Computational Imaging and Vision 43

Isaac Amidror

Mastering
the Discrete Fourier
Transform in One,
Two or Several
Dimensions
Pitfalls and Artifacts
Mastering the Discrete Fourier Transform in One,
Two or Several Dimensions: Pitfalls and Artifacts
Computational Imaging and Vision

Managing Editor

MAX VIERGEVER
Utrecht University, Utrecht, The Netherlands

Series Editors

GUNILLA BORGEFORS, Centre for Image Analysis, SLU, Uppsala, Sweden

DANIEL CREMERS, Technische Universität München, München, Germany
RACHID DERICHE, INRIA, Sophia Antipolis, France
KATSUSHI IKEUCHI, Tokyo University, Tokyo, Japan
REINHARD KLETTE, University of Auckland, Auckland, New Zealand
ALES LEONARDIS, ViCoS, University of Ljubljana, Ljubljana, Slovenia
STAN Z. LI, CASIA, Beijing & CIOTC, Wuxi, China
DIMITRIS N. METAXAS, Rutgers University, New Brunswick, NJ, USA
HEINZ-OTTO PEITGEN, CeVis, Bremen, Germany
JOHN K. TSOTSOS, York University, Toronto, Canada

This comprehensive book series embraces state-of-the-art expository works and advanced
research monographs on any aspect of this interdisciplinary field.

Topics covered by the series fall in the following four main categories:

• Imaging Systems and Image Processing

• Computer Vision and Image Understanding
• Visualization
• Applications of Imaging Technologies

Only monographs or multi-authored books that have a distinct subject area, that is where each
chapter has been invited in order to fulfill this purpose, will be considered for the series.

Volume 43

For further volumes:

http://www.springer.com/series/5754
Isaac Amidror

Mastering the Discrete Fourier

Transform in One, Two or
Several Dimensions
Pitfalls and Artifacts
Isaac Amidror
School of Computer and Communication Sciences
Peripheral Systems Laboratory
Ecole Polytechnique Fédérale de Lausanne
Lausanne, Switzerland

ISSN 1381-6446
ISBN 978-1-4471-5166-1 ISBN 978-1-4471-5167-8 (eBook)
DOI 10.1007/978-1-4471-5167-8
Springer London Heidelberg New York Dordrecht

Library of Congress Control Number: 2013941876

Mathematics Subject Classification (2010): 65T50, 42A99, 42B99

© Springer-Verlag London 2013

This work is subject to copyright. All rights are reserved 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. Exempted from this legal reservation are brief excerpts in connection with
reviews or scholarly analysis or material supplied specifically for the purpose of being entered and executed
on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication or parts
thereof is permitted only under the provisions of the Copyright Law of the Publisher’s location, in its current
version, and permission for use must always be obtained from Springer. Permissions for use may be obtained
through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution under the
respective Copyright Law.
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.
While the advice and information in this book are believed to be true and accurate at the date of publication,
neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or
omissions that may be made. The publisher makes no warranty, express or implied, with respect to the
material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

To my parents
 An expert is a man who has made all the mistakes,
which can be made, in a very narrow field.

Niels Bohr [Mackay91 p. 35]

Front cover image: Aliasing due to insufficient sampling rate

may give unexpected shapes both in the signal domain and
in the spectral domain. See Fig. 5.17(k).
 Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 The discrete Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 A brief historical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 The scope of the present book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Overview of the following chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 About the graphic presentation of sampled signals and discrete data . . . . . . 7
1.5.1 Graphic presentations in the 1D case . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5.2 Graphic presentations in the 2D case . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5.3 Graphic presentations in the MD case . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 About the exercises and the internet site . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2. Background and basic notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 The continuous Fourier transform: definitions and notations . . . . . . . . . . . . 15
2.3 The discrete Fourier transform: definitions and notations . . . . . . . . . . . . . . . 17
2.4 Rules for deriving new Fourier transforms from already known ones . . . . . 21
2.4.1 Rules for the 1D continuous Fourier transform . . . . . . . . . . . . . . . . 22
2.4.2 Rules for the 2D continuous Fourier transform . . . . . . . . . . . . . . . . 23
2.4.3 Rules for the MD continuous Fourier transform . . . . . . . . . . . . . . . 25
2.4.4 Rules for the 1D discrete Fourier transform . . . . . . . . . . . . . . . . . . 26
2.4.5 Rules for the 2D discrete Fourier transform . . . . . . . . . . . . . . . . . . 28
2.4.6 Rules for the MD discrete Fourier transform . . . . . . . . . . . . . . . . . 29
2.5 Graphical development of the DFT — a three-stage process . . . . . . . . . . . . 31
2.6 DFT as an approximation to the continuous Fourier transform . . . . . . . . . . 36
2.7 The use of DFT in the case of periodic or almost-periodic functions . . . . . . 41
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3. Data reorganizations for the DFT and the IDFT . . . . . . . . . . . . . . . . . . . . . . . 45

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Reorganization of the output data of the DFT . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Reorganization of the input data of the DFT . . . . . . . . . . . . . . . . . . . . . . . . . 47

vii
viii Contents

3.4 Data reorganizations in the case of IDFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4. True units along the axes when plotting the DFT . . . . . . . . . . . . . . . . . . . . . . 69

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 True units for the input array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 True units for the output array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.3.1 The particular case of periodic functions . . . . . . . . . . . . . . . . . . . . . 74
4.4 True units for the DFT element values (heights along the vertical axis) . . . . 78
4.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5. Issues related to aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2 Aliasing in the one dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.3 Examples of aliasing in the one dimensional case . . . . . . . . . . . . . . . . . . . . . 95
5.4 Aliasing in two or more dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.5 Examples of aliasing in the multidimensional case . . . . . . . . . . . . . . . . . . . . 114
5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.7 Signal-domain aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

6. Issues related to leakage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.2 Leakage in the one dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.3 Examples of leakage in the one dimensional case . . . . . . . . . . . . . . . . . . . . . 148
6.4 Errors due to signal-domain truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.5 Spectral impulses that fall between output array elements . . . . . . . . . . . . . . . 162
6.6 Leakage in two or more dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6.6.1 The case of 2D 1-fold periodic functions . . . . . . . . . . . . . . . . . . . . 170
6.6.2 The case of 2D 2-fold periodic functions . . . . . . . . . . . . . . . . . . . . 172
6.6.3 The general MD case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.7 Signal-domain leakage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

7. Issues related to resolution and range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.2 The choice of the array size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Contents ix

7.3 The choice of the sampling interval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

7.4 The choice of the sampling range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.5 The choice of the frequency step and of the frequency range . . . . . . . . . . . . 189
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

8. Miscellaneous issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.2 Representation of discontinuities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.3 Phase related issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.4 Symmetry related issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
8.5 Jaggies on sharp edges as aliasing or reconstruction phenomena . . . . . . . . . 210
8.6 Sub-Nyquist artifacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
8.7 Displaying considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
8.8 Numeric precision considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

Appendices

A. Impulses in the continuous and discrete worlds . . . . . . . . . . . . . . . . . . . . . . . 247

A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
A.2 Continuous-world impulses vs. discrete-world impulses . . . . . . . . . . . . . . . 247
A.3 Impulses in the spectral domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
A.4 Impulses in the signal domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

B. Data extensions and their effects on the DFT results . . . . . . . . . . . . . . . . . . . 265

B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
B.2 Method 1: Extending the input data by adding new values beyond the
original range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
B.3 Method 2: Extending the input data by denser sampling within the
original range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
B.4 Method 3: Extending the input data by adding zeroes after each value
(zero packing) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
B.5 Method 4: Extending the input data by replicating it . . . . . . . . . . . . . . . . . . 269
B.6 Method 5: Extending the input data by replicating each of its elements . . . . 269
B.7 Method 6: Extending the input data by adding zeroes beyond its original
range (zero padding) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
B.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
x Contents

C. The roles of p and q and their interconnections . . . . . . . . . . . . . . . . . . . . . . . . 273

C.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
C.2 The one dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
C.3 Generalization of p and q to the multidimensional case . . . . . . . . . . . . . . . . 280
C.3.1 Multidimensional generalization in the continuous world . . . . . . . . 281
C.3.2 Multidimensional generalization in the discrete world . . . . . . . . . . 285

D. Miscellaneous remarks and derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299

D.1 The periodicity of the input and output arrays of the DFT . . . . . . . . . . . . . . 299
D.2 Explanation of the element order in the DFT output array . . . . . . . . . . . . . . 300
D.3 The beating effect in a sum of cosines with similar frequencies . . . . . . . . . . 302
D.4 Convolutions with a sinc function which have no effect . . . . . . . . . . . . . . . . 306
D.5 The convolution of a square pulse with a sinc function . . . . . . . . . . . . . . . . 307
D.6 A more detailed discussion on Remark A.2 of Sec. A.3 . . . . . . . . . . . . . . . . 310
D.7 The effect of the sinc lobes on its convolution with a
sharp-edged function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
D.8 On the order of applying the sampling and truncation operations . . . . . . . . 315
D.9 Relation of the DFT to the CFT and to Fourier series . . . . . . . . . . . . . . . . . 317
D.9.1 Examples illustrating the relation of the DFT to the CFT and
to Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
D.10 The 2D spectrum of a rotated bar passing through the origin . . . . . . . . . . . 332

E. Glossary of the main terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335

E.1 About the glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
E.2 Terms in the signal domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
E.3 Terms in the spectral domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
E.4 Miscellaneous terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345

List of the main relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353

List of notations and symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355

List of abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
 Preface

The continuous Fourier transform (CFT) and its discrete counterpart, the discrete
Fourier transform (DFT), are probably among the most widely used mathematical tools in
science and engineering. Their fields of application are so diverse and numerous that it
would be practically impossible to list them exhaustively in a short paragraph. It is not
surprising, therefore, that almost anyone studying or working in a scientific or
technological discipline had or will have the opportunity to use them one day or another.

The present book originates from the author’s own experience as a frequent user of the
DFT. Although several excellent textbooks already exist for teaching the DFT and its
applications, it turns out that many practical issues of great importance to the users can
still hardly be found in the existing literature. Most of the existing textbooks do not
explain in sufficient detail the various pitfalls and artifacts that very often await users of
the DFT, while in other books these topics are not easy to find, being dispersed between
several sections and chapters. The situation in the multidimensional (MD) case is even
worse, since many books defer this subject to a later chapter and only treat it there briefly
as a straightforward extension of the one-dimensional case, without considering the
particularities of the DFT pitfalls and artifacts in the MD case. Moreover, one can rarely
find in the literature a detailed discussion on practical questions such as how to place the
true unit values along the axes when plotting the DFT results, or similar practical issues
that are often acquired only through the user’s own experience, sometimes simply by trial
and error. The aim of the present book is, therefore, to close these gaps in the existing
literature.

This book is not intended to be a self-contained textbook for teaching the basics of the
Fourier theory, a subject which is already covered by many excellent books. It rather
intends to assist users who already have a basic knowledge of the DFT, but are not
sufficiently familiar with its practical limitations, artifacts and pitfalls. The present book
can be used, therefore, as a useful additional reference for students and researchers in
various branches of science and engineering who have to work on applications of DFT.
For this reason we have preferred throughout this book an informal style based on a
pictorial, intuitive approach over a rigorous mathematical treatment, and we have
intentionally avoided a purist’s approach based on theorems and proofs. And just like the

xi
xii Preface

author’s previous books on the theory of the moiré phenomenon, this volume, too,
contains a large number of illustrative examples and figures, some of which are visually
fascinating and even spectacular.

The material covered in this book includes a wide range of subjects, ranging from the
aliasing and leakage phenomena to less widely known topics such as symmetry related
issues and the DFT artifacts that may result thereof, or the influence of the various
possible types of data extension on the DFT results. It is our aim in this book to attract the
reader’s attention to the many possible pitfalls and sources of error in the use of DFT,
including potential errors due to the different CFT or DFT definitions being used in
different books and software packages, the need for input and output data reorganizations,
etc. And most importantly, we show the users how to correctly interpret the DFT results
they obtain, and how to distinguish between true spectral contents and the various artifacts
that are only due to DFT.

This book is intended for students, scientists and engineers wishing to widen their
knowledge of the DFT and its practical use. In particular, it will be very useful for
“naive” users from various scientific or technical disciplines who have to use the DFT for
their respective applications. The reader will find in this book not only a theoretical
explanation of the DFT artifacts and pitfalls in question, but also practical recipes
accompanied by many examples for the correct use of DFT in one or several dimensions.
The prerequisite mathematical background is limited to an elementary familiarity with
calculus and with the continuous and discrete Fourier theory.

This concentrated treatment of the DFT artifacts and pitfalls in a single volume is,
indeed, new, and it aims to make this book a valuable source of information for the widest
possible range of DFT users. We hope this book will not only help to “demystify” this
subject, but also encourage readers to further deepen their interest in the Fourier theory,
which is undoubtedly one of the most beautiful and enlightening branches of mathematics.

The material in this book is based on the author’s personal research at the EPFL (Ecole
Polytechnique Fédérale de Lausanne). This work would have never been possible without
the support and the excellent research environment provided by the EPFL. The author
wishes to express his gratitude to Prof. Roger D. Hersch, the head of the Peripheral
Systems Laboratory of the EPFL, for his encouragement throughout the different stages
of this work. Many thanks are also due to the publishers for their continued helpfulness
and availability throughout the publishing cycle.
 Chapter 1
Introduction

1.1 The discrete Fourier transform

The discrete Fourier transform (DFT) is the discrete-world counterpart of the
continuous Fourier transform (CFT). The DFT is widely used as a practical and efficient
computing tool for calculating numerically the Fourier transform (i.e. the frequency
spectrum) of functions or signals [Brigham88 pp. xiv, 1–3, 98]. In many circumstances
the values of our given signal are only known on a discrete grid (for example, if the signal
values have been measured at discrete intervals or obtained by a digital computer). In such
cases using DFT is the natural way for computing the Fourier transform of the given data.
But even when the given signal is a continuous function whose analytical expression is
fully known, DFT often remains the most convenient way for getting a visual glimpse at
its spectrum, especially when the analytic calculation of the continuous Fourier transform
proves to be too laborious or impractical.
One has to remember, however, that although the use of DFT as a numerical
approximation to the continuous Fourier transform is extremely simple and attractive,
there is a price to pay for this convenience: The DFT results are often “soiled” by some
artifacts which are not part of the underlying continuous Fourier transform, and which
may be quite hard to detect. Moreover, the use of DFT also requires caution due to some
inherent pitfalls that may await an inexperienced user. A basic discussion on these artifacts
and pitfalls can be found dispersed in the literature, and yet, most of these subjects are
usually learned by DFT users through their own experience, sometimes simply by trial
and error, and often they remain in the form of personal notes, recipes or “tricks”.
Moreover, these questions have been treated in the literature mostly for the one
dimensional (1D) case. But in many disciplines, including applications in optics and in
image processing, it is rather the multidimensional (MD) case, and in particular the two
dimensional (2D) case, that is being used. Although most DFT artifacts and pitfalls in the
MD case can be explained theoretically as a multidimensional extension of their 1D
counterparts, there still exist many practical issues that are specifically of interest in the 2D
or MD case.1
Our aim in this book is twofold: To show how to avoid such DFT artifacts, whenever
possible; and if they cannot be avoided, at least to show how to identify them in time, in
order not to interpret them erroneously as spectral properties of the given function. Our

1 Note that the multidimensional Fourier theory is indeed more rich, versatile and interesting than its 1D
counterpart. For example, some 2D Fourier theorems, such as the rotation theorem, the shear theorem, the
separable product theorem and the projection-slice theorem [Bracewell95 pp. 157–159, 166, 499], have
no 1D equivalents.

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 1
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_1,
© Springer-Verlag London 2013
2 1. Introduction

explanations are given from the user’s point of view rather than from the mathematical
point of view of a DFT theoretician. We favour a pictorial, intuitive approach rather than a
purely theoretic one, since an intuitive understanding may often be extremely helpful in
practical engineering situations. We mainly mention the 1D and 2D cases, due to their
particular importance in applications, and also because they lend themselves more easily to
a pictorial presentation that is the key for intuitive understanding. But our discussion
covers the general multidimensional case, too.

Remark 1.1: In most applications, DFT is implemented using one of the numerous
variants of the very efficient Fast Fourier Transform (FFT) algorithm [Briggs95, Chapter
10; Brigham88, Chapter 8; Nussbaumer82, Chapters 4–5], but it can be also computed
using other methods (see, for example, [Kammler07, Sec. 6.1]). Because the DFT results
are basically independent of the specific algorithm being used for their computation, we
prefer to use here the generic term DFT, although our discussion is also applicable to all
cases where an FFT algorithm is being used. Q

1.2 A brief historical background

The early seeds of what we call today the Fourier theory can be traced back to 18th-
century mathematicians such as Euler, D. Bernoulli, Lagrange, and others [Briggs95 pp.
2–4]. Nevertheless, it is widely accepted that the first landmark contribution to this theory
can be attributed to Jean Baptiste Joseph Fourier in his famous paper on the propagation
of heat in solid bodies, which was presented to the French Academy of Sciences (Institut
de France) in December 1807. This paper claimed, for the first time, that any arbitrary
function, whether continuous or discontinuous, can be represented as an infinite series of
sines and cosines. In spite of the skepticism and controversy aroused by this paper at that
time, an upgraded version thereof finally earned Fourier the Academy’s grand prize in
January 1812, yet with some reservations. But it was only a decade later that Fourier’s
work was first published in his book Théorie Analytique de la Chaleur [Fourier22]
(meaning in English: Analytic Theory of Heat). His prize-winning paper was finally also
published by the Academy, in two parts, in 1824 and 1826 [Fourier24, Fourier26].

Fourier’s work was pursued in the 19th and 20th centuries by famous mathematicians
such as Poisson, Dirichlet, Riemann, Lesbegue and others, who further extended its scope
but also established it on a more rigorous mathematical basis by addressing challenging
questions such as convergence, integration, etc.

But as already mentioned, the continuous Fourier transform cannot always be computed
analytically, either due to the nature of the problem at hand or due to the complexity of the
integrals, which cannot always be expressed in closed form. And indeed, a discrete version
of the Fourier theory, which offers a more practical numerical approach, has been
developed in parallel to its continuous counterpart. Eighteenth century mathematicians
such as Euler and Lagrange already used the discretized approach in their work, but it
1. Introduction 3

seems that the earliest explicit formulas of the DFT appear in a paper by Clairaut
published in 1754 [Briggs95 p. 4] (although Clairaut’s work was based on cosines alone,
and can be considered therefore as a discrete cosine transform).

The use of the Fourier theory expanded rapidly over the years and found many new
fields of application. But even the computation of the more practical discrete approach
proved to be a formidable task when done by hand. Special-purpose mechanical analog
computers were invented for this purpose at the end of the 19th century (see, for example,
[Kammler07 pp. 86–87, 294]), and when the first digital computers appeared in the middle
of the 20th century they were soon utilized for numeric DFT calculations, too.

It is important to note, however, that even when powerful computers are available, a
straightforward application of the DFT formulas incurs a significant computation cost,
since the number of operations required for an N-point DFT is proportional to N2. As a
simple example, a personal computer that performs 107 operations per second will do a
1000-point DFT in one tenth of a second, which seems fast. But if we need to apply such
a DFT to a 2000×2000-point matrix or image, which is a routine task in modern image
processing, this would already require some minutes. Obviously, before the advent of
modern digital computers such a computation would simply be prohibitive.

This illustrates, indeed, the overwhelming importance of the scientific revolution that was
brought about by the fast Fourier transform (FFT), a very efficient algorithm for the
computation of the DFT which was published in 1965 by J. W. Cooley and J. W. Tukey
[Cooley65]. Using this algorithm, the number of operations required for an N-point DFT
is reduced from N2 to Nlog2N (up to some constant factors), which is a very significant
improvement. For example, a 1024-point FFT may require about 10,000 operations rather
than about 1,000,000 operations, which is an improvement by a factor of 100; and
furthermore, this factor itself (N / log2N) continuously increases with N. The dramatic
reduction in computation cost thanks to the FFT algorithm made it possible to do Fourier
analysis on a digital computer, even when large-scale data arrays are involved.

Ironically, however, it turns out that the FFT algorithm, one of the most influential
computational tools of the 20th century, was already used by Gauss in a paper dated to
1805, and which appeared in 1866 after his death in his collected works [Heideman85;
Briggs95 pp. 5, 381; Kammler07 p. 295].

Today, the Fourier theory and in particular its discrete counterpart are used in virtually
any imaginable branch of science and technology. For example, one could mention
applications in fields as far apart as optics, electricity, acoustics, communications, signal
processing, imaging, biological engineering, hydrodynamics, heat propagation, mechanics,
geophysics, spectroscopy, statistics, mathematics, and the list is still long.

A more detailed historical account on the Fourier transform and on the DFT can be
found, for example, in [Prestini04]. [Brigham88] gives a survey of various applications of
the DFT as well as an extensive bibliography on the subject. An overview on the life of
4 1. Introduction

Fourier and his main mathematical achievements can be found, for example, in
[Bracewell86 pp. 462–464], in Chapter 1 of [Prestini04], or in dedicated books such as
[Herivel75].

1.3 The scope of the present book

This book is written for people who already know and use the discrete Fourier
transform, but are not yet sufficiently familiar with its practical limitations, pitfalls and
artifacts. For this reason, no attempt is made here to explain the fundamentals of the
Fourier theory and of the DFT. Many excellent textbooks already exist on these subjects,
and readers who wish to review basic notions of the Fourier theory are encouraged to refer
to any of these books. To mention just a few, one may consider classical works such as
Bracewell’s or Gaskill’s books [Bracewell86; Bracewell95; Gaskill78], and for the
discrete case Brigham’s and Briggs’ books on the DFT [Brigham88; Briggs95].

In particular, we do not explain here the implementation of DFT by the very efficient
FFT algorithm; this well-known subject can be found in many existing textbooks such as
[Nussbaumer82], [Loan92], [Brigham88, Chapter 8] or in the excellent overview provided
in Chapter 10 of [Briggs95]. All our discussions in this book are valid for the DFT in
general, independently of the algorithm being used to compute it.

Because this book is addressed to a wide audience in various scientific and engineering
branches, a pictorial, intuitive approach supported by mathematics is preferred over a
rigorous mathematical treatment. In many cases we give informal demonstrations rather
than formal proofs, or defer detailed derivations to an appendix.

Furthermore, it is beyond the scope of this book to state rigorously pure mathematical
issues such as the precise conditions under which each formula holds, the precise class of
functions which satisfy the required convergence conditions, etc. In fact, a large number of
functions encountered in engineering problems possess Fourier transforms only in a
generalized sense (e.g. the sine and cosine functions, the step function, etc.). Following the
practice adopted by [Bracewell86], [Gaskill78], and many other authors, such generalized
Fourier transforms are manipulated in the same fashion as ordinary transforms, and we
make no distinction between the two: It is to be understood that the term “Fourier
transform” means “generalized Fourier transform” if we are dealing with functions that
violate the existing conditions (see, for example, [Gaskill78 p. 184]). Some good
references are given, though, to help the interested readers find a more detailed
mathematical discussion on these questions.

Finally, it should be noted that although we occasionally use questions related to image
processing to illustrate our discussion, this book has not been written with any specific
DFT application in mind. In fact, our principal aim is to present the material in a general,
application-independent way. Consequently, a full discussion on the various applications
of the DFT remains beyond the scope of the book. A survey on the different fields of
1. Introduction 5

application of the DFT in science and technology can be found in [Brigham88, pp. 1–3].
The short, incomplete list at the end of Sec. 1.2 above may give the reader a small glimpse
into the diversity of the various fields in which the DFT has found applications.

1.4 Overview of the following chapters

Chapter 2 provides the background and the basic notions for the entire book. It starts
with an overview of the various definitions and notations that are being used in the
literature for the continuous Fourier transform and for its discrete counterpart. Although
the different variants used in the literature are essentially equivalent in terms of the Fourier
theory, each of them gives slightly different results, and one should be aware of this fact to
avoid any possible confusion, especially when working with several textbooks, references
or Fourier software packages. After sorting out the notations issue, Chapter 2 reviews the
basic rules which govern, respectively, the continuous and the discrete Fourier transforms,
both in the 1D case and in the 2D and MD cases. Then it proceeds with a brief
presentation of the graphical development of the DFT, and finally it reviews the main
potential pitfalls that may await us when using the DFT as an approximation to the CFT,
and which are the subject of the following chapters.

Chapter 3 deals with the data reorganizations that may be needed before or after the
application of the DFT or the inverse DFT (IDFT). As we will see, the discrete input or
output sequences are not necessarily organized the way we would expect them to be in the
continuous world. Therefore, in order to avoid misinterpretations or other potential pitfalls
and errors, one may need to reorganize these discrete sequences correctly before
interpreting or displaying them. This is explained first for the 1D case and then for the 2D
and MD cases, and illustrated by several examples and figures.

Chapter 4 explains how to interpret the DFT results in terms of true units along the
axes. The input and output arrays of the DFT consist of a list of values (real or complex
numbers), that are indexed by an integer running for example between 0,...,N–1 (or, in the
MD case, by M such indices). But when presenting or plotting the data we would like to
mark along the axes the true values in terms of the units that are being used in the original
continuous function (meters, seconds, etc.) or in its spectrum (cycles per unit, or Hz).
Furthermore, there also exists the question of how to assign the true units along the
vertical axis, i.e. what is the connection between the values obtained in the output array of
the DFT and the true heights in the spectrum of our original function. Each of these
questions is treated separately, and various examples and figures are provided to illustrate
the discussion.

Chapter 5 is dedicated to the most important and well-known DFT artifact, aliasing.
Aliasing is one of the main sources of discrepancy between the DFT results and the
continuous Fourier transform. As in the previous chapters we start with the simpler 1D
case, and only then we proceed to the more general (and more interesting) MD case.
6 1. Introduction

Many illustrative examples and figures are also provided, some of which are quite
surprising and sometimes even spectacular.

In Chapter 6 we focus our attention to the second major source of discrepancy between
the DFT results and the continuous Fourier transform, the leakage artifact. Once again we
start with the 1D case, and only then we proceed to the more general MD case. Both cases
are accompanied by examples and figures to illustrate the discussion.

Then, in Chapter 7 we explain how to choose optimal values for the various parameters
that are related to the resolution and the range of the data, both in the signal and in the
spectral domains. We discuss the significance of each of these parameters, and show the
interrelations between them.

Finally, in Chapter 8 we group together some further questions and potential pitfalls that
may await the unaware DFT users, but which do not fit into any of the previous chapters.
These include, among other topics, the representation of discontinuities in the input and
output arrays of the DFT; issues that are related to the phase when using the polar (i.e.
magnitude and phase) representation of the complex-valued spectrum; symmetry related
issues and the DFT artifacts that may result thereof; artifacts that may occur due to
sampling and reconstruction of continuous-world signals (jaggies and sub-Nyquist
artifacts); and the various displaying considerations that should be taken into account in
order to obtain an optimal graphic representation of the DFT results.

The main body of the book is accompanied by several appendices:

In Appendix A we briefly review the mathematical concept of an impulse, both in the

continuous and in the discrete worlds, and we explain how impulses should be treated in
the discrete world in order to obtain a good approximation to the continuous Fourier
transform.

Appendix B examines the most straightforward ways to extend the input data when
higher precision is required, and explains how each of these ways influences the DFT
results (including the possible effects on the DFT artifacts). It is shown that some of these
methods are indeed useful and advantageous, depending on one’s aims and on the nature
of the data, while others are rather misleading and may introduce significant artifacts.

In Appendix C we study in more detail the discrete-world properties of cases involving

periodic functions. In particular, we discuss the discrete counterparts of the period and the
frequency, explain their roles and their interconnections in the signal and spectral domains,
and provide their multidimensional generalization.

In Appendix D we group together various issues, including the derivations of several

results that we preferred for different reasons not to include in the main body of the work.

And finally, Appendix E provides a glossary of the most important terms that have been
used in the present book.
1. Introduction 7

1.5 About the graphic presentation of sampled signals and discrete data

Because graphic presentation of discrete data is essential to the understanding of many

issues throughout this book, we find it necessary to give here a short introductory section
on this subject, before delving into our actual work in the following chapters.

Many different ways exist for presenting graphically discrete data such as sampled
signals, but in this section we mainly focus on the methods being used in this book. We
start with the 1D case, where we have to plot the discrete function yk = g(xk) with constant
xk increments of Δx. Then we continue with the 2D case, where we have to plot the discrete
function zk,l = g(xk,yl) with constant xk and yl increments of Δ x and Δ y, respectively,
and finally we proceed to the MD case, in which we have to plot the discrete function
zk = g(xk), xk being an M-dimensional discrete vector with constant increments along each
of its M components. Further details can be found in references on the visual display of
quantitative data such as [Cleveland93; Cleveland94; Wright07], or in the user manuals of
mathematical software packages such as Mathematica®, Matlab®, Maple®, etc.

1.5.1 Graphic presentations in the 1D case

In the 1D case, the most obvious way to plot a discrete function yk = g(xk) consists of
the symbol plot, where each successive point (xk,yk) is graphically represented by a symbol
(such as a dot or any other shape) at the corresponding coordinates; see Fig. 1.1(a).2 This
straightforward plotting method gives good visual results when the discrete data to be
plotted varies slowly, so that the eye can easily follow the successive dots in the plot.
However, when the given discrete data varies more rapidly, as shown on the right-hand
side of Fig. 1.1(a), the eye may have some difficulties in following the successive dots in
their right order. In such cases it may be useful to add to the symbol plot straight line
segments that connect between the successive points, giving a connected symbol plot as in
Fig. 1.1(b). Furthermore, if the successive points (xk,yk) are to be drawn densely, it may be
advantageous to drop the dots altogether and only draw the connecting lines, resulting in a
connected line plot, as in Fig. 1.1(c). This can give an illusion of a continuous plot (which
may be similar to the continuous graph of the underlying continuous unsampled function,
inasmuch as such a function exists3). Although this illusion could mislead one to believe
that the plotted data is continuous, this should not really be a problem when we are aware
that the graph in question represents a discrete signal. And indeed, this drawing method is
widely used throughout this book, and dots are only used when confusion may arise.
Note, however, that in this pseudo-continuous plotting method, when the points (xk,yk) to
be plotted are not sufficiently dense, abrupt jumps in the signal may be rendered by
slanted ramps rather than by sharp vertical transitions, so that an isolated impulse or a
sharp-edged pulse may look, respectively, as an isosceles triangle or trapezium (Fig. 1.2).

2 Note that we reserve the term point to the element (xi,yi) itself or its location in the plot, and use the
term dot for the actual symbol that represents this point graphically.
3 Such an underlying continuous function containing all the in-between data may exist, for example, if

our data represents the temperature yk sampled at the discrete moments xk; but it does not exist if our
data represents, for instance, the number of cars sold per agent or per country.
8 1. Introduction

Another useful graphic improvement of the symbol plot consists of adding to the latter a
vertical line plot, where each of the points (xk,yk) is connected to the horizontal axis by a
thin vertical line (Fig. 1.1(d)); it is also possible to omit the dots altogether and draw the
vertical line plot alone (Fig. 1.1(e)).4 These successive vertical lines tend to fill the area
below the graph of the discrete signal and guide the eye along the successive (xk,yk) points,
without really connecting them by a continuous line. Yet another possible variant consists
of replacing the vertical lines by adjacent vertical rectangles or boxes, like in a histogram.
However, this is not recommended for reasons that we will see in Sec. 8.5.

1.5.2 Graphic presentations in the 2D case

We now proceed to the 2D case, where we have to plot the discrete function zk,l = g(xk,yl)
with constant xk and yl increments. Here the situation is indeed more interesing, since each
discrete point must be represented by three values: its spatial coordinates xk and yl and its
value zk,l. Basically, there are two families of methods which allow us to graphically
represent such cases: (1) planar presentations, and (2) pseudo-3D presentations.

(1) In the family of planar presentation methods the paper’s surface represents a 2D
matrix or grid whose elements correspond to the spatial locations (coordinates) xk and
yl of our discrete points. The values zk,l are then represented in the plot either by using
gray levels (or colour tones), giving a density plot, or alternatively by drawing contour
lines like in a topographical map, giving a contour plot.
In a density plot, each point xk,yl is represented in the plot by a small square pixel
(corresponding to one matrix element or grid element) whose gray level represents its
value zk,l. The gray level scale can be chosen in various ways, but usually black
represents the lowest possible value, white represents the highest possible value, and
intermediate gray levels represent all the in-between values. Thus, if zk,l can only take
values between 0 and 1 (as is often the case in the image or signal domain), black may
represent the value 0 and white then represents the value 1. But if the values of zk,l can
take both positive and negative values (as is often the case when drawing the real or
imaginary valued parts of the spectral domain), black may represent the most negative

Figure 1.1: Comparison of various graphic methods for plotting 1D discrete

data, using the signal yk = cos(2πfxk), xk = k/8, k ∈Z with f = 0.5
(left-hand column) and with f = 3.75 (right-hand column). Both
of these two signals are plotted in each row using a different
graphic presentation method: (a) Symbol plot. (b) Connected
symbol plot (i.e. combination of symbol plot and connected line
plot). (c) Connected line plot. (d) Combination of symbol plot
and vertical line plot. (e) Vertical line plot.

4 In the present book vertical line plots (without dots) are used, for example, in Fig. B.1.
1. Introduction 9

2 2

1 1

-4 -2 2 4
x -4 -2 2 4
x
-1 -1

-2 -2

(a)
2 2

1 1

-4 -2 2 4
x -4 -2 2 4
x
-1 -1

-2 -2

(b)
2 2

1 1

-4 -2 2 4
x -4 -2 2 4
x
-1 -1

-2 -2

(c)
2 2

1 1

-4 -2 2 4
x -4 -2 2 4
x
-1 -1

-2 -2

(d)
2 2

1 1

-4 -2 2 4
x -4 -2 2 4
x
-1 -1

-2 -2

(e)
10 1. Introduction

2 2

1 1

-4 -2 2 4
x -4 -2 2 4
x
-1 -1

-2 -2

Figure 1.2: When plotted using a connected line plot (see the right-hand side
drawing), a discrete impulse may look as an isosceles triangle and a
sharp-edged pulse may look as an isosceles trapezium.

value, white then represents the highest positive value, and zero is represented by a
mid-gray tone.5 When the attribution of gray shades to the different zk,l values is not
obvious, one may also add a legend (see, for example, Fig. 3.6). This is also true when
using colours to represent the z k,l values.6 Finally, note that when the drawing
resolution is sufficiently high, the density plot looks almost continuous. It should be
remembered, however, that in all cases each discrete point is represented in the density
plot by a small square pixel.7 We will see the significance of this fact in Sec. 8.5.
Another point to remember when using density plots is the fact that they are very
sensitive to various types of printer-dependent artifacts, which may appear in the
printed plot as vertical or horizontal bands, gray-level undulations, or other
irregularities such as moiré effects. This may be particularly misleading in figures that
contain large spans of constant gray levels, like Fig. 3.6 or Fig. 4.7, where such
printing artifacts may be mistakenly considered as true intensity fluctuations of the
functions being plotted. For more details see also Sec. 8.7.
In a contour plot, on the other hand, the zk,l values are represented by means of level
lines, like in a topographical map. In order to resolve the relief ambiguity (hill/valley)

5 Note that in figures showing side by side the input and the output of a 2D Fourier transform (or DFT),
the density plots used for the input data and for the output data may need different gray level scale
conventions. This is the case, for example, in many figures in [Amidror09] and [Amidror07], where the
2D signal consists of a black/white structure such as a binary line grating or dot screen but the spectrum
includes both positive and negative values. In the present book, in order to avoid such situations, we
usually prefer to represent 2D binary 0/1-valued structures in the signal domain by gray and white rather
than by black and white, respectively, and to reserve black for negative values (in the spectrum). See, for
example, the binary line gratings in Figs. 5.14–5.16, 5.18, 8.6–8.9, etc.
6 When colours are being used (which is not the case in this book), a good practice consists of using the

same colour conventions as in geographical maps, where colours gradually vary from dark green (for the
lowest zk,l values), through clearer green and yellow tones (for middle-range zk,l values), and then
clearer and darker brown tones (for the heighest zk,l values). This colour scale can be further extended
according to the needs, for example by adding dark blue or violet tones in the lowest scale end, and dark
red or purple tones in the highest scale end.
7 This is also true, of course, when digitally plotting a continuous 2D function, since any continuous
function is numerically evaluated by the computer on a discrete grid.
1. Introduction 11

y y
4. 4.

2. 2.

0 0

-2. -2.

-4.
-4. -2. 0 2. 4.
x -4.
-4. -2. 0 2. 4.
x

(a) (b)

Figure 1.3: Comparison of two planar presentation methods for plotting 2D

discrete data, using the signal zk,l = sinc(xk) sinc(yl), xk = k/8, yl = l/8,
k,l ∈Z. (a) Density plot. (b) Contour plot (note that the level lines are
obtained by interpolation, and they are drawn with continuous curves
although the data is discrete).

in such a contour plot we can superpose it on top of the corresponding density plot,
which better conveys the scale of the zk,l values. Note that the level lines are obtained
by interpolations, and they are drawn with continuous curves even if they represent in
reality discrete data.
(2) The second family of methods for the graphic presentation of 2D discrete functions
consists of pseudo-3D presentations. In this case, the spatial locations xk and yl of our
discrete points are no longer represented along the standard x and y axes of the paper
surface. Instead, one tries to draw, using the rules of perspective, an oblique view of a
3D rigid model which corresponds to the given data, where the z k,l values are
represented by heights along the model’s z axis rather than by gray or colour shades.
Here, too, there exist several possible variants: For example, one may represent the zk,l
values by thin vertical lines (impulses) of the corresponding heights that emanate from
the model’s x,y plane, as in Fig. 1.4(b); by using 3D vertical boxes rather than vertical
lines (like in a 3D histogram), as in Fig. 1.4(c); or by a continuous mesh or wireframe
that represents the surface generated by the points (xk,yl,zk,l) in the 3D space, as in Fig.
1.4(d). These variants are, in fact, simple 2D generalizations of methods that we have
already discussed in the 1D case. Note, however, that the 2D generalization of the
symbol plot (i.e. representing each point (xk,yl,zk,l) by a single dot in the 3D space) is
not really effective, since such a drawing does not clearly convey the 3D location of
each point (see Fig. 1.4(a)).
12 1. Introduction

z z
1 1
y y
2. 2.
0 0
1. 1.
0 0
-2. -2.
-1. -1. -1. -1.
0 0
1. -2. 1. -2.
(a) 22. (b) 2
2.
x x

z z
1 1
y y
2. 2.
0 0
1. 1.
0 0
-2. -2.
-1. -1. -1.
-1.
0 0
1. -2. 1. -2.
(c) 22. (d) 2
2.
x x

Figure 1.4: Comparison of various pseudo-3D presentation methods for plotting

2D discrete data, using the signal z k,l = sinc(x k) sinc(y l), x k = k/4,
yl = l/4, k,l ∈Z. (a) Symbol plot. (b) Vertical line plot (nailbed); note
that a dot has been added here at the basis of each vertical line, to
clearly show its location in the x,y plane as well as its vertical direction
(positive or negative). (c) Box plot. (d) Mesh or wireframe plot.

Pseudo-3D presentations have the advantage of being very intuitive, but on the other
hand they are limited to relatively small data sets, since beyond a certain element
density the plot may become too dark or even completely black. Moreover, they often
suffer from difficulties in showing the data in the “rear” side of the 3D model,
depending on the oblique viewing angle being used.

In the present book we usually represent 2D discrete data by simple gray-level density
plots consisting of juxtaposed square gray-level pixels. It should be stressed, however, as
already mentioned in point (1) above, that the attribution of gray levels to the zk,l values
may vary according to the case. Furthermore, as explained later in Sec. 8.7, when drawing
1. Introduction 13

Fourier or DFT spectra it is commonplace to truncate the range of zk,l values being
actually plotted, so that the entire range of gray levels can be used for representing in detail
the most interesting part of the zk,l range.

1.5.3 Graphic presentations in the MD case

Proceeding now to higher dimensions, it is clear that discrete functions zk = g(xk) of

three or more dimensions cannot be directly plotted in a single drawing. Rather, they
can be represented by two or more drawings showing various projections, i.e. partial
dimensional representations of the MD data (just as we do, for example, when we plot a
2D complex-valued function by drawing separately its real-valued and its imaginary-
valued parts).

1.6 About the exercises and the internet site

Having presented the graphic methods we use to depict discrete data, we conclude this
introductory chapter with some further practical information about the present book.

First of all, for the benefit of the readers we have added at the end of each chapter a
section containing a few problems and exercises. Some of these are not merely routine
exercises, but rather intriguing or even challenging problems. Their aim is to aid the
assimilation of the material covered by the chapter, but also to develop new insights
beyond it. We therefore highly encourage readers to dedicate some time for reviewing
these problems.

This book is also accompanied by an internet site, which can be found at the following
address:
http://lspwww.epfl.ch/publications/books/dft/

This site will provide useful complementary information on the book, including various
updates or even printable versions of some of the key figures. This site will also include a
corrigendum section, where all the typos and errors to be discovered in this volume will be
posted along with their respective corrections.

* * *

Finally, a word about our notations. Throughout this book we adopt the following
notational conventions:

Sec. 3.2 — Section 2 of Chapter 3.

Sec. A.2 — Section 2 of Appendix A.

14 1. Introduction

Fig. 3.2 — Figure 2 of Chapter 3.

Fig. A.2 — Figure 2 of Appendix A.

(3.2) — Equation or formula 2 of Chapter 3.

(A.2) — Equation or formula 2 of Appendix A.

Similar conventions are also used for enumerating tables, examples, propositions,
remarks, etc.; for instance, Example 3.2 is the second example of Chapter 3.

The mathematical symbols and notations used in the present volume are listed at the end,
following the list of the main relations. A glossary of the main terms is provided in
Appendix E.
 Chapter 2
Background and basic notions

2.1 Introduction
In this chapter we review the main definitions and notations that are being used in the
literature for the continuous Fourier transform and for its discrete counterpart.
Unfortunately, no standard definition exists for the Fourier transform, and one may find in
the literature several slightly different variants. Although these variants are essentially
equivalent in terms of the Fourier theory, each of them gives slightly different results, and
one should be aware of this fact to avoid any possible confusion, especially when working
with several textbooks, references or Fourier software packages.
We start in Sec. 2.2 with the continuous Fourier transform (CFT), and then we proceed
in Sec. 2.3 to the discrete Fourier transform (DFT). In both cases we review some of the
most commonly used definitions, and present the notations that will be used in the sequel.
In Sec. 2.4 we list the basic rules which govern the CFT and the DFT, respectively, in the
1D, 2D and MD cases. Then, in Sec. 2.5 we briefly present the graphical development of
the DFT, and in Sec. 2.6 we review the potential pitfalls that may await us when using the
DFT as an approximation to the CFT, and which are the subject of the following chapters.
Finally, in Sec. 2.7 we provide a first glimpse into cases having impulsive spectra.

2.2 The continuous Fourier transform: definitions and notations

Let us start with a brief reminder on the definition of the continuous Fourier transform.
The continuous Fourier transform can be defined in several slightly different ways, all of
which are in use in the literature (see, for example, [Bracewell86 pp. 6–8; Gaskill78 pp.
181–182; Brigham88 pp. 22–23; Champeney73 pp. 8–10; Cartwright90 p. 98;
Kammler07 pp. 63–64]). According to the common convention used in optics, the Fourier
transform of a possibly complex-valued function g(x) of a real variable x is defined by:
∞
G(u) =
∫ –∞
g(x) e–i2π ux dx –∞ < u < ∞ (2.1)

where G(u), too, is a possibly complex-valued function of a real variable u. The inverse
transform is given by:
∞
g(x) =
∫ –∞
G(u) ei2π ux du –∞ < x < ∞ (2.2)

Here, the variable x may represent a quantity such as distance or time, and it is measured
accordingly in terms of units such as meters, seconds, etc. The variable u represents
frequency, and it is measured in terms of cycles per unit of x, such as cycles per meter,
I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 15
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_2,
© Springer-Verlag London 2013
16 2. Background and basic notions

cycles per second (also called Hertz), etc. In the following we will often use “Hertz” (Hz)
as a generic frequency unit, i.e. as a shorthand notation for “cycles per unit of x”.

Remark 2.1: It is important to note that in order for the Fourier transform and its inverse
to exist, i.e. in order that the integrals converge, the functions in question must satisfy
some conditions; for example, they must be absolutely integrable on the entire range of
integration. Nevertheless, it turns out that a large number of functions encountered in
engineering problems do not strictly satisfy these conditions, and they possess Fourier
transforms only in a generalized sense (we may mention, for example, the sine and cosine
functions, the step function, etc.). Following the practice adopted by [Bracewell86],
[Gaskill78], and many other authors, such generalized Fourier transforms are manipulated
in the same fashion as ordinary transforms, and we make no distinction between the two:
It is to be understood that the term “Fourier transform” means “generalized Fourier
transform” if we are dealing with functions that violate the existing conditions (see, for
example, [Gaskill78 p. 184] or [Bracewell86 pp. 8–13]). In particular, periodic functions
are treated like any other function, with the understanding that their continuous-world
spectrum consists, in fact, of impulses (whose amplitudes reflect the coefficients of the
Fourier series decomposition). A rigorous mathematical discussion on issues such as
convergence conditions, generalized Fourier transforms, etc. is beyond the scope of this
book; the interested readers can find more detailed information in references such as
[Kammler07 pp. 48–58 and Chapter 7], [Champeney87 Chapters 6–9], [Champeney73
Sec. 2.3], [Papoulis62 Chapter 2], etc. Q

In an alternative definition, the factor 2π appearing in the exponential part is lumped

together with u, giving the following definition of the Fourier transform and of its inverse:
∞
G(ω) =
∫ –∞
g(x) e–iωx dx –∞ < ω < ∞ (2.3)
∞
g(x) = 1
2π ∫ –∞
G(ω) eiωx dω –∞ < x < ∞ (2.4)

For the sake of symmetry, some authors prefer to split the constant factor 1 of Eq.
2π
(2.4) and to distribute it equally between the direct and the inverse transforms, giving:
∞
G(ω) = 1
2π ∫ –∞
g(x) e–iωx dx –∞ < ω < ∞ (2.5)
∞
g(x) = 1
2π ∫ G(ω) eiωx dω
–∞
–∞ < x < ∞ (2.6)

Note that while in the definition given by Eqs. (2.1)–(2.2) u stands for the spectral
frequency measured in cycles per unit of x (meter, second, etc.), in the definitions given by
Eqs. (2.3)–(2.4) or (2.5)–(2.6) ω is the angular frequency ω = 2πu measured in radians
per unit of x [Gaskill78 p. 182; Cartwright90 p. 98].

Finally, each of these definitions has also a variant in which the signs in the exponential
part of the direct transform and of its inverse are interchanged.
2. Background and basic notions 17

It is clear that each of these definitions gives for the same function g(x) a slightly
different Fourier transform; but all of the definitions work equally well — provided that
one always sticks to the same definition. Conversion rules from one convention to another
can be found, for example, in [Bracewell86 p. 7] or in [Champeney73 Sec. 2.2]. In the
present work we will use the definition provided by Eqs. (2.1)–(2.2), unless explicitly
mentioned otherwise.

All of these definitions can be also extended to the multidemensional case (see, for
example, [Champeney73 pp. 40–41]). Thus, in the 2D case Eqs. (2.1)–(2.2) become:
∞ ∞
G(u,v) =
∫ ∫
–∞ –∞
g(x,y) e–i2π(ux+vy) dx dy –∞ < u,v < ∞ (2.7)
∞ ∞
g(x,y) =
∫ ∫
–∞ –∞
G(u,v) ei2π(ux+vy) du dv –∞ < x,y < ∞ (2.8)

and in the general MD case they become (see, for example, [Marks09 Sec. 8.4]):
∞
G(u) =
∫ g(x) e
–∞
∞
–i2π u·x dx (2.9)

g(x) =
∫ G(u) e
–∞
i2π u·x du (2.10)

where x, u ∈RM, u·x is their scalar product, and the integration between –∞ and ∞ is done
along each of the M dimensions.

2.3 The discrete Fourier transform: definitions and notations

Returning now to the discrete case, the DFT and the inverse DFT (IDFT) are often
defined as follows [Nussbaumer82 p. 81]:1
N– 1
Gn = ∑ gk e–i2π kn/N n = 0, ... ,N–1 (2.11)
k=0
N– 1
gk = 1 ∑ Gn ei2π kn/N k = 0, ... ,N–1 (2.12)
N n=0

where gk and Gn are sequences of N possibly complex-valued numbers. Note that unlike
their continuous counterparts, the discrete definitions are valid in all cases, and they are not
subject to any particular conditions [Kammler07 p. 38].2

But just as in the continuous Fourier transform, in the discrete case, too, many slightly
different definitions can be found in the literature. For example, in some references (such

1 The constant factor 1N is required for the sake of reversibility, i.e. in order to get back by the IDFT
exactly the original input array that was fed to the DFT [Nussbaumer pp. 80–81], [Smith03 p. 577].
The reason it appears can be also understood from Remark 2.2 below.
2 This does not yet mean, however, that Eqs. (2.11)–(2.12) can approximate the continuous Fourier

transform in all cases, with no conditions on the original underlying function g(x). We will return to
this question in Sec. 2.6 below.
18 2. Background and basic notions

as [Bracewell86 p. 358] or [Kammler07 p. 196]) the constant 1 appears in the DFT

N
rather than in the IDFT, while in other references it is replaced in a more symmetric way
by the constant 1 in both DFT and IDFT (like in [Smith07 p. 109]). In each of these
N
variants the signs in the exponential parts of the DFT and of the IDFT can be also
interchanged. And even the range of the summation may be defined in various ways (for
example, in [Bracewell86 p. 362] all indices vary between –N/2 and N/2–1, while in
[Briggs95 p. 23] the indices vary between –N/2+1 and N/2).3

Also, in order to stress the connection between the discrete transforms and their
continuous counterparts,4 many authors prefer to use g(k) and G(n), i.e. the discrete
(sampled) version of the continuous functions g(x) and G(u), rather than gk and Gn:
N– 1
G(n) = ∑ g(k) e–i2π kn/N n = 0, ... ,N–1 (2.13)
k=0
N– 1
g(k) = 1 ∑ G(n) ei2π kn/N k = 0, ... ,N–1 (2.14)
N n=0

The input step of the DFT (i.e. the sampling interval Δx) and the frequency step of the
DFT, Δu, may be explicitly inserted into the definition, as follows:
N– 1
G(nΔu) = ∑ g(kΔx) e–i2π kn/N n = 0, ... ,N–1
k=0
N– 1
g(kΔx) = 1 ∑ G(nΔu) ei2π kn/N k = 0, ... ,N–1
N n=0
1 (see
And if we note that the frequency step Δu of the DFT output is given by Δu = NΔx
Eq. (4.4)), we obtain, like in [Brigham88 p. 97]:
N– 1
G( n ) = ∑ g(kΔx) e–i2π kn/N n = 0, ... ,N–1 (2.15)
NΔx k=0
N– 1
g(kΔx) = 1 ∑ G( n ) ei2π kn/N k = 0, ... ,N–1 (2.16)
N n=0 NΔx

In the two last equation pairs the integer k in g( ) has been multiplied by the step Δx (the
sampling interval), while in G( ) the index n has been multiplied by the corresponding
1
frequency step Δu = NΔx . The advantage in doing so is that in this notation the discrete
abscissa of g( ), kΔx, is no longer a simple integer index, but consists of true sampling
points along the axis, measured in terms of the true units (seconds, meters, etc.); and
similarly, the discrete abscissa of G( ), nΔu or equivalently NΔxn , is not merely an integer
index, but indeed consists of frequencies measured in cycles per second, meter, etc.
[Bracewell86 p. 362].5 This notation may be less convenient, however, in other
circumstances, for example when expressing the discrete counterparts of the classical
n
Fourier theorems; and indeed, when discussing the DFT theorems k Δ x and NΔx are
usually replaced by k and n for convenience of notation (see [Brigham88 p. 107]).

3 On the various possible indexing methods, their equivalence, and the reasons for choosing one method
rather than another, see for example [Briggs95 p. 24 and Sec. 3.1].
4 This assumed connection will be justified in Remark 2.2 and Secs. 2.5, 2.6 below.
5 A fuller and more precise discussion on this subject is provided in Sec. 4.3; see in particular Eq. (4.11)
and the paragraph which follows it.
2. Background and basic notions 19

If we further denote by xk = kΔx the k-th sampling point and by un = nΔu = n the n-th
NΔx
frequency sample measured in cycles per unit we obtain yet another variant:
N– 1
G(un) = ∑ g(xk) e–i2π xkun n = 0, ... ,N–1 (2.17)
k=0
N– 1
g(xk) = 1 ∑ G(un) ei2π xkun k = 0, ... ,N–1 (2.18)
N n=0
In some other references (such as [Smith07 p. 2]) the factor 2π appearing in the
exponential part is lumped together with un, giving the following definition of the DFT
and of its inverse (where xk = kΔx is the k-th sampling point and ωn = 2πn is the n-th
NΔx
frequency sample measured in radians per unit):
N– 1
G(ωn) = ∑ g(xk) e–ixkωn n = 0, ... ,N–1 (2.19)
k=0
N– 1
g(xk) = 1 ∑ G(ωn) eixkωn k = 0, ... ,N–1 (2.20)
N n=0

Obviously, just as in the continuous case, each of these definitions gives for the same
input a slightly different DFT; but all of the definitions work equally well — provided that
one always sticks to the same definition. Conversion rules from one convention to another
can be easily established, like in the continuous case.

In the present work we will usually use the definition provided by Eqs. (2.13)–(2.14),
unless mentioned otherwise. Note, however, that this notation does not explicitly
distinguish between the original continuous functions g, G and their discrete counterparts.
If the distinction is not sufficiently clear from the context, for example when it is desired
to compare the value of the original function G to the value of its DFT approximation at
the same point, we will denote the discrete counterparts of the functions g, G by g, G, like
in [Brigham88 p. 95]. Alternatively, one may use the sequence notation gk, Gn as in Eqs.
(2.11)–(2.12), like in [Briggs95 pp. 22–23, xv]. Yet another possibility consists of using
square brackets for the discrete case: g[k], G[n], like in [Kammler07 p. 5].

Remark 2.2: Note that the DFT can be obtained from the continuous Fourier transform
as a numerical approximation to the integrals in Eqs. (2.1)–(2.2), using the rectangular
rule for approximating integrals (see Sec. D.9 in Appendix D or [Press02 p. 508]):6
N/2–1
G(nΔu) ≈ Δx ∑ g(kΔx) e–i2π kΔxnΔu n = – N , ... ,N –1 (2.21)
k=–N/2 2 2
N/2–1
g(kΔx) ≈ Δu ∑ G(nΔu) ei2π kΔxnΔu k = – N , ... ,N –1 (2.22)
n=–N/2 2 2
Because Δu = 1 (as we will see below in Eq. (4.4)), this means, indeed, using Eqs.
NΔx
(2.15)–(2.16), that the DFT and the IDFT approximate their continuous counterparts up to
a constant factor of Δx or 1 , respectively:6
Δx
6 The choice of the index ranges here is explained in Secs. 2.5 and D.9 (see in particular Remark D.2). As
we will see, due to the N-element periodicity of the input and output arrays of the DFT and IDFT, all
index ranges of N consecutive integers are equivalent, so once the best match with the underlying CFT
is obtained we can change the index ranges and choose the most convenient ones depending on the case.
20 2. Background and basic notions

N/2–1
G(nΔu) ≈ Δx G(nΔu), where G(nΔu) = ∑ g(kΔx) e–i2π kn/N n = – N , ... ,N –1
k=–N/2 2 2
N/2–1
1
g(kΔx) ≈ Δx g(kΔx), where g(kΔx) = 1 ∑ G(nΔu) ei2π kn/N k = – N2 , ... ,N2 –1 Q
N n=–N/2

Remark 2.3: For the sake of completeness, it should be mentioned here that the DFT can
also be represented in an elegant way using matrix notation. Since its input and output are
arrays of N elements (real or complex numbers), they can be considered as N-element
vectors; and thanks to the linearity property of the DFT (see rule 1 in Sec. 2.4.3 below), it
can be represented as a product of such an N-element vector and an N×N matrix. The
matrix representation of the DFT is explained, for example, in [Briggs95 pp. 31–33] or in
[Smith07 Sec. 6.12]. This matrix representation gives a new algebraic insight into the
properties of the DFT as a linear operator, but we will not need it in the present book.
Note that just like the explicit DFT formulas that we have seen above, the matrix
representation, too, may be subject to some slight form variations. Q

All of the variants of the DFT definition can be easily extended to the multidimensional
case, too (see, for example, [Rosenfeld82 pp. 21–22], [Brigham88 Sec. 11.2] or
[Bracewell95 pp. 167–168]). For example, in the 2D case Eqs. (2.13)–(2.14) become:7
N– 1 N– 1
G(m,n) = ∑ ∑ g(k,l) e–i2π(mk+nl)/N m, n = 0, ... ,N–1 (2.23)
k=0 l=0
N– 1 N– 1
g(k,l) = ( 1 ) ∑ ∑ G(m,n) ei2π(mk+nl)/N
2
k, l = 0, ... ,N–1 (2.24)
N m=0 n=0

and in the general MD case they become (see, for example, [Marks09 Sec. 8.6.3]):

G(n) = ∑ g(k) e–i2π n·k/N n0, ... ,nM–1 = 0, ... ,N–1 (2.25)
k

g(k) = ( 1 )M ∑ G(n) ei2π n·k/N k0, ... ,kM–1 = 0, ... ,N–1 (2.26)
N n

where k = (k0, ... ,kM–1), n = (n0, ... ,nM–1), n·k is their scalar product, and the summations
go between 0,...,N–1 for each element of k or n, respectively. More generally, if each
dimension j has a different length (number of elements) Nj, then the MD case becomes:

G(n) = ∑ g(k) e–i2π k·(n/N) nj = 0, ... ,Nj –1, j = 0, ... ,M–1 (2.27)
k

1
g(k) =
N0 N M–1
∑n G(n) ei2π n·(k/N) kj = 0, ... ,Nj –1, j = 0, ... ,M–1 (2.28)

where the summations go between 0,...,Nj –1 for each element of k or n, N = (N0, ... ,NM–1),
and the vector division is performed element-wise: v/w = (v0/w0, ... ,vM–1/wM–1).

Finally, for the benefit of the DFT users, we list below the DFT and IDFT definitions as
they are implemented in some widely used mathematical software packages. We only give

7 For the sake of simplicity we assume here that the number of elements is N along each of the
dimensions. But the definitions given here can be also extended to the more general case where the
number of elements along each dimension is different, as shown below for the MD case.
2. Background and basic notions 21

here the 1D forms, the MD definitions being straightforward extensions as shown above.
For the sake of clarity, we stick here to the same letter conventions as above, even if the
letters being used in the respective package documentations are different than ours.

• In Mathematica® [Wolfram96 p. 868] the definitions being used are:

N
1
Gn = ∑
N k=1
gk ei2π (k–1)(n–1)/N n = 1, ... ,N (2.29)
N
1
gk = ∑
N n=1
Gn e–i2π (k–1)(n–1)/N k = 1, ... ,N (2.30)

• In Matlab® [Matlab02 p. 1-48] the definitions are:

N– 1
G(n+1) = ∑ g(k+1) e–i2π kn/N n = 0, ... ,N–1 (2.31)
k=0
N– 1
g(k+1) = 1 ∑ G(n+1) ei2π kn/N k = 0, ... ,N–1 (2.32)
N n=0

• In Maple® [Maplesoft] the definitions used by default are:

N
1
Gn = ∑
N k=1
gk e–i2π (k–1)(n–1)/N n = 1, ... ,N (2.33)
N
1
gk = ∑
N n=1
Gn ei2π (k–1)(n–1)/N k = 1, ... ,N (2.34)

The reason for the index shift of +1 or –1 in the respective cases is that the index range
associated with an array in these software packages is 1,...,N while the DFT indices are
supposed to run between 0,...,N–1.

2.4 Rules for deriving new Fourier transforms from already known ones

The most straightforward way to find the Fourier transform of a given function is by
using directly the definition formula, for example Eq. (2.1). However, in many cases the
computation of the integral is quite tedious. Fortunately, things may become much simpler
thanks to some general rules which tell us how we may obtain new Fourier transforms
from already known ones or from their combinations. These rules are well known in the
literature and they can be found along with their proofs in virtually any book on the
Fourier theory, usually under the name of Fourier transform properties or Fourier
transform theorems. Many of these rules subsist also in the discrete case, while others
must be adapted accordingly; but there are also rules which only apply to the continuous
case and rules which only apply to the discrete case. For the sake of convenience we
provide here (without proofs) a list of some of the most useful rules that we have collected
from various sources as cited below. We start with the continuous 1D case, then we
proceed to the more interesting 2D and MD cases, and finally we give their 1D, 2D and
MD counterparts for the discrete case. Note, however, that these lists are non-exhaustive,
and further rules can be found in the cited references. Whenever required, we have added
remarks to draw the reader’s attention to some particularities or potential sources of error.
22 2. Background and basic notions

2.4.1 Rules for the 1D continuous Fourier transform

Assume that the functions g(x) and h(x) have the Fourier transforms G(u) and H(u),
respectively, and that a, b and f are real numbers. Then we have the following general rules
(see, for example, [Bracewell86 Chapter 6], [Gaskill78 Sec. 7.3], [Brigham88 Chapter 3]
and [Chu08 pp. 175–180]):

Rule name Function Fourier transform Remarks

1. Linearity ag(x) + bh(x) aG(u) + bH(u)

1a. Scaling ag(x) aG(u)

1b. Addition g(x) + h(x) G(u) + H(u)

1
2. Dilation g(ax) a G(u/a) a≠0

2a. Frequency dilation 1 g(x/a) G(au) a≠0

a
3. Reflection g(–x) G(–u)

4. Inversion G(x) g(–u)

5. Complex conjugation g*(x) G*(–u)

6. Shift (or translation) g(x – a) e–i2π au G(u)

6a. Frequency shift ei2π fx g(x) G(u – f)

7. Convolution g(x)*h(x) G(u)H(u)

7a. Product g(x)h(x) G(u)*H(u)

8. Cross correlation g(x)+h(x) G(u)H(–u) (a)

= g(x)*h(–x)

8a. Autocorrelation g(x)+g(x) |G(u)|2

= g(x)*g(–x)

9. Real part Re[g(x)] 1

2 G(u) + 12 G*(–u)

9a. Imaginary part Im[g(x)] 1

2 G(u) – 12 G*(–u)

Remarks:
(a) The cross-correlation rule deserves particular attention in order to avoid confusion.
First of all, it should be remembered that cross correlation is not commutative,
meaning that h(x) + g(x) generally differs from g(x) + h(x). There also exists in the
2. Background and basic notions 23

literature an alternative convention for the definition of cross correlation, in which the
roles of g and h are inversed (see, for example, [Coulon84 pp. 81–82]): g(x) + h(x) =
g(–x) * h(x). In this case the resulting Fourier transform becomes G(–u)H(u).
Unfortunately, these two conventions are not compatible, and one should be aware of
this source of confusion when using different references or software packages.
Furthermore, in the case of complex-valued functions g(x) and h(x) one can also
define their complex cross correlation to be: g(x) + h(x) = g(x) * h*(–x), in which
case the resulting Fourier transform is G(u)H*(u). Note that if h(x) is a real-valued
function then H(u) is Hermitian [Bracewell86 p. 15] and therefore H*(u) = H(–u), so
that the complex cross correlation indeed reduces into real-valued cross correlation.
But in the complex case, too, some authors use the alternative convention in which the
roles of g and h are inversed (see, for example, [Weisstein99 p. 352]): g(x) + h(x) =
g*(–x) * h(x); in this case the Fourier transform becomes G *(u)H(u). For further
details see, for example, [Gaskill78 pp. 172–173, 200], [Amidror07 pp. 413–414].

Note that the rules given above are adapted to the Fourier transform as defined in Eqs.
(2.1)–(2.2); but as one could expect, some of these rules may undergo slight changes
when other conventions are being used. For example, the rules which correspond to the
definition based on angular frequency ω = 2π u (Eqs. (2.5)–(2.6)) can be found in
references such as [Chu08 Sec. 5.7], [Poularikas96 Sec. 2.2] or [Champeney73 Sec. 2.4].
Moreover, the names of the rules may also vary between different sources; for example,
our dilation rule [Kammler07 p. 138] is called similarity in [Bracewell86 p. 101], while in
[Brigham88 pp. 32–33] it is called scaling.

A similar list of rules can be also formulated for the ICFT. For example, it follows from
rules 7 and 7a above that the ICFT of G(u)* H(u) is g(x)h(x), and that the ICFT of
G(u)H(u) is g(x)*h(x). If one uses an alternative definition of the CFT such as the one
based on angular frequency, the ICFT rules, too, should be adapted accordingly.

2.4.2 Rules for the 2D continuous Fourier transform

Assume that the functions g(x,y) and h(x,y) have the Fourier transforms G(u,v) and
H(u,v), respectively, and that a, b, f1 and f2 are real numbers. Then we have the following
rules (see, for example, [Bracewell95 pp. 154–166], [Gaskill78 Sec. 9.3]):

Rule name Function Fourier transform Remarks

1. Linearity ag(x,y) + bh(x,y) aG(u,v) + bH(u,v)

1a. Scaling ag(x,y) aG(u,v)

1b. Addition g(x,y) + h(x,y) G(u,v) + H(u,v)

2. Dilation g(ax,by) 1 G(u/a, v/b) a,b ≠ 0

ab
1
2a. Frequency dilation
ab
g(x/a, y/b) G(au,bv) a,b ≠ 0
24 2. Background and basic notions

3. Reflection g(–x,–y) G(–u,–v)

4. Inversion G(x,y) g(–u,–v)

5. Complex conjugation g*(x,y) G*(–u,–v)

6. Shift (or translation) g(x – a, y – b) e–i2π (au + bv) G(u,v)

6a. Frequency shift ei2π (f1x + f2y) g(x,y) G(u – f1, v – f2)

7. Convolution g(x,y)**h(x,y) G(u,v)H(u,v)

7a. Product g(x,y)h(x,y) G(u,v)**H(u,v)

8. Cross correlation g(x,y)++h(x,y) G(u,v)H(–u,–v) (a)

= g(x,y)**h(–x,–y)

8a. Autocorrelation g(x,y)++g(x,y) |G(u,v)|2

= g(x,y)**g(–x,–y)

9. Real part Re[g(x,y)] 1

2 G(u,v) + 12 G*(–u,–v)

9a. Imaginary part Im[g(x,y)] 1

2 G(u,v) – 12 G*(–u,–v)

10. Separable product g(x)h(y) G(u)H(v) (b),(c)

11. Rotation g(x cosα – y sinα, G(u cosα – v sinα, (b)

x sinα + y cosα) u sinα + v cosα)

12. Shear g(x + by, y) G(u, v – bu) (b)

Remarks:
(a) The same remark about cross correlation as in the 1D case applies here, too.
(b) These 2D rules have no non-trivial 1D counterparts.
(c) See an example in Sec. D.10 of Appendix D.

As in the 1D case, the rules given above are adapted to the Fourier transform as defined
in Eqs. (2.7)–(2.8); when using other conventions some of the rules may undergo
respective minor changes.

Further rules for the 2D case (such as the compound shear rule, the affine rule, etc.) can
be found in [Bracewell95 pp. 158–166] and [Gaskill78 Sec. 9-3]. Note that the 2D case is
indeed richer and more interesting than its 1D counterpart: Not only it allows for more
flexibility with the 1D rules, which can be now applied along any of the axes or along
both; but it also offers new 2D rules which have no equivalents in the 1D case, like rules
10–12 above.
2. Background and basic notions 25

2.4.3 Rules for the MD continuous Fourier transform

The generalization of the 2D rules to the MD case is usually straightforward. Assume

that the functions g(x) and h(x) have the Fourier transforms G(u) and H(u), respectively,
where x = (x0, ... ,xM–1), u = (u0, ... ,uM–1) ∈RM; that a, b are real numbers; that a = (a0, ...
,a M–1), f = (f0, ... ,fM–1) ∈R M ; and that v ·w is the scalar product of the two vectors in
question. Then we have the following rules (see [Marks09, Sec. 8.4; Stein71, Chapter 1]):

Rule name Function Fourier transform Remarks

1. Linearity ag(x) + bh(x) aG(u) + bH(u)

1a. Scaling ag(x) aG(u)

1b. Addition g(x) + h(x) G(u) + H(u)

2. Dilation g(a0x0, ... ,aM–1xM–1) 1

G(u0/a0, ... ,uM–1/aM–1)
a0 a M–1
2a. Frequency dilation 1 g(x0/a0, ... ,xM–1/aM–1) G(a0u0, ... ,aM–1uM–1)
a0 a M–1
3. Reflection g(–x) G(–u)

4. Inversion G(x) g(–u)

5. Complex conjugation g*(x) G*(–u)

6. Shift (or translation) g(x – a) e–i2π a·u G(u)

6a. Frequency shift ei2π f·x g(x) G(u – f)

7. Convolution g(x)*h(x) G(u)H(u)

7a. Product g(x)h(x) G(u)*H(u)

8. Cross correlation g(x)+h(x) G(u)H(–u) (a)

= g(x)*h(–x)

8a. Autocorrelation g(x)+g(x) |G(u)|2

= g(x)*g(–x)

9. Real part Re[g(x)] 1

2 G(u) + 12 G*(–u)

9a. Imaginary part Im[g(x)] 1

2 G(u) – 12 G*(–u)

10. Separable product g(x0, ... ,xj–1)h(xj, ... ,xM–1) G(u0, ... ,uj–1)H(uj, ... ,uM–1)

10a. Separable products g0(x0)...gM–1(xM–1) G0(u0)...GM–1(uM–1) (b)

11. Affine g(Ax) 1 G(A–Tu) (c)

A
26 2. Background and basic notions

Remarks:
(a) The same remark about cross correlation as in the 1D case applies here, too.
(b) See an example in Sec. D.10 of Appendix D.
(c) This rule is an MD generalization of the 2D rules 2, 11, 12 and of all similar 2D rules
having the form g(a1x+b1y, a2x+b2y). Here, A is an M×M matrix, |A| is its determinant,
and A–T is the transpose of the inverse matrix. Of course, x and u should be treated
here as column vectors in order for the products Ax and A–Tu to make sense.

Note that the MD rules given above are adapted to the Fourier transform as defined in
Eqs. (2.9)–(2.10). The MD rules which correspond to the definition based on angular
frequency ω = 2π u can be found, for example, in [Cartwright90 pp. 115–117],
[Poularikas96 pp. 149–155] or [Champeney73 Sec. 3.6].

2.4.4 Rules for the 1D discrete Fourier transform

Many of the rules of the 1D continuous Fourier transform subsist in the discrete case,
too; but some of them must be slightly amended or even completely redefined in order to
be adapted to the discrete case. For example, the convolution, cross correlation and
autocorrelation rules may remain unchanged if we agree to replace the respective
operations by their cyclical discrete counterparts (see, for example, [Brigg95 pp. 78–89]).
Other rules, such as the dilation rule, must be entirely redefined; but there also exist new
rules which have no equivalent in the continuous case, and vice versa (see Problem 2-5).
Note that just as in the continuous case, the discrete rules may depend on the choice of
the DFT definition. The rules presented in the list below are adapted to the convention
used in Eqs. (2.13)–(2.14); but when using the convention in which the constant 1
N
appears in the DFT definition rather than in the IDFT definition, like in [Bracewell86 p.
358] or [Kammler07 p. 196], some of the rules must be modified by factor N. When the
rules obtained in this alternative convention are different, we give them in the “Alt. DFT”
column. Note that a third convention also exists, which uses 1 in both DFT and IDFT
N
definitions [Smith07 p. 109]. In this case, too, an appropriate change should be made in
those rules which depend on the DFT definition.
Assume that the sequences g(k) and h(k) have the discrete Fourier transforms G(n) and
H(n), respectively, that a and b are real numbers, and that j is an integer. Then we have the
following general rules (see, for example, [Brigham88 Sec. 6.5], [Smith07 Sec. 7.4],
[Briggs95 Secs. 3.2–3.3]):

Rule name Function DFT Alt. DFT Remarks

1. Linearity ag(k) + bh(k) aG(n) + bH(n) Same

1a. Scaling ag(k) aG(n) Same
1b. Addition g(k) + h(k) G(n) + H(n) Same
2. Background and basic notions 27

2. Dilation Complicated; see, for example, [Kammler07 pp. 205–208]

3. Reflection g(–k) G(–n) Same

4. Inversion G(k) Ng(–n) 1 g(–n)

N
5. Complex conjugation g*(k) G*(–n) Same

6. Shift (or translation) g(k – j) e–i2π jn/N G(n) Same

6a. Frequency shift ei2π jk/N g(k) G(n – j) Same

7. Cyclic convolution g(k)*h(k) G(n)H(n) NG(n)H(n)

7a. Product g(k)h(k) 1 G(n) H(n) G(n)*H(n)

N *
8. Cross correlation g(k)+h(k) G(n)H(–n) NG(n)H(–n) (a)
= g(k)*h(–k)
8a. Autocorrelation g(k)+g(k) |G(n)|2 N|G(n)|2
= g(k)*g(–k)
9. Real part Re[g(k)] 1
2 G(n) + 12 G*(–n) Same

9a. Imaginary part Im[g(k)] 1

2 G(n) – 12 G*(–n) Same

10. Periodicity g(k + jN) G(n) Same (b)

10a. Spectral periodicity g(k) G(n + jN) Same (b)

11. Stretch (zero packing) StretchL[g(k)] RepeatL[G(n)] 1 Repeat [G(n)] (c)

L
L
11a. Repeat RepeatL[g(k)] LStretchL[G(n)] StretchL[G(n)] (c)

Remarks:
(a) The same remark about cross correlation as in the continuous case applies here, too.
(b) In other words, the periodicity rules say that g(k + jN) = g(k) and G(n + jN) = G(n)
for any integer j. These rules have no equivalent in the continuous case, where the
functions g(x) and G(u) are not necessarily periodic.
(c) These rules have no equivalent in the list given above for the continuous case. The
discrete operator StretchL[g(k)] (also known as zero packing) receives a sequence g(k)
of length N and generates a sequence of length LN by adding L–1 zeroes after each
element of g(k). Similarly, the discrete operator RepeatL[g(k)] receives a sequence g(k)
of length N and generates a sequence of length LN by repeating the original sequence
g(k) L–1 more times. A full explanation of these rules can be found, for example, in
[Smith07 pp. 157–158] or for the alternative DFT definition in [Bracewell86 p. 370],
[Kammler07 pp. 201–203].
28 2. Background and basic notions

Just as in the continuous case, here, too, the names of the rules may vary between
different sources. For example, our reflection rule [Kammler07 p. 199] is called in
[Bracewell86 p. 366] reversal while in [Nussbaumer82 p. 82] it is called symmetry.

Various other rules which are valid for the discrete case alone can be found, for example,
in [Smith07 Sec. 7.4], [Briggs95 Sec. 3.3] or [Kammler07 pp. 201–212].

Note that a similar list of rules can be also deduced for the IDFT. For example, it
follows from rules 7 and 7a above that the IDFT of G(n)*H(n) is Ng(k)h(k), and that the
IDFT of G(n)H(n) is g(k)*h(k) [Cí Ń zŃ ek86 Sec. 4.2]. If one uses an alternative definition
of the DFT (such as the definition in which the constant 1 appears in the DFT rather
N
than in the IDFT), the IDFT rules, too, should be adapted accordingly.

2.4.5 Rules for the 2D discrete Fourier transform

The extension of most of the 1D DFT rules to the 2D case is straightforward. Assume
that the sequences g(k,l) and h(k,l) have the discrete Fourier transforms G(m,n) and
H(m,n), respectively, that a and b are real numbers, and that j1 and j2 are integers. Then the
following general rules can be stated (see, for example, [Briggs95 pp. 149, 173]):8

Rule name Function DFT Alt. DFT Remarks

1. Linearity ag(k,l) + bh(k,l) aG(m,n) + bH(m,n) Same

1a. Scaling ag(k,l) aG(m,n) Same

1b. Addition g(k,l) + h(k,l) G(m,n) + H(m,n) Same

2. Dilation Complicated

3. Reflection g(–k,–l) G(–m,–n) Same

4. Inversion G(k,l) N2g(–m,–n) 1 g(–m,–n)

N2
5. Complex conjugation g*(k,l) G*(–m,–n) Same

6. Shift (or translation) g(k – j1, l – j2) e–i2π (j1m + j2n)/N G(m,n) Same

6a. Frequency shift ei2π (j1k + j2l)/N g(k,l) G(m – j1, n – j2) Same

7. Cyclic convolution g(k,l)**h(k,l) G(m,n)H(m,n) N2G(m,n)H(m,n)

7a. Product g(k,l)h(k,l) 1 G(m,n) H(m,n) G(m,n)**H(m,n)

N2
**

8 For the sake of simplicity we assume here that the number of elements is N along each of the two
dimensions, but the rules given here can be also extended to the more general case where the number of
elements along each dimension is different.
2. Background and basic notions 29

8. Cross correlation g(k,l)++h(k,l) G(m,n)H(–m,–n) N2G(m,n)H(–m,n) (a)

= g(k,l)**h(–k,–l)

8a. Autocorrelation g(k,l)++g(k,l) |G(m,n)|2 N2|G(m,n)|2

= g(k,l)**g(–k,–l)

9. Real part Re[g(k,l)] 1

2 G(m,n) + 12 G*(–m,–n) Same

9a. Imaginary part Im[g(k,l)] 1

2 G(m,n) – 12 G*(–m,–n) Same

10. Separable product g(k)h(l) G(m)H(n) Same

11. Periodicity g(k + j1N, l + j2N) G(m,n) Same (b)

11a. Spectral periodicity g(k,l) G(m + j1N, n + j2N) Same (b)

12. Stretch StretchL[g(k,l)] RepeatL[G(m,n)] 1 Repeat [G(m,n)] (c)

L
L2
12a. Repeat RepeatL[g(k,l)] L2 StretchL[G(m,n)] StretchL[G(m,n)] (c)

Remarks:
(a) The same remark about cross correlation as in the continuous case applies here, too.
(b) In other words, the periodicity rules say that g(k + j1N, l + j2N) = g(k,l) and G(m + j1N,
n + j2 N) = G(m,n) for any integers j1 , j2 . These rules have no equivalent in the
continuous case, where the functions g(x,y) and G(u,v) are not necessarily periodic.
(c) These rules have no equivalent in the list given above for the 2D continuous case. The
discrete operator StretchL[g(k,l)] (also known as zero packing) receives a 2D matrix
g(k,l) of N×N elements and generates a 2D matrix of LN×LN elements by replacing
each element of the original matrix by an L×L matrix having the original element at
the first position and zeroes everywhere else. Similarly, the discrete operator
RepeatL[g(k,l)] receives a 2D matrix g(k,l) of N×N elements and generates a 2D matrix
of LN×LN elements by repeating the original matrix to both directions into a block
consisting of L×L copies of the origimal matrix.

2.4.6 Rules for the MD discrete Fourier transform

The generalization of the 2D rules to the MD case is usually straightforward. Assume

that the MD sequences g(k) and h(k) have the Fourier transforms G(n) and H(n),
respectively, where k = (k0, ... ,kM–1), n = (n0, ... ,nM–1) ∈ZM; that a, b are real numbers; that
a = (a0, ... ,aM–1), j = (j0, ... ,jM–1) ∈ZM; and that v·w is the scalar product of the two vectors
in question. Then we have the following rules:9

9 For the sake of simplicity we assume here that the number of elements is N along each of the
dimensions, but the rules given here can be also extended to the more general case where the number of
elements along each dimension is different.
30 2. Background and basic notions

Rule name Function DFT Alt. DFT Remarks

1. Linearity ag(k) + bh(k) aG(n) + bH(n) Same

1a. Scaling ag(k) aG(n) Same

1b. Addition g(k) + h(k) G(n) + H(n) Same

2. Dilation Complicated

3. Reflection g(–k) G(–n) Same

4. Inversion G(k) NMg(–n) 1 g(–n)

NM
5. Complex conjugation g*(k) G*(–n) Same

6. Shift (or translation) g(k – j) e–i2π j·n/N G(n) Same

6a. Frequency shift g(k) ei2π j·k/N G(n – j) Same

7. Cyclic convolution g(k)*h(k) G(n)H(n) NMG(n)H(n)

7a. Product g(k)h(k) 1 G(n) H(n) G(n)*H(n)

NM *
8. Cross correlation g(k)+h(k) G(n)H(–n) NMG(n)H(–n) (a)
= g(k)*h(–k)

8a. Autocorrelation g(k)+g(k) |G(n)|2 NM|G(n)|2

= g(k)*g(–k)

9. Real part Re[g(k)] 1

2 G(n) + 12 G*(–n) Same

9a. Imaginary part Im[g(k)] 1

2 G(n) – 12 G*(–n) Same

10. Separable product g(k0, ... ,kj–1) G(n0, ... ,nj–1) Same
h(kj, ... ,kM–1) H(nj, ... ,nM–1)

10a. Separable products g0(x0)...gM–1(xM–1) G0(u0)...GM–1(uM–1) Same

11. Periodicity g(k + jN) G(n) Same (b)

11a. Spectral periodicity g(k) G(n + jN) Same (b)

12. Stretch (zero packing) StretchL[g(k)] RepeatL[G(n)] 1 Repeat [G(n)] (c)

L
LM
12a. Repeat RepeatL[g(k)] LMStretchL[G(n)] StretchL[G(n)] (c)

Remarks:
(a) The same remark about cross correlation as in the continuous case applies here, too.
2. Background and basic notions 31

(b) In other words, the periodicity rules say that g(k + jN) = g(k) and G(n + jN) = G(n)
for any integer vector j. These rules have no equivalent in the continuous case, where
the functions g(x) and G(u) are not necessarily periodic.
(c) The discrete operator StretchL[g(k)] (also known as zero packing) receives an MD
sequence g(k) of N elements to each direction and generates an MD sequence of LN
elements to each direction by replacing each element of the original sequence by a
sequence of L elements to each direction having the original element at the first
position and zeroes everywhere else. Similarly, the discrete operator RepeatL[g(k)]
receives an MD sequence g(k) of N elements to each direction and generates an MD
sequence of LN elements to each direction by repeating the original sequence to all
directions into an MD block with L copies of the original sequence to each direction.

2.5 Graphical development of the DFT — a three-stage process

As already mentioned in Remark 2.2, the DFT definition can be derived as a numerical
approximation to the definition of the continuous Fourier transform. But in fact, the DFT
can be developed in many different ways. For example, Chapter 2 in [Briggs95] shows in
detail how the DFT can be obtained as an approximation to the Fourier transform, as an
approximation to the Fourier series coefficients of a periodic function, as a least-squares
approximation to N given data pairs by a sum of sines and cosines, and as the Fourier
transform of an impulse train. There also exists in the literature a very appealing graphical
approach illustrating the development of the DFT from the CFT (see, for example,
[Brigham88 Chapter 6], [Gonzalez87 pp. 95–97], [Bergland69 pp. 45–47] or [Cochran67
pp. 47–48]). Although this graphical approach may be considered less rigorous, it is most
beneficial from the user’s point of view since it offers an intuitive understanding of the
theory, without recurring to detailed mathematical expressions and formulations. In the
present section we will review this graphical approach very briefly; further details can be
found, for example, in the references mentioned above.

Consider Fig. 2.1. The first row of this figure schematically shows a “suitably well
behaved” function g(x)10 and its continuous Fourier transform G(u); and the last row of
the figure shows their discrete counterparts, namely the sequence g(k Δ x) and its DFT
G(nΔu). The in-between rows of the figure illustrate the intermediate steps which lead us
from the continuous case to the discrete one. As shown in the figure, the DFT
approximation is obtained from the original CFT by means of three simple operations:

1. Sampling the original continuous function.

2. Truncating the sampled function into a finite range.
3. Sampling the resulting continuous spectrum.

10Namely, a function whose continuous Fourier transform can be expected to be reasonably well
approximated by DFT; see Sec. 2.6.
32 2. Background and basic notions

The first stage is illustrated by rows (a)–(c) of our figure. Row (a) shows the original
function g(x) and its CFT G(u). Row (b) shows a sampling impulse train with sampling
interval of Δx and impulse height of 1, along with its CFT, which is an impulse train with
impulse interval of 1/Δx and impulse height of 1/Δx (see, for example, [Brigham88 p. 25]).
As shown in row (c), the sampled version of the original signal g(x) is obtained by
multiplying the signals in rows (a) and (b). The resulting spectrum (CFT) is, therefore,
according to the convolution theorem, the convolution of the spectra in rows (a) and (b),
i.e. a periodic repetition of replicas of G(u), as shown indeed in row (c). Note that due to
this convolution the resulting spectrum in row (c) is now periodic with period F = 1/Δx,
unlike the original spectrum G(u) in row (a) which is not necessarily periodic. The
possible overlapping between consecutive replicas in this spectrum (which is, in reality,
additive) is the source of the aliasing artifact of the DFT, as we will see in Chapter 5.

The second stage is illustrated by rows (c)–(e) of the figure. Row (d) shows a
rectangular truncation window of height 1 and length R = NΔx (i.e. N sampling intervals)
and its CFT, which is a narrow sinc function with amplitude R (see, for example,
[Brigham88 p. 24]). As shown in row (e), the truncated version of the sampled signal is
obtained by multiplying the signals in rows (c) and (d). The resulting spectrum (CFT) is
again, according to the convolution theorem, the convolution of the spectra in rows (c) and
(d); as we can see in row (e), this spectrum is a somewhat distorted version of the
spectrum of the sampled signal shown in row (c). This distortion is the source of the
leakage artifact of the DFT, that will be discussed in detail in Chapter 6.

We now proceed to the third stage, which is illustrated by rows (e)–(g) of our figure. So
far our spectral-domain function is still continuous, as shown in row (e), and we therefore
need to sample it. Row (f) shows in the spectral domain a sampling impulse train with
sampling interval of Δu = 1/R (see point (iii) below) and impulse height of 1, along with its
inverse CFT in the signal domain, which is an impulse train with impulse interval of
1/Δu = R and impulse height of R. As shown in row (g), the sampled version of the CFT
of row (e) is obtained by multiplying the spectra in rows (e) and (f). The resulting signal
is, according to the convolution theorem, the convolution of the signals in rows (e) and (f),
namely a periodic repetition of replicas of the sampled and truncated signal, as shown
indeed in the signal domain of row (g). Note that due to this signal-domain convolution
the resulting signal in row (g) is now periodic with period R = NΔx, unlike the original
function g(x) in row (a) which is not necessarily periodic. And as we have already seen
above, the resulting DFT, shown in the spectral domain of row (g), is periodic with period
F = 1/Δx. Although the signal-domain period R and the spectral-domain period F are not
necessarily equal, each of them consists of exactly N sampling intervals (respectively,
R = NΔx and F = NΔu; see point (iii) below).

A few remarks concerning this process are now in order:

(i) Formally, all the rows of Fig. 2.1 subsist in the continuous world (including the
impulse trains, the sampled functions and the sampled spectra), and the Fourier
transform used in all rows is the CFT. But row (g) can be also interpreted as being
2. Background and basic notions 33

discrete, by simply ignoring the continuous zero-valued intervals between the

individual impulse locations along the x and u axes (see, for example, [Granlund95 p.
192]). Similarly, row (c) can be also interpreted as a continuous Fourier transform of a
discrete signal11 [Kammler07 p. 11; Smith07 pp. 207–209]; or, considering this row
in the opposite direction, as a Fourier series development of the periodic signal shown
in the right-hand column.
(ii) Although the original function g(x) and its CFT G(u) in row (a) are not necessarily
periodic, their resulting discrete counterparts shown in row (g) are periodic. As we
have seen, this happens due to the replications caused by the spectral-domain
convolution in row (c) and the signal-domain convolution in row (g). If g(x) and/or
G(u) are periodic, their periods may or may not coincide with the DFT periodicity; the
consequences will be discussed later, for example in Chapter 6 and Sec. 7.4.
(iii) The spectral-domain sampling interval in row (f) is not chosen arbitrarily. We take it
to be Δu = 1/R in order that the impulse interval in the signal domain of row (f), and
hence the period of the discrete signal in row (g), be exactly equal to the length of the
truncation window, 1/Δu = R. This choice also guarantees that each period of the DFT
contains exactly N discrete elements, just like each period of the discrete signal g(k)
(since one spectral period, F = 1/Δx, contains N intervals of length Δu = 1/R = 1/NΔx).
(iv) The second step (signal-domain truncation) is required here since the length of the
input array is limited, meaning that when we sample the original continuous-world
signal g(x) we must stop the sampling operation at the input array borders. Note,
however, that although the length of the output array is limited, too, a dual spectral-
domain truncation is not included here. The reason is that unlike the input sampling
process, the DFT operation does not necessarily stop at the borders of the output
array; rather, if higher frequencies are present, they will be folded over cyclically to the
other end of the output array, causing spectral domain aliasing (see Chapter 5).12
(v) On the flexibility in the order we perform the operations in the figure, see Sec. D.8.
(vi) We have used here Fig. 2.1 as a didactic tool for illustrating graphically the
relationship between the CFT and the DFT. Note, however, that in this figure the
original function g(x) is symmetric about the origin, which suggests a symmetric
truncation in row (d). As we can see in row (g), this implies a centered DFT definition
where the integer summation index varies between –N/2 to N/2–1. But when g(x) is
causal (i.e. g(x) = 0 for x < 0 [Smith07 p. 133]), it may be more natural to use in row

11Also called in the literature Discrete-time Fourier transform [Smith03 p. 145; Smith07 p. 208].
12 From another point of view, our second step (signal-domain truncation) is used here to avoid
overlapping of far-away tails of the original signal with the signal’s replicas (that are generated by the
signal-domain convolution in stage 3), a problem known as signal-domain aliasing (see Sec. 5.7). In
other words, without the signal-domain truncation in row (e) of Fig. 2.1 we would have in the signal
domain of row (g) an overlapping problem analogous to the spectral-domain overlapping in row (c). If
we also wish to avoid the overlapping problem in the spectral domain of row (c), known as spectral-
domain aliasing, a dual truncation step (spectral-domain truncation) may be added before step 1 (see
end of Sec. 5.2). Adding this spectral-domain truncation step to our 3-step process would make this
process more symmetric; but this new step is not completely harmless (see Sec. 6.7), and the decision
whether or not to add it is left to the users. If this spectral-domain truncation of G(u) is indeed desired,
it will be typically done by low-pass filtering the continuous input signal g(x) prior to its sampling.
34 2. Background and basic notions

(d) a truncation window going between x = 0 and x = R, as shown in Fig. 2.2. This
leads naturally to a non-centered DFT definition where the integer summation index
varies between 0 and N–1, as in Eqs. (2.13)–(2.14) above.13 In the most general case,
where g(x) is neither symmetric about the origin nor causal (see Fig. 2.3), there is no
natural reason to prefer either of the truncation windows; we will therefore stick to the
non-centered one, which leads to the indexing convention used in Eqs. (2.13)–(2.14).
In fact, this non-centered indexing is the standard convention used by virtually all
software implementations of the DFT (see, for example, the partial list given at the end
of Sec. 2.3); furthermore, this convention is used by such DFT implementations in all
circumstances, including for symmetric functions g(x). The practical implications of
this fact are very important, and they will be discussed at length in Chapter 3.
(vii) A final remark concerns the amplitudes of the resulting signals in the last row.
Although we started in the first row with a signal g(x) and a CFT whose maximum
amplitudes are c and d, respectively, we see that the corresponding amplitudes in the

Figure 2.1: A schematic drawing illustrating graphically the development of the DFT
from the CFT. (a) The original function g(x) and its CFT, G(u). (b) A
sampling impulse train with sampling interval of Δx and impulse height of 1,
and its CFT, which is an impulse train with impulse interval of 1/Δx and
impulse height of 1/Δx. (c) The sampled version of the original signal g(x) is
obtained by multiplying the signals in rows (a) and (b). The resulting
spectrum (CFT) is, therefore, the convolution of the spectra in rows (a) and
(b), i.e. a periodic repetition of replicas of G(u). (d) A rectangular truncation
window of height 1 and length R = NΔx (i.e. N sampling intervals) and its
CFT, which is a narrow sinc function with amplitude R. (e) The truncated
version of the sampled signal is obtained by multiplying the signals in rows
(c) and (d); the resulting spectrum (CFT) is the convolution of the spectra in
rows (c) and (d). This spectrum is a somewhat distorted version of the
spectrum of the sampled signal shown in (c). (f) A spectral-domain
sampling impulse train with sampling interval of Δu = 1/R and impulse
height of 1, along with its inverse CFT (in the signal domain), which is an
impulse train with impulse interval of 1/Δu = R and impulse height of R.
(g) The sampled version of the CFT of row (e) is obtained by multiplying
the spectra in rows (e) and (f). The resulting signal in the signal domain is
the convolution of the signals in rows (e) and (f), which is a periodic
repetition of replicas of the sampled and truncated signal. Row (g) shows
indeed the discrete counterparts of the function g(x) and of its CFT G(u)
from row (a), namely, the corresponding discrete sequence g(kΔx) and its
DFT. The two pairs of vertical dotted lines going down throughout the
figure indicate the extent of the input and output arrays of the DFT. Note
that this figure is provided here for didactic reasons only; most software
implementations of the DFT use non-centered truncation and indexing, as
shown in Figs. 2.2 and 2.3.
13Note, however, that in this case the CFT of the truncation window has a non-zero imaginary part due
to the shift theorem (see point 6 in Sec. 2.4.1 above); this does not have a major impact on the results
(see [Brigham88 pp. 94, 100]) since it only affects the phase of the spectrum, but not its magnitude.
2. Background and basic notions 35

Signal domain Spectral domain

g(x) c G(u) d

x u
0 0
(a)
1 1
Δx
... ... ... ...
x u
0 Δx – Δ1x 0 1
Δx
(b)
1d
c Δx

x u
–1 – 1 0 1 1
0 Δx Δx 2 Δx 2 Δx Δx
(c)

1 R

x u
– R 0 R 0
2 2
(d)
Rd
c Δx

x u
R R – Δ1x – 1 0 1 1
–
2
0 2 2 Δx 2 Δx Δx
(e)
1
R
... ... ... ...
x u
–R 0 R 0
Δu = R1
(f)
g(kΔx) G(nΔu) Rd
Rc Δx

x u
R R – 1 0 1
–
2
0 2 Δx 2 Δx 2 Δx Δu = R1
(g)
N elements N elements
36 2. Background and basic notions

discrete case, as shown in the last row of the figure, are scaled by R and R/Δx = N,
respectively. These scaling factors with respect to the original continuous Fourier pair
shown in row (a) occur due to the CFTs in rows (b) and (d) and the ICFT in row (f),
which introduce amplitude scaling factors as indicated in each of these rows. The
factor R can be eliminated in both domains, for example, by choosing in row (d) a
truncation window having the height of 1/R rather than 1, whose spectrum has the
amplitude 1 rather than R. In this case, in the signal domain there will be no scaling
difference between rows (a) and (g), but in the spectral domain a scaling factor of
F = 1/Δx will still exist. This agrees, indeed, with the results presented in Remark 2.2.

2.6 DFT as an approximation to the continuous Fourier transform

We have seen above that the DFT can be considered as an approximation to the
continuous Fourier transform. However, it is important to note right away that the discrete
theory of the DFT stands in its own right, independently of its continuous counterpart.
This discrete-world theory is in itself precise and self-consistent, and it exactly describes
the behaviour of the DFT operator on actual data sequences in the discrete world [Smith07
p. xi]. And yet, the main interest in the DFT comes from the facts that it can be used to
approximate the CFT, and that its computation is much easier. While the computation of
the CFT using the continuous-world definitions provided in Sec. 2.2 can be very
laborious, the introduction of the FFT algorithms in the mid 1960s made the computation
of the DFT very efficient and computer friendly (see Sec. 1.2). It is therefore important to
see to what extent the DFT can really approximate the CFT of the continuous-world
function underlying the data samples. Clearly, the DFT can be only an approximation,
since it is based on a finite number N of input values (samples of the original function),
and it yields the same number N of output values (discrete frequencies). But are these
discrete values themselves precise samples of the continuous Fourier transform? It turns
out that in many cases the approximation is not perfect, and there are several reasons to it.
First of all, as we have seen above, the DFT imposes a cyclic behaviour on both of its
input and output data sequences, in the sense that each of them should be understood as a
single N-element period (or “truncation window”) from a periodic sequence which
repeatedly extends to both directions ad infinitum (even if the original underlying data was
not periodic). The reasons for this periodicity are explained in Sec. D.1 of Appendix D.

Figure 2.2: Same as Fig. 2.1, except that here the original function g(x) is not
symmetric about the origin, but rather causal (i.e. zero for any x < 0).
In such cases it is more suitable to use in row (d) a truncation window
going between x = 0 and x = R. As shown in (g), this leads naturally to
a non-centered DFT definition where the integer summation index
varies between 0 and N–1, like in Eqs. (2.13)–(2.14).
2. Background and basic notions 37

Signal domain Spectral domain: Magnitude

c 1 g(x) G(u) d

x u
0 0
(a)
1 1
Δx
... ... ... ...
x u
0 Δx – Δ1x 0 1
Δx
(b)
1d
c Δx

x u
0 Δx –1 – 1 0 1 1
Δx 2 Δx 2 Δx Δx
(c)

1 R

x u
0 R R 0
2
(d)
Rd
c Δx

x u
– R 0 R R – Δ1x – 1 0 1 1
2 2 2 Δx 2 Δx Δx
Δx (e)
R 1

... ... ... ...

x u
–R 0 R 0
Δu = R1
(f)
g(kΔx) G(nΔu) Rd
Rc Δx

x u
Δx –R – R 0 R R Δu = R1 – 21Δx 0 1
2 Δx
1
2 2 Δx
(g)
N elements N elements
38 2. Background and basic notions

Hence, it is only inside this N-element window that the DFT of the sampled input data can
be expected to approximate the CFT; but even there, it will not approximate the CFT of the
underlying signal, but rather that of its periodic version (see, for example, [Granlund95
Sec. 4.9.2]). Second, as we will see in Chapters 3 and 4, the DFT results are not obtained
to scale and usually not even in the right order, so before plotting or interpreting the
results we must reorganize them and scale them correctly. And as we will see in Chapters
5 and on, the approximation provided by the DFT is limited by some inherent DFT-related
artifacts, such as aliasing and leakage, that are precisely caused by the sampling, repetition
and truncation involved. Some of these artifacts can be avoided or at least reduced by a
skillful use of the DFT, but in any case one should be aware of their existence in order to
avoid possible pitfalls and misinterpretations.14

As we can see, the discrete approximation of continuous Fourier transforms by DFT has
its limitations. And indeed, as will be shown in the following chapters, the DFT cannot be
expected to give a reasonably accurate approximation unless (a) the original function g(x)
and its continuous Fourier transform G(u) are suitably localized15 (or periodic); (b) our
sampling interval is suitably small; and (c) we use sufficiently many sample points. To see
what happens when these conditions are violated, consider, for example, a function such as
g(x) = cos(π x2) (whose oscillations become denser as we go away from the origin). This
function clearly violates condition (a), and indeed, it is easy to see even intuitively that any
attempt to use the DFT for approximating the continuous Fourier transform in this case is
hopeless, although g(x) does have a continuous Fourier transform: G(u) = 1 (cos(π u2) +
2
sin(π u2)) = cos(π u2 – π4 ) (see [Amidror09 p. 259], or [Kammler07 pp. 165–166, 420,
557] for the more general Fresnel function g(x) = eiπ x2 = cos(π x2) + isin(π x2)).16 As

Figure 2.3: Same as Figs. 2.1–2.2, illustrating this time the more general case
where the original function g(x) is neither symmetric about the
origin nor causal. In such cases there is no natural reason to prefer
in row (d) a centered or a non-centered truncation window (if both
resulting sampling ranges are equally representative of g(x)).
Nevertheless, we choose here a non-centered truncation window like
in Fig. 2.2 because it corresponds to the non-centered indexing
standard used in virtually all software applications of the DFT.

14 Note that the issue of the DFT artifacts only arises when we consider the DFT as a tool for
approximating the continuous Fourier transform. As mentioned above, the theory of the DFT in itself
(where the input and output arrays are just considered as number sequences, and not as approximations
to anything else) is exact, and it suffers from no flaws or inconsistencies.
15 Where “localized” roughly means “decaying to both directions”, i.e. having small tails; this means, in
other words, that g(x) must be both “almost time-limited” and “almost band-limited” [Chu08 Sec.
5.8.1; Zayed93 p. 9]. As we will see in Chapters 5 and 6, violation of the first condition may lead to a
significant leakage artifact, and violation of the latter may lead to a significant aliasing artifact.
16 Note, however, that when properly defined, the discrete Fresnel function g(k) = ei2π k2/N does have a
meaningful DFT, as shown in [Kammler07 pp. 214–215].
2. Background and basic notions 39

Signal domain Spectral domain: Magnitude

c 1 g(x) G(u) d

x u
0 0
(a)
1 1
Δx
... ... ... ...
x u
0 Δx – Δ1x 0 1
Δx
(b)
1d
c Δx

x u
0 Δx –1 – 1 0 1 1
Δx 2 Δx 2 Δx Δx
(c)

1 R

x u
0 R R 0
2
(d)
Rd
c Δx

x u
– R 0 R R – Δ1x – 1 0 1 1
2 2 2 Δx 2 Δx Δx
Δx (e)
R 1

... ... ... ...

x u
–R 0 R 0
Δu = R1
(f)
g(kΔx) G(nΔu) Rd
Rc Δx

x u
Δx –R – R 0 R R Δu = R1 – 21Δx 0 1
2 Δx
1
2 2 Δx
(g)
N elements N elements
 40 2. Background and basic notions

Signal domain Spectral domain

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k n
2 g(x) = cos(πx )
2
2 CFT
• • • Sampled • • • DFT
1 1

0 0

-1 -1

-2
-4. -3. -2. -1. 0 1. 2. 3. 4. x -2
-4. -3. -2. -1. 0 1. 2. 3. 4. Hz

(a) N = 64, Δx = 1/8

Signal domain Spectral domain

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k n
2 g(x) = cos(πx )
2
2
• • • Sampled
1 1

0 0

-1 -1 CFT
• • • DFT
-2
-4. -3. -2. -1. 0 1. 2. 3. 4. x -2
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. Hz

(b) N = 32, Δx = 1/4

Figure 2.4: The continuous function g(x) = cos(π x2) and its CFT G(u) = cos(π u2 – π4 )
(plotted in continuous lines), and their discrete counterparts within a certain
limited range (plotted by black dots). (a) If the sampling interval Δx is
sufficiently small with respect to the chosen range of g(x) and we use
sufficiently many sample points, the DFT gives a reasonable approximation
to a limited range of the CFT G(u). (b) Otherwise, the aliasing (and
leakage) errors become so overwhelming that they completely destroy the
approximation. (The reason the frequency range of the DFT in (b) is
smaller than in (a) is explained by Eq. (7.2).)

shown in Fig. 2.4, the best we can do using DFT in this case is to approximate a portion
of this function, for example between x = –4 and x = 4 (admitting that beyond this range
this portion repeats itself ad infinitum, as imposed by the DFT properties). If our
sampling interval is sufficiently small and we use sufficiently many sample points, as
shown in part (a) of the figure, the DFT may still give a reasonable approximation to the
central portion of the CFT G(u). Otherwise the aliasing errors become so overwhelming
that they completely destroy the approximation, as shown in part (b) of the figure.17

17 Fig. 8.4 in [Kammler07 p. 487] shows the effect of aliasing on the signal domain in a similar case.
2. Background and basic notions 41

Mathematical estimations of the errors between the continuous Fourier transform and its
DFT approximation can be found in the literature (see, for example, Sec. 2.7 and Chapter
6 in [Briggs95] or Chapters 12 and 13 in [Morrison94]). Here, however, we will rather be
interested in understanding the nature and the origin of these errors. Our aim in the
following chapters is not to review the theory of the DFT, a subject about which the
literature already abounds, but rather to point out some of the subtle issues that may arise
when using the DFT as an approximation to the Fourier transform (including in the
special case of periodic functions, as explained in more detail in the following section). An
understanding of these subtle issues is necessary to guarantee a correct utilization of this
important tool, and a correct interpretation of its results.

