quaternion and clifford fourier transforms and wavelets
TRANSCRIPT
Trends in Mathematics
Quaternion and Clifford Fourier Transforms and Wavelets
Eckhard HitzerStephen J. Sangwine Editors
Trends in Mathematics
Trends in Mathematics is a series devoted to the publication of volumes arisingfrom conferences and lecture series focusing on a particular topic from anyarea of mathematics. Its aim is to make current developments available to thecommunity as rapidly as possible without compromise to quality and to archivethese for reference.
Proposals for volumes can be submitted using the Online Book Project Submis-sion Form at our website www.birkhauser-science.com.
Material submitted for publication must be screened and prepared as follows:
All contributions should undergo a reviewing process similar to that carriedout by journals and be checked for correct use of language which, as a rule, isEnglish. Articles without proofs, or which do not contain any significantly newresults, should be rejected. High quality survey papers, however, are welcome.
We expect the organizers to deliver manuscripts in a form that is essentiallyready for direct reproduction. Any version of TEX is acceptable, but the entirecollection of files must be in one particular dialect of TEX and unified accordingto simple instructions available from Birkhauser.
Furthermore, in order to guarantee the timely appearance of the proceedings itis essential that the final version of the entire material be submitted no later thanone year after the conference.
For further volumes:http://www.springer.com/series/4961
Editors
Eckhard HitzerStephen J. Sangwine
Quaternion and Clifford FourierTransforms and Wavelets
Editors Eckhard Hitzer Stephen J. Sangwine Department of Material Science School of Computer Science International Christian University Tokyo, Japan
ISBN 978-3-0348-0602-2 ISBN 978-3-0348-0603-9 (eBook) DOI 10.1007/978-3-0348-0603-9 Springer Basel Heidelberg New York Dordrecht London Library of Congress Control Number: 2013939066 Mathematics Subject Classification (2010): 11R52, 15A66, 42A38, 65T60, 42C40, 68U10, 94A08, 94A12 ยฉ Springer Basel 2013 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the Publisherโs location, in its current version, and permission for use must always be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution under the respective Copyright Law. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect to the material contained herein. Printed on acid-free paper Springer Basel is part of Springer Science+Business Media (www.springer.com)
University of Essex Colchester, United Kingdom
and Electronic Engineering
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
F. Brackx, E. Hitzer and S.J. SangwineHistory of Quaternion and CliffordโFourier Transformsand Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Part I: Quaternions
1 T.A. EllQuaternion Fourier Transform: Re-tooling Image andSignal Processing Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 E. Hitzer and S.J. SangwineThe Orthogonal 2D Planes Split of Quaternions andSteerable Quaternion Fourier Transformations . . . . . . . . . . . . . . . . . . . . . . . 15
3 N. Le Bihan and S.J. SangwineQuaternionic Spectral Analysis of Non-Stationary ImproperComplex Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4 E.U. Moya-Sanchez and E. Bayro-CorrochanoQuaternionic Local Phase for Low-level Image ProcessingUsing Atomic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5 S. Georgiev and J. MoraisBochnerโs Theorems in the Framework of Quaternion Analysis . . . . . . 85
6 S. Georgiev, J. Morais, K.I. Kou and W. SproรigBochnerโMinlos Theorem and Quaternion Fourier Transform . . . . . . . . 105
Part II: Clifford Algebra
7 E. Hitzer, J. Helmstetter and R. AblamowiczSquare Roots of โ1 in Real Clifford Algebras . . . . . . . . . . . . . . . . . . . . . . . 123
8 R. Bujack, G. Scheuermann and E. HitzerA General Geometric Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
vi Contents
9 T. Batard and M. BerthierCliffordโFourier Transform and Spinor Representationof Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
10 P.R. Girard, R. Pujol, P. Clarysse, A. Marion,R. Goutte and P. DelachartreAnalytic Video (2D + ๐ก) Signals Using CliffordโFourier Transformsin Multiquaternion GrassmannโHamiltonโClifford Algebras . . . . . . . . . 197
11 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. HeiseGeneralized Analytic Signals in Image Processing: Comparison,Theory and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
12 R. Soulard and P. CarreColour Extension of Monogenic Wavelets with Geometric Algebra:Application to Color Image Denoising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
13 S. BernsteinSeeing the Invisible and Maxwellโs Equations . . . . . . . . . . . . . . . . . . . . . . . . 269
14 M. BahriA Generalized Windowed Fourier Transform inReal Clifford Algebra ๐ถโ0,๐ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
15 Y. Fu, U. Kahler and P. CerejeirasThe BalianโLow Theorem for the WindowedCliffordโFourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
16 S. Li and T. QianSparse Representation of Signals in Hardy Space . . . . . . . . . . . . . . . . . . . . 321
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, viiโxcโ 2013 Springer Basel
Preface
One hundred and seventy years ago (in 1843) W.R. Hamilton formally introducedthe four-dimensional quaternions, perceiving them as one of the major discoveriesof his life. One year later, in 1844, H. Grassmann published the first version of hisAusdehnungslehre, now known as Grassmann algebra, without any dimensionallimitations. Circa thirty years later (in 1876) W.K. Clifford supplemented theGrassmann product of vectors with an inner product, which fundamentally unifiedthe preceding works of Hamilton and Grassmann in the form of Cliffordโs geometricalgebras or Clifford algebras. A Clifford algebra is a complete algebra of a vectorspace and all its subspaces, including the measurement of volumes and dihedralangles between any pair of subspaces.
To work in higher dimensions with quaternion and Clifford algebras allows usto systematically generalize known concepts of symmetry, phase, analytic signaland holomorphic function to higher dimensions. And as demonstrated in the cur-rent proceedings, it successfully generalizes Fourier and wavelet transformationsto higher dimensions. This is interesting both for the development of analysis inhigher dimensions, as well as for a broad range of applications in multi-dimensionalsignal, image and color image processing. Therefore a wide variety of readers frompure mathematicians, keen to learn about the latest developments in quaternionand Clifford analysis, to physicists and engineers in search of dimensionally ap-propriate and efficient tools in concrete applications, will find many interestingcontributions in this book.
The contributions in this volume originated as papers in a session on Quater-nion and CliffordโFourier transforms and wavelets of the 9th International Con-ference on Clifford Algebras and their Applications (ICCA9), which took placefrom 15th to 20th July 2011 at the Bauhaus-University in Weimar, Germany. Thesession was organized by the editors of this volume.
After the conference we asked the contributors to prepare expanded versionsof their works for this volume, and many of them agreed to participate. The ex-panded submissions were subjected to a further round of reviews (in addition tothe original reviews for the ICCA9 itself) in order to ensure that each contributionwas clearly presented and worthy of publication. We are very grateful to all thosereviewers whose efforts contributed significantly to the quality of the final chaptersby asking the authors to revise, clarify or to expand on points in their drafts.
viii Preface
The contributions have been edited to achieve as much uniformity in presen-tation and notation as can reasonably be achieved across the somewhat differenttraditions that have arisen in the quaternion and Clifford communities. We hopethat this volume will contribute to a growing unification of ideas across the ex-panding field of hypercomplex Fourier transforms and wavelets.
The book is divided into two parts: Chapters 1 to 6 deal exclusively withquaternions โ, while Chapters 7 to 16 mainly deal with Clifford algebras ๐ถโ๐,๐,but sometimes include high-dimensional complex as well as quaternionic resultsin several subsections. This is natural, since complex numbers (โ โผ= ๐ถโ0,1) andquaternions (โ โผ= ๐ถโ0,2) are low-dimensional Clifford algebras, and often appearas subalgebras, e.g., โ โผ= ๐ถ๐+2,0, โ
โผ= ๐ถ๐+3,0, etc. The first chapter was writtenespecially for this volume to provide some background on the history of the subject,and to show how the contributions that follow relate to each other and to priorwork. We especially thank Fred Brackx (Ghent/Belgium) for agreeing to contributeto this chapter at a late stage in the preparation of the book.
The quaternionic part begins with an exploration by Ell (Chapter 1) of theevolution of quaternion Fourier transform (QFT) definitions as a framework forproblems in vector-image and vector-signal processing, ranging from NMR prob-lems to applications in colour image processing. Next, follows an investigation byHitzer and Sangwine (Chapter 2) into a steerable quaternion algebra split, whichleads to: a local phase rotation interpretation of the classical two-sided QFT, effi-cient fast numerical implementations and the design of new steerable QFTs.
Then Le Bihan and Sangwine (Chapter 3) perform a quaternionic spectralanalysis of non-stationary improper complex signals with possible correlation ofreal and imaginary signal parts. With a one-dimensional QFT they introduce ahyperanalytic signal closely linked to the geometric features of improper com-plex signals. In the field of low level image processing Moya-Sanchez and Bayro-Corrochano (Chapter 4) employ quaternionic atomic functions to enhance geo-metric image features and to analytically express image processing operations likelow-pass, steerable and multiscale filtering, derivatives, and local phase computa-tion.
In the next two chapters on quaternion analysis Georgiev and Morais (Chap-ter 5) characterize a class of quaternion Bochner functions generated via a quater-nion FourierโStieltjes transform and generalize Bochnerโs theorem to quaternionfunctions. In Chapter 6 Georgiev, Morais, Kou and Sproรig study the asymptoticbehavior of the QFT, apply the QFT to probability measures, including positivedefinite measures, and extend the classical BochnerโMinlos theorem to the frame-work of quaternion analysis.
The Clifford algebra part begins with Chapter 7 by Hitzer, Helmstetter andAblamowicz, who establish a detailed algebraic characterization of the continuousmanifolds of (multivector) square roots of โ1 in all real Clifford algebras ๐ถโ๐,๐,including as examples detailed computer generated tables of representative squareroots of โ1 in dimensions ๐ = ๐+ ๐ = 5, 7 with signature ๐ = ๐โ ๐ = 3(mod 4).
Preface ix
Their work is fundamental for any form of CliffordโFourier transform (CFT) usingmultivector square roots of โ1 instead of the complex imaginary unit. Based onthis Bujack, Scheuermann and Hitzer (Chapter 8) introduce a general (Clifford)geometric Fourier transform covering most CFTs in the literature. They prove arange of standard properties and specify the necessary conditions in the transformdesign.
A series of four chapters on image processing begins with Batard and Ber-thierโs (Chapter 9) on spinorial representation of images focusing on edge- andtexture detection based on a special CFT for spinor fields, that takes into accountthe Riemannian geometry of the image surface. Then Girard, Pujol, Clarysse, Mar-ion, Goutte and Delachartre (Chapter 10) investigate analytic signals in Cliffordalgebras of ๐-dimensional quadratic spaces, and especially for three-dimensionalvideo (2D + ๐ ) signals in (complex) biquaternions (โผ= ๐ถโ3,0). Generalizing fromthe right-sided QFT to a rotor CFT in ๐ถโ3,0, which allows a complex fast Fouriertransform (FFT) decomposition, they investigate the corresponding analytic videosignal including its generalized six biquaternionic phases. Next, Bernstein, Bou-chot, Reinhardt and Heise (Chapter 11) undertake a mathematical overview of gen-eralizations of analytic signals to higher-dimensional complex and Clifford analysistogether with applications (and comparisons) for artificial and real-world imagesamples.
Soulard and Carre (Chapter 12) define a novel colour monogenic wavelettransform, leading to a non-marginal multiresolution colour geometric analysis ofimages. They show a first application through the definition of a full colour imagedenoising scheme based on statistical modeling of coefficients.
Motivated by applications in optical coherence tomography, Bernstein (Chap-ter 13) studies inverse scattering for Dirac operators with scalar, vector and quater-nionic potentials, by writing Maxwellโs equations as Dirac equations in Cliffordalgebra (i.e., complex biquaternions). For that she considers factorizations of theHelmholtz equation and related fundamental solutions; standard- and FaddeevโsGreen functions.
In Chapter 14 Bahri introduces a windowed CFT for signal functions ๐ :โ๐ โ ๐ถโ0,๐, and investigates some of its properties. For a different type of win-dowed CFT for signal functions ๐ : โ๐ โ ๐ถโ๐,0, ๐ = 2, 3(mod 4), Fu, Kahler andCerejeiras establish in Chapter 15 a BalianโLow theorem, a strong form of Heisen-bergโs classical uncertainty principle. They make essential use of Clifford framesand the CliffordโZak transform.
Finally, Li and Qian (Chapter 16) employ a compressed sensing technique inorder to introduce a new kind of sparse representation of signals in a Hardy spacedictionary (of elementary wave forms) over a unit disk, together with examplesillustrating the new algorithm.
We thank all the authors for their enthusiastic participation in the projectand their enormous patience with the review and editing process. We further thankthe organizer of the ICCA9 conference K. Guerlebeck and his dedicated team for
x Preface
their strong support in organizing the ICCA9 session on Quaternion and CliffordโFourier Transforms and Wavelets. We finally thank T. Hempfling and B. Hellriegelof Birkhauser Springer Basel AG for venturing to accept and skillfully accompanythis proceedings with a still rather unconventional theme, thus going one morestep in fulfilling the 170 year old visions of Hamilton and Grassmann.
Eckhard HitzerTokyo, Japan
Stephen SangwineColchester, United Kingdom
October 2012
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, xiโxxviicโ 2013 Springer Basel
History of Quaternion and CliffordโFourierTransforms and Wavelets
Fred Brackx, Eckhard Hitzer and Stephen J. Sangwine
Abstract. We survey the historical development of quaternion and CliffordโFourier transforms and wavelets.
Mathematics Subject Classification (2010). Primary 42B10; secondary 15A66,16H05, 42C40, 16-03.
Keywords. Quaternions, Clifford algebra, Fourier transforms, wavelet trans-forms.
The development of hypercomplex Fourier transforms and wavelets has taken placein several different threads, reflected in the overview of the subject presented inthis chapter. We present in Section 1 an overview of the development of quaternionFourier transforms, then in Section 2 the development of CliffordโFourier trans-forms. Finally, since wavelets are a more recent development, and the distinctionbetween their quaternion and Clifford algebra approach has been much less pro-nounced than in the case of Fourier transforms, Section 3 reviews the history ofboth quaternion and Clifford wavelets.
We recognise that the history we present here may be incomplete, and thatwork by some authors may have been overlooked, for which we can only offer ourhumble apologies.
1. Quaternion Fourier Transforms (QFT)
1.1. Major Developments in the History of the Quaternion Fourier Transform
Quaternions [51] were first applied to Fourier transforms by Ernst [49, ยง 6.4.2]and Delsuc [41, Eqn. 20] in the late 1980s, seemingly without knowledge of theearlier work of Sommen [90, 91] on CliffordโFourier and Laplace transforms fur-ther explained in Section 2.2. Ernst and Delsucโs quaternion transforms were two-dimensional (that is they had two independent variables) and proposed for ap-plication to nuclear magnetic resonance (NMR) imaging. Written in terms of two
xii F. Brackx, E. Hitzer and S.J. Sangwine
independent time variables1 ๐ก1 and ๐ก2, the forward transforms were of the followingform2:
โฑ(๐1, ๐2) =
โโซโโ
โโซโโ
๐(๐ก1, ๐ก2)๐๐๐1๐ก1๐๐๐2๐ก2d๐ก1d๐ก2 . (1.1)
Notice the use of different quaternion basis units ๐ and ๐ in each of the two ex-ponentials, a feature that was essential to maintain the separation between thetwo dimensions (the prime motivation for using a quaternion Fourier transforma-tion was to avoid the mixing of information that occurred when using a complexFourier transform โ something that now seems obvious, but must have been lessso in the 1980s). The signal waveforms/samples measured in NMR are complex, sothe quaternion aspect of this transform was essential only for maintaining the sep-aration between the two dimensions. As we will see below, there was some unusedpotential here.
The fact that exponentials in the above formulation do not commute (witheach other, or with the โsignalโ function ๐), means that other formulations arepossible3, and indeed Ell in 1992 [45, 46] formulated a transform with the twoexponentials positioned either side of the signal function:
โฑ(๐1, ๐2) =
โโซโโ
โโซโโ
๐๐๐1๐ก1๐(๐ก1, ๐ก2) ๐๐๐2๐ก2 d๐ก1d๐ก2 . (1.2)
Ellโs transform was a theoretical development, but it was soon applied to thepractical problem of computing a holistic Fourier transform of a colour image [84]in which the signal samples (discrete image pixels) had three-dimensional values(represented as quaternions with zero scalar parts). This was a major changefrom the previously intended application in nuclear magnetic resonance, becausenow the two-dimensional nature of the transform mirrored the two-dimensionalnature of the image, and the four-dimensional nature of the algebra used followednaturally from the three-dimensional nature of the image pixels.
Other researchers in signal and image processing have followed Ellโs formu-lation (with trivial changes of basis units in the exponentials) [27, 24, 25], but aswith the NMR transforms, the quaternion nature of the transforms was appliedessentially to separation of the two independent dimensions of an image (Bulowโswork [24, 25] was based on greyscale images, that is with one-dimensional pixelvalues). Two new ideas emerged in 1998 in a paper by Sangwine and Ell [86].These were, firstly, the choice of a general root ๐ of โ1 (a unit quaternion withzero scalar part) rather than a basis unit (๐, ๐ or ๐) of the quaternion algebra,
1The two independent time variables arise naturally from the formulation of two-dimensionalNMR spectroscopy.2Note, that Georgiev et al. use this form of the quaternion Fourier transform (QFT) in Chapter 6to extend the BochnerโMinlos theorem to quaternion analysis. Moreover, the same form of QFTis extended by Georgiev and Morais in Chapter 5 to a quaternion FourierโStieltjes transform.3See Chapter 1 by Ell in this volume with a systematic review of possible forms of quaternionFourier transformations.
Quaternion, CliffordโFourier & Wavelet Transforms History xiii
and secondly, the choice of a single exponential rather than two (giving a choiceof ordering relative to the quaternionic signal function):
โฑ(๐1, ๐2) =
โโซโโ
โโซโโ
๐๐(๐1๐ก1+๐2๐ก2)๐(๐ก1, ๐ก2)d๐ก1d๐ก2 . (1.3)
This made possible a quaternion Fourier transform of a one-dimensional signal:
โฑ(๐) =โโซ
โโ๐๐๐๐ก๐(๐ก)d๐ก . (1.4)
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 threemutually orthogonal transducers, such as a vector geophone.)
Very little has appeared in print about the interpretation of the Fourier coef-ficients resulting from a quaternion Fourier transform. One interpretation is com-ponents of different symmetry, as explained by Ell in Chapter 1. Sangwine and Ellin 2007 published a paper about quaternion Fourier transforms applied to colourimages, with a detailed explanation of the Fourier coefficients in terms of ellipticalpaths in colour space (the ๐-dimensional space of the values of the image pixels ina colour image) [48].
1.2. Splitting Quaternions and the QFT
Following the earlier works of Ernst, Ell, Sangwine (see Section 1.1), and Bulow[24, 25], Hitzer thoroughly studied the quaternion Fourier transform (QFT) appliedto quaternion-valued functions in [54]. As part of this work a quaternion split
๐ยฑ =1
2(๐ ยฑ ๐๐๐), ๐ โ โ, (1.5)
was devised and applied, which led to a better understanding of ๐บ๐ฟ(โ2) trans-formation properties of the QFT spectrum of two-dimensional images, includingcolour images, and opened the way to a generalization of the QFT concept to afull spacetime Fourier transformation (SFT) for spacetime algebra ๐ถโ3,1-valuedsignals.
This was followed up by the establishment of a fully directional (opposedto componentwise) uncertainty principle for the QFT and the SFT [58]. Indepen-dently Mawardi et al. [77] established a componentwise uncertainty principle forthe QFT.
The QFT with a Gabor window was treated by Bulow [24], a study whichhas been continued by Mawardi et al. in [1].
Hitzer reports in [59] initial results (obtained in co-operation with Sangwine)about a further generalization of the QFT to a general form of orthogonal 2Dplanes split (OPS-) QFT, where the split (1.5) with respect to two orthogonalpure quaternion units ๐, ๐ is generalized to a steerable split with respect to any two
xiv F. Brackx, E. Hitzer and S.J. Sangwine
pure unit quaternions ๐, ๐ โ โ, ๐2 = ๐2 = โ1. This approach is fully elaboratedupon in a contribution to the current volume (see Chapter 2). Note that theCayleyโDickson form [87] of quaternions and the related simplex/perplex split[47] are obtained for ๐ = ๐ = ๐ (or more general ๐ = ๐ = ๐), which is employed inChapter 3 for a novel spectral analysis of non-stationary improper complex signals.
