butterworth Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
8.5 Chebyshev Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
8.6 Chebyshev Type II IIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . 358
8.7 Elliptic Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
9 Quantization Effect of Digital Filter Coefficients . . . . . . . . . . . . . . . 377
9.1 Number Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
9.2 Fixed-Point Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
9.2.1 Fixed-Point Quantization by Rounding . . . . . . . . . . . . 379
9.2.2 Fixed-Point Quantization Using Two’s Complement
Truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
9.2.3 Fixed-Point Quantization Using Magnitude
Truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
xiv Contents

9.3 Coefficient Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383

9.4 Limit Cycle Oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
9.5 Cascade Form of a Higher Order Filters . . . . . . . . . . . . . . . . . . 397
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
10 Multirate Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
10.1 Multirate Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
10.1.1 Downsampling Operation . . . . . . . . . . . . . . . . . . . . . . 405
10.1.2 Upsampling Operation . . . . . . . . . . . . . . . . . . . . . . . . 409
10.2 Noble Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
10.2.1 Noble Identity for Downsampling Operation . . . . . . . . 413
10.2.2 Noble Identity for Upsampling Operation . . . . . . . . . . . 415
10.3 Polyphase Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
10.4 Filter Bank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
10.4.1 Two-Channel Filter Bank . . . . . . . . . . . . . . . . . . . . . . 420
10.4.2 Relationship Between Analysis and Synthesis Filters . . 421
10.4.3 Two-Channel Filter Bank Without Filters . . . . . . . . . . . 424
10.4.4 Three-Channel Filter Bank Without Filters . . . . . . . . . . 426
10.5 Tree-Structured Filter Bank . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
10.6 Transmultiplexer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
11 Adaptive Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
11.1 Wiener Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
11.1.1 Wiener Filter in Frequency Domain . . . . . . . . . . . . . . . 447
11.2 Adaptive Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
11.2.1 LMS Adaptive Filter . . . . . . . . . . . . . . . . . . . . . . . . . 452
11.2.2 Normalized LMS Algorithm . . . . . . . . . . . . . . . . . . . . 457
11.2.3 Sign LMS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 459
11.3 RLS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
12 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
12.1 Case Study 1: Speech Recognition Using MFCC
(Mel-Frequency Cepstral Coefficient) . . . . . . . . . . . . . . . . . . . . 469
12.1.1 Speaker Identification . . . . . . . . . . . . . . . . . . . . . . . . . 470
12.1.2 Speaker Verification System . . . . . . . . . . . . . . . . . . . . 471
12.1.3 Mel-Frequency Cepstral Coefficient (MFCC) Feature . . 472
12.2 Case Study 2: QRS Detection in ECG Signal Using
Pan-Tomkins Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
12.2.1 ECG Signal Preprocessing . . . . . . . . . . . . . . . . . . . . . 481
12.2.2 Moving Window Integration . . . . . . . . . . . . . . . . . . . . 484
12.2.3 Fiducial Mark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
12.2.4 Decision Rule Approach . . . . . . . . . . . . . . . . . . . . . . . 486
12.3 Case Study 3: Power Quality Disturbance Detection . . . . . . . . . 488
12.3.1 Generation of Power Quality Disturbance . . . . . . . . . . 489
 12.3.2Simulation of Power Quality Disturbance . . . . . . . . . . . 491
12.3.3Time-Frequency Representation of Power Quality
Disturbance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
12.3.4 Time-Scale Representation of Power Quality
Disturbance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
Chapter 1: Generation Of Continuous-Time Signals . . . . . . . . . . . . . . 497
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 497
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 499
Chapter 2: Sampling and Quantization of Signals . . . . . . . . . . . . . . . . 500
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 500
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 501
Chapter 3: Generation and Operation on Discrete-Time Sequence . . . . 501
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 501
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 503
Chapter 4: Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 503
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 506
Chapter 5: Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 506
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 508
Chapter 6: Filter Design Using Pole-Zero Placement Method . . . . . . . . 508
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 508
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 509
Chapter 7: FIR Filter Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 509
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 511
Chapter 8: Infinite Impulse Response Filter . . . . . . . . . . . . . . . . . . . . . 511
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 511
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 513
Chapter 9: Quantization Effect of Digital Filter Coefficients . . . . . . . . . 513
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 513
Answer to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 514
Chapter 10: Multirate Signal Processing . . . . . . . . . . . . . . . . . . . . . . . 515
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 515
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 517
Chapter 11: Adaptive Signal Processing . . . . . . . . . . . . . . . . . . . . . . . 517
Answers to PreLab Questions . . . . . . . . . . . . . . . . . . . . . . . . . 517
Answers to Objective Questions . . . . . . . . . . . . . . . . . . . . . . . . 518

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
About the Authors

S. Esakkirajan is a Professor in Instrumentation and Control Engineering Depart-

ment at PSG College of Technology, where he has been a faculty member since
2004. He did B.Sc. Physics from Sadakathullah Appa College, Palayamkottai, B.
Tech. in Instrumentation Engineering from Cochin University of Science and Tech-
nology, M.E. in Applied Electronics from PSG College of Technology, and Ph.D. in
the area of Image Processing from Anna University. He has successfully guided four
research scholars towards their Ph.D. in the area of signal processing. He has
published papers in reputed journals and conferences. His research interest includes
digital signal processing and digital image processing.

T. Veerakumar is an Associate Professor in the Department of Electronics and

Communication Engineering, National Institute of Technology, Goa. He graduated
with a B.E. in Electronics and Communication Engineering from RVS College of
Engineering Technology, Dindigul. Then, he did an M.E. degree in Applied Elec-
tronics from PSG College of Technology, Coimbatore, and a Ph.D. in Image
Denoising from Anna University, Chennai. He co-authored the textbook titled
Digital Image Processing and Digital Signal Processing, published by Tata
McGraw Hill. In addition, he has published around 60 research articles in reputed
Journals and Conferences. His area of interest includes Signal and Image Processing,
Biomedical Image Processing, Object Detection, and Tracking.

Badri Narayan Subudhi received M. Tech. in Electronics and System Communi-

cation from the National Institute of Technology, Rourkela, India, in 2008–2009. He
worked on his Ph.D. from Machine Intelligence Unit, Indian Statistical Institute,
Kolkata, India, in 2014 (degree from Jadavpur University). Currently, he is serving
as an Associate Professor at the Indian Institute of Technology Jammu, India. Prior
to this, he was working as an Assistant Professor at NIT Goa from July 2014 to
November 2017. He received CSIR senior research fellowship for the year
2011–2015. He was nominated as the Young Scientist Awardee by Indian Science
Congress Association for the year 2011–2012. He was awarded the Young Scientist
Travel grant award from DST, Government of India, and Council of Scientific and
xvii
xviii About the Authors

Industrial Research, India, in 2011. He received the Bose-Ramagnosi Award for the
year 2010 from DST, Government of India, under India-Trento Programme for
Advanced Research (ITPAR). He was a visiting scientist at the University of Trento,
Italy, during August 2010 to February 2011. His research interests include Video
Processing, Image Processing, Medical Image Processing, Machine Learning, Pat-
tern Recognition, and Remote Sensing Image Analysis. He co-authored the textbook
titled Digital Signal Processing, published by Tata McGraw Hill. He has published
around 80 research papers in reputed journals and conferences. He is a senior
member of IEEE.
Chapter 1
Generation of Continuous-Time Signals

Learning Objectives
After completing this chapter, the reader should be able to
• Simulate and visualize periodic continuous-time signals.
• Simulate, visualize and interpret non-stationary signals.
• Simulate and visualize standard continuous-time test signals.
• Simulate and visualize continuous-time pulse signals.

Roadmap of the Chapter

This section discusses the flow of contents in this chapter. The objective of this
chapter is to generate different types of continuous-time signals, pulse waveforms.
The representation of different signals generated in this chapter is given in the form
of a flow diagram, which is given below:

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 1
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_1
2 1 Generation of Continuous-Time Signals

Generation and visualization of

Continuous-Time Signals

Periodic signals Non-stationary Pulse signals Standard test Other

signals signal signals
Rectangular
Sine wave
Sine wave with pulse Dirac Delta Gaussian
Cosine wave time-varying function function
Triangular
Square wave frequency Unit Step
pulse Sinc
Chirp signal function function
Complex
exponential
function

Sawtooth wave

Triangular wave

PreLab Questions
1. Give a few examples of real-world signals, which are continuous in nature.
2. Mention the built-in functions available in ‘numpy’ library in python to generate
data points of specific length to define the independent variable like time.
3. Explain the significance of a sinusoidal signal in signal processing.
4. What do you understand by the term ‘phase’ of a signal?
5. Give a few examples of multidimensional signals.
6. Cite an example where the signal or a process can be modelled as a real
exponential function.
7. Mention a few significant features of complex exponential signals.
8. Mention the salient features of the ‘sinc’ function in signal processing. Is it an
even or odd function?
9. Distinguish between stationary and non-stationary signal. Give examples of
each category of signal.
10. List a few significant properties of the Gaussian function (signal).

1.1 Continuous-Time Signal

A signal corresponds to a physical quantity that varies with time, space, etc. Signals
are represented mathematically as a function of one or more independent variables.
The continuous-time signals are defined for a continuum of values of the indepen-
dent variable. The continuous-time signal is generally represented as x(t). Speech
signal as a function of time is an example of continuous-time signal. The signal can
1.1 Continuous-Time Signal 3

be either deterministic or random. Deterministic signals can be described by math-

ematical functions or expressions. In this chapter, the objective is to generate
different types of continuous-time periodic signals, like sinusoidal signal, complex
exponential signal, square wave, etc.; non-stationary signals, like chirp signal;
standard test signals, like Dirac delta; unit step signal, etc.

1.1.1 Continuous-Time Periodic Signal

A periodic signal is one which repeats itself in an identical manner. Examples of

continuous-time periodic signals include sinusoidal signal, complex exponential
signal, square wave and sawtooth wave. In this section, python codes are developed
to generate a sinusoidal signal, three-phase sinusoidal signal, complex exponential
signal, etc. Also, sinusoids are Eigen functions of linear system. Continuous-time
sinusoids are described by an amplitude, frequency and phase. Continuous-time
sinusoids with distinct frequencies are always distinct.
Experiment 1.1 Generation of Sinusoidal Signal
The aim of this experiment is to generate sinusoidal signal. Sinusoidal signals are
periodic functions, which are based on the sine or cosine function. The expression
for the sinusoidal signal is given by

xðt Þ = A sinð2πft þ ϕÞ ð1:1Þ

In the above equation, ‘A’ represents the amplitude of the signal, ‘f’ denotes the
frequency of the signal and ‘ϕ’ indicates the phase of the signal. To generate
sinusoidal signal, one should define three parameters: amplitude, frequency and
phase. The independent-variable is ‘time (t)’. In amplitude modulation, the ampli-
tude of the carrier is changed in accordance with the message, while the frequency
and phase are kept constant. In frequency modulation, the frequency of the carrier is
changed in accordance with the signal, while the amplitude and phase are kept
constant. In phase modulation, the phase of the carrier is changed in accordance with
the signal, while the amplitude and frequency are kept constant.
The steps involved in the generation of sinusoidal signal are summarized below:
Step 1: Defining the independent variable
The built-in function ‘np.linspace()’ is used to generate the independent
variable, which is the time axis.
Step 2: Defining the parameters of the sine wave
In this step, the three parameters of sine wave, namely, amplitude, frequency
and phase are defined.
Step 3: Generation of sinusoidal signal
In this step, the mathematical expression to generate a sine wave is given by
4 1 Generation of Continuous-Time Signals

Table 1.1 Built-in functions used in the program

S. No. Built-in function used Purpose
1 np.sin() To generate sinusoidal function
2 np.linspace() To generate equally interval data points in an interval
3 plt.subplot() To plot more than one figure in the same plot

#Experiment 1: Generation of sinusoidal signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Defining the independent variable
t=np.linspace(0,1,1000)
#Step 2: Defining the parameters of sine wave
A=5 #Amplitude of sine wave
f=5 #Frequency of sine wave
ph=0 #Phase of sine wave
#Step 3: Expression of sine wave
x=A*np.sin(2*np.pi*f*t+ph)
#Step 4: Plotting the sine wave
plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('A={}V,F={} Hz,$\phi={}^\circ$'.format(A,f,ph))

Fig. 1.1 Python code to generate sinusoidal signal

xðt Þ = A sinð2πft þ ϕÞ

Step 4: Plotting the sinusoidal signal

The built-in function plt.plot() is used to plot the generated signal. While
plotting the waveform, it is important to mention that the label of x and y axes
using plt.xlabel() and plt.ylabel() command. The command plt.title() is used to
display the title of the plot.

Built-In Libraries
The built-in libraries used in the program are (1) Numpy and (2) Matplotlib. The
‘numpy’ is a general purpose array-processing package. In this program, the numpy
library is used to create array (np.linspace), and it is used to perform mathematical
function (np.sin). Matplotlib is a data visualization library used to visualize the
generated sinusoidal signal. The built-in functions used in the program is given in
Table 1.1.
The python code used to generate sinusoidal waveform is shown in Fig. 1.1, and
the corresponding output is shown in Fig. 1.2.
Inference
From Fig. 1.1, the following inferences can be made with respect to python code:
1.1 Continuous-Time Signal 5

Fig. 1.2 Result of the python code shown in Fig. 1.1

1. The libraries used in the program are (a) Numpy and (b) Matplotlib.
2. The built-in function ‘np.linspace()’ is used to generate the independent variable,
which is the time axis. In this program, 1000 data points are generated between
‘0’ and ‘1’.
From Fig. 1.2, it is possible to infer the following:
1. The phase of the signal is ‘0’; this implies that the waveform starts from the
origin.
2. The amplitude of the sine wave is 5 V. The waveform oscillates between -5
and +5.
3. The frequency of the generated waveform is 5 Hz. The number of oscillations per
second is 5.
Tasks
1. Write a python code to mark the peak of the sinusoidal signal.
2. Write a python code to compute the number of zero crossing of the sine wave.
Experiment 1.2 Sinusoidal Signal with Different Phase
In this experiment, the objective is to generate sine wave of amplitude = 1 V,
frequency = 5 Hz and four different phase angles, namely, 0°, 90°, 180° and 270°.
The python code, which does this task, is shown in Fig. 1.3, and the corresponding
output is shown in Fig. 1.4. The built-in libraries used in the program are (1) Numpy
and (2) Matplotlib.
6 1 Generation of Continuous-Time Signals

#Generation of sine wave of different phase angles

import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(0,1,100)
#Parameters of sine wave
A=1 #Amplitude
f=5 #Frequency
phi=[0,90,180,270] #Phase
#Generation of sine wave
for i in range(len(phi)):
x=A*np.sin(2*np.pi*f*t+phi[i]*np.pi/180)
#Plotting the result
plt.subplot(2,2,i+1)
plt.plot(t,x),plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)')
plt.title('$\Phi ={}^\circ $'.format(phi[i]))
plt.tight_layout()

Fig. 1.3 Python code to generate sine wave with different phase angle

) = 0° ) = 90°
1 1
Amplitude (V)

Amplitude (V)

0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time (t) Time (t)

) = 180° ) = 270°
1 1
Amplitude (V)

Amplitude (V)

0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time (t) Time (t)

Fig. 1.4 Sine wave with different phase angle

Inference
From the python code shown in Fig. 1.3, the following inferences can be made. The
phase angle is varied as 0°, 90°, 180° and 270°. The amplitude of the sine wave is
fixed as 1 V and the frequency is fixed as 5 Hz.
1.1 Continuous-Time Signal 7

Fig. 1.5 Python code to

generate three-phase #Generation of three phase sine wave
sinusoidal signals import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(0,1,100)
#Parameters of sine wave
A=1 #Amplitude
f=5 #Frequency
#Three different phases of sine wave
phi_1,phi_2,phi_3=0, 120, 240
x1=A*np.sin(2*np.pi*f*t+phi_1*np.pi/180)
x2=A*np.sin(2*np.pi*f*t+phi_2*np.pi/180)
x3=A*np.sin(2*np.pi*f*t+phi_3*np.pi/180)
#Plotting the result
plt.plot(t,x1,'b',t,x2,'r',t,x3,'g')
plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)')
plt.title('Three phase sinusoidal signal')
plt.legend(['Phase-1','Phase-2','Phase-3'],loc=1)

From Fig. 1.4, it is possible to infer that the starting point of the waveform is
different for different phase angle. The phase parameter determines the time loca-
tions of the maxima and minima of the sinusoid.
Tasks
1. Write a python code to generate a sinusoidal signal whose phase is varying in a
random manner. Assume the phase angle ‘Φ’ to follow uniform distribution in the
range -1 to +1.
2. Write a python code to generate a sinusoidal signal, whose frequency is varying
in a random manner. Assume the frequency ‘f’ to follow uniform distribution in
the range -1 to +1.
Experiment 1.3 Generation of Three-Phase Sinusoidal Signal
The expressions for three-phase sinusoidal signals are given by

x1 ðt Þ = A sinð2πft Þ ð1:2Þ
x2 ðt Þ = A sinð2πft - 120 ° Þ ð1:3Þ
x3 ðt Þ = A sinð2πft - 240 ° Þ ð1:4Þ

The amplitude and frequency of the three waveforms are equal. The phase shift
between the signals is 120°. The python code, which generates the three-phase
sinusoidal waveforms, is shown in Fig. 1.5, and the corresponding output is
shown in Fig. 1.6.
8 1 Generation of Continuous-Time Signals

Fig. 1.6 Three-phase sinusoidal signals

Inference
From the python code to generate the three-phase sinusoidal signal, it is possible to
observe that the amplitude of each signal is 1 V and frequency is 5 Hz. The phase
shift between the signals is 120°.
Task
1. Change the value of amplitude A1, A2 and A3 of three-phase sinusoidal signal in
the python code given in Fig. 1.5, and comment on the output waveform.

1.1.2 Exponential Function

Exponential function is of two types: (1) real exponential function and (2) complex
exponential function. The real exponential function can be either an increasing
function or it could be a decreasing function. The price of petrol is an example of
exponentially increasing function. Radioactive decay is an example of exponentially
decaying function.
Complex exponential function is of interest in signal processing. Complex
exponential function is the basis function of Fourier transform. It is possible to
obtain sine wave and cosine wave from complex exponential function. It is an Eigen
function for a linear time-invariant system.
Experiment 1.4 Generation of Real Exponential Signal
The general expression for the real exponential signal is given by
1.1 Continuous-Time Signal 9

Table 1.2 List of built-in functions used in the program

S. No. Built-in function used Purpose
1 np.exp() To generate exponential function
2 np.linspace() To generate equally interval data points in an interval
3 plt.subplot() To plot more than one figure in the same plot

#Real exponential function

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Defining the time axis
t=np.linspace(-1,1,1000)
#Step 2: Defining the parameter 'alpha'
a,b=2,-2
#Step 3: Generation of function
x1=np.exp(a*t)
x2=np.exp(b*t)
#Step 4: Plotting of the function
plt.subplot(2,1,1),plt.plot(t,x1)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Exponentially growing function')
plt.subplot(2,1,2),plt.plot(t,x2)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Exponentially decaying function')
plt.tight_layout()

Fig. 1.7 Python code to generate real exponential functions

xðt Þ = Ceαt ð1:5Þ

where ‘C’ and ‘α’ are real. If ‘α’ is greater than zero, it is exponentially growing
function. If ‘α’ is less than zero, it is exponentially decreasing or decaying function.
The built-in functions used in the program are given in Table 1.2.
The python code, which generates exponentially growing and decaying function
for α = 2 and α = -2, is shown in Fig. 1.7, and the corresponding output is shown in
Fig. 1.8.
Inference
In exponentially growing function, the value of the function (amplitude of the
function) increases with an increase in time. In contrast, in exponentially decaying
function, the amplitude of the function decays with respect to time.
Task
1. Write a python code to generate two real exponential functions (one growing and
another decaying) with different amplitudes, and add these two functions. Com-
ment on the observed result.
Experiment 1.5 Forward Characteristics of PN Junction Diode
The equation of current through the diode is given by
10 1 Generation of Continuous-Time Signals

Fig. 1.8 Exponentially growing and decaying functions

VD
I D = I s eηV T - 1 ð1:6Þ

In Eq. (1.6), Is represents the reverse saturation current; VD is the voltage drop
across the diode and ID is the current through the diode; VT is the volt-equivalent of
temperature, which is 26 mV at room temperature; and η is the ideality factor, which
is material dependent. The python code, which simulates the V–I characteristics of
PN junction diode by assuming η = 1, VT = 26 mV, and Is = 1 mA is shown in
Fig. 1.9, and the corresponding output is shown in Fig. 1.10.
Inference
From the forward characteristics shown in Fig. 1.10, it is possible to observe that the
diode current increases after crossing the threshold voltage, generally termed ‘knee
voltage’. If one considers the current through the diode as a function, then the
function is an exponentially growing function.
Experiment 1.6 Radioactive Decay Function
The equation of radioactive decay is given by N(t) = N0e-λt. The python code, which
implements this equation, is shown in Fig. 1.11, and the corresponding output is
shown in Fig. 1.12.
1.1 Continuous-Time Signal 11

#Forward characteristics of PN junction diode

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Defining the voltage axis
v=np.arange(0,1,0.001)
#Step 2: Parameters
vt=0.026 #Volt-equivalent of temp.
i_s=1/1000 #Reverse saturation current
n=1 #ideality factor
#Step 3: Equation of current through diode
i=i_s*(np.exp(v/(n*vt))-1)
#Step 4: Plotting the characteristics
plt.plot(v,i),plt.xlabel('Forward voltage')
plt.ylabel('Forward current')
plt.title('Forward characteristics of PN junction diode')

Fig. 1.9 Python code to plot the forward characteristics of PN junction diode

le13 Forward characteristics of PN junction diode

5

4
Forward current

0
0.0 0.2 0.4 0.6 0.8 1.0
Forward voltage

Fig. 1.10 Forward characteristics of PN junction diode

Inference
From Fig. 1.12, it is possible to observe that the radioactive decay activity can be
modelled by an exponentially decaying function.
Experiment 1.7 Complex Exponential Function
Generate two complex exponential signals x1(t) = ejΩt and x2(t) = e-jΩt. Here the
frequency of the signal is fixed as f = 5 Hz. After signal generation, extract the
magnitude and phase of the two signals, and comment on the observed output.
The built-in functions used in the python program are given in Table 1.3.
12 1 Generation of Continuous-Time Signals

Fig. 1.11 Python code for

radioactive decay function #Radioactive decay
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Defining the time axis
t=np.linspace(0,80,10)
#Step 2: Parameters
A0=400 #Initial value
T=24
#Step 3: Equation of current through diode
A=A0*np.exp(-t/T)
#Step 4: Plotting the characteristics
plt.plot(t,A),plt.xlabel('Time (Hours)')
plt.ylabel('Counts per second')
plt.title('Radio active decay')

Radio active decay

400

350
300
Counts per second

250

200
150

100
50

0
0 10 20 30 40 50 60 70 80
Time (Hours)

Fig. 1.12 Result of python code shown in Fig. 1.11

Table 1.3 Python built-in functions used in Experiment 1.7

S. No. Built-in function used Purpose
1 np.exp() To generate an exponential function
2 np.abs() To obtain the magnitude value of a complex number
3 np.angle() To obtain the phase value of the complex number
4 np.linspace() To generate equally interval data points in an interval
5 plt.subplot() To plot more than one figure in the same plot
1.1 Continuous-Time Signal 13

#Complex exponential signals

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of signals x1(t) and x2(t)
t=np.linspace(-1,1,100)
f=5
x1=np.exp(1j*2*np.pi*f*t)
x2=np.exp(-1j*2*np.pi*f*t)
#Step 2: Plotting the signals, its magnitude, and phase
plt.subplot(3,2,1),plt.plot(t,x1)
plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)')
plt.title('$e^{j\Omega t}$'),plt.subplot(3,2,2),plt.plot(t,x2)
plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)')
plt.title('$e^{-j\Omega t}$'),plt.subplot(3,2,3),plt.plot(t,np.abs(x1))
plt.xlabel('Time (t)'),plt.ylabel('Magnitude (V)')
plt.title('|$e^{j\Omega t}$|'),plt.subplot(3,2,4),plt.plot(t,np.abs(x2))
plt.xlabel('Time (t)'),plt.ylabel('Magnitude (V)')
plt.title('|$e^{-j\Omega t}$|'), plt.subplot(3,2,5),plt.plot(t,np.angle(x1)*360/(2*np.pi))
plt.xlabel('Time (t)'),plt.ylabel('$Phase(^\circ$)')
plt.title('$\Phi(x_1(t))$'),plt.subplot(3,2,6),plt.plot(t,np.angle(x2)*360/(2*np.pi))
plt.xlabel('Time (t)'),plt.ylabel('$Phase(^\circ$)'),plt.title('$\Phi(x_2(t))$')
plt.tight_layout()

Fig. 1.13 Generation of complex exponential signals

ej:t e-j:t
Amplitude (V)

Amplitude (V)

1 1
0 0
–1 –1
–1 0 1 –1 0 1
Time (t) Time (t)

|e j:t| |e-j:t|
Magnitude (V)

Magnitude (V)

1.05 1.05
1.00 1.00
0.95 0.95
–1 0 1 –1 0 1
Time (t) Time (t)

)(x1(t)) )(x2(t))
Phase( ° )

Phase( ° )

100 100
0 0
–100 –100
–1 0 1 -1 0 1
Time (t) Time (t)

Fig. 1.14 Magnitude and phase of complex exponential signals

14 1 Generation of Continuous-Time Signals

#Generation of sinusoidal signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of rotating phasor
t=np.linspace(0,1,100)
f=5
x1=np.exp(1j*2*np.pi*f*t)
x2=np.exp(-1j*2*np.pi*f*t)
#Step 2: Generation of sine and cosine wave
x_cos=(x1+x2)/2
x_sin=(x1-x2)/(2*1j)
#Step 3: Plotting the result
plt.subplot(2,1,1),plt.plot(t,x_cos),plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)')
plt.title('Cosine wave'),plt.subplot(2,1,2),plt.plot(t,x_sin,’r’)
plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)'),plt.title('Sine wave')
plt.tight_layout()

Fig. 1.15 Generation of sinusoidal signals from complex exponential functions

The python code, which performs the task mentioned above, is shown in
Fig. 1.13, and the corresponding output is shown in Fig. 1.14.
Inferences
From Figs. 1.13 and 1.14, the following inferences can be made:
1. Two complex exponential signals x1(t) = ej2πft x2(t) = e-j2πft with the frequency
value f = 5 Hz are generated. The signals x1(t) and x2(t) look alike.
2. The magnitude and phase responses of the two signals are plotted. The magnitude
of the signal x1(t) = ejΩt is given by jx1 ðt Þj = cos 2 ðΩt Þ þ sin 2 ðΩt Þ = 1.
Similarly, the magnitude of the signal x2(t) = e-jΩt is given by
j x2 ð t Þ j = cos 2 ðΩt Þ þ sin 2 ðΩt Þ = 1. Thus, the magnitudes of the two signals
are alike.
sinðΩt Þ
3. The phase of the signal x1(t) = ejΩt is expressed as ϕðx1 ðt ÞÞ = tan - 1 cosðΩt Þ .
-1
Upon simplifying the expression, we get ϕ(x1(t)) = tan (tan(Ωt)), which results
in ϕ(x1(t)) = Ωt. The phase of the signal x2(t) = e-jΩt is expressed as
sinðΩt Þ
ϕðx1 ðt ÞÞ = tan - 1 - cosðΩt Þ . Upon simplifying the expression, we get ϕ-
-1
(x1(t)) = tan (- tan (Ωt)), which results in ϕ(x1(t)) = - Ωt. The phases of
the two signals are different. This implies that the signal ej2πft and e-j2πft repre-
sents two phasors, rotating in the opposite direction.
Experiment 1.8 Generation of ‘Sine’ and ‘Cosine’ Functions from ‘Complex
Exponential Function’
This experiment aims to prove that sinusoidal signal can be generated through two
phasors rotating in the opposite direction. Mathematically it is expressed as
1.1 Continuous-Time Signal 15

Cosine wave
1
Amplitude (V)
0

–1
0.0 0.2 0.4 0.6 0.8 1.0
Time (t)
Sine wave
1
Amplitude (V)

–1
0.0 0.2 0.4 0.6 0.8 1.0
Time (t)

Fig. 1.16 Sine and cosine wave from complex exponential functions

- jΩt - jΩt
cosðΩt Þ = e þe2 and sinðΩt Þ = e -2je . The python code, which generates the
jΩt jΩt

cosine and sine wave using a rotating phasor, is shown in Fig. 1.15, and the
corresponding output is shown in Fig. 1.16.
Inference
From Fig. 1.16, it is possible to observe that there is a phase difference of 90°
between the sine and cosine waveforms.
Task
1. Add the square of the sine and cosine wave obtained in Experiment 1.8, and plot
the resultant waveform. Comment on the observed output. [Hint:
sin2(θ) + cos2(θ) = 1]
Experiment 1.9 Modulating Sinusoidal Signal with an Exponential Signal
This experiment discusses the sinusoidal signal multiplied with a growing and
decaying real exponential signal. The python code, which accomplishes this task,
is shown in Fig. 1.17, and the corresponding output is shown in Fig. 1.18.
Inference
The following inferences can be made from Fig. 1.18
1. The sinusoidal signal amplitude varies between -1 and +1.
2. Upon multiplying the sinusoidal signal with growing exponential, the amplitude
value increases; hence, the plot is shown in the range -5 to +5.
3. Upon multiplying the sinusoidal signal with decaying exponential, the amplitude
of the input sinusoidal signal decreases, which is shown between -0.5 and +0.5.
16 1 Generation of Continuous-Time Signals

#Multiplying sinusoidal signal with an exponential signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Defining the time axis
t=np.linspace(0,1,1000)
#Step 2: Defining the parameter
a=2 #Parameter for exponentially growing function
b=-2 #Parameter for exponentially decaying function
#Step 3: Generation of function
x1=np.sin(2*np.pi*5*t)
x2=np.exp(a*t)*np.sin(2*np.pi*5*t)
x3=np.exp(b*t)*np.sin(2*np.pi*5*t)
#Step 4: Plotting of the function
plt.subplot(3,1,1),plt.plot(t,x1)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Sinusoidal signal')
plt.subplot(3,1,2),plt.plot(t,x2)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Exponentially increasing sinusoidal signal')
plt.subplot(3,1,3),plt.plot(t,x3)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Exponentially decaying sinusoidal signal')
plt.tight_layout()

Fig. 1.17 Python code to modulate sinusoidal signal by exponential signal

Fig. 1.18 Result of python code shown in Fig. 1.17

1.2 Non-stationary Signal 17

1.2 Non-stationary Signal

A stationary signal is one whose statistical characteristics do not change with respect
to time. If the signal characteristics change with respect to time, then it is a non-
stationary signal. Example of non-stationary signal is a chirp signal whose frequency
varies with respect to time. Most of the real-world signals, like the alarm sound from
the clock or the sound of an ambulance, are non-stationary. In this section, few
stationary and non-stationary signals are generated.
Experiment 1.10 Generation of Stationary and Non-stationary Signal
This experiment deals with the generation of stationary and non-stationary signal.
The expression for stationary signal is given by

x1 ðt Þ = sinð2πft Þ ð1:7Þ

The above expression generates a sinusoidal signal of specific frequency. The

expression for non-stationary signal is given by

x2 ðt Þ = sin 2πft2 ð1:8Þ

The frequency of the signal changes with respect to time; hence, it is considered
as non-stationary. The python code, which generates the two signals and the
corresponding output, is shown in Fig. 1.19, and Fig. 1.20.

#Stationary and non-stationary signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of signals
t=np.linspace(0,1,1000)
f=5
x1=np.sin(2*np.pi*f*t)
x2=np.sin(2*np.pi*f*t**2)
#Step 2: Plotting of signals
plt.subplot(2,1,1),plt.plot(t,x1),plt.xlabel('Time (t)'),
plt.ylabel('Amplitude'),plt.title('Stationary signal')
plt.subplot(2,1,2),plt.plot(t,x2,'r'),plt.xlabel('Time (t)'),
plt.ylabel('Amplitude'),plt.title('Non-stationary signal')
plt.tight_layout()

Fig. 1.19 Python code to generate stationary and non-stationary signal

18 1 Generation of Continuous-Time Signals

Stationary signal
1
Amplitude

–1
0.0 0.2 0.4 0.6 0.8 1.0
Time (t)
Non-stationary signal
1
Amplitude

–1
0.0 0.2 0.4 0.6 0.8 1.0
Time (t)

Fig. 1.20 Result of python code shown in Fig. 1.19

Inference
By observing Fig. 1.20, it is possible to conclude that the frequency of the stationary
signal does not change with respect to time. On the other hand, the frequency of the
non-stationary signal increases with time increases.
Experiment 1.11 Generation of Non-stationary Sinusoidal Signal
The objective of this experiment is to append sinusoidal signals of different frequen-
cies. Signal-1 is generated with 5 Hz frequency appearing first, DC signal next and
10 Hz frequency occurs last. Signal-2 is obtained by interchanging the first and last
part of signal-1, which means high frequency occurs first and low frequency occurs
next. The python code, which performs this task, is shown in Fig. 1.21, and the
corresponding output is shown in Fig. 1.22.
Inference
The signals in Fig. 1.22 are considered as non-stationary, because the signal fre-
quency varies with respect to time. Signal-1 and Signal-2 contain the same frequency
components at different instants.
Experiment 1.12 Generation of Chirp Signal
The objective of this experiment is to generate chirp signal. The chirp signal can be
considered as a frequency swept sinusoidal signal. Four different methods of fre-
quency sweep are (1) linear, (2) quadratic, (3) logarithmic and (4) hyperbolic. In this
experiment, the frequency sweep is from 10 Hz to 1 Hz, as considered. The python
code, which generates the chirp signals, are shown in Fig. 1.23, and the
corresponding output is shown in Fig. 1.24.
1.2 Non-stationary Signal 19

import numpy as np
import matplotlib.pyplot as plt
t1=np.linspace(0,1,100)
#Defining signal frequencies
f1,f2,f3=0,5,10
#Generation of signal-1
x1=np.sin(2*np.pi*f2*t1)
#Generation of signal-2
x2=np.sin(2*np.pi*f1*t1)
x3=np.sin(2*np.pi*f3*t1)
x=np.concatenate([x1,x2,x3])
y=np.concatenate([x3,x2,x1])
#Plotting the result
t=np.linspace(0,1,300)
plt.subplot(2,1,1),plt.plot(t,x),plt.xlabel('Time(t)'),plt.ylabel('Amplitude (V)')
plt.title('Signal-1'),plt.subplot(2,1,2),plt.plot(t,y)
plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)'),plt.title('Signal-2')
plt.tight_layout()

Fig. 1.21 Generation of non-stationary signal

Signal-1
1
Amplitude (V)

–1
0.0 0.2 0.4 0.6 0.8 1.0
Time (t)
Signal-2
1
Amplitude (V)

–1
0.0 0.2 0.4 0.6 0.8 1.0
Time (t)

Fig. 1.22 Result of non-stationary signals

Inference
From Fig. 1.24, it is possible to observe that the frequency varies with respect to time
in all four types of chirp signals; hence, they are considered non-stationary signals.
20 1 Generation of Continuous-Time Signals

#Generation of chirp signals

import numpy as np
import matplotlib.pyplot as plt
from scipy.signal import chirp
#Step 1: Generation of chirp signal
t=np.linspace(0,10,10000)
x1= chirp(t, f0=10, f1=1, t1=10, method='linear')
x2= chirp(t, f0=10, f1=1, t1=10, method='quadratic')
x3= chirp(t, f0=10, f1=1, t1=10, method='logarithmic')
x4= chirp(t, f0=10, f1=1, t1=10, method='hyperbolic')
#Step 2: Plotting the signals
plt.subplot(2,2,1),plt.plot(t,x1),plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)'),
plt.title('Linear chirp'),plt.subplot(2,2,2),plt.plot(t,x2),plt.xlabel('Time (t)'),
plt.ylabel('Amplitude (V)'),plt.title('Quadratic chirp'),plt.subplot(2,2,3),
plt.plot(t,x3),plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)'),
plt.title('Logarithmic chirp'),plt.subplot(2,2,4),plt.plot(t,x4),plt.xlabel('Time (t)'),
plt.ylabel('Amplitude (V)'),plt.title('Hyperbolic chirp')
plt.tight_layout()

Fig. 1.23 Generation of chirp signals

Linear chirp Quadratic chirp

1 1
Amplitude (V)

Amplitude (V)

0 0

–1 –1
0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0
Time (t) Time (t)

Logarithmic chirp Hyperbolic chirp

1 1
Amplitude (V)

Amplitude (V)

0 0

–1 –1
0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0
Time (t) Time (t)

Fig. 1.24 Simulation result of chirp signals

1.3 Non-sinusoidal Waveform 21

1.3 Non-sinusoidal Waveform

The non-sinusoidal waveform generation considered in this section includes square

waveform, triangular waveform, sawtooth waveform, sinc and Gaussian signals.

1.3.1 Square Waveform

A square waveform is a non-sinusoidal periodic waveform. A square wave repre-

sents a sudden variation from ‘ON’ to ‘OFF’ state and vice versa. Duty cycle is the
percentage of time a square wave remains high versus low over one period. Square
waves are useful in modelling digital signals. Sine wave contains single frequency,
whereas square wave contains a very wide bandwidth of frequencies.
Experiment 1.13 Generation of Square Waveform
The objective is to generate square wave with different duty cycle. The duty option
refers to which fraction of the whole duty cycle the signal will be in its ‘high’ state.
The python code, which generates square wave of frequency 5 Hz, is shown in
Fig. 1.25, and the corresponding output is shown in Fig. 1.26.
Inference
From Fig. 1.26, it is possible to interpret that the generated waveform is a square
waveform of a fundamental frequency of 5 Hz. With increase in the duty, the ‘ON
time’ of the generated square wave increases. The square waveform takes only
binary value, which is either +1 or -1. The state change from +1 to -1 and -1 to
+1 occurs immediately.

#Square wave with a different duty cycle

import numpy as np
from scipy import signal
import matplotlib.pyplot as plt
t=np.linspace(0,1,100)
f=5
duty=[0.15,0.25,0.5,0.75]
for i in range(len(duty)):
x=signal.square(2*np.pi*f*t,duty[i])
plt.subplot(2,2,i+1)
plt.plot(t,x),plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)')
plt.ylim(-2,2),plt.title('Square wave (duty={})'.format(duty[i]))
plt.tight_layout()

Fig. 1.25 Generation of a square wave

22 1 Generation of Continuous-Time Signals

Square wave (duty=0.15) Square wave (duty=0.25)

2 2
Amplitude (V)

Amplitude (V)
1 1

0 0
–1 –1

–2 –2
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time (t) Time (t)

Square wave (duty=0.5) Square wave (duty=0.75)

2 2
Amplitude (V)

Amplitude (V)
1 1

0 0
–1 –1

–2 –2
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time (t) Time (t)

Fig. 1.26 Result of python code shown in Fig. 1.25

1.3.2 Triangle and Sawtooth Waveform

Triangle and sawtooth waveforms are useful for exploring non-linearity in the
circuit. A triangle waveform has uniform rise and fall time, whereas in a sawtooth
waveform, the rise and fall times are markedly different.
Experiment 1.14 Generation of Sawtooth and Triangular Waveforms
The python code, which generates the sawtooth waveform of frequency 5 Hz, is
shown in Fig. 1.27, and the corresponding output is shown in Fig. 1.28.
Inference
When the width is 0.5, the sawtooth waveform turns out to be a triangular waveform.
In square waveform, the state change from -1 to +1 and from +1 to -1 occurs
instantaneously, whereas, in triangular waveform, the change of state from -1 to +1
and from +1 to -1 occurs gradually.

1.3.3 Sinc Function

A sinc function is represented as

1.3 Non-sinusoidal Waveform 23

#Sawtooth wave with different width

import numpy as np
from scipy import signal
import matplotlib.pyplot as plt
t=np.linspace(0,1,1000)
f=5
width=[0.1,0.2,0.5,0.8]
for i in range(len(width)):
x=signal.sawtooth(2*np.pi*f*t,width[i])
plt.subplot(2,2,i+1)
plt.plot(t,x),plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)')
plt.ylim(-2,2),plt.title('Sawtooth wave (width={})'.format(width[i]))
plt.tight_layout()

Fig. 1.27 Python code to generate sawtooth waveform

Sawtooth wave (width=0.1) Sawtooth wave (width=0.2)

2 2
Amplitude (V)

Amplitude (V)

1 1

0 0
–1 –1

–2 –2
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time (t) Time (t)

Sawtooth wave (width=0.5) Sawtooth wave (width=0.8)

2 2
Amplitude (V)

Amplitude (V)

1 1
0 0
–1 –1

–2 –2
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time (t) Time (t)

Fig. 1.28 Simulation result of sawtooth waveform

sin πt
xð t Þ = , -1<t<1 ð1:9Þ
πt

A sinc function is an even function with a unit area. It is a symmetric function

with respect to the origin. Fourier transform of sinc function will result in rectangular
function and vice versa. Thus, sinc function is the impulse response of the ideal
lowpass filter.
24 1 Generation of Continuous-Time Signals

Fig. 1.29 Python code to

generate sinc function #Generation of sinc function
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Defining the independent variable
t=np.linspace(-10,10,1000)
#Step 2: Generating sinc function
x=np.sinc(t)
#Step 3: Plotting the sinc function
plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Sinc function')
plt.tight_layout()

Sinc function
1.0

0.8

0.6
Amplitude

0.4

0.2

0.0

–0.2

–10.0 –7.5 –5.0 –2.5 0.0 2.5 5.0 7.5 10.0

Time

Fig. 1.30 Result of python code shown in Fig. 1.29

Experiment 1.15 Generation of Sinc Function

The expression for sinc function is given by sin cðt Þ = sinπtðπtÞ. In this experiment, the
sinc function is generated using the built-in function (sinc()) available in numpy
library. The python code, which generates the sinc function, is shown in Fig. 1.29,
and the corresponding output is shown in Fig. 1.30.
Inference
From Fig. 1.30, the following observations can be made:
1. Sinc function has a main lobe and side lobes.
2. The sinc function is symmetric with respect to the origin. It is an even function.
3. The sinc function attains the maximum value at the origin.
1.3 Non-sinusoidal Waveform 25

Task
1. Write a python code to prove that sinc function is an even function.

1.3.4 Pulse Signal

A rectangular pulse can be considered as a positive going edge, followed by negative

going one. Convolution of two rectangular pulses results in a triangular pulse.
Experiment 1.16 Generation of Rectangular and Triangular Pulse Signal
1, jt j < 1
The expression for rectangular pulse is given by xðt Þ = . It can be
0, otherwise
considered as a positive going edge followed by negative going one. Rectangular
pulse represents a drastic variation from level 0 to 1 and from 1 to 0. The expression
jt j
1 - , for jt j < T
for the triangular pulse is xðt Þ = T . The triangular pulse rep-
0, otherwise
resents a gradual variation from level 0 to 1 and from 1 to 0.
The python code, which generates the rectangular and triangular pulse signal, is
shown in Fig. 1.31, and the corresponding output is shown in Fig. 1.32.
Inference
Figure 1.32 shows that rectangular pulse exhibits a drastic change in amplitude from
0 to 1 V, whereas triangular pulse exhibits a gradual variation in amplitude from 0 to
1 V. In later section, it will be proved that the convolution of two rectangular pulses
will result in a triangular pulse.

#Generation of rectangular and triangular pulse signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of signals
t=np.linspace(-2,2,100)
rect_pulse=abs(t)<1 #Rectangular pulse
tri_pulse=(1 - abs(t)) * (abs(t) < 1) #Triangular pulse
#Step 2: Plotting of the pulse signals
plt.subplot(2,1,1),plt.plot(t,rect_pulse)
plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)'),plt.title('Rectangular pulse')
plt.subplot(2,1,2),plt.plot(t,tri_pulse)
plt.xlabel('Time (t)'),plt.ylabel('Amplitude (V)'),plt.title('Triangular pulse')
plt.tight_layout()

Fig. 1.31 Generation of rectangular and triangular pulse

26 1 Generation of Continuous-Time Signals

Fig. 1.32 Rectangular and triangular pulse signal

Task
1. Write a python code to illustrate the fact that convolution of two rectangular pulse
signals results in a triangular pulse.

1.3.5 Gaussian Function

The Gaussian function is expressed as

1 ðt - μÞ2
xð t Þ = p e - 2σ2 ð1:10Þ
2π σ

where ‘μ’ represents the mean and ‘σ’ represents the standard deviation. Fourier
transform of a Gaussian function results in another Gaussian function. The product
of two Gaussian functions is a Gaussian function. Gaussian window is an optimal
window for time-frequency localization. Smoothening by Gaussian function is
widely employed in image processing.
Experiment 1.17 Generation of Gaussian Function
The Gaussian function is widely used in signal processing, image processing and
communication fields. The expression for Gaussian function with the mean value ‘μ’
ðt - μ Þ2
and standard deviation ‘σ’ is given by xðt Þ = p1 e - 2σ 2 . This experiment aims to
2π σ
1.3 Non-sinusoidal Waveform 27

#Generation of Gaussian function

import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(-10,10,1000)
mu=0 #Mean value
sigma=[0.01,0.5,1,10] #Standard deviation
for i in range(len(sigma)):
k=1/np.sqrt(2*np.pi*sigma[i])
x=k*np.exp(-np.power(t-mu,2.)/2*np.power(sigma[i],2.))
plt.subplot(2,2,i+1),plt.plot(t,x),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('$\sigma={} $'.format(sigma[i]))
plt.tight_layout()

Fig. 1.33 Gaussian function for different values of standard deviation

Fig. 1.34 Gaussian function for different values of standard deviation

generate Gaussian function for different values of standard deviation, namely, 0.01,
0.5, 1 and 10. The mean value is taken as zero. The python code, which generates the
Gaussian function, is shown in Fig. 1.33, and the corresponding output is shown in
Fig. 1.34.
Inference
From Fig. 1.34, it is possible to observe the following facts:
28 1 Generation of Continuous-Time Signals

Fig. 1.35 Python code to

generate sinusoidal note #Hearing sinusoidal signal
import numpy as np
import sounddevice as sd
f=1000 #Signal frequency
fs=8000 #Sampling rate
t=np.linspace(0,1,fs)
x=np.sin(2*np.pi*f*t)
sd.play(x,fs)

1. The Gaussian function is characterized by two parameters, which are mean and
standard deviation.
2. The mean value of the Gaussian function is zero; hence, the maximum value
occurs at the origin.
3. With an increase in the value of standard deviation, the narrower the Gaussian
function.
Task
1. Write a python code to prove that the multiplication of two Gaussian functions
results in a Gaussian function.
Experiment 1.18 Hearing a Sinusoidal Signal
Human ears can hear sound in the frequency range from 20 Hz to 20 kHz. In this
experiment, sine wave of particular frequency is heard as a tone. The sampling
frequency is chosen as 8000 Hz, and the signal frequency is chosen as 1000 Hz. The
library functions used are (1) Numpy and (2) Sounddevice. The built-in function in
sound device library (sd.play) is used to play the sound. The python code, which
generates the sinusoidal tone, is shown in Fig. 1.35. The user can hear the audio
using headphone.
Inference
From Fig. 1.35, the following inferences can be made:
1. The signal frequency is 1000 Hz, and the sampling frequency is 8000 Hz.
2. The library used to hear the audio is ‘sounddevice’ library.
3. The built-in function (sd.play) is used to hear the audio.
Task
1. Human ear can hear an audio signal whose frequency is between 20 Hz and
20 kHz. Generate 10 Hz sinusoidal waveform; try to hear the waveform. It should
not be audible. Now increase the frequency of sine wave to 100 Hz; now it should
be possible to hear the sinusoid as a single note.
Experiment 1.19 Hearing Amplitude Modulated Sinusoidal Signal
The impact of modulating the amplitude of the sinusoidal signal is observed in this
experiment. In this experiment, the amplitude of the sinusoidal signal is modulated
by both exponentially decaying and growing functions. The python code, which
1.3 Non-sinusoidal Waveform 29

Fig. 1.36 Amplitude

modulated sinusoidal signal #Hearing amplitude modulated sinusoidal signal
import numpy as np
import sounddevice as sd
f=1000 #Signal frequency
fs=8000 #Sampling rate
a=-5 #Decaying factor
b=5 #Growing factor
t=np.linspace(0,1,fs)
x1=np.exp(a*t)*np.sin(2*np.pi*f*t)
x2=np.exp(b*t)*np.sin(2*np.pi*f*t)
sd.play(x1,fs)
sd.wait()
sd.play(x2,fs)

Table 1.4 Built-in functions used in Experiment 1.19

S. No. Built-in function used Purpose
1 np.exp() To generate an exponential function
2 np.sin() To generate a sinusoidal function
3 sd.play() To play the audio signal
4 sd.wait() To pause the audio signal

performs this task is shown in Fig. 1.36. The built-in functions used in the program
are summarized in Table 1.4.
Inference
The following inferences can be made from Fig. 1.36:
1. The signal ‘x1’ refers to a sine wave modulated by an exponentially decaying
function.
2. The signal ‘x2’ refers to a sine wave modulated by an exponentially growing
function.
Experiment 1.20 Generation of Amplitude Modulated Signal
In amplitude modulation, the amplitude of the carrier signal is varied in accordance
with the message signal. The expression for amplitude modulated signal is given by

xðt Þ = ð1 þ m sinð2πf m t ÞÞ sinð2πf c t Þ ð1:11Þ

In the above expression, ‘m’ denotes the modulation index, fm represents the
frequency of the modulating signal and fc denotes frequency of the carrier signal.
The python code, which generates the amplitude modulated signal for different
modulating indices, is shown in Fig. 1.37, and the corresponding output is shown
in Fig. 1.38.
Inference
From Figs. 1.37 and 1.38, the following inferences can be drawn:
30 1 Generation of Continuous-Time Signals

#Amplitude modulation
import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(0,1,1000)
fm=10 #Frequency of modulating signal
fc=100 #Frequency of carrier signal
message=np.sin(2*np.pi*fm*t)
carrier=np.sin(2*np.pi*fc*t)
m=[0.25,0.5,1,1.5] #modulation index
for i in range(len(m)):
mod_sig=(1+m[i]*message)*carrier
plt.subplot(2,2,i+1),plt.plot(t,mod_sig)
plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Modulated signal with m={}'.format(m[i]))
plt.tight_layout()

Fig. 1.37 Python code to generate amplitude modulated signal

Fig. 1.38 Amplitude modulated signal with different modulation indices

1. The frequency of the message signal is 10 Hz; the frequency of the carrier signal
is 100 Hz. The modulation index is varied as 0.25, 0.5, 1.0 and 1.5.
2. It is possible to observe that the amplitude of the carrier is changed in accordance
with the message signal.
1.3 Non-sinusoidal Waveform 31

3. Modulation index less than one corresponds to under modulated signal. Modu-
lation index greater than one corresponds to over modulated signal. Modulation
index equal to one corresponds to perfect modulation.
Exercises
1. Generate the following sinusoidal signal x(t) = A sin (2πft + ϕ) with the
amplitude A = 2 V, frequency f = 10 Hz and phase ϕ = 0. Let the length of
the signal be 100 samples. Store this signal in your system in a particular folder
along with the time stamp in an Excel sheet. From the Excel sheet, read the data
and the time stamp and plot the signal.
2. Write a python code to generate the sinusoidal signal of 1 V amplitude, 5 Hz
frequency and phase ϕ = 0. Mark the positive peak of the waveform. That is the
positive peak of the waveform should be marked with ‘x’ mark.
3. Write a python code to compute the number of zero crossings of sine wave of 2 V
amplitude, 5 Hz frequency and phase ϕ = 0.
4. Write a python code to generate the seven nodes ‘sa’, ‘re’, ‘ga’ and ‘ma’. Use the
sounddevice library to play the seven notes.
ðt - μÞ2
5. Generate the Gaussian function, which is given by xðt Þ = p2π
1
σ
e - 2σ2 for differ-
ent mean values μ = 0, 1, 2, 4 with the fixed standard deviation value σ = 1. Use
subplot to plot the generated Gaussian functions.
Objective Questions
1. What will the signal’s length be if the following code is executed?

A. 10
B. 50
C. 75
D. 100
2. What will be the magnitude of the variables ‘x’ and ‘y’ if the following code
segment is executed?

A. 1 and 0, respectively
B. -1 and 1, respectively
C. 0 and 1, respectively
D. 1 and -1, respectively
3. What will be the output plot if the following segment of code is executed?
32 1 Generation of Continuous-Time Signals

A. DC signal of magnitude 1
B. DC signal of magnitude 5
C. Sine wave of frequency 5 Hz
D. Cosine wave of frequency 5 Hz
4. What will be stored in the variable ‘z’ if executing the following code segment?

A. Sine wave of 5 Hz frequency

B. Cosine wave of 5 Hz frequency
C. Square wave of 5 Hz frequency
D. Sawtooth wave of 5 Hz frequency
5. The phase difference between each signal in a three-phase sinusoidal signal is
A. 45°
B. 90°
C. 120°
D. 240°
6. What will be stored in the variable ‘z’ if executing the following code segment?

A. Phase angle of the signal ‘x’

B. Magnitude of the signal ‘x’
C. Frequency of the signal ‘x’
D. Number of zero crossings of the signal ‘x’
7. The audible frequency range for human beings is
A. 10 Hz to 100 kHz
B. 20 Hz to 20 kHz
C. 1 to 1000 Hz
D. 200 Hz to 2 MHz
8. What will the signal’s length be if the following code segment is executed?
Bibliography 33

A. 1000
B. 2000
C. 4000
D. 8000
9. Identify the statement that is WRONG with respect to sinc function
A. Sinc function is an even function.
B. Sinc function is an odd function.
C. Fourier transform of sinc function will result in a rectangular function.
D. Sinc function can be used for signal interpolation.
10. The magnitude of the function x(t) = e-jΩt is
A. 1
B. 0
C. -1
D. Infinity

Bibliography

1. Alan V. Oppenheim, and Alan S. Willsky. “Signals and Systems”, Prentice Hall, 1996.
2. Simon Haykin, and Bary Van Veen, “Signals and Systems”, Wiley, 2005.
3. Hwei P. Hsu, “Signals and Systems”, Schaum’s outline series, McGraw Hill Education, 2017.
4. Charles L. Phillips, John M. Parr, and Eve A. Riskin, “Signals, Systems, and Transforms”,
Pearson, 2013.
5. Mark Lutz, “Learning Python”, O’Reilly Media, 2013.
Chapter 2
Sampling and Quantization of Signals

Learning Objectives
After reading this chapter, the reader is expected to
• Simulate and visualize standard discrete-time signals.
• Simulate and visualize arbitrary discrete-time signals.
• Perform different mathematical operations on discrete-time signals.
• Implement convolution and correlation operations and interpret the obtained
results.

Roadmap of the Chapter

The contents discussed in this chapter are given as a flow diagram. The objective is
to convert the continuous-time signal into a discrete-time signal. Two important
processes in converting the continuous-time signal into a discrete-time signal are
(1) sampling and (2) quantization. Also, reconstructing the original signal from the
sampled signal is another important task in signal processing. This chapter explores
these three processes in detail.

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 35
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_2
36 2 Sampling and Quantization of Signals

Continuous-Time Signal to Discrete-Time Signal

Sampling of Signal Quantization of Signal Reconstruction of Signal

Time-domain
view of sampling Zero order Hold
Uniform Non-uniform
quantization quantization
Frequency-domain First order Hold
view of sampling
Mid-tread
µ-law
quantizer
companding Ideal Interpolation

Mid-rise A-law
quantizer companding

PreLab Questions
1. Mention the steps involved in converting the analogue signal into a digital
signal.
2. A real-valued signal is known to be bandlimited. The maximum frequency
content in the signal is fmax. What is the guideline given by the sampling theorem
with respect to the choice of sampling frequency such that from the samples, the
signal can be reconstructed without aliasing?
3. What is the impact of sampling a bandlimited signal with too low a sampling
frequency?
4. Is it possible to reconstruct a periodic square wave of fundamental frequency
5 Hz from its samples? Explain your answer.
5. Mention the reason for aliasing to occur while sampling the signals?
6. What is the meaning of sampling the signal x(t)? What is the meaning of the
terms (a) sampling rate and (b) sampling interval?
7. A signal has a bandwidth of 5 kHz. What is the Nyquist rate of the signal?
8. Why quantization is considered as a non-linear phenomenon?
9. Why quantization is considered as irreversible phenomenon?
10. What is signal reconstruction? Mention different types of signal reconstruction
strategies.

2.1 Sampling of Signal

Sampling is basically taking a specific instant of the signal. In time domain, it is

visualized as passing the signal through a switch. Sampling can be considered as
multiplying the continuous-time signal x(t) with train of impulse c(t). The train of
impulse will take a value of either one or zero; hence, the multiplication of the signal
2.1 Sampling of Signal 37

#Generation of comb function

import numpy as np
import matplotlib.pyplot as plt
Fs=100
t = np.arange(0, 2, 1/Fs)
c=np.zeros(len(t))
T = 0.1
c[::int(Fs*T)]=1
plt.stem(t,c),plt.xlabel('Time'),plt.ylabel('Ampliude'),plt.title('c(t)')

Fig. 2.1 Python code to generate comb function

c(t)
1.0

0.8

0.6
Ampliude

0.4

0.2

0.0

0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00

Time

Fig. 2.2 Result of python code shown in Fig. 2.1

x(t) with a train of impulse can be regarded as passing the signal x(t) through a
switch. The expression for a train of impulse is given by

1
cðt Þ = δðt - nT Þ ð2:1Þ
n= -1

The function c(t) takes a value of one whenever t = nT; else, it takes a value of
zero.
Experiment 2.1 Generation of a Train of Impulse Function
The python code, which generates the train of impulse function or comb function, is
given in Fig. 2.1, and the corresponding output is shown in Fig. 2.2.
38 2 Sampling and Quantization of Signals

Inference
1. From Fig. 2.1, it is possible to observe that the variable ‘T’ (Sampling interval)
decides the distance between consecutive samples.
2. From Fig. 2.2, it is possible to confirm that the comb function c(t) takes a value of
either ‘1’ or ‘0’. Whenever c(t) = 1, the signal x(t) samples will be collected.
Experiment 2.2 Frequency Domain View of Comb Function
The time-domain expression for the comb function is given by

1
cðt Þ = δðt - nT Þ ð2:2Þ
n= -1

Upon taking the Fourier transform of the comb function, we get

1
2π
C ð ΩÞ = δðΩ - kΩs Þ ð2:3Þ
T k= -1

This experiment aims to prove that Fourier transform of a train of impulse will
result in a train of impulse function. Here, two comb functions (train of impulse
function), namely, c1(t) and c2(t) are generated. In the comb function c1(t), the
spacing between consecutive impulses is 0.1 s, whereas in the comb function c2(t),
the spacing between successive impulses is 0.05 s. Upon taking Fourier transform of
these two comb functions, the corresponding magnitude spectra |C1( f )| and |C2( f )|
are obtained. In the magnitude spectrum (|C1( f )|), the spacing between successive
peaks is 1/0.1 = 10, whereas in the magnitude spectrum (|C2( f )|), the spacing
between successive peaks is 1/0.05 = 20. The python code that performs this task
is given in Fig. 2.3, and the corresponding output is shown in Fig. 2.4.
Inferences From Fig. 2.4, the following inferences can be drawn:
1. The spacing between two successive samples in the comb function c1(t) is 0.1 s.
2. The spacing between two consecutive peaks in C1( f ) is 10 Hz.
3. The spacing between two successive samples in the comb function c2(t) is 0.5 s.
4. The spacing between two consecutive peaks in C2( f ) is 20 Hz.
5. This experiment illustrates the fact that time and frequency are inversely related to
each other. That is, compression in one domain is equivalent to expansion in other
domain and vice versa.
6. The Fourier transform of a train of impulse function results in a train of impulse
function.
Task
1. Write a python code to generate a function expressed as x½m] = 1
M x
M -1 2π
ej M km , - 10 < m < 10 for M = 1 and M = 2, and comment on the observed
k=0
result.
2.1 Sampling of Signal 39

#Fourier transform of train of impulse

import numpy as np
import matplotlib.pyplot as plt
from scipy.fft import fft,fftshift
#Step 1: Generation of comb functions
Fs=100
t = np.arange(0, 2, 1/Fs)
f = np.linspace(-Fs/2, Fs/2, len(t), endpoint=False)
T1 = 0.1
c1=np.zeros(len(t))
c1[::int(Fs*T1)]=1
T2=0.05
c2=np.zeros(len(t))
c2[::int(Fs*T2)]=1
#Step 2: Fourier transform of comb function
C1=fftshift(fft(c1))
C2=fftshift(fft(c2))
#Step 3: Plotting the result
plt.subplot(2,2,1),plt.stem(t,c1),plt.xlabel('Time'),plt.ylabel('Ampliude'),
plt.title('$c_1(t)$'),plt.subplot(2,2,2),plt.plot(f, np.abs(C1)/len(C1))
plt.xlabel('Frequency'),plt.ylabel('Magnitude'),plt.title('$|C_1(f)|$')
plt.subplot(2,2,3),plt.stem(t,c2),plt.xlabel('Time'),plt.ylabel('Ampliude'),
plt.title('$c_2(t)$'),plt.subplot(2,2,4),plt.plot(f, np.abs(C2)/len(C2))
plt.xlabel('Frequency'),plt.ylabel('Magnitude'),plt.title('$|C_2(f)|$')
plt.tight_layout()

Fig. 2.3 Python code to obtain the spectrum of comb function

2.1.1 Violation of Sampling Theorem

The sampling theorem gives the guideline regarding the choice of the sampling rate.
According to the sampling theorem, a continuous-time signal with frequencies no
higher than fmax (Hz) can be reconstructed exactly from its samples if the samples are
taken at a rate greater than 2fmax. That is, fs ≥ 2fmax. Violation of the sampling
theorem results in an aliasing, which can be visualized in both the time and
frequency domains.
Experiment 2.3 Illustration of Aliasing in Time Domain
In this experiment, the aliasing is visualized in time domain. The analogue signal to
be sampled is represented as x(t) = sin (2πft + ϕ). The frequency of the signal x(t) is
10 Hz, and the phase angle is zero. This signal is sampled at four different sampling
frequencies 8, 15, 50 and 100 Hz. Obviously, the first two sampling frequencies
( fs = 8 and 15 Hz) are less than the criteria specified by the sampling theorem. This
will result in aliasing. The impact of aliasing is visualized in this experiment. The
40 2 Sampling and Quantization of Signals

Fig. 2.4 Fourier transform of comb functions

#Aliasig in time domain

import numpy as np
import matplotlib.pyplot as plt
f=10 #Signal frequency
fs=[8,15,50,100] #Sampling frequencies
for i in range(len(fs)):
t=np.arange(0,1,1/fs[i])
x=np.sin(2*np.pi*f*t)
plt.subplot(2,2,i+1)
plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('$F_s={} $ Hz'.format(fs[i]))
plt.tight_layout()

Fig. 2.5 Python code which illustrates aliasing in time domain

python code that performs this task is shown in Fig. 2.5, and the corresponding
output is shown in Fig. 2.6.
Inferences
From Fig. 2.6, the following inferences can be made:
1. The sampling frequency of 8 Hz is insufficient to capture all the information in the
signal. The frequency of the sampled signal is given by f′ = f - fs. This implies f
2.1 Sampling of Signal 41

Fs = 8 Hz Fs = 15 Hz
1
0.5
Amplitude

Amplitude
0 0.0

–0.5
–1
0.0 0.2 0.4 0.6 0.8 0.00 0.25 0.50 0.75
Time Time

Fs = 50 Hz Fs = 100 Hz
1 1
Amplitude

Amplitude
0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time Time

Fig. 2.6 Result of python code shown in Fig. 2.5

′ = 10 - 8 = 2 Hz. This is the reason that the signal obtained using fs = 8 Hz

resembles that of 2 Hz sinusoidal signal.
2. If the sampling frequency is chosen as 15 Hz, then the Nyquist interval is between
-7.5 and 7.5 Hz. The signal frequency is not within the Nyquist interval; hence,
the frequency of the sampled signal is f′ = f - fs. Upon substituting the value, we
get f′ = 10 - 15 = -5 Hz. This is the reason that the signal obtained using
fs = 15 Hz resembles that of a 5 Hz sinusoidal signal.
3. For the choice of sampling frequency as 50 and 100 Hz, signal frequency lies well
within the Nyquist interval. Hence, no aliasing exists in these cases. As a result,
the 10 Hz signal appeared as 10 Hz for fs = 50 and 100 Hz.
Experiment 2.4 Aliasing in the Time Domain
Generate two sinusoidal signals with a frequency of 1 and 6 Hz. Use a sufficiently
high sampling frequency to plot the generated signal. Now use the sampling
frequency as 5 Hz to plot the 6 Hz frequency component sinusoidal signal, and
comment on the observed output. Illustration of this experiment is shown in Fig. 2.7.
The steps involved in the python code implementation of this experiment are as
follows:
Step 1: Generation of sine wave of 1 and 6 Hz sinusoidal signals. Let it be
represented by the variables ‘x1’ and ‘x2’. ‘x1’ represents a 1 Hz sine wave,
and ‘x2’ represents a 6 Hz sine wave. The sampling frequency chosen is 100 Hz
( fs = 100 Hz), which is sufficient to represent these two signals without
ambiguity.
42 2 Sampling and Quantization of Signals

Generate two sinusoidal signals

with 1 Hz and 6 Hz frequency

Use sufficient sampling Use sampling frequency as 5 Hz to plot

frequency to plot the signal the 6 Hz frequency sinusoidal signal

Fig. 2.7 Illustration of Experiment 2.4

import numpy as np
import matplotlib.pyplot as plt
#Step 1: To generate x1 and x2
f1=1 #Signal frequency
f2=6
fs=100
t=np.arange(0,1,1/fs)
x1=np.sin(2*np.pi*f1*t)
x2=np.sin(2*np.pi*f2*t)
#Step 2: New sampling frequency is 5 Hz
fs1=5
t1=np.arange(0,1.1,1/fs1)
x3=np.sin(2*np.pi*f1*t1)
#Step 3: Plotting the result
plt.plot(t,x1,'k--',t,x2,'k'),#plt.plot(t,x2,'k')
plt.stem(t1,x3,'r'),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.legend(['1 Hz Sine wave','6 Hz Sine wave','Sampling with 5 Hz']),
plt.title('Aliasing in Time Domain')
plt.tight_layout()

Fig. 2.8 Python code to illustrate aliasing in time domain

Step 2: Now, the new sampling frequency chosen is 5 Hz. That is, f′ = 5 Hz. This
sampling frequency is used to represent a 6 Hz sine wave, which is stored in the
variable ‘x3’. It is well-known that 5 Hz is insufficient to represent a sine wave of
6 Hz frequency. Because of aliasing, the new frequency will appear at 1 Hz.
Step 3: From the samples taken using f′ = 5 Hz, it is not possible to distinguish
between 1 and 6 Hz sine waves. This phenomenon is termed as ‘aliasing’. This
occurs due to spectral folding.
The python code used to illustrate this concept is shown in Fig. 2.8, and the
corresponding output is shown in Fig. 2.9.
Inferences
From Fig. 2.9, the following inferences can be made:
2.1 Sampling of Signal 43

Aliasing in Time Domain

1.00 1 Hz Sine wave
6 Hz Sine wave
0.75
Sampling with 5 Hz
0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 2.9 Result of python code shown in Fig. 2.8

1. The solid line shows a sine wave of 6 Hz frequency. The dotted line represents a
sine wave of 1 Hz frequency. Since the sampling frequency is 100 Hz, both
waveforms appear as desired without ambiguity.
2. The new sampling frequency is chosen as 5 Hz. This sampling frequency is used
to represent a 6 Hz sine wave. This sampling frequency is insufficient to represent
the 6 Hz. Represent a 6 Hz sine wave; the sampling frequency should be greater
than 12 Hz. From the discrete samples, it is not possible to interpret whether the
samples are taken from a 6 Hz sine wave or a 1 Hz sine wave. This ambiguity is
termed as aliasing, which arises due to spectral folding.
Experiment 2.5 Illustration of Aliasing in Frequency Domain
The python code, which demonstrates the phenomenon of aliasing in the frequency
domain, is shown in Fig. 2.10. This experiment generates the signal x-
(t) = sin (10πt) + sin (30πt) using two different sampling rates: fs = 50 Hz and
fs = 25 Hz.
Inferences
The following inferences can be made from Fig. 2.11.
1. The frequency components present in the signal x(t) are f1 = 5 Hz and f2 = 15 Hz.
2. When the sampling rate is 50 Hz, the peak in the magnitude spectrum appears
correctly at f1 = 5 Hz and f2 = 15 Hz.
3. On the other hand, if the sampling rate is chosen as fs = 25 Hz, there is no change
with respect to f1 = 5 Hz frequency component, whereas the frequency compo-
nent f2 = 15 Hz appears as f2 = 10 Hz. Observing a 15 Hz frequency component
signal as a 10 Hz frequency component is termed as aliasing.
44 2 Sampling and Quantization of Signals

#Sampling theorem
import numpy as np
import matplotlib.pyplot as plt
from scipy.fft import fft,fftfreq
#Step 1: Generate the two signals
f1=5
f2=15
fs=[25,50]
N=256
for i in range(len(fs)):
T=1/fs[i]
t=np.linspace(0,N*T,N)
x=np.sin(2*np.pi*f1*t)+np.sin(2*np.pi*f2*t)
X=fft(x)
f_axis=fftfreq(N,T)[0:N//2]
plt.subplot(2,1,i+1)
plt.plot(f_axis,2/N*np.abs(X[0:N//2]))
plt.xlabel('$\omega$-->'),plt.ylabel('|X($\omega$)|'),
plt.title(r'Spectrum corresponding to $f_s = {} Hz$'.format(fs[i]))
plt.tight_layout()

Fig. 2.10 Python code to illustrate the concept of aliasing in frequency domain

Spectrum corresponding to fs = 25Hz

|X(w)|

0.5

0.0
0 2 4 6 8 10 12
w ->
Spectrum corresponding to fs = 50Hz
1.0
|X(w)|

0.5

0.0
0 5 10 15 20 25
w ->

Fig. 2.11 Illustration of aliasing in the frequency domain

Task
1. Change the value of the sampling frequency ( fs) in the python code given in
Fig. 2.10, and observe the changes in the output spectrum.
2.1 Sampling of Signal 45

Fig. 2.12 Hearing aliasing

effect #Hearing aliasing
import sounddevice as sd
import numpy as np
import matplotlib.pyplot as plt
fs=1500
dur=1
T=1/fs
t=np.linspace(0,1,dur*fs)
x1=np.sin(2*np.pi*500*t)
x2=np.sin(2*np.pi*1000*t)
x=np.concatenate([x1,x2])
sd.play(x,fs)

Experiment 2.6 Hearing Aliasing

In this experiment, two sinusoidal tones of frequency f1 = 500 Hz and f2 = 1000 Hz
are generated with sampling frequency fs = 1500 Hz. Let x1(t) and x2(t) represent the
two tones. The maximum signal frequency is 1000 Hz. The minimum sampling rate
required is fs = 2000 Hz. Unfortunately, fs is chosen as 1500 Hz. As a result, 1500 Hz
will be heard as 500 Hz. The python code, which illustrates this concept, is given in
Fig. 2.12.
Inference
As per the code shown in Fig. 2.12, two sinusoidal tones of frequencies 500 and
1000 Hz are generated. These two tones are appended. Instead of hearing two notes,
only one note corresponding to the frequency 500 Hz is heard. This is due to the
violation of the sampling theorem. Due to improper sampling, tone of 1000 Hz is
heard as a tone of 500 Hz. To overcome the impact of aliasing, the sampling
frequency has to be chosen properly.
Task
1. Modify the sampling frequency as 8000 Hz and observe its impact.

2.1.2 Quantization of Signal

Quantization is mapping a large set of values to a smaller set of values. It can be

broadly classified into (1) uniform and (2) non-uniform quantization. A uniform
quantizer splits the mapped input signal into quantization steps of equal size. The
uniform scalar quantization can be broadly classified into (1) mid-tread and (2) mid-
rise quantizer.
If ‘N’ bits are used to represent the value of the signal x[n], then there are 2N
distinct values that x[n] can assume. If the xmin and xmax are the minimum and
46 2 Sampling and Quantization of Signals

maximum values taken by the signal x[n], then the dynamic range of the signal is
calculated by

Dynamic range = xmax - xmin ð2:4Þ

2.1.2.1 Mid-Tread Quantizer

The relationship between the input and output of a mid-tread uniform quantizer is
given by

x ½ n] 1
y½n] = Q × þ ð2:5Þ
Q 2

In the above equation, x[n] represents the input signal to be quantized and y[n]
represents the quantized signal, ‘Q’ denotes the quantization step size and the
symbol bc denotes flooring operation. The expression for quantization step size
can be computed by

Dynamic range
Q= ð2:6Þ
L

where ‘dynamic range’ represents the difference between the maximum and mini-
mum value of the signal and ‘L’ denotes the number of reconstruction levels.
The expression for the number of reconstruction levels is given by

L = 2b ð2:7Þ

In the above expression, ‘b’ is the number of bits used to represent the signal.
Experiment 2.7 Transfer Characteristics of Mid-Tread Quantizer
The aim of this experiment is to plot the transfer characteristics of mid-tread
quantizer for different bit-rate. The bit-rate (b) chosen is b = 1, 2, 4 and 8. The
python code, which performs this task, is shown in Fig. 2.13, and the corresponding
output is shown in Fig. 2.14.
Inferences
The following inferences can be drawn from Figs. 2.13 and 2.14, which are
summarized below:
1. From Fig. 2.13, it is possible to observe that the input signal is represented as the
variable ‘x’ and the quantized signal Q(x) is represented as ‘y’. The input signal
‘x’ varies from -20 to +20; hence, the dynamic range of ‘x’ is 40.
2. Figure 2.13 shows that the number of bits used to represent the input signal is
varied as 1, 2, 4 and 8. It is represented as the variable ‘b’ in the code.
2.1 Sampling of Signal 47

#Transfer characteristics of mid-tread quantizer

import numpy as np
import matplotlib.pyplot as plt
x=np.linspace(-20,20)
DR=np.max(x)-np.min(x) #Dynamic range
b=[1,2,4,8] #Bits
for i in range(len(b)):
L=2**b[i] #Reconstruction level
q=DR/L
#Mid-tread quantizer
y=np.sign(x)*q*np.floor((abs(x)/q)+(1/2))
plt.subplot(2,2,i+1),plt.plot(x,y),plt.xlabel('x'),plt.ylabel('Q(x)')
plt.title('Quantizer with b={}' .format(b[i]))
plt.tight_layout()

Fig. 2.13 Python code for transfer characteristics of mid-tread quantizer

Quantizer with b=1 Quantizer with b=2

20 20
Q(x)

Q(x)

0 0

–20 –20
–20 –10 0 10 20 –20 –10 0 10 20
x x

Quantizer with b=4 Quantizer with b=8

20 20
Q(x)

Q(x)

0 0

–20 –20
–20 –10 0 10 20 –20 –10 0 10 20
x x

Fig. 2.14 Transfer characteristics of mid-tread quantizer

3. From Fig. 2.14, it is possible to observe that the transfer characteristics of a

uniform quantizer is similar to that of a stair-step waveform at low bit rate.
4. At high bit rate (b = 8), the relationship between the input signal (x) and the
quantized signal (Q(x)) is a straight line. This implies that the output follows the
input; hence, the error due to quantization will be zero.
5. From Fig. 2.14, it is possible to observe that the number of reconstruction levels
depends on the number of bits used to represent the signal.
48 2 Sampling and Quantization of Signals

#Uniform Quantization
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generate the input signal
t=np.linspace(0,1,100)
x=signal.sawtooth(2*np.pi*5*t)
#Step 2: Parameters of the quantizer
DR=np.max(x)-np.min(x) #Dynamic range
b=[1,2,4,8] #Number of bits
for i in range(len(b)):
L=2**b[i] #Quantization level
q=DR/(L) #Quantization step size
#Step 3: To obtain the quantized signal
y=np.sign(x)*q*np.floor((abs(x)/q)+(1/2))
plt.figure(i+1)
plt.plot(t,x,'b',t,y,'r'),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.legend(['Input signal','Quantized Signal'],loc='upper right')
plt.title('Quantization with b={}'.format(b[i]))
plt.tight_layout()

Fig. 2.15 Python code to perform uniform mid-tread quantization of the signal

6. The stair tread in a ladder is the horizontal walking surface of an individual step.
From Fig. 2.14, it is possible to observe that mid-tread quantizer has a zero-
valued reconstruction level.
Tasks
1. Write a python code to plot the error signal. The error signal is the difference
between the input and quantized signals. Comment on the observed output.
2. Write a python code to illustrate the fact that quantization error follows a uniform
distribution.
Experiment 2.8 Quantization of Input Sawtooth Signal Using Mid-Tread
Quantizer
The objective of this python experiment is to perform uniform mid-tread quantiza-
tion of input sawtooth signal of 5 Hz frequency for different bit rate. The number of
bits used to represent the input signal varies as 1, 2, 4 and 8. With an increase in the
number of bits used to represent the signal, the quantized signal resembles the input
signal. The python code to verify this experiment is shown in Fig. 2.15, and its
simulation result is displayed in Fig. 2.16.
Inferences
The following are the inferences can be drawn from Fig. 2.16:
1. The input signal to be quantized is a sawtooth signal whose fundamental fre-
quency is 5 Hz.
2.1 Sampling of Signal 49

Fig. 2.16 Result of uniform mid-tread quantization

2. The input signal will be uniformly quantized by mid-tread quantizer for different
bit rates.
3. It is possible to observe that the quantized signal resembles the input signal with
an increase in bit-rate.

2.1.3 Mid-Rise Quantizer

The relationship between the input and output of mid-rise uniform quantizer is given
by

x ½ n] 1
y ½ n] = Q × þ ð2:8Þ
Q 2

In the above equation, x[n] represents the input signal to be quantized and y[n]
represents the quantized signal, ‘Q’ denotes the quantization step size and the
symbol bc denotes flooring operation.
Experiment 2.9 Transfer Characteristics of Mid-Rise Quantizer
The aim of this experiment is to plot the transfer characteristics of mid-rise quantizer
for different bit-rate. The bit-rate (b) chosen is b = 1, 2, 4 and 8. The python code,
50 2 Sampling and Quantization of Signals

Fig. 2.17 Python code for

transfer characteristics of #Transfer characteristics of mid-rise quantizer
mid-rise quantizer import numpy as np
import matplotlib.pyplot as plt
x=np.linspace(-20,20)
DR=np.max(x)-np.min(x) #Dynamic range
b=[1,2,4,8] #Bits
for i in range(len(b)):
L=2**b[i] #Reconstruction level
q=DR/L
#Mid-rise quantizer
y=np.sign(x)*q*(np.floor((abs(x)/q))+(1/2))
plt.subplot(2,2,i+1)
plt.plot(x,y),plt.xlabel('x'),plt.ylabel('Q(x)')
plt.title('Quantizer with b={}' .format(b[i]))
plt.tight_layout()

Quantizer with b=1 Quantizer with b=2

20
20
Q(x)
Q(x)

0 0

–20
–20

–20 –10 0 10 20 –20 –10 0 10 20

x x

Quantizer with b=4 Quantizer with b=8

20 20
Q(x)

Q(x)

0 0

–20 –20
–20 –10 0 10 20 –20 –10 0 10 20
x x

Fig. 2.18 Transfer characteristics of mid-rise quantizer

which performs this task, is shown in Fig. 2.17, and the corresponding output is
shown in Fig. 2.18.
Inferences
From Fig. 2.18, it is possible to observe that the reconstruction level rises to the next
level at the origin; hence, it is termed as ‘mid-rise quantizer’. It is also possible to
observe that with the bit rate increase, the output follows the input. In other words,
the quantizer error is minimal with a bit rate increase.
2.1 Sampling of Signal 51

#Uniform mid-rise quantizer

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generate the input signal
t=np.linspace(0,1,100)
x=signal.sawtooth(2*np.pi*5*t)
#Step 2: Parameters of the quantizer
DR=np.max(x)-np.min(x) #Dynamic range
b=[1,2,4,8] #Number of bits
for i in range(len(b)):
L=2**b[i] #Quantization level
q=DR/(L) #Quantization step size
#Step 3: To obtain the quantized signal
y=np.sign(x)*q*(np.floor((abs(x)/q))+(1/2))
plt.figure(i+1)
plt.plot(t,x,'b',t,y,'r'),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.legend(['Input signal','Quantized Signal'],loc='upper right')
plt.title('Quantization with b={}'.format(b[i]))
plt.tight_layout()

Fig. 2.19 Python code to perform uniform mid-rise quantization

Task
In the python code given in Fig. 2.17, replace ‘np.floor()’ by ‘np.ceil()’ function, and
comment on the change in the transfer characteristics.
Experiment 2.10 Quantization of Input Sawtooth Signal Using Mid-Rise
Quantizer
The objective of this experiment is to perform uniform mid-rise quantization of the
input sawtooth signal for different bit rate. The python code, which performs this
task, is shown in Fig. 2.19, and the corresponding output is shown in Fig. 2.20.
Inference
From Fig. 2.20, it is possible to interpret that with the increase in the number of bits
used to represent the signal, the quantized signal resembles the input signal. In other
words, the error due to quantization will be minimum with the increase in the number
of bits used to represent the signal.
Experiment 2.11 Quantization of Speech Signal
The objective of this experiment is to analyse the performance of uniform mid-tread
quantizer for the speech signal. The experiment consists of two steps. Reading the
speech signal from a given location is the first step, and performing uniform
midtread-quantization of the input speech signal for different bit rates is the second
step. The python code, which does this task, is shown in Fig. 2.21, and the
corresponding output is shown in Figs. 2.22 and 2.23.
52 2 Sampling and Quantization of Signals

Fig. 2.20 Results of uniform mid-rise quantizer

Inference
The following inference can be made from this experiment:
1. The input speech signal belongs to the uttered word ‘Hello’.
2. The quantized signal resembles the original speech signal with the increase in the
number of bits of the quantizer.
Experiment 2.12 Uniform Mid-Tread Quantization of Image
In this experiment, a greyscale image, whose intensity varies gradually from black to
white, is generated first. This image is subjected to uniform quantization with bit
rates 1, 2, 4 and 8. The python code, which performs this task, is shown in Fig. 2.24,
and the corresponding output is shown in Fig. 2.25.
Inferences
The following inferences can be drawn from this experiment:
1. The grey level of the input image varies gradually from black to white.
2. The input image is quantized uniformly with a bit rate of b = 1, 2, 4 and 8. When
b = 1, the number of grey levels used to represent the image is minimum. The
quantized image is different from the input image.
3. With the increase in the number of bits used to represent the pixel value, the
quantized image resembles the input image.
2.1 Sampling of Signal 53

#Uniform quantization of speech signal

from scipy.io import wavfile
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Reading of speech waveform
samplerate, x = wavfile.read('C:\\Users\\Admin\\Desktop\\speech1.wav')
duration = x.shape[0] / samplerate
t = np.linspace(0, duration, x.shape[0])
plt.figure(1)
plt.plot(t,x,'k',linewidth=2)
plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Input speech signal')
#Step 2: Performing uniform quantization of the signal
DR=np.max(x)-np.min(x) #Dynamic range
b=[1,2,4,8] #Number of bits
for i in range(len(b)):
L=2**b[i] #Quantization level
q=DR/L #Quantization step size
#Step 3: To obtain the quantized signal
y=np.floor(x/q)*q-(q/2)
plt.figure(2)
plt.subplot(2,2,i+1)
plt.plot(t,y,'k',linewidth=2),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Quantized signal with b={}'.format(b[i]))
plt.tight_layout()

Fig. 2.21 Performing uniform mid-tread quantization of the speech signal

Input speech signal

0.2

0.0
Amplitude

–0.2

–0.4

–0.6
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
Time

Fig. 2.22 Input speech signal

54 2 Sampling and Quantization of Signals

Quantized signal with b=1 Quantized signal with b=2

–0.25
0.00
Amplitude

Amplitude
–0.50
–0.25
–0.75
–0.50
–1.00
–0.75
0 1 2 0 1 2
Time Time

Quantized signal with b=4 Quantized signal with b=8

0.25
0.25
0.00
Amplitude

Amplitude
0.00
–0.25
–0.25
–0.50
–0.50

0 1 2 0 1 2
Time Time

Fig. 2.23 Uniformly quantized speech signal for different bit-rate

#Uniform mid-tread quantization of image

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of test image
img=np.zeros([256,256])
img[:,0:256]=np.arange(0,256,1)
plt.figure(1)
plt.imshow(img,cmap='gray')
plt.title('Input image')
#Step 2: Parameters of uniform quantizer
DR=np.max(img)-np.min(img) #Dynamic range
b=[1,2,4,8] #Number of bits
for i in range(len(b)):
L=2**b[i] #Quantization level
q=DR/(L) #Quantization step size
#Step 3: To obtain the quantized signal
y=np.sign(img)*q*np.floor((abs(img)/q)+(1/2))
plt.figure(2)
plt.subplot(2,2,i+1)
plt.imshow(y,cmap='gray')
plt.title('b={} '.format(b[i]))
plt.tight_layout()

Fig. 2.24 Uniform mid-tread quantization of the image

2.2 Non-uniform Quantization 55

b=1 b=2
0 0

100 100

200 200

0 100 200 0 100 200

b=4 b=8
0
Input image
0 0
50

100 100
100

150
200 200
200

250
0 50 100 150 200 250
0 100 200 0 100 200

Fig. 2.25 Input and Output of uniform mid-tread quantizer

Task
1. Generate a 256 × 256 image in which half of the pixels are white (grey level 255)
and half of the pixels are black (grey level 0). The columns 0 to 127 is white,
whereas column 128 to 256 is black. Try to quantize this image for different bit
rate and comment on the observed result.

2.2 Non-uniform Quantization

One way to construct non-uniform quantizer is to perform companding.

Companding = Compression + Expanding
The three steps involved in companding are (1) compression, (2) uniform quan-
tization and (3) expanding. In the first step, the input signal is applied to a logarith-
mic function, and the output of this function is given to a uniform quantizer. Finally,
the inverse of the logarithmic function is applied to the output of the quantizer. There
are two standards for non-uniform quantizer companding. They are (1) μ-law
companding for North America and (2) A-law companding for Europe.
The μ-law compression expression in terms of the input signal x(t) is expressed as

ð1 þ μjxjÞ
x1 ðt Þ = sgnðxÞ ln ð2:9Þ
lnð1 þ μÞ
56 2 Sampling and Quantization of Signals

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generate the input signal
t1=np.linspace(0,1,100)
x=signal.sawtooth(2*np.pi*5*t1)
#Step 2: Mu law Encoding (Non-uniform encoding)
mu=255 # 8 bit Quantization
y1=np.sign(x)*((np.log(1+(mu*abs(x))))/np.log(1+mu))
plt.figure(1)
plt.plot(t1,x,'b',t1,y1,'g'),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.legend(['Input signal','Encoded'],loc='upper right')
plt.title('Degree of Compression with mu={}'.format(mu))
#Step 3: Parameters of the quantizer
DR=np.max(y1)-np.min(y1) #Dynamic range
b=[1,2,4,8] #Number of bits
for i in range(len(b)):
L=2**b[i] #Quantization level
q=DR/(L) #Quantization step size
#Step 3: To obtain the quantized signal
y2=np.sign(y1)*q*np.floor((abs(y1)/q)+(1/2))
y=np.sign(y2)*(((1+mu)**(abs(y2))-1)/mu)
plt.figure(i+2)
plt.plot(t1,y2,'r',t1,y),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.legend(['Quantized Before decoding','Non-Uniform Quantized'],loc='upper right')
plt.title('Quantization with b={} and mu={}'.format(b[i],mu))
plt.tight_layout()

Fig. 2.26 Python code for μ-law companding

In the above expression, ‘μ’ is the compression parameter, which is 255 for the
USA and Japan. During compression, the least significant bits of large amplitude
values are discarded.
Experiment 2.13 μ-Law Companding
The python code which performs μ-law companding is shown in Fig. 2.26, and the
corresponding output is shown in Figs. 2.27 and 2.28.
Inference
The input signal to be companded is a sawtooth signal. The fundamental frequency
of a sawtooth signal is 5 Hz. Figure 2.27 illustrates the signal to be encoded using
μ-law companding with μ = 255. Here the signal is basically compressed before
passing it to the uniform quantizer. Figure 2.28 shows the uniform quantizer results
for different bit-rate values. With increase in bit-rate, the quantized signal resembles
the input signal.
2.2 Non-uniform Quantization 57

Degree of Compression with mu=255

1.00 Input signal
0.75 Encoded

0.50
0.25
Amplitude

0.00
–0.25
–0.50
–0.75
–1.00
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 2.27 Encoded signal using μ-law companding

Fig. 2.28 Quantized signal

Experiment 2.14 Error Due to Quantization

Quantization is basically mapping a large set of values to a smaller set of values. It is
a non-linear and irreversible process. Quantization leads to loss of information. The
loss of information due to quantization can be considered as an error. The error
signal is considered as the difference between the quantized signal (y[n]) and the
input signal (x[n]). The objective of this experiment is to quantize the input
58 2 Sampling and Quantization of Signals

#Error due to quantization

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generate the input signal
t=np.linspace(0,1,100)
x=np.sin(2*np.pi*5*t)
#Step 2: Parameters of the quantizer
DR=np.max(x)-np.min(x) #Dynamic range
b=[1,2,4,8] #Number of bits
for i in range(len(b)):
L=2**b[i] #Quantization level
q=DR/(L) #Quantization step size
#Step 3: To obtain the quantized signal
y=np.sign(x)*q*(np.floor((abs(x)/q))+(1/2))
#Step 4: Obtain the error signal
e=y-x
#Plot the error signal
plt.subplot(2,2,i+1), plt.plot(e),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Error signal for b={}'.format(b[i]))
plt.tight_layout()

Fig. 2.29 Error due to quantization

sinusoidal signal of 5 Hz frequency for different bit rate. Then, plot the error signal
for different bit-rate. The python code, which performs this task, is shown in
Fig. 2.29, and the corresponding output is shown in Fig. 2.30.
Inferences
From Fig. 2.30, the following inferences can be made:
1. The error signal is oscillatory in nature. The magnitude of the error signal varies
between positive and negative values.
2. The magnitude of the error signal decreases with increase in bit-rate of the
quantizer.
3. Error due to quantization is inevitable; hence, quantization is considered as
irreversible phenomenon.
Experiment 2.15 Probability Density Function of Quantization Error
From the previous experiment, it is possible to confirm that error is inevitable in
quantization process. The objective of this experiment is to prove that quantization
error follows a uniform distribution. The steps followed in this experiment are
displayed in Fig. 2.31.
The python code which performs the task mentioned above is shown in Fig. 2.32,
and the corresponding output is shown in Fig. 2.33.
Inferences
The following inferences can be drawn from Fig. 2.33:
2.3 Signal Reconstruction 59

Error signal for b=1 Error signal for b=2

0.5
0.2
Amplitude

Amplitude
0.0 0.0

–0.2
–0.5
0 50 100 0 50 100
Time Time
Error signal for b=4 Error signal for b=8

0.05
0.0025
Amplitude

0.00 Amplitude 0.0000

–0.0025
–0.05

0 50 100 0 50 100
Time Time

Fig. 2.30 Error signal for different bit-rate of the quantizer

Fig. 2.31 Flow diagram of

Experiment 2.13
Generate the 5 Hz sine wave.

Quantize the sine wave with bit rate (b) = 4

Obtain the error between input and

quantized sine wave

Plot the histogram of the error signal

1. The quantization error follows uniform distribution in the range (-Δ/2, Δ/2),
where ‘Δ’ is the quantization step size.
2. In this example, the value of ‘Δ’ is 0.125; hence, Δ/2 value is 0.0625.

2.3 Signal Reconstruction

Signal reconstruction is an attempt to obtain the continuous-time signal from the

samples. This is also termed as interpolation. Different types of interpolation
schemes include (1) zero-order hold interpolation, (2) first-order hold or linear
interpolation and (3) ideal interpolation.
60 2 Sampling and Quantization of Signals

#PDF of quantized error signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generate the input signal
t=np.linspace(0,1,1000)
x=np.sin(2*np.pi*5*t)
#Step 2: Parameters of the quantizer
DR=np.max(x)-np.min(x) #Dynamic range
b=4
L=2**b #Quantization level
q=DR/(L) #Quantization step size
#Step 3: Quantize the input signal
y=np.sign(x)*q*(np.floor((abs(x)/q))+(1/2))
#Step 3: Obtain the error signal
e=y-x
#Step 4: Plot the histogram of the error signal
plt.hist(e,10),plt.xlabel('e'),plt.ylabel('$P_e(e)$')
plt.title('PDF of error signal')

Fig. 2.32 Python code for PDF of quantization error

PDF of error signal

120

100

80
P c (e)

60

40

20

0
–0.06 –0.04 –0.02 0.00 0.02 0.04 0.06
e

Fig. 2.33 Histogram plot quantization error

2.3.1 Zero-Order Hold Interpolation

A zero-order hold (ZoH) system is a form of simple interpolation, where a line of

zero-slope connects discrete samples. The zero-order hold maintains the signal level
of the previous pulse until the next pulse arrives. The reconstructed signal will
resemble a staircase curve. This is depicted in Fig. 2.34.
2.3 Signal Reconstruction 61

Fig. 2.34 Zero-order hold

interpolation
x(t)

Ts 2Ts 3Ts 4Ts 5Ts t

Fig. 2.35 Impulse response

h(t)
of ZoH interpolation
function 1

0 T t

#Zero-order hold interpolation

import numpy as np
import matplotlib.pyplot as plt
from scipy.interpolate import interp1d
#Step 1: Generation of sine wave
t=np.linspace(0,2*np.pi,10)
x=np.sin(t)
#Step 2: Performing zero-order hold interpolation
f=interp1d (t,x,kind='previous')
#Step 3: Plotting the results
t1=np.linspace(0,2*np.pi,500)
plt.plot(t1,f(t1),'k--'),plt.stem(t,x,'r'),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Sine wave')
plt.legend(['ZOH interpolation','Sine wave samples'],loc=1)

Fig. 2.36 Python code of zero-order hold interpolation

The impulse response of a zero-order hold is shown in Fig. 2.35.

The transfer function of zero-order hold function is given by

1 - e - Ts
H ðsÞ = ð2:10Þ
s

Experiment 2.16 Zero-Order Hold Interpolation

The python example, which performs zero-order hold interpolation of the sinusoidal
signal, is shown in Fig. 2.36, and the corresponding output is shown in Fig. 2.37. In
the scipy package, the built-in function ‘interp1d’ performs the zero-order hold
interpolation.
62 2 Sampling and Quantization of Signals

Sine wave
1.00 ZOH interpolation
0.75 Sine wave samples

0.50
0.25
Amplitude

0.00
–0.25
–0.50
–0.75
–1.00
0 1 2 3 4 5 6
Time

Fig. 2.37 Result of python code shown in Fig. 2.36

x(t) h(t)
1

Ts 2Ts 3Ts 4Ts 5Ts t 0 T 2T t

(a) First-order hold interpolation (b) Impulse response

Fig. 2.38 First-order hold interpolation. (a) First-order hold interpolation. (b) Impulse response

Inference
From Fig. 2.37, it is possible to interpret that zero-order hold interpolation converts
the input signal into a piece-wise constant signal. It is possible to observe disconti-
nuity in the zero-order hold interpolated signal.

2.3.2 First-Order Hold Interpolation

In first-order hold (FoH) interpolation, the signal samples are connected by a straight
line. This idea is illustrated in Fig. 2.38a.
The first-order hold performs linear interpolation between samples. The impulse
response of first-order hold is shown in Fig. 2.38b.
The transfer function of first-order hold is expressed as
2.3 Signal Reconstruction 63

#First-order hold interpolation

import numpy as np
import matplotlib.pyplot as plt
from scipy.interpolate import interp1d
#Step 1: Generation of sine wave
t=np.linspace(0,2*np.pi,10)
x=np.sin(t)
#Step 2: Performing zero-order hold interpolation
f=interp1d(t,x,kind='linear')
#Step 3: Plotting the results
t1=np.linspace(0,2*np.pi,10)
plt.plot(t1,f(t1),'k--'),plt.stem(t,x,'r'),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Sine wave')
plt.legend(['FOH interpolation','Sine wave samples'],loc=1)
plt.tight_layout()

Fig. 2.39 Python code to perform first-order hold interpolation

Sine wave
1.00 FOH interpolation
0.75 Sine wave samples

0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0 1 2 3 4 5 6
Time

Fig. 2.40 Result of python code shown in Fig. 2.39

2
1 - e - sT
H ðsÞ = ð2:11Þ
s

Experiment 2.17 First-Order Hold Interpolation

The python code to illustrate first-order hold interpolation is shown in Fig. 2.39, and
the corresponding output is shown in Fig. 2.40.
From Fig. 2.40, it is possible to interpret that first-order hold interpolation
attempts to connect the sample points through a straight line.
64 2 Sampling and Quantization of Signals

Inferences
The following inference can be drawn from Fig. 2.40:
1. The zero-order hold yields a staircase approximation of the signal.
2. The first-order hold yields a linear approximation of the signal.
3. The first-order hold connects the samples with straight lines.

2.3.3 Ideal or Sinc Interpolation

The expression for continuous-time signal obtained using sinc interpolation is

expressed as

1
t - nT s
xð t Þ = x½n] sin c ð2:12Þ
n= -1
Ts

The sinc function is a symmetric function which is square integrable. The decay
of the sinc function is slow. The sinc function has infinite support; hence, it is termed
as ideal interpolation. The sinc interpolation produces the smoothest possible inter-
polation of the samples.
Experiment 2.18 Ideal or Sinc Interpolation of a Sinusoidal Signal
The python code, which performs the ideal interpolation of the sine waveform, is
shown in Fig. 2.41, and the corresponding output is shown in Fig. 2.42.
Inference
The sinc interpolation produces the smoothest possible interpolation of the samples.
Experiment 2.19 Comparison of Zero-Order Hold and Sinc Interpolation
The python code, which performs the zero-order hold and sinc interpolation of a
given sinusoidal signal, is shown in Fig. 2.43, and the corresponding output is in
Fig. 2.44.
Inference
By observing Fig. 2.44, it is possible to infer that sinc interpolation smooths the
successive samples in the sine wave when compared to zero-order hold interpolation
method.
Exercises
1. Write a python code to demonstrate the phenomenon of aliasing in the frequency
domain for which the signal x(t) = sin (20πt) + sin (50πt) is generated using two
different sampling rates: fs = 100 Hz and fs = 25 Hz. Plot the corresponding
spectrum and comment on the observed result.
2.3 Signal Reconstruction 65

#Ideal or sinc interpolation

import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(0,2*np.pi,10)
t1=np.linspace(0,2*np.pi,100)
x=np.sin(t)
def sinc_interp(x, s, u):
if len(x) != len(s):
raise ValueError('x and s must be the same length')
T = s[1] - s[0]
sincM = np.tile(u, (len(s), 1)) - np.tile(s[:, np.newaxis], (1, len(u)))
y = np.dot(x, np.sinc(sincM/T))
return y
y=sinc_interp(x,t,t1)
plt.plot(t1,y,'r--'),plt.stem(t,x,'k'),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Sinc interpolation')
plt.legend(['Ideal interpolation','Sine wave samples'],loc=1)
plt.tight_layout()

Fig. 2.41 Python code to perform sinc interpolation

Sinc interpolation
1.00 Ideal interpolation
0.75 Sine wave samples

0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0 1 2 3 4 5 6
Time

Fig. 2.42 Result of python code shown in Fig. 2.41

2. Generate a sinusoidal signal of 5 Hz frequency. Quantize this signal using

uniform mid-rise quantizer with bit-rate, b = 1, 2 and 4. Use a subplot to plot
the input signal and the quantized signal.
3. Consider an analogue signal x(t) = cos (2πt) + cos (14πt) + cos (18πt), where ‘t’
is in seconds. If this signal is sampled at fs = 8 Hz, then it will be aliased with the
66 2 Sampling and Quantization of Signals

#Ideal and sinc interpolation

import numpy as np
import matplotlib.pyplot as plt
from scipy.interpolate import interp1d
t=np.linspace(0,2*np.pi,10)
t1=np.linspace(0,2*np.pi,500)
x=np.sin(t)
#Zero-order hold interpolation
f=interp1d(t,x,kind='previous')
#Sinc interpolation
def sinc_interp(x, s, u):
if len(x) != len(s):
raise ValueError('x and s must be the same length')
T = s[1] - s[0]
sincM = np.tile(u, (len(s), 1)) - np.tile(s[:, np.newaxis], (1, len(u)))
y = np.dot(x, np.sinc(sincM/T))
return y
y=sinc_interp(x,t,t1)
plt.plot(t1,f(t1),'b:'),plt.plot(t1,y,'k--'),plt.stem(t,x,'r'),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Comparison of interpolation methods')
plt.legend(['ZOH interpolation','Sinc interpolation','Sine wave samples',],loc=1)

Fig. 2.43 Comparison of zero-order hold and sinc interpolation

Comparison of interpolation methods

1.00 ZOH interpolation
0.75 Sinc interpolation
Sine wave samples
0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0 1 2 3 4 5 6
Time

Fig. 2.44 Result of ZOH and sinc interpolation

2.3 Signal Reconstruction 67

signal, which is expressed as x′(t) = 3 cos (2πt). Plot x(t) and x′(t) on the same
graph to verify the signals inter at the sampling instants.
4. Write a python code to generate a sinusoidal signal of fundamental frequency
1300 Hz and sampling frequency fs = 8 kHz. Hear this tone. Now downsample
this signal by a factor of 2 and hear the tone. Comment on the heard tones.
5. Write a python code to generate a sinusoidal signal of 10 Hz frequency. Quantize
this signal using 4-bit uniform mid-tread quantizer. Use a subplot to plot the
input, quantized and error signals. Comment on the observed output.
Objective Questions
1. What will be the output if the following code is executed?

A. 4.5
B. 4.0
C. 4.25
D. 5.0
2. What will be the output if the following code is executed?

A. -4.0
B. -5.0
C. -4.25
D. -5.25
3. The following python code segment produces

A. Zero-order hold interpolation

B. Linear interpolation
C. Polynomial interpolation
D. Sinc interpolation
4. Fourier transform of train of impulse function results in
A. Train of step function
B. Train of impulse
C. Sinc function
D. Triangular function
68 2 Sampling and Quantization of Signals

5. A sinusoidal signal of the form x(t) = sin(2πft), where ‘f = 5 Hz’ is sampled at the
rate fs = 100 Hz to obtain the discrete-time sequence x[n]. The expression for the
signal x[n] is

π
x½n] = sin n
2
π
x½n] = sin n
4
π
x½n] = sin n
5
π
x½n] = sin n
10

6. Assertion: Quantization is an irreversible process.

Reason: Quantization is many-to-one mapping:
A. Assertion and reason are true.
B. Assertion is wrong; reason is true.
C. Assertion is true; reason is wrong.
D. Assertion and reason are wrong.
7. Statement 1: Quantization is a non-linear phenomenon
Statement 2: Quantization is an irreversible phenomenon
A. Statement 1 and 2 are false.
B. Statement 1 and 2 are true.
C. Statement 1 is true; statement 2 is false.
D. Statement 1 is false; statement 2 is true.
8. The transfer function of zero-order hold is

H ðsÞ = 1
1
H ðsÞ =
s
1 - e - sT
H ðsÞ =
s
H ðsÞ = 1 - e - sT

9. An analogue voltage in the range 0–4 V is divided into 32 equal intervals. The
quantization step size of this uniform quantizer is
A. 0.0625
B. 0.125
C. 0.25
D. 0.5
2.3 Signal Reconstruction 69

10. If ‘Δ’ represents the quantization step size of a uniform quantizer, the expres-
sion for mean square quantization error is
Δ2
A.
2
Δ2
B.
4
Δ2
C.
8
Δ2
D.
12
11. The quantization error follows
A. Normal distribution
B. Uniform distribution
C. Chi-square distribution
D. Exponential distribution
12. The transfer function of first-order hold is
1 - e - sT
A. H ðsÞ =
s
1
B. H ðsÞ =
s
2
1 - e - sT
C. H ðsÞ =
s
- sT
D. H ðsÞ = 1 - e
13. The signal to be quantized takes the value in the range (-1,1). The dynamic
range of the signal is
A. 1
B. -1
C. 0
D. 2
14. If fs represents the sampling frequency, then the expression for Nyquist fre-
quency is
A. fs
B. fs/2
C. fs/4
D. fs/8
70 2 Sampling and Quantization of Signals

15. The quantization step size of a two-bit quantizer which accepts the input signal,
which varies from 0 to 2 V, is
A. 0.125
B. 0.25
C. 0.5
D. 0.75

Bibliography

1. Alan V. Oppenheim, and Ronald W. Schafer, “Discrete-Time Signal Processing”, Pearson, 2009.
2. Michael Roberts, and Govind Sharma, “Fundamentals of Signals and Systems”, McGraw Hill
Education, 2017.
3. John G. Proakis, and Dmitris G. Manolakis, “Digital Signal Processing: Principles, Algorithms
and Applications”, Pearson Education, 2007.
4. Barrie Jervis, Emmanuel Ifeachor, “Digital Signal Processing: A Practical Approach”,
Pearson, 2001.
5. Allen B. Downey, “Think DSP: Digital Signal Processing in Python”, O’ Reilly Media, 2016.
Chapter 3
Generation and Operation on Discrete-Time
Sequence

Learning Objectives
After completing this chapter, the reader is expected to
• Generate standard discrete-time sequences like unit sample, unit step, unit ramp
sequences, etc.
• Perform operations like folding, shifting and scaling on the discrete-time
sequence.
• Perform linear convolution and circular convolution between discrete-time
sequences.
• Perform autocorrelation and cross-correlation between discrete-time sequences.