2.7 The use of DFT in the case of periodic or almost-periodic functions

So far, we have considered the use of DFT as a digital approximation to the continuous
Fourier transform. It should be mentioned, however, that DFT can be also used to obtain a
digital approximation to the Fourier series development of a periodic function. This
should not be surprising, since the special case of periodic functions is already included in
the continuous Fourier theory: Simply, the spectrum of a periodic function is not
continuous, but rather consists of an impulse train where the distance between successive
impulses equals the frequency f = 1/T of the periodic function, and the impulse amplitudes
correspond (up to a possible constant scaling factor) to the Fourier series coefficients (see,
for example, [Papoulis68 p. 107] or [Amidror09 p. 377]). Therefore, once we get by DFT
an approximation to the impulsive spectrum of our given periodic function, we can use the
resulting impulse amplitudes (up to a constant scaling factor) as an approximation to the
corresponding Fourier series coefficients.

The main condition for this to work properly (i.e. for the DFT to correctly approximate
the impulse train which makes up the continuous-world spectrum of our periodic
function) is that the sampling range length R = NΔx must be exactly equal to one period of
our given function [Smith07 pp. 109, 210–212], or to an integer number of periods. As we
will see in Chapter 6, this condition guarantees that the impulses of the continuous-world
spectrum fall exactly on sampling points of the DFT spectrum, so that no leakage occurs.
A further condition is that the given periodic function should be band limited, and the
sampling rate must be at least twice the maximum frequency of our periodic function; as
we will see in Chapter 5, this condition guarantees that the resulting impulses are not
corrupted by aliasing. Under these conditions the resulting impulse amplitudes in the
DFT spectrum should be essentially identical (up to a certain scaling factor and a small
discretization error18) to the original continuous-world impulse amplitudes. This is
explained in detail, along with several examples, in Sec. D.9 of Appendix D and in

18Note that a discretization error is unavoidable here (just as in the DFT approximation to CFT), since
the DFT approximates the integrals of the Fourier coefficients in the same way as numerical integration
does; this is explained in detail in [Briggs95 pp. 53–54 and Sec. 2.4].
42 2. Background and basic notions

Appendix A. Note that for any real-valued periodic function, the maximum number of
impulses (Fourier series coefficients) that can be thus approximated by the DFT cannot
exceed N/2, where N is the array length; the other N/2 elements of the output belong to the
negative frequencies, and they are simply a Hermitian-symmetric copy of the positive
frequencies. In the most general case, where the given periodic function is complex-valued
and has no particular symmetry properties, so that its negative spectral frequencies are
independent of its positive frequencies, the maximum number of impulses (Fourier series
coefficients) that can be approximated by the DFT cannot exceed N.

A more detailed and rigorous discussion on the DFT approximation to Fourier series
coefficients can be found in [Brigham88, Sec. 5.2], and in particular in [Briggs95, Sec.
2.4]. This last reference also clearly explains why if the given periodic function contains
discontinuities (even discontinuities that fall between the two end points of the sampled
data), the DFT input at such points must be the average of the values to both sides of the
discontinuity (see also Sec. 8.2 below).

A detailed discussion on the errors in the approximation of the Fourier series

coefficients by DFT can be found in [Briggs95 pp. 180–192]. A discussion on the errors
in the inverse DFT that reconstructs the original periodic function from the Fourier series
coefficients can be found in [Briggs95 pp. 212–215].

Finally, note that DFT can be also used to approximate the spectrum of an almost-
periodic function (i.e. a function whose continuous-world spectrum is purely impulsive
but its impulse locations are not commensurable; see [Amidror09, Appendix B]).
However, in this case it is impossible for all the impulses in the continuous-world
spectrum to fall exactly on sampling points of the DFT spectrum: Otherwise, the impulse
locations would be commensurable and the given function would be periodic. This means
that leakage cannot be avoided here and the DFT approximation cannot be as good as in
the periodic case.

PROBLEMS

2-1. The different Fourier transform definitions. As we have seen in Sec. 2.2, the continuous
Fourier transform can be defined in several slightly different ways, all of which are
frequently used in the literature. Some of the differences between these definitions are
just a matter of notation (such as the different letters being used for the function names
or for the variables), but other differences are more substantial and may give, indeed,
slightly different Fourier transforms for the same original function. Such are, for
example, the choice of the constant factors that precede the integral in the definition,
and the location of the minus sign in the exponential part (it may appear either in the
direct transform or in the inverse transform).
(a) Make a list of the main definitions you can find in various sources, and try to see
what are the advantages and the shortcomings of each of them. What could be the
reasons that lead authors to prefer one definition over the others? (See, for example,
2. Background and basic notions 43

[Bracewell86 pp. 6–8; Gaskill78 pp. 181–182; Brigham88 pp. 22–23; Champeney73
pp. 8–10, Cartwright90 p. 98].)
(b) Is it possible and practical to impose a universally agreed convention, in order to
eliminate once and for all the inconvenience caused by the coexistance of several
different and non-compatible versions in the literature?
2-2. Conversion rules between the different Fourier transform definitions. Suppose you have
found in the literature a table containing various functions g(x) and their Fourier
transforms G(u). Indeed, classical references such as [Erdélyi54], [Champeney73 pp.
20–39], [Bracewell86, Chapter 21], [Oberhettinger90], [Gradshteyn07, Sec. 17.2], etc.
provide a vast selection of such Fourier pairs, which could be used as a good starting
point when trying to find the analytic CFT expression of a given function. But before
using such a table, you should always check under which conventions it has been
established. It often happens that the conventions used in the reference in question are
not the same as those you are using; such a mismatch can be, indeed, a source of many
pitfalls and errors for an unaware user. To avoid such pitfalls, identify which of the
different Fourier transform variants you are using, and always stick to the same
convention. It is also useful to establish the conversion rules for going back and forth
between your favorite convention and the other main conventions; a small selection of
such conversion rules can be found, for example, in [Bracewell68 p. 7] or in
[Champeney73 pp. 8–10]. Such conversion rules will allow you to easily convert
Fourier transforms from one convention to another, and to safely use Fourier
transform tables that are provided in various sources by different authors.
2-3. Conversion rules between the different DFT definitions. The same is also true for the
Discrete Fourier transform. Unfortunately, here, too, no standard definition exists, and
one may find in the literature several slightly different variants. Although these variants
are essentially equivalent in terms of the Fourier theory, each of them may give slightly
different results, and one should be aware of this fact to avoid any possible confusion,
especially when working with several textbooks, references or Fourier software
packages. For this end, just as you did in the previous problem, identify which variant
of the DFT you are using, and always stick to the same convention. Then, establish the
conversion rules for going back and forth between your favorite convention and the
other main conventions. This will allow you to easily convert discrete Fourier
transforms from one convention to another, to work with different DFT software
packages, and to use DFT tables that are provided in the literature by different authors.
Also, make sure that the DFT variant you are using is compatible with your favorite
CFT variant; for example, if you have chosen a CFT definition with the minus sign in
the exponential part of the inverse transform, it would be wise (if possible) to use the
same convention in the DFT software, too.
2-4. What would you expect to get if you cannot follow the advice given at the end of the
previous problem, and you must use a DFT software package having a different
exponent-sign convention than your CFT? (See also the end of Sec. 3.6 and Footnote
7 in Sec. 4.4).
2-5. Transformation and combination rules for the continuous and for the discrete Fourier
transforms. The Fourier theory benefits from a certain number of rules which tell us
how we may obtain new Fourier transforms from already known ones or from their
combinations (see Sec. 2.4). These transformation and combination rules prove to be
very useful, since they allow us to find the Fourier transforms of many new functions
quite easily, without having to use the explicit definition formula to compute each
Fourier transform from scratch. As we have seen in Sec. 2.4.4, many of these rules for
44 2. Background and basic notions

the continuous Fourier transform subsist in the discrete case, too; but some of them
must be slightly amended or even completely redefined in order to be adapted to the
discrete case. Furthermore, there also exist discrete-world rules which have no
equivalent in the continuous case, and continuous-world rules that cannot be extended
to the discrete case. Can you mention a few rules from each of these categories?
2-6. Transformation and combination rules for the continuous and for the discrete Fourier
transforms (continued). One example of a continuous-world rule that cannot be easily
transferred to the discrete world is the dilation rule. Can you explain the difficulty?
(see, for example, [Kammler07 pp. 205–208]). Another well-known 2D example is the
rotation theorem, which says that when a 2D function g(x,y) is rotated by angle θ , its
CFT G(u,v), too, undergoes a rotation by the same angle [Bracewell95 p. 157]. Why
doesn’t this rule have a discrete-world equivalent?
2-7. Symmetry rules for the continuous and for the discrete Fourier transforms. An
extremely nice feature of both continuous and discrete Fourier transforms is that they
possess many symmetry properties, that prove to be very helpful. For example, we give
below some of the most useful symmetry rules for the 1D CFT:
(1) If g(x) is real-valued and even, then its CFT G(u), too, is real-valued and even.
(2) If g(x) is real-valued and odd, then its CFT G(u) is imaginary-valued and odd.
(3) If g(x) is even, then its CFT G(u), too, is even.
(4) If g(x) is odd, then its CFT G(u), too, is odd.
(5) If g(x) is real-valued, then its CFT G(u) is Hermitian.
(6) If g(x) is imaginary-valued, then its CFT G(u) is Anti-Hermitian.
The main symmetry rules for the 1D continuous case can be found, for example, in
[Bracewell86 pp. 14–16]. Their discrete counterparts can be found in [Bracewell86 p.
366] or in [Briggs95 pp. 74–78]. Note that sometimes the symmetry rules are listed
together with the transformation and combination rules (see, for example, the table
provided in [Briggs95 pp. 88–89]).
(a) Are there any difficulties in transferring these symmetry rules from the continuous
world to the discrete world or vice versa?
(b) Can we reformulate these symmetry rules using “if and only if” rather than “if”?
Hint: See, for example, [Cooley69 p. 79] or [Cooley70 p. 317].
(c) How can these symmetry rules be extended to the 2D or MD cases? Can you think
of any difficulties in transferring such 2D or MD symmetry rules from the continuous
world to the discrete world? Hint: See Sec. 8.4.
2-8. Do the transformation and combination rules (see Problem 2-5 and Sec. 2.4) vary
depending on the Fourier transform definition being used, or are they independent of
the definitions? Do the symmetry rules mentioned in the previous problem vary
depending on the chosen Fourier transform definition? Would you give the same
answers in the discrete case, too?
2-9. Analog vs. digital. It is often said that the real world is analog, while computations are
digital. In particular, analog signals and spectra belong to the real world, while their
digital counterparts are the results of computation.
(a) Is this always true?
(b) Is it possible to approach continuous-world signals and spectra at will by
sufficiently increasing the resolution of digital computation?
2-10. What is the difference between digital and discrete? and what is the difference
between analog and continuous? Hint: See in the glossary.
 Chapter 3
Data reorganizations for the DFT and the IDFT

3.1 Introduction
While the continuous Fourier transform takes as its input a function g(x) and returns as
its output another function G(u), the discrete Fourier transform takes as its input a
sequence of N values and returns as its output another sequence of N values. However, if
we consider the input and output sequences of the DFT as a digital approximation to the
continuous-world functions g(x) and G(u), we must be aware of an important difference
between the continuous and discrete cases, which may be a source of many pitfalls and
errors. While the continuous functions g(x) and G(u) are generally defined throughout a
continuous domain along their x or u axes, the input and output sequences of the DFT are
only defined on a discrete domain consisting of N consecutive points, and outside this
domain they are considered to behave periodically. Furthermore, for reasons we will see
below the discrete values which correspond to x = 0 or u = 0 are not necessarily located in
the center of the input or output sequences of the DFT. Therefore, in order to present or
plot these discrete sequences the way they would be presented in the continuous world,
one may need to reorganize them, i.e. to reorder the sequence elements differently (or
more precisely, to shift the N-element sequence along its periodic replications until the
elements corresponding to x = 0 or u = 0 fall where they are expected to be with respect to
their continuous-world counterparts). This is also true in the 2D and MD cases, where the
same reasoning applies along each dimension individually.
In this chapter we explain the data reorganizations that may be required for
understanding the DFT and IDFT results, both in the 1D and in the MD cases. We start
in Sec. 3.2 with the reorganization of the output data of the DFT. Then in Sec. 3.3 we
proceed to the reorganization of the input data of the DFT, and in Sec. 3.4 we discuss the
required reorganizations in the case of the IDFT. Several illustrative examples are given in
Sec. 3.5, and finally Sec. 3.6 provides some concluding remarks.

3.2 Reorganization of the output data of the DFT

A typical 1D DFT algorithm receives as input a 1D array of N complex numbers (say,
samples of a given continuous complex- or real-valued function y = g(x)), and returns as
its output a 1D array of N complex numbers indexed by n = 0,...,N–1, which contains the
resulting discrete frequency spectrum. However, the frequency values in the resulting
output array are not ordered in the conventional way we use to plot the spectrum: The first
element in the array (n = 0) corresponds to the frequency 0 (that is traditionally called DC
because it represents in electrical transmission theory the direct current component
I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 45
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_3,
© Springer-Verlag London 2013
46 3. Data reorganizations for the DFT and the IDFT

[Smith03 pp. 152, 633]); the second element (n = 1) corresponds to f1, the smallest
frequency above 0 in the discrete spectrum (see Sec. 4.3); and the next elements represent
the following frequencies of the discrete spectrum, i.e. 2f1, 3f1, etc. However, this is only
true until n = N/2–1; the output array elements starting from n = N/2 and until the last
element, n = N–1, correspond to the negative frequencies of the spectrum. Thus, the last
element of the output array, n = N–1, corresponds to the negative frequency –f1, which
should conventionally be plotted just before the frequency 0; its predecessor, the element
n = N–2, corresponds to the frequency –2f1; and finally, the element n = N/2 corresponds
to –(N/2)f1, the most negative frequency in the spectrum.1 This is illustrated in Fig. 3.1(b)
and in Fig. 3.2(a). It should be mentioned that this complication only occurs when using a
DFT definition with non-centered indices, like Eqs. (2.13)–(2.14), in which the indices
vary between 0 and N–1 rather than between –N/2 and N/2–1 [Briggs95 p. 67]. But
because this is the standard convention used in virtually all software applications of the
DFT (see, for example, the partial list given at the end of Sec. 2.3), we have to accept this
fact and get used to it. The reason for this ordering of the output is explained in detail in
Sec. D.2 of Appendix D. It follows, therefore, that in order to plot the resulting spectrum
in the conventional way, with the zero frequency in the center and the negative frequencies
to its left, we have to reorganize the output array of the DFT by interchanging (swapping)
its first half (the segment containing the elements n = 0,...,N/2–1) and its second half (the
segment containing the elements n = N/2,...,N–1), as shown in Figs. 3.1(c) and 3.2(a).2
This reorganization of the spectrum is not usually done by the DFT algorithm
[Brigham88 p. 169], and it is up to the user to perform it when plotting the DFT results.3

Of course, whether or not to reorganize the plotted data in the conventional way may be
a question of personal preference. As long as we remember that the results for n > N/2
actually relate to negative frequencies, we should encounter no interpretation problems; but
in order to avoid misunderstanding of the DFT results by people who may be unaware of
this non-conventional ordering, it is usually recommended to reorganize the data whenever
plotting the DFT results. If one prefers not to do so, a possible alternative solution is to
plot a true frequency scale beneath the scale of the output index n, as done, for example, in
Fig. 9.1 in [Brigham88 pp. 168–169]. For more details on plotting the true units along the
axes, see Chapter 4.

1 In the vast majority of DFT implementations N is a power of two, or at least an even number. In these
cases the output array returned by the DFT consists of the frequency 0, followed by N/2–1 positive
frequencies (f1, ... ,(N/2–1)f1), and then by N/2 negative frequencies (–(N/2)f1, ... ,–f1). Thus, the output
array contains one more negative frequency than positive frequencies (although thanks to the cyclic
nature of the DFT the element n = N/2 corresponds in fact to both of the positive and negative
frequencies ±(N/2)f1 [Press02 p. 510]). In the rare cases where N is odd the number of positive and
negative frequencies in the output array of the DFT is identical, and the reorganization of the array
should be done accordingly [Briggs95 Sec. 3.1].
2 Note that this is also equivalent to a cyclic shift of the array (i.e. a shift with wraparound) through N/2
positions in either direction.
3 For example, the Matlab® mathematical software package provides a dedicated routine called fftshift

that can be manually called, if required, for reordering the output of the routines fft and fft2
[Matlab02 p. 1-50]. But many other applications do not even provide this manual tool, and the
reordering operation must be explicitly programmed by the user himself.
3. Data reorganizations for the DFT and the IDFT 47

A 2D DFT algorithm is the 2D generalization of its 1D counterpart: it receives as input a

2D array of N×N complex numbers (say, samples of a given continuous 2D complex- or
real-valued function z = g(x,y)), and returns as its output a 2D array of N×N complex
numbers, usually indexed by m = 0,...,N–1 and n = 0,...,N–1, which contains the discrete
2D frequency spectrum. The 2D DFT algorithm works by applying the 1D DFT in two
steps, first along each row in the horizontal direction, and then once again along each of
the resulting columns in the vertical direction. Therefore, the reorganization required in the
2D case consists of swapping the first half and the second half of each horizontal row, and
then swapping the first half and the second half of each of the resulting vertical columns.
This results in a reshuffle of the four quadrants of the output matrix provided by the 2D
DFT algorithm as shown in Fig. 3.2(b) [Brigham88 pp. 244–247; Bracewell86 p. 380].4

This can be easily extended to any MD DFT, if we remember that the MD DFT
algorithm is based on successive applications of the 1D DFT along each of the M
dimensions. The required MD reorganization consists therefore of swapping the first half
and the second half of each row along the first dimension, then along the second
dimension, and so on until the M-th dimension.

3.3 Reorganization of the input data of the DFT

While the reorganization of the output data (the discrete frequency values) obtained by
the DFT is widely discussed in the literature, it is less commonly mentioned that
sometimes a reorganization is also required in the input data. This is, again, a consequence
of the non-centered DFT definition (i.e. the DFT definition with non-centered indices)
which is used by virtually all DFT software applications. Indeed, just as a non-centered
DFT gives the frequency 0 in the first element of its output array, so does it expect to get
the sample belonging to x = 0 (the origin) in the first element of its input array (see Eqs.
(2.13)–(2.14)). But if we wish to sample a function g(x) within a symmetric range –r...r
and feed it to the DFT, the sample belonging to x = 0 will not be located in the first
position of the input array (k = 0), but rather in the position k = N/2. Therefore, in such
cases a reorganization (or rather unreorganization5) of the input array must be performed
before feeding it to the DFT, as illustrated in Fig. 3.3.6 More generally, if the sample

4 As pointed out in [Brigham88 p. 244], this shuffle consists of a right circular shift of the four
quadrants through two quadrants. It should be stressed, however, that no rotation is involved, and each
quadrant remains in its original orientation.
5 Note that in this case the original input array is already organized (i.e. ready to be plotted), so that we
rather need to unreorganize it (i.e. to adapt it to the DFT convention). But because unreorganization
involves precisely the same swapping as reorganization, the term reorganization can be used in both
cases. For the sake of clarity, we will also use the more explicit but non-standard terms “DFT-
organization” and “display-organization” whenever confusion may arise.
6 Without this reorganization, the input function will be considered by the DFT as being laterally shifted,
resulting in a corresponding change in the spectrum according to the shift theorem (see point 6 in Sec.
2.4.4 or [Brigham88 p. 107, 114]). But when the complex-valued spectrum is expressed in terms of
magnitude (absolute value) and phase, this lateral shift will only affect the phase, while the magnitude
will remain unchanged; see Sec. 8.3.
 48 3. Data reorganizations for the DFT and the IDFT

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
-x
g(x) = e step(x)
1
• • • Sampled
0.5

-0.5

-1
0. 1. 2. 3. 4. 5. 6. 7. 8.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT 1 • • • DFT
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
output reorg. output reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

Figure 3.1: Example illustrating the need for DFT output reorganization
when plotting the DFT results. (a) The continuous function
g(x) = e–x step(x), represented here by a continuous line, and its
discrete counterpart, represented by black dots, which is
obtained by sampling g(x) within the range 0...8 with a step of
Δx = 0.25, so that N = 32. (b) The CFT (continuous Fourier
 3. Data reorganizations for the DFT and the IDFT 49

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
output reorg. output reorg.
0.5 and vert. sca. 0.5 and vert. sca.

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(d)

transform) of g(x), G(u) = 1/(1+[2π u]2) – 2π ui/(1+[2 π u]2)

[Bracewell86 p. 418], represented here by a continuous line,
and its discrete counterpart, represented by black dots, as
obtained by applying DFT to the discrete sequence shown
in (a), without any corrections. (c) Same as (b), after
reorganization of the DFT output. (d) Same as (c), after having
applied to the DFT output the required vertical scaling
correction, too, as explained in Sec. 4.4; note the close match
between the corrected DFT and the CFT. In each of the plots
the horizontal axis at the bottom corresponds to the continuous
function being plotted, and the horizontal axis at the top
corresponds to the integer index of the discrete data sequence
that is represented by the dots. (Note that in (a) the value of the
sampled function at the discontinuity point of g(x), x = 0, is
neither 0 nor 1, but rather the midvalue 0.5; see Sec. 8.2.)

belonging to x = 0 appears in any other position of the input array, the array should be
cyclically shifted so as to bring x = 0 to the position k = 0; this may be considered as a
non-standard reorganization. But if the sample belonging to x = 0 is already located in
the first element of the input array, like in Fig. 3.1, then no reorganization of the input
array is necessary. This could be the case, for example, when the input array contains a
signal that was obtained from measurements that can be considered as starting from x = 0
[Smith07 p. 133], or when sampling a causal signal, i.e. a signal that is only non-zero for
x ≥ 0, such as the function g(x) = e–x step(x) of Fig. 3.1 [Brigham88 pp. 167–170]. The
output array of the DFT, on its part, should always be reorganized, as explained above, for
display purposes. Further examples illustrating the need for input and output
reorganizations in 1D and 2D cases are given in Sec. 3.5. One may also see Example 9.3
and Fig. 9.5 in [Brigham88 pp. 175–177], or for the 2D case, Example 11.7 and Fig. 11.9
in [Brigham88 pp. 248–250]. It is important to note, however, that while a wrongly
reorganized DFT output only gives a wrongly displayed output, a wrongly reorganized
DFT input gives indeed a wrong DFT output (see also Sec. 3.6).
 50 3. Data reorganizations for the DFT and the IDFT

Before reorganization After reorganization

I II II I
• •
• • • •
• • • • • n • • • • • n
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
... N N ... N ... ... N
0 2 –1 2 N –1 2 N –1 0 2 –1

(a) The 1D case

Before reorganization After reorganization

n n
N
N –1 7 2 –1 7
6 6 II I
...

...

5 IV III 5
N
2 4 0 4
N–
2 1 3 N –1 3
2 I II 2
...

...

1 1 III IV
N
0 0 2 0
m m
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
... N N ... N ... ... N
0 2 –1 2 N –1 2 N –1 0 2 –1

(b) The 2D case

Figure 3.2: Graphical illustration of the output reorganization operation: (a) in the 1D
case, and (b) in the 2D case. Both cases are drawn here at a very low
resolution, with N = 8, so that the individual elements (dots in (a) and
square pixels in (b); see Sec. 1.5) can be clearly seen. In the 1D case the
reorganization consists of swapping the first and second halves of the
array (which is equivalent to a cyclical shift of N/2). In higher dimensions
the reorganization consists of swapping half spaces along each dimension
(which is equivalent to a cyclical shift of N/2 along each dimension).
3. Data reorganizations for the DFT and the IDFT 51

Finally, note that if we are only interested in the power spectrum or in the amplitude
spectrum of the given data (which is the case, for instance, when our input consists of
random or statistical data) then no input reorganizations are necessary, since the input
reorganization (being a cyclical shift of the DFT input array) only affects the phase of the
spectrum, in accordance with the discrete counterpart of the shift theorem (see point 6 in
Sec. 2.4.4 or [Brigham88 pp. 107, 114]). But this change in the phase does affect, of
course, the real-valued part and the imaginary-valued part of the spectrum.

3.4 Data reorganizations in the case of IDFT

We now arrive to the inverse DFT (IDFT), which is used to retrieve the signal-domain
data back from the frequency-domain data. Because the IDFT is computed with the same
basic algorithm as the DFT (see, for example, [Press02 pp. 511–513] for the 1D case and
[Press02 pp. 527–529] for the MD case), the reorganization issue may also arise when
using the IDFT [Brigham88 Sec. 9.4]. Let us try to clarify this point.

Obviously, due to the reversibility of the DFT, applying IDFT to the output of the DFT
must give back the original data. In other words, if we apply the IDFT to the output array
obtained by the corresponding DFT algorithm, without any reorganizations, we obtain the
same input array that was originally fed to the DFT, so that no reorganizations are
required here. And indeed, it is important to emphasize that a spectral-domain
reorganization is usually needed only when we wish to plot the resulting output array,
because this array is not ordered in the conventional way we use to display our data. But if
we need to further process the spectral data obtained by DFT and then apply the IDFT to
bring it back into the signal-domain, processing should proceed with the non-reorganized
(i.e. DFT-organized) data. In other words, whenever we wish to plot the intermediate

Figure 3.3: Example in which both input and output reorganizations are required
for correctly plotting the DFT results. (a) The same continuous
function g(x) = e–x step(x) as in Fig. 3.1, represented by a continuous
line; this time, however, its discrete counterpart, represented by the
black dots, is obtained by sampling g(x) within the symmetric range
–4...4 (with the same step Δx = 0.25, so that again N = 32). (b) The
CFT (continuous Fourier transform) of g(x), G(u) = 1/(1+[2πu]2) –
2πui/(1+[2πu]2) [Bracewell86 p. 418], represented by a continuous
line, and its discrete counterpart, represented by black dots, as
obtained by applying DFT to the discrete sequence shown in (a),
without any corrections. (c) Same as (b), after reorganization of the
DFT output. Unlike in Fig. 3.1(c), this output reorganization is
not yet sufficient in the present case. (d) Same as (b), but this
time only the input array has been reorganized before applying
the DFT. (e) Same as (b), but this time with both input and output
 52 3. Data reorganizations for the DFT and the IDFT

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
-x
g(x) = e step(x)
1
• • • Sampled
0.5

-0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT 1 • • • DFT
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
output reorg. output reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

reorganizations. (f) Same as (e), after having applied to the DFT

output the required vertical scaling correction, too, as explained
in Sec. 4.4; note the close match between the corrected DFT and
the CFT. In each of the plots the horizontal axis at the bottom
corresponds to the continuous function being plotted, and the
 3. Data reorganizations for the DFT and the IDFT 53

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 1
• • • DFT after • • • DFT after
input reorg. input reorg.
0.5 0.5

0 0

-0.5 -0.5

-1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(d)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(e)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 and vert. sca. 0.5 and vert. sca.

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(f)

horizontal axis at the top corresponds to the integer index of the

discrete data sequence that is represented by the dots. (Note that in (a)
the value of the sampled function at the discontinuity point of g(x),
x = 0, is neither 0 nor 1, but rather the midvalue 0.5; this is explained
in Sec. 8.2.)
54 3. Data reorganizations for the DFT and the IDFT

spectral data we need to prepare a display-organized copy of the array, that will only be
used for the drawings, but processing itself should continue with the DFT-organized data.
Note that IDFT should be always applied to the DFT-organized data (see for the 1D case
[Brigham88 pp. 195–199] and for the 2D case [Brigham88 pp. 252–253]). Furthermore,
the signal-domain result obtained by the IDFT contains the value belonging to x = 0 in its
first element, so if the result is expected to be symmetric about the origin, it should be
display-organized before being plotted.

3.5 Examples

Let us illustrate the above discussions by a few simple examples.

Example 3.1: 1D DFT of a symmetrically sampled function:

Consider Fig. 3.3, in which the original continuous function is the same as in Fig. 3.1,
namely g(x) = e–x step(x). The only difference between the two figures is in the way g(x)
has been sampled: While in Fig. 3.1 the sample belonging to x = 0 (the origin) is the first
element of the samples array, at the position k = 0, in Fig. 3.3 g(x) has been sampled in a
symmetric way, with the sample belonging to x = 0 in the center of the samples array, at
the position k = N/2.7 In the case of our given function g(x) there is no apparent reason for
using this symmetric way of sampling, but when g(x) is a symmetric function, such as the
function g(x) = 12 e–|x| (Fig. 3.4) or g(x) = rect(x) (Fig. 3.5), this is indeed the most natural
and usually preferred sampling way. However, this symmetric way of sampling may give
the unaware DFT user an unpleasant surprise: Although he may have performed the
output reorganization before plotting the DFT results as required, the plotted results do
not correspond to the continuous Fourier transform of g(x) (see part (c) in Figs. 3.3–3.5,
and compare with part (c) in Fig. 3.1 where this problem does not occur). As explained in
Sec. 3.3, this happens because the non-centered DFT algorithm being used expects to get
the sample belonging to x = 0 (the origin of the continuous function) in the first element
of the input array, whereas in symmetric sampling the sample belonging to x = 0 is located
in the center of the input array, at the position k = N/2.8 In such cases, in order for the
DFT results to match the CFT we need to reorganize the input array, too, namely, to
cyclically shift it (with wraparound) so that the element belonging to x = 0 be in the first
position, and the elements belonging to negative x values be placed at the right end of the
sequence. Obviously, this is not the natural way we are used to visualize the sampled
signal, and it may look particularly awkward in familiar functions such as g(x) = rect(x).
But as we can see in Figs. 3.3–3.5, when applying the DFT to the reorganized input array,

7 Note that the continuous function g(x) in Fig. 3.3 is exactly the same as in Fig. 3.1, and not a shifted
version thereof. Otherwise, the continuous Fourier transform here wouldn’t remain the same as in Fig.
3.1; rather, it would behave as predicted by the shift theorem (point 6 in Sec. 2.4.1).
8 Note that this is equivalent to a cyclical shift of N/2 elements in the input array, so that the DFT output
indeed behaves as predicted by the discrete counterpart of the shift theorem (see point 6 in Sec. 2.4.4 or
[Brigham88 p. 107, 114]) — while the continuous Fourier transform remains unchanged (since g(x) has
not been shifted).
3. Data reorganizations for the DFT and the IDFT 55

the DFT results (after the output reorganization) do match the corresponding CFT, up to a
certain vertical scaling factor that will be discussed in detail in Sec. 4.4. (In fact, after
performing this vertical scaling correction there may still remain some visible discrepancy
between the CFT and the DFT results, as we can clearly see for example in Fig. 3.3(f) or
in Fig. 3.5(f); this discrepancy is due to the aliasing and leakage artifacts, inherent to the
DFT, that will be discussed later in Chapters 5 and 6.) Q

Example 3.2: 2D DFT of a symmetrically sampled function:

The 2D counterpart of Fig. 3.5 is shown in Fig. 3.6, where the original continuous
function is the 2D unit rectangle function, g(x,y) = rect(x) rect(y). Just as in the 1D case
(Fig. 3.5), it is natural and commonplace to sample this continuous function in a
symmetric way, with the sample belonging to (x,y) = (0,0) in the center of the 2D sampling
array, at the position (N/2,N/2). However, just as in the 1D case, this symmetric way of
sampling may give trouble to an unexperienced DFT user: Once again, even if the output
reorganization is correctly performed before plotting the DFT results, the resulting plot
does not correspond to the continuous Fourier transform of g(x,y), which is the 2D sinc
function G(u,v) = sinc(u) sinc(v). But as shown in Fig. 3.6(e), if we remember to
reorganize the input array before applying the 2D DFT (so that the sample corresponding
to (x,y) = (0,0) appears in the first position of the 2D array, (k,l) = (0,0), and the 3 negative
parts of the rectangle are moved to the 3 other corners of the array, as shown in Fig.
3.2(b)) then the 2D DFT results (of course, after reorganization of the output array) are
indeed correct. Q

Example 3.3: Spectral manipulation (filtering) of a signal (see Fig. 3.7):

Suppose that we are given the signal g(x) = rect(x/2), whose spectrum is G(u) =
2sinc(2u), and that we wish to low-pass filter it by multiplying its spectrum G(u) by a
filter such as F(u) = rect(u/2). We start by sampling the original signal g(x). Because this
function is symmetric about the origin, we sample it within a symmetric range, say,
–r...r. This input array is ready for being plotted (see the signal domain plot in the top row
of Fig. 3.7), but before we pass it to the DFT we first have to DFT-organize it so that the
sample belonging to x = 0 be indeed in the first position, k = 0. We now feed the DFT-
organized input array to the DFT, and obtain in the output array a sampled version of the
spectrum G(u) = sinc(u), whose frequency u = 0 is located in the first position of the
array, n = 0. If we wish to plot this intermediate result, we can make a copy of this output
array, display-organize it, and pass it to the plotting routine (see the spectral domain plots
in the top row of Fig. 3.7); but we keep the original DFT-organized output array for the
following processing steps.
We now consider the filter F(u). Because this function is symmetric about the origin, we
sample it within a symmetric range, as shown in the second row of Fig. 3.7, and we then
DFT-organize it so that the frequency u = 0 appears in the first position. Now, we
multiply element-by-element the sampled spectrum of G(u) and the sampled filter (both of
which contain the frequency u = 0 in the first position) to obtain the filtered spectrum.
Note that in order for this element-by-element multiplication to be meaningful, both arrays
 56 3. Data reorganizations for the DFT and the IDFT

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
-|x|
g(x) = 0.5e
1
• • • Sampled
0.5

-0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT 1 • • • DFT
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
output reorg. output reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

Figure 3.4: Same as Fig. 3.3, but this time with a symmetric continuous function
g(x) = 12 e–|x|. In such symmetric cases it is indeed natural to sample
g(x) in a symmetric way, as shown in (a); but as explained in
Example 3.1, this implies the need to reorganize the input array of the
 3. Data reorganizations for the DFT and the IDFT 57

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 1
• • • DFT after • • • DFT after
input reorg. input reorg.
0.5 0.5

0 0

-0.5 -0.5

-1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(d)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(e)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 and vert. sca. 0.5 and vert. sca.

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(f)

DFT, too. In each of the plots the horizontal axis at the bottom
corresponds to the continuous function being plotted, and the
horizontal axis at the top corresponds to the integer index of the
discrete data sequence that is represented by the dots.
 58 3. Data reorganizations for the DFT and the IDFT

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
g(x) = rect(x)
1
• • • Sampled
0.5

-0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT 1 • • • DFT
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
output reorg. output reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

Figure 3.5: Same as Fig. 3.3, but this time using the familiar symmetric continuous
function g(x) = rect(x). In this case, too, it is natural to sample g(x) in a
symmetric way, as shown in (a); and as explained in Example 3.1, this
implies the need to reorganize the input array of the DFT, too. In each of the
plots the horizontal axis at the bottom corresponds to the continuous
 3. Data reorganizations for the DFT and the IDFT 59

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 1
• • • DFT after • • • DFT after
input re. input reorg.
0.5 0.5

0 0

-0.5 -0.5

-1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(d)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(e)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 and vert. sca. 0.5 and vert. sca.

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(f)

function being plotted, and the horizontal axis at the top corresponds to the
integer index of the discrete data sequence that is represented by the dots.
(Note that in (a) the value of the sampled function at the two discontinuity
points of g(x), x = –0.5 and x = 0.5, is neither 0 nor 1, but rather the
midvalue 0.5; this is explained in Sec. 8.2.)
60 3. Data reorganizations for the DFT and the IDFT

must have the same length and the same resolution, namely, the frequency step Δu and the
frequency range must be the same. If we wish to plot the resulting filtered spectrum, we
can make a copy of this output array, display-organize it, and pass it to the plotting routine
(see the spectral domain plots in the bottom row of Fig. 3.7); but we keep the original
DFT-organized output array for the following processing steps.
Finally, we feed the DFT-organized filtered spectrum to our IDFT, in order to obtain
back in the signal domain the filtered version of our original sampled signal. Once again,
the output array of the IDFT contains the sample belonging to x = 0 in its first element,
k = 0. But because the resulting filtered signal is expected to have the same symmetric
range –r...r as the original signal, it is clear that if we wish to plot it we have to send to the
plotting routine a display-organized copy of this output array (see the signal domain plot
in the bottom row of Fig. 3.7). Q

Example 3.4: Convolution by spectral multiplication [Brigham88 Ch. 10; Smith07 pp.
152–154; Oppenheim99 pp. 576–588] (see Fig. 3.8):
Suppose that we need to find the convolution of the two given signals g1(x) and g2(x). A
simple and efficient way for doing such convolutions is provided by the convolution
theorem, according to which the Fourier transform of the convolution is equal to the
product of the two individual Fourier transforms (see point 7 in Sec. 2.4.4). Using this
method, all that we have to do is to sample the two given signals, then feed each of them to
the DFT, multiply the two DFT spectra, and apply IDFT to the resulting product.9
Just as in the previous example, the sampled input signals may need to be DFT-
organized before processing starts, if they are not provided with their origin (x = 0) in the
first input array element. And once again, all the spectral processing (multiplication of the
two spectra, and then application of the IDFT) should be done using the DFT-organized
spectra; only when we wish to plot any of the spectra do we have to make a display-
organized copy thereof in order to pass it to the plotting routine. Q

3.6 Discussion
In conclusion, the reorganization of the input and/or output data is necessary since the
data ordering used by the DFT algorithm with non-centered indices is not the same as in
our usual convention for displaying or manipulating data, where 0 appears in the center,
the negative values to its left and the positive values to its right. When the input data
(obtained by sampling) does not correspond to the convention used by the DFT algorithm,
in which the value belonging to x = 0 appears in the first element of the array, a DFT-
organization is required. And when the output data provided by the DFT does not
correspond to our displaying conventions, we have to display-organize the data before
plotting it. These rules are obviously true for the 2D and MD cases, too.

9 Note, however, that in order that the cyclic convolution obtained by DFT correctly correspond to the
original continuous convolution g1(x)*g2(x) some precautions may be needed, such as zero padding of
the sampled input data. For further details, see Sec. B.7 in Appendix B or the provided references.
3. Data reorganizations for the DFT and the IDFT 61

Based on this insight, it should be clear by now in which cases a reorganization is

required. What happens, however, if we forget to reorganize the data, or if, on the contrary,
we perform a superfluous reorganization that is not required? If the wrongly reorganized
data is the output of the DFT, its plot will simply appear wrongly reorganized (see, for
example, part (b) in Fig. 3.1). But if the input of the DFT is wrongly reorganized, how will
this affect the DFT results? The answer is given, for the 2D case, in references such as
[Briggs95 p. 150], [Gonzalez87 p. 77] or [Rosenfeld82 pp. 26–27]: When applying DFT
to such a wrongly reorganized input array, each element in the output array is multiplied
by (–1)m+n, resulting in a checkerboard of positive and negative values as shown, for
example, in Fig. 3.6(c).10 This effect results from the shift theorem (see point 6 in Sec.
2.4.5) for a shift of N/2 elements along each dimension. The extension of this result to
DFTs of any other dimension is straightforward (see, for example, part (c) in Figs. 3.3–
3.5 for the 1D case). Therefore, if we expect to obtain a rather smooth result, but our
DFT-based processing yields a violently jumpy result consisting of alternating positive
and negative values (which may sometimes look completely dark when densely plotted),
one may reasonably suspect that this is related to a reorganization problem in the input of
the DFT in question.

Figure 3.6: A 2D generalization of Fig. 3.5, plotted at the same low resolution so that the
individual pixels can be clearly seen. The continuous rectangular function
g(x,y) = rect(x) rect(y) is sampled in a symmetric way, but once again this
implies the need to reorganize the input array of the DFT, too. (a) The
discrete counterpart of the rectangular function g(x,y), as obtained by
sampling it within the symmetric range –4...4 along both axes (with the same
step Δ x = Δ y = 0.25, so that again N = 32). (b) The 2D DFT of the 2D
discrete data sequence shown in (a), without any corrections. (c) Same as
(b), after reorganization of the DFT output. As explained in Sec. 3.3, this
output reorganization is not yet sufficient in the present case. (d) Same as
(b), but this time only the input array has been reorganized before applying
the DFT. (e) Same as (b), but this time with both input and output
reorganizations. (f) Same as (e), after having applied to the DFT output the
required vertical scaling correction, too, as explained in Sec. 4.4. In each of
the plots the horizontal axis at the bottom and the vertical axis at the left hand
side correspond to the continuous function, while the horizontal axis at the
top and the vertical axis at the right hand side correspond to the integer
indices of the 2D discrete data sequence. But unlike in the 1D case (Fig.
3.5), the original continuous function g(x,y) and its CFT cannot be added
here to the same plots for comparison. See Sec. 1.5.2 for further details on
the use of 2D density plots and on the possible printing artifacts they may
contain, and which should not be confused with true gray lavel fluctuations.
(Note that in (a) the value of the sampled function at the discontinuity points
along the rectangle’s edges is neither 0 nor 1 but rather the midvalue 0.5,
except for the 4 corners where the value is 0.25; see Sec. 8.2.)

10If the complex-valued spectrum is expressed in terms of magnitude (absolute value) and phase, this
effect will only concern the phase, while the magnitude will remain unchanged.
62 3. Data reorganizations for the DFT and the IDFT

Signal domain
0 4 8 12 16 20 24 28 31
4.
31 Legend:
3. 28
1
2. 24

1. 20

0 16 0
-1. 12

-2. 8
-1
-3. 4

-4. 0
-4. -3. -2. -1. 0 1. 2. 3. 4.

(a)
Spectral domain: Real part Spectral domain: Imaginary part
0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
2. 2.
31 31
28 28

1. 24 1. 24

20 20

0 16 0 16

12 12

-1. 8 -1. 8

4 4

-2. 0 -2. 0
-2. -1. 0 1. 2. -2. -1. 0 1. 2.

(b)
Spectral domain: Real part Spectral domain: Imaginary part
0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
2. 2.
31 31
28 28

1. 24 1. 24

20 20

0 16 0 16

12 12

-1. 8 -1. 8

4 4

-2. 0 -2. 0
-2. -1. 0 1. 2. -2. -1. 0 1. 2.

(c)
3. Data reorganizations for the DFT and the IDFT 63

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
2. 2.
31 31
28 28

1. 24 1. 24

20 20

0 16 0 16

12 12

-1. 8 -1. 8

4 4

-2. 0 -2. 0
-2. -1. 0 1. 2. -2. -1. 0 1. 2.

(d)
Spectral domain: Real part Spectral domain: Imaginary part
0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
2. 2.
31 31
28 28

1. 24 1. 24

20 20

0 16 0 16

12 12

-1. 8 -1. 8

4 4

-2. 0 -2. 0
-2. -1. 0 1. 2. -2. -1. 0 1. 2.

(e)
Spectral domain: Real part Spectral domain: Imaginary part
0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
2. 2.
31 31
28 28

1. 24 1. 24

20 20

0 16 0 16

12 12

-1. 8 -1. 8

4 4

-2. 0 -2. 0
-2. -1. 0 1. 2. -2. -1. 0 1. 2.

(f)
64 3. Data reorganizations for the DFT and the IDFT

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

2 2 2
DFT
1 1 1

-8 -4 4 8
x -4 -2 2 4 Hz -4 -2 2 4 Hz
-1 -1 -1

-2 -2 -2

×
2 2
1 1

-4 -2 2 4 Hz -4 -2 2 4 Hz
-1 -1
-2 -2

=
2 2 2
1 1 1

-8 -4 4 8
x -4 -2 2 4 Hz -4 -2 2 4 Hz
-1 -1 -1
IDFT
-2 -2 -2

Figure 3.7: Example of spectral manipulation (filtering) by DFT. First row: the
sampled version of the given function g(x) = rect(x/2) and its
spectrum G(u) = 2sinc(2u) as obtained by DFT. Second row: the
spectral low-pass filter F(u) = rect(u/2). Third row: the product of
the spectra G(u) and F(u), and its inverse DFT back in the signal
domain, showing the low-pass filtered version of the given signal
g(x). As explained in Example 3.3, both the signal-domain and
frequency-domain plots have been drawn here from a display-
organized copy of the data, but the actual processing is done using
the DFT-organized data (which is not shown in the figure). The
array size in all of the plots here is N = 128. For the sake of
convenience the discrete data is drawn using connected line plots
rather than symbol plots (see Sec. 1.5.1).

Finally, one should always be aware of the precise DFT definition being used. For
example, if the DFT definition includes a positive exponent sign in the direct transform (as
is the case, for example, in the Mathematica® software package [Wolfram96 p. 868])
while the CFT results being used have been obtained from a CFT definition with a
3. Data reorganizations for the DFT and the IDFT 65

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

2 2 2
DFT
1 1 1

-4 -2 2 4
x -8 -4 4 8 Hz -8 -4 4 8 Hz
-1 -1 -1

-2 -2 -2

×
2 2 2
DFT
1 1 1

-4 -2 2 4
x -8 -4 4 8 Hz -8 -4 4 8 Hz
-1 -1 -1

-2 -2 -2

=
2 2 2
1 1 1

-4 -2 2 4
x -8 -4 4 8 Hz -8 -4 4 8 Hz
-1 -1 -1
IDFT
-2 -2 -2

Figure 3.8: Example of convolution by spectral multiplication using DFT. First

row: a sampled version of the first given function g1(x) = rect(x/2)
and its spectrum G1(u) = 2sinc(2u) as obtained by DFT. Second
row: a sampled version of the second given function g2(x) = rect(x)
and its spectrum G2(u) = sinc(u) as obtained by DFT. Third row:
the product of the spectra G 1(u) and G 2(u), and its inverse DFT
back in the signal domain, showing the convolution g1(x)*g1(x). As
explained in Example 3.4, both the signal-domain and frequency-
domain plots have been drawn here from a display-organized copy
of the data, but the actual processing is done using the DFT-
organized data (which is not shown in the figure). The array size in
all of the plots here is N = 128. For the sake of convenience the
discrete data is drawn using connected line plots rather than symbol
plots (see Sec. 1.5.1).

negative exponent sign in the direct transform (as in [Bracewell86 p. 6]), then the DFT
results we obtain are reversed with respect to the CFT. In such cases the user must
remember to invert the resulting DFT output, in addition to all the corrections mentioned
so far (see Sec. 8.3 for more details).
66 3. Data reorganizations for the DFT and the IDFT

PROBLEMS

3-1. A student wishes to find the Fourier transform of a rotated b/w line grating, where black
is represented by the value 0 and white is represented by the value 1 (the central part of
this infinite continuous-world structure is shown in Fig. 3.9). Being too lazy to derive
this continuous Fourier transform by hand, he decides to use DFT instead. For this end,
he samples the given continuous-world image into a 64×64 discrete array, and passes
this array to a 2D DFT routine in his preferred mathematical software package. The
result obtained by plotting the 2D output array of the DFT is shown in Fig. 3.10. Help
our student to correctly interpret the results he has obtained. Hint: A commom pitfall
you should be aware of when plotting your results is that software packages do not
always choose correctly the vertical range to be plotted if you do not explicitly specify
such a range. For example, they may mistakenly interpret the weak background noise
as being the signal of interest, and treat the stronger signal values themselves as outliers
that should be truncated. But a bad vertical range may also result from the user’s
inappropriate choice. If the vertical range being used is too large, the resulting plot
may seem to be too flat or even “empty”; and if it is too small, the background noise
may appear so strong that it completely obscures the truly interesting data.
3-2. Having better adjusted the vertical range of his plot, the student obtains the DFT plot
whose real-valued part is shown in Fig. 3.11 (the imaginary-valued part is zero). What
do you think of this result? In particular, has it been reorganized correctly?
3-3. Output array reorganization. In order to get familiarized with the way your favorite
DFT software package works, try to apply DFT to a few input arrays of your choice
and to interpret the resulting output array. How can you tell if the output array you get
is already reorganized or not? Hint: It is usually easy to identify the DC element in the
resulting output array, since it is often (but not always) stronger than the other
elements; furthermore, if your input sequence is purely real-valued, the output array
must be Hermitian, meaning that its real-valued part must be symmetric to both sides of
the DC element. If you see that the DC is located in the first element of the output
array, you may guess that the output array is not reorganized. But if the DC is located
in the center of the array and surrounded by symmetric elements to both sides (in
terms of their real-valued part), you may guess that the output array is already
reorganized.
3-4. Input array reorganization. Apply your favorite DFT software package to the same
input arrays as in the previous problem, but this time after having reorganized them.
How does the input reorganization affect the DFT results? When is the input
reorganization necessary?
3-5. Missing or superfluous input array reorganization. Suppose you forgot to reorganize
the input array, or that, on the contrary, you have reorganized it although no such
reorganization was required.
(a) What do you expect to see in the DFT output? Hint: See Sec. 3.6.
(b) It is easy to design an input signal that remains unchanged after the input
reorganization. For example, in the 1D case any signal whose first and second array
halves are identical will do. What are the respective conditions in the 2D and MD cases?
(c) In cases where the input signal remains unchanged after the input reorganization, a
missing or superfluous input-array reorganization will have no effect on the DFT
output. How does this agree with your answer in part (a)?
3-6. Missing or superfluous input array reorganization. What do you expect to get in
Examples 3.3 and 3.4 if you forget to correctly reorganize the input arrays?
3. Data reorganizations for the DFT and the IDFT 67

8.

4.

-4.

-8.
-8. -4. 0 4. 8.

Figure 3.9: A rotated line grating (see Problem 3-1).

0.2
60
0

-0.2 40
0

20 20
40

60 0

Figure 3.10: The DFT plot obtained by the student in Problem 3-1.

2
1
60
0
-1
-2 40
0

20 20
40

60 0

Figure 3.11: The DFT plot obtained by the student in Problem 3-2.
68 3. Data reorganizations for the DFT and the IDFT

3-7. When do you expect the output array of the DFT to be purely real-valued, and when
will it be complex valued?
3-8. Consider Figs. 3.2(b) and 3.6. While in Fig. 3.6 the origin is located in the middle of a
pixel and the two axes, too, pass through the middle of pixels, in Fig. 3.2(b) the axes
seem to pass between pixels. How do you explain this contradiction? Is any of the two
figures flawed by an error? Hint: There exist different ways to interpret pixels of the
discrete world with respect to the corresponding continuous-world image. A pixel is a
2D (or MD) entity having a side dimension of Δ x units (in the signal domain) or Δ u
units (in the spectral domain), but in fact each pixel represents one ideal sampling
point of the continuous-world space. The question is, therefore, which point within the
pixel corresponds to the true sampled point in the continuous world: is it the center of
the pixel? or its bottom left point? or any other point? There exist different possible
conventions, and each of them is valid and useful. But confusion may occur if we are
not aware of the convention being used. The lines in Fig. 3.2(b) are simply separation
lines that illustrate the division of the 2D array into 4 quadrants; but these lines do not
intend to represent the axes. If we admit that the center of each pixel corresponds to a
continuous-world sampling point, then in Fig. 3.2(b), too, the two main axes will pass
through the center of the bottom-left pixel of the first quadrant, which corresponds to
the continuous-world origin. The true axes will be therefore shifted by half a pixel with
respect to the shown separation lines. See also the 1D equivalent in Fig. 3.2(a), where
the vertical line in the center is located half a pixel away from the true vertical axis.
3-9. Figs. 3.1 and 3.3 show the DFT results obtained for the same original continuous-
world function g(x) = e–x step(x). The difference between the two cases is in the choice
of the sampling range: in Fig. 3.1 g(x) has been sampled within the range 0...8, while
in Fig. 3.3 g(x) has been sampled within the symmetric range –4...4.
(a) Explain the differences between the resulting DFT spectra in both cases, and their
fidelity to the original continuous-world spectrum G(u).
(b) What would you expect to get in an intermediate case where g(x) is sampled within
the range –2...6, so that the discontinuity is located elsewhere in the input array? You
may try to use the shift theorem (point no. 6 in Sec. 2.4.1 or 2.4.4).
3-10. Can you explain the input and output data reorganizations of the DFT using the
figures of Sec. 2.5, which illustrate the graphic development of the DFT from the CFT?
3-11. Can the rules of Sec. 2.4.4 help in the explanation of the input and output data
reorganizations? Hint: See Footnote 6 in Sec. 3.3.
3-12. Multiplication of spectra. Consider Figs. 3.7 and 3.8, in both of which the spectrum in
the third row is obtained as a product of the spectra in the first and second rows.
Trying to compute this product, a student argues that by multiplying the real-valued
parts of the first two rows one gets the real-valued part of the third row, and by
multiplying the imaginary-valued parts of the first two rows he gets the imaginary-
valued part of the third row. Is this reasoning correct? Under what conditions?
3-13. In order to make sure that you master the data reorganizations explained in this
chapter, apply your favorite DFT software package to some functions g(x) whose CFT
G(u) you already know, and see if the plotted results you get correspond (up to some
scaling factor) to the expected plot of G (u). For this end it may be convenient to
choose your Fourier pairs g(x), G (u) from a Fourier transform table which also
provides their graphical representations (such as the tables in [Bracewell86, Chapter 21]
or [Champeney73 pp. 20–39]). Try also to use the IDFT in order to get g(x) back
from G(u). If your DFT results do not yet fully match the expected continuous-world
results, try this exercise once again after you have read the following chapters.
 Chapter 4
True units along the axes when plotting the DFT

4.1 Introduction
As already mentioned above, the DFT is often used as a tool for the numerical
calculation of the Fourier transform (the frequency spectrum) of a given continuous
function. For this end, the original continuous function in question must be sampled, and
the resulting N-element array (or in the MD case, N×...×N-element array) is fed as input to
the DFT. The resulting N-element (or N×...×N-element) output array contains the discrete
counterpart of the desired frequency spectrum, and after the appropriate reorganization
(see Chapter 3) it can be plotted to illustrate graphically the resulting spectrum.
However, there still exists a significant obstacle on our way before we can graphically
represent the results obtained by the DFT in a meaningful way: Both the input and output
arrays of the DFT consist of a list of values (real or complex numbers), that are indexed
by an integer running for example between 0,...,N–1 (or, in the MD case, by M such
indices).1 But in our graphical plot we would like to mark along the axes the true values in
terms of the units that are being used in the original continuous function (meters, seconds,
etc.) or in its spectrum (cycles per unit, or Hz). For the input array this is usually not a
problem, at least if the sampling of the original continuous function is done by the user
himself, so that he knows the correspondence between the input array indices and the true
coordinates of the original function. However, in the case of the output array, it is not
always evident how to correctly replace in our graphical plot the integer indices of the
reorganized output array (say, 0,...,N–1 along each axis) by the true frequencies.
Furthermore, there also exists the question of how to assign the true units along the
vertical axis, i.e. what is the connection between the values obtained in the output array of
the DFT and the true heights in the spectrum of our original function. In the following
three sections we will treat each of these questions separately. Some illustrative examples
are provided at the end of the chapter, in Sec. 4.5.

4.2 True units for the input array

Although the connection between the true values along the axis of the given function and
the index of the DFT’s input array is usually known, let us illustrate the situation with a
simple 1D example. Suppose we are given the following continuous symmetric function
(see Fig. 4.1(a)):

1 Note that in some applications array indices run between 1,...,N rather than between 0,...,N–1. This is the
case, for example, in the Mathematica® software package [Wolfram96 p. 41].

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 69
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_4,
© Springer-Verlag London 2013
70 4. True units along the axes when plotting the DFT

⎧ sin(πx) x≠0
g(x) = sinc(x) = ⎨ πx (4.1)
⎩ 1 x=0
If we wish to feed this function to the DFT, we first need to sample it to create the
N-element input array. Since our function g(x) is symmetric, we may want to sample it
within the symmetric range –R/2...R/2, where R is a chosen real number, so that the indices
k = 0,...,N–1 of the input array correspond to the sampled values x = –R/2 ,..., x = R/2.
However, because the input array’s length N is usually even (most often a power of 2), it
follows that if we assign to k = 0 the point x = –R/2 and to k = N–1 the point x = R/2, the
center of our symmetric function (the point x = 0) will not be sampled. If this is
undesirable, the best solution is to assign to k = 0 the point x = –R/2 and to k = N/2 the
point x = 0;2 in this case we will have one less sample in the positive end of the x axis,
since the maximum positive value being sampled, which is assigned to k = N–1, is
x = R/2 – R/N rather than x = R/2. The index k = 0,...,N–1 therefore corresponds to the x
values of –R/2, ... ,R/2 – R/N, and the step Δ x between successive x values (i.e. the
sampling interval) is given, thanks to the following simple connection [Gonzalez87 p. 96]:

R = NΔx (4.2)

by Δx = R/N. Therefore, the correspondence between the k-th element in the input array
and the true x value is given in this example by:

x = –R/2 + kR/N (4.3)

A simple verification of this relationship shows, indeed, that for the first array element,
k = 0, we have x = –R/2; for the element k = N/2 we have x = 0; and for the last element
k = N–1 we have x = R/2 – R/N.

Figure 4.1: True units along the axes when plotting the DFT input and output. (a) The
continuous function g(x) = sinc(x) (represented by a continuous line), and
its discrete counterpart (represented by black dots) that is obtained by
sampling g(x) within the symmetric range –4...4 with a step of Δx = 0.25,
so that N = 32. (b) The CFT of g(x), G(u) = rect(u) [Bracewell86 p. 415]
(represented by a continuous line), and its discrete counterpart
(represented by black dots) as obtained by applying DFT to the discrete
sequence shown in (a), after both input and output reorganizations.
(c) Same as (b), after having applied to the DFT output the required
vertical scaling correction, too, as explained in Sec. 4.4; note the close
match between the corrected DFT and the CFT. In each of the plots the

2 Having the value x = 0 at the index k = N/2, i.e. as the first element in the second half of the input
array, is indeed consistent with our reorganization convention: As we have seen in Chapter 3, the
frequency 0 is also located at the same index n = N/2 in the output array (after its reorganization). Note
that keeping the same conventions for the input and output arrays is desireable due to the reversibility
of the Fourier transform (and of the DFT), since our original function g(x) can be also regarded, up to
an axis inversion, as the spectrum of the output function G(u) (see point 4 in Sec. 2.4.1).
 4. True units along the axes when plotting the DFT 71

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
g(x) = sinc(x)
1
• • • Sampled
0.5

-0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 and v. sca. 0.5 and vert. sca.

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

horizontal axis at the top corresponds to the integer index of the discrete
data sequence that is represented by the dots, and the horizontal axis at the
bottom corresponds to the true units, i.e. the units of the respective
continuous function. (Note that in (c) the value obtained by the DFT at the
discontinuity points of G(u), u = –0.5 and u = 0.5, is neither 0 nor 1, but
rather the midvalue 0.5; this is explained in Sec. 8.2.)
72 4. True units along the axes when plotting the DFT

Note, however, that if we wish to reorganize this input array before feeding it to the DFT
(see Sec. 3.3), these indices will have to be modified accordingly.

Finally, it should be noted that in cases where the array indices vary between 1,...,N
rather than between 0,...,N–1, like in the Mathematica® software package, the above
arithmetics should be adapted accordingly.

4.3 True units for the output array

We now arrive to the more interesting question of how to convert output array indices to
true frequencies. A first hint to the answer has been given following Eqs. (2.15)–(2.16).
The full answer is as follows (see, for example, the brief discussion in [Bracewell86 p.
358] or [Bracewell95 p. 168], or the more detailed explanation in [Briggs95 Sec. 1.2]):

Let Δx be the sampling interval of the given continuous signal-domain function, i.e. the
interval (in terms of seconds, length units, etc.) between two consecutive sampling points.
The Fourier transform of a function sampled with an interval of Δx is periodic with the
period 1 (see Sec. 2.5 or Sec. D.1 in Appendix D), and since this period corresponds
Δx
to N spectral-domain elements in the DFT output we have NΔu = 1 , where Δu is the fre-
Δx
quency step of the discrete spectrum obtained by the DFT [Gonzalez87 p. 96]. There-
fore f1, the first frequency above 0 in the discrete spectrum, is given in terms of cycles per
unit of x, which we denote here by Hz, by (see also Problem 4-6):3

f1 = Δu = 1 (4.4)
NΔx
Note that f1 is equal to the frequency step Δu of the discrete spectrum obtained by the
DFT, so that the next frequencies in the output array correspond, indeed, to 2f1, 3f1, etc.
Similarly, in the N×N 2D case we have [Brigham88 p. 243; Gonzalez87 p. 100]:

Δu = 1 , Δv = 1 (4.5)
NΔx NΔy
These equations show that along each dimension the frequency sampling interval is
simply the reciprocal value of the corresponding sampling range of the signal (see Eq.
(4.2)); it is also inversely proportional to the sampling interval of the given signal. This
can be generalized to the MD case without difficulty.

Remark 4.1: It is important to note that f1 is not the sampling frequency of the input
signal; the latter is given in terms of cycles per unit of x (or in Hz) by:

fs = 1 (4.6)
Δx
[Smith07 pp. 2, 115]. As we can see, it follows that fs = Nf1 = NΔu. Q

1
3 Note that Δu = NΔx is often said to be the spectral resolution of the DFT. However, this term should be
used with care, since as we will see in Chapter 6 the effective resolution of the DFT is often limited by
the leakage effect [Brigham88 p. 180]. On the frequency unit “Hz” and its use here, see in the Glossary.
4. True units along the axes when plotting the DFT 73

Let us now show explicitly, based on Eq. (4.4), the connection between the element
index in the reorganized output array and the real frequency in Hz. As we have seen in
Sec. 3.2, the indices n = 0,...,N–1 of the reorganized output array correspond to the
frequencies –(N/2)f1, ... ,(N/2–1)f1 in steps of f1. Thus, the first element in the reorganized
output array, n = 0, corresponds to the frequency –(N/2)f1 = – 1 ; the element n = N/2
2Δx
corresponds to the frequency 0; the element n = N/2 + 1 corresponds to the frequency f1;
the element n = N/2 + 2 corresponds to the frequency 2f1; and the last element, n = N–1,
corresponds to the frequency (N/2–1)f1 = 1 – f1. The general formula is, therefore:
2Δx
fHz = –(N/2)f1 + nf1

= (n – N/2)f1 (4.7)
and using Eq. (4.4) above:
fHz = (n – N/2) 1 (4.8)
NΔx
This relationship allows us, when we graphically plot the output array obtained by the
DFT (after its reorganization), to replace the default integer units along the axes (i.e. the
array index n = 0,...,N–1) by the correponding true frequencies. See, for example, Figs.
3.1(d) or 4.1(c) for the 1D case, and Fig. 3.6(f) for the 2D case.

It is important to stress here that the lowest negative and highest positive frequencies
provided by the DFT are (see also Problem 4-6):4

fmin = –(N/2)f1 = – 1 fmax = (N/2–1)f1 = 1 – f1 (4.9)

2Δx 2Δx
or in other words, by virtue of Eq. (4.6):

fmin = –12 fs fmax = 12 fs – Δu (4.10)

In the MD case this applies, of course, along each of the M axes separately.

Beyond these frequencies the DFT assumes periodicity of the spectrum (see Sec. 2.5 or
Sec. D.1 in Appendix D). So what happens if the original continuous function contains
frequencies which exceed these limits? We will see the answer in Chapter 5.

Finally, note that the following relationship, that is sometimes found in the literature:

fHz = nf1 = n (4.11)

NΔx

4 In fact, due to the cyclic (periodic) nature of the DFT the element n = 0 (after the reorganization of the
1
output array) corresponds to both of the positive and negative frequencies ±(N/2)f1 = ± 2Δx [Press02 p.
1
510], so that we can say that fmax = –fmin = 2Δx . This is also in line with the well known sampling
theory result which says that when sampling g(x) with a sampling interval of Δx, no frequencies beyond
1
± 2Δx may appear in the resulting spectrum [Castleman79 pp. 235–236; Briggs95 p. 9; Harris98 p. 713;
Smith07 p. 229] (or, in other words, by virtue of Eq. (4.6), that discrete data can only contain
frequencies up to half of the sampling frequency fs [Smith03 pp. 42, 149]). Note that the frequency
1 1
2 f s = 2Δx is often called the Nyquist frequency, but we prefer to avoid using this term because of its
ambiguity. As explained in [Smith03 pp. 41–42 and 638], it turns out that different authors use this
term in four different meanings: the heighest frequency contained in a (band-limited) signal; twice this
frequency; the sampling frequency; or half of the sampling frequency.
74 4. True units along the axes when plotting the DFT

may be misleading due to the uncertainty about the meaning of its index n: Clearly, it is
not the index 0,...,N–1 of the output array — neither before nor after its reorganization. In
fact, this relationship holds for n = –N/2, ... ,N/2–1, but not for n = 0,...,N–1. This also
explains why some authors prefer to use, as we have already noted in Sec. 2.3, a variant of
the DFT definition where the indices vary between –N/2 and N/2–1 (see, for example,
[Briggs95 p.67; Bracewell86 p. 362]).

4.3.1 The particular case of periodic functions

We terminate this section with an example illustrating the particular case in which g(x) is
periodic. This will help us to better understand the relationships between the signal and
spectral domains in the continuous world and their discrete-world counterparts.

Suppose we are given the continuous function g(x) = cos(2π fx), i.e. a cosine with
frequency f. Its period is, therefore, P = 1/f. Because this is a symmetric function, we may
want to sample it within a symmetric range –R/2...R/2, as explained in Sec. 4.2 (see Fig.
4.2(a)). The number of periods of g(x) within this range is, of course:

q = R/P = R f (4.12)

The same number of periods will also occur within the range 0,...,N–1 of the sampled
version of our function, in the input array of the DFT. This means that in the discrete case
the length of each period of the cosine is given, in terms of the array index, by:

p = N/q (4.13)

This is, indeed, the discrete counterpart of the continuous-world relationship P = 1/f.

The Fourier transform of our continuous cosine function consists of two impulses at the
frequencies –f and f [Brigham88 pp. 19–20; Bracewell86 p. 412]. By inverting Eq. (4.8):

n = N/2 + NΔxfHz (4.14)

Figure 4.2: (a) The continuous periodic function g(x) = cos(2πfx) with frequency f = 0.5
(represented by a continuous line), and its discrete counterpart (represented
by black dots) that is obtained by sampling g(x) within the symmetric range
–4...4 with a step of Δx = 0.25, so that N = 32 and q = 4. (b) The CFT of
g(x), G(u) = 12 δ(u–f) + 12 δ(u+f) [Bracewell86 p. 412] (represented by a
continuous line), and its discrete counterpart (represented by black dots) as
obtained by applying DFT to the discrete sequence shown in (a), after both
input and output reorganizations. (c) Same as (b), after having applied to the
DFT output the required vertical scaling correction, too, as explained below in
Sec. 4.4; the vertical discrepancy between the corrected DFT and the CFT is
explained in Sec. A.3 of Appendix A. In each of the plots the horizontal axis
 4. True units along the axes when plotting the DFT 75

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
1
g(x) = cos(2πfx) with f = 0.5
0.5 • • • Sampled
0

-0.5

-1

-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
n n
4.0 4.0

3.0 3.0
CFT CFT
2.0 2.0 • • • DFT after
• • • DFT after
i&o reorg. i&o reorg.
1.0 1.0

0.0 0.0

-1.0 Hz -1.0 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 n 0 4 8 12 16 20 24 28 31 n
4.0 4.0

3.0 3.0
CFT CFT
2.0 • • • DFT after 2.0 • • • DFT after
i&o reorg. i&o reorg.
1.0 and v. sca. 1.0 and vert. sca.

0.0 0.0

-1.0 Hz -1.0 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

at the top corresponds to the integer index of the discrete data sequence that
is represented by the dots, and the horizontal axis at the bottom corresponds
to the true units of the respective continuous function. In terms of the true,
continuous-world units the spectral impulses are located at u = ±f = ±0.5 Hz,
and in terms of the reorganized output array of the DFT they are located,
according to Eq. (4.16), at the indices n = 16 ± q = 16 ± 4.
76 4. True units along the axes when plotting the DFT

we see that the frequencies ±f of these two impulses correspond in the reorganized output
array of the DFT to the array indices:
n = N/2 ± NΔx f (4.15)
or by Eq. (4.2):
= N/2 ± R f
This means by Eq. (4.12) that the array indices of the two impulses are given by:
n = N/2 ± q (4.16)

It follows, therefore, as we can see in Fig. 4.2(c), that the impulses of our cosine
function will be located in the reorganized output array of the DFT q elements away to
each side of the DC (which is itself located, after the reorganization, at the index n = N/2).
If we now replace our function g(x) by a periodic rectangular wave having the same period
P, whose spectrum contains impulses at all the higher harmonics kf, too, these new
impulses will also appear in the discrete output array, spaced by a q-element step, as
shown in Fig. 4.3. (For the sake of comparison, in the case of Fig. 3.5 we had one single
period throughout the entire sampling range, meaning that q = 1; and indeed, all the
harmonic impulses appeared there in the output array in contiguous locations. Further
details on the number of periods being used and its significance are given in Remark 7.1.)

Note that in the present discussion we assumed that q is an integer number, so that each
impulse in the spectrum of g(x) falls exactly on an output array element. We will see later,
in Sec. 6.2, what happens when q is not precisely an integer. A more detailed discussion
on the roles of p and q in the signal and spectral domains and on the relationship between
them is provided in Appendix C.

Figure 4.3: Same as Fig. 4.2, but with a different periodic function. (a) The continuous
square wave g(x) with frequency f = 0.5 Hz (represented by a continuous
line), and its discrete counterpart (represented by black dots) that is obtained
by sampling g(x) within the symmetric range –4...4 with a step of Δx = 0.25,
so that N = 32 and q = 4.5 (b) The CFT of g(x), i.e. the infinite impulse train
G(u) = a sinc(bu) ∑δ(u–nf) [Amidror09 p. 22] (represented by continuous
lines), and its discrete counterpart (represented by black dots) as obtained by
applying DFT to the discrete sequence shown in (a), after both input and
output reorganizations. (c) Same as (b), after having applied to the DFT
output the required vertical scaling correction, too, as explained in Sec. 4.4;
the vertical discrepancy between the corrected DFT and the CFT is explained
in Sec. A.3 of Appendix A. In each of the plots the horizontal axis at the
top corresponds to the integer index of the discrete data sequence that is

5 Note that at this low resolution the sampled version of our square wave is indistinguishable from a
sampled triangular wave. See also Problem 4-10.
 4. True units along the axes when plotting the DFT 77

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k

1
g(x) = wave(x)
0.5 • • • Sampled

-0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
2.0
n 2.0
n
CFT CFT
• • • DFT after • • • DFT after
i&o reorg. i&o reorg.
1.0 1.0

0.0 0.0

Hz Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 n 0 4 8 12 16 20 24 28 31 n
2.0 2.0
CFT
• • • DFT after
i&o reorg.
1.0 1.0 and vert. sca.

0.0 0.0

Hz Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

represented by the dots, and the horizontal axis at the bottom corresponds to
the true units, i.e. the units of the respective continuous function. In terms of
the true, continuous-world units the spectral impulses are located at u = ±kf
= ±0.5k Hz for all integer k, and in terms of the reorganized output array of
the DFT they are located, according to Eq. (4.16), at the indices n = 16 ± 4k.
78 4. True units along the axes when plotting the DFT

4.4 True units for the DFT element values (heights along the vertical axis)

When we use DFT to compute and plot the Fourier transform of a given continuous-
world function, we obviously wish that the values of the DFT output array agree with the
values of the continuous Fourier transform and, in particular, we want them to be correctly
located along the vertical axis in terms of the true units. However, in most cases the DFT
output values do not exactly correspond to the expected values of the continuous Fourier
transform. This discrepancy may have several origins.

First of all, as we have seen in Remark 2.2, in order to obtain the true values (heights) of
the continuous Fourier transform, one needs to multiply the values obtained in the output
array of the DFT by the scaling factor Δ x (the sampling interval used to sample the
original signal-domain function).6 In the 2D case the scaling factor should be (Δx)2, and in
the general MD case the scaling factor is (Δx)M.

However, this is not yet all. As we have seen in Chapter 2, both the continuous Fourier
transform and the DFT can be defined in several slightly different ways, that may vary, in
particular, in the value of the constant factor being used. For this reason, we must know
which definitions are being used for the continuous Fourier transform and for our DFT
algorithm, in order to take into account the different constant factors being used.7

Let us consider a few examples. Suppose that the continuous Fourier transform is
defined by Eq. (2.1); in this case, the Fourier transform of the sinc function given in Eq.
(4.1) is the unit square pulse function [Bracewell86 pp. 100–101], whose height is 1 (see
Fig. 4.1). If the DFT program being used is based on the definition given by Eq. (2.13) or
(2.15), the resulting plot obtained by drawing the (reorganized) output array will show a
square pulse function whose height is 1/Δx. Therefore, as explained above, in order to
correctly represent the original continuous Fourier transform, the DFT result must be
multiplied by the sampling interval Δ x. But if our DFT algorithm is based on the
symmetric definition having a constant factor 1 in both direct and inverse transforms (as
N
is the case in the Mathematica® software package [Wolfram96 p. 868]), the DFT will give
us a square pulse function whose height is Δx1 N . In this case, in order to adapt the output
of our DFT algorithm to the continuous Fourier transform definition being used, we have,
therefore, to scale the values of the DFT output array accordingly, i.e. to multiply them by
Δx N . This is, indeed, the case in our figures; for example, in Fig. 4.1 the scaling factor we
applied to the DFT output in (b) to obtain the match in (c) is Δx N = 0.25√32 ≈ 1.414. If,
however, we use a DFT implementation with a factor of 1 in the direct transform, we will
N
6 One may provide two explanations for this fact: (1) As a result of signal-domain sampling, the
frequency domain has been scaled by factor 1/Δ x [Brigham88 pp. 77–79]; therefore, to recover the
original spectrum, a scaling by factor Δx is required (see also Sec. 2.5 above). (2) The rectangular rule
for the numeric integration of the continuous Fourier transform provides the same sum as the DFT,
multiplied by the factor Δx [Brigham88 pp. 100–101]; see also Remark 2.2 in Chapter 2.
7 Note also that using a DFT definition with reversed signs in the exponential part will result in a
reversed list of values in the output [Wolfram96 p. 868]; more precisely, the positive and negative
frequencies will be interchanged (since a sign reversal of the exponent can be assimilated with a sign
reversal in the index n within the exponent, where due to the cyclic nature of the DFT we have
–n = N – n [Smith07 pp. 118–119]). See Sec. 8.3 for more details.
4. True units along the axes when plotting the DFT 79

rather have to multiply the DFT output values by the factor ΔxN = R. We will henceforth
call the required scaling correction, which takes into account Δx as well as the other scaling
factors in question (depending on the continuous and discrete definitions being used), the
standard amplitude scaling correction.

It is important to note that a similar amplitude scaling correction is also required when
using the inverse DFT (IDFT). But because in this case the input array consists of the
frequency domain and the output array consists of the signal domain, Δx must be replaced
here by Δu (the sample interval in the frequency domain). Note that applying the corrected
IDFT to the corrected DFT output must give back the original data. Similarly, applying an
uncorrected IDFT to the uncorrected DFT output gives back the original input array, too.

As we can see, the standard amplitude scaling correction is a necessary condition for
obtaining the correctly scaled Fourier transform values from the brute results of the DFT,
and for correctly placing the true units along the vertical axis. However, even this is not yet
all. It turns out that there also exist inherent DFT-related reasons that may cause, in some
cases, significant discrepancies between the results obtained by DFT and the true Fourier
transform, as clearly seen, for example, in Figs. 4.1(c) and 3.5(f). These artifacts of the
DFT (known as aliasing and leakage) will be discussed in the following chapters.

Finally, there exists one more case that deserves particular attention, because it defies the
standard amplitude scaling corrections described above even when no DFT artifacts
(aliasing or leakage) are present. This case involves all functions whose frequency spectra
contain impulses, such as the constant function g(x) = 1 whose spectrum is G(u) = δ(u),
i.e. a unit impulse at the origin; the cosine function g(x) = cos(2πfx) whose spectrum is
G(u) = 12 δ(u–f) + 12 δ(u+f), i.e. a pair of half-unit impulses8 at the frequencies ±f; or any
other periodic function (or even the sum of an aperiodic function with a periodic function).
As shown in Figs. 4.2, 4.3 and Fig. A.2 of Appendix A, it turns out that in such impulsive
spectra the strength of the impulses obtained by the DFT is not properly rectified by the
standard amplitude scaling correction. This should not really be surprising, since the
notion of the amplitude (or height) of an impulse is quite elusive already in the continuous
case, so the equivalence with its discrete counterpart is not really obvious. A detailed
discussion on this point is provided in Appendix A.

4.5 Examples

Let us illustrate the above discussion by some further examples.

Example 4.1: The same discrete signal representing different continuous functions:
Let us consider Fig. 4.4(a), and compare it with Fig. 4.1(a). The 32-element discrete
signals shown in these two figures are strictly identical, as we can clearly see by

8 The reason we use the term “half-unit impulses” rather than “impulses of amplitude 12 ” is explained in
Appendix A.
80 4. True units along the axes when plotting the DFT

comparing them element-by-element between k = 0 and k = 31 (see the horizontal axis

representing the index k at the top of the figures). However, while in Fig. 4.1(a) this
discrete signal represents the continuous function g(x) = sinc(x) that has been sampled
within the range –4...4 with a step of Δx = 0.25, in Fig. 4.4(a) the same discrete signal
corresponds to the continuous function h(x) = sinc(2x) that has been sampled within the
range –2...2 with a step of Δx = 0.125. Because both discrete signals are identical, it is not
surprising that their DFTs are identical, too (see the black dots in Figs. 4.1(b) and 4.4(b)).
However, the CFTs of the continuous functions g(x) = sinc(x) and h(x) = sinc(2x) are
obviously different, as we can see, indeed, by comparing the continuous lines in the same
two figures: G(u) = rect(u) while H(u) = 12 rect(u/2). The difference between the two cases
is in the scaling of the frequency axis, as well as in a vertical scaling factor. Both of these
are automatically taken care of when we plot the DFT results according to the explanations
given in Secs. 4.3 and 4.4 (i.e. when drawing the true frequency axis according to Eq.
(4.8), and applying the standard amplitude scaling correction). And indeed, when we
follow these steps we obtain in both cases a correct discrete graphic representation of the
underlying continuous Fourier transform, as shown in Figs. 4.1(c) and 4.4(c). The
different interpretations of the same DFT results in these figures (the different true
frequency axis and standard amplitude scaling correction) are simply a consequence of the
different sampling range length R and sampling step Δx in Figs. 4.1(a) and 4.4(a).
As we clearly see in this example, when we change the continuous-world interpretation
of the same DFT input signal, the DFT output signal remains, of course, unchanged, but
its continuous-world interpretation does change accordingly. Q

Example 4.2: The same underlying continuous function as in Example 4.1, using
different sampling range and sampling step:
Let us consider again the same continuous function h(x) = sinc(2x) as in Fig. 4.4(a), but
this time we take its 32 discrete samples using the same sampling range –4...4 and the
same sampling step of Δx = 0.25 as in Fig. 4.1(a). This is shown in Fig. 4.5. In this case,
the range of the true frequency axis and the standard amplitude scaling correction that we
obtain according to Secs. 4.3 and 4.4 remain, of course, the same as in Fig. 4.1; but the
discrete input signals that are fed to the DFT, and hence the DFT results, are not the same
in the two figures (compare Fig. 4.1 and Fig. 4.5). The final graphic result, however, when
we plot the DFT according to the explanations given in Secs. 4.3 and 4.4, is correct in
both cases, as we can see in Figs. 4.1(c) and 4.5(c). Q

Figure 4.4: (a) The same discrete signal as in Fig. 4.1(a), but this time it represents
the continuous function h(x) = sinc(2x) that has been sampled within the
range –2...2 with a step of Δx = 0.125, so that again N = 32. (b) The CFT
of h(x), H(u) = 12 rect(u/2) (represented by a continuous line), and its
discrete counterpart (represented by black dots) as obtained by applying
DFT to the discrete sequence shown in (a), after both input and output
 4. True units along the axes when plotting the DFT 81

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
h(x) = sinc(2x)
1
• • • Sampled
0.5

-0.5

-1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -4. -3. -2. -1. 0 1. 2. 3. 4.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 and v. sca. 0.5 and vert. sca.

0 0

-0.5 -0.5

-1 Hz -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -4. -3. -2. -1. 0 1. 2. 3. 4.
(c)

reorganizations. Note that the DFT output itself is exactly the same as in
Fig. 4.1(b), and only the interpretation of the true continuous-world units
along the frequency axis is different, as explained in Sec. 4.3. (c) Same as
(b), after having applied to the DFT output the required vertical scaling
correction, too, as explained in Sec. 4.4; note the close match between the
corrected DFT and the CFT.
82 4. True units along the axes when plotting the DFT

Example 4.3: The same discrete signal representing different continuous functions — in
the periodic case:
Consider Fig. 4.6(a). The 32-element discrete signal shown in this figure is strictly
identical to that of Fig. 4.2(a), as we can clearly see by comparing them element-by-
element between k = 0 and k = 31 (see the horizontal axis representing the index k at the
top of the figures). However, while in Fig. 4.2(a) this discrete signal represents the
continuous function g(x) = cos(2πfx) with f = 12 Hz that has been sampled within the range
–4...4 with a step of Δx = 0.25, in Fig. 4.6(a) the same discrete signal corresponds to the
continuous function h(x) = cos(2πfx) with f = 1 Hz that has been sampled within the range
–2...2 with a step of Δx = 0.125. Because both discrete signals are identical, it is not
surprising that their DFTs are identical, too (see the black dots in Figs. 4.2(b) and 4.6(b)).
However, the continuous Fourier transforms of the continuous functions g(x) and h(x) are
obviously different, as we can see, indeed, by comparing the continuous lines in the same
two figures: G(u) = 12 δ(u–12 ) + 12 δ(u+ 12 ) while H(u) = 12 δ(u–1) + 12 δ(u+1).
And indeed, as expected, when we plot the DFT results according to the explanations
given in Secs. 4.3 and 4.4, the true frequency units are correctly affixed to the DFT results
in both of the figures. But in both cases the standard amplitude scaling correction is not
yet sufficient for the discrete impulse heights to match their continuous-world
counterparts: As we can see in Figs. 4.2(c) and 4.6(c), in both cases the DFT results still
must be multiplied by the factor 1/R, where R = N Δ x is the sampling range. This is
explained in detail in Sec. A.3 of Appendix A. Q

Example 4.4: A 2D example:

Consider now Fig. 4.7(a) and compare it with Fig. 3.6(a). The 32×32-element discrete
signals shown in these two figures are strictly identical, as we can clearly see by
comparing them element-by-element within the 2D range k = 0,...,31, l = 0,...,31 (see the
axes representing the discrete indices k and l at the top and at the right hand side of the
figures). However, while in Fig. 3.6(a) this 2D discrete signal represents the continuous
function g(x,y) = rect(x) rect(y) that has been sampled within the range –4...4 along both
axes with a step of Δx = Δy = 0.25, in Fig. 4.7(a) the same discrete signal corresponds to
the continuous function h(x,y) = rect(2x) rect(2y) that has been sampled along both axes
within the range –2...2 with a step of Δx = Δy = 0.125. Because both discrete signals are
identical, it is not surprising that their DFTs are identical, too (see Figs. 3.6(e) and 4.7(b)).

Figure 4.5: (a) The same continuous signal h(x) = sinc(2x) as in Fig. 4.4(a), but this
time it has been sampled within the range –4...4 with a step of
Δx = 0.25, so that again N = 32. (b) The CFT of h(x), H(u) = 12 rect(u/2)
(represented by a continuous line), and its discrete counterpart
(represented by black dots) as obtained by applying DFT to the
 4. True units along the axes when plotting the DFT 83

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
h(x) = sinc(2x)
1
• • • Sampled
0.5

-0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 n 1.5 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 0.5

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

1.5
0 4 8 12 16 20 24 28 31 n 1.5
0 4 8 12 16 20 24 28 31 n
CFT CFT
1 • • • DFT after 1 • • • DFT after
i&o reorg. i&o reorg.
0.5 and v. sca. 0.5 and vert. sca.

0 0

-0.5 -0.5

-1 Hz -1 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

discrete sequence shown in (a), after both input and output

reorganizations. (c) Same as (b), after having applied to the DFT
output the required vertical scaling correction, too, as explained in Sec.
4.4; note the close match between the corrected DFT and the CFT. See
Example 4.2.
84 4. True units along the axes when plotting the DFT

However, the continuous Fourier transforms of the continuous functions g(x,y) and h(x,y)
are obviously different: G(u,v) = sinc(u) sinc(v) while H(u,v) = 12 sinc(u/2) 12 sinc(v/2). The
difference between the two cases is in the scaling of the frequency axes, as well as in a
vertical scaling factor. Both of these are automatically taken care of when we plot the DFT
results according to the explanations given in Secs. 4.3 and 4.4 (i.e. when drawing the true
frequency axes according to Eq. (4.8), and applying the standard amplitude scaling
correction). And indeed, when we follow these steps we obtain in both cases a correct
discrete graphic representation of the underlying CFT, as shown in Figs. 3.6(f) and 4.7(c).
The different interpretations of the same DFT results in these figures (the different true
frequency axes and standard amplitude scaling correction) are simply a consequence of
the different sampling range and sampling steps in Figs. 3.6(a) and 4.7(a). Q

PROBLEMS

4-1. Our student from Problem 3-1 has finally obtained, following the explanations of
Chapter 3, a correctly reorganized plot of the DFT spectrum (whose real-valued part is
shown in Fig. 4.8). However, he soon realizes that the values marked along the two
planar axes of the DFT spectrum are not true frequencies in terms of Hz, but rather
integer numbers specifying the element’s position (index) within the DFT output
array. Furthermore, the values along the vertical axis are far away from the expected
CFT values. What can he do in order to solve each of these problems?
4-2. Suppose our student has finally obtained the true values along all the axes (see Fig.
4.9). Now, he would like to double the frequency range obtained by the DFT along
each of the u and v axes. What would you suggest him to do?
4-3. If our student desired to double the resolution of the DFT output along the u and v
axes instead of doubling the frequency range, what would you suggest him to do?
4-4. If our student wanted to double both the frequency range and the resolution along
each of the u and v axes, what would you suggest him to do?
4-5. The function g(x) = cos(2π fx) is periodic with period P = 1/ f; it therefore has f periods
per unit (for example, within the interval 0...1). Plot a sampled version of this function
within an input array of length N = 128, N = 64 or N = 32, with true function values
along both axes. Compute the DFT of these input arrays, and plot the resulting output
arrays with true units along both axes (frequency in Hz along the horizontal axis, and
true spectral values along the vertical axis, as explained in Secs. 4.3 and 4.4).
(a) What is the useful range of the cosine parameter f (the cosine frequency) for which

Figure 4.6: (a) The same discrete periodic signal as in Fig. 4.2(a), but this time it
represents the continuous periodic function h(x) = cos(2πfx) with frequency
f = 1 that has been sampled within the range –2...2 with a step of Δ x =
0.125, so that again N = 32. (b) The CFT of h(x), H(u) = 12 δ(u–1) + 12 δ(u+1)
(represented by a continuous line), and its discrete counterpart (represented
 4. True units along the axes when plotting the DFT 85

Signal domain
0 4 8 12 16 20 24 28 31
1.5 k
1
h(x) = cos(2πfx) with f = 1
0.5 • • • Sampled
0

-0.5

-1

-1.5
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
x
(a)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
n n
4.0 4.0

3.0 3.0
CFT CFT
2.0 2.0 • • • DFT after
• • • DFT after
i&o reorg. i&o reorg.
1.0 1.0

0.0 0.0

-1.0 Hz -1.0 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -4. -3. -2. -1. 0 1. 2. 3. 4.
(b)

Spectral domain: Real part Spectral domain: Imaginary part

0 4 8 12 16 20 24 28 31 n 0 4 8 12 16 20 24 28 31 n
4.0 4.0

3.0 CFT 3.0 CFT

• • • DFT after • • • DFT after
2.0 i&o reorg. 2.0 i&o reorg.
and v. sca. and vert. sca.
1.0 1.0

0.0 0.0

-1.0 Hz -1.0 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -4. -3. -2. -1. 0 1. 2. 3. 4.
(c)

by black dots) as obtained by applying DFT to the discrete sequence shown

in (a), after both input and output reorganizations. (c) Same as (b), after
having applied to the DFT output the required vertical scaling correction,
too, as explained in Sec. 4.4; the vertical discrepancy between the corrected
DFT and the CFT is explained in Sec. A.3 of Appendix A.
86 4. True units along the axes when plotting the DFT

your plots make sense and correctly represent the continuous-world function g(x) and
its spectrum G(u), when using N = 128, N = 64 or N = 32?
(b) What should you do if you need the function g(x) to have a higher or a lower
frequency f than the limit values you suggested in (a)?
(c) Do the impulse heights obtained in your DFT spectra according to the explanations
of Sec. 4.4 correctly represent the impulse heights in G(u)? Hint: See Appendix A.
4-6. The longest full sinusoidal cycle that can be captured by the N given data points
consists of N samples per cycle. The smallest frequency above zero in the DFT
spectrum, f1 , is the reciprocal of this value, namely: f1 = 1/N cycles per sample (i.e.
cycles per Δx seconds), meaning that f1 = NΔx1
Hz (Eq. (4.4)). Can you find in a similar
way the largest frequency in the DFT spectrum? Hint: The smallest full sinusoidal cycle
that can be captured by the given data points consists of 2 samples per cycle.
4-7. The explanations provided in Secs. 4.2 and 4.3 are given under the assumption that the
input and output array elements are indexed by integers running between 0,...,N–1.
However, in many software packages the array indices must run between 1,...,N. Can
you adapt the results of Secs. 4.2 and 4.3 to this case?
4-8. We have seen in Sec. 4.4 that in functions whose frequency spectra contain impulses,
such as all periodic functions, the height of the impulses obtained by the DFT is not
properly rectified by the standard amplitude scaling correction (even if no aliasing and
leakage artifacts are present). This is clearly shown in Figs. 4.2(c) and 4.6(c) for the
function g(x) = cos(2π fx). This issue is discussed in detail in Appendix A. Redo Figs.
4.2(c) and 4.6(c) after having read Appendix A, and verify that when applying the
adjustments as explained there, the DFT spectrum matches the expected CFT values.
4-9. In all the examples provided in this chapter, the underlying original continuous-world
functions and their respective continuous-world spectra were always known, so that we
could easily compare the DFT results with the true continuous-world values and see if
they match correctly. However, in many real-world situations the original functions and
their spectra are not known, and all that we have at hand is a given discrete signal. For
example, suppose your input data was obtained by astronomical measurements. Do
you expect to have any difficulties in applying the results of the present chapter to the
plots of your signal and of its DFT spectrum? Will you be able to detect discrepancies
due to DFT artifacts such as aliasing and leakage, like in Figs. 4.1(c) and 3.5(f)?

Figure 4.7: (a) The same 2D discrete signal as in Fig. 3.6(a), but this time it
represents the 2D continuous function h(x,y) = rect(2x) rect(2y) that
has been sampled within the range –2...2 along both axes with a step
of Δ x = Δ y = 0.125, so that again N = 32. (b) The DFT of the
discrete signal shown in (a), after both input and output
reorganizations. Note that the DFT output itself is exactly the same
as in Fig. 3.6(e), and only the interpretation of the true continuous-
world units along both frequency axes is different, as explained in
Sec. 4.3. (c) Same as (b), after having applied to the DFT output the
required vertical correction, too, as explained in Sec. 4.4. Although
the CFTs G(u,v) = sinc(u) sinc(v) and H(u,v) = 12 sinc(u/2) 12 sinc(v/2)
are not shown in Figs. 3.6 and 4.7, one can verify that the 2D DFT
plot in each of these figures closely matches the respective CFT.
4. True units along the axes when plotting the DFT 87

Signal domain
0 4 8 12 16 20 24 28 31
2.
31 Legend:
1.5 28
1
1. 24

0.5 20

0 16 0
-0.5 12

-1. 8

-1.5 4
-1

-2. 0
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.

(a)
Spectral domain: Real part Spectral domain: Imaginary part
0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
4. 4.
31 31
28 28

2. 24 2. 24

20 20

0 16 0 16

12 12

-2. 8 -2. 8

4 4

-4. 0 -4. 0
-4. -2. 0 2. 4. -4. -2. 0 2. 4.

(b)
Spectral domain: Real part Spectral domain: Imaginary part
0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
4. 4.
31 31
28 28

2. 24 2. 24

20 20

0 16 0 16

12 12

-2. 8 -2. 8

4 4

-4. 0 -4. 0
-4. -2. 0 2. 4. -4. -2. 0 2. 4.

(c)
88 4. True units along the axes when plotting the DFT

4-10. Fig. 4.3 shows a square wave g(x) with period P = 2 (drawn by a continuous line) and
its discrete counterpart (represented by black dots) that is obtained by sampling g(x)
within the symmetric range –4...4 with a step of Δ x = 0.25, so that N = 32 and q = 4.
However, as mentioned in Footnote 5 there, at this low resolution the sampled version
of our square wave is indistinguishable from a sampled triangular wave. Knowing that
the CFT of a square wave and the CFT of a triangular wave are not identical, how can
you interpret the resulting DFT spectrum? The same question can be also asked for the
non-periodic counterpart of g(x) which only contains the central pulse around x = 0,
and whose spectrum is continuous (non-impulsive). Hint: The problems related to
insufficient sampling resolution are treated in Chapter 5; see also Problem 5-6.

2
1
60
0
-1
-2 40
0

20 20
40

60 0

Figure 4.8: The DFT plot obtained by the student in Problem 4-1.

0.6 v
0.4 2.
0.2
1.
0
-2. 0
-1.
0 -1.

1.
-2.
2. u
Figure 4.9: The DFT plot obtained by the student in Problem 4-2.
 Chapter 5
Issues related to aliasing

5.1 Introduction
When the DFT is used as a numerical approximation to the Fourier transform of an
underlying continuous function, some differences can be expected between the two. For
example, it is clear that the DFT can only give us the Fourier transform values at the N
discrete points for which the computation is done. However, even at these points there may
appear in some cases quite significant discrepancies, in spite of the adjustments and
scalings discussed above in Chapters 3 and 4. This is not a result of rounding errors in the
calculations, which are normally quite insignificant (see Sec. 8.8). The main reason for
these discrepancies resides in the discrete and finite nature of the DFT, which implies that
the underlying continuous function must be sampled, and that it can only be considered
within a finite range. Both of these two operations, sampling and finite-range windowing
(or truncation), are inherent to DFT, as it can be seen by comparing Eqs. (2.1)–(2.2) with
Eqs. (2.13)–(2.14). Each of these operations involves some loss of data, and introduces its
own discrepancy between the values obtained by DFT and the true values of the CFT.
These two types of error, both considered as DFT artifacts, are named respectively
aliasing and leakage. In the present chapter we will focus on issues related to aliasing;
leakage-related considerations will be discussed in Chapter 6.

5.2 Aliasing in the one dimensional case

Aliasing is a phenomenon which may occur due to the discretization (sampling) of the
underlying input function. In fact, aliasing is not only limited to DFT applications, and it
may be encountered wherever sampling is used, whether or not the sampled data is then
subjected to DFT. Discretizing an analog signal requires that the signal’s value be
sampled often enough to define the waveform unambiguously. According to the sampling
theorem, the sampling frequency must be at least twice the highest frequency present in
the signal for its waveform to be defined completely.1 In other words, there must be at
least two samples per cycle for any frequency component (sine or cosine) of the signal. If
the sampling points are not taken as densely as required (a situation often called
undersampling), they will fail to follow the high-frequency fine details of the signal, thus
leading to aliasing. In the signal domain this is expressed by the existence of a smoother,
lower-frequency signal, known as alias, which can be traced through all the sampled

1 There exist in the literature many different variants of the sampling theorem, which are basically all
equivalent. This version of the theorem appears, for example, in [Ramirez85 p. 115], [Gaskill78 p. 269]
or [Weisstein99 p. 1593].

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 89
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_5,
© Springer-Verlag London 2013
90 5. Issues related to aliasing

points and “mimic” or “masquerade” the behaviour of the original signal on its sampled
values (see Fig. 5.1 for a periodic example and Fig. 5.2 for an aperiodic example). Only
based on these too far-apart sampling points, the original signal is undistinguishable from

x
-4 -2 2 4

-1

(a)
-2

x
-4 -2 2 4

-1

(b)
-2

Figure 5.1: Illustration of the signal-domain interpretation of aliasing: a periodic

case. (a) Sampling the continuous periodic signal g(x) = cos(2πfx)
having frequency f = 2 at a rate higher than twice the maximum signal
frequency gives a correct discrete representation of the original
signal. (b) Sampling the same continuous periodic signal at a rate
lower than twice the maximum signal frequency gives a false, aliased
lower-frequency signal (here: a cosine function with frequency fA = 14 ;
see Example 5.4 in Sec. 5.3) that mimics the original signal on its
sampled values. Both signals are drawn with continuous lines, and
their sampled values are indicated by black dots.
5. Issues related to aliasing 91

its lower-frequency aliased signal, since both of them coincide on all of these sampling
points. In other words, aliasing means that high-frequency components in the original
signal appear incorrectly as lower frequencies in the sampled result.

x
-4 -2 2 4

-1

(a)
-2

x
-4 -2 2 4

-1

(b)
-2

Figure 5.2: Same as Fig. 5.1, showing this time an aperiodic case. (a) Sampling a
continuous aperiodic signal (here: the function g(x) = sinc(x)) at a
rate higher than twice the maximum signal frequency gives a correct
discrete representation of the original signal. (b) Sampling the same
continuous aperiodic signal at a rate lower than twice the maximum
signal frequency gives a false, aliased lower-frequency signal (here:
the function gA(x) = 32 sinc(34 x) – 12 sinc(12 x); see Example 5.1 in Sec.
5.3) that mimics the original signal on its sampled values. Both
signals are drawn with continuous lines, and their sampled values are
indicated by black dots.
92 5. Issues related to aliasing

This is the interpretation of aliasing in terms of the signal domain; but the aliasing
phenomenon can be also interpreted from the viewpoint of the spectral domain. Remember
that sampling in the signal domain is equivalent to multiplying the original continuous
signal g(x) by Δ1x III(x/Δx), a periodic unit-height impulse comb (impulse train) having an
impulse interval of Δx (see Fig. 5.3).2 The Fourier transform of this sampling impulse
comb is itself an impulse comb with impulse interval of fs = Δ1x (see Remark 4.1 in Sec.
4.3) and impulse height of Δ1x [Castleman79 pp. 227–228; Bracewell86 p. 414]:

F [ Δ1x III(x/Δx)] = III(Δx u) (5.1)

Therefore, by virtue of the convolution theorem, the effect on the spectral domain of
sampling g(x) is that G(u), the original spectrum of g(x), is now convolved with impulse
comb (5.1). This means, as explained in Sec. 2.5 and shown again in Fig. 5.3(c), that the
spectrum of the sampled version of g(x) consists of infinitely many identical replicas of
the spectrum G(u), which are scaled by Δ1x (in terms of amplitude) and centered about all
the integer multiples of the sampling frequency fs = Δ1x (see, for example, [Brigham88 pp.
79–81] or [Castleman79 pp. 227–229]):3

F [ Δ1x III(x/Δx) g(x)] = F [ Δ1x III(x/Δx)] * F [g(x)] = III(Δx u) * G(u) (5.2)

where * is the convolution operator.

Note that at this stage our discussion on aliasing is still completely general, and we do
not yet consider the particular case of aliasing that occurs in the context of DFT. Thus, if
the spectrum G(u) of the original function g(x) is continuous, like in Fig. 5.3(a), so is the
periodic spectrum (5.2) of the sampled version of g(x), as shown in Fig. 5.3(c).

Let us now denote by fG the highest frequency in G(u). As long as the sampling
frequency fs = Δ1x is at least twice fG, as required by the sampling theorem, the replicas of
G(u) are sufficiently far from each other (Fig. 5.3(c)). But if the sampling frequency fs is
lower than twice fG,4 every two neighbouring replicas of G(u) will somewhat overlap, as
shown in Fig. 5.3(d); note that this overlapping is additive, due to the properties of
convolution. This means in particular that frequencies from neighbouring replicas will
penetrate into the main replica (the replica which is centered about the origin) and appear

2 Thanks to the impulse property δ(x/a) = |a|δ(x) [Bracewell86 p. 76] the impulse height of III(x/Δx) is
1
Δx [Bracewell86 p. 414; Bracewell95 pp. 118–119]; its unit-height counterpart is Δx III(x/Δx). Similarly,
1
the impulse height of III(Δx u) is Δx. See Footnote 3 in Sec. A.2 on the use of the term impulse height.
3 This means that the Fourier transform of the sampled version of a continuous signal g(x) is a periodic
function. If no aliasing occurs (see below), each period of this spectrum is equal (within a constant
factor Δx) to G(u), the Fourier transform of g(x); see Fig. 5.3(c). What is the spectral meaning of such
non-decaying high frequencies that extend to both directions of the spectrum ad infinitum? The fact that
the spectrum is periodic simply means that the corresponding signal-domain function (our sampled
function) is discrete. In fact, our sampled function in the signal domain can be seen as the impulsive
inverse Fourier transform of its periodic spectrum. Remember that the Fourier transform P(u) of a
periodic function p(x) is an impulsive discrete function (see, for example, [Papoulis68 p. 107],
[Brigham88 pp. 77–79] or [Amidror09, Sec. A.2]); the situation in our case is similar, except that the
roles of the signal domain and the frequency domain are interchanged (see Chapter 5 in [Brigham88]).
4 This happens, for example, if g(x) is not band limited, i.e. if f = ∞. Note that the vertical dotted lines
G
going down throughout Fig. 5.3 help to follow the frequencies fG, 2fG etc. in all rows of the figure.
5. Issues related to aliasing 93

Signal domain Spectral domain

g(x) G(u)
c d

x u
0 – 2 fG –fG 0 fG 2 fG
1 x (a)
Δx III(Δx ) III(Δx u)
1 1
Δx fs = Δ1x

x u
0 –fs 0 fs > 2 fG
Δx
(b)
1
c Δx d
fs = Δ1x

x –fs fs
u
0 –fs 0 fs > 2 fG
Δx 2 2
(c)
1
c Δx' d fs' = Δ1x
'

x –fs' fs'
u
0 Δx' –fs' 0 fs' < 2 fG
2 2
(d)

Figure 5.3: Schematic illustration of the Fourier-domain interpretation of aliasing. (a) An

original continuous signal g(x) and its continuous-world spectrum G(u).
(b) The sampling impulse train and its continuous-world spectrum. (c) The
sampled signal is the product of signals (a) and (b), and therefore its
continuous-world spectrum is the convolution of their spectra. Note the
replicas of the original spectrum, that are located at all integer multiples of the
sampling frequency fs. (d) If the sampling rate is lower than twice the
maximum signal frequency fG (say, f 's), then every two neighbouring replicas
of G(u) overlap (additively), giving false aliased frequencies in the overlapped
zones. The aliased parts of the spectrum in (d) are shown by thick dashed
lines, which are simply the sum of the original overlapped replicas. The
horizontal arrows in the spectra (c) and (d) indicate the part of the infinite
continuous-world spectrum that will be represented in the DFT output should
we subject the sampled signal to DFT rather than to CFT: As shown in Sec.
4.3, the frequency range of the DFT extends between minus half and plus
half of the corresponding sampling frequency. Note that in terms of the DFT
frequency range, aliasing can be seen as a cyclical foldover where frequencies
which exceed beyond one end of the range re-enter at the opposite end. The
frequencies fG, 2fG etc. can be followed through all rows thanks to the vertical
dotted lines (note that these lines are not the same as in Figs. 2.1–2.3).
94 5. Issues related to aliasing

in the spectrum as parasite, false lower frequencies, known as aliased frequencies. Note
that in this case, too, all of the replicas are identical; however, none of them (including the
main replica) remains identical to the original spectrum G(u) itself, since they have all been
corrupted (additively overlapped) by frequencies (positive or negative) belonging to their
neighbouring replicas (see Fig. 5.3(d)).