2. CliffordโFourier Transformations in CliffordโsGeometric Algebra
W.K. Clifford introduced (Clifford) geometric algebras in 1876 [28]. An introduc-tion to the vector and multivector calculus, with functions taking values in Cliffordalgebras, used in the field of CliffordโFourier transforms (CFT) can be found in[53, 52]. A tutorial introduction to CFTs and Clifford wavelet transforms can befound in [55]. The Clifford algebra application survey [65] contains an up to datesection on applications of Clifford algebra integral transforms, including CFTs,QFTs and wavelet transforms4.
2.1. How Clifford Algebra Square Roots of โ1 Lead to CliffordโFourierTransformations
In 1990 Jancewicz defined a trivector Fourier transformation
โฑ3{๐}(๐) =โซโ3
๐(x)๐โ๐3xโ ๐๐3x, ๐3 = ๐1๐2๐3, ๐ : โ3 โ ๐ถโ3,0, (2.1)
for the electromagnetic field5 replacing the imaginary unit ๐ โ โ by the centraltrivector ๐3, ๐
23 = โ1, of the geometric algebra ๐ถโ3,0 of three-dimensional Euclidean
space โ3 = โ3,0 with orthonormal vector basis {๐1, ๐2, ๐3}.In [50] Felsberg makes use of signal embeddings in low-dimensional Clifford
algebras โ2,0 and โ3,0 to define his CliffordโFourier transform (CFT) for one-dimensional signals as
โฑ๐๐1 [๐ ](๐ข) =
โซโ
exp (โ2๐๐2๐ข๐ฅ) ๐(๐ฅ) ๐๐ฅ, ๐2 = ๐1๐2, ๐ : โโ โ, (2.2)
where he uses the pseudoscalar ๐2 โ ๐ถโ2,0, ๐22 = โ1. For two-dimensional signals6
he defines the CFT as
โฑ๐๐2 [๐ ](๐ข) =
โซโ2
exp (โ2๐๐3 < ๐ข, ๐ฅ >) ๐(๐ฅ) ๐๐ฅ, ๐ : โ2 โ โ2, (2.3)
4Fourier and wavelet transforms provide alternative signal and image representations. See Chap-
ter 9 for a spinorial representation and Chapter 16 by Li and Qian for a sparse representation ofsignals in a Hardy space dictionary (of elementary wave forms) over a unit disk.5Note also Chapter 11 in this volume, in which Bernstein considers optical coherence tomography,formulating the Maxwell equations with the Dirac operator and Clifford algebra.6Note in this context the spinor representation of images by Batard and Berthier in Chapter 9of this volume. The authors apply a CFT in ๐ถโ3,0 to the spinor represenation, which uses inthe exponential kernel an adapted choice of bivector, that belongs to the orthonormal frame
of the tangent bundle of an oriented two-dimensional Riemannian manifold, isometrically im-mersed in โ3.
Quaternion, CliffordโFourier & Wavelet Transforms History xv
where he uses the pseudoscalar ๐3 โ ๐ถโ3,0. It is used amongst others to introducea concept of two-dimensional analytic signal. Together with Bulow and Sommer,Felsberg applied these CFTs to image stucture processing (key-notion: structuremultivector) [50, 24].
Ebling and Scheuermann [44, 43] consequently applied to vector signal pro-cessing in two- and three dimensions, respectively, the following two-dimensionalCFT
โฑ2{๐}(๐) =โซโ2
๐(x)๐โ๐2xโ ๐๐2x, ๐ : โ2 โ โ2, (2.4)
with CliffordโFourier kernel
exp (โ๐1๐2(๐1๐ฅ1 + ๐2๐ฅ2)), (2.5)
and the three-dimensional CFT (2.1) of Jancewicz with CliffordโFourier kernel
exp (โ๐1๐2๐3(๐1๐ฅ1 + ๐2๐ฅ2 + ๐3๐ฅ3)). (2.6)
An important integral operation defined and applied in this context by Eblingand Scheuermann was the Clifford convolution. These CliffordโFourier transformsand the corresponding convolution theorems allow Ebling and Scheuermann foramongst others the analysis of vector-valued patterns in the frequency domain.
Note that the latter Fourier kernel (2.6) has also been used by Mawardi andHitzer in [78, 63, 78] to define their CliffordโFourier transform of three-dimensionalmultivector signals: that means, they researched the properties of โฑ3{๐}(๐) of(2.1) in detail when applied to full multivector signals ๐ : โ3 โ ๐ถโ3,0. This includedan investigation of the uncertainty inequality for this type of CFT. They subse-quently generalized โฑ3{๐}(๐) to dimensions ๐ = 3(mod 4), i.e., ๐ = 3, 7, 11, . . .,
โฑ๐{๐}(๐) =โซโ๐
๐(x)๐โ๐๐xโ ๐๐๐x, ๐ : โ๐ โ ๐ถโ๐,0, (2.7)
which is straightforward, since for these dimensions the pseudoscalar ๐๐ = ๐1 . . . ๐๐squares to โ1 and is central [64], i.e., it commutes with every other multivectorbelonging to ๐ถโ๐,0. A little less trivial is the generalization of โฑ2{๐}(๐) of (2.4) to
โฑ๐{๐}(๐) =โซโ๐
๐(x)๐โ๐๐xโ ๐๐๐x, ๐ : โ๐ โ ๐ถโ๐,0, (2.8)
with ๐ = 2(mod 4), i.e., ๐ = 2, 6, 10 . . ., because in these dimensions the pseu-doscalar ๐๐ = ๐1 . . .๐๐ squares to โ1, but it ceases to be central. So the relativeorder of the factors in โฑ๐{๐}(๐) becomes important, see [66] for a systematicinvestigation and comparison.
In the context of generalizing quaternion Fourier transforms (QFT) via alge-bra isomorphisms to higher-dimensional Clifford algebras, Hitzer [54] constructeda spacetime Fourier transform (SFT) in the full algebra of spacetime ๐ถโ3,1, whichincludes the CFT (2.1) as a partial transform of space. Implemented analogously(isomorphicaly) to the orthogonal 2D planes split of quaternions, the SFT permitsa natural spacetime split, which algebraically splits the SFT into right and leftpropagating multivector wave packets. This analysis allows to compute the effect
xvi F. Brackx, E. Hitzer and S.J. Sangwine
of Lorentz transformations on the spectra of these wavepackets, as well as a 4Ddirectional spacetime uncertainty formula [58] for spacetime signals.
Mawardi et al. extended the CFT โฑ2{๐}(๐) of (2.4) to a windowed CFT in[76]. Fu et al. establish in Chapter 15 a strong version of Heisenbergโs uncertaintyprinciple for Gabor-windowed CFTs.
In Chapter 8 in this volume, Bujack, Scheuermann, and Hitzer, expand thenotion of CliffordโFourier transform to include multiple left and right exponentialkernel factors, in which commuting (or anticommuting) blades, that square toโ1, replace the complex unit ๐ โ โ, thus managing to include most practicallyused CFTs in a single comprehensive framework. Based on this they have alsoconstructed a general CFT convolution theorem [23].
Spurred by the systematic investigation of (complex quaternion) biquater-nion square roots of โ1 in ๐ถโ3,0 by Sangwine [85], Hitzer and Ablamowicz [62]systematically investigated the explicit equations and solutions for square roots ofโ1 in all real Clifford algebras ๐ถโ๐,๐, ๐+ ๐ โค 4. This investigation is continued inthe present volume in Chapter 7 by Hitzer, Helmstetter and Ablamowicz for allsquare roots of โ1 in all real Clifford algebras ๐ถโ๐,๐ without restricting the valueof ๐ = ๐ + ๐. One important motivation for this is the relevance of the Cliffordalgebra square roots of โ1 for the general construction of CFTs, where the imagi-nary unit ๐ โ โ is replaced by a
โโ1 โ ๐ถโ๐,๐, without restriction to pseudoscalarsor blades.
Based on the knowledge of square roots of โ1 in real Clifford algebras ๐ถโ๐,๐,[60] develops a general CFT in ๐ถโ๐,๐, wherein the complex unit ๐ โ โ is replacedby any square root of โ1 chosen from any component and (or) conjugation classof the submanifold of square roots of โ1 in ๐ถโ๐,๐, and details its properties, in-cluding a convolution theorem. A similar general approach is taken in [61] for theconstruction of two-sided CFTs in real Clifford algebras ๐ถโ๐,๐, freely choosing twosquare roots from any one or two components and (or) conjugation classes of thesubmanifold of square roots of โ1 in ๐ถโ๐,๐. These transformations are thereforegenerically steerable.
This algebraically motivated approach may in the future be favorably com-bined with group theoretic, operator theoretic and spinorial approaches, to bediscussed in the following.
2.2. The CliffordโFourier Transform in the Light of Clifford Analysis
Two robust tools used in image processing and computer vision for the analysisof scalar fields are convolution and Fourier transformation. Several attempts havebeen made to extend these methods to two- and three-dimensional vector fieldsand even multi-vector fields. Let us give an overview of those generalized Fouriertransforms.
In [25] Bulow and Sommer define a so-called quaternionic Fourier transformof two-dimensional signals ๐(๐ฅ1, ๐ฅ2) taking their values in the algebra โ of realquaternions. Note that the quaternion algebra โ is nothing else but (isomorphicto) the Clifford algebra ๐ถโ0,2 where, traditionally, the basis vectors are denoted
Quaternion, CliffordโFourier & Wavelet Transforms History xvii
by ๐ and ๐, with ๐2 = ๐2 = โ1, and the bivector by ๐ = ๐๐. In terms of these basisvectors this quaternionic Fourier transform takes the form
โฑ๐[๐ ](๐ข1, ๐ข2) =
โซโ2
exp (โ2๐๐๐ข1๐ฅ1) ๐(๐ฅ1, ๐ฅ2) exp (โ2๐๐๐ข2๐ฅ2) ๐๐ฅ. (2.9)
Due to the non-commutativity of the multiplication in โ, the convolution theoremfor this quaternionic Fourier transform is rather complicated, see also [23].
This is also the case for its higher-dimensional analogue, the so-called Clif-fordโFourier transform7 in ๐ถโ0,๐ given by
โฑ๐๐[๐ ](๐ข) =
โซโ๐
๐(๐ฅ) exp (โ2๐๐1๐ข1๐ฅ1) . . . exp (โ2๐๐๐๐ข๐๐ฅ๐) ๐๐ฅ. (2.10)
Note that for ๐ = 1 and interpreting the Clifford basis vector ๐1 as the imaginaryunit ๐, the CliffordโFourier transform (2.10) reduces to the standard Fourier trans-form on the real line, while for ๐ = 2 the quaternionic Fourier transform (2.9) isrecovered when restricting to real signals.
Finally Bulow and Sommer also introduce a so-called commutative hyper-complex Fourier transform given by
โฑโ[๐ ](๐ข) =
โซโ๐
๐(๐ฅ) exp(โ2๐โ๐
๐=1๏ฟฝ๏ฟฝ๐๐ข๐๐ฅ๐
)๐๐ฅ (2.11)
where the basis vectors (๏ฟฝ๏ฟฝ1, . . . , ๏ฟฝ๏ฟฝ๐) obey the commutative multiplication rules๏ฟฝ๏ฟฝ๐ ๏ฟฝ๏ฟฝ๐ = ๏ฟฝ๏ฟฝ๐๏ฟฝ๏ฟฝ๐ , ๐, ๐ = 1, . . . ,๐, while still retaining ๏ฟฝ๏ฟฝ2
๐ = โ1, ๐ = 1, . . . ,๐. Thiscommutative hypercomplex Fourier transform offers the advantage of a simpleconvolution theorem.
The hypercomplex Fourier transforms โฑ๐, โฑ๐๐ and โฑโ enable Bulow andSommer to establish a theory of multi-dimensional signal analysis and in partic-ular to introduce the notions of multi-dimensional analytic signal8, Gabor filter,instantaneous and local amplitude and phase, etc.
In this context the CliffordโFourier transformations by Felsberg [50] for one-and two-dimensional signals, by Ebling and Scheuermann for two- and three-dimensional vector signal processing [44, 43], and by Mawardi and Hitzer for gen-eral multivector signals in ๐ถโ3,0 [78, 63, 78], and their respective kernels, as alreadyreviewed in Section 2.1, should also be considered.
The above-mentioned CliffordโFourier kernel of Bulow and Sommer
exp (โ2๐๐1๐ข1๐ฅ1) โ โ โ exp (โ2๐๐๐๐ข๐๐ฅ๐) (2.12)
was in fact already introduced in [19] and [89] as a theoretical concept in theframework of Clifford analysis. This generalized Fourier transform was furtherelaborated by Sommen in [90, 91] in connection with similar generalizations ofthe Cauchy, Hilbert and Laplace transforms. In this context also the work of Li,McIntosh and Qian should be mentioned; in [72] they generalize the standard
7Note that in this volume Mawardi establishes in Chapter 14 a windowed version of the CFT
(2.10).8See also Chapter 10 by Girard et al. and Chapter 11 by Bernstein et al. in this volume.
xviii F. Brackx, E. Hitzer and S.J. Sangwine
multi-dimensional Fourier transform of a function in โ๐, by extending the Fourierkernel exp
(๐โจ๐, ๐ฅโฉ)
to a function which is holomorphic in โ๐ and monogenic9
in โ๐+1.In [15, 16, 18] Brackx, De Schepper and Sommen follow another philoso-
phy in their construction of a CliffordโFourier transform. One of the most funda-mental 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 pseudo-differential operator, by embedding Euclidean space into a Clifford algebra, onecan realize
โโฮ๐ as the Dirac operator โ๐ฅ. In this way Clifford analysis sponta-neously refines harmonic analysis. In the same order of ideas, Brackx et al. decidedto not replace nor to improve the classical Fourier transform by a Clifford analysisalternative, since a refinement of it automatically appears within the language ofClifford analysis. The key step to making this refinement apparent is to interpretthe standard Fourier transform as an operator exponential:
โฑ = exp(โ๐
๐
2โ)=
โโ๐=0
1
๐!
(โ๐
๐
2
)๐โ๐ , (2.13)
where โ is the scalar operator
โ =1
2
(โฮ๐ + ๐2 โ๐). (2.14)
This expression links the Fourier transform with the Lie algebra ๐ฐ๐ฉ2 generated
by ฮ๐ and ๐2 = โฃ๐ฅโฃ2 and with the theory of the quantum harmonic oscillatordetermined by the Hamiltonian โ 1
2
(ฮ๐ โ ๐2
). Splitting the operator โ into a
sum of Clifford algebra-valued second-order operators containing the angular Diracoperator ฮ, one is led, in a natural way, to a pair of transforms โฑโยฑ , the harmonicaverage of which is precisely the standard Fourier transform:
โฑโยฑ = exp
(๐๐๐
4
)exp
(โ ๐๐ฮ
2
)exp
(๐๐
4
(ฮ๐ โ ๐2
)). (2.15)
For the special case of dimension two, Brackx et al. obtain a closed form forthe kernel of the integral representation of this CliffordโFourier transform leadingto its internal representation
โฑโยฑ [๐ ](๐) = โฑโยฑ [๐ ](๐1, ๐2) =1
2๐
โซโ2
exp(ยฑ๐12(๐1๐ฅ2 โ ๐2๐ฅ1)
)๐(๐ฅ) ๐๐ฅ , (2.16)
which enables the generalization of the calculation rules for the standard Fouriertransform both in the ๐ฟ1 and in the ๐ฟ2 context. Moreover, the CliffordโFouriertransform of Ebling and Scheuermann
โฑ๐[๐ ](๐) =
โซโ2
exp (โ๐12(๐ฅ1๐1 + ๐ฅ2๐2)) ๐(๐ฅ) ๐๐ฅ , (2.17)
9See also in this volume Chapter 4 by Moya-Sanchez and Bayro-Corrochano on the applicationof atomic function based monogenic signals.
Quaternion, CliffordโFourier & Wavelet Transforms History xix
can be expressed in terms of the CliffordโFourier transform:
โฑ๐[๐ ](๐) = 2๐โฑโยฑ [๐ ](โ๐2,ยฑ๐1) = 2๐โฑโยฑ [๐ ](ยฑ๐12๐) , (2.18)
taking into account that, under the isomorphism between the Clifford algebras๐ถโ2,0 and ๐ถโ0,2, both pseudoscalars are isomorphic images of each other.
The question whether โฑโยฑ can be written as an integral transform is an-swered positively in the case of even dimension by De Bie and Xu in [39]. Theintegral kernel of this transform is not easy to obtain and looks quite complicated.In the case of odd dimension the problem is still open.
Recently, in [35], De Bie and De Schepper have studied the fractional CliffordโFourier transform as a generalization of both the standard fractional Fourier trans-form and the CliffordโFourier transform. It is given as an operator exponential by
โฑ๐ผ,๐ฝ = exp
(๐๐ผ๐
2
)exp (๐๐ฝฮ) exp
(๐๐ผ
2
(ฮ๐ โ ๐2
)). (2.19)
For the corresponding integral kernel a series expansion is obtained, and, in thecase of dimension two, an explicit expression in terms of Bessel functions.
The above, more or less chronological, overview of generalized Fourier trans-forms in the framework of quaternionic and Clifford analysis, gives the impressionof a medley of ad hoc constructions. However there is a structure behind some ofthese generalizations, which becomes apparent when, as already slightly touchedupon above, the Fourier transform is linked to group representation theory, inparticular the Lie algebras ๐ฐ๐ฉ2 and ๐ฌ๐ฐ๐ญ(1โฃ2). This unifying character is beautifullydemonstrated by De Bie in the overview paper [34], where, next to an extensivebibliography, also new results on some of the transformations mentioned belowcan be found. It is shown that using realizations of the Lie algebra ๐ฐ๐ฉ2 one is leadto 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 reinventedseveral times as well in mathematics as in physics, and is attributed to Namias[81], Condon [30], Bargmann [2], Collins [29], Moshinsky and Quesne [80];for a detailed overview of the theory and recent applications of the fractionalFourier transform we refer the reader to [82];
(ii) the Dunkl transform, see, e.g., [42], where the symmetry is reduced to thatof a finite reflection group;
(iii) the radially deformed Fourier transform, see, e.g., [71], which encompassesboth the fractional Fourier and the Dunkl transform;
(iv) the super Fourier transform, see, e.g., [33, 31], which is defined in the contextof superspaces and is invariant under the product of the orthogonal with thesymplectic group.
Realizations of the Lie algebra ๐ฌ๐ฐ๐ญ(1โฃ2), on the contrary, need the framework ofClifford analysis, and lead to:
xx F. Brackx, E. Hitzer and S.J. Sangwine
(v) the CliffordโFourier transform and the fractional CliffordโFourier transform,both already mentioned above; meanwhile an entire class of CliffordโFouriertransforms has been thoroughly studied in [36];
(vi) the radially deformed hypercomplex Fourier transform, which appears as aspecial case in the theory of radial deformations of the Lie algebra ๐ฌ๐ฐ๐ญ(1โฃ2),see [38, 37], and is a topic of current research, see [32].
3. Quaternion and Clifford Wavelets
3.1. Clifford Wavelets in Clifford Analysis
The interest of the Ghent Clifford Research Group for generalizations of the Fouriertransform in the framework of Clifford analysis, grew out from the study of the mul-tidimensional Continuous Wavelet Transform in this particular setting. Clifford-wavelet theory, however restricted to the continuous wavelet transform, was initi-ated by Brackx and Sommen in [20] and further developed by N. De Schepper in herPhD thesis [40]. The Clifford-wavelets originate from a mother wavelet not only bytranslation and dilation, but also by rotation, making the Clifford-wavelets appro-priate for detecting directional phenomena. Rotations are implemented as specificactions on the variable by a spin element, since, indeed, the special orthogonalgroup SO(๐) is doubly covered by the spin group Spin(๐) of the real Cliffordalgebra ๐ถโ0,๐. The mother wavelets themselves are derived from intentionally de-vised orthogonal polynomials in Euclidean space. It should be noted that theseorthogonal polynomials are not tensor products of one-dimensional ones, but gen-uine multidimensional ones satisfying the usual properties such as a Rodriguesformula, recurrence relations, and differential equations. In this way multidimen-sional Clifford wavelets were constructed grafted on the Hermite polynomials [21],Laguerre polynomials [14], Gegenbauer polynomials [13], Jacobi polynomials [17],and Bessel functions [22].