Road Map of the Chapter

This chapter aims to generate different discrete-time signals or sequences and
perform various mathematical operations on the discrete-time signal. The flow of
the concept in this chapter is illustrated in the form of a block diagram, which is
given below:

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 71
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_3
72 3 Generation and Operation on Discrete-Time Sequence

Discrete-Time Sequence

Generation of DT Mathematical operation

signals on DT signals

Standard Arbitrary Folding Shifting Scaling

DT signals DT signals

Unit sample Sinusoidal

sequence (δ[n]) sequence
Unit step Signum Convolution Correlation
sequence (u[n]) sequence
Unit ramp Sinc Linear Convolution Autocorrelation
sequence (r[n]) sequence
Circular Convolution Cross-correlation
Exponential
sequence (x[n])

PreLab Questions
1. What are the steps involved in converting the continuous-time signal into a
discrete-time signal?
2. Mention different forms of representations of discrete-time signals?
3. Mention a few standard discrete-time sequences.
4. Mention the significant features of the unit sample sequence (δ[n]).
5. State the condition for the discrete-time signal to be periodic.
6. Distinguish between energy and power signal.
7. What are the various mathematical operations that can be performed on discrete-
time signals?
8. When a discrete-time signal is said to be (a) an even signal (b) an odd signal?
Give an example for each class of signal. Also, give an example of a signal
which is neither even nor odd.
9. Give an example of an energy and power signal. Also, give an example of a
discrete-time signal which is neither energy nor power signal.
10. Explain in your own word regarding the significance of convolution operation in
signal processing.
11. What is the relationship between convolution and correlation? Mention two
applications of correlation.
3.1 Generation of Discrete-Time Signals 73

3.1 Generation of Discrete-Time Signals

This section deals with the generation of different types of discrete-time signals like
unit sample signal, unit step signal, unit ramp signal, real and complex exponential
signals. The following section discusses about different mathematical operations that
could be performed on discrete-time signals.
Experiment 3.1 Generation of the Unit Sample Sequence
The mathematical expression of the unit sample sequence (δ[n]) is given by

1, if n = 0
δ ½ n] = ð3:1Þ
0, Otherwise

This experiment discusses the generation of unit sample sequence using ‘if’ and
‘else’ conditions in python platform. The python code to generate unit sample
sequence using ‘if’ and ‘else’ conditions is shown in Fig. 3.1, and the corresponding
output is shown in Fig. 3.2. The built-in functions used in the program are given in
Table 3.1.
Inference
It is possible to observe that unit sample sequence takes a value of ‘1’ at ‘n’ equal to
zero and zero at other instances of ‘n’.
Experiment 3.2 Generation of Unit Sample Sequence Using the Logical
Operation
This experiment deals with the logical operation used to generate unit sample
sequence, and the python code for this experiment is shown in Fig. 3.3, and the
corresponding output is shown in Fig. 3.4.
Inference
The statement (x = (n == 0)) given in Fig. 3.3 implies that the variable ‘x’ takes a
value of ‘1’ if n = 0, and it takes a value of ‘0’ for all the other values of ‘n’.

Fig. 3.1 Python code to

generate unit sample #Python code to generate unit sample sequence
sequence import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generating the sequence
n=np.arange(-10,11) #Define the x-axis
x=[1 if i==0 else 0 for i in n] #Unit sample sequence
#Step 2: Plotting the sequence
plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$\delta[n]$')
plt.xticks(n)
74 3 Generation and Operation on Discrete-Time Sequence

d [n]
1.0

0.8
Amplitude

0.6

0.4

0.2

0.0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.2 Unit sample sequence

Table 3.1 Built-in functions used in unit sample signal generation

S. No. Built-in function used Purpose
1 np.arange() To generate evenly spaced values within a given interval
2 plt.stem() To plot the discrete-time signal
3 plt.xticks() To get or set the current tick locations and labels of the x-axis

Fig. 3.3 Logical operation

to generate unit sample #Python code to generate unit sample sequence
sequence import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generating the sequence
n=np.arange(-10,11,1) #Define the x-axis
x=(n==0) #Unit sample sequence
#Step 2: Plotting the sequence
plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$\delta[n]$'),plt.xticks(n)

Experiment 3.3 Generation of Unit Sample Sequence Using the Built-In Func-
tion from the Scipy Library
The built-in function in scipy library ‘unit_impulse’ can be used to generate unit
sample sequence. The python code, which generates unit sample sequence using the
built-in function from the scipy library, is shown in Fig. 3.5, and the corresponding
output is shown in Fig. 3.6.
3.1 Generation of Discrete-Time Signals 75

d [n]
1.0

0.8

0.6
Amplitude

0.4

0.2

0.0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.4 Output of python code shown in Fig. 3.3

Fig. 3.5 Unit sample

sequence generation using import matplotlib.pyplot as plt
scipy library import numpy as np
from scipy import signal
n=np.arange(-5,6)
x=signal.unit_impulse(len(n), 'mid')
plt.stem(n, x),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$\delta[n]$'),plt.xticks(n)

Inference
From Figs. 3.5 and 3.6, it is possible to confirm that unit sample sequence can be
generated using the scipy library with the built-in command of ‘signal.unit_impulse’.
Experiment 3.4 Generation of Unit Step Sequence
The mathematical expression of the unit step sequence is written as

1, if n ≥ 0
u½n] = ð3:2Þ
0, Otherwise

In this experiment, the unit step sequence is generated using two methods. In the
first method, ‘if’ and ‘else’ conditions are used to generate unit step sequence. The
second method uses logical operation to generate unit step signal. The python code,
which generates unit step signal using two different methods, is shown in Fig. 3.7,
and the corresponding output is shown in Fig. 3.8.
76 3 Generation and Operation on Discrete-Time Sequence

d [n]
1.0

0.8

0.6
Amplitude

0.4

0.2

0.0

–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->

Fig. 3.6 Result of python code shown in Fig. 3.5

#Genertion of unit step signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generating the sequence
n=np.arange(-10,11,1) #Define the x-axis
#Method 1
x1=[1 if i>=0 else 0 for i in n] # if and else
#Method 2
x2=(n>=0) #Logical operation
#Plotting the result
plt.subplot(2,1,1),
plt.stem(n,x1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('u[n]'),plt.xticks(n)
plt.subplot(2,1,2),plt.stem(n,x2),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('u[n]'),plt.xticks(n)
plt.tight_layout()

Fig. 3.7 Python code to generate unit step signal

Inference
From Fig. 3.8, it is possible to interpret that both methods yield the same result,
which is a unit step signal. The unit step signal exhibits a sudden change in state from
logic 0 to logic 1 instantaneously.
Experiment 3.5 Generation of the Unit Ramp Signal
The mathematical expression of the unit ramp sequence (r[n]) is written as
3.1 Generation of Discrete-Time Signals 77

Fig. 3.8 Result of python code shown in Fig. 3.7

n, if n ≥ 0
r ½ n] = ð3:3Þ
0, Otherwise

The python code, which generates unit ramp signal using two methods, is
discussed in this experiment. In method 1, ‘if’ and ‘else’ conditions generate unit
ramp signals, whereas in method 2, logical operation is used to generate unit ramp
signals. The python code, which generates unit ramp signal using the two methods,
is shown in Fig. 3.9, and the corresponding output is shown in Fig. 3.10.
Inference
From Fig. 3.10, it is possible to observe that the ramp signal generated using ‘if’ and
‘else’ condition and ‘logical operation’ are alike. Unlike step signal, the ramp signal
gradually increases from low to high value.
Task
1. Write a python code to generate unit ramp signal from unit step signal.
Experiment 3.6
From unit sample signal generates unit step signal, and from unit step signal
generates unit ramp signal.
The relationship between unit sample (δ[n]) and unit step (u[n]) sequence is given
by
78 3 Generation and Operation on Discrete-Time Sequence

Fig. 3.9 Python code to

generate unit ramp signal #Generation of unit ramp signal
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generating the sequence
n=np.arange(-10,11,1) #Define the x-axis
#Two methods to generate unit ramp signal
x1=[i if i>=0 else 0 for i in n] #Unit ramp sequence
x2=n*(n>=0) #Logical operation
#Plotting the result
plt.subplot(2,1,1),plt.stem(n,x1),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('r[n]'),plt.xticks(n)
plt.subplot(2,1,2),plt.stem(n,x2),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('r[n]'),plt.xticks(n)
plt.tight_layout()

Fig. 3.10 Result of python code shown in Fig. 3.9

n
u½ n] = δ½k ] ð3:4Þ
k= -1

and
3.1 Generation of Discrete-Time Signals 79

Fig. 3.11 Flow chart

depicting the problem Generation of unit sample sequence (δ[n])
statement of Experiment 3.6

Derive unit step sequence (u[n]) from unit

sample sequence (δ[n])

Derive unit ramp sequence (r[n]) from unit

step sequence (u[n])

δ½n] = u½n] - u½n - 1] ð3:5Þ

The relationship between the unit ramp and unit step sequence is given by

r½n] = nu½n] ð3:6Þ

The flow chart, which depicts the objective of this experiment, is shown in
Fig. 3.11.
From the flow chart, the unit sample sequence is generated first. From unit sample
sequence, unit step sequence is obtained by repeated addition. From unit step
sequence, unit ramp sequence is derived. The python code, which performs the
above-mentioned task, is shown in Fig. 3.12, and the corresponding output is shown
in Fig. 3.13.
Inferences
From the python code shown in Fig. 3.12, it is possible to infer that unit step
sequence is obtained by repeatedly adding the unit sample sequence. The unit
ramp sequence is obtained by weighting the unit step signal by a factor of ‘n’.
From this example, it is possible to infer that any arbitrary signal x[n] can be obtained
from the unit sample sequence by scaling and shifting operations.
Task
1. Write a python code to generate a unit sample signal from the unit step signal.
Experiment 3.7 Generation of Real Exponential Sequence
The expression for a real exponential signal is given by

x½n] = αn ð3:7Þ

where α must be a real value. The aim of this experiment is to generate real
exponential sequence for four different values of ‘α’, namely, α = 0.5, α = -
0.5, α = 1.0 and α = - 1.0. The python code, which performs this task, is shown in
Fig. 3.14, and the corresponding output is shown in Fig. 3.15.
80 3 Generation and Operation on Discrete-Time Sequence

#Generation of test sequences from unit sample sequence

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of unit sample sequence
n=np.arange(-10,11)
x=[1 if i==0 else 0 for i in n] #delta[n]
#Step 2: Unit step sequence from unit sample sequence
y=np.zeros_like(n)
for k in range(len(x)):
y[k]=np.sum(x[:k+1])
#Step 3: Unit ramp sequence from unit step sequence
z=n*y
#Step 4: Plotting the result
plt.subplot(3,1,1),plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$\delta[n]$'),plt.xticks(n),plt.yticks(x),
plt.subplot(3,1,2),plt.stem(n,y),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('u[n]'),plt.xticks(n),plt.yticks(y),
plt.subplot(3,1,3),plt.stem(n,z),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('r[n]'),plt.xticks(n),
plt.tight_layout()

Fig. 3.12 Python code to generate test signals from unit sample sequence

Fig. 3.13 Result of python code shown in Fig. 3.12

3.1 Generation of Discrete-Time Signals 81

#Python code to generate real exponential sequences

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6,1) #Define the x-axis
alpha=[0.5, -0.5, 1.0, -1.0]
for i in range(len(alpha)):
x=alpha[i]**n #Real exponential sequence
plt.subplot(2,2,i+1)
plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('Amplitude')
plt.title(r'$\alpha$={}'.format(alpha[i]))
plt.xticks(n)
plt.tight_layout()

Fig. 3.14 Python code to generate real exponential signal

α=0.5 α=–0.5
30
Amplitude

Amplitude

20 0

10 –20
0
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->

α=1.0 α=–1.0
1.0 1
Amplitude

Amplitude

0.5 0

0.0 –1
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->

Fig. 3.15 Result of python code shown in Fig. 3.14

Inferences
The following inference can be made from this experiment:
1. If the value of α is 0 < α < 1, then the signal x[n] decreases in magnitude. This is
evident by observing the first subplot for α = 0.5.
2. If the value of α is -1 < α < 0, then the signal x[n] alternates in sign but
decreases in magnitude. This is evident by viewing the second subplot in
Fig. 3.15 for α = - 0.5.
3. For α = 1.0, there is no oscillation and the amplitude is always one.
82 3 Generation and Operation on Discrete-Time Sequence

#Python code to generate complex exponential sequences

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6,1) #Define the x-axis
omega_degree=[0, 90, 180, 270]
omega_radians=np.deg2rad(omega_degree)
for i in range(len(omega_radians)):
x=np.exp(1j*omega_radians[i]*n) #Complex exponential sequences
plt.subplot(2,2,i+1)
plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('Amplitude')
plt.title(r'$\omega={}^\circ$'.format(omega_degree[i]))
plt.xticks(n)
plt.tight_layout()

Fig. 3.16 Python code to generate complex exponential sequences

4. For α = - 1.0, the signal x[n] toggles. This is the highest frequency in digital
sequence.
Task
1. Obtain the real exponential sequence for α = 2 and comment on the nature of the
signal. Here the term ‘nature’ refers to whether the signal is a bounded or not.
Experiment 3.8 Generation of Complex Exponential Signal
The general form of complex exponential signal is given by

x½n] = ejωn ð3:8Þ

where ‘ω’ represents the angular frequency in radians. The python code to generate
complex exponential sequences for four different values of ‘ω’ such as ω =
0, π2 , π, 3π
2 is given in Fig. 3.16, and the corresponding output is shown in
Fig. 3.17.
Inferences
The following inference can be drawn from this experiment:
1. When ω = 0, the frequency is zero, the amplitude of the signal is constant and
there is no variation in the signal. This is termed as DC signal. For a DC signal,
the frequency is zero.
2. With increase in the value of ‘ω’, the oscillation exhibited by the signal increases.
At ω = π, the signal takes alternate values of +1 and -1. It is the highest
frequency in the digital signal.
Task
1. Write a python code to prove the fact that digital frequency ‘ω’ is unique in the
range 0 to 2π or from –π to π.
3.1 Generation of Discrete-Time Signals 83

w = 0° w = 90°
1.0 1
Amplitude

Amplitude
0.5 0

0.0 –1
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->

w = 180° w = 270°
1 1
Amplitude

Amplitude
0 0

–1 –1
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->

Fig. 3.17 Complex exponential sequences for different values of ‘ω’

#Generation of signum function

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6)
x=np.sign(n)
plt.stem(n,x)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.xticks(n)
plt.yticks(x),plt.title('Signum function')

Fig. 3.18 Python code to generate signum function

Experiment 3.9 Generation of Signum Function

Signum function is defined as a mathematical function that gives the sign of a real
number. The signum function f : R → R is defined as

1, if n > 0
sgn½n] = 0, if n = 0 ð3:9Þ
- 1, if n < 0

The python code to generate signum function is shown in Fig. 3.18, and the
corresponding output is shown in Fig. 3.19.
84 3 Generation and Operation on Discrete-Time Sequence

Signum function
1

Amplitude

–1
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->

Fig. 3.19 Result of python code shown in Fig. 3.18

Inference
From Fig. 3.19, it is possible to observe that the signum function takes only three
values, which are -1, 0 and 1; whenever n < 0, the signum function takes the value
of -1. At n = 0, the signum function takes a value of ‘0’. For the positive values of
‘n’, the signum function takes the value of +1.
Task
1. Is it possible to obtain signum function from unit step function? If yes, write a
python code to generate discrete signum signal from unit step signal.

3.2 Mathematical Operation on Discrete-Time Signals

This section discusses various mathematical operations that are performed on

discrete-time signals. The basic mathematical operations that could be performed
on the discrete-time signals are given in Fig. 3.20.

3.2.1 Amplitude Modification on DT Signal

The different signal operations that come under amplitude modification are
discussed in this section.
(a) Amplitude scaling
If x[n] is the input signal, the scaling of the signal x[n] by a factor of ‘A’ is
represented as
3.2 Mathematical Operation on Discrete-Time Signals 85

Mathematical operation on DT signal

Amplitude modification Time modification

Amplitude Amplitude Product of Addition of Time Time Time

scaling shifting two signals two signals scaling shifting reversal

Fig. 3.20 Different mathematical operations on DT signal

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generating the input signal
n=np.arange(-10,11,1)
x=(n==0)
#Obtaining the output signals for different values of 'A'
A=[2,0.5,1] #Three different values of factor 'A'
y1=A[0]*x
y2=A[1]*x
y3=A[2]*x
#Step 2: Plotting the result
plt.subplot(2,2,1),plt.stem(n,x),plt.yticks([0,2]),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('x[n]'),plt.subplot(2,2,2),plt.stem(n,y1)
plt.yticks([0,2]),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$y_1[n]$'),plt.subplot(2,2,3),plt.stem(n,y2)
plt.yticks([0,2]),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$y_2[n]$'),plt.subplot(2,2,4),plt.stem(n,y3),plt.yticks([0,2]),
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$y_3[n]$')
plt.tight_layout()

Fig. 3.21 Python code to perform amplitude scaling

y½n] = Ax½n] ð3:10Þ

If A > 1, the operation is called as amplification, A < 1 represents attenuation.

If A = 1, the output follows the input, it is called as input follower or buffer.

Experiment 3.10 Amplitude Scaling

Generate unit sample signal and perform the amplitude scaling for three different
values of A, namely: A = 2, A = 0.5 and A = 1. Plot the input and output signal and
comment on the observed output.
The python code, which performs the above-mentioned task, is shown in
Fig. 3.21, and the corresponding output is shown in Fig. 3.22.
86 3 Generation and Operation on Discrete-Time Sequence

x[n] Y1[n]
2 2
Amplitude

Amplitude
0 0
–10 –5 0 5 10 –10 –5 0 5 10
n--> n-->

Y2[n] Y3[n]
2 2
Amplitude

Amplitude

0 0
–10 –5 0 5 10 –10 –5 0 5 10
n--> n-->

Fig. 3.22 Result of python code shown in Fig. 3.21

Inference
The following inferences can be made from this experiment:
1. From Fig. 3.22, it is possible to observe that y1[n] is the amplified version of x[n],
and y2[n] is the attenuated version of x[n]. If the gain is unity, the output follows
the input, which is evident from the output y3[n].
2. This example illustrates the scaling of the amplitude axis for different values of
the factor ‘A’.
Task
1. Write a python code to illustrate the fact that amplitude scaling changes the
energy of the signal.

(b) Amplitude Shifting

If x[n] is the input signal, the amplitude shifting of the signal x[n] by a factor
of ‘C’ is represented as

y ½ n] = x ½ n] ± C ð3:11Þ

Experiment 3.11 Amplitude Shifting (DC Offset)

Let x[n] represent the discrete-time sinusoidal signal, and perform the DC offset of
this signal x[n] to obtain the signals y1[n] = x[n] + C and y2[n] = x[n] - C. The value
3.2 Mathematical Operation on Discrete-Time Signals 87

#DC offset
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of input sinusoidal sequence
t=np.linspace(0,1,100)
x=np.sin(2*np.pi*5*t)
#Step 2: Perform DC offset
offset=[5,-5]
y1=x+offset[0]
y2=x+offset[1]
#Step 3: Ploting the input and output signals
plt.subplot(3,1,1),plt.stem(t,x),
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('x[n]')
plt.subplot(3,1,2),plt.stem(t,y1)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$y_1$[n]')
plt.subplot(3,1,3),plt.stem(t,y2)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$y_2$[n]')
plt.tight_layout()

Fig. 3.23 Python code which performs DC offset

of ‘C’ for this experiment is to be chosen as 5. Write a python code to perform this
task and comment on the observed output.
The python code, which performs the above-mentioned task, is shown in
Fig. 3.23, and the corresponding output is shown in Fig. 3.24.
Inference
By observing Fig. 3.24, it is possible to infer that the reference for signal y1[n] is
+5 V, whereas the reference for signal y2[n] is -5 V. This is termed as DC offset.
Task
1. Does amplitude shifting affect the energy of the signal? Write a python code to
answer this question.

(c) Product of Two Signals

The product of two signals x1[n] and x2[n] is represented by

y½n] = x1 ½n] × x2 ½n] ð3:12Þ

The amplitude of the resultant signal y[n] gets modified. For example,
consider

x1 ½n] = sinð2πf 1 nÞ ð3:13Þ

x2 ½n] = cosð2πf 2 nÞ ð3:14Þ

Substituting Eqs. (3.13) and (3.14) in Eq. (3.12), we get

88 3 Generation and Operation on Discrete-Time Sequence

x[n]

Amplitude
1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
n-->
y1[n]
Amplitude

0
0.0 0.2 0.4 0.6 0.8 1.0
n-->
y 2[n]
Amplitude

–5
0.0 0.2 0.4 0.6 0.8 1.0
n-->

Fig. 3.24 Result of python code shown in Fig. 3.23

y½n] = sinð2πf 1 nÞ × cosð2πf 2 nÞ ð3:15Þ

Using the formula

1
sin A cos B = f sinðA þ BÞ þ sinðA - BÞg ð3:16Þ
2

Equation (3.15) can be written as

1
y ½ n] = f sin 2π ðf 1 þ f 2 Þn þ sin 2π ðf 1 - f 2 ÞnÞg ð3:17Þ
2

The amplitude of the output signal is different from the input signal x[n].

Experiment 3.12 Product of Two Signals

Obtain the product of the two signals given by x1[n] = sin (2πf1n) and
x2[n] = sin (2πf2n). In this example, consider f1 = f2 = 5 Hz. Using the relation
(3.17), the expression for the output signal is given by
y½n] = 12 f sin 2π ðf 1 þ f 2 Þn þ sin 2π ðf 1 - f 2 ÞnÞg. In this case, f1 = f2 = 5 Hz;
hence, the expression for the output signal is given by y½n] = 12 f sin 2π ð10ÞnÞg.
The frequency of the resultant signal should be 10 Hz, whereas its amplitude is
reduced by half. The python code, which performs this task, is shown in Fig. 3.25,
and the corresponding output is shown in Fig. 3.26.
Inference
The following inferences can be drawn from this experiment:
3.2 Mathematical Operation on Discrete-Time Signals 89

#Product of two signals

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of input signals
t=np.arange(0,100,1)
Fs=100
x=np.sin(2*np.pi*(5/Fs)*t)
y=np.cos(2*np.pi*(5/Fs)*t)
#Step 2: Product of the two signals
z=np.multiply(x,y)
#Step 3: Plotting the result
plt.subplot(3,1,1),plt.stem(t,x),plt.xlabel('n-->'),plt.ylabel('Ampltitude'),plt.title('$x_1[n]$')
plt.subplot(3,1,2),plt.stem(t,y),plt.xlabel('n-->'),plt.ylabel('Ampltitude'),plt.title('$x_2[n]$'),
plt.subplot(3,1,3),plt.stem(t,z),plt.yticks([-1,1]),plt.xlabel('n-->'),plt.ylabel('Ampltitude'),
plt.title('$y[n]$')
plt.tight_layout()

Fig. 3.25 Python code to obtain the product of the two signals

x1[n]
Amplitude

1
0
–1
0 20 40 60 80 100
n-->
x2[n]
Amplitude

1
0
–1
0 20 40 60 80 100
n-->
y [n]
Amplitude

–1
0 20 40 60 80 100
n-->

Fig. 3.26 Result of python code shown in Fig. 3.25

1. From Fig. 3.25, two signals of the same frequency are generated and multiplied.
2. From Fig. 3.26, it is possible to observe that x1[n] is a sine wave and x2[n] is a
cosine wave. The resultant signal y[n] is a sinusoidal signal with a frequency of
10 Hz, whereas the amplitude of the output waveform is reduced by a factor
of half.

(d) Signal Addition

90 3 Generation and Operation on Discrete-Time Sequence

#Addition of two signals

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of input signals
n=np.arange(-10,11,1)
x=(n>=0)
y=(n>=0)
#Step 2: Addition of the two signals
z=np.add(x.astype('float32'),y.astype('float32'))
#Step 3: Plotting the result
plt.subplot(3,1,1),plt.stem(n,x),plt.xticks(n)
plt.yticks([0,2]),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$x_1[n]$'),plt.subplot(3,1,2),plt.stem(n,y),plt.xticks(n)
plt.yticks([0,2]),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$x_2[n]$'),plt.subplot(3,1,3),plt.stem(n,z),plt.xticks(n)
plt.yticks([0,2]),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$y[n]$')
plt.tight_layout()

Fig. 3.27 Python code to perform addition of two signals

The signal addition results in a change in the amplitude of the signal. Two
signals x1[n] and x2[n] are added together to obtain the resultant output signal y
[n], which is given by

y½n] = x1 ½n] þ x2 ½n] ð3:18Þ

Experiment 3.13 Signal Addition

In this example, let x1[n] = u[n] and x2[n] = u[n]. The signal y[n] is the addition of
two unit step signals. The python code which performs this task is shown in
Fig. 3.27, and the corresponding output is shown in Fig. 3.28.
Inferences
The following inferences are drawn from these Figs. 3.27 and 3.28:
1. By observing Fig. 3.27, it is possible to observe that the result of logical operation
is converted to float using the command ‘.astype(‘float32’)’.
2. By observing Fig. 3.28, the inputs x1[n] and x2[n] are unit step signal, whose
amplitude takes value from 0 to 1, whereas the amplitude of the output signal y[n]
has variation from 0 to 2.
3. This experiment illustrates the fact that the amplitude of the signal can be changed
by signal addition operation.
3.2 Mathematical Operation on Discrete-Time Signals 91

x1[n]

Amplitude
2

0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->
x2[n]
Amplitude

0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->
y[n]
Amplitude

0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.28 Result of python code shown in Fig. 3.27

Task
1. Write a python code to illustrate the fact that ‘signal addition is a commutative
operation’.

3.2.1.1 Time Scaling Operation

Time scaling operations can be classified into two types, namely, (1) downsampling
and (2) upsampling.
(a) Downsampling
The downsampling of the signal x[n] by a factor of ‘M’ is represented as

y½n] = x½Mn] ð3:19Þ

where ‘M’ is an integer. Here ‘M - 1’ samples will be discarded between two

consecutive samples. Downsampling by a factor of ‘2’ is represented as

y½n] = x½2n] ð3:20Þ

Experiment 3.14 Downsampling

This experiment discusses the downsampling operation on the input signal. The
python code to perform downsampling by a factor of ‘2’ is shown in Fig. 3.29, and
the corresponding output is shown in Fig. 3.30.
92 3 Generation and Operation on Discrete-Time Sequence

#Downsampling by a factor of M
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generating the input signal
n=np.arange(-10,11,1)
x=n
M=2 #Downsampling factor
m=np.arange(n[0]/2,(n[-1]/2)+1,1)
#Step 2: Performing downsampling operation
y=x[::M]
#Step 3: Plotting the input and downsampled signal
plt.subplot(2,1,1),plt.stem(n,x),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('x[n]'),plt.xticks(n)
plt.subplot(2,1,2),plt.stem(m,y),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('y[n]'),plt.xticks(n)
plt.tight_layout()

Fig. 3.29 Python code to perform downsampling operation

x[n]
10
Amplitude

–10
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->
y[n]
10
Amplitude

–10
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.30 Result of downsampling operation

Inferences
The following inferences can be drawn from this experiment:
1. By observing Fig. 3.30, the number of samples in the input signal x[n] is
21, whereas the number of samples in the output signal y[n] is 11.
2. Downsampling leads to a reduction in the number of samples.
3.2 Mathematical Operation on Discrete-Time Signals 93

Task
1. Write a python code to prove the fact that downsampling is an irreversible
operation. That is, it is not possible to obtain the original signal from the
downsampled signal because downsampling results in loss of signal samples.

(b) Upsampling
The upsampling of the signal x[n] by a factor of ‘L’ is represented by

n
y ½ n] = x ð3:21Þ
L

The upsampling operation is basically inserting ‘L - 1’ zeros between two

consecutive samples. For L = 2, the above expression can be written as

n
y½n] = x
2

Experiment 3.15 Upsampling

This experiment deals with the upsampling process of discrete-time signal. The
python code, which performs the upsampling operation by a factor of 2, is shown
in Fig. 3.31, and the corresponding output is shown in Fig. 3.32.
Inference
The following observations can be made from this experiment:
By observing Fig. 3.32, it is possible to observe that in the case of upsampling by
a factor of 2, one zero is inserted between successive samples. Generally, when
upsampling by a factor of ‘L’, ‘L - 1’ zeros will be inserted between successive
samples. Also, it shows that the number of samples in the output increases to almost
L times than the number of samples in the input signal.
Task
1. Write a python code to illustrate the fact that ‘Upsampling is a reversible
operation’. It is possible to obtain the original signal from the upsampled signal.

3.2.1.2 Time Shifting Operation

The time shifting operation can be broadly classified into two types: (1) delay
operation and (2) advance operation.
(a) Delay operation
The delaying of the input signal by a factor of ‘k’ units is expressed as
94 3 Generation and Operation on Discrete-Time Sequence

#Upsampling by a factor of 2
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generating the input signal
L=2 #Upsampling factor
n=np.arange(-5,6,1)
N=len(n)
m=np.arange(-N+1,N+1,1)
x=np.ones(N)
#Step 2: Upsampling the input signal
y=np.zeros(L*N)
y[::2]=x
#Step 3: Plotting the input and output signal
plt.subplot(2,1,1),plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('x[n]'),
plt.xticks(m),plt.subplot(2,1,2),plt.stem(m,y),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('y[n]'),plt.xticks(m)
plt.tight_layout()

Fig. 3.31 Python code performs upsampling by a factor of 2

x[n]
1.0
Amplitude

0.5

0.0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10 11
n-->
y[n]
1.0
Amplitude

0.5

0.0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10 11
n-->

Fig. 3.32 Result of upsampling by a factor of 2

y ½ n] = x ½ n - k ] ð3:22Þ

where 'k' must be a positive integer.

(b) Advance operation
The advance of the input signal x[n] by a factor of ‘k’ units is expressed as
3.2 Mathematical Operation on Discrete-Time Signals 95

Fig. 3.33 Delay and

advance of unit step #Delay and advance operation
sequence import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-10,11)
x1=(n>=0) #u[n]
k=5
x2=(n>=k) #Delay of u[n]
x3=(n>=-k)#Advance of u[n]
plt.subplot(3,1,1),plt.stem(n,x1),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('u[n]')
plt.subplot(3,1,2),plt.stem(n,x2),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('u[n-5]')
plt.subplot(3,1,3),plt.stem(n,x3),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('u[n+5]')
plt.tight_layout()

y ½ n] = x ½ n þ k ] ð3:23Þ

where ‘k’ must be a positive integer.

Experiment 3.16 Time Shifting Operation

This experiment performs both delay and advance operations by a factor of 'k' units
on the unit step signal. First, the unit step signal is generated; then, it is delayed by a
factor of 5 units. The unit step signal is advanced by the factor of 5 units. The python
code, which performs this task, is shown in Fig. 3.33, and the corresponding output
is shown in Fig. 3.34.
Inference
This experiment illustrates the concept of shifting operation on the signal. Delay of
the signal u[n] by a factor of ‘5’ units results in u[n - 5], whereas advance of the
signal u[n] by a factor of 5 units results in u[n + 5]. It is to be observed that shifting
operation on the signal will not alter the energy of the signal.
Task
1. Write a python code to illustrate the fact that the signal energy is unaltered due to
signal shifting.

3.2.1.3 Time Reversal Operation

The time reversal of the signal x[n] is denoted as x[-n]. This refers to flipping the
signal x[n] from left to right and right to left. It can be considered as a signal
reflection about the origin. A discrete-time signal can be reversed in time by
96 3 Generation and Operation on Discrete-Time Sequence

u[n]
Amplitude 1

0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

u[n-5]
Amplitude

0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

u[n+5]
Amplitude

0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.34 Delay and advance of unit step sequence

changing the sign of the independent variable for all instances. Two different ways to
perform time reversal operation in python are given below.
Experiment 3.17 Time Reversal Without Built-In Function
This experiment deals with the time reversal operation using python without built-in
function. In this method, the signal x[n] is flipped from left to right using the
command" x[::-1]", the python code which performs the task of time reversal is
shown in Fig. 3.35, and the corresponding output is shown in Fig. 3.36.
Inference
Figure 3.36 clearly indicates that the left side of the input signal is moved into the
right side of the output signal and the right side of the input signal is moved into the
left side of the output signal.
Experiment 3.18 Time Reversal Using Built-In Function
This experiment tries to obtain the time reversal using a python built-in function. In
this method, the built-in function ‘np.fliplr()’ is used to perform a time reversal
operation. The python code, which performs this task, is shown in Fig. 3.37, and the
corresponding output is shown in Fig. 3.38.
Inference
This experiment confirms that the time reversal can be done using ‘np.fliplr’ built-in
function.
Task
1. Write a python code to illustrate that flipping operation does not alter the signal’s
energy.
3.3 Convolution 97

#Time reversal operation

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-10,11,1)
x=(n)
y=x[::-1]
plt.subplot(2,1,1),plt.stem(n,x)
plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('x[n]')
plt.subplot(2,1,2),plt.stem(n,y)
plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('y[n]')
plt.tight_layout()

Fig. 3.35 Method-1 to perform time reversal operation

x[n]
10
Amplitude

–10
–10–9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->
y[n]
10
Amplitude

–10
–10–9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.36 Result of python code shown in Fig. 3.35

3.3 Convolution

Convolution is an important operation in digital signal processing, because many

DSP algorithms use convolution operations in one form or other. The most common
application of convolution operation is filtering. It can be used for signal enhance-
ment. The relationship between the input and output of a linear time-invariant system
shown in Fig. 3.39.
The relationship between the input and output of the system is given by
98 3 Generation and Operation on Discrete-Time Sequence

#Time reversal operation

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-10,11,1)
x=(n)
y=np.fliplr([x])[0]
plt.subplot(2,1,1),plt.stem(n,x)
plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('x[n]')
plt.subplot(2,1,2),plt.stem(n,y)
plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('y[n]')
plt.tight_layout()

Fig. 3.37 Method-2 to perform time reversal operation

x[n]
10
Amplitude

–10
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

y[n]
10
Amplitude

–10
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.38 Result of python code shown in Fig. 3.37

y½n] = x½n] * h½n] ð3:24Þ

In the above expression, ‘*’ denotes the convolution operation. The above
expression can be written as

1
y½n] = x½k ]h½n - k ] ð3:25Þ
k= -1

Convolution obeys commutative property; hence, the above equation can be

expressed as
3.3 Convolution 99

x[n] h[n] y[n]

Fig. 3.39 Representation of the LTI system

1
y½n] = h½k ]x½n - k ] ð3:26Þ
k= -1

Experiment 3.19 Convolution of Given Signal with Unit Sample Signal

This experiment illustrates the fact that the convolution of any signal (x[n]) with unit
sample signal (δ[n]) will result in the same signal x[n]. This is expressed as

x½n] * δ½n] = x½n] ð3:27Þ

The python code, which illustrates the above concept, is shown in Fig. 3.40, and
the corresponding output is shown in Fig. 3.41.
Inferences
The following inferences can be drawn from this experiment:
1. From Fig. 3.41, the input signal (x[n]) generated is a triangular signal.
2. The impulse response (h[n]) is unit sample signal (δ[n]).
3. The signal x[n] is convolved with unit sample signal to obtain the output signal y
[n]. It can be observed that the output signal y[n] resembles the input signal x[n].
Experiment 3.20 Convolution of the Signal x[n] with Shifted Unit Sample Signal
This experiment illustrates the fact that the signal x[n] can be shifted by convolving it
with δ[n ± k]. Convolving the signal x[n] with δ[n - k] results in delaying the signal
x[n] by a factor of ‘k’. Convolving the signal x[n] with δ[n + k] results in advancing
the signal x[n] by a factor of ‘k’. This is expressed as

x½n] * δ½n ± k] = x½n ± k] ð3:28Þ

The python code, which performs this task, is shown in Fig. 3.42, and the
corresponding output is shown in Fig. 3.43.
Inferences
The task performed by the python program is summarized in Fig. 3.44.
1. The input signal x[n] is applied to two systems with impulse responses
h1[n] = δ[n - k] and h2[n] = δ[n + k] to obtain the output signals y1[n] and
y2[n] respectively.
2. By comparing the input signal x[n] with the output signal y1[n], it is possible to
observe that the output signal y1[n] is a shifted version (delayed version) of the
input signal x[n].
100 3 Generation and Operation on Discrete-Time Sequence

#Convolution with unit sample sequence

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6)
N=len(n)
n1=np.arange(-N+1,N)
#Step 1: Generation of triangular signal
x=5-np.abs(n)
#Step 2: Generation of unit sample signal
h=(n==0)
#Step 3: Perform the convolution
y=np.convolve(x,h,mode='full')
#Step 4: Displaying the result
plt.subplot(3,1,1),plt.stem(n,x),plt.xticks(n),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('x[n]'),plt.subplot(3,1,2),plt.stem(n,h),
plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
plt.subplot(3,1,3),plt.stem(n1,y),plt.xticks(n1),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('y[n]')
plt.tight_layout()

Fig. 3.40 Convolution of the signal x[n] with unit sample signal δ[n]

x[n]
Amplitude

0
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->
h[n]
Amplitude

0
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->
y[n]
Amplitude

0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 3.41 Result of python code shown in Fig. 3.40

3. By comparing the input signal x[n] with the output signal y2[n], it is possible to
observe that the output signal y2[n] is a shifted version (advanced version) of the
input signal x[n].
4. This experiment illustrates the fact that signal shifting can be accomplished using
convolution operation.
3.3 Convolution 101

#Convolution with shifted unit sample sequence

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6)
N=len(n)
n1=np.arange(-N+1,N)
#Step 1: Generation of triangular signal
x=5-np.abs(n)
#Step 2: Generation of shifted unit sample signals
k=5
h1=(n==k)
h2=(n==-k)
#Step 3: Perform the convolution
y1=np.convolve(x,h1,mode='full')
y2=np.convolve(x,h2,mode='full')
#Step 4: Displaying the result
plt.subplot(3,2,1),plt.stem(n,x),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('x[n]')
plt.subplot(3,2,2),plt.stem(n,x),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('x[n]')
plt.subplot(3,2,3),plt.stem(n,h1),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_1[n]$')
plt.subplot(3,2,4),plt.stem(n,h2),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_2[n]$')
plt.subplot(3,2,5),plt.stem(n1,y1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$y_1[n]$'),plt.subplot(3,2,6),plt.stem(n1,y2),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('$y_2[n]$')
plt.tight_layout()

Fig. 3.42 Python code to perform convolution of signal x[n] with shifted unit sample signal

Task
1. Repeat the above experiment with a rectangular pulse signal instead of a
triangular one.
Experiment 3.21 Commutative Property of Convolution
The motive of this experiment is to prove the commutative property of convolution.
The commutative property of convolution is expressed as

x½n] * h½n] = h½n] * x½n] ð3:29Þ

The python code to illustrate the commutative property of convolution is given in

Fig. 3.45, and the corresponding output is shown in Fig. 3.46.
Inferences
The following inferences can be drawn from Fig. 3.46:
102 3 Generation and Operation on Discrete-Time Sequence

x[n] x[n]

Amplitude
Amplitude
5 5

0 0
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->
h1[n] h2[n]

Amplitude
Amplitude

1 1

0 0
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->

y1[n] y2[n]

Amplitude
Amplitude

5 5

0 0
–10 –5 0 5 10 –10 –5 0 5 10
n--> n-->

Fig. 3.43 Result of python code shown in Fig. 3.42

Fig. 3.44 Task performed

y1[n]
by the python example h1[n]=δ[n – k]
x[n]

h2[n]=δ[n + k] y2[n]

1. The input signal x[n] is a pulse signal. Similarly, the signal h[n] is a pulse signal.
The signals x[n] and h[n] are the same.
2. The signal y1[n] is obtained by convolving x[n] with h[n], whereas the signal y2[n]
is obtained by convolving h[n] with x[n]. From Fig. 3.46, the signals y1[n] and
y2[n] are the same.
3. This experiment illustrates that convolution is commutative. Also, the convolu-
tion of two pulse signals results in a triangular signal.
Task
1. In the above experiment, let L1 and L2 be the length of the signals x[n] and h
[n]. Then, the length of the convolved signal is L1 + L2 - 1. Write a python code
to illustrate that linear convolution results in stretching the length of the signal.
Experiment 3.22 Associative Property of Convolution
The associative property of convolution is expressed as

ðx½n] * h1 ½n]Þ * h2 ½n] = x½n] * ðh1 ½n] * h2 ½n]Þ ð3:30Þ

To illustrate this property, the input signal x[n] chosen is x[n] = ejπn, which
toggles between +1 and -1. The impulse response h1[n] = δ[n - k] and the impulse
3.3 Convolution 103

#Commutative property of convolution

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6)
N=len(n)
n1=np.arange(-N+1,N)
#Step 1: Generating x[n]
x=np.array([0,0,0,0,1,1,1,0,0,0,0])
#Step 2: Generating h[n]
h=x
#Step 3: Obtaining the outputs
y1=np.convolve(x,h,mode='full')
y2=np.convolve(h,x,mode='full')
#Sep 4: Plotting the results
plt.subplot(2,2,1),plt.stem(n,x),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('x[n]')
plt.subplot(2,2,2),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('h[n]'),plt.subplot(2,2,3),plt.stem(n1,y1)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$y_1[n]$= x[n]*h[n]')
plt.subplot(2,2,4),plt.stem(n1,y2)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$y_2[n]$=h[n]*x[n]')
plt.tight_layout()

Fig. 3.45 Python code to illustrate the commutative property of convolution

x[n] h[n]
1.0 1.0
Amplitude
Amplitude

0.5 0.5

0.0 0.0
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5.0 –2.5 0.0 2.5 5.0
n--> n-->

y1[n]= x[n]*h[n] y2[n]=h[n]*x[n]

Amplitude

Amplitude

2 2

0 0
–10 –5 0 5 10 –10 –5 0 5 10
n--> n-->

Fig. 3.46 Result of python code shown in Fig. 3.45

104 3 Generation and Operation on Discrete-Time Sequence

#Associative property of convolution

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6)
#Step 1: Generation of triangular signal
x=np.exp(1j*np.pi*n)
#Step 2: Generation of shifted unit sample signals
k=5
h1=(n==k)
h2=(n==-k)
#Step 3: Perform the convolution
u1=np.convolve(x,h1,mode='full')
y1=np.convolve(u1,h2,mode='full')
v1=np.convolve(h1,h2,mode='full')
y2=np.convolve(x,v1,mode='full')
N=len(y1)
n1=np.arange(-N/2,N/2)
#Step 4: Displaying the result
plt.figure(1),plt.subplot(3,1,1),plt.stem(n,x),plt.xticks(n),plt.xlabel('n'),
plt.ylabel('Amplitude'),plt.title('x[n]'),plt.subplot(3,1,2),plt.stem(n,h1),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_1$[n]'),plt.subplot(3,1,3),
plt.stem(n,h2),plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_2$[n]')
plt.tight_layout()
plt.figure(2),plt.subplot(2,1,1),plt.stem(n1,y1),
plt.title('$y_1[n]$=(x[n]*$h_1$[n])*$h_2$[n]'), plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.subplot(2,1,2),plt.stem(n1,y1)
plt.title('$y_2[n]$=x[n]*($h_1$[n])*$h_2$[n])'),plt.xlabel('n-->'),plt.ylabel('Amplitude')
plt.tight_layout()

Fig. 3.47 Python code to illustrate associative property of convolution

response h2[n] = δ[n + k]. The python code, which illustrates the associative
property of the convolution operation, is given in Fig. 3.47, and the corresponding
outputs are shown in Figs. 3.48 and 3.49, respectively.
Inferences
The following are the inferences from this experiment:
1. The input signal x[n] = (-1)n, -5 ≤ n ≤ 5. The impulse response h1[n] = δ[n -
5] and h2[n] = δ[n + 5], which is shown in Fig. 3.48.
2. The output y1[n] = (x[n]*h1[n])*h2[n], whereas the output y2[n] = x[n]*(h1[n]
*h2[n]). From Fig. 3.49, it is possible to observe that the output y1[n] = y2[n],
which shows that associative property of convolution is verified.
Experiment 3.23 Distributive Property of Convolution
The distributive property of convolution is expressed as
3.3 Convolution 105

x[n]
Amplitude
1
0
–1
–5 –4 –3 –2 –1 0 1 2 3 4 5
n

h1[n]
Amplitude

0
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->

h2[n]
Amplitude

0
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->

Fig. 3.48 Input signal and impulse response

y1[n]=(x[n]*h1[n])*h2[n]
1
Amplitude

–1
–15 –10 –5 0 5 10 15
n-->
y2[n]=x[n]*(h1[n])*h2[n])
1
Amplitude

–1
–15 –10 –5 0 5 10 15
n-->

Fig. 3.49 Output signal

x½n] * ðh1 ½n] þ h2 ½n]Þ = x½n] * h1 ½n] þ x½n] * h2 ½n] ð3:31Þ

For illustration, the signal x[n] is chosen as x[n] = δ[n + 1] + 2δ[n] + δ[n - 1],
h1[n] = δ[n + 1] + δ[n] + δ[n - 1] and h2[n] = - δ[n + 1] - δ[n - 1] such that
h1[n] + h2[n] results in unit sample signal. The python code, which illustrates the
distributive of convolution, is shown in Fig. 3.50, and the corresponding outputs are
shown in Figs. 3.51 and 3.52, respectively.
106 3 Generation and Operation on Discrete-Time Sequence

Inferences
1. From Fig. 3.51, it is possible to observe that the input signal and the impulse
responses are all finite-duration signals. The input signal is expressed as x
[n] = δ[n + 1] + 2δ[n] + δ[n - 1]. The impulse responses are given by
h1[n] = δ[n + 1] + δ[n] + δ[n - 1] and h2[n] = - δ[n + 1] - δ[n - 1].
2. The sum of the impulse responses results in a unit sample signal, which is
expressed as h1[n] + h2[n] = δ[n]. Also, convolution of any input signal x[n]
with unit sample signal results in the same signal, which is expressed as x
[n] * δ[n] = x[n]. Because of this property, the output signal y1[n] is same as
the input signal x[n].
3. By observing the output signals y1[n] and y2[n], it is possible to infer
y1[n] = y2[n], which implies that the distributive property of convolution is
illustrated through this experiment.
Experiment 3.24 Convolution of a Square Wave with Lowpass Filter
Coefficient
In this experiment, a square wave of fundamental frequency 5 Hz is generated. It is
then passed through moving average filter with M = 5, 7, 9, and 11. The block
diagram of the experiment performed is shown in Fig. 3.53.
The impulse response of lowpass filter (moving average filter) is given by

M-1
1
h½n] = δ½n - k ] ð3:32Þ
M k=0

In this experiment, the value of ‘M’ is chosen as 5, 7, 9 and 11.

The expression for the output signal is given by

y½n] = x½n] * h½n]

The python code which accomplishes this task is shown in Fig. 3.54, and the
corresponding output is shown in Figs. 3.55 and 3.56.
Inferences
The following inferences can be drawn from Figs. 3.55 and 3.56:
1. The input to the system is a square wave of a fundamental frequency 5 Hz.
2. The system is passed through lowpass filter to obtain a triangular waveform.
3. By observing the input and output waveform, it is possible to observe that the
system converts drastic change (square waveform) to a gradual change (sawtooth
waveform). The system basically performs lowpass filtering of the input signal.
4. The extent of smoothing is governed by the value of ‘M’. Increasing the value of
‘M’ increases the extent of smoothing the input signal.
 3.3

#Distributive property of convolution

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6)
Convolution

#Step 1: Generation of input signal

x=np.array([0,0,0,0,1,2,1,0,0,0,0])
#Step 2: Generation of h1 and h2
k=5
h1=np.array([0,0,0,0,1,1,1,0,0,0,0])
h2=np.array([0,0,0,0,-1,0,-1,0,0,0,0])
#Step 3: Perform the convolution
h=h1+h2
y1=np.convolve(x,h,mode='full')
y2=np.convolve(x,h1,mode='full')+np.convolve(x,h2,mode='full')
N=len(y1)
n1=np.arange(-N/2,N/2)
#Step 4: Displaying the result
plt.figure(1),plt.subplot(3,1,1),plt.stem(n,x),plt.xticks(n),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('x[n]'),plt.subplot(3,1,2),plt.stem(n,h1),plt.xticks(n)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_1$[n]'),plt.subplot(3,1,3),
plt.stem(n,h2),plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_2$[n]')
plt.tight_layout()
plt.figure(2),plt.subplot(2,1,1),plt.stem(n1,y1),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('$y_1[n]$=(x[n]*$h_1$[n])*$h_2$[n]')
plt.subplot(2,1,2),plt.stem(n1,y1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$y_2[n]$=x[n]*($h_1$[n])*$h_2$[n])')
plt.tight_layout()

Fig. 3.50 Python code to illustrate distributive property of convolution

107
108 3 Generation and Operation on Discrete-Time Sequence

x[n]

Amplitude
2

0
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->
h1[n]
Amplitude

0
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->
h2[n]
Amplitude

–1
–5 –4 –3 –2 –1 0 1 2 3 4 5
n-->

Fig. 3.51 Plot of input signal and the impulse responses

y1[n]=(x[n]*h1[n])*h2[n]
2
Amplitude

0
–10.0 –7.5 –5.0 –2.5 0.0 2.5 5.0 7.5 10.0
n-->
y2[n]=x[n]*(h1[n])*h2[n])
2
Amplitude

0
–10.0 –7.5 –5.0 –2.5 0.0 2.5 5.0 7.5 10.0
n-->

Fig. 3.52 Plot of the output signals

Fig. 3.53 Block diagram of Input signal Output signal

problem statement Low pass filter
Square wave ?

Task
1. In the above experiment, replace the square wave input with sine wave with a
spike signal. That is a sine wave with an abrupt change in amplitude in a few time
instants. Now pass this sine wave through the moving average filter and comment
on the observed signal.
3.3 Convolution 109

#Low pass filtering of square wave

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of input signal
t=np.linspace(0,1,100)
x=signal.square(2*np.pi*5*t)
#Step 2: Generation of low pass filter coefficient
M=[5,7,9,11]
fig1=plt.figure(1)
plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Input signal')
for i in range(len(M)):
h=1/M[i]*np.ones(M[i])
#Step 3: Obtaining the output signal
y=np.convolve(x,h,mode='full')
fig2=plt.figure(2)
#Step 4: Plotting the results
plt.subplot(2,2,i+1),plt.plot(t,y[0:len(t)]), plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Output signal for M={}'.format(M[i]))
plt.tight_layout()

Fig. 3.54 Python code to perform lowpass filtering of square wave

Input signal
1.00
0.75
0.50
0.25
Amplitude

0.00
–0.25
–0.50
–0.75
–1.00
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 3.55 Input square waveform

Experiment 3.25 Convolution of a Square Wave with Highpass Filter

Coefficient
In this experiment, the square wave is passed through highpass filter whose impulse
response is h[n] = {1/2, -1/2}. The highpass filter is basically a change detector.
When a square wave is fed to highpass filter, the resultant waveform is a spike
110 3 Generation and Operation on Discrete-Time Sequence

Output signal for M=5 Output signal for M=7

1 1
Amplitude

Amplitude
0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time Time
Output signal for M=9 Output signal for M=11
1
0.5
Amplitude

Amplitude
0 0.0

–0.5
–1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time Time

Fig. 3.56 Lowpass filtered square waveform

waveform. The python code, which performs this task, is shown in Fig. 3.57, and the
corresponding output is shown in Fig. 3.58.
Inferences
The following inferences can be made from this experiment:
1. From Fig. 3.57, it is possible to infer that the input signal is a square wave, the
impulse response of highpass filter is h[n] = {1/2, -1/2}.
2. From Fig. 3.58, it is possible to observe that the output waveform is a spike
waveform. It is due to the fact that differentiation of a constant is zero. In a square
wave, major portion is constant in magnitude; hence, differentiation of a constant
is zero. Highpass filter is a change detector; hence, it gives spike waveform as the
output for the input square waveform.
Task
1. Generate sine wave of 5 Hz frequency. Add white noise, which follows normal
distribution to this sine wave. Now pass this noisy sine wave through highpass
filter. Plot the clean sine wave, noisy sine wave and highpass filtered signal. Write
a python code to answer the query ‘Does highpass filter tend to amplify the
noise?’
3.4 Correlation 111

#Square wave through high pass filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of input signal
t=np.linspace(0,1,100)
x=signal.square(2*np.pi*5*t)
#Step 2: Generation of high pass filter coefficient
h=np.array([0.5,-0.5])
#Step 3: To obtain the output signal
y=np.convolve(x,h,mode='full')
#Step 4: Plotting the input and output signal
plt.subplot(3,1,1),plt.plot(t,x),plt.xlabel('time'),plt.ylabel('Amplitude')
plt.title('Input signal'),plt.subplot(3,1,2),plt.stem(h),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('h[n]'),plt.subplot(3,1,3),
plt.plot(t,y[0:len(t)]),plt.xlabel('time'),plt.ylabel('Amplitude'),plt.title('Output signal')
plt.tight_layout()

Fig. 3.57 Python code to perform highpass filtering of square wave

Input signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
time
h[n]
Amplitude

0.5
0.0
–0.5
0.0 0.2 0.4 0.6 0.8 1.0
n-->
Output signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
time

Fig. 3.58 Spike waveform obtained by differentiating input square wave

3.4 Correlation

Correlation is a tool to find the relative similarity between two signals. Correlation
has two variants, namely: autocorrelation and cross-correlation. Autocorrelation
involves the correlation of a signal with itself. Cross-correlation is performed
when two different signals are correlated with one another.
The expression for autocorrelation of the sequence x[n] is given by
112 3 Generation and Operation on Discrete-Time Sequence

r xx ðlÞ = x½l] * x½ - l] ð3:33Þ

Equation (3.33) gives the relationship between correlation and convolution.

Convolving the folded version of the sequence x[n] with the signal x[n] results in
autocorrelation. Equation (3.33) can be expressed as

1
r xx ðlÞ = x½n]x½n þ l] ð3:34Þ
n= -1

Some of the properties of the autocorrelation function are summarized below:

1. Autocorrelation function is an even function. It is expressed as rxx(-l) = rxx(l).
2. Autocorrelation attains its maximum value at zero lag. It is expressed as rxx(0) ≥ |
rxx(k)| for all ‘k’.
The cross-correlation between two signals x[n] and y[n] is expressed as

r xy ðlÞ = x½l] * y½ - l] ð3:35Þ

The above equation can be expressed as

1
rxy ðlÞ = x½k]y½l þ k] ð3:36Þ
k= -1

Experiment 3.26 Autocorrelation and Cross-correlation of Sine

and Cosine Waves
In this experiment, two signals, namely, sine wave and cosine wave of frequency
5 Hz, are generated. Then, the autocorrelation between the sinewave and cosine
wave and the cross-correlation between sine and cosine wave is computed. The
results of autocorrelated and cross-correlated signals are plotted. The python code,
which performs the above-mentioned task, is shown in Fig. 3.59, and the
corresponding output is shown in Fig. 3.60.
Inferences
The following observation can be made from this experiment:
1. The autocorrelation between the sine waves is represented by rxx(l). The auto-
correlation result is observed to be even symmetric. The maximum value is
obtained at zero lag.
2. The autocorrelation between the cosine waves is represented by ryy(l ). The
autocorrelation is an even symmetric function with the maximum value obtained
at zero lag.
3. The cross correlation between sine and cosine waves is not even symmetric. Also,
it is possible to observe that rxy(l) is not equal to ryx(l).
3.4 Correlation 113

#Autcorrelation and cross-correlation

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of sine and cosine wave
t=np.linspace(0,1,100)
f=5
x=np.sin(2*np.pi*f*t)
y=np.cos(2*np.pi*f*t)
N=len(x)
#Step 2: Perform autocorreation and cross-correlation
rxx=np.correlate(x,x,mode='full')
ryy=np.correlate(y,y,mode='full')
rxy=np.correlate(x,y,mode='full')
ryx=np.correlate(y,x,mode='full')
lag = np.arange(-N+1,N)
#Step 3: Plot the results
plt.subplot(2,2,1),plt.plot(lag,rxx),plt.xlabel('Lag'),plt.ylabel('Autocorrelation')
plt.title('$r_{xx}(l)$'),plt.subplot(2,2,2),plt.plot(lag,ryy)
plt.xlabel('Lag'),plt.ylabel('Autocorrelation'),plt.title('$r_{yy}(l)$')
plt.subplot(2,2,3),plt.plot(lag,rxy),plt.xlabel('Lag'),plt.ylabel('Cross correlation')
plt.title('$r_{xy}(l)$'),plt.subplot(2,2,4),plt.plot(lag,ryx),plt.xlabel('Lag'),
plt.ylabel('Cross correlation'),plt.title('$r_{yx}(l)$')
plt.tight_layout()

Fig. 3.59 Autocorrelation and cross-correlation between signals

4. The autocorrelation and cross-correlation are used to find the relative similarity
between the two signals.
Tasks
1. Write a python code to illustrate the fact that maximum value of autocorrelation
occurs at zero lag.
2. Write a python code to illustrate the fact that correlation can be performed in
terms of convolution. That is convolution of a signal with its folded version
results in autocorrelation.
Experiment 3.27 Autocorrelation of Sine Wave to Itself and Noisy Signal
In this experiment, sine wave of 5 Hz is generated. It is stored as the variable ‘x’. The
sine wave is then corrupted by random noise, which follows normal distribution to
obtain the signal ‘y’. The autocorrelation of clean sine wave is obtained as rxx(l ), and
the cross-correlation between the clean and noisy sine wave is obtained as rxy(l ). The
python code, which performs this task, is shown in Fig. 3.61, and the corresponding
output is shown in Fig. 3.62.
114 3 Generation and Operation on Discrete-Time Sequence

rxx(I) ryy(I)
50 50
Autocorrelation

Autocorrelation
0 0

–50
–100 –50 0 50 100 –100 –50 0 50 100
Lag Lag

rxy(I) ryx(I)
50 50
Cross correlation

Cross correlation
0 0

–50 –50
–100 –50 0 50 100 –100 –50 0 50 100
Lag Lag

Fig. 3.60 Autocorrelation and cross-correlation results

Inferences
1. In Fig. 3.62, x(t) represents clean sine wave of 5 Hz frequency, and y(t) represents
noisy sine wave. The noisy sine wave is obtained by adding random noise to the
clean sine wave.
2. In Fig. 3.62, rxx(l) represents the autocorrelation of a clean sine wave. The
autocorrelation function exhibits even symmetry, with the maximum value
occurring at zero lag.
3. In Fig. 3.62, rxy(l ) represents the cross-correlation between clean and noisy sine
waves. The cross-correlation is not exhibiting even symmetry relation. Compar-
ing rxx(l ) and rxy(l), the maximum value is obtained in autocorrelation function.
Thus, the autocorrelation reveals the relative similarity between the signals.
Experiment 3.28 Delay Estimation Using Autocorrelation
In this experiment, unit step sequence (signal ×1) is generated, it is then shifted by a
factor of ‘5’ units to the right to obtain the signal ×2. The autocorrelation of the
signal ×1 to itself and the correlation between the signals ×1 and ×2 are used to
estimate the delay. The python code, which performs this function, is shown in
Fig. 3.63, and the corresponding output is shown in Fig. 3.64.
Inference
Upon displaying the result, the answer in the variable ‘td’ is ‘5’, which is a measure
of delay between the two signals x1[n] and x2[n]. Thus, autocorrelation can be used
to measure or estimate the delay between the two signals.
3.4 Correlation 115

#Autocorrelation and cross-correlation

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of sine and cosine wave
t=np.linspace(0,1,100)
f=5
x=np.sin(2*np.pi*f*t)
#Step 2: Generation of noisy signal
w=2.5*np.random.randn(len(t))
y=x+w
N=len(x)
#Step 2: Perform autocorreation and cross-correlation
rxx=np.correlate(x,x,mode='full')
rxy=np.correlate(x,y,mode='full')
lag = np.arange(-N+1,N)
#Step 3: Plot the results
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Sine wave (x(t))'),plt.subplot(2,2,2),plt.plot(t,y)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Noisy sine wave(y(t))')
plt.subplot(2,2,3),plt.plot(lag,rxx),plt.xlabel('Lag'),plt.ylabel('Autocorrelation')
plt.title('$r_{xx}(l)$'),plt.subplot(2,2,4),plt.plot(lag,rxy)
plt.xlabel('Lag'),plt.ylabel('Cross correlation'),plt.title('$r_{xy}(l)$')
plt.tight_layout()

Fig. 3.61 Python code to perform autocorrelation of clean and noisy sine wave

Exercises
1. Generate the following sequences (a) x1[n] = δ[n + 1] + δ[n - 1]
(b) x1[n] = δ[n + 1] - δ[n - 1] (c) x3[n] = δ[n] + 2δ[n - 1] + δ[n - 2] and
(d) x4[n] = δ[n] - δ[n - 1] + δ[n - 2], and plot it using a subplot, which consists
of two rows and two columns. The time index should vary from -5 to +5.
2. Write a python code to generate the finite length discrete-time signals
(a) x1[n] = u[n] - u[n - 5], (b) x2[n] = δ[n], (c) x3[n] = u[n + 5] - u[n - 5]
n, 0 ≤ n ≤ 5
and (d) x4 ½n] = in the interval -10 ≤ n ≤ 10. Use subplot to plot
0, otherwise
the generated signals.
π
3. Generate a complex exponential signal x½n] = ej4n , - 10 ≤ n ≤ 10. Perform the
following: (a) Extract the real and imaginary part of this signal. (b) Reconstruct
the signal x[n] from the real and imaginary parts using the relation x[n] = Re {x
[n]} + j Im {x[n]}.
π
4. Generate a complex exponential signal of the form x½n] = ej8n , - 10 ≤ n ≤ 10.
*
Obtain the signal y[n], which is expressed as y[n] = x[n] × x [n], and comment
on the nature of the signal y[n].
5. Write a python code to generate the following sequences:
116 3 Generation and Operation on Discrete-Time Sequence

Sine wave(x(t)) Noisy sine wave(y(t))

1
5

Amplitude
Amplitude

0 0

–5
–1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time Time

rxx(I) rxy(I)
50

Cross correlation
Autocorrelation

25

0 0

–25

–100 –50 0 50 100 –100 –50 0 50 100

Lag Lag

Fig. 3.62 Autocorrelation and cross-correlation of clean and noisy sine wave

#Delay estimation using correlation

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-20,21)
#Step 1: Generation of unit step signal
x1=(n>=0)
#Step 2: Delay signal by a factor of 5 units
x2=(n>=5)
N=len(x1)
lag=np.arange(-N+1,N)
#Step 3: Perform autocorrelation of signal x1
rxx=np.correlate(x1,x1,mode='full')
#Step 4: Perform the cross-correlation between x1 and x2
ryx=np.correlate(x2,x1,mode='full')
#Step 4: Estimate the delay
td=np.argmax(ryx)-np.argmax(rxx)
#Step 5: Plot the signal and its delayed version
print('Time delay={}'.format(td))
plt.subplot(2,1,1),plt.stem(n,x1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$x_1[n]$')
plt.subplot(2,1,2),plt.stem(n,x2),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$x_2[n]$')
plt.tight_layout()

Fig. 3.63 Python code to perform delay estimation

3.4 Correlation 117

x1[n]
1.0
Amplitude
0.5

0.0
–20 –15 –10 –5 0 5 10 15 20
n-->
x2[n]
1.0
Amplitude

0.5

0.0
–20 –15 –10 –5 0 5 10 15 20
n-->

Fig. 3.64 Signal and its delayed version

(a) x1[n] = δ[n] + δ[n - 1] + δ[n - 2] + δ[n - 3] + δ[n - 4] + δ[n - 5] + δ[n -

6] + δ[n - 7]
(b) x2[n] = δ[n] - δ[n - 1] + δ[n - 2] - δ[n - 3] + δ[n - 4] - δ[n - 5] + δ[n -
6] - δ[n - 7]
(c) x3[n] = δ[n] + δ[n - 1] - δ[n - 2] - δ[n - 3] + δ[n - 4] + δ[n - 5] - δ[n -
6] - δ[n - 7]
(d) x4[n] = δ[n] + δ[n - 1] + δ[n - 2] + δ[n - 3] - δ[n - 4] - δ[n - 5] - δ[n -
6] - δ[n - 7]
Compute the energy of these sequences and comment on the obtained
result.
6. Sketch the following signals in the range -5 ≤ n ≤ 5 (a) x1[n] = 2nδ[n - 2]
(b) x2[n] = n[δ[n + 2] + δ[n - 2]].
5 - jnj, jnj ≤ 5
7. Generate the signal x½n] = in the range -10 ≤ n ≤ 10. Extract
0, otherwise
the even and odd part of the signal. Try to reconstruct the signal from the even
and odd part and comment on the observed output.
8. Write a python code to demonstrate the following facts:
(a) Product to two even signals is an even signal.
(b) Product of two odd signals is an even signal.
(c) Product of an even and odd signal is odd signal.
9. Read a speech signal and perform the autocorrelation of the speech signal, and
observe whether the autocorrelation function is an even function.
10. Read a ‘male’ and ‘female’ voice. Perform the following
118 3 Generation and Operation on Discrete-Time Sequence

(a) Autocorrelation of the male voice (x)

(b) Autocorrelation of the female voice ( y)
(c) Cross-correlation between male and female voice
(d) Cross-correlation between female and male voice
Comment on the observed output.
Objective Questions
1. The python code segment shown below generates

A. Unit sample signal

B. Unit step signal
C. Unit ramp signal
D. Real exponential signal
2. The value of the signal ‘x’ shown in the following python code is high at n =?

A. -1
B. -2
C. 0
D. 2
3. If the variable ‘x’ contains the signal of interest, then the variable ‘y’ in the
following python code returns

A. Maximum value of the signal

B. Minimum value of the signal
C. Energy of the signal
D. Power of the signal
4. The signal generated in the variable ‘x’ after executing the following segment of
code is

A. x[n] = δ[n] + δ[n – 1] – δ[n + 1]

B. x[n] = δ[n + 1] + δ[n] + δ[n – 1]
C. x[n] = δ[n + 1] + δ[n] – δ[n – 1]
D. x[n] = δ[n + 1] + 2δ[n] + δ[n – 1]
3.4 Correlation 119

5. The signal generated in the variable ‘x’ after executing the following segment of
code is

A. Unit sample sequence

B. Unit step sequence
C. Unit ramp sequence
D. Real exponential sequence
6. What would be the energy of the signal ‘x’ which is stored in variable ‘E’ if the
following code segment is executed?

A. 1J
B. 2J
C. 3J
D. 4J
7. What operation is performed on the input signal ‘x’ if the following segment of
code is executed?

A. Convolution of signal ‘x’ with itself

B. Correlation of the signal ‘x’ with itself
C. Power spectral estimation of the signal ‘x’
D. Energy density estimation of signal ‘x’
8. A square wave is fed to a lowpass filter, the resulting signal is
A. Sine wave
B. Cosine wave
C. Triangular wave
D. Inverted square wave
9. The energy of the signal is unaltered by the following mathematical operation
A. Downsampling of the signal by a factor of ‘M’
B. Upsampling the signal by a factor of ‘L’
C. Amplitude scaling
D. Folding of the signal
120 3 Generation and Operation on Discrete-Time Sequence

10. The energy of the signal is unaltered by the following mathematical operation:
A. Downsampling of the signal by a factor of ‘M’
B. Upsampling the signal by a factor of ‘L’
C. Delaying or advancing the signal by a factor of ‘k’
D. Amplitude scaling of the signal
11. Upsampling by a factor of ‘L’ inserts
A. ‘L’ zeros between successive samples
B. ‘L – 1’ zeros between successive samples
C. ‘L + 1’ zeros between successive samples
D. ‘L + 2’ zeros between successive samples
12. If a discrete-time signal x[n] obeys the relation x[-n] = x[n], then the signal is
A. Odd signal
B. Even signal
C. Either even or odd signal
D. Neither even nor odd signal
13. Sum of elements of finite duration discrete-time odd signal is
A. Infinite
B. One
C. Zero
D. Always negative
14. The python code shown below generates the following signal in the variable ‘x’

A. u[n]
B. u[-n]
C. u[n + 5]
D. u[n – 5]
15. The product of two odd signal results in
A. Odd signal
B. Even signal
C. Either even or odd signal depending on the length of the signals
D. Neither even nor odd signal
16. Identify the statement which is FALSE
A. Autocorrelation is finding the relative similarity of the signal to itself.
B. Autocorrelation is an even function.
3.4 Correlation 121

C. Autocorrelation attains its maximum value at zero lag.

D. Auto correlation is an odd function.
17. What will be the fundamental period of the signal ‘x’ if the following python
code is executed?

A. 1
B. 2
C. 3
D. 4
18. Assertion: Highpass filter act as change detector
Reason: Highpass filter has the ability to detect the change in the input signal
A. Both assertion and reason are true.
B. Assertion is true, reason is false.
C. Assertion is false, reason may be true.
D. Both assertion and reason are false.
19. What will be the length of the signal ‘y’ if the following code segment is
executed?

A. 11
B. 21
C. 31
D. 41
20. What will be the impulse response (h[n]) if the following code segment is
executed?

A. h[n] = δ[n]
B. h[n] = δ[n - 1]
C. h[n] = u[n]
D. h[n] = u[n - 1]
122 3 Generation and Operation on Discrete-Time Sequence

21. Identify the statement that is WRONG with respect to ‘folding’ or ‘time
reversal’ operation
A. Folding operation does not alter the energy of the signal.
B. Folding increases the length of the signal.
C. If the folded version of the signal is equal to the signal itself, then the signal
is even signal.
D. If the folded version of the signal is equal to the signal itself, then the signal
is odd signal.
22. If x[n] is a unit step signal, then the following signal (y[n]) generated from x[n] is

A. Unit sample signal

B. Unit step signal
C. Unit ramp signal
D. Real exponential signal
23. The fundamental frequency of the signal generated by executing the following
code is

A. ω = π/2 rad/sample
B. ω = π rad/sample
C. ω = π/4 rad/sample
D. ω = π/8 rad/sample

Bibliography

1. Lonnie C. Ludeman, “Fundamentals of Digital Signal Processing”, John Wiley and Sons, 1986.
2. S. Esakkirajan, T. Veerakumar and Badri N Subudhi, “Digital Signal Processing”, McGraw
Hill, 2021.
3. Sophocles Orfanidis, “Introduction to Signal Processing”, Pearson, 1995.
4. Hwei P. Hsu, “Signals and Systems”, Schaum’s outline series, McGraw Hill Education, 2017.
5. Maurice Charbit, “Digital Signal Processing with Python Programming”, Wiley-ISTE, 2017.
Chapter 4
Discrete-Time Systems

Learning Objectives
After reading this chapter, the reader is expected to
• Obtain the impulse response and step response of the discrete-time system.
• Plot the magnitude and phase response of the discrete-time system.
• Plot the pole-zero plot of the discrete-time system.
• Verifying the linearity, time-invariance, causal and stable properties of the
discrete-time system.

Roadmap of the Chapter

This chapter begins with different types of representations of discrete-time system,
including difference equation, block diagram and state-space. Properties of a
discrete-time system which includes linearity, time-invariance, causal and stable
are verified with python illustration. In this chapter, discrete-time system responses,
including impulse response, step response and frequency response, are plotted, and
the obtained results are interpreted.

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 123
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_4
124 4 Discrete-Time Systems

Discrete-Time System

Representation Properties of Response of

of DT System DT System DT System

Difference equation Impulse response

Linear/Non-linear
Block diagram
Time-invariant/Time variant
Transfer function Step response
State-space Causal/Non-causal

Stable/Unstable Frequency response

Pole-zero plot
Impulse response Invertible/Non-invertible

PreLab Questions
1. List different forms of representation of discrete-time system.
2. When a discrete-time system is said to be a relaxed system?
3. When a discrete-time system is said to be linear? Give a few examples of linear
system.
4. Is it always true that the cascade of two non-linear systems will result in a non-
linear system? Justify your answer.
5. Why real-time (real-world) systems are considered as causal systems?
6. ‘All memoryless systems are causal, whereas all the causal systems are not
memoryless’. Justify this statement.
7. Is it always true that a cascade of two time-variant systems will result in a time-
variant system? Justify your answer.
8. When a discrete-time system is said to be invertible? Give an example of the
invertible and non-invertible systems.
9. Is it possible to test the causality and stability of a linear time-invariant discrete-
time system from its impulse response? If so, how?
10. Distinguish between static and dynamic discrete-time systems. Cite an example
for static and dynamic discrete-time systems.
11. When is a discrete-time system said to be non-recursive, and when is it said to be
recursive? Give examples for each class of the discrete-time system.
12. The pole-zero plot of a discrete-time system exhibits a zero at z = 1; what can
you infer about this system?
13. When is a discrete-time system invertible? Give an example of a discrete-time
system, which is invertible, and an example of a non-invertible discrete-time
system.
14. Mention two advantages of state-space representation of the discrete-time
system.
4.2 Representation of DT Systems 125

15. What do you understand by the statement ‘Discrete-time system is characterised

by its impulse response (h[n])’?

4.1 Discrete-Time System

Discrete-time (DT) system accepts a discrete-time signal as input and generates a

discrete-time signal as the output. The input to the discrete-time system is termed as
‘excitation’ and the output of the system is termed as ‘response’. The block diagram
of DT system is shown in Fig. 4.1.

4.2 Representation of DT Systems

Different forms of representation of DT systems include (1) block diagram, (2) dif-
ference equation, (3) transfer function, (4) pole-zero plot, (5) state-space, etc. Python
illustration with respect to different forms of representation of discrete-time system
and python examples to obtain different DT system responses, including impulse,
step, magnitude and phase responses, are discussed in this chapter.

4.2.1 Difference Equation Representation of Discrete-Time

Linear Time-Invariant System

The relationship between the input and output of a discrete-time linear time-invariant
(LTI) system is expressed in terms of linear constant coefficient difference equation
(LCCDE) as

N -1 M -1
ak y½ n - k ] = bk x ½ n - k ] ð4:1Þ
k=0 k=0

where {ak} and {bk} are the output and input coefficients respectively. The above
equation represents the fact that weighted sum of input is equal to the weighted sum
of output. Equation (4.1) can be expanded as

Fig. 4.1 Block diagram of

discrete-time system Input DT System Output
x[n] y[n]
126 4 Discrete-Time Systems

a0 y ½ n] þ a 1 y ½ n - 1] þ ⋯ þ a N y ½ n - N ] = b 0 x ½ n] þ b 1 x ½ n - 1] þ ⋯
þ bM x½n - M ] ð4:2Þ

If a0 = 1, the above expression can be written as

y½n] þ a1 y½n - 1] þ ⋯ þ aN y½n - N ] = b0 x½n] þ b1 x½n - 1] þ ⋯

þ b M x½ n - M ] ð4:3Þ

Equation (4.3) can be expressed as

y½n] = b0 x½n] þ b1 x½n - 1] þ ⋯ þ bM x½n - M ]

- f a1 y ½ n - 1] þ ⋯ þ a N y ½ n - N ] g ð4:4Þ

If the current output is not a function of the previous output of the system, the
system is said to be a non-recursive system. The input-output relationship of a
non-recursive system is given by

y½n] = b0 x½n] þ b1 x½n - 1] þ ⋯ þ bM x½n - M ] ð4:5Þ

An example of non-recursive system is a finite impulse response (FIR) filter. In

the case of infinite impulse response (IIR) filter, the current output is a function of the
current input, previous input and previous output.
Experiment 4.1 Solution of Difference Equation with Zero Initial Condition
This experiment discusses solving the difference equation with zero initial condi-
tions. The relationship between the input and output of the discrete-time system is
given by

1
y ½ n] = y½n - 1] þ x½n] ð4:6Þ
2

Let us consider the input to the system as a unit step signal; hence, the above
equation can be written as

1
y ½ n] = y½n - 1] þ u½n] ð4:7Þ
2

Substituting n = 0 in the expression (4.7), we get

1
y½ 0] = y½ - 1] þ u½ 0 ]
2

If the initial condition is zero, y[-1] = 0 and u[0] = 1; hence, the above equation
can be written as
4.2 Representation of DT Systems 127

y½0] = 1

Substituting n = 1 in Eq. (4.7), we get

1
y½1] = y½ 0] þ u½ 1]
2

Substituting the value of y[0] as 1 and u[1] = 1 in the above equation, we get

1 3
y½1] = ×1 þ 1=
2 2

Substituting n = 2 in Eq. (4.7), we get

1 3
y½ 2] = × þ1
2 2

Simplifying the above equation, we get

5
y½ 2] =
4

The output of the system is given by

3 5
y½n] = 1, , ,⋯
2 4

The python code, which obtains the output of the system, is given in Fig. 4.2, and
the corresponding output is shown in Fig. 4.3.

import numpy as np
from scipy import signal
import matplotlib.pyplot as plt
n=np.arange(0,4)
x=np.ones(len(n)) #Input
#Defining the system
num=[1]
den=[1,-1/2]
#Obtaining the output
y=signal.lfilter(num,den,x)
#Displaying the result
plt.stem(y),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('y[n]')
plt.tight_layout()

Fig. 4.2 Python code to obtain the output of discrete-time system

128 4 Discrete-Time Systems

y[n]

1.75
1.50
1.25
Amplitude

1.00
0.75
0.50
0.25
0.00
0.0 0.5 1.0 1.5 2.0 2.5 3.0
n-->

Fig. 4.3 Simulation result

Inference
The output of the system obtained using python example is in agreement with the
theoretical result.
Task
1. Write a python code to obtain the impulse response of a discrete-time system,
whose input and output are related as y[n] = x[n] + 2y[n - 1]. Also, try to answer
the question “whether the impulse response is absolutely summable or not”.
Experiment 4.2 Solution of Difference Equation with Initial Condition
This experiment deals with solving the difference equation with the initial condition.
The relationship between the input and output of a discrete-time system is given by

1
y½n] = y½n - 1] þ x½n] ð4:8Þ
2
n
If x½n] = 14 u½n] and y[-1] = 1.
n
Substituting x½n] = 14 u½n] in Eq. (4.8), we get

1 1 n
y½n] = y ½ n - 1] þ u½ n] ð4:9Þ
2 4

Substituting n = 0 in Eq. (4.9), we get

1 1 0
y½0] = y½ - 1] þ u½ 0]
2 4

1 0
Substituting y[-1] = 1 and 4 u½0] = 1 in the above equation, we get
4.2 Representation of DT Systems 129

import numpy as np
from scipy import signal
n=np.arange(0,4)
x=(1/4)**n
num=[1]
den=[1,-1/2]
y1=signal.lfiltic(num,den,y=[1.])
y=signal.lfilter(num,den,x,zi=y1)
print(y[0]) [1.5 1. 0.5625 0.296875]

(a) Python Code (b) Result

Fig. 4.4 Python code to obtain the solution of difference equation. (a) Python Code (b) Result

1 3
y½0] = ×1 þ 1=
2 2

Substituting n = 1 in Eq. (4.9), we get

1 1 1
y½1] = y½ 0] þ u½ 1]
2 4

Substituting y[0] = 3/2 and u[1] = 1 in the above equation, we get

1 3 1
y½ 1] = × þ =1
2 2 4

Substituting n = 2 in Eq. (4.9), we get

1 1 2
y½2] = y½ 1] þ u½ 2]
2 4

Substituting y[1] = 1 in the above expression, we get

1 1 9
y½2] = ×1 þ = = 0:5625:
2 16 16

The python code, which obtains the solution of the difference equation with a
non-zero initial condition, is given in Fig. 4.4a, and the corresponding output is
shown in Fig. 4.4b.
Inference
The built-in function ‘lfiltic’ and ‘lfilter’ is used to obtain the solution of difference
equation. The experimental result is in agreement with the theoretical result.
130 4 Discrete-Time Systems

4.2.2 State-Space Model of a Discrete-Time System

The state-space model describes the system’s dynamics through two equations,
namely: ‘state equation’ and ‘output equation’. The state equation describes how
the input influences the state, and the output equation describes how the state and the
input directly influence the output. It is to be noted that the state-space representa-
tions of a particular system’s dynamics are not unique. The two equations are given
below

x½k þ 1] = Ax½k] þ Bu½k] ð4:10Þ

y½k] = Cx½k ] þ Du½k] ð4:11Þ

where u[k] 2 ℜm is the input, y[k] 2 ℜ p is the output and x[k] 2 ℜn is the state
vector. 'A' is the system matrix, 'B' and 'C' are the input and output matrices, and 'D' is
the feed forward matrix.

4.2.2.1 State-Space to Transfer Function

Taking Z-transform on both sides of Eq. (4.10), we get

zXðzÞ - zx0 = AXðzÞ þ BUðzÞ

Assuming an zero initial condition (x0), the above equation can be expressed as

zX ðzÞ = AX ðzÞ þ BU ðzÞ

The above equation can be expressed as

ðzI - AÞX ðzÞ = BU ðzÞ

The expression for X(z) is given by

X ðzÞ = BðzI - AÞ - 1 U ðzÞ ð4:12Þ

Taking Z-transform on both sides of the Eq. (4.11), we get

Y ðzÞ = CX ðzÞ þ DU ðzÞ ð4:13Þ

Substituting the expression for X(z) from Eq. (4.12) in Eq. (4.13), we get
4.2 Representation of DT Systems 131

Y ðzÞ = CBðzI - AÞ - 1 U ðzÞ þ DU ðzÞ

The expression for the transfer function is given by

Y ðzÞ
H ðzÞ = = CðzI - AÞ - 1 B þ D ð4:14Þ
U ðzÞ

Thus, the transfer function of the system is represented in terms of state-space

model.
Experiment 4.3 State-Space to Transfer Function
The state-space representation of discrete-time system is given by

x½k þ 1] = Ax½k] þ Bu½k]

and

y½k ] = Cx½k] þ Du½k],

-1 2 1
where A = , B= , C = [-1 2] and D = [1]. Obtain the transfer
1 0 0
function of the system using python.
The relationship between the transfer function and state-space representation is
given by

H ðzÞ = CðzI - AÞ - 1 B þ D

Step 1: To determine (zI - A)-1

z 0 -1 2
zI - A = -
0 z 1 0

Upon simplifying the above equation, we get

zþ1 -2
zI - A =
-1 z
z 2
z 2 þ z-2 z2 þ z - 2
ðzI - AÞ - 1 =
1 zþ1
z2 þ z - 2 z2 þ z - 2

Step 2: To determine H(z)

The expression for the transfer function H(z) is given by
132 4 Discrete-Time Systems

Fig. 4.5 Python code to

obtain the transfer function #State-space to transfer function
from state-space model from scipy import signal
import numpy as np
#Step 1: Defining the state-space model
A=[[-1,2],[1,0]]
B=[[1], [0]]
C = [[-1, 2]]
D=1
#Step 2: Obtaining the transfer function
[num,den]=signal.ss2tf(A,B,C,D)
print('numerator=',num)
print('denominaor=',den)

Fig. 4.6 Transfer function

of the discrete-time system numerator= [[1 0 0]]
denominaor= [ 1. 1. -2.]

H ðzÞ = CðzI - AÞ - 1 B þ D
z 2
z2 þ z - 2 z2 þ z - 2 1
H ðzÞ = ½ - 1 2] þ ½ 1]
1 zþ1 0
z2 þ z - 2 z2 þ z - 2

Simplifying the above expression, we get

-z þ 2 2z 1
H ðzÞ = þ ½ 1]
z2 þ z - 2 z2 þ z - 2 0

Upon simplifying the above equation, the transfer function of the system is
given by

-z þ 2
H ðzÞ = þ1
z2 þ z - 2

The transfer function of the system is given by

z2
H ðzÞ =
z2 þ z - 2

The built-in function ‘ss2tf’ available in scipy library can be used to obtain the
transfer function of the system from the state-space representation. The python
code, which performs this task, is shown in Fig. 4.5, and the corresponding output
is shown in Fig. 4.6.
4.2 Representation of DT Systems 133

Fig. 4.7 State-space model

from the transfer function #Transfer function to state-space
import numpy as np
from scipy import signal
#Step 1: Defining the transfer function
num=[1, 0, 0]
den=[1,1,-2]
#Step 2: Obtaining the state-space model
A,B,C,D=signal.tf2ss(num,den)
print("A=",A,"\n","B=",B,"\n","C=",C,"\n","D=",D)

Fig. 4.8 Result of python

code shown in Fig. 4.7 A= [[-1. 2.]
[ 1. 0.]]
B= [[1.]
[0.]]
C= [[-1. 2.]]
D= [[1.]]

Upon executing the code shown in Fig. 4.5, the transfer function of the system
obtained is given in Fig. 4.6.
Inference
The transfer function obtained using the built-in function ‘ss2tf’ is in agreement with
the theoretical result.
Experiment 4.4 Transfer Function to State-Space
The objective of this experiment is to obtain the state-space representation of the
2
discrete-time system, whose transfer function is given by H ðzÞ = z2 þzz - 2. As per the
previous experiment, the value of the state-space model parameters should be
-1 2 1
A= , B= , C = ½ - 1 2 ] and D = [1]. The built-in function
1 0 0
‘tf2ss’ in scipy library can be used to obtain the state-space representation of
discrete-time system from the transfer function. The python code, which performs
this task, is shown in Fig. 4.7, and the corresponding output is shown in Fig. 4.8.
Inference
From Fig. 4.8, it is possible to observe that the state-space model parameters are
-1 2 1
A= , B= , C = ½ - 1 2 ] and D = [1].
1 0 0
Task
1. What will be the value of the state-space parameter ‘c’ if the numerator and the
denominator polynomial of the transfer function are same? For example,
2
þzþ1
H ðzÞ = zz2 þzþ1.
134 4 Discrete-Time Systems

4.2.3 Impulse Response and Step Response of Discrete-Time

System

Impulse response is the reaction of the discrete-time system to unit sample input
signal, whereas step response is the reaction of the system to unit step input signal.
The discrete-time system is completely characterised by its impulse response. The
meaning is, if one knows the impulse response of the system, it is possible to infer
whether the system is causal and stable from it.
Figure 4.9 depicts the input-output relationship of a discrete-time system in which
x[n] represents the input signal, h[n] represents the impulse response and y[n]
represents the output of the system. If the system is a LTI system, then the output
of the system is expressed as

y½n] = x½n]* h½n] ð4:15Þ

In the above expression, ‘*’ indicates the convolution operation. If the input to the
system is unit sample signal, then the output of the system is given by

y½n] = δ½n]* h½n] ð4:16Þ

Convolution of any signal with unit sample signal results in the same signal;
hence, the above equation can be expressed as

y½n] = h½n] ð4:17Þ

Thus, the impulse response of the system is the reaction of an LTI system to unit
sample input signal.
The reaction of LTI system to unit step input signal is termed as step response of
the system. It is denoted as s[n]. The relationship between step response (s[n]) and
impulse response (h[n]) is given by

n
s½n] = h½k ] ð4:18Þ
k= -1

Experiment 4.5 Impulse and Step Responses of the System

This experiment deals with the computation of impulse and step responses from
the LCCDE. An LTI discrete-time system is defined by the difference equation
y[n] = x[n] + y[n - 1]. Plot the impulse response and step response of the
system.

Fig. 4.9 Input-output of a

x[n] h[n] y[n]
discrete-time system
4.2 Representation of DT Systems 135

#Impulse and step response of LTI DT system

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of unit sample and unit step signal
x1=np.zeros(100)
x1[0]=1
x2=np.ones(100)
#Step 2: Define the system
num=[1]
den=[1,-1]
#Step 3: Obtaining the impulse and step response
h=signal.lfilter(num,den,x1)
s=signal.lfilter(num,den,x2)
#Plotting the result
plt.subplot(2,1,1),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]')
plt.title('Impulse response (h[n])')
plt.subplot(2,1,2),plt.stem(s),plt.xlabel('n-->'),plt.ylabel('s[n]')
plt.title('Step response (s[n])')
plt.tight_layout()

Fig. 4.10 Python code to obtain the impulse and step response of the system

Impulse response is the response of the system to unit sample input signal, and
step response is the response of the system to unit step input signal. The python code,
which computes the impulse and step response of the system, is shown in Fig. 4.10,
and the corresponding output is shown in Fig. 4.11.
Inferences
The following inferences can be made from this experiment:
1. From Fig. 4.11, it is possible to observe that the impulse response of the system is
unit step signal. This implies h[n] = u[n]. The impulse response of the system is
not absolutely summable; hence, this system is not stable system.
2. The step response of the system is a ramp signal. This implies s[n] = nu[n].
Task
1. In the above experiment if h[n] = δ[n], what will be the step response of the
system?
Experiment 4.6 Computation of Impulse and Step Responses of the System
from the Difference Equation
This experiment also discusses the computation of the impulse and step response
from the difference equation. Let us consider the discrete-time LTI system, whose
difference equation is given by y[n] - 0.5y[n - 1] = x[n] + x[n - 1]. The python
code, which obtains the impulse and step responses of the given discrete-time
system, is given in Fig. 4.12, and the corresponding output is shown in Fig. 4.13.
136 4 Discrete-Time Systems

Impulse response (h[n])

1.0
h[n]

0.5

0.0
0 20 40 60 80 100
n-->
Step response (s[n])
100
s[n]

50

0
0 20 40 60 80 100
n-->

Fig. 4.11 Impulse response and step response of the given system

Inferences
The following inferences can be drawn from this experiment:
1. The impulse response decays to zero. The impulse response is absolutely sum-
mable; hence, the given system is a stable system.
2. The step response settles to a finite value after a short span.
Task
1. In the above experiment, from the impulse response, will it be possible to
comment on the stability of the system? Write a python code to find whether
the impulse response is absolutely summable or not?
Experiment 4.7 Computation of Step Response from the Impulse Response
This experiment discusses the computation of step response from the impulse
response. The impulse response of the system is unit step signal. The step response
is obtained by repeatedly adding the impulse response. The python command ‘np.
cumsum’ can be used to obtain the step response of the system from impulse
response of the system. The relationship between input and output of a linear
time-invariant system is given by

y½n] = x½n]* h½n] ð4:19Þ

It is given that the impulse response is unit step signal; hence, h[n] = u[n]. Also,
unit step response of the system implies that the input to the system is unit step
signal; hence, x[n] = u[n]. The output of the system is given by
4.2 Representation of DT Systems 137

#Impulse and step responses of LTI DT system

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of unit sample and unit step signal
x1=np.zeros(100)
x1[0]=1
x2=np.ones(100)
#Step 2: Define the system
num=[1,1]
den=[1,-0.5]
#Step 3: Obtaining the impulse and step response
h=signal.lfilter(num,den,x1)
s=signal.lfilter(num,den,x2)
#Plotting the result
plt.subplot(2,1,1),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]')
plt.title('Impulse response (h[n])')
plt.subplot(2,1,2),plt.stem(s),plt.xlabel('n-->'),plt.ylabel('s[n]')
plt.title('Step response (s[n])')
plt.tight_layout()

Fig. 4.12 Python code to obtain the impulse and step response of the system

Impulse response (h[n])

1.5

1.0
h[n]

0.5
0.0
0 20 40 60 80 100
n-->
Step response (s[n])
4
s[n]

0
0 20 40 60 80 100
n-->

Fig. 4.13 Result of python code shown in Fig. 4.12

y½n] = u½n]* u½n] ð4:20Þ

The above equation can be simplified as

138 4 Discrete-Time Systems

#Step response from the impulse response

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Impulse response of the system
h=np.ones(10)
#Step 2: Step response of the system
s=np.cumsum(h,axis=0)
#Step 3: Plotting the result
plt.subplot(2,1,1),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]')
plt.title('Impulse response (h[n])')
plt.subplot(2,1,2),plt.stem(s),plt.xlabel('n-->'),plt.ylabel('s[n]')
plt.title('Step response (s[n])')
plt.tight_layout()

Fig. 4.14 Step response from impulse response

y½n] = ðn þ 1Þu½n]

The python code, which obtains the step response from the impulse response, is
given in Fig. 4.14, and the corresponding output is shown in Fig. 4.15.
Inferences
The following inferences can be made from this experiment:
1. The impulse response of the system is unit step signal.
2. The step response of the system is (n + 1)u[n], which is similar to that of a ramp
signal.
3. The step response obtained using ‘np.cumsum’ command is in agreement with the
theoretical result.
Experiment 4.8 Impulse Response from Step Response
This experiment tries to obtain the impulse response from the step response. If s[n]
represents the step response of discrete-time system, then the impulse response of the
system is given by

h½n] = s½n] - s½n - 1] ð4:21Þ

In this experiment, the step response of the discrete-time system is chosen as unit
step signal. This implies s[n] = u[n]. Upon taking the impulse response as per
Eq. (4.21), one should obtain

h ½ n ] = u ½ n ] - u ½ n - 1] ð4:22Þ

The above equation can be simplified as

4.2 Representation of DT Systems 139

Impulse response (h[n])

1.0
h[n]

0.5

0.0
0 2 4 6 8
n-->
Step response (s[n])
10
s[n]

0
0 2 4 6 8
n-->

Fig. 4.15 Impulse response and step response of the system

h½n] = δ½n] ð4:23Þ

The impulse response of the system is obtained as unit sample signal. The python
code, which obtains the impulse response from the step response, is shown in
Fig. 4.16, and the corresponding output is shown in Fig. 4.17.
Inferences
From this experiment the following inferences can be drawn:
1. From Fig. 4.17, the step response of the system is unit step sequence.
2. From the step response, the impulse response is derived which is unit sample
signal, which is in agreement with the theoretical result.

4.2.4 Pole-Zero Plot of Discrete-Time System

The pole-zero plot of a discrete-time system is plotted in the Z-plane. The position on
the complex plane is represented by rejω. The transfer function of the discrete-time
system is given by

BðzÞ
H ðzÞ = ð4:24Þ
AðzÞ

The zeros are value of ‘z’ for which B(z) = 0. In other words, zeros are the
complex frequencies that make the overall gain of the transfer function is zero. The
140 4 Discrete-Time Systems

#Impulse response from step response

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Step response of the system
s=np.ones(100)
s1=np.zeros(len(s)+1)
s1[1:]=s
h=s-s1[0:len(s)] #s[n]-s[n-1]
plt.subplot(2,1,1),plt.stem(s),plt.xlabel('n-->'),plt.ylabel('s[n]'),
plt.title('Step response (s[n])')
plt.subplot(2,1,2),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]'),
plt.title('Impulse response (h[n])')
plt.tight_layout()

Fig. 4.16 Impulse response from the step response

Step response (s[n])

1.0
s[n]

0.5

0.0
0 20 40 60 80 100
n-->
Impulse response (h[n])
1.0
h[n]

0.5

0.0
0 20 40 60 80 100
n-->

Fig. 4.17 Result of python code shown in Fig. 4.16

poles are the values of ‘z’ for which A(z) = 0. The poles are the complex frequencies
that make the overall gain of the transfer function is infinite. The poles and zeros in
the Z-plane are indicated by the symbol ‘x’ and ‘o’, respectively.
Experiment 4.9 Plotting the Zeros of Non-recursive System
The objective of this experiment is to plot the zeros of the non-recursive system.
Consider two discrete-time system with the transfer function H1(z) = 1 - z-1 and
H2(z) = 1 + z-1. The built-in function ‘tf2zpk’ in scipy library is utilized to plot the Z-
plane of the two systems. The python code does this task is shown in Fig. 4.18, and
the corresponding output is shown in Fig. 4.19.
4.2 Representation of DT Systems 141

#Pole-zero plot
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#To plot the unit circle
theta=np.linspace(0,2*np.pi,100)
#Defining system-1
num1, den1=[1,-1],[1]
z1,p1,k1=signal.tf2zpk(num1,den1)
#Defining system-2
num2,den2=[1,1],[1]
z2,p2,k2=signal.tf2zpk(num2,den2)
#To plot unit circle
plt.subplot(1,2,1),plt.plot(np.real(z1),np.imag(z1),'ko')
plt.plot(np.real(p1),np.imag(p1),'rx'),plt.plot(np.cos(theta),np.sin(theta))
plt.title('Z-plane of system-1'),plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$')
plt.subplot(1,2,2), plt.plot(np.real(z2),np.imag(z2),'ko'),
plt.plot(np.real(p2),np.imag(p2),'rx'),plt.plot(np.cos(theta),np.sin(theta)),
plt.title('Z-plane of system-2'),plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$')
plt.tight_layout()

Fig. 4.18 Python code to plot the Z-plane of the given system

Inferences
The following inferences can be made for discrete-time systems 1 and 2:
1. System-1 has zero at z = 1. This implies that the zero occurs at ω = 0. This zero
will block all low frequency components; hence, the system will act like a high
pass filter.
2. System-2 has zero at z = -1. This implies that the zero occurs at ω = π. The
system-2 will block all high frequency components. Thus, the system-2 will act
like a low pass filter.
Task
1. From the pole-zero plot, will it be possible to find whether systems-1 and -2 are
minimum phase system or not?
Experiment 4.10 Plot the Magnitude and Phase Responses of Non-recursive
System
The objective of this experiment is to plot the magnitude and phase responses of the
given non-recursive systems. Consider two discrete-time system with the transfer
function H1(z) = 1 - z-1 and H2(z) = 1 + z-1. The built-in function ‘freqz’ in scipy
library can be used to obtain the magnitude and phase response of the system. The
python code, which does this task, is shown in Fig. 4.20, and the corresponding
output is shown in Fig. 4.21.
142 4 Discrete-Time Systems

Z-plane of system-1 Z-plane of system-2

1.00 1.00

0.75 0.75

0.50 0.50

0.25 0.25

0.00 0.00
jw

jw
–0.25 –0.25

–0.50 –0.50

–0.75 –0.75

–1.00 –1.00
–1 0 1 –1 0 1
s s

Fig. 4.19 Result of python code shown in Fig. 4.18

Inferences
The following inferences are made from this experiment:
1. From the magnitude responses of system-1 and system-2, it is possible to infer
that system-1 acts like a high pass filter and system-2 acts like low pass filter.
2. From the phase responses of the two systems, it is possible to infer that both the
systems exhibit linear phase characteristics in the pass band. This means that the
phase angle varies linearly with respect to frequency. The linear phase response
implies that the system will not exhibit phase distortion. The linear phase
characteristics is an important attribute of finite impulse response filter.
Task
1. In the above experiment, what is the purpose of the command ‘np.unwrap’? What
will happen if this command is not included in the program?

4.3 Responses of Discrete-Time System

Response of discrete-time system refer to how the discrete-time system react to

different types of test signals. The response of discrete-time system to unit sample
input signal is referred to impulse response of the system. The response of discrete-
time system unit step input signal is referred as step response of the system. The
response of discrete-time system to complex exponential signal is referred as
frequency response of the system. The frequency response of the system comprises
of magnitude response and phase response.
4.3 Responses of Discrete-Time System 143

#Magnitude and phase response of discrete-time systems

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#To plot the unit circle
theta=np.linspace(0,2*np.pi,100)
#Defining system-1
num1,den1=[1,-1],[1]
#Defining system-2
num2,den2=[1,1],[1]
#Magnitude and phase response of systems
w1,H1=signal.freqz(num1,den1)
w2,H2=signal.freqz(num2,den2)
#Plotting the magnitude and phase response of the systems
plt.subplot(2,2,1),plt.plot(w1,20*np.log10(np.abs(H1)))
plt.xlabel('$\omega$'),plt.ylabel('Magnitude'),plt.title('Magnitude response-System1')
plt.subplot(2,2,2),plt.plot(w2,20*np.log10(np.abs(H2)))
plt.xlabel('$\omega$'),plt.ylabel('Magnitude'),plt.title('Magnitude response-System2')
plt.subplot(2,2,3),plt.plot(w1,np.unwrap(np.angle(H1)))
plt.xlabel('$\omega$'),plt.ylabel('Phase'),plt.title('Phase response-System1')
plt.subplot(2,2,4),plt.plot(w2,np.unwrap(np.angle(H2)))
plt.xlabel('$\omega$'),plt.ylabel('Phase'),plt.title('Phase response-System2')
plt.tight_layout()

Fig. 4.20 Python code to plot the magnitude and phase response of the system

Magnitude response-System1 Magnitude response-System2

0 0
Magnitude

Magnitude

–20 –20

–40 –40
0 1 2 3 0 1 2 3
w w
Phase response-System1 Phase response-System2
1.5 0.0
–0.5
Phase

Phase

1.0
0.5 –1.0

0.0 –1.5
0 1 2 3 0 1 2 3
w w

Fig. 4.21 Result of python code shown in Fig. 4.20

144 4 Discrete-Time Systems

Experiment 4.11 Responses of Discrete-Time System

This experiment deals the computation of impulse response, pole-zero plot and
frequency response of the discrete-time system which is given by h[n] = anu
[n]. Let us choose the value of a is 1. The python code, which obtains the impulse
response, pole-zero plot, magnitude and phase response of the discrete-time system,
is shown in Fig. 4.22, and the corresponding output is shown in Fig. 4.23.
Inferences
The following inferences can be drawn from this experiment:
1. The impulse response of the signal is obtained as h[n] = u[n]. It is possible to
observe that the impulse response of the system takes a value of ‘1’. The impulse
response is not absolutely summable. Hence, the system is not stable.
2. The pole-zero plot indicates a pole at z = 1. This means that the pole is placed at
ω = 0. For the discrete-time system to be stable, the pole should lie inside the unit
circle. Here the pole lies on the unit circle. Hence, the system is not BIBO stable.
The order of the system is one.
3. The magnitude response indicates that the gain of the system at low frequency is
high and it decreases with increase in frequency.
4. The phase angle varies linearly with respect to frequency.

#Responses of discrete-time system

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
h=np.zeros(100)
h[0]=1
#Defining system 1
num,den=[1],[1,-1]
#Obtaining the impulse responses of the system
h1=signal.lfilter(num,den,h)
#Obtaining the pole-zero plot of the system
z,p,k=signal.tf2zpk(num,den)
#To obtain the frequency responses of the three systems
w,H=signal.freqz(num,den)
theta=np.linspace(0,2*np.pi,100)
plt.subplot(2,2,1),plt.stem(h1),plt.xlabel('n-->'),plt.ylabel('h[n]'),
plt.title('Impulse response (h[n])')
plt.subplot(2,2,2),plt.plot(w,20*np.log10(np.abs(H)))
plt.xlabel('$\omega$-->'),plt.ylabel('|H(j$\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,2,3),plt.plot(np.real(z),np.imag(z),'ko'),
plt.plot(np.real(p),np.imag(p),'rx'),plt.plot(np.cos(theta),np.sin(theta))
plt.title('Z-plane'),plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$')
plt.subplot(2,2,4),plt.plot(w,np.angle(H))
plt.xlabel('$\omega$-->'),plt.ylabel('$\u2220$H(j$\omega$)'),plt.title('Phase response')
plt.tight_layout()

Fig. 4.22 Python code to obtain the responses of the discrete-time system
4.4 Different Representations and Response of Unit Delay Discrete-Time System 145

Impulse response (h[n]) Magnitude response

1.0 40

|H((jw)|
h[n]

0.5 20

0
0.0
0 25 20 75 100 0 1 2 3
n--> w -->
Z-plane Phase response
1 0

<H(jw)
0
jw

–1

–1
–1.0 –0.5 0.0 0.5 1.0 0 1 2 3
s w -->

Fig. 4.23 Result of python code shown in Fig. 4.22

Task
1. From the impulse response and the pole-zero plot, will it be possible to comment
on the stability of the system?

4.4 Different Representations and Response of Unit Delay

Discrete-Time System

Consider a discrete-time system, which introduces unit delay between the input and
output signal. The different representations of unit delay system are discussed first.
(a) Block diagram representation
The block diagram, which represents unit delay system, is given in Fig. 4.24.
(b) Transfer function representation
The transfer function representation of unit delay system is given by

H ðzÞ = z - 1 ð4:25Þ

(c) Impulse response of unit delay system

The impulse response of discrete-time system is obtained by taking inverse Z-
transform of the transfer function. It is given by
146 4 Discrete-Time Systems

x[n] y[n]
z-1

Fig. 4.24 Block diagram representation of unit delay system

h½ n] = Z - 1 fH ð z Þ g ð4:26Þ

Substituting the expression of H(z) from Eq. (4.25) in Eq. (4.26), we get

h½ n] = Z - 1 z - 1

Upon taking inverse Z-transform, the expression for impulse response is

obtained as

h½n] = δ½n - 1] ð4:27Þ

(d) Step response of the system

The step response of the system is the response of the system to unit step input
signal. For a linear time-invariant system, the relationship between the input and
output signal is given by

y½n] = x½n] * h½n] ð4:28Þ

In the above equation, ‘*’ represents the convolution. To obtain the step
response of the system, x[n] = u[n] and h[n] = δ[n - 1], which is obtained from
the previous result.

s½n] = u½n] * δ½n - 1]

The above equation can be simplified by using the property of delta function
as

s ½ n ] = u ½ n - 1]

Experiment 4.12 Unit Delay DT System Analysis

The objective of this experiment is to realize unit delay system using python. The
program consists of two sections. First section obtains the response of the unit delay
system like, impulse response, step response, magnitude and phase response. The
next section deals with exciting the unit delay system with a sinusoidal input signal
and obtaining the output signal. The output signal should be a delayed (one unit
4.4 Different Representations and Response of Unit Delay Discrete-Time System 147

#Unit delay system

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step1: Defining the system
num,den=[0,1],[1]
#Part1: Obtaining the responses of discrete-time system
#Impulse response of the system
h1=np.zeros(10)
h1[0]=1
h=signal.lfilter(num,den,h1)
#Step response of the system
s1=np.ones(10)
s=signal.lfilter(num,den,s1)
#Magnitude and phase response of the system
w,H=signal.freqz(num,den)
#Plotting different responses of the system
plt.figure(1),plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]'),
plt.title('Unit delay system (h[n])'),plt.subplot(2,2,2),plt.plot(w,np.abs(H))
plt.xlabel('$\omega$-->'),plt.ylabel('|H(j$\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,2,3),plt.stem(s),plt.xlabel('n-->'),plt.ylabel('s[n]'),
plt.title('Step response'),plt.subplot(2,2,4),plt.plot(w,np.angle(H))
plt.xlabel('$\omega$-->'),plt.ylabel('$\u2220$H(j$\omega$)'),plt.title('Phase response')
plt.tight_layout()
#Part2: Input and output of the system
t=np.linspace(0,1,100)
x=np.sin(2*np.pi*5*t)
y=signal.lfilter(num,den,x)
plt.figure(2),plt.plot(t,x,t,y),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.legend(['Input','Output'],loc=1),plt.title('Input and output of Unit delay system')
plt.tight_layout()

Fig. 4.25 Python code to obtain the response of unit delay system

delay) version of the input signal. The python code, which accomplishes this task, is
shown in Fig. 4.25, and the corresponding outputs are shown in Figs. 4.26 and 4.27,
respectively.
Inferences
From this experiment following inferences can be made:
1. From Fig. 4.25, it is possible to observe that the python code consists of two
sections. Section 1 obtains the responses of the unit delay system, which include
(a) impulse response, (b) step response, (c) magnitude response and (d) phase
response.
2. The section of python code simulates sinusoidal signal of 5 Hz frequency as the
input signal, and it is fed to the unit delay system to obtain the output signal. The
input is represented by the variable ‘x’, and the variable represents the output ‘y’.
148 4 Discrete-Time Systems

Unit delay system (h[n]) Magnitude response

1.0 1.05

|H(jw)|
h[n]

0.5 1.00

0.0 0.95
0 2 4 6 8 0 1 2 3
n--> w -->
Step response Phase response
1.0 0

<H(jw)
s[n]

0.5
–2
0.0
0 2 4 6 8 0 1 2 3
n--> w -->

Fig. 4.26 Responses of unit delay system

Input and output of Unit delay system

1.00 Input
Output
0.75

0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 4.27 Input and output of unit delay system

3. From Fig. 4.26, the impulse response of the filter is obtained as h[n] = δ[n - 1],
which is in agreement with the theoretical result. The step response of the system
is given by s[n] = u[n - 1], which is in agreement with the theoretical result. The
magnitude response of the unit delay system is unity, which implies that the
system will not affect the magnitude of the input signal. The phase response
varies linearly with respect to frequency.
4.5 Properties of Discrete-Time System 149

4. Figure 4.27 depicts the input and output signal. The input signal is a sinusoidal
signal of 5 Hz frequency. From this figure, it is possible to observe that the output
signal is a delayed version of the input signal. The delay between the input and
output signal is one unit, which justifies the term the system is a unit delay
system.
Task
1. From the impulse response, will it be possible to comment on the stability of the
system? If so, state whether the system is stable or unstable.

4.5 Properties of Discrete-Time System

The properties of discrete-time system discussed in this section include (1) linearity
property, (2) time shift property, (3) causality and (4) stability.