This spectral explanation of aliasing still holds if our sampled signal undergoes DFT
rather than a continuous Fourier transform — the main differences being (1) that the
resulting DFT spectrum is obviously discrete, and (2) that it only contains frequencies in
the range between – 12 fs and 12 fs, where the main replica is located; see Eq. (4.10), and the
frequency range indicated by the horizontal arrows in Figs. 5.3(c),(d).5 A further,
consequential difference is that in the DFT spectrum aliasing looks as if it were a cyclical
foldover, meaning that the frequencies of G(u) that fall beyond 12 fs or – 12 fs re-enter into the
opposite end of the DFT spectrum. In fact, as we can see in the infinite, full-range
spectrum of the continuous-world case (Fig. 5.3(d)), the frequencies which penetrate into
the main replica really belong to the neighbouring replicas; but if we only consider the
DFT spectrum (or the restriction of the continuous spectrum to the corresponding range
– 12 fs...12 fs which contains the main replica) it can be said, indeed, that the frequencies which
exceed from one end re-enter cyclically into the other end. These two viewpoints are
completely equivalent thanks to the periodicity of the spectrum of the sampled signal, i.e.
thanks to the convolution in Eq. (5.2): Because all the replicas in the continuous-world
spectrum are identical, we do not really care whether the intruding frequencies come from
the exceeding frequencies at the other end of the same replica or from the exceeding
frequencies of a neighbouring replica.6 In any case, whichever viewpoint we adopt, the
intruding frequencies due to aliasing penetrate into the frequency range of the DFT
spectrum and additively overlap with the true frequencies that are really located there. This
cyclical foldover (additive wraparound) may corrupt the values in the DFT spectrum with
respect to the original spectrum G(u), starting from the far borders of the DFT spectrum
(i.e. ±12 fs) inward; see, for example, Fig. 5.6 in Example 5.1 or Fig. 5.7 in Example 5.2.
This is, indeed, the spectral-domain interpretation of the aliasing artifacts in the DFT.

Note that if G(u) (and hence the resulting discrete DFT spectrum) is rather smooth and
non-impulsive,7 as in Fig. 5.3(d), this foldover usually modifies the shape of the DFT
spectrum only slightly, and its existence may be quite difficult to detect in practice unless

5 Remember that both the input and output sequences of the DFT should be understood as a single
N-element period (or “truncation window”) from a periodic sequence which repeatedly extends to both
directions ad infinitum (even if the original underlying data was not periodic).
6 Note that in some references foldover due to aliasing is described (and even graphically depicted) as
mirroring of the frequencies that exceed beyond 12 fs (or – 12 fs) backward from the same end of the DFT
spectrum, i.e. as folding around the spectrum edges backward. While this interpretation makes no
substantial difference in the case of real-valued 1D signals, in other 1D cases or in 2D or MD cases it
should be avoided (see, for example, the 2D DFT in Fig. 5.13(a)). The correct interpretation of the
foldover is that whatever exceeds from one side of the DFT spectrum re-enters in a cyclical way from the
opposite side. See also Problem 5-7.
7 The DFT spectrum is said to be “smooth and non-impulsive”, in spite of its obvious discrete nature, if
the values of its successive discrete elements vary in a continuous, non-impulsive way.
5. Issues related to aliasing 95

the true values of the original spectrum G(u) are known in advance. However, when G(u)
is a purely impulsive spectrum (i.e. when the original function g(x) is periodic or almost
periodic [Amidror09, Sec. B.6]) aliasing may be much easier to detect, since the cyclical
foldover introduces false “stray impulses” in the DFT results, where no impulses are
present in the correct Fourier transform. This is illustrated, for the 1D case, in Fig. 5.11
(see Example 5.5 in Sec. 5.3 below).

Remark 5.1: There exists a particular case in which no such “stray impulses” appear in
the DFT spectrum of a periodic function g(x), even if strong aliasing occurs. This happens
if the distance q between the origin and the first harmonic impulse of g(x) in the DFT
spectrum (in terms of output-array elements; see Sec. 4.3.1) satisfies N = pq for an integer
p (note that q itself does not necessarily need to be integer; see Fig. C.2 in Appendix C).
In this case higher harmonic impulses which exceed from one side of the DFT spectrum
will fall exactly on top of true impulse locations (harmonic frequencies) in the other side
of the spectrum, and their heights will be simply summed up. More details on this subject
can be found in Appendix C. Q

Remark 5.2: It is important to note that aliasing can be avoided completely only if the
given signal g(x) is band limited; under this condition, aliasing will be absent if the
sampling frequency is at least twice the highest frequency present in g(x). But in practice,
such situations rarely exist. If g(x) is not band limited (this happens, for example, for any
finite-duration function g(x) [Brigham88 p. 103]), the only solution is to reduce aliasing
as much as possible, as explained below. Q

Aliasing can be reduced in two different ways. One way consists of sampling the given
signal at a high enough rate (i.e. using a sufficiently small sampling interval Δx), so that
aliasing becomes negligible. The other possibility is to low-pass filter the signal prior to
sampling in order to make it band limited (or almost band limited); this allows us to avoid
(or at least reduce) aliasing by using a sufficiently high sampling rate [Brigham88 p. 86;
Cartwright90 p. 205]. Any combination of these two methods can also be used.

More details on the optimal choice of the sampling interval Δx (and the tradeoffs that are
involved in this choice) can be found in Sec. 7.3.

For further reading, there exists a large choice of useful references that treat the aliasing
phenomenon from various different points of view. One may consult, for example,
[Bracewell86 pp. 197–198], [Brigham88 pp. 172–173], [Ramirez85 pp. 115–123],
[Briggs95 pp. 95–98] or [Smith07 Sec. 7.2.14].

5.3 Examples of aliasing in the one dimensional case

Let us illustrate the above discussion by a few 1D examples. The first examples concern
aperiodic functions g(x), whose spectrum (CFT) G(u) is continuous, and the remaining
examples concern periodic functions, whose spectrum is purely impulsive.
96 5. Issues related to aliasing

Example 5.1: Aliasing due to undersampling in the case of the function g(x) = sinc(ax):
Figure 5.4 shows a particular case of the generic Fig. 5.3 in which the function being
sampled is g(x) = sinc(ax), whose CFT is the square pulse G(u) = 1a rect(u/a) (see part (a)
of the figure). As long as the sampling frequency fs = Δ1x is at least twice fG = a/2, the
highest frequency in G(u), the signal g(x) is sampled sufficiently often as stipulated by the
sampling theorem, and the replicas of G(u) in the spectrum of the sampled signal are far
enough from each other to avoid overlapping. This means, indeed, that no aliasing occurs
(see part (c) of the figure). But when the sampling frequency is smaller than 2fG, say, f 's,
the signal is no longer sufficiently often sampled, and every two neighbouring replicas in
the spectral domain somewhat overlap (see part (d) of the figure, and notice there the
additive nature of this spectral overlapping). This is the spectral domain manifestation of
the aliasing. The signal domain interpretation of this aliasing is that only based on our too
far-apart sampling points, the original signal g(x) is undistinguishable from its alias gA(x),
the signal whose spectrum GA(u) consists of the central period (marked by an arrow) of
the periodic spectrum of Fig. 5.4(d). This signal-domain effect is shown in Fig. 5.2(b) for
the simple case in which a = 1 and hence fG = 12 , and the sampling frequency is 34 .
It is interesting to note that in the present example the explicit expression of the aliased
signal gA(x) can be easily found. For this end, we note that its spectrum GA(u), i.e. the
central period of the periodic spectrum in Fig. 5.4(d), is simply a 2/a-units high rect
function extending between – 12 f 's...12 f 's (where f 's is the corresponding sampling
frequency) minus an 1/a-units high rect function with a slightly narrower width extending
between –b...b where b = 12 f 's – (f G – 12 f 's ) = f 's – fG . By taking the inverse Fourier
transform of this difference of rect functions we find that gA(x) is simply a difference of
two sinc functions. For example, in the case shown in Fig. 5.2(b) where g(x) = sinc(x),
fG = 12 and f 's = 34 , GA(u) can be seen as a 2-units high rect function extending between
– 38 ...38 minus a 1-unit high rect function extending between – 14 ...14 (see Fig. 5.5(b)). It
follows therefore that the aliased function that mimics g(x) on the too far-apart sampling
points of Fig. 5.2(b) is expressed by gA(x) = 32 sinc(34 x) – 12 sinc(12 x). Note that gA(x) differs
from our original signal g(x) = sinc(x) only in its high frequencies. And indeed, if we
carefully compare the spectrum of gA(x) (Fig. 5.5(b)) with the spectrum of the original
signal g(x) (Fig. 5.5(a)), we see that: (1) The maximum frequency in GA(u) is slightly
lower than fG, the maximum frequency in G(u); and (2) The highest frequencies in GA(u)
are simply a result of the additive overlap with the neighbouring replicas. We see,
therefore, that gA(x) is indeed a lower-frequency alias of g(x): lower frequency due to point
(1); and alias because at our too low sampling frequency f 's = 34 the sampled versions of
g(x) and of gA(x) have precisely the same spectrum, i.e. the infinite periodic spectrum
which is shown in Fig. 5.5(c), so they are undistinguishable from each other.8

8 It should be noted for the record that g A (x) is not the only function that shares with g(x) the same
values at the given sampling points; for example, (g(x) + g A(x))/2 will also have the same values at
these sampling points, and the same goes for any other weighted mean of g(x) and gA(x). In fact, there
exist infinitely many functions that share the same values at these sampling points. Because their
sampled versions are identical, the Fourier transforms of their sampled versions are identical, too, and
they are all equal to the infinite periodic spectrum shown in Fig. 5.5(c). However, gA(x) has the lowest
maximum frequency, since it does not exceed half of the sampling frequency f 's (while g(x) does).
5. Issues related to aliasing 97

Signal domain Spectral domain

g(x) = sinc(ax)
1 G(u) = 1a rect(au )
1
a
x u
0 – 2 fG –fG 0 fG = a2 2 fG
1 x (a)
Δx III(Δx ) III(Δx u)
1 1
Δx fs = Δ1x

x u
0 –fs 0 fs > 2 fG
Δx
(b)
1 fs = Δ1x
1
a Δx
x –fs fs
u
0 –fs 0 fs > 2 fG
Δx 2 2
(c)
1 2 fs' = Δ1x
aΔx' '
1
aΔx'
x –fs' fs'
u
0 –fs' 0 fs' < 2 fG
Δx' (d)
2 2

Figure 5.4: A particular case of Fig. 5.3 showing the signal g(x) = sinc(ax), whose
CFT is the square pulse G(u) = 1a rect(u/a). The sampled version of g(x)
and its CFT are shown in (c). (d) When the sampling rate is lower than
twice the maximum signal frequency f G (say, f 's ), every two
neighbouring replicas of G(u) overlap additively on their borders,
giving in the overlapped zones false aliased spectral values with respect
to the original continuous-world spectrum G(u) shown in (a). The
horizontal arrows in the spectra (c) and (d) indicate the frequency range
of the DFT, which always extends between minus half and plus half of
the corresponding sampling frequency. See also a concrete example in
Fig. 5.5.

So far we have only discussed the continuous Fourier transform, so that the spectrum of
the sampled version of g(x) (see Figs. 5.4(c),(d)) is periodic (extending throughout the
range –∞...∞) and continuous (being a periodic repetition of the continuous spectrum
G(u)). Assume now that our sampled signal undergoes DFT rather than a continuous
 98 5. Issues related to aliasing

Fourier transform. As we have seen in Sec. 5.2, this has two main consequences: (1) The
resulting spectrum is now discrete; and (2) it only contains frequencies in the range
between – 12 fs and 12 fs, and includes therefore only frequencies located in the main replica of
the periodic spectrum belonging to the continuous Fourier transform.

Signal domain Spectral domain

g(x) = sinc(x) G(u) = rect(u)

1
1
x u
0 1
– – 1 0 1 1
2 2
(a)

gA(x) = 32 sinc( 34 x) – 12 sinc(12 x) GA(u) = 2rect( 43 u) – rect(2u)

1 2
1
x u
0 1
– – 3 – 1 – 1 0 1 1 3 1
4 2 4 4 2 4
(b) – 3 3
8 8

fs' = Δ1x
2 '
1 Δx'
1
Δx'
x u
0 1
– – 3 – 1 – 1 0 1 1 1
4 2 4 4 2
Δx' (c) fs' 3
3 fs'= 34
–
8 2 8
=

Figure 5.5: The spectral-domain interpretation of the aliasing shown in Fig. 5.2(b).
(a) The function g(x) = sinc(x) and its continuous Fourier transform
G(u) = rect(u). (b) The aliased function gA(x) = 32 sinc(34 x) – 12 sinc(12 x) and
its continuous Fourier transform GA(u) = 2rect(43 u) – rect(2u). As shown
in Fig. 5.2(b), the sampled versions of g(x) and gA(x) at the sampling
rate of f 's = 34 are identical; this implies that the spectra of both sampled
signals are identical, too, but here we show the reason directly in the
spectral domain: When convolving G(u) or GA(u) with an impulse comb
having impulse intervals of f 's = 34 (and impulse heights of 1/Δx' = f 's),
the periodic spectra we obtain are exactly identical, as shown here in (c).
Note that gA(x) is indeed a lower-frequency alias of g(x): Its maximum
frequency is exactly 12 f 's = 38 (see the spectrum GA(u) in (b)), whereas the
maximum frequency of g(x) is fG = 12 (see the spectrum G(u) in (a)). The
horizontal arrow in (c) indicates the frequency range of the DFT.
5. Issues related to aliasing 99

Figure 5.6 shows the sampled version of the signal g(x) = sinc(2x) as well as its
resulting DFT spectrum (after having performed all the required adjustments as discussed
in Chapters 3 and 4), for several gradually decreasing values of the sampling frequency fs
(and of the array length N). As we can see in rows (a)–(c) of the figure, as long as the
sampling frequency fs is higher than twice fG = 1 (the maximum frequency of g(x)), the
resulting adjusted DFT matches well the expected CFT of g(x), G(u) = 12 rect(u/2). Row (d)
of the figure shows the situation in the limit case when the sampling frequency fs is
exactly twice fG = 1, and rows (e) and (f) show what happens when fs < 2fG. As we can see,
in the last two rows the DFT results no longer match the continuous spectrum G(u) =
2 rect(u/2), due to the existence of aliasing. Note that because the DFT spectrum only
1

contains frequencies between – 12 fs...12 fs, i.e. frequencies in the central replica of the periodic
spectrum of Fig. 5.4(d), the overlapping of the neighbouring replicas due to aliasing looks
in the DFT spectrum as a cyclical foldover. In other words, frequencies that exceed from
one end of the DFT spectrum re-enter in the opposite end and are summed up with the
true spectral values belonging to the frequencies that are really located at that part of the
spectrum. Q

Example 5.2: Aliasing due to undersampling in the case of the function g(x) = sinc2(x):
This example shows another particular case of the generic Fig. 5.3 in which the function
being sampled is g(x) = sinc2(x), whose CFT is the triangle G(u) = tri(u) [Bracewell86 p.
415]. The DFT of the sampled version of this function is shown in Fig. 5.7 for the same
gradually decreasing values of the sampling frequency fs (and of the array length N) as in
Fig. 5.6. Once again, as long as the sampling frequency fs = Δ1x is at least twice fG = 1 (the
highest frequency in G(u)), the signal g(x) is sampled sufficiently often as stipulated by
the sampling theorem, and the replicas of G(u) in the spectrum of the sampled signal are
far enough from each other to avoid overlapping, so that no aliasing occurs. But when
fs < 2fG, the signal is no longer sufficiently often sampled, and every two neighbouring
replicas in the spectral domain somewhat overlap. This gives a cyclical foldover effect in
the DFT spectrum (see parts (e)–(f) of the figure, and notice there the additive nature of
this spectral overlapping).
Note that in the present example the aliased frequencies have a softer and less drastic
impact on the resulting spectrum than in the case shown in Example 5.1. This is, indeed,
the situation in most of the real-world cases, where the aliased frequencies only slightly
distort the shape of the resulting DFT (starting from both ends of the adjusted DFT
spectrum inward, toward the center). The existence of such aliasing may be difficult to
detect — unless we know in advance the precise shape of the expected continuous-world
spectrum. Q

Example 5.3: An example whose spectrum is not band limited, and contains both real-
valued and imaginary-valued parts (Fig. 5.8):
In the examples discussed so far the functions g(x) were symmetric about the origin, so
that their spectra were purely real-valued [Bracewell86 pp. 14–15]. The present example
illustrates the situation when both the real-valued and imaginary-valued parts of the
 100 5. Issues related to aliasing

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
1.5 k 1.5 n
g(x) = sinc(2x) CFT
1 1
• • • Sampled • • • DFT after
i&o reorg.
0.5 0.5 and vert. sca.
0 0

-0.5 -0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, fs = 8

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x) = sinc(2x)
1 1
• • • Sampled
0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(b) N = 32, fs = 4

Signal domain Spectral domain: Real part

1.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 k 1.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 n
g(x) = sinc(2x)
1 • • • Sampled 1

0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
x and vert. sca. Hz
-1 -1
-4. -3. -2. -1. 0 1. 2. 3. 4. -1.125 -0.5625 0 0.5625 1.125
(c) N = 18, fs = 2.25

Figure 5.6: The sampled version of the signal g(x) = sinc(2x) and its resulting DFT (after
all the required adjustments as discussed in Chapters 3 and 4). For the sake
of comparison, the original continuous function g(x) and its CFT G(u) =
2 rect(u/2) are also shown, using continuous lines. The successive rows of the
1

figure show the situation for gradually decreasing values of the sampling
 5. Issues related to aliasing 101

Signal domain Spectral domain: Real part

0 2 4 6 8 10 12 14 15 0 2 4 6 8 10 12 14 15
1.5 k 1.5 n
g(x) = sinc(2x)
1 1
• • • Sampled
0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

(d) N = 16, fs = 2

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
1.5 k 1.5 n
g(x) = sinc(2x)
1 1
• • • Sampled
0.5 0.5

0 0
CFT of the CFT
sampled g(x): • • • DFT after
-0.5 -0.5
aliased zone i&o reorg.
(see Fig. 5.4(d)) and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
0
Hz
-0.75 -0.375 0.375 0.75
(e) N = 12, fs = 1.5

Signal domain Spectral domain: Real part

1.5
0 1 2 3 4 5 6 7 8 9 k 1.5
0 1 2 3 4 5 6 7 8 9 n
g(x) = sinc(2x)
1 • • • Sampled 1

0.5 0.5

0 0
CFT of the CFT
sampled g(x): • • • DFT after
-0.5 -0.5
aliased zone i&o reorg.
(see Fig. 5.4(d)) and vert. sca.
-1 x -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -0.625 -0.3125 0 0.3125 0.625
(f) N = 10, fs = 1.25

frequency fs (and of the array length N). In rows (a)–(c) fs > 2fG and no
aliasing occurs. In row (d) fs = 2fG, while in rows (e)–(f) fs < 2fG and aliasing
appears in the form of a cyclical foldover: Frequencies which exceed from
one end of the DFT spectrum re-enter in the opposite end and are additively
superposed on top of the existing spectral values there. See Example 5.1.
 102 5. Issues related to aliasing

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
1.5 k 1.5 n
g(x) =sinc2(x) CFT
1 1
• • • Sampled • • • DFT after
i&o reorg.
0.5 0.5 and vert. sca.
0 0

-0.5 -0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, fs = 8

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x) =sinc2(x)
1 1
• • • Sampled
0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(b) N = 32, fs = 4

Signal domain Spectral domain: Real part

1.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 k 1.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 n
g(x) =sinc2(x)
1 • • • Sampled 1

0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
x and vert. sca. Hz
-1 -1
-4. -3. -2. -1. 0 1. 2. 3. 4. -1.125 -0.5625 0 0.5625 1.125
(c) N = 18, fs = 2.25

Figure 5.7: Same as Fig. 5.6, for the signal g(x) = sinc2(x) whose CFT is
G(u) = tri(u). The successive rows of the figure show the
situation for gradually decreasing values of the sampling
frequency fs (and of the array length N). In rows (a)–(c) fs > 2fG
and no aliasing occurs. In row (d) fs = 2fG, while in rows (e)–(f)
 5. Issues related to aliasing 103

Signal domain Spectral domain: Real part

0 2 4 6 8 10 12 14 15 0 2 4 6 8 10 12 14 15
1.5 k 1.5 n
g(x) = sinc2(x)
1 1
• • • Sampled
0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

(d) N = 16, fs = 2

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
1.5 k 1.5 n
g(x) = sinc2(x)
1 1
• • • Sampled
0.5 0.5

0 0
CFT of the CFT
sampled g(x): • • • DFT after
-0.5 -0.5
aliased zone i&o reorg.
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
0
Hz
-0.75 -0.375 0.375 0.75
(e) N = 12, fs = 1.5

Signal domain Spectral domain: Real part

1.5
0 1 2 3 4 5 6 7 8 9 k 1.5
0 1 2 3 4 5 6 7 8 9 n
g(x) = sinc2(x)
1 • • • Sampled 1

0.5 0.5

0 0
CFT of the CFT
sampled g(x): • • • DFT after
-0.5 -0.5
aliased zone i&o reorg.
x and vert. sca. Hz
-1 -1
-4. -3. -2. -1. 0 1. 2. 3. 4. -0.625 -0.3125 0 0.3125 0.625
(f) N = 10, fs = 1.25

fs < 2fG and aliasing appears in the form of a cyclical foldover:

Frequencies which exceed from one end of the DFT spectrum
re-enter in the opposite end and are additively superposed on top
of the existing spectral values there. See Example 5.2.
 104 5. Issues related to aliasing

Signal domain
0 8 16 24 32 40 48 56 63
1.5 k
g(x) = e-xstep(x)
1
• • • Sampled
0.5

-0.5

4. x
-1
-4. -3. -2. -1. 0 1. 2. 3.

Spectral domain: Real part Spectral domain: Imaginary part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
1.5 n 1.5 n
CFT CFT
1 1
• • • DFT after • • • DFT after
i&o reorg. i&o reorg.
0.5 and vert. sca. 0.5 and vert. sca.
0 0

-0.5 -0.5

4. Hz 4. Hz
-1 -1
-4. -3. -2. -1. 0 1. 2. 3. -4. -3. -2. -1. 0 1. 2. 3.

(a) N = 64, fs = 8

Figure 5.8: The sampled version of the signal g(x) = e–x step(x) and its resulting
DFT (after all the required adjustments as discussed in Chapters 3
and 4). For the sake of comparison, the original continuous function
g(x) and its CFT G(u) = 1/(1+[2π u]2) – 2π ui/(1+[2π u]2) are also
shown, using continuous lines. In this case the DFT spectrum
contains both real- and imaginary-valued parts, because g(x) is not
symmetric. Furthermore, because this function is not band limited,
aliasing cannot be totally avoided here by increasing the sampling
frequency fs. Part (a) of the figure shows the situation when N = 64
and fs = 8, and part (b) on the opposite page shows the situation
when N = 10 and fs = 1.25. Aliasing is clearly stronger in the latter
case, as indicated by the greater discrepancy between the DFT and
the original continuous spectrum G(u). See Example 5.3.

spectrum are non-zero. Consider the function g(x) = e–x step(x) whose CFT is G(u) =
1/(1+[2πu]2) – 2πui/(1+[2πu]2) [Bracewell86 p. 418]. Note that this function is not band
limited (meaning that its maximum frequency is fG = ∞), so that aliasing cannot be avoided
here by increasing the sampling frequency fs. Figure 5.8 shows the sampled version of
this function and its adjusted DFT for two different values of the sampling frequency (and
of the array length N). As we can see in part (a) of the figure, as long as the sampling
 5. Issues related to aliasing 105

Signal domain
0 1 2 3 4 5 6 7 8 9
1.5 k
g(x) = e-xstep(x)
1
• • • Sampled
0.5

-0.5

4. x
-1
-4. -3. -2. -1. 0 1. 2. 3.

Spectral domain: Real part Spectral domain: Imaginary part

0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
1.5 n 1.5 n
CFT
1 1
• • • DFT after
i&o reorg.
0.5 0.5 and vert. sca.
0 0
CFT
-0.5 • • • DFT after -0.5
i&o reorg.
and vert. sca.
-1 Hz -1 Hz
-0.625 -0.3125 0 0.3125 0.625 -0.625 -0.3125 0 0.3125 0.625

(b) N = 10, fs = 1.25

frequency fs = Δ1x is reasonably high, the signal is reasonably well sampled (obviously,
except in the sharp peak at the origin), and the replicas of G(u) in the spectrum of the
sampled signal do not significantly corrupt each other (although they obviously do
overlap). But when the sampling frequency is too low, as shown in part (b) of the figure,
the signal is no longer reasonably well sampled, and the overlapping between every two
neighbouring replicas in the spectral domain becomes much more significant, giving a
stronger cyclical foldover effect in the DFT spectrum. Note that because this function is
not band limited, each of the replicas in the spectrum of its sampled version extends ad
infinitum to both directions. This means that the main replica (the one contained in the
DFT spectrum) is additively contaminated not only by its two direct neighbours, but
indeed by all of the infinitely many replicas. Nevertheless, the quantitative influence of far-
away replicas is marginal here due to the decaying nature of the original spectrum G(u),
and most of the aliasing distortion is due to the intrusion of the closest replicas. Q

Example 5.4: Aliasing due to undersampling in the case of the simple periodic function
g(x) = cos(2πfx):
Fig. 5.9 shows a particular case of generic Fig. 5.3 in which the function being sampled
is g(x) = cos(2πfx) with frequency f (see part (a) of the figure). The CFT of this function
106 5. Issues related to aliasing

consists of a pair of impulses that are located at a distance f to both sides of the origin:
G(u) = 12 δ(u–f) + 12 δ(u+f) [Bracewell86 p. 412]. As long as the sampling frequency fs = Δ1x
is at least twice fG = f (the highest frequency in G(u)), the signal g(x) is sampled
sufficiently often as stipulated by the sampling theorem, and the replicas of G(u) in the
spectrum of the sampled signal are far enough from each other to avoid overlapping. This
means, indeed, that no aliasing occurs (see part (c) of the figure). But when the sampling
frequency is smaller than 2fG, say, f 's, the signal is no longer sufficiently often sampled,
and every two neighbouring replicas in the spectral domain somewhat overlap (see part (d)
of the figure, and note there how each impulse pair penetrates into the territory of its two
neighbouring replicas; note in particular the two false, lower-frequency impulses that
penetrate into the main replica). This is the spectral domain manifestation of the aliasing.
The signal domain manifestation of this aliasing is also straightforward: Only based on
our too far-apart sampling points, the cosine g(x) is undistinguishable from its alias gA(x),
the cosine whose spectrum GA(u) consists of the central replica (marked by an arrow) of
the periodic spectrum of Fig. 5.9(d). This signal-domain manifestation of the aliasing can
be clearly seen in Fig. 5.1(b). Note that in this case, because the original signal consists of
a single frequency, the aliasing effect manifests itself in a more spectacular way than usual,
in the form of a highly visible lower-frequency cosine (compare Fig. 5.1(b) with Fig.
5.2(b), where the aliasing is less conspicuous). The frequency fA of this aliased cosine
gA(x) = cos(2πfAx) is located at fs (or f 's) minus f (see the replica around f 's in Fig. 5.9(d)):
fA = fs – f (5.3)
For example, if the original cosine g(x) has the frequency f = 2 and the sampling
frequency is fs = Δ1x = 94 (which is indeed lower than 2f), the resulting aliased cosine gA(x)
has the frequency fA = 94 – 2 = 14 . Note that in the case shown in Fig. 5.1(b), in which the
original cosine has the same frequency f = 2 but the sampling frequency is only fs = 74
(which is again lower than 2f) we obtain fA = 74 – 2 = – 14 , which is obviously equivalent to
its positive-sign counterpart fA = 14 . The negative sign of the aliased frequency is obtained
here because the sampling frequency fs = 74 is smaller than the frequency f = 2 of the
original cosine, meaning that the aliased impulse is located here in the other side of the
origin, unlike in Fig. 5.9(d). Note that when the resulting aliased frequency fA falls close to
the spectrum origin the resulting low-frequency aliased cosine may be considered as a
sampling moiré effect; and indeed, Eq. (5.3) above is precisely the well-known formula of
the moiré frequency between two parallel gratings (see for example Eq. (2.11) in
[Amidror09 p. 20]). We will return to this subject later, when we discuss the more
interesting 2D case (see Examples 5.7 and on in Sec. 5.5, and the discussion in Sec. 5.6).
So far we have only discussed the continuous Fourier transform, so that the spectrum of
the sampled cosine contains an infinite number of replicas of the impulse pair G(u) and
extends throughout the range –∞...∞, as shown in Figs. 5.9(c),(d). But if our sampled
cosine undergoes DFT rather than a continuous Fourier transform, the resulting DFT
spectrum will only contain frequencies in the range – 12 fs...12 fs, where the main replica is
located (see the frequency range indicated by the horizontal arrow in Figs. 5.9(c),(d)).
Figure 5.10 shows, indeed, the DFT of the sampled cosine function with f = 1 (after all the
5. Issues related to aliasing 107

Signal domain Spectral domain

g(x) = cos(2πfx)
G(u) = 12 δ(u – f) + 12 δ(u + f)
1
1
2
x
0 u
– 2 fG –fG 0 fG = f 2 fG
1 x (a)
Δx III(Δx ) 1
III(Δx u)
1
Δx fs = Δ1x
.. . .. .
x u
0 –fs 0 fs > 2 fG
Δx
(b)

1 fs = Δ1x
1
2Δx
x
0 –fs fs
u
–fs 0 fs > 2 fG
Δx 2 2
fs – f f s+ f
(c)
= Replicas of
the impulse pair 1
1 fs' =
1 Δx'
2Δx'
x .. . .. .
0 –fs' fs'
u
Δx' –fs' 0 fs' < 2 fG
2 2 fs' + f
(d) fs' – f

Figure 5.9: A particular case of Fig. 5.3 showing the periodic signal g(x) = cos(2πfx),
whose CFT consists of a pair of impulses that are located at a distance f to
both sides of the origin: G(u) = 12 δ(u–f) + 12 δ(u+f). The sampled version of
g(x) and its CFT are shown in (c). (d) When the sampling rate is lower
than twice the maximum signal frequency fG (say, f 's), each replica of G(u)
overlaps additively with the two neighbouring replicas on their borders,
meaning that its impulse pair intrudes into the territory of the two
neighbouring replicas. In particular, two new false lower-frequency
impulses appear now in the main replica; this impulse pair belongs to the
new, false aliased cosine that becomes visible in the signal domain. The
horizontal arrows in the spectra (c) and (d) indicate the part of the infinite
continuous-world spectrum that will be represented in the DFT output
should we subject the sampled signal to DFT rather than to CFT; this
frequency range extends between minus half and plus half of the
corresponding sampling frequency. Note that in terms of the DFT
frequency range the aliasing effect can be viewed as a cyclical foldover,
where frequencies which exceed beyond one end of the range re-enter
(additively) at the opposite end.
 108 5. Issues related to aliasing

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 1.5 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5
-0.5 • • • DFT after
-1 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, fs = 8

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k 1.5 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5 • • • DFT after
-0.5
-1 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(b) N = 32, fs = 4

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 k 1.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5 • • • DFT after
-0.5
-1 i&o reorg.
x and vert. sca. Hz
-1.5 -1
-4. -3. -2. -1. 0 1. 2. 3. 4. -1.125 -0.5625 0 0.5625 1.125
(c) N = 18, fs = 2.25

Figure 5.10: Same as Fig. 5.6, for the signal g(x) = cos(2πfx) with frequency
f = 1 whose CFT is G(u) = 12 δ(u–1) + 12 δ(u+1). In rows (a)–(c)
fs > 2fG and no aliasing occurs. Row (d) shows the limit case
where fs = 2fG; note that because the array length N is even, the
DFT spectrum contains one less sample in the positive
frequency side (see Footnote 1 in Sec. 3.2), and therefore the
 5. Issues related to aliasing 109

Signal domain Spectral domain: Real part

0 2 4 6 8 10 12 14 15 0 2 4 6 8 10 12 14 15
k 1.5 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT of the CFT
-0.5 sampled g(x):
-0.5 • • • DFT after
-1 aliased zone i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

(d) N = 16, fs = 2

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
k 1.5 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT of the CFT
-0.5 sampled g(x): • • • DFT after
-0.5
-1 aliased zone i&o reorg.
(see Fig. 5.9(d)) and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
0
Hz
-0.75 -0.375 0.375 0.75
(e) N = 12, fs = 1.5

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 k 1.5
0 1 2 3 4 5 6 7 8 9 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT of the CFT
-0.5 sampled g(x): • • • DFT after
-0.5
-1 aliased zone i&o reorg.
(see Fig. 5.9(d)) and vert. sca.
-1.5 x -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -0.625 -0.3125 0 0.3125 0.625
(f) N = 10, fs = 1.25

impulse at the right end of the DFT is folded over and summed
up on top of its negative-frequency counterpart at the left end of
the DFT. In rows (e)–(f) fs < 2fG, so each of the two impulses
folds over and re-enters from the opposite side of the DFT
spectrum, thus simulating a false, lower-frequency cosine
function. See Example 5.4.
 110 5. Issues related to aliasing

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k n
2
g(x) 0.6
1.5 • • • Sampled 0
0.4 1
1 2
0.5 0.2 3
4 5 6 7
0 0
CFT
-0.5 -0.2 • • • DFT after
-1 i&o reorg.
-0.4
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, fs = 8

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k n
2
g(x) 0.6
1.5 • • • Sampled
0.4
1
0.5 0.2

0 0
CFT
-0.5 -0.2 • • • DFT after
-1 i&o reorg.
-0.4
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(b) N = 32, fs = 4

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 k 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 n
2
g(x) 0.6
1.5 • • • Sampled 0
0.4 1
1
0.2
2
0.5 3 4 5 6
0 0
CFT
-0.5 -0.2 • • • DFT after
-1 i&o reorg.
-0.4 and vert. sca.
-1.5 x Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -1.125 -0.5625 0 0.5625 1.125
(c) N = 18, fs = 2.25

Figure 5.11: Same as Fig. 5.6, for the periodic triangular signal g(x) = triangwave(2x)
with period P = 2 (i.e. frequency f = 12 ). As explained in Example 5.5, the
CFT of g(x), G(u), is an impulse train with impulse interval Δ u = f = 12
whose amplitude is modulated by half the CFT of a single triangular period
tri(2x), namely by 14 sinc2(12 u). Note in (c) how the impulses of the third
 5. Issues related to aliasing 111

Signal domain Spectral domain: Real part

0 2 4 6 8 10 12 14 15 0 2 4 6 8 10 12 14 15
k n
2
g(x) 0.6
1.5 • • • Sampled
0.4
1
0.5 0.2

0 0
CFT
-0.5 -0.2 • • • DFT after
-1 i&o reorg.
-0.4
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

(d) N = 16, fs = 2

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
k n
2
g(x) 0.6
1.5 • • • Sampled
0.4
1
0.5 0.2

0 0
CFT
-0.5 -0.2 • • • DFT after
-1 i&o reorg.
-0.4
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x 0
Hz
-0.75 -0.375 0.375 0.75
(e) N = 12, fs = 1.5

Signal domain Spectral domain: Real part

0 1 2 3 4 5 6 7 8 9 k 0 1 2 3 4 5 6 7 8 9 n
2
g(x) 0.6
1.5 • • • Sampled
0.4
1
0.5 0.2

0 0
CFT
-0.5 -0.2 • • • DFT after
-1 i&o reorg.
-0.4 and vert. sca.
-1.5 x Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -0.625 -0.3125 0 0.3125 0.625
(f) N = 10, fs = 1.25

harmonic and on exceed from one end of the DFT spectrum and re-enter
from the other end in the form of false in-between stray impulses. (This is
not visible in (b), because in this case the folded-over impulses fall exactly
on top of existing impulses, and they only corrupt their amplitudes.) In
(d)–(f) aliasing is already so strong that G(u) is no longer recognizable.
112 5. Issues related to aliasing

required adjustments as discussed in Chapters 3 and 4 and in Appendix A), for several
gradually decreasing values of the sampling frequency fs (and of the array length N). As
we can see in rows (a)–(c) of the figure, as long as the sampling frequency fs is higher
than twice fG = 1, the resulting adjusted DFT matches well the expected CFT of g(x),
G(u) = 12 δ(u–1) + 12 δ(u+1). Row (d) of the figure shows the situation in the limit case
when the sampling frequency fs is exactly twice fG = 1, and rows (e) and (f) show what
happens when fs < 2fG. As we can see, in the last two rows the DFT results no longer
match the continuous spectrum G(u), due to the existence of aliasing. Because the DFT
spectrum only contains frequencies between – 12 fs...12 fs, i.e. frequencies in the central replica
of the periodic spectrum of Fig. 5.4(d), the overlapping of the neighbouring replicas due
to aliasing looks in the DFT spectrum as a cyclical foldover, so that the impulse that
exceeds from one end of the DFT spectrum re-enters in the opposite end and appears
there as a false, lower-frequency stray impulse. Note that in this case no true impulses
appear in the DFT, meaning that the DFT spectrum we obtain is entirely flawed by
aliasing; this explains again, this time from the spectral-domain point of view, why the
aliasing effect in this case is so spectacular. Q

Example 5.5: Aliasing due to undersampling in a periodic triangular wave (Fig. 5.11):
In this example we consider a more interesting periodic function, whose CFT contains
not only the first harmonic impulses as in the previous example, but all their higher
harmonics too. As we know from Fourier theory, the CFT of the triangular function tri(x)
is sinc2(u) [Bracewell86 p. 415], so that the CFT of its narrower counterpart tri(2x)
extending between – 12 and 12 is 12 sinc2(12 u). Now, our periodic triangular wave g(x) =
triangwave(2x) can be seen as a convolution of the function tri(2x) with the infinite
impulse train t(x) = 12 III(12 x) having period P = 2 and amplitude 1. The CFT of t(x) is the
impulse train T(u) = III(2u), which has impulse intervals of f = 1/P = 12 and amplitude 12
[Bracewell86 p. 414]. Therefore, the CFT G(u) of our periodic triangular wave g(x) =
triangwave(2x) consists of the product of 12 sinc2(12 u) and III(2u), which is an infinite
impulse train that samples the “envelope” 14 sinc2(12 u) at impulse intervals of f = 12 (see, for
example, [Bracewell86 p. 213]). Note that unlike the cosine function discussed in the
previous example our triangular wave g(x) is not band limited, and its spectrum contains
an infinite number of gradually decaying harmonic impulses to both sides of the origin.
Figure 5.11 shows the sampled version of our periodic triangular wave g(x), as well as
its DFT spectrum (after all the required adjustments as discussed in Chapters 3 and 4 and
in Appendix A), for several gradually decreasing values of the sampling frequency fs (and
of the array length N). As we can see in parts (a)–(b) of the figure, as long as the sampling
frequency fs = Δ1x is reasonably high, the signal g(x) is reasonably well sampled, and the
replicas of G(u) in the spectrum of the sampled signal do not significantly corrupt each
other (although they obviously do overlap). But when the sampling frequency is too low,
as shown in parts (c)–(f) of the figure, the signal is no longer reasonably well sampled,
and the overlapping between every two neighbouring replicas in the spectral domain
becomes much more significant, giving a stronger cyclical foldover effect in the DFT
spectrum. Note that because the spectrum G(u) is impulsive, the overlap (or cyclical
5. Issues related to aliasing 113

foldover) due to aliasing implies that impulses that exceed from one end of the DFT
spectrum re-enter in the opposite end and appear there as false, lower-frequency stray
impulses. This can be clearly seen when comparing the DFT spectrum in parts (a) and (c)
of Fig. 5.11: While in (a) all the impulses up to the 7th harmonic appear in their correct
locations, in (c) only the impulses up to the second harmonic appear in their correct
locations, and the impulses of the third harmonic and on fold over and re-enter from the
opposite side of the DFT spectrum in the form of new, in-between stray impulses. (Note,
however, that in some cases, like in part (b) of the figure, the aliased stray impulses are not
visible because they fall precisely on top of true impulses, and they only corrupt their
amplitudes; see Remark 5.1 in Sec. 5.2). In parts (d)–(f) of the figure aliasing is already
so strong that the impulses of the original spectrum G(u) are no longer recognizable.
Finally, note that because our function is not band limited, each of the replicas in the
spectrum of its sampled version extends ad infinitum to both directions. This means that
the main replica (the one contained in the DFT spectrum) is additively contaminated not
only by its two direct neighbours, but by all of the infinitely many replicas. This explains,
indeed, the small differences in the impulse amplitudes that we see when comparing the
DFT spectra in rows (a)–(f) of Fig. 5.11: These amplitude differences are due to the
different ways in which impulses of the infinitely many replicas fold over and fall on top
of existing impulses in each of the rows. Nevertheless, the quantitative influence of far-
away replicas is marginal because of the decaying nature of the original spectrum G(u),
and most of the aliasing distortion is due to the intrusion of the closest replicas. Q

5.4 Aliasing in two or more dimensions

The above discussion about aliasing can be also extended to the 2D or MD cases. In
such cases the sampling in the signal domain and the corresponding convolution in the
spectral domain (see Eq. (5.2)) are simply extended to M dimensions:9
F [( Δ1x )M III(x/Δx) g(x)] = F [( Δ1x )M III(x/Δx)] * F [g(x)] = III(Δx u) * G(u) (5.4)
Figure 5.12 shows schematically how aliasing occurs due to undersampling in the 2D
case. The explanation in the MD case, too, is similar.

It should be noted, however, that although Eq. (5.4) is a straightforward generalization of

Eq. (5.2), the multidimensional case is much more rich and diversified than the 1D case.
For example, the effect of a rotation about the origin in the 2D case on the aliasing (see
below) has no equivalent in the 1D case. A discussion on sampling and aliasing in the 2D
continuous world can be found, for example, in [Rosenfeld82, Sec. 4.1] or [Bracewell95,
Chapter 7], and the connection of such 2D aliasing with the moiré effect is explained in
[Amidror09, Sec. 2.13]. More details on the sampling theorem and on aliasing in the MD
case can be found in [Petersen62], [Marks91, Sec. 6.2] or [Zayed93, Sec. 3.3].

9 We assume here that the same sampling interval Δx is used along all the M dimensions. If required, an
appropriate generalization may allow us to use a different sampling interval along each dimension.
114 5. Issues related to aliasing

5.5 Examples of aliasing in the multidimensional case

To better illustrate the aliasing phenomenon in the multidimensional DFT, let us

consider a few 2D examples with various spectral properties. Note that all our 2D DFT
figures are drawn as density plots (see Sec. 1.5.2, as well as Footnote 14 on p. 219).

Example 5.6: Consider the 2D case shown in Fig. 5.13(a). Here, the given continuous-
world function g(x,y) is an infinite vertical bar of height 1 and width τ that is centered
about the origin and rotated by angle θ; its spectrum G(u,v) is an infinite, continuous
straight line-impulse (blade) passing through the origin at the perpendicular direction,
whose amplitude is modulated by a sinc function (see Sec. D.10 in Appendix D). The
spectrum of this function is not band limited since it contains frequencies going ad
infinitum to both sides of the origin, and therefore it largely exceeds the maximum
frequency that can be represented in the output array of the DFT. As we can see in Fig.
5.13(a), wherever the line-impulse exceeds the borders of the DFT spectrum it simply
folds over and re-enters into the DFT spectrum from the opposite side. Note that no parts
of the line-impulse are thereby lost, and its folded-over parts simply pursue their course, at
the same angle θ, starting from the opposite end of the DFT spectrum. But because in the
continuous world the line-impulse extends to both directions ad infinitum, it follows that
in the discrete world the folded-over blade exceeds the borders of the DFT spectrum again
and again, each time re-entering and pursuing its course from the opposite side.10 This

Figure 5.12: Schematic illustration of aliasing due to undersampling in the 2D case.

This figure is a 2D counterpart of the generic 1D case shown in Fig. 5.3.
(a) An original continuous band-limited 2D signal g(x,y) and its
continuous-world spectrum G(u,v). (b) The sampling impulse nailbed and
its continuous-world spectrum. (c) The sampled signal is the product of
signals (a) and (b), and therefore its continuous-world spectrum is the
convolution of their spectra. Note the replicas of the original spectrum, that
are located at all integer linear combinations of the sampling frequencies us
and vs. (d) If the sampling rate along any of the two main directions is
lower than twice the maximum signal frequency in that direction, uG or vG,
then every two neighbouring replicas of G(u,v) overlap (additively) along
this direction, giving false aliased frequencies in the overlapped zones. The
value of the spectrum in the aliased areas in (d) is simply the sum of the
values of the overlapped replicas. The horizontal and vertical arrows in the
spectra (c) and (d) indicate the part of the infinite continuous-world
spectrum that will be represented in the 2D DFT output should we subject
the sampled signal to DFT rather than to CFT: The frequency range of the
DFT extends along each of the axes between minus half and plus half of
the corresponding sampling frequency. See Problem 5-9 for cases where
the original signal g(x,y) is not band-limited.

10In fact, if we consider the infinite 2D continuous-world spectrum explaining the sampling of g(x,y)
(see Problem 5-9 and Fig. 5.21), these multiple cycles of foldover are really caused by overlapping of
the main replica by farther away replicas of the line impulse. But again, just as in the 1D case (see
p. 94), from the point of view of the DFT this is fully equivalent to the folding-over interpretation.
5. Issues related to aliasing 115

Signal domain Spectral domain

y v
g(x,y) G(u,v)
vG
x u
–2uG –uG 0 uG 2uG
–vG

(a) v
y
.. .. .. .. .. .. .. .. .. .. .. .. .. . vs . .
.. .. .. .. .. .. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. .. .. .. .. .. .. x . . . u
.. .. .. .. .. .. .. .. .. .. .. .. .. –us 0 us >2uG

.. .. .. .. .. .. .. .. .. .. .. .. ..
............. (b)
. –vs . .
v

y
vs

.. ... ... ... ... ... ..

vs
2

.. .. .. .. .. .. .. x u
. .. .. .. .. .. . –us –us us us >2uG
2
0 2
–vs
2

–vs
(c)

v
y

... v's
..... v's

..... x
2
u
..... –v's –u's 0 u's< 2uG
... 2
–u's –v's u's
2 2

(d)
116 5. Issues related to aliasing

can be clearly seen in the DFT spectrum if the amplitude of the line-impulse does not fade
out too quickly (see Fig. 5.13(a)).11 See also Problem 5-9 and Fig. 5.21.

Note that if the line-impulse runs along one of the two main axes of the DFT spectrum
or along one of its two main diagonals, i.e. if θ = k·45° with an integer k, the line-impulse
parts which exceed from one side of the DFT spectrum fall, when they re-enter from the
opposite side, exactly on top of the same line-impulse, and their amplitudes are simply
summed up (see Fig. 5.13(c)). In this case no “stray line-impulses” appear in the
spectrum even if strong aliasing occurs. In fact, a similar cyclical self-overlapping of the
line-impulse occurs whenever the angle θ has a rational slope, i.e. whenever we have
tanθ = mn with integer m,n, except that in this general case the folded-over line-impulse
may first complete one or more folding-over cycles through the DFT spectrum before it
ends up falling over itself (so that some “stray line-impulses” may still be visible in the
DFT spectrum). This happend, for example, in Fig. 5.13(a), where m = 4 and n = 1. Q

Figure 5.13: Aliasing in 2D DFT: a case whose spectrum consists of a 1D

line-impulse (see Example 5.6). The original continuous-world
function in the signal domain is a rotated, 1-valued straight bar
of width τ . Its continuous-world spectrum is a perpendicular
infinite line-impulse passing through the origin, whose
amplitude is modulated by a sinc function (see Sec. D.10).
Wherever this spectral line-impulse exceeds the boundaries of
the DFT spectrum, it folds over and re-enters from the opposite
side. The three first rows only differ in their rotation angle θ:
(a) θ = 14.036° (which has a rational slope of 4:1, since
tanθ = 1/4); (b) θ = 10°; (c) θ = 45° (which has a rational slope
of 1:1). When the line-impulse runs along the diagonal of the
DFT spectrum, as in case (c), its parts which exceed from one
side re-enter from the other side and fall there exactly on top of
the same line-impulse; their amplitudes are simply summed up.
A similar cyclical self-overlapping of the line-impulse occurs
whenever θ has a rational slope, except that in this more general
case the folded-over line-impulse may first complete one or
more folding-over cycles through the DFT spectrum before it
ends up falling over itself. This happens, indeed, in case (a), but
not in case (b) which does not have a rational slope. Case (d) is
similar to case (a), but differs in the width τ of the straight bar;
note that the bar width only influences in the spectral domain
the opening of the sinc function that modulates the amplitude of
the line impulse. The small oscillations around the center of the
DFT spectrum are due to leakage (see Chapter 6). See also
Problem 5-9 and Fig. 5.21.

11This case clearly illustrates the fact that the spectral foldover due to aliasing does not consist of
mirroring back from the same edge of the DFT spectrum, as already mentioned in Footnote 6. The
correct interpretation of the foldover is that whatever exceeds from one side of the DFT spectrum re-
enters in a cyclical way from the opposite side.
5. Issues related to aliasing 117

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(a)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(b)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(c)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(d)
118 5. Issues related to aliasing

Example 5.7: 1D impulsive spectrum (impulse comb) over the u,v plane (Fig. 5.14):
In this example we consider the periodic counterpart of Example 5.6, i.e. a periodic line-
grating that is centered about the origin and rotated by angle θ (see Fig. 5.14). The
spectrum of this continuous-world function is no longer a continuous blade but rather a
subset thereof, namely, an impulse comb consisting of the DC impulse at the spectrum
origin and infinitely many equidistant harmonic impulses that are all located on the same
straight line along the angle θ.
Just as in Example 5.6, the spectrum of this continuous-world function is not band
limited, and it largely exceeds the frequency range that is covered by the DFT spectrum.
And indeed, in this case, too, wherever the impulse comb exceeds the borders of the DFT
spectrum it simply folds over and re-enters into the DFT spectrum from the opposite side;
and once again, if the impulse amplitudes do not fade out too quickly the impulse comb
may be seen folding over even more than once, as is the case in Fig. 5.14.
Note, however, that in this case the folded-over impulses do not necessarily coincide
with existing impulse locations even if we have tanθ = mn with integer m,n. For this to
happen, we also need to impose a condition on the vector p which determines the period
(in terms of array elements) of our rotated line grating: it must be purely integer. We will
return to this question in Appendix C (see Remarks C.2 and C.5 there for full details).
Finally, note that if any of the folded-over impulses due to aliasing happens to fall close
enough to the spectrum origin, its new frequency there is so small that it may correspond
back in the signal domain to a new, visible low-frequency parasite structure (see Fig.
5.14(d)). This parasite structure does not exist in our original function (the continuous-
world line grating), and it only appears in its sampled version due to aliasing (i.e. due to
the undersampling of the original line grating). Such low-frequency structures are known,
particularly in the 2D case, as sampling moirés; this subject is explained in greater detail
in [Amidror09, Sec. 2.13]. Q

Figure 5.14: Aliasing in 2D DFT: a case whose spectrum consists of a 1D impulse-

comb (see Example 5.7). The original continuous-world function in the
signal domain is a rotated line grating, whose spectrum is a rotated
infinite impulse comb. Whenever this impulse-comb exceeds from one
side of the DFT spectrum due to aliasing it folds over and re-enters from
the opposite side. The only difference between the four rows is in the
rotation angle θ: (a) θ = 14.036° (which has a rational slope of 4:1, since
tanθ = 1/4); (b) θ = 10°; (c) θ = 8°; and (d) θ = 3°. Note that in each
case several cycles of foldover are clearly visible in the DFT spectrum.
In case (d) folded-over impulses fall very close to the origin, and indeed,
they correspond to a new lower-frequency parasite structure (sampling
moiré) that is clearly visible in the signal domain, i.e. in the sampled line
grating. The cross-like oscillations which surround impulses in the DFT
spectrum are due to leakage (see Chapter 6). See Footnote 5 in Sec. 1.5.
5. Issues related to aliasing 119

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(a)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(b)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(c)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(d)
120 5. Issues related to aliasing

Example 5.8: 1D impulsive spectrum (three isolated collinear impulses) over the u,v
plane (Fig. 5.15):
Consider a raised cosine function of the form g(x,y) = 12 cos(2π fx) + 12 that undergoes a
small rotation by angle θ = 3° about the origin. The spectrum of this continuous-world
function consists of three impulses: the DC impulse at the origin (the Fourier transform of
the constant 12 ), plus a pair of impulses (the Fourier transform of the cosine itself) that are
located at a distance f to both sides of the origin along the direction θ = 3°. Because the
function g(x,y) is band limited, its sampled version is indeed alias-free as long as the
cosine frequency f is sufficiently low with respect to the sampling frequency (see Fig.
5.15(a)–(c)). But if we increase the cosine frequency f until the sampling rate is no longer
sufficient, the two impulses of the cosine will exceed the DFT spectrum borders and re-
enter from the opposite side as aliased impulses. In the signal domain, this aliasing means
that a lower-frequency cosine mimics (and takes the place of) the original high-frequency
cosine. And if the folded-over impulses happen to fall close enough to the spectrum
origin, the corresponding low-frequency aliased cosine back in the signal domain becomes
so prominent that it may be considered as a sampling moiré effect. Such a case is shown
in Fig. 5.15(d). Note that Fig. 5.15(c), too, has in the signal domain a similar but less
visible low-frequency structure; but this structure is not a true moiré effect (note that it is
not accompanied by corresponding low-frequency impulses in the spectrum). This
pseudo-moiré artifact is explained in Sec. 8.6. Q

Figure 5.15: Aliasing in 2D DFT: a case whose spectrum consists of three isolated
collinear impulses. The original continuous-world function in the signal
domain is a raised cosine of the form g(x,y) = 12 cos(2π fx) + 12 that has been
slightly rotated by angle θ = 3°. Its spectrum is the sum of the Fourier
transform of the cosinusoidal term, which consists of two symmetric
impulses of amplitude 14 located at a distance f to both sides of the origin
along the direction θ, plus the Fourier transform of the additional constant
2 , which is an impulse of amplitude 2 at the origin (the DC impulse). In all
1 1

of the four rows we have N = 128, R = 16, F = 8, Δx = 1/F = 1/8 (by Eq.
(7.4)) and Δu = 1/R = 1/16 Hz (by Eq. (7.3)), and the sampling frequency
is fs = 1/Δx = 8 Hz (by Eq. (4.6)). The only difference between the four
rows is in the period P of the cosine: (a) P = 1.05, so that f = 1/P = 0.95
Hz; (b) P = 1.05 / 2, so that f = 1.90 Hz; (c) P = 1.05 / 4, so that f = 3.81
Hz; (d) P = 1.05 / 8, so that f = 7.62 Hz. As we can see, in rows (a)–(c) all
the impulses are located within the range of the DFT spectrum, and no
aliasing (foldover) occurs. But in row (d) the cosine impulses exceed the
DFT borders, fold-over and end up close to the spectrum origin; and
indeed, they correspond to a new low-frequency parasite structure
(sampling moiré) that is clearly visible in the signal domain, instead of the
sampled cosine. The low-frequency structure that is visible in row (c) is not
a true moiré effect (note that there are no corresponding low-frequency
impulses in the DFT spectrum); this artifact is a pseudo-moiré effect, as
explained in Sec. 8.6. See also Footnote 5 in Sec. 1.5.2.
5. Issues related to aliasing 121

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(a)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(b)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(c)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(d)
122 5. Issues related to aliasing

Example 5.9: 2D non-impulsive spectrum over the u,v plane (Fig. 5.16):
Aliasing may also occur in 2D cases in which the continuous-world spectrum is smooth
and non-impulsive. For example, the spectrum of the square pulse g(x,y) = rect(x/τ,y/τ) =
rect(x/τ) rect(y/τ) is the 2D sinc function G(u,v) = τ 2 sinc(τu) sinc(τv) which extends
throughout the entire u,v plane. Because this spectrum is not band limited it largely
exceeds the frequency range that is covered by the 2D DFT spectrum, so that aliasing in
this case is unavoidable. But just as in the 1D case (see Fig. 3.5(f)), since G(u,v) is rather
smooth and non-impulsive, the foldover due to aliasing only slightly modifies the shape of
the DFT spectrum, and it may be quite difficult to detect unless the true values of the
original spectrum G(u,v) are known in advance. This is illustrated by Fig. 5.16(a), in
which the existence of aliasing in the DFT spectrum is not easy to detect visually. Note
that aliasing is easier to detect in the slightly rotated version of this case shown in Fig.
5.16(b): if aliasing were not present, the rotated 2D sinc spectrum would have been
perfectly symmetric with respect to its new rotated axes. Q

Example 5.10: A curvilinear line-impulse over the u,v plane (Fig. 5.17):
Another interesting example of aliasing involving a line impulse in the spectral domain is
shown in Fig. 5.17. In this case the continuous-world function g(x,y) is a circular
cosinusoidal grating with radial period T, which extends throughout the entire x,y plane. Its
spectrum G(u,v) consists of a circular line-impulse (impulsive ring) of radius f = 1/T
which is centered about the origin.12 The only difference between the twelve rows of the
figure is in the radial period T of the circular grating. As long as the radial period is
sufficiently large, the sampling rate is sufficient for capturing the main features of the
circular grating and there is no aliasing (see rows (a)–(d) in the figure). In such cases the
impulsive ring in the spectrum is fully contained within the frequency range of the DFT
spectrum, and no foldover occurs. The signal-domain counterpart is that the circular
cosinusoidal grating can be easily recognized in the sampled figure, and only some minor
distortions can be observed.13 But when the radial period T gets smaller the sampling rate
becomes insufficient. In such cases (see rows (e)–(l) in the figure) the spectral impulsive
ring exceeds the borders of the DFT spectrum and folds-over, so it is not surprising that
in the signal domain the structure of the circular grating is lost and a lower-frequency
structure takes its place. And if the folded-over impulse ring falls sufficiently close to the
spectrum origin (see rows (h)–(j) and (l)), the new aliased low-frequencies near the origin
manifest themselves in the form of a strong sampling moiré effect that completely
overshadows the original structure of the circular grating.
When the folded-over frequencies fall exactly on the spectrum origin, they correspond
to a singular moiré, i.e. a moiré having frequency zero. For example, in Fig. 5.17(i)
the moiré effect is singular along the two main axes, i.e. along the directions from which

12 More details on this continuous-world function and on the analytic expression of its spectrum can be
found in [Amidror97].
13 Note that these distortions are not caused by aliasing, since in rows (a)–(d) of Fig. 5.17 there is no

aliasing. The minor distortions in the signal domain, particularly visible in row (d), are discussed in
Problem 5-18; the minor ripple artifact in the DFT spectra is due to leakage (see Chapter 6).
5. Issues related to aliasing 123

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(a)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(b)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(c)

Figure 5.16: Aliasing in 2D DFT: a case whose spectrum is smooth and

continuous in all directions (see Example 5.9). (a) The
spectrum of a square pulse in the continuous-world is a 2D sinc
function. Because this function extends throughout the entire
u,v plane it is clear that the DFT spectrum in this case is flawed
by aliasing; and yet, aliasing artifacts are not easy to detect
visually. (b) Aliasing is easier to detect in a slightly rotated
version of this function; note that without aliasing the rotated
sinc function would have been perfectly symmetric with respect
to its rotated axes. (c) In the 45° rotated version of this function
aliasing is again not easy to detect. See also Footnote 5 in Sec.
1.5.2 concerning the gray levels used in the figure.
124 5. Issues related to aliasing

the folded-over impulse-ring touches the spectrum origin. But when the folded-over
frequencies fall just slightly away from the spectrum origin the moiré effect has low,
visible frequencies and it becomes very prominent. A more detailed explanation on this
subject and on the geometric shapes (hyperbolic, parabolic or elliptic) of the moiré fringes
that are obtained in this case can be found in [Amidror09 Sec. 10.7.5].
Note that in Fig. 5.17 the aliasing effect may look as a mirroring-back or a reflection of
the exceeding parts of the spectral rings from the same edge of the DFT spectrum. But
once again, the correct interpretation of the foldover is that whatever exceeds from one side
of the DFT spectrum re-enters in a cyclical way from the opposite side. Q

Example 5.11: Several curvilinear line-impulses over the u,v plane (Fig. 5.18):
A further example of aliasing involving line impulses in the spectral domain is shown in
Fig. 5.18. In this case the continuous-world function g(x,y) is a 0/1-valued circular grating
with radial period T, which extends throughout the entire x,y plane. Its spectrum G(u,v)
consists of an infinite series of concentric circular line-impulses (impulsive rings) whose
radiuses are positive integer multiples of f = 1/T.14 Because this continuous-world
spectrum is not band limited it is clear that aliasing here is unavoidable. And indeed, Fig.
5.18 clearly shows the folding-over of the spectral rings which exceed the borders of the
DFT spectrum due to aliasing. The only difference between the twelve rows of the figure
is in the radial period T of the circular grating. As long as the radial period is sufficiently
large, the sampling rate is “reasonably sufficient” for capturing the main features of the
circular grating (see rows (a)–(d) in the figure). In such cases the main impulsive ring in

Figure 5.17: Aliasing in 2D DFT: a case whose spectrum consists of a circular line-
impulse (see Example 5.10). In this case the original continuous-world
function in the signal domain is a circular cosinusoidal grating with
radial period T which extends throughout the entire x,y plane: g(x,y) =
cos(2π x 2 + y 2 /T). Its spectrum consists of a circular impulsive ring of
radius f = 1/T about the origin.14 In all of the twelve rows we have
N = 128, R = 16, F = 8, Δx = 1/F = 1/8 (by Eq. (7.4)) and Δu = 1/R =
1/16 Hz (by Eq. (7.3)), and the only difference between the rows is in
the radial period T of the circular grating: (a) T = 1, i.e. 8 array
elements; (b) T = 0.75, i.e. 6 array elements; (c) T = 0.5, i.e. 4 array
elements; (d) T = 0.3, i.e. 2.4 array elements; (e) T = 0.25, i.e. 2 array
elements; (f) T = 0.2, i.e. 1.6 array elements; (g) T = 0.15, i.e. 1.2 array
elements; (h) T = 0.13, i.e. 1.04 array elements; (i) T = 0.125, i.e. 1
array element; (j) T = 0.12, i.e. 0.96 array elements; (k) T = 0.1, i.e. 0.8
array elements; (l) T = 0.09, i.e. 0.72 array elements. As long as the
radial period T is sufficiently large, the sampling rate is sufficient for
capturing the main features of the circular grating (see rows (a)–(d)).
But when the radial period T gets smaller the sampling rate becomes
insufficient, as in rows (e)–(l). See also Problems 5-16 and 5-18.

14More details on this continuous-world function and on the analytic expression of its spectrum can be
found in [Amidror97]. See also Footnote 5 in Sec. 1.5.2 concerning the gray levels used in the figure.
5. Issues related to aliasing 125

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(a)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(b)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(c)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(d)
126 5. Issues related to aliasing

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(e)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(f)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(g)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(h)
5. Issues related to aliasing 127

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(i)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(j)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(k)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(l)
128 5. Issues related to aliasing

the spectrum (i.e. the inner, first-harmonic ring, which contains the most significant
spectral information about g(x,y)) is fully contained within the frequency range of the DFT
spectrum, and only the higher-harmonic impulsive rings exceed the borders and fold-over
due to the aliasing. The signal-domain counterpart is that the circular grating can be easily
recognized in the sampled figure, and only some minor aliasing distortions can be
observed (compare rows (a)–(d) with the same rows in Fig. 5.17 where aliasing does not
occur at all). But when the radial period T gets smaller the sampling rate becomes
insufficient. In such cases (see rows (e)–(l) in the figure) the main spectral impulsive ring
itself is folded over, so it is not surprising that in the signal domain the circular grating is
so much corrupted that it can no longer be recognized. As we can see, when aliasing is
moderate it disturbs only slightly, and the structures of the circular grating and of its
spectrum are still recognizable in the discrete world. But when aliasing gets stronger its
effect becomes devastating and it completely destroys the structure of the circular grating.
And if the main impulse ring itself is folded-over and falls sufficiently close to the
spectrum origin (see rows (h)–(j) and (l)), the new aliased low-frequencies near the origin
manifest themselves in the form of a strong sampling moiré effect that completely
overshadows the original structure of the circular grating.

Figure 5.18: Aliasing in 2D DFT: a case with several concentric circular line-impulses
in the spectral domain (see Example 5.11). In this case the original
continuous-world function is a 0/1-valued circular grating with the same
radial period T as in Fig. 5.17, which extends throughout the entire x,y
plane. Its spectrum consists of a DC impulse at the origin plus an infinite
series of concentric impulsive rings whose radiuses are positive integer
multiples of f = 1/T.14 Because this spectrum is not band limited it is clear
that aliasing here is unavoidable. And indeed, we can clearly see in the
figure the folding-over of the spectral rings which exceed the borders of
the DFT spectrum due to aliasing. In all of the twelve rows we have
N = 128, R = 16, F = 8, Δx = 1/F = 1/8 (by Eq. (7.4)) and Δu = 1/R = 1/16
Hz (by Eq. (7.3)), and the only difference between the rows is in the radial
period T of the circular grating: (a) T = 1, i.e. 8 array elements;
(b) T = 0.75, i.e. 6 array elements; (c) T = 0.5, i.e. 4 array elements;
(d) T = 0.3, i.e. 2.4 array elements; (e) T = 0.25, i.e. 2 array elements;
(f) T = 0.2, i.e. 1.6 array elements; (g) T = 0.15, i.e. 1.2 array elements;
(h) T = 0.13, i.e. 1.04 array elements; (i) T = 0.125, i.e. 1 array element;
(j) T = 0.12, i.e. 0.96 array elements; (k) T = 0.1, i.e. 0.8 array elements;
(l) T = 0.09, i.e. 0.72 array elements. As long as the radial period T is
sufficiently large, the sampling rate is “reasonably sufficient” for
capturing the main features of the circular grating (see rows (a)–(d)). But
when the radial period T gets smaller the sampling rate becomes
insufficient, as in rows (e)–(l). As we can see, when aliasing is low the
structures of the original grating and of its spectrum are still recognizable
in the discrete world. But when the sampling rate is no longer sufficient
aliasing becomes so devastating that it completely alters in the discrete
world the structures of the circular grating.
5. Issues related to aliasing 129

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(a)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(b)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(c)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(d)
130 5. Issues related to aliasing

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(e)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(f)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(g)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(h)
5. Issues related to aliasing 131

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(i)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(j)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(k)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(l)
132 5. Issues related to aliasing

Unlike in the previous example where no higher-harmonic impulsive rings exist, in the
present example sampling moiré effects may also occur when a higher-harmonic
impulsive ring folds over and falls close to the spectrum origin. But the higher the
harmonic of the ring in question, the lower the intensity of the resulting moiré effect. For
example, compare the singular moiré effect of row (c), which is generated by the foldover
of the 4-th harmonic ring, with that of row (i), which is generated by the first harmonic
ring. Q

5.6 Discussion

The main point we learn from the examples above is that in all cases, frequencies which
exceed from one side of the DFT spectrum due to aliasing wrap-around and re-enter
additively from the opposite side. This is, indeed, a generalization of the 1D case, in which
foldover due to aliasing occurs along the u axis. Furthermore, as long as the main
frequencies of the given function (like the first harmonic impulses in the case of a periodic
function) do not exceed beyond the borders of the DFT spectrum and are not folded-over,
the original structure of the given function is basically preserved in its sampled version (of
course, if some of the less important frequencies do exceed and fold over, some minor
distortions may occur; for example, compare Figs. 5.17(a)–(d) with Figs. 5.18(a)–(d)).
But when the main frequencies of the given function exceed from the borders of the DFT
spectrum, it is clear that the original structure of the given function is no longer preserved
in its sampled version, and it is altered (aliased) by a parasite lower-frequency structure.
Finally, if any of the folded-over frequencies falls close enough to the spectrum origin,
this aliased frequency may manifest itself in the signal domain in the form of a parasite
low-frequency sampling moiré effect. And if the folded-over frequencies which fall close
to the spectrum origin are main frequencies of the given function, as in Figs. 5.18(h)–(j)
and (l), the resulting sampling moiré effect is particularly strong and dominant.

Note that aliasing takes longer to appear along diagonal directions: While along the
main axes frequencies begin to exceed from the DFT spectrum starting from half of the
sampling frequency, along the diagonals frequencies start to exceed slightly farther away
(for example, in the 2D case we have a factor of 2 ; see Fig. 5.12). However, many of
these gained non-aliased higher frequencies are infested by strong pseudo-moiré artifacts
(see Sec. 8.6 and Fig. 8.15) and are therefore not really advantageous.

Remark 5.3: At this point, we may ask ourselves the following interesting question: How
can we tell if a certain frequency in the DFT spectrum (for example, a certain impulse) is
indeed a true spectral frequency of the given function, or just an aliasing artifact (a folded-
over frequency)? A simple method which often allows to distinguish between true and
aliased frequencies in the multidimensional case is based on the rotation theorem. This
continuous-world theorem states that if a function g(x,y) has the Fourier transform G(u,v),
then applying a rotation to g(x,y) will simply result in the spectral domain in an identical
rotation of G(u,v) [Bracewell95 p. 157]. In the discrete world, however, this theorem
5. Issues related to aliasing 133

remains valid only for the true frequencies in the DFT spectrum,15 but not for aliased
frequencies. For example, Figs. 5.14(a)–(d) show what happens in the DFT spectrum to
the impulse comb when we apply DFT to g(x,y) after having rotated it by a small angle θ
(note that we continue using the same unrotated sampling grid as before the rotation). As
we can see, all the true impulses of G(u,v) simply undergo a rotation by angle θ, while the
folded-over (aliased) impulses move in a different way. (In fact these impulses, too,
undergo a rotation by angle θ, but their center of rotation is not the origin of the DFT
spectrum, but rather the center of the replica to which they actually belong, which is
obviously located outside the range of the DFT spectrum.) Other examples illustrating the
effect of rotation on true and aliased frequencies in the DFT spectrum are shown in Fig.
5.13 and in Fig. 5.16. Note, however, that this method does not work in cases such as
Figs. 5.17 and 5.18, which are circularly symmetric and hence invariant under rotations.
A similar test can be also obtained by applying to g(x,y) a stretching transformation,
since according to the dilation theorem (see point 2 in Sec. 2.4.2) the Fourier transform of
1
the continuous-world function g(ax,by) is ab G(u/a,v/b) [Bracewell95 p. 154]; however, in
this case shrinking of g(x,y) results in an expansion of G(u,v) and vice versa. Note that this
test can be also used in the 1D case, in which the method based on rotations is not
applicable since rotations have no 1D equivalent. The same effect can be also obtained by
varying the sampling rate that is being used to sample the continuous-world input function
(see Sec. B.3 in Appendix B). Q

Remark 5.4: (MD counterpart of Remark 5.2): Just like in the 1D case, in the
multidimensional case, too, aliasing can be avoided completely only if the given signal g(x)
is band limited; under this condition, aliasing will be absent if the sampling frequency
along each of the axes is at least twice the highest frequency present in g(x). But in
practice, such situations rarely occur. If g(x) is not band limited (this happens, for
example, for any finite-duration function g(x)), the only solution is to reduce aliasing as
much as possible, using the same methods as in the 1D case. Q

Another issue that is often associated with aliasing in the 2D or MD cases concerns the
“jaggies” or “staircasing” which appear in the discrete world along slanted or curved
lines due to their sampling. The interesting question of how precisely these jaggies are
related to aliasing is treated separately in Sec. 8.5.

5.7 Signal-domain aliasing

The use of DFT as an approximation to the CFT involves sampling both in the input and
in the output (see Sec. 2.5). Therefore, although aliasing is usually considered as an
artifact which occurs due to undersampling in the signal domain, a similar phenomenon

15With the understanding that after the rotation an impulse may end up in a non-integer location within
the DFT spectrum, or even outside the range of the DFT spectrum. The first case will simply result in
leakage (see Chapter 6), but in the second case the impulse in question will be folded-over and become
an aliased impulse.
134 5. Issues related to aliasing

may also occur as a result of undersampling in the spectral domain, thanks to the
frequency sampling theorem — the frequency-domain dual of the classical sampling
theorem [Thornhill83]. This phenomenon, known as signal-domain aliasing (or possibly
in the 1D case time-domain aliasing), may be regarded under two different angles:

The first point of view is based on the inverse DFT. Due to the duality between DFT
and the IDFT, it follows that all our previous discussions on the aliasing phenomenon
under DFT remain true under IDFT, too, except that the roles of the input and output
arrays are reversed. For example, each row in Fig. 5.7 shows in the right-hand column the
result obtained by applying DFT to the discrete sequence shown in the left-hand column.
But the same figure can be also interpreted the other way around, where the left-hand
column is the result of applying IDFT to the discrete sequence shown in the right-hand
column. Now, just as in the DFT, the sampling which occurs in the input of the IDFT (the
sampling of the continuous-world spectrum G(u)) causes a periodic replication of the
output of the IDFT; but this time, the replication occurs in the signal domain. Therefore, if
the input array of the IDFT does not sample the continuous-world spectrum G(u) densely
enough (which is not the case in Fig. 5.7), the replicas of g(x) back in the signal domain
will overlap each other (additively) and cause signal-domain aliasing. Stated in other
words, if a continuous-world frequency function G(u) is sampled with sampling interval
Δu, and if g(x), the inverse CFT of G(u), does not effectively die out by the distance (or
time) R = 1/Δu (see Eq. (7.3)), then the IDFT of the sampled version of G(u) will be
signal-aliased [Thornhill83 p. 234].

The second possible point of view for considering signal-domain aliasing does not
explicitly involve the use of the IDFT. As we have seen in Sec. 2.5, the DFT operation can
be obtained from the continuous-world Fourier transform by applying three successive
steps: Sampling of the signal domain (which results in periodic replications in the spectral
domain); truncating the signal domain to a finite range of length R (which results in
convolving the spectral domain by the function R sinc(R u)); and finally sampling of the
spectral domain (which results in periodic replications in the signal domain). And indeed,
just as the sampling operation in the signal domain may result in aliasing if it is not
performed densely enough, as required by the sampling theorem, so the sampling
operation in the frequency domain may result in aliasing if it is not performed densely
enough, as required this time by the frequency sampling theorem. This last aliasing affects

Figure 5.19: When the original continuous-world function g(x) exceeds beyond the
borders of the input array, we can, by taking into account the signal-
domain aliasing (cyclical foldover in the input array of the DFT), reduce
the mismatch between the DFT spectrum and the original continuous-
world spectrum G(u). (a) When the input function g(x) = sinc(x) is
truncated at the borders of the DFT input array without taking care of the
foldover, a mismatch occurs between the DFT results and G(u) (this
 5. Issues related to aliasing 135

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(x)
1 • • • Sampled 1

0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
-1 x -1 and vert. sca. Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(a) No input foldover

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(x)
1 • • • Folded and 1
sampled
0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
-1 x -1
and vert. sca. Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b) One input foldover

Signal domain Spectral domain: Real part

1.5
0 4 8 12 16 20 24 28 31 k 1.5
0 4 8 12 16 20 24 28 31 n
g(x)=sinc(x)
1 • • • Folded and 1
sampled
0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
-1 x -1
and vert. sca. Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c) Several input foldovers

discrepancy is precisely the leakage artifact that is discussed in Chapter

6; see Fig. 6.5). (b) Allowing a single additive foldover to each side of
the input array already improves the DFT results. (c) Allowing several
foldover cycles to each side further reduces the discrepancy. Allowing
an infinite number of additive foldovers will result in a perfect match
between the DFT results and G(u).
136 5. Issues related to aliasing

the signal domain, and causes the input signal that should be forwarded to the DFT to
differ from the original continuous-world function g(x) due to the overlapping of the
signal-domain replicas. It should be noted, however, that when sampling the original
function g(x) in order to forward it to the DFT we usually neglect this signal-domain
overlapping, and rather truncate g(x) at the borders of the sampling range. This omission
results in a certain discrepancy in the resulting DFT with respect to the original
continuous-world spectrum G(u), as shown in Fig. 5.19; this discrepancy is, in fact,
precisely the leakage artifact that is discussed in Chapter 6.

In conclusion, both points of view lead to the following result, formulated here directly
for the general MD case:

Remark 5.5: (Signal-domain aliasing): If the original function g(x) exceeds beyond the
sampling range of the DFT, it should be extended cyclically rather than being truncated —
just as G(u) is treated when it exceeds beyond the range of the DFT output. In other
words, any part of g(x) that exceeds beyond the borders of the DFT input array should be
folded over and re-enter (additively) from the opposite side. Ignoring this point (which is
precisely what happens by default when g(x) undergoes standard sampling and truncation,
as we have seen in Sec. 2.5) will give imprecise results in the DFT and in its inverse DFT,
although these errors are usually rather minor. Q

Finally, another related point should be mentioned in this context. Consider row (e) of
Fig. 5.6. It is obvious that in order to get, back in the signal domain, the same discrete
sequence as the one shown in the left-hand column, we must provide as input to the IDFT
the aliased signal shown in the right-hand column by dashed lines. Failing to do so will
result in a distorted signal in the output array of the IDFT, that will not match the signal
shown in the left-hand side of the figure. In this case, however, it would be more
appropriate to use the term input-domain aliasing, since the aliasing (foldover) effect in
the input array of the IDFT occurs, in fact, in the frequency domain, and it cannot be
considered as signal-domain aliasing.

PROBLEMS

5-1. Our student from Problem 4-1 has finally obtained a correct DFT plot, after doing all
the required adjustments as discussed in Chapters 3, 4 and Appendix A (see Fig. 4.9).
In order to better see the details of this DFT plot, he decides to plot only the lower part
of the vertical scale, i.e. up to the value 0.1 (see Fig. 5.20). This allows him to realize
that his DFT plot significantly differs from the expected CFT results. For example,
instead of getting a single comb of impulses located along a rotated straight line
passing through the origin, the DFT spectrum contains impulses that are dispersed
throughout the entire spectral plane.
(a) Can you find the analytic expression of the CFT? Hint: See, for example,
[Amidror09 pp. 21–25]. You may assume that the grating’s period is 1, the width of
each black line is 40% of the period, and the rotation angle is θ = 15°.
5. Issues related to aliasing 137

(b) How do you explain the results obtained in the DFT spectrum? Can you give the
precise interpretation of each of the impulses in the DFT spectrum?
(c) Can you guess why the impulses in the DFT spectrum (after all the required
adjustments) are not sharp as in the CFT, and look rather smeared and leaked-out?
Hint: See Chapter 6.
5-2. What would you expect to see in the DFT spectrum of Fig. 5.20 if the rotation angle of
the original line grating were slightly increased or decreased?
5-3. What would you expect to see in the DFT spectrum of Fig. 5.20 if the period of the
original line grating were slightly increased or decreased?
5-4. Looking at Figs. 5.1 and 5.2, a student asks the following question: When a continuous
function g(x) is being sampled, there always exist many other continuous-world
functions that can be traced through the resulting sample points, and thus mimic the
behaviour of g(x) on its sampled values and be considered as aliases. This is true even
if the sampling frequency satisfies the requirements of the sampling theorem, like in
Figs. 5.1(a) and 5.2(a). This means that aliased signals exist in all cases, even when the
requirements of the sampling theorem are satisfied. So why do we say that aliasing
does not occur when these requirements are satisfied?
(a) How would you answer the student’s question?
(b) Suppose that g 1(x) is a continuous function that passes through the same sampling
values as g(x). Is it necessarily a lower-frequency alias of g(x)? Hint: See Example 5.1
and Footnote 8 there.
(c) What can you say about the DFT spectra of g(x) and g1(x)?
5-5. What would you expect to see if in the successive rows of Fig. 5.6 we gradually
reduced the sampling frequency fs but left the array length N unchanged?
5-6. After having studied the present chapter, consider again Problem 4-10. Can you
provide a detailed explanation like in the examples given in Sec. 5.3?
5-7. Foldover. The interpretation of the foldover effect due to aliasing is that “whatever
exceeds from one side of the DFT spectrum re-enters in a cyclical way from the
opposite side”. But as mentioned in Footnote 6 in Sec. 5.2, some references describe

0.1 v
0.05 2.
0
-0.05 1.
-0.1
-2. 0
-1.
0 -1.

1.
-2.
2.
u
Figure 5.20: The DFT plot obtained by the student in Problem 5-1.
138 5. Issues related to aliasing

the foldover effect (and even graphically depict it) as a mirroring of the frequencies
that exceed beyond 12 fs (or – 12 fs) backward from the same end of the DFT spectrum, i.e.
as folding around the spectrum edges backward (see for example the figures in
[Press02 p. 507], [Ramirez85 p. 121] or [Ramirez74 p. 102]). Explain why this
interpretation is equivalent to ours in the case of 1D real-valued signals. What happens
in the case of complex or purely imaginary 1D signals? Why should this interpretation
be avoided in 2D or MD cases? Hint: See, for example, the 2D DFT in Fig. 5.13(a).
5-8. For any 1D real-valued signal, the leftmost element of the reorganized DFT plot, which
corresponds to the lowest negative frequency provided by the DFT (and symmetrically
also to the highest positive frequency; see Footnote 4 in Sec. 4.3) must be purely real
valued.
(a) How do you explain this fact? Hint: See Sec. 8.4 and point (1) in Sec. D.2.
(b) This means that for any 1D real-valued signal having a continuous CFT, the
imaginary-valued part of the DFT spectrum must gradually go down to zero toward
both ends of the DFT spectrum, even if the CFT itself behaves differently (see, for
example, Fig. 5.8(b)). This imposes an obvious discrepancy between the CFT and the
corresponding DFT. Can this discrepancy be also explained as a result of aliasing? In
particular, can it occur when no aliasing is present?
5-9. Fig. 5.12 schematically shows what happens in the continuous-world spectrum when
sampling a band-limited function g(x,y) whose maximum frequencies to both
directions are u G and vG, respectively. Row (c) of this figure shows the sampled signal
and its continuous-world spectrum when the sampling frequencies us and vs are higher
than 2u G and 2v G , respectively, so that the replicas of G(u,v) in the spectrum do not
overlap and no aliasing occurs. Row (d) of the figure shows the situation when the
sampling frequencies u s and vs are slightly lower than 2u G and 2vG, and the replicas of
G(u,v) in the spectrum slightly overlap with their direct neighbours. While clearly
illustrating the spectral overlapping-effect of aliasing for a band-limited function
g(x,y), this figure is no longer useful when g(x,y) is not band limited and its maximum
frequencies are u G = ∞ and v G = ∞, like in the case of the infinite rotated bar g(x,y)
discussed in Example 5.6. Draw a schematic figure that clearly shows what happens in
the continuous-world spectrum when this function is sampled along both directions
with sampling rates of u s and v s, showing in particular the overlapping replicas of the
infinite line impulse G(u,v). Compare your continuous-world spectrum with the DFT
spectrum of this case shown in Fig. 5.13. Hint: See Fig. 5.21.
5-10. Suppose that you redo Fig. 5.17, but this time instead of keeping the sampling
resolution constant and reducing the radial period of the circular cosine function, you
keep the radial period of the function constant and reduce the sampling resolution.
What would you expect to get in the signal domain and in the spectral domain?
5-11. Sometimes, the aliasing phenomenon due to insufficient sampling resolution gives
unexpected, spectacular or even “artistic” results. For instance, Fig. 5.17 shows the
evolution of the shapes produced by aliasing both in the signal and in the spectral
domains, as we gradually reduce the radial period of a circular cosine function (and
hence increase the effect of aliasing). Can you explain the geometric shapes obtained
in these figures? Hint: The explanation in this case is quite easy in the spectral domain.
If you wish to understand the shapes obtained in the signal domain, you may consult,
for example, [Amidror09 Sec. 10.7.5].
5-12. Fig. 5.22 shows some other unexpected results of aliasing due to insufficient
sampling rates. Can you explain what you see in the signal and spectral domains in
each case? Hint: Compare with the non-aliased versions shown in Fig. 5.23. As for the
5. Issues related to aliasing 139

.. .
vs ..

..
–us
.. 0
..
us
u

.. ..
–vs ..

Figure 5.21: Schematic view of the continuous-world spectrum of the sampled, infinite
rotated bar discussed in Example 5.6. The original continuous bar g(x,y)
was sampled to both directions with the sampling frequencies of us and vs,
respectively, and therefore the spectrum of g(x,y), the infinite sinc-
modulated line impulse G(u,v), is replicated here about each integer
multiple (mus, nvs), m,n ∈Z. Each of the squares represents the central
part of a replica of G(u,v); note, however, that each replica contains an
infinite line impulse, and extends ad infinitum to all directions, overlapping
(additively) all the other replicas on its way. The highlighted square in the
center, belonging to the main replica, indicates the part of the continuous-
world spectrum that will be included in the 2D DFT output should we
subject our sampled signal to DFT rather than to CFT. The frequency
range of the DFT extends along each of the axes between minus half and
plus half of the corresponding sampling frequency. The DFT counterpart
of this continuous-world spectrum is shown in Fig. 5.13(a).

formation of circular or hyperbolic moiré shapes in the sampled signal in Fig. 5.22(a),
see, for example, analog discussions in Sec. 10.7.4 of [Amidror09] and Example H.4
in Appendix H of [Amidror07]. As for the foldover effect in the DFT spectrum in Fig.
5.22(b), see a similar case in Fig. 5.17. And as for the formation of a 2D periodic
structure in the sampled signal in Fig. 5.22(c), and hence the formation of an
140 5. Issues related to aliasing

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(a)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(b)
8. 4. 4.

0 0 0

-8. -4. -4.

-8. 0 8. -4. 0 4. -4. 0 4.
(c)

Figure 5.22: Sometimes aliasing due to an insufficient sampling rate may give
quite unexpected shapes. Each row shows the sampled version of a
given continuous 2D function g(x,y) and the resulting DFT. In all of
the 3 cases we have used N = 128, a sampling range of R = 16, and
a sampling interval of Δ x = 1/8. The functions being sampled
are, respectively: (a) g(x,y) = sin(2π f [x 3 – y 3 ]/8) with f = 1.
(b) g(x,y) = π d 2 somb(d x 2 + y 2 ) = d J 1 ( π d x 2 + y 2 ) / (2 x 2 + y 2 )
with d = 38.5.4 (c) g(x,y) = cos(2π f [x2 + y2]) with f = 1; note in this
last case the unexpected periodicity of the signal domain, and the
resulting impulsive nature of the DFT spectrum. Non-aliased
counterparts of these cases, which better reflect the continuous-
world functions and their expected CFTs, are shown in Fig. 5.23.
See Problem 5-12.
5. Issues related to aliasing 141

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

4. 8. 8.

0 0 0

-4. -8. -8.

-4. 0 4. -8. 0 8. -8. 0 8.
(a)
0.5 64. 64.

0 0 0

-0.5 -64. -64.

-0.5 0 0.5 -64. 0 64. -64. 0 64.

(b)
4. 8. 8.

0 0 0

-4. -8. -8.

-4. 0 4. -8. 0 8. -8. 0 8.
(c)

Figure 5.23: A more faithful counterpart of Fig. 5.22, which is obtained by using higher
sampling rates. In this figure we still use N = 128, like in Fig. 5.22, but in
rows (a) and (c) we use the sampling range R = 8 and the sampling
interval Δx = 1/16, and in row (b) we use the sampling range R = 1 and
the sampling interval Δx = 1/128. The functions being sampled are the
same as in Fig. 5.22: (a) g(x,y) = sin(2π f [x 3 – y 3 ]/8) with f = 1.
(b) g(x,y) = π4 d 2 somb(d x 2 + y 2 ) = d J 1 ( π d x 2 + y 2 ) / (2 x 2 + y 2 ) with
d = 38.5 (compare also with the second row of Fig. 6.8, after a domain
reversal). (c) g(x,y) = cos(2π f [x2 + y2]) with f = 1. Unlike in Fig. 5.22, in
the present figure the sampling frequencies being used are sufficiently high
for capturing all (or almost all) the frequencies present in the signal
domain. Hence the DFT spectrum reflects correctly the expected CFTs and
no aliasing is visible (except in row (c), where weak aliasing still exists).
142 5. Issues related to aliasing

impulsive DFT spectrum (although both g(x,y) and G(u,v) are originally continuous
and aperiodic, as shown in Fig. 5.23(c)), see Sec. 10.8 in [Amidror09].
5-13. Generate other examples involving aliasing (both 1D and 2D cases) using various
functions, and vary in each of them the degree of aliasing by gradually modifying the
sampling resolution. Observe the effects of the gradually varying aliasing, both in the
signal and in the spectral domains. Beside being fun, this should develop your intuition
as for the many different forms that aliasing may take, and allow you to easily
recognize aliasing artifacts as such, and to avoid any possible misinterpretations.
5-14. Interpretation of the DFT results. In a paper published in 1970 [Cooley70, Figures 1
and 2] the DFT approximation to the CFT of the continuous-world function g(x) =
e –x step(x) is plotted in a different way than in our Fig. 3.1 (see Chapter 3). It is
explained in the paper that the significant deviation one sees there between the DFT
results and the continuous curve of the CFT at the higher frequencies is due to the
aliasing present in the DFT, and that by taking a smaller sampling step Δx this error can
be reduced and a reasonable approximation to the CFT can be obtained for a wider
frequency range. A student claims that the significant deviation observed in the plots
there is not due to aliasing, but rather due to a misinterpretation of the unreorganized
output data. What is your opinion? Hint: You may also compare with Fig. 9.1 in
[Brigham88 pp. 168–169].
5-15. Can both the original high frequency and its low-frequency alias be perceived
simultaneously in a sampled signal? We have seen that if aliasing occurs when
sampling a continuous signal, a new lower-frequency signal appears that mimics the
true high-frequency signal (and actually takes its place). For example, looking at Fig.
5.15(d) we no longer see in the signal domain the original high-frequency 2D cosine,
but we do see instead a new low-frequency cosinusoidal structure (alias) that mimics it.
On the other hand, in Fig. 5.1(b), which explains this phenomenon, we clearly see both
the original high-frequency structure and the new low-frequency structure that mimics
it through the same sampling points. How do you explain this difference? Hint: In real-
world situations we are given the sampled signal alone, and unlike in Fig. 5.1(b), the
original continuous-world signal is not shown in the background. And indeed, if we
delete all the continuous lines from Fig. 5.1(b) and leave there only the black dots (the
sampled signal), our eye will interpret these dots as a low-frequency signal, and it will
not perceive the original high-frequency. The same result is also true in the case of
aperiodic functions, like in Fig. 5.2(b), but it is less obvious there since our eye (or
brain) has more difficulties in the interpretation of aperiodic structures (as compared to
periodic structures in which a simple pattern repeats regularly).
5-16. Interestingly, figures such as 5.17(f) offer an opportunity to see within a single figure
both aliased and non-aliased versions of the same sampled signal: Observe in the
spectral domain the angular sectors in which the impulsive ring already exceeds the
DFT borders and the sectors where it does not, and compare in the signal domain the
structures of these sectors. Does this contradict our findings in the previous problem?
5-17. Suppose you are given a random or pseudo-random figure, such as an image
consisting of randomly positioned dots. What would you expect to see in the signal and
in the spectral domains as you gradually reduce the sampling resolution of this image?
5-18. Figure 5.17(d) contains some visible distortions in its signal domain, although it is
free of aliasing (unlike Fig. 5.18(d)). Student A claims that these are printing artifacts
due to an interference with the halftone screen that is used for printing this gray level
figure (see Sec. 8.7). Student B, however, claims that these distortions are sub-Nyquist
artifacts (see Sec. 8.6). What is your opinion? How can you verify it?
 Chapter 6
Issues related to leakage

6.1 Introduction
After having discussed in Chapter 5 the aliasing artifact, we arrive now to the second
major source of discrepancy between the DFT results and the continuous Fourier
transform, the leakage artifact. Once again we start with the 1D case, and only then we
proceed to the more general multidimensional case.

6.2 Leakage in the one dimensional case

Unlike the continuous Fourier transform, the DFT can only take into account a finite
range of the original, underlying function (compare, for example, Eqs. (2.1)–(2.2) with
Eqs. (2.13)–(2.14)). This fact, however, is not as innocent as it may first seem, and it can
cause in the DFT output another type of discrepancy, known as leakage.
Leakage is a frequency-smearing artifact that may occur in the DFT spectrum due to the
finite-range truncation of the underlying signal. This truncation operation can be viewed as
a multiplication of the underlying function g(x) by a rectangular “truncation window”
w(x) whose length equals the sampling range length R = NΔx (see Eq. (4.2)), namely,
w(x) = rect(x/R).1 From the viewpoint of the signal domain, this truncation may cause a
discontinuity between the last and the first elements of the input array (because data is
considered by the DFT as cyclical), and thus introduce into the DFT spectrum new high
frequencies that do not exist in the original signal (see, for example, [Brigham88 pp. 102–
103] or [Cartwright90 pp. 212–213], and Figs. 6.3(b)–(d) below). However, as is often
the case, this phenomenon can be more easily explained from the viewpoint of the spectral
domain, still in the continuous world. By virtue of the convolution theorem, the signal-
domain multiplication of g(x) with w(x) = rect(x/R) causes in the frequency domain a
convolution of the original Fourier transform G(u) with the Fourier transform of the
truncation window w(x), which is given by (see, for example, [Brigham88 pp. 13–14]):2
W(u) = F [rect(x/R)] = R sinc(R u) (6.1)
1 This truncation can be done either before or after sampling; see Sec. D.8 in Appendix D.
2 It is assumed here, for the sake of simplicity, that the range R of the truncation window is symmetric
about the origin, namely, that the original continuous function g(x) is sampled within a symmetric range
–R/2...R/2. This allows us to assume that the Fourier transform of the truncation window is purely real
valued, as in Eq. (6.1). If the rectangular truncation window is located in a different position, its
spectrum is no longer purely real valued (see, for example, [Brigham88 pp. 15–16]); but according to
the shift theorem the magnitude (absolute value) of its spectrum remains unchanged, and only the phase
of its spectrum is affected by a linear increment. However, as mentioned in [Brigham88 p. 100], had we
considered the complex spectrum instead of Eq. (6.1), similar results would have been obtained.

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 143
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_6,
© Springer-Verlag London 2013
144 6. Issues related to leakage

Signal domain Spectral domain

g(x) G(u)

x u
–fG 0 fG
(a)

w(x) = rect(x/R) W(u) = R sinc(Ru)

R
1

x u
– R 0 R 0
2 2
(b)

g(x)w(x) G(u)*W(u)

x u
– R 0 R 0
2 2
(c)

Figure 6.1: Schematic illustration showing the effect of truncation on the spectrum
of a signal g(x) in the continuous world. (a) An original continuous
signal g(x) and its continuous-world spectrum G(u). (b) The truncating
function w(x) = rect(x/R) and its continuous Fourier transform W(u) =
R sinc(R u). (c) The truncated version of g(x) extending between –R/2
and R/2 is a product of g(x) and w(x); therefore its spectrum is a
convolution of G(u) and W(u). By comparing the spectra of the
original signal in (a) and of its truncated version in (c) we see that the
Fourier-domain effect of the truncation consists of smearing or
smoothing out the original spectrum, and adding a ripple effect (i.e.
decaying oscillations)

As shown in Fig. 6.1, this convolution may affect the spectrum in two ways with respect
to the original spectrum G(u): it may smear or smooth out the spectrum due to the main
lobe of the sinc function, and it may add a ripple effect (i.e. decaying oscillations) due to
the sidelobes of the sinc function (see, for example, [Brigham88 pp. 90–91 and 98–103]).
These effects occur in the DFT spectrum, too, since the DFT operates on the truncated
version of g(x), and thus yields discrete samples of G(u)* W(u) rather than discrete
samples of G(u). This is, indeed, the origin of the leakage artifact in the DFT output.
6. Issues related to leakage 145

Signal domain Spectral domain

g(x) = cos(2πfx)
G(u) = 12 δ(u – f) + 12 δ(u + f)
1
1
2
x
0 u
–f 0 f
(a)

w(x) = rect(x/R) W(u) = R sinc(Ru)

R
1

x u
– R 0 R 0
2 2
(b)

g(x)w(x) G(u)*W(u)

R
1 2

x u
– R 0 R –f 0 f
2 2
(c)

Figure 6.2: A particular case of the generic Fig. 6.1, showing the effect of
truncation on the spectrum of the function g(x) = cos(2πfx) in the
continuous world. (a) The function g(x) and its continuous Fourier
transform G(u), which consists of a pair of half-unit impulses at the
frequencies ±f. (b) The truncating function w(x) = rect(x/R) and its
continuous Fourier transform W(u) = R sinc(R u). (c) The truncated
version of g(x) extending between –R/2 and R/2 is a product of g(x)
and w(x); therefore its spectrum is a convolution of G(u) and W(u).
By comparing the spectra of the original function in (a) and of its
truncated version in (c) we see that the Fourier-domain effect of the
truncation consists of vertical scaling by R and smearing each of
the two impulses into a narrow sinc function that is centered about
the original impulse location, and whose oscillations gradually
decay to both directions. See also Problem 6-23.

Note that just as in the case of aliasing, if the original spectrum G(u) (and hence the
resulting discrete DFT spectrum, too) is rather smooth and non-impulsive, leakage does
not have a spectacular effect on the shape of the DFT spectrum, and its existence may be
quite difficult to detect unless the true values of the original spectrum G(u) are known in
advance. However, if the original spectrum G(u) is impulsive (i.e. if the given signal g(x) is
146 6. Issues related to leakage

periodic or almost periodic [Amidror09, Sec. B.6]) leakage is much easier to detect, since
it replaces each impulse in the DFT spectrum by the function R sinc(R u) of Eq. (6.1), so
that the originally sharp impulses may smear or leak into adjacent frequencies [Brigham88
pp. 101–103]. This is, indeed, the origin of the term “leakage” [Brigham88 pp. 103, 178;
Press02 p. 557]. This leakage effect has two main visible consequences on an impulsive
DFT spectrum (see Fig. 6.2): (a) it broadens or smears each of the spectral impulses in
accordance with the main lobe of the function R sinc(R u) (thus causing an effective
reduction in the spectral resolution); and (b) it adds a ripple effect which spreads out to
both sides of each original impulse into areas where no spectral components originally
existed, due to the sidelobes of the function R sinc(R u). A further consequence consists of
scaling the impulse height by R. See Example 6.1 below for more details.

Another case in which the leakage effect is clearly visible concerns signals g(x) whose
spectra contain sharp one-way transitions or discontinuities resembling step functions.3 In
such cases leakage may manifest itself in the DFT spectrum in the form of an overshoot to
each side of the discontinuity, that is followed by gradually decaying ripple (see Example
6.2 below and Fig. 6.5 there).4 Once again, this effect is best understood in the
continuous-world spectrum, by observing the convolution of a spectrum G(u) having a
sharp discontinuity with the sinc function R sinc(R u): The overshoots and the ripple are
simply due to the main lobe and the gradually decaying sidelobes of the sinc.5 As we can
see, this effect, too, results from the truncation of the signal-domain function g(x) by the
rectangular window rect(x/R). This overshoot-and-ripple phenomenon is, indeed, another
facet of the leakage artifact, which manifests itself this time in areas of the DFT spectrum
having sharp one-way transitions or discontinuities.6 See Example 6.2 below for a detailed
illustration of such a case.