Taking the dimension ๐ to be even, say๐ = 2๐, introducing a complex struc-ture, i.e., an SO(2๐)-element squaring up to โ1, and considering functions withvalues in the complex Clifford algebra โ2๐, so-called Hermitian Clifford analysisoriginates as a refinement of standard or Euclidean Clifford analysis. It should benoticed that the traditional holomorphic functions of several complex variables ap-pear as a special case of Hermitian Clifford analysis, when the function values arerestricted to a specific homogeneous part of spinor space. In this Hermitian settingthe standard Dirac operator, which is invariant under the orthogonal group O(๐),is split into two Hermitian Dirac operators, which are now invariant under theunitary group U(๐). Also in this Hermitian Clifford analysis framework, multidi-mensional wavelets have been introduced by Brackx, H. De Schepper and Sommen[11, 12], as kernels for a Hermitian Continuous Wavelet Transform, and (general-ized) Hermitian CliffordโHermite polynomials have been devised to generate thecorresponding Hermitian wavelets [9, 10].
Quaternion, CliffordโFourier & Wavelet Transforms History xxi
3.2. Further Developments in Quaternion and Clifford Wavelet Theory
Clifford algebra multiresolution analysis (MRA) has been pioneered by M. Mitrea[79]. Important are also the electromagnetic signal application oriented develop-ments of Clifford algebra wavelets by G. Kaiser [70, 67, 68, 69].
Quaternion MRA Wavelets with applications to image analysis have beendeveloped in [92] by Traversoni. Clifford algebra multiresolution analysis has beenapplied by Bayro-Corrochano [5, 3, 4] to: Clifford wavelet neural networks (infor-mation processing), also considering quaternionic MRA, a quaternionic waveletphase concept, as well as applications to (e.g., robotic) motion estimation andimage processing.
Beyond this Zhao and Peng [94] established a theory of quaternion-valuedadmissible wavelets. Zhao [93] studied Clifford algebra-valued admissible (continu-ous) wavelets using the complex Fourier transform for the spectral representation.Mawardi and Hitzer [74, 75] extended this to continuous Clifford and CliffordโGabor wavelets in ๐ถโ3,0 using the CFT of (2.1) for the spectral representation.They also studied a corresponding Clifford wavelet transform uncertainty prin-ciple. Hitzer [56, 57] generalized this approach to continous admissible Cliffordalgebra wavelets in real Clifford algebras ๐ถโ๐,0 of dimensions ๐ = 2, 3(mod 4),i.e., ๐ = 2, 3, 6, 7, 10, 11, . . .. Restricted to ๐ถโ๐,0 of dimensions ๐ = 2(mod 4) thisapproach has also been taken up in [73].
Kahler et al. [26] treated monogenic (Clifford) wavelets over the unit ball.Bernstein studied Clifford continuous wavelet transforms in ๐ฟ0,2 and ๐ฟ0,3 [6], aswell as monogenic kernels and wavelets on the three-dimensional sphere [7]. Bern-stein et al. [8] further studied Clifford diffusion wavelets on conformally flat cylin-ders and tori. In the current volume Soulard and Carre extend in Chapter 12 thetheory and application of monogenic wavelets to colour image denoising.
References
[1] M. Bahri, E. Hitzer, R. Ashino, and R. Vaillancourt. Windowed Fourier transformof two-dimensional quaternionic signals. Applied Mathematics and Computation,216(8):2366โ2379, June 2010.
[2] V. Bargmann. On a Hilbert space of analytic functions and an associated integraltransform. Communications on Pure and Applied Mathematics, 14(3):187โ214, Aug.1961.
[3] E. Bayro-Corrochano. Multi-resolution image analysis using the quaternion wavelettransform. Numerical Algorithms, 39(1โ3):35โ55, 2005.
[4] E. Bayro-Corrochano. The theory and use of the quaternion wavelet transform. Jour-nal of Mathematical Imaging and Vision, 24:19โ35, 2006.
[5] E. Bayro-Corrochano and M.A. de la Torre Gomora. Image processing using the qua-ternion wavelet transform. In Proceedings of the Iberoamerican Congress on PatternRecognition, CIARPโ 2004, pages 612โ620, Puebla, Mexico, October 2004.
[6] S. Bernstein. Clifford continuous wavelet transforms in ๐0,2 and ๐0,3. In T.E. Simos,G. Psihoyios, and C. Tsitouras, editors, NUMERICAL ANALYSIS AND APPLIED
xxii F. Brackx, E. Hitzer and S.J. Sangwine
MATHEMATICS: International Conference on Numerical Analysis and AppliedMathematics, volume 1048 of AIP Conference Proceedings, pages 634โ637, Psalidi,Kos, Greece, 16โ20 September 2008.
[7] S. Bernstein. Spherical singular integrals, monogenic kernels and wavelets on the 3Dsphere. Advances in Applied Clifford Algebras, 19(2):173โ189, 2009.
[8] S. Bernstein, S. Ebert, and R.S. Krausshar. Diffusion wavelets on conformally flatcylinders and tori. In Simos et al. [88], pages 773โ776.
[9] F. Brackx, H. De Schepper, N. De Schepper, and F. Sommen. The generalized Her-mitean CliffordโHermite continuous wavelet transform. In T.E. Simos, G. Psihoyios,and C. Tsitouras, editors, NUMERICAL ANALYSIS AND APPLIED MATHE-MATICS: International Conference of Numerical Analysis and Applied Mathemat-ics, volume 936 of AIP Conference Proceedings, pages 721โ725, Corfu, Greece, 16โ20September 2007.
[10] F. Brackx, H. De Schepper, N. De Schepper, and F. Sommen. Generalized HermiteanCliffordโHermite polynomials and the associated wavelet transform. MathematicalMethods in the Applied Sciences, 32(5):606โ630, 2009.
[11] F. Brackx, H. De Schepper, and F. Sommen. A Hermitian setting for wavelet analysis:the basics. In Proceedings of the 4th International Conference on Wavelet Analysisand Its Applications, 2005.
[12] F. Brackx, H. De Schepper, and F. Sommen. A theoretical framework for waveletanalysis in a Hermitean Clifford setting. Communications on Pure and Applied Anal-ysis, 6(3):549โ567, 2007.
[13] F. Brackx, N. De Schepper, and F. Sommen. The CliffordโGegenbauer polynomialsand the associated continuous wavelet transform. Integral Transforms and SpecialFunctions, 15(5):387โ404, 2004.
[14] F. Brackx, N. De Schepper, and F. Sommen. The CliffordโLaguerre continuouswavelet transform. Bulletin of the Belgian Mathematical Society โ Simon Stevin,11(2):201โ215, 2004.
[15] F. Brackx, N. De Schepper, and F. Sommen. The CliffordโFourier transform. Journalof Fourier Analysis and Applications, 11(6):669โ681, 2005.
[16] F. Brackx, N. De Schepper, and F. Sommen. The two-dimensional CliffordโFouriertransform. Journal of Mathematical Imaging and Vision, 26(1):5โ18, 2006.
[17] F. Brackx, N. De Schepper, and F. Sommen. CliffordโJacobi polynomials and theassociated continuous wavelet transform in Euclidean space. In Qian et al. [83], pages185โ198.
[18] F. Brackx, N. De Schepper, and F. Sommen. The Fourier transform in Cliffordanalysis. Advances in Imaging and Electron Physics, 156:55โ201, 2009.
[19] F. Brackx, R. Delanghe, and F. Sommen. Clifford Analysis, volume 76. Pitman,Boston, 1982.
[20] F. Brackx and F. Sommen. The continuous wavelet transform in Clifford analysis.In F. Brackx, J.S.R. Chisholm, and V. Soucek, editors, Clifford Analysis and itsApplications, pages 9โ26. Kluwer, 2001.
[21] F. Brackx and F. Sommen. The generalized CliffordโHermite continuous wavelettransform. Advances in Applied Clifford Algebras, 11(S1):219โ231, Feb. 2001.
Quaternion, CliffordโFourier & Wavelet Transforms History xxiii
[22] F. Brackx and F. Sommen. CliffordโBessel wavelets in Euclidean space.MathematicalMethods in the Applied Sciences, 25(16โ18):1479โ1491, Nov./Dec. 2002.
[23] R. Bujack, G. Scheuermann, and E. Hitzer. A general geometric Fourier transformconvolution theorem. Advances in Applied Clifford Algebras, 23(1):15โ38, 2013.
[24] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[25] T. Bulow and G. Sommer. Hypercomplex signals โ a novel extension of the ana-lytic signal to the multidimensional case. IEEE Transactions on Signal Processing,49(11):2844โ2852, Nov. 2001.
[26] P. Cerejeiras, M. Ferreira, and U. Kahler. Monogenic wavelets over the unit ball.Zeitschrift fur Analysis und ihre Anwendungen, 24(4):841โ852, 2005.
[27] V.M. Chernov. Discrete orthogonal transforms with data representation in compo-sition algebras. In Proceedings Scandinavian Conference on Image Analysis, pages357โ364, Uppsala, Sweden, 1995.
[28] W.K. Clifford. On the classification of geometric algebras. In R. Tucker, editor,Math-ematical Papers (1882), pages 397โ401. Macmillian, London, 1876. Found amongstCliffordโs notes, of which the forewords are the abstract that he communicated tothe London Mathematical Society on 10th March 1876 (see A. Diek, R. Kantowski,www.nhn.ou.edu/หski/papers/Clifford/history.ps).
[29] S.A. Collins. Lens-system diffraction integral written in terms of matrix optics. Jour-nal of the Optical Society of America, 60:1168โ1177, 1970.
[30] E.U. Condon. Immersion of the Fourier transform in a continuous group of functionaltransformations. Proceedings of the National Academy of Sciences USA, 23:158โ164,1937.
[31] K. Coulembier and H. De Bie. Hilbert space for quantum mechanics on superspace.Journal of Mathematical Physics, 52, 2011. 063504, 30 pages.
[32] H. De Bie. The kernel of the radially deformed Fourier transform. PreprintarXiv.org/abs/1303.2979, Mar. 2013.
[33] H. De Bie. Fourier transform and related integral transforms in superspace. Journalof Mathematical Analysis and Applications, 345:147โ164, 2008.
[34] H. De Bie. Clifford algebras, Fourier transforms and quantum mechanics. Mathe-matical Methods in the Applied Sciences, 35(18):2198โ2225, 2012.
[35] H. De Bie and N. De Schepper. The fractional CliffordโFourier transform. ComplexAnalysis and Operator Theory, 6(5):1047โ1067, 2012.
[36] H. De Bie, N. De Schepper, and F. Sommen. The class of CliffordโFourier transforms.Journal of Fourier Analysis and Applications, 17:1198โ1231, 2011.
[37] H. De Bie, B. รrsted, P. Somberg, and V. Soucek. The Clifford deformation ofthe Hermite semigroup. Symmetry, integrability and geometry-methods and appli-cations, vol. 9, 010, 2013. doi: 10.3482/SIGMA.2013.010
[38] H. De Bie, B. รrsted, P. Somberg, and V. Soucek. Dunkl operators and a fam-ily of realizations of ๐ฌ๐ฐ๐ญ(1โฃ2). Transactions of the American Mathematical Society,364:3875โ3902, 2012.
xxiv F. Brackx, E. Hitzer and S.J. Sangwine
[39] H. De Bie and Y. Xu. On the CliffordโFourier transform. International MathematicsResearch Notices, 2011(22):5123โ5163, 2011.
[40] N. De Schepper. Multidimensional Continuous Wavelet Transforms and GeneralizedFourier Transforms in Clifford Analysis. PhD thesis, Ghent University, Belgium,2006.
[41] M.A. Delsuc. Spectral representation of 2D NMR spectra by hypercomplex numbers.Journal of magnetic resonance, 77(1):119โ124, Mar. 1988.
[42] C.F. Dunkl. Differential-difference operators associated to reflection groups. Trans-actions of the American Mathematical Society, 311:167โ183, 1989.
[43] J. Ebling and G. Scheuermann. Clifford Fourier transform on vector fields. IEEETransactions on Visualization and Computer Graphics, 11(4):469โ479, July 2005.
[44] J. Ebling and J. Scheuermann. Clifford convolution and pattern matching on vectorfields. In Proceedings IEEE Visualization, volume 3, pages 193โ200, Los Alamitos,CA, 2003. IEEE Computer Society.
[45] T.A. Ell. Hypercomplex Spectral Transformations. PhD thesis, University of Min-nesota, June 1992.
[46] T.A. Ell. Quaternion-Fourier transforms for analysis of 2-dimensional linear time-invariant partial-differential systems. In Proceedings of the 32nd Conference on De-cision and Control, pages 1830โ1841, San Antonio, Texas, USA, 15โ17 December1993. IEEE Control Systems Society.
[47] T.A. Ell and S.J. Sangwine. Decomposition of 2D hypercomplex Fourier transformsinto pairs of complex Fourier transforms. In M. Gabbouj and P. Kuosmanen, editors,Proceedings of EUSIPCO 2000, Tenth European Signal Processing Conference, vol-ume II, pages 1061โ1064, Tampere, Finland, 5โ8 Sept. 2000. European Associationfor Signal Processing.
[48] T.A. Ell and S.J. Sangwine. Hypercomplex Fourier transforms of color images. IEEETransactions on Image Processing, 16(1):22โ35, Jan. 2007.
[49] R.R. Ernst, G. Bodenhausen, and A. Wokaun. Principles of Nuclear Magnetic Res-onance in One and Two Dimensions. International Series of Monographs on Chem-istry. Oxford University Press, 1987.
[50] M. Felsberg. Low-Level Image Processing with the Structure Multivector. PhD thesis,Christian-Albrechts-Universitat, Institut fur Informatik und Praktische Mathematik,Kiel, 2002.
[51] W.R. Hamilton. On a new species of imaginary quantities connected with a theory ofquaternions. Proceedings of the Royal Irish Academy, 2:424โ434, 1843. Online: http://www.maths.tcd.ie/pub/HistMath/People/Hamilton/Quatern1/Quatern1.html.
[52] E. Hitzer. Multivector differential calculus. Advances in Applied Clifford Algebras,12(2):135โ182, 2002.
[53] E. Hitzer. Vector differential calculus. Memoirs of the Faculty of Engineering, FukuiUniversity, 50(1):109โ125, 2002.
[54] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[55] E. Hitzer. Tutorial on Fourier transformations and wavelet transformations in Clif-ford geometric algebra. In K. Tachibana, editor, Lecture Notes of the International
Quaternion, CliffordโFourier & Wavelet Transforms History xxv
Workshop for โComputational Science with Geometric Algebraโ (FCSGA2007), pages65โ87, Nagoya University, Japan, 14โ21 February 2007.
[56] E. Hitzer. Clifford (geometric) algebra wavelet transform. In V. Skala and D. Hilden-brand, editors, Proceedings of GraVisMa 2009, pages 94โ101, Plzen, Czech Repub-lic, 2โ4 September 2009. Online: http://gravisma.zcu.cz/GraVisMa-2009/Papers_2009/!_2009_GraVisMa_proceedings-FINAL.pdf.
[57] E. Hitzer. Real Clifford algebra ๐๐(๐, 0), ๐ = 2, 3(mod 4) wavelet transform. In Simoset al. [88], pages 781โ784.
[58] E. Hitzer. Directional uncertainty principle for quaternion Fourier transforms. Ad-vances in Applied Clifford Algebras, 20(2):271โ284, 2010.
[59] E. Hitzer. OPS-QFTs: A new type of quaternion Fourier transforms based on theorthogonal planes split with one or two general pure quaternions. In T.E. Simos,G. Psihoyios, C. Tsitouras, and Z. Anastassi, editors, NUMERICAL ANALYSISAND APPLIED MATHEMATICS ICNAAM 2011: International Conference on Nu-merical Analysis and Applied Mathematics, volume 1389 of AIP Conference Proceed-ings, pages 280โ283, Halkidiki, Greece, 19โ25 September 2011.
[60] E. Hitzer. The Clifford Fourier transform in real Clifford algebras. In K. Guerlebeck,T. Lahmer, and F. Werner, editors, Proceedings 19th International Conference onthe Application of Computer Science and Mathematics in Architecture and CivilEngineering, Weimar, Germany, 4โ6 July 2012.
[61] E. Hitzer. Two-sided Clifford Fourier transform with two square roots of โ1 in ๐ถโ๐,๐.In Proceedings of the 5th Conference on Applied Geometric Algebras in ComputerScience and Engineering (AGACSE 2012), La Rochelle, France, 2โ4 July 2012.
[62] E. Hitzer and R. Ablamowicz. Geometric roots of โ1 in Clifford algebras ๐ถโ๐,๐ with๐ + ๐ โค 4. Advances in Applied Clifford Algebras, 21(1):121โ144, 2010. Publishedonline 13 July 2010.
[63] E. Hitzer and B. Mawardi. Uncertainty principle for the Clifford geometric alge-bra ๐๐(3, 0) based on Clifford Fourier transform. In T.E. Simos, G. Sihoyios, andC. Tsitouras, editors, International Conference on Numerical Analysis and AppliedMathematics 2005, pages 922โ925, Weinheim, 2005. Wiley-VCH.
[64] E. Hitzer and B. Mawardi. Uncertainty principle for Clifford geometric algebras๐ถโ๐,0, ๐ = 3(mod 4) based on Clifford Fourier transform. In Qian et al. [83], pages47โ56.
[65] E. Hitzer, T. Nitta, and Y. Kuroe. Applications of Cliffordโs geometric algebra.Advances in Applied Clifford Algebras, 2013. accepted.
[66] E.M.S. Hitzer and B. Mawardi. Clifford Fourier transform on multivector fields anduncertainty principles for dimensions ๐ = 2(mod 4) and ๐ = 3(mod 4). Advances inApplied Clifford Algebras, 18(3-4):715โ736, 2008.
[67] G. Kaiser. Communications via holomorphic Green functions. In Clifford Analysisand its Applications, Kluwer NATO Science Series, 2001.
[68] G. Kaiser. Complex-distance potential theory, wave equations, and physical wavelets.Mathematical Methods in the Applied Sciences, 25:1577โ1588, 2002. Invited paper,Special Issue on Clifford Analysis in Applications.
[69] G. Kaiser. Huygensโ principle in classical electrodynamics: a distributional approach.Advances in Applied Clifford Algebras, 22(3):703โ720, 2012.
xxvi F. Brackx, E. Hitzer and S.J. Sangwine
[70] G. Kaiser, E. Heyman, and V. Lomakin. Physical source realization of complex-source pulsed beams. Journal of the Acoustical Society of America, 107:1880โ1891,2000.
[71] T. Kobayashi and G. Mano. Integral formulas for the minimal representation of๐(๐, 2). Acta Applicandae Mathematicae, 86:103โ113, 2005.
[72] X. Li. On the inverse problem for the Dirac operator. Inverse Problems, 23:919โ932,2007.
[73] B. Mawardi, S. Adji, and J. Zhao. Clifford algebra-valued wavelet transform onmultivector fields. Advances in Applied Clifford Algebras, 21(1):13โ30, 2011.
[74] B. Mawardi and E. Hitzer. Clifford algebra ๐ถ๐(3, 0)-valued wavelets and uncertaintyinequality for Clifford Gabor wavelet transformation. Preprints of Meeting of theJapan Society for Industrial and Applied Mathematics, 16โ18 September 2006.
[75] B. Mawardi and E. Hitzer. Clifford algebra ๐ถ๐(3, 0)-valued wavelet transformation,Clifford wavelet uncertainty inequality and Clifford Gabor wavelets. InternationalJournal of Wavelets, Multiresolution and Information Processing, 5(6):997โ1019,2007.
[76] B. Mawardi, E. Hitzer, and S. Adji. Two-dimensional Clifford windowed Fouriertransform. In E. Bayro-Corrochano and G. Scheuermann, editors, Applied GeometricAlgebras in Computer Science and Engineering, pages 93โ106. Springer, London,2010.
