Quaternion and Clifford Fourier Transforms describes the development of quaternion and Clifford Fourier transforms in Cl
133 109 62MB
English Pages 472 [475] Year 2021
Table of contents :
Cover
Half Title
Title Page
Copyright Page
Contents
Foreword
Preface
Acknowlegements
Chapter 1: Introduction
1.1. BRIEF HISTORICAL NOTES
1.1.1. Definitions related to Clifford’s geometric algebras
1.2. QUATERNION FOURIER TRANSFORMS (QFT)
1.2.1. Major developments in the history of the quaternion Fourier transform
1.2.2. Towards splitting quaternions and the QFT
1.3. DEVELOPMENT OF CLIFFORD FOURIER TRANSFORMATIONS IN CLIFFORD’S GEOMETRIC ALGEBRA
1.3.1. On the role of Clifford algebra square roots of 1 for Clifford Fourier transformations
1.3.2. Clifford Fourier transform in the light of Clifford analysis
1.4. DEVELOPMENT OF QUATERNION AND CLIFFORD WAVELETS
1.4.1. Towards Clifford wavelets in Clifford analysis
1.4.2. Further developments in quaternion and Clifford wavelet theory
Chapter 2: Clifford algebra
2.1. AXIOMS OF CLIFFORD’S GEOMETRIC ALGEBRA
2.1.1. Axioms for geometric algebra Rp,q = Cl(p, q)
2.1.1.1. Algebra over field of real numbers
2.1.1.2. Definition (1): using quadratic form
2.1.1.3. Definition (2): by basic multiplication rules
2.1.1.4. Definition (3): compact form
2.1.1.5. Grade r subspaces
2.1.1.6. Reverse and principal reverse
2.1.2. Geometric algebra R2 = Cl(2, 0)
2.1.2.1. Complex numbers
2.1.2.2. Reflections and rotations
2.1.2.3. Twodimensional point groups
2.1.2.4. Clifford’s geometric algebra G3 = Cl(3, 0) of R3
2.2. QUADRATIC FORMS IN CLIFFORD’S GEOMETRIC ALGEBRA
2.2.1. Definition of geometric algebra with quadratic form
2.2.2. Examples of quadratic forms and associated geometric algebras
2.2.3. Geometric algebras with quadratic form of signature p = q
2.2.3.1. A new interpretation of the geometric algebra of the Minkowski plane
2.2.3.2. Generalizing to the geometric mother algebra with p = q = n
2.3. CLIFFORD’S GEOMETRIC PRODUCT AND DERIVED PRODUCTS
2.3.1. The geometric product of multivectors
2.3.2. The scalar product
2.3.3. The outer product
2.3.3.1. The cross product of three dimensions
2.3.3.2. Linear dependence and independence
2.3.4. Right and left contraction
2.4. DETERMINANTS IN GEOMETRIC ALGEBRA
2.4.1. Determinant definition
2.4.2. Adjoint and inverse linear maps
2.5. GRAMSCHMIDT ORTHOGONALIZATION IN GEOMETRIC ALGEBRA
2.6. IMPORTANT CLIFFORD GEOMETRIC ALGEBRAS
2.6.1. Overview of Clifford’s geometric algebra of the Euclidean plane
2.6.2. Example of Cl(2, 0)
2.6.2.1. Aspects of algebraic unification and vector inverse
2.6.2.2. On geometric operations and transformations
2.6.2.3. Concepts of vectors, kvectors and multivectors
2.6.2.4. Higher dimensional types of inner and outer products for multivectors
2.6.3. Overview of Clifford’s geometric algebra of the 3D Euclidean space
2.6.3.1. Explicit multiplication table and important subalgebras of Cl(3, 0)
2.6.3.2. Concepts of grade structure of Cl(3, 0) and duality
2.6.3.3. Concept of blade subspaces of Cl(3, 0)
2.6.3.4. Useful algebraic formulas in Clifford’s geometric algebras
2.6.4. Extending Clifford’s geometric algebra to spacetime (STA)
2.6.5. Geometric modeling with conformal extension
2.6.5.1. On points, planes and motors in Cl(4, 1)
2.6.5.2. Modeling geometric objects in Cl(4, 1) with blades
2.6.6. On Clifford analysis
2.7. HOW IMAGINARY NUMBERS BECOME REAL IN CLIFFORD ALGEBRAS
2.7.1. What is an imaginary number?
2.7.2. What about imaginary eigenvalues and complex eigenvectors?
2.7.2.1. Two real dimensions
2.7.2.1.1. Complex treatment
2.7.2.1.2. Real explanation
2.7.2.2. Three real dimensions
2.7.2.2.1. Complex treatment of three dimensions
2.7.2.2.2. Real explanation for three dimensions
2.7.3. From imaginary numbers to multivector square roots of 1 in Clifford geometric algebras Cl(p, q) with p + q
Quaternion and Clifford Fourier Transforms
Quaternion and Clifford Fourier Transforms
Eckhard Hitzer
International Christian University, College of Liberal Arts
First edition published 2022 by CRC Press 6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 334872742 and by CRC Press 2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN © 2022 Eckhard Hitzer CRC Press is an imprint of Taylor & Francis Group, LLC Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, access www.copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 9787508400. For works that are not available on CCC please contact [email protected] Trademark notice: Product or corporate names may be trademarks or registered trademarks and are used only for identification and explanation without intent to infringe. ISBN: 9780367774660 (hbk) ISBN: 9781032026589 (pbk) ISBN: 9781003184478 (ebk) DOI: 10.1201/9781003184478 Typeset in LatinModern by KnowledgeWorks Global Ltd. Access the Support Material (color figures) at www.routledge.com/9780367774660
Contents Foreword
xvii
Preface
xix
Acknowlegements
xxi
Chapter 1.1 1.2
1 Introduction BRIEF HISTORICAL NOTES
2
1.1.1
3
1.2.2
2.1
Major developments in the history of the quaternion Fourier transform
4
Towards splitting quaternions and the QFT
6 6
On the role of Clifford algebra square roots of −1 for Clifford Fourier transformations
7
Clifford Fourier transform in the light of Clifford analysis
9
1.3.2
Chapter
4
CLIFFORD FOURIER TRANSFORMS IN CLIFFORD’S GEOMETRIC ALGEBRA
1.3.1
1.4
Definitions related to Clifford’s geometric algebras
QUATERNION FOURIER TRANSFORMS (QFT)
1.2.1
1.3
1
QUATERNION AND CLIFFORD WAVELETS
12
1.4.1
Towards Clifford wavelets in Clifford analysis
12
1.4.2
Further developments in quaternion and Clifford wavelet theory
13
2 Clifford algebra AXIOMS OF CLIFFORD ALGEBRA
2.1.1
15 15
Axioms for geometric algebra Rp,q = Cl(p, q)
15
2.1.1.1
Algebra over field of real numbers
15
2.1.1.2
Definition (1): using quadratic form
17
2.1.1.3
Definition (2): by basic multiplication rules
17
2.1.1.4
Definition (3): compact form
19 v
vi Contents
2.1.2
2.1.1.5
Grade r subspaces
19
2.1.1.6
Reverse and principal reverse
19
Geometric algebra R2 = Cl(2, 0)
20
2.1.2.1
Complex numbers
20
2.1.2.2
Reflections and rotations
20
2.1.2.3
Twodimensional point groups
2.1.2.4 2.2
23
2.2.1
Definition of geometric algebra with quadratic form
23
2.2.2
Examples of quadratic forms and associated geometric algebras
24
Geometric algebras with quadratic form of signature p = q
25
2.2.3.1 2.2.3.2
2.5 2.6
A new interpretation of the geometric algebra of the Minkowski plane
25
Generalizing to the geometric mother algebra with p=q=n
26
CLIFFORD PRODUCT AND DERIVED PRODUCTS
27
2.3.1
The geometric product of multivectors
27
2.3.2
The scalar product
31
2.3.3
The outer product
32
2.3.3.1
The cross product of three dimensions
33
2.3.3.2
Linear dependence and independence
33
2.3.4 2.4
23
QUADRATIC FORMS IN CLIFFORD’S GEOMETRIC ALGEBRA
2.2.3
2.3
Clifford’s geometric algebra G3 = Cl(3, 0) of R
22 3
Right and left contraction
35
DETERMINANTS IN GEOMETRIC ALGEBRA
37
2.4.1
Determinant definition
37
2.4.2
Adjoint and inverse linear maps
38
GRAMSCHMIDT ORTHOGONALIZATION IN GEOMETRIC ALGEBRA
39
IMPORTANT CLIFFORD GEOMETRIC ALGEBRAS
40
2.6.1
Overview of Clifford’s geometric algebra of the Euclidean plane 40
2.6.2
Example of Cl(2, 0)
41
2.6.2.1
Aspects of algebraic unification and vector inverse
42
2.6.2.2
On geometric operations and transformations
43
2.6.2.3
Concepts of vectors, kvectors and multivectors
45
2.6.2.4
Higher dimensional types of inner and outer products for multivectors
46
Contents vii
2.6.3
Overview of Clifford’s geometric algebra of the 3D Euclidean space 2.6.3.1
Explicit multiplication table and important subalgebras of Cl(3, 0)
48
2.6.3.2
Concepts of grade structure of Cl(3, 0) and duality
49
2.6.3.3
Concept of blade subspaces of Cl(3, 0)
50
2.6.3.4
Useful algebraic formulas in Clifford’s geometric algebras
54
2.6.4
Extending Clifford’s geometric algebra to spacetime (STA)
54
2.6.5
Geometric modeling with conformal extension
55
2.6.5.1
On points, planes and motors in Cl(4, 1)
55
2.6.5.2
Modeling geometric objects in Cl(4, 1) with blades
56
2.6.6 2.7
47
On Clifford analysis
57
HOW IMAGINARY NUMBERS BECOME REAL IN CLIFFORD ALGEBRAS
57
2.7.1
What is an imaginary number?
57
2.7.2
What about imaginary eigenvalues and complex eigenvectors? 60 2.7.2.1
2.7.2.2
2.7.3
2.7.4
Two real dimensions
60
2.7.2.1.1 Complex treatment
60
2.7.2.1.2 Real explanation
61
Three real dimensions
63
2.7.2.2.1 Complex treatment of three dimensions
63
2.7.2.2.2 Real explanation for three dimensions
64
From imaginary numbers to multivector square roots of −1 in Clifford geometric algebras Cl(p, q) with p + q ≤ 4
67
2.7.3.1
Notation for multivector square root computations
68
2.7.3.2
Geometric multivector square roots of −1
70
2.7.3.3
Summary on multivector square roots of −1 in Clifford algebras Cl(p, q), n = p + q ≤ 4
70
General theory of square roots of −1 in real Clifford algebras 74 2.7.4.1
Introductory remarks
74
2.7.4.2
Background and problem formulation
75
2.7.4.3
Square roots of −1 in M(2d, R)
78
2.7.4.4
2
Square roots of −1 in M(2d, R )
80
2.7.4.5
Square roots of −1 in M(d, H)
82
2.7.4.6
Square roots of −1 in M(d, H2 )
83
2.7.4.7
Square roots of −1 in M(2d, C)
84
viii Contents
2.7.4.8
2.7.4.9 2.8
Chapter
Square roots of −1 in conformal geometric algebra Cl(4, 1)
90
2.7.4.8.1 Ordinary square roots of −1 in Cl(4, 1) with k = 0
90
2.7.4.8.2 Exceptional square roots of −1 in Cl(4, 1) with k = 1
91
2.7.4.8.3 Exceptional square roots of −1 in Cl(4, 1) with k = −1
91
2.7.4.8.4 Exceptional square roots of −1 in Cl(4, 1) with k = ±2
91
Summary on general multivector square roots of −1 92
QUATERNIONS AND GEOMETRY OF ROTATIONS IN 3, 4 DIM.
93
2.8.1
Introduction to quaternions
93
2.8.2
Motivation for quaternion split
95
2.8.2.1
Splitting quaternions and knowing what it means
95
2.8.2.2
Even one pure unit quaternion can do a nice split
96
2.8.3
General orthogonal twodimensional planes split (OPS)
97
2.8.4
Coxeter on quaternions and reflections
98
2.8.5
Quaternion geometry of rotations analyzed by 2D OPS
100
2.8.6
Summary on quaternions and rotation geometry
103
3 Geometric calculus
105
3.1
INTRODUCTORY NOTES ON VECTOR DIFFERENTIAL CALCULUS 106
3.2
BRIEF OVERVIEW OF VECTOR DIFFERENTIAL CALCULUS
106
3.3
GEOMETRIC ALGEBRA FOR DIFFERENTIAL CALCULUS
108
3.4
VECTOR DIFFERENTIAL CALCULUS
112
3.5
SUMMARY ON VECTOR DIFFERENTIAL CALCULUS
139
Chapter 4.1
4 Quaternion Fourier transforms FUNDAMENTALS OF QUATERNION FOURIER TRANSFORMS (QFT)
4.1.1
141
141
Quaternion Fourier transform and the ±split
141
4.1.1.1
Background
142
4.1.1.1.1 Quaternions, split and module
142
The quaternion Fourier transform
143
4.1.1.2.1 Rewriting and splitting functions
144
4.1.1.2
Contents ix
4.1.1.2.2 Useful properties of the QFT
145
2
4.1.1.3
4.1.1.4
4.1.1.5 4.1.2
4.1.1.2.3 Example: GL(R ) transformation properties of the QFT
147
The rightside quaternion Fourier transform (QFTr)
148
4.1.1.3.1 Properties of the QFTr
149
2
Understanding the GL(R ) transformation properties of the QFT
151
4.1.1.4.1 Geometric interpretation and coordinate independent formulation of GL(R2 ) transformations of the QFT
152
4.1.1.4.2 Explicit examples: stretches, reflections & rotations
154
Summary on QFT and the ±split
154
QFT and orthogonal 2D planes split of quaternions
155
4.1.2.1
Background
155
4.1.2.2
Orthogonal planes split of quaternions with two orthonormal pure unit quaternions
155
General orthogonal 2D planes split
156
4.1.2.3
4.1.2.3.1 Orthogonal 2D planes split using two linearly independent pure unit quaternions 157
4.1.2.4
4.1.2.5
4.1.2.3.2 Orthogonal 2D planes split using one pure unit quaternion
159
4.1.2.3.3 Geometric interpretation of left and right exponential factors in f , g
164
4.1.2.3.4 Determination of f, g for given steerable pair of orthogonal 2D planes
164
New QFT forms: OPSQFTs with two pure unit quaternions f, g
166
4.1.2.4.1 Generalized OPS leads to new steerable type of QFT
166
4.1.2.4.2 Two phase angle version of QFT
168
Involutions and QFTs involving quaternion conjugation
169
4.1.2.5.1 Involutions involving quaternion conjugations
169
4.1.2.5.2 New steerable QFTs with quaternion conjugation and two pure unit quaternions f, g
173
x Contents
4.1.2.6 4.2
4.1.2.5.3 Local geometric interpretation of the QFT with quaternion conjugation
173
4.1.2.5.4 Phase angle QFT with respect to f, g, including quaternion conjugation
174
Summary on QFT and OPS of quaternions
175
PROPERTIES OF QUATERNION FOURIER TRANSFORM
176
4.2.1
Uncertainty principle for quaternion Fourier transform
176
4.2.1.1
Uncertainty principle
176
4.2.1.1.1 The Heisenberg uncertainty principle in physics
176
4.2.1.1.2 Directional uncertainty principle in Clifford geometric algebra
177
4.2.1.1.3 Componentwise uncertainty principle for rightsided QFT
178
Quaternion Fourier transform (QFT)
178
4.2.1.2.1 Simple complex forms for QFT of f±
179
4.2.1.2.2 Preparations for the full directional QFT uncertainty principle
179
4.2.1.3
Directional QFT uncertainty principle
180
4.2.1.4
Summary on directional QFT uncertainty principle 182
4.2.1.2
4.2.2
Quaternion Fourier transform and convolution
182
4.2.2.1
Background
182
4.2.2.1.1 Quaternions, commutators, notation orthogonal planes split
183
4.2.2.1.2 Review of general orthogonal twodimensional planes split (OPS)
183
4.2.2.2
The general twosided QFT
185
4.2.2.3
Convolution and Mustard convolution
186
4.2.2.3.1 Expressing the convolution in terms of the Mustard convolution
187
4.2.2.3.2 Expressing the Mustard convolution in terms of the convolution
189
Summary on QFT and convolution
191
4.2.2.4 4.2.3
Quaternionic Wiener Khintchine type theorems and convolution
192
4.2.3.1
Background
192
4.2.3.2
Convolution, crosscorrelation and autocorrelation of quaternion signals
193
Contents xi
4.2.3.3 4.2.3.4 4.2.3.5 4.3
The WienerKhintchine theorem for quaternion signals using the steerable twosided QFT
194
New spectral representation for convolution of quaternion signals
197
Summary on QFT WienerKhintchine theorems
199
SPECIAL QUATERNION FOURIER TRANSFORMS
199
4.3.1
Windowed quaternion Fourier transform (QWFT)
199
4.3.1.1
Background
199
4.3.1.2
Quaternionic windowed Fourier transform
202
4.3.1.2.1 Twodimensional WFT
202
4.3.1.2.2 Definition of the QWFT
203
4.3.1.2.3 Examples of the QWFT
205
4.3.1.2.4 Properties of the QWFT
207
4.3.1.3
Heisenberg’s uncertainty principle for the QWFT
209
4.3.1.4
Application of the QWFT
211
4.3.1.5
Summary on windowed QFT
215
Quaternionic Fourier Mellin transform
215
4.3.2.1
Background
215
4.3.2.2
The quaternionic Fourier Mellin transformations (QFMT)
215
4.3.2.2.1 Robert Hjalmar Mellin (1854–1933)
215
4.3.2.2.2 QFMT derivatives and power scaling
218
4.3.2
4.3.2.2.3 QFMT Plancherel and Parseval theorems 218 4.3.2.3 4.3.2.4 4.3.3
4.4
Symmetry and kernel structures of 2D FMT, FT, QFT, QFMT
219
Summary on quaternionic Fourier Mellin transformation
219
The Quaternion domain Fourier transform
221
4.3.3.1
Background
221
4.3.3.1.1 About quaternions H
223
4.3.3.1.2 Quaternions and reflections and rotations in three and four dimensions
224
4.3.3.1.3 Quaternion domain functions
225
4.3.3.2
The quaternion domain Fourier transform
226
4.3.3.3
Properties of the QDFT
228
4.3.3.4
Summary on QDFT
232
FROM QFT TO VOLUMETIME FT, SPACETIME FT
233
xii Contents
Chapter 5.1
4.4.1
QFT generalization to volumetime functions
233
4.4.2
Generalization to full spacetime algebra functions
235
4.4.3
Spacetime Fourier transform of f± split parts and physical interpretation
237
4.4.4
Directional 4D spacetime uncertainty principle
238
4.4.5
Summary of QFT generalization to spacetime FT
238
5 Clifford Fourier transforms OVERVIEW OF CLIFFORD FOURIER TRANSFORMS
241
5.1.1
Clifford Fourier transform (CFT)
242
5.1.1.1
Fourier transform in R
242
5.1.1.2
Clifford geometric algebra Fourier transform in G3 = Cl(3, 0)
243
5.1.2
Overview of basic CFT properties
245
5.1.3
Clifford Fourier transform and differentiation
245
5.1.4
Convolution and CFT
246
5.1.5
Plancherel and Parseval theorems
247
5.1.6
Precision limits – minimal variance products for geometric algebra Fourier transforms
247
Summary of CFT overview
249
5.1.7 5.2
241
ONESIDED CLIFFORD FOURIER TRANSFORMS
5.2.1
249
Onesided Clifford Fourier transforms in Cl(n, 0), n = 2 (mod 4) and n = 3 (mod 4)
249
5.2.1.1
249
Preliminaries
5.2.1.1.1 Clifford’s geometric algebra Gn = Cl(n, 0) of Rn 250
5.2.1.2
5.2.1.1.2 Multivector functions, vector differentiation and convolution in the context of onesided CFTs in Cl(n, 0)
251
Onesided Clifford Fourier transform (CFT)
255
5.2.1.2.1 Properties of the CFT for n = 2, 3 (mod 4) expressed independent of in commutations 256 5.2.1.2.2 Properties of the CFT for n = 2, 3 (mod 4) dependent on in commutations 261
5.2.2
5.2.1.3
Uncertainty principle
265
5.2.1.4
Summary on CFTs in Cl(n, 0), n = 2 (mod 4) and n = 3 (mod 4)
268
Onesided Clifford Fourier transforms in Cl(p, q)
268
Contents xiii
5.2.2.1
Background
268
5.2.2.1.1 Principal reverse and modulus in Cl(p, q) 270 5.2.2.1.2 Square roots of −1 in Clifford algebras 5.2.2.2
The onesided Clifford Fourier transform in Cl(p, q) 270
5.2.2.3
Properties of the onesided CFT in Cl(p, q)
271
5.2.2.3.1 Linearity, shift, modulation, dilation and powers of f, g, steerability
271
5.2.2.3.2 CFT inversion, moments, derivatives, Plancherel, Parseval
274
5.2.2.3.3 Convolution
277
Summary on onesided Clifford Fourier transforms in Cl(p, q)
278
Operator exponential approach to onesided Clifford Fourier transform
279
5.2.3.1
Motivation and background
279
5.2.3.2
Clifford analysis
280
5.2.3.3
The Clifford Fourier transform
283
5.2.2.4 5.2.3
5.3
TWOSIDED CLIFFORD FOURIER TRANSFORMS
5.3.1
290
Twosided Clifford Fourier transform with two square roots of −1 in Cl(p, q)
290
5.3.1.1
Background
290
5.3.1.2
The ± split with respect to two square roots of −1 291
5.3.1.3
General twosided Clifford Fourier transforms
292
5.3.1.4
Properties of the general twosided CFT
293
5.3.1.4.1 Linearity, shift, modulation, dilation and powers of f, g
294
5.3.1.4.2 CFT inversion, moments, derivatives, Plancherel, Parseval
294
5.3.1.4.3 Convolution
295
Reflection on twosided Clifford Fourier transform with two square roots of −1 in Cl(p, q)
296
5.3.1.5 5.3.2
270
Twosided (Clifford) geometric Fourier transform with multiple square roots of −1 in Cl(p, q)
297
5.3.2.1
Background
297
5.3.2.2
Definition of the (Clifford) geometric Fourier transform
297
5.3.2.3
General GFT properties
299
5.3.2.4
Bilinearity
301
xiv Contents
5.3.3
5.4
5.3.2.5
Products with invertible factors
302
5.3.2.6
Separable geometric Fourier transform
306
5.3.2.7
Summary on twosided (Clifford) geometric Fourier transform with multiple square roots of −1 in Cl(p, q)
313
Geometric Fourier transform and trigonometric transform with multiple square roots of −1 in Cl(p, q)
314
5.3.3.1
Background
314
5.3.3.2
The trigonometric transform (TT)
315
5.3.3.3
The true nature of separable geometric Fourier transforms
317
5.3.3.4
The explicitly invertible geometric Fourier transform 321
5.3.3.5
Summary on GFT and TT with multiple square roots of −1 in Cl(p, q)
325
CLIFFORD FOURIER TRANSFORM AND CONVOLUTION
325
5.4.1
Onesided CFT and convolution
325
5.4.1.1
Background and motivation
325
5.4.1.2
Convolution and steerable Mustard convolution
328
5.4.1.2.1 Expressing the convolution in terms of the Mustard convolution
329
5.4.1.2.2 Expressing the Mustard convolution in terms of the convolution
330
5.4.1.2.3 Single convolution product identities for classical and Mustard convolutions
331
Summary on onesided CFT and convolution
333
5.4.1.3 5.4.2
Twosided CFT with two square roots of −1 in Cl(p, q) and convolution
333
5.4.2.1
Background
333
5.4.2.2
Convolution and steerable Mustard convolution
337
5.4.2.2.1 Expressing the convolution in terms of the Mustard convolution
337
5.4.2.2.2 Expressing the Mustard convolution in terms of the convolution
341
Summary on twosided CFT with two square roots of −1 in Cl(p, q) and convolution
344
Twosided CFT with multiple square roots of −1 in Cl(p, q) and convolution
344
5.4.3.1
Background and notation
344
5.4.3.2
Convolution theorem for not coorthogonal exponents 346
5.4.2.3 5.4.3
Contents xv
5.4.3.3
5.5
Short convolution theorem for not coorthogonal exponents
348
5.4.3.4
Coorthogonality and bases
351
5.4.3.5
Products with basis blades
353
5.4.3.6
Geometric convolution theorem
357
5.4.3.7
Summary of twosided CFT with multiple square roots of −1 and convolution
363
SPECIAL CLIFFORD FOURIER TRANSFORMS
364
5.5.1
Windowed Clifford Fourier transform (CWFT)
364
5.5.1.1
Background
364
5.5.1.2
2D Clifford windowed Fourier transform
366
5.5.1.2.1 Definition of the CWFT
367
5.5.1.2.2 Properties of the CWFT
368
5.5.1.2.3 Examples of the CWFT
373
5.5.1.3
Comparison of CFT and CWFT
374
5.5.1.4
Summary on windowed CFT
375
5.5.2
Clifford Fourier Mellin transform
375
5.5.2.1
Background
375
5.5.2.2
The Clifford Fourier Mellin transformations (CFMT) 377 5.5.2.2.1 CFMT derivatives and power scaling
379
5.5.2.2.2 CFMT Plancherel and Parseval theorems 379
5.5.3
Chapter
5.5.2.3
Symmetry of the CFMT
380
5.5.2.4
Summary on Clifford Fourier Mellin transform
380
The quest for conformal geometric algebra Fourier transformations
380
5.5.3.1
Background
380
5.5.3.2
Prerequisites from conformal geometric algebra of the Euclidean plane
381
5.5.3.3
Clifford Fourier transformations in Cl(3, 1)
382
5.5.3.4
Summary on quest for conformal geometric algebra Fourier transformations
384
6 Relating QFTs and CFTs
387
6.1
BACKGROUND
387
6.2
388
6.3
GENERAL GEOMETRIC FOURIER TRANSFORM ¨ CFT DUE TO SOMMEN AND BULOW
6.4
COLOR IMAGE CFT
389
388
xvi Contents
6.5
TWOSIDED CFT
389
6.6
QUATERNION FOURIER TRANSFORM (QFT)
389
6.7
QUATERNION FOURIER STIELTJES TRANSFORM
390
6.8
QUATERNION FOURIER MELLIN TRANSFORM, CLIFFORD FOURIER MELLIN TRANSFORM 390
6.9
VOLUMETIME CFT AND SPACETIME CFT
390
6.10 ONESIDED CFTS
391
6.11 PSEUDOSCALAR KERNEL CFTS
391
6.12 QUATERNION AND CLIFFORD LINEAR CANONICAL TRANSFORMS
392
6.13 SUMMARY INTERRELATIONSHIP OF QFTS, CFTS
393
Appendix
A.1
A Square roots of −1 MAPLE, CauchySchwarz, uncertainty equality 395 SUMMARY OF ROOTS OF −1 IN C`(p, q) ∼ = M(2d, C) FOR d =
1, 2, 4
395
A.2
MULTIVECTOR CAUCHYSCHWARZ INEQUALITY
395
A.3
UNCERTAINTY EQUALITY FOR GAUSSIAN MV FUNCTIONS
397
Bibliography
401
Index
429
Foreword I have known the author of this book since 2010, and worked with him in person and remotely throughout the decade since. Eckhard Hitzer is an expert on Clifford algebras and in this book he has brought together a wide body of mathematical knowledge on the subject of hypercomplex Fourier transforms including those based on the algebra of quaternions (a special case of Clifford algebra), and on Clifford algebras in general. This represents a comprehensive collection of ideas and results by many authors over many years, as evidenced by the extensive list of references. They are brought together here in a consistent style and notation which will make this book a useful reference. This book should provide a sound mathematical basis for further research on Clifford Fourier transforms, both theoretical, and applied to problems in science and engineering. Stephen J. Sangwine Colchester, United Kingdom October 2020
xvii
Preface Dedicated to: Eriko, Jesaja and Joshua
Lift up your eyes to the heavens, look at the earth beneath; the heavens will vanish like smoke, the earth will wear out like a garment and its inhabitants die like flies. But my salvation will last forever, my righteousness will never fail. Isaiah 51:6 (NIV)
Jesus said, Heaven and earth will pass away, but My words will by no means pass away. Luke 21:33 (NKJV)
Der Mensch lebt und bestehet nur eine kleine Zeit, und alle Welt vergehet mit ihrer Herrlichkeit. Es ist nur Einer ewig und an allen Enden und wir in seinen H¨anden. Matthias Claudius
This book describes the development of quaternion and Clifford Fourier transforms in Clifford geometric algebra over the last 30 years. It begins with a historical introduction, Chapter 1, to quaternion Fourier transforms, Clifford Fourier transforms and their refinement in the form of wavelets. Then, Chapter 2 begins with the axiomatic formulation of Clifford algebras, the important role of quadratic forms, the fundamental Clifford (geometric) product of vectors and multivectors and convenient products derived from it, determinants in Clifford algebra, the method of GramSchmidt orthonormalization, a brief guide to the Clifford algebras of the plane, Euclidean space, spacetime and conformal geometric algebra. Chapter 2 concludes with the study of imaginary numbers in real Clifford algebras generalized to multivector square roots of −1, and the introduction
xix
xx Preface
to quaternions, of interest in its own and frequently appearing as subalgebra of higher dimensional Clifford algebras. Chapter 3 provides an elementary introduction to the basics of differential calculus with vectors. Chapter 4 is devoted to quaternion Fourier transforms (QFT). In many respects, due to its low dimensionality of four, and its nature as division algebra, quaternions are easier to handle as higher dimensional Clifford algebras, but they already display the essential feature of noncommutativity. While QFTs are by themselves of great interest, their knowledge helps to better grasp the treatment of Clifford Fourier transforms in Chapter 5. Chapter 4 first outlines the fundamentals of QFTs, including their important orthogonal twodimensional planes split. It then describes properties of the QFT such as uncertainty, convolution, Wiener Khinchine type theorems, special types of QFTs (windowed, FourierMellin, quaterniondomain), and the physicsrelated generalization to volumetime and spacetime Fourier transforms. Chapter 5 first provides an overview of Clifford Fourier transforms (CFT), studies the important case of CFTs in threedimensional Euclidean space, continues with generalization to onesided CFTs in quadratic space Clifford algebras Cl(p, q), and generalization to twosided CFTs. This is followed by the treatment of properties of CFTs such as uncertainty and convolution. Finally in this chapter, special CFTs (windowed, FourierMellin, conformal geometric algebra) are expounded. The last chapter, Chapter 6, endeavors to provide a holistic view on quaternionic and Clifford Fourier transforms, some important applied forms, and further extensions such as to Fourier Stieltjes and linear canonical transforms that appear in the recent literature. Appendix A provides a summary of Clifford square roots of −1, proofs the multivector CauchySchwartz inequality and the uncertainty equality in case of Gaussian multivector functions. Even though Chapters 2 and 3 lay the foundations of Clifford algebra and quaternions, and geometric calculus, we try to keep later chapters on more special topics reasonably self contained, so that readers with basic familiarity with quaternions and Clifford algebra will hopefully be able to begin reading directly in the chapter and section of their particular interest, without frequently needing to skip back and forth. Finally, Fourier transforms are a prime example of integral transforms. Readers who look for more background on integral transforms are referred to textbooks like [10, 87, 278].
Acknowlegements I gratefully acknowledge the fruitful scientific collaboration with the following colleagues: • R. Ablamowicz • S. Adji • R. Ashino • F. Brackx • R. Bujack • D. Eelbode • A. Hayashi • J. Helmstetter • M. Hlawitschka • I. Hotz • B. Mawardi • R. Nagaoka • S. J. Sangwine • G. Scheuermann • R. Vaillancourt Many results of our joint work are also described in this book. I thank my colleagues at the Department of Natural Science of International Christian University, especially in the Physics Major, for strongly supporting my current sabbatical research year. During this year with Airborne Hydro Mapping (AHM) GmbH in Innsbruck, Austria, I thank for excellent collaboration my colleagues at AHM; and my precious neighbours Evi and Hermann Jakschitz, Cinzia Zotti, Leopoldo d’Agostino, Johann Salchner and his family, the family of Benedikt Mair and Brigitte Mader, and Helga managing the Ladele, who provide a sort of home away from home. xxi
CHAPTER
1
Introduction
. . . for geometry, you know, is the gate of science, and the gate is so low and small that one can only enter it as a little child. (William K. Clifford [152]) But the gate to life is narrow and the way that leads to it is hard, and there are few people who find it. . . . I assure you that unless you change and become like children, you will never enter the Kingdom of heaven. (Jesus Christ [224])
At the beginning, we survey the historical development of quaternion and Clifford Fourier transforms and wavelets. This chapter is mainly based on [51]. The first quotation above stems from Clifford himself, who initially was a theologian and then became an atheist. But somehow his view of science was strongly colored by what Jesus taught as the Gospel about the Kingdom of God.1 To agree or disagree on what Clifford believed is a matter of faith and not of science. But I quite like his point, that geometry is like a gateway to a new understanding of science. The development of hypercomplex Fourier transforms and wavelets has taken place in several different threads, reflected in the overview of the subject presented in this chapter. We give in Section 1.2 an overview of the development of quaternion Fourier transforms, then in Section 1.3 the development of Clifford Fourier transforms. Additionally, wavelets are a more recent refinement of Fourier transforms, and the distinction between their quaternion and Clifford algebra approach has been much 1 It is interesting to note that a parallel to this exists even in the Japanese tea ceremony: “. . . enter the teahouse. The sliding door is only thirty six inches high. Thus all who enter must bow their heads and crouch. This door points to the reality that all are equal in tea, irrespective of status or social position.”[223]. The present form of the tea ceremony was established by Sen Rikyu in the 16th century. His wife is said to have been a secret Christian (Kirishitan), some think even Sen Rikyu was.
DOI: 10.1201/97810031844781
1
2 Quaternion and Clifford Fourier Transforms
less pronounced than in the case of Fourier transforms, Section 1.4 reviews the history of both quaternion and Clifford wavelets, although within this book we will not be able to treat wavelets in any depth. We recognize that the history presented here is bound to be incomplete, and that work by some authors will have been overlooked, for which we can only offer our humble apologies.
1.1
BRIEF HISTORICAL NOTES
W.K. Clifford (1845–1879), a young English Goldsmid professor of applied mathematics at the University College of London, published in 1878 in the American Journal of Mathematics Pure and Applied a nine page long paper on Applications of Grassmann’s Extensive Algebra [77]. In this paper, the young genius Clifford, standing on the shoulders of two giants of algebra: W.R. Hamilton (1805–1865), the inventor of quaternions [155, 156], and H.G. Grassmann (1809–1877), the inventor of extensive algebra [146], added the measurement of length and angle to Grassmann’s abstract and coordinate free algebraic methods for computing with a space and all its subspaces. Clifford thus unified and generalized in his geometric algebras (= Clifford algebras) the works of Hamilton and Grassmann by finalizing the fundamental concept of directed numbers [304]. Any Clifford algebra Cl(V ) is generated from an innerproduct2 vector space (V , a · b : a, b ∈ V 7→ R) by Clifford’s geometric product setting3 the geometric product4 of any vector with itself equal to their inner product: aa = a · a. We indeed have the universal property [244, 283] that any isometry5 from the vector space V into an innerproduct algebra6 A over the field7 K can be uniquely extended to an isometry8 from the Clifford algebra Cl(V ) into A. The Clifford algebra Cl(V ) is the unique associative and multilinear algebra with this property. Thus if we wish to generalize methods from algebra, analysis, calculus, differential geometry (etc.) of real numbers, complex numbers and quaternion algebra to vector spaces and multivector spaces (which include additional elements representing 2D up to nD subspaces, i.e. plane elements up to hypervolume elements), the study of Clifford algebras becomes unavoidable. Indeed, repeatedly and independently a long list of Clifford algebras, their subalgebras and in Clifford algebras embedded algebras (like octonions [247]) of many spaces have been studied and applied historically, often under different names. 2
The inner product defines the measurement of length and angle. This setting amounts to an algebra generating relationship. 4 No product sign will be introduced, simple juxtaposition implies the geometric product just like 2x = 2 × x. 5 A Kisometry between two innerproduct spaces is a Klinear mapping preserving the inner products. 6 A Kalgebra is a Kvector space equipped with an associative and multilinear product. An innerproduct Kalgebra is a Kalgebra equipped with an inner product structure when taken as Kvector space. 7 Important fields are real R and complex numbers C, etc. 8 That is a Klinear homomorphism preserving the inner products, i.e. a Klinear mapping preserving both the products of the algebras when taken as rings, and the inner products of the algebras when taken as innerproduct vector spaces. 3
Introduction 3
Some of these algebras are complex numbers (and the complex number plane), hyperbolic numbers (split complex numbers, real tessarines), dual numbers, quaternions, biquaternions (complex quaternions), dual quaternions, Pl¨ ucker coordinates, bicomplex numbers (commutative quaternions, tessarines, Segre quaternions), Pauli algebra (space algebra), Dirac algebra (spacetime algebra, Minkowski algebra), algebra of physical space, paravector algebra, spinor algebra, Lie algebras, Cartan algebra, versor algebra, rotor algebra, motor algebra, Clifford bracket algebra, conformal algebra, algebra of differential forms, etc. Section 1.1.1 can also be skipped by readers less interested in mathematical definitions. 1.1.1 Definitions related to Clifford’s geometric algebras
Definition of an algebra [316]: Let A be a vector space over the reals R with an additional binary operation from A × A to A, denoted here by ◦ (x ◦ y is the product of any x, y ∈ A). Then A is an algebra over R if the following identities hold ∀ x, y, z ∈ A, and “scalars” α, β ∈ R : (1,2) Left and right distributivity: (x+y)◦z = x◦z +y ◦z, x◦(y +z) = x◦y +x◦z. (3) Compatibility with scalars: (αx)◦(βy) = (αβ)(x◦y). This means that x◦y is bilinear. The binary operation is often referred to as multiplication in A, which is not necessarily associative. Definition of inner product space [2]: An inner product space is a vector space V over R together with an inner product map h., .i : V × V → R, that satisfies ∀ x, y, z ∈ V and ∀ α ∈ R: (1) Symmetry: hx, yi = hy, xi. (2) Linearity in the first argument: hαx, yi = αhx, yi, hx + y, zi = hx, zi + hy, zi. Note: We do not assume positive definiteness. Definition of inner product algebra: An inner product algebra is an algebra equipped with an inner product A × A → R. Definition of Clifford’s geometric algebra (GA) [131,247]: Let {e1 , e2 , . . . , ep , ep+1 , . . . , ep+q , ep+q+1 , . . ., en }, with n = p + q + r, e2k = εk , εk = +1 for k = 1, . . . , p, εk = −1 for k = p + 1, . . . , p + q, εk = 0 for k = p + q + 1, . . . , n, be an orthonormal base of the inner product vector space Rp,q,r with a geometric product according to the multiplication rules ek el + el ek = 2εk δk,l ,
k, l = 1, . . . n,
(1.1.1)
where δk,l is the Kronecker symbol with δk,l = 1 for k = l, and δk,l = 0 for k 6= l. This noncommutative product and the additional axiom of associativity generate the 2n dimensional Clifford geometric algebra Cl(p, q, r) = Cl(Rp,q,r ) = Clp,q,r = Gp,q,r = Rp,q,r over R. The set {eA : A ⊆ {1, . . . , n}} with eA = eh1 eh2 . . . ehk , 1 ≤ h1 < . . . < hk ≤ n, e∅ = 1, forms a graded (blade) basis of Cl(p, q, r). The grades k range from 0 for scalars, 1 for vectors, 2 for bivectors, s for svectors, up to n for pseudoscalars. The vector space Rp,q,r is included in Cl(p, q, r) as the subset of 1vectors. The general elements of Cl(p, q, r) are real linear combinations of basis blades eA , called Clifford numbers, multivectors or hypercomplex numbers. Note: The definition of Clifford’s GA is fundamentally coordinate system independent, i.e. coordinate free. Equation (1.1.1) is fully equivalent to the coordinate
4 Quaternion and Clifford Fourier Transforms
independent definition: aa = a · a, ∀ a ∈ Rp,q,r [161]. This even applies to Clifford analysis (Geometric Calculus). Clifford algebra is thus ideal for computing with geometrical invariants [244]. A review of five different ways to define GA, including one definition based on vector space basis element multiplication rules, and one definition focusing on Cl(p, q, r) as a universal associative algebra, is given in chapter 14 of the textbook [247]. In general hAik denotes the grade k part of A ∈ Cl(p, q, r). The parts of grade 0, (s − k), (k − s), and (s + k), respectively, of the geometric product of a kvector Ak ∈ Cl(p, q, r) with an svector Bs ∈ Cl(p, q, r) Ak ∗ Bs := hAk Bs i0 ,
Ak cBs := hAk Bs is−k ,
Ak bBs := hAk Bs ik−s ,
(1.1.2)
Ak ∧ Bs := hAk Bs ik+s ,
are called scalar product, left contraction (zero for s < k), right contraction (zero for k < s), and (associative) outer product, respectively. These definitions extend by linearity to the corresponding products of general multivectors. The various derived products of (1.1.2) are related, e.g. by (A ∧ B)cC = Ac(BcC),
∀ A, B, C ∈ Cl(p, q, r).
(1.1.3)
Note that for vectors a, b in Rp,q,r ⊂ Cl(p, q, r) we have ab = acb + a ∧ b,
acb = abb = a · b = a ∗ b,
(1.1.4)
where a · b is the usual inner product of Rp,q,r . For r = 0 we often denote Rp,q = Rp,q,0 , and Cl(p, q) = Cl(p, q, 0). For Euclidean vector spaces (n = p) we use Rn = Rn,0 = Rn,0,0 , and Cl(n) = Cl(n, 0) = Cl(n, 0, 0). The even grade subalgebra of Cl(p, q, r) is denoted by Cl+ (p, q, r), the kvectors of its basis have only even grades k. Every kvector B that can be written as the outer product B = b1 ∧ b2 ∧ . . . ∧ bk of k vectors b1 , b2 , . . . , bk ∈ Rp,q,r is called a simple kvector or blade. Definition of outermorphism [161,248]: An outermorphism is the unique extension to Cl(V ) of a vector space map for all a ∈ V , f : a 7→ f (a) ∈ V 0 , and is given by the mapping B = b1 ∧ b2 ∧ . . . ∧ bk → f (b1 ) ∧ f (b2 ) ∧ . . . ∧ f (bk ) in Cl(V 0 ), for every blade B in Cl(V ).
1.2 1.2.1
QUATERNION FOURIER TRANSFORMS (QFT) Major developments in the history of the quaternion Fourier transform
Quaternions were first applied to Fourier transforms by Ernst [130, Section 6.4.2] and Delsuc [102, Eqn. 20] in the late 1980s, seemingly without knowledge of the earlier work of Sommen [306, 307] on Clifford Fourier and Laplace transforms further explained in Section 1.3.2. Ernst and Delsuc’s quaternion transforms were twodimensional (that is they had two independent variables) and proposed for application to nuclear magnetic resonance (NMR) imaging. Written in terms of two independent
Introduction 5
time variables9 t1 and t2 , the forward transforms were of the following form: Z∞ Z∞
F (ω1 , ω2 ) =
f (t1 , t2 )eiω1 t1 ejω2 t2 dt1 dt2 .
(1.2.1)
−∞ −∞
Notice the use of different quaternion basis units i and j in each of the two exponentials, a feature that was essential to maintain the separation between the two dimensions (the prime motivation for using a quaternion Fourier transformation was to avoid the mixing of information that occurred when using a complex Fourier transform – something that now seems obvious, but must have been less so in the 1980s). The signal waveforms/samples measured in NMR are complex, so the quaternion aspect of this transform was essential only for maintaining the separation between the two dimensions. As we will see below, there was some unused potential here. The fact that exponentials in the above formulation do not commute (with each other, or with the ‘signal’ function f ), means that other formulations are possible, and indeed Ell in 1992 [123, 124] formulated a transform with the two exponentials positioned either side of the signal function: Z∞ Z∞
F (ω1 , ω2 ) =
eiω1 t1 f (t1 , t2 ) ejω2 t2 dt1 dt2 .
(1.2.2)
−∞ −∞
Ell’s transform was a theoretical development, but it was soon applied to the practical problem of computing a holistic Fourier transform of a color image [291] in which the signal samples (discrete image pixels) had threedimensional values (represented as quaternions with zero scalar parts). This was a major change from the previously intended application in nuclear magnetic resonance, because now the twodimensional nature of the transform mirrored the twodimensional nature of the image, and the fourdimensional nature of the algebra used followed naturally from the threedimensional nature of the image pixels. Other researchers in signal and image processing have followed Ell’s formulation (with trivial changes of basis units in the exponentials) [56, 57, 72], but as with the NMR transforms, the quaternion nature of the transforms was applied essentially to separation of the two independent dimensions of an image (B¨ ulow’s work [56,57] was based on greyscale images, that is with onedimensional pixel values). Two new ideas emerged in 1998 in a paper by Sangwine and Ell [292]. These were, firstly, the choice of a general root µ of −1 (a unit quaternion with zero scalar part) rather than a basis unit (i, j or k) of the quaternion algebra, and secondly, the choice of a single exponential rather than two (giving a choice of ordering relative to the quaternionic signal function): Z∞ Z∞
F (ω1 , ω2 ) =
eµ(ω1 t1 +ω2 t2 ) f (t1 , t2 ) dt1 dt2 .
(1.2.3)
−∞ −∞ 9 The two independent time variables arise naturally from the formulation of twodimensional NMR spectroscopy.
6 Quaternion and Clifford Fourier Transforms
This made possible a quaternion Fourier transform of a onedimensional signal: Z∞
F (ω) =
eµωt f (t) dt .
(1.2.4)
−∞
Such a transform makes sense only if the signal function has quaternion values, suggesting applications where the signal has three or four independent components. (An example is vibrations in a solid, such as rock, detected by a sensor with three mutually orthogonal transducers, such as a vector geophone.) Very little has appeared in print about the interpretation of the Fourier coefficients resulting from a quaternion Fourier transform. One interpretation are components of different symmetry, as explained by Ell in [128]. Sangwine and Ell in 2007 published a paper about quaternion Fourier transforms applied to color images, with a detailed explanation of the Fourier coefficients in terms of elliptical paths in color space (the ndimensional space of the values of the image pixels in a color image) [127]. 1.2.2 Towards splitting quaternions and the QFT
Following the earlier works of Ernst, Ell, Sangwine (see Section 1.2.1) and B¨ ulow [56, 57], Hitzer thoroughly studied the quaternion Fourier transform (QFT) applied to quaternionvalued functions in [184]. As part of this work a quaternion split 1 q± = (q ± iqj), q ∈ H, (1.2.5) 2 was devised and applied, which lead to a better understanding of GL(R2 ) transformation properties of the QFT spectrum of 2D images, including color images, and opened the way to a generalization of the QFT concept to a full spacetime Fourier transformation (SFT) for spacetime algebra Cl(3, 1)valued signals. This was followed up by the establishment of a fully directional (opposed to component wise) uncertainty principle for the QFT and the SFT [187]. Independently Mawardi et al. [255] established a componentwise uncertainty principle for the QFT. The QFT with a Gabor window was treated by B¨ ulow [56], a study which has been continued by Mawardi et al. in [256]. Hitzer reports in [194] initial results (obtained in cooperation with Sangwine) about a further generalization of the QFT to a general form of orthogonal 2D planes split (OPS) QFT, where the split (1.2.5) with respect to two orthogonal pure quaternion units i, j is generalized to a steerable split with respect to any two pure unit quaternions. This approach is fully elaborated upon in [195, 205].
1.3
DEVELOPMENT OF CLIFFORD FOURIER TRANSFORMATIONS IN CLIFFORD’S GEOMETRIC ALGEBRA
Regarding the literature, an introduction to the vector and multivector calculus used in the field of Clifford Fourier transforms (CFT) can be found in [174,175]. A tutorial introduction to CFTs and Clifford wavelet transforms can be found in [183]. The Clifford algebra application survey [202] contains an up to date section on applications of Clifford algebra integral transforms, including CFTs, QFTs and wavelet transforms.
Introduction 7
1.3.1 On the role of Clifford algebra square roots of −1 for Clifford Fourier transformations
In 1990, Jancewicz [222] defined a trivector Fourier transformation F3 {g}(ω) =
Z
g(x)e−i3 x·ω d3 x,
R3
g : R3 → Cl(3, 0),
i3 = e1 e2 e3 ,
(1.3.1)
for the electromagnetic field replacing the imaginary unit i ∈ C by the trivector i3 of the geometric algebra Cl(3, 0) of threedimensional Euclidean space R3 = R3,0 with orthonormal vector basis {e1 , e2 , e3 }. In [132] Felsberg makes use of signal embeddings in low dimensional Clifford algebras R2,0 and R3,0 to define his CliffordFourier transform (CFT) for onedimensional signals as F1f e [f ](u)
Z
=
exp (−2πi2 u x) f (x) dx,
i2 = e1 e2 ,
f : R → R,
(1.3.2)
R
where he uses the pseudoscalar i2 ∈ Cl(2, 0). For twodimensional signals he defines the CFT as F2f e [f ](u)
Z
= R2
exp (−2πi3 < u, x >) f (x) dx,
f : R2 → R2 ,
(1.3.3)
where he uses the pseudoscalar i3 ∈ Cl(3, 0). It is used a.o. to introduce a concept of twodimensional analytic signal. Together with B¨ ulow, and Sommer, Felsberg applied these CFTs to image structure processing (keynotion: structure multivector) [56, 132]. Ebling and Scheuermann [117,118] consequently applied to vector signal processing in two and three dimensions, respectively, the following twodimensional CFT F2 {f }(ω) =
Z R2
f (x)e−i2 x·ω d2 x,
f : R2 → R2 ,
(1.3.4)
with Clifford Fourier kernel exp (−e1 e2 (ω1 x1 + ω2 x2 )),
(1.3.5)
and the threedimensional CFT (1.3.1) of Jancewicz with Clifford Fourier kernel exp (−e1 e2 e3 (ω1 x1 + ω2 x2 + ω3 x3 )).
(1.3.6)
An important integral operation defined and applied in this context by Ebling and Scheuermann was the Clifford convolution. These CliffordFourier transforms and the corresponding convolution theorems allow Ebling and Scheuermann for a.o. the analysis of vectorvalued patterns in the frequency domain. Note that the latter Fourier kernel (1.3.6) has also been used by Mawardi and Hitzer in [181,251,251] to define their CliffordFourier transform of threedimensional multivector signals: that means, they researched the properties of F3 {g}(ω) in detail when applied to full multivector signals g : R3 → Cl(3, 0). This included an
8 Quaternion and Clifford Fourier Transforms
investigation of the uncertainty inequality for this type of CFT. They subsequently generalized F3 {g}(ω) to dimensions n = 3(mod 4), i.e. n = 3, 7, 11, . . ., Fn {g}(ω) =
Z
g(x)e−in x·ω dn x,
Rn
g : Rn → Cl(n, 0),
(1.3.7)
which is straight forward, since for these dimensions the pseudoscalar in = e1 . . . en squares to −1 and is central [182], i.e. it commutes with every other multivector belonging to Cl(n, 0). A little less trivial is the generalization of F2 {f }(ω) to Fn {f }(ω) =
Z Rn
f (x)e−in x·ω dn x,
f : Rn → Cl(n, 0),
(1.3.8)
with n = 2(mod 4), i.e. n = 2, 6, 10 . . ., because in these dimensions the pseudoscalar in = e1 . . . en squares to −1, but it ceases to be central. So the relative order of the factors in Fn {f }(ω) becomes important, see [185] for a systematic investigation and comparison. In the context of generalizing quaternion Fourier transforms (QFT) via algebra isomorphisms to higher dimensional Clifford algebras, Hitzer [184] constructed a spacetime Fourier transform (SFT) in the full algebra of spacetime Cl(3, 1), which includes the CFT (1.3.1) as a partial transform of space. Implemented analogous (isomorphic) to the orthogonal 2D planes split of quaternions, the SFT permits a natural spacetime split, which algebraically splits the SFT into right and left propagating multivector wave packets. This analysis allows to compute the effect of Lorentz transformations on the spectra of these wavepackets, as well as a 4D directional spacetime uncertainty formula [187] for spacetime signals. Mawardi et al. extended the CFT F2 {f }(ω) to a windowed CFT in [254]. Fu et al. establish in [136] a strong version of Heisenberg’s uncertainty principle for Gaborwindowed CFTs. In [61] Bujack, Scheuermann and Hitzer expand the notion of Clifford Fourier transform to include multiple left and right exponential kernel factors, in which commuting (or anticommuting) blades, that square to −1, replace the complex unit i ∈ C, thus managing to include most practically used CFTs in a single comprehensive framework. Based on this, they have also constructed a general CFT convolution theorem [59]. Spurred by the systematic investigation of (complex quaternion) biquaternion square roots of −1 in Cl(3, 0) by Sangwine [294], Hitzer and Ablamowicz [193] systematically investigated the explicit equations and solutions for square roots of −1 in all real Clifford algebras Cl(p, q), p + q ≤ 4. This investigation has been continued by Hitzer, Helmstetter and Ablamowicz [203] for all square roots of −1 in all real Clifford algebras Cl(p, q) without restricting the value of n = p + q. One important motivation for this is the relevance of the Clifford algebra square roots of −1 for the general construction of CFTs, where the imaginary unit i ∈ C is replaced by a √ −1 ∈ Cl(p, q), without restriction to pseudoscalars or blades. Based on the knowledge of square roots of −1 in real Clifford algebras Cl(p, q), [201] develops a general CFT in Cl(p, q), wherein the complex unit i ∈ C is replaced by any square root of −1 chosen from any component and (or) conjugation class of the
Introduction 9
submanifold of square roots of −1 in Cl(p, q), and details its properties, including a convolution theorem. A similar general approach is taken in [196] for the construction of twosided CFTs in real Clifford algebras Cl(p, q), freely choosing two square roots from any one or two components and (or) conjugation classes of the submanifold of square roots of −1 in Cl(p, q). These transformations are therefore generically steerable. This algebraically motivated approach may eventually be favourably combined with group theoretic, operator theoretic and spinorial approaches, to be reviewed in the following. 1.3.2 Clifford Fourier transform in the light of Clifford analysis
Two robust tools used in image processing and computer vision for the analysis of scalar fields are convolution and Fourier transformation. Several attempts have been made to extend these methods to two and threedimensional vector fields and even multivector fields. Let us give an overview of those generalized Fourier transforms. In [57] B¨ ulow and Sommer define a socalled quaternionic Fourier transform of twodimensional signals f (x1 , x2 ) taking their values in the algebra H of real quaternions. Note that the quaternion algebra H is nothing else but (isomorphic to) the Clifford algebra R0,2 = Cl(0, 2) where, traditionally, the basis vectors are denoted by i and j, with i2 = j 2 = −1, and the bivector by k = ij. In terms of these basis vectors, this quaternionic Fourier transform takes the form F q [f ](u1 , u2 ) =
Z R2
exp (−2πiu1 x1 ) f (x1 , x2 ) exp (−2πju2 x2 ) dx
(1.3.9)
Due to the noncommutativity of the multiplication in H, the convolution theorem for this quaternionic Fourier transform is rather complicated, see also [59]. This is also the case for its higher dimensional analogue, the socalled Clifford Fourier transform given by cl
F [f ](u) =
Z Rm
f (x) exp (−2πe1 u1 x1 ) . . . exp (−2πem um xm ) dx
(1.3.10)
Note that for m = 1 and interpreting the Clifford basis vector e1 as the imaginary unit i, the Clifford Fourier transform (1.3.10) reduces to the standard Fourier transform on the real line, while for m = 2 the quaternionic Fourier transform (1.3.9) is recovered when restricting to real signals. Finally, B¨ ulow and Sommer also introduce a socalled commutative hypercomplex Fourier transform given by h
F [f ](u) =
Z Rm
f (x) exp −2π
Pm
˜ j uj xj j=1 e
dx
(1.3.11)
˜ m ) obey the commutative multiplication rules where the basis vectors (˜ e1 , . . . , e ˜j e ˜k = e ˜k e ˜ j , j, k = 1, . . . , m, while still retaining e ˜ 2j = −1, j = 1, . . . , m. This e commutative hypercomplex Fourier transform offers the advantage of a simple convolution theorem.
10 Quaternion and Clifford Fourier Transforms
The hypercomplex Fourier transforms F q , F cl and F h enable B¨ ulow and Sommer to establish a theory of multidimensional signal analysis and in particular to introduce the notions of multidimensional analytic signal, Gabor filter, instantaneous and local amplitude and phase, etc. In this context the Clifford Fourier transformations by Felsberg [132] for one and twodimensional signals, by Ebling and Scheuermann for two and threedimensional vector signal processing [117,118], and by Mawardi and Hitzer for general multivector signals in Cl(3, 0) [181, 251, 251], and their respective kernels, as already reviewed in Section 1.3.1, should also be considered. The above mentioned Clifford Fourier kernel of B¨ ulow and Sommer exp (−2πe1 u1 x1 ) · · · exp (−2πem um xm )
(1.3.12)
was in fact already introduced in [33] and [305] as a theoretical concept in the framework of Clifford analysis. This generalized Fourier transform was further elaborated by Sommen in [306, 307] in connection with similar generalizations of the Cauchy, Hilbert and Laplace transforms. In this context also the work of Li, McIntosh and Qian should be mentioned; in [243] they generalize the standard multidimensional D E Fourier transform of a function in Rm , by extending the Fourier kernel exp i ξ, x to a function which is holomorphic in Cm and monogenic in Rm+1 . In [40,42,46] Brackx, De Schepper and Sommen follow another philosophy in their construction of a Clifford Fourier transform. One of the most fundamental features of Clifford analysis is the factorization of the Laplace operator. Indeed, whereas in general the square root of the Laplace operator is only a pseudodifferential√operator, by embedding Euclidean space into a Clifford algebra, one can realize −∆m as the Dirac operator ∂x . In this way Clifford analysis spontaneously refines harmonic analysis. In the same order of ideas, Brackx et al. decided to not replace nor to improve the classical Fourier transform by a Clifford analysis alternative, since a refinement of it automatically appears within the language of Clifford analysis. The key step to making this refinement apparent is to interpret the standard Fourier transform as an operator exponential: ∞ X π 1 π F = exp −i H = −i 2 k! 2 k=0
k
Hk ,
(1.3.13)
where H is the scalar operator 1 −∆m + r2 − m (1.3.14) 2 This expression links the Fourier transform with the Lie algebra sl2 generated by ∆m and r2 = x2 and with the theory of the quantum harmonic oscillator determined by the Hamiltonian − 12 ∆m − r2 . Splitting the operator H into a sum of Clifford algebra–valued second order operators containing the angular Dirac operator Γ, one is led, in a natural way, to a pair of transforms FH± , the harmonic average of which is precisely the standard Fourier transform:
H=
FH± = exp
iπm 4
exp ∓
iπΓ 2
exp
iπ ∆m − r 2 4
(1.3.15)
Introduction 11
For the special case of dimension 2, Brackx et al. obtain a closed form for the kernel of the integral representation of this Clifford Fourier transform leading to its internal representation 1 FH± [f ](ξ) = FH± [f ](ξ 1 , ξ 2 ) = 2π
Z R2
exp ±e12 (ξ 1 x2 − ξ 2 x1 ) f (x) dx
(1.3.16)
which enables the generalization of the calculation rules for the standard Fourier transform both in the L1 and in the L2 context. Moreover, the CliffordFourier transform of Ebling and Scheuermann F e [f ](ξ) =
Z R2
exp (−e12 (x1 ξ 1 + x2 ξ 2 )) f (x) dx
(1.3.17)
can be expressed in terms of the CliffordFourier transform: F e [f ](ξ) = 2π FH± [f ](∓ξ 2 , ±ξ 1 ) = 2π FH± [f ](±e12 ξ),
(1.3.18)
taking into account that, under the isomorphism between the Clifford algebras R2,0 and R0,2 , both pseudoscalars are isomorphic images of each other. The question whether FH± can be written as an integral transform is answered positively in the case of even dimension by De Bie and Xu in [90]. The integral kernel of this transform is not easy to obtain and looks quite complicated. In the case of odd dimension the problem is still open. Later, in [92], De Bie and De Schepper have studied the fractional Clifford Fourier transform as a generalization of both the standard fractional Fourier transform and the Clifford Fourier transform. It is given as an operator exponential by iαm = exp 2
Fα,β
iα exp (iβΓ) exp ∆m − r2 . 2
(1.3.19)
For the corresponding integral kernel a series expansion is obtained, and, in the case of dimension 2, an explicit expression in terms of Bessel functions. The above, more or less chronological, overview of generalized Fourier transforms in the framework of quaternionic and Clifford analysis gives the impression of a medley of ad hoc constructions. However, there is a structure behind some of these generalizations, which becomes apparent when, as already slightly touched upon above, the Fourier transform is linked to group representation theory, in particular the Lie algebras sl2 and osp(12). This unifying character is beautifully demonstrated by De Bie in the overview paper [94], where, next to an extensive bibliography, also new results on some of the transformations mentioned below can be found. It is shown that using realizations of the Lie algebra sl2 , one is lead to scalar generalizations of the Fourier transform, such as: (i) The fractional Fourier transform, which is, as the standard Fourier transform, invariant under the orthogonal group; this transform has been reinvented several times as well in mathematics as in physics, and is attributed to Namias [272], Condon [81], Bargmann [16], Collins [79], Moshinsky and Quesne [266]; for a detailed overview of the theory and recent applications of the fractional Fourier transform we refer the reader to [276];
12 Quaternion and Clifford Fourier Transforms
(ii) The Dunkl transform, see e.g. [113], where the symmetry is reduced to that of a finite reflection group; (iii) The radially deformed Fourier transform, see e.g. [235], which encompasses both the fractional Fourier and the Dunkl transform; (iv) The super Fourier transform, see e.g. [83, 88], which is defined in the context of superspaces and is invariant under the product of the orthogonal with the symplectic group. Realizations of the Lie algebra osp(12), on the contrary, need the framework of Clifford analysis, and lead to: (v) The Clifford–Fourier transform and the fractional Clifford–Fourier transform, both already mentioned above; meanwhile an entire class of Clifford–Fourier transforms has been thoroughly studied in [91]; (vi) The radially deformed hypercomplex Fourier transform, which appears as a special case in the theory of radial deformations of the Lie algebra osp(12), see [93, 96], and is a topic of current research, see [98].
1.4 1.4.1
DEVELOPMENT OF QUATERNION AND CLIFFORD WAVELETS Towards Clifford wavelets in Clifford analysis
The interest of the Belgian Ghent Clifford Research Group for generalizations of the Fourier transform in the framework of Clifford analysis, grew out from the study of the multidimensional Continuous Wavelet Transform in this particular setting. Clifford wavelet theory, however restricted to the continuous wavelet transform, was initiated by Brackx and Sommen in [34] and further developed by N. De Schepper in her PhD thesis [105]. The Clifford wavelets originate from a mother wavelet not only by translation and dilation, but also by rotation, making the Clifford wavelets appropriate for detecting directional phenomena. Rotations are implemented as specific actions on the variable by a spin element, since, indeed, the special orthogonal group SO(m) is doubly covered by the spin group Spin(m) of the real Clifford algebra R0,m . The mother wavelets themselves are derived from intentionally devised orthogonal polynomials in Euclidean space. It should be noted that these orthogonal polynomials are not tensor products of one–dimensional ones, but genuine multidimensional ones satisfying the usual properties such as a Rodrigues formula, recurrence relations and differential equations. In this way multidimensional Clifford wavelets were constructed grafted on the Hermite polynomials [35], Laguerre polynomials [38], Gegenbauer polynomials [39], Jacobi polynomials [43] and Bessel functions [36]. Taking the dimension m to be even, say m = 2n, introducing a complex structure, i.e. an SO(2n)–element squaring up to −1, and considering functions with values in the complex Clifford algebra C2n , socalled Hermitian Clifford analysis originates as a refinement of standard or Euclidean Clifford analysis. It should be noticed that the traditional holomorphic functions of several complex variables appear as a special
Introduction 13
case of Hermitian Clifford analysis, when the function values are restricted to a specific homogeneous part of spinor space. In this Hermitian setting the standard Dirac operator, which is invariant under the orthogonal group O(m), is split into two Hermitian Dirac operators, which are now invariant under the unitary group U(n). Also in this Hermitian Clifford analysis framework, multidimensional wavelets have been introduced by Brackx, H. De Schepper and Sommen [41, 44], as kernels for a Hermitian Continuous Wavelet Transform, and (generalized) Hermitian Clifford Hermite polynomials have been devised to generate the corresponding Hermitian wavelets [45, 47]. 1.4.2 Further developments in quaternion and Clifford wavelet theory
Clifford algebra multiresolution analysis (MRA) has been pioneered by M. Mitrea [264]. Important are also the electromagnetic signal application oriented developments of Clifford algebra wavelets by G. Kaiser [228–230, 232]. Quaternion MRA wavelets with applications to image analysis have been developed in [314] by Traversoni. Clifford algebra multiresolution analysis has been applied by BayroCorrochano [22–24] to: Clifford wavelet neural networks (information processing), also considering quaternionic MRA, a quaternionic wavelet phase concept, as well as applications to (e.g. robotic) motion estimation and image processing. Beyond this Zhao and Peng [335] established a theory of quaternionvalued admissible wavelets. Zhao [336] studied Clifford algebravalued admissible (continuous) wavelets using the complex Fourier transform for the spectral representation. Mawardi and Hitzer [252,253] extended this to continuous Clifford and CliffordGabor wavelets in Cl(3, 0) using the CFT of (1.3.1) for the spectral representation. They also studied a corresponding Clifford wavelet transform uncertainty principle. Hitzer [188,189] generalized this approach to continuous admissible Clifford algebra wavelets in real Clifford algebras Cl(n, 0) of dimensions n = 2, 3(mod 4), i.e. n = 2, 3, 6, 7, 10, 11, . . . Restricted to Cl(n, 0) of dimensions n = 2(mod 4) this approach has also been taken up in [257]. K¨ahler et al. [227] treated monogenic (Clifford) wavelets over the unit ball. Bernstein studied Clifford continuous wavelet transforms in L0,2 and L0,3 [27], as well as monogenic kernels and wavelets on the 3D sphere [28]. Bernstein et al. [29] further studied Clifford diffusion wavelets on conformally flat cylinders and tori. Soulard and Carr´e extend in [309] the theory and application of monogenic wavelets to color image denoising. De Martino and Diki [103] establish a special one dimensional quaternion shorttime Fourier transform (QSTFT). Its construction is based on a so called slice hyperholomorphic SegalBargmann transform.
CHAPTER
2
Clifford algebra
We study fundamental aspects of Clifford algebra, survey its modern formulation, look at how imaginary numbers are dealt with in real multivector algebra, and introduce useful aspects of quaternion based geometry. Sections 2.1 to 2.5 aim to highlight several fundamental aspects of Clifford’s geometric algebra (GA). They are mainly based on unpublished notes for a linear algebra course with some geometric algebra content at the University of the Air in Japan broadcast between 2004 and 2008 [177], and cover questions which often arise when first dealing with Clifford’s geometric algebra, and on how familiar linear algebra concepts are dealt with. Because Clifford’s geometric algebra serves later in this book as the algebraic grammar for hypercomplex integral transforms, it is important to become familiar with this essential extension of Grassmann algebra. Section 2.6 provides an overview of important geometric algebra examples of the Euclidean plane, Euclidean threedimensional space, spacetime and of conformal geometric algebra. Then Section 2.7 studies how real multivector elements of geometric algebra replace the many roles (e.g. in Fourier transform kernels) of imaginary numbers, providing a real geometric interpretation at the same time. Finally, Section 2.8 studies in detail the role quaternions play for rotations in three and four dimensions.
2.1 2.1.1
AXIOMS OF CLIFFORD’S GEOMETRIC ALGEBRA Axioms for geometric algebra Rp,q = Cl(p, q)
2.1.1.1 Algebra over field of real numbers
The set R of real numbers forms a field. For the addition of a, b, c ∈ R we have the properties of a+b=b+a commutativity (a + b) + c = a + (b + c) associativity a+0=a zero 0 a + (−a) = 0 opposite −a of a.
DOI: 10.1201/97810031844782
15
16 Quaternion and Clifford Fourier Transforms
For the multiplication of a, b, c ∈ R we have the properties of (a + b)c = bc + ac a(b + c) = ab + ac distributivity (ab)c = a(bc) associativity 1a = a unity 1 6= 0 aa−1 = 1 inverse a−1 for a = 6 0. ab = ba commutativity. Definition 2.1.1. An algebra over R is a linear space A over R together with a bilinear map (implying distributivity) A × A → A, (a, b) → ab. Hence, for Gn to be a Clifford geometric algebra over the real ndimensional Euclidean vector space Rn , the geometric product of elements A, B, C ∈ Gn must satisfy the following axioms: Axiom 2.1.2. Addition is commutative: A + B = B + A.
(2.1.1)
Axiom 2.1.3. Addition and the geometric product are associative: (A + B) + C = A + (B + C),
A(BC) = (AB)C,
(2.1.2)
(A + B)C = AC + BC.
(2.1.3)
and distributive: A(B + C) = AB + AC,
Axiom 2.1.4. There exist unique additive and multiplicative identities 0 and 1 such that: A + 0 = A, 1A = A. (2.1.4) Axiom 2.1.5. Every A in Gn has an additive inverse: A + (−A) = 0.
(2.1.5)
Axiom 2.1.6. For any nonzero vector a in Gn the square of a is equal to a unique positive scalar a2 , that is aa = a2 = a2 > 0. (2.1.6) Depending on the signature of the underlying vector space Rp,q , for defining Cl(p, q), zero and negative squares of vectors will also occur. Axiom 2.1.7. Every simple kvector, Ak = a1 a2 ...ak , can be factorized into pairwise orthogonal vector factors, which satisfy: ai aj = −aj ai ,
i, j = 1, 2, ..., k
and
i= 6 j.
(2.1.7)
Clifford algebra 17
2.1.1.2 Definition (1): using quadratic form
Let V = (Rn , Q) be a real vector space with a nondegenerate quadratic form Q : V → R of signature (p, q) with p + q = n. It is conventional to use the abbreviation Rp,q = Cl(p, q). For vectors a, b ∈ Rp,q we associate with Q the symmetric bilinear form habi = 21 [Q(a + b) − Q(a) − Q(b)]. Definition 2.1.8. An associative algebra over the field R with unity 1 is the geometric algebra Rp,q of the nondegenerate quadratic form Q (signature (p, q) and p + q = n) on Rn , which contains copies of the field R and of the vector space Rp,q as distinct subspaces so that (1) a2 = Q(a) for any a ∈ Rp,q ; (2) Rp,q generates Rp,q as an algebra over the field R; (3) Rp,q = Cl(p, q) is not generated by any proper subspace of Rp,q . Note the deliberate choice of unity 1! If q = 0 the second index is often omitted Rn = Rn,0 or Cln = Cl(n, 0).
(2.1.8)
Using an orthonormal basis {e1 , e2 , . . . , en } for Rp,q , the condition (1) can be expressed as e2k = 1, 1 ≤ k ≤ p, e2k = −1, p < k ≤ n, ek el = −el ek ,
(2.1.9)
k < l.
(3) is only needed for signatures p − q = 1 mod 4 where (e1 e2 . . . en )2 = 1. The basic multiplication rules for the basis vectors of an orthonormal basis can also be used to define a geometric algebra. This was the approach taken by Clifford himself in 1878 [77] and 1882 [78]. 2.1.1.3
Definition (2): by basic multiplication rules
Historically the first multiplication rule for vectors of the linear space Rn of importance for us is the bilinear outer (exterior) product of vectors. For a fixed orthonormal basis {e1 , e2 , . . . , en } of Rn Grassmann introduced “bivectors” ek ∧ el = −el ∧ ek , k = 6 l ek ∧ el = 0, k = l. The set of all bivectors {ek ∧ el k < l} forms a basis of a new linear space 2 Rn of ! n dimension . The name outer (or exterior) product stems from the fact, that it 2 assigns to any pair of vectors of the original space Rn an element of a different vector V space 2 Rn . Grassmann then defines multivectors. To any rtuple {v1 , v2 , . . . , vr } of vectors vk ∈ Rn , he assigns the multivector of grade r (or rank r, German: Stufe r) V
v1 ∧ v2 ∧ . . . ∧ vr .
18 Quaternion and Clifford Fourier Transforms
v1 ∧ v2 ∧ . . . ∧ vr is set zero if the vk are linearly dependent. The grade r multivectors change sign, under the permutation of any two adjacent vectors v1 ∧ . . . ∧ vk ∧ vk+1 ∧ . . . ∧ vr = −v1 ∧ . . . ∧ vk+1 ∧ vk ∧ . . . ∧ vr and form their own vector space r Rn . It follows that the set of all rvectors of the form ek1 ∧ ek2 ∧ . . . ∧ ekr , {k1 , k2 , . . . , kr } rsubset of {1, 2, . . . , n} V
!
n forms a basis of the new vector space R of dimension . r V The full exterior algebra (or Grassmann algebra) is the vector space Rn of dimension 2n with the rvectors ek1 ∧ ek2 ∧ . . . ∧ ekr , 1 ≤ r ≤ n as a basis. For defining the associative bilinear geometric product of vectors, Clifford used Grassmann’s multiplication rule of orthogonal vectors, Vr
n
ek el = −el ek , k 6= l, but set the geometric product of a vector with itself to be a real number ek ek = ±1. The associative algebra of dimension 2n so defined is the geometric algebra Rp,q = Cl(p, q). p says how many basis vectors of an orthogonal basis of the linear space Rn have positive square and the rest q = n − p have negative square. V The geometric product operating on multivectors u, v, w ∈ Rp,q can be exV pressed with the help of the left contraction ucv ∈ Rp,q (a) xcy = 12 (xy + yx) ∈ R (b) xc(u ∧ v) = (xcu) ∧ v + uˆ ∧ (xcv) (c) (u ∧ v)cw = uc(vcw) for x, y ∈ Rp,q and the grade involution uˆ =
X
(−1)r huir ,
r
with huir the grade rvector parts of u. (a) shows that the left contraction generalizes the usual inner product of 1vectors. (b) indicates that the vector x acts like a “derivation”. With the help of the left contraction, the geometric product of a vector V x ∈ Rp,q and a general multivector u ∈ Rp,q can be written as xu = xcu + x ∧ u. Demanding associativity and linearity, the geometric product extends to all of Rp,q . This gives the bilinear map which equips Grassmann’s exterior algebra vector space V p,q R to become the geometric algebra Rp,q = Cl(p, q). V
Clifford algebra 19
2.1.1.4 Definition (3): compact form
Definition 2.1.9 (Clifford’s geometric algebra [131, 247]). Let {e1 , e2 , . . ., ep , ep+1 , . . ., en }, with n = p + q, e2k = εk , εk = +1 for k = 1, . . . , p, εk = −1 for k = p + 1, . . . , n, be an orthonormal base of the inner product vector space Rp,q with a geometric product according to the multiplication rules ek el + el ek = 2εk δk,l ,
k, l = 1, . . . n,
(2.1.10)
where δk,l is the Kronecker symbol with δk,l = 1 for k = l, and δk,l = 0 for k = 6 l. This noncommutative product and the additional axiom of associativity generate the 2n dimensional Clifford geometric algebra Cl(p, q) = Cl(Rp,q ) = Clp,q = Gp,q = Rp,q over R. The set {eA : A ⊆ {1, . . . , n}} with eA = eh1 eh2 . . . ehk , 1 ≤ h1 < . . . < hk ≤ n, e∅ = 1, forms a graded (blade) basis of Cl(p, q). The grades k range from 0 for scalars, 1 for vectors, 2 for bivectors, s for svectors, up to n for pseudoscalars. The vector space Rp,q is included in Cl(p, q) as the subset of 1vectors. The general elements of Cl(p, q) are real linear combinations of basis blades eA , called Clifford numbers, multivectors or hypercomplex numbers. 2.1.1.5 Grade r subspaces
All elements u ∈ Rp,q = Cl(p, q) are defined as sums of terms of length r, summing over all r = 0 . . . n. Terms of constant length r X
huir =
uk1 ...kr ek1 . . . ekr
1≤k1 r we can reverse the order of all vector factors in the geometric product AB and get s(s−1)/2 e e ] hA (−1)r(r−1)/2 hBr As i. (2.3.20) s Br i = hBr As i = (−1) The powers of (−1) are due to the anticommutativity of the vector factors of the simple grade components of the homogeneous multivectors As and Br . Now we can apply the same argument as for the case s ≤ r and see in general that the scalar product is only nonvanishing if r = s. But if r = s the powers of (−1) cancel each other and we find that As ∗ Br = Br ∗ As . (2.3.21) That is the scalar product of two homogeneous mulitvectors is symmetric. By linearity, this extends to the scalar product of arbitrary multivectors n X
n X
n X
n X
s=1
r=1
s=1
s=1
A∗B =(
hAis ) ∗ (
= B ∗ A.
hBir ) =
hAis ∗ hBis =
hBis ∗ hAis (2.3.22)
32 Quaternion and Clifford Fourier Transforms
The scalar product inherits linearity from the geometric product. As an example let us consider the scalar product of a simple grade s−vector As with its reverse Aes Aes ∗ As = hAes As i = has . . . a2 a1 a1 . . . as−1 as i = a12 a22 . . . as2 ,
(2.3.23)
where we used the associativity of the geometric product, which allows us convenient pairwise multiplication of vector factors. Remember that the square of each vector is a scalar. In case that the linear space Rn has positive signature (q = 0) the above result will be positive and by linearity we can define a positive magnitude for a general multivector A as n X
A2 =
e s ∗ hAis ≥ 0. hAi
(2.3.24)
s=1
2.3.3 The outer product
The outer product in the Grassmann algebra [146] historically precedes the definition of geometric algebras by Clifford [77, 78]. But Clifford did not intend to do away with it, he rather wanted to unify Grassmann algebra and Hamilton’s algebra of quaternions in a single algebraic framework. It is therefore not surprising, that the geometric product of two arbitrary multivectors comprises the outer product as the sum over the maximum grade parts of the result n X
n X
n X n X
s=1
r=1
s=1 r=1
A∧B =(
hAis ) ∧ (
hBir ) =
hhAis hBir ir+s .
(2.3.25)
The outer product inherits both linearity and associativity from the geometric product. As defined by Grassmann, the outer product of two vectors a, b is antisymmetric a ∧ b = −b ∧ a.
(2.3.26)
Calculating the reverse of two homogeneous simple grade s and grade r multivectors As and Br we see that f e ^ A^ s ∧ Br = hAs Br ir+s = hBr As ir+s
= (−1)s(s−1)/2 (−1)r(r−1)/2 hBr As ir+s .
(2.3.27)
On the other hand, we also have (s+r)(s+r−1)/2 ^ hA hAs Br ir+s . s Br ir+s = (−1)
(2.3.28)
Equating both right sides again and taking care of the powers of (−1) we end up with the symmetry formula for the outer product of homogeneous multivectors As ∧ Br = (−1)rs Br ∧ As ,
(2.3.29)
which includes the antisymmetry of the outer product of vectors as a special case for r = s = 1.
Clifford algebra 33
2.3.3.1 The cross product of three dimensions
In conventional threedimensional vector analysis, frequent use is made of an antisymmetric product of vectors, which results in a third vector perpendicular to the two vector factors, with the length equal to the area of the parallelogram spanned by these two vectors and with the orientation given by the socalled right hand rule. The names used for this product are: vector product, cross product or outer product and it is often indicated by an xshaped product sign a × b = (a2 b3 − a3 b2 )e1 + (a3 b1 − a1 b3 )e2 + (a1 b2 − a2 b1 )e3 .
(2.3.30)
Historically, this product has actually been derived from Grassmann’s outer product of vectors by mapping the resulting grade 2 bivector area element to the dual vector perpendicular to it. This is done by multiplication with the pseudoscalar volume element I3 . a × b ≡ −I3 a ∧ b, a ∧ b = I3 a × b. (2.3.31) For an orthonormal basis {e1 , e2 , e3 } of the Euclidean space R3 we have I3 = e1 e2 e3 . Let us conclude this subsection with an example: a = e1 + 3e2 , b = e2 + 4e3 ,
(2.3.32)
a × b = −I3 (a ∧ b) = −I3 (e1 + 3e2 ) ∧ (e2 + 4e3 ) = −(e1 e2 e3 )(e1 e2 + 4e1 e3 + 12e2 e3 ) = −e1 e2 e3 e1 e2 − 4e1 e2 e3 e1 e3 − 12e1 e2 e3 e2 e3 = e3 − 4e2 + 12e1 ,
(2.3.33)
where we have used orthonormality of the vectors {e1 , e2 , e3 } and that ej ek = −ek ej for k 6= j. To recover the outer product bivector we simply multiply again with I3 to get a ∧ b = I3 (a × b) = e1 e2 e3 (e3 − 4e2 + 12e1 ) = e1 e2 + 4e1 e3 + 12e2 e3 .
(2.3.34)
It is important to note that the cross product vector construction is limited to three dimensions. It does not work in two dimensions, because no third perpendicular dimension is available and it does not work in four and higher dimensions, because the perpendicular space is then two or higher dimensional, i.e. a unique perpendicular vector can no longer be defined. In contrast to this, Grassmann’s original bivector construction is not bound to the dimensionality of the space. 2.3.3.2
Linear dependence and independence
The nonzero outer product of a set of r linearly independent vectors uniquely determines an rdimensional subspace of Rn spanned by these vectors [161]. Let us therefore proof the following proposition: a1 ∧ a2 ∧ . . . ∧ ar = 0 ⇐⇒ a1 , a2 , . . . , ar ∈ Rn linearly dependent.
(2.3.35)
34 Quaternion and Clifford Fourier Transforms
(⇐) Let a1 , a2 , . . . , ar ∈ Rn be a set of linearly dependent vectors. Without loss of generality we assume that a1 =
r X
αk ak ,
(2.3.36)
k=2
with at least one of the coefficients αk 6= 0. Inserting the sum for a1 we get a1 ∧ a2 ∧ . . . ∧ ar =
r X
!
αk ak
∧ a2 ∧ . . . ∧ ar = 0,
(2.3.37)
k=2
because each term of the sum contains an expression of the form ak ∧ ak = 0. (⇒) Let the outer product of r vectors a1 , a2 , . . . , ar ∈ Rn be zero: a1 ∧ a2 ∧ . . . ∧ ar = 0. But let us assume, that Ar−1 = a2 ∧ . . . ∧ ar = 6 0, and that the vectors {a2 , . . . , ar } are orthogonal. Then Ar−1 = a2 a3 . . . ar
(2.3.38)
The following definitions will be convenient for showing that a1 is a linear combination of a2 , . . . , ar . The multiplicative inverse of each vector with respect to the geometric product is given by ak ak−1 = 2 , k = 2, . . . , r (2.3.39) ak and the inverse of Ar−1 is then given by the reversely ordered product of the inverse vectors −1 −1 −1 A−1 (2.3.40) r−1 = ar ar−1 . . . a2 We then have −1 −1 −1 −1 −1 Ar−1 A−1 r−1 = a2 . . . ar−1 ar ar ar−1 . . . a2 = a2 . . . ar−1 ar−1 . . . a2
= . . . = 1.
(2.3.41)
According to our assumption, the full geometric product of a1 and Ar−1 is a1 Ar−1 = a1 cAr−1 + a1 ∧ Ar−1 = a1 cAr−1
(2.3.42)
and therefore −1 a1 = a1 Ar−1 A−1 r−1 = (a1 cAr−1 )Ar−1
= = =
r X k=2 r X k=2 r X k=2
End of the proof.
−1 (−1)r−k (a1 cak )ar−1 ar−1 . . . a×k . . . a2−1 a2 . . . ak . . . ar
−1 (a1 cak )ar−1 ar−1 . . . a×k . . . a2−1 a2 . . . a×k . . . ar ak
(a1 cak )ak .
(2.3.43)
Clifford algebra 35
A corollary of the above proposition is, that a1 ∧ a2 ∧ . . . ∧ ar 6= 0 ⇐⇒ a1 , a2 , . . . , ar ∈ Rn linearly independent.
(2.3.44)
Another corollary is that for a set of linearly independent vectors a1 , a2 , . . . , ar ∈ Rn we have for a ∈ Rn : a ∈ Span{a1 , a2 , . . . , ar } ⇐⇒ a ∧ a1 ∧ a2 ∧ . . . ∧ ar = 0.