4.5.1 Linearity Property

A discrete-time system is linear if it obeys superposition principle. According to the

superposition principle, the system should obey both the homogeneity and additivity
properties. According to the homogeneity property, scaling of the input should result
in scaling of the output. Both these properties are expressed in Table 4.1.
For a discrete-time system to be linear, the system should be a relaxed system, and
it should obey superposition principle. For a relaxed system, zero input should result
in zero output.
Experiment 4.13 Testing the Linearity Property of Discrete-Time System
This experiment tries to check whether the given discrete-time system is a linear or
nonlinear. The relationship between the input and output of the system is given by y
[n] = nx[n]. For the system to be linear, the response of the system to the weighted
sum of input is equal to the sum of the weighted responses. The python code, which
examines whether the given system is linear or not, is given in Fig. 4.28, and the
corresponding output is shown in Fig. 4.29.
Inferences
The following inferences can be made from this experiment:
From Fig. 4.29, it is possible to observe that the response of the system to the
weighted sum of input is equal to the sum of the weighted responses. The system
obeys superposition principle; hence, it is a linear system.
Task
1. The relationship between the input and output of the system is given by y[n] = nx
[n] + 5. An offset being added. Will it affect the linearity of the system? Write a
python code to illustrate that modified system is a non-linear system.
150 4 Discrete-Time Systems

Table 4.1 Superposition principle

S. No Property Representation Meaning
1 Homogeneity Scaling of the
x[n] Linear y[n] αx[n] Linear αy[n]
property input x[n] by a
system system
factor ‘α’ results in
scaling of the out-
put y[n] by the
same factor ‘α’
2 Additivity Response of the
property x1[n] Linear y1[n] system to sum of
system
inputs is equal to
x1[n]+ x2[n] Linear y1[n]+ y2[n]
sum of individual
system responses
x2[n] Linear y2[n]
system

#Test for linearity property

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-10,11,1)
#Step1: Defining the two inputs
x1=(n==1)
x2=(n==2)
#Step2: Defining the scaling factors
alpha,beta=2,4
y1=n*x1
y2=n*x2
#Step3: Response due to weighted sum of input
y_1=n*(alpha*x1+beta*x2)
#Step3:Sum of weighted response
y_2=alpha*y1+beta*y2
#Step4:Plotting the results
plt.subplot(2,1,1),plt.stem(n,y_1),plt.xticks(n),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Response due to weighted sum of inputs')
plt.subplot(2,1,2),plt.stem(n,y_2),plt.xticks(n),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Sum of weighted responses')
plt.tight_layout()

Fig. 4.28 Python code to test the linearity of the given discrete-time system
4.5 Properties of Discrete-Time System 151

Response due to weighted sum of inputs

7.5
Amplitude
5.0
2.5
0.0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->
Sum of weighted responses
7.5
Amplitude

5.0
2.5
0.0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 4.29 Result of python code shown in Fig. 4.28

4.5.2 Time-Invariant and Time-Variant System

A DT system is said to be time-invariant or shift-invariant if its input-output

characteristics do not change with time. This implies time shift of the input causes
a corresponding shift in the output. This implies that the system response is inde-
pendent of time.

Property Representation Meaning

Time Shift in the input signal x[n]
shift x[n] Time- y[n] x[n-k] Time- y[n-k] by a factor of ‘k’ should
invariant invariant
property result in shift in the output
system system
signal y[n] by the same fac-
tor ‘k’

Experiment 4.14 Testing the Time-Invariant Property of Discrete-Time System

The aim of this experiment is to check the given DT system function is time varying
or time invarying system by python code. The relationship between the input and
output of the discrete-time system is given by y[n] = nex[n]. A discrete-time system is
time-invariant, if a time shift in the input signal should result in a time shift in the
output signal. The python code, which tests the time-invariance property of the given
discrete-time system, is given in Fig. 4.30, and the corresponding output is shown in
Fig. 4.31.
Inference
From Fig. 4.31, the output due to time shift in the input is not equal to the time shift
in the output; hence, the system is time-variant.
152 4 Discrete-Time Systems

#Test for time-invariance property

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-10,11,1)
#Step1: Defining the input
x=(n==0)
k=5 #Shift parameter
#Step2: Time shift in the input
x1=(n==k)
y1=n*np.exp(x1)
#Step 3: Time shift in the output
y2=(n-k)*np.exp(x1)
#Step4:Plotting the results
plt.subplot(2,1,1),plt.stem(n,y1),plt.xticks(n),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Output due to time shift in the input')
plt.subplot(2,1,2),plt.stem(n,y2),plt.xticks(n),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Time shift in the output')
plt.tight_layout()

Fig. 4.30 Python code to test the time-invariance property

Output due to time shift in the input

10
Amplitude

–10
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->
Time shift in the output
Amplitude

–10

–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 4.31 Result of python code shown in Fig. 4.30

Task
1. The relationship between the input and output of the system is given by
y[n] = x[-n]. Examine whether the system is time-variant or not.
4.5 Properties of Discrete-Time System 153

4.5.3 Causal and Non-causal System

A discrete-time system is causal if it is non-anticipatory. The output of the system

should not be dependent on the future value of the input. A DT-LTI system is causal
if the impulse response of the system is zero for n < 0.
Experiment 4.15 Check the DT System Is Causal or Non-causal
This experiment uses python to discuss whether the given DT system is causal or
not. Let us consider four discrete-time systems with the impulse responses:
h1[n] = δ[n + 1] + δ[n] + δ[n - 1], h2[n] = δ[n] + δ[n - 1] + δ[n - 2], h3[n] = u
[n] - u[n - 1] and h4[n] = δ[n + 2] + δ[n + 1] + δ[n] for -10 ≤ n ≤ 10. The python
code, which plots the impulse responses of the above-mentioned discrete-time
systems, is shown in Fig. 4.32, and the corresponding output is shown in Fig. 4.33.
Inferences
For the discrete-time system to be causal, the impulse response should be equal to
zero for n < 0.
1. The impulse response of system-1 is non-zero for n < 0; hence, discrete-time
system-1 is a non-causal system.
2. The impulse response of system-2 is zero for n < 0; hence, discrete-time system-2
is a causal system.

#Impulse response of DT systems

import numpy as np
import matplotlib.pyplot as plt
N = 10
n= np.arange(-N, N + 1,1, dtype = float)
#Defining the impulse response of the four systems
h1=np.zeros(2*N+1,dtype='float')
h2=np.zeros(2*N+1,dtype='float')
h3=np.zeros(2*N+1,dtype='float')
h4=np.zeros(2*N+1,dtype='float')
h1=[1 if (i==-1) | (i==0)|(i==1) else 0 for i in n]
h2=[1 if (i==0) | (i==1)|(i==2) else 0 for i in n]
h3=[1 if (i==0) else 0 for i in n]
h4=[1 if (i==-2) | (i==-1)|(i==0) else 0 for i in n]
#Plotting the impulse response
plt.subplot(2,2,1),plt.stem(n,h1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$h_1[n]$'),plt.subplot(2,2,2),plt.stem(n,h2),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('$h_2[n]$'),plt.subplot(2,2,3),plt.stem(n,h3),
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_3[n]$'),plt.subplot(2,2,4),
plt.stem(n,h4),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_4[n]$')
plt.tight_layout()

Fig. 4.32 Python code to plot the impulse response of the system
154 4 Discrete-Time Systems

h1[n] h2[n]
1.0 1.0
Amplitude

Amplitude
0.5 0.5

0.0 0.0
–10 –5 0 5 10 –10 –5 0 5 10
n--> n-->

h3[n] h4[n]
1.0 1.0
Amplitude

Amplitude
0.5 0.5

0.0 0.0
–10 –5 0 5 10 –10 –5 0 5 10
n--> n-->

Fig. 4.33 Impulse responses of the discrete-time systems

3. The impulse response of system-3 is zero for n < 0; hence, discrete-time system-3
is a causal system.
4. The impulse response of system-4 is non-zero for n < 0; hence, discrete-time
system-4 is a non-causal system.
5. From the impulse response, it is possible to infer whether the discrete-time system
is causal or not.
Task
1. The relationship between the input and output of the system is expressed as y
[n] = x[-n]. Examine whether the system is causal or not?

4.5.4 Stability of Discrete-Time System

A discrete-time system is stable if the following criterion are met.

(a) BIBO stability criterion: A discrete-time system is stable if bounded input
results in bounded output.
(b) Stability criterion with respect to impulse response: A discrete-time system is
stable if the impulse response of the system is absolutely summable.
(c) Stability with respect to position of pole: A discrete-time system is stable if the
pole of the discrete-time system lies within the unit circle.
4.5 Properties of Discrete-Time System 155

It is to be noted that all the above-mentioned criteria are not independent criteria.
It means that one criterion implies the other.
Experiment 4.16 BIBO Stability Criterion
The aim of this experiment is to obtain the given DT system is stable or not. The
relationship between the input and output of a linear time-invariant discrete-time
system is given by y[n] = x[n] + y[n - 1]. The relationship between the input and
output of the system is given by

y ½ n] = x ½ n] þ y ½ n - 1]

Taking Z-transform on both sides of the above equation, we get

Y ð zÞ = X ð zÞ þ z - 1 Y ðz Þ

The transfer function of the system is obtained as

Y ðzÞ 1
H ðzÞ = =
X ð zÞ 1 - z - 1

The python code, which applies the unit step input signal to the above-mentioned
system, is given in Fig. 4.34, and the corresponding output is shown in Fig. 4.35.
Inference
The following inferences can be made from this experiment:
Figure 4.36 shows that the input signal (x[n]) is a unit step signal, a bounded input
signal. By observing the output signal (y[n]) is not a bounded signal. The output
signal is a ramp signal, which is not bounded. This shows that bounded input signal
to the system does not result in bounded output signal. Hence, the system is not
BIBO stable.
Experiment 4.17 Stability Criterion Based on the Impulse Response
This experiment discusses the stability of the DT system to be checked from the
impulse response of it. Let us consider two discrete-time LTI systems with impulse
n
responses h1 ½n] = 12 u½n] and h2[n] = (2)nu[n].
For the system to be stable, the impulse response should be absolutely summable.
The impulse response of system-1 is absolutely summable; hence, it is stable. On the
other hand, the impulse response of the system-2 is not absolutely summable; hence,
it is unstable. The python code, which plots the impulse response and obtains the
absolute sum of the impulse response of the above-mentioned discrete-time LTI
systems, is given in Fig. 4.36, and the corresponding output is shown in Figs. 4.37
and 4.38.
156 4 Discrete-Time Systems

#BIBO stability criterion

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step1: Defining the system
num,den=[1],[1,-1]
#Step 2: Generation of unit step input signal
N=50
n=np.arange(N)
x=np.ones(N)
#Step 3: Obtain the output of the system
y=signal.lfilter(num,den,x)
#Step 4: Plot the input and output of the system
plt.subplot(2,1,1),plt.stem(n,x),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('x[n]'),plt.subplot(2,1,2),
plt.stem(n,y),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('y[n]')
plt.tight_layout()

Fig. 4.34 Python code to check the BIBO criterion of discrete-time LTI system

x[n]
1.0
Amplitude

0.5

0.0
0 10 20 30 40 50
n-->
y[n]

40
Amplitude

20

0
0 10 20 30 40 50
n-->

Fig. 4.35 Input-output of the discrete-time LTI system

Inferences
The following inferences can be made from this experiment:
1. From Fig. 4.37, the impulse response of system-1 is absolutely summable,
whereas the impulse response of system-2 is not absolutely summable. Hence,
system-1 is a stable system, whereas system-2 is not a stable system.
4.5 Properties of Discrete-Time System 157

#Stability based on impulse response

import numpy as np
import matplotlib.pyplot as plt
N=50
n=np.arange(N)
#Step 1: Defining the impulse response of the two systems
h1=0.5**n
h2=2.0**n
#Step 2: Obtaining the absolute sum of the impulse response
print('The absolute sum of impulse response of system 1 is:',np.sum(abs(h1)))
print('The absolute sum of impulse response of system 2 is:',np.sum(abs(h2)))
#Step 3: Plotting the impulse response of the two systems
plt.subplot(1,2,1),plt.stem(n,h1)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_1[n]$')
plt.subplot(1,2,2),plt.stem(n,h2)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_2[n]$')
plt.tight_layout()

Fig. 4.36 Python code to test the stability of discrete-time system based on the impulse response

The absolute sum of impulse response of system 1 is: 1.9999999999999982

The absolute sum of impulse response of system 2 is: 1125899906842623.0

Fig. 4.37 The absolute sum of impulse response of the two systems

2. From Fig. 4.38, the impulse response of system-1 (h1[n]) is converging, whereas
the impulse response of system-2 (h2[n]) is diverging. Therefore, system-1 is
stable, and system-2 is unstable.
Experiment 4.18 Stability Based on the Location of Poles of the Discrete-Time
System
This experiment discusses the verification of the stability of the DT system based on
the location of the poles of the DT system. Let us consider the transfer function of
two discrete-time LTI systems given by H 1 ðzÞ = ð1 - 0:2z - 1 Þ2ð1 - 0:4z - 1 Þ and
H 2 ðzÞ = ð1 - 2z - 1 Þ1ð1 - 4z - 1 Þ. For the discrete-time system to be stable, the poles should
lie within the unit circle. The poles of system-1 defined the transfer function H1(z)
lies within the unit circle; hence, the system is stable, whereas the poles of the
system-2 defined by the transfer function H2(z) lies outside the unit circle; hence, the
system is unstable. The python code to obtain the pole-zero plot of the above-
mentioned discrete-time systems is given in Fig. 4.39, and the corresponding output
is shown in Fig. 4.40.
158 4 Discrete-Time Systems

h1[n] h2[n]
1e14
1.0
5
0.8
4
0.6
Amplitude

Amplitude
3

0.4
2

0.2 1

0.0 0
0 20 40 0 20 40
n--> n-->

Fig. 4.38 Plot of impulse responses of the discrete-time systems

Inference
From Fig. 4.40, the poles of discrete-time system-1 lies within the unit circle; hence,
the system is stable. The poles of discrete-time system-2 lies outside the unit circle;
hence, the system is unstable.
Task
1. Comment on the stability of a discrete-time system, whose current output
depends on current and past input signal values. Write a python code to validate
your answer.

4.5.5 Invertibility of Discrete-Time System

A discrete-time system is invertible if distinct input results in distinct output.

Experiment 4.19 Examining the Invertibility of Discrete-Time System
This experiment tries to examine the invertibility of two discrete-time systems,
whose input-output relationship is given by (1) system 1: y[n] = x[2n] and (2) system
2: y[n] = x[n/2]. The python code, which examines the invertibility of discrete-time
system, is shown in Fig. 4.41, and the corresponding output is shown in Figs. 4.42
and 4.43.
4.5 Properties of Discrete-Time System 159

#Pole-zero plot of discrete-time systems

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Defining system 1
num1,den1=[2],[1,-0.6,0.032]
#Defining system 2
num2,den2=[1],[1,-6,8]
#Obtaining the pole-zero plot of the system
z1,p1,k1=signal.tf2zpk(num1,den1)
z2,p2,k2=signal.tf2zpk(num2,den2)
theta=np.linspace(0,2*np.pi,100)
re=0;
for i in range(len(p1)):
if (p1[i].real>1) or (p1[i].imag>1):re=re+1;
else:re=0;
if re==0:print('System is Stable')
else:print('System is Unstable')
re1=0;
for i in range(len(p2)):
if (p2[i].real>1) or (p2[i].imag>1):re1=re1+1;
else:re1=0;
if re1==0:print('System is Stable')
else:print('System is Unstable')
#Plotting the pole-zero plot
plt.subplot(2,1,1),plt.plot(np.real(z1),np.imag(z1),'ko'),plt.plot(np.real(p1),
np.imag(p1),'rx'),plt.plot(np.cos(theta),np.sin(theta)),plt.title('Z-plane of system-1')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.subplot(2,1,2),
plt.plot(np.real(z2),np.imag(z2),'ko'),plt.plot(np.real(p2),np.imag(p2),'rx')
plt.plot(np.cos(theta),np.sin(theta)),plt.title('Z-plane of system-2'),
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$')
plt.tight_layout()

Fig. 4.39 Python code to obtain the pole-zero plot of the given discrete-time systems

Inferences
The following are the inferences:
1. Two discrete-time systems considered in this example are the following: (1) Sys-
tem-1, downsampling by a factor of 2, and (2) system-2, upsampling by a factor
of 2.
2. The input signals considered to excite the discrete-time signals are denoted as
x1[n] and x2[n]. x1[n] is a DC signal, whereas x2[n] is the highest frequency digital
signal. The output signals of system-1 for the inputs x1[n] and x2[n] are denoted as
s11[n] and s22[n], respectively.
160 4 Discrete-Time Systems

Z-plane of system-1
1

jw
–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
s
Z-plane of system-2
1

jw
System is Stable
–1
System is Unstable –1 0 1 2 3 4
s

Fig. 4.40 Result and pole-zero plot of the discrete-time systems

#Invertibility of DT system
import numpy as np
import matplotlib.pyplot as plt
N=10
n=np.arange(N)
x1=np.ones(N)
x2=np.exp(1j*np.pi*n)
#System 1 output
s11=x1[::2]
s12=x2[::2]
#System 2 output
s21=np.zeros(2*len(x1))
s21[::2]=x1
s22=np.zeros(2*len(x2))
s22[::2]=x2
plt.figure(1),plt.subplot(2,2,1),plt.stem(n,x1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$x_1[n]$'),plt.subplot(2,2,2),plt.stem(s11),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('$s_{11}[n]$'),plt.subplot(2,2,3),plt.stem(n,x2)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$x_2[n]$'),plt.subplot(2,2,4),
plt.stem(s12),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$s_{12}[n]$')
plt.tight_layout()
plt.figure(2),plt.subplot(2,2,1),plt.stem(n,x1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$x_1[n]$'),plt.subplot(2,2,2),plt.stem(s21),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('$s_{21}[n]$'),plt.subplot(2,2,3),plt.stem(n,x2)
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$x_2[n]$'),plt.subplot(2,2,4),
plt.stem(s22),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$s_{22}[n]$')
plt.tight_layout()

Fig. 4.41 Python code to examine the invertibility of the given discrete-time systems

3. The output of discrete-time system-2 for the input signals x1[n] and x2[n] is
denoted as s21[n] and s22[n], respectively.
4.5 Properties of Discrete-Time System 161

x1[n] s11[n]
1.0 1.0
Amplitude

Amplitude
0.5 0.5

0.0 0.0
0 2 4 6 8 0 1 2 3 4
n--> n-->

x2[n] s12[n]
1 1.0
Amplitude

Amplitude
0 0.5

–1 0.0
0 2 4 6 8 0 1 2 3 4
n--> n-->

Fig. 4.42 Input-output signals of discrete-time system-I

4. Figure 4.42 shows the input and output signals corresponding to system-1
(downsample by a factor of 2). From Fig. 4.42, it is possible to observe that the
input signals x1[n] and x2[n] are different but the output s11[n] and s12[n] are the
same. This shows that the system produces same output for distinct inputs; hence,
the system-1 is a non-invertible system.
5. Figure 4.43 shows the input and output signals corresponding to system-2
(upsample by a factor of 2). From Fig. 4.43, it is possible to observe that the
system produces different output for distinct inputs. Hence, the system-2 is an
invertible system.
Task
1. Write a python code to prove that cascade connection of accumulator and
backward difference system results in an invertible system. System-1 is an
accumulator, whose input-output relation is given by y[n] = x[n] + y[n - 1];
system-2 is a backward system, whose difference equation is given by y[n] = x
[n] - x[n - 1].
Exercises
1. Write a python code to plot the impulse response of the discrete-time system
whose input-output relationship is given by y½n] = x½n] þ 12 y½n - 1]. From the
impulse response plot, will it be possible to comment on the stability of the
system?
2. Write a python code to obtain the state-space representation of the discrete-time
z2
system whose transfer function is given by H ðzÞ = z2 þ4z - 2.
162 4 Discrete-Time Systems

x1[n] s21[n]
1.0 1.0
Amplitude

Amplitude
0.5 0.5

0.0 0.0
0 2 4 6 8 0 5 10 15
n--> n-->

x2[n] s22[n]
1 1
Amplitude

Amplitude
0 0

–1 –1
0 2 4 6 8 0 5 10 15
n--> n-->

Fig. 4.43 Input-output signals of discrete-time system-II

3. The state-space representation of discrete-time system is given by x[k + 1] = Ax

0 1 0
[k] + Bu[k] and y[k] = Cx[k] + Du[k], where A = 1 5 , B= ,
- - 1
6 6
C = ½ 1 0 ] and D = [0]. Write a python code to obtain the transfer function of
the system.
4. The impulse response of a discrete-time system h[n] = (α)n, - 10 ≤ n ≤ 10. Plot
the impulse response for α = 14 , 34 , 1, 2. Use subplot to plot the impulse
responses and comment on the observed result.
5. The relationship between the input and output of a discrete-time system is given
by y½n] = x n2 . The system is a linear, time-variant system. Write a python code
to validate the property.
6. Plot the pole-zero plot of the discrete-time systems whose transfer functions are
1þ12z - 1
given by H 1 ðzÞ = 1 þ 12 z - 1 , H 2 ðzÞ = 1
1 - 34z - 1
and H 3 ðzÞ = 1 - 34z - 1
. Comment on
the observed output.
7. Write a python code to obtain the magnitude and phase response of discrete-time
system whose input-output relationship is given by
y½n] = 13 fx½n] þ x½n - 1] þ x½n - 2]g. Examine whether the system exhibits lin-
ear phase characteristics from the phase response.
8. The impulse response of a discrete-time system is given by h[n] = δ[n] - δ[n -
1]. Use subplot to plot the impulse and step responses of the system. Comment
on the observed result.
4.5 Properties of Discrete-Time System 163

9. Two discrete-time systems with impulse responses h1[n] = u[n] and

h2[n] = δ[n] - δ[n - 1] are connected in cascade. Write a python code to plot
the magnitude and phase responses of the cascaded system and comment on the
observed result.
10. Two discrete-time systems with transfer functions H1(z) = X(z2) and
1 1
H 2 ðzÞ = 1
2 X z2 þ X - z2 are connected in cascade. A sine wave of
5 Hz frequency is fed to the cascaded system. What will be the output of the
system? Write a python code to plot the input and output signals and comment
on the observed result.
Objective Questions
n
1. If ‘h[n]’ represents the impulse response of the system, then y½n] = h½ k ]
k= -1
represents
A. Magnitude response of the system
B. Phase response of the system
C. Shifted impulse response of the system
D. Step response of the system
2. If the variable ‘h’ represents the impulse response of the system, then the
variable ‘y’ in the following code results in

A. Magnitude response of the system

B. Phase response of the system
C. Shifted impulse response of the system
D. Step response of the system
3. A discrete-time system is linear if it obeys
A. Superposition theorem
B. Thevenin’s theorem
C. Tellegen’s theorem
D. Norton’s theorem
4. Assertion: Causal systems are non-anticipatory system.
Reason: In causal system, the current output will not depend on the future
value of the input.
A. Both assertion and reason are true.
B. Assertion is true; reason is false.
C. Assertion is false; reason may be true.
D. Both assertion and reason are false.
164 4 Discrete-Time Systems

5. Identify the system which is NOT a relaxed system:

A. y[n] = nx[n]
B. y[n] = x[-n]
C. y[n] = Ax[n]
D. y[n] = Ax[n] + B
6. The relationship between the input and output of a discrete-time is given by y
[n] = αx2[n] + βx[n] + γ. For the system to be linear
A. α=0
B. β=0
C. γ=0
D. α = 0 and γ = 0
7. The impulse response of discrete-time linear, time-invariant system is given by
n
h½n] = 1 - 12 u½n]. The system is
A. Causal and stable system
B. Non-causal and stable system
C. Causal and unstable system
D. Non-causal and unstable system
8. A linear time-invariant discrete-time system is given by y[n] = Ax2[n] + Bx[n -
1] + Cy[n - 1]. For the system to be static system.
A. A=0
B. B=0
C. C=0
D. B=C=0
9. The transfer function (H(z)) of the system derived from the state-space model is
expressed as
A. C(zI - A)-1B + D
B. D(zI - A)-1B + C
C. D(zI - B)-1A + C
D. C(zI - B)-1A + D
10. The transfer function of a discrete-time system is represented as H(z) = 1 - z-1.
The system has zero at
A. ω=0
B. ω = π/4
C. ω = π/2
D. ω=π
Bibliography 165

z-1
11. The transfer function of a discrete-time system is given by H ðzÞ = .
ðz - 12Þðzþ34Þ
The poles of the system are at
A. - 12 , þ 34
B. 12 , - 34
C. 12 , 34
D. - 12 , - 34
12. The transfer function of a discrete-time system is represented as H(z) = z-1. The
impulse response of the system is given by
A. h[n] = δ[n]
B. h[n] = δ[n - 1]
C. h[n] = δ[n - 2]
D. h[n] = δ[n - 3]
13. Among the transfer function of the discrete-time systems, identify the system
which is NOT BIBO stable system:
A. H ðzÞ = 1
1 - 14z - 1
B. H ðzÞ = 1
1 - 12z - 1
C. H ðzÞ = 1
1 - 34z - 1
D. H ðzÞ = 1
1 - z-1

14. Among the input-output relationship of a given discrete-time systems, identify

the system which is NOT a relaxed system:
A. y[n] = nx[n]
B. y[n] = x[-n]
C. y[n] = ex[n]
D. y[n] = x2[n]

Bibliography

1. Charles L. Phillips, John M. Parr, and Eve A. Riskin, “Signals, Systems, and Transforms”,
Pearson, 2013.
2. David J. Defatta, Joseph G. Lucas, and Villiam S. Hodgkiss, “Digital Signal Processing: A
System Design Approach”, Wiley India Pvt Ltd, 2009.
3. Simon Haykin, and Bary Van Veen, “Signals and Systems”, Wiley, 2005.
4. S. Esakkirajan, T. Veerakumar and Badri N Subudhi, “Digital Signal Processing”, McGraw
Hill, 2021.
5. Jose Unpingco, “Python for Signal Processing”, Springer, 2013.
Chapter 5
Transforms

Learning Objectives
After completing this chapter, the reader is expected to
• Compute the forward and inverse Z-transform.
• Analyse discrete-time system using Z-transform and discrete-time Fourier
transform.
• Compute the spectrum of continuous-time and discrete-time signals.
• Plot and infer the spectrogram of stationary and non-stationary signals.
• Plot and interpret the scalogram of non-stationary signal.

Roadmap of the Chapter

Different transforms discussed in this chapter are given below as a flow diagram.
Transforms are widely used for signal as well as system study. To analyse discrete-
time system, Z-transform is widely used. DTFT can be considered as an evaluation
of Z-transform on a unit circle. DTFT is used to obtain the frequency response of the
system.

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 167
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_5
168 5 Transforms

Transform domain analysis of

signals and systems

Analysis of Representation of
discrete-time systems signals

Z-transform

Discrete-time Frequency domain Time-Frequency Time-Scale

Fourier representation representation representation
transform
Fourier Continuous
transform Short-Time Wavelet
Fourier transform
Discrete
transform
Cosine Discrete
transform Wavelet
transform

Signals can be analysed entirely in the frequency domain. Example of transform,

which gives frequency domain representation, includes Fourier transform and dis-
crete cosine transform. Short-time Fourier transform gives a joint time-frequency
representation of the signal. STFT is an effective tool to analyse the non-stationary
signal. Example of timescale representation includes continuous and discrete wave-
let transform, which are effective in providing multi-resolution (MRA) analysis of
the signals.
PreLab Questions
1. What do you understand by the term ‘spectrum’?
2. What do you mean by unilateral and bilateral Z-transform?
3. What is the region of convergence (ROC) in the context of Z-transform?
4. Mention the essential condition for the function to be called as basis function?
What is the basis of Fourier transform?
5. When applying Fourier analysis to a signal, under which circumstances should
Fourier series analysis be employed, and under which circumstances Fourier
transform be employed?
6. What is the relationship between discrete-time Fourier transform (DTFT) and Z-
transform?
7. What do you mean by a double-sided spectrum of a signal?
8. What is the need of transform in signal analysis?
9. If the DFT of the signal x[n] has to be real, what property should the signal x[n]
satisfy?
10. Compare Fourier transform and short-time Fourier transform.
11. Mention the significant features of discrete cosine transform?
12. Mention one significant advantage of wavelet transform over short-time Fourier
transform.
5.2 Z-Transform 169

5.1 Introduction to Transform

Transform is a tool to analyse signals and systems. Signals are converted from time
or spatial domain to frequency domain using transform. Frequency domain is used to
describe the signal in terms of frequency components. Each frequency has its own
amplitude and phase. From the spectrum, it is possible to interpret the frequencies
present in the signal. Thus, the time domain and the frequency domain representation
of the signal are equivalent. It is possible to transform the signal from time domain to
frequency domain and vice versa without any loss of information. Mathematically,
transform takes the inner product of the signal with the basis function. The inner
product is one way of quantifying the similarity or the dissimilarity of two signals.

5.2 Z-Transform

The Z-transform is a powerful tool to analyse linear, time-invariant discrete-time

systems. The Z-transform for discrete-time signals is the counterpart of the Laplace
transform for continuous-time signals. It simplifies the solution of discrete-time
problems by converting LTI difference equations to algebraic equations and convo-
lution to multiplication. The Z-transform of a discrete-time signal x[n] is defined as

1
X ðzÞ = x½n]z - n ð5:1Þ
n= -1

The above expression is often termed as two-sided Z-transform. Here, z is a

complex variable. The Z-transform of right-sided sequence is expressed as

1
X ð zÞ = x½n]z - n ð5:2Þ
n=0

The Z-transform of left-sided sequence is expressed as

-1
X ðzÞ = x½n]z - n ð5:3Þ
n= -1

Region of Convergence The region of convergence of the Z-transform is the value

of z for which X(z) is finite. The region of convergence allows the unique inversion
of the Z-transform. The ROC depends on the signal x[n] being transformed. The
ROC helps to characterize the system as causal or stable.
170 5 Transforms

5.2.1 Z-Transform of Standard Test Sequences

The Z-transform of the standard test sequences is tabulated in Table 5.1.

Experiment 5.1 Z-Transform of the Unit Sample and Unit Step Sequences
Using sympy Package
This experiment computes the Z-transform of the test sequences like unit sample and
unit step sequences. The python code, which computes the Z-transform of test
sequences, is shown in Fig. 5.1.
Inferences
The following inferences can be made from this experiment:
1. In this experiment sympy library package is utilized to compute the Z-transform of
the test sequences.
2. The python command sym.summation is used for the summation computation,
and sym.KroneckerDelta command is used to define the unit impulse sequence.
3. After executing the python code given in Fig. 5.1, the user has to enter ‘1’ for the
computation of Z-transform of unit sample sequence and ‘2’ for the computation
of Z-transform of unit step sequence.
4. The simulation result of this experiment is shown in Fig. 5.2. From this figure, it
is evident that the simulation result is on par with the theoretical result.

Task
1. Write a python code to obtain the Z-transform of x[n] = nu[n].
Experiment 5.2 Z-Transform of Unit Sample and Unit Step Sequences Using
lcapy Package
This experiment discusses the lcapy package, which can be used to compute the Z-
transform of unit sample and unit step sequences. The python code is shown in
Fig. 5.3.
Inferences
Upon executing the code shown in Fig. 5.3, the result obtained is 1 and 1/(1 - 1/z),
which is in agreement with the theoretical result. The python command delta defines
the unit impulse sequence, 'us' gives the unit step sequence and 'ZT' obtains the Z-
transform.
Experiment 5.3 Z-Transform of x[n] = ejnu[n] and x[n] = cos(n)
This experiment deals with the computation of Z-transform of the given input
sequences x[n] = ejnu[n] and x[n] = cos(n). The python code, which computes the
Z-transform of x[n] = ejnu[n] and x[n] = cos(n), is shown in Fig. 5.4a, and the
corresponding output is shown in Fig. 5.4b.
Inferences
1. From Fig. 5.4a, it is possible to observe that the package lcapy is used to define
the exponential function. lcapy is a Python package for linear circuit analysis.
5.2 Z-Transform 171

Table 5.1 Z-Transform of the standard test sequences

Sequence Transform ROC
δ[n] 1 8z
δ[n - m] z-m 8z except 0 if m > 0 or infinity if
m<0
u[n] z 1
z - 1 or 1 - z - 1
|z| > 1
-u[-n - 1] z 1
z - 1 or 1 - z - 1
|z| < 1
nu[n] z -1
or ð1 -z z - 1 Þ2 |z| > 1
ðz - 1Þ2
(n + 1)u[n] z2
or ð1 - 1z - 1 Þ2 |z| > 1
ðz - 1Þ2
2
n u[n] zðzþ1Þ z -1
ð1þz Þ -1
|z| > 1
ðz - 1Þ3
or ð1 - z - 1 Þ3
n3u[n] zðz2 þ4zþ1Þ z - 1 ð1þ4z - 1 þz - 2 Þ |z| > 1
ðz - 1Þ4
or ð1 - z - 1 Þ4
n4u[n] zðz3 þ11z2 þ11zþ1Þ z - 1 ð1þ11z - 1 þ11z - 2 þz - 3 Þ |z| > 1
ðz - 1Þ5
or ð1 - z - 1 Þ5
n5u[n] zðz4 þ26z3 þ66z2 þ26zþ1Þ |z| > 1
ðz - 1Þ6
or
z - 1 ð1þ26z - 1 þ66z - 2 þ26z - 3 þz - 4 Þ
ð1 - z - 1 Þ6
(-1) u[n] n z
zþ1 or 1
1þz - 1
|z| > 1
-(-1)nu[-n - z
zþ1 or 1
1þz - 1
|z| < 1
1]
anu[n] z
z-a
1
or 1 - az -1 |z| > a
|n|
a 1 - a2
ð1 - az - 1 Þð1 - azÞ
a < jzj < 1
a

an - 1u[n - 1] 1 z-1 |z| > a

z - a or ð1 - az - 1 Þ
(-a) u[n] n z 1
zþa or 1þaz - 1
|z| > a
-anu[-n - 1] z 1
z - a or 1 - az - 1
|z| < a
(n - 1)
-a u[-n] 1 z-1 |z| < a
z - a or 1 - az - 1
an[u[n] - u[n - 1 - aN z - N |z| > 0
1 - az - 1
N]]
-nbnu[-n - 1] zb
or bz - 1 |z| < b
ðz - bÞ2 ð1 - bz - 1 Þ
2

nanu[n] za -1
or ð1 -azaz - 1 Þ2 |z| > a
ðz - aÞ2
cos(ω0n)u[n] z - z cos ω0
2 -1
or 1 -12z--z1 coscos ω0 |z| > 1
z2 - 2z cos ω0 þ1 ω0 þ z - 2
sin(ω0n)u[n] z sin ω0 z-1
sin ω0
or 1 - 2z - 1 cos |z| > 1
z2 - 2z cos ω0 þ1 ω0 þ z - 2
[a sin ω0n]u[n]
n az sin ω0 -1
or 1 - 2azaz- 1 cossinωω0 þ0 a2 z - 2 |z| > |a|
z2 - 2az cos ω0 þa2
[a cos ω0n]u[n]
n
zðz - a cos ω0 Þ ð1 - az cos ω0 Þ
-1
|z| > |a|
z2 - 2az cos ω0 þa2 or 1 - 2az - 1 cos ω0 þa2 z - 2

2. Upon executing the commands shown in Fig. 5.4a, the result obtained is shown in
Fig. 5.4b. The result of python code is in agreement with the theoretical result.
172 5 Transforms

#Z-transform of unit sample and step sequences

import sympy as sym
n = sym.symbols('n', integer=True)
z = sym.symbols('z', complex=True)
S=int(input("Enter : (1= Unit Impulse, 2=Unit Step) : "));
if (S==1):
X = sym.summation(sym.KroneckerDelta(n, 0) * z**(-n), (n, -sym.oo, sym.oo));
print('X(z) = ', X)
elif(S==2):
X = sym.summation(1*z**-n,(n,0,sym.oo));
print('X(z) = ', X)
else:
print('Please enter the correct number')

Fig. 5.1 Python code for Z-transform of unit sample sequence

Enter : (1= Unit Impulse, 2=Unit Step) : 1

X(z) = 1
Enter : (1= Unit Impulse, 2=Unit Step) : 2
X(z) = Piecewise((1/(1 - 1/z), 1/Abs(z) < 1), (Sum(z**(-n), (n , 0, oo)), True))
Enter : (1= Unit Impulse, 2=Unit Step) : 3
Please enter the correct number

Fig. 5.2 Simulation result of the python code given in Fig. 5.1

Fig. 5.3 Python code for Z-

transform of unit sample and #Z-transform of unit sample and unit step signal
step sequences from lcapy import n,delta,us
x =delta(n)
Xz=x.ZT()
print(Xz)
x1 = us(n)
Yz=x1.ZT()
print(Yz)

n
Experiment 5.4 Z-Transform of x½n] = 12 u½n]
This experiment discusses the python code to obtain the Z-transform of x½n] =
1 n
2 u½n] and the corresponding output, which is shown in Fig. 5.5.

Inferences
1. Figure 5.5 shows that the us variable is called from the lcapy package as a unit
step sequence and multiplied by (1/2)n to get x[n].
2. The Z-transform of x[n] is obtained using ‘ZT’ python command, and the result is
displayed in Fig. 5.5. This result confirms the theoretical result.
5.3 Inverse Z-Transform 173

import lcapy
from lcapy import n
x=lcapy.exp(1j*n)
y=lcapy.cos(n)
Xz=x.ZT()
print(Xz)
Yz=y.ZT() z/(z - exp(j))
print(Yz) z*(z - cos(1))/(z**2 - 2*z*cos(1) + 1)

(a) Pyhton code (b) Simulation result

Fig. 5.4 Python code to Experiment 5.3. (a) Pyhton code. (b) Simulation result

from lcapy import n,us

x = (1/2)**n*us(n)
Xz=x.ZT()
print(Xz) 2*z/(2*z - 1)

Fig. 5.5 Python code and its simulation result

Task
3 n
1. Write a python code to obtain the Z-transform of x½n] = 4 u½n].

5.3 Inverse Z-Transform

This section discusses some of the experiments related to the inverse Z-transform.
Experiment 5.5 Inverse Z-Transform of X(z) = z-1
The python code computes the inverse Z-transform of X(z) = z-1, and the
corresponding output is shown in Fig. 5.6.
Inferences
1. From Fig. 5.6a, it is possible to infer that ‘IZT’ python command is used to obtain
the inverse Z-transform.
2. After executing the python code given in Fig. 5.6a, the result obtained is shown in
Fig. 5.6b. This result is in agreement with the theoretical result.
Task
1. Write a python code to obtain the inverse Z-transform of X(z) = z-4.
Experiment 5.6 Inverse Z-Transform of X(z) = 1/1 - z-1
The python code, which computes the inverse Z-transform of X(z) = 1/1 - z-1 and
the corresponding output, is shown in Fig. 5.7. From Fig. 5.7b, it is possible to
174 5 Transforms

#Inverse z-transform of z^(-1)

from lcapy import z
X=z**(-1)
x=X.IZT()
print(x) Piecewise((UnitImpulse(n - 1), n >= 0))

(a) Python Code (b) Simulation result

Fig. 5.6 Python code to obtain the inverse z-transform of X(z) = z-1 and its result. (a) Python code.
(b) Simulation result

#Inverse Z-transform
import sympy
import lcapy
from lcapy import z
X=1/(1-z**(-1))
x=X.IZT()
print(x) Piecewise((1, n >= 0))

(a) Python Code (b) Simulation result

Fig. 5.7 Python code to obtain the inverse Z-transform of X(z) = 1/1 - z-1 and its result. (a)
Python code (b) Simulation result

observe that the result obtained using python code is in agreement with the theoret-
ical result.
Inferences
The inverse Z-transform of X(z) = 1/1 - z-1 will be u[n], and the simulation result of
the python code given in Fig. 5.7a is shown in Fig. 5.7b. This result is in agreement
with the theoretical result.
Task
1. Write a python code to compute the inverse Z-transform of X(z) = 1/1 - z-2.

5.4 Family of Fourier Series and Transforms

Based on the nature of the signal, the Fourier family can be classified into Fourier
series or Fourier transform. Fourier series is an effective tool to analyse the periodic
signal. If the signal is aperiodic, Fourier transform can be used to analyse the signal.
Fourier transform can be viewed as the Fourier series when the period ‘T’ tends to
infinity. The Fourier transform is a generalization of the Fourier series representation
of functions. The Fourier series is limited to periodic functions, while Fourier
5.4 Family of Fourier Series and Transforms 175

Table 5.2 Family of Fourier series and transforms

Nature of the signal
Continuous/ Periodic/
S. No. discrete aperiodic Fourier family
1 Continuous Periodic Fourier series
2 Continuous Aperiodic Continuous-time Fourier transform (CTFT)
3 Discrete Periodic Discrete-time Fourier series
4 Discrete Aperiodic Discrete-time Fourier transform (DTFT)

Fig. 5.8 Spectrum of Signal Spectrum

continuous-time signal CTFT
x(t) X(Ω)

transform can be used for periodic and aperiodic functions. The family of Fourier
series and transform is given in Table 5.2.

5.4.1 Continuous-Time Fourier Transform (CTFT)

The continuous-time Fourier transform (CTFT) of the signal x(t) is represented as

X ð ΩÞ = xðt Þe - jΩt dt ð5:4Þ

-1

It can be interpreted as taking the inner product of the signal x(t) with the basis
function e-jΩt. This is represented as

XðΩÞ = hxðtÞ, e - jΩt i ð5:5Þ

Equations (5.4) and (5.5) are termed as ‘analysis equation’. The result of
continuous-time Fourier transform is termed as ‘spectrum’, which is illustrated in
Fig. 5.8. The equations reveal that how an arbitrary signal x(t) can be expanded as a
sum of elementary harmonic functions. The elementary harmonic functions are
termed as the basis function. The Fourier transform uses complex exponentials of
various frequencies as its basis function.
CTFT is a complex function of ‘Ω’ in the range - 1 < Ω < 1. CTFT exists if
the signal x(t) satisfies Dirichlet conditions, which are given below:
1. The signal x(t) has a finite number of discontinuities and a finite number of
maxima and minima in any finite interval.
2. The signal x(t) must be absolutely integrable, which is represented as
176 5 Transforms

Fig. 5.9 Forward and

inverse CTFT

1
jxðt Þjdt < 1.
-1

Inverse CTFT refers to obtaining the signal from the spectrum, which is also
called as ‘synthesis equation’. The inverse CTFT is given by

1
1
xð t Þ = X ðΩÞejΩt dΩ ð5:6Þ
2π
-1

From Eq. (5.6), it is possible to interpret that Fourier synthesis formula recon-
structs a signal using a set of scaled complex exponentials.
Analysis refers to the decomposition of the signal into its constituent components
specifying the weights of the basis functions in the expansion. Synthesis refers to the
reconstruction of the signal from the basis functions chosen to represent the signal.
The analysis and synthesis function of CTFT of the signal x(t) is illustrated in
Fig. 5.9.
(a) Forward Fourier Transform
The continuous-time Fourier transform of unit impulse signal is given by

δ ð ΩÞ = δðt Þe - jΩt dt ð5:7Þ

-1

Upon simplifying the above equation, we get

δðΩÞ = 1 ð5:8Þ
5.4 Family of Fourier Series and Transforms 177

From Eq. (5.8), it is possible to interpret that unit impulse contains a compo-
nent at every frequency. Another way to interpret the result is to make up δ(t);
one needs infinite number of equal frequency components.
(b) Inverse Fourier transform of δ(Ω)
The inverse CTFT of δ(Ω) is given by

1
1
xð t Þ = X ðΩÞejΩt dΩ ð5:9Þ
2π
-1

Substituting X(Ω) = δ(Ω) in the above expression, we get

1
-1 1
F fδðΩÞg = δðΩÞejΩt dΩ ð5:10Þ
2π
-1

Using the sampling property of the impulse signal, the above expression can
be simplified as

1
F - 1 f δ ð ΩÞ g = ð5:11Þ
2π

From the above expression, it is possible to interpret that Fourier transform of

a constant signal is

1 $ 2πδðΩÞ

Thus, Fourier transform of a DC signal results in an impulse signal in the

frequency domain.

Experiment 5.7 Computation of Forward CTFT of the Impulse Signal

and Inverse CTFT of the Resultant Forward CTFT
This experiment discusses the computation of forward CTFT of the impulse signal
and inverse CTFT of the resultant forward CTFT. The python code that obtains the
unit impulse signal spectrum is shown in Fig. 5.10, and the corresponding output is
in Fig. 5.11.
Inferences
From Fig. 5.11, it is possible to observe that Fourier transform of an impulse
function is a constant function in the frequency domain. The impulse function is a
compact function in time domain, whereas its spectrum exists in all frequencies.
Thus compression in time domain is equivalent to expansion in frequency domain
and vice versa.
178 5 Transforms

import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(-5,6,100)
w=np.linspace(-50,60,1000)
x=(t==0)
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('t-->'),plt.ylabel('$\u03B4[t]$'),plt.title('Input Signal')
y1=np.zeros(len(w))
for i in range(len(t)):
y=x[i]*np.exp(-1j*w*t[i])
y1=y1+y
y2=np.zeros(len(t))
for i in range(len(w)):
y3=y1[i]*np.exp(1j*w[i]*t)
y2=y2+y3
plt.subplot(2,2,2),plt.plot(w,np.abs(y1),linewidth=3),plt.title('Magnitude response')
plt.xlabel('$Ω$-->'),plt.ylabel('|\u03B4(${Ω}$|')
plt.subplot(2,2,3),plt.plot(w,np.angle(y1),linewidth=3),plt.title('Phase response')
plt.xlabel('$Ω$-->'),plt.ylabel('$\phi({Ω})$')
plt.subplot(2,2,4),plt.plot(t,y2/np.max(y2)),plt.xlabel('t-->'),plt.ylabel('$\u03B4[t]$'),
plt.title('Reconstructed Signal')
plt.tight_layout()

Fig. 5.10 Python code to obtain the spectrum of unit impulse signal

Input Signal Magnitude response

1.0 1.05
|δ(Ω)|

0.5 1.00
d [t]

0.0 0.95
–5 0 5 –50 0 50
t--> Ω-->

Phase response Reconstructed Signal

0.05 1.0
f (Ω)

0.00 0.5
d [t]

–0.05 0.0
–50 0 50 –5 0 5
Ω--> t-->

Fig. 5.11 Result of python code shown in Fig. 5.10

5.4 Family of Fourier Series and Transforms 179

Task
1. Modify the above code to obtain the Fourier transform of x(t) = δ(t - 5) and
comment on the observed result.

5.4.2 Fourier Transform of Standard Test Signals

This section focuses on obtaining the spectrum of standard test signals using CTFT.
The standard test signals include sinusoidal signal, Gaussian function and pulse
signal.
Experiment 5.8 CTFT of the Complex Exponential Signal (ejΩ0 t and e - jΩ0 t )
In this experiment, the objective is to obtain the spectrum of the signal ejΩ0 t and
e - jΩ0 t . Both the spectrum should be given an impulse corresponding to the fre-
quency ‘Ω0’. The python code which obtains the spectrum of the signals ejΩ0 t and
e - jΩ0 t is shown in Fig. 5.12 and the corresponding output is shown in Fig. 5.13. To
show the change in the spectrum between the signals ejΩ0 t and e - jΩ0 t , double-sided
spectrum is drawn instead of single-sided spectrum.

import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(-50,50,1000)
w=np.linspace(-5,5,100)
yy=np.exp(1j*(np.pi/4)*t)
xx=np.exp(-1j*(np.pi/4)*t)
plt.subplot(2,2,1),plt.plot(t,yy,linewidth=2),plt.xlabel('t-->'),plt.ylabel('x$_1$(t)')
plt.title('e$^{jΩot}$')
plt.subplot(2,2,2),plt.plot(t,xx,linewidth=2),plt.xlabel('t-->'),plt.ylabel('x$_2$(t)')
plt.title('e$^{-jΩot}$')
y1=np.zeros(len(w))
y2=np.zeros(len(w))
for i in range(len(t)):
y=yy[i]*np.exp(-1j*w*t[i])
y1=y1+y
z=xx[i]*np.exp(-1j*w*t[i])
y2=y2+z
plt.subplot(2,2,3),plt.plot(w,np.abs(y1)/len(t),linewidth=2)
plt.xlabel('$Ω$-->'),plt.ylabel('|X$_1$(${Ω}$|'),plt.title('Double sided Spectrum')
plt.subplot(2,2,4),plt.plot(w,np.abs(y2),linewidth=2)
plt.xlabel('$Ω$-->'),plt.ylabel('|X$_2$(${Ω}$)|'),plt.title('Double sided Spectrum')
plt.tight_layout()

Fig. 5.12 Python code to obtain the spectrum of complex exponential signal
180 5 Transforms

ejΩot e–jΩot
1 1
x1(t)

x2(t)
0 0

–1 –1
–50 –25 0 25 50 –50 –25 0 25 50
t--> t-->

Double sided Spectrum Double sided Spectrum

0.5 500
|X1(Ω)|

|X2(Ω)|

0.0 0
–5.0 –2.5 0.0 2.5 5.0 –5.0 –2.5 0.0 2.5 5.0
Ω--> Ω-->

Fig. 5.13 Spectrum of the complex exponential signals

Inferences
From Fig. 5.13, it is possible to observe that both the signals ejΩ0 t and e - jΩ0 t produce
single impulse at Ω = Ω0 and at Ω = - Ω0. In this case, the value of the frequency is
10 Hz; hence, it is possible to observe impulse at π=4 Hz and at - π=4 Hz, respectively,
for the signal ejΩ0 t and e - jΩ0 t .
Task
1. Obtain the CTFT of the signal xðt Þ = ejΩ0 t þ e - jΩ0 t and comment on the observed
result.
Experiment 5.9 Fourier Transform of x(t) = cos(Ωt)
According to Euler’s formula, the cos(Ωt) can be expressed as

ejΩt þ e - jΩt
cosðΩt Þ = ð5:12Þ
2

Hence, the signal x(t) is expressed as

1 jΩt
cosðΩt Þ = e þ e - jΩt ð5:13Þ
2

Taking Fourier transform on both sides, we get

5.4 Family of Fourier Series and Transforms 181

#Spectrum of Cosine wave

import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(-50,50,1000)
w=np.linspace(-5,5,1000)
yy=np.cos((np.pi/4)*t)
plt.subplot(2,1,1),plt.plot(t,yy,linewidth=1.5),plt.xlabel('t-->'),plt.ylabel('x$_1$(t)')
plt.title('cos(Ω$_o$t)')
y1=np.zeros(len(w))
for i in range(len(t)):
y=yy[i]*np.exp(-1j*w*t[i])
y1=y1+y
plt.subplot(2,1,2),plt.plot(w,np.abs(y1)/len(t),linewidth=1.5)
plt.xlabel('$Ω$-->'),plt.ylabel('|X$_1$(${Ω}$)|'),plt.title('Double sided Spectrum')
plt.tight_layout()

Fig. 5.14 Python code to obtain the spectrum of cosine wave

1
FTf cosðΩt Þg = FT ejΩt þ FT e - jΩt ð5:14Þ
2

From the previous example,

FT ejΩt = 2πδðΩ - Ω0 Þ ð5:15Þ

FT e - jΩt = 2πδðΩ þ Ω0 Þ ð5:16Þ

Substituting Eqs. (5.15) and (5.16) in Eq. (5.14), we get

1
FTf cosðΩt Þg = f2πδðΩ - Ω0 Þ þ 2πδðΩ þ Ω0 Þg ð5:17Þ
2

Simplifying the above expression, we get

FTf cosðΩt Þg = π fδðΩ - Ω0 Þ þ δðΩ þ Ω0 Þg ð5:18Þ

Thus, the spectrum of the cosine signal has two impulses placed symmetrically at
the frequency of the cosine and its negative.
The python code, which obtains the spectrum of cosine wave, is shown in
Fig. 5.14, and the corresponding output is shown in Fig. 5.15.
Inferences
From Fig. 5.15, it is possible to observe that the Fourier transform of the cosine
signal has two impulses placed symmetrically at the frequency of the cosine and its
negative which is in agreement with the theoretical result.
182 5 Transforms

cos(Ωot)
1
x1(t)

–1
–40 –20 0 20 40
t-->

Double sided Spectrum

0.4
|X1(Ω)|

0.2

0.0
–4 –2 0 2 4
Ω-->

Fig. 5.15 Result of python code shown in Fig. 5.14

Fig. 5.16 Representation of

the signal x(t) x(t) 1

-τ 0 τ t

Task
1. Write a python code to illustrate the fact that the magnitude spectrum of sine wave
and cosine wave of same amplitude, frequency and phase are alike.
1, jt j < τ
Experiment 5.10 CTFT of the Signal xðt Þ =
0, otherwise
The given signal is a rectangular pulse. It is shown in Fig. 5.16.
The expression for the CTFT of the signal x(t) is given by

X ð ΩÞ = xðt Þe - jΩt dt ð5:19Þ

-1

In this case, the signal exists from –τ to τ; hence, the limit of integration is
modified as
5.4 Family of Fourier Series and Transforms 183

X ð ΩÞ = xðt Þe - jΩt dt ð5:20Þ

-τ

In the limit –τ to τ, the value taken by the signal x(t) is one; hence, the above
equation can be expressed as

X ð ΩÞ = 1 × e - jΩt dt ð5:21Þ
-τ

Upon performing the integration, we get

τ
e - jΩt
X ð ΩÞ = ð5:22Þ
- jΩ -τ

Substituting the upper and lower limits in the above expression, we get

e - jΩτ - ejΩτ
X ð ΩÞ = ð5:23Þ
- jΩ

The above equation can be written as

ejΩτ - e - jΩτ
X ð ΩÞ =
jΩ

The above equation can be simplified as

sinðΩτÞ
X ð ΩÞ = 2
Ω

Multiplying and dividing the above equation by ‘τ’, we get

sinðΩτÞ
X ðΩÞ = 2τ = 2τ sin cðΩτÞ ð5:24Þ
Ωτ

From the above expression, it is possible to conclude that Fourier transform of a

rectangular function will result in a sinc function.
The objective is to write a python code to generate two rectangular functions with
different width. Pass these two rectangular functions through Fourier transform to
obtain their spectra. The python code, which generates two rectangular functions of
different width and their corresponding spectra, is shown in Fig. 5.17, and the
corresponding output is obtained in Fig. 5.18.
184 5 Transforms

import numpy as np
import matplotlib.pyplot as plt
t=np.linspace(-50,50,1000)
w=np.linspace(-5,5,100)
yy=(abs(t)<15)
xx=(abs(t)<2)
plt.subplot(2,2,1),plt.plot(t,yy,linewidth=1.5),plt.xlabel('t-->'),plt.ylabel('x$_1$(t)')
plt.title('Rectangular Function-1')
plt.subplot(2,2,2),plt.plot(t,xx,linewidth=1.5),plt.xlabel('t-->'),plt.ylabel('x$_2$(t)')
plt.title('Rectangular Function-2')
y1=np.zeros(len(w))
y2=np.zeros(len(w))
for i in range(len(t)):
y=yy[i]*np.exp(-1j*w*t[i])
y1=y1+y
z=xx[i]*np.exp(-1j*w*t[i])
y2=y2+z
plt.subplot(2,2,3),plt.plot(w,np.abs(y1)/len(t),linewidth=1.5)
plt.xlabel('$Ω$-->'),plt.ylabel('|X$_1$(${Ω}$)|'),plt.title('Spectrum-1')
plt.subplot(2,2,4),plt.plot(w,np.abs(y2),linewidth=1.5)
plt.xlabel('$Ω$-->'),plt.ylabel('|X$_2$(${Ω}$)|'),plt.title('Spectrum-2')
plt.tight_layout()

Fig. 5.17 Python code to obtain the spectrum of rectangular function

Rectangular Function-1 Rectangular Function-2

1.0 1.0
x1(t)

x2(t)

0.5 0.5

0.0 0.0
–50 –25 0 25 50 –50 –25 0 25 50
t--> t-->

Spectrum-1 Spectrum-2
40
0.2
|X1(Ω)|

|X2(Ω)|

20
0.1

0.0 0
–5.0 –2.5 0.0 2.5 5.0 –5.0 –2.5 0.0 2.5 5.0
Ω--> Ω-->

Fig. 5.18 Result of python code shown in Fig. 5.17

5.4 Family of Fourier Series and Transforms 185

Inferences
The following inferences can be obtained by observing Fig. 5.18:
1. Two rectangular functions, rectangular function-1 and rectangular function-2, are
generated.
2. The width of rectangular function-1 is larger than the width of rectangular
function-2.
3. The spectrum of the rectangular function is observed to be a sinc function.
4. The main lobe width of spectrum-1 is narrower when compared to the main lobe
width of spectrum-2.
5. This example illustrates the fact that compression in the time domain leads to
expansion in the frequency domain and vice versa.
Task
1. Instead of rectangular pulse, obtain the magnitude spectrum of triangular pulse
signal and comment on the observed result. Triangular pulse can be obtained by
convolving two rectangular pulse signals.
1, jΩj < Ω0
Experiment 5.11 Inverse CTFT of X ðΩÞ =
0, otherwise
The expression for inverse continuous-time Fourier transform is given by

1
1
xð t Þ = X ðΩÞejΩt dΩ ð5:25Þ
2π
-1

The spectrum exists from -Ω0 to Ω0; hence, the limit of integration has to be
changed. This is represented as

Ω0
1
xð t Þ = X ðΩÞejΩt dΩ ð5:26Þ
2π
- Ω0

In the interval from -Ω0 to Ω0, the value of the spectrum is unity. This is
expressed as

Ω0
1
xðt Þ = 1 × ejΩt dΩ ð5:27Þ
2π
- Ω0

Upon performing the integration, we get

186 5 Transforms

Ω0
1 ejΩt
xðt Þ = ð5:28Þ
2π jt - Ω0

Substituting the upper and lower limits, we get

1 ejΩ0 t - e - jΩ0 t
xð t Þ = ð5:29Þ
2π jt

The above equation can be written as

1 ejΩ0 t - e - jΩ0 t
xð t Þ = ð5:30Þ
πt 2j

The above equation can be expressed as

sinðΩ0 t Þ
xð t Þ = ð5:31Þ
πt

Thus, inverse Fourier transform of a rectangular function results in a sinc

function. Comparing this example with the previous example, it is possible to
write that rectangular function and sinc function are dual functions in Fourier
domain.
The aim of this experiment is to prove that rectangular and sinc functions are dual
functions in the Fourier domain. In the previous experiment, it is possible to prove
that Fourier transform of rectangular function results in sinc function. In this
experiment, the objective is to prove that Fourier transform of sinc function will
result in a rectangular function. Execute the python code given in Fig. 5.19 and enter
the number ‘1’. The simulation result of this python code is shown in Fig. 5.20.
Inferences
From Fig. 5.20, it is possible to observe that Fourier transform of sinc function
results in a rectangular function. Also, it is possible to infer that compression in one
domain (time) corresponds to expansion in another domain (frequency) and vice
versa.
Task
1. What is the reason for ringing effect observed in the magnitude spectrum of sinc
signal? Is there any way to minimize the ringing effect?
Experiment 5.12 CTFT of a Gaussian Function
The objective of this experiment is to prove that Fourier transform of a Gaussian
function results in a Gaussian function. The expression for Gaussian function with
mean μ and standard deviation σ is given by
5.4 Family of Fourier Series and Transforms 187

#Fourier transform of Sinc and Gaussian functions

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of sinc function
t=np.linspace(-5,5,1000)
w=np.linspace(-60,60,1000)
S=int(input("Enter : (1 = Sinc, 2 = Gaussian) : "));
if (S==1):
x1=np.sinc(t);
x2=np.sinc(2*t)
y1=np.zeros(len(w))
y2=y1;
for i in range(len(t)):
yx1=x1[i]*np.exp(-1j*w*t[i]);
yx2=x2[i]*np.exp(-1j*w*t[i]);
y1=y1+yx1;#Step 2: Spectrum of sinc function
y2=y2+yx2;#Step 2: Spectrum of sinc function
elif(S==2):
mu,sigma1,sigma2=0,0.1,0.5; #Mean and sigma values
x1=np.exp(-np.power(t - mu, 2.) / (2 * np.power(sigma1, 2.)));
x2=np.exp(-np.power(t - mu, 2.) / (2 * np.power(sigma2, 2.)));
y1=np.zeros(len(w))
y2=y1;
for i in range(len(t)):
yx1=x1[i]*np.exp(-1j*w*t[i]);
yx2=x2[i]*np.exp(-1j*w*t[i]);
y1=y1+yx1;#Step 2: Spectrum of Gaussian function
y2=y2+yx2;#Step 2: Spectrum of Gaussian function
else:
print('Please enter the correct number')
x1,x2,y1,y2=np.zeros(len(t)),np.zeros(len(t)),np.zeros(len(w)),np.zeros(len(w));
#Step 3: Plotting the results
plt.subplot(2,2,1),plt.plot(t,x1),plt.xlabel('t-->'),plt.ylabel('x$_1$(t)'),plt.title('Signal-1'),
plt.subplot(2,2,2),plt.plot(t,x2),plt.xlabel('t-->'),plt.ylabel('x$_2$(t)'), plt.title('Signal-2')
plt.subplot(2,2,3),plt.plot(w,np.abs(y1)),plt.xlabel('$Ω$-->'),plt.ylabel('|X$_1$(${Ω}$)|'),
plt.title('Spectrum of x$_1$(t)'),plt.subplot(2,2,4),plt.plot(w,np.abs(y2)), plt.xlabel('$Ω$-->'),
plt.ylabel('|X$_2$(${Ω}$)|'),plt.title('Spectrum of x$_2$(t)')
plt.tight_layout()

Fig. 5.19 Python code to obtain the spectrum of sinc and Gaussian function

1 ðt - μÞ2
xðt Þ = p e - 2σ2 ð5:32Þ
σ 2π

If the mean of the Gaussian function is zero, the above expression is given by

1 t2
xðt Þ = p e - 2σ2 ð5:33Þ
σ 2π

Differentiating both sides with respect to t, we get

188 5 Transforms

Signal-1 Signal-2
1.0 1.0

0.5 0.5
x1(t)

x2(t)
0.0 0.0

–5.0 –2.5 0.0 2.5 5.0 –5.0 –2.5 0.0 2.5 5.0
t--> t-->

Spectrum of x1(t) Spectrum of x2(t)

100
40

|X2(Ω)|
|X1(Ω)|

50
20

0 0
–50 0 50 –50 0 50
Ω--> Ω-->

Fig. 5.20 Result of python code shown in Fig. 5.19

dxðt Þ 1 t2 - 2t
= p e - 2σ2 × ð5:34Þ
dt σ 2π 2σ 2

The above equation can be simplified as

dxðt Þ 1 t2 -t
= p e - 2σ2 × 2 ð5:35Þ
dt σ 2π σ

Substituting Eq. (5.32) in Eq. (5.35), we get

dxðt Þ -t
= xð t Þ × 2 ð5:36Þ
dt σ

The above equation can be rearranged as

dxðt Þ 1
= - 2 txðt Þ ð5:37Þ
dt σ

Taking Fourier transform on both sides, we get

5.4 Family of Fourier Series and Transforms 189

dxðt Þ 1
FT = - 2 FTftxðt Þg ð5:38Þ
dt σ

Using the following fact

dxðt Þ
FT = jΩX ðΩÞ ð5:39Þ
dt
dX ðΩÞ
FTftxðt Þg = j ð5:40Þ
dΩ

Substituting Eqs. (5.39) and (5.40) in Eq. (5.38), we get

1 dX ðΩÞ
jΩX ðΩÞ = - ×j ð5:41Þ
σ2 dΩ

Simplifying the above expression, we get

1 dX ðΩÞ
ΩX ðΩÞ = - × ð5:42Þ
σ2 dΩ

Upon rearranging the terms, we get

dX ðΩÞ
dΩ
= - σ2Ω ð5:43Þ
X ð ΩÞ

Taking integral on both sides, we get

dX ðΩÞ
dΩ
= - σ2 ΩdΩ ð5:44Þ
X ð ΩÞ

Upon integration, we get

Ω2
ln½X ðΩÞ] - ln½X ð0Þ] = - σ 2 ð5:45Þ
2

Since the mean value of the Gaussian signal is assumed to be zero, the above
equation can be written as

Ω2
ln½X ðΩÞ] = - σ 2 ð5:46Þ
2

Taking exponential on both sides, we get

190 5 Transforms

Signal-1 Signal-2
1.0

0.5
x1(t)

x2(t)
0.5

0.0 0.0
–5.0 –2.5 0.0 2.5 5.0 –5.0 –2.5 0.0 2.5 5.0
t--> t-->

Spectrum of x1(t) Spectrum of x2(t)

2.4 20
|X1(Ω)|

|X2(Ω)| 10
2.2
0
–50 0 50 –50 0 50
Ω--> Ω-->

Fig. 5.21 Spectra of Gaussian functions

σ 2 Ω2
X ð ΩÞ = e - 2 ð5:47Þ

From the above expression, it is possible to interpret that Fourier transform of a

Gaussian function results in a Gaussian function.
In this experiment, two Gaussian signals are generated with zero mean and
standard deviation as σ 1 = 0.01 and σ 2 = 0.1, respectively. After generating the
two Gaussian signals, their spectra are obtained by taking the Fourier transform. The
python code, which does this task, is shown in Fig. 5.19. After executing this code,
enter the number ‘2’. The simulation result is shown in Fig. 5.21.
Inference
1. From Fig. 5.21, it is possible to observe that two Gaussian functions with zero
mean and standard deviation σ 1 = 0.01 and σ 2 = 0.1 are generated.
2. Gaussian function-1 (x1) has a narrow spread, whereas Gaussian function-2 (x2)
has a wider spread.
3. Upon obtaining the spectra, it is possible to infer the fact that if the signal spread
is narrow in time domain (Gaussian function-1), the corresponding spectrum has
wide spread (spectrum of Gaussian function-1).
4. On the other hand, if the Gaussian function has wide spread in time domain
(Gaussian function-2), its spectrum is narrower (spectrum of Gaussian function-
2).
5. This illustrates the fact that ‘Compression in one domain leads to expansion in
another domain and vice-versa’.
5.4 Family of Fourier Series and Transforms 191

Task
1. In the above experiment, Signal-1 and Signal-2 are Gaussian functions. Now
multiply Signal-1 and Signal-2 to obtain Signal-3. Obtain the spectrum of Signal-
3 and comment on the observed result.

5.4.3 Discrete-Time Fourier Transform (DTFT)

Discrete-time Fourier transform is a transformation that maps the discrete-time

signal into a complex valued function, which is given by

1
X ejω = x½n]e - jωn ð5:48Þ
n= -1

DTFT is a way to represent the frequency content of discrete-time signal.

The magnitude and phase form of DTFT representation is given by

X ejω = X ejω ejΦðe Þ

jω
ð5:49Þ

In the above expression, |X(ejω)| represents the magnitude of DTFT, and Φ(ejω)
represents the phase of DTFT. The magnitude spectrum determines the relative
presence of a sinusoid in the signal x[n], whereas the phase spectrum determines
how the sinusoids line up relative to one another to form the signal x[n]. The
condition for the existence of DTFT is that the signal x[n] should be absolutely
summable. The signal x[n] is absolutely summable if it obeys the following
condition:

1
jx½n]j < 1 ð5:50Þ
n= -1

The expression inverse discrete-time Fourier transform (IDTFT) expression is

given by

π
1
x ½ n] = X ejω ejωn dω ð5:51Þ
2π
-π

j nj < N
1,
Experiment 5.13 DTFT of x½n] =
0, otherwise
The signal x[n] represents a rectangular pulse. The DTFT of x[n] is given by
192 5 Transforms

# Python code for DTFT of rectangular pulse signal

import numpy as np
import matplotlib.pyplot as plt
n=np.arange(-5,6)
w=np.arange(-3*np.pi,3*np.pi,0.1)
x=(n>=0)
y=(n<=3)
z=x*y
plt.subplot(3,1,1),plt.stem(n,z),plt.xlabel('n-->'),plt.ylabel('x[n]')
y1=np.zeros(len(w))
for i in range(len(n)):
y=z[i]*np.exp(-1j*w*n[i])
y1=y1+y
print(y1)
plt.subplot(3,1,2),plt.plot(w,np.abs(y1),linewidth=3),plt.title('Magnitude response')
plt.xlabel('$\omega$-->'),plt.ylabel('|X(${j\omega)}$|')
plt.subplot(3,1,3),plt.plot(w,np.angle(y1),linewidth=3),plt.title('Phase response')

Fig. 5.22 Python code for Experiment 5.13

N
X ejω = 1 × e - jωn
n= -N

aNþ1 - a - N
N
a = , jaj < 1 , the above
Using the summation formula n
a-1
n= -N 2N þ 1, a = 1
equation can be written as

e - jωðNþ1Þ - ejωN
X ejω = , e - jω < 1
e - jω - 1

The above equation can be simplified as

e - jωN e - jω - ejωN
X ejω =
e - jω - 1

The above equation can be written as

sin N þ 12 ω
X ejω = if e - jω < 1
sin ω2

This shows that Fourier transform of a rectangular pulse signal will result in a sinc
function.
Python code for the DTFT of rectangular pulse is given in Fig. 5.22, and its
corresponding output is shown in Fig. 5.23.
5.4 Family of Fourier Series and Transforms 193

x[n]
0
–4 –2 0 2 4
n-->
Magnitude response
|X(jw )|

2.5
0.0
–10.0 –7.5 –5.0 –2.5 0.0 2.5 5.0 7.5 10.0
w -->
Phase response
2.5
f (jw )

0.0
–2.5
–10.0 –7.5 –5.0 –2.5 0.0 2.5 5.0 7.5 10.0
w -->

Fig. 5.23 Simulation result of python code given in Fig. 5.22

Fig. 5.24 Discrete-time h[n]

LTI system
x[n] DT-LTI y[n]
X(e jω) System Y(e jω)
H(e jω)

Inferences
From Fig. 5.23, it is possible to infer that the magnitude response of a rectangular
function is a sinc function, which is in agreement with the theoretical result.
Task
1. In the above experiment, x[n] is a rectangular pulse signal. What will be the
impact of increasing the width of the signal x[n] in the magnitude and phase
responses?

5.4.4 Analysis of Discrete-Time LTI System Using DTFT

The block diagram of discrete-time LTI system with the input signal x[n], impulse
response h[n] and the output signal y[n] is shown in Fig. 5.24.
The relationship between the input and output of the system if it is LTI is given by

y½n] = x½n] * h½n] ð5:52Þ

Upon taking DTFT on both sides of the above equation, we get

194 5 Transforms

Y ejω = X ejω H ejω ð5:53Þ

Equation (5.53) is obtained using the fact that convolution in time domain is
equivalent to multiplication in the Fourier domain.
The frequency response of the system from Eq. (5.53) can be expressed as

Y ðejω Þ
H ejω = ð5:54Þ
X ðejω Þ

The frequency response of the system is a combination of magnitude and phase

responses. This is expressed as

H ejω = H ejω ejϕðe Þ

jω
ð5:55Þ

The frequency response defines how a complex exponential is changed in

amplitude and phase by a system.
Experiment 5.14 Computation of the Magnitude and Phase Responses
of Discrete-Time System Using DTFT
This experiment discusses the computation of magnitude and phase responses of DT
system using DTFT. Let us consider two discrete-time systems and its impulse
responses given by h1 ½n] = 12 , 12 and h2 ½n] = 12 , - 12 . The python code, which
obtains the magnitude and phase responses of the two systems, is shown in Fig. 5.25,
and the corresponding output is shown in Fig. 5.26.
Inferences
From Fig. 5.26, the following inferences can be made:
1. The magnitude response of system-1 shows that the system behaves like a
lowpass filter.
2. The magnitude response of system-2 shows that the system behaves like a
highpass filter.
3. The phase responses of both these systems reveal that both systems exhibit linear
phase characteristics.
4. The response of the two systems is in agreement with the theoretical result.
Task
1. From the magnitude response, it is possible to observe that the roll-off rate is not
sharp? What has to be done to improve the roll-off rate?
5.4 Family of Fourier Series and Transforms 195

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Impulse response of the two systems
h1=[0.5,0.5]
h2=[0.5,-0.5]
#Step 2: Obtaining the frequency response
w1, H1 = signal.freqz(h1,1)
w2, H2 = signal.freqz(h2,1)
angle_1 = np.unwrap(np.angle(H1))
angle_2 = np.unwrap(np.angle(H2))
#Step3 : Plotting the responses
plt.subplot(2,3,1),plt.stem(h1),plt.xlabel('n-->'),plt.ylabel('h$_1$[n]')
plt.title('Impulse response'),plt.subplot(2,3,2),plt.plot(w1, 10 * np.log10(abs(H1)))
plt.xlabel('$\omega$-->'),plt.ylabel('|X(${j\omega}$)|'),plt.title('Magnitude response')
plt.subplot(2,3,3),plt.plot(w1,(angle_1)),plt.xlabel('$\omega$--
>'),plt.ylabel('$\phi({j\omega})$')
plt.title('Phase response'),plt.subplot(2,3,4),plt.stem(h2),plt.xlabel('n-->'),
plt.ylabel('h$_2$[n]'),plt.title('Impulse response'),plt.subplot(2,3,5),
plt.plot(w1, 10 * np.log10(abs(H2))),plt.xlabel('$\omega$-->'),
plt.ylabel('|X(${j\omega}$)|'),plt.title('Magnitude response')
plt.subplot(2,3,6),plt.plot(w1,(angle_2)),plt.xlabel('$\omega$-->'),
plt.ylabel('$\phi({j\omega})$'),plt.title('Phase response')
plt.tight_layout()

Fig. 5.25 Python code to obtain the magnitude and phase response of the systems

Impulse response Magnitude response Phase response

0 0
0.4
|X(jw)|

–10
h1[n]

f (jw)

0.2 –1
–20
0.0
0 1 0 2 0 2
n--> w--> w-->

Impulse response Magnitude response Phase response

0.5 0

1
|X(jw)|

–10
f (jw)
h2[n]

0.0

–20
–0.5 0
0 1 0 2 0 2
n--> w--> w-->

Fig. 5.26 Responses of the two systems

196 5 Transforms

5.4.5 Discrete Fourier Transform

Discrete Fourier transform (DFT) represents a signal in terms of sinusoids. For a

discrete-time signal of length N, the basis functions are sinusoids of length N.
Discrete Fourier transform is used when the signal is discrete-time and periodic
only. In practice, it calculates the frequency domain representation of aperiodic
signals in a given time interval, by assuming their periodic extension. The discrete
Fourier transform (DFT) of the signal x[n] of length N is given by

N -1
x½n]e - j N kn , k = 0, 1, 2, . . . , N - 1
2π
X ½k ] = ð5:56Þ
n=0

The above expression can be written in the form of

N -1
X ½k ] = x½n]W kn
N ð5:57Þ
n=0

where

WN = e - jN
2π
ð5:58Þ

The signal is reconstructed by using the inverse discrete Fourier transform, which
is defined as

N -1
1 2π
x½n] = X ½k]ej N kn ð5:59Þ
N k=0

The forward transform is generally known as ‘analysis’, and the inverse transform
is called as ‘synthesis’.
Experiment 5.15 Plotting the Twiddle Factor for N = 8
The aim of this experiment is to plot the twiddle factor of DFT with the length N = 8.
The python code, which plots the twiddle factor or phase factor for N = 8, is shown
in Fig. 5.27, and the corresponding output is shown in Fig. 5.28.
Inference
From Fig. 5.28, it is possible to observe that for the choice of N = 8, the unit circle is
divided into eight equal portions.

(a) DFT matrix

The DFT matrix of order N is given by
5.4 Family of Fourier Series and Transforms 197

Fig. 5.27 Python code to

plot the twiddle factor for import numpy as np
N=8 import matplotlib.pyplot as plt
n=8
for k in range(0,n):
z=np.exp(2*np.pi*1j*k/n)
plt.plot([0,np.real(z)],[0,np.imag(z)])
x=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(x),np.sin(x),color='gray')
plt.axis('square')

Fig. 5.28 Result of python

1.00
code shown in Fig. 5.27
0.75

0.50

0.25

0.00

–0.25

–0.50

–0.75

–1.00
–1.0 –0.5 0.0 0.5 1.0

1 1 1 ⋯ 1
1.ðN - 1Þ
1 W 1.1
N W 1.2
N ⋯ WN
2.ðN - 1Þ
WN = 1 W 2.1
N W 2.2
N ⋯ WN ð5:60Þ
⋮ ⋮ ⋮ ⋯ ⋮
ðN - 1Þ.1 ðN - 1Þ.2 ðN - 1Þ.ðN - 1Þ
1 WN WN ⋯ WN

Substituting N = 2, the DFT matrix of order 2 is given by

1 1
W2 = ð5:61Þ
1 -1

Substituting N = 4, the DFT matrix of order 4 is given by

198 5 Transforms

Fig. 5.29 Python code to

obtain 4 point DFT matrix import numpy as np
np.set_printoptions(precision=2, suppress=True)
N=4;
n=np.arange(0,N,1)
k1=np.outer(n, n)
D=np.exp(-1j*2*np.pi*k1/N)
print('{} point DFT Matrix'.format(N))
print(D)

Fig. 5.30 4 point DFT

matrix 4 point DFT Matrix

[[ 1.+0.j 1.+0.j 1.+0.j 1.+0.j]

[ 1.+0.j 0.-1.j -1.-0.j -0.+1.j]

[ 1.+0.j -1.-0.j 1.+0.j -1.-0.j]

[ 1.+0.j -0.+1.j -1.-0.j 0.-1.j]]

1 1 1 1
1 -j -1 j
W4 = ð5:62Þ
1 -1 1 -1
1 j -1 -j

Experiment 5.16 Computation of 4 Point DFT Matrix

This experiment deals with the computation of 4 point DFT matrix using python.
The python code of 4 point DFT matrix generation is given in Fig. 5.29, and its
corresponding simulation output is shown in Fig. 5.30.
The built-in function dft available in scipy.linalg can be used to obtain the DFT
matrix. The python code to obtain the DFT matrix of order N = 4 is shown in
Fig. 5.31a, and the corresponding output is shown in Fig. 5.31b.
Inferences
The python code to compute the 4 point DFT matrix is shown in Fig. 5.29. From this
figure, it is possible to observe that the DFT computation formula is implemented in
python. Also, it is possible to see that ‘np.outer’ python command is used to generate
outer product of two vectors (1 for time index (n) and other for frequency index (k)).
From this Fig. 5.31, it is possible to observe that dft python command called from the
linalg library package to compute the DFT matrix. In both these methods, the
simulation result is in agreement with the theoretical result.
5.4 Family of Fourier Series and Transforms 199

from scipy.linalg import dft [[ 1.+0.j 1.+0.j 1.+0.j 1.+0.j]

import numpy as np
[ 1.+0.j 0.-1.j -1.-0.j -0.+1.j]
np.set_printoptions(precision=2, suppress=True)
N=4 [ 1.+0.j -1.-0.j 1.+0.j -1.-0.j]
W= dft(N)
print(W) [ 1.+0.j -0.+1.j -1.-0.j 0.-1.j]]

(a) Python code (b) Simulation result

Fig. 5.31 Python code to obtain the DFT matrix and its result. (a) Python code. (b) Simulation
result

from scipy.linalg import dft

from numpy.linalg import matrix_rank
import numpy as np
np.set_printoptions(precision=2, suppress=True)
N=4
W= dft(N)
rank=matrix_rank(W)
print("Rank of the matrix is {}".format(rank)) Rank of the matrix is 4

(a) Python code (b) Simulation result

Fig. 5.32 Python code to compute the rank of the matrix and its result. (a) Python code. (b)
Simulation result

Experiment 5.17 Computation of the Rank of DFT Matrix of Order 4

The python code to obtain the rank of 4 × 4 DFT matrix is shown in Fig. 5.32a, and
the corresponding output is shown in Fig. 5.32b.
Inference
From Fig. 5.32b, it is possible to observe that the rank of 4 × 4 DFT matrix is
4. Similarly, it is possible to prove that the rank of N × N DFT matrix is N. Full rank
of the matrix indicates that all rows and columns are linearly independent.
Task
1. Investigate the nature of DFT matrix. Find whether the DFT matrix is unitary
or not.
Experiment 5.18 Computattion of DFT of a Sequence Using DFT Matrix
This experiment deals with the computation of DFT of a input sequence using DFT
matrix. Let us consider the input sequence x[n] = {1, 1, 1, 1}. The python code to
obtain 4 point DFT of the sequence x[n] is given in Fig. 5.33, and the corresponding
output obtained is {4, 0, 0, 0}.
200 5 Transforms

Fig. 5.33 Python code to

obtain DFT of a sequence from scipy.linalg import dft
using DFT matrix import numpy as np
np.set_printoptions(precision=2, suppress=True)
N=4
x=np.ones(N)
W= dft(N)
X=W@x # matrix multiplication
print(np.abs(X))

Fig. 5.34 Python code to

compute inverse DFT using from scipy.linalg import dft
DFT matrix import numpy as np
np.set_printoptions(precision=2, suppress=True)
N=4
X=[4,0,0,0]
W= dft(N)
x=(W@X)/N
print(np.abs(x))

Inference
The signal x[n] is a DC signal, which is given by x[n] = {1, 1, 1, 1}. The DFT of the
sequence x[n] is obtained as X[k] = {4, 0, 0, 0}. The maximum energy of the
sequence x[n] is packed into one transform coefficient. This is through energy
compaction property of DFT. In Fig. 5.33, the symbol (‘@’) denotes matrix multi-
plication in python.
Experiment 5.19 Computation of Inverse DFT Through DFT Matrix
Let us consider the DFT coefficients X[k] = {4, 0, 0, 0}. The python code to compute
the inverse DFT of X[k] = {4, 0, 0, 0} is given in Fig. 5.34. After executing this code,
we get the result of {1, 1, 1, 1}, which is in agreement with the theoretical result.
Inference
The inverse DFT of X[k] = {4, 0, 0, 0} is x[n] = {1, 1, 1, 1}. The experimental result
is in agreement with the fact that DFT is invertible. This is to inform that the inverse
DFT computation is done by the forward DFT matrix, which can be seen in
Fig. 5.34.

5.4.6 Properties of DFT

Discrete Fourier transform is applied to discrete-time signal x[n] that are zero for
n < 0 and n ≥ N. However, the discrete-time signal x[n] must be considered a
periodic signal. Therefore, some of the DFT properties are based on modulo N or
5.4 Family of Fourier Series and Transforms 201

#Linearity property of DFT

import numpy as np
import matplotlib.pyplot as plt
x1=[1,1,1,1,1,1,1,1]
x2=[1,-1,1,-1,1,-1,1,-1]
a,b=5,10
x3=np.add(np.multiply(a,x1),np.multiply(b,x2))
N=len(x2)
n=np.arange(0,N,1)
k=np.arange(0,N,1)
k1=np.outer(n, k)
D=np.exp(-1j*2*np.pi*k1/N)#DFT matrix
X1=np.dot(D,x1)
X2=np.dot(D,x2)
X3=np.dot(D,x3)
X4=a*X1+b*X2
plt.subplot(2,2,1),plt.stem(n,x1),plt.xlabel('n-->'),plt.ylabel('x$_1$[n]')
plt.title('x$_1$[n]'), plt.subplot(2,2,2),plt.stem(n,x2),plt.xlabel('n-->'),
plt.ylabel('x$_2$[n]'),plt.title('x$_2$[n]')
plt.subplot(2,2,3), plt.stem(k,X1),plt.xlabel('k-->'),plt.ylabel('X$_1$[k]')
plt.title('X$_1$[k]'),plt.subplot(2,2,4),plt.stem(k,X2),plt.xlabel('k-->'),
plt.ylabel('X$_2$[k]'),plt.title('X$_2$[k]'),plt.tight_layout()
plt.figure(2),plt.subplot(3,1,1),plt.stem(n,x3),plt.xlabel('n-->'),
plt.ylabel('x$_3$[n]'),plt.title('x$_3$[n]=a*x$_1$[n]+b*x$_2$[n]')
plt.subplot(3,1,2),plt.stem(k,X3),plt.xlabel('k-->'),plt.ylabel('X$_3$[k]')
plt.title('X$_3$[k]'),plt.subplot(3,1,3),plt.stem(k,X4),plt.xlabel('k-->'),
plt.ylabel('X$_4$[k]'),plt.title('X$_4$[k]=a*X$_1$[k]+b*X$_2$[k]')
plt.tight_layout()

Fig. 5.35 Python code for linearity property

mod N operation. The modulo operation yields a division’s remainder or signed

remainder after dividing one number by another. For example, (5 mod 2), the result
will be ‘1’ (i.e. remainder value getting 5 divided by 2).
Experiment 5.20 Verification of the Linearity Property of DFT
The DFT of a linear combination of two sequences is the linear combination of the
DFT of the individual sequences. The DFT property is given by

DFTfαx1 ½n] þ βx2 ½n]g = αX 1 ½k ] þ βX 2 ½k ] ð5:63Þ

The python code, which verifies the linearity property of DFT, is shown in
Fig. 5.35, and the corresponding output is shown in Fig. 5.36.
Inferences
The following inferences can be made from this experiment:
202 5 Transforms

x1[n] x2[n] x3[n]=a*x1[n]+b*x2[n]

1.0 1

x3[n]
10
0
x1[n]

x2[n]
0.5 0 0 1 2 3 4 5 6 7
n-->
0.0 –1 X3[k]
0 2 4 6 0 2 4 6

X3[k]
n--> n--> 50
X1[k] X2[k] 0
0 1 2 3 4 5 6 7
k-->
X1[k]

X2[k]
5 5 X4[k]=a*X1[k]+b*X2[k]

X4[k]
50
0 0 0
0 2 4 6 0 2 4 6 0 1 2 3 4 5 6 7
k--> k--> k-->

Fig. 5.36 Verification of the linearity property of DFT

1. From Fig. 5.35, the input signals chosen to prove the linearity property of DFT
are x1[n] = {1, 1, 1, 1, 1, 1, 1}, which is a DC signal, and x2[n] = {1, -1, 1, -1, -
1, 1, -1} which is an AC signal. The scaling factors chosen in this example are
a = 5, b = 10.
2. From Fig. 5.36, it is possible to observe the following facts:
(a) DFT of the DC signal x1[n] is represented as X1[k], which exhibits peak at
k = 0. DFT of the signal x2[n] shows the peak at k = 4.
(b) The DFT of ax1[n] + bx2[n] is equal to aX1[k] + bX2[k]. This implies that DFT
obeys homogeneity and additivity properties; hence, it is a linear transform.
Experiment 5.21 Verification of Circular Shift Property of DFT
The circular shift property of DFT is expressed as

DFT
x½ðn - mÞmodN ] $ e - j N km X ½k]
2π
ð5:64Þ

The python code to illustrate the circular shift property of DFT is shown in
Fig. 5.37, and the corresponding output is shown in Fig. 5.38.
Inferences
After executing the python code given in Fig. 5.37, the result of input signal x[n] is
{4, 3, 2, 1}, and the circularly shifted sequence with k = 2 is obtained as {2, 1, 4, 3}.
The DFT of the input sequence x[n] and its circularly shifted version is shown in
Fig. 5.38. From this figure, it is possible to confirm that the magnitude spectrum of
both sequences is the same, whereas the phase spectrum is different. This indicates
that ‘time shift in the time domain corresponds to phase shift in the frequency
domain’.
Experiment 5.22 Verification of the Parseval’s Relationship of DFT
According to Parseval’s relation, energy in time domain is equivalent to energy in
frequency domain.
5.4 Family of Fourier Series and Transforms 203

# Python code for circular shift property

import numpy as np
import matplotlib.pyplot as plt
x=[4,3,2,1]
z=[]
k=2# Circular shifting factor
for i in range(len(x)):
m=np.mod((i-k),len(x))
y=x[m]
z=np.append(z,y)
N=len(x)
n=np.arange(0,N,1)
k=np.arange(0,N,1)
k1=np.outer(n, k)
D=np.exp(-1j*2*np.pi*k1/N)#DFT matrix
X1=np.dot(D,x) # DFT computation of x[n]
X2=np.dot(D,z) # DFT computation of x[n-k]
plt.figure(1),plt.subplot(3,1,1),plt.stem(n,x),plt.xlabel('n-->')
plt.ylabel('x[n]'),plt.title('Input sequence')
plt.subplot(3,1,2),plt.stem(k,np.abs(X1)),plt.xlabel('k-->')
plt.ylabel('|X[k]|'),plt.title('Magnitude Response')
plt.subplot(3,1,3),plt.stem(k,(np.angle(X1))),plt.xlabel('k-->')
plt.ylabel('$\phi$[k]'),plt.title('Phase Response'),plt.tight_layout()
plt.figure(2),plt.subplot(3,1,1),plt.stem(n,z),plt.xlabel('n-->')
plt.ylabel('z[n]'),plt.title('Circularly shifted Sequence')
plt.subplot(3,1,2),plt.stem(k,np.abs(X2)),plt.xlabel('k-->')
plt.ylabel('|Z[k]|'),plt.title('Magnitude Response')
plt.subplot(3,1,3),plt.stem(k,(np.angle(X2))),plt.xlabel('k-->')
plt.ylabel('$\phi$[k]'),plt.title('Phase Response'),plt.tight_layout()

Fig. 5.37 Python code to illustrate circular shifting property of DFT

Input sequence Circularly shifted sequence

2.5
x[n]

2.5
z[n]

0.0 0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
n--> n-->
Magnitude response Magnitude response
10 10
|x[k]|

|z[k]|

0 0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
k--> k-->
Phase response Phase response
0.5 2.5
f [k]
f [k]

0.0 0.0
–0.5 –2.5
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
k--> k-->

Fig. 5.38 Simulation result of Python code given in Fig. 5.37

204 5 Transforms

Fig. 5.39 Python code for

Parseval’s relation # Python code for Parseval's relation
import numpy as np
np.set_printoptions(precision=2, suppress=True)
x=[1,2,3,4];
N=len(x);
n=np.arange(0,N,1);
k=np.arange(0,N,1);
k1=np.outer(n, k)
D=np.exp(-1j*2*np.pi*k1/N)# DFT matrix
print('Input Sequence x[n]: ', x)
y=np.sum((np.abs(x)**2))
print('\u03A3|x[n]|^2: ', y)
Y1=np.dot(D,x) # DFT Computation
print('DFT output X[k]: ', Y1)
Y=np.sum((np.abs(Y1)**2))/N
print('\u03A3|X[k]|^2: ', Y)

Fig. 5.40 Result of

Parseval’s relation property Input Sequence x[n]: [1, 2, 3, 4]
Σ|x[n]|^2: 30
DFT output X[k]: [10.+0.j -2.+2.j -2.-0.j -2.-2.j]
Σ|X[k]|^2: 30.0

N -1 N -1
1
jx½n]j2 = j X ½ k ]j 2 ð5:65Þ
n=0
N k=0

The python code to verify the Parseval’s relation of DFT is given in Fig. 5.39, and
its simulation output is shown in Fig. 5.40.
Inference
From Fig. 5.40, it is possible to confirm that the energy in time domain and the
energy in the frequency domain are the same. Hence, converting the time domain
signal into the frequency domain using DFT always preserves energy. Therefore, the
perfect reconstruction of the original signal from the frequency components is
possible for the DFT.

5.4.7 Limitations of Fourier Transform

The basis function of Fourier transform is a complex exponential, which oscillates

for all the time. This means that Fourier transform describes frequency components
in the signal averaged over all the time. It is difficult for the Fourier transform to
represent signals that are localized in time. Hence, Fourier transform is not an
5.4 Family of Fourier Series and Transforms 205

Sinusodial signal of Sinusodial signal of

10 Hz frequency 10.5 Hz frequency

Fourier transform

Magnitude spectrum-1 Magnitude spectrum-2

Fig. 5.41 Block diagram of problem statement

effective tool to analyse non-stationary signals. To overcome this drawback, time

localization in Fourier transform can be achieved by windowing the signal over
which the signal is nearly stationary, which leads to the development of short-time
Fourier transform (STFT).
Experiment 5.23 Limitation of Fourier Transform
The objective of this experiment is to prove that Fourier transform cannot estimate
fractional frequencies. Fourier transform of signal with fractional frequencies results
in the spreading of the spectrum to other frequencies, which are not present in the
signal. This fact is verified in this experiment. In this experiment, two sinusoidal
signals of frequency 10 and 10.5 Hz are generated, and the magnitude spectrum of
the generated signals is obtained by taking the Fourier transform of the generated
signals. This is shown in Fig. 5.41.
The python code, which performs this task, is shown in Fig. 5.42, and the
corresponding output is shown in Fig. 5.43.
Inference
From Fig. 5.43, the following inference can be drawn:
1. Fourier transform of 10 Hz sinusoidal signal has a peak exactly at 10 Hz.
2. Fourier transform of 10.5 Hz frequency component sinusoidal signal does not
show peak at 10.5 Hz. Instead, it resulted in spreading of the spectrum to other
frequencies. To avoid the spreading of the spectrum to other frequencies, the
value of N has to be increased.
Task
1. Increase the value of N in the python code shown in Fig. 5.42 from 100 to 256 and
512 and comment on the observed output.
206 5 Transforms

#Fourier transform of fractional frequency component

import numpy as np
import matplotlib.pyplot as plt
from scipy.fft import fft,fftfreq
# Step 1: Signal generation
fs=100
T=1/fs
f1=10 #10 Hz frequency component
f2=10.5 #10.5 Hz frequency component
N=100
t=np.linspace(0,N*T,N)
x1=np.sin(2*np.pi*f1*t)
x2=np.sin(2*np.pi*f2*t)
#Step 2: Spectrum of the signals
faxis=fftfreq(N,T)[0:N//2]
X1=fft(x1)
X2=fft(x2)
#Step 3: Ploting the result
plt.subplot(2,1,1),plt.stem(faxis,2/N*np.abs(X1)[0:N//2])
plt.xlabel('Frequency (Hz)'),plt.ylabel('Magnitude'),
plt.title('Spectrum of 10 Hz sine wave')
plt.subplot(2,1,2),plt.stem(faxis,2/N*np.abs(X2)[0:N//2])
plt.xlabel('Frequency (Hz)'),plt.ylabel('Magnitude'),
plt.title('Spectrum of 10.5 Hz sine wave')
plt.tight_layout()

Fig. 5.42 Fourier transform of fractional frequency component signal

5.5 Discrete Cosine Transform (DCT)

Discrete cosine transform was developed by Ahmed, Rao and Natarajan in the year
1974. DCT is a unitary transform, and it is not a discrete version of the cosine
functions. The DCT has better energy compaction than DFT; hence, it is widely used
in signal compression. DCT is employed in JPEG compression standard. DCT is
based on the DFT with imposed even symmetry through reflection; hence, DCT is a
real-valued transform.
The formula to compute forward discrete cosine transform is given by

N -1
ð2n þ 1Þπk
X ½k ] = αðkÞ x½n] cos ð5:66Þ
n=0
2N

where
5.5 Discrete Cosine Transform (DCT) 207

Spectrum of 10 Hz sine wave

1.0
Magnitude
0.5

0.0
0 10 20 30 40 50
Frequency (Hz)
Spectrum of 10.5 Hz sine wave
0.75
Magnitude

0.50
0.25

0.00
0 10 20 30 40 50
Frequency (Hz)

Fig. 5.43 Result of python code shown in Fig. 5.42

1
, for k = 0
N
αð k Þ = ð5:67Þ
2
, Otherwise
N

The formula to compute inverse discrete cosine transform is given by

N -1
ð2n þ 1Þπk
x½n] = αðk Þ X ½k] cos ð5:68Þ
k=0
2N

where

1
, for k = 0
N
αðkÞ = ð5:69Þ
2
, Otherwise
N

Experiment 5.24 Computation of Forward and Inverse Discrete Cosine Trans-

form of a Given Signal
This experiment deals with the computation of the forward and inverse DCT of a
given input signal. The python code to compute the forward and inverse DCT of
signal is given in Fig. 5.44, and its simulation result is shown in Fig. 5.45. From this
figure, it is possible to infer that the DCT output is always real value, and there is no
phase component in it.
208 5 Transforms

# DCT and IDCT python implementation

import numpy as np
import matplotlib.pyplot as plt
from scipy import fft
np.set_printoptions(precision=2, suppress=True)
n=np.arange(0,40,1)
k=np.arange(0,40,1)
x=np.sin(2*np.pi*(5/100)*(n))+np.sin(2*np.pi*(15/100)*(n));
plt.figure(1),plt.subplot(3,1,1),plt.stem(n,x),plt.xlabel('n--
>'),plt.ylabel('x[n]'),plt.title('Input Signal')
y1=np.zeros(len(k))
alpha=np.zeros(len(k));
for i in range(len(n)):
if k[i]==0:
alpha[i]=np.sqrt(1/len(n));
else:
alpha[i]=np.sqrt(2/len(n));
y=alpha[i]*x[i]*np.cos(((2*i)+1)*np.pi*k/(2*len(n)))
y1=y1+y
plt.subplot(3,1,2),plt.stem(k,y1),plt.xlabel('k-->'),plt.ylabel('X[k]'),plt.title('DCT output')
y2=np.zeros(len(n))
for i in range(len(k)):
if k[i]==0:
alpha[i]=np.sqrt(1/len(n));
else:
alpha[i]=np.sqrt(2/len(n));
x1=alpha[i]*y1[i]*np.cos(((2*i)+1)*np.pi*k/(2*len(n)))
y2=y2+x1
y2=(1/2)*y2
plt.subplot(3,1,3),plt.stem(n,y2),plt.xlabel('n-->'),plt.ylabel('y[n]'),plt.title('IDCT output')
plt.tight_layout()
y3=fft.dct(x);# Built in command for DCT
z=fft.idct(y3);#Built in Command for IDCT
plt.figure(2),plt.subplot(3,1,1),plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('x[n]'),
plt.title('Input Signal'),plt.subplot(3,1,2),plt.stem(k,y1),plt.xlabel('k-->'),plt.ylabel('X[k]'),
plt.title('DCT output Using in-built'),plt.subplot(3,1,3),plt.stem(n,y2),plt.xlabel('n-->'),
plt.ylabel('y[n]'),plt.title('IDCT output using in-built')

Fig. 5.44 Python code for forward and inverse DCT

Inferences
The following inferences can be made from this experiment:
1. The input is real valued mulitiple sinusoidal signal, and the DCT output of the
input signal shows that most of the DCT coefficients are zero, which indicates
that the signal is highly correlated, and then the few DCT coefficients are used to
represent the signal.
5.5 Discrete Cosine Transform (DCT) 209

Input Signal Input Signal

1 1

x[n]
x[n]

0 0
–1 –1
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
n--> n-->
DCT output DCT output Using in-built
2.5 2.5

x[k]
x[k]

0.0 0.0
–2.5 –2.5
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
k--> k-->
IDCT output IDCT output using in-built

y[n]
y[n]

0 0
–1 –1
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
n--> n-->

(a) Using DCT and IDCT formula (b) Using Built-in Command

Fig. 5.45 Simulation result. (a) Using DCT and IDCT formula. (b) Using built-in command

Input Sawtooth signal

Discrete Fourier Transform (DFT) Discrete Cosine Transform (DCT)

Nullifying a few coefficients of DFT Nullifying a few coefficients of DCT

Inverse DFT Inverse DCT

Reconstructed signal Reconstructed signal

Fig. 5.46 Problem statement

2. Inverse DCT is used to reconstruct the original signal from the DCT coefficients.
From Fig. 5.45, it is possible to observe that the reconstructed signal is exactly the
same as the original signal.
3. The same result is verified with the built-in commands ( fft.dct and fft.idct).
Task
1. Investigate whether DCT matrix entries are all real.
Experiment 5.25 Comparison Between Discrete Fourier Transform
and Discrete Cosine Transform
The objective of this experiment is to compare the performance of discrete Fourier
transform (DFT) with discrete cosine transform (DCT). This is done by taking DFT
and DCT of the sawtooth signal. After taking both DFT and DCT, the last ten
coefficients are nullified. Then, inverse DFT and inverse DCT of the modified
coefficients are taken to obtain the reconstructed signal. This is depicted in Fig. 5.46.
210 5 Transforms

#Comparison of Fourier and DCT

from scipy.fftpack import dct, idct
from scipy.fftpack import fft, ifft
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of sawtooth signal
t=np.linspace(0,1,100)
x=signal.sawtooth(2*np.pi*5*t)
#Step 2: Modifying DFT coefficients
X1=fft(x)
X1[89:99]=0
y1=ifft(X1)
#Step 3: Modifying DCT coefficients
X=dct(x)
X[89:99]=0
y2=idct(X)
#Step 4: Plotting the results
plt.subplot(3,1,1),plt.stem(x),plt.xlabel('n-->'),plt.ylabel('x[n]')
plt.title('Input Sawtooth signal')
plt.subplot(3,1,2),plt.stem(y1),plt.xlabel('n-->'),plt.ylabel('y$_1$[n]')
plt.title('Reconstructed signal using DFT')
plt.subplot(3,1,3),plt.stem(y2),plt.xlabel('n-->'),plt.ylabel('y$_2$[n]')
plt.title('Reconstructed signal using DCT')
plt.tight_layout()

Fig. 5.47 Python code to compare DCT with Fourier transform

The python code, which performs this task, is shown in Fig. 5.47, and the
corresponding output is shown in Fig. 5.48.
Inferences
From Fig. 5.48, it is possible to interpret that the reconstructed signal obtained using
DCT is better than DFT. DCT has better energy compaction than the DFT. This
means that DCT can pack signal energy into a few coefficients.