[77] B. Mawardi, E. Hitzer, A. Hayashi, and R. Ashino. An uncertainty principlefor quaternion Fourier transform. Computers and Mathematics with Applications,56(9):2411โ2417, 2008.
[78] B. Mawardi and E.M.S. Hitzer. Clifford Fourier transformation and uncertainty prin-ciple for the Clifford algebra ๐ถโ3,0. Advances in Applied Clifford Algebras, 16(1):41โ61, 2006.
[79] M. Mitrea. Clifford Wavelets, Singular Integrals, and Hardy Spaces, volume 1575 ofLecture notes in mathematics. Springer, Berlin, 1994.
[80] M. Moshinsky and C. Quesne. Linear canonical transformations and their unitaryrepresentations. Journal of Mathematical Physics, 12:1772โ1780, 1971.
[81] V. Namias. The fractional order Fourier transform and its application to quantummechanics. IMA Journal of Applied Mathematics, 25(3):241โ265, 1980.
[82] H. Ozaktas, Z. Zalevsky, and M. Kutay. The Fractional Fourier Transform. Wiley,Chichester, 2001.
[83] T. Qian, M.I. Vai, and Y. Xu, editors. Wavelet Analysis and Applications, Appliedand Numerical Harmonic Analysis. Birkhauser Basel, 2007.
[84] S.J. Sangwine. Fourier transforms of colour images using quaternion, or hypercom-plex, numbers. Electronics Letters, 32(21):1979โ1980, 10 Oct. 1996.
[85] S.J. Sangwine. Biquaternion (complexified quaternion) roots of -1. Advances in Ap-plied Clifford Algebras, 16(1):63โ68, June 2006.
[86] S.J. Sangwine and T.A. Ell. The discrete Fourier transform of a colour image. In J.M.Blackledge and M.J. Turner, editors, Image Processing II Mathematical Methods, Al-gorithms and Applications, pages 430โ441, Chichester, 2000. Horwood Publishing forInstitute of Mathematics and its Applications. Proceedings Second IMA Conferenceon Image Processing, De Montfort University, Leicester, UK, September 1998.
Quaternion, CliffordโFourier & Wavelet Transforms History xxvii
[87] S.J. Sangwine and N. Le Bihan. Quaternion polar representation with a complexmodulus and complex argument inspired by the CayleyโDickson form. Advances inApplied Clifford Algebras, 20(1):111โ120, Mar. 2010. Published online 22 August2008.
[88] T.E. Simos, G. Psihoyios, and C. Tsitouras, editors. NUMERICAL ANALYSIS ANDAPPLIED MATHEMATICS: International Conference on Numerical Analysis andApplied Mathematics 2009, volume 1168 of AIP Conference Proceedings, Rethymno,Crete (Greece), 18โ22 September 2009.
[89] F. Sommen. A product and an exponential function in hypercomplex function theory.Applicable Analysis, 12:13โ26, 1981.
[90] F. Sommen. Hypercomplex Fourier and Laplace transforms I. Illinois Journal ofMathematics, 26(2):332โ352, 1982.
[91] F. Sommen. Hypercomplex Fourier and Laplace transforms II. Complex Variables,1(2โ3):209โ238, 1983.
[92] L. Traversoni. Quaternion wavelet problems. In Proceedings of 8th International Sym-posium on Approximation Theory, Texas A& M University, Jan. 1995.
[93] J. Zhao. Clifford algebra-valued admissible wavelets associated with admissiblegroup. Acta Scientarium Naturalium Universitatis Pekinensis, 41(5):667โ670, 2005.
[94] J. Zhao and L. Peng. Quaternion-valued admissible wavelets associated with the 2DEuclidean group with dilations. Journal of Natural Geometry, 20(1/2):21โ32, 2001.
Fred BrackxDepartment of Mathematical AnalysisFaculty of EngineeringGhent UniversityGalglaan 2B-9000 Gent, Belgiume-mail: [email protected]
Eckhard HitzerCollege of Liberal ArtsDepartment of Material ScienceInternational Christian University181-8585 Tokyo, Japane-mail: [email protected]
Stephen J. SangwineSchool of Computer Science
and Electronic EngineeringUniversity of EssexWivenhoe ParkColchester, CO4 3SQ, UKe-mail: [email protected]
Part I
Quaternions
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 3โ14cโ 2013 Springer Basel
1 Quaternion Fourier Transform: Re-toolingImage and Signal Processing Analysis
Todd Anthony Ell
โDid you ask a good question today?โ โ Janet Teig
Abstract. Quaternion Fourier transforms (QFTโs) provide expressive powerand elegance in the analysis of higher-dimensional linear invariant systems.But, this power comes at a cost โ an overwhelming number of choices in theQFT definition, each with consequences. This chapter explores the evolutionof QFT definitions as a framework from which to solve specific problems invector-image and vector-signal processing.
Mathematics Subject Classification (2010). Primary 11R52; secondary 42B10.
Keywords. Quaternion, Fourier transform.
1. Introduction
In recent years there has been an increasing recognition on the part of engineersand investigators in image and signal processing of holistic vector approaches tospectral analysis. Generally speaking, this type of spectral analysis treats the vec-tor components of a system not in an iterated, channel-wise fashion but insteadin a holistic, gestalt fashion. The Quaternion Fourier transform (QFT) is one suchanalysis tool.
One of the earliest documented attempts (1987) at describing this type ofspectral analysis was in the area of two-dimensional nuclear magnet resonance.Ernst, et al. [6, pp. 307โ308] briefly discusses using a hypercomplex Fourier trans-form as a method to independently adjust phase angles with respect to two fre-quency variables in two-dimensional spectroscopy. After introducing the conceptthey immediately fall back to an iterated approach leaving the idea unexplored.For similar reasons, Ell [2] in 1992 independently explored the use of QFTs as atool in the analysis of linear time-invariant systems of partial differential equations(PDEs). Ell specifically โdesignedโ a quaternion Fourier transform whose spectral
4 T.A. Ell
operators allowed him to disambiguate partial derivatives with respect to two dif-ferent independent variables. Ellโs original QFT was given by
๐ป [๐๐,๐๐] =
โซโ2
๐โ๐๐๐กโ (๐ก, ๐) ๐โ๐๐๐๐๐ก๐๐ , (1.1)
where ๐ป [๐๐,๐๐] โ โ (the set of quaternions), ๐ and ๐ are Hamiltonโs hypercom-plex operators, and โ (๐ก, ๐) : โร โโ โ (the set of reals). The partial-differentialequivalent spectral operators for this transform are given by
โ
โ๐กโ (๐ก, ๐)โ ๐๐๐ป [๐๐,๐๐] ,
โ
โ๐โ (๐ก, ๐)โ ๐ป [๐๐,๐๐]๐๐. (1.2)
These two differentials have clearly different spectral signatures in contrast to thetwo-dimensional iterated complex Fourier transform where
โ
โ๐กโ (๐ก, ๐ )โ ๐๐๐ป [๐๐, ๐๐] ,
โ
โ๐โ (๐ก, ๐ )โ ๐๐๐ป [๐๐, ๐๐] , (1.3)
especially when ๐ = ๐, at which point the complex spectral domain responses areindistinguishable. This was the first step towards stability analysis in designingcontrollers for systems described by PDEs.
The slow adoption of QFTs at the present time by the engineering communityis due in part to their lack of practical understanding of its properties. This slowadoption is further exacerbated by the variety of transform definitions available.But, in the middle of difficulty lies opportunity. Instead of attempting to find thesingle best QFT (which cannot meet every design engineerโs needs) we provideinstead the means to allow the designer to select the definition most appropriateto his specific problem. That means, allow him to re-tool for the analysis problemat hand.
For example, when QFTs were later applied to colour-image processing [4],where each colour pixel in an image is treated as a 3-vector with basis {๐, ๐,๐} โ โ,it became apparent that there was no preferential association of colour-space axeswith either the basis or the QFTโs exponential-kernel axis. This lead to the nextgeneralization of the QFT defined as
โฑ+ [๐, ๐] =
โซโ2
๐โ๐(๐๐ก+๐๐)๐ (๐ก, ๐ ) ๐๐ก๐๐ , (1.4)
where the transform kernel axis ๐ is any pure unit quaternion, i.e.,
๐ โ {๐๐ฅ+ ๐๐ฆ + ๐๐ง โ โ โฃ ๐ฅ2 + ๐ฆ2 + ๐ง2 = 1}
so that ๐2 = โ1. Still later it was realized [9] that since there is no preferreddirection of indexing the imageโs pixels then the sign on the transform kernel isalso arbitrary, so that a forward QFT could also be defined as
โฑโ [๐, ๐] =
โซโ2
๐+๐(๐๐ก+๐๐)๐ (๐ก, ๐ ) ๐๐ก๐๐ , (1.5)
and the two definitions could be intermixed without concern of creating a non-causal set of image processing filters. This led to several simplifications of a spectralform of the vector correlation operation on two images [8].
1. Quaternion Fourier Transform 5
Bearing in mind such diverse application of various QFTs, the focus of thiswork is to detail as broad a set of QFT definitions as possible, and where known,some of the issues associated with applying them to problems in signal and im-age processing. It also includes a review of approaches taken to define the inter-relations between the various QFT definitions.
2. Preliminaries
To provide a basis for discussion this section gives nomenclature, basic facts onquaternions, and some useful subsets and algebraic equations.
2.1. Just the Facts
The quaternion algebra over the reals โ, denoted by
โ = {๐ = ๐0 + ๐๐1 + ๐๐2 + ๐๐3 โฃ ๐0, ๐1, ๐2, ๐3 โ โ} , (2.1)
is an associative non-commutative four-dimensional algebra, which obeys Hamil-tonโs multiplication rules
๐๐ = ๐ = โ๐๐, ๐๐ = ๐ = โ๐๐, ๐๐ = ๐ = โ๐๐, (2.2)
๐2 = ๐2 = ๐2 = ๐๐๐ = โ1. (2.3)
The quaternion conjugate is defined by
๐ = ๐0 โ ๐๐1 โ ๐๐2 โ ๐๐3 , (2.4)
which is an anti-involution, i.e., ๐ = ๐, ๐+ ๐ = ๐ + ๐, and ๐๐ = ๐ ๐. The norm ofa quaternion is defined as
โฃ๐โฃ = โ๐๐ =โ
๐20 + ๐21 + ๐22 + ๐23 . (2.5)
Using the conjugate and norm of ๐, one can define the inverse of ๐ โ โ โ {0} as
๐โ1 = ๐/โฃ๐โฃ2. (2.6)
Two classical operators on quaternions are the vector- and scalar-part, ๐ [.] and๐[.], respectively; these are defined as
๐ [๐] = ๐๐1 + ๐๐2 + ๐๐3, ๐ [๐] = ๐0. (2.7)
2.2. Useful Subsets
Various subsets of the quaternions are of interest and used repeatedly throughoutthis work. The 3-vector subset of โ is the set of pure quaternions defined as
๐ [โ] = {๐ = ๐๐1 + ๐๐2 + ๐๐3 โ โ } . (2.8)
The set of pure, unit length quaternions is denoted ๐3โ, i.e.,
๐3โ =
{๐ = ๐๐1 + ๐๐2 + ๐๐3 โ โ โฃ ๐21 + ๐22 + ๐23 = 1
}. (2.9)
6 T.A. Ell
Each element of ๐3โcreates a distinct copy of the complex numbers because ๐2 =
โ1, that is, each creates an injective ring homomorphism from โ to โ. So, foreach ๐ โ ๐3
โ, we associate a complex sub-field of โ denoted
โ๐ ={๐ผ+ ๐ฝ๐; โฃ ๐ผ, ๐ฝ โ โ,๐ โ ๐3
โ
}. (2.10)
2.3. Useful Algebraic Equations
In various quaternion equations the non-commutativity of the multiplication causesdifficulty, however, there are algebraic forms which assist in making simplifications.The following three defined forms appear to be the most useful.
Definition 2.1 (Even-Odd Form). Every ๐ : โ2 โ โ can be split into even andodd parts along the ๐ฅ- and ๐ฆ-axis as
๐ (๐ฅ, ๐ฆ) = ๐ee (๐ฅ, ๐ฆ) + ๐eo (๐ฅ, ๐ฆ) + ๐oe (๐ฅ, ๐ฆ) + ๐oo (๐ฅ, ๐ฆ) (2.11)
where ๐eo denotes the part of ๐ that is even with respect to ๐ฅ and odd with respectto ๐ฆ, etc., given as
๐ee (๐ฅ, ๐ฆ) =14 (๐ (๐ฅ, ๐ฆ) + ๐ (โ๐ฅ, ๐ฆ) + ๐ (๐ฅ,โ๐ฆ) + ๐ (โ๐ฅ,โ๐ฆ)) ,
๐eo (๐ฅ, ๐ฆ) =14 (๐ (๐ฅ, ๐ฆ) + ๐ (โ๐ฅ, ๐ฆ)โ ๐ (๐ฅ,โ๐ฆ)โ ๐ (โ๐ฅ,โ๐ฆ)) ,
๐oe (๐ฅ, ๐ฆ) =14 (๐ (๐ฅ, ๐ฆ)โ ๐ (โ๐ฅ, ๐ฆ) + ๐ (๐ฅ,โ๐ฆ)โ ๐ (โ๐ฅ,โ๐ฆ)) ,
๐oo (๐ฅ, ๐ฆ) =14 (๐ (๐ฅ, ๐ฆ)โ ๐ (โ๐ฅ, ๐ฆ)โ ๐ (๐ฅ,โ๐ฆ) + ๐ (โ๐ฅ,โ๐ฆ)) .
(2.12)
Definition 2.2 (Symplectic Form [3]). Every ๐ = ๐0 + ๐๐1 + ๐๐2 + ๐๐3 โ โ can berewritten in terms of a new basis of operators {๐1,๐2,๐3} as
๐ = ๐โฒ0 + ๐1๐โฒ1 + ๐2๐
โฒ2 + ๐3๐
โฒ3 = (๐โฒ0 + ๐1๐
โฒ1) + (๐โฒ2 + ๐1๐
โฒ3)๐2 , (2.13)
where ๐1๐2 = ๐3, ๐1,2,3 โ ๐3โ, hence they form an orthogonal triad.
Remark 2.3. The mapping {๐1, ๐2, ๐3} โ {๐โฒ1, ๐โฒ2, ๐โฒ3} is a change in basis from{๐, ๐,๐} to {๐1,๐2,๐3} via
๐โฒ0 = ๐0, ๐โฒ๐ = โ 12 (๐ [๐]๐๐ + ๐๐๐ [๐]) , ๐ = {1, 2, 3} . (2.14)
Remark 2.4. The symplectic form essentially decomposes a quaternion with re-spect to a specific complex sub-field. That is
๐ = (๐โฒ0 + ๐1๐โฒ1) + (๐โฒ2 + ๐1๐
โฒ3)๐2 = ๐1 + ๐2๐2 , (2.15)
where ๐1,2 โ โ๐1. The author coined the terms simplex and perplex parts of ๐, for
๐1 and ๐2, respectively.
Remark 2.5. The symplectic form works for any permutation of the basis{๐1,๐2,๐3} so that the simplex and complex parts can be taken from any complexsub-field โ๐๐
Further, the swap rule applies to the last term, i.e., ๐2๐2 = ๐2๐2,where the over bar denotes both quaternion and complex sub-field conjugation.
1. Quaternion Fourier Transform 7
Definition 2.6 (Split Form [7]). Every ๐ โ โ can be split as
๐ = ๐+ + ๐โ, ๐ยฑ = 12 (๐ ยฑ ๐1๐๐2) , (2.16)
where ๐1๐2 = ๐3, and ๐1,2,3 โ ๐3โ.
Remark 2.7. The split form allows for the explicit ordering of factors with respectto the operators. So, for example, ๐ = ๐0 + ๐1๐1 + ๐2๐2 + ๐3๐3 becomes
๐ยฑ = {(๐0 ยฑ ๐3) + ๐1 (๐1 ยฑ ๐2)} 1ยฑ ๐3
2
=1ยฑ ๐3
2{(๐0 ยฑ ๐3)โ ๐2 (๐1 ยฑ ๐2)} . (2.17)
Eulerโs formula holds for quaternions, so any unit length quaternion can bewritten as cos ๐+๐ sin ๐ = ๐๐๐, for ๐ โ โ and ๐ โ ๐3
โ. Here ๐ is referred to as the
(Eigen-) axis and ๐ as the (Eigen-) phase angle. Although in general ๐๐ ๐๐ โ= ๐๐+๐
for ๐, ๐ โ โ, their exponential product is a linear combination of exponentials ofthe sum and difference of their phase angles. This can be written in two ways asshown in the following two propositions.
Proposition 2.8 (Exponential Split). Let ๐1,2 โ ๐3โand ๐, ๐ โ โ, then
๐๐1๐๐๐2๐ = ๐๐1(๐โ๐) 1 + ๐3
2+ ๐๐1(๐+๐) 1โ ๐3
2(2.18)
and
๐๐1๐๐๐2๐ =1 + ๐3
2๐๐2(๐โ๐) +
1โ ๐3
2๐๐2(๐+๐) , (2.19)
where ๐1๐2 = ๐3 and ๐3 โ ๐3โ.
Proof. Application of split form to the exponential product. โก
Proposition 2.9 (Exponential Modulation). Let ๐1,2 โ ๐3โand ๐, ๐ โ โ, then
๐๐1๐๐๐2๐ = 12
(๐๐1(๐+๐) + ๐๐1(๐โ๐)
)โ 1
2๐1
(๐๐1(๐+๐) โ ๐๐1(๐โ๐)
)๐2 (2.20)
and
๐๐1๐๐๐2๐ = 12
(๐๐2(๐+๐) + ๐๐2(๐โ๐)
)โ 1
2๐1
(๐๐2(๐+๐) โ ๐๐2(๐โ๐)
)๐2 . (2.21)
Proof. Direct application of Eulerโs formula and trigonometric identities. โก
Remark 2.10. The sandwich terms (i.e., ๐1(.)๐2) in the exponential equationsintroduce 4-space rotations into the interpretation of the product [1]. For if ๐ =๐1๐๐2, then ๐ is a rotated version of ๐ about the (๐1,๐2)-plane by ๐
2 .
8 T.A. Ell
3. Quaternion Fourier Transforms
The purpose of this section is to enumerate a list of possible definitions for a quater-nion Fourier transform. This is followed by a discussion regarding various operatorproperties used in the engineering fields that require simple Fourier transform pairsbetween the non-transformed operation and the equivalent Fourier domain oper-ation, i.e., the so-called operator pairs as seen in most engineering textbooks onFourier analysis. Finally, a discussion on how the inter-relationship between QFTdefinitions are explored, not so as to reduce them to a single canonical form, butto provide the investigator a tool to cross between definitions when necessary soas to gain insight into operator properties.
3.1. Transform Definitions
Although there has been much use of the QFT forms currently in circulation, thereare however more available. Not all โdegrees-of-freedomโ have been exploited. Thenon-commutativity of the quaternion multiplication gave rise to the left- and right-handed QFT kernels. The infinite number of square-roots of โ1 (the cardinality of๐โ) gave rise to the two-sided, or sandwiched kernel. One concept left unexploredis the implication of the exponential product of two quaternions, i.e., ๐๐๐๐ โ= ๐๐+๐.When this is also taken into account, the list expands to eight distinct QFTs asenumerated in Table 1.
Table 1. QFT kernel definitions for ๐ : โ2 โ โ.
Left Right Sandwich
Single-axis ๐โ๐1(๐๐ฅ+๐๐ฆ) ๐(.) ๐(.) ๐โ๐1(๐๐ฅ+๐๐ฆ) ๐โ๐1๐๐ฅ ๐(.) ๐โ๐1๐๐ฆ
Dual-axis ๐โ(๐1๐๐ฅ+๐2๐๐ฆ) ๐(.) ๐(.) ๐โ(๐1๐๐ฅ+๐2๐๐ฆ) โ
Factored ๐โ๐1๐๐ฅ๐โ๐2๐๐ฆ ๐(.) ๐(.) ๐โ๐1๐๐ฅ๐โ๐2๐๐ฆ ๐โ๐1๐๐ฅ ๐(.) ๐โ๐2๐๐ฆ
Depending on the value space of ๐(๐ฅ, ๐ฆ), the available number of distinctQFT forms changes. Table 1 shows the options when ๐ : โ2 โ โ. However, if๐ : โ2 โ โ, then all chirality options (left, right, and sandwiched) collapse to thesame form, leaving three distinct choices: single-axis and factored and un-factoreddual-axis forms.