(2.3.45)
Up to an arbitrary nonzero scalar factor, each homogeneous simple multivector is therefore in one to one correspondence with a subspace of Rn . This is the reason, why the operations of join, intersection (meet) and projection on subspaces can be easily be expressed by geometric products of corresponding homogeneous simple multivectors. The pseudoscalar In of a geometric algebra Rn corresponds to the space Rn itself. 2.3.4 Right and left contraction
We have already encountered the left and right contractions of two grade 1 vectors and the left contraction of a vector and a homogeneous grade r multivector. But the left and right contractions can be generalized to apply to arbitrary multivectors. The most general approach is to define left and right contractions solely in terms of the outer product and the scalar product. In this way, the left contraction can be defined as [110] C ∗ (AcB) ≡ (C ∧ A) ∗ B,
∀A, B, C ∈ Rn ,
(2.3.46)
∀A, B, C ∈ Rn .
(2.3.47)
and the right contraction as (BbA) ∗ C ≡ B ∗ (A ∧ C),
Note that both left and right contractions are linear, because the scalar product and the outer product used in these definitions are linear. Decomposing the multivectors A, B, C grade by grade we get A=
n X
Ak , B =
k=1
X
Bl , C =
X
(2.3.48)
Cm .
m
l
Inserting this into the expression for the left contraction we get by linearity of both the outer and the scalar product (C ∧ A) ∗ B =
X
(Cm ∧ Ak ) ∗ Bl =
X
hCm Ak Bl=m+k i0
m,k
k,l,m
!
=
X
X
m,k
m
hCm hAk Bm+k im i0 =
= C ∗ (AcB).
Cm ∗
X
hAk Bm+k im
k
(2.3.49)
This gives an explicit expression for the left contraction in terms of the grade parts of A and B X AcB = hhAik hBil il−k . (2.3.50) k,l
36 Quaternion and Clifford Fourier Transforms
In analogy to this we get for the right contraction BbA =
X
hhBil hAik il−k .
(2.3.51)
k,l
Note that in both cases m = l − k ≥ 0, i.e. combinations of l and k with l < k do not contribute. It is now straightforward to see that the contractions of two homogeneous multivectors of the same grade give their scalar product hAik chBik = hAik bhBik = hhAik hBik ik−k=0 = hAik ∗ hBik ,
(2.3.52)
and that, e.g. the left contraction of a vector a with a homogeneous grade r multivector Br will give a homogeneous grade r − 1 multivector acBr = haBr ir−1 .
(2.3.53)
In order to derive convenient explicit formulas for the calculation of the contractions of homogeneous simple multivectors, we will show the following formula: Ar c(Bs cCt ) = (Ar ∧ Bs )cCt ,
(2.3.54)
where Ar , Bs and Ct are supposed to be homogeneous simple multivectors of grade r, s and t, respectively. For the left side to be nonzero, we must have t ≥ s and t − s > r which is equivalent to t > r + s. To perform the proof, we scalar multiply with an arbitrary multivector D ∈ Rn from the left to get by repeated application of the defining relationship of the left contraction D ∗ [Ar c(Bs cCt )] = (D ∧ Ar ) ∗ (Bs cCt ) = (D ∧ Ar ∧ Bs ) ∗ Ct = [D ∧ (Ar ∧ Bs )] ∗ Ct = D ∗ [(Ar ∧ Bs )cCt ], ∀D ∈ Rn .
(2.3.55)
And hence Ar c(Bs cCt ) = (Ar ∧ Bs )cCt . Note that for conducting the proof no conditions on the values of the grades r, s and t had to be made. The left contraction ”takes care” of this. Let us now look in detail at the left contraction of two homogeneous simple multivectors Ar and Bs of grades r and s, respectively, given by the following vector factorizations Ar = a1 ∧ a2 ∧ . . . ∧ ar , Bs = b1 ∧ b2 ∧ . . . ∧ bs .
(2.3.56)
With repeated application of the formula, which we have just proved, we can rewrite the left contraction of Ar and Bs as Ar cBs = (a1 ∧ a2 ∧ . . . ∧ ar )c(b1 ∧ b2 ∧ . . . ∧ bs ) = ((a1 ∧ a2 ∧ . . . ∧ ar−1 ) ∧ ar )c(b1 ∧ b2 ∧ . . . ∧ bs ) = (a1 ∧ a2 ∧ . . . ∧ ar−1 )c(ar c(b1 ∧ b2 ∧ . . . ∧ bs )) = ... = = a1 c(a2 c(. . .c(ar c(b1 ∧ b2 ∧ . . . ∧ bs )) . . .)).
(2.3.57)
Clifford algebra 37
By repeated application of a previously shown formula (b ×k means to omit bk ) acBs =
r X
(−1)k+1 acbk (b1 ∧ b2 ∧ . . . ∧ b×k ∧ . . . ∧ bs ),
(2.3.58)
k=1
and by reordering we finally reach a very explicit formula for Ar cBs Ar cBs =
X
(j1 . . . js )Ar c(bj1 ∧ . . . ∧ bjr )(bjr+1 ∧ . . . ∧ bjs ).
(2.3.59)
j1 0. 11 Twofold covering means, that there are always two elements ±A in Pin(2, 0), Spin(2, 0) and Spin+ (2, 0), representing one element in O(2, 0), SO(2, 0) and SO+ (2, 0), respectively.
Clifford algebra 45
2.6.2.3 Concepts of vectors, k vectors and multivectors
A general element in Cl(2, 0), also called multivector can be represented as M = m0 + m1 e1 + m2 e2 + m12 e12 ,
m0 , m1 , m2 , m12 ∈ R.
(2.6.29)
Like real and imaginary parts of a complex number, we have a scalar part hM i0 of grade 0, a vector part hM i1 of grade 1 and a bivector part hM i2 of grade 2 M = hM i0 + hM i1 + hM i2 ,
(2.6.30)
with hM i0 = m0 ,
hM i1 = m1 e1 + m2 e2 ,
hM i2 = m12 e12 .
(2.6.31)
The set of all grade k elements, 0 ≤ k ≤ 2, is denoted Clk (2, 0). Grade extraction h. . .ik , 0 ≤ k ≤ 2, allows to do many useful computations, like e.g. the angle between two vectors (atan = tan−1 = arctan) θa,b = atan
habi2 e−1 12 habi0
(2.6.14)
=
atan
sin θa,b . cos θa,b
(2.6.32)
The symmetric scalar part of the geometric product of two multivectors M, N ∈ Cl(2, 0) is also called their scalar product and because of its fundamental importance denoted with a special product sign ∗ M ∗ N = hM N i0 = m0 n0 + m1 n1 + m2 n2 − m12 n12 = N ∗ M,
(2.6.33)
which is easy to compute using the multiplication table Table 2.3, since only the diagonal entries of Table 2.3 are scalar. Using the reversion (ef 12 = eg 1 e2 = e2 e1 = −e1 e2 = −e12 ) f = m0 + m1 e1 + m2 e2 − m12 e12 , M (2.6.34) we get the norm M  of a multivector as f = hM M fi0 = m2 + m2 + m2 + m2 , M 2 = M ∗ M 0 1 2 12
(2.6.35)
which for vectors M ∈ Cl1 (2, 0) is identical to the length of a vector, and for even grade subalgebra elements M ∈ Cl+ (2, 0) to the modulus of complex numbers, and for pure bivectors like a ∧ b = habi2 gives the area content of the parallelogram spanned by the vectors a, b ∈ R2 a ∧ b = a b  sin θa,b  =  det(a, b).
(2.6.36)
The reversion operation maps ef 12 = −e12 , it is therefore the equivalent of complex conjugation in the isomorphism Cl+ (2, 0) ∼ = C, fully consistent with its use in the norm (2.6.35).
46 Quaternion and Clifford Fourier Transforms
2.6.2.4 Higher dimensional types of inner and outer products for multivectors
The grade extraction also allows us to generalize the inner product of vectors, which maps two grade one vectors a, b ∈ R2 = Cl1 (2, 0) to a grade zero scalar a · b = habi(1−1=0) ∈ R = Cl0 (2, 0),
(2.6.37)
and therefore lowers the grade by 1. In contrast, the outer product with a vector raises the grade by 1, i.e. a ∧ b = habi(1+1=2) ∈ Cl2 (2, 0).
(2.6.38)
In general, the left contraction (symbol c) of a kvector Ak = hAik with a lvector Bl = hBil is defined as Ak cBl = hAk Bl i(l−k) , (2.6.39) which is zero if 0 > l − k, i.e. if k > l. Figuratively speaking (like with projections), we can only contract objects of the same or lower dimension from the left onto an object on the right. The right contraction (symbol b) is defined as Ak bBl = hAk Bl i(k−l) ,
(2.6.40)
which is zero if k − l < 0, i.e. if k < l. Figuratively speaking, we can only contract objects of the same or lower dimension from the right onto an object on the left. Both contractions are bilinear and can thus be extended to contractions of multivectors AcB = AbB =
2 X 2 X
hhAik hBil i(l−k) ,
k=0 l=0 2 2 X X
hhAik hBil i(k−l) .
(2.6.41) (2.6.42)
k=0 l=0
The reversion changes the order of factors and therefore relates left and right contractions by e b A, e e c A. e (AcB)∼ = B (AbB)∼ = B (2.6.43) In general, the associative outer product of a kvector Ak = hAik with a lvector Bl = hBil is defined as the maximum grade part of the geometric product Ak ∧ Bl = hAk Bl i(l+k) ,
(2.6.44)
where Ak ∧ Bl = 0 for l + k > 2, because in Cl(2, 0) the highest grade possible is 2. For example, the following outer products are obtained (α, β ∈ R, a, b, c ∈ R2 ) α ∧ β = αβ,
α ∧ a = a ∧ α = αa,
a ∧ b ∧ c = 0,
(2.6.45)
where the last identity is specific to Cl(2, 0), it does generally not hold in GAs of higher dimensional vector spaces. The last identity is due to the fact, that in a plane every third vector c can be expressed by linear combination of two linearly independent vectors a, b. If a, b would not be linearly independent, then already a ∧ b = 0, i.e. a k b. The outer product in 2D is of great advantage, because the cross product of vectors does only exist in 3D, not in 2D. We next treat the GA of R3 .
Clifford algebra 47
2.6.3 Overview of Clifford’s geometric algebra of the 3D Euclidean space
The Clifford algebra Cl(R3 ) = Cl(3, 0) of threedimensional (3D) Euclidean space R3 is arguably the by far most thoroughly studied and applied geometric algebra (GA). In physics, it is also known as Pauli algebra, since Pauli’s spin matrices provide a 2×2 matrix representation. This shows how GA unifies classical with quantum mechanics. Given an orthonormal vector basis {e1 , e2 , e3 } of R3 , the eightdimensional (23 = 8) Clifford algebra Cl(R3 ) = Cl(3, 0) has a basis of one scalar, three vectors, three bivectors and one trivector {1, e1 , e2 , e3 , e23 , e31 , e12 , e123 },
(2.6.46)
where as before e23 = e2 e3 , e123 = e1 e2 e3 , etc. All basis bivectors square to −1, and the product of two basis bivectors gives the third e23 e31 = e21 = −e12 , etc.
(2.6.47)
Therefore the even subalgebra Cl+ (3, 0) with basis12 {1, −e23 , −e31 , −e12 } is indeed found to be isomorphic to quaternions {1, i, j, k}. This isomorphism is not incidental. As we have learned already for Cl(2, 0), also in Cl(3, 0), the even subalgebra is the algebra of rotors (rotation operators) or spinors, and describes rotations in the same efficient way as do quaternions [322]. We therefore gain a real geometric interpretation of quaternions, as the oriented bivector side faces of a unit cube, with edge vectors {e1 , e2 , e3 }. In Cl(3, 0) a reflection at a plane (=hyperplane) is specified by the plane’s normal vector a ∈ R3 x0 = −a−1 xa, (2.6.48) the proof is identical to the one in (2.6.18) for Cl(2, 0). The combination of two such reflections leads to a rotation by α = 2θa,b x00 = R−1 xR,
1
R = ab = abeθa,b ia,b = abe 2 αia,b ,
(2.6.49)
where ia,b = a ∧ b/(a ∧ b) specifies the oriented unit bivector of the plane spanned by a, b ∈ R3 . The unit trivector i3 = e123 also squares to −1 i23 = e1 e2 e3 e1 e2 e3 = −e1 e2 e1 e3 e2 e3 = e1 e2 e1 e2 e3 e3 = (e1 e2 )2 (e3 )2 = −1,
(2.6.50)
where we only used that the permutation of two orthogonal vectors in the geometric product produces a minus sign. Hence i−1 3 = −i3 . We further find, that i3 commutes with every vector, e.g. e1 i3 = e1 e1 e2 e3 = e23 ,
(2.6.51)
i3 e1 = e1 e2 e3 e1 = −e1 e2 e1 e3 = e1 e1 e2 e3 = e23 , 12 The minus signs are only chosen, to make the product of two bivectors identical to the third, and not minus the third.
48 Quaternion and Clifford Fourier Transforms
and the like for e2 i3 = i3 e2 , e3 i3 = i3 e3 . If i3 commutes with every vector, it also commutes with every bivector a ∧ b = 12 (ab − ba), hence i3 commutes with every element of Cl(3, 0), a property which is called central in mathematics. The central subalgebra spanned by {1, i3 } is isomorphic to complex numbers C. i3 changes bivectors into orthogonal vectors e23 i3 = e2 e3 e1 e2 e3 = e1 e223 = −e1 , etc.
(2.6.52)
This is why the GA Cl(3, 0) is isomorphic to complex quaternions, one form of biquaternions n
o
−1 −1 1, e1 = e23 i−1 3 , e2 = e31 i3 , e3 = e12 i3 , e23 , e31 , e12 , i3 .
(2.6.53)
Yet writing the basis in the simple product form (2.6.46), fully preserves the geometric interpretation in terms of scalars, vectors, bivectors and trivectors, and allows to reduce all products to elementary geometric products of basis vectors, which is used in computational optimization schemes for GA software like Galoop [166]. 2.6.3.1 Explicit multiplication table and important subalgebras of Cl(3, 0)
For the full multiplication table of Cl(3, 0) we still need the geometric products of vectors and bivectors. By changing labels in Table 2.3 (1 ↔ 3 or 2 ↔ 3), we get that e2 e23 = −e23 e2 = e3 , e1 e31 = −e31 e1 = −e3 ,
e3 e23 = −e23 e3 = −e2 , e3 e31 = −e31 e3 = e1 ,
(2.6.54)
which shows that in general a vector and a bivector, which includes the vector, anticommute. The products of a vector with its orthogonal bivector always give the trivector i3 e1 e23 = e23 e1 = i3 ,
e2 e31 = e31 e2 = i3 ,
e3 e12 = e12 e3 = i3 ,
(2.6.55)
which also shows that in general vectors and orthogonal bivectors necessarily commute. Commutation relationships therefore clearly depend on both orthogonality properties and on the grades of the factors, which can frequently be exploited for computations even without the explicit use of coordinates. Table 2.2 gives the multiplication table of Cl(3, 0). The elements on the left most column are to be multiplied from the left in the geometric product with the elements in the top row. Every subtable of Table 2.2, that is closed under the geometric product represents a subalgebra of Cl(3, 0). We naturally find Table 2.3 as a subtable, because with R2 ⊂ R3 we necessarily have Cl(2, 0) ⊂ Cl(3, 0). In general, any pair of orthonormal vectors will generate a 4D subalgebra of Cl(3, 0) specific to the plane spanned by the two vectors. We also recognize the subtable of the even subalgebra with basis {1, e23 , e31 , e12 } isomorphic to quaternions. Then there is the subtable of {1, e123 } of the central subalgebra isomorphic to C. Any element of Cl(3, 0) squaring to −1 generates a 2D subalgebra isomorphic to C.
Clifford algebra 49
This fundamental observation leads to the generalization of the conventional complex Fourier transformation (FT) to Clifford FTs, sometimes called geometric algebra FTs [184]. Moreover, we also obtain generalizations of real and complex (dual) wavelets to Clifford wavelets, which include quaternion wavelets. We further have closed subtables with elements {1, e1 , e23 , e123 = i3 }, which are fully commutative and correspond to tessarines, also called bicomplex numbers, Segre quaternions, commutative quaternions or Cartan subalgebras. In general, we can specify any unit vector u ∈ R3 , u2 = 1, and get a commutative tessarine subalgebra with basis {1, u, ui3 , i3 }, which includes besides 1 and the vector u itself, the bivector ui3 orthogonal to u, and the oriented unit volume trivector i3 . Again GA provides a clear and useful geometric interpretation of tessarines and their products. Knowledge of the geometric interpretation is fundamental in order to see these algebras in nature and in problems at hand to be solved, and to identify settings, where a restriction to subalgebra computations may save considerable computation costs. 2.6.3.2 Concepts of grade structure of Cl(3, 0) and duality
A general multivector in Cl(3, 0), can be represented as M = m0 + m1 e1 + m2 e2 + m3 e3 + m23 e23 + m31 e31 + m12 e12 +m123 e123 ,
m0 , . . . , m123 ∈ R.
(2.6.56)
We have a scalar part hM i0 of grade 0, a vector part hM i1 of grade 1, a bivector part hM i2 of grade 2, and a trivector part hM i3 of grade 3 M = hM i0 + hM i1 + hM i2 + hM i3 , hM i0 = m0 ,
(2.6.57)
hM i1 = m1 e1 + m2 e2 + m3 e3 ,
hM i2 = m23 e23 + m31 e31 + m12 e12 ,
hM i3 = m123 e123 .
The set of all grade k elements, 0 ≤ k ≤ 3, is denoted Clk (3, 0). The multiplication table of Cl(3, 0), Table 2.2, reveals that multiplication with i3 (or i−1 3 = −i3 ) consistently changes an element of grade k, 0 ≤ k ≤ 3, into an element of grade 3−k, i.e. scalars to trivectors (also called pseudoscalars) and vectors to bivectors, and vice versa. This means that the geometric product of a multivector 13 M ∈ Cl(3, 0) and the pseudoscalar i3 (or i−1 in the left or 3 = −i3 ) always results right contraction M i3 = M ci3 , i3 M = i3 bM. (2.6.58) Mapping grades k, 0 ≤ k ≤ 3, to grades 3 − k is known as duality or Hodge duality. Because of its usefulness and importance it gets the symbol ∗ as an upper index M ∗ := M i−1 3 = −m0 i3 − m1 e23 − m2 e31 − m3 e12 + m23 e1 + m31 e2 + m12 e3 + m123 .
(2.6.59)
13 In the context of blade subspaces, whenever a blade B contains another blade A as factor, then the geometric product is reduced to left or right contraction: AB = AcB, BA = BbA.
50 Quaternion and Clifford Fourier Transforms
Duality (2.6.59) in Cl(3, 0) changes the outer product of vectors into the cross product (e1 ∧ e2 )∗ = e1 e2 (−e123 ) = e3 = e1 × e2 , (e2 ∧ e3 )∗ = e1 ,
(e3 ∧ e1 )∗ = e2 .
(2.6.60)
a ∧ b = i3 (a × b),
(2.6.61)
Therefore we can use (a ∧ b)∗ = a × b,
to translate well known results of standard 3D vector algebra into GA. Yet we emphasize, that the cross product a × b only exists in 3D, where a bivector has a unique (orthogonal) dual vector. The outer product has the advantage, that it can be universally used in all dimensions, because geometrically speaking the parallelogram spanned by two linearly independent vectors is always well defined, independent of the dimension of the embedding space. Duality14 also relates outer product and contraction of any A, B ∈ Cl(3, 0) A ∧ B ∗ = (AcB)∗ , ∗
A∗ ∧ B = (AbB)∗ ,
∗
∗
AcB = A bB = (A ∧ B) .
(2.6.62) (2.6.63)
For vectors a, b ∈ R3 this can be rewritten as a ∧ b∗ = a∗ ∧ b = (a · b)∗ ,
acb∗ = a∗ bb = (a ∧ b)∗ .
(2.6.64)
Inserting the duality definition this gives a ∧ (i3 b) = (i3 a) ∧ b = i3 (a · b),
ac(i3 b) = (i3 a)bb = i3 (a ∧ b),
(2.6.65)
which are handy relations in GA computations. Let us do two simple examples. First, for a = b = e1 . Then we have i3 a = i3 b = e23 , a · b = 1, and get a ∧ (i3 b) = e1 ∧ e23 = i3 , (i3 a) ∧ b = e23 ∧ e1 = i3 ,
i3 (a · b) = i3 .
(2.6.66)
Second, for a = e1 , b = e2 . Then we have i3 a = e23 , i3 b = e31 , a ∧ b = e12 , and get ac(i3 b) = e1 ce31 = he1 e31 i1 = −e3 , (i3 a)bb = e23 be2 = −e3 ,
i3 (a ∧ b) = i3 e12 = −e3 .
(2.6.67)
2.6.3.3 Concept of blade subspaces of Cl(3, 0)
A vector a ∈ R3 can be used to define a hyperplane H(a) by H(a) = {x ∈ R3  x · a = xca = 0},
(2.6.68)
14 Equations (2.6.62) apply in all Clifford algebras, but (2.6.63) needs some modification if the pseudoscalar is not central.
Clifford algebra 51
which is an example of an inner product null space (IPNS) definition. It can also be used to define a line L(a) by L(a) = {x ∈ R3  x ∧ a = 0},
(2.6.69)
which shows an outer product null space (OPNS) definition. We can also use two linearly independent vectors to span a plane P (a ∧ b). By that we mean, that x = αa + βb, α, β ∈ R, if and only if, x ∧ a ∧ b = 0. Let us briefly check that. Assume x = αa+βb, α, β ∈ R, then x∧a∧b = (αa+βb)∧a∧b = α(a∧a)∧b−βa∧(b∧b) = 0. On the other hand, if x has a component not in the plane spanned by a, b, then x∧a∧b results in a nonzero trivector, the oriented parallelepiped volume spanned by x, a, b. Hence every plane spanned by two linearly independent vectors a, b ∈ R is in OPNS P = P (a ∧ b) = {x ∈ R3  x = αa + βb, ∀ α, β ∈ R} = {x ∈ R3  x ∧ a ∧ b = 0}.
(2.6.70)
Three linearly independent vectors a, b, c ∈ R3 always span R3 , and their outer product with any fourth vector x ∈ R3 is zero by default (because in Cl(3, 0) no 4vectors exist). Therefore formally R3 = {x ∈ R3  x ∧ a ∧ b ∧ c = 0}. These geometric facts motivate the notion of blade. A blade A is an element of Cl(3, 0), that can be written as the outer product A = a1 ∧ . . . ∧ ak of k linearly independent vectors a1 , . . . , ak ∈ R3 , 0 ≤ k ≤ 3. Another name for kblade A is simple kvector. As we have just seen, every kblade fully defines and characterizes a kdimensional vector subspace V (A) ⊂ R3 . Such subspaces are therefore often simply called blade subspaces. The GA of a blade subspace Cl(V (A)) is a natural subalgebra of Cl(3, 0). Examples are Cl(V (e1 )) = Cl(1, 0) ⊂ Cl(3, 0), Cl(V (e12 )) = Cl(2, 0) ⊂ Cl(3, 0) and Cl(V (e123 ) = R3 ) = Cl(3, 0) ⊆ Cl(3, 0), etc. The formulas for projection and rejection of vectors can now be generalized to the projection and rejection of subspaces, represented by their blades in their OPNS representations, where the grade of a blade shows the dimension of the blade subspace. The projection of a blade A onto a blade B is given by PB (A) = (AcB)B −1 ,
(2.6.71)
where the left contraction automatically takes care of the fact that it would not be meaningful to project a higher dimensional subspace V (A) onto a lower dimensional subspace V (B). This projection formula is a first example of how GA allows to solve geometric problems by means of elementary algebraic products, rather than the often cumbersome conventional solution of systems of linear equations. The rejection (orthogonal part) of a blade A from a second blade B is given by PB⊥ (A) = (A ∧ B)B −1 .
(2.6.72)
If we interpret a general multivector A ∈ Cl(3, 0) as a weighted sum of blades, where the weights can also be interpreted as blade volumes, then by linearity the
52 Quaternion and Clifford Fourier Transforms
projection PB (A) [and rejection PB⊥ (A)] formula applied to a multivector A yields the weighted sum of blades, each projected onto [rejected from] the blade B. Let us consider the example of A = e1 + e3 + e12 + e23 and B = e12 , B −1 = −e12 . We get PB (A) = [(e1 + e3 + e12 + e23 )ce12 ](−e12 ) = [e2 + 0 − 1 + 0](−e12 ) = e1 + e12 ,
(2.6.73)
which is exactly what we expect, because the vector e1 is in the plane B = e12 , e3 is orthogonal to e12 and is projected out, the bivector component e12 is also in the plane B = e12 , but the bivector component e23 is perpendicular to e12 and is correctly projected out. Let us also compute the rejection PB⊥ (A) = [(e1 + e3 + e12 + e23 ) ∧ e12 ](−e12 ) = [0 + e3 e12 + 0 + 0](−e12 ) = e3 ,
(2.6.74)
which is again correct, because only the vector e3 is orthogonal to the plane B = e12 , whereas e.g. e23 is not fully orthogonal15 , because it has the vector e2 in common with B = e12 . The general duality of the outer product and the left contraction in (2.6.62) and (2.6.63) has the remarkable consequence of duality of IPNS and OPNS subspace representations, because for all blades A ∈ Cl(3, 0) xcA = 0 ⇔ x ∧ A∗ = 0,
∀ x ∈ R3 ,
(2.6.75)
i.e. because in general (xcA)∗ = x ∧ A∗ for all x, A ∈ Cl(3, 0). And because of (A∗ )∗ = A(−i3 )2 = −A, we also have xcA∗ = 0 ⇔ x ∧ A = 0,
∀ x ∈ R3 .
(2.6.76)
So we can either represent a subspace in OPNS or IPNS by a blade A ∈ Cl(3, 0) as VOPNS (A) = {x ∈ R3  x ∧ A = 0} = {x ∈ R3  xcA∗ = 0} = VIPNS (A∗ ).
(2.6.77)
We also have the relationship VIPNS (A) = {x ∈ R3  xcA = 0} = {x ∈ R3  x ∧ A∗ = 0} = VOPNS (A∗ ).
(2.6.78)
In application contexts both representations are frequently used and it should always be taken care to clearly specify if a blade is meant to represent a subspace in the 15
To include e23 one can simply compute A − PB (A) = e3 + e23 .
Clifford algebra 53
OPNS or the IPNS representation. In GA software, like CLUCalc or CLUViz [281], this is done by an initial multivector interpretation command. The outer product itself joins (or unifies) disjoint (orthogonal) blade subspaces, e.g., the two lines L(a) and L(b) are unified by the outer product to the plane P (a∧b) in (2.6.70). The left contraction has the dual property to cut out one subspace contained in a larger one and leave only the orthogonal complement. For example, if A = e1 , B = e12 , then AcB = e1 ce12 = e2 , the orthogonal complement of V (A) in V (B). In general in the OPNS, the orthogonal complement of V (A) in V (B) is therefore given by V (AcB). Let us assume two blades A, B ∈ Cl(3, 0), which represent two blade subspaces V (A), V (B) ⊂ R3 with intersection V (M ) = V (A) ∩ V (B), given by the common blade factor M ∈ Cl(3, 0) (called meet) A = A0 M = A0 ∧ M,
B = M B0 = M ∧ B0,
(2.6.79)
where A0 , B 0 are the orthogonal complement blades of M in A, B, respectively, A0 = AM −1 = AbM −1 ,
B 0 = M −1 B = M −1 cB,
(2.6.80)
where we freely use the fact that scalar factors, like M −2 = ±M −2 ∈ R of M −1 = M M −2 , are not relevant for determining a subspace (in both OPNS and IPNS), i.e. V (M ) = V (λM ), ∀ λ ∈ R \ {0}. The join (union) J of any two blade subspaces V (A), V (B) can therefore be represented as J = A0 ∧ M ∧ B 0 = A ∧ B 0 = A0 ∧ B. (2.6.81) Inserting the above contraction formulas for A0 , B 0 we get J = A ∧ (M −1 cB) = (AbM −1 ) ∧ B.
(2.6.82)
In turn, we can use the join J (or equivalently its inverse blade J −1 ) to compute M . The argument is from set theory. If we cut out B from J (or J −1 ), then only A0 (or A0−1 ) will remain as the orthogonal complement of B in J (or J −1 ). Cutting out A0 (or A0−1 ) from A itself leaves only the meet 16 (intersection) M M = A ∨ B = A0−1 cA = (BcJ −1 )cA = Bb(J −1 bA),
(2.6.83)
The right most form M = Bb(J −1 bA) is obtained by cutting A out of from J (or J −1 ), and using the resulting B 0−1 = J −1 bA, to get M as orthogonal complement of B 0 in B. For example, if we assume A = e12 , B = e23 , then the join is obviously J = e123 = i3 . This allows to compute the meet as M = A ∨ B = (BcJ −1 )cA = (e23 c(−e123 ))ce12 = e1 ce12 = e2 ,
(2.6.84)
which represents the line of intersection L(e2 ) of the two planes P (e12 ) and P (e23 ). 16
The symbol ∨ stems from GrassmannCayley algebra.
54 Quaternion and Clifford Fourier Transforms
2.6.3.4 Useful algebraic formulas in Clifford’s geometric algebras
Let A, B ∈ Cl(p, q, r). Some authors use angular brackets without index to indicate the scalar part of a multivector, e.g. hABi = hABi0 = hBAi0 .
(2.6.85)
For reversion (an involution) we have e A, e (AB)∼ = B
a ˜ = a,
e 0 = hAi0 = hAi∼ , hAi 0
k(k−1)/2 hAi∼ hAik . k = (−1)
(2.6.86)
Apart from reversion, there is the automorphism main involution (or grade involution) which maps all vectors a → a ˆ = −a, and therefore s a\ 1 . . . as = (−1) a1 . . . as ,
(2.6.87)
i.e. even blades are invariant under the grade involution, odd grade blades change sign. The composition of reversion and grade involution is called Clifford conjugation g ˆ = (d ˜ A = (A) A),
a1 . . . as = (−1)s as . . . a1 ,
hAik = (−1)k(k+1)/2 hAik .
(2.6.88)
Applied to a multivector M ∈ Cl(3, 0) we would get f = hM i0 + hM i1 − hM i2 − hM i3 , M c = hM i0 − hM i1 + hM i2 − hM i3 , M
M = hM i0 − hM i1 − hM i2 + hM i3 .
(2.6.89)
The following algebraic identities [162] are frequently applied ac(b ∧ c) = (acb)c − (acc)b = (a · b)c − (a · c)b Cl(3,0)
=
−a × (b × c),
(2.6.90)
ac(b ∧ c ∧ d) = (b ∧ c ∧ d)ba, = (b ∧ c)(a · d) − (b ∧ d)(c · a) + (c ∧ d)(b · a), a∧b∧c∧d
Cl(3,0)
=
0.
(2.6.91) (2.6.92)
2.6.4 Extending Clifford’s geometric algebra to spacetime (STA)
The GA Cl(1, 3) of flat Minkowski spacetime 17 R1,3 [108], has the 24 = 16D basis (e0 represents the time dimension) {1, e0 , e1 , e2 , e3 , σ1 = e10 , σ2 = e20 , σ3 = e30 , e23 , e31 , e12 , e123 , e230 , e310 , e120 , e0123 = I} 17
Note that some authors prefer opposite signature Cl(3, 1).
(2.6.93)
Clifford algebra 55
of one scalar, 4 vectors (basis of R1,3 ), 6 bivectors, 4 trivectors and one pseudoscalar. The 4 vectors fulfill eµ ceν = ηµ,ν = diag(+1, −1, −1, −1)µ,ν , 0 ≤ µ, ν ≤ 3. The commutator of two bivectors always gives a third bivector, e.g., 1 (e10 e20 − e20 e10 ) = −e12 , . . . 2
(2.6.94)
Exponentiating all bivectors gives a group of rotors, the Lorentz group, its Lie algebra is the commutator algebra of the six bivectors. The even grade subalgebra Cl+ (1, 3) with 8D basis (l = 1, 2, 3): {1, {σl }, {Iσl }, I} is isomorphic to Cl(3, 0). ~ = E1 σ1 + E2 σ2 + E3 σ3 , B ~ = B1 σ1 + B2 σ2 + Assigning electromagnetic field vectors E ~ + I B. ~ All four Maxwell equations then unify B3 σ3 , the Faraday bivector is F = E into one ∇F = J, (2.6.95) with vector derivative ∇ = 3µ,ν=0 η µ,ν eν ∂µ = 3µ=0 eµ ∂µ . The Dirac equation of quantum mechanics for the electron becomes in STA simply P
P
∇ψIσ3 = mψe0 ,
(2.6.96)
where the spinor ψ is an even grade multivector, and m the mass of the electron. The electron spinor acts as a spacetime rotor resulting in observables, e.g. the current vector e J = ψe0 ψ. (2.6.97) 2.6.5 Geometric modeling with conformal extension 2.6.5.1 On points, planes and motors in Cl(4, 1)
In order to linearize translations as in (2.6.100) we need the conformal model [9,244, 246] of Euclidean space (in Cl(4, 1)), which adds to x ∈ R3 ⊂ R4,1 two nullvector dimensions for the origin e 0 and infinity e ∞ , e 0 ce ∞ = −1, to obtain by a nonlinear embedding in R4,1 the homogeneous18 conformal point 1 X = x + x 2 e∞ + e0 , 2
e 20 = e 2∞ = X 2 = 0,
Xce ∞ = −1.
(2.6.98)
The condition X 2 = 0 restricts the point manifold to the socalled nullcone of R4,1 , similar to the light cone of special relativity. The second condition Xce ∞ = e 0 ce ∞ = −1 further restricts all points to a hyperplane of R4,1 , similar to points in projective geometry. The remaining point manifold is again 3D. See [108, 111, 190, 244, 248, 281, 282, 304] for details and illustrations. We can always move from the Euclidean representation x ∈ R3 to the conformal point X ∈ R4,1 by adding the e 0 and e ∞ components, or by dropping them (projection (2.6.71) with i3 , or rejection (2.6.72) with E = e ∞ ∧ e 0 ). The contraction of two conformal points gives their Euclidean distance and therefore a plane m equidistant from two points A = a + 21 a 2 e ∞ + e 0 , B = b + 12 b 2 e ∞ + e 0 18
Unique up to a nonzero scalar factor.
56 Quaternion and Clifford Fourier Transforms
as 1 XcA = − (x − a)2 2 ⇒ Xc(A − B) = 0, m = A − B ∝ n + d e ∞ ,
(2.6.99)
where n ∈ R3 is a unit normal to the plane and d its signed scalar distance from the origin, “∝” means proportional. Reflecting at two parallel planes m, m0 with distance t/2 ∈ R3 we get the translation operator (translator by t ) 1 X 0 = m0 m X mm0 = Tt−1 XTt , Tt := mm0 = 1 + te ∞ . 2
(2.6.100)
Reflection at two nonparallel planes m, m0 yields the rotation around the m, m0 intersection by twice the angle subtended by m, m0 . Group theoretically the conformal group C(3) [304] is isomorphic to O(4, 1) [315] and the Euclidean group E(3) is the subgroup of O(4, 1) leaving infinity e ∞ invariant. Now general translations and rotations are both linearly represented by geometric products of invertible vectors (called Clifford monomials, Lipschitz elements, versors or simply motion operators = motors). The commutator algebra of the bivectors of Cl(4, 1) constitutes the Lie algebra of all conformal transformations (exponentials of bivectors) of R3 . Derivatives with respect to these bivector motion parameters allow motor optimization, used in pose estimation, structure and motion estimation, motion capture and airborne laser strip adjustment (lidar) [219]. 2.6.5.2
Modeling geometric objects in Cl(4, 1) with blades
Computer vision, computer graphics and robotics are interested in the intuitive geometric OPNS meaning of blades (outer products of conformal points P1 , . . . , P4 ) in conformal GA (P p = point pair) P p = P1 ∧ P2 ,
Circle = P1 ∧ P2 ∧ P3 ,
Sphere = P1 ∧ P2 ∧ P3 ∧ P4 ,
(2.6.101)
homogeneously representing the point pair {P1 , P2 }, the circle through {P1 , P2 , P3 }, and the sphere with surface points {P1 , P2 , P3 , P4 }. Abstractly these are 0D, 1D, 2D and 3D spheres S with center c ∈ R3 , radius r ≥ 0, and Euclidean carrier blade directions D: 1 for points P , distance d ∈ R3 of P2 from P1 , circle plane bivector ic , and sphere volume trivector is ∝ e123 = i3 , see [190]. 1 S = D ∧ c + [ (c2 + r2 )D − c(ccD)]e ∞ + De 0 + (Dbc)E, 2
(2.6.102)
with the origininfinity bivector E = e ∞ ∧ e 0 . A point X ∈ R4,1 is on the sphere S, if and only if S ∧ X = 0. By duality this is equivalent to XcS ∗ = 0, which is equivalent in Euclidean terms to (x − c)2 = r2 , where x is the Euclidean part of X , c ∈ R3 the center of S, and r the radius of S. These objects are stretched to infinity (flattened) by wedging with e ∞ F = S ∧ e ∞ = D ∧ ce ∞ − DE = Dc ⊥ e ∞ − DE,
(2.6.103)
Clifford algebra 57
where c ⊥ ∈ R3 indicates the support vector (shortest distance from the origin): p for finite–infinite point pair P ∧ e ∞ , c ⊥ for the line P1 ∧ P2 ∧ e ∞ and the plane P1 ∧ P2 ∧ P3 ∧ e ∞ , and 0 for the 3D space R3 itself (F = −is E ∝ I5 ). All geometric entities can be extracted easily (as derived and illustrated in [190]) D = −F bE,
r2 =
S Sb , D2
c = D−1 [S ∧ (1 + E)]bE.
(2.6.104)
The conformal model is therefore also a complete super model for projective geometry. In the dual IPNS representation spheres (points for r = 0) and planes become vectors in R4,1 1 Sphere∗ = C − r2 e ∞ , 2
P lane∗ = n + de ∞
(2.6.105)
where C is the conformal center point, n := c ⊥ /c ⊥  ∝ ic i3 , “∝” means proportional. A point X ∈ R4,1 is on the plane P lane∗ , if and only if XcP lane∗ = 0, which is equivalent in Euclidean terms to x · n = d, where x is the Euclidean part of X , n ∈ R3 the oriented unit normal vector of P lane∗ , and d its signed scalar distance from the origin. Points are spheres with r = 0, circles and lines are intersection bivectors Circle∗ = S1∗ ∧ S2∗ ,
Line∗ = P lane∗1 ∧ P lane∗2 ,
(2.6.106)
of two sphere, and two plane vectors, respectively. Inversion at a sphere becomes X → SXS = S ∗ XS ∗ , inversion at two concentric spheres gives scaling. For example, the conformal center point C of a sphere S is C = Se ∞ S. 2.6.6 On Clifford analysis
Within geometric calculus (Clifford analysis), compare Chapter 3, we can define quaternionic and Clifford FTs and wavelet transforms, with applications in image and signal processing, and multivector wave packet analysis [184]. We also obtain a single fundamental theorem of multivector calculus, which unifies a host of classical theorems of integration for path independence, Green’s, Stokes’ and Gauss’ divergence theorems. Monogenic functions f , ∇f = 0, generalize complex analytic functions, and allow us to generalize Cauchy’s famous integral theorem for analytic functions in the complex plane to n dimensions [303].
2.7 2.7.1
HOW IMAGINARY NUMBERS BECOME REAL IN CLIFFORD ALGEBRAS What is an imaginary number?
The current subsection tries to provide a light introduction to the fundamental question of the nature of imaginary numbers. It is based on notes that first appeared in [178]. The previous Japanese emperor (Emperor Showa) is said to have asked this question. Today many students and scientists still ask it, but the traditional canon of
58 Quaternion and Clifford Fourier Transforms
mathematics at school and university needs to be widened for the answer. We find it in the works of Hamilton, Grassmann and Clifford. Hamilton introduced quaternions i, j, k, with i2 = j 2 = k2 = ijk = −1, ij = −ji = k,
jk = −kj = i,
ki = −ik = j,
(2.7.1)
for threedimensional rotations. Grassmann invented the outer product of oriented line segments (vectors) a, b to give the directed oriented area of the enclosed parallelogram: a ∧ b = −b ∧ a. (2.7.2) Clifford unified their work with the geometric product ab = a · b + a ∧ b,
(2.7.3)
leading to geometric algebras. In two dimensions we have orthogonal, unit vectors e1 , e2 as vector space basis with e21 = e22 = 1, e1 · e2 = 0. (2.7.4) The associative geometric multiplication of the oriented directed unit square i = e1 e2
(2.7.5)
gives: ii = e1 e2 e1 e2 = e1 (e2 e1 )e2 = e1 (e2 · e1 + e2 ∧ e1 )e2 = e1 (0 − e1 ∧ e2 )e2 = −e1 e1 e2 e2 = −1.
(2.7.6)
Remark 2.7.1. We only used e2 · e1 = e1 · e2 = 0 and
e2 e1 = e2 ∧ e1 = −e1 ∧ e2 = −e1 e2 .
(2.7.7)
So the square of the oriented unit area i is −1. Enough to satisfy the emperor’s curiosity! But today’s politicians ask for an application. As an answer we calculate: ie1 = e1 e2 e1 = −e1 e1 e2 = −e2 ,
(2.7.8)
ie2 = e1 e2 e2 = e1 ,
(2.7.9)
and which is a clockwise 90◦ rotation. We can also calculate (NB: the order!) e1 i = e1 e1 e2 = e1 ,
(2.7.10)
e2 i = e2 e1 e2 = −e1 e2 e2 = −e1 ,
(2.7.11)
and
Clifford algebra 59
which is an anticlockwise (mathematically positive) 90◦ rotation. For a general rotation in two dimensions, we simply add trigonometric coefficients: a(cos α + i sin α)
(2.7.12)
rotates the real vector a by α radians. Now even a politician can rotate vectors without using (or even knowing) matrices. Of what use may the geometric product be for some new advanced technology venture business? As an application to laser beam optics let us imagine a laser beam with direction vector a hitting a mirror surface element approximated with unit normal vector n, (n 2 = 1). We can write a in components parallel and perpendicular to n: a = ak + a⊥. (2.7.13) Now a k ∧ n = 0,
(2.7.14)
because parallel vectors span no parallelogram, and a ⊥ · n = 0,
(2.7.15)
because of perpendicularity. So we must have a k n = a k · n + 0 = n · a k + 0 = na k ,
(2.7.16)
a ⊥ n = 0 + a ⊥ ∧ n = 0 − n ∧ a ⊥ = −na ⊥ .
(2.7.17)
and Reflection only changes the sign of a k . Therefore a 0 = −a k + a ⊥ = −nn(a k − a ⊥ ) = −n(na k − na ⊥ ) = −n(a k n + a ⊥ n) = −n(a k + a ⊥ )n = −nan,
(2.7.18)
is the reflected vector. In a cavity we may want to trace many reflections at a sequence of surface elements with normal vectors n 1 , n 2 , . . . n s which simply results in a 0 = (−1)s n s . . . n 2 n 1 an 1 n 2 . . . n s .
(2.7.19)
Nanoscience is a modern buzz word. On this scale mechanics meets quantum mechanics. Geometric algebra provides complete tools for both. From elementary geometry we know that two reflections at planes with normal vectors n, m enclosing the angle θ/2 result in a rotation by angle θ: a 0 = mn a nm.
(2.7.20)
The general rotation operator (rotor) is R = nm = n · m + n ∧ m = cos(θ/2) + i sin(θ/2) = eiθ/2
(2.7.21)
with unit area element i in the n, m rotation plane. Two rotations are given by the
60 Quaternion and Clifford Fourier Transforms
geometric product of two rotors RR0 . A second θ0 = 2π (equivalent to 360◦ ) rotation produces RR0 = Re2πi/2 = Reπi = R(cos π + i sin π) = R(−1 + i0) = −R.
(2.7.22)
The rotor R itself behaves therefore like the first known quantum particle, i.e. the electron described by a Pauli spinor y = ρ1/2 R.
(2.7.23)
We summarize, that Clifford’s geometric algebra answers fundamental questions, which the traditional canon of mathematics taught at schools and universities cannot. It further provides great methodological simplifications and geometric insight in applications to physics, molecular geometry, image processing, computer graphics, robotics, quantum computing, etc., compare [202]. 2.7.2 What about imaginary eigenvalues and complex eigenvectors?
This subsection is mainly based on [171]. For further reading on the same topic see also [172, 173]. We first review how antisymmetric matrices in two dimensions yield imaginary eigenvalues and complex eigenvectors. We will see how this carries on to rotations by means of the Cayley transformation. Then a real geometric interpretation is given to the eigenvalues and eigenvectors by means of real geometric algebra. The eigenvectors are seen to be two component eigenspinors which can be further reduced to underlying vector duplets. The eigenvalues are interpreted as rotation operators, which rotate the underlying vector duplets. We then extend and generalize the treatment to three dimensions. The idea for this treatment arose from a linear algebra problem on antisymmetric matrices for undergraduate engineering students. I wrote it, looking for a real geometric understanding of the imaginary eigenvalues and complex eigenvectors. Being already familiar with geometric algebra [152, 159, 161, 162] it was natural to try to apply it in this context. 2.7.2.1 Two real dimensions 2.7.2.1.1
Complex treatment
Any antisymmetric matrix in two real dimensions is
proportional to U=
0 −1 1 0
!
(2.7.24)
.
The characteristic polynomial equation of the matrix U is −λ −1 U − λE = 1 −λ
= λ2 + 1 = 0,
i.e.
λ2 = −1.
(2.7.25)
The classical way to√solve this equation is to postulate an imaginary entity j to be the root of −1: j = −1. This leads to many interesting consequences, yet any real
Clifford algebra 61
geometric meaning of this imaginary quantity is left obscure. The two eigenvalues are therefore the imaginary unit j and −j. λ2 = −j.
λ1 = j,
(2.7.26)
The corresponding complex eigenvectors x1 and x2 are →
U x1 = λ1 x1 = jx1 U x2 = λ2 x2 = −jx2
x1 =
→
1 −j
!
1 j
!
x2 =
, .
(2.7.27)
The Cayley transformation [271] C(−kU ), with k = (1 − cos ϑ)/ sin ϑ C(−kU ) = (E + (−kU ))−1 (E − (−kU )) = E − cos ϑ − sin ϑ sin ϑ cos ϑ
=
2 (−kU − (−kU )2 ) 1 + k2
!
(2.7.28)
allows to describe twodimensional rotations. The third expression of equation (2.7.28) shows that U and C(−kU ) must have the same eigenvectors x1 and x2 . The corresponding eigenvalues of C(−kU ) can now easily be calculated from (2.7.28) as λc1 = 1 +
2kλ1 (1 + kλ1 ) , 1 + k2
λc2 = 1 +
2kλ2 (1 + kλ2 ) . 1 + k2
(2.7.29)
Inserting λ1 = j and λ2 = −j we obtain the complex eigenvalues of the twodimensional rotation C(−kU ) as λc1 = cos ϑ + j sin ϑ,
λc2 = cos ϑ − j sin ϑ.
(2.7.30)
We now face the question what the imaginary and complex eigenvalues and the complex eigenvectors of U and the rotation C(−kU ) mean in terms of purely real geometry. In order to do this let us turn to the real geometric algebra R2 of a real twodimensional vector space R2 [152, 161, 162]. Real explanation Instead of postulating the imaginary unit j we now solve the characteristic polynomial equation (2.7.25) using both orientations of the oriented unit area element i of R2 : λ1 = i, λ2 = −i. (2.7.31) 2.7.2.1.2
The corresponding “eigenvectors” x1 and x2 will then be: x1 =
1 −i
!
,
x2 =
1 i
!
.
(2.7.32)
62 Quaternion and Clifford Fourier Transforms
As before, the “eigenvectors” of the Cayley transformation C(−kU ) will be the same. And the eigenvalues of C(−kU ) now become: λc1 = cos ϑ + i sin ϑ,
λc2 = cos ϑ − i sin ϑ.
(2.7.33)
We can now take the first step in our real explanation and identify the two “eigenvectors” x1 and x2 as twocomponent spinors with the entries: x11 = 1, x12 = −i and x21 = 1, x22 = i. Now we want to better understand what the realorientedunitareaelement eigenvalues λ1 , λ2 as well as λc1 and λc2 do when multiplied with the twocomponent eigenspinors x1 and x2 . Every spinor can be understood to be the geometric product of two vectors. We therefore choose an arbitrary, but fixed reference vector z from the vector space R2 . For simplicity let us take z to be z = σ 1 , assuming {σ 1 , σ 2 } to be the orthonormal basis of R2 . We can then factorize the spinor components of the eigenspinors x1 and x2 to: x11 = x21 = 1 = σ 1 σ 1 ,
x12 = −i = σ 2 σ 1 ,
x22 = i = −σ 2 σ 1 .
(2.7.34)
Note that we always factored out z to the right. In two real dimensions it now seems natural to adopt the following interpretation: The eigenspinor x1 corresponds (modulus the geometric multiplication from the right with z = σ 1 ) to the real vector pair {σ 1 , σ 2 }, whereas x2 corresponds to the real vector pair {σ 1 , −σ 2 }. Multiplication with λ1 from the left as in U x1 = λ1 x1 = i
x11 x12
!
=
ix11 ix12
!
,
(2.7.35)
results in x11 → ix11 = (−σ 2 )σ 1 ,
x12 → ix12 = σ 1 σ 1 .
(2.7.36)
That is the multiplication with λ1 = i from the left transforms the vector pair {σ 1 , σ 2 } to the new pair {σ 2 , σ 1 }, which is a simple rotation by −90 degrees. Here the noncommutative nature of the geometric product is important. The analogous calculation for λ2 x2 = −ix2 shows that the pair {σ 1 , −σ 2 }, which corresponds to x2 is transformed to {σ 2 , −σ 1 }, i.e. it is rotated by +90 degree. I will now treat C(−kU )x1 = λc1 x1 and C(−kU )x2 = λc2 x2 in the same way. x11 → λc1 x11 = (cos ϑ + i sin ϑ)σ 21 = σ 1 (cos ϑ − i sin ϑ)σ 1 = (σ 1 R(−ϑ))σ 1 ,
(2.7.37)
where R(−ϑ) is the rotation operator by −ϑ. For the second component x12 we have x12 → λc1 x12 = (cos ϑ + i sin ϑ)(σ 2 σ 1 ) = (σ 2 R(−ϑ))σ 1 .
(2.7.38)
The action of λc1 on x1 means therefore a rotation of the corresponding vector pair {σ 1 , σ 2 } by −ϑ. The analogous calculations for λc2 x2 show that λc1 rotates the vector pair
Clifford algebra 63
{σ 1 , −σ 2 }, which corresponds to x2 , into {σ 1 R(ϑ), −σ 2 R(ϑ)}. This corresponds to a rotation of the vector pair by +ϑ. Summarizing the twodimensional situation, we see that the complex eigenvectors x1 and x2 may rightfully be interpreted as twocomponent eigenspinors with underlying vector pairs. The multiplication of these eigenspinors with the unitorientedareaelement eigenvalues λ1 and λ2 means a real rotation of the underlying vector pairs by −90 and +90 degrees, respectively. Whereas the multiplication with λc1 and λc2 means a real rotation of the underlying vector pairs by −ϑ and +ϑ, respectively. Now all imaginary eigenvalues and complex eigenvectors of antisymmetric matrices in two real dimensions have a real geometric interpretation. Let us examine next how this carries on to three dimensions. 2.7.2.2 Three real dimensions Complex treatment of three dimensions Any antisymmetric matrix in three real dimensions is proportional to a matrix of the form
2.7.2.2.1
0 −c b 0 −a , U = c −b a 0
(2.7.39)
with a2 + b2 + c2 = 1. The characteristic polynomial equation of the matrix U is −λ −c b U − λE = c −λ −a −b a −λ
= λ(λ2 + a2 + b2 + c2 ) = 0.
(2.7.40)
If we use the condition that a2 + b2 + c2 = 1, this simplifies and breaks up into the two equations λ21,2 = −1, λ3 = 0. (2.7.41) That means we have one eigenvalue λ3 equal to zero and for the other two eigenvalues λ1 , λ2 we have the same condition (2.7.25) as in the twodimensional case for the matrix of equation (2.7.24). It is therefore clear that in the conventional treatment one would again assign19 λ1 = j and λ2 = −j. The corresponding complex eigenvectors are:
1 − a2 −ab + jc −ac − jb . . x1 = −ab − jc = 1 − b2 = −bc + ja , 2 −ac + jb −bc − ja 1−c x2 = cc(x1 ),
(2.7.42)
where cc(.) stands for the usual complex conjugation, i.e. cc(j) = −j. The symbol 19
Here an incompleteness of the conventional treatment becomes obvious. A priori there is no reason to assume that the solutions to the characteristic polynomial equations in two and three dimensions (2.7.25) and (2.7.41) must geometrically be the same.
64 Quaternion and Clifford Fourier Transforms
. = expresses that all three given forms are equivalent up to the multiplication with a scalar (complex) constant. The eigenvector that corresponds to λ3 simply is:
a x3 = b . c
(2.7.43)
The fact that λ3 = 0 simply means that the matrix U projects out any component of a vector parallel to x3 . U maps the threedimensional vector space therefore to a plane perpendicular to x3 containing the origin. ϑ The Cayley transformation [271] C(−kU ) with k = 1−cos sin ϑ now describes rotations in three dimensions: C(−kU ) = (E + (−kU ))−1 (E − (−kU )) 2 = E− (−kU − (−kU )2 ) = 2 2 1 + k (a + b2 + c2 ) 1 + (1 − cos ϑ)(1 − a2 ) c sin ϑ + ab(1 − cos ϑ) −b sin ϑ + ac(1 − cos ϑ)
(2.7.44)
−c sin ϑ + ab(1 − cos ϑ) b sin ϑ + ac(1 − cos ϑ) 1 + (1 − cos ϑ)(1 − b2 ) −a sin ϑ + bc(1 − cos ϑ) . a sin ϑ + bc(1 − cos ϑ) 1 + (1 − cos ϑ)(1 − b2 )
The vector x3 is the rotation axis. The expression for C(−kU ) after the second equal sign in (2.7.44) clearly shows that the eigenvectors of U and C(−kU ) agree in three dimensions as well. The general formula for calculating the eigenvalues λc of C(−kU ) from the eigenvalues λ of U reads: λc = 1 +
2kλ(1 + kλ) 1 + k 2 (a2 + b2 + c2 )
a2 +b2 +c2 =1
=
1+
2kλ(1 + kλ) . 1 + k2
(2.7.45)
Inserting λ1 , λ2 and λ3 in this formula yields: λc1 = cos ϑ + j sin ϑ, λc2 = cos ϑ − j sin ϑ,
λc3 = 1.
(2.7.46)
We see that in three dimensions the complex eigenvectors (2.7.42) contain more structure and the explicit form of the Cayley transformation (2.7.44) gets rather unwieldy. If we follow the treatment of the twodimensional case given in Section 2.7.2.1.2, then we need to replace the imaginary unit j in the eigenvalues λ1 , λ2 and in the eigenvectors x1 , x2 by an element of the real threedimensional geometric algebra R3 [152,159,161]. In principle, there are two different choices: The volume element i or any twodimensional unit area element like e.g. i1 = σ 2 σ 3 , i2 = σ 3 σ 1 or i3 = σ 1 σ 2 . ({σ 1 , σ 2 , σ 3 } denotes an orthonormal basis in R3 .) While both interpretations are possible, let me argue for the second possibility: We have seen in Section 2.7.2.2.1 that the multiplication of U with a vector always 2.7.2.2.2
Real explanation for three dimensions
Clifford algebra 65
projects out the component of this vector parallel to x3 so that the y on the right hand side of equations like U x = y is necessarily a vector in the twodimensional plane perpendicular to x3 containing the origin. Thus it seems only natural to interpret the square root of −1 in the solution of equation (2.7.41) to be the oriented unit area element i = ai1 + bi2 + ci3 characteristic for the plane perpendicular to x3 containing the origin as opposed to the volume element element i or any other twodimensional unit area element. I will show in the following, that this leads indeed to a consistent interpretation. e 1 = −i and Using this area element i we have λ1 = i, λ2 = λ
−ac − ib −ab + ic 1 − a2 . . x1 = −ab − ic = 1 − b2 = −bc + ia , 1 − c2 −bc − ia −ac + ib e1, x2 = x
(2.7.47)
where the tilde operation marks the reverse of geometric algebra. As in the twodimensional case I interpret the three components of each “eigenvector” as spinorial components, i.e. elementary geometric products of two vectors. (In the following we will therefore use the expression threecomponent eigenspinor instead of “eigenvector”.) I again arbitrarily fix one vector from the i plane (the plane perpendicular to x3 ) as a reference vector z with respect to which I will factorize the three component eigenspinors x1 and x2 . With regard to the first representation of the eigenspinors x1 and x2 we choose to set z = σ 1k = σ 1 · ii−1 = (1 − a2 )σ 1 − abσ 2 − acσ 3 .
(2.7.48)
Using a2 + b2 + c2 = 1, the square z2 = 1 − a2 is seen to be the first component spinor of the first representation of x1 and x2 as given in (2.7.47). Next we will use the inverse of z z−1 = σ 1 −
ab ac σ2 − σ3 2 1−a 1 − a2
(2.7.49)
in order to factorize the two other component spinors n2 z = −ab − ic and n3 z = −ac + ib of x1 in (2.7.47) as well. A somewhat cumbersome calculation20 renders n2 = n2 zz−1 = σ 2k , n3 = n3 zz−1 = σ 3k .
(2.7.50)
Summarizing these calculations we have (setting n1 = z = σ 1k ):
1 − a2 n1 z σ 1k σ 1k x1 = −ab − ic = n2 z = σ 2k σ 1k . −ac + ib n3 z σ 3k σ 1k
(2.7.51)
20 A good way to speed up and verify such calculations is geometric algebra software, such as [14] and others.
66 Quaternion and Clifford Fourier Transforms
The other two equivalent representations of x1 given in (2.7.47) can be written as:
σ 1k σ 2k −ab + ic 1 − b2 = σ 2k σ 2k , −bc − ia σ 3k σ 2k
σ 1k σ 3k −ac − ib −bc + ia = σ 2k σ 3k . 2 1−c σ 3k σ 3k
(2.7.52)
We see that this simply corresponds to a different choice of the reference vector z, as z0 = σ 2k and as z00 = σ 3k , respectively. In general all possible ways to write x1 correspond to different choices of z from the i plane. The geometric product Rz−1 z0 = z−1 z0 for any two such reference vectors z and z0 gives the rotation operation to rotate one choice of threecomponent eigenspinor representation x1 (z) into the other x1 (z0 ) = x1 (z)Rz−1 z0 . As for two dimensions on page 62 we could also try to interpret x2 by factoring out a reference vector z to the right. But since according to equation (2.7.47) x2 is simply the reverse of x1 , it seems not really needed for a real interpretation. Doing it nevertheless, yields less handy expressions. So all we need to give a real geometric interpretation for the threecomponent eigenspinors x1 (and x2 ) is the triplet (σ 1k , σ 2k , σ 3k ) of projections of the three basis vectors σ 1 , σ 2 and σ 3 onto the i plane. Multiplying this triplet with any vector z, element of the i plane, from the right (from the left) yields all representations of x1 (and x2 ). After successfully clarifying the real interpretation of the “complex eigenvectors” in terms of a real vector space R3 vector triplet, we turn briefly to the interpretation of the eigenvalues. The real oriented plane unit area element eigenvalues λ1 = i and ˜ 1 = −i yield via equation (2.7.45) the eigenvalues of the Cayley transformation λ2 = λ C(−kU ) as: λc1 = cos ϑ + i sin ϑ,
˜ c1 = cos ϑ − i sin ϑ, λc2 = λ
λc3 = 1.
(2.7.53)
The action of λc1 on x1 and λc2 on x2 , respectively, give
n1 z λc1 n1 z C(−kU )x1 = λc1 x1 = λc1 n2 z = λc1 n2 z , n3 z λc1 n3 z and
(2.7.54)
zn1 zλc1 n1 C(−kU )x2 = λc2 x2 = λc2 zn2 = zλc1 n2 . zn3 zλc1 n3
(2.7.55)
e c2 = zλc1 , e 1 and that λc2 z = zλ In equation (2.7.55), we have used the facts that x2 = x since z is element of the i plane. We can therefore consistently interpret the both λc1 x1 and λc2 x2 as one and the same rotation of the vector triplet (n1 , n2 , n3 ) = (σ 1k , σ 2k , σ 3k ) by the angle −ϑ in a right handed sense around the axis x3 in the i plane. But we are equally free to alternatively view it as a +ϑ rotation (in the i plane) of the reference vector z
Clifford algebra 67
instead. No further discussion for the eigenvalues λ1 , λ2 of U is needed, since these are special cases of λc1 , λc2 with ϑ = π/2. The third eigenvalue of the Cayley transformation C(−kU ) is λc3 = 1, which means that any component parallel to x3 will be invariant under multiplication with C(−kU ). So far geometric algebra has served us as an investigative tool in order to gain a consistent real geometric vector space interpretation of imaginary eigenvalues and complex eigenvectors of antisymmetric matrices in two and three dimensions. But it is equally possible to pretend not to know about the antisymmetric matrices and their eigenvalues and eigenvectors in the first place, and synthetically construct relationships in geometric algebra which give all the counterparts found in our investigative (analytical) approach so far. As shown in [172] this necessitates in three dimensions the use of the twosided spinorial description [159, 162] of rotations. 2.7.3 From imaginary numbers to multivector square roots of −1 in Clifford geometric algebras Cl(p, q) with p + q ≤ 4
It is known that Clifford (geometric) algebra21 offers a geometric interpretation for square roots of −1 in the form of blades that square to −1, e.g. unit bivectors and the unit pseudoscalar in Cl(3, 0). Furthermore, this extends to a geometric interpretation of quaternions as the side face bivectors of a unit cube H ∼ = Cl+ (3, 0). Research has been done [294] on the biquaternion roots of −1, abandoning the restriction to blades. Biquaternions are isomorphic to the Clifford (geometric) algebra Cl(3, 0) of R3 . All these roots of −1 find immediate applications in the construction of new types of geometric Clifford Fourier transformations. In the current section we show how to explicitly extend this research to general algebras Gp,q = Cl(p, q). We fully present the geometric roots of −1 for the Clifford (geometric) algebras with p + q ≤ 4. The description is mainly based on [193], which contains fully detailed algebraic proofs. Historically, Grassmann invented the antisymmetric outer product of vectors, that regards the oriented parallelogram area spanned by two vectors as a new type of number, commonly called bivector. The bivector represents its own plane, because outer products with vectors in the plane vanish. In three dimensions the outer product of three linearly independent vectors defines a socalled trivector with the magnitude of the volume of the parallelepiped spanned by the vectors. Its orientation (sign) depends on the handedness of the three vectors. In the Clifford algebra [162] of R3 the three bivector side faces of a unit cube {e 1 e 2 , e 2 e 3 , e 3 e 1 } oriented along the three coordinate directions {e 1 , e 2 , e 3 } correspond to the three quaternion units i, j and k. Like quaternions, these three bivectors square to minus one and generate the rotations in their respective planes. Beyond that Clifford algebra allows to extend complex numbers to higher 21
In his original publication [77] Clifford first used the term geometric algebras. Subsequently in mathematics the new term Clifford algebras [247] has become the proper mathematical term. For emphasizing the geometric nature of the algebra, some researchers continue [58,161,162] to use the original term geometric algebra(s).
68 Quaternion and Clifford Fourier Transforms
dimensions [33, 161] and systematically generalize our knowledge of complex numbers, holomorphic functions and quaternions. It has found rich applications in symbolic computation, physics, robotics, computer graphics, etc. [56, 58, 132, 244]. Since bivectors and trivectors in the Clifford algebras of Euclidean vector spaces square to minus one, we can use them to create new geometric kernels for Fourier transformations. This leads to a large variety of new Fourier transformations, which all deserve to be studied in their own right [58, 117, 182, 184, 185, 243, 251, 254, 256, 259, 286, 290]. We will treat both Euclidean (positive definite metric) and nonEuclidean (indefinite metric) vector spaces. We know from Einstein’s special theory of relativity that nonEuclidean vector spaces are of fundamental importance in nature [158]. Therefore this section is about finding square roots of −1 in a nondegenerate Clifford algebra Gp,q = Cl(p, q). 2.7.3.1 Notation for multivector square root computations
The associative geometric product of two vectors a, b ∈ Rp,q , p + q = n is defined as the sum of their symmetric inner product (scalar) and their antisymmetric outer product (bivector) ab = a · b + a ∧ b. (2.7.56) We recall [247] a real Clifford algebra Gp,q = Cl(p, q) as the linear space of all elements generated by the associative (and distributive) bilinear geometric product of vectors of an inner product vector space Rp,q , p + q = n over the field of reals R. A Clifford algebra includes the field of reals R and the vector space Rp,q as grade zero and grade one elements, respectively. Clifford algebras in one, two and three dimensions have the following basis blades of grade 0 (scalars), grade 1 (vectors), grade 2 (bivectors) and grade 3 (trivectors) {1, e 1 , e 2 , e 3 , e23 , e31 , e12 , e123 },
(2.7.57)
where we use abbreviations e12 = e 1 e 2 , e23 = e 2 e 3 , e31 = e 3 e 1 , e123 = e 1 e 2 e 3 . Every multivector can be expanded in terms of these basis blades with real coefficients. We give examples for M ∈ Gp,q , n = p + q = 1, 2, 3: M = α + βe 1 ,
(2.7.58)
0
(2.7.59)
00
(2.7.60)
M = α + b1 e 1 + b2 e 2 + βe12 , M = α + b1 e 1 + b2 e 2 + b3 e 3 + c1 e23 + c2 e31 + c3 e12 + βe123 . The general notation for the quadratic form of basis vectors in Rp,q is: (
~ek2
= εk =
+1 for 1 ≤ k ≤ p, −1 for p + 1 ≤ k ≤ p + q = n.
(2.7.61)
We therefore always have ~ek4 = ε2k = 1, and we abbreviate Gp = Gp,0 . We follow the convention that inner and outer products have priority over the geometric product, which saves writing a number of brackets. Therefore, ~a · ~b ~c equals (~a · ~b )~c and not ~a · (~b~c), etc.
Clifford algebra 69
We will frequently use the following basic formulas of Clifford algebra in the rest of this work. The symmetric part of the geometric product of any two vectors ~a, ~b is the inner product (contraction, scalar product) 1 ~ ~ (~ab + b~a) = ~a · ~b = h~a~bi0 = h~a~bi. 2
(2.7.62)
Likewise the inner product (contraction, scalar product) of any two bivectors c, c0 is symmetric 1 0 (2.7.63) (c c + c0 c) = c · c0 = hc c0 i0 = hc c0 i. 2 The antisymmetric part of the geometric product of any two vectors ~a, ~b is the outer product (bivector) 1 ~ ~ (~ab − b~a) = ~a ∧ ~b = h~a~bi2 . (2.7.64) 2 The inner product (left contraction) of a vector ~a with a bivector c is antisymmetric 1 ~a · c = (~ac − c~a) = h~aci1 . 2
(2.7.65)
Let In = Πnk=1~ek be the unit oriented pseudoscalar of Gp,q , n = p+q. Let Ar , Bs ∈ Gp,q be two blades of grade r and s, respectively. Then we have the following general rules [244]. The inner product (left contraction [110]) is related to the outer product by22 (Ar · Bs )In = Ar ∧ (Bs In ), (Ar ∧ Bs )In = Ar · (Bs In ),
if r ≤ s;
(2.7.66)
if r + s ≤ n,
r, s > 0.
(2.7.67)
Two blades Ar , Bs ∈ Gp,q are called orthogonal if and only if their inner product is zero Ar ⊥ Bs ⇐⇒ Ar · Bs = 0. (2.7.68) With (2.7.66) follows that for r ≤ s, Ar ⊥ Bs
⇐⇒
Ar ∧ (Bs In ) = 0
⇐⇒
fs = 0, Ar ∧ B
(2.7.69)
fs = Bs I −1 is the dual of Bs with I −1 = ±In . Likewise (2.7.67) shows that where B n n for r + s ≤ n, r, s > 0 fs Ar ⊥ B
⇐⇒
Ar ∧ Bs = 0.
(2.7.70)
Example 2.7.2. Let ~b, c ∈ Gp,q , p+q = 3 be a vector ~b and a bivector c with vanishing outer product. Then by (2.7.70) the dual vector ~c = ec is always perpendicular to ~b independent of the signature of the underlying vector space Rp,q , p + q = 3, ~b ∧ c = 0 22
⇐⇒
~b · ~c = 0
⇐⇒
~b ⊥ ~c.
(2.7.71)
In order to avoid a discussion of deviating definitions of the inner product for r = 0 or s = 0, we exclude scalars in (2.7.67), but depending on the definition of the inner product (or contraction), a single general formula for all grades exists. For example, for the left contraction [110] P A, B ∈ Gp,q , AcB = hhAir hBis is−r we have the two formulas (A ∧ B)In = Ac(BIn ) and r,s (AcB)In = A ∧ (BIn ).
70 Quaternion and Clifford Fourier Transforms
2.7.3.2 Geometric multivector square roots of −1
Definition 2.7.3 (Geometric root of −1). A geometric multivector square root (geometric root) of −1 is a multivector A ∈ Gp,q with A2 = AA = −1.
(2.7.72)
An immediate application of this definition is the generalization of the famous Euler formula to geometric roots A of −1 eϕA = cos ϕ + A sin ϕ.
(2.7.73)
For example, Lounesto considers cos ϕ + e12 sin ϕ in G2 in [247, Page 29]. Theorem 2.7.4. Every multivector square root A of −1 is subject to n+1 = p+q +1 gradewise constraints: A2 = hAAi = −1,
(2.7.74)
and hAAik = 0,
1 ≤ k ≤ n,
(2.7.75)
where hAAik denotes the kth vector part of AA, and hAAi = hAAi0 . We point out that hAAi is identical to the scalar product A ∗ A of [161]. In the following we call the scalar equation (2.7.74) the root equation of Gp,q and (2.7.75) the constraints. Depending on the value of k, each kvector constraint represents n k scalar equations. We will sometimes conveniently split up a kvector constraint equation and still call the resulting partial equations constraints. Table 2.5 on page 72 lists all geometric roots of −1 of Gp,q , n = p + q = 4. We point out, that similar to Table 2.4, also in Table 2.5 all root equations of the third column result from the general n = 4 root equation A2 = hA2 i = −1, simply by inserting the case conditions and constraints of columns one and two. 2.7.3.3 Summary on multivector square roots of −1 in Clifford algebras Cl(p, q), n =
p+q ≤4
Table 2.4 lists all geometric roots of −1 for Clifford algebras Gp,q , n = p + q ≤ 3, and Table 2.5 does the same for Clifford algebras Gp,q , n = p + q = 4. The content of both tables has been checked with the MAPLE package CLIFFORD [5]. The solutions for G3 included in Table 2.4 correspond to the biquaternion roots of −1 found in [294]. Overall the calculations in [193] and the results demonstrate how in Clifford algebras extensive calculations can be done without referring to coordinates [1, 161]. In the case of Gp,q , n = p + q = 4, one arbitrarily selects one nonisotropic vector ~e4 for suitably splitting the algebra in order to use well developed techniques for algebras Gp,q , n = p + q = 3. In the end it is always possible to express the results in coordinates as in Table 2.4. However, this considerably blows up the expressions and blurs the mostly (p, q)signature independent form of the root equations of the
Clifford algebra 71
Geometric roots of −1 for Clifford algebras Gp,q , n = p + q ≤ 3. The multivectors are denoted for n = 1 by α + βe 1 , for n = 2 by α + b1 e 1 + b2 e 2 + βe12 and for n = 3 by α + b1 e 1 + b2 e 2 + b3 e 3 + c1 e23 + c2 e31 + c3 e12 + βe123 . Source: [193, Tab. 1]. Table 2.4
n Cases
Solutions A and root equations
1
No solution for G1 A = ±e1 for G0,1
2
3
α=0
β 2 = b21 ε2 + b22 ε1 + ε1 ε2 2 2 b1 + b2 + 1 for G2 2 2 2 −b1 + b2 − 1 for G1,1 β = −b2 − b2 + 1 for G 0,2 1 2
α 6= 0
No solution
Constraint: α=β=0
0 = b ∧ c = b1 c1 + b2 c2 + b3 c3 −1 = ~b2 + c2 −1 = b21 ε1 + b22 ε2 + b23 ε3 − c21 ε2 ε3 − c22 ε3 ε1 − c23 ε1 ε2 b21 + b22 + b23 − (c21 + c22 + c23 ) for G3 b21 − b22 − b23 − (c21 − c22 − c23 ) for G1,2 −1 = b21 + b22 − b23 + (c21 + c22 − c23 ) for G2,1 −(b21 + b22 + b23 ) − (c21 + c22 + c23 ) for G0,3
α = 0, β = 6 0 A = ±e123 for G3 , G1,2 No solution for G2,1 , G0,3 α 6= 0
No solution
72 Quaternion and Clifford Fourier Transforms
Geometric roots of −1 for Clifford algebras Gp,q , n = p + q = 4. The multivectors are denoted by α + ~b + c + β e123 + (α0 + ~b0 + c0 + β 0 e123 ) e 4 , for details see equ. (68) in [193]. Source: [193, Tab. 2].
Table 2.5
Case
Subcase / Constraints
α= 6 0 α = 0, α0 = 6 0
c = α10 ~b0 ∧ ~b, c0 = α10 (β 0~b0−ε4 β~b) e123
Solutions and root equations No solution ~b2 + 10 2 (~b0 ∧ ~b)2 + β 2 e2 + ε4 α0 2 − ε4~b02 123 α +ε4 10 2 (β 0~b0 − ε4 β~b)2 e2 − ε4 β 0 2 e2 123
α
123
= −1 α = 0, β = β 0 = 0 α0 = 0, c · c0 = 0, ~b0 = 0 ~b · c0 = 0, ~b ∧ c = 0 β = 0, β 0 6= 0 c = − β10 ~b · c0 e−1 123 β= 6 0 ~b = 0, c = 0
~b2 + c2 + ε4 c0 2 = −1
~b2 + 10 2 (~b · c0 )2 e2 + ε4 c0 2 − ε4 β 0 2 e2 123 123 β = −1 β 2 e2123 + ε4 c0 2 − ε4 β 0 2 e2123 = −1
α = 0, ~b = 0, β = 0 α0 = 0, β 0 = 0, c · c0 = 0, c2 − ε4~b02 + ε4 c0 2 = −1 ~b0 = ~b0 · c0 = 0, ~b0 ∧ c = 0 6 0 ~b = 0, β 6= 0 1 ~0 0 2 2 2 2 02 ~ 02 c = − εβ4 ~b0 · c0 e−1 123 , β 2 (b · c ) e123 + β e123 − ε4 b + ε4 c β0 = 0 = −1 ~b 6= 0, β = 0 c · c0 = 0, ~b ∧ c = 0, ~b0 = γ~b, γ
β 0 = 0, ~b · c0 = 0, (1 − ε4 γ 2 ) ~b2 + c2 + ε4 c0 2 = −1 ∈ R \ {0}
~b 6= 0, β 6= 0 β 0 6= 0, ~b0 = ε4 ββ0 ~b, c = − 10 ~b · c0 e−1 β
123
2
(1 − ε4 ββ0 2 )~b2 + +ε4 c0 2 −
1 ~ (b · c0 )2 e2123 β02 ε4 β 0 2 e2123 = −1
+ β 2 e2123
Clifford algebra 73
families of geometric roots of −1. In the case of Gp,q , n = p + q = 4, in Table 2.5, we have not expressed the results in coordinates, because then the table would extend over several pages. Open questions that are partly answered in the next Section 2.7.4 are: • How can the graded structure of Gp,q be used best in the calculation of higher order geometric multivector square roots of −1? This also includes a question how to best use, for this type of computation, invariance of the equation AA = −1 under Clifford algebra (anti) automorphisms such as grade involution, reversion or conjugation, and under symmetries of the root equation. For example, under the grade involution, AA = −1 ⇐⇒ AˆAˆ = −1.
(2.7.76)
Another example would be a rotor R symmetry AA = hAAi = −1 ⇐⇒ R
−1
ARR
−1
AR = hR
−1
ARR
(2.7.77) −1
ARi = hR
−1
AARi = hAAi = −1.
• The interesting relationship with families of idempotents of Clifford geometric algebras [3]. • What is the relationship with combinatorics? • Expansion of this work to Clifford algebras Gp,q in arbitrary dimension n = p+q. For this purpose, it will be appropriate to use the modulo eight periodicity of Clifford algebras and the isomorphisms with matrix rings. Central elements squaring to −1 would be of particular importance as then they can be used in place of the imaginary i. • The further use of Clifford algebra computation software like CLIFFORD for MAPLE and other packages [4, 5, 169]. Of special interest in physics are the Clifford algebras of Minkowski spacetime, sometimes called [158] spacetime algebras G3,1 and G1,3 . Table 2.5 contains the complete set of all geometric roots of −1 for these algebras, so in particular all possible geometric multivector elements that may take on the role of the imaginary unit i in quantum mechanics, which is e.g. fundamental for the description of spin and for wave propagation. Finally, the door is now wide open to construct all possible new types of Clifford Fourier transformations (CFT) [290] for multivector fields with domains and image domains ranging over the full Clifford algebras involved or subalgebras and subspaces thereof. In particular, all known Fourier transformations will find their place in this new general framework. The close relationship of wavelet transformations [188, 189, 253] and windowed transformations [254] to Fourier transformations shows that also in these fields new mathematics is to be expected. Examples of CFTs working with noncentral replacements of the imaginary unit i are the quaternion FT (QFT) [56, 58, 184, 256], and the CFT [117, 185] where i is
74 Quaternion and Clifford Fourier Transforms
replaced by pseudoscalars in Gn , n = 2 (mod 4). This shows that in principle every geometric root of −1, be it central or not, gives rise to its own geometric FT. Regarding the noncentral geometric roots of −1, the example of the QFT shows that the noncommutativity may indeed be of advantage for obtaining more information about the symmetry and the physical nature of signals thus processed. 2.7.4 General theory of square roots of −1 in real Clifford algebras
It is well known that Clifford (geometric) algebra offers a geometric interpretation for square roots of −1 in the form of blades that square to minus 1. This extends to a geometric interpretation of quaternions as the side face bivectors of a unit cube. Systematic research has been done [294] on the biquaternion roots of −1, abandoning the restriction to blades. Biquaternions are isomorphic to the Clifford (geometric) algebra C`(3, 0) of R3 . Further research on general algebras C`(p, q) has explicitly derived the geometric roots of −1 for p+q ≤ 4, as we just discussed in the Subsection 2.7.3, [193]. The current Subsection overcomes this dimension limit and by applying the Clifford algebra to matrix algebra isomorphisms in order to algebraically characterize the continuous manifolds of square roots of −1 found in the different types of Clifford algebras, depending on the type of associated ring (R, H, R2 , H2 or C). At the end of the book in Appendix A.1 explicit computer generated tables of representative square roots of −1 are given for all Clifford algebras with n = 5, 7, and s = 3 (mod 4) with the associated ring C. This includes, e.g., C`(0, 5) important in Clifford analysis, and C`(4, 1) which in applications is at the foundation of conformal geometric algebra. All these roots of −1 are immediately useful in the construction of new types of geometric Clifford Fourier transformations. This section is mainly based on [203]. 2.7.4.1 Introductory remarks
Beyond bivectors, quaternions and rotations, Clifford algebra allows to extend complex numbers to higher dimensions [33,161] and systematically generalize our knowledge of complex numbers, holomorphic functions and quaternions into the realm of Clifford analysis. It has found rich applications in symbolic computation, physics, robotics, computer graphics, etc. [56, 58, 112, 132, 244]. Since bivectors and trivectors in the Clifford algebras of Euclidean vector spaces square to minus one, we can use them to create new geometric kernels for Fourier transformations. This leads to a large variety of new Fourier transformations, which all deserve to be studied in their own right [58, 117, 182, 184, 185, 194, 243, 251, 254, 256, 259, 286, 290]. In this section, we will treat square roots of −1 in Clifford algebras C`(p, q) of both Euclidean (positive definite metric) and nonEuclidean (indefinite metric) nondegenerate vector spaces, Rn = Rn,0 and Rp,q , respectively. We know from Einstein’s special theory of relativity that nonEuclidean vector spaces are of fundamental importance in nature [158]. They are further, e.g., used in computer vision and robotics [112] and for general algebraic solutions to contact problems [244]. Therefore this section is about characterizing square roots of −1 in all Clifford algebras C`(p, q), extending previous limited research on C`(3, 0) in [294] and C`(p, q), n = p + q ≤ 4
Clifford algebra 75
Manifolds of square roots f of −1 in C`(2, 0) (left), C`(1, 1) (center), and ∼ C`(0, 2) = H (right). The square roots are f = α+b1 e1 +b2 e2 +βe12 , with α, b1 , b2 , β ∈ R, α = 0 and β 2 = b21 e22 + b22 e21 + e21 e22 . Source: [203, Fig. 1].
Figure 2.4
in [193]. The manifolds of square roots of −1 in C`(p, q), n = p + q = 2, compare Table 1 of [193], are visualized in Figure 2.4. First, we introduce necessary background knowledge of Clifford algebras and matrix ring isomorphisms and explain in more detail how we will characterize and classify the square roots of −1 in Clifford algebras in Section 2.7.4.2. Next, we treat section by section (in Sections 2.7.4.3 to 2.7.4.7) the square roots of −1 in Clifford algebras which are isomorphic to matrix algebras with associated rings R, H, R2 , H2 and C, respectively. The term associated means that the isomorphic matrices will only have matrix elements from the associated ring. The square roots of −1 in Section 2.7.4.7 with associated ring C are of particular interest, because of the existence of classes of exceptional square roots of −1, which all include a nontrivial term in the central element of the respective algebra different from the identity. Section 2.7.4.7 therefore includes a detailed discussion of all classes of square roots of −1 in the algebras C`(4, 1), the isomorphic C`(0, 5), and in C`(7, 0). Finally, we add Appendix A.1 with tables of square roots of −1 for all Clifford algebras with n = 5, 7, and s = 3 (mod 4). The square roots of −1 in Section 2.7.4.7 and in Appendix A.1 were all computed with the Maple package CLIFFORD [5], as explained in [204]. 2.7.4.2 Background and problem formulation
Let C`(p, q) be the algebra (associative with unit 1) generated over R by p + q elements ek (with k = 1, 2, . . . , p + q) with the relations e2k = 1 if k ≤ p, e2k = −1 if k > p and eh ek + ek eh = 0 whenever h 6= k, see [247]. We set the vector space dimension n = p + q and the signature s = p − q. This algebra has dimension 2n , and its even subalgebra C`0 (p, q) has dimension 2n−1 (if n > 0). We are concerned with square roots of −1 contained in C`(p, q) or C`0 (p, q). If the dimension of C`(p, q) or C`0 (p, q) is ≤ 2, it is isomorphic to R ∼ = C`(0, 0), R2 ∼ = C`(1, 0) or C ∼ = C`(0, 1), and 2 it is clear that there is no square root of −1 in R and R = R × R, and that there are
76 Quaternion and Clifford Fourier Transforms
two squares roots i and −i in C. Therefore we only consider algebras of dimension ≥ 4. Square roots of −1 have been computed explicitly in [294] for C`(3, 0), and in [193] for algebras of dimensions 2n ≤ 16. An algebra C`(p, q) or C`0 (p, q) of dimension ≥ 4 is isomorphic to one of the five matrix algebras: M(2d, R), M(d, H), M(2d, R2 ), M(d, H2 ) or M(2d, C). The integer d depends on n. According to the parity of n, it is either 2(n−2)/2 or 2(n−3)/2 for C`(p, q), and, either 2(n−4)/2 or 2(n−3)/2 for C`0 (p, q). The associated ring (either R, H, R2 , H2 or C) depends on s in this way23 : Table 2.6
Associated rings in matrix algebra isomorphisms. Source: [203, p. 4]. s mod 8 Associated ring for C`(p, q)
0 R
Associated ring for C`0 (p, q) R
2
1 R
2
R
2
3
R C
4 H
C H H
2
5 H
2
H
6
7
H C C R
Therefore we shall answer this question: What can we say about the square roots of −1 in an algebra A that is isomorphic to M(2d, R), M(d, H), M(2d, R2 ), M(d, H2 ), or, M(2d, C)? They constitute an algebraic submanifold in A; how many connected components24 (for the usual topology) does it contain? Which are their dimensions? This submanifold is invariant by the action of the group Inn(A) of inner automorphisms 25 of A, i.e. for every r ∈ A, r2 = −1 ⇒ f (r)2 = −1 ∀f ∈ Inn(A). The orbits of Inn(A) are called conjugacy classes26 ; how many conjugacy classes are there in this submanifold? If the associated ring is R2 or H2 or C, the group Aut(A) of all automorphisms of A is larger than Inn(A), and the action of Aut(A) in this submanifold shall also be described. We recall some properties of A that do not depend on the associated ring. The group Inn(A) contains as many connected components as the group G(A) of invertible elements in A. We recall that this assertion is true for M(2d, R) but not for M(2d + 1, R) which is not one of the relevant matrix algebras. If f is an element of A, let Cent(f ) be the centralizer of f , that is, the subalgebra of all g ∈ A such that f g = gf. The conjugacy class of f contains as many connected components27 as G(A) if (and 23
Compare chapter 16 on matrix representations and periodicity of 8, as well as Table 1 on p. 217 of [247]. 24 Two points are in the same connected component of a manifold, if they can be joined by a continuous path inside the manifold under consideration. (This applies to all topological spaces satisfying the property that each neighbourhood of any point contains a neighbourhood in which every pair of points can always be joined by a continuous path.) 25 An inner automorphism f of A is defined as f : A → A, f (x) = a−1 xa, ∀x ∈ A, with given fixed a ∈ A. The composition of two inner automorphisms g(f (x)) = b−1 a−1 xab = (ab)−1 x(ab) is again an inner automorphism. With this operation the inner automorphisms form the group Inn(A), compare [323]. 26 The conjugacy class (similarity class) of a given r ∈ A, r2 = −1 is {f (r) : f ∈ Inn(A)}, compare [321]. Conjugation is transitive, because the composition of inner automorphisms is again an inner automorphism. 27 According to the general theory of groups acting on sets, the conjugacy class (as a topological space) of a square root f of −1 is isomorphic to the quotient of G(A) and Cent(f ) (the subgroup
Clifford algebra 77
only if) Cent(f ) G(A) is contained in the neutral28 connected component of G(A), and the dimension of its conjugacy class is T
dim(A) − dim(Cent(f )).