5.6 Short-Time Fourier Transform

The short-time Fourier transform of the signal x(t) is given by

5.6 Short-Time Fourier Transform 211

Input Sawtooth signal

1

x[n]
0
–1
0 20 40 60 80 100
n-->
Reconstructed signal using DFT
0.5
y1[n]

0.0
–0.5
0 20 40 60 80 100
n-->
Reconstructed signal using DCT
200
y2[n]

0
–200
0 20 40 60 80 100
n-->

Fig. 5.48 Result of python code shown in Fig. 5.47

Fig. 5.49 Block diagram of

Stationary signal X1(τ, Ω)
the problem statement STFT
x1(t)

Non-stationary X2(τ, Ω)
STFT
signal x2(t)

X ðτ, ΩÞ = xðt Þw* ðt - τÞe - jΩt dt ð5:70Þ

-1

where ‘w(t)’ is the window function and ‘τ’ is the centre of the window. Equation
(5.70) can be interpreted as ‘STFT provides two-dimensional representation of the
one-dimensional signal x(t)’. Narrow window provides good time resolution but
poor frequency resolution, whereas wider window provides good frequency resolu-
tion but poor time resolution. According to the Heisenberg uncertainty principle, it is
difficult to obtain both good time and frequency resolutions at the same resolution.
Experiment 5.26 STFT of Stationary and Non-stationary Signal
In this experiment, STFT of stationary and non-stationary signals are obtained, and
their results are interpreted. The stationary signal x1(t) is generated using the formula
x1(t) = sin (2πft); the non-stationary signal x2(t) is generated using the formula
x2(t) = sin (2πft2). The built-in function available in scipy library stft is used to
obtain time-frequency representation of the two signals. The block diagram of the
problem statement is given in Fig. 5.49.
212 5 Transforms

#STFT of stationary and non-stationary signals

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of a stationary signal
fs=100
T,N,f=1/fs, 100, 5
t=np.linspace(0,N*T,N)
x1=np.sin(2*np.pi*f*t)#Step 2: Generation of non-stationary signal
x2=np.sin(2*np.pi*f*t**2)#Step 3: Obtaining the STFT of signals
f1,t1,z1=signal.stft(x1,fs,'hamming',64)
f2,t2,z2=signal.stft(x2,fs,'hamming',64)
#Step 4: Plotting the results
plt.subplot(2,2,1),plt.plot(t,x1),plt.xlabel('t-->'),plt.ylabel('x$_1$(t)'),
plt.title('Signal-1'),plt.subplot(2,2,2),plt.plot(t,x2)
plt.xlabel('t-->'),plt.ylabel('x$_2$(t)'),plt.title('Signal-2')
plt.subplot(2,2,3),plt.pcolormesh(t1,f1,np.abs(z1),shading='gouraud')
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),plt.title('STFT of Signal-1')
plt.subplot(2,2,4),plt.pcolormesh(t2,f2,np.abs(z2),shading='gouraud')
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),plt.title('STFT of Signal-2')
plt.tight_layout()

Fig. 5.50 Python code to obtain time-frequency representation of stationary and non-stationary
signals

The python code to implement the task is done in four steps. First step deals with
the generation of stationary signals, and second step deals with the generation of
non-stationary signals. Obtaining the STFT of the two signals is done in the third
step. Finally, the results are plotted in the fourth step. The python code, which
performs the abovementioned task, is given in Fig. 5.50, and the corresponding
output is shown in Fig. 5.51.
Inferences
From Fig. 5.51, the following inferences can be drawn:
1. Signal-1 and Signal-2 are stationary and non-stationary signals, respectively. For
Signal-1, the frequency does not change with respect to time; hence, it is
stationary, for Signal-2, the frequency increases with an increase in time;
hence, it is non-stationary.
2. STFT provides time-frequency representation of the signal. STFT of Signal-1 is a
horizontal line indicating that Signal-1 has one frequency component at all times.
The STFT of Signal-2 shows the gradual variation of frequency with respect to
time. With respect to time, the frequency changes, which is depicted in the
spectrogram plot.
Experiment 5.27 Impact of Choice of Window Length
In this experiment, the built-in function available in ‘matplot’ library plt.specgram is
used to analyse the impact of the choice of width of the window in STFT. The
5.6 Short-Time Fourier Transform 213

Signal-1 Signal-2
1 1

x2[t]
x1[t]

0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
t--> t-->

STFT of Signal-1 STFT of Signal-2

Frequency (w -->)

Frequency (w -->)
40 40

20 20

0 0
0.0 0.5 1.0 0.0 0.5 1.0
Time (t-->) Time (t-->)

Fig. 5.51 Result of python code shown in Fig. 5.50

Non-stationary signal with

discontinuity

STFT STFT STFT

Window width=16 Window width=32 Window width=128

Good time resolution, Average time and Good frequency resolution,

poor frequency resolution frequency resolution poor time resolution

Fig. 5.52 Problem statement

objective of this experiment is to verify the fact that a shorter window gives a good
time resolution and a wider window gives a good frequency resolution. To demon-
strate this fact, a non-stationary signal is generated. This signal has 5, 0 and 10 Hz
frequency components. In 0 Hz or DC, a discontinuity is introduced. The disconti-
nuity is the increase in the amplitude of the signal from 1 to 2 V. This non-stationary
signal is analysed using a spectrogram of different window widths, namely, 16, 32
and 128. The problem statement is depicted in Fig. 5.52.
The python code, which performs this task, is shown in Fig. 5.53, and the
corresponding output is shown in Fig. 5.54.
214 5 Transforms

#Effect of window length of STFT

import numpy as np
import matplotlib.pyplot as plt
#Step1: Signal generation
fs=100
T,N=1/fs, 100;
#Frequency components of the signal
f1,f2,f3=5, 0, 10;
t1=np.linspace(0,N*T,N)
t=np.linspace(0,N*T,3*N)
x1=np.sin(2*np.pi*f1*t1)
x2=np.sin(2*np.pi*f2*t1)
x3=np.sin(2*np.pi*f3*t1)
x=np.concatenate([x1,x2,x3])
x[150:160]=2 #Discontinuity
#Step 2: Plotting the signal and its spectrogram
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('t-->'),plt.ylabel('x(t)')
plt.title('Signal')
plt.subplot(2,2,2),plt.specgram(x, Fs=fs, NFFT=16, noverlap=1,window =None)
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),
plt.title('Window length=16')
plt.subplot(2,2,3),plt.specgram(x, Fs=fs, NFFT=32, noverlap=1,window =None)
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),
plt.title('Window length=32')
plt.subplot(2,2,4),plt.specgram(x, Fs=fs, NFFT=128, noverlap=1,window =None)
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),
plt.title('Window length=128')
plt.tight_layout()

Fig. 5.53 Python code to analyse the impact of window width in STFT

Inferences
From Fig. 5.54, the following inferences can be drawn:
1. The signal is a non-stationary signal with three frequency components 5 Hz, 0 Hz
(DC component) and 10 Hz, respectively. There is a discontinuity in the signal in
the DC component. The discontinuity refers to an abrupt change in amplitude
from 1 to 2 V.
2. The STFT of the signal is obtained for different window widths, namely, 16, 32
and 128.
3. The spectrogram corresponding to window width 16 gives good time informa-
tion. The occurrence of discontinuity at a particular instant is clearly visible in the
spectrogram with a window width of 16. But the frequency resolution is poor.
The two frequency components present in the signal, namely, 5 and 10 Hz, are not
visible in the spectrogram with a window width of 16. That is, a shorter window
gives good time resolution but poor frequency resolution.
5.6 Short-Time Fourier Transform 215

Signal Window length=16

2

Frequency (w -->)
40
1
x(t)

0 20

–1 0
0.00 0.25 0.50 0.75 1.00 1 2
t--> Time (t-->)

Window length=32 Window length=128

Frequency (w -->)

Frequency (w -->)
40 40

20 20

0 0
1 2 1 2
Time (t-->) Time (t-->)

Fig. 5.54 Result of python code shown in Fig. 5.53

4. When the window width is 32, the frequency components present in the signal,
namely, 5 and 10 Hz, are partially visible. When the window width is 32, the
average time and frequency resolution are obtained.
5. When the window width is 128, the frequency resolution is good. From this
spectrogram, it is possible to identify 5 and 10 Hz frequency components. But it is
not possible to locate the discontinuity present in the signal. This means that time
resolution is poor.
6. In a nut shell, shorter window gives good time resolution but poor frequency
resolution, whereas a wider window gives good frequency resolution but poor
time resolution.
Task
1. Repeat this experiment by choosing different types of window functions like
Bartlett and Kaiser for specified value of ‘β’.
Experiment 5.28 Choice of Window Function in Resolving Two Close Fre-
quency Components
The objective of this experiment is to analyse the choice of window function in
resolving two close frequency components of the input signal. The input signal is the
addition of two sinusoidal signals of frequencies 5 and 8 Hz. The spectrogram of this
signal is obtained for different choices of window functions like rectangular win-
dow, Blackman window and Kaiser window. The impact of window choices in
frequency resolution is analysed in this experiment. The python code, which per-
forms this task, is shown in Fig. 5.55, and the corresponding output is shown in
Fig. 5.56.
216 5 Transforms

#Choice of window function

import numpy as np
import matplotlib.pyplot as plt
#Step1: Signal generation
fs,f1,f2=100, 5, 8;
t=np.linspace(0,1,100)
x1=np.sin(2*np.pi*f1*t)
x2=np.sin(2*np.pi*f2*t)
x=x1+x2
#Step 2: Generation of window functions
NFFT=64
win1 =np.ones((NFFT)) #Rectangular window
win2=np.blackman(NFFT) #Blackman window
beta=1
win3 = np.kaiser(NFFT,beta)
#Step 3: Plotting the results
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('t-->'),plt.ylabel('x(t)')
plt.title('Signal'),plt.subplot(2,2,2),
plt.specgram(x, Fs=fs, NFFT=64, noverlap=1,window = win1)
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),plt.title('Rectangular
window')
plt.subplot(2,2,3),plt.specgram(x, Fs=fs, NFFT=64, noverlap=1,window = win2)
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),plt.title('Blackman window')
plt.subplot(2,2,4),plt.specgram(x, Fs=fs, NFFT=64, noverlap=1,window = win3)
plt.xlabel('Time (t-->)'),plt.ylabel('Frequency ($\omega$-->)'),plt.title('Kaiser window')
plt.tight_layout()

Fig. 5.55 Window function and frequency resolution

Inferences
The following inference can be made from this experiment:
1. From the python code, it is possible to observe that the signal consists of two
frequency components, 5 and 8 Hz, that are added to obtain the input signal
whose time-frequency representation for different windows is obtained.
2. Rectangular window is able to resolve two closely spaced frequency components.
3. Blackman window has a wider main lobe; hence, it could not resolve the
frequency components present in the signal.
4. Kaiser window successfully resolves the frequency components present in the
signal for the choice of β = 1.
5. If the main lobe width of the window is small, then good frequency resolution
could be obtained. Side lobes affect the extent to which adjacent frequency
components leak into the adjacent frequency bins.
Experiment 5.29 Comparison of FT with STFT
This experiment aims to compare Fourier transform with short-time Fourier trans-
form in analysing non-stationary signal. The non-stationary signal considered in this
example has three frequency components, namely, 5, 0 and 15 Hz. In non-stationary
5.6 Short-Time Fourier Transform 217

Signal Rectangular window

2

Frequency (w -->)
40
x(t)

0
20

–2 0
0.00 0.25 0.50 0.75 1.00 0.2 0.4 0.6
t--> Time (t-->)

Blackman window Kaiser window

Frequency (w -->)
Frequency (w -->)

40 40

20 20

0 0
0.2 0.4 0.6 0.2 0.4 0.6
Time (t-->) Time (t-->)

Fig. 5.56 Choice of window and its impact in frequency resolution

Non-stationary signal-1 Non-stationary signal-2

FT STFT FT STFT

Fig. 5.57 Comparison of FT and STFT

Signal-1, 5 Hz signals appear first, followed by the DC and 15 Hz frequency

components. In non-stationary Signal-2, 15 Hz frequency components appear first
followed by DC and 5 Hz frequency components. For these two signals, Fourier
transform and short-time Fourier transform are taken. This objective is illustrated in
Fig. 5.57.
The python code, which implements this task, is given in Fig. 5.58 and the
corresponding output is shown in Fig. 5.59.
Inference
From Fig. 5.59, the following inferences can be drawn:
1. Signal-1 and Signal-2 are non-stationary signals, because the frequency of these
two signals changes with respect to time.
2. The frequency components present in Signal-1 and Signal-2 are 5, 0 and 10 Hz. In
Signal-1, the 5 Hz frequency component appears first, followed by the 0 Hz
frequency component and the 15 Hz frequency component. In Signal-2, 15 Hz
218 5 Transforms

#Comparison of FT and STFT

import numpy as np
import matplotlib.pyplot as plt
from scipy.fft import fft,fftfreq
from scipy import signal
#Step1: Signal generation
fs=100
T=1/fs
N,f1,f2,f3=100,5,0,15
t1=np.linspace(0,N*T,N)
t=np.linspace(0,N*T,3*N)
x1=np.sin(2*np.pi*f1*t1)
x2=np.sin(2*np.pi*f2*t1)
x3=np.sin(2*np.pi*f3*t1)
x=np.concatenate([x1,x2,x3])
y=np.concatenate([x3,x2,x1])
plt.subplot(3,2,1),plt.plot(t,x),plt.xlabel('t-->'),plt.ylabel('x$_1$(t)'),plt.title('Signal-1')
plt.subplot(3,2,2),plt.plot(t,y),plt.xlabel('t-->'),plt.ylabel('x$_2$(t)'),plt.title('Signal-2')
#Step 2: Obtain the spectrum
faxis=fftfreq(3*N,T)[0:3*N//2]
X=fft(x)
Y=fft(y)
#Step 3: Plotting the result
plt.subplot(3,2,3),plt.plot(faxis,2/N*np.abs(X)[0:3*N//2])
plt.xlabel('Frequency ($\omega$-->)'),plt.ylabel('|X$_1$($\omega$)|'),plt.title('Spectrum-1')
plt.subplot(3,2,4),plt.plot(faxis,2/N*np.abs(Y)[0:3*N//2])
plt.xlabel('Frequency ($\omega$-->)'),plt.ylabel('|X$_2$($\omega$)|'),plt.title('Spectrum-2')
#Step 4: STFT of the signals
f1,t1,z1=signal.stft(x,fs,'hamming',1024)
f2,t2,z2=signal.stft(y,fs,'hamming',1024)
plt.subplot(3,2,5),plt.pcolormesh(t1, f1, np.abs(z1),shading='gouraud')
#plt.pcolormesh(t1,f1,np.abs(z1),shading='flat')
plt.xlabel('Time (t-->)'),plt.ylabel('Freq($\omega$-->)'),plt.title('STFT of Signal-1')
plt.subplot(3,2,6),plt.pcolormesh(t2,f2,np.abs(z2),shading='gouraud')
plt.xlabel('Time (t-->)'),plt.ylabel('Freq ($\omega$-->)'),plt.title('STFT of Signal-2')
plt.tight_layout ()

Fig. 5.58 Python code to compare FT with STFT

frequency component appears first, then followed by 0 Hz and finally by 5 Hz

frequency component.
3. From Fig. 5.59, it is possible to interpret that spectrum-1 and spectrum-2 are
alike. That is magnitude spectrum of Fourier transform cannot distinguish Signal-
1 and Signal-2. The reason is Fourier transform is an effective tool for the
frequency representation of the stationary signal, but it does not provide time
information.
4. STFT of Signal-1 shows time-frequency representation of the signal. STFT of
Signal-1 indicates that 5 Hz frequency component appears first, followed by
15 Hz frequency component. In STFT of Signal-2, it is possible to observe that
5.7 Continuous Wavelet Transform (CWT) 219

Signal-1 Signal-2
1 1

x2(t)
x1(t)

0 0
–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
t--> t-->

Spectrum-1 Spectrum-2
1 1
|X1(w)|

|X2(w)|
0 0
0 20 40 0 20 40
Frequency (w -->) Frequency (w -->)

STFT of Signal-1 STFT of Signal-2

Freq(w -->)

Freq(w -->)
50 50

0 0
0 1 2 3 0 1 2 3
Time (t-->) Time (t-->)

Fig. 5.59 Result of comparison of FT with STFT

15 Hz frequency component appears first, followed by 5 Hz frequency

component.
5. It is possible to interpret that STFT is effective in analysing non-stationary
signals.

5.6.1 Shortcoming of STFT

The width of the window cannot be changed once it is fixed. This implies that STFT
provides fixed resolution. For the multi-resolution representation of the signal,
wavelet transform is employed.

5.7 Continuous Wavelet Transform (CWT)

Wavelets are oscillatory functions of finite duration. Wavelet transform provides

timescale relationship of the signal. The continuous wavelet transform of the signal
f(t) is expressed as
220 5 Transforms

1
1 t-b
W f ða, bÞ = p f ðt Þψ * dt ð5:71Þ
a a
-1

In the above equation, f(t) represents signal of interest, ψ(t) denotes ‘mother
wavelet’, b is the shifting parameter and a is the scaling parameter. The above
equation can be written as

W f ða, bÞ = f ðt Þ, ψ a,b ðt Þ ð5:72Þ

The above equation indicates that wavelet transform is basically taking inner
product of the function f(t) with the ‘daughter wavelet’ ψ a, b(t). The daughter
wavelets are derived from the mother wavelet ψ(t) using the relation

1 t-b
ψ a,b ðt Þ = p ψ ð5:73Þ
a a

5.7.1 Continuous Wavelets Family

A variety of continuous wavelets filter are currently in use. They are (1) Haar,
(2) Mexican Hat, (3) Morlet, (4) Complex Morlet, (5) Gaussian, (6) Shannon and
(7) Daubechies. The wavelet family and its mathematical expression are given in
Table 5.3.
Experiment 5.30 Detection of Discontinuity in the Signal Using CWT
The objective of this experiment is to detect the discontinuity present in the signal
using continuous wavelet transform (CWT). The built-in function cwt available in
the library pywt is utilized in this experiment. The three steps followed in this
experiment are the following: Step 1: generating signal with discontinuity; Step 2:
obtaining timescale relationship using CWT, in which the wavelet chosen for this
study is Gaussian wavelet; and Step 3: plotting the signal and the corresponding
scalogram. The python code which performs this task is shown in Fig. 5.60, and the
corresponding output is shown in Fig. 5.61.
Inferences
From Fig. 5.61, the following inferences can be drawn:
1. The input signal is a smooth sinusoidal signal with a sharp discontinuity at a
particular location.
2. Upon observing the CWT result, it is possible to interpret that discontinuity
occurs at 90th sample of the sinusoidal signal, which has 200 samples of data.
3. Thus, CWT is capable of detecting the discontinuity present in the signal.
5.8 Discrete Wavelet Transform 221

Table 5.3 List of wavelet family

Python command
Name of pywt.Wavelet
wavelet Mathematical expression (‘wavelet_name’)
Haar 1 ‘haar’
1, 0≤t<
2
ψ ðt Þ = 1
-1 ≤t<1
2
0, otherwise
Mexican Hat ψ ðt Þ = p 2p4 ð1 - t 2 Þ exp - t2 ‘mexh’
3 π 2

Morlet ψ ðt Þ = exp - t2
cosð5t Þ ‘morl’
2

Complex ψ ðt Þ = p2 exp - t2
expðj2πCtÞ ‘cmor’
Morlet πB B
Where B is Bandwidth and C is centre frequency
Gaussian ψ(t) = C exp (-t2) ‘gaus’
wavelet where C is an order-dependent normalization
constant
p
Shannon ðπBt Þ
ψ ðt Þ = B sinπBt expðj2πCt Þ ‘shan’
wavelet Where B is Bandwidth and C is centre frequency

Task
1. Repeat the experiment for different choices of mother wavelet and comment on
the observed result.

5.7.2 Drawback of CWT

CWT is a redundant representation because of continuous values taken by scaling

and shifting parameters. Overcoming the problem of redundant representation, a
discrete wavelet transform was proposed.

5.8 Discrete Wavelet Transform

The discrete wavelet transform decomposes the signal into approximation and detail.
The process is further iterated by decomposing the approximation with the assump-
tion that much of the signal energy is in approximation. This idea is illustrated in
Fig. 5.62.
In Fig. 5.62, L1 corresponds to first-level decomposition, where the signal is
decomposed into approximation and detail. In the second-level of decomposition
(L2), the approximation obtained in L1 is further decomposed into approximation and
detail. In the third-level of decomposition (L3), the approximation of level L2 is
222 5 Transforms

#Discontinuity detection using CWT

import pywt
import numpy as np
import matplotlib.pyplot as plt
#Step 1: Signal generation
#t=np.linspace(0,1,200)
t=np.arange(0,200,1);
x=np.sin(2*np.pi*5*t/len(t))
x[90]=10
#Step 2: CWT of the signal
scale=np.arange(1,5)
coef,freqs=pywt.cwt(x,scale,'gaus1')
plt.subplot(2,1,1),plt.plot(t,x),plt.xlabel('t-->'),plt.ylabel('x(t)')
plt.title('Signal with discontinuity')
#Step 3: Plotting the reslt
plt.subplot(2,1,2),
plt.imshow(abs(coef),extent=[0,200,30,1],interpolation='bilinear',cmap='winter',
aspect='auto',vmax=abs(coef).max(),vmin=-abs(coef).max())
plt.gca().invert_yaxis()
plt.xticks(np.arange(0,201,20))
plt.xlabel('Time (t-->)'),plt.ylabel('Freq Scale ($\omega$-->)'),
plt.title('CWT of the signal')
plt.tight_layout()
plt.show()

Fig. 5.60 Discontinuity detection using CWT

Signal with discontinuity

10

5
x(t)

0
0 25 50 75 100 125 150 175 200
t-->
CWT of the signal
Freq Scale (w -->)

30

20

10

0 20 40 60 80 100 120 140 160 180 200

Time (t-->)

Fig. 5.61 CWT of signal with discontinuity

5.8 Discrete Wavelet Transform 223

Signal

Approximation Detail
L1

Approximation Detail
L2

Approximation Detail
L3

Fig. 5.62 Wavelet decomposition

Square wave with

Square wave
discontinuity

L1

Approximation Detail Approximation Detail

Inverse Wavelet Transform Inverse Wavelet Transform

Reconstructed signal Reconstructed signal

Fig. 5.63 Pictorial representation of problem statement

decomposed further into approximation and detail. This is done assuming that most
of the signal energy is in approximation.
Experiment 5.31 Detection of Discontinuity in Signal Using DWT
The objective of this experiment is to compare the first-level approximation and
detail of DWT coefficient of a signal with the coefficient of the signal with discon-
tinuity. Here discontinuity refers to sudden changes in the amplitude of the signal.
The problem statement is depicted in Fig. 5.63.
From Fig. 5.63, it is possible to observe that two signals are considered in this
experiment, Signal-1 is a square wave, whereas Signal-2 is a square wave with
discontinuity. L1 in the figure represents the first-level of decomposition. Upon first-
level of decomposition, the signal is split into approximation and detail. Upon taking
an inverse discrete wavelet transform, it is possible to reconstruct the signal. The
224 5 Transforms

#DWT of a signal with discontinuity

import pywt
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Signal generation
f=5
n=np.arange(0,100,1)
x1=signal.square(2*np.pi*f*n/len(n)) #Square wave
x2=signal.square(2*np.pi*f*n/len(n)) #Square wave with discontinuity
x2[50]=5
#Step 2: DWT of the signal
cA,cD=pywt.dwt(x1,'db1')
cA1,cD1=pywt.dwt(x2,'db1')
#Step 3: Inverse DWT
y1=pywt.idwt(cA,cD,'db1')
y2=pywt.idwt(cA1,cD1,'db1')
#Step 4: Plotting the result
plt.subplot(3,2,1),plt.stem(n,x1),
plt.xlabel('n-->'),plt.ylabel('x$_1$[n]'),plt.title('Signal-1')
plt.subplot(3,2,2),plt.stem(n,x2),
plt.xlabel('n-->'),plt.ylabel('x$_2$[n]'),plt.title('Signal-2')
WC=np.concatenate([cA,cD])
WC1=np.concatenate([cA1,cD1])
plt.subplot(3,2,3),plt.stem(n,WC),plt.title('First level Decomposition')
plt.subplot(3,2,4),plt.stem(n,WC1),plt.title('First level Decomposition')
plt.subplot(3,2,5),plt.stem(n,y1),
plt.xlabel('n-->'),plt.ylabel('y$_1$[n]'),plt.title('Reconstructed signal-1')
plt.subplot(3,2,6),plt.stem(n,y2),
plt.xlabel('n-->'),plt.ylabel('y$_2$[n]'),plt.title('Reconstructed signal-2')
plt.tight_layout()

Fig. 5.64 Python code to compute the DWT CWT of signal with discontinuity

python code that performs this task mentioned above is shown in Fig. 5.64, and the
corresponding output is in Fig. 5.65.
Inferences
From Fig. 5.65, the following inferences are drawn:
1. Signal-1 is a square wave with 5 Hz fundamental frequency; Signal-2 is a square
wave with discontinuity.
2. The first-level decomposition of the signal gives approximation and detail coef-
ficients. For Signal-1, the approximation coefficient is similar to the signal,
whereas the detail coefficient is almost zero. For Signal-2, the discontinuity is
captured in detail coefficient.
3. Upon taking inverse DWT, the reconstructed signals are obtained, which resem-
bles the input signal. Thus, DWT is a reversible transform.
5.8 Discrete Wavelet Transform 225

Signal-1 Signal-2
1 5
x1[n]

x2[n]
0
–1 0
0 25 50 75 100 0 25 50 75 100
n--> n-->

First level Decomposition First level Decomposition

1
2.5
0
–1 0.0
0 25 50 75 100 0 25 50 75 100

Reconstructed signal-1 Reconstructed signal-2

1 5
y1[n]

0 y2[n]
–1 0
0 25 50 75 100 0 25 50 75 100
n--> n-->

Fig. 5.65 Result of python code shown in Fig. 5.64

4. In this experiment, it is possible to observe that discontinuity is captured in the

detail coefficient.
Task
1. Repeat this experiment for different choices of mother wavelet and comment on
the observed result.
Experiment 5.32 Denoising of Signal Using DWT and Thresholding Approach
The objective of this experiment is to donoise the signal using discrete wavelet
transform and inverse discrete wavelet transform. The input signal (sawtooth signal)
is corrupted by white noise, which follows normal distribution of zero mean and
variance of 0.125. The wavelet decomposition of the noisy signal is performed using
the built-in function ‘wavedec’ available in ‘pywavelet’ library. The wavelet chosen
for decomposition is db2, and the level of decomposition chosen is 3. After wavelet
decomposition, the detail coefficients are thresholded using the built-in function
‘pywt.threshold’. The choice of threshold is soft ‘thresholding’. After thresholding,
the modified wavelet coefficients are reconstructed using the built-in function
wavedec to obtain the reconstructed (filtered) signal. The python code, which
performs this task, is shown in Fig. 5.66, and the corresponding output is shown
in Fig. 5.67.
Inferences
The following inferences can be made from this experiment:
1. The input signal (clean signal) is a sawtooth signal of 5 Hz fundamental
frequency.
226 5 Transforms

#Denoising of signals using DWT

import pywt
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Signal generation
f=5
t=np.arange(0,50,1)
x1=signal.sawtooth(2*np.pi*f*t/len(t))
#Step 2: Adding noise to the clean signal
n=np.random.normal(0,0.125,len(x1))
x=x1+n
wavelet = 'db2'
level =3
# Step 3: Perform wavelet decomposition
coeffs = pywt.wavedec(x, wavelet, level=level)
# Step 4: Define threshold for filtering
threshold = 0.75 * np.max(coeffs[-1])
# Step 5: Perform wavelet thresholding
coeffs_filtered = [pywt.threshold(c, threshold, mode='soft') for c in coeffs]
# Step 6: Reconstruct filtered signal
y= pywt.waverec(coeffs_filtered, wavelet)
#Step 7: Plotting the result
plt.subplot(3,1,1),plt.plot(t,x1),plt.xlabel('n-->'),plt.ylabel('x[n]')
plt.title('Clean signal')
plt.subplot(3,1,2),plt.plot(t,x),plt.xlabel('n-->'),plt.ylabel('z[n]')
plt.title('Noisy signal')
plt.subplot(3,1,3),plt.plot(t,y),plt.xlabel('n-->'),plt.ylabel('y[n]')
plt.title('Filtered signal')
plt.tight_layout()

Fig. 5.66 Python code to perform denoising of the signal

2. The noisy signal is obtained by adding white noise, which follows normal
distribution to the input signal.
3. The noisy signal is decomposed using db2 wavelet. The level of decomposition is
three.
4. After wavelet decomposition, the detail coefficients are thresholded using soft
thresholding to minimize the impact of noise. It is generally believed that much of
the signal energy will be in low-frequency regions and noise will reside in high-
frequency regions.
5. The inverse wavelet transform of the modified wavelet coefficients is performed
to obtain the filtered signal.
6. From Fig. 5.67, it is possible to interpret that the impact of noise is less in filtered
signal when compared to noisy signal.
5.8 Discrete Wavelet Transform 227

Clean signal

x[n] 0
–1
0 10 20 30 40 50
n-->
Noisy signal
1
z[n]

0
–1
0 10 20 30 40 50
n-->
Filtered signal
1
y[n]

0 10 20 30 40 50
n-->

Fig. 5.67 Denoising of sawtooth signal

Tasks
1. The experiment can be repeated by the following: (a) Choose different wavelet
family other than ‘db2’. (b) The level of decomposition can be changed.
(c) Instead of soft thresholding, hard thresholding can be tried.
Exercises
1. Write a python code to obtain the Z-transform of the following sequences:
(a) x1[n] = δ[n - 5] (b) x2[n] = u[n] - u[n - 1] (c) x3[n] = nu[n]
(d) x4[n] = sin (ω0n).
2. Write a python code to compute the inverse Z-transform of (a) X1(z) = z-2
(b) X 2 ðzÞ = ð1 - 1z - 1 Þ2 .
3. Write a python code to compute the magnitude and phase responses of the
system, whose transfer function is given by H ðzÞ = 1 -1z - 1 .
4. Let the signal x[n] represent 100 samples of 5 Hz sine wave. Now increase the
length of the signal by padding 50 sample values of zeros to x[n]. Zero padding
is done at the end of 100 samples of x[n]. Let the zero padded signal be denoted
as y[n]. Write a python code to plot the spectrum of the signal x[n] and y
[n]. Comment on the observed result.
5. Obtain the DFT of the sequences x1[n] = {1, 1, 1, 1} and x2[n] = {1, -1, 1, -1}.
Plot their magnitude responses and comment on the observed result.
6. Generate a square wave of 5 Hz fundamental frequency. Write a python code to
plot the spectrum of the square wave and comment on the observed result.
7. Let x[n] represent 100 samples of 5 Hz sine wave. Let y[n] represent 100 samples
of 5 Hz cosine wave. Take Fourier transform of x[n] and y[n] to obtain X[k] and
Y[k]. Extract the magnitude and phase components of X[k] and Y[k]. Now
interchange the phase of X[k] with Y[k]. After phase interchange, take inverse
228 5 Transforms

Fourier transform to obtain x′[n] and y′[n]. Use subplot to plot the signals x[n], y
[n], x′[n] and y′[n] and comment on the observed result.
8. Generate a linear chirp signal whose frequency varies from 10 to 1 Hz in 10 s.
Plot the spectrum and spectrogram of this chirp signal and comment on the
observed result.
9. Write a python code to verify the fact that a shorter window gives good time
resolution and a wider window gives good frequency resolution in short-time
Fourier transform.
10. Generate sinusoidal signal with momentary interruption. Apply CWT to identify
the momentary interruption present in the signal.
Objective Questions
1. The region of convergence of unit sample signal (δ[n]) is
A. Entire Z-plane except z = 0
B. Entire Z-plane except z = infinity
C. Entire Z-plane
D. Entire Z-plane except z = 0 and z = infinity
2. Convolution in time domain is equivalent to
A. Addition in Z-domain
B. Subtraction in Z-domain
C. Multiplication in Z-domain
D. Division in Z-domain
3. For a discrete-time system to be stable
A. Pole should lie inside the unit circle.
B. Pole should lie outside the unit circle.
C. Pole should lie on the unit circle.
D. Pole can lie anywhere in the z-plane.
4. Z-transform of x[n] = nu[n] is
A. X ðzÞ = z
z-1
B. X ðzÞ = z
ðz - 1Þ2
2
C. X ðzÞ = z
z-1
D. X ðzÞ = z
ðz - 1Þ3

5. Let X(z) be the Z-transform of the signal x[n]. If X ðzÞ = z

z - 1, then lim x½n] is
n→1

A. 0
B. 1
C. -1
D. Infinite
5.8 Discrete Wavelet Transform 229

n n
6. Let x½n] = 25 u½n] - 52 u½ - n - 1]. Let X(z) be the Z-transform of the given
signal x[n]; then, the region of convergence of its Z-transform is
A. 25 < jzj < 1
B. 52 < jzj < 1
C. 25 < jzj < 52
D. - 1 < jzj < 5
2

7. The transfer function of an LTI system is

A. Linear function of ‘z’
B. Rational function of ‘z’
C. Logarithmic function of ‘z’
D. Exponential function of ‘z’
8. The basis function of Fourier transform is
A. Triangular function
B. Rectangular function
C. Complex exponential function
D. Prolate spheroidal function
9. Fourier transform of a Gaussian function will result in
A. Triangular function
B. Rectangular function
C. Sinc function
D. Gaussian function
10. Fourier transform of a rectangular function will result in
A. Triangular function
B. Rectangular function
C. Sinc function
D. Gaussian function
11. The 2 × 2 DFT matrix is given by
1 1
A.
1 1
1 1
B.
1 -1
1 0
C.
0 1
1 0
D.
0 -1
230 5 Transforms

Bibliography

1. Ronald N. Bracewell, “Fourier Transform and its Applications”, McGraw Hill, 1978.
2. Alexander D. Poularikas, and Richard C. Dorf, “Transforms and Applications Handbook”,
Wiley, 2021.
3. Martin Vetterli, and Jelena Kovacevic, “Wavelets and Subband Coding”, CreateSpace Indepen-
dent Publishing Platform, 2013.
4. Ronald L. Allen, Duncan W. Mills, “Signal Analysis: Time, Frequency, Scale, and Structure”,
Wiley-IEEE Press, 2004.
5. B. P. Lathi, “Signals, Systems and Communication”, B.S publication, 2001.
Chapter 6
Filter Design Using Pole-Zero Placement
Method

Learning Objectives
After reading this chapter, the reader is expected to
• Design, implement and analyse first-order infinite impulse response filter.
• Design, implement and analyse the moving average filter.
• Design and analyse digital resonator.
• Design and analyse notch filter and comb filter.
• Design and analyse all-pass filter.

Roadmap of the Chapter

Digital filters can be considered as a linear time-invariant system that accepts input
and gives modified input as the output. Based on the input-output relation, discrete-
time systems can be classified as autoregressive system (AR), moving average
system (MA) and autoregressive moving average (ARMA) system. If the current
output is a function of the current input and past outputs, then the system is
autoregressive. If the current output of the system is a function of current input
and past inputs, the system is a moving average system. If the current output is a
function of both past input and past output, the system is autoregressive moving
average system. An example of autoregressive system is the IIR filter and notch
filter. Digital resonator and M-point moving average systems are examples of MA
system. All-pass filter is an example of an ARMA system. This is depicted below.

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 231
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_6
232 6 Filter Design Using Pole-Zero Placement Method

Digital Filter as LTI system

Autoregressive System Moving Average System Autoregressive Moving

(AR system) (MA system) Average System (ARMA)

First-order IIR Digital

filter All pass filter
M-point Moving
Notch filter Average Filter

The filters discussed in this chapter include the IIR filter, moving average filter,
digital resonator, notch filter, comb filter and all-pass filter.
PreLab Questions
1. If h[n] represents the impulse response of a lowpass filter, what would be the
behaviour of the filter whose impulse response is (-1)nh[n]?
2. When a discrete-time system is said to be a minimum phase system?
3. What do you understand by the term ‘poles’ and ‘zeros’ of a system?
4. How is the stability of the discrete-time system related to (a) location of poles of
the system and (b) impulse response of the system?
5. What is the basic principle involved in the design of digital filter using pole-zero
placement method?
6. What do you understand by the term ‘delay equalizer’ or ‘phase equalizer’?
7. Mention two applications of notch filter.
8. A square wave is fed as input to M-point moving average filter. What would be
the output of M-point moving average filter?
9. The relationship between the input and output of a digital filter is given by y
[n] = αx[n] + βx[n - 1] + γx[n - 2]. Is this a finite impulse response filter (FIR)
or infinite impulse response filter (IIR)? Justify your choice.
10. What is a pole-zero plot? What information one gets by interpreting the pole-
zero plot?

6.1 First-Order IIR Filter

This section begins with the design of a first-order IIR filter. The transfer function of
a first-order IIR filter is given by

1
H ðzÞ = ð6:1Þ
1 - p1 z - 1

If the pole lies on the unit circle ( p1 = 1), the transfer function of the filter is given
by
6.1 First-Order IIR Filter 233

1
H ðzÞ = ð6:2Þ
1 - z-1

Step 1: Magnitude response of the filter

From the magnitude response of the filter, it is possible to observe the filter
behaviour at low and high frequencies. The frequency response of the system is
obtained by substituting z = ejω in Eq. (6.2), we get

1
H ejω = ð6:3Þ
1 - e - jω

The above equation can be expressed as

1
H ejω = ð6:4Þ
1 - cosðωÞ þ j sinðωÞ

The magnitude response is obtained as

1
H ejω = ð6:5Þ
ð1 - cosðωÞÞ2 þ sin 2 ðωÞ

Upon simplifying the above equation

1
H ejω = ð6:6Þ
2 - 2 cosðωÞ

When ω = 0, the magnitude response tends to infinity, and when ω = π, the

magnitude response is given by jH ðejω Þj = 12. The filter passes low-frequency
components and attenuates high-frequency components. The filter behaves like a
lowpass filter.
Step 2: Impulse response of the filter
The impulse response of the filter is obtained by taking inverse Z-transform of
the transfer function. The impulse response of the filter is given by

h½ n] = Z - 1 fH ð z Þ g ð6:7Þ

Upon substituting Eq. (6.2) in Eq. (6.7), we get

1
h½n] = Z - 1 = u½ n] ð6:8Þ
1 - z-1

The impulse response of the filter is the unit step function. The unit step
function is not absolutely summable; hence, the filter is not stable.
234 6 Filter Design Using Pole-Zero Placement Method

Table 6.1 Built-in function in the design of first-order IIR filter

Built-in
function Library Use
freqz signal. To obtain the frequency response of the filter, which is a combination
scipy of magnitude and phase responses
tf2zpk signal. To obtain the poles, zeros and gain of the filter
scipy

Experiment 6.1 Characteristics of First-Order IIR Filter

The experiment is about obtaining the characteristics of the filter like (a) impulse
response, (b) pole-zero plot and (c) magnitude and phase responses of the filter. The
built-in functions used to obtain the responses of the filter are given in Table 6.1
The python code which obtains the characteristics of the first-order IIR filter is
given in Fig. 6.1, and the corresponding output is shown in Fig. 6.2.
The python code consists of five steps which are given as S1 to S5. Step 1 (S1)
generation of an impulse to obtain the impulse response of the filter. Step 2 (S2)
defines the system in terms of the transfer function of the system. Step 3 (S3) deals
with obtaining the impulse response of the filter for which the input to the filter is
unit sample signal. Step 4 (S4) deals with obtaining the frequency response of the
filter, and Step 5 (S5) deals with plotting the characteristics of the filter. The result
obtained upon execution of the code shown in Fig. 6.1 is given in Fig. 6.2.
Inferences
From Fig. 6.2, the following inferences can be obtained:
1. The impulse response of the filter is obtained as a unit step function, which is in
agreement with the theoretical result.
2. From the pole-zero plot, it is possible to observe that the pole lies on the unit
circle.
3. From the frequency response, it is possible to observe that the filter behaves like a
lowpass filter.
4. For a discrete-time system to be stable, the impulse response should be absolutely
summable. From Fig. 6.2, the impulse response is a unit step function that is not
absolutely summable; hence, the given filter is not stable.
5. For discrete-time system to be stable, the poles should lie within the unit circle.
From the pole-zero plot, it is possible to observe that the pole lies on the unit
circle; hence, the filter is BIBO stable.
Task
1. Repeat this experiment for H ðzÞ = 1þz1 - 1 , and comment on the observed result.
Will the pole-position change the nature of the system?
Experiment 6.2 Input-Output of First-Order IIR Filter
In order to understand the behaviour of the filter, two types of inputs are fed to the
filter. Input 1 is a DC signal, whereas input 2 is a high-frequency signal. The python
6.1 First-Order IIR Filter 235

#Characteristics of first-order IIR filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#S1: Generation of impulse input
x=np.zeros(100)
x[0]=1
#S2: Define the system
num,den=[1],[1,-1]
#S3: To obtain the impulse response
h=signal.lfilter(num,den,x)
#S4: Characteristics of the first order IIR filter
fs=100
w,H=signal.freqz(num,den)
z,p,k=signal.tf2zpk(num,den)
#S5: Plotting the result
plt.figure(1),plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]'),
plt.title('Impulse Response')
plt.subplot(2,2,2),plt.plot((w/np.pi)*fs/2,20*np.log10(np.abs(H))),
plt.xlabel('$\omega$-->'),plt.ylabel('|H($j\omega$)|'),
plt.title('Magnitude Response'),plt.subplot(2,2,3),plt.xlabel('$\sigma$'),
plt.ylabel('$j\omega$'),plt.title('Pole Zero Plot')
plt.plot(np.real(z),np.imag(z),'ko'),plt.plot(np.real(p),np.imag(p),'rx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta))
plt.subplot(2,2,4),plt.plot(w,np.angle(H)),plt.xlabel('$\omega$-->'),
plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase Response')
plt.tight_layout()

Fig. 6.1 Python code to obtain the characteristics of first order IIR filter

code which deals with the response of the filter for these two different types of inputs
is shown in Fig. 6.3, and the corresponding output is shown in Fig. 6.4.
Inferences
From Fig. 6.4, the following inferences can be drawn:
1. y1[n] is the output signal corresponding to the input signal x1[n]. Here x1[n] is a
DC signal. The signal x1[n] is generated from the expression x1[n] = ejωn by
substituting ω = 0. From the output signal y1[n], it is possible to interpret that the
filter amplifies x1[n], which is a DC signal.
2. y2[n] is the output signal corresponding to the input signal x2[n]. Here x2[n] is an
AC signal. The signal x2[n] is generated from the expression x2[n] = ejωn by
substituting ω = π. From the output signal y2[n], it is possible to interpret that the
filter blocks x2[n], which is a high-frequency signal.
236 6 Filter Design Using Pole-Zero Placement Method

Impulse Response Magnitude Response

1.0
40

|H(jw)_
20
h[n]

0.5

0
0.0
0 25 50 75 100 0 20 40
n--> w -->

Pole Zero Plot Phase Response

1 0

H(jw)
0
jw

–1

–1
–1.0 –0.5 0.0 0.5 1.0 0 1 2 3
s w -->

Fig. 6.2 Characteristics of first-order IIR filter

#Input-Output of first-order IIR filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Signal generation
n=np.arange(-5,6)
omega1=0
omega2=np.pi
x1=np.exp(1j*omega1*n)
x2=np.exp(1j*omega2*n)
#S2: Define the system
num,den=[1],[1,-1]
#S3: To obtain the output
y1=signal.lfilter(num,den,x1)
y2=signal.lfilter(num,den,x2)
#S4: Plotting the input and output of the filter
plt.subplot(2,2,1),plt.stem(n,x1),plt.xticks(n),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('$x_1[n]$'),plt.subplot(2,2,2),plt.stem(n,y1),plt.xticks(n),
plt.xlabel('n-->'),plt.ylabel('Amplitude'), plt.title('$y_1[n]$')
plt.subplot(2,2,3),plt.stem(n,x2),plt.xticks(n),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('$x_2[n]$'),plt.subplot(2,2,4),plt.stem(n,y2),plt.xticks(n),
plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$y_2[n]$')
plt.tight_layout()

Fig. 6.3 Python code to obtain the output of the filter

6.1 First-Order IIR Filter 237

x1[n] y1[n]
1.0
10
Amplitude

Amplitude
0.5 5

0.0 0
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->

x2[n] y2[n]
1 0.0

Amplitude
Amplitude

0 –0.5

–1 –1.0
–5 –4 –3 –2 –1 0 1 2 3 4 5 –5 –4 –3 –2 –1 0 1 2 3 4 5
n--> n-->

Fig. 6.4 Input and output of first-order IIR filter

3. When the poles are at position p = 1, the filter amplifies the DC signal and blocks
the high-frequency signal. The filter behaves like a lowpass filter.
Experiment 6.3 Impact of Pole Position on the Magnitude and Impulse
Responses of First-Order IIR Filter
The objective of this experiment is to analyse the impact of pole position on the
impulse and magnitude response of first-order IIR filter whose transfer functions are
given by H 1 ðzÞ = 1 - 0:25z
1
- 1 and H 2 ðzÞ = 1þ0:25z - 1 . The python code which obtains
1

the impulse response and the magnitude response of the two filters is given in
Fig. 6.5, and the corresponding output is shown in Fig. 6.6.
Inferences
From Fig. 6.6, the following inferences can be drawn:
1. The pole of System-1 whose transfer function is given by H 1 ðzÞ = 1 - 0:25z
1
- 1 lies

on the positive half of the Z-plane.

2. The impulse response of System-1 is observed to be an exponentially decreasing
function.
3. From the magnitude response of System-1, it is possible to observe that System-1
behaves like a lowpass filter.
4. The pole of System-2 whose transfer function is given by H 2 ðzÞ = 1þ0:25z1
- 1 lies

on the negative half of Z-plane.

5. From the magnitude response of System-2, it is possible to infer that the System-2
behaves like a highpass filter.
238 6 Filter Design Using Pole-Zero Placement Method

#Impact of pole position on the behaviour of the system

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Defining systems
num,den1,den2=[1],[1,-0.25],[1,0.25]
#Obtaining the magnitude response
w1,H1=signal.freqz(num,den1)
z1,p1,k1=signal.tf2zpk(num,den1)
w2,H2=signal.freqz(num,den2)
z2,p2,k2=signal.tf2zpk(num,den2)
#To generate imupluse input
x=np.zeros(15)
x[0]=1
h1=signal.lfilter(num,den1,x)
h2=signal.lfilter(num,den2,x)
plt.subplot(2,3,1),plt.plot(np.real(z1),np.imag(z1),'ko'),plt.xlabel('$\sigma$'),
plt.ylabel('$j\omega$'),plt.title('Pole Zero Plot')
plt.plot(np.real(p1),np.imag(p1),'rx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta))
plt.subplot(2,3,2),plt.stem(h1),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_1[n]$')
plt.subplot(2,3,3),plt.plot(w1,20*np.log10(np.abs(H1)))
plt.xlabel('$\omega$-->'),plt.ylabel('|$H_1(j\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,3,4),plt.plot(np.real(z2),np.imag(z2),'ko')
plt.plot(np.real(p2),np.imag(p2),'rx'),plt.xlabel('$\sigma$'),
plt.ylabel('$j\omega$'),plt.title('Pole Zero Plot')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta))
plt.subplot(2,3,5),plt.stem(h2),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('$h_2[n]$')
plt.subplot(2,3,6),plt.plot(w2,20*np.log10(np.abs(H2)))
plt.xlabel('$\omega$-->'),plt.ylabel('|$H_2(j\omega$)|'),plt.title('Magnitude response')
plt.tight_layout()

Fig. 6.5 Python code to analyse the impact of pole position on the behaviour of the filter

6. The impulse response of System-1 is given by h1[n] = (0.25)nu[n], whereas the

impulse response of System-2 is given by h2[n] = (-1)nh1[n]. If h1[n] acts as a
lowpass filter, then h2[n] behaves like a highpass filter.
7. This experiment concludes that the pole position changes the behaviour of the
filter from lowpass to highpass filter.
Task
1. Repeat the above experiment for the system whose transfer function is given by
H ðzÞ = 1 - 12z - 1 , and comment on the observed result.
6.2 Moving Average filter 239

Pole Zero Plot h1[n] Magnitude response

1 1.0
2

Amplitude

|H1(jw)_
0 u 0.5
jw

–1 0.0 –2
–1 0 1 0 10 0 2
s n--> w -->

Pole Zero Plot h2[n] Magnitude response

1 1.0
2
Amplitude

|H2(jw)_
0.5
u
jw

0 0
0.0
–1 –2
–1 0 1 0 10 0 2
s n--> w -->

Fig. 6.6 Result of python code shown in Fig. 6.5

6.2 Moving Average filter

The relationship between the input and output of M-point moving average filter is
given by

M-1
1
y ½ n] = x½ n - k ] ð6:9Þ
M k=0

For a three-point moving average filter, M = 3, substituting M = 3 in Eq. (6.9),

we get the input-output relationship as

1
y½n] = fx½n] þ x½n - 1] þ x½n - 2]g ð6:10Þ
3

From Eq. (6.10), it is possible to interpret that equal weightage is given to x[n], x
[n - 1] and x[n - 2]. This type of system is termed as ‘moving average system’. This
system performs the weighted average of three input samples x[n], x[n - 1] and x
[n - 2]; hence, it is termed as ‘moving average filter’.
Experiment 6.4 Characteristics of Moving Average Filter
This experiment tries to obtain the characteristics of a moving average filter using
python. The python code, which obtains the characteristics of three-point moving
average filter, is shown in Fig. 6.7, and the corresponding output is shown in
Fig. 6.8.
240 6 Filter Design Using Pole-Zero Placement Method

#Characteristics of Moving average filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#S1: Generation of impulse input
x=np.zeros(100)
x[0]=1
#S2: Define the three-point Moving average system
num,den=[1/3,1/3,1/3],[1]
#S3: To obtain the impulse response
h=signal.lfilter(num,den,x)
#S4: Characteristics of the first-order IIR filter
fs=100
w,H=signal.freqz(num,den)
z,p,k=signal.tf2zpk(num,den)
#S5: Plotting the result
plt.figure(1),plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),
plt.title('h[n]'),plt.subplot(2,2,2),plt.plot((w/np.pi)*fs/2,20*np.log10(np.abs(H))),
plt.xlabel('$\omega$-->'),plt.ylabel('|H($j\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,2,3),plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole Zero Plot')
plt.plot(np.real(z),np.imag(z),'ko'),plt.plot(np.real(p),np.imag(p),'rx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.subplot(2,2,4),plt.plot(w,np.angle(H)),
plt.xlabel('$\omega$-->'),plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase response')
plt.tight_layout()

Fig. 6.7 Python code to obtain the characteristics of three-point moving average filter

h[n] Magnitude Response

0
Amplitude

|H(jw)_

0.2 –25

–50
0.0
0 25 50 75 100 0 20 40
n--> w -->

Pole Zero Plot Phase Response

1
H(jw)

0
0
jw

–1 –2
–1.0 –0.5 0.0 0.5 1.0 0 1 2 3
s w -->

Fig. 6.8 Characteristics of three-point moving average filter

6.2 Moving Average filter 241

Input Signal
(Square waveform)

Moving average filter Moving average filter Moving average filter

(M = 51) (M = 71) (M = 91)

Output 1 Output 2 Output 3

Fig. 6.9 Problem statement illustration

Inferences
From Fig. 6.8, the following inferences can be drawn:
1. The impulse response of a moving average filter is finite. If the impulse response
of the system is absolutely summable, then the discrete-time system is a stable
system. Thus, three-point moving average system is inherently stable.
2. Moving average filter is an all-zero filter.
3. From the magnitude response, it is possible to observe that three-point moving
average filter act as a lowpass filter.
4. From the phase response, it is possible to conclude that a moving average filter
exhibits linear phase characteristics in the pass band.
Task
1. Repeat the above experiment for a five-point moving average filter and six-point
moving average filter, and comment on the observed output. What change do you
observe in the pole-zero plot for M = 5 and M = 6?
Experiment 6.5 Impact of the Order of Moving Average Filter
The objective of this experiment is to observe the impact of the order of the moving
average filter with respect to the extent of filtering. This objective is shown in
Fig. 6.9. From Fig. 6.9, it is possible to interpret that the input signal to the three
moving average filters of orders 51, 71 and 91 is a square wave. The reason for
choosing square wave as input is that it exhibits sharp transition between ‘ON’ and
‘OFF’ state. The python code, which implements the task shown in Fig. 6.9, is given
in Fig. 6.10, and the corresponding output is shown in Fig. 6.11.
Inferences
The following inferences can be drawn from Fig. 6.11:
1. The input signal to the moving average filter is a square wave. The input signal
exhibits sudden transitions between states ‘0’ and ‘1’.
2. The input signal is passed through 3 moving average filters of order 51, 71 and
91.
242 6 Filter Design Using Pole-Zero Placement Method

#Impact of the order of Moving average filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generating the input square waveform
t=np.linspace(0,1,1000)
x=signal.square(2*np.pi*5*t)
#Step 2: Defining the MA filters
num1=1/51*np.ones(51)
num2=1/71*np.ones(71)
num3=1/91*np.ones(91)
den=[1]
#Step 3: Obtaining the outputs
y1=signal.lfilter(num1,den,x)
y2=signal.lfilter(num2,den,x)
y3=signal.lfilter(num3,den,x)
#Step 4: Plotting the results
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Input signal'),plt.subplot(2,2,2),plt.plot(t,y1)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Filtered signal (M=51)')
plt.subplot(2,2,3),plt.plot(t,y2),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Filtered signal (M=71)'),plt.subplot(2,2,4),plt.plot(t,y3)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Filtered signal (M=91)')
plt.tight_layout()

Fig. 6.10 Python code to obtain the results of moving average filter

3. The square wave is transformed into a triangular wave for the moving average
filter of order 91. The square wave, when passed through an integrator (lowpass
filter), results in a triangular wave. The triangular wave exhibits gradual variation
between the states ‘0’ and ‘1’.
4. The extent of smoothing increases with an increase in the order of the moving
average filter.

6.3 M-Point Exponentially Weighted Moving Average

Filter (EWMA)

The relationship between the input and output of M-point exponentially weighted
moving average filter is given by
6.3 M-Point Exponentially Weighted Moving Average Filter (EWMA) 243

Input signal Filtered signal (M=51)

1 1
Amplitude

Amplitude
0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time Time

Filtered signal (M=71) Filtered signal (M=91)

1 1
Amplitude

Amplitude
0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time Time

Fig. 6.11 Results of moving average filter

M -1
y ½ n] = C αk x½n - k] ð6:11Þ
k=0

In the above expression, ‘C’ is the normalization constant, and ‘α’ is the expo-
nential weighting factor, where 0 < α < 1.
To Find the Expression for Normalization Constant (C) One way to obtain the
value of ‘C’ is that it should preserve the DC gain. If the input is a constant signal (x
[n] = K ), if the filter preserves the DC component of the signal, then the output is
also expected to be ‘K’. Substituting x[n] = K and y[n] = K in Eq. (6.11), we get

M-1
K =C αk K ð6:12Þ
k=0

The above equation can be written as

M -1
K = CK αk ð6:13Þ
k=0

From the above expression, the expression for the constant ‘C’ can be written as
244 6 Filter Design Using Pole-Zero Placement Method

M -1
C αk = 1 ð6:14Þ
k=0

The expression for the constant ‘C’ is written as

1
C= M-1
ð6:15Þ
αk
k=0

Using summation formula

M-1
1 - αM
αk = ð6:16Þ
k=0
1-α

Substituting Eq. (6.16) in Eq. (6.15), we have

1-α
C= ð6:17Þ
1 - αM

The expression for M-point exponentially weighted moving average filter is given
by

M -1
1-α
y ½ n] = α k x½ n - k ] ð6:18Þ
1 - αM k=0

If ‘α’ value is closer to 1, then M-point exponentially weighted moving average

filter will behave like an M-point moving average filter.
Experiment 6.6 Comparing the Impulse Response of MA and EWMA Filter
The objective of this experiment is to compare the impulse responses of the moving
average filter with the exponentially weighted moving average filter for M = 3 and
5. The python code, which plots the impulse response of the moving average filter
and exponentially weighted average filter, is shown in Fig. 6.12, and the
corresponding output is shown in Fig. 6.13.
Inferences
From Fig. 6.13, it is possible to interpret the following:
1. The MA filter gives equal weightage to all the input sample values.
2. The EWMA filter gives more weightage to the current input sample and less
weightage to the past input samples.
6.3 M-Point Exponentially Weighted Moving Average Filter (EWMA) 245

#Impulse response of MA and EWMA

import numpy as np
import matplotlib.pyplot as plt
M1, M2=3,5
alpha=0.5
h1=1/M1*np.ones(M1) #MA filter for M=3
h2=1/M2*np.ones(M2) #MA filter for M=5
C1=(alpha-1)/(alpha**M1-1)
C2=(alpha-1)/(alpha**M2-1)
h3=C1*np.array([1,alpha,alpha**2]) #EWMA filter for M=3
h4=C2*np.array([1,alpha,alpha**2,alpha**3,alpha**4])#EWMA filter for M=5
plt.subplot(2,2,1),plt.stem(h1),plt.xlabel('n-->'),plt.ylabel('$h_1[n]$'),
plt.title('$h_1[n]$ of MA filter for M = 3')
plt.subplot(2,2,2),plt.stem(h3),plt.xlabel('n-->'),plt.ylabel('$h_3[n]$'),
plt.title('$h_3[n]$ of EWMA filter for M = 3')
plt.subplot(2,2,3),plt.stem(h2),plt.xlabel('n-->'),plt.ylabel('$h_2[n]$'),
plt.title('$h_2[n]$ of MA filter for M = 5')
plt.subplot(2,2,4),plt.stem(h4),plt.xlabel('n-->'),plt.ylabel('$h_4[n]$'),
plt.title('$h_4[n]$ of EWMA filter for M = 5'),plt.tight_layout()

Fig. 6.12 Python code which obtains the impulse response of MA and EWMA filter

h1[n] of MA filter for M = 3 h3[n] of EWMA filter for M = 3

0.50
0.2
h1[n]

h3[n]

0.25

0.0 0.00
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
n--> n-->

h2[n] of MA filter for M = 5 h4[n] of EWMA filter for M = 5

0.2
0.4
h2[n]

h4[n]

0.1
0.2

0.0 0.0
0 1 2 3 4 0 1 2 3 4
n--> n-->

Fig. 6.13 Impulse response of MA and EWMA filter

246 6 Filter Design Using Pole-Zero Placement Method

6.4 Digital Resonator

A resonator is designed to have its strongest response to match certain input signal.
Resonators find application in communication receivers, AM/FM demodulators, etc.
A digital resonator is a two-pole bandpass filter with a pair of complex-conjugate
poles near the unit circle to create a resonant peak at the desired frequency. The
digital resonator has a large magnitude response in the vicinity of the pole location. If
one pole is located at p1 = rejω, then the other pole will be at p2 = re-jω, where
0 < r < 1. The expression for the transfer function of the system is given by

1
H ðzÞ = ð6:19Þ
ð1 - p1 z - 1 Þð1 - p2 z - 1 Þ

Substituting p1 = rejω and p2 = re-jω in the above expression, we get

1
H ðzÞ =
ð1 - rejω z - 1 Þð1 - re - jω z - 1 Þ

Simplifying the above expression, we get

1
H ðzÞ =
1 - re - jω z - 1 - rejω z - 1 þ r 2 z - 2

The above equation can be expressed as

1
H ðzÞ =
1 - rz - 1 ½e - jω þ ejω ] þ r2 z - 2

Simplifying the above equation, we get

1
H ðzÞ = ð6:20Þ
1 - 2rz - 1 cosðωÞ þ r 2 z - 2

Experiment 6.7 Digital Resonator

This experiment analyses the concept of digital resonator using python. The python
illustration of digital resonator with two complex conjugate poles occurring at
r = 0.98 is shown in Fig. 6.14, and the corresponding output is shown in Fig. 6.15.
Inferences
The following inferences can be made from this experiment:
1. From the input and output signals, it is possible to observe from Fig. 6.15 that the
input is a unit sample signal. The system is excited with an impulse signal. The
output of the system produces an oscillation.
6.4 Digital Resonator 247

#Digital resonator
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Defining the system
r,fs=0.98,100 #Sampling frequency
fn,fc=fs/2, 5 #Cutoff frequency
w=2*np.pi*(fc/fn)
b=[1]
a=[1,-2*r*np.cos(w),r**2]
#Generating the input and obtaining the response
x=np.zeros(25) #Input to the resonator
x[0]=1
y=signal.lfilter(b,a,x) #Output of resonator
plt.subplot(3,2,1),plt.stem(x),plt.xlabel('n-->'),plt.ylabel('x[n]'),plt.title('Input signal')
plt.subplot(3,2,2),plt.stem(y),plt.xlabel('n-->'),plt.ylabel('y[n]'),plt.title('Output signal')
#Impulse response of the system
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(3,2,3),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('y1[n]'),plt.title('Impulse response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(3,2,5),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx',ms=7.5)
plt.xlabel('Real part'),plt.ylabel('Imaginary part'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, h = signal.freqz(b,a)
plt.subplot(3,2,4),plt.plot(w, 10 * np.log10(abs(h))),plt.xlabel('$\omega$ [rad/sample]'),
plt.ylabel('$|H(e^{j\omega})|$ in [dB]'),plt.title('Magnitude response'),plt.subplot(3,2,6),
plt.plot(w,np.unwrap(np.angle(h))),plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Degree')
plt.title('Phase response'),plt.tight_layout()

Fig. 6.14 Python code to implement digital resonator

2. From the magnitude response, it is possible to observe that the system behaves
like a narrow bandpass filter.
3. From the pole-zero plot, it is possible to observe that two complex conjugate
poles occur very closer to the unit circle.
4. From the phase response, it is possible to observe that the system exhibits
non-linear phase characteristics.
248 6 Filter Design Using Pole-Zero Placement Method

Input signal Output signal

1 1
x[n]

y[n]
0
0 –1
0 10 20 0 10 20
n--> n-->

Impulse response Magnitude response

|H(ejw)| in [dB]
1 10
y1[n]

0
–1 0
0 10 20 0 1 2 3
n--> w>rad/sample]
Imaginary part

Pole-zero plot Phase response

1
Degree 0
0
–1 –2
–1.0 –0.5 0.0 0.5 1.0 0 1 2 3
Real part w>rad/sample]

Fig. 6.15 Result of python code shown in Fig. 6.14

Task
1. Repeat the above experiment for r = 1 and comment on the observed result.

6.5 Notch Filter

Notch filter has two complex conjugate zeros placed on the unit circle to create a null
at a desired frequency. A notch filter has the ability to reject one particular frequency.
A pair of complex conjugate zeros on the unit circle produces a null in the frequency
response, which results in the rejection of one particular frequency. Let the conjugate
zeros be represented as z1 = rejω and z2 = re-jω. If the zeros occur on the unit circle,
then r = 1. The transfer function of such a system is given by

H ð z Þ = 1 - z1 z - 1 1 - z 2 z - 1 ð6:21Þ

Substituting z1 = ejω and z2 = e-jω in the above expression, we get

H ðzÞ = 1 - ejω z - 1 1 - e - jω z - 1

The above equation can be written as

6.5 Notch Filter 249

Input signal Notch filter with 5 Hz Output signal

cut-off frequency
Addition of 2 Hz and 5 Hz Sine wave of 2 Hz
sine waves frequency

Fig. 6.16 Problem illustration

H ðzÞ = 1 - e - jω z - 1 - ejω z - 1 þ z - 2

The transfer function of the system is expressed as

H ðzÞ = 1 - z - 1 e - jω þ ejω þ z - 2

The above equation can be expressed as

H ðzÞ = 1 - 2z - 1 cosðωÞ þ z - 2 ð6:22Þ

Experiment 6.8 Notch Filter

This experiment discusses the design of a notch filter to eliminate one particular
frequency component. The input signal to the notch filter is the addition of two sine
waves of frequency components, 2 and 5 Hz. The notch filter cut-off frequency is
5 Hz. It is expected that the notch filter will eliminate 5 Hz frequency component so
that the filtered signal will have only 2 Hz frequency component. This is illustrated in
Fig. 6.16.
The python code which performs the above-mentioned task is shown in Fig. 6.17,
and the corresponding output is shown in Fig. 6.18.
Inferences
From Fig. 6.18, the following inferences can be drawn
1. The input signal to the notch filter is an addition of 2 and 5 Hz sine waves.
2. The output of the notch filter clearly shows that it is a 2 Hz sine wave. This means
that the notch filter has filtered 5 Hz sinusoidal component only.
3. The impulse response shows that the filter designed has a finite impulse response.
4. From the magnitude response, it is possible to observe that the notch occurs at
5 Hz.
5. From the pole-zero plot, it is possible to observe two conjugate zeros.
6. From the phase response, it is possible to observe that the phase response varies
linearly with respect to frequency. Therefore, the designed filter exhibits linear
phase characteristics.
Experiment 6.9 Design of Notch Filter Using Built-In Function
The built-in function ‘iirnotch’ available in ‘scipy’ library can be used to design a
notch filter. The python code to design a notch filter for cut-off frequency of 50 Hz
250 6 Filter Design Using Pole-Zero Placement Method

#Notch filter
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generating the input signal
f1,f2=2,5 # 1Hz and 5Hzfrequency component
n=np.arange(0,100)
x1=np.sin(2*np.pi*f1*n/100)
x2=np.sin(2*np.pi*f2*n/100)
x=x1+x2 #Input signal has 1 Hz and 5 Hz component
#Step 2: Design of notch filter
r,fs,fc=0.99,100,5 # Sampling, Cutoff frequencies
w=2*np.pi*(fc/fs)
b=[1,-2*np.cos(w),1]
a=[1]
#Step 3: Obtaining the output
y=signal.lfilter(b,a,x) #Output of resonator
plt.subplot(3,2,1),plt.stem(x),plt.xlabel('n-->'),plt.ylabel('x[n]'),plt.title('Input signal'),
plt.subplot(3,2,2),plt.stem(y),plt.xlabel('n-->'),plt.ylabel('y[n]'),plt.title('Output signal')
#Impulse response of the system
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(3,2,3),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('y1[n]'),plt.title('Impulse response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(3,2,5),plt.plot(circle.real, circle.imag, 'k--'),plt.plot(z.real, z.imag, 'ro', ms=7.5),
plt.plot(p.real, p.imag, 'rx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, h = signal.freqz(b,a)
plt.subplot(3,2,4),plt.plot(0.5*fs*w/np.pi, 10 * np.log10(abs(h)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('$|H(j\omega)|$ in [dB]')
plt.title('Magnitude response'),plt.subplot(3,2,6),
plt.plot(0.5*fs*w/np.pi,np.unwrap(np.angle(h))),plt.xlabel('$\omega$ [rad/sample]'),
plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase response'),plt.tight_layout()

Fig. 6.17 Python code which performs notch filtering of the input signal

and a sampling frequency of fs = 1000 Hz is shown in Fig. 6.19, and the

corresponding characteristics are shown in Fig. 6.20.
Inferences
From Fig. 6.20, it is possible to observe the following facts:
1. The impulse response is of infinite duration. The impulse response is not
symmetric.
6.6 All-Pass Filter 251

Input signal Output signal

2
0 0.25
x[n]

y[n]
–2 0.00
0 25 50 75 100 0 25 50 75 100
n--> n-->

Impulse response Magnitude response

|H(jw)| in [dB]
0 0
y1[n]

–25
–2
0 10 20 0 20 40
n--> w [rad/sample]

Pole-zero plot Phase response

1 H(jw) 2.5
0
jw

–1 0.0
–1.0 –0.5 0.0 0.5 1.0 0 20 40
s w [rad/sample]

Fig. 6.18 Result of the python code shown in Fig. 6.17

2. The pole-zero plot shows two complex conjugate poles and zeros on the unit
circle. The presence of poles on the unit circle indicates that the stability of the
filter is not guaranteed.
3. The magnitude response shows that the 50 Hz notch frequency is the cut-off
frequency.
4. The phase response indicates that it is non-linear. Since the impulse response is
not symmetric, the phase response is not linear.

6.6 All-Pass Filter

The transfer function of first-order all-pass filter is given by

z-1 - a
H ðzÞ = , where jaj < 1 ð6:23Þ
1 - az - 1

The frequency response of the system is obtained by substituting z = ejω in the

above equation, we get

e - jω - a
H ejω = ð6:24Þ
1 - ae - jω
252 6 Filter Design Using Pole-Zero Placement Method

#Characteristics of notch filter

import matplotlib.pyplot as plt
import numpy as np
from scipy import signal
#Step 1: Design of notch filter
fs, fc, QF = 1000, 50, 10 # Sampling, Cut-off frequency and Quality factor
w0=fc/(fs/2)
b, a = signal.iirnotch(w0, QF)
#Step 2: Plotting the characteristics
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]'),plt.title('Impulse response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 100)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'gx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, H = signal.freqz(b,a)
plt.subplot(2,2,3),plt.plot(w/np.pi*fs/2, 10*np.log(abs(H)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('$|H(j\omega)|$ in [dB]')
plt.title('Magnitude response'),plt.subplot(2,2,4),plt.plot(w/np.pi*fs/2,np.unwrap(np.angle(H)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('$\u2220H(j\omega)$')
plt.title('Phase response'),plt.tight_layout()

Fig. 6.19 Python code to design notch filter using built-in function

Using Euler’s formula e-jω = cos (ω) - j sin (ω), the above expression can be
written as

cosðωÞ - j sinðωÞ - a
H ejω = ð6:25Þ
1 - a½cosðωÞ - j sinðωÞ]

Now the expression for squared magnitude response is given by

2 ðcosðωÞ - aÞ2 þ sin 2 ðωÞ

H ejω =
ð1 - a cosðωÞÞ2 þ a2 sin 2 ðωÞ

The above equation can be simplified as

2 cos 2 ðωÞ þ a2 - 2a cosðωÞ þ sin 2 ðωÞ

H ejω =
1 þ a2 cos 2 ðωÞ - 2a cosðωÞ þ a2 sin 2 ðωÞ

Using the fact that sin2(ω) + cos2(ω) = 1, the above equation can be simplified as
6.6 All-Pass Filter 253

Impulse response Pole-zero plot

1.0 1
h[n]

0.5 0

jw
0.0 –1
0 10 20 –1.0 –0.5 0.0 0.5 1.0
n--> s

Magnitude response Phase response

0
1
|H(jw)| in [dB]

–10

H(jw)
0

–20 –1

0 200 400 0 200 400

w [rad/sample] w [rad/sample]

Fig. 6.20 Characteristics of a notch filter

2 1 þ a2 - 2a cosðωÞ
H ejω =
1þ a2 cos 2 ðωÞ þ sin 2 ðωÞ - 2a cosðωÞ

Upon simplifying the above expression, we get

2 1 þ a2 - 2a cosðωÞ
H ejω = =1 ð6:26Þ
1 þ a2 - 2a cosðωÞ

Thus, the magnitude response of the all-pass filter is unity. This means all-pass
filters pass all frequency components of the input signal.
Experiment 6.10 All-Pass Filter
This experiment discusses the python implementation of all-pass filter. The python
implementation of first-order all-pass filter with the value of ‘a = 0.5’ is shown in
Fig. 6.21, and the corresponding output is shown in Fig. 6.22.
Inferences
From Fig. 6.22, the following inferences can be drawn:
1. The input to the all-pass filter is a sine wave of 5 Hz frequency.
2. The output of the all-pass filter is almost the same as the input signal. Thus, all-
pass filter passes all the frequency components of the input signal.
3. The impulse response is not finite. It slowly reaches the value of zero. Hence, it is
an IIR filter.
254 6 Filter Design Using Pole-Zero Placement Method

#First order all-pass filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generating the input signal
f=5 # frequency
n=np.arange(0,50)
x=np.sin(2*np.pi*f*n/50)
#Step 2: Design of all-pass filter
a1=0.5
b,a=[-a1,1],[1,-a1]
#Step 3: Obtaining the output
y=signal.lfilter(b,a,x) #Output of resonator
plt.subplot(3,2,1),plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('x[n]'),plt.title('Input signal')
plt.subplot(3,2,2),plt.stem(n,y),plt.xlabel('n-->'),plt.ylabel('y[n]'),plt.title('Output signal')
#Impulse response of the system
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(3,2,3),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]'),plt.title('Impulse
response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(3,2,4),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'gx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, h = signal.freqz(b,a)
plt.subplot(3,2,5),plt.plot(w,np.abs(h))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('$|H(j\omega)|$')
plt.title('Magnitude response'),plt.subplot(3,2,6),
plt.plot(w,np.unwrap(np.angle(h))),plt.xlabel('$\omega$ [rad/sample]'),
plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase response'),plt.tight_layout()

Fig. 6.21 First-order all-pass filter

4. The magnitude response indicates that the filter gain is one for all frequency
components.
5. From the pole-zero plot, a pole lies at 0.5, whereas a zero lies at 2. The given
system is not a minimum phase system.
6. The phase response of the system is non-linear. The phase is not varying linearly
with respect to frequency.
Task
1. Repeat the above experiment for a = 0.25 and a = 0.75, and comment on the
observed results.
6.7 Comb Filter 255

Input signal Output signal

1 1
0
x[n]

y[n]
0
–1 –1
0 20 40 0 20 40
n--> n-->

Impulse response Pole-zero plot

1
0.5
h[n]

jw
0.0
–0.5 –1
0 10 20 –1 0 1 2
n--> s
Magnitude response Phase response
1.05 0.0
‘H(jw)
|H(jw)|

1.00
–2.5
0.95
0 1 2 3 0 1 2 3
w [rad/sample] w [rad/sample]

Fig. 6.22 Result of first-order all-pass filter

Fig. 6.23 Block diagram x[n] y[n]

representation of comb filter

z-L

6.7 Comb Filter

A comb filter is a notch filter with a number of equally spaced nulls. The block
diagram representing the comb filter is shown in Fig. 6.23.
The relationship between the input and output of the comb filter is expressed as

y½n] = x½n] þ x½n - L] ð6:27Þ

Taking Z-transform on both sides of the above equation, we get

Y ðzÞ = X ðzÞ þ z - L X ðzÞ ð6:28Þ

The above equation can be expressed as

Y ðzÞ = X ðzÞ 1 þ z - L

The transfer function can be expressed as

256 6 Filter Design Using Pole-Zero Placement Method

Y ðz Þ
H ðzÞ = = 1 þ z-L ð6:29Þ
X ð zÞ

The frequency response of the system is obtained by substituting z = ejω in the

above expression, we get

H ejω = 1 þ e - jωL ð6:30Þ

6.7.1 Location of Poles and Zeros of Comb Filter

From the expression of the transfer function given in Eq. (6.30), it is possible to
interpret; there is a pole of multiplicity ‘L’ at the origin. The location of zeros is
obtained by equating the numerator of the transfer function to zero, which results in

z-L = - 1

The above equation can be expressed as

e - jωL = ejð2kþ1Þπ

From the above expression

ð2k þ 1Þπ
ωk = ð6:31Þ
L

The zeros of the FIR filter are uniformly spaced 2π L radians apart around the unit
circle starting at ω = Lπ . For odd ‘L’, there is a zero at ω = π.
Experiment 6.11 Comb Filter
The objective of this experiment is to plot the pole-zero pattern of Comb filter for
even and odd values of L. The odd value is chosen as L = 5, and the even value is
chosen as L = 6. The python code which plots the pole-zero plot of comb filters is
shown in Fig. 6.24, and the corresponding output is shown in Fig. 6.25.
Inferences
From the pole-zero plot, which is shown in Fig. 6.25, the following inferences can
be made:
1. The zeros are uniformly spaced 2π L radians apart around the unit circle.
2. For odd values of ‘L’, there is a zero at ω = π.
3. The poles lie at the origin, which implies that the filters are inherently stable.
6.7 Comb Filter 257

#Pole-zero plot of comb filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
h1=[1,0,0,0,0,1] # Comb filter for N=5 (Odd)
h2=[1,0,0,0,0,0,1] #Comb filter for N=6 (Even)
z1, p1, k1 = signal.tf2zpk(h1,1) #Pole-zero for N=5
z2, p2, k2 = signal.tf2zpk(h2,1) #Pole-zero for N=6
#Plotting the pole-zero plot
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,1),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z1.real, z1.imag, 'ro', ms=7.5)
plt.plot(p1.real, p1.imag, 'gx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),
plt.title('Pole-zero plot for L = 5'),plt.grid()
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z2.real, z2.imag, 'ro', ms=7.5)
plt.plot(p2.real, p2.imag, 'gx', ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),
plt.title('Pole-zero plot for L = 6'),plt.grid(),plt.tight_layout()

Fig. 6.24 Pole-zero plot of comb filter

Pole-zero plot for L = 5

1

0
jZ

–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
V

Pole-zero plot for L = 6

1

0
jZ

–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
V

Fig. 6.25 Pole-zero plot of comb filters for odd and even values of ‘L’
258 6 Filter Design Using Pole-Zero Placement Method

Clean Noisy signal Moving Filtered signal

sinusoidal average filter
signal

Random signal

Fig. 6.26 Block diagram representation of problem statement

Clean ECG Noisy signal Notch Filtered signal

signal filter

50 Hz sinusoidal
signal

Fig. 6.27 Block diagram of problem statement

Exercises
1. Generate sinusoidal signal of 5 Hz frequency. Add random noise which follows
uniform distribution to the clean sinusoidal signal. Pass the noisy signal to the
moving average filter and comment on the observed result. The block diagram of
the problem statement is shown in Fig. 6.26.
2. Generate three tones of frequencies 500, 1000 and 1500 Hz. Append the three
tones together as one signal. Now pass this signal to a notch filter, which will
block the frequency component of a specific frequency (say 1000 Hz). Hear the
input and output signals and comment on your observation.
3. Design a notch filter to minimize 50 Hz powerline interference in ECG signal.
Read an ECG signal which is stored in ‘mat’ file format. Add 50 Hz powerline
interference to the clean ECG signal to generate noisy signal. Pass the noisy
signal to the notch filter to minimize the powerline interference. Plot the clean,
noisy and filtered signals and comment on the observed result. The problem
statement is depicted in the form of a block diagram and is shown in Fig. 6.27.
4. Generate 10 Hz square waveform. Design a comb filter to eliminate 10 Hz
frequency component in this signal and its odd harmonics.
5. Generate sine wave of frequencies 5 and 10 Hz. Add these two waveforms. Now
pass this signal through a notch filter, which should eliminate the 5 Hz frequency
component, so that the output signal contains a 10 Hz frequency component.
6.7 Comb Filter 259

Objective Questions
1. The filter which is used to reject one particular frequency is
A. Lowpass filter
B. Highpass filter
C. All-pass filter
D. Notch filter
2. The filter which can be used as a delay equalizer is
A. Lowpass filter
B. Highpass filter
C. All-pass filter
D. Notch filter
3. Cascading of lowpass and highpass filter will result in
A. Lowpass filter
B. Highpass filter
C. Band pass filter
D. All-pass filter
4. The filter which is used to minimize the impact of power line interference is
A. Lowpass filter
B. Highpass filter
C. All-pass filter
D. Notch filter
5. The impulse response of three-point moving average filter is given by
h½ n] = 3 f - δ½n] - δ½n - 1] - δ½n - 2g
1
A.
B. h½ n] = 3 f - δ½n] - δ½n - 1] þ δ½n - 2g
1

C. h½ n] = 3 fδ½n] - δ½n - 1] þ δ½n - 2g

1

D. h½ n] = 3 f δ ½ n ] þ δ ½ n - 1 ] þ δ ½ n - 2g
1

6. The filter which has the ability to remove fundamental frequency and its
harmonics is
A. Notch filter
B. Comb filter
C. Lowpass filter
D. Highpass filter
7. The impulse response of a digital filter is given by h[n] = δ[n] + δ[n - 8]. The
filter behaves like a
A. All-pass filter
B. Highpass filter
260 6 Filter Design Using Pole-Zero Placement Method

C. Comb filter
D. Notch filter
8. Statement 1: Stable filters are always causal
Statement 2: Causal filters are always stable:
A. Both statements are wrong.
B. Both statements are true.
C. Statement 1 is true, and Statement 2 is wrong.
D. Statement 1 is wrong, and Statement 2 is true.
9. Assertion: Moving average filter attenuates quick change in the signal.
Reason: Moving average filter is a lowpass filter.
A. Both assertion and reason are wrong.
B. Assertion is true, reason is wrong.
C. Assertion is wrong, reason may be true.
D. Both assertion and reason are true.
- jω - j2ω - j3ω
10. The frequency response of lowpass filter is given by H ðejω Þ = 1þe þe4 þe .
Using frequency shift, the lowpass filter can be converted to a highpass filter.
The impulse response of the highpass filter is
A. h[n] = 0.25{δ[n] + δ[n - 1] + δ[n - 2] + δ[n - 3]}
B. h[n] = 0.25{δ[n] + δ[n - 1] - δ[n - 2] - δ[n - 3]}
C. h[n] = 0.25{-δ[n] - δ[n - 1] - δ[n - 2] - δ[n - 3]}
D. h[n] = 0.25{δ[n] - δ[n - 1] + δ[n - 2] - δ[n - 3]}
11. Statement 1: Digital resonator generates sinusoidal signal of specific frequency.
Statement 2: Digital resonator has complex conjugate pole located on the unit
circle.
A. Statement 1 is correct, and Statement 2 is wrong.
B. Statement 1 is wrong, and Statement 2 is correct.
C. Both Statements 1 and 2 are correct.
D. Both Statements 1 and 2 are wrong.
12. In the design of a simple digital filter using pole-zero placement:
Statement 1: To suppress a frequency component, locate a zero at this
frequency on the unit circle.
Statement 2: To amplify a frequency, locate a pole at this frequency inside the
unit circle.
A. Statement 1 is correct, and Statement 2 is wrong.
B. Statement 1 is wrong, and Statement 2 is correct.
C. Both Statements 1 and 2 are wrong.
D. Both Statements 1 and 2 are correct.
13. The transfer function of a linear time-invariant system is expressed as
H ðzÞ = BAððzzÞÞ.
Bibliography 261

Statement 1: Zeros are roots of the polynomial B(z).

Statement 2: Poles are roots of the polynomial A(z).
A. Statement 1 is correct, and Statement 2 is wrong.
B. Statement 1 is wrong, and Statement 2 is correct.
C. Both Statements 1 and 2 are wrong.
D. Both Statements 1 and 2 are correct.

Bibliography

1. Ashok Ambardar, “Digital Signal Processing: A Modern Introduction”, Cengage Learning

India, 2007.
2. Sanjit K. Mitra, “Digital Signal Processing: A Computer-Based Approach”, McGraw Hill
Education, 2013.
3. Boaz Porat, “A Course in Digital Signal Processing”, John Wiley and Sons, 1996.
4. Dimitris G. Manolakis, Vinay K. Ingle, “Applied Digital Signal Processing: Theory and Prac-
tice”, Cambridge University Press, 2011.
5. L. B. Jackson, “Signals, Systems and Transforms”, Addison-Wesley, 1991.
Chapter 7
FIR Filter Design

Learning Objectives
After completing this chapter, the reader is expected to
• Analyse the characteristics of Type-I, Type-II, Type-III and Type-IV FIR filters.
• Design and analyse window-based finite impulse response filter.
• Design and analyse frequency sampling based finite impulse response filter.
• Design and analyse optimal finite impulse response filter.

Roadmap of the Chapter

This chapter discusses the type of FIR filters and its characteristic. Also, it gives
detail about the designs of the FIR filter. The roadmap of this chapter is given in the
form of flowchart below.

FIR filter

Types of FIR filter Design of FIR filter

Type – I FIR filter Window-based FIR

filter design
Type – II FIR filter
Frequency sampling based
FIR filter design
Type – III FIR filter

Optimal FIR filter design

Type – IV FIR filter

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 263
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_7
264 7 FIR Filter Design

PreLab Questions
1. What is the difference equation relating the input and output of a finite impulse
response filter? What are the inferences that could be made from the difference
equation?
2. On what basis are FIR filters classified as Type-I, Type-II, Type-III and Type-IV
FIR filters?
3. When is a FIR filter coefficient said to exhibit (a) even symmetry and (b) odd
symmetry?
4. What is the condition for the digital filter to exhibit linear phase characteristics?
5. What is the advantage of ‘linear phase’ characteristics of digital filter?
6. What is the relationship between the group delay and the phase response of the
FIR filter?
7. What is the relationship between the order (M ) and the number of coefficients
(N ) of FIR filters?
8. Why FIR filter is considered as an ‘inherently stable’ filter?
9. List four advantages of FIR filter.
10. Mention different methods of design of FIR filter.

7.1 FIR Filter

FIR stands for finite impulse response. The coefficients of FIR filter are either
symmetric or anti-symmetric. Due to the symmetric nature of FIR filter coefficients,
it exhibits linear phase characteristics. Because of linear phase characteristics, the
FIR filter has no phase distortion. Also, FIR filter exhibits constant group delay. In
FIR filter, the current output is a function of the current and previous inputs. This
implies that FIR filters are non-recursive filters; hence, they are inherently stable.
FIR filter is an all-zero filter, and the zeros occur in conjugate reciprocal pair.

7.2 Classification of FIR Filter

Based on nature of symmetry and the number of coefficients, FIR filter can be
classified as Type-I, Type-II, Type-III and Type-IV. The classification is given in
Table 7.1.

Table 7.1 Classification of FIR filter

S. No. Nature of symmetry Number of coefficients Type of FIR filter
1 Even symmetry Odd Type-I
2 Even symmetry Even Type-II
3 Odd symmetry Odd Type-III
4 Odd symmetry Even Type-IV
7.2 Classification of FIR Filter 265

#Type-I FIR filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1 : Type-I FIR filter
b,a=[1,2,3,2,1],[1]
#Impulse response of the system
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Impulse response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--'),plt.plot(z.real, z.imag, 'ro', ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, h = signal.freqz(b,a)
plt.subplot(2,2,3),plt.plot(w, np.abs(h)),plt.xlabel('$\omega$-->'),
plt.ylabel('|H(j$\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,2,4),plt.plot(w,np.unwrap(np.angle(h))), plt.xlabel('$\omega$-->'),
plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase response'),plt.tight_layout()

Fig. 7.1 Python code to obtain the characteristics of Type-I FIR filter

Experiment 7.1 Characteristics of Type-I FIR Filter

The objective of this experiment is to plot the characteristics of Type-I FIR filter.
Here, the impulse response of Type-I FIR filter is chosen as h[n] = {1, 2, 3, 2, 1}.
The filter coefficients satisfy even symmetry, and the number of coefficients is odd;
hence, it belongs to Type-I FIR filter. The python code, which obtains the impulse
response, magnitude response, phase response and pole-zero plot, is shown in
Fig. 7.1, and the corresponding output is shown in Fig. 7.2. The built-in functions
used in the program are given in Table 7.2.
Inferences
From Fig. 7.2, the following inferences can be drawn:
1. From the impulse response, it is possible to observe that the impulse response is
of finite duration, and the filter coefficients exhibit even symmetry.
2. From the pole-zero plot, the filter is an all-zero filter. The zeros occur in
conjugate pair.
3. From the magnitude response, it is possible to observe that the filter behaves like a
lowpass filter.
4. From the phase response, it is possible to observe that the filter exhibits a linear
phase characteristic in the passband.
266 7 FIR Filter Design

Impulse response Pole-zero plot

3 1
Amplitude

2
0

jw
1

0 –1
0 10 20 –1.0 –0.5 0.0 0.5 1.0
n--> s

Magnitude response Phase response

0
7.5
–2

‘H(jw)
|H(jw)|

5.0
–4
2.5

0.0 –6
0 1 2 3 0 1 2 3
w --> w -->

Fig. 7.2 Characteristics of Type-I FIR filter

Table 7.2 Built-in functions used in the program

Built-in
S. No. function Library Application
1 signal. Scipy To obtain the output of LTI system
lfilter
2 signal. Scipy To obtain the zeros, poles and gain of the LTI system
tf2zpk
3 signal. Scipy To obtain the frequency response of the LTI system. Frequency
freqz response is a combination of magnitude and phase responses
4 abs Numpy To obtain the magnitude response of the system
5 angle Numpy To obtain the phase response of the system

Task
1. Write a python code to illustrate that Type-I FIR filter is versatile; (i.e.) it can be
used as a lowpass, highpass, bandpass and band reject filter.
Experiment 7.2 Characteristics of Type-II FIR Filter
Type-II FIR filter exhibits even symmetry with an even number of coefficients. The
Type-II FIR filter coefficients considered in this experiment is h[n] = {1, 2, 2, 1}.
The python code to obtain the impulse response, pole-zero plot, magnitude and
phase responses is shown in Fig. 7.3, and the corresponding output is shown in
Fig. 7.4.
Inferences
From Fig. 7.4, the following inferences can be made:
7.2 Classification of FIR Filter 267

#Type-II FIR filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1 : Type-I FIR filter
b,a=[1,2,2,1],[1]
#Impulse response of the system
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Impulse response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--'),plt.plot(z.real, z.imag, 'ro', ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, h = signal.freqz(b,a)
plt.subplot(2,2,3),plt.plot(w, np.abs(h)),plt.xlabel('$\omega$--
>'),plt.ylabel('|H(j$\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,2,4),plt.plot(w,np.unwrap(np.angle(h))), plt.xlabel('$\omega$-->'),
plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase response'),plt.tight_layout()

Fig. 7.3 Python code to obtain the characteristics of Type-II FIR filter

Impulse response Pole-zero plot

2 1
Amplitude

1 0
jZ

0 –1
0 10 20 –1.0 –0.5 0.0 0.5 1.0
n--> s

Magnitude response Phase response

6 0

–1
‘H(jw)

4
|H(jw)|

2 –2

0 –3
0 1 2 3 0 1 2 3
w --> w -->

Fig. 7.4 Characteristics of Type-II FIR filter

268 7 FIR Filter Design

#Type-III FIR filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1 : Type-I FIR filter
b,a=[1,2,0,-2,-1],[1]
#Impulse response of the system
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Impulse response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--'),plt.plot(z.real, z.imag, 'ro', ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, h = signal.freqz(b,a)
plt.subplot(2,2,3),plt.plot(w, np.abs(h)),plt.xlabel('$\omega$-->'),
plt.ylabel('|H(j$\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,2,4),plt.plot(w,np.unwrap(np.angle(h))), plt.xlabel('$\omega$-->'),
plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase response'),plt.tight_layout()

Fig. 7.5 Python code to obtain the characteristics of Type-III FIR filter

1. From the impulse response, it is possible to observe that the impulse response is
of finite duration, and it exhibits even symmetry.
2. From the pole-zero plot, it is possible to observe that Type-II FIR filter is an
all-zero filter. The magnitude which is zero at ω = π indicates that Type-II FIR
filter cannot be used as a highpass filter.
3. From the magnitude response, it is possible to conclude that the filter behaves like
a lowpass filter.
4. From the phase response, it is possible to confirm that the Type-II FIR filter
exhibits linear phase characteristics in the passband.
Task
1. Generate x[n] = ejπn, 0 < n ≤ 100. Pass this signal through Type-II FIR filter
whose impulse response is h[n] = {1, 2, 2, 1}. Use subplot to plot the input and
output signals and comment on the observed result.
Experiment 7.3 Characteristics of Type-III FIR Filter
This experiment discusses the analysis of the characteristics of Type-III FIR filter
using python. The python code, which obtains the characteristics of Type-III FIR
filter, is shown in Fig. 7.5, and the corresponding output is shown in Fig. 7.6. The
7.2 Classification of FIR Filter 269

Impulse response Pole-zero plot

2 1
Amplitude

jw
0

–2 –1
0 10 20 –1.0 –0.5 0.0 0.5 1.0
n--> s

Magnitude response Phase response

4 0

‘H(jw)
|H(jw)|

2 –2

–4
0
0 1 2 3 0 1 2 3
w --> w -->

Fig. 7.6 Characteristics of Type-III FIR filter

impulse response chosen for Type-III FIR filter characteristics is {1, 2, 0, -2, -1},
and it satisfies both odd symmetry and the number of coefficients is odd.
Inferences
From Fig. 7.6, the following inferences can be drawn:
1. The impulse response shows odd symmetry with an odd number of coefficients.
The duration of the impulse response is finite.
2. From the pole-zero plot, it is possible to observe that the magnitude value is zero
at ω = 0 and ω = π. This implies that Type-III FIR filter cannot be used as a
lowpass and a highpass filters.
3. From the magnitude response, it is possible to observe that the filter can act as a
bandpass filter only.
4. From the phase response, it is possible to infer that the filter exhibits linear phase
characteristics in the passband.
Task
1. Write a python code to illustrate the fact that cascading of lowpass and highpass
filters will result in a bandpass filter.
Experiment 7.4 Characteristics of Type-IV FIR Filter
This experiment tries to analyse the characteristics of Type-IV FIR filter using
python. The python code, which obtains the characteristics of Type-IV FIR filter,
is shown in Fig. 7.7, and the corresponding output is shown in Fig. 7.8. The impulse
response chosen for this illustration is h[n] = {1, 2, -2, -1}. The impulse response
exhibits odd symmetry with an even number of filter coefficients.
270 7 FIR Filter Design

#Type-IV FIR filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1 : Type-I FIR filter
b,a=[1,2,-2,-1],[1]
#Impulse response of the system
h_1=np.zeros(25)
h_1[0]=1
h=signal.lfilter(b,a,h_1)
plt.subplot(2,2,1),plt.stem(h),plt.xlabel('n-->'),
plt.ylabel('Amplitude'),plt.title('Impulse response')
# Pole-zero plot
z, p, k = signal.tf2zpk(b, a)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--'),plt.plot(z.real, z.imag, 'ro', ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot'),plt.grid()
#Magnitude and phase response
w, h = signal.freqz(b,a)
plt.subplot(2,2,3),plt.plot(w, np.abs(h)),plt.xlabel('$\omega$-->'),
plt.ylabel('|H(j$\omega$)|'),plt.title('Magnitude response')
plt.subplot(2,2,4),plt.plot(w,np.unwrap(np.angle(h))), plt.xlabel('$\omega$-->'),
plt.ylabel('$\u2220H(j\omega)$'),plt.title('Phase response'),plt.tight_layout()

Fig. 7.7 Python code to obtain the characteristics of Type-IV FIR filter

Impulse response Pole-zero plot

2 1
Amplitude

jw

0 0

–2 –1
0 10 20 –2 –1 0 1
n--> s

Magnitude response Phase response

4
‘H(jw)

0
|H(jw)|

2
–2
0
0 1 2 3 0 1 2 3
w --> w -->

Fig. 7.8 Characteristics of Type-IV FIR filter

7.3 Design of FIR Filter 271

Inferences
From Fig. 7.8, the following inferences can be made:
1. The impulse response plot reveals that Type-IV FIR filter impulse response is
anti-symmetric with an even number of coefficients.
2. From the pole-zero plot, it is possible to confirm that Type-IV FIR filter has a zero
at ω = 0; hence, it cannot be used as a lowpass filter.
3. The magnitude response resembles that of a bandpass filter.
4. The phase response plot reveals that Type-IV FIR filter exhibits linear phase
characteristics in the passband.
Experiment 7.5 Comparison of Type-I, Type-II, Type-III and Type-IV FIR
Filters with Respect to Their Location of Zeros
This experiment compares all four types of FIR filters with respect to their location
of zeros using python. The python code used to plot the pole-zero plot of Type-I,
Type-II, Type-III and Type-IV FIR filters is shown in Fig. 7.9, and the
corresponding output is shown in Fig. 7.10. In this program, the impulse responses
of four FIR filters are defined first. Then, the pole, zero and gain of each type of FIR
filter are obtained using the built-in function ‘tf2zpk’ available in ‘scipy’ package.
Then, the extracted poles and zeros are plotted. The automatic location of zeros in
Type-I, Type-II, Type-III and Type-IV FIR filters is given in Table 7.3.
Inferences
From Fig. 7.10, the following inferences can be drawn:
1. For Type-I FIR filter, there is no zero at ω = 0 and ω = π. It can be used as a
versatile filter.
2. Type-II FIR filter has a zero at ω = π. It cannot be used as a highpass filter.
3. Type-III FIR filter has zero at ω = 0 and ω = π. It cannot be used as both lowpass
and highpass filters.
4. Type-IV FIR filter has zero at ω = 0. It cannot be used as lowpass filter.
5. In general, zeros of FIR filter occur in conjugate, reciprocal pairs.
Task
1. Write a python code to illustrate the fact that all four types of FIR filters are
inherently stable filters. Hint: For a discrete-time system to be stable, the impulse
response should be absolutely summable.

7.3 Design of FIR Filter

The design of FIR filter starts with specification. The specification can be either in
time domain or frequency domain. In time domain, the desired impulse response is
given as specification. In frequency domain, the specification involves magnitude
and phase response. Three prominent methods to design FIR filters are (1) window-
based method, (2) frequency sampling method and (3) optimal method.
272 7 FIR Filter Design

#Pole-zero plot of different types of FIR filter

import matplotlib.pyplot as plt
import numpy as np
from scipy import signal
#Defining four types of FIR filter
h1,h2,h3,h4=[1,2,5,2,1],[1,2,2,1],[1,2,0,-2,-1],[1,2,-2,-1]
#Poles and zeros of the filter
z1, p1, k1 = signal.tf2zpk(h1,1)
z2, p2, k2 = signal.tf2zpk(h2,1)
z3, p3, k3 = signal.tf2zpk(h3,1)
z4, p4, k4 = signal.tf2zpk(h4,1)
#Ploting the pole-zero plot
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,1),plt.plot(circle.real, circle.imag, 'k--'),plt.plot(z1.real, z1.imag, 'ro', ms=7.5)
plt.plot(p1.real, p1.imag, 'rx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot(Type-I)'),plt.grid()
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z2.real, z2.imag, 'ro', ms=7.5),plt.plot(p2.real, p2.imag, 'gx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot(Type-II)'),plt.grid()
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,3),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z3.real, z3.imag, 'ro', ms=7.5),plt.plot(p3.real, p3.imag, 'gx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot(Type-III)'),plt.grid()
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,4),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z4.real, z4.imag, 'ro', ms=7.5),plt.plot(p4.real, p4.imag, 'gx',ms=7.5)
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot(Type-IV)'),plt.grid()
plt.tight_layout()

Fig. 7.9 Python code to obtain the pole-zero plots of different types of FIR filter

7.3.1 Steps in Window-Based FIR Filter Design

The steps followed in FIR filter design using Windows are summarized below:
1. The filter design starts with the specification of the filter in terms of desired
frequency response.
2. The desired impulse response (hd[n]) is obtained from the desired frequency
response using inverse discrete-time Fourier transform.
3. Multiply the desired impulse response with the selected window function.
4. Delay the windowed impulse response by a factor of ‘τ’ to get the causal FIR filter
coefficients.
5. The process is complete if the frequency response is satisfied as per the specifi-
cation. If the frequency specifications are not satisfied, increase the filter order
and repeat the steps.
7.3 Design of FIR Filter 273

Pole-zero plot(Type-I) Pole-zero plot(Type-II)

2 1

0 0
jw

jw
–2 –1
–1.0 –0.5 0.0 0.5 1.0 –1.0 –0.5 0.0 0.5 1.0
s s

Pole-zero plot(Type-III) Pole-zero plot(Type-IV)

1 1

0 0
jw

jw

–1 –1
–1.0 –0.5 0.0 0.5 1.0 –2 –1 0 1
s s

Fig. 7.10 Pole-zero plot of FIR filters

Table 7.3 Automatic location of zeros in different types of FIR filter

Type of Automatic zero
FIR filter location Inference
Type-I – Type-I FIR filter is a versatile filter; it can be used to design
lowpass, highpass, bandpass and band reject filters
Type-II Zero at ω = π Type-II FIR filter cannot be used as a highpass filter
Type-III Zero at ω = 0 Type-III FIR filter cannot be used as lowpass and highpass
and ω = π filters
Type-IV Zero at ω = 0 Type-IV FIR filter cannot be used as lowpass filter

7.3.2 Window-Based FIR Lowpass Filter

The expression for the impulse response of the ideal lowpass filter is given by

ωc ω
hd ½ n] = sin c c n ð7:1Þ
π π

From Eq. (7.1), it is possible to observe that the filter is neither causal nor finite in
duration. To make it finite duration, the desired impulse response is multiplied with
the window function. The mathematical expression for the impulse response multi-
plied with a rectangular window of length ‘N’ is given by
274 7 FIR Filter Design

#Impulse response of ideal filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N, omega = 50,1.0
n = np.arange(-N/2,N/2)
rect_win=np.ones(N)
hd = omega/np.pi * np.sinc(n*omega/np.pi)
h=hd*rect_win
plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')

Fig. 7.11 Python code to obtain the impulse response of the ideal filter

h[n]

0.30

0.25

0.20
Amplitude

0.15

0.10

0.05

0.00

–0.05

–20 –10 0 10 20
n-->

Fig. 7.12 Impulse response of an ideal filter

h½n] = hd ½n] × rectN ½n] ð7:2Þ

Experiment 7.6 Window-Based Design of Ideal Lowpass Filter

This experiment discusses the ideal lowpass FIR filter design using windowing
method. The python code, which obtains the impulse response of an ideal filter, is
shown in Fig. 7.11, and the corresponding output is shown in Fig. 7.12.
Inferences
The following inferences can be drawn from Fig. 7.12:
1. The impulse response of the ideal filter is non-causal.
2. If the impulse response is non-causal, the filter is not physically realizable.
7.3 Design of FIR Filter 275

#Comparison of ideal and practical filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N, omega = 32, 1.0
n = np.arange(0,N)
rect_win=np.ones(N)
hd = omega/np.pi * np.sinc(n*omega/np.pi)
hd1 = omega/np.pi * np.sinc((n-(N-1)/2)*omega/np.pi)
h1=hd1*rect_win
h=hd*rect_win
w,H=signal.freqz(h)
w1,H1=signal.freqz(h1)
plt.subplot(2,3,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude')
plt.title('h[n]:IF'),plt.subplot(2,3,2),plt.plot(w,20*np.log10(np.abs(H)))
plt.xlabel('$\omega$-->'),plt.ylabel('$|H(e^{j\omega})|$')
plt.title('Magnitude response:IF'),plt.subplot(2,3,3),plt.plot(w,np.unwrap(np.angle(H)))
plt.xlabel('$\omega$-->'),plt.ylabel('$\phi(e^{j\omega})$'),plt.title('Phase response:IF')
plt.subplot(2,3,4),plt.stem(n,h1),plt.xlabel('n-->'),plt.ylabel('Amplitude')
plt.title('h[n]:PF'),plt.subplot(2,3,5),plt.plot(w1,20*np.log10(np.abs(H1)))
plt.xlabel('$\omega$-->'),plt.ylabel('$|H(e^{j\omega})|$'),plt.title('Magnitude response:PF')
plt.subplot(2,3,6),plt.plot(w1,np.unwrap(np.angle(H1)))
plt.xlabel('$\omega$-->'),plt.ylabel('$\phi(e^{j\omega})$'),plt.title('Phase response:PF')
plt.tight_layout()

Fig. 7.13 Comparison of ideal and practical filter

3. Practically realizable filters have passband and stopband ripples and a non-zero
transition band.
4. For practical filter, a delay is necessary to capture most of the signal energy in
causal time.
5. Delay in time-domain is accomplished by multiplying the spectrum with a
complex exponential. The magnitude response is multiplied by e-jωτ, which
results in the time shift of the impulse response. This is discussed in the subse-
quent section.
Task
1. In the above experiment, increase the number of coefficients of the filter to
100, observe the filter’s impulse response and comment on the observed result.
Experiment 7.7 Comparison of Ideal and Practical Lowpass Filter
The impulse response of an ideal filter is non-causal; hence, it is not physically
realizable. To design a practical filter, the impulse response of the ideal filter has to
be delayed to make it causal. Delay in the time-domain is accomplished by multi-
plying the spectrum with a complex exponential. The magnitude response is multi-
plied by e-jωτ, which results in the time shift of the impulse response.
This python illustration compares the ideal FIR filter with the practical FIR filter.
The python code, which performs the comparison, is shown in Fig. 7.13, and the
corresponding output is shown in Fig. 7.14.
276 7 FIR Filter Design

h[n]:IF Magnitude response:IF Phase response:IF

0.0
–5
Amplitude

0.2

|H(e jw)|
–0.5

f (e jw)
–10
–1.0
0.0
–15
0 20 0 2 0 2
n--> w --> w -->

h[n]:PF Magnitude response:PF Phase response:PF

–0 0
Amplitude

0.2
|H(e jw)|

I(e jw)
–50 –10
0.0

0 20 0 2 0 2
n--> w --> w -->

Fig. 7.14 Characteristics of Ideal and practical filters

In Fig. 7.14, IF and PF denote ideal and practical filters, respectively.

Inferences
The following inferences can be made from this experiment:
1. By comparing the impulse response of the ideal and practical filters, it is possible
to observe that the practical filter exhibits a symmetric impulse response. In
contrast, the impulse response of the ideal filter is not symmetric.
2. From the magnitude responses, it is possible to observe that the ideal filter
exhibits ripples, whereas, in the practical filter, the ripples in the magnitude
response are less.
3. Practical filter exhibits linear phase characteristics, whereas the ideal filter phase
response is not linear.
4. From this experiment, it is possible to conclude that the impulse response should
be symmetric for the phase response to be linear.

7.3.3 Window-Based FIR Highpass Filter

The desired impulse response of window-based highpass FIR filter is given by

7.3 Design of FIR Filter 277

#FIR high pass filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Desired impulse response hd[n]
N,omega = 50,np.pi/4
n = np.arange(0,N)
rect_win=np.ones(N)
hd =np.sinc(n-(N-1)/2)-(omega/np.pi * np.sinc((n-(N-1)/2)*omega/np.pi))
h=hd*rect_win #Step 2: Windowed impulse response h[n]
w,H=signal.freqz(h) #Step 3: Frequency response of ideal filter
#Step 4:Impulse response of the filter
plt.subplot(2,2,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
z, p, k = signal.tf2zpk(h,1) #Step 5: Pole-zero plot of the filter
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot')
#Step 6: Magnitude response of the filter
plt.subplot(2,2,3),plt.plot(w,20*np.log10(np.abs(H)))
plt.xlabel('$\omega$-->'),plt.ylabel('$|H(e^{j^\omega})|$'),plt.title('Magnitude response')
#Step 7: Phase response of the filter
plt.subplot(2,2,4),plt.plot(w,np.unwrap(np.angle(H)))
plt.xlabel('$\omega$-->'),plt.ylabel('$\phi(e^{j^\omega})$'),plt.title('Phase response')
plt.tight_layout()

Fig. 7.15 Python code to obtain the characteristics of highpass filter

ωc ω
hd ½n] = sin cðn - τÞ - sin c c ðn - τÞ ð7:3Þ
π π

From Eq. (7.3), it is possible to know that the desired impulse response is a sinc
function that is not of finite duration. The desired impulse response must be
multiplied by the window function w[n] to make the finite impulse response. This
is expressed as

h½n] = hd ½n] × w½n] ð7:4Þ

Experiment 7.8 Window-Based FIR Highpass Filter

This experiment discusses about the study of characteristics of FIR highpass filter
design using the windowing method. The python code, which obtains the charac-
teristics of a highpass filter, is given in Fig. 7.15, and the corresponding output is
shown in Fig. 7.16.
Inferences
The following inferences can be made from this experiment:
278 7 FIR Filter Design

h[n] Pole-zero plot

0.25 1
Amplitude

0.00 0

jw
–0.25 –1

0 20 40 –1 0 1
n--> s

Magnitude response Phase response

0 0

–20
|H(ejw)|

–20

–40 f (ejw) –40

–60
0 1 2 3 0 1 2 3
w --> w -->

Fig. 7.16 Characteristics of highpass filter

1. Figure 7.16 shows that the code was written to simulate the desired impulse
response, which is given in Eq. (7.3). The desired impulse response is multiplied
by the window to get the actual response. The window chosen in this illustration
is a rectangular window.
2. The ‘scipy’ library is used here to obtain the frequency response and the pole-zero
plot of the filter. The built-in function ‘tf2zpk’ is utilized to obtain the pole-zero
plot, whereas the built-in function ‘freqz’ is used here to obtain the frequency
response of the filter.
3. From the magnitude response shown in Fig. 7.16, it is possible to observe that the
filter is a highpass filter that exhibits linear phase characteristics in the passband.
4. From the pole-zero plot, it is possible to observe that the zeros of FIR filter occur
in conjugate reciprocal pair.
Task
1. In the python code given in Fig. 7.15, try to use windows like Hamming,
Hanning, Bartlett and Blackman window, and observe the changes in magnitude
and phase response.

7.3.4 Window-Based FIR Bandpass Filter

The expression for the desired impulse response of the bandpass filter is given by
7.3 Design of FIR Filter 279

#Characteristics of bandpass filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Desired impulse response hd[n]
N,omega_1,omega_2 = 16,np.pi/4,np.pi/2
n = np.arange(0,N)
rect_win=np.ones(N)
hd =(omega_2/np.pi * np.sinc((n-(N-1)/2)*omega_2/np.pi))-(omega_1/np.pi * np.sinc((n-(N-
1)/2)*omega_1/np.pi))
h=hd*rect_win #Step 2: Windowed impulse response h[n]
w,H=signal.freqz(h) #Step 3: Frequency response of ideal filter
#Step 4:Impulse response of the filter
plt.subplot(2,2,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
z, p, k = signal.tf2zpk(h,1) #Step 5: Pole-zero plot of the filter
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot')
#Step 6: Magnitude response of the filter
plt.subplot(2,2,3),plt.plot(w,20*np.log10(np.abs(H)))
plt.xlabel('$\omega$-->'),plt.ylabel('$|H(e^{j^\omega})|$'),plt.title('Magnitude response')
#Step 7: Phase response of the filter
plt.subplot(2,2,4),plt.plot(w,np.unwrap(np.angle(H)))
plt.xlabel('$\omega$-->'),plt.ylabel('$\phi(e^{j^\omega})$')
plt.title('Phase response'),plt.tight_layout()

Fig. 7.17 Python code to obtain the characteristics of the bandpass filter

ωc2 ω ω ω
hd ½n] = sin cð c2 ðn - τÞÞ - c1 sin cð c1 ðn - τÞÞ ð7:5Þ
π π π π

Here ‘ωc1’ and ‘ωc2’ are pass band frequencies and ωc2 > ωc1. The desired
impulse response must be multiplied by the window function w[n] to get a finite
impulse response. This is expressed as

h½n] = hd ½n] × w½n] ð7:6Þ

Experiment 7.9 Window-Based FIR Bandpass Filter

This experiment discusses the FIR bandpass filter design using a windowing
approach. The python code, which obtains the characteristics of bandpass filter
with the cut-off frequencies ωc1 = π4 radians/sample and ωc2 = π2 radians/sample, is
shown in Fig. 7.17, and the corresponding output is shown in Fig. 7.18.
Inferences
From Fig. 7.18, the following observations can be made:
280 7 FIR Filter Design

h[n] Pole-zero plot

0.2 1
Amplitude

0.0

jw
0

–0.2 –1
0 5 10 15 –1 0 1 2 3
n--> s

Magnitude response Phase response

0
0
–50
|H(ejw)|

–100 I(ejw) –5

0 1 2 3 0 1 2 3
w --> w -->

Fig. 7.18 Characteristics of the bandpass filter

1. The lower cut-off frequency chosen is ωc1 = π4 which is equal to 0.785 radians/
sample, and the upper cut-off frequency chosen is ωc2 = π2 which is equal to
1.57 radians/sample. The magnitude response shows the passband between 0.785
and 1.57, and the gain drops beyond the cut-off frequency.
2. From the impulse response plot, it is possible to observe that the impulse response
is symmetric in nature.
3. From the phase response, it is possible to observe that the filter exhibits linear
phase characteristics in the passband. The linear phase is due to the symmetric
nature of the impulse response.
4. The pole-zero plot shows that the filter is an all-zero filter with the zeros occurring
in a conjugate reciprocal manner.
Task
1. In the above program, try to use windows like Hamming, Hanning, Bartlett and
Blackman, and observe the change in magnitude and phase response.