If neither ๐ฅ nor ๐ฆ are time-like, so that causality of the solution is not afactor, then the number of given QFTs doubles. Variations created by conjugat-ing the quaternion-exponential kernel of both the forward and inverse transformare usually a matter of convention โ the signs must be opposites. For non-causalsystems, however, the sign on the kernel can be taken both ways; each definingits own forward transform from the spatial to spatial-frequency domain. To dis-tinguish the two versions of the forward transform, one is called forward the other
1. Quaternion Fourier Transform 9
reverse. Of course, the inverse transform is still obtained by conjugating the cor-responding forward (or reverse) kernel. Hence, one may define the single-axis, left-and right-sided, forward and reverse transforms as follows1.
Definition 3.1 (Single-axis, Left-sided QFT). The single-axis, left-sided, forward(โฑ+๐ฟ) and reverse (โฑโ๐ฟ) QFTs are defined as
โฑยฑ๐ฟ [๐ (๐ฅ, ๐ฆ)] =
โซโซโ2
๐โ๐1(๐๐ฅ+๐๐ฆ)๐ (๐ฅ, ๐ฆ) ๐๐ฅ๐๐ฆ =๐นยฑ๐ฟ [๐, ๐] . (3.1)
Definition 3.2 (Single-axis, Right-sided QFT). The single-axis, right-sided, forward(โฑ+๐ ) and reverse (โฑโ๐ ) QFTs are defined as
โฑยฑ๐ [๐ (๐ฅ, ๐ฆ)] =
โซโซโ2
๐ (๐ฅ, ๐ฆ) ๐โ๐1(๐๐ฅ+๐๐ฆ)๐๐ฅ๐๐ฆ =๐นยฑ๐ [๐, ๐] . (3.2)
All of the entries in Table 1 exploit the fact that unit length complex numbersact as rotation operators within the complex plane. There is, however, anotherrotation operator โ the 3-space rotation operator for which quaternions are famous.Table 2 lists additional definitions under the provision that ๐ takes on valuesrestricted to ๐ [โ]. Note the factor of 1
2 in the kernel exponent, this is included sothat the frequency scales between the various definitions align.
Table 2. QFT kernel definitions exclusively for ๐ : โ2 โ ๐ [โ].
3-Space Rotator
Single-axis ๐โ๐1๐๐ฅ/2 ๐(.) ๐+๐1๐๐ฆ/2
Dual-axis ๐โ(๐1๐๐ฅ+๐2๐๐ฆ)/2 ๐(.) ๐+(๐1๐๐ฅ+๐2๐๐ฆ)/2
Dual-axis, factored ๐โ๐1๐๐ฅ/2๐โ๐2๐๐ฆ/2 ๐(.) ๐+๐2๐๐ฆ/2๐+๐1๐๐ฅ/2
Taking all these permutations in mind, one arrives at 22 unique QFT defini-tions.
3.2. Functional Relationships
There are several properties used in complex Fourier transform (CFT) analysis thatone hopes will carry over to the QFT in some fashion. These are listed in Table3 from which we will discuss the challenges which arise in QFT analysis. In whatfollows let ๐ (๐ฅ, ๐ฆ)โ ๐น [๐, ๐] denote transform pairs, i.e., โฑ [๐ (๐ฅ, ๐ฆ)] = ๐น [๐, ๐] isthe forward (or reverse) transform and โฑโ1 [๐น (๐, ๐)] = ๐ (๐ฅ, ๐ฆ) is its inversion.
Inversion. Every transform should be invertible. Although this seems obvious,there are instances where a given transform is not. For example, if the re-striction of ๐ : โ2 โ ๐ [โ] were not imposed on the inputs of Table 2, then
1Note, that in (3.1) and (3.2) the arguments ๐ฅ and ๐ฆ of ๐ in โฑยฑ๐ฟ,๐ [๐ (๐ฅ, ๐ฆ)] are shown for clarity,
but are actually dummy arguments, which are integrated out. A more mathematical notationwould be โฑยฑ๐ฟ,๐ {๐}(๐, ๐) = ๐นยฑ๐ฟ,๐ (๐, ๐).
10 T.A. Ell
Table 3. Fourier transform โฑ properties. [๐ผ, ๐ฝ, ๐พ, ๐ฟ โ โ]
Property Definition
Inversion โฑโ1 [โฑ [๐ (๐ฅ, ๐ฆ)]] = ๐ (๐ฅ, ๐ฆ)
Linearity ๐ผ๐(๐ฅ, ๐ฆ) + ๐ฝ๐(๐ฅ, ๐ฆ)
Complex Degenerate (๐1 = ๐ and ๐ : โ2 โ โ๐) โ (QFTโผ=CFT)
Convolution ๐ โ ๐(๐ฅ, ๐ฆ) = (?)
Correlation ๐ โ ๐(๐ฅ, ๐ฆ) = (?)
Modulation ๐๐1๐0๐ฅ๐(.), ๐๐2๐0๐ฅ๐(.), ๐(.)๐๐1๐0๐ฆ, ๐(.)๐๐2๐0๐ฆ, etc.
Scaling ๐(๐ฅ/๐ผ, ๐ฆ/๐ฝ)
Translation ๐(๐ฅโ ๐ฅ0, ๐ฆ โ ๐ฆ0)
Rotation ๐(๐ฅ cos๐ผโ ๐ฆ sin๐ผ, ๐ฅ sin๐ผ+ ๐ฆ cos๐ผ)
Axis-reversal ๐(โ๐ฅ, ๐ฆ), ๐(๐ฅ,โ๐ฆ), ๐(โ๐ฅ,โ๐ฆ)
Re-coordinate ๐(๐ผ๐ฅ+ ๐ฝ๐ฆ, ๐พ๐ฅ+ ๐ฟ๐ฆ)
Conjugation ๐(๐ฅ, ๐ฆ)
Differentials โโ๐ฅ ,
โโ๐ฆ ,
โ2
โ๐ฅโ๐ฆ , etc.
every transform of that table would cease to be invertible. This is because anyreal-valued function, or the scalar part of full quaternion valued functions,commute with the kernel factors which then vanish from under the integral.
Linearity. ๐ผ๐(๐ฅ, ๐ฆ)+๐ฝ๐(๐ฅ, ๐ฆ)โ ๐ผ๐น [๐, ๐]+๐ฝ๐บ[๐, ๐] where ๐ผ, ๐ฝ โ โ. A quick checkverifies that this property holds for all proposed QFT definitions.
Complex Degenerate. For the single-axis transforms, if ๐1 = ๐ and ๐ : โ2 โ โ๐,then the QFT should ideally degenerate to the twice iterated complex Fouriertransform. This degenerate property cannot apply to the dual axis, factoredforms of Tables 1 and 2 since if ๐1 = ๐2 = ๐, these forms reduce to theirsingle-axis versions.
Convolution (Faltung) theorem. Rarely does the standard, complex transform typepair ๐ โ ๐ (๐ฅ, ๐ฆ) โ ๐น [๐, ๐]๐บ [๐, ๐] exist in such a simple form for the QFT.Even the very definition of convolution needs an update since ๐ โ ๐ โ= ๐ โ ๐when ๐ and ๐ are โ-valued. The definition is altered again based on which ofthe two functions is translated, i.e., is the integrand ๐(๐ฅ โ ๐ฅโฒ, ๐ฆ โ ๐ฆโฒ)๐(๐ฅ, ๐ฆ)or ๐(๐ฅ, ๐ฆ)๐(๐ฅโ๐ฅโฒ, ๐ฆโ ๐ฆโฒ). This gives rise to at least four distinct convolutiondefinitions. This will also alter the spectral operator pair.
Further, if ๐ is an input function, then ๐ is typically related to theimpulse response of a system. But, if ๐ : โ2 โ ๐ [โ], is a single impulseresponse sufficient to describe such a system? Or does it take at least twoorthogonally oriented impulses, say ๐1๐ฟ(๐ฅ, ๐ฆ) and ๐2๐ฟ(๐ฅ, ๐ฆ), where ๐1 โฅ ๐2
and ๐ฟ(๐ฅ, ๐ฆ) is the Dirac delta function?
1. Quaternion Fourier Transform 11
Correlation. Consider the correlation definition of two โ-valued functions ๐ and๐ (let ๐, ๐ : โโ โ for simplicity of discussion)
๐ โ ๐ (๐ก) =
โซโ
๐ (๐) ๐ (๐ โ ๐ก) ๐๐ =
โซโ
๐ (๐ + ๐ก) ๐ (๐) ๐๐, (3.3)
where substituting ๐ โ ๐ก = ๐ โฒ yields the second form. For the correlation ofreal-valued functions this is entirely sufficient.
However, for โ-valued functions a conjugation operation is required toensure the relation of the autocorrelation functions (๐ โ ๐) to the powerspectrum as required by the Wiener-Khintchine theorem. This effectivelyensures that the power spectrum of a complex auto-correlation is โ-valued.The complex extension to the cross-correlation function can then be given as
๐ โ ๐ (๐ก) =
โซโ
๐ (๐) ๐ (๐ โ ๐ก) ๐๐ , (3.4)
or alternatively as
๐ โ ๐ (๐ก) =
โซโ
๐ (๐) ๐ (๐ + ๐ก) ๐๐ . (3.5)
In general, the literature does not give significance to the direction of theshifted signal (๐ ยฑ ๐ก). However, in the case of vector correlation matchingproblems, such as colour image registration, direction is fundamental.
Taking this into consideration, for โ-valued functions the equivalentcorrelation could be either
๐ โ ๐ (๐ฅ, ๐ฆ) =
โซโ2
๐ (๐ฅโฒ, ๐ฆโฒ) ๐ (๐ฅโฒ โ ๐ฅ, ๐ฆโฒ โ ๐ฆ)๐๐ฅโฒ๐๐ฆโฒ, (3.6)
or
๐ โ ๐ (๐ฅ, ๐ฆ) =
โซโ2
๐ (๐ฅโฒ + ๐ฅ, ๐ฆโฒ + ๐ฆ) ๐ (๐ฅโฒ, ๐ฆโฒ)๐๐ฅโฒ๐๐ฆโฒ, (3.7)
depending on which shift direction is required. For more details see [9, 4].Note that the correlation result is not necessarily โ-valued.
Modulation. There are multiple types of frequency modulation that need to beaddressed. The modulating exponential can be applied from the left or right,can be driven as a function of either input parameter (i.e., ๐ฅ or ๐ฆ), andbe pointing along one of the kernel axes (i.e., ๐1 or ๐2). Some options aredetailed in the Karnaugh map of Table 4.
In summary, when addressing the QFT operator properties one often needsto regress back to the basic operator definitions and their underlying assumptions,then verify they are still valid for generalization to quaternion forms. Either theoperator definition itself needs to be modified (as in the case of correlation) orthe number of permutations on the definition increases (as in convolution andmodulation).
12 T.A. Ell
Table 4. Frequency Modulations
Left Right
๐ฅ๐๐2๐0๐ฅ๐(.) ๐(.)๐๐2๐0๐ฅ ๐2
๐๐1๐0๐ฅ๐(.) ๐(.)๐๐1๐0๐ฅ
๐1
๐ฆ๐๐1๐0๐ฆ๐(.) ๐(.)๐๐1๐0๐ฆ
๐๐2๐0๐ฆ๐(.) ๐(.)๐๐2๐0๐ฆ ๐2
3.3. Relationships between Transforms
At the heart of all methods for determining inter-relationships between variousQFTs is a decomposing process, of either the input function ๐(.) or the exponential-kernel, so that their parts can be commuted into an alternate QFT form. Ell andSangwine [5] used the symplectic form to link the single-axis, left and right, forwardand reverse forms of the QFT via simplex and perplex complex sub-fields. Yeh[10] reworked these relationships and made further connections to the dual-axis,factored form QFT, but instead used even-odd decomposition of the input function.This approach essentially split each QFT into cosine and sine QFTs. Hitzerโs [7]approach was to use the split form to factor the input function and kernel intofactors with respect to the hypercomplex operators, so as to manipulate the resultto an alternate QFT.
The inter-relationships between the various transform definitions not onlygive insight into the subsequent Fourier analysis, they are also used to simplifyoperator pairs. For example, the inter-relationships between the single-axis formsas given in Definitions 3.2 and 3.1 were used with the symplectic form (Def. 2.2)by Ell and Sangwine [5] to arrive at operator pairs for the convolution operator.Let the single-sided convolutions be defined as follows.
Definition 3.3 (Convolution [5]). The left- and right-sided convolution are defined,respectively, as
โ๐ฟ โ ๐ (๐ฅ, ๐ฆ) =
โซโซโ2
โ๐ฟ (๐ฅโฒ, ๐ฆโฒ) ๐ (๐ฅโ ๐ฅโฒ, ๐ฆ โ ๐ฆโฒ) ๐๐ฅโฒ๐๐ฆโฒ,
๐ โ โ๐ (๐ฅ, ๐ฆ) =
โซโซโ2
๐ (๐ฅโ ๐ฅโฒ, ๐ฆ โ ๐ฆโฒ)โ๐ (๐ฅโฒ, ๐ฆโฒ) ๐๐ฅโฒ๐๐ฆโฒ.(3.8)
Now, let the QFT of the input function ๐ be symplectically decomposed withrespect to ๐1 as
โฑยฑ๐ฟ [๐ (๐ฅ, ๐ฆ)] = ๐นยฑ๐ฟ1 [๐, ๐] + ๐นยฑ๐ฟ2 [๐, ๐]๐2
and
โฑยฑ(๐ฟ,๐ ) [โ๐ (๐ฅ, ๐ฆ)] = ๐ปยฑ(๐ฟ,๐ )๐ [๐, ๐] ,
1. Quaternion Fourier Transform 13
then the right-convolution operator can be written as
โฑยฑ๐ฟ [๐ โ โ๐ ] (๐, ๐) = ๐นยฑ๐ฟ1 [๐, ๐]๐ปยฑ๐ฟ๐ [๐, ๐] + ๐นยฑ๐ฟ2 [๐, ๐]๐2๐ป
โ๐ฟ๐ [๐, ๐] .
Note the use of both forward and reverse QFT transforms. Such a compact oper-ator formula would not be possible without the intermixing of QFT definitions.
4. Conclusions
The three currently defined quaternion Fourier transforms have been shown tobe incomplete. By careful consideration of the underlying reasons for those threeforms, this list has been extended to no less than twenty-two unique definitions.Future work may show that some of these definitions hold little of practical valueor, without loss of generality, they may be reduced to but a few. The shift fromiterated, channel-wise vector analysis to gestalt vector-image and vector-signalanalysis shows promise. This promise raises several challenges:
1. Are there other, more suitable quaternion Fourier transform definitions?2. Can these transforms be reduced to a salient few?3. Are there additional decomposition methods, like the even-odd, split, and
symplectic discussed herein, which can be used?4. All the decomposition methods used to simplify the operator formulas are
at odds with the very gestalt, holistic approach espoused, can this be doneotherwise?
These questions will be the focus of future efforts.
Acknowledgment
Many thanks to Iesus propheta a Nazareth Galilaeae.
References
[1] H.S.M. Coxeter. Quaternions and reflections. The American Mathematical Monthly,53(3):136โ146, Mar. 1946.
[2] T.A. Ell. Hypercomplex Spectral Transformations. PhD thesis, University of Min-nesota, June 1992.
[3] T.A. Ell and S.J. Sangwine. Decomposition of 2D hypercomplex Fourier transformsinto pairs of complex Fourier transforms. In M. Gabbouj and P. Kuosmanen, editors,Proceedings of EUSIPCO 2000, Tenth European Signal Processing Conference, vol-ume II, pages 1061โ1064, Tampere, Finland, 5โ8 Sept. 2000. European Associationfor Signal Processing.
[4] T.A. Ell and S.J. Sangwine. Hypercomplex Wiener-Khintchine theorem with ap-plication to color image correlation. In IEEE International Conference on ImageProcessing (ICIP 2000), volume II, pages 792โ795, Vancouver, Canada, 11โ14 Sept.2000. IEEE.
[5] T.A. Ell and S.J. Sangwine. Hypercomplex Fourier transforms of color images. IEEETransactions on Image Processing, 16(1):22โ35, Jan. 2007.
14 T.A. Ell
[6] R.R. Ernst, G. Bodenhausen, and A. Wokaun. Principles of Nuclear Magnetic Res-onance in One and Two Dimensions. International Series of Monographs on Chem-istry. Oxford University Press, 1987.
[7] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[8] C.E. Moxey, S.J. Sangwine, and T.A. Ell. Vector phase correlation. Electronics Let-ters, 37(25):1513โ1515, Dec. 2001.
[9] C.E. Moxey, S.J. Sangwine, and T.A. Ell. Hypercomplex correlation techniques forvector images. IEEE Transactions on Signal Processing, 51(7):1941โ1953, July 2003.
[10] M.-H. Yeh. Relationships among various 2-D quaternion Fourier transforms. IEEESignal Processing Letters, 15:669โ672, Nov. 2008.
Todd Anthony EllEngineering FellowGoodrich, Sensors and Integrated Systems14300 Judicial RoadBurnsville, MN 55306, USAe-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 15โ39cโ 2013 Springer Basel
2 The Orthogonal 2D Planes Split ofQuaternions and Steerable QuaternionFourier Transformations
Eckhard Hitzer and Stephen J. Sangwine
Abstract. The two-sided quaternionic Fourier transformation (QFT) was in-troduced in [2] for the analysis of 2D linear time-invariant partial-differentialsystems. In further theoretical investigations [4, 5] a special split of quater-nions was introduced, then called ยฑsplit. In the current chapter we analyzethis split further, interpret it geometrically as an orthogonal 2D planes split(OPS), and generalize it to a freely steerable split of โ into two orthogonal2D analysis planes. The new general form of the OPS split allows us to findnew geometric interpretations for the action of the QFT on the signal. Thesecond major result of this work is a variety of new steerable forms of theQFT, their geometric interpretation, and for each form, OPS split theorems,which allow fast and efficient numerical implementation with standard FFTsoftware.
Mathematics Subject Classification (2010). Primary 16H05; secondary 42B10,94A12, 94A08, 65R10.
Keywords. Quaternion signals, orthogonal 2D planes split, quaternion Fouriertransformations, steerable transforms, geometric interpretation, fast imple-mentations.
1. Introduction
The two-sided quaternionic Fourier transformation (QFT) was introduced in [2] forthe analysis of 2D linear time-invariant partial-differential systems. Subsequentlyit has been applied in many fields, including colour image processing [8]. This ledto further theoretical investigations [4, 5], where a special split of quaternions wasintroduced, then called the ยฑsplit. An interesting physical consequence was thatthis split resulted in a left and right travelling multivector wave packet analysis,when generalizing the QFT to a full spacetime Fourier transform (SFT). In thecurrent chapter we investigate this split further, interpret it geometrically and
16 E. Hitzer and S.J. Sangwine
generalize it to a freely steerable1 split of โ into two orthogonal 2D analysis planes.For reasons to become obvious we prefer to call it from now on the orthogonal 2Dplanes split (OPS).
The general form of the OPS split allows us to find new geometric interpre-tations for the action of the QFT on the signal. The second major result of thiswork is a variety of new forms of the QFT, their detailed geometric interpretation,and for each form, an OPS split theorem, which allows fast and efficient numericalimplementation with standard FFT software. A preliminary formal investigationof these new OPS-QFTs can be found in [6].
The chapter is organized as follows. We first introduce in Section 2 severalproperties of quaternions together with a brief review of the ยฑ-split of [4, 5].In Section 3 we generalize this split to a freely steerable orthogonal 2D planessplit (OPS) of quaternions โ. In Section 4 we use the general OPS of Section 3 togeneralize the two-sided QFT to a new two-sided QFT with freely steerable analysisplanes, complete with a detailed local geometric transformation interpretation. Thegeometric interpretation of the OPS in Section 3 further allows the construction ofa new type of steerable QFT with a direct phase angle interpretation. In Section 5we finally investigate new steerable QFTs involving quaternion conjugation. Theirlocal geometric interpretation crucially relies on the notion of 4D rotary reflections.