3 This includes also a discontinuity between the last value of the DFT spectrum and its first value, due to
the cyclical nature of the DFT.
4 This phenomenon is essentially equivalent to (or more precisely, it is a Fourier dual of) the Gibbs

phenomenon [Smith07 p. 108, Footnote 7]. The Gibbs phenomenon occurs when we reconstruct a
continuous periodic function p(x) with sharp transitions, such as a rectangular wave, from a truncated
part of its Fourier series representation that does not include the higher frequencies [Bracewell86 pp.
209–211]. This gives an overshoot and ripple effect to each side of any discontinuity in p(x). And
indeed, our overshooting effect here can be seen as a discrete counterpart of the Gibbs phenomenon,
where both the signal domain and the spectral domain are discrete. An illustrated demonstration of this
phenomenon in the DFT context can be found, for example, in [Smith03 pp. 218–219]. The similarity
of our case with the Gibbs phenomenon becomes clearer if we interchange the roles of the signal domain
and of the spectral domain, so that the truncation affects the high frequencies in the spectral domain and
the overshoots and ripple occur in the signal domain, as in the classical Gibbs phenomenon. Note that
the explanation of the classical Gibbs phenomenon in the continuous world is also based on truncation
with a rectangular window in one domain, that gives convolution with a sinc function in the other
domain (see, for example, [Bracewell86 pp. 209–211]).
5 In fact, as one can see by making convolution trials between a step function and a gradually truncated

sinc function, the main lobe of the sinc function only rounds the corners of the step function. The
overshoots are obtained when we add to the main lobe of the sinc function the first sidelobe to each of
its sides; and the decaying ripple effect beyond the overshoots is obtained when we add also the higher-
order sidelobes of the sinc function (see Sec. D.7 in Appendix D and Fig. D.6 there). A similar
discussion can be found in [Oppenheim99 p. 473] and Fig. 7.23 therein.
6 As explained in [Bracewell86 p. 209], the function Rsinc(Ru) has unit area, so in places where G(u) is
slowly varying the result of the convolution will be in close agreement with G(u).
6. Issues related to leakage 147

Remark 6.1: Just like aliasing, leakage is inherent to DFT, and in practice it can rarely be
avoided completely. And yet, there exist two particular cases in which leakage does not
occur, in spite of the finite-range truncation of the underlying signal [Brigham88 pp. 98–
101 and 103–104]: (a) when the given function g(x) is periodic, and the range length R of
the truncation window (i.e. the sampling range of g(x)) is an exact integer multiple of the
period of g(x); or (b) when g(x) is a finite-duration function, and it is fully included within
the sampling range R so that it is not affected by the rectangular window rect(x/R). Note
that case (a) is the only periodic case where each of the impulses in the DFT spectrum is
perfectly sharp, meaning that (i) it falls precisely on an element of the output array, and
(ii) its width is exactly one frequency sample (a single element of the output array); see
Fig. 6.3 and compare the different rows there. And indeed, as we will see in Remark 6.3
below, this is precisely the dual formulation of case (a) in terms of the spectral domain.7,8
In all cases other than (a) and (b) leakage is unavoidable, and the only solution is,
therefore, to reduce leakage as much as possible. Q
Just as in the case of aliasing (see Remark 5.2), leakage, too, can be reduced either by a
judicious choice of the DFT parameters, or by applying an appropriate treatment to the
input signal prior to the application of the DFT; but unfortunately, the remedies here are
not the same as in the case of aliasing reduction. The first way to reduce leakage consists
of increasing the length of the rectangular truncation window (i.e. the sampling range
length R), so that its spectrum (Eq. (6.1)) becomes a narrower sinc function and better
approaches an impulse. The more closely this sinc function approximates an impulse, the
less smearing and ripple is introduced by the convolution with it [Brigham88 p. 90].9 The
other possible technique for reducing leakage (that can also be used when the range length
R cannot be increased) consists of truncating the input signal more softly, i.e. multiplying
it before the application of the DFT with a windowing function that is smoother than the
sharp-edged rectangular window rect(x/R) [Brigham88 p. 180]. Such windowing
functions have the advantage that their Fourier transforms have smaller sidelobes than
R sinc(R u), the Fourier transform of rect(x/R). A list of such windowing functions (also
called tapering functions, weighting functions or apodisation functions) can be found in
the literature, along with a full description of their respective properties, advantages and
drawbacks (see, for example, [Harris78], [Nuttall81], [Brigham88, Sec. 9.2] or [Press02
pp. 558–563]). In general the use of such windowing functions reduces the sidelobes, and

7 To see this, note that the fact that the period P of g(x) satisfies R = qP (i.e. P = R/q) for a certain
integer q means, in terms of the frequency domain, that the frequency f of g(x) satisfies f = 1/P = q/R =
qΔu (using Eq. (7.3)), i.e. that f is an exact integer multiple of the frequency step Δu, and falls exactly
on an output array element of the DFT. And vice versa, if the period P of g(x) does not exactly satisfy
R = qP for any integer q (so that there may exist a cyclical discontinuity between the beginning and the
end of the input array), it means in terms of the frequency domain that the frequency f of g(x) does not
fall exactly on an output array element of the DFT. We will return to this point in greater detail in Sec.
6.5; more information on the roles of q and p (the discrete counterpart of P) is provided in Appendix C.
8 On the dual formulation of case (b) in terms of the spectral domain, see Sec. D.4 in Appendix D.
9 In fact, we will see in Sec. 6.4 that the width of the main lobe of Rsinc(Ru) is always two DFT output
array elements, and the width of each of its sidelobes is always one array element. However, by
increasing the range length R we decrease the frequency step Δu (see Eq. (7.3)), so that although the
lobe widths remain unchanged in terms of array elements, in terms of true frequencies in the continuous-
world spectrum their bandwidths become narrower.
148 6. Issues related to leakage

hence allows to detect weaker spectral features that were hidden by the sidelobes; this,
however, at the expense of broadening the main lobe and therefore decreasing the spectral
resolution, i.e. the capacity to distinguish or resolve closely spaced peaks in the spectrum
[Brigham88 p. 180]. This tradeoff between sidelobe level and main lobe bandwidth means
that there is no single, universal best solution, and in each application one needs to find the
most suitable windowing function depending on the case.10

Remark 6.2: By combining Remark 5.2 and Remark 6.1 we see that the only case where
both aliasing and leakage are completely absent happens when the given signal g(x) is
both band limited and periodic, the sampling rate being used is at least twice the largest
frequency component of g(x), and the range length R of the truncation window (i.e. the
sampling range of g(x)) is an exact integer multiple of the period of g(x). This is, indeed,
the only case where the (reorganized) DFT and the continuous Fourier transform perfectly
coincide on the N discrete points of the DFT spectrum (within a scaling constant, as
already explained above in Sec. 4.3) [Brigham88 p. 101]. Note that case (b) of Remark
6.1 is excluded here, since a finite-duration function g(x) cannot be band limited
[Brigham88 p. 103], so that case (b) cannot be aliasing-free. Q

It is interesting to note also that fighting against leakage requires increasing the
sampling range length R of the given signal g(x), while fighting against aliasing requires,
as we have seen in Remark 5.2, a denser sampling of the signal (the use of a higher
sampling rate, i.e. a smaller sampling interval Δx). Each of these goals requires, therefore,
its own increase in the number N of samples; but if for some reason N cannot be freely
increased, one has to choose between reducing aliasing or reducing leakage. In other
words, because for a given N there is a tradeoff between the sampling resolution of g(x)
and its sampling range, there is also a tradeoff between aliasing and leakage.

More details on the optimal choice of the sampling range length R (and the tradeoffs that
are involved in this choice) can be found in Sec. 7.4.

Interested readers can find many references that treat the leakage phenomenon from
various different points of view. For example, one may mention [Smith03 pp. 174–176],
[Smith07 pp. 105–108], [Ramirez85 pp. 102–109], [Brigham88 pp. 98–107, 176–179],
[Cartwright90 pp. 206–208] and [Briggs95 pp. 98–99].

6.3 Examples of leakage in the one dimensional case

Let us illustrate the above discussion by two 1D examples, one featuring a periodic
function g(x) (whose spectrum G(u) is therefore purely impulsive), and the other featuring
an aperiodic function g(x) (whose spectrum G(u) is continuous).

10 Note that in some references such as [Weisstein99 pp. 55–56] this tapering process is called
apodization, literally meaning “foot removal”. The “foot” refers here to the trailing sequence of
sidelobes; the origin of this term is more obvious in the 2D case (see Sec. 6.6 below).
6. Issues related to leakage 149

Example 6.1: Leakage due to truncation in the periodic function g(x) = cos(2πfx):
Fig. 6.2 shows a particular case of the generic Fig. 6.1 in which the original continuous
function is g(x) = cos(2πfx), whose CFT consists of a pair of half-unit impulses at the
frequencies ±f (see part (a) of the figure): G(u) = 12 δ(u–f) + 12 δ(u+f) [Brigham88 pp. 19–
20]. The truncation of g(x) into the finite range –R/2...R/2 can be seen as a multiplication
of g(x) by the truncating window w(x) = rect(x/R), a square pulse of length R and height 1
which extends between –R/2 and R/2. The truncating window w(x) and its CFT W(u) =
R sinc(R u) are shown in part (b) of the figure, and the resulting truncated function
g(x)w(x) and its spectrum are shown in part (c) of the figure. By virtue of the convolution
theorem, the spectrum (CFT) of the truncated function g(x)w(x) is the convolution of the
spectra G(u) and W(u), which is given by (see [Bracewell86 p. 416]):
G(u)*W(u) = 12 R sinc(R(u – f)) + 12 R sinc(R(u + f)) (6.2)
As we can see in part (c) of the figure, the effect of the signal-domain truncation on the
spectral domain consists of vertically scaling each of the two impulses by R, and smearing
it into a narrow sinc function that is centered about the original impulse location, and
whose oscillations gradually decay to both directions.
So far we have only discussed the continuous Fourier transform, so that the spectrum of
the truncated function g(x)w(x) is continuous (see Fig. 6.2(c)). Assume now that instead
of considering g(x)w(x) and its CFT, we sample g(x)w(x) and subject it to DFT (note that
the truncating function w(x) corresponds here to the finite range of g(x) that is taken into
account by the DFT). As we have seen in Sec. 5.2, the use of DFT has two main
consequences: (1) The resulting spectrum is now discrete; and (2) it only contains
frequencies in the range between – 12 fs and 12 fs (where fs is the sampling frequency), and
includes therefore only frequencies located in the main replica of the periodic spectrum of
the continuous Fourier transform.

Figure 6.3: Illustration of the leakage artifact in the case of the periodic signal
g(x) = cos(2π fx). The figure shows the sampled version of the
signal g(x), after its truncation into the range –4...4 that is included
in the input array, and the resulting DFT (after all the required
adjustments as discussed in Chapters 3 and 4 and in Appendix A).
For the sake of comparison, the original continuous function g(x)
and its CFT G(u) = 12 δ(u–f) + 12 δ(u+f) are also shown, drawn by
continuous lines. The successive rows of the figure show the
situation for a few gradually increasing values of the cosine
frequency f, varying between f = 0.5 in row (a) and f = 0.75 in row
(f). Note that the sampling frequency fs = 4 and the array length
N = 32 remain unchanged in all cases (they have been chosen
sufficiently high so as to guarantee that no aliasing artifacts occur
here). Similarly, the signal domain range length R = 8, the signal
domain step Δ x = 14 and the spectral domain step Δ u = 18 remain
unchanged, too. In rows (a), (e) and (f) the cosine frequency f is
 150 6. Issues related to leakage

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k 1.5 n
2
g(x)=cos(2πfx) CFT
1.5 • • • Sampled 1 • • • DFT after
1 i&o reorg.
0.5 and vert. sca.
0.5
0 0
-0.5
-0.5
-1
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(a) f = 1/2 = 4Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k 1.5 n
2
g(x)=cos(2πfx) CFT
1.5 • • • Sampled 1 • • • DFT after
1 i&o reorg.
0.5 and vert. sca.
0.5
0 0
-0.5 CFT of
-0.5
-1 g(x)w(x)
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(b) f = 17/32 = 4.25Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 k 1.5
0 4 8 12 16 20 24 28 31 n
2 g(x)=cos(2πfx) CFT
1.5 • • • Sampled 1 • • • DFT after
1 i&o reorg.
0.5 and vert. sca.
0.5
0 0
-0.5 CFT of
-0.5
-1 g(x)w(x)
-1.5 x -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c) f = 18/32 = 4.5Δu

exactly an integer multiple of the frequency step of the DFT

spectrum, Δu, so that the two impulses are represented in the DFT
results correctly. But in rows (b)–(d) the cosine frequency f falls in-
between elements of the DFT spectrum, resulting in a significant
leakage effect: each of the two impulses is smeared between its two
neighbouring output array elements, and also shows gradually
 6. Issues related to leakage 151

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k 1.5 n
2
g(x)=cos(2πfx) CFT
1.5 • • • Sampled 1 • • • DFT after
1 i&o reorg.
0.5 and vert. sca.
0.5
0 0
-0.5 CFT of
-0.5
-1 g(x)w(x)
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(d) f = 19/32 = 4.75Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k 1.5 n
2
g(x)=cos(2πfx) CFT
1.5 • • • Sampled 1 • • • DFT after
1 i&o reorg.
0.5 and vert. sca.
0.5
0 0
-0.5
-0.5
-1
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(e) f = 20/32 = 5Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 k 1.5
0 4 8 12 16 20 24 28 31 n
2 g(x)=cos(2πfx)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5
-0.5 • • • DFT after
-1 i&o reorg.
-1.5 x -1
and vert. sca. Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(f) f = 3/4 = 6Δu

decaying oscillations to both directions. The dashed lines in the

spectra of rows (b)–(d) correspond to the CFT of g(x)w(x), i.e. to
the continuous function G(u)*W(u) (Eq. (6.2)). Note the perfect
agreement between these dashed lines and the black dots that
represent the DFT results (after the adjustments according to
Chapters 3 and 4 and Appendix A). See Example 6.1.
152 6. Issues related to leakage

Fig. 6.3 shows the sampled version of our signal g(x) = cos(2πfx) after its truncation to
the range –4...4 that is covered by the N sampling points of the DFT input array, as well as
the resulting DFT spectrum (after all the required adjustments discussed in Chapters 3
and 4 and in Appendix A). The different rows of the figure show the sampled signal and
its DFT for several slowly increasing values of the cosine frequency f, going from f = 0.5
in row (a) to f = 0.75 in row (f). Note that the sampling frequency fs = 4 as well as the
array length N = 32 are kept identical in all of the rows; we have chosen them to be
sufficiently high so as to guarantee that no aliasing artifacts occur (note that our starting
point here, shown in Fig. 6.3(a), is identical to Fig. 5.10(b), with f = 0.5 rather than f = 1).
The signal domain range length R = 8, the signal domain step Δx = 14 and the spectral
domain step Δu = 18 , too, are identical in all the rows. As we can see in rows (a), (e) and (f),
whenever the cosine frequency f falls exactly on an integer multiple of the frequency step
Δu, each of the cosine impulses falls exactly on an element of the DFT output array and is
represented there correctly (after the required adjustments as discussed in Chapters 3 and
4 and in Appendix A). The dual signal-domain condition in such cases is simply that the
sampled cosine function has exactly an integer number of full periods within the range
included by the input array, R = NΔx (see the left-hand side of rows (a), (e) and (f)). What
happens now when the frequency f of our original cosine function is not exactly an integer
multiple of the frequency step Δu? In the signal domain this condition means that the
truncated part of our original signal g(x) = cos(2πfx), i.e. the finite range of g(x) that is
included in the input array of the DFT, does not cover an integer number of full periods of
g(x) (see the left-hand side of rows (b)–(d) in our figure). Although the input array of the
DFT remains periodic, as always, the resulting periodic signal (i.e. the periodic repetition
of the truncated signal) significantly differs from our original periodic signal g(x), due to
the discontinuities between both ends of the input array. Therefore, its DFT cannot be
equivalent to the CFT of g(x), and we would expect it to have additional frequency
components due to these discontinuities. And indeed, the spectral domain counterpart of
our condition is that neither of the two impulses having the frequencies ±f is precisely
represented by a single element of the DFT spectrum; instead, each of them is smeared
between the two closest frequency elements of the DFT spectrum, and also shows a ripple
effect which gradually decays to both directions. As we have seen above, this effect results
from the truncation that our original function g(x) = cos(2πfx) undergoes in the signal
domain: The truncation into the range –R/2...R/2 that is taken into account by the DFT
input array is equivalent in the spectral domain to a convolution by the function W(u) =
R sinc(R u), meaning that the function being represented by the DFT results is not G(u),
but rather the convolution G(u)*W(u) as given in Eq. (6.2) above. And indeed, looking at
the DFT spectra in rows (b)–(d) of the figure we note the perfect agreement between the
DFT results represented by the black dots and the continuous function G(u)*W(u) which
is plotted there by dashed lines. This is, indeed, the explanation of the leakage artifact in
the DFT spectrum of our given function g(x) = cos(2πfx).
So why such a smearing effect does not occur in rows (a), (e) and (f), where the cosine
frequency f is exactly an integer multiple of the frequency step Δu (and the input array of
the DFT contains exactly an integer number of full periods of g(x) = cos(2πfx))? The
6. Issues related to leakage 153

answer is that G(u) is indeed convolved with W(u) = R sinc(R u) in all of the cases, because
the same truncation is operated in all of them. However, when the cosine frequency f is
exactly an integer multiple of the frequency step Δu the functions W(u±f) = R sinc(R(u±f))
have their peaks exactly at an integer multiple of Δu, where the true impulse is located, and
all the other points of the DFT spectrum, i.e. the other integer multiples of Δ u, fall
precisely on a zero point of these sinc functions. Clearly, in such cases the convolution of
the spectral impulses with the narrow sinc function W(u) = R sinc(R u) has no visible effect
on the DFT output (except for an impulse height scaling by R, which can be corrected as
explained in Appendix A). This is clearly shown in Fig. 6.4. More details on the precise
effects of the truncation and on impulses that fall in-between elements of the DFT
spectrum are given in Secs. 6.4 and 6.5 below (see also references such as [Brigham88
pp. 98–103]). Q

Example 6.2: Leakage due to truncation in the aperiodic function g(x) = sinc(2ax):
Let us now consider the aperiodic function g(x) = sinc(2ax), whose CFT is G(u) =
1
2arect(2au ), i.e. a square pulse of height 2a1 extending between –a and a. Once again, the
truncation of g(x) into a finite range –R/2...R/2 can be seen as a multiplication of g(x) by
the truncating window w(x) = rect(x/R), a square pulse of length R and height 1 which
extends between –R/2 and R/2. By virtue of the convolution theorem, the spectrum (CFT)
of the truncated function g(x)w(x) is the convolution of the spectra G(u) and W(u) =
R sinc(R u), which is given by (see Sec. D.5 in Appendix D):
G(u)*W(u) = 1
2aπ
[Si(πR(u+a)) – Si(πR(u–a))] (6.3)
where Si(u) is the sine integral function. This convolution with W(u) means that the effect
of the signal-domain truncation on the spectral domain consists of smoothing out the
original square pulse G(u); moreover, the nature of the sine integral function implies the
addition of an overshoot to each side of the two sharp transitions of G(u), followed by
gradually decreasing oscillations to both directions.
So far we have only discussed the continuous Fourier transform, so that the spectrum of
the truncated function g(x)w(x) is continuous. Assume now that instead of considering
g(x)w(x) and its CFT, we sample g(x)w(x) and subject it to DFT (note that the truncating
function w(x) corresponds here to the finite range of g(x) that is taken into account by the
DFT). Fig. 6.5 shows the sampled version of our signal g(x) = sinc(2ax) after its
truncation to the range –4...4 that is covered by the N sampling points of the DFT input
array, as well as the resulting DFT spectrum (after all the required adjustments discussed
in Chapters 3 and 4). The different rows of the figure show the sampled signal and its
DFT for several slowly increasing values of the parameter a, going from a = 0.5 in row (a)
to a = 0.75 in row (f). Note that as in the previous example, the sampling frequency fs = 4
as well as the array length N = 32 are kept identical in all of the rows; we have chosen
them to be sufficiently high so as to guarantee that no aliasing artifacts occur (note that
our starting point with a = 0.5, shown in Fig. 6.5(a), is identical to Fig. 4.1 in Chapter 4).
Similarly, the signal domain range length R = 8, the signal domain step Δx = 14 and the
spectral domain step Δu = 18 , too, remain identical in all the rows.
154 6. Issues related to leakage

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 31
1.0

0.5

-0.5
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(a)

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 31
1.0

0.5

-0.5
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(b)
Figure 6.4: A magnified view of the DFT spectra in rows (a) and (c) of Fig. 6.3 (after all
the required adjustments). In each case the continuous black line represents
the original continuous spectrum G(u), consisting of two half-unit impulses,
and the dashed line represents the continuous function G(u)*W(u) of Eq.
(6.2); note the perfect agreement between the dashed line and the DFT
results represented by the black dots. (a) When the cosine frequency is
exactly a multiple of the spectral domain step Δu (here: f = 4Δu), each of the
two spectral impulses falls exactly on an element of the DFT spectrum, and
so does the corresponding peak of G(u)*W(u); moreover, all the zero points
of G(u)*W(u) fall exactly on the successive elements of the DFT spectrum
(i.e. on all the other integer multiples of Δu), so that W(u) does not affect the
DFT results (up to vertical scaling). (b) When the cosine frequency is not
exactly a multiple of Δu (here: f = 4.5Δu), each spectral impulse falls in-
between frequency elements of the DFT spectrum, and “leaks” into its two
closest neighbours, and into gradually decaying oscillations to both sides.
6. Issues related to leakage 155

Let us now compare in Fig. 6.5 the DFT results with the original spectrum G(u), which
is plotted with a continuous line. As we can see, the DFT smoothes out each of the sharp
transitions of G(u), and also adds an overshoot to each side of the transition, followed by
gradually decaying oscillations to both directions, much like the Gibbs phenomenon (see
Footnote 4 in Sec. 6.2). This effect results, as we have seen above, from the truncation that
our original function g(x) = sinc(2ax) undergoes in the signal domain: In order to be
forwarded to the DFT, g(x) must be first sampled and truncated into the range –R/2...R/2
that is taken into account by the DFT input array. This truncation is equivalent in the
spectral domain to a convolution by the function W(u) = R sinc(R u), meaning that the
function being represented by the DFT results is not G(u), but rather the convolution
G(u)*W(u) given in Eq. (6.3) above. And indeed, looking at the DFT spectrum in each
row of the figure we note the perfect agreement between the DFT results represented by
the black dots and the continuous function G(u)*W(u) which is plotted there by dashed
lines. This is, indeed, the explanation of the leakage artifact in the DFT spectrum of our
given function g(x) = sinc(2ax).
Note that unlike in Fig. 6.3, in the present case all the rows of the figure suffer from a
significant leakage effect, both when the pulse borders ±a fall exactly on an integer
multiple of the frequency step Δu and when they fall in-between elements of the DFT
spectrum. From the spectral domain point of view, the reason is that unlike the function
G(u)*W(u) of the previous example (Eq. (6.2)), the function G(u)*W(u) in the present
case (Eq. (6.3)) does not have zeroes at the integer multiples of Δu, so that even when the
sharp transitions of G(u) are centered exactly on a frequency that exists in the DFT output
we still get decaying oscillations throughout the spectrum. From the signal-domain point
of view, the explanation is based on the fact that our given signal g(x) = sinc(2ax) is
aperiodic: Since the input array of the DFT is periodic, as always, the resulting signal (i.e.
the periodic repetition of the truncated signal) significantly differs from our original
aperiodic signal g(x). Therefore, we cannot expect its DFT to be equivalent to the CFT of
g(x), and this, independently of whether or not the pulse borders ±a fall exactly on an
integer multiple of the frequency step Δu. Q

Figure 6.5: Illustration of the leakage artifact in the case of the aperiodic signal
g(x) = sinc(2ax). The figure shows the sampled version of the
signal g(x), after its truncation into the range –4...4 that is included
in the input array, and the resulting DFT (after all the required
adjustments as discussed in Chapters 3 and 4). For the sake of
comparison, the original continuous function g(x) and its CFT G(u)
= 2a1 rect(2au ) (a square pulse of height 2a1 extending between –a and a)
are also shown, drawn by continuous lines. The successive rows of
the figure show the situation for a few gradually increasing values
of the parameter a, varying between a = 0.5 in row (a) and a = 0.75
in row (f). Note that the sampling frequency fs = 4 and the array
length N = 32 remain unchanged in all cases (they have been chosen
 156 6. Issues related to leakage

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0

• • • DFT after CFT of

-0.5 -0.5
i&o reorg. g(x)w(x)
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(a) a = 1/2 = 4Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0

• • • DFT after CFT of

-0.5 -0.5
i&o reorg. g(x)w(x)
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(b) a = 17/32 = 4.25Δu

Signal domain Spectral domain: Real part

1.5
0 4 8 12 16 20 24 28 31 k 1.5
0 4 8 12 16 20 24 28 31 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0

• • • DFT after CFT of

-0.5 -0.5
i&o reorg. g(x)w(x)
and vert. sca.
-1 x -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c) a = 18/32 = 4.5Δu

sufficiently high so as to guarantee that no aliasing artifacts

occur here). Similarly, the signal domain range length R = 8,
the signal domain step Δx = 14 and the spectral domain step
Δu = 18 remain unchanged, too. The dashed lines in the spectra
 6. Issues related to leakage 157

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0

• • • DFT after CFT of

-0.5 -0.5
i&o reorg. g(x)w(x)
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(d) a = 19/32 = 4.75Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0

• • • DFT after CFT of

-0.5 -0.5
i&o reorg. g(x)w(x)
and vert. sca.
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(e) a = 20/32 = 5Δu

Signal domain Spectral domain: Real part

1.5
0 4 8 12 16 20 24 28 31 k 1.5
0 4 8 12 16 20 24 28 31 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0

• • • DFT after CFT of

-0.5 -0.5
i&o reorg. g(x)w(x)
and vert. sca.
-1 x -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(f) a = 3/4 = 6Δu

correspond to the CFT of g(x)w(x), i.e. to the continuous function

G(u)*W(u) (Eq. (6.3)). Note the perfect agreement between these
dashed lines and the black dots that represent the DFT results (after
the adjustments according to Chapters 3 and 4). See Example 6.2.
158 6. Issues related to leakage

0 2 4 6 8 10 12 14 16 17 20 22 24 26 28 30 31

-2
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.

Figure 6.6: The continuous function W(u) = R sinc(R u), drawn here for R = 8,
and its sampled version with N = 32.

In the two following sections we treat in greater detail some topics that are related to the
leakage phenomenon; then, in Sec. 6.6 we proceed to the multidimensional case.

6.4 Errors due to signal-domain truncation

As we have seen above, leakage occurs due to the truncation of the original signal g(x)
by the rectangular truncation window w(x) = rect(x/R) whose length equals the entire
sampling range of the given signal, R = NΔx. This multiplication in the signal domain
implies in the spectral domain a convolution of the original Fourier transform G(u) with
the Fourier transform of rect(x/R), which is given by W(u) = R sinc(R u) (Eq. (6.1)). Let us
now study more closely the properties of this sinc function (see Fig. 6.6).

First of all, we note that R sinc(R u) is a narrow sinc function, whose main lobe extends
between –1/R and 1/R with maximum amplitude of R at the origin, and whose k-th sidelobe
extends between k/R and (k+1)/R for k = 1, 2, ... (or symmetrically between k/R and
(k–1)/R for k = –1, –2, ... in the negative side). The main lobe and the sidelobes are
delimited by the zero-crossing points of this function, which occur at the points k/R for all
integer values of k except for 0 (the origin). We see, therefore, that the width of the main
lobe of this function (between the two zero-crossing points –1/R and 1/R) is exactly 2/R.
What does this mean in terms of the DFT output array elements? Knowing that the entire
frequency range length F of the DFT spectrum extends over N elements, it follows (e.g.
using the rule of three [Harris98 p. 15]) that a frequency span of 2/R extends over:
6. Issues related to leakage 159

w = 2N output array elements

RF
and using the relation RF = N (see Eq. (7.2)) we obtain:

w = 2 output array elements (6.4)

Similarly, because the width of each sidelobe of the function R sinc(R u) (i.e. the distance
between its two bordering zero-crossing points) is 1/R, half of the width of the main lobe,
it follows that the width of each sidelobe corresponds exactly to one output array element.

We see, therefore, that the main lobe and the sidelobes of the function R sinc(R u), which
are the source of the leakage effect in the DFT spectrum, always have the width of 2 or 1
array elements, respectively, independently of the choice of the DFT parameters R, N, Δx,
etc. And indeed, this is clearly visible in detailed figures illustrating the leakage of spectral
impulses, like Fig. 6.4. This means that the smearing of a spectral impulse due to the main
lobe of the function R sinc(R u) with which the impulse is being convolved may extend to
two array elements.11 Similarly, each of the sidelobes being added to either side of the
spectral impulse extends over exactly one array element. Nevertheless, the global effect of
leakage extends over a multitude of consecutive sidelobes with gradually decreasing
amplitudes. This gives the typical impression of a leaking or melting-down impulse with a
decaying “tail” that trails away to each side of the smeared impulse. Note that the
individual one-element sidelobes of each trailing tail will have alternating positive and
negative amplitudes. But if we plot the magnitude (i.e. the absolute value) of this DFT
spectrum (or its squared version, the power spectrum), we will get a continuously decaying
positive-valued tail to each side of the smeared impulse (compare Figs. 6.14 and 6.16).

It should be noted, however, that this analysis of the lobe widths due to the leakage
phenomenon is basically valid for cases with impulsive spectra, where each spectral
impulse will be individually smeared by R sinc(R u) (compare Figs. 6.4 and 6.6). When the
original spectrum G(u) is not impulsive, its convolution with R sinc(R u) will still give a
smearing effect with visible sidelobes, but the widths of these lobes may be different. As
we have seen in Example 6.2 and in Fig. 6.5, this is the case when G(u) is a rectangular
function; in this case the convolution of G(u) with R sinc(R u) gives a new function (see
Eq. (6.3)) whose lobe widths are no longer exact multiples of output array elements.
Although the typical behaviour of leakage still remains the same (smearing and ripple), the
precise lobe-width analysis given above is not necessarily preserved.

Thus far, we have described here in greater detail the leakage artifact of the DFT. What
happens, however, when the signal-domain function g(x) is truncated by a rectangular
window whose length is less than N elements (i.e. less than the entire span R of
the sampling range), so that all the remaining input array elements are set to zero?
Suppose, for example, that we use a truncation window w1(x) whose length is N/4 (or

11Note that this smearing causes an effective 2-fold reduction in the spectral resolution, although the
1
theoretical spectral resolution of the DFT spectrum, given by the frequency step Δu = NΔx of Eq. (4.4)
remains unchanged [Cartwright90 pp. 207–208].
 160 6. Issues related to leakage

Signal domain
0 8 16 24 32 40 48 56 63
1.5 k
g(x) = e-xstep(x)
1
• • • Sampled
0.5

-0.5

4. x
-1
-4. -3. -2. -1. 0 1. 2. 3.

Spectral domain: Real part Spectral domain: Imaginary part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
1.5 n 1.5 n
CFT CFT
1 1
• • • DFT after • • • DFT after
i&o reorg. i&o reorg.
0.5 and vert. sca. 0.5 and vert. sca.
0 0

-0.5 -0.5

4. Hz 4. Hz
-1 -1
-4. -3. -2. -1. 0 1. 2. 3. -4. -3. -2. -1. 0 1. 2. 3.

(a)

Figure 6.7: The effect of excessive signal-domain truncation on the DFT spectrum.
(a) The original continuous function g(x) = e–x step(x) and its CFT
G(u) = 1/(1+[2πu]2) – 2πui/(1+[2πu]2) [Bracewell86 p. 418], plotted
by continuous lines, and their DFT counterparts, plotted by black dots.

equivalently R/4, assuming that the sampling interval Δx remains unchanged; see Eq.
(4.2)). In this case the convolution in the spectral domain will be done with the sinc
function W1(u) = R/4 sinc(u R/4), whose lobes are 4 times wider than the lobes of the sinc
function R sinc(R u) that generates the leakage phenomenon. This will cause in the DFT
spectrum a 4-fold wider smearing: Each impulse will be smeared by the main lobe of the
new sinc function into 8 output array elements about its true location, and moreover, each
individual sidelobe will be 4 elements wide, thus causing a visible, wavy ripple effect. Such
a ripple effect in the DFT spectrum due to an excessive signal-domain truncation may also
occur when g(x) is not impulsive (see, for example, Fig. 6.7), but then the lobe widths may
be somewhat different.12

12 Because the truncation of the given function g(x) causes a sharp discontinuity in the signal domain,
the addition of a large spectral range of high frequencies in its DFT spectrum is not really surprising
[Smith07 p. 108].
 6. Issues related to leakage 161

Signal domain
0 6 12 19 25 32 38 44 51 57 63
1.5 k
g(x) = e-xstep(x)
1
• • • Truncated &
sampled
0.5

-0.5

4. x
-1
-4. -3. -2. -1. 0 1. 2. 3.

Spectral domain: Real part Spectral domain: Imaginary part

0 6 12 19 25 32 38 44 51 57 63 0 6 12 19 25 32 38 44 51 57 63
1.5 n 1.5 n
CFT CFT
1 1
• • • DFT of • • • DFT of
truncated & truncated &
0.5 sampled 0.5 sampled
0 0

-0.5 -0.5

4. Hz 4. Hz
-1 -1
-4. -2. 0 2. -4. -2. 0 2.

(b)

(b) The same original function g(x) and its CFT G(u), plotted by
continuous lines, but this time the black dots represent the truncated
function g(x)w1(x) and its DFT.

It should be stressed, however, that although the explanation of this phenomenon is

similar to that of leakage, there exists a fundamental difference between the two: The
N-element windowing effect that causes leakage is inherent to DFT, and it cannot be
avoided. But the windowing effect due to a truncation of the given N-element signal within
its sampling range (i.e. using a truncation window of less than N elements) can be avoided
— and in fact, such a truncation should never be done unless its consequences in terms of
the DFT spectrum are intended and well understood.13

13 Note that this truncation in the signal domain is in fact the Fourier dual of the low-pass filtering
of g(x) by truncating its DFT spectrum and taking the inverse DFT of the result. As shown in Fig. 3.7
in Chapter 3, low-pass filtering gives, indeed, a “leaked out” (i.e. low-pass filtered) version of g(x); the
difference between the two cases is, however, that the roles of the signal and spectral domains are
interchanged.
162 6. Issues related to leakage

6.5 Spectral impulses that fall between output array elements

A related question may be asked now, to which we have already partially replied in
Example 6.1: What happens to an impulse δ(u–a) of the original continuous spectrum
G(u) when it does not fall in the discrete world exactly on an array element n of the DFT
output, but rather between two elements?

Let us formulate this question more precisely. As we already know (see Sec. 4.3), each
of the N elements in the output array of the DFT corresponds to a continuous-world
frequency of nΔu Hz for a certain integer n in the range –N/2,...,N/2–1, where Δu is the
frequency step of the discrete spectrum. The DFT spectrum is therefore defined only for
these discrete frequencies, namely, nΔu = n 1 = nfs /N Hz. If we apply DFT to a periodic
NΔx
signal g(x) whose period P is exactly the sampling range length R = NΔx or some integer
divisor of R, meaning that P = R/q for a certain integer q, 1 ≤ q < N/2, then the frequency
of g(x), f = 1/P = q/R = qΔu (see Eq. (7.3)), as well as its harmonics, fall indeed exactly on
elements of the DFT output array. What happens, however, when we apply DFT to a
periodic signal whose period is not exactly an integer divisor of R, so that its frequency f is
not exactly an integer multiple of the discrete frequency step Δu?

For example, consider the continuous input function g(x) = cos(2πfx) of Example 6.1,
whose continuous Fourier transform consists of a pair of half-unit impulses at the
frequencies ±f: G(u) = 12 δ(u–f) + 12 δ(u+f). As we already know, in order to apply DFT to
the function g(x) we first need to sample it (say, at a sufficiently high sampling rate so as
to avoid aliasing), and to truncate it by the rectangular window w(x) = rect(x/R) where R is
the chosen sampling range length. This signal-domain truncation gives in the spectral
domain a convolution of the original two-impulse spectrum G(u) with the sinc function
W(u) = R sinc(R u), which is the Fourier transform of w(x). In other words, the DFT of the
original function consists of samples of G(u)*W(u) rather than samples of G(u). We have
seen above that the sinc function R sinc(R u) consists of a narrow main lobe extending
between the two zero-crossing points –1/R and 1/R, and of sidelobes which extend to both
directions between all the following zero-crossing points k/R and (k+1)/R for k = 1, 2, ...
(or in the negative side between k/R and (k–1)/R for k = –1, –2, ...). In terms of output
array elements, the main lobe of this sinc function extends between the integer points –1
and 1, and its sidelobes extend to both directions between all the integer pairs k and k+1
for k = 1, 2, ... (or, respectively, between k and k–1 for k = –1, –2, ...).

Now, if the frequency f of our given cosine function is exactly an integer multiple of the
frequency step Δu = 1 = 1/R (see Eqs. (4.4) and (4.2)), so that the two impulses of
NΔx
G(u) fall exactly on an output array element, the convolution of G(u) with R sinc(R u),
which places a replica of R sinc(R u) about each of the two impulses, will cause no errors in
the DFT spectrum with respect to G(u) (except for multiplying the impulse amplitude by
R; see Sec. A.3 in Appendix A). The reason is that all the other output array elements will
fall exactly on the zero-crossing points of the replicas of R sinc(R u) (see Fig. 6.4(a)). The
resulting DFT spectrum will therefore contain exactly two sharp impulses, as shown in
this figure.
6. Issues related to leakage 163

If, however, the frequency f of our given cosine function is not an integer multiple of
Δu = 1 = 1/R, so that the two impulses of G(u) fall between output array elements, the
NΔx
convolution of each of these impulses with R sinc(R u) will cause in the DFT spectrum a
leakage effect, as shown in Fig. 6.4(b): Since the two-element wide main lobe of the sinc
is now located between two output array elements, both of these array elements will fall
inside the main lobe of the sinc function and have positive values. And moreover, each of
the other output array elements to the left and to the right will fall inside one of the one-
element wide sidelobes of the sinc replica, and hence they will have alternating negative
and positive non-zero values. The resulting DFT spectrum will therefore have additional
non-zero values at frequencies other than ±f: It will include two smeared (two-element
wide) impulses located at the output array elements nearest to the true cosine frequencies f
and –f, each of them having long “tails” which trail away to both directions throughout
the entire range of the DFT spectrum, as shown in Fig. 6.4(b). Note that the effect of
leakage is quantifiable; a mathematical expression giving the exact leakage value generated
by the original frequency f at each of the output array elements can be found, for example,
in [Smith07 pp. 105–106] or in [Cartwright90 p. 206]. Furthermore, it is shown in
[Cartwright90 p. 208] that the power (i.e. the square of the amplitude) of the original
frequency f is the sum of the powers of the output array elements among which it is split
(this is not true, however, for the amplitudes). Leakage can be seen, therefore, as taking
power from the frequency component f existing in the original continuous signal g(x), and
transferring it to other frequency components of the DFT output array, which do not exist
in the original continuous signal [Ramirez85 p. 107].

Leakage, when it occurs, affects in the general case both the real and the imaginary parts
of the DFT spectrum. Also, when converting the DFT spectrum to the polar form, leakage
is carried through the conversion to affect both magnitude and phase [Ramirez85 p. 107].

It is important to stress, however, that in spite of this significant discrepancy with respect
to the original spectrum G(u), the DFT spectrum is not wrong in itself — it is exactly what
it should be for the given discrete signal [Ramirez85 p. 107]; simply, it contains an
inherent leakage effect that did not exist in the spectrum G(u) in the continuous case. In
other words, from the viewpoint of the continuous case leakage is an “error”; but from
the viewpoint of the N samples provided to the DFT the result is correct.

Although we used in this discussion the specific example of a cosine function, the same
leakage effect occurs in the DFT spectrum for any spectral impulse that falls between
output array elements, and more generally, for any frequency component f (impulsive or
not) that does not fall exactly on an output array element. Note, in particular, that if a
general periodic signal g(x) is acquired so that the sampling range R contains an integer
number of periods, then each of the harmonics, too, has an integer number of periods in R.
Thus, each of the spectral components of g(x) will fall exactly on an output array element
of the DFT, rather than between elements. But if g(x) is an almost-periodic function (see,
for example, Appendix B in [Amidror09]) which is made up of discrete frequencies that
are not harmonically related (such as f and 2f), some frequency components may exhibit
164 6. Issues related to leakage

leakage while other don’t; such signals are often found in vibration studies where the
signal is the result of several vibration sources [Ramirez85 p. 109].
As we can see, the answer to the question in the beginning of this section can be
formulated as follows:

Remark 6.3: The fact that the frequency f of a given periodic function g(x) falls exactly
on an output array element is equivalent, in terms of the signal domain, to the fact that the
period P of the function g(x) precisely satisfies R = qP for a certain integer q, so that g(x)
satisfies the leakage-free condition (a) of Remark 6.1. Indeed, if the frequency f of g(x) is
an integer multiple of Δu, we have f = qΔu = q/R (using Eq. (7.3) below), so that the period
P = 1/f of g(x) satisfies P = R/q for a certain integer q; this is precisely condition (a) of
Remark 6.1. And vice versa, if the frequency f of g(x) does not fall exactly on an output
array element, this means in terms of the signal domain that the period P of g(x) does not
exactly satisfy R = qP for any integer q, meaning that leakage will occur in the DFT
spectrum. More details on the role of q can be found in Appendix C. Q
We conclude, therefore, with the following result (see also Remark C.3 in Appendix C):

Remark 6.4: A periodic function g(x) is perfectly cyclical (seamlessly wraparound)

within the input array if and only if its impulses in the DFT spectrum fall exactly on output
array elements (which means, in turn, that no leakage occurs in the DFT spectrum). Q
We see, as a consequence, that if g(x) is a periodic function which suffers from leakage
in the DFT spectrum since q is non-integer, then there are two equivalent ways to rectify
the problem: either by resampling g(x) within a new range R = qP where q is integer; or by
modifying N to be an integer multiple of p (the period of g(x) in terms of input array
elements), so that by virtue of pq = N (Eq. (C.2)) q becomes integer.14

In real life, where signals are usually much more complex than simple cosines and
contain many more components of different frequencies, it is extremely unlikely that a
smooth match occurs between the beginning and the end of the sampled signal, or that all
the frequency components fall precisely on output array elements of the DFT. Spectral
leakage will therefore almost certainly affect the spectrum of any real-world signal.
Remark 6.5: (Signal-domain impulses that fall between array elements): Due to the
duality between the signal and frequency domains under DFT, the input array can be seen
as the inverse DFT of the output array. Therefore, our discussion above applies also to
impulses in the input array of the DFT. Suppose we are given a continuous-world signal
g(x) that contains one or more impulses, and that we want to find its spectral
representation using DFT. For this end, we first need to sample g(x) in order to obtain the
input array that will be forwarded to the DFT. Obviously, the best would be to sample g(x)
such that all its impulses fall precisely on a discrete element of the input array. But what
happens when this is not possible, for example, if the impulse locations in g(x) are not

14This is not always possible since many DFT algorithms require that the array length N be a power of
2, or at least an even number.
6. Issues related to leakage 165

harmonically related (say, x = 1 and x = 2)? In such cases one (or more) of the impulses
of g(x) will actually fall between elements of the input array. Due to the duality between
DFT and the inverse DFT, it follows that such an in-between impulse should be obtained
in the input array by sampling the narrow signal-domain sinc function F sinc(F x) after
centering it on the true impulse location and scaling its amplitude. And if g(x) contains
several in-between impulses, their respective signal-domain sinc functions should be
summed up additively, just like in Fig. 6.4 (but this time in the signal domain). This may
be considered as signal-domain leakage. Other possible solutions such as shifting the
in-between impulse to the closest input array location, or even worse, ignoring it altogether
(which is the default solution when g(x) undergoes a standard “brute force” sampling)
will give imprecise results in the DFT as well as in the inverse DFT of these results. We
will return to the subject of signal-domain leakage in greater detail in Sec. 6.7 below. Q

Table 6.1 provides a synoptic review of the main properties of the aliasing and leakage
artifacts of the DFT.

6.6 Leakage in two or more dimensions

The above discussion about leakage can be extended to the multidimensional case, too.
A short discussion on the 2D case can be found, for example, in [Brigham88 pp.
250–253]. In the 2D case the default truncation window is w(x,y) = rect(x/R,y/R) =
rect(x/R) rect(y/R), whose spectrum is the 2D sinc function W(u,v) = R2 sinc(R u) sinc(R v).
This function has a narrow 2D main lobe centered at the origin, and a narrow “tail” of
decaying sidelobes trailing off ad infinitum to both sides of the main lobe along each of
the u and v axes; its sidelobes elsewhere in the u,v plane, away from both axes, are rather
negligible (see Fig. 6.8(b)). This is, indeed, a straightforward generalization of the 1D sinc
function R sinc(R u) of Eq. (6.1) (Fig. 6.6). However, as is often the case, here too the
situation in the MD case is richer and more diversified than in the 1D case: While in the
1D case the various alternative window functions that can be used instead of the default
truncation window rect(x/R) basically differ only in their gradual slopes on the borders of
the window, in 2D or MD cases window functions may also differ in their spatial
geometric shape. For example, in the 2D case one may replace the default rectangular
truncation window w(x,y) = rect(x/R,y/R), where R is the length of the sampling range
along each dimension, by a circular truncation window with diameter R. This truncation
window only differs from w(x,y) in its 2D geometric shape, but it still has a sharp 0/1
transition along its borders (see Fig. 6.8(c)):

wc(x,y) = rect( x 2 + y 2 /R)

The Fourier transform of this function, shown in Fig. 6.8(d), is expressed in terms of
the “sombrero” function, or equivalently in terms of the Bessel function J1 [Gaskill78 pp.
72–73, 329; Bracewell95 pp. 150–151]:

Wc(u,v) = π4 R2 somb(R u 2 + v 2 ) = R J1(πR u 2 + v 2 ) / (2 u 2 + v 2 )

166 6. Issues related to leakage

where:
⎧ 2J1(πr) /πr r ≠ 0
somb(r) = ⎨
⎩1 r=0

The function Wc(u,v) has the advantage over W(u,v) of being circular, with no protruding
tails along the axes. But although W(u,v) and Wc(u,v) are the most obvious choices, there
exist in the 2D or MD cases infinitely many other sharped-edged truncation functions,
that only differ in their geometric shape (triangular, hexagonal, etc.). And then, just as in
the 1D case, one may also devise based on each of these sharp-edged functions many
other variants with smoother transitions between 0 and 1.

Artifact: Aliasing Leakage

Properties:

Effect of the artifact in the A lower frequency signal Windowing, mismatch

signal domain: mimics the original signal between last and first values
on its sampled values

Effect of the artifact in the All the signal frequencies Smearing, ripple
spectral domain: beyond half of the sampling
frequency undergo a
cyclical foldover back into
the lower frequency range

Cause of the artifact: Undersampling of the given Truncation of the given

signal signal

Particular cases where the See Remark 5.2 See Remark 6.1
artifact does not occur:

Adequate choice of DFT Smaller Δx Larger R

parameters for reducing the ( ⇒ Larger F) ( ⇒ Smaller Δu)
artifact: (For fixed N: ⇒ Smaller R) (For fixed N: ⇒ Smaller F)

Appropriate treatment to be Low pass filtering; see also Smoother windowing

applied to the given signal Remark 6.10 in Sec. 6.7 (tapering); see also
(prior to DFT) for Remark 6.10 in Sec. 6.7
reducing the artifact:

Table 6.1: Synoptic overview of the main properties of aliasing and leakage.
6. Issues related to leakage 167

1
y 1
v
6
1
4
0 0
0.5 2
0 -6 0
-1
-4 -2
-0.5 -2
0 -0.5 0 -4
2
0.5 4 -6
(a) 1
-1 (b) 6
x u

1
y v
1
1 6
0 4
0.5 0
2
0 -6 0
-1
-4 -2
-0.5 -2
0 -0.5 0 -4
2
0.5 4 -6
(c) 1
-1 (d) 6
x u

Figure 6.8: (a) The default 2D truncation window, w(x,y) = rect(x/R,y/R), drawn
here for R = 1. (b) Its continuous Fourier transform is the narrow 2D
sinc function W(u,v) = R2 sinc(R u) sinc(R v), the 2D counterpart of the
function R sinc(R u) of Fig. 6.6. (c) The circular 2D truncation
window wc(x,y) = rect( x 2 + y 2 /R), drawn here for R = 1. (d) Its
continuous Fourier transform is the narrow “sombrero” function
W c (u,v) = π4 R2 somb(R u 2 + v 2 ) = R J 1 ( π R u 2 + v 2 ) / (2 u 2 + v 2 );
note the circular symmetry of this function.

As we have already seen in the 1D case, the form that leakage takes depends on the form
of the Fourier transform of the window being used. And indeed, by changing the shape of
the window we can change the shape of the leakage [Ramirez85 p. 109]. This fact is even
more striking in the multidimensional case. For example, Figs. 6.9(a),(b) show the 2D
DFT of the same periodic function; notice the different shapes of the leakage effect due to
the different window shapes being used.

The shapes of the leakage effect in the 2D case are indeed so remarkable, that some
leakage-related terms in the literature have been coined after these shapes. For example,
the term ringing that is often used as a synonym to ripple is inspired from 2D DFT
168 6. Issues related to leakage

spectra like in Fig. 6.9(b), where the ripple effect takes the form of concentric decaying
rings; interestingly, the term ringing is often used instead of ripple even in 1D cases,
where no rings are involved. Another term that is inspired from the 2D leakage shape in
DFT spectra like Fig. 6.9(a) is apodisation, which is employed in some references such as
[Weisstein99 pp. 55–56] to denote the windowing or tapering process used for leakage
reduction. Apodisation literally means “removing the feet”, referring to the protruding
“feet” (or rather tails) of the 2D-sinc leakage effect that occurs due to the 2D rectangular
truncation window rect(x/R,y/R) used by default.

Remark 6.6: It is interesting to note that leakage may also affect folded-over (i.e. aliased)
elements in the DFT spectrum, such as folded-over impulses that do not happen to fall
exactly on a DFT output element. And conversely, aliasing may also affect leaked-out
elements in the DFT spectrum if they happen to extend beyond the DFT borders (see, for
example, Fig. 6.9(a)). Note also that leakage may affect different elements in the DFT
spectrum in a different way; for example, in an impulsive DFT spectrum each impulse
may be leaked to a different extent, depending on its precise location between elements of
the DFT output array. Note in particular that the DC impulse is never affected by leakage,
since it is always located precisely on the first output array element (before the
reorganization), which corresponds to the spectrum origin. Q

Let us now see what is the multidimensional counterpart of Remark 6.1, namely, what
are the conditions in the 2D (or MD) case under which a given function g(x) does not
have leakage artifacts in its DFT spectrum. The generalization of condition (b) in Remark
6.1 is rather straightforward, so we will concentrate here on the generalization of condition
(a), which concerns periodic functions.

Figure 6.9: The shape of the 2D leakage artifact depends on the truncation window
being used. This is illustrated here using the original function g(x,y) =
cos(2π f [xcosθ + ysinθ]), a rotated cosinusoidal wave of frequency f,
whose CFT consists of an impulse pair. The two first rows show the
effect of two different truncation windows on the DFT of the function
g(x,y). (a) The default truncation window w(x,y), consisting of an R×R
square pulse, gives a cross-like leakage artifact around each impulse
that does not fall exactly on a discrete element in the DFT spectrum
(compare the shape of this leakage artifact with the shape of W(u,v) in
Fig. 6.8(b)). Note that since each of the leakage crosses trails away ad
infinitum, it wraps around the DFT border and re-enters from the
opposite side due to aliasing (see Chapter 5). (b) The circular
truncation window wc(x,y), consisting of a circular cylinder of radius
R/2, gives a sombrero-like leakage artifact around each impulse that
does not fall exactly on a discrete element in the DFT spectrum
(compare the shape of this leakage artifact with the shape of Wc(u,v) in
Fig. 6.8(d)). (c) Leakage can be completely avoided in this case by
slightly changing the angle θ and the frequency f of g(x,y) so that both
of its impulses fall exactly on a discrete element of the DFT output.
6. Issues related to leakage 169

Signal domain Spectral domain: Real part

4. 8.

2. 4.

0 0

-2. -4.

-4. -8.
-4. -2. 0 2. 4. -8. -4. 0 4. 8.
(a)
4. 8.

2. 4.

0 0

-2. -4.

-4. -8.
-4. -2. 0 2. 4. -8. -4. 0 4. 8.
(b)
4. 8.

2. 4.

0 0

-2. -4.

-4. -8.
-4. -2. 0 2. 4. -8. -4. 0 4. 8.
(c)
170 6. Issues related to leakage

6.6.1 The case of 2D 1-fold periodic functions

We start with the simplest possible 2D case, that of a 1-fold periodic function over the
x,y plane. For example, consider the cosine function with frequency of f cycles per unit
(i.e. period of P = 1/f units) along the x direction:

g(x,y) = cos(2π fx) (6.5)

After a rotation by angle θ counterclockwise this cosine function becomes (Fig. 6.9(a)):

g(x,y) = cos(2π f [xcosθ + ysinθ]) (6.6)

The period of this 2D function is given by the period vector P in the x,y plane which
emanates from the origin and whose length is P units along the angle θ:

P = (Pcosθ, Psinθ) (6.7)

The Fourier transform of this 1-fold periodic function consists of two impulses that are
located in the u,v plane to both sides of the origin, at the points (u,v) = ±(fcosθ, fsinθ).
Note that if our 1-fold periodic function g(x,y) is more complex than a simple cosine, its
spectrum will contain higher harmonic impulses, too, that are located in the u,v plane at
integer multiples of the first harmonic impulses, i.e. at the points (u,v) = ±k(fcosθ, fsinθ)
for integer values of k. The vector

f = (fcosθ, fsinθ) (6.8)

that connects the origin of the continuous-world spectrum to the location of the first
harmonic impulse, (fcosθ, fsinθ), is called the frequency vector of the 1-fold periodic
function g(x,y). Note that because in the 1-fold periodic case the vectors P and f are
collinear, we also have in this case the connections f = P/P 2, P = f/f 2 and P·f = 1.

Let us now consider p and q, the discrete-world counterparts of the period vector P and
the frequency vector f, which are expressed in terms of input array elements or output
array elements, respectively. We call these discrete-world vectors the p-vector and the
q-vector. If the sampling interval in the signal domain is Δx along both dimensions,15 the
length of the period vector P is given in terms of input-array elements by p = P/Δ x.
Similarly, if the length of each output-array element in the DFT spectrum is Δu Hz to both
directions, the length of the frequency vector f is given in terms of output-array elements
by q = f /Δu. We therefore have, in terms of input or output array elements, respectively:

p = (pcosθ, psinθ) = P/Δx (6.9)

q = (qcosθ, qsinθ) = f /Δu (6.10)

Note that since p and q are collinear the 1D relation pq = Pf/(ΔxΔu) = N remains valid
along angle θ (using Eq. (4.5)), and we also have q = pN/p2, p = qN/q2 and p·q = N.

15For the sake of simplicity we assume here that the array length N and the sampling interval Δx remain
the same in all dimensions; but an equivalent expression can be also formulated for the general case.
6. Issues related to leakage 171

N = 16

. q n
N = 16

Figure 6.10: Schematic illustration of the the DFT spectrum of a 2D 1-fold periodic
function whose impulses fall exactly on elements (“pixels”) of the
output array. In such cases, the q-vector q, which emanates from the
spectrum origin and points to the first harmonic impulse of the
spectrum, has exactly an integer number of array elements along each
of its Cartesian components: q = (m,n). Each array element in the DFT
spectrum represents a frequency step of Δu in the continuous-world
spectrum, so the corresponding frequency vector in the continuous-
world spectrum is f = (mΔu, nΔu). See also Problem 3-8.

Now, a straightforward generalization of the 1D case suggests that a 2D 1-fold periodic

function g(x,y) is free from leakage artifacts if and only if its impulses are located precisely
on 2D output array elements, i.e. exactly on “pixels” of the 2D image that represents the
2D DFT spectrum. This happens, of course, when each of the two Cartesian components
of the frequency vector f is exactly an integer multiple of the frequency step Δu, or in other
words, when the two Cartesian components of the corresponding discrete-world q-vector
q are integer numbers, say, m and n (see Fig. 6.10). This means that:

(1) The rotation angle θ of g(x,y) has a rational slope of tanθ = mn , and
(2) The frequency of g(x,y) along this angle is given, in terms of DFT spectrum pixels, by:

q = m2 + n 2

We will call 1-fold periodic functions g(x,y) which satisfy these conditions rational
1-fold periodic functions. The q-vector q = (qcosθ, qsinθ) of such a function, which
connects the origin (0,0) to the pixel (m,n) in the 2D DFT spectrum, will be called a
172 6. Issues related to leakage

rational q-vector. The coordinates of such a vector are m and n array elements, both of
which are integer numbers; this can be easily extended to any MD case, too.16

We see, therefore, that a 1-fold periodic function g(x,y) has no leakage artifacts in the 2D
DFT spectrum if and only if it is rational, i.e. if and only if its q-vector q is rational. As
shown in Appendix C, this is precisely the condition required for the 1-fold periodic
function g(x,y) to be seamlessly wraparound within the N×N input array of the DFT. This
wraparound property means, in other words, that if we pave the plane with juxtaposed
replicas of this 2D array, there will be no discontinuities at the array borders and the
periodicity will be perfect and seamless throughout the entire plane (see Fig. C.9).

Note that in terms of the signal domain, the fact that the q-vector q is rational (i.e.
satisfies conditions (1) and (2) above) means that the rotation angle of our 1-fold periodic
function g(x,y) is tanθ = mn , and that its period length along this direction is, in terms of
input-array elements (see Eq. (4.13)):

p = N/q = N/ m 2 + n 2 (6.11)

This does not yet mean, however, that the p-vector p of this function is also rational (i.e.
extends along an integer number of input-array pixels both horizontally and vertically). In
order that the p-vector p be rational, it must itself satisfy condition (2). We will return to
this point in greater detail in Apendix C, where we discuss the implications of the
rationality of the p-vectors and/or of their corresponding q-vectors.

Finally, note that condition (2) can be also translated from the discrete-world language
of array elements (or pixels) into the continuous-world language of true frequency units in
terms of Hz (see the 1D equivalent in Sec. 4.2): The length of a single pixel along each
dimension of the DFT output array is given by Δu = 1 (Eq. (4.5)). Therefore, condition
NΔx
(2) says that the frequency f in terms of Hz along the direction of the frequency vector f
must have the form:

f = qΔu = m 2 + n 2 Δu = m 2 + n 2 / (NΔx) (6.12)

for some integers m and n. Its components along the u and v axes are mΔu and nΔu.

6.6.2 The case of 2D 2-fold periodic functions

We now proceed to the more general case of 2D periodic functions, that of a 2-fold
periodic function over the x,y plane. This time the function g(x,y) has two independent
period vectors P1 and P2 rather than one in the x,y plane, and two frequency vectors f1 and
f2 rather than one in the u,v plane. Therefore, its spectrum may contain impulses at all the
locations k1f1 + k2f2 with integer values of k1 and k2. Just as in the 1-fold periodic case,
here too we consider the discrete-word counterparts of the period vectors P1 and P2, the

16Note that by abuse of language we may also say that the corresponding frequency vector f in the
continuous world is a rational frequency vector, although its coordinates are not integer numbers m and
n but rather mΔq and nΔq Hz.
6. Issues related to leakage 173

p-vectors p 1 and p 2, which are expressed in terms of input-array elements; and the
discrete-world counterparts of the frequency vectors f1 and f2, the q-vectors q1 and q2,
which are expressed in terms of output-array elements:

pi = Pi / Δx (6.13)

q i = fi / Δ u (6.14)

Now, a 2-fold periodic function g(x,y) generates no leakage artifacts in the 2D DFT
spectrum if and only if both of its q-vectors q 1 and q 2 are rational, namely, when
tanθ1 = n1/m1 and q1 = m 12 + n 12 , and tanθ2 = n2/m2 and q2 = m 22 + n 22 . Note that in
the case of 2-fold periodic functions the two main periodicity directions are not the same
as the two main frequency directions, but rather orthogonal to them; see Appendix C and
Fig. C.5 there. Nevertheless, the graphic interpretation of these conditions in terms of the
signal domain remains the same as in the 1-fold periodic case, namely, that our 2-fold
periodic function g(x,y) is perfectly cyclical (seamlessly wraparound) within the N×N
input array of the DFT. We will return to this point in greater detail in Appendix C.

6.6.3 The general MD case

The discussions above can be also extended to the MD n-fold periodic case (see
Appendix C). We finally obtain the following MD generalization of Remark 6.1:

Remark 6.7: Just like in the 1D setting, in the multidimensional setting, too, there exist
two particular cases in which leakage does not occur in spite of the finite-range windowing
of the underlying signal: (a) when the given function g(x) is n-fold periodic with
1 ≤ n ≤ M, and it is perfectly cyclical (seamlessly wraparound) within its sampling range
(truncation window) RM;17 or (b) when g(x) is a finite-duration function, and it is fully
included within the sampling range R M so that it is not truncated by the rectangular
window rect(x/R). Note that in terms of the spectral domain, case (a) means that each of
the impulses of the n-fold periodic function g(x) in the DFT spectrum is perfectly sharp,
i.e. that (i) it falls precisely on an element of the output array, and (ii) its width is exactly
one frequency sample (a single element of the output array). This is, indeed, the dual
formulation of case (a) in terms of the spectral domain, as shown in Appendix C. In all
cases other than (a) and (b) leakage is unavoidable, and the only solution is to reduce
leakage as much as possible, using the same methods as in the 1D setting. Q

Note that the conditions of Remark 6.7 do not prevent aliasing; but if aliasing does
occur — for example in the form of impulses that exceed beyond the borders of the DFT
spectrum and re-enter from the opposite side — it will not be accompanied by leakage, so
that each impulse, aliased or not, will fall exactly on one single pixel of the DFT spectrum.
The conditions on g(x) for having neither aliasing nor leakage in the DFT spectrum are
given in the following multidimensional counterpart of Remark 6.2:

17 This also means that the input array NM seamlessly paves the entire M-dimensional discrete space.
174 6. Issues related to leakage

Remark 6.8: By combining Remark 5.4 and Remark 6.7 we see that the only case where
both aliasing and leakage are completely absent occurs when the given signal g(x) is both
band limited and periodic, the sampling rate being used in each dimention is at least twice
the largest frequency component of g(x) in that dimension, and the function g(x) is
perfectly cyclical (seamlessly wraparound) within its sampling range (truncation window)
RM. This is, indeed, the only case where the (reorganized) DFT and the continuous Fourier
transform perfectly coincide on the NM discrete points of the DFT spectrum (within a
scaling constant, as already explained above in Sec. 4.3). Note that case (b) of Remark 6.7
is excluded here, since a finite-duration function g(x) cannot be band limited [Brigham88
p. 103], so that case (b) cannot be aliasing-free. Q

6.7 Signal-domain leakage

Just as in the case of aliasing, it follows from the duality between DFT and the inverse
DFT that all our previous discussions on the leakage phenomenon under DFT remain true
under IDFT, too, except that the roles of the input and output arrays are reversed. For
example, each row in Fig. 6.11 shows in the right-hand column the result obtained by
applying DFT to the discrete sequence shown in the left-hand column. But the same
figure can be also interpreted the other way around, where the left-hand column is the
result of applying IDFT to the discrete sequence shown in the right-hand column. Now,
just as in the DFT, if the truncation which occurs in the input of the IDFT (the truncation
of the continuous-world spectrum G(u) into the finite range length F) is inadequate, it will
cause leakage in the output of the IDFT; but this time, the leakage occurs in the signal
domain, and we can therefore call it signal-domain leakage, or in the 1D case time-domain
leakage [Smith96]. Stated in other words, if a continuous-world frequency function G(u)
is truncated within a finite range length F in order to be forwarded to IDFT, and
(a) G(u) is periodic (meaning that g(x) is impulsive) and its truncation into the finite range
length F does not give exactly an integer number of periods (see Figs. 6.11(b)–(c)); or
(b) G(u) is aperiodic and it does not effectively die out by the borders of this range;
then the IDFT of the sampled version of G(u) will suffer from signal-domain leakage.

However, the same signal-domain leakage in the very same figure can be also regarded
from a second point of view, which does not explicitly require the use of the IDFT:
Suppose we are given a continuous-world signal g(x) whose continuous-world spectrum
is G(u), and that we try to approximate this continuous-world Fourier transform by DFT.
As we already know, the DFT can only take into account a finite range of frequencies.
Therefore, if the original, underlying continuous-world spectrum G(u) does not die out by
the borders of this range, G(u) should be truncated before the transition to the discrete
world (in order to avoid frequency foldover due to aliasing; see Sec. 5.2). This truncation
operation can be viewed, still in the continuous world, as a multiplication of the underlying
spectrum G(u) by a rectangular “truncation window” W(u) whose length equals the DFT
frequency range length F = NΔu, namely, W(u) = rect(u/F). This multiplication in the
6. Issues related to leakage 175

spectral domain is equivalent in the signal domain to a convolution of the input signal g(x)
with the inverse Fourier transform of the truncation window W(u), w(x) = F sinc(F x). This
convolution represents, indeed, a signal-domain leakage of g(x). Note, however, that the
situation is easier to understand from the spectral-domain point of view: Signal-domain
leakage occurs if the truncation of the continuous-world spectrum G(u) by W(u) is
inadequate (non-integer multiple of the period, if G(u) is periodic; or too narrow, if G(u) is
aperiodic). And conversely, signal-domain leakage does not occur (a) if G(u) is periodic
and it is truncated exactly on an integer multiple of its period; or (b) if G(u) is aperiodic
and its frequency range is fully included within that of the DFT spectrum (compare with
conditions (a) and (b) in Remark 6.1, which speaks about spectral-domain leakage).18

Example 6.3: Signal-domain leakage when G(u) is periodic:

Consider Fig. 6.3, and suppose that we are interested in the inverse situation: what input
should we supply to the DFT in order to obtain in its output the truncated cosines shown
in the signal domain of Fig. 6.3? In this case, shown in Fig. 6.11, the given continuous-
world signal is g(x) = 12 δ(x–f) + 12 δ(x+f), and the resulting continuous-world spectrum is
G(u) = cos(2πfu). The 3 rows of the figure only differ in the value of the cosine frequency
f. The difficulty here, when we proceed to the discrete world, is twofold: From the signal
domain point of view, how should we sample the impulse pair g(x) when the impulses do
not fall exactly on sampling points, i.e. when they fall in-between input array elements, like
in rows (b) and (c) of Fig. 6.11? From the point of view of the spectral domain the dual
question is, how can we obtain in the DFT output a cosine function which is not exactly
truncated into an integer number of its periods, like in rows (b) and (c)?19 The answer to
both questions is the same: Instead of sampling into the input array of the DFT the
function g(x) itself, we have to sample the convolution g(x) * F sinc(F x), which
corresponds in the spectral domain to the truncation of G(u) into the frequency range
length F. From the signal domain point of view, this convolution represents, indeed,
signal-domain leakage, as we can see in the discrete input signal in rows (b) and (c). In
row (a), however, where the impulses do fall on input-array elements, no leakage effect
appears in the discrete input signal, since in this case the convolution simply gives back
the original impulse pair g(x) (up to a vertical scaling by F), as explained in Fig. 6.4(a) and
Sec. 6.5. But although this convolution is practically superfluous in row (a), its effect of
vertically scaling the input impulses by F should not be forgotten, in order to obtain the
correct cosine amplitude in the DFT spectrum. This scaling by F (amplitude correction) is
discussed in more detail in Sec. A.4 of Appendix A.
It is interesting to note, however, that even after this convolution, the resulting DFT
spectrum still suffers from a small discrepancy with respect to the continuous cosine
spectrum G(u). This tiny difference is best visible in the leftmost point of the DFT
spectrum in rows (b) and (c) of Fig. 6.11. It occurs due to the truncation of the function

18 Note that from the signal-domain point of view the situation is less easy to understand, since the
convolution of g(x) with F sinc(F x) does not always cause g(x) to leak, as we have just seen in
conditions (a) and (b). See also Sec. D.4 in Appendix D.
19 The duality between these two questions is explained in detail in Sec. 6.5.
176 6. Issues related to leakage

g(x) * F sinc(F x) at both ends of the input array (see the left-hand column in Fig. 6.11),
i.e. due to conventional spectral-domain leakage, as explained in the beginning of the
present chapter. Q

In conclusion, we have the following result:

Remark 6.9: (Signal-domain leakage): If the continuous-world frequency function G(u)

does not effectively die out by the borders of the finite range F that is taken care of by the
DFT, then the input signal g(x) should be convolved with F sinc(F x) before it is forwarded
to the DFT. This convolution represents signal-domain leakage in g(x). Ignoring this point
(which is precisely what happens by default when g(x) undergoes standard sampling and
truncation before it is forwarded to the DFT) will result in a discrepancy between the
resulting DFT and the original continuous-world spectrum G(u). When G(u) is aperiodic
this discrepancy will occur since the DFT output will not correspond to a truncated
version of G(u) but rather to a folded-over (aliased) version of G(u). And when G(u) is
periodic, this discrepancy will occur if impulses of g(x) are not located exactly on input
array elements, thus giving a bad sampling of g(x) and hence a bad DFT output. Q

We see, therefore, that thanks to the duality between the signal and frequency domains,
leakage and aliasing occur in both domains. Moreover, it turns out that they are very
closely related — in fact, they are duals of each other, as explicitly stated by the following
result (note that signal-domain aliasing has already been discussed in Sec. 5.7):

Figure 6.11: Example of signal-domain leakage that occurs when the continuous-world
spectrum G(u) is periodic, so that the continuous-world input function g(x)
is impulsive. In the present example we have G(u) = cos(2π fu) and
therefore g(x) = 12 δ(x–f) + 12 δ(x+f); this is exactly the inverse of the situation
shown in Fig. 6.3. In row (a) the truncation at the borders of the DFT
spectrum corresponds exactly to an integer number of periods of G(u); and
in terms of the signal domain, the two impulses of g(x) fall exactly on input
array elements. In this case no signal-domain leakage is present. In rows
(b)–(c) the truncation at the borders of the DFT spectrum no longer
corresponds to an integer number of periods; and in terms of the signal
domain, the impulses of g(x) fall in-between input array elements. Note that
the correct cosine is obtained in the DFT spectrum because we have
sampled in the input array the convolution of the impulse pair g(x) with
F sinc(F x) rather than the impulse pair g(x) itself, to reflect the spectral-
domain truncation to a frequency range of length F. This convolution
precisely represents signal-domain leakage. All rows are shown here after
the required adjustments as discussed in Chapters 3 and 4 and in Appendix
A. In each case discrete signals are represented by black dots, while the
original continuous-world functions (the impulse pair g(x) in the signal
domain and the cosine function G(u) in the spectral domain) are drawn with
 6. Issues related to leakage 177

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
4
k n
2 CFT
3
g(x)*8sinc(8x) 1.5 • • • DFT after
• • • Sampled reorg. & v. sc.
1
2
g(x) 0.5
1 0
0 -0.5
-1
-1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
x -1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) f = 1/2 = 4Δx

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
k n
4 2 CFT
3 1.5 • • • DFT after
g(x)*8sinc(8x) reorg. & v. sc.
2
• • • Sampled 1
0.5
1 g(x) 0
0 -0.5
-1
-1
-2. -1. 0 1. 2.
x -1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz
-1.5 -0.5 0.5 1.5
(b) f = 17/32 = 4.25Δx

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 k 0 4 8 12 16 20 24 28 31 n
4 2 CFT
g(x)*8sinc(8x) 1.5 • • • DFT after
3
• • • Sampled reorg. & v. sc.
1
2
g(x) 0.5
1 0
0 -0.5
-1
-1
x -1.5 Hz
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2. -4. -3. -2. -1. 0 1. 2. 3. 4.
(c) f = 18/32 = 4.5Δx

continuous lines. The dashed lines in the signal domain correspond to the
convolution g(x) * F sinc(F x). Note that the amplitude at the two peaks of
the convolution g(x) * F sinc(F x) is F = 8 times higher than the amplitude 12
of the original continuous-world impulse pair g(x). A similar vertical scaling
(by R = 8) occurs in the spectral domain of Fig. 6.3, but it is not visible there
since it has been corrected by the adjustments due to Appendix A.
178 6. Issues related to leakage

Remark 6.10: (Duality of leakage and aliasing in both domains [Smith96 p. 741]):
Leakage in one domain (signal or frequency) removes (or reduces) aliasing in the other
domain; and aliasing in one domain (signal or frequency) removes (or reduces) leakage in
the other domain. Q

For example, although low-pass filtering (which is used to reduce or eliminate aliasing;
see end of Sec. 5.2 or last row in Table 6.1) is seldom referred to as signal-domain
leakage, it is exactly analogous: The portions of the original signal g(x) at any point which
are contributed by high frequencies, such as sharp edges or peaks, are effectively leaked to
neighbouring points of the signal, since the truncation of the spectrum G(u) to a frequency
range length F is equivalent in the signal domain to a convolution of the signal g(x) with
F sinc(F x). Note that if the low-pass filter is physically realizable, the resulting leakage is
one-way, i.e. leakage is always to later points in time and never to earlier points in time
[Smith96 p. 741].

This example illustrates, indeed, how signal-domain leakage reduces (or eliminates)
frequency-domain aliasing. Similarly, Fig. 5.19 illustrates how aliasing (foldover) in the
signal domain reduces (or eliminates) leakage in the spectral domain. As we can see,
Remark 6.10 sheds new light on Table 6.1 and on its interpretation.

Note, however, that allowing signal-domain aliasing as a remedy to frequency-domain

leakage, as shown in Fig. 5.19, does not work in the case of periodic signals g(x). This is
quite logical, since a poor in-between location of an impulse in the DFT spectrum cannot
be healed by simply allowing folding-over of the periodic function in the signal domain.
For example, if g(x) = cos(2πfx), then its signal-domain folding-over simply gives in the
signal domain a sum of cosines having the same frequency f but different phases. But this
will only influence in the DFT spectrum the height of the resulting impulses (and their
complex-valued phase, if g(x) is not symmetric), but not their locations. Each impulse will
therefore remain located at the same point in-between DFT output array elements, and
hence it will continue suffering from leakage. For this reason, the term “leakage” as used
in Remark 6.10 above does not include the periodic variant of leakage, where the function
g(x) or G(u) is purely periodic but it is not truncated precisely on an integer multiple of its
period (like in Figs. 6.3(b),(c) or 6.11(b),(c), respectively). Nevertheless, if the function is
not purely periodic, its folding-over may still be used to reduce leakage in the other
domain [Smith96 pp. 744–745].

Finally, another related point should be mentioned in this context. Consider row (a) of
Fig. 6.5. It is obvious that in order to get, back in the signal domain, the same discrete
sequence as the one shown in the left-hand column, we must provide as input to the IDFT
the leaked signal shown in the right-hand column by dashed lines. Failing to do so will
result in a distorted signal in the output array of the IDFT, that will not match the signal
shown in the left-hand side of the figure. In this case, however, it would be more
appropriate to use the term input-domain leakage, since the leakage effect in the input
array of the IDFT occurs, in fact, in the frequency domain, and it cannot be considered as
signal-domain leakage.
6. Issues related to leakage 179

PROBLEMS

6-1. Can the DC impulse in a DFT spectrum leak out? Why? Verify your answer by
checking the DFT figures in the present chapter and in other examples of your own.
6-2. Is it possible that within the same impulse comb in a 1D DFT spectrum some of the
impulses be leaked out and some other impulses not? Hint: What happens in a 1D
comb when q = 4? when q = 3.5? and when q = 3.2?
6-3. Part (a) of Remark 6.1 says that when the given function g(x) is periodic, and the range
length R of the truncation window is an exact integer multiple of the period of g(x), no
leakage occurs. A student claims that Fig. 6.2 contradicts this result, since in this figure
the cosine function has been truncated at exactly an integer multiple of the period, and
yet, the impulses in the right-hand part of row (c) are clearly smeared and leaked-out
(compare with row (a) of the figure). What is your opinion?
6-4. Rational vectors. Fig. 6.12 shows a table with all the 2D rational vectors (rational
combinations of integers m and n; see Fig. 6.10) for m = 1,...,16 and n = 1,...,11. Each
cell in the table gives the ratio n/m, the corresponding rational angle θ = arctan(n/m),
and the value of q = m 2+n 2 . What is the significance of this table? Hint: This table
may come in handy when you wish to design a 2D discrete periodic example with
rational q-vectors, like Fig. C.9(b). Note that in Fig. C.9(b) we have used m = 7, n = 2
so that the rotation angle of our 2D cosine function was θ = arctan(2/7) = 15.9454°
and its frequency was, by Eq. (6.12), f = m 2+n 2 / R = 7.2801 / 16 = 0.455 Hz. This
table only shows the possible rational vectors for small values of m and n; but it can be
extended to any desired values of m and n.
6-5. Is it possible to find a rational vector whose angle is exactly 15°? 30°? 45°? Explain.
6-6. Suppose you need to plot for a new publication a 2D cosine function with frequency
f = 1 and rotation angle of θ = 15°, along with its spectrum. In order to obtain using
your 2D DFT a clear spectral representation of the two impulses without awkward
leakage effects, you agree to use a close rational approximation of the specified
parameters f = 1 and θ = 15°. Suppose that you wish to plot the function within the
range –8...8 along both axes, so that R = 16. This means that in terms of array
elements (pixels) you have q = f / Δ u = f R = 16. Which combination of m and n would
you choose in the table? Explain.
6-7. What would our student from Problem 5-1 obtain in his DFT plot if he modified the
parameters of the given line grating (P = 1, θ = 15°) according to the results of the
previous problem? Hint: See Fig. 6.13. Once the impulses in the DFT spectrum are
sharp and no longer suffer from leakage, would you expect their heights to be the
same as in the CFT? Explain.
6-8. If our student chooses in the table another close rational approximation, he will get
again sharp impulses in the DFT spectrum, but their locations will be somewhat
different than before. Will the heights of the impulses remain unchanged?
6-9. Several figures in this book have been designed using the table of Fig. 6.12, in order to
avoid leakage artifacts in the DFT spectrum. Consider, for example, Fig. 8.7, in which
all the impulses in the DFT spectrum are sharp and clean. Given that the array size used
in this figure is N = 64 to both directions, can you find out the values of m and n in this
case?
6-10. How can the table in Fig. 6.12 be extended to the 3D case?
6-11. Identification of leakage in DFT spectra. How can you tell if a certain frequency in
the DFT spectrum is indeed a true spectral frequency of the given spectrum, or just a
180 6. Issues related to leakage

leakage artifact? Remark 5.3 in Chapter 5 provides a simple method that often allows
to distinguish between true and aliased frequencies in the multidimensional case, based
on the rotation theorem (or on the dilation theorem). Can you think of a similar
method for the leakage artifact? Consider, for example, Fig. 6.9(a); how will it be
modified if the original continuous-world function g(x,y) is slightly rotated before
sampling? Will the cross-like structure around each impulse be rotated, too, like any
true spectral element of G(u,v)?
6-12. Having obtained the DFT spectrum of a single rotated b/w line grating, our student
now wishes to study the spectrum of a superposition of two such gratings having
slightly different rotation angles.
(a) What is the CFT of such a grating superposition? Hint: See, for example,
[Amidror09 Secs. 2.6, 6.5].
(b) What would you expect to see in the DFT spectrum of such a grating
superposition? In particular, is it possible to avoid leaked-out impulses?
6-13. Limitations of the DFT. Based on his experience so far, our student comes to the
conclusion that after all, for his specific needs, the use of DFT is not the best option:
Due to the various artifacts inherent to the DFT (such as aliasing and leakage), impulses
in the DFT spectrum do not always represent true spectral entities in the continuous-
world spectrum; and even when they do, their values (heights) do not necessarily give a
reliable estimation of their true values in the CFT. Even after applying all the required
adjustments as explained in Chapters 3, 4 and Appendix A, a perfect match with the
CFT is only possible in simple cases where the DFT artifacts can be avoided or easily
mastered. In more complex situations, the DFT can offer a good qualitative
approximation to the continuous-world spectrum, but not a fully reliable quantitative
evaluation. What is your opinion on this conclusion?
6-14. Can you think of an alternative method, not based on DFT, that could help our
student in finding the impulse locations and heights in the spectrum of a superposition
of two (or even more) periodic gratings? Hint: Due to the 1-fold periodicity of each
grating, the superposition of n gratings corresponds in the spectral domain to a
convolution of the n respective impulse combs. The impulse locations and heights in
such cases can be easily determined using the properties of impulse comb
convolutions. This can be done, for example, by a short computer program that
simulates the behaviour of the continuous-world comb convolution. See, for example,
Problem 5-22 in [Amidror09 p. 148].
6-15. Can you think of other Fourier or spectral analysis situations where the DFT is not the
optimal tool? Hint: In general, in most situations where the computation of the CFT is
feasible — be it by directly using the definition formulas, by using Fourier-pair tables
in the literature, by applying transformation and combination rules to already known
Fourier pairs, or by any other means (such as the comb-convolution approach in the
last problem) — the results you obtain will be more accurate than by using DFT. But
there also exist various other circumstances where the DFT approach is not optimal.
This is the case, for example, in the spectral analysis of sound, music, radar or video
signals, etc., where the spectral content varies with time or with spatial location, either
continuously or discretely. In such cases an approach based on joint-domain analysis
(such as joint time / frequency analysis in the 1D case or joint spatial / spatial-frequency
analysis in the 2D case) may be preferred [Gröchenig01; Poularikas96, Chapter 12]. In
applications where compression or multiresolution are sought, the wavelet approach
may be more judicious [Mallat91; Strang93; Rao98; Poularikas96, Chapter 10;
Bovik05, Chapters 4.2 and 5.4]. Thus, for each application one should use the method
6. Issues related to leakage 181

n
m· 1 2 3 4 5 6 7 8 9 10 11
1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0 11.0
1 45.0Ň
1.4142
63.4349Ň 71.5651Ň 75.9638Ň 78.6901Ň 80.5377Ň 81.8699Ň 82.875Ň
2.2361 3.1623 4.1231 5.099 6.0828 7.0711 8.0623
83.6598Ň 84.2894Ň 84.8056Ň
9.0554 10.0499 11.0454

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5
2 26.5651Ň 45.0Ň
2.2361 2.8284
56.3099Ň 63.4349Ň 68.1986Ň 71.5651Ň 74.0546Ň 75.9638Ň 77.4712Ň 78.6901Ň 79.6952Ň
3.6056 4.4721 5.3852 6.3246 7.2801 8.2462 9.2195 10.198 11.1803

0.3333 0.6667 1.0 1.3333 1.6667 2.0 2.3333 2.6667 3.0 3.3333 3.6667
3 18.4349Ň 33.6901Ň 45.0Ň
3.1623 3.6056 4.2426
53.1301Ň 59.0362Ň 63.4349Ň 66.8014Ň 69.444Ň
5.0 5.831 6.7082 7.6158 8.544
71.5651Ň 73.3008Ň 74.7449Ň
9.4868 10.4403 11.4018

0.25 0.5 0.75 1.0 1.25 1.5 1.75 2.0 2.25 2.5 2.75
4 14.0362Ň 26.5651Ň 36.8699Ň 45.0Ň
4.1231 4.4721 5.0 5.6569
51.3402Ň 56.3099Ň 60.2551Ň 63.4349Ň 66.0375Ň 68.1986Ň 70.0169Ň
6.4031 7.2111 8.0623 8.9443 9.8489 10.7703 11.7047

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2
5 11.3099Ň 21.8014Ň 30.9638Ň 38.6598Ň 45.0Ň
5.099 5.3852 5.831 6.4031 7.0711
50.1944Ň 54.4623Ň 57.9946Ň 60.9454Ň 63.4349Ň 65.556Ň
7.8102 8.6023 9.434 10.2956 11.1803 12.083

0.1667 0.3333 0.5 0.6667 0.8333 1.0 1.1667 1.3333 1.5 1.6667 1.8333
6 9.4623Ň 18.4349Ň 26.5651Ň 33.6901Ň 39.8056Ň 45.0Ň
6.0828 6.3246 6.7082 7.2111 7.8102 8.4853
49.3987Ň 53.1301Ň 56.3099Ň 59.0362Ň 61.3895Ň
9.2195 10.0 10.8167 11.6619 12.53

0.1429 0.2857 0.4286 0.5714 0.7143 0.8571 1.0 1.1429 1.2857 1.4286 1.5714
7 8.1301Ň 15.9454Ň 23.1986Ň 29.7449Ň 35.5377Ň 40.6013Ň 45.0Ň
7.0711 7.2801 7.6158 8.0623 8.6023 9.2195 9.8995
48.8141Ň 52.125Ň
10.6301 11.4018
55.008Ň
12.2066
57.5288Ň
13.0384

0.125 0.25 0.375 0.5 0.625 0.75 0.875 1.0 1.125 1.25 1.375
8 7.125Ň
8.0623
14.0362Ň 20.556Ň
8.2462 8.544
26.5651Ň 32.0054Ň 36.8699Ň 41.1859Ň 45.0Ň
8.9443 9.434 10.0 10.6301 11.3137
48.3665Ň 51.3402Ň 53.9726Ň
12.0416 12.8062 13.6015

0.1111 0.2222 0.3333 0.4444 0.5556 0.6667 0.7778 0.8889 1.0 1.1111 1.2222
9 6.3402Ň 12.5288Ň 18.4349Ň 23.9625Ň 29.0546Ň 33.6901Ň 37.875Ň
9.0554 9.2195 9.4868 9.8489 10.2956 10.8167 11.4018
41.6335Ň 45.0Ň
12.0416 12.7279
48.0128Ň 50.7106Ň
13.4536 14.2127

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1
10 5.7106Ň 11.3099Ň 16.6992Ň 21.8014Ň 26.5651Ň 30.9638Ň 34.992Ň
10.0499 10.198 10.4403 10.7703 11.1803 11.6619 12.2066
38.6598Ň 41.9872Ň 45.0Ň
12.8062 13.4536 14.1421
47.7263Ň
14.8661

0.0909 0.1818 0.2727 0.3636 0.4545 0.5455 0.6364 0.7273 0.8182 0.9091 1.0
11 5.1944Ň 10.3048Ň 15.2551Ň 19.9831Ň 24.444Ň
11.0454 11.1803 11.4018 11.7047 12.083
28.6105Ň 32.4712Ň 36.0274Ň 39.2894Ň 42.2737Ň 45.0Ň
12.53 13.0384 13.6015 14.2127 14.8661 15.5563

0.0833 0.1667 0.25 0.3333 0.4167 0.5 0.5833 0.6667 0.75 0.8333 0.9167
12 4.7636Ň 9.4623Ň 14.0362Ň 18.4349Ň 22.6199Ň 26.5651Ň 30.2564Ň 33.6901Ň 36.8699Ň 39.8056Ň 42.5104Ň
12.0416 12.1655 12.3693 12.6491 13.0 13.4164 13.8924 14.4222 15.0 15.6205 16.2788

0.0769 0.1538 0.2308 0.3077 0.3846 0.4615 0.5385 0.6154 0.6923 0.7692 0.8462
13 4.3987Ň 8.7462Ň 12.9946Ň 17.1027Ň 21.0375Ň 24.7751Ň 28.3008Ň 31.6075Ň 34.6952Ň 37.5686Ň 40.2364Ň
13.0384 13.1529 13.3417 13.6015 13.9284 14.3178 14.7648 15.2643 15.8114 16.4012 17.0294

0.0714 0.1429 0.2143 0.2857 0.3571 0.4286 0.5 0.5714 0.6429 0.7143 0.7857
14 4.0856Ň 8.1301Ň 12.0948Ň 15.9454Ň 19.6538Ň 23.1986Ň 26.5651Ň 29.7449Ň 32.7352Ň 35.5377Ň 38.1572Ň
14.0357 14.1421 14.3178 14.5602 14.8661 15.2315 15.6525 16.1245 16.6433 17.2047 17.8045

0.0667 0.1333 0.2 0.2667 0.3333 0.4 0.4667 0.5333 0.6 0.6667 0.7333
15 3.8141Ň 7.5946Ň 11.3099Ň 14.9314Ň 18.4349Ň 21.8014Ň 25.0169Ň 28.0725Ň 30.9638Ň 33.6901Ň 36.2538Ň
15.0333 15.1327 15.2971 15.5242 15.8114 16.1555 16.5529 17.0 17.4929 18.0278 18.6011

0.0625 0.125 0.1875 0.25 0.3125 0.375 0.4375 0.5 0.5625 0.625 0.6875
16 3.5763Ň 7.125Ň
16.0312 16.1245
10.6197Ň 14.0362Ň 17.354Ň
16.2788 16.4924 16.7631
20.556Ň
17.088
23.6294Ň 26.5651Ň 29.3578Ň 32.0054Ň 34.5085Ň
17.4642 17.8885 18.3576 18.868 19.4165

Figure 6.12: A table showing all the 2D rational combinations for m = 1,...,16 and
n = 1,...,11. Each cell gives the ratio n/m, the corresponding rational
angle θ = arctan(n/m), and the value of q = m 2 + n 2 . The diagonal
guiding lines show the angles of 15°, 30°, 45°, 60° and 75°. In order
to compare with Fig. 6.10, it may be helpful to rotate this table by 90°
counterclockwise, so that m increases to the right and n increases
going upward.
182 6. Issues related to leakage

which is best adapted to the problem. You may have a glimpse at the various alternative
analysis methods in [Poularikas96] or in other handbooks on signal or image
processing.
6-16. Is it possible for an impulse in a 2D or MD DFT spectrum to leak in one dimension
only? Hint: Consider Fig. 6.9(c). How would the DFT spectrum look like if each of the
two cosine impulses fell between two neighbouring pixels in the horizontal direction?
and if it fell between two neighbouring pixels in the vertical direction? Part (a) of the
figure shows what happens when the impulses fall between two pixels both horizontally
and vertically.
6-17. As we have seen in Sec. 6.6.1, a rotated line grating such as in Fig. 5.14(a) can be
made fully leakage-free by slightly modifying its angle and frequency to be rational.
Suppose that we have obtained such a leakage-free line grating, whose impulses in the
2D DFT spectrum are therefore perfectly sharp (like in Fig. 8.7(a)). Let us take now
only the central line of this grating, the one that passes through the origin (like in Fig.
8.8(a)). This line is a subset of the leakage-free line grating, and it clearly has a rational
angle. Will it be leakage free? Hint: As we can clearly see, Fig. 8.8(a) is not leakage
free, although the line has a rational angle; does this contradict our results in Sec. 6.6?
Will the 2D DFT input array containing this line segment pave seamlessly the entire x,y
plane? In which cases it will, and in which cases it will not? Consider in particular the
rotation angle of the line.
6-18. A student looks at Figs. 8.6(b),(c) just after studying Sec. 3.6 in Chapter 3, and comes
to the conclusion that the checkerboard-like nature of the spectra in these figures is
due to a reorganization problem in the input array of the DFT (either missing
reorganization or superfluous reorganization).
(a) What is your opinion?
(b) How will the cross-shaped leakage artifact of an impulse in a 2D DFT spectrum
look like in case of missing or superfluous input array reorganization? Generate and
plot a few such cases using your favorite DFT software package and verify your
answer. Hint: Compare Fig. 4.8 with its non-reorganized counterpart, Fig. 3.11 (note
that in Fig. 3.11 no output array reorganization was performed, either).

0.1 v
0.05 2.
0
-0.05 1.
-0.1
-2. 0
-1.
0 -1.

1.
-2.
2.
u
Figure 6.13: The DFT plot obtained by the student in Problem 6-7.
6. Issues related to leakage 183

6-19. Student A applies DFT to a cosine function and obtains the result shown in Fig. 6.14.
How do you explain this result? And how do you explain the scale along the vertical
axis?
6-20. Student B applies DFT to the same cosine function as Student A, but obtains the result
shown in Fig. 6.15. How do you explain this result? Hint: See Chapter 3.
6-21. Student C applies DFT to the same cosine function as student A, but obtains the result
shown in Fig. 6.16. How do you explain this result? Hint: See Sec. 6.4.
6-22. What are the main adverse effects of leakage? Hint: Limitation of the frequency
resolution; significant alteration of the spectrum amplitudes; a strong spectral peak
may hide behind its leaked tails other weaker peaks, thus limiting the possibility of
detecting spectral peaks; wrong graphical representation of the underlying continuous-
world function (if the aim of using DFT is a visualization of the CFT).

Hz
-2 -1 1 2

-2

-4

Figure 6.14: The DFT plot obtained by the student in Problem 6-19.

Hz
-2 -1 1 2

-2

-4

Figure 6.15: The DFT plot obtained by the student in Problem 6-20.
184 6. Issues related to leakage

Hz
-2 -1 1 2

-2

-4

Figure 6.16: The DFT plot obtained by the student in Problem 6-21.

6-23. The effects of convolution with the function R sinc(R u). Fig. 6.1 shows that the effect
of convolving a function G (u) with the function W (u) = R sinc(R u) consists of
smearing or smoothing out the original function G(u), and adding a ripple effect (i.e.
decaying oscillations). Fig. 6.2 shows a particular case of Fig. 6.1 in which G (u)
consists of a pair of impulses located at u = ±f. As explained in this figure, the effect of
convolution with W(u) = R sinc(R u) consists in this case of vertical scaling by R and
smearing each of the two impulses into a narrow sinc function, whose oscillations
gradually decay to both directions. A sharp-eyed student notices that while the
smearing and the ripple effects of the convolution with W (u) = R sinc(R u) are
explicitly mentioned in both of the figures, the vertical scaling by R is only mentioned
in Fig. 6.2, but not in Fig. 6.1. Is this merely an unintentional omission? If not, how do
you explain this difference? Hint: Looking at Fig. D.6(a) in Appendix D we see,
indeed, that the convolution of a non-impulsive function G(u) with W(u) = R sinc(R u)
does not cause a vertical scaling of G(u) by R even when R ≠ 1 (note that in Fig. D.6(a)
we have R = 4). Our student may find some elements of the response in Remark D.1
(see Sec. D.5 in Appendix D), and in Footnote 6 in Sec. 6.2.
6-24. The k frequencies f 1 ,...,f k are said to be harmonically related if each of them is an
integer multiple of a fundamental frequency f0, namely, if f1 = m 1 f0, ... ,fk = m k f0 with
m 1 ,...,m k ∈ Z (see, for example, [Manolakis11 p. 137]). In other references, the k
frequencies f1,...,fk are said to be commensurable if they are linearly dependent over Z,
i.e. if there exist integer numbers n 1 ,...,n k not all zero for which n 1 f1 + ... + n k fk = 0
(see, for example, [Amidror09 p. 113]). Are both terms equivalent? Hint: When k = 2 it
is easy to see that f1 = m 1 f0 , f2 = m 2 f0 implies that f0 = f1 /m 1 = f2 /m 2 and hence that
m2 f1 – m1 f2 = 0, and vice versa. Can you extend this to cases with k > 2?
 Chapter 7
Issues related to resolution and range

7.1 Introduction
In this chapter we explain how to choose optimal values for the various parameters that
are related to the resolution and the range of the DFT data, both in the signal and in the
spectral domains. We discuss the significance of these parameters, and show the
interconnections between them. Although the discussion is provided here for the 1D case,
it can be readily extended to any multidimensional case, too.

7.2 The choice of the array size

It is clear that a larger array size N allows us to increase the length R of the range within
which our given signal g(x) is sampled, and/or to increase the sampling rate of the signal
(i.e. to reduce the sampling interval Δx being used). This is clearly expressed by Eq. (4.2):
R = NΔx. Both of these options are, of course, welcome; and moreover, as we have seen in
Chapters 5 and 6, they can also contribute to the reduction of DFT artifacts such as
aliasing and leakage.
Since the array size N is common to both input and output arrays of the DFT, an
increase in the size of N also affects the DFT spectrum. And indeed, as we can see from
Eqs. (7.1)–(7.4) below, a larger array size N allows for an increase in the frequency range
length F of the DFT spectrum, and/or for an increase in the resolution of the DFT
spectrum (i.e. a smaller frequency step Δu). These are, indeed, the frequency-domain
counterparts of a reduction in the sampling interval Δx and of an increase in the sampling
range length R, respectively (see Eqs. (7.3) and (7.4) below).
In general, if the array size N is not limited by extraneous computational, storage or
other constraints, it should be taken sufficiently large to allow for the optimal values of R
and Δ x (or their spectral-domain counterparts, Δ u and F). For example, if the input
function g(x) has a finite duration, it would be desirable that the entire span of the function
be covered by R, in order to avoid truncation; and if g(x) is rather jumpy, it would be
advantageous to use a small sampling interval Δx. In any case, using a large size of N has
no major shortcomings (except, possibly, in terms of processing time or storage), so it is
almost always advisable to choose a sufficiently large value of N.
It should be remembered, however, that the choice of N may also have some
consequences in terms of the graphic representation of the input and output data (on a
computer display, a printed plot, etc.). For instance, in the 2D case, in order to reduce

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 185
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_7,
© Springer-Verlag London 2013
186 7. Issues related to resolution and range

jaggies along slanted lines in the signal domain (for example, when the input signal
consists of a slightly rotated 0/1-valued line grating like in Fig. 5.14) it may be desirable
to use an array of at least 512×512 elements. But on the other hand, if the displayed
elements (“pixels”) become too small, single-element details such as isolated impulses in
the DFT spectrum (like in the spectral domain of Fig. 5.14) may become too small and
therefore hardly visible in the 2D plot. If we want to make visible fine-detailed features
such as isolated impulses in the spectrum or isolated pixels in a 2D image, we need to
choose the array size accordingly, taking into account the physical dimensions of the final
displayed plot.

It is useful to note at this point that in many cases we do not really need to plot the entire
range used by the DFT, and we can, instead, only plot the most interesting part of the data.
This is indeed a good practice, since it allows us to show in greater detail the most
important part of the data and to omit all the rest; this can be done both in the input and in
the output arrays of the DFT. (Note however that for didactic reasons we prefer not use
this technique in the figures of this book, and unless mentioned otherwise we always show
the entire ranges of the input and output data.)

A final point to remind here is that in many FFT algorithms N must be a power of 2, so
if the actual data length N does not satisfy this requirement it may be necessary to increase
N to the next closest power of 2. Various possible ways of filling the input array when
increasing N, as well as their effects on the DFT results, are discussed in Appendix B.

7.3 The choice of the sampling interval

The smaller the sampling interval Δx, the larger the frequency range covered by the DFT
(see Eq. (4.9) or Eq. (7.4) below). On the other hand, for any given value of N, the larger
the sampling interval Δx, the higher the frequency resolution in the spectrum (meaning that
Δu is smaller) and hence, the better the ability to distinguish between closely spaced peaks
or impulses in the spectrum. This is a direct consequence of Eq. (4.4).

Note also that the choice of the sampling interval Δx may influence the aliasing in the
resulting DFT spectrum (see Chapter 5). If Δ x is chosen small enough so that the
sampling frequency fs = 1/Δx (Eq. (4.6)) is at least twice the highest frequency present in
the input signal, no aliasing will occur. But even when such a choice is not possible (e.g. if
the input signal is not band limited), still the smaller Δx, the less significant is the resulting
aliasing.

As we can see, we are confronted here with conflicting considerations: On the one hand,
there are good reasons for choosing Δx as small as possible. The tradeoff is, however, that
for a fixed N, smaller Δ x values imply a higher value of Δ u (i.e. a coarser spectral
resolution; see Eq. (4.4)), as well as a smaller sampling range length R (see Eq. (4.2)),
which may, in turn, increase the leakage effect. To solve this conflicting situation, N can be
increased as Δx is decreased in order to keep R and Δu unchanged.
7. Issues related to resolution and range 187

Finally, it is sometimes useful to make a first estimation of the optimal sampling interval
empirically: If possible, one can perform a series of DFT computations with successively
reduced sampling interval Δx, and choose the case which provides the best compromise
[Brigham88 p. 172].

7.4 The choice of the sampling range

Another parameter that can be considered is the sampling range of the original function,
i.e. the range xmin...xmax within which we sample the original function with a sampling
interval Δx, in order to fill the N-elements long input array of the DFT. If we denote the
length of the sampling range by R = xmax – xmin, we get (see Eq. (4.2)):

R = NΔx

(Note that since we have N samples, one could have argued that there exist only N–1
sampling intervals between them, so that we actually have R = (N–1) Δ x. But if we
remember that the input array is considered by the DFT as cyclical, and admit that each of
the N samples occupies along the continuous x axis an identical segment of width Δx, we
see indeed that R = NΔx.)

This means that any two of the three parameters N, Δx and R can be freely chosen, fixing
automatically the third one. And yet, even if N and Δx have already been chosen (thus
fixing R as well), we still have the freedom of choosing which R-units long part of the
original function, along the continuous x axis, will be sampled into our N-element long
input array. For example, as we have seen in Sec. 4.2, if the original function is symmetric
about the origin, it is natural to sample it within the almost symmetric range [–R/2...R/2);
but other choices can be also justified depending on the function at hand. In all cases, as
we have seen in Chapter 2, the sampled part of the function will be considered by the DFT
as cyclical, i.e. as an N-element period from a periodic sequence which repeatedly extends
to both directions ad infinitum, even if the original underlying function was not periodic.1

It should be noted that if the chosen sampling range is not sufficiently wide, so that it
truncates significant parts of the input function, a typical leakage artifact may occur in the
DFT spectrum (such as the ripple effect shown in Fig. 6.5). Therefore, if the input signal
g(x) is known to die out at a certain distance from the origin, it is highly recommended to
choose R such that most of the non-zero span of g(x) be included, as far as possible
without truncation. Similarly, if the input signal g(x) is known to be periodic and its period
is known, it is highly recommended to choose R to be equal to this period. As we have
seen in Chapter 6, this guarantees that no leakage artifacts occur in the resulting DFT

1 Note that choosing a different range having the same length R can be sometimes assimilated with a
cyclic shift in the input data. According to the shift theorem (see point 6 in Sec. 2.4.4 or [Brigham88
pp. 107, 114]), this results in a change in the phase of the spectrum, while its magnitude (absolute
value) remains unchanged (see also Secs. 3.3 and 8.3). However, this is no longer true if the new range
implies a different truncation of the input function so that the input data is not purely shifted.
188 7. Issues related to resolution and range

spectrum, so that all the impulses will be sharp. Finally, if the input signal g(x) is not
periodic and it does not die out outside a certain range, the best we can do is to choose a
sampling range that is as far as possible representative of what g(x) looks like along the x
axis. Intuitively, our aim when choosing the sampling range should be to capture the most
of the given signal, or at least the most significant parts of it.

Remark 7.1: In fact, if g(x) is periodic, Remark 6.1 guarantees that leakage does not
occur whenever the sampling range R is any exact integer multiple of the period. But in
reality, there is usually no advantage in choosing a sampling range R that consists of
several full periods of the input signal, since all of the periods are identical, and the
additional periods will provide no new information about the signal. Instead, it would
usually be better to decrease the sampling step Δx and get within the same N-element input
array a more detailed sampling of a smaller range R consisting of a single period of g(x).
Sampling m full periods of g(x) rather than one (within the same input array of length N)
will cause a loss of input information, because the new repeatedly sampled periods will
come to the detriment of a much denser sampling which could be obtained within a single
period. In terms of the DFT spectral domain, sampling m full periods of g(x) rather than
one will result in packing m–1 zeroes after each element in the output array (see the repeat
theorem in point 11a of Sec. 2.4.4). But since the output array length remains N we are
only left in the output array with N/m rather than N significant (i.e. non zero-packed)
spectral values. In other words, when sampling m full periods rather than one (within the
same input array of length N) we lose spectral information, since the higher Fourier series
coefficients will be pushed outside of the output array and lost.2
On the other hand, having several periods of g(x) within the sampling range R may have
some graphical advantages when plotting the input and output arrays of the DFT of g(x).
Consider, for example, Figure 7.1(a) which shows the input and the output arrays of the
DFT in the case of a periodic triangular function g(x). As we can see, the input array of the
DFT contains exactly 4 sampled periods of g(x), and indeed, in the output array each
impulse (Fourier series coefficient) is followed by 3 zero elements. In the signal domain,
having more than one period allows to clearly convey the periodic nature of g(x); and in
the spectral domain, the resulting zero packing allows to clearly see the impulsive nature of
the resulting spectrum. Had we chosen to include only one sampled period in the same N-
element input array, as shown in Fig. 7.1(c), the detail we would supply to the DFT about
g(x) would be 4 times finer, and hence we would get 4 times more Fourier series
coefficients in the DFT output. But on the other hand, the DFT output would then seem to
be compact and “continuous”, and it would not as clearly convey graphically the
impulsive nature of the spectrum (compare the DFT spectra in Figs. 7.1(a)–(c)). See also
Sec. D.9 in Appendix D on the interpretation of the DFT results as an approximation to a
CFT or as an approximation to a Fourier series decomposition, and the relations between

2 Note, however, that if sampling m full periods rather than one is also accompanied by an m-fold
extension of the input and output arrays from N to mN elements, so that we have no loss in terms of the
sampling density of g(x), then the resulting zero packing in the output array will not cause any loss of
information in the DFT spectrum, either; see Sec. B.5 in Appendix B, and row (e) in Fig. B.1.
7. Issues related to resolution and range 189

them. Appendix C provides further information on the number of periods in the sampling
range and its interconnections with the DFT spectrum. Q

Let us now consider the influence of widening the range R on the output of the DFT.
Doubling R (and Δx) while N remains fixed reduces by half the frequency range obtained
in the output array of the DFT (see Eq. (4.9)) but also the frequency step Δu (see Eq.
(4.4)). In other words, the N discrete frequencies included in the output array will be twice
denser but they will only cover half of the original frequency range. But if required (and
possible), N can be increased as R is increased in order to keep Δx unchanged.