(2.7.78)
Note that for invertible g ∈ Cent(f ) we have g −1 f g = f . Besides, let Z(A) be the center of A, and let [A, A] be the subspace spanned by all [f, g] = f g − gf . In all cases A is the direct sum of Z(A) and [A, A]. For example,29 Z(M(2d, R)) = {a1  a ∈ R} and Z(M(2d, C)) = {c1  c ∈ C}. If the associated ring is R or H (that is for even n), then Z(A) is canonically isomorphic to R, and from the projection A → Z(A) we derive a linear form Scal : A → R. When the associated ring30 is R2 or H2 or C, then Z(A) is spanned by 1 (the unit matrix31 ) and some element ω such that ω 2 = ±1. Thus, we get two linear forms ScalixScal!scalar coefficient and Spec such that Scal(f )1 + Spec(f )ω is the projection of f in Z(A) for every f ∈ A. Instead of ω we may use −ω and replace Spec with −Spec. The following assertion holds for every f ∈ A: The trace of each multiplication32 g 7→ f g or g 7→ gf is equal to the product tr(f ) = dim(A) Scal(f ).
(2.7.79)
The word “trace” (when nothing more is specified) means a matrix trace in R, which is the sum of its diagonal elements. For example, the matrix M ∈ M(2d, R) with P elements mkl ∈ R, 1 ≤ k, l ≤ 2d has the trace tr(M ) = 2d k=1 mkk [220]. We shall prove that in all cases Scal(f ) = 0 for every square root of −1 in A. Then, we may distinguish ordinary square roots of −1, and exceptional ones. In all cases the ordinary square roots of −1 constitute a unique33 conjugacy class of dimension dim(A)/2 which has as many connected components as G(A), and they satisfy the equality Spec(f ) = 0 if the associated ring is R2 or H2 or C. The exceptional square roots of −1 only exist34 if A ∼ = M(2d, C). In M(2d, C) there are 2d conjugacy classes of exceptional square roots of −1, each one characterized by an equality Spec(f ) = k/d with ±k ∈ {1, 2, . . . , d} [see Section 2.7.4.7], and their dimensions are < dim(A)/2 [see equation (2.7.99)]. For instance, ω (mentioned of stability of f ). Quotient means here the set of left handed classes modulo the subgroup. If the subgroup is contained in the neutral connected component of G(A), then the number of connected components is the same in the quotient as in G(A). See also [70]. 28 Neutral means to be connected to the identity element of A. 29 A matrix algebra based proof is, e.g., given in [312]. 30 This is the case for n (and s) odd. Then the pseudoscalar ω ∈ C`(p, q) is also in Z(C`(p, q)). 31 The number 1 denotes the unit of the Clifford algebra A, whereas the bold face 1 denotes the unit of the isomorphic matrix algebra M. 32 These multiplications are bilinear over the center of A. 33 Let A be an algebra M(m, K) where K is a division ring. Thus two elements f and g of A induce Klinear endomorphisms f 0 and g 0 on Km ; if K is not commutative, K operates on Km on the right side. The matrices f and g are conjugate (or similar) if and only if there are two Kbases B1 and B2 of Km such that f 0 operates on B1 in the same way as g 0 operates on B2 . This theorem allows us to recognize that in all cases but the last one (with exceptional square roots of −1), two square roots of −1 are always conjugate. 34 The pseudoscalars of Clifford algebras whose isomorphic matrix algebra has ring R2 or H2 square to ω 2 = +1.
78 Quaternion and Clifford Fourier Transforms
above) and −ω are central square roots of −1 in M(2d, C) which constitute two conjugacy classes of dimension 0. Obviously, Spec(ω) = 1. For symbolic computer algebra systems (CAS), like MAPLE, there exist Clifford algebra packages, e.g., CLIFFORD [5], which can compute idempotents [3] and square roots of −1. This will be of especial interest for the exceptional square roots of −1 in M(2d, C). Regarding a square root r of −1, a Clifford algebra is the direct sum of the subspaces Cent(r) (all elements that commute with r) and the skewcentralizer SCent(r) (all elements that anticommute with r). Every Clifford algebra multivector has a unique split by this Lemma. Lemma 2.7.5. Every multivector A ∈ C`(p, q) has, with respect to a square root r ∈ C`(p, q) of −1, i.e., r−1 = −r, the unique decomposition 1 A± = (A ± r−1 Ar), 2
A = A+ + A− ,
A+ r = rA+ ,
A− r = −rA− .
(2.7.80)
Proof. For A ∈ C`(p, q) and a square root r ∈ C`(p, q) of −1, we compute 1 1 r−1 =−r 1 (rr−1 Ar ± rA) A± r = (A ± r−1 Ar)r = (Ar ± r−1 A(−1)) = 2 2 2 1 = ±r (A ± r−1 Ar). 2
For example, in Clifford algebras C`(n, 0) [185] of dimensions n = 2 mod 4, Cent(r) is the even subalgebra C`0 (n, 0) for the unit pseudoscalar r, and the subspace C`1 (n, 0) spanned by all kvectors of odd degree k, is SCent(r). The most interesting case is M(2d, C), where a whole range of conjugacy classes becomes available. These results will therefore be particularly relevant for constructing Clifford Fourier transformations using the square roots of −1. 2.7.4.3 Square roots of −1 in M(2d, R)
Here A = M(2d, R), whence dim(A) = (2d)2 = 4d2 . The group G(A) has two connected components determined by the inequalities det(g) > 0 and det(g) < 0. For the case d = 1 we have, e.g., the algebra C`(2, 0) isomorphic to M(2, R). The basis {1, e1 , e2 , e12 } of C`(2, 0) is mapped to (
!
!
!
!)
1 0 0 1 1 0 0 −1 , , , 0 1 1 0 0 −1 1 0
.
The general element α + b1 e1 + b2 e2 + βe12 ∈ C`(2, 0) is thus mapped to α + b2 −β + b1 β + b1 α − b2
!
(2.7.81)
Clifford algebra 79
in M(2, R). Every element f of A = M(2d, R) is treated as an Rlinear endomorphism of V = R2d . Thus, its scalar component and its trace (2.7.79) are related as follows: tr(f ) = 2dScal(f ). If f is a square root of −1, it turns V into a vector space over C (if the complex number i operates like f on V ). If (e1 , e2 , . . . , ed ) is a Cbasis of V , then (e1 , f (e1 ), e2 , f (e2 ), . . . , ed , f (ed )) is a Rbasis of V , and the 2d×2d matrix of f in this basis is ! ! 0 −1 0 −1 ,..., diag . (2.7.82) 1 0 1 0 
{z d
}
Consequently all square roots of −1 in A are conjugate. The centralizer of a square root f of −1 is the algebra of all Clinear endomorphisms g of V (since i operates like f on V ). Therefore, the Cdimension of Cent(f ) is d2 and its Rdimension is 2d2 . Finally, the dimension (2.7.78) of the conjugacy class of f is dim(A)−dim(Cent(f )) = 4d2 −2d2 = 2d2 = dim(A)/2. The two connected components of G(A) are determined by the sign of the determinant. Because of the next lemma, the Rdeterminant of every T element of Cent(f ) is ≥ 0. Therefore, the intersection Cent(f ) G(A) is contained in the neutral connected component of G(A) and, consequently, the conjugacy class of f has two connected components like G(A). Because of the next lemma, the Rtrace of f vanishes (indeed its Ctrace is di, because f is the multiplication by the scalar i: f (v) = iv for all v) whence Scal(f ) = 0. This equality is corroborated by the matrix written above. We conclude that the square roots of −1 constitute one conjugacy class with two connected components of dimension dim(A)/2 contained in the hyperplane defined by the equation Scal(f ) = 0. (2.7.83) Before stating the lemma that here is so helpful, we show what happens in the easiest case d = 1. The square roots of −1 in M(2, R) are the real matrices a c b −a
!
!
with
a c b −a
!
a c = (a2 + bc) 1 = −1; b −a
(2.7.84)
hence a2 + bc = −1, a relation between a, b, c which is equivalent to (b − c)2 = (b + c)2 + 4a2 + 4 ⇒ (b − c)2 ≥ 4 ⇒ b − c ≥ 2 (one component) or c − b ≥ 2 (second component). Thus, we recognize the two connected components of square roots of −1: The inequality b ≥ c + 2 holds in one connected component, and the inequality c ≥ b + 2 in the other one, compare Figure 2.5. In terms of C`(2, 0) coefficients (2.7.81) with b − c = β + b1 − (−β + b1 ) = 2β, we get the two component conditions simply as β ≥ 1 (one component),
β ≤ −1 (second component).
(2.7.85)
Rotations (det(g) = 1) leave the pseudoscalar βe12 invariant (and thus preserve the two connected components of square roots of −1), but reflections (det(g 0 ) = −1) change its sign βe12 → −βe12 (thus interchanging the two components). Because of the previous argument involving a complex structure on the real space
80 Quaternion and Clifford Fourier Transforms
Figure 2.5
Two components of square roots of −1 in M(2, R). Source: [203, Fig. 2].
V , we conversely consider the complex space Cd with its structure of vector space over R. If (e1 , e2 , . . . , ed ) is a Cbasis of Cd , then (e1 , ie1 , e2 , ie2 , . . . , ed , ied ) is a Rbasis. Let g be a Clinear endomorphism of Cd (i.e. a complex d × d matrix), let trC (g) and detC (g) be the trace and determinant of g in C, and trR (g) and detR (g) its trace and determinant for the real structure of Cd . Exercise 2.7.6. For d = 1 an endomorphism of C1 is given by a complex number g = a + ib, a, b ∈ R. Its matrix representation is according to (2.7.82) !2
!
a −b b a
with
a −b b a
!
!
1 0 0 −1 = (a − b ) + 2ab . 0 1 1 0 2
2
(2.7.86)
!
Then we have trC (g) = a + ib, trR
a −b = 2a = 2 0 : identity,
det(g) > 0,
det(g 0 ) < 0 : (c1 , c2 ), (c3 , c4 ),
det(g) < 0,
det(g 0 ) > 0 : (c1 , c3 ), (c2 , c4 ),
det(g) < 0,
det(g 0 ) < 0 : (c1 , c4 ), (c2 , c3 ).
(2.7.89)
Beside the identity permutation, Inn(A) gives the three permutations that permute two elements and also the other two ones. The automorphisms outside Inn(A) are (f, f 0 ) 7→ (gf 0 g −1 , g 0 f g 0−1 ) for some (g, g 0 ) ∈ G(A).
(2.7.90)
If det(g) and det(g 0 ) have opposite signs, it is easy to realize that this automorphism induces a circular permutation on the four connected components of square roots of (−1, −1): If det(g) and det(g 0 ) have the same sign, this automorphism leaves globally invariant two connected components, and permutes the other two ones. For example, for d = 1 the automorphisms (2.7.90) outside Inn(A) permute the components (2.7.87) of square roots of (−1, −1) in M(2, R2 ) as follows: det(g) > 0,
det(g 0 ) > 0 : (c1 ), (c2 , c3 ), (c4 ),
det(g) > 0,
det(g 0 ) < 0 : c1 → c2 → c4 → c3 → c1 ,
det(g) < 0,
det(g 0 ) > 0 : c1 → c3 → c4 → c2 → c1 ,
det(g) < 0,
det(g 0 ) < 0 : (c1 , c4 ), (c2 ), (c3 ).
(2.7.91)
Consequently, the quotient of the group Aut(A) by its neutral connected component is isomorphic to the group of isometries of a square in a Euclidean plane. 2.7.4.5 Square roots of −1 in M(d, H)
Let us first consider the easiest case d = 1, when A = H, e.g., of C`(0, 2). The square roots of −1 in H are the quaternions ai + bj + cij with a2 + b2 + c2 = 1. They constitute a compact and connected manifold of dimension 2. Every square root f of −1 is conjugate with i, i.e. there exists v ∈ H : v −1 f v = i ⇔ f v = vi. If we set v = −f i + 1 = a + bij − cj + 1 we have f v = −f 2 i + f = f + i = (f (−i) + 1)i = vi. v is invertible, except when f = −i. But i is conjugate with −i because ij = j(−i), hence, by transitivity f is also conjugate with −i. Here A = M(d, H), whence dim(A) = 4d2 . The ring H is the algebra over R generated by two elements i and j such that i2 = j 2 = −1 and ji = −ij. We identify C with the subalgebra generated by36 i alone. The group G(A) has only one connected component. We shall soon prove that 36
This choice is usual and convenient.
Clifford algebra 83
every square root of −1 in A is conjugate with i1. Therefore, the submanifold of square roots of −1 is a conjugacy class, and it is connected. The centralizer of i1 in A is the subalgebra of all matrices with entries in C. The Cdimension of Cent(i1) is d2 , its Rdimension is 2d2 , and, consequently, the dimension (2.7.78) of the submanifold of square roots of −1 is 4d2 − 2d2 = 2d2 = dim(A)/2. Here V = Hd is treated as a (unitary) module over H on the right side: The product of a line vector t v = (x1 , x2 , . . . , xd ) ∈ V by y ∈ H is t v y = (x1 y, x2 y, . . . , xd y). Thus, every f ∈ A determines an Hlinear endomorphism of V : The matrix f multiplies the column vector v = t (x1 , x2 , . . . , xd ) on the left side v 7→ f v. Since C is a subring of H, V is also a vector space of dimension 2d over C. The scalar i always operates on the right side (like every scalar in H). If (e1 , e2 , . . . , ed ) is an Hbasis of V , then (e1 , e1 j, e2 , e2 j, . . . , ed , ed j) is a Cbasis of V . Let f be a square root of −1, then the eigenvalues of f in C are +i or −i. If we treat V as a 2d vector space over C, it is the direct (Clinear) sum of the eigenspaces V + = {v ∈ V  f (v) = vi} and V − = {v ∈ V  f (v) = −vi},
(2.7.92)
representing f as a 2d×2d Cmatrix w.r.t. the Cbasis of V , with Cscalar eigenvalues (multiplied from the right): λ± = ±i. Since ij = −ji, the multiplication v 7→ vj permutes V + and V − , as f (v) = ±vi is mapped to f (v)j = ±vij = ∓(vj)i. Therefore, if (e1 , e2 , . . . , er ) is a Cbasis of V + , then (e1 j, e2 j, . . . , er j) is a Cbasis of V − , consequently (e1 , e1 j, e2 , e2 j, . . . , er , er j) is a Cbasis of V , and (e1 , e2 , . . . , er=d ) is an Hbasis of V . Since f by f (ek ) = ek i for k = 1, 2, . . . , d operates on the Hbasis (e1 , e2 , . . . , ed ) in the same way as i1 on the natural Hbasis of V , we conclude that f and i1 are conjugate. Besides, Scal(i1) = 0 because 2i1 = [j1, ij1] ∈ [A, A], thus i1 ∈ / Z(A). Whence,37 Scal(f ) = 0 for every square root of
− 1.
(2.7.93)
These results are easily verified in the above example of d = 1 when A = H. 2.7.4.6 Square roots of −1 in M(d, H2 )
Here, A = M(d, H2 ) = M(d, H) × M(d, H), whence dim(A) = 8d2 . The group G(A) has only one connected component (see Footnote 35). The square roots of (−1, −1) in A are pairs of two square roots of −1 in M(d, H). Consequently, they constitute a unique conjugacy class which is connected and its dimension is 2 × 2d2 = 4d2 = dim(A)/2. For every (f, f 0 ) ∈ A we can write Scal(f )+Scal(f 0 ) = 2 Scal(f, f 0 ) and, similarly to (2.7.88), Scal(f ) − Scal(f 0 ) = 2 Spec(f, f 0 ) if ω = (1, −1); (2.7.94) whence Scal(f, f 0 ) = Spec(f, f 0 ) = 0 if (f, f 0 ) is a square root of (−1, −1), compare with (2.7.93). 37 Compare the definition of Scal(f ) in Section 2.7.4.2, remembering that in the current section the associated ring is H.
84 Quaternion and Clifford Fourier Transforms
The group Aut(A) has two38 connected components; the neutral component is Inn(A), and the other component contains the swap automorphism (f, f 0 ) 7→ (f 0 , f ). The simplest example is d = 1, A = H2 , where we have the identity pair (1, 1) representing the scalar part, the idempotents (1, 0), (0, 1), and ω as the pair (1, −1). A = H2 is isomorphic to C`(0, 3). The pseudoscalar ω = e123 has the square ω 2 = +1. The center of the algebra is {1, ω}, and includes the idempotents ± = 12 (1±ω), 2± = ± , + − = − + = 0. The basis of the algebra can thus be written as {+ , e1 + , e2 + , e12 + , − , e1 − , e2 − , e12 − } where the first (and the last) four elements form a basis of the subalgebra C`(0, 2) isomorphic to H. 2.7.4.7 Square roots of −1 in M(2d, C)
The lowest dimensional example for d = 1 is the Pauli matrix algebra A = M(2, C) isomorphic to the geometric algebra C`(3, 0) of the 3D Euclidean space and C`(1, 2). The C`(3, 0) vectors e1 , e2 , e3 correspond onetoone to the Pauli matrices !
0 1 , σ1 = 1 0
!
!
0 −i σ2 = , i 0
1 0 σ3 = , 0 −1
(2.7.95)
!
with σ1 σ2 = iσ3 = tral pseudoscalar e123 following idempotents
i 0 . The element ω = σ1 σ2 σ3 = i1 represents the cen0 −i of C`(3, 0) with square ω 2 = −1. The Pauli algebra has the
1 = σ12 = 1,
1 0 = (1 + σ3 ), 2
−1 = 0 .
(2.7.96)
The idempotents correspond via f = i(2 − 1),
(2.7.97)
to the square roots of −1: !
i 0 f1 = i1 = , 0 i
!
i 0 f0 = iσ3 = , 0 −i
!
f−1
−i 0 = −i1 = , 0 −i
(2.7.98)
where by complex conjugation f−1 = f1 . Let the idempotent 00 = 12 (1 − σ3 ) correspond to the matrix f00 = −iσ3 . We observe that f0 is conjugate to f00 = σ1−1 f0 σ1 = σ1 σ2 = f0 using σ1−1 = σ1 but f1 is not conjugate to f−1 . Therefore, only f1 , f0 , f−1 lead to three distinct conjugacy classes of square roots of −1 in M(2, C). Compare [204] for the corresponding computations with CLIFFORD for Maple. In general, if A = M(2d, C), then dim(A) = 8d2 . The group G(A) has one connected component. The square roots of −1 in A are in bijection with the idempotents [3] according to (2.7.97). According39 to (2.7.97) and its inverse = 12 (1 − if ) 38
Compare Footnote 35. On the other hand it is clear that complex conjugation always leads to f− = f+ , where the overbar means complex conjugation in M(2d, C) and Clifford conjugation in the isomorphic Clifford algebra C`(p, q). So either the trivial idempotent − = 0 is included in the bijection (2.7.97) of idempotents and square roots of −1 or alternatively the square root of −1 with Spec(f− ) = −1 is obtained from f− = f+ . 39
Clifford algebra 85
the square root of −1 with Spec(f− ) = k/d = −1, i.e. k = −d (see below), always corresponds to the trivial idempotent − = 0, and the square root of −1 with Spec(f+ ) = k/d = +1, k = +d, corresponds to the identity idempotent + = 1. If f is a square root of −1, then V = C2d is the direct sum of the eigenspaces40 associated with the eigenvalues i and −i. There is an integer k such that the dimensions of the eigenspaces are respectively d + k and d − k. Moreover, −d ≤ k ≤ d. Two square roots of −1 are conjugate if and only if they give the same integer k. Then, all elements of Cent(f ) consist of diagonal block matrices with 2 square blocks of (d + k) × (d + k) matrices and (d − k) × (d − k) matrices. Therefore, the Cdimension of Cent(f ) is (d + k)2 + (d − k)2 . Hence the Rdimension (2.7.78) of the conjugacy class of f : 8d2 − 2(d + k)2 − 2(d − k)2 = 4(d2 − k 2 ). (2.7.99) Also, from the equality tr(f ) = (d + k)i − (d − k)i = 2ki we deduce that Scal(f ) = 0 and that Spec(f ) = (2ki)/(2di) = k/d if ω = i1 (whence tr(ω) = 2di). As announced on page 77, we consider that a square root of −1 is ordinary if the associated integer k vanishes, and that it is exceptional if k = 6 0 . Thus the following assertion is true in all cases: the ordinary square roots of −1 in A constitute one conjugacy class of dimension dim(A)/2 which has as many connected components as G(A), and the equality Spec(f ) = 0 holds for every ordinary square root of −1 when the linear form Spec exists. All conjugacy classes of exceptional square roots of −1 have a dimension < dim(A)/2. All square roots of −1 in M(2d, C) constitute (2d + 1) conjugacy classes41 which are also the connected components of the submanifold of square roots of −1 because of the equality Spec(f ) = k/d, which is conjugacy class specific. When A = M(2d, C), the group Aut(A) is larger than Inn(A) since it contains the complex conjugation (that maps every entry of a matrix to the conjugate complex number). It is clear that the class of ordinary square roots of −1 is invariant by complex conjugation. But the class associated with an integer k other than 0 is mapped by complex conjugation to the class associated with −k. In particular, the complex conjugation maps the class {ω} (associated with k = d) to the class {−ω} associated with k = −d. All these observations can easily verified for the above example of d = 1 of the Pauli matrix algebra A = M(2, C). For d = 2 we have the isomorphism of A = M(4, C) with C`(0, 5), C`(2, 3) and C`(4, 1). While C`(0, 5) is important in Clifford analysis, C`(4, 1) is both the geometric algebra of the Lorentz space R4,1 and the conformal geometric algebra of 3D Euclidean geometry. Its set of square roots of −1 is therefore of particular practical interest. Exercise 2.7.8. Let C`(4, 1) ∼ = A where A = M(4, C) for d = 2. The C`(4, 1) 40 The following theorem is sufficient for a matrix f in M(m, K), if K is a (commutative) field. The matrix f is diagonalizable if and only if P (f ) = 0 for some polynomial P that has only simple roots, all of them in the field K. (This implies that P is a multiple of the minimal polynomial, but we do not need to know whether P is or is not the minimal polynomial.) 41 Two conjugate (similar) matrices have the same eigenvalues and the same trace. This suffices to recognize that 2d + 1 conjugacy classes are obtained.
86 Quaternion and Clifford Fourier Transforms
1vectors can be represented42 by the following matrices:
0 1 0 0 0 1 0 −1 0 0 e1 = , e2 = 0 0 0 −1 0 0 0 0 0 1
0 0 0 0 e4 = 1 0 0 −1
1 0 0 0
0 0 0 1
0 −i 0 0 0 i 0 0 0 0 , , e = 1 3 0 0 0 −i 0 0 i 0 0
0 0 −1 0 1 0 0 0 0 1 0 −1 . , e5 = 1 0 0 0 0 0 0 −1 0 0 0 0
(2.7.100)
We find five conjugacy classes of roots fk of −1 in C`(4, 1) for k ∈ {0, ±1, ±2}: four exceptional and one ordinary. Since fk is a root of p(t) = t2 + 1 which factors over C into (t − i)(t + i), the minimal polynomial mk (t) of fk is one of the following: t − i, t + i or (t − i)(t + i). Respectively, there are three classes of characteristic polynomial ∆k (t) of the matrix Fk in M(4, C) which corresponds to fk , namely, (t − i)4 , (t + i)4 and (t − i)n1 (t + i)n2 , where n1 + n2 = 2d = 4 and n1 = d + k = 2 + k, n2 = d − k = 2 − k. As predicted by the above discussion, the ordinary root corresponds to k = 0 whereas the exceptional roots correspond to k = 6 0. 1. For k = 2, we have ∆2 (t) = (t − i)4 , m2 (t) = t − i, and so F2 = diag(i, i, i, i) which in the above representation (2.7.100) corresponds to the nontrivial central element f2 = ω = e12345 . Clearly, Spec(f2 ) = 1 = kd ; Scal(f2 ) = 0; the Cdimension of the centralizer Cent(f2 ) is 16; and the Rdimension of the conjugacy class of f2 is zero as it contains only f2 since f2 ∈ Z(A). Thus, the Rdimension of the class is again zero in agreement with (2.7.99). 2. For k = −2, we have ∆−2 (t) = (t + i)4 , m−2 (t) = t + i, and F−2 = diag(−i, −i, −i, −i) which corresponds to the central element f−2 = −ω = −e12345 . Again, Spec(f−2 ) = −1 = kd ; Scal(f−2 ) = 0; the Cdimension of the centralizer Cent(f−2 ) is 16 and the conjugacy class of f−2 contains only f−2 since f−2 ∈ Z(A). Thus, the Rdimension of the class is again zero in agreement with (2.7.99). 3. For k 6= ±2, we consider three subcases when k = 1, k = 0 and k = −1. When k = 1, then ∆1 (t) = (t − i)3 (t + i) and m1 (t) = (t − i)(t + i). Then the root F1 = diag(i, i, i, −i) corresponds to 1 f1 = (e23 + e123 − e2345 + e12345 ). 2 Note that Spec(f1 ) = 42
1 2
=
k d
(2.7.101)
so f1 is an exceptional root of −1.
For the computations of this example in the Maple package CLIFFORD we have used the identification i = e23 . Yet the results obtained for the square roots of −1 are independent of this setting (we can alternatively use, e.g., i = e12345 , or the imaginary unit i ∈ C), as can easily be checked for f1 of (2.7.101), f0 of (2.7.102) and f−1 of (2.7.103) by only assuming the standard Clifford product rules for e1 to e5 .
Clifford algebra 87
When k = 0, then ∆0 (t) = (t − i)2 (t + i)2 and m0 (t) = (t − i)(t + i). Thus the root of −1 in this case is F0 = diag(i, i, −i, −i) which corresponds to just f0 = e123 .
(2.7.102)
Note that Spec(f0 ) = 0 thus f0 = e123 is an ordinary root of −1. When k = −1, then ∆−1 (t) = (t − i)(t + i)3 and m−1 (t) = (t − i)(t + i). Then, the root of −1 in this case is F−1 = diag(i, −i, −i, −i) which corresponds to 1 f−1 = (e23 + e123 + e2345 − e12345 ). 2
(2.7.103)
Since Scal(f−1 ) = − 12 = kd , we gather that f−1 is an exceptional root. As expected, we can also see that the roots ω and −ω are related via the grade involution whereas f1 = −f˜−1 where ˜ denotes the reversion in C`(4, 1). Exercise 2.7.9. Let C`(0, 5) ∼ = A where A = M(4, C) for d = 2. The C`(0, 5) 43 1vectors can be represented by the following matrices:
−i 0 −i 0 0 0 −1 0 0 0 −i 0 1 0 0 0 0 0 e1 = , e = , e2 = 0 0 0 0 −1 0 0 −i 3 0 0 0 0 −i 0 0 0 1 0
0 0 0 −1 0 0 0 0 1 0 e4 = , e = 1 0 0 0 5 −i 0 0 −1 0 0
0 i 0 0
0 0 0 0 , i 0 0 −i
0 −i 0 0 0 i . 0 0 0 i 0 0
(2.7.104)
Like for C`(4, 1), we have five conjugacy classes of the roots fk of −1 in C`(0, 5) for k ∈ {0, ±1, ±2}: four exceptional and one ordinary. Using the same notation as in Example 2.7.8, we find the following representatives of the conjugacy classes. 1. For k = 2, we have ∆2 (t) = (t − i)4 , m2 (t) = t − i and F2 = diag(i, i, i, i) which in the above representation (2.7.104) corresponds to the nontrivial central element f2 = ω = e12345 . Then, Spec(f2 ) = 1 = kd ; Scal(f2 ) = 0; the Cdimension of the centralizer Cent(f2 ) is 16; and the Rdimension of the conjugacy class of f2 is zero as it contains only f2 since f2 ∈ Z(A). Thus, the Rdimension of the class is again zero in agreement with (2.7.99). 2. For k = −2, we have ∆−2 (t) = (t + i)4 , m−2 (t) = t + i and F−2 = diag(−i, −i, −i, −i) which corresponds to the central element f−2= − ω = 43
For the computations of this example in the Maple package CLIFFORD we have used the identification i = e3 . Yet the results obtained for the square roots of −1 are independent of this setting (we can alternatively use, e.g., i = e12345 or the imaginary unit i ∈ C), as can easily be checked for f1 of (2.7.105), f0 of (2.7.106) and f−1 of (2.7.107) by only assuming the standard Clifford product rules for e1 to e5 .
88 Quaternion and Clifford Fourier Transforms
−e12345 . Again, Spec(f−2 ) = −1 = kd ; Scal(f−2 ) = 0; the Cdimension of the centralizer Cent(f−2 ) is 16 and the conjugacy class of f−2 contains only f−2 since f−2 ∈ Z(A). Thus, the Rdimension of the class is again zero in agreement with (2.7.99). 3. For k 6= ±2, we consider three subcases when k = 1, k = 0 and k = −1. When k = 1, then ∆1 (t) = (t − i)3 (t + i) and m1 (t) = (t − i)(t + i). Then the root F1 = diag(i, i, i, −i) corresponds to 1 f1 = (e3 + e12 + e45 + e12345 ). 2 Since Spec(f1 ) =
1 2
(2.7.105)
= kd , f1 is an exceptional root of −1.
When k = 0, then ∆0 (t) = (t − i)2 (t + i)2 and m0 (t) = (t − i)(t + i). Thus the root of −1 is this case is F0 = diag(i, i, −i, −i) which corresponds to just f0 = e45 .
(2.7.106)
Note that Spec(f0 ) = 0 thus f0 = e45 is an ordinary root of −1. When k = −1, then ∆−1 (t) = (t − i)(t + i)3 and m−1 (t) = (t − i)(t + i). Then, the root of −1 in this case is F−1 = diag(i, −i, −i, −i) which corresponds to 1 f−1 = (−e3 + e12 + e45 − e12345 ). 2
(2.7.107)
Since Scal(f−1 ) = − 12 = kd , we gather that f−1 is an exceptional root. Again we can see that the roots f2 and f−2 are related via the grade involution whereas f1 = −f˜−1 where ˜ denotes the reversion in C`(0, 5). Exercise 2.7.10. Let C`(7, 0) ∼ = A where A = M(8, C) for d = 4. We have nine conjugacy classes of roots fk of −1 for k ∈ {0, ±1, ±2 ± 3 ± 4}. Since fk is a root of a polynomial p(t) = t2 + 1 which factors over C into (t − i)(t + i), its minimal polynomial m(t) will be one of the following: t − i, t + i or (t − i)(t + i) = t2 + 1. Respectively, each conjugacy class is characterized by a characteristic polynomial ∆k (t) of the matrix Mk ∈ M(8, C) which represents fk . Namely, we have ∆k (t) = (t − i)n1 (t + i)n2 , where n1 + n2 = 2d = 8 and n1 = d + k = 4 + k and n2 = d − k = 4 − k. The ordinary root of −1 corresponds to k = 0 whereas the exceptional roots correspond to k= 6 0. 8 8
z } {
1. When k = 4, we have ∆4 (t) = (t − i) , m4 (t) = t − i and F4 = diag(i, . . . , i) which in the representation used by CLIFFORD [5] corresponds to the nontrivial central element f4 = ω = e1234567 . Clearly, Spec(f4 ) = 1 = kd ; Scal(f4 ) = 0; the Cdimension of the centralizer Cent(f4 ) is 64; and the Rdimension of the conjugacy class of f4 is zero since f4 ∈ Z(A). Thus, the Rdimension of the class is again zero in agreement with (2.7.99).
Clifford algebra 89
2. When k = −4, we have ∆−4 (t) = (t + i)8 , m−4 (t) = t + i and F−4 = 8
z
}
{
diag(−i, . . . , −i) which corresponds to f−4 = −ω = −e1234567 . Again, Spec(f−4 ) = −1 = kd ; Scal(f−4 ) = 0; the Cdimension of the centralizer Cent(f ) is 64 and the conjugacy class of f−4 contains only f−4 since f−4 ∈ Z(A). Thus, the Rdimension of the class is again zero in agreement with (2.7.99). 3. When k = 6 ±4, we consider seven subcases when k = ±3, k = ±2, k = ±1 and k = 0. When k = 3, then ∆3 (t) = (t − i)7 (t + i) and m3 (t) = (t − i)(t + i). Then the 7
z } {
root F3 = diag(i, . . . , i, −i) corresponds to 1 f3 = (e23 − e45 + e67 − e123 + e145 − e167 + e234567 + 3e1234567 ). 4 Since Spec(f3 ) =
3 4
(2.7.108)
= kd , f3 is an exceptional root of −1.
When k = 2, then ∆2 (t) = (t − i)6 (t + i)2 and m2 (t) = (t − i)(t + i). Then the 6
z } {
root F2 = diag(i, . . . , i, −i, −i) corresponds to 1 f2 = (e67 − e45 − e123 + e1234567 ). 2 Since Spec(f2 ) =
1 2
(2.7.109)
= kd , f2 is also an exceptional root.
When k = 1, then ∆1 (t) = (t − i)5 (t + i)3 and m1 (t) = (t − i)(t + i). Then the 5
z } {
root F1 = diag(i, . . . , i, −i, −i, −i) corresponds to 1 f1 = (e23 − e45 + 3e67 − e123 + e145 + e167 − e234567 + e1234567 ). 4 Since Spec(f1 ) =
1 4
(2.7.110)
= kd , f1 is another exceptional root.
When k = 0, then ∆0 (t) = (t − i)4 (t + i)4 and m0 (t) = (t − i)(t + i). Then the root F0 = diag(i, i, i, i, −i, −i, −i, −i) corresponds to 1 f0 = (e23 − e45 + e67 − e234567 ). 2
(2.7.111)
Since Spec(f0 ) = 0 = kd , we see that f0 is an ordinary root of −1. When k = −1, then ∆−1 (t) = (t − i)3 (t + i)5 and m−1 (t) = (t − i)(t + i). Then 5
z
}
{
the root F−1 = diag(i, i, i, −i, . . . , −i) corresponds to 1 f−1 = (e23 − e45 + 3e67 + e123 − e145 − e167 − e234567 − e1234567 ). (2.7.112) 4
90 Quaternion and Clifford Fourier Transforms
Thus, Spec(f−1 ) = − 14 =
k d
and so f−1 is another exceptional root.
When k = −2, then ∆−2 (t) = (t − i)2 (t + i)6 and m−2 (t) = (t − i)(t + i). Then 6
z
}
{
the root F−2 = diag(i, i, −i, . . . , −i) corresponds to 1 f−2 = (e67 − e45 + e123 − e1234567 ). (2.7.113) 2 Since Spec(f−2 ) = − 21 = kd , we see that f−2 is also an exceptional root. When k = −3, then ∆−3 (t) = (t − i)(t + i)7 and m−3 (t) = (t − i)(t + i). Then 7
z
}
{
the root F−3 = diag(i, −i, . . . , −i) corresponds to 1 f−3 = (e23 − e45 + e67 + e123 − e145 + e167 + e234567 − 3e1234567 ). (2.7.114) 4 Again, Spec(f−3 ) = − 34 = kd and so f−3 is another exceptional root of −1. As expected, we can also see that the roots ω and −ω are related via the reversion whereas f3 = −f¯−3 , f2 = −f¯−2 , f1 = −f¯−1 where ¯ denotes the conjugation in C`(7, 0). 2.7.4.8 Square roots of −1 in conformal geometric algebra Cl(4, 1)
We pay special attention to the square roots of −1 in conformal geometric algebra Cl(4, 1), because of the enormous practical importance of this algebra in applications to robotics, computer graphics, robot and computer vision, virtual reality, visualization and the like [202]. See Table 2.7 for representative exceptional (k 6= 0) square roots of −1 in conformal geometric algebra Cl(4, 1) of threedimensional Euclidean space [203]. k
fk
∆k (t)
2
ω = e12345
(t − i)4
1
1 2 (e23
0
e123
−1
1 2 (e23
+ e123 − e2345 + e12345 ) (t − i)3 (t + i) (t − i)2 (t + i)2 + e123 + e2345 − e12345 ) (t − i)(t + i)3
−2 −ω = −e12345
(t + i)4
Square roots of −1 in conformal geometric algebra Cl(4, 1) ∼ = M(4, C), d = 2, with characteristic polynomials ∆k (t). See [203] for details. Source: modified version of [203, Tab. 2].
Table 2.7
Ordinary square roots of −1 in Cl(4, 1) with k = 0
In the algebra basis of Cl(4, 1) there are nine blades which represent ordinary square roots of −1:
2.7.4.8.1
e5 ,
e234 , e134 , e124 , e123 ,
e2345 , e1345 , e1245 , e1235 .
(2.7.115)
Clifford algebra 91
But remembering the work in [193], we know that even if we only look at the subalgebras Cl(4, 0) or Cl(3, 1), which do not contain the pseudoscalar e12345 , and contain therefore only ordinary square roots of −1 for Cl(4, 1), we have long parametrized expressions for ordinary square roots of −1. But because of the high dimensionality it may not be easy to compute a complete expression for the whole 16D submanifold of ordinary square roots of −1 in Cl(4, 1) by hand. 2.7.4.8.2 Exceptional square roots of −1 in Cl(4, 1) with k = 1 In this case we can generalize Table 2.7 to patches of the 12dimensional submanifold of exceptional square roots of −1 in Cl(4, 1). In the future a complete parametrized expression obtained, e.g., with Clifford for Maple would be very desirable. We begin with the general expression
f1 = (
1+u 1−u E+ ) ω, 2 2
ω = e12345 ,
(2.7.116)
where we assume that E, u ∈ Cl(4, 1), E 2 = u2 = +1. This makes the expressions 1±u 2 1±u 1±u 2 = 2 . In the following we put forward certain values become idempotents 2 for E and u which will yield linearly independent patches of the 12dimensional √ submanifold of −1. • E = ve5 , v ∈ R4 , v 2 = 1, u ∈ R3⊥v , u2 = 1 gives a 3D × 2D = 6D submanifold. As a concrete example in this submanifold we can e.g. set v = e4 , u = e1 and get 1 1 f1 = [(1 + e1 )e45 + 1 − e1 ]ω = [e45 + e145 + 1 − e1 ]ω. (2.7.117) 2 2 √ • E = e1234 , u ∈ R4 , u2 = 1 gives a 3D submanifold of −1. A concrete example is e.g. u = e1 , then 1 1 f1 = [(1 + e1 )e1234 + 1 − e1 ]ω = [e1234 + e234 + 1 − e1 ]ω. 2 2
(2.7.118)
• E = v, v ∈ R4 , v 2 = 1, u = e1234 gives another 3D submanifold. A concrete example is e.g. v = e1 and gives 1 1 f1 = [(1 + e1234 )e1 + 1 − e1234 ]ω = [e1 − e234 + 1 − e1234 ]ω. 2 2 Exceptional square roots of −1 in Cl(4, 1) with k = −1 analogous to k = +1 by starting with
2.7.4.8.3
f−1 = (
1−u 1+u E− )ω, 2 2
ω = e12345 .
(2.7.119)
This is completely
(2.7.120)
Exceptional square roots of −1 in Cl(4, 1) with k = ±2
The exceptional square roots of −1 are zerodimensional in this case and therefore uniquely given by 2.7.4.8.4
f±2 = ±e12345 .
(2.7.121)
92 Quaternion and Clifford Fourier Transforms
2.7.4.9 Summary on general multivector square roots of −1
We proved that in all cases Scal(f ) = 0 for every square root of −1 in A isomorphic to C`(p, q). We distinguished ordinary square roots of −1, and exceptional ones. In all cases the ordinary square roots f of −1 constitute a unique conjugacy class of dimension dim(A)/2 which has as many connected components as the group G(A) of invertible elements in A. Furthermore, we have Spec(f ) = 0 (zero pseudoscalar part) if the associated ring is R2 , H2 or C. The exceptional square roots of −1 only exist if A ∼ = M(2d, C) (see Section 2.7.4.7). For A = M(2d, R) of Section 2.7.4.3, the centralizer and the conjugacy class of a square root f of −1 both have Rdimension 2d2 with two connected components, pictured in Figure 2.5 for d = 1. For A = M(2d, R2 ) = M(2d, R)×M(2d, R) of Section 2.7.4.4, the square roots of (−1, −1) are pairs of two square roots of −1 in M(2d, R). They constitute a unique conjugacy class with four connected components, each of dimension 4d2 . Regarding the four connected components, the group Inn(A) induces the permutations of the Klein group whereas the quotient group Aut(A)/Inn(A) is isomorphic to the group of isometries of a Euclidean square in 2D. For A = M(d, H) of Section 2.7.4.5, the submanifold of the square roots f of −1 is a single connected conjugacy class of Rdimension 2d2 equal to the Rdimension of the centralizer of every f . The easiest example is H itself for d = 1. For A = M(d, H2 ) = M(2d, H) × M(2d, H) of Section 2.7.4.6, the square roots of (−1, −1) are pairs of two square roots (f, f 0 ) of −1 in M(2d, H) and constitute a unique connected conjugacy class of Rdimension 4d2 . The group Aut(A) has two connected components: the neutral component Inn(A) connected to the identity and the second component containing the swap automorphism (f, f 0 ) 7→ (f 0 , f ). The simplest case for d = 1 is H2 isomorphic to C`(0, 3). For A = M(2d, C) of Section 2.7.4.7, the square roots of −1 are in bijection to the idempotents. First, the ordinary square roots of −1 (with k = 0) constitute a conjugacy class of Rdimension 4d2 of a single connected component which is invariant under Aut(A). Second, there are 2d conjugacy classes of exceptional square roots of −1, each composed of a single connected component, characterized by equality Spec(f ) = k/d (the pseudoscalar coefficient) with ±k ∈ {1, 2, . . . , d}, and their Rdimensions are 4(d2 −k 2 ). The group Aut(A) includes conjugation of the pseudoscalar ω 7→ −ω which maps the conjugacy class associated with k to the class associated with −k. The simplest case for d = 1 is the Pauli matrix algebra isomorphic to the geometric algebra C`(3, 0) of 3D Euclidean space R3 , and to complex biquaternions [294]. Section 2.7.4.7 includes explicit examples for d = 2: C`(4, 1) and C`(0, 5), and for d = 4: C`(7, 0). Appendix A.1 summarizes the square roots of −1 in all C`(p, q) ∼ = M(2d, C) for d = 1, 2, 4. [204] contains details on how square roots of −1 can be computed using the package CLIFFORD for Maple. Among the many possible applications of these findings, the possibility of new integral transformations in Clifford analysis is very promising. This field thus obtains essential algebraic information, which can e.g., be used to create steerable
Clifford algebra 93
transformations, which may be steerable within a connected component of a submanifold of square roots of −1.
2.8 2.8.1
QUATERNIONS AND THE GEOMETRY OF ROTATIONS IN THREE AND FOUR DIMENSIONS Introduction to quaternions
The electromagnetic field equations were originally formulated by J. C. Maxwell [258] in the language of Hamilton’s quaternions [155]. Later, among many other applications, quaternions began to play an important role in aerospace engineering [238], color signal processing [127], and in material science for texture analysis (understood as the distribution of crystallographic orientations of a polycrystalline sample [328]) [12, 261]. During the last decade the general orthogonal planes split (OPS) with respect to any two pure unit quaternions f, g ∈ H, f 2 = g 2 = −1, including the case f = g, has proved extremely useful for the construction and geometric interpretation of general classes of doublekernel quaternion Fourier transformations (QFT) [205]. Applications include color image processing, where the OPS with f = g = the greyline, naturally splits a pure quaternionic threedimensional color signal into luminance and chrominance components. Yet it is found independently in the quaternion geometry of rotations [261], that the pure quaternion units f, g and the analysis planes, which they define, play a key role in the geometry of rotations, and the geometrical interpretation of integrals related to the spherical Radon transform of probability density functions of unit quaternions, as relevant for texture analysis in crystallography. In our current section, we further investigate these connections. Gauss, Rodrigues and Hamilton’s [273, 274] fourdimensional (4D) quaternion algebra H is defined over R with three imaginary units i, j , k and multiplication laws: ij = −ji = k, 2
2
jk = −kj = i,
ki = −ik = j,
2
i = j = k = ijk = −1.
(2.8.1)
The explicit form of a quaternion q ∈ H is q = qr + qi i + qj j + qk k ∈ H,
qr , qi , qj , qk ∈ R.
(2.8.2)
Quaternions are isomorphic to the Clifford geometric algebra Cl0,2 of R0,2 , and to + the even subalgebra Cl3,0 of the Clifford geometric algebra Cl3,0 of R3 , i.e. H is isomorphic to the algebra of rotation operators in Cl(3, 0): + H∼ . = Cl0,2 ∼ = Cl3,0
(2.8.3)
+ Cl3,0 has, with an orthonormal basis {e 1 , e 2 , e 3 } of R3 , the four dimensional basis
{1, e 32 = e 3 e 2 , e 13 = e 1 e 3 , e 21 = e 2 e 1 }.
(2.8.4)
94 Quaternion and Clifford Fourier Transforms
The quaternion conjugate (equivalent to Clifford conjugation in Cl(3, 0)+ and Cl(0, 2)) is defined as q = qr − qi i − qj j − qk k,
pq = q p,
(2.8.5)
which leaves the scalar part qr unchanged. This leads to the norm of q ∈ H q =
√
qq =
q
qr2 + qi2 + qj2 + qk2 ,
pq = p q .
(2.8.6)
The part 1 (2.8.7) q = V (q) = q − qr = (q − q) = qi i + qj j + qk k, 2 is called a pure quaternion or vector part, it squares to the negative number −(qi2 + qj2 + qk2 ). Every unit quaternion ∈ S 3 (i.e. q = 1) can be written as: q = qr + qi i + qj j + qk k = qr +
q
b = cos α + q b sin α qi2 + qj2 + qk2 q
b ), = exp(α q
(2.8.8)
where cos α = qr , b= q
sin α =
q
qi2 + qj2 + qk2 ,
qi i + qj j + qk k q = q , q qi2 + qj2 + qk2
(2.8.9)
and b 2 = −1, q
b ∈ S2. q
(2.8.10)
The left and right inverse of a nonzero quaternion is q −1 = q/ q2 = q/(qq).
(2.8.11)
The real scalar part (grade zero selection [161] in Clifford geometric algebra) is 1 S(q) = hqi0 = qr = (q + q), 2
(2.8.12)
with symmetries ∀p, q ∈ H: S(pq) = S(qp) = pr qr − pi qi − pj qj − pk qk ,
S(q) = S(q) ,
(2.8.13)
leads to a cyclic multiplication symmetry S(pqs) = S(spq) = S(qsp) ,
∀q, r, s ∈ H.
(2.8.14)
∀p, q ∈ H, α, β ∈ R.
(2.8.15)
We further have linearity S(αp + βq) = α S(p) + β S(q) = αpr + βqr ,
Clifford algebra 95
The scalar part and the quaternion conjugate allow the definition of the R4 inner product of two quaternions p, q as p · q = S(pq) = pr qr + pi qi + pj qj + pk qk ∈ R.
(2.8.16)
Accordingly, we interpret in this section the four quaternion coefficients as coordinates in R4 . In this interpretation selecting any twodimensional plane subspace44 and its orthogonal complement twodimensional subspace allows to split fourdimensional quaternions H into pairs of orthogonal twodimensional planes (compare Theorem 3.5 of [205]). Dealing with rotations we include general rotations in R4 . Definition 2.8.1 (Orthogonality of quaternions). Two quaternions p, q ∈ H are orthogonal p ⊥ q, if and only if S(pq) = 0. 2.8.2 Motivation for quaternion split 2.8.2.1 Splitting quaternions and knowing what it means
We deal with a split of quaternions, motivated by the consistent appearance of two terms in the quaternion Fourier transform [184] F{f }(u, v) =
Z
e−ixu f (x, y)e−jyv dxdy.
(2.8.17)
R2
This observation45 (note that in the following always i is on the left, and j is on the right) and that every quaternion can be rewritten as q = qr + qi i + qj j + qk k = qr + qi i + qj j + qk ij,
(2.8.18)
motivated the quaternion split 46 with respect to the pair of orthonormal pure unit quaternions i, j q = q+ + q− ,
1 q± = (q ± iqj). 2
(2.8.19)
Using (2.8.1), the detailed results of this split can be expanded in terms of real components qr , qi , qj , qk ∈ R, as q± = {qr ± qk + i(qi ∓ qj )}
1±k 1±k = {qr ± qk + j(qj ∓ qi )}. 2 2
(2.8.20)
The analysis of these two components leads to the following Pythagorean modulus identity [187]. 44 The notion of twodimensional plane employed here is thus different from a twodimensional plane in R3 . The latter can be characterized by a unit bivector area element of the plane, which corresponds via the isomorphism Cl(3, 0)+ ∼ = H to a pure unit quaternion. This difference in interpretation means also that despite of the isomorphism Cl(3, 0)+ ∼ = H, the notion and expression of rotations in R4 cannot be automatically carried over to rotation operators of Cl(3, 0). Only in the case when rotations are restricted to the threedimensional subset of pure quaternions, then Hamilton’s original R3 interpretation of these rotations is obvious. 45 Replacing e.g. i → j , j → k throughout would merely change the notation, but not the fundamental observation. 46 Also called OPS as explained below.
96 Quaternion and Clifford Fourier Transforms
Lemma 2.8.2 (Modulus identity). For q ∈ H, q2 = q− 2 + q+ 2 . Lemma 2.8.3 (Orthogonality of OPS split parts [187]). Given any two quaternions p, q ∈ H and applying the OPS split of (2.8.19) the resulting two parts are orthogonal, i.e. p+ ⊥ q− and p− ⊥ q+ , S(p+ q− ) = 0,
S(p− q+ ) = 0.
(2.8.21)
Next, we discuss the map i( )j, which will lead to an adapted orthogonal basis of H. We observe, that iqj = q+ − q− , i.e. under the map i( )j the q+ part is invariant, but the q− part changes sign. Both parts are twodimensional (2.8.20), and by Lemma 2.8.3 they span two completely orthogonal planes, therefore also the name OPS. The q+ plane has the orthogonal quaternion basis {i − j = i(1 + ij), 1 + ij = 1 + k}, and the q− plane has orthogonal basis {i + j = i(1 − ij), 1 − ij = 1 − k}. All four basis quaternions (if normed: {q1 , q2 , q3 , q4 }) {i − j, 1 + ij, i + j, 1 − ij},
(2.8.22)
form an orthogonal basis of H interpreted as R4 . Moreover, we obtain the following geometric picture on the left side of Figure 2.6. The map i()j rotates the q− plane by 180◦ around the twodimensional q+ axis plane. This interpretation of the map i()j is in perfect agreement with Coxeter’s notion of halfturn [84]. In agreement with its geometric interpretation, the map i( )j is an involution, because applying it twice leads to identity i(iqj)j = i2 qj 2 = (−1)2 q = q. (2.8.23) We have the important exponential factor identity eαi q± eβj = q± e(β∓α)j = e(α∓β)i q± .
(2.8.24)
This equation should be compared with the kernel construction of the quaternion Fourier transform (QFT). The equation is also often used in our present context for values α = π/2 or β = π/2. Finally, we note the interpretation [205] of the QFT integrand e−ix1 ω1 h(x) e−jx2 ω2 as a local rotation by phase angle −(x1 ω1 + x2 ω2 ) of h− (x) in the twodimensional q− plane, spanned by {i + j, 1 − ij}, and a local rotation by phase angle −(x1 ω1 −x2 ω2 ) of h+ (x) in the twodimensional q+ plane, spanned by {i − j, 1+ ij}. This concludes the geometric picture of the OPS of H (interpreted as R4 ) with respect to two orthonormal pure quaternion units. 2.8.2.2 Even one pure unit quaternion can do a nice split
Let us now analyze the involution i( )i. The map i( )i gives iqi = i(qr + qi i + qj j + qk k)i = −qr − qi i + qj j + qk k.
(2.8.25)
The following OPS with respect to the single quaternion unit i gives 1 q± = (q ± iqi), 2
q+ = qj j + qk k = (qj + qk i)j,
q− = qr + qi i,
(2.8.26)
Clifford algebra 97
where the q+ plane is twodimensional and manifestly orthogonal to the twodimensional q− plane. The basis of the two planes are (if normed: {q1 , q2 }, {q3 , q4 }) q+ basis: {j, k},
q− basis: {1, i}.
(2.8.27)
The geometric interpretation of i( )i as Coxeter halfturn is perfectly analogous to the case i( )j. This form (2.8.26) of the OPS is identical to the quaternionic simplex/perplex split applied in quaternionic signal processing, which leads in color image processing to the luminosity/chrominance split [127]. 2.8.3 General orthogonal twodimensional planes split (OPS)
Assume in the following an arbitrary pair of pure unit quaternions f, g, f 2 = g 2 = −1. The orthogonal 2D planes split (OPS) is then defined with respect to any two pure unit quaternions f, g as 1 q± = (q ± f qg) 2
f qg = q+ − q− ,
=⇒
(2.8.28)
i.e. under the map f ()g the q+ part is invariant, but the q− part changes sign. Both parts are twodimensional, and span two completely orthogonal planes. For f = 6 ±g the q+ plane is spanned by two orthogonal quaternions {f − g, 1 + f g = −f (f −g)}, the q− plane is, e.g., spanned by {f +g, 1−f g = −f (f +g)}. For g = f a fully orthonormal fourdimensional basis of H is (R acts as rotation operator (rotor)) {1, f, j 0 , k0 } = R−1 {1, i, j, k}R,
R = i(i + f ),
(2.8.29)
and the two orthogonal twodimensional planes basis: q+ basis: {j 0 , k0 },
q− basis: {1, f }.
(2.8.30)
Note the notation for normed vectors in [261] {q1 , q2 , q3 , q4 } for the resulting total orthonormal basis of H. Lemma 2.8.4 (Orthogonality of two OPS planes). Given any two quaternions q, p and applying the OPS with respect to any two pure unit quaternions f, g we get zero for the scalar part of the mixed products Sc(p+ qe− ) = 0,
Sc(p− qe+ ) = 0.
(2.8.31)
Note, that the two parts x± can be represented as x± = x+f
1 ± fg 1 ∓ fg 1 ± fg 1 ∓ fg + x−f = x+g + x−g , 2 2 2 2
(2.8.32)
with commuting and anticommuting parts x±f f = ±f x±f , etc. Next we mention the possibility to perform a split along any given set of two (twodimensional) analysis planes. It has been found, that any twodimensional plane in R4 determines in an elementary way an OPS split and vice versa, compare Theorem 3.5 of [205].
98 Quaternion and Clifford Fourier Transforms
Geometric pictures of the involutions i( )j and f ()g as half turns. Source: [212, Fig. 1].
Figure 2.6
Let us turn to the geometric interpretation of the map f ()g. It rotates the q− plane by 180◦ around the q+ axis plane. This is in perfect agreement with Coxeter’s notion of halfturn [84], see the right side of Figure 2.6. The following identities hold eαf q± eβg = q± e(β∓α)g = e(α∓β)f q± .
(2.8.33)
This leads to a straightforward geometric interpretation of the integrands of the quaternion Fourier transform (OPSQFT) with two pure quaternions f, g, and of the orthogonal 2D planes phase rotation Fourier transform [205]. We can further incorporate quaternion conjugation, which consequently provides a geometric interpretation of the QFT involving quaternion conjugation of the signal function. For d = eαg , t = eβf the map d (f) t represents a rotaryreflection in four dimensions with pointwise invariant line d + t, a rotaryreflection axis d − t: e in the plane ⊥ ^ d (d − t) t = −(d − t), and rotation angle Γ = π − arccos S dt {d + t, d − t}. (The derivation of Γ will be shown later.) We obtain the following Lemma. Lemma 2.8.5. For OPS q± = 12 (q ± f qg), and left and right exponential factors we have the identity eαg q± eβf = q± e(β∓α)f = e(α∓β)g q± . (2.8.34) 2.8.4 Coxeter on quaternions and reflections
The fourdimensional angle Θ between two unit quaternions p, q ∈ H, p = q = 1, is defined by cos Θ = Sc(pqe). (2.8.35) The right and left Clifford translations are defined by Coxeter [84] as q → q 0 = qa,
q → q 00 = aq,
a = ebaΘ ,
b2 = −1. a
(2.8.36)
Both Clifford translations represent turns by constant angles Θq,q0 = Θq,q00 = Θ. We b: analyze the following special cases, assuming the split q± w.r.t. f = g = a
Clifford algebra 99
b = i, aq− = q− a = (q− a)− , is a mathematically positive (anti clockwise) • For a rotation in the q− plane {1, i}.
• Similarly, aq+ = (aq+ )+ , is a mathematically positive rotation in the q+ plane {j, k}. eq+ = (q+ a)+ is a mathematically negative rotation (clockwise) • Finally, q+ a = a by Θ in the q+ plane {j, k}. b. For Next, we compose Clifford translations, assuming the split q± w.r.t. f = g = a b aΘ b2 the unit quaternion a = e , a = −1 we find that
q → aqa = a2 q− + q+
(2.8.37)
is a rotation only in the q− plane by the angle 2Θ, and q → aq˜ a = q− + a2 q+
(2.8.38)
is a rotation only in the q+ plane by the angle 2Θ. Let us now revisit Coxeter’s Lemma 2.2 in [84]: For any two quaternions a, b ∈ H, a = b, ar = br , we can find a y ∈ H such that ay = yb.
(2.8.39)
We now further ask for the set of all y ∈ H such that ay = yb? Based on the OPS, ˆ the answer is straightforward. For a = aeΘˆa , b = aeΦb , Φ = ±Θ, a ˆ2 = ˆb2 = −1 we 1 use the split q± = 2 (q ± a ˆqˆb) to obtain: • For Θ = Φ: The set of all y spans the q− plane. Moreover, aq+ b = q+ ,
q+ b = a ˜q+ ,
aq+ = q+˜b,
aq− b = a2 q− = q− b2 , aq− = q− b.
(2.8.40)
• For Θ = −Φ: The set of all y spans the q+ plane. Moreover, aq− b = q− ,
q− b = a ˜q− ,
aq− = q−˜b,
aq+ b = a2 q+ = q+ b2 , aq+ = q+ b.
(2.8.41)
Let us turn to a reflection in a hyperplane. Theorem 5.1 in [84] says: The reflection in the hyperplane ⊥ a ∈ H: Sc(a˜ q ) = 0, a2 = 1, a = aeΘˆa , a ˆ2 = −1, is represented by q → −a˜ q a. (2.8.42) bq a b) to We analyze the situation using the OPS. We define the split q± = 12 (q ± a obtain ea = −a. q+ → −aqf a → −aa (2.8.43) + a = q+ , π
and for a0 = ae− 2 aˆ a0 → −aae0 a = a0 .
(2.8.44)
100 Quaternion and Clifford Fourier Transforms
We further consider a general rotation. Theorem 5.2 in [84] states: The general ˆ rotation through 2Φ (about a plane) is q → aqb, a = aeΦˆa , b = aeΘb , Φ = ±Θ, a ˆ2 = ˆb2 = −1. We again apply the OPS. We define the split q± = 12 (q ± a ˆqˆb) to obtain: • For Θ = Φ: Rotation of q− plane by 2Φ around q+ plane. • For Θ = −Φ: Rotation of q+ plane by 2Φ around q− plane. Let us illustrate this with an example: a ˆ = ˆb = i, Φ = −Θ, aq− b = q− .
(2.8.45)
aq+ b = ajb = jb2 = je−2θi = j cos 2Φ − k sin 2Φ,
(2.8.46)
For q+ = j : a rotation in the q+ plane around the q− plane. Note, that the detailed analysis of general q → aqb, q± = 21 (q ± a ˆqˆb), a = b = 1, can be found in [205]. As for the rotary inversion, we follow the discussion in [205], sec. 5.1, but add a simple formula for determining the rotation angle. The rotary inversion is given by, d, t ∈ H, d = t = 1, q → dqet. For d 6= ±t, [d, t] = dt − td, we obtain two vectors in ˜ with dvg ˜ the rotation plane v1,2 = [d, t](1 ± dt), 1,2 t = −v1,2 dt. The angle Γ of rotation can therefore be simply found from cos Γ = Sc(
1 ˜ = cos(π − γ), ve1 dve1 t) = Sc(−dt) v1 2
(2.8.47)
˜ with γ the angle between d and t : cos γ = dt. 2.8.5 Quaternion geometry of rotations analyzed by 2D OPS
According to [261] the circle C(q1 , q2 ) of all unit quaternions,47 which rotate g → f f 6= ±g is given by t 1 − fg t g e 2 = q1 e 2 g , t ∈ [0, 2π), 1 − f g g = f, q = f + g . q(t)g q(t) 2 f + g
q(t) =
(2.8.48)
f,g The twodimensional OPS q± = 21 (q ± f qg) tells us, that all q(t), t ∈ [ 0, 2π ) are elements of the q− plane. And indeed, f q− g = −q− for all q− ∈ H leads to −1 f = q− gq− , 47
(2.8.49)
Note that apart from the shortest (and longest) rotation(s) along a geodesic connecting two end points of two unit vectors attached to the center of a unit sphere, any circle on the unit sphere, which includes the end points gives another trajectory of (two) rotation(s) between the two endpoints with rotation axis through the center of the unit sphere and the center of the circle on the unit sphere.
Clifford algebra 101
for all q− in the q− plane. Note, that this is valid for all f, g ∈ H, f 2 = g 2 = −1, even for f = ±g ! We therefore get a one line proof, which at the same time generalizes from the unit circle to the whole plane. Meister and Schaeben [261] state that for q ∈ C(q1 , q2 ): f q, qg, f qg ∈ C(q1 , q2 ). This can easily be generalized to the whole q− plane, because (f q− )− = f q− ,
(q− g)− = q− g,
(f q− g)− = f q− g.
(2.8.50)
We can use the exponential form, and show that the circle C(q1 , q2 ) parametrization of (34), (35) in [261] is a specialization of the general relation t
t
e 2 f q− = q− e 2 g
(2.8.51)
which means that the two parametrizations are element wise identical. Now we look at the quaternion circles for the rotations g → ±f . Prop. 5 of [261] states: Two circles C(q1 , q2 ) = G(g, f ) and C(q3 , q4 ) = G(g, −f ) = G(−g, f ), representing all rotations g → f and g → −f , respectively, are orthonormal to each other. Here four orthogonal unit quaternions are defined as: q1 =
1 − fg f +g 1 + fg f −g , q2 = , q3 = , q4 = . 1 − f g f + g 1 + f g f − g
(2.8.52)
We provide a simple proof: We already know that all q1 , q2 span the q− plane of the f,g split q± = 12 (q ± f qg), and all q3 , q4 span the q+ plane. And that f q± g = ±q±
⇐⇒
−1 f = q± (∓g)q± ,
QED.
(2.8.53)
Note, the proof is again much faster than in [261]. We see that G(g, f ) = f,g f,g f,g f,g {q− /q− , ∀q ∈ H}, and G(g, −f ) = {q+ /q+ , ∀q ∈ H}. For later use, we translate the notation of [261] (38),(39): n3 = −n1 =
[f, g] , [f, g]
n4 = q4 ,
n2 = n4 n1 ,
n4 = n1 n2 ,
n1 = n2 n4 ,
(2.8.54)
which shows that {n1 , n2 , n4 } is a right handed set of three orthonormal pure quaternions, obtained by rotating {i, j, k}. The two circles G(g, f ), G(g 0 , f ) do not intersect for g 6= g 0 , see Cor. 1(i) of [261]. We provide a simple proof: Assume ∃1 q ∈ H : f qg = −q, f qg 0 = −q for g 6= g 0 . Then f qg = f qg 0
⇔
g = g0
⇒
G(g, f )
\
G(g 0 , f ) = ∅.
(2.8.55)
QED. Cor. 1 (iii) of [261] further states that for every 3D rotation R and given g0 , g02 = −1 we can always find f , f 2 = −1, such that R is represented by a (unit) quaternion q in G(g0 , f ). We can equivalently ask for f , such that q representing the f,g0 rotation R is ∈ q− plane. We find f qg0 = −q
⇔
f = qg0 q −1 .
(2.8.56)
102 Quaternion and Clifford Fourier Transforms
Small circles and tangential plane projection. Adapted from Figs. 2 and 3 of [261]. Source: [212, Fig. 2].
Figure 2.7
The left side of Figure 2.7 shows two small circles C(g, ρ), C(f, ρ) ⊂ S 2 [261]. We now analyze the mapping between pairs of small circles. A small circle with center g f,g and radius ρ is defined as C(g, ρ) = {g 0 ∈ §2 : g · g 0 = cos ρ}, and all q ∈ q− plane map C(g, ρ) to the small circle C(f, ρ) of the same radius (a slight generalization of [261], Prop. 6), with the correspondence q(t)g 0 (u)q(t)−1 = f 0 (u + 2t), q(t) = q1 etg ,
u
u
g 0 (u) = e 2 g g00 e− 2 g ,
u
u
f 0 (u) = e 2 f f00 e− 2 f
(2.8.57)
starting with the corresponding circle points q1 g00 = f00 q1 . We provide the following direct proof: We repeatedly apply (2.8.33) to obtain u
u
u
u
u
u
q1 g00 = f00 q1 ⇔ e 2 f q1 g00 e− 2 g = e 2 f f00 q1 e− 2 g u
u
⇔ q1 e 2 g g00 e− 2 g = e 2 f f00 e− 2 f q1 u
u
u
u
u
u
u
u
⇔ etf q1 e 2 g g00 e− 2 g = etf e 2 f f00 e− 2 f e−tf etf q1 ⇔ q1 etg e 2 g g00 e− 2 g = etf e 2 f f00 e− 2 f e−tf q1 etg .
(2.8.58)
QED. Note, that this proof is much shorter than in [261], and we do not need to use addition theorems. We consider the projection onto the tangential plane of a S 2 vector, see the right side of Figure 2.7. Assume v, r ∈ S 2 . Note, that (v)T (r) = v − (v · r)r of [261] can be simplified to (v)T (r) = V (vr)r−1 , valid for all pure (nonunit) quaternions r. Finally we consider the torus theorem for all maps g → small circle C(f, 2Θ). We slightly reformulate the theorem Prop. 13 of [261]. We will use the twodimensional OPS with respect to f, g ∈ S 2 , and the corresponding orthonormal basis {q1 , q2 , q3 , q4 } of (2.8.52). The theorem says, that the circle C(q1 , q2 ) ∈ q− plane: q− (s) = q1 exp(sg/2), s ∈ [0, 2π), represents all rotations g → f , while the orthogonal circle C(q3 , q4 ) ∈ q+ plane: q+ (t) = q3 exp(−tg/2), t ∈ [0, 2π), represents
Clifford algebra 103
all rotations g → −f . Then the spherical torus T (q− (s), q+ (t); Θ) is defined as the set of quaternions s, t ∈ [0, 2π),
q(s, t; Θ) = q− (2s) cos Θ + q+ (2t) sin Θ,
Θ ∈ [0, π/2],
(2.8.59)
and represents all rotations g → C(f, 2Θ) ⊂ S 2 . In particular, the set q(s, −s; Θ) maps g for all s ∈ [0, 2π) onto f00 in the f, g plane with g · f00 = cos(η − 2Θ), g · f = cos η, q(s, −s; Θ)gq(s, −s; Θ)−1 = f00
∀s ∈ [0, 2π).
(2.8.60)
Moreover, for arbitrary s0 ∈ [0, 2π), the set q(s0 , t − s0 ; Θ) (or equivalently q(s0 + t, s0 ; Θ)) maps g → f 0 ∈ C(f, 2Θ), which results from positive rotation (counterclockwise) of f00 about f by the angle t ∈ [0, 2π), t
t
q(s0 , t − s0 ; Θ)gq(s0 , t − s0 ; Θ)−1 = e 2 f f00 e− 2 f
∀s0 ∈ [0, 2π).
(2.8.61)
We state the following direct proof of the torus theorem. q(s, t; Θ) = q− (2s) cos Θ + q+ (2t) sin Θ = q1 esg cos Θ + q3 e−tg sin Θ = (q1 cos Θ + q3 e−(t+s)g sin Θ)esg = (cos Θ + (q3 /q1 )e−(t+s)f sin Θ)q1 esg = (cos Θ + (−n1 )e−(t+s)f sin Θ)q1 esg = (cos Θ + e(t+s)f (−n1 ) sin Θ)esf q1 0
= e−n1 Θ esf q1 ,
n01 = e(t+s)f (−n1 ),
(n01 )2 = −1.
(2.8.62)
We observe, that n01 is n1 rotated around f by angle s + t. Application to g gives 0
0
q(s, t; Θ)g q(s, t; Θ)−1 = e−n1 Θ f en1 Θ ,
(2.8.63)
so geometrically g is rotated into f = esf q1 g q1−1 e−sf , which in turn is rotated around n01 on the circle C(f, 2Θ). For t = −s obviously q(s, −s; Θ)gq(s, −s; Θ)−1 = e−n1 Θ f en1 Θ = f00 ,
(2.8.64)
is a rotation in the f, g plane of g into f00 , with g · f00 = cos η − 2Θ. We further note, that for s = s0 , t → t − s0 : n01 = etf (−n1 ), such that 0
0
q(s0 , t − s0 ; Θ)gq(s0 , t − s0 ; Θ)−1 = e−n1 Θ f en1 Θ = etf f00 e−tf ,
(2.8.65)
describes the small circle C(f, 2Θ). QED. Our proof is very compact, obtained by direct computation of monomial results, which in turn permit direct geometric interpretation. 2.8.6 Summary on quaternions and rotation geometry
We have exposed the geometric understanding of the general OPS split of quaternions [205] into two orthogonal planes (R4 interpretation). Moreover, we have consolidated the OPS with the geometric understanding by Altmann [12, 322], Coxeter [84], and Meister and Schaeben [261].
CHAPTER
3
Geometric calculus
Now faith is being sure of what we hope for and certain of what we do not see. This is what the ancients were commended for. By faith we understand that the universe was formed at God’s command, so that what is seen was not made out of what was visible. [157] The German 19th century mathematician H. Grassmann had the clear vision that his . . . extension theory (now developed to geometric calculus) . . . forms the keystone of the entire structure of mathematics. [146]
We introduce geometric vector differential calculus. This chapter is mainly based on [174]. We treat the fundamentals of the vector differential calculus part of universal geometric calculus. Geometric calculus is seen to simplify and unify the structure and notation of mathematics for all of science and engineering, and for technological applications. In order to make the treatment reasonably selfcontained, we also note the most important geometric algebra relationships, which are necessary for vector differential calculus. Readers, who need more than this summary are advised to turn back to Chapter 2, or for a comprehensive less technical introduction to Clifford’s geometric algebras we refer to [200]. We further recommend the two excellent college level textbooks [248, 249]. After that differentiation by vectors is introduced and a host of major vector differential and vector derivative relationships is proven explicitly in a very elementary step by step approach. The chapter is thus intended to serve as reference material, giving details, which are usually skipped in more advanced discussions of the subject matter.
DOI: 10.1201/97810031844783
105
106 Quaternion and Clifford Fourier Transforms
3.1
INTRODUCTORY NOTES ON VECTOR DIFFERENTIAL CALCULUS
The algebraic grammar of the universal form of (geometric) calculus, as envisioned by H. Grassmman, is geometric algebra (or Clifford algebra). A brief overview of vector differential calculus is given in Section 3.2. Next, the basic geometric algebra necessary for this is summarized for notation and reference in Section 3.3. Finally, Section 3.4 develops vector differential calculus with the help of few simple definitions. This approach is generically coordinate free, and fully shows both the concrete and abstract geometric and algebraic beauty of Grassmann’s “keystone” of mathematics. Applications of geometric algebra are surveyed in [202]. We demonstrate the proofs for all common formulas of vector differential calculus in an elementary step by step fashion. Where other texts (e.g. [162],[161]) tend both to skip “elementary steps,” and to presume, that the reader would be smart enough to fill in the gaps himself. I put the emphasis therefore on thorough proofs and not on comments, interpretations or application. Following the same approach as taken here, the step from vector to multivector differential calculus is taken in [175].
3.2
BRIEF OVERVIEW OF VECTOR DIFFERENTIAL CALCULUS
We start with a brief overview without the proofs (that are given in the following sections). Note that we use in the current section the principal reverse of Section 2.1.1.6. Multivector valued functions f : Rp,q → Clp,q , p + q = n, have 2n blade components (fA : Rp,q → R) X f (x) = fA (x)e A . (3.2.1) A
We define the inner product of two R Z
(f, g) =
p,q
g dn x = f (x)g(x)
Rp,q
→ Clp,q functions f, g by
X
Z
eA efB
A,B
Rp,q
fA (x)gB (x) dn x.
(3.2.2)
In (3.2.2), the inner product ( , )L2 (Rp,q ;Clp,q ) satisfies the following conditions[33] (f, g + h)L2 (Rp,q ;Clp,q ) = (f, g)L2 (Rp,q ;Clp,q ) + (f, h)L2 (Rp,q ;Clp,q ) , ˜ (f, λg)L2 (Rp,q ;Cl ) = (f, g)L2 (Rp,q ;Cl ) λ, p,q
(f λ, g)L2 (Rp,q ;Clp,q )
p,q
˜ L2 (Rp,q ;Cl ) , = (f, g λ) p,q
] (f, g)L2 (Rp,q ;Clp,q ) = (g, f )L2 (Rp,q ;Clp,q ) .
(3.2.3)
where f, g ∈ L2 (Rp,q ; Clp,q ), and the constant multivector λ ∈ Clp,q . The symmetric scalar part is hf, gi =
Z Rp,q
g dn x = f (x) ∗ g(x)
XZ A
Rp,q
fA (x)gA (x) dn x,
(3.2.4)
Geometric calculus 107
and the norm for functions in L2 (Rn ; Cln,0 ) = {f : Rn → Cln,0  kf k < ∞} as 2
kf k = h(f, f )i =
Z
f (x)2 dn x.
Rn
Z
(5.2.148) f (x) ∗ f˜(x)d3 x =
=
Z
Rn
X
Rn
fA2 (x)d3 x.
(3.2.5)
A
Example 3.2.1. For g = af, f, g ∈ L2 (R3 ; Cl3,0 ), a ∈ R3 we get because of f i0 = haf f˜ai0 = ha2 f f˜i0 = a2 f ∗ f˜ haf af kaf k2L2 (R3 ;Cl3,0 )
Z
=
a f (x) ∗ f˜(x)d3 x = 2
Z
R3
a2
R3
X
fA2 (x)d3 x.
(3.2.6)
A
Definition 3.2.2 (Clifford module). Let Clp,q be the real Clifford algebra of the quadratic space Rp,q . A Clifford algebra module L2 (Rp,q ; Clp,q ) is defined by L2 (Rp,q ; Clp,q ) = {f : Rp,q −→ Clp,q  kf kL2 (Rp,q ;Clp,q ) < ∞}.
(3.2.7)
e of a An alternative definition of Clifford module uses the principal reverse N e = multivector N ∈ Cl(p, q) of Section 2.1.1.6. For M, N ∈ Cl(p, q) we get M ∗ N P e A MA NA . Two multivectors M, N ∈ Cl(p, q) are orthogonal if and only if M ∗N = 0. The modulus M  of a multivector M ∈ Cl(p, q) is defined as f= M 2 = M ∗ M
X
MA2 .
(3.2.8)
A
The possibility to differentiate with respect to any multivector (representing a geometric object or a transformation operator) is essential for solving optimization problems in GA. We can then define the vector differential [161,174] of f for any constant a ∈ Rp,q as f (x + a) − f (x) , (3.2.9) a · ∇f (x) = lim →0 where a · ∇ is scalar. The vector derivative ∇ can be expanded in terms of the basis vectors e k as ∇ = ∇x =
n X
e k ∂k ,
k=1
∂k = e k · ∇ =
∂ . ∂xk
(3.2.10)
Both a · ∇ and ∇ are coordinate independent. Replacing the vectors a, x by multivectors gives the multivector derivative [161, 175]. Examples. The five multivector functions f1 = x,
f2 = x 2 ,
f5 = log r,
r = x − x 0,
f3 = x, r = r,
f4 = x · hAik ,
0 ≤ k ≤ n, (3.2.11)
108 Quaternion and Clifford Fourier Transforms
have the following vector differentials [174] a · ∇f1 = a, a · ∇f2 = 2a · x, a · ∇f3 = a · ∇f4 = a · hAik ,
a · ∇f5 =
a·x , x
a·r . r2
(3.2.12)
This leads to the vector derivatives [174] ∇f1 = ∇ · x = 3, (n = 3), ∇f2 = 2x, ∇f3 = x/x, ∇f4 = khAik ,
∇f5 = r −1 = r/r2 .
(3.2.13)
We can compute the derivative from the differential for ∇a : regard x = constant, a = variable, and compute ∇f = ∇a (a · ∇f ) (3.2.14) The vector derivative obeys sum rules and product rulesixvector derivative!product rule (overdots indicate functions to be differentiated). But noncommutativity leads ˙ g˙ = ˙ g. to modifications, because ∇f 6 f∇ ˙ The product rule is ˙ f˙)g + ∇f ˙ g˙ = (∇ ˙ f˙)g + ∇(f g) = (∇
n X
e k f (∂k g).
(3.2.15)
k=1
The chain rules for the vector differential and the vector derivative of f (x) = g(λ(x)), λ(x) ∈ R, are a · ∇f = {a · ∇λ(x)}
∂g , ∂λ
∇f = (∇λ)
∂g . ∂λ
(3.2.16)
Example: For a = e k (1 ≤ k ≤ n) we have e k · ∇f = ∂k f = (∂k λ)
∂g . ∂λ
(3.2.17)
GA thus provides a new formalism for differentiation on vector manifolds and for mappings between surfaces, including conformal mapping [304].
3.3
GEOMETRIC ALGEBRA FOR DIFFERENTIAL CALCULUS
This section is a basic summary (without proofs) of important relationships in geometric algebra. For further details we refer to Chapter 2. This summary mainly serves as a reference section (including notation) for the vector differential calculus to be developed. Most of the relationships below can also be found in the synopsis of geometric algebra and in chapters 1 and 2 of [162], as well as in chapter 1 of [161], together with relevant proofs. Beyond that [162] and [161] follow a much more didactic approach for complete newcomers to geometric algebra. G(I) is the full geometric algebra over all vectors in the ndimensional unit pseudoscalar I = ~e1 ∧ ~e2 ∧ . . . ∧ ~en . An ≡ G 1 (I) is the ndimensional vector subspace of grade1 elements in G(I) spanned by ~e1 , ~e2 , . . . , ~en . For vectors ~a, ~b, ~c ∈ An ≡ G 1 (I) and scalars α, β, λ, τ ; G(I) has the fundamental properties of
Geometric calculus 109
• Associativity ~a(~b~c) = ~a(~b~c),
~a + ~b + ~c = (~a + ~b) + ~c,
(3.3.1)
• Commutativity α~a = ~aα,
~a + ~b = ~b + ~a,
(3.3.2)
• Distributivity ~a(~b + ~c) = ~a~b + ~a~c,
(~b + ~c)~a = ~b~a + ~c~a,
(3.3.3)
• Linearity α(~a + ~b) = α~a + α~b = (~a + ~b)α,
(3.3.4)
• Scalar square (vector length ~a) ~a2 = ~a~a = ~a · ~a = ~a2 .
(3.3.5)
The geometric product ~a~b is related to the (scalar) inner product ~a · ~b and to the (bivector or 2vector) outer product ~a ∧ ~b by ~a~b = ~a · ~b + ~a ∧ ~b,
(3.3.6)
with 1 (3.3.13) ~a · ~b = (~a~b + ~b~a) = ~a~b − ~a ∧ ~b = ~b · ~a = h~a~bi0 , 2 1 (3.3.13) ~a ∧ ~b = (~a~b − ~b~a) = −~b ∧ ~a = ~a~b − ~a · ~b = h~a~bi2 . 2
(3.3.7) (3.3.8)
The inner and the outer product are both linear and distributive ~a · (α~b + β~c) = α~a · ~b + β~a · ~c, ~a ∧ (α~b + β~c) = α~a ∧ ~b + β~a ∧ ~c.
(3.3.9) (3.3.10)
A unit vector a ˆ in the direction of ~a is a ˆ=
~a ~a
with
The inverse of a vector is ~a−1 =
a ˆ2 = a ˆa ˆ = 1,
~a = a ˆ~a.
1 ~a a ˆ = 2 = . ~a ~a ~a
(3.3.11)
(3.3.12)
A multivector A can be uniquely decomposed into its homogeneous grade k parts (h ik grade k selector): A = hAi0 + hAi1 + hAi2 +... + hAik +... +  {z } scalar
 {z }
vector
 {z }
bivector
 {z }
kvector
hAin  {z }
pseudoscalar
(3.3.13)
110 Quaternion and Clifford Fourier Transforms
If A is homogeneous of grade k one often simply writes A = hAik = Ak .
(3.3.14)
Grade selection is invariant under scalar multiplication λ hAik = hλAik .
(3.3.15)
The consistent definition of inner and outer products of vectors ~a and rvectors Ar is 1 ~aAr − (−1)r Ar~a 2 1 = ~aAr + (−1)r Ar~a . 2
~a · Ar ≡ h~aAr ir−1 = ~a ∧ Ar ≡ h~aAr ir+1
(3.3.16) (3.3.17)
By linearity the full geometric product of a vector and a multivector A is then ~aA = ~a · A + ~a ∧ A.
(3.3.18)
This extends to the distributive multiplication with arbitrary multivectors A and B ~a(A + B) = ~aA + ~aB.
(3.3.19)
The inner and outer products of homogeneous multivectors Ar and Bs are defined ([161], p. 5, (1.21), (1.22)) as Ar · Bs ≡ hAr Bs ir−s for r, s > 0,
(3.3.20)
Ar · Bs ≡ 0 for r = 0 or s = 0,
(3.3.21)
Ar ∧ Bs ≡ hAr Bs ir+s ,
(3.3.22)
Ar ∧ λ = λ ∧ Ar = λAr for scalar λ.
(3.3.23)
The inner (and outer) product is again linear and distributive (λAr ) · Bs = Ar · (λBs ) = λ(Ar · Bs ) = λAr · Bs ,
(3.3.24)
Ar · (Bs + Ct ) = Ar · Bs + Ar · Ct ,
(3.3.25)
λ(Bs + Ct ) = λBs + λCt .
(3.3.26)
The reverse of a multivector is Ae =
n X
(−1)k(k−1)/2 hAik .
(3.3.27)
k=1
[161] uses a dagger instead of the tilde. Special examples are e = λ, λ
e = ~a, ~a
] ~a ∧ ~b = ~b ∧ ~a = −~a ∧ ~b,
(3.3.28)
Geometric calculus 111
The scalar magnitude A of a multivector A is 2
A ≡
e∗A A {z }
=
hAi20
+
n X
e r · hAir , hAi
(3.3.29)
r=1
scalar product
where the separate term hAi20 is in particular due to the definition of the inner product in [161], p. 5, (1.21). The magnitude allows to define the inverse1 for simple kblade vectors Ae , with A−1 A = AA−1 = 1. (3.3.30) A−1 ≡ A2 Alternative ways to express ~a ∈ An ≡ G 1 (I) are I ∧ ~a = 0
or
I~a = I · ~a.