2. Orthogonal Planes Split of Quaternions withTwo Orthonormal Pure Unit Quaternions
Gauss, Rodrigues and Hamiltonโs four-dimensional (4D) quaternion algebra โ isdefined over โ with three imaginary units:
๐๐ = โ๐๐ = ๐, ๐๐ = โ๐๐ = ๐, ๐๐ = โ๐๐ = ๐,
๐2 = ๐2 = ๐2 = ๐๐๐ = โ1. (2.1)
Every quaternion can be written explicitly as
๐ = ๐๐ + ๐๐๐+ ๐๐๐ + ๐๐๐ โ โ, ๐๐, ๐๐, ๐๐, ๐๐ โ โ, (2.2)
and has a quaternion conjugate (equivalent2 to Clifford conjugation in ๐ถโ+3,0 and
๐ถโ0,2)
๐ = ๐๐ โ ๐๐๐โ ๐๐๐ โ ๐๐๐, ๐๐ = ๐ ๐, (2.3)
which leaves the scalar part ๐๐ unchanged. This leads to the norm of ๐ โ โ
โฃ๐โฃ =โ
๐๐ =โ
๐2๐ + ๐2๐ + ๐2๐ + ๐2๐, โฃ๐๐โฃ = โฃ๐โฃ โฃ๐โฃ . (2.4)
The part V(๐) = ๐ โ ๐๐ = 12 (๐ โ ๐) = ๐๐๐+ ๐๐๐ + ๐๐๐ is called a pure quaternion,
and it squares to the negative number โ(๐2๐ +๐2๐ +๐2๐). Every unit quaternion (i.e.,
1Compare Section 3.4, in particular Theorem 3.5.2This may be important in generalisations of the QFT, such as to a space-time Fourier transformin [4], or a general two-sided CliffordโFourier transform in [7].
2. Orthogonal 2D Planes Split 17
โฃ๐โฃ = 1) can be written as:
๐ = ๐๐ + ๐๐๐+ ๐๐๐ + ๐๐๐ = ๐๐ +โ
๐2๐ + ๐2๐ + ๐2๐ ๐(๐)
= cos๐ผ+ ๐(๐) sin๐ผ = ๐๐ผ๐(๐),(2.5)
where
cos๐ผ = ๐๐, sin๐ผ =โ
๐2๐ + ๐2๐ + ๐2๐,
๐(๐) =V(๐)
โฃ๐โฃ =๐๐๐+ ๐๐๐ + ๐๐๐โ
๐2๐ + ๐2๐ + ๐2๐
, and ๐(๐)2= โ1. (2.6)
The inverse of a non-zero quaternion is
๐โ1 =๐
โฃ๐โฃ2 =๐
๐๐. (2.7)
The scalar part of a quaternion is defined as
S(๐) = ๐๐ =1
2(๐ + ๐), (2.8)
with symmetries
S(๐๐) = S(๐๐) = ๐๐๐๐ โ ๐๐๐๐ โ ๐๐๐๐ โ ๐๐๐๐, S(๐) = S(๐) , โ๐, ๐ โ โ, (2.9)
and linearity
S(๐ผ๐+ ๐ฝ๐) = ๐ผ S(๐) + ๐ฝ S(๐) = ๐ผ๐๐ + ๐ฝ๐๐ , โ๐, ๐ โ โ, ๐ผ, ๐ฝ โ โ. (2.10)
The scalar part and the quaternion conjugate allow the definition of the โ4 innerproduct3 of two quaternions ๐, ๐ as
S(๐๐) = ๐๐๐๐ + ๐๐๐๐ + ๐๐๐๐ + ๐๐๐๐ โ โ. (2.11)
Definition 2.1 (Orthogonality of quaternions). Two quaternions ๐, ๐ โ โ are or-thogonal ๐ โฅ ๐, if and only if the inner product S(๐๐) = 0.
The orthogonal4 2D planes split (OPS) of quaternions with respect to theorthonormal pure unit quaternions ๐, ๐ [4, 5] is defined by
๐ = ๐+ + ๐โ, ๐ยฑ =1
2(๐ ยฑ ๐๐๐). (2.12)
Explicitly in real components ๐๐, ๐๐, ๐๐ , ๐๐ โ โ using (2.1) we get
๐ยฑ = {๐๐ ยฑ ๐๐ + ๐(๐๐ โ ๐๐)}1ยฑ ๐
2=
1ยฑ ๐
2{๐๐ ยฑ ๐๐ + ๐(๐๐ โ ๐๐)}. (2.13)
This leads to the following new Pythagorean modulus identity [5]
Lemma 2.2 (Modulus identity). For ๐ โ โ
โฃ๐โฃ2 = โฃ๐โโฃ2 + โฃ๐+โฃ2 . (2.14)
3Note that we do not use the notation ๐ โ ๐, which is unconventional for full quaternions.4Compare Lemma 2.3.
18 E. Hitzer and S.J. Sangwine
Lemma 2.3 (Orthogonality of OPS split parts). Given any two quaternions ๐, ๐ โ โ
and applying the OPS of (2.12) the resulting parts are orthogonal
S(๐+๐โ) = 0, S(๐โ๐+) = 0, (2.15)
i.e., ๐+ โฅ ๐โ and ๐โ โฅ ๐+.
In Lemma 2.3 (proved in [5]) the second identity follows from the first byS(๐ฅ) = S(๐ฅ) , โ๐ฅ โ โ, and ๐โ๐+ = ๐+๐โ.
It is evident, that instead of ๐, ๐, any pair of orthonormal pure quaternionscan be used to produce an analogous split. This is a first indication, that the OPS of(2.12) is in fact steerable. We observe, that ๐๐๐ = ๐+โ๐โ, i.e., under the map ๐( )๐the ๐+ part is invariant, the ๐โ part changes sign. Both parts are according to (2.13)two-dimensional, and by Lemma 2.3 they span two completely orthogonal planes.The ๐+-plane is spanned by the orthogonal quaternions {๐ โ ๐, 1 + ๐๐ = 1 + ๐},whereas the ๐โ-plane is, e.g., spanned by {๐+ ๐, 1โ ๐๐ = 1โ ๐}, i.e., we have thetwo 2D subspace bases
๐+-basis: {๐โ ๐, 1 + ๐๐ = 1 + ๐}, ๐โ-basis: {๐+ ๐, 1โ ๐๐ = 1โ ๐}. (2.16)
Note that all basis vectors of (2.16)
{๐โ ๐, 1 + ๐๐, ๐+ ๐, 1โ ๐๐} (2.17)
together form an orthogonal basis of โ interpreted as โ4.The map ๐( )๐ rotates the ๐โ-plane by 180โ around the 2D ๐+ axis plane. Note
that in agreement with its geometric interpretation, the map ๐( )๐ is an involution,because applying it twice leads to identity
๐(๐๐๐)๐ = ๐2๐๐2 = (โ1)2๐ = ๐. (2.18)
3. General Orthogonal 2D Planes Split
We will study generalizations of the OPS split by replacing ๐, ๐ by arbitrary unitquaternions ๐, ๐. Even with this generalization, the map ๐( )๐ continues to be aninvolution, because ๐2๐๐2 = (โ1)2๐ = ๐. For clarity we study the cases ๐ โ= ยฑ๐,and ๐ = ๐ separately, though they have a lot in common, and do not always needto be distinguished in specific applications.
3.1. Orthogonal 2D Planes Split Using Two Linearly IndependentPure Unit Quaternions
Our result is now, that all these properties hold, even if in the above considerationsthe pair ๐, ๐ is replaced by an arbitrary pair of linearly independent nonorthogonalpure quaternions ๐, ๐, ๐2 = ๐2 = โ1, ๐ โ= ยฑ๐. The OPS is then re-defined withrespect to the linearly independent pure unit quaternions ๐, ๐ as
๐ยฑ =1
2(๐ ยฑ ๐๐๐). (3.1)
2. Orthogonal 2D Planes Split 19
Equation (2.12) is a special case with ๐ = ๐, ๐ = ๐. We observe from (3.1), that๐๐๐ = ๐+ โ ๐โ, i.e., under the map ๐( )๐ the ๐+ part is invariant, but the ๐โ partchanges sign
๐๐ยฑ๐ =1
2(๐๐๐ ยฑ ๐2๐๐2) =
1
2(๐๐๐ ยฑ ๐) = ยฑ1
2(๐ ยฑ ๐๐๐) = ยฑ๐ยฑ. (3.2)
We now show that even for (3.1) both parts are two-dimensional, and span twocompletely orthogonal planes. The ๐+-plane is spanned
5 by the orthogonal pair ofquaternions {๐ โ ๐, 1 + ๐๐}:
S((๐ โ ๐)(1 + ๐๐)
)= S((๐ โ ๐)(1 + (โ๐)(โ๐)))
= S(๐ + ๐๐๐ โ ๐ โ ๐2๐
) (2.9)= S
(๐ + ๐2๐ โ ๐ + ๐
)= 2S(๐ โ ๐) = 0, (3.3)
whereas the ๐โ-plane is, e.g., spanned by {๐ + ๐, 1 โ ๐๐}. The quaternions ๐ +๐, 1โ ๐๐ can be proved to be mutually orthogonal by simply replacing ๐ โ โ๐ in(3.3). Note that we have
๐(๐ โ ๐)๐ = ๐2๐ โ ๐๐2 = โ๐ + ๐ = ๐ โ ๐,
๐(1 + ๐๐)๐ = ๐๐ + ๐2๐2 = ๐๐ + 1 = 1 + ๐๐,(3.4)
as well as๐(๐ + ๐)๐ = ๐2๐ + ๐๐2 = โ๐ โ ๐ = โ(๐ + ๐),
๐(1โ ๐๐)๐ = ๐๐ โ ๐2๐2 = ๐๐ โ 1 = โ(1โ ๐๐).(3.5)
We now want to generalize Lemma 2.3.
Lemma 3.1 (Orthogonality of two OPS planes). Given any two quaternions ๐, ๐ โโ and applying the OPS (3.1) with respect to two linearly independent pure unitquaternions ๐, ๐ we get zero for the scalar part of the mixed products
S(๐+๐โ) = 0, S(๐โ๐+) = 0. (3.6)
We prove the first identity, the second follows from S(๐ฅ) = S(๐ฅ).
S(๐+๐โ) =1
4S((๐+ ๐๐๐)(๐ โ ๐๐๐)) =
1
4S(๐๐ โ ๐๐๐๐๐๐ + ๐๐๐๐ โ ๐๐๐๐)
(2.10), (2.9)=
1
4S(๐๐ โ ๐๐ + ๐๐๐๐ โ ๐๐๐๐) = 0. (3.7)
Thus the set
{๐ โ ๐, 1 + ๐๐, ๐ + ๐, 1โ ๐๐} (3.8)
forms a 4D orthogonal basis of โ interpreted by (2.11) as โ4, where we have forthe orthogonal 2D planes the subspace bases:
๐+-basis: {๐ โ ๐, 1 + ๐๐}, ๐โ-basis: {๐ + ๐, 1โ ๐๐}. (3.9)
5For ๐ = ๐, ๐ = ๐ this is in agreement with (2.13) and (2.16)!
20 E. Hitzer and S.J. Sangwine
We can therefore use the following representation for every ๐ โ โ by means of fourreal coefficients ๐1, ๐2, ๐3, ๐4 โ โ
๐ = ๐1(1 + ๐๐) + ๐2(๐ โ ๐) + ๐3(1 โ ๐๐) + ๐4(๐ + ๐), (3.10)
where
๐1 = S(๐(1 + ๐๐)โ1
), ๐2 = S
(๐(๐ โ ๐)โ1
),
๐3 = S(๐(1 โ ๐๐)โ1
), ๐4 = S
(๐(๐ + ๐)โ1
).
(3.11)
As an example we have for ๐ = ๐, ๐ = ๐ according to (2.13) the coefficients for thedecomposition with respect to the orthogonal basis (3.8)
๐1 =1
2(๐๐ + ๐๐), ๐2 =
1
2(๐๐ โ ๐๐), ๐3 =
1
2(๐๐ โ ๐๐), ๐4 =
1
2(๐๐ + ๐๐). (3.12)
Moreover, using
๐ โ ๐ = ๐(1 + ๐๐) = (1 + ๐๐)(โ๐), ๐ + ๐ = ๐(1โ ๐๐) = (1โ ๐๐)๐, (3.13)
we have the following left and right factoring properties
๐+ = ๐1(1 + ๐๐) + ๐2(๐ โ ๐) = (๐1 + ๐2๐)(1 + ๐๐)(3.14)
= (1 + ๐๐)(๐1 โ ๐2๐),
๐โ = ๐3(1โ ๐๐) + ๐4(๐ + ๐) = (๐3 + ๐4๐)(1โ ๐๐)(3.15)
= (1โ ๐๐)(๐3 + ๐4๐).
Equations (3.4) and (3.5) further show that the map ๐( )๐ rotates the ๐โ-plane by 180โ around the ๐+ axis plane. We found that our interpretation of themap ๐( )๐ is in perfect agreement with Coxeterโs notion of half-turn in [1]. Thisopens the way for new types of QFTs, where the pair of square roots of โ1 involveddoes not necessarily need to be orthogonal.
Before suggesting a generalization of the QFT, we will establish a new set ofvery useful algebraic identities. Based on (3.14) and (3.15) we get for ๐ผ, ๐ฝ โ โ
๐๐ผ๐๐๐๐ฝ๐ = ๐๐ผ๐๐+๐๐ฝ๐ + ๐๐ผ๐๐โ๐๐ฝ๐,
๐๐ผ๐๐+๐๐ฝ๐ = (๐1 + ๐2๐)๐๐ผ๐ (1 + ๐๐)๐๐ฝ๐ = ๐๐ผ๐ (1 + ๐๐)๐๐ฝ๐(๐1 โ ๐2๐), (3.16)
๐๐ผ๐๐โ๐๐ฝ๐ = (๐3 + ๐4๐)๐๐ผ๐ (1โ ๐๐)๐๐ฝ๐ = ๐๐ผ๐ (1โ ๐๐)๐๐ฝ๐(๐3 + ๐4๐).
Using (3.14) again we obtain
๐๐ผ๐ (1 + ๐๐) = (cos๐ผ+ ๐ sin๐ผ)(1 + ๐๐)
(3.14)= (1 + ๐๐)(cos๐ผโ ๐ sin๐ผ) = (1 + ๐๐)๐โ๐ผ๐,
(3.17)
where we set ๐1 = cos๐ผ, ๐2 = sin๐ผ for applying (3.14). Replacing in (3.17)โ๐ผโ ๐ฝwe get
๐โ๐ฝ๐(1 + ๐๐) = (1 + ๐๐)๐๐ฝ๐, (3.18)
2. Orthogonal 2D Planes Split 21
Furthermore, replacing in (3.17) ๐ โ โ๐ and subsequently ๐ผโ ๐ฝ we get
๐๐ผ๐ (1โ ๐๐) = (1 โ ๐๐)๐๐ผ๐,
๐๐ฝ๐(1โ ๐๐) = (1 โ ๐๐)๐๐ฝ๐.(3.19)
Applying (3.14), (3.16), (3.17) and (3.18) we can rewrite
๐๐ผ๐๐+๐๐ฝ๐(3.16)= (๐1 + ๐2๐)๐
๐ผ๐ (1 + ๐๐)๐๐ฝ๐(3.17)= (๐1 + ๐2๐)(1 + ๐๐)๐(๐ฝโ๐ผ)๐
(3.14)= ๐+๐(๐ฝโ๐ผ)๐, (3.20)
or equivalently as
๐๐ผ๐๐+๐๐ฝ๐(3.16)= ๐๐ผ๐ (1 + ๐๐)๐๐ฝ๐(๐1 โ ๐2๐)
(3.18)= ๐(๐ผโ๐ฝ)๐(1 + ๐๐)(๐1 โ ๐2๐)
(3.14)= ๐(๐ผโ๐ฝ)๐๐+. (3.21)
In the same way by changing ๐ โ โ๐, ๐ฝ โ โ๐ฝ in (3.20) and (3.21) we can rewrite
๐๐ผ๐๐โ๐๐ฝ๐ = ๐(๐ผ+๐ฝ)๐๐โ = ๐โ๐(๐ผ+๐ฝ)๐. (3.22)
The result is therefore
๐๐ผ๐๐ยฑ๐๐ฝ๐ = ๐ยฑ๐(๐ฝโ๐ผ)๐ = ๐(๐ผโ๐ฝ)๐๐ยฑ. (3.23)
3.2. Orthogonal 2D Planes Split Using One Pure Unit Quaternion
We now treat the case for ๐ = ๐, ๐2 = โ1. We then have the map ๐( )๐ , and theOPS split with respect to ๐ โ โ, ๐2 = โ1,
๐ยฑ =1
2(๐ ยฑ ๐๐๐). (3.24)
The pure quaternion ๐ can be rotated by ๐ = ๐(๐+ ๐), see (3.27), into the quater-nion unit ๐ and back. Therefore studying the map ๐( )๐ is, up to the constantrotation between ๐ and ๐ , the same as studying ๐( )๐ . This gives
๐๐๐ = ๐(๐๐ + ๐๐๐+ ๐๐๐ + ๐๐๐)๐ = โ๐๐ โ ๐๐๐+ ๐๐๐ + ๐๐๐. (3.25)
The OPS with respect to ๐ = ๐ = ๐ gives
๐ยฑ =1
2(๐ ยฑ ๐๐๐), ๐+ = ๐๐๐ + ๐๐๐ = (๐๐ + ๐๐๐)๐, ๐โ = ๐๐ + ๐๐๐, (3.26)
where the ๐+-plane is two-dimensional and manifestly orthogonal to the 2D ๐โ-plane. This form (3.26) of the OPS is therefore identical to the quaternionic sim-plex/perplex split of [3].
For ๐ = ๐ the ๐โ-plane is always spanned by {1, ๐}. The rotation operator๐ = ๐(๐+๐), with squared norm โฃ๐ โฃ2 = โฃ๐(๐+๐)โฃ2 = โฃ(๐+๐)โฃ2 = โ(๐+๐)2, rotates๐ into ๐ according to
๐ โ1๐๐ =๐
โฃ๐ โฃ2 ๐๐ =(๐+ ๐)๐๐๐(๐+ ๐)
โ(๐+ ๐)2=
(๐+ ๐)๐(๐(โ๐) + 1)๐
(๐+ ๐)2
=(๐+ ๐)(๐ + ๐)๐
(๐+ ๐)2= ๐.
(3.27)
22 E. Hitzer and S.J. Sangwine
The rotation ๐ leaves 1 invariant and thus rotates the whole {1, ๐} plane into the๐โ-plane spanned by {1, ๐}. Consequently ๐ also rotates the {๐,๐} plane into the๐+-plane spanned by {๐โฒ = ๐ โ1๐๐ , ๐โฒ = ๐ โ1๐๐ }. We thus constructively obtainthe fully orthonormal 4D basis of โ as
{1, ๐, ๐โฒ,๐โฒ} = ๐ โ1{1, ๐, ๐,๐}๐ , ๐ = ๐(๐+ ๐), (3.28)
for any chosen pure unit quaternion ๐ . We further have, for the orthogonal 2Dplanes created in (3.24) the subspace bases:
๐+-basis: {๐โฒ,๐โฒ}, ๐โ-basis: {1, ๐}. (3.29)
The rotation ๐ (an orthogonal transformation!) of (3.27) preserves the fun-damental quaternionic orthonormality and the anticommutation relations
๐๐โฒ = ๐โฒ = โ๐โฒ๐, ๐โฒ๐ = ๐โฒ = โ๐๐โฒ ๐โฒ๐โฒ = ๐ = โ๐โฒ๐โฒ. (3.30)
Hence
๐๐๐ = ๐(๐+ + ๐โ)๐ = ๐+ โ ๐โ, i.e., ๐๐ยฑ๐ = ยฑ๐ยฑ, (3.31)
which represents again a half-turn by 180โ in the 2D ๐โ-plane around the 2D๐+-plane (as axis).