Note, however, that doubling R has no influence on the amplitude of the DFT spectrum
(if the standard amplitude scaling correction explained in Sec. 4.4 is adapted to the new Δx
value). The only exceptions are cases involving an impulsive spectrum (where the impulse
heights do depend on R, as explained in Appendix A); and cases where doubling the
sampling interval Δx may introduce aliasing artifacts, as explained in Chapter 5.

As we can see, the choice of the sampling range is not always trivial, and it may involve
several conflicting considerations. Furthermore, it may also have some unexpected
consequences on the DFT output, as we have seen, for example, in Secs. 3.3 and 3.5.

Finally, it is worthwhile to note here, just as we did above in the case of Δ x, that
sometimes an estimation of the optimal sampling range can be also obtained empirically:
If possible, one can perform a series of DFT computations with successively increased
sampling range length R, and choose the case which provides the best compromise.

7.5 The choice of the frequency step and of the frequency range

The frequency-domain counterparts of the sampling interval Δx and the sampling range
length R are, respectively, the frequency step (or frequency interval) Δu and the frequency
range length F.

As we have seen in Eq. (4.4) above, the frequency step Δu is related to the sampling step
(sampling interval) Δx by the relationship Δu = 1 .
NΔx
As for the frequency range length F = fmax – fmin of the DFT spectrum, it is given by the
frequency-domain counterpart of Eq. (4.2) (see [Briggs95 p. 21] or [Gonzalez87 p. 96]):

F = NΔu (7.1)

(Note that from Eq. (4.9) we have F = fmax – fmin = 2 1 – f1 = NΔu – Δu = (N–1)Δu, but
2Δx
using the same considerations as in the beginning of Sec. 7.4 we can indeed say that
F = NΔu.)

The frequency range length F is also related to the sampling range length R of the input
signal by (see, for example, [Briggs95 p. 22]):
190 7. Issues related to resolution and range

RF = N (7.2)

This can be obtained from Eq. (4.5) by substituting therein Δx and Δu as obtained from
Eqs. (4.2) and (7.1); or alternatively, by multiplying Eqs. (4.2) and (7.1) and then using
Eq. (4.5). Inserting this result into Eqs. (7.1) and (4.2), respectively, we also obtain:

Δu = 1/R (7.3)

and: Δx = 1/F (7.4)

This means that the resolution in each of the two domains, the signal domain and the
spectral domain, is inversely determined by (i.e. reciprocal to) the range length of the
other domain. Note also that Eq. (4.5) means that the resolutions of the two domains are
inversely proportional. Similarly, Eq. (7.2) means that the range lengths of the two
domains (R and F) are inversely proportional.

Thanks to this duality between the signal-domain and frequency-domain parameters it is

clear that the frequency step Δu and the frequency range length F are fully determined by
their signal-domain counterparts, the sampling interval Δx and the sampling range length
R. However, in some cases it may be advantageous to consider things in terms of the
frequency domain parameters, or even in terms of both domains. Table 7.1 summarizes
the connections between the main signal-domain and frequency-domain parameters and
shows how they influence each other. As an illustration, one may refer to Figs. A.4–A.7 in
Appendix A.

For example, suppose we wish to refine (decrease) the frequency step Δu. As we can see
in Table 7.1, this can be done in two different ways: either by keeping N unchanged and
increasing R and Δ x, ot by keeping Δ x unchanged and increasing N and R (see also
[Briggs95 pp. 55–56]). In fact, as we can see in the table, such a double choice exists for
any paramerer change we may wish to perform — since any parameter change (increase
or decrease) appears in the column of the respective parameter twice, in two different rows.

Figure 7.1: The effect of choosing a sampling range length R that includes several full
periods of the given periodic function g(x). This is illustrated here using the
periodic triangular signal g(x) = triangwave(2x), which has the period P = 2
(i.e. frequency f = 12 ). Each row shows the sampled version of g(x) as well as
its resulting DFT (after all the required adjustments as discussed in Chapters
3, 4 and Appendix A). For the sake of comparison, the original continuous
function g(x) and its CFT G(u) are also shown in the same plots, using
continuous lines. As explained in Example 5.5, the CFT of g(x), G(u), is an
impulse train with impulse interval Δu = f = 12 whose amplitude is modulated
by half the CFT of a single triangular period tri(2x), namely by 14 sinc2(12 u).
(a) When the sampling range length is R = 8 (i.e. 4 periods of g(x)).
(b) When the sampling range length is R = 4 (i.e. two periods of g(x)).
 7. Issues related to resolution and range 191

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 0.4 n
2
g(x)
0.3 0
1.5 • • • Sampled 1
1 0.2
2
0.5 0.1 3
4 5 6 7
0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, R = 8

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 0.4 n
2
g(x)
0.3
1.5 • • • Sampled
1 0.2

0.5 0.1

0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
and vert. sca.
-1.5
-2. -1. 0 1. 2.
x -8. -6. -4. -2. 0 2. 4. 6. 8.
Hz
-1.5 -0.5 0.5 1.5
(b) N = 64, R = 4

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 k 0.4
0 8 16 24 32 40 48 56 63 n
2
g(x)
0.3
1.5 • • • Sampled
1 0.2

0.5 0.1

0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
x and vert. sca. Hz
-1.5
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1. -16. -12. -8. -4. 0 4. 8. 12. 16.
(c) N = 64, R = 2

(c) When the sampling range length is R = 2 (i.e. a single period of g(x)).
Note that if row (c) is understood as the DFT of the non-periodic function
h(x) = tri(2x), whose CFT is H(u) = 12 sinc2(12 u), then the impulse height
correction of Appendix A (division by R = 2) should not be applied to the
DFT output. This agrees, indeed, with Remark D.3 in Sec. D.9 of Appendix
D.
192 7. Issues related to resolution and range

Finally, let us note that Remark 4.1 provides a further useful relationship involving the
frequency range length F — it turns out that F is equal to the sampling frequency fs:

fs = NΔu = F (7.5)

This result can be also obtained by Eq. (7.4), which gives us F = 1/Δx. According to Eq.
(4.6) this is precisely the sampling frequency fs of the input signal, so that indeed F = fs.

The list of the main relations that is provided at the end of this book (before the list of
notations and symbols) can be used as a quick reference to the various relations that exist
between the DFT parameters in the signal and frequency domains.

Case Signal domain Frequency domain Examples/

No. N R Δx F Δu Figures
1 unchanged doubled doubled halved halved Fig. A.4
2 doubled doubled unchanged unchanged halved Fig. A.5
3 doubled unchanged halved doubled unchanged Fig. A.6
4 unchanged halved halved doubled doubled
5 halved halved unchanged unchanged doubled
6 halved unchanged doubled halved unchanged Fig. A.7

Table 7.1: The interconnections between the various DFT parameters in the image
and frequency domains, based on the duality established by Eqs. (4.2),
(4.4), (7.1), (7.2). The table shows how doubling (or halving) any of
these parameters may influence the others. Cases 4–6 are reciprocal to
cases 1–3.

PROBLEMS

7-1. Student A says that by reducing the sampling interval Δ x in the signal domain we
improve the quality of the sampling, and therefore we also improve the quality of the
frequency resolution in the spectral domain. Student B argues that this reasoning
contradicts Eq. (4.4), Δx Δ u = 1/N, according to which improving Δx gives a worse Δ u,
and vice versa. Can you help them find a way out of this contradiction?
7-2. The interconnections between the DFT parameters Δx, Δu, R, F and N. Make a list of
the various relations you can think of between the DFT parameters Δx, Δu, R, F and N.
Note that some of these relations involve three parameters, like R = N Δ x, but others
involve only two, like Δ u = 1/R. Explain the significance of each of the relations in
your list. (See also the list of the main relations at the end of this book, before the list
of notations and symbols.)
7. Issues related to resolution and range 193

7-3. Consider Eq. (7.3), Δ u = 1/R. How can it be that the DFT spectrum resolution Δ u does
not depend on N, but only on R? Is it possible that for all values of N we get the same
spectrum resolution Δu if R remains unchanged? Hint: See row 3 in Table 7.1.
7-4. What are the 2D and MD counterparts of each of the relations between the DFT
parameters Δx, Δu, R, F and N? Are there any new relations in the 2D or MD cases that
have no equivalents in the 1D case?
7-5. Explain Remark 7.1 and Footnote 2 therein in light of the stretch and the repeat
theorems (see points 11 and 11a in Sec. 2.4.4). What is the equivalent in the 2D case?
7-6. We have seen in Remark 7.1 that when sampling m full periods of an original periodic
signal g(x) rather than one (within the same input array of length N) we lose spectral
information, since the higher Fourier series coefficients will be pushed outside of the
output array and lost (see also Fig. 7.1). A student argues that because the output array
of the DFT is cyclical, the Fourier series coefficients that are pushed outside the end of
the output array are not really lost: Rather, they re-enter from the opposite end of the
output array, just as we have seen in Chapter 5 (see, for example, Fig. 5.11). What is
your opinion?
7-7. Looking at Fig. 7.1, Student A comes to the conclusion that given a periodic signal
g(x), by increasing the number of its periods in the input array of the DFT one obtains
a better spectral resolution in the DFT output. For example, in row (b) of Fig. 7.1 the
64-element DFT spectrum shows the frequency range between –8...8 Hz, while in row
(a) the same 64-element DFT spectrum shows in a higher resolution the smaller
frequency range between –4...4 Hz. This also agrees with row 1 in Table 7.1. However,
Student B claims that no gain of spectral information can be obtained by sampling a
larger number of periods which are all strictly identical. Furthermore, since the
sampling resolution in row (a) of Fig. 7.1 is lower than in row (b), no gain of spectral
information occurs when passing from (b) to (a), but rather a loss of spectral
information. What is your opinion?
7-8. Student A claims that Fig. 7.1 does not respect the repeat theorem as given in point 11a
of Sec. 2.4.4, since in this figure the number of repetitions L does not influence the
impulse heights in the spectral domain. Student B says that this could be explained if
Fig. 7.1 were based on the alternative DFT definition, as in the column “Alt DFT” in
Sec. 2.4.4, but student A reminds him that in the present book we are not using the
alternative DFT definition. Is Fig. 7.1 flawed by an error?
7-9. Most rows in Table 7.1 provide in the last column an example (a figure in this book)
which illustrates the case in question. Complete the table by providing figures of your
own to illustrate the remaining rows of the table.
7-10. From row 4 of Table 7.1 we learn that by improving the sampling resolution in the
signal domain (using a smaller sampling interval Δ x) we lose in spectral resolution,
since the resulting frequency interval Δ u is larger. How can it be? See also Problem
7-1.
7-11. Suppose you wish to find the DFT spectrum of an unknown signal that you obtained
from some astronomical measurements. How would you choose the various DFT
parameters?
7-12. Suppose you suspect that there exists some hidden periodicity in the astronomical
signal of the previous problem.
(a) How can you check if your assumption is correct?
(b) How can you determine the periodicity in question (or, equivalently, the
corresponding frequency)?
194 7. Issues related to resolution and range

(c) How can you optimize your DFT plot if you know the periodicity (or the
frequency) in question?
(d) How can you optimize the DFT plot if there exist in the given signal several
different periodicities?
7-13. Twin peaks. Suppose that you observe in the DFT output a peak having two very close
summits. While this “M-shaped” peak may truly be part of the desired spectrum, one
may also suspect that such a spectral peak reflects in fact the existence of two closely
located individual impulses that are simply merged together due to insufficient spectral
resolution (see, for example, Fig. 9-3 in [Smith03 p. 173]). How would you increase
the spectral resolution in order to resolve this uncertainty?
7-14. The interconnections between the sampling frequency fs and the other DFT
parameters. Make a list of the various relations you can think of between the sampling
frequency fs and the DFT parameters Δ x, Δ u, R, F and N. Explain the significance of
each of the relations in your list.
7-15. What are the 2D and MD counterparts of each of the relations between the sampling
frequency fs and the DFT parameters Δ x, Δ u, R, F and N? Are there any new relations
in the 2D or MD cases that have no equivalents in the 1D case?
7-16. Is it possible to sample a 1D signal using a non-uniform sampling interval (for
example, logarithmic sampling) and apply DFT to the resulting sampled signal? What
continuous-world spectrum would you expect this DFT to match? Hint: On the
extension of the 1D DFT to irregularly spaced data see, for example, [Press02, Sec.
13.8], [Bagchi99], [Marvasti01, Chapter 7], [Henson92] and [Dutt93].
7-17. Is it possible to sample a 2D signal using a circular or radial sampling pattern, and
apply DFT to the resulting sampled signal? Hint: On the computation of 2D DFT on
irregular grids, and more particularly in the case of polar coordinates, see [Briggs95
pp. 292–299]. On the extension of the sampling theorem to polar coordinates see
[Stark79], [Stark82]. See also [Marvasti01], which is an excellent source of information
on non-uniform sampling and non-uniform DFT in 1D and 2D settings, and includes
many further references.
7-18. The CFT is said to preserve a given coordinate transformation if the application of
that coordinate transformation to any CFT input affects the CFT output (the spectrum)
by exactly the same coordinate transformation. For example, the 2D CFT preserves
rotations since rotating the CFT input g(x,y) by angle θ results in an equal rotation of
the spectrum G(u,v). The 2D CFT is therefore a rotation-preserving transform (see in
the Glossary the difference between rotation-preserving and rotation-invariant).
(a) Does the CFT (1D, 2D or MD) generally preserve linear (or non-linear) coordinate
transformations? Hint: Consider, for example, the dilation rule in Secs. 2.4.1–2.4.3.
See also Sec. 10.3 in [Amidror09].
(b) What about the DFT? Hint: Unfortunately, even in the rare cases where the DFT
does preserve a given coordinate transformation, the DFT usually loses this property.
For example, although 2D CFT preserves all rotations, DFT only preserves rotations
through integer multiples of 90° (see, for example, [Park09]).
(c) What happens under 2D CFT or 2D DFT to the polar coordinate transformation?
7-19. Suppose you are given a continuous 2D signal having a circular symmetry (such as
the functions shown in Fig. 6.8(c) or Fig. 6.8(d)). A student suggests to apply to this
signal a polar coordinate transformation, then sample the signal in the r,θ space using a
standard orthogonal sampling grid along the r and θ axes, apply DFT to this 2D
discrete signal, and finally apply an inverse coordinate transformation to obtain the
spectral samples of the original 2D signal. What is your opinion?
 Chapter 8
Miscellaneous issues

8.1 Introduction

In this chapter we group together some further questions and potential pitfalls that may
await the unaware DFT users, but which do not fit into any of the previous chapters. In
Sec. 8.2 we discuss the representation of discontinuities in the discrete world, in the input
and output arrays of the DFT. In Sec. 8.3 we briefly review issues and artifacts that are
related to the phase when using the polar (i.e. magnitude and phase) representation of the
complex-valued spectrum. Then, in Sec. 8.4 we discuss symmetry related issues and the
DFT artifacts that may result thereof. In Sec. 8.5 and 8.6 we discuss two artifacts that may
occur due to sampling and reconstruction of continuous-world signals, namely, jaggies
and sub-Nyquist artifacts, respectively. These two discrete-world artifacts are not specific
to DFT, but they do occur in DFT applications, too, and we seize the opportunity offered
by their explanation here to illustrate some questions related to DFT spectra and their
continuous-world counterparts. Then, in Sec. 8.7 we review the main displaying
considerations that should be taken into account in order to obtain a correct graphic
representation of the DFT results, and finally in Sec. 8.8 we review very briefly some
numeric precision considerations.

8.2 Representation of discontinuities

Consider the function g(x) = sinc(x), whose spectrum is the unit square pulse G(u) =
rect(u). If we look carefully at the DFT of g(x) (see Fig. 8.1), we note that at both of the
pulse discontinuities (i.e. the transitions from 0 to 1 and then from 1 back to 0) the DFT
output contains a point having the midvalue height 12 . It turns out that this is, indeed, the
general rule in discontinuities: the value of a function at a discontinuity is simply taken to
be the midvalue [Brigham88 p. 167], both in the signal and in the spectral domains.

The “midvalue rule” which defines the behaviour at discontinuities is fundamental and
omnipresent in the Fourier theory. This rule already exists in the two main branches of the
continuous-world Fourier theory, namely, in the CFT and in the Fourier series
development of periodic functions. In the case of CFT this rule is required, when g(x)
contains discontinuities, in order for the inverse Fourier transform to hold [Brigham88 pp.
22, 167; Bracewell86 p. 7]. For this reason, if a continuous-world function g(x) has a
discontinuity at the point x, it is always recommended to explicitly define its value there to
be equal to the midvalue 12 [g(x+) + g(x–)], even if this subtlety seems to be superfluous as

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 195
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_8,
© Springer-Verlag London 2013
196 8. Miscellaneous issues

it does not really affect the evaluation of the Fourier integral in Eq. (2.1).1 This midvalue
rule is even more significant in the case of Fourier series, since it turns out that the Fourier
series associated with a periodic function p(x) converges at any discontinuity point x of
p(x) to the midvalue 12 [p(x+) + p(x–)] (see [Bracewell86 p. 205; Briggs95 pp. 37–38]).

It is therefore not really surprising that a similar rule exists in the discrete world, too, in
the case of DFT. And indeed, as already mentioned in references such as [Briggs95 pp.
38, 93–95] and [Brigham88 p. 167], the midvalue rule should always be observed when a
function is sampled for input to the DFT. For example, when sampling the function g(x) =
rect(x) (see Fig. 8.2(a)), the two sampling points that fall on its discontinuities x = –12 and
x = 12 should be assigned the midvalue 12 , and only their neighbouring samples will actually
take the values of 0 and 1, respectively.2 The same is also true in the 2D and MD cases.
For example, as shown in Fig. 3.6(a), the value of the 2D unit rectangle function g(x,y) =
rect(x) rect(y) (i.e. the 2D counterpart of g(x) = rect(x)) at the discontinuities along its four
edges is neither 0 nor 1 but rather the midvalue 12 , except at the four corners where the
value is 12 · 12 = 14 (since g(x,y) is the product of the two perpendicular bars g1(x,y) = rect(x)
and g2(x,y) = rect(y), each of which has the value 12 along its two edges).

Furthermore, because the DFT input is considered to be cyclical (i.e. one N-element
period from a periodic discrete signal that repeats itself to both directions ad infinitum),
the first element and the last element of the input array are in fact neighbours. For this
reason, the midvalue rule should also be applied to a discontinuity which occurs between
the two ends of the input array. For example, when sampling the function g(x) =
e–x step(x) (see Figs. 3.1, 3.3), the sampled value at the discontinuity point x = 0 should be
defined as the midvalue 12 [g(x+) + g(x–)] = 12 , even if the discontinuity is located at the left
end of the input array [Brigham88 p. 168]. This is also true in the 2D and MD cases. And
due to the reversibility of the DFT, it is clear that the same midvalue rule applies both to
the signal and to the spectral domains, whenever a discontinuity occurs.

Note, however, that the midvalue rule is not systematically followed in the DFT
literature: Some references such as [smith03 pp. 212–217] do not apply it, while others
use it selectively (for example, [Brigham88] applies this rule in the 1D case [Brigham88
pp. 122, 168], but not in the 2D case [Brigham88 p. 242]). And indeed, because this rule
only concerns the value of few points of the sampled signal, it may be legitimately asked
whether this subtlety is really so important, and what happens if we forget to apply it. Let
us answer this question with the help of a concrete example.

Consider Fig. 8.2. Row (a) of this figure shows a sampled version of the unit square
pulse function g(x) = rect(x), which correctly takes into account the midvalue rule at both

1 Note that in the context of CFT some references such as [Bracewell86 pp. 52–53] do not really insist
on this midvalue rule, while others do, like [Brigham88 pp. 13–18] or [Gaskill78 pp. 42–43] for the
1D case and [Gaskill78 pp. 67, 71] for the 2D case.
2 Note that this gradual transition at a discontinuity has nothing to do with the ramp-like graphic

presentation of a sharp transition in the signal when drawing it as a connected line plot (Sec. 1.5.1).
While the former is a true property of the discrete signal, the latter is just a graphical plotting artifact.
 8. Miscellaneous issues 197

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(x)
1 • • • Sampled 1

0.5 0.5

0 0
CFT
-0.5 -0.5 • • • DFT after
i&o reorg.
-1 x -1
and vert. sca. Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.

Figure 8.1: The continuous-world function g(x) = sinc(x) and its spectrum G(u) =
rect(u), and their discrete counterparts in the input and output arrays of the
DFT. Two interesting phenomena can be observed here in the DFT
spectrum: (1) The pulse discontinuities in the DFT spectrum (both the
transitions from 0 to 1 and from 1 to 0) contain an array element having
the midvalue height 12 . (2) Because the sampling range length R = 8 in the
signal domain is not sufficiently large, it truncates significant parts of the
sinc function and leaves them outside the scope of the input array.
Consequently, a typical ripple effect (similar to the Gibbs phenomenon)
occurs about both sharp transitions of the square pulse in the spectral
domain, due to leakage. This ripple effect is discussed in Sec. 5.7 (see
Fig. 5.19) and in Example 6.2.

of the discontinuity points x = ±12 . As we can see in the spectral domain of row (a), the
resulting DFT output (after the corrections of Chapters 3 and 4) is a discrete function that
closely approximates the CFT G(u) = sinc(u), up to some aliasing error (which is
unavoidable since our function g(x) is not band limited).3 Row (b) of the figure shows
what happens to the DFT output when we do not apply the midvalue rule in the input
array, and the discontinuity jumps directly from 0 to 1 and vice versa. As we can see, the
resulting DFT spectrum in this case differs from the CFT G(u) much more than in row
(a): Note the significant error in the height of the resulting DFT and in the location of its
zero crossing points. Row (c) of the figure shows another discrete variant of g(x) having
direct jumps from 0 to 1 and back, but this time, unlike in row (b), the discrete samples at
the discontinuities have the value 0 rather than 1. Note that in this case, too, the resulting
DFT spectrum significantly differs from the CFT G(u) = sinc(u). Another example
illustrating the errors that are introduced to the DFT output when we do not apply the
midvalue rule can be found in [Briggs95 pp. 93–95].

3 In order to obtain in the DFT spectrum the exact, unaliased function G(u) = sinc(u), we have to take in
the DFT input the rippled version of the unit pulse, which is precisely its band-limited (i.e. low-pass
filtered) version. The DFT of this function is simply a truncated (and therefore unaliased) version of the
true function G(u) = sinc(u). This can be illustrated by Fig. 8.1 if we reverse the roles of its signal and
spectral domains.
198 8. Miscellaneous issues

It should be noted, however, that in the discrete case the behaviour of the midvalue rule is
often more subtle than what we have seen so far. To see this, let us study in more detail the
situation in the discrete case, when a discontinuity occurs in the spectral domain. The
reason for this choice is that the behaviour at the discontinuity in the spectral domain is
automatically taken care of by the DFT itself. Thus, thanks to the reversibility of the DFT,
we can learn what the DFT does in spectral-domain discontinuities, in order to apply the
same rule when a discontinuity occurs in the input domain. Let us consider for this end
Fig. 6.5 in Chapter 6. Note that in all rows of this figure the sampling interval is
Δx = 0.25, and the frequency step between consecutive elements of the output array is
Δu = 0.125. Row (a) of this figure is identical to Fig. 8.1, and shows the sampled version
of the function g(x) = sinc(x), as well as the corresponding DFT output array (after
applying the corrections of Chapters 3 and 4). This output array is the DFT
approximation to the continuous-world spectrum of g(x), the unit square pulse G(u) =
rect(u). And indeed, we see that both of the pulse discontinuities (i.e. the transitions from
0 to 1 and then from 1 back to 0) fall on an array point whose height is the midvalue 12 .
More precisely, if the cutoff frequency of the square pulse, 12 Hz, falls exactly on the n-th
element of the output array, then the n-th element itself will contain the midvalue 12 , and
only its two neighbours, the elements n–1 and n+1, will contain the values 0 and 1.4
Consider now row (c) of Fig. 6.5, which shows the function g(x) = sinc(2ax) and its CFT
G(u) = 2a1 rect(2au ) (a square pulse of height 2a1 extending between –a and a), as well as their
discrete counterparts, when a = 18/32. In this case the cutoff frequency is not exactly
a = 12 = 4Δu as in row (a), but rather a = 18/32 = 4.5Δu. This means that in this case the
cutoff frequency does not fall exactly on an element of the output array but rather halfway
between two consecutive output array elements. Similarly, in rows (b) and (d), where a
equals 17/32 = 4.25Δu and 19/32 = 4.75Δu, respectively, the cutoff frequency falls again

Figure 8.2: The continuous-world function g(x) = rect(x) and its spectrum G(u) =
sinc(u), and their discrete counterparts in the input and output arrays of
the DFT. (a) Using the correct discrete representation of the input
function, where each of the pulse discontinuities (the transitions from 0 to
1 and from 1 to 0) contains an array element having the midvalue height 12 .
Note that the DFT spectrum is slightly corrupted by aliasing, as we can
see by comparing it with the true CFT, G(u) = sinc(u), which is drawn by
a continuous line. (b) Using a wrong discrete representation of the input
function g(x) = rect(x) where both of the pulse discontinuities take the
value 1, so that the transitions are done directly from 0 to 1 and from 1 to
0. (c) Using another wrong discrete representation of the input function
where both of the pulse discontinuities take the value 0. Note that the DFT
spectra obtained in rows (b) and (c) are significantly more corrupted than

4 Possibly with some discrepancies due to a typical ripple effect, similar to the Gibbs phenomenon, that
occurs in both sharp transitions of the square pulse in the spectral domain, due to leakage. This ripple
effect is discussed in Sec. 5.7 (see Fig. 5.19) and in Example 6.2.
 8. Miscellaneous issues 199

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x) = rect(x) CFT
1 1
• • • Sampled • • • DFT after
correctly input reorg.
0.5 0.5 and vert. sca.
0 0

-0.5 -0.5

-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

(a)

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x) = rect(x) CFT
1 • • • Sampled 1 • • • DFT after
incorrectly input reorg.
0.5 0.5 and vert. sca.

0 0

-0.5 -0.5 CFT of

g1(x) = rect(x/1.25)
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1. 0 1. 2.
Hz
-1.5 -0.5 0.5 1.5
(b)

Signal domain Spectral domain: Real part

1.5
0 4 8 12 16 20 24 28 31 k 1.5
0 4 8 12 16 20 24 28 31 n
g(x) = rect(x) CFT
1 • • • Sampled 1 • • • DFT after
incorrectly output reorg.
0.5 0.5 and vert. sca.

0 0

-0.5 -0.5 CFT of

g2(x) = rect(x/0.75)
-1 x -1 Hz
-4. -3. -2. -1. 0 1. 2. 3. 4. -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
(c)

in row (a): their heights as well as their zero crossing points are incorrect.
The array length (number of samples) being used in all of the rows is
N = 32 and the sampling range length is R = 8, so that the sampling
interval is Δx = 8/32 = 0.25. The frequency step in the DFT output is,
according to Eq. (7.3), Δu = 1/R = 0.125.
200 8. Miscellaneous issues

between two consecutive elements, but this time not exactly halfway between them but
rather closer to one or to the other, depending on the case. As we can see, when the
discontinuity falls between two consecutive array elements, there exists no longer an
array element having the average value 12 ; instead, the two neighbouring array elements
between which the discontinuity is located get intermediate height values between the
heights 0 and 1, in accordance with the relative distance of these array elements from the
discontinuity.

We see, therefore, that the behaviour near a discontinuity is more subtle in the discrete
world than in the continuous world: While in the continuous world the discontinuity point
itself is included in the domain of the function, and takes there exactly the average value, in
the discrete case the discontinuity (of the underlying continuous-world function) may fall
either exactly on a discrete array element or anywhere between two neighbouring array
elements, and the function height at these array elements is determined accordingly.

As we see in Fig. 6.5, this discrete-world midvalue rule is automatically taken care of by
the DFT whenever a discontinuity occurs in its output array. But thanks to the reversibility
of the DFT, it is clear that the situation must remain the same when we reverse the roles of
the input and the output arrays. Therefore, when the discontinuity occurs in the input array
of the DFT, the very same rule must be observed. But this time, it is our responsibility
to apply this rule when we generate the input array, meaning that we must sample
the input discontinuities correctly; otherwise the resulting DFT will be somewhat
corrupted.

Returning to Fig. 8.2, we can now understand why the difference between the resulting
DFT spectrum and the CFT G(u) = sinc(u) is bigger in row (b) than in row (a): Simply,
the discrete samples in the signal domain of row (b) do not represent correctly the
continuous function g(x) = rect(x). Indeed, just as the DFT automatically applies the
midvalue rule to its output array, so does it also expect that the same rule be satisfied in its
input array. Therefore, in row (b), the DFT assumes that the discontinuities in its input
array actually occur between the consecutive samples having the values 0 and 1 (or 1 and
0). This means that the discrete samples in the signal domain of row (b) are not interpreted
by the DFT as samples of a continuous-world square pulse having the width 1, but rather
as samples of a slightly wider continuous-world square pulse that extends half a sampling
interval further to each direction. The width of this square pulse is therefore not 1 but
rather 1 + Δx, which means in our case 1.25 (since Δx = 0.25). The continuous-world
function represented by this sampled signal is therefore not g(x) = rect(x), but rather
g1(x) = rect(x/1.25), whose CFT is G1(u) = 1.25sinc(1.25u) rather than G(u) = sinc(u).
And indeed, the DFT spectrum in row (b) corresponds well to the function G1(u), which is
plotted in our figure using a dashed curve. Note that the small discrepancy between our
DFT and G 1(u) is basically due to aliasing, which remains unavoidable in the DFT
spectrum because g1(x), too, is not band limited. Similarly, in row (c) the discrete samples
in the signal domain are not interpreted as a continuous-world square pulse having the
width 1, but rather as a slightly narrower continuous-world square pulse that extends half
8. Miscellaneous issues 201

a sampling interval less to each direction. The width of this square pulse is therefore
1 – Δx = 0.75. Consequently, the continuous-world function represented by this sampled
signal is not g(x) = rect(x), but rather g 2 (x) = rect(x/0.75), whose CFT is G 2 (u) =
0.75sinc(0.75u) rather than G(u) = sinc(u). And indeed, the DFT spectrum in row (c)
matches well (up to some minor error mainly due to aliasing) the function G2(u), which is
plotted in our figure using a dashed curve.

Note, however, that this explanation of the situation in rows (b) and (c) of Fig. 8.2 is
only a close approximation. In fact, the correct sampling of a square pulse of width 1 + Δx
or 1 – Δx, whose discontinuities fall between input array elements, should not assign to the
two array elements adjacent to a discontinuity the values 0 and 1. Rather, these elements
should get some intermediate values, depending on their relative distance from the true
discontinuity point (see Fig. 6.5). This suggests that in rows (b) and (c) of Fig. 8.2 the
small discrepancy between the DFT results and the continuous spectra G1(u) or G2(u) is
not only due to aliasing, but also due to this rather negligible sampling inaccuracy.

In conclusion, the discrete counterpart of the midvalue rule can be formulated as

follows: If a discontinuity of a continuous-world spectrum G(u) falls exactly on the n-th
discrete element of the DFT output array, then this element takes the midvalue height, and
only the elements n–1 and n+1 really reflect the spectrum heights G(u–) and G(u+) to
both sides of the jump. But if the discontinuity of the continuous-world spectrum falls
between the discrete elements n and n+1, then these discrete elements take intermediate
values between G(u–) and G(u+) depending on the relative distance of these two elements
from the true discontinuity point. This rule is automatically applied by the DFT in the
output array whenever a discontinuity occurs in the spectrum (as we have seen in Fig.
6.5). But when it is our role to sample a continuous-world function having one or more
discontinuities as input to the DFT, the application of this midvalue rule during the
sampling process is clearly less obvious. The best we can do in this case is to check where
the discontinuity falls: If it falls very close to an input array element, then we should
assign to this element the midvalue height. But if the discontinuity falls between the input
array elements n and n+1, we should reasonably estimate (interpolate) the intermediate
heights at these two input array elements in accordance with their respective distance from
the discontinuity point between them. Brute-force sampling without taking the midvalue
rule into account will assuredly infect the DFT results with some errors, which may be
more or less significant depending on the case.

Remark 8.1: It is important to note, however, that the midvalue rule does not concern the
case of impulses. If we carefully observe the DFT spectrum of a function such as g(x) =
cos(2πfx) (see Fig. 4.2 or Fig. A.2 in Appendix A) we will see that its impulses jump
directly up and down, with no midvalues to the left and to the right of the peak. This could
be indeed expected, since the width of an impulse (that falls precisely on an array element)
is a single array element; see Appendix A for more details on impulses in the continuous
and discrete worlds. Q
202 8. Miscellaneous issues

8.3 Phase related issues

Like any other complex-valued entity, the spectrum G(u) of a function g(x) can be
represented either by its real-valued and its imaginary-valued parts:
G(u) = Re[G(u)] + i Im[G(u)] (8.1)

or, using the polar notation, by its magnitude (also called absolute value or modulus) and
its phase (also called argument):

G(u) = Abs[G(u)] · ei Arg[G(u)] (8.2)

where: Abs[G(u)] = Re[G(u)]2 + Im[G(u)]2 (8.3)

Arg[G(u)] = arctan Im[G(u)] (8.4)

Re[G(u)]
(Note that Re[G(u)], Im[G(u)], Abs[G(u)] and Arg[G(u)] are all real-valued functions of
the real variable u.) Both of these representations can be used, of course, in the continuous
case as well as in its discrete counterpart.

Although the representation of the spectrum in terms of its real-valued part and its
imaginary-valued part is rather straightforward, often the polar representation in terms of
magnitude and phase is more convenient and easier to understand [Smith03 pp. 163–164].
For example, a shift of a in the signal-domain function g(x) usually changes both
Re[G(u)] and Im[G(u)], in a rather obscure way; but in the polar representation, thanks to
the shift theorem, the magnitude-spectrum Abs[G(u)] remains unchanged and only the
phase-spectrum Arg[G(u)] is affected by a linear increment of ϕ(u) = 2πua [Smith03 p.
188–190], [Smith07 pp. 149–150].5

However, in spite of its usefulness, the graphic presentation of the phase spectrum
suffers from several shortcomings. For example, we can mention its 2π ambiguity, and its
high numeric instability in zones where the magnitude is very small, which causes violent
but meaningless oscillations. These and other pitfalls related to the phase have already
been explained in detail in the literature, and the interested readers are encouraged to
consult references such as [Smith03 pp. 164–168] or [Smith07 pp. 180–181].

Finally, let us mention here one more point that is also related to the phase of the DFT
spectrum. As we have seen in Chapter 2, there exist in the literature many variants of the
CFT and DFT definitions, which may differ, among other things, in the sign of the
exponential part. Using a DFT definition with reversed signs in the exponential part will
result in a reversed list of values in the output [Wolfram96 p. 868]; more precisely, the
positive and negative frequencies will be interchanged (since a sign reversal of the

5 Note that the phase of the spectrum G(u) is a straight line through the origin u = 0 if and only if the
signal-domain function g(x) is symmetric around some point x0. If the point of symmetry is x0 = 0 (i.e.
the origin) the slope of the phase line is zero, and the phase of the spectrum is identically zero; if x0 > 0
the slope of the phase line is negative, and if x0 < 0 the slope is positive.
8. Miscellaneous issues 203

exponent can be assimilated with a sign reversal in the index n within the exponent, where
due to the cyclic nature of the DFT we have –n = N – n [Smith07 pp. 118–119]).

Note that if the original input signal is real-valued (which is most often the case for a
real-world signal), then its spectrum is Hermitian, meaning that the real-valued part of the
spectrum is even while its imaginary-valued part is odd. In this case, using a DFT
definition with the opposite exponential sign convention will simply change the sign of the
imaginary-valued part of the resulting spectrum, without affecting its real-valued part (this
is a straightforward consequence of the output frequency reversal mentioned above, when
the spectrum is Hermitian). And if we prefer to use the polar representation of the
spectrum, the sign inversion in the exponent will only cause a sign inversion in the phase
of the resulting spectrum, without affecting its magnitude (this simply results from Eqs.
(8.3)–(8.4)). Thus, when the DFT implementation being used is based on a different
exponential sign convention the user must remember to change the sign of the phase (or
the imaginary part) of the resulting DFT, in addition to all the other required corrections
according to Chapters 3 and 4. This is the case, for example, when the DFT algorithm we
use is based on a DFT definition with a positive exponent sign in the direct transform (like
in the Mathematica® software package [Wolfram96 p. 868]), while the corresponding
continuous-world results being used have been obtained using a CFT definition with a
negative exponent sign in the direct transform (as in [Bracewell86 p. 6]).

8.4 Symmetry related issues

The continuous Fourier transform has some very useful symmetry properties (see, for
example, [Bracewell86 pp. 14–16]). In particular, if the given function g(x) is even (i.e.
symmetric about the origin: g(–x) = g(x)) then its spectrum G(u) is also even, and if g(x) is
odd (antisymmetric about the origin: g(–x) = –g(x)) then G(u) is odd, too. Furthermore, if
g(x) is real-valued (which is often the case for a real-world signal), then its spectrum is
Hermitian (meaning that its real-valued part is even and its imaginary-valued part is odd,
which is succinctly expressed by G(–u) = G*(u)); and in particular:

(a) If g(x) is real-valued and even, so is its spectrum G(u);

(b) If g(x) is real-valued and odd, its spectrum G(u) is imaginary-valued and odd.

These useful symmetry rules are also preserved in the discrete world, when we consider
the DFT (after the usual reorganizations) rather than the continuous Fourier transform;
see, for example, [Bracewell86 p. 366] or [Smith07 pp. 142–148]. Interestingly, however,
the situation in the discrete case is somewhat more subtle. We illustrate this by the
following two examples:

As a first example, note that a symmetric rect function in the discrete world must have an
odd number of 1-valued elements across its width, namely, the center of the rectangle
(which corresponds to x = 0), and to each of its sides an equal number of neighbouring
204 8. Miscellaneous issues

1-valued elements (usually followed by an element with the midvalue 12 , as we have seen in
Sec. 8.2). Any other possibility will result in a non purely-real DFT.6

As a second example, note that in the continuous world when g(x) is real valued, G(0) is
real valued, too (since the Hermitian nature of G(u) implies that Im[G(0)] = –Im[G(0)],
which means that Im[G(0)] = 0). However, in the discrete world it turns out that when the
input signal of the DFT is real valued, not only the output element corresponding to the
zero frequency is real valued, but also the output element representing the highest (or
lowest) frequency in the spectrum, which is located N/2 elements away from the zero-
frequency element [Briggs95 pp. 77, 120]; see, for example, the leftmost spectral element
in Figs. 3.1(d) and 3.3(f). This is explained in Problem 8-18. It follows therefore that if
the DFT spectrum does not have a discontinuity on its border, its imaginary part must
gradually go down to zero when approaching the spectrum borders, even if its continuous
counterpart behaves differently (see, for example, Fig. 5.8(b), where this deviation of the
DFT with respect to the CFT is explained by aliasing).

Let us now proceed to the multidimensional case. Once again, as we have done above in
the 1D case, we will first consider the continuous world, and only then return to the
discrete world and the DFT.

From the point of view of symmetry, the 2D and MD cases are much more interesting
than the 1D case, since they have many more possible types of symmetry. To start with, a
multidimensional function may have a different line-symmetry along each of its axes. For
example, in the 2D case, the function g(x,y) = y is even in the x direction (symmetric with
respect to the y axis) and odd in the y direction (antisymmetric with respect to the x axis):
g(–x,y) = g(x,y), g(x,–y) = –g(x,y); while the function g(x,y) = ey has only one symmetry (it
is even in the x direction). However, in the multidimensional case there may also exist line-
symmetries with respect to lines other than the main axes; and in addition, there may also

Figure 8.3: Symmetry properties of the multidimensional Fourier transform of a real-

valued signal g(x). This figure was generated by 2D DFT with N = 64.
(a),(b) If g(x) is real-valued and centrosymmetric, so is its spectrum G(u).
(c) If g(x) is real-valued and anticentrosymmetric, its spectrum G(u) is
imaginary-valued and anticentrosymmetric. (d) If g(x) is real-valued but
has any other symmetry (or no symmetry at all), its spectrum G(u) has a
centrosymmetric real-valued part as well as an anticentrosymmetric
imaginary-valued part (this is the multidimensional generalization of a
Hermitian function). The functions used here are as follows (with f = 1):
(a) g(x,y) = cos(2π f [xcos θ + ysin θ ]); (b) g(x,y) = cos(2π f x 2 + y 2 );
(c) g(x,y) = sin(2π f [x3 – y3]/16); (d) g(x,y) = cos(2π f [y – 0.15x 2]). The
spectra of these 2D functions are explained in [Amidror09, Chapter 10].

6 Interestingly, this is true independently of the array length N, be it an even or an odd number (see
Problem 8-5).
8. Miscellaneous issues 205

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

4. 4. 4.

0 0 0

-4. -4. -4.

-4. 0 4. -4. 0 4. -4. 0 4.
(a)
4. 4. 4.

0 0 0

-4. -4. -4.

-4. 0 4. -4. 0 4. -4. 0 4.
(b)
4. 4. 4.

0 0 0

-4. -4. -4.

-4. 0 4. -4. 0 4. -4. 0 4.
(c)
4. 4. 4.

0 0 0

-4. -4. -4.

-4. 0 4. -4. 0 4. -4. 0 4.
(d)
206 8. Miscellaneous issues

exist point-symmetries about a given point, for example about the origin. And indeed,
when we come to generalize the symmetry properties of the Fourier transform, it turns out
that the MD counterpart of an even function is a centrosymmetric function (namely,
a function having central point-symmetry: g(–x) = g(x)), and the MD counterpart
of an odd function is an anticentrosymmetric function (namely, a function having
central antisymmetry: g(–x) = –g(x)).7 A few examples of 2D centrosymmetric or
anticentrosymmetric functions are shown in Fig. 8.3. As illustrated in this figure, if g(x) is
real valued, then the real-valued part of its spectrum G(u) is centrosymmetric and its
imaginary-valued part is anticentrosymmetric (this is, indeed, the multidimensional
generalization of a Hermitian function); and in particular, the MD generalization of
properties (a) and (b) above is:

(c) If g(x) is real-valued and centrosymmetric, so is its spectrum G(u);

(d) If g(x) is real-valued and anticentrosymmetric, its spectrum G(u) is imaginary-valued

and anticentrosymmetric.

These symmetry rules are basically preserved by the multidimensional DFT, too (after
the usual reorganizations, that bring the origin back to the center). However, once again, in
the discrete case we should pay attention to some particular subtleties that have no
equivalent in the continuous case.

Consider, for example, Fig. 8.7(a) and Fig. 8.8(a) in the following section. As we can
see, Fig. 8.8(a) consists of a single white line taken from the center of Fig. 8.7(a), and
both of the figures seem to be symmetric about the origin (namely, centrosymmetric). And
yet, although the DFT spectrum of Fig. 8.7(a) is purely real-valued, as expected, it turns
out that the DFT spectrum of Fig. 8.8(a) does have a non-zero imaginary-valued part
(which is not shown in this figure). Why does Fig. 8.7(a) have a purely real-valued DFT
spectrum, as expected, but Fig. 8.8(a) does not?

One possible explanation could be that our discrete approximation of the continuous,
slanted line in Fig. 8.8(a) is not perfect, so that the jaggies to both sides of the discrete line
are not exactly centrosymmetric. While this could be, indeed, a good reason for having a
non-zero imaginary-valued part in the DFT spectrum, a close inspection of Figs. 8.7(a)
and 8.8(a) shows that the jaggies in both of them have been generated correctly and in a
centrosymmetric manner, so this explanation is not relevant in our case.

To better understand this phenomenon, consider Fig. 8.4. As shown in this figure, when
the size N of the input and output arrays of the DFT is an even number (which is usually
the case, since most FFT algorithms assume that N is a power of two), the origin cannot be
located exactly in the center of the array. After the usual reorganizations (wherever
required), the origin in both input and output arrays is located in the first row and the first

7 Line-symmetries, too, have an impact on the resulting Fourier transform, but an exhaustive list of all
possible combinations in the multidimensional case would be rather tedious and impractical; each of
these cases can be elaborated individually whenever needed [Gaskill78 p. 313]. See also [Komrska79].
8. Miscellaneous issues 207

column of the first quadrant, i.e. in the position N/2 + 1 along each of the array’s axes.
This means that along each of the axes the origin has N/2 negative neighbours, but only
N/2 – 1 positive neighbours. Therefore, a perfect symmetry (or antisymmetry) about the
origin can be obtained if we keep the most negative row and the most negative column
unused, i.e. filled with zeroes (see Fig. 8.4(e),(g)). But because of the periodic nature of
the DFT, the most negative row or column can be also identified with the missing row or
column in the positive end. Therefore, as shown in Fig. 8.4(c),(d), if the input array
containing the centrosymmetric stucture is perfectly wraparound, so that the most negative
row is indeed identical to the expected (but missing) N+1-th row and the most negative
column is identical to the expected (but missing) N+1-th column, then the signal is still
centrosymmetric and has a purely real-valued DFT spectrum even if the most negative
row and column are not zero. (Note, however, that this does not work in the anti-
centosymmetric case, in which the only solution for having a purely imaginary-valued
DFT spectrum is to zero the most negative row and column of the input array).

Figure 8.4: Subtleties in the discrete counterpart of the multidimensional Fourier

symmetry rules. This figure was generated by 2D DFT with N = 32. (a) A
centrosymmetric discrete signal consisting of two 1-valued samples (white
pixels) to the left of the origin and two 1-valued samples to its right, and the
corresponding 2D DFT. Note that the sample at the origin of the signal
domain has been assigned the value of –1 in order that its location be
clearly indicated by a black pixel. (b) A similar centrosymmetric discrete
signal, but this time with 15 1-valued samples to each side of the origin.
Because N = 32, the 15-th sample to the right of the origin is already on the
border, while to the left of the origin there still remains room for one
additional sample. (c) If we set this additional sample, too, to 1, our signal
contains 16 1-valued samples to the left of the origin but only 15 1-valued
samples to its right. And yet, the resulting discrete signal is still considered
as centrosymmetric, and indeed its DFT spectrum has a zero imaginary-
valued part. The reason is that thanks to the periodic nature of the DFT, this
leftmost sample corresponds simultaneously to the –16-th sample to the left
of the origin and to the (missing) 16-th sample to the right of the origin.
(d) A more complex 2D discrete signal that is still considered as
centrosymmetric thanks to the periodic nature of the DFT. Note that its
DFT spectrum has a zero imaginary-valued part. (e),(f) Signals similar to
(b),(c) except that they are not purely horizontal. In this case, as we can see
in (f), if we add the leftmost sample, too, the resulting signal is no longer
centrosymmetric, and it clearly contains one more sample to the left. Unlike
in (c), this additional sample cannot be identified here with a sample that
exceeds to the right, since these two samples would not match in the vertical
direction. Because the discrete signal in (f) is not centrosymmetric, its DFT
spectrum does have a non-zero imaginary-valued part. (g),(h) Similar to
(e),(f), but this time the problem occurs both in the horizontal and in the
vertical directions. More information on the pattern in row (d) is provided
in Problem 8-7.
208 8. Miscellaneous issues

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(a)
8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(b)
8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(c)
8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(d)
8. Miscellaneous issues 209

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(e)
8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(f)
8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(g)
8. 1. 1.

0 0 0

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(h)
210 8. Miscellaneous issues

And indeed, the sharp-eyed reader may have noticed that in some of our 2D DFT
figures (but not in all of them) we have intentionally zeroed the last (most negative) row in
the bottom of the 2D input array, and possibly also the leftmost (most negative) column.
We have done this in order to keep the DFT spectrum purely real-valued if its continuous
counterpart is indeed purely real-valued (namely, in order to avoid abnormal situations
where the Fourier symmetry rules of the continuous-world are violated in the DFT
spectrum). This is the case, for example, in Figs. 5.13, 5.14, 5.15 and 8.8(a), but not in
other cases such as Figs. 5.16, 5.17 and 8.7(a), in which the DFT spectrum is anyway
purely real-valued (because of the wraparound nature of these signals). Other examples
can be found in Fig. C.8, too.
It is interesting to note that this artifact does not occur in the 1D case. This is explained
by Figs. 8.4(a)–(c), which correspond, in fact, to a 1D DFT along the horizontal axis. This
is probably the reason why this DFT artifact is rarely, if ever, mentioned in the literature.

8.5 Jaggies on sharp edges as aliasing or reconstruction phenomena8

It is well known in computer graphics that the approximation of continuous-world

shapes by pixels on a discrete raster grid (discretization) may cause the appearance of
jagged edges (also known as “jaggies” or “staircasing effects”) on the shape’s borders
[Foley90 p. 14], along slanted or curved lines. These jaggies are most prominent on sharp
edges such as black/white transitions (which are represented mathematically by 0/1
transitions), like in Fig. 8.5(a). But they can be also visible in smoother gray-level
transitions, like in Fig. 8.5(b), where each pixel is represented by a constant intermediate
gray level corresponding to a value between 0 and 1.9 These jaggies are particularly
objectionable at low resolutions (i.e. low sampling rates). In the computer graphics
literature, and in particular in the field of digital typography, these jaggies are often
considered as aliasing effects (see, for example, [Crow77], [Foley90 pp. 14–15, 628] or
[Rubinstein88 pp. 45–48]); and moreover, the methods devised for reducing the visibility
of these jaggies are often called anti-aliasing methods [Foley90 pp. 132–142, 617–646,
817–818, 909, 965–976].10

Having reviewed in Chapter 5 the various manifestations of aliasing, such as

masquerading lower frequencies (in the signal domain) or overlapping of replicas (in the
spectral domain), how can jaggies be related to aliasing phenomena? In particular, it is

8 This section concerns artifacts that occur due to signal sampling or reconstruction, and it is not
specifically related to DFT. Note that the term reconstruction artifact that is explained here is also used
in the following section.
9 The number of available gray levels is limited, but it may differ from case to case.
10 In digital typography, anti-aliasing is performed either by using gray-level pixels at the edges of the
strokes [Rubinstein88 pp. 111–115, 311], if individual pixels are allowed to have gray levels (for
example on computer displays); or, alternatively, by using the dentation technique [Rubinstein88 pp.
81–82, 314], in cases where individual pixels can only be set on or off and no intermediate levels are
available (like in most printing devices). Both methods are aimed at blurring the jaggies and making
them less visible when viewed from a reasonable distance.
8. Miscellaneous issues 211

(a) (b)

Figure 8.5: Jaggies in discrete-world images: (a) Along sharp edges (here:
slanted black lines forming the letter A). (b) Along smooth
gray-level shapes (here: a circular cosinusoidal wave).

clear that if jaggies are indeed aliasing phenomena, they must be represented in some way
or another in the spectral domain, too; but how exactly? In the present section we will try
to eludidate these and other similar questions. Note, however, that throughout this section
we assume that the jaggies in question are sufficiently big to be perceived by the naked
eye, and we do not discuss issues related to the human visual system, such as visual angle
resolution, viewing distance, modulation transfer functions (MTFs), etc. More details
about human vision and its modelization can be found in references such as [Wandell95],
[Daly92], [Pratt91, Sec. 2.4], [Boff86], [Cornsweet70], etc.

Let us consider, as a simple illustrative example, a periodic line grating that is slightly
rotated by angle θ (see Example 5.7). The spectrum of the original continuous-world
grating, before sampling, is an infinite comb whose impulses are located along a straight
line passing through the origin at the angle θ, where the impulse interval along this line
corresponds to the frequency of the grating. This spectrum is obviously not band limited.
After sampling the original continuous-world grating, as we have seen in Chapter 5, the
continuous-world spectrum of the resulting sampled grating consists of the original comb
plus infinitely many replicas thereof, that are centered about each impulse of a 2D impulse
nailbed (this nailbed being itself the Fourier transform of the sampling nailbed).11 The
DFT spectrum, on its part (see Fig. 8.6), contains only the central part of this continuous-

11More precisely, the replicas are centered about the points (kfs, lfs) where fs is the sampling frequency
in both directions and k,l ∈Z. See Fig. 5.12.
212 8. Miscellaneous issues

world spectrum: it extends along both axes between the frequencies –12 fs and 12 fs, where fs is
the sampling frequency (see Sec. 4.3).12 This means that the spectrum of the sampled
grating, both in the continuous-world and in its DFT counterpart (Fig. 8.6), contains many
new impulses belonging to the new replicas. If any of these new impulses falls close
enough to the spectrum origin, like in Fig. 8.6(c) (note the two encircled impulses), a new
low-frequency structure (moiré effect) becomes visible in our sampled grating, as shown
by the arrows in the image domain (for more details on sampling moiré effects see, for
example, [Amidror09, Sec. 2.13]). But even if the new impulses do not fall close to the
spectrum origin, like in Fig. 8.6(b), they still must correspond to some new structures in
the sampled grating, which did not exist in the original, continuous grating. And indeed, it
turns out that these new impulses represent new frequencies which correspond to the
jaggedness of the sampled grating.

Figure 8.6: A simple example illustrating aliasing in the 2D case, where the
spectrum consists of a 1D impulse-comb (see also Fig. 5.14). The
original continuous-world function in the signal domain is a rotated
line grating, whose spectrum is a rotated infinite impulse comb
through the spectrum origin whose impulse interval corresponds to
the frequency of the grating (see, for example, [Amidror09 pp. 23–
25]). After sampling, the spectrum of the resulting sampled grating
consists of the original comb plus infinitely many replicas of this
comb, that are centered about the points (kfs, lfs), where fs is the
sampling frequency in both directions and k,l ∈Z. This means that
the continuous-world spectrum of the sampled grating contains
infinitely many new impulses belonging to the new replicas. But the
DFT spectrum only includes the center of this continuous-world
spectrum, between the frequencies –12 f s and 12 fs to both directions
(see also Problems 8-10 and 8-11). In each of the rows of this
figure the left-hand column shows the sampled, discrete-world
signal and the right-hand column shows its spectrum as obtained
by DFT. The only difference between the three rows is in the
rotation angle θ: (a) θ = 0°; (b) θ = 14.036° (so that tanθ = 1/4);
and (c) θ = 3°. In case (c), due to the small rotation angle, impulses
of the replicas fall very close to the spectrum origin (see the two
impulses surrounded by circles, which are very slightly rotated with
respect to the vertical axis). And indeed, these impulses correspond
to a new low-frequency parasite structure (sampling moiré) that is
clearly visible in the sampled line grating (see the arrows in the
signal domain). The cross-like oscillations which surround impulses
in the spectra of (b) and (c) are due to the leakage artifact (see
Chapter 6). Note that both in the image and in the spectral domains
mid-gray represents zero, black is negative and white is positive (see
Footnote 5 in Sec. 1.5.2).

12As explained in Chapter 5, the impulses of the other replicas which fall within the range of the DFT
spectrum can be also viewed as a cyclical continuation of the main comb, that folds-over each time it
exceeds the border of the DFT spectrum and re-enters from the opposite side.
8. Miscellaneous issues 213

Signal domain Spectral domain

8. 4.

0 0

-8. -4.
-8. 0 8. -4. 0 4.
(a)
8. 4.

0 0

-8. -4.
-8. 0 8. -4. 0 4.
(b)
8. 4.

0 0

-8. -4.
-8. 0 8. -4. 0 4.
(c)
214 8. Miscellaneous issues

To see this, consider Fig. 8.7, which is intentionally drawn at a lower resolution so that
the individual jaggies as well as the individual impulses in the spectrum become clearly
visible. Fig. 8.7(a) shows the sampled grating and its spectrum as obtained by DFT. As
we can see, the spectrum contains the main, original impulse comb (the slightly rotated
comb passing through the origin, which is indicated by arrows) plus impulses belonging
to its new replicas due to sampling. In order to see the influence of these new impulses, let
us zero all of them leaving in the spectrum only the impulses belonging to the original
comb. The corresponding structure back in the signal domain is obtained by taking the
inverse DFT of this spectrum. The results of this manipulation are shown in Fig. 8.7(b).
As we can see in this figure, the effect on the signal domain of eliminating the impulses of
the new replicas (the aliased impulses) consists of smoothing out the jagged edges of the
sampled grating (see the gray level pixels along the line edges). Note, however, that this
does not yet completely eliminate the jaggies; as will be shown below, the residual jaggies
that we can still see in Fig. 8.7(b) under a magnifying glass are in fact reconstruction
artifacts that occur due to the square pixels being used to draw the sampled signal.
Finally, Fig. 8.7(c) shows what happens if we zero in the spectrum of Fig. 8.7(a) the
impulses of the original comb and keep only the new impulses that are due to sampling.
Once again, the corresponding structure back in the signal domain is obtained by taking
the inverse DFT of this spectrum. Note that the signal-domain structure in Fig. 8.7(c) is
simply the difference between the jagged grating of Fig. 8.7(a) and its smoothed-out
version of Fig. 8.7(b); this difference corresponds, indeed, to the net effect of the jaggies
themselves on the line edges.

Figure 8.7: Jaggies along the edges of a sampled, slightly rotated line grating
and their spectral representation. (a) The sampled grating and its
spectrum, as obtained by DFT, between the frequencies –12 fs and 12 fs
in both directions; the two arrows indicate the main comb. (b) The
impulses of the main comb alone, after zeroing all the remaining
aliased impulses, and the corresponding signal-domain structure
obtained by inverse DFT. (c) The aliased impulses alone, after
zeroing the impulses of the main comb, and the corresponding
signal-domain counterpart obtained by inverse DFT. As shown in
(b), the effect on the signal domain of eliminating the impulses of
the other replicas (the aliased impulses) consists of smoothing out
the jagged edges of the sampled grating (see the gray level pixels
along the line edges). The signal-domain structure in (c) is simply
the difference between the jagged grating of (a) and its smoothed-
out version of (b); this difference corresponds, indeed, to the net
effect of the jaggies themselves on the line edges. Note that in this
figure there are no leakage artifacts because this case satisfies
condition (a) of Remark 6.7. Note also the overshoots (like in the
Gibbs phenomenon) which occur to both sides of each of the line
edges in (b).
8. Miscellaneous issues 215

Signal domain Spectral domain

4. 4.

0 0

-4. -4.
-4. 0 4. -4. 0 4.
(a)
4. 4.

0 0

-4. -4.
-4. 0 4. -4. 0 4.
(b)
4. 4.

0 0

-4. -4.
-4. 0 4. -4. 0 4.
(c)
216 8. Miscellaneous issues

These Fourier-based considerations are most easily observed in periodic structures like
Fig. 8.7 since the spectra in such structures are purely impulsive, and the aliased impulses
can be easily identified and manipulated. However, it is clear that jaggies do not only occur
in periodic structures. To illustrate a simple aperiodic case consider Fig. 8.8, which shows
a single straight line that is slightly rotated by angle θ. As we have seen in Example 5.6,
the continuous-world spectrum of this aperiodic structure, before sampling, consists of a
straight line-impulse (a “blade”) passing through the origin at the angle θ , whose
amplitude is modulated by a sinc function (see also Sec. D.10 in Appendix D). After
sampling the original continuous-world line, the continuous-world spectrum of the
sampled line consists of the original line-impulse plus infinitely many replicas thereof.
These replicas are centered about each impulse of a 2D impulse nailbed, which is islelf the
Fourier transform of the sampling nailbed. More precisely, these replicas are centered
about the points (kfs, lfs), where fs is the sampling frequency in both directions and k,l ∈Z.
This continuous-world spectrum of the sampled line is shown schematically in Fig. 5.21.
We now proceed to the discrete world, following the same reasoning as in Fig. 8.7. Fig.
8.8(a) shows the sampled line and its spectrum as obtained by DFT. As we can see, the
DFT spectrum contains the main, original line-impulse (indicated by arrows) plus line-
impulses belonging to its new replicas due to sampling. In order to see the influence of
these new spectral elements, Fig. 8.8(b) shows the main line-impulse alone, after zeroing
all its replicas in the DFT spectrum, and the corresponding signal-domain counterpart
obtained by inverse DFT. Finally, Fig. 8.8(c) shows only the replicas of the line-impulse,
after zeroing the main line-impulse itself, and the corresponding signal-domain
counterpart obtained by inverse DFT. Like in Fig. 8.7, the effect on the signal domain of
eliminating the aliased elements from the spectrum consists of smoothing out the jagged
edges of the sampled line.

Figure 8.8: Same as in Fig. 8.7, except that this time the original continuous-
world signal is not a slightly rotated line grating, but just a single
white line taken from this grating (the central one). As explained in
Example 5.6, the continuous-world spectrum of this line consists of
a straight line-impulse passing through the origin whose amplitude
is modulated by a sinc function, and which is slightly rotated by the
same angle. (a) The sampled line and its DFT spectrum; the two
arrows indicate the main line-impulse. (b) The main line-impulse,
after zeroing all its replicas in the DFT spectrum, and the
corresponding signal-domain counterpart obtained by inverse DFT.
(c) The replicas of the line-impulse, after zeroing the main line-
impulse, and the corresponding signal-domain counterpart obtained
by inverse DFT. Like in Fig. 8.7, the effect on the signal domain of
eliminating the aliased elements from the spectrum consists of
smoothing out the jagged edges of the sampled line. The ripple
effect in the DFT spectrum is due to leakage; leakage artifacts are
present here because this case, unlike that of Fig. 8.7, does not
satisfy the conditions of Remark 6.7.
8. Miscellaneous issues 217

Signal domain Spectral domain

4. 4.

0 0

-4. -4.
-4. 0 4. -4. 0 4.
(a)
4. 4.

0 0

-4. -4.
-4. 0 4. -4. 0 4.
(b)
4. 4.

0 0

-4. -4.
-4. 0 4. -4. 0 4.
(c)
218 8. Miscellaneous issues

Signal domain Spectral domain

8. 4.

0 0

-8. -4.
-8. 0 8. -4. 0 4.

Figure 8.9: A sampled vertical line grating may have a sampling moiré effect
(note the periodically repeating thicker vertical lines) thanks to new
aliased low frequencies that appear due to an interaction with the
sampling frequency. Compare with Fig. 8.6(a), where no sampling
moiré occurs. Note that in both cases no jaggies are present.

Note that in both of Figs. 8.7 and 8.8 the jaggies occur on slanted edges. In both cases,
if the lines in the signal domain are purely vertical or horizontal no jaggies appear on their
edges. And indeed, in such cases the new replicas in the spectral domain due to sampling
fall back along the original replica, and no new frequencies are generated that correspond
to jaggies (see Fig. 8.6(a) and Fig. 8.9, as well as Problem 8-11).

As we can see, there is indeed strong evidence in favour of considering jaggies as

aliasing phenomena. Furthermore, just like the previously discussed facets of aliasing
(masquerading lower-frequencies in the signal domain, or overlapping of replicas in the
spectral domain; see Chapter 5), jaggies, too, tend to become less prominent when we
increase the sampling resolution or when we apply low-pass filtering prior to the sampling
so that sharp transitions become smoother [Crow77].

But on the other hand, there are also good reasons against considering jaggies as
aliasing phenomena. For example, there exist many structures that present strong aliasing,
but have no jaggies at all (for instance, as shown in Fig. 8.9, a sampled horizontal or
vertical line grating may have a sampling moiré effect due to new aliased low frequencies
without having jaggies). And conversely, jaggies may also exist when there is no aliasing
at all — for example in a 2D plot of a slightly rotated cosinusoidal grating, whose
frequency is far below half of the sampling frequency (see Fig. 8.10). How can this be
explained?
8. Miscellaneous issues 219

Signal domain Spectral domain

4. 2.

2. 1.

0 0

-2. -1.

-4. -2.
-4. -2. 0 2. 4. -2. -1. 0 1. 2.

Figure 8.10: Low-resolution figure showing a slightly rotated cosine wave of

frequency f = 0.88 and its spectrum. Note that jaggies are visible
although the sampling frequency fs = 4 is higher than 2f, and no
aliasing occurs in the sampling process.

The key for understanding this question resides in the ambiguity of the term aliasing.
This term is used in the literature for two different effects (see, for example, pp. 343–344
in [Glassner95]): (1) the distortion that is produced by poor sampling, and which causes
the sampled signal to become indistinguishable from a lower frequency signal (an alias);
and (2) the distortion that is produced by poor reconstruction, which causes the signal
that is reconstructed from the samples to be different than the original continuous signal
(again, an alias).13 To resolve this ambiguity, we will henceforth call the first, classical
effect aliasing due to sampling or sampling aliasing, while the second effect will be called
aliasing due to reconstruction or simply reconstruction error (other terms being
sometimes used in the literature are prealiasing and postaliasing, respectively [Mitchell88
p. 222]). Let us explain this in more detail.

While Sampling is the process that converts a continuous signal to a discrete one,
reconstruction is the process that recreates a continuous signal from its samples [Foley90
Sec. 14.10.5; Mitchell88].14 According to the sampling theorem, all the information in the
original continuous signal is preserved in its sampled version if the sampling frequency is

13 As shown below, both of these problems originate from failures in the correct application of the
sampling theorem: the failure to fulfil the required condition on the frequencies leads to sampling
aliasing; while the failure to approximate ideal reconstruction leads to reconstruction error.
14 Note that theoretically a sampled signal consists of zero-width impulses (of varying heights), and not

of real-world “pixels” having square or circular shapes. It is precisely the reconstruction process that
brings back the “flesh” arround each of the sampled “bones”.
220 8. Miscellaneous issues

Signal domain Spectral domain

x u
0 –fs –fG 0 fG fs
(a)

x u
0 –fs 0 fs
(b)

x u
0 0
(c)

Figure 8.11: Continuation of Fig. 5.3, showing schematically the signal- and spectral-
domain representations of an ideal reconstruction process, that gives back
exactly the original continuous signal we had before sampling. (a) The
sampled signal, as in Fig. 5.3(c). (b) The ideal reconstruction function
(a narrow sinc function) according to the sampling theorem and its
spectrum, which is a rect function (a 1-valued pulse) that extends from
–12 f s to 12 fs. (c) The perfectly reconstructed signal (convolution of the
signals (a) and (b)) and its spectrum (product of the spectra (a) and (b)).
Note that the reconstructed signal in row (c) is indeed identical to the
original continuous signal shown in row (a) of Fig. 5.3.

at least twice that of the heighest frequency contained in the signal. Under this condition,
the theorem guarantees that the original continuous signal can be perfectly reconstructed
by multiplying the spectrum of the sampled signal with a rect function that cuts off all
frequencies beyond half of the sampling frequency, or equivalently, by convolving the
sampled signal with the Fourier transform of this rect function, i.e. with the corresponding
sinc function [Brigham88 p. 83]. This is shown in Figs. 5.3 and 8.11; note that the
explanation is easier to understand in the spectral domain (the right-hand side of the
figures). However, in practice, the reconstruction of a sampled signal is never done by
convolving its impulses with an infinite-range sinc function, as stipulated by the sampling
8. Miscellaneous issues 221

Signal domain Spectral domain

x u
0 –fs –fG 0 fG fs
(a)

x u
0 –fs 0 fs
(b)

x u
0 –fs 0 fs
(c)

Figure 8.12: Continuation of Fig. 5.3, showing schematically the signal- and spectral-
domain representations of a non-ideal reconstruction process, that causes
reconstruction artifacts. (a) The sampled signal, as in Fig. 5.3(c). (b) A
non-ideal reconstruction function (a rect function representing a square
pixel) and its spectrum, which is a sinc function that extends ad infinitum.
(c) The reconstructed signal (convolution of the signals (a) and (b)) and
its spectrum (product of the spectra (a) and (b)). Row (c) is not identical
to row (a) in Fig. 5.3, and it contains reconstruction artifacts (note the
debris from the neighbouring replicas that give new high-frequency noise
in the resulting reconstructed signal).

theorem.15 Instead, each impulse of the sampled signal is typically represented (i.e.
convolved) by a pulse function whose width equals the distance between two consecutive
samples [Foley90 p. 641]. For example, in the 2D case (like in computer displays or in
digital printing devices) each impulse of the sampled signal to be displayed is convolved
with a single pixel shape (which may be, depending on the device, a square dot, a circular

15 Note that the condition required by the sampling theorem, that the signal being sampled is band
limited, implies that in the signal domain the signal has infinite duration [Zwillinger96 p. 535].
Similarly, the reconstruction of the continuous signal according to the sampling theorem is done by
convolving the sampled signal with an unrealizable, infinite sinc function. These assumptions yield a
mathematical model that is only an idealized approximation to the real world. The sampling theorem’s
claim that perfect reconstruction is possible is mathematically correct for the model, but only an
approximation for real-world signals (see, for example, [Chu08 p. 182] or [Wikipedia12]).
222 8. Miscellaneous issues

dot, etc.). The resulting reconstructed signal may therefore have highly visible jaggies
along its borders (see Fig. 8.10), and possibly also sharp transitions between the values of
neighbouring pixels. These elements correspond to new high frequencies in the
reconstructed signal that are not present in the samples, and which do not exist in the
original, continuous signal, either (see the spectral domain explanation in Fig. 8.12). This
process results in reconstruction errors, since when the resolution of the display device is
not sufficient the reconstructed image differs significantly from the original continuous
image (it is pixelized and jagged), and what we actually see is an alias. This kind of
aliasing has nothing to do with undersampling and it is solely due to poor
reconstruction;16 and indeed, as we have already mentioned above, this phenomenon may
occur even in cases which are completely alias-free (in the classical sense).

Let us now return to our question concerning the origin of the jaggies. As we can see in
Figs. 8.7 and 8.8, in many cases jaggies result indeed from sampling aliasing, and they are
represented in the DFT spectrum by aliased elements (this is particularly evident when
sampling a 2D signal involving slanted sharp transitions, which is not band limited and
therefore suffers from aliasing). Clearly, jaggies are further amplified by reconstruction
aliasing (due to poor reconstruction of the signal). However, they can be even generated
by reconstruction aliasing if they have not already been generated due to sampling
aliasing. This is the case, for example, in slightly rotated low-frequency cosinusoidal
gratings, which are clearly band limited and have no sampling aliasing (the impulse-
samples along the corrugations of the sampled cosinusoidal grating are gradually
attenuated in such a way that no jaggies are apparent). In cases like this, if reconstruction
is correctly done, as stipulated by the sampling theorem, no jaggies should appear in the
reconstructed signal; and if jaggies do appear — like in Fig. 8.10 — this is only due to
reconstruction aliasing. And indeed, this observation led some sources to the conclusion
that jaggies are not aliasing artifacts as often claimed in the literature, but rather
reconstruction artifacts (see Problem 8-12).

As we can see, it follows from our discussion that jaggies have mixed origins: in some
cases like in Figs. 8.7 and 8.8 they are due to sampling aliasing (although even then they
may be further amplified by poor reconstruction); while in other cases, like in Fig. 8.10,
they are a pure product of poor reconstruction. In the first case jaggies clearly manifest
themselves in the DFT spectrum, but in the latter case they do not. This is, indeed, the
origin of the confusion that prevails in the literature with regards to jaggies and aliasing.

Remark 8.2: It is important to stress that the new high frequencies due to reconstruction
aliasing are not represented in the DFT spectrum of the sampled signal, since they only
occur in a later stage, during the reconstruction of the continuous signal (whence the name
postaliasing). 17 And yet, the issue of reconstruction does have an impact on our

16 Note, in particular, that the new high frequencies that are generated by poor reconstruction are not
corrupting the low frequency components, so no sampling aliasing is actually taking place.
17 Of course, they would have been represented in the continuous-world spectrum of the reconstructed
signal, had we cared to produce such a spectrum.
8. Miscellaneous issues 223

discussion on DFT and its artifacts: Each time we plot a discrete function or its DFT, be it
in the 1D or MD case, we implicitly make a choice regarding the precise way the discrete
values should be graphically plotted (i.e. reconstructed). For example, in the 1D case we
may represent individual samples by isolated dots, by vertical spikes or bars emanating
from the horizontal axis, or even as a curve that connects neighbouring points and
somewhat approaches the familiar look of the original continuous function. In the 2D
case, too, sampled functions may be represented in various ways. The most usual way
(that we tacitly utilize in all our 2D figures here) uses adjacent square pixels, each having a
distinct constant gray level, to represent the sampled values. These square pixels are
clearly visible in cases where the array size N is low, but when N is sufficiently high (with
respect to the figure’s dimensions and the viewing distance) we may have the illusion of
dealing with a continuous function (see, for example, Fig. 8.3). However, the various
graphical techniques that may be used to plot the sampled functions and their DFT spectra
only influence the visual presentation of the plots, and they do not add or remove
frequencies from the DFT spectra themselves. A more detailed discussion on the graphic
presentation of sampled signals can be found in Sec. 1.5. Q

Finally, let us return to the question with which we opened this section: What is the
connection between the sampling-induced jaggies and the classical manifestations of
aliasing (masquerading lower-frequencies in the signal domain, or spectral replications in
the spectral domain)? Indeed, the most notorious signal-domain manifestation of aliasing
consists of cases where new low-frequency structures become visible due to the sampling
process, as shown in Fig. 5.1(b) for the 1D case and in Fig. 8.6(c) for the 2D case. These
new low-frequency structures are simply sampling moirés. But the spectral-domain
replicas due to the sampling process may also generate new frequencies that are much
higher than this. These new high frequencies contribute to the microstructure details of the
sampled signal and give it its typical jagged look, as shown for example in Figs. 8.6(b) or
8.7.

8.6 Sub-Nyquist artifacts18

We have seen in Chapter 5 that when an aliased frequency due to sampling falls close
enough to the center of the signal’s Fourier spectrum, a low-frequency sampling moiré
effect may become visible in the signal domain, i.e. in the sampled signal.

Interestingly, however, a conspicuous low-frequency structure may also appear in the

sampled signal when no aliasing occurs, so that no aliased frequencies fall close to the
center of the signal’s spectrum. It is even more surprising to realize that such artifacts
occur even within the frequency range in which sampling is supposed to be safe by virtue
of the sampling theorem. And to crown it all, these conspicuous low frequencies are not

18This section concerns artifacts that occur due to signal sampling or reconstruction, and it is not
specifically related to DFT. Note that the term reconstruction artifact that is being used here was
already explained in the previous section.
224 8. Miscellaneous issues

even represented in the spectral domain. This happens, for example, when the signal being
sampled is a cosine function g(x) = cos(2πfx) whose frequency f is slightly below half of
the sampling frequency (12 fs), i.e. close to the maximum frequency that can be represented
in the DFT spectrum (see Eq. 4.10). But because this phenomenon occurs here when the
cosine frequency f is below 12 fs (see Fig. 8.13(f)) it is clear by virtue of the sampling
theorem that this artifact is not caused by aliasing. For this reason it is sometimes called a
sub-Nyquist distortion [Williams00], where the Nyquist frequency means half of the
sampling frequency.19 This phenomenon has already been reported in several references
[Mitchell88 pp. 222, 225; Foley90 p. 642; Gomes97 pp. 206–207; Jain98 pp. 126–127;
Williams00; Fielding06], but it still remains poorly understood, sometimes just being
dismissed as a moiré effect or as a reconstruction artifact, without further explanations.

To better understand this phenomenon, let us consider Fig. 8.13. The rows of this figure
show a series of cosine functions g(x) = cos(2πfx) that only differ in their frequency f. All
of these cosines have been sampled with the same sampling frequency fs. When the
frequency f of the continuous cosine precisely equals 12 fs (or in other words, when the
cosine period equals two sampling intervals; see row (g) of the figure), the sampling is
done exactly twice per period of the cosine; hence, the cosine is sampled on its successive
maxima and minima, giving the sampled values 1, -1, 1, -1... But when the original
continuous function g(x) = cos(2 π fx) has a frequency f just slightly below 12 fs , say
f = 12 fs – ε (so that no aliasing occurs; see row (f) of the figure), sampling is done slightly
more often than twice per period of the cosine. Hence, the successive sampling points fall
at first close to the extrema of the cosine and have values close to 1, -1, 1, -1... (for
example, around the origin), but farther away, due to the slight frequency difference, the
successive sampling points fall near the zero-crossing points of the cosine and have values
close to 0, 0, 0, 0... These different regions of the sampled signal repeat themselves
periodically20 at a much lower frequency than f or fs, and indeed they look in the signal
domain like a moiré effect, as shown in row (f) of the figure. Let us study this
phenomenon in more detail, using a concrete example.

Example 8.1: A 1D sub-Nyquist artifact:

Consider the case shown in Fig. 8.13(f). Here, the sampling frequency is fs = 8 Hz, so
that the maximum frequency allowed by the sampling theorem without causing aliasing is
2 f s = 4 Hz (note that this is also the maximum frequency representable in the DFT
1

spectrum, as we have seen in Eq. (4.10)). The frequency of the continuous cosine function
being sampled in this case is f = 3.75 Hz, so that ε = 12 fs – f = 0.25 Hz. As we can see in
the figure, the resulting beating effect has a period of 2, i.e. a frequency of 12 Hz, which is
precisely 2ε. Why is the low frequency of the beating effect twice the frequency difference
ε = 12 fs – f rather than simply ε? To better understand this consider Fig. 8.14, which shows

19 As already mentioned in Footnote 4 in Sec. 4.3, we usually prefer to avoid using the term “Nyquist
frequency” because of its ambiguity. The name of the present artifact is an exception, but it should not
lead to any confusions. See also Problem 8-14 about the name of this artifact.
20 Note that the distance between successive zero-crossing points of the cosine is the same as the distance
between successive extrema.
8. Miscellaneous issues 225

in the continuous-world spectrum what happens when we sample a continuous cosine

function g(x) = cos(2πfx) having the frequency f = 12 fs – ε using the sampling frequency fs.
The reason we prefer to use here continuous-world spectra is that they are easier to
understand than DFT spectra (such as in Fig. 8.13), since their frequency range is not
limited, and they do not suffer from wraparound and foldover of higher frequencies as
DFT spectra do.
Fig. 8.14(a) shows the continuous-world spectrum of our original, unsampled cosine
function, which consists of an impulse pair at the frequencies 12 fs – ε and –12 fs + ε, while
Fig. 8.14(b) shows the continuous-world spectrum of its sampled counterpart. As we
already know from Chapter 5, if the CFT of g(x) is G(u), then the continuous-world
spectrum of the sampled version of g(x) consists of infinitely many replicas of the original
spectrum G(u), which are centered about all the integer multiples of the sampling
frequency fs (see Eq. (5.2) and Fig. 5.9). Fig. 8.14(b) shows only three of these replicas:
the original one (which is identical, up to a certain amplitude scaling factor, to row (a) of
the figure) plus its two nearest neighbours, that are centered about fs and –fs. As we can
see in Fig. 8.14(b), these new replicas, that are introduced due to the sampling, add to the
continuous-world spectrum a new impulse pair that is located very close to the impulse

Figure 8.13: A series of cosine functions g(x) = cos(2πfx) having gradually increasing
frequency f, and their spectra (as obtained by DFT, after all the required
adjustments as explained in Chapters 3, 4 and Appendix A). For the sake of
clarity, consecutive samples (dots) are connected by straight line segments;
the original, continuous cosines are drawn by thinner curves. The array
length (number of samples) being used is N = 64, the sampling range length
is R = 8, and therefore the sampling interval is Δx = 8/64 = 0.125 and the
sampling frequency is fs = 8 Hz. The only difference between the rows is in
the cosine frequency f: (a) f = 0.25 Hz; (b) f = 0.5 Hz; (c) f = 2 Hz;
(d) f = 3 Hz; (e) f = 3.5 Hz; (f) f = 3.75 Hz; (g) f = 4 Hz, namely f = 12 fs;
(h) f = 4.25 Hz; (i) f = 7.5 Hz; (j) f = 7.75 Hz. In row (g) the cosine
frequency equals half of the sampling frequency, i.e. the maximum
frequency that can be represented in the DFT spectrum. Note that in this
case the positive frequency impulse at 4 Hz is already aliased and falls on
top of the negative frequency impulse at –4 Hz. This happens because, as we
have seen in Sec. 3.2, when N is even the DFT spectrum contains in its
positive end one element less than in its negative end, so that the minimum
frequency –12 fs is indeed included in the DFT spectrum, but the maximum
frequency 12 fs is already folded over. In rows (h)–(j) the cosine frequency f
exceeds 12 fs, meaning that aliasing occurs (note that the corresponding
impulses exceed the boundaries of the DFT spectrum and re-enter from the
opposite end). In particular, when the folded-over impulses are sufficiently
close to the spectrum origin (see rows (i),(j)) a strong low-frequency
sampling moiré becomes more prominent and visible than the original
cosine function itself. Note, however, that the low-frequency beating effects
which appear in rows (e),(f),(h) are not true moirés (there are no
corresponding low-frequency impulses in the respective spectra).
226 8. Miscellaneous issues

Signal domain Spectral domain

2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(a)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(b)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(c)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(d)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(e)
8. Miscellaneous issues 227

Signal domain Spectral domain

2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(f)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(g)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(h)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(i)
2 1

1
0.5

-4 -2 2 4
x
-1 -4 -2 2 4
Hz

-2
-0.5
(j)
228 8. Miscellaneous issues

pair of our original cosine, i.e. at the frequencies 12 fs + ε and –12 fs – ε. This new impulse
pair corresponds to a newly added cosine in the signal domain, whose frequency is
2 fs + ε. But as shown in Sec. D.3 of Appendix D, the sum of two cosines with very close
1

frequencies f1 and f2 gives a beating effect (pseudo moiré) whose frequency is f2 – f1; and
indeed, the beating effect we obtain here has the frequency of (12 fs + ε ) – (12 fs – ε) = 2ε, as
we have seen above. In the particular case of Fig. 8.13(f) this beating effect has the
frequency of 12 Hz, i.e. a period of 2; note that this low-frequency artifact in our sampled
signal is more prominent than the original cosine itself. This beating effect is indeed a
sub-Nyquist artifact and not an aliasing artifact, since the frequency of the original
continuous cosine being sampled here, f = 12 fs – ε = 3.75 Hz, is slightly lower than the
maximum frequency allowed by the sampling theorem, 12 fs = 4 Hz. Q

Now, because the frequency of our cosine function is below 12 fs, this function can be
perfectly reconstructed from its sampled version, as stipulated by the sampling theorem.
This is done, as shown in Figs. 8.14(c)–(d), by multiplying the spectrum of Fig. 8.14(b)
with a rect function (a 1-valued pulse) extending from –12 fs to 12 fs (or equivalently, in the
signal domain, by sinc-function interpolation, i.e. by convolving the sampled version of
the cosine signal with the sinc function that is the inverse Fourier transform of the above
rect function; see Fig. 8.11). This ideal reconstruction removes the new impulse pair from
the spectrum, as shown in Fig. 8.14(d), and hence the corresponding new cosine
disappears from the signal domain, so that no pseudo moiré is generated. But as we have
seen in Sec. 8.5, such an ideal reconstruction rarely occurs in real-world situations. More
realistic reconstruction methods include: (i) zero-order interpolation (also called nearest-
neighbour interpolation), which is equivalent to convolving the sampled signal with a
narrow square pulse or “pixel”, like in Fig. 8.12 (see [Abdou82 pp. 668–669] or
[Stucki79 pp. 184–185]); or (ii) linear (i.e. first-order) interpolation (connecting
consecutive samples by straight line segments, like in Fig. 8.13), which is equivalent to

Figure 8.14: Explanation in the continuous-world spectrum of the sub-Nyquist artifact

that occurs when sampling a continuous cosine function g(x) = cos(2πfx)
whose frequency f is just slightly below half of the sampling frequency
(12 fs). (a) The continuous-world spectrum of the original cosine. (b) The
continuous-world spectrum of the sampled cosine. Note the replicas of
the original spectrum, which are centered about the sampling frequency fs
and each of its integer multiples. Thanks to the two neighbouring replicas,
the spectrum (b) of the sampled cosine represents, in fact, a sum of our
original continuous cosine, whose frequency is slightly below 12 fs, with a
new continuous cosine whose frequency is slightly above 12 fs. (Impulses
of higher frequencies, that are also present in the spectrum (b), are
required for the correct spectral representation of the sampled cosine, but
for the sake of our discussion here they can be ignored). (c), (d) Because
the frequency of our cosine function is below 12 fs, it can be perfectly
reconstructed from its sampled version as stipulated by the sampling
theorem, by multiplying the spectrum (b) with a rect function (a 1-valued
8. Miscellaneous issues 229

Spectral domain

ε ε

–fs fs
u
0
2 – f f 2
(a)

ε ε ε ε

–fs fs
u
– fs 0 fs
2 2
(b)

–fs fs
u
– fs 0 fs
2 2
(c)

ε ε

–fs fs
u
– fs 0 fs
2 2
(d)

ε ε ε ε

–fs fs
u
– fs 0 fs
2 2
(e)

pulse) extending from –12 fs to 12 fs. (e) When reconstructed by multiplying

the spectrum (b) with a non-ideal substitute of the ideal rect function (c),
debris of the new replicas that appeared in (b) due to the sampling may still
subsist in the spectrum. The sub-Nyquist artifact that appears in the
sampled signal (see Fig. 8.13(f)) is simply the beating modulation effect
that occurs in the sum of the original cosine having the frequency f = 12 fs – ε
and the new cosine having the very close frequency of 12 fs + ε. As shown in
(b), this beating effect is generated by the sampling operation; but it
becomes actually visible due to the non-ideal reconstruction. Note that the
low beating frequency itself is not present in the spectrum, meaning that it
is not a true moiré effect but rather a pseudo-moiré effect.
230 8. Miscellaneous issues

convolving the sampled signal with a narrow triangular pulse (see [Abdou82 p. 669] or
[Stucki79 pp. 186–188]). But none of these alternative interpolation methods can perfectly
reconstruct the original signal, since, as stipulated by the sampling theorem, sinc-function
interpolation is the only perfect reconstruction method. In terms of the spectral domain,
both of these alternative reconstruction methods are equivalent to multiplying the spectrum
(b) with a non-ideal substitute of the ideal rect function. But when doing such a
multiplication, the new replicas that appeared in row (b) due to the sampling will not be
completely removed, and some debris thereof may still subsist in the spectrum, as shown
in Fig. 8,14(e). In such cases, a low-frequency sub-Nyquist artifact may indeed become
visible in the sampled signal (see Fig. 8.13(f)); this is simply the beating modulation effect
that occurs in the sum of the newly generated cosine (due to the new impulses in the
spectrum) and the original cosine function.

It is important to note, however, that although this beating effect (sub-Nyquist artifact)
has a visible, low frequency of 2ε, it is not a true moiré, since no corresponding impulse
having the low frequency 2ε appears in the spectrum. In terms of the signal domain, too,
our beating effect (Fig. 8.13(f)) clearly differs from a true sampling moiré effect (Figs.
8.13(i),(j)): Unlike the latter, the former is not really a discrete cosinusoidal signal, but
rather a discrete modulated signal having alternating positive and negative values; and it is
only its global envelope shape that gives to this beating effect its cosinusoidal look. This
is, in fact, a modulation or “pseudo-moiré” phenomenon, just like the beating effect
which occurs in the sum of any two cosines with slightly different frequencies (see Sec.
D.3 in Appendix D). A similar sub-Nyquist artifact may also appear at other alias-free
combinations of signal and sampling frequencies (see, for example, [Williams00]).
Furthermore, such phenomena may also occur in 2D cases (see Fig. 8.15 or [Fielding06
pp. 132–134]) or in MD cases. Sec. D.3 of Appendix D provides the generalization of

Figure 8.15: 2D sub-Nyquist artifacts (pseudo-moiré effects) due to sampling in the

2D case. Each row shows a 2D cosine function g(x,y) = cos(2π[ux + vy])
with a different frequency vector f = (u,v), as indicated by the impulse
locations in the respective DFT spectra. Note that none of these frequency
vectors exceeds the borders of the DFT spectrum (i.e. the frequency range
–12 fs ... 12 fs in each dimension), meaning that no aliasing occurs in any of
the cases. And yet, as shown in rows (b)–(d), when the frequency vectors
are just slightly below half of the sampling frequency (which is also the
maximum frequency that can be represented in the DFT spectrum), a
strong low-frequency beating effect may appear, which is even more
prominent than the original sampled cosine itself. This beating effect has a
much lower frequency (i.e. a much larger period) than the original cosine,
and sometimes, like in row (c), even a different orientation. These artifacts
are not true moirés (note that there are no corresponding low-frequency
impulses in the respective spectra). The sampling frequency used in all of
the rows is identical (4 Hz in each of the two directions); the array size is
64×64 and the sampling range is –8...8 to each direction.
8. Miscellaneous issues 231

Signal domain Spectral domain

8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(a)
8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(b)
8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(c)
8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(d)
232 8. Miscellaneous issues

pseudo moirés to the 2D and MD cases, where the frequencies are 2D or MD vectors
rather than scalars. A more detailed discussion on pseudo-moiré effects in 1D, 2D or MD
can be found in [Amidror09a].

Example 8.2: A 2D sub-Nyquist artifact:

To take a concrete 2D example, let us study Fig. 8.15(c) in more detail. Here, the
sampling frequency to both directions is fs = 4 Hz, so that the maximum frequency
allowed by the 2D sampling theorem without causing aliasing is 12 fs = 2 Hz in each of the
two directions (note that this is also the maximum frequency representable in each
direction of the 2D DFT spectrum, as we have seen in Eq. (4.10)). The frequency of the
2D continuous cosine function g(x,y) = cos(2π[ux + vy]) being sampled in this case is
f1 = (1.875,0.25) Hz. To see this, note that in terms of 2D DFT elements (pixels) the two
corresponding impulses are located in the 2D DFT spectrum of Fig. 8.15(c) in the
discrete locations ±(30,4) pixels. Since 32 pixels (i.e. N/2) correspond to the maximum
frequency 12 fs = 2 Hz, it follows, indeed, that 30 pixels in the horizontal direction
correspond to 1.875 Hz, and 4 pixels in the vertical direction correspond to 0.25 Hz. Let
us now denote the difference between our cosine frequency f1 and the frequency (12 fs,0)
by (εu,εv). We have, therefore, (εu,εv) = (12 fs,0) – f1 = (2,0) – (1.875,0.25) = (0.125,–0.25),
where f1 = (12 fs,0) – (εu,εv).

Figure 8.16: Explanation in the continuous-world spectrum of the sub-Nyquist artifact

that occurs in Fig. 8.15(c) when sampling a continuous 2D cosine
function g(x,y) = cos(2π [ux + vy]) whose frequency f1 = (u,v) is just
slightly below half of the sampling frequency 12 fs in the horizontal
direction. (a) The continuous-world spectrum of the original cosine.
(b) The continuous-world spectrum of the sampled cosine. Note the
impulse pair of the original cosine, at the frequencies ±f1, and its replicas
which are centered about all the points (kfs, lfs) where fs is the sampling
frequency in both directions and k,l ∈Z. Thanks to the two neighbouring
replicas to the right and to the left of the original replica, the spectrum (b)
of the sampled cosine represents, in fact, a sum of two very close 2D
cosines: Our original continuous cosine, whose frequency f1 is slightly
below 12 fs, and a new continuous cosine, whose frequency f2 is slightly
above 12 fs. (Impulses of higher frequencies, that are also present in the
spectrum (b), are required for the correct spectral representation of the
sampled signal, but for the sake of our discussion here they can be
ignored). The sub-Nyquist artifact that appears in the sampled signal (see
Fig. 8.15(c)) is simply the 2D beating modulation effect that occurs in the
sum of the two very close 2D cosines having the frequencies f1 and f2.
(c) Schematic view of the frequency vectors ±f1, ±f2, their sums and their
differences. Note that the low beating frequency f2 – f1 itself, which is
clearly visible in the signal domain of Fig. 8.15(c), has no corresponding
impulse (black dot) in the spectrum (b), meaning that it is not a true moiré
effect. The line segments in (a)–(c) have been added to clarify the
geometric relations, but they are not really parts of the spectra.
8. Miscellaneous issues 233

fs
Borders of the corresponding
DFT spectrum
1f
2 s

f1
• u
– fs • 1f
2 s fs
–f1
–1
2 fs

– fs
(a)

• • •
fs
• • •
1f
2 s
–f2 f1
• • •
u
• – fs • • fs
–f1 1f
f2
2 s
– 1f
2 s

• • •
– fs
• • •

(b)

f1– f2
–f2 f1
– f1– f2 • • f1 + f2
u
–f
s • • fs
–f1 f2
f2 – f1

(c)
234 8. Miscellaneous issues

As we already know, continuous-world spectra are often easier to understand than DFT
spectra, since their frequency range is not limited, and they do not suffer from wrap-
around and foldover. For this reason, let us consider the continuous-world spectrum of
our 2D cosine rather than the DFT spectrum of Fig. 8.15(c). This continuous-world
spectrum is shown in Fig. 8.16. Row (a) of this figure shows the continuous-world
spectrum of the original 2D cosine before sampling, and row (b) shows the continuous-
world spectrum of the same 2D cosine after its 2D sampling. As we have seen in Chapter
5, the continuous-world spectrum of the sampled function consists of infinitely many
replicas of the original spectrum, that are centered about all the points (kfs, lfs) where fs is
the sampling frequency in both directions and k,l ∈Z (see Eq. (5.4) in Sec. 5.4). Fig.
8.16(b) shows only 9 of these replicas, namely, the original one (which is identical, up to a
certain amplitude scaling factor, to row (a) of the figure) plus its 8 nearest neighbours, that
are centered about the points (fs,0), (–fs,0), (0,fs), (0,–fs), (fs,fs), (fs,–fs), (–fs,fs) and (–fs,–fs).
As we can see in Fig. 8.16(b), the first two of these 8 new replicas, the ones that are
located to the right and to the left of the original replica, add to the continuous-world
spectrum a new impulse pair that is located very close to the original impulse pair ±f1 of
our cosine, i.e. at the frequencies f2 = (12 fs,0) + (εu,εv) and –f2 = –(12 fs,0) – (εu,εv). This new
impulse pair corresponds to a newly added cosine in the signal domain, whose frequency
is, in the case of Fig. 8.15(c), f2 = (2.125,–0.25) Hz (or (34,–4) pixels, in terms of the 2D
DFT elements, i.e. slightly outside the DFT borders). But as shown in Sec. D.3 of
Appendix D, the sum of two 2D cosines with very close frequencies f1 = (12 fs,0) – (εu,εv)
and f2 = (12 fs,0) + (εu,εv) gives a beating effect (pseudo moiré) whose frequency is f2 – f1.
And indeed, the beating effect we obtain here has the frequency of f 2 – f 1 =
((12 fs,0) + (εu,εv)) – ((12 fs,0) – (εu,εv)) = 2(εu,εv), namely, in the particular case shown in Fig.
8.15(c), (0.25,–0.5) Hz. In terms of the signal domain (see the left hand side of Fig.
8.15(c)), the period of this beating effect is, therefore, reciprocal to the length of the
frequency vector f2 – f1, namely 1/ 0.252+0.52 ≈ 1.79 (see Eq. (2.8) in [Amidror09 p.
18]) in the direction of the frequency vector f2 – f1 (although this frequency vector has no
corresponding impulse in the DFT spectrum, meaning that this beating effect is not a true
moiré). This beating effect is indeed a sub-Nyquist artifact and not an aliasing artifact,
since the frequency f1 of the original continuous cosine being sampled here does not
exceed the maximum frequency allowed by the 2D sampling theorem, 12 fs = 2 Hz to each
direction. Q

Now, just as in the 1D case shown in Figs. 8.14(c)–(d), because the frequency of our
original 2D cosine function is below 12 fs in each direction, this function can be perfectly
reconstructed from its sampled version, as stipulated by the 2D sampling theorem. This is
done by multiplying the spectrum of Fig. 8.16(b) with a 2D rect function extending from
–12 fs to 12 fs in both directions (or equivalently, in the signal domain, by 2D sinc-function
interpolation [Bracewell95 pp. 252–254], i.e. by convolving the sampled version of the
2D cosine signal with the 2D sinc function that is the inverse Fourier transform of the
above 2D rect function). This ideal reconstruction removes the new impulse pair from the
spectrum, and hence the corresponding new cosine disappears from the signal domain and
no pseudo moiré can be generated. But as we have seen in Sec. 8.5, such an ideal
8. Miscellaneous issues 235

reconstruction rarely occurs in real-world situations. A more realistic reconstruction

method may consist of 2D zero-order (i.e. nearest-neighbour) interpolation, which is
equivalent to convolving the sampled signal with a narrow 2D pulse or “pixel” (see
[Bracewell95 pp. 247–248] or [Stucki79 pp. 184–185]). This alternative reconstruction
method means, in terms of the spectral domain, multiplication of the spectrum (b) with a
non-ideal substitute of the ideal rect function. But when doing such a multiplication, the
new replicas that appeared in row (b) due to sampling will not be completely removed, and
some debris thereof may still subsist in the spectrum. In such cases, a low-frequency sub-
Nyquist artifact may indeed become visible in the sampled signal (see Fig. 8.15(c)); this is
simply the beating modulation effect that occurs in the sum of the newly generated cosine
(due to the new impulses in the spectrum) and the original cosine function.

As we have seen above, the beating effects (sub-Nyquist artifacts) which are visible in
Figs. 8.15(b)–(d) are not true moiré effects, since no corresponding low-frequency
impulses appear in their spectra. These beating effects are, in fact, modulation or pseudo-
moiré phenomena, just like the beating effect which occurs in the sum of any two cosines
with slightly different frequencies (Sec. D.3 in Appendix D). However, just as shown in
Fig. 8.13 for the 1D case, here, too, by judiciously modifying the frequency of the 2D
cosine being sampled we can obtain a true sampling moiré effect, having exactly the same
period and orientation as our beating effect. Figs. 8.17(b)–(d) show, indeed, for each of
the 2D sub-Nyquist artifacts of Figs. 8.15(b)–(d), its corresponding true sampling moiré.
This is obtained by sampling 2D cosines with higher frequencies than in Fig. 8.15, as we
will see below, while keeping the sampling frequency unchanged.

Example 8.3: A true 2D sampling moiré:

Let us study in more detail the true sampling moiré effect shown in Fig. 8.17(c), and
compare it with its pseudo-moiré counterpart of Fig. 8.15(c) that we have already analyzed
in Example 8.2.
Here, too, we prefer to use the continuous-world spectrum rather than the DFT spectrum
of Fig. 8.17(c). This continuous-world spectrum is shown in Fig. 8.18. Row (a) of this
figure shows the continuous-world spectrum of our original 2D cosine before sampling.
Note that the frequency vector f1 of this original 2D cosine is exactly twice the frequency
vector f1 of the original 2D cosine we used in Fig. 8.16(a); as we will see below, this
choice is not accidental. Row (b) of our figure shows the continuous-world spectrum of
the same 2D cosine as in row (a), after its sampling (note that we are using here the same
sampling frequency fs as in Fig. 8.16 to both directions). As we already know, this
spectrum consists of infinitely many replicas of the original spectrum, that are centered
about all the points (kfs, lfs) where fs is the sampling frequency in both directions and
k,l ∈Z (see Eq. (5.4) in Sec. 5.4). Just like in Fig. 8.16(b), row (b) of our figure shows
only 9 of these replicas, namely, the original one (which is identical, up to a certain
amplitude scaling factor, to row (a)) plus its 8 nearest neighbours, that are centered about
the points (fs,0), (–fs,0), (0,fs), (0,–fs), (fs,fs), (fs,–fs), (–fs,fs) and (–fs,–fs). But unlike in Fig.
8.16(b), this time the impulse pair belonging to each of the replicas is not confined inside
236 8. Miscellaneous issues

the replica itself. This happens because in our present case the frequency f1 of the original
2D cosine is larger than half of the sampling frequency (12 fs), meaning that we are
confronted here with aliasing. In terms of the signal domain this means that our sampling
rate is not sufficient for capturing all details of our given 2D cosine, so that the sampled
result will be mimicked by a lower frequency alias, like in Fig. 5.15(d).
If we look now at the main replica in Fig. 8.18(b), which also represents the territory of
the corresponding DFT spectrum, we see that it contains, in fact, only two impulses, both
of which exceed from the horizontal neighbouring replicas. These two impulses penetrate
into the very center of the main replica, and fall there close to the origin. These new low-
frequency impulses (that are due to the aliasing which occurs while sampling our original
2D cosine), correspond, indeed, to a new low-frequency moiré effect, which is clearly seen
in Fig. 8.17(c). It is important to note, however, that unlike the low frequency which is
visible in Fig. 8.15(c), the sampling artifact that we see in Fig. 8.17(c) is indeed a true
sampling moiré, and not a sub-Nyquist artifact: Unlike its counterpart of Fig. 8.15(c), it
does have a corresponding low-frequency impulse pair in the spectrum (the impulses
indicated in Fig. 8.18 by ±fM); and furthermore, in terms of the signal domain, it consists
of a true 2D low-frequency cosinusoidal signal rather than of a modulated signal made of
rapid alternating positive and negative values, as in Fig. 8.15(c). Compare also to the 1D
equivalent shown in Fig. 8.13, in which row (f) shows a sub-Nyquist artifact, and row (i)
shows its true sampling-moiré counterpart having the same frequency (and periodicity).
It is interesting to note, however, that in spite of the different nature of the low-frequency
sampling artifacts in Figs. 8.15 and 8.17 (respectively, sub-Nyquist artifacts or true
sampling moiré effects), their frequency (or period) as well as their orientation are exactly
the same in both figures. The reason is that when designing Figs. 8.17(b)–(d), we have
used in each of them an original 2D continuous cosine whose frequency vector f1 is
exactly twice the frequency vector of the corresponding cosine in Fig. 8.15 (namely, twice

Figure 8.17: True 2D sampling moirés due to sampling in the 2D case (compare with
the similar 2D sub-Nyquist artifacts shown in Fig. 8.15). Just like in Fig.
8.15, each row shows a 2D cosine function g(x,y) = cos(2π[ux + vy]) with
a different frequency vector f = (u,v). But this time, in each of the rows
(b)–(d) the impulses of the cosine function do exceed the borders of the
DFT spectrum (i.e. the frequency range –12 fs ... 12 fs in each dimension), and
thus they fold over and re-enter into the opposite end of the DFT
spectrum as aliased impulses. Since these aliased impulses fall at low
frequencies close to the origin of the DFT spectrum, they are
indistinguishable in the signal domain from the true signals having these
low-frequency impulses, and give there highly visible low-frequency
sampling moiré effects. Note that just like in Fig. 8.15 here, too, the
sampling frequency is identical in all of the rows (4 Hz in each of the two
directions); the array size is 64×64 and the sampling range is –8...8 to
each direction.
8. Miscellaneous issues 237

Signal domain Spectral domain

8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(a)
8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(b)
8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(c)
8. 2.