7.3.5 Window-Based FIR Band Reject Filter

The expression for the desired impulse response of the band reject/stop filter is given
by
7.3 Design of FIR Filter 281

#Characteristics of band-reject filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Desired impulse response hd[n]
N,omega_1,omega_2 = 21,np.pi/4,np.pi/2
n = np.arange(0,N)
rect_win=np.ones(N)
hd =(omega_1/np.pi * np.sinc((n-(N-1)/2)*omega_1/np.pi))+np.sinc(n-(N-1)/2)-(omega_2/np.pi *
np.sinc((n-(N-1)/2)*omega_2/np.pi))
h=hd*rect_win#Step 2: Windowed impulse response h[n]
w,H=signal.freqz(h) #Step 3: Frequency response of ideal filter
#Step 4:Impulse response of the filter
plt.subplot(2,2,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
z, p, k = signal.tf2zpk(h,1)#Step 5: Pole-zero plot of the filter
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot')
#Step 6: Magnitude response of the filter
plt.subplot(2,2,3),plt.plot(w,20*np.log10(np.abs(H))),plt.xlabel('$\omega$-->'),
plt.ylabel('$|H(e^{j^\omega})|$'),plt.title('Magnitude response')
#Step 7: Phase response of the filter
plt.subplot(2,2,4),plt.plot(w,np.unwrap(np.angle(H))),plt.xlabel('$\omega$-->'),
plt.ylabel('$\phi(e^{j^\omega})$'),plt.title('Phase response'),plt.tight_layout()

Fig. 7.19 Python code to obtain the characteristics of band reject filter

ωc1 ω ω ω
hd ½n] = sin cðn - τÞ þ sin c c1 ðn - τÞ - c2 sin c c2 ðn - τÞ ð7:7Þ
π π π π

Here ‘ωc1’ and ‘ωc2’ are stop band frequencies and ωc2 > ωc1. To make the
impulse response finite, the desired impulse response must be multiplied by the
window function w[n]. This is expressed as

h½n] = hd ½n] × w½n] ð7:8Þ

Experiment 7.10 Window-Based FIR Band Reject Filter

This experiment deals with the FIR band reject filter design using windowing
method. The python code, which obtains the characteristics of band reject filter, is
shown in Fig. 7.19, and the corresponding output is shown in Fig. 7.20. The cut-off
frequencies chosen are ωc1 = π4 radians/sample and ωc2 = π2 radians/sample.
Inferences
The following inferences can be drawn from this experiment:
1. The lower cut-off frequency chosen is ωc1 = π4, which is equal to 0.785 radians/
sample, and the upper cut-off frequency chosen is ωc2 = π2, which is equal to
282 7 FIR Filter Design

h[n] Pole-zero plot

0.75
1
0.50
Amplitude

jw
0
0.25

0.00 –1

0 5 10 15 20 –1 0 1
n--> s

Magnitude response Phase response

0 0

–25
|H(ejw )|

f (ejw )
–10

–50
–20
–75
0 1 2 3 0 1 2 3
w --> w -->

Fig. 7.20 Characteristics of band reject filter

1.57 radians/sample. The magnitude response shows that the frequency between
0.785 and 1.57 is attenuated.
2. From the impulse response plot, it is possible to observe that the impulse response
is symmetric in nature.
3. From the phase response, it is possible to observe that the filter exhibits linear
phase characteristics in the passband. The linear phase is due to symmetric nature
of the impulse response.
4. The pole-zero plot shows that the filter is an all-zero filter with the zeros occurring
in conjugate reciprocal manner.
Task
1. In the above program, try to use windows like Hamming, Hanning, Bartlett and
Blackman, and observe the change in magnitude and phase responses.

7.3.6 Design of FIR Filter Using Built-In Function

The built-in function ‘firwin’ available in ‘scipy’ library is used here to generate FIR
filter coefficients using window-based method. The input to the built-in function is
the number of coefficients of the filter, cut-off frequency and window type.
7.3 Design of FIR Filter 283

#Characteristics of Low pass filter using firwin command

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N,fs,LPF_cutoff=20,100,5
n=np.arange(N)
w_LPF=LPF_cutoff/(fs/2)
h=signal.firwin(N,w_LPF,window='hamming')
w,H=signal.freqz(h)
z, p, k = signal.tf2zpk(h,1)
plt.subplot(2,2,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot')
plt.subplot(2,2,3),plt.plot((fs * 0.5 / np.pi) * w,20*np.log10(np.abs(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$|H(jf)|$ in dB'),plt.title('Magnitude response')
#Step 7: Phase response of the filter
plt.subplot(2,2,4),plt.plot((fs * 0.5 / np.pi) * w,np.unwrap(np.angle(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$\phi(jf)$')
plt.title('Phase response'),plt.tight layout()

Fig. 7.21 Built-in function ‘firwin’ to obtain the characteristics of lowpass filter

Experiment 7.11 Design of FIR Lowpass Filter Using a Built-In Function

This experiment intends to obtain the FIR filter coefficients using the built-in
function ‘firwin’. After obtaining the coefficients, the characteristics of FIR filter,
like impulse response, magnitude response, phase response and pole-zero plot, are
plotted. The python code, which performs this task, is shown in Fig. 7.21, and the
corresponding output is shown in Fig. 7.22.
Inferences
From Fig. 7.21, the following observations can be obtained:
1. The built-in function ‘firwin’ available in ‘scipy’ library is used here to obtain the
filter coefficients.
2. The specifications of the lowpass filter are (a) number of coefficients = 20, cut-off
frequency = 5 Hz, sampling frequency = 100 Hz and window chosen is
‘Hamming’ window.
From Fig. 7.22, the following observations can be made:
1. The impulse response consists of 20 coefficients. From the impulse response, it is
possible to observe that the coefficients are symmetric.
2. From the pole-zero plot, it is possible to observe that the zeros occur in conjugate
reciprocal pair.
284 7 FIR Filter Design

h[n] Pole-zero plot

0.10 1
Amplitude

jw
0.05
–1
0.00
0 5 10 15 –1.0 –0.5 0.0 0.5 1.0
n--> s

Magnitude response Phase response

0 0
|H(jf)| in dB

–50

f (jf)
–5
–100

0 20 40 0 20 40
Frequency (f) (Hz) Frequency (f) (Hz)

Fig. 7.22 Characteristics of lowpass filter

3. From the magnitude response, it is possible to observe that the gain drops after the
cut-off frequency of 5 Hz.
4. From the phase response, it is possible to confirm that linear phase characteristic
is obtained in the passband.
Experiment 7.12 Design of FIR Highpass Filter Using the Built-In Function
This experiment aims to obtain the highpass filter coefficients with a slight change in
the code which generates the lowpass filter. The keyword ‘pass-zero = false’ helps
one to obtain the highpass filter. The python code, which performs this task, is
shown in Fig. 7.23, and the corresponding output is shown in Fig. 7.24.
Inferences
The following inferences can be made from this experiment:
1. From Fig. 7.23, it is possible to observe that the keyword ‘pass_zero = False’
allows one to obtain the coefficients of the highpass filter.
2. The characteristic of highpass filter is shown in Fig. 7.24. From this figure, it is
possible to observe that beyond the cut-off frequency of 5 Hz, the gain reaches a
value of 0 dB, and the phase response is linear curve.
Experiment 7.13 Design of FIR Bandpass Filter Using the Built-In Function
The objective of this experiment is to design a bandpass filter, which will pass signal
in the frequency range 10–20 Hz. The sampling frequency chosen is 100 Hz. The
order of the filter is 50, and the window chosen is ‘Hamming window’.
7.3 Design of FIR Filter 285

#Characteristics of high pass filter using "firwin" command

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N,fs,HPF_cutoff=21,100,5
n=np.arange(N)
w_HPF=HPF_cutoff/(fs/2)
h=signal.firwin(N,w_HPF,window='hamming',pass_zero=False)
w,H=signal.freqz(h)
z, p, k = signal.tf2zpk(h,1)
plt.subplot(2,2,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot')
plt.subplot(2,2,3),plt.plot((fs * 0.5 / np.pi) * w,20*np.log10(np.abs(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$|H(jf)|$ in dB'),plt.title('Magnitude response')
#Step 7: Phase response of the filter
plt.subplot(2,2,4),plt.plot((fs * 0.5 / np.pi) * w,np.unwrap(np.angle(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$\phi(jf)$')
plt.title('Phase response'),plt.tight_layout()

Fig. 7.23 Python code to obtain the characteristics of highpass filter

h[n] Pole-zero plot

2
Amplitude

0.5
0
jw

0.0 –2
0 5 10 15 20 –2 –1 0 1
n--> s

Magnitude response Phase response

0 0
|H(jf)| in dB

–5 –10
f (jf)

–10 –20

–15 –30
0 20 40 0 20 40
Frequency (f) (Hz) Frequency (f) (Hz)

Fig. 7.24 Characteristics of highpass filter

286 7 FIR Filter Design

#Characteristics of bandpass filter using "firwin" command

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N,fs,lf,uf=20,100,10,20 #Specifications
n=np.arange(0,N)
w_LCF=lf/(fs/2)
w_UCF=uf/(fs/2)
h=signal.firwin(N,[w_LCF,w_UCF],window='hamming',pass_zero=False)
w,H=signal.freqz(h)
z, p, k = signal.tf2zpk(h,1)
plt.subplot(2,2,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot')
plt.subplot(2,2,3),plt.plot((fs * 0.5 / np.pi) * w,20*np.log10(np.abs(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$|H(jf)|$ in dB'),plt.title('Magnitude response')
#Step 7: Phase response of the filter
plt.subplot(2,2,4),plt.plot((fs * 0.5 / np.pi) * w,np.unwrap(np.angle(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$\phi(jf)$')
plt.title('Phase response'),plt.tight_layout()

Fig. 7.25 Python code to obtain the characteristics of bandpass filter

The python code which generates the filter coefficient corresponding to the
desired bandpass filter is shown in Fig. 7.25, and the corresponding filter character-
istics are shown in Fig. 7.26.
Inferences
The following inferences can be drawn from this experiment:
1. From Fig. 7.26, it is possible to observe from the magnitude response that the
filter passes a band of frequencies from 10 to 20 Hz.
2. From the impulse response, it is possible to observe that the impulse response of
the filter is finite and symmetric.
3. The filter also exhibits linear phase characteristics in the passband. This is due to
the symmetric nature of the impulse response.
4. From the pole-zero plot, it is possible to observe that the zeros occur in conjugate
reciprocal pair.
Experiment 7.14 Design of FIR Band Reject Filter Using the Built-In Function
This experiment discusses the design of FIR band reject filter using built-in function.
The band reject filter is obtained from bandpass filter design using the key term
‘pass_zero = True’. The python code which obtains the coefficient of the band reject
filter is shown in Fig. 7.27, and the corresponding output is shown in Fig. 7.28.
7.3 Design of FIR Filter 287

h[n] Pole-zero plot

0.2 1
Amplitude

0.1
0

jw
0.0

–0.1 –1
0 5 10 15 0 2 4
n--> s

Magnitude response Phase response

0 5
|H(jf)| in dB

–0

f (jf)
–50
–5

–100 –10

0 20 40 0 20 40
Frequency (f) (Hz) Frequency (f) (Hz)

Fig. 7.26 Characteristics of bandpass filter

#Characteristics of band reject filter using "firwin" command

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N,fs,lf,uf=25,100,10,30
n=np.arange(0,N)
w_LCF=lf/(fs/2)
w_UCF=uf/(fs/2)
h=signal.firwin(N,[w_LCF,w_UCF],window='hamming',pass_zero=True)
w,H=signal.freqz(h)
z, p, k = signal.tf2zpk(h,1)
plt.subplot(2,2,1),plt.stem(n,h),plt.xlabel('n-->'),plt.ylabel('Amplitude'),plt.title('h[n]')
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,2,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(z.real, z.imag, 'ro', ms=7.5),plt.plot(p.real, p.imag, 'rx')
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot')
plt.subplot(2,2,3),plt.plot((fs * 0.5 / np.pi) * w,20*np.log10(np.abs(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$|H(jf)|$ in dB'),plt.title('Magnitude response')
#Step 7: Phase response of the filter
plt.subplot(2,2,4),plt.plot((fs * 0.5 / np.pi) * w,np.unwrap(np.angle(H)))
plt.xlabel('Frequency (f) (Hz)'),plt.ylabel('$\phi(jf)$')
plt.title('Phase response'),plt.tight_layout()

Fig. 7.27 Python code to generate the band reject filter coefficients and its characteristics
288 7 FIR Filter Design

h[n] Pole-zero plot

0.6
1
0.4
Amplitude

jw
0.2 0

0.0 –1

0 10 20 –1 0 1
n--> s

Magnitude response Phase response

0 0
|H(jf)| in dB

–25 –10

f (jf)
–50
–20
–75
–30
0 20 40 0 20 40
Frequency (f) (Hz) Frequency (f) (Hz)

Fig. 7.28 Characteristics of band reject filter

Inferences
The following observations can be made from this experiment:
1. Figure 7.27 shows that the keyword ‘pass-zero = True’ is used here to convert the
bandpass filter to a band reject filter.
2. From the magnitude response shown in Fig. 7.28, it is possible to confirm that this
filter blocks the frequency band from 10 to 30 Hz.

7.3.7 Window Functions

In this section, the window function is visualized in both time domain and frequency
domain. Different types of window functions include rectangular, triangular, Ham-
ming, Hanning, Kaiser, etc. The main lobe width of the window function controls the
transition bandwidth, whereas the height of the side lobe controls the passband and
stopband ripples. A rectangular window has the narrowest main lobe; hence, it gives
sharpest transition. Compared to rectangular window, the Hamming and Hanning
windows are smoother. By tapering the window smoothly to zero, the sidelobes can
be reduced in amplitude, which will reduce the ripple, but the trade-off is a larger
main lobe. A linear phase response can be achieved if the window function is
symmetric. Some windows allow controlled trade-offs between sidelobe amplitude
and main lobe width. One such window is the Kaiser window.
7.3 Design of FIR Filter 289

#Window functions in time and frequency domain

import numpy as np
import matplotlib.pyplot as plt
from scipy.fft import fft,fftshift
N = 51 #Length of the window
n =np.arange(-(N-1)/2, (N-1)/2)
#Defining different window functions
w_Rect = n-n+1; #Rectangular
w_Bart = 1 - 2* abs(n)/(N-1) #Bartlett
w_Han = 0.5 + 0.5 * np.cos(2*np.pi*n/(N-1)); #Hanning
w_Hamm = 0.54 + 0.46 * np.cos(2*np.pi*n/(N-1)); #Hamming
#Spectrum of window
W_Rect=fftshift(fft(w_Rect,1024)/len(w_Rect))
W_Bart =fftshift(fft(w_Bart,1024)/len(w_Bart))
W_Han=fftshift(fft(w_Han,1024)/len(w_Bart))
W_Hamm=fftshift(fft(w_Hamm,1024)/len(w_Bart))
plt.figure(1),plt.subplot(2,2,1),plt.stem(n,w_Rect),plt.xlabel('n-->'),plt.ylabel('w[n]'),
plt.title('Rectangular window'),plt.subplot(2,2,3),plt.stem(n,w_Bart),plt.xlabel('n-->')
plt.ylabel('w[n]'),plt.title('Bartlett window')
freq = np.linspace(-0.5, 0.5, len(W_Rect))
plt.subplot(2,2,2), plt.plot(freq, 20 * np.log10(W_Rect)),plt.axis([-0.5, 0.5, -120, 0]),
plt.xlabel('Normalized frequency'),plt.ylabel('Magnitude [dB]'),
plt.title('Spectrum of rectangular window'),
plt.subplot(2,2,4), plt.plot(freq, 20 * np.log10(W_Bart)),plt.xlabel('Normalized frequency'),
plt.ylabel('Magnitude [dB]'),plt.title('Spectrum of Bartlett window')
plt.axis([-0.5, 0.5, -120, 0]),plt.tight_layout()
plt.figure(2),plt.subplot(2,2,1),plt.stem(n,w_Han),plt.xlabel('n-->'),plt.ylabel('w[n]'),
plt.title('Hanning window'),plt.subplot(2,2,3),plt.stem(n,w_Hamm),plt.xlabel('n-->')
plt.ylabel('w[n]'),plt.title('Hamming window'),plt.subplot(2,2,2),
plt.plot(freq, 20 * np.log10(W_Han)),plt.xlabel('Normalized frequency'),
plt.ylabel('Magnitude [dB]'),plt.title('Spectrum of Hanning window'),plt.axis([-0.5, 0.5, -120, 0]),
plt.subplot(2,2,4), plt.plot(freq, 20 * np.log10(W_Hamm)),plt.xlabel('Normalized frequency'),
plt.ylabel('Magnitude [dB]'),plt.title('Spectrum of Hamming window')
plt.axis([-0.5, 0.5, -120, 0]),plt.tight_layout()

Fig. 7.29 Python code to plot the window functions in time and frequency domain

Experiment 7.15 Plotting Windows in the Time and Frequency Domain

The python code which plots the window function in the time and frequency domain
is shown in Fig. 7.29, and the corresponding output is shown in Figs. 7.30 and 7.31.
Different window functions considered in this example include (1) rectangular,
(2) triangular or Bartlett, (3) Hanning and (4) Hamming window.
Inference
From Figs. 7.30 and 7.31, the following inferences can be made:
1. Four different types of windows chosen are (a) rectangular, (b) triangular,
(c) Hamming and (d) Hanning.
2. Fourier transform of different types of window functions results in the sinc
functions.
290 7 FIR Filter Design

Rectangular window Spectrum of rectangular window

1.0 0

Magnitude [dB]
–50
w[n]

0.5

–100
0.0
–20 0 20 –0.4 –0.2 0.0 0.2 0.4
n--> Normalized frequency
Bartlett window Spectrum of Bartlett window
1.0 0

Magnitude [dB]
–50
w[n]

0.5

–100
0.0
–20 0 20 –0.4 –0.2 0.0 0.2 0.4
n--> Normalized frequency

Fig. 7.30 Window functions in time and frequency domain

Hanning window Spectrum of Hanning window

1.0 0
Magnitude [dB]

–50
w[n]

0.5

–100
0.0
–20 0 20 –0.4 –0.2 0.0 0.2 0.4
n--> Normalized frequency

Hamming window Spectrum of Hamming window

1.0 0
Magnitude [dB]

–50
w[n]

0.5

–100
0.0
–20 0 20 –0.4 –0.2 0.0 0.2 0.4
n--> Normalized frequency

Fig. 7.31 Window functions in time and frequency domain

3. The spectrum of different windows differs with respect to main lobe width and
side lobe height.
7.4 Frequency Sampling-Based FIR Filter Design 291

7.4 Frequency Sampling-Based FIR Filter Design

The steps followed in frequency sampling method of FIR filter design are summa-
rized below:
1. The design step starts with a prescribed magnitude response.
2. The prescribed magnitude response is sampled at enough points.
3. Take the inverse Fourier transform of the samples obtained in step (2). This will
result in the filter’s impulse response.

Experiment 7.16 Frequency Sampling-Based FIR Filter Design

The built-in function ‘firwin2’ available in ‘scipy’ library is used here to generate
FIR filter coefficients. The following python code helps one to obtain the coefficients
of Type-I, Type-II, Type-III and Type-IV FIR filters. It is known that Type-I and
Type-II FIR filter exhibits even symmetry with odd and even number of coefficients,
respectively. Type-III, and Type-IV FIR filter exhibits odd symmetry with odd and
even number of coefficients, respectively. The python code, which obtains the
response of four types of FIR filter, is shown in Fig. 7.32, and the corresponding
output is shown in Fig. 7.33.
Inferences
From Fig. 7.32, the following observations can be made:
1. The built-in function ‘firwin2’ is used here to generate the FIR filter coefficients.
In the program, the variables ‘f1 to f4’ represent the desired normalized frequency
in the range 0 to 1. The variables ‘m1 to m4’ represent the desired magnitude
response. ‘1’ represents passband, and ‘0’ represents the stopband. In the pro-
gram, ‘N1 to N4’ represents the number of coefficients of the filter. For Type-I
and Type-III, the number of coefficients has to be odd. For Type-II and Type-IV,
the number of coefficients has to be even.
2. The keyword ‘asymmetric = false’ implies even symmetry, and ‘asymmet-
ric = true’ represents odd symmetry. Type-I and Type-II FIR filters exhibit
even symmetry, whereas Type-III and Type-IV FIR filters exhibit odd symmetry.
From Fig. 7.33, the following interpretations can be made:
1. Type-I and Type-II FIR filters act as lowpass filter; Type-III FIR filter act as
bandpass filter. Type-IV FIR filter act as a highpass filter.
2. All four types of FIR filters exhibit linear phase characteristics in the passband.
3. It is to be noted that Type-II FIR filter cannot be used as a highpass filter. Type-III
FIR filter cannot be used as a lowpass and highpass filters. Type-IV FIR filter
cannot be used as a lowpass filter.
292 7 FIR Filter Design

#Characteristics of FIR filter using frequency sampling method

import numpy as np
from scipy import signal
import matplotlib.pyplot as plt
#Type-I FIR filter
f1 = [0,0.5,0.5,1]; #desired frequencies(w/pi)
m1 = [1,1,0,0]; #desired magnitude at f
N1 = 51 #samples
h1 = signal.firwin2(N1, f1, m1,antisymmetric=False)
w1, H1 = signal.freqz(h1)
plt.figure(1),plt.subplot(2,2,1),plt.plot(w1,(10 * np.log10(abs(H1)))),
plt.ylabel('$|H(e^{j^\omega})|$ in dB'),plt.xlabel('$\omega/\pi$'),
plt.title('Type-I FIR filter'),plt.subplot(2,2,2),plt.plot(w1,np.unwrap(np.angle(H1)))
plt.ylabel('$\phi(e^{j^\omega})$'),plt.xlabel('$\omega/\pi$'),plt.title('Type-I FIR filter')
#Type-II FIR filter
f2 = [0,0.6,0.6,1]; #desired frequencies(w/pi)
m2 = [1,1,0,0]; #desired magnitude at f
N2 = 50 #samples
h2 = signal.firwin2(N2, f2, m2,antisymmetric=False)
w2, H2 = signal.freqz(h2)
plt.subplot(2,2,3),plt.plot(w2,(10 * np.log10(abs(H2)))),
plt.ylabel('$|H(e^{j^\omega})|$ in dB'),plt.xlabel('$\omega/\pi$'),
plt.title('Type-II FIR filter'),plt.subplot(2,2,4),plt.plot(w2,np.unwrap(np.angle(H2)))
plt.ylabel('$\phi(e^{j^\omega})$'),plt.xlabel('$\omega/\pi$'),plt.title('Type-II FIR
filter'),plt.tight_layout()
#Type-III FIR filter
f3 = [0,0.2,0.4,0.6,0.8,1]; #desired frequencies(w/pi)
m3 = [0,0,1,0,0,0]; #desired magnitude at f
N3= 101 #samples
h3 = signal.firwin2(N3, f3, m3,antisymmetric=True)
w3, H3 = signal.freqz(h3);
plt.figure(2),plt.subplot(2,2,1),plt.plot(w3,(10 * np.log10(abs(H3)))),
plt.ylabel('$|H(e^{j^\omega})|$ in dB'),plt.xlabel('$\omega/\pi$'),
plt.title('Type-III FIR filter'),plt.subplot(2,2,2),plt.plot(w3,np.unwrap(np.angle(H3)))
plt.ylabel('$\phi(e^{j^\omega})$'),plt.xlabel('$\omega/\pi$'),plt.title('Type-III FIR filter')
#Type-IV FIR filter
f4 = [0,0.6,0.6,1]; #desired frequencies(w/pi)
m4 = [0,0,1,1]; #desired magnitude at f
N4 = 150 #samples
h4 = signal.firwin2(N4, f4, m4,antisymmetric=True)
w4, H4 = signal.freqz(h4);
plt.subplot(2,2,3),plt.plot(w4,(10 * np.log10(abs(H4)))),plt.ylabel('$|H(e^{j^\omega})|$ in dB'),
plt.xlabel('$\omega/\pi$'),plt.title('Type-IV FIR filter'),plt.subplot(2,2,4),
plt.plot(w4,np.unwrap(np.angle(H4))),plt.ylabel('$\phi(e^{j^\omega})$'),
plt.xlabel('$\omega/\pi$'),plt.title('Type-IV FIR filter'), plt.tight_layout()

Fig. 7.32 Python code to obtain the characteristics of four types of FIR filter using frequency
sampling method
7.5 Design of Optimal FIR filter 293

Type-I FIR filter Type-I FIR filter Type-III FIR filter Type-III FIR filter
0 0 0
|H(ejw )| in dB

|H(ejw )| in dB
0

I(ejw )

I(ejw )
–20 –20
–100 –50
–40 –40
0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
w /p w /p w /p w /p
Type-II FIR filter Type-II FIR filter Type-IV FIR filter Type-IV FIR filter
0 0 0 0
|H(ejw )| in dB

|H(e )| in dB
I(ejw )

I(ejw )
–25 –25 –25 –50

jw
–50 –50 –50
–100
0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
w /p w /p w /p w /p

Fig. 7.33 Magnitude and phase responses of four types of FIR filter

7.5 Design of Optimal FIR filter

The optimal equiripple FIR filter design is due to Parks and McClellan. The optimal
method provides an FIR filter coefficient representing the best approximation to the
desired frequency response in a Chebyshev sense. The term optimal can be defined
in various ways. The Parks-McClellan package uses the Remez exchange algorithm
to optimize the filter design by selecting the impulse response, which minimizes the
peak ripple in the passband and stopband. The filter designed by this approach is
termed as ‘equiripple’ filter. It is also termed as ‘minimax filter’ because the
maximum ripple deviation is minimized in the optimization procedure.
Experiment 7.17 Design of Optimal FIR Filter
The built-in functions available in ‘scipy’ library like, ‘remez’ and ‘firls’ can be used
to obtain the coefficients of optimal filter using Remez exchange algorithm and the
least square approach, respectively.
The aim of this experiment is to design the lowpass, highpass, bandpass and band
reject filters using the built-in function ‘remez’ available in ‘scipy’ library. The
python code, which generates the filter coefficients and plots the magnitude
responses of these four filters, is shown in Fig. 7.34, and the corresponding output
is shown in Fig. 7.35.
Inferences
From Fig. 7.34, the following observations can be made:
1. The sampling frequency of the four filters is kept at 1000 Hz, the number of taps
of the filter of all four filters is kept as 125 and the transition width of the four
filters is kept as 25 Hz.
2. The lowpass filter cut-off frequency is kept at 100 Hz. This means that the filter
should pass all frequencies till 100 Hz and block frequency components greater
than 100 Hz.
3. The cut-off frequency of a highpass filter is fixed as 200 Hz.
4. For the bandpass and band reject filters, the lower and upper cut-off frequencies
are fixed as 100 Hz and 200 Hz, respectively.
294 7 FIR Filter Design

#Filter design using built-in function remez

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
fs,N,trans_width=1000,125,25 #Low pass filter design
fc_lp=100 #LPF cut off frequency
h_lp= signal.remez(N, [0, fc_lp, fc_lp + trans_width, 0.5*fs], [1, 0], Hz=fs)
w1, H_lp = signal.freqz(h_lp,1)
plt.subplot(2,2,1),plt.plot(0.5*fs*w1/np.pi, 20*np.log10(np.abs(H_lp)))
plt.xlabel('Frequency (Hz)'),plt.ylabel('Gain (dB)'),plt.title('Magnitude response of LPF')
#High pass filter design
fc_hp = 200.0 # High pass filter cut off frequency
h_hp = signal.remez(N, [0, fc_hp - trans_width, fc_hp, 0.5*fs],[0, 1], Hz=fs)
w2, H_hp = signal.freqz(h_hp, [1])
plt.subplot(2,2,2),plt.plot(0.5*fs*w2/np.pi, 20*np.log10(np.abs(H_hp)))
plt.xlabel('Frequency (Hz)'),plt.ylabel('Gain (dB)'),plt.title('Magnitude response of HPF')
#Band pass filter
band = [100, 200] # Desired pass band, Hz
edges = [0, band[0] - trans_width, band[0], band[1], band[1] + trans_width, 0.5*fs]
h_bpf = signal.remez(N, edges, [0, 1, 0], Hz=fs)
w3, H_bpf = signal.freqz(h_bpf,1)
plt.subplot(2,2,3),plt.plot(0.5*fs*w3/np.pi, 20*np.log10(np.abs(H_bpf)))
plt.xlabel('Frequency (Hz)'),plt.ylabel('Gain (dB)'),plt.title('Magnitude response of BPF')
#Band reject filter
h_brf = signal.remez(N, edges, [1, 0, 1], Hz=fs)
w4, H_brf = signal.freqz(h_brf,1)
plt.subplot(2,2,4),plt.plot(0.5*fs*w4/np.pi, 20*np.log10(np.abs(H_brf)))
plt.xlabel('Frequency (Hz)'),plt.ylabel('Gain (dB)')
plt.title('Magnitude response of BRF'),plt.tight_layout()

Fig. 7.34 Python code to obtain the filter coefficients using the built-in function ‘remez’

From Fig. 7.35, it is possible to observe that the magnitude responses of the four
filters are as per the specification.

7.6 Applications of FIR Filter

The coefficients of FIR filters are either symmetric or anti-symmetric. FIR filter
exhibits linear phase characteristics, because of which, there is no phase distortion.
The group delay of FIR filter is constant. Since the poles of FIR filter occur at the
origin, FIR filters are inherently stable. Because of these characteristics, FIR filters
are used in many areas of signal processing like multirate signal processing, adaptive
signal processing, etc. In multirate signal processing, FIR filters are preferred to
design perfect reconstruction filter bank. In adaptive signal processing, FIR filters
are preferred in system identification, adaptive notch filter, inverse system model-
ling, echo cancellation and variety of such applications. In this section, two simple
7.6 Applications of FIR Filter 295

Magnitude response of LPF Magnitude response of HPF

0 0

Gain (dB)
Gain (dB)

–50 –50

–100
–100
0 200 400 0 200 400
Frequency (Hz) Frequency (Hz)
Magnitude response of BPF Magnitude response of BRF
0 0

Gain (dB)
Gain (dB)

–50 –50

–100
–100
0 200 400 0 200 400
Frequency (Hz) Frequency (Hz)

Fig. 7.35 Magnitude responses of the filters

Window-based FIR low pass

Input signal Output signal
filter with cut-off frequency
5Hz and 15Hz sine wave 5Hz sine wave
8Hz

Fig. 7.36 Block diagram of problem statement

applications are discussed. One is signal separation, and the other is signal
denoising.
Experiment 7.18 Separation of Signals Using FIR Filter
The signal x(t) is an addition of two signals x1(t) and x2(t). The frequencies of the two
signals x1(t) and x2(t) are 5 Hz and 15 Hz, respectively. The signal x(t) is now passed
through a lowpass filter whose cut-off frequency is 8 Hz, the order of the filter is
20 and the window chosen is Hanning.
The problem statement is depicted in the form of block diagram, which is shown
in Fig. 7.36.
The python code, which performs lowpass filtering of the input sine wave with
5 and 15 Hz frequency components, is shown in Fig. 7.37, and the corresponding
output is in Fig. 7.38.
Inferences
From Fig. 7.37, the following observations can be made:
296 7 FIR Filter Design

#Low pass filtering of sine wave

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
from scipy.fftpack import fft,fftfreq
#Step 1: Defining the parameters of sine wave
f1,f2,ph=5,15,0 #Frequency of signal 1 and 2 and phase
#Step 2: Defining the sampling frequency and number of points in FFT
fs, N=100, 256 #Sampling frequency
T=1/fs #Sampling period
#Step 3: Generation of sine wave
t=np.linspace(0,N*T,N)
x1=np.sin(2*np.pi*f1*t+ph)
x2=np.sin(2*np.pi*f2*t+ph)
x=x1+x2
#Step 4: Design of LPF
f_cut=8 # Cut-off frequency
Nyquist_freq, numtaps=fs/2, 21
f_cutoff=f_cut/Nyquist_freq
h=signal.firwin(numtaps, f_cutoff,window='hann')
y=signal.lfilter(h,1,x) #Step 5: Filtering of the signal
f_axis=fftfreq(N,T)[0:N//2] #Step 5: Obtaining the spectrum of input and output signal
X=fft(x)
Y=fft(y)
#Step 6: Plotting the results
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Input Signal'),
plt.subplot(2,2,2),plt.plot(t,y),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Filtered Signal'),plt.subplot(2,2,3),plt.plot(f_axis,2/N*np.abs(X[0:N//2]))
plt.xlabel('Frequency'),plt.ylabel('Magnitude'),plt.title('Spectrum of input Signal')
plt.subplot(2,2,4),plt.plot(f_axis,2/N*np.abs(Y[0:N//2]))
plt.xlabel('Frequency'),plt.ylabel('Magnitude'),plt.title('Spectrum of filtered Signal')
plt.tight_layout()

Fig. 7.37 Python code to perform lowpass filtering of the input signal

1. The input signal is the sum of two sine wave frequencies, 5 and 15 Hz.
2. Window-based FIR filter is designed with a cut-off frequency of 8 Hz, the number
of taps is 21 and the window chosen is Hanning. The sampling frequency chosen
is 100 Hz.
3. ‘The built-in function ‘firwin’, which is available in ‘scipy’ package, is used to
design the filter.
From Fig. 7.38, the following observations can be made:
1. The input signal is a mixture of 5 and 15 Hz sine wave.
2. The output signal is a lowpass filtered signal which retains a 5 Hz sine wave.
3. The spectrum of the input signal shows peaks corresponding to 5 and 15 Hz
frequency components.
7.6 Applications of FIR Filter 297

Input Signal Filtered Signal

1 0.5
Amplitude

Amplitude
0 0.0

–1 –0.5

0 1 2 0 1 2
Time Time

Spectrum of input Signal Spectrum of filtered Signal

1.0
0.6
Magnitude

Magnitude
0.5 0.4

0.2
0.0 0.0
0 20 40 0 20 40
Frequency Frequency

Fig. 7.38 Lowpass filtering using window-based FIR filter

4. The spectrum of the output signal shows peak at 5 Hz, which implies that the filter
allows 5 Hz frequency component of the input signal, and it blocks the 15 Hz
frequency component of the input signal.
Experiment 7.19 Denoising of the Signal Using FIR Filter
The signal x(t) is a 5 Hz sine wave. This signal x(t) is corrupted by white noise,
which follows uniform distribution in the range [0, 1]. The noisy signal is then
passed through FIR lowpass filter. The FIR filter coefficients are generated using the
windowing technique. Plot results of the clean, noisy and filtered signals.
The python code which performs the above-mentioned task is shown in Fig. 7.39,
and the corresponding output is shown in Fig. 7.40.
Inferences
The following observations can be made from this experiment:
1. From Fig. 7.39, it is possible to observe that the built-in function ‘random.
uniform’ available in numpy library is used here to generate uniformly distributed
random noise in the interval 0 to 1.
2. The random noise is added to pure sine wave to create noisy sine wave. The noisy
sine wave is then filtered using FIR filter, obtained using the built-in function
‘firwin’ available in scipy library.
3. From Fig. 7.40, it is possible to observe that the clean sine wave has a frequency
of 5 Hz. It is then corrupted by random noise to create noisy sine wave. From the
filtered signal, it is possible to observe that the impact of noise is minimized.
298 7 FIR Filter Design

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of clean signal
t=np.linspace(0,1,100)
x=np.sin(2*np.pi*5*t)
n=np.random.uniform(0,1,100) #Step 2: Uniform random noise
x1=x+n #Step 3: Noisy sine wave
h=signal.firwin(21,.2,pass_zero=True)#FIR filter coefficients
y=signal.filtfilt(h,1,x1) #Step 4: Filtered signal
#Step 5: Plotting the result
plt.subplot(3,1,1),plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Clean signal'),plt.subplot(3,1,2),plt.plot(t,x1)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Noisy signal')
plt.subplot(3,1,3),plt.plot(t,y),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Filtered signal'),plt.tight_layout()

Fig. 7.39 Python code which performs filtering of noisy signal

Clean signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
Time
Noisy signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
Time
Filtered signal
Amplitude

1
0
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 7.40 Filtering of signal corrupted by uniform random noise

Exercises
1. Design a linear phase lowpass filter that satisfies the following magnitude
π
1, jωj <
response H ðe Þ =
jω 4 . For N = 5 and N = 7. Assume the window
0, otherwise
to be rectangular window. Plot the magnitude response of the filter.
7.6 Applications of FIR Filter 299

Table 7.4 Comparison of different window functions

ðN - 1Þ ðN - 1Þ
S.I No. Name of the window Time domain expression - 2 ≤n≤ 2
1 Rectangular window w[n] = 1
2jnj
2 Bartlett window or Triangular w½n] = 1 - N -1
window
3 Hamming window w½n] = 0:54 þ 0:46 cos 2πn
N -1
4 Hanning window w½n] = 0:5 þ 0:5 cos N - 1
2πn

5 Blackman window -1 þ
w½n] = 0:42 þ 0:5 cos N2πn 0:08 cos 4πn
N -1

2. Design a length 7 linear phase highpass filter using windowing method with the
cut-off frequency ωc = π3 radians/sample. Assume the window to be Bartlett
window. Plot the magnitude and phase response of the filter.
3. Design a length 5 linear phase bandpass filter with lower cut-off frequency
ωc1 = 0.25π radians/sample and upper cut-off frequency ωc2 = 0.75π radians/
sample. Assume the window to be rectangular window. Plot the magnitude and
phase response of the filter.
4. Design a length 7 linear phase band reject filter with lower cut-off frequency
ωc1 = 0.15π radians/sample and upper cut-off frequency ωc2 = 0.45π radians/
sample. Assume the window to be Hamming window. Plot the magnitude and
phase response of the filter.
5. The signal x(t) is an addition of two signals x1(t) and x2(t). The frequencies of the
two signals x1(t) and x2(t) are 5 Hz and 10 Hz, respectively. The signal x(t) is
now passed through a highpass filter whose cut-off frequency is 8 Hz, the order
of the filter is 10 and the window chosen is Hamming window. Plot the input
signal and the filtered signal and comment on the observed result.
6. The time domain expression for different window functions is given in the
following Table 7.4.
Write a python code to plot the above window functions and comment on the
observed result. Assume the value of N as 31.
7. Write a python code to design a length 9 linear phase highpass filter using
windowing method with the cut-off frequency ωc = 4 radians/sample. Assume
the window to be Bartlett window. Plot the pole-zero plot of the filter, and
observe that the zeros of the filter occur in conjugate pair.
8. The impulse response of 5-tap linear phase lowpass filter is given by
h1[n] = {0.159, 0.225, 0.25, 0.225, 0.159}. Derive another filter from this
lowpass filter, whose impulse response is given by h2[n]. The relationship
between the impulse responses is given by h2[n] = (-1)nh1[n]. Plot the magni-
tude responses of the two filters and comment on the observed result.
9. Design a lowpass FIR filter using a frequency sampling technique having cut-off
frequency of π/2 radians/sample. The length of the filter is 21. Plot the magni-
tude response of the filter.
10. Design a digital FIR lowpass filter with the following specifications:
(a) Passband cut-off frequency: fp = 1 kHz. (b) Stopband cut-off frequency:
300 7 FIR Filter Design

fs = 4 kHz. (c) Passband ripple: Rp = 0.25 dB. (d) Stopband attenuation:

Rs = 0.25 dB. (e) Sampling frequency: fs = 20 kHz. Use subplot to plot the
magnitude response, phase response, impulse response and pole-zero plot of the
filter.

Objective Questions
1. Assertion: FIR filter exhibits linear phase characteristics.
Reason: The coefficients of FIR filter are either symmetric or anti-symmetric:
A. Both assertion and reason are true.
B. Assertion is true; reason is false.
C. Assertion is false; reason may be true.
D. Both assertion and reason are false.
2. If ‘N’ represents the number of coefficients of the FIR filter, then the group delay
of the filter is expressed as
A. τg = N2
B. τg = N2 - 1
C. τg = N -2
1

D. τg = N
3. Identify the statement that is FALSE with respect to FIR filter
A. FIR filter is all-zero filter.
B. FIR filter is all-pole filter.
C. FIR filter is inherently stable filter.
D. Group delay of FIR filter is constant.
4. The filter which exhibits even symmetry with odd number of coefficient is
A. Type-I FIR filter
B. Type-II FIR filter
C. Type-III FIR filter
D. Type-IV FIR filter
5. The built-in function available in scipy library to design window-based FIR filter
is
A. signal.firwin()
B. signal.firwin2()
C. signal.remez()
D. signal.firls()
6. The built-in function available in scipy library to design frequency sampling-
based FIR filter is
A. signal.firwin(),
B. signal.firwin2()
C. signal.remez()
D. signal.firls()
7.6 Applications of FIR Filter 301

7. The following python command h = signal.firwin(5,0.5) generates

A. Five coefficients of lowpass filter
B. Four coefficients of lowpass filter
C. Five coefficients of highpass filter
D. Four coefficients of highpass filter
8. The type of FIR filter that has zero at ω = 0 and at ω = π is
A. Type-I FIR filter
B. Type-II FIR filter
C. Type-III FIR filter
D. Type-IV FIR filter
9. Type-II FIR filter cannot be used as
A. Lowpass filter
B. Highpass filter
C. Band pass filter
D. Band reject filter
10. The frequency response of a linear phase filter is given by H(ejω) = e-j4ωR(ω),
where R(ω) represents the magnitude response. The group delay of the filter is
A. 1
B. 2
C. 3
D. 4
11. If a zero occurs at z0 of a real-valued linear phase filter than the other zeros are at
A. z10 only
B. z*0 only
C. z10 and at z1* only
0
1 *
D. z 0 , z0 and at z1* .
0

12. Let h[n] represent the impulse response of lowpass filter and then the impulse
response (-1)nh[n] represent
A. Lowpass filter
B. Highpass filter
C. Band pass filter
D. Band reject filter
13. Match the following

Transfer function of the filter Type of FIR filter

(P) H1(z) = 1 + 2z-1 + z-2 (i) Type-I FIR filter
(Q) H2(z) = 1 - z-1 (ii) Type-II FIR filter
(continued)
302 7 FIR Filter Design

Transfer function of the filter Type of FIR filter

(R) H3(z) = 1 + z-1 (iii) Type-III FIR filter
(S) H4(z) = 1 - z-2 (iv) Type-IV FIR filter

A. P-(i), Q-(ii), R-(iii), S-(iv)

B. P-(i), Q-(iv), R-(ii), S-(iii)
C. P-(iv), Q-(iii), R-(ii), S-(i)
D. P-(iii), Q-(iv), R-(ii), S-(i)
14. Upon executing the following python code, what will be the impulse response
(h) of the filter?

A. h = {1,2,0,2,1}
B. h = {1,2,0,-2,-1}
C. h = {1,2,0,1,2}
D. h = {1,2,0,-1,-2}
15. The impulse response of a filter is obtained using the following python code.
The filter is

A. Type-I FIR filter

B. Type-II FIR filter
C. Type-III FIR filter
D. Type-IV FIR filter

Bibliography

1. Vijay Madisetti, “The Digital Signal Processing Handbook”, CRC Press, 1997.
2. Paulo S. R. Diniz, Eduardo A.B. da Silva and Sergio L. Netto, “Digital Signal Processing:
System Analysis and Design”, Cambridge University Press, 2010.
3. T.W. Parks, and C.S. Burrus, “Digital Filter Design”, John Wiley and Sons, 1987.
4. Wai-Kai Chen, “Passive, Active and Digital Filters”, CRC Press, 2006.
5. Robert J. Schilling and Sandra L. Harris, “Fundamentals of Digital Signal Processing using
MATLAB”, Cengage Learning, 2012.
Chapter 8
Infinite Impulse Response Filter

Learning Objectives
After reading this chapter, the reader is expected to:
• Design and analyse Butterworth filter.
• Design and analyse Chebyshev and inverse Chebyshev filters.
• Design and analyse elliptic filter.
• Implement different mapping techniques to convert analogue filter into an equiv-
alent digital filter.

Roadmap of the Chapter

This chapter starts with the types of infinite impulse response (IIR) filter and
discusses the different mapping methods for converting analogue filters into digital
filters. Finally, the design of IIR filters is discussed in this chapter. Roadmap of this
chapter is illustrated below.

IIR Filter

Types IIR Filter Mapping Techniques Design of IIR Filter

Butterworth Filter Backward Difference Analog filter

Chebyshev Filter Impulse Invariant Digital filter

Inverse Chebyshev Filter Bilinear Transformation

Elliptic Filter Matched Z transform

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 303
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_8
304 8 Infinite Impulse Response Filter

PreLab Questions
1. What is a recursive filter?
2. Examine whether IIR filter is recursive or not. Justify your answer.
3. What is a ripple in the filter’s frequency response?
4. List the different types of IIR filters based on the ripples.
5. Why the Butterworth filter is termed as ‘maximally flat frequency response’
filter?
6. Mention the techniques to convert the analogue filter transfer function into an
equivalent digital filter transfer function.
7. Which criterion is important in mapping an analogue filter into an equivalent
digital filter?
8. List the steps involved in obtaining the digital filter transfer function from the
analogue filter transfer function using the impulse invariant technique.
9. What are the drawbacks of impulse invariant technique?
10. How does the bilinear transformation technique avoid aliasing while performing
the mapping process?
11. What is frequency warping with respect to bilinear transformation technique?
Suggest a solution to overcome the frequency warping problem in bilinear
transformation technique.
12. Tabulate the difference between Butterworth filter, Chebyshev filter, inverse
Chebyshev and elliptic filter with respect to (a) ripples in passband and
stopband, (b) transition width and (c) order of the filter required to meet the
filter specification.
13. Elaborate on the steps involved in the design of a digital IIR filter.

8.1 IIR Filter

In IIR filter, the current output is a function of the current and previous inputs and
past outputs. The relationship between the input and output of an IIR filter is given
by

y½n] = b0 x½n] þ b1 x½n - 1] þ ⋯ þ bM x½n - M] - a1 y½n - 1] þ a2 y½n - 2]

þ ⋯ þ aN y½n - N] ð8:1Þ

From Eq. (8.1), it is possible to observe that the current output is a function of
current and previous inputs and past outputs. Thus, IIR filters are ‘recursive filters’.
In a recursive filter, the current output depends on both the input and previously
calculated outputs. The word ‘recursive’ literally means ‘running back’ and refers to
previously calculated output values that go back into calculating the current output
along with input values.
Upon taking Z-transform of the input-output relation given in Eq. (8.1), the
transfer function expression for IIR filter is given by
8.1 IIR Filter 305

M -k
k = 0 bk z
H ðzÞ = N ð8:2Þ
1þ -k
k = 1 ak z

The impulse response of the IIR filter can be obtained upon taking inverse Z-
transform of the transfer function. As the name suggests, the impulse response is not
of finite duration. The impulse response is not guaranteed to be either symmetric or
anti-symmetric; hence, it is not possible to obtain linear phase characteristics in the
IIR filter. The group delay is not constant in IIR filter. If the pole of the IIR filter lies
outside the unit circle, the filter is unstable. This implies that stability is not
guaranteed in IIR filter. The main advantage of IIR filter is that it is possible to
meet the filter specification with the minimum number of coefficients.
Experiment 8.1 Computation of Impulse Response h[n] of the Recursive Filter
The filter’s impulse response h[n] can be obtained from the input and output relation.
Let us consider the linear constant coefficient difference equation

1
y½n] þ y½n - 1] = x½n] ð8:3Þ
2

where y[n] denotes the output and x[n] represents the input of the equation. From
Eq. (8.3), it is possible to observe that the current output (y[n]) is a function of
current input (x[n]) and past output (y[n - 1]). The impulse response of this filter can
be computed by replacing y[n] as h[n] and x[n] as δ[n]. Hence, Eq. (8.3) can be
rewritten as

1
h½n] þ h½n - 1] = δ½n]
2

The above equation can be rewritten as

1
h½n] = δ½n] - h½n - 1]
2

Assume that the system is initially at rest (i.e. y[n] = 0 for n < 0) and substituting
n = 0, 1, 2, 3, . . ., in the above equation, we get

1
h½0] = δ½0] = h ½ 0 - 1] = 1 - 0 = 1
2
1 1 1 1
h½1] = δ½1] - h½1 - 1] = 0 - h½0] = - × 1 = -
2 2 2 2
1 1 1 1 1
h½2] = δ½2] - h½2 - 1] = 0 - h½1] = - × - =
2 2 2 2 4
306 8 Infinite Impulse Response Filter

Impulse response
1.0

0.8

0.6

0.4

h[n]
0.2

0.0

–0.2

–0.4

–4 –2 0 2 4 6 8 10
n-->
(a) Python code (b) Simulation result

Fig. 8.1 Python code and its result of impulse response computation. (a) Python code. (b)
Simulation result

1 1 1 1 1
h½3] = δ½3] - h½3 - 1] = 0 - h½2] = - × = -
2 2 2 4 8
1 n
h½ n] = - u½n]
2

The python code to obtain the impulse response from the input and output
coefficients mentioned in the difference equation is shown in Fig. 8.1a, and the
corresponding output is shown in Fig. 8.1b.
Inference
1. From Fig. 8.1a, it is possible to observe that the ‘scipy’ library is used for the
filtering, and ‘signal.lfilter’ is used for the computation of impulse response from
the input and output coefficients of the filter equation.
2. From Fig. 8.1b, it is possible to infer that the simulation result is in agreement
with the theoretical result.
Task
1. Write a python code to obtain the impulse response of the filter whose difference
equation is given by y½n] - 12 y½n - 1] = x½n] and comment on the observed output.

8.2 Mapping Techniques in the Design of IIR Filter

Two common approaches in the design of IIR filters are:

Approach 1: Design an analogue IIR filter to meet the given design requirement and
convert the analogue filter into an equivalent digital filter using mapping tech-
niques like backward difference method, impulse invariant technique (IIT),
bilinear transformation technique (BLT), matched Z-transform technique and
so on.
8.2 Mapping Techniques in the Design of IIR Filter 307

Approach 2: IIR filter is designed using an algorithmic design procedure by solving

a set of linear and non-linear equations using a computer or dedicated hardware.
This section focuses on the first approach in which analogue filters are converted
into an equivalent digital filter using mapping techniques. While performing the
mapping techniques, care must be taken to map stable analogue filter into a stable
digital filter.

8.2.1 Backward Difference Method

The relationship between ‘s’ domain and ‘z’ domain using the backward difference
method is given by

1 - z-1
s= ð8:4Þ
T

and

1
z= ð8:5Þ
1 - sT

This section displays the mapping of analogue filter into equivalent digital filter
using backward difference method.
Experiment 8.2 Mapping of S-Plane to Z-Plane Using Backward Difference
Method
The main objective of this experiment is to prove that stable analogue filter will be
mapped to a stable digital filter using backward difference method. The python code,
which performs the mapping from S-plane to Z-plane, is given in Fig. 8.2.
Inference
1. From Fig. 8.3a, it is possible to infer that points in the left half of S-plane
(i.e. σ ≤ 0) are mapped into inside and on the unit circle in Z-plane.
2. This confirms that the stable analogue filter can be mapped into a stable digital
filter using the backward difference mapping technique.
3. On the other hand, from Fig. 8.3b, it is evident that the points on the right side of
the S-plane (i.e. σ > 0) are mapped into outside the unit circle of the Z-plane.

Experiment 8.3 Conversion of Analogue Filter to Digital Filter Using Backward

Difference Method
This experiment discusses the conversion of analogue filter to digital filter using
backward difference approach. Let us consider the transfer function of the analogue
filter is H ðsÞ = ðs2 þ3sþ2
1
Þ. Assume a sampling period is 0.1 s. The relationship
between ‘S-domain’ and ‘Z-domain’ in the backward difference method is given by
308 8 Infinite Impulse Response Filter

#Mapping between S to Z plane

import numpy as np
import matplotlib.pyplot as plt
omega1=np.linspace(-15, 15, 20)
omega2 = np.array([1 in range(len(omega1))])
omega2=omega1
sigma1=np.linspace(-10, 0, 20)
sigma2 = np.array([1 in range(len(sigma1))])
sigma2 = sigma1
#np.linspace(-10, 0, 100)
S=[[0 for i in range(len(sigma2))] for j in range(len(omega2))]
Z=S
T=0.1
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k'),
for j in range(len(sigma2)):
for i in range(len(omega2)):
S[i][j]=complex(sigma1[i],omega1[j])
Z[i][j]=1/(1-(S[i][j]*(T)))
plt.subplot(2,1,1),plt.plot(sigma1[i],omega1[j],'bx')
plt.title('s-plane'), plt.xlabel('$\sigma$'),plt.ylabel('$j\u03A9})$')
plt.subplot(2,1,2),plt.plot(np.real(Z[i][j]),np.imag(Z[i][j]),'rx')
plt.title('z-plane'), plt.xlabel('$\sigma$'),plt.ylabel('$j\omega})$')
plt.tight_layout()

Fig. 8.2 Python code for mapping S to Z plane using backward difference method. (a) σ ≤ 0. (b)
σ>0

s-plane s-plane
10
10
j:

j:

0 0
–10
–10
–10 –8 –6 –4 –2 0 2 4 6 8 10
s s
z-plane z-plane
1 1

0
jw

jw

–1 –1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00 –1.0 –0.5 0.0 0.5 1.0
s s

(a) s ≤0 (b) s > 0

Fig. 8.3 Result of python code given in Fig. 8.2

8.2 Mapping Techniques in the Design of IIR Filter 309

1
-------------
s^2 + 3 s + 2
import control as ss
s1 = ss.tf(1, [1,3,2]) 0.007576 z^2 - 2.22e-16 z + 2.22e-16
print(s1) ------------------------------------
yd = s1.sample(0.1, method='backward_diff') z^2 - 1.742 z + 0.7576
print(yd) dt = 0.1

(a) Python program (b) Simulation result

Fig. 8.4 Python code and simulation result for backward difference mapping method. (a) Python
program. (b) Simulation result

1 - z-1
s=
T

Substituting the value of T = 0.1 in the above expression, we get

1 - z-1
s=
0:1

The above expression can be simplified as

s = 10 - 10z - 1

Substituting the above relation in the transfer function of the analogue filter, we
get

1
H ðzÞ =
ð10 - 10z - 1 Þ2 þ 3ð10 - 10z - 1 Þ þ 2

Simplifying the above expression, we get

0:007576z2
H ðzÞ =
z2 - 1:742z þ 0:7576

The frequency responses of both the analogue and digital filters are computed
using python code, shown in Fig. 8.4a, and its corresponding simulation result is
shown in Fig. 8.4b. Before execution of the python code, the ‘python-control’
package must be installed using the pip command, which is given by ‘pip install
control’. The new python commands used in this python program are (1) xx.tf and
(2) yy.sample. The result of the python code given in Fig. 8.4a is shown in Fig. 8.4b.
Inference
The following inferences can be made from this experiment:
310 8 Infinite Impulse Response Filter

1. From the simulation result shown in Fig. 8.4b, it is evident that the denominator
polynomial function is exactly matched with the theoretical result.
2. The denominator polynomial corresponds to the poles of the system. Thus,
analogue filter is mapped into an equivalent digital filter, which is in agreement
with the theoretical result.
Experiment 8.4 Mapping Stable Analogue Filter to a Stable Digital Filter Using
Backward Difference Method
This experiment deals with mapping a stable analogue filter to a stable digital filter
using the backward difference method. The transfer function of a stable second-order
filter considered in this example is given by

4
H ðsÞ =
s2 þ 2:82s þ 4

This filter is converted into an equivalent digital filter H(z) using backward
difference method. The python code to verify the conversion of a second-order
stable analogue filter into a stable digital filter is given in Fig. 8.5, and the simulation
result is depicted in Fig. 8.6. This python code will work for the second-order filter
only. The simulation result of this experiment is shown in Fig. 8.6.
Inference
From Fig. 8.6b, it is possible to observe that the poles of H(s) lie left half of the S-
plane, confirming that the analogue filter is stable. Similarly, from this figure, it is
possible to know that the poles of H(z) lie inside the unit circle. Hence, the backward
difference mapping method preserves the stability criterion during mapping.

8.2.2 Impulse Invariant Technique

In impulse invariant technique, the digital filter is designed by sampling the impulse
response of the analogue filter. The pole at s = sp is mapped to a pole at z = esp T in
the digital filter. Impulse invariant technique performs many-to-one mapping; hence,
it suffers from an aliasing problem. Thus, impulse invariant technique is useful if the
analogue filter is band-limited. The step followed in impulse invariant technique is
given in Fig. 8.7.
The impulse response of the analogue filter is represented by h(t). It is sampled to
get h[nT]. Upon taking Z-transform of the sampled impulse response, the transfer
function of the digital filter is obtained, which is represented as H(z).
Experiment 8.5 Mapping of S-Plane to Z-Plane Using Impulse Invariant
Technique
This experiment deals with mapping the S-plane to Z-plane using impulse invariant
technique.
Case 1: Mapping the points on the j Ω axis of the S-plane
8.2 Mapping Techniques in the Design of IIR Filter 311

import numpy as np
import control as ss
from scipy import signal
import matplotlib.pyplot as plt
num=[4]
den=[1,2*np.sqrt(2),4]
T=0.1
fsam=1/T
s1=ss.tf(num,den)
print('H(s) =', s1)
a1=1
b1=-(2+den[1]*T)/(1+(den[1]*T)+den[2]*(T**2))
c1=1/(1+(den[1]*T)+den[2]*(T**2))
num1=[num[0]*c1*(T**2),0,0]
den1=[a1,b1,c1]
yd = s1.sample(T, method='backward_diff')
print('Using Built in function: H(z) =',yd)
s2=ss.tf(num1,den1,T)
print('H(z) =', s2)
ps,zs=ss.pzmap(s1)
# Pole-zero plot
plt.subplot(2,1,1),plt.plot(ps.real, ps.imag, 'kx', ms=10),plt.xlabel('$\sigma$'),
plt.ylabel('$j\Omega$'),plt.title('Pole-zero plot of H(s)'),plt.grid()
z, p, k = signal.tf2zpk(num1,den1)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k--')
plt.plot(p.real, p.imag, 'rx', ms=7.5),plt.xlabel('$\sigma$'),
plt.ylabel('$j\omega$'),plt.title('Pole-zero plot of H(z)'),plt.grid()
plt.tight_layout(),plt.show()

Fig. 8.5 Python code for mapping analogue filter to digital filter

The python code for mapping the points on the j Ω axis onto unit circle is given in
Fig. 8.8, and its corresponding output is shown in Fig. 8.9. Figure 8.9 confirms that
the points on the j Ω axis in the S-plane are mapped onto a unit circle in the Z-plane.
Case 2: Mapping the left half of the S-plane
The python code for mapping the points in left half of S-plane are mapped into
inside the unit circle is given in Fig. 8.10, and its corresponding output is shown in
Fig. 8.11. Figure 8.11 confirms that the points in the left half of S-plane are mapped
into within the unit circle of the Z-plane.
Case 3: Mapping right half of S-plane
The points in the right half of S-plane are mapped into outside the unit circle in the
Z-plane. The python code for mapping the points in right half of S-plane are mapped
into outside the unit circle is given in Fig. 8.12, and its corresponding output is
312 8 Infinite Impulse Response Filter

Pole-zero plot of H(s)

j:
–1

–1.48 –1.46 –1.44 –1.42 –1.40 –1.38 –1.36 –1.34

s
Pole-zero plot of H(z)
1

0
jw
–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
s

(a) Transfer function (b) Pole zero plot

Fig. 8.6 Simulation result. (a) Transfer function. (b) Pole-zero plot

Impulse response Sampling Sampled impulse Z-transform Transfer function of

of analog filter h(t) response h(nT) digital filter H(z)

Fig. 8.7 Steps in impulse invariant technique

#Mapping between S to Z plane using IIT

import numpy as np
import matplotlib.pyplot as plt
omega1=np.linspace(-15, 15, 50)
sigma1=np.zeros(len(omega1))
S=sigma1+1j*omega1
T=1
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k'),
plt.subplot(2,1,1),plt.plot(sigma1,omega1,'bx')
plt.title('S-plane'),plt.xlabel('$\sigma$'),plt.ylabel('$j\Omega$')
z=np.exp(S*T)
plt.subplot(2,1,2),plt.plot(np.real(z),np.imag(z),'rx')
plt.title('Z-plane'),plt.xlabel('$\sigma$'),plt.ylabel('$j\omega})$')
plt.tight_layout()

Fig. 8.8 Python code for mapping points on the j Ω axis onto the unit circle

shown in Fig. 8.13. From Fig. 8.13, it is possible to infer that the points in the right
half of the S-plane are mapped outside the unit circle in the Z-plane.
Thus, a stable analogue filter can be mapped to an equivalent stable digital filter
using the impulse invariant technique.
8.2 Mapping Techniques in the Design of IIR Filter 313

S-plane

10

0
j:

–10

–0.04 –0.02 0.00 0.02 0.04

s
Z-plane
1

0
jw

–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
s

Fig. 8.9 Result of the python code given in Fig. 8.8

import numpy as np
import matplotlib.pyplot as plt
omega1=np.linspace(-15, 15, 30)
omega2 = np.array([1 in range(len(omega1))])
omega2=omega1
sigma1=np.linspace(-10, -0.1, 30)
sigma2 = np.array([1 in range(len(sigma1))])
sigma2 = sigma1
S=[[0 for i in range(len(sigma2))] for j in range(len(omega2))]
Z=S
T=0.9
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k'),
for j in range(len(sigma2)):
for i in range(len(omega2)):
S[i][j]=complex(sigma1[i],omega1[j])
Z[i][j]=np.exp(S[i][j]*(T))
plt.subplot(2,1,1),plt.plot(sigma1[i],omega1[j],'bx')
plt.title('S-plane'), plt.xlabel('$\sigma$'),plt.ylabel('$j\Omega$')
plt.subplot(2,1,2),plt.plot(np.real(Z[i][j]),np.imag(Z[i][j]),'rx')
plt.title('Z-plane'), plt.xlabel('$\sigma$'),plt.ylabel('$j\omega})$')
plt.tight layout()

Fig. 8.10 Python code of mapping left half of S-plane

314 8 Infinite Impulse Response Filter

S-plane

10

0
j:

–10

–10 –8 –6 –4 –2 0
s
Z-plane
1

0
jw

–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
s

Fig. 8.11 Result of python code given in Fig. 8.10

import numpy as np
import matplotlib.pyplot as plt
omega1=np.linspace(-2, 2, 30)
omega2 = np.array([1 in range(len(omega1))])
omega2=omega1
sigma1=np.linspace(0.1, 2, 30)
sigma2 = np.array([1 in range(len(sigma1))])
sigma2 = sigma1
S=[[0 for i in range(len(sigma2))] for j in range(len(omega2))]
Z=S
T=1
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k'),
for j in range(len(sigma2)):
for i in range(len(omega2)):
S[i][j]=complex(sigma1[i],omega1[j])
Z[i][j]=np.exp(S[i][j]*(T))
plt.subplot(2,1,1),plt.plot(sigma1[i],omega1[j],'bx')
plt.title('S-plane'),plt.xlabel('$\sigma$'),plt.ylabel('$j\Omega$')
plt.subplot(2,1,2),plt.plot(np.real(Z[i][j]),np.imag(Z[i][j]),'rx')
plt.title('Z-plane'),plt.xlabel('$\sigma$'),plt.ylabel('$j\omega})$')
plt.tight_layout()

Fig. 8.12 Python code of mapping right half of S-plane

8.2 Mapping Techniques in the Design of IIR Filter 315

S-plane
2

0
j:

–2
0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00
s
Z-plane
5

0
jw

–5

–2 0 2 4 6
s

Fig. 8.13 Result of python code given in Fig. 8.12

Experiment 8.6 Many-to-One Mapping in the Impulse Invariant Mapping

Technique
This experiment tries to prove that many to one mapping will have happened in the
impulse invariant mapping technique. The python code to verify the many-to-one
mapping for the impulse invariant technique is given in Fig. 8.14, and its
corresponding output is shown in Fig. 8.15. In Fig. 8.15, the symbol ‘◊’ represents
points that are mapped between the analogue frequency range from - Tπ to Tπ and the
digital frequency, whereas the symbol ‘x’ represents the points that are mapped
between analogue frequency in the range - 3π 3π
T to T the digital frequency.

Inferences
The following inferences can be made from this experiment:
1. Overlapping of the symbols in Fig. 8.15 implies that impulse invariant technique
is basically a many-to-one mapping.
2. Many-to-one mapping leads to an ‘aliasing problem’ in impulse invariant
technique.
3. Hence, impulse invariant technique is suitable for the design of lowpass and
bandpass filters.
4. It is not advisable to use it in highpass and bandstop filters design.
Experiment 8.7 Conversion of Analogue Filter into a Digital Filter Using IIT
This experiment discusses the conversion of an analogue filter to a digital filter using
impulse invariant technique. Let us consider the transfer function of the analogue
0:5ðsþ3Þ
filter H ðsÞ = ðsþ1 Þðsþ4Þ to be converted into the transfer function of the digital filter
H(z) using the impulse invariant technique. Assume the sampling frequency to be
20 Hz.
316 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
T=0.5
omega1=np.linspace(-np.pi/T, np.pi/T, 15)
omega2 = np.array([1 in range(len(omega1))])
omega2=omega1
omega11=np.linspace(-3*(np.pi/T), 3*(np.pi/T), 15)
omega21 = np.array([1 in range(len(omega11))])
omega21=omega11
sigma1=np.linspace(0, 0, 15)
sigma2 = np.array([1 in range(len(sigma1))])
sigma2 = sigma1
S=[[0 for i in range(len(sigma2))] for j in range(len(omega2))]
S1=[[0 for i in range(len(sigma2))] for j in range(len(omega21))]
Z,Z1=S,S1
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k'),
for j in range(len(sigma2)):
for i in range(len(omega2)):
S[i][j]=complex(sigma1[i],omega1[j])
Z[i][j]=np.exp(S[i][j]*(T))
S1[i][j]=complex(sigma1[i],omega11[j])
Z1[i][j]=np.exp(S1[i][j]*(T))
plt.subplot(2,1,1),plt.plot(sigma1[i],omega1[j],'bx'),plt.title('S-plane')
plt.xlabel('$\sigma$'),plt.ylabel('$j\Omega$')
plt.subplot(2,1,2),plt.plot(np.real(Z[i][j]),np.imag(Z[i][j]),'rd',markersize=12)
plt.subplot(2,1,2),plt.plot(np.real(Z1[i][j]),np.imag(Z1[i][j]),'kx',markersize=6)
plt.title('Z-plane'),plt.xlabel('$\sigma$'),plt.ylabel('$j\omega})$')
plt.tight_layout()

Fig. 8.14 Python code for many-to-one mapping of impulse invariance technique

Step 1: By using partial fraction expansion, the given analogue transfer function
H(s) can be written as

A B
H ðsÞ = þ
ð s þ 1 Þ ð s þ 4Þ

From the above expression, it is possible to write as

Aðs þ 4Þ þ Bðs þ 1Þ = 0:5ðs þ 3Þ ð8:6Þ

Substituting s = -4 in the above equation, we get Bð- 3Þ = - 0:5, B = 1=6.

Substituting s = -1 in Eq. (8.6), we get 3A = 1, A = 1=3.
Substituting the values of ‘A’ and ‘B’ in the expression of H(s), we get
8.2 Mapping Techniques in the Design of IIR Filter 317

S-plane
5

0
j:

–5
–0.04 –0.02 0.00 0.02 0.04
s
Z-plane
1

0
jw

–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
s

Fig. 8.15 Result of python code given in Fig. 8.14

1=3 1=6
H ðsÞ = þ
ð s þ 1 Þ ð s þ 4Þ

Step 2: Transfer function H(z)

The digital transfer function H(z) using impulse invariant technique is given by

N
Ai
H ðzÞ =
i=1
1 - esi T z - 1

In this case, N = 2, the expression for H(z) is given by

2
Ai A1 A2
H ðzÞ = = þ
i=1
1 - esi T z - 1 1 - es1 T z - 1 1 - es2 T z - 1

Substituting A1 = 1/3, A2 = 1/6, s1 = - 1, s2 = - 4 and T = 1/20 = 0.05 in the

above expression, we get

1=3 1=6
H ðzÞ = þ
ð1 - e - 0:05 z - 1 Þ ð1 - e - 0:2 z - 1 Þ

Python code to verify the theoretical result with the simulation result is given in
Fig. 8.16, and its corresponding output is shown in Fig. 8.17. From Fig. 8.16, the
following new tool imported for the simulation, they are (1) sympy, (2) control and
(3) scipy. In addition, the following new python commands are used for this
simulation are (1) symbols—used to define the symbol (z^(-1) and ‘+’),
318 8 Infinite Impulse Response Filter

Fig. 8.16 Python code for

analogue filter to digital import control as ss
filter conversion using import numpy as np
impulse invariance method from sympy import symbols
from scipy import signal
z=symbols('z^-1')
z1=symbols(' + ')
s1 = ss.tf([0.5,1.5], [1,5,4])
print('H(s) ='),print(s1)
T=1/20
y=['+']
yy=signal.residue([0.5,1.5],[1,5,4])
A=[0 for i in range(len(yy[0]))]
S=[0 for i in range(len(yy[1]))]
A[0]=yy[0][0]
S[0]=yy[1][0]
y1=np.append((A[0]/(1-np.exp(S[0]*T)*z)),y)
for i in range(len(yy[0])-1):
A[i+1]=yy[0][i+1]
S[i+1]=yy[1][i+1]
y1=np.append(y1,(A[i+1]/(1-np.exp(S[i+1]*T)*z)))
print('H(z) = '),print(y1)

H(s) =

0.5 s + 1.5
-------------
s^2 + 5 s + 4

H(z) =
[0.333333333333333/(1 - 0.951229424500714*z^-1) '+'
0.166666666666667/(1 - 0.818730753077982*z^-1)]

Fig. 8.17 Simulation result of python code given in Fig. 8.16

(2) residue—used to obtain the residues and poles. The result of transfer function in
S domain and Z domain is displayed in Fig. 8.17.
Note: This python code converts the analogue filter into the digital filter,
which has distinct poles.
Inference
From Fig. 8.17, it is possible to observe that the simulation result is on par with the
theoretical result.
8.2 Mapping Techniques in the Design of IIR Filter 319

import control as ss
import numpy as np
import matplotlib.pyplot as plt
from sympy import symbols
from scipy import signal
z=symbols('z^-1')
num,den=[1,2],[1,6,5]
s1 = ss.tf(num, den)
print('H(s) =',s1)
T=1/20
zs,ps,ks=signal.tf2zpk(num,den)
r,p,k=signal.residue(num,den)
theta = np.linspace(0, np.pi*2, 500)
zz=np.zeros(len(theta))
plt.subplot(2,1,1),plt.plot(zz, theta, 'k--'),plt.plot(ps.real,ps.imag, 'bx',ms=10)
plt.plot(zs.real,zs.imag, 'go',ms=10),plt.xlabel('$\sigma$'),
plt.ylabel('$j\Omega$'),plt.title('Pole-zero plot H(s)')
y1=np.append((r[0]/(1-np.exp(p[0]*T)*z)),(r[1]/(1-np.exp(p[1]*T)*z)))
print('H(z) = ',y1)
z=np.exp(p*T)
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k--'),
plt.plot(z.real,z.imag, 'rx',ms=10),
plt.xlabel('$\sigma$'),plt.ylabel('$j\omega$'),plt.title('Pole-zero plot H(z)')

Fig. 8.18 Python code to map poles in S-domain to Z-domain using IIT

Experiment 8.8 Conversion of a Stable Analogue Filter to a Stable Digital Filter

Using the Impulse Invariance Technique (IIT)
This experiment is to verify the stability of the analogue filter to digital filter during
conversion using impulse invariant technique. The python code to verify the stability
of the analogue filter to digital conversion using IIT is given in Fig. 8.18. Here, we
have considered the second-order stable analogue filter with two poles at -5 and -1
and a zero at -2. This stable analogue filter is converted into a digital filter using the
impulse invariance method. The simulation result of the python code, which is given
in Fig. 8.18, is shown in Fig. 8.19.
Inference
1. From Fig. 8.19, it is possible to observe that the poles of the analogue filter lay left
half of S-plane; it is evident that the analogue filter is stable.
2. Similarly, the pole-zero plot of H(z) is shown in Fig. 8.19. From the figure, it is
possible to observe that all poles are lying inside the unit circle.
3. This implies that the digital filter is also stable. Therefore, the stability is retained
while mapping the analogue filter into the digital filter using the impulse invari-
ance method.
320 8 Infinite Impulse Response Filter

Pole-zero plot H(s)

6

4
j:

2
0
–5 –4 –3 –2 –1 0
s
Pole-zero plot H(z)
1

0
jw

–1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
s

Fig. 8.19 Pole-zero plot

8.2.3 Bilinear Transformation Technique (BLT)

In bilinear transformation technique, the mapping from S-plane to Z-plane is done

using the relation

2 1 - z-1
s= ð8:7Þ
T 1 þ z-1

Unlike impulse invariant technique (IIT), there is no aliasing problem in bilinear

transformation technique. The relationship between the analogue frequency (Ω) and
digital frequency (ω) in BLT is given in Eq. (8.5)

2 ω
Ω= tan ð8:8Þ
T 2

The above equation can also be expressed as

ΩT
ω = 2 tan - 1 ð8:9Þ
2

From Eqs. (8.8) and (8.9), it is possible to infer that there exists a non-linear
relationship between analogue and digital frequency in bilinear transformation
technique, which is termed as warping.
8.2 Mapping Techniques in the Design of IIR Filter 321

import numpy as np
import matplotlib.pyplot as plt
T=[1,0.5,0.25,0.1]
omega=np.linspace(-15,15,20)
for i in range(len(T)):
domega=2*np.arctan(omega*(T[i]/2))
plt.plot(omega,domega,'-*'),plt.title('BLT Mapping')
plt.xlabel('$\Omega$ rad/sec'),plt.ylabel('$\omega$ rad/sample')
plt.legend(['$T={}$ Sec'.format(T[0]),'$T={}$ Sec'.format(T[1]),'$T={}$
Sec'.format(T[2]),'$T={}$ Sec'.format(T[3])],loc=0)
plt.tight_layout()

Fig. 8.20 Python code to display the relationship between analogue and digital frequency

BLT Mapping
3
T = 1 sec
T = 0.5 sec
2
T = 0.25 sec
T = 0.1 sec
1
w rad/sample

–1

–2

–3
–15 –10 –5 0 5 10 15
: rad/sec

Fig. 8.21 Relationship between analogue and digital frequencies

Experiment 8.9 Display the Relationship Between Analogue Frequency

and Digital Frequency Using Bilinear Transformation (BLT) Technique
This experiment displays the relationship between analogue frequency and digital
frequency using BLT. The python code given in Fig. 8.20 gives the relationship
between analogue and digital frequency using BLT. The simulation result of the
python code given in Fig. 8.20 is shown in Fig. 8.21. The sampling intervals are
considered as [1, 0.5, 0.25, 0.1]. The relationship between analogue and digital
frequency for different sampling intervals is shown in Fig. 8.21.
Inferences
1. From this figure, it is possible to observe that with the value of T > 0.1, the
mapping between analogue and digital frequency is linear at the low-frequency
range and non-linear at the high-frequency range.
322 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
omega1=np.linspace(-15, 15, 20)
omega2 = np.array([1 in range(len(omega1))])
omega2=omega1
#sigma1=np.linspace(0, 0, 20)# for sigma=0
#sigma1=np.linspace(-15, -0.5, 20)# for sigma < 0
sigma1=np.linspace(0.5, 5, 20)# for sigma > 0
sigma2 = np.array([1 in range(len(sigma1))])
sigma2 = sigma1
S=[[0 for i in range(len(sigma2))] for j in range(len(omega2))]
Z=S
T=0.9
theta = np.linspace(0, np.pi*2, 500)
circle = np.exp(1j*theta)
plt.subplot(2,1,2),plt.plot(circle.real, circle.imag, 'k'),
for j in range(len(sigma2)):
for i in range(len(omega2)):
S[i][j]=complex(sigma1[i],omega1[j])
Z[i][j]=(1+(S[i][j]*(T/2)))/(1-(S[i][j]*(T/2)))
plt.subplot(2,1,1),plt.plot(sigma1[i],omega1[j],'bx')
plt.title('S-plane'), plt.xlabel('$\sigma$'),plt.ylabel('$j\Omega$')
plt.subplot(2,1,2), plt.plot(np.real(Z[i][j]),np.imag(Z[i][j]),'rx')
plt.title('Z-plane'), plt.xlabel('$\sigma$'),plt.ylabel('$j\omega})$')
plt.tight_layout()

Fig. 8.22 Python code for frequency mapping using BLT

2. This non-linear relationship is termed as ‘warping’.

3. In order to overcome this warping effect, prewarping is necessary for bilinear
transformation technique.
4. When T = 0.1, the mapping is almost linear for both low- and high-frequency
ranges.
Experiment 8.10 Illustration of BLT Preserves Stability Criterion
The objective of this experiment is to prove that stable analogue filter will be mapped
to a stable digital filter using BLT. The stability conditions are verified with the
python code. The python code maps S-plane into an equivalent Z-plane using BLT.
The S-plane is represented as s = σ + jΩ, whereas the Z-plane is represented as
z = rejω. Three different cases considered in this example are σ = 0, σ < 0 and σ > 0.
The python code to verify the stability preservation of BLT is given in Fig. 8.22.
Varying the values of σ = 0, σ < 0 and σ > 0 in the python code, the mapping results
are shown in Fig. 8.23.
8.2 Mapping Techniques in the Design of IIR Filter 323

S-plane S-plane
10 10
j:

j:
0 0
–10 –10
–0.04 –0.02 0.00 0.02 0.04 –14 –12 –10 –8 –6 –4 –2 0
s s
Z-plane Z-plane
1 1
jw

jw
0 0

–1 –1
–1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00 –1.00 –0.75 –0.50 –0.25 0.00 0.25 0.50 0.75 1.00
s s
as=0 bs<0
S-plane
10
j:

0
–10
1 2 3 4 5
s
Z-plane
5
jw

–5
–4 –3 –2 –1 0 1 2
s
cs>0
Fig. 8.23 Simulation result of python code given in Fig. 8.22. (a) σ = 0. (b) σ < 0. (c) σ > 0

Inferences
1. When σ = 0 indicate jΩ in the S-plane. From Fig. 8.23a, it is possible to observe
that the entire jΩ axis is mapped to points on the unit circle in the Z-plane.
2. When σ < 0 represent the left half of S-plane. All the points in the LHS of S-plane
are mapped into within the unit circle of the Z-plane.
3. When σ > 0 are the points in the RHS of the S-plane. These points are mapped
into the outside of the unit circle in the Z-plane.
4. An analogue filter is stable if the poles lie in the left half of S-plane. A digital filter
is stable if the poles lie within the unit circle. The BLT ensures that the stable
analogue filter will be mapped as a stable digital filter.

8.2.4 Matched Z-Transform Technique

In matched Z-transform technique, the poles and zeros of the analogue filter are
mapped to Z-plane using the relation z = esT. The matched Z-transform gives the
same pole location as impulse invariant technique but a different zero location.
324 8 Infinite Impulse Response Filter

Experiment 8.11 Conversion of Analogue Filter to Equivalent Digital Filter

Using Matched Z-Transform Method
This experiment discusses the conversion of an analogue filter into its equivalent
digital filter using matched Z transform. Let us consider a second-order analogue
ðsþ3Þ
filter transfer function is H ðsÞ = ðsþ1 Þðsþ4Þ to be converted into digital filter using
matched Z-transform. Assume T = 0.1 s. The matched Z transform (MZT) method
directly maps the poles and zeros of an analogue filter into the poles and zeros of the
digital filter. The transfer function of the analogue filter is given as

M
k = 1 ðs - zk Þ
H ðsÞ = N ð8:10Þ
k = 1 ð s - pk Þ

The transfer function of the equivalent digital filter is computed by replacing the
term (s - zk) with ð1 - ezk T z - 1 Þ. The transfer function of the equivalent digital filter
can be written as

M zk T - 1
k = 1 ð1 - e z Þ
H ðzÞ = N ð8:11Þ
k = 1 ð1 - e
p k T z - 1Þ

where T is the sampling period. The transfer function of the analogue filter consid-
ered in this example is given by

ðs þ 3Þ
H ðsÞ =
ð s þ 1 Þ ð s þ 4Þ

From the above equation, the zeros and poles are computed as z1 = -3, and
p1 = -1, p2 = -4, respectively.
Substituting the values of zeros, poles and T = 0.1 in Eq. (8.11), we get

ð1 - e - 0:3 z - 1 Þ
H ðzÞ =
ð1 - e - 0:1 z - 1 Þð1 - e - 0:4 z - 1 Þ

Simplifying the above equation, we get

ð1 - 0:7408z - 1 Þ
H ðzÞ =
ð1 - 0:9048z - 1 Þð1 - 0:6703z - 1 Þ

Further simplifying the above equation, we get

zðz - 0:7408Þ
H ðzÞ =
ðz2 - 1:5751z þ 0:6065Þ
8.2 Mapping Techniques in the Design of IIR Filter 325

import numpy as np
from scipy import signal
import control as ss
T=0.1
z=[-3]
p =[-1, -4] # analog poles
b_s=signal.convolve([1],[1,-z[0]])
a_s=signal.convolve([1,-p[0]],[1,-p[1]])
Hs=ss.tf(b_s,a_s)
print('Transfer function H(s) =',Hs)
b_z=signal.convolve([1,0], [1, -(np.exp(z[0]*T))])
a_z = signal.convolve([1, -(np.exp(p[0]*T))], [1, -(np.exp(p[1]*T))])
Hz=ss.tf(b_z,a_z,T)
print('Transfer function H(z) =',Hz)

Fig. 8.24 Python code for Experiment 8.11

Fig. 8.25 Result of the

python code is given in Transfer function H(s) =
Fig. 8.24 s+3
-------------
s^2 + 5 s + 4

Transfer function H(z) =

z^2 - 0.7408 z
----------------------
z^2 - 1.575 z + 0.6065

dt = 0.1

Python code to verify this experiment is given in Fig. 8.24, and the corresponding
simulation result is shown in Fig. 8.25.
Inferences
1. Figure 8.24 shows that the signal.convole command is used here to compute the
product of two zeros or two poles term.
2. This python code indicates that direct mapping has existed between analogue
filter poles to digital filter poles and analogue filter zeros to digital filter zeros.
3. From Fig. 8.25, it is possible to observe that the transfer function of the digital
filter is on par with the theoretical result.
326 8 Infinite Impulse Response Filter

8.3 Analog Frequency Transformation

A normalized lowpass filter can be transformed into a desired lowpass, highpass,

bandpass or bandstop filter by frequency transformation technique. Table 8.1 sum-
marizes different analogue frequency transformations.
Experiment 8.12 Analogue Frequency Transformation
The objective of this experiment is to convert a normalized first-order lowpass
Butterworth filter into an equivalent bandpass and band reject filter with the lower
and upper cut-off frequencies of 3 and 5 rad/s, respectively, using the frequency
transformation method. A normalized first-order lowpass Butterworth filter transfer
function is given by H ðsÞ = ðsþ1 Þ. Convert this filter into an equivalent (1) bandpass
1

filter and (2) band reject filter with the lower and upper cut-off frequencies of 3 and
5 rad/s, respectively, using the frequency transformation method.
Step 1: Converting the prototype filter into its equivalent bandpass filter
The prototype filter can be converted into its equivalent band pass filter by using
the frequency transformation s → ssðΩþΩ u Ωl
2

u - Ωl Þ
. In this experiment, the value of Ωu = 5
rad/s and Ωl = 3 rad/s. Hence, the s can be computed as

s2 þ 15 s2 þ 15
s→ =
sð2Þ 2s

s2 þ15
Now replacing s in the normalized lowpass filter transfer function by s → 2s ,
we get

1
H 1 ðsÞ = s2 þ15
2s þ1

Simplifying the above expression, we get

2s
H 1 ðsÞ =
s2 þ 2s þ 15

Step 2: Obtaining the transfer function of the band reject filter

Table 8.1 Analogue fre- S. No. Type of transformations Transformation

quency transformation
1 Lowpass to lowpass s → Ωsc
Ωc
2 Lowpass to highpass s→ s
s2 þΩu Ωl
3 Lowpass to bandpass s→ sðΩu - Ωl Þ
sðΩu - Ωl Þ
4 Lowpass to bandstop s→ s2 þΩu Ωl
8.3 Analog Frequency Transformation 327

import numpy as np
import matplotlib.pyplot as plt
import control as ss
num,den=[1],[1,1]
num1,den1=[2, 0],[1,2,15]
num2,den2=[1,0,15],[1,2,15]
s = ss.tf(num, den)
print(s)
s1 = ss.tf(num1, den1)
print(s1)
s2 = ss.tf(num2, den2)
print(s2)
omega1=np.linspace(0, 15, 100)
mag, phase, omega1=ss.freqresp(s, omega1)
mag1, phase1, omega1=ss.freqresp(s1, omega1)
mag2, phase2, omega1=ss.freqresp(s2, omega1)
plt.subplot(2,1,1),plt.plot(omega1,mag,'-.',omega1,mag1,'--',omega1,mag2,'-')
plt.title('Magnitude response'),plt.legend(['Prototype','BPF','BRF'],loc=0)
plt.xlabel('$\Omega$ in rad/sec'),plt.ylabel('|$H(j\Omega)$|')
plt.subplot(2,1,2),plt.plot(omega1,phase,'-.',omega1,phase1,'--',omega1,phase2,'-')
plt.title('Phase response'),plt.legend(['Prototype','BPF','BRF'],loc=0)
plt.xlabel('$\Omega$ in rad/sec'),plt.ylabel('∠$H(j\Omega)$)')
plt.tight_layout()

Fig. 8.26 Python code to obtain the magnitude and phase response

The given prototype filter can be converted into a band reject filter using the
frequency transformation s → ssð2ΩþΩ
u - Ωl Þ
u Ωl
. In this problem, the value of Ωu = 5 rad/s
and Ωl = 3 rad/s. Hence, the equivalent s is calculated as

s ð Ωu - Ω l Þ 2s
s→ = 2
s 2 þ Ωu Ωl s þ 15

Now replacing s in the normalized lowpass filter transfer function by s → 2s

s2 þ15,
we get

1
H 2 ðsÞ =
2s
s2 þ15 þ1

Simplifying the above expression, we get

s2 þ 15
H 2 ðsÞ =
s2 þ 2s þ 15

The following python code helps us to understand that the prototype filter H(s) is
converted into bandpass filter H1(s) and band reject filter H2(s). The python code is
given in Fig. 8.26, and the corresponding simulation result is shown in Fig. 8.27.
328 8 Infinite Impulse Response Filter

Magnitude response
1.0
Prototype

|H(j:)|
BPF
0.5 BRF

0.0
0 2 4 6 8 10 12 14
: in rad/sec
Phase response

1 Prototype
‘H(j:) BPF
0 BRF
–1

0 2 4 6 8 10 12 14
: in rad/sec

(a) Transfer function (b) Frequency response

Fig. 8.27 Simulation result of the python code given in Fig. 8.26. (a) Transfer function. (b)
Frequency response

Inferences
1. The transfer function of prototype normalized lowpass filter, desired bandpass
filter and band reject filter are shown in Fig. 8.27a, which is in agreement with the
theoretical result.
2. From the magnitude response shown in Fig. 8.27b, it is possible to observe that
the filter response is in agreement with the desired result.
3. From the phase response, it is possible to observe that the phase response is
non-linear.

8.4 Butterworth Filter

The squared magnitude response of Nth order Butterworth lowpass filter is given by

2 1
H N ejΩ = 2N
ð8:12Þ
Ω
1þ Ωc

where Ωc is cut-off frequency and N denotes the order of the filter. Butterworth filter
exhibits maximally flat response in both passband and stopband. Therefore, these
filters are called maximally flat filters or flat-to-flat filters. The salient features of
lowpass Butterworth filter are:
1. The magnitude response is a monotonically decreasing function of frequency.
2. The maximum gain occurs at Ω = 0.
8.4 Butterworth Filter 329

import numpy as np
import matplotlib.pyplot as plt
omega=np.linspace(0, 5, 100)
omegac=np.array([2])
N=[1,2,3,4,5]
for i in range(len(N)):
H=1/(1+(omega/omegac)**(2*N[i]))
plt.plot(omega,np.abs(H)),plt.title('Squared Magnitude Response')
plt.xlabel('$\Omega$ in rad/sec'),
plt.ylabel('|$H(j\Omega)|^2$')
plt.legend(['N = 1','N =2','N = 3','N = 4','N = 5'])
plt.tight_layout()

Fig. 8.28 Python code for squared magnitude response of Butterworth lowpass filter

3. The first (2N - 1) derivatives of an Nth order lowpass Butterworth filter are zero
at Ω = 0. Hence, Butterworth filters are termed as maximally flat magnitude
filters.
4. The high-frequency roll off of a Nth order Butterworth filter is 20N dB/decade.

Experiment 8.13 Magnitude Response of Butterworth Filter for Different

Filter Order
The objective of this experiment is to obtain the squared magnitude response of the
Butterworth lowpass filter for different orders. The python code to generate a
squared magnitude response of different orders of Butterworth lowpass filter is
given in Fig. 8.28, and the corresponding result is shown in Fig. 8.29.
Inferences
1. The squared magnitude response of Butterworth lowpass filter with different
orders is shown in Fig. 8.29. Here, the cut-off frequency is chosen as 2 rad/s,
and the orders are varied from 1 to 5.
2. From Fig. 8.29, it is possible to observe that the transition width decreases when
the order of the filter (N) increases. Also, there is no ripple in the passband and
stopband.
3. The squared magnitude response is a monotonically decreasing function of
frequency.
Experiment 8.14 Computing the Order of Butterworth Filter
The objective of this experiment is to compute the order of Butterworth filter for the
given filter specifications. The given specifications are as follows: (1) The passband
gain at 2 rad/s is 0 dB. (2) The stopband attenuation at 5 rad/s is 30 dB. (3) Passband
cut-off frequency Ωp = 2 rad/s. (4) Stopband cut-off frequency is Ωs = 5 rad/s.
The order of Butterworth filter can be computed using the following formula
330 8 Infinite Impulse Response Filter

Squared Magnitude Response

1.0 N=1
N=2
N=3
0.8
N=4
N=5
0.6
|H(j:)|2

0.4

0.2

0.0
0 1 2 3 4 5
: in rad/sec

Fig. 8.29 Squared magnitude response of Butterworth lowpass filter

Ap
10 - 10 - 1
log A
- s
10 10 -1
N= ð8:13Þ
Ωp
2 log Ωs

The symbol d.e denotes ceiling operator. Substituting the values given in the
specifications in the above equation, we get
-3
10 - 10 - 1
log - 30
10 10 - 1
N=
2 log 25

The order of the filter is calculated as

logð0:001Þ -3
N= = = d3:77e = 4
2 × ð- 0:3979Þ - 0:7958

The python code to compute the order of the Butterworth filter is given in
Fig. 8.30.
Inference
After executing the python code is given in Fig. 8.30. The result is Order of the filter
(N) = 4.0. It is in agreement with the theoretical result.
8.4 Butterworth Filter 331

Fig. 8.30 Python code for

Butterworth filter order import numpy as np
calculation Ap, As=-3, -30
omega_p,omega_s=2, 5
num1=10**(-Ap/10)-1
num2=10**(-As/10)-1
num=np.log(num1/num2)
den=2*np.log(omega_p/omega_s)
N=np.ceil(num/den)
print('Order of the filter (N) =')

Fig. 8.31 Python code to

generate transfer function of import control as ss
Butterworth filter import numpy as np
from scipy import signal
N=np.array([1,2,3,4,5,6,7,8],dtype=int)
for i in range(len(N)):
b,a=signal.butter(N[i],1,'low', analog=True)
s1=ss.tf(b,a)
print('H(s) for N ={} '.format(N[i]))
print(s1)

Task
1. Change the value of Ap and As in the python code, given in Fig. 8.30, execute the
program and observe the result of the order of the filter.
Experiment 8.15 Transfer Function of Normalized Butterworth Filter
of Different Orders
The objective of this experiment is to obtain the transfer function of normalized
Butterworth lowpass filter with different orders of the filter. The python code to
obtain the transfer function of Butterworth filter is given in Fig. 8.31. The simulation
result is given in Fig. 8.32.
Inference
1. Using the python code given in Fig. 8.31, Nth order normalized Butterworth
lowpass filter transfer function can be obtained.
2. The simulation result of the python code given in Fig. 8.31 is shown in Fig. 8.32,
and the maximum value of the order N is chosen as 8.
3. From Fig. 8.32, it possible to see that the transfer function of the normalized
Butterworth lowpass filter from first order to eighth order.
Experiment 8.16 Design of Butterworth Filter for a Given Specifications
The aim of this experiment is to design a Butterworth filter using bilinear transfor-
mation technique (BLT) that has a passband gain of 0 to -3 dB cut-off frequency of
2 kHz, and attenuation of at least 20 dB for frequencies greater than 5 kHz. Assume
the sampling frequency to be 20 kHz and sampling period T = 1 s.
332 8 Infinite Impulse Response Filter

H(s) for N =1
1
-----
s+1
H(s) for N =2
1
-----------------
s^2 + 1.414 s + 1
H(s) for N =3
1
---------------------
s^3 + 2 s^2 + 2 s + 1
H(s) for N =4
1
-----------------------------------------
s^4 + 2.613 s^3 + 3.414 s^2 + 2.613 s + 1
H(s) for N =5
1
-----------------------------------------------------
s^5 + 3.236 s^4 + 5.236 s^3 + 5.236 s^2 + 3.236 s + 1
H(s) for N =6
1
-----------------------------------------------------------------
s^6 + 3.864 s^5 + 7.464 s^4 + 9.142 s^3 + 7.464 s^2 + 3.864 s + 1
H(s) for N =7
1
---------------------------------------------------------------------------
s^7 + 4.494 s^6 + 10.1 s^5 + 14.59 s^4 + 14.59 s^3 + 10.1 s^2 + 4.494 s + 1
H(s) for N =8
1
-----------------------------------------------------------------------------------------
s^8 + 5.126 s^7 + 13.14 s^6 + 21.85 s^5 + 25.69 s^4 + 21.85 s^3 + 13.14 s^2 + 5.126 s + 1

Fig. 8.32 Simulation result of the python code given in Fig. 8.31

The specifications given in this experiment as follows:

1. Sampling frequency fsamp = 20 kHz
2. Pass band gain Ap = -3 dB
3. Stop band attenuation As = -20 dB
4. Pass band frequency fp = 2 kHz
5. Stop band frequency fs = 5 kHz
Converting the pass band and stop band frequencies from Hz to radians per
sample, it is necessary to compute the ωp and ωs from the frequency fp and fs,
which are given below
8.4 Butterworth Filter 333

fp
ωp = 2π
f samp

Substituting the values of fp and fsamp in the above expression, we get

2
ωp = 2π = 0:2π
20

Now the expression for ωs is given by

fs
ωs = 2π
f samp

Substituting the values of fs and fsamp in the above expression, we get

5
ωs = 2π = 0:5π
20

Step 1: Prewarping
The prewarping process must be done for the bilinear transformation technique to
preserve one-to-one mapping in the frequency transformation from digital to
analogue.
The corresponding analogue frequencies Ωp and Ωs are obtained using
prewarping technique which is given by

2 ωp
Ωp = tan
T 2

Substituting the value of ωp = 0.2π and T = 1 in the above expression, we get

0:2π
Ωp = 2 tan = 0:650
2

Similarly, the value of Ωs is computed as

0:5π
Ωs = 2 tan =2
2

Step 2: To determine the order of the filter

The expression for the order of Butterworth filter is given by
334 8 Infinite Impulse Response Filter

Ap
-
10 10 -1
log A
- s
10 10 -1
N=
Ωp
2 log Ωs

Substituting the value of Ap = -3 dB, As = -20 dB, Ωp = 0.650 and Ωs = 2 in

the above expression, we get

10ð10Þ - 1
3

log
10ð Þ - 1
20
10
N= 0:650
2 log 2

Simplifying the above expression, we get

N = d2:0463e ≈ 3

The order of the filter is calculated as 3.

Step 3: To determine the cut-off frequency
The expression for cut-off frequency is given by

Ωp
Ωc = 1
A
- 10p 2N
10 -1

Substituting Ap = -3 dB, Ωp = 0.650 and N = 3 in the above expression, we get

3 - 16
Ωc = 0:650 1010 - 1

Simplifying the above expression, the value of Ωc is computed as

Ωc = 0:650

Step 4: Transfer function of normalized lowpass filter

The transfer function of normalized lowpass filter for N = 3 is given using the
Butterworth polynomial as

1
H lp ðsÞ =
BðsÞ

The Butterworth polynomial B(s) for N = 3 is (s + 1)(s2 + s + 1). Substituting this

in the above expression, we get
8.4 Butterworth Filter 335

1
H lp ðsÞ =
ð s þ 1 Þ ð s 2 þ s þ 1Þ

Step 5: Converting normalized lowpass filter to the desired lowpass filter using
frequency transformation

H ðsÞ = H lp ðsÞ s = Ωsc

1
H ðsÞ =
ð s þ 1Þ ð s 2 þ s þ 1 Þ s = Ωsc

Substituting the value of Ωc = 0.650 (from Step 3), the above expression is given
by

1
H ðsÞ = 2
s
0:650 þ1 s
0:650 þ 0:650
s
þ1

Simplifying the above expression, we get

0:2751
H ðsÞ =
s3 þ 1:301s2 þ 0:8459s þ 0:2751

Step 6: Converting the analogue filter into an equivalent digital filter

The digital equivalent of the analogue filter using BLT Technique is given by

H ðzÞ = H ðsÞjs = 2 z - 1
T zþ1

Substituting the expression for H(s) from Step 5, we get

0:2751
H ðzÞ =
s3 þ 1:301s2 þ 0:8459s þ 0:2751 -1
s = T2 zzþ1

Substituting T = 1 s, the above expression can be written as

0:2751
H ðzÞ = 3 2
-1 -1 -1
2 zzþ1 þ 1:301 2 zzþ1 þ 0:8459 2 zzþ1 þ 0:2751

Simplifying the above expression, we get

0:0181z3 þ 0:0544z2 þ 0:0544z þ 0:0181

H ðzÞ =
z3 - 1:759z2 þ 1:182z - 0:2778
336 8 Infinite Impulse Response Filter

import numpy as np
#import matplotlib.pyplot as plt
from scipy import signal
import control as ss
# Specifications of Filter
fsam=20000 # Sampling frequency
fp=2000 # Pass band frequency
fs=5000 # Stop abnd frequency
Ap, As, Td=3,20, 1
wp=2*np.pi*(fp/fsam) # pass band freq in radian per sample
ws=2*np.pi*(fs/fsam) # Stop band freq in radian per sample
# prewarping process
omega_p=(2/Td)*np.tan(wp/2)
omega_s=(2/Td)*np.tan(ws/2)
#Computation of order and normalized cut-off frequency
N, omega_c=signal.buttord(omega_p,omega_s,Ap,As,analog=True)
print('Order of the Filter N =', N)
print('Cut-off frequency= {:.4f} rad/s '. format(omega_c))
# Computation of H(s)
b, a=signal.butter(N,omega_c,'low', analog=True)
s1 = ss.tf(b, a)
print('Transfer function H(s)=',s1)
bz, az=signal.bilinear(b, a, Td)
z1 = ss.tf(bz,az,Td)
print('Transfer function H(z)=',z1)

Fig. 8.33 Python code to verify the result of Experiment 8.16

Inference
1. The theoretical result is verified with the python code, which is shown in
Fig. 8.33. The built-in function signal.buttord helps to obtain the order and
cut-off frequency of the analogue filter.
2. The numerator and denominator coefficients of the analogue filter are computed
using signal.butter built-in function.
3. The analogue filter coefficients are converted into digital filter coefficients using
bilinear transformation, the built-in function used for bilinear transformation is
signal.bilinear.
4. After executing the python code, which is given in Fig. 8.33, the simulation result
is shown in Fig. 8.34. From this figure, it is possible to observe that the simulation
result is on par with the theoretical result.

Task
1. Reduce the gap between passband and stopband cut-off frequencies and observe
the order of the filter.
8.4 Butterworth Filter 337

Order of the Filter N = 3

Cut-off frequency= 0.6504 rad/s
Transfer function H(s)=
0.2751
-----------------------------------
s^3 + 1.301 s^2 + 0.8459 s + 0.2751
Transfer function H(z)=
0.01813 z^3 + 0.0544 z^2 + 0.0544 z + 0.01813
---------------------------------------------
z^3 - 1.759 z^2 + 1.182 z - 0.2778

Fig. 8.34 Simulation result of the python code is given in Fig. 8.33

Experiment 8.17 Design of Butterworth Filter for Given Specifications

The objective of this experiment is to obtain the Butterworth lowpass filter coeffi-
cients, transfer function and frequency response for the following filter specifica-
tions: (1) order of the filter (N ) = 2, sampling frequency (Fs) = 8 kHz and (2) cut-off
frequency ( fc) = 2 kHz. Use BLT method for transformation. The python code for
this experiment is given in Fig. 8.35, and its corresponding output is shown in
Figs. 8.36 and 8.37. Figure 8.37 gives the magnitude and phase responses of the
Butterworth filter.
Inference
1. From the magnitude response, it is possible to observe that the filter behaves like a
lowpass filter.
2. The gain drops beyond 2000 Hz, which is in agreement with the filter
specification.
3. It is also possible to observe that the phase response is not linear.
Experiment 8.18 Frequency Transformation
The objective of this experiment is to convert the normalized Butterworth analogue
filter into desired Butterworth analogue filter using frequency transformation. Using
frequency transformation technique, the normalized lowpass filter is converted into
desired lowpass, highpass, bandpass and band reject filter. The filter considered for
frequency transformation is a second-order normalized lowpass filter. This filter is
converted to a desired lowpass and highpass filter for a cut-off frequency of 2 rad/s.
For the bandpass and band reject filter, the lower and upper cut-off frequency is
considered as 3 and 5 rad/s, respectively. The python code for this experiment is
given in Fig. 8.38. The corresponding result is shown in Figs. 8.39 and 8.40.
338 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
# Specifications of Filter
fsam=8000 # Sampling frequency in Hz
fc=2000 # cut off frequency in Hz
N, T = 2, 1/fsam
wc1=2*np.pi*fc # Cut off frequency in rad/sec
print('Cut-off frequency (in rad/sec)=', wc1)
wc = (2/T)*np.tan(wc1*T/2) # Prewarp the analog frequency
# Design analog Butterworth filter using signal.butter function
b, a = signal.butter(N, wc, 'low', analog='True')
s1 = ss.tf(b,a)
print('Transfer function H(s)=',s1)
# Perform bilinear Transformation
bz, az = signal.bilinear(b, a, fs=fsam)
# Print numerator and denomerator coefficients of the filter
print('Numerator Coefficients:', bz),print('Denominator Coefficients:', az)
z1 = ss.tf(bz,az,T)
print('Transfer function H(z)=',z1)
# Compute frequency response of the filter using signal.freqz function
wz, hz = signal.freqz(bz, az, 512)
fig = plt.figure(figsize=(10, 8))
Mag = 20*np.log10(abs(hz)) # Calculate Magnitude in dB
Freq = wz*fsam/(2*np.pi) # Calculate frequency in Hz
# Plot Magnitude response
sub1 = plt.subplot(2, 1, 1),sub1.plot(Freq, Mag, 'r', linewidth=2),sub1.axis([1, fsam/2, -60, 5])
sub1.set_title('Magnitude Response', fontsize=15),
sub1.set_xlabel('Frequency [Hz]', fontsize=15),sub1.set_ylabel('Magnitude [dB]', fontsize=15)
sub1.grid()
# Plot phase angle
sub2 = plt.subplot(2, 1, 2)
Phase = np.unwrap(np.angle(hz))*180/np.pi # Calculate phase angle in degree from hz
sub2.plot(Freq, Phase, 'g', linewidth=2),sub2.set_ylabel('Phase (degree)', fontsize=15)
sub2.set_xlabel(r'Frequency (Hz)', fontsize=15),sub2.set_title(r'Phase response', fontsize=15)
sub2.grid(),plt.subplots_adjust(hspace=0.5),fig.tight_layout(),plt.show()

Fig. 8.35 Python code for Experiment 8.17

From Fig. 8.39, it is possible to observe that the transfer function of normalized
Butterworth lowpass filter is on par with the theoretical transfer function, which is
given by

1
H ðsÞ =
s2 þ 1:414s þ 1

Here, we have considered the order of the filter is (N = 2), the cut-off frequency of
lowpass and highpass filter is chosen as 2. The highest degree of the denominator
polynomial function of lowpass and highpass filters is 2, whereas the highest degree
of the denominator polynomial function of bandpass and band reject filters is 4. The
8.4 Butterworth Filter 339

Cut-off frequency (in rad/sec)= 12566.370614359172

Transfer function H(s)=
2.56e+08
----------------------------
s^2 + 2.263e+04 s + 2.56e+08
Numerator Coefficients: [0.29289322 0.58578644 0.29289322]
Denominator Coefficients: [ 1.00000000e+00 -2.04583550e-16 1.71572875e-01]
Transfer function H(z)=
0.2929 z^2 + 0.5858 z + 0.2929
------------------------------
z^2 - 2.046e-16 z + 0.1716
dt = 0.000125

Fig. 8.36 Simulation result of python code is given in Fig. 8.35

Fig. 8.37 Magnitude and phase responses

340 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
N, omega_c, wn = 2, 2, [3,5] # Order, cutoff freq. of (LPF and HPF),(BPF and BSF),
omega1=np.linspace(0, 10, 100)
b, a = signal.butter(N, 1, 'low', analog=True)
b1, a1 = signal.butter(N, omega_c, 'low', analog=True)
b2, a2 = signal.butter(N, omega_c, 'high', analog=True)
b3, a3 = signal.butter(N, wn, 'bandpass', analog=True)
b4, a4 = signal.butter(N, wn, 'bandstop', analog=True)
s1 = ss.tf(b, a)
print('Normalized Butterworth filter H(s)=', s1)
mag, phase, omega1=ss.freqresp(s1, omega1)
plt.figure(1),plt.plot(omega1,np.abs(mag)),plt.xlabel('$\Omega$-->'),
plt.ylabel('$|H(j\Omega)|$'),plt.title('Magnitude response of Normalized LPF')
s2 = ss.tf(b1, a1)
print('Desired Butterworth LPF H1(s)=', s2)
mag1, phase1, omega1=ss.freqresp(s2, omega1)
plt.figure(2),plt.subplot(2,2,1),plt.plot(omega1,np.abs(mag1))
plt.xlabel('$\Omega$-->'),plt.ylabel('$|H_1(j\Omega)|$'),plt.title('Desired LPF')
s3 = ss.tf(b2, a2)
print('Desired Butterworth HPF H2(s)=', s3)
mag2, phase2, omega1=ss.freqresp(s3, omega1)
plt.subplot(2,2,2),plt.plot(omega1,np.abs(mag2)),plt.xlabel('$\Omega$-->'),
plt.ylabel('$|H_2(j\Omega)|$'),plt.title('Desired HPF')
s4 = ss.tf(b3, a3)
print('Desired Butterworth BPF H3(s)=', s4)
mag3, phase3, omega1=ss.freqresp(s4, omega1)
plt.subplot(2,2,3),plt.plot(omega1,np.abs(mag3)),plt.xlabel('$\Omega$-->'),
plt.ylabel('$|H_3(j\Omega)|$'),plt.title('Desired BPF')
s5 = ss.tf(b4, a4)
print('Desired Butterworth BSF H4(s)=', s5)
mag4, phase4, omega1=ss.freqresp(s5, omega1)
plt.subplot(2,2,4),plt.plot(omega1,np.abs(mag4)),plt.xlabel('$\Omega$-->'),
plt.ylabel('$|H_4(j\Omega)|$'),plt.title('Desired BSF')
plt.tight layout()

Fig. 8.38 Python code to convert normalized Butterworth filter to desired filters

bandpass and band reject filter contains rising and falling transition widths, each
transition width takes second-order roll-off rate; hence, the order of the filter is 4.
Inferences
1. The magnitude responses of the normalized lowpass and desired lowpass,
highpass, bandpass and bandstop filters are shown in Fig. 8.40.
2. From Fig. 8.40a, it is clearly understood that the cut-off frequency is 2 Hz.
3. From Fig. 8.40b, the cut-off frequency of lowpass and highpass filters is 2 Hz,
while bandpass and band reject filters are 3 and 5 Hz.
8.4 Butterworth Filter 341

Normalized Butterworth filter H(s)=

1
-----------------
s^2 + 1.414 s + 1

Desired Butterworth LPF H1(s)=

4
-----------------
s^2 + 2.828 s + 4

Desired Butterworth HPF H2(s)=

s^2
-----------------
s^2 + 2.828 s + 4

Desired Butterworth BPF H3(s)=

4 s^2
----------------------------------------
s^4 + 2.828 s^3 + 34 s^2 + 42.43 s + 225

Desired Butterworth BSF H4(s)=

s^4 + 30 s^2 + 225
----------------------------------------
s^4 + 2.828 s^3 + 34 s^2 + 42.43 s + 225

Fig. 8.39 Transfer functions of normalized and desired Butterworth filters

Magnitude response of Normalized LPF Desired LPF Desired HPF

1.0 1.0
1.0
|H1(j:)|

|H2(j:)|

0.5 0.5
0.8
0.0 0.0
0.6 0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0
|H(j:)|

:--> :-->
0.4 Desired BPF Desired BSF
1.0 1.0
|H3(j:)|

|H4(j:)|

0.2
0.5 0.5

0.0 0.0 0.0

0 2 4 6 8 10 0.0 2.5 5.0 7.5 10.0 0.0 2.5 5.0 7.5 10.0
:--> :--> :-->

(a) Normalized low pass filter (b) Desired filters

Fig. 8.40 Magnitude responses of normalized and desired Butterworth filters. (a) Normalized
lowpass filter. (b) Desired filters

4. The conclusion that can be made from this experiment is that the desired filter can
be obtained from the normalized lowpass filter using analogue frequency
transformation.
342 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
# Specifications of Filter
fsam=8000 # Sampling frequency in Hz
fc1,fc2=[1500,2500],[1000,3000] # cut off frequency in Hz
T = 1/fsam
wcp1=2*np.pi*fc1[0] # Pass band Cut off frequency in rad/sec
wcp2=2*np.pi*fc1[1] # Pass band Cut off frequency in rad/sec
wcs1=2*np.pi*fc2[0] # Stop band Cut off frequency in rad/sec
wcs2=2*np.pi*fc2[1] # Stop band Cut off frequency in rad/sec
pwcp1 = (2/T)*np.tan(wcp1*T/2) # Prewarp the analog frequency
pwcp2 = (2/T)*np.tan(wcp2*T/2) # Prewarp the analog frequency
pwcs1 = (2/T)*np.tan(wcs1*T/2) # Prewarp the analog frequency
pwcs2 = (2/T)*np.tan(wcs2*T/2) # Prewarp the analog frequency
N,wn=signal.buttord([pwcp1,pwcp2],[pwcs1,pwcs2],1,30,True)
print('Order of the filter (N) = ',N)
# Design analog Butterworth filter using signal.butter function
b, a = signal.butter(N, wn, 'bandpass', analog='True')
s1 = ss.tf(b,a)
print('Transfer function H(s)=',s1)
# Perform bilinear Transformation
bz, az = signal.bilinear(b, a, fs=fsam)
z1 = ss.tf(bz,az,T)
print('Transfer function H(z)=',z1)
# Compute frequency response of the filter using signal.freqz function
wz, hz = signal.freqz(bz, az, 512)
fig = plt.figure(figsize=(10, 8))
Mag = 10*np.log10(abs(hz)) # Calculate Magnitude in dB
Freq = wz*fsam/(2*np.pi) # Calculate frequency in Hz
# Plot Magnitude response
sub1 = plt.subplot(2, 1, 1)
sub1.plot(Freq, Mag, 'r', linewidth=2),sub1.axis([1, fsam/2, -60, 5])
sub1.set_title('Magnitude Response', fontsize=15),
sub1.set_xlabel('Frequency [Hz]', fontsize=15),sub1.set_ylabel('Magnitude [dB]', fontsize=15)
sub1.grid()
# Plot phase angle
sub2 = plt.subplot(2, 1, 2)
Phase = np.unwrap(np.angle(hz))*180/np.pi # Calculate phase angle in degree from hz
sub2.plot(Freq, Phase, 'g', linewidth=2),sub2.set_ylabel('Phase (degree)', fontsize=15)
sub2.set_xlabel(r'Frequency (Hz)', fontsize=15),sub2.set_title(r'Phase response', fontsize=15)
sub2.grid(),plt.subplots_adjust(hspace=0.5),fig.tight_layout(),plt.show()

Fig. 8.41 Python code for Butterworth bandpass filter design

Experiment 8.19 Design of Butterworth Bandpass Filter

The objective of this experiment is to write a python code to design a Butterworth
digital bandpass filter for the following specifications: (1) Passband frequencies are
1500 Hz and 2500 Hz. (2) Stopband frequencies are 1000 Hz and 3000 Hz.
(3) Sampling frequency (Fs) = 8 kHz. (4) Passband ripple is 1 dB and stopband
attenuation is 30 dB. Use BLT method for transformation.
8.4 Butterworth Filter 343

Order of the filter (N) = 5

Transfer function H(s)=
8.041e+20 s^5
-------------------------------------------------------------------------------------------------------------------
s^10 + 4.91e+04 s^9 + 2.485e+09 s^8 + 6.857e+13 s^7 + 1.753e+18 s^6 + 2.947e+22
s^5 + 4.487e+26 s^4 + 4.494e+30 s^3 + 4.17e+34 s^2 + 2.109e+38 s + 1.1e+42

Transfer function H(z)=

0.005376 z^10 - 0.02688 z^8 - 7.894e-18 z^7 + 0.05376 z^6 - 3.552e-17 z^5 -
0.05376 z^4 - 7.894e-18 z^3 + 0.02688 z^2 - 0.005376
-------------------------------------------------------------------------------------------------------------------
z^10 + 5.999e-16 z^9 + 2.156 z^8 + 3.663e-15 z^7 + 2.281 z^6 + 1.516e-15 z^5 + 1.297
z^4 + 1.516e-15 z^3 + 0.3952 z^2 + 2.684e-16 z + 0.05031
dt = 0.000125

Fig. 8.42 Simulation result of python code is given in Fig. 8.41

Fig. 8.43 Magnitude and phase responses

344 8 Infinite Impulse Response Filter

Table 8.2 Built-in functions used in the program

S. Built-in
No. function Library Use
1 tan Numpy To obtain the tan value
2 buttord Scipy To obtain the order and cut-off frequency of the Butterworth filter
3 butter Scipy To obtain the coefficients of Butterworth filter
4 bilinear Scipy To convert analogue transfer function into an equivalent digital
transfer function using the bilinear transformation mapping
5 freqz Scipy To obtain the frequency response of the digital filter

The python code for this experiment is given in Fig. 8.41, and its corresponding
output is shown in Figs. 8.42 and 8.43. Figure 8.43 gives the magnitude and phase
response of the Butterworth digital bandpass filter. The libraries used in this program
are (1) numpy, (2) matplotlib, (3) scipy and (4) control. The built-in functions used in
this program are summarized in Table 8.2.
Inference
From the magnitude response shown in Fig. 8.43, it is possible to observe that the
passband of the digital filter is 1500–2500 Hz. This is in agreement with the
specification of the filter. The phase response of the filter is non-linear.
Task
1. Change the passband cut-off frequencies of the bandpass filter and see the
changes in the magnitude and phase responses.