Figures 1 and 2 illustrate this decomposition for the case where ๐ is theunit pure quaternion 1โ
3(๐ + ๐ + ๐). This decomposition corresponds (for pure
quaternions) to the classical luminance-chrominance decomposition used in colourimage processing, as illustrated, for example, in [3, Figure 2]. Three hundredunit quaternions randomly oriented in 4-space were decomposed. Figure 1 showsthe three hundred points in 4-space, projected onto the six orthogonal planes{๐, ๐โฒ}, {๐, ๐โฒ}, {๐,๐โฒ}, {๐โฒ, ๐โฒ}, {๐โฒ,๐โฒ}, {๐โฒ, ๐โฒ} where ๐ = 1 and ๐โฒ = ๐ , as given in(3.28). The six views at the top show the ๐+-plane, and the six below show the๐โ-plane.
Figure 2 shows the vector parts of the decomposed quaternions. The basisfor the plot is {๐โฒ, ๐ โฒ,๐โฒ}, where ๐โฒ = ๐ as given in (3.28). The green circles showthe components in the {1, ๐} plane, which intersects the 3-space of the vector partonly along the line ๐ (which is the luminance or grey line of colour image pixels).The red line on the figure corresponds to ๐ . The blue circles show the componentsin the {๐โฒ,๐โฒ} plane, which is entirely within the 3-space. It is orthogonal to ๐ andcorresponds to the chrominance plane of colour image processing.
The next question is the influence the current OPS (3.24) has for left andright exponential factors of the form
๐๐ผ๐๐ยฑ ๐๐ฝ๐ . (3.32)
We learn from (3.30) that
๐๐ผ๐๐ยฑ๐๐ฝ๐ = ๐(๐ผโ๐ฝ)๐๐ยฑ = ๐ยฑ๐(๐ฝโ๐ผ)๐ , (3.33)
which is identical to (3.23), if we insert ๐ = ๐ in (3.23).
2. Orthogonal 2D Planes Split 23
0.5 0 0.5
0.5
0
0.5
e
iยด
0.5 0 0.5
0.5
0
0.5
ejยด
0.5 0 0.5
0.5
0
0.5
e
kยด0.5 0 0.5
0.5
0
0.5
iยด
jยด
0.5 0 0.5
0.5
0
0.5
jยด
kยด
0.5 0 0.5
0.5
0
0.5
kยดiยด
0.5 0 0.5
0.5
0
0.5
e
iยด
0.5 0 0.5
0.5
0
0.5
e
jยด
0.5 0 0.5
0.5
0
0.5
e
kยด
0.5 0 0.5
0.5
0
0.5
iยด
jยด
0.5 0 0.5
0.5
0
0.5
jยด
kยด
0.5 0 0.5
0.5
0
0.5
kยด
iยด
Figure 1. 4D scatter plot of quaternions decomposed using the or-thogonal planes split of (3.24) with one unit pure quaternion ๐ = ๐โฒ =1โ3(๐+ ๐ + ๐) = ๐.
24 E. Hitzer and S.J. Sangwine
0.50
0.5
0.5
0
0.5
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
jยด
iยด
kยด
Figure 2. Scatter plot of vector parts of quaternions decomposed usingthe orthogonal planes split of (3.24) with one pure unit quaternion ๐ =๐โฒ = 1โ
3(๐+ ๐ + ๐) = ๐. The red line corresponds to the direction of ๐ .
Next, we consider ๐ = โ๐, ๐2 = โ1. We then have the map ๐( )(โ๐), andthe OPS split with respect to ๐,โ๐ โ โ, ๐2 = โ1,
๐ยฑ =1
2(๐ ยฑ ๐๐(โ๐)) =
1
2(๐ โ ๐๐๐). (3.34)
Again we can study ๐ = ๐ first, because for general pure unit quaternions ๐ theunit quaternion ๐ can be rotated by (3.27) into the quaternion unit ๐ and back.Therefore studying the map ๐( )(โ๐) is up to the constant rotation ๐ of (3.27) thesame as studying ๐( )(โ๐). This gives the map
๐๐(โ๐) = ๐(๐๐ + ๐๐๐+ ๐๐๐ + ๐๐๐)(โ๐) = ๐๐ + ๐๐๐โ ๐๐๐ โ ๐๐๐. (3.35)
The OPS with respect to ๐ = ๐, ๐ = โ๐ gives
๐ยฑ =1
2(๐ ยฑ ๐๐(โ๐)), ๐โ = ๐๐๐ + ๐๐๐ = (๐๐ + ๐๐๐)๐, ๐+ = ๐๐ + ๐๐๐, (3.36)
where, compared to ๐ = ๐ = ๐, the 2D ๐+-plane and the 2D ๐โ-planes appear in-terchanged. The form (3.36) of the OPS is again identical to the quaternionic sim-plex/perplex split of [3], but the simplex and perplex parts appear interchanged.
For ๐ = โ๐ the ๐+-plane is always spanned by {1, ๐}. The rotation ๐ of(3.27) rotates ๐ into ๐ and leaves 1 invariant and thus rotates the whole {1, ๐}
2. Orthogonal 2D Planes Split 25
plane into the ๐+-plane spanned by {1, ๐}. Consequently, ๐ of (3.27) also rotatesthe {๐,๐} plane into the ๐โ-plane spanned by {๐โฒ = ๐ โ1๐๐ , ๐โฒ = ๐ โ1๐๐ }.
We therefore have for the orthogonal 2D planes created in (3.34) the subspacebases:
๐+-basis: {1, ๐}, ๐โ-basis: {๐โฒ,๐โฒ}. (3.37)
We again obtain the fully orthonormal 4D basis (3.28) of โ, preservingthe fundamental quaternionic orthonormality and the anticommutation relations(3.30).
Hence for (3.34)
๐๐(โ๐) = ๐(๐+ + ๐โ)(โ๐) = ๐+ โ ๐โ, i.e., ๐๐ยฑ(โ๐) = ยฑ๐ยฑ, (3.38)
which represents again a half-turn by 180โ in the 2D ๐โ-plane around the 2D๐+-plane (as axis).
The remaining question is the influence the current OPS (3.34) has for leftand right exponential factors of the form
๐๐ผ๐๐ยฑ๐โ๐ฝ๐ . (3.39)
We learn from (3.30) that
๐๐ผ๐๐ยฑ๐โ๐ฝ๐ = ๐(๐ผโ๐ฝ)๐๐ยฑ = ๐ยฑ๐โ(๐ฝโ๐ผ)๐ , (3.40)
which is identical to (3.23), if we insert ๐ = โ๐ in (3.23).
For (3.23) therefore, we do not any longer need to distinguish the cases ๐ โ=ยฑ๐ and ๐ = ยฑ๐. This motivates us to a general OPS definition for any pair of purequaternions ๐, ๐, and we get a general lemma.
Definition 3.2 (General orthogonal 2D planes split). Let ๐, ๐ โ โ be an arbitrarypair of pure quaternions ๐, ๐, ๐2 = ๐2 = โ1, including the cases ๐ = ยฑ๐. Thegeneral OPS is then defined with respect to the two pure unit quaternions ๐, ๐ as
๐ยฑ =1
2(๐ ยฑ ๐๐๐). (3.41)
Remark 3.3. The three generalized OPS (3.1), (3.24), and (3.34) are formallyidentical and are now subsumed in (3.41) of Definition 3.2, where the values ๐ = ยฑ๐are explicitly included, i.e., any pair of pure unit quaternions ๐, ๐ โ โ, ๐2 = ๐2 =โ1, is admissible.
Lemma 3.4. With respect to the general OPS of Definition 3.2 we have for left andright exponential factors the identity
๐๐ผ๐๐ยฑ๐๐ฝ๐ = ๐ยฑ๐(๐ฝโ๐ผ)๐ = ๐(๐ผโ๐ฝ)๐๐ยฑ. (3.42)
26 E. Hitzer and S.J. Sangwine
3.3. Geometric Interpretation of Left and Right Exponential Factors in ๐ , ๐
We obtain the following general geometric interpretation. The map ๐( )๐ alwaysrepresents a rotation by angle ๐ in the ๐โ-plane (around the ๐+-plane), the map๐ ๐ก( )๐๐ก, ๐ก โ โ, similarly represents a rotation by angle ๐ก๐ in the ๐โ-plane (aroundthe ๐+-plane as axis). Replacing6 ๐ โ โ๐ in the map ๐( )๐ we further find that
๐๐ยฑ(โ๐) = โ๐ยฑ. (3.43)
Therefore the map ๐( )(โ๐) = ๐( )๐โ1, because ๐โ1 = โ๐, represents a rotationby angle ๐ in the ๐+-plane (around the ๐โ-plane), exchanging the roles of 2Drotation plane and 2D rotation axis. Similarly, the map ๐๐ ( )๐โ๐ , ๐ โ โ, representsa rotation by angle ๐ ๐ in the ๐+-plane (around the ๐โ-plane as axis).
The product of these two rotations gives
๐ ๐ก+๐ ๐๐๐กโ๐ = ๐(๐ก+๐ )๐2 ๐๐๐(๐กโ๐ )
๐2 ๐ = ๐๐ผ๐๐๐๐ฝ๐,
๐ผ = (๐ก+ ๐ )๐
2, ๐ฝ = (๐กโ ๐ )
๐
2,
(3.44)
where based on (2.5) we used the identities ๐ = ๐๐2 ๐ and ๐ = ๐
๐2 ๐.
The geometric interpretation of (3.44) is a rotation by angle ๐ผ + ๐ฝ in the๐โ-plane (around the ๐+-plane), and a second rotation by angle ๐ผโ ๐ฝ in the ๐+-plane (around the ๐โ-plane). For ๐ผ = ๐ฝ = ๐/2 we recover the map ๐( )๐, and for๐ผ = โ๐ฝ = ๐/2 we recover the map ๐( )๐โ1.
3.4. Determination of ๐, ๐ for Given Steerable Pair of Orthogonal 2D Planes
Equations (3.9), (3.29), and (3.37) tell us how the pair of pure unit quaternions๐, ๐ โ โ used in the general OPS of Definition 3.2, leads to an explicit basis forthe resulting two orthogonal 2D planes, the ๐+-plane and the ๐โ-plane. We nowask the opposite question: how can we determine from a given steerable pair oforthogonal 2D planes in โ the pair of pure unit quaternions ๐, ๐ โ โ, which splitsโ exactly into this given pair of orthogonal 2D planes?
To answer this question, we first observe that in a 4D space it is sufficientto know only one 2D plane explicitly, specified, e.g., by a pair of orthogonal unitquaternions ๐, ๐ โ โ, โฃ๐โฃ = โฃ๐โฃ = 1, and without restriction of generality ๐2 = โ1,i.e., ๐ can be a pure unit quaternion ๐ = ๐(๐). But for ๐ = cos๐ผ + ๐(๐) sin๐ผ,compare (2.6), we must distinguish S(๐) = cos๐ผ โ= 0 and S(๐) = cos๐ผ = 0, i.e., of๐ also being a pure quaternion with ๐2 = โ1. The second orthogonal 2D plane isthen simply the orthogonal complement in โ to the ๐, ๐-plane.
Let us first treat the case S(๐) = cos๐ผ โ= 0. We set
๐ := ๐๐, ๐ := ๐๐. (3.45)
6Alternatively and equivalently we could replace ๐ โ โ๐ instead of ๐ โ โ๐.
2. Orthogonal 2D Planes Split 27
With this setting we get for the basis of the ๐โ-plane
๐ + ๐ = ๐๐+ ๐๐ = 2S(๐) ๐,
1โ ๐๐ = 1โ ๐๐๐๐ = 1โ ๐2๐2 = 1 + ๐2
= 1 + cos2 ๐ผโ sin2 ๐ผ+ 2๐(๐) cos๐ผ sin๐ผ
= 2 cos๐ผ(cos๐ผ+ ๐(๐) sin๐ผ) = 2 S(๐) ๐.
(3.46)
For the equality ๐๐๐๐ = ๐2๐2 we used the orthogonality of ๐, ๐, which means thatthe vector part of ๐ must be orthogonal to the pure unit quaternion ๐, i.e., it mustanticommute with ๐
๐๐ = ๐๐, ๐๐ = ๐๐. (3.47)
Comparing (3.9) and (3.46), the plane spanned by the two orthogonal unit quater-nions ๐, ๐ โ โ is indeed the ๐โ-plane for S(๐) = cos๐ผ โ= 0. The orthogonal ๐+-planeis simply given by its basis vectors (3.9), inserting (3.45). This leads to the pair oforthogonal unit quaternions ๐, ๐ for the ๐+-plane as
๐ =๐ โ ๐
โฃ๐ โ ๐โฃ =๐๐โ ๐๐
โฃ(๐โ ๐)๐โฃ =๐โ ๐
โฃ๐โ ๐โฃ๐ = ๐(๐)๐, (3.48)
๐ =1 + ๐๐
โฃ1 + ๐๐โฃ =๐ โ ๐
โฃ๐ โ ๐โฃ๐ = ๐๐ = ๐(๐)๐๐ = ๐(๐)๐๐๐ = ๐(๐)๐๐2 = โ๐(๐)๐, (3.49)
where we have used (3.13) for the second, and (3.47) for the sixth equality in(3.49).
Let us also verify that ๐, ๐ of (3.45) are both pure unit quaternions using(3.47)
๐2 = ๐๐๐๐ = (๐๐)๐๐ = โ1, ๐2 = ๐๐๐๐ = (๐๐)๐๐ = โ1. (3.50)
Note, that if we would set in (3.45) for ๐ := โ๐๐, then the ๐, ๐-plane wouldhave become the ๐+-plane instead of the ๐โ-plane. We can therefore determine bythe sign in the definition of ๐, which of the two OPS planes the ๐, ๐-plane is torepresent.
For both ๐ and ๐ being two pure orthogonal quaternions, we can again set
๐ := ๐๐โ ๐2 = ๐๐๐๐ = โ๐2๐2 = โ1, ๐ := ๐๐ = โ๐๐ = โ๐, (3.51)
where due to the orthogonality of the pure unit quaternions ๐, ๐ we were able touse ๐๐ = โ๐๐. In this case ๐ = ๐๐ is thus also shown to be a pure unit quaternion.Now the ๐โ-plane of the corresponding OPS (3.34) is spanned by {๐, ๐}, whereasthe ๐+-plane is spanned by {1, ๐}. Setting instead ๐ := โ๐๐ = ๐๐ = ๐ , the ๐โ-plane of the corresponding OPS (3.24) is spanned by {1, ๐}, wheras the ๐+-planeis spanned by {๐, ๐}.
We summarize our results in the following theorem.
Theorem 3.5. (Determination of ๐, ๐ from given steerable 2D plane) Given any2D plane in โ in terms of two unit quaternions ๐, ๐, where ๐ is without restriction
28 E. Hitzer and S.J. Sangwine
of generality pure, i.e., ๐2 = โ1, we can make the given plane the ๐โ-plane of theOPS ๐ยฑ = 1
2 (๐ ยฑ ๐๐๐), by setting
๐ := ๐๐, ๐ := ๐๐. (3.52)
For S(๐) โ= 0 the orthogonal ๐+-plane is fully determined by the orthogonal unitquaternions
๐ = ๐(๐)๐, ๐ = โ๐(๐)๐. (3.53)
where ๐(๐) is as defined in (2.6). For S(๐) = 0 the orthogonal ๐+-plane with basis{1, ๐} is instead fully determined by ๐ = โ๐ = ๐๐.
Setting alternatively
๐ := ๐๐, ๐ := โ๐๐. (3.54)
makes the given ๐, ๐-plane the ๐+-plane instead. For S(๐) โ= 0 the orthogonal ๐โ-plane is then fully determined by (3.54) and (3.9), with the same orthogonal unitquaternions ๐ = ๐(๐)๐, ๐ = โ๐(๐)๐ as in (3.53). For S(๐) = 0 the orthogonal๐โ-plane with basis {1, ๐} is then instead fully determined by ๐ = ๐ = ๐๐.
An illustration of the decomposition is given in Figures 3 and 4. Again, threehundred unit pure quaternions randomly oriented in 4-space have been decom-posed into two sets using the decomposition of Definition 3.2 and two unit purequaternions ๐ and ๐ computed as in Theorem 3.5. ๐ was the pure unit quaternion1โ3(๐ + ๐ + ๐) and ๐ was the full unit quaternion 1โ
2+ 1
2 (๐ โ ๐). ๐ and ๐ were
computed by (3.53) as ๐ = ๐(๐)๐ and ๐ = โ๐(๐)๐.Figure 3 shows the three hundred points in 4-space, projected onto the six
orthogonal planes {๐, ๐}, {๐, ๐}, {๐, ๐}, {๐, ๐}, {๐, ๐}, {๐, ๐} where the orthonormal4-space basis {๐, ๐, ๐, ๐} = {(๐ โ ๐)/โฃ๐ โ ๐โฃ, (1 + ๐๐)/โฃ1 + ๐๐โฃ, (๐ + ๐)/โฃ๐ + ๐โฃ, (1โ๐๐)/โฃ1โ ๐๐โฃ}. The six views at the top show the ๐+-plane, and the six below showthe ๐โ-plane. Figure 4 shows the vector parts of the decomposed quaternions.
4. New QFT Forms: OPS-QFTs with Two PureUnit Quaternions ๐, ๐
4.1. Generalized OPS Leads to New Steerable Type of QFT
We begin with a straightforward generalization of the (double-sided form of the)QFT [4, 5] in โ by replacing ๐ with ๐ and ๐ with ๐ defined as
Definition 4.1. (QFT with respect to two pure unit quaternions ๐, ๐)Let ๐, ๐ โ โ, ๐2 = ๐2 = โ1, be any two pure unit quaternions. The quaternionFourier transform with respect to ๐, ๐ is
โฑ๐,๐{โ}(๐) =โซโ2
๐โ๐๐ฅ1๐1โ(๐) ๐โ๐๐ฅ2๐2๐2๐, (4.1)
where โ โ ๐ฟ1(โ2,โ), ๐2๐ = ๐๐ฅ1๐๐ฅ2 and ๐,๐ โ โ2.
2. Orthogonal 2D Planes Split 29
jยด
0.5 0 0.5
0.5
0
0.5
a
b
0.5 0 0.5
0.5
0
0.5
ac
0.5 0 0.5
0.5
0
0.5
a
d
0.5 0 0.5
0.5
0
0.5
b
c
0.5 0 0.5
0.5
0
0.5
c
d
0.5 0 0.5
0.5
0
0.5
db
0.5 0 0.5
0.5
0
0.5
a
b
0.5 0 0.5
0.5
0
0.5
a
c
0.5 0 0.5
0.5
0
0.5
a
d
0.5 0 0.5
0.5
0
0.5
b
c
0.5 0 0.5
0.5
0
0.5
c
d
0.5 0 0.5
0.5
0
0.5
d
b
Figure 3. 4D scatter plot of quaternions decomposed using the orthog-onal planes split of Definition 3.2.
30 E. Hitzer and S.J. Sangwine
0.6 0.4 0.2 0 0.2 0.4 0.6
0.5
0
0.5
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
c
b
d
Figure 4. Scatter plot of vector parts of quaternions decomposed usingthe orthogonal planes split of Definition 3.2.
Note, that the pure unit quaternions ๐, ๐ in Definition 4.1 do not need to beorthogonal, and that the cases ๐ = ยฑ๐ are fully included.
Linearity of the integral (4.1) allows us to use the OPS split โ = โโ + โ+
โฑ๐,๐{โ}(๐) = โฑ๐,๐{โโ}(๐) + โฑ๐,๐{โ+}(๐)= โฑ๐,๐
โ {โ}(๐) + โฑ๐,๐+ {โ}(๐),
(4.2)
since by their construction the operators of the Fourier transformation โฑ๐,๐, andof the OPS with respect to ๐, ๐ commute. From Lemma 3.4 follows
Theorem 4.2 (QFT of โยฑ). The QFT of the โยฑ OPS split parts, with respect totwo unit quaternions ๐, ๐, of a quaternion module function โ โ ๐ฟ1(โ2,โ) have thequasi-complex forms
โฑ๐,๐ยฑ {โ} = โฑ๐,๐{โยฑ} =
โซโ2
โยฑ๐โ๐(๐ฅ2๐2โ๐ฅ1๐1)๐2๐ฅ
=
โซโ2
๐โ๐(๐ฅ1๐1โ๐ฅ2๐2)โยฑ๐2๐ฅ .