0 0

-8. -2.
-8. 0 8. -2. 0 2.
(d)
238 8. Miscellaneous issues

the frequency in the same orientation). To see how this works, let us consider the
particular case of Fig. 8.17(c). As we can see in the corresponding continuous spectrum
shown in Fig. 8.18(c), the two horizontally neighbouring replicas of the impulse pair ±f1
give close to the spectrum origin two new low-frequency impulses, fM = f2 – f1 and
–f M = f 1 – f 2, where f 2 = (fs,0) represents the sampling frequency in the horizontal
direction. If we compare Fig. 8.18(c) with its sub-Nyquist counterpart in Fig. 8.16(c), we
see that the basic geometry in both cases remains the same, with one major difference:
While in Fig. 8.18(c) the low-frequency vector f2 – f1 corresponds indeed to an impulse,
in Fig. 8.16(c) this low-frequency vector, which falls exactly in the same location, is not
endowed with an impulse of its own, and it therefore only represents there a pseudo-moiré
effect (sub-Nyquist artifact) having exactly the same frequency and orientation. Note that
in the 1D case of Fig. 8.13, too, the frequency of the original cosine being sampled in row
(i), which gives a true sampling moiré, is exactly twice the frequency of the original cosine
being sampled in row (f), and which gives the equivalent sub-Nyquist artifact.
Note that in the present moiré effect (Fig. 8.18) the frequency f2 does not correspond to
a second cosine signal, as it does in the sub-Nyquist artifact (Fig. 8.16) or in a classical
superposition moiré effect (see Fig. 2.2 in [Amidror09 p. 14]). Rather, f2 corresponds
here to the sampling frequency vector in the horizontal direction, (fs,0). For this reason our
present moiré effect is said to be a sampling moiré effect [Amidror09 p. 432]. Note that
the frequency f 2 itself does not appear in the DFT spectrum; as we can see in the
continuous-world spectrum (Fig. 8.18(b)) this frequency is indeed located beyond the
borders of the DFT spectrum. Q

Figure 8.18: Explanation in the continuous-world spectrum of the true moiré effects
that occur in Fig. 8.17(c) when sampling a continuous 2D cosine function
g(x,y) = cos(2π[ux + vy]) whose frequency f1 = (u,v) is exactly twice than
in Fig. 8.15(c). Compare with the continuous-world spectrum explanation
of Fig. 8.15(c), which is provided in Fig. 8.16. (a) The continuous-world
spectrum of the original cosine. (b) The continuous-world spectrum of
the sampled cosine. Note the impulse pair of the original cosine, at the
frequencies ±f1, and its replicas which are centered about all the points
(kfs, lfs) where fs is the sampling frequency in both directions and k,l ∈Z.
Thanks to the two neighbouring replicas to the right and to the left of the
original replica, the spectrum (b) of the sampled cosine includes very
close to its origin two new impulses that are only due to the sampling.
These new impulses, noted by ±fM, correspond to a new low-frequency
structure in the signal domain, which is indeed a true sampling moiré
effect, and not a sub-Nyquist artifact. (c) Schematic view of the frequency
vectors ±f1, ±f2, their sums and their differences. Note that unlike in Fig.
8.16(c), the low frequency f2 – f1 itself, which is clearly visible in the
signal domain of Fig. 8.17(c), does have a corresponding impulse (black
dot) in the spectrum (b), meaning that in this case it is a true moiré effect.
The line segments in (a)–(c) have been added to clarify the geometric
relations, but they are not really parts of the spectra. The gray area in (c)
shows the geometry of Fig. 8.16(c), for the sake of comparison.
8. Miscellaneous issues 239

fs
Borders of the corresponding
DFT spectrum
1f
2 s
f1
•
1f
u
– fs 2 s fs
•
–f1
–1
2 fs

– fs
(a)

• • •
fs
• • •
1f
2 s
– fM f1
• • •
u
– fs fs
• • •f 1f
–f1 M
2 s
– 1f
2 s
• • •
– fs
• • •

(b)

– fM = f1 – f2 f1 f1+ f2
• • •
– f2 f2 = (fs ,0)
u
–f fs
s
• • •
– f1– f2 –f1 fM = f2 – f1
(c)
240 8. Miscellaneous issues

Another interesting point to be mentioned here is a direct result of the periodicity of the
DFT spectrum. As we have seen in Sec. 2.5, this periodicity reflects, in fact, the infinitely
many replicas of the range –12 fs ... 12 fs along each direction, which appear in the continuous-
world spectrum due to the sampling of the input signal (see Fig. 2.1 for the 1D case).
Now, according to Sec. D.3 of Appendix D, any two neighbouring impulse pairs in the
continuous-world spectrum (which represent in the signal domain a sum of two very close
cosine functions) may cause a pseudo-moiré phenomenon (a modulation effect) between
the two similar cosines. The sub-Nyquist artifact is, indeed, a particular case of this
phenomenon, as we clearly see in the continuous-world spectrum (Fig. 8.16(b)). The
difficulty in realizing this in the DFT spectrum (Fig. 8.15(c)) is simply due to the fact that
these neighbouring impulses are located there far away from each other, close to opposite
borders of the DFT spectrum. The fact that these impulses are close neighbours is
explained, indeed, by the periodicity of the DFT spectrum. But this fact is much easier to
understand in the continuous-world spectrum, where these impulses really appear next to
each other (although they fall there to both sides of a boundary between two neighbouring
replicas). We can therefore formulate the following remark:

Remark 8.3: DFT spectra are usually more difficult to understand than their continuous-
world counterparts, since their frequency range is limited and they may suffer from
foldover of higher frequencies, or from other DFT artifacts such as leakage. But sub-
Nyquist artifacts show us that DFT spectra may be difficult to understand even when they
are free from artifacts such as aliasing and leakage: Impulses that are close neighbours in
the continuous-world spectrum (such as f1 and f2 in Fig. 8.16(b)) may still appear in the
corresponding DFT spectrum far away from each other, in opposite ends of the spectrum.
And indeed, as we have seen in the examples above, a glance at the continuous-world
spectrum may often shed new light into the DFT spectrum, and help in its correct
interpretation. Q

Remark 8.4: Another pitfall relating to DFT spectra originates from a possible ambiguity
in the use of the terms “wraparound” and “foldover”. See Problem 8-17 and Remark
C.10 at the end of Appendix C. Q

In conclusion, the surprising fact that low-frequency artifacts may exist in a sampled
signal even when the sampling rate is located within the frequency range in which
sampling is supposed to be safe by virtue of the sampling theorem, is based, in fact, on a
misconception. “Safe” in terms of the sampling theorem does not mean that under the
specified conditions (i.e. when the sampling rate is at least twice the highest frequency
present in the signal) the sampled signal is perfectly faithful to the original continuous-
world signal, and does not show new parasite structures due to sampling. The sampling
theorem simply states that under the specified conditions there is no aliasing, and the
original continuous signal can be perfectly reconstructed from its sampled version by
convolution with a specified sinc function. But no guarantee is provided that the sampled
signal looks to us exactly the same as the original one, and that it has no new apparent
artifacts. This is true, of course, for signals of all dimensions, 1D, 2D or MD.
8. Miscellaneous issues 241

8.7 Displaying considerations

We have seen throughout the previous chapters several displaying considerations, that
are helpful and sometimes even crucial for the correct interpretation of the DFT results.
These include, among others, the reorganization issues (Chapter 3); the use of true units
along the axes (Chapter 4); the optimal choice of the DFT parameters N, Δx, Δu, etc.
(Chapter 7); the size of the final drawing, which may also affect the visibility of detail in
the displayed data (Sec. 7.2); the possibility of plotting only the most interesting part of
the data (Sec. 7.2); the choice of the number of periods to be taken into consideration, in
case of a periodic function (Remark 7.1); and the reconstruction issues related to the
plotting method being used to display the data (Sec. 8.5).

Another important displaying consideration concerns the amplitude of the DFT

spectrum (or of its magnitude). Plots of the DFT spectrum or magnitude tend to be
visually dominated by very low frequencies and in particular by the zero frequency (DC);
these spectral components are often so strong that the entire DFT plot (after
reorganization) seems to consist of a single spot in the center, surrounded by weak,
negligible data. This is, of course, highly undesirable, since most of the interesting
information usually resides at frequencies farther away from the center. One possible way
to deal with this problem (that we have indeed used in some of our figures) is to truncate
the exceeding values of the DC and of the surrounding low frequencies, in order to adapt
the scaling of the vertical axis to the values of the farther and more interesting parts of the
spectrum. Another possible solution is to use a non-linear scale, such as a logarithmic
scale, for displaying the DFT values along the vertical axis (or even along other axes).

When choosing the most judicious plotting method for displaying the given data and the
DFT results, one should also consider the various printer-dependent printing artifacts that
may occur. Typical examples are irregularities that may appear when printing large
uniform gray level spans, or moiré effects which may occur when printing a periodic
structure (due to an interference with the halftone screen or the underlying periodic grid
used by the printing device [Amidror09 p. 432]). Note that moiré artifacts are not only
limited to printed plots, and they may also occur when viewing periodic structures on
computer displays (again, due to their discrete nature).

A more detailed discussion on the graphic presentation of the DFT input and output
signals in one, two or more dimensions is provided in Sec. 1.5.

8.8 Numeric precision considerations

As we already know, a multitude of errors of various origins may occur when using the
DFT to approximate the continuous Fourier transform. Of most significant importance are
the errors due to the aliasing and leakage artifacts, as explained in detail in Chapters 5 and
6. What happens, however, when the required conditions are met and no leakage and
aliasing occur? Does the DFT in such cases perfectly approximate the continuous Fourier
242 8. Miscellaneous issues

transform of the function underlying the data samples? The answer is, of course, no: The
DFT can only be an approximation, since it is only based on a finite number N of input
values (samples of the original function), and it yields the same number N of output values
(discrete frequencies). But are these discrete values themselves precise samples of the
continuous Fourier transform? The answer is, again, no, even if no aliasing and leakage
artifacts occur. The reason can be best understood when we investigate the mechanism in
which the DFT approximates the continuous Fourier transform. As already mentioned in
Remark 2.2, the DFT can be derived from the continuous Fourier transform as a
numerical approximation to the integrals of Eqs. (2.1)–(2.2), using the rectangular (or the
trapezoidal) rules. These rules are simple, straightforward methods for numerically
calculating integrals; they approximate the area below the function curve by a sum of N–1
narrow rectangles (or trapezoids) of width Δx (see, for example, [Harris98 pp. 567–568]
or [Spiegel68 p. 95]). This approximation process unavoidably contains certain
discretization errors that arise because the step Δx is small but not zero. Letting N→∞ and
Δx→0 reduces this error, but in any realizable case some discretization error will subsist.
A more detailed discussion on the inherent errors in the approximation of the Fourier
transform by DFT can be found in [Briggs95 pp. 53–55], and for the case of the inverse
DFT in [Briggs95 pp. 56–59 and Sec. 6.8].

Another potential source of error in the computation of the DFT is due to the precision
of the hardware and software being used; for example, one may consider the effects of
word length and of round-off or quantization errors on the precision of the DFT results.
But when modern computers are being used, this source of error can be considered
as marginal and it can be usually neglected (except, perhaps, when considering very
unstable situations like the phase of the DFT spectrum in areas where the absolute value
of the DFT is almost zero (see Sec. 8.3)). A discussion on such numeric precision
considerations can be found, for example, in [Oppenheim75, Sec. 9.5].

PROBLEMS

8-1. Representation of a rect function in the discrete world. As we have seen in Sec. 8.2, the
correct discrete representation of the rect function should take into account the
midvalue rule, as shown in Fig. 8.2(a). Without taking this subtlety into account, the
resulting DFT will not give the expected sinc function but rather a somewhat corrupted
version thereof (see Figs. 8.2(b),(c)). However, even if we do take into account the
midvalue rule, the resulting DFT does not perfectly fit the expected sinc function,
because of some unavoidable aliasing effects (remember that the rect function is not
band limited). Suppose that you insist on getting in your DFT output a good match
with the expected continuous-world sinc function, i.e. a faithful sampled version of the
sinc function which is not flawed by aliasing or other artifacts. How should the rect
function be represented in the input signal in order to obtain this goal? Hint: Have a
look at Fig. 8.1.
8-2. Representation of a rect function in the discrete world (continued). Figs. 8.1 and 8.2
show several different discrete representations of the rect function. Make a list of the
8. Miscellaneous issues 243

various representations, explain each of them and discuss the resulting DFT spectra.
Can you think of other possible representations and their impact on the resulting DFT?
Which of the different representations are most used in practice? You may have a look
at the figures in various books or internet resources.
8-3. The discrete counterpart of the midvalue rule. We have seen in Sec. 8.2 that if a
discontinuity of a continuous-world function falls between the discrete elements n and
n+1, then these discrete elements take intermediate values which depend on their
relative distance from the true discontinuity point between them. This rule is applied
automatically by the DFT in the output array whenever a discontinuity occurs in the
continuous-world spectrum. But when it is our role to sample a continuous-world
function having a discontinuity, the application of this midvalue rule during the
sampling process is clearly less obvious. The best we can do in this case is to
reasonably estimate (interpolate) the intermediate values at the two input array elements
n and n+1 in accordance with their respective distance from the discontinuity point
between them. However, interpolation can be done in many different ways, each giving
slightly different results. What kind of interpolation should we apply here to get the
most accurate results?
8-4. The original continuous-world function being sampled in Fig. 8.2(a) is a rect function
whose width is 1. However, the sampled signal we obtain in this case could be also
considered as the sampled version of a continuous-world trapezoid (to see this, connect
each of the midvalue points to its two direct neighbouring points by a straight line). It
is obvious that this trapezoid and the rect function have different continuous Fourier
transforms, but since they are represented in the discrete world by the same discrete
signal, they will have identical DFT spectra. How do you explain this contradiction?
8-5. We have seen in Sec. 8.4 that a symmetric rect function in the discrete world must have
an odd number of 1-valued elements across its width, namely, the center of the
rectangle (which corresponds to x = 0), and to each of its sides an equal number of
neighbouring 1-valued elements (usually followed by an element with the midvalue 12 ,
as we have seen in Sec. 8.2). Any other possibility will result in a non purely-real DFT.
Interestingly, this is true independently of the array length N, be it an even or an odd
number. For example, using the Mathematica® software package, which is capable of
computing the DFT of an input sequence of any given length N [Wolfram96 p. 1093],
we find that for N = 4 the DFT of the sequence {1,1,0,1} is purely real-valued,
{1.5,0.5,–0.5,0.5}, while the DFT of the sequence {1,0,0,1} is complex. Similarly, for
N = 5, the DFT of the sequence {1,1,0,0,1} is purely real, {1.34,0.72,–0.28,–0.28,0.72},
while the DFT of the sequence {1,1,0,1,1} is complex. As we can see, although the
input sequences {1,0,0,1} and {1,1,0,1,1} look symmetric, in terms of the DFT they are
not considered as even. Explain these results. Note that all the sequences are given here
as they are actually fed to or received from the DFT, i.e. with no reorganizations or
amplitude corrections; in particular, their first element corresponds to x = 0 (in the
signal domain) or to the DC (in the spectral domain). Obviously, before plotting these
sequences or interpreting them as CFT approximations, reorganization and amplitude
corrections should be done in order to avoid confusion.
8-6. As mentioned at the end of Sec. 8.4, in order to avoid abnormal situations where 2D
continuous-world Fourier symmetry rules are violated in the 2D discrete case, we have
intentionally zeroed the last row and/or the leftmost column in the 2D input array
whenever required. This was done, for example, in Figs. 5.13, 5.14, 5.15 and 8.8(a),
but not in other cases such as Figs. 5.16, 5.17 and 8.7(a). Discuss each of these cases
and explain why a correction was or was not required there.
244 8. Miscellaneous issues

8-7. Can you guess how we have generated the relatively complicated yet perfectly
symmetric pattern shown in row (d) of Fig. 8.4? Hint: This is simply a lower-resolution
counterpart of Fig. 5.18(e), where the circular nature of the original continuous-world
function being sampled is rendered unrecognizable due to undersampling (aliasing).
8-8. Jaggies and aliasing. Consider Fig. 8.7(a). The continuous-world spectrum of a
rotated grating consists of an infinite impulse comb passing through the origin, which
is rotated by the same angle as the grating. But as we can see in this figure, the DFT
counterpart of this spectrum contains also many impulses which are not located along
this comb. Student A says that from the point of view of aliasing, these impulses are
simply folded-over continuations of the main impulse comb which exceed from one
end or the other of the DFT spectrum (due to the insufficient sampling rate) and re-
enter from the opposite end. Student B, however, says that from the point of view of the
jaggies, the impulses in question simply represent the jaggies which are present in the
sampled grating of Fig. 8.7(a), and which do not exist in the continuous-world grating.
Student C claims that both of these interpretations are correct, and that this is indeed
the reason why such jaggies are considered in the literature as aliasing effects (both
jaggies and aliasing being side effects of the same sampling process). What is your
opinion? Can you extend your reasoning to Fig. 8.8? And to Figs. 8.5(a),(b)?
8-9. Suppose that the grating of Fig. 8.7(a) was not rotated, so that its lines were perfectly
vertical. In this case, no jaggies are visible in the discrete, sampled image (see Fig.
8.6(a) and Fig. 8.9). Can you explain this case, too, using the same reasoning?
8-10. The spectra shown in Figs. 8.6–8.9 were all obtained by DFT, and therefore they only
include the frequency range going between –12 fs and 12 fs along each axis, where fs is the
sampling frequency in both directions. However, the spectral domain explanation of
2D sampling is easier to understand in the continuous-world spectrum, which extends
throughout the u,v plane and clearly shows the replicas due to the convolution with the
2D nailbed corresponding to the 2D sampling frequencies (see Fig. 5.12). Draw the
continuous-world counterpart of each of the cases shown in Figs. 8.6–8.9 and explain
what you see in the signal and spectral domains in each of these cases.
8-11. In particular, consider the continuous-world spectrum that corresponds to the
unrotated vertical grating shown in Fig. 8.6(a) (the spectral domain shown in this
figure is the DFT counterpart of this continuous-world spectrum). As we already know,
the continuous-world spectrum of an unrotated vertical grating consists of a comb of
impulses along the u axis, whose amplitudes are modulated by a sinc function. Due to
the 2D sampling, this continuous-world spectrum is convolved with a 2D nailbed whose
impulse distance to both directions equals the sampling frequency fs. This convolution
generates replicas of our comb throughout the u,v plane, away from the u axis, too. A
student asks the following question: Since in this case there are no jaggies, the sampled
grating (Fig. 8.6(a)) looks exactly the same as the original unsampled grating. How
can it be that their continuous-world spectra are nevertheless different? Hint: Although
no jaggies appear in this case, the sampled and unsampled vertical gratings are not
really identical: the former only consists of impulses, while in the latter the lines are
really continuous. Note that what we see in the signal domain of Fig. 8.6(a) is not
really the sampled grating, but rather its reconstructed version, where each impulse is
convolved with a square pixel. This reconstructed signal looks, indeed, very similar to
the original unsampled grating (up to a small possible difference in the line widths, due
to their rounding to the closest multiple of full square pixels).
8-12. Fig. 8.10 shows a slightly rotated low-frequency cosinusoidal grating, which is clearly
band limited and has no sampling aliasing. If reconstruction is done correctly, as
8. Miscellaneous issues 245

stipulated by the sampling theorem, no jaggies should appear in the reconstructed

signal; and if jaggies do appear — as in Fig. 8.10 — this is only due to poor
reconstruction, which causes reconstruction aliasing. And indeed, a student says that
jaggies are not aliasing artifacts, as often claimed in the literature, but rather
reconstruction artifacts (see also [Gomes97 p. 197; Wolberg90 pp. 107–108]). What is
your opinion? Hint: Note that in cases like Figs. 8.7 and 8.8 ideal reconstruction alone
will not eliminate the jaggies. To see this, remember that ideal reconstruction is
equivalent to multiplying the spectrum of the sampled signal with a rect function that
cuts off all the frequencies beyond half of the sampling frequency. But looking at the
spectrum in Figs. 8.7 and 8.8 we see that this multiplication will not eliminate the stray
impulses that are present within this range and the artifacts they represent in the signal
domain.
8-13. Sub-Nyquist artifacts. Although in Sec. 8.6 we insist on the fact that sub-Nyquist
artifacts may occur at frequencies below half of the sampling frequency, so they are
certainly not caused by aliasing, Fig. 8.13 shows that a similar artifact may also occur
above half of the sampling frequency, where aliasing does occur (see row (h) in the
figure, and compare it with row (f)). Note that in both cases no impulses belonging to
the new, visible low frequency in question is present in the spectrum. In Sec. 8.6 we
concentrated on the explanation of the case shown in row (f) of the figure, but the
explanation of the case shown in row (h) is quite similar. Can you give the detailed
explanation for this case?
8-14. Sub-Nyquist artifacts. As we saw in the previous problem, sub-Nyquist artifacts may
also occur at frequencies above half of the sampling frequency, so that the term “sub-
Nyquist artifacts” is not really optimal. Can you think of a better term, for example a
term reflecting the main property of these artifacts, the fact that they do not possess in
the spectrum impulses that correspond to their frequencies?
8-15. Sub-Nyquist artifacts. After carefully studying Sec. 8.6, a student points out that the
sub-Nyquist artifact is not generated by reconstruction errors, as sometimes suggested
in the literature. This beating effect already exists in the sampled signal itself, much
before the reconstruction stage. Simply, while ideal reconstruction would eliminate this
beating effect from the reconstructed signal, non-ideal reconstructions fail to eliminate
it (or even may further amplify it), so that this sampling artifact remains visible in the
reconstructed continuous signal, too. In other words, reconstruction errors are not the
cause of this beating effect, but they only allow it to subsist after the reconstruction
step. What is your opinion?
8-16. Looking at Fig. 8.15(d), student A says that the frequency of the given 2D cosine
(which is represented by the two impulses at the corners of the DFT sprctrum) exceeds
beyond half of the sampling frequency, 12 fs = 2 Hz, since the distance of each of these
impulses from the origin is almost 12 2 fs, which is higher than 12 fs. Student B, however,
says that our 2D cosine does not exceed half of the sampling frequency, since in both
of its impulses, located at ±(u,v), each of the Cartesian components u and v is below 12 fs.
What is your opinion?
8-17. Foldover vs. wraparound. We have seen in Remark 8.4 that the terms foldover and
wraparound should be used with care in order to avoid confusion. Explain the exact
meaning of each of these terms, the difference between them, and the precise
circumstances in which they can be used without leading to confusion.
Hint: We have seen in Sec. 2.5 and in Chapter 5 that the sampling of an original
continuous signal generates in the continuous-world spectrum new frequencies due to
the replication of the original spectrum. When aliasing occurs, frequencies from the
246 8. Miscellaneous issues

new replicas penetrate additively into their neighbouring replicas, and in particular into
the main replica about the origin (which corresponds also to the territory of the DFT
spectrum). However, as we have seen in Chapter 5, this additive intrusion into
neighbouring replicas can be also considered from the point of view of the DFT
spectrum as folding over, namely: “impulses (or other spectral elements) that exceed
from one end of the DFT spectrum re-enter from the opposite end”. This can be seen,
for example, in Figs. 8.18(b), 5.13, etc. This interpretation as folding over is, indeed,
convenient and helpful for the explanation of aliasing in the DFT spectrum, but it
should not be confused with wraparound. Consider Fig. 8.16(b), where no aliasing
occurs, and no new impulses penetrate into the territory of the main replica (and hence
into the DFT spectrum). Because of the periodicity of the DFT spectrum along both
axes, one can say that each of the identical replicas in the continuous-world spectrum
(due to sampling) wraps around and coincides with the main replica (which represents
the DFT spectrum). However, in this case the term folding over should be avoided since
no addition is involved here.
Let us illustrate the difference using Figs. 8.16 and 8.18. Consider first Fig. 8.18(b),
and suppose that the original continuous-world spectrum already contained impulses
of its own at the locations ±fM. After sampling, because of the aliasing effect, replicas of
the impulses ±f 1 penetrate into the main replica, and fall there on top of the already
existing impulses ±f M . As already mentioned, this can be also considered as folding
over. Since the replicas due to sampling are generated by a convolution operation
(more precisely, the convolution of the original spectrum with a 2D nailbed containing
all the integer combinations of the sampling frequency vectors), our impulse
superposition due to folding over implies that the amplitudes of the coinciding
impulses be summed up. Let us now return to Fig. 8.16(b), where no aliasing occurs
and no impulses penetrate into neighbouring replicas. In this case, too, it is legitimate
to say that due to the periodicity of the DFT spectrum along both axes all the other
replicas wrap around and coincide with the main replica (the DFT territory). But of
course, an existing impulse and a wrapped-around impulse which coincide within the
main replica are not summed up: They are simply identified with each other (or
construed as congruent modulo the DFT spectrum periodicity21), meaning that each of
them represents in fact the same entity in the DFT. In this case the use of the term
wraparound is legitimate, but not the use of the term foldover (which implies the
addition of amplitudes, and is only to be used in the case of aliasing). See also Remark
C.10 at the end of Appendix C.
8-18. We saw in Sec. 8.4 that in the continuous world, when g(x) is real valued, G(0) is real
valued, too (since the Hermitian nature of G(u) implies that Im[G(0)] = –Im[G(0)],
which means in turn that Im[G(0)] = 0). However, in the discrete world it turns out that
when the input signal of the DFT is real valued, not only the output element
corresponding to the zero frequency is real valued, but also the output element
representing the highest (or lowest) frequency in the spectrum, which is located N/2
elements away from the zero-frequency element [Briggs95 pp. 77, 120]; see, for
example, the leftmost spectral element in Figs. 3.1(d) and 3.3(f). How can you
demonstrate this? Hint: See point (1) in Sec. D.2 of Appendix D.

21A number a is said to be congruent to b modulo p if p divides the difference a – b [Weisstein99 p.

303], i.e. if a – b is an integer multiple of p (which means that a – b = kp for some integer k). Thus, if
F is the frequency range length of the DFT spectrum, then the continuous-world frequencies f2 and f1
are congruent mofulo F if f2 = f1 + kF for some integer k. Congruence of frequency vectors f2 = (u2,v2)
and f1 = (u1,v1) modulo the 2D DFT spectrum range (F,F) is interpreted component-wise. The extension
to the MD case is straightforward, too.
 Appendix A
Impulses in the continuous and discrete worlds

A.1 Introduction
The mathematical concept of an impulse is widely utilized in many disciplines, including
physics, electricity, optics, etc. It is used to represent entities that are concentrated in a
single point, such as a point mass, a point charge, etc. Although such entities do not exist
in the real world, this theoretical concept is very useful since it may considerably simplify
derivations [Bracewell86 Chapter 5; Brigham88 Appendix A].

A.2 Continuous-world impulses vs. discrete-world impulses

In the continuous case, an impulse δ(u) (also known as a Dirac delta function) is
defined by (see [Bracewell86 p. 70; Jain89 p.12; Wolberg90 p. 15; Chu08 p. 188]):1

δ(u) = 0 u≠0
∞
(A.1)
and:
∫ –∞
δ(u) du = 1

This impulse, or more generally its shifted counterpart δ(u–a) that is located at the point
u = a, is therefore considered to have a unit integral (or area); but since it is concentrated
on a single point u = a and has there zero width it must have an infinite amplitude. Thanks
to its sifting property, when this impulse samples a given function g(u), as in g(u)δ(u–a),
its integral (area) takes the value g(a), while its amplitude remains, of course, infinitely
high [Bracewell86 pp. 74–77]. By convention, an impulse δ(u–a) is often represented
graphically as a spike of unit height; more generally, impulses are plotted as spikes of
height equal to their integral [Bracewell86 p. 75].2 For this reason the Fourier transform
of the constant function g(x) = 1, G(u) = δ(u), is graphically represented by a spike of
height 1 at the origin, and the Fourier transform of the periodic function g(x) = cos(2πfx),
G(u) = 12 δ(u–f) + 12 δ(u+f), is graphically represented by two spikes of height 12 located at
the frequencies ±f [Brigham88 pp. 19–20; Bracewell86 p. 412]. In reality each of these
impulses is infinitely high, but according to this graphical convention the spike heights
indicate the value of the integral. This convention is very practical, because it really
conveys the most important attribute of the impulse, i.e. its integral (and not its height,

1 Note that δ(u) is not a function in the usual mathematical sense of this term; it is normally defined as
the limit of the integral of a suitable series of functions, as explained, for example, in [Bracewell86,
Chapter 5].
2 Such a spike can be graphically represented by a vertical line of the corresponding length, with or
without an arrow at its end (see, for example, Fig. 2.1).

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 247
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_9,
© Springer-Verlag London 2013
248 Appendix A: Impulses in the continuous and discrete worlds

which is always infinite). But this convention may be somewhat dangerous, since it may
induce one to erroneously believe that the spike’s height represents the amplitude (or
height) of the impulse, which is obviously wrong.3

These concepts can be also extended without difficulty to the 2D and MD cases (see, for
example, [Bracewell86 pp. 84–85]).

In the discrete case, however, the situation is somewhat different. A discrete impulse
(also known as a Kronecker delta function) is defined by (see, for example, [Jain89 p. 12;
Wolberg90 p. 15; Smith07 pp. 122, 184; Chu08 p. 219]):4
⎧1 n=0
δ(n) = ⎨ (A.2)
⎩0 otherwise
Therefore, in the output array of the DFT, the discrete counterpart of the shifted
continuous-world impulse δ(u–a) is an array of zero-valued elements, with a single
1-valued element that is located in the array position n which corresponds to the
continuous-world frequency u = a.5 Thus, a discrete impulse consists of one non-zero
array element. (Concerning the “width” of such an element see Footnote 14 in Sec. 8.5
for the 2D case; its extension to the 1D and MD cases is straightforward.)

By convention, a discrete impulse cδ(n) is usually represented graphically as a spike of

height c. Such a spike can be drawn as a vertical line of the corresponding length, with or
without an arrow at its end (see, for example, Fig. B.1). Alternatively, when plotting our
discrete data as a symbol plot or as a connected line plot, the discrete impulse is
represented by a single dot or by a narrow isosceles triangle of height c, respectively (see
Sec. 1.5.1 and Fig. 1.2). Note that in Appendix A we will only use connected line plots.

A.3 Impulses in the spectral domain

Having seen the definition and the graphical representation of impulses in the discrete
world, what can be said now about the height of a discrete impulse in the DFT spectrum?
Consider, for example, the continuous function g(x) = 1, whose spectrum is G(u) = δ(u).
The discrete counterpart of g(x) consists of a sequence of N samples all having the value
of 1; if we take its DFT we obtain a sequence consisting of zeroes except for the element

3 Sometimes the term “impulse amplitude” is used, by abuse of language, referring to the spike’s height
in the plot (i.e. the value of the integral). This practice can be tolerated as long as it does not lead to
confusion. A better choice would be “impulse strength”, as in [Bracewell86 p. 76; Castleman79 p. 228].
4 In fact, in the context of the DFT, where both the input and the output arrays are considered to be

periodic with period N (the array length), the discrete impulse, too, may be considered as periodic with
period N. In this case, the condition in the first row becomes: “n = 0 or a multiple of N” like in
[Briggs95 p. 28].
5 It is assumed here that this frequency falls exactly on an output array element of the DFT and not
between two elements (i.e. that the value of n which corresponds to the frequency u = aHz according to
Eq. (4.14) is precisely integer), in order to avoid leakage artifacts (see Sec. 6.5). Otherwise, the correct
discrete counterpart of the impulse is a sampled, narrow sinc function (see Remark 6.5).
 Appendix A: Impulses in the continuous and discrete worlds 249

Sampled g(x) DFT: Real part DFT: Imaginary part

16 16
3 N = 64, R = 8
12 12
2
1 8 8

-4 -2 2 4
x 4 4

-1
-4 -2 2 4 Hz -4 -2 2 4 Hz

16 16
3
N = 64, R = 16
12 12
2
1 8 8

-8 -4 4 8
x 4 4

-1
-2 -1 1 2 Hz -2 -1 1 2 Hz

Figure A.1: The discrete counterpart of the function g(x) = 1 and its DFT spectrum
(after the corrections of Chapters 3 and 4). The array length (number
of samples) being used is N = 64. Top row: with a sampling range
length of R = 8 (i.e. sampling interval Δx = 8/64 = 0.125); bottom row:
with a sampling range length of R = 16 (i.e. sampling interval Δx =
16/64 = 0.25). Note the difference in the impulse heights between the
two DFT spectra. In the continuous case this impulse is represented in
the spectrum by a spike of height 1.

representing the DC, whose height has a finite positive value that depends on the DFT
definition being used (see, for example, [Bracewell86 p. 364]; [Weaver89 p. 250]). After
applying the standard amplitude scaling correction (see Sec. 4.4) one could expect the
height of this DC impulse to be 1, in agreement with the conventional graphic
representation of its continuous counterpart as a spike of height 1. However, as shown in
Fig. A.1, it turns out that the height of the discrete DC impulse we obtain by the DFT
(after the standard amplitude scaling correction) is not 1 but rather R = NΔx, the length of
the range within which the original function has been sampled to form our N-elements
input array (i.e. xmax – xmin).

As a second example, consider the continuous function g(x) = cos(2π fx), whose
spectrum is G(u) = 12 δ(u–f) + 12 δ(u+f). Even if this function is chosen and sampled so as to
avoid DFT artifacts (aliasing and leakage), as explained in Chapters 5 and 6, the impulse
heights obtained by the DFT after performing the standard amplitude scaling correction
are not 12 , as could be expected from the conventional graphic representation used in the
continuous case. Rather, just as in our first example, the height of each impulse is further
multiplied by R = NΔx, the length of the range within which the original function has been
sampled to form our N-elements input array (see Fig. A.2).
 250 Appendix A: Impulses in the continuous and discrete worlds

Sampled g(x) DFT: Real part DFT: Imaginary part

8 8
2 N = 64, R = 8
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

4 Hz 4 Hz
-2
-4 -2 2 -4 -2 2

N = 64, R = 16 8 8
2
1 6 6

-8 -4 4 8
x 4 4

-1 2 2

2 Hz 2 Hz
-2
-2 -1 1 -2 -1 1

Figure A.2: The discrete counterpart of the function g(x) = cos(2πfx) (with f = 12 )
and its DFT spectrum (after the corrections of Chapters 3 and 4).
The array length being used is N = 64. Top row: with a sampling
range length of R = 8 (i.e. sampling interval Δx = 8/64 = 0.125);
bottom row: with a sampling range length of R = 16 (i.e. sampling
interval Δ x = 16/64 = 0.25). Note the difference in the impulse
heights between the two DFT spectra. In the continuous case these
impulses are represented in the spectrum by spikes of height 12 .

This rule applies, indeed, to the DFT of all functions having impulsive spectra. Similarly,
in the 2D case any impulse amplitude in the DFT output is multiplied by R2 = N2Δx2, and
in the general MD case the factor is RM = NMΔxM.

It follows, therefore, that the height of the discrete impulse in the DFT spectrum is not
fixed, since it varies according to the choice of R. This may seem quite surprising, because
in functions having non-impulsive spectra the height of the DFT spectrum (after the
standard amplitude scaling correction) corresponds to the true height of the original
underlying continuous spectrum (except for possible aliasing or leakage distortions, if
any), and it does not depend on the choice of R. For example, as shown in Fig. A.3, the
height of the DFT spectrum of the function sinc(2x), after the standard amplitude scaling
correction, is 12 independently of the choice of R.

In order to understand this seemingly abnormal behaviour of the DFT in cases involving
impulsive spectra, let us recall the following Fourier-theory results. In the case of the
continuous Fourier transform we have [Bracewell86 p. 136]:
∞
G(0) =
∫ –∞
g(x) dx (A.3)
 Appendix A: Impulses in the continuous and discrete worlds 251

Sampled g(x) DFT: Real part DFT: Imaginary part

8 8
2 N = 64, R = 8
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

4 Hz 4 Hz
-2
-4 -2 2 -4 -2 2

8 8
2 N = 64, R = 16
1 6 6

-8 -4 4 8
x 4 4

-1 2 2

2 Hz 2 Hz
-2
-2 -1 1 -2 -1 1

Figure A.3: The discrete counterpart of the function g(x) = sinc(2x) and its non-
impulsive DFT spectrum (after the corrections of Chapters 3 and 4).
The array length being used is N = 64. Top row: with a sampling range
length of R = 8 (i.e. sampling interval Δx = 8/64 = 0.125); bottom row:
with a sampling range length of R = 16 (i.e. sampling interval
Δx = 16/64 = 0.25). Note that the height of the two DFT spectra is 12 ,
just as in the continuous spectrum, and it does not depend on R.

This result is obtained from Eq. (2.1) for the frequency u = 0. (Note, however, that if we
define the continuous Fourier transform using other conventions, such as Eq. (2.5), we
may have here an additional constant factor preceding the integral.)

The discrete counterpart of Eq. (A.3) can be easily derived by inserting n = 0 in Eq.
(2.13):
N– 1
G(0) = ∑ g(k) (A.4)
k=0
(Note that here, too, a constant factor may be needed if we use a different definition of the
DFT; for example, with the definition used in the Mathematica® software package
[Wolfram96 p. 868] the required constant factor is 1 , and with the definition used in
N
[Bracewell86 p. 358] the constant factor is 1 .)
N
Eq. (A.3) means that in the continuous case the value of the spectrum at the DC
frequency u = 0 equals the area under the original signal-domain function. Its discrete
counterpart, Eq. (A.4), means that the value of the DFT spectrum at the discrete DC
frequency (i.e. at the array element representing the frequency 0) simply equals the sum of
all the samples of the signal-domain function in the input array.
 252 Appendix A: Impulses in the continuous and discrete worlds

Sampled g(x) DFT: Real part DFT: Imaginary part

8 8
2 N = 64, R = 8
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

4 Hz 4 Hz
-2
-4 -2 2 -4 -2 2

N = 64, R = 16 8 8
2
1 6 6

-8 -4 4 8
x 4 4

-1 2 2

2 Hz 2 Hz
-2
-2 -1 1 -2 -1 1

Figure A.4: The discrete counterpart of the function g(x) = sinc(2x) + cos(2πfx)
(with f = 12 ) and its DFT spectrum (after the corrections of Chapters
3 and 4). The array length being used is N = 64. Top row: with a
sampling range length of R = 8 (i.e. sampling interval Δx = 8/64 =
0.125); bottom row: with a sampling range length of R = 16 (i.e.
sampling interval Δx = 16/64 = 0.25). Note that the height of the
square pulse in both DFT spectra is 12 , just as in the continuous
spectrum, and it does not depend on R, while the height of the two
impulses does depend on R. Compare with Figs. A.2 and A.3.

Returning now to our first example above, the function g(x) = 1, we see by Eq. (A.4) that
N– 1
the value of the DC in the output array of the DFT6 is ∑ 1 = N (or with the DFT
k=0
N– 1
1 1 N = N ; or with the DFT definition
definition used by Mathematica®: ∑1=
N k=0 N
N– 1
used in [Bracewell86 p. 358]: 1 ∑ 1 = 1). And indeed, after applying the standard
N k=0
amplitude scaling correction of Chapter 4 (which corresponds, in the cited cases, to a
multiplication by Δx, N Δx or NΔx, respectively), the value of the DC becomes, as we have
seen above, NΔx = R (rather than 1, the height of the DC spike in the continuous case
according to the graphic convention of Sec. A.2).

Proceeding to our second example, g(x) = cos(2πfx), let us recall the following theorem
[Weaver89 p. 248]: If the DFT of the N-element sequence g(k) is the sequence G(m), then
the DFT of the sequence g(k) cos(2πkn/N) is the sequence 12 G(m+n) + 12 G(m–n). This is,

6 Remember that the DC is provided by the DFT in the first element (n = 0) of the output array; if
required, a reorganization of the output array can be done later for display purposes (see Chapter 3).
 Appendix A: Impulses in the continuous and discrete worlds 253

Sampled g(x) DFT: Real part DFT: Imaginary part

8 8
2 N = 64, R = 8
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

4 Hz 4 Hz
-2
-4 -2 2 -4 -2 2

N = 128, R = 16 8 8
2
1 6 6

-8 -4 4 8
x 4 4

-1 2 2

-2
-4 -2 2 4 Hz -4 -2 2 4 Hz

Figure A.5: Same as Fig. A.4, but in the bottom row the sampling range length
has been doubled to R = 16 without modifying the sampling
interval Δx, so that N has been doubled to N = 128.

in fact, the discrete counterpart of the modulation theorem [Bracewell86 p. 108]. By

applying this theorem to the constant function g(k) = 1 of our first example it follows that
the DFT of the function cos(2π kn/N), k = 0,...,N–1 consists of two impulses that are
located ±n elements away from the zero frequency, each having half the value of the DC
impulse of the constant function g(k) = 1, i.e., depending on the DFT definition being
used, N/2, N /2 or 12 (see, for example, [Brigham88 pp. 176–177]; [Weaver89 pp. 250–
251]). This means that in all cases, after applying the standard amplitude scaling
correction of Chapter 4, the height of the two impulses in the DFT spectrum is 12 R (rather
than 12 , the height of the two spikes in the spectrum of the continuous case). A similar
reasoning is also true for the sine function g(x) = sin(2πfx).7

Now, in order to understand what happens in the general impulsive case, we recall that
any function having an impulsive spectrum (i.e. any periodic, almost periodic or constant
function) can be represented as a Fourier series. This Fourier series is, in fact, a sum of
cosine or sine functions (that contribute to the spectrum the impulses of the corresponding
frequencies) and possibly a constant function (that contributes to the spectrum the
DC impulse). This explains, indeed, based on our previous results for the constant, cosine
and sine functions, why in the general impulsive case, too, the impulses in the DFT
spectrum (after the standard amplitude scaling correction) are R times higher than in the
corresponding CFT plot.

7 This time using the sine counterpart of the above mentioned theorem, according to which the DFT of
the sequence g(k) sin(2πkn/N) is the sequence 12 G(m+n) – 12 G(m–n) [Weaver89 p. 248].
 254 Appendix A: Impulses in the continuous and discrete worlds

Sampled g(x) DFT: Real part DFT: Imaginary part

8 8
2 N = 64, R = 8
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

4 Hz 4 Hz
-2
-4 -2 2 -4 -2 2

N = 128, R = 8 8 8
2
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

-2
-8 -4 4 8 Hz -8 -4 4 8 Hz

Figure A.6: Same as Fig. A.4, but in the bottom row N has been doubled to N = 128
without modifying the sampling range length R = 8, so that the sampling
interval has been halved to Δx = 8/128 = 0.0625.

In contrast, in functions whose spectra are non-impulsive, an increase in the range length
R causes an equivalent increase in the width of the DFT spectrum,8 and therefore it does
not affect the height of the DFT spectrum (see, for example, Fig. A.3).

The different behaviour of impulsive and non-impulsive spectra obtained by DFT is best
illustrated by a combination of both cases in a single function. Consider, for example, the
continuous-world function g(x) = sinc(2x) + cos(2πfx). Thanks to the addition theorem
the continuous-world spectrum of this function is the sum of the individual spectra of its
two terms, namely, a square pulse of height 12 within the frequency range –1 ≤ u ≤ 1 plus a
pair of impulses at the frequencies u = ±f which are represented by spikes of height 12 on
top of the square pulse, and whose heads reach therefore the total height 1: G(u) =
2 rect(u/2) + [2 δ(u–f) + 2 δ(u+f)]. When the function g(x) is sampled and fed to the DFT,
1 1 1

the result (after reorganization and application of the standard amplitude scaling
correction) shows, indeed, a square pulse of height 12 within the range –1 ≤ u ≤ 1 plus an
impulse pair to both sides of the origin (see Fig. A.4). But as shown in this figure, while
the height of the square pulse remains 12 independently of the range length R being used in
the signal domain, the total height of each of the two impulses is 12 R + 12 , which does
depend on the choice of R.

8 Not in terms of true frequencies in Hz, but in terms of the index n of the output array, i.e. in terms of
array elements (since by increasing R and hence Δx, for a given array length N, the frequency step in the
output array is reduced in accordance with Eq. (4.4)). See the influence of modifying R on the output of
the DFT in Sec. 7.4.
 Appendix A: Impulses in the continuous and discrete worlds 255

Sampled g(x) DFT: Real part DFT: Imaginary part

8 8
2 N = 64, R = 8
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

4 Hz 4 Hz
-2
-4 -2 2 -4 -2 2

N = 32, R = 8 8 8
2
1 6 6

-4 -2 2 4
x 4 4

-1 2 2

-2
-2 -1 1 2 Hz -2 -1 1 2 Hz

Figure A.7: Same as Fig. A.4, but in the bottom row the the sampling interval
has been doubled to Δx = 0.25 without modifying the sampling
range length R = 8, so that N has been halved to N = 32.

In fact, this behaviour is quite logical: as we can see, doubling R without changing N
results in doubling the width of the square pulse in the DFT spectrum (in terms of the
array index n, but not in terms of the real frequency in Hz); but when the spectrum
consists of impulses, whose width cannot be increased, it is clear that the height of the
impulses must be doubled.

Finally, Figs. A.5–A.7 show the behavoiur of the DFT of the same function as in Fig.
A.4, g(x) = sinc(2x) + cos(2πfx), under other possible parameter changes based on Eq.
(4.2), R = NΔx (see Table 7.1): While Fig. A.4 shows what happens when R is doubled
but N remains unchanged (so that Δx is doubled), Fig. A.5 shows what happens when R is
doubled but Δx remains unchanged (so that N is doubled). Similarly, Fig. A.6 shows what
happens when N is doubled but R remains unchanged (so that Δx is halved), and Fig. A.7
shows what happens when Δx is doubled but R remains unchanged (so that N is halved).
Note that in the two last cases the impulse height is not affected, since the value of R
remains unchanged.

Remark A.1: (Impulses in the spectral domain): The fact that the impulse amplitudes in
the DFT spectrum (after the application of the standard amplitude scaling correction of
Sec. 4.4) are R times higher than in the corresponding CFT means that in order to obtain
the expected results in the DFT spectrum, we need to divide the amplitude of each of the
impulses by R. However, there exists an important difference between this amplitude
correction and the standard amplitude scaling correction of Chapter 4: While the standard
 256 Appendix A: Impulses in the continuous and discrete worlds

Sampled g(x) DFT: Real part DFT: Imaginary part

1 1 1
N = 64, R = 8 F=8
0.5 0.5 0.5
1/8
-4 -2 2 4
x -4 -2 2 4 Hz -4 -2 2 4 Hz
-0.5 -0.5 -0.5

-1 -1 -1

1 1 1
N = 64, R = 16 F=4
0.5 0.5 0.5
1/4

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-0.5 -0.5 -0.5

-1 -1 -1

Figure A.8: Example with an impulse in the signal domain (see Sec. A.4):
The discrete counterpart of g(x) = δ(x) and its DFT spectrum (after
the corrections of Chapters 3 and 4). The array length being used is
N = 64. Top row: with a sampling range length of R = 8 (i.e.
sampling interval Δx = 8/64 = 0.125); bottom row: with a sampling
range length of R = 16 (i.e. sampling interval Δx = 16/64 = 0.25).
Note the amplitude difference between the two DFT spectra. In the
continuous case the corresponding spectrum amplitude is 1.

amplitude scaling correction should be applied in all cases to the entire DFT output, the
present impulse amplitude correction only applies to impulses in the DFT output (and
their leaked tails, if any; see Remark A.2 below). Thus, when the DFT output is hybrid
and contains both continuous and impulsive elements, like in Figs. A.4–A.7, the present
amplitude correction should be only applied to the impulses themselves (and their leaked
tails, if any), but not to the continuous spectral elements. For example, in the above
mentioned figures, the division by R should only be applied to the two elements of the
output array that contain impulses, and even there it should be done with care, since this
division only concerns the amplitude contribution of the impulse itself (in the present
example, 12 R), but not that of the underlying continuous element (in the present example, 12 ).
If this selective amplitude correction within the DFT output array is not feasible, no
major harm will be caused — one should simply remember that the resulting impulse
amplitudes in the DFT output will be R times higher than in the corresponding CFT.9 Q

9 In simple hybrid cases where the input signal is known to be the sum of a periodic function and an
aperiodic function, it would be easier to apply the DFT to each of the two functions separately and then
sum up the two resulting DFT spectra (thanks to the addition theorem). Here, the division by R can be
applied to the entire DFT spectrum of the periodic function, before adding up the two DFT spectra.
 Appendix A: Impulses in the continuous and discrete worlds 257

Sampled g(x) DFT: Real part DFT: Imaginary part

1 1 1
N = 64, R = 8 F=8
0.5 0.5 0.5

-4 -2 2 4
x -4 -2 2 4 Hz -4 -2 2 4 Hz
-0.5 -0.5 -0.5

-1 -1 -1

1 1 1
N = 64, R = 16 F=4
0.5 0.5 0.5

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-0.5 -0.5 -0.5

-1 -1 -1

Figure A.9: Example with impulses in the signal domain (see Sec. A.4): The discrete
counterpart of g(x) = 12 δ(x–a) + 12 δ(x+a) (with a = 12 ) and its DFT
spectrum (after the corrections of Chapters 3 and 4). The array length
being used is N = 64. Top row: with a sampling range length of R = 8
(i.e. sampling interval Δx = 8/64 = 0.125); bottom row: with a sampling
range length of R = 16 (i.e. sampling interval Δx = 16/64 = 0.25). Note
the amplitude difference between the two DFT spectra. In the continuous
case the corresponding spectrum amplitude is 1.

Remark A.2: The particular behaviour of impulses in the DFT spectrum can be also
explained from a different point of view. As we have seen in Chapter 6, unlike the
continuous Fourier transform, the DFT can only take into account a finite range of the
original, underlying function g(x), meaning that g(x) must be truncated before the
application of the DFT. This truncation operation can be viewed, still in the continuous
world, as a multiplication of the underlying function g(x) by a rectangular “truncation
window” w(x) whose length equals the sampling range R = NΔx, namely, w(x) = rect(x/R).
The spectral-domain counterpart of this truncation is that the DFT does not approximate
the original continuous-world Fourier transform G(u), but rather the convolution of G(u)
with the Fourier transform of the truncation window w(x), W(u) = R sinc(R u). As long as
G(u) is non-impulsive, the convolution with W(u) will only smooth out G(u) and add a
ripple effect (see Fig. 6.1, and a worked-out example in Sec. D.5). However, when G(u)
includes impulses, the convolution with W(u) will replace each impulse ciδ(u–ai) by an
additive replica of R sinc(R u) that is scaled by the impulse amplitude ci and centered about
the impulse location ai (Fig. 6.2). As shown in Sec. 6.5, if the continuous-world impulse
is located exactly on an element of the DFT output array, giving a sharp discrete-world
impulse in the DFT spectrum, this replica of R sinc(R u) will simply multiply the discrete
impulse amplitude ci by R, but have no effect on the other elements of the output array.
 258 Appendix A: Impulses in the continuous and discrete worlds

Sampled g(x) DFT: Real part DFT: Imaginary part

8 8 8
N = 64, R = 8 F=8
6 6 6
4 4 4
2 2 2

-4 -2 2 4
x -4 -2 2 4 Hz -4 -2 2 4 Hz
-2 -2 -2

8 8 8
N = 64, R = 16 F=4
6 6 6
4 4 4
2 2 2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-2 -2 -2

Figure A.10: The discrete counterpart of the signal g(x) = c[12 δ(x–a) + 12 δ(x+a)] +
cos(2πfx) (with a = 12 , c = 8 and f = 12 ) and its DFT spectrum (after the
corrections of Chapters 3 and 4). The array length (number of samples)
being used is N = 64. Top row: with a sampling range length of R = 8
(i.e. sampling interval Δx = 8/64 = 0.125); bottom row: with a sampling
range length of R = 16 (i.e. sampling interval Δx = 16/64 = 0.25). Note
that in the DFT spectrum the two impulses are R times higher than in the
continuous-world spectrum, whereas the amplitude of the cosine (the
Fourier transform of the impulse pair in the signal domain) is 1/F times
that of the continuous-world spectrum. Compare with Figs. A.2 and A.9.
This figure is based on the same concept as Fig. A.4 (the combination of
impulsive and non-impulsive signals in the same example), but in the
present case both the signal and the spectral domains contain such a
combination. See Sec. A.4.

And if the continuous-world impulse is located between elements of the DFT output array,
so that the resulting discrete impulse is smoothed out and smeared due to leakage, the
multiplication by R (and hence the amplitude dependence on the choice of the sampling
range length R) will affect both the smoothed out impulse and its leaked tails, but not the
other components of the DFT output array. This explains, indeed, why the impulse
amplitude in the DFT spectrum is multiplied by R with respect to the original continuous-
world spectrum G(u), and why this dependence of the DFT on the choice of R affects only
impulses (and their leaked tails, if any) but not other components of G(u), as indeed
illustrated in Fig. A.4. More details on this subject are provided in Sec. D.6. Q

The particular behaviour of impulses in the DFT spectrum is explained from yet another
point of view in Sec. D.9 of Appendix D, and several further examples are given there in
Sec. D.9.1.
 Appendix A: Impulses in the continuous and discrete worlds 259

Sampled g(x) DFT: Real part DFT: Imaginary part

1.5 1.5 1.5

N = 64, R = 8 F=8
1 1 1

0.5 0.5 0.5

-4 -2 2 4
x -4 -2 2 4 Hz -4 -2 2 4 Hz
-0.5 -0.5 -0.5

1.5
N = 64, R = 16 4 F=4 4
1 3 3
2 2
0.5
1 1

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -1
-0.5

Figure A.11: Example where both the signal domain and the spectral domain are purely
impulsive (see Sec. A.4): The discrete counterpart of the comb signal
g(x) = III(2x) and of its continuous-world spectrum G(u) = 12 III(u/2) as
obtained by DFT (after the corrections of Chapters 3, 4). The array length
being used is N = 64. Top: with a sampling range length of R = 8 (i.e.
sampling interval Δ x = 8/64 = 0.125); bottom: with a sampling range
length of R = 16 (i.e. sampling interval Δx = 16/64 = 0.25). Note that in
the DFT spectrum the impulses are R/F times higher than in the
continuous-world spectrum G(u) = 12 III(u/2), in which the impulse heights
are 1 [Bracewell86 p. 414].

A.4 Impulses in the signal domain

So far, we have discussed the behaviour of impulses in the spectral domain. What
happens when impulses appear in the signal domain?

Consider, for example, the continuous-world case of g(x) = δ(x), whose spectrum is the
constant function G(u) = 1. The discrete counterpart of g(x) is a sequence consisting of
zeroes except for the element at the origin, whose height is 1. As shown in Fig. A.8, if we
take the DFT of this sequence we obtain (after applying the standard amplitude scaling
correction) a constant sequence whose height is 1/F, where F = NΔu is the frequency
range length (the range length of the non-impulsive domain).

As a second example, consider the continuous-world case of g(x) = 12 δ(x–a) + 12 δ(x+a),

whose spectrum is G(u) = cos(2π au). The discrete counterpart of g(x) is a sequence
consisting of zeroes except for two 12 -valued elements that are located at the same distance
a to both sides of the origin (we assume here that a is an integer multiple of the sampling
interval Δx, so that the two impulses fall exactly on a sampling point). Now, as shown in
260 Appendix A: Impulses in the continuous and discrete worlds

Fig. A.9, if we take the DFT of this sequence we obtain (after applying the standard
amplitude scaling correction) a cosinusoidal signal whose amplitude is 1/F, which is again
dependent on the frequency range length F.

The different behaviour under DFT of impulsive and non-impulsive signals in the signal
and in the spectral domains can be best illustrated by a combination of all cases in a single
signal. Consider, for example, the continuous-world signal g(x) = c[12 δ(x–a) + 12 δ(x+a)] +
cos(2πfx), which consists of a pair of impulses of height 12 c located at the points x = ±a
plus a continuous cosine of frequency f and amplitude 1. Thanks to the addition theorem
the continuous-world spectrum of this signal is the sum of the individual spectra of its
two terms, namely, a cosine with amplitude c and frequency a (the spectrum of the impulse
pair) plus a pair of impulses at the frequencies u = ±f (the spectrum of cos(2π fx)),
which is represented by two spikes of height 12 on top of the cosine: G(u) = c cos(2πau) +
[12 δ(u–f) + 12 δ(u+f)]. The discrete counterpart of g(x) is a sequence consisting of samples
of the signal cos(2πfx) plus two c-valued spikes that are added to the cosine at a distance a
to both sides of the origin (again, we assume here that a is an integer multiple of the
sampling interval Δx, so that the two impulses fall exactly on a sampling point). When this
input signal is fed to the DFT, the result (after reorganization and application of the
standard amplitude scaling correction) shows indeed a cosine signal plus an impulse pair
to both sides of the origin, as expected (see Fig. A.10). But as shown in this figure, the
amplitude of the cosine in the DFT spectrum is multiplied by 1/F with respect to the
continuous-world spectrum (like in Fig. A.9), while the height of the impulse pair is
multiplied by R with respect to the continuous-world spectrum (just as in Fig. A.2). This
confirms, indeed, the behaviour under DFT of impulsive signals in both domains, as
explained above. And it also confirms that when two signals are summed together in the
signal domain each of them still behaves in the DFT of the sum exactly as it would when it
undergoes DFT alone; this is most spectacular if the individual behaviour under DFT of
each of the summed-up signals is different.

In conclusion, the behaviour under DFT of discrete impulses in the signal domain is
dual to their behaviour in the spectral domain: When impulses appear in the spectral
domain, their heights in the DFT spectrum are multiplied by R with respect to the
continuous-world spectrum (as shown in Figs. A.1 and A.2); but when impulses appear in
the signal domain, the height of their individual contribution to the DFT spectrum
is multiplied by 1/F with respect to the continuous-world spectrum (see Figs. A.8 and
A.9).

Remark A.3: (Impulses in the signal domain — signal-domain counterpart of Remark

A.1): The fact that the DFT spectrum of an impulsive signal (after the application of
the standard amplitude scaling correction of Sec. 4.4) is F times smaller than the
corresponding CFT means that in order to obtain the expected DFT spectrum, we need to
perform a multiplication by F. Once again, this treatment should only be applied to the
spectral components that belong to the impulsive input in question (note that this treatment
should be applied even if these input impulses suffer from signal-domain leakage). If such
Appendix A: Impulses in the continuous and discrete worlds 261

a selective correction is not feasible, the respective elements in the DFT output will simply
have a different amplitude than in the CFT. Q

Remark A.4: (Signal-domain counterpart of Remark A.2): The particular behaviour of

impulses in the DFT input can be also explained from a different point of view. As we
already know, the DFT can only take into account a finite range of frequencies. Therefore,
if the original, underlying continuous-world spectrum G(u) does not die out at the borders
of this range, G(u) should be truncated before the transition to the discrete world (in order
to avoid frequency foldover due to aliasing). This truncation operation can be viewed, still
in the continuous world, as a multiplication of the underlying spectrum G(u) by a
rectangular “truncation window” W(u) whose length equals the DFT frequency range
length F = N Δ u, namely, W(u) = rect(u/F). The signal-domain counterpart of this
truncation is that the input signal of the DFT should not be g(x) itself, but rather the
convolution of g(x) with the inverse Fourier transform of the truncation window W(u),
w(x) = F sinc(F x). As long as g(x) is non-impulsive, the convolution with w(x) will only
smooth out g(x) and add a ripple effect (see a worked-out example in Sec. D.5 of
Appendix D). However, when g(x) includes impulses, the convolution with w(x) will
replace each impulse ciδ(x–ai) by an additive replica of F sinc(F x) that is vertically scaled
by the impulse amplitude ci and centered about the impulse location ai. But unlike in the
non-impulsive case, this tacitly implies a vertical scaling by F. If the original continuous-
world impulse is located exactly on an element of the DFT input array, this convolution
will simply multiply the discrete impulse amplitude ci by F, but have no effect on the
other elements of the input array. And if the continuous-world impulse is located
between elements of the DFT input array, the convolution will scale up the impulse by F,
but also force it to leak out. In both cases, the multiplication by F (and hence the amplitude
dependence on the choice of the DFT frequency range length F) will only affect the
impulses (and their leaked tails, if any), but not the other components of the DFT input
array. It follows, therefore, that all impulses in the DFT input, either leaked out or not,
should be scaled up by F with respect to the original continuous-world function g(x) (see,
for example, the left-hand column in Fig. 6.11). This multiplication by F of the impulse
heights in the DFT input explains, indeed, thanks to the scaling theorem (see rule 1a in
Sec. 2.4.4) why the corresponding spectral components in the DFT output should be
multiplied by F in order to correctly represent the CFT G(u), as we have seen above. Q

The particular behaviour of impulses in the DFT input can be explained from yet
another point of view, which is analogous to the explanation provided in Sec. D.9 of
Appendix D.

We have already seen above what happens when both the signal and the spectral
domains are hybrid, i.e. contain a sum of impulsive and non-impulsive signals. What
happens, however, when both domains are purely impulsive? This happens in the case of
the continuous-world comb signal g(x) = III(x), for which we have the following
continuous-world Fourier pair [Bracewell86 pp. 214, 414]:

III(x) ↔ III(u)
262 Appendix A: Impulses in the continuous and discrete worlds

or more generally, by virtue of the dilation theorem [Bracewell86 p. 122]:

1
III(ax) ↔ a III(u/a)

Note, however, that the spike heights in III(ax) are 1a , and the spike heights in III(u/a) are
|a| (see, for example, [Gaskill78 pp. 60–61] or [Bracewell86 p. 414]). So what happens to
such continuous-world signals under DFT?

The discrete counterpart of the signal g(x) = III(ax) is a sequence consisting of zeroes
except for all the elements that are located at integer multiples of x/a, which contain the
value 1a (again, we assume here that these locations are multiples of the sampling
interval Δx, so that all impulses fall exactly on sampling points). Now, as shown in Fig.
A.11, if we take the DFT of this sequence we obtain (after applying the standard amplitude
scaling correction) a signal corresponding to the continuous-world spectrum 1a III(u/a),
whose amplitude is further multiplied by R/F. This means that in this case the DFT
spectrum is multiplied both by R and by 1/F with respect to the continuous-world
spectrum.

Figure A.12 shows this doubly-impulsive case, in which both the signal and the spectral
domains are purely impulsive, as the limit of a gradual sequence of impulsive signals with
non-impulsive spectra. (For the sake of comparison, a similar continuous-world
counterpart of this figure can be found in [Brigham88 p. 21].)

Interestingly, the “abnormality” in the behaviour of impulses in the discrete case can be
partially resolved if we modify our definition of a discrete impulse (Eq. (A.2)) by
multiplying the Kronecker delta function by a factor S, which stands for the length of the
other domain, i.e. R = NΔx or F = NΔu, depending on the case:10
⎧S n=0
δ(n) = ⎨ (A.5)
⎩0 otherwise

This way the factor S is already incorporated in the discrete impulse definition, and
therefore the Fourier pairs obtained by DFT (after applying the standard amplitude scaling
correction of Chapter 4) will be the same as in the continuous case; for example, the DFT
of the constant sequence consisting of 1-valued elements will be δ(n), and vice versa. Note
that this new definition of the discrete impulse is closer to the definition of the continuous-
world impulse (Eq. (A.1)), in which the height of the impulse at the point u = 0 is infinite,
so that it is equal, indeed, to the infinite range of definition (the continuous counterpart of
R or F). However, although this definition resolves our notational problem (in the sense
that the continuous and discrete Fourier pairs involving impulses are expressed now
exactly in the same way), it does not eliminate the actual dependence on the value of R or F
in DFTs involving impulses, as shown in our figures throughout this appendix.

10 This multiplication of the impulse amplitude by R or by F corresponds, indeed, to the effect of

convolving the impulse with the function R sinc(R u) or F sinc(F u), respectively, due to the truncation
in the other domain (see Remarks A.2 and A.4). As shown in Fig. 6.4(a), when the impulse falls exactly
on an array element, this convolution only affects the impulse height.
 Appendix A: Impulses in the continuous and discrete worlds 263

Sampled g(x) DFT: Real part DFT: Imaginary part

1.5
4 4
1 3 3
2 2
0.5
1 1

-7.5 -5 -2.5 2.5 5 7.5

x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -1
-0.5

1.5
4 4
1 3 3
2 2
0.5
1 1

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -1
-0.5

1.5
4 4
1 3 3
2 2
0.5
1 1

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -1
-0.5

1.5
4 4
1 3 3
2 2
0.5
1 1

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -1
-0.5

1.5
4 4
1 3 3
2 2
0.5
1 1

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -1
-0.5

Figure A.12: A sequence showing how the discrete counterpart of the comb signal
g(x) = III(2x) is obtained by adding more and more impulse pairs. The
DFT of each intermediate case is still non-impulsive (a sum of
cosines), but it gradually converges to an impulsive comb, as shown in
the last row. In each intermediate row the amplitude of the DFT
spectrum (after the corrections of Chapters 3, 4) is 1/F times that of
the continuous-world spectrum, due to the impulsive nature of the
signal domain. But in the last row, where both the signal domain and
the spectral domain are impulsive, the DFT spectrum is R/F = 4 times
that of its continuous-world counterpart G(u) = 12 III(u/2), in which the
impulse heights are 1 [Bracewell86 p. 414].
264 Appendix A: Impulses in the continuous and discrete worlds

Table A.1 below provides a summary of the main results that are presented in Secs. A.3
and A.4.

Finally, it should be noted that the amplitude corrections as discussed in this appendix
have been applied to the figures throughout the book (under the name “amplitude
correction due to Appendix A”), except when otherwise mentioned. Obviously, these
amplitude corrections have not been applied to the figures of Appendix A itself, since
these figures precisely intend to explain the amplitude discrepancies which occur between
the DFT results and the expected continuous-world results without the application of
these amplitude corrections (and hence, to show why these corrections are indeed
necessary for having a good match). Note, however, that the corrections due to Chapter 3
(reorganizations) and due to Chapter 4 (the standard amplitude scaling correction) have
been applied to all figures in this appendix, too.

Case Behaviour Examples

Both signal and The amplitude of the DFT spectrum (after the Fig. A.3
spectral domains are standard amplitude scaling correction of Sec. 4.4)
non-impulsive: matches (up to possible aliasing and leakage
distortions) that of the CFT.

Spectral domain is The impulse amplitudes in the DFT spectrum Figs. A.1, A.2
purely impulsive: (after the standard amplitude scaling correction
of Sec. 4.4) are R times higher than in the
corresponding CFT.

Signal domain is The amplitude of the DFT spectrum of an Figs. A.8, A.9
purely impulsive: impulsive signal (after the standard amplitude
scaling correction of Sec. 4.4) is 1/F of the
amplitude of the corresponding CFT.

Both signal and The amplitude of the DFT spectrum (after the Fig. A.11
spectral domains are standard amplitude scaling correction of Sec. 4.4)
is multiplied by R/F with respect to the
purely impulsive:
corresponding CFT.

Hybrid cases: The DFT of a sum of various components is the Figs. A.4–A.7,
sum of the DFTs of the individual components, A.10
by virtue of the addition theorem.

Table A.1: Synoptic overview of the cases discussed in Secs. A.3 and A.4.
Impulse amplitudes in the continuous world are to be understood
according to the graphic convention of Sec. A.2.
 Appendix B
Data extensions and their effects on the DFT results

B.1 Introduction
1D DFT algorithms receive as input a 1D array of N complex numbers (say, samples of
a given continuous complex- or real-valued function y = g(x)), and return as their output a
1D array of N complex numbers which contains the discrete frequency spectrum.
However, when the result obtained by DFT with N data values is not yet satisfactory, one
may wonder if a larger number of data values could improve the results. In fact, there exist
several possible ways to extend the input data for the DFT, each of which yields quite
different results. In this appendix we examine 6 of the most simple ways of extending the
input data, and explain how they influence the DFT results (including the possible effects
on the DFT artifacts). As we will see, some of these 6 methods are indeed useful and
advantageous, depending on one’s aims and also on the nature of the data, while others are
rather misleading and may introduce significant artifacts. Note that although the
discussion in this appendix is presented for the 1D case, its generalization to 2D or MD
DFT is straightforward.

B.2 Method 1: Extending the input data by adding new values beyond the
original range
Assume that we need to find the DFT of a function g(x) which extends to infinity, such
as g(x) = e–x2. It is clear that any number N of samples limits us to a finite sampling range;
but if we maintain the sampling interval Δx fixed and take more samples of g(x) beyond
the last original sample, a larger part of g(x) will be sampled and submitted to DFT.
Suppose that we take KN samples rather than N; how will it affect the DFT results?
It follows from Eq. (4.5) that increasing N by a factor of K while Δx remains unchanged
decreases the frequency interval Δu by the same factor. This means that the resolution of
the new DFT spectrum resulting from this extention is densified K times. In fact, all values
of the original DFT spectrum will still appear in the new transformed signal as before,1
but each original value will be followed by K–1 new values, giving more frequency details
(a higher spectral resolution) in the new DFT spectrum. However, due to Eq. (4.9), the
frequency range covered by the extended transformed signal does not change, and the
maximum frequency of the transformed signal remains the same as before. This is
schematically illustrated in Fig. B.1(b).

1 Up to a possible constant scaling factor, if the DFT definition being used contains a scaling factor such
as 1/N or 1/ N (see Chapter 2).

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 265
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_10,
© Springer-Verlag London 2013
266 Appendix B: Data extensions and their effects on the DFT results

This extention method may be useful when a higher resolution of the spectrum is
desired, for instance in order to better distinguish between some closely spaced spectral
peaks. It is interesting to note how the DFT artifacts in the original transformed signal are
affected by this extention: leakage (see Chapter 6) will be somewhat reduced (since the
truncation window has been enlarged and therefore its sinc-shaped transform has been
narrowed; see, for example, Fig. 2.1(d) in Chapter 2); but aliasing (see Chapter 5) will be
carried over to the new transformed signal and will not be modified (except for
manifesting itself at a higher resolution).

B.3 Method 2: Extending the input data by denser sampling within the original
range

A second way to increase the size of the DFT consists of decreasing the sampling
interval Δx, and taking more samples of the original function within the same range length
R as before. If the original signal is extended from N to KN values by such an increase of
the sampling rate, the transformed signal will also be extended K times, but this time by
increasing K times the frequency range length F (and the maximum frequency fmax
obtained by the DFT), while keeping the spectral resolution Δu unchanged (see Eqs. (7.4),
(4.10) and (7.3), respectively). In other words, the N values of the original DFT spectrum
will continue to represent the same frequencies as before, but new values will appear
beyond them, which correspond to higher frequencies that did not appear in the original
DFT spectrum (see Fig. B.1(c)).

This does not mean, however, that the N original values in the DFT spectrum will remain
exactly as before. First of all, they may be affected by a possible scaling factor, depending
on the DFT definition being used. But even more importantly, any aliasing artifacts which
occurred in the DFT spectrum before the extention, due to insufficient frequency range

Figure B.1: Schematic illustration of the 6 extension methods discussed in Appendix B.

Left column: the given input data sequence; right column: its reorganized
DFT (absolute value only). (a) The original N-points input data sequence
and its reorganized DFT. (b)–(g) The extended data sequences of KN
points according to methods 1–6, and their respective reorganized DFT.
Note that if one uses a DFT definition which contains a scaling factor (such
as 1/N or 1/ N; see Sec. 2.3 in Chapter 2), appropriate amplitude scaling
should be applied to the resulting DFT spectra. All cases are drawn with
respect to the true continuous-world signal-domain and frequency-domain
axes, which are shown at the bottom of the figure. (This figure could be
alternatively drawn with respect to the discrete axes k and n, i.e. with the
same element intervals in all cases, so that all rows would have the same
geometric length. But then the connection with the real-world interpretation
of the results would be lost.)
 Appendix B: Data extensions and their effects on the DFT results 267

Signal domain Spectral domain

0 N–1 0 N–1
(a) Original DFT

0 KN–1 0 KN–1
(b) Extension method 1

0 KN–1 0 KN–1
(c) Extension method 2

0 KN–1 0 KN–1
(d) Extension method 3

0 KN–1 0 KN–1
(e) Extension method 4

0 KN–1 0 KN–1
(f) Extension method 5

0 KN–1 0 KN–1
(g) Extension method 6

Signal domain axis: Frequency axis:

–Kx max –x max 0 x max Kx max –Kf max –fmax 0 fmax Kf max
268 Appendix B: Data extensions and their effects on the DFT results

(or, equivalently, insufficient sampling rate in the signal domain), will be moved in the
extended DFT spectrum to a new location: either to the correct frequency, if it is now
included in the extended frequency range (where it will no longer be considered as an
artifact), or, otherwise, to a new aliased location. This extention method is therefore useful
for eliminating aliasing artifacts, or at least moving them farther away in the spectrum
where they no longer interfere with the spectral values within the range of interest. But
even if aliasing artifacts can not completely be eliminated, the mere fact that they can be
displaced within the DFT spectrum by varying the sampling rate can be used as an
indicator helping to distinguish them from the real data in the spectrum (see Remark 5.3 in
Sec. 5.6).

Note that this extention method is in fact the inverse of the first method; and indeed, its
effect can be understood by exchanging roles between the given signal and the
transformed signal in the first method (since the given signal can be seen as the DFT of
the transformed signal, up to a scaling factor and a sign inversion [Bracewell86 p. 364]).
Increasing the resolution K times in one of the two domains (signal or frequency) causes a
K-fold range extension in the other. This is illustrated in Fig. B.1 (a)–(c).

B.4 Method 3: Extending the input data by adding zeroes after each value (zero
packing)

When the underlying continuous function is not known and additional samples cannot
be taken, neither beyond the original range as in the first method nor between existing
samples as in the second method, there still exist several ways to artificially extend the
given signal from N values to KN. One possible method is by adding K–1 zeroes after
each of the original values in the given signal. However, since no new information
concerning the underlying function has been added to the original signal by this extention
method, we cannot expect to gain any new spectral information (because the information
in one domain is completely equivalent to the information in the other domain, one being
obtainable from the other by DFT or IDFT). Not surprisingly, therefore, this extention
method has only a replicating effect, and the new transformed signal will consist of K
identical, consecutive copies of the original transformed signal (all of which possibly
scaled by a constant scaling factor). This is formally expressed by point 11 in Sec. 2.4.4
(called the similarity theorem in [Bracewell86 p. 370], or the stretch theorem in [Smith07
p. 157–158] and in [Cartwright90 pp. 204–205]). Note that along with the original DFT
spectrum, all the artifacts which are included in it will be replicated, as well.

We see, therefore, that in this method the frequency resolution in the transformed signal
is not changed, and although the frequency range is increased K times, no new spectral
information is really obtained (see Fig. B.1(d)).
Appendix B: Data extensions and their effects on the DFT results 269

B.5 Method 4: Extending the input data by replicating it

Another possible method to extend the given signal from N to KN values is by simply
replicating it K–1 more times. Obviously, in this case, too, no gain of information can be
expected in the spectral domain. In fact, this extending method is the inverse of the
previous method, and its effect on the transformed signal, within a possible constant
scaling, is merely the insertion of K–1 zeroes after each original value (this can be
formally obtained by an inverse application of the similarity theorem in [Bracewell86 p.
370] or the stretch theorem in [Smith07 pp. 157–158]; see point 11a in Sec. 2.4.4). The
frequency range of the transformed signal is not changed by this operation, nor does the
frequency represented by each of its original N elements, now being separated by zeroes
in the transformed signal. Although the DFT spectrum contains now more values per
frequency unit, no new spectral information has been obtained. The benefits of such an
extention of the DFT are therefore rather limited. This method is illustrated in Fig. B.1(e).
A similar case of data replication and its spectral consequences (albeit within the original
array length N) is presented in Remark 7.1 and Fig. 7.1.

B.6 Method 5: Extending the input data by replicating each of its elements

Still another way to extend the given signal from N to KN data values is obtained by
replicating each of its values K–1 more times. Here, too, no gain of information can be
expected in the spectral domain, since no new information has been added to the original
signal concerning the underlying function. As illustrated in Fig. B.1(f), the result of this
extention method in the spectral domain is a K-fold replication of the original transformed
signal, which is then modulated (i.e. multiplied) by a sinc-shaped envelope. This
multiplication reflects the fact that the value replications in the signal domain are in fact a
convolution with a K-values wide rect function; according to the convolution theorem,
this implies in the spectral domain a multiplication with the Fourier transform of this
function, which is the sinc function in question. Note that the zeroes of the sinc fall exactly
on the central frequency (the replicated DC) in all the new duplicates, reducing them to
zero.

If the underlying function g(x) is a piecewise constant function such as a train of pulses
or a histogram, then at first sight this method and method 2 (see Sec. B.3) seem to be
identical; however, they conceptually differ since in the present method all the new in-
between values are just identical copies of a neighboring value, while in method 2 they
more finely approximate the up and down transitions of g(x). The use of the present
method should normally be avoided, due to the significant distortions it introduces to the
resulting DFT spectrum.
270 Appendix B: Data extensions and their effects on the DFT results

B.7 Method 6: Extending the input data by adding zeroes beyond its original
range (zero padding)

This is a variant of extention method 1 (see Sec. B.2), which obviously coincides with it
if the underlying function g(x) has non-zero values only within a limited range which is
already included in the original sampled signal. If, however, g(x) has non-zero values
beyond the range covered by the original signal, too, then although the effect of extending
the signal K times by adding extra zeroes beyond the original N samples may look similar
to the effect of method 1, the new intermediate values obtained in the transformed signal
will not give any new informative values (see Fig. B.1(g)).

In fact, appending zeroes to the original signal rather than appending new samples from
the underlying function g(x) can be considered as a multiplication of the periodic
extension of the signal by a rectangular wave with a period of KN elements and pulse
width of N elements (the original signal length). This multiplication in the signal domain
implies in the transformed signal a convolution with the transform of the rectangular wave,
which is a sampled sinc function whose frequency step is K times smaller than in the
original signal. In other words, what we obtain in the transformed signal is an addition of
K–1 intermediate values after each value of the original transformed signal, but unlike in
method 1, where the intermediate values really give finer frequency values, here the
intermediate values simply interpolate the values of the original transformed signal with
the sinc function mentioned above [Brigham88 pp. 170–172; Smith07 pp. 133–134, 159–
160; Bracewell86 p. 369, the Packing Theorem].

In conclusion, although the smooth results obtained by the present method give an
illusion of an increased spectral resolution, in fact no additional resolution has been gained
(unless the underlying function g(x) is really zero in the interval where the new zeroes are
appended [Brigham88 p. 172]). Note, however, that padding with zeroes is not always
useless. For example, this method can be used advantageously as a convenient tool for
interpolating a given sampled function by means of DFT [Brigham88 pp. 198–199]: We
first compute the DFT of the given sampled function, and then add padding zeroes beyond
its original frequency range (i.e. after the index n = N/2 of the non-organized output
array). And indeed, the inverse DFT of this extended spectrum gives, back in the signal
domain, a more densely interpolated version of the original signal. Padding the original
signal with zeroes is also necessary when we wish to find the convolution of two finite-
duration signals using DFT, by virtue of the convolution theorem. When convolution is
computed with the aid of DFT, both of the given signals are treated as being periodic
(recall that DFT considers both its input and its output data as an N-element period of a
periodic function), and therefore the resulting convolution is cyclical. This means that the
outer ends of both input signals inevitably penetrate (additively) into the other side of the
resulting convolution. This generates a cyclical wrap-around effect which corrupts the
resulting convolution with respect to the expected non-cyclical convolution. This wrap-
around effect can be avoided by padding both of the signals to be convolved by a
Appendix B: Data extensions and their effects on the DFT results 271

Method
Operations on the DFT input: Results in the DFT output:**
number:

1 K-fold extension by adding new values K-fold increase in the spectral resolution;
beyond the original range. frequency range unchanged.

K-fold extension by performing a K times K-fold increase in the frequency range;

2 denser sampling within the original range. spectral resolution unchanged.

3* K-fold extension by adding K–1 zeroes K-fold replication of the transformed signal.
after each value.

K-fold extension by replicating the entire Insertion of K–1 zeroes after each original
4* given signal K–1 more times. value in the transformed signal.

K-fold extension by replicating each value K-fold replication of the transformed signal,
5* of the given signal K–1 more times. and modulation by a sinc function.

K-fold extension by padding zeroes beyond Insertion of K–1 sinc-interpolation values

6* the original range. after each original value in the transformed
signal.

* These cases are “dummy” extension methods which do not really add new spectral information in the
transformed signal.
** An appropriate scaling factor may be also applied in each case, depending on the DFT definition being
used.

Table B.1: Summary of the 6 extension methods.

sufficiently long span of zeroes before applying the DFT convolution. A detailed and
illustrated explanation of this subject can be found, for example, in [Bergland69 p. 48] or
in [Brigham88, Chapter 10] for the 1D case, and in [Brigham88 Sec. 11.4] for the 2D
case.

It may be noted in this context that in cases where the original data must be extended in
order to fit into an input array length N which is a power of 2 (for the sake of using FFT),
adding zeroes might not always be the best method. If the addition of trailing zeroes
causes jumps in the original data (including between its last and first elements, considering
the data as cyclical due to the periodic nature of DFT), a large spectral range of high
frequencies will be introduced into the transformed signal, which may cause in the
resulting DFT spectrum an oscillatory behaviour due to the leakage artifact (see Chapter
6). Aliasing artifacts (see Chapter 5) may be also caused, since the sampling frequency of
the input signal may not be sufficient to capture the high frequencies due to the jumps.
Padding with an appropriate non-zero value in order to prevent significant jumps in the
signal could be more judicious in such cases [Bracewell86 pp. 375–376].
272 Appendix B: Data extensions and their effects on the DFT results

B.8 Conclusions

A summary of the different extension methods which have been discussed above is
given in Table B.1 and in Fig. B.1 for quick reference. Understanding the advantages and
shortcomings of each method may help the DFT user to select the most appropriate
method for his needs, and to avoid pitfalls and false interpretations of the DFT results.

It should be noted that although only the one-dimensional case has been discussed here,
similar results can be also obtained for two or higher dimensional DFT, by applying the
same considerations to each dimension separately.
 Appendix C
The roles of p and q and their interconnections

C.1 Introduction

As we have seen throughout this book, and in particular in Chapters 5 and 6, cases in
which the original input function is periodic are particularly interesting, among other
reasons due to the impulsive nature of their spectra. In this appendix we study in more
detail some properties of such periodic cases in the discrete world, and provide their
multidimensional generalization.

C.2 The one dimensional case

Suppose we are given a continuous periodic function g(x) whose frequency is f, so that
its period is P = 1/f. As we already know, in order to apply DFT to the function g(x) we
first need to truncate it into a finite range of length R and sample it within this range. The
resulting N values serve as the input array of the DFT, and the DFT spectrum is obtained
in the N values of the output array. The number of periods of g(x) within the sampling
range of length R is, of course:1
q = R/P = R f (C.1)

The same number of periods will also appear within the range 0,...,N–1 of the discrete
(sampled) version of our function, in the input array of the DFT. This means that in the
discrete world the length of each period of our function is, in terms of array elements:

p = N/q (C.2)
This is, indeed, the discrete counterpart of the continuous-world relation P = 1/f.
Formulated in a more symmetric way, the discrete counterpart of Pf = 1 is, therefore:

pq = N (C.3)

We dwell here on the symmetric form because, as we will see later in Sec. C.3, this is the
form that lends itself best to multidimensional generalization.

Remark C.1: Note the dual role of q in the signal and spectral domains: In the signal
domain, q is the number of periods of g(x) within the sampling range length R (or within
the input array length N). But in terms of the DFT spectrum, q is the distance of the first

1 On the advantages and shortcomings of the inclusion of more than one period of g(x) within the
sampling range (the input array of the DFT), see Remark 7.1.

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 273
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_11,
© Springer-Verlag London 2013
274 Appendix C: The roles of p and q and their interconnections

harmonic impulse from the origin (i.e. the frequency of g(x) in terms of array elements),
as well as the distance between successive harmonic impulses (again, in terms of array
elements). Note that q needs not necessarily be integer; as explained in Chapter 6, when q
is non-integer the DFT spectrum suffers from a leakage artifact. Similarly, p, too, plays a
double role: In the signal domain p is the length of one period of g(x) in terms of array
elements; but in the DFT spectrum p is the number of q-element intervals, i.e. the number
of harmonic impulses for which there is room (without exceeding and folding over) in the
DFT spectrum of the periodic function g(x).2 Note that p, too, is not necessarily integer
(see below); in fact, the only restriction on p and q is that they must satisfy pq = N (Eq.
(C.3)), where N is obviously an integer number. See Example C.1 and Figs. C.2, C.3. Q

As we see, p and q are given in terms of continuous-world parameters by the relations:

p = F/f , q = R/P (C.4)

whose respective discrete-world counterparts are simply (see Eq. (C.3)):

p = N/q , q = N/p

Furthermore, the connection between the frequency f of our periodic function g(x), in
terms of Hz, and its discrete counterpart q, in terms of array elements, is given by:

q = f /Δu (C.5)

(this can be also obtained from Eqs. (C.1) and (7.3)). Similarly, we have also the
following connection between the period P in terms of continuous-world units and its
discrete-world counterpart p in terms of array elements:

p = P/Δx (C.6)

The nature of p (integer or non-integer) has particular significance if high harmonic

impulses of the periodic function g(x) fall beyond the borders of the DFT spectrum, i.e. if
our sampled version of g(x) suffers from aliasing. Using our present terminology we can
now reformulate Remark 5.1 as follows:

Remark C.2: There exists a particular case in which no “stray impulses” appear in the
DFT spectrum of a periodic function g(x) even if strong aliasing occurs. This happens if p
is integer, i.e. if the distance q between the origin and the first harmonic impulse of the
periodic function g(x) in the DFT spectrum (in terms of output array elements) satisfies
N = pq for an integer p (see Figs. C.2(a) and (c)). In this case, higher harmonic impulses
which exceed from one end of the DFT spectrum due to aliasing (foldover) will fall
exactly on top of true impulse locations (harmonic frequencies)3 in the opposite end of the
DFT spectrum; their heights will be simply summed up. Q

2 We count here all the harmonic impulses of g(x), even those whose amplitude is zero. See Footnote 3.
3 To be more precise, we speak here of impulse locations in terms of the algebraic support of the
spectrum of our periodic function, which consists of all the integer multiples of the basic frequency f of
g(x), even if the corresponding impulse amplitudes happen to be zero [Amidror09 p. 110].
Appendix C: The roles of p and q and their interconnections 275

g(x) 4 P

x
-8 -4 4 8

-1
(a)

0.5

G(u) 0.4 f

0.3

0.2

0.1

Hz
-4 -2 2 4

-0.1
(b)

Figure C.1: (a) The continuous-world periodic function g(x) = 0.8cos(2πfx) +

0.7cos(2π2fx) + 0.6cos(2π3fx) + 0.5cos(2π4fx) + 0.4cos(2π5fx) +
0.3cos(2π6fx) (see Example C.1). In the particular case plotted here
the frequency and period of the function are f = 0.5 and P = 1/f = 2.
(b) The continuous-world spectrum of g(x), G(u). The black dots in
(a) correspond to the sampled counterpart of g(x) with a sampling
step of Δx = 0.25. Note that this sampling rate is not sufficient to
capture the detail of the oscillations of g(x) (see below the x axis), and
it causes visible aliasing; the same sampled signal and its aliased
DFT spectrum are shown in row (a) of Fig. C.2.

Example C.1: The effect of p and q on the input and output arrays of the DFT of a
periodic function:
Consider the following periodic function with frequency f and period P = 1/f (Fig. C.1):
g(x) = 0.8cos(2π fx) + 0.7cos(2π 2fx) + 0.6cos(2π 3fx)
+ 0.5cos(2π 4fx) + 0.4cos(2π 5fx) + 0.3cos(2π 6fx)
276 Appendix C: The roles of p and q and their interconnections

This periodic function has been specially designed to best illustrate our discussion here:
It has 6 non-zero harmonics (f, 2f, 3f, 4f, 5f, 6f) with gradually decreasing amplitudes, so
that its impulses can be easily identified in the spectral domain. Fig. C.2 shows the
discrete counterpart of our continuous-world function g(x) as well as its resulting DFT
(after all the required adjustments as discussed in Chapters 3, 4 and Appendix A), for 4
slightly different values of f. In all cases the array length being used is N = 64, the
sampling range length is R = 16 and the sampling interval is Δx = R/N = 0.25; therefore in
the spectral domain the frequency range length is F = 1/Δx = 4 (by Eq. (7.4)) and the
frequency step is Δu = 1/R = 0.0625 Hz (by Eq. (7.3)). The only difference between the
four rows of the figure is in the frequency f of the function g(x), which has been chosen in
each case so as to illustrate one of the four possible combinations of p and q (given that
pq = N): (a) f = 0.5, so that by Eq. (C.1) q = Rf = 16·0.5 = 8 and p = N/q = 8 (both p and
q are integers); (b) f = 0.375, so that q = Rf = 16·0.375 = 6 and p = N/q = 10.666 (q is

Figure C.2: The effect of p and q of Eq. (C.3) on the input and output arrays of the
DFT of a periodic function. This is illustrated here using the periodic
function shown in Fig. C.1, g(x) = 0.8cos(2π fx) + 0.7cos(2π 2fx) +
0.6cos(2π3fx) + 0.5cos(2π4fx) + 0.4cos(2π5fx) + 0.3cos(2π6fx), whose
frequency and period are, respectively, f and P = 1/f. Each row shows the
sampled version of g(x) as well as its resulting DFT (after all the required
adjustments as discussed in Chapters 3, 4 and Appendix A), for a slightly
different value of f. The periodic function g(x) consists of 6 harmonics
(f, 2f, 3f, 4f, 5f, 6f) with gradually decreasing amplitudes, whose impulses
are clearly identifiable in the spectral domain. The first column shows the
input array (signal domain), and the second and third columns show the
output array (real and imaginary parts of the DFT spectrum). The array
length being used is N = 64, the sampling range length is R = 16 and the
sampling interval is Δx = R/N = 0.25; therefore in the spectral domain the
frequency range length is F = 1/ Δ x = 4 and the frequency step is
Δu = 1/R = 0.0625 Hz. The only difference between the four rows is in
the frequency f of the function g(x), which has been chosen in each case
so as to illustrate one of the four possible combinations of p and q (given
that pq = N): (a) f = 0.5, so that by Eq. (C.4) p = F/f = 4/0.5 = 8 and
q = N /p = 8 (both p and q are integers); (b) f = 0.375, so that
p = F/f = 4/0.375 = 10.666 and q = N/p = 6 (q is integer but not p);
(c) f = 0.666, so that p = F/f = 4/0.666 = 6 and q = N/p = 10.666 (p is
integer but not q); (d) f = 0.471, so that p = F/f = 4/0.471 = 8.5 and
q = N/p = 7.53 (neither p nor q are integers). Whenever q is integer (rows
(a) and (b)) the spectral impulses are sharp and no leakage occurs. In row
(b) all the 6 harmonic impulses are clearly visible; note that the 6-th
harmonic impulse has been folded over into the other end of the spectrum
due to aliasing. In row (a) the 3 highest harmonic impulses have been
folded over due to aliasing, but unlike in row (b) they fall exactly on top
of existing low-frequency impulses, and their amplitudes are summed up.
In rows (c) and (d) q is non-integer, so that leakage occurs in the DFT
spectrum. Just like in the first two rows, in row (d) all the 6 harmonic
impulses are clearly visible (the last two being folded-over due to
 Appendix C: The roles of p and q and their interconnections 277

Sampled g(x) DFT: Real part DFT: Imaginary part

p
q
3 0.6 0.6

2 0.4 0.4

1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(a) p, q integers

p 0.6 0.6
3
q
2 0.4 0.4

1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(b) q integer, p not

p
q
3 0.6 0.6

2 0.4 0.4

1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(c) p integer, q not

p
3 0.6 0.6
q
2 0.4 0.4

1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(d) p, q not integers

aliasing), while in (c) the folded-over impulses fall exactly on top of

existing low-frequency impulses, and their amplitudes are summed up.
Note that the high-frequency “grass” at the bottom of the original
continuous-world signal g(x) (see Fig. C.1(a)) is mimicked here in all
rows by a lower frequency grass, due to the insufficient sampling rate
and the resulting aliasing effect. The shape of the grass differs from row
to row because the aliasing is different in each case; the irregular shape
of the grass in rows (b) and (d) is due to the non-integer value of p. All
the discrete signals and DFT spectra are drawn here as connected
symbol plots (see Sec. 1.5.1).
 278 Appendix C: The roles of p and q and their interconnections

Sampled g(x) DFT: Real part DFT: Imaginary part

3 0.6 q 0.6

2 0.4 0.4
p
1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(a) p, q integers

3 0.6 q 0.6

2 0.4 0.4
p
1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(b) q integer, p not

3 0.6 0.6
q
2 0.4 0.4
p
1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(c) p integer, q not

3 0.6 0.6

0.4
q 0.4
2
p
1 0.2 0.2

-8 -4 4 8
x -2 -1 1 2 Hz -2 -1 1 2 Hz
-1 -0.2 -0.2

(d) p, q not integers

Figure C.3: Same as Fig. C.2, except that the function g(x) consists here of the first
harmonic alone: g(x) = 0.8cos(2πfx). The roles of p and q in terms of the
signal domain are clearly visible: p is the length of one period of g(x) in
terms of array elements, while q represents the number of periods of g(x)
within the input array. In rows (a) and (b) q is exactly integer, so that the
input array is perfectly cyclical (seamlessly wraparound); but this is not the
case in rows (c) and (d). In terms of the spectral domain, the impulses in
rows (a) and (b) are located exactly on an output array element because q is
integer; but in rows (c) and (d) the impulses are located between output
array elements, causing a significant leakage effect.
Appendix C: The roles of p and q and their interconnections 279

integer but not p); (c) f = 0.666, so that q = Rf = 16·0.666 = 10.666 and p = N/q = 6 (p is
integer but not q); (d) f = 0.471, so that q = Rf = 16·0.471 = 7.53 and p = N/q = 8.5
(neither p nor q are integers). As we can see in the figure, whenever q is integer (cases (a)
and (b)) the spectral impulses are sharp and no leakage occurs in the DFT spectrum. But
when q is non-integer (cases (c) and (d)) the spectral impulses fall between output-array
elements, so that leakage does occur. On the other hand, whenever p is integer (cases (a)
and (c)), there is room in the DFT spectrum exactly for an integer number of harmonic
impulses (this is also true in cases where some or even most of the harmonic impulses
have zero amplitude, as in the case of the function g(x) = cos(2π fx); see Fig. C.3). This is
particularly significant when aliasing occurs, since in this case all the folded-over impulses
due to aliasing fall exactly on top of existing lower-frequency impulse locations (see
Remark C.2). But when p is non-integer (cases (b) and (d)), there is room in the DFT
spectrum for a non-integer number of harmonic impulses. In this case, if aliasing occurs,
the impulses that are folded-over due to aliasing fall between existing low-frequency
impulse locations, and they become visible in the DFT spectrum as stray impulses.
Note that our function g(x) has been intentionally designed to have its highest harmonic
impulses beyond the range of the DFT spectrum (which means that g(x) suffers here from
aliasing, i.e. insufficient sampling rate in the signal domain): This allows us to show how
the exceeding harmonic impulses behave in each of the four cases. A best grasp of the
intended continuous-world spectrum (Fig. C.1(b)) is obtained in case (b), where all the
impulses are sharp (no leakage occurs), and the folded-over (aliased) impulses do not
overlap with existing impulses, so they are clearly visible. In case (d), too, the folded-over
impulses do not overlap with existing impulses, but they are more difficult to identify due
to the leakage artifact. In cases (a) and (c), however, the folded-over impulses overlap
(additively) with existing lower-frequency impulses in the other end of the DFT spectrum,
so that no intruding stray impulses are visible in spite of the folding over. In such cases
the only effect of aliasing on the DFT spectrum is in modifying the amplitudes of the
affected low-frequency impulses. Q

For the sake of comparison we also provide here Fig. C.3, which is identical to Fig. C.2
except that the periodic function it shows contains only the first harmonic of g(x), so that
no aliasing occurs.

We terminate with a consequence of Remark C.1, which is, in fact, a reformulation of

Remarks 6.3 and 6.4 (as we will see soon, this result is more interesting in the MD case):

Remark C.3: A periodic function g(x) is perfectly cyclical (seamlessly wraparound)

within the input array, meaning that it has no discontinuities on the array boundaries when
the input array is repeated periodically to pave the entire 1D space, if and only if its
impulses in the DFT spectrum are located exactly on output array elements (so there is no
leakage), i.e. if and only if q is integer. Q

Table C.1 provides a synoptic review of the roles of p and q in the 1D case, both in the
signal and in the spectral domains.
280 Appendix C: The roles of p and q and their interconnections

Effect on the input Effect on the output Examples

array (signal domain) array (DFT spectrum)

q is integer Integer number of periods of Impulse intervals consist of Figs.

g(x) within the input array an integer number of array 6.3(a),(e),(f),
(sampling range); input elements; impulses are
C.2(a),(b) and
array is perfectly cyclical located exactly on array
(seamlessly wraparound). elements; no leakage.
C.3(a),(b)

q is non-integer Non-integer number of Impulse intervals consist of Figs.

periods of g(x) within the a non-integer number of 6.3(b),(c),(d),
input array (sampling array elements; impulses are C.2(c),(d) and
range); input array is not located between array
C.3(c),(d)
perfectly cyclical. elements; leakage.

p is integer Period length of g(x) is Integer number of harmonic Figs. C.2(a),(c)

exactly an integer number of impulse locations within and C.3(a),(c)
array elements. the output array (DFT
spectrum). In particular, if
aliasing occurs, folded-over
impulses fall exactly on top
of lower-harmonic impulse
locations, and their
amplitudes are summed up.

p is non-integer Period length of g(x) is not Non-integer number of Figs. C.2(b),(d)

an integer number of array harmonic impulse locations and C.3(b),(d)
elements. within the output array
(DFT spectrum). In
particular, if aliasing occurs,
folded-over impulses fall
between lower-harmonic
impulse locations, giving
visible stray impulses.

Table C.1: The effects of p and q of Eq. (C.3) on the input and output arrays of the
DFT of a periodic function g(x). Note that the effects of p and q are
orthogonal (i.e. independent of each other), and all the 4 combinations of
p and q may exist, as shown in Fig. C.2.

C.3 Generalization of p and q to the multidimensional case

As it could be expected, the situation in the 2D case is more complex and more
interesting than in the 1D case. This is already true in the continuous world: First of all,
the period and the frequency of a 2D periodic function g(x,y) are no longer scalars but
Appendix C: The roles of p and q and their interconnections 281

rather vectors in the x,y space and in the spectral u,v space, respectively. Second, a 2D
periodic function g(x,y) may be either 1-fold periodic or 2-fold periodic, so that it may
have either one or two period-vectors in the x,y space, and hence either one or two
frequency-vectors in the u,v space. Furthermore, it turns out that the relationship between
the period-vectors and the frequency-vectors in the 2D case is no longer as simple as the
relationship between the period P and the frequency f in the 1D case. For example, if
g(x,y) is 2-fold periodic, then its two period-vectors P 1 and P 2 do not have the same
orientation as the respective frequency-vectors f1 and f2 in the spectral domain; rather, P1 is
orthogonal to f2, and P2 is orthogonal to f1. It can be expected, therefore, that the situation
in the discrete world, too, will not be as simple as in the 1D case. So how can we
generalize the scalars p and q and the relationship pq = N between them to the
multidimensional case?

As we already know, pq = N is the discrete counterpart of the continuous-world

relationship Pf = 1 between the period P and the frequency f of a 1D periodic function
g(x). Let us start with the multidimensional generalization of the continuous-world
relationship, Pf = 1. Then, based on the experience we gain from this case, we will proceed
to the analogous generalization in the discrete world.

C.3.1 Multidimensional generalization in the continuous world

The multidimensional generalization of the continuous-world relationship Pf = 1 can be

found, for example, in [Rosenfeld82 p. 75]; a more detailed review is also provided in
Secs. A4–A.5 of Appendix A in [Amidror09]. Let us briefly reformulate here the main
continuous-world results.

Definition C.1: If P1,...,Pn ∈RM are n fundamental period-vectors of the n-fold periodic
function g(x1,...,xM) in the MD case (where obviously 1 ≤ n ≤ M), then the corresponding
frequency-vectors f1,...,fn ∈R M in the spectral domain are defined by the following
relationship:4
⎧1 if i = j
Pi · fj = ⎨ (C.7)
⎩0 if i ≠ j
where i, j = 1,...,n. Q

Note that because Eq. (C.7) is symmetric, it can also be used the other way around to
determine the period-vectors P 1,...,P n given the frequency-vectors f1,...,fn. The only
requirement in both cases is that the vectors P 1,...,Pn as well as the vectors f1,...,fn be
independent; the existence of such sets of n independent vectors is indeed guaranteed
because g(x1,...,xM) is n-fold periodic.

4 As illustrated in Figs. C.5 and C.6 (and explained in greater detail in Sec. A.3.4 of Appendix A in
[Amidror09]), the period-vectors P1,...,Pn as well as the frequency-vectors f1,...,fn are not unique. For
example, every 2D 2-fold periodic function has infinitely many sets of period-vectors (and of frequency-
vectors), all of which are equivalent and define the same periodicity. But each set of period-vectors
P 1,...,P n defines by Eq. (C.7) its own unique set of frequency-vectors f1,...,fn. This set of frequency
vectors is said to be reciprocal to its corresponding set of period-vectors.
282 Appendix C: The roles of p and q and their interconnections

y v
y'
u1 = f1 cosθ1
v1 = f1 sinθ1
T1 2 f1
x'
f1 •
v1 •
θ1
θ1
x • u1 u
–
•
f1
•
–2 f1
f1 = 1/T1

(a) (b)

Figure C.4: Schematic plot of a 1-fold periodic function g(x,y) in the signal
domain (a), and of the corresponding spectrum (impulse-comb)
G(u,v) in the spectral domain (b).

The geometric interpretation of this definition is as follows:

(a) The second condition in Eq. (C.7) means that the vector fj is perpendicular to all the
vectors Pi with i ≠ j; this determines the direction of the line through the origin of RM
on which f j is situated. (For example, in the 3D case f 1 is situated on the line
emanating from the origin of R3 perpendicularly to the plane spanned by P2 and P3).
(b) The first condition in Eq. (C.7) determines the precise length and direction of the
vector fj on the line defined in (a). Pj·fj = 1 means that the length of fj on this line is
reciprocal to the length of the projection of Pj on the same line: Since by [Vygodski73
p. 142] P · f = | f | |proj(P)f | (where proj(P)f denotes the projection of P on f), and
since we have here P· f = 1, it follows that | f | = 1/|proj(P)f |. Furthermore, since we have
Pj · fj = 1 > 0 the direction of the vector fj on the same line is determined such that the
angle between fj and Pj is sharp [ibid.].

Hence, the two conditions of Eq. (C.7) fully determine each of the n vectors fj.

Note that when M = 1 and n = 1 Eq. (C.7) simply gives the 1D relationship Pf = 1.
Another simple particular case of Eq. (C.7) is that of a 2D 1-fold periodic function g(x,y),
such as a line grating on the x,y plane that is possibly rotated by angle θ (see Fig. C.4); in
this case M = 2 and n = 1, so that the function g(x,y) has a single period-vector P = (x,y)
and a single frequency-vector f = (u,v), both having the same direction θ and reciprocal
lengths along this direction. This case is discussed in more detail in the beginning of Sec.
6.6.1. Yet another particular case of Eq. (C.7) is that of a 2D 2-fold periodic function
g(x,y), in which M = 2 and n = 2; we will return to this case in Example C.2 below.
Appendix C: The roles of p and q and their interconnections 283

Based on Eq. (C.7) the reciprocity between the period-vectors Pi and the frequency-
vectors fi can be also expressed in matrix notation:5
P1 P1·f1 P1·fn 1 0
(f1,...,fn) = =
Pn Pn·f1 Pn·fn 0 1

Therefore, if the matrix (f1,...,fn) is invertible, i.e., non-singular (which is true iff the vectors
f1,...,fn are linearly independent over R [Birkhoff77 pp. 237–238], a condition that is
guaranteed in our case because g(x1,...,xM) is n-fold periodic), then:
P1
–1
= (f1,...,fn)
Pn

and by writing both sides as columns we obtain:

P1 f1 –T f1 P1 –T
= or: = (C.8)
Pn fn fn Pn

where “–T” means the transpose of the inverse matrix. If we denote these two matrices in
short by P and F, then Eq. (C.8) becomes P = F –T or F = P –T.