(3.3.31)
The projection of ~a into An ≡ G 1 (I) is PI (~a) = P (~a) ≡
n X
k
~a ~ak · ~a =
k=1
n X
~ak~ak · ~a,
(3.3.32)
1 if j = k, . 0 if j = 6 k,
(3.3.33)
k=1
where ~ak is the reciprocal frame defined by ( k
~a ~aj =
δjk
= Kronecker delta =
A general convention is that inner products ~a · ~b and outer products ~a ∧ ~b have priority over geometric products ~a~b, e.g. ~e = (~a · ~b)(~c ∧ d)~ ~ e. ~a · ~b~c ∧ d~
(3.3.34)
The projection of a multivector B on a subspace described by a simple mvector (mblade) Am = ~a1 ∧ ~a2 ∧ . . . ∧ ~am is s≤m
PA (B) ≡ (B · A) · A−1 = A−1 · (A · B( s)) , 
{z
general
PA (hBi0 ) ≡ hBi0 ,
}

{z
degree dependent
(3.3.35)
}
PA (hBin ) ≡ hBin · AA−1 ,
(3.3.36)
the exceptions for scalars hBi0 and pseudoscalars hBin being again due to the definition of the inner product in [161], p. 6, (1.21). A projection of one factor of an inner product has the effect to project the other factor as well ~a · P (~b) = P (~a) · P (~b) = P (~a) · ~b.
(3.3.37)
For a multivector B ∈ G(Am ), with A = Am we have (~a ∧ B) · A = (~a ∧ B)A = ~a · (BA), 1
if
~a ∧ A = 0.
(3.3.38)
Every multivector can be inverted, as long as its determinant is nonzero, see [216, 300].
112 Quaternion and Clifford Fourier Transforms
Reordering rules for products of homogeneous multivector are Ar · Bs = (−1)r(s−r) Bs · Ar
for
r ≤ s,
Ar ∧ Bs = (−1)rs Bs ∧ Ar .
(3.3.39) (3.3.40)
Elementary combinations that occur often are ~a · (~b ∧ ~c) = (~a · ~b)~c − (~a · ~c)~b = ~a · ~b~c − ~a · ~c~b, (3.3.41) ~ = ~a · (~b · (~c ∧ d) ~ = (~a · d)( ~ ~b · ~c) − (~a · ~c)(~b · d), ~ (~a ∧ ~b) · (~c ∧ d) (3.3.42) 2 2 2 2 (~a ∧ ~b) = (~a ∧ ~b) · (~a ∧ ~b) = (~a · ~b) − ~a ~b = −(~b ∧ ~a) · (~a ∧ ~b) = −~a ∧ ~b2 ,
(3.3.43)
and the cyclic Jacobi identity ~a · (~b ∧ ~c) + ~b · (~c ∧ ~a) + ~c · (~a ∧ ~b) = 0.
(3.3.44)
The commutator product of multivectors A, B is 1 A × B ≡ (AB − BA). 2
(3.3.45)
(~a ∧ ~b) × A = ~a~b · A − A · ~a~b = ~a~b ∧ A − A ∧ ~a~b.
(3.3.46)
One useful identity using it is
The commutator product is to be distinguished from the cross product, which is strictly limited to the threedimensional Euclidean case with unit pseudoscalar I3 : ~a × ~b = (~b ∧ ~a)I3 = −(~a ∧ ~b)I3 = (~a ∧ ~b)I3−1 .
(3.3.47)
For more on basic geometric algebra I refer to [162], [161], to section 3 of [173], and to Chapter 2 in the present book.
3.4
VECTOR DIFFERENTIAL CALCULUS
This section shows how to differentiate functions on linear subspaces of the universal geometric algebra G by vectors. It has wide applications particularly to mechanics and physics in general [162]. Separate concepts of gradient, divergence and curl merge into a single concept of vector derivative, united by the geometric product. The relationship of differential and derivative is clarified. The Taylor expansion (Prop. 116) is applied to important examples, yielding e.g. the Legendre polynomials (Prop. 124). The adjoint (Def. 3.4.57) and the integrability (Prop. 3.4.42, etc.) of multivector functions are defined and discussed. Throughout this section a number of basic differentials and derivations are performed explicitly illustrating ease and power of the calculus. Since my emphasis here is on explicit step by step proofs, I refer the reader, who
Geometric calculus 113
is interested in the philosophy, comments and interpretation to the literature in the final Reference section. As for the notation: Prop. 3.4.7 refers to proposition 3.4.7 of this section. Def. 3.4.13 refers to definition 3.4.13 of this section. (3.3.6) refers to equation number (3.3.6) in the previous section on basic geometric algebra, etc. Standard definitions of continuity and scalar differentiability apply to multivectorvalued functions, because the scalar product determines a unique “distance” A − B between two elements A, B ∈ G(I). Definition 3.4.1. Directional derivative. F = F (~x) multivectorvalued function of a vector variable ~x defined on an ndimensional vector space An = G 1 (I), I unit pseudoscalar, ~a ∈ An : ~ = dF (~x + ~aτ ) = lim F (~x + ~aτ ) − F (~x) . ~a · ∂F (3.4.1) τ →0 dτ τ Nomenclature: derivative of F in the direction ~a, ~aderivative of F . ([162] uses ~ [161] uses ∂ ≡ ∂.) ~ ∇ ≡ ∂, Proposition 3.4.2. Distributivity w.r.t. vector argument. Assume that the directional derivative of F is continuous for all direction vectors of an arbitrary basis of An .” ~ = ~a · ∂F ~ + ~b · ∂F, ~ (~a + ~b) · ∂F ∀ ~a, ~b ∈ An . (3.4.2) Proof. ~ ~ = lim F (~x + τ~a + τ b) − F (~x) (~a + ~b) · ∂F τ →0 τ ( ) F (~x + τ~a + τ~b) − F (~x + τ~b) + F (~x + τ~b) − F (~x) = lim τ →0 τ ~ (~x + τ~b) + ~b · ∂F ~ = ~a · ∂F ~ + ~b · ∂F. ~ = lim ~a · ∂F τ →0
(3.4.3)
Proposition 3.4.3. For scalar λ ~ = λ(~a · ∂F ~ ). (λ~a) · ∂F
(3.4.4)
Proof. ~ (λ~a) · ∂F
Def.3.4.1
=
F (~x + λ~aτ ) − F (~x) . τ →0 τ lim
(3.4.5)
Case 1: λ 6= 0 : F (~x + ~a(λτ )) − F (~x) λτ 0 F (~ x + ~ a τ ) − F (~x) Def.3.4.1 λ6=0, rp. ~ ), = λ lim = λ(~a · ∂F 0 0 τ →0 τ with rp. = reparametrization: τ → τ 0 . Case 2: λ = 0 : F (~x) − F (~x) ~ Def.3.4.1 ~ ). (λ~a) · ∂F = lim = 0 = 0 (~a · ∂F τ →0 τ ~ (λ~a) · ∂F
Def.3.4.1
=
lim λ
τ →0
(3.4.6)
(3.4.7)
114 Quaternion and Clifford Fourier Transforms
Proposition 3.4.4. Distributivity w.r.t. multivectorvalued function. ~ + G) = ~a · ∂F ~ + ~a · ∂G, ~ ~a · ∂(F
(3.4.8)
with G = G(~x), F = F (~x), multivectorvalued functions of a vector variable ~x. In the notation of Def. 3.4.13: F + G = F + G. (3.4.9) Proof. F (~x + ~a(τ )) + G(~x + ~a(τ )) − F (~x) − G(~x) ~ + G) Def.3.4.1 ~a · ∂(F = lim τ →0 τ G(~x + ~a(τ )) − G(~x) F (~x + ~a(τ )) − F (~x) + lim = lim τ →0 τ →0 τ τ Def.3.4.1 ~ ~ = ~a · ∂F + ~a · ∂G. (3.4.10)
Proposition 3.4.5. Product rule. ~ G) = (~a · ∂F ~ )G + F (~a · ∂G). ~ ~a · ∂(F
(3.4.11)
In the notation of Def. 13: F G = F G + F G.
(3.4.12)
Proof. F (~x + ~aτ )G(~x + ~aτ ) − F (~x)G(~x) τ F (~x + ~aτ )G(~x + ~aτ ) − F (~x)G(~x + ~aτ ) + F (~x)G(~x + ~aτ ) − F (~x)G(~x) = lim τ →0 τ G(~x + ~aτ ) − G(~x) F (~x + ~aτ ) − F (~x) G(~x + ~aτ ) + F (~x) = lim τ →0 τ τ Def.3.4.1 ~ ) lim G(~x + ~aτ ) + F (~a · ∂G) ~ = (~a · ∂F
~ G) ~a · ∂(F
Def.3.4.1
=
lim
τ →0
τ →0
~ )G + F (~a · ∂G). ~ = (~a · ∂F
(3.4.13)
Proposition 3.4.6. Grade invariance. ~ ik = h~a · ∂F ~ ik . ~a · ∂hF
(3.4.14)
~a · ∂~ is therefore said to be a scalar differential operator. Proof. ~ ik ~a · ∂hF
hF (~x + ~aτ )ik − hF (~x)ik F (~x + ~aτ ) − F (~x) = lim τ →0 τ →0 τ τ k Def.3.4.1 ~ = h~a · ∂F ik . (3.4.15) Def.3.4.1
=
lim
Geometric calculus 115
Proposition 3.4.7. Scalar chain rule. ~ = (~a · ∂λ) ~ dF ~a · ∂F dλ
(3.4.16)
with F = F (λ(~x)), λ = λ(~x) a scalar valued function. Proof. Using the Taylor expansions: F (λ + τ ∆λ) = F (λ) + τ ∆λ (3.4.28)
λ(~x + τ~a)
=
dF τ 2 (∆λ)2 d2 F + + ..., dλ 2 dλ2
2 ~ x) + τ (~a · ∂) ~ 2 λ(~x) + . . . , λ(~x) + τ~a · ∂λ(~ 2
(3.4.17) (3.4.18)
we have F (λ(~x + ~aτ )) − F (λ(~x)) ~ (λ(~x)) Def.3.4.1 ~a · ∂F = lim τ →0 τ ~ x)) − F (λ(~x)) F (λ(~x) + τ~a · ∂λ(~ (3.4.18) = lim τ →0 n τ o ~ x) dF − F (λ(~x)) F (λ(~ x )) + τ ~ a · ∂λ(~ (3.4.17) dλ = lim τ →0 τ dF ~ = (~a · ∂λ) . dλ
(3.4.19)
Proposition 3.4.8. Identity. ~a · ∂~ ~x = ~a.
(3.4.20)
Proof. F (~x) = ~x,
~x + τ~a − ~x ~ x Def.3.4.1 = lim ~a = ~a. ~a · ∂~ = lim τ →0 τ →0 τ
(3.4.21)
Proposition 3.4.9. Constant function. For A independent of ~x : ~ = 0. ~a · ∂A
(3.4.22)
Proof. F (~x) = A,
A−A ~ Def.3.4.1 ~a · ∂A = lim = 0. τ →0 τ
(3.4.23)
Proposition 3.4.10. Vector length. ~ x = ~a · ~x = ~a · xˆ, ~a · ∂~ ~x with xˆ =
~ x ~ x
unit vector in the direction of ~x, see (3.3.11).
(3.4.24)
116 Quaternion and Clifford Fourier Transforms
Proof. ~ x2 ~a · ∂~
Prop. 3.4.5
~ x)~x + ~x(~a · ∂~ ~ x) Prop.=3.4.8 ~a~x + ~x~a = 2~a · ~x, (~a · ∂~
~ x2 ~a · ∂~
Prop. 3.4.5
~ x)~x + ~x(~a · ∂~ ~ x) = 2~x(~a · ∂~ ~x), (~a · ∂~ ~ x2 = ~a · ∂~ ~x2 , ⇒ ~a · ∂~
=
=
~x2 = ~x2 ⇒
~a · ~x = ~x(~a · ∂~ ~x)
⇒
~a · ~x ~a · ∂~ ~x = = ~a · xˆ. ~x
Proposition 3.4.11. Direction function. xˆxˆ ∧ ~a ~a − ~a · xˆxˆ = . ~a · ∂~ xˆ = ~x ~x
(3.4.25)
(3.4.26)
Proof. ~x ~a · ∂~ 1 + ~x~a · ∂~ ~x ~x d ~x1 Prop. 3.4.10 ~a a −1 Props. 3.4.7, 3.4.8 ~ ~ = + ~x(~a · ∂~x) = + ~x(~a · xˆ) 2 ~x d~x ~x ~x 1 ~a − ~a · xˆxˆ xˆxˆ~a − xˆxˆ · ~a ~a − xˆ(~a · xˆ) = = = ~x ~x ~x ~x xˆ(ˆ x · ~a + xˆ ∧ ~a) − xˆxˆ · ~a xˆxˆ ∧ ~a = = . (3.4.27) ~x ~x
~x ~a · ∂~ xˆ = ~a · ∂~ ~x
Prop. 3.4.5
=
Proposition 3.4.12. Taylor expansion. ~ (~x) = F (~x + ~a) = exp(~a · ∂)F
∞ ~ k X (~a · ∂) k=0
k!
F (~x).
(3.4.28)
Proof. Note that this proof is done without referring to Props. 3.4.7 to 3.4.11! G(τ ) ≡ F (~x + ~aτ ) dG(0) dF (~x + ~aτ ) Def.3.4.1 ~ ⇒ = τ =0 = ~a · ∂F dτ dτ d2 G(0) d dF (~x + ~aτ ) dF (~x + ~aτ ) Def.3.4.1 ⇒ = τ =0 = ~a · ∂~ τ =0 2 dτ dτ dτ dτ Def.3.4.1 ~ (~x) = (~a · ∂) ~ 2 F (~x). = ~a · ∂~ ~a · ∂F (3.4.29) General:
dk G(0) dτ k
~ k F (~x) . The Taylor series for G is: = (~a · ∂)
G(1) = G(0 + 1) = G(0) + ⇒
G(1) = F (~x + ~a) =
∞ X 1 dk G(0) dG(0) 1 d2 G(0) + + . . . = dτ 2 dτ 2 k! dτ k k=0
∞ X 1
k! k=0
~ k F (~x) = exp(~a · ∂)F ~ (~x). (~a · ∂)
(3.4.30)
Geometric calculus 117
Definition 3.4.13. Continuously differentiable, differential. ~ (~y ) exists and is a F is continuously differentiable at ~x, if for each fixed ~a, ~a · ∂F continuous function of ~y for each ~y in a neighborhood of ~x. ~ (~x) is If F is defined and continuously differentiable at ~x, then, for fixed ~x, ~a · ∂F a linear function of ~a, the (first) differential of F . ~ (~x). F (~a, ~x) = F~a (~x) ≡ ~a · ∂F
(3.4.31)
([162], p. 107 uses F 0 ≡ F .) Suppressing ~x, or for fixed ~x : ~ F = F (~a) = F~a ≡ ~a · ∂F.
(3.4.32)
Proposition 3.4.14. Linearity. F (~a + ~b) = F (~a) + F (~b),
λ scalar : F (λ~a) = λF (~a).
(3.4.33)
Proof. Propositions 3.4.2 and 3.4.3. Proposition 3.4.15. Linear approximation. For ~r = ~x − ~x0  sufficiently small: F (~x) − F (~x0 ) ≈ F (~x − ~x0 ) = F (~x) − F (~x0 ).
(3.4.34)
Proof. ~ 2 F (~x0 ) + . . . ~ (~x0 ) + 1 (~r · ∂) F (~x0 ) + ~r · ∂F 2 2 ~ (~x0 ) + ~r (ˆ ~ 2 F (~x0 ) + . . . = F (~x0 ) + ~rˆ r · ∂F r · ∂) 2 ~rk ~ k F (~x0 ) + . . . (ˆ r · ∂) (3.4.35) + k!
F (~x) = F (~x0 + ~r)
with rˆ ≡
~ r ~ r .
Prop. 3.4.12
=
For sufficiently small ~r :
~ (~x0 ) = F (~x0 ) + (~x − ~x0 ) · ∂F ~ (~x0 ) F (~x) = F (~x0 ) + ~r · ∂F Def. 3.4.13
=
F (~x0 ) + F (~x − ~x0 , ~x0 )
Prop. 3.4.14
Def. 3.4.13
=
F (~x0 ) + F (~x − ~x0 )
F (~x0 ) + F (~x) − F (~x0 )
(3.4.36)
⇒ F (~x) − F (~x0 ) ≈ F (~x − ~x0 ) = F (~x) − F (~x0 ).
(3.4.37)
=
Proposition 3.4.16. Chain rule. d dF ~ (~x(t)) = ~x(t) · ∂ F (~x) . dt dt ~ x=~ x(t)
(3.4.38)
118 Quaternion and Clifford Fourier Transforms
Proof. Using the Taylor expansion τ 2 d2 d ~x(t) + ~x(t) + . . . , dt 2 dt2 dF F (~x(t + τ )) − F (~x(t)) (~x(t)) = lim τ →0 dt τ d F (~ x (t) + τ dt ~x(t)) − F (~x(t)) Taylor = lim τ →0 τ d Def. 3.4.1 . ~x(t) · ∂~ F (~x) = dt ~ x=~ x(t) ~x(t + τ ) = ~x(t) + τ
(3.4.39)
(3.4.40)
Definition 3.4.17. Vector derivative. Differentiation of F by its argument ~x ~ ∂~~x F (~x) = ∂F,
(3.4.41)
~ assumed to with the differential operator ∂, (i) have the algebraic properties of a vector in An ≡ G1 (I), I unit pseudoscalar; and (ii) that ~a · ∂~~x with ~a ∈ An is ~a · ∂~~x F as in Def. 3.4.1. Proposition 3.4.18. Algebraic properties of ∂~~x . I ∧ ∂~~x I ∂~~x ∂~~x
(3.3.32)
=
(3.3.31)
=
0.
(3.4.42)
I · ∂~~x
(3.4.43)
~ak~ak · ∂~~x ,
(3.4.44)
(3.3.31)
=
n X k=1
where the ~ak express the algebraic vector properties and the ~ak · ∂~~x the scalar differential properties. ~ for a scalar Definition 3.4.19. Gradient. The vector field f~ = f~(~x) = ∂~~x Φ(~x) = ∂Φ function Φ = Φ(~x) is called the gradient of Φ. Proposition 3.4.20. 3dimensional cross product. For ~b independent of ~x ∈ A3 ≡ G 1 (I3 ): ~ x × ~b) = ~a × ~b. ~a · ∂(~ (3.4.45) Only here × means the 3dimensional cross product (3.3.47), not the commutator product in (3.3.45) or in Proposition 3.4.81.
Geometric calculus 119
Proof. ~ x × ~b) (3.3.47) ~a · ∂(~ = −~a · ∂~ (~x ∧ ~b)I3 Prop. 3.4.5
=
~ x ∧ ~b)]I3 − (~x ∧ ~b)~a · ∂I ~ 3 [−~a · ∂(~
I3 = const.,Prop. 3.4.9
=
Prop. 3.4.6
=
Prop. 3.4.5
=
~ x ∧ ~b)]I3 (3.3.8) [−~a · ∂(~ = [−~a · ∂~ ~x~b ]I3 D
E
2
D
E
~ x~b − ~a · ∂~
2
D
~ x)~b − (~a · ∂~
Props. 3.4.8, 3.4.9
=
D
I3
E 2
E 2
I3
(3.3.8)
E
− ~a~b
D
I3 − ~x(~a · ∂~~b)
2
I3 = −(~a ∧ ~b)I3
(3.3.47)
=
~a × ~b.
(3.4.46)
Proposition 3.4.21. ~a · ∂~ ~x · hAir = ~a · hAir ,
(3.4.47)
A ∈ G(I) independent of ~x. By linearity ~a · ∂~ (~x · A) = ~a · A.
(3.4.48)
Proof. (3.3.16) ~a · ∂~ ~x · hAir = ~a · ∂~ h~x hAir ir−1 Prop. 3.4.5
=
Prop. 3.4.6
=
D
D
~ x) hAi + ~x ~a · ∂~ hAi (~a · ∂~ r r
Props. 3.4.8, 3.4.9
=
h~a hAir ir−1
(3.3.16)
=
E
~ x hAi ) ~a · ∂(~ r
r−1
E r−1
~a · hAir .
(3.4.49)
Proposition 3.4.22. ~ x · (~x ∧ ~b)] = ~a · (~x ∧ ~b) + ~x · (~a ∧ ~b). ~a · ∂[~
(3.4.50)
Proof. ~ x · (~x ∧ ~b)] (3.3.41) ~ x2~b − ~x · ~b~x] ~a · ∂[~ = ~a · ∂[~ Prop. 3.4.4,3.4.5
=
~ x2 )~b + ~x2~a · ∂~~b − [~a · ∂(~ ~ x · ~b)]~x − ~x · ~b(~a · ∂~ ~ x) (~a · ∂~
Proof 3.4.10, Prop. 3.4.9,3.4.21,3.4.8
=
(3.3.41)
=
~a · ~x~b − ~a · ~b~x + ~x · ~a~b − ~x · ~b~a
~a · (~x ∧ ~b) + ~x · (~a ∧ ~b).
(3.4.51)
Proposition 3.4.23. For ~x0 independent of ~x and r ≡ ~r = ~x − ~x0  : ~ = ~a · ~r = ~a · rˆ, ~a · ∂r r where rˆ =
~ r r
.
(3.4.52)
120 Quaternion and Clifford Fourier Transforms
Proof. Compare [162], p. 681. r2 = (~x − ~x0 )(~x − ~x0 ), then ~ 2 ~a · ∂r
Prop. 3.4.5
=
~ x − ~x0 )](~x − ~x0 ) + (~x − ~x0 )[~a · ∂(~ ~ x − ~x0 )] [~a · ∂(~
Prop. 3.4.4,3.4.9
=
~ 2 ~a · ∂r
~a(~x − ~x0 ) + (~x − ~x0 )~a = 2~a · ~r,
Prop. 3.4.5
~ 2r(~a · ∂r) ⇒ ~ = ~a · ~r (3.3.11) = ~a · rˆ. ~a · ∂r r
⇒
=
(3.4.53)
~ 2~a · ~r = 2r(~a · ∂r) (3.4.54)
Proposition 3.4.24. ~ r = rˆrˆ ∧ ~a . ~a · ∂ˆ r
(3.4.55)
Proof. Compare [162], p. 681. ~ r = ~a · ∂~ ~a · ∂ˆ
~r r
Prop. 3.4.7
=
Prop. 3.4.8
=
1 ~ 1 ~a · ∂~r + ~r~a · ∂~ r r 1 rˆ 1 ~ 3.4.23 1 ~ Prop. 3.4.8, ~a · ∂(~x − ~x0 ) − ~r 2 ~a · ∂r = ~a − ~a · rˆ r r r r rˆrˆ~a − rˆ(ˆ r · ~a) (3.3.8) rˆrˆ ∧ ~a = . (3.4.56) r r
Prop. 3.4.5
=
Proposition 3.4.25. r ∧ ~a2 ~ r · ~a) = ˆ ~a · ∂(ˆ . r
(3.4.57)
Proof. ~ r · ~a) = ~a · ∂~ ~r · ~a Prop.=3.4.5 1 ~a · ∂(~ ~ r · ~a) + ~r · ~a(~a · ∂) ~ 1 ~a · ∂(ˆ r r r ~ r · ~ a 1 (ˆ r · ~a)2 Prop. 3.4.7, 3.4.23 1 ~ x · ~a − ~x0 · ~a) − ~a · ∂(~ ~ a · r ˆ r ˆ ∧ ~ a ~ a · ~ a − = r r2 r r 2 2 2 r ∧ ~a) · (ˆ r ∧ ~a) (3.3.43) ˆ r ∧ ~a2 ˆ ~a − (ˆ r · ~a) (3.3.43) (ˆ (3.3.11) r = = . (3.4.58) = r r r
Proposition 3.4.26. ~ r ∧ ~a) = ~a · ∂(ˆ
rˆ · ~a~a ∧ rˆ . r
(3.4.59)
Geometric calculus 121
Proof. ~ r ∧ ~a) (3.3.8) ~ r~a − rˆ · ~a) Prop.=3.4.4 ~a · ∂(ˆ ~ r~a) − ~a · ∂(ˆ ~ r · ~a) ~a · ∂(ˆ = ~a · ∂(ˆ r ∧ ~a2 ~ r)~a + rˆ~a · ∂~ ~ a − ˆ ~a · ∂(ˆ r ˆrˆ ∧ ~a (~a ∧ rˆ)(~a ∧ rˆ) Prop. 3.4.9, 3.4.24, (3.3.43) r = ~a − r r ~ a ∧ r ˆ r ˆ ~ a − (~ a ∧ r ˆ )(ˆ r ∧ ~ a ) (3.3.39), (3.3.40) = r a ∧ rˆ rˆ~a − (ˆ r ∧ ~a) (3.3.7) ~a ∧ rˆrˆ · ~a (3.3.3) ~ = = r r rˆ · ~a~a ∧ rˆ . = r Prop. 3.4.5, 3.4.25
=
(3.4.60)
Proposition 3.4.27. r ∧ ~a ~ r ∧ ~a = − rˆ · ~aˆ ~a · ∂ˆ . r
(3.4.61)
Proof. ~ r ∧ ~a2 ~a · ∂ˆ
Prop. 3.4.5
~ r ∧ ~a2 ~a · ∂ˆ
(3.3.43)
=
=
~ r ∧ ~a, 2ˆ r ∧ ~a~a · ∂ˆ
(3.4.62)
~ r ∧ ~a) · (~a ∧ rˆ)] (3.3.43) ~ r2~a2 − (ˆ ~a · ∂[(ˆ = ~a · ∂[ˆ r · ~a)]
Prop. 3.4.5
=
~ a2 − 2(ˆ ~ r · ~a) ~a · ∂~ r · ~a)~a · ∂(ˆ
ˆ r ∧ ~a2 r (ˆ r · ~ a )ˆ r ∧ ~ a 2 ~ r ∧ ~a = −2 ⇒ 2ˆ r ∧ ~a~a · ∂ˆ r (ˆ r · ~ a )ˆ r ∧ ~ a  ~ r ∧ ~a = − ⇒ ~a · ∂ˆ . r Prop. 3.4.9, 3.4.25
=
−2(ˆ r · ~a)
(3.4.63) (3.4.64) (3.4.65)
Proposition 3.4.28. ~a · ∂~
1 1 1 = − ~a . ~r ~r ~r
(3.4.66)
Proof. 1 ~a · ∂~ ~r
~r Prop. 3.4.5 1 ~ r + ~r~a · ∂~ 1 ~a · ∂~ 2 = ~a · ∂~ 2 r r r2 2 Prop. 3.4.7 1 ~ Prop.=3.4.23 1 ~a − 2 ~r ~a · rˆ = ~a + ~r(− 3 )~a · ∂r 2 r r r2 r3 1 1 (3.3.7) 1 1 1 1 11 (3.3.12), (3.3.11) 1 1 = ~a − 2 ~a · = ~a − ~a − ~a ~r ~r ~r ~r ~r ~r ~r ~r ~r ~r 1 1 = − ~a . ~r ~r (3.3.12)
=
(3.4.67)
122 Quaternion and Clifford Fourier Transforms
Proposition 3.4.29. ~a · ∂~
~a · rˆ 1 = −2 3 . 2 r r
(3.4.68)
Proof. 1 ~a · ∂~ 2 r
(3.3.7)
= −
2 ~ ~a · ∂r r3
Prop. 3.4.23
=
−
2 ~a · rˆ. r3
(3.4.69)
Proposition 3.4.30. 2 r ∧ ~a2 1 ~ 2 1 = 3(~a · rˆ) − ˆ (~a · ∂) . 2 r2 r4
(3.4.70)
Proof. 1 ~ 21 (~a · ∂) 2 r2
1 ~ ~a · rˆ ~a · ∂(−2 3 ) 2 r 1 1 Prop. 3.4.5 ~ a · rˆ) = −(~a · ∂~ 3 )~a · rˆ − 3 ~a · ∂(~ r r a · rˆ 1 ˆ r · ~a2 Prop. 3.4.7,3.4.23,3.4.25 3~ = ~ a · r ˆ − ~ a r4 r3 r 2 2 3(~a · rˆ) − ˆ r · ~a = . r4 Prop. 3.4.29
=
(3.4.71)
Proposition 3.4.31. 3 1 r ∧ ~a2~a · rˆ ~ 3 1 = −4(~a · rˆ) + 4ˆ (~a · ∂) . 6 r2 r5
(3.4.72)
Proof. 2 1 r · ~a2 ~ 3 1 = 1 (~a · ∂) ~ 1 (~a · ∂) ~ 2 1 Prop.=3.4.30 1 (~a · ∂) ~ 3(~a · rˆ) − ˆ (~a · ∂) 6 r2 3 2 r2 3 r4 ~ a · rˆ) − 2ˆ ~ r ∧ ~a a · rˆ)~a · ∂(~ r ∧ ~a~a · ∂ˆ Prop. 3.4.5,3.4.7 1 3 2(~ = 3 r4 1 1 ~ + [3(~a · rˆ)2 − ˆ r ∧ ~a2 ](−4) 5 ~a · ∂r 3 r a2 a a · rˆ) ˆr∧~ + 23 ˆ r ∧ ~aˆ r · ~a ˆr∧~ Prop. 3.4.23,3.4.25,3.4.27 2(~ r r = r4 4 1 − [3(~a · rˆ)2 − ˆ r ∧ ~a2 ] 5 ~a · rˆ 3 r 2(~a · rˆ)ˆ r ∧ ~a2 + 23 ˆ r ∧ ~a2 rˆ · ~a − 4(~a · rˆ)3 + 43 ˆ r ∧ ~a2 rˆ · ~a = r5 3 2 −4(~a · rˆ) + 4ˆ r ∧ ~a ~a · rˆ = . (3.4.73) 5 r
Geometric calculus 123
Proposition 3.4.32. ~a · ∂~ log r =
~a · ~r = ~a · ~r−1 . r2
(3.4.74)
Proof. ~a · ∂~ log r
Prop. 3.4.7
=
1 ~ ~a · ∂r r
Prop. 3.4.23
=
~a · rˆ r
(3.3.11)
=
~a · ~r . r2
(3.4.75)
Proposition 3.4.33. For integer k and ~r if k < 0 : ~ r2k = 2k~a · ~r ~r2(k−1) . ~a · ∂~
(3.4.76)
Proof. ~ r2k = ~a · ∂r ~ 2k ~a · ∂~
Prop. 3.4.7
=
~ 2kr2k−1~a · ∂r
Prop. 3.4.23
=
2kr2k−1~a · rˆ
= 2kr2k−2~a · ~r = 2k~a · ~r ~r2(k−1) .
(3.4.77)
Proposition 3.4.34. For integer k and ~r 6= 0 if 2k + 1 < 0 : ~ r2k+1 = ~r2k (~a + 2k~a · rˆrˆ). ~a · ∂~
(3.4.78)
Proof. ~ r2k+1 = ~a · ∂(~ ~ r2k ~r) Prop.=3.4.5 ~a · ∂(~ ~ r2k )~r + ~r2k~a · ∂~ ~r ~a · ∂~ Prop. 3.4.8,3.4.33
=
2k~a · ~r~r2(k−)1 + ~r2k~a = ~r2k (~a + 2k~a · rˆrˆ).
Proposition 3.4.35. Taylor expansion of
(3.4.79)
1 ~ x−~a .
1 1 1 1 1 1 1 = + ~a + ~a ~a + . . . . ~x − ~a ~x ~x ~x ~x ~x ~x
(3.4.80)
124 Quaternion and Clifford Fourier Transforms
Proof. 1 ~x − ~a ~ k (−~a · ∂)
Prop. 3.4.12
=
~ −~a·∂
e
∞ ~ k1 1 X (−~a · ∂) = , ~x k=0 k! ~x
1 ~ k−1 (−~a · ∂) ~ 1 Prop.=3.4.28 (−~a · ∂) ~ k−1 1 ~a 1 = (−~a · ∂) ~x ~x ~x ~x 1 1 ~ k−2 (−~a · ∂)[ ~ ~a ] = (−~a · ∂) ~x ~x 1 1 1 1 Prop. 3.4.5 k−2 ~ = (−~a · ∂) [(−~a · ∂~ )~a + ~a(−~a · ∂~ )] ~x ~x ~x ~x Prop. 3.4.28 ~ k−2 1 ~a 1 ~a 1 = 2(−~a · ∂) ~x ~x ~x Props. 3.4.5,3.4.28 ~ k−3 1 ~a 1 ~a 1 ~a 1 = . . . = 2 3(−~a · ∂) ~x ~x ~x ~x 1 Props. 3.4.5,3.4.28 ~ k−k (~a 1 )k = . . . = k!(−~a · ∂) ~x ~x 1 1 k = k! (~a ) ~x ~x 1 ~ k 1 = 1 (~a 1 )k ⇒ (−~a · ∂) k! ~x ~x ~x ∞ X 1 1 1 k ⇒ = (~a ) . ~x − ~a k=0 ~x ~x
(3.4.81)
(3.4.82) (3.4.83) (3.4.84)
Proposition 3.4.36. Legendre Polynomials. The Legendre polynomials Pn (homogeneous and of grade n) are defined by: ∞ ∞ X 1 Pn (ˆ x~a) X Pn (~x~a) ≡ = . n+1 ~x − ~a n=0 ~x ~x2n+1 n=0
(3.4.85)
The explicit first four polynomials are: P0 (~x~a) = 1,
(3.4.86)
P1 (~x~a) = ~x · ~a, 1 1 P2 (~x~a) = [3(~x · ~a)2 − ~a2~x2 ] = (~x · ~a)2 + (~x ∧ ~a)2 , 2 2 1 3 3 2 2 P3 (~x~a) = [5(~x · ~a) − 3~a ~x ~x · ~a] = (~x · ~a)3 + ~x · ~a(~x ∧ ~a)2 , 2 2 n n n Pn (~x~a) = ~x Pn (ˆ x~a) = ~x ~a Pn (ˆ xa ˆ).
(3.4.87) (3.4.88) (3.4.89) (3.4.90)
Proof. ∞ ~ n 1 X 1 (−~a · ∂) Prop. 3.4.12 −~a·∂ ~ F (~x − ~a) = = e F (~x) = ~x − ~a n! ~x n=0 1 1 1 ~ 2 1 − 1 (~a · ∂) ~ 3 1 + ... = − ~a · ∂~ + (~a · ∂) ~x ~x 2 ~x 6 ~x ⇒ P0 (~x~a) = 1,
(3.4.91) (3.4.92)
Geometric calculus 125
−~a · ∂~
1 ~x
1 ~ x Prop.=3.4.10 1 ~a · xˆ ~a · ∂~ ~x2 ~x2 a · ~x Prop. 3.4.11 ~ = , ~x3 ⇒ P1 (~x~a) = ~a · ~x, Prop. 3.4.7
=
(3.4.93) (3.4.94)
1 ~ 2 1 = − 1 (~a · ∂)(−~ ~ ~ 1 = − 1 (~a · ∂) ~ ~a · ~x (~a · ∂) a · ∂) 2 ~x 2 ~x 2 ~x3 1 1 1 1 Prop. 3.4.5 ~ a · ~x) = − [~a · ∂~ 3 ]~a · ~x − ~a · ∂(~ 2 ~x 2 ~x3 1 −3 Prop. 3.4.7,3.4.21 ~ x]~a · ~x − 1 1 ~a · ~a = − [ 4 ~a · ∂~ 2 ~x 2 ~x3 1 1 2 Prop. 3.4.10 1 3 = ~a · xˆ(~a · ~x) − ~a 4 2 ~x 2 ~x3 (~a · ~x)2 + 12 {(~a · ~x)2 − ~x2~a2 } a · ~x)2 − ~x2~a2 (3.3.11),(3.3.5) 1 3(~ = = 2 ~x5 ~x5 a · ~x)2 + 12 {~a ∧ ~x}2 (3.3.43) (~ = (3.4.95) ~x5 1 ⇒ P2 (~x~a) = [3(~a · ~x)2 − ~x2~a2 ] 2 1 = (~a · ~x)2 + {~a ∧ ~x}2 , (3.4.96) 2 2 2 2 1 ~ 3 1 = − 1 ~a · ∂[ ~ 1 (~a · ∂) ~ 2 1 ] = − 1 ~a · ∂[ ~ 1 3(~a · ~x) − ~x ~a ] − (~a · ∂) 6 ~x 3 2 ~x 3 2 ~x5 1 1 Props. 3.4.4,3.4.5 = − (~a · ∂~ 5 )[3(~a · ~x)2 − ~x2~a2 ] 6 ~x 1 ~ a · ~x) − ~a2~a · ∂~ ~ x2 ] − [3 2(~a · ~x)~a · ∂(~ 6~x5 −5 Props. 3.4.5,3.4.7,3.4.21,3.4.23 ~ x)[3(~a · ~x)2 − ~x2~a2 ] (~a · ∂~ = − 6~x6 1 − [6(~a · ~x)~a · ~a − ~a2 2(~a · ~x)] 6~x5 5 ~a · ~x 4~a2~a · ~x Prop. 3.4.10 2 2 2 = [3(~ a · ~ x ) − ~ x ~ a ] − 6~x6 ~x 6~x5 15(~a · ~x)3 − 5~x2~a2~a · ~x − 4~x2~a2~a · ~x = 6~x7 1 = [5(~a · ~x)3 − 3~x2~a2~a · ~x] 2~x7 1 3 = 7 [(~a · ~x)3 + ~a · ~x((~a · ~x)2 − ~x2~a2 )] ~x 2 3 (3.3.43) 1 = [(~a · ~x)3 + ~a · ~x(~a ∧ ~x)2 ] (3.4.97) 7 ~x 2
126 Quaternion and Clifford Fourier Transforms
1 P3 (~x~a) = [5(~a · ~x)3 − 3~a2~x2~x · ~a] 2 3 = (~a · ~x)3 + ~x · ~a(~a ∧ ~x)2 . 2
⇒
(3.4.98)
Homogeneity of degree n of the Pn : ~ n 1 def. Pn (~x~a) (−~a · ∂) = n! ~x ~x2n+1
⇒
~ n 1 = n! Pn (~x~a) (−~a · ∂) ~x ~x2n+1
(3.4.99)
and ~ n+1 1 = (n + 1)! Pn+1 (~x~a) (−~a · ∂) ~x ~x2(n+1)+1 (n + 1)! Pn+1 (~x~a) ~ Pn (~x~a) ⇒ = (−~a · ∂) 2(n+1)+1 n! ~x ~x2n+1 1 1 Prop. 3.4.5 ~ n (~x~a) = [−~a · ∂~ 2n+1 ]Pn (~x~a) − 2n+1 ~a · ∂P ~x ~x 1 Prop. 3.4.7 2n + 1 ~ x − ~ n (~x~a) Pn (~x~a)~a · ∂~ ~a · ∂P = ~x2n+2 ~x2n+1 ~a · ~x 1 Prop. 3.4.10 2n + 1 ~ n (~x~a) Pn (~x~a) − 2n+1 ~a · ∂P = 2n+2 ~x ~x ~x 2n + 1 1 ~ n (~x~a). = 2n+3 Pn (~x~a)~a · ~x − 2n+1 ~a · ∂P ~x ~x
(3.4.100)
Pn (~x~a)~a · ~x is a homogeneous function of degree n + 1, if we assume Pn to be homogeneous of degree n : Pn (~x~a)
assume
=
n X
αk (~x · ~a)k ~xn−k~an−k ,
(3.4.101)
k=0
αk = const., which is especially true for n = 0, 1, 2, 3. The right term ~ n (~x~a) = ~a · ∂~ ~a · ∂P
n X
αk (~x · ~a)k ~xn−k~an−k
k=0 Props. 3.4.4,3.4.5,3.4.9
=
n n X
~ x · ~a)k ]~xn−k~an−k αk [~a · ∂(~
k=0
~ xn−k ]~an−k +αk (~x · ~a)k [~a · ∂~ Props. 3.4.5,3.4.7
=
o
n n X
~ x · ~a)]~xn−k~an−k αk [k(~x · ~a)k−1~a · ∂(~
k=0
~ x]~an−k +αk (~x · ~a)k [(n − k)~xn−k−1~a · ∂~ Props. 3.4.10,3.4.21
=
n n X
αk [k(~x · ~a)k−1~a · ~a]~xn−k~an−k
k=0 k
· ~x n−k ]~a ~x
a n−k−1 ~
+αk (~x · ~a) [(n − k)~x
o
Geometric calculus 127
=
n X
(
αk k(~x · ~a)k−1
k=0
~x2 n−k 2 n−k ~x ~a ~a ~x2
+αk (n − k)(~x · ~a) =
1 ~x2
xn−k−1 k+1 ~ ~x
~x n−k ~a ~x
)
n n X
αk k(~x · ~a)k−1 ~xn−k+2~an−k+2
k=0
+αk (n − k)(~x · ~a)k+1 ~xn−k~an−k
o
(3.4.102)
~ n (~x~a), to be homogeneous of degree n + 1. Hence yields ~x2 ~a · ∂P (Polynomial homogeneous of degree n + 1) (n + 1)! Pn+1 (~x~a) = . 2(n+1)+1 n! ~x ~x2(n+1)+1
(3.4.103)
By induction every Pn will therefore be homogeneous of degree n. This and the (~ x~a) explicit expressions above for ~xPn+1 2(n+1)+1 fully prove for all n : Pn (~x~a) = ~xn Pn (ˆ x~a) = ~xn ~an Pn (ˆ xa ˆ).
(3.4.104)
Definition 3.4.37. Redefinition of differential, overdots. ~ F (~a) = ~a · ∂F
(3.3.7)
=
1 ~ ~˙ aF˙ ), (~a∂F + ∂~ 2
(3.4.105)
where the overdots indicate, that only F is to be differentiated and not ~a . Proposition 3.4.38. For ~a ∈ / An ≡ G 1 (I), P = PI : ~a · ∂~~x = ~a · P (∂~~x ) = P (~a) · ∂~~x ,
P (~a) ∈ An .
(3.4.106)
Proof. Prop. 3.4.18
~a · ∂~~x
=
~a · P (∂~~x )
Prop. 3.4.18
=
n X
~a · ~ak (~ak · ∂~~x )
k=1
=
n X
P (~a) · ~ak (~ak · ∂~~x ) = P (~a) · ∂~~x .
(3.4.107)
k=1
Proposition 3.4.39. ~ F (~a) = F (P (~a)) = P (~a) · ∂F.
F (~a) = 0, if P (~a) = 0.
(3.4.108)
Proof. F (~a) ⇒
Def. 3.4.13, Prop. 3.4.38
=
F (P (~a))
F (~a) = 0, if P (~a) = 0.
Def. 3.4.13
=
~ P (~a) · ∂F, (3.4.109)
128 Quaternion and Clifford Fourier Transforms
Proposition 3.4.40. Differential of composite functions. For F (~x) = G(f (~x)) and f : ~x ∈ An = G 1 (I) → f (~x) ∈ A0n = G 1 (I 0 ),
(3.4.110)
we have ~ ~ = f · ∂G, ~a · ∂F
F (~a) = G(f (~x), f (~x, ~a))
(Def. 3.4.13).
(3.4.111)
The differential of composite functions is the composite of differentials : F (~x, ~a) = G(f (~x), f (~x, ~a))
(explicit).
(3.4.112)
Proof. Using Taylor expansion (Prop. 3.4.12): ~ (~x) + 1 τ 2 (~a · ∂) ~ 2 f (~x) + . . . , f (~x + τ~a) = f (~x) + τ~a · ∂f 2 ~ ~a · ∂G(f (~x))
(3.4.113)
d G(f (~x + τ~a)) dτ τ =0 Taylor (Prop. 3.4.12), Def. 3.4.13 d = G(f (~x) + τ f (~a)) dτ τ =0 Def. 3.4.1 0 ~ = f (~a) · ∂~x0 G(~x )
Def. 3.4.13
=
~ x0 =f (~ x)
= f (~a) · ∂~ G (evaluation at corresponding points.)
(3.4.114)
Definition 3.4.41. Second differential. ˙ F~a~b = ~b · ∂~ ~a · ∂~ F˙ (~x).
(3.4.115)
˙ Suppressing ~x : F~a~b = ~b · ∂~ ~a · ∂~ F˙ . Proposition 3.4.42. Integrability condition. F~a~b = F~b~a .
(3.4.116)
The second differential is a symmetric bilinear function of its differential arguments ~a, ~b . Proof.
Def. 3.4.1 ˙ ˙ dF˙ (~x + τ~a) F~a~b (~x) ≡ ~b · ∂~ ~a · ∂~ F˙ (~x) = ~b · ∂~ dτ τ =0
d2 F˙ (~x + τ~a + σ~b) = dσdτ τ =0,σ=0 = lim lim
σ→0 τ →0
F (~ x+τ~a+σ~b)−F (~ x+σ~b) τ
σ
which is symmetric under (~a, τ ) ↔ (~b, σ). Hence
−
F (~ x+τ~a)−F (~ x) τ
,
(3.4.117)
Geometric calculus 129
˙ ˙ F~a~b (~x) = ~b · ∂~ ~a · ∂~ F˙ (~x) = ~a · ∂~ ~b · ∂~ F˙ (~x) = F~b~a (~x).
(3.4.118)
The bilinearity follows from the linearity in each argument (Props. 3.4.2, 3.4.3, 3.4.14). Proposition 3.4.43. Differential of identity function. ~a · ∂~~x~x = P (~a) = ∂~~x (~x · ~a).
(3.4.119)
Proof. We first prove the first identity: Prop. 3.4.38 Prop. 3.4.8 ~a · ∂~~x~x = PI (~a) · ∂~~x~x = PI (~a).
(3.4.120)
Regarding the second identity, especially for base vectors ~ak ∈ An = G 1 (I): ~ak · ∂~~x~x = PI (~ak ) = ~ak , Prop. 3.4.18 ∂~~x (~x · ~a) =
X
Prop. 3.4.21
X
(3.4.121)
~ak~ak · ∂~~x (~x · ~a)
k
=
~ak~ak · ~a = PI (~a).
(3.4.122)
k
Proposition 3.4.44. Operator identity. ∂~~x = PI (∂~~x ) = ∂~~a (~a · ∂~~x ).
(3.4.123)
Proof. Proposition 3.4.18 and 3.4.43. Proposition 3.4.45. Derivative from differential. ∂~~x F (~x) = ∂~~a (~a · ∂~~x )F (~x) = ∂~~a F (~x, ~a)
(3.4.124)
Proof. Proposition 3.4.44 and Definition 3.4.13. Definition 3.4.46. ~ = ∂~~x F (~x) Prop.=3.4.45 ∂~~a F (~x, ~a) = ∂~ F , ∂F
(3.4.125)
where ∂~ is the derivative with respect to the differential argument ~a of F (~x, ~a). Proposition 3.4.47. ~ = ∂~ · F + ∂~ ∧ F. ∂F
(3.4.126)
Proof. Vector property (Prop. 3.4.18) of ∂~ = ∂~~x and (3.3.18). Proposition 3.4.48. Gradient. For scalar F = Φ(~x) : ∂~ · Φ = 0,
~ = ∂~ ∧ Φ = Φ˙ ∂. ~˙ ∂Φ
(3.4.127)
130 Quaternion and Clifford Fourier Transforms
Proof. (3.3.21), Prop. 3.4.47 and (3.3.23). Remark 3.4.49. In Prop. 3.4.48 the special definition of Hestenes and Sobczyk[161] in (3.3.20) and (3.3.21) for the inner product becomes important. It should be possible to make it more intuitive by replacing the inner product with the contraction [110]. Definition 3.4.50. Divergence and curl. • Divergence of F : ∂~ · F , • Curl of F : ∂~ ∧ F . ~ .) (Full vector derivative of F : ∂F Proposition 3.4.51. Vector derivative of sums. ~ + G) = ∂F ~ + ∂G. ~ ∂(F
(3.4.128)
Proof. Prop. 3.4.4 ~ 3.4.46 ~ ~ + G) Def.= = ∂(F + G) ∂(F ∂(F + G) Prop. 3.4.18
=
X
~ak~ak · ∂~~a (F + G)
k Prop. 3.4.4
=
X
~ak (~ak · ∂~~a F + ~ak · ∂~~a G)
k distributivity, (3.3.19)
=
X
~ak~ak · ∂~~a F +
=
~ak~ak · ∂~~a G
k
k Prop. 3.4.18, Def. 3.4.46
X
~ + ∂G. ~ ∂F
(3.4.129)
Note that geometric multiplication is distributive with respect to addition. Proposition 3.4.52. Vector derivative of products. ~ G) = ∂~˙ F˙ G + ∂F ~˙ G. ˙ ∂(F
(3.4.130)
Proof. Prop. 3.4.5 ~ 3.4.46 ~ ~ G) Def.= ∂(F ∂(F G) = ∂(F G + F G) Def. 3.4.46, Prop. 3.4.51
~ G = ∂~ F G + ∂F Def. 3.4.46 ~˙ ˙ ~˙ G. ˙ = ∂ F G + ∂F
(3.4.131)
The third equality is a special case of Prop. 3.4.51, if we take the definition of ∂~ in Def. 3.4.46 into account. The last term is to be interpreted as:
~˙ G˙ = ∂~~y F (~x)G(~y ) ∂F
~ y =~ x
.
(3.4.132)
Geometric calculus 131
Proposition 3.4.53. ~ x2 = 2~x. ∂~
(3.4.133)
Proof. X X ~ x2 = (∂~ ~ x)~x + ∂~ ~˙ x~x˙ Prop.=3.4.18 ( ~ak~ak · ∂~~x~x)~x + ∂~ ~ak ~x~ak · ∂~~x~x k Prop. 3.4.43
=
X
(
~ak~ak )~x +
k (3.3.32)
=
X
k
~ak ~x~ak
(3.3.7)
= 2
k
X
~ak ~x · ~ak
k
2~x.
(3.4.134)
Proposition 3.4.54. ~ x = xˆ. ∂~
(3.4.135)
Proof. ~ x Prop.=3.4.18 ∂~
X
~ak~ak · ∂~~x ~x
Prop. 3.4.10
=
X
k
~ak~ak · xˆ
(3.3.32)
=
xˆ.
(3.4.136)
k
Proposition 3.4.55. For F = F (~x) : ~ = xˆ dF . ∂F d~x
(3.4.137)
Proof. ~ ~a · ∂F ⇒
Prop. 3.4.7
=
~ ∂F
~ x dF ~a · ∂~ d~x
Def. 3.4.46
=
Prop. 3.4.54
=
~a · xˆ
dF d~x
~ = ∂~~a~a · xˆ dF ∂~~a (~a · ∂)F d~x
(3.4.138)
Prop. 3.4.43
=
xˆ
dF . d~x
(3.4.139)
Definition 3.4.56. Sides of differentiation. Only right side differentiation: ~ F ∂G
Def. 3.4.46
=
˙ ˙ F ∂~ G = F ∂~ G.
(3.4.140)
Left and right side differentiation (another form of the product rule Prop. 3.4.52): ˙ ~˙ + F ∂~˙ G. ˙ F˙ ∂~ G˙ = F˙ ∂G
(3.4.141)
Definition 3.4.57. Adjoint. For f : ~x ∈ An = G 1 (I) → f (~x) ∈ A0n = G 1 (I 0 ),
(3.4.142)
~ · ~a0 ) f (~a0 ) ≡ ∂(f
(3.4.143)
the entity is the adjointof f or explicitly: f (~x, ~a0 ) ≡ ∂~~a
h
(~a · ∂~~x )f (~x) · ~a0
i
Def. 3.4.37
=
∂~~a [f (~x, ~a) · ~a0 ].
(3.4.144)
132 Quaternion and Clifford Fourier Transforms
Proposition 3.4.58. ~ · ~a0 ), f (~a0 ) = ∂(f
(3.4.145)
f (~x, ~a0 ) = ∂~~x (f (~x) · ~a0 ).
(3.4.146)
or explicitly: Proof. f (~x, ~a0 )
Def. 3.4.57
=
Prop. 3.4.45
=
h
∂~~a
i
(~a · ∂~~x )f (~x) · ~a0 = ∂~~a (~a · ∂~~x ) [f (~x) · ~a0 ]
∂~~x (f (~x) · ~a0 ) .
(3.4.147)
(Compare [161] p. 50; [107] p. 23 (1.109), p. 24 (1.118), and p. 104 (5.11).) Proposition 3.4.59. Linearity of adjoint. For scalar α ∈ R : f (~a0 + ~b0 ) = f (~a0 ) + f (~b0 ),
f (α ~a0 ) = αf (~a0 ).
(3.4.148)
Proof. Linearity of the inner product (3.3.9). Proposition 3.4.60.
P f (~a0 ) = f (~a0 ).
(3.4.149)
Proof.
P f (~a0 )
Def. 3.4.57, Prop. 3.4.58
=
P ∂~~x (f (~x) · ~a0 ) = P (∂~~x )f (~x) · ~a0
Prop. 3.4.58 f (~a0 ). ∂~~x f (~x) · ~a0 =
(3.4.150)
f (~a0 ) = ∂~~a f (~a) · ~a0 = f (P 0 (~a0 )),
(3.4.151)
Prop. 3.4.44
=
Proposition 3.4.61.
with P 0 the projection into the range of f and f , i.e. into A0n ≡ G 1 (I 0 ). Proof. f (~a0 )
Def. 3.4.57
=
Prop. 3.4.58 ∂~~a f (~a) · ~a0 = ∂~~x f (~x) · ~a0 = ∂~~x [P 0 f (~x)] · ~a0
Prop. 3.4.58 = ∂~~x f (~x) · P 0 (~a0 ) = f (P 0 (~a0 )).
(3.4.152)
Proposition 3.4.62. P f P 0 (~a0 ) = f (~a0 ),
P 0 f P (~a) = f (~a).
(3.4.153)
Proof. First identity: Definition 3.4.57 and Proposition 3.4.58. Second identity: Proposition 3.4.39 and because the range of f and f is A0n = G i (I 0 ).
Geometric calculus 133
Proposition 3.4.63. Change of variables. For F (~x) = G(f (~x)), i.e. ~x → f (~x) : ∂~~x F (~x) = f (∂~~x0 )G(~x0 ),
(3.4.154)
i.e. ∂~~x = f (∂~~x0 ) . Proof. Prop. 3.4.44 ~ ∂~~x F (~x) = ∂~a [~a · ∂~~x F (~x)] = ∂~~a [~a · ∂~~x G(f (~x))] Prop. 3.4.40
=
∂~~a [f (~a) · ∂~~x0 G(~x0 )] 
Def. 3.4.57
=
{z
scalar
~ x0 =f (~ x)
}
= [∂~~a f (~a) · ∂~~x0 G(~x0 )]
[f (∂~~x0 )G(~x0 )]~x0 =f (~x) = f (∂~~x0 )G(~x0 ).
~ x0 =f (~ x)
(3.4.155)
Proposition 3.4.64. Second derivative. ∂~~x2 F (~x) = ∂~~b ∂~~a F~a~b = (∂~~b · ∂~~a + ∂~~b ∧ ∂~~a )F~a~b .
(3.4.156)
Proof. Def. 3.4.13 Prop. 3.4.54 ~ ∂~~x F (~x) = ∂~a ~a · ∂~~x F (~x) = ∂~~a F~a ⇒ ∂~ 2 F (~x) = ∂~~x ∂~~x F (~x) = ∂~~x ∂~~a F~a
(3.4.157)
~ x
Prop. 3.4.54
=
~b · ∂~˙~x
∂~~b [
Def. 3.4.41 ~ ~ ∂~~a F˙~a ] = ∂~~b ∂~~a [(~b · ∂~~x )F~a ] = ∂~b ∂~a F~a~b
 {z }
scalar operator Prop. 3.4.18
=
(∂~~b · ∂~~a + ∂~~b ∧ ∂~~a )F~a~b .
(3.4.158)
Proposition 3.4.65. Integrability condition for vector derivative. ∂~~x ∧ ∂~~x = 0
⇔
F~a~b = F~b~a .
(3.4.159)
Proof. (⇒) ∂~~x ∧ ∂~~x = 0 ⇒
(3.4.160)
(3.3.42) 0 = (~a ∧ ~b) · (∂~~x ∧ ∂~~x )F = ~a · [~b · (∂~~x ∧ ∂~~x )]F = ~a · [(~b · ∂~~x )∂~~x − ∂~~x (~b · ∂~~x )]F = [(~b · ∂~~x )(~a · ∂~~x )F − (~a · ∂~~x )(~b · ∂~~x )F ] Def. 3.4.41
=
F~a~b − F~b~a ,
(3.4.161)
(⇐) Integrability (Prop. 3.4.42) means: F~a~b = F~b~a . Obviously 1 Prop. 3.4.42 1 ~ ~ = (∂~ ∂~a − ∂~~a ∂~~b )F~a~b = ∂~~b ∧ ∂~~a F~a~b 0 = (∂~~b ∂~~a F~a~b − ∂~~a ∂~~b F~b~a ) 2 2 b Prop. 3.4.64 ~ = ∂~x ∧ ∂~~x F (~x) (3.4.162) ⇒
∂~~x ∧ ∂~~x = 0.
(3.4.163)
134 Quaternion and Clifford Fourier Transforms
Proposition 3.4.66. Laplacian. Integrability of F ⇔ ∂~~x2 = ∂~~x · ∂~~x . Proof. Integrability of F ⇔
Prop. 3.4.65
⇔
∂~~x2
∂~~x ∧ ∂~~x = 0
(3.3.6), (3.3.18)
=
∂~~x · ∂~~x + ∂~~x ∧ ∂~~x = ∂~~x · ∂~~x .
(3.4.164)
Proposition 3.4.67. ∂~~x ∧ ~x = 0.
(3.4.165)
Proof. 1 Prop. 3.4.53 1 ~ = (∂~x ∧ ∂~~x )~x2 ∂~~x ∧ ~x = ∂~~x ∧ (2~x) 2 2
Prop. 3.4.65
=
0,
(3.4.166)
because F = ~x2 is integrable. Proposition 3.4.68. (3.3.33) ∂~~x~x = ∂~~x · ~x = n.
(3.4.167)
Proof. Prop. 3.4.47 ~ Prop. 3.4.67 ~ Prop. 3.4.18 ∂~~x~x = ∂~x · ~x + ∂~~x ∧ ~x = ∂~x · ~x =
n X
~ak~ak · ∂~~x · ~x
k=1
= =
n X k=1 n X
Prop. 3.4.43 ~ak · (~ak · ∂~~x )~x =
n X
~ak · P (~ak )
k=1
~ak · ~ak = n.
(3.4.168)
k=1
Proposition 3.4.69. ∂~~x ~xk = k~xk−2~x.
(3.4.169)
Proof. ∂~~x ~xk
Prop. 3.4.55
=
xˆ
d~xk = xˆk~xk−1 = k~xk−2 ~xˆ x = k~xk−2~x. d~x
(3.4.170)
Proposition 3.4.70. ∂~~x
~ x
~xk
=
n−k . ~xk
(3.4.171)
Geometric calculus 135
Proof. ∂~~x
~ x
~xk
1 ∂~~x (~x) k + ~x∂~~x ~x−k ~x 1 1 Prop. 3.4.68,3.4.69 = n k + ~x(−k)~x−k−2~x = n k − k~x−k−2~x2 ~x ~x n−k n . (3.4.172) = k − k~x−k = ~x ~xk Prop. 3.4.52
=
Proposition 3.4.71. ~x ∂~~x log ~x = 2 = ~x−1 . ~x
(3.4.173)
Proof. x 1 Prop. 3.4.55 d log ~ ∂~~x log ~x = xˆ = xˆ ~x ~x
(3.3.12)
=
~x ~x2
(3.3.12)
=
~x−1 .
(3.4.174)
Proposition 3.4.72. For A = P (A) = hAir : ˙ ∂~~x (~x˙ · A) = A · ∂~~x~x = rA.
(3.4.175)
Proof. If A is a simple rblade, then 1 ˙ (3.3.39) ˙ ˙ ∂~~x (~x˙ · A) = ∂~~x (A · ~x˙ )(−1)r−1 = [∂~~x A~x˙ − (−1)r 2
(∂~~x~x)
A](−1)r−1
 {z }
=n,Prop. 3.4.68
1 ˙ ˙ = [∂~~x A~x˙ − (−1)r A(∂~~x~x)](−1)r−1 = (∂~~x · A)~x˙ (−1)r−1 2 (3.3.39) = (A · ∂~~x )~x, (3.3.30)
⇒
˙ A−1 ∂~~x (~x˙ · A) = A−1 (A · ∂~~x )~x Prop. 3.4.18
r X
Prop. 3.4.43
r X
=
(3.3.35)
=
PA (∂~~x )~x
~ak~ak · ∂~~x ~x =
~ak~ak
r X
~ak (~ak · ∂~~x )~x
k=1
k=1
=
(3.4.176)
(3.3.33)
=
r,
(3.4.177)
k=1
⇒
˙ ∂~~x (~x˙ · A) = (A · ∂~~x )~x = rA.
(3.4.178)
Last step: Multiplication with A from the left. Distributivity (3.3.19), (3.3.25) gives the same result even for nonsimple rvectors A. Proposition 3.4.73. For A = P (A) = hAir : ˙ ∂~~x (~x˙ ∧ A) = A ∧ ∂~~x~x = (n − r)A.
(3.4.179)
136 Quaternion and Clifford Fourier Transforms
Proof. If A is a simple rblade, then 1 ˙ (3.3.40) ˙ ˙ ∂~~x (~x˙ ∧ A) = ∂~~x (A ∧ ~x˙ )(−1)r = [∂~~x A~x˙ + (−1)r 2
(∂~~x~x)
A](−1)r
 {z }
=n,Prop. 3.4.68
1 ˙ ˙ = [∂~~x A~x˙ + (−1)r A(∂~~x~x)](−1)r = (∂~~x ∧ A)~x˙ (−1)r 2 (3.3.40) = (A ∧ ∂~~x )~x,
(3.4.180)
˙ A−1 ∂~~x (~x˙ ∧ A) = A−1 (A ∧ ∂~~x )~x = A−1 I −1 I(A ∧ ∂~~x )~x
(3.3.30)
⇒
Ipseudoscalar, (3.3.40)
=
A−1 I −1 [{z} I · (∂~~x ∧ A)(−1)r ]~x n

{z
}
r+1 grades
(3.3.39)
=
A−1 I −1 [(∂~~x ∧ A) · I]~x(−1)r+(r+1)(n−r−1)
(3.3.38)
A−1 I −1 [∂~~x · (AI)]~x(−1)r+(r+1)(n−r)−r−1
(3.3.39)
A−1 I −1 [(AI) ·∂~~x ]~x(−1)(r+1)(n−r)−1+n−r−1
=
=
 {z } n−r
(3.3.40)
=
−1 −1
A I
[(IA) ·∂~~x ]~x (−1)r(n−r)+r(n−r)  {z } n−r

{z
}
=+1
(3.3.35), (3.3.36) AA−1 =1,(3.3.30)
⇒
Proof of Prop. 3.4.72 = PIA (∂~~x )~x = n − r. ˙ ∂~~x (~x˙ ∧ A) = (A ∧ ∂~~x )~x = (n − r)A.
(3.4.181) (3.4.182)
Last step: Multiplication with A from the left. The distributive rule for the inner product gives the same result even for nonsimple A. Proposition 3.4.74. For A = P (A) = hAir : n
X ˙ ∂~~x A~x˙ = ~ak A~ak = (−1)r (n − 2r)A.
(3.4.183)
k=1
Proof. n
n
k=1 n X
k=1
X Prop. 3.4.18 X k ˙ ˙ ˙ ~ak A(~ak · ∂~~x )~x˙ ∂~~x A~x˙ = ~a (~ak · ∂~~x )A~x˙ = Prop. 3.4.43
=
~ak A~ak ,
(3.4.184)
k=1 (3.3.39), (3.3.40) ˙ ˙ ˙ ∂~~x A~x˙ = ∂~~x [A · ~x˙ + A ∧ ~x˙ ] = ∂~~x [~x˙ · A(−1)r−1 + ~x˙ ∧ A(−1)r ] ˙ ˙ = (−1)r [−∂~~x (~x˙ · A) + ∂~~x (~x˙ ∧ A)] Props. 3.4.72, 3.4.73
=
(−1)r [−rA + (n − r)A]
= (−1)r (n − 2r)A.
(3.4.185)
Geometric calculus 137
Proposition 3.4.75. For ~a = ~a(~x) , ~b = ~b(~x) : ∂~~x (~a · ~b) = ~a · ∂~~x~b + ~b · ∂~~x~a − ~a · (∂~~x ∧ ~b) − ~b · (∂~~x ∧ ~a).
(3.4.186)
Proof. Prop. 3.4.18, (3.3.41) ˙ ˙ ˙ ˙ ~a · (∂~~x ∧ ~b) = ~a · ∂~~x~b − ∂~~x (~b · ~a), (3.3.41) ~b · (∂~~x ∧ ~a) Prop. 3.4.18, ~b · ∂~˙~x~a˙ − ∂~˙~x (~a˙ · ~b), = addition ~ ˙ ˙ ˙ ⇒ ∂~x (~a · ~b) = ∂~~x (~a˙ · ~b) + ∂~~x (~b · ~a) = ~a · ∂~~x~b + ~b · ∂~~x~a − ~a · (∂~~x ∧ ~b) − ~b · (∂~~x ∧ ~a).
(3.4.187) (3.4.188)
(3.4.189)
Definition 3.4.76. Lie bracket. For ~a = ~a(~x), ~b = ~b(~x), the Lie bracket is defined as: [~a, ~b ] = ~a · ∂~~x ~b − ~b · ∂~~x ~a . (3.4.190) Proposition 3.4.77. For ~a = ~a(~x), ~b = ~b(~x), we get the following expression for the Lie bracket : [~a, ~b ] = ∂~~x · (~a ∧ ~b) − ~b∂~~x · ~a + ~a∂~~x · ~b. (3.4.191) Proof. Prop. 3.4.18, (3.3.41) ∂~~x · (~a ∧ ~b) = (∂~~x · ~a)~b − (∂~~x · ~b)~a = ~b(∂~~x · ~a) + (~a · ∂~~x )~b − ~a(∂~~x · ~b) − (~b · ∂~~x )~a Def. 3.4.76
= ~a · ∂~~x ~b − ~b · ∂~~x ~a = ∂~~x · (~a ∧ ~b) − ~b(∂~~x · ~a) + ~a(∂~~x · ~b).
(3.4.193)
Proposition 3.4.78. For ~a = ~a(~x), ~b = ~b(~x), ~c = ~c(~x) : ˙ ˙ (~c ∧ ~b) · (∂~~x ∧ ~a) = ~b · ∂~~x~a˙ · ~c − ~c · ∂~~x~a˙ · ~b = ~b · ∂~~x (~a · ~c) − ~c · ∂~~x (~a · ~b) + [~c, ~b] · ~a,
(3.4.194)
⇒
[~a, ~b ]
(3.4.192)
where the Lie bracket [~c, ~b] is defined according to Def. 3.4.76. Proof. (3.3.42) (~c ∧ ~b) · (∂~~x ∧ ~a) = ~c · ~b · (∂~~x ∧ ~a)
Proof 3.4.75
=
˙ ˙ ~c · ~b · ∂~~x~a˙ − ∂~~x (~a˙ · ~b))
˙ ˙ = ~b · ∂~~x~a˙ · ~c − ~c · ∂~~x~a˙ · ~b ˙ ˙ ˙ = ~b · ∂~~x (~a · ~c) − ~b · ∂~~x (~a · ~c˙) − ~c · ∂~~x (~a · ~b) + ~c · ∂~~x (~a · ~b) ˙ ˙ ˙ = ~b · ∂~~x (~a · ~c) − ~c · ∂~~x (~a · ~b) + ~c · ∂~~x (~b · ~a) − ~b · ∂~~x (~c˙ · ~a) Def. 3.4.76 ~
=
b · ∂~~x (~a · ~c) − ~c · ∂~~x (~a · ~b) + [~c, ~b] · ~a.
(3.4.195)
138 Quaternion and Clifford Fourier Transforms
Proposition 3.4.79. For ~a = ~a(~x), ~b = ~b(~x) : ˙ ˙ ˙ ˙ ~a · (∂~~x ∧ ~b) = ~b · (∂~~x ∧ ~a) + ∂~~x · (~a ∧ ~b) = (~a ∧ ∂~~x ) · ~b + ~a · ∂~~x~b − ~a∂~~x · ~b.
(3.4.196)
Proof. It follows from the Jacobi identity (3.3.44) that ˙ ˙ ˙ ˙ ~a · (∂~~x ∧ ~b) + ∂~~x · (~b ∧ ~a) + ~b · (~a ∧ ∂~~x ) = 0. ⇒
(3.4.197)
˙ ˙ ˙ ˙ ~a · (∂~~x ∧ ~b) = ~b · (∂~~x ∧ ~a) + ∂~~x · (~a ∧ ~b) (3.3.39),(3.3.41)
=
(~a ∧ ∂~~x ) · ~b + ~a · ∂~~x ~b − ~a ∂~~x · ~b.
(3.4.198)
Proposition 3.4.80. For ~a = ~a(~x), ~b = ~b(~x) : ˙ ˙ ˙ ˙ ˙ ˙ ∂~~x · (~a˙ ∧ ~b) = ~a˙ · (∂~~x ∧ ~b) − ~b · (∂~~x ∧ ~a˙ ) = (~b ∧ ∂~~x ) · ~a + ~a · (∂~~x ∧ ~b) − (~a ∧ ∂~~x ) · ~b − ~b · (∂~~x ∧ ~a).
(3.4.199)
Proof. It follows from the Jacobi identity (3.3.44) that ˙ ˙ ˙ ˙ ˙ ˙ ∂~~x · (~a˙ ∧ ~b) + ~a˙ · (~b ∧ ∂~~x ) + ~b · (∂~~x ∧ ~a˙ ) = 0. ⇒
(3.4.200)
˙ ˙ ˙ ˙ ˙ ˙ ∂~~x · (~a˙ ∧ ~b) = ~a˙ · (∂~~x ∧ ~b) − ~b · (∂~~x ∧ ~a˙ ) ˙ ˙ ˙ ˙ ˙ ˙ = ~a˙ · (∂~~x ∧ ~b) + ~a · (∂~~x ∧ ~b) − ~b · (∂~~x ∧ ~a) − ~b · (∂~~x ∧ ~a˙ ) (3.3.39)
=
(~b ∧ ∂~~x ) · ~a + ~a · (∂~~x ∧ ~b) − (~a ∧ ∂~~x ) · ~b − ~b · (∂~~x ∧ ~a).
(3.4.201)
Proposition 3.4.81. ˙ ˙ ˙ ˙ A × (∂~~x ∧ ~b) = A · ∂~~x~b − ∂~~x~b · A = A ∧ ∂~~x~b − ∂~~x~b ∧ A.
(3.4.202)
with the commutator product A × B of multivectors A, B, introduced in (3.3.45). Proof. (3.3.46) (3.3.45) ˙ ˙ ˙ ˙ A × (∂~~x ∧ ~b) = −(∂~~x ∧ ~b) × A = −∂~~x~b · A + A · ∂~~x~b (3.3.46) ˙ ˙ = −∂~~x~b ∧ A + A ∧ ∂~~x~b.
(3.4.203)
Geometric calculus 139
Proposition 3.4.82. For A = hAir = A(~x), B = hBis = B(~x) : ˙ A˙ ∧ ∂~~x ∧ B˙ = (−1)r ∂~~x ∧ (A ∧ B) = A ∧ ∂~~x ∧ B + (−1)r+s(r+1) B ∧ ∂~~x ∧ A.
(3.4.204)
Proof. ˙ A˙ ∧ ∂~~x ∧ B˙
(3.3.40)
=
(−1)r ∂~~x ∧ (A ∧ B),
(3.4.205)
˙ ˙ A˙ ∧ ∂~~x ∧ B˙ = A ∧ ∂~~x ∧ B + (A˙ ∧ ∂~~x ) ∧ B (3.3.40) ˙ ˙ ∧B = A ∧ ∂~~x ∧ B + (−1)r (∂~~x ∧ A) (3.3.40)
=
A ∧ ∂~~x ∧ B + (−1)r+s(r+1) B ∧ (∂~~x ∧ A).
(3.4.206)
The following properties and definitions are needed to develop the subject of Clifford Fourier transforms. Differentiating twice with the vector derivative, we get the differential Laplacian operator ∇2 . We can write ∇2 = ∇ · ∇ + ∇ ∧ ∇. But for integrable functions ∇ ∧ ∇ = 0. In this case we have ∇2 = ∇ · ∇. Proposition 3.4.83. (integration of parts) Z
g(x)[a · ∇h(x)]d3 x =
Z
R3
g(x)h(x)d2 x
R2
a·x=∞
−
Z
[a · ∇g(x)]h(x)d3 x
R3
a·x=−∞
We illustrate proposition 3.4.83 by inserting a = e3 , i.e. Z R3
3
g(x)[∂3 h(x)]d x =
x3 =∞
Z R2
g(x)h(x)dx1 dx2
− x3 =−∞
Z R3
[∂3 g(x)]h(x)d3 x,
which is nothing but the usual integration of parts formula for the partial derivative ∂3 h(x).
3.5
SUMMARY ON VECTOR DIFFERENTIAL CALCULUS
Section 3 first summarized important geometric algebra relationships, which are necessary for the thorough and explicit development of the vector differential calculus part of universal geometric calculus. It then showed how to differentiate multivector functions by a vector, including the results of standard vector analysis. The vector differential relationships are proven in a very explicit step by step way, enabling the reader, who is unfamiliar with the algebraic techniques to get complete comprehension. It may thus also serve as important reference material for studying and applying vector differential calculus in connection with Clifford Fourier transforms. A corresponding thorough treatment of multivector differential calculus can be found in [175].
CHAPTER
4
Quaternion Fourier transforms
We study quaternion Fourier transforms (QFT). The twosided QFT was introduced in [124] for the analysis of twodimensional (2D) linear timeinvariant partialdifferential systems. In further theoretical investigations [184, 187] a special split of quaternions was introduced, then called ±split. We develop this split further, interpret it geometrically as orthogonal 2D planes split (OPS), and generalize it to a freely steerable split of H into two orthogonal 2D analysis planes. The new general form of the OPS split allows us to find new geometric interpretations for the action of the QFT on the signal. The second major result we present is a variety of new steerable forms of the QFT, their geometric interpretation, and for each form OPS split theorems, which allow fast and efficient numerical implementation with standard FFT software. We continue by studying properties of the QFT, like the uncertainty principle, convolution, and a Wiener Khinchine type theorem for the QFT. We also show how QFTs can be further specialized to a windowed QFT, a quaternionic FourierMellin transform, and generalized to a quaternion domain FT, a volumetime FT, and a spacetime FT.
4.1 4.1.1
FUNDAMENTALS OF QUATERNION FOURIER TRANSFORMS (QFT) Quaternion Fourier transform and the ±split
We first treat the quaternionic Fourier transform (QFT) applied to quaternion fields and investigate QFT properties useful for applications. Different forms of the QFT lead us to different Plancherel theorems. We relate the QFT computation for quaternion fields to the QFT of real signals. We study the general linear (GL) transformation behavior of the QFT with matrices, Clifford geometric algebra and with examples. (Later, in Section 4.4 we reach wideranging noncommutative multivector FT generalizations of the QFT. Examples are the new volumetime and spacetime algebra Fourier transformations.) The content of this section is mainly based on [184].
DOI: 10.1201/97810031844784
141
142 Quaternion and Clifford Fourier Transforms
4.1.1.1 Background
We strive to deepen the understanding of the quaternionic Fourier transform (QFT) applied to quaternion fields f : R2 → H, and not only to real signals f : R2 → R. We present QFT properties useful for applications to partial differential equations, image processing and optimized numerical implementations. We show how different forms of the QFT allow to establish different scalar and quaternion valued Plancherel theorems. We explain how to reduce the computation for quaternion fields to the case of real signal computations, and on the other hand how results for real signals can be generalized to quaternion fields. The third major focus of this section is on deriving the behavior of the QFT under GL(R2 ) automorphisms. To do this we split the QFT appropriately, and work with invariant techniques of Clifford geometric algebra [161] to establish and understand the automorphism behavior. Details are brought to light by looking at the examples of stretches (dilations), reflections and rotations. Together with isomorphisms (to Clifford subalgebras) we finally arrive at wideranging generalizations of the QFT. These new noncommutative multivector Fourier transforms operate on functions from domain spaces Rm,n (with m, n ∈ N0 ) to Clifford algebras Cl(m, n) or subalgebras thereof. To practically demonstrate the method, we work out generalizations to volumetime and to spacetime algebra Fourier transformations, and provide some physical interpretation in Section 4.4. Quaternions were introduced in Section 2.8.1, part of Section 2.8 on the geometric meaning of quaternions. We remind the reader (compare Section 2.8.2) that in some applications it proves convenient to replace k with k = ij and write a quaternion as 4.1.1.1.1
Quaternions, split and module
q = qr + iqi + qj j + iqk j,
(4.1.1)
neatly keeping all i to the left and all j to the right of each term. A second convenient form is the split q = q+ + q− ,
1 q± = (q ± iqj). 2
(4.1.2)
Explicitly in real components qr , qi , qj , qk ∈ R using (2.8.1) the split (4.1.2) produces: q± = {qr ± qk + i(qi ∓ qj )}
1±k 1±k = {qr ± qk + j(qj ∓ qi )}. 2 2
(4.1.3)
For quaternionvalued functions f, g : R2 → H we can define the quaternionvalued inner product Z
(f, g) =
f (x) g˜(x) d2 x ,
with
d2 x = dxdy,
(4.1.4)
R2
with symmetric real scalar part [56] 1 hf, gi = [(f, g) + (g, f )] = 2
Z R2
hf (x) g˜(x)i0 d2 x .
(4.1.5)
Quaternion Fourier transforms 143
Both (4.1.4) and (4.1.5) lead to the L2 (R2 ; H)norm kf k =
q
(f, f ) =
q
hf, f i =
Z
f (x)2 d2 x .
(4.1.6)
R2
A quaternion module L2 (R2 ; H) is then defined as L2 (R2 ; H) = {f f : R2 → H, kf k < ∞}.
(4.1.7)
4.1.1.2 The quaternion Fourier transform
Before defining the quaternion Fourier transform (QFT), we briefly outline its relationship with Clifford Fourier transformations, see also Chapter 1. Brackx et al. [33] extended the Fourier transform to multivector valued functiondistributions in Cl0,n with compact support. A related applied approach for hypercomplex Clifford Fourier transformations1 in Cl0,n was followed by B¨ ulow et. al. [58]. By extending the classical trigonometric exponential function exp(j x∗ξ) (where ∗ denotes the scalar product of x ∈ Rm with ξ ∈ Rm , j the imaginary unit) in [243,259], McIntosh et. al. generalized the classical Fourier transform. Applied to a function of m real variables this generalized Fourier transform is holomorphic in m complex variables and its inverse is monogenic in m + 1 real variables, thereby effectively extending the function of m real variables to a monogenic function of m + 1 real variables (with values in a complex Clifford algebra). This generalization has significant applications to harmonic analysis, especially to singular integrals on surfaces in Rm+1 . Based on this approach Kou and Qian obtained a Clifford PayleyWiener theorem and derived Shannon interpolation of bandlimited functions using the monogenic sinc function [286, and references therein]. The Clifford PayleyWiener theorem also allows to derive leftentire (leftmonogenic in the whole Rm+1 ) functions from square integrable functions on Rm with compact support. The real ndimensional volume element in = e1 e2 . . . en of Cln,0 over the field of the reals R has been used in [117, 182, 185, 251] to construct and apply Clifford Fourier transformations for n = 2, 3 (mod 4) with kernels exp(−in x ∗ u), x, u ∈ Rn . This in has a clear geometric interpretation. Note that i2n = −1 for n = 2, 3 (mod 4). Ell [124] defined the quaternion Fourier transform (QFT) for application to 2D linear timeinvariant systems of PDEs. Ell’s QFT belongs to the growing family of Clifford Fourier transformations because of (2.8.3). But the left and right placement of the exponential factors in Definition 4.1.1 distinguishes it. Later the QFT was applied extensively to 2D image processing, including color images [56, 58, 124]. This spurred research into optimized numerical implementations [132, 280]. Ell [124] and others [56,69] also investigated related commutative hypercomplex Fourier transforms like in the commutative subalgebra of Cl4,0 with subalgebra basis {1, e 12 , e 34 , e 1234 }, e 212 = e 234 = −1,
1
e 21234 = +1.
(4.1.8)
This is the kind of Clifford Fourier transform to which we will refer in Section 4.1.1.3.
144 Quaternion and Clifford Fourier Transforms
Definition 4.1.1 (Quaternion Fourier transform (QFT)). The quaternion Fourier transform2 fˆ : R2 → H of f ∈ L2 (R2 ; H), x = xe1 +ye2 ∈ R2 and u = ue1 +ve2 ∈ R2 is defined3 as Z ˆ f (u) = e−ixu f (x) e−jyv d2 x . (4.1.9) R2
The QFT can be inverted by f (x) =
1 (2π)2
Z
eixu fˆ(u) ejyv d2 u ,
(4.1.10)
R2
with d2 u = dudv. Let f : R2 → H ( or f ∈ L2 (R2 ; H) ). Using four R → R ( or L (R ; R) ) real component functions fr , fi , fj and fk we can decompose and rewrite f with (4.1.1) as
4.1.1.2.1
Rewriting and splitting functions 2
2
2
f = fr + fi i + fj j + fk k = fr + ifi + fj j + ifk j.
(4.1.11)
We can also split the functions f [similar to q± in (4.1.2)] into f = f+ + f− ,
1 f+ = (f + if j), 2
1 f− = (f − if j). 2
(4.1.12)
According to (4.1.3) the two components f± can also be rewritten as f± = {fr ± fk + i(fi ∓ fj )}
1±k 1±k = {fr ± fk + j(fj ∓ fi )}. 2 2
(4.1.13)
As an example, let us consider the split of the product of exponential functions under the QFT integral in (4.1.9). Using Euler’s formula and trigonometric addition theorems the split leads to K = e−ixu e−jyv = K+ + K− , 1 ± k −j(yv∓xu) 1±k K± = e−i(xu∓yv) = e . 2 2
2
R
(4.1.14)
We also assume always that R2 f (x) d2 x exists as well. But we do not explicitly write this condition again in the rest of the paper. Strictly speaking, the integral definition of Def. 4.1.1 works T only 1 2 2 2 for f ∈ L1 (R2 ; H). But one can first define the QFT on the dense subset L (R ; H) L (R ; H), T 2 2 1 2 and then use the continuity of the Fourier transform on L (R ; H) L (R ; H), due to Plancherel’s theorem Theorem 4.1.2, to define the QFT on L2 (R2 ; H), see e.g. [135]. A similar argument applies for all quaternionic and Clifford Fourier transforms considered in this monograph. 3 For real signals f ∈ L2 (R2 ; R) the detailed relationship of the QFT of definition 4.1.1 with the conventional scalar FT, i.e. with the even cospart and the odd sinpart are given on pp. 191 and 192 of [58]. With the help of (4.1.24) this can easily be extended to the full QFT of quaternionvalued f ∈ L2 (R2 ; H).