8.5 Chebyshev Filter

The squared magnitude response of Chebyshev filter is given by

2 1
H ejΩ = ð8:14Þ
1þ ε2 T 2N ðΩÞ

The parameter ε sets the ripple amplitude. Chebyshev filters can be classified into
two types, namely, Type I Chebyshev filters and Type II Chebyshev filters. Type I
Chebyshev filter exhibits ripple in passband, whereas Type II Chebyshev filter
exhibits ripple in stopband. Chebyshev polynomial TN(x) for different order is
given in Table 8.3.
Experiment 8.20 Plotting Chebyshev Polynomial for Different Order
This experiment tries to plot the Chebyshev polynomial functions of different order
using python. The python code is used here to plot the Chebyshev polynomial
functions of different order and the corresponding output is shown in Figs. 8.44
8.5 Chebyshev Filter 345

Table 8.3 Chebyshev poly- Order (N ) Polynomial function

nomial function
0 T0(x) = 1
1 T1(x) = x
2 T2(x) = 2x2 - 1
3 T3(x) = 4x3 - 3x
4 T4(x) = 8x4 - 8x2 + 1
5 T5(x) = 16x5 - 20x3 + 5x
6 T6(x) = 32x6 - 48x4 + 18x2 - 1
7 T7(x) = 64x7 - 112x5 + 56x3 - 7x
8 T8(x) = 128x8 - 256x6 + 160x4 - 32x2 + 1

import numpy as np
import matplotlib.pyplot as plt
x=np.linspace(-1,1,50)
T0=np.ones(len(x)) # zeroth degree polynomial
T1=x # First degree polynomial
T2=2*x*T1-T0 # Second degree polynomial
T3=2*x*T2-T1 # Third degree polynomial
T4=2*x*T3-T2 # Fourth degree polynomial
T5=2*x*T4-T3 # Fifth degree polynomial
plt.subplot(3,2,1),plt.plot(x,T0,'k--',linewidth=3.5), plt.xlabel('x'), plt.ylabel('T_0(x)'),
plt.grid(),plt.subplot(3,2,2),plt.plot(x,T1,'k--',linewidth=3.5), plt.xlabel('x'),
plt.ylabel('T_1(x)'),plt.grid(),plt.subplot(3,2,3),plt.plot(x,T2,'k--',linewidth=3.5), plt.xlabel('x'),
plt.ylabel('T_2(x)'),plt.grid(),plt.subplot(3,2,4),plt.plot(x,T3,'k--',linewidth=3.5),
plt.xlabel('x'),plt.ylabel('T_3(x)'),plt.grid(), plt.subplot(3,2,5),plt.plot(x,T4,'k--',linewidth=3.5),
plt.xlabel('x'),plt.ylabel('T_4(x)'),plt.grid(),plt.subplot(3,2,6),plt.plot(x,T5,'k--',linewidth=3.5),
plt.xlabel('x'),plt.ylabel('T_5(x)'),plt.grid(),plt.tight_layout()

Fig. 8.44 Python code to plot the Chebyshev polynomial

and 8.45, respectively. From Fig. 8.45, it is possible to infer that the Chebyshev
polynomial function is created based on the recursive formula, which is given by

T N ðxÞ = 2xT N - 1 ðxÞ - T N - 2 ðxÞ ð8:15Þ

Inference
Figure 8.45 shows that whenever N = 1, 3, 5, etc., the graph passes through the
origin. For N = 0, 2, 4, etc., the graph does not pass through the origin.
Task
1. Write a python code to plot the Chebyshev polynomial function of order 8. Men-
tion the number of zero crossings that exist in it.
346 8 Infinite Impulse Response Filter

1.05 1
T_0(x)

T_1(x)
1.00 0

0.95 –1
–1.0 –0.5 0.0 0.5 1.0 –1.0 –0.5 0.0 0.5 1.0
x x
1 1
T_2(x)

T_3(x)
0 0

–1 –1
–1.0 –0.5 0.0 0.5 1.0 –1.0 –0.5 0.0 0.5 1.0
x x

1 1
T_4(x)

0 T_5(x) 0

–1 –1
–1.0 –0.5 0.0 0.5 1.0 –1.0 –0.5 0.0 0.5 1.0
x x

Fig. 8.45 Chebyshev polynomials of different order

Experiment 8.21 Design of Type I Chebyshev Filter

This experiment discusses the design of a Type I Chebyshev filter using bilinear
transformation technique (BLT) that has a passband of 0 to -2 dB cut-off frequency
of 3 kHz and attenuation of at least 20 dB for frequencies greater than 5 kHz.
Assume the sampling frequency to be 15 kHz. Assume T = 1 s.
Given data
1. Sampling frequency fsamp = 15 kHz.
2. Gain in pass band Ap = -2 dB.
3. Stop band attenuation As = -20 dB.
4. Pass band cut off frequency fp = 3 kHz.
5. Stop band cut off frequency fs = 5 kHz.
Step 1: Prewarping
First, it is necessary to compute the ωp and ωs from the frequency fp and fs, which
are given below

fp
ωp = 2π
f samp

Substituting the values of fp and fsamp in the above expression, we get

8.5 Chebyshev Filter 347

3
ωp = 2π = 0:4π
15

Now the expression for ωs is given by

fs
ωs = 2π
f samp

Substituting the values of fs and fsamp in the above expression, we get

5
ωs = 2π = 0:667π
15

The corresponding analogue frequencies Ωp and Ωs are obtained using

prewarping by using BLT relation as

2 ωp
Ωp = tan
T 2

Substituting the value of ωp = 0.4π and T = 1 in the above expression, we get

2 0:4π
Ωp = tan = 1:4531
1 2

Similarly, the value of Ωs is computed as

2 0:667π
Ωs = tan = 3:4633
1 2

Step 2: To determine the passband ripple factor (ε)

The passband ripple factor is calculated as

10ð 10 Þ - 1
Ap
ε=

ε = 10ð10Þ - 1 = 10ð0:2Þ - 1
2

p p
ε = 1:5849 - 1 = 0:5849 = 0:7648

Step 3: To determine the order of the filter

The expression for the order of Chebyshev filter is given by
348 8 Infinite Impulse Response Filter

cosh - 1 10 - 10 - 1
As
1
ε2
N≥
Ωs
cosh - 1 Ωp

Substituting the value of ε = 0.7648, As = -20 dB, Ωp = 1.4531 and

Ωs = 3.4633, in the above expression, we get

- 20
cosh - 1 1
ð0:7648Þ2
10 - 10 -1
N≥
cosh - 1 3:4633
1:4531

The above expression is simplified as

cosh - 1 1
ð0:7648Þ2
102 - 1
N≥
cosh - 1 ð2:3834Þ

The above expression is further simplified as

cosh - 1 1
ð0:7648Þ2
ð100 - 1Þ
N≥
1:5144

The further simplification of the above expression, we get

cosh - 1 99
0:5849
N≥
1:5144

The above expression is simplified as

p
cosh - 1 169:2597
N≥
1:5144

The simplified version of the above equation, we get

3:2574
N≥ = 2:1510 = 3
1:5144

From the above result, the order of the filter is computed as three (N = 3).
Step 4: Computation of the left half of the S plane poles
From the order of filter 3, 6 poles will be there in this filter. Only 3 poles will be
calculated for further process. That is, those poles must be lying left half of S-plane.
8.5 Chebyshev Filter 349

The computation of the poles are given by

ð2k - 1Þπ 1 1
sk = σ k þ jΩk = - sin sinh sinh - 1
2N N ε
ð8:16Þ
ð2k - 1Þπ 1 1
þj cos cosh sinh - 1
2N N ε

In the above equation substituting k = 1, 2, 3, we may get 3 numbers of poles.

For k = 1, the above equation can be written as

ð2:1 - 1Þπ 1 1
s1 = σ 1 þ jΩ1 = - sin sinh sinh - 1
2×3 3 0:7648
ð2:1 - 1Þπ 1 1
þj cos cosh sinh - 1
2×3 3 0:7648

The above equation can be simplified as

ð2:1 - 1Þπ 1 ð2:1 - 1Þπ 1

s1 = - sin sinh ð1:0831Þ þ j cos cosh ð1:0831Þ
2×3 3 2×3 3

The above equation further simplified as

π π
s1 = - sin sinhð0:3610Þ þ j cos coshð0:3610Þ
2×3 2×3

Simplifying the above equation, we get

π π
s1 = - sin ð0:3689Þ þ j cos ð1:0659Þ
6 6

We know that sin π6 = 0:5 and cos π

2 = 0:8660, substituting this results in the
above equation, we get

s1 = - ð0:5Þð0:3689Þ þ jð0:8660Þð1:0659Þ
s1 = - 0:1845 þ j0:9231

Substituting k = 2, in Eq. (8.16), we get

ð2:2 - 1Þπ 1 1
s2 = σ 2 þ jΩ2 = - sin sinh sinh - 1
2×3 3 0:7648
ð2:2 - 1Þπ 1 1
þj cos cosh sinh - 1
2×3 3 0:7648

The above equation can be simplified as

350 8 Infinite Impulse Response Filter

ð4 - 1Þπ 1 ð4 - 1Þπ 1
s2 = - sin sinh ð1:0831Þ þ j cos cosh ð1:0831Þ
2×3 3 2×3 3

The above equation further simplified as

3π 3π
s2 = - sin sinhð0:3610Þ þ j cos coshð0:3610Þ
6 6

Simplifying the above equation, we get

π π
s2 = - sin ð0:3689Þ þ j cos ð1:0659Þ
2 2
π π
We know that sin 2 = 1 and cos 2 = 0, substituting this results in the above
equation, we get

s2 = - 0:3689

Substituting k = 3, in Eq. (8.16), we get

ð2:3 - 1Þπ 1 1
s3 = σ 3 þ jΩ3 = - sin sinh sinh - 1
2×3 3 0:7648
ð2:3 - 1Þπ 1 1
þj cos cosh sinh - 1
2×3 3 0:7648

The above equation can be simplified as

ð6 - 1Þπ 1 ð6 - 1Þπ 1
s3 = - sin sinh ð1:0831Þ þ j cos cosh ð1:0831Þ
2×3 3 2×3 3

The above equation further simplified as

5π 5π
s3 = - sin sinhð0:3610Þ þ j cos coshð0:3610Þ
6 6

Simplifying the above equation, we get

5π 5π
s3 = - sin ð0:3689Þ þ j cos ð1:0659Þ
6 6

The simplification of the above result, we get

8.5 Chebyshev Filter 351

s3 = - 0:1845 - j0:9231

The results of this step are s1 = - 0.1845 + j0.9231, s2 = - 0.3689 and s3 = -

0.1845 - j0.9231.
Step 5: Calculate the normalized frequency transfer function (i.e. Ωp = 1)
The normalized frequency transfer function is obtained by

ð- 1ÞN s1 × s2 × ⋯ × sN
H ð sÞ = K
ðs - s1 Þðs - s2 Þ⋯ðs - sN Þ

where

1
p for N is Even
K= 1 þ ε2
1 for N is odd

In this example, the order of the filter is 3; it shows that odd, hence the gain K = 1.
The transfer function is formed as

ð- 1Þ3 × s1 × s2 × s3
H ðsÞ = K
ðs - s1 Þðs - s2 Þðs - s3 Þ

Substituting the values of s1 = - 0.1845 + j0.9231, s2 = - 0.3689 and s3 = -

0.1845 - j0.9231 in the above equation, we get

ð- 1Þ × ð- 0:1845 þ j0:9231Þ × ð- 0:3689Þ × ð- 0:1845 - j0:9231Þ

H ðsÞ = K
ðs - ð- 0:1845 þ j0:9231ÞÞðs - ð- 0:3689ÞÞðs - ð- 0:1845 - j0:9231ÞÞ

The above equation can be simplified as

0:3269
H ðsÞ =
ðs þ 0:1845 - j0:9231ÞÞðs þ 0:3689ÞÞðs þ 0:1845 þ j0:9231ÞÞ

The above equation can be further simplified as

0:3269
H ðsÞ =
ðs3 þ 0:7378s2 þ 1:0222s þ 0:3269Þ

Step 6: Calculate the transfer function of the desired frequency

The desired passband cut-off frequency of the filter is Ωp = 1.4531. The transfer
function is given by
352 8 Infinite Impulse Response Filter

s 0:3269
Ha =
1:4531 s 3
þ 0:7378 s 2
þ 1:0222 s
þ 0:3269
1:4531 1:4531 1:4531

The above result can be further simplified as

s 1:003
Ha = 3
1:4531 ðs þ 1:072s2 þ 2:158s þ 1:003Þ

Step 7: Converting the analogue filter into an equivalent digital filter

The digital equivalent of the analogue filter using bilinear transformation tech-
nique is given by

H ðzÞ = H a ðsÞjs = 2 1 - z - 1
T 1þz - 1

Substituting the expression for H(s) from Step 5, we get

1:003
H ðzÞ =
ðs3 þ 1:072s2 þ 2:158s þ 1:003Þ -1
s = 21 - z- 1
1þz

The above expression can be written as

1:003
H ðzÞ = 3 2
-z -1 -z -1 -z -1
2 11þz -1 þ 1:072 2 11þz -1 þ 2:158 2 11þz -1 þ 1:003

Simplifying the above expression, we get

0:05696z3 þ 0:1709z2 þ 0:1709z þ 0:05696

H ðzÞ =
z3 - 1:191z2 þ 1:045z - 0:399

The python code for this experiment is given in Fig. 8.45. The built-in function
signal.cheb1ord is used here to compute the order and cut-off frequency of the filter.
The built-in function signal.cheby1 is used here to obtain the numerator and denom-
inator coefficients of filter, ss.tf is used for the computation of transfer function and
signal.bilinear helps to convert analogue filter coefficients into digital filter coeffi-
cients using BLT approach.
Inferences
1. After executing the python code given in Fig. 8.46, the simulation results are
shown in Figs. 8.47 and 8.48.
2. The order of the filter is calculated as (N = 3); the cut-off frequency is obtained as
1.4531 rad/s, which is equivalent to passband cut-off frequency, and the analogue
8.5 Chebyshev Filter 353

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
# Specifications of Filter
fsam=15000 # Sampling frequency
fp=3000 # Passband frequency
fs=5000 # Stopband frequency
Ap, As, Td=2,20, 1
wp=2*np.pi*(fp/fsam) # passband freq in radian per sample
ws=2*np.pi*(fs/fsam) # Stopband freq in radian per sample
# prewarping process
omega_p=(2/Td)*np.tan(wp/2)
omega_s=(2/Td)*np.tan(ws/2)
N, omega_c=signal.cheb1ord(omega_p,omega_s,Ap,As,analog=True)
print('Order of the Filter N =', N)
print('Cut-off frequency= {:.4f} rad/s'. format(omega_c))
# Computation of H(s)
b_s, a_s=signal.cheby1(N,Ap,omega_c,'low', analog=True)
s1 = ss.tf(b_s, a_s)
print('Transfer function H(s)=',s1)
bz, az=signal.bilinear(b_s, a_s, Td)
z1 = ss.tf(bz,az,Td)
print('Transfer function H(z)=',z1)
ws, hs = signal.freqz(bz, az) # Calculate Magnitude from hz in dB
Mag = 20*np.log10(abs(hs)) # Calculate phase angle in degree from hz
Phase = np.unwrap(np.arctan2(np.imag(hs), np.real(hs)))*(180/np.pi)
# Calculate frequency in Hz from wz
Freq = ws*fsam/(2*np.pi)
# Plot filter magnitude and phase responses using subplot.
fig = plt.figure(figsize=(10, 6))# Plot Magnitude response
sub1 = plt.subplot(2, 1, 1)
sub1.plot(Freq, Mag, 'r', linewidth=2)
sub1.axis([1, fsam/2, -100, 5])
sub1.set_title('Magnitude Response', fontsize=15)
sub1.set_xlabel('Frequency [Hz]', fontsize=15)
sub1.set_ylabel('Magnitude [dB]', fontsize=15)
sub1.grid()
# Plot phase angle
sub2 = plt.subplot(2, 1, 2)
sub2.plot(Freq, Phase, 'g', linewidth=2),sub2.set_ylabel('Phase (degree)', fontsize=15)
sub2.set_xlabel(r'Frequency (Hz)', fontsize=15)
sub2.set_title(r'Phase response', fontsize=15)
sub2.grid(),plt.subplots_adjust(hspace=0.5),fig.tight_layout(),plt.show()

Fig. 8.46 Python code to get transfer function of Chebyshev lowpass filter
354 8 Infinite Impulse Response Filter

Order of the Filter N = 3

Cut-off frequency= 1.4531 rad/s
Transfer function H(s)=
1.003
---------------------------------
s^3 + 1.072 s^2 + 2.158 s + 1.003
Transfer function H(z)=
0.05696 z^3 + 0.1709 z^2 + 0.1709 z + 0.05696
---------------------------------------------
z^3 - 1.191 z^2 + 1.045 z - 0.399
dt = 1

Fig. 8.47 Simulation result

Fig. 8.48 Magnitude and phase response

and digital filters transfer functions are the same as the theoretical result, which is
shown in Fig. 8.47.
3. The frequency response of the Chebyshev Type I filter is shown in Fig. 8.48. This
figure clearly shows that the passband frequency is up to 3 kHz.

Experiment 8.22 Design of Chebyshev Type I Digital Highpass Filter

This experiment deals with the design of a Chebyshev Type I digital highpass filter
for the following specifications: (1) Passband frequency is 2500 Hz. (2) Stopband
frequency is 1500 Hz. (2) Sampling frequency (Fs) = 8 kHz. (4) Passband ripple is
8.5 Chebyshev Filter 355

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
# Specifications of Filter
fsam=8000 # Sampling frequency in Hz
fc1=1500 # Stop band cut-off frequency in Hz
fc2=2500# Pass band cut-off frequency in Hz
Ap,As,T =3,40, 1/fsam
wc1=2*np.pi*fc1 # Stopband Cut off frequency in rad/sec
wc2=2*np.pi*fc2 # Passband Cut off frequency in rad/sec
#print('Cut-off frequency (in rad/sec)=', wc1)
pwc1 = (2/T)*np.tan(wc1*T/2) # Prewarp the analog frequency
pwc2 = (2/T)*np.tan(wc2*T/2) # Prewarp the analog frequency
# Design analog Butterworth filter using signal.butter function
N, Wn = signal.cheb1ord(pwc2, pwc1, Ap, As,analog=True)
print('Order of the filter (N) = ',N)
b, a = signal.cheby1(N, Ap, Wn, 'high',analog=True)
s1 = ss.tf(b,a)
print('Transfer function H(s)=',s1)
# Perform bilinear transformation
bz, az = signal.bilinear(b, a, fs=fsam)
# Print numerator and denominator coefficients of the filter
print('Numerator Coefficients:', bz)
print('Denominator Coefficients:', az)
z1 = ss.tf(bz,az,T)
print('Transfer function H(z)=',z1)
# Compute frequency response of the filter using signal.freqz function
wz, hz = signal.freqz(bz, az, 512)
fig = plt.figure(figsize=(10, 8))
Mag = 20*np.log10(abs(hz)) # Calculate Magnitude in dB
Freq = wz*fsam/(2*np.pi) # Calculate frequency in Hz
# Plot Magnitude response
sub1 = plt.subplot(2, 1, 1)
sub1.plot(Freq, Mag, 'r', linewidth=2),sub1.axis([1, fsam/2, -60, 5])
sub1.set_title('Magnitude Response', fontsize=15),sub1.set_xlabel('Frequency [Hz]', fontsize=15)
sub1.set_ylabel('Magnitude [dB]', fontsize=15),sub1.grid()
# Plot phase angle
sub2 = plt.subplot(2, 1, 2)
# Calculate phase angle in degree from hz
Phase = np.unwrap(np.angle(hz))*180/np.pi
sub2.plot(Freq, Phase, 'g', linewidth=2),sub2.set_ylabel('Phase (degree)', fontsize=15)
sub2.set_xlabel(r'Frequency (Hz)', fontsize=15),sub2.set_title(r'Phase response', fontsize=15)
sub2.grid(),plt.subplots_adjust(hspace=0.5),fig.tight_layout(),plt.show()

Fig. 8.49 Python code for Butterworth bandpass filter design

3 dB and stop attenuation is 40 dB. Use BLT method for transformation. The python
code for this experiment is given in Fig. 8.49, and its corresponding output is shown
in Figs. 8.50 and 8.51. Figure 8.50 gives the magnitude and phase responses of the
Chebyshev Type I digital highpass filter.
Inference
From this experiment, the following inferences can be made:
356 8 Infinite Impulse Response Filter

Order of the filter (N) = 4

Transfer function H(s)=
0.7079 s^4
-------------------------------------------------------------
s^4 + 5.476e+04 s^3 + 3.788e+09 s^2 + 4.512e+13 s + 1.858e+18
Numerator Coefficients: [ 0.01208527 -0.04834108 0.07251162 -0.04834108
0.01208527]
Denominator Coefficients: [1. 2.12648692 2.50060541 1.60804914 0.50706501]
Transfer function H(z)=
0.01209 z^4 - 0.04834 z^3 + 0.07251 z^2 - 0.04834 z + 0.01209
-------------------------------------------------------------
z^4 + 2.126 z^3 + 2.501 z^2 + 1.608 z + 0.5071
dt = 0.000125

Fig. 8.50 Simulation result of python code is given in Fig. 8.49

Magnitude Response
0
Magnitude [dB]

–10
–20
–30
–40
–50
–60
500 1000 1500 2000 2500 3000 3500 4000
Frequency [Hz]
Phase response
0
–50
Phase (degree)

–100
–150
–200
–250
–300
–350
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency [Hz]

Fig. 8.51 Magnitude and phase responses

1. From the magnitude response, it is possible to observe that passband gain is high
after the frequency of 2000 Hz, and there is a ripple in the passband.
2. From the phase response, it is possible to confirm that the curve is not linear;
hence, it cannot provide the linear phase in the output.
8.5 Chebyshev Filter 357

Tasks
1. Write a python code to obtain the magnitude and phase response of Chebyshev
Type I bandpass filter with the passband frequencies of 1500–3000 Hz. Choose
the order of the filter that is 3.
2. Write a python code to obtain the magnitude and phase response of Chebyshev
Type I band reject filter with the stopband frequencies of 1500–3000 Hz. Choose
the order of the filter that is 3.

1.05
|H(j:)|

1.00
0.95
–3 –2 –1 0 1 2 3
:-->
|H1(j:)|

0
–3 –2 –1 0 1 2 3
:-->
1-|H1(j:)|

0
–3 –2 –1 0 1 2 3
:-->

Fig. 8.52 First step procedure

1.0

0.8
1-|H1(j1/:)|

0.6

0.4

0.2

0.0
–3 –2 –1 0 1 2 3
:-->

Fig. 8.53 Second step procedure

358 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
N, rs, omega_c=3, 30, [5]
omega1=np.linspace(0, 10, 100)
b_s, a_s=signal.cheby2(N,rs,omega_c,'low', analog=True)
s1 = ss.tf(b_s, a_s)
print('H(s) = ', s1)
z1=np.roots(b_s)
print('Zeros : ', z1)
mag, phase, omega1=ss.freqresp(s1, omega1)
plt.figure,plt.plot(omega1,np.abs(mag))
plt.xlabel('$\Omega$-->'),plt.ylabel('$|H(j\Omega)|$')
plt.title('Magnitude response of Chebyshev Type II LPF')
plt.tight_layout()

Fig. 8.54 Python code for Chebyshev Type II LPF

8.6 Chebyshev Type II IIR Filter

The Chebyshev Type II IIR filter has an equiripple in stopband and monotonic
response in passband, which is inverse of Chebyshev Type I IIR filter; hence, it is
also called as ‘inverse Chebyshev filter’. Two-step procedures can obtain the
frequency response of this filter. In the first step, subtract the frequency response
of Chebyshev Type I filter (H(ω)) from 1, which is illustrated in Fig. 8.52. In the
second step, convert the ω by ω1 . This result will give the frequency response of
Chebyshev Type II IIR filter, which is shown in Fig. 8.53.
From Fig. 8.53, it is possible to observe that the frequency response has mono-
tonic in the passband and an equiripple in the stop band.
Using the two step procedures, the mathematical expression for the frequency
response of the Chebyshev Type II filter can be written as

1
H ðΩÞ = 1 - ð8:17Þ
1 þ ε2 T 2N ð1=ΩÞ

Simplifying the above expression, we get

=Þ
1Ω
H ðΩÞ = ε2 T 2N ð8:18Þ
1þ ε2 T 2N ð1=ΩÞ

From the above expression, it is possible to understand that the Chebyshev Type
II filter has zeros as the numerator, and all the zeros lie on the imaginary axis.
8.6 Chebyshev Type II IIR Filter 359

Magnitude response of Chebyshev Type II LPF

1.0

0.8

H(s) = 0.6

|H(j:)|
0.4746 s^2 + 15.82
0.4
---------------------------------
s^3 + 4.883 s^2 + 11.81 s + 15.82 0.2

0.0
Zeros : [-0.+5.77350269j 0.-5.77350269j] 0 2 4 6 8 10
:-->

(a) Transfer function and zeros (b) Magnitude response

Fig. 8.55 Simulation result of the python code is given in Fig. 8.54. (a) Transfer function and
zeros. (b) Magnitude response

Experiment 8.23 Design of Chebyshev Type II Lowpass Filter

This experiment discusses the design of Chebyshev Type II lowpass filter using
python with a stopband frequency of 5 Hz. Assume the order of the filter is 3, and
stopband attenuation is 30 dB.
The python code to design a Chebyshev Type II lowpass filter is shown in
Fig. 8.54, and its corresponding simulation result is shown in Fig. 8.55. From
Fig. 8.54, it is possible to infer that the signal.cheby2 built-in command can be
used to obtain the numerator and denominator coefficients. After executing the
python code in Fig. 8.54, the transfer function of Chebyshev Type II lowpass filter
is shown in Fig. 8.55a. This figure confirms that the Chebyshev Type II filter has
zeros, which occur on the imaginary axis. The magnitude response of the Chebyshev
Type II lowpass filter is shown in Fig. 8.55b. This figure confirms the monotonic
response in the passband and equiripple in the stopband.
Inference
From Fig. 8.55b, it is possible to confirm that the ripple exists in the stopband and
monotonic response in the passband.
Task
1. Write a python code to plot the magnitude response of Chebyshev Type II
highpass filter with the lower cut-off frequency of 5 Hz. Assume the order of
the filter is 3 and stopband attenuation is 30 dB.
Experiment 8.24 Design of Chebyshev Type II Digital Bandstop Filter
This experiment deals with the design of a Chebyshev Type II digital bandstop filter
for the following specifications: (1) Stopband frequencies are from 1500 to 2500 Hz
with an attenuation of 40 dB. (2) Passband frequencies are below 1000 Hz and above
3000 Hz with the passband ripple of 3 dB. Sampling frequency (Fs) = 8 kHz. Use
BLT method for transformation. The python code for this experiment is given in
Fig. 8.56, and its corresponding output is shown in Figs. 8.57 and 8.58. Figure 8.58
gives the magnitude and phase responses of the Chebyshev Type II digital bandstop
filter.
360 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
# Specifications of Filter
fsam=8000 # Sampling frequency in Hz
fc1,fc2=[1500,2500],[1000,3000] # cut off frequency in Hz
Ap, As, T = 2, 40, 1/fsam
wcp1=2*np.pi*fc1[0] # Pass band Cut off frequency in rad/sec
wcp2=2*np.pi*fc1[1] # Pass band Cut off frequency in rad/sec
wcs1=2*np.pi*fc2[0] # Stop band Cut off frequency in rad/sec
wcs2=2*np.pi*fc2[1] # Stop band Cut off frequency in rad/sec
pwcs1 = (2/T)*np.tan(wcp1*T/2) # Prewarp the analog frequency
pwcs2 = (2/T)*np.tan(wcp2*T/2) # Prewarp the analog frequency
pwcp1 = (2/T)*np.tan(wcs1*T/2) # Prewarp the analog frequency
pwcp2 = (2/T)*np.tan(wcs2*T/2) # Prewarp the analog frequency
N,wn=signal.cheb2ord([pwcp1,pwcp2],[pwcs1,pwcs2],Ap,As,analog=True)
print('Order of the filter (N) = ',N)
# Design analog Chebyshev Type 2 filter using signal.cheby2 function
b, a = signal.cheby2(N, As, wn, 'bandstop', analog='True')
s1 = ss.tf(b,a)
print('Transfer function H(s)=',s1)
# Perform bilinear transformation
bz, az = signal.bilinear(b, a, fs=fsam)
z1 = ss.tf(bz,az,T)
print('Transfer function H(z)=',z1)
# Compute frequency response of the filter using signal.freqz function
wz, hz = signal.freqz(bz, az, 512)
fig = plt.figure(figsize=(10, 8))
Mag = 10*np.log10(abs(hz)) # Calculate Magnitude in dB
Freq = wz*fsam/(2*np.pi) # Calculate frequency in Hz
# Plot Magnitude response
sub1 = plt.subplot(2, 1, 1)
sub1.plot(Freq, Mag, 'r', linewidth=2),sub1.axis([1, fsam/2, -60, 5])
sub1.set_title('Magnitude Response', fontsize=15),
sub1.set_xlabel('Frequency [Hz]', fontsize=15),sub1.set_ylabel('Magnitude [dB]', fontsize=15)
sub1.grid()
# Plot phase angle
sub2 = plt.subplot(2, 1, 2)
Phase = np.unwrap(np.angle(hz))*180/np.pi # Calculate phase angle in degree from hz
sub2.plot(Freq, Phase, 'g', linewidth=2),sub2.set_ylabel('Phase (degree)', fontsize=15)
sub2.set_xlabel(r'Frequency (Hz)', fontsize=15),sub2.set_title(r'Phase response', fontsize=15)
sub2.grid(),plt.subplots_adjust(hspace=0.5),fig.tight_layout(),plt.show()

Fig. 8.56 Python code for Chebyshev Type II bandstop filter design

Inferences
1. From Fig. 8.58, the magnitude response shows no ripple in the passbands and a
ripple in the stop band.
2. From the phase response, it is possible to confirm that IIR filters do not have
linear phase characteristics; hence, the phase response is not linear.
8.6 Chebyshev Type II IIR Filter 361

Order of the filter (N) = 4

Transfer function H(s)=
s^8 + 1.249e+09 s^6 + 5.145e+17 s^4 + 8.183e+25 s^2 + 4.295e+33
-------------------------------------------------------------------------------------------------------------------
s^8 + 6.844e+04 s^7 + 3.59e+09 s^6 + 1.094e+14 s^5 + 2.338e+18 s^4 + 2.802e+22 s^3 +
2.353e+26 s^2 + 1.148e+30 s + 4.295e+33

Transfer function H(z)=

0.1535 z^8 - 5.028e-11 z^7 + 0.4981 z^6 - 1.309e-10 z^5 + 0.7012 z^4 - 1.309e-10 z^3 +
0.4981 z^2 - 5.028e-11 z + 0.1535
-------------------------------------------------------------------------------------------------------------------
z^8 - 1.944e-10 z^7 + 0.3006 z^6 - 9.283e-11 z^5 + 0.5759 z^4 - 6.545e-11 z^3 + 0.0987
z^2 - 9.75e-12 z + 0.02925
dt = 0.000125

Fig. 8.57 Simulation result of python code is given in Fig. 8.56

Fig. 8.58 Magnitude and phase responses

362 8 Infinite Impulse Response Filter

Task
1. Write a python code to design a Chebyshev Type II digital bandpass filter for the
following specifications: (a) Passband frequencies are from 1500 to 2500 Hz with
an attenuation of 3 dB. (b) Stopband frequencies are below 1000 Hz and above
3000 Hz with the passband ripple of 40 dB. Sampling frequency (Fs) = 8 kHz.
Use BLT method for transformation.

8.7 Elliptic Filter

The elliptic filter has an equiripple in both the passband and stopband. It is also
called as ‘Cauer filter’. The elliptic filter gives minimal error between the desired and
ideal filter response for a given set of error tolerances. The elliptical filter maximizes
the rate of transition between the frequency response passband and stopband, at the
expense of equiripple in both bands. Also, it increases ringing in the step response.
The square magnitude response of ‘nth’ order elliptic filter is given by

1
jH ðjΩÞj2 = ð8:19Þ
1þ ε2 R2n ξ, Ω=Ωc

In the above expression, Rn is the nth order elliptic rational function, ‘ε’ is the
ripple factor and ‘ξ’ is the selectivity factor. The ripple factor specifies the passband
ripple, and a combination of the ripple factor and the selectivity factor specifies the
stopband ripple. As the ripple in the stopband approaches zero, the filter tends to
become Chebyshev Type I filter. As the ripple in the passband approaches zero, the
filter tends to become Chebyshev Type II filter. If both passband and stopband
ripples approach zero, then the filter tends to become a Butterworth filter. Elliptic
filter meets a standard magnitude specification with lower filter order than other filter
approximations.
Experiment 8.25 Design of an Elliptic Lowpass Filter
This experiment tries to design an elliptic lowpass filter with a passband frequency of
5 Hz. Assume the order of the filter is 3, passband ripple is 2 dB and stopband
attenuation is 30 dB. The python code to design an elliptic lowpass filter is shown in
Fig. 8.59, and its corresponding simulation result is shown in Fig. 8.60. From
Fig. 8.59, it is possible to infer that the signal.ellip built-in command is used here
to obtain the numerator and denominator coefficients. After executing the python
code given in Fig. 8.59, the transfer function of elliptic lowpass filter is shown in
Fig. 8.60a. This figure confirms that the elliptic filter has zeros, and those zeros occur
on the imaginary axis. The magnitude response of the elliptic lowpass filter is shown
in Fig. 8.60b. From this figure, it is possible to observe the equiripples in both the
passband and stopband.
8.7 Elliptic Filter 363

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
N, rp,rs, omega_c=3, 2, 20, [5]
omega1=np.linspace(0, 10, 100)
b_s, a_s=signal.ellip(N,rp,rs,omega_c,'low', analog=True)
s1 = ss.tf(b_s, a_s)
print('H(s) = ', s1)
z1=np.roots(b_s)
print('Zeros : ', z1)
mag, phase, omega1=ss.freqresp(s1, omega1)
plt.figure,plt.plot(omega1,np.abs(mag))
plt.xlabel('$\u03A9$-->'),plt.ylabel('$|H({j\u03A9})|$')
plt.title('Magnitude response of Elliptic LPF')
plt.tight_layout()

Fig. 8.59 Python code for elliptic LPF

Magnitude response of Elliptic LPF

1.0

0.8

H(s) = 0.6
|H(j:)|

1.394 s^2 + 60.14

0.4
--------------------------------
s^3 + 3.63 s^2 + 26.75 s + 60.14 0.2

0.0
Zeros : [-0.+6.5683962j 0.-6.5683962j] 0 2 4 6 8 10
:-->

(a) Transfer function and zeros (b) Magnitude response

Fig. 8.60 Simulation result of the python code is given in Fig. 8.59. (a) Transfer function and
zeros. (b) Magnitude response

Inference
From the simulation result of this experiment, it is possible to observe that ripples
present in both the passband and stopband, and also it infers that the width of the
transition band is reduced.
Task
1. Write a python code to design an elliptic highpass filter with a lower cut-off
frequency of 5 Hz. Assume the order of the filter is 3, passband ripple is 2 dB and
stopband attenuation is 30 dB.
364 8 Infinite Impulse Response Filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
import control as ss
# Specifications of Filter
fsam=8000 # Sampling frequency in Hz
fc1,fc2=[1500,2500],[1000,3000] # cut off frequency in Hz
Ap, As, T = 2, 40, 1/fsam
wcp1=2*np.pi*fc1[0] # Pass band Cut off frequency in rad/sec
wcp2=2*np.pi*fc1[1] # Pass band Cut off frequency in rad/sec
wcs1=2*np.pi*fc2[0] # Stop band Cut off frequency in rad/sec
wcs2=2*np.pi*fc2[1] # Stop band Cut off frequency in rad/sec
pwcs1 = (2/T)*np.tan(wcp1*T/2) # Prewarp the analog frequency
pwcs2 = (2/T)*np.tan(wcp2*T/2) # Prewarp the analog frequency
pwcp1 = (2/T)*np.tan(wcs1*T/2) # Prewarp the analog frequency
pwcp2 = (2/T)*np.tan(wcs2*T/2) # Prewarp the analog frequency
N,wn=signal.ellipord([pwcp1,pwcp2],[pwcs1,pwcs2],Ap,As,analog=True)
print('Order of the filter (N) = ',N)
# Design analog Elliptic filter using signal.ellip function
b, a = signal.ellip(N, Ap, As, wn, 'bandstop', analog='True')
s1 = ss.tf(b,a)
print('Transfer function H(s)=',s1)
# Perform bilinear transformation
bz, az = signal.bilinear(b, a, fs=fsam)
z1 = ss.tf(bz,az,T)
print('Transfer function H(z)=',z1)
# Compute frequency response of the filter using signal.freqz function
wz, hz = signal.freqz(bz, az, 512)
fig = plt.figure(figsize=(10, 8))
Mag = 10*np.log10(abs(hz)) # Calculate Magnitude in dB
Freq = wz*fsam/(2*np.pi) # Calculate frequency in Hz
# Plot Magnitude response
sub1 = plt.subplot(2, 1, 1)
sub1.plot(Freq, Mag, 'r', linewidth=2),sub1.axis([1, fsam/2, -60, 5])
sub1.set_title('Magnitude Response', fontsize=15),
sub1.set_xlabel('Frequency [Hz]', fontsize=15),sub1.set_ylabel('Magnitude [dB]', fontsize=15)
sub1.grid()
# Plot phase angle
sub2 = plt.subplot(2, 1, 2)
Phase = np.unwrap(np.angle(hz))*180/np.pi # Calculate phase angle in degree from hz
sub2.plot(Freq, Phase, 'g', linewidth=2),sub2.set_ylabel('Phase (degree)', fontsize=15)
sub2.set_xlabel(r'Frequency (Hz)', fontsize=15),sub2.set_title(r'Phase response', fontsize=15)
sub2.grid(),plt.subplots_adjust(hspace=0.5),fig.tight_layout(),plt.show()

Fig. 8.61 Python code for elliptic bandstop filter design

Experiment 8.26 Design of Elliptic Digital Bandstop Filter

The objective of this experiment is to design an elliptic digital bandstop filter for the
following specifications: (1) Stopband frequencies are from 1500 to 2500 Hz with an
attenuation of 40 dB. (2) Passband frequencies are below 1000 Hz and above
3000 Hz with the passband ripple of 3 dB. Sampling frequency (Fs) = 8 kHz. Use
BLT method for transformation. The python code for this experiment is given in
Fig. 8.61, and its corresponding output is shown in Figs. 8.62 and 8.63. Figure 8.63
8.7 Elliptic Filter 365

Order of the filter (N) = 3

Transfer function H(s)=
s^6 + 9.409e+08 s^4 + 2.409e+17 s^2 + 1.678e+25
---------------------------------------------------------------------------------------------
s^6 + 9.278e+04 s^5 + 2.858e+09 s^4 + 1.391e+14 s^3 + 7.317e+17 s^2 + 6.081e+21 s + 1.678e+25

Transfer function H(z)=

0.1338 z^6 - 3.28e-11 z^5 + 0.3241 z^4 - 5.86e-11 z^3 + 0.3241 z^2 - 3.28e-11 z + 0.1338
----------------------------------------------------------------------------------------
z^6 - 1.122e-10 z^5 - 0.5182 z^4 - 1.985e-11 z^3 + 0.7377 z^2 + 7.843e-12 z - 0.3037
dt = 0.000125

Fig. 8.62 Simulation result of python code is given in Fig. 8.61

Fig. 8.63 Magnitude and phase responses

shows the magnitude and phase response of the elliptic (Cauer) digital bandstop
filter.
Inference
From this experiment, the following observations can be made:
366 8 Infinite Impulse Response Filter

#Magnitude response of analog IIR filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Specification of the filter
N=10 #Order of the filter
wn=100 #Cut-off frequency
rp=5 #Pass band ripple
rs=40 #Stop band ripple
#Step 2: Magnitude response of Butterworth filter
b1,a1=signal.butter(N,wn,'low',analog='true')
w1,H1=signal.freqs(b1,a1)
#Magnitude response of Type-1 Chebyshev filter
b2,a2=signal.cheby1(N,rp,wn,'low',analog='true')
w2,H2=signal.freqs(b2,a2)
#Magnitude response of Type-II Chebyshev filter
b3,a3=signal.cheby2(N,rs,wn,'low',analog='true')
w3,H3=signal.freqs(b3,a3)
#Magnitude response of Elliptic filter
b4,a4=signal.ellip(N,rp,rs,wn,'low',analog='true')
w4,H4=signal.freqs(b4,a4)
#Step 3: Plotting the magnitude responses
plt.subplot(2,2,1),plt.semilogx(w1,20*np.log10(abs(H1)),'k-')
plt.xlabel('Frequency'),plt.ylabel('Magnitude[dB]'),plt.title('Butterworth filter')
plt.subplot(2,2,2),plt.semilogx(w2,20*np.log10(abs(H2)),'k-')
plt.xlabel('Frequency'),plt.ylabel('Magnitude[dB]'),plt.title('Chebyshev (type 1) filter')
plt.subplot(2,2,3),plt.semilogx(w3,20*np.log10(abs(H3)),'k-')
plt.xlabel('Frequency'),plt.ylabel('Magnitude[dB]'),plt.title('Chebyshev (type 2) filter')
plt.subplot(2,2,4),plt.semilogx(w4,20*np.log10(abs(H4)),'k-')
plt.xlabel('Frequency'),plt.ylabel('Magnitude[dB]'),plt.title('Elliptic filter')
plt.tight_layout()

Fig. 8.64 Python code to obtain the magnitude response of four analogue IIR filters

1. The magnitude response is having ripples in the passband as well as stopband.

2. The phase response is nonlinear; hence, IIR filters do not have linear phase
characteristics.
Experiment 8.27 Comparing the Magnitude Response of Butterworth,
Chebyshev (Type I), Chebyshev (Type II) and Elliptic Filters
The objective of this experiment is to compare the magnitude response of analogue
IIR filters, namely, Butterworth, Chebyshev (Type I), Chebyshev (Type II) and
elliptic filters for the following lowpass filter specifications: (1) order of the fil-
ter = 10, (2) cut-off frequency = 100 Hz, (3) passband ripple = 5 dB and
(4) stopband ripple = 40 dB. The python code, which obtains the magnitude
response of the four filters, is shown in Fig. 8.64, and the corresponding output is
shown in Fig. 8.65.
The built-in functions used in the program are summarized in Table 8.4.
8.7 Elliptic Filter 367

Magnitude response
0

–20

–40
Magnitude[dB]

–60

–80
Butterworth
–100
Chebyshev Type I
Chebyshev Type II
–120 Elliptic

100 101 102 103

Frequency

Fig. 8.65 Result of python code shown in Fig. 8.64

Table 8.4 Built-in function used

S. Built-in
No. function Library Purpose of built-in function
1 butter Scipy To obtain the coefficients of analogue/digital Butterworth
filter
2 cheby1 Scipy To obtain the coefficients of analogue/digital Type I
Chebyshev filter
3 cheby2 Scipy To obtain the coefficients of analogue/digital Type II
Chebyshev filter
4 ellip Scipy To obtain the coefficients of analogue/digital elliptic filter

Inferences
From Fig. 8.65, the following inferences can be made:
1. Butterworth filter exhibits maximally flat frequency response without ripples in
passband and stopband.
2. Type I Chebyshev filter exhibits ripples in only passband.
3. Type II Chebyshev filter exhibits ripples in stopband.
4. Elliptic filter exhibits ripples in both passband and stopband.
5. The roll-off rate of elliptic filter is better than Butterworth and Chebyshev filters.
Experiment 8.28 Comparing the Order of Butterworth, Chebyshev (Type I and
Type II) and Elliptic Filters for the Same Filter Specification
The objective of this experiment is to compute the order of Butterworth, Type I
Chebyshev, Type II Chebyshev and elliptic filter for the following specifications
using bilinear transformation technique (BLT). -3 dB cut-off frequency at 5 Hz, and
an attenuation of 40 dB at 10 Hz, use bilinear transformation technique. Assume the
368 8 Infinite Impulse Response Filter

#Order of different IIR filters

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Filter specification
fsample=1000
f_pass, f_stop, g_pass, g_stop, Td =5, 10, 3, 40, 1
wp=f_pass/(fsample/2)
ws=f_stop/(fsample/2)
#Step 2: Pre-warping
omega_pass=(2/Td)*np.tan(wp/2)
omega_stop=(2/Td)*np.tan(ws/2)
#Step 3: Computing the order of different filters
N1,w1=signal.buttord(omega_pass,omega_stop,g_pass,g_stop)
N2,w2=signal.cheb1ord(omega_pass,omega_stop,g_pass,g_stop)
N3,w3=signal.cheb2ord(omega_pass,omega_stop,g_pass,g_stop)
N4,w4=signal.ellipord(omega_pass,omega_stop,g_pass,g_stop)
#Step 4: Plotting the order of different filters
filter_name=['Butterworth','Cheby1','Cheby2','Elliptic']
order1=[N1,N2,N3,N4]
plt.bar(filter_name,order1),plt.title('Order of different filters')
for i, v in enumerate(order1):
plt.text(i, v+0.1, str(v),color = 'blue', fontweight = 'bold')
plt.show()

Fig. 8.66 Comparing the order of different IIR filters

Order of different filters

7
7

5 5
5

3
3

0
Butterworth Cheby1 Cheby2 Elliptic

Fig. 8.67 Comparison of the order of IIR filters

8.7 Elliptic Filter 369

# Low pass Butterworth filter

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Signal generation
t1=np.linspace(0,1,100)
f1, f2, f3=5, 0, 10
x1=np.sin(2*np.pi*f1*t1)
x2=np.sin(2*np.pi*f2*t1)
x3=np.sin(2*np.pi*f3*t1)
x=np.concatenate([x1,x2,x3])
t=np.linspace(0,1,len(x))
plt.figure(1),plt.subplot(2,3,1),plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Input signal')
#Step 2: Design of filter
N=[2,5,10,20,25] #Order of the filter
fsamp=100 #Sampling frequency
f_cut=8 #Cut-off frequency
fn=fsamp/2
wc=f_cut/fn
for i in range(len(N)):
b, a = signal.butter(N[i],wc,'low')
#Step 3: Obtaining the output
plt.figure(1), plt.subplot(2,3,i+2)
y=signal.lfilter(b,a,x)
plt.plot(t,y),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Filtered signal (N={})'.format(N[i]))
plt.tight_layout()

Fig. 8.68 Lowpass Butterworth filter to filter sinusoidal signal

sampling frequency to be 1000 Hz. The python code, which computes the order of
different IIR filter, is shown in Fig. 8.66, and the corresponding output is shown in
Fig. 8.67.
Inferences
From Fig. 8.67, it is possible to observe the following facts:
1. The order of Butterworth filter to meet the given filter specification is 7.
2. The order of Chebyshev filter to meet the given filter specification is 5.
3. The order of elliptic filter to meet the given filter specification is 3.
4. The order of Butterworth filter is higher than the order of Chebyshev and elliptic
filters.
5. Elliptic filter meets the given filter specification with a minimum order.
370 8 Infinite Impulse Response Filter

Input signal Filtered signal (N=2) Filtered signal (N=5)

1 1 1

Amplitude
Amplitude

Amplitude
0 0 0

–1 –1 –1
0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0
Time Time Time
Filtered signal (N=10) Filtered signal (N=20) Filtered signal (N=25)
1 1 1

Amplitude
Amplitude

Amplitude

0 0 0

–1 –1 –1
0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0
Time Time Time

Fig. 8.69 Butterworth lowpass filter output

Experiment 8.29 Filtering of Sinusoidal Signal Using Butterworth Filter of a

Different Order
This experiment aims to filter a sinusoidal signal consisting of three frequencies: 5, 0
and 10 Hz. This signal is to be filtered by a Butterworth lowpass filter of different
orders, namely, 2, 5, 10, 20 and 25. The cut-off and sampling frequencies chosen are
8 and 100 Hz, respectively. The python code, which implements the above-
mentioned task is shown in Fig. 8.68, and the corresponding outputs are shown in
Fig. 8.69.
Inferences
The following inferences can be made from Fig. 8.69:
1. The sinusoidal input signal contains three frequencies, namely, 5, 0 and 10 Hz.
2. The input signal is passed through lowpass Butterworth filter of order 2, 5, 10, 20
and 25.
3. The filter retains 5 Hz frequency, and it blocks 10 Hz frequency component.
4. The extent of filtering increases with an increase in the order of the filter. The
10 Hz frequency component is blocked effectively when the filter order is
20 and 25.
5. The group delay increases with an increase in the order of the filter. The delay can
be observed in the filtered signal with orders 20 and 25.
Task
1. Use the highpass filter with cut-off frequency of 10 Hz to repeat Experiment 8.29.
Compare the result with the result of Experiment 8.29.
8.7 Elliptic Filter 371

#Comparing the performances of IIR filters

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Signal generation
t1=np.linspace(0,1,100)
f1,f2,f3=5,0,10
x1=np.sin(2*np.pi*f1*t1)
x2=np.sin(2*np.pi*f2*t1)
x3=np.sin(2*np.pi*f3*t1)
x=np.concatenate([x1,x2,x3])
t=np.linspace(0,1,300)
plt.figure(1),plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Input signal')
#Step 2: Design of filter
N=5 #Order of the filter
fsamp=100 #Sampling frequency
f_cut=8 #Cut-off frequency
fn=fsamp/2
wc=f_cut/fn
b1, a1 = signal.butter(N,wc,'low')
b2, a2 = signal.cheby1(N,3,wc,'low')
b3, a3 = signal.cheby2(N,40,wc,'low')
b4, a4 = signal.ellip(N,3,40,wc,'low')
#Step 3: Obtaining the output of the filter
y1=signal.lfilter(b1,a1,x)
y2=signal.lfilter(b2,a2,x)
y3=signal.lfilter(b3,a3,x)
y4=signal.lfilter(b4,a4,x)
#Step 4: Plotting the output signals
plt.figure(2)
plt.subplot(2,2,1),plt.plot(t,y1),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Butterworth filter output')
plt.subplot(2,2,2),plt.plot(t,y2),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Chebyshev (Type 1) filter output')
plt.subplot(2,2,3),plt.plot(t,y3),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Chebyshev (Type2) filter output')
plt.subplot(2,2,4),plt.plot(t,y4),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Elliptic filter output')
plt.tight_layout()

Fig. 8.70 Python code to compare the performances of IIR filters

Experiment 8.30 Filtering of Sinusoidal Signal with Different IIR Filters

of the Same Order
This experiment aims to filter a sinusoidal signal consisting of three frequencies: 5, 0
and 10 Hz. This signal is to be filtered by a Butterworth, Chebyshev (Type I),
Chebyshev (Type II) and elliptic lowpass filters. The order of the filter is fixed as
372 8 Infinite Impulse Response Filter

Input signal

1.00

0.75

0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 8.71 Input signal

Output of Butterworth filter Output of Chebyshev (Type1) filter

1 1

0 0

–1 –1
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

Output of Chebyshev (Type2) filter Output of Elliptic filter

1
0.5

0.0
0

–0.5

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

Fig. 8.72 Filtered signal using different IIR filters

5. The cut-off frequency and the sampling frequency chosen are 8 and 100 Hz,
respectively.
The python code, which implements the above-mentioned task, is shown in
Fig. 8.70, and the corresponding outputs are shown in Figs. 8.71 and 8.72.
8.7 Elliptic Filter 373

Inferences
From Fig. 8.71, it is possible to observe that the input signal has three frequency
components, namely 5, 0 and 10 Hz. This
1. Input signal is passed through lowpass Butterworth, Chebyshev (Type I and Type
II) and elliptic filters.
2. Butterworth filter performs smoothing action without ripples. But the extent of
filtering is poor. Significant high-frequency components appear along with low
frequency components.
3. In Chebyshev and elliptic filters, the high-frequency components are filtered
effectively; however, the filtering action is not smooth. Ripples appear in the
filtered output signal.
Exercises
1. The transfer function of analogue filter is given by H ðsÞ = sþ10
10
. Write a python
code to obtain the transfer function of the equivalent digital filter using
(a) impulse invariant technique and (b) bilinear transformation technique.
Assume the sampling period to be 1 s.
2. A normalized first-order lowpass Butterworth filter transfer function is given by
H ðsÞ = ðsþ1
1
Þ. Write a python code to convert this filter into desired highpass
filter with a cut-off frequency of 5 rad/s. Using subplot, plot the magnitude
response of normalized filter and the desired highpass filter.
3. Write a python code to obtain the order and cut-off frequency of analogue
Butterworth filter that has -3 dB cut-off frequency of 20 rad/s and 10 dB of
attenuation at 40 rad/s. Plot the magnitude response of the filter.
4. Design an analogue bandpass filter that has the following characteristics:
(a) -3 dB upper and lower cut-off frequency of 100 Hz and 10 kHz
(b) Stopband attenuation of 30 dB at 50 Hz and 25 kHz
(c) Monotonic frequency response
Plot the magnitude response of the above-mentioned filter.
5. Write a python code to design a digital lowpass Butterworth filter for the
following specification using impulse invariant technique (a) -3 dB cut-off
frequency at 250 Hz and (b) magnitude of frequency response down at least
10 dB at 500 Hz. Assume the sample to be 1000 samples/s. Plot the magnitude
response of the filter.
6. Write a python code to design a digital filter using the bilinear transformation
technique for the following specification: (a) maximally flat frequency response
with -3 dB cut-off at 10 rad/s, (b) 30 dB of attenuation at all frequencies greater
than 20 rad/s and (c) assuming the sampling frequency to be 1000 samples/s.
Plot the magnitude response of the filter.
7. Write a python code to design a Type I Chebyshev filter using bilinear trans-
formation technique (BLT) that has a passband of 0 to -3 dB cut-off frequency
at 5 kHz and attenuation of at least 30 dB for frequencies greater than 10 kHz.
374 8 Infinite Impulse Response Filter

Assume the sampling frequency to be 20 kHz. Assume T = 1 s. Plot the

magnitude response of the filter.
8. Write a python code to design a Chebyshev Type II digital bandstop filter for the
following specifications: (a) Stopband frequencies are from 1000 to 2000 Hz
with an attenuation of 40 dB. (b) Passband frequencies are below 800 Hz and
above 2500 Hz with the passband ripple of 3 dB. Sampling frequency
(Fs) = 8 kHz. Use BLT method for transformation. Plot its frequency response.
9. Write a python code to design an elliptic lowpass filter with the passband
frequency of 10 Hz. Assume the order of the filter is 2, passband ripple 3 dB
and stopband attenuation is 40 dB.
10. Obtain the order of Butterworth filter, Type I Chebyshev filter, Type II
Chebyshev filter and elliptic filter that has -3 dB bandwidth of 10 Hz and an
attenuation of 30 dB at 20 Hz using bilinear transformation technique. Assume
the sampling frequency to be 1500 Hz.
Objective Questions
1. The filter has maximally flat response at both passband and stopband is called as
A. Elliptic filter
B. Butterworth filter
C. Chebyshev filter
D. Inverse Chebyshev filter
2. The filter has monotonic response at passband and ripple at stopband
A. Elliptic filter
B. Butterworth filter
C. Chebyshev filter
D. Inverse Chebyshev filter
3. The filter has ripple at passband and monotonic response at stopband
A. Elliptic filter
B. Butterworth filter
C. Chebyshev filter
D. Inverse Chebyshev filter
4. The filter has ripple at both passband and stopband
A. Elliptic filter
B. Butterworth filter
C. Chebyshev filter
D. Inverse Chebyshev filter
5. Mapping between S-plane to Z-plane using approximation derivative method is
1 - z-1
A. s = T
1þz - 1
B. s = T
8.7 Elliptic Filter 375

C. s = T
1 - z-1
D. s = T
1þz - 1

6. Identify the wrong statement

A. Impulse invariant method exists one to one mapping between Ω and ω.
B. Impulse invariant method does not exist one to one mapping between Ω and
ω.
C. Impulse invariant method retains the stability while converting analogue
filter into digital filter.
D. Impulse invariant method is appropriate for the design of lowpass and
bandpass filters only.
7. Mapping between S-plane to Z-plane using bilinear transformation method is
1þz - 1
A. s = 2
T 1 - z-1
1 - z-1
B. s = 2
T 1þz - 1
1þz - 1
C. s = T
2 1 - z-1
1 - z-1
D. s = T
2 1þz - 1

8. Identify the correct statement

A. The order (N ) of the Chebyshev polynomial function TN(x) is even; the
graph will pass through origin.
B. The order (N ) of the Chebyshev polynomial function TN(x) is even, and the
amplitude of the TN(x) is equal to zero at x equal to zero.
C. The order (N ) of the Chebyshev polynomial function TN(x) is odd, and the
amplitude of the TN(x) is equal to non-zero at x equal to zero.
D. The order (N ) of the Chebyshev polynomial function TN(x) is even; the
graph will not pass through origin.
9. The poles of Chebyshev Type I analogue filter are located on an S-plane in the
form of
A. Circle
B. Parabola
C. Ellipse
D. Hyperbola
10. The Chebyshev type II analogue filter has
A. Zeros as the numerator and all the zeros lie on the real axis only.
B. Zeros as the numerator and all the zeros lie on the imaginary axis only.
C. Zeros as the numerator and all the zeros lie on both real and imaginary axes.
D. None of the above.
376 8 Infinite Impulse Response Filter

Bibliography

1. Harry Y.F. Lam, “Analog and Digital Filters: Design and Realization”, Prentice Hall, 1979.
2. Andreas Antoniou, “Digital Filters: Analysis, Design and Signal Processing Applications”,
McGraw Hill, 2018.
3. Dietrich Schlichtharle, “Digital Filters: Basics and Design”, Springer, 2011.
4. Leland B. Jackson, “Digital Filters and Signal Processing”, Springer, 1988.
5. Thomas Haslwanter, “Hands-on Signal Analysis with Python”, Springer, 2022.
Chapter 9
Quantization Effect of Digital Filter
Coefficients

Learning Objectives
After completing this chapter, the reader is expected to
• Understand and implement approximation of numbers through flooring, ceiling
and rounding operations.
• Analyse the impact of quantizing the finite impulse response filter coefficients.
• Analyse the impact of quantizing the infinite impulse response filter coefficients.
• Demonstrate limit cycle oscillation due to quantization of IIR filter coefficients.

Roadmap of the Chapter

The roadmap of this chapter is shown below. This chapter discusses the effect of
fixed-point representations of digital systems and the effect of quantization using
rounding, two’s complement and magnitude truncation approaches. Also, it dis-
cusses the impact of the finite word length effect of FIR and IIR filters.

Finite word length effect

Fixed point Representation Floating point Representation

Quantization Errors

Overflow

Limit Cycle Oscillations

Coefficient Quantization Effect

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 377
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_9
378 9 Quantization Effect of Digital Filter Coefficients

PreLab Questions
1. What is the finite word length effect in DSP?
2. List out the difference between fixed-point and floating-point number
representations.
3. What is the significance of sign-magnitude representation?
4. What are quantization and quantization errors?
5. What is rounding, and how does it the quantization process?
6. What do you mean by two’s complement truncation in quantization, and what is
the range of quantization error?
7. What is magnitude truncation, and can it suppress the limit cycle oscillation?
8. What is overflow? And mention the types of overflow.
9. What do you mean by limit cycle oscillation, and how it exists in signal
processing?
10. What do you mean by coefficient quantization, and how it affects the stability of
the result?

9.1 Number Representation

Numeric representation and type of arithmetic profoundly influence the performance

of DSP system. The two forms of representation of numbers are fixed-point repre-
sentation and floating-point representation. Fixed-point arithmetic represents num-
bers in a fixed range with a finite number of bits of precision. Numbers outside the
specified range will either saturate or wrap around. In general, fixed-point represen-
tation is preferred for high speed and lower cost. Floating-point arithmetic represents
every number in two parts (a) a mantissa and (b) an exponent. Floating-point
representation has a higher dynamic range, and there is no need for scaling, which
makes it attractive for complex algorithms.
The implementation of digital filters involves the use of finite precision arith-
metic. This leads to quantizing the filter coefficients and the results of the arithmetic
operations. Such quantization operations are non-linear and cause a filter response
substantially different from the response of the underlying infinite precision model.

9.2 Fixed-Point Quantization

Quantization is the process of approximating a quantity X into a quantity Q(X). It is

approximately equal to X, but it has some distortion errors. Quantization exists when
it represents the real numbers by nearest integers and is to reduce the word length of
a binary representation of X by reducing the number of bits after the binary point.
The relationship between X and Q(X) is called as ‘quantization characteristics’.
Quantization commonly comes across in digital signal processing in the form of
9.2 Fixed-Point Quantization 379

1. Rounding
2. Two’s complement truncation
3. Magnitude truncation

9.2.1 Fixed-Point Quantization by Rounding

If X is an original value to be quantized using the rounding method, and the

quantized output is denoted as Qr(X). If the quantized value has N bits to the right
of the binary point, then the quantization step size is Δ = 2-N. In general, rounding
selects the quantized integer nearest to the original value. But the error between the
quantized and original value is not more than ± Δ2 . Hence, the rounding error is
denoted as εr = Qr(X) - X. Then, the rounding error is ranging from - Δ2 ≤ εr ≤ Δ2 .
The error resulting from quantization can be modelled as a random variable uni-
formly distributed over the appropriate error range. Hence, the round off error can be
considered error-free calculations that have been corrupted by additive white noise.
The mean value of the rounding error is zero and variance of the rounding error is
σ2εr = Δ negativethinmathspace=negativethinmathspace12 .
2

Experiment 9.1 Perform the Fixed-Point Quantization Using the Rounding

Method
This experiment deals with quantizing value using a rounding approach. The first
and foremost method of quantization is rounding, which is mathematically written as

X r = Δ × roundðx=ΔÞ ð9:1Þ

where x denotes the value to be quantized, Δ represents the step size and is computed
as Δ = 2-N, N denotes the number of bits and Xr represents the quantized output
using rounding. The python code for computing quantization using the rounding
approach is given in Fig. 9.1, and its corresponding simulation result is shown in
Fig. 9.2.

# This program performs quantization using rounding

import numpy as np
h=float(input('Enter the value to be Quantized: '));
B=int(input('Enter the Number of Bits (N): '));
Q = 1/(2**(B))
Qhr=Q*np.round(h/Q)#Rounding
print('The input unquantized value : ',h)
print('The Quantized result using Rounding: ', Qhr)

Fig. 9.1 Python code for quantization using rounding

380 9 Quantization Effect of Digital Filter Coefficients

Fig. 9.2 Simulation result

of python code given in Enter the value to be Quantized: 0.126
Fig. 9.1 Enter the Number of Bits (N): 2
The input unquantiz ed value : 0.126
The Quantized result using Rounding: 0.25

Inference
1. From Fig. 9.2, it is possible to infer that the value to be quantized is ‘0.126’ and
the number of bits that can be used to represent the value is ‘2’.
2. The two bits can represent the maximum of 4 levels (0–0.25, 0.25–0.5, 0.5–0.75
and 0.75–1.0).
3. Therefore, the input ‘0.126’ is quantized, and the result is 0.25 (0.126 is greater
than 0.125).
Task
1. Enter the value to be quantized is 5.15, and 8 bits are used for the quantization.
Observe the quantized result.

9.2.2 Fixed-Point Quantization Using Two’s Complement

Truncation

The two’s complement truncation always gives the quantized value less than or
equal to the original value. Hence, the truncation error is denoted as εt = Qt(X) - X.
Then, the truncation error is ranging from -Δ ≤ εt ≤ 0. The error resulting from
quantization can be modelled as a random variable uniformly distributed over the
appropriate error range. Hence, the two’s complement truncation error can be
considered error-free calculations corrupted by additive white noise. The mean
value of the two’s complement truncation error can be obtained as μεt = - Δ2 . The
variance of the two’s complement truncation error is calculated as σ 2εt = Δ2=12.
Experiment 9.2 Perform Fixed-Point Quantization Using Two’s Complement
This experiment deals with quantizing the value using two’s complement truncation
approach. The quantization using two’s complement truncation method is denoted as

X t = Δ × bðx=ΔÞc ð9:2Þ

where bc denotes flooring operation. Xt represents the quantized result using trun-
cation method. The python code, which computes the fixed-point quantization of a
number using two’s complement truncation method, is shown in Fig. 9.3, and its
simulation result is depicted in Fig. 9.4.
9.2 Fixed-Point Quantization 381

# This program performs quantization using 2’s complement truncation

import numpy as np
h=float(input('Enter the value to be Quantized: '));
B=int(input('Enter the Number of Bits (N): '));
Q = 1/(2**(B))
Qht=Q*np.floor(h/Q)#2's Complement truncation
print('The input unquantized value : ',h)
#print('The Quantized result using Rounding: ', Qhr)
print('The Quantized result using 2s Complement truncation: ', Qht)

Fig. 9.3 Python code for quantization using two’s complement truncation

Enter the value to be Quantized: 0.126

Enter the Number of Bits (N): 2
The input unquantiz ed value : 0.126
The Quantized result using 2s Complement truncation: 0.0

Fig. 9.4 Simulation result of python code given in Fig. 9.3

Inference
From Fig. 9.4, it is possible to infer that the value to be quantized is 0.126, and the
resultant value is 0 for N = 2.
Task
1. Choose the proper value of N and execute the python code given in Fig. 9.3 to get
0.126 as the output.

9.2.3 Fixed-Point Quantization Using Magnitude Truncation

Magnitude truncation gives the quantized value less than the original value for
X > 0, and the quantized result is greater than original for X < 0. Hence, the
magnitude truncation error is denoted as εmt = Qmt(X) - X.
Then, the magnitude truncation error is ranging from -Δ ≤ εmt ≤ 0 for X > 0 and
0 ≤ εmt ≤ Δ for X < 0. The mean value of the magnitude truncation error is obtained
as 0, and variance of the magnitude truncation error is calculated as σ 2εmt = Δ2=3. The
specific advantage of magnitude truncation lies in its inherent capability of limit
cycle suppression.
382 9 Quantization Effect of Digital Filter Coefficients

# This program performs quantization using Magnitude truncation

import numpy as np
h=float(input('Enter the value to be Quantized: '));
B=int(input('Enter the Number of Bits (N): '));
Q = 1/(2**(B))
# Magnitude Truncation
if h > 0:
Qhmt=Q*np.floor(h/Q)
else:
Qhmt=Q*np.ceil(h/Q)
print('The input unquantized value : ',h)
print('The Quantized result using Magnitude truncation: ', Qhmt)

Fig. 9.5 Python code for quantization using magnitude truncation

Enter the value to be Quantized: 0.126

Enter the Number of Bits (N): 2
The input unquantiz ed value : 0.126
The Quantized result using M agnitude truncation: 0.0

Fig. 9.6 Simulation result of python code given in Fig. 9.5

Experiment 9.3 Perform Fixed-Point Quantization Using Magnitude

Truncation
The quantization using magnitude truncation is represented as

Δ × bðx=ΔÞc, for x ≥ 0
X mt = ð9:3Þ
Δ × dðx=ΔÞe, for x < 0

where de denotes ceiling operation. Xmt represents the quantized result using mag-
nitude truncation. The python code of this experiment is shown in Fig. 9.5, and the
simulation result is displayed in Fig. 9.6.
Inference
The following inferences can be made from this experiment:
1. Figure 9.6 shows that the value to be quantized is 0.126, and the number of bits is
chosen as 2. The result of quantization using the magnitude truncation is 0. This
simulation result is on par with the theoretical result.
2. The selection of the number of bits is essential for quantization.
3. The quantization method is crucial for the hardware implementation of the digital
systems.
9.3 Coefficient Quantization 383

Task
1. Execute the python code given in Fig. 9.5, and enter the value to be quantized as
‘0.126’ and obtain the minimum value of ‘N’, which will give the quantized result
that is the same as the input value.

9.3 Coefficient Quantization

The filter coefficients can be obtained using filter design approaches based on the
given set of filter specifications. These filter coefficients are represented as a system
function of the filter H(z), and they can be used in signal processing applications.
These filter coefficients may be integer or non-integer numbers. If the filter coeffi-
cient is an integer, then the finite precision format is enough to represent it in digital
computation. Otherwise, the coefficient is non-integer, then infinite precision format
is necessary to represent it for the accurate result of the applications. However, in
real-time scenario the hardware setup in the application may not be able to store the
value of filter coefficients as it is due to the limited register size or finite precision
processor. The representation of the filter coefficients from infinite precision to finite
number precision may give coefficients quantization. This coefficient quantization
can change the location of the filter poles and zeros. As a result, after implementing a
filter, it may observe that the frequency response of the filter is quite different from
that of the original design. The coefficients obtained by design methods are real or
complex. These coefficients are often stored in a finite length register for real-world
digital signal processing. The coefficients are often rounded to accommodate it in the
finite length register. This causes a rounding error, which will influence the filter
characteristics. The frequency response of the quantized filter coefficients will differ
from the desired frequency response. Sensitivity of the filter response to coefficient
quantization is dependent on the type of filter structures. Detail about this will be
discussed later in this chapter.
Experiment 9.4 Effect of Quantization Using the Rounding Approach of FIR
Filters
This experiment deals with the effect of coefficient quantization of FIR filters. Here,
the method to quantize the filter coefficients is considered as rounding method. The
FIR filter coefficients are computed by using the built in python command ‘a =
signal.firwin(n, cutoff = 0.25, window = "hamming")’. Here ‘n’ denotes the number
of the filter coefficients, ‘cutoff = 0.25’ represents cut-off frequency of the filter is
0.25π rad/sample and the ‘window = "hamming"’ indicates the window function
used for the FIR filter design. The python code for this experiment is shown in
Fig. 9.7. In this figure, ‘B = 2’ denotes the number of bits used to quantize each filter
coefficient. While increasing the value of B = 3, 4, 5, 6, . . ., the quantized filter’s
impulse and frequency response will approach the original filter’s impulse and
384 9 Quantization Effect of Digital Filter Coefficients

from scipy import signal

import numpy as np
import matplotlib.pyplot as plt
n = 31
n1=np.arange(0,n);
a = signal.firwin(n, cutoff = 0.25, window = "hamming")
B = 16;# Number of Bits
Q = 1/(2**(B))
Qhr=Q*np.round(a/Q)#Rounding
#Obtaining the Frequency response
w,H=signal.freqz(a)
wq,Hq=signal.freqz(Qhr)
#Obtaining the pole-zero plot
z,p,k=signal.tf2zpk(a,1)
zq,pq,kq=signal.tf2zpk(Qhr,1)
#Plotting the responses
plt.figure(1),plt.subplot(2,2,1),plt.stem(a),plt.xlabel('n-->'),plt.ylabel('h[n]')
plt.title('h[n]'),plt.subplot(2,2,2),plt.plot((w/np.pi),20*np.log10(np.abs(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('|H($\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(z),np.imag(z),'ro'),plt.plot(np.real(p),np.imag(p),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.xlabel('Real part'),plt.ylabel('Imaginary part'),
plt.title('Pole-zero plot'),plt.subplot(2,2,4),plt.plot((w/np.pi),np.unwrap(np.angle(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi(e^{jw})$')
plt.tight_layout()
plt.figure(2),plt.subplot(2,2,1),plt.stem(Qhr),plt.xlabel('n-->'),plt.ylabel('$h_q[n]$')
plt.title('$h_q[n]$ with N = {} bits'.format(B))
plt.subplot(2,2,2),plt.plot((w/np.pi),20*np.log10(np.abs(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('$|H_q(\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(zq),np.imag(zq),'ro')
plt.plot(np.real(pq),np.imag(pq),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.xlabel('Real part'),plt.ylabel('Imaginary part'),
plt.title('Pole-zero plot'),plt.subplot(2,2,4),plt.plot((w/np.pi),np.unwrap(np.angle(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi_q(e^{jw})$')
plt.tight_layout()

Fig. 9.7 Python code to analyse coefficient quantization effect of FIR filter

frequency response. The simulation result of the python code given in Fig. 9.7 is
displayed in Fig. 9.8.
Inferences
From Fig. 9.8, the following observations can be made:
1. Figure 9.8a displays the impulse response of the original FIR filter components
and its magnitude, phase and pole-zero plot.
2. Figure 9.8b shows the number of bits selected as 2 for quantizing FIR filter
coefficients and its impulse response, magnitude, phase and pole-zero plot. From
this figure, it is possible to infer that the impulse response of the quantized FIR
filter looks like a rectangular pulse, which is completely deviated from the
original impulse response (i.e.) sinc function. Magnitude and phase responses
9.3 Coefficient Quantization 385

Fig. 9.8 Simulation result of the python code is shown in Fig. 9.7. (a) Original (unquantized). (b)
Quantized with N = 2 bits. (c) Quantized with N = 4 bits. (d) Quantized with N = 16 bits

differ completely from the original magnitude and phase responses. In the pole-
zero plot, some zeros occur at the origin, whereas in the original pole-zero plot,
there is no zero at the origin; it indicates that after the filter coefficients, a lot of
coefficients become zero due to less number of bits allocated for the
representation.
3. From Fig. 9.8c, d, when the number of bits (N ) is chosen as 4 and 16, the impulse
response of the quantized FIR filter approaches the original impulse response, and
the magnitude and phase response approaches the original one.
4. From these figures, it is possible to confirm that allocating the number of bits to
represent the filter coefficients plays a significant role in DSP computations.
Experiment 9.5 Verify the Effect of Quantization Using the Two’s Complement
Truncation Approach of FIR Filters
This experiment discusses the effect of quantization using two’s complement trun-
cation method of FIR filter coefficients. The python code to perform the two’s
complement truncation-based quantization of FIR filter coefficients is shown in
Fig. 9.9. In this experiment, we have chosen the order of the FIR filter as 31, the
window function is considered ‘Hamming’, and the cut-off frequency is 0.25π
rad/samples. The finite number of bits required to represent each filter coefficient
386 9 Quantization Effect of Digital Filter Coefficients

# This program verifies the effect of coefficient quantization of FIR filters

# Twos complement truncation
from scipy import signal
import numpy as np
import matplotlib.pyplot as plt
n = 31
n1=np.arange(0,n);
a = signal.firwin(n, cutoff = 0.25, window = "hamming")
B = 16;# Number of Bits
Q = 1/(2**(B))
Qhr=Q*np.floor(a/Q)# Twos complement truncation
#Obtaining the Frequency response
w,H=signal.freqz(a)
wq,Hq=signal.freqz(Qhr)
#Obtaining the pole-zero plot
z,p,k=signal.tf2zpk(a,1)
zq,pq,kq=signal.tf2zpk(Qhr,1)
#Plotting the responses
plt.figure(1),plt.subplot(2,2,1),plt.stem(a),plt.xlabel('n-->'),plt.ylabel('h[n]')
plt.title('h[n]'),plt.subplot(2,2,2),plt.plot((w/np.pi),20*np.log10(np.abs(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('|H($\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(z),np.imag(z),'ro'),plt.plot(np.real(p),np.imag(p),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta))
plt.xlabel('Real part'),plt.ylabel('Imaginary part'),plt.title('Pole-zero plot')
plt.subplot(2,2,4),plt.plot((w/np.pi),np.unwrap(np.angle(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi(e^{jw})$')
plt.tight_layout()
plt.figure(2),plt.subplot(2,2,1),plt.stem(Qhr),plt.xlabel('n-->'),plt.ylabel('$h_q[n]$')
plt.title('$h_q[n]$ with N = {} bits'.format(B))
plt.subplot(2,2,2),plt.plot((w/np.pi),20*np.log10(np.abs(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('$|H_q(\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(zq),np.imag(zq),'ro'),plt.plot(np.real(pq),np.imag(pq),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.xlabel('Real part'),plt.ylabel('Imaginary part'),
plt.title('Pole-zero plot'),plt.subplot(2,2,4),plt.plot((w/np.pi),np.unwrap(np.angle(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi_q(e^{jw})$')
plt.tight_layout()

Fig. 9.9 Python code for two’s complement truncation

is selected as 2, 4 and 16. The simulation result of the python code given in Fig. 9.9
is shown in Fig. 9.10.
Inferences
The following observations can be drawn from Fig. 9.10:
1. When the number of bits N = 2, the impulse response of the FIR filter is
completely deviated from the original impulse response of the FIR filter. Simi-
larly, the magnitude and phase responses differ from the original one. From the
pole-zero plot, the locations of the zeros of quantized filter are dislocated from the
original positions.
9.3 Coefficient Quantization 387

Fig. 9.10 Simulation result of the python code given in Fig. 9.9. (a) Original filter. (b) Quantized
filter with N = 2 bits. (c) Quantized filter with N = 4 bits. (d) Quantized filter with N = 16 bits

2. When the number of bits is chosen as 4, the impulse response is still not the same
as the original one. From this, 4 bits for each filter coefficient are insufficient for
the finite precision.
3. When the number of bits is considered as 16. The impulse, magnitude and phase
responses are exactly the same as the original one. This is also evident from the
pole-zero plot.
4. From this experiment, we must understand that allocating the number of bits to
represent the filter coefficients is very important for the FIR filter design. Even
though the FIR filter is stable after the quantization, the filter response is not as
expected.
Experiment 9.6 Verify the Effect of Quantization Using the Magnitude Trun-
cation Approach of FIR Filters
This experiment explores the effect of quantization using magnitude truncation of
FIR filter coefficients. The magnitude truncation is another type of quantization
approach that can represent the infinite precision to the finite precision of filter
coefficients. The python code to perform the magnitude truncation of the FIR filter
coefficients is shown in Fig. 9.11. In this experiment, we have chosen the order of
FIR filter is 31, the window function is considered ‘Hamming’ and the cut-off
frequency is 0.25π rad/samples. The finite number of bits required to represent
388 9 Quantization Effect of Digital Filter Coefficients

# This program verifies the effect of coefficient quantization of FIR filters

# Magnitude truncation
from scipy import signal
import numpy as np
import matplotlib.pyplot as plt
n = 31
n1=np.arange(0,n);
a = signal.firwin(n, cutoff = 0.25, window = "hamming")
B = 16;# Number of Bits
Q = 1/(2**(B))
Qhmt=np.zeros(len(a))
# Magnitude Truncation
for i in range(len(a)):
if a[i] > 0:
Qhmt[i]=Q*np.floor(a[i]/Q)
else:
Qhmt[i]=Q*np.ceil(a[i]/Q)
#Obtaining the Frequency response
w,H=signal.freqz(a)
wq,Hq=signal.freqz(Qhmt)
#Obtaining the pole-zero plot
z,p,k=signal.tf2zpk(a,1)
zq,pq,kq=signal.tf2zpk(Qhmt,1)
#Plotting the responses
plt.figure(1),plt.subplot(2,2,1),plt.stem(a),plt.xlabel('n-->'),plt.ylabel('h[n]'),plt.title('h[n]')
plt.subplot(2,2,2),plt.plot((w/np.pi),20*np.log10(np.abs(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('|H($\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(z),np.imag(z), 'ro'),plt.plot(np.real(p),np.imag(p), 'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta))
plt.xlabel('Real part'),plt.ylabel('Imaginary part'),plt.title('Pole-zero plot')
plt.subplot(2,2,4),plt.plot((w/np.pi),np.unwrap(np.angle(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi(e^{jw})$')
plt.tight_layout()
plt.figure(2),plt.subplot(2,2,1),plt.stem(Qhmt),plt.xlabel('n-->'),plt.ylabel('$h_q[n]$')
plt.title('$h_q[n]$ with N = {} bits'.format(B))
plt.subplot(2,2,2),plt.plot((w/np.pi),20*np.log10(np.abs(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('$|H_q(\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(zq),np.imag(zq), 'ro')
plt.plot(np.real(pq),np.imag(pq), 'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.xlabel('Real part'),plt.ylabel('Imaginary part'),
plt.title('Pole-zero plot'),plt.subplot(2,2,4),plt.plot((w/np.pi),np.unwrap(np.angle(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi_q(e^{jw})$')
plt.tight_layout()

Fig. 9.11 Python code for magnitude truncation

each filter coefficient is chosen as 2, 4 and 16. The simulation result of the python
code given in Fig. 9.11 is shown in Fig. 9.12.
Inferences
From Fig. 9.12, the following inferences can be made:
1. When the number of bits is chosen as N = 2, the impulse response of the
quantized filter has only one non-zero element, and all other elements are zero.
9.3 Coefficient Quantization 389

Fig. 9.12 Simulation result of the python code given in Fig. 9.11. (a) Original filter. (b) Quantized
filter with N = 2 bits. (c) Quantized filter with N = 4 bits. (d) Quantized filter with N = 16 bits

From the pole-zero plot, it is possible to infer that all the zeros of the FIR filters
are placed at the origin, and it confirms that improper allocation of bits will lead to
error frequency response.
2. While choosing the value N is 4, the impulse response of the quantized filter looks
like a triangular function; it is not a Gaussian function. Also, it shows that all the
zeros of the FIR filters are squeezed into only four locations.
3. If N = 16, the impulse, magnitude, and phase responses of the quantized filter are
very close to the responses of the original filter. Also, the pole-zero plot is similar
to the original one.
Experiment 9.7 Verify the Effect of Quantization Using the Rounding
Approach of Butterworth IIR Filter
This experiment discusses the quantization of Butterworth filter coefficients com-
puted using the bilinear transformation technique (BLT) that has a passband gain of
0 to -3 dB, cut-off frequency of 2 kHz and an attenuation of at least 20 dB for
frequencies greater than 5 kHz. Assume the sampling frequency to be 20 kHz.
Butterworth filter coefficients are quantized using the rounding approach. The
python code to verify the concept of this experiment is shown in Fig. 9.13. Here
the number of bits to be allocated to represent each coefficient of the IIR Butterworth
filter is chosen as 4, 8, 12 and 16. The simulation results of the python code given in
Fig. 9.13 are shown in Fig. 9.14.
390 9 Quantization Effect of Digital Filter Coefficients

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
# Specifications of Filter
fsam, fp, fs, Ap, As, Td=20000, 2000, 5000, 3, 20, 1 # Sampling frequency
wp=2*np.pi*(fp/fsam) # pass band freq in radian per sample
ws=2*np.pi*(fs/fsam) # Stop band freq in radian per sample
# prewarping process
omega_p=(2/Td)*np.tan(wp/2)
omega_s=(2/Td)*np.tan(ws/2)
#Computation of order and normalized cut-off frequency
N, omega_c=signal.buttord(omega_p,omega_s,Ap,As,analog=True)
print('Order of the Filter N =', N),print('Cut-off frequency= {:.4f} rad/s '. format(omega_c))
# Computation of H(s)
b, a=signal.butter(N,omega_c,'low', analog=True)
bz, az=signal.bilinear(b, a, Td)
n=15
n1=np.arange(0,n);
x=(n1==0)
y=signal.lfilter(bz,az,x)
W,H = signal.freqz(bz,az)
#Obtaining the pole-zero plot
z,p,k=signal.tf2zpk(bz,az)
B = 2;# Number of Bits
Q = 1/(2**(B))
Qbr=Q*np.round(bz/Q)#Rounding
Qar=Q*np.round(az/Q)#Rounding
yr=signal.lfilter(Qbr,Qar,x)
Wq,Hq = signal.freqz(Qbr,Qar)
zq,pq,kq=signal.tf2zpk(Qbr,Qar)
#Plotting the responses
plt.figure(1),plt.subplot(2,2,1),plt.stem(n1,y),plt.xlabel('n-->'),plt.ylabel('h[n]'),plt.title('h[n]')
plt.subplot(2,2,2),plt.plot((W/np.pi),20*np.log10(np.abs(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('|H($\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(z),np.imag(z),'ro'),plt.plot(np.real(p),np.imag(p),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta))
plt.xlabel('Real part'),plt.ylabel('Imaginary part'),plt.title('Pole-zero plot')
plt.subplot(2,2,4),plt.plot((W/np.pi),np.unwrap(np.angle(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi(e^{jw})$')
plt.tight_layout()
plt.figure(2),plt.subplot(2,2,1),plt.stem(n1,yr),plt.xlabel('n-->'),plt.ylabel('$h_q[n]$')
plt.title('$h_q[n]$ with N = {} bits'.format(B))
plt.subplot(2,2,2),plt.plot((Wq/np.pi),20*np.log10(np.abs(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('$|H_q(\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(zq),np.imag(zq),'ro'),plt.plot(np.real(pq),np.imag(pq),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.xlabel('Real part'),plt.ylabel('Imaginary part'),
plt.title('Pole-zero plot'),plt.subplot(2,2,4),plt.plot((Wq/np.pi),np.unwrap(np.angle(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi_q(e^{jw})$')
plt.tight_layout()

Fig. 9.13 IIR filter coefficients quantization using rounding

9.3 Coefficient Quantization 391

Fig. 9.14 Simulation results of Experiment 9.7. (a) Infinite precision. (b) Finite precision (N = 2
bits). (c) Finite precision (N = 4 bits). (d) Finite precision (N = 8 bits). (e) Finite precision (N = 12
bits). (f) Finite precision (N = 16 bits)

Inferences
The following inferences can be made from Fig. 9.14:
1. When the number of bits allocated to represent each coefficient of the IIR filter is
2, the impulse response of the quantized filter becomes zero, and the pole-zero
plot, magnitude and phase responses confirm it.
2. The number of bits chosen as 4, 8 and 12, the impulse responses of the quantized
filter are not the same as the impulse response of the original filter, which reflects
in the pole-zero plot, magnitude and phase responses.
392 9 Quantization Effect of Digital Filter Coefficients

3. When the number of bits selected is 16, the impulse response of the quantized
filter is exactly the same as the impulse response of the original filter. Also, the
pole-zero plot, magnitude and phase responses are on par with the original result.
Task
Repeat the Experiment 9.7 with the filter quantization using two’s complement
truncation approach.
Experiment 9.8 Chebyshev Type I IIR Filter Coefficients Quantization Effect
Let us consider a Chebyshev Type I IIR filter designed using the bilinear transfor-
mation technique (BLT) that has a passband gain of 0 to -3 dB, cut-off frequency of
2 kHz and an attenuation of at least 20 dB for frequencies greater than 5 kHz.
Assume the sampling frequency to be 20 kHz. This experiment deals with the
quantization effect of IIR filter coefficients. The order and coefficients of Chebyshev
Type I filter are computed based on the given specifications. These filter coefficients
are quantized by either two’s complement or magnitude truncation method. The
performance of the infinite precision and finite precision is displayed. The python
code to quantize the IIR filter coefficients is given in Fig. 9.15, and the simulation
result is shown in Fig. 9.16.
When executing the python code given in Fig. 9.15, first enter the type of
quantization ‘1’ for two’s complement truncation and ‘2’ for magnitude truncation.
The simulation result is shown in Fig. 9.16.
Inference
From Fig. 9.16, it is possible to observe the quantization effect of IIR filter. The
impulse and magnitude responses are zero for the low-bit representation of filter
coefficients. When the high-bit representation of filter coefficients, the result of
impulse and magnitude responses approaches the original responses.
Task
1. Repeat Experiment 9.8 with the filter quantization using two’s complement
truncation approach.
2. Repeat Experiment 9.8 with the filter quantization using the rounding approach.

9.4 Limit Cycle Oscillations

A digital filter is a non-linear system affected by the quantization of the arithmetic

operations. This non-linearity of the digital filter may give stable output under
infinite precision arithmetic for specific input. Also, it may give unstable output
under finite precision arithmetic for specific input signals. This type of instability
usually results in an oscillatory periodic output called ‘limit cycle oscillation’. These
limit cycle oscillations do not have FIR filters because they do not have a feedback
path. But it will exist in IIR filters because it has a feedback path. The limit cycle
oscillation is broadly classified into (1) granular and (2) overflow.
9.4 Limit Cycle Oscillations 393

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
# Specifications of Filter
fsam, fp, fs, Ap, As, Td=20000, 2000, 5000, 3, 30, 0.9 # Sampling frequency
wp=2*np.pi*(fp/fsam) # pass band freq in radian per sample
ws=2*np.pi*(fs/fsam) # Stop band freq in radian per sample
# prewarping process
omega_p=(2/Td)*np.tan(wp/2)
omega_s=(2/Td)*np.tan(ws/2)
#Computation of order and normalized cut-off frequency
N, omega_c=signal.cheb1ord(omega_p,omega_s,Ap,As,analog=True)
print('Order of the Filter N =', N),print('Cut-off frequency= {:.4f} rad/s '. format(omega_c))
# Computation of H(s)
b, a=signal.cheby1(N, Ap, omega_c,'low', analog=True)
bz, az=signal.bilinear(b, a, Td)
n=15
n1=np.arange(0,n);
x=(n1==0)
y=signal.lfilter(bz,az,x)
W,H = signal.freqz(bz,az)
#Obtaining the pole-zero plot
z,p,k=signal.tf2zpk(bz,az)
B = 8;# Number of Bits
Q = 1/(2**(B))
QT=int(input('Enter the type of Qunatization:1-2s Complement truncation; 2-Magnitude truncation: '))
import sys
if (QT == 1):
Qbr=Q*np.floor(bz/Q)# Twos complement truncation
Qar=Q*np.floor(az/Q)# Twos complement truncation
elif (QT == 2):
Qbr=np.zeros(len(bz)) # Magnitude Truncation
Qar=np.zeros(len(az)) # Magnitude Truncation
for i in range(len(bz)):
if bz[i] > 0:
Qbr[i]=Q*np.floor(bz[i]/Q)
else: Qbr[i]=Q*np.ceil(bz[i]/Q)
for j in range(len(az)):
if az[j] > 0:
Qar[j]=Q*np.floor(az[j]/Q)
else: Qar[j]=Q*np.ceil(az[j]/Q)
else:
print('"Please select the proper quantization method"');
sys.exit()
yr=signal.lfilter(Qbr,Qar,x)
Wq,Hq = signal.freqz(Qbr,Qar)
zq,pq,kq=signal.tf2zpk(Qbr,Qar)
#Plotting the responses
plt.figure(1),plt.subplot(2,2,1),plt.stem(n1,y),plt.xlabel('n-->'),plt.ylabel('h[n]')
plt.title('h[n]')
plt.subplot(2,2,2),plt.plot((W/np.pi),20*np.log10(np.abs(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('|H($\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(z),np.imag(z),'ro')
plt.plot(np.real(p),np.imag(p),'kx')
theta=np.linspace(0,2*np.pi,100)

Fig. 9.15 Python code for IIR filter coefficients quantization

394 9 Quantization Effect of Digital Filter Coefficients

plt.plot(np.cos(theta),np.sin(theta))
plt.xlabel('Real part'),plt.ylabel('Imaginary part'),plt.title('Pole-zero plot')
plt.subplot(2,2,4),plt.plot((W/np.pi),np.unwrap(np.angle(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi(e^{jw})$')
plt.tight_layout()
plt.figure(2)
plt.subplot(2,2,1),plt.stem(n1,yr),plt.xlabel('n-->'),plt.ylabel('$h_q[n]$')
plt.title('$h_q[n]$ with N = {} bits'.format(B))
plt.subplot(2,2,2),plt.plot((Wq/np.pi),20*np.log10(np.abs(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('$|H_q(\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(zq),np.imag(zq),'ro')
plt.plot(np.real(pq),np.imag(pq),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.xlabel('Real part'),plt.ylabel('Imaginary part'),
plt.title('Pole-zero plot'),plt.subplot(2,2,4),plt.plot((Wq/np.pi),np.unwrap(np.angle(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi_q(e^{jw})$')
plt.tight_layout()

Fig. 9.15 (continued)

Experiment 9.9 Limit Cycle Oscillation in IIR Filter

This experiment tries to verify that the limit cycle oscillation can be occurred in IIR
filter due to coefficient quantization.
Let us consider a second-order recursive system which is given by

7 5
y ½ n] - y½n - 1] þ y½n - 2] = δ½n]
8 8

The input and output coefficients of the recursive system are quantized using the
rounding fixed-point 3-bit quantization approach. The quantized result of the recur-
sive system is written as

7 5
yq ½n] = Qr δ½n] þ y½n - 1] - y½n - 2]
8 8

The python code for this above equation is given in Fig. 9.17. From this figure, it
is possible to confirm that the number of bits used in the quantization is 3, and the
rounding quantization approach is implemented. After executing the python code
given in Fig. 9.17, the obtained result is shown in Fig. 9.18. From this figure, it is
possible to infer that the output is oscillated and confirms that the limit cycle
oscillation exists in this filter.
Inferences
From this experiment, the following inferences can be made:
1. The finite arithmetic operation in the digital implementation of a recursive system
may introduce limit cycle oscillations in the final output. This causes due to the
feedback that exists in the recursive system.
9.4 Limit Cycle Oscillations 395

Fig. 9.16 Simulation results. (a) Original filter. (b) Quantized with two’s complement. (c) Mag-
nitude truncation
396 9 Quantization Effect of Digital Filter Coefficients

import numpy as np
import matplotlib.pyplot as plt
B = 3;# Number of Bits
Q = 1/(2**(B))
N=100
x=np.zeros(N)
y1=np.zeros(N)
y=np.zeros(N)
n=np.arange(0,len(x))
x[0]=1#5/8
y1[-1], y[-1]=0,0
y1[-2], y[-2]=0,0
for i in range(len(x)):
Y=x[i]+((7/8)*y1[i-1])-(5/8)*y1[i-2]
y[i]=x[i]+((7/8)*y[i-1])-(5/8)*y[i-2]
y1[i]=(Q*np.round(Y/Q))
out=y1
plt.subplot(311),plt.stem(n,x),plt.xlabel('n-->'),plt.ylabel('x[n]'),plt.title('Input')
plt.subplot(312),plt.stem(n,y),plt.xlabel('n-->'),plt.ylabel('y[n]'),plt.title('Infinite precision Output')
plt.subplot(313),plt.stem(n,y1),plt.xlabel('n-->'),plt.ylabel('y1[n]'),plt.title('Finite arithmetic Output')
plt.tight_layout()

Fig. 9.17 Python code for limit cycle oscillation

Input
1
x[n]

0
0 20 40 60 80 100
n-->
Infinite precision Output
1
y[n]

0
0 20 40 60 80 100
n-->
Finite arithmetic Output
1
y1[n]

0
0 20 40 60 80 100
n-->

Fig. 9.18 Simulation result

2. After some time, the infinite precision arithmetic operation settles its output as
zero, whereas the finite precision result does not become zero.
3. Since the feedback path does not exist in the non-recursive system, the limit cycle
oscillation does not occur.
9.5 Cascade Form of a Higher Order Filters 397

Task
1. Execute the python code given in Fig. 9.17, and determine the minimum value
that will eliminate the limit cycle oscillation in Experiment 9.9 (i.e. change the
value of B (4, 5, 6, . . .)).

9.5 Cascade Form of a Higher Order Filters

The sensitivity reduction to coefficient quantization can be reduced by splitting high-

order filters into lower-order filters. This cascade form can split the higher-order
filters into multiple lower-order filters.
Experiment 9.10 Verify the Cascade Structure of FIR Filter May Reduce
the Quantization Effect of FIR Filter
This experiment explains the concept of the reduction of the quantization effect of
FIR filter using cascade structure implementation. The higher-order FIR filter is
decomposed into multiple second-order FIR filters, and those filter’s coefficients are
quantized with finite precision. The result of the multiple second-order filters are
combined to get a final output. The final output is always similar to the infinite
precision result of the higher-order FIR filter. The python code to verify this
experiment is given in Fig. 9.19, and the corresponding simulation result is displayed
in Fig. 9.20. From Fig. 9.19, the ‘signal.tf2sos’ python command is used here to
decompose the higher-order filter into multiple second-order filters.
The command ‘signal.sos2tf’ is used here to convert the multiple second-order
filters into higher-order ones.
Inferences
The following inferences can be made from Fig. 9.20:
1. The frequency response of the infinite precision FIR filter coefficients and finite
precision (N = 8 bits) FIR filter coefficients are not the same.
2. However, the frequency response of the cascade realization filter with finite
precision looks similar to the original one.
3. From this experiment, it is possible to confirm that the effect of quantization of
the FIR filter can be reduced with the help of cascade realization. In the cascade
realization, the higher-order filter is decomposed into multiple second-order
filters, and these lower-order filters are quantized with finite precision. Finally,
all these quantized filter coefficients are combined to get a higher-order filter.
4. Note that the number of bits (N = 8) used to represent the higher-order filter is the
same as for the lower-order filters.
5. This realization will help when the hardware is limited in the length of the
registers. For example, the hardware is an 8-bit register length, and to represent
the filter coefficients, it needs more than 8 bits. Then, the cascade realization will
help to represent all the filter coefficients with 8-bit precision.
398 9 Quantization Effect of Digital Filter Coefficients

from scipy import signal

import numpy as np
import matplotlib.pyplot as plt
n = 15
n1=np.arange(0,n);
a = signal.firwin(n, cutoff = 0.25, window = "hamming")
A = signal.tf2sos(a,1)#Decomposing higher order filter to second order filters
A1=A
B = 8;# Number of Bits
Q = 1/(2**(B))
Qhr=Q*np.round(a/Q)#Rounding
w,H=signal.freqz(a) #Obtaining the Frequency response
wq,Hq=signal.freqz(Qhr) #Obtaining the Frequency response
z,p,k=signal.tf2zpk(a,1) #Obtaining the pole-zero plot
zq,pq,kq=signal.tf2zpk(Qhr,1) #Obtaining the pole-zero plot
for i in range(len(A)):
A1[i][0:3]=Q*np.round(A[i][0:3]/Q)
cA=signal.sos2tf(A1)
cwq,cHq=signal.freqz(cA[0][0:n]) #Obtaining the Frequency response
czq,cpq,ckq=signal.tf2zpk(cA[0][0:n],1) #Obtaining the pole-zero plot
plt.figure(1),plt.subplot(2,2,1),plt.stem(a),plt.xlabel('n-->'),plt.ylabel('h[n]')
plt.title('h[n]'),plt.subplot(2,2,2),plt.plot((w/np.pi),20*np.log10(np.abs(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),plt.title('|H($\omega$)|')
plt.subplot(2,2,3),plt.plot(np.real(z),np.imag(z),'ro'),plt.plot(np.real(p),np.imag(p),'kx')
theta=np.linspace(0,2*np.pi,100)
plt.plot(np.cos(theta),np.sin(theta)),plt.xlabel('Real part'),plt.ylabel('Imaginary part'),
plt.title('Pole-zero plot'),plt.subplot(2,2,4),plt.plot((w/np.pi),np.unwrap(np.angle(H)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi(e^{jw})$')
plt.tight_layout()
plt.figure(2),plt.subplot(2,2,1),plt.stem(Qhr),plt.xlabel('n-->'),plt.ylabel('$h_q[n]$')
plt.title('$h_q[n]$ with N = {} bits'.format(B))
plt.subplot(2,2,2),plt.plot((wq/np.pi),20*np.log10(np.abs(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),
plt.title('$|H_q(\omega$)|'),plt.subplot(2,2,3),plt.plot(np.real(zq),np.imag(zq),'ro')
plt.plot(np.real(pq),np.imag(pq),'kx'),plt.plot(np.cos(theta),np.sin(theta))
plt.xlabel('Real part'),plt.ylabel('Imaginary part'),plt.title('Pole-zero plot')
plt.subplot(2,2,4),plt.plot((wq/np.pi),np.unwrap(np.angle(Hq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title('$\Phi_q(e^{jw})$')
plt.tight_layout()
plt.figure(3),plt.subplot(2,2,1),plt.stem(cA[0][0:n]),plt.xlabel('n-->'),plt.ylabel('$h_q[n]$')
plt.title('Cascade $h_q[n]$ with N = {} bits'.format(B))
plt.subplot(2,2,2),plt.plot((cwq/np.pi),20*np.log10(np.abs(cHq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Magnitude'),
plt.title(' Cascade $|H_q(\omega$)|'),plt.subplot(2,2,3),plt.plot(np.real(czq),np.imag(czq),'ro')
plt.plot(np.real(cpq),np.imag(cpq),'kx'),plt.plot(np.cos(theta),np.sin(theta))
plt.xlabel('Real part'),plt.ylabel('Imaginary part'),plt.title('Pole-zero plot')
plt.subplot(2,2,4),plt.plot((cwq/np.pi),np.unwrap(np.angle(cHq)))
plt.xlabel(r'$\omega$(x$\pi$rad/sample)'),plt.ylabel('Phase'),plt.title(' Cascade $\Phi_q(e^{jw})$')
plt.tight_layout()

Fig. 9.19 Python code for cascade realization of FIR filter

Exercises
1. Write a python code to convert decimal numbers to binary with fixed-point
representation.
 9.5
Cascade Form of a Higher Order Filters

Fig. 9.20 Simulation result. (a) Infinite precision output. (b) Finite precision output. (c) Cascade with finite precision output
399
400 9 Quantization Effect of Digital Filter Coefficients

2. Write a python code to convert binary representation to decimal representation.

3. Write a python code to verify how the saturation overflow stabilizes the arith-
metic operation in digital implementation.
4. Write a python code to plot the quantization characteristics curve of the rounding
approach.
5. Write a python code to plot the quantization characteristics curve of two’s
complement truncation approach.
6. Write a python code to plot the quantization characteristics curve of the truncation
approach.
7. Write a python code to verify that the limit cycle oscillation does not occur in FIR
filter coefficients quantization.
8. Write a python code to verify the magnitude truncation can inherently suppress
the limit cycle oscillation.
Objective Type Questions
1. Input to the python command is a floating-point number 1.45, and the output is
2. Identify the suitable python command.
A. numpy.round
B. numpy.floor
C. numpy.ceil
D. numpy.array
2. Identify the suitable python command whose input is a floating-point number of
1.45 and the output is 1.
A. numpy.floor
B. numpy.ceil
C. numpy.float
D. numpy.array
3. The input to the python command ‘np.ceil’ is 0.9, and the output is
A. 1.0
B. 0.5
C. 0.0
D. 0.9
4. The python command used to find the order of the Butterworth filter is
A. signal.buttord
B. signal.chebord
C. signal.butter
D. signal.bilinear
5. The python command used to obtain the coefficients of Butterworth filter is
A. signal.butter
B. signal.buttord
Bibliography 401

C. signal.butter1
D. signal.buttord1
6. The formula to quantize the input value ‘x’ and step size ‘Δ’ by rounding
approach is
A. Δ × round Δx
B. Δ × ceil Δx
C. Δ × floor Δx
D. x × round Δx
7. The formula to quantize the input value ‘x’ and step size ‘Δ’ by two’s compli-
ment truncation approach is
A. Δ × round Δx
B. Δ × ceil Δx
C. Δ × floor Δx
D. x × round Δx
8. The formula to quantize the input value ‘x’ is greater than ‘0’ and step size ‘Δ’
by magnitude truncation approach is
A. Δ × floor Δx
B. Δ × ceil Δx
C. Δ × round Δx
D. x × round Δx
9. Limit cycle oscillation does not occur in
A. Recursive system
B. IIR filter
C. Stable filter
D. FIR filter
10. The effect of quantization in a higher-order FIR filter is reduced by using
A. Parallel realization
B. Cascade realization
C. Direct form I realization
D. Lattice realization

Bibliography

1. Bernard Widrow, Istvan Kollar, “Quantization Noise: Roundoff Error in Digital Computation,
Signal Processing, Control and Communications”, Cambridge University Press, 2008.
402 9 Quantization Effect of Digital Filter Coefficients

2. N.S. Jayant, Peter Noll, “Digital Coding of Waveforms: Principles and Applications to Speech
and Video”, Prentice Hall India, 1984.
3. Lawrence R. Rabiner, and Bernard Gold, “Theory and Applications of Digital Signal
Processing”, Prentice-Hall, 1975.
4. James H. McClellan, Ronald W. Schafer, and Mark A. Yoder, “DSP First”, Prentice Hall, 1998.
5. Allen Gersho and Robert M. Gray, “Vector Quantization and Signal Compression”,
Springer, 1991.
Chapter 10
Multirate Signal Processing

Learning Objectives
After completing this chapter, the reader is expected to
• To perform sampling rate conversion using multirate operators.
• Time-domain and frequency-domain view of multirate operators.
• Demonstrate Type I and Type II polyphase decomposition.
• Signal decomposition using perfect reconstruction filter bank.
• Implementation of crosstalk free two-channel transmultiplexer.