Example C.2: Let us consider the 2D 2-fold periodic case (M = 2, n = 2); see Fig. C.5. If
the two period-vectors are P1 = (x1,y1) and P2 = (x2,y2), then:
f1 –T x1 y1 –T y2 -x2
= P1 = = 1 (C.9)
f2 P2 x2 y2 x1y2 – y1x2 -y1 x1
where A = x1y2 – y1x2 is the area of the parallelogram A defined by the two vectors P1
and P 2: A = P 1× P 2 = x1y2 – y1x2. Note that A is, in fact, the determinant of the matrix
P x1 y1 6
P= 1 = .
P2 x2 y2
The two frequency-vectors f1 = (u1,v1) and f2 = (u2,v2) are given, therefore, by:
f1 = 1 (y2, –x2)
A
(C.10)
f2 = 1 (–y1, x1)
A

Consequently, we obtain here the following interesting properties [Amidror09 p. 388]:

5 Note that each vector in the following expressions represents an M-tuple of coordinates, and therefore
each entity in parentheses is, in fact, a matrix. If n < M this matrix is not square, but we can always
choose a new basis in RM along the n axes of periodicity. When expressed using this basis each of the
vectors will have n non-zero coordinates, so if we omit the trailing M – n zeros each vector becomes an
n-tuple and all the matrices are therefore square with n×n elements. For example, in the 2D 1-fold
periodic case (see Fig. C.4) the original 2D vectors become 1D along the axis of the 1-fold periodicity.
6 It is interesting to note that in linear algebra determinants play the role of area (or volume) functions

[Lay03 pp. 204–209]. For example, the determinant A of the matrix P gives the area of the parallelogram
that is determined by the rows of the matrix, i.e. by the vectors P1 = (x1,y1) and P2 = (x2,y2).
284 Appendix C: The roles of p and q and their interconnections

y v
y' x''
2f 2
• u1 = f1 cosθ1
T1 • • v1 = f1 sinθ1
• •
y'' f2
v2 •B u2 = f2 cosθ2
x' 2 f1
T2 •
y2 f1 • v2 = f2 sinθ2
P2 • v1 •
A x1
x2 x – f1 • u2 u1 u
y1
P1 –2 f1 • •
• • f1 = 1/T1
–f 2
• f2 = 1/T2
• •
• –2f 2 •
•

(a) (b)

Figure C.5: Schematic plot of a 2-fold periodic (skew-periodic) function g(x,y) in the
signal domain (a), and its skewed impulse-nailbed G(u,v) in the spectral
domain (b). The gray parallelogram A in the signal domain represents a
one-period element (tile) of g(x,y). P 1 and P 2 are segments of this
parallelogram that coincide with the period-vectors P1 and P2.

y v
y' x''
2f 2
•
T1 • •
• f2 •
y'' x' •
T2 f'2 = f2 – f1 2 f1
• •
P'2 A' • B' f'1 = f 1
•
P'1 x – f1 • u
–2 f1 • •
• •
–f 2
•
• •
• –2f 2 •
•

(a) (b)

Figure C.6: A different set of fundamental period-vectors P'1, P'2 and fundamental
frequency vectors f'1, f'2 for the same function g(x,y) as in Fig. C.5. The
parallelograms A' and B' they define have the same respective areas as
the parallelograms A and B in Fig. C.5, but they are almost square.
Appendix C: The roles of p and q and their interconnections 285

(1) Orthogonality: f1 ⊥ P2 and f2 ⊥ P1.

(2) Vector lengths: |P1| = x12 + y12
|P2| = x22 + y22
1
and from Eq. (C.10): |f1| = x22 + y22 = 1 |P2|
A A
1
|f2| =
A
x1 + y1 = A1 |P1|
2 2

This means that the length ratio between the two pairs of fundamental vectors is
preserved reciprocally:
f 1 P2
=
f 2 P1
For example, if in the signal domain the length of P1 is twice the length of P2, then in
the spectrum the length of f1 is half the length of f2.
(3) Another interesting result concerns the areas of the period-parallelogram A that is
defined by the vectors P1 and P2 in the signal domain and the frequency-parallelogram
B that is defined by the vectors f1 and f2 in the spectrum (see Fig. C.5):
A = P1×P2 = x1y2 – y1x2
1
B = f 1 × f2 = u 1v 2 – v 1u 2 = 2 (x1y2 – y1x2) = A1
A

This means that the areas of the parallelograms A and B are, indeed, reciprocal:7
AB = 1
Due to these reciprocity relations which prevail between the period-vectors and the
frequency-vectors these two sets of vectors are said to be reciprocal.
Finally, it is easy to verify that the mixed scalar products of the period and frequency
vectors give, indeed:
y2 y1
P 1· f 1 = x 1 – y1 x2 = A1 A = 1 P1· f2 = –x1 + y1 x1 = 0
A A A A
y2 x2 y1 x1
P 2· f 1 = x 2 – y2 =0 P2· f2 = –x2 + y2 = A1 A = 1
A A A A

in accordance with Eq. (C.7). Q

C.3.2 Multidimensional generalization in the discrete world

Having generalized in the continuous world the 1D relationship Pf = 1 to the MD case,

we are ready to proceed now to the discrete world. The approach we take in the discrete
world is very similar, except that this time we start from the 1D relationship pq = N, which

7 The generalization of this result to the MD case is based on determinants: As suggested by the previous
footnote, the volume A defined by the vectors P1,...,Pn is the determinant of the matrix P of Eq. (C.8)
(see, for example, under “parallelepiped” in [Weisstein99 p. 1314]). Similarly, the volume B defined by
–
the vectors f1,...,fn in the spectral domain is the determinant of the matrix F = P T, which equals, indeed,
1/A. We therefore have in the general MD case, too, AB = 1. This result is already found (in a slightly
different form due to the use of different Fourier conventions) in Appendix B of [Petersen62].
286 Appendix C: The roles of p and q and their interconnections

is the particular case that we expect to get when M = 1 and n = 1. Let us denote by
p1,...,pn the counterparts in the MD input array of the continuous-world period vectors
P1,...,Pn, and by q1,...,qn the counterparts in the MD output array (DFT spectrum) of the
continuous-world frequency vectors f1,...,fn (see Fig. C.7). We will henceforth call these
discrete-world vectors p-vectors and q-vectors, respectively. Note that the coordinates of
all the p- and q-vectors are specified in terms of array elements. We therefore obtain:

Definition C.2: If p1,...,pn ∈RM are n fundamental p-vectors of the discrete (sampled)
version of the n-fold periodic function g(x1,...,xM ) in the MD case (where obviously
1 ≤ n ≤ M), then the corresponding q-vectors q1,...,qn ∈R M in the spectral domain are
defined by the following relationship:8
⎧N if i = j
p i · qj = ⎨ (C.11)
⎩0 if i ≠ j
where i, j = 1,...,n. Q

Because Eq. (C.11) is symmetric, it can also be used the other way around to determine
the p-vectors p1,...,pn given the q-vectors q1,...,qn. The only requirement in both cases is,
again, that the vectors p1,...,pn as well as the vectors q1,...,qn be independent; the existence
of such sets of n independent vectors is guaranteed since our function is n-fold periodic.

This definition is coherent with its continuous-world counterpart, Definition C.1, since
just like in the 1D case (see Eqs. (C.5) and (C.6)), in the MD case, too, the connection
between the continuous-world vectors and their discrete-world counterparts is given by:

pi = Pi / Δx (C.12)

q j = fj / Δ u (C.13)

so that by using Eq. (4.4) we obtain, indeed, Eq. (C.11):

p i · q j = (Pi · fj) / ΔxΔu = N Pi · fj

Furtheremore, we also have here the following MD generalization of Eq. (C.4):

p i = F Pi = N P i (C.14)
R
N
q i = R fi = f i (C.15)
F
Note that in definition C.2 we say that p1,...,pn and q1,...,qn are vectors in RM rather than
in ZM, although we are dealing here with the discrete world. The reason is that, just as in
the 1D case, both the p-vectors and the q-vectors need not necessarily have integer
coordinates; the only restriction which applies is that they must satisfy Eq. (C.11) where
obviously N is integer. We will return to this point in Remark C.4.

8 Just as in the continuous case (Definition C.1), the p-vectors p1,...,pn as well as the q-vectors q1,...,qn
are not unique; but each set of p-vectors p1,...,pn defines by Eq. (C.11) its own unique set of q-vectors
q1,...,qn. This set of q-vectors is said to be reciprocal to its corresponding set of p-vectors.
Appendix C: The roles of p and q and their interconnections 287

g(x,y) G(u,v)
y v
8. . . . . 1.. . . . .
. . . . . . . . .
4. . . . . 0.5. . . . .
P2 A . . . . . . . . .
-8. -4. 4. 8.
x .
-1.
. .
-0.5
. f.B.
2
.
0.5
. .
1.
u
P1
. . . . .f . 1
. . .
-4. . . . . -0.5. . . . .
. . . . . . . . .
-8. . . . . -1.. . . . .
P1 = (4.,0.) f 1 = (1/4., 0.)
(a)
P2 = (0.,4.) f 2 = (0., 1/4.)
A = 16. A B = 1. B = 1/16.

y v
0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
8. 1.
31 31
6. 28 28

4. 24 0.5 24

2. p C 20 20
2
q2 D
0 16 0 16
p1 q1
-2. 12 12

-4. 8 -0.5 8

-6. 4 4

-8. 0 -1. 0
-8. -6. -4. -2. 0 2. 4. 6. 8. x -1. -0.5 0 0.5 1. u
p1 = (8,0) q1 = (4,0)
(b)
p2 = (0,8) N = 32 q2 = (0,4)
2
C = 64 C D = 64·16 = N D = 16

Figure C.7: (a) A 2D 2-fold periodic function g(x,y) and its spectrum G(u,v) in the
continuous world. (b) Their discrete-world counterparts using 2D DFT.
Compare the period vectors P1, P2 and the frequency vectors f1, f2 in (a)
with their respective discrete-world counterparts in (b), p1, p2 and q1, q2,
whose coordinates are specified in terms of array elements. We use in
(b) N = 32 in order that the individual array elements be sufficiently large
and their locations easily identified. The 2D input array corresponds to
the continuous-world range –8...8 along both axes, and the 2D DFT
spectrum covers the frequency range –1...1 Hz along both axes, so that
R = 16 and F = 2.
288 Appendix C: The roles of p and q and their interconnections

Note that when M = 1 and n = 1 Eq. (C.11) simply gives, as expected, the 1D
relationship of Eq. (C.3): pq = N. Another simple particular case of Eq. (C.11) is that of a
2D 1-fold periodic function g(x,y), such as a 2D cosine wave or a line grating on the x,y
plane that is possibly rotated by angle θ; in this case M = 2 and n = 1, so that the function
g(x,y) has in the discrete world a single p-vector p = (x,y) and a single q-vector q = (u,v),
both having the same direction θ and reciprocal lengths along this direction (see also Sec.
6.6.1). Yet another particular case of Eq. (C.11) is that of a 2D 2-fold periodic function
g(x,y), in which M = 2 and n = 2; we will return to this case in Example C.3 below.

Based on Eq. (C.11) the reciprocity between the p-vectors pi and the q-vectors qi can be
also expressed in matrix notation:9
p1 p 1·q 1 p 1·q n N 0
(q1,...,qn) = =
pn p n·q 1 p n·q n 0 N
and therefore the discrete counterpart of Eq. (C.8) is:10
p1 q1 –T q1 p1 –T

= N or: = N (C.16)
pn qn qn pn
where “–T” means the transpose of the inverse matrix. If we denote these two matrices in
short by S and Q, then Eq. (C.16) becomes S = N Q –T or Q = N S –T.

Example C.3: The discrete-world counterpart of Example C.2 above:

Let us consider the 2D 2-fold periodic case (M = 2, n = 2) in the discrete world. If the two
p-vectors are given in terms of array elements by p1 = (a1,b1) and p2 = (a2,b2), then:11
q1 p1 –T –T
= N a1 b1 N ba21 -a
b12
q2 = N p2
= (C.17)
a2 b2 a 1b 2 – b 1a 2 -ba21 ba21
where C = a1b2 – b1a2 is the area (in terms of array elements) of the parallelogram C
defined by the two vectors p1 and p2: C = p 1× p 2 = a1b2 – b1a2. Note that C is, in fact, the
p1 a b
determinant of the matrix S = p = 1 1 (see Footnote 6 in Sec. C.3.1).
2 a2 b2
Therefore, the two q-vectors q1 = (c1,d1) and q2 = (c2,d2) are given in terms of array
elements by:
q1 = N (b2, –a2)
C
q = N (–b , a ) (C.18)
2 1 1
C

Consequently, we obtain here the following properties (compare with Example C.2):

9 Note, again, that each vector in the following expressions represents an M-tuple of coordinates, and
therefore each entity in parentheses is, in fact, a matrix. And once again, if n < M we can choose a new
basis in RM along the n axes of periodicity, in which the vectors become n-tuples and the matrices are
square with n×n elements. These n axes are also used for sampling the given function.
10 This is a generalization of the 1D case, in which pq = N (Eq. (C.3)) gives p = N/q or q = N/p.
11 Although a ,b ,c ,d are given in terms of array elements, they are not necessarily integer (Remark C.4).
i i i i
Appendix C: The roles of p and q and their interconnections 289

(1) Orthogonality: q1 ⊥ p2 and q2 ⊥ p1.

(2) Vector lengths (in terms of array elements):
|p1| = a12 + b12
|p2| = a22 + b22
N
and from Eq. (C.18): |q1| = a 22 + b 22 = N |p2|
C C
N
|q2| = a 12 + b 12 = N |p1|
C C

This means that the length ratio between the two pairs of p- and q-vectors is preserved
reciprocally:
q1 p2
=
q2 p1
For example, if in the input array the length of p1 is twice the length of p2, then in the
DFT spectrum the length of q1 is half the length of q2.
(3) The area of the period-parallelogram C that is defined by the vectors p1 and p2 in the
input array and the area of the frequency-parallelogram D that is defined by the vectors
q1 and q2 in the DFT spectrum are given, in terms of array elements, by:
C = p 1× p 2 = a 1b 2 – b 1a 2
D = q1× q2 = c1d2 – d1c2 = ( N )2(a1b2 – b1a2) = N2 (C.19)
C C

This means that the areas of the parallelograms C and D in terms of array elements are
reciprocal, in the sense that:12
CD = N2 (C.20)
This is, indeed, a generalization of the discrete 1D relationship pq = N.
Due to these reciprocity relations which prevail between the p-vectors and the q-vectors
these two sets of vectors are said to be reciprocal.
Finally, it is easy to verify that the mixed scalar products of the p- and q-vectors give,
indeed:
p 1· q1 = N (a1b2 – b1a2) = N C = N p 1· q2 = N (–a1b1 + b1a1) = 0
C C C

p 2· q1 = CN (a2b2 – b2a2) = 0 p 2· q2 = CN (–a2b1 + b2a1) = CN C = N

in accordance with Eq. (C.11).
The connections between C and D and their continuous-world counterparts A and B can
be obtained using Eqs. (C.14) and (C.15). Thus, for the case of n = 2 we get:
C = p1× p2 = F 2 P 1× P2 = F 2 A = F 2 / B (C.21)
D = q 1× q 2 = R f 1× f2 = R B = R / A
2 2 2
(C.22)

12The generalization of this discrete-world result to the MD case is CD = N n (where 1 ≤ n ≤ M is the

number of p-vectors p1,...,pn), as opposed to AB = 1 in the continuous-world case of Example C.2.
290 Appendix C: The roles of p and q and their interconnections

Note also that the surface areas of the input array and of the output array are N2 array
elements, which is obviously an integer number. However, the area of each period-
parallelogram in the input array, as well as the number of such period-parallelograms that
pave the entire input array, may be either integer or not, depending on the case. The same
is also true for the frequency-parallelograms in the output array. For example, in Fig.
C.8(b) we have N = 32, C = 9·4 = 36 and D = 3.555·8 = 28.444; and indeed, CD =
36·28.444 = 322. Note that if the area of a single period-parallelogram in the input array is
C array elements, then the number of period-parallelograms C within the N×N-element
input array is precisely N 2 / C = D. And similarly, if the area of a single frequency-
parallelogram in the output array is D array elements, then the number of frequency-
parallelograms D within the N×N-element output array is precisely N2 / D = C. This is,
indeed, a 2D generalization of the roles of p and q in the 1D case (see Remark C.1). Q
It should be mentioned that the generalization of Eqs. (C.21) and (C.22) to the n-fold
periodic MD case is straightforward, using determinants rather than vector products:9
C = det(p1,...,pn) = F n det(P1,...,Pn) = F n A = F n / B (C.23)
D = det(q1,...,qn) = Rn det(f1,...,fn) = Rn B = Rn / A (C.24)
Based on Definition C.2 we can now formulate the generalization of Remark C.1 to the
MD n-fold periodic case with 1 ≤ n ≤ M (as a simple visual illustration consider, for
example, the 3D 2-fold periodic case or the 3D 3-fold periodic case):13

Remark C.4: Note the dual role of the vectors q1,...,qn in the signal and spectral domains:
In the signal domain, the vectors q1,...,qn determine (via their determinant D; see Footnotes
6 and 7 in Sec. C.3.1) the number of period-parallelepipeds of the MD n-fold periodic
function g(x1,...,xM), 1 ≤ n ≤ M, that:
(a) Fill the periodic n-dimensional sub-volume Rn of the sampling range volume RM (since
by Eq. (C.24) we have D = Rn / A); or, equivalently,
(b) Fill the periodic n-dimensional sub-volume Nn of the input array volume NM (since by
the general version of Eq. (C.20) we have D = Nn / C).
But in terms of the DFT spectrum, the vectors q 1,...,q n give the locations of the n
fundamental first-harmonic impulses (i.e. the n frequency-vectors of g(x1,...,xM) in terms of
array elements), as well as the distances between successive harmonic impulses (again, in
terms of array elements). Note that q1,...,qn need not necessarily be integer; when qj is
non-integer in its i-th coordinate the DFT spectrum suffers from a leakage artifact along
its i-th coordinate.
Similarly, the vectors p 1,...,p n , too, play a double role: In the signal domain they
determine the n-dimensional period-parallelepiped of g(x 1,...,x M ) in terms of array

13Once again, if n < M we can choose a new basis in R M along the n axes of periodicity, in which the
vectors become n-tuples. Thus, if n < M we consider the restriction of RM to the n-dimensional subspace
Rn in which the given n-fold periodic function is indeed periodic. To avoid complications, we assume
that these n axes are also used for sampling the given function. Note however that this assumption on
the sampling axes excludes here discrete cases that are not sampled along the axes of periodicity, like
the 2D 1-fold periodic cases shown in Fig. C.9, although p = N/q is still valid there (see Sec. 6.6.1).
Appendix C: The roles of p and q and their interconnections 291

elements. But in the DFT spectrum the vectors p1,...,pn determine (via their determinant C;
see Footnotes 6 and 7 in Sec. C.3.1) the number of frequency-parallelepipeds that:
(a) Fill the n-dimensional sub-volume Fn of the DFT frequency range volume FM (since
by Eq. (C.23) we have C = Fn / B); or equivalently
(b) Fill the n-dimensional sub-volume Nn of the output array volume NM (since by the
general version of Eq. (C.20) we have C = Nn / D).

Figure C.8: The effect of the p-vectors and the q-vectors of Eq. (C.11) on the input and
output arrays of the DFT of a periodic function in the multidimensional
case. Each row shows a 2D 2-fold periodic function (dot lattice made of
1-valued impulses) as well as its DFT spectrum (real and imaginary parts),
after all the required adjustments as discussed in Chapters 3, 4 and
Appendix A. We use here N = 32, so that the individual array elements are
sufficiently large and their locations can be easily identified. Each 2D input
array corresponds to the continuous-world range –8...8 along both axes,
and each 2D DFT spectrum covers the frequency range –1...1 Hz along
both axes, so that R = 16 and F = 2. But all the vectors are given here in
terms of array elements (pi and qi rather than Pi and fi). Note that in this
figure all the p-vectors are purely integer (meaning that all the impulses of
the input lattices are exactly located on input-array elements), but the
q-vectors in the resulting DFT spectra illustrate various integer, non-integer
and hybrid cases, as follows (we use italics to highlight the non-integer
coordinates):
(a) p1 = (8, 0), p2 = (0, 4) so that q1 = (4, 0), q2 = (0, 8): both q-vectors have
purely integer coordinates (so no leakage occurs in the DFT spectrum); the
2D input array is perfectly cyclical along both directions.
(b) p 1 = (9, 0), p 2 = (0, 4) so that q 1 = (3.555, 0), q 2 = (0, 8): q 1 is non-
integer in its first coordinate (so leakage occurs in the horizontal direction);
the 2D input array is perfectly cyclical (seamlessly wraparound) in the
vertical direction but not in the horizontal direction.
(c) p 1 = (8, 1), p 2 = (0, 8) so that q 1 = (4, 0), q 2 = (–0.5, 4): q 2 is non-
integer in its first coordinate; the effect is similar to that of case (b).
(d) p1 = (5, 1), p2 = (0, 4) so that q1 = (6.4, 0), q2 = (–1.6, 8): both q1 and
q2 are non-integer in their first coordinate; the effect is again as in case (b).
(e) p 1 = (5, 0), p 2 = (1, 4) so that q 1 = (6.4, –1.6), q 2 = (0, 8): q 1 is non-
integer in both coordinates (so that leakage occurs in both directions); the
2D input array is not perfectly cyclical in any of the two directions.
(f) p1 = (9, 0), p2 = (0, 5) so that q1 = (3.555, 0), q2 = (0, 6.4): q1 is non-
integer in its first coordinate and q2 is non-integer in its second coordinate;
the effect is similar to that of case (e).
(g) p1 = (8, –1), p2 = (1, 4) so that q1 = (3.88, –0.98), q2 = (0.98, 7.76): q1
and q2 are non-integer in both of their coordinates; the effect is similar to
that of case (e).
Note that in cases (c), (e) and (g) the DFT spectrum has a non-zero
imaginary-valued part although in the continuous-world counterparts the
spectrum is purely real-valued (as the infinite image-domain signal is
centrosymmetric). This artifact of the DFT is explained in Sec. 8.4.
292 Appendix C: The roles of p and q and their interconnections

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 1. 1.

q2
p2
0 0 0
p1 q1

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(a)
8. 1. 1.

q2
p2
0 0 0
p1 q1

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(b)
8. 1. 1.

p2
q2
0 p1 0 0
q1

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(c)
8. 1. 1.

q2
p2
0 0 0
p1 q1

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(d)
Appendix C: The roles of p and q and their interconnections 293

Spectral domain: Spectral domain:

Signal domain
Real part Imaginary part

8. 1. 1.

q2
p2
0 0 0
p1 q1

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(e)
8. 1. 1.

p2 q2
0 0 0
p1 q1

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(f)
8. 1. 1.

q2
p2
0 0 0
p1 q1

-8. -1. -1.

-8. 0 8. -1. 0 1. -1. 0 1.
(g)

Note that the vectors p1,...,pn, too, need not necessarily be integer; in fact, the only
restriction on p 1,...,p n and q 1,...,q n is that they must satisfy Eq. (C.11), where N is
obviously an integer number. Q

The nature of the p-vectors (integer or non-integer) has particular significance if high
harmonic impulses of the periodic function g(x) fall beyond the borders of the DFT
spectrum, i.e. if our sampled version of g(x) suffers from aliasing. Using our present
terminology we can now formulate the MD generalization of Remark C.2 as follows:13
294 Appendix C: The roles of p and q and their interconnections

Remark C.5: There exists a particular case in which no “stray impulses” appear in the
DFT spectrum of a periodic function g(x) even if strong aliasing occurs. This happens if
all the p-vectors have purely integer coordinates (see, for example, Fig. C.8). In this case
higher harmonic impulses which exceed from one side of the DFT spectrum due to
aliasing will fall exactly on top of true impulse locations (harmonic frequencies)14 in the
other side of the DFT spectrum; their heights will be simply summed up. Q

We can also formulate now the multidimensional generalization of Remark C.3 (this is,
in fact, a reformulation of Remarks 6.3 and 6.4 for the MD case):

Remark C.6: An MD n-fold periodic function g(x) is perfectly cyclical (seamlessly

wraparound) within the input array, meaning that it has no discontinuities on any of the
array boundaries when the input array is repeated periodically to pave the entire MD
space, if and only if its impulses in the DFT spectrum are located exactly on output array
elements (so there is no leakage), i.e. if and only if all the q-vectors have purely integer
coordinates. Q

This is clearly illustrated for the 2D case by Fig. C.9. Note that this result is indeed
quite intuitive, in view of the fact that the DFT assumes periodicity of its input data,
namely, that its input array should always be understood as a single period from a periodic
sequence which repeatedly extends ad infinitum in all directions: When the input array
seamlessly paves the entire space, this discrete paved space corresponds well to the
underlying infinite periodic function in the continuous world (like in Fig. C.9(b)). It is
therefore not surprising that in this case the DFT spectrum corresponds well to the
continuous-world spectrum, whose impulses are perfectly sharp. But when there are
discontinuities along the borders, the resulting paved space is artificially divided into
N×N-element squares (or N×...×N-element MD cubes) which do not seamlessly join
when paving the entire space, and the discrete paved space no longer corresponds to the
underlying infinite periodic function in the continuous world (see Fig. C.9(a)). In such
cases we cannot expect the DFT spectrum to perfectly agree with the continuous
spectrum. And indeed, this gives in the DFT spectrum unsharp, leaked impulses (each
impulse being convolved by an MD sinc function, as explained in Chapter 6).

Figure C.9: A 2D example illustrating Remark C.6: The 2D input array (containing
our sampled periodic function) is perfectly cyclical (seamlessly
wraparound) if and only if the impulses in the DFT spectrum are located
exactly on output array elements (so there is no leakage), i.e. if and only
if all the q-vectors have purely integer coordinates. The continuous-
world function being sampled here is a rotated cosine, whose spectrum

14 To be more precise, we speak here of impulse locations in terms of the algebraic support of the
spectrum of our periodic function, which consists of all the integer linear combinations of the basic
frequency vectors f1,...,fn of g(x), even if the corresponding impulse amplitudes happen to be zero
[Amidror09 p. 110].
Appendix C: The roles of p and q and their interconnections 295

Spectral domain:
Signal domain
Real part

-2. 0 2.
8. 2. 2.

0 0 0

-8. -2. -2.

-8. 0 8. -2. 0 2.
(a)

-8. 0 8. -2. 0 2.
8. 2. 2.

0 0 0

-8. -2. -2.

-2. 0 2.
(b)

consists of an impulse pair centered about the origin. In both rows we

use N = 64, R = 16, F = 4, Δx = 16/64 = 0.25 and Δu = 4/64 = 0.0625.
In the first row the cosine frequency is f = 0.5 and its angle is θ = 16°;
the second row is a rational approximation of the first row with
q = (7,2), so that f = 0.455 and θ = arctan(2/7) = 15.9454°.
296 Appendix C: The roles of p and q and their interconnections

Note, however, that in the multidimensional setting there may also exist hybrid cases that
behave differently along each of the M coordinates. To include these hybrid cases, too, we
can further extend Remark C.6 as follows:

Remark C.7: The MD input array is perfectly cyclical (seamlessly wraparound) along
the i-th coordinate if and only if the impulses in the MD DFT spectrum are located exactly
on array elements along their i-th coordinate (so there is no leakage in this direction), i.e.
if and only if the i-th coordinate of all the q-vectors is integer. And vice versa, it is enough
that one q-vector has a non-integer i-th coordinate in order that leakage will occur along
this direction, and in order that the input array will not be perfectly cyclical (seamlessly
wraparound) along the i-th coordinate. Q

As an illustration, consider Figs. C.8–C.9 which show examples belonging to various

hybrid cases based on 2D 2-fold periodic or 2D 1-fold periodic functions, along with their
DFT spectra.

Remark C.8: As explained in Sec. 6.6, a vector q in the output array is called rational (or
equivalently purely integer) if all its coordinates in terms of output array elements are
integer. Similarly, a vector p in the input array is called rational if all its coordinates in
terms of input array elements are integer. This allows us to reformulate Remarks C.5–C.7
above in terms of rational vectors in the input or output array. Q

Remark C.9: If all the p-vectors are purely integer then the period-parallelepipeds have
an integer volume; and similarly, if all the q-vectors are purely integer then the frequency-
parallelepipeds have an integer volume. However, the converse claims are not necessarily
true: The volume of a parallelepiped may be integer even if not all of its generating vectors
are purely integer. For example, in the 2D case the area of the parallelogram defined by
a b
the vectors p1 = (a1,b1) and p2 = (a2,b2) is given by the determinant of the matrix a1 b1 ,
2 2
namely, a1b2 – b1a2. It is clear that if both vectors are purely integer so is the determinant;
but if b1 = 0, then the determinant will be integer even if a2 is not an integer. Q

Table C.2 offers a synoptic view of the roles of the p-vectors and the q-vectors in the
MD case, both in the signal and in the spectral domains.

Finally, in order to avoid any possible confusion, let us repeat here the distinction
between the following four attributes that are largely used throughout the above
discussions, and which are all related to periodicity:

Remark C.10: Periodic, folded-over, wrapped-around and cyclical:

A function (or a structure) is said to be periodic if it is composed of a finite segment
(called period) that repeats itself in all directions ad infinitum. For ecample, the 1D
function g(x) = cos(2πfx) and the 2D nailbed g(x,y) = III(x,y) are periodic. As we have
seen in Secs. 2.5 and D.1, the input and output arrays of the DFT are periodic: each of
them should be understood as a single period from a periodic sequence which repeatedly
extends ad infinitum in all directions.
Appendix C: The roles of p and q and their interconnections 297

Effect on the input Effect on the output Examples

array (signal domain) array (DFT spectrum)

All q-vectors are Input array is perfectly Rational angles and Figs. C.8(a)
purely integer cyclical (seamlessly frequencies; all impulses are and C.9(b)
wraparound) along located exactly on array
all dimensions (see elements; no leakage
Remarks C.6 and C.7). (see Sec. 6.6).

At least one Input array is not perfectly Irrational angle or frequency Figs. C.8(b)–(g)
q-vector qj has a cyclical (seamlessly wrap- along the j-th q-vector; and C.9(a)
non-integer i-th around) along the i-th impulses are located
coordinate (see between array elements
coordinate
Remarks C.6 and C.7). along the i-th coordinate;
leakage along the i-th
coordinate (see Sec. 6.6).

All p-vectors are Rational angles and If aliasing occurs, folded- Figs. C.8(a)–(g)
purely integer periods; period area (or over impulses fall exactly
volume) of g(x) is exactly on top of lower-harmonic
an integer number of array impulse locations, and their
elements (see Remark C.9). amplitudes are summed up
(see Remark C.5).

At least one Irrational angle or period If aliasing occurs, folded-

p-vector pj has a along the j-th p-vector; over impulses fall between
non-integer i-th period area (or volume) of lower-harmonic impulse
g(x) may be a non-integer locations, causing visible
coordinate
number of array elements stray impulses
(see Remark C.9). (see Remark C.5).

Table C.2: Multidimensional generalization of Table C.1: The effects of the p-vectors
and of the q-vectors of Eq. (C.11) on the input and output arrays of the MD
DFT of an MD n-fold periodic function g(x1,...,xM), 1 ≤ n ≤ M.13 Note that
the effects of the p- and q-vectors are orthogonal (i.e. independent of each
other), and just as in the 1D case each of the 4 combinations may exist.

The term foldover or folding-over is used when frequencies of the continuous spectrum
(or any other components thereof, such as impulses, etc.) fall beyond the boundaries of
the DFT spectrum, due to aliasing. Because of the periodicity of the DFT spectrum (the
output array), such exceeding entities re-enter from the opposite end of the DFT spectrum
and give the impression of folding over (see, for example, row (b) in Fig. C.2). It is
important to note that any folded-over entity is additively overlapped (summed up) on top
of the spectral components that were originally present at that location. This happens, as
we have seen in Chapter 5, due to the additive nature of the convolution with an impulse
comb or nailbed.
298 Appendix C: The roles of p and q and their interconnections

The term perfectly cyclical or seamlessly wraparound is used when a periodic function
is sampled into our input or output array (1D, 2D or MD), and its periodicity exactly fits
into the array, so that when this array is periodically repeated in all directions, no seams
appear between its neighbouring copies (see Remarks C.3 and C.6). This is illustrated by
the signal domain in Figs. 6.3(a),(e),(f) for the 1D case, or in Fig. C.9(b) for the 2D case.
Confusingly, here too one may say that “every element which exceeds from one side of
the array re-enters from the opposite side of the array”; however, unlike folding-over,
wraparound is not additive, and it should only be understood as a geometric consideration.
When the array in question is not seamlessly wraparound (see, for example, Fig.
C.9(a)), it still repeats itself periodically, like any input or output array of the DFT.
Simply, it will contain jumps or discontinuities along the seams between neighbouring
copies of the array. But because of this periodic repetition, we can still say that all the
neighbouring copies of the array wrap around and coincide with the main copy, meaning
that each of them represents, in fact, the same entity (see Problem 8-17). In this case, too, it
can be said that whatever exceeds from one side of the array re-enters from the opposite
side (thanks to the periodicity) — but of course, elements which exceed beyond the
boundaries may suffer from jumps or discontinuities.
Note that the term wraparound is also used in expressions such as “shift with
wraparound” (see, for example, Footnote 2 in Sec. 3.2), referring to the cyclical nature of
the operation (meaning, again, that whatever exceeds from one side of the entity in
question re-enters from the opposite side). Sometimes the term “additive wraparound” is
also used as a synonym to foldover.
Finally, the term cyclical is more ambiguous, since it can be used as a synonym to
periodic or to wraparound, depending on the context. Q
 Appendix D
Miscellaneous remarks and derivations

D.1 The periodicity of the input and output arrays of the DFT
As we have seen in Sec. 2.5, the DFT imposes a cyclic behaviour on both of its input
and output data sequences, in the sense that each of them should be understood as a single
N-element period (or “truncation window”) from a periodic sequence which repeatedly
extends to both directions ad infinitum (even if the underlying continuous data before
sampling was not periodic). In the informal graphical development of the DFT, as
presented in Sec. 2.5, this periodicity is explained by the fact that sampling (i.e.
multiplication with an infinite impulse train) in one domain is equivalent to a convolution
with the reciprocal infinite impulse train in the other domain. Thus, sampling in the signal
domain results in a periodic behaviour in the frequency domain, and sampling in the
frequency domain results in a periodic behaviour in the signal domain (see Fig. 2.1).
More formally, the reason for this periodicity is that the exponential part in the definition
of the DFT (for example, in Eq. (2.13)) is periodic in the index n with period N.1 To see
this, let n = r + N; we have, therefore:

e–i2π kn/N = e–i2π k(r+N)/N = e–i2π kr/N e–i2π k

= e–i2π kr/N
since e–i2π k = cos(2πk) – i sin(2πk) = 1 for any integer k [Brigham88 p. 96]. This implies
that for any integer j:
G(n + jN) = G(n) (D.1)
Similarly, the periodicity of the exponential part in the definition of the IDFT (for
example, in Eq. (2.14)) implies that for any integer j:
g(k + jN) = g(k) (D.2)
Therefore, if we desire to extend the definition of g(k) and G(n) beyond the original
domains given by k = 0, ... ,N–1 and n = 0, ... ,N–1, this extension must be governed by
Eqs. (D.1) and (D.2). This means that the extensions in the signal and spectral domains
are periodic repetitions of the original N-element arrays g(k), k = 0, ... ,N–1 and G(n),
n = 0, ... ,N–1, respectively.
The 2D generalization of this reasoning is straightforward, as shown for example
in [Rosenfeld82 pp. 22–24]: The two-fold periodicity of the exponential part in the

1 Note that no similar periodicity exists in the exponential part of the various definitions of the CFT (for
example, in Eq. (2.1)).

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 299
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_12,
© Springer-Verlag London 2013
300 Appendix D: Miscellaneous remarks and derivations

definition of the 2D DFT, for example in Eq. (2.23), implies that for any integers j1
and j2:
G(m + j1N, n + j2N) = G(m,n) (D.3)
and similarly, in the 2D IDFT we have for any integers j1 and j2:
g(k + j1N, l + j2N) = g(k,l) (D.4)

Therefore, if we desire to extend the definition of g(k,l) and G(m,n) beyond the original
domains given by k,l = 0, ... ,N–1 and m,n = 0, ... ,N–1, this extension must be governed
by Eqs. (D.3) and (D.4). This means that the extensions in the signal and spectral
domains are 2D periodic repetitions of the original N×N-element arrays g(k,l), k,l =
0, ... ,N–1 and G(m,n), m,n = 0, ... ,N–1, respectively.

This result can be readily extended to cases having a different number of elements along
each dimension, and to cases with any higher number of dimensions.

D.2 Explanation of the element order in the DFT output array

We have seen in Chapter 3 that the elements in the output array of the DFT are not
obtained in the usual order, so that a reorganization of the output array is required if we
wish to plot the DFT results in the conventional way. In the present section we give the
explanation for the 1D case; the extension to the 2D and MD cases can be obtained by
applying these considerations to each dimension separately.

As we have seen in Sec. 3.2, 1D DFT algorithms receive as input a 1D array of N

complex numbers (say, samples of a given continuous complex- or real-valued function
y = g(x)), and return as their output a 1D array of N complex numbers, indexed by n =
0,...,N–1, which contains the resulting discrete frequency spectrum. However, the
frequency values in the resulting output array are not ordered in the conventional way we
use to plot the spectrum: The first element in the array (n = 0) corresponds to the
frequency 0; the second element (n = 1) corresponds to f1, the smallest frequency above 0
in the discrete spectrum; and the next elements represent the following frequencies of the
discrete spectrum, i.e. 2f1, 3f1, etc. However, this is only true until n = N/2–1; the output
array elements starting from n = N/2 and until the last element, n = N–1, correspond to the
negative frequencies of the spectrum. Thus, the last element of the output array, n = N–1,
corresponds to the negative frequency –f1, which should conventionally be plotted just
before the frequency 0; its predecessor, the element n = N–2, corresponds to the frequency
–2f1; and finally, the element n = N/2 corresponds to –(N/2)f1, the most negative frequency
in the spectrum.2 It follows, therefore, that in order to plot the resulting spectrum in the
conventional way, with the zero frequency in the center and the negative frequencies to its

2 In the vast majority of DFT algorithms N is a power of two, or at least an even number. In the rare cases
where N is odd the reorganization of the array should be adapted accordingly [Briggs95 Sec. 3.1].
Appendix D: Miscellaneous remarks and derivations 301

left, we have to reorganize the output array of the DFT by interchanging (swapping) its
first half (the segment containing the elements n = 0,...,N/2–1) and its second half (the
segment containing the elements n = N/2,...,N–1).3 This reorganization of the spectrum is
not usually done by the DFT algorithm [Brigham88 p. 169], and it is up to the user to
perform it when plotting the DFT results.

Why does the DFT behave this way? First of all, it should be mentioned that this
complication only occurs when using a DFT definition with non-centered indices, like
Eqs. (2.13)–(2.14), in which the indices vary between 0 and N–1 rather than between –N/2
and N/2–1 [Briggs95 p. 67]; but because this is the standard convention used in virtually
all software applications of the DFT (see, for example, the partial list given at the end of
Sec. 2.3), we have to accept this fact and get used to it. So let us see now in detail why the
DFT with non-centered indices behaves this way. This ordering is a consequence of the
N-element periodicity of the output array of the DFT (see Sec. D.1), which implies that
G(N–n) = G(–n). But as shown in Fig. D.1, the periodicity alone does not yet explain this
ordering, since it could still allow many other possibilities — for example, the wrong
output array ordering shown in Fig. D.1(b), in which the heighest DFT frequency is
located in the last element of the output array, N–1. In order to obtain the correct ordering,
shown in Fig. D.1(c), and to exclude all the other possibilities, we need therefore an
additional condition, as explained below.

(1) If the input signal is purely real, which is indeed the case in most situations, then both
G(u) and its DFT counterpart are Hermitian (see, for example, [Bracewell86] pp. 15
and 366, respectively). This additional information indeed enforces the correct ordering
of the DFT output array shown in Fig. D.1(c): Because we are using the non-centered
definition of the DFT (Eq. (2.13)) it follows that the first element in the DFT output
array is n = 0, which corresponds to the frequency 0 (i.e. to the origin of the CFT).
Then, the Hermitian nature of the DFT output implies that for any integer n:

G(n) = G*(–n)

and due to the periodicity with period N:

= G*(N–n)

which means that Re[G(n)] = Re[G(N–n)] and Im[G(n)] = –Im[G(N–n)]. In particular

we have:
G(1) = G*(N–1)
G(2) = G*(N–2)

G(N ) = G*(N– N ) = G* ( N )
2 2 2
3 Note that this is also equivalent to a cyclic shift of the array (i.e. a shift with wraparound) through N/2
positions in either direction.
302 Appendix D: Miscellaneous remarks and derivations

(see, for example, Fig. 3.1(b)). This means that the ordering of the DFT output is,
indeed, as shown in Fig. D.1(c).
(2) However, even if the input signal is not purely real, meaning that the Hermitian
property no longer holds, the DFT output array is still ordered in the same way. To see
this, remember that when sampling a continuous function g(x) with a sampling interval
Δx, no frequencies beyond ± 2Δx 1
may appear in the resulting DFT spectrum [Smith07
p. 229; Harris98 p. 713; Briggs95 p. 9; Castleman79 p. 235]. In other words, by virtue
of Eq. (4.6), the DFT spectrum can only contain frequencies up to half of the sampling
frequency fs [Smith03 p. 42], i.e. up to (N/2)f1. And indeed, as we have seen in
Eq. (4.9), fmax = (N/2–1)f1 and fmin = –(N/2)f1. Hence, the maximum positive frequency
is located in the position n = N/2–1 of the output array and the minimum negative
frequency is located in the negative position n = –N/2 (see Fig. D.1(c)). Since we are
using a non-centered definition of the DFT that is formulated for values of n in the
interval 0,...,N–1, negative values of n are not included. But by virtue of the N-element
periodicity of the DFT we see that the elements in the positions N/2, ... ,N–1 of the
output array are indeed a copy of the elements in the negative positions –N/2, ... ,–1.

Note that the explanation given in point (2) is more general than that given in point (1)
since it covers all cases, including those with purely real input signals. This means that the
explanation provided in point (1) is redundant, but we still mention it here because it
appears quite frequently in the literature (see, for example, [Gonzalez87 p. 78]).

D.3 The beating effect in a sum of cosines with similar frequencies

Let us consider the sum of two cosines having close frequencies f1 and f2. Based on the
well-known trigonometric identity [Harris98 p. 284; Spiegel68 p. 17]:
cosα + cosβ = 2cos α + β cos α – β
2 2
we obtain:
cos(2π f1x) + cos(2π f2x) = 2 cos(2π f1 + f2 x) cos(2π f1 – f2 x) (D.5)
2 2
If the frequencies f1 and f2 are similar, this cosine product corresponds to a modulation
effect, where the cosine with the higher frequency f1 + f2 represents the carrier and the
2
cosine with the low frequency f1 – f2 represents the envelope (see Fig. D.2(a)).4 However,
2
because the envelope periodicity is in fact half of one full cosine period (see Fig. D.2(b))
it turns out that the envelope’s frequency is equal to the frequency difference f1 – f2.

Thus, the sum of two cosines with very close frequencies f1 and f2 gives a beating effect
(pseudo moiré) whose frequency is f1 – f2. Note, however, that this beating is not a true

4 If the original frequencies are very different, say f1 << f2, the cosine sum no longer resembles such a
modulation, but rather takes the form of a low-frequency cosine wave of frequency f1 having a high-
frequency perturbation of frequency f2 (or in other words, a high-frequency cosine wave of frequency f2
riding on a low-frequency cosine wave of frequency f1).
Appendix D: Miscellaneous remarks and derivations 303

Spectral domain

G(u)

u
0
(a)
Matching zone
... ... ...
DFT counterpart ... ... ...
of G(u) — wrong ..... ..... .....
................. ............ ............ ........
N-element period: ......... ......... u
0 f1 Nf1
(b)
Matching zone
. . .
DFT counterpart
..... ..... ..... ..... ..... .....
of G(u) — correct .... .....
...............
. .....
...............
. .....
N-element period: ....... ......
u
0 f1 Nf Nf1
2 1
(c)
The N discrete elements
of the DFT output array

Figure D.1: The element order in the DFT output array. Suppose we are given a
continuous-world function g(x) whose spectrum (continuous Fourier
transform) is G(u), as shown in (a). The discrete counterpart of G(u)
(the output array of the DFT) consists of N complex elements, and it
is periodic with period N. This periodicity means that the DFT output
corresponds to the continuous Fourier transform (a) within only one
sequence of N output elements; but outside this range this sequence
of N elements is repeated ad infinitum, and it no longer matches G(u).
In principle, the sequence of N elements that do match G(u) could be
situated in many different ways, for example as shown in (b) or in (c).
However, we show that the only possible ordering of the DFT output
array is indeed the one shown in (c). This means that when using a
non-centered definition of the DFT, indexed by n = 0, ... ,N–1, the
elements in the positions n = N/2, ... ,N–1 of the DFT output array are
indeed a copy of the elements that are located in the figure in the
negative positions –N/2, ... ,–1. Note that even within the “matching
zone” there still may exist a discrepancy between the DFT results and
G(u) due to aliasing and leakage (see Chapters 5 and 6).

moiré effect, since the corresponding spectrum contains no impulse having the new low
frequency f1 – f2: According to the addition rule (see point 1b in Sec. 2.4.1), the Fourier
transform of the cosine sum (D.5) only contains the frequencies that are contributed by
the two original cosines themselves, namely, ±f1 and ±f2.
304 Appendix D: Miscellaneous remarks and derivations

4
(a)

-15 -10 -5 5 10 15

-2

1
Envelope period:
f1- f2
-4

4
(b)

-15 -10 -5 5 10 15

-2

2
Cosine period:
f1- f2
-4

4
(c)

-15 -10 -5 5 10 15

1 -2 1
f1 f2

-4

Figure D.2: (a) The sun of two cosinusoidal waves with similar frequencies f1 and f2,
namely, cos(2π f1x) + cos(2π f2x). (b) The envelope of the modulated
wave (a) is given by the cosine 2 cos(2π f1 – f2 x), but its period is half of
the period of this cosine. The frequency 2of the envelope is therefore
twice f1 – f2 , namely f1 – f2. For the sake of completeness, we show in (c)
the two2 original cosines being used in the present figure: The cosine
cos(2π f1x) with f1 = 1 is plotted by a continuous line (left), while the
cosine cos(2π f 2 x) with f 2 = 1.1 is dashed (right); both curves are
overprinted in the central part of (c) to allow a better understanding of
their sum in (a).
Appendix D: Miscellaneous remarks and derivations 305

(a) (b) (c)

(d) (e) (f)

Figure D.3: 2D cosinusoidal signals (a) and (b) having frequencies f1 and f2 and
their sum (c) in the signal domain; their respective spectra are the
impulse pairs shown in (d) and (e) and their sum (f). These spectra
were obtained by applying 2D DFT to each of the first-row images.
Note the beating effect in the sum (c), whose frequency is f1 – f2.
This is a pseudo moiré effect since it does not have corresponding
impulses in the spectrum (f).

The situation in the 2D case is similar, but a little more complex since the carrier and the
envelope waves in the 2D case may have different orientations (as we can see, for example,
in Fig. D.3(c)). To simplify the presentation in the 2D case, we adopt the following vector
notation:
cos(2π fi [xcosθi + ysinθi ]) = cos(2π fi ·x)

where fi = f i ·(cosθi , sinθi ) and x = (x,y). Using this vector notation the generalization of
Eq. (D.5) to the 2D case of Fig. D.3 is:

cos(2π f1·x) + cos(2π f2·x) = 2 cos(2π f1 + f2 ·x) cos(2π f1 – f2 ·x) (D.6)

2 2
306 Appendix D: Miscellaneous remarks and derivations

Thus, here too, the sum of the two cosinusoidal layers gives a product of cosines. If the
original cosine frequencies f1 and f2 are similar, this cosine product corresponds to a 2D
modulation effect, where the cosine with the higher frequency f1 + f2 represents the carrier
2
and the cosine with the low frequency f1 – f2 represents the envelope (see Fig. D.3(c)).
2
However, once again, because the envelope periodicity is only half of the cosine period it
turns out that the envelope’s frequency is equal to the frequency difference f1 – f2. And
since, just as in the 1D case, the spectrum of the 2D cosine sum (D.6) (see Fig. D.3(f))
does not contain an impulse having the new low frequency f1 – f2, it follows that this
beating is not a true moiré effect, but just a pseudo-moiré. For further details on 2D moiré
effects and their spectra, see Sec. 2.3 in [Amidror09].

These considerations can be extended in a similar way to the general MD case, too.

A more detailed discussion on the beating effect (pseudo moiré) that is generated in the
sum of cosines with similar frequencies can be found in [Amidror09a, Sec. 3].

D.4 Convolutions with a sinc function which have no effect

As shown in Fig. 6.1, the truncation of a function g(x) into a finite sampling range of
length R, namely, the multiplication of g(x) by the windowing function w(x) = rect(x/R), is
equivalent, in terms of the spectral domain, to a convolution of the CFT of g(x), G(u), with
the CFT of w(x), W(u) = R sinc(Ru). This convolution with W(u) is, indeed, the source of
the leakage artifact that G(u) undergoes in the spectral domain (see, for example, Figs.
6.1–6.5). However, an interesting question may arise at this point: Consider the particular
case in which the given function g(x) has a finite duration which is fully included within
the sampling range. Obviously, there exist infinitely many different functions g(x) that
satisfy this condition. It is clear that in all these cases the truncation into the sampling
range will not affect g(x), since g(x)w(x) = g(x). This implies in the spectral domain that
G(u)*W(u) = G(u), meaning that in all these cases the convolution of G(u) with W(u) =
R sinc(Ru) gives back G(u), so that G(u) is not affected by this convolution and does not
suffer from leakage (see case (b) in Remark 6.1). The question is: how can it be that in
such cases the convolution with W(u) does not modify (smooth out) G(u)? Indeed, it may
be surprising to see that in all these different cases the convolution with the sinc function
W(u) has no effect whatsoever, not even smoothing out or adding ripple. The situation
here is obvious when we consider the product of g(x) with the truncation window w(x); but
it remains quite surprising when we consider the dual question in terms of convolution.

To take a concrete example, consider any function g(x) = rect(x/a) with a ≤ R. It is clear
that in all these cases we have g(x)w(x) = g(x); but the dual result, which says that
a sinc(au)* R sinc(Ru) = a sinc(au), is not a-priori obvious. Similarly, given a finite-
duration function h(x), any rectangular windowing function wr(x) = rect(x/r) that is larger
than h(x) will give exactly the same product h(x)wr(x) = h(x) (i.e. no effect). But once
again, this is not obvious to see when considering the convolution in the other domain.
Appendix D: Miscellaneous remarks and derivations 307

Similar questions may arise in many different situations and in various disguises (see,
for example, case (b) in Remark 6.1, or, with a domain reversal, in Sec. 6.7). This is indeed
a good demonstration of the power of the Fourier theory, which may often transform
complicated questions in one domain into trivial ones in the dual domain.

D.5 The convolution of a square pulse with a sinc function

In this section we provide the analytic derivation of the convolution of the functions
u
W(u) = R sinc(Ru) and G(u) = 1 rect(2a ) (see Example 6.2 in Chapter 6).
2a
∞
C(u) = W(u)*G(u) =
∫ –∞
W(z)G(u–z) dz

∞
= R
2a ∫ –∞
sinc(Rz) h(u–z) dz

where the function h(z) = rect( z ) has the value 1 within the range –a...a, and is zero
2a
everywhere else. Because h(z) is symmetric about the origin, h(–z) has the value 1 in the
same range –a...a, and therefore the function we actually have inside the integral, h(u–z) =
h(–z+u), has the value 1 in the range –a–u...a–u. We therefore have:
a–u
= R
2a ∫ –a–u
sin(πRz)
πRz
dz

This integral can be simplified by using the variable substitution t = πRz which implies
also dz = 1 dt. Following this variable substitution our integration limits:
πR
–a–u ≤ z ≤ a–u
change into: πR(–a–u) ≤ t ≤ πR(a–u)

and we therefore obtain:

πR(a–u)

∫
R sin(t) 1
= dt
2a πR
πR(–a–u) t

0 πR(a–u)

∫ ∫
1 sin(t) sin(t)
= dt + 1 dt
2aπ t 2aπ t
πR(–a–u) 0

πR(u+a) πR(u–a)

∫ ∫
= 1 sin(t) dt – 1 sin(t) dt
2aπ t 2aπ t
0 0

Now, using the definition of the sine integral function [Weisstein99 p. 1646]:
x
Si(x) =
∫ sin(t)
t
0
dt
308 Appendix D: Miscellaneous remarks and derivations

-1

-2

-3
-30 -20 -10 0 10 20 30

(a)

1.5

0.5

-0.5

-1
-1.5 -1 -0.5 0 0.5 1 1.5

(b)

Figure D.4: (a) Graphic representation of the function Si(x). (b) Graphic
representation of the functions G(u) = 2a1 rect(2au ) and G(u)*W(u)
of Eq. (D.7) for the case of a = 1/2.

(see Fig. D.4(a)) we finally get:

= 1
2aπ
[Si(πR(u+a)) – Si(πR(u–a))] (D.7)

As shown in Fig. D.4(b), this function resembles a smoothed-out version of G(u)

(which is a square pulse of height 2a1 extending between –a and a), but it has an overshoot
Appendix D: Miscellaneous remarks and derivations 309

to each side of the two transitions, followed by gradually decaying oscillations to both
directions (see also the dashed lines in the spectra shown in Fig. 6.4).

Note that the oscillations of the sine integral function are the reason for the overshoot
and ripple artifacts which occur in signal processing in various circumstances, such as
when using a truncated sinc filter as an approximation to the ideal low-pass filter (see
Example 3.3 and Fig. 3.7 in Chapter 3), or the well known Gibbs phenomenon
[Kammler07 pp. 44–45; Bracewell86 pp. 209–211].

We terminate this section with an interesting remark that is used later in Sec. D.6:

Remark D.1: Although the function G(u)*W(u) of Eq. (D.7) is just a smoothed-out and
slightly oscillating version of the original function G(u) = 2a1 rect(2au ), there still exists
between them a major difference in their behaviour when a tends to zero: As a→0, the
value of G(u) at u = 0 tends in the limit to infinity, but the value of G(u)*W(u) at u = 0
tends in the limit to the finite value R.
The first part of this claim is rather obvious, since the value of G(u) at u = 0 is 2a1 , which
tends to infinity when a→0. To show the second part of the claim, we note from Eq. (D.7)
that at u = 0 the value of C(u) = G(u)*W(u) is:

C(0) = [Si(πRa) – Si(–πRa)]

1
2aπ
= 1 [Si(πRa) + Si(πRa)]
2aπ

= 1 Si(πRa)
πa

We want to show now that:

Si(π Ra)
lim =R
a→0 πa

As suggested by the Wolfram Alpha® website [WolframAlpha09], we start by factoring

out constants:
Si(π Ra) Si(π Ra)
lim = 1 lim a
a→0 πa π a→0

This is a limit of the type 0/0. Applying l’Hopital’s rule we have:

dSi(πRa)
= 1 lim da
π a→0 da
da

= 1 lim πR sin(πRa)
π a→0 π Ra

= 1 lim πR sinc(Ra) (where sinc(x) = sin(πx) )

π a→0 πx

= R lim sinc(Ra)
a→0
310 Appendix D: Miscellaneous remarks and derivations

Using the continuity of sinc(x) at x = 0:

= R sinc( lim Ra)

a→0

= R sinc(R lim a)
a→0

= R sinc(0)
This gives, indeed:
=R
which was to be demonstrated. Q

The significance of this result is explained in Sec. D.6 below.

D.6 A more detailed discussion on Remark A.2 of Sec. A.3

As we have seen in Chapter 6, unlike the continuous Fourier transform, the DFT can
only take into account a finite range of the original, underlying function g(x), meaning that
g(x) must be truncated before the application of the DFT. This truncation operation can be
viewed as a multiplication of the underlying function g(x) by a rectangular truncation
window w(x) whose length equals the sampling range R = NΔx, namely, w(x) = rect(x/R).5
The spectral-domain counterpart of this truncation is that the DFT does not approximate
the original Fourier transform G(u), but rather the convolution of G(u) with the Fourier
transform of the truncation window w(x), W(u) = R sinc(R u). As long as G(u) is non-
impulsive, the convolution with W(u) will smooth out G(u) and add a ripple effect (see a
worked-out example in Sec. D.5). However, when G(u) includes impulses, the convolution
with W(u) will replace each impulse ciδ(u–ai) by an additive replica of R sinc(R u) that is
scaled by the impulse amplitude ci and centered about the impulse location ai. As we have
seen in Sec. 6.5, if the continuous-world impulse is located exactly on an element of the
DFT output array, giving a sharp discrete-word impulse in the DFT spectrum, this replica
of R sinc(R u) will simply multiply the discrete impulse amplitude ci by R, but have no
effect on the other elements of the output array. And if the continuous-world impulse is
located between elements of the DFT output array, so that the resulting discrete impulse is
smoothed out and smeared due to leakage, the multiplication by R (and hence the
dependence on the choice of the sampling range length R) will affect both the smoothed
out impulse and its leaked tails, but again it will have no other effect on the DFT output
array. This explains, indeed, why the impulse amplitude in the DFT spectrum is multiplied
by R with respect to the original continuous-world spectrum G(u), and why this
dependence of the DFT on the choice of R affects only impulses (and their leaked tails, if
any) but no other components of G(u).

5 This truncation can be done either before or after sampling; see Sec. D.7 below. In practice, this is
simply done by sampling the underlying function g(x) within a finite domain.
Appendix D: Miscellaneous remarks and derivations 311

To better understand this point, it may be instructive to see what happens in the DFT
when a non-impulsive continuous spectrum such as G(u) = 2a1 rect(2au ) gradually becomes
narrower and taller and approaches an impulse. Figures D.5(a)–(f) on the following pages
show a sequence of continuous-world functions g(x) = sinc(2ax) with gradually
decreasing values of the parameter a, going from a = 1/2 in (a) to a = 1/64 in (f), along
with their respective continuous Fourier transforms G(u) = 2a1 rect(2au ) and their DFT
counterparts. Each of the cases (a)–(f) is shown twice, with a truncation window of length
R = 8 (top row) or R = 16 (bottom row); the number of sampling points remains N = 32 in
all cases. As the parameter a tends to zero, the value of G(u) at the point u = 0 tends to
infinity (since it equals 2a1 ); but the value of the convolution G(u)*W(u) (Eq. (D.7)) at the
point u = 0 does not tend to infinity but rather to R (see the derivation in Remark D.1
above). And indeed, as we clearly see in the figures, when a→0 the CFT G(u) = 2a1 rect(2au )
(a pulse of height 2a1 extending between –a and a, plotted in the spectra by continuous
lines) tends in the limit into a continuous-world impulse at u = 0 [Bracewell86 pp. 69–72];
but on the other hand, when a→0 the function G(u)* W(u), plotted in the spectra by
dashed lines, becomes a sinc-like function whose peak at u = 0 tends to R. Since the DFT
of the original function consists of samples of G(u)*W(u) rather than samples of G(u), it
follows that the discrete-world impulse that is obtained in the DFT when a→0 has the
height R (while the corresponding continuous-world impulse is conventionally drawn in
the spectrum as a unit-height spike, although its height at u = 0 is infinite; see Sec. A.2 of
Appendix A). This implies, in turn, that the height of our discrete-world impulse depends
on the choice of the truncation range length R. Note that this multiplication of the height
by R in the discrete world (and the resulting dependence on the value of R) affect only
impulses, but not continuous components in the spectrum: As we can see by comparing
the top and bottom rows in each of Figs. D.5(a)–(d), as long as the DFT is not impulsive
its height follows the height of the CFT, G(u) = 2a1 rect(2au ), and is therefore independent of
R (except for small fluctuations due to the ripple effect). But as we see in Figs. D.5(e)–(f),
as soon as a becomes sufficiently small to make the DFT impulsive, or more precisely, as
soon as the continuous-world pulse width 2a (extending between –a and a) becomes
equal or smaller than the DFT frequency step Δu = 1/R (see Eq. (7.3)), the height of the
resulting DFT impulse at u = 0 no longer follows the height of the CFT G(u) but starts to
approach R, meaning that now it does depend on the choice of R. Note that the discrete
impulse height in the DFT spectrum reaches the exact value R only when a = 0; in this
limit case the original signal-domain function g(x) = sinc(2ax) becomes a straight
horizontal line g(x) = 1, and in the DFT spectrum all the samples have the height 0, except
for the sample corresponding to u = 0 which has the height R. Fig. D.5(f), in which
a = 1/64, is already quite close to this limit case, both in the signal and in the spectral
domains.

Table D.1 summarizes the behaviour when a→0 of the CFT G(u) = 2a1 rect(2au ) and of the
corresponding DFT (which follows, as we have seen, the behaviour of G(u)*W(u) rather
than that of G(u)).
 312 Appendix D: Miscellaneous remarks and derivations

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0

• • • DFT after CFT of

-0.5 -0.5 g(x)w(x), i.e.
i&o reorg.
and vert. sca. G(u) *W(u)
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

R = 8, Δu = 1/8, p.w. = 2a > Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 1.5 n
g(x)=sinc(2ax)
1 • • • Sampled 1

0.5 0.5
CFT
0 0
• • • DFT after CFT of
-0.5 -0.5 i&o reorg. g(x)w(x), i.e.
and vert. sca. G(u) *W(u)
-1
-8. -6. -4. -2. 0 2. 4. 6. 8.
x -1
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

R = 16, Δu = 1/16, p.w. = 2a > Δu

(a) a = 1/2

Figure D.5: A gradual transition from a non-impulsive DFT spectrum to an

impulsive DFT spectrum, illustrating in particular the transition
from R-independent to R-dependent DFT spectra (see Sec. D.6).
Parts (a)–(f) of the figure show a series of continuous-world
functions g(x) = sinc(2ax), along with their respective continuous
Fourier transforms G(u) = 2a1 rect(2au ), for six gradually decreasing
values of a, going from a = 1/2 in (a) to a = 1/64 in (f). In each
case, the continuous lines correspond to the continuous-world
function g(x) and to its CFT G(u), and the black dots correspond
to the sampled version of g(x) and to its DFT spectrum. The
dashed lines in the spectra correspond to the CFT of g(x)w(x)
where w(x) = rect(x/R), i.e. to the continuous-world convolution
G(u) * W (u) (see Eq. (D.7)), which is the continuous-world
function being sampled by the discrete elements of the DFT
spectrum. Each of the cases (a)–(f) is shown twice, with a
truncation window of length R = 8 (top row) or R = 16 (bottom
 Appendix D: Miscellaneous remarks and derivations 313

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 3 n

1 2

0.5 1
CFT
0 0

• • • DFT after CFT of

-0.5 g(x)=sinc(2ax) -1
i&o reorg. g(x)w(x), i.e.
• • • Sampled and vert. sca. G(u) *W(u)
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -2
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

R = 8, Δu = 1/8, p.w. = 2a > Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 3 n

1 2

0.5 1
CFT
0 0
• • • DFT after CFT of
-0.5 g(x)=sinc(2ax) -1
i&o reorg. g(x)w(x), i.e.
• • • Sampled and vert. sca. G(u) *W(u)
-1
-8. -6. -4. -2. 0 2. 4. 6. 8.
x -2
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

R = 16, Δu = 1/16, p.w. = 2a > Δu

(b) a = 1/4

row); the number of sampling points remains N = 32 in all cases.

In cases (a)–(d) the dashed lines as well as the dots representing
the DFT spectrum follow the continuous-world spectrum G(u), up
to some minor oscillations (which correspond to the ripple effect
due to leakage, as explained in Chapter 6). But when a gets smaller,
in cases (e)–(f), the dashed lines and the DFT dots break away
from the continuous-world pulse G(u) (whose height continues to
grow to infinity), and instead of following it they rather tend to the
height R (and hence they become dependent on the choice of the
truncation range length R). This break away starts when the
continuous-world pulse width (p.w. = 2a) becomes smaller than the
DFT frequency step Δu = 1/R. And indeed, note that when R = 8
(top row) this break away begins in case (e), i.e. starting from
a = 1/32, but when R = 16 (bottom row), where the pulse width is
twice wider (in terms of DFT output elements), the break away only
starts in case (f), i.e. when a = 1/64.
 314 Appendix D: Miscellaneous remarks and derivations

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 6 n

1 4

0.5 2
CFT
0 0

• • • DFT after CFT of

-0.5 g(x)=sinc(2ax) -2
i&o reorg. g(x)w(x), i.e.
• • • Sampled and vert. sca. G(u) *W(u)
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -4
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

R = 8, Δu = 1/8, p.w. = 2a > Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 6 n

1 4

0.5 2
CFT
0 0
• • • DFT after CFT of
-0.5 g(x)=sinc(2ax) -2 i&o reorg. g(x)w(x), i.e.
• • • Sampled and vert. sca. G(u) *W(u)
-1
-8. -6. -4. -2. 0 2. 4. 6. 8.
x -4
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

R = 16, Δu = 1/16, p.w. = 2a > Δu

(c) a = 1/8

D.7 The effect of the sinc lobes on its convolution with a sharp-edged function

We have seen in the previous sections and in Chapter 6 that the convolution of a sharp-
edged function such as a square pulse with the sinc function R sinc(R x) results in an
overshoot to each side of the discontinuity, followed by a gradually decaying ripple effect.
It is interesting to see what is the contribution of the main lobe of the sinc function and
what is the contribution of its following sidelobes to the shape of the resulting
convolution.

As one can see by making a series of convolutions between a step function and a
gradually truncated sinc function (see, for example, Fig. D.6), the main lobe of the sinc
function only rounds the corners of the step function. The overshoots are obtained when
we add the first sidelobe to each side of the main lobe of the sinc function; and the
decaying ripple effect beyond the overshoots is obtained when we also add the higher
sidelobes of the sinc function. A similar discussion can be found in [Oppenheim99 p.
473] and Fig. 7.23 therein.
 Appendix D: Miscellaneous remarks and derivations 315

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k n
10
1 8
5
0.5
CFT
0 0

• • • DFT after CFT of

-0.5 g(x)=sinc(2ax) i&o reorg. g(x)w(x), i.e.
-5
• • • Sampled and vert. sca. G(u) *W(u)
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

R = 8, Δu = 1/8, p.w. = 2a = Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k n
10
1 8
5
0.5
CFT
0 0
• • • DFT after CFT of
-0.5 g(x)=sinc(2ax) i&o reorg. g(x)w(x), i.e.
-5
• • • Sampled and vert. sca. G(u) *W(u)
-1
-8. -6. -4. -2. 0 2. 4. 6. 8.
x -1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

R = 16, Δu = 1/16, p.w. = 2a > Δu

(d) a = 1/16

D.8 On the order of applying the sampling and truncation operations

We have seen that in order to apply DFT to an original continuous function g(x), we
must first sample g(x) and truncate it into a finite range. In most texts explaining the
relationship between CFT and DFT (see, for example, [Brigham88 pp. 89–97] or
[Gonzalez87 pp. 95–97]) these two operations are performed in the above mentioned
order: First sampling is applied to g(x), and only then a finite range truncation is applied to
the sampled signal. This approach is adopted in the present book, too, as one can see in
Sec. 2.5.

It may be useful to note, however, that the order we apply these two operations, sampling
and truncation, is immaterial, and it is perfectly admissible to first truncate g(x), giving
the truncated, continuous signal g(x)w(x) and its spectral domain counterpart G(u)*W(u),
and only then sample g(x)w(x) (which gives in the spectral domain a convolution
of G(u)*W(u) with an impulse train, i.e. an infinite number of shifted replicas of
 316 Appendix D: Miscellaneous remarks and derivations

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k n
20
1 16
10
0.5 8
CFT
0 0

• • • DFT after CFT of

-0.5 g(x)=sinc(2ax) i&o reorg. g(x)w(x), i.e.
-10
• • • Sampled and vert. sca. G(u) *W(u)
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

R = 8, Δu = 1/8, p.w. = 2a < Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k n
20
1 16
10
0.5
CFT
0 0
• • • DFT after CFT of
-0.5 g(x)=sinc(2ax) i&o reorg. g(x)w(x), i.e.
-10
• • • Sampled and vert. sca. G(u) *W(u)
-1
-8. -6. -4. -2. 0 2. 4. 6. 8.
x -1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

R = 16, Δu = 1/16, p.w. = 2a = Δu

(e) a = 1/32

G(u)*W(u)). This can be formally explained by the fact that both multiplication and
convolution are commutative: in the signal domain we have
[g(x) III(x/Δx)] w(x) = [g(x) w(x)] III(x/Δx) (D.8)
where Δx is the sampling interval, and in the spectral domain we have
[G(u) * Δx III(Δx u)] * W(u) = [G(u) * W(u)] * Δx III(Δx u) (D.9)
where the spectral-domain impulse train Δ x III(Δ x u) is the Fourier transform of the
sampling impulse train III(x/Δx) (see Eqs. (5.1)–(5.2) in Sec. 5.2).

Hence, whichever order we choose for applying the sampling and truncation operations
to the continuous-world signal g(x) will give the same results, and the discrete counterpart
of the resulting continuous-world spectrum (D.9), the DFT spectrum, will also be the
same in both cases. This result is tacitly used, for example, in the beginning of Sec. 6.2.
For more details on the freedom in choosing the order we perform the various operations
in Fig. 2.1, see the cubic commutative diagram in [Kammler07 Sec. 1.4].
 Appendix D: Miscellaneous remarks and derivations 317

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 40 n
32 CFT
1 30 • • • DFT after
i&o reorg.
0.5 20 and vert. sca.
CFT of
0 10
8 g(x)w(x), i.e.
G(u) *W(u)
-0.5 g(x)=sinc(2ax) 0
• • • Sampled
-1
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -10
-2. -1.5 -1. -0.5 0 0.5 1. 1.5 2.
Hz

R = 8, Δu = 1/8, p.w. = 2a < Δu

Signal domain Spectral domain: Real part

0 4 8 12 16 20 24 28 31 0 4 8 12 16 20 24 28 31
1.5 k 40 n
32
CFT
1 30 • • • DFT after
i&o reorg.
0.5 20 and vert. sca.
16
CFT of
0 10 g(x)w(x), i.e.
G(u) *W(u)
-0.5 g(x)=sinc(2ax) 0
• • • Sampled
-1
-8. -6. -4. -2. 0 2. 4. 6. 8.
x -10
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1.
Hz

R = 16, Δu = 1/16, p.w. = 2a < Δu

(f) a = 1/64

D.9 Relation of the DFT to the CFT and to Fourier series

We have seen in Chapter 2 that although the discrete theory of the DFT stands in its
own right, the DFT can be also obtained as a numerical approximation to the CFT (see
Remark 2.2). Let us briefly review this result.

Suppose we are given a non-periodic continuous-world function g(x), which satisfies

g(x) ≈ 0 outside a certain finite range 0...R along the x axis (the case of periodic functions
will be discussed later in this section). We wish to approximate the Fourier transform
G(u) from a sequence of N samples of g(x) taken within this finite range with a sampling
interval of Δx = R/N:

g(kΔx) k = 0, ... ,N–1 (D.10)

With N input samples of g(x) we can only expect to get N independent output values
(frequencies). Therefore, anticipating that the frequency step obtained in the DFT output is
318 Appendix D: Miscellaneous remarks and derivations

Height of the CFT Height of G(u)*W(u)

G(u) at u = 0 and of the DFT at u = 0

1 1
p.w. = 2a > Δu 2a
~ 2a
1 1
p.w. = 2a = Δu 2a
~ 2a (~R)
1
p.w. = 2a < Δu 2a
~R

Table D.1: A summary of the behaviour of the continuous spectrum G(u) =

2a rect(2a ) and of the corresponding DFT spectrum when a→0. As
1 u

long as the continuous pulse width (p.w. = 2a) is ≥ Δ u, the

convolution G(u)*W(u) (Eq. (D.7)), and therefore the DFT, too,
give a good approximation to the function G(u), except for some
fluctuations due to overshoot and ripple effects. But when
p.w. < Δ u, the convolution G(u)* W(u) and the DFT no longer
follow G (u): While at u = 0 G (u) equals 2a1 , G(u) * W(u) and
the DFT tend at u = 0 to R, and consequently they become
R-dependent.

Δu = 1 (Eq. (4.4)), let us try to numerically approximate the Fourier transform G(u) at
NΔx
the following N discrete frequencies (for the sake of simplicity we assume here that N is
even):6
nΔu = n n = – N , ... ,N –1 (D.11)
NΔx 2 2
The reasons for choosing this range of n will be explained soon in Remark D.2. For the
time being, note that the extreme frequencies obtained with this range of n, namely
– 1 and 1 – 1 , correspond exactly to the lowest and highest frequencies that can be
2Δx 2Δx NΔx
obtained when sampling g(x) at N points that are spaced by a sampling interval of Δx (see
Sec. 4.3).

Now, for any value of n within this range we have from the definition of the CFT (Eq.
(2.1)):
∞
G(nΔu) =
∫ –∞
g(x) e–i2π nΔux dx

Because we assumed that g(x) ≈ 0 outside our finite sampling range 0...R, we have:
R
≈
∫ 0
g(x) e–i2π nΔux dx

6 In the vast majority of DFT implementations N is a power of two, or at least an even number. In the rare
cases where N is odd the index range should be adapted accordingly [Briggs95 Sec. 3.1].
Appendix D: Miscellaneous remarks and derivations 319

Given sharp pulse Truncated 4sinc(4x) Their convolution

4 4 4
3 3 3
2 2 2
1 * 1 = 1

-2 -1 1 2 -2 -1 1 2 -2 -1 1 2
-1 -1 -1
(a)
4 4 4
3 3 3
2 2 2
1 * 1 = 1

-2 -1 1 2 -2 -1 1 2 -2 -1 1 2
-1 -1 -1
(b)

4 4 4
3 3 3
2 2 2
1 * 1 = 1

-2 -1 1 2 -2 -1 1 2 -2 -1 1 2
-1 -1 -1
(c)

Figure D.6: Convolution of a sharp pulse (consisting of two step functions) with
a gradually truncated R sinc(R x) function, showing the influence of
the different sinc lobes on the resulting convolution. All parts of the
figure subsist in the signal domain. We have used here the function
4sinc(4x) for the sake of clarity, but higher values of R give the same
results. As shown in row (c), the main lobe of the sinc function only
rounds the corners of each sharp transition of the pulse. Row (b)
shows that the overshoots to both sides of each sharp transition are
obtained when we add next to the main lobe of the sinc function the
first neighbouring sidelobe to each direction. Row (a) shows that the
decaying ripple effect beyond the overshoots is obtained when we
also add the higher sidelobes of the sinc function.

Let us approximate this integral by a discrete sum, using the rectangular rule for
approximating integrals [Press02 p. 508; Ramirez85 p. 67]:7

7 A more detailed and rigorous explanation can be found in [Briggs95, Sec. 2.2], but it is based there on
a different variant of the DFT definition, and on the trapezoidal rule rather than the rectangular rule for
approximating integrals (these rules of numerical integration can be found, for example, in [Harris98 pp.
567–568], [Kreyszig93 Sec. 18.5] or [Spiegel68 p. 95]).
320 Appendix D: Miscellaneous remarks and derivations

N– 1
≈ Δx ∑ g(kΔx) e–i2π nΔukΔx
k=0

Since we are approximating G(u) at discrete frequencies with a frequency step Δu = 1

NΔx
we get:
N– 1
= Δx ∑ g(kΔx) e–i2π kn/N
k=0

But this discrete sum corresponds exactly to the definition of the DFT as given in Eq.
(2.15). We therefore obtain:

= Δx G(nΔu) (D.12)

This means, indeed, that G(nΔu) ≈ Δx G(nΔu) for any integer n in the range between
– N , ... ,N –1, as claimed in Remark 2.2. In other words: The DFT approximates the CFT
2 2
at N discrete frequencies, up to a constant factor of Δx.

Remark D.2: (The choice of the index ranges):

As we can see from Secs. 2.5, D.2 and the figures therein, the DFT can approximate the
CFT G(u) in the spectral domain only within the index range n = – N , ... ,N –1 (compare
2 2
the first and last rows in any of the Figs. 2.1, 2.2 or 2.3). This explains why we have
chosen this specific range for the index n. Note that for any given N, the question of which
N-integer range should be chosen for the index n is immaterial for the DFT definition
itself (because of the cyclical nature of the DFT, as explained in Sec. D.1). But it is
important when we consider the DFT as an approximation to the CFT, in which case the
index range for the frequency samples should be indeed n = – N , ... ,N –1 (as in Remark
2 2
2.2). Nevertheless, if we regard the DFT as a periodic repetition of N-element
replicas, meaning that the index n is considered modulo this range, then any other range of
N consecutive integers may also be used for the index n, including the simple range
n = 0, ... ,N–1 as in Eqs. (2.11)–(2.20). In such cases the resulting N frequency values will
still be the same as before, but their order in the output array of the DFT will be cyclically
shifted. See also Sec. 3.2 about the output data array of the DFT and its reorganization.
As for the range of the signal-domain index k, its choice basically depends on the nature
and the domain of the original continuous-world function g(x), as we can see in Figs. 2.1–
2.3. In the general case where the function g(x) is not symmetric (see Figs. 2.2–2.3) it is
reasonable to choose k = 0, ... ,N–1, as we did in Eq. (D.10). But here, too, due to the
cyclical nature of the DFT input (which is periodic with an N-element period), the
formal numerical approximation above can be done for any range of the index k
consisting of N consecutive integers. For example, in [Briggs95, Chapter 2] the
development is done using the range k = – N + 1, ... ,N , while in [Press02 p. 508] it is done
2 2
using the range k = 0, ... ,N–1. See also Sec. 3.3 about the input data array of the DFT and
its reorganization. Q

Let us now proceed to the case of Fourier series, which occurs if the given continuous-
world function is periodic. Here, too, a similar result can be formulated to express the
Appendix D: Miscellaneous remarks and derivations 321

relation of the DFT to the spectrum of the periodic function, which consists, in this case,
of the Fourier series coefficients of g(x). Let us briefly review this case, too.

Suppose that p(x) is a periodic continuous-world function with period T. Its Fourier
series expansion is given by (see, for example, [Briggs95 p. 33], [Kammler07 pp. 5–6]):
∞
p(x) ~ ∑ cn ei2π nx/T (D.13)
n=–∞
where the n-th Fourier series coefficients cn are:

cn = 1
T T ∫ p(x) e –i2π nx/T dx (D.14)

What is the spectrum of the function p(x)? The Fourier series expansion of p(x) can
serve us here as a link between the original function p(x) in the signal domain and its
Fourier transform P(u) in the spectral domain. This is based on the fact that the Fourier
transform (the spectrum) of the function g(x) = ei2π x/T is G(u) = δ(u – 1 ), namely, an
T
impulse at the frequency f = T1 [Bracewell86 p. 101]. Consequently, the spectrum of
each term in the Fourier series (D.13) consists of a single impulse at the frequency of n/T,
whose amplitude (real or complex) is given by the corresponding Fourier coefficient cn. It
follows therefore (under the appropriate convergence conditions) that the spectrum P(u) of
the periodic function p(x) is an impulse-comb with step f = T1 , whose n-th impulse is
located at the frequency u = nf and has the amplitude cn. The general form of the spectrum
of a periodic function p(x) with period T is, therefore [Papoulis68 p. 107]:
∞ ∞
P(u) = ∑ cn δ(u – n/T) = ∑ cn δ(u – nf) (D.15)
n=–∞ n=–∞
where δ(u) is the impulse symbol. Note that the range of the index n in this sum may be
infinite, depending on the given function p(x).

Now, to find the relation of the DFT to the spectrum of a given periodic function g(x),
we proceed in a similar way as above. First of all, we must limit ourselves once again to a
finite set of only N discrete frequencies; and indeed, for the same reasons as before, we
choose for the index n the range n = – N , ... ,N –1. For any value of n within this range we
2 2
have from the definition of the Fourier series coefficient (Eq. (D.14)):

cn = 1
T T ∫ g(x) e –i2π nx/T dx

As before, we approximate this integral over the range T by a discrete sum, using the
rectangular rule for approximating integrals:8
N– 1
≈ 1 Δx ∑ g(kΔx) e–i2π nkΔx/T
T k=0
Note that in order for this approximation to make sense, our sampling range length,
R = NΔx, should be equal to one period T of our given function. We therefore have:

8 A more detailed and rigorous explanation can be found in [Briggs95, Sec. 2.4], but it is based there on
a different variant of the DFT definition, and on the trapezoidal rule rather than the rectangular rule for
approximating integrals.
322 Appendix D: Miscellaneous remarks and derivations

N– 1
= 1 Δx ∑ g(kΔx) e–i2π kn/N
R k=0
As we can see, this discrete sum corresponds exactly to the definition of the DFT as
given in Eq. (2.15). We therefore obtain:

= Δx G(nΔu) (D.16)
R
Δx
This means, indeed, that cn ≈ G(nΔu) for any integer n in the range – N , ... ,N –1. In
R 2 2
other words: The DFT approximates N Fourier series coefficients of g(x), up to a
Δx
constant factor of (which equals 1 , since R = NΔx).
R N
It is interesting to note at this point that if we take for the discrete approximation of the
integral a range length R = NΔx that is not equal to the period T of the given function, we
cannot expect to obtain a good discrete approximation to cn. And indeed, in terms of the
DFT theory we have seen in Chapter 6, if R ≠ T a leakage error may occur and
considerably deteriorate the approximation (see Remark 6.1).9

In conclusion, we see that the relation between a CFT and its DFT approximation is
given by:
N– 1
G(nΔu) ≈ Δx G(nΔu), where G(nΔu) = ∑ g(kΔx) e–i2π kn/N n = – N , ... ,N –1
k=0 2 2

And similarly, the relation between a Fourier series and its DFT approximation is:
N– 1
Δx
cn ≈ G(nΔu), where G(nΔu) = ∑ g(kΔx) e–i2π kn/N n = – N , ... ,N –1
R k=0 2 2
It is important to note, however, that when using a version of the DFT definition that
includes the constant 1 (as is the case, for example, in [Briggs95]), the scaling factors Δx
N
and Δx that precede the DFT G(nΔu) in the above relations must be multiplied by N (in
R
order to allow for a partial factor of 1 thereof to be taken out and incorporated into the
N
DFT itself without causing any changes). In this case we get the following relations (like
in [Briggs95 pp. 22–23 and 40]):
N– 1
G(nΔu) ≈ R G(nΔu), where G(nΔu) = 1 ∑ g(kΔx) e–i2π kn/N n = – N , ... ,N –1
N k=0 2 2
and
N– 1
cn ≈ G(nΔu), where G(nΔu) = 1 ∑ g(kΔx) e–i2π kn/N n = – N , ... ,N –1
N k=0 2 2

These results are succinctly summarized in Table D.2. Further insights into the imter-
connections between the DFT, the CFT and the Fourier series, and on the ways they
approach each other under various limiting considerations, can be found in [Briggs95 pp.

9 In fact, according to Remark 6.1 leakage does not occur whenever the sampling range length R is an
exact integer multiple of the period T. But if the integer multiple m in question is bigger than 1,
meaning that our sampling range R contains m > 1 full periods of g(x) rather than one, then the
impulses in the N-element output array of the DFT will be spaced by m–1 zeroes, and the number of the
approximated coefficients cn we obtain in the N-element DFT output will be therefore reduced from N to
N/m (see Remark 7.1).
Appendix D: Miscellaneous remarks and derivations 323

Relation between DFT Relation between DFT

and CFT and Fourier series

Using DFT definition Δx

G(nΔu) ≈ Δx G(nΔu) cn ≈ G(nΔu)
without 1 R
N

Using DFT definition

G(nΔu) ≈ R G(nΔu) cn ≈ G(nΔu)
with 1
N

Table D.2: The relations between CFT or Fourier series and their DFT
approximations.

211–212]. Similar interconnections for the IDFT can be found in [Briggs95 pp. 218–
222]. Note, however, that the DFT and IDFT definitions used there are different than ours.

Remark D.3: As suggested by Table D.2, the following relation exists in the continuous
world between the CFT and the Fourier series (where Δu stands for 1/R):
G(nΔu) = R cn (D.17)
Note that this underlying relation holds in each of the two rows of Table D.1; and
indeed, this relation exists in the continuous world in its own right, independently of the
DFT theory. To see this, consider a continuous-world function g(x) that is spatially
limited, meaning that g(x) is zero outside a range of length R. The Fourier transform
evaluated at the frequency u = nΔu = n/R is given by:

G(n/R) =
∫ g(x) e
R
–i2π nx/R dx

On the other hand, consider the periodic function p(x) that is defined as a periodic
extension of g(x) with the period R. The integral for the Fourier coefficients of p(x) on the
interval R is given, according to Eq. (D.14), by:

cn = 1
R R ∫ p(x) e –i2π nx/R dx

We see, therefore, that in the case of a function g(x) that is zero outside the interval R, the
CFT of g(x) and the Fourier series coefficients of p(x), the periodic extension of g(x), are
indeed related by Eq. (D.17) [Brigham88 Sec. 5.2; Briggs95 pp. 40–41, 197; Kammler07
p. 179]. If outside the interval R the original function g(x) only satisfies g(x) ≈ 0 rather
than g(x) = 0, then relation (D.17) turns into an approximation:
G(nΔu) ≈ R cn (D.18)
Note that Eq. (D.17) concerns the relation between the Fourier coefficients cn of a
periodic function p(x) with period R and the Fourier transform G(u) of the spatially-
324 Appendix D: Miscellaneous remarks and derivations

limited function g(x) that consists of a single period of p(x) about the origin. The relation
between the Fourier coefficients cn of the periodic function p(x) and the Fourier transform
P(u) of p(x) itself is given by Eq. (D.15); in this case no scaling factor is present, and we
have, indeed:
P(nΔu) = cn (D.19)
as explained in the paragraph preceding Eq. (D.15). Q

Example D.1: To take a concrete example, consider the triangular function tri(x), which
extends between –1 and 1; its CFT is sinc2(u) [Bracewell86 p. 415]. According to the
dilation theorem (see point 2 in Sec. 2.4.1), its two-fold narrower counterpart g(x) =
tri(2x), which extends between –12 and 12 , has the CFT G(u) = 12 sinc2(12 u). Now, as shown in
Fig. D.7, the periodic triangular wave p(x) = triangwave(2x) that is defined as a periodic
extension of g(x) with the period R = 2, can be seen as a convolution of the function g(x) =
tri(2x) with the infinite impulse train t(x) = 12 III(12 x) having period P = 2 and amplitude 1.
The CFT of t(x) is the impulse train T(u) = III(2u), which has impulse intervals of f =
1/P = 12 and amplitude 12 [Bracewell86 p. 414]. Therefore, the CFT P(u) of our periodic
triangular wave p(x) = triangwave(2x) consists of the product of G(u) = 12 sinc2(12 u) and
T(u) = III(2u), which is an infinite impulse train that samples the “envelope” 14 sinc2(12 u) at
impulse intervals of f = 12 (see, for example, [Bracewell86 p. 213]). And indeed, we see that
the impulse amplitudes in the spectrum of p(x), namely, its Fourier series coefficients cn,
are identical to the corresponding values in the spectrum of g(x) up to a scaling factor of
R = 2, in accordance with Eq. (D.17). Q

Eq. (D.17) says, in fact, that it is possible to compute the Fourier series coefficients cn of
p(x) using the CFT of g(x), i.e. by calculating G(n/R) [Kammler07 p. 179]. In this case,
the result we obtain is not cn but rather Rcn: each impulse in the spectrum is multiplied by
R (which is, in fact, the period of the periodic function p(x) whose Fourier series we are
computing). This fact sheds, indeed, a new light on our results in Appendix A.3, where we
have seen that the impulse amplitudes in the DFT spectrum (after the standard amplitude
scaling correction) are R times higher than the true impulse amplitudes c n in the
continuous-world spectrum.

Stating it in other words, remember that the standard amplitude scaling correction that
we should always apply to the DFT output according to Sec. 4.4 includes a multiplication
by Δ x. This is, indeed, in accordance with our result G(n Δ u) ≈ Δ x G(n Δ u) from
Eq. (D.12). This multiplication by Δ x is to be done in all cases, including when the
spectrum G(u) contains impulses.10 We therefore must realize, by virtue of our result
Δx
c n ≈ G(nΔu) (see Eq. (D.16)), that if the spectrum G(u) includes impulses, then their
R
amplitudes (which are, according to Eq. (D.15), the Fourier series coefficients cn) will
appear in the DFT output, after the standard amplitude scaling correction, R times higher
than their true amplitudes in the continuous-world spectrum G(u).

10As shown, for example, in Fig. A.4, in certain cases the spectrum G(u) may be hybrid and contain
symultaneously both continuous and impulsive components; the multiplication of the DFT output
values by Δx will affect both the impulsive and the non-impulsive parts of the DFT spectrum.
Appendix D: Miscellaneous remarks and derivations 325

Signal domain Spectral domain

g(x) = tri(2x) G(u) = 12 sinc2(u/2)

1 1
2

x u
–1 0 1 –2 0 2
2 2
(a)

t(x) = 12 III(x/2) T(u) = III(2u)

1 1
2

x u
–2 0 2 –2 –1 0 1 2
(b)

p(x) = g(x)*t(x) = triangwave(2x) P(u) = G(u) T(u)

1
1
4
x u
–2 –1 0 1
1 2 –2 –1 0 1 2
2 2
(c)

Figure D.7: A concrete continuous-world example illustrating Remark D.3. (a) The
continuous function g(x) = tri(2x), which extends in the range between –12
and 12 , and its CFT G(u) = 12 sinc2(12 u). (b) The impulse train t(x) = 12 III(12 x),
having period P = 2 and amplitude 1, and its CFT T(u) = III(2u), which
is an impulse train with impulse intervals of f = 1/P = 12 and amplitude 12 .
(c) The periodic triangular wave p(x) = triangwave(2x) is a convolution
of the signals g(x) and t(x) of rows (a) and (b), and therefore its CFT
P(u) is the product of the CFTs G(u) and T(u) of rows (a) and (b).

D.9.1 Examples illustrating the relation of the DFT to the CFT and to Fourier series

Let us now consider a few examples that will help us better understand the relationship
between the DFT of the function g(x) and the DFT of its periodic extension p(x). These
examples illustrate, in particular, the various amplitude corrections that must be applied to
the DFT output in order that it may match the corresponding real-world spectra (assuming
that no DFT errors such as aliasing or leakage are present).

Fig. D.8 shows the continuous function g(x) = tri(2x) and its CFT G(u) = 12 sinc2(12 u), as
well as their discrete counterparts that are obtained by sampling g(x) and applying DFT
(after applying the corrections of Chapters 3 and 4). All the rows of the figure show the
326 Appendix D: Miscellaneous remarks and derivations

same functions g(x) and G(u), although the sampling range length R and the sampling step
Δx in each row are different. The resulting DFT matches the original spectrum G(u) in all
of the rows, thanks to the standard amplitude scaling correction of Chapter 4, which
includes a multiplication of the DFT values by Δx, in accordance with Eq. (D.12).

It is interesting to note that the very same N-element input arrays as in Figs. D.8(a)–(c)
can be also interpreted in a different way, as shown in Figs. D.9(a)–(c). Fig. D.9(a) is
identical to Fig. D.8(a), and shows the same triangular signal g(x) within a range of length
R = 8. Fig. D.9(b), although it shows exactly the same DFT input array as Fig. D.8(b), is
interpreted here as a two-fold dilated version of g(x), namely g(x/2), within the same range
length of R = 8. Similarly, Fig. D.9(c), although it shows exactly the same DFT input
array as Fig. D.8(c), is interpreted here as a 4-fold dilated version of g(x), namely g(x/4),
within the same range length of R = 8. As we can see in Fig. D.9, the continuous-world
spectra obtained in the three rows are clearly different from their counterparts in Fig. D.8;
this happens due to the dilation theorem of the CFT which states that g(x/a) has the CFT
|a|G(au) (see point 2 in Sec. 2.4.1). This theorem is not applied in Fig. D.8 since all of
the rows there show exactly the same original function g(x), and no dilation is involved.

Let us now consider the DFT results in both figures. Given that the input arrays in rows
(a)–(c) of Fig. D.8 are identical to their counterparts in Fig. D.9, how can the DFT results
match the CFT results in both figures, hence giving different results for identical inputs?
This is taken care of by the standard amplitude scaling correction of Chapter 4 (which
includes the multiplication of the DFT output values by the sampling step Δx). Consider,
for example, row (b) in Figs. D.8 and D.9. Because the input arrays forwarded to the DFT
are identical in both cases, it is obvious that the output arrays obtained by the DFT are also
identical. However, in Fig. D.8(b) this raw data is then multiplied by Δx = 4/64 = 1/16,
while in Fig. D.9(b) the same output data is rather multiplied by Δx = 8/64 = 1/8. Hence
the final DFT output amplitudes in Fig. D.9(b) are twice higher than in Fig. D.8(b), so
that in both cases the corrected DFT results indeed match the respective CFTs.

Consider now Fig. D.10, which shows the periodic counterpart of Fig. D.8. The
periodic function p(x) = triangwave(2x) is the periodic extension of g(x) = tri(2x) with the
period P = 2, and its CFT P(u) is, as we have seen in Example D.1 above, an infinite
impulse train that samples the “envelope” 12 G(u) = 14 sinc2(12 u) at impulse intervals of f = 12 .
Just as in Fig. D.8, all the rows of our figure show the same functions p(x) and P(u),
although the sampling range length R and the sampling step Δx in each row are different.
The resulting DFT matches the original spectrum P(u) in all of the rows, thanks to the
standard amplitude scaling correction of Chapter 4 (which includes a multiplication of the
DFT results by Δx) and the impulse amplitude correction of Appendix A (the division of
the DFT impulse amplitudes by R), in accordance with Eq. (D.16).

Remark D.4: Note that the input arrays that are forwarded to the DFT in Fig. D.8(c) and
in Fig. D.10(c) are strictly identical. The difference between the two figures is only in the
interpretation given to this data: While in Fig. D.8(c) the input data is considered to be a
sampled version of the aperiodic triangular function g(x) = tri(2x), in Fig. D.10(c) the
Appendix D: Miscellaneous remarks and derivations 327

same input data is considered to be a sampled version of a single period of the periodic
function p(x) = triangwave(2x), the periodic extension of g(x) with period P = 2.
Consequently, the very same raw DFT output that is obtained in the two cases is
interpreted differently in each case: in Fig. D.8(c) it is considered as an approximation to
the continuous CFT of g(x), G(u) = 12 sinc2(12 u), while in Fig. D.10(c) it is interpreted as an
approximation to the Fourier series of p(x), which is an impulse train that samples the
“envelope” 12 G(u) = 14 sinc2(12 u) at impulse intervals of f = 12 . Thus, although the raw output
data that is obtained by the DFT in both cases is identical, in the first case it will only
undergo the standard amplitude scaling correction of Chapter 4 (multiplication by Δx),
while in the second case it will also undergo the impulse amplitude correction of Appendix
A (division of the impulse amplitudes by R = 2). Note that since this output array looks
“continuous”, unlike its clearly impulsive counterparts in Figs. D.10(a),(b), its
interpretation as a continuous CFT or as a discrete impulse train (Fourier series) only
depends on whether we consider the input array as a sampled version of g(x) or as a
sampled version of its periodic extension p(x). In the second case, we must remember to
apply the impulse amplitude correction of Appendix A (division by R), in accordance with
Eq. (D.17). Q

Just as in the case of Fig. D.8, it is interesting to note that in the periodic case, too, the
very same N-element input arrays of Fig. D.10 can be also interpreted in a different way.
This is shown in Figs. D.11(a)–(c). Fig. D.11(a) is identical to Fig. D.10(a), and shows
the same periodic triangular signal p(x) within a range of length R = 8, covering 4 periods
of p(x). Fig. D.11(b), although it shows exactly the same DFT input and output arrays as
Fig. D.10(b), is interpreted here as a two-fold dilated version of p(x), namely p(x/2), within
the same range length of R = 8 (which covers now only two periods of p(x/2)). Similarly,
Fig. D.11(c), although it shows exactly the same discrete data as Fig. D.10(c), is
interpreted here as a four-fold dilated version of p(x), namely p(x/4), within the same range
length of R = 8 (which covers this time only one period of p(x/4)). Note that in the case of
Fourier series, the dilation theorem (i.e. the spectral-domain behaviour when the original
periodic function undergoes a dilation) is not the same as in the case of CFT (point 2 in
Sec. 2.4.1), and it does not involve an amplitude change in the Fourier series coefficients
[Kammler07 p. 187]. In the discrete DFT spectra of Fig. D.11 this is taken care of by the
amplitude corrections of Chapter 4 (the multiplication by Δx) and of Appendix A (the
division by R), since in all rows of Fig. D.11 both Δx and R remain unchanged. Note that
in Fig. D.10, both Δx and R vary in each row; but since they vary proportionally (because
R = NΔx, and N remains unchanged in all rows), the total amplitude correction (multi-
plication by Δx and division by R) remains the same in the three rows, just as in Fig. D.11.

Of course, if g(x) is not periodic, then the height of its CFT does depend on the dilation
factor of the given function, as shown in Fig. D.9. Furthermore, because the DFT
spectrum is no longer impulsive, the impulse amplitude correction of Appendix A (the
division of the impulse amplitudes in the DFT spectrum by R) should not be applied, and
therefore the amplitude correction of the DFT output depends on Δx rather than on the
ratio Δx/R as in the periodic case.
 328 Appendix D: Miscellaneous remarks and derivations

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 1.5 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5
-0.5 • • • DFT after
-1 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, R = 8

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 1.5 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5 • • • DFT after
-0.5
-1 i&o reorg.
and vert. sca.
-1.5
-2. -1. 0 1. 2.
x -1
-8. -6. -4. -2. 0 2. 4. 6. 8.
Hz
-1.5 -0.5 0.5 1.5
(b) N = 64, R = 4

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 k 1.5
0 8 16 24 32 40 48 56 63 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5 • • • DFT after
-0.5
-1 i&o reorg.
x and vert. sca. Hz
-1.5 -1
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1. -16. -12. -8. -4. 0 4. 8. 12. 16.
(c) N = 64, R = 2

Figure D.8: The continuous function g(x) = tri(2x) and its CFT G(u) = 12 sinc2(12 u),
represented by continuous lines, and their discrete counterparts that are
obtained by sampling g(x) and applying DFT (including the corrections
of Chapters 3 and 4), represented by black dots. All the rows of the
figure show the same functions g(x) and G(u), although the sampling
range length R and the sampling step Δx in each row are different (note
that N = 64 in all cases): (a) R = 8, Δx = 8/64 = 1/8; (b) R = 4, Δx = 4/64
= 1/16; (c) R = 2, Δx = 2/64 = 1/32.
 Appendix D: Miscellaneous remarks and derivations 329

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 1.5 n
2
g(x)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5
-0.5 • • • DFT after
-1 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, R = 8

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 1.5 n
2
g(x/2)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5 • • • DFT after
-0.5
-1 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -1
-4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(b) N = 64, R = 8

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 k 1.5
0 8 16 24 32 40 48 56 63 n
2
g(x/4)
1.5 • • • Sampled 1
1
0.5
0.5
0 0
CFT
-0.5 • • • DFT after
-0.5
-1 i&o reorg.
x and vert. sca. Hz
-1.5 -1
-4. -3. -2. -1. 0 1. 2. 3. 4. -4. -3. -2. -1. 0 1. 2. 3. 4.
(c) N = 64, R = 8

Figure D.9: A different interpretation of the very same discrete input arrays as in Fig.
D.8, where the sampling range length and the sampling step Δx remain
identical in the three rows (R = 8, Δx = 8/64 = 1/8), but each row represents
a differently dilated version of g(x), and its corresponding CFT: (a) g(x) =
tri(2x); G(u) = 12 sinc2(12 u). (b) g(x/2) = tri(x); 2G(2u) = sinc2(u). (c) g(x/4) =
tri(x/2); 4G(4u) = 2sinc2(2u). Although the discrete input array in each row
is identical to its counterpart in Fig. D.8, the resulting discrete spectra do
match the different CFTs thanks to the standard amplitude scaling
correction of Chapter 4, which includes a multiplication by Δx.
 330 Appendix D: Miscellaneous remarks and derivations

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 0.4 n
2
p(x)
0.3 0
1.5 • • • Sampled 1
1 0.2
2
0.5 0.1 3
4 5 6 7
0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, R = 8

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 0.4 n
2
p(x)
0.3
1.5 • • • Sampled
1 0.2

0.5 0.1

0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
and vert. sca.
-1.5
-2. -1. 0 1. 2.
x -8. -6. -4. -2. 0 2. 4. 6. 8.
Hz
-1.5 -0.5 0.5 1.5
(b) N = 64, R = 4

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 k 0.4
0 8 16 24 32 40 48 56 63 n
2
p(x)
0.3
1.5 • • • Sampled
1 0.2

0.5 0.1

0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
x and vert. sca. Hz
-1.5
-1. -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1. -16. -12. -8. -4. 0 4. 8. 12. 16.
(c) N = 64, R = 2

Figure D.10: The continuous periodic function p(x) = triangwave(2x) (a periodic

extension of g(x) = tri(2x) of Fig. D.8) and its impulsive CFT P(u),
represented by continuous lines, and their discrete counterparts that are
obtained by sampling p(x) and applying DFT (including the corrections
of Chapters 3, 4 and Appendix A), represented by black dots. All the
rows of the figure show the same functions p(x) and P(u), although the
sampling range length R and the sampling step Δ x in each row are
different (note that N = 64 in all cases): (a) R = 8, Δ x = 8/64 = 1/8;
(b) R = 4, Δx = 4/64 = 1/16; (c) R = 2, Δx = 2/64 = 1/32.
 Appendix D: Miscellaneous remarks and derivations 331

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 0.4 n
2
p(x)
0.3 0
1.5 • • • Sampled 1
1 0.2
2
0.5 0.1 3
4 5 6 7
0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(a) N = 64, R = 8

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 0 8 16 24 32 40 48 56 63
k 0.4 n
2
p(x/2)
0.3
1.5 • • • Sampled
1 0.2

0.5 0.1

0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
and vert. sca.
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4.
x -4. -3. -2. -1. 0 1. 2. 3. 4.
Hz

(b) N = 64, R = 8

Signal domain Spectral domain: Real part

0 8 16 24 32 40 48 56 63 k 0.4
0 8 16 24 32 40 48 56 63 n
2
p(x/4)
0.3
1.5 • • • Sampled
1 0.2

0.5 0.1

0 0
CFT
-0.5 -0.1
• • • DFT after
-1 -0.2 i&o reorg.
x and vert. sca. Hz
-1.5
-4. -3. -2. -1. 0 1. 2. 3. 4. -4. -3. -2. -1. 0 1. 2. 3. 4.
(c) N = 64, R = 8

Figure D.11: A different interpretation of the very same input and output DFT arrays
as in Fig. D.10, where the sampling range length remains identical in the
three rows (R = 8) but each row represents a differently dilated version
of p(x), and its corresponding impulsive CFT: (a) p(x) = triangwave(2x);
(b) p(x/2) = triangwave(x); (c) p(x/4) = triangwave(x/2). Note that unlike
in Figs. D.8–D.9, all the spectra in Figs. D.10–D.11 have the same
amplitude.
332 Appendix D: Miscellaneous remarks and derivations

Given signal: g(x) Given signal: p(x)

(not periodic) (a single period)

Fig. D.8(c): DFT Fig. D.10(c): DFT

Not dilated;
output multiplied by output multiplied by
R = 2, Δx = 1/32
Δx Δx/R

Fig. D.9(c): CFT scaled Fig. D.11(c): CFT not

Dilated by a = 4;
by a; DFT output scaled; DFT output
R = 8, Δx = 1/8
multiplied by Δx multiplied by Δx/R

Table D.3: The four different interpretations of the same input data in row (c)
of Figs. D.8 – D.11, and the postprocessing treatment that should
be applied to the DFT output in each case.