Quaternion Fourier transforms 145
Properties of the quaternion Fourier transform (QFT) of quaternion functions (Quat. Funct.) f, g ∈ L2 (R2 ; H), with x, u ∈ R2 , constants α, β ∈ {q q = qr + qi i, qr , qi ∈ R}, α 0 , β 0 ∈ {q q = qr + qj j, qr , qj ∈ R}, a, b ∈ R \ {0}, x 0 = x0 e 1 + y0 e 2 , u 0 = u0 e 1 + v0 e 2 ∈ R2 and m, n ∈ N0 . Source: [184, Tab. 1]. Table 4.1
Property Left linearity Right linearity xShift Modulation Dilation4 Part. deriv. Powers5 of x, y Powers5 of i, j Plancherel5 Parseval6
4.1.1.2.2
Quat. Funct. αf (x)+β g(x) f (x)α 0 + g(x)β 0 f (x − x 0 ) eixu0 f (x) ejyv0 f (a xe 1 + b ye 2 ) ∂ m+n ∂xm ∂y n f (x) xm y n f (x) im f (x) j n hf, gi = kf k =
QFT αfˆ(u)+ βˆ g (u) 0 ˆ f (u)α + gˆ(u)β 0 e−ix0 u fˆ(u) e−jy0 v fˆ(u − u 0 ) v 1 ˆ u ab f ( a e 1 + b e 2 ) (iu)m fˆ(u)(jv)n m+n im ∂u∂ m ∂vn fˆ(u) j n im fˆ(u) j n 1 ˆ ˆi (2π)2 hf , g 1 ˆ 2π kf k
We first show a new Plancherel theorem with
Useful properties of the QFT
respect to the scalar product (4.1.5). Theorem 4.1.2 (QFT Plancherel). The scalar product (4.1.5) of two quaternion module functions f, g ∈ L2 (R2 ; H) is given by the scalar product of the of the corresponding QFTs fˆ and gˆ 1 hf, gi = hfˆ, gˆi. (4.1.15) (2π)2 Proof. For f, g ∈ L2 (R2 ; H) we calculate the scalar product (4.1.5) hf, gi =
Z R2
hf (x)˜ g (x)i0 d2 x
1 h eiux fˆ(u)ejvy d2 u g˜(x)i0 d2 x (2π)2 R2 R2 Z Z 1 ˆ(u) = h f ejvy g˜(x)eiux d2 xi0 d2 u 2 (2π)2 R2 R Z Z 1 hfˆ(u)[ e−iux g(x)e−jvy d2 x]∼ i0 d2 u = 2 (2π)2 R2 R Z 1 1 = hfˆ(u)g˜ˆ(u)i0 d2 u = hfˆ, gˆi. (2π)2 R2 (2π)2 Z
Z
=
4
B¨ ulow [56] omits the absolute value signs for the determinant of the transformation. Theorems 4.1.4, 4.1.5 and 4.1.2. 6 Corollary 4.1.3. 5
(4.1.16)
146 Quaternion and Clifford Fourier Transforms
In the second equality of (4.1.16) we replaced f with its inverse QFT expression (4.1.10). In the third equality we exchanged the order of integration and we used the cyclic symmetry (2.8.14). For the fourth equality we simply pulled the reversion outside the square brackets [. . .] and obtained the QFT gˆ(u), which proves (4.1.15) according to (4.1.5). For g = f the Plancherel theorem 4.1.2 has a QFT Parseval theorem (also called Rayleigh’s theorem) as a direct corollary. Corollary 4.1.3 (QFT Parseval). The L2 (R2 ; H)norm of a quaternion module function f ∈ L2 (R2 ; H) is given by the L2 (R2 ; H)norm of its QFT multiplied by 1/(2π) 1 ˆ kf k. 2π This leads to the following observations: kf k =
(4.1.17)
• The way we obtained the Parseval theorem of cor. 4.1.3 is much simpler than the proofs in [56, 124]. • For twodimensional linear timeinvariant partial differential systems the Parseval theorem provides an appropriate method to measure controller performance. • In signal processing it states that the signal energy is preserved by the QFT. For solving PDEs with quaternionic (or real) coefficient polynomials in x, y ∈ R2 we show the following two theorems. In this context we note again that every quaternionic (or real) coefficient polynomial in the variables x, y ∈ R2 can be brought into a form having factors of i ∈ H to the left side of each term and factors of j ∈ H to the right side of each term (compare (4.1.11)). Theorem 4.1.4 (Powers of x, y). The QFT of a quaternion module function xm y n f (x) ∈ L2 (R2 ; H), x = xe1 + ye2 ∈ R2 , f ∈ L2 (R2 ; H), m, n ∈ N0 is given by ∂ m+n ˆ m y nf (u) = i m x\ f (u) j n . (4.1.18) ∂um ∂v n Proof. The proof is done by induction. It is trivial for m = n = 0. c according to (4.1.9) For m = 1, n = 0 we calculate the QFT of xf Z c (u) = xf
∂ −ixu e f (x) e−jyv d2 x ∂u Z ∂ ∂ e−ixu f (x) e−jyv d2 x = i fˆ(u). =i ∂u R2 ∂u −ixu
e
−jyv 2
xf (x) e
R2
Z
d x=
i
R2
(4.1.19)
∂ −ixu In second equality we used ∂u e = −ixe−ixu and i(−i) = 1. Completely analogous for m = 0, n = 1 we find
c (u) = yf
∂ ∂v
Z R2
e−ixu f (x) e−jyv d2 x j =
∂ ˆ f (u)j . ∂v
(4.1.20)
Because of noncommutativity j appears to the right of fˆ. Induction over m, n ∈ N completes the proof.
Quaternion Fourier transforms 147
Theorem 4.1.5 (Powers of i, j ). The QFT of a quaternion module function im f (x)j n ∈ L2 (R2 ; H), f ∈ L2 (R2 ; H), m, n ∈ N0 is given by m i\ f j n (u) = im fˆ(u) j n .
(4.1.21)
Proof. Similar to the left and right linearities of Table 4.1, theorem 4.1.5 follows directly from the definition 4.1.1 of the QFT, using the commutation relationships exp(−ixu)im = im exp(−ixu) and
exp(−jyv)j n
= j n exp(−jyv).
(4.1.22)
For every f ∈ L2 (R2 ; H) we can always rewrite f = fr + fi i + fj j + fk k as in (4.1.11) to the form f = fr + ifi + fj j + ifk j. (4.1.23) Accordingly we now can make the following two important observations: • Theorem 4.1.5 reduces the computation of the QFT of any f ∈ L2 (R2 ; H) to the computation of four QFTs of the real functions fr , fi , fj , fk ∈ L2 (R2 ; R) as in fˆ = fˆr + ifˆi + fˆj j + ifˆk j. (4.1.24) • On the other hand theorem 4.1.5 reveals that every theorem for the QFT of real functions g ∈ L2 (R2 ; R) immediately results via (4.1.24) in a corresponding theorem for quaternion module functions f ∈ L2 (R2 ; H). We simply need to apply the theorem for the QFT of real functions to each of the four real component functions fr , fi , fj , fk ∈ L2 (R2 ; R). This fact is rather useful, because often in image processing theorems are only established for real image signals [56]. Example: GL(R2 ) transformation properties of the QFT
To give an example for the second observation at the end of Section 4.1.1.2.2 we use it to generalize the general linear real nonsingular transformation property of the QFT of real 2D functions f ∈ L2 (R2 ; R) of [56] to quaternion module functions f ∈ L2 (R2 ; H). This property of real 2D signals states that for 4.1.1.2.3
x 0 = Ax = (ax + by)e 1 + (cx + dy)e 2
(4.1.25)
with nonsingular real transformation matrix A=
a b c d
!
(4.1.26)
the QFT of a real signal f : R2 → R is7 n o  det B ˆ f\ (Ax)(u) = f (B+ u) + fˆ(B− u) + i fˆ(B+ u) − fˆ(B− u) j . 2 7
B¨ ulow [56] omits the absolute value signs for the determinant of the transformation.
(4.1.27)
148 Quaternion and Clifford Fourier Transforms
In (4.1.27) the two linear real nonsingular transformations B+ and B− have corresponding matrices and the (same) determinant B+ = A
−1 T
,
1 B− = det A
d c b a
!
,
det B = det B+ = det B− = (det A)−1 .
(4.1.28)
We can now establish the generalization from f ∈ L2 (R2 ; R) to f ∈ L2 (R2 ; H) functions. Theorem 4.1.6. The QFT of a quaternionmodule function f ∈ L2 (R2 ; H) with a GL(R2 ) transformation A of its vector arguments (4.1.25) is also given by (4.1.27). Proof. We only sketch the proof, because writing out all expressions explicitly would consume too much space: • Applying (4.1.27) and (4.1.28) to each component of (4.1.24) and • Rearranging the sum (of 16 terms) yields the validity of (4.1.27) together with (4.1.28) also for quaternionvalued f ∈ L2 (R2 ; H). • It is again crucial that in each term all factors i are always kept to the left and all factors j are always kept to the right.
We remark that resorting to matrices and matrix manipulations is geometrically not very intuitive, so in Section 4.1.1.4 an alternative more geometric approach is taken to derive the transformation properties of general f ∈ L2 (R2 ; H). This geometric approach has far reaching consequences for the generalization of the QFT, exploited in later sections. But before geometrically reanalyzing QFT transformation properties we look at the following variant of the QFT with some desirable properties not valid for the QFT of Definition 4.1.1. 4.1.1.3 The rightside quaternion Fourier transform (QFTr)
We observe that it is not possible to establish a general Plancherel theorem for the QFT of the inner product (f, g) of (4.1.4), because the product (4.1.4) lacks the cyclic symmetry (2.8.14) applied in the proof of theorem 4.1.2. To obtain a Plancherel theorem it is therefore either necessary to modify the symmetry properties of the inner product as in (4.1.5) or to modify the QFT itself. In this section we explore the second possibility. Definition 4.1.7 (Rightside QFT (QFTr)). The rightside quaternion Fourier .
transform8 f : R2 → H of f ∈ L2 (R2 ; H), x = xe1 +ye2 ∈ R2 and u = ue1 +ve2 ∈ R2 is defined as .
Z
f (u) =
f (x) e−ixu e−jyv d2 x
with d2 x = dxdy.
R2 8
For the rightside QFT applied to L1 and L2 functions, see Footnote 2 on page 144.
(4.1.29)
Quaternion Fourier transforms 149
Properties of the right sided quaternion Fourier transform (QFTr) of quaternion functions (Quat. Funct.) f, g ∈ L2 (R2 ; H), with x, u ∈ R2 , constants α, β ∈ R, α 0 , β 0 ∈ H, a, b ∈ R \ {0}, x 0 = x0 e 1 + y0 e 2 , u 0 = u0 e 1 + v0 e 2 ∈ R2 and m, n ∈ N. Source: [184, Tab. 2]. Table 4.2
Property
Quat. Funct.
QFTr
Linearity9
αf (x)+β g(x)
αf (u)+ β g(u)
Left linearity xShift10
α 0 f (x)+ β 0 g(x) f (x − x 0 )
α 0 f (u)+ β 0 g(u) F. {f e−ix0 u }(u) e−jy0 v
Dilation
f (a xe 1 + b ye 2 )
Part. deriv.11
u f (u)(jv)n
Powers12 of x, y
−m ∂ m+n ∂xm ∂y n f (x)i −m m n
. u 1 ab f ( a e 1 . m
x y f (x)i
Powers13 of i, j
im j n f (x)
i j f (u)
.
14
Plancherel
(f, g) =
Plancherel15
hf, gi =
Parseval
kf k =
.
.
.
+ vb e 2 )
. n ∂ m+n ∂um ∂v n f (u) j . m n . . 1 (2π)2 (f , g) . . 1 (2π)2 hf , gi . 1 2π kf k
The QFTr is known as Clifford Fourier transform [33, 58], because of the isomorphism H ∼ = Cl0,2 . Further freedoms in alternative definitions would be to exchange the order of the exponentials in (4.1.29) or to wholly shift both exponential factors to the left side instead. The former would simply exchange the roles of i and j , but the latter would not serve our purpose as will soon become clear. The QFTr can be inverted [33, 58] using f (x) =
1 (2π)2
Z
.
f (u) ejyv eixu d2 u,
(4.1.30)
R2
with d2 u = dudv. Attention needs to be paid to the reversed order of the exponential factors in (4.1.30) compared to (4.1.29). For general f, g ∈ L2 (R2 ; H) left linearity and dilation properties of Table 4.1 hold. The left linearity coefficients can now be fully quaternionic constants α0 , β 0 ∈ H. 4.1.1.3.1
9 10
Properties of the QFTr
The positions of the real scalars α, β before or after the functions f, g do not matter. Only for quaternion module functions f ∈ L2 (R2 ; H) with if = f i, i.e. f = fr + .
ifi with fr , fi ∈ L2 (R2 ; R) do we get F. {f (x − x 0 )}(u) = e−ix0 u f (u) e−j y0 v . 11
m+n
.
m n ∂ Only for if = f i do we get F. { ∂x m ∂y n f }(u) = (iu) f (u)(jv) . m+n
.
n ∂ Only for if = f i do we get F. {xm y n f }(u) = im ∂u m ∂v n f (u) j . 13 Here the powers of i, j law is a direct consequence of the left linearity. 14 Compare theorem 4.1.8. 15 A direct consequence of symmetrizing theorem 4.1.8. 12
150 Quaternion and Clifford Fourier Transforms
But xshift, partial derivative, and powers of xm y n properties need to be modified as in Table 4.2. Regarding (2.8.1) it is clear that if = f i holds iff f = fr +fi i, fr , fi ∈ R, which is slightly more general than the restriction of [56] to f = fr ∈ R. A modulation property analogous to the one in Table 4.1 does not hold. It is obstructed by the noncommutativity of the exponential factors exp(jyv0 ) exp(ixu) = 6 exp(ixu) exp(jyv0 ).
(4.1.31)
For a powers of i, j property to hold for the QFTr, we need to shift the factors j n also to the left of the quaternion function f (x). For fully general quaternionvalued f, g ∈ L2 (R2 ; H) we can establish for the QFTr the following quaternionvalued Plancherel theorem based on the inner product (4.1.4). Theorem 4.1.8 (QFTr Plancherel). The (quaternionvalued) inner product (4.1.4) of two quaternion module functions f, g ∈ L2 (R2 ; H) is given by the inner product of .
.
the corresponding QFTrs f and g (f, g) =
1 . . (f , g). (2π)2
(4.1.32)
Proof. For f, g ∈ L2 (R2 ; H) we calculate the inner product (4.1.4) Z
(f, g) =
f (x)˜ g (x)d2 x
R2 . 1 f (u)ejvy eiux d2 u g˜(x)d2 x 2 (2π) R2 R2 Z . Z 1 f (u) ejvy eiux g˜(x)d2 xd2 u = 2 (2π)2 R2 R Z Z . 1 f (u)[ g(x)e−iux e−jvy d2 x]∼ d2 u = 2 (2π)2 R2 R Z . 1 1 . . ˜. 2 = f (u) g(u)d u = (f , g). (2π)2 R2 (2π)2
Z
Z
=
(4.1.33)
In the second equality of (4.1.33) we replaced f with its inverse QFTr expression (4.1.30). In the third equality we exchanged the order of integration. For the fourth equality we simply pulled the reversion outside the square brackets [. . .] and obtained . the QFTr g(u), which proves (4.1.32) according to (4.1.4).
Quaternion Fourier transforms 151
For g = f theorem 4.1.8 has a corresponding QFTr Parseval theorem as a direct corollary. Corollary 4.1.9 (QFTr Parseval). The L2 (R2 ; H)norm of a quaternion module .
function f ∈ L2 (R2 ; H) is given by the L2 (R2 ; H)norm of its QFTr f multiplied by 1/(2π) 1 . 1 ˆ kf k = kf k = kf k. (4.1.34) 2π 2π Proof. The first identity follows from setting g = f in theorem 4.1.8 (QFTr Plancherel). The second identity follows from comparing with corollary 4.1.3 (QFT Parseval). To facilitate the use of the QFTr and comparison with the QFT (Table 4.1) we list the main QFTr properties in Table 4.2. 4.1.1.4 Understanding the GL(R2 ) transformation properties of the QFT
We begin with noting that the matrix transformation law (4.1.27), derived by B¨ ulow 2 2 [56] for real signals f ∈ L (R ; R), and generalized in Theorem 4.1.6 of Section 4.1.1.2.3 to quaternionvalued signals16 f ∈ L2 (R2 ; H), with four terms on the right side, allows no straightforward geometric interpretation. Yet a clear geometric interpretation is not only needed in many applications, such an interpretation is also very instructive in order to successfully generalize the QFT to higher dimensions. Toward this aim we observe that the split (4.1.14) of the exponentials K under the QFT integral results in two (single exponential) complex kernels K± with complex units i (or j ) apart from the right (or left) factor (1 ± k)/2. This and the known elegant monomial transformation properties of complex Fourier transforms (also preserved in the Clifford FT of [251]) motivates us to geometrically reanalyze the GL(R2 ) transformation properties of the QFT of f ∈ L2 (R2 ; H) in terms of its two components f± as given in (4.1.12). Theorem 4.1.10 (QFT of f± ). The QFT of the f± split parts of a quaternion module function f ∈ L2 (R2 , H) have the complex forms fˆ± =
16
Z R2
f± e−j(yv∓xu) d2 x =
Z R2
e−i(xu∓yv) f± d2 x .
(4.1.35)
Remember that B¨ ulow [56] proved his transformation law only for real signals. But in Theorem 4.1.6 of Section 4.1.1.2.3, we used (4.1.24) and Theorem 4.1.5 to generalize from real signals f ∈ L2 (R2 , R) to quaternion valued signals f ∈ L2 (R2 , H).
152 Quaternion and Clifford Fourier Transforms
Proof. 1 ± k −jyv 2 e d x 2 R2 Z 1 ± k −jyv 2 = {fr ± fk + i(fi ∓ fj )} e−ixu e d x 2 R2 Z 1 ± k −j(yv∓xu) 2 d x = e {fr ± fk + i(fi ∓ fj )} R2  2 {z }
fˆ± =
Z
e−ixu {fr ± fk + i(fi ∓ fj )}
=K±
(4.1.13)
Z
=
R2
f± e−j(yv∓xu) d2 x =
Z R2
e−i(xu∓yv) f± d2 x ,
(4.1.36)
where for the third equality we did a number of quaternion algebra manipulations, involving Euler’s formula and trigonometric addition theorems. The last equality of (4.1.36) follows analogously by replacing f± with the third expression in (4.1.13), etc. We learn from the third line of (4.1.36) that the behavior of the two parts (4.1.14) under automorphisms A ∈ GL(R2 ) also determines the automorphism properties of the QFTs fˆ± , where due to theorem 4.1.5 the QFT operation and the split operation (4.1.12) commute. 4.1.1.4.1 Geometric interpretation and coordinate independent formulation of GL(R2 ) transformations of the QFT We begin with noting that according to the polar decom
position theorem [162] every automorphism A ∈ GL(R2 ) has a unique decomposition A = T R = RS, where R is a rotation and T and S are symmetric with positive and negative eigenvalues. Positive eigenvalues correspond to stretches by the eigenvalue in the direction of the eigenvector. Negative eigenvalues correspond to reflections at the line (hyperplane) normal to the eigenvector, composed with stretches by the absolute value of the eigenvalue in the direction of the eigenvector. Stretches (positive eigenvalues) D ∈ GL(R2 ) were already fully treated in [56] (compare also Table 4.1). Rotations correspond to two reflections [68,85] at lines subtending half the angle of the resulting rotation Rab = Ua Ub . The elementary transformations that compose all automorphisms A ∈ GL(R2 ) are therefore stretches and reflections. In geometric algebra reflections Un at a hyperplane (line in 2D) through the origin can be characterized by normal vectors n Un x = −n −1 xn.
(4.1.37)
The length of n does not matter. Un preserves (reverses) the component parallel (perpendicular) to the hyperplane of reflection. With the vectors x = xe 1 + ye 2 , u = ue 1 + ve 2 we now rewrite coordinate free17 17
The fact that the reflection Ue 1 with the special hyperplane normal to vector e 1 is needed stems from the arbitrary initial association of the e 1 coordinate product xu with i and of the e 2 coordinate product yv with j .
Quaternion Fourier transforms 153
the angles in the exponentials of fˆ± as −xu + yv = x · (Ue 1 u),
xu + yv = x · u.
(4.1.38)
Hence we get for the QFTs of f± fˆ+ =
Z R2
f+ e−j x ·(Ue 1 u ) d2 x,
fˆ− =
Z
f− e−j x ·u d2 x.
R2
(4.1.39)
The QFT of f− is analogous to a complex 2D Fourier transform, only in general f− and the exponential factor do not commute. The QFT of f+ is similar except for the reflection Ue 1 . We are now in a position to apply any automorphism A ∈ GL(R2 ) to the spatial argument of the f± components of any f ∈ L2 (R2 , H). We begin with \ f− (Ax)(u) =
Z
f− (Ax)e−j x ·u d2 x
R2
=  det A−1 
Z R2
f− (z)e−j z
·(A−1
z =Ax =
Z R2
−1 f− (z)e−j(A z )·u  det A−1 d2 z
u ) d2 z =  det A−1  fˆ (A−1 u), −
(4.1.40)
where A−1 indicates the adjoint automorphism of A−1 . The absolute value of the determinant det A−1 needs to be used, because of the interchange of integration boundaries for a negative determinant. We continue with Z
\ f+ (Ax)(u) = z =Ax =
Z R2
R2
f+ (Ax)e−j x ·(Ue 1 u ) d2 x
−1 f+ (z)e−j(A z )·(Ue 1 u )  det A−1 d2 z
−1
=  det A 
Z R2
=  det A−1 
Z R2
f+ (z)e−j z ·(A
−1 U
e 1 u ) d2 z
f+ (z)e−j z ·(Ue 1 Ue 1 A
−1 U
e 1 u ) d2 z
=  det A−1  fˆ+ (Ue 1 A−1 Ue 1 u),
(4.1.41)
which is very similar to the previous calculation for fˆ− . The only difference is that in line 4 we insert 1 = Ue 1 Ue 1 before A−1 , and that the argument of the transformed fˆ+ now has the reflected version Ue 1 A−1 Ue 1 of the adjoint inverse transformation A−1 . Recombining fˆ+ and fˆ− we get from (4.1.40) and (4.1.41) Theorem 4.1.11 (GL(R2 ) transformation properties of the QFT). The QFT of a quaternion module function f ∈ L2 (R2 ; H) with a GL(R2 ) transformation A of its vector argument is given by f\ (Ax)(u) =  det A−1  { fˆ− (A−1 u) + fˆ+ (Ue1 A−1 Ue1 u) } .
(4.1.42)
Theorem 4.1.11 corresponds exactly to equation (4.1.27) with (4.1.28), if the matrix expression (4.1.26) is used for the automorphism A and if the f± split formulas (4.1.12) are used. The four terms of (4.1.27) together with all the matrices involved therefore get in theorem 4.1.11 a clear geometric interpretation. In order to be even more explicit we specify below the full geometric algebra expressions for stretches, reflections and rotations.
154 Quaternion and Clifford Fourier Transforms
4.1.1.4.2 Explicit examples: stretches, reflections & rotations To deepen our geometrical understanding we now look at stretches, reflections (and rotations) which compose every general automorphism A ∈ GL(R2 ). Stretches expressed by As x = axe 1 + bye 2 , with a, b ∈ R \ {0}, result because of Ue 1 As Ue 1 = As in
v 1 ˆu ˆ −1 f\ (As x)(u) =  det A−1 f ( e 1 + e 2 ). s  f (As u) = ab a b
(4.1.43)
Reflections in hyperplanes normal to a expressed by Ua x = −a −1 xa , with  det Ua  = 1 , Ua = Ua , Ue 1 Ua Ue 1 = Ua 0 and a 0 = Ue 1 a result in f\ (Ua x)(u) = fˆ− (Ua u) + fˆ+ (Ua 0 u).
(4.1.44)
Finally rotations (equivalent to two reflections at lines subtending half the rotation angle) expressed by Rab x = Ub Ua x , with  det Rab  = 1 , R−1 ab = Rba , −1 0 0 and Ua Ub = Ue 1 Rab Ue 1 , result in \x)(u) = fˆ(U Ua x)(u) = fˆ− (Ua U u) + fˆ+ (Ua 0 U 0 u) f (R ab b b b ˆ+ (Ue R−1 Ue u) = fˆ− (R−1 u) + f 1 ab ab 1
(4.1.45)
In two dimensions18 the formula for rotations of the spatial argument of a quaternion module function f subject to the QFT can be further simplified to \x)(u) in=2D fˆ− (R−1 u) + fˆ+ (R u) , f (R ab ab ab
(4.1.46)
because in two dimensions we have Ue 1 R−1 ab Ue 1 = Rab . Theorems 4.1.10 and 4.1.11 together with their clear geometric interpretation with the help of geometric algebra pave the way for wideranging generalizations of the QFT of Definition 4.1.1. In this book we cannot fully treat all possible generalizations. But in order to demonstrate the method, we show in the Section 4.4 how to generalize the QFT to a new general noncommutative Fourier transformation of functions from spacetime R3,1 to the spacetime algebra [158] of R3,1 , i.e. to the Clifford geometric algebra Cl(3, 1). An intermediate step will be the generalization to a new Fourier transform of functions from spacetime R3,1 to a volumetime subalgebra of the spacetime algebra. 4.1.1.5 Summary on QFT and the ±split
We employed a convenient rewriting of quaternions only in terms of i and j , keeping one to the left and the other to the right; and a quaternion split, which in spacetime applications (Section 4.4) is closely related the choice of the time direction. This allowed us to investigate a range of properties of the QFT, last but not least the behavior of the QFT under general linear automorphisms. 18 In Section 4.4, we generalize theorem 4.1.11 to higher dimensions, but for rotations the expression for fˆ+ on the right hand side of (4.1.46) will in general not be valid for higher dimensions.
Quaternion Fourier transforms 155
General coordinate free formulation in combination with quaternion to Clifford subalgebra isomorphisms opens the door to a wide range of QFT generalizations to these Clifford geometric algebras, studied further in Section 4.4. 4.1.2 QFT and orthogonal 2D planes split of quaternions
The current treatment is based on [194, 205]. 4.1.2.1 Background
The twosided quaternionic Fourier transformation (QFT) was introduced in [124] for the analysis of 2D linear timeinvariant partialdifferential systems. Subsequently it has been applied in many fields, including colour image processing [291]. This led to further theoretical investigations [184, 187], where a special split of quaternions was introduced, then called ±split, see Sections 2.8.2, and 4.1.1. An interesting physical consequence was that this split resulted a in left and right traveling multivector wave packet analysis, when generalizing the QFT to a full spacetime Fourier transform (SFT), see Section 4.4. In the current section we investigate this split further, interpret it geometrically and generalize it to a freely steerable split of H into two orthogonal 2D analysis planes. For reasons to become obvious we prefer to call it from now on the orthogonal 2D planes split (OPS). The general form of the OPS split allows us to find new geometric interpretations for the action of the QFT on the signal. The second major result of this work are a variety of new forms of the QFT, their detailed geometric interpretation, and for each form, OPS split theorems, which allow fast and efficient numerical implementation with standard FFT software. A preliminary formal investigation of these new OPSQFTs can also be found in [194]. This section is organized as follows. We first introduce in Section 4.1.2.2 several properties of quaternions together with a brief review of the ±split of [184, 187]. In Section 4.1.2.3, we generalize this split to a freely steerable orthogonal 2D planes split (OPS) of quaternions H. In Section 4.1.2.4, we use the general OPS of Section 4.1.2.3 to generalize the two sided QFT to a new two sided QFT with freely steerable analysis planes, complete with a detailed local geometric transformation interpretation. The geometric interpretation of the OPS in Section 4.1.2.3 further allows the construction of a new type of steerable QFT with a direct phase angle interpretation. In Section 4.1.2.5, we finally investigate new steerable QFTs involving quaternion conjugation. Their local geometric interpretation crucially relies on the notion of 4D rotation reflections. 4.1.2.2
Orthogonal planes split of quaternions with two orthonormal pure unit quaternions
For an introduction to quaternions we refer to Section 2.8.1. We remind the reader of the orthogonal 2D planes split (OPS) of quaternions with respect to the orthonormal pure unit quaternions i, j [184,187], see also Section
156 Quaternion and Clifford Fourier Transforms
2.8.2, defined by 1 q± = (q ± iqj). 2 Explicitly in real components qr , qi , qj , qk ∈ R using (2.8.1) we get q = q+ + q− ,
1±k 1±k = {qr ± qk + j(qj ∓ qi )}. 2 2 This leads to the following new Pythagorean modulus identity [187] q± = {qr ± qk + i(qi ∓ qj )}
(4.1.47)
(4.1.48)
Lemma 4.1.12 (Modulus identity). For q ∈ H q2 = q− 2 + q+ 2 .
(4.1.49)
Lemma 4.1.13 (Orthogonality of OPS split parts). Given any two quaternions p, q ∈ H and applying the OPS of (4.1.47) the resulting parts are orthogonal S(p+ q− ) = 0,
S(p− q+ ) = 0,
(4.1.50)
i.e. p+ ⊥ q− and p− ⊥ q+ . In Lemma 4.1.13 (proved in [187]) the second identity follows from the first by S(x) = S(x) , ∀x ∈ H, and p− q+ = q+ p− . It is evident, that instead of i, j any pair of orthonormal pure quaternions can be used to produce an analogous split. This is a first indication, that the OPS of (4.1.47) is in fact steerable. We observe, that iqj = q+ − q− , i.e. under the map i( )j the q+ part is invariant, the q− part changes sign. Both parts are according to (4.1.48) twodimensional, and by Lemma 4.1.13 they span two completely orthogonal planes. The q+ plane is spanned by the orthogonal quaternions {i − j, 1 + ij = 1 + k}, whereas the q− plane is e.g. spanned by {i + j, 1 − ij = 1 − k}, i.e. we have the two 2D subspace basis q+ basis: {i − j, 1 + ij = 1 + k},
q− basis: {i + j, 1 − ij = 1 − k}.
(4.1.51)
Note that all basis vectors of (4.1.51) {i − j, 1 + ij, i + j, 1 − ij}
(4.1.52) 4
together form an orthogonal basis of H interpreted as R . The map i( )j rotates the q− plane by 180◦ around the 2D q+ axis plane. Note that in agreement with its geometric interpretation, the map i( )j is an involution, because applying it twice leads to identity i(iqj)j = i2 qj 2 = (−1)2 q = q.
(4.1.53)
4.1.2.3 General orthogonal 2D planes split
We will study generalizations of the OPS split by replacing i, j by arbitrary unit quaternions f, g. Even with this generalization, the map f ( )g continues to be an involution, because f 2 qg 2 = (−1)2 q = q. For clarity we study the cases f 6= ±g, and f = g separately, though they have a lot in common, and do not always need to be distinguished in concrete applications. In the following we further extend the treatment given in Section 2.8.3.
Quaternion Fourier transforms 157
4.1.2.3.1
Orthogonal 2D planes split using two linearly independent pure unit quaternions
Our result is now, that all these properties hold, even if in the above considerations the pair i, j is replaced by an arbitrary pair of linearly independent nonorthogonal pure quaternions f, g, f 2 = g 2 = −1, f 6= ±g. The OPS is then redefined with respect to the linearly independent pure unit quaternions f, g as 1 q± = (q ± f qg). 2
(4.1.54)
Equation (4.1.47) is a special case with f = i, g = j. We observe from (4.1.54), that f qg = q+ − q− , i.e. under the map f ( )g the q+ part is invariant, but the q− part changes sign 1 1 1 f q± g = (f qg ± f 2 qg 2 ) = (f qg ± q) = ± (q ± f qg) = ±q± . 2 2 2
(4.1.55)
We now show that even for (4.1.54) both parts are twodimensional, and span two completely orthogonal planes. The q+ plane is spanned by the orthogonal pair of quaternions {f − g, 1 + f g}:
S (f − g)(1 + f g) = S((f − g)(1 + (−g)(−f )))
= S f + f gf − g − g 2 f
(2.8.13)
=
S f + f 2g − g + f
= 2 S(f − g) = 0,
(4.1.56)
whereas the q− plane is e.g. spanned by {f + g, 1 − f g}. The quaternions f + g, 1 − f g can be proved to be mutually orthogonal by simply replacing g → −g in (4.1.56). Note that we have f (f − g)g = f 2 g − f g 2 = −g + f = f − g, f (1 + f g)g = f g + f 2 g 2 = f g + 1 = 1 + f g,
(4.1.57)
as well as f (f + g)g = f 2 g + f g 2 = −g − f = −(f + g), f (1 − f g)g = f g − f 2 g 2 = f g − 1 = −(1 − f g).
(4.1.58)
We now want to generalize Lemma 4.1.13 to Lemma 4.1.14 (Orthogonality of two OPS planes). Given any two quaternions q, p ∈ H and applying the OPS (4.1.54) with respect to two linearly independent pure unit quaternions f, g we get zero for the scalar part of the mixed products S(p+ q− ) = 0,
S(p− q+ ) = 0.
(4.1.59)
We prove the first identity, the second follows from S(x) = S(x). S(p+ q− ) =
1 1 S((p + f pg)(q − gqf )) = S(pq − f pggqf + f pgq − pgqf ) 4 4 (2.8.15),(2.8.13) 1 = S(pq − pq + pgqf − pgqf ) = 0. (4.1.60) 4
158 Quaternion and Clifford Fourier Transforms
Thus the set {f − g, 1 + f g, f + g, 1 − f g}
(4.1.61)
forms a 4D orthogonal basis of H interpreted by (2.8.16) as R4 , where we have for the orthogonal 2D planes the subspace basis: q+ basis: {f − g, 1 + f g},
q− basis: {f + g, 1 − f g}.
(4.1.62)
We can therefore use the following representation for every q ∈ H by means of four real coefficients q1 , q2 , q3 , q4 ∈ R q = q1 (1 + f g) + q2 (f − g) + q3 (1 − f g) + q4 (f + g),
(4.1.63)
where
q2 = S q(f − g)−1 ,
q4 = S q(f + g)−1 .
q1 = S q(1 + f g)−1 , q3 = S q(1 − f g)−1 ,
As an example we have for f = i, g = j according to (4.1.48) 1 q1 = (qr + qk ), 2
1 q2 = (qi − qj ), 2
1 q3 = (qr − qk ), 2
1 q4 = (qi + qj ). 2
(4.1.64)
Moreover, using f − g = f (1 + f g) = (1 + f g)(−g),
f + g = f (1 − f g) = (1 − f g)g,
(4.1.65)
we have the following left and right factoring properties q+ = q1 (1 + f g) + q2 (f − g) = (q1 + q2 f )(1 + f g) = (1 + f g)(q1 − q2 g),
(4.1.66)
q− = q3 (1 − f g) + q4 (f + g) = (q3 + q4 f )(1 − f g) = (1 − f g)(q3 + q4 g).
(4.1.67)
Equations (4.1.57) and (4.1.58) further show that the map f ( )g rotates the q− plane by 180◦ around the q+ axis plane. We found that our interpretation of the map f ( )g is in perfect agreement with Coxeter’s notion of halfturn in [84]. This opens the way for new types of QFTs, where the pair of square roots of −1 involved does not necessarily need to be orthogonal. Before suggesting a generalization of the QFT, we will establish a new set of very useful algebraic identities. Based on (4.1.66) and (4.1.67) we get for α, β ∈ R eαf qeβg = eαf q+ eβg + eαf q− eβg , eαf q+ eβg = (q1 + q2 f )eαf (1 + f g)eβg = eαf (1 + f g)eβg (q1 − q2 g), eαf q− eβg = (q3 + q4 f )eαf (1 − f g)eβg = eαf (1 − f g)eβg (q3 + q4 g).
(4.1.68)
Using (4.1.66) again we obtain eαf (1 + f g) = (cos α + f sin α)(1 + f g) (4.1.66)
=
(1 + f g)(cos α − g sin α) = (1 + f g)e−αg ,
(4.1.69)
Quaternion Fourier transforms 159
where we set q1 = cos α, q2 = sin α for applying (4.1.66). Replacing in (4.1.69) −α → β we get e−βf (1 + f g) = (1 + f g)eβg , (4.1.70) Furthermore, replacing in (4.1.69) g → −g and subsequently α → β we get eαf (1 − f g) = (1 − f g)eαg ,
eβf (1 − f g) = (1 − f g)eβg ,
(4.1.71)
Applying (4.1.66), (4.1.68), (4.1.69) and (4.1.70) we can rewrite eαf q+ eβg
(4.1.68)
(q1 + q2 f )eαf (1 + f g)eβg
(4.1.66)
q+ e(β−α)g ,
= =
(4.1.69)
=
(q1 + q2 f )(1 + f g)e(β−α)g (4.1.72)
or equivalently as eαf q+ eβg
(4.1.68)
eαf (1 + f g)eβg (q1 − q2 g)
(4.1.66)
e(α−β)f q+ .
=
=
(4.1.70)
=
e(α−β)f (1 + f g)(q1 − q2 g) (4.1.73)
In the same way by changing g → −g, β → −β in (4.1.72) and (4.1.73) we can rewrite eαf q− eβg = e(α+β)f q− = q− e(α+β)g .
(4.1.74)
eαf q± eβg = q± e(β∓α)g = e(α∓β)f q± .
(4.1.75)
The result is therefore
Orthogonal 2D planes split using one pure unit quaternion We now treat the case for g = f, f 2 = −1. We then have the map f ( )f , and the OPS split with respect to f ∈ H, f 2 = −1, 1 (4.1.76) q± = (q ± f qf ). 2 The pure quaternion i can be rotated by R = i(i+f ), see (4.1.79), into the quaternion unit f and back. Therefore studying the map i( )i is up to the constant rotation between i and f the same as studying f ( )f . This gives 4.1.2.3.2
iqi = i(qr + qi i + qj j + qk k)i = −qr − qi i + qj j + qk k.
(4.1.77)
The OPS with respect to f = g = i gives 1 q± = (q ± iqi), 2
q+ = qj j + qk k = (qj + qk i)j,
q− = qr + qi i,
(4.1.78)
where the q+ plane is twodimensional and manifestly orthogonal to the 2D q− plane. This form (4.1.78) of the OPS is therefore identical to the quaternionic simplex/perplex split of [127], see an application in Fig. 4.1 from [127]. For g = f the q− plane is always spanned by {1, f }. The rotation operator
160 Quaternion and Clifford Fourier Transforms
√ Simplex/perplex split with gray line f = g = (i + j + k)/ 3. Photo: E. Hitzer. Image processing: S.J. Sangwine. Top left: Original. Top right: q− part (luminance). Bottom left: q+ part (qi ↔ qj ) (chrominance). Bottom right: Sum.
Figure 4.1
R = i(i + f ), with norm square R2 = i(i + f )2 = (i + f )2 = −(i + f )2 , rotates i into f according to R−1 iR =
(i + f )iii(i + f ) (i + f )i(i(−f ) + 1)f R iR = = 2 2 R −(i + f ) (i + f )2 (i + f )(f + i)f = = f. (i + f )2
(4.1.79)
The rotation R leaves 1 invariant and thus rotates the whole {1, i} plane into the q− plane spanned by {1, f }. Consequently R also rotates the {j, k} plane into the q+ plane spanned by {j 0 = R−1 jR, k0 = R−1 kR}. We thus constructively obtain the fully orthonormal 4D basis of H as {1, f, j 0 , k0 } = R−1 {1, i, j, k}R,
R = i(i + f ),
(4.1.80)
for any chosen pure unit quaternion f . Remark 4.1.15. Note, that there is a gauge freedom in this split by changing R → R exp(f ϕ/2), ϕ ∈ [0, 2π), i.e. a rotation freedom in the q+ plane. The units {1, f, j 0 , k0 } form another equivalent representation of quaternions H. We further have for the orthogonal 2D planes created in (4.1.76) the subspace basis: q+ basis: {j 0 , k0 }, q− basis: {1, f }. (4.1.81)
Quaternion Fourier transforms 161
Figure 4.2 4D scatter plot of quaternions decomposed using the orthogonal planes split of (4.1.76) with one unit pure quaternion f = i0 = √13 (i + j + k) = g. Left: q+ component; right: q− component. Source: [205, Fig. 1].
The rotation R (an orthogonal transformation!) of (4.1.79) preserves the fundamental quaternionic orthonormality and the anticommutation relations f j 0 = k0 = −j 0 f,
k0 f = j 0 = −f k0
j 0 k0 = f = −k0 j 0 .
(4.1.82)
Hence f qf = f (q+ + q− )f = q+ − q− , i.e. f q± f = ±q± ,
(4.1.83)
which represents again a half turn by 180◦ in the 2D q− plane around the 2D q+ plane (as axis). Equivalently, the q− part commutes with f , whereas the q+ part anticommutes q− f = f q− , q+ f = −f q+ . (4.1.84) Figures 4.2 and 4.3 illustrate this decomposition for the case where f is the unit pure quaternion √13 (i + j + k). This decomposition corresponds (for pure quaternions) to the classical luminancechrominance decomposition used in colour image processing, as illustrated, for example, in [127, Figure 2]. Three hundred unit quaternions randomly oriented in 4space were decomposed. Figure 4.2 shows the three hundred points in 4space, projected onto the six orthogonal planes {e, i0 }, {e, j 0 }, {e, k0 }, {i0 , j 0 }, {j 0 , k0 }, {k0 , i0 } where e = 1 and i0 = f , as given in (4.1.80). The six views on the left show the q+ plane, and the six on the right show the q− plane. Figure 4.3 shows the vector parts of the decomposed quaternions. The basis for the plot is {i0 , j 0 , k0 }, where i0 = f as given in (4.1.80). The green circles show the components in the {1, f } plane, which intersects the 3space of the vector part only along the line f (which is the luminance or grey line of colour image pixels). The red line on the figure corresponds to f . The blue circles show the components in
162 Quaternion and Clifford Fourier Transforms
Scatter plot of vector parts of quaternions decomposed using the orthogonal planes split of (4.1.76) with one pure unit quaternion f = i0 = √13 (i + j + k) = g. The red line corresponds to the direction of f . Source: [205, Fig. 2]. Figure 4.3
the {j 0 , k0 } plane, which is entirely within the 3space. It is orthogonal to f and corresponds to the chrominance plane of colour image processing. The next question is the influence the current OPS (4.1.76) has for left and right exponential factors of the form eαf q± eβf . (4.1.85) We learn from (4.1.82) that eαf q± eβf = e(α∓β)f q± = q± e(β∓α)f ,
(4.1.86)
which is identical to (4.1.75), if we insert g = f in (4.1.75). Next, we consider g = −f, f 2 = −1. We then have the map f ( )(−f ), and the OPS split with respect to f, −f ∈ H, f 2 = −1, 1 1 q± = (q ± f q(−f )) = (q ∓ f qf ). (4.1.87) 2 2 Again we can study f = i first, because for general pure unit quaternions f the unit quaternion i can be rotated by (4.1.79) into the quaternion unit f and back. Therefore studying the map i( )(−i) is up to the constant rotation R of (4.1.79) the same as studying f ( )(−f ). This gives the map iq(−i) = i(qr + qi i + qj j + qk k)(−i) = qr + qi i − qj j − qk k.
(4.1.88)
The OPS with respect to f = i, g = −i gives 1 q± = (q ± iq(−i)), 2
q− = qj j + qk k = (qj + qk i)j,
q+ = qr + qi i,
(4.1.89)
Quaternion Fourier transforms 163
where, compared to f = g = i, the 2D q+ plane and the 2D q− plane appear interchanged. The form (4.1.89) of the OPS is again identical to the quaternionic simplex/perplex split of [127], but the simplex and perplex parts appear interchanged. For g = −f the q+ plane is always spanned by {1, f }. The rotation R of (4.1.79) rotates i into f and leaves 1 invariant and thus rotates the whole {1, i} plane into the q+ plane spanned by {1, f }. Consequently, R of (4.1.79) also rotates the {j, k} plane into the q− plane spanned by {j 0 = R−1 jR, k0 = R−1 kR}. We therefore have for the orthogonal 2D planes created in (4.1.87) the subspace basis: q+ basis: {1, f }, q− basis: {j 0 , k0 }. (4.1.90) We again obtain the fully orthonormal 4D basis (4.1.80) of H, preserving the fundamental quaternionic orthonormality and the anticommutation relations (4.1.82). Hence for (4.1.87) f q(−f ) = f (q+ + q− )(−f ) = q+ − q− , i.e. f q± (−f ) = ±q± ,
(4.1.91)
which represents again a half turn by 180◦ in the 2D q− plane around the 2D q+ plane (as axis). The remaining question is the influence the current OPS (4.1.87) has for left and right exponential factors of the form eαf q± e−βf .
(4.1.92)
eαf q± e−βf = e(α∓β)f q± = q± e−(β∓α)f ,
(4.1.93)
We learn from (4.1.82) that
which is identical to (4.1.75), if we insert g = −f in (4.1.75). For (4.1.75) therefore, we do not any longer need to distinguish the cases f = 6 ±g and f = ±g. This motivates us to a general OPS definition for any pair of pure quaternions f, g, and we get a general lemma. Definition 4.1.16 (General orthogonal 2D planes split). Let f, g ∈ H be an arbitrary pair of pure quaternions f, g, f 2 = g 2 = −1, including the cases f = ±g. The general OPS is then defined with respect to the two pure unit quaternions f, g as 1 q± = (q ± f qg). 2
(4.1.94)
Remark 4.1.17. The three generalized OPS (4.1.54), (4.1.76) and (4.1.87) are formally identical and are now subsumed in (4.1.94) of Definition 4.1.16, where the values g = ±f are explicitly included, i.e. any pair of pure unit quaternions f, g ∈ H, f 2 = g 2 = −1, is admissible. Lemma 4.1.18. With respect to the general OPS of Definition 4.1.16 we have for left and right exponential factors the identity eαf q± eβg = q± e(β∓α)g = e(α∓β)f q± .
(4.1.95)
164 Quaternion and Clifford Fourier Transforms
4.1.2.3.3 Geometric interpretation of left and right exponential factors in f , g We obtain the following general geometric interpretation. The map f ( )g always represents a rotation by angle π in the q− plane (around the q+ plane), the map f t ( )g t , t ∈ R, similarly represents a rotation by angle tπ in the q− plane (around the q+ plane as axis). Replacing g → −g in the map f ( )g we further find that
f q± (−g) = ∓q± .
(4.1.96)
Therefore the map f ( )(−g) = f ( )g −1 , because g −1 = −g, represents a rotation by angle π in the q+ plane (around the q− plane), exchanging the roles of 2D rotation plane and 2D rotation axis. Similarly, the map f s ( )g −s , s ∈ R, represents a rotation by angle sπ in the q+ plane (around the q− plane as axis). The product of these two rotations gives π
π
f t+s qg t−s = e(t+s) 2 f qe(t−s) 2 g = eαf qeβg ,
π α = (t + s) , 2
π β = (t − s) , (4.1.97) 2
π
π
where based on (2.8.10) we used the identities f = e 2 f and g = e 2 g . The geometric interpretation of (4.1.97) is a rotation by angle α+β in the q− plane (around the q+ plane), and a second rotation by angle α − β in the q+ plane (around the q− plane). For α = β = π/2 we recover the map f ( )g, and for α = −β = π/2 we recover the map f ( )g −1 . 4.1.2.3.4 Determination of f, g for given steerable pair of orthogonal 2D planes Equations (4.1.62), (4.1.81), and (4.1.90) tell us how the pair of pure unit quaternions f, g ∈ H used in the general OPS of Definition 4.1.16, leads to an explicit basis for the resulting two orthogonal 2D planes, the q+ plane and the q− plane. We now ask the opposite question: How can we determine from a given steerable pair of orthogonal 2D planes in H the pair of pure unit quaternions f, g ∈ H, which splits H exactly into this given pair of orthogonal 2D planes? To answer this question, we first observe that in a 4D space it is sufficient to know only one 2D plane explicitly, specified e.g. by a pair of orthogonal unit quaternions a, b ∈ H, a = b = 1, and without restriction of generality b2 = −1, i.e. b can be a pure unit quaternion b = µ(b). But for a = cos α + µ(a) sin α, compare (2.8.10), we must distinguish S(a) = cos α 6= 0 and S(a) = cos α = 0, i.e. of a also being a pure quaternion with a2 = −1. The second orthogonal 2D plane is then simply the orthogonal complement in H to the a, b plane. Let us first treat the case S(a) = cos α = 6 0. We set
f := ab,
g := ab.
(4.1.98)
With this setting we get for the basis of the q− plane f + g = ab + ab = 2 S(a) b, 1 − f g = 1 − abab = 1 − a2 b2 = 1 + a2 = 1 + cos2 α − sin2 α + 2µ(a) cos α sin α = 2 cos α(cos α + µ(a) sin α) = 2 S(a) a.
(4.1.99)
Quaternion Fourier transforms 165
For the equality abab = a2 b2 we used the orthogonality of a, b, which means that the vector part of a must be orthogonal to the pure unit quaternion b, i.e. it must anticommute with b ab = ba, ba = ab. (4.1.100)
Comparing (4.1.62) and (4.1.99), the plane spanned by the two orthogonal unit quaternions a, b ∈ H is indeed the q− plane for S(a) = cos α 6= 0. The orthogonal q+ plane is simply given by its basis vectors (4.1.62), inserting (4.1.98). This leads to the pair of orthogonal unit quaternions c, d for the q+ plane as c= d=
f −g ab − ab a−a = = b = µ(a)b, f − g (a − a)b a − a
(4.1.101)
1 + fg f −g = g = cg = µ(a)bg = µ(a)bab = µ(a)ab2 = −µ(a)a, (4.1.102) 1 + f g f − g
where we have used (4.1.65) for the second, and (4.1.100) for the sixth equality in (4.1.102). Let us also verify that f, g of (4.1.98) are both pure unit quaternions using (4.1.100) f 2 = abab = (aa)bb = −1,
g 2 = abab = (aa)bb = −1.
(4.1.103)
Note, that if we would set in (4.1.98) for g := −ab, then the a, b plane would have become the q+ plane instead of the q− plane. We can therefore determine by the sign in the definition of g, which of the two OPS planes the a, b plane is to represent. For both a and b being two pure orthogonal quaternions, we can again set f := ab ⇒ f 2 = abab = −a2 b2 = −1,
g := ab = −ab = −f,
(4.1.104)
where due to the orthogonality of the pure unit quaternions a, b we were able to use ba = −ab. In this case f = ab is thus also shown to be a pure unit quaternion. Now the q− plane of the corresponding OPS (4.1.87) is spanned by {a, b}, whereas the q+ plane is spanned by {1, f }. Setting instead g := −ab = ab = f , the q− plane of the corresponding OPS (4.1.76) is spanned by {1, f }, whereas the q+ plane is spanned by {a, b}. We summarize our results in Theorem 4.1.19. (Determination of f, g from given steerable 2D plane) Given any 2D plane in H in terms of two unit quaternions a, b, where b is without restriction of generality pure, i.e. b2 = −1, we can make the given plane the q− plane of the OPS q± = 12 (q ± f qg), by setting f := ab,
g := ab.
(4.1.105)
For S(a) = 6 0 the orthogonal q+ plane is fully determined by the orthogonal unit quaternions c = µ(a)b, d = −µ(a)a. (4.1.106)
166 Quaternion and Clifford Fourier Transforms
4D scatter plot of quaternions decomposed using the orthogonal planes split of Definition 4.1.16. Left: q+ component; right: q− component. Source: [205, Fig. 3].
Figure 4.4
where µ(a) is as defined in (2.8.10). For S(a) = 0 the orthogonal q+ plane with basis {1, f } is instead fully determined by f = −g = ab and (4.1.90). Setting alternatively f := ab, g := −ab (4.1.107) makes the given a, b plane the q+ plane instead. For S(a) = 6 0 the orthogonal q− plane is then fully determined by (4.1.107) and (4.1.62), with the same orthogonal unit quaternions c = µ(a)b, d = −µ(a)a as in (4.1.106). For S(a) = 0 the orthogonal q− plane with basis {1, f } is then instead fully determined by f = g = ab and (4.1.81). An illustration of the decomposition is given in Figures 4.4 and 4.5. Again, three hundred unit pure quaternions randomly oriented in 4space have been decomposed into two sets using the decomposition of Definition 4.1.16 and two unit pure quaternions f and g computed as in Theorem 4.1.19. b was the pure unit quaternion √1 (i + j + k) and a was the full unit quaternion √1 + 1 (i − j). c and d were 2 3 2 computed by (4.1.106) as c = µ(a)b and d = −µ(a)a. Figure 4.4 shows the three hundred points in 4space, projected onto the six orthogonal planes {c, d}, {c, b}, {c, a}, {d, b}, {b, a}, {a, d} where the orthonormal 4space basis {c, d, b, a} = {(f − g)/f − g, (1 + f g)/1 + f g, (f + g)/f + g, (1 − f g)/1 − f g}. The six views on the left show the q+ plane, and the six on the right show the q− plane. Figure 4.5 shows the vector parts of the decomposed quaternions. 4.1.2.4 New QFT forms: OPSQFTs with two pure unit quaternions f, g
We note that the versions of the QFT defined here on L1 can also be defined on L2 as explained in Footnote 2 on page 144. We begin with a straightforward generalization of the (doublesided form of the) QFT [184, 187] in H by replacing i → f and j → gdefined as 4.1.2.4.1
Generalized OPS leads to new steerable type of QFT
Definition 4.1.20. (QFT with respect to two pure unit quaternions f, g)
Quaternion Fourier transforms 167
Scatter plot of vector parts of quaternions decomposed using the orthogonal planes split of Definition 4.1.16. Source: [205, Fig. 4].
Figure 4.5
Let f, g ∈ H, f 2 = g 2 = −1, be any two pure unit quaternions. The quaternion Fourier transform with respect to f, g is F f,g {h}(ω) =
Z
e−f x1 ω1 h(x) e−gx2 ω2 d2 x,
(4.1.108)
R2
where h ∈ L1 (R2 , H), d2 x = dx1 dx2 and x, ω ∈ R2 . Note, that the pure unit quaternions f, g in Definition 4.1.20 do not need to be orthogonal, and that the cases f = ±g are fully included. Linearity of the integral (4.1.108) allows us to use the OPS split h = h− + h+ F f,g {h}(ω) = F f,g {h− }(ω) + F f,g {h+ }(ω) = F−f,g {h}(ω) + F+f,g {h}(ω), (4.1.109) since by their construction the operators of the Fourier transformation F f,g , and of the OPS with respect to f, g commute. From Lemma 4.1.18 follows Theorem 4.1.21 (QFT of h± ). The QFT of the h± OPS split parts, with respect to two unit quaternions f, g, of a quaternion module function h ∈ L1 (R2 , H) have the quasicomplex forms F±f,g {h} = F f,g {h± } Z
= R2
h± e−g(x2 ω2 ∓x1 ω1 ) d2 x =
Z R2
e−f (x1 ω1 ∓x2 ω2 ) h± d2 x .
(4.1.110)
Remark 4.1.22. The quasicomplex forms in Theorem 4.1.21 allow to establish discretized and fast versions of the QFT of Definition 4.1.20 as sums of two complex discretized and fast Fourier transformations (FFT), respectively.
168 Quaternion and Clifford Fourier Transforms
Geometric interpretation of integrand of QFTf,g in Definition 4.1.20 in terms of local phase rotations in q± planes. Source: [205, Fig. 5]. Figure 4.6
We can now give a geometric interpretation of the integrand the QFTf,g in Definition 4.1.20 in terms of local phase rotations, compare Section 4.1.2.3.3. The integrand product e−f x1 ω1 h(x) e−gx2 ω2 (4.1.111) means to locally rotate by the phase angle −(x1 ω1 + x2 ω2 ) in the q− plane, and by phase angle −(x1 ω1 − x2 ω2 ) = x2 ω2 − x1 ω1 in the orthogonal q+ plane, compare Figure 4.6, which depicts two completely orthogonal planes in four dimensions. Based on Theorem 4.1.19 the two phase rotation planes (analysis planes) can be freely steered by defining the two pure unit quaternions f, g used in Definition 4.1.20 according to (4.1.105) or (4.1.107). Two phase angle version of QFT The above newly gained geometric understanding motivates us to propose a further new version of the QFTf,g , with a straightforward two phase angle interpretation. 4.1.2.4.2
Definition 4.1.23. (Phase angle QFT with respect to f, g) Let f, g ∈ H, f 2 = g 2 = −1, be any two pure unit quaternions. The phase angle quaternion Fourier transform with respect to f, g is f,g FD {h}(ω) =
Z
1
1
e−f 2 (x1 ω1 +x2 ω2 ) h(x) e−g 2 (x1 ω1 −x2 ω2 ) d2 x.
(4.1.112)
R2
where again h ∈ L1 (R2 , H), d2 x = dx1 dx2 and x, ω ∈ R2 . The geometric interpretation of the integrand of (4.1.112) is a local phase rotation by angle −(x1 ω1 + x2 ω2 )/2 − (x1 ω1 − x2 ω2 )/2 = −x1 ω1 in the q− plane, and a second local phase rotation by angle −(x1 ω1 + x2 ω2 )/2 + (x1 ω1 − x2 ω2 )/2 = −x2 ω2 in the q+ plane, compare Section 4.1.2.3.3. If we apply the OPSf,g split to (4.1.112) we obtain the following theorem. Theorem 4.1.24 (Phase angle QFT of h± ). The phase angle QFT of Definition
Quaternion Fourier transforms 169
4.1.23 applied to the h± OPS split parts, with respect to two pure unit quaternions f, g, of a quaternion module function h ∈ L1 (R2 , H) leads to the quasicomplex expressions f,g f,g {h+ } = {h} = FD FD+
Z
f,g FD− {h}
Z
=
f,g FD {h− }
R2
= R2
h+ e+gx2 ω2 d2 x = −gx1 ω1 2
h− e
Z R2
Z
d x= R2
e−f x2 ω2 h+ d2 x,
(4.1.113)
e−f x1 ω1 h− d2 x,
(4.1.114)
Note that based on Theorem 4.1.19 the two phase rotation planes (analysis planes) are again freely steerable. Theorem 4.1.24 allows to establish discretized and fast versions of the phase angle QFT of Definition 4.1.23 as sums of two complex discretized and fast Fourier transformations (FFT), respectively. The maps f ( )g considered so far did not involve quaternion conjugation q → q. In the following we investigate maps which additionally quaternion conjugate the argument, i.e. of type f ( )g, which are also involutions. 4.1.2.5 Involutions and QFTs involving quaternion conjugation 4.1.2.5.1
Involutions involving quaternion conjugations
The simplest case is quater
nion conjugation itself q → q = qr − qi i − qj j − qk k,
(4.1.115)
which can be interpreted as a reflection at the real line qr . The real line through the origin remains pointwise invariant, while every other point in the 3D hyperplane of pure quaternions is reflected to the opposite side of the real line. The related involution q → −q = −qr + qi i + qj j + qk k, (4.1.116) is the reflection at the 3D hyperplane of pure quaternions (which stay invariant), i.e. only the real line is changed into its negative qr → −qr . Similarly any pure unit quaternion factor like i in the map q → i qi = −qr + qi i − qj j − qk k,
(4.1.117)
leads to a reflection at the (pointwise invariant) line through the origin with direction i, while the map q → −i qi = qr − qi i + qj j + qk k, (4.1.118) leads to a reflection at the invariant 3D hyperplane orthogonal to the line through the origin with direction i. (4.1.119) q → f qf, leads to a reflection at the (pointwise invariant) line with direction f through the origin, while the map q → −f qf, (4.1.120) leads to a reflection at the invariant 3D hyperplane orthogonal to the line with direction f through the origin.
170 Quaternion and Clifford Fourier Transforms
Next we turn to a map of the type q → −eαf qeαf .
(4.1.121)
Its set of pointwise invariants is given by q = −eαf qeαf ⇔ e−αf q = −qeαf ⇔ e−αf q + qeαf = 0
⇔ S qeαf = 0 ⇔ q ⊥ eαf .
(4.1.122)
eαf → −eαf e−αf eαf = −eαf .
(4.1.123)
We further observe that
The map −a( )a, with unit quaternion a = eαf , therefore represents a reflection at the invariant 3D hyperplane orthogonal to the line through the origin with direction a. Similarly, the map a( )a, with unit quaternion a = eαf , then represents a reflection at the (pointwise invariant) line with direction a through the origin. The combination of two such reflections (both at 3D hyperplanes, or both at lines), given by unit quaternions a, b, leads to a rotation −b−aqab = baqab = baqab = rqs, r = ba,
s = ab,
r = b a = 1 = s ,
(4.1.124)
in two orthogonal planes, exactly as studied in Section 4.1.2.3.3. The combination of three reflections at 3D hyperplanes, given by unit quaternions a, b, c , leads to −c[−b−aqab]c = d qt,
d = −cba, t = abc,
d = c b a = t = 1.
(4.1.125)
The product of the reflection map −q of (4.1.116) with d qt leads to −dqt, a double rotation as studied in Sections 4.1.2.4. Therefore d ( ) t represents a rotary reflection. The three reflections −aqa, −bqb, −cqc have the intersection of the three 3D hyperplanes as a resulting common pointwise invariant line, which is d + t, because d (d + t) t = d tt + d dt = d + t.
(4.1.126)
In the remaining 3D hyperplane, orthogonal to the pointwise invariant line through the origin in direction d + t, the axis of the rotary reflection is d (d − t) t = −d tt + d dt = −d + t = −(d − t).
(4.1.127)
We now also understand that a sign change of d → −d (compare three reflections at three 3D hyperplanes −c[−b(−aqa)b]c with three reflections at three lines +c[+b(+aqa)b]c) simply exchanges the roles of pointwise invariant line d + t and rotary reflection axis d − t.
Quaternion Fourier transforms 171
Next, we seek for an explicit description of the rotation plane of the rotary reflection d ( ) t. We find that for the unit quaternions d = eαg , t = eβf the commutator [d, t] = dt − td = eαg eβf − eβf eαg = (gf − f g) sin α sin β,
(4.1.128)
is a pure quaternion, because gf − f g = f g − gf = −(gf − f g).
(4.1.129)
Moreover, [d, t] is orthogonal to d and t, and therefore orthogonal to the plane spanned by the pointwise invariant line d + t and the rotary reflection axis d − t, because S [d, t]d = S dtd − td d = 0, S([d, t]t) = 0. (4.1.130) We obtain a second quaternion in the plane of d + t, d − t by applying the rotation reflection to [d, t] d [d, t] t = −d[d, t]t = −[d, t]dt, (4.1.131) because d is orthogonal to the pure quaternion [d, t]. We can construct an orthogonal basis of the plane of the rotation reflection d ( ) t by computing the pair of orthogonal quaternions v1,2 = [d, t] ∓ d [d, t] t = [d, t] ± [d, t]dt = [d, t](1 ± dt).
(4.1.132)
For finally computing the rotation angle, we need to know the relative length of the two orthogonal quaternions v1 , v2 of (4.1.132). For this it helps to represent the unit quaternion dt as dt = eγu , γ ∈ R, u ∈ H, u2 = −1. (4.1.133) We then obtain for the length ratio r2 =
1 + dt2 v1 2 (1 + dt)(1 + td) 1 + dttd + dt + td = = = 2 2 v2  1 − dt (1 − dt)(1 − td) 1 + dttd − dt − td 2 + 2 cos γ 1 + cos γ = = . 2 − 2 cos γ 1 − cos γ
(4.1.134)
By applying the rotation reflection d ( ) t to v1 and decomposing the result with respect to the pair of orthogonal quaternions in the rotation reflection plane (4.1.132) we can compute the rotation angle. Applying the rotation reflection to v1 gives d v1 t = d [d, t] − d[d, t]t t = d[d, t](1 + dt)t = d(1 + td)[d, t]t = d(1 + td)(−[d, t])t = −[d, t](dt + dtdt).
(4.1.135)
The square of dt is (dt)2 = (cos γ + u sin γ)2 = −1 + 2 cos γ [cos γ + u sin γ] = −1 + 2 cos γ dt.
(4.1.136)
172 Quaternion and Clifford Fourier Transforms
We therefore get d v1 t = −[d, t](dt − 1 + 2 cos γdt) = [d, t](1 − (1 + 2 cos γ)dt) = a1 v1 + a2 rv2 ,
(4.1.137)
1 − (1 + 2 cos γ)dt = a1 (1 + dt) + a2 r(1 − dt),
(4.1.138)
and need to solve
which leads to d v1 t = − cos γv1 + sin γrv2 = cos(π − γ)v1 + sin(π − γ)rv2 .
(4.1.139)
The rotation angle of the rotation reflection d ( ) t in its rotation plane v1 , v2 is therefore Γ = π − γ, γ = arccos S dt . (4.1.140) In terms of d = eαg , t = eβf we get dt = cos α cos β − g sin α cos β + f cos α sin β − gf sin α sin β.
(4.1.141)
And with the angle ω between g and f 1 1 1 gf = (gf + f g) + (gf − f g) = S(gf ) + [g, f ] 2 2 2 [g, f ] = − cos ω − sin ω , [g, f ]
(4.1.142)
we finally obtain for γ the scalar part S dt as
S dt = cos γ = cos α cos β + cos ω sin α sin β = cos α cos β − S(gf ) sin α sin β.
(4.1.143)
In the special case of g = ±f , S(gf ) = ∓1, i.e. for ω = 0, π, we get from (4.1.143) that
S dt = cos α cos β ± sin α sin β = cos α cos β + sin(±α) sin β = cos(±α − β),
(4.1.144)
and thus using (4.1.140) the rotation angle would become Γ = π − (±α − β) = π ∓ α + β.
(4.1.145)
Yet (4.1.140) was derived assuming [d, t] = 6 0. But direct inspection shows that (4.1.145) is indeed correct: For g = ±f the plane d + t, d − t is identical to the 1, f plane. The rotation plane is thus a plane of pure quaternions orthogonal to the 1, f plane. The quaternion conjugation in q 7→ d q t leads to a rotation by π and the left and right factors lead to further rotations by ∓α and β, respectively. Thus (4.1.145) is verified as a special case of (4.1.140) for g = ±f . By substituting in Lemma 4.1.18 (α, β) → (−β, −α), and by taking the quaternion conjugate we obtain the following Lemma.
Quaternion Fourier transforms 173
Lemma 4.1.25. Let q± = 21 (q ± f qg) be the OPS of Definition 4.1.16. For left and right exponential factors we have the identity eαg q± eβf = q± e(β∓α)f = e(α∓β)g q± .
(4.1.146)
4.1.2.5.2
New steerable QFTs with quaternion conjugation and two pure unit quaternions f, g We therefore consider now the following new variant of the (doublesided form of the) QFT [184, 187] in H (replacing both i → g and j → f , and using quaternion conjugation). It is essentially the quaternion conjugate of the new QFT of Definition 4.1.20, but because of its distinct local transformation geometry deserves separate treatment.
Definition 4.1.26. (QFT with respect to f, g, including quaternion conjugation) Let f, g ∈ H, f 2 = g 2 = −1, be any two pure unit quaternions. The quaternion Fourier transform with respect to f, g, involving quaternion conjugation, is Fcg,f {h}(ω) = F f,g {h}(−ω) =
Z
e−gx1 ω1 h(x) e−f x2 ω2 d2 x,
(4.1.147)
R2
where h ∈ L1 (R2 , H), d2 x = dx1 dx2 and x, ω ∈ R2 . Linearity of the integral in (4.1.147) of Definition 4.1.26 leads to the following corollary to Theorem 4.1.21. Corollary 4.1.27 (QFT Fcg,f of h± ). The QFT Fcg,f (4.1.147) of the h± = 12 (h ± f hg) OPS split parts, with respect to any two unit quaternions f, g, of a quaternion module function h ∈ L1 (R2 , H) have the quasicomplex forms Fcg,f {h± }(ω) = F f,g {h± }(−ω) Z
= R2
−f (x2 ω2 ∓x1 ω1 ) 2
h± e
Z
d x= R2
e−g(x1 ω1 ∓x2 ω2 ) h± d2 x .
(4.1.148)
Note, that the pure unit quaternions f, g in Definition 4.1.26 and Corollary 4.1.27 do not need to be orthogonal, and that the cases f = ±g are fully included. Corollary 4.1.27 leads to discretized and fast versions of the QFT with quaternion conjugation of Definition 4.1.26. It is important to note that the roles (sides) of f, g appear exchanged in (4.1.147) of Definition 4.1.26 and in Corollary 4.1.27, although the same OPS of Definition 4.1.16 is applied to the signal h as in Sections 4.1.2.3 and 4.1.2.4. This role change is due to the presence of quaternion conjugation in Definition 4.1.26. Note that it is possible to first apply (4.1.147) to h, and subsequently split the integral with the OPSg,f Fc,± = 12 (Fc ± gFc f ), where the particular order of g from the left and f from the right is due to the application of conjugation in (4.1.148) to h± after h is split with (4.1.94) into h+ and h− . Regarding the local geometric interpretation of the QFT with quaternion conjugation of 4.1.2.5.3
Local geometric interpretation of the QFT with quaternion conjugation
174 Quaternion and Clifford Fourier Transforms
Definition 4.1.26 we need to distinguish the following cases, depending on [d, t] and on whether the left and right phase factors d = e−gx1 ω1 ,
t = e−f x2 ω2 ,
(4.1.149)
attain scalar values ±1 or not. Let us first assume that [d, t] = 6 0, which by (4.1.128) is equivalent to g 6= ±f , and sin(x1 ω1 ) = 6 0, and sin(x2 ω2 ) 6= 0. Then we have the generic case of a local rotation reflection with pointwise invariant line of direction
rotation axis in direction
d + t = e−gx1 ω1 + e−f x2 ω2 ,
(4.1.150)
d − t = e−gx1 ω1 − e−f x2 ω2 ,
(4.1.151)
rotation plane with basis (4.1.132), and by (4.1.140) and (4.1.143) the general rotation angle
Γ = π − arccos S dt ,
S dt = cos(x1 ω1 ) cos(x2 ω2 ) − S(gf ) sin(x1 ω1 ) sin(x2 ω2 ).
(4.1.152)
Whenever g = ±f , or when sin(x1 ω1 ) = 0 (x1 ω1 = 0, π[mod 2π], i.e. d = ±1), we get for the pointwise invariant line in direction d + t the simpler unit quaternion 1 direction expression e− 2 (±x1 ω1 +x2 ω2 )f , because we can apply 1
1
1
1
eαf + eβf = e 2 (α+β)f (e 2 (α−β)f + e 2 (β−α)f ) = e 2 (α+β)f 2 cos
α−β , 2
(4.1.153)
1
and similarly for the rotation axis d − t we obtain the direction e− 2 (±x1 ω1 +x2 ω2 +π)f, whereas the rotation angle is by (4.1.145) simply Γ = π ± x1 ω1 − x2 ω2 .
(4.1.154)
For sin(x2 ω2 ) = 0 (x2 ω2 = 0, π[mod 2π], i.e. t = ±1), the pointwise invariant 1 line in direction d + t simplifies by (4.1.153) to e− 2 (x1 ω1 +x2 ω2 )g , and the rotation axis 1 with direction d − t simplifies to e− 2 (x1 ω1 +x2 ω2 +π)g , whereas the angle of rotation is by (4.1.145) simply Γ = π + x1 ω1 − x2 ω2 . (4.1.155) Even when quaternion conjugation is applied to the signal h we can propose a further new version of the QFTg,f c , with a straight forward two phase angle interpretation. The following definition to some degree ignores the resulting local rotation reflection effect of combining quaternion conjugation and left and right phase factors of Section 4.1.2.5.3, but depending on the application context, it may nevertheless be of interest in its own right. 4.1.2.5.4
Phase angle QFT with respect to f, g , including quaternion conjugation
Quaternion Fourier transforms 175
Definition 4.1.28. (Phase angle QFT with respect to f, g, including quaternion conjugation) Let f, g ∈ H, f 2 = g 2 = −1, be any two pure unit quaternions. The phase angle quaternion Fourier transform with respect to f, g, involving quaternion conjugation, is g,f f,g {h}(−ω1 , ω2 ) {h}(ω) = FD FcD
Z
1
1
e−g 2 (x1 ω1 +x2 ω2 ) h(x) e−f 2 (x1 ω1 −x2 ω2 ) d2 x.
=
(4.1.156)
R2
where again h ∈ L1 (R2 , H), d2 x = dx1 dx2 and x, ω ∈ R2 . Based on Lemma 4.1.25, one possible geometric interpretation of the integrand of (4.1.156) is a local phase rotation of h+ by angle −(x1 ω1 − x2 ω2 )/2 + (x1 ω1 + x2 ω2 )/2 = +x2 ω2 in the q+ plane, and a second local phase rotation of h− by angle −(x1 ω1 − x2 ω2 )/2 − (x1 ω1 + x2 ω2 )/2 = −x1 ω1 in the q− plane. This is expressed in the following corollary to Theorem 4.1.24. Corollary 4.1.29 (Phase angle QFT of h± , involving quaternion conjugation). The phase angle QFT with quaternion conjugation of Definition 4.1.28 applied to the h± OPS split parts, with respect to any two pure unit quaternions f, g, of a quaternion module function h ∈ L1 (R2 , H) leads to the quasicomplex expressions g,f g,f FcD {h+ }(ω) = FD {h+ }(−ω1 , ω2 )
Z
= R2
h+ e
+f x2 ω2 2
Z
d x= R2
e−gx2 ω2 h+ d2 x ,
(4.1.157)
g,f g,f FcD {h− }(ω) = FD {h− }(−ω1 , ω2 )
Z
= R2
h− e−f x1 ω1 d2 x =
Z R2
e−gx1 ω1 h− d2 x .
(4.1.158)
Note that based on Theorem 4.1.19 the two phase rotation planes (analysis planes) are again freely steerable. Corollary 4.1.29 leads to discretized and fast versions of the phase angle QFT with quaternion conjugation of Definition 4.1.28. 4.1.2.6 Summary on QFT and OPS of quaternions
The involution maps i( )j and more general f ( )g have led us to explore a range of similar quaternionic maps q 7→ aqb and q 7→ aqb, where a, b are taken to be unit quaternions. Geometric interpretations of these maps as reflections, rotations and rotary reflections in 4D can mostly be found in [84], see also Section 2.8. We have further developed these geometric interpretations to gain a complete local transformation geometric understanding of the integrands of the proposed new quaternion Fourier transformations (QFTs) applied to general quaternionic signals h ∈ L1 (R2 , H). This new geometric understanding is also valid for the special cases of the hitherto well known left sided, right sided and left and right sided (twosided) QFTs of [124, 127, 184, 291] and numerous other references.
176 Quaternion and Clifford Fourier Transforms
Our newly gained geometric understanding itself motivated us to study new types of QFTs with specific geometric properties. The investigation of these new types of QFTs with the generalized form of the orthogonal 2D planes split of Definition 4.1.16 lead to important QFT split theorems , which allow the use of discrete and (complex) Fourier transform software for efficient discretized and fast numerical implementations. Finally, our geometric interpretation of old and new QFTs paves the way for new applications, e.g., regarding steerable filter design for specific tasks in image, color image and signal processing, etc.
4.2 4.2.1
PROPERTIES OF QUATERNION FOURIER TRANSFORM Uncertainty principle for quaternion Fourier transform
This section derives a directional uncertainty principle for quaternion valued functions subject to the quaternion Fourier transformation. This can be generalized to establish directional uncertainty principles in Clifford geometric algebras with quaternion subalgebras. We demonstrate this later in Section 4.4 with the example of a directional spacetime algebra function uncertainty principle related to multivector wave packets. Our treatment is based on [187]. See [255] for a strictly coordinate wise approach. The Heisenberg uncertainty principle and quaternions are both fundamental for quantum mechanics, including the spin of elementary particles. The quaternion Fourier transform (QFT) [124, 184] is used in image and signal processing. It allows to formulate a component wise uncertainty principle [56, 255, 256]. In Clifford geometric algebras, which generalize real and complex numbers and quaternions to higher dimensions, the vector differential allows to formulate a more general directional uncertainty principle [185, 251]. The present text formulates a directional uncertainty principle for quaternion functions. As prerequisite for its proof we further investigate the split of quaternions introduced in [184, 265]. 4.2.1.1 Uncertainty principle The Heisenberg uncertainty principle in physics Photons (quanta of light) scattering off particles to be detected cause minimal position momentum uncertainty during detection. The uncertainty principle has played a fundamental role in the development and understanding of quantum physics. It is also central for information processing [287]. In quantum physics it states, e.g., that particle momentum and position cannot be simultaneously measured with arbitrary precision. The hypercomplex (quaternion or more general multivector) function f (x) would represent the spatial part of a separable wave function and its Fourier transform (QFT, CFT, SFT, ...) F{f }(ω) the same wave function in momentum space (compare [108, 299, 319]). The variance in space Rn (in physics often n = 3) would then be calculated as (1 ≤ k ≤ n) 4.2.1.1.1
(∆xk )2 =
Z
hf (x)(ek · x)2 f˜(x)i dn x =
Rn
Z
(ek · x)2 f (x)2 dn x,
Rn
where it is customary to set without loss of generality the mean value of ek · x to
Quaternion Fourier transforms 177
zero [319]. The variance in momentum space would be calculated as (1 ≤ l ≤ n) 1 e }(ω)i dn ω hF{f }(ω)(el · ω)2 F{f (2π)n Rn Z 1 (el · ω)2 F{f }(ω)2 dn ω. (2π)n Rn Z
2
(∆ωl )
= =
Again the mean value of el · ω is customarily set to zero, it merely corresponds to a phase shift [319]. Using our mathematical units, the positionmomentum uncertainty relation of quantum mechanics is then expressed by (compare e.g. with (4.9) of [299, page 86]) 1 (4.2.1) ∆xk ∆ωl = δk,l F, 2 where δk,l is the usual Kronecker symbol. NoteR that we have not normalized the squares of the variances by division with F = Rn f (x)2 dn x, therefore the extra factor F on the right side of (4.2.1). Further explicit examples from image processing can be found in [132]. In general in Fourier analysis such conjugate entities correspond to the variances of a function and its Fourier transform which cannot both be simultaneously sharply localized (e.g. [74,287]). Material on the classical uncertainty principle for the general case of L2 (Rn ) without the additional condition limx→∞ x2 f (x) = 0 can be found in [250] and [147]. Felsberg [132] even notes for two dimensions: In 2D however, the uncertainty relation is still an open problem. In [145] it is stated that there is no straightforward formulation for the 2D uncertainty relation. Let us now briefly define the notion of directional uncertainty principle in Clifford geometric algebra. Directional uncertainty principle in Clifford geometric algebra From the view point of Clifford geometric algebra an uncertainty principle gives us information about how the variance of a multivector valued function and the variance of its Clifford Fourier transform are related. We can shed the restriction to the parallel (k = l) and orthogonal (k = 6 l) cases of (4.2.1) by looking at the x ∈ Rn variance in an arbitrary but fixed direction a ∈ Rn and at the ω ∈ Rn variance in an arbitrary but fixed direction b ∈ Rn . We are now concerned with Clifford geometric algebra multivector functions f : Rn → Gn , n = 2, 3 (mod 4) with GA FT F{f }(ω) such that19 4.2.1.1.2
Z
f (x)2 dn x = F < ∞.
(4.2.2)
Rn
The directional uncertainty principle [185, 251] with arbitrary constant vectors a, b ∈ Rn means that 1 F 19
1 (a · x) f (x) d x n (2π)n F R
Z
2
2
n
Z
(b · ω)2 F{f }(ω)2 dn ω ≥
Rn
(a · b)2 . 4
Regarding the definition of the GA FT (CFT) on L2 , see Footnote 2 on page 144.
(4.2.3)
178 Quaternion and Clifford Fourier Transforms
Equality (minimal bound) in (4.2.3) is achieved for optimal Gaussian multivector functions 2 f (x) = C0 e−α x , (4.2.4) C0 ∈ Gn arbitrary const. multivector, 0 < α ∈ R. So far there seems to be no directional uncertainty principle for quaternion functions over R2 subject to the QFT. But there already exists a component wise uncertainty principle. 4.2.1.1.3
Componentwise uncertainty principle for rightsided QFT
For the right sided
QFT20 Fr {f }: R2 → H of f ∈ L1 (R2 ; H) given by [56, 255] Fr {f }(ω) =
Z
f (x)e−iω1 x1 e−j ω2 x2 d2 x,
(4.2.5)
R2
where x = x1 e1 + x2 e2 , ω = ω1 e1 + ω2 e2 , and the quaternion exponential product e−iω1 x1 e−j ω2 x2 is the quaternion Fourier kernel, it is possible to establish a component uncertainty principle (k = 1, 2) Z R2
x2k f (x)2 d2 x
Z R2
ωk2 Fr {f }(ω)2 d2 ω ≥
(2π)2 4
Z
f (x)2 d2 x
2
,
(4.2.6)
R2
where f ∈ L2 (R2 ; H) is now a quaternionvalued signal such that both (1+xk )f (x) ∈ L2 (R2 ; H) and ∂x∂ k f (x) ∈ L2 (R2 ; H). It is further possible to prove [255] that equality holds if and only if f is a Gaussian quaternion function 2
2
f (x) = C0 e−α1 x1 −α2 x2 ,
const. C0 ∈ H,
0 < α1 , α2 ∈ R.
(4.2.7)
Remark 4.2.1. Now we want to address the very important question: Is it also possible to establish a full directional uncertainty principle for the QFT? In the following we will try to answer this question after investigating some relevant properties of the doublesided QFT, which is slightly different from (4.2.5). 4.2.1.2 Quaternion Fourier transform (QFT)
From now on we only consider the doublesided QFT F{f }: R2 → H of f ∈ L1 (R2 ; H) in the form21 , F{f }(ω) = fˆ(ω) =
Z
e−ix1 ω1 f (x) e−jx2 ω2 d2 x.
(4.2.8)
R2
Linearity allows to also split up the QFT itself as F{f }(ω) = F{f− + f+ }(ω) = F{f− }(ω) + F{f+ }(ω).
20 21
For the QFT on L2 , see Footnote 2 on page 144. For the QFT on L2 , see Footnote 2 on page 144.
(4.2.9)
Quaternion Fourier transforms 179
Simple complex forms for QFT of f± The QFT of the f± split parts of a quaternion function f ∈ L2 (R2 , H) have simple complex forms [184]
4.2.1.2.1
fˆ± =
Z R2
f± e−j(x2 ω2 ∓x1 ω1 ) d2 x =
Z R2
e−i(x1 ω1 ∓x2 ω2 ) f± d2 x.
(4.2.10)
We can rewrite this free of coordinates as fˆ− =
Z
−j x ·ω 2
R2
f− e
d x,
fˆ+ =
Z R2
f+ e−j x ·(U1 ω) d2 x,
(4.2.11)
where the reflection U1 ω changes component ω1 → −ω1 . That this applies to ω1 is due to the choice of the kernel in (4.2.8). Preparations for the full directional QFT uncertainty principle Now we continue to lay the ground work for the full directional QFT uncertainty principle. We begin with the following lemma. 4.2.1.2.2
Lemma 4.2.2 (Integration of parts). With the vector differential a·∇ = a1 ∂1 +a2 ∂2 , with arbitrary constant a ∈ R2 , g, h ∈ L2 (R2 , H) Z
g(x)[a · ∇h(x)]d2 x
R2
Z
=
a·x=∞
−
g(x)h(x)dx R
a·x=−∞
Z
[a · ∇g(x)]h(x)d2 x.
(4.2.12)
R2
Remark 4.2.3. The proof of Lemma 4.2.2 works very similar to the proof of integration of parts in [251]. We therefore don’t repeat it here. For a quaternion function f and its QFT F{f } itself we also get important modulus identities. Lemma 4.2.4 (Modulus identities). Due to q2 = q− 2 + q+ 2 of Lemma 4.1.12 we get for f : R2 → H the following identities f (x)2 = f− (x)2 + f+ (x)2 ,
(4.2.13)
F{f }(ω)2 = F{f− }(ω)2 + F{f+ }(ω)2 .
(4.2.14)
We further establish formulas for the vector differentials of the QFTs of the split function parts f− and f+ . Lemma 4.2.5 (QFT of vector differentials). Using the split f = f− + f+ we get the QFTs of the split parts. Let b ∈ R2 be an arbitrary constant vector. F{b · ∇f− }(ω) = b · ωF{f− }(ω) j,
(4.2.15)
F{b · ∇f+ }(ω) = b · (U1 ω) F{f+ }(ω) j,
(4.2.16)
F{(U1 b) · ∇f+ }(ω) = b · ω F{f+ }(ω) j.
(4.2.17)
180 Quaternion and Clifford Fourier Transforms
Remark 4.2.6. The previously explained complex forms of the QFTs make the proof of Lemma 4.2.5 very similar to the case n = 2 of [185]. Noncommutativity must be duly taken into account! Lemma 4.2.7 (Schwarz inequality). Two quaternion functions g, h ∈ L2 (R2 , H) obey the following Schwarz inequality 1 g(x) d x h(x) d x ≥ 2 2 4 R R
Z
2 2
Z
2 2
Z
˜ g(x)h(x) + h(x)˜ g (x)d2 x
2
R2
Z
2 ˜ Sc(g(x)h(x))d x
=
2
(4.2.18)
.
R2
Remark 4.2.8. The proof of Lemma 4.2.7 can be based on the following inequality (0 < ∈ R) Z (4.2.19) [g(x) + h(x)][g(x) + h(x)]∼ d2 x ≥ 0. R2
4.2.1.3 Directional QFT uncertainty principle
In this section we state the directional QFT uncertainty principle and prove it step by step. Theorem 4.2.9 (Directional QFT UP). For two arbitrary constant vectors a = a1 e1 + a2 e2 ∈ R2 , b = b1 e1 + b2 e2 ∈ R2 (selecting two directions), and f ∈ L2 (R2 , H), x1/2 f ∈ L2 (R2 , H) we obtain Z
2
2 2
(a · x) f (x) d x
R2
Z
(b · ω)2 F{f }(ω)2 d2 ω
R2
i (2π)2 h (a · b)2 F−2 + (a · b0 )2 F+2 , 4
≥
(4.2.20)
with the energies Z
F± =
R2
b0 = U1 b = −b1 e1 + b2 e2 ∈ R2 .
f± (x)2 d2 x,
(4.2.21)
Proof. We now prove the directional QFT uncertainty principle by the following direct calculation. Z
2
2 2
(a · x) f (x) d x
Z
R2
(b · ω)2 F{f }(ω)2 d2 ω
R2 ±split
Z
2
=
ZR
2
(a · x) f− (x) d x + 2
R2
Lem.4.2.5h = ...