Roadmap of the Chapter

The roadmap of this chapter is given below. From this figure, it is possible to observe
that this chapter begins with multirate operators. Downsampling and upsampling
operations are discussed in detail. Polyphase decomposition involving
downsampling operation is termed as Type I, and polyphase decomposition involv-
ing upsampling operation is termed as Type II polyphase decomposition. Subband
decomposition enables signals to be divided into different frequency regions.
Subband decomposition is done through a filter bank. In this chapter, two-channel,
three-channel and tree-structured filter banks are discussed. Finally, this chapter
concludes with the design of two-channel crosstalk free transmultiplexer.

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 403
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_10
404 10 Multirate Signal Processing

Multirate Signal Processing

Multirate Polyphase Signal Transmultiplexer

Operators Decomposition decomposition

Downsampling Type-I
Filter bank

Upsampling Type-II

2 Channel M Channel
Filter bank Filter bank

PreLab Questions
1. Mention the need to change the sampling rate of the signal.
2. What is a sampling rate converter?
3. Mention two basic operations in sampling rate conversion or multirate signal
processing.
4. What is the need for a filter before the downsampling operation? What is the
name of the filter?
5. What is the name given to the filter after upsampling operation? What is the
purpose of this filter?
6. Write the time-domain and frequency-domain expression for downsampling by
a factor of ‘M’.
7. Write the time-domain and frequency-domain expression for upsampling by a
factor of ‘L’.
8. Mention the three significant properties of the downsampling operation.
9. Why upsampling operation is considered as a linear time-variant operation?
10. What is the condition for interchanging of upsampling by a factor of ‘L’ and
downsampling by a factor of ‘M’ operation?
11. Why is downsampling by a factor of ‘M’ followed by upsampling by a factor of
‘M’ considered an idempotent operation?
12. What is the objective of polyphase decomposition? Mention the types of
polyphase decomposition.
13. What is a filter bank? Mention two applications of the filter bank.
14. Mention the threat involved in perfect reconstruction in a two-channel filter
bank? Mention the ways to overcome this threat.
15. What is a transmultiplexer? Mention its application.
10.1 Multirate Operators 405

10.1 Multirate Operators

Multirate operators are used to change the sampling rate of the signal digitally by
either by removing (deletion) of samples or inserting zeros between successive
samples. Two basic multirate operators are (1) downsampling operator and
(2) upsampling operator. The downsampling operation is used to decrease the
sampling rate of the signal, whereas upsampling operator is used to increase the
sampling rate of the signal.

10.1.1 Downsampling Operation

Downsampling operation reduces the sampling rate by a factor of ‘M’. The

downsampling operation by a factor of ‘M’ is shown in Fig. 10.1.
The time-domain expression for downsampling by a factor of ‘M’ is given by

y½n] = x½nM ] ð10:1Þ

From Eq. (10.1), it is possible to interpret that the output signal consists of every
Mth element of the input signal. The transform domain expression for downsampling
by a factor of ‘M’ is given by

M -1
1 1
Y ð zÞ = X zM W kM ð10:2Þ
M k=0

Downsampling produces expansion in the frequency-domain giving rise to

‘aliasing’. Aliasing will occur in the output signal if the input signal x[n] is not
bandlimited, which will lead to loss of information. In order to overcome the aliasing
problem, a filter is employed before downsampling operation, which is termed as
‘anti-aliasing filter’. The combination of downsampler with anti-aliasing filter is
termed as ‘decimator’. This concept is illustrated in Fig. 10.2. To maintain the
bandwidth of the input signal, the cut-off frequency of the filter (anti-aliasing)
H(z) is chosen as (π/M ) always, which is basically a low pass filter.

Fig. 10.1 Downsampling x[n] y[n]

operation M

Fig. 10.2 Decimation by a x[n] y[n]

factor of ‘M’ H(z) M
Fs Fs Fs/M
406 10 Multirate Signal Processing

#Illustration of downsampling operation

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of input signal x[n]
n=np.linspace(-10,10,21)
x=np.exp(1j*np.pi*n)
#Step 2: Downsampling the input signal
M=2
y=x[::M]
#Step 3: Plotting the results
n1=np.linspace(min(n)/M,max(n)/M,len(y))
plt.subplot(2,1,1),plt.stem(n,x),plt.xticks(range(-10,11))
plt.xlabel('n-->'),plt.ylabel('x[n]'),plt.title('Input Signal (x[n])')
plt.subplot(2,1,2),plt.stem(n1,y),plt.xticks(range(-10,11))
plt.xlabel('n-->'),plt.ylabel('y[n]'),
plt.title('Downsampled Signal (y[n]) by (M={})'.format(M))
plt.tight_layout()

Fig. 10.3 Python code which performs downsampling operation

Input Signal (x[n])

1

0
x[n]

–1
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->
Downsampled Signal (y[n]) by (M=2)
1.0
y[n]

0.5

0.0
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 10.4 Result of python code shown in Fig. 10.3

Experiment 10.1 Downsampling Operation in Time Domain

The objective of this experiment is to illustrate downsampling operation in time
domain. The signal to be downsampled is expressed as x[n] = ejπn, - 10 ≤ n ≤ 10.
The signal is to be downsampled by a factor of 2 to obtain the output signal y[n]. The
python code, which performs this task, is shown in Fig. 10.3, and the corresponding
result is shown in Fig. 10.4.
10.1 Multirate Operators 407

X[k] Y[k]
2

Fig. 10.5 Pictorial representation of problem statement

Inferences
From Fig. 10.3, it is possible to infer the following:
1. The input signal is x[n] = ejπn in the range -10 to 10. The input signal toggles
between +1 and -1. This is the highest frequency in the digital signal.
2. The input signal is downsampled by a factor of 2 to obtain the output signal y[n].
From Fig. 10.4, the following fact can be inferred:
1. The input signal is the highest frequency in digital signal. The signal toggles
between -1 and 1 and vice versa.
2. The signal x[n] is downsampled by a factor of 2 to obtain the output signal y[n],
which is a DC signal.
3. Downsampling operation has the ability to convert the highest frequency digital
signal to a DC signal.
4. The number of samples in the input signal is 21, whereas the number of samples
in the output signal is 11. Thus, the downsampling operation reduces the number
of samples in the input signal by a factor of ‘M’. In this case, the value of ‘M’ is 2.
Task
1. Change the value of the M = 4 in the python code given in Fig. 10.3, and observe
the result and comment on it.
Experiment 10.2 Spectrum of Downsampled Signal
The objective of this experiment is to obtain the spectrum of the downsampled
signal, and compare it with respect to the spectrum of the input signal. To accom-
plish this task, a sine wave of 5 Hz signal is generated. This signal is represented as x
[n], and its corresponding spectrum is X[k]. The signal x[n] is downsampled by a
factor of 2 to obtain the output signal y[n], and its corresponding spectrum is Y
[k]. The objective is to compare these two spectrums. The problem statement is
illustrated in Fig. 10.5.
The python code, which performs this task, is shown in Fig. 10.6, and the
corresponding output is shown in Fig. 10.7.
Inferences
From Fig. 10.6, the following inferences can be made:
1. The python code generates the sum of sine waves of 0, 2.5, 5 Hz frequency, and it
is stored in the variable ‘x’.
2. The input sine wave is downsampled by a factor of 2, and the result is stored in
the variable ‘y’.
408 10 Multirate Signal Processing

#Spectrum of downsampled signal

import numpy as np
import matplotlib.pyplot as plt
from scipy.fftpack import fft,fftfreq
#Step 1: Generating the input signal
f1,f2,fs,N=2.5,5,75,512
M,T=2,1/fs
t=np.linspace(0,N*T,N)
x1=0.5*np.ones(len(t))# DC signal
x=x1+np.sin(2*np.pi*f1*t)+np.sin(2*np.pi*f2*t)#
#Step 2: Obtaining the downsampled signal
y=x[::M]
t1=np.linspace(0,N*T/M,len(y))
#Step 3: Obtaining the spectrum of the input signal and downsampled signal
X=fft(x,N)
Y=fft(y,N)
f_axis=fftfreq(N,T)[0:N//M]
#Step4: Plotting the result
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Input signal'),plt.subplot(2,2,2),plt.plot(t1,y),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Downsampled signal')
plt.subplot(2,2,3),plt.plot(f_axis,2/N*np.abs(X[0:N//M])),plt.xlabel('$\omega$-->'),
plt.ylabel('|X($j\omega$)|'),plt.title('Spectrum of input signal')
plt.subplot(2,2,4),plt.plot(f_axis,2/N*np.abs(Y[0:N//M])),plt.xlabel('$\omega$-->'),
plt.ylabel('|Y($j\omega$)|'),plt.title('Spectrum of downsampled signal')
plt.tight_layout()

Fig. 10.6 Python code to plot the spectrum of downsampled signal

3. Using the built-in function ‘fft’ and ‘fftfreq’ in ‘scipy’ library, the spectrum of the
input signal ‘x’ and the output signal ‘y’ is obtained and stored in the variable ‘X’
and ‘Y’, respectively.
4. The magnitude spectrum of the input and output signal is obtained using the built-
in function ‘abs’, which is available in ‘numpy’ library.
From Fig. 10.7, the following inferences can be drawn:
1. From the plot of the input signal and output signal, it is possible to observe that
the length of the output signal is lesser than the length of the input signal.
2. The magnitude spectrum of the input signal has peaks at 0, 2.5 and 5 Hz, which
shows that the frequencies of the input signal are 0, 2.5 and 5 Hz.
3. The magnitude spectrum of the output signal has peaks at 0, 5 and 10 Hz, which
shows that the frequencies of the output signal are 0, 5 and 10 Hz. That is, the
bandwidth of the downsampled spectrum increased by 2 because of the
downsampling factor chosen as 2.
4. This experiment reveals that compression in the time-domain is equivalent to
expansion in the frequency domain.
10.1 Multirate Operators 409

Input signal Downsampled signal

2 2

Amplitude
Amplitude

1 1
0 0
–1 –1

0 2 4 6 0 1 2 3
Time Time

Spectrum of input signal Spectrum of downsampled signal

1.0
0.4
|X(jw)|

|Y(jw)|
0.5
0.2

0.0 0.0
0 10 20 30 0 10 20 30
w --> w -->

Fig. 10.7 Simulation result

Fig. 10.8 Upsampling x[n] y[n]

operation L

Task
1. Change the value of the M = 4 in the python code given in Fig. 10.6, and observe
the result and comment on it.

10.1.2 Upsampling Operation

Upsampling operation increases the sampling rate by a factor of ‘L’. The upsampling
operation by a factor of ‘L’ is shown in Fig. 10.8.
The time-domain expression for upsampling by a factor of ‘L’ is given by

n
x , n = L, 2L, . . .
y ½ n] = L ð10:3Þ
0, otherwise

From Eq. (10.3), it is possible to interpret that upsampling by a factor of ‘L’ in the
time-domain is accomplished by inserting ‘L - 1’ zeros between successive samples
of the input signal x[n]. This will increase the length of the input signal; hence,
410 10 Multirate Signal Processing

Fig. 10.9 Interpolation by a x[n] y[n]

factor of ‘L’ L H(z)
Fs LFs LFs

#Illustration of upsampling operation

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of input signal x[n]
N=5
n=np.arange(N)
x=np.ones(N)
#Step 2: Upsampling the input signal
L=2
y=np.zeros(L*N)
y[::L]=x
# #Step 3: Plotting the results
n1=np.arange(L*N)
plt.subplot(2,1,1),plt.stem(n,x),plt.xticks(range(0,N))
plt.xlabel('n-->'),plt.ylabel('x[n]'),plt.title('Input Signal (x[n])')
plt.subplot(2,1,2),plt.stem(n1,y),plt.xticks(range(0,L*N+1))
plt.xlabel('n-->'),plt.ylabel('y[n]'),
plt.title('Upsampled Signal (y[n]) by L={}'.format(L))
plt.tight_layout()

Fig. 10.10 Python code to perform upsampling operation

upsampling operation can also be termed as ‘expansion operation’. The frequency-

domain expression for upsampling by a factor of ‘L’ is given by

Y ð zÞ = X zL ð10:4Þ

The above equation can be expressed as

Y ejω = X ejωL ð10:5Þ

The upsampler introduces spectral images. A filter is employed after the

upsampler to remove the spectral images. Such type of filter is termed as ‘anti-
imaging’ filter. This is shown in Fig. 10.9.
The cut-off frequency of the filter H(z) is chosen as π/L, which is basically a
lowpass filter.
Experiment 10.3 Upsampling Operation in the Time Domain
In this experiment, the input signal x[n] is upsampled by a factor of 2 to obtain the
output signal y[n]. The python code which performs this task is shown in Fig. 10.10,
and the corresponding output is shown in Fig. 10.11.
10.1 Multirate Operators 411

Input Signal (x[n])

1.0
x[n]

0.5

0.0
0 1 2 3 4
n-->
Upsampled Signal (y[n]) by L=2
1.0
y[n]

0.5

0.0
0 1 2 3 4 5 6 7 8 9 10
n-->

Fig. 10.11 Upsampling operation result

X[k] Y[k]
L

Fig. 10.12 Pictorial representation of problem statement

Inferences
The following inferences can be made from this experiment:
1. From Fig. 10.11, it is possible to observe that the length of the input signal x[n] is
5, whereas the length of the output signal y[n] is 10.
2. Thus, upsampling by a factor of 2 (L ) inserts one zero (L - 1) between two
successive samples of the input signal x[n].
3. Therefore, upsampling is a ‘length stretching operation’.
Experiment 10.4 Spectrum of Upsampled Signal
The objective of this experiment is to obtain the spectrum of upsampled signal and to
compare it with the spectrum of input signal. In order to accomplish this task, the
following steps are carried out:
• Generate input signal sum of sine waves with the frequency of 0, 4 and 10 Hz.
• Pass this signal through a system that upsamples the input signal by a factor (L ) of
2 to obtain the output signal.
• Plot the spectrums of the input and output signal, and comment on the observed
result.
The pictorial representation of the problem statement is shown in Fig. 10.12.
From Fig. 10.12, X[k] represents the spectrum of the input signal, and Y[k] represents
the spectrum of the output signal.
412 10 Multirate Signal Processing

#Spectrum of upsampled signal

import numpy as np
import matplotlib.pyplot as plt
from scipy.fftpack import fft,fftfreq
#Step 1: Generating the input signal
f1,f2,fs,N=4,10,50,512
T=1/fs
t=np.linspace(0,N*T,N)
x1=0.5*np.ones(len(t))# DC signal
x=x1+np.sin(2*np.pi*f1*t)+np.sin(2*np.pi*f2*t)#
#x=np.sin(2*np.pi*f*t)
#Step 2: Obtaining the upsampled signal
n = len(x)
L = 2 # Upsample_factor
y = np.zeros(L*n-(L-1))
y[::L] = x
t1=np.linspace(0,N*T*L,len(y))
#Step 3: Obtaining the spectrum of input and downsampled signal
X=fft(x,N)
Y=fft(y,N)
f_axis=fftfreq(N,T)[0:N//(2*L)]
#Step4: Plotting the result
plt.subplot(2,2,1),plt.plot(t,x),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Input signal'),plt.subplot(2,2,2),plt.plot(t1,y),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Upsampled signal by L={}'.format(L))
plt.subplot(2,2,3),plt.plot(f_axis,2/N*np.abs(X[0:N//(2*L)])),plt.xlabel('$\omega$-->'),
plt.ylabel('|X($j\omega$)|'),plt.title('Spectrum of input signal')
plt.subplot(2,2,4),plt.plot(f_axis,2/N*np.abs(Y[0:N//(2*L)])),plt.xlabel('$\omega$-->'),
plt.ylabel('|Y($j\omega$)|'),plt.title('Spectrum of upsampled signal by L={}'.format(L))
plt.tight_layout()

Fig. 10.13 Python code to plot the spectrum of upsampled signal

The python code, which plots the spectrum of the input and upsampled signals, is
shown in Fig. 10.13, and the corresponding output is shown in Fig. 10.14.
Inferences
From Fig. 10.14, the following inferences can be drawn:
1. From the plot of the input signal and output signal, it is possible to observe that
the length of the output signal is more than the length of the input signal.
2. The magnitude spectrum of input signal has peaks at 0, 4 and 10 Hz, which shows
that the frequencies of the input signal are 0, 4 and 10 Hz.
3. The magnitude spectrum of the output signal has peaks at 0, 2 and 5 Hz, which
shows that the frequencies of the output signal are 0, 2 and 5 Hz. That is, the
bandwidth of the upsampled spectrum is decreased by 2 because of upsampling
factor chosen as 2.
10.2 Noble Identity 413

Input signal Upsampled signal by L=2

2 2
Amplitude

Amplitude
0 0

0.0 2.5 5.0 7.5 10.0 0 5 10 15 20

Time Time
Spectrum of input signal Spectrum of upsampled signal by L=2
1.0
0.4
|X(jw)|

|Y(jw)|
0.5
0.2

0.0 0.0
0 5 10 0 5 10
w --> w -->

Fig. 10.14 Spectrum of input and upsampled signal

x[n] v1[n] y1[n] x[n] v2[n] y2[n]

M H(z) H(zM) M

Fig. 10.15 Noble identity for downsampling operation

4. This experiment reveals the fact that expansion in the time-domain is equivalent
to compression in the frequency domain.
Task
1. Change the value of the L = 4 in the python code given in Fig. 10.13, and
comment on the observed spectrum result.

10.2 Noble Identity

Noble identities describe the way to reverse the order of multirate operators and
filtering.

10.2.1 Noble Identity for Downsampling Operation

The noble identity of the downsampling operation is depicted in Fig. 10.15.

414 10 Multirate Signal Processing

x[n] v1[n] y1[n] x[n] v2[n] y2[n]

2 H(z) H(z2) 2

Fig. 10.16 Noble identity for downsampling

#Proof of noble identity for downsampling

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Define the signal x[n]
n=np.arange(-10,11)
x=np.ones(len(n))
M=2
#Function to perform downsampling and upsampling operation
def downsample(x,M):
y=x[::M]
return(y)
def upsample(x,L):
n=len(x)
y=np.zeros(n*L)
y[::L]=x
return(y)
#Obtaining the signal v1[n]
v1=downsample(x,M) #Downsampling of x[n] by a factor of two
h=signal.firwin(5,0.5) #Defining the filter
y1=signal.lfilter(h,1,v1);#Obtaining the signal y1[n]
h1=upsample(h,2);#Equivalent to H(z^2) in time domain
v2=signal.lfilter(h1,1,x)#Obtaining the signal v2[n]
y2=downsample(v2,M)#Obtaining the signal y2[n]
#Plotting the results
plt.subplot(2,1,1),plt.stem(np.abs(y1))
plt.xlabel('n-->'),plt.ylabel('$y_1[n]$'),plt.title('$y_1[n]$')
plt.subplot(2,1,2),plt.stem(np.abs(y2))
plt.xlabel('n-->'),plt.ylabel('$y_2[n]$'),plt.title('$y_2[n]$')
plt.tight_layout()

Fig. 10.17 Python illustration regarding noble identity for downsampling operation

Experiment 10.5 Python Illustration of Noble Identity for Downsampling

Operation
The block diagram for the noble identity for downsampling operation considered for
python illustration is shown in Fig. 10.16.
In this experiment, downsampling factor is chosen as 2. In Fig. 10.16, H(z2) in
time-domain represents upsampling of the filter coefficient h[n] by a factor of 2. The
python code, which illustrates the noble identity for downsampling operation, is
shown in Fig. 10.17, and the corresponding output is shown in Fig. 10.18.
10.2 Noble Identity 415

y1[n]
1.0
y1[n]

0.5

0.0
0 2 4 6 8 10
n-->
y2[n]
1.0
y2[n]

0.5

0.0
0 2 4 6 8 10
n-->

Fig. 10.18 Result of python code shown in Fig. 10.17

x[n] v1[n] y1[n] x[n] v2[n] y2[n]

L H(zL) H(z) L

Fig. 10.19 Noble identity for upsampling operation

Inference
From Fig. 10.18, it is possible to observe that the output y1[n] is equal to the output
y2[n]; thus, the noble identity for the downsampling operation is verified.

10.2.2 Noble Identity for Upsampling Operation

The noble identity for upsampling operation is shown in Fig. 10.19

Experiment 10.6 Python Illustration of Noble Identity for Upsampling
Operation
This experiment attempts to prove the noble identity for upsampling operation for
L = 2. The python code which performs this task is shown in Fig. 10.20, and the
corresponding output is shown in Fig. 10.21.
Inference
Figure 10.21 shows that the signals y1[n] and y2[n] are identical, which means that
the noble identity for upsampling operation has been verified for L = 2.
416 10 Multirate Signal Processing

#Noble identity for upsampling operation

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Define the signal x[n]
n=np.arange(-10,11)
x=np.ones(len(n))
L=2
#Step 2: Function to perform upsampling operation
def upsample(x,L):
n=len(x)
y=np.zeros(n*L)
y[::L]=x
return(y)
v1=upsample(x,L)#Step 3: Obtaining the signal v1[n]
h=signal.firwin(5,0.5)
h1=upsample(h,L)
y1=signal.lfilter(h1,1,v1)#Step 4: Obtaining the signal y1[n]
v2=signal.lfilter(h,1,x)#Step 5: Obtaining the signal v2[n]
y2=upsample(v2,L)#Step 6: Obtaining the signal y2[n]
#Step 7: Plotting the result
plt.subplot(2,1,1),plt.stem(y1)
plt.xlabel('n-->'),plt.ylabel('$y_1[n]$'),plt.title('$y_1[n]$')
plt.subplot(2,1,2),plt.stem(y2)
plt.xlabel('n-->'),plt.ylabel('$y_2[n]$'),plt.title('$y_2[n]$')
plt.tight_layout()

Fig. 10.20 Python code to illustrate noble identity for upsampling operation

y1[n]
1.0
y1[n]

0.5

0.0
0 10 20 30 40
n-->
y2[n]
1.0
y2[n]

0.5

0.0
0 10 20 30 40
n-->

Fig. 10.21 Result of python code shown in Fig. 10.20

10.3 Polyphase Decomposition 417

10.3 Polyphase Decomposition

Polyphase decomposition refers to the strategy through which the multirate operators
can be used to decompose the system function H(z) into its polyphase representation.
Polyphase decomposition can be broadly classified into (1) Type I polyphase
decomposition and (2) Type II polyphase decomposition.
Experiment 10.7 Python Illustration of Type I Polyphase Decomposition
This python illustration aims to prove the Type I polyphase decomposition illus-
trated in Fig. 10.22 for the downsampling factor of M = 2. In Fig. 10.22, H(z)
represents the filter, whereas E0(z) and E1(z) represent the polyphase components of
H(z). The objective is to prove y1[n] is equal to y2[n]. The filter chosen in this
illustration is a finite impulse response filter designed using the windowing
technique.
The python code which implements the Type I polyphase decomposition is
shown in Fig. 10.23, and the corresponding output is shown in Fig. 10.24.
Inference
From Fig. 10.24, it is possible to observe that the output y1[n] is equal to the output
y2[n]. Thus, the Type I polyphase decomposition is verified.
Experiment 10.8 Type II Polyphase Decomposition
Type II polyphase decomposition deals with the upsampling operation. Upsampling
operation introduces multiple copies of the original signal spectrum, which is termed
as ‘imaging’. The Type II polyphase decomposition structure is shown in Fig. 10.25.
The python illustration of Type II polyphase decomposition is shown in
Fig. 10.26, and the corresponding output is shown in Fig. 10.27.
Inference
From Fig. 10.27, it is possible to observe that the output y1[n] is equal to y2[n]. This
implies that Type II polyphase decomposition is verified.

x[n] v1[n] w1[n] y2[n]

x[n] u[n] y1[n] 2 E0(z)
H(z) 2 =
z-1
v2[n] w2[n]
2 E1(z)

Fig. 10.22 Type I polyphase decomposition

418 10 Multirate Signal Processing

#Type-I Polyphase decomposition

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Defining the input signal
x=np.ones(8)
h=signal.firwin(8,0.5)
u=signal.lfilter(h,1,x)
y1=u[::2]
#Step 2 Polyphase decomposition
e0=h[0::2] # Obtaining E0(z)
e1=h[1::2] # Obtaining E1(z)
x1=np.zeros(len(x)+1)
x1[1:]=x
v1=x[::2]
v2=x1[::2]
w1=signal.lfilter(e0,1,v1)
w2=signal.lfilter(e1,1,v2)
y2=w1+w2[0:len(w1)]
#Step 3: Plotting the result
plt.subplot(2,1,1),plt.stem(y1)
plt.xlabel('n-->'),plt.ylabel('$y_1[n]$'),plt.title('$y_1[n]$')
plt.subplot(2,1,2),plt.stem(y2)
plt.xlabel('n-->'),plt.ylabel('$y_2[n]$'),plt.title('$y_2[n]$')
plt.tight_layout()

Fig. 10.23 Python code to illustrate Type I polyphase decomposition

y1[n]
1.0
y1[n]

0.5

0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
n-->
y2[n]
1.0
y2[n]

0.5

0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
n-->

Fig. 10.24 Result of python code shown in Fig. 10.23

10.4 Filter Bank 419

x[n]
R0(z) 2
x[n] y1[n] z-1
2 H(z) ≡

R1(z) 2 y2[n]

Fig. 10.25 Type II polyphase decomposition structure

#Type-II polyphase decomposition

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Defining the input signal
x=np.ones(8)
h=signal.firwin(8,0.5)
#Step 2: Polyphase components of H(z)
r0=h[1::2] # Obtaining R0(z)
r1=h[0::2] # Obtaining R1(z)
#Step 3: Obtaining the output y1[n]
u=np.zeros(2*len(x))
u[::2]=x
y1=signal.lfilter(h,1,u)
#Step 3: Obtaining the output y2[n]
v1=signal.lfilter(r0,1,x)
v2=signal.lfilter(r1,1,x)
w1=np.zeros(2*len(v1))
w1[::2]=v1
w2=np.zeros(2*len(v2))
w2[::2]=v2
w11=np.zeros(len(w1)+1)
w11[1:]=w1
y2=w2+w11[0:len(w2)]
#Step 4: Plotting the output signals
plt.subplot(2,1,1),plt.stem(y1),plt.xlabel('n-->'),plt.ylabel('$y_1[n]$'),
plt.title('$y_1[n]$'),plt.subplot(2,1,2),plt.stem(y2)
plt.xlabel('n-->'),plt.ylabel('$y_2[n]$'),plt.title('$y_2[n]$')
plt.tight_layout()

Fig. 10.26 Python code to demonstrate Type II polyphase decomposition

10.4 Filter Bank

Filter bank is group of filters arranged in a specific fashion. Filter bank is used to split
the signal into different frequency bands, which are termed as ‘subband coding’.
While splitting the signal into various frequency bands, the signal characteristics are
420 10 Multirate Signal Processing

y1[n]

y1[n] 0.50

0.25

0.00
0 2 4 6 8 10 12 14
n-->
y2[n]
0.50
y2[n]

0.25

0.00
0 2 4 6 8 10 12 14
n-->

Fig. 10.27 Output of python code shown in Fig. 10.26

H0(z) 2 2 G0(z)
y[n]
x[n]

H1(z) 2 2 G1(z)

Analysis section Synthesis section

Fig. 10.28 Structure of two channel filter bank

different in each band, and different bits can be used for coding the subband signals.
This idea is used in speech and image coding. Based on the number of paths
available for the input signal, the filter bank can be broadly classified into (1) two-
channel filter bank and (2) M-channel filter bank.

10.4.1 Two-Channel Filter Bank

Two-channel filter bank has two sections, namely, (1) analysis section and (2) syn-
thesis section, which is depicted in Fig. 10.28. The input signal fed into the two-
channel filter bank is x[n], and the output signal received from the two-channel filter
bank is y[n]. The channel represents the medium through which the data is trans-
mitted. In Fig. 10.28, the filters in the analysis section are represented as H0(z) and
H1(z). If H0(z) represents the lowpass filter, then H1(z) represents the highpass filter.
The corresponding filters in the synthesis section are G0(z) and G1(z), respectively.
10.4 Filter Bank 421

H0(z) 1 3 5 7
2 2 G0(z)
y[n]
x[n]
9
2 4 6 G1(z) 8
H1(z) 2 2

Fig. 10.29 Two-channel filter bank

For perfect reconstruction, the output signal y[n] has to be the delayed version of
the input signal x[n]. The different threats for perfect reconstruction are (1) aliasing
problem due to downsampling operation, (2) amplitude distortion and (3) phase
distortion. Proper choice of analysis and synthesis filters will overcome the above-
mentioned threats and achieves perfect reconstruction.
Experiment 10.9 Python Implementation of Two-Channel Filter Bank
The structure of the two-channel filter bank which is implemented in this experiment
is shown in Fig. 10.29. In Fig. 10.29, different nodes are marked as 1–8.
The input signal x[n] is a sinusoidal signal of 5 Hz frequency. The choice of
analysis and synthesis filters are H 0 ðzÞ = 12 þ 12 z - 1 , H 1 ðzÞ = 12 - 12 z - 1 ,
G0(z) = 1 + z-1 and G1(z) = - 1 + z-1. The python code which implements this
two-channel filter bank is given in Fig. 10.30, and the corresponding output is shown
in Fig. 10.31.
Inferences
1. From Fig. 10.30, it is possible to observe that the filters chosen for the analysis
section are h0 ½n] = 12 , 12 and h1 ½n] = 12 , - 12 . The filters have only two
coefficients. h0[n] act as lowpass filter, whereas h1[n] act as high pass filter.
2. The variables chosen in the python code, as shown in Fig. 10.30 like x1, x2, . . .,
x8, are in line with the nodes shown in Fig. 10.29.
3. From Fig. 10.31, it is possible to observe that the output signal follows the input
signal with one sample delay. That is, perfect reconstruction is achieved through
the proper choice of filters.

10.4.2 Relationship Between Analysis and Synthesis Filters

Let the analysis filter be expressed as

H ðzÞ = H 0 ðzÞ ð10:6Þ

The expression for H1(z) in terms of H(z) is given by

422 10 Multirate Signal Processing

#Two-channel filter bank

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Functions to perform downsampling and upsampling
def downsample(x,M):
y=x[::M]
return(y)
def upsample(x,L):
y=np.zeros(L*len(x))
y[::L]=x
return(y)
#Step 1: Define the filters
h0=np.array([0.5,0.5])
h1=np.array([0.5,-0.5])
g0,g1=2*h0,-2*h1
#Step 2: Generate the input signal
f,fs,N=5,100,256
T=1/fs;
t=np.linspace(0,N*T,N)
x=np.sin(2*np.pi*f*t)
#Step 3: Traversing the path
x1=signal.lfilter(h0,1,x)
x2=signal.lfilter(h1,1,x)
x3=downsample(x1,2)
x4=downsample(x2,2)
x5=upsample(x3,2)
x6=upsample(x4,2)
x7=signal.lfilter(g0,1,x5)
x8=signal.lfilter(g1,1,x6)
y=x7+x8
plt.plot(t,x,'b',t,y,'r--',linewidth=2),plt.legend(['Input','Output'],loc=4)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Input and Output signals')
plt.tight_layout()

Fig. 10.30 Python code for two-channel filter bank

H 1 ðzÞ = H ð- zÞ ð10:7Þ

The synthesis filters are expressed as

G0 ðzÞ = 2H ðzÞ ð10:8Þ

G1 ðzÞ = - 2H ð- zÞ ð10:9Þ

From Eqs. (10.6) to (10.9), it is possible to infer the following

10.4 Filter Bank 423

Input and output signals

1.00

0.75

0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75 Input
–1.00 Output

0.0 0.5 1.0 1.5 2.0 2.5

Time

Fig. 10.31 Plot of input and output signals of a two-channel filter bank

• Instead of designing four filters (2 for analysis and 2 for synthesis). It is sufficient
to design one prototype filter H(z). All the other filters are obtained as a modified
version of the prototype filter.
• If H(z) acts as lowpass filter, then H(-z) acts as highpass filter and vice versa.
Experiment 10.10 Relationship Between Analysis and Synthesis Filters
This experiment tries to obtain the magnitude response of the analysis and synthesis
filters. The python code to obtain the magnitude response of the analysis and
synthesis filters is given in Fig. 10.32, and the corresponding output is shown in
Fig. 10.33.
Inferences
From the magnitude response, the following inferences can be drawn:
1. The analysis filter h0[n] behaves like a lowpass filter.
2. The analysis filter h1[n] behaves like a highpass filter.
3. The synthesis filter g0[n] behaves like a lowpass filter.
4. The synthesis filter g1[n] behaves like a highpass filter.
5. The filters h0[n] and h1[n] are complementary to each other.
6. The filters g0[n] and g1[n] are complementary to each other.
Experiment 10.11 Phase Responses of Analysis and Synthesis Filters
This experiment discusses the phase responses of analysis and synthesis filters of
two-channel filter bank. The python code to obtain the phase response of the analysis
and the synthesis filters are given in Fig. 10.34, and the corresponding output is
shown in Fig. 10.35.
424 10 Multirate Signal Processing

#Magnitude response of the analysis and synthesis filters

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
h0=[0.5,0.5]
h0=np.array(h0)
h1=np.array([0.5,-0.5])
g0,g1=2*h0,-2*h1
#Frequency response of four filters
w0, H0 = signal.freqz(h0,1)
w1, H1 = signal.freqz(h1,1)
w2, H2 = signal.freqz(g0,1)
w3, H3 = signal.freqz(g1,1)
#Plotting the result
plt.subplot(2,2,1),plt.plot(w0, 10 * np.log10(abs(H0)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Magnitude(dB)')
plt.title('$|H_0(e^{j\omega})|$')
plt.subplot(2,2,2),plt.plot(w1, 10 * np.log10(abs(H1)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Magnitude(dB)')
plt.title('$|H_1(e^{j\omega})|$')
plt.subplot(2,2,3),plt.plot(w2, 10 * np.log10(abs(H2)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Magnitude(dB)')
plt.title('$|G_0(e^{j\omega})|$')
plt.subplot(2,2,4),plt.plot(w3, 10 * np.log10(abs(H3)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Magnitude(dB)')
plt.title('$|G_1(e^{j\omega})|$')
plt.tight_layout()

Fig. 10.32 Python code to obtain the magnitude response of analysis and synthesis filters

Inferences
1. From the phase response of the analysis and synthesis filters, it is possible to infer
that the filter exhibits linear phase characteristics in the pass band.
2. Because of linear phase characteristics, phase distortion can be avoided.
3. The filters will exhibit constant group delay.

10.4.3 Two-Channel Filter Bank Without Filters

This is a special case of two-channel filter bank in which delay is introduced instead
of filters. For example, if H0(z) = 1, H1(z) = z-1, G0(z) = z-1 and G1(z) = 1, then the
structure of two-channel filter bank is modified as in Fig. 10.36.
The relationship between the input and output in the frequency-domain is
expressed as
10.4 Filter Bank 425

|H0(ejw)| |H1(ejw)|
0 0
Magnitude(dB)

Magnitude(dB)
–10 –10

–20 –20

0 1 2 3 0 1 2 3
w [rad/sample] w [rad/sample]

|G0(ejw)| |G1(ejw)|
Magnitude(dB)

Magnitude(dB)
0 0

–10 –10

–20 –20
0 1 2 3 0 1 2 3
w [rad/sample] w [rad/sample]

Fig. 10.33 Magnitude responses of analysis and synthesis filters

Y ðzÞ = z - 1 X ðzÞ ð10:10Þ

Upon taking inverse Z-transform, the relationship between the input and output is
given by

y ½ n] = x ½ n - 1] ð10:11Þ

Experiment 10.12 Python Illustration of Two-Channel Filter Bank Without

Filters
This experiment deals with the illustration of two-channel filter bank without filters.
The python code, which depicts filter bank without filters is shown in Fig. 10.37, and
the corresponding output is shown in Fig. 10.38.
Inference
From Fig. 10.38, it is possible to observe that the output signal is a delayed version of
the input signal. Hence, perfect reconstruction is achieved. There is one sample delay
between the input and output signal.
426 10 Multirate Signal Processing

#Phase response of analysis and synthesis filters

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
h0=[0.5,0.5]
h0=np.array(h0)
h1=np.array([0.5,-0.5])
g0=2*h0
g1=-2*h1
#Frequency response of four filters
w0, H0 = signal.freqz(h0,1)
w1, H1 = signal.freqz(h1,1)
w2, H2 = signal.freqz(g0,1)
w3, H3 = signal.freqz(g1,1)
#Plotting the result
plt.subplot(2,2,1),plt.plot(w0,np.unwrap(np.angle(H0)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Degree')
plt.title('$\Phi_{h0}{(e^{j\omega})}$')
plt.subplot(2,2,2),plt.plot(w1,np.unwrap(np.angle(H1)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Degree')
plt.title('$\Phi_{h1}{(e^{j\omega})}$')
plt.subplot(2,2,3),plt.plot(w2,np.unwrap(np.angle(H2)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Degree')
plt.title('$\Phi_{g0}{(e^{j\omega})}$')
plt.subplot(2,2,4),plt.plot(w3,np.unwrap(np.angle(H3)))
plt.xlabel('$\omega$ [rad/sample]'),plt.ylabel('Degree')
plt.title('$\Phi_{g1}{(e^{j\omega})}$')
plt.tight_layout()

Fig. 10.34 Python code to obtain the phase response of analysis and synthesis filters

10.4.4 Three-Channel Filter Bank Without Filters

The structure of three-channel filter bank is shown in Fig. 10.39. From this figure, it
is possible to infer that there are three channels and each channel contains an analysis
and synthesis filters.
If H0(z) = 1, H1(z) = z-1 and H2(z) = z-2, G0(z) = z-2, G1(z) = z-1 and
G2(z) = 1, the structure of three-channel filter bank for this choice of filters is
given in Fig. 10.40.
The frequency-domain relationship between the input and output signal is given
by

Y ðzÞ = z - 2 X ðzÞ ð10:12Þ

10.4 Filter Bank 427

Фh0(ejw) Фh1(ejw)
0
Degree

Degree
1
–1

0
0 1 2 3 0 1 2 3
w [rad/sample] w [rad/sample]

Фg0(ejw) Фg1(ejw)
0 0

Degree
Degree

–1 –2

0 1 2 3 0 1 2 3
w [rad/sample] w [rad/sample]

Fig. 10.35 Phase response of the analysis and synthesis filters

1 3 5 7
1 2 2 z-1

x[n] y[n]
2 4 6 8
z-1 2 2 1

Fig. 10.36 Filter bank without filters

Upon taking inverse Z-transform, the time-domain relationship between the input
and output signal is given by

y ½ n] = x ½ n - 2] ð10:13Þ

Experiment 10.13 Illustration of Three-Channel Filter Bank Without Filters

This experiment implements three-channel filter bank without filters using python,
and the python code is shown in Fig. 10.41, and the corresponding output is shown
in Fig. 10.42.
Inference
1. From Fig. 10.42, it is possible to observe that the output signal is a delayed
version of the input signal.
428 10 Multirate Signal Processing

#Filterbank without filters

import numpy as np
import matplotlib.pyplot as plt
#Step 1: Generation of sine wave
t=np.linspace(0,1,200)
x=np.sin(2*np.pi*5*t)
#Downsample x
x1=x[::2]
#Upsample x1
x2=np.zeros(2*len(x1))
x2[::2]=x1
#Introduce a delay to get x3
delay=1
x3=np.zeros(len(x)+delay)
x3[delay:]=x
x4=x3[::2]
x5=np.zeros(2*len(x4))
x5[::2]=x4
x6=np.zeros(len(x2)+delay)
x6[delay:]=x2
y=x6+x5[0:len(x6)]
plt.plot(t,x,t,y[0:len(t)]),plt.legend(["Input", "Output"], loc ="upper right"),
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Input and Output signal')
plt.xlim((0, 1)),plt.ylim((-1, 1))

Fig. 10.37 Python code which implements filter bank without filters

2. There is two sample delay between the input and output signal, which is in
agreement with the theoretical result.

10.5 Tree-Structured Filter Bank

The structure of uniform tree-structured filter bank is given in Fig. 10.43. The
numbers after the block are used to understand the sequence of the process of the
input signal. The same numbers are used as variables in the python code to
understand the sequence of process.
Experiment 10.14 Tree-Structured Filter Bank
This experiment illustrates the concept of tree-structured filter bank using python.
The python code for tree-structured filter bank is shown in Fig. 10.44, and the
corresponding output is shown in Fig. 10.45.
Inference
The following inferences can be made from this experiment:
10.5 Tree-Structured Filter Bank 429

Fig. 10.38 Result of filter bank without filters

1 4 7 10
H0(z) 3 3 G0(z)
5 8 11
x[n] 2
y[n]
H1(z) 3 3 G1(z)
3 6 9 12
H2(z) 3 3 G2(z)

Fig. 10.39 Structure of three-channel filter bank

1 4 7 10
1 3 3 z-2

2 5 8 11
x[n] y[n]
z-1 3 3 z-1
3 6 9 12
z-2 3 3 1

Fig. 10.40 Three-channel filter bank for the choice of filters

430 10 Multirate Signal Processing

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
def downsample(x,M):
y=x[::M]
return(y)
def upsample(x,L):
y=np.zeros(L*len(x))
y[::L]=x
return(y)
#Define the filters
h0,h1,h2=[1],[0,1],[0,0,1]
g0,g1,g2=[0,0,1],[0,1],[1]
#Input signal
t=np.linspace(0,1,200)
x=signal.sawtooth(2*np.pi*5*t)
x1=signal.lfilter(h0,1,x)
x2=signal.lfilter(h1,1,x)
x3=signal.lfilter(h2,1,x)
x4=downsample(x1,3)
x5=downsample(x2,3)
x6=downsample(x3,3)
x7=upsample(x4,3)
x8=upsample(x5,3)
x9=upsample(x6,3)
x10=signal.lfilter(g0,1,x7)
x11=signal.lfilter(g1,1,x8)
x12=signal.lfilter(g2,1,x9)
y=x10+x11+x12
plt.plot(t,x,t,y[0:len(t)]),plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.title('Input and Output signal'),
plt.legend(['Input','Output'],loc=1), plt.tight_layout()

Fig. 10.41 Three-channel filter bank without filters

1. The input signal to tree-structured filter bank is a sawtooth signal of 5 Hz

frequency. The output signal is also a sawtooth signal.
2. The output signal is a delayed version of the input signal. Thus, tree-structured
filter bank obeys the perfect reconstruction criterion.
3. Perfect reconstruction is achieved through the proper choice of analysis and
synthesis filters.
10.6 Transmultiplexer 431

Input and Output signal

1.00 Input
0.75 Output

0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 10.42 Input and output signals of three-channel filter bank without filters

5 7 13 15
H0(z) 2 2 G0(z) 18
1 3 17 19
H0(z) 2 2 G0(z)
6 8 14
H1(z) 2 2 G1(z)
16 y[n]
x[n]
20
9 11 22
H0(z) 2 2 G0(z) 25
2 4 24
H1(z) 2 2 G1(z
10 26
12 21
H1(z) 2 2 G1(z)
23

Analysis section Synthesis section

Fig. 10.43 Structure of tree-structured filter bank

10.6 Transmultiplexer

A transmultiplexer converts time division multiplexing (TDM) signals to frequency

division multiplexing (FDM) and vice versa. A technique for sending multiple
signals through the same physical medium is to use different portions of the available
frequency spectrum. Frequency division multiplexing refers to the process of spec-
tral separation to permit the simultaneous transmission of signals from multiple
users. In frequency division multiplexing, all the signals operate at the same time
with different frequencies. In time-division multiplexing, all the signals operate with
the same frequency at different times. The operation of converting from one form of
432 10 Multirate Signal Processing

#Tree structured filter bank

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
def downsample(x,M): #Function to perform downsampling operation
y=x[::M]
return(y)
def upsample(x,L): #Function to perform upsampling operation
y=np.zeros(L*len(x))
y[::L]=x
return(y)
#Step 1: Define the filters
h0=[0.5,0.5]
h0=np.array(h0)
h1=signal.qmf(h0)
g0,=2*h0,-2*h1
#Step 2: Generate the input signal
t=np.linspace(0,1,100)
x=signal.sawtooth(2*np.pi*5*t)
#Step 3: Traversing the path(1: Analysis section)
x1=signal.lfilter(h0,1,x)
x2=signal.lfilter(h1,1,x)
x3=downsample(x1,2)
x4=downsample(x2,2)
x5=signal.lfilter(h0,1,x3)
x6=signal.lfilter(h1,1,x3)
x7=downsample(x5,2)
x8=downsample(x6,2)
x9=signal.lfilter(h0,1,x4)
x10=signal.lfilter(h1,1,x4)
x11=downsample(x9,2)
x12=downsample(x10,2)
#2: Synthesis section
x13=upsample(x7,2)
x14=upsample(x8,2)
x15=signal.lfilter(g0,1,x13)
x16=signal.lfilter(g1,1,x14)
x17=x15+x16
x18=upsample(x17,2)
x19=signal.lfilter(g0,1,x18)
x20=upsample(x11,2)
x21=upsample(x12,2)
x22=signal.lfilter(g0,1,x20)
x23=signal.lfilter(g1,1,x21)
x24=x22+x23
x25=upsample(x24,2)
x26=signal.lfilter(g1,1,x25)
y=x19+x26
plt.plot(t,x,'k',t,y,'r'),plt.legend(['Input','Output'],loc=1),plt.xlabel('Time'),
plt.ylabel('Amplitude'),plt.title('Input-Output waveform')

Fig. 10.44 Python code for uniform tree-structured filter bank

10.6 Transmultiplexer 433

Input-Output wavefrom
1.00 Input
Output
0.75

0.50

0.25
Amplitude

0.00

–0.25

–0.50

–0.75

–1.00
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 10.45 Input-output waveform of tree-structured filter bank

1 3 6 8
x[0] 2 G0(z) H0(z) 2 v[0]

5
2 4 7 9
x[1] 2 G1(z) H1(z) 2 v[1]

Fig. 10.46 Structure of two-channel transmultiplexer

multiplexing to another is termed as ‘transmultiplexing’. The structure of

two-channel transmultiplexer is shown in Fig. 10.46.
Proper choice of filters will avoid the problem of crosstalk in two-channel
transmultiplexer.
Experiment 10.15 Implementation of Two-Channel Transmultiplexer
This experiment discusses the implementation of two-channel transmultiplexer
using python. The python code, which implements a two-channel transmultiplexer,
is shown in Fig. 10.47, and the corresponding output is shown in Fig. 10.48.
Figure 10.47 shows that the signal x[0] is a sine wave of 5 Hz frequency, and the
signal x[1] is a cosine wave of 5 Hz frequency. The transmitted signals are sine and
cosine waves. The variables ‘n0 to n9’ in the python code shown in Fig. 10.46 align
with the nodes depicted in Fig. 10.47. From Fig. 10.48, it is possible to observe that
the received signals y[0] is a sine wave similar to the signal x[0], and the signal y[1] is
a cosine wave similar to the transmitted signal x[1].
434 10 Multirate Signal Processing

#Two-channel transmultiplexer
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
def downsample(x,M):
y=x[::M]
return(y)
def upsample(x,L):
y=np.zeros(L*len(x))
y[::L]=x
return(y)
#Step 1 defining x0 and x1
f,fs,N,N1=5,100,256,128;
T=1/fs
t=np.linspace(0,N*T,N)
x0=np.sin(2*np.pi*f*t)
x1=np.cos(2*np.pi*f*t)
#Step 2: Define the filters
g0,g1,h0,h1=[0, 1, 1], [0, -1, 1], [0.5, 0.5], [0.5, -0.5];
#Step3 Tracing the structure
n1=upsample(x0,2) #At node 1
n2=upsample(x1,2) #At node 2
n3=signal.lfilter(g0,1,n1) #At node 3
n4=signal.lfilter(g1,1,n2) #At node 4
n5=n3+n4 #At node 5
n6=signal.lfilter(h0,1,n5) #At node 6
n7=signal.lfilter(h1,1,n5) #At node 7
n8=downsample(n6,2) #At node 8
n9=downsample(n7,2) #At node 9
plt.subplot(2,1,1),plt.plot(t,x0,t,n8),plt.legend(['Transmitted(x0)','Received(y0)'],loc=1)
plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.subplot(2,1,2),plt.plot(t,x1,t,n9),plt.legend(['Transmitted(x1)','Received(y1)'],loc=1)
plt.xlabel('Time'),plt.ylabel('Amplitude')
plt.suptitle('Transmitted and received signals in transmultiplexer')
plt.tight_layout()

Fig. 10.47 Python code for two-channel transmultiplexer

Inferences
The following inferences can be drawn from Fig. 10.48:
1. The received signals are similar to the transmitted signal without distortion; thus,
crosstalk problem is avoided. The proper choice of synthesis and analysis filter
avoids the crosstalk problem.
2. Perfect reconstruction of transmultiplexer achieves complete crosstalk cancella-
tion and is distortion-free.
10.6 Transmultiplexer 435

Fig. 10.48 Transmitted and received signals in transmultiplexer

Experiment 10.16 Audio Signal Transmission

This experiment tests the process of two-channel transmultiplexer using audio signal
as an input. Instead of transmitting sine wave and cosine wave as the input through
transmultiplexer, pass male and female voices through two-channel
transmultiplexer, and observe whether the male and female voices can be received
in the output without distortion.
The python code which performs this task is shown in Fig. 10.49, and the
corresponding output is shown in Fig. 10.50.
Inferences
From Fig. 10.49, it is possible to observe the following:
1. The library ‘sounddevice’ is used to play the speech signal.
2. The built-in function ‘wavfile’ from ‘scipy’ library is used to read the speech
signal.
3. The two signals fed to the transmultiplexer are (a) x0 is a male voice and (b) x1 is
a female voice.
4. The output (variable name ‘n8’) represents the received male voice corresponding
to the transmitted male voice (x0).
5. The output (variable name ‘n9’) represents the received female voice
corresponding to the transmitted female voice (x1).
6. The transmitted and the received male and female voices are plotted and heard.
436 10 Multirate Signal Processing

#Transmultiplexer for speech signal

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
from scipy.io import wavfile
import sounddevice as sd
#Functions to perform downsampling and upsampling
def downsample(x,M):
y=x[::M]
return(y)
def upsample(x,L):
y=np.zeros(L*len(x))
y[::L]=x
return(y)
#Step 1 Reading the speech signals
fs, x0 = wavfile.read('Male.wav')
fs, x1 = wavfile.read('Female.wav')
x0=x0[:,0]
x1=x1[:,1]
#Step 2: Define the filters
g0,g1,h0,h1=[0,1, 1], [0,-1, 1], [0.5, 0.5], [0.5, -0.5];
#Step3 Tracing the structure
n1=upsample(x0,2)
n2=upsample(x1,2)
n3=signal.lfilter(g0,1,n1)
n4=signal.lfilter(g1,1,n2)
n5=n3+n4
n6=signal.lfilter(h0,1,n5)
n7=signal.lfilter(h1,1,n5)
n8=downsample(n6,2)
n9=downsample(n7,2)
plt.subplot(2,2,1),plt.plot(x0),plt.title('Transmitted male voice')
plt.subplot(2,2,2),plt.plot(x1),plt.title('Transmitted female voice')
plt.subplot(2,2,3),plt.plot(n8),plt.title('Received male voice')
plt.subplot(2,2,4),plt.plot(n9),plt.title('Received female voice')
plt.tight_layout()
#Hearing the audio signal
sd.play(x0,fs) #Transmitted male voice
sd.wait()
sd.play(n8,fs) #Received male voice
sd.wait()
sd.play(x1,fs) #Transmitted female voice
sd.wait()
sd.play(n9,fs) #Received female voice

Fig. 10.49 Python code for Experiment 10.16

10.6 Transmultiplexer 437

Transmitted male voice Transmitted female voice

0.2 0.2

0.0 0.0

–0.2
–0.2
0 50000 100000 0 50000 100000
Received male voice Received female voice
0.2 0.2

0.0 0.0

–0.2
–0.2
0 50000 100000 0 50000 100000

Fig. 10.50 Transmitted and received speech signals through two-channel transmultiplexer

Fig. 10.51 Multirate

system
System
x[n] y[n]
2 2

From Fig. 10.50, it is possible to observe that the received speech signal resem-
bles the transmitted speech signal, which confirms that the two-channel
transmultiplexer is free from cross-talk.
Exercises
1. Write a python code to simulate the comb signal whose expression is given by

M-1
1 2π
x ½ n] = ej M kn , 0 ≤ n ≤ 10
M k=0

For M = 1, 2, 3 and 4, comment on the observed output.

2. Write a python code to prove that the downsampling operation obeys the
superposition principle. The objective is to prove that downsampling operation
obeys both additivity and homogeneity properties.
3. Write a python code to prove that upsampling by a factor of 2 is a time-varying
operation.
4. Generate a square wave of 5 Hz fundamental frequency. Downsample this
signal by a factor of 2. Plot the spectrum of the input and downsampled square
waves, and comment on the observed result.
5. Write a python code to prove the fact that the output signal y[n] is identical to the
input signal x[n] for the multirate system shown in Fig. 10.51.
438 10 Multirate Signal Processing

Fig. 10.52 Multirate

system
System
x[n] y[n]
2 2

H0(z)
1 3 5 G0(z) 7
2 2
y[n]
x[n]
9
2 4 [0,0,0…0] 6 G1(z) 8
H1(z) 2 2

Analysis Section Synthesis Section

Fig. 10.53 Two-channel filter bank

6. Write a python code to implement the following multirate system depicted in

Fig. 10.52.
Plot the input and output signal and comment on the observed output.
7. Read an audio signal. Downsample it by a factor of 2. Hear the original and
downsampled audio signal and comment on the observation.
8. Read an audio signal. Upsample it by a factor of 2. Hear the original and
upsampled audio signal and comment on the observation.
9. The input to the two-channel filter bank shown in Fig. 10.53 is a square of 5 Hz
fundamental frequency. The impulse response of the analysis and synthesis
filters are h0[n] = {0.5, 0.5}, h1[n] = {0.5, -0.5}, g0[n] = {1, 1} and g1[n] =
{-1, 1}. The high pass filter section has to be masked. The bitstream from node
4 has to be strings of zeros. The impact of masking the high frequency compo-
nent has to be analysed.
Plot the input and output signal and comment on the observed output.
10. Record two voice signals, namely, x[0] and x[1], with a sampling frequency of
8000 Hz. The signal x[0] corresponds to the word ‘YES’, and the signal x
[1] corresponds to the word ‘NO’. That is, the signal x[0] and x[1] are recorded
voice signals with the word ‘YES’ and ‘NO’, respectively. Pass these two voice
signals to crosstalk free transmultiplexer to obtain the output signal y[0] and y
[1]. Comment on the observed output.

Objective Questions
1. Downsampling by a factor of ‘M’ is a
A. Linear, time-invariant operation
B. Linear, time-variant operation
C. Non-linear, time-invariant operation
D. Non-linear, time-variant operation
2. The time-domain expression for downsampling by a factor of ‘M’ is given by
10.6 Transmultiplexer 439

A. y½n] = x Mn
B. y[n] = x[nM]
C. y[n] = x[n]M
1
D. y½n] = x½n]M
3. The input signal x[n] is upsampled by a factor of ‘L’; then, the result is
downsampled by the same factor ‘L’ to obtain the signal y[n]. The relationship
between y[n] and x[n] is given by
A. y[n] = x[nL]
B. y[n] = x[n]
C. y½n] = x Ln
D. y[n] = x[n]L

4. The input signal x½n] = 1, 1, 1, 1 is passed through downsampling by a factor

"
of 2. The result of downsampling operation is then passed through upsampling
by a factor of 2 to obtain the output signal y[n]. The expression for the output
signal is

A. y½n] = 1, 1, 1, 1
"

B. y½n] = 1, 0, 1, 0
"

C. y½n] = 1, - 1, 1, - 1
"
D. y[n] = {0, 0, 0, 0}
5. Which of the following operation is an example of an idempotent operation
A. Upsampling followed by downsampling
B. Downsampling followed by upsampling
C. Downsampling followed by downsampling
D. Upsampling followed by upsampling
6. Which of the following is an example of an identity operation
A. Upsampling followed by downsampling
B. Downsampling followed by upsampling
C. Downsampling followed by downsampling
D. Upsampling followed by upsampling
7. Which of the following results in idempotent operation
A. Downsampling by a factor of ‘M’ followed by upsampling by a factor of ‘M’
B. Upsampling by a factor of ‘M’ followed by downsampling by the same
factor
440 10 Multirate Signal Processing

C. Downsampling followed by unit delay operation

D. Upsampling followed by unit delay operation
8. If the variable ‘x’ contain the input signal, the python command y=x[::2] results
in
A. Upsampling of the input signal by a factor of 2
B. Downsampling of the input signal by a factor of 2
C. Delaying of input signal by a factor of 2
D. Advance of input signal by a factor of 2
9. A function ‘operation’ is given below. The function accepts the input signal (x)
and gives an output signal ( y). What is the relationship between the input and
output signal?

A. Output signal ‘y’ is downsampled by a factor of 2

B. Output signal ‘y’ is upsampled by a factor of 2
C. Output signal ‘y’ is delayed by a factor of 2
D. Output signal ‘y’ is advanced by a factor of 2
10. Downsampling is a
A. Linear, time-invariant operation
B. Linear, time-variant operation
C. Non-linear, time-invariant operation
D. Non-linear, time-variant operation
11. Interchanging of upsampling by a factor of ‘L’ and downsampling by a factor of
‘M’ is possible if and only if
A. L and M are of same value
B. L and M should be odd number
C. L and M should be even number
D. L and M are relatively prime

Bibliography

1. Ronald E. Crochiere, and Lawrence R. Rabiner, “Multirate Digital Signal Processing”,

Pearson, 1983.
2. P. P. Vaidyanathan, “Multirate Systems and Filter Banks”, Prentice Hall, 1993.
Bibliography 441

3. N. J. Fliege, “Multiratge Digital Signal Processing: Multirate Systems, Filterbanks, Wavelets”,

John Wiley and Sons, 1999.
4. Bruce W. Suter, “Multirate and Wavelet Signal Processing”, Academic Press, 1997.
5. Vikram M Gadre, and Aditya S. Abhyankar, “Multiresolution and Multirate Signal Processing:
Introduction, Principles and Applications”, McGraw Hill, 2017.
Chapter 11
Adaptive Signal Processing

Learning Objectives
After reading this chapter, the reader is expected to
• Implement and analyse the Wiener filter.
• Write a python code to implement the LMS algorithm and its variants.
• Perform system identification using the LMS algorithm.
• Perform inverse system modelling using the NLMS algorithm.
• Implement adaptive line enhancer using the LMS algorithm and its variants.
• Implement the RLS algorithm.

Roadmap of the Chapter

The roadmap of this chapter is depicted below. This chapter starts with the Wiener
filter, least mean square (LMS) algorithm and its variant approaches for adaptive
signal processing applications like system identification and signal denoising. Next,
the RLS algorithm is discussed with the suitable python code.

Filter

Optimum Filter Adaptive Filter

Wiener Filter
LMS Algorithm RLS Algorithm

NLMS Algorithm

Sign LMS Algorithm

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 443
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_11
444 11 Adaptive Signal Processing

PreLab Questions
1. List out the valid differences between the optimal filter and the adaptive filter.
2. What is an adaptive filter? How it differs from the ordinary filter.
3. Examples of adaptive filter.
4. When are adaptive filters preferred?
5. List out the performance measures of the adaptive filter.
6. What is an LMS algorithm?
7. What do you mean by least square estimation?
8. List out the variants of LMS algorithm.
9. How the step size impacts the LMS algorithm?
10. What is the RLS algorithm, and how it differs from LMS?

11.1 Wiener Filter

Wiener filter is the mean square error (MSE) optimal stationary linear filter for signal
corrupted by additive noise. The Wiener filter computation requires the assumption
that the signal and noise are in the random process. The general block diagram of the
Wiener filter is shown in Fig. 11.1. The main objective of the Wiener filter is to
obtain the filter coefficient of the LTI filter, which can provide the final output (y[n])
as much as the minimum MSE between the output and the desired signal or target (d
[n]). In Fig. 11.1, s[n] denotes the original signal, which is a clean signal, and it is
corrupted by additive noise η[n] to give the signal x[n]. The parameters of the filter
have to be designated has to be designed in such a way that the output of the filter y
[n] should resemble the desired signal d[n] such that the error ‘e[n]’ is minimum.
The expression for the optimal Wiener filter is given by

hopt = R - 1 p ð11:1Þ

The above expression is termed as ‘Wiener-Hopf’ expression, which is named

after American-born Norbert Wiener and Austrian-born Eberhard Hopf. The expres-
sion for optimal filter depends on the autocorrelation matrix (R) of the observed
signal (x[n]) and the cross-correlation vector ( p) between the observed signal (x[n])
and the desired signal (d[n]). hopt denotes the optimal filter coefficients.
Experiment 11.1 Wiener Filtering
The aim of this experiment is to implement the Wiener filtering using python. Here
the optimal filter coefficients are obtained using the Wiener-Hopf equation given in

Fig. 11.1 Block diagram of η[n] d[n]

Wiener filter

s[n] x[n] y[n] e[n]

h[n]
11.1 Wiener Filter 445

#Wiener filter
import numpy as np
from numpy.random import randn
import matplotlib.pyplot as plt
from scipy.linalg import toeplitz
from scipy import signal
#Step 1: Generation of signal s[n]
t=np.linspace(0,1,100)
s=np.sin(2*np.pi*5*t)
Ns=len(s)
#Step 2: Generation of random noise
# n=randn(len(t))*0.1
n=np.random.normal(0,.2,len(s))
#Step 3: Observed signal x[n]
x=s+n
#Step 4: Autocorrelation of observed signal
rxx=np.correlate(x,x,mode='full')
#Step 5: Cross-correlation between desired and observed signal
rsx=np.correlate(s,x,mode='full')
#Step 6: Deciding the length of the filter
Nh=11
#Step 7: Trimming the autocorrelation and cross-correlation values
rxx1=rxx[Ns-1:Ns+Nh-1]
rsx1=rsx[Ns-1:Ns+Nh-1]
#Step 8: Obtaining the autocorrelation matrix
Rx=toeplitz(rxx1)
#Step 9: Inverse of the autocorrelation matrix
Rx1=np.linalg.inv(Rx)
#Step 10: Obtaining the filter coefficient
w1=np.matmul(Rx1,rsx1)
#Step 11: Filtering the noisy signal
y=signal.lfilter(w1,1,x)
plt.subplot(3,1,1),plt.plot(t,s),plt.xlabel('t-->'),plt.ylabel('Amplitude'),
plt.title('Clean signal'),plt.subplot(3,1,2),plt.plot(t,x),plt.xlabel('t-->'),
plt.ylabel('Amplitude'),plt.title('Noisy signal'),plt.subplot(3,1,3), plt.plot(t,y)
plt.xlabel('t-->'),plt.ylabel('Amplitude'),plt.title('Filtered signal'),plt.tight_layout()

Fig. 11.2 Python code for Wiener filtering

Eq. (11.1). The python code for Wiener filter is shown in Fig. 11.2. Simulation result
of the python code given in Fig. 11.2 is depicted in Fig. 11.3.
The built-in functions used in python code shown in Fig. 11.2 is summarized in
Table 11.1.
Inference
From Fig. 11.3, it can be made the following observations:
446 11 Adaptive Signal Processing

Clean signal

Amplitude
1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
t-->
Noisy signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
t-->
Filtered signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
t-->

Fig. 11.3 Simulation result of Wiener filter

Table 11.1 Built-in functions used in the python code given in Fig. 11.2
S. No. Objective Built-in function Library
1 To generate a clean sinusoidal signal of 5 Hz frequency np.sin() Numpy
2 To add white noise, which follows normal distribution to np.random.nor- Numpy
clean signal mal()
3 To perform autocorrelation np.correlate() Numpy
4 To obtain the inverse of the matrix np.linalg.inv() Scipy
5 To perform convolution signal.lfilter() Scipy

1. The input or clean signal frequency is 5 Hz, and it is a smooth sine waveform.
2. The additive noise added signal as input to the Wiener filter, and it is a distorted
signal.
3. The filtered signal is not a smooth sine waveform. However, this waveform is far
better than the noisy signal. Hence, the Wiener filter has a capability to minimize
the impact of additive noise in a signal.
Task
1. Change the value of standard deviation in random noise generation python
command ‘np.random.normal(0,.2,len(s))’ given in Fig. 11.2. Execute and
make the appropriate changes in this python code to get ‘filtered signal’ as similar
as ‘clean signal’.
Experiment 11.2 Wiener Filter Using Built-In Function
This experiment performs the Wiener filtering using built-in function in ‘scipy’
library. The built-in function is available in the ‘scipy’ library ‘wiener’ can be
used to filter out the noisy components. In this experiment, noise-free sinusoidal
signal of 5 Hz frequency is generated. The clean signal is corrupted by adding
11.1 Wiener Filter 447

Table 11.2 Steps followed and built-in functions

S. No. Objective Built-in function Library
1 To generate a clean sinusoidal signal of 5 Hz frequency np.sin() Numpy
2 To add white noise, which follows normal distribution to np.random.nor- Numpy
clean signal mal()
3 To minimize the impact of noise using Wiener filter signal.wiener() Scipy

#Wiener filter
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
#Step 1: Generation of clean signal
t=np.linspace(0,1,100)
s=np.sin(2*np.pi*5*t)
#Step 2: Adding noise
n=np.random.normal(0,.2,len(s))
x=s+n
#Step 3: Wiener filter
y=signal.wiener(x)
#Step 4: Plotting the results
plt.subplot(3,1,1),plt.plot(t,s),
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Clean signal')
plt.subplot(3,1,2),plt.plot(t,x),plt.xlabel('Time'),plt.ylabel('Amplitude'),
plt.title('Noisy signal'),plt.subplot(3,1,3),plt.plot(t,y)
plt.xlabel('Time'),plt.ylabel('Amplitude'),plt.title('Filtered signal')
plt.tight_layout()

Fig. 11.4 Wiener filtering using built-in function

random noise, which follows the normal distribution with zero mean and 0.2
standard deviation. The corrupted signal is then passed through the Wiener filter to
minimize the impact of noise. The steps followed along with the built-in functions
used in the program are given in Table 11.2.
The python code which performs this task is shown in Fig. 11.4, and the
corresponding output is shown in Fig. 11.5.
Inference
From Fig. 11.5, it is possible to infer that the impact of noise is minimized after
passing the noisy signal through Wiener filter.

11.1.1 Wiener Filter in Frequency Domain

From Wiener-Hopf equation, the expression for the optimal Wiener filter is given by
448 11 Adaptive Signal Processing

Clean signal

Amplitude
1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
Time
Noisy signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
Time
Filtered signal
Amplitude

1
0
–1
0.0 0.2 0.4 0.6 0.8 1.0
Time

Fig. 11.5 Result of Wiener filtering

hopt = R - 1 p ð11:2Þ

The above equation can be expressed as

p
hopt = ð11:3Þ
R

In the above expression, ‘p’ represents the cross-correlation between desired

signal and the observed signal, and ‘R’ represents the autocorrelation of the observed
signal. Taking Fourier transform on both sides of Eq. (11.3), we get

FTfpg
FT hopt = ð11:4Þ
FTfRg

According to the Wiener-Khinchin theorem, Fourier transform of autocorrelation

function gives power spectral density. Using this theorem, Eq. (11.4) is expressed as

Sdx ðejω Þ
H ejω = ð11:5Þ
Sxx ðejω Þ

In Eq. (11.5), H(ejω) represents the frequency response of the Wiener filter,
Sdx(ejω) represents the cross-power spectral density estimation between desired and
observed signal and Sxx(ejω) represents the power spectral density of the observed
signal.
11.1 Wiener Filter 449

Desired Clean signal s[n]

signal
Observed
White noise signal
d[n]
Noisy signal ƞ[n] x[n]

Power Spectral
Density S xx (e jZ )
Power Spectral
Density S xd (e jZ )

Wiener filter
S dx (e jZ )
H ( e jZ )
S xx (e jZ )

Inverse Fourier Transform

Impulse response
h[n]

Fig. 11.6 Wiener filter in frequency domain

Experiment 11.3 Wiener Filter in Frequency Domain

The steps followed in the implementation of Wiener filter in frequency domain are
given in Fig. 11.6. The noisy signal is obtained by adding white noise, which follows
normal distribution to the clean signal. The observed signal is a clean signal with
white noise added to it. The power spectral density of the observed signal is
represented by Sxx(ejω). The power spectral density between the desired and
observed signal is represented by Sdx(ejω). The Wiener filter is obtained in the
frequency domain using the relation H ðejω Þ = SSdxxx ððeejω ÞÞ. Here the desired signal is the
jω

clean signal s[n]. Upon taking inverse Fourier transform of H(ejω), the impulse
response of the Wiener filter is obtained.
The python code used to implement the Wiener filter in frequency domain is
shown in Fig. 11.7, and the corresponding output is in Fig. 11.8.
The built-in functions used in the program and its purpose are given in Table 11.3.
Inference
From Fig. 11.8, the following observations can be made:
450 11 Adaptive Signal Processing

#Wiener filter in frequency domain

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
from matplotlib import patches
t=np.linspace(0,1,100)
s=np.sin(2*np.pi*5*t) #Step1: Generation of clean signal s[n]
n=np.random.normal(0,0.1,len(t)) #Step 2: Generation of noise
x=s+n #Step 3: Generation of observed signal x[n]
Nh=25
f,Pxx=signal.csd(x,x,nperseg=Nh) #Step 4: Power spectral density of observed signal
f,Psx=signal.csd(s,x,nperseg=Nh) #Step 5: PSD of desired and observed signal
H=Psx/Pxx #Step 6: Wiener filter in frequency domain
h=np.fft.irfft(H) #Step 7: Wiener filter in time domain
w, H1 = signal.freqz(h, 1)
y=signal.filtfilt(h,1,x) #Step 8: Filtered signal
plot1 = plt. figure(1)
bx=plt.subplot(3,1,1)
bx.plot(t,s),bx.set(title='Clean signal',xlabel='Time',ylabel='Amplitude')
bx=plt.subplot(3,1,2)
bx.plot(t,x),bx.set(title='Noisy signal',xlabel='Time',ylabel='Amplitude')
bx=plt.subplot(3,1,3)
bx.plot(t,y),bx.set(title='Filtered signal',xlabel='Time',ylabel='Amplitude')
plt.tight_layout()
plot2 = plt. figure(2)
#Pole-zero plot of the filter
ax = plt.subplot(2,2,3);
unit_circle = patches.Circle((0,0),radius = 1 , fill = False,color='black',ls='solid',alpha = 0.1)
ax.add_patch(unit_circle),ax.axhline(0,color='black',alpha = 0.5)
ax.axvline(0,color='black',alpha = 0.5)
b,a = h,[1]
z,p,k = signal.tf2zpk(b,a)
ax.plot(np.real(z),np.imag(z),'or',label='zeros')
ax.plot(np.real(p),np.imag(p),'xb',label = 'poles')
ax.set(title='Zeros and poles',xlabel='$\sigma$', ylabel='$j\omega$'),ax.legend(loc = 2),ax.grid()
ax = plt.subplot(2,2,1)
ax.stem(h),ax.set(title='Impulse response',xlabel='n-->',ylabel='Amplitude')
ax = plt.subplot(2,2,2)
ax.plot(w/np.pi,20*np.log10(abs(H1))),
ax.set(title='Magnitude response',xlabel='w',ylabel='Magnitude')
ax=plt.subplot(2, 2, 4)
ax.plot(w/np.pi, 180/np.pi*np.unwrap(np.angle(H1)))
ax.set(title='Phase response',xlabel='w',ylabel='Phase'),plt.tight_layout()

Fig. 11.7 Python code to implement Wiener filter in frequency domain

1. The impact of noise is minimized by applying the Wiener filter.

2. The impulse response of the Wiener filter is not symmetric; hence, the phase
response of the filter is not a linear curve.
3. From the magnitude response, it is possible to observe that the filter is a lowpass
filter, and it performs smoothing actions to minimize the impact of noise.
11.2 Adaptive Filter 451

Clean signal Impulse response Magnitude response

Amplitude

1 0

Magnitude
Amplitude
0 0.2
–1 –20
0.0 0.2 0.4 0.6 0.8 1.0 0.1
Time 0.0 –40
Noisy signal
Amplitude

1 0 10 20 0.0 0.5 1.0

0 n--> w
–1
Zeros and poles Phase response
0.0 0.2 0.4 0.6 0.8 1.0 1 0
Time zeros

Phase
x poles
Filtered signal –500
Amplitude

jZ
1
0 –1000
–1 –1
0.0 0.2 0.4 0.6 0.8 1.0 –1 0 1 0.0 0.5 1.0
Time s w

Fig. 11.8 Result and characteristics of Wiener filtering

Table 11.3 Built-in functions used in this experiment

Built-in
S. No. Objective function Library
1 To generate clean sinusoidal signal of 5 Hz frequency np.sin() Numpy
2 To add white noise which follows normal distribution to np.random. Numpy
clean signal normal()
3 To compute the power spectral density signal.csd() Scipy
4 To compute the impulse response of the filter from the np.fft.irfft() Numpy
frequency response
5 To obtain the frequency response of the filter signal.freqz() Scipy
6 To obtain the poles, zeros and the gain of the filter from the signal.tf2zpk() Scipy
transfer function

Fig. 11.9 General block d[n]

diagram of adaptive filtering

y[n]
x[n] Adaptive filter -
e[n]

4. From the pole-zero plot, it is possible to observe that poles and zeros lie within the
unit circle; hence the filter is stable.

11.2 Adaptive Filter

The adaptive filter is a non-linear filter, which updates the value of the filter
coefficients based on some specific criterion. The general block diagram of the
adaptive filter is shown in Fig. 11.9. From this figure, it is possible to observe that
452 11 Adaptive Signal Processing

the filter coefficients are updated based on the error, e[n] between the output of the
filter y[n] and reference data d[n]. Examples of adaptive filters are LMS filter and
RLS filter.

11.2.1 LMS Adaptive Filter

The LMS is a least mean square algorithm that works based on the stochastic
gradient descent approach to adapt the estimate based on the current error. The
estimate is called the weight or filter coefficient. The weight or filter coefficient
update equation of the LMS algorithm is given by.

w½n þ 1] = w½n] þ μx½n]e½n] ð11:6Þ

where w[n + 1] represents the new weight or updated weight, w[n] denotes the old
weight, μ indicates the step size or learning rate, x[n] is the input signal or data and
the error signal e[n] = d[n] - y[n]. d[n] is the reference data or target data, and y[n] is
the actual output of the adaptive filter of the system.
Experiment 11.4 Implementation of LMS Algorithm
This experiment discusses the implementation of LMS algorithm for adaptive
filtering using python. The python code to define the LMS algorithm as a function
is shown in Fig. 11.10. This code can be called a function in the different applica-
tions of the LMS algorithm, which will be discussed in the subsequent experiments.
From Fig. 11.10, it is possible to see that the weight updation formula of the LMS
algorithm given in Eq. (11.6) exists in it.
Inference
1. From Fig. 11.10, it is possible to observe that the LMS algorithm is written as a
function, and it can be called a signal processing application whenever needed.

# This python code for LMS algorithm

def LMS_algorithmm(x,mu,N,t):
# x = input data, mu = step size, t = reference data, N = Filter length
N1=len(x)
w = np.zeros(N) # Initial filter
e = np.zeros(N1-N)
for n in range(0, N-F):
xn = x[n+N:n:-1]
en = t[n+N] - np.dot(xn,w) # Error
w = w + mu * en * xn # Update filter (LMS algorithm)
e[n] = en # Record error
return w,e

Fig. 11.10 Python code for LMS algorithm

11.2 Adaptive Filter 453

Fig. 11.11 Block diagram x[n]

of system identification

System w(n)

d[n] y[n] Adaptive

∑ algorithm

e[n]

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N1 = 500 # Size of the Input data
N = 25 # Filter size
n_iter=[10,50,100,150]# it must be less than (N1-N)
x = np.random.randn(N1) # Input to the filter
h = signal.firwin(N, 0.25) # FIR filter to be identified
t = signal.convolve(x, h) # Target/desired signal
t = t + 0.01 * np.random.randn(len(t)) # with added noise
mu = 0.04 # LMS step size
plt.figure(),plt.title('Filter to be Identified'),plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]')
for i in range(0,len(n_iter)):
[w,e]=LMS_algorithmm(x,mu,N,t,n_iter[i]);
plt.figure(),plt.title('Error signal at iteration %d' % n_iter[i])
plt.stem(e),plt.xlabel('n-->'),plt.ylabel('e[n]')
plt.figure(),plt.title('Identified Filter at iteration %d' % n_iter[i])
plt.stem(w),plt.xlabel('n-->'),plt.ylabel('w[n]')

Fig. 11.12 Python code for unknown system identification

2. The inputs to the LMS function are ‘x’, ‘mu’, ‘N’ and ‘t’. ‘x’ denotes the input
data, ‘mu’ represents step size, ‘t’ denotes the reference data or target data and ‘N’
indicates the length of the adaptive filter.
3. The outputs from this LMS function are ‘w’, which denotes the adaptive filter
coefficients, and ‘e’ is an error between the estimate and target data.
Experiment 11.5 System Identification Using LMS Algorithm
This experiment deals with unknown system identification using the LMS algorithm.
Let us consider the unknown system as an FIR filter with a length of 25. In this
experiment, the output filter coefficients are obtained by using LMS algorithm with
different number of iterations. The block diagram of the system identification is
shown in Fig. 11.11. The python code to find the unknown system using the LMS
algorithm is given in Fig. 11.12, and its simulation result is shown in Fig. 11.13.
 454

Filter to be Identified Identified Filter at iteration 10 Identified Filter at iteration 25

0.25
0.04
0.20 0.10
0.02
0.15 0.05
0.00
0.10

h[n]
w[n]
w[n]
–0.02 0.00
0.05
–0.04 –0.05
0.00
–0.06 –0.10
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
n--> n--> n-->
Identified Filter at iteration 50 Identified Filter at iteration 100 Identified Filter at iteration 150
0.25
0.25
0.15 0.20
0.20
0.10 0.15 0.15

0.10 0.10
w[n]

w[n]

w[n]
0.05
11

0.05 0.05
0.00
0.00 0.00
–0.05
–0.05
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
n--> n--> n-->

Fig. 11.13 Simulation results of Experiment 11.5

Adaptive Signal Processing
11.2 Adaptive Filter 455

x[n] Unknown
Adaptive filter
system
- y[n]

e[n]
+ d[n]
Delay

Fig. 11.14 Inverse system modelling using adaptive filter

Figure 11.12 indicates that the number of iterations is considered as 10, 50,
100 and 150, and the length of the unknown FIR filter is chosen as 25. The input
to the LMS algorithm is a random signal with a length of 500 samples. The targeted
or desired or reference data is obtained by convolving the input random signal with
the unknown FIR filter coefficients along with the random noise.
Note that the inputs to the LMS algorithm ([w,e]=LMS_algorithmm(x,mu,N,t,
n_iter[i])) are random signal (x), learning rate (mu), length of the filter (N), a
reference signal (t) and number or iteration (n_iter). Also, note that the filter
coefficients (h) are not given as input to the LMS algorithm. The outputs of the
LMS algorithm are error signal (e) and identified filter output (w).
The simulation result of the python code given in Fig. 11.12 is displayed in
Fig. 11.13.
Inference
From Fig. 11.13, it is possible to observe that the adaptive filter result approaches the
original filter coefficients while increasing the number of iterations.
Task
Increase/decrease the length of the FIR filter and fix the number of iterations is 50.
Comment on the observed result.
Experiment 11.6 Inverse System Modelling Using LMS Algorithm
This experiment discusses the inverse system modelling using LMS algorithm. The
general block diagram of inverse system modelling using adaptive filter is shown in
Fig. 11.14. From this figure, it is possible to understand that the unknown system and
the adaptive filter are connected in a cascade form, and the delayed version of the
input signal act as a reference signal. The aim of adaptive filtering in this experiment
is to obtain the inverse system of the unknown system so that y[n] and d[n] will be
similar. If y[n] and d[n] are similar, then the adaptive filter is equal to the inverse of
the unknown system.
In communication systems, inverse system modelling is used as channel equal-
ization. In such scenario, the adaptive filter is termed as ‘equalizer’. Adaptive
equalizer can combat intersymbol interference. Intersymbol interference arises
because of the spreading of a transmitted pulse due to the dispersive nature of the
channel.
456 11 Adaptive Signal Processing

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
from scipy.fft import fft
mu,W=0.04,2.2 # learning rate,Channel Capacity
filt_order,t_samples,delay,trial=7,200,4,1000
noise_var,data_var=0.001,1
for i in range(0,trial):
inp=np.zeros(filt_order)
data=np.zeros(filt_order+t_samples)
v=np.zeros(filt_order+t_samples)
w=np.zeros(filt_order)
#Generation of random data and random noise
for j in range(filt_order-1,t_samples+filt_order):
data[j]=np.fix(np.random.rand(1)+0.5)*2-1
v[j]=np.fix(np.random.rand(1)+0.5)*2*np.sqrt(noise_var)-np.sqrt(noise_var)
# Impusle response of the channel
h=np.zeros(3)
for j in range(0,3):
h[j]=(1/2)*(1+np.cos(2*np.pi/W)*(j-(3-1)))
C_out=signal.convolve(h,data) # Output from Channel
Err_square=np.zeros(len(C_out))
data=np.append(np.zeros(len(h)-1), data)
v=np.append(np.zeros(len(h)-1), v)
C_outn=C_out+v;
[w,e]=LMS_algorithmm(C_outn,mu,filt_order,data,len(C_outn)-filt_order);
e=np.append(e,np.zeros(filt_order))
Err_square=Err_square+(e**2)
mse=Err_square/trial
plt.figure,plt.subplot(2,2,1),plt.stem(h),plt.title('Impulse Resp. of Channel filter')
plt.xlabel('n-->'),plt.ylabel('h1[n]'),plt.subplot(2,2,2),plt.stem(w),
plt.title('Impulse Resp. of Inverse filter'),plt.xlabel('n-->'),plt.ylabel('h2[n]')
cas=signal.convolve(w,h);#Cascade operation
mag=fft(cas);#Frequency Response
plt.subplot(2,2,3),plt.stem(cas),plt.title('Impulse Resp. of Cascaded filter'),plt.xlabel('n-->'),
plt.ylabel('h1[n]*h2[n]'),plt.subplot(2,2,4),plt.plot(np.abs(mag)),
plt.title('Mag. Resp. of Cascaded filter'),plt.xlabel('$\omega$-->'), plt.ylabel('|H($\omega$)|'),
plt.ylim(0,10),plt.tight_layout()

Fig. 11.15 Python code for Inverse system modelling

The impulse response of the channel is given by

1 2π
1 þ cos ð n - 2Þ , n = 1, 2, 3
h½ n] = 2 W ð11:7Þ
0, otherwise

In the above equation, ‘W’ represents the channel capacity. Higher value of ‘W’
implies that the channel is more complex.
The python code to obtain the inverse of unknown system using LMS algorithm
is given in Fig. 11.15, and its corresponding simulation result is shown in Fig. 11.16.
11.2 Adaptive Filter 457

Impulse Resp. of Channel filter Impulse Resp. of Inverse filter

1.5
0.5
1.0
h1[n]

h2[n]
0.5 0.0

0.0
0.0 0.5 1.0 1.5 2.0 0 2 4 6
n--> n-->
Impulse Resp. of Cascaded filter Mag. Resp. of Cascaded filter
1.0 10.0
7.5
h1[n]*h2[n]

|H(w)|
0.5 5.0

2.5
0.0
0.0
0 2 4 6 8 0 2 4 6 8
n--> w -->

Fig. 11.16 Simulation result of inverse system modelling

Inference
From Fig. 11.16, it is possible to perceive the following facts
1. The impulse response of the cascaded system is an impulse. This implies that the
cascade of channel filter and its inverse system results in an identity system.
2. The Fourier transform of an impulse response will result in a flat spectrum. This is
obvious by observing the spectrum of the cascaded system.
Task
1. Increase the order of the adaptive filter and obtain the impulse response of the
inverse system.

11.2.2 Normalized LMS Algorithm

The weight updation formula for the normalized LMS algorithm is given by

β
w½n þ 1] = wT ½n] þ e½n]x½n] ð11:8Þ
k x k2 þ c

where ‘β’ is a positive constant, which controls the convergence speed of the
algorithm. ‘c’ is a small regularization parameter; it is added with the norm of the
signal x[n] to avoid the divide by zero error.
458 11 Adaptive Signal Processing

# This code for NLMS algorithm

def NLMS_algorithmm(x,N,t,beta,c,n_iter):
# x = input data, N = Filter length t = reference data,
# beta = Convergence parameter, c = regularization constant,
# n_iter = number of iteration
N1=len(x)
w = np.zeros(N) # Initial filter
e = np.zeros(N1-N)
for n in range(0, n_iter):
xn = x[n+N:n:-1]
en = t[n+N] - np.dot(xn,w) # Error
mu=beta/((xn*(np.transpose(xn)))+c)#Learning rate update
w = w + mu * en * xn # Update filter (NLMS algorithm)
e[n] = en # Record error
return w,e

Fig. 11.17 Python code for NLMS algorithm

Experiment 11.7 Normalized LMS (NLMS) Algorithm

The python code for the normalized LMS algorithm is given in Fig. 11.17.
Inference
1. From Fig. 11.17, it is possible to observe that it is in the form of a function, and it
can be called for the adaptive signal processing applications whenever required.
2. Also, it is possible to know that step size or learning rate is not given as a direct
input to the function.
3. The step size is calculated using the input data, β and ‘c’.
Experiment 11.8 Inverse System Modelling Using NLMS Algorithm
This experiment is a repetition of the inverse system modelling experiment, which
was discussed earlier. Here, Experiment 11.6 is repeated with the same specifica-
tions, and NLMS is used for adaptive filtering instead of LMS algorithm. The python
code of this experiment is shown in Fig. 11.18, and its corresponding simulation
result is displayed in Fig. 11.19.
Inference
The following conclusions can be made from this experiment:
1. From this Fig. 11.19, it is possible to conclude that the cascade of channel and
inverse filter gives the impulse response as unit impulse sequence.
2. The magnitude response confirms that the cascaded filter spectrum is a dc.
3. Therefore, the channel filter and the adaptive filter are inverse to each other.
11.2 Adaptive Filter 459

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
from scipy.fft import fft
c,beta,W=1.5,0.25,2.2 # learning rate,Channel Capacity
filt_order,t_samples,delay,trial=7,200,4,1500
noise_var,data_var=0.001,1
for i in range(0,trial):
inp=np.zeros(filt_order)
data=np.zeros(filt_order+t_samples)
v=np.zeros(filt_order+t_samples)
w=np.zeros(filt_order)
#Generation of random data and random noise
for j in range(filt_order-1,t_samples+filt_order):
data[j]=np.fix(np.random.rand(1)+0.5)*2-1
v[j]=np.fix(np.random.rand(1)+0.5)*2*np.sqrt(noise_var)-np.sqrt(noise_var)
# Impusle response of the channel
h=np.zeros(3)
for j in range(0,3):
h[j]=(1/2)*(1+np.cos(2*np.pi/W)*(j-(3-1)))
C_out=signal.convolve(h,data) # Output from Channel
Err_square=np.zeros(len(C_out))
data=np.append(np.zeros(len(h)-1), data)
v=np.append(np.zeros(len(h)-1), v)
C_outn=C_out+v;
[w,e]=NLMS_algorithmm(C_outn,filt_order,data,beta,c,len(C_outn)-filt_order);
e=np.append(e,np.zeros(filt_order))
Err_square=Err_square+(e**2)
mse=Err_square/trial
plt.figure,plt.subplot(2,2,1),plt.stem(h),plt.title('Impulse Resp. of Channel filter')
plt.xlabel('n-->'),plt.ylabel('h1[n]'),plt.subplot(2,2,2),plt.stem(w),
plt.title('Impulse Resp. of Inverse filter'),plt.xlabel('n-->'),plt.ylabel('h2[n]')
cas=signal.convolve(w,h);#Cascade operation
mag=fft(cas);#Frequency Response
plt.subplot(2,2,3),plt.stem(cas),plt.title('Impulse Resp. of Cascaded filter')
plt.xlabel('n-->'),plt.ylabel('h1[n]*h2[n]'),plt.subplot(2,2,4),plt.plot(np.abs(mag)),
plt.title('Mag. Resp. of Cascaded filter'),plt.xlabel('$\omega$-->'),
plt.ylabel('|H($\omega$)|'),plt.ylim(0,10),plt.tight_layout()

Fig. 11.18 Python code for Experiment 11.8

11.2.3 Sign LMS Algorithm

The weight updation formula for Sign LMS algorithm is given by

w½n þ 1] = w½n] þ μ signfe½n]x½n]g ð11:9Þ

where ‘sign’ indicates the sign of the number, ‘w[n + 1]’ represents new weight and
‘e[n]’ denotes the error signal between target and estimated signal.
460 11 Adaptive Signal Processing

Impulse Resp. of Channel filter Impulse Resp. of Inverse filter

1.5
0.5
1.0
h1[n]

h2[n]
0.0
0.5

0.0 –0.5
0.0 0.5 1.0 1.5 2.0 0 2 4 6
n--> n-->
Impulse Resp. of Cascaded filter Mag. Resp. of Cascaded filter
1.0 10.0
7.5
h1[n]*h2[n]

|H(w)|
0.5 5.0

2.5
0.0
0.0
0 2 4 6 8 0 2 4 6 8
n--> w -->

Fig. 11.19 Simulation result of the python code given in Fig. 11.18

Adaptive
Algorithm
v[n]

+ y[n]
x[n] FIR d̂ [n] +
∑ z-k ∑
s[n] + Filter
-
d[n]

Fig. 11.20 Block diagram of adaptive line enhancer

Experiment 11.9 Adaptive Line Enhancer Using Sign LMS Algorithm

This experiment discusses the python implementation of adaptive line enhancer
using sign LMS algorithm. The block diagram of adaptive line enhancer is shown
in Fig. 11.20. From this figure, it is possible to observe that input to the FIR filter is a
noisy version of the input signal (x[n]), and the final output (y[n]) is the enhanced
input signal or noise-free signal. The aim of this experiment is to remove the noisy
components present in the input signal using sign LMS adaptive algorithm. The
python code for the “sign LMS algorithm” is given in Fig. 11.21 as a function.
The python code for adaptive line enhancer using sign LMS is given in
Fig. 11.22. In this experiment, the input signal has 500, 2000 and 3500 Hz frequen-
cies. The sampling frequency is considered as 8000 Hz. The input signal is added
with the external random noise, which is the input to the adaptive filter. The number
11.2 Adaptive Filter 461

# This Code for Sign LMS algorithm

def Sign_LMS_algorithmm(x,mu,N,t,n_iter):
# x = input data, mu = step size, t = reference data, N = Filter length
# n_iter = number of iteration
N1=len(x)
w = np.zeros(N) # Initial filter
e = np.zeros(N1-N)
for n in range(0, n_iter):
xn = x[n+N:n:-1]
en = t[n+N] - np.dot(xn,w) # Error
w = w + mu * np.sign(en * xn) # Update filter (LMS algorithm)
e[n] = en # Record error
return w,e

Fig. 11.21 Python code for Sign LMS algorithm

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
from scipy.fft import fft
f1,f2,f3,Fs=500,2000,3500,8000 # Signal and sampling freq
T=1/Fs
t=np.arange(0,1,T)
noise=np.random.randn(len(t));
d=np.sin(2*np.pi*f1*t)+np.sin(2*np.pi*f2*t)+np.sin(2*np.pi*f3*t)+noise;
delay,N,mu=10,25,0.001 # Delay,Filter length and step size
x=np.append(np.zeros(delay),d);
[w,e]=Sign_LMS_algorithmm(x,mu,N,d,len(t)-N)
y1=signal.convolve(w,x)
mag_x=fft(x)/len(x);#Frequency Response
mag_y=fft(y1)/len(y1);#Frequency Response
plt.figure(),plt.subplot(2,2,1),plt.plot(x),plt.title('Input noisy signal')
plt.xlabel('t-->'),plt.ylabel('x(t)')
plt.subplot(2,2,2),plt.plot(y1),plt.title('Denoised signal')
plt.xlabel('t-->'),plt.ylabel('y(t)')
plt.subplot(2,2,3),plt.plot(np.abs(mag_x[0:4000])),plt.title('Spectrum of noisy signal')
plt.xlabel('$\omega$-->'),plt.ylabel('|X($\omega$)|')
plt.subplot(2,2,4),plt.plot(np.abs(mag_y[0:4000])),plt.title('Spectrum of denoised signal')
plt.xlabel('$\omega$-->'),plt.ylabel('|Y($\omega$)|')
plt.tight_layout()

Fig. 11.22 Python code for adaptive line enhancer using sign LMS

of delay is chosen as 10, and length of the adaptive FIR filter is fixed as 25. The main
objective of this experiment is to recover or enhance the original signal from the
noisy input data using sign LMS algorithm. The simulation result of the python code
462 11 Adaptive Signal Processing

Input noisy signal Denoised signal

5 2

0 0
x(t)

y(t)
–2
–5
0 2000 4000 6000 8000 0 2000 4000 6000 8000
t--> t-->
Spectrum of noisy signal Spectrum of denoised signal
0.4 0.4

|Y(ω)|
|X(ω)|

0.2 0.2

0.0 0.0
0 1000 2000 3000 4000 0 1000 2000 3000 4000
ω--> ω-->

Fig. 11.23 Simulation result of the adaptive line enhancer using sign LMS

given in Fig. 11.22 is shown in Fig. 11.23. From the magnitude spectrum, it is
possible to observe that the noise impact is reduced by the sign LMS algorithm.
Inference
From this experiment, the following observations can be drawn:
1. From Fig. 11.23, the magnitude response of the noisy signal indicates that the
signal has three unique frequency components and noisy components.
2. The magnitude response of denoised signal has three spikes, and the impact of the
noisy components is lesser than the input magnitude response.
Task
1. Do the suitable adjustments in the parameters used in the python code given in
Fig. 11.22 to reduce the effect of noise in the denoised or enhanced signal?

11.3 RLS Algorithm

Recursive least square (RLS) is an adaptive algorithm based on the idea of least
squares. The block diagram of the adaptive filter based on RLS algorithm is shown in
Fig. 11.24. From the figure x[n] is the input to the filter, d[n] is the desired signal and
the difference between the desired signal and the output of the filter is the error signal
e[n]. Forgetting factor is used in RLS algorithm to remove or minimize the influence
of old measurements. A small forgetting factor reduces the influence of old samples
and increases the weight of new samples; as a result, a better tracking can be realized
at the cost of a higher variance of the filter coefficients. A large forgetting factor
11.3 RLS Algorithm 463

Fig. 11.24 Block diagram d[n]

of adaptive filter based on
RLS algorithm x[n] y[n] - +
Variable filter w[n]
e[n]

RLS adaptive algorithm

Initialization P[0] G 1 I ; w(0) 0

P[n  1]x[n]
Computation of gain k [ n]
O  x T [n]P[n  1]x[n]

Computation of error eˆ[n] d [n]  wT (n  1) x[n]

Weight updation w(n) w(n  1)  k[n]eˆ[n]

Updation of Inverse
Correlation matrix P[n] ^
O1 P[n  1]  k[n]x T [n]P[n  1] `

Fig. 11.25 Flow chart of sequence of steps in RLS algorithm

keeps more information about the old samples and has a lower variance of the filter
coefficients, but it takes a longer time to converge.
Let us define the a priori error as e½n] = d½n] - wT ½n - 1]x½n] and the weight
updation formula for the RLS algorithm is given by

P½n - 1]x½n]e½n]
w½n] = w½n - 1] þ ð11:10Þ
λ þ xT ½n]P½n - 1]x½n]

P½n - 1]x½n]
If k ½n] = λþxT ½n]P½n - 1]x½n] represents the gain, then the above expression can be
written as

w½n] = w½n - 1] þ k ½n]e½n] ð11:11Þ

The flow chart of the sequence of steps followed in RLS algorithm is shown in
Fig. 11.25. From the flow chart, it is possible to observe that the algorithm is
iterative. Proper initialization of filter coefficients is necessary for convergence.
464 11 Adaptive Signal Processing

# This Code for RLS algorithm

def RLS_algorithmm(x,lamda,delta,N,t,n_iter):
# x = input data, lamda = Forgetting factor, delta = Regularization parameter
# t = reference data, N = Filter length, n_iter = number of iteration
N1=len(x)
w = np.zeros(N) # Initial filter
w=np.transpose(w)
e = np.zeros(N1-N)
P=np.eye(N)/delta
x=np.transpose(x)
for n in range(0, n_iter):
xn = x[n+N:n:-1]
k1=np.dot(P,xn)
k2=np.dot(np.transpose(xn),P)
k3=np.dot(k2,xn)
k =k1/(lamda+k3)
en = t[n+N] - np.dot(np.transpose(w),xn);# Error
w = w + np.dot(k,np.conjugate(en)) # Update filter (RLS algorithm)
P=(1/lamda)*P
e[n] = en # Record error
return w,e

Fig. 11.26 Python code for RLS algorithm

Experiment 11.10 Implementation of RLS Algorithm

This experiment discusses the implementation of RLS algorithm using python. The
python code for RLS algorithm is given in Fig. 11.26, and it is in the form of a
function so that this function can be used for different applications.
Experiment 11.11 Adaptive Line Enhancer Using RLS Algorithm
This experiment is a repetition of Experiment 11.9; instead of sign LMS, RLS
algorithm is used to filter out the noisy component present in the input signal. The
python code for this experiment is given in Fig. 11.27, and its corresponding
simulation result is displayed in Fig. 11.28.
Inference
From Fig. 11.28, it is possible to confirm that the magnitude response of the filtered
or denoised output is better than the magnitude response of the noisy input. There-
fore, RLS algorithm can act as an adaptive line enhancer.
Experiment 11.12 Comparison of System Identification with Different Adaptive
Filters
The main objective of this experiment is to compare the simulation result of different
adaptive algorithms like LMS, NLMS, Sign LMS and RLS for the system identifi-
cation process. The python code to compare the simulation results of system
identification is given in Fig. 11.29, and its simulation results are shown in
Fig. 11.30.
11.3 RLS Algorithm 465

import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
from scipy.fft import fft
f1,f2,Fs=500,2000,8000 # Signal and sampling freq
T,lamda,delta=1/Fs,1.9,0.05
t=np.arange(0,1,T)
noise=np.random.randn(len(t));
d=np.sin(2*np.pi*f1*t)+np.sin(2*np.pi*f2*t)+noise;
delay,N=10,50 # Delay,Filter length
x=np.append(np.zeros(delay),d);
[w,e]=RLS_algorithmm(x,lamda,delta,N,d,len(d)-N)
y1=signal.convolve(w,x)
mag_x=fft(x)/len(x);#Frequency Response
mag_y=fft(y1)/len(y1);#Frequency Response
plt.figure(),plt.subplot(2,2,1),plt.plot(x),plt.title('Input noisy signal')
plt.xlabel('t-->'),plt.ylabel('x(t)')
plt.subplot(2,2,2),plt.plot(y1),plt.title('Denoised signal')
plt.xlabel('t-->'),plt.ylabel('y(t)')
plt.subplot(2,2,3),plt.plot(np.abs(mag_x[0:4000])),plt.title('Spectrum of noisy signal')
plt.xlabel('$\omega$-->'),plt.ylabel('|X($\omega$)|')
plt.subplot(2,2,4),plt.plot(np.abs(mag_y[0:4000])),plt.title('Spectrum of denoised signal')
plt.xlabel('$\omega$-->'),plt.ylabel('|Y($\omega$)|')
plt.tight_layout()

Fig. 11.27 Python code for adaptive line enhancer using RLS

Inference
From Fig. 11.30, it is possible to observe that proper selection of the adaptive filter
parameters like step size or learning rate, forgetting factor and regularization plays a
major role in using the adaptive filtering algorithm for the system identification
application in signal processing.
Task
Write a python code to compare the simulation result of different adaptive algo-
rithms like LMS, NLMS, sign LMS and RLS for adaptive line enhancement
application in signal processing.
Exercises
1. Execute the python code given in Fig. 11.12 and compare the estimated filter ‘w’
with the original filter coefficients ‘h’ for different length of the filter. Also,
execute the same python code and comment on the convergence of the LMS
algorithm with different values of learning rate ‘mu’, including negative value.
2. Use the python code for the sign LMS algorithm given in Fig. 11.22 to compute
the impulse response of the inverse filter and comment on the role of learning rate.
3. Modify the sign LMS algorithm based on the equation of the sign regressor
algorithm is given by w[n + 1] = w[n] + μe[n] sign {x[n]}, and compute the
impulse response of the inverse filter and comment on the simulation result.
466 11 Adaptive Signal Processing

Input noisy signal Denoised signal

5
2
x(t)

y(t)
0 0

–2

0 2000 4000 6000 8000 0 2000 4000 6000 8000

t--> t-->
Spectrum of noisy signal Spectrum of denoised signal
0.4

|Y(ω)|
|X(ω)|

0.2 0.2

0.0 0.0
0 1000 2000 3000 4000 0 1000 2000 3000 4000
ω--> ω-->

Fig. 11.28 Simulation result of the python code given in Fig. 11.27

# Python code for the comparison of adaptive algorithms for system identification
import numpy as np
import matplotlib.pyplot as plt
from scipy import signal
N1 = 1500 # Size of the Input data
N = 25 # Filter size
n=np.arange(0,N,1)
n_iter=200# it must be less than (N1-N)
x = np.random.randn(N1) # Input to the filter
h = signal.firwin(N, 0.25) # FIR filter to be identified
t = signal.convolve(x, h) # Target/desired signal
t = t + 0.01 * np.random.randn(len(t)) # with added noise
mu,mu1,beta,c,lamda,delta = 0.05,0.0005,0.05,1.5,1,0.25 # LMS step size
plt.figure(1),plt.title('Filter to be Identified')
plt.stem(h),plt.xlabel('n-->'),plt.ylabel('h[n]')
[w,e]=LMS_algorithmm(x,mu,N,t,n_iter);
[w1,e1]=NLMS_algorithmm(x,N,t,beta,c,n_iter)
[w2,e2]=Sign_LMS_algorithmm(x,mu1,N,t,n_iter)
[w3,e3]=RLS_algorithmm(x,lamda,delta,N,t,n_iter)
plt.figure(2),plt.subplot(2,2,1),plt.stem(n,w,'g'),plt.xlabel('n-->'),plt.ylabel('w[n]')
plt.title('Identified by LMS'),plt.subplot(2,2,2),plt.stem(n,w1,'k'),plt.xlabel('n-->'),
plt.ylabel('w[n]'),plt.title('Identified by NLMS'),plt.subplot(2,2,3),
plt.stem(n,w2,'r'),plt.xlabel('n-->'),plt.ylabel('w[n]'),plt.title('Identified by Sign LMS')
plt.subplot(2,2,4),plt.stem(n,w3,'b'),plt.xlabel('n-->'),plt.ylabel('w[n]')
plt.title('Identified by RLS'),plt.tight_layout()

Fig. 11.29 Python code for unknown system identification

11.3 RLS Algorithm 467

Identified by LMS Identified by NLMS

0.2 0.2

w[n]
w[n]
0.1 0.1
Filter to be Identified

0.25 0.0 0.0

0.20
0 10 20 0 10 20
0.15
n--> n-->
h[n]

0.10
Identified by Sign LMS Identified by RLS
0.05
0.00 0.2
0.02

w[n]

w[n]
0 5 10 15 20 25 0.1
n-->
0.00
0.0
0 10 20 0 10 20
n--> n-->

Fig. 11.30 Simulation result of the python code given in Fig. 11.29

4. Modify the sign LMS algorithm based on the equation of sign-sign LMS algo-
rithm is given by w[n + 1] = w[n] + μ sign {e[n]} sign {x[n]}, and compute the
impulse response of the inverse filter and comment on the simulation result.
5. Use the python code for RLS algorithm given in Fig. 11.26 to obtain the inverse
filter coefficients and comment on the simulation result. Also, comment on the
selection of the forgetting factor and regularization parameter.
Objective Questions
1. The filter which is based on the minimum mean square error criterion, is
A. Wiener filter
B. Window-based FIR filter
C. Frequency sampling-based FIR filter
D. Savitsky Golay filter
2. If ‘R’ is the autocorrelation matrix of the observed signal and ‘p’ represents the
cross-correlation between the desired signal and the observed signal, then the
expression for the Wiener-Hopf equation is
A. wopt = R × p
B. wopt = R + p
C. wopt = R - p
D. wopt = p/R
3. The weight update expression of the standard LMS algorithm is
A. w(n + 1) = w(n) + μx[n]e[n]
B. w(n + 1) = w(n) - μx[n]e[n]
C. w(n + 1) = w(n) + μx[n]e2[n]
D. w(n + 1) = w(n) - μx[n]e2[n]
468 11 Adaptive Signal Processing

4. If μ refers to the step size and λ refers to the eigen value of the autocorrelation
matrix, then the condition for convergence of LMS algorithm is given by
A. 0 < μ < 2
λmin
B. 0 < μ < 2
λmax
C. 0 < μ < 2
λ2max
D. 0 < μ < 2
λ2min

5. Statement 1: Wiener filter is based on the statistics of the input data.

Statement 2: Wiener filter is an optimal filter with respect to minimum mean
absolute error
A. Statements 1 and 2 are true
B. Statement 1 is correct, and Statement 2 is wrong
C. Statement 1 is wrong, Statement 2 is correct
D. Statements 1 and 2 are wrong
6. The filter which changes its characteristics in accordance with the environment is
termed as
A. Optimal filter
B. Non-linear filter
C. Adaptive filter
D. Linear filter

Bibliography

1. Simon Haykin, “Adaptive Filter Theory”, Pearson, 2008.

2. Bernard Widrow, Samuel D. Stearns, “Adaptive Signal Processing”, Pearson, 2002.
3. Dimitris G. Manolakis, Vinay K. Ingle, and Stephen M. Kogon, “Statistical and Adaptive Signal
Processing: Spectral Estimation, Signal Modeling, Adaptive Filtering and Array Processing”,
Artech House Publishers, 2005.
4. Behrouz F. Boroujey, “Adaptive Filters: Theory and Applications”, Wiley -Blackwell, 2013.
5. Alexandar D. Poularikas, “Adaptive Filtering”, CRC Press, 2015.
Chapter 12
Case Study

Learning Objectives
After completing this chapter, the reader should be familiar with the following
• Applications of signal processing in speech signals
• ECG signal analysis
• Power line signal analysis

Roadmap of the Chapter

The case study discussed in this section focuses on the application of signal
processing algorithms in the field of electrical and electronics engineering. Three
case studies discussed in this section are (1) speech recognition, (2) QRS detection
algorithm in ECG (3) power line signal analysis. Transform domain analysis of
speech signal is discussed in the first case study. Analysis of ECG signal is the focus
of second case study. Identification of different types of faults in power line signal is
done in the third case study.