Finally, as a further extension of Remark D.4, note that row (c) in all of the four figures
D.8–D.11 contains exactly the same discrete input data, but in each case it is interpreted in
a different way, and therefore it undergoes a different postprocessing treatment in order to
match the corresponding CFT. This is succinctly summarized in Table D.3.

D.10 The 2D spectrum of a rotated bar passing through the origin

In Example 5.6 (see Sec. 5.5) we discuss the 2D DFT of the continuous-world function
g(x,y) consisting of an infinite bar of height 1 that is centered about the origin and rotated
by angle θ. Let us show here that the 2D CFT of g(x,y), G(u,v), consists of a continuous
straight line-impulse (blade) passing through the origin at the perpendicular direction, and
whose amplitude is modulated by a sinc function.

We start with the simple case where θ = 0. The infinite vertical bar function extending
along the y axis is expressed by the 2D function g(x,y) = rect(x). Note that the function
rect(x) over the x,y plane is independent of y, and therefore g(x,y) remains constant along
the y direction. Similarly, the function sinc(u) over the u,v plane is independent of v, and
therefore it remains constant along the v direction; therefore, in order to obtain a sinc-
shaped blade (line impulse) along the u axis we have to multiply sinc(u) by δ(v).

Now, because G(u,v) = sinc(u)δ(v) is a product of a function of u and a function of v

(both of which being considered throughout the u,v plane), we can apply the separable
product theorem (see point 10 in Sec. 2.4.2 or [Bracewell95 p. 166]): Knowing that
Appendix D: Miscellaneous remarks and derivations 333

Signal domain Spectral domain

g(x) = rect(x/a) G(u) = a sinc(au)

a
1

x u
a a 0
–
2 0 2
(a)

g(x,y) = rect(x/a)
G(u,v) = a sinc(au)δ(v)
y
1 v
a Blade

a a x u
–
2
0 2 0

(b)

Figure D.12: (a) The 1D Fourier transform of the 1D function g(x) = rect(x/a) is
G(u) = a sinc(a u). (b) Therefore, the 2D Fourier transform of the
straight bar g(x,y) = rect(x/a) along the y axis of the x,y plane is a
vertical sinc-shaped blade (line impulse) G(u,v) = a sinc(a u)δ(v)
extending along the u axis of the u,v plane.

rect(x) ↔ sinc(u)

and 1 ↔ δ(v)

are 1D Fourier transform pairs [Bracewell86 pp. 412, 415], we obtain that

g(x,y) = rect(x) ↔ G(u,v) = sinc(u)δ(v)

is a 2D Fourier transform pair. More generally, for any 2D function independent of y

g(x,y) = g1(x) we have G(u,v) = G1(u)δ(v) [Bracewell95 p. 166; Bracewell86 p. 253]. For
example, as shown in Fig. D.12, if g(x,y) = rect(x/a) then G(u,v) = a sinc(a u)δ(v).
334 Appendix D: Miscellaneous remarks and derivations

Finally, using the rotation theorem (see point 11 in Sec. 2.4.2 or [Bracewell95 p. 157])
we see that if we rotate the vertical bar of Fig. D.12 by angle θ, its Fourier transform will
also undergo a rotation by the same angle θ. Note that the rotated blade in the spectrum
always remains perpendicular to the rotated bar in the signal domain.
 Appendix E
Glossary of the main terms

E.1 About the glossary

Because the DFT has a very wide range of applications in many different fields of
science and technology (see end of Sec. 1.2), it is not surprising that the terminology used
in the vast DFT literature is far from being consistent and uniform. In many cases
different authors use different terms for the same entities, and what is even worse, the
same terms are often used in different meanings by different authors. As a few examples
among many others, let us mention here the various Fourier theorems (or “rules”; see
Sec. 2.4), whose names greatly vary between different sources. For instance, the inversion
rule [Kammler07 p. 199] is called symmetry rule in [Brigham88 p. 107], while in
[Bracewell86 pp. 364–365] it is called reciprocity and the term symmetry is reserved to
properties related to oddness and evenness [Bracewell86 p. 366]. Similarly, the reflection
rule [Kammler07 p. 199] is called in [Bracewell86 p. 366] reversal while in
[Nussbaumer82 p. 82] the term being used is symmetry (again!). Even the aliasing artifact
is often called foldover, and the leakage artifact is sometimes called ringing [Brigham88
pp. 94, 106].
Obviously, in such an interdisciplinary domain as the Fourier theory it would be quite
impossible to adopt a universally acceptable standardization of the terms, because of the
different needs and traditions in the various fields involved.1 Nevertheless, even without
having such far-reaching pretensions, we were obliged to make our own terminological
choices in a systematic and coherent way, in order to avoid confusion and ambiguity in
our own work. We tried to be consistent in our terminology throughout this work, even if
it forced us to assign to some terms a somewhat different meaning than one would expect
(depending on his own background, of course). A typical example is the systematic termi-
nology we adopted for the different classes of continuous and discrete signals. Here, too,
the terms in the literature largely vary between the different sources (compare for instance
[Oppenheim99 pp. 8–9], [Bracewell95 p. 80], [Smith03 p. 11], [Mitra93 pp. 57–60],
[Gomes97 pp. 16, 112], [PhysicsForums12], [Goodman05 p. 284], [Zayed93 p. 4], etc.).
We included in the present glossary all the terms for which we felt a clear definition was
desirable to avoid any risk of ambiguity. Note, however, that this glossary is not ordered
1 For example, the term density has very different meanings in almost any imaginable field of science or
technology. A non-exhaustive list, just for the sake of illustration, may include: density of matter (in
physics), density of population (in statistics), probability density (in probability), spectral density (in
spectral analysis), density of a set (in mathematic topology), ink or colour density (in printing and
colorimetry), etc. Another similar example is the term phase. It should be also noted that terminology
sometimes tends to change over the years or according to fashions. As an example, the term function
convolution in modern literature appears in older publications as function composition [Zygmund68
p. 36], the resultant of two functions, or even using the German term Faltung [Hardy68 p. 10].

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 335
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8_13,
© Springer-Verlag London 2013
336 Appendix E: Glossary of the main terms

alphabetically; rather, we preferred to group the various terms according to subjects. We

hope this should help the reader not only to clearly see the meaning of each individual
term by itself, but also to put it in relation with other closely related terms (which would be
completely dispersed throughout the glossary if an alphabetical order were preferred).
Note that terms in the glossary can be found alphabetically through the general index at
the end of the book.

E.2 Terms in the signal domain

signal —
A function of one or more variables such as time, distance, 2D or MD space, etc.,
which represents a certain phenomenon of interest (physical, electrical,
astronomical, statistical, etc.). For example, the temperature measured at a given
point as function of time (either continuous or discrete) is a 1D signal; and any
gray-level image or even colour image is a 2D signal. More precisely, a gray-level
image is a scalar-valued signal of two variables, z = g(x,y), while a colour image is
a vector-valued signal of two variables, for example (r,g,b) = g(x,y). Note that we
often use the terms “signal” and “function” interchangeably as synonyms.

digitization (or digitizing) —

The process of converting a continuous signal (such as voltage, sound, a 2D
analog image, etc.) into digital form (see, for example, [Russ95 p. 16]). This
process consists of two steps, which can be performed in any order: discretization
(sampling), and quantization. Note that each of these steps involves some loss of
information. The resulting digital signal is therefore said to be a digital
approximation (or a digital version) of the underlying continuous signal. It should
be noted, however, that in some references (such as [Jähne95 pp. 19, 40–41]) the
term digitization is used in the sense we reserve here to the term discretization.

discretization (or sampling) —

The process of recording the values of a continuous signal (such as voltage, sound,
a 2D analog image, etc.) at a given finite or countable set of points, usually with
regular intervals, through the signal’s domain of definition (for example, along the
time axis, within space, etc.). Each single recording is called a sample. For
example, in order to display an analog image on a monitor, the image must be first
discretized (scanned or sampled) into a matrix of M ×N pixels (samples).
Discretization affects the domain of the given signal, and converts it from a
continuous set into a discrete set. Note that discretization involves some loss of
information. A sampled signal is therefore said to be a discrete approximation (or
a discrete version) of the underlying continuous signal.

quantization —
The process of rounding off the values of a continuous signal (or of the samples
of a continuous signal) into a finite or countable set of predefined values, usually
Appendix E: Glossary of the main terms 337

with regular intervals (such as integer numbers, decimals, or any other values
resulting from the finite precision of the measuring or computing equipment being
used, for instance an 8-bit sensor or a 16-bit processor). As a concrete example,
gray-level images that are displayed on a monitor are usually quantized into 256
gray levels, so that the value of the image at each of its discrete elements (pixels)
can be stored in one byte (8 bits). Quantization affects the range of the given
signal, and converts it from a continuous set into a discrete set. Note that
quantization involves some loss of information. A quantized signal is therefore
said to be a quantized approximation of the underlying continuous signal.

continuous signal 2 (or function; not to be confused with analog signal) —

Unfortunately, this term is used in the literature in two different but equally useful
and legitimate meanings (see, for example, [Gomes97 pp. 16, 109, 110] or
[Glassner95 pp. 128–129]). In mathematical branches such as analysis or
topology, a continuous function is a continuous-valued function of continuous
variable(s); the value of such a function varies continuously (i.e. without
interruptions) when its independent variable(s) vary continuously. However, in
digital signal processing (and in our context, too) the same term is being used in a
much wider sense: A signal is said to be continuous if it has a continuous domain
and a continuous range; but this does not yet imply that the signal has no
interruptions. In other words, all we require from a signal in order to be continuous
in this wider sense is that its variables as well as its amplitude be defined on
continuous sets, rather than being limited to a certain number of discrete values. In
most cases the intended meaning of the term should be understood from the
context. But when a clear distinction between these two meanings is desired, one
can use for the less restrictive meaning the longer but more precise term
continuous-domain continuous-range signal, or alternatively continuous-world
signal. Note that when referring to the restrictive meaning, we often say so
explicitly.

continuous-domain signal (or function; also called continuous-time signal) —

A signal having a continuous domain. Such signals are often called in the literature
continuous-time signals (see, for example, [Oppenheim99 p. 8] or [Glassner95
Sec. 4.2.1]); but this term may be misleading in 2D or MD signals, or even in 1D
signals whose independent variable is not time but any other entity (such as
distance, etc.).

continuous-range signal (or function; also called continuous-amplitude signal or

continuous-valued signal) —
A signal having a continuous range.

2 According to another convention, which is not compatible with our terminology, the term discrete
implies a discrete domain; the term continuous implies a continuous domain; digital implies a discrete
range; and analog implies a continuous range. In this convention, all four combinations are possible: a
signal can be discrete and digital; discrete and analog; continuous and digital; or continuous and analog
[PhysicsForums12].
338 Appendix E: Glossary of the main terms

digital signal 2 (or function) —

A discrete-valued signal of discrete variable(s). A digital signal has a discrete
domain and a discrete range. A continuous signal that undergoes digitization, i.e.
both discretization (sampling) and quantization, is a digital signal (see, for
example, [Oppenheim99 pp. 8–9] or [Gomes97 p. 16]). Note that after
discretization and quantization a digital signal may also undergo a third stage of
encoding, in which it is converted into a certain encoded form such as a binary
signal (a sequence of zeroes and ones) or any other encoded form that may be
needed for its processing [Hayes12 pp. 101, 106; Gomes97 pp. 13, 133–137].

discrete-domain signal (or function; also called discrete-time signal or simply

discrete signal 2) —
A signal having a discrete domain. Such signals are often called in the literature
discrete-time signals (see, for example, [Oppenheim99 p. 8] or [Glassner95 Sec.
4.2.2]); but this term may be misleading in 2D or MD signals, or even in 1D
signals whose independent variable is not time but any other entity (such as
distance, etc.). While many discrete-domain signals are obtained by discretization
(sampling) of original continuous signals, not all discrete-domain signals are
obtained in this manner. Trivial examples are the number of cars sold per day, the
annual number of births per year in a given country, etc.

discrete-range signal (or function; also called discrete-amplitude signal, discrete-valued

signal or quantized signal) —
A signal having a discrete range. A continuous signal that undergoes quantization
is a discrete-range signal.

continuous-domain discrete-range signal (or function) —

A signal having a continuous domain but a discrete range. A continuous signal that
undergoes quantization but not discretization (sampling) is a continuous-domain
discrete-range signal. For example, a staircase function having a finite or countable
number of steps (see, for example, Fig. 8.12(c)) is a continuous-domain discrete-
range function.

discrete-domain continuous-range signal (or function) —

A signal having a discrete domain but a continuous range. A continuous signal that
undergoes discretization (sampling) but not quantization is a discrete-domain
continuous-range signal. Note that most of the mathematical developments in the
field of digital signal processing, including the DFT theory, deal in fact with this
class of signals, although in practice such signals are represented in digital
computers by their quantized approximations, i.e. by digital signals (see, for
example, [Mitra93 p. 58]).

discontinuous signal (or function) —

A signal having one or more discontinuities (interruptions). Note that any signal
having a discrete domain or a discrete range is necessarily discontinuous, but a
Appendix E: Glossary of the main terms 339

discontinuous signal does not necessarily have a discrete domain or a discrete

range. For example, the signal:
⎧x x≤0
g(x) = ⎨
⎩x+2 x>0
is discontinuous, but it is neither a discrete-domain signal nor a discrete-range
signal. This is, indeed, an example of a signal which is neither strictly continuous
(in the topological sense) nor discrete or digital; but since it still subsists in the
continuous world (i.e., it is defined on a continuum and its range is basically a
continuum), we can say it is a continuous-world signal. Note that signals such as
g(x), staircase functions, and histograms (see, for example, Fig. 8.12(c)) are
discontinuous even if they can be graphically drawn by a continuous, uninterrupted
line, i.e. without raising and shifting the pen. The reason is that instantaneous
vertical jumps are always considered as discontinuities, since by definition a
function cannot have more than one value (scalar or vectorial) at any given point.
Hence, the function contains only one single point along the vertical jump, even if
the jump is represented graphically by a vertical line.

continuous-world signal (or function) —

A signal which subsists in the continuous world, but still may have one or more
discontinuities. For example, the signal g(x) given above is neither strictly
continuous (in the topological sense) nor discrete or digital, but since it still
subsists in the continuous world (i.e., it is defined on a continuum and its range is
basically a continuum), we say it is a continuous-world signal. This term is a
convenient softening of the term continuous signal (in the strict topological sense),
since it still allows for the existence of some discontinuities, as often happens in
the real world. This term can be used as a synonym to the term continuous signal
(in the wider sense), whenever we wish to avoid confusion with the strict
(topological) sense of the term. Note that a continuous-world signal may be
discrete; for instance, there exist continuous-world signals that are purely
impulsive (see, for example, remark (i) in Sec. 2.5).

discrete-world signal (or function) —

A signal which subsists in the discrete world, as opposed to a continuous-world
signal. Theoretically, a discrete-world signal is a discrete-domain function (i.e. a
function having a discrete domain); but when processed by computer, it is in
practice a digital function (since its range, too, is discrete). Note that a purely
impulsive signal may be considered, depending on its nature, as a discrete-world
signal or as a continuous-world signal (see, for example, remark (i) in Sec. 2.5): If
the signal is undefined between its impulse locations, its domain is discrete and the
signal is considered as a discrete-world signal. But if the signal takes the value
zero throughout the intervals between its impulse locations, then its domain is
continuous, and the signal is considered as a continuous-world signal. Similarly, a
continuous-domain discrete-range signal such as a staircase fonction (see, for
340 Appendix E: Glossary of the main terms

example, Fig. 8.12(c)) is considered as a continuous-world signal, again, because

its domain is continuous.

real-world signal (or entity) —

A signal or entity that is realizable in the real world (unlike conceptual or
theoretical entities such as signals based on Cantor sets [Weisstein99 p. 187],
Penrose stairs [Weisstein99 p. 1335], etc.). This definition is admittedly vague,
just like that of the term physically realizable (as explained, for example, in
[Bracewell86 p. 271]).

analog signal 2 (or function; not to be confused with continuous signal) —

A signal representing some real-world (physical, electrical, etc.) data that varies
with respect to some variable such as time, distance, height, spatial location, etc.
Examples are the temperature measured at a given location as function of time; the
barometric pressure measured at a given time as function of the altitude; the
temperature at a given time as function of the location (within a 2D plane or a 3D
volume); etc. Theoretically, any real-world signal is analog, including digital
signals (since digital signals, too, are actually represented by a varying voltage or
current, and they are only interpreted as digital signals). From this point of view, it
can be said that a signal is always analog, but the information it carries may be
interpreted, depending on the case, as continuous or digital (in the former case
small fluctuations in the signal are meaningful, whereas in the latter they are not
taken into consideration).3 The term “analog signal” originally comes from the
fact that such a signal is analogous to the physical phenomenon it comes to
express; for example, a signal representing barometric pressure varies up and
down (in terms of voltage, current, etc.) in analogy with the original pressure it
comes to represent.

period (or repetition-period of a function p) —

A number T ≠ 0 such that for any x ∈R, p(x +T) = p(x). In the case of a 2-fold
periodic function p(x,y), a double period (or period parallelogram) of p(x,y) is any
parallelogram A which tiles the x,y plane so that p(x,y) repeats itself identically on
any of these tiles (see also Sec. A.3.4 in Appendix A of [Amidror09]). If the
function p is discrete, R is replaced by Z. The MD generalization is similar, too.

period-vector (of a periodic function p(x,y)) —

A non-zero vector P = (x0,y0) such that for any (x,y)∈R2, p(x + x0, y + y0) = p(x,y). If
there exist two non-collinear vectors P1, P2 having this property, p(x,y) is said to be
2-fold periodic; in this case, for any point x ∈R 2 the points x, x + P 1, x + P 2,

3 Note, however, that the term analog is often used as the opposite of the term digital (for example, an
analog display or watch as opposed to a digital one). This is also the case in the expressions analog to
digital conversion and digital to analog conversion. Interestingly, in some references like
[Oppenheim99 p.8] an analog signal is simply a synonym to continuous-domain signal, while in other
references the term analog designates signals having continuous range (see the previous footnote).
Finally, in yet other references like [Oppenheim75 p. 7] or [Mitra93 p. 57] the term analog refers to
signals having both continuous domain and continuous range.
Appendix E: Glossary of the main terms 341

x + P1+ P2 define a period parallelogram of p(x,y). If the function p is discrete, then

R is replaced by Z. The MD generalization is straightforward.

periodic function —
A function having a period. Note that a 2D function p(x,y) can be 2-fold periodic
(such as p(x,y) = cos(x) + cos(y)) or only 1-fold periodic (such as p(x,y) = cos(x)).
Similarly, an MD function can be 1-fold periodic, 2-fold periodic, ... , or M-fold
periodic. In the continuous world, periodic functions have impulsive spectra whose
impulse locations are commensurable.

almost-periodic function —
See Appendix B in [Amidror09]. In the continuous world, almost-periodic
functions have impulsive spectra whose impulse locations are not commensurable.

aperiodic function —
A function that is neither periodic nor almost periodic. In the continuous world,
aperiodic functions have non-impulsive spectra.

grating (or line-grating) —

A pattern consisting of parallel lines. Unless otherwise mentioned it will be
assumed that a grating is periodic and consists of equally wide parallel, straight
lines that are separated by equal spaces. For example, a binary grating is a grating
with a square-wave intensity profile consisting of white lines (with a constant value
of 1) on a black background (whose value is 0).

E.3 Terms in the spectral domain

spectrum (or frequency spectrum) —

The frequency decomposition of a given function or signal, which specifies the
contribution of each frequency to the function or signal in question. The frequency
spectrum is obtained by taking the Fourier transform of the given function or
signal.

continuous-world spectrum (not to be confused with continuous spectrum) —

This term refers to a spectrum in the continuous world, as opposed to a discrete-
world spectrum. The spectrum obtained when applying a continuous Fourier
transform to a continuous-world signal is a continuous-world spectrum; note,
however, that this spectrum is not necessarily continuous (in the strict topological
sense), and it can even be purely impulsive if the given continuous-world signal is
periodic.

discrete-world spectrum (not to be confused with discrete spectrum) —

This term refers to a spectrum in the discrete world, as opposed to a continuous-
world spectrum. The sequence obtained when applying DFT to a discrete-world
342 Appendix E: Glossary of the main terms

signal is a discrete-world spectrum. Theoretically, a discrete-world spectrum is a

discrete-domain function (i.e. a function having a discrete domain); but when
processed by computer, it is in practice a digital function (since its range, too, is
discrete). Note that a discrete-world spectrum is often a discrete approximation to
an original continuous spectrum; but sometimes the continuous-world counterpart
of a discrete-world spectrum is not continuous, and it can be even purely impulsive
(if the given continuous-world signal is periodic).

continuous spectrum (not to be confused with continuous-world spectrum) —

A spectrum that is expressed by a continuous function (in the strict topological
sense of this term). Note that continuous-world spectra are not necessarily
continuous spectra. It should be also noted that a discrete (impulsive) continuous-
world function may have a continuous spectrum; for example, the continuous-
world spectrum of g(x) = 12 δ(x–f) + 12 δ(x+f) is the continuous function G(u) =
cos(2πfu) [Bracewell86 p. 413].

discrete spectrum (not to be confused with discrete-world spectrum) —

A spectrum that is expressed by a discrete function. Note that some continuous-
world spectra are discrete; this is the case if the given continuous-world signal is
periodic.

Hz (pronounced Hertz) —
A unit of frequency, named after the German physicist Heinrich Hertz. The unit
Hz is defined in the literature as “cycles per second” (see, for example, [Briggs95
p. 18]). But for the sake of convenience we sometimes use this term more loosely
as a generic frequency unit, i.e. as a compact shorthand notation for “cycles per
signal-domain unit”. This allows us to label all frequency axes (both in 1D and
MD cases, including axes showing the DFT output array after the adaptations
discussed in Sec. 4.3) in terms of Hz, rather than in terms of the more exact but
too cumbersome “cycles per signal-domain unit”. However, to avoid confusion,
this abuse of language should not be adopted in real life, and the frequency unit Hz
should only be used in its original, narrow sense, i.e. when the signal domain is
indeed measured in seconds. For example, if the signal domain is measured in
meters the spectral-domain frequency unit that one should use is “cycles per
meter”, and if the signal domain is measured in terms of minutes the spectral-
domain frequency unit should be “cycles per minute”.

visibility circle —
A circle around the spectrum origin whose radius represents the cutoff frequency,
i.e., the threshold frequency beyond which fine detail is no longer detected by the
eye. Obviously, its radius depends on several factors such as the viewing distance,
the light conditions, etc. It should be noted that the visibility circle is just a first-
order approximation. In fact, the sensitivity of the human eye is a continuous
2D bell-shaped function [Daly92 p. 6], with a steep “crater” in its center
(representing frequencies which are too small to be perceived), and “notches” in
Appendix E: Glossary of the main terms 343

the diagonal directions (owing to the drop in the eye sensibility in the diagonal
directions [Ulichney88 pp.79–84]). This term basically applies to the 2D case.

frequency vector —
A vector in the 2D (or MD) spectrum which represents the geometric location of a
2D (or MD) frequency in the spectrum (see, for example, Sec. 2.2 and Fig. 2.1 in
[Amidror09]).

pulse (not to be confused with impulse) —

A 1D or MD mathematical function representing any phenomenon which only has
non-zero values within a short, finite interval. For example, the function rect(ax)
represents a square (or rectangular) pulse of height 1 that is centered about the
origin of the x axis (see Fig. 3.5(a)). Its 2D counterpart rect(ax,ay) =
rect(ax) rect(ay) represents a 2D pulse of height 1 that is centered over the origin
of the x,y plane (see Fig. 3.6(a) or Fig. 6.8(a)).

impulse (not to be confused with pulse) —

A 1D or MD mathematical entity representing any phenomenon whose existence
is concentrated in a single point, i.e. which has non-zero value at a single point
(typically at the coordinate zero, i.e. at the origin). In the continuous world the
impulse is also known as a Dirac delta function, and it has zero width and an
infinite amplitude. In the discrete world the discrete impulse is also known as a
Kronecker delta function, and it has the width of one discrete element and an
amplitude of 1. See Sec. A.2 in Appendix A for further details.

DC impulse —
The impulse that is located on the spectrum origin. This impulse represents the
frequency of zero, which corresponds to the constant component in the Fourier
series decomposition of a periodic function. The amplitude of the DC impulse
corresponds to the intensity of this constant component. This impulse is
traditionally called the DC impulse because it represents in electrical transmission
theory the direct current component, i.e., the constant term in the frequency
decomposition of an electric wave. See, for example, [Smith03 pp. 152, 633].

compound impulse —
An impulse in the spectrum which is composed of several distinct impulses that
happen to fall on the same location and hence “fuse down” into a single impulse.
The amplitude of a compound impulse is the sum of the amplitudes (real or
complex) of the individual impulses from which it is composed. See Sec. 6.4 in
[Amidror09].

line-impulse —
A generalized 2D function which is impulsive along a 1D line through the plane,
and null everywhere else. A line-impulse can be graphically illustrated as a
“blade” whose behaviour is continuous along its 1D line support but impulsive in
344 Appendix E: Glossary of the main terms

the perpendicular direction. Note that the amplitude of a line-impulse does not
necessarily die out away from its center, and it may even rapidly oscillate between
two constant values.

curvilinear impulse —
A generalized 2D function which is impulsive along a 1D curvilinear path through
the plane, and null everywhere else. A curvilinear impulse can be graphically
illustrated as a curvilinear “blade” whose behaviour is continuous along its 1D
curvilinear support but impulsive in the perpendicular direction.

comb (or impulse-comb, Dirac-comb, impulse-train) —

An infinite train of equally spaced impulses located on a straight line in the
spectrum. A 1-fold periodic function p(x,y) is represented in the 2D spectrum by a
comb centered on the spectrum origin. The step and the direction of this comb
represent the frequency and the orientation of the periodic function; its impulse
amplitudes, which are given by the Fourier series development of the periodic
function, determine its intensity profile.

nailbed (or impulse-nailbed) —

An infinite 2D train of equally spaced impulses located in the spectrum on a
dot-lattice (either square-angled or skewed). A 2-fold periodic function p(x,y) is
represented in the 2D spectrum by an impulse nailbed centered on the spectrum
origin. The steps and the two main directions of this nailbed represent the
frequency and the orientation of the two main directions of the function’s 2D
periodicity; the impulse amplitudes, which are given by the 2D Fourier series
development of the periodic function, determine its intensity profile.

impulsive spectrum —
A spectrum which only consists of impulses, i.e., whose support consists of a
finite or at most denumerably infinite number of points. All periodic and almost-
periodic functions have impulsive spectra.

line-spectrum —
A spectrum which consists of line impulses (see, for example, Fig. 5.13).

wake —
A 2D continuous surface which trails off from an impulsive element (line-impulse,
curvilinear impulse, etc.), gradually dying out as it goes away from the impulsive
element in question. The amplitude of the wake may be considered as negligible
with respect to that of its generating impulsive element. As an example, the
spectrum of the cosinusoidal circular grating cos(2π f x 2 + y 2 ) is a circular
impulse ring of radius f, with a particular dipole-like impulsive behaviour on the
perimeter of the circle and a negative, continuous wake which gradually trails off
toward the center (see Example 10.6 in [Amidror09, Sec. 10.3] and Fig. 10.4(d)
there).
Appendix E: Glossary of the main terms 345

Nyquist frequency —
We sometimes use this term to denote half of the sampling frequency, which is the
highest frequency that discrete data (and the DFT spectrum) may contain [Smith03
pp. 42, 149]. Note, however, that we usually prefer to avoid using this term because
of its ambiguity. As explained in [Smith03 pp. 41–42 and 638], it turns out that
different authors use this term in four different meanings: the heighest frequency
contained in a (band-limited) signal; twice this frequency; the sampling frequency;
or half of the sampling frequency.

E.4 Miscellaneous terms

domain of a function (or domain of definition) —

The set of all values on which the function is defined, i.e. the set of values through
which the independent variable(s) of the function may run. For example, the
domain of the function g(x) = x2 is R, and the domain of its discrete counterpart
g(k) = k2 is Z. Note, however, that we sometimes use the term “domain” in a
wider sense, like in the terms signal domain and spectral domain of a Fourier
transform.

range of a function —
The set of all values that the function can take when its independent variable(s) run
throughout the domain of definition of the function. For example, the range of the
function g(x) = x 2 is R + = [0...∞), and the range of its discrete counterpart
g(k) = k 2 is Z + = 0, 1, 2, ... Note, however, that we sometimes use the term
“range” in a wider sense: For example, we may consider the range of the variable
x (i.e. the set of values along which it may run); the frequency range of the DFT
(which extends between minus half and plus half of the sampling frequency); the
sampling range of a continuous function g(x) (i.e. the range along the x axis within
which the function has been sampled), etc. As illustrated by the last example, this
wider sense of the term “range” may actually refer to the domain of a function (or
a subset thereof) rather than to its range; but usually this ambiguity should not
cause confusion.

point (not to be confused with dot) —

A single infinitely small geometric location within space, such as a point x along a
line, a point (x,y) within a plane, etc. Note that outside the strict mathematical
context the word “point” can be also used in a larger, figurative sense (such as a
point in a list of items, stressing a point in a discussion, etc.).

dot (not to be confused with point) —

A graphical symbol that is used to represent a point pictorially. A dot may have
any geometric shape (circular •, square Q , etc.).
346 Appendix E: Glossary of the main terms

vector (or point in a vector space) —

An element of the vector space in question (Rn, the u,v plane, etc.). We always
consider vectors as radius-vectors attached to the origin, and we do not distinguish
between a vector and a point in the vector space (i.e. the end point or the head of
the vector).

continuous —
An entity that is defined throughout a continuum such as R, Rn, an interval [a...b]
within R where a < b, etc., and has no jumps. Examples are a continuous line, a
continuous set, a continuous signal or function (in the strict topological sense), etc.

discrete —
A subset D of Rn is called discrete if there exists a number d > 0 so that for any
points a,b ∈D the distance between a and b is larger than d. Note, however, that
the term discrete is also used as the opposite of continuous. For example, periodic
functions have discrete spectra. See also the related terms discrete signal (in Sec.
E.2), discrete spectrum and discrete-world spectrum (in Sec. E.3).

binary (grating, etc.) —

A structure which contains only two values, usually 0 and 1.

commensurable (or commensurate) —

Two vectors v1,v2 ∈ Rn (or real numbers in R) are called commensurable if there
exist non-zero integers m,n for which nv2 = mv1, i.e. v2 = mn v1 (so that both v1 and
v2 can be measured as integer multiples of the same length unit, say 1n v1). Note that
two numbers x,y ∈ R are commensurable iff their ratio x/y is rational. More
generally, k vectors v1,...,vk ∈ Rn (or real numbers in R) are called commensurable
if they are linearly dependent over Q (which is identical to linear dependence over
Z). See also Problem 6-24.

incommensurable (or incommensurate) —

Two vectors v1,v2 ∈ R n (or real numbers in R) are called incommensurable if
there do not exist non-zero integers m,n for which nv2 = mv1, i.e. v2 = mn v1. Note
that two numbers x,y ∈ R are incommensurable iff their ratio x/y is irrational.
More generally, k vectors v 1 ,...,v k ∈ R n (or real numbers in R) are called
incommensurable if they are linearly independent over Q (which is identical to
linear independence over Z).

real-valued part; imaginary-valued part (of a complex-valued entity) —

If c = a + bi is a complex-valued entity, a is called its real-valued part, and
denoted by Re[c], and b is called its imaginary-valued part, and denoted by Im[c].
Note that one also finds in the literature (and sometimes in the present book, too)
the terms “real part” and “imaginary part”; but these terms may be misleading,
because both a and b are indeed real-world entities (real numbers, integers, real-
valued functions, etc.), and none of them is more real or more imaginary than the
Appendix E: Glossary of the main terms 347

other. Similarly, one should use the term complex-valued entity rather than
“complex entity”, since such entities are not necessarily more complex than other
ones.

phase (of a complex-valued entity) —

Any complex-valued entity c can be represented either by its real-valued part and
its imaginary-valued part, i.e. by c = Re[c] + Im[c] i, or alternatively, using the
polar notation, by its magnitude (also called absolute value or modulus) and its
phase (also called argument), i.e. by c = Abs[c] · ei Arg[c], where:

Abs[c] = Re[c]2 + Im[c]2 and Arg[c] = arctan Im[c]

Re[c]
Note that Re[c], Im[c], Abs[c] and Arg[c] are all real-valued entities.

phase (of a periodic function, etc.) —

If p(x) is a periodic function with period T, shifting it by T (or by any integer
multiple of T) along the x axis leaves the function unchanged: p(x + T) = p(x). But
any other shift of 0 < a < T gives a shifted version of p(x), p(x + a), which is said
to differ from p(x) in its phase. For further details see [Amidror09, Sec. C.4 in
Appendix C and Secs. 7.1–7.5 in Chapter 7].

Hermitian function (or sequence) —

A complex-valued function or sequence having a symmetric (i.e. even) real-valued
part and an antisymmetric (i.e. odd) imaginary-valued part (see, for example,
[Bracewell86 p. 16]). Thus, if g(x) is Hermitian, then g*(x) = g(–x) and g(x) =
g*(–x) [Bracewell86 p. 16]. The multidimensional generalization of a Hermitian
function is discussed in Sec. 8.4; in particular, if g(x) is Hermitian then g*(x) =
g(–x) and g(x) = g*(–x). Hermitian functions (or sequences) are named after the
French mathematician Charles Hermit. Note that in some references such entities
are called conjugate symmetric or conjugate even (see, for example, [Briggs95 p.
76]), and the term “Hermitian” is reserved to the symmetry properties which
prevail between the DFT and the IDFT [Briggs95 p. 74].

anti-Hermitian function (or sequence) —

A complex-valued function or sequence having an antisymmetric (i.e. odd) real-
valued part and a symmetric (i.e. even) imaginary-valued part (see, for example,
[Bracewell86 p. 15]). Thus, if g(x) is anti-Hermitian, then g*(x) = –g(–x) and
g(x) = –g*(–x) [Bracewell86 p. 22]. Note that in some references such entities are
called conjugate antisymmetric or conjugate odd (see, for example, [Hayes12
p. 4]). The multidimensional generalization of an anti-Hermitian function is
obtained in a similar way to that of a Hermitian function; in particular, if g(x) is
anti-Hermitian then g*(x) = –g(–x) and g(x) = –g*(–x).

separable function (of two or more variables) —

A function f(x,y) of the variables x and y is said to be separable if it can be
presented as (or separated into) a product of a function of x and a function of y:
348 Appendix E: Glossary of the main terms

f(x,y) = g(x)·h(y) [Gaskill78 pp. 16–17; Cartwright90 p. 117]. Note, however, that
we use this term in a slightly larger sense: A 2D function f(x,y) is separable if it can
be presented as a product of two independent 1D functions. Therefore, although
f(x,y) = g(x)·h(y) may no longer be separable (in the narrower sense) after it has
undergone a rotation or a skewing transformation, we will still consider it
as separable (with respect to the rotated or skewed axes x' and y': f(x',y') =
g(x')·h(y')). Note that in the MD case more possible variants exist, as illustrated for
example by rules 10 and 10a in Sec. 2.4.3.

inseparable function (of two or more variables) —

A function that is not separable. For example, the function representing a square
white dot is separable: rect(x,y) = rect(x)·rect(y), while the function representing a
circular white dot is inseparable.

bounded function —
A function g is said to be bounded if there exists a finite, positive constant c for
which |g(x)| ≤ c for any x in the domain of definition of g. For example, the
function g(x) = cos(x) is bounded (with c = 1). Note that a function which always
has a finite value is not necessarily bounded; for example, the function g(x) = x has
a finite value at any point x in its domain of definition, but it is unbounded.

unbounded function —
A function that is not bounded. For example, the function g(x) = 1/x is unbounded.

finite-duration function (also called finite-length, support-limited, spatially-limited,

time-limited or compactly-supported function) —
A function g is said to be a finite-duration function if there exists a finite, positive
constant c for which g(x) = 0 for |x| > c (see, for example, [Brigham88 p. 86] or
[Briggs95 p. 197]). For example, the function g(x) = x for x ∈[0...1] and g(x) = 0
otherwise is a finite-duration function. A function g(x) that is only defined on a
finite interval such as [a...b], where a < b, a,b ∈R, is also a finite-duration function
[Chu08 p. 51]; an example of such a function is g(x) = arccos(x), which is only
defined within the interval [–1...1]. The 2D or MD extensions of a finite-duration
function are straightforward. Note that finite-duration functions are often called in
the literature time-limited functions; but this term may be misleading in 2D or MD
functions, or even in 1D functions whose independent variable is not time but any
other entity (such as distance, etc.).

infinite-duration function (also called infinite-length function) —

A function that is not a finite-duration function. For example, the function g(x) =
cos(x) is an infinite-duration function.

band-limited function —
A function g is said to be a band-limited function if its spectrum G is a finite-
duration function. In other words, a function g(x) is band limited if its Fourier
Appendix E: Glossary of the main terms 349

transform G(u) is zero everywhere except over a finite interval of the variable u
(see, for example, [Bracewell86 p. 189]). For example, the function g(x) = sinc(x)
is band limited since its spectrum, G(u) = rect(u), is only non-zero over a finite
interval. The 2D or MD extensions of a band-limited function are straightforward.

causal function —
A function that is identically zero for negative values of its independent variable. In
other words, a function g(x) is said to be causal if g(x) = 0 for all x < 0 [Marks09
p. 12]. For example, the function g(x) = e–x step(x) (see Fig. 3.3(a)) is causal. A
short discussion on the origin of this term can be found in [Smith03 p. 130] or in
[Gaskill78 pp. 140–141]. For our needs, an MD causal function is a function that
is causal with respect to each of its M dimensions.

non-causal function —
A function that is not causal (see, for example, [Gaskill78 p. 141]).

transient function —
A function g(x) is said to be transient if it dies out for large |x| (see, for example,
[Castleman79 p. 164]).

1D —
One dimensional (function, periodicity, spectrum, impulse comb, etc.). Strictly, a
one dimensional entity is an entity that has only one dimension (such as a straight
line). However, by abuse of language we often use this term to designate a 2D
entity that only varies along one dimension, while its other dimension is constant.
For example, we say that an impulse comb such as in Fig. C.4(b) is 1D, even
though it subsists in the 2D u,v spectrum, because it varies only in one dimension,
while in the orthogonal direction it remains constantly zero. Similarly, we often say
that a 1-fold periodic structure such as a line grating (see, for example, Fig. C.4(a))
is a 1D structure, even though it actually spreads in the 2D x,y space, because it
only varies along one dimension but remains constant along the orthogonal
direction (i.e. along the individual grating lines). Such a 1D periodic structure is, in
fact, a constant extension of a true one-dimensional structure (such as a square
wave) into the second dimension. Note that we sometimes use the terms “1D
periodic” and “1-fold periodic” interchangeably as synonyms.

2D —
Two dimensional (function, periodicity, spectrum, impulse nailbed, etc.). Strictly,
any entity that has two dimensions, including a 1-fold periodic structure such as a
line grating, is a 2D entity. However, by abuse of language we often use this term
to designate a 2D entity that indeed varies along two dimensions. For example,
we say that a function such as g(x,y) = rect(x)rect(y) is 2D, because it varies along
two dimensions. Note that we sometimes use the terms “2D periodic” and
“2-fold periodic” interchangeably as synonyms.
350 Appendix E: Glossary of the main terms

moiré effect (or moiré phenomenon) —

A visible phenomenon which occurs when repetitive structures (such as line-
gratings, dot-screens, etc.) are superposed or sampled. It consists of a new pattern
which is clearly observed in the superposition or in the sampled structure, although
it does not appear in any of the original structures. The moiré effect is represented
in the spectrum of the superposition by new low frequencies which do not exist in
any of the individual spectra (the spectra of the original structures).

pseudo moiré —
A visible moiré-like phenomenon which occurs in some cases when repetitive
structures are superposed or sampled. Unlike a true moiré effect, a pseudo moiré is
only a modulation artifact, and it is not accompanied by corresponding low
frequencies in the spectrum of the superposition. For example, the new structure
which becomes visible in the sum of two cosines having similar frequencies is a
pseudo moiré (see Sec. D.3 in Appendix D).

sub-Nyquist artifact —
The term we use for a pseudo moiré effect that may occur when sampling a
periodic signal whose frequency is typically (but not exclusively) just below the
Nyquist frequency (i.e. just below half of the sampling frequency). See Sec. 8.6
and Problem 8-14.

singular moiré (or singular state, singular superposition) —

A configuration of the superposed or sampled structures in which the period of the
moiré in question becomes infinitely large (i.e., its frequency becomes 0), and
hence it can no longer be seen in the superposition. See Sec. 2.9 in [Amidror09].

macrostructure, microstructure (within a layer superposition or a sampled layer) —

The superposition of two or more layers (gratings, screens, etc.) may generate new
structures which appear in the superposition but not in the original layers. Similar
structures may also appear when sampling an original layer. These new structures
can be classified into two categories: the macrostructures, i.e., the moiré effects
proper, which are much coarser than the detail of the original layers; and the
microstructures, i.e., the tiny geometric forms which are almost as small as the
periods of the original layers, and are normally visible only from a close distance
or through a magnifying glass (see Chapter 8 in [Amidror09]).

jaggies (or staircasing effects) —

The term we use for the jagged edges that often appear when discretizing 2D
shapes, usually (but not exclusively) along borders with sharp 0/1 or black/white
transitions. Jaggies tend to appear especially along slanted or curved edges. See
Sec. 8.5.

wraparound —
See Remark C.10 in Appendix C and Problem 8-17.
Appendix E: Glossary of the main terms 351

operator —
An operator is a mathematical entity that gets as its input one or more functions
and returns as its output a function. For example, S[g(x),h(x)] = g(x) + h(x) is a
binary operator (i.e. an operator that acts on two functions) which returns for any
two given functions a third function, namely, their sum; D[g(x)] = g'(x) is a unary
operator (i.e. an operator that acts on a single function) which returns for any given
function its derivative; and F [g(x)] = G(u) is the unary operator that gives for any
(suitable) function g(x) its spectrum (Fourier transform) G(u).

transform (not to be confused with transformation) —

A transform is a unary operator, i.e. a mathematical entity that maps any input
function into an output function (including generalized functions such as impulses,
etc.). For example, the continuous Fourier transform returns for any (suitable)
function y = g(x) another function, w = G(u), which is called the spectrum of the
original function. Note, however, that this definition is too wide, since it includes
operators that are not usually considered as transforms such as D[g(x)] = g'(x)
(the operator which returns for any given function its derivative) or P[g(x,y)] = g(x)
(the operator which returns for each 2D function its 1D restriction to the x axis).
Strictly speaking, the term transform is reserved to integral transforms (such as the
Fourier transform), as well as their discrete counterparts (like the DFT). Thus,
while every transform is a unary operator, not all unary operators are transforms.
See also Sec. G.4 in [Amidror07] and in particular Footnote 2 there concerning
some terminological subtleties. Note that a transform does not necessarily work
for any possible input function; the set of functions for which it is defined is called
its domain of definition (much like the domain of definition of a function).

transformation (or mapping; not to be confused with transform) —

A transformation is a function which maps a space such as R, R2, Z, Z2, etc. into
another such space or into itself. Thus, for each element of the original space it
returns an element of the target space. For example, the function g(x) = 2x that
maps any point x to 2x is a transformation from R onto R; and the transformation
g(x,y) = x which projects any point (x,y) to the x axis is a transformation from R2
onto R . More specifically, a coordinate transformation is a continuous
transformation from a given space (such as R2, etc.) onto itself, which maps lines
parallel to the main axes into distinct, continuous and not self-intersecting lines or
curves (see also Appendices B–D in [Amidror07]). For example, a rotation by
angle θ is a coordinate transformation in R2.

functional —
A functional is a mathematical entity whose input is a function but whose output is
a number. Thus, a functional returns a number for any given function within its
domain of definition. For example, the definite integral between 0 and 1 is a
functional that returns for any suitable function g(x) the area enclosed between the
curve of the function, the x axis, and the two vertical lines x = 0 and x = 1.
352 Appendix E: Glossary of the main terms

rotation-invariant (transform, etc.; not to be confused with rotation-preserving) —

A transform is said to be rotation-invariant (or invariant under rotations) if
applying an arbitrary rotation transformation to any input of the transform does
not affect the output of the transform. More generally, a transform is said to be
invariant under a given coordinate transformation if applying the given
coordinate transformation to any input of the transform does not affect the output
of the transform. For example, the magnitude of the Fourier transform (as well as
the power spectrum) are invariant under shifts in the spatial domain (i.e. they are
shift-invariant), since shifts only affect the phase of the spectrum but not its
magnitude (or power).

rotation-preserving (transform, etc.; not to be confused with rotation-invariant) —

A transform is said to be rotation-preserving (or to preserve rotations) if applying
an arbitrary rotation transformation to any input of the transform affects the output
of the transform by exactly the same rotation. In other words, this happens if the
transform P in question commutes with rotation: P [R[g(x,y)]] = R[P [g(x,y)]],
namely, if the transform of a rotated function equals the rotated transform of
the original function. More generally, a transform is said to preserve a given
coordinate transformation if applying the given coordinate transformation to any
input of the transform results in applying the same coordinate transformation to
the output of the transform. The coordinate transformation is then said to be
preserved by the transform in question. For example, the 2D CFT preserves any
rotation transformation (i.e. it is a rotation-preserving transform), since rotating
the CFT input g(x,y) by angle θ results in an equal rotation of the spectrum G(u,v)
(thanks to the rotation theorem; see, for example, [Bracewell95 p. 157]). Note that
some references use in this context the term invariant, saying that the 2D CFT is
rotation-invariant, or invariant under rotation transformations (see, for example,
[Prestini04 p. 229] or [Granlund95 p. 235]); but this abuse of language may cause
confusion with the original meaning of the term invariant.

sinc (or cardinal sine) —

A very important function in Fourier theory, defined as (see [Bracewell86 p.62]):
⎧ sin(πx) x≠0
sinc(x) = ⎨ πx
⎩ 1 x=0
Confusingly, however, many references use a slightly different definition:
⎧ sin(x) x≠0
sinc(x) = ⎨ x
⎩ 1 x=0
(see, for example, [Briggs95 pp. 96–97]). The integral of the sinc function between
0 and x (using the second definition) is known as the sine integral function Si(x)
[Bracewell86 p. 62]; see also Sec. D.5 in Appendix D. Note, however, that some
references use the integral of the first definition of the sinc function, but usually
without calling it Si(x) (see, for example, [Kammler07 p. 44]).
 List of the main relations

The following list summarizes the main relations between the signal and frequency
domain parameters in the 1D and MD cases. Each relation is accompanied by its equation
number in the text.

1. Relations between Δx, Δu, R, F and N (see also Table 7.1 in Chapter 7):

R = NΔx (4.2)

F = NΔu (7.1)

Δu = 1 or Δx Δu = 1/N (4.5)
NΔx
RF = N (7.2)

Δu = 1/R (7.3)

Δx = 1/F (7.4)

The 2D or MD counterpart of each of these relations is simply obtained by applying the

given 1D relation in each dimension separately.

2. Relations involving the sampling frequency fs and the DFT minimal and maximal
frequencies:

fs = 1 (4.6)
Δx
fs = NΔu = F (7.5)

fmin = –12 fs fmax = 12 fs – Δu (4.10)

The 2D or MD counterpart of each of these relations is obtained by applying the given

1D relation in each dimension separately.

3. If the signal g(x) is periodic with period P and frequency f then we have, in addition:

Pf = 1 Sec. C.2

and its discrete counterpart:

pq = N (C.2)

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 353
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8,
© Springer-Verlag London 2013
354 List of the main relations

Note that the 2D or MD counterparts of these two relations are not a coordinate-wise
extension of the 1D relations, and they are given, as explained in Appendix C, by:
⎧1 if i = j
Pi · fj = ⎨ (C.7)
⎩0 if i ≠ j
⎧N if i = j
p i · qj = ⎨ (C.11)
⎩0 if i ≠ j
where i, j = 1,...,n (n being the number of dimensions enjoying periodicity, 1 ≤ n ≤ M), and
where P1,...,Pn ∈RM, f1,...,fn ∈RM, p1,...,pn ∈RM and q1,...,qn ∈RM.

In the 1D case we also have the following connections between the discrete-world
parameters p and q and their continuous-world counterparts P and f:

p = F/f = FP = N P (C.4)
R
N
q = R/P = Rf = f (C.4)
F
p = P/Δx (C.6)

q = f /Δu (C.5)

The 2D or MD counterpart of each of these relations is obtained by applying the

corresponding 1D relation to each dimension separately (note that the vectors pi and P i
are collinear, just like the vectors qi and fi ):

p i = F Pi = N P i (C.14)
R
N
q i = R fi = f i (C.15)
F
p i = Pi /Δx (C.12)

q i = fi /Δu (C.13)

where i = 1,...,n.

Note that in these MD relations it is assumed that the same parameters Δx, Δu, R, F and
N are used in all of the dimensions; if this is not the case, we may simply append the index
i to each of these parameters, too.
 List of notations and symbols

This list consists of the main symbols used in the text. They appear here with a very
brief description and a reference to the page in which they are first used or defined.
Obvious symbols such as ‘+’, ‘–’, etc. have not been included.

Symbol Short description Page

x, y Variables (coordinates) of the continuous signal domain 17

u, v Variables (coordinates) of the continuous spectral domain 17

x Vector variable in the MD continuous signal domain 17

u Vector variable in the MD continuous spectral domain 17

g(x) 1D continuous-world function, input of the 1D CFT 15

G(u) 1D continuous-world function, output of the 1D CFT 15

g(x,y) 2D continuous-world function, input of the 2D CFT 17

G(u,v) 2D continuous-world function, output of the 2D CFT 17

g(x) MD continuous-world function, input of the MD CFT 17

G(u) MD continuous-world function, output of the MD CFT 17

k, l Variables (coordinates) of the discrete signal domain 18, 20

n, m Variables (coordinates) of the discrete spectral domain 18, 20

k Vector variable in the MD discrete signal domain 20

n Vector variable in the MD discrete spectral domain 20

g(k) 1D discrete-world function, input of the 1D DFT 18

G(n) 1D discrete-world function, output of the 1D DFT 18

g(k,l) 2D discrete-world function, input of the 2D DFT 20

G(m,n) 2D discrete-world function, output of the 2D DFT 20

g(k) MD discrete-world function, input of the MD DFT 20

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 355
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8,
© Springer-Verlag London 2013
356 List of notations and symbols

G(n) MD discrete-world function, output of the MD DFT 20

g(xk) 1D discrete-world function (alternative notation) 7

g(xk,yl) 2D discrete-world function (alternative notation) 7

g(xk) MD discrete-world function (alternative notation) 7

g Discrete approximation of the continuous-world function g 19, 20

G Discrete approximation of the continuous-world spectrum G 19, 20

R Length of the sampling range in the input signal 70, 187

F Length of the frequency range in the DFT spectrum 158, 189

M Input / output space dimension of the CFT or DFT 17

N Input / output array length (number of elements) in the 1D DFT 17

N×N Input / output array size (number of elements) in the 2D DFT 29, 47

N×...×N Input / output array size (number of elements) in the MD DFT 69

P Period of a 1D periodic function 74, 273

f Frequency of a 1D periodic function 74, 273

P Period of an MD periodic function 281

f Frequency of an MD periodic function 281

f1 The first frequency above 0 in the DFT spectrum 46, 72

fs The sampling frequency of the DFT input signal 72

fG The highest frequency in the spectrum G(u) 92

ω Angular frequency 16

Δx, Δy Sampling interval along the x or y axis 18, 72

Δu, Δv Frequency step along the u or v axis 18, 72

p Period of a 1D periodic function in terms of array elements 74, 273

q Frequency of a 1D periodic function in terms of array elements 74, 273

p Period of an MD periodic function in terms of array elements 170, 286

q Frequency of an MD periodic function in terms of array elements 170, 286

List of notations and symbols 357

fM Frequency vector of a moiré effect 236

p(x) A 1D periodic function 321

P(u) The spectrum of p(x) 321

g(x+) A one-sided limit lim + g(x + a) 195

a→0
g(x –) A one-sided limit lim g(x + a) 195
a→0–
* 1D convolution (or cyclic convolution) 22, 27

** 2D convolution (or cyclic convolution) 24, 28

* MD convolution (or cyclic convolution) 25, 30

+ 1D correlation (or cyclic correlation) 22, 27

++ 2D correlation (or cyclic correlation) 24, 29

+ MD correlation (or cyclic correlation) 25, 30

a* Complex conjugate of a 22

a→0 The variable a tends to 0 309

a << b Much smaller than 302

θ Angle 114

ε An arbitrarily small, positive real number 224

P, F, A Matrices 25

|P | Determinant of matrix P 25

P –1 The inverse of an invertible matrix P 283

P –T The inverse transpose of an invertible matrix P 25, 283

a, b, r, s, t Real numbers (sometimes also used as integer numbers) 23

a, b, v, w Vectors 346

i, j, k, l, m, n Integer numbers or indices 33

i, j, k, l, m, n Integer vectors (vectors of integers) 29

i (In complex numbers): the imaginary unit, –1 15

cn Fourier series coefficients; impulse amplitudes 321

358 List of notations and symbols

Z The set of all integer numbers (positive, negative, and 0) 8

Q The set of all rational numbers 346

R The set of all real numbers 283

ZM The M-dimensional integer space 29

RM The M-dimensional Euclidean space 25

a...b The number range from a to b in R 48, 68

[a...b] The number range from a to b in R (alternative notation) 348

[a...b) The number range from a up to (but not including) b in R 187

k,...,l The number range from k to l in Z 17

Re[ ] The real-valued part of a complex entity 22, 202

Im[ ] The imaginary-valued part of a complex entity 22, 202

Abs[ ] The magnitude of a complex entity 202

Arg[ ] The phase of a complex entity 202

a·b Multiplication 196, 276

v ·w Scalar product of two vectors 17, 20

v×w Vector product of two vectors 285

v /w Element-wise vector division 20

|a| The absolute value of the number a (real or complex) 22

|v| The length (= Euclidean norm) of the vector v 282, 285

v⊥w v is perpendicular to w 285

proj(v)w The projection of v on w 282

G(u) = F [g(x)] G(u) is the Fourier transform of g(x) 92, 143

g(x) ↔ G(u) g(x) and G(u) are a Fourier pair 35, 333

~a Approximately a 318

~ Is Fourier series of ... 321

≈ Approximately equal 19, 20

List of abbreviations 359

δ(u) The impulse symbol (Dirac delta function) 247

δ(n) Discrete impulse (Kronecker delta function) 248

III(x) An impulse train (comb) 92, 261

step(x) The step function: 0 for x < 0, 0.5 at x = 0, and 1 for x > 0 48

rect(x) A square pulse: 1 in the range –0.5 < x < 0.5, 0.5 at x = ±0.5,
and 0 elsewhere 58

tri(x) A triangular pulse: 1 – |x| in the range –1 < x < 1, and 0 elsewhere 99

triangwave(x) A periodic triangular wave 112

sinc(x) sin(πx) for x ≠ 0, and 1 for x = 0 70, 352

πx
somb(r) The sombrero function 165, 166

Jn(x) Bessel function of the first kind of order n [Spiegel68 p. 136] 165

Si(x) The sine integral function 153, 307

Q End of example, proof, etc. 16

List of abbreviations

Symbol Short description Page

CFT Continuous Fourier transform 1

ICFT Inverse continuous Fourier transform 23

DFT Discrete Fourier transform 1

IDFT Inverse discrete Fourier transform 5

FFT Fast Fourier transform 2, 3

DC The impulse at the spectrum origin (i.e., at frequency zero) 45

Hz Hertz (unit of frequency) 5, 16

360 List of abbreviations

1D 1-dimensional 1

2D 2-dimensional 1

MD M-dimensional 1

iff If and only if 283

b/w Black and white 66

p.w. Pulse width 318

 References

[Abdou82] I. E. Abdou and K. Y. Wong, “Analysis of linear interpolation schemes for bi-level
image applications,” IBM Jour. of Research and Development, Vol. 26, No. 6, 1982, pp.
667–680.
[Amidror97] I. Amidror, “Fourier spectrum of radially periodic images,” Jour. of the Optical Society
of America A, Vol. 14, No. 4, 1997, pp. 816–826.

[Amidror07] I. Amidror, The Theory of the Moiré Phenomenon, Volume II: Aperiodic Layers.
Springer, Dordrecht, 2007.

[Amidror09] I. Amidror, The Theory of the Moiré Phenomenon, Volume I: Periodic Layers. Springer,
London, 2009 (second edition).

[Amidror09a] I. Amidror and R. D. Hersch, “The role of Fourier theory and of modulation in the
prediction of visible moiré effects,” Journal of Modern Optics, Vol. 56, 2009, pp. 1103–
1118.

[Bagchi99] S. Bagchi and S. K. Mitra, The Nonuniform Discrete Fourier Transform and its
Applications in Signal Processing. Kluwer, Boston, 1999.

[Bergland69] G. D. Bergland, “A guided tour of the fast Fourier transform,” IEEE spectrum, Vol. 6,
1969, pp. 41–52; also reprinted in Digital Signal Processing, L. R. Rabiner and C. M.
Rader (Eds.), IEEE Press, NY, 1972, pp. 228–239.

[Birkhoff77] G. Birkhoff and S. Mac Lane, A Survey of Modern Algebra. Macmillan Publishing, NY,
1977 (fourth edition).

[Boff86] K. R. Boff, L. Kaufman and J. P. Thomas, Handbook of Perception and Human

Performance (Vol. 2), John Wiley & sons, NY, 1986.
[Bovik05] A. Bovik (Ed.), Handbook of Image and Video Processing. Elsevier, Amsterdam, 2005
(second edition).

[Bracewell86] R. N. Bracewell, The Fourier Transform and its Applications. McGraw-Hill Publishing
Company, Reading, NY, 1986 (second edition).

[Bracewell95] R. N. Bracewell, Two Dimensional Imaging. Prentice Hall, NJ, 1995.

[Briggs95] W. L. Briggs and V. E. Henson, The DFT: An Owner’s Manual for the Discrete Fourier
Transform. SIAM, Philadelphia, 1995.

[Brigham88] E. O. Brigham, The Fast Fourier Transform and Its Applications. Prentice-Hall, NJ,
1988.

[Cartwright90] M. Cartwright, Fourier Methods for Mathematicians, Scientists and Engineers. Ellis
Horwood, UK, 1990.

[Castleman79] K. R. Castleman, Digital Image Processing. Prentice Hall, NJ, 1979.

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 361
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8,
© Springer-Verlag London 2013
362 References

[Champeney73] D. C. Champeney, Fourier Transforms and their Physical Applications. Academic

Press, London, 1973.

[Champeney87] D. C. Champeney, Fourier Theorems. Cambridge University Press, UK, 1987.

Ń zek]
[ Cí Ń Ń zŃ ek, Discrete Fourier Transforms and Their Applications. Adam Hilger, Bristol,
V. Cí
1986.

[Chu08] E. Chu, Discrete and Continuous Fourier Transforms: Analysis, Applications and Fast
Algorithms. CRC, Boca Raton, 2008.

[Cleveland93] W. S. Cleveland, Visualizing Data. AT&T Bell Laboratories, NJ, 1993.

[Cleveland94] W. S. Cleveland, The Elements of Graphing Data. AT&T Bell Laboratories, NJ, 1994
(revised edition).

[Cochran67] W. T. Cochran et al., “What is the fast Fourier transform?” IEEE Transactions on
Audio and Electroacoustics, Vol. 15, June 1967, pp. 45–55; also reprinted in Digital
Signal Processing, L. R. Rabiner and C. M. Rader (Eds.), IEEE Press, NY, 1972, pp.
240–250.

[Cooley65] J. W. Cooley and J. W. Tukey, “An algorithm for the machine calculation of complex
Fourier series,” Mathematics of Computation, Vol. 19, 1965, pp. 297–301.

[Cooley69] J. W. Cooley, P. A. W. Lewis and P. D. Welch, “The finite Fourier transform,” IEEE
Transactions on Audio and Electroacoustics, Vol. 17, June 1969, pp. 77–85; also
reprinted in Digital Signal Processing, L. R. Rabiner and C. M. Rader (Eds.), IEEE
Press, NY, 1972, pp. 251–259.

[Cooley70] J. W. Cooley, P. A. W. Lewis and P. D. Welch, “The fast Fourier transform algorithm:
programming considerations in the calculation of sine, cosine and Laplace transforms,”
Journal of Sound and Vibration, Vol. 12, July 1970, pp. 315–337; also reprinted in
Digital Signal Processing, L. R. Rabiner and C. M. Rader (Eds.), IEEE Press, NY,
1972, pp. 271–293.

[Cornsweet70] T. N. Cornsweet, Visual Perception. Harcourt Brace Jovanovich, Florida, 1970.

[Coulon84] F. de Coulon, Theorie et Traitement des Signaux. Presses Polytechniques Romandes,

Lausanne, 1984 (in French).

[Crow77] F. C. Crow, “The aliasing problem in computer-generated shaded images,”

Communications of the ACM, Vol. 20, No. 11, 1977, pp. 799–805.

[Daly92] S. Daly, “The visible differences predictor: an algorithm for the assessment of image
fidelity,” Human Vision, Visual Processing and Digital Display III, SPIE proceedings,
Vol. 1666, USA, 1992, pp. 2–15.

[Dutt93] A. Dutt and V. Rokhlin, “Fast Fourier transforms for nonequispaced data,” SIAM
Journal on Scientific Computing, Vol. 14, No. 6, November 1993, pp. 1368–1393.

[Erdélyi54] A. Erdélyi (Ed.), Tables of Integral Transforms (Vol. 1). McGraw-Hill, NY, 1954.

[Fielding06] G. Fielding, R. Hsu, P. Jones and C. DuMont, “Aliasing and reconstruction distortion
in digital intermediates,” SMPTE Motion Imaging Journal, Vol. 115, No. 4, 2006, pp.
128–136.
[Foley90] J. D. Foley, A. van Dam, S. K. Feiner and J. F. Hughes, Computer Graphics: Principles
and Practice. Addison-Wesely, Reading, Massachusetts, 1990 (second edition).

[Fourier22] J. B. J. Fourier, Théorie Analytique de la Chaleur. Didot, Paris, 1822 (in French).
References 363

[Fourier24] J. B. J. Fourier, “Théorie du mouvement de la chaleur dans les corps solides,” Mémoires
de l’Académie des Sciences de l’Institut de France, Vol. 4, 1824, pp. 185–556 (in
French).

[Fourier26] J. B. J. Fourier, “Théorie du mouvement de la chaleur dans les corps solides,” Mémoires
de l’Académie des Sciences de l’Institut de France, Vol. 5, 1826, pp. 153–246 (in
French).

[Gaskill78] J. D. Gaskill, Linear Systems, Fourier Transforms, and Optics. John Wiley & Sons,
NY, 1978.

[Glassner95] A. S. Glassner, Principles of Digital Inage Synthesis, Volume 1. Morgan Kaufmann,

San Francisco, 1995.

[Gomes97] J. Gomes and L. Velho, Image Processing for Computer Graphics. Springer, NY, 1997.

[Gonzalez87] R. C. Gonzalez and P. Wintz, Digital Image Processing. Addison-Wesley, Reading, Ma,
1987 (second edition).

[Goodman05] J. W. Goodman, Introduction to Fourier Optics. Roberts & Company, Colorado, 2005
(third edition).

[Gradshteyn07] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products. Academic
Press, San Diego, 2007 (seventh edition).

[Granlund95] G. H. Granlund and H. Knutsson, Signal Processing for Computer Vision. Kluwer,
Dordrecht, 1995.

[Gröchenig01] K. Gröchenig, Foundations of Time-Frequency Analysis. Birkhäuser, Boston, 2001.

[Hardy68] G. H. Hardy and W. W. Rogosinski, Fourier Series. Cambridge University Press,
London, 1968 (reprint of third edition of 1956).

[Harris78] F. J. Harris, “On the use of windows for harmonic analysis with the discrete Fourier
transform,” Proceedings of the IEEE, Vol. 66, 1978, pp. 51–83.

[Harris98] J. W. Harris and H. Stocker, Handbook of Mathematics and Computational Science.

Springer, NY, 1998.

[Hayes12] M. H. Hayes, Digital Signal Processing. Schaum’s Outline Series, McGraw-Hill, NY,
2012 (second edition).

[Heideman85] M. T. Heideman, D. H. Johnson and C. S. Burrus, “Gauss and the history of the fast
Fourier transform,” Archive for History of Exact Sciences, Vol. 34, No. 3, 1985, pp.
265–277. An earlier version was published in IEEE ASSP Magazine, Vol. 1, No. 4,
1984, pp. 14–21.

[Henson92] V. E. Henson, “DFTs on irregular grids: The anterpolated DFT,” Technical Report NPS-
MA-92-006, Naval Postgraduate School, Monterey, CA, 1992. Available online at
http://www.dtic.mil/dtic/tr/fulltext/u2/a255187.pdf (accessed Oct. 2, 2012).

[Herivel75] J. Herivel, Joseph Fourier: The Man and the Physicist. Clarendon Press, Oxford, 1975.

[Jähne95] B. Jähne, Digital Image Processing: Concepts, Algorithms and Scientific Applications.
Springer, Berlin, 1995 (third edition).

[Jain89] A. K. Jain, Fundamentals of Digital Image Processing. Prentice Hall, NJ, 1989.

[Kammler07] D. W. Kammler, A First Course in Fourier Analysis. Cambridge University Press,

Cambridge, 2007 (revised edition).
364 References

[Komrska79] J. Komrska, “Symmetry of Fraunhofer diffraction expressed by series expansions of

Fourier transforms,” Optica Acta, Vol. 26, No. 2, 1979, pp. 173–195.

[Kreyszig93] E. Kreyszig, Advanced Engineering Mathematics. John Wiley & Sons, NY, 1993
(seventh edition).

[Lay03] D. C. Lay, Linear Algebra and its Applications. Addison-Wesley, Boston, 2003 (third
edition).

[Loan92] C. van Loan, Computational Frameworks for the fast Fourier transform. SIAM,
Philadelphia, 1992.

[Mackay91] A. L. Mackay, A Dictionary of Scientific Quotations. Institute of Physics, Bristol, 1991

(second edition).

[Mallat91] S. Mallat, A Wavelet Tour of Signal Processing. Academic Press, San Diego, 1991
(second edition).

[Manolakis11] D. Manolakis and V. Ingle, Applied Digital Signal Processing. Cambridge University
Press, UK, 2011.

[Maplesoft] DiscreteTransforms/FourierTransform – Maple® Help.

http://www.maplesoft.com/
support/help/Maple/view.aspx?path=DiscreteTransforms/FourierTransform
(accessed January 19, 2012).

[Marks91] R. J. Marks II, Introduction to Shannon Sampling and Interpolation Theory. Springer,
NY, 1991. http://marksmannet.com/RobertMarks/REPRINTS/1999_Introduction
ToShannonSamplingAndInterpolationTheory.pdf (accessed May 30, 2012).

[Marks09] R. J. Marks II, Handbook of Fourier Analysis and its Applications. Oxford University
Press, NY, 2009.

[Marvasti01] F. Marvasti (Ed.), Nonuniform Sampling: Theory and Practice. Kluwer, NY, 2001.

[Matlab02] Signal Processing Toolbox for Use With Matlab®: User’s Guide. The MathWorks Inc.,
MA, 2002 (fourth printing).

[Mitchell88] D. P. Mitchell and A. N. Netravali, “Reconstruction filters in computer graphics,”

Proceedings of SIGGRAPH 1988, Computer Graphics, Vol. 22, No. 4, 1988, pp. 221–
228.

[Mitra93] S. K. Mitra and J. F. Kaiser, Handbook for Digital Signal Processing. Wiley, NY, 1993.

[Morrison94] N. Morrison, Introduction to Fourier Analysis. Wiley, NY, 1994.

[Nussbaumer82] H. J. Nussbaumer, Fast Fourier Transforms and Convolution algorithms. Springer,

Berlin, 1982 (second edition).

[Nuttall81] A. H. Nuttall, “Some windows with very good sidelobe behavior,” IEEE transactions
on Acoustics, Speech and Signal Processing, Vol. 29, 1981, pp. 84–91.

[Oberhettinger90] F. Oberhettinger, Tables of Fourier Transforms and Fourier Transforms of

Distributions. Springer, Berlin, 1990.

[Oppenheim75] A. V. Oppenheim and R. W. Schafer, Digital Signal Processing. Prentice Hall, London,
1975.

[Oppenheim99] A. V. Oppenheim and R. W. Schafer with J. R. Buck, Discrete-Time Signal Processing.

Prentice Hall, London, 1999 (second edition).

[Papoulis62] A. Papoulis, The Fourier Integral and its Applications. McGraw-Hill, NY, 1962.
References 365

[Papoulis68] A. Papoulis, Systems and Transforms with Applications in Optics. McGraw-Hill, NY,
1968.

[Park09] W. Park et al., “Accurate image rotation using Hermite expansions,” IEEE Transactions
on Image Processing, Vol. 18, No. 9, September 2009, pp. 1988–2003.

[Petersen62] D. P. Petersen and D. Middleton, “Sampling and reconstruction of wave-number-limited

functions in N-dimensional Euclidean spaces,” Information and Control, Vol. 5, 1962,
pp. 279–323.

[PhysicsForums12] Physics Forums, 2012, “Analog and digital, continuous and discrete.”
http://www.physicsforums.com/showthread.php?t=479118 (accessed October 2,
2012).
[Poularikas96] A. D. Poularikas (Ed.), The Transforms and Applications Handbook. CRC, Boca
Raton, 1996.

[Pratt91] W. K. Pratt, Digital Image Processing. John Wiley & Sons, NY, 1991 (second edition).

[Press02] W. A. Press, S. A. Teukolski, W. T. Vetterling and B. P. Flannery, Numerical recipes in

C++. Cambridge University Press, UK, 2002 (second edition).

[Prestini04] E. Prestini, The Evolution of Applied Harmonic Analysis: Models of the Real World.
Birkhäuser, Boston, 2004.

[Ramirez74] R. W. Ramirez, “The fast Fourier transform’s errors are predictable, therefore
manageable,” Electronics, Vol. 47, June 1974, pp. 96–102.

[Ramirez85] R. W. Ramirez, The FFT: Fundamentals and Concepts. Prentice Hall, NJ, 1985.

[Rao98] R. M. Rao and A. S. Bopardikar, Wavelet Transforms: Introduction to Theory and

Applications. Addison-Wesley, MA, 1998.

[Rosenfeld82] A. Rosenfeld and A. C. Kak, Digital Picture Processing (Vol.1). Academic Press,
Florida, USA, 1982 (second edition).

[Rubinstein88] R. Rubinstein, Digital Typography: An Introduction to Type and Composition for

Computer System Design. Addison-Wesely, Reading, Massachusetts, 1988.

[Russ95] J. C. Russ, The Image Processing Handbook. CRC Press, Boca Raton, 1995 (second
edition).

[Smith96] C. C. Smith, J. F. Dahl and R. J. Thornhill, “The duality of leakage and aliasing and
improved digital spectral analysis techniques,” Journal of Dynamic Systems,
Measurement and Control, Vol. 118, 1996, pp. 741–747.

[Smith03] S. W. Smith, Digital Signal Processing: A Practical Guide for Engineers and
Scientists. Newnes, USA, 2003.

[Smith07] J. O. Smith III, Mathematics of the Discrete Fourier Transform (DFT) with Audio
Applications. W3K Publishing, WA, 2007 (second edition).

[Spiegel68] M. R. Spiegel, Mathematical Handbook of Formulas and Tables. Schaum’s Outline

Series, McGraw-Hill, NY, 1968.
[Stark79] H. Stark, “Sampling theorems in polar coordinates,” Jour. of the Optical Society of
America, Vol. 69, No. 11, November 1979, pp. 1519–1525.
[Stark82] H. Stark, “Polar sampling theorems of use in optics,” Applications of Mathematics in
Modern Optics, SPIE proceedings, Vol. 358, USA, 1982, pp. 24–30.
366 References

[Stein71] E. M. Stein and G. Weiss, Fourier Analysis on Euclidean Spaces. Princeton University
Press, NJ, 1971.
[Strang93] G. Strang, “Wavelet transforms versus Fourier transforms,” Bulletin of the American
Mathematical Society (New Series), Vol. 28, No. 2, April 1993, pp. 288–305.

[Stucki79] P. Stucki (Ed.), Advances in Digital Image Processing, Plenum Press, NY, 1979.

[Thornhill83] R. J. Thornhill and C. C. Smith, “Time aliasing: A digital data processing

phenomenon,” Journal of Dynamic Systems, Measurement and Control, Vol. 105, 1983,
pp. 232–237.

[Ulichney88] R. Ulichney, Digital Halftoning. MIT Press, USA, 1988.

[Vygodski73] M. Vygodski, Aide-Mémoire de Mathématiques Supérieures. Mir, Moscow, 1973

(French translation).

[Wandell95] B. A. Wandell, Foundations of Vision. Sinauer, Massachusetts, 1995.

[Weaver89] H. J. Weaver, Theory of Discrete and Continuous Fourier Analysis. Wiley, NY, 1989.

[Weisstein99] E. W. Weisstein, CRC Concise Encyclopedia of Mathematics. CRC, Boca Raton, 1999.

[Williams00] G. L. Williams, “Sub-Nyquist distortions in sampled data, waveform recording, and

video imaging,” NASA Technical Memorandum TM-2000-210381, Ohio, 2000.
Available online at http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/
20000120584_2000177329.pdf (accessed Oct. 2, 2012).

[Wikipedia12] Wikipedia®, 2012, “Aliasing.” http://en.wikipedia.org/wiki/Aliasing (accessed

October 2, 2012).

[Wolberg90] G. Wolberg, Digital Image Warping. IEEE Computer Society Press, California, 1990.

[Wolfram96] S. Wolfram, The Mathematica® Book. Cambridge University Press, UK, 1996 (third
edition).

[WolframAlpha09] WolframAlpha® LLC, 2009. http://www.wolframalpha.com/input/?i=limit

(accessed August 24, 2011).

[Wright07] H. Wright, Introduction to Scientific Visualization. Springer, London, 2007.

[Zayed93] A. I. Zayed, Advances in Shannon’s Sampling Theory. CRC, Boca Raton, 1993.

[Zwillinger96] D. Zwillinger (Ed.), CRC Standard Mathematical Tables and Formulae. CRC, Boca
Raton, 1996 (30th edition).

[Zygmund68] A. Zygmund, Trigonometric Series (Vol. 1). Cambridge University Press, US, 1968
(corrected reprint of second edition of 1959).
 Index

Page numbers followed by the letter “g” indicate entries in the glossary (Appendix E).

A analog function (or signal), 44, 340g

addition theorem (or rule), see under theorems angular frequency, 16
(rules) for the CFT; theorems (rules) for the anti-aliasing, 210
DFT see also aliasing: avoiding or
affine theorem (or rule), see under theorems eliminating
(rules) for the CFT anticentrosymmetric, 204–210
alias, 89–91, 96, 137, 142, 219, 222 anti-Hermitian function (or signal), 44, 347g
aliased frequency, 94, 99, 132–133 antisymmetric function (or signal), 203–207
aliasing, 5, 55, 89–142, 166, 210–223, 244, see also odd function (or signal)
274–280, 293–294 aperiodic function (or signal), 341g
adverse effects (summary), 166 apodisation, 147, 168
avoiding or eliminating, 95, 133, 166, 178, artifacts:
268 DFT artifacts, see under DFT
detecting, 94–95, 99, 122–123, 132–133, printing artifacts, see under printing
268 reconstruction artifacts, see under
duality with leakage in the other domain, reconstruction
178 autocorrelation theorem (or rule), see under
in 1D, 89–113, 274–280 theorems (rules) for the CFT; theorems
in 2D or MD, 113–132, 293–294 (rules) for the DFT
origin of the term, 89–91
reducing, 95, 133, 166, 178, 268 B
signal-domain interpretation, 89–91, 96, 106, background noise, 66
142, 166, 210–223 band-limited function (or signal), 92, 95, 138,
spectral-domain interpretation, 92–95, 96, 148, 174, 197, 348g
106, 166, 210–223 bar (vertical or rotated), 114–117, 216, 332–334
surprising shapes, 138–141 beating (modulation) effect, 224–240, 302–306
that occurs in the signal domain, see signal- Bessel function, 165
domain aliasing binary, 346g
see also foldover; jaggies blade, see line impulse
aliasing free, see aliasing: avoiding or bounded function (or signal), 348g
eliminating box plot, 11–12
almost band-limited function (or signal), 38, 95
almost periodic function (or signal), 41–42, C
163–164, 341g causal function (or signal), 33, 36, 49, 349g
almost time-limited function (or signal), 38 centrosymmetric, 204–210
amplitude spectrum, 51 CFT, 1
see also magnitude analytic calculation, 1, 2

I. Amidror, Mastering the Discrete Fourier Transform in One, Two or Several Dimensions: Pitfalls 367
and Artifacts, Computational Imaging and Vision 43, DOI 10.1007/978-1-4471-5167-8,
© Springer-Verlag London 2013
368 Index

CFT (continued) continuous-time function (or signal), 337g

combination rules, see theorems (rules) for continuous-valued function (or signal), 337g
the CFT continuous-world function (or signal), 339g
conversion rules between the various continuous-world spectrum, 341g
definitions, see under conversion rules contour plot, 8–11
definition, 15 conversion rules:
definition variants, 15–17, 42–43, 78–79 between different CFT definitions, 17, 43
existence conditions, 16 between different DFT definitions, 19, 43
generalized, 4, 16 convolution by DFT, 60, 65
history, 2 convolution of square pulse and sinc, 307–310
inverse, 15–17, 23 convolution theorem (or rule), see under
MD definition, 17 theorems (rules) for the CFT
of a periodic function, 41–42, 74–77, 92 see also cyclic convolution under theorems
of an almost periodic function, 41–42 (rules) for the DFT
of an aperiodic function, 41 convolution with sinc having no effect, 306–
properties, see theorems (rules) for the CFT; 307
symmetry rules for the CFT coordinate transformation, 194, 351g
reversibility, 70 cross correlation theorem (or rule), see under
rules, see theorems (rules) for the CFT theorems (rules) for the CFT; theorems
symmetry rules, see symmetry rules for the (rules) for the DFT
CFT curvilinear line impulse, 122, 124, 344g
tables of CFT pairs, 43, 68 see also impulsive ring
theorems, see theorems (rules) for the CFT cyclic convolution theorem (or rule), see under
transformation rules, see theorems (rules) for theorems (rules) for the DFT
the CFT see also convolution under theorems (rules)
2D and MD definitions, 17 for the CFT
checkerboard effect (due to wrong input cyclic foldover, see foldover
reorganization), 61, 182 cyclic shift, 46, 49–51, 54, 187, 301
circular grating, 122–132 cyclic (within a discrete array), see wraparound
circular symmetry, 133, 167, 194
comb, see impulse train D
combination rules for the CFT, see theorems data extension methods, 265–272
(rules) for the CFT method 1: adding new values beyond the
combination rules for the DFT, see theorems original range, 265–266, 271
(rules) for the DFT method 2: denser sampling within the
commensurable, 42, 184, 346g original range, 266–268, 271
commutativity, 316 method 3: adding zeroes after each value
compactly-supported function (or signal), 348g (zero packing), 268, 271
complex conjugation theorem (or rule), see method 4: replicating the entire original
under theorems (rules) for the CFT; range, 269, 271
theorems (rules) for the DFT method 5: replicating each of the original
compound impulse, 343g elements, 269, 271
congruent, 246 method 6: adding zeroes beyond the original
connected line plot, 7–9, 64, 65, 196, 248 range (zero padding), 270–271
connected symbol plot, 7–9, 277 DC, 45, 66, 128, 168, 179, 241, 249–253, 343g
continuous, 346g density plot, 8–11
continuous function (or signal), 44, 337g DFT, 1
continuous-amplitude function (or signal), 337g alternatives to, 180–182
continuous-domain discrete-range function (or amplitude scaling, 34–36
signal), 338g applications, 1, 3, 5
continuous-domain function (or signal), 337g signal convolution, 60, 65
continuous Fourier transform, see CFT signal filtering, 55, 64
continuous-range function (or signal), 337g artifacts, 1, 5, 6, 38, 89, 166, 180, 203–210,
continuous spectrum, 342g 291 see also aliasing; leakage
Index 369

DFT (continued) DFT (continued)

as approximation to CFT, 1, 19–20, 36– true units along the axes, see under true
41, 55, 317–320, 322–323, 325–332 units
as approximation to Fourier series, 41–42, uses, see DFT: applications
321–323, 326–332 2D and MD definitions, 20
behaviour at a discontinuity, see digital function (or signal), 44, 338g
discontinuity; midvalue rule at digital typography (aliasing and anti-aliasing),
discontinuities 210
combination rules, see theorems (rules) for digitization, 336g
the DFT dilation theorem (or rule), see under theorems
conversion rules between the various (rules) for the CFT; theorems (rules) for the
definitions, see under conversion rules DFT
definition, 17 Dirac comb, see impulse train
definition variants, 17–21, 78–79 Dirac delta function, 247, 343g
frequency range, 40, 73, 84, 93, 97–98, 107, see also impulse: in the continuous world
139, 158, 189–192, 212, 259–264, 266 discontinuity, 42, 49, 53, 59, 61, 71, 146, 195–
graphical development, 31–39 201
highest frequency, 73, 138, 266 see also midvalue rule at discontinuities
history, 3 discontinuous function (or signal), 338g
in mathematical software packages, 20–21, discrepancy between DFT and CFT, 5, 6, 55,
34, 46 78–79, 89, 104, 136, 138, 142, 163, 175,
index ranges, 18, 19, 21, 33–34, 45–47, 69– 195–201, 203–210
76, 300–303, 320 in periodic functions, 74, 76, 79, 82, 85
indexing conventions, see DFT: index discrete, 346g
ranges discrete-amplitude function (or signal), 338g
input array, see input array discrete-domain function (or signal), 338g
inverse, 5, 17–21, 28, 51–54, 60, 64–65, 134, discrete-domain continuous-range function (or
136, 174, 178 signal), 338g
limitations of the DFT, 180 discrete Fourier transform, see DFT
lowest negative frequency, 73, 138, 204, 246 discrete function (or signal), 44, 338g
lowest positive frequency, 72, 138 discrete-range function (or signal), 338g
matrix notation, 20 discrete spectrum, 342g
maximum frequency, 73, 138, 266 discrete-time function (or signal), 338g
MD definition, 20 discrete-valued function (or signal), 338g
minimum frequency, 73, 138, 204, 246 discrete-world function (or signal), 339g
non-uniform, 194 discrete-world spectrum, 341g
of a periodic function, 41–42, 74–77, 95 discretization, 336g
of an almost periodic function, 41–42 display considerations, see plotting
output array, see output array considerations
periodicity, 31–33, 94, 196, 240, 299–300 domain of a function, 345g
pitfalls, 1, 5, 6, 45 domain of definition, 345g
see also under pitfalls dot, 7, 50, 345g
properties, see theorems (rules) for the DFT; see also point
symmetry rules for the DFT drawing methods for discrete data, see graphic
reversibility, 51, 70, 134, 174, 198–201 presentation of discrete data
rules, see theorems (rules) for the DFT
spectral resolution, 72, 84, 146, 148, 159, E
186, 190, 194, 265–266 error:
symmetry rules, see symmetry rules for the discretization error, 41, 242
DFT due to signal-domain truncation, 158–161
theorems, see theorems (rules) for the DFT estimation of the DFT approximation
transformation rules, see theorems (rules) for errors, 41
the DFT reconstruction error, see under reconstruction
370 Index

error (continued) graphic development of the DFT, see under

rounding error, 89, 242 DFT
see also numeric precision; pitfalls; graphic presentation of an impulse, 247–248
discrepancy between DFT and CFT graphic presentation of discrete data, 7–13, 223,
even function (or signal), 44, 203–206 241
extension of the input data, see data extension in 1D, 7–9, 223
methods in 2D, 8–13, 223
in MD, 13
F grating, see line grating; circular grating
fast Fourier transform see FFT gray levels, 8, 61–63, 122, 210–211
FFT, 2, 3, 4, 36, 186, 206
filter: H
low pass, 55, 64, 95, 161, 178, 197, 309 harmonic, 76, 95, 110–113, 118, 128, 132, 162,
finite-duration function (or signal), 133, 147– 170, 274, 276
148, 173–174, 185, 306, 348g harmonically related, 163, 165, 184
finite-length function (or signal), 348g Hermitian function (or signal), 44, 66, 203–206,
foldover, 93, 94–95, 99–101, 103, 107, 112, 301, 347g
274–280 Hertz, see Hz
in the signal domain, 136 hybrid spectrum, see under spectrum
in 2D, 114–133, 138–142, 212 Hz, 16, 69, 72, 342g
misinterpretation of, 94, 116, 136–138
vs. wraparound, 240, 245–246, 296–298 I
see also aliasing ICFT, see CFT: inverse
Fourier, Jean Baptiste Joseph, 2 IDFT, see DFT: inverse
Fourier series, 41–42, 188, 321–324, 326–332 imaginary part, 346g
Fresnel function, 38 imaginary part theorem (or rule), see under
frequency: theorems (rules) for the CFT; theorems
highest frequency in the DFT, 73, 86 (rules) for the DFT
lowest negative frequency in the DFT, 73 imaginary-valued part, 346g
lowest positive frequency in the DFT, 72, 86 impulse, 247–264, 343g
of a periodic function, 74, 162, 273 compound impulse, 343g
discrepe counterpart, see q (discrete curvilinear impulse, see curvilinear line
counterpart of the frequency of a impulse
periodic function) graphic representation, see graphic
sampling frequency, see sampling frequency presentation of an impulse
frequency dilation theorem (or rule), see under in the continuous world, 41, 74–77, 79, 107,
theorems (rules) for the CFT 120, 145
frequency interval, see frequency step properties, 92, 247
frequency range, see DFT: frequency range in the discrete world, 201, 248
frequency resolution, see DFT: spectral height of a discrete impulse in the
resolution signal domain, 259–264
frequency sample, 19 height of a discrete impulse in the
frequency shift theorem (or rule), see under spectral domain, 248–258, 310–318,
theorems (rules) for the CFT; theorems 324, 326–332
(rules) for the DFT line impulse, see line impulse
frequency spectrum see spectrum sharp (not leaked), 147, 162, 171–173, 179,
frequency step, 18, 72, 189–192 182, 248, 257, 261, 276–280, 294
frequency vector, 170–173, 281–285, 343g stray impulse, 95, 112, 113, 274, 279,
function, 1, 336g 294
functional, 351g that falls between discrete array elements,
162–165, 175–177, 182, 258, 261, 276–
G 280, 294
generalized Fourier transform, see under CFT impulse comb, see impulse train
Gibbs phenomenon, 146, 155, 197, 309 impulse nailbed, see impulse train: 2D or MD
Index 371

impulse train, 32, 41, 92, 112, 118, 179, 211– leakage free, see leakage: avoiding or
215, 261–263, 344g eliminating
2D or MD, 113, 211, 344g legend, 10, 62, 87
impulsive ring, 122–132 length:
see also curvilinear impulse of the DFT input and output arrays, 45–47,
impulsive spectrum, see under spectrum 69–70, 148, 164, 185–186
incommensurable, 346g see also data extension methods
infinite-duration function (or signal), 348g of the frequency range, see DFT: frequency
infinite-length function (or signal), 348g range
input array, 45, 69 of the sampling range, see sampling range
length of, see under length level lines, 10–11
periodicity, see DFT: periodicity line (single straight line), see bar (vertical or
reorganization, see under reorganization rotated)
true units, see under true units line grating, 66–67, 118–119, 136–137, 179–
input step, see sampling interval 180, 211–218, 341g
inseparable function (or signal), 348g line impulse, 114–117, 216–217, 343g
interpolation, 201, 228,–230, 234–235, 243, circular, see impulsive ring
270 curvilinear, see curvilinear line impulse
invariant, 133, 194, 352g line spectrum, 344g
inverse continuous Fourier transform, see CFT: line symmetry, 204–206
inverse linearity theorem (or rule), see under theorems
inverse discrete Fourier transform, see DFT: (rules) for the CFT; theorems (rules) for the
inverse DFT
inversion theorem (or rule), see under theorems lobe, see main lobe; sidelobes
(rules) for the CFT; theorems (rules) for the
DFT M
macrostructure, 350g
J magnitude, 34, 47, 61, 143, 159, 163, 187, 202
jaggies, 133, 186, 206, 210–223, 244–245, 350g main lobe, 144–148, 158–159, 162–163, 165,
314, 319
K matrix, 20
Kronecker delta function, 248, 262, 343g matrix notation for the DFT, 20
see also impulse: in the discrete world MD, 1, 113, 165, 204–210, 281–296
MD n-fold periodic function, 281–297, 341g
L mesh plot, 11–12
leakage, 6, 55, 89, 143–184, 276–280, 290–297 microstructure, 223, 350g
adverse effects (summary), 166, 183 midvalue rule at discontinuities, 195–201
avoiding or eliminating, 147, 164, 166, 168, CFT, 195–196, 200
171–173, 178–182 DFT, 195–201, 242–243
detecting, 145–146, 179–180 Fourier series, 195–196
duality with aliasing in the other domain, modulation beating effect, see beating
178 (modulation) effect
in 1D, 143–158, 276–280 modulation theorem (or rule), see under
in 2D or MD, 165–174, 290–297 theorems (rules) for the DFT
origin of the term, 146 moiré effect, 350g
quantification of, 163 see also sampling moiré effect; pseudo moiré
reducing, 147, 166, 178 effect; singular moiré effect
shapes, 165–169
signal-domain interpretation, 143, 152, 155, N
166 nailbed, see impulse train: 2D or MD
spectral-domain interpretation, 143–146, non-causal function (or signal), 249g
149–155, 166 numeric integration rules:
that occurs in the signal domain, see signal- rectangular rule, 19, 78, 242, 319, 321
domain leakage trapezoidal rule, 242, 319, 321
372 Index

numeric precision, 241–242 pitfalls (continued)

see also under error pitfalls due to differences in CFT or DFT
Nyquist frequency, 73, 224, 345g definitions being used, 15, 43, 64–65,
78–79, 202–203
O pitfalls due to different definitions of the
odd function (or signal), 44, 203–206 sinc function, 352g
operator, 27, 29, 31, 92, 351g pitfalls due to inconsistent terminology, 335
organization, see reorganization pitfalls related to a bad vertical range choice
DFT-organization, 47, 51–60, 64–65 in the plot, 66–67, 241
display-organization, 47, 54, 55–60, 64–65 pitfalls related to cross correlation, 22–23
outliers, 66 pitfalls related to input reorganization, 47–
output array, 45, 69 51, 54–55, 61
length of, see under length pitfalls related to output reorganization, 45–
periodicity, see DFT: periodicity 46, 49, 61
reorganization, see under reorganization pitfalls related to symmetry, 203–210, 291
true units, see under true units pitfalls related to the computation of a
overshoot, 146, 153, 155, 214, 308–309, 314, continuous-world convolution by using
319 DFT, 60, 270–271
pitfalls related to the multiplication of
P spectra, 68
p (discrete counterpart of the period of a pitfalls related to the polar representation of
periodic function), 74–76, 147, 273–280 complex-valued entities, 202–203
in 2D or MD (p-vector), 118, 170–173, pitfalls related to the position of pixels
280–298 (centered or not), see under pixel
rational p-vector, 172, 296 pitfalls related to the representation of
parasite frequency, see aliased frequency discontinuities, 195–201
period, 340g stray impulses, see under impulse
number of full periods in the input array, true units along the axes, see under true
188–191, 273 units
of a periodic function, 74, 147, 162, 273 see also discrepancy between DFT and CFT;
discrete counterpart, see p (discrete DFT: artifacts; plotting considerations
counterpart of the period of a periodic pixel, 10, 12, 50, 61, 68, 171, 186, 214, 221–
function) 223, 235
radial period, 122, 124 centered or not, 68, 171, 186
period vector, 170–173, 281–285, 340g position of pixels with respect to the axes,
periodic function (or signal), 341g 68, 171, 186
periodicity theorem (or rule), see under planar graphic presentation, 8–11
theorems (rules) for the DFT plotting considerations, 241
phase, 34, 47, 51, 61, 143, 163, 187, 202–203, choice of the array size, 185–186, 223
347g choice of the vertical range to be plotted,
pitfalls: 66–67, 241
checkerboard effect (due to wrong input number of full periods in the input array,
reorganization), 61, 182 188–191
confusion between foldover and wraparound, output data reorganization, 46, 51–54, 66–68
see under foldover plotting a partial range, 186
DFT pitfalls, see under DFT printing or plotting artifacts, 10, 61, 122,
impulse height (amplitude), 247–264 196, 241
inappropriate data extension methods, 265– reconstruction issues, 223
272 true units along the axes, see under true
misinterpretation of foldover, see under units
foldover see also graphic presentation of discrete data
opposite ends of input or output arrays are plotting methods for discrete data, see graphic
in fact neighbours, 196, 240, 271 presentation of discrete data
Index 373

point, 7, 345g, 346g reconstruction, 219–223, 228–230, 234, 240

see also dot ideal, 220, 228–230, 234, 240, 245
point symmetry, 206 non-ideal, 221, 228–230
polar coordinates, 194 reconstruction aliasing (or error), 219–
polar representation of a complex-valued entity, 223
202 reconstruction artifacts, 214, 221–223, 244–
postaliasing, 219, 222 245
power spectrum, 51, 159 rect function, see pulse: square (or rectangular)
prealiasing, 219 pulse
preserving, 194, 352g rectangular pulse, see pulse: square (or
printer artifacts, see under printing rectangular) pulse
printing: reflection theorem (or rule), see under theorems
printing or plotting artifacts, 10, 61, 142, (rules) for the CFT; theorems (rules) for the
196, 241 DFT
printing considerations, see plotting reordering, see reorganization
considerations reorganization, 5, 45, 60–61, 66–68
product theorem (or rule), see under theorems input reorganization, 47–51, 54–55, 66–
(rules) for the CFT; theorems (rules) for the 68
DFT output reorganization, 45–47, 66–68, 300–
projection-slice theorem (or rule), see under 303
theorems (rules) for the CFT reorganizations in the IDFT, 51–54
pseudo moiré effect, 120, 132, 228–240, 302– wrong (superfluous or missing) input
306, 350 reorganization, 49, 61, 66
pseudo-3D graphic presentation, 8, 11–12 wrong (superfluous or missing) output
pulse, 343g reorganization, 49, 61, 66
square (or rectangular) pulse, 58, 198–199 see also under organization
triangular pulse, 324–329 repeat theorem (or rule), see under theorems
(rules) for the DFT
Q ringing, 167–168
q (discrete counterpart of the frequency of a see also ripple
periodic function), 74–76, 95, 147, 162, ripple, 144–146, 152, 167–168, 197, 309, 314,
164, 179, 273–280 319
in 2D or MD (q-vector), 170–173, 280–298 rotation, 116–123, 132–133, 137, 170, 179, 194
rational q-vector, 171–173, 296 rotation invariant, 194, 352g
quantization, 336g–337g rotation preserving, 194, 352g
quantized function (or signal), 338g rotation theorem (or rule), see under theorems
(rules) for the CFT
R rounding error, see under error
radial period, see under period rules:
random data, 51, 142 midvalue rule at discontinuities, see
range of a function, 345g midvalue rule at discontinuities
rational angle, see rational slope rules for numeric integration, see under
rational approximation, 179, 295 numeric integration rules
rational slope, 116, 118, 171–172, 179–181 rules for the CFT, see theorems (rules) for
rational vector, 171–172, 179, 296 the CFT
real part, 346g rules for the DFT, see theorems (rules) for
real part theorem (or rule), see under theorems the DFT
(rules) for the CFT; theorems (rules) for the symmetry rules for the CFT, see symmetry
DFT rules for the CFT
real-valued part, 346g symmetry rules for the DFT, see symmetry
real-world function (or signal), 340g rules for the DFT
374 Index

S spectrum (continued)
sampled signals: impulsive, 41–42, 74–77, 79, 95, 118–121,
graphic presentation, see graphic presentation 140–142, 145–146, 216, 248–258, 283,
of discrete data 344g
sampling, 336g line spectrum, 344g
in the signal domain, 31–32, 92 of a periodic function, 41–42, 321
in the spectral domain, 31–32 of an almost-periodic function, 41–42
non uniform, 194 of an aperiodic function, 41
sampling frequency, 72–73, 92, 192, 224 square pulse, see pulse: square (or rectangular)
sampling interval, 18, 70, 72, 148, 186–187 pulse
sampling moiré effect, 106, 113, 118, 120, 122– square wave function, 76–77, 88
132, 212–213, 218, 225, 235–239, 350g staircasing effect, see jaggies
sampling range, 47–49, 70, 72, 148, 187–189, standard amplitude scaling correction, 79, 80,
249–258 325–332
sampling theorem, 89, 219–223, 224–234, 240 in the case of impulsive spectra, 79, 82, 85,
spectral domain counterpart of, 134–136 248–264, 310–318, 324, 326–332
scalar product, 17, 20, 281, 285, 286, 289, 305 stray impulse, see under impulse
scaling theorem (or rule), see under theorems stretch theorem (or rule), see under theorems
(rules) for the CFT; theorems (rules) for the (rules) for the DFT
DFT sub-Nyquist artifact, 223–240, 245, 350g
separable function (or signal), 347g support-limited function (or signal), 348g
separable product theorem (or rule), see under symbol plot, 7–9, 11–12
theorems (rules) for the CFT; theorems symmetric function (or signal), 33–35, 203–207
(rules) for the DFT see also even function (or signal)
separable products theorem (or rule), see under symmetry related artifacts, 203–210, 243
theorems (rules) for the CFT; theorems symmetry rules for the CFT, 44, 203, 206
(rules) for the DFT symmetry rules for the DFT, 44, 203–210
sharp transition, 146, 153, 155, 165
shear theorem (or rule), see under theorems T
(rules) for the CFT tables of CFT pairs, see under CFT
shift theorem (or rule), see under theorems tapering function, see windowing function
(rules) for the CFT; theorems (rules) for the terminology, 335
DFT test for detecting aliased frequencies, 132–133
sidelobes, 144–148, 158–159, 162–163, 165, theorems (rules) for the CFT, 15–17, 43–44,
314, 319 335
signal, 1, 336g addition, 22, 23, 25, 254, 260
signal-domain aliasing, 33, 133–136, 178 affine, 25–26
signal-domain leakage, 165, 174–178 autocorrelation, 22, 24, 25
sinc function, 70, 352g complex conjugation, 22, 24, 25
sine integral function, 153, 307–308, 352g convolution, 22, 24, 25
singular moiré effect, 122–124, 132, 350g cross correlation, 22, 24, 25
size of the input and output arrays, see under dilation, 22, 23, 25, 44, 133
length frequency dilation, 22, 23, 25
smearing, 144–146, 152, 159, 163 frequency shift, 22, 24, 25
sombrero function, 165–167 imaginary part, 22, 24, 25
spatially-limited function (or signal), 348g inversion, 22, 24, 25
spectral periodicity theorem (or rule), see under linearity, 22, 23, 25
theorems (rules) for the DFT product, 22, 24, 25
spectral resolution, see DFT: spectral resolution projection-slice, 1
spectrum, 1, 341g real part, 22, 24, 25
continuous, 41, 342g reflection, 22, 24, 25
hybrid (having both continuous and rotation, 1, 24, 44, 132–133, 334
impulsive components), 254–260, 324 scaling, 22, 23, 25
Index 375

theorems (rules) for the CFT (continued) truncation, 31–34, 134–136, 143–161, 174, 257,
separable product, 1, 24, 25, 332–333 261
separable products, 25 excessive, 159–161
shear, 1, 24 truncation window, 34, 36, 143, 149, 153, 174,
shift (or translation), 22, 24, 25, 34, 54, 143 257, 261
theorems (rules) for the DFT, 26–31, 43–44,
335 U
addition, 26, 28, 30 unbounded function (or signal), 348g
autocorrelation, 27, 29, 30 underlying continuous function, 7, 36, 86, 89,
complex conjugation, 27, 28, 30 143, 183, 187, 200, 336g
cross correlation, 27, 29, 30 undersampling, 89, 96, 99, 105, 110–112
cyclic convolution, 27, 28, 30, 60, 65 in the spectral domain, 134
dilation, 27, 28, 30, 44 in 2D, 114–115
frequency shift, 27, 28, 30 units:
imaginary part, 27, 29, 30 in the signal domain, 15–16, 18
inversion, 27, 28, 30 in the spectral domain, 15–16, 18
linearity, 26, 28, 30 see also true units
modulation, 252–253
periodicity, 27, 29, 30–31 V
product, 27, 28, 30 vector, 17, 20, 346g
real part, 27, 29, 30 frequency, see frequency vector
reflection, 27, 28, 30 period, see period vector
repeat, 27, 29, 30–31, 188 p-vector, see under p
rotation (explanation why it does not work), q-vector, see under q
132–133, 194 rational, see rational vector
scaling, 26, 28, 30 vector division, 20
separable product, 29, 30 vector product, 283, 285, 289
separable products, 30 vector-valued function (or signal), 336g
spectral periodicity, 27, 29, 30–31 vertical line plot, 8–9, 11–12
shift (or translation), 27, 28, 30, 47, 54, 61, visibility circle, 342g
187
stretch (or zero packing), 27, 29, 30–31 W
time-limited function (or signal), see finite- wake, 344g
duration function (or signal) windowing function, 147–148
transform, 351g in 2D or MD, 165–169
transformation, 351g wireframe plot, 11–12
transformation rules for the CFT, see theorems wraparound, 46, 54, 94, 164, 168, 172–174, 207,
(rules) for the CFT 278–280, 294–298
transformation rules for the DFT, see theorems vs. foldover, 240, 245–246, 296–298
(rules) for the DFT
transient function (or signal), 349g Z
transition rules between different CFT zero crossing points, 158, 197, 199
definitions, see under conversion rules zero packing, 27, 29, 30–31, 188, 268
transition rules between different DFT zero packing theorem (or rule), see stretch under
definitions, see under conversion rules theorems (rules) for the DFT
translation theorem (or rule), see shift under zero padding, 60, 270–271
theorems (rules) for the CFT; theorems
(rules) for the DFT 0–9
triangular pulse, see under pulse 1D, 1, 89, 113, 165, 349g
triangular wave function, 76, 88, 110, 324–332 2D, 1, 113, 165, 204–210, 280–281, 349g
true units: 2D 1-fold periodic function, 170–172, 281–282,
for the input array, 18, 69–72 341g
for the output array, 18, 72–74, 84, 86 2D 2-fold periodic function, 172–173, 281, 284,
for the vertical axis, 78–79, 84, 86 288, 341g