2 2
Z
2 2
R2
R2
2 2
Z R2
2
+ R2
0
2 2
(b · ω) F{f+ }(ω) d ω
F{(b · ∇)f− }(ω)(−j)2 d2 ω
Z
(a · x) f+ (x) d x
(b · ω) F{f− }(ω) d ω +
i Z
2
2 2
F{(b · ∇)f+ }(ω)(−j) d ω
Quaternion Fourier transforms 181
Parsev. =
Z
Z
2 2
R2
(a · x)f− (x) d x + 2
2 2
(b · ∇)f− (x) d x +
R2
Z R2
Z
+ R2
Z
+ R2
Z
+ R2
(a · x)f− (x)2 d2 x
Z
(a · x)f+ (x)2 d2 x
Z
(a · x)f− (x)2 d2 x
Z
2 2
(a · x)f+ (x) d x
"Z Schwarz ≤ (2π)2 R2
R2
Z
2 2
(b · ∇)f+ (x) d x
(b 0 · ∇)f+ (x)2 d2 x (b 0 · ∇)f+ (x)2 d2 x 2 2
R2
nZ R2
R2
(b · ∇)f− (x) d x
2 2 e Sc a · xf+ b · ∇f+ d x
Sc a · xf− b 0 · ∇fe+ d2 x {z

2
Sc a · xf+ b · ∇fe− d x {z

(Lem.4.1.13)
"Z
1 a · x b 0 · ∇(f+ fe+ )d2 x 2 R2
+
R2
nZ R2
b · ∇(a · x) f− 2 d2 x 
{z
}
=a ·b
{z
2 #
o2
b 0 · ∇(a · x) f+ 2 d2 x 
#
2
Z
Z
o2
}
1 a · x b · ∇(f− fe− )d2 x 2 2 R
+
o2
}
(Lem.4.1.13)
=0
Int. by p. (2π)2 hn = 4
0
=0
nZ
= (2π)
(b · ∇)f− (x)2 d2 x
R2
2
R2
R2
0
2 2 e Sc a · xf− b · ∇f− d x
+
+
R2
Z
Z
+
(a · x)f+ (x) d x
R2
Z
(2π) = (2π)2
2 2
0
o2 i
}
=a ·b i (2π)2 h = (a · b)2 F−2 + (a · b 0 )2 F+2 . 4
(4.2.22)
After using the Schwarz inequality of Lemma 4.2.7 we further used 1 1 Sc f− b · ∇fe− = b · ∇ f− fe− , Sc f+ b · ∇fe+ = b · ∇ f+ fe+ . 2 2
(4.2.23)
182 Quaternion and Clifford Fourier Transforms
4.2.1.4 Summary on directional QFT uncertainty principle
We first studied the ± split of quaternions and its effects on the doublesided QFT. Based on this we have established a new directional QFT uncertainty principle. Finally, see [255] for a strictly coordinate wise approach. 4.2.2 Quaternion Fourier transform and convolution 4.2.2.1 Background
In this treatment we use the general twosided quaternion Fourier transform (QFT), and relate the classical convolution of quaternionvalued signals over R2 with the Mustard convolution. A Mustard convolution can be expressed in the spectral domain as the point wise product of the QFTs of the factor functions. In full generality do we express the classical convolution of quaternion signals in terms of finite linear combinations of Mustard convolutions, and vice versa the Mustard convolution of quaternion signals in terms of finite linear combinations of classical convolutions. The treatment in this section is based on [213]. The twosided quaternionic Fourier transformation (QFT) of Section 4.1.1 was introduced in [124] for the analysis of 2D linear timeinvariant partialdifferential systems. Subsequently it has been further studied in [56] and applied in many fields, including color image processing [291], edge detection and image filtering [104, 293], watermarking [17], pattern recognition [268, 308], quaternionic multiresolution analysis (a generalization of discrete wavelets) [24], speech recognition [25], noise removal from video images [225], and efficient and robust image feature detection [153]. This led to further theoretical investigations [184, 187], where a special split of quaternions was introduced, then called ±split. An interesting physical consequence was that this split resulted in a left and right traveling multivector wave packet analysis, when generalizing the QFT to a full spacetime Fourier transform (SFT), compare Section 4.4. Later [194, 205, 212] this split has been analyzed further, interpreted geometrically and generalized to a freely steerable split of H into two orthogonal 2D analysis planes, then appropriately named orthogonal 2D planes split (OPS), see Section 4.1.2. A key strength of the classical complex Fourier transform is its easy and fast application to filtering problems. The convolution of a signal with its filter function becomes in the spectral domain a point wise product of the respective Fourier transformations. This is not the case for the convolution of two quaternionvalued signals over R2 , due to noncommutativity. Yet it is possible to define from the point wise product of the QFTs of two quaternion signals a new type of convolution, called Mustard convolution [270]. This Mustard convolution can be expressed in terms of sums of classical convolutions and vice versa. For the leftsided QFT this has recently been carried out in [97]. We expand this work in full generality to the twosided QFT, making significant and efficient use of the twosided orthogonal planes split of quaternions. This Section is organized as follows. Section 4.2.2.1.1 introduces quaternion commutators, notation and briefly reviews the general orthogonal twodimensional planes
Quaternion Fourier transforms 183
split of quaternions. Section 4.2.2.2 introduces to the general form of the twosided version of the quaternion Fourier transform, and a related mixed exponentialsine transform. Section 4.2.2.3 introduces the Mustard type convolutions based on the QFT and contains the main results of this paper. That is the formulation of the QFT of the classical convolution of quaternion signals in Theorem 4.2.17, and specialized to the simpler case of only one transform axis in Corollary 4.2.18. Moreover, the expression of the classical convolution of quaternion signals in terms of the Mustard type convolutions is given in Theorem 4.2.19, for only one transform axis in Corollary 4.2.21, and using only standard Mustard convolutions fully general and explicit in Theorem 4.2.22. Finally the Mustard convolution is fully expanded in terms of classical convolutions in Theorem 4.2.23. 4.2.2.1.1
Quaternions, commutators, notation orthogonal planes split
The quaternion
algebra and its properties are described in Section 2.8.1.
The commutator of any two quaternions p, q ∈ H is a pure quaternion (because S(pq) = S(qp)) defined as [p, q] = pq − qp. (4.2.24) For example, [i, j] = ij − ji = 2k.
(4.2.25)
The commutator [f, g] of any two pure unit quaternions f, g gives therefore another pure quaternion (or zero for f = αg, α ∈ R). It appears in the commutation of exponentials, e.g. eαf eβg = eβg eαf + [f, g] sin(α) sin(β), (4.2.26) which (in the same form for general multivector square roots of −1) has been used in [209] in order to derive a general convolution theorem for Clifford Fourier transformations. We furthermore note the useful anticommutation relationships g[f, g] = −[f, g]g,
f [f, g] = −[f, g]f,
(4.2.27)
eβg [f, g] = [f, g]e−βg .
(4.2.28)
and therefore eαf [f, g] = [f, g]e−αf ,
We will subsequently adopt the following notation for reflecting the argument of functions ([97], Notation 2.4, page 583) Notation 4.2.10 (Argument reflection). For a function h : R2 → H and a multiindex φ = (φ1 , φ2 ) with φ1 , φ2 ∈ {0, 1} we set hφ = h(φ1 ,φ2 ) (x) := h((−1)φ1 x1 , (−1)φ2 x2 ).
(4.2.29)
Review of general orthogonal twodimensional planes split (OPS) This section provides only a brief review and clarifies notation used in the rest of Section 4.2.2. For details we refer to Section 4.1.2. Assume in the following an arbitrary pair of pure unit quaternions f, g, f 2 = 4.2.2.1.2
184 Quaternion and Clifford Fourier Transforms
g 2 = −1. The orthogonal 2D planes split (OPS) is then defined with respect to any two pure unit quaternions f, g as Definition 4.2.11 (General orthogonal 2D planes split [205]). Let f, g ∈ H be an arbitrary pair of pure quaternions f, g, f 2 = g 2 = −1, including the cases f = ±g. The general OPS is then defined with respect to the two pure unit quaternions f, g as 1 q± = (q ± f qg). 2
(4.2.30)
f qg = q+ − q− ,
(4.2.31)
Note that i.e. under the map f ()g the q+ part is invariant, but the q− part changes sign. Both parts are twodimensional, and span two completely orthogonal planes. For f 6= ±g the q+ plane is spanned by two orthogonal quaternions {f − g, 1 + f g = −f (f − g)}, the q− plane is e.g. spanned by {f + g, 1 − f g = −f (f + g)}. For g = f a fully orthonormal fourdimensional basis of H is (R acts as rotation operator (rotor)) {1, f, j 0 , k0 } = R−1 {1, i, j, k}R,
R = i(i + f ),
(4.2.32)
and the two orthogonal twodimensional planes basis: q+ basis: {j 0 , k0 = f j 0 },
q− basis: {1, f }.
(4.2.33)
Note the notation for normed vectors in [261] {q1 , q2 , q3 , q4 } for the resulting total orthonormal basis of H. In (4.2.33), the q− part commutes with f and is also known as simplex part of q, whereas the q+ anticommutes with f , and −q+ j 0 is known as perplex part of q, see [127]. Lemma 4.2.12 (Orthogonality of two OPS planes [205]). Given any two quaternions q, p and applying the OPS with respect to any two pure unit quaternions f, g we get zero for the scalar part of the mixed products Sc(p+ q − ) = 0,
Sc(p− q + ) = 0.
(4.2.34)
Next we mention the possibility to perform a split along any given set of two (twodimensional) analysis planes. It has been found, that any twodimensional plane in R4 determines in an elementary way an OPS split and vice versa, compare Theorem 3.5 of [205]. Let us turn to the geometric interpretation of the map f ()g. It rotates the q− plane by 180◦ around the q+ axis plane. This is in perfect agreement with Coxeter’s notion of halfturn [84], see the right side of Figure 2.6. The following identities hold eαf q± eβg = q± e(β∓α)g = e(α∓β)f q± ,
(4.2.35)
as well as their quaternion conjugate (with α → −α, β → −β) eβg q± eαf = q± e(α∓β)f = e(β∓α)g q± .
(4.2.36)
Quaternion Fourier transforms 185
This leads to a straightforward geometric interpretation of the integrands of the quaternion Fourier transform (QFT or OPSQFT) with two pure quaternions f, g [205]. Particularly useful cases of (4.2.35) are (α, β) = (π/2, 0) and (0, π/2): f q± = ∓q± g,
q± g = ∓f q± .
(4.2.37)
We further note, that with respect to any pure unit quaternion f ∈ H, f 2 = −1, every quaternion A ∈ H can be similarly split into commuting and anticommuting parts [203]. Lemma 4.2.13 (Commuting and anticommuting with pure unit quaternion [203]). Every quaternion A ∈ H has, with respect to any pure unit quaternion f ∈ H, f 2 = −1, i.e. f −1 = −f, the unique decomposition denoted by 1 1 A+f = (A + f −1 Af ), A−f = (A − f −1 Af ) 2 2 A = A+f + A−f , A+f f = f A+f , A−f f = −f A−f .
(4.2.38)
Note, that in Lemma 4.2.13 the commuting part A+f is also known as simplex part of A, and the anticommuting part is up to a pure quaternion factor the perplex part of A, see [127]. 4.2.2.2 The general twosided QFT
Definition 4.2.14. (QFT with respect to two pure unit quaternions f, g [205]) Let f, g ∈ H, f 2 = g 2 = −1, be any two pure unit quaternions. The quaternion Fourier transform (QFT) with respect to f, g is F{h}(ω) = F f,g {h}(ω) =
Z
e−f x1 ω1 h(x) e−gx2 ω2 d2 x,
(4.2.39)
R2
where h ∈ L1 (R2 , H), d2 x = dx1 dx2 and x, ω ∈ R2 . The QFT can be inverted with h(x) = F −1 {h}(x) =
1 (2π)2
Z
ef x1 ω1 F{h}(ω) egx2 ω2 d2 ω,
(4.2.40)
R2
with d2 ω = dω1 dω2 . Remark 4.2.15. Note, that the general pair of pure unit quaternions f, g in Definition 4.2.14 includes orthogonal pairs f ⊥ g, nonorthogonal pairs f 6⊥ g, and parallel pairs f = ±g (only one transform axis). In the rest of this paper the theorems will be valid for fully general pairs f, g, if not otherwise specified. To avoid clutter we often omit the indication of the pair f, g as in F = F f,g . Linearity of the integral (4.2.39) allows us to use the OPS split h = h− + h+ F f,g {h}(ω) = F f,g {h− }(ω) + F f,g {h+ }(ω) = F−f,g {h}(ω) + F+f,g {h}(ω), (4.2.41) since by its construction the operators of the QFT F f,g , and of the OPS with respect to f, g commute. From (4.2.35) follows
186 Quaternion and Clifford Fourier Transforms
Theorem 4.2.16 (QFT of h± [205]). The QFT of the h± OPS split parts, with respect to two linearly independent unit quaternions f, g, of a quaternion module function h ∈ L1 (R2 , H) have the quasicomplex forms F±f,g {h} = F f,g {h± } Z
Z
−g(x2 ω2 ∓x1 ω1 ) 2
= R2
d x=
h± e
R2
e−f (x1 ω1 ∓x2 ω2 ) h± d2 x .
(4.2.42)
We further define for later use the following two mixed exponentialsine Fourier transforms F f,±s {h}(ω) =
Z R2
F ±s,g {h}(ω) =
Z R2
e−f x1 ω1 h(x)(±1) sin(−x2 ω2 )d2 x,
(4.2.43)
(±1) sin(−x1 ω1 )h(x)e−gx2 ω2 d2 x.
(4.2.44)
With the help of sin(−x1 ω1 ) =
g f −f x1 ω1 (e − ef x1 ω1 ), sin(−x2 ω2 ) = (e−gx2 ω2 − egx2 ω2 ), 2 2
(4.2.45)
we can rewrite the above mixed exponentialsine Fourier transforms in terms of the QFTs of Definition 4.2.14 as g g F f,±s {h} = ±(F f,g {h } − F f,−g {h }), 2 2 f f F ±s,g {h} = ±(F f,g { h} − F −f,g { h}). 2 2
(4.2.46) (4.2.47)
We further note the following useful relationships using the argument reflection of Notation 4.2.10 F −f,g {h} = F f,g {h(1,0) }, F f,−g {h} = F f,g {h(0,1) }, F −f,−g {h} = F f,g {h(1,1) },
(4.2.48)
and similarly F −s,g {h} = F s,g {h(1,0) }.
F f,−s {h} = F f,s {h(0,1) },
(4.2.49)
4.2.2.3 Convolution and Mustard convolution
We define the convolution of two quaternion signals a, b ∈ L1 (R2 ; H) as Z
(a ? b)(x) =
a(y)b(x − y)d2 y,
(4.2.50)
R2
provided that the integral exists. The Mustard convolution [62, 270] of two quaternion signals a, b ∈ L1 (R2 ; H) is defined as (a ?M b)(x) = (F f,g )−1 (F f,g {a}F f,g {b}) (4.2.51)
Quaternion Fourier transforms 187
provided that the integral exists. The Mustard convolution has the conceptual and computational advantage to simply yield as spectrum in the QFT Fourier domain the point wise product of the QFTs of the two signals, just as the classical complex Fourier transform. We additionally define a further type of exponentialsine Mustard convolution as (a ?M s b)(x) = (F f,g )−1 (F f,s {a}F s,g {b}).
(4.2.52)
In the following two subsections we will first express the convolution (4.2.50) in terms of the Mustard convolution (4.2.51) and vice versa. Expressing the convolution in terms of the Mustard convolution In [97] Theorem 4.1 on page 584 expresses the classical convolution of two quaternion functions with the help of the general leftsided QFT as a sum of 40 Mustard convolutions. In our approach we use Theorem 5.12 on page 327 of [209], which expresses the convolution of two Clifford signal functions (higher dimensional generalizations of quaternion functions) in the Clifford Fourier domain with the help of the general twosided Clifford Fourier transform (CFT), the latter is in turn a generalization of the QFT. We restate this theorem here again, specialized for quaternion functions and the QFT of Definition 4.2.14.
4.2.2.3.1
Theorem 4.2.17 (QFT of convolution). Assuming a general pair of unit pure quaternions f, g, the general twosided QFT of the convolution (4.2.50) of two functions a, b ∈ L1 (R2 ; H) can then be expressed as F f,g {a ? b} = F f,g {a+f }F f,g {b+g } + F f,−g {a+f }F f,g {b−g } + F f,g {a−f }F −f,g {b+g } + F f,−g {a−f }F −f,g {b−g } +F
f,s
{a+f }[f, g]F
s,g
{b+g } + F
f,−s
{a+f }[f, g]F
s,g
(4.2.53) {b−g }
+ F f,s {a−f }[f, g]F −s,g {b+g } + F f,−s {a−f }[f, g]F −s,g {b−g }. Note that due to the commutation properties of (4.2.43) and (4.2.44) we can place the commutator [f, g] also inside the exponentialsine transform terms as e.g. in F f,s {a+f }[f, g]F s,g {b+g } = F f,s {a+f [f, g]}F s,g {b+g } = F f,s {a+f }F s,g {[f, g]b+g }.
(4.2.54)
For the special case of parallel unit pure quaternions f = ±g, the commutator vanishes [f, g] = 0, and we get the following corollary. Note that in this case b±g = b±f . Corollary 4.2.18 (QFT of convolution with f k g). Assuming a parallel pair of unit pure quaternions f = ±g, the general twosided QFT of the convolution (4.2.50) of two functions a, b ∈ L1 (R2 ; H) can be expressed as F f,g {a ? b} = F f,g {a+f }F f,g {b+f } + F f,−g {a+f }F f,g {b−f } + F f,g {a−f }F −f,g {b+f } + F f,−g {a−f }F −f,g {b−f }.
(4.2.55)
188 Quaternion and Clifford Fourier Transforms
By applying the inverse QFT, we can now easily express the convolution of two quaternion signals a ? b, in terms of only eight Mustard convolutions (4.2.51) and (4.2.52). Theorem 4.2.19 (Convolution in terms of two types of Mustard convolution). Assuming a general pair of unit pure quaternions f, g, the convolution (4.2.50) of two quaternion functions a, b ∈ L1 (R2 ; H) can be expressed in terms of four Mustard convolutions (4.2.51) and four exponentialsine Mustard convolutions (4.2.52) as (0,1)
(1,0)
(0,1)
(1,0)
a ? b = a+f ?M b+g + a+f ?M b−g + a−f ?M b+g + a−f ?M b−g (0,1)
+ a+f ?M s [f, g]b+g + a+f ?M s [f, g]b−g (1,0)
(0,1)
(1,0)
+ a−f ?M s [f, g]b+g + a−f ?M s [f, g]b−g .
(4.2.56)
Remark 4.2.20. We use the convention, that terms such as a+f ?M s [f, g]b+g , should be understood with brackets a+f ?M s ([f, g]b+g ), which are omitted to avoid clutter. Assuming f k g, the standard Mustard convolution is sufficient to express the classical convolution. Corollary 4.2.21 (Convolution in terms of Mustard convolution with parallel axis). Assuming a parallel pair of unit pure quaternions f k g, the convolution (4.2.50) of two quaternion functions a, b ∈ L1 (R2 ; H) can be expressed in terms of four Mustard convolutions (4.2.51) as (0,1)
(1,0)
(0,1)
(1,0)
a ? b = a+f ?M b+f + a+f ?M b−f + a−f ?M b+f + a−f ?M b−f
(4.2.57)
Furthermore, applying (4.2.46) and (4.2.47), we can expand the terms in (4.2.53) with exponentialsine transforms into sums of products of QFTs. For example, the first term gives F f,s {a+f }[f, g]F s,g {b+g } 1 f,g F {a+f g} − F f,−g {a+f g} F f,g {f [f, g]b+g } − F −f,g {f [f, g]b+g } = 4 1 (1,0) F{a+f g}F{f [f, g]b+g } − F {a+f g}F{f [f, g]b+g } = 4 (0,1)
(0,1)
(1,0)
−F{a+f g}F{f [f, g]b+g } + F{a+f g}F{f [f, g]b+g } .
(4.2.58)
This now allows us in turn to express the quaternion signal convolution purely in terms of standard Mustard convolutions Theorem 4.2.22 (Convolution in terms of Mustard convolution). Assuming a general pair of unit pure quaternions f, g, the convolution (4.2.50) of two quaternion functions a, b ∈ L1 (R2 ; H) can be expressed in terms of twenty standard Mustard
Quaternion Fourier transforms 189
convolutions (4.2.51) as (0,1)
(1,0)
(0,1)
(1,0)
a ? b = a+f ?M b+g + a+f ?M b−g + a−f ?M b+g + a−f ?M b−g (1,0)
(0,1)
(0,1)
(1,0)
+14 ( a+f g ?M f cb+g − a+f g ?M f cb+g − a+f g ?M f cb+g + a+f g ?M f cb+g (0,1)
(0,1)
(1,0)
(1,0)
+a+f g ?M f cb−g − a+f g ?M f cb−g − a+f g ?M f cb−g + a+f g ?M f cb−g (1,0)
(0,1)
(1,0)
(0,1)
+a−f g ?M f cb+g − a−f g ?M f cb+g − a−f g ?M f cb+g + a−f g ?M f cb+g (0,1)
(1,0)
(0,1)
(1,0)
+a−f g ?M f cb−g − a−f g ?M f cb−g − a−f g ?M f cb−g + a−f g ?M f cb−g ), (4.2.59) with the abbreviation c = [f, g]. Now we will simply write out the Mustard convolution (4.2.51) and simplify it until only standard convolutions (4.2.50) remain. In this subsection, we will use the general OPS split of Definition 4.2.11. Our result should be compared with the Theorem 2.5 on page 584 of [97] for the leftsided QFT with 32 classical convolutions for expressing the Mustard convolution of quaternion functions (and for the twosided QFT in [62], Section 4.4.2, with 16 classical convolutions, stated in a different but apparently equivalent form to Theorem 4.2.23). We begin by writing the Mustard convolution (4.2.51) of two quaternion functions a, b ∈ L2 (R2 , H) 4.2.2.3.2
Expressing the Mustard convolution in terms of the convolution
1 (2π)2
a ?M b(x) = =
1 (2π)2
Z
Z
e−f x1 ω1
R2
ef x1 ω1 F{a}(ω)F{b}(ω)egx2 ω2 d2 ω
R2
Z
e−f y1 ω1 a(y)e−gy2 ω2 d2 y
R2
Z
e−f z1 ω1 b(z)e−gz2 ω2 d2 zegx2 ω2 d2 ω
R2
=
1 (2π)2
Z R2
Z R2
Z R2
ef (x1 −y1 )ω1 (a+ (y) + a− (y))e−gy2 ω2
e−f z1 ω1 (b+ (z) + b− (z))eg(x2 −z2 )ω2 d2 yd2 zd2 ω.
(4.2.60)
2 ω2 Next, we use the identities (4.2.35) in order to shift the Rinner factor e−gy to the R R RRR −f z1 ω1 left and e to the right, respectively. We abbreviate R2 R2 R2 to .
a ?M b(x) = =
1 (2π)2
+
1 (2π)2
+
1 (2π)2
+
1 (2π)2
(4.2.61)
ZZZ
ef (x1 −y1 )ω1 ef y2 ω2 a+ (y)b+ (z)egz1 ω1 eg(x2 −z2 )ω2 d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 )ω1 ef y2 ω2 a+ (y)b− (z)e−gz1 ω1 eg(x2 −z2 )ω2 d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 )ω1 e−f y2 ω2 a− (y)b+ (z)egz1 ω1 eg(x2 −z2 )ω2 d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 )ω1 e−f y2 ω2 a− (y)b− (z)e−gz1 ω1 eg(x2 −z2 )ω2 d2 yd2 zd2 ω
Furthermore, we abbreviate the inner function products as ab±± (y, z) := a± (y)b± (z), and apply the general OPS split of Definition 4.2.11 once again to obtain
190 Quaternion and Clifford Fourier Transforms
ab±± (y, z) = [ab±± (y, z)]+ + [ab±± (y, z)]− = ab±± (y, z)+ + ab±± (y, z)− . We omit the square brackets and use the convention that the final OPS split indicated by the final ± index should be performed last. This allows to further apply (4.2.35) again in order to shift the factors e±gz1 ω1 eg(x2 −z2 )ω2 to the left. We end up with the following eight terms a ?M b(x) = =
1 (2π)2
1 + (2π) 2 1 + (2π) 2 1 + (2π) 2 1 + (2π) 2 1 + (2π) 2 1 + (2π) 2 1 + (2π) 2
ZZZ
ef (x1 −y1 −z1 )ω1 ef (y2 −(x2 −z2 ))ω2 ab++ (y, z)+ d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 +z1 )ω1 ef (y2 +(x2 −z2 ))ω2 ab++ (y, z)− d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 +z1 )ω1 ef (y2 −(x2 −z2 ))ω2 ab+− (y, z)+ d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 −z1 )ω1 ef (y2 +(x2 −z2 ))ω2 ab+− (y, z)− d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 −z1 )ω1 ef (−y2 −(x2 −z2 ))ω2 ab−+ (y, z)+ d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 +z1 )ω1 ef (−y2 +(x2 −z2 ))ω2 ab−+ (y, z)− d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 +z1 )ω1 ef (−y2 −(x2 −z2 ))ω2 ab−− (y, z)+ d2 yd2 zd2 ω
ZZZ
ef (x1 −y1 −z1 )ω1 ef (−y2 +(x2 −z2 ))ω2 ab−− (y, z)− d2 yd2 zd2 ω.
(4.2.62)
We now only show explicitly how to simplify the second triple integral, the others follow the same pattern. ZZZ
1 (2π)2
=
1 (2π)2
ef (x1 −y1 +z1 )ω1 ef (y2 +(x2 −z2 ))ω2 [a+ (y)b+ (z)]− d2 yd2 zd2 ω
ZZ Z
ef (x1 −y1 +z1 )ω1 dω1 R
Z
ef (y2 +(x2 −z2 ))ω2 dω2 [a+ (y)b+ (z)]− d2 yd2 z
R
ZZ
δ(x1 − y1 + z1 )δ(y2 + (x2 − z2 ))[a+ (y)b+ (z1 , z2 )]− d2 yd2 z
Z
[a+ (y)b+ (−(x1 − y1 ), x2 + y2 )]− d2 y
= = R2
Z
= R2
Z
= R2
[a+ (y)b+ (−(x1 − y1 ), −(−x2 − y2 ))]− d2 y (1,1)
[a+ (y)b+ (x1 − y1 , −x2 − y2 )]− d2 y (1,1)
= [a+ ? b+ (x1 , −x2 )]− . (1,1)
(4.2.63)
Note that a+ ? b+ (x1 , −x2 ) means to first apply the convolution to the pair of (1,1) functions a+ and b+ , and only then to evaluate them with the argument (x1 , −x2 ). (1,1) So in general a+ ? b+ (x1 , −x2 ) 6= a+ ? b+ (−x1 , x2 ). Simplifying the other seven triple integrals similarly we finally obtain the desired decomposition of the Mustard convolution (4.2.51) in terms of the classical convolution.
Quaternion Fourier transforms 191
Theorem 4.2.23 (Mustard convolution in terms of standard convolution). The Mustard convolution (4.2.51) of two quaternion functions a, b ∈ L1 (R2 ; H) can be expressed in terms of eight standard convolutions (4.2.50) as a ?M b(x) = (1,1)
= [a+ ? b+ (x)]+ + [a+ ? b+ (x1 , −x2 )]− (1,0)
(0,1)
+ [a+ ? b− (x)]+ + [a+ ? b− (x1 , −x2 )]− (0,1)
(1,0)
+ [a− ? b+ (x1 , −x2 )]+ + [a− ? b+ (x)]− (1,1)
+ [a− ? b− (x1 , −x2 )]+ + [a− ? b− (x)]− .
(4.2.64)
Remark 4.2.24 (Levels of computation). Equation (4.2.64) involves five levels of operation: primary (inner) and secondary (outer) OPS, argument reflections before (pre) and after (post) the actual convolution, and the convolution itself in each of the eight terms. Remark 4.2.25 (Efficiency of notation and interpretation). If we would explicitly insert according Definition 4.2.11 a± = 12 (a±f ag) and b± = 12 (b±f bg), and similarly explicitly insert the second level OPS split [. . .]± , we would (potentially) obtain up to a maximum of 64 terms22 . It is therefore obvious how straightforward, significant and efficient the use of the general OPS split is in this context. Efficiency first of all with respect to concise (compact) notation, which in turn may assist geometric (or physical) interpretation in concrete applications. Regarding derivation efficiency, we needed two pages to derive (4.2.64), but in order to strike the balance with sufficient detail for reasonably selfcontained comprehension, this level of detail may be justified. Whether the compact eight term form of (4.2.64) is advantageous for actual numerical computations, is an open question, which requires application to concrete representatives problems, e.g. in the area of quaternionic image processing (see e.g. [17,25,56,104,127,153,225,268,291,293,308]). And it may depend on the concrete hardware architecture, e.g. how many parallel channels of computation (e.g. number of parallel GPUs) are available, whether quaternion operations are hard coded or need breaking down in elementary real (or matrix) computations; and whether optimized software packages like the precompiler GAALOP [167] or the MATLAB package QFMT [297] would be used, etc. 4.2.2.4 Summary on QFT and convolution
In Section 4.2.2, we have briefly reviewed the general orthogonal twodimensional planes split of quaternions, the general twosided quaternion Fourier transform, and introduced a related mixed quaternionic exponentialsine Fourier transform. We defined the notions of convolution of two quaternion valued functions over R2 , the Mustard convolution (with its QFT as the point wise product of the QFTs of the factor functions), and a special Mustard convolution involving the point wise products 22
Note that in [62], Section 4.4.2, a fully explicit form is given with only 16 terms.
192 Quaternion and Clifford Fourier Transforms
of mixed exponentialsine QFTs. The main results are: an efficient decomposition of the classical convolution of quaternion signals in terms of eight Mustard type convolutions. For the special case of parallel pure unit quaternion axis in the QFT, only four terms of the standard Mustard convolution are sufficient. Even in the case of two general pure unit quaternion axis in the QFT, the classical convolution of two quaternion signals can always be fully expanded in terms of standard Mustard convolutions. Finally we showed how to fully generally expand the Mustard convolution of two quaternion signals in terms of eight classical convolutions. In view of the many applications of the QFT explained in the introduction, we expect the results described to be of great interest, especially for filter design and feature extraction in signal and color image processing. 4.2.3 Quaternionic Wiener Khintchine type theorems and convolution 4.2.3.1 Background
In this part we use the general twosided quaternion Fourier transform (QFT) of Section 4.1.1 as generalized in Section 4.1.2, in order to derive WienerKhintchine type theorems for the crosscorrelation and for the autocorrelation of quaternion signals. Furthermore, we show how to derive a new four term spectral representation for the convolution of quaternion signals. The treatment in this section is based on [214]. The twosided quaternionic Fourier transformation (QFT) was introduced in [124] for the analysis of 2D linear timeinvariant partialdifferential systems. Subsequently it has been further studied in [56] and applied in many fields, including color image processing [291], edge detection and image filtering [104, 293], watermarking [17], pattern recognition [268, 308], quaternionic multiresolution analysis (a generalization of discrete wavelets) [24], speech recognition [25], noise removal from video images [225], and efficient and robust image feature detection [153]. This led to further theoretical investigations [184, 187], where a special split of quaternions was introduced, then called ±split. An interesting physical consequence was that this split resulted in a left and right traveling multivector wave packet analysis, when generalizing the QFT to a full spacetime Fourier transform (SFT). Later [194, 205, 212] this split has been analyzed further, interpreted geometrically and generalized to a freely steerable split of H into two orthogonal 2D analysis planes, then appropriately named orthogonal 2D planes split (OPS). A hypercomplex (quaternionic) WienerKhintchine theorem for a pair of discrete quaternionic signals applied to color image correlation can be found in [126, 267], using a rightsided twodimensional quaternionic Fourier transform. In our contribution we generalize the WienerKhintchine theorem to continuous quaternion signals, using the general steerable twosided QFT (Theorem 4.2.27). Furthermore a corollary (Corollary 4.2.28) presents the autocorrelation of a quaternion signal. Finally we use a similar approach to derive a new four term spectral representation of the convolution of two quaternion signals (Theorem 4.2.29). This Section is organized as follows. The following paragraphs introduce quaternions and the twodimensional orthogonal planes split of quaternions. Then
Quaternion Fourier transforms 193
Section 4.2.2.2 explains the general steerable twosided quaternion Fourier transform. Section 4.2.3.2 defines the notions of convolution, crosscorrelation and autocorrelation applied to quaternion signal functions. Section 4.2.3.3 derives the WienerKhintchine theorem for quaternion signals based on the steerable twosided QFT, which gives quaternionic spectral representations for the crosscorrelation and the autocorrelation of quaternion signals. Finally, Section 4.2.3.4 derives a new spectral representation of the convolution of two quaternion signals. For extensive treatment of quaternions, we refer to Section 2.8.1. Here we only provide some extra notation relevant to Section 4.2.3. Regarding quaternion commutators, we refer to Section 4.2.2.1.1. Also the argument reflection Notation 4.2.10 should be noted. We further recall, that with respect to any pure unit quaternion f ∈ H, f 2 = −1, every quaternion A ∈ H can be similarly split into commuting and anticommuting parts [203], see Lemma 4.2.13. Remark 4.2.26. Note, that the general pair of pure unit quaternions f, g in the general Definition 4.2.14 of a quaternionic FT includes orthogonal pairs f ⊥ g, nonorthogonal pairs f 6⊥ g, and parallel pairs f = ±g (only one transform axis). In the rest of this section the theorems will be valid for fully general pairs f, g, if not otherwise specified. To avoid clutter we often omit the indication of the pair f, g as in F = F f,g . We further note the following useful relationships using the argument reflection of Notation 4.2.10 F −f,g {h} = F f,g {h(1,0) }, F f,−g {h} = F f,g {h(0,1) }, F −f,−g {h} = F f,g {h(1,1) }.
(4.2.65)
4.2.3.2 Convolution, crosscorrelation and autocorrelation of quaternion signals
We define the convolution of two quaternion signals a, b ∈ L1 (R2 ; H) as Z
2
a(y)b(x − y)d y =
(a ? b)(x) = R2
Z
a(x − y)b(y)d2 y,
(4.2.66)
R2
provided that the integral exists. We define the crosscorrelation of two quaternion signals a, b ∈ L1 (R2 ; H) as Z
(a ?c b)(x) =
2
a(y)b(y − x)d y =
R2
Z
a(y + x)b(y)d2 y,
(4.2.67)
R2
provided that the integral exists. We define the autocorrelation of a quaternion signal a ∈ L1 (R2 ; H) as Z
(a ?a a)(x) =
2
a(y)a(y − x)d y =
R2
provided that the integral exists.
Z R2
a(y + x)a(y)d2 y,
(4.2.68)
194 Quaternion and Clifford Fourier Transforms
4.2.3.3 The WienerKhintchine theorem for quaternion signals using the steerable twosided QFT
For establishing the WienerKhintchine theorem for quaternion signals using the steerable twosided QFT we begin with the cross correlation (4.2.67) and apply the inverse twosided QFT of (4.1.10), where we set bb = F f,g {b}: Z
(a ?c b)(x)
a(y)b(y − x)d2 y
= R2
Z
a(y)F −1 {bb}(y − x)d2 y
= R2
= = = =
Z
Z
1 (2π)2
ef (y1 −x1 )ω1 bb(ω)eg(y2 −x2 )ω2 d2 ωd2 y
a(y) R2
R2
Z
1 (2π)2
Z
a(y) R2
Z
1 (2π)2
Z
R2
a(y)e−gy2 ω2 egx2 ω2 bb(ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω
R2
Z
1 (2π)2
e−g(y2 −x2 )ω2 bb(ω)e−f (y1 −x1 )ω1 d2 yd2 ω
R2
Z
R2
R2
[a− (y) + a+ (y)]
e−gy2 ω2 egx2 ω2 [ bb− (ω) + bb+ (ω)]e−f y1 ω1 ef x1 ω1 d2 yd2 ω =
Z
1 (2π)2
+ + +
Z
a− (y)e−gy2 ω2 egx2 ω2 bb− (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω
R2
R2
Z
Z
1 (2π)2 1 (2π)2 1 (2π)2
R2
Z R2
Z R2
R2
Z R2
Z R2
a− (y)e−gy2 ω2 egx2 ω2 bb+ (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω a+ (y)e−gy2 ω2 egx2 ω2 bb− (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω a+ (y)e−gy2 ω2 egx2 ω2 bb+ (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω. (4.2.69)
In the following we transform the last four integrals I1 , I2 , I3 , I4 of (4.2.69) one by f,g one, applying (4.2.35) and (4.2.36) repeatedly, and setting ac {a± }. ± =F I1 =
1 (2π)2
=
1 (2π)2
= = =
1 (2π)2 1 (2π)2 1 (2π)2
Z R2
Z R2
Z R2
Z R2
Z R2
Z R2
Z R2
Z R2
Z R2
a− (y)e−gy2 ω2 egx2 ω2 bb− (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω a− (y)e−gy2 ω2 egx2 ω2 e−gy1 ω1 bb− (ω)ef x1 ω1 d2 yd2 ω a− (y)e−gy1 ω1 e−gy2 ω2 d2 yegx2 ω2 bb− (ω)ef x1 ω1 d2 ω e−f y1 ω1 a− (y)e−gy2 ω2 d2 ybb− (ω)ef (x1 ω1 +x2 ω2 ) d2 ω
f x·ω 2 b ac d ω. − (ω) b− (ω)e
(4.2.70)
Quaternion Fourier transforms 195
Computing the second integral we further apply Notation 4.2.10 and similarly get I2 =
1 (2π)2
=
1 (2π)2
= = = =
1 (2π)2 1 (2π)2 1 (2π)2 1 (2π)2
Z
Z
R2
Z
R2
Z
R2
Z
R2
Z
R2
Z
R2
Z
R2
Z R2
Z R2
R2
a− (y)e−gy2 ω2 egx2 ω2 bb+ (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω a− (y)e−gy2 ω2 egx2 ω2 e+gy1 ω1 bb+ (ω)ef x1 ω1 d2 yd2 ω a− (y)e+gy1 ω1 e−gy2 ω2 d2 yegx2 ω2 bb+ (ω)ef x1 ω1 d2 ω e+f y1 ω1 a− (y)e−gy2 ω2 d2 ybb+ (ω)ef (x1 ω1 −x2 ω2 ) d2 ω
f (x1 ω1 −x2 ω2 ) 2 b d ω ac − (−ω1 , ω2 ) b+ (ω)e (1,0) ac (ω) bb+ (ω)ef (x1 ω1 −x2 ω2 ) d2 ω. −
(4.2.71)
Computing the third integral we get I3 =
1 (2π)2
=
1 (2π)2
= = = =
1 (2π)2 1 (2π)2 1 (2π)2 1 (2π)2
Z
Z
R2
Z
R2
Z
R2
Z
R2
Z
R2
R2
Z
Z
R2
R2
Z R2
Z R2
a+ (y)e−gy2 ω2 egx2 ω2 bb− (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω a+ (y)e−gy2 ω2 egx2 ω2 e−gy1 ω1 bb− (ω)ef x1 ω1 d2 yd2 ω a+ (y)e−gy1 ω1 e−gy2 ω2 d2 yegx2 ω2 bb− (ω)ef x1 ω1 d2 ω e+f y1 ω1 a+ (y)e−gy2 ω2 d2 ybb− (ω)ef (x1 ω1 +x2 ω2 ) d2 ω
f x·ω 2 b ac d ω + (−ω1 , ω2 ) b− (ω)e (1,0) ac (ω) bb− (ω)ef x·ω d2 ω. +
(4.2.72)
Computing the fourth integral we get I4 =
1 (2π)2
=
1 (2π)2
= = =
1 (2π)2 1 (2π)2 1 (2π)2
Z
Z
R2
R2
Z
Z
R2
Z R2
Z R2
Z R2
R2
Z R2
Z R2
a+ (y)e−gy2 ω2 egx2 ω2 bb+ (ω)e−f y1 ω1 ef x1 ω1 d2 yd2 ω a+ (y)e−gy2 ω2 egx2 ω2 e+gy1 ω1 bb+ (ω)ef x1 ω1 d2 yd2 ω
a+ (y)e+gy1 ω1 e−gy2 ω2 d2 yegx2 ω2 bb+ (ω)ef x1 ω1 d2 ω e−f y1 ω1 a+ (y)e−gy2 ω2 d2 ybb+ (ω)ef (x1 ω1 −x2 ω2 ) d2 ω
f (x1 ω1 −x2 ω2 ) 2 b ac d ω. + (ω) b+ (ω)e
(4.2.73)
196 Quaternion and Clifford Fourier Transforms
Collecting the results of the four integrals I1 , I2 , I3 , I4 , we obtain two relevant forms of the crosscorrelation (a ?c b)(x) =
1 (2π)2
Z R2
(1,0) gx2 ω2 b [ac b− (ω)ef x1 ω1 + ac (ω) egx2 ω2 bb+ (ω)ef x1 ω1 − (ω) e −
(1,0) gx2 ω2 b b+ (ω)ef x1 ω1 ]d2 ω + ac (ω) egx2 ω2 bb− (ω)ef x1 ω1 + ac + (ω) e +
=
1 (2π)2
Z
R2
(1,0) c [ac (ω)]egx2 ω2 bb− (ω)ef x1 ω1 + − (ω) + a +
(1,0) gx2 ω2 b + [ac (ω) + ac b+ (ω)ef x1 ω1 d2 ω − + (ω)]e
=
1 (2π)2
Z
R2
(1,0) c [ac (ω)]e−f x1 ω1 bb− (ω)e−gx2 ω2 + − (ω) + a +
(1,0) −f x1 ω1 b + [ac (ω) + ac b+ (ω)e−gx2 ω2 d2 ω, − + (ω)]e
(4.2.74)
and alternatively (a ?c b)(x) =
1 (2π)2
Z R2
(1,0) c [ac (ω)]bb− (ω)ef x·ω − (ω) + a +
(1,0) f (x1 ω1 −x2 ω2 ) b + [ac (ω) + ac d2 ω. − + (ω)]b+ (ω)e
(4.2.75)
We note, that the last form appears most similar to the conventional complex WienerKhintchine theorem. Formal identity with the conventional complex case is obtained b± = ac for a+ = b+ = 0. Setting b = a, and noting that with Theorem 4.2.16 a ± , we get important identities for the autocorrelation (a ?a a)(x) =
1 (2π)2
+ =
1 (2π)2
+ =
1 (2π)2
Z
Z
Z
(1,0) c b− (ω)e−gx2 ω2 + [ac (ω)]e−f x1 ω1 a − (ω) + a + R2 (1,0) −f x1 ω1 a −gx2 ω2 d2 ω c b [ac (ω) + a (ω)]e (ω)e − + + (1,0) c b− (ω)ef x·ω [ac (ω)]a − (ω) + a + R2 (1,0) f (x1 ω1 −x2 ω2 ) b c [ac (ω) + a (ω)] a (ω)e d2 ω + − +
R2
(1,0) 2 f x·ω c [ac (ω)ac − (ω) + a + − (ω)]e
(1,0) 2 f (x1 ω1 −x2 ω2 ) c + [ac (ω)ac d2 ω. − + (ω) + a + (ω) ]e
(4.2.76)
Formal identity with the conventional complex autocorrelation is obtained for a+ = 0. We summarize our results in one theorem and one corollary. Theorem 4.2.27 (WienerKhintchine theorem for crosscorrelation of quaternionic signals). The crosscorrelation of two quaternionic signals a, b ∈ L1 (R2 ; H) can be expressed as
Quaternion Fourier transforms 197
(a ?c b)(x) =
1 (2π)2
Z
R2
(1,0) c [ac (ω)]egx2 ω2 bb− (ω)ef x1 ω1 + − (ω) + a +
(1,0) gx2 ω2 b b+ (ω)ef x1 ω1 d2 ω + [ac (ω) + ac − + (ω)]e
=
1 (2π)2
Z
R2
(1,0) c [ac (ω)]e−f x1 ω1 bb− (ω)e−gx2 ω2 + − (ω) + a +
(1,0) −f x1 ω1 b + [ac (ω) + ac b+ (ω)e−gx2 ω2 d2 ω − + (ω)]e
=
1 (2π)2
Z
R2
(1,0) c [ac (ω)]bb− (ω)ef x·ω − (ω) + a +
(1,0) f (x1 ω1 −x2 ω2 ) b + [ac (ω) + ac d2 ω. − + (ω)]b+ (ω)e
(4.2.77)
Corollary 4.2.28 (WienerKhintchine theorem for autocorrelation of quaternionic signals). The crosscorrelation of a quaternionic signal a ∈ L1 (R2 ; H) can be expressed as (a ?a a)(x) =
1 (2π)2
Z
R2
(1,0) c b− (ω)e−gx2 ω2 + [ac (ω)]e−f x1 ω1 a − (ω) + a +
(1,0) −f x1 ω1 a b+ (ω)e−gx2 ω2 d2 ω + [ac (ω) + ac − + (ω)]e
=
1 (2π)2
+
Z
(1,0) f x·ω 2 c [ac (ω)ac − (ω)]e − (ω) + a + R2 (1,0) 2 f (x1 ω1 −x2 ω2 ) c d2 ω. [ac (ω)ac + (ω) + a + (ω) ]e −
(4.2.78)
4.2.3.4 New spectral representation for convolution of quaternion signals
It is known, that due to the noncommutativity of quaternion signals and the quaternionic kernels of the QFT, the spectral representation of the convolution may include many terms. See e.g. Theorem 4.6 in [213] for a representation involving as many as twenty terms. But the expressions in (4.2.66) and (4.2.67) for convolution and crosscorrelation of two quaternion signals are related by (a ? b)(x) = (a(1,1) ?c b)(x).
(4.2.79)
This indeed raises the expectation of finding similar to Theorem 4.2.27 a new short and compact spectral representation for the convolution by reflecting the argument of the first signal a → a(1,1) and inserting the quaternion conjugate of the second signal b → b. But in doing this all operations in the derivation of Theorem 4.2.27 would have to be carefully checked for their validity, which is not as trivial as it may seem at first. Therefore starting all over with the very definition of the convolution of two quaternion signals (4.2.66) seems to be better and clearly reveals why the result is both similar, yet also characteristically different from the above crosscorrelation result.
198 Quaternion and Clifford Fourier Transforms
In the following and throughout the current section we will apply two different splits to the signal functions a and b, respectively defined as a± = 12 (a ± gaf ),
b± = 12 (b ± f bg),
(4.2.80)
note in particular the different positions of f and g. Two different splits are needed, because in the convolution the second signal is not quaternion conjugated. To avoid clutter we will use the notation bb = F f,g {b}, and for the inverse transform we use a f,g lower index: F−1 . Z
a(y)b(x − y)d2 y
(a ? b)(x) = R2
Z
f,g b a(y)F−1 {b}(x − y)d2 y
= R2
Z
Z
=
a(y) Z
Z
a(y)e−f y1 ω1 ef x1 ω1 bb(ω)e−gy2 ω2 egx2 ω2 d2 ωd2 y
= R2
R2
Z
=
ef (x1 −y1 )ω1 bb(ω)eg(x2 −y2 )ω2 d2 ωd2 y
R2
R2
Z
[a− (y) + a+ (y)]e−f y1 ω1 ef x1 ω1 [bb− (ω) R2 R 2 −gy2 ω2 gx2 ω2 2 2 e
e
Z
Z
R2
R2
Z
Z
+ Z
R2
Z
+ R2
Z
R2
Z
+ R2
Z
R2
Z
R2
R2
Z
Z
+ Z
R2
Z
+ R2
Z
R2
Z
+ R2
a− (y)e−f y1 ω1 ef x1 ω1 bb+ (ω)e−gy2 ω2 egx2 ω2 d2 ωd2 y a+ (y)e−f y1 ω1 ef x1 ω1 bb− (ω)e−gy2 ω2 egx2 ω2 d2 ωd2 y a+ (y)e−f y1 ω1 ef x1 ω1 bb+ (ω)e−gy2 ω2 egx2 ω2 d2 ωd2 y
e−gy1 ω1 a− (y)e−f y2 ω2 d2 yef x1 ω1 bb− (ω)egx2 ω2 d2 ω
= R2
d ωd y
a− (y)e−f y1 ω1 ef x1 ω1 bb− (ω)e−gy2 ω2 egx2 ω2 d2 ωd2 y
= R2
+ bb+ (ω)]
R2
e−gy1 ω1 a− (y)ef y2 ω2 d2 yef x1 ω1 bb+ (ω)egx2 ω2 d2 ω egy1 ω1 a+ (y)e−f y2 ω2 d2 yef x1 ω1 bb− (ω)egx2 ω2 d2 ω egy1 ω1 a+ (y)ef y2 ω2 d2 yef x1 ω1 bb+ (ω)egx2 ω2 d2 ω
Quaternion Fourier transforms 199
Z
F g,f {a− }(ω)ef x1 ω1 bb− (ω)egx2 ω2 d2 ω
= R2
Z
(0,1)
+ R2
Z
(1,0)
+ R2
Z
F g,f {a+ }(ω)ef x1 ω1 bb− (ω)egx2 ω2 d2 ω (1,1)
+ R2
Z
F g,f {a− }(ω)ef x1 ω1 bb+ (ω)egx2 ω2 d2 ω
F g,f {a+ }(ω)ef x1 ω1 bb+ (ω)egx2 ω2 d2 ω (1,0)
F g,f {a− + a+ }(ω)ef x1 ω1 bb− (ω)egx2 ω2 d2 ω
= R2
Z
(0,1)
+ R2
F g,f {a−
(1,1)
+ a+ }(ω)ef x1 ω1 bb+ (ω)egx2 ω2 d2 ω,
(4.2.81)
where we repeatedly applied that bb± exp(βg) = exp(∓βf )bb± , and a± exp(αf ) = exp(∓αg)a± , and note the use of F g,f {a± }. We therefore obtain the following theorem. Theorem 4.2.29 (Spectral representation of convolution of quaternion signals). The convolution of two quaternionic signals a, b ∈ L1 (R2 ; H) can be expressed as Z
(1,0)
F g,f {a− + a+ }(ω)ef x1 ω1 bb− (ω)egx2 ω2 d2 ω
(a ? b)(x) = R2
Z
+ R2
(0,1)
F g,f {a−
(1,1)
+ a+ }(ω)ef x1 ω1 bb+ (ω)egx2 ω2 d2 ω,
(4.2.82)
where bb = F f,g {b}, a± = 12 (a ± gaf ), and b± = 21 (b ± f bg). 4.2.3.5 Summary on QFT WienerKhintchine theorems
In Section 4.2.3, we defined the notions of convolution and crosscorrelation of two quaternion valued functions over R2 , and the autocorrelation of a quaternion valued function over R2 . The main results are a WienerKhintchine theorem for the crosscorrelation of two quaternion signals, based on the steerable twodimensional QFT, and as a corollary for the autocorrelation. Furthermore we obtain a new four term long spectral representation of the convolution of two quaternion signals. In view of the many applications of the QFT referred to in Section 4.2.3.1, we expect these results to be of great interest, especially for filter design and feature extraction in signal and color image processing.
4.3
SPECIAL QUATERNION FOURIER TRANSFORMS
4.3.1 4.3.1.1
Windowed quaternion Fourier transform (QWFT) Background
In this section, we generalize the classical windowed Fourier transform (WFT) to quaternionvalued signals, calling it the quaternionic windowed Fourier transform (QWFT). Using the spectral representation of the quaternionic Fourier transform
200 Quaternion and Clifford Fourier Transforms
(QFT), we derive several important properties such as reconstruction formula, reproducing kernel, isometry and orthogonality relation. Taking the Gaussian function as window function we obtain quaternionic Gabor filters which play the role of coefficient functions when decomposing the signal in the quaternionic Gabor basis. We apply the QWFT properties and the (rightsided) QFT to establish a Heisenberg type uncertainty principle for the QWFT. Finally, we briefly introduce an application of the QWFT to a linear timevarying system. This section is based on [256], but for lack of space and because most proofs are fairly elementary, we omit proofs. The interested reader is invited to try the proofs himself, or simply look them up in [256]. One of the basic problems encountered in signal representations using the conventional Fourier transform (FT) is the ineffectiveness of the Fourier kernel to represent and compute location information. One method to overcome such a problem is the windowed Fourier transform (WFT). Some authors [147, 148, 318] have extensively studied the WFT and its properties from a mathematical point of view. In [234, 337] the WFT has been successfully applied as a tool of spatialfrequency analysis which is able to characterize the local frequency at any location in a fringe pattern. On the other hand, the quaternionic Fourier transform (QFT), which, as we have seen, is a nontrivial generalization of the real and complex Fourier transform (FT) using the quaternion algebra [156], has been of interest to researchers for some years (see, for example, [25, 56,58, 124, 184, 187, 255, 280]). They found that many FT properties still hold and others have to be modified. Based on the (rightsided) QFT, one may extend the WFT to quaternion algebra while enjoying similar properties as in the classical case. The idea of extending the WFT to the quaternion algebra setting has already been studied by B¨ ulow and Sommer [56, 58]. They introduced a special case of the QWFT known as quaternionic Gabor filters. They applied these filters to obtain a local twodimensional quaternionic phase. Their generalization is obtained using the inverse (twosided) quaternion Fourier kernel. Hahn [154] constructed a FourierWigner distribution of 2D quaternionic signals which is in fact closely related to the QWFT. In [254], the extension of the WFT to Clifford (geometric) algebra was discussed. This extension used the kernel of the Clifford Fourier transform (CFT) [185]. In general a CFT replaces the complex imaginary unit i ∈ C by a geometric root [193, 294] of −1, i.e. any element of a Clifford (geometric) algebra squaring to −1. The main goal of this section is to thoroughly study the generalization of the classical WFT to quaternion algebra, which we call the quaternionic windowed Fourier transform (QWFT), and investigate important properties of the QWFT such as (specific) shift, reconstruction formula, reproducing kernel, isometry and orthogonality relation. We emphasize that the QWFT proposed in the present work is significantly different from [254] in the definition of the exponential kernel. In the present approach, we use the kernel of the (rightsided) QFT. We present several examples to show the differences between the QWFT and the WFT. Using the (rightsided) QFT properties and its uncertainty principle [255] we establish a generalized QWFT uncertainty principle. We will also study an application of the QWFT to a linear timevarying system.
Quaternion Fourier transforms 201
The organization of the treatment is as follows. In the remainder of this background section we briefly describe notation and review the (rightsided) QFT. In Section 4.3.1.2, we discuss the basic ideas for the construction of the QWFT and derive several important properties of the QWFT using the (rightsided) QFT, including the Heisenberg uncertainty principle in Section 4.3.1.3 We also give some examples of the QWFT. In Section 4.3.1.4, an application of the QWFT to a linear time varying system is presented. We will make use of an inner product for two functions f, g : R2 −→ H as follows: hf, giL2 (R2 ;H) =
Z
f (x)g(x) d2 x,
(4.3.1)
R2
where the overline indicates the quaternion conjugation of the function. In particular, if f = g, we obtain the associated norm kf kL2 (R2 ;H) =
1/2 hf, f iL2 (R2 ;H)
Z
2
2
f (x) d x
=
1/2
.
(4.3.2)
R2
As a consequence of the inner product (4.3.1) we obtain the quaternion CauchySchwarz inequality Schf, gi ≤ kf kL2 (R2 ;H) kgkL2 (R2 ;H) ,
∀f, g ∈ L2 (R2 ; H).
(4.3.3)
Now we recall the (rightsided) QFT needed later in section 4.3.1.2 to establish the QWFT. Definition 4.3.1 (Rightsided QFT). The (right sided) quaternion Fourier transform (QFT) of f ∈ L1 (R2 ; H) is the function23 Fq {f }: R2 → H given by Fq {f }(ω) =
Z
f (x)e−iω1 x1 e−j ω2 x2 d2 x,
(4.3.4)
R2
where x = x1 e1 + x2 e2 , ω = ω1 e1 + ω2 e2 , and the quaternion exponential product e−iω1 x1 e−j ω2 x2 is the quaternion Fourier kernel. Theorem 4.3.2 (Inverse QFT). Suppose that f ∈ L2 (R2 ; H) and Fq {f } ∈ L1 (R2 ; H). Then the QFT of f is an invertible transform and its inverse is given by Z 1 −1 Fq [Fq {f }](x) = f (x) = Fq {f }(ω)ej ω2 x2 eiω1 x1 d2 ω, (4.3.5) (2π)2 R2 where the quaternion exponential product ej ω2 x2 eiω1 x1 is called the inverse (rightsided) quaternion Fourier kernel. Detailed information about the QFT and its properties can be found in [56, 58, 124, 184, 255].
23
For the QFT on L2 , see Footnote 2 on page 144.
202 Quaternion and Clifford Fourier Transforms 1
1 0.5
0.5 0 0 −0.5 −0.5 5
−1 5 0 x2
−5
−5
0
0
5
x2
x1
−5
0
−5
5
x1
Representation of complex Gabor filter for σ1 = σ2 = 1, u0 = v0 = 1 in the spatial domain with its real part (left) and imaginary part (right). Source: [256, Fig. 1].
Figure 4.7
4.3.1.2 Quaternionic windowed Fourier transform
This section generalizes the classical WFT to quaternion algebra. Using the definition of the (rightsided) QFT described before, we extend the WFT to the QWFT. We shall later see how some properties of the WFT are extended in the new definition. For this purpose we briefly review the twodimensional WFT. Twodimensional WFT The FT is a powerful tool for the analysis of stationary signals but it is not well suited for the analysis of nonstationary signals because it is a global transformation with poor spatial localization [337]. However, in practice, most natural signals are nonstationary. In order to characterize a nonstationary signal properly, the WFT is commonly used. 4.3.1.2.1
Definition 4.3.3 (WFT). The WFT of a twodimensional real signal f ∈ L2 (R2 ) with respect to the window function g ∈ L2 (R2 ) \ {0} is given by Gg f (ω, b) =
1 (2π)2
Z R2
f (x) gω ,b (x) d2 x,
(4.3.6)
where the window daughter function gω ,b is called the windowed Fourier kernel defined by √ (4.3.7) gω ,b (x) = g(x − b)e −1 ω ·x . Equation (4.3.6) shows that the image of a WFT is a complex 4D coefficient function. Most applications make use of the Gaussian window function g which is nonnegative and well localized around the origin in both spatial and frequency domains. The Gaussian window function can be expressed as g(x, σ1 , σ2 ) = e−[(x1 /σ1 )
2 +(x
2 /σ2 )
2
]/2 ,
(4.3.8)
where σ1 and σ2 are the standard deviations of the Gaussian function and determine the width of the window. We call (4.3.7), for fixed ω = ω 0 = u0 e1 + v0 e2 , and
Quaternion Fourier transforms 203
b1 = b2 = 0, a complex Gabor filter as shown in Figure 4.7 if g is the Gaussian function (4.3.8), i.e. √
gc,ω 0 (x, σ1 , σ2 ) = e
−1 (u0 x1 +v0 x2 )
g(x, σ1 , σ2 ).
(4.3.9)
In general, when the Gaussian function (4.3.8) is chosen as the window function, the WFT in (4.3.6) is called Gabor transform. We observe that the WFT localizes the signal f in the neighborhood of x = b. For this reason, the WFT is often called short time Fourier transform. B¨ ulow [56] extended the complex Gabor filter √ −1(u0 x1 +v0 x2 ) (4.3.9) to quaternion algebra by replacing the complex kernel e with the inverse (twosided) quaternion Fourier kernel eiu0 x1 ej v0 x2 . His extension then takes the form 4.3.1.2.2
Definition of the QWFT
gq (x, σ1 , σ2 ) = eiu0 x1 ej v0 x2 e−[(x1 /σ1 )
2 +(x
2 /σ2 )
2
]/2 ,
(4.3.10)
which he called quaternionic Gabor filter 24 as shown in Figure 4.8 and applied it to get the local quaternionic phase of a twodimensional real signal. BayroCorrochano et al. [25] also used quaternionic Gabor filters for the preprocessing of twodimensional speech representations. The extension of the WFT to quaternion algebra using the (twosided) QFT is rather complicated, due to the noncommutativity of quaternion functions. Alternatively, we use the (rightsided) QFT to define the QWFT. We therefore introduce the following general QWFT of a twodimensional quaternion signal f ∈ L2 (R2 ; H) in Def. 4.3.5. Definition 4.3.4. A quaternion window function is a function φ ∈ L2 (R2 ; H) \{0} such that x1/2 φ(x) ∈ L2 (R2 ; H) too. We call φω ,b (x) =
1 j ω2 x2 iω1 x1 e e φ(x − b), (2π)2
(4.3.11)
a quaternionic window daughter function. If we fix ω = ω 0 , and b1 = b2 = 0, and take the Gaussian function as the window function of (4.3.11), then we get the quaternionic Gabor filter shown in Figure 4.9, gq (x, σ1 , σ2 ) = 24
2 2 1 ej v0 x2 eiu0 x1 e−[(x1 /σ1 ) +(x2 /σ2 ) ]/2 . 2 (2π)
(4.3.12)
If we would have interchanged the order of the two exponentials in Definition 4.3.1, which we are always free to do, then (4.3.10) and (4.3.12) would agree fully, except for the factor (2π)−2 . Figures 4.8 and 4.9 illustrate the two different kinds of quaternionic Gabor filters that arise. The differences can be made obvious by decomposition of the two exponential products eiu0 x1 ej v0 x2 and ej v0 x2 eiu0 x1 . The imaginary ipart of Figure 4.8 is the imaginary jpart of Figure 4.9 and vice versa. Note also that the imaginary kparts of Figures 4.8 and 4.9 are essentially the same, because they only have different signs.
204 Quaternion and Clifford Fourier Transforms 1
1 0.5
0.5 0 0 −0.5 −0.5 5
−1 5 0 x
−5 2
0
5
0
−5
x
x1
1
0.4
0.5
0.2
0
0
−0.5
−0.2
−1 5
−0.4 5 0 x2
−5
5
0
−5
−5
x1
−5
−5
x1
0 x2
x1
0
−5
2
0
5
5
B¨ ulow’s quaternionic Gabor filter (4.3.10) (σ1 = σ2 = 1, u0 = v0 = 1) in the spatial domain with real part (top left) and imaginary ipart (top right), jpart (bottom left), and kpart (bottom right). Source: [256, Fig. 2].
Figure 4.8
Definition 4.3.5 (QWFT). Denote the QWFT on L2 (R2 ; H) by Gφ . Then the QWFT of f ∈ L2 (R2 ; H) is defined by f (x) −→ Gφ f (ω, b) = hf, φω ,b iL2 (R2 ;H) Z
= R2
f (x) φω ,b (x) d2 x
1 f (x) ej ω2 x2 eiω1 x1 φ(x − b)d2 x (2π)2 R2 Z 1 f (x) φ(x − b)e−iω1 x1 e−j ω2 x2 d2 x. (2π)2 R2 Z
= =
(4.3.13)
Please note that the order of the exponentials in (4.3.13) is fixed because of the noncommutativity of the product of quaternions. Changing the order yields another quaternion valued function which differs by the signs of the terms. Equation (4.3.13) clearly shows that the QWFT can be regarded as the (rightsided) QFT (compare (5.5.35)) of the product of a quaternionvalued signal f and a shifted and quaternion conjugate version of the quaternion window function or as an inner product (4.3.1) of f and the quaternionic window daughter function. In contrast to the QFT basis e−iω1 x1 e−j ω1 x2 which has an infinite spatial extension, the QWFT basis φ(x − b) e−iω1 x1 e−j ω1 x2 has a limited spatial extension due to the local quaternion window function φ(x − b). The energy density is defined as the modulus square of the QWFT (4.3.13) given
Quaternion Fourier transforms 205 1
1 0.5
0.5 0 0 −0.5 −0.5 5
−1 5 0 x
−5 2
0
−5
0
5 x
x1
1
0.4
0.5
0.2
0
0
−0.5
−0.2
−1 5
−0.4 5 0 −5
x2
0
−5
x1
5
5
0
−5
−5
x1
−5
−5
x1
2
0 x2
5
0
The real part (top left) and imaginary ipart (top right), jpart (bottom left), and kpart (bottom right) of a quaternionic Gabor filter (σ1 = σ2 = 1, u0 = v0 = 1) in the spatial domain. Source: [256, Fig. 3].
Figure 4.9
by Z 2 1 − i ω1 x1 − j ω2 x2 2 Gφ f (ω, b) = f (x) φ(x − b)e e d x . (2π)4 R2
2
(4.3.14)
Equation (4.3.14) is often called a spectrogram which measures the energy of a quaternionvalued function f in the positionfrequency neighborhood of (b, ω). A good choice for the window function φ is the Gaussian quaternion function because, according to Heisenberg’s uncertainty principle, the Gaussian quaternion signal can simultaneously minimize the spread in both spatial and quaternionic frequency domains, and it is smooth in both domains. The uncertainty principle can be written in the following form [255] 1 4gx1 4gx2 4gω1 4gω2 ≥ , (4.3.15) 4 where 4gxk , k = 1, 2, are the effective spatial widths of the quaternion function g and 4gωk , k = 1, 2, are its effective bandwidths. For illustrative purposes, we shall discuss examples of the QWFT. We begin with a straightforward example. 4.3.1.2.3
Examples of the QWFT
Example 4.3.6. Consider the twodimensional first order Bspline window function (see [313]) defined by (
φ(x) =
1, 0,
if −1 ≤ x1 ≤ 1 and −1 ≤ x2 ≤ 1, otherwise.
(4.3.16)
206 Quaternion and Clifford Fourier Transforms
Obtain the QWFT of the function defined as follows: (
f (x) =
ex1 +x2 , 0,
if −∞ < x1 < 0 and −∞ < x2 < 0, otherwise.
(4.3.17)
By applying the definition of the QWFT we have m2 m1 1 ex1 +x2 e−iω1 x1 e−j ω2 x2 dx1 dx2 , Gφ f (ω, b) = (2π)2 −1+b1 −1+b2 m1 = min(0, 1 + b1 ), m2 = min(0, 1 + b2 ).
Z
Z
(4.3.18)
Simplifying (4.3.18) yields m1 m2 1 ex1 (1−iω1 ) ex2 (1−j ω2 ) d2 x (4.3.19) (2π)2 −1+b1 −1+b2 Z m1 Z m2 1 ex1 (1−iω1 ) dx1 ex2 (1−j ω2 ) dx2 (2π)2 −1+b1 −1+b2 m 1 ex2 (1−j ω2 ) m2 1 x1 (1−iω1 ) e (1 − iω ) 1 (2π)2 −1+b1 (1 − jω2 ) −1+b2 (em1 (1−iω1 ) − e(−1+b1 )(1−iω1 ) )(em2 (1−j ω2 ) − e(−1+b2 )(1−j ω2 ) ) . (2π)2 (1 − iω1 − jω2 + kω1 ω2 )
Z
Gφ f (ω, b) = = = =
Z
Using the properties of quaternions we obtain Gφ f (ω, b) =
(em1 (1−iω1 ) − e(−1+b1 )(1−iω1 ) )(em2 (1−j ω2 ) − e(−1+b2 )(1−j ω2 ) ) × (2π)2 (1 + iω1 + jω2 − kω1 ω2 ) × . (4.3.20) (1 + ω12 + ω22 + ω12 ω22 )
Example 4.3.7. Given the window function of the twodimensional Haar function defined by 1, for 0 ≤ x1 < 1/2 and 0 ≤ x2 < 1/2, (4.3.21) φ(x) = −1, for 1/2 ≤ x1 < 1 and 1/2 ≤ x2 < 1, 0, otherwise, 2
2
find the QWFT of the Gaussian function f (x) = e−(x1 +x2 ) . From Definition 4.3.5 we obtain 1 f (x)φ(x − b)e−iω1 x1 e−j ω2 x2 d2 x (2π)2 R2 Z 1/2+b1 Z 1/2+b2 1 2 −x21 −iω1 x1 e e dx e−x2 e−j ω2 x2 dx2 = 1 2 (2π) b1 b2 Z 1+b1 Z 1+b2 1 2 2 − e−x1 e−iω1 x1 dx1 e−x2 e−j ω2 x2 dx2 . 2 (2π) 1/2+b1 1/2+b2 (4.3.22) Z
Gφ f (ω, b) =
Quaternion Fourier transforms 207
By completing squares, we have Gφ f (ω, b) =
1 (2π)2
1/2+b1
Z
e−(x1 +iω1 /2)
2 −ω 2 /4 1
dx1
b1 1/2+b2
Z
e−(x2 +j ω2 /2)
2 −ω 2 /4 2
(4.3.23)
dx2
b2
1 − (2π)2
Z
1+b1
−(x1 +iω1 /2)2 −ω12 /4
e
1+b2
Z
e−(x2 +j ω2 /2)
dx1
Making the substitutions y1 = x1 + i ω21 and y2 = x2 + j we immediately obtain 2
2
e−(ω1 +ω2 )/4 Gφ f (ω, b) = (2π)2
−(ω12 +ω22 )/4
=
e
e
Z
1/2+b1 +iω1 /2
Z
−y12
1+b1 +iω1 /2
−y12
(−e
ω2 2
dy1
b2 +j ω2 /2
Z
e−y1 dy1
2
e−y2 dy2
1+b2 +j ω2 /2
1/2+b2 +j ω2 /2
1/2+b1 +iω1 /2
Z
) dy1 +
2
e−y2 dy2
! −y12
e
dy1
0
0 b2 +j ω2 /2
Z
×
−y22
(−e
Z
Z
−
1/2+b2 +j ω2 /2
) dy2 +
0 1/2+b1 +iω1 /2
−y12
(−e Z
! −y22
e
dy2
0
Z
1+b1 +iω1 /2
! −y12
) dy1 +
e
dy1
0
0
×
in the above expression
1/2+b2 +j ω2 /2
Z
2
1/2+b1 +iω1 /2
b1 +iω1 /2
" Z
e
b1 +iω1 /2
−(ω12 +ω22 )/4
(2π)2
(2π)2
dx2 .
1/2+b2
1/2+b1
−
2 −ω 2 /4 2
1/2+b2 +j ω2 /2
−y22
(−e
1+b2 +j ω2 /2
Z
) dy2 +
0
!# −y22
e
dy2
.
0
(4.3.24) Equation (4.3.24) can be written in the form 2
2
e−(ω1 +ω2 )/4 √ Gφ f (ω, b) = (2 π)3
i 1 i −erf b1 + ω1 + erf + b1 + ω1 2 2 2 j j 1 × −erf b2 + ω2 + erf + b2 + ω2 2 2 2 1 i i − −erf + b1 + ω1 + erf 1 + b1 + ω1 2 2 2 1 j j × −erf + b2 + ω2 + erf 1 + b2 + ω2 , 2 2 2
where erf(x) =
√2 π
Rx 0
(4.3.25)
2
e−t dt.
4.3.1.2.4 Properties of the QWFT In this subsection, we describe the properties of the QWFT. We must exercise care in extending the properties of the WFT to the QWFT because of the general noncommutativity of quaternion multiplication. We will find most of the properties of the WFT are still valid for the QWFT, however with some modifications.
208 Quaternion and Clifford Fourier Transforms
Theorem 4.3.8 (Left linearity). Let φ ∈ L2 (R2 ; H) be a quaternion window function. The QWFT of f, g ∈ L2 (R2 ; H) is a left linear operator, which means [Gφ (λf + µg)](ω, b) = λGφ f (ω, b) + µGφ g(ω, b),
(4.3.26)
for arbitrary quaternion constants λ, µ ∈ H. Remark 4.3.9. Restricting the constants in Theorem 4.3.8 to λ, µ ∈ R we get both left and right linearity of the QWFT. Theorem 4.3.10 (Parity). Let φ ∈ L2 (R2 ; H) be a quaternion window function. Then we have GP φ {P f }(ω, b) = Gφ f (−ω, −b), (4.3.27) where P φ(x) = φ(−x), ∀φ ∈ L2 (R2 ; H). Theorem 4.3.11 (Specific shift). Let φ be a quaternion window function. Assume that f = f0 + if1 and φ = φ0 + iφ1 . (4.3.28) Then we obtain Gφ Tx0 f (ω, b) = e−iω1 x0 (Gφ f (ω, b − x0 )) e−j ω2 y0 ,
(4.3.29)
where Tx0 denotes the translation operator by x0 = x0 e1 + y0 e2 , i.e. Tx0 f = f (x − x0 ). Equation (4.3.29) describes that if the original function f (x) is shifted by x0 , its window function will be shifted by x0 , the frequency will remain unchanged, and the phase will be changed by the left and right phase factors e−iω1 x0 and e−j ω2 y0 . Remark 4.3.12. Like for the (rightsided) QFT, the usual form of the modulation property of the QWFT does not hold [184, 255]. It is obstructed by the noncommutativity of the quaternion exponential product factors e−iω1 x1 e−j ω2 x2 = 6 e−j ω2 x2 e−iω1 x1 .
(4.3.30)
The following theorem tells us that the QWFT is invertible, that is, the original quaternion signal f can be recovered simply by taking the inverse QWFT. Theorem 4.3.13 (Reconstruction formula). Let φ be a quaternion window function. Then every 2D quaternion signal f ∈ L2 (R2 ; H) can be fully reconstructed by (2π)2 f (x) = kφk2L2 (R2 ;H)
Z
Z
R2
R2
Gφ f (ω, b)φω ,b (x) d2 b d2 ω.
(4.3.31)
Set Cφ = kφk2L2 (R2 ;H) and assume that 0 < Cφ < ∞. Then, the reconstruction formula (4.3.31) can also be written as (2π)2 Gφ f (ω, b) φω ,b d2 b d2 ω C φ R2 R2 Z Z (2π)2 hf, φω ,b iL2 (R2 ;H) φω ,b d2 b d2 ω. C φ R2 R2 Z
f (x) = =
Z
More properties of the QWFT are given in the following theorems.
(4.3.32)
Quaternion Fourier transforms 209
Theorem 4.3.14 (Orthogonality relation). Let φ be a quaternion window function and f, g ∈ L2 (R2 ; H) arbitrary. Then we have Z
Z
R2
R2
hf, φω ,b iL2 (R2 ;H) hg, φω ,b i =
L2 (R2 ;H)
d2 ω d2 b
Cφ hf, giL2 (R2 ;H) . (2π)2
(4.3.33)
As an easy consequence of the previous theorem, we immediately obtain the following corollary. Corollary 4.3.15. If f, φ ∈ L2 (R2 ; H) are two quaternionvalued signals, then Z
Z
R2
R2
Gφ f (ω, b)2 d2 b d2 ω =
1 kf k2L2 (R2 ;H) kφk2L2 (R2 ;H) . (2π)2
(4.3.34)
In particular, if the quaternion window function is normalized to kφkL2 (R2 ;H) = 1, then (4.3.34) becomes Z
Z
R2
R2
Gφ f (ω, b)2 d2 b d2 ω =
1 kf k2L2 (R2 ;H) . (2π)2
(4.3.35)
Equation (4.3.35) shows that the QWFT is an isometry from L2 (R2 ; H) into 1 L (R2 ; H). In other words, up to a factor of (2π) 2 the total energy of a quaternionvalued signal computed in the spatial domain is equal to the total energy computed in the quaternionic windowed Fourier domain, compare (4.3.14) for the corresponding energy density. 2
Theorem 4.3.16 (Reproducing kernel). Let be φ ∈ L2 (R2 ; H) be a quaternion window function. If Kφ (ω, b; ω 0 , b0 ) =
(2π)2 hφω ,b , φω 0 ,b0 iL2 (R2 ;H) , Cφ
(4.3.36)
then Kφ (ω, b; ω 0 , b0 ) is a reproducing kernel, i.e. Gφ f (ω 0 , b0 ) =
Z R2
Z R2
Gφ f (ω, b)Kφ (ω, b; ω 0 , b0 ) d2 ω d2 b
(4.3.37)
The above properties of the QWFT are summarized in Table 4.3. 4.3.1.3 Heisenberg’s uncertainty principle for the QWFT
The classical uncertainty principle of harmonic analysis states that a nontrivial function and its Fourier transform can not both be simultaneously sharply localized [76, 319]. In quantum mechanics an uncertainty principle asserts one can not at the same time be certain of the position and of the velocity of an electron (or any particle). That is, increasing the knowledge of the position decreases the knowledge of the velocity or momentum of an electron. This section extends the uncertainty
210 Quaternion and Clifford Fourier Transforms
Properties of the QWFT of f, g ∈ L2 (R2 ; H), where λ, µ ∈ H are constants and x0 = x0 e1 + y0 e2 ∈ R2 . Source: [256, Tab. 1].
Table 4.3
Property Left linearity Parity
Quat. Function λf (x) +µg(x) GP φ {P f }(ω, b)
Specific shift
f (x − x0 )
QWFT λGφ f (ω, b)+ µGφ g(ω, b) Gφ f (−ω, −b) e−iω1 x0 G f (ω, b − x ) e−j ω2 y0 , φ
0
if f = f0 + if1 and φ = φ0 + iφ1 Formula Orthogonality
kφk2 2
L (R2 ;H)
(2π)2
hf, giL2 (R2 ;H) =
Reconstruction f (x) =
R
R
R2 R2 hf, φω ,b iL2 (R2 ;H) hg, φω ,b i 2 2 d2 ω d2 b L (R ;H) R R (2π)2 R2 R2 Gφ f (ω, b) kφk2 2 2 L (R ;H)
φω ,b (x) d2 b d2 ω Isometry
1 2 (2π)2 kf kL2 (R2 ;H)
Reproducing Kernel
Gφ f (ω 0 , b0 ) =
=
Gφ f (ω, b)2 d2 b d2 ω, if kφkL2 (R2 ;H) = 1 R
R
R
R2 R2
R
Gφ f (ω, b) Kφ (ω, b; ω 0 , b0 ) d2 ω d2 b, Kφ (ω, b; ω 0 , b0 ) = R2 R2
(2π)2 φ2 2 2
L (R ;H)
hφω ,b , φω 0 ,b0 iL2 (R2 ;H)
principle which is valid for the (rightsided) QFT [255] to the setting of the QWFT. A directional QFT uncertainty principle has been studied in [187]. In [255] a componentwise uncertainty principle for the QFT establishes a lower bound on the product of the effective widths of quaternionvalued signals in the spatial and frequency domains. This uncertainty can be written in the following form. Theorem 4.3.17 (QFT uncertainty principle). Let f ∈ L2 (R2 ; H) be a quaternionvalued function. If Fq {f }(ω) ∈ L2 (R2 ; H) too, then we have the inequality (no summation over k, k = 1, 2) Z R2
x2k f (x)2 d2 x
Z R2
ωk2 Fq {f }(ω)2 d2 ω ≥
(2π)2 4
Z
f (x)2 d2 x
2
.
(4.3.38)
R2
Equality holds if and only if f is the Gaussian quaternion function, i.e. 2
2
f (x) = C0 e−(a1 x1 +a2 x2 ) ,
(4.3.39)
where C0 is a quaternion constant and a1 , a2 are positive real constants. Applying the Parseval theorem for the QFT [184] to the righthand side of (4.3.38) we get the following corollary.
Quaternion Fourier transforms 211
Corollary 4.3.18. Under the above assumptions, we have Z R2
x2k Fq−1 [Fq {f }](x)2
Z
2
d x R2
ωk2 Fq {f }(ω)2 d2 ω
≥
1 4π
Z R2
Fq {f }(ω)2 d2 ω
2
.
(4.3.40)
Let us now establish a generalization of the Heisenberg type uncertainty principle for the QWFT. From a mathematical point of view this principle describes how the spatial extension of a twodimensional quaternion function relates to the bandwidth of its QWFT. Theorem 4.3.19 (QWFT uncertainty principle). Let φ ∈ L2 (R2 ; H) be a quaternion window function and let Gφ f ∈ L2 (R2 ; H) be the QWFT of f such that ωk Gφ f ∈ L2 (R2 ; H), k = 1, 2. Then for every f ∈ L2 (R2 ; H) we have the following inequality: Z
Z
R2
R2
ωk2 Gφ f (ω, b)2
2
2
1/2 Z
d ωd b R2
x2k f (x)2
2
1/2
d x
1 (4.3.41) kf k2L2 (R2 ;H) kφkL2 (R2 ;H) . 4π In order to prove this theorem, we need to introduce the following lemma. ≥
Lemma 4.3.20. Under the assumptions of Theorem 4.3.19, we have for k = 1, 2 that kφk2L2 (R2 ;H) Z (2π)4
R2
x2k f (x)2 d2 x =
Z R2
Z R2
x2k Fq−1 {Gφ f (ω, b)}(x)2 d2 xd2 b. (4.3.42)
Remark 4.3.21. According to the properties of the QFT and its uncertainty principle, Theorem 4.3.19 does not hold for summation over k. If we introduce summation, 1 1 we would have to replace the factor 4π on the right hand side of (4.3.41) by 2π . 4.3.1.4 Application of the QWFT
The WFT plays a fundamental role in the analysis of signals and linear timevarying (TV) systems [140,147,313]. The effectiveness of the WFT is a result of its providing a unique representation for the signals in terms of the windowed Fourier kernel. It is natural to ask whether the QWFT can also be applied to such problems. This section briefly discusses the application of the QWFT to study twodimensional linear TV systems (see Figure 4.10). We may regard the QWFT as a linear TV bandpass filter element of a filterbank spectrum analyzer and, therefore, the TV spectrum obtained by the QWFT can also be interpreted as the output of such a linear TV bandpass filter element. For this purpose let us introduce the following definition. Definition 4.3.22. Consider a twodimensional linear TV system with h(·, ·, ·) denoting the quaternion impulse response of the filter. The output r(·, ·) of the linear TV system is defined by Z
r(ω, b) = R2
2
f (x)h(ω, b, b − x) d x =
Z
f (b − x)h(ω, b, x) d2 x,
R2
where f (·) is a twodimensional quaternion valued input signal.
(4.3.43)
212 Quaternion and Clifford Fourier Transforms f
h
r
Figure 4.10 Block diagram of a twodimensional linear timevarying system. Source: [256, Fig. 4].
We then obtain the transfer function R(·, ·) of the quaternion impulse response h(·, ·, ·) of the TV filter as Z
R(ω, b) = R2
h(ω, b, α) e−iω1 α1 e−j ω2 α2 d2 α, α = α1 e1 + α2 e2 ∈ R2 .
(4.3.44)
The following simple theorem (compare to Ghosh and Sreenivas [140]) relates the QWFT to the output of a linear TV bandpass filter. Theorem 4.3.23. Consider a linear TV bandpass filter. Let the TV quaternion impulse response h1 (·, ·, ·) of the filter be defined by h1 (ω, b, α) =
1 φ(−α) e−iω1 (b1 −α1 ) e−j ω2 (b2 −α2 ) , (2π)2
(4.3.45)
where φ(·) is the quaternion window function. The output r1 (·, ·) of the TV system is equal to the QWFT of the quaternion input signal f (x). This shows that the choice of the quaternion impulse response of the filter will determine a characteristic output of the linear TV systems. For example, if we translate the TV quaternion impulse response h1 (·, ·, ·) by b0 = b01 e1 + b02 e2 , i.e. h1 (ω, b, α) → h1 (ω, b, α − b0 ) =
1 φ(−(α − b0 )) e−iω1 (b1 −(α1 −b01 )) (2π)2 × e−j ω2 (b2 −(α2 −b02 )) , (4.3.46)
then the output is according to Theorem 4.3.11 r1,b0 (ω, b) = e−iω1 b01 Gφ f (ω, b − b0 ) e−j ω2 b02 .
(4.3.47)
In this case, we assumed that the input f i = if and the window function φi = iφ. Theorem 4.3.24. Consider a linear TV bandpass filter with the TV quaternion impulse response h2 (·, ·, ·) defined by h2 (ω, b, α) = e−iω1 (b1 −α1 ) e−j ω2 (b2 −α2 ) ,
(4.3.48)
If the input to this system is the quaternion signal f (x), its output r2 (ω) = r2 (ω, ·) is, independent of the bargument, equal to the QFT of f : r2 (ω) = Fq {f }(ω).
(4.3.49)
Quaternion Fourier transforms 213
0.03
0.02
0.02 0 0.01 0 10
−0.02 10
ω2
10 0
0 −10
−10
10
ω2 0
ω1
−10
0.02
0.01
0
0
−0.02 10
0
ω1
0
ω1
−10
−0.01 10
ω2 0
10 0 −10
−10
ω1
ω2 0
10 −10
−10
The real part (top left) and imaginary ipart (top right), jpart (bottom left), and kpart (bottom right) of the output r2 (·) in Example 4.3.25, i.e. the QFT (4.3.51) of (4.3.50). Source: [256, Fig. 5].