12.1 Case Study 1: Speech Recognition Using MFCC

(Mel-Frequency Cepstral Coefficient)

Speech is the easiest and most widely used way of communication between humans.
The interaction between a human and a computer is typical in the current scenario in
the communication field. Communication between humans and computers can be
made possible only with the help of hardware devices like keyboards, touch screens,
mice, etc. However, humans prefer a more natural form of interaction than hardware
devices. The speech signal is the most profound means of communication human
beings use. For the human to human interaction, voice is the most significant feature,
which helps us to recognize the speaker and extract the information from the
speaker. The speech recognition system can be used to create documents from

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 469
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0_12
470 12 Case Study

Speaker
Modelling Speaker
Speech Feature Model
Signal Extraction Database
Pattern
Matching

Decision Decision/
Logic Output

Fig. 12.1 General block diagram of the speech recognition system

speech, saving more time and reducing the burden on a human. In general, voice
samples contain more information, including the person’s gender and age. We can
distinguish whether the voice belongs to a male or female, child or adult, based on
the voice samples. Also, sometimes it reflects the state of mind of the speaker. The
voice recognition method uses some of the information in the voice and identifies the
speaker. Voice recognition is a technique that detects a voice sample from unique
properties that may be acoustic or phonetic.
The general block diagram of speech recognition system contains feature extrac-
tion, speaker modelling and pattern matching method to identify the speaker, which
is illustrated in Fig. 12.1. In the first stage, a speech sample is considered an input to
the system. Here the speech sample would be noise-free; to remove the noise
components in the speech signal, preprocessing method can be used. The
preprocessed speech signal is the input to the feature extraction process in the second
stage of the system. Using the feature extraction approaches, some properties of
speech, like acoustic or phonetic features, are extracted. Finally, in the third stage,
training and testing of the speech recognition model is developed, which will give
the final decision of the system (i.e.,) which speech belongs to whom.

12.1.1 Speaker Identification

The process of identifying a speaker’s voice from a group of speakers is called

speaker identification. In this process, the voice of the input speaker is verified with
the voice stored already in the database, and best match can be obtained using a
pattern-matching algorithm. If the voice does not match the voices stored in the
database, then the voice is a new one. The new voice can be updated in the database.
The general block diagram of the speaker identification system is shown in Fig. 12.2.
12.1 Case Study 1: Speech Recognition Using MFCC (Mel-Frequency. . . 471

Similarity

Reference Model Identification

Input Voice Feature Speaker - 1 Maximum
Result
Sample Extraction Selection
(Speaker Id)
⁞

Similarity

Reference Model
Speaker - N

Fig. 12.2 Block diagram of speaker identification system

Input Feature Verification Result

Similarity Decision
Speech Extraction (Accepted/Rejected)

Reference
Model Speakers
(Database)

Fig. 12.3 General block diagram of a speaker verification system

12.1.2 Speaker Verification System

The process of authenticating a speaker based on the characteristics of voice samples

is called speaker verification. In this system, the final output will be either accepted
or rejected. The main applications of this system are military, aircraft and voice-
verified authentication areas. The block diagram of the speaker verification system is
shown in Fig. 12.3. From this figure, it is possible to confirm that the major blocks in
the speaker verification system are feature extraction, similarity identification and
decision-maker.
Feature extraction is the common block for both speaker recognition and verifi-
cation systems. In general, feature extraction approaches help us to extract some
good features and characteristics from the voice samples. The mel-frequency
cepstral coefficient (MFCC) is the most widely used feature extraction method.
More detail about the MFCC is discussed in the next section.
472 12 Case Study

12.1.3 Mel-Frequency Cepstral Coefficient (MFCC) Feature

The step-by-step procedure to compute the MFCC feature is discussed in this

section. This procedure contains six major steps involved in the computation of
MFCC; they are (1) pre-emphasis, (2) sampling and windowing, (3) fast Fourier
transform (FFT), (4) mel filter bank, (5) logarithmic function and (6) discrete cosine
transform. The block diagram of MFCC computation is depicted in Fig. 12.4.

12.1.3.1 Pre-emphasis

The voice samples are passed through a highpass filter, which is mathematically
expressed as

y½n] = x½n] - ax½n - 1] ð12:1Þ

where x[n] is the input voice samples, ‘a’ is the filter constant and it takes the value
between 0.9 to 1.0 and y[n] denotes the filtered voice samples. In this pre-emphasis
process, the input voice samples are passed through a highpass filter, and the filtered
output will emphasize the high-frequency component present in the input voice
samples.
The python code to read, normalize and display audio files is shown in Fig. 12.5,
and its simulation results are shown in Fig. 12.6.
From Fig. 12.6, it is possible to observe that the amplitude of the original audio
signal is [-20,000, 20,000], whereas in the normalized audio, the amplitude varies
from -1 to 1.

12.1.3.2 Sampling and Windowing

The speech signal is divided into small segments with a duration of 20–30 ms, which
are called ‘frames’. While splitting the input voice samples may be allowed to
overlap between the successive segments. Windowing can be used to avoid discon-
tinuity between consecutive segments. Also, the windowing technique smooths the
extreme samples in both starting and ending of the segments. The commonly used
windowing function is Hamming or Hanning. The process of windowing the input
sequence is mathematically written as

Input Voice Pre- Sampling & Mel Filter MFCC

DFT Log DCT
Samples emphasis Windowing Bank

Fig. 12.4 Block diagram of MFCC computation

12.1 Case Study 1: Speech Recognition Using MFCC (Mel-Frequency. . . 473

import numpy as np
from scipy.io import wavfile
import scipy.fftpack as fft
from scipy.signal import get_window
import IPython.display as ipd
import matplotlib.pyplot as plt
audio1 = "DSP_UV.wav"
sample_rate, audio = wavfile.read(audio1)
ipd.Audio(audio1)
audio2=audio[50000:100000]
duration = len(audio2)/sample_rate
print(f"Sample rate: {sample_rate}Hz")
print(f"Audio duration: {duration}s")
t = np.linspace(0,duration,len(audio2))
plt.figure(figsize=(15,6)),plt.plot(t,audio2),plt.xlabel("Time (s)"),plt.ylabel("Amplitude")
plt.title("Original Audio in Time domain"),plt.show()
#Normalizing to amplitude ranging between +1 and -1
normalizedAudio = audio2/np.max(np.abs(audio2))
plt.figure(figsize=(15,6)),plt.plot(t,normalizedAudio),
plt.xlabel("Time"),plt.ylabel("Amplitude")
plt.title("Normalized Audio in Time domain"),plt.show()

Fig. 12.5 Python code for read, normalize and display audio file

Original Audio in Time domain

20000
Amplitude

–20000
0.0 0.2 0.4 0.6 0.8 1.0
Time (s)
Normalized Audio in Time domain
1
Amplitude

–1
0.0 0.2 0.4 0.6 0.8 1.0
Time (s)

Fig. 12.6 Plot of original and normalized audio signal

474 12 Case Study

def frame_audio(normalizedAudio, FFT_size=2048, hop_size=10, sample_rate=8000):

audio = np.pad(normalizedAudio, int(FFT_size / 2), mode='reflect')
frame_len = np.round(sample_rate * hop_size / 1000).astype(int)
frame_num = int((len(audio) - FFT_size) / frame_len) + 1
frames = np.zeros((frame_num,FFT_size))
for n in range(frame_num):
frames[n] = audio[n*frame_len:n*frame_len+FFT_size]
return frames
hop_size = 15 #ms
FFT_size = 2048
audio_framed = frame_audio(audio, FFT_size=FFT_size, hop_size=hop_size,
sample_rate=sample_rate)
print(f"Framed audio shape: {audio_framed.shape}")
window = get_window("hann", FFT_size, fftbins=True)
plt.figure,plt.subplot(3,1,1),plt.plot(window)
plt.title("Hanning Window"),plt.xlabel("Samples"),plt.ylabel("Amplitude"),plt.grid(True)
audio_win = audio_framed*window
plt.subplot(3,1,2),plt.plot(audio_framed[72])
plt.xlabel("Samples"),plt.ylabel("Amplitude"),plt.title("Before Windowing")
plt.subplot(3,1,3),plt.plot(audio_win[72]),plt.xlabel("Samples"),plt.ylabel("Amplitude")
plt.title("After Windowing"),plt.tight_layout(),plt.show()

Fig. 12.7 Python code for framing and windowing audio signal

y½n] = x½n] × w½n] ð12:2Þ

where w[n] represents the windowing function.

The python code for framing and windowing of normalized audio signal is shown
in Fig. 12.7. From this figure, it is possible to understand that the first part is the
framing/segmenting/partitioning of the audio signal. Then Hanning window is used
for the windowing operation on the partitioned audio signal in the second part. The
simulation result of the python code given in Fig. 12.7 is shown in Fig. 12.8. From
this Fig. 12.8, it is evident the importance of the windowing concept. The
windowing method is used to smoothen the initial and end of the audio signal frame.

12.1.3.3 Discrete Fourier Transform (DFT)

DFT is a well-known transform to convert the time-domain information of speech

signal into frequency-domain information. Also, it extracts useful information/some
features of the speech signal without losing the information present in it.
The python code to convert a time-domain audio signal into a frequency-domain
magnitude spectrum is shown in Fig. 12.9. ‘FFT’ library is used here to compute the
12.1 Case Study 1: Speech Recognition Using MFCC (Mel-Frequency. . . 475

Hanning Window

Amplitude
1

0
0 500 1000 1500 2000
Samples
Before Windowing
Amplitude

0
–10000
0 500 1000 1500 2000
Samples
After Windowing
Amplitude

5000
0
–5000
0 500 1000 1500 2000
Samples

Fig. 12.8 Result of Hanning window and windowing audio signal

audio_winT = np.transpose(audio_win)
audio_fft = np.empty((int(1 + FFT_size // 2), audio_winT.shape[1]), dtype=np.complex64, order='F')
for n in range(audio_fft.shape[1]):
audio_fft[:, n] = fft.fft(audio_winT[:, n], axis=0)[:audio_fft.shape[0]]
audio_fft = np.transpose(audio_fft)
frameNo = 72
f_axis = fft.fftfreq(audio_framed[frameNo].size,1/sample_rate)[0:audio_framed[frameNo].size//2]
plt.figure
plt.plot(f_axis,2/audio_framed[frameNo].size*np.abs(audio_fft[frameNo][0:audio_framed[0].size//2]))
plt.title(f"FFT of Frame:{frameNo}")
plt.xlabel("Frequency"),plt.ylabel("Magnitude")
plt.show()
audio_power = np.square(np.abs(audio_fft))
print(audio_power.shape)
freq_min = 0
freq_high = sample_rate / 2
mel_filter_num = 10
print(f"Minimum frequency: {freq_min}"),print(f"Maximum frequency: {freq_high}")

Fig. 12.9 Magnitude spectrum computation using FFT

magnitude spectrum. The output of the python code given in Fig. 12.9 is displayed in
Fig. 12.10. Frame number 72 is displayed.
476 12 Case Study

DFT Magnitude Spectrum of Frame:72

1600

1400

1200

1000
Magnitude

800

600

400

200

0
0 5000 10000 15000 20000 25000
Frequency in (Hz)

Fig. 12.10 Magnitude spectrum of frame number 72

Frequency

Fig. 12.11 Frequency response of mel-frequency bandpass filter

12.1.3.4 Mel-Frequency Bandpass Filter

The mel-frequency bandpass filter is a triangular-shaped multiple bandpass filter.

DFT obtains the magnitude spectrum of the speech signal, and this spectrum is
multiplied with the set of triangular bandpass filters to smoothen the magnitude
spectrum of speech signal, which is expressed as

Y ½k ] = X ½k ] × H ½k ] ð12:3Þ

where H[k] is the magnitude spectrum of the triangular bandpass filter. The sample
triangular bandpass filter frequency responses are shown in Fig. 12.11. This figure
shows a set of triangular filters that are used to compute weighted sum of filter
spectral components so that the output of process approximates to a mel scale. Each
filter’s magnitude response is triangular in shape, and the magnitude is unity at the
centre frequency and decreases linearly to zero at the centre frequency of two
adjacent filters. The final output is the sum of its filtered spectral components.
12.1 Case Study 1: Speech Recognition Using MFCC (Mel-Frequency. . . 477

def freq_to_mel(freq):
return 2595.0 * np.log10(1.0 + freq / 700.0)

def met_to_freq(mels):
return 700.0 * (10.0**(mels / 2595.0) - 1.0)

def get_filter_points(fmin, fmax, mel_filter_num, FFT_size, sample_rate=8000):

fmin_mel = freq_to_mel(fmin)
fmax_mel = freq_to_mel(fmax)
print("MEL min: {0}".format(fmin_mel))
print("MEL max: {0}".format(fmax_mel))
mels = np.linspace(fmin_mel, fmax_mel, num=mel_filter_num+2)
freqs = met_to_freq(mels)
return np.floor((FFT_size + 1) / sample_rate * freqs).astype(int), freqs
filter_points, mel_freqs = get_filter_points(freq_min, freq_high, mel_filter_num, FFT_size,
sample_rate=sample_rate)
print(f"Filter Points : {filter_points}")

def get_filters(filter_points, FFT_size):

filters = np.zeros((len(filter_points)-2,int(FFT_size/2+1)))
for n in range(len(filter_points)-2):
filters[n, filter_points[n] : filter_points[n + 1]] = np.linspace(0, 1, filter_points[n + 1] -
filter_points[n])
filters[n, filter_points[n + 1] : filter_points[n + 2]] = np.linspace(1, 0, filter_points[n + 2] -
filter_points[n + 1])
return filters
filters = get_filters(filter_points, FFT_size)
plt.figure
for filter in filters:
plt.plot(filter)
plt.title("Mel Filterbank"),plt.xlabel("Frequency"),plt.ylabel("Weights")
plt.show()
enorm = 2.0 / (mel_freqs[2:mel_filter_num+2] - mel_freqs[:mel_filter_num])
filters *= enorm[:, np.newaxis]
plt.figure
for n in range(filters.shape[0]):
plt.plot(filters[n])
plt.title("Normalized Mel Filterbank"),plt.xlabel("Frequency"),plt.ylabel("Weights")
plt.show()

Fig. 12.12 Python code for design of mel-filter bank

The python code to design a mel-filter bank and normalized mel-filter bank is
shown in Fig. 12.12. The simulation result of the python code given in Fig. 12.12 is
shown in Fig. 12.13.
Figure 12.13 shows that the gain of the filters in the mel-filter bank is unity,
whereas, in the normalized mel-filter bank, the gain of the filters is different.
478 12 Case Study

Fig. 12.13 Mel-filter bank and normalized mel-filter bank

audio_filtered = np.dot(filters, np.transpose(audio_power))

audio_log = 10.0 * np.log10(audio_filtered)
plt.figure,plt.subplot(2,1,1),plt.plot(audio_filtered),
plt.xlabel("Time in (sec)"),plt.ylabel("Amplitude"),plt.title("Filtered Signal")
plt.subplot(2,1,2),plt.plot(audio_log),
plt.xlabel("Frequency in (Hz)"),plt.ylabel("Magnitude")
plt.title("Log Power Spectrum")
plt.tight_layout()

Fig. 12.14 Python code for the computation of cepstral components

12.1.3.5 Log Operation

The logarithmic function is used to compute cepstral components from the filtered
acoustic signal. The python code to compute the cepstral components is given in
Fig. 12.14, and its simulation result is shown in Fig. 12.15. This figure makes it
possible to understand the use of the logarithmic function for the MFCC computa-
tion. The amplitude of the filtered signal is too high; the role of the logarithmic
function is to reduce the amplitude level.

12.1.3.6 Discrete Cosine Transform (DCT)

The final mel-frequency cepstral coefficient is obtained by taking DCT on the

cepstral component, which is the output of the logarithmic function. The mathemat-
ical expression of the computation MFCC is given by
12.2 Case Study 2: QRS Detection in ECG Signal Using Pan-Tomkins Algorithm 479

1ell
Filtered Signal

Amplitude 1.0

0.5

0.0
0 2 4 6 8
Time in (sec)
Log Power Spectrum
100
Magnitude

50

0 2 4 6 8
Frequency in (Hz)

Fig. 12.15 Plot of filtered signal and its cepstral components

L-1
πn 1
C ½ n] = Y ½k] cos k- ð12:4Þ
k=0
N 2

where n = 0, 1, 2, . . ., N - 1 and N denote the number of triangular bandpass filters.

L represents the number of mel-scale cepstral coefficients. The primary use of DCT
is to extract the output of bandpass filters to generate mel scale coefficients, and also
it converts the frequency-domain spectrum into a time-domain signal. The outcome
of the DCT is called ‘mel-scale cepstral coefficients’ (MFCC). These MFCC act as a
feature of the voice signal, which helps for the different speech signal processing
applications.
The python code to compute the DCT of the logarithmic function output is given
in Fig. 12.16. The simulation result of the python code, which is shown in Fig. 12.16,
is displayed in Fig. 12.17.
These features fed into the different classification methods, like SVM, KNN,
random forest, etc., to identify and recognize the speakers.

12.2 Case Study 2: QRS Detection in ECG Signal Using

Pan-Tomkins Algorithm

Electrocardiogram (ECG or EKG) is the electrical indication of the contractile

process of the heart. Every heart contraction produces an electrical impulse captured
by electrodes placed on the skin. ECG gives information about the heart rate, rhythm
and morphology. ECG is characterized by a periodic wave sequence of P, QRS, J, T
and U wavelets associated with each heartbeat. The QRS complex has a high clinical
480 12 Case Study

def dct(dct_filter_num, filter_len):

basis = np.empty((dct_filter_num,filter_len))
basis[0, :] = 1.0 / np.sqrt(filter_len)
samples = np.arange(1, 2 * filter_len, 2) * np.pi / (2.0 * filter_len)
for i in range(1, dct_filter_num):
basis[i, :] = np.cos(i * samples) * np.sqrt(2.0 / filter_len)
return basis
dct_filter_num = 12
dct_filters = dct(dct_filter_num, mel_filter_num)
cepstral_coefficents = np.dot(dct_filters, audio_log)
cepstral_coefficents.shape
print(f"MFCCs : {cepstral_coefficents[:, 1:dct_filter_num+1]}")
plt.figure(figsize=(10,6))
c = plt.imshow(cepstral_coefficents[:,1:dct_filter_num+1], aspect='auto',
origin='lower',cmap='Spectral');
plt.title("Mel Frequency Ceptral Coefficiernt"),plt.ylabel("MFCC"),plt.xlabel("Frames")
plt.tight_layout(),plt.colorbar(c)

Fig. 12.16 Python code for the computation of DCT

Mel Frequency Ceptral Coefficiernt

80
10 60

8 40

20
MFCC

6
0

4 –20

–40
2
–60
0 –80
0 2 4 6 8 10
Frames

Fig. 12.17 MFCC output

significance, and its detection is the first stage of ECG signal processing. From the
position of the QRS complex, it is possible to obtain the positions of P and T waves.
The normal ECG signal with different intervals of wavelets is shown in Fig. 12.18.
From this figure, it is possible to know that, in particular, QRS complex as compared
to the other waves has the steepest slope, has the highest amplitude in most cases,
12.2 Case Study 2: QRS Detection in ECG Signal Using Pan-Tomkins Algorithm 481

Fig. 12.18 Normal ECG showing different waves

Bandpass Filter Differentiator Moving window

ECG Signal and Squaring Integrator
(5-15) Hz

Fiducial Mark and

QRS Output Threshold Selection

Fig. 12.19 Block diagram QRS Detection

lasts for less than 0.2 s, has a peak at R and is preceded by a P wave and succeeded
by a T wave for a normal ECG.
QRS detection consists of three major processing steps; they are (1) linear digital
filtering, (2) non-linear transformation and decision rule algorithms. Let us discuss
the Pan and Tompkins QRS detection algorithm in this case study. This algorithm
starts with the linear process, which includes a bandpass filter, a derivative operation
and moving window integration. The second step is the non-linear transformation,
which uses amplitude squaring. The final stage is a decision rule algorithm, which
includes adaptive thresholds and QRS detection. The block diagram of QRS com-
plex detection in ECG using the Pan and Tompkins algorithm is given in Fig. 12.19.

12.2.1 ECG Signal Preprocessing

The normal ECG signals are time-varying with small amplitude ranging from 10 μV
to 5 mV. The typical amplitude of the ECG signal is 1 mV and their frequencies vary
from 0.05 to 100 Hz. The ECG signal is mainly concentrated in the 0.05–35 Hz
482 12 Case Study

range. For the ECG signal analysis, the system requires a noise-free ECG signal to
get an accurate prediction. However, ECG signals are affected by various noises and
artifacts practically. The ECG analysis system’s first step is to remove its noise by
using the filter.

12.2.1.1 Bandpass filter

The bandpass filter is used to reduce the effect of muscle noise, powerline interfer-
ence, baseline wander and T wave interference. The desirable passband frequency to
maximize the QRS energy is approximately 5–15 Hz. In this case study, the
Butterworth filter is used with order 3, and passband frequency [0.5, 15]
Hz. Instead of bandpass filter, cascaded lowpass and highpass filters may be
preferred. Filtering the ECG signal, ‘butter’ and ‘lfilter’ python commands are
used here. The python code to read the ECG data and noise removal is given in
Fig. 12.20, and its corresponding output is shown in Fig. 12.21.
From Fig. 12.21, it is possible to observe that the raw ECG signal has shifted the
amplitude to 1 mV, whereas the filtered ECG signal amplitude is between -1
and +1.

# QRS peak detection in ECG using Pan-Tomkins algorithm

import numpy as np
import matplotlib.pyplot as plt
from scipy.signal import butter, lfilter
x0=np.loadtxt('ecg_data_1.csv',skiprows=1, delimiter=',')
signal_freq,f_low,f_high,filt_order = 250,0.5,15.0,3
y0=x0[:,1];
qrs_peak_value,noise_peak_value,threshold_value = 0.0,0.0,0.0;
refractory_period = 120 # Change proportionally when adjusting frequency (in samples).
qrs_peak_filtering_factor,noise_peak_filtering_factor,qrs_noise_diff_weight = 0.125,0.125,0.25;
# Detection results.
qrs_peaks_indices = np.array([], dtype=int)
noise_peaks_indices = np.array([], dtype=int)
#Bandpass filtering
Fs = 0.5 * signal_freq
low, high = f_low / Fs, f_high / Fs;
b, a = butter(filt_order, [low, high], btype="band")
y1 = lfilter(b, a, y0)#Band pass filtering
plt.figure(1),plt.subplot(2,1,1),plt.plot(y0),plt.title('Raw ECG Signal')
plt.xlabel('Time-->'),plt.ylabel('Amplitude in mV'),plt.subplot(2,1,2),plt.plot(y1),
plt.title('Filtered ECG Signal'),plt.xlabel('Time-->'),plt.ylabel('Amplitude in mV')
plt.tight_layout()

Fig. 12.20 Python code for linear filtering

12.2 Case Study 2: QRS Detection in ECG Signal Using Pan-Tomkins Algorithm 483

Raw ECG Signal

Amplitude in mV
3

1
0 500 1000 1500 2000 2500
Time-->
Filtered ECG Signal
Amplitude in mV

–1
0 500 1000 1500 2000 2500
Time-->

Fig. 12.21 Raw and filtered ECG signals

12.2.1.2 Derivative Process

After bandpass filtering, the ECG signal is differentiated to provide QRS complex
slope information. ‘np.ediff1d’ python command is used here to obtain the derivative
of the filtered ECG signal. The first-order derivative equation can be written as

x ½ n - 1] - x ½ n]
y ½ n] = ð12:5Þ
2

where x[n] denotes the input signal and y[n] represents the derivative output signal.

12.2.1.3 Squaring Operation

After the derivative, the resultant signal is squared point by point. The squaring
operation can be written as

y ½ n] = x2 ½ n ] ð12:6Þ

The squaring operation makes all the data points positive and does non-linear
amplification of the derivative output emphasizing the higher frequencies.
484 12 Case Study

12.2.2 Moving Window Integration

Moving window integration helps to obtain the waveform feature information in

addition to the slope of the R wave. The mathematical equation to perform moving
window integration is given by

1
y ½ n] = fx½n - ðN - 1Þ] þ x½n - ðN - 2Þ] þ ⋯ þ x½n]g ð12:7Þ
N

where N is the number of samples in the width of the integration window. In this case
study, the length of the integration window is chosen as 15. The length of the moving
window integration (N ) is important for QRS wave detection. Generally, the value of
N should be approximately the same as the widest possible QRS complex. If the
window length (N ) is large, the integration waveform will merge the QRS complex
and T wave together. If the window length (N ) is small, some QRS complexes will
produce several peaks in the integration waveform. This may cause difficulty in the
subsequent QRS detection. The python code for the derivative process, squaring and
moving window integration is shown in Fig. 12.22, and its corresponding output is
displayed in Fig. 12.23.

12.2.3 Fiducial Mark

The QRS wave corresponds to the rising edge of the integrated waveform. The time
duration of the rising edge is equal to the width of the QRS wave. A fiducial mark for
the temporal location of the QRS wave can be obtained from the rising edge, and the
desired waveform point is marked as peak of the R wave. The python code for the

# Python code for derivative, squaring, and moving window integration

y2=np.ediff1d(y1);#Derivative
y3=y2**2;#Squaring
integral_window=15;
y4=np.convolve(y3, np.ones(integral_window));#Moving Window Integration
plt.figure(2),plt.subplot(3,1,1),plt.plot(y2),plt.title('Derivative ECG output')
plt.xlabel('Time-->'),plt.ylabel('Amp. in mV')
plt.subplot(3,1,2),plt.plot(y3),plt.title('Squared Derivative output')
plt.xlabel('Time-->'),plt.ylabel('Amp. in mV')
plt.subplot(3,1,3),plt.plot(y4),plt.title('Moving window integration output')
plt.xlabel('Time-->'),plt.ylabel('Amp. in mV')
plt.tight_layout()

Fig. 12.22 Python code for derivative, squaring and moving window integration
12.2 Case Study 2: QRS Detection in ECG Signal Using Pan-Tomkins Algorithm 485

Derivative ECG output

Amp. in mV
0.25
0.00
–0.25
0 500 1000 1500 2000 2500
Time-->
Squared Derivative output
Amp. in mV

0.05

0 500 1000 1500 2000 2500

Time-->
Moving window integration output
Amp. in mV

0.5

0.0
0 500 1000 1500 2000 2500
Time-->

Fig. 12.23 Simulation result of python code is given in Fig. 12.22

# Fiducial mark - peak detection on integrated measurements.

spacing=50
kk = y4.size
y5 = np.zeros(kk + 2 * spacing)
y5[:spacing] = y4[0] - 1.e-6
y5[-spacing:] = y4[-1] - 1.e-6
y5[spacing:spacing + kk] = y4
peak_candidate = np.zeros(kk)
peak_candidate[:] = True
for s in range(spacing):
start = spacing - s - 1
h_b = y5[start: start + kk] # before
start = spacing
h_c = y5[start: start + kk] # central
start = spacing + s + 1
h_a = y5[start: start + kk] # after
peak_candidate = np.logical_and(peak_candidate, np.logical_and(h_c > h_b, h_c > h_a))
ind = np.argwhere(peak_candidate)
ind = ind.reshape(ind.size)#detected_peaks_indices
detected_peaks_vals=y4[ind]

Fig. 12.24 Python code for QRS peak detection

peak detection of the integrated ECG measurement is given in Fig. 12.24. From this
figure, it is possible to observe that the position of the peak value results in the
variable ‘ind’ and the detected peak value can be obtained in the variable
‘detected_peaks_vals’.
486 12 Case Study

12.2.4 Decision Rule Approach

The decision rule consists of adaptive threshold selection. The thresholds are
adjusted automatically based on the noise in the ECG signal. The adaptive two
thresholds (Th1 and Th2) are calculated using the equation given below.

Th1 = NPK þ 0:25ðSPK - NPKÞ

ð12:8Þ
Th2 = 0:5Th1

where NPK represents the running estimate of noise peak and SPK denotes the
running estimate of the signal peak, which are computed as

SPK = 0:125 Pk þ 0:875SPK if Pk is signal peak

NPK = 0:125 Pk þ 0:875 NPK if Pk is noise peak

where Pk denotes peak. A peak is a local maximum determined by observing when

the signal changes direction within a predefined time interval. The SPK is a peak the
algorithm has already established to be a QRS complex. The NPK is any peak that is
not related to the QRS. Here, the thresholds Th1 and Th2 are based on running
estimates of SPK and NPK. When a new peak is detected, it must first be classified as
a signal peak or noise peak. The peak is a signal peak if the peak exceeds Th1, and
the QRS is obtained using Th2. The python code for threshold selection and the final
QRS detection is given in Fig. 12.25.
The simulation result of the python code given in Fig. 12.25 is shown in
Fig. 12.26. From this figure, it is possible to observe that the R peak of the integrated
ECG signal is detected, and the R peak of the filtered ECG is also shown.
Inferences
The following inferences can be made from Figs. 12.21, 12.23 and 12.26.
1. The raw ECG signal is affected by the baseline wander (i.e. the base x-axis of a
signal moves up and down rather than straight).
2. The bandpass filtered signal shows that the baseline wander is removed and the x-
axis of the signal is in a straight line.
3. The derivated ECG signal output highlights the positive and negative peaks of the
ECG signal very clearly. The squared ECG signal displays that all the negative
peaks are brought up into the positive peak. These results can be found in
Fig. 12.23.
4. The QRS peak marked moving window integrated ECG signal and the final R
peak marked filtered ECG signal are displayed, which can be found in Fig. 12.26.
12.2 Case Study 2: QRS Detection in ECG Signal Using Pan-Tomkins Algorithm 487

for ind, detected_peaks_val in zip(ind, detected_peaks_vals):

try:
last_qrs_index = qrs_peaks_indices[-1]
except IndexError:
last_qrs_index = 0
if ind - last_qrs_index > refractory_period or not qrs_peaks_indices.size:
if detected_peaks_val > threshold_value:
qrs_peaks_indices = np.append(qrs_peaks_indices, ind)
qrs_peak_val = qrs_peak_filtering_factor * detected_peaks_val + \
(1 - qrs_peak_filtering_factor) * qrs_peak_value
else:
noise_peaks_indices = np.append(noise_peaks_indices, ind)
noise_peak_value = noise_peak_filtering_factor * detected_peaks_val + \
(1 - noise_peak_filtering_factor) * noise_peak_value
threshold_value = noise_peak_value + \
qrs_noise_diff_weight * (qrs_peak_value - noise_peak_value)
qrs_peaks_indices_fin=qrs_peaks_indices-8;
plt.figure(3),plt.subplot(2,1,1),plt.plot(y4),plt.title('Integrated ECG with QRS peak marked')
plt.scatter(x=qrs_peaks_indices, y=y4[qrs_peaks_indices], c="red", s=10, zorder=2)
plt.xlabel('Time-->'),plt.ylabel('Amplitude in mV')
plt.subplot(2,1,2),plt.plot(y1),plt.title('Filterd ECG with R peak marked')
plt.scatter(x=qrs_peaks_indices_fin, y=y1[qrs_peaks_indices_fin], c="red", s=10, zorder=2)
plt.xlabel('Time-->'),plt.ylabel('Amplitude in mV')
plt.tight_layout()

Fig. 12.25 Python code for adaptive threshold selection and QRS detection

Integrated ECG with QRS peak marked

0.6
Amplitude in mV

0.4

0.2

0.0
0 500 1000 1500 2000 2500
Time-->
Filterd ECG with R peak marked
Amplitude in mV

–1
0 500 1000 1500 2000 2500
Time-->

Fig. 12.26 Simulation result of the python code given in Fig. 12.25
488 12 Case Study

12.3 Case Study 3: Power Quality Disturbance Detection

Power quality refers to maintaining a sinusoidal power distribution bus voltage at

rated magnitude and frequency. The significant increase in non-linear load and the
increased usage of semiconductor devices, lighting controls, solid-state switching
devices, inverters and relaying equipment are causing non-linear loads, which lead to
power quality disturbances. The basic power quality disturbances (PQD) are voltage
sag, voltage swell, voltage interruption, harmonics, flickers, etc. A combination of
these disturbances can occur simultaneously. Table 12.1 summarizes the causes and
impact of power quality disturbances.
Effective detection and recognition of power quality disturbances are necessary to
ensure the reliability of electric power quality. This section focuses on the simulation
of the power quality disturbance and analysis of power quality disturbance signals,
which is depicted in Fig. 12.27.
The first step in this direction is to generate different types of power quality
disturbances. Mathematical models can be developed for different types of power

Table 12.1 Causes and effects of power quality disturbances

Power
quality
S. No. disturbance Causes Effects
1 Voltage sag 1. Inductive load 1. Tripping of sensi-
2. Switching on and off of large loads tive equipment
2. Tripping of motors
2 Voltage 1. Capacitor switching 1. Damage to insula-
swell 2. Switch off large loads tion and windings
2. Damage to power
supplies
3 Harmonics 1. Non-linear loads 1. Malfunctioning of
2. Rectifier type equipment relays and equipment
2. Capacitor failure
4 Momentary 1. Equipment failure 1. Loss of supply volt-
interruption 2. Control malfunction age to consumer
equipment
2. Shutdown of
computers
5 Flicker 1. Machinery with rapid fluctuations in load 1. Misoperation of
current or voltage relays and contactors
2. Loads that cause voltage fluctuation include 2. Neurological prob-
arc welding machines, arc furnaces, etc. lems in humans

Signal Processing
Simulation of PQD Analysis of PQD
Techniques

Fig. 12.27 Objectives of this section

12.3 Case Study 3: Power Quality Disturbance Detection 489

Table 12.2 Mathematical model of power quality disturbances

S. No. PQ disturbance Mathematical model Parameters
1 Voltage sag A(1 - α(u(t - t1) - u(t - t2)))sin 0.1 < α < 0.9
(ωt)
2 Voltage swell A(1 + α(u(t - t1) - u(t - t2)))sin 0.1 < α < 0.9
(ωt)
3 Harmonics Asin(ωt) + 7n = 3 αn sinðnωt Þ 0.05 < αn < 0.15
4 Momentary A(1 - α(u(t - t1) - u(t - t2)))sin 0.9 < α < 1
interruption (ωt)
5 Flicker (1 + λ sin(kωt)) * sin(ωt) 0.1< λ < 0.2;
5 < k < 50

quality disturbances. The second step is to analyse different types of power quality
disturbances using time-frequency and time-scale representation.
The mathematical models of different types of power quality disturbances are
given in Table 12.2.

12.3.1 Generation of Power Quality Disturbance

The generation of various power quality disturbances like voltage sag, voltage swell
and momentary interruption are simulated, and their results are plotted. The python
code which generates various power quality disturbances is shown in Fig. 12.28, and
the corresponding output is shown in Fig. 12.29.
The following observations can be drawn from Fig. 12.29:
1. The amplitude of a pure sine wave varies from -1 to +1. It is a sine wave of 50 Hz
frequency.
2. During the power quality disturbance ‘sag’, the amplitude of sine wave decreases
for a brief period of time.
3. During the power quality disturbance ‘swell’, the amplitude of sine wave
increases for a brief period of time.
4. During ‘momentary interruption’, the amplitude of sine wave approaches zero
value for a brief period of time.
5. During power quality disturbance, the characteristics of the signal (amplitude of
the signal) vary with respect to time; hence, the power quality disturbances can be
considered as a non-stationary signal.
490 12 Case Study

#Power Quality Disturbance

import numpy as np
import matplotlib.pyplot as plt
A,fs,f,N,ph = 1,1000,50,200,0
T = 1/fs
t = np.linspace(0,N*T,N)
#Pure sine wave
pure_sine=np.sin(2*np.pi*f*t+ ph)
#Power quality disturbance
sag = np.sin(2*np.pi*f*t+ ph) - 0.5*np.sin(2*np.pi*f*t+ ph)*((t<0.15)&(t>0.08))
swell = np.sin(2*np.pi*f*t+ ph) + 0.5*np.sin(2*np.pi*f*t+ ph)*((t<0.15)&(t>0.08))
mi = np.sin(2*np.pi*f*t+ ph) - 0.98*np.sin(2*np.pi*f*t+ ph)*((t<0.15)&(t>0.08))
plt.subplot(2,2,1),plt.plot(t,pure_sine),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Pure sine wave')
plt.subplot(2,2,2),plt.plot(t,sag),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Sag')
plt.subplot(2,2,3),plt.plot(t,swell),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Swell')
plt.subplot(2,2,4),plt.plot(t,mi),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Momentary interruption')
plt.tight_layout()

Fig. 12.28 Python code to simulate power quality disturbances

Pure sine wave Sag

1 1
Amplitude
Amplitude

0 0

–1 –1
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20
Time Time

Swell Momentary interruption

1
1
Amplitude
Amplitude

0 0

–1
–1
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20
Time Time

Fig. 12.29 Result of python code shown in Fig. 12.28

12.3 Case Study 3: Power Quality Disturbance Detection 491

import numpy as np
import matplotlib.pyplot as plt
A,fs,f,N,ph = 1,1000,50,200,0
T = 1/fs
t = np.linspace(0,N*T,N)
#Pure sine wave
pure_sine=np.sin(2*np.pi*f*t+ ph)
#Power quality disturbance
har = np.sin(2*np.pi*f*t+ ph)+0.2*np.sin(2*np.pi*3*f*t+ ph)+0.3*np.sin(2*np.pi*5*f*t+ ph)
lamda=0.2
k=50
flicker=(1+lamda*np.sin(k*2*np.pi*f*t+ ph)) * np.sin(2*np.pi*f*t+ ph)
#Plotting the results
plt.subplot(3,1,1),plt.plot(t,pure_sine),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Pure sine wave')
plt.subplot(3,1,2),plt.plot(t,har),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Harmonics')
plt.subplot(3,1,3),plt.plot(t,flicker),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Flicker')
plt.tight_layout()

Fig. 12.30 Python code to simulate harmonics and flicker

Pure sine wave

Amplitude

1
0
–1
0.000 0.025 0.050 0.075 0.100 0.125 0.150 0.175 0.200
Time
Harmonics
Amplitude

1
0
–1
0.000 0.025 0.050 0.075 0.100 0.125 0.150 0.175 0.200
Time
Flicker
Amplitude

1
0
–1
0.000 0.025 0.050 0.075 0.100 0.125 0.150 0.175 0.200
Time

Fig. 12.31 Result of python code shown in Fig. 12.30

12.3.2 Simulation of Power Quality Disturbance

The other power quality disturbances like ‘harmonics’ and ‘flicker’ are simulated
using python. The python code which performs this task is shown in Fig. 12.30, and
the corresponding output is shown in Fig. 12.31.
492 12 Case Study

From Fig. 12.31, the following inferences can be drawn

1. A harmonic is an integer multiple of fundamental frequency. The fundamental
frequency of the sinusoidal signal is 50 Hz. The odd harmonics of the signal are
added to the original sinusoidal signal to obtain the harmonic signal.
2. The term ‘flicker’ implies the effect the fluctuations in electric voltage have on
electrical lighting devices. Loads, like arc furnaces, saw mills, welding machines
and high-powered engines with fast stop-and-start cycle, can give rise to the
phenomenon of flicker.

12.3.3 Time-Frequency Representation of Power Quality

Disturbance

Time-frequency representation is a good tool to analyse non-stationary signal.

Spectrogram is a square magnitude of short-time Fourier transform (STFT). STFT
gives two-dimensional representation of a one-dimensional signal. In this section,
the built-in function ‘plt.specgram’ is used here to obtain the time-frequency repre-
sentation of different types of power quality disturbance. The choice of the window
function and the width of the window function is important in obtaining good time-
frequency representation so that one obtains good time and frequency resolution.
The python code to obtain the time-frequency plot of normal sinusoidal waveform
and sinusoidal waveform with sag is shown in Fig. 12.32, and the corresponding
output is shown in Fig. 12.33.

#Time-Frequency representation of sag

import numpy as np
import matplotlib.pyplot as plt
A,fs,f,N,ph = 1,1000,50,256,0
T = 1/fs
t = np.linspace(0,N*T,N)
#Pure sine wave
pure_sine=np.sin(2*np.pi*f*t+ ph)
#Generation of sag
sag = np.sin(2*np.pi*f*t+ ph) - 0.5*np.sin(2*np.pi*f*t+ ph)*((t<0.15)&(t>0.08))
#Plotting the signal and its STFT
plt.subplot(2,2,1),plt.plot(t,pure_sine),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Pure sine wave')
plt.subplot(2,2,2),plt.plot(t,sag),plt.xlabel('Time'), plt.ylabel('Amplitude'),plt.title('Sag')
plt.subplot(2,2,3),plt.specgram(pure_sine, Fs=fs, NFFT=32, noverlap=1,window =None)
plt.xlabel('Time'),plt.ylabel('Frequency'),plt.title('Spectrogram of sine wave')
plt.subplot(2,2,4),plt.specgram(sag, Fs=fs, NFFT=32, noverlap=1,window =None)
plt.xlabel('Time'),plt.ylabel('Frequency'),plt.title('Spectrogram of Sag')
plt.tight_layout()

Fig. 12.32 STFT of sinusoidal signal and sinusoidal signal with ‘sag’
12.3 Case Study 3: Power Quality Disturbance Detection 493

Pure sine wave Sag

1 1
Amplitude

Amplitude
0 0

–1 –1
0.0 0.1 0.2 0.0 0.1 0.2
Time Time
Spectrogram of sine wave Spectrogram of Sag

400 400
Frequency

200 Frequency 200

0 0
0.05 0.10 0.15 0.20 0.05 0.10 0.15 0.20
Time Time

Fig. 12.33 Result of python code shown in Fig. 12.32

From Fig. 12.33, it is possible to infer that the short-time Fourier transform of
pure sinusoidal signal shows a horizontal line at 50 Hz, which indicates that the
generated sinusoidal signal has a 50 Hz frequency component. The STFT of the sag
waveform clearly indicates the starting and ending of the sag in the sinusoidal
waveform.
The python code to obtain the time-frequency representation of a sinusoidal
signal with momentary interruption is given in Fig. 12.34, and its simulation result
is depicted in Fig. 12.35.
Figure 12.35 represents the time-frequency representation of momentary inter-
ruption in a power line signal. Momentary interruption refers to zeroing of the
amplitude of the sinusoidal signal for a brief period of time. From Fig. 12.35, it is
possible to observe that the time-frequency representation of momentary interruption
is different from the time-frequency representation of the pure sinusoidal waveform.
Thus, time-frequency representation clearly distinguishes pure sinusoidal signal
from momentary interruption.

12.3.4 Time-Scale Representation of Power Quality

Disturbance

Time-scale representation can be obtained using wavelet transform. Wavelet trans-

form has the ability to perform multi-resolution analysis of the signal. In this section,
494 12 Case Study

#Time-Frequency representation of MI
import numpy as np
import matplotlib.pyplot as plt
A,fs,f,N,ph = 1,1000,50,256,0
T = 1/fs
t = np.linspace(0,N*T,N)
#Pure sine wave
pure_sine=np.sin(2*np.pi*f*t+ ph)
#Generation of sag
mi = np.sin(2*np.pi*f*t+ ph) - 0.98*np.sin(2*np.pi*f*t+ ph)*((t<0.15)&(t>0.08))
#Plotting the signal and its STFT
plt.subplot(2,2,1),plt.plot(t,pure_sine),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Pure sine wave')
plt.subplot(2,2,2),plt.plot(t,mi),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Momentary interruption')
plt.subplot(2,2,3),plt.specgram(pure_sine, Fs=fs, NFFT=16, noverlap=1,window =None)
plt.xlabel('Time'),plt.ylabel('Frequency'),plt.title('Spectrogram of sine wave')
plt.subplot(2,2,4),plt.specgram(mi, Fs=fs, NFFT=16, noverlap=1,window =None)
plt.xlabel('Time'),plt.ylabel('Frequency'),plt.title('Spectrogram of MI')
plt.tight_layout()

Fig. 12.34 Python code for Time-frequency representation of momentary interruption

Pure sine wave Momentary interruption

1 1
Amplitude

Amplitude

0 0

–1 –1
0.0 0.1 0.2 0.0 0.1 0.2
Time Time
Spectrogram of sine wave Spectrogram of MI

400 400
Frequency

Frequency

200 200

0 0
0.1 0.2 0.1 0.2
Time Time

Fig. 12.35 Time-frequency representation of momentary interruption

12.3 Case Study 3: Power Quality Disturbance Detection 495

#Scalogram of Momentary Interruption

import numpy as np
import matplotlib.pyplot as plt
import pywt
A,fs,f,N,ph = 1,1000,50,200,0
T = 1/fs
t = np.linspace(0,N*T,N)
#Pure sine wave
pure_sine=np.sin(2*np.pi*f*t+ ph)
#Momentary Interruption (MI)
mi = np.sin(2*np.pi*f*t+ ph) - 0.98*np.sin(2*np.pi*f*t+ ph)*((t<0.15)&(t>0.08))
#Scalogram
scale = [1.,2.]
coef1,freqs1=pywt.cwt(pure_sine,scale,'gaus1')
coef2,freqs2=pywt.cwt(mi,scale,'gaus1')
#Plotting the result
plt.subplot(2,2,1),plt.plot(t,pure_sine),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('Pure sine wave')
plt.subplot(2,2,2),plt.plot(t,mi),plt.xlabel('Time'), plt.ylabel('Amplitude'),
plt.title('MI'),plt.subplot(2,2,3),
plt.imshow(abs(coef1),extent=[0,200,10,1],interpolation='bilinear',cmap='bone',
aspect='auto',vmax=abs(coef1).max(),vmin=-abs(coef1).max())
plt.gca().invert_yaxis(),plt.xticks(np.arange(0,201,25))
plt.xlabel('Time'),plt.ylabel('Scale'), plt.title('Scalogram of Sinewave'),plt.subplot(2,2,4),
plt.imshow(abs(coef2),extent=[0,200,10,1],interpolation='bilinear',cmap='bone',
aspect='auto',vmax=abs(coef2).max(),vmin=-abs(coef2).max())
plt.gca().invert_yaxis(),plt.xticks(np.arange(0,201,25))
plt.xlabel('Time'),plt.ylabel('Scale'), plt.title('Scalogram of MI')
plt.tight_layout()

Fig. 12.36 Python code to obtain the scalogram of momentary interruption

time-scale representation of power quality disturbance is obtained using continuous

wavelet transform. The library ‘pywavelet’ is used here to obtain the scalogram of
the signal. Scalogram represents the square magnitude to continuous wavelet trans-
form. The choice of the wavelet and the scale are important in obtaining good time-
scale representation of the signal. The python code which obtains the time-scale
representation of ‘momentary interruption (MI)’ is shown in Fig. 12.36, and the
corresponding output is shown in Fig. 12.37.
From Fig. 12.37, it is possible to observe that the scalogram of signal with
momentary interruption is different from the scalogram of a normal sinusoidal
signal. This implies that continuous wavelet transform at the proper scale can
distinguish power quality disturbances from the normal signal.
496 12 Case Study

Pure sine wave MI

1 1
Amplitude

Amplitude
0 0

–1 –1
0.0 0.1 0.2 0.0 0.1 0.2
Time Time

Scalogram of Sinewave Scalogram of MI

10.0 10.0

7.5 7.5
Scale

Scale
5.0 5.0

2.5 2.5

0 25 50 75 100 125 150 175 200 0 25 50 75 100 125 150 175 200
Time Time

Fig. 12.37 Scalogram of momentary interruption

Bibliography

1. Lawrence Rabiner and Ronald Schafer, “Theory and Applications of Digital Speech Processing”,
Pearson, 2010.
2. Rangaraj M. Rangayyan, “Biomedical Signal Analysis: A Case-Study Approach”, Wiley-
Blackwell, 2002.
3. Tompkins Willis J, “Biomedical Digital Signal Processing: C-Language Examples and Labora-
tory Experiments for the IBMR PC”, Prentice Hall India Learning Private Limited, 1998.
4. Waldemar Rebizant, Janusz Szafran, and Andrzej Wiszniewski, “Digital Signal Processing in
Power System Protection and Control”, Springer, 2011.
5. S. Esakkirajan, T. Veerakumar and Badri N Subudhi, “Digital Signal Processing”, McGraw
Hill, 2021.
Appendix

Chapter 1: Generation Of Continuous-Time Signals

Answers to PreLab Questions

1. A continuous-time signal can have infinite number of values in a range. Room

temperature as a function of time is considered to be continuous-time signal.
Speech signal is considered as a continuous-time signal.
2. The built-in functions in the numpy library to create an array of numbers to
generate independent variable like time are (a) np.linspace() and (b) np.arange.
The linspace is a built-in function available in numpy library to create an evenly
spaced sequence of numbers in a specified interval. The syntax of linspace is:
np.linspace(start, stop, num, endpoint, retstep, dtype)
In the above syntax, ‘start’ represents the starting value of sequence, ‘stop’
represents the end value of the sequence and ‘num’ represents the number of
values to generate, and it has to be non-negative. The end point can be either
‘true’ or ‘false’. If it is ‘true’, the stop is the last sample, it is ‘false’ then end
point value is excluded. The ‘retstep’ can be either true or false. If it is true,
return (samples, step), where step is the spacing between samples. The ‘dtype’ is
the data type of the output array. If ‘dtype’ is not specified, then it infers the data
type from the other input arguments.
Example: np.linspace(-1,1,5) returns ‘array([-1. , -0.5, 0. , 0.5, 1. ])’ where
‘-1’ is the start value, ‘1’ is the end value. Five sample values are generated,
including -1 and +1. The difference between sample values is uniform.
Note: Similar to ‘np.linspace’, we have ‘np.logspace’, which is used to create
an array of evenly spaced numbers on a log scale value.
The syntax of np.arangebuilt-in function is given by
np.arange(start, stop ,step, dtype)
The ‘start’ and ‘stop’ represent the beginning and the end value of the
interval. The ‘step’ denotes the step size of the interval. The ‘dtype’ represent

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 497
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0
498 Appendix

the data type of the output array. The length of the array can be computed using
the command ceil((stop-start)/step).
Example: np.arange(0,1,0.1) generates array of numbers as array([0. , 0.1,
0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]). The length of the array is 10. By default,
'np.arange' command does not allow to include the end point value.
3. Most of the real-world phenomenon like motion of pendulum, under damped
spring-mass system can be expressed as sinusoidal signal. The sinusoidal signal
is a periodic signal, which varies smoothly with respect to time. According to
Fourier series, it is possible to represent periodic signals as sum of sinusoids.
Also, sinusoidal signals are eigen functions of linear time-invariant systems.
4. The term ‘phase’ refers to position of the waveform with respect to the origin.
The phase of the signal is measured in degrees or radians.
5. Multidimensional signals require more than one independent variable to repre-
sent the signal. Examples of multidimensional signal include (a) grey scale
image, (b) colour image and (c) video. Grey scale image is represented as
f(x, y), where ‘x’ and ‘y’ are termed as spatial variable. The colour image is
represented as f(x, y, λ), where ‘λ’ represents colour information. The video
signal is basically sequence of image, which is represented as f(x, y, λ, t). The
video signal is characterized by both spatial and temporal information.
6. (a) The equation of current through the diode is given as
VD
I D = I s eηV T - 1

In the above expression, VD is the voltage across the diode and ID is the
current through the diode; VT is the volt-equivalent of temperature, which is
26 mV at room temperature, and η is the ideality factor, which is material
dependent. Thus, the current through the diode is modelled as an exponential
function. In this case, it is an exponentially growing function.
(b) The equation for radioactive decay is expressed as

A = A0 e - λt

where ‘A’ is the ending activity and A0 is the initial activity, λ = 0:693
T1 , where
2
T 12 is the half-life period of the element. Thus, the radioactive decay is modelled
as exponentially decaying function.
7. Few significant features of complex exponential function are
Appendix 499

(a) Complex exponential function is the basis function of Fourier transform.

(b) It is a complex valued signal that simultaneously encapsulates both sine and
cosine signal by posting them on the real and imaginary components of the
complex signal.
(c) Complex exponentials are Eigen functions of continuous-time linear time-
invariant system.
8. Sinc function is mathematically defined as

sinðπt Þ
sin cðt Þ =
πt

The sinc function is an even function sinc(-t) = sin c(t). Few significant
features of sinc function are
(a) Sinc functions are used in the interpolation of signals
(b) Sinc and rectangular functions are dual function. Fourier transform of sinc
function results in rectangular function and vice versa.
9. A stationary signal is one whose statistical characteristics do not vary with
respect to time. Example is x(t) = A sin (2πft + ϕ). Here the frequency of the
signal does not change with respect to time. It is considered as stationary. A non-
stationary signal is one whose statistical characteristics change with respect to
time. Example of non-stationary signal includes y(t) = A sin (2πft2). Here the
frequency of the signal changes with respect to time. It is considered as
non-stationary. Chirp signal is considered as non-stationary signal.
10. The Gaussian function is characterized by two parameters, which are (a) mean
and (b) standard deviation. Few significant features of Gaussian functions are
(a) Gaussian functions are used as smoothing functions. The extent of smooth-
ing is governed by the standard deviation.
(b) Fourier transform of a Gaussian function result in another Gaussian
function.

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10
Key D C A B C B B D B A
500 Appendix

Chapter 2: Sampling and Quantization of Signals

Answers to PreLab Questions

1. The steps involved in converting the analogue signal into a digital signal are
(a) sampling, (b) quantization and (c) encoding. Before sampling, it is necessary
to ensure that the signal to be sampled is bandlimited. Sampling converts a
continuous-time signal into a discrete-time signal. Quantization is basically
mapping a large set of values to a smaller set of values. In quantization, the
discrete-time signal is converted to a quantized signal. In encoding, the quan-
tized signal is converted to a digital code.
2. The sampling theorem specifies the minimum sampling rate so that the sampled
signal can be reconstructed from its samples without aliasing problem. In order
to reconstruct the signal from the samples without an aliasing problem, the
sampling frequency must be greater than twice the maximum frequency content
of the signal. This is expressed as fs ≥ 2fmax, where fs represents the sampling
frequency and fmax represents the maximum frequency content of the signal.
3. Suppose the signal is a periodic and ideal interpolation is employed. In that case,
all spectral components less than fs/2 (where fs represents the sampling fre-
quency) are reconstructed perfectly, but all higher-frequency spectral compo-
nents are aliased to a frequency less than fs/2.
4. The square wave is not a bandlimited signal. It is not possible to reconstruct the
square wave from its samples.
5. Two prominent reasons for aliasing while performing sampling are
(a) undersampling and (b) signal which is not a bandlimited signal. Here
undersampling implies that the sampling rate fs < 2fmax, where fmax is the
maximum signal frequency.
6. Sampling is basically taking specific instants of the signal. The sampling rate (fs)
is the number of samples per second. Sampling interval (Ts) is the time-interval
between two consecutive samples.
7. Nyquist rate = 2B = 2 × 5 kHz = 10 kHz.
8. Quantization is mapping a large set of values to a smaller set of values. It will not
obey the superposition principle; hence, it is considered as non-linear
phenomenon.
9. Quantization maps a large set of values to a smaller set of values. It is not a one-
to-one mapping; hence, error is inevitable, and it is considered as irreversible
phenomenon. The meaning is that it is difficult to get the original signal exactly
after quantization.
10. The process of converting sampled data sequences to a continuous-time signal is
termed as signal reconstruction. Different strategies include (a) zero-order hold,
(b) first-order hold or linear interpolation and (c) ideal or sinc interpolation.
Appendix 501

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Key B A A B D A B C B D B C D B C

Chapter 3: Generation and Operation on Discrete-Time

Sequence

Answers to PreLab Questions

1. Two important steps involved in converting continuous-time signal into a

discrete-time signal are (a) sampling and (b) quantization.
2. Different forms of representation of discrete-time signals are (a) graphical form,
(b) functional form, (c) sequential form and (d) tabular form.
3. Some standard discrete-time sequences are (a) unit sample sequence (δ[n]),
(b) unit step sequence (u[n]), (c) unit ramp sequence (r[n]) and (d) exponential
sequence, which can be broadly classified as real exponential sequence and
complex exponential sequence.
4. Some of the salient features of unit sample sequence are:
(a) Any arbitrary signal x[n] can be expressed in terms of scaled and shifted
versions of unit sample sequences. This is expressed as
1
x ½ n] = x½k]δ½n - k ].
k= -1
(b) Convolution of signal x[n] with unit sample signal will result in the signal x
[n]. This is expressed as x[n] * δ[n] = x[n].
(c) If unit sample sequence is applied to linear time-invariant discrete-time
system, then the output of the system is termed as the impulse response of
the system. This is illustrated in Fig. A.1. From Fig. A.1, it is possible to
observe that the input to LTI discrete-time system is unit sample signal;
then, the output of the system is termed as the impulse response of the
system.

5. It is to be noted that linear time-invariant discrete-time system is characterized

by its impulse response. That is, by knowing the impulse response of the system,
it is possible to know the properties of the system like causality and stability.

Fig. A.1 Impulse response x[n] = δ[n] y[n] = h[n]

of the system LTI System
502 Appendix

Table A.1 Descriptions of energy and power signals

S. no. Energy signal Power signal
1 For a discrete-time energy signal, the For a discrete-time power signal, the
energy is finite and non-zero power is finite and non-zero.
2 Non-periodic signals are energy signals Periodic signals are power signals
3 Power of energy signal is zero Energy of power signal is infinite
4 Examples of DT energy signal are Examples of power signals are
x[n] = anu[n], |a| < 1 Unit step signal
x[n] = u[n] - u[n – 1] x[n] = Asin(ωon + φ)

x[n] y[n]
h[n]

Fig. A.2 Linear time-invariant system

ω
6. Discrete-time signal x[n] is periodic if it obeys the condition 2π = Nk . Here, ‘ω’
represents the angular frequency, and ‘N’ represents the fundamental period.
7. The energy and power signal descriptions are summarized in Table A.1.
8. Various mathematical operations that can be performed on DT signal include
(a) folding or time reversal; (b) shifting operation, which include delay and
advance operation; and (c) scaling operation, which could be time scaling
operation, like downsampling and upsampling, and amplitude scaling operation.
9. A DT signal is even if it obeys the condition x[-n] = x[n]. A DT signal is odd if
x[-n] = -x[n]. Example of even signal is cosine signal, whereas example of
odd signal is sinusoidal signal. Example of signal which is neither even and nor
odd includes unit step signal and unit ramp signal.
10. An energy signal has finite energy, whereas power signals have finite average
power. There are certain signals, which are neither energy nor power signal.
Example of finite energy signal is x[n] = (1/2)nu[n]. Example of power signal is
unit step signal. Example of signal which is neither energy nor power signal is
unit ramp signal.
11. Convolution is one of the most important operations in signal processing. The
three main mathematical operations involved in convolution are
(a) multiplication, (b) addition and (c) shifting operation. Convolution basically
performs filtering operation. It is represented in Fig. A.2
In Figure A.2, the input and output signals are represented as x[n] and y[n]
respectively. The impulse response of linear time-invariant system is denoted as
h[n]. The nature of filtering is decided by h[n]. If h[n] = {0.5, 0.5}, the system
behaves like a lowpass filter; on the other hand if h[n] = {0.5, -0.5}, the system
behaves like a highpass filter. The nature of filtering is decided by the impulse
response of the system. The expression for the output of the system is given by y
[n] = x[n] * h[n], where ‘*’ denotes the convolution operation.
The correlation between two signals x1[n] and x2[n] is given by
Appendix 503

r x1x2 ðlÞ = x1 ½n] * x2 ½ - n]

Convolving the folded version of sequence x2[n] with the sequence x1[n]
results in correlation of the signal. The correlation can be broadly classified into
(a) autocorrelation and (b) cross-correlation. Autocorrelation is finding the
relative similarity of the signal with itself.
Applications of correlation are summarized below:
(a) Correlation is used to find the relative similarity between signals.
(b) Fourier transform of autocorrelation function gives the power spectral
density of the signal. This is regarded as Wiener-Khinchin theorem.
(c) Correlation can be used to estimate the pitch of the speech signal.
(d) Correlation can be used for template matching.

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Key B B C B B C B C D C B B C D B

Q. No. 16 17 18 19 20 21 22 23 24 25
Key B B A B A D C B

Chapter 4: Discrete-Time Systems

Answers to PreLab Questions

1. Different forms of representation of discrete-time systems are (a) block diagram,

(b) difference equation, (c) transfer function, (d) impulse response, (e) pole-zero
plot and (f) state-space.
2. A discrete-time system is said to be a relaxed system if zero input results in zero
output. If x[n] = 0, then the corresponding output y[n] should be zero.
3. A discrete-time system is linear if it obeys superposition theorem. Superposition
theorem implies (a) homogeneity property and (b) additivity property.
According to homogeneity property, scaling of the input results in scaling of
the output. According to additivity property, the response of the system to sum
of inputs must be equal to sum of the individual responses. Examples of linear
504 Appendix

system are as follows: (a) y[n] = x[-n], (b) y[n] = nx[n], (c) y[n] = Ax[n], (d) y
[n] = x[Mn] and (e) y[n] = x[n/L].
4. Cascade of two non-linear systems may result in a linear system. For example,
consider the cascade of two systems as shown below:

x[n] v[n] y[n]

System 1 System 2

For system 1: The relationship between the input and output is given by v
[n] = log {x[n]}.
For system 2: The relationship between the input and output is given byy
[n] = exp {v[n]}.
It can be observed that both System-1 and System-2 are non-linear system,
whereas the cascaded system is a linear system. This implies that cascading of
two non-linear systems need not be always non-linear.
5. A discrete-time system is causal, if its current output should not depend on the
future value of the input. A real-world system cannot react to future input;
hence, they are considered as causal systems.
6. A system is memory-less if the current output of the system depends on the
current input. All the memoryless systems are causal. A causal system is
non-anticipatory. For a causal system, the current output will not depend on
the future value of the input. For a causal system, the current output depends on
the past input. If the system output depends on the past input, it is a memoryless
system. Hence, all memoryless systems are causal, whereas all causal systems
are not memoryless. For example, y[n] = Ax[n] is a memory less system, which
is also causal. Consider the system y[n] = Ax[n] + Bx[n - 1]; the system is
causal, whereas it is not memory less.
7. Consider the cascade of two discrete-time time varying system, which is
depicted below.

x[n] v[n] y[n]

System 1 System 2

System 1: The system performs upsampling of the input signal by a factor of

2. The relationship between the input and output of the system is given by v[n] =
x[n/2]. Upsampling by a factor of 2 is a time-varying system.
System 2: System 2 performs downsampling of the input signal by a factor of
2. The relationship between the input and output of the system is given by y[n] =
v[2n]. Downsampling by a factor of 2 is a time-varying system.
Cascaded system: Cascading of upsampling by a factor of 2 followed by
downsampling of two is an identity system. Thus, the cascaded system is a time-
invariant system.
Appendix 505

Hence, it can be concluded that cascading of two time-varying discrete-time

system need not always result in time-varying system.
8. A discrete-time system is said to be invertible if ‘distinct input should in distinct
output’. Example of invertible system is y[n] = x[n/2], which is upsampling of
the input signal by a factor of 2. Example of non-invertible system is y[n] = x
[2n], which is downsampling of the input signal by a factor of 2.
9. It is possible to test the causality and stability of linear time-invariant system
from its impulse response. A linear time-invariant system is causal if its impulse
response is zero for n < 0. A linear time-invariant discrete time system is stable
if its impulse response is absolutely summable.
10. A discrete-time system is static if current output depends only on current input.
A discrete-time system is dynamic if the current output depends on current
input, past input and past output. Example of static system is y[n] =
x2[n]. Example of dynamic system is y[n] = x[n] + y[n -1].
11. A discrete-time system is said to be non-recursive if the current output of the
system depends on the current input and the past input. A discrete-time system is
said to be recursive if the current output depends on the previous output of the
system.
Example of non-recursive system: y[n] = 0.5x[n] + 0.5x[n - 1]. For this
system, the current output does not depend on the past output, hence it is
non-recursive.
Example of recursive system: y[n] = x[n] + 0.5y[n - 1]. It is a recursive
system because the current output is a function of previous output.
12. A zero at z = 1 is equivalent to zero at ω = 0. A zero at ω = 0, will block all low
frequency component. The system will block DC component of the signal. The
system will behave like a highpass filter.
13. A discrete-time system is invertible if distinct input leads to distinct output.
Example of a discrete-time system which is invertible is ‘accumulator system’.
The impulse response of accumulator is h[n] = u[n], whereas the impulse
response of the inverse of the accumulator system is given by h[n] = δ[n] -
δ[n - 1]. Example of discrete-time system which is non-invertible is
downsampler system whose input-output relationship is given by y[n] = x[2n].
14. State-space representation is an application for multiple input and multiple
output systems. State-space approach can be used to model non-linear and
time-varying systems.
15. From the impulse response of the discrete-time system, it is possible to infer
whether the system is causal and stable. A discrete-time system is causal if its
impulse response is zero for n < 0. A discrete-time system is stable if its impulse
response is absolutely summable.
506 Appendix

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Key D D A A D D C D A A B B C D C

Chapter 5: Transforms

Answers to PreLab Questions

1. Spectrum is a compact representation of the frequency content of a signal that is

composed of sinusoids.
2. The unilateral or one-sided Z-transform differs from the bilateral or double-sided
Z-transform in that the summation is carried out only over non-negative values
of time index (n), whereas the bilateral transform includes the both negative and
positive values of time index (n). The unilateral Z-transforms can be used to
analyse causal systems that are specified by linear constant coefficient difference
equations with non-zero initial conditions.
3. The range of variation of ‘z’ for which Z-transform converges is called region of
convergence of Z-transform.
4. The basis function has to be an orthogonal function. The basis function of
Fourier transform is complex exponential function.
5. Applying discrete Fourier transform is equivalent to applying discrete Fourier
series on a periodically extended finite aperiodic signal. Discrete Fourier series
is applied when the signal under analysis is periodic, and discrete Fourier
transform is applied when the signal under analysis is aperiodic.
6. DTFT is Z-transform evaluated on a unit circle. The expression for Z-transform
is given by

1
X ðzÞ = x½n]z - n
n= -1

Substituting z = rejω in the above expression, we get

1
-n
X ejω = x½n] rejω
n= -1

For a unit circle, r = 1; hence, the above expression can be written as

Appendix 507

1
X ejω = x½n]e - jωn
n= -1

Thus, DTFT is Z-transform evaluated on a unit circle.

7. Double-sided spectrum of a signal composed of sinusoid is expressed as

X k j2πf k t X *k - j2πf k t
N
xð t Þ = X 0 þ e þ e
k=1
2 2

The set of pairs ð0, X 0 Þ f 1 , 12 X 1 , - f 1 , 12 X *1 , . . . , f k , 12 X k , - f k , 12 X *k

indicates the size and relative phase of sinusoidal component contributing at
frequency fk. This is termed as frequency-domain representation of the signal.
8. Transform is a tool to analyse the signals and systems. Signals are converted
from time or spatial domain to frequency domain using transform. Frequency
domain is used to describe the signal with respect to frequency. Each frequency
has its own amplitude and phase. From the spectrum, it is possible to interpret
the frequencies are present in the signal. Thus, the time-domain and the
frequency-domain representation of the signal are equivalent. It is possible to
transform the signal from time domain to frequency domain and vice versa
without any loss of information. Mathematically, transform is taking the inner
product of the signal with the basis function. The inner product is one way of
quantifying the similarity or the dissimilarity of two signals.
9. The signal x[n] must be conjugate symmetric. This is expressed as x[n] = x*[-
n].
10. The basis function of Fourier transform is complex exponential, which oscillates
for all the time. Fourier transform describes the frequency components in the
signal averaged over all the time. It is difficult for the Fourier transform to
represent signals that are localized in time. Fourier transform is not an effective
tool to analyse non-stationary signals.
Time localization in Fourier transform can be achieved by windowing the
signal over which the signal is nearly stationary, which leads to the development
of short-time Fourier transform (STFT). It can be represented to non-stationary
signal and gives both time and frequency resolutions. However, the time and
frequency resolutions are fixed based on the windowing signal.
11. The DCT provides a decomposition of any discrete time signal as a weighted
sum of basis functions and these basis functions are cosines. If x[n] is a real for
all ‘n’, then the DCT output X[k] is real for all ‘k’. The DCT has excellent energy
compaction for many real-world signals (i.e. signals with high correlation
among neighbouring samples).
12. Wavelet transform has the ability to perform multi-resolution analysis of the
signal, whereas STFT cannot perform multi-resolution analysis. STFT provides
time-frequency representation of the signal, whereas wavelet transform provides
time-scale representation of the signal.
508 Appendix

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10 11
Key C C B B B C B C D C B

Chapter 6: Filter Design Using Pole-Zero Placement Method

Answers to PreLab Questions

1. If h[n] represents the impulse response of the lowpass filter, then the filter whose
impulse response is (-1)nh[n] will act as a highpass filter.
2. A discrete-time system with transfer functionH(z) is a minimum phase system if
the following conditions are met.
(a) All the zeros of the system are inside the unit circle centred about the origin.
(b) All the poles of the system are inside the unit circle centred about the origin.
(c) The numerator and the denominator of the transfer function (H(z)) have
equal orders of ‘z’.
3. The transfer function of a discrete-time system is given by H(z) = B(z)/A(z). The
frequencies for which the values of the denominator and numerator become zero
in a transfer function are called poles and zeros. Poles are the roots of the
denominator of a transfer function. Similarly, zeros are the roots of the numer-
ator of the transfer function. For a discrete-time system to be stable, the poles of
the system should lie within the unit circle.
4. For a discrete-time system to be stable, (a) the poles of the system should lie
within the unit circle, and (b) the impulse response should be absolutely
summable.
5. The basic principle underlying the pole-zero placement method is to locate poles
near points of the unit circle corresponding to frequencies to be emphasized and
locate zeros near the frequencies to be deemphasized. All the poles should be
placed within the unit circle for the filter to be stable. All complex zeros and
poles must occur in complex conjugate pairs for the filter coefficients to be real.
6. All-pass filters can be used as delay equalizer or phase equalizer. When an all-
pass filter is placed in cascade with a system that has an undesired phase
response, a phase equalizer is designed to compensate for the poor phase
characteristics of the system such that the cascaded system will exhibit linear
phaseresponse.
7. Notch filter is used to eliminate one particular frequency. It is used to minimize
power line interference in biomedical equipment. It can be used in radio
receivers to remove unwanted interfering frequencies.
Appendix 509

8. Moving average filter basically performs lowpass filtering of the input signal.
Lowpass filter converts drastic variation in the signal to a gradual variation.
When a square wave is fed as input to the M-point moving average filter, the
output will be a triangular wave. Square wave exhibits sharp transition between
‘ON’ and ‘OFF’ state. Triangular waveform exhibits gradual variation between
‘ON’ and ‘OFF’ state.
9. From the input-output relation, it is possible to observe that the current output is
a function of current input and previous input; hence, the given filter is finite
impulse response (FIR) filter.
10. Pole-zero plot is a two-dimensional plot with x-axis as the real part and y-axis as
the imaginary part. The pole-zero plot shows the unit circle with zeros marked as
‘0’, and poles are indicted with the symbol ‘x’. Zeros and poles near the unit
circle are expected to have a strong influence on the magnitude response of the
filter.

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10 11 12 13
Key D C C D D B C A D D C D D

Chapter 7: FIR Filter Design

Answers to PreLab Questions

1. The difference equation relating the input and output of an FIR filter is given by

M
y ½ n] = bk x½ n - k ]
k=0

From the difference equation, the following inferences can be drawn:

(a) The current output of FIR filter depends on current input and previous
inputs.
(b) The number of previous outputs necessary to compute the current output is
termed as the order of FIR filter.
(c) Since the current output is not a function of previous output, FIR filter is
considered as a ‘non-recursive filter’. It can also be termed as ‘all-zero’ filter
with the poles at the origin.
(d) The filter coefficients are denoted as ‘bk’. The nature of filtering depends on
‘bk’.
510 Appendix

2. Based on symmetry and number of coefficients, the FIR filters are classified as
Type I, Type II, Type III and Type IV FIR filters
(a) Type I FIR filter: Even symmetry with odd number of coefficients
(b) Type II FIR filter: Even symmetry with even number of coefficients
(c) Type III FIR filter: Odd symmetry with odd number of coefficients
(d) Type IV FIR filter: Odd symmetry with even number of coefficients
3. An FIR filter with impulse response ‘h[n]’ is said to exhibit even symmetry if h
[n] = h[N – 1 - n], where ‘N’ is the number of coefficients of FIR filter. An FIR
filter is said to exhibit odd symmetry if h[n] = -h[N – 1 - n].
4. A digital filter exhibits linear phase characteristics if its impulse response is
either symmetric or anti-symmetric.
5. If a digital filter exhibits linear phase characteristics, then it will not introduce
phase distortion. All the frequency components of the input signal will pass
through the filter with constant delay so that there will not be any phase
distortion.
6. The relationship between group delay (τg) and the phase response (ϕ(ejω)) of the
filter is given by τg = - dω d
ðϕðejω ÞÞ.
7. Order of FIR filter (M) is the number of previous input samples necessary to
compute the current output. If ‘M’ denotes the order of the FIR filter and ‘N’
denotes the number of coefficients of FIR filter, then the relationship between
‘M’ and ‘N’ is given by N = M + 1.
8. FIR filter is an ‘all-zero’ filter with the poles occurring at the origin. For a digital
system to be stable, the pole should lie within the unit circle. Since the pole of
FIR filter occurs at the origin, the FIR filter is an inherently stable filter.
9. The advantages of FIR filter are:
(a) FIR filter exhibits linear phase characteristics; hence, there will not be phase
distortion
(b) The group delay of FIR filter is constant; hence, all the frequency compo-
nent of the input signal passes through FIR filter with equal delay.
(c) The poles of FIR filter lie at the origin; hence, FIR filter is inherently stable
filter
(d) The coefficients of FIR filter are either symmetric or anti-symmetric in
nature. Symmetricity of filter coefficient leads to linear phaseresponse of
the filter.
10. FIR filters can be designed using (a) windowing method, (b) frequency sampling
method and (c) optimal method.
Appendix 511

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Key A C B A A B A C B D D B B B D

Chapter 8: Infinite Impulse Response Filter

Answers to PreLab Questions

1. In a recursive filter, the present output depends on both the inputs and previously
calculated outputs.
2. The IIR filter is a recursive filter. The present output of the IIR filter depends on
input and past output. Hence, the IIR filter is a recursive filter.
3. Ripple is the fluctuations in the passband or stopband of the filter’s frequency
response. It is expressed in decibels.
4. Based on the ripples in the frequency response of the IIR filter, the filters are
classified as (a) Butterworth filter, (b) Chebyshev filter (Type I), (c) inverse
Chebyshev filter (Type II) and (d) elliptic filter.
5. The Butterworth filter’s magnitude response is monotonically decreased at all
frequencies, and also, there are no local maxima or minima in both the passband
and stopband. Hence, it is also called as flat-flat filters.
6. Mapping is a technique used in IIR filter design for converting analogue filter
into digital filter. The different types of mapping techniques are (a) backward
difference method, (b) impulse invariant technique (IIT), (c) bilinear transfor-
mation technique (BLT) and (d) matched Z-transformtechnique.
7. While converting an analogue filter into an equivalent digital filter, it is neces-
sary that stable analogue filter should be mapped to a stable digital filter. The
points of the analogue filter in the left half S-plane must be mapped into inside
the unit circle in the Z-plane to preserve the stability of the filter. Therefore, all
the mapping techniques must preserve the stability of the filter.
8. The steps involved in obtaining the transfer function of a digital filter using the
impulse invariant technique are summarized below:
(a) Obtain the transfer function of an analogue filter (H(s)), which has to be
converted into an equivalent digital filter (H(z)).
(b) From the analogue transfer functionH(s), get the impulse response h(t) using
the inverse Laplace transform.
(c) Apply the sampling process on impulse response h(t) to get h[nT].
(d) Take Z-transform of the sampled impulse response h[nT] to get the equiv-
alent transfer functionH(z).
512 Appendix

Table A.2 Comparison of IIR filters

Type of IIR Ripple in Ripple in Transition Order of the filter to meet the
S. no. filter passband stopband width given specification
1 Butterworth NO NO Widest Highest
filter
2 Chebyshev YES NO Narrower Lower
filter
3 Inverse NO YES Narrower Lower
Chebyshev
filter
4 Elliptic filter YES YES Narrowest Lowest

9. The drawbacks of impulse invariant technique are listed below:

(a) It is more suitable for all pole filters and does not consider the system’s
zeros.
(b) The mapping of analogue frequency ‘Ω’ to digital frequency ‘ω’ is ‘many-
to-one’; hence, aliasing problem exists.
(c) Due to the presence of aliasing, the impulse invariant method is appropriate
for the design of lowpass and bandpass filters only. It is not a suitable
technique for the design of highpass and band reject filters.

10. The bilinear transformation is a conformal mapping that maps the ‘jΩ’ axis of
the S-plane into the unit circle of the Z-plane only once. Therefore, it can avoid
aliasing problems.
11. The bilinear transformation technique maps the analogue frequency and digital
frequency in a non-linear fashion. The relationship between the analogue and
digital frequency is given by ω = 2 tan - 1 ΩT 2 . The non-linear relationship
between analogue and digital frequency is termed as ‘frequency warping’. To
overcome the frequency warping problem, ‘prewarping’ technique is used.
Prewarping will preserve the edge frequencies but not the exact shape of the
magnitude response.
12. Comparison of different types of IIR filter is given in Table A.2
13. The following steps are involved in the IIR filter design:
Step 1: Convert the digital filter specifications into an equivalent analogue filter
specification.
Step 2: Convert the analogue filter specifications to normalized lowpass proto-
type specifications.
Step 3: Design a normalized lowpass prototype filter by using any one of the
analogue filters: (a) Butterworth filter, (b) Chebyshev Type I filter,
(c) Chebyshev Type II filter or (d) elliptic filter.
Step 4: Use the analogue transformation technique to convert the normalized
lowpass prototype filter into the desired analogue filter.
Appendix 513

Step 5: Use the mapping technique to convert the desired analogue filter into a
desired digital one.

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10
Key B D C A A A B D C B

Chapter 9: Quantization Effect of Digital Filter Coefficients

Answers to PreLab Questions

1. The finite word length introduces an error that can affect the performance of the
DSP system. The finite word length has limited precision, and it is not sufficient
to represent the filter coefficients accurately. This causes errors between the
original filter coefficients and finite word length coefficients. The finite number
of bits is used in the arithmetic operations in DSP, which is insufficient to give
the proper result.
2. In fixed point arithmetic representation, the numbers are represented in a fixed
range with a finite number of bits of precision. The numbers beyond the fixed
range are either saturated or wrapped around. It is preferred for high speed and
low cost.
In floating-point arithmetic representation, every number is represented in two
parts (a) mantissa and (b) exponent. Floating-point representation has a higher
dynamic range and no need for scaling. It can be used to perform more complex
algorithms in it.
3. The numbers can be represented in binary format, which contains ‘0s’ and ‘1s’.
In sign-magnitude representation, most significant bit (MSB) is used to denote
the number as positive or negative. It is called as sign bit, and the remaining bits
are used to represent the number, which is called as magnitude bits.
4. Quantization is a process in which a quantity X is approximated into a quantity
Q(X). The approximated value will have a distortion between X and Q(X), called
quantization error.
5. Rounding is a method to perform quantization operation. It selects the quantized
value nearest to the original value. The error between the quantized and original
value will not exceed ±(Δ/2). Here ‘Δ’ denotes step size, which is obtained by Δ
= 2-B, and B is the number of binary bits.
514 Appendix

6. The two’s complement truncation is another method to perform the quantization

operation. This method always gives a resultant quantized value less than or
equal to the original value. The truncation error will to (-Δ to 0).
7. The magnitude truncation is another approach to perform the quantization
operation. The result of the quantized value is always less than the original
value for X > 0, and the quantized result is always greater than the original value
for X < 0. The advantage of magnitude truncation is that it can inherently
suppress the limit cycle oscillation.
8. If the dynamic range of the signals crosses the word length limit, then the
overflow exists. The different types of overflows are (a) saturation, (b) zeroing
and (c) two’s complement.
Saturation: If the input value crosses the maximum/minimum limit (X/-X),
the output will be X/-X.
Zeroing: The output will be zero if the input value exceeds the maximum
limit.
Two’s complement: It is a periodic continuation of the 45° straight line. The
advantage of this method has the capability of correcting the intermediate
overflows automatically.
9. A limit cycle oscillation is a low-level oscillation that can exist in a stable filter
due to the non-linearity associated with the quantization operation, like rounding
or truncation of the arithmetic calculations in the filtering operation. This limit
cycle oscillation is also termed a multiplier round-off limit cycle. These limit
cycle oscillations do not occur in non-recursive FIR filters.
10. Infinite precision is needed to represent the filter coefficients. However, the finite
number of bits represents the filter coefficients in the real world. Therefore, the
representation of the filter coefficients from infinite precision to finite number
precision may introduce coefficient quantization. Due to the finite precision
representation of filter coefficients, the response of the quantized filter may
deviate from the response of the original filter.

Answer to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10
Key C A A A A A C A D B
Appendix 515

Chapter 10: Multirate Signal Processing

Answers to PreLab Questions

1. When two devices operating at different sampling rate are to be interconnected, it

is necessary to change the sampling rate of the signal. Often there is a mismatch
between the sampling rates of the recording and playback system. The sampling
rate of audio signals in compact disc is 44.1 kHz, whereas the sampling rate of
audio signals in digital audio tape is 48 kHz. Sampling rate conversion is required
for the interconnection of compact disc with digital audio tape.
2. A sampling rate converter is a device or software that accepts digital input
signals at one sampling rate and outputs a digital signal at a different sampling
rate. This is shown below.

x[n] y[n]
Sampling rate converter
fs f’s

Block diagram of sample rate converter

3. The two basic operations in multirate signal processing are (a) downsampling
and (b) upsampling. Downsampling reduces the sampling rate of the input
signal, whereas upsampling increases the sampling rate of the input signal.
4. Downsampling in time domain may lead to spectral overlap in the frequency
domain, which is termed as ‘aliasing’. To overcome aliasing, a lowpass filter is
used before downsampling. This filter is termed as ‘anti-aliasing filter’. This is
depicted below.

x[n] Anti-aliasing y[n]

filter M

Decimation operation

To overcome the problem of aliasing, the cut-off frequency of the anti-aliasing

filter is chosen as fs/M.
5. Upsampling by a factor of ‘L’ is the process of inserting ‘L - 1’ zeros between
successive samples. Upsampling in time domain results in the creation of
multiple copies of the original spectrum in the frequency domain. The anti-
aliasing filter is a lowpass filter that retains only the original spectrum and
removes multiple copies of the original spectrum. The block diagram of anti-
imaging filter with upsampling operation is shown below. The sampling fre-
quency of the anti-imaging filter is chosen as fs/L.

x[n] Anti-imaging y[n]

L filter

Interpolation operation
516 Appendix

6. The time-domain expression for downsampling by a factor of ‘M’ is given by y

[n] = x[nM]. The frequency-domain expression for downsampling by a factor of
M-1
X zM e - j M k .
1 2π
‘M’ is given by Y ðzÞ = 1
M
k=0
7. The time-domain expression for upsampling by a factor of ‘L’ is given by y[n] =
x[n/L]. The frequency-domain expression is given by Y(z) = X(zL).
8. The three significant properties of downsampling operation are
(a) Downsampling is a linear operation.
(b) Downsampling is a time-variant operation.
(c) Downsampling is an irreversible operation.
9. (a) Upsampling is a linear operation, because it obeys additivity and homoge-
neity properties. Thus, upsampling obeys the superposition principle; hence, it is
a linear operation.
(b) Upsampling has varying responses to the same input at different instants
of time; hence, it is considered as time-variant operation.
10. Downsampling by a factor ‘M’ and upsampling by a factor of ‘L’ are inter-
changeable if ‘L’ and ‘M’ are relatively prime.
11. Idempotent operations are operations, which can be applied multiple times
without changing the result. Downsampling by a factor of ‘M’ followed by
upsampling by the same factor gives the same result if the operation is repeated
many times; hence, it is considered as idempotent operation.
12. Polyphase decomposition ensures that filtering operations are performed at the
lowest possible sampling rate in the system, which reduces the computational
complexity and the overall system’s cost. Polyphase decomposition can be
classified as (a) Type I polyphase decomposition and (b) Type II polyphase
decomposition.
13. Filter bank is a group of filters arranged in a specific fashion. Filter bank is used
for subband decomposition of the signal. It is useful for signal denoising and
signal compression.
14. The main threats for perfect reconstruction in a two-channel filter bank are
(a) aliasing problem, (b) amplitude distortion and (c) phase distortion. Perfect
reconstruction can be achieved by proper choice of analysis and synthesis filters.
15. Transmultiplexer is a multiple input-multiple output system (MIMO). It uses
multirate operators and filters to combine ‘M’ signals for transmission across a
channel and then recovers the ‘M’ input signals at the receiver end. The
separation of signals should be perfect, and the recovery of each signal should
be performed without leakage of signal from one channel to another, which is
generally termed as crosstalk. The proper choice of filters can avoid the crosstalk
problem.
Appendix 517

Answers to Objective Questions

Q. No. 1 2 3 4 5 6 7 8 9 10 11
Key B B B B B A A B B B D

Chapter 11: Adaptive Signal Processing

Answers to PreLab Questions

1. In optimal filtering, the input and the desired signals are available for a given
time window, and the optimal parameters of the filter are computed only once. In
adaptive filtering, the input and the desired signal are provided to the algorithm,
and the algorithm computes the parameters of the filter and is updated; hence, it
is iterative in nature. Adaptive filter does not require previous knowledge of the
signal statistics.
2. An adaptive filter is a filter with filter coefficients that are non-constants. The
filter coefficients are adjusted based on some specific criterion defined to
optimize the filter’s performance. In an ordinary filter, the filter coefficients
are constant and do not vary with respect to specific criteria.
3. Some examples of the adaptive filter include Wiener filter, least mean square
filter, RLS filter, etc.
4. The adaptive filters are generally preferred in the following contexts:
(a) The filter characteristics are necessary to be changed or adapted to specific
conditions.
(b) Spectral overlap between the signal and noise
(c) If the noise present in the signal is unknown or varies with time.
5. The performance measures of the adaptive filter are rate of convergence,
misadjustment, tracking, robustness, computational complexity, filter structure,
numerical stability and accuracy.
6. The least mean square (LMS) algorithm is an adaptive filter method that uses a
gradient-based method of steepest descent to obtain the least mean square error
between the output and the reference data. It is an iterative procedure that
corrects the weight vector (filter coefficient) in the direction of the negative of
the gradient vector, which eventually leads to the minimum mean square error.
7. The cost function of least square estimation is defined as the sum of weighted
error squares. The least square estimation is to minimize the error of the filter
output to the reference signal. In this process, statistical modelling is not
involved directly.
8. The variants of LMS algorithm include (a) normalized LMS, (b) leaky LMS,
(c) block LMS and (d) sign LMS.
518 Appendix

9. The step size parameter plays a vital role in the LMS algorithm. The larger the
step size value, the faster the adaption, increasing residual MSE. Also, it affects
the stability of the algorithm. Therefore, the step size selection cannot be
arbitrarily large.
10. The recursive least squares is an adaptive filter algorithm, which recursively
obtains the filter coefficients that minimize a least squares cost function relating
to the input. In this algorithm, input signals are considered deterministic,
whereas input signals are considered stochastic in the LMS. As a result, the
RLS algorithm converges faster than the LMS algorithm.

Answers to Objective Questions

Q. No. 1 2 3 4 5 6
Key A D A B A C
Index

A 296, 297, 300, 336, 344, 352, 366, 367,

Adaptive filter, 444, 451–465, 468, 517, 518 408, 435, 445–447, 449, 451, 492, 497
Aliasing, 36, 39, 40, 42–45, 64, 310, 315, 405, Butterworth filter, 304, 326, 328–344, 362, 367,
421, 500, 512, 515, 516 369, 370, 373, 374, 389, 400, 482, 511,
All-pass filter, 231, 232, 251–255, 259, 508 512
Analog frequency transformation, 326–328
Analog signal, 36, 39, 65, 500
Anti-aliasing filter, 405, 515 C
Anti-imaging filter, 515 Causal system, 124, 153, 154, 163, 504, 506
Associative property, 102, 104 Chebyshev filter, 304, 344–358, 367, 369, 373,
Autocorrelation, 71, 111–118, 120, 121, 444, 374, 511, 512
446, 448, 467, 468, 503 Coefficient quantization, 378, 383–392, 394,
Autoregressive moving average system 397, 514
(ARMA), 231 Comb filter, 231, 232, 255–261
Autoregressive system (AR), 231 Commutative property, 98, 101, 103
Continuous-time Fourier transform (CTFT),
175–180, 182, 185, 186
B Continuous-time signal, 1–31, 35, 36, 39, 59,
Band pass filter, 259, 301, 326 64, 72, 169, 175, 497–518
Basis function, 8, 168, 169, 175, 176, 196, 204, Continuous wavelet transform (CWT),
229, 499, 506, 507 219– 222, 224, 228
Bilinear transformation technique (BLT), 304, Control library
306, 320–323, 331, 333, 335, 337, 342, control.tf(), 325
346, 347, 352, 355, 359, 362, 364, 367, Convolution, 25, 26, 71, 72, 97–110, 112, 113,
373, 374, 389, 392, 511, 512 119, 134, 146, 169, 194, 228, 446, 501,
Block diagram, 71, 106, 108, 123, 125, 145, 502
146, 193, 205, 211, 255, 258, 295, 414, Correlation, 72, 111–122, 502, 503, 507
444, 451, 453, 455, 460, 462, 463, 470–
472, 481, 503, 515
Built-in function, 2–5, 9, 11, 12, 24, 28, 29, 61, D
73, 74, 96, 129, 132, 133, 140, 141, 198, Difference equation, 123, 125–129, 134, 135,
211, 212, 220, 225, 234, 249, 252, 265, 161, 169, 264, 305, 306, 503, 506, 509
266, 271, 278, 282–288, 291, 293, 294, Digital resonator, 231, 232, 246–247, 260

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 519
S. Esakkirajan et al., Digital Signal Processing,
https://doi.org/10.1007/978-981-99-6752-0
520 Index

Digital signal, 21, 36, 82, 97, 159, 378, 383, Frequency resolution, 211, 213–217, 228, 492,
407, 500, 515 507
Discrete Cosine transform (DCT), 168, Frequency response, 123, 142, 144, 167, 194,
206– 210, 472, 478–480, 507 233, 234, 248, 251, 256, 260, 266, 272,
Discrete Fourier transform (DFT), 168, 278, 293, 301, 304, 309, 328, 337, 344,
196– 204, 206, 209, 210, 227, 229, 354, 358, 362, 367, 373, 374, 383, 384,
474– 476, 478–479, 506 389, 397, 448, 451, 476, 511
Discrete-Time Fourier Transform (DTFT), 167, Frequency sampling method, 271, 291, 292,
168, 175, 191–194, 272, 506, 507 510
Discrete-time system
system properties, 123, 149–161
system representation, 123, 125–142 G
system response, 123, 125, 142–144 Gaussian function, 2, 26–33, 179, 186, 187,
Discrete wavelet transform, 168, 221–229 190, 191, 229, 389, 499
Distributive property, 104, 106, 107
Double-sided spectrum, 168, 179, 507
Downsampling, 91–93, 119, 120, 159, H
403– 408, 413–415, 417, 421, 422, 432, Hamming window, 284, 289, 299
436–440, 502, 504, 505, 515, 516 Harmonics, 175, 258, 259, 488, 489, 491, 492
Dual function, 186, 499 Highpass filter, 141, 142, 421, 438
Homogeneity property, 149, 150, 503, 516

E
Elliptic filter, 304, 362–375, 511, 512 I
Energy, 72, 86, 87, 95, 96, 117–120, 122, 200, Ideal interpolation, 59, 64, 500
202, 204, 206, 210, 221, 223, 226, 275, Impulse invariant technique (IIT), 304, 306,
482, 502, 507 310–320, 323, 373, 511, 512
Even signal, 72, 117, 120, 122, 502 Impulse response, 23, 61, 99, 128, 193, 233,
264, 305, 384, 417, 449
Infinite impulse response filter (IIR), 126,
F 231– 238, 253, 304–373, 377, 389–394,
Filter bank, 294, 403, 404, 419–433, 472, 516 401, 511–513
Finite impulse response filter Inverse transform, 196
type-I FIR filter, 264–266, 271, 273, 510 Inverse Z-transform, 145, 146, 167, 173–174,
type-II FIR filter, 266–268, 271, 273, 291, 227, 233, 305, 425, 427
510
type-III FIR filter, 268, 269, 271, 273, 291,
510 L
type-IV FIR filter, 269–271, 273, 291, 510 lcapy
Finite word length effect, 377, 378 lcapy.exp(), 170–172
FIR filter design lcapy.sin(), 173
frequency sampling based FIR filter design, Least Mean Square (LMS) algorithm, 443, 444,
291 452, 453, 455, 456, 458–462, 465, 467,
optimal FIR filter design, 293–294 468, 517, 518
window based FIR filter design, 272, 467 Linear phase, 142, 162, 194, 241, 249, 264,
First-order hold (FoH), 59, 62–64, 69 265, 268, 269, 271, 276, 278, 280, 282,
First order system, 232–239 284, 286, 288, 291, 294, 298–301, 305,
Fourier transform, 8, 23, 26, 38, 40, 67, 168, 356, 360, 366, 424, 508, 510
174, 175, 177, 181, 183, 186, 190, 204, Linear system, 3, 149, 503, 504
205, 210, 216, 218, 227, 229, 289, 448, Linear time-invariant (LTI) system, 8, 97,
449, 457, 499, 503, 506, 507 125–129, 136, 146, 231, 260, 498, 499,
Frequency, 3, 36, 82, 139, 168, 233, 271, 315, 502, 505
383, 405, 446 Low pass filter, 141, 142, 405
Index 521

M Non-stationary signal, 1–3, 17–19, 167, 168,

Magnitude response, 142, 144, 147, 148, 163, 205, 211–214, 216, 217, 219, 489, 492,
193, 194, 227, 233, 237, 241, 246, 247, 499, 507
249, 251–254, 265, 266, 268, 269, 271, Normalized LMS algorithm, 457–458
275, 276, 278, 280, 282–284, 286, 288, Notch filter, 231, 232, 248–253, 255, 258–260,
291, 293–295, 298–301, 328–330, 337, 294, 508
340, 341, 344, 356, 359, 360, 362, 363, Numpy
366, 373, 374, 392, 423–425, 450, 458, np.abs(), 12
462, 464, 476, 509, 511, 512 np.append(), 203
Mapping technique np.arange(), 74
backward difference method, 306, 307, 309, np.arctan(), 321
310, 511 np.ceil(), 51
bilinear transformation technique (BLT), np.concatenate(), 214
306, 320–323, 511 np.cos(), 51
impulse invariant technique (IIT), 306, np.cumsum(), 51
310–320, 511 np.dot(), 51
matched Z-transform technique, 306, np.exp(), 9, 12, 29
323–325, 511 np.floor(), 51
Matplotlib np.imag(), 51
plt.axis(), 197 np.kaiser(), 51
plt.bar(), 368 np.linspace(), 3–5, 9, 12, 497
plt.grid(), 247, 250, 252, 254, 257, 265, 267, np.log(), 51
268, 270, 272, 311, 345 np.max(), 51
plt.imshow(), 54 np.ones(), 51
plt.pcolormesh(), 212 np.outer(), 51
plt.plot(), 4 np.random(), 51
plt.show(), 222, 311, 338, 342, 353, 355, np.real(), 51
360, 364, 368, 473–475, 477 np.roots(), 51
plt.specgram(), 212, 492 np.sign(), 51
plt.stem(), 74 np.sin(), 4, 29, 446, 447, 451
plt.subplot(), 4, 9, 12 np.sinc(), 51
plt.tight_layout(), 6, 39, 76, 127, 178, 235, np.sum(), 51
265, 308, 384, 406, 445, 478 np.tan(), 51
plt.title(), 4 np.tile(), 51
plt.xlabel(), 4 np.unwrap(), 51
plt.xticks(), 74 np.zeros(), 51
plt.ylabel(), 4
plt.yticks(), 80
Matrix, 196–200, 209, 229, 444, 446, 467, 468 O
Mid-rise quantizer, 45, 49–55, 65 Odd signals, 72, 117, 120, 122, 502
Mid-tread quantizer, 46–49, 51, 55, 67 Optimal filter, 293, 444, 468
Minimum mean square error, 467, 517
Minimum phase system, 141, 232, 254, 508
Moving average system (MA system), 231, P
239, 241 Parseval’s theorem, 202, 204
Multirate operator, 403, 405–413, 417, 516 Perfect reconstruction, 204, 294, 404, 421, 425,
Multi-resolution analysis, 493, 507 430, 434, 516
Periodic signal, 3–8, 174, 200, 498, 502
Phase interchange, 227
N Phase response, 14, 123, 125, 141–144,
Noble identity, 413–416 146–148, 162, 163, 193–195, 227, 234,
Noise, 110, 113, 114, 226, 258, 297, 298, 444, 241, 247, 249, 251, 254, 264–266, 268,
446, 447, 450, 455, 460, 462, 470, 482, 269, 271, 276, 278, 280, 282–284, 288,
486 293, 299, 300, 327, 328, 337, 339, 343,
Non-linear system, 124, 149, 392, 504 344, 354, 356, 357, 360, 361, 365, 366,
522 Index

384–387, 389, 391, 392, 423, 424, 426, Recursive system, 394, 401, 505
427, 450, 508, 510 RLS algorithm, 443, 444, 462–468, 518
Poles, 139, 140, 144, 154, 157, 158, 165, 228, Rounding, 377–380, 383, 389, 390, 392, 394,
232, 234, 237, 238, 246, 247, 251, 254, 400, 401, 513, 514
256–261, 266, 271, 294, 305, 310, 318,
319, 323–325, 348, 349, 375, 383, 451,
508–510, 512 S
Pole-zero plot, 123–125, 139–142, 144, 145, Sampling, 35–64, 291, 299, 307, 321, 472–474,
157, 159, 160, 162, 232, 234, 241, 247, 500, 511, 515, 516
249, 251, 254, 256, 257, 265, 266, 268, Sampling frequency, 28, 36, 39–45, 69, 250,
269, 271–273, 278, 280, 282, 283, 286, 283, 284, 293, 296, 300, 315, 331, 332,
299, 300, 312, 319, 320, 384–387, 389, 337, 342, 346, 354, 359, 362, 364, 369,
391, 392, 451, 503, 509 370, 372–374, 389, 392, 438, 460, 500,
Polyphase decomposition, 403, 404, 417–419, 515
516 Sampling rate conversion, 404, 515
Pulse signal, 1, 25–26, 101, 102, 179, 185, 192, Sampling theorem, 36, 39–45, 500
193 Sawtooth signal, 48, 51, 56, 209, 225, 227, 430
Pulse train, 36–39 Scaling operation, 91–93, 502
Python, 2–12, 14–18, 21–31, 37–52, 56, 58, Scipy
60–65, 67, 73–107, 109–118, 120, 121, signal.bilinear(), 353
123, 125, 127–129, 131–133, 135–147, signal.butter(), 331
149–153, 155–163, 170–174, 177–179, signal.buttord(), 336
181–184, 186–188, 190, 192–205, 207, signal.cheby1(), 366
208, 210–218, 220, 221, 223–228, signal.cheby1ord(), 353
234–242, 244–253, 256, 265–272, 274, signal.cheby2(), 358
275, 277–279, 281, 283–287, 289, signal.convolve(), 325
291–299, 301, 302, 306–319, 321–323, signal.ellip(), 363
325, 327–332, 336–340, 342–345, 352, signal.ellipord(), 364
353, 355–367, 369–374, 379–389, signal.filtfilt(), 298
392–394, 396–398, 400, 406–410, signal.firwin(), 300
412–428, 432–438, 440, 443–447, 449, signal.firwin2(), 300
450, 452, 453, 455, 456, 458–462, signal.freqz(), 451
464–467, 472–475, 477–480, 482–487, signal.lfilter(), 446
489–493, 495 signal.remez(), 300
Python code for time-frequency representation, signal.residue(), 318
492 signal.sawtooth(), 23
Pywavelet signal.sos2tf(), 397, 398
pywt.cwt(), 222 signal.square(), 21
pywt.dwt(), 224 signal.ss2tf(), 132
pywt.idwt(), 224 signal.stft(), 212
pywt.threshold(), 226 signal.tf2ss(), 133
pywt.wavedec(), 226 signal.tf2zpk(), 451
pywt.waverec(), 226 Scipy.fftpack
fft(), 206
fftfreq(), 206
Q ifft(), 210
Quantization, 35–64, 68–70, 377–397, 400, Shifting operation, 79, 93–95, 502
401, 500–501, 513–514 Short-time Fourier transform (STFT), 168, 205,
210–219, 228, 492, 493, 507
Sinc signal, 186
R Single sided spectrum, 179
Rectangular pulse, 25, 26, 101, 182, 185, Sinusoidal signal, 2–8, 14–18, 28, 29, 31, 32,
191–193, 384 41, 58, 61, 64, 65, 67, 68, 86, 89, 147,
Index 523

149, 179, 205, 208, 215, 220, 228, 258, U

260, 369–371, 421, 446, 447, 451, 492, Uniform quantization, 52, 55
493, 495, 498, 502 Unit ramp signals, 73, 76–78, 118, 122, 502
Sound device library as sd Unit sample signal, 73, 74, 77, 79, 85, 99–101,
sd.play(), 29 105, 106, 118, 122, 134, 139, 228, 234,
sd.wait(), 29 246, 501
Square wave, 3, 21, 32, 36, 106, 108–111, 119, Unit step signal, 3, 73, 75–77, 79, 84, 90, 95,
223, 224, 227, 232, 241, 242, 437, 500, 118, 122, 126, 135, 136, 138, 155, 502
509 Unstable system, 149, 157, 158
Stable system, 135, 136, 156, 164, 165, 241 Upsampling, 91, 93, 94, 119, 120, 159,
Step response, 123, 134–140, 142, 146–148, 403–405, 409–417, 437, 439, 440, 502,
162, 163, 362 504, 505, 515, 516
Subband, 403, 419, 420, 516
Sympy
sympy.KroneckerDelta(), 170 V
sympy.summation(), 172 Voice, 117, 435, 438, 469–472, 479
sympy.symbols(), 172 Voice recognition, 470
System identification, 294, 443, 453, 464–466

W
T Warping, 304, 320, 322, 512
Time-frequency representation, 168, 211, 212, Wavelet, 219, 220, 225, 479, 480, 493, 495
216, 218, 492–494, 507 Wavelet transform, 168, 219, 220, 493, 507
Time-invariant system, 164, 504 White noise, 110, 225, 226, 297, 379, 380, 446,
Time resolution, 211, 213–215, 228 447, 449, 451
Time reversal, 95–98, 122, 502 Wiener filter, 443–451, 467, 468, 517
Time-variant system, 124, 151–152, 162 Wiener-Hopf equation, 444, 447, 467
Transfer function, 61, 62, 68, 69, 125, 130–133, Window, 26, 211, 271, 383, 467
139–141, 145, 155, 157, 161–165, 227,
229, 232–234, 237, 238, 246, 248, 249,
251, 255, 256, 260, 302, 304, 305, 307, Z
309, 310, 312, 315–318, 324–328, 331, Zero input limit cycle oscillation, 392–397, 514
334, 337, 338, 341, 344, 351–354, 359, Zero-order hold (ZoH), 59–64, 66–68
362, 363, 373, 451, 503, 508, 511 Zero padding, 227
Transform, 167–226, 405 Zeros, 5, 37, 73, 126, 187, 232, 264, 319, 383,
Transmultiplexer, 403, 431–437, 516 405, 447
Truncation, 377–383, 385–388, 392, 395, 400, Z-plane, 139–141, 228, 237, 307, 308,
401, 514 310–312, 320, 322, 323, 374, 375, 511,
Two-channel filter bank, 404, 420–425, 438, 512
516 Z-transform, 130, 155, 167–172, 228, 255, 304,
306, 310, 323, 324, 506, 507, 511