Figure 4.11
Example 4.3.25. Given the TV quaternion impulse response defined by (4.3.48). Find the output r2 (·) (see Figure 4.11) of the following input (
f (x) =
e−(x1 +x2 ) , 0,
if x1 ≥ 0 and x2 ≥ 0, otherwise.
(4.3.50)
From Theorem 4.3.24, we obtain the QFT of f ∞ ∞ 1 e−x1 (1+iω1 ) e−x2 (1+j ω2 ) d2 x 2 (2π) 0 0 ∞ ∞ 1 −1 −1 −iω1 x1 −x1 −j ω2 x2 −x2 e e e e (2π)2 1 + iω1 0 (1 + jω2 ) 0 1 1 (2π)2 1 + iω1 + jω2 + kω1 ω2 1 1 − iω1 − jω2 − kω1 ω2 . (4.3.51) (2π)2 1 + ω12 + ω22 + ω12 ω22
Z
r2 (ω) = = = =
Z
Example 4.3.26. Consider the TV quaternion impulse response defined by (4.3.45) with respect to the first order twodimensional Bspline window function (4.3.16) in Example 4.3.6. Find the output r1 (·, ·) (see Figure 4.12) of the input (4.3.50) defined in Example 4.3.25.
214 Quaternion and Clifford Fourier Transforms
0.02
0.01
0.01 0 0 −0.01 10
−0.01 10
ω2
10 0
0 −10
−10
ω20
ω1
0.01
0.01
0
0
−0.01 10
10 −10
0
ω1
0
ω1
−10
−0.01 10 10
ω20
0 −10
−10
ω2 0
ω1
10 −10
−10
The real part (top left) and imaginary ipart (top right), jpart (bottom left), and kpart (bottom right) of the output r1 (ω, b = 0) in Example 4.3.26, i.e. the QWFT (4.3.52) of (4.3.45) with b1 = b2 = 0. Source: [256, Fig. 6].
Figure 4.12
With m1 = max(0, −1 + b1 ), m2 = max(0, −1 + b2 ), Theorem 4.3.23 gives r1 (ω, b) = Gφ f (ω, b) =
1 (2π)2
Z
1+b1
m1
e−x1 e−ix1 ω1 dx1
(4.3.52) Z
1+b2
e−x2 e−j x2 ω2 dx2
m2
1+b1 1+b2 −1 −1 −x1 (1+iω1 ) −x2 (1+j ω1 ) = e e (2π)2 (1 + iω1 ) (1 + jω2 ) m1 m2 −m1 (1+iω1 ) −(1+b1 )(1+iω1 ) −m2 (1+j ω2 ) −(1+b2 )(1+j ω2 ) (e −e )(e −e ) . = 2 (2π) (1 + iω1 + jω2 + kω1 ω2 )
For the sake of simplicity, we take the parameters b1 = b2 = 0 ⇒ m1 = m2 = 0, to obtain (1 − e−(1+iω1 ) )(1 − e−(1+j ω2 ) ) (2π)2 (1 + iω1 + jω2 + kω1 ω2 ) 1 − e−1 cos ω1 − e−1 cos ω2 + e−2 cos ω1 cos ω2 = (2π)2 (1 + iω1 + jω2 + kω1 ω2 ) i(e−1 sin ω1 − e−2 sin ω1 cos ω2 ) + j(e−1 sin ω2 − e−2 cos ω1 sin ω2 ) + (2π)2 (1 + iω1 + jω2 + kω1 ω2 ) −2 k e sin ω1 sin ω2 + . (4.3.53) 2 (2π) (1 + iω1 + jω2 + kω1 ω2 )
r1 (ω, b = 0) =
We may regard the QFT (4.3.51) as the QWFT with an infinite window function. Since the integration domain of the QWFT (4.3.52) is smaller than that of the QFT (4.3.51), the QWFT output (4.3.52) is more localized in the base space than the QFT
Quaternion Fourier transforms 215
output of (4.3.51). In addition, according to the PaleyWiener theorem the QWFT output of (4.3.52) is very smooth. This means that it provides accurate information on the output r(·, ·) due to the local window function φ. 4.3.1.5 Summary on windowed QFT
Using the basic concepts of quaternion algebra and the (rightsided) QFT we introduced the windowed QFT (QWFT). Important properties of the QWFT, such as left linearity, parity, reconstruction formula, reproducing kernel, isometry and orthogonality relation, were demonstrated. Because of the noncommutativity of multiplication in the quaternion algebra H, not all properties of the classical WFT can be established for the QWFT, such as general shift and modulation properties. This generalization also enables us to construct quaternionic Gabor filters (compare to B¨ ulow [56, 58]), which can extend the applications of the 2D complex Gabor filters to texture segmentation and disparity estimation [56]. We have established a new uncertainty principle for the QWFT. This principle is founded on the QWFT properties and the uncertainty principle for the (rightsided) QFT. We also applied the QWFT to a linear timevarying (TV) system. We showed that the output of a linear TV system can result in a QFT or a QWFT of the quaternion input signal, depending on the choice of the quaternion impulse response of the filter. 4.3.2 Quaternionic Fourier Mellin transform 4.3.2.1 Background
In this part we generalize the classical Fourier Mellin transform [109], which transforms functions f representing, e.g., a gray level image defined over a compact set of R2 . The resulting quaternionic Fourier Mellin transform (QFMT) applies to functions f : R2 → H, for which f  is summable over R∗+ × S1 under the measure dθ dr r . R∗+ is the multiplicative group of positive and nonzero real numbers. We investigate the properties of the QFMT similar to the investigation of the quaternionic Fourier Transform (QFT) in [184, 187]. This section is based on [198]. 4.3.2.2
The quaternionic Fourier Mellin transformations (QFMT)
Robert Hjalmar Mellin (1854–1933) Robert Hjalmar Mellin (1854–1933) [260], Figure 4.13, was a Finnish mathematician, a student of G. MittagLeffler and K. Weierstrass. He became the director of the Polytechnic Institute in Helsinki, and in 1908 first professor of mathematics at Technical University of Finland. He was a fervent fennoman with fiery temperament, and cofounder of the Finnish Academy of Sciences. He became known for the Mellin transform with major applications to the evaluation of integrals, see [284], which lists 1624 references. During his last 10 years, he tried to refute Einstein’s theory of relativity as logically untenable.
4.3.2.2.1
216 Quaternion and Clifford Fourier Transforms
Definition 4.3.27 (Classical Fourier Mellin transform (FMT)). ∀(v, k) ∈ R × Z,
1 2π
M{h}(v, k) =
∞
Z
Z
2π
h(r, θ)r−iv e−ikθ dθ
0
0
dr , r
(4.3.54)
where h : R2 → R denotes a function representing, e.g., a gray level image defined over a compact set of R2 . Well known applications are to shape recognition (independent of rotation and scale), image registration and similarity. We now define the generalization of the FMT to quaternionic signals. Definition 4.3.28 (Quaternionic Fourier Mellin transform (QFMT)). Let f, g ∈ H : f 2 = g 2 = −1 be any pair of pure unit quaternions. The quaternionic Fourier Mellin transform (QFMT) is given by ∀(v, k) ∈ R × Z, ˆ k) = M{h}(v, k) = 1 h(v, 2π
∞
Z
2π
Z
0
r−f v h(r, θ)e−gkθ dθ
0
dr , r
(4.3.55)
where h : R2 → H denotes a function from R2 into the algebra of quaternions H, such ∗ that h is summable over R∗+ × S1 under the measure dθ dr r . R+ is the multiplicative group of positive and nonzero real numbers. For f = i, g = j we have the special case ∀(v, k) ∈ Z × R, ˆ k) = M{h}(v, k) = 1 h(v, 2π
Z
∞
Z
2π
r−iv h(r, θ)e−jkθ dθ
0
0
dr . r
(4.3.56)
Note, that the ± split and the QFMT commute: M{h± } = M{h}± .
(4.3.57)
Theorem 4.3.29 (Inverse QFMT). The QFMT can be inverted by h(r, θ) = M−1 {h}(r, θ) =
Figure 4.13
1 2π
Z
∞
X
ˆ k) egkθ dv. rf v h(v,
−∞ k∈Z
Robert Hjalmar Mellin (1854–1933). Image: Wikipedia.
(4.3.58)
Quaternion Fourier transforms 217
The proof uses 1 X gk(θ−θ0 ) e = δ(θ − θ0 ), 2π k∈Z 1 2π
2π
Z
rf v = ef v ln r ,
ef v(ln(r)−s) dv = δ(ln(r) − s).
(4.3.59)
0
We now investigate the basic properties of the QFMT. First, left linearity: For α, β ∈ {q  q = qr + qf f, qr , qf ∈ R}, ˆ 1 (v, k) + β h ˆ 2 (v, k). m(r, θ) = αh1 (r, θ) + βh2 (r, θ) =⇒ m(v, ˆ k) = αh
(4.3.60)
Second, right linearity: For α0 , β 0 ∈ {q  q = qr + qg g, qr , qg ∈ R}, ˆ 2 (v, k)β 0 . ˆ 1 (v, k)α0 + h ˆ k) = h m(r, θ) = h1 (r, θ)α0 + h2 (r, θ)β 0 =⇒ m(v,
(4.3.61)
The linearity of the QFMT leads to M{h}(v, k) = M{h− + h+ }(v, k) = M{h− }(v, k) + M{h+ }(v, k),
(4.3.62)
which gives rise to the following theorem. Theorem 4.3.30 (Quasicomplex FMT like forms for QFMT of h± ). The QFMT of h± parts of h ∈ L2 (R2 , H) have simple quasicomplex forms25 1 ∞ 2π dr h± r±gv e−gkθ dθ 2π 0 0 r Z ∞ Z 2π dr 1 r−f v e±f kθ h± dθ = . 2π 0 0 r
M{h± } =
Z
Z
(4.3.63)
Theorem 4.3.30 allows to use discrete and fast software to compute the QFMT based on a pair of complex FMT transformations. For the two split parts of the QFMT, we have the following lemma. Lemma 4.3.31 (Modulus identities). Due to q2 = q− 2 + q+ 2 we get for f : R2 → H the following identities h(r, θ)2 = h− (r, θ)2 + h+ (r, θ)2 , M{h}(v, k)2 = M{h− }(v, k)2 + M{h+ }(v, k)2 .
(4.3.64)
Further properties are scaling and rotation: For m(r, θ) = h(ar, θ+φ), a > 0, 0 ≤ φ ≤ 2π, ˆ k)egkφ . b m(v, k) = af v h(v, (4.3.65) Moreover, we have the following magnitude identity: ˆ k), b m(v, k) = h(v, 25
For quaternionic Fourier transforms on L1 and L2 , see Footnote 2 on page 144.
(4.3.66)
218 Quaternion and Clifford Fourier Transforms
i.e. the magnitude of the QFMT of a scaled and rotated quaternion signal m(r, θ) = h(ar, θ + φ) is identical to the magnitude of the QFMT of h. Equation (4.3.66) forms the basis for applications to rotation and scale invariant shape recognition and image registration. This may now be extended to color images, since quaternions can encode colors RGB in their i, j, k components. The reflection at the unit circle (r → 1r ) leads to 1 m(r, θ) = h( , θ) r
=⇒
ˆ b m(v, k) = h(−v, k).
(4.3.67)
Reversing the sense of rotation (θ → −θ) yields m(r, θ) = h(r, −θ)
=⇒
ˆ −k). b m(v, k) = h(v,
(4.3.68)
Regarding radial and rotary modulation we assume m(r, θ) = rf v0 h(r, θ) egk0 θ ,
v0 ∈ R, k0 ∈ Z.
(4.3.69)
Then we get ˆ − v0 , k − k0 ). b m(v, k) = h(v 4.3.2.2.2
that
d d ln r
QFMT derivatives and power scaling
=
d r dr
(4.3.70)
We note for the logarithmic derivative
= r∂r , ˆ k), M{(r∂r )n h}(v, k) = (f v)n h(v,
n ∈ N.
(4.3.71)
Applying the angular derivative with respect to θ we obtain ˆ k)(gk)n , M{∂θn h}(v, k) = h(v,
n ∈ N.
(4.3.72)
Finally, power scaling with ln r and θ leads to ˆ k) g n , M{(ln r)m θn h}(v, k) = f m ∂vm ∂kn h(v, 4.3.2.2.3
QFMT Plancherel and Parseval theorems
m, n ∈ N.
(4.3.73)
For the QFMT we have the fol
lowing two theorems. Theorem 4.3.32 (QFMT Plancherel theorem). The scalar part of the inner product of two functions h, m : R2 → H is ˆ mi. b hh, mi = hh,
(4.3.74)
Theorem 4.3.33 (QFMT Parseval theorem). Let h : R2 → H. Then ˆ khk = khk,
ˆ 2 = kh ˆ + k2 + kh ˆ − k2 . khk2 = khk
(4.3.75)
Quaternion Fourier transforms 219
4.3.2.3 Symmetry and kernel structures of 2D FMT, FT, QFT, QFMT
The QFMT of real signals analyzes symmetry. The following notation will be used.26 The function hee is even with respect to (w.r.t.) r → 1r ⇐⇒ ln r → − ln r, i.e. w.r.t. the reflection at the unit circle, and even w.r.t. θ → −θ, i.e. w.r.t. reversing the sense of rotation (reflection at the θ = 0 line of polar coordinates in the (r, θ)plane). Similarly we denote by heo evenodd symmetry, by hoe oddeven symmetry, and by hoo oddodd symmetry. Let h be a real valued function R2 → R. The QFMT of h results in ˆ k) = h ˆ ee (v, k) + h ˆ eo (v, k) + h ˆ oe (v, k) + h ˆ oo (v, k) . h(v, 
{z
}
real part

{z
f part
}

{z
gpart
}

{z
f gpart
(4.3.76)
}
The QFMT of a real signal therefore automatically separates components with different combinations of symmetry w.r.t. reflection at the unit circle and reversal of the sense of rotation. The four components of the QFMT kernel differ by radial and angular phase shifts, see the left side of Figure 4.17. The symmetries of r → 1/r (reflection at yellow unit circle), and θ → −θ (reflection at green line) can be clearly seen on the right of Figure 4.17. Figure 4.18 shows real the component of the QFMT kernel for various values of v, k, demonstrating various angular and radial resolutions. Figure 4.19 shows the real component of the QFMT kernel for v = k = 4 at three different scales. Similar patterns appear at all scales. Figure 4.14 shows the kernels of complex 2D Fourier transform (FT) e−i(ux+vy) , i ∈ C, and the QFT e−iux e−jvy , i, j ∈ H, taken from [56], which treats applications to 2D gray scale images. Corresponding applications to color images can be found in [127]. The 2D FT is intrinsically 1D, the QFT is intrinsically 2D, which makes it superior in disparity estimation and 2D texture segmentation, etc. Figure 4.15 compares the kernels (real parts) of 2D complex FMT and the QFMT. Obviously the 2D QFMT can analyze genuine 2D textures better than the 2D complex FMT. Finally, Figure 4.16 compares the kernels of the QFT (left) and the QFMT (right). The scale invariant feature of the QFMT is obvious. Compared with the left side of Figure 4.15, the QFMT is obviously the linear superposition of two quasicomplex FMTs with opposite winding sense, as shown in Theorem 4.3.30. 4.3.2.4 Summary on quaternionic Fourier Mellin transformation
The algebra of quaternions allows to construct a variety of quaternionic FourierMellin transformations (QFMT), dependent on the choice of f, g ∈ H, f 2 = g 2 = −1. Further variations would be to place both kernel factors initially at the left or right of the signal h(r, θ). The whole QFMT concept can easily be generalized to Clifford algebras Cl(p, q), based on the general theory of square roots of −1 in Cl(p, q), see Section 2.7. The modulus of the transform is scale and rotation invariant. Preceded by 2D FT or by QFT, this allows translation, scale and rotation invariant object description. 26
In this section we assume g 6= ±f , but a similar study is possible for g = ±f .
220 Quaternion and Clifford Fourier Transforms
Figure 4.14
Left: 2D FT is intrinsically 1D. Right: QFT is intrinsically 2D. Source:
[56].
Left: Kernel of FMT. Right: Kernel of QFMT. k, v ∈ {0, 1, 2, 3, 4}. Source: [198, Fig. 5].
Figure 4.15
Quaternion Fourier transforms 221
Left: QFT kernel, right: QFMT kernel. Top row: q+ parts: 1 + f g and f − g components. Bottom row: q− parts: 1 − f g and f + g components. Source: [198, Fig. 6].
Figure 4.16
A diverse range of applications can therefore be imagined: Color object shape recognition, color image registration, application to evaluation of hypercomplex integrals, etc.
We will encounter an extension of the QFMT to Clifford algebras in Section 5.5.2. The future may bring extensions for the QFMT to windowed and wavelet transforms, discretization and numerical implementations. 4.3.3 The Quaternion domain Fourier transform 4.3.3.1 Background
So far quaternion Fourier transforms have been mainly defined over R2 as signal domain space, see e.g. Sections 4.1 and 4.2. But it seems natural to define a quaternion Fourier transform for quaternion valued signals over quaternion domains. This quaternion domain Fourier transform (QDFT) transforms quaternion valued signals (e.g. electromagnetic scalarvector potentials, color data, spacetime data, etc.) defined over a quaternion domain (spacetime or other 4D domains) from a quaternion “position” space to a quaternion “frequency” space. The QDFT uses the full potential provided by hypercomplex algebra in higher dimensions and may moreover be useful for solving quaternion partial differential equations or functional equations, and in crystallographic texture analysis. We define the QDFT and analyze its main properties, including quaternion dilation, modulation and shift properties, Plancherel and Parseval identities, covariance under orthogonal transformations, transformations of coordinate polynomials and differential operator polynomials, transformations of derivative and Dirac derivative operators, as well as signal width related to band width uncertainty relationships. This section is based on [211]. See also [208, 210].
222 Quaternion and Clifford Fourier Transforms
Left: Four components of the QFMT kernel (v = 2, k = 3). Right: Symmetries of four components of the QFMT kernel (v = 2, k = 3). Source: [198, Fig. 7].
Figure 4.17
Left: High angular resolution. Center: High radial resolution. Right: High radial and angular resolution. Source: [198, Fig. 8]. Figure 4.18
Figure 4.19
Illustration of QFMT scaling. Source: [198, Fig. 9].
Quaternion Fourier transforms 223
Generally speaking, quaternion Fourier transforms (QFT) are since over 20 years a mathematically well researched and frequently applied subject [51]. Yet interesting enough most publications on QFTs concentrate on transformations for signals with domain R2 . First motivated by private communication with T.L. Saaty related to quaternion valued functions over the domain of quaternions, we establish here a genuine Fourier transform with a quaternionic kernel operating on such functions. This section begins by introducing quaternions and their relevant properties, including quaternion domain functions in Section 4.3.3.1.1. For more details on quaternions, we refer the reader to Section 2.8.1. The quaternion domain Fourier transform (QDFT) is defined in Section 4.3.3.2. Many fundamental properties of the QDFT are investigated in Section 4.3.3.3. The conclusions in Section 4.3.3.4 give an outlook into the wide area of possible applications and the rich possibilities of studying related transforms for quaternion domain signals. About quaternions H Quaternion multiplication pq can be alternatively represented by the following matrix vector multiplication [326] 4.3.3.1.1
Sc(pq) (pq)i (pq)j (pq)k
=
pr −pi −pj −pk qr pi pr −pk pj qi pj pk pr −pi qj pk −pj pi pr qk
.
(4.3.77)
The determinant of the above matrix P (p) is simply det P (p) = p4 .
(4.3.78)
If we interpret the four real coefficients of x ∈ H, xr , xi , xj , xk ∈ R as coordinates in R4 , with infinitesimal volume element d4 x = dxr dxi dxj dxk , then the substitution z = ax, a ∈ H, yields z = ax
⇒
d4 z = a4 d4 x,
d4 x = a−4 d4 z,
(4.3.79)
assuming a 6= 0 for the last identity. For the transformation z = axb, a, b, x ∈ H, we set y = xb and then d4 z = a4 d4 y.
(4.3.80)
y˜ = ˜b˜ x,
(4.3.81)
Quaternion conjugation leads to such that −d4 y = d4 y˜ = ˜b4 d4 x˜ = −b4 d4 x,
(4.3.82)
because ˜b = b, d x˜ = dxr (−dxi )(−dxj )(−dxk ) = −d x and similarly d y˜ = −d4 y. Hence d4 z = a4 d4 y = a4 b4 d4 x = ab4 d4 x. (4.3.83) 4
4
4
As expected the rotation (4.3.88) does not change the infinitesimal volume element z = axa−1
⇒
d4 x = aa−1 4 d4 x = d4 x.
(4.3.84)
224 Quaternion and Clifford Fourier Transforms
We follow [151] in defining the following derivative operators ∂e = ∂xr + ∂xi i + ∂xj j + ∂xk k,
(4.3.85)
∂ = ∂xr − ∂xi i − ∂xj j − ∂xk k,
(4.3.86)
where ∂xr = ∂/∂xr , etc. We further define the threedimensional Dirac operator D = ∂e − ∂xr = ∂xi i + ∂xj j + ∂xk k,
∂e = ∂xr + D.
(4.3.87)
Quaternions and reflections and rotations in three and four dimensions The geometry of reflections and rotations in three and four dimensions, expressed in the language of quaternions is discussed in [84, 205, 212, 261], see also Section 2.8. We now briefly recall how important orthogonal transformations in threedimensional and fourdimensional Euclidean space can be expressed by means of quaternions. A threedimensional rotation of the vector part x of the quaternion x ∈ H by the ˆ , leaving the scalar part xr invariant, is given by angle 2α around the axis a 4.3.3.1.2
x0 = axa−1 ,
a = eαˆa ,
ˆ 2 = −1. a
(4.3.88)
For example a = cos α + k sin α = exp(kα) rotates x = i to x0 = aia−1 = ekα ie−kα = e2kα i = (cos 2α + k sin 2α)i = cos(2α)i + sin(2α)j.
(4.3.89)
We further note, that the transformation x0 = axb,
a = eαˆa ,
b = eβˆa ,
(4.3.90)
rotates the x− part by the angle α + β in the q− plane (determined according to ˆ ), and rotates the x+ part by α − β in the q+ plane. Section 4.1.2.3.2, setting f = a The 4D reflection at the real line is given by quaternion conjugation x → x˜, leaving the real line pointwise invariant. The 4D reflection at the 3D hyperplane of pure quaternions is therefore given by x → −˜ x, leaving the 3D hyperplane of pure quaternions pointwise invariant. A reflection at a (pointwise invariant) general line in R4 in the direction of the unit quaternion a ∈ H, a = 1, is given by x → a˜ xa. A reflection at the (pointwise invariant) threedimensional hyperplane orthogonal to the direction in four dimensions given by the unit quaternion a, a = 1, is given by x → −a˜ xa. A general rotation in R4 is given by x → axb,
a, b ∈ H,
a = b = 1.
(4.3.91)
To understand the geometry of this rotation [205], we rewrite the unit quaternions a, b as ˆ a = eαˆa , b = eβ b . (4.3.92) ˆ define two orthogonal twodimensional rotation ˆ and b The pure unit quaternions a
Quaternion Fourier transforms 225
ˆ because the ˆ 6= b, planes in R4 , where without restriction of generality we assume a a,b ˆ ˆ = b has already been discussed in (4.3.90). The q+ plane with orthogonal case a basis and projection a,b ˆ 1+a ˆ ˆ b}, q+ basis : {ˆ a − b,
1 a,b ˆ ˆ q b), q+ = (q + a 2
(4.3.93)
a,b and the orthogonal q− plane orthogonal basis and projection a,b ˆ 1−a ˆ ˆ b}, q− basis : {ˆ a + b,
1 a,b ˆ ˆ q b), q− = (q − a 2
(4.3.94)
a,b a,b such that q = q+ + q− , for all q ∈ H. The transformation x → axb of (4.3.91) then a,b a,b means geometrically a rotation by the angle α − β in the q+ plane (around the q− a,b a,b plane as axis) and a rotation by the angle α + β in the q− plane (around the q+ plane as axis). This also tells us, that for α = β the rotation degenerates to a single a,b twodimensional rotation by 2α in the q− plane, and for α = −β it degenerates to a,b a single twodimensional rotation by 2α in the q+ plane. A general rotary reflection (rotation reflection) in R4 is given by
x → a˜ xb,
a, b ∈ H,
a = b = 1.
(4.3.95)
This rotary reflection has the pointwise invariant line through a + b. In the remaining threedimensional hyperplane, orthogonal to the a + b line, the axis of the rotary ^ reflection is the line in the direction a − b, because a(a − b)b = −(a − b). The rotation plane of the rotary reflection is spanned by the two orthogonal quaternions v1,2 = [a, b](1 ± a ˜b), [a, b] = ab − ba, and the angle of rotation is Γ = π − acrccos(Sc(˜ ab)), [205]. 4.3.3.1.3
Quaternion domain functions
Every real valued quaternion domain func
tion f maps H → R: f : x 7→ f (x) ∈ R,
∀x ∈ H.
(4.3.96)
Every quaternion valued quaternion domain function f maps H → H, its four coefficient functions fr , fi , fj , fk , are in turn real valued quaternion domain functions: f : x 7→ f (x) = fr (x) + fi (x)i + fj (x)j + fk (x)k ∈ H.
(4.3.97)
Quaternion valued quaternion domain functions have been historically studied in [137, 275, 298, 311], and applications are described in [151]. We define for two functions f, g : H → H the following quaternion valued inner product Z f (x)˜ g (x)d4 x
(f, g) =
(4.3.98)
H
with d4 x = dxr dxi dxj dxk ∈ R. Note that quaternion conjugation yields ] (f, g) = (g, f ).
(4.3.99)
226 Quaternion and Clifford Fourier Transforms
This means that the real scalar part of the inner product (f, g) is symmetric 1 hf, gi = [(f, g) + (g, f )] = 2
Z
hf (x)˜ g (x)i0 d4 x ∈ R,
hf, gi = hg, f i.
(4.3.100)
H
We further define the L2 (H; H)norm 27 as kf k =
q
(f, f ) =
q
hf, f i =
sZ
f (x)2 d4 x ≥ 0.
(4.3.101)
H
The quaternion domain module L2 (H; H) is the set of all finite L2 (H; H)norm functions L2 (H; H) = {f f : H → H, kf k ≤ ∞}. (4.3.102) The convolution of two functions f, g ∈ L2 (H; H) is defined as (f ∗ g)(x) =
Z
f (y)g(x − y)d4 y.
(4.3.103)
H
For unit norm signals f ∈ L2 (H; H), kf k = 1, we define the effective spatial width or spatial uncertainty (or signal width) of f in the direction of the unit quaternion a ∈ H, a = 1, as the square root of the variance of the energy distribution of f along the aaxis s Z
∆xa = k(x · a)f k =
(x · a)2 f (x)2 d4 x.
(4.3.104)
H
Also for unit norm signals f , we define the effective spatial width (spatial uncertainty) as the square root of the variance of the energy distribution of f ∆x = kxf k =
sZ
x2 f (x)2 d4 x.
(4.3.105)
H
4.3.3.2 The quaternion domain Fourier transform
Since the traditional quaternion Fourier transform (QFT) [56, 124, 184] is only defined for real or quaternion valued signals over the domain R2 , we newly define the quaternion domain Fourier transform (QDFT) for h ∈ L1 (H; H) as28 ˆ F{h}(ω) = h(ω) =
1 (2π)2
Z
h(x)e−Ix·ω d4 x,
(4.3.106)
H
with x, ω ∈ H, d4 x = dxr dxi dxj dxk ∈ R, and some constant I ∈ H, I 2 = −1. The constant unit pure quaternion I can be chosen specific for each problem. Note that the QDFT of (4.3.106) is steerable due to the free choice of the unit pure quaternion unit I ∈ S 2 . 27
Note that in equation (13) of [184] the square root is missing over the integral in the definition of the L2 (R2 ; H)norm. 28 For quaternionic Fourier transforms on L2 , see Footnote 2 on page 144.
Quaternion Fourier transforms 227
This QDFT definition is left linear ˆ F{αh + βg}(ω) = αh(ω) + βˆ g (ω),
(4.3.107)
for g, h ∈ L1 (H; H) and constants α, β ∈ H. See (4.3.126) for linear combinations of signals with constant quaternion coefficients multiplied from the right. Applying the orthogonal planes split of Section 4.1.2.3.2 to the signal function ˆ we find h = h+ + h− and to the QDFT h ˆ ˆ + (ω) + h ˆ − (ω), h(ω) =h Z 1 −Ix·ω 4 ˆh+ (ω) = 1 h+ (x)e d x= e+Ix·ω h+ (x)d4 x, (2π)2 H (2π)2 H Z Z 1 −Ix·ω 4 ˆh− (ω) = 1 h− (x)e d x= e−Ix·ω h− (x)d4 x. (2π)2 H (2π)2 H
(4.3.108)
Z
(4.3.109) (4.3.110)
Example 4.3.34. Following the suggestion of T. L. Saaty, we QDFT transform the functional quaternion equation29 h(ax) = bh(x),
h : H → H,
(4.3.111)
with quaternion constants a, b ∈ H. We define the auxiliary function ha (x) = h(ax) and compute 1 1 ∼ −1 h(ax)e−Ix·ω d4 x = h(ax)e−ISc((ax) a˜ ω) d4 x 2 2 (2π) H (2π) H Z 1 −1 z=ax ˆ a−1 ω), = h(z)e−ISc(˜za˜ ω) a−4 d4 z = a−4 h(˜ (4.3.112) (2π)2 H
ˆ a (ω) = h
Z
Z
where we used that x · ω = Sc(˜ xω) = Sc(˜ xa ˜a ˜−1 ω) = Sc((ax)∼ a ˜−1 ω) = (ax) · (˜ a−1 ω). For a = α ∈ R we get
(4.3.113)
ˆ α (ω) = 1 h( ˆ ω ), h 4 α α
(4.3.114)
ˆ a (ω) = h(aω), ˆ h
(4.3.115)
and for a ∈ H, a = 1, we get
because for a ∈ H, a = 1, we have a ˜−1 = a. Using relationship (4.3.112) and left linearity we arrive at the QDFT of (4.3.111) ˆ a−1 ω) = bh(ω), ˆ a−4 h(˜
(4.3.116)
ˆ a−1 ω) = a4 bh(ω), ˆ h(˜
(4.3.117)
or equivalently 29
The simplest solutions of this equation take the form h(x) = cxd, bc = ca, with quaternion constants c, d ∈ H. (I thank the anonymous reviewer for this hint.) In the complex domain T. L. Saaty has developed interesting solutions [289].
228 Quaternion and Clifford Fourier Transforms
which seems neither less nor more complicated to solve than the original equation (4.3.111). But note, that for a ∈ H, a = 1, equations (4.3.111) and (4.3.117) become identical, because (4.3.117) then reads ˆ ˆ h(aω) = bh(ω),
(4.3.118)
i.e. then (4.3.111) is invariant under the QDFT operator. An application of (4.3.112) is the fourdimensional inversion at the origin x → −x results in ˆ −1 (ω) = h(−ω). ˆ h (4.3.119) The QDFT can separate the two components of a “complex” signal f : H → R + iR, f (x) = fr (x) + ifi (x), into even and odd components with respect to the inversion x → −x. Let f (x) = fr (x) + ifi (x) = fre (x) + fro (x) + i(fie (x) + fio (x)),
(4.3.120)
with 1 fre (−x) = fre (x) = (fr (x) + fr (−x)), 2 1 o o fr (−x) = −fr (x) = (fr (x) − fr (−x)), 2 e e fi (x) = fi (−x), fio (x) = −fio (x).
(4.3.121)
Using the steerability of the QDFT (4.3.106) we select for I = j (we could also set I = k or any other pure quaternion ⊥ i) and have by linearity fˆ(ω) = fˆre (ω) + fˆro (ω) + i(fˆie (ω) + fˆio (ω)) Z
=
Z
fre (x) cos(x · ω)d4 ω + j
H
fro (x) sin(x · ω)d4 ω
H
Z
+i
fie (x) cos(x
4
· ω)d ω + k
H
Z
fio (x) sin(x · ω)d4 ω.
(4.3.122)
H
Compare [240] for a similar approach to the symmetry analysis of signals f : R → C. 4.3.3.3 Properties of the QDFT
Properties of the QDFT that can easily be established are inversion 1 h(x) = (2π)2
Z
+Ix·ω 4 ˆ h(x)e d ω,
(4.3.123)
H
a shift theorem for g(x) = h(x − a), constant a ∈ H, −Ia·ω ˆ gˆ(ω) = h(ω)e ,
(4.3.124)
and a modulation theorem for m(x) = h(x)eIx·ω0 , constant ω0 ∈ H, ˆ − ω0 ). m(ω) ˆ = h(ω
(4.3.125)
Quaternion Fourier transforms 229
Linear combinations with constant quaternion coefficients α, β ∈ H from the right lead due to (4.1.84) to ˆ ˆ F{hα + gβ} = h(ω)α ˆ(ω)β+ + gˆ(−ω)β− . + + h(−ω)α− + g
(4.3.126)
We define gl (x) = ∂xl h(x), l ∈ {r, i, j, k} for the partial derivative of the signal function h and obtain its QDFT as ˆ gbl (ω) = h(ω)Iω l.
(4.3.127)
ˆ ∂d xr h(ω) = ωr h(ω)I.
(4.3.128)
For example for l = r we obtain
This leads to the QDFT of the derivative operators (4.3.85) and (4.3.86) d m ˆ ∂˜m h(ω) = ω m h(ω)I ,
m m h(ω) = ω ˆ e m h(ω)I ∂d ,
m ∈ N.
(4.3.129)
Applying the derivative operators from the right to the signal function h we further obtain dm (ω)I m , h∂˜m (ω) = hω
dm (ω) = hd e m (ω)I m , h∂ ω
d
m ∈ N.
(4.3.130)
QDFT transformations of the Dirac operator D applied from the left and right, respectively, give m m h(ω) = ω m h(ω)I ˆ [ D ,
dm (ω)I m , m (ω) = hω [ hD
m ∈ N.
(4.3.131)
where the pure quaternion part of the quaternion frequency ω is ω = ω − ωr . The QDFT of mfold powers of coordinates xl , l ∈ {r, i, j, k}, m ∈ N, times the signal function h leads to (dual to (4.3.127)) m m h(ω) = ∂ m h(ω)I ˆ xd . ωl l
(4.3.132)
For example for l = r we obtain ˆ xd r h(ω) = ∂ωr h(ω)I.
(4.3.133) m
j mk r mi If P (xr , xi , xj , xk ) = mr ,mi ,mj ,mk αmr ,mi ,mj ,mk xm r xi xj xk , with quaternion coefficients αmr ,mi ,mj ,mk ∈ H, is a polynomial of the four coordinates {xr , xi , xj , xk }, then the QDFT yields
P
X
αmr
F{P (xr , xi , xj , xk )h}(ω) = mr mi mj mk ˆ ∂ ∂ ∂ h(ω)I mr +mi +mj +mk . ,m ,m ,m ∂ i
j
k
ωr
ωi
ωj
ωk
(4.3.134)
mr ,mi ,mj ,mk
For example for P (x) = a · x = ar xr + ai xi + aj xj + ak xk we obtain ˆ F{(a · x)h}(ω) = (a · ∂˜ω )h(ω)I,
(4.3.135)
230 Quaternion and Clifford Fourier Transforms
with ∂˜ω = ∂ωr + ∂ωi i + ∂ωj j + ∂ωk k and a · ∂˜ω = ar ∂ωr + ai ∂ωi + aj ∂ωj + ak ∂ωk . We have the dual to (4.3.134) result that
X
αmr ,mi ,mj
F{P (∂xr , ∂xi , ∂xj , ∂xk )h}(ω) = mr mi mr +mi +mj +mk ˆ ωi ωj mj ωk mk h(ω)I , ,m ωr k
(4.3.136)
mr ,mi ,mj ,mk
with the special case (dual to (4.3.135)) ˜ ˆ F{(a · ∂)h}(ω) = (a · ω)h(ω)I.
(4.3.137)
Note that (4.3.136) shows how the QDFT (with t = x0 , x1 = xi , x2 = xj , x3 = xk ) can be used to treat important partial differential equations in physics, e.g. the heat equation, wave equation, KleinGordon equation, the Maxwell equations in vacuum, free particle Schr¨odinger and Dirac equations [320, 324, 325, 327]. Equation (4.3.134) leads further (dual to left side of (4.3.129)) to, c ˆ xh(ω) = ∂˜h(ω)I,
m m h(ω) = ∂ ˜m h(ω)I ˆ xd ,
m ∈ N.
(4.3.138)
Multiplying instead with the quaternion conjugate x˜ we obtain (dual to right side of (4.3.129)) ˆ x˜ch(ω) = ∂ h(ω)I,
m m h(ω) = ∂ m h(ω)I ˆ x˜d ,
m ∈ N.
(4.3.139)
Taking only the pure vector part of x, x = x − xr we obtain (dual to (4.3.131)) c ˆ xh(ω) = Dω h(ω)I,
m m h(ω) = D m h(ω)I ˆ xd , ω
m ∈ N,
(4.3.140)
where Dω = ∂ωi i + ∂ωj j + ∂ωk k. We further obtain the following QDFT Plancherel identity, which expresses, that the quaternion valued inner product (4.3.98) of two quaternion domain module functions f, g ∈ L2 (H; H) is given by the quaternion valued inner product of the corresponding QDFTs fˆ and gˆ (f, g) = (fˆ, gˆ). (4.3.141) As corollaries we get the corresponding QDFT Plancherel identity for the scalar inner product of equation (4.3.100) hf, gi = hfˆ, gˆi, (4.3.142) as well as the QDFT Parseval identity kf k = kfˆk.
(4.3.143)
The QDFT Parseval identity means, that the QDFT preserves the signal energy when applied in signal processing. We now define analogous to (4.3.104) for unit norm signals f ∈ L2 (H; H), kf k = 1, the effective spectral width (or band width) of f in the direction of the unit quaternion
Quaternion Fourier transforms 231
a ∈ H, a = 1, as the square root of the variance of the frequency spectrum of f along the aaxis ∆ωa = k(ω · a)fˆk =
sZ
(ω · a)2 fˆ(ω)2 d4 ω.
(4.3.144)
H
We further define the effective spectral width (frequency uncertainty) as the square root of the variance of the energy distribution of fˆ ∆ω = kω fˆk =
sZ
ω2 fˆ(ω)2 d4 ω.
(4.3.145)
H
We can now state the directional uncertainty principle for the QDFT of unit norm signals f ∈ L2 (H; H), kf k = 1 as ∆xa ∆ωb ≥
a · b . 2
(4.3.146)
The uncertainty principle takes the form ∆x∆ω ≥ 1.
(4.3.147)
Equality holds in (4.3.146) and (4.3.147) for Gaussian signals 2
f (x) = Ce−cx ,
0 < c ∈ R,
(4.3.148)
with constant factor C ∈ H. The QDFT of the convolution (4.3.103) of two functions f, g ∈ L2 (H; H) results in \ (f ∗ g)(ω) = (2π)2 [fˆ(ω)ˆ g− (ω) + fˆ(−ω)ˆ g+ (ω)]. (4.3.149) Note that for gˆ+ (ω) = 0 or if fˆ(ω) = fˆ(−ω) we obtain \ (f ∗ g)(ω) = (2π)2 fˆ(ω)ˆ g (ω).
(4.3.150)
The QDFT of the convolution of the two functions f, g ∈ L2 (H; H) in opposite order results in \ (g ∗ f )(ω) = (2π)2 [ˆ g (ω)fˆ− (ω) + gˆ(−ω)fˆ+ (ω)], (4.3.151) and is usually different from (4.3.149), because of the general noncommutativity of f, g ∈ L2 (H; H). An application of the QDFT convolution (4.3.149) is, e.g., the fast convolution (via simple multiplication of the QDFTs in the Fourier domain) of a quaternion domain signal f : H → H with a pair of complex filters g1 (x) = g1,r (x) + g1,i (x)i = g− (x), g2 (x) = g2,r (x) + g2,i (x)i = g+ (x)(−j), choosing I = i in (4.3.106). Next, we study the covariance properties of the QDFT under orthogonal transformations. We find that a threedimensional rotation (4.3.88) of the argument g(x) = h(a−1 xa) leads to ˆ −1 ωa). gˆ(ω) = h(a (4.3.152)
232 Quaternion and Clifford Fourier Transforms
The reflection at the pointwise invariant real scalar line x → x˜, g(x) = h(˜ x) gives ˆ ω e ). gˆ(ω) = −h(
(4.3.153)
The reflection at the threedimensional hyperplane of pure quaternions x → −˜ x, g(x) = h(−˜ x) results in ˆ ω e ). gˆ(ω) = −h(− (4.3.154) The reflection at the pointwise invariant line through a ∈ H, a = 1, x → a˜ xa, g(x) = h(a˜ xa) gives ˆ a−1 ω ˆ ω ea e a), gˆ(ω) = −h(˜ ˜−1 ) = −h(a (4.3.155) because a ˜−1 = a for a = 1. The reflection at the threedimensional hyperplane orthogonal to the line through a ∈ H, a = 1, x → −a˜ xa, g(x) = h(−a˜ xa) results in ˆ e a). gˆ(ω) = −h(−a ω
(4.3.156)
A general fourdimensional rotation in R4 , x → axb, a, b ∈ H, a = b = 1, g(x) = h(axb) leads to ˆ gˆ(ω) = h(aωb). (4.3.157) We have thus studied the behavior of the QDFT under all point group transformations in three and four dimensions (reflections, rotations, rotary reflections, inversions), which are of importance in crystallography. We note, that quaternions have already been employed for the description of crystallographic symmetry in [12] and for the description of root systems of finite groups in three and four dimensions in [99]. 4.3.3.4 Summary on QDFT
We first briefly reviewed quaternion algebra, its expression in terms of matrices and vectors, quaternion derivatives and the Dirac derivative, the orthogonal planes split of quaternions with respect to a single unit pure quaternion, and the description of threedimensional and fourdimensional orthogonal transformations. Then we defined quaternion domain functions, the quaternion module L2 (H; H), convolution of quaternion domain functions, and their effective spatial width (uncertainty). We established the steerable quaternion domain Fourier transform (QDFT) with a free choice a single constant pure unit quaternion in the kernel. We examined the properties of left and right linearity, orthogonal plane split property, and gave an example of the QDFT applied to a functional equation. Further properties studied are the inverse QDFT, shift and modulation theorems, the QDFT of quaternion coordinate polynomials30 multiplied with quaternion domain signals, as well as products with powers of the signal argument x, and the corresponding dual properties (polynomials of partial differential operators, quaternion derivatives and Dirac derivatives). We found that the QDFT can separate the symmetry components of complex signals, 30
Note that real and complex polynomial generated moment invariants have been successfully used for translation, rotation and scale invariant normalized moment description of vector field features, including flow fields [63–65].
Quaternion Fourier transforms 233
and can be applied to many partial differential equations in physics. Quaternion noncommutativity means, that multiplication from the right and left need to be distinguished carefully. Next we established Parseval and Plancherel identities, uncertainty principles and convolution properties for the QDFT. The convolution allows, e.g. fast filtering with pairs of complex filters. Finally we studied the covariance properties of the QDFT under orthogonal transformations of the signal arguments, which may a.o. be of importance for applications in crystallography. We expect that this relatively new quaternionic Fourier transformation may find rich applications in mathematics (e.g. higher dimensional holomorphic functions [151]) and physics, including relativity and spacetime physics, in threedimensional color field processing, neural signal processing, space color video processing, crystallography, quaternion analysis, and for the solution of many types of quaternionic differential equations. We further expect that the QDFT can be successfully extended to localized transforms, e.g. quaternion domain window Fourier transforms, and continuous quaternionic wavelets and quaternionic ridgelets [67]31 . Further research should be done into operator versions of the QDFT, and its related linear canonical transforms, which may open up many further areas of interesting applications, including quantum physics. Especially for applications, discretization and fast implementation with pairs of complex fast Fourier transforms will be of great interest.
4.4
GENERALIZING THE QFT TO A VOLUMETIME FT AND A SPACETIME FT
We begin by recalling isomorphisms of quaternion algebra to subalgebras of higher order Clifford algebras [247] such as H ∼ = Cl(0, 2) ∼ = Cl+ (3, 0). Such isomorphisms together with the generalized GL(Rn,m ) transformation laws for {fˆ± (Ax)}(u) allow us now to generalize the QFT to higher dimensions. This indeed opens up a vast new field of related multivector Fourier transforms, which are in general noncommutative. 4.4.1 QFT generalization to volumetime functions
One of these quaternion algebra to Clifford subalgebra isomorphisms that is of particular relevance in physics exists with a subalgebra of the spacetime algebra [158] Cl3,1 , which is of prime importance in physics, and in applications where time matters as well (motion in time, video sequences, flow fields, global navigation satellite systems (GNSS) ...). The quaternion subalgebra allows to introduce generalizations of the QFT to functions in these higher order Clifford geometric algebras. For example, it allows to generalize the QFT to a spacetime FT [184]. The spacetime algebra is the geometric algebra of R3,1 . We express this isomorphism by introducing an orthonormal (grade 1) vector basis for R3,1 {e 0 , e 1 , e 2 , e 3 }, 31
−e 20 = e 21 = e 22 = e 23 = 1.
(4.4.1)
Note that in [188] real Clifford Fourier transforms served as spectral representation for the construction and study of Clifford wavelets. Analogously we expect that the QDFT can serve as suitable spectral representation in the construction of quaternionic wavelets and ridgelets.
234 Quaternion and Clifford Fourier Transforms
Using this vector basis of R3,1 , the spatial unit volume trivector i3 and total fourdimensional (hyper volume) pseudoscalar i4 can be expressed by i3 = e 1 e 2 e 3 ,
i23 = −1,
i4 = e 0 e 1 e 2 e 3 ,
i24 = −1.
(4.4.2)
We emphasize the fact that the vector e 0 , the 3D volume trivector i3 , and the 4D pseudoscalar i4 , all square to minus one and anticommute mutually. Examining the geometric algebra multiplication laws of e 0 , i3 and i4 shows indeed that the arising volumetime subalgebra Vt of the spacetime algebra is isomorphic (see Sections 4.1 and 4.2 of [142]) to the quaternion algebra H Vt = span{1, e 0 , i3 , i4 } ∼ = H = span{1, i, j, k},
(4.4.3)
Note especially that ∗ i3 = −e 0 e 0 i3 = e 0 (−i4 ) = e 0 i−1 4 = e0 ,
(4.4.4)
which shows that i3 is dual to e 0 in Cl3,1 . The time direction e 0 determines therefore the complimentary 3D space with pseudoscalar i3 as well! Based on the isomorphism (4.4.3) we now define a Fourier transform for volumetime module functions f ∈ L2 (R3,1 ; Vt ). Definition 4.4.1 (volumetime Fourier transform (VtFT)). The volumetime ◦
Fourier transform f : R3,1 → Vt of volumetime module functions32 f ∈ L2 (R3,1 ; Vt ), with spacetime vectors x = te0 + ~x ∈ R3,1 , ~x = xe1 + ye2 + ze3 ∈ R3 , and spacetime frequency vectors u = se0 + ~u ∈ R3,1 , ~u = ue1 + ve2 + we3 ∈ R3 is defined as ◦
Z
e−e0 ts f (x) e−i3 ~x·~u d4 x ,
f (u) =
(4.4.5)
R3,1
with the differential spacetime integration volume d4 x = dtdxdydz. Remark 4.4.2. The 3D integration part Z
f (x) e−i3 ~x·~ω d3~x
(4.4.6)
in (4.4.5) and (4.4.17) fully corresponds to the GA FT in Cl(3, 0), as defined in (5.1.7), compare [185, 251]. The VtFT can be inverted in close analogy to (4.1.10) by using 1 f (x) = (2π)4
Z
◦
ee 0 ts f (u) ei3 ~x·~u d4 u ,
R3,1
with d4 u = dsdudvdw.
32
For the Fourier transform on L1 and L2 , see Footnote 2 on page 144.
(4.4.7)
Quaternion Fourier transforms 235
The f± split (4.1.12) of the QFT combined with the isomorphism (4.4.3) now yields for volumetime module functions f ∈ L2 (R3,1 ; Vt ) f = f+ + f− ,
1 f+ = (f + e 0 f i3 ), 2
1 f− = (f − e 0 f i3 ). 2
(4.4.8)
Rewriting the split (4.4.8) with the duality relation (4.4.4) to 1 f± = (f ± e 0 f e ∗0 ) 2
(4.4.9)
shows that it naturally only depends on the physical spacetime split, i.e. on the choice of the time direction e 0 . Applying our new VtFT of Definition 4.4.1 to the split functions f± of (4.4.9) results in a VtFT formula which corresponds to Theorem 4.3.30 ◦
◦
◦
f = f + + f −, ◦
f± =
Z R3,1
f± e−i3 ( ~x·~u∓ ts ) d4 x =
Z R3,1
e−e 0 ( ts∓~x·~u ) f± d4 x .
(4.4.10)
◦
Note especially that the f + part in (4.4.10) has the kernel with the flat Minkowski metric ts − ~x · ~u in the exponent. (Compare Section 4.4.3 for further interpretation.) Definition 4.4.1 preserves the form of the GL transformation properties of Section 4.1.1.4. We get the GL(R3,1 ) transformation properties of (4.4.5) simply by inserting in Theorem 4.1.11 transformations A ∈ GL(R3,1 ) and replacing Ue 1 by Ue 0 . Theorem 4.4.3 (GL(R3,1 ) transformation properties of the VtFT). The VtFT of a Vt module function f ∈ L2 (R2 ; Vt ) with a GL(R3,1 ) transformation A of its vector argument is given by ◦
◦
{f (Ax)}◦ (u) =  det A−1  { f − (A−1 u) + f + (Ue0 A−1 Ue0 u) }.
(4.4.11)
In physical applications proper Lorentz transformations with  det A = 1 are most relevant, so the  det A−1  factor in (4.4.11) can then naturally be omitted. For all kinds of applications it is of interest to know whether we can push the QFT generalization established by the VtFT for volumetime module functions f ∈ L2 (R3,1 ; Vt ) even further, i.e. if even more general spacetime algebra functions can be treated meaningfully with the VtFT. That this is indeed the case will be shown in the next subsection. 4.4.2 Generalization to full spacetime algebra functions
We now explain how we can drop in the VtFT Definition 4.4.1 the restriction to volumetime functions f ∈ L2 (R3,1 ; Vt ). The key to this is found in the commutativity of the unit volume trivector i3 of the right side exponential factor in (4.4.5) with all spatial vectors {e 1 , e 2 , e 3 } i3 e k = e k i3 ,
1 ≤ k ≤ 3.
(4.4.12)
236 Quaternion and Clifford Fourier Transforms
This directly leads us to the right linearity of the VtFT {f α}◦ (u) =
Z
e−e 0 ts f (x) α e−i3 ~x·~u d4 x
R3,1 ◦
Z
e−e 0 ts f (x) e−i3 ~x·~u d4 x α = f (u) α,
= R3,1
∀ const. α ∈ Cl3,0 ,
(4.4.13)
where Cl(3, 0) is the eightdimensional Clifford geometric algebra of R3,0 , i.e. the 3D space subalgebra of Cl(3, 1) spanned by {1, e 1 , e 2 , e 3 , e 2 e 3 , e 3 e 1 , e 1 e 2 , i3 }.
(4.4.14)
Naturally this right linearity also holds for the inverse transformation Z 1 e 0 ts f◦ (u)α ei3 ~x·~u d4 u ∀ const. α ∈ Cl . f (x) α = e 3,0 (2π)4 R3,1
(4.4.15)
Now all 16 basis multivectors of Cl3,1 can be obtained by successive geometric multiplications of 1 and e 0 (or alternatively of i3 and i4 , etc.) with the three spatial vectors {e 1 , e 2 , e 3 } from the right33 {1, e 1 , e 2 , e 3 , e 2 e 3 , e 3 e 1 , e 1 e 2 , i3 , e 0 , e 0 e 1 , e 0 e 2 , e 0 e 3 , e 0 e 2 e 3 , e 0 e 3 e 1 , e 0 e 1 e 2 , e 0 i3 }.
(4.4.16)
We now have laid all the groundwork for the full spacetime algebra generalization of the VtFT of Definition 4.4.1:
Definition 4.4.4 (Spacetime Fourier transform (SFT)). The SFT f : R3,1 → Cl(3, 1) of a (16 dimensional) spacetime algebra Cl(3, 1) module function34 f ∈ L2 (R3,1 ; Cl(3, 1)) with spacetime vectors x = te0 +~x ∈ R3,1 , ~x = xe1 +ye2 +ze3 ∈ R3 , and spacetime frequency vectors u = se0 + ~u ∈ R3,1 , ~u = ue1 + ve2 + we3 ∈ R3 is defined by Z F {f }(u) = f (u) = e−e0 ts f (x) e−i3 ~x·~u d4 x , (4.4.17) SF T
R3,1
4
with d x = dtdxdydz . Because of (4.4.13) Definition 4.4.4 is fully compatible with Definition 4.4.1, since (4.4.17) is nothing but a (right) linear combination of (4.4.5). To show this, we can use Cl(3, 1) ∼ = Vt ⊗ {1, e 1 , e 2 , e 3 }, or we can e.g. rewrite a general = Vt ⊗ Cl(3, 0)+ ∼ spacetime algebra module function f ∈ L2 (R3,1 ; Cl3,1 ) as a (right) linear combination of four volumetime subalgebra module L2 (R3,1 ; Vt ) functions f = fs + f1 e 1 + f2 e 2 + f3 e 3 + f23 e 2 e 3 + f31 e 3 e 1 + f12 e 1 e 2 + f123 i3 + f0 e 0 + f01 e 0 e 1 + f02 e 0 e 2 + f03 e 0 e 3 + f023 e 0 e 2 e 3 + f031 e 0 e 3 e 1 + f012 e 0 e 1 e 2 + f4 i4 = fs + f0 e 0 + f123 i3 + f4 i4 + {f1 + f01 e 0 + f23 i3 + f023 i4 } e 1 + + {f2 + f02 e 0 + f31 i3 + f031 i4 } e 2 + {f3 + f03 e 0 + f12 i3 + f012 i4 } e 3 .
(4.4.18)
+ Cl3,1 is also isomorphic to the tensor product Vt ⊗ Cl3,0 , with Vt defined as in Section 4.4.1 + and Cl3,0 defined as in Section 4.1.1.1.1. (See [142], Sections 4.1 and 4.2.) 34 For the Fourier transform on L1 and L2 , see Footnote 2 on page 144. 33
Quaternion Fourier transforms 237
The four L2 (R3,1 ; Vt ) functions of (4.4.18) are {fs + f0 e 0 + f123 i3 + f4 i4 , f1 + f01 e 0 + f23 i3 + f023 i4 , f2 + f02 e 0 + f31 i3 + f031 i4 , f3 + f03 e 0 + f12 i3 + f012 i4 }, where all 16 coefficient functions {fs , f0 , f1 , · · · , f4 } belong to L2 (R3,1 , R). Because of (4.4.15) the general SFT of Clifford module L2 (R3,1 ; Cl3,1 ) functions of Definition 4.4.4 is also invertible Z 1 f (x) = ee 0 ts f (u) ei3 ~x·~u d 4 u. (4.4.19) 4 (2π) R3,1 4.4.3 Spacetime Fourier transform of f± split parts and physical interpretation
Further application of analogous (right) linearity arguments also yield that the split (4.4.8) and (4.4.9) can also be applied to general multivector Clifford module functions f ∈ L2 (R3,1 ; Cl3,1 ). In (4.4.8) and (4.4.9), we can thus simply replace the L2 (R3,1 ; Vt ) functions by L2 (R3,1 ; Cl3,1 ) functions35 . This carries on to the general SFTs of the split functions f± , which are formally identical to (4.4.10) if we again replace the L2 (R3,1 ; Vt ) functions by L2 (R3,1 ; Cl3,1 ) functions. We can therefore rewrite the SFT (4.4.17) for f ∈ L2 (R3,1 ; Cl3,1 ) as
Z
f = f+ + f− = Z
= R3,1
R3,1
Z
−i3 ( ~ x·~ u− ts ) 4
d x+
f+ e
R3,1
e−e 0 ( ts−~x·~u ) f+ d4 x +
Z R3,1
f− e−i3 ( ~x·~u+ ts ) d4 x
e−e 0 ( ts+~x·~u ) f− d4 x.
(4.4.20)
Complex spacetime Fourier transformations, with exp{−i(~x · ~u − ts)} (where i ∈ C) as the related complex kernel, are e.g. used for electromagnetic fields in spatially dispersive media [8] or in electromagnetic wavelet theory [231]. In physics f+ can be interpreted as (time dependent) multivector amplitude of a rightward (forward) moving wave packetixmultivector wave packet!left, right and f− as that of a leftward (backward) moving wave packet. We therefore see that these physically important wave packets arise absolutely naturally from elementary purely algebraic considerations. But we emphasize that both the noncommutative multivector structure and the geometric interpretation (e.g. of i3 as oriented 3D spatial volume trivector) go beyond conventional treatment. We get the consequent generalization of Theorem 4.1.11, i.e. the GL(R3,1 ) transformation properties of the SFT in the form of Theorem 4.4.5 (GL(R3,1 ) transformation properties of the SFT). The SFT of a Cl3,1 module function f ∈ L2 (R2 ; Cl3,1 ) with a GL(R3,1 ) transformation A of its vector argument is given by
{f (Ax)} (u) =  det A−1  { f − (A−1 u) + f + (Ue0 A−1 Ue0 u) }.
(4.4.21)
This concludes our brief example of a higher dimensional multivector generalization of the QFT for L2 (R2 ; H) functions to a SFT for L2 (R3,1 ; Cl3,1 ) functions. We again emphasize that mathematically many other generalizations are in fact possible and we expect a number of them to be of great utility in applications. 35
Again the f± split (4.4.9) solely depends on the choice of time direction e 0 .
238 Quaternion and Clifford Fourier Transforms
4.4.4 Directional 4D spacetime uncertainty principle
We now generalize the directional QFT uncertainty principle of Section 4.2.1 to spacetime algebra multivector functions. Theorem 4.4.6 (Directional 4D spacetime uncertainty principle). For two arbitrary constant spacetime vectors a, b ∈ R3,1 (selecting two directions), and f ∈ L2 (R3,1 , G3,1 ), x1/2 f ∈ L2 (R3,1 , G3,1 ) we obtain Z R3,1
2
2 4
(at t − ~a · ~x) f (x) d x ≥
Z R3,1
(bt ωt − ~b · ω ~ )2 F{f }(ω)2 d4 ω
i (2π)4 h (at bt − ~a · ~b)2 F−2 + (at bt + ~a · ~b)2 F+2 , 4
with energies of the left and right traveling wave packets: Z
F± =
R2
f± (x)2 d2 x.
(4.4.22)
Remark 4.4.7. The proof is strictly analogous to the proof of the directional QFT uncertainty principle inR Section 4.2.1. In that proof we simply need to replace the inteR gral R2 by the integral R3,1 , the infinitesimal 2D area element d2 x by the infinitesimal 4D volume element d4 x, etc. Applying the Parseval theorem in four dimensions leads to the factor of (2π)4 instead of (2π)2 . 4.4.5 Summary of QFT generalization to spacetime FT
As already pointed out at the end of Section 4.1.1, general coordinate free formulation in combination with quaternion to Clifford subalgebra isomorphisms opens the door to a wide range of QFT generalizations to these Clifford geometric algebras. These noncommutative multivector Fourier transforms act on functions from Rm,n , m, n ∈ N0 to Clifford geometric algebras Clm,n (or appropriate subalgebras). We demonstrated this by establishing two multivector Fourier transforms: the volumetime and the spacetime Fourier transforms. There the quaternion ± split corresponds to the relativistic split of spacetime into time and space, e.g. the rest frame of an observer. The split of the SFT corresponds then to analyzing a spacetime multivector function in terms of left and right traveling multivector wave packets [184]. This generalization to higher dimensional algebras is also possible for the new directional QFT uncertainty principle. We demonstrated this with the physically most relevant generalization to spacetime algebra functions resulting in a (left and right traveling) multivector wave packet analysis. The energies of these wave packets determine together with two arbitrary spacetime directions (one spacetime vector and one relativistic wave vector) the resulting uncertainty threshold. Regarding the proliferation of higher dimensional theories in theoretical physics, and the model of Euclidean space in the socalled conformal geometric algebra Cl(4, 1), which has recently become of interest for applications in computer graphics,
Quaternion Fourier transforms 239
computer vision and geometric reasoning [111, 244], even higher dimensional generalizations of the directional QFT uncertainty principle may be of interest in the future. They await application, e.g. in the fields of dynamic fluid and gas flows, seismic analysis, to electromagnetic phenomena, in short wherever spatial data are recorded with time. We expect other generalizations of the QFT obtained by the same methods to be of great potential use as well. The SFT has been further studied in [218] and [122].
CHAPTER
5
Clifford Fourier transforms
In the field of applied mathematics the Fourier transform has developed into an important tool. It is a powerful method for solving partial differential equations. The Fourier transform provides also a technique for signal analysis where the signal from the original domain is transformed to the spectral or frequency domain. In the frequency domain many characteristics of the signal are revealed. But how to extend the Fourier transform to Clifford’s geometric algebras1 ? This central chapter begins with an introduction to the Clifford Fourier transforms (CFT) in Cl(3, 0), the geometric algebra of threedimensional Euclidean space R3 , one of the most studied and applied versions of the CFT. Next is the generalization to onesided CFTs in ndimensional Euclidean space Clifford algebras Cl(n, 0) and quadratic space Clifford algebras Cl(p, q). After that an alternative operator exponential approach to onesided CFTs is presented. Then we turn our attention to twosided CFTs in Clifford algebras Cl(p, q), which makes sense due to the noncommutativity of Clifford algebras. Distinctions are made based on the number of square roots of −1 (and therefore kernel factors) involved. Of further interest are the uncertainty and convolution properties of CFTs. Finally, special CFTs (windowed, FourierMellin and in conformal geometric algebra) are investigated. Note that a thorough understanding of the CFT is also essential for constructing Clifford wavelets.
5.1
OVERVIEW OF CLIFFORD FOURIER TRANSFORMS
In this overview we mainly follow the tutorial [183]. We concentrate on introducing Clifford Fourier transforms (CFT) taking multivector functions with values in Cl(3, 0) as example. In this overview we omit proofs, because as it turns out the CFT in Cl(3, 0) is a special case of the CFT in Cl(n, 0) treated in full detail in Section 5.2.1. We proceed as follows. 1 In the geometric algebra literature [161] instead of the mathematical notation Cl(p, q) = Clp,q the notation Gp,q is widely in use. It is convention to abbreviate Gn,0 to Gn . We use the words Clifford algebra and geometric algebra interchangeably, similarly the notions of geometric algebra FT and Clifford FT, and we use both notations Clp,q , Cl(p, q) and Gp,q . Nowadays geometric algebra is often understood as Clifford algebra together with geometric interpretation based on the underlying vector space and its subspaces.
DOI: 10.1201/97810031844785
241
242 Quaternion and Clifford Fourier Transforms
First, the basic concept of multivector functions and their vector derivative in geometric algebra (GA) is briefly reviewed. Second, beginning with the Fourier transform on a scalar function we generalize to a real Fourier transform on GA multivectorvalued functions (f : R3 → Cl(3, 0)). Third, we show a set of important properties of the Clifford Fourier transform (CFT) on Cl(3, 0) multivectorvalued functions such as differentiation properties, and the Plancherel theorem. We round off the treatment of the CFT (at the end of this overview) by applying the Clifford Fourier transform properties for showing an uncertainty principle for Cl(3, 0) multivector functions. Historically, Brackx et al. [33] extended the Fourier transform to multivectorvalued functiondistributions in Cl(0, n) with compact support. A related applied approach for hypercomplex Clifford Fourier transformations in Cl(0, n) was followed by B¨ ulow et. al. [58]. By extending the classical trigonometric exponential function exp(j x∗ξ) (where ∗ denotes the scalar product of x ∈ Rm with ξ ∈ Rm , j the imaginary unit) in [243,259], McIntosh et. al. generalized the classical Fourier transform. Applied to a function of m real variables this generalized Fourier transform is holomorphic in m complex variables and its inverse is monogenic in m+1 real variables, thereby effectively extending the function of m real variables to a monogenic function of m + 1 real variables (with values in a complex Clifford algebra). This generalization showed significant applications to harmonic analysis, especially to singular integrals on surfaces in Rm+1 . Based on this approach Kou and Qian obtained a Clifford PayleyWiener theorem and derived Shannon interpolation of bandlimited functions using the monogenic sinc function [286, and references therein]. The Clifford PayleyWiener theorem also allows to derive leftentire (leftmonogenic in the whole Rm+1 ) functions from square integrable functions on Rm with compact support. In this introductory overview we adopt and expand the generalization of the Fourier transform in Clifford geometric algebra G3 = Cl(3, 0) as suggested by Ebling and Scheuermann [117]. We explicitly introduce properties of the real2 Clifford geometric algebra Fourier transform (CFT), which are useful to establish the uncertainty principle for G3 multivector functions. 5.1.1 Clifford Fourier transform (CFT)
We first present the classical Fourier transform in R and generalize it to Clifford’s geometric algebra G3 = Cl(3, 0). Generalizations to other dimensions can be found in [182, 184, 185]. 5.1.1.1 Fourier transform in R
Popoulis [277] defined the Fourier transform and its inverse as follows:
2
The meaning of real in this context is that we use the oriented three dimensional volume element i3 = e123 of the geometric algebra G3 = Cl(3, 0) over the field of the reals R to construct the kernel of the Clifford Fourier transformation of Definition 5.1.3. This i3 has a clear geometric interpretation.
Clifford Fourier transforms 243
Table 5.1
Properties of the traditional Fourier transform. Source: [251, Tab. 2]. Property Linearity Delay Shift Scaling Convolution Derivative Parseval theorem
Function αf (x)+βg(x) f (x − a) eiω0 x f (x) f (ax) (f ? g)(x) f (n) (x) R 2 R f (x) dx
Fourier Transform αF{f }(ω)+ βF{g}(ω) e−iωa F{f }(ω) F{f }(ω − ω0 ) 1 ω a F{f }( a ) F{f }(ω) F{g}(ω) n (iω) F{f }(ω) R 1 2 2π R F{f }(ω) dω
Definition 5.1.1. For an integrable function3 f ∈ L2 (R), the Fourier transform of f is the function F{f }: R → C given by F{f }(ω) =
Z
f (x) e−iωx dx,
(5.1.1)
R
where i, i2 = −1, is the unit imaginary, and exp(−iωx) = cos(ωx) + i sin(ωx). The function F{f }(ω) has the general form F{f }(ω) = A(ω) + iB(ω) = C(ω)eiφ(ω) .
(5.1.2)
C(ω) is called the Fourier spectrum of f (t), C 2 (ω) its energy spectrum, and φ(ω) its phase angle. Definition 5.1.2. If F{f }(ω) ∈ L2 (R) and f ∈ L2 (R), the inverse Fourier transform is given by F −1 [F{f }(ω)] = f (x) =
1 2π
Z
F{f }(ω) eiωx dω.
(5.1.3)
R
The following Table 5.1 summarizes some basic properties of the Fourier transform. 5.1.1.2 Clifford geometric algebra Fourier transform in G3 = Cl(3, 0)
Here we only give a brief overview. Consider a multivector valued function f (x) in G3 , i.e. f : R3 → G3 = Cl(3, 0) where x is a vector variable. f (x) can be decomposed as f (x) =
X
fA (x)eA = f0 (x) + f1 (x)e1 + f2 (x)e2 + f3 (x)e3
A
+ f12 (x) e12 + f31 (x)e31 + f23 (x)e23 + f123 (x)e123 , 3
For the Fourier transform on L1 and L2 , see Footnote 2 on page 144.
(5.1.4)
244 Quaternion and Clifford Fourier Transforms
where the fA are eight realvalued functions. Equation (5.1.4) can also be written as (compare Table 2.2) f (x) = [f0 (x) + f123 (x)i3 ] + [f1 (x) + f23 (x)i3 ] e1 + [f2 (x) + f31 (x)i3 ] e2 + [f3 (x) + f12 (x)i3 ] e3 .
(5.1.5)
Equation (5.1.5) can be regarded as a set of four complex functions. This motivates the extension of the Fourier transform to G3 multivector functions f . We will call this the Clifford Fourier transform (CFT). Alternatively to (5.1.7), B¨ ulow et. al. [58] extended the real Fourier transform to the ndimensional geometric algebra G0,n = Cl(0, n). This variant of the Clifford Fourier transform of a multivector valued function in G0,n is given by F{f }(ω) =
Z
f (x) Rn
n Y
e−ek 2πωk xk dn x,
(5.1.6)
k=1
where x=
k=n X k=1
xk ek ,
ω=
k=n X
ωk e k ,
and
ei · ej = −δij
i, j = 1, 2, ..., n.
k=1
Yet in the following we will adopt (compare [117]) Definition 5.1.3. The Clifford Fourier transform4 of f (x) is the function F{f }: R3 → G3 given by Z f (x) e−i3 ω ·x d3 x, (5.1.7) F{f }(ω) = R3
Note that we can write ω = ω1 e1 + ω2 e2 + ω3 e3 , x = x1 e1 + x2 e2 + x3 e3 with e1 , e2 , e3 the basis vectors of R3 and d3 x =
dx1 ∧ dx2 ∧ dx3 i3
(5.1.8)
Theorem 5.1.4. The Clifford Fourier transform F{f } of f ∈ L2 (R3 , G3 ), R 2 3 R3 kf k d x < ∞ is invertible and its inverse is calculated by F −1 [F{f }](x) = f (x) =
1 (2π)3
Z
F{f }(ω) ei3 ω ·x d3 ω.
(5.1.9)
R3
Equation (5.1.9) is called the Clifford Fourier integral theorem. It describes how to get from the transform back to the original function f .
4
For the Fourier transform on L1 and L2 , see Footnote 2 on page 144.
Clifford Fourier transforms 245
5.1.2 Overview of basic CFT properties
We summarize some important properties of the Clifford Fourier transform which are similar to the traditional scalar Fourier transform properties. Most can be proven via substitution of variables. Details of proofs are given in Section 5.2.1 for the more general case of Cl(n, 0). Linearity. If f (x) = αf1 (x) + βf2 (x) for constants α and β, f1 (x), f2 (x) ∈ G3 then by construction F{f }(ω) = αF{f1 }(ω) + βF{f2 }(ω).
(5.1.10)
Delay property. If the argument of f (x) is offset by a constant vector a, i.e. fd (x) = f (x − a), then F{fd }(ω) = e−i3 ω ·a F{f }(ω).
(5.1.11)
Scaling property. Let a be a positive scalar constant, then the Clifford Fourier transform of the function fa (x) = f (ax) becomes F{fa }(ω) =
1 ω F{f }( ). a3 a
(5.1.12)
Shift property. If ω 0 ∈ R3 and f0 (x) = f (x)ei3 ω 0 ·x , then F{f0 }(ω) = F{f }(ω − ω 0 )
(5.1.13)
The shift property shows that the multiplication by ei3 ω 0 ·x shifts the CFT of the multivector function f (x) so that it becomes centered on the point ω = ω 0 in the frequency domain. 5.1.3 Clifford Fourier transform and differentiation
The CFT differentiation properties also resemble that of the traditional scalar Fourier transform of Table 5.1. For details of proofs we refer to Section 5.2.1 on the more general case of Cl(n, 0). The Clifford Fourier transform of the vector differential of f (x) is F{a · ∇f }(ω) = i3 a · ω F{f }(ω).
(5.1.14)
The vector differential in the a direction is a · ∇f (x) = i3 F −1 [a · ωF{f }](x). This proves (5.1.14). Setting a = ek we get for a partial derivative of f (x) F{∂k f }(ω) = i3 ωk F{f }(ω),
k = 1, 2, 3.
(5.1.15)
Similarly, we can find the derivatives of second order, i.e. F{a · ∇ b · ∇f }(ω) = −a · ω b · ω F{f }(ω).
(5.1.16)
246 Quaternion and Clifford Fourier Transforms
For a = ek , b = el we therefore get F{∂k ∂l }(ω) = −ωk ωl F{f }(ω)
k, l = 1, 2, 3.
(5.1.17)
If x is a vector variable, then F{xf (x)}(ω) = i3 ∇ω F{f }(ω)
(5.1.18)
The Clifford Fourier transform of a · x f (x) gives F{a · x f (x)}(ω) = i3 a · ∇ω F{f }(ω).
(5.1.19)
For a = ek (k = 1, 2, 3) we get F{xk f (x)}(ω) = i3
∂ F{f }(ω). ∂ωk
(5.1.20)
The Clifford Fourier transform of the vector derivative is F{∇f }(ω) = i3 ωF{f }(ω)
(5.1.21)
F{∇2 f }(ω) = −ω 2 F{f }(ω)
(5.1.22)
and of the Laplace operator
In general5 we get F{∇m f } = (i3 ω)m F{f }(ω),
m ∈ N.
(5.1.23)
5.1.4 Convolution and CFT
The most important property of the Clifford Fourier transform for signal processing applications is the convolution theorem. For detailed proofs we refer to Section 5.2.1 on the more general case of Cl(n, 0). Because of the nonAbelian geometric product we have the following definition: Definition 5.1.5. Let f and g be multivector valued functions and both have Clifford Fourier transforms, then the convolution of f and g is denoted f ?g, and defined by Z
(f ?g)(x) =
f (y)g(x − y) d3 y,
(5.1.24)
R3
Theorem 5.1.6. The Clifford Fourier transform of the convolution of f (x) and g(x) is equal to the product of the Clifford Fourier transforms of f (x) and g(x), i.e F{f ?g}(ω) = F{f }(ω)F{g}(ω),
(5.1.25)
5 This general formula should prove very useful for transforming partial differential equations (more precisely: vector derivative equations) into algebraic equations.
Clifford Fourier transforms 247
5.1.5 Plancherel and Parseval theorems
Just as in the case of the traditional scalar Fourier transform, the Plancherel theorem in the geometric algebra G3 = Cl(3, 0) relates two multivector functions with their Clifford Fourier transforms. For detailed proofs we refer to Section 5.2.1 on the more general case of Cl(n, 0), fe means the reverse of f , see Section 2.1.1.6. Theorem 5.1.7. Assume that f1 (x), f2 (x) ∈ G3 with Clifford Fourier transform F{f1 }(ω) and F{f2 }(ω) respectively, then hf1 (x) f^ 2 (x)iV =
1 ^ hF{f1 }(ω)F{f 2 }(ω)iV , (2π)3
(5.1.26)
where we define the volume integral hf1 (x) f^ 2 (x)iV =
Z R3
3 f1 (x)f^ 2 (x)d x
(5.1.27)
In particular, with f1 (x) = f2 (x) = f (x), we get the (multivector) Parseval theorem, i.e. 1 ^ hf (x) fg (x)iV = hF{f }(ω) F{f }(ω)iV . (5.1.28) (2π)3 Note that equation (5.1.26) is multivector valued. This theorem holds for each grade k of the multivectors on both sides of equation (5.1.26) hhf1 (x) f^ 2 (x)iV ik =
1 ^ hhF{f1 }(ω) F{f 2 }(ω)iV ik , (2π)3
k = 0, 1, 2, 3.
(5.1.29)
For k = 0 and according to equations (5.1.27) and (5.2.147), the (scalar) Parseval theorem becomes Z Z 1 2 3 kF{f }(ω)k2 d3 ω. (5.1.30) kf (x)k d x = (2π)3 R3 R3 Because of the similarity with equation (5.1.2) we call R3 kf (x)k2 d3 x the energy of f . A summary of the properties of the Clifford Fourier transform (CFT) for Cl(3, 0) can be found in Table 5.2. R
5.1.6
Precision limits – minimal variance products for geometric algebra Fourier transforms
The uncertainty principle plays an important role in the development and understanding of quantum physics. It is also central for information processing [287]. In quantum physics it states, e.g., that particle momentum and position cannot be simultaneously known. In Fourier analysis such conjugate entities correspond to a function and its Fourier transform which cannot both be simultaneously sharply localized. Furthermore, much work (e.g. [74, 287]) has been devoted to extending the uncertainty principle to a function and its Fourier transform. From the view point of geometric algebra an uncertainty principle gives us information about how a multivector valued function and its Clifford Fourier transform are related. For detailed proofs we refer to Section 5.2.1 on the more general case of Cl(n, 0).
248 Quaternion and Clifford Fourier Transforms
Table 5.2
Properties of the Clifford Fourier transform (CFT). Source: [251, Tab. 3].
Property Linearity Delay Shift Scaling Convolution Vec. diff.
Vec. deriv. Plancherel T. sc. Parseval T.
Multivector Function αf (x)+β g(x) f (x − a) ei3 ω 0 ·x f (x) f (ax) (f ?g)(x) a · ∇f (x) a · x f (x) xf (x) ∇m f (x) hf1 (x)f] 2 (x)iV =
CFT αF{f }(ω)+ βF{g}(ω) e−i3 ω ·a F{f }(ω) F{f }(ω − ω 0 ) 1 ω a3 F{f }( a ) F{f }(ω) F{g}(ω) i3 a · ωF{f }(ω) i3 a · ∇ω F{f }(ω) i3 ∇ω F{f }(ω) (i3 ω)m F{f }(ω) 1 ^ 3 hF{f1 }(ω)F{f2 }(ω)iV (2π) 1 (2π)3
kf (x)k2 d3 x =
R
R3
R
R3
kF{f }(ω)k2 d3 ω
Theorem 5.1.8. Let f be a multivector valued function in G3 which has the Clifford R Fourier transform F{f }(ω). Assume R3 kf (x)k2 d3 x = F < ∞, then the following inequality holds for arbitrary constant vectors a, b: Z
(a · x)2 kf (x)k2 d3 x
Z
R3
(b · ω)2 kF{f }(ω)k2 d3 ω ≥ (a · b)2
R3
(2π)3 2 F 4
(5.1.31)
Choosing b = ±a, with a2 = 1 we get the following uncertainty principle, i.e. Z
2
2
3
(a · x) kf (x)k d x
Z
R3
(a · ω)2 kF{f }(ω)k2 d3 ω ≥
R3
(2π)3 2 F . 4
(5.1.32)
In (5.1.32) equality holds for Gaussian multivector valued functions (See appendix A.3) 2 f (x) = C0 e−k x (5.1.33) where C0 ∈ G3 is a constant multivector, 0 < k ∈ R. Theorem 5.1.9. For a · b =0, we get Z
(a · x)2 kf (x)k2 d3 x
R3
Z
(b · ω)2 kF{f }(ω)k2 d3 ω ≥ 0.
(5.1.34)
R3
Note that with x2 =
3 X k=1
x2k =
3 X
(ek · x)2 ,
k=1
ω2 =
3 X l=1
ωl2 =
3 X
(el · ω)2
(5.1.35)
l=1
we can extend the formula of the uncertainty principle to Theorem 5.1.10. Under the same assumptions as in theorem 5.1.8, we obtain Z R3
2
2
3
x kf (x)k d x
Z R3
ω 2 kF{f }(ω)k2 d3 ω ≥ 3
(2π)3 2 F . 4
(5.1.36)
Clifford Fourier transforms 249
5.1.7 Summary of CFT overview
We introduced how the (real) Clifford geometric algebra Fourier transform (CFT) extends the traditional Fourier transform on scalar functions to multivector functions. Basic properties and rules for differentiation, convolution, the Plancherel and Parseval theorems were given for this new CFT. It is known that the Fourier transform is successfully applied to solving physical equations such as the heat equation, wave equations, etc. Therefore geometric algebra and the Clifford Fourier transform can now be applied to solve such problems involving scalar, vector, bivector and pseudoscalar fields. A close relative of the geometric algebra CFT is the quaternion Fourier transform (QFT) of Chapter 4, since quaternions are nothing but the bivector subalgebra of the geometric algebra of threedimensional Euclidean space R3 . Based on the QFT far reaching generalizations to high dimensional geometric algebra are possible, including the geometric algebra of spacetime [184], see also Section 4.4. We also gave uncertainty principles in the geometric algebra G3 = Cl(3, 0) which describe how a multivectorvalued function and its Clifford geometric algebra Fourier transform relate. In the field of information theory and image processing these geometric algebra uncertainty principles establish bounds or limits for the effective width times frequency extension of processed signals or images.
5.2
ONESIDED CLIFFORD FOURIER TRANSFORMS
We now generalize the CFT for functions valued in Cl(3, 0) of Section 5.1 to higher dimensions. 5.2.1 Onesided Clifford Fourier transforms in Cl(n, 0), n = 2 (mod 4) and n =
3 (mod 4) 5.2.1.1 Preliminaries
We define a generalized real Fourier transform on Clifford multivectorvalued functions (f : Rn → Cl(n, 0), n = 2, 3 (mod 4)). Next, we show a set of important properties of the Clifford Fourier transform on Cl(n, 0), n = 2, 3 (mod 4) such as differentiation properties, and the Plancherel theorem, independent of special commutation properties. Then, we develop and utilize commutation properties for giving explicit formulas for f xm , f ∇m and for the Clifford convolution. Finally, we apply Clifford Fourier transform properties for proving an uncertainty principle for Cl(n, 0), n = 2, 3 (mod 4) multivector functions. The treatment in this section is based on [185]. Historically, Brackx et al. [33] extended the Fourier transform to multivector valued functiondistributions in Cl0,n with compact support. They also showed some properties of this generalized Fourier transform. A related applied approach for hypercomplex Clifford Fourier transformations in Cl0,n was followed by B¨ ulow et. al. [58]. In [243], Li et. al. extended the Fourier transform holomorphically to a function of m complex variables. For more background and context, we refer to Chapter 1.
250 Quaternion and Clifford Fourier Transforms
In this section, we adopt and expand6 to Gn , n = 2, 3 (mod 4) the generalization of the Fourier transform in Clifford geometric algebra G3 previously suggested by Ebling and Scheuermann [117], based on [132]. To avoid ambiguities we recall that n = 2 (mod 4) ⇔ n = 2 + 4k, k ∈ N, n = 3 (mod 4) ⇔ n = 3 + 4l, l ∈ N.
(5.2.1)
We explicitly show detailed properties of the real7 Clifford geometric algebra Fourier transform (CFT). As a first application we subsequently use some of these properties to define and prove the uncertainty principle for Gn multivector functions. 5.2.1.1.1
Clifford’s geometric algebra Gn = Cl(n, 0) of Rn
Readers not familiar with
Clifford’s geometric algebras are referred to Chapter 2. Let us consider now and in the following an orthonormal vector basis {e1 , e2 , . . ., en } of the real ndimensional Euclidean vector space Rn with n = 2, 3 (mod 4). The geometric algebra over Rn denoted by Gn = Cl(n, 0) then has the graded 2n dimensional basis {1, e1 , e2 , . . . , en , e12 , e31 , e23 , . . . , in = e1 e2 . . . en }.
(5.2.2)
Remark 5.2.1. The fact that we begin by introducing orthonormal bases for both the vector space Rn and for its associated geometric algebra Gn is only because we assume readers to be familiar with these concepts. As is wellknown, the definitions of vector spaces and geometric algebras are generically basis independent [161]. The definition of the vector derivative of Section 5.2.1.1.2 is basis independent, too. Only when we introduce the infinitesimal scalar volume element for integration over Rn in Section 5.2.1.2 and in the proof of the last Theorem 5.2.59 in [251] do we use a basis explicitly. In the latter case it may well be possible to formulate a basis independent proof. All results derived in this section are therefore manifestly invariant (independent of the use of coordinate systems), apart from the proof of Theorem 5.2.59. The squares of vectors are positive definite scalars (Euclidean metric) and so are all even powers of vectors x2 ≥ 0,
xm ≥ 0 for m = 2m0 , m0 ∈ N.
(5.2.3)
Therefore given a multivector M ∈ Gn xm M = M xm , m = 2m0 , m0 ∈ N.
(5.2.4)
6 In the following we mean with n = 2, 3 (mod 4) that n = 2 (mod 4) or n = 3 (mod 4), i.e. with (5.2.1) that n ∈ {2, 3, 6, 7, 10, 11, . . .}. For further details in the case of n = 3 compare Section 5.1. In the geometric algebra literature [161] instead of the mathematical notation Clp,q the notation Gp,q is widely in use. It is convention to abbreviate Gn,0 to Gn . 7 The meaning of real in this context is, that we use the oriented ndimensional unit volume element in of the geometric algebra Gn over the field of the reals R to construct the kernel of the Clifford Fourier transformation of Definition 5.2.21. This in has a clear geometric interpretation, e.g. as ndimensional hypercube with side length one, and in an orthonormal basis of Rn it can be factorized as in = e1 e2 . . . en .
Clifford Fourier transforms 251
Note that for n = 2, 3 (mod 4) i2n = −1,
m
i−1 n = −in ,
2 im n = (−1)
for m = 2m0 , m0 ∈ Z,
(5.2.5)
similar to the complex imaginary unit. fi, we obtain the As a consequence of the square norm of M , kM k2 = hM M multivector CauchySchwarz inequality e i2 ≤ kM k2 kN k2 hM N
∀ M, N ∈ Gn .
(5.2.6)
5.2.1.1.2 Multivector functions, vector differentiation and convolution in the context of onesided CFTs in Cl(n, 0) We only include here a very brief summary of vector dif
ferential calculus. For details and proofs we refer to Chapter 3. This section deals throughout with multivector valued functions f : Rn → Cl(n, 0). Remark 5.2.2. The vector differential a · ∇ is a scalar operator, therefore the left and right vector differentials8 agree, i.e. ˙ f˙(x) = f˙(x) a · ∇. ˙ a·∇
(5.2.7)
The basis independent vector derivative ∇ is defined in [161, 174] to have the algebraic properties of a grade one vector in Rn and to obey equation (3.2.9) for all vectors a ∈ Rn . Exercise 5.2.3. Here we give a set of multivector functions f : R6 → G6 , their vector differentials and vector derivatives [174]. We assume that e1256 = e1 e2 e5 e6 , constant x0 ∈ R6 , r = x − x0 , r = krk, and r−1 = krrk2 . f1 = x, 2
∇f1 = 6,
a · ∇f2 = 2a · x,
f2 = x , f3 = kxk,
a · ∇f1 = a,
a · ∇f3 = a · x/kxk,
f4 = x · e1256 , f5 = log r,
(5.2.8)
∇f2 = 2x, ∇f3 = x/kxk,
a · ∇f4 = a · e1256 , −1
a · ∇f5 = a · r ,
∇f4 = 4e1256 , −1
∇f5 = r .
(5.2.9) (5.2.10) (5.2.11) (5.2.12)
For convenience, we now summarize some properties of the vector derivative, including its application from the left and the right. Proposition 5.2.4 (Left and right linearity). ∇(f + g) = ∇f + ∇g ,
(f + g)∇ = f ∇ + g∇ .
(5.2.13)
Proposition 5.2.5. For f (x) = g(λ(x)), λ(x) ∈ R, a · ∇f = f a · ∇ = {a · ∇λ(x)} 8
∂g . ∂λ
(5.2.14)
The point symbols specify on which function the vector derivative is supposed to act.
252 Quaternion and Clifford Fourier Transforms
Proposition 5.2.6 (Left and right derivative from differential). ∇f = ∇a (a · ∇f ),
f ∇ = (a · ∇f )∇a .
(5.2.15)
Proposition 5.2.7 (Left and right product rules). ˙ f˙)g + ∇f ˙ g˙ = (∇ ˙ f˙)g + ∇a f (a · ∇g) . ∇(f g) = (∇
(5.2.16)
˙ + f˙g ∇ ˙ = f (g˙ ∇) ˙ + (a · ∇f ) g ∇a . (f g)∇ = f (g˙ ∇)
(5.2.17)
Note that the multivector functions f and g in (5.2.16) and (5.2.17) do not necessarily commute. Exercise 5.2.8. For two functions f, g : R2 → G2 , f = x, g = x · e12 we calculate [174] ˙ + (a · ∇f ) g ∇a (f g)∇ = f (g˙ ∇) = x[(x · e12 )∇] + (a · ∇x)(x · e12 )∇a = x(−2e12 ) + a(x · e12 )∇a = −2xe12 + e1 (x · e12 )e1 + e2 (x · e12 )e2 = −2xe12 .
(5.2.18)
Because ∇2 is a scalar operator, the left and right Laplace derivatives agree, i.e. ∇ f = f ∇2 . More generally all even powers of the left and right vector derivative agree ∇m f = f ∇m for m = 2m0 , m0 ∈ N. (5.2.19) 2
Proposition 5.2.9 (Integration of parts). Z
f (x)[a · ∇g(x)] dn x =
Rn
Z
f (x)g(x) dn−1 x
Rn−1
a·x=∞
−
Z
[a · ∇f (x)]g(x) dn x.
(5.2.20)
Rn
a·x=−∞
Remark 5.2.10. Proposition 5.2.9 reduces to the familiar coordinate form, if we insert for a the grade 1 basis vectors ek , 1 ≤ k ≤ n of (5.2.2), because e k · ∇ = ∂k
and
e k · x = xk .
(5.2.21)
We also note that because of (5.2.4) even powers of vectors commute with multivector valued functions f ∈ L2 (Rn , Gn ) xm f = f xm for m = 2m0 , m0 ∈ N.
(5.2.22)
Theorem 5.2.11. For all geometric algebras Gn , n ∈ N we have for f : (Rn → Gn ), a, x, ω ∈ Rn and λ = ±ω · x ± in ω·x a · ∇ω f (x)(± i−1 = a · xf (x)e± in ω·x . n )e
(5.2.23)
For f (x) = 1 we get ± in ω·x a · ∇ω (± i−1 = a · x e± in ω·x . n )e
(5.2.24)
Clifford Fourier transforms 253
Proof. ± in ω·x ± in ω·x a · ∇ω f (x)(± i−1 = f (x)(± i−1 n )e n ) a · ∇ω e
∂ein λ in λ = f (x) i−1 n a · x in e ∂λ in λ = a · xf (x) i−1 = a · xf (x) e± in ω·x. n in e
= f (x)(± i−1 n ) a · ∇ω (±ω · x)
(5.2.25)
For the second equality we used proposition 5.2.5, and for the third equality proposition 21 of [174]. Exercise 5.2.12. Functions f : R6 → G6 , like sin(x) and ei6 ω·x with i6 = ˜ e1 e2 e3 e4 e5 e6 , i−1 6 = i6 = −i6 can be defined by power series. An example for (5.2.23) is therefore a · ∇ω sin(x)(−i6 )ei6 ω·x = a · x sin(x)ei6 ω·x . (5.2.26) By exchanging the roles of x and ω in (5.2.24) we obtain Corollary 5.2.13. ± in ω·x a · ∇(± i−1 = a · ω e± in ω·x . n )e
(5.2.27)
Applying proposition 5.2.6 to corollary 5.2.13 and multiplying both sides with ± in from the right we get Corollary 5.2.14. ∇e± in ω·x = ω (±in ) e± in ω·x ,
(5.2.28)
e± in ω·x ∇ = e± in ω·x (±in ) ω.
(5.2.29)
and its reverse Corollary 5.2.15. Theorem 5.2.16. For all geometric algebras Gn , n ∈ N we have for f : (Rn → Gn ), a, x, ω ∈ Rn , and λ = ±ω · x ± in ω·x ∇ω f (x)(± i−1 = xf (x)e± in ω·x . n )e
(5.2.30)
± in ω·x ∇ω (± i−1 = x e± in ω·x . n )e
(5.2.31)
For f (x) = 1 we get Proof. We first proof (5.2.30). ± in ω·x ± in ω·x ∇ω f (x)(± i−1 = ∇a [a · ∇ω f (x)(± i−1 ] n )e n )e
= ∇a (a · x)f (x) e± in ω·x = xf (x) e± in ω·x .
(5.2.32)
For the first equality we used proposition 5.2.6, for the second equality Theorem 5.2.11, and for the third equality proposition 72 of [174].
254 Quaternion and Clifford Fourier Transforms
Exercise 5.2.17. Similar to example 5.2.12 functions f : R7 → G7 , like cos(x) and ˜ ei7 ω·x with i7 = e1 e2 e3 e4 e5 e6 e7 , i−1 7 = i7 = −i7 can be defined by power series. An example for (5.2.30) is therefore ∇ω cos(x) i7 e−i7 ω·x = x cos(x)ei7 ω·x .
(5.2.33)
The reverse of (5.2.31) gives Corollary 5.2.18. ± in ω·x e± in ω·x (± i−1 x. n ) ∇ω = e
(5.2.34)
Convolution is an important operation for smoothing images and for edge detection in image processing. The Clifford Convolution of multivector valued functions is defined for arbitrary n. Definition 5.2.19 (Clifford Convolution). The Clifford Convolution of f, g ∈ L2 (Rn , Gn ) is defined as Z
(f ? g)(x) =
f (y)g(x − y)dn y.
(5.2.35)
Rn
Exercise 5.2.20. As an example let us compute the convolution of two exponential functions f, g : R2 → G2 , f (x) = e2 exp(−i2 ω · x), g(x) = 3e1 exp(i2 ω 0 · x) with ω, ω 0 ∈ R2 , i2 = e1 e2 , i2 e1,2 = −e1,2 i2 , Z R2
= 3e2 e1 e
i 2 ω 0 ·x
Z
e R2
e2 e−i2 ω·y 3e1 ei2 ω ·(x−y) d2 y 0
(f ? g)(x) =
i2 (ω−ω 0 )·y 2
d y = −3(2π)2 i2 ei2 ω ·x δ(ω − ω 0 ). 0
(5.2.36)
Exchanging the order of the functions we get (g ? f )(x) = 3(2π)2 i2 e−i2 ω ·x δ(ω − ω 0 ), 0
(5.2.37)
illustrating the general noncommutativity (f ? g) 6= (g ? f ) due to the geometric product. Note that the following identity, which follows from the substitution of variables (z = x − y), is valid for all dimensions n. Let f, g ∈ L2 (Rn , Gn ) then Z Rn
n
f (x − y)g(y)d y =
Z
f (z)g(x − z)dn z.
(5.2.38)
Rn
Ebling and Scheuermann [117] distinguish between right and left convolution. They are right that products of multivector valued functions do not commute (compare example 5.2.20), so after e.g. a linear and shiftinvariant (LSI) multivector filter is chosen it matters if a multivector image function is multiplied with the filter from the right or from the left. But because of (5.2.38) we only define one kind of convolution and leave it up to particular applications which factor is taken as multivector filter and which for the multivector image function, etc. The CFT formulas of the convolution which we derive for n = 2, 3 (mod 4) are valid for whatever choice is made in applications.
Clifford Fourier transforms 255
5.2.1.2 Onesided Clifford Fourier transform (CFT)
Definition 5.2.21. The Clifford Fourier transform9 of f (x) is the function F{f }: Rn → Gn , n = 2, 3 (mod 4) given by Z
f (x) e−in ω·x dn x,
(5.2.39)
dn x = dx1 ∧ dx2 ∧ . . . ∧ dxn i−1 n
(5.2.40)
F{f }(ω) =
Rn
with x, ω ∈ Rn . Note that is scalar valued (dxk = dxk ek ∈ Rn , k = 1, 2, . . . , n, no summation). Because for n = 3 (mod 4) in commutes with every element of Gn , the Clifford Fourier kernel e−in ω·x will also commute with every element of Gn . This is not the case for n = 2 (mod 4). Exercise 5.2.22. We give an example for an integrable function f : Rn → Gn , the ndimensional rect(x) function, which can be given in terms of the real scalar rect(x) function with x ∈ R as rect(x) =
n Y
rect(xk )ek = in
n Y
rect(xk ).
(5.2.41)
k=1
k=1
The CFT of rect(x) gives F{rect}(ω) =
Z
rect(x)e−in ω·x dn x
Rn
Z
= in
n Y
Rn k=1
rect(xk )e−in ω·x dn x = in
n Y k=1
sinc(
ωk ). 2π
(5.2.42)
Theorem 5.2.23.R The Clifford Fourier transform F{f } of f ∈ L2 (Rn , Gn ), n = 2, 3 (mod 4), Rn kf k2 dn x < ∞ is invertible and its inverse is calculated by F
−1
1 [F{f }](x) = f (x) = (2π)n
Z
F{f }(ω) ein ω·x dn ω.
(5.2.43)
Rn
For a full proof of Theorem 5.2.23 in dimension n = 3 that can be generalized straight forwardly to dimensions n = 2, 3 (mod 4) see [251]. Though Definition 5.2.21 and Theorem 5.2.23 are the same for the dimensions n = 2 (mod 4) and n = 3 (mod 4), care has to be taken of the general noncommutativity of in for n = 2 (mod 4). However it turns out, that many properties of the CFT for n = 2, 3 (mod 4) can be expressed independent of the commutation properties of in , if sufficient care is taken to avoid commuting in with other multivectors (except scalars and powers of in itself). Exceptions are the CFT of the Clifford convolution, f xm and f ∇m , which need to be studied dimension dependent. 9 Compare e.g. [221], article 160 for the precise conditions on the existence of Fourier integrals. For the Fourier transform on L1 and L2 , see Footnote 2 on page 144.
256 Quaternion and Clifford Fourier Transforms
We therefore continue with a general section on investigating properties of the CFT for n = 2 (mod 4) and n = 3 (mod 4) aiming at expressions that do not depend on the commutation properties of in . This will be followed by a section on the properties of the CFT of the Clifford convolution, f xm and f ∇m . This second section will also include one table each for n = 2 (mod 4) and n = 3 (mod 4) that summarize all the properties of the CFT studied in this text and fully utilize the commutation properties of in . 5.2.1.2.1 Properties of the CFT for n = 2, 3 (mod 4) expressed independent of in commutations The properties of the CFT we will treat now are linearity, scaling, delay,
shift, transformations of powers of the vector differential, of left and right powers of the vector derivative, of the vector variable x ∈ Rn , and finally the Plancherel and Parseval theorems. The unique feature of our study is the independence of theorem formulations and proofs on the commutation properties of the pseudoscalars in in dimensions n = 2, 3 (mod 4). If not otherwise stated, n is assumed to be n = 2, 3 (mod 4) in the remainder of this section. Theorem 5.2.24 (Left linearity). For f (x) = αf1 (x)+βf2 (x) with constants α, β ∈ Gn , and functions f1 (x), f2 (x) ∈ Gn we have F{f }(ω) = αF{f1 }(ω) + βF{f2 }(ω).
(5.2.44)
Proof. Follows from the linearity of the geometric product and the integration involved in the Definition 5.2.21 of the CFT. Remark 5.2.25. Restricting the constants in Theorem 5.2.24 to α, β ∈ R we get both left and right linearity of the CFT. Theorem 5.2.26 (Scaling). Let a ∈ R, a 6= 0 be a scalar constant, then the Clifford Fourier transform of the function fa (x) = f (ax) becomes F{fa }(ω) =
1 ω F{f }( ). an a
(5.2.45)
Proof. Follows from variable substitution u = ax. Theorem 5.2.27 (Shift in space domain, delay). If the argument of f (x) ∈ Gn is offset by a constant vector a ∈ Rn , i.e. fd (x) = f (x − a), then F{fd }(ω) = F{f }(ω)e−in ω·a .
(5.2.46)
Proof. Definition 5.2.21 gives F{fd }(ω) =
Z
f (x − a)e−in ω·x dn x.
Rn
We substitute t for x − a in the above expression, and get with dn x = dn t F{fd }(ω) = This proves (5.2.46).
Z Rn
f (t) e−in ω·t e−in ω·a dn t = F{f }(ω)e−in ω·a .
(5.2.47)
Clifford Fourier transforms 257
Exercise 5.2.28. Using example 5.2.22 we can calculate the CFT of a shifted ndimensional rect(x) function with center at a = 3e2 , ω2 = ω · e2 as Z
F{rectd }(ω) =
Rn
n Y
rect(x − 3e2 )e−in ω·x dn x = in e−3in ω 2
sinc(
k=1
ωk ). 2π
(5.2.48)
Theorem 5.2.29 (Shift in frequency domain). If ω 0 ∈ Rn and f0 (x) = f (x) ein ω 0 ·x , then F{f0 }(ω) = F{f }(ω − ω 0 ) (5.2.49) Proof. Using Definition 5.2.21 and simplifying it we obtain F{f0 }(ω) =
Z Rn
f (x)e−in (ω−ω0 )·x dn x = F{f }(ω − ω 0 ).
(5.2.50)
The CFT F{f }(ω − ω 0 ) is centered on the point ω = ω 0 in the frequency domain. Theorem 5.2.30 (Powers of x ∈ Rn from left). m F{xm f (x)}(ω) = ∇m ω F{f }(ω) in ,
m ∈ N.
(5.2.51)
Proof. We first proof Theorem 5.2.30 for m = 1. Direct calculation leads to F{xf (x)}(ω) =
Z
xf (x) e−in ω·x dn x =
Rn
= ∇ω
Z Rn
Z Rn
∇ω f (x) in e−in ω·x dn x
f (x) e−in ω·x dn x in = ∇ω F{f }(ω) in ,
(5.2.52)
where we have used Definition 5.2.21 and (5.2.30) of Theorem 5.2.16. We therefore have F{xf (x)}(ω) = ∇ω F{f }(ω) in . (5.2.53) Repeating this process m − 1 times we get m F{xm f (x)}(ω) = ∇m ω F{f }(ω) in ,
m ∈ N.
(5.2.54)
Exercise 5.2.31. The CFT of a Gaussian function f (x) = exp(−x2 ), x ∈ Rn is again a Gaussian function F{f }(ω) =
Z
n
2
e−x e−in ω·x dn x = π 2 e−
ω2 4
(5.2.55)
Rn
The CFT of its first moment is therefore according to Theorem 5.2.30 and propositions 5.2.5 and 5.2.6 F{xf }(ω) =
Z
xe Rn
−x2 −in ω·x n
e
n 2
2
− ω4
d x = π ∇ω e
n
ω2 π2 in = − ω in e− 4 . 2
(5.2.56)
258 Quaternion and Clifford Fourier Transforms
Theorem 5.2.32 (xm from right). m
F{f (x) x }(ω) =
Z Rn
−in ω·x n f (x)∇m d x im ωe n,
m ∈ N.
(5.2.57)
Proof. Direct calculation leads to F{f (x) x}(ω) =
Z
f (x) x e−in ω·x dn x
Rn
Z
= Rn
f (x) ∇ω in e
−in ω·x
n
Z
d x= Rn
f (x)∇ω e−in ω·x dn x in
(5.2.58)
where we have used Definition 5.2.21 and (5.2.31) of Theorem 5.2.16. Replacing f m − 1 times by f x and converting the additional right factor x each time into a derivative of e−in ω·x leads to the full proof of the theorem. Remark 5.2.33. In the next section, we will use Theorem 5.2.32 and the dimension dependent commutation properties of in to derive final formulas for the CFTs of f (x) xm , m ∈ N. Theorem 5.2.34. F{(a · x)m f (x)}(ω) = (a · ∇ω )m F{f }(ω) im n,
m ∈ N.
(5.2.59)
Proof. We first proof Theorem 5.2.34 for m = 1. F{a · x f (x)}(ω)
Z
=
a · x f (x) e−in ω·x dn x
Rn
Z
=
f (x) a · x e−in ω·x dn x
Rn Theor. 5.2.11
Z
=
Rn
= =
f (x) a · ∇ω in e−in ω·x dn x
a · ∇ω
Z Rn
f (x) e−in ω·x dn x in
a · ∇ω F{f }(ω) in .
(5.2.60)
Repeatedly inserting a · xf for f in (5.2.60) we obtain Theorem 5.2.34 for every m ∈ N. Inserting b · xf with b ∈ Rn for f in (5.2.60) we obtain the following corollary. Corollary 5.2.35. F{a · x b · xf (x)}(ω) = − a · ∇ω b · ∇ω F{f }(ω).
(5.2.61)
Theorem 5.2.36 (Vector differential). The Clifford Fourier transform of the mth power vector differential of f (x) is F{(a · ∇)m f }(ω) = (a · ω)m F{f }(ω) im n,
m ∈ N.
(5.2.62)
Clifford Fourier transforms 259
Proof. We first proof Theorem 5.2.36 for m = 1. 1 a·∇ F{f }(ω) ein ω·x dn ω (2π)n Rn Z 1 F{f }(ω) a · ∇ ein ω·x dn ω (2π)n Rn Z 1 F{f }(ω) a · ωin ein ω·x dn ω (2π)n Rn F −1 [a · ω F{f } in ](x). Z
a · ∇f (x)
= = Cor. 5.2.13
=
=
(5.2.63)
Application of the inverse CFT Theorem 5.2.23 proves Theorem 5.2.36 for m = 1 F{a · ∇f (x)} = a · ω F{f } in .
(5.2.64)
By repeatedly replacing f with a · ∇f in (5.2.64) we obtain Theorem 5.2.36 for all m ∈ N. Theorem 5.2.37 (Left vector derivative). The Clifford Fourier transform of the mth power vector derivative of f (x) is F{∇m f }(ω) = ω m F{f }(ω) im n,
m ∈ N.
(5.2.65)
Proof. We first proof Theorem 5.2.37 for m = 1. According to proposition 5.2.6 we can calculate the derivative from the differential of Theorem 5.2.36 F{∇f (x)} = F{∇a [a · ∇f (x)]} = ∇a F{a · ∇f (x)} = ∇a (a · ω) F{f } in = ω F{f } in
(5.2.66)
By repeatedly replacing f with ∇f in (5.2.66) we obtain Theorem 5.2.37 for all m ∈ N. Exercise 5.2.38. Using the CFT of a Gaussian function f (x) = exp(−x2 ) of (5.2.55) we can calculate the CFT of its third vector derivative with (5.2.65) as F{∇3 f }(ω) =
Z Rn
n
2
∇3 e−x e−in ω·x dn x = −π 2 ω 3 in e−
ω2 4
.
(5.2.67)
We now prove a theorem, which we will use in the next section together with the dimension dependent commutation properties of in to derive final formulas for the CFT of powers of the right vector derivative f (x) ∇m , m ∈ N. Theorem 5.2.39 (∇m from right). f (x)∇m =
1 (2π)n
Z Rn
in ω·x m n F{f }(ω) im ω d ω. n e
(5.2.68)
260 Quaternion and Clifford Fourier Transforms
Proof. We first proof Theorem 5.2.39 for m = 1. 1 F{f }(ω) ein ω·x dn ω∇ (2π)3 Rn Z 1 = F{f }(ω) ein ω·x ∇ dn ω (2π)3 Rn Z 1 F{f }(ω) ein ω·x in ω dn ω = (2π)3 Rn Z 1 F{f }(ω) in ein ω·x ω dn ω, = (2π)3 Rn Z
f (x)∇ =
(5.2.69)
where we used Corollary 5.2.15 for the third equality. Repeating the application of the vector derivative ∇ from the right m − 1 times to both sides of (5.2.69) completes the proof. Next we will prove a Plancherel theorem and deduce a scalar Parseval theorem, which we need in the last section on the uncertainty principles. Theorem 5.2.40 (Plancherel). Assume that f1 (x), f2 (x) ∈ Gn with Clifford Fourier transform F{f1 }(ω) and F{f2 }(ω) respectively, then Z Rn
n f1 (x) f^ 2 (x) d x =
1 (2π)n
Z Rn
n ^ F{f1 }(ω)F{f 2 }(ω) d x.
(5.2.70)
Proof. Direct calculation yields Z Rn
n f1 (x) f] 2 (x) d x
1 n = [ F{f1 }(ω) ein ω·x dn ω] f] 2 (x)d x (2π)n Rn Rn Z ∼ Z 1 −in ω·x n = F{f1 }(ω) f2 (x) e d x dn ω (2π)n Rn Rn Z 1 n ^ F{f1 }(ω)F{f = 2 }(ω)d ω. (2π)n Rn Z
Z
(5.2.71)
Note that Theorem 5.2.40 is multivector valued. It holds for each grade k, 0 ≤ k ≤ n of the multivectors on both sides of equation (5.2.70). We therefore have Corollary 5.2.41. h
Z Rn
n f1 (x) f^ 2 (x) d xik =
1 h (2π)n
Z Rn
n ^ F{f1 }(ω)F{f 2 }(ω) d xik .
(5.2.72)
Note further, that with f1 (x) = f2 (x) = f (x), we get the following multivector version of the Parseval theorem, i.e. Theorem 5.2.42 (Multivector Parseval). Z Rn
f (x) fg (x) dn x =
1 (2π)n
Z Rn
^ F{f }(ω) F{f }(ω) dn x.
(5.2.73)
Clifford Fourier transforms 261
Exercise 5.2.43. According to (5.2.55) and left linearity of Theorem 5.2.24 the CFT of the function f (x) = (1 + e1 ) exp(−x2 ), x ∈ R2 is F{f }(ω) =
Z R2
ω2 4
(5.2.74)
e−2x d2 x = (1 + e1 )π.
(5.2.75)
2
(1 + e1 )e−x e−i2 ω·x d2 x = (1 + e1 )πe−
Inserting this f into (5.2.73) gives on the left side Z R2
f (x) fg (x) d2 x = (1 + e1 )2
Z R2
2
We can check (5.2.73) by inserting (5.2.74) on the right side. The scalar part of the result is π, the vector part is πe1 . The scalar part of Theorem 5.2.42 together with (5.2.147), gives us the scalar Parseval theorem Theorem 5.2.44 (Scalar Parseval). Z Rn
kf (x)k2 dn x =
1 (2π)n
Z
kF{f }(ω)k2 dn ω.
(5.2.76)
Rn
Properties of the CFT for n = 2, 3 (mod 4) dependent on in commutations Now we concentrate on properties of the CFT, which need to make use of the commutation properties of the unit oriented pseudoscalar in ∈ Gn in order to be fully developed. In this category, we have for m ∈ N the CFTs of f xm , f ∇m and the CFT of the Clifford convolution with distinct expressions for the dimensions of n = 2 (mod 4) and n = 3 (mod 4). Before we proceed, we note that for n = 3 (mod 4) the pseudoscalar in commutes with all elements of the algebra. For the case of n = 2 (mod 4) we first establish a theorem and two corollaries.
5.2.1.2.2
Theorem 5.2.45. Any odd grade multivector Ar ∈ Gn , r = 2s + 1, s ∈ N, s < anticommutes with in for n = 2 (mod 4) Ar in = −in Ar .
n 2
(5.2.77)
Any even grade multivector Ar ∈ Gn , r = 2s, s ∈ N, s ≤ n/2 commutes with in for n = 2 (mod 4) Ar in = +in Ar . (5.2.78) Proof. For the case of n = 2 we have [58] for a vector a ∈ R2 i2 a = −ai2 .
(5.2.79)
For the general case of n = 2 (mod 4) we can factorize in for any vector a ∈ Rn such that a 2 in = in−2 ˆ ba ˆ, ˆ b∗a ˆ=ˆ bc in−2 = a ˆ c in−2 = 0, a ˆ= , ˆ b = 1. a
(5.2.80)
262 Quaternion and Clifford Fourier Transforms
in−2 will therefore be of grade 0 (mod 4), represent a subspace10 of Rn perpendicular to a and therefore commute with a. According to (5.2.79) and (5.2.80) the twoblade ˆ ba ˆ anticommutes with a and hence in a = in−2 ˆ ba ˆ a = −in−2 a ˆ ba ˆ = −ain−2 ˆ ba ˆ = −ain .
(5.2.81)
Any odd grade multivector Aodd can be written as a sum over homogeneous odd grade parts. These parts can in turn be decomposed into sums of odd grade blades, which can be factorized into products of an odd number of vectors [58, 179]. Since a single vector anti commutes with in , a geometric product of an odd number of vectors will also anticommute with in and hence by linearity any odd grade multivector will anticommute with in . This proves (5.2.77). Similarly any even grade multivector Aeven can be written as a sum over homogeneous even grade parts. These parts can in turn be decomposed into sums of even grade blades, which can be factorized into geometric products of an even number of vectors [58, 179]. The even number of commutations with an even number of vector factors leads via linearity to the total commutation relationship (5.2.78). Based on Theorem 5.2.45 we derive two useful corollaries. Corollary 5.2.46. For n = 2 (mod 4) and a ∈ R we have for even m ∈ N
and for odd m ∈ N
(a in )m = am
(5.2.82)
(a in )m = am in
(5.2.83)
(a in )2 = a in a in = aa(−in in ) = a2 ,
(5.2.84)
Proof. We have where we used (5.2.77) for the second equality. Using (5.2.84) m/2 times [(m − 1)/2 times] we arrive at equations (5.2.82) [and (5.2.83)]. Corollary 5.2.47. Let the odd grade part of a general multivector A ∈ Gn be Aodd = hAiodd and the even grade part be Aeven = hAieven . Then for n = 2 (mod 4) we have
and for λ ∈ R
A in = Aodd in + Aeven in = −in Aodd + in Aeven ,
(5.2.85)
Aein λ = e−in λ Aodd + e+in λ Aeven ,
(5.2.86)
ein λ A = Aodd e−in λ + Aeven e+in λ .
(5.2.87)
and Proof. Corollary 5.2.47 follows from Theorem 5.2.45 and the fact that ein λ is a power series of in .
10
A subspace in the sense of outer product null space.
Clifford Fourier transforms 263
Theorem 5.2.48 (Powers of x ∈ Rn from right). For n = 2 (mod 4) we have m F{f (x) xm }(ω) = F{f }((−1)m ω) ∇m ω in ,
m ∈ N.
(5.2.88)
For n = 3 (mod 4) we have m F{f (x) xm }(ω) = im n F{f }(ω) ∇ ω ,
m ∈ N.
(5.2.89)
Proof. We first proof Theorem 5.2.48 for n = 2 (mod 4). We start with Theorem 5.2.32 and apply corollary 5.2.47 to commute the vector derivative ∇ω to the right of the integral m
F{f (x) x }(ω) =
Z Rn
Z
f (x)e−in (−1)
= Rn
m ω·x
−in ω·x n f (x)∇m d x im ω e n
m m m m dn x ∇m ω in = F{f }((−1) ω) ∇ω in .
(5.2.90)
The proof of (5.2.89) with n = 3 (mod 4) is the same, except that the sign of ω in the exponent does not change and that we can freely commute in to the left. Theorem 5.2.49 (Right vector derivative). The Clifford Fourier transform of the mth power vector derivative (applied from the right) of f (x) is for n = 2 (mod 4) F{f ∇m }(ω) = F{f }((−1)m ω) ω m im n,
m ∈ N,
(5.2.91)
and for n = 3 (mod 4) m F{f ∇m }(ω) = im n F{f }(ω) ω ,
m ∈ N.
(5.2.92)
Proof. We first proof Theorem 5.2.49 for n = 2 (mod 4). We start with Theorem 5.2.39 and apply corollary 5.2.47 to commute the vector ω m with ein ω·x 1 in ω·x m n F{f }(ω) im ω d ω n e (2π)n Rn Z 1 m in (−1)m ω·x n F{f }(ω) im d ω = n ω e (2π)n Rn Z 1 m in ω·x n = F{f }((−1)m ω) im d ω n (−ω) e (2π)n Rn = F −1 [F{f }((−1)m ω) ω m im n ](x),
f (x)∇m =
Z
(5.2.93)
where for odd m we substituted −ω → ω for the third equality. For the fourth equality we applied apply corollary 5.2.47 once more to commute ω m and im n . Equation (5.2.91) is obtained by applying the inverse CFT Theorem 5.2.23 to both sides of (5.2.93). Once again the proof of (5.2.92) with n = 3 (mod 4) is the same, except that the sign of ω in the exponent does not change and that we can freely commute in to the left.
264 Quaternion and Clifford Fourier Transforms
Remark 5.2.50. Theorem 5.2.37 and (5.2.91) show that all signs of the right hand sides of all five lines in the derivative Theorem 5.7 in [117] are wrong. Line three there contains further errors. For even m we get from Theorems 5.2.30 and 5.2.48, and from (5.2.5) Corollary 5.2.51. m
F{xm f (x)}(ω) = F{f (x) xm }(ω) = (−1) 2 ∇m ω F{f }(ω).
(5.2.94)
We further get for even m from Theorems 5.2.37 and 5.2.49, and from (5.2.5) Corollary 5.2.52. m
F{∇m f (x)} = F{f (x) ∇m } = (−1) 2 ω m F{f }(ω).
(5.2.95)
Theorem 5.2.53 (CFT of Clifford Convolution). For n = 2 (mod 4), f, g ∈ L2 (Rn , Gn ), and godd (geven ) the odd (even) grade part of g we have F{f ?g}(ω) = F{f }(−ω) F{godd }(ω) + F{f }(ω) F{geven }(ω).
(5.2.96)
For n = 3 (mod 4) we have F{f ?g}(ω) = F{f }(ω) F{g}(ω).
(5.2.97)
Proof. For n = 2 (mod 4) we have F{f ?g}(ω) =
Z
Z
[ Rn
f (y)g(x − y)dn y]e−in ω·x dn x
Rn
Z
Z
= Rn
Rn
Z
Z
=
g(z)e−in ω·(y+z) dn z]dn y
f (y)[ Rn
Rn
Z
Z
=
f (y) Rn
Z
=
g(x − y)e−in ω·x dn x]dn y
f (y)[
Rn
[e+in ω·y godd (z) + e−in ω·y geven (z)]e−in ω·z dn zdn y
f (y)e−in (−ω)·y dn yF{godd }(ω) +
Rn
Z
f (y)e−in ω·y dn yF{geven }(ω)
Rn
= F{f }(−ω) F{godd }(ω) + F{f }(ω) F{geven }(ω).
(5.2.98)
For the third equality we used the variable substitution z = x − y and for the fourth equality we used Corollary 5.2.47. The corresponding proof in [252] for n = 3, can be generalized straight forwardly to n = 3 (mod 4). This part of the proof is also strictly invariant. Remark 5.2.54. The above proof of Theorem 5.2.53 for n = 2 (mod 4) depends on the in commutation relationships. But on the other hand, it has the advantage of being manifestly invariant, since no coordinate system needed to be introduced.
Clifford Fourier transforms 265
Properties of the Clifford Fourier transform (CFT) with n = 3 (mod 4). Multivector functions (Multiv. Funct.) f, g, f1 , f2 ∈ L2 (Rn , Cl(n, 0)), constants α, β ∈ Cl(n, 0), 0 6= a ∈ R, a, ω 0 ∈ Rn and m ∈ N. Source: [185, Tab. 1.1]. Table 5.3
Property Left lin. xShift ωShift Scaling Convolution Vec. diff. Powers of x Vec. deriv. Plancherel sc. Parseval
Multiv. Funct. αf (x)+β g(x) f (x − a) ein ω 0 ·x f (x) f (ax), a ∈ R \ {0} (f ?g)(x) (a · ∇)m f (x) (a · x)m f (x) xm f (x) f (x)xm ∇m f (x) f (x)∇m R ] n n f1 (x)f2 (x) d x RR
Rn
kf (x)k2 dn x
CFT αF{f }(ω)+ βF{g}(ω) e−in ω ·a F{f }(ω) F{f }(ω − ω 0 ) 1 ω an F{f }( a ) F{f }(ω) F{g}(ω) m im n (a · ω) F{f }(ω) m in (a · ∇ω )m F{f }(ω) m im n ∇ω F{f }(ω) m im n F{f }(ω) ∇ω m m in ω F{f }(ω) m im n F{f }(ω) ω R 1 n ^ n F{f1 }(ω)F{f2 }(ω) d ω n (2π) 1 (2π)n
RR
Rn
kF{f }(ω)k2 dn ω
Theorem 5.2.53 correctly generalizes the results for n = 2 in [117] to n = 2 (mod 4). Comparing Theorem 5.2.53 with the convolution Theorem 5.6 of [117] in two dimensions, we see that the fourth line of convolution Theorem 5.6 in [117] must be wrong. On the right hand side of this formula the dot over the vector filter ˙ function h under the CFT indicating h(x) = h(−x) is incorrect. Because of (5.2.38) h should also have no dot, in agreement with the correct dotfree vector filter function f on the right hand side of line two of Theorem 5.6 in [117]. In order to give an overview of the properties of the CFT we list its properties for n = 3 (mod 4) in Table 5.3 and for n = 2 (mod 4) in Table 5.4. Comparing the tables, the differences caused by the different commutation rules for the pseudoscalars in in n = 3 (mod 4) and n = 2 (mod 4) dimensions are obvious. In Table 5.3 the positions of in and of exponentials ein λ , λ ∈ R are arbitrary. In Table 5.4 the pseudoscalar in and its exponentials ein λ , λ ∈ R cannot be freely commuted. 5.2.1.3 Uncertainty principle
The uncertainty principle plays an important role in the development and understanding of quantum physics. It is also central for information processing [287]. In quantum physics it states e.g. that conjugate properties like particle momentum and position cannot be be simultaneously measured accurately. In Fourier analysis such conjugate entities correspond to a function and its Fourier transform which cannot both be simultaneously sharply localized. Furthermore much work (e.g. [74,287]) has been devoted to extending the uncertainty principle to a function and its Fourier transform. Yet Felsberg [132] notes even for two dimensions: In two dimensions
266 Quaternion and Clifford Fourier Transforms
Properties of the Clifford Fourier transform (CFT) with n = 2 (mod 4). Multivector functions (Multiv. Funct.) f, g, f1 , f2 ∈ L2 (Rn , Cl(n, 0)), constants α, β ∈ Cl(n, 0), 0 6= a ∈ R, a, ω 0 ∈ Rn and m ∈ N. Source: [185, Tab. 1.2].
Table 5.4
Property Left lin. xShift ωShift Scaling Convolution
Multiv. Funct. αf (x)+β g(x) f (x − a) f (x) ein ω 0 ·x f (ax), a ∈ R \ {0} (f ?g)(x)
Vec. diff.
(a · ∇)m f (x) (a · x)m f (x) xm f (x) f (x) xm ∇m f (x) f (x)∇m R ] n n f1 (x)f2 (x) d x
Powers of x Vec. deriv. Plancherel sc. Parseval
RR
Rn
CFT αF{f }(ω)+ βF{g}(ω) F{f }(ω) e−in ω·a F{f }(ω − ω 0 ) 1 ω an F{f }( a ) F{f }(−ω) F{godd }(ω) +F{f }(ω) F{geven }(ω) (a · ω)m F{f }(ω) im n (a · ∇ω )m F{f }(ω) im n m ∇m F{f }(ω) i ω n m F{f }((−1)m ω) ∇m ω in ω m F{f }(ω) im n F{f }((−1)m ω) ω m im n R 1 n ^ F{f }(ω) F{f n n 1 2 }(ω) d ω (2π) 1 (2π)n
kf (x)k2 dn x
RR
Rn
kF{f }(ω)k2 dn ω
however, the uncertainty relation is still an open problem. In [145] it is stated that there is no straightforward formulation for the twodimensional uncertainty relation. From the view point of geometric algebra an uncertainty principle gives us information about how the variations of a multivector valued function and its Clifford Fourier transform are related. The theorems and the corollary below have all been proven with great detail for the case of n = 3 in [251]. The key steps of the proofs there involve the CFT of the vector differential of Theorem 5.2.36 and the (scalar) Parseval Theorem 5.2.44, the Cauchy Schwarz inequality (5.2.6) for multivectors, and finally the coordinate free integration of parts formula of Proposition 5.2.9. Otherwise the proofs are very analogous, and do not involve dimension dependent in commutations. Therefore we don’t repeat them here, we only list the resulting formulas. Theorem 5.2.55 (Directional uncertainty principle). Let f be a multivector valued function Rin Gn , n = 2, 3 (mod 4), which has the Clifford Fourier transform F{f }(ω). Assume Rn kf (x)k2 dn x = F < ∞, then the following inequality holds for arbitrary constant vectors a, b: Z
(a · x)2 kf (x)k2 dn x
Rn
≥ (a · b)2
1 2 F . 4
1 (2π)n
Z
(b · ω)2 kF{f }(ω)k2 dn ω
Rn
(5.2.99)
Clifford Fourier transforms 267
Proof. Applying the results stated so far we have11 1 (a · x) kf (x)k d x n (2π)n R
Z
2
n
2
tab. 5.3,5.4, line 6
(b · ω)2 kF{f }(ω)k2 dn ω
Rn
1 (a · x) kf (x)k d x (2π)n Rn
Z
Z
kb · ∇f (x)k2 dn x
Z
=
Z
sc. Parseval
2
n
2
2
2
n
(a · x) kf (x)k d x
=
Z
Rn footnote 11
Z
(5.2.6)
Z
kF{b · ∇f }(ω)k2 dn ω
Rn
Rn
a · x kf (x)k kb · ∇f (x)k dn x
≥
2
Rn
a · xhfg (x) b · ∇f (x)i dn x
≥
2
Rn
Z
a · xhfg (x) b · ∇f (x)i dn x
≥
2
(5.2.100)
.
Rn
Because of (5.2.147) and (5.2.148) (b · ∇)kf k2 = 2hfe (b · ∇)f i,
(5.2.101)
we furthermore obtain Z
1 (b · ω)2 kF{f }(ω)k2 dn ω (2π)n Rn 2 1 a · x (b · ∇kf k2 ) dn x 2 Z
(a · x)2 kf (x)k2 dn x
Rn
Z
≥ Rn
1 4
Prop. 5.2.9
=
Z
2 n−1
a · xkf (x)k d
b·x=∞
x
Rn−1
b·x=−∞
−
Z
2
n
2
[(b · ∇)(a · x)] kf (x)k d x
Rn
2
1 0−a·b kf (x)k2 dn x) n 4 R 2 1 2 (a · b) F . 4
= =
Z
(5.2.102)
Choosing b = ±a, i.e. b k a, with a2 = 1 we get from Theorem 5.2.55 the following Corollary 5.2.56 (Uncertainty principle). 1 (a · x) kf (x)k d x (2π)n Rn
Z
11
2
φ, ψ : Rn → C,
2
R Rn
n
φ(x)2 dn x
Z
R Rn
(a · ω)2 kF{f }(ω)k2 dn ω ≥
Rn
ψ(x)2 dn x ≥ (
R Rn
φ(x)ψ(x) dn x)2
1 2 F . 4
(5.2.103)
268 Quaternion and Clifford Fourier Transforms
Remark 5.2.57. In (5.2.103) equality holds for Gaussian multivector valued functions 2 f (x) = C0 e−k x , (5.2.104) where C0 ∈ Gn is an arbitrary but constant multivector, 0 < k ∈ R. This follows from the observation that we have for the f of (5.2.104) −2k a · x f = a · ∇f.
(5.2.105)
Theorem 5.2.58. For a · b = 0, i.e. b ⊥ a, we get 1 (a · x) kf (x)k d x n (2π)n R
Z
2
2
n
Z
(b · ω)2 kF{f }(ω)k2 dn ω ≥ 0.
(5.2.106)
Rn
Theorem 5.2.59. Under the same assumptions as in Theorem 5.2.55, we obtain Z
x2 kf (x)k2 dn x
Rn
1 (2π)n
1 ω 2 kF{f }(ω)k2 dn ω ≥ n F 2 . 4 Rn
Z
(5.2.107)
Remark 5.2.60. To prove Theorem 5.2.59 we first insert x2 = nk=1 (ek · x)2 , P ω 2 = nl=1 (el · ω)2 . After that we apply (5.2.103) and (5.2.106) depending on the relative directions of the vectors ek and el . P
5.2.1.4 Summary on CFTs in Cl(n, 0), n = 2 (mod 4) and n = 3 (mod 4)
The (real) Clifford Fourier transform extends the traditional Fourier transform on scalar functions to Gn = Cl(n, 0) multivector functions with n = 2, 3 (mod 4) over the vector space domain Rn . Basic properties and rules for differentiation, convolution, the Plancherel and Parseval theorems have been demonstrated in a manifestly invariant fashion. We then presented an uncertainty principle in the geometric algebra Gn , which describes how a multivectorvalued function and its Clifford Fourier transform relate. In many fields the Fourier transform is successfully applied to solving physical equations such as heat and wave equations, in optics, in signal and image processing, etc. Therefore, we can apply geometric algebra and the Clifford Fourier transform to solve such problems involving the whole range of kvector fields (k = 0, 1, 2, . . . , n) in geometric algebras Gn = Cl(n, 0) with n = 2, 3 (mod 4). The calculations will be real, have clear geometric interpretations and manifestly invariant. The use of coordinate bases is optional. 5.2.2 Onesided Clifford Fourier transforms in Cl(p, q)
We now generalize the onesided CFTs to multivector valued functions in Cl(p, q). 5.2.2.1 Background
We will now use the comprehensive research [193, 203] on the manifolds of square roots of −1 in real Clifford’s geometric algebras Cl(p, q) in order to construct the Clifford Fourier transform. Basically in the kernel of the complex Fourier transform
Clifford Fourier transforms 269
the imaginary unit j ∈ C is replaced by a square root of −1 in Cl(p, q). The Clifford Fourier transform (CFT) thus obtained generalizes previously known and applied CFTs [117, 132, 185], which replaced j ∈ C only by blades (usually pseudoscalars) squaring to −1. A major advantage of real Clifford algebra CFTs is their completely real geometric interpretation. We study (left and right) linearity of the CFT for constant multivector coefficients ∈ Cl(p, q), translation (xshift) and modulation (ωshift) properties and signal dilations. We show an inversion theorem. We establish the CFT of vector differentials, partial derivatives, vector derivatives and spatial moments of the signal. We also derive Plancherel and Parseval identities as well as a general convolution theorem. The treatment in this section is based on [201]. We note that quaternion, Clifford and geometric algebra Fourier transforms (QFT, CFT, GAFT) [184, 185, 194, 202] have already proven very useful tools for applications in nonmarginal color image processing, image diffusion, electromagnetism, multichannel processing, vector field processing, shape representation, linear scale invariant filtering, fast vector pattern matching, phase correlation, analysis of nonstationary improper complex signals, flow analysis, partial differential systems, disparity estimation, texture segmentation, as spectral representations for Clifford wavelet analysis, etc. All these Fourier transforms essentially analyze scalar, vector and multivector signals in terms of sine and cosine waves with multivector coefficients. For this purpose the imaginary unit j ∈ C in ejφ = cos φ + j sin φ can be replaced by any square root of −1 in a real Clifford algebra Cl(p, q). The replacement by pure quaternions and blades with negative square [61,184] has already yielded a wide variety of results with a clear geometric interpretation. It is wellknown that there are elements other than blades, squaring to −1. Motivated by their special relevance for new types of CFTs, they have by now been studied thoroughly [193, 203, 294]. We therefore tap into these new results on square roots of −1 in Clifford algebras and fully general construct CFTs, with one general square root of −1 in Cl(p, q). Our new CFTs form therefore a more general class of CFTs, subsuming and generalizing previous results12 . A further benefit is, that these new CFTs become fully steerable within the continuous Clifford algebra submanifolds of square roots of −1. We thus obtain a comprehensive new mathematical framework for the investigation and application of Clifford Fourier transforms together with new properties (full steerability). Regarding the question of the most suitable CFT for a certain application, we are only just beginning to leave the terra cognita of familiar transforms to map out the vast array of possible CFTs in Cl(p, q). This part is organized as follows. We first recall in Section 5.2.2.1.1 some key notions of Clifford algebra, multivector signal functions and the recent results on square roots of −1 in Clifford algebras. Next, in Section 5.2.2.2 we define the central notion of Clifford Fourier transforms with respect to any square root of −1 in Clifford algebras Cl(p, q). Then we study in Section 5.2.2.3 (left and right) linearity of the 12
This is only the first step towards generalization. The noncommutativity of the geometric product of multivectors makes it meaningful to investigate CFTs with several kernel factors to both sides of the signal function. Each kernel factor may use a different square root of −1, compare Section 5.3.
270 Quaternion and Clifford Fourier Transforms
CFT for constant multivector coefficients ∈ Cl(p, q), translation (xshift) and modulation (ωshift) properties, and signal dilations, followed by an inversion theorem. We establish the CFT of vector differentials, partial derivatives, vector derivatives and spatial moments of the signal. We also show Plancherel and Parseval identities as well as a general convolution theorem. For an introduction to Clifford algebras we refer to Chapter 2, for vector differential calculus to Chapter 3. The principal reverse13 , compare Section 2.1.1.6, of every basis element eA ∈ Cl(p, q), 1 ≤ A ≤ 2n , has the property
5.2.2.1.1
Principal reverse and modulus in Cl(p, q)
ef A ∗ eB = δAB ,
1 ≤ A, B ≤ 2n ,
(5.2.108)
where the Kronecker delta δAB = 1 if A = B, and δAB = 0 if A = 6 B. For the vector space Rp,q this leads to a reciprocal basis el , 1 ≤ l, k ≤ n ( l
e := eel = εl el ,
l
l
e ∗ ek = e · ek =
1, 0,
for l = k . for l = 6 k
(5.2.109)
e = For M, N ∈ Cl(p, q) we get M ∗ N A MA NA . Two multivectors M, N ∈ e = 0. The modulus M  of a multivector Cl(p, q) are orthogonal if and only if M ∗ N M ∈ Cl(p, q) is defined as P
f= M 2 = M ∗ M
X
MA2 .
(5.2.110)
A
Square roots of −1 in Clifford algebras are extensively studied in Section 2.7 of this book. With respect to any square root i ∈ Cl(p, q) of −1, i2 = −1, every multivector A ∈ Cl(p, q) can be split into commuting A+i and anticommuting A−i parts, see Lemma 2.7.5 and [203]. 5.2.2.1.2
Square roots of −1 in Clifford algebras
5.2.2.2 The onesided Clifford Fourier transform in Cl(p, q)
The general Clifford Fourier transform (CFT), to be introduced now, can be understood as a generalization of known CFTs [185] to a general real Clifford algebra setting. Most previously known CFTs use in their kernels specific square roots of −1, like bivectors, pseudoscalars, unit pure quaternions or blades [61]. For an introduction to known CFTs see [51], and for their various applications see [202]. We will remove all these restrictions on the square root of −1 used in a CFT.14
13
Note that in the current section, we use the principal reverse throughout. But depending on the context another involution or antiinvolution of Clifford algebra may be more appropriate for specific Clifford algebras, or for the purpose of a specific geometric interpretation. 14 For example, the use of the square root i = f1 of Table 2.7 would lead to a new type of CFT, which has so far not been studies anywhere in the literature, apart from [201].
Clifford Fourier transforms 271
Definition 5.2.61 (CFT with respect to one square root of −1). Let i ∈ Cl(p, q), i2 = −1, be any square root of −1. The general Clifford Fourier transform15 (CFT) of f ∈ L1 (Rp,q ; Cl(p, q)), with respect to i is i
F {f }(ω) =
Z
f (x) e−iu(x,ω) dn x,
(5.2.111)
Rp,q
where dn x = dx1 . . . dxn , x, ω ∈ Rp,q , and u : Rp,q × Rp,q → R. Since square roots of −1 in Cl(p, q) populate continuous submanifolds in Cl(p, q), the CFT of Definition 5.2.61 is generically steerable within these manifolds, see (5.2.127). In Definition 5.2.61, the square roots i ∈ Cl(p, q) of −1 may be from any component of any conjugacy class. The choice of the geometric product in the integrand of (5.2.111) is very important. Because only this choice allowed, e.g. in [117], to define and apply a holistic vector field convolution, without loss of information. 5.2.2.3 Properties of the onesided CFT in Cl(p, q)
We now study important properties of the general CFT of Definition 5.2.61. The proofs in this section may seem deceptively similar to standard proofs of properties of the classical complex Fourier transform. But the inherent noncommutativity of the geometric product of multivectors, makes it necessary to carefully respect the order of factors. Already for the first property of left and right linearity in (5.2.112) and (5.2.113), respectively, the order of factors leads to crucial differences. We therefore give detailed proofs of all properties. 5.2.2.3.1 Linearity, shift, modulation, dilation and powers of f, g , steerability Regarding left and right linearity of the general CFT of Definition 5.2.61 we can establish with the help of Lemma 2.7.5 that for h1 , h2 ∈ L1 (Rp,q ; Cl(p, q)), and constants α, β ∈ Cl(p, q)
F i {αh1 + βh2 }(ω) = αF i {h1 }(ω) + βF i {h2 }(ω), i
(5.2.112)
−i
i
F {h1 α + h2 β}(ω) = F {h1 }(ω)α+i + F {h1 }(ω)α−i + F i {h2 }(ω)β+i + F −i {h2 }(ω)β−i .
(5.2.113)
Proof. Based on Lemma 2.7.5 we have α = α+i + α−i ,
α+i i = iα+i ,
α−i i = −iα−i
⇒ αe−iu = (α+i + α−i )e−iu = α+i e−iu + α−i e−iu = e−iu α+i + e−(−i)u α−i ,
(5.2.114)
and similarly β = β+i + β−i , 15
βe−iu = e−iu β+i + e−(−i)u β−i .
For the Fourier transform on L1 and on L2 , see Footnote 2 on page 144.
(5.2.115)
272 Quaternion and Clifford Fourier Transforms
We apply Definition 5.2.61 and get Z
F i {αh1 + βh2 }(ω) =
Rp,q
{αh1 + βh2 } e−iu dn x
= αF i {h1 }(ω) + βF i {h2 }(ω).
(5.2.116)
By inserting (5.2.114) and (5.2.115) into Definition 5.2.61 we can further derive F i {h1 α + h2 β}(ω) = F i {h1 }(ω)α+i + F −i {h1 }(ω)α−i + F i {h2 }(ω)β+i + F −i {h2 }(ω)β−i .
(5.2.117)
For i power factors in ha,b (x) = ia h(x)ib , a, b ∈ Z, we obtain as an application of linearity F i {ha,b }(ω) = ia F i {h}(ω)ib . (5.2.118) Regarding the xshifted function h0 (x) = h(x − x0 ) we obtain with constant x0 ∈ Rp,q , assuming linearity of u(x, ω) in its vector space argument x, F i {h0 }(ω) = F i {h}(ω) e−iu(x0 ,ω) .
(5.2.119)
Proof. We assume linearity of u(x, ω) in its vector space argument x. Inserting h0 (x) = h(x − x0 ) in Definition 5.2.61 we obtain i
F {h0 }(ω) =
Z Rp,q
Z
=
h(x − x0 ) e−iu(x,ω) dn x h(y) e−iu(y+x0 ,ω) dn y
Rp,q
Z
=
h(y) e−iu(y,ω) e−iu(x0 ,ω) dn y
Rp,q
Z
=
h(y) e−iu(y,ω) dn y e−iu(x0 ,ω)
Rp,q
= F i {h}(ω) e−iu(x0 ,ω) ,
(5.2.120)
where we have substituted y = x − x0 for the second equality, we used the linearity of u(x, ω) in its vector space argument x for the third equality, and that e−iu(x0 ,ω) is independent of y for the fourth equality. For the purpose of modulation we make the special assumption, that the function u(x, ω) is linear in its frequency argument ω. Then we obtain for hm (x) = h(x) e−iu(x,ω0 ) , and constant ω 0 ∈ Rp,q the modulation formula F i {hm }(ω) = F i {h}(ω + ω 0 ).
(5.2.121)
Clifford Fourier transforms 273
Proof. We assume, that the function u(x, ω) is linear in its frequency argument ω. Inserting hm (x) = h(x) e−iu(x,ω0 ) in Definition 5.2.61 we obtain F i {hm }(ω) =
Z Rp,q
Z
hm (x) e−iu(x,ω) dn x h(x) e−iu(x,ω0 ) e−iu(x,ω) dn x
= Rp,q
Z
h(x) e−iu(x,ω+ω0 ) dn x
= Rp,q
= F i {h}(ω + ω 0 ),
(5.2.122)
where we used the linearity of u(x, ω) in its frequency argument ω for the third equality. Regarding dilations, we make the special assumption, that for constants a1 , . . . , an Pn k 0 0 0 ∈ R \ {0}, and x0 = k=1 ak x e k , we have u(x , ω) = u(x, ω ), with ω = Pn k 0 k=1 ak ω e k . We then obtain for hd (x) = h(x ) that 1 F i {h}(ω d ), F {hd }(ω) = a1 . . . an  i
ωd =
n X 1
a k=1 k
ω k ek .
(5.2.123)
For a1 = . . . = an = a ∈ R \ {0} this simplifies under the same special assumption to F i {hd }(ω) =
1 i 1 F {h}( ω). n a a
(5.2.124)
e = Note, that the above assumption would, e.g., be fulfilled for u(x, ω) = x ∗ ω Pn Pn k k x ω = x ω . k=1 k=1 k k P Proof. We assume for constants a1 , . . . , an ∈ R \ {0}, and x0 = nk=1 ak xk ek , that P we have u(x0 , ω) = u(x, ω 0 ), with ω 0 = nk=1 ak ω k ek . Inserting hd (x) = h(x0 ) in Definition 5.2.61 we obtain i
F {hd }(ω) =
Z Rp,q
Z
hd (x) e−iu(x,ω) dn x h(x0 ) e−iu(x,ω) dn x
= Rp,q
1 0 h(y) e−iu(y ,ω) dn y a1 . . . an  Rp,q Z 1 = h(y) e−iu(y,ωd ) dn y a1 . . . an  Rp,q 1 = F i {h}(ω d ), a1 . . . an  Z
=
(5.2.125)
n n 1 k k 0 where we substituted y = x0 = k=1 ak x e k and x = k=1 ak y e k = y for the third equality. Note that in this step each negative ak < 0, 1 ≤ k ≤ n, leads to a factor a1k  , because the negative sign is absorbed by interchanging the resulting integration boundaries {+∞, −∞} to {−∞, +∞}. For the fourth equality we applied P the assumption u(y 0 , ω) = u(y, ω 0 ), and defined ω d = ω 0 = nk=1 ak ω k ek .
P
P
274 Quaternion and Clifford Fourier Transforms
Within the same conjugacy class of square roots of −1 the CFTs of Definition 5.2.61 are related by the following equation, and therefore steerable. Let i, i0 ∈ Cl(p, q) be any two square roots of −1 in the same conjugacy class, i.e. i0 = a−1 ia, a ∈ Cl(p, q), a being invertible. As a consequence of this relationship we also have 0
e−i u = a−1 e−iu a, ∀u ∈ R.
(5.2.126)
This in turn leads to the following steerability relationship of all CFTs with square roots of −1 from the same conjugacy class: 0
F i {h}(ω) = F i {ha−1 }(ω)a,
(5.2.127)
where ha−1 means to multiply the signal function h by the constant multivector a−1 ∈ Cl(p, q). For establishing an inversion formula, moment and derivative properties, Plancherel and Parseval identities, certain assumptions about the phase function u(x, ω) need to be made. In principle these assumptions could be made based on the desired properties of the resulting CFT. One possibility is, e.g., to assume 5.2.2.3.2
CFT inversion, moments, derivatives, Plancherel, Parseval
n X
e = u(x, ω) = x ∗ ω
xl ω l =
l=1
n X
xl ωl ,
(5.2.128)
l=1
which will be assumed for the current subsection. We then get the following inversion formula16 i h(x) = F−1 {F i {h}}(x) =
1 (2π)n
Z
F i {h}(ω) eiu(x,ω) dn ω,
(5.2.129)
Rp,q
where dn ω = dω1 . . . dωn , x, ω ∈ Rp,q . For the existence of (5.2.129) we need F i {h} ∈ L1 (Rp,q ; Cl(p, q)). Proof. By direct computation we find 1 F i {h}(ω) eiu(x,ω) dn ω (2π)n Rp,q Z Z 1 h(y) e−iu(y,ω) eiu(x,ω) dn ydn ω = (2π)n Rp,q Rp,q Z Z 1 = h(y) eiu(x−y,ω) dn ωdn y (2π)n Rp,q Rp,q Z Z Pn 1 i (xm −ym )ωm n m=1 = h(y) e d ωdn y (2π)n Rp,q Rp,q Z
i Note, that we show the inversion symbol −1 as lower index in F−1 , in order to avoid a possible confusion by using two upper indice. The inversion could also be written with the help of the CFT i −i 1 itself as F−1 = (2π) . nF 16
Clifford Fourier transforms 275
=
1 (2π)n
Z
Z
h(y) Rp,q
Z
=
h(y) Rp,q
Rp,q n Y
n Y
ei(xm −ym )ωm dn ωdn y
m=1
δ(xm − ym )dn y
m=1
= h(x),
(5.2.130)
where we have inserted Definition 5.2.61 for the first equality, used the linearity of u according to (5.2.128) forR the second equality, as well as inserted (5.2.128) for the 1 i(xm −ym )ωm dωm = δ(xm − ym ), 1 ≤ m ≤ n, for the fifth third equality, and that 2π Re equality. Additionally, we get the transformation law for partial derivatives h0l (x) = ∂xl h(x), 1 ≤ l ≤ n, for h piecewise smooth and integrable, and h, h0l ∈ L1 (Rp,q ; Cl(p, q)) as F i {h0l }(ω) = ωl F i {h}(ω)i, for 1 ≤ l ≤ n. (5.2.131) Proof. We have F i {h0l }(ω) =
Z Rp,q
Z
= Rp,q
Z
= Rp,q
=−
h0l (x) e−iu(y,ω) dn x ∂xl h(x) e−iu(y,ω) dn x ∂xl h(x) e−i
Z
l=1
Rp,q
=−
Pn
Z
h(x) ∂xl e−i h(x) e−i
Rp,q i
Pn l=1
xl ωl n
d x
Pn l=1
xl ωl
dn x
xl ωl n
d x(−iωl )
= ωl F {h}(ω)i,
(5.2.132)
where we inserted u of (5.2.128) for the third equality and performed integration by parts for the fourth equality. The vector derivative of h ∈ L1 (Rp,q ; Cl(p, q)) with h0l ∈ L1 (Rp,q ; Cl(p, q)) gives therefore due to the linearity (5.2.112) of the CFT integral n X
F i {∇h}(ω) = F i {
el h0l }(ω) = ωF i {h}(ω)i.
(5.2.133)
l=1
For the transformation of the spatial moments with hl (x) = xl h(x), 1 ≤ l ≤ n, h, hl ∈ L1 (Rp,q ; Cl(p, q)), we obtain F i {hl }(ω) = ∂ωl F i {h}(ω)i.
(5.2.134)
276 Quaternion and Clifford Fourier Transforms
Proof. We compute i −hl (x)i = h(x)(−ixl ) = F−1 {F i {h}}(x)(−ixl ) Z 1 = F i {h}(ω) eiu(x,ω) dn ω(−ixl ) (2π)n Rp,q Z Pn 1 i xω i n l=1 l l (−ix )d ω = F {h}(ω) e l (2π)n Rp,q Z Pn 1 i xω i l=1 l l =− F {h}(ω) ∂ e dn ω ω l (2π)n Rp,q Z h i Pn 1 i ∂ F {h}(ω) ei l=1 xl ωl dn ω = ω l (2π)n Rp,q
h
i
i = F−1 ∂ωl F i {h} (x),
(5.2.135)
where we used the inversion formula (5.2.129) for the second equality, integration by parts for the sixth equality and (5.2.129) again for the seventh equality. Moreover, by applying the CFT F i to both sides of (5.2.135) we finally obtain F i {hl (−i)}(ω) = ∂ωl F i {h}(ω) ⇔ F i {hl }(ω) = ∂ωl F i {h}(ω)i,
(5.2.136)
because F i {hl (−i)} = F i {hl }(−i). Note that in (5.2.136) the notation (−i) indicates a constant right side multivector factor and not an argument of the function hl . For the spatial vector moment we obtain due to the linearity (5.2.112) of the CFT integral F i {xh}(ω) = F i {
n X
el xl h}(ω) = ∇ω F i {h}(ω)i,
(5.2.137)
l=1
Note that for Cl(p, q) ∼ = M(2d, C) or M(d, H) or M(d, H2 ), or for i being a blade in Cl(p, q) ∼ = M(2d, R) or M(2d, R2 ), we have ei = −i. We assume this for the CFT i F in the following Plancherel and Parseval identities. For the functions h1 , h2 , h ∈ L2 (Rp,q ; Cl(p, q)) we obtain the Plancherel identity hh1 , h2 i =
1 hF i {h1 }, F i {h2 }i, (2π)n
(5.2.138)
as well as the Parseval identity khk =
1
i
F {h}
. (2π)n/2
(5.2.139)
Proof. We only need to proof the Plancherel identity, because the Parseval identity follows from it by setting h1 = h2 = Rh and by taking the square root on both sides. R Assume that ei = −i. We abbreviate = Rp,q , and compute hF i {h1 }, F i {h2 }i Z
=
hF i {h1 }(ω)[F i {h2 }(ω)]∼ idn ω
Z Z Z
=
hh1 (x) e−iu(x,ω) dn x[h2 (y) e−iu(y,ω) dn y]∼ idn ω
Clifford Fourier transforms 277
Z Z Z
n n n ^ hh1 (x) e−iu(x,ω) e−eiu(y,ω) h 2 (y) d yid xd ω
= Z Z Z
n n n ^ hh1 (x) e−iu(x,ω) eiu(y,ω) h 2 (y) d ω d yid x
= Z Z Z
n n n ^ hh1 (x) e−iu(x−y,ω) h 2 (y) d ω d yid x
= n
Z Z Z
= (2π)
= (2π)n
Z Z
= (2π)n
Z
hh1 (x)
hh1 (x)
e−i
n Y
Pn m=1
(xm −ym )ωm
(2π)n
n n n ^ h 2 (y) d ω d yid x
n n ^ δ(xm − ym ) h 2 (y) d yid x
m=1 n ^ hh1 (x)h 2 (x)id x
= (2π)n hh1 , h2 i,
(5.2.140)
where we inserted (3.2.4) for the first equality, the Definition 5.2.61 of the CFT F i for the second equality, applied the principal reverse for the third equality, and the symmetry of the scalar product and that ei = −i for the fourth equality, the linearity of u according to (5.2.128) for the fifth equality, inserted the explicit forms of u R i(x −y )ω 1 m m m dωm = δ(xm − ym ), of (5.2.128) for the sixth equality, and that 2π Re 1 ≤ m ≤ n, for the seventh equality, and again (3.2.4) for the last equality. Division of both sides with (2π)n finally gives the Plancherel identity (5.2.138). 5.2.2.3.3
Convolution
We define the convolution of two multivector signals a, b ∈
L1 (Rp,q ; Cl(p, q)) as Z
(a ? b)(x) =
a(y)b(x − y)dn y.
(5.2.141)
Rp,q
We assume that the function u is linear with respect to its first argument. The CFT of the convolution (5.2.141) can then be expressed as F i {a ? b}(ω) = F −i {a}(ω)F i {b−i }(ω) + F i {a}(ω)F i {b+i }(ω).
(5.2.142)
Proof. We now proof (5.2.142). F i {a ? b}(ω) Z
=
(a ? b)(x) e−iu(x,ω) dn x
Rp,q
Z
Z
= Rp,q
Z
Z
= Rp,q
Z Z
Z
a(y)b(z)dn y e−iu(y,ω) e−iu(z,ω) dn z
Rp,q
Z
= Rp,q
a(y)b(z)dn y e−iu(y+z,ω) dn z
Rp,q
= Rp,q
a(y)b(x − y)dn y e−iu(x,ω) dn x
Rp,q
Rp,q
a(y)[b+i (z) + b−i (z)]dn y e−iu(y,ω) e−iu(z,ω) dn z,
(5.2.143)
278 Quaternion and Clifford Fourier Transforms
where we used the substitution z = x − y, x = y + z. To simplify (5.2.143) we expand the inner expression of the integrand to obtain a(y)[b+i (z) + b−i (z)] e−iu(y,ω) = a(y)[e−iu(y,ω) b+i (z) + e+iu(y,ω) b−i (z)] = a(y)e−iu(y,ω) b+i (z) + a(y)e+iu(y,ω) b−i (z).
(5.2.144)
Reinserting (5.2.144) into (5.2.143) we get F i {a ? b}(ω) Z
=
−iu(y,ω) n
a(y)e Rp,q
Z
d y Rp,q
Z
+ Rp,q
b+i (z)e−iu(z,ω) dn z
a(y)e+iu(y,ω) dn y
Z Rp,q
b−i (z)e−iu(z,ω) dn z
= F i {a}(ω)F i {b+i }(ω) + F −i {a}(ω)F i {b−i }(ω).
(5.2.145)
We point out that the above convolution theorem of equation (5.2.142) is a special case of a more general convolution theorem recently derived in [59]. 5.2.2.4 Summary on onesided Clifford Fourier transforms in Cl(p, q)
We have established a comprehensive new mathematical framework for the investigation and application of Clifford Fourier transforms (CFTs) together with new properties. Our new CFTs form a more general class of CFTs, subsuming and generalizing previous results. We have applied new results (see Section 2.7) on square roots of −1 in Clifford algebras to fully general construct CFTs, with a general square root of −1 in real Clifford algebras Cl(p, q). The new CFTs are fully steerable within the continuous Clifford algebra submanifolds of square roots of −1. We have thus left the terra cognita of familiar transforms to outline the vast array of possible onesided CFTs in Cl(p, q). We defined the central notion of the Clifford Fourier transform with respect to any square root of −1 in real Clifford algebras. Finally, we investigated important properties of these new CFTs: linearity, shift, modulation, dilation, moments, inversion, partial and vector derivatives, Plancherel and Parseval formulas, as well as a convolution theorem. Regarding numerical implementations, usually 2n complex Fourier transformations (FTs) are sufficient. In some cases this can be reduced to 2(n−1) complex FTs, e.g. when the square root of −1 is a pseudoscalar. Further algebraic studies may widen the class of CFTs, where 2(n−1) complex FTs are sufficient. Numerical implementation is then possible with 2n (or 2(n−1) ) discrete complex FTs, which can also be fast Fourier transforms (FFTs), leading to fast CFT implementations. A wellknown example of a CFT is the quaternion FT (QFT) [56,58,124,127,184, 194,291] of Chapter 4, which is particularly used in applications to partial differential systems, color image processing, filtering, disparity estimation (two images differ by
Clifford Fourier transforms 279
local translations) and texture segmentation. Another example is the spacetime FT, see Section 4.4, which leads to a multivector wave packet analysis of spacetime signals (e.g. electromagnetic signals), applicable even to relativistic signals [184, 187]. Depending on the choice of the phase functions u(x, ω) the multivector basis coefficient functions of the CFT result carry information on the symmetry of the signal, similar to the special case of the QFT [56]. The convolution theorem allows to design and apply multivector valued filters to multivector valued signals. Research on the application of CFTs with general square roots of −1 is ongoing. Further results, including special choices of square roots of −1 for certain applications are to be expected. 5.2.3 Operator exponential approach to onesided Clifford Fourier transform 5.2.3.1 Motivation and background
In this section, we study Clifford Fourier transforms (CFT) of multivector functions taking values in Clifford’s geometric algebra, hereby using techniques coming from Clifford analysis (the multivariate function theory for the Dirac operator). In these CFTs on multivector signals, the complex unit i ∈ C is replaced by a multivector square root of −1, see Section 2.7, which may be a pseudoscalar in the simplest case. For these integral transforms we derive an operator representation expressed as the Hamilton operator of a harmonic oscillator. This section is based on [121]. See also [120]. Historically,17 the Clifford Fourier transform (CFT) referred to above was originally introduced by B. Jancewicz [222] for electromagnetic field computations in Clifford’s geometric algebra G3 = Cl(3, 0) of R3 , replacing the imaginary complex unit i ∈ C by the central unit pseudoscalar i3 ∈ G3 , which squares to minus one. This type of CFT was subsequently expanded to G2 , instead using i2 ∈ G2 , and applied to image structure computations by M. Felsberg, see e.g. [132]. Subsequently J. Ebling and G. Scheuermann [117,118] applied both CFTs to the study of vector fields, as they occur in physical flows in dimensions 2 and 3. E. Hitzer and B. Mawardi [185] extended these CFTs to general higher dimensions n = 2, 3(mod 4) and studied their properties, including the physically uncertainty principle for multivector fields. Independently De Bie et al [90] showed how Fourier transforms can be generalized to Clifford algebras in arbitrary dimensions, by introducing an operator form for the complex Fourier transform and generalizing it in the framework of Clifford analysis. This approach relies on a particular realization for the Lie algebra sl(2), closely connected to the Hamilton operator of a harmonic oscillator, which makes it possible to introduce a fractional Clifford Fourier transform (see [92]) and to study radial deformations (see [95, 96]). In the present section, we show how the CFT of [185] on multivector fields in S(Rn , Gn ) can be written in the form of an exponential of the Hamilton operator of the harmonic oscillator in n dimensions and a constant phase factor depending 17
For more historical details, we refer to Chapter 1.
280 Quaternion and Clifford Fourier Transforms
on the underlying dimension. The proof of this result relies on the properties of monogenic functions of degree k and (in particular) of CliffordHermite functions. The computation of an integral transformation is thus replaced by the application of a differential operator, which has profound physical meaning. We therefore expect the content of this section to be valuable not only as another way to represent and compute the CFT, but beyond its mathematical aspect to shed new light on the closely related roles of the harmonic oscillator and the CFT in nature, in particular in physics, and in its wide field of technical applications. Note that in quantum physics the Fourier transform of a wave function is called the momentum representation, and that the multiplication of the wave function with the exponential of the Hamilton operator multiplied with time represents the transition between the Schr¨odinger and the Heisenberg representations [299]. The fact that the exponential of the Hamilton operator can also produce the change from position to momentum representation in quantum mechanics adds a very interesting facet to the picture of quantum mechanics. The section is structured as follows. Section 5.2.3.2 introduces the field of Clifford analysis (a function theory for the Dirac operator). Section 5.2.3.3 finally reviews the notion of CFT and derives the operator representation of the CFTs we are looking for by exponentiating the Hamiltonian of a harmonic oscillator, hereby using a multivector function representation in terms of monogenic polynomials. For a systematic introduction to Clifford’s geometric algebras, we refer to Chapter 2. We remind the reader, that the reverse of M ∈ Gn = Cl(n, 0) is defined by the antiautomorphism f= M
n X
(−1)
k(k−1) 2
hM ik .
(5.2.146)
k=0
The square norm of M is defined by fi , kM k2 = hM M
(5.2.147)
f, in terms of the realvalued (scalar) inner which can also be expressed as M ∗ M product e := hM N ei . M ∗N (5.2.148)
Remark 5.2.62. For vectors a, b ∈ Rn ⊂ Gn the inner product is identical with the scalar product (5.2.148), e = a∗b. a·b=a∗b As a consequence we obtain the multivector CauchySchwarz inequality e i2 ≤ kM k2 kN k2 hM N
∀ M, N ∈ Gn .
(5.2.149)
5.2.3.2 Clifford analysis
We add a brief excursion into Clifford analysis. For vector differential calculus, we refer to Chapter 3. The function theory for the operator ∇, often denoted by means
Clifford Fourier transforms 281
of ∂ x in the literature (and referred to as the Dirac operator), is known as Clifford analysis. This is a multivariate function theory, which can be described as a higherdimensional version of complex analysis or a refinement of harmonic analysis on Rn . The latter is a simple consequence of the fact that ∇2 yields the Laplace operator, the former refers to the fact that the functions on which ∇ acts take their values in the geometric algebra Gn , which then generalizes the algebra C of complex numbers. Remark 5.2.63. Note that in classical Clifford analysis, for which we refer to e.g. [33, 100, 141, 149], one usually works with the geometric algebra (also known as a Clifford algebra) of signature (0, n), which obviously lies closer to the idea of having complex units (n noncommuting complex units e2j = −1, to be precise). However, in this paper we have chosen to work with the geometric algebra Gn associated to the Euclidean signature (n, 0) to stay closer to the situation as it is used in physics. It is important to add that this has little influence on the final conclusions, as most of the results in Clifford analysis (in particular the ones we need in this paper) can be formulated independent of the signature. Clifford analysis can then essentially be described as the function theory for the Dirac operator, in which properties of functions f (x) ∈ ker ∇ are studied. In this section, we will list a few properties. For the main part, we refer to the aforementioned references, or the excellent overview paper [101]. An important definition, in which the analogues of holomorphic powers z k are introduced, is the following: Definition 5.2.64. For all integers k ≥ 0, the vector space Mk (Rn , Gn ) is defined by means of Mk (Rn , Gn ) := Polk (Rn , Gn ) ∩ ker ∇. This is the vector space of khomogeneous monogenics on Rn , containing polynomial null solutions for the Dirac operator. As the Dirac operator ∇ is surjective on polynomials, see e.g. [100], one easily finds for all k ≥ 1 that ! n n n+k−2 dk := dim Mk (R , Gn ) = 2 (5.2.150) k It is trivial to see that d0 = dim Gn . In view of the fact that the operator ∇ factorizes the Laplace operator, null solutions of which are called harmonics, it is obvious that each monogenic polynomial is also harmonic. Indeed, denoting the space of khomogeneous (Gn valued) harmonic polynomials by means of Hk (Rn , Gn ), Mk (Rn , Gn ) ⊂ Hk (Rn , Gn ).
(5.2.151)
As ∇2 is a scalar operator, one can even decompose monogenic polynomials into harmonic ones (using the basis for Gn ): Mk (x) =
X A
(A)
(A)
eA Mk (x) ⇒ ∇2 Mk (x) = 0.
(5.2.152)
282 Quaternion and Clifford Fourier Transforms
This will turn out to be an important property, since we will use a crucial algebraic characterization of the space of realvalued harmonic polynomials in the next section. Another crucial decomposition, which will then be used in the next section, is the socalled Fischer decomposition for Gn valued polynomials on Rn (see e.g. [100] for a proof): Theorem 5.2.65. For any k ∈ N, the space Pk (Rn , Gn ) decomposes into a direct sum of monogenic polynomials: Pk (Rn , Gn ) =
k M
xj Mk−j (Rn , Gn ).
(5.2.153)
j=0
We will often rely on the fact that the operator ∇ and the onevector x ∈ Gn (considered as a multiplication operator, acting on Gn valued functions) span the Lie superalgebra osp(1, 2), which appears as a Howe dual partner for the spin group (see [48] for more details). This translates itself into a collection of operator identities, P and we hereby list the most crucial ones for what follows. Note that Ex := j xj ∂xj denotes the Euler operator on Rm , and {A, B} = AB +BA (resp. [A, B] = AB −BA) denotes the anticommutator (resp. the commutator) of two operators A and B: {x, ∇} = 2 Ex + {x, x} = 2x2 {∇, ∇} = 2∇2
n 2
[∇2 , x2 ] = 4 Ex + [∇, x2 ] = 2x [∇2 , x] = 2∇
n 2
In particular, there also exists an operator which anticommutes with the generators x and ∇ of osp(1, 2), see [13]. This operator, which is known as the Scasimir operator in abstract representation theory (it factorizes the Casimir operator), is related to the Gamma operator from Clifford analysis. To define this latter operator, we need a polar decomposition of the Dirac operator. Due to the change of signature, mentioned in the remark above, we will find a Gamma operator which differs from the one obtained in e.g. [100] by an overall minus sign. In order to retain the most important properties of this operator, we will therefore introduce a new notation Γ∇ , for the operator defined below18 : x∇ = Ex + Γ∇ :=
n X j=1
xj ∂xj +
X
eij (xi ∂xj − xj ∂xi ).
(5.2.154)
i k For the transformation of the spatial moments with hl (x) = xl h(x), 1 ≤ l ≤ n, h, hl ∈ L1 (Rp,q ; Cl(p, q)), we obtain (
F f,g {hl }(ω) =
f ∂ωl F f,g {h}(ω), ∂ωl F f,g {h}(ω) g,
for l ≤ k . for l > k
(5.3.22)
Moreover, for the functions h1 , h2 , h ∈ L2 (Rp,q ; Cl(p, q)) we obtain the Plancherel identity 1 hF f,g {h1 }, F f,g {h2 }i, (5.3.23) hh1 , h2 i = (2π)n as well as the Parseval identity khk = 5.3.1.4.3
Convolution
1
f,g F {h}
.
(2π)n/2
(5.3.24)
We define the convolution of two multivector signals a, b ∈
L1 (Rp,q ; Cl(p, q)) as Z
(a ? b)(x) =
a(y)b(x − y)dn y.
(5.3.25)
Rp,q
For establishing the general twosided CFT of the convolution, we need the identity eαf eβg = eβg eαf + [f, g] sin(α) sin(β),
[f, g] = f g − gf.
(5.3.26)
We further define the following two mixed exponentialsine transforms F
f,±s
{h}(ω) =
Z
e−f u(x,ω) h(x)(±1) sin(−v(x, ω))dn x,
(5.3.27)
(±1) sin(−u(x, ω))h(x)e−gv(x,ω) dn x.
(5.3.28)
Rp,q
F ±s,g {h}(ω) =
Z Rp,q
296 Quaternion and Clifford Fourier Transforms
We assume that the functions u, v are both linear with respect to their first argument. The general twosided CFT of the convolution (5.3.25) can then be expressed as F f,g {a ? b}(ω) = F f,g {a+f }(ω)F f,g {b+g }(ω) + F f,−g {a+f }(ω)F f,g {b−g }(ω) +F f,g {a−f }(ω)F −f,g {b+g }(ω) + F f,−g {a−f }(ω)F −f,g {b−g }(ω) +F f,s {a+f }(ω)[f, g]F s,g {b+g }(ω) + F f,−s {a+f }(ω)[f, g]F s,g {b−g }(ω) +F f,s {a−f }(ω)[f, g]F −s,g {b+g }(ω) + F f,−s {a−f }(ω)[f, g]F −s,g {b−g }(ω). (5.3.29) 5.3.1.5 Reflection on twosided Clifford Fourier transform with two square roots of −1 in
Cl(p, q)
We have by now established a comprehensive new mathematical framework for the investigation and application of Clifford Fourier transforms (CFTs) together with new properties. Our new CFTs form a more general class of CFTs, subsuming and generalizing previous results. We have applied new results on square roots of −1 in Clifford algebras to fully general construct CFTs, with two general square roots of −1 in Clifford algebras Cl(p, q), see Section 2.7. The new CFTs are fully steerable within the continuous Clifford algebra submanifolds of square roots of −1. We have thus left the terra cognita of familiar transforms to outline the vast array of possible twosided CFTs in Cl(p, q), even more general then the onesided CFTs of Section 5.2.2. We showed how the ± split or orthogonal 2D planes split of quaternions can be generalized to split multivector signal functions with respect to a general pair of square roots of −1 in Clifford algebra. Next, we defined the central notion of general twosided Clifford Fourier transforms with respect to any two square roots of −1 in Clifford algebra. Finally, we investigated important properties of these new CFTs: linearity, shift, modulation, dilation, moments, inversion, derivatives, Plancherel and Parseval formulas, as well as a convolution theorem. Regarding numerical implementations, Theorem 5.3.4 shows that 2n complex Fourier transformations (FTs) are sufficient. In some cases this can be reduced to 2(n−1) complex FTs, e.g., when one of the two square roots of −1 is a pseudoscalar. Further algebraic studies may widen the class of CFTs, where 2(n−1) complex FTs are sufficient. Numerical implementation is then possible with 2n (or 2(n−1) ) discrete complex FTs, which can also be fast Fourier transforms (FFTs), leading to fast CFT implementations. A wellknown example of a CFT with two square roots of −1 are the quaternion FTs (QFTs) [56,58,124,127,184,194,291], compare Chapter 4, which are particularly used in applications to partial differential systems, color image processing, filtering, disparity estimation (two images differ by local translations) and texture segmentation. Another example is the spacetime FT, which leads to a multivector wave packet analysis of spacetime signals (e.g. electromagnetic signals), applicable even to relativistic signals [184, 187], see Section 4.4. Depending on the choice of the phase functions u(x, ω) and v(x, ω), the multivector basis coefficient functions of the CFT result carry information on the symmetry of the signal, similar to the special case of the QFT [56].
Clifford Fourier transforms 297
The convolution theorem allows to design and apply multivector valued filters to multivector valued signals. 5.3.2 Twosided (Clifford) geometric Fourier transform with multiple square roots of −1 in Cl(p, q) 5.3.2.1 Background
Here we introduce one single straightforward definition of a general geometric Fourier transform covering many versions in the literature. We show which constraints are additionally necessary to obtain certain features such as linearity or a shift theorem. The section is based on [61]. The kernel of the Fourier transform consists of the complex exponential function. With the square root of minus one, the imaginary unit i, as part of the argument it is periodic and therefore suitable for the analysis of oscillating systems. Geometric algebras usually contain continuous submanifolds of geometric square roots of −1 [193, 203], see Section 2.7. Each multivector has a natural geometric interpretation so the generalization of the Fourier transform to multivector valued functions in the geometric algebras is very reasonable. It helps to interpret the transform, apply it in a targetoriented way to the specific underlying problem and it allows a new point of view in fields like fluid mechanics. For a historical overview of the development of a wide variety of Clifford Fourier transforms, we refer to Chapter 1. The analysis of their similarities reveals a lot about their qualities, too. We concentrate on this matter and try to summarize most of them in one general definition. Here we focus on continuous geometric Fourier transforms over flat spaces Rp,q in their integral representation. That way their finite, regular discrete equivalents as used in computational signal and image processing can be intuitively constructed and direct applicability to the existing practical issues and easy numerical manageability are ensured. 5.3.2.2 Definition of the (Clifford) geometric Fourier transform
For background on Clifford geometric algebra, we refer to Chapter 2. For the sake of brevity we now want to refer to arbitrary multivectors A=
n X
X
aj1 ...jk ej1 . . . ejk ∈ G p,q = Cl(p, q),
(5.3.30)
k=0 1≤j1