(4.3)
Remark 4.3. The quasi-complex forms in Theorem 4.2 allow us to establish dis-cretized and fast versions of the QFT of Definition 4.1 as sums of two complexdiscretized and fast Fourier transformations (FFT), respectively.
2. Orthogonal 2D Planes Split 31
โโ
+
โ
โ
+
+
โ
Figure 5. Geometric interpretation of integrand of QFT๐,๐ in Defini-tion 4.1 in terms of local phase rotations in ๐ยฑ-planes.
We can now give a geometric interpretation of the integrand of the QFT๐,๐ inDefinition 4.1 in terms of local phase rotations, compare Section 3.3. The integrandproduct
๐โ๐๐ฅ1๐1โ(๐) ๐โ๐๐ฅ2๐2 (4.4)
represents a local rotation by the phase angle โ(๐ฅ1๐1 + ๐ฅ2๐2) in the ๐โ-plane,and by the phase angle โ(๐ฅ1๐1 โ ๐ฅ2๐2) = ๐ฅ2๐2 โ ๐ฅ1๐1 in the orthogonal ๐+-plane, compare Figure 5, which depicts two completely orthogonal planes in fourdimensions.
Based on Theorem 3.5 the two phase rotation planes (analysis planes) canbe freely steered by defining the two pure unit quaternions ๐, ๐ used in Definition4.1 according to (3.52) or (3.54).
4.2. Two Phase Angle Version of QFT
The above newly gained geometric understanding motivates us to propose a furthernew version of the QFT๐,๐, with a straightforward two phase angle interpretation.
Definition 4.4. (Phase angle QFT with respect to ๐, ๐)Let ๐, ๐ โ โ, ๐2 = ๐2 = โ1, be any two pure unit quaternions. The phase anglequaternion Fourier transform with respect to ๐, ๐ is
โฑ๐,๐๐ท {โ}(๐) =
โซโ2
๐โ๐12 (๐ฅ1๐1+๐ฅ2๐2)โ(๐) ๐โ๐
12 (๐ฅ1๐1โ๐ฅ2๐2)๐2๐. (4.5)
where again โ โ ๐ฟ1(โ2,โ), ๐2๐ = ๐๐ฅ1๐๐ฅ2 and ๐,๐ โ โ2.
The geometric interpretation of the integrand of (4.5) is a local phase rotationby angle โ(๐ฅ1๐1 + ๐ฅ2๐2)/2 โ (๐ฅ1๐1 โ ๐ฅ2๐2)/2 = โ๐ฅ1๐1 in the ๐โ-plane, and asecond local phase rotation by angle โ(๐ฅ1๐1+๐ฅ2๐2)/2+(๐ฅ1๐1โ๐ฅ2๐2)/2 = โ๐ฅ2๐2
in the ๐+-plane, compare Section 3.3.If we apply the OPS๐,๐ split to (4.5) we obtain the following theorem.
32 E. Hitzer and S.J. Sangwine
Theorem 4.5 (Phase angle QFT of ๐ยฑ). The phase angle QFT of Definition 4.4applied to the โยฑ OPS split parts, with respect to two pure unit quaternions ๐, ๐, ofa quaternion module function โ โ ๐ฟ1(โ2,โ) leads to the quasi-complex expressions
โฑ๐,๐๐ท+{โ} = โฑ๐,๐
๐ท {โ+} =โซโ2
โ+๐+๐๐ฅ2๐2๐2๐ฅ =
โซโ2
๐โ๐๐ฅ2๐2โ+๐2๐ฅ , (4.6)
โฑ๐,๐๐ทโ{โ} = โฑ๐,๐
๐ท {โโ} =โซโ2
โโ๐โ๐๐ฅ1๐1๐2๐ฅ =
โซโ2
๐โ๐๐ฅ1๐1โโ๐2๐ฅ. (4.7)
Note that based on Theorem 3.5 the two phase rotation planes (analysisplanes) are again freely steerable.
Theorem 4.5 allows us to establish discretized and fast versions of the phaseangle QFT of Definition 4.4 as sums of two complex discretized and fast Fouriertransformations (FFT), respectively.
The maps ๐( )๐ considered so far did not involve quaternion conjugation ๐ โ๐. In the following we investigate maps which additionally conjugate the argument,
i.e., of type ๐( )๐, which are also involutions.
5. Involutions and QFTs Involving Quaternion Conjugation
5.1. Involutions Involving Quaternion Conjugations
The simplest case is quaternion conjugation itself
๐ โ ๐ = ๐๐ โ ๐๐๐โ ๐๐๐ โ ๐๐๐, (5.1)
which can be interpreted as a reflection at the real line ๐๐. The real line through theorigin remains pointwise invariant, while every other point in the 3D hyperplaneof pure quaternions is reflected to the opposite side of the real line. The relatedinvolution
๐ โ โ๐ = โ๐๐ + ๐๐๐+ ๐๐๐ + ๐๐๐, (5.2)
is the reflection at the 3D hyperplane of pure quaternions (which stay invariant),i.e., only the real line is changed into its negative ๐๐ โ โ๐๐.
Similarly any pure unit quaternion factor like ๐ in the map
๐ โ ๐ ๐๐ = โ๐๐ + ๐๐๐โ ๐๐๐ โ ๐๐๐, (5.3)
leads to a reflection at the (pointwise invariant) line through the origin with di-rection ๐, while the map
๐ โ โ๐ ๐๐ = ๐๐ โ ๐๐๐+ ๐๐๐ + ๐๐๐, (5.4)
leads to a reflection at the invariant 3D hyperplane orthogonal to the line throughthe origin with direction ๐. The map
๐ โ ๐ ๐๐, (5.5)
leads to a reflection at the (pointwise invariant) line with direction ๐ through theorigin, while the map
๐ โ โ๐ ๐๐, (5.6)
2. Orthogonal 2D Planes Split 33
leads to a reflection at the invariant 3D hyperplane orthogonal to the line withdirection ๐ through the origin.
Next we turn to a map of the type
๐ โ โ๐๐ผ๐๐๐๐ผ๐ . (5.7)
Its set of pointwise invariants is given by
๐ = โ๐๐ผ๐๐๐๐ผ๐ โ ๐โ๐ผ๐๐ = โ๐๐๐ผ๐ โ ๐โ๐ผ๐๐ + ๐๐๐ผ๐ = 0
โ S(๐๐๐ผ๐
)= 0 โ ๐ โฅ ๐๐ผ๐ .
(5.8)
We further observe that
๐๐ผ๐ โ โ๐๐ผ๐๐โ๐ผ๐๐๐ผ๐ = โ๐๐ผ๐ . (5.9)
The map โ๐( )๐, with unit quaternion ๐ = ๐๐ผ๐ , therefore represents a reflectionat the invariant 3D hyperplane orthogonal to the line through the origin withdirection ๐.
Similarly, the map ๐( )๐, with unit quaternion ๐ = ๐๐ผ๐ , then represents areflection at the (pointwise invariant) line with direction ๐ through the origin.
The combination of two such reflections (both at 3D hyperplanes, or both atlines), given by unit quaternions ๐, ๐, leads to a rotation
โ๐โ๐๐๐๐ = ๐๐๐๐๐ = ๐๐๐๐๐ = ๐๐๐ ,
๐ = ๐๐, ๐ = ๐๐, โฃ๐โฃ = โฃ๐โฃ โฃ๐โฃ = 1 = โฃ๐ โฃ , (5.10)
in two orthogonal planes, exactly as studied in Section 3.3.
The combination of three reflections at 3D hyperplanes, given by unit quater-nions ๐, ๐, ๐, leads to
โ๐[โ๐โ๐๐๐๐]๐ = ๐ ๐๐ก, ๐ = โ๐๐๐, ๐ก = ๐๐๐, โฃ๐โฃ = โฃ๐โฃ โฃ๐โฃ โฃ๐โฃ = โฃ๐กโฃ = 1. (5.11)
The product of the reflection map โ๐ of (5.2) with ๐ ๐๐ก leads to โ๐๐๐ก, a double
rotation as studied in Section 4. Therefore ๐ ( ) ๐ก represents a rotary reflection(rotation reflection). The three reflections โ๐๐๐, โ๐๐๐, โ๐๐๐ have the intersectionof the three 3D hyperplanes as a resulting common pointwise invariant line, whichis ๐+ ๐ก, because
๐ (๐+ ๐ก) ๐ก = ๐ ๐ก๐ก+ ๐ ๐๐ก = ๐+ ๐ก. (5.12)
In the remaining 3D hyperplane, orthogonal to the pointwise invariant line throughthe origin in direction ๐+ ๐ก, the axis of the rotary reflection is
๐ (๐โ ๐ก) ๐ก = โ๐ ๐ก๐ก+ ๐ ๐๐ก = โ๐+ ๐ก = โ(๐โ ๐ก). (5.13)
We now also understand that a sign change of ๐ โ โ๐ (compare three reflec-
tions at three 3D hyperplanes โ๐[โ๐(โ๐๐๐)๐]๐ with three reflections at three lines
+๐[+๐(+๐๐๐)๐]๐) simply exchanges the roles of pointwise invariant line ๐ + ๐ก androtary reflection axis ๐โ ๐ก.
34 E. Hitzer and S.J. Sangwine
Next, we seek for an explicit description of the rotation plane of the rotary
reflection ๐ ( ) ๐ก. We find that for the unit quaternions ๐ = ๐๐ผ๐, ๐ก = ๐๐ฝ๐ the com-mutator
[๐, ๐ก] = ๐๐กโ ๐ก๐ = ๐๐ผ๐๐๐ฝ๐ โ ๐๐ฝ๐๐๐ผ๐ = (๐๐ โ ๐๐) sin๐ผ sin๐ฝ, (5.14)
is a pure quaternion, because
๐๐ โ ๐๐ = ๐๐ โ ๐๐ = โ(๐๐ โ ๐๐). (5.15)
Moreover, [๐, ๐ก] is orthogonal to ๐ and ๐ก, and therefore orthogonal to the planespanned by the pointwise invariant line ๐+ ๐ก and the rotary reflection axis ๐โ ๐ก,because
S([๐, ๐ก]๐
)= S(๐๐ก๐โ ๐ก๐ ๐
)= 0, S
([๐, ๐ก]๐ก
)= 0. (5.16)
We obtain a second quaternion in the plane orthogonal to ๐ + ๐ก, and ๐ โ ๐ก, byapplying the rotary reflection to [๐, ๐ก]
๐ [๐, ๐ก] ๐ก = โ๐[๐, ๐ก]๐ก = โ[๐, ๐ก]๐๐ก, (5.17)
because ๐ is orthogonal to the pure quaternion [๐, ๐ก]. We can construct an or-
thogonal basis of the plane of the rotary reflection ๐ ( ) ๐ก by computing the pair oforthogonal quaternions
๐ฃ1,2 = [๐, ๐ก]โ ๐ [๐, ๐ก] ๐ก = [๐, ๐ก]ยฑ [๐, ๐ก]๐๐ก = [๐, ๐ก](1 ยฑ ๐๐ก). (5.18)
For finally computing the rotation angle, we need to know the relative length ofthe two orthogonal quaternions ๐ฃ1, ๐ฃ2 of (5.18). For this it helps to represent theunit quaternion ๐๐ก as
๐๐ก = ๐๐พ๐ข, ๐พ โ โ, ๐ข โ โ, ๐ข2 = โ1. (5.19)
We then obtain for the length ratio
๐2 =โฃ๐ฃ1โฃ2โฃ๐ฃ2โฃ2 =
โฃ1 + ๐๐กโฃ2โฃ1โ ๐๐กโฃ2 =
(1 + ๐๐ก)(1 + ๐ก๐)
(1 โ ๐๐ก)(1โ ๐ก๐)=
1 + ๐๐ก๐ก๐+ ๐๐ก+ ๐ก๐
1 + ๐๐ก๐ก๐โ ๐๐กโ ๐ก๐
=2 + 2 cos ๐พ
2โ 2 cos ๐พ=
1 + cos ๐พ
1โ cos ๐พ.
(5.20)
By applying the rotary reflection ๐ ( ) ๐ก to ๐ฃ1 and decomposing the result withrespect to the pair of orthogonal quaternions in the rotary reflection plane (5.18)we can compute the rotation angle. Applying the rotary reflection to ๐ฃ1 gives
๐ ๐ฃ1 ๐ก = ๐ [๐, ๐ก]โ ๐[๐, ๐ก]๐ก ๐ก = ๐[๐, ๐ก](1 + ๐๐ก)๐ก = ๐(1 + ๐ก๐)[๐, ๐ก]๐ก
= ๐(1 + ๐ก๐)(โ[๐, ๐ก])๐ก = โ[๐, ๐ก](๐๐ก+ ๐๐ก๐๐ก).(5.21)
The square of ๐๐ก is
(๐๐ก)2 = (cos ๐พ + ๐ข sin ๐พ)2 = โ1 + 2 cos ๐พ [cos ๐พ + ๐ข sin ๐พ]
= โ1 + 2 cos ๐พ ๐๐ก.(5.22)
2. Orthogonal 2D Planes Split 35
We therefore get
๐ ๐ฃ1 ๐ก = โ[๐, ๐ก](๐๐กโ 1 + 2 cos๐พ๐๐ก) = [๐, ๐ก](1โ (1 + 2 cos๐พ)๐๐ก)
= ๐1๐ฃ1 + ๐2๐๐ฃ2,(5.23)
and need to solve
1โ (1 + 2 cos ๐พ)๐๐ก = ๐1(1 + ๐๐ก) + ๐2๐(1 โ ๐๐ก), (5.24)
which leads to
๐ ๐ฃ1 ๐ก = โ cos๐พ๐ฃ1 + sin ๐พ๐๐ฃ2 = cos(๐ โ ๐พ)๐ฃ1 + sin(๐ โ ๐พ)๐๐ฃ2. (5.25)
The rotation angle of the rotary reflection ๐ ( ) ๐ก in its rotation plane ๐ฃ1, ๐ฃ2 istherefore
ฮ = ๐ โ ๐พ, ๐พ = arccosS(๐๐ก). (5.26)
In terms of ๐ = ๐๐ผ๐, ๐ก = ๐๐ฝ๐ we get
๐๐ก = cos๐ผ cos๐ฝ โ ๐ sin๐ผ cos๐ฝ + ๐ cos๐ผ sin๐ฝ โ ๐๐ sin๐ผ sin๐ฝ. (5.27)
And with the angle ๐ between ๐ and ๐
๐๐ =1
2(๐๐ + ๐๐) +
1
2(๐๐ โ ๐๐) = S(๐๐) +
1
2[๐, ๐ ]
= โ cos๐ โ sin๐[๐, ๐ ]
โฃ[๐, ๐ ]โฃ ,(5.28)
we finally obtain for ๐พ the scalar part S(๐๐ก)as
S(๐๐ก)= cos ๐พ = cos๐ผ cos๐ฝ + cos๐ sin๐ผ sin๐ฝ
= cos๐ผ cos๐ฝ โ S(๐๐) sin๐ผ sin๐ฝ.(5.29)
In the special case of ๐ = ยฑ๐ , S(๐๐) = โ1, i.e., for ๐ = 0, ๐, we get from(5.29) that
S(๐๐ก)= cos๐ผ cos๐ฝ ยฑ sin๐ผ sin๐ฝ = cos๐ผ cos๐ฝ + sin(ยฑ๐ผ) sin๐ฝ
= cos(ยฑ๐ผโ ๐ฝ), (5.30)
and thus using (5.26) the rotation angle would become
ฮ = ๐ โ (ยฑ๐ผโ ๐ฝ) = ๐ โ ๐ผ+ ๐ฝ. (5.31)
Yet (5.26) was derived assuming [๐, ๐ก] โ= 0. But direct inspection shows that (5.31)is indeed correct: For ๐ = ยฑ๐ the plane ๐ + ๐ก, ๐ โ ๐ก is identical to the 1, ๐ plane.The rotation plane is thus a plane of pure quaternions orthogonal to the 1, ๐ plane.The quaternion conjugation in ๐ ๏ฟฝโ ๐ ๐ ๐ก leads to a rotation by ๐ and the left andright factors lead to further rotations by โ๐ผ and ๐ฝ, respectively. Thus (5.31) isverified as a special case of (5.26) for ๐ = ยฑ๐ .
By substituting in Lemma 3.4 (๐ผ, ๐ฝ)โ (โ๐ฝ,โ๐ผ), and by taking the quater-nion conjugate we obtain the following lemma.
36 E. Hitzer and S.J. Sangwine
Lemma 5.1. Let ๐ยฑ = 12 (๐ ยฑ ๐๐๐) be the OPS of Definition 3.2. For left and right
exponential factors we have the identity
๐๐ผ๐ ๐ยฑ๐๐ฝ๐ = ๐ยฑ๐(๐ฝโ๐ผ)๐ = ๐(๐ผโ๐ฝ)๐ ๐ยฑ. (5.32)
5.2. New Steerable QFTs with Quaternion Conjugation andTwo Pure Unit Quaternions ๐, ๐
We therefore consider now the following new variant of the (double-sided form ofthe) QFT [4, 5] in โ (replacing both ๐ with ๐ and ๐ with ๐ , and using quaternionconjugation). It is essentially the quaternion conjugate of the new QFT of Def-inition 4.1, but because of its distinct local transformation geometry it deservesseparate treatment.
Definition 5.2. (QFT with respect to ๐, ๐, including quaternion conjugation)Let ๐, ๐ โ โ, ๐2 = ๐2 = โ1, be any two pure unit quaternions. The quaternionFourier transform with respect to ๐, ๐, involving quaternion conjugation, is
โฑ๐,๐๐ {โ}(๐) = โฑ๐,๐{โ}(โ๐) =
โซโ2
๐โ๐๐ฅ1๐1โ(๐) ๐โ๐๐ฅ2๐2๐2๐, (5.33)
where โ โ ๐ฟ1(โ2,โ), ๐2๐ = ๐๐ฅ1๐๐ฅ2 and ๐,๐ โ โ2.
Linearity of the integral in (5.33) of Definition 5.2 leads to the followingcorollary to Theorem 4.2.
Corollary 5.3 (QFT โฑ๐,๐๐ of โยฑ). The QFT โฑ๐,๐
๐ (5.33) of the โยฑ = 12 (โ ยฑ ๐โ๐)
OPS split parts, with respect to any two unit quaternions ๐, ๐, of a quaternionmodule function โ โ ๐ฟ1(โ2,โ) have the quasi-complex forms
โฑ๐,๐๐ {โยฑ}(๐) = โฑ๐,๐{โยฑ}(โ๐) =
โซโ2
โยฑ๐โ๐(๐ฅ2๐2โ๐ฅ1๐1)๐2๐ฅ
=
โซโ2
๐โ๐(๐ฅ1๐1โ๐ฅ2๐2)โยฑ๐2๐ฅ .
(5.34)
Note, that the pure unit quaternions ๐, ๐ in Definition 5.2 and Corollary5.3 do not need to be orthogonal, and that the cases ๐ = ยฑ๐ are fully included.Corollary 5.3 leads to discretized and fast versions of the QFT with quaternionconjugation of Definition 5.2.
It is important to note that the roles (sides) of ๐, ๐ appear exchanged in(5.33) of Definition 5.2 and in Corollary 5.3, although the same OPS of Definition3.2 is applied to the signal โ as in Sections 3 and 4. This role change is due tothe presence of quaternion conjugation in Definition 5.2. Note that it is possibleto first apply (5.33) to โ, and subsequently split the integral with the OPS๐,๐
โฑ๐,ยฑ = 12 (โฑ๐ยฑ ๐โฑ๐๐), where the particular order of ๐ from the left and ๐ from the
right is due to the application of conjugation in (5.34) to โยฑ after โ is split with(3.41) into โ+ and โโ.
2. Orthogonal 2D Planes Split 37
5.3. Local Geometric Interpretation of the QFT with Quaternion Conjugation
Regarding the local geometric interpretation of the QFT with quaternion conju-gation of Definition 5.2 we need to distinguish the following cases, depending on[๐, ๐ก] and on whether the left and right phase factors
๐ = ๐โ๐๐ฅ1๐1 , ๐ก = ๐โ๐๐ฅ2๐2 , (5.35)
attain scalar values ยฑ1 or not.Let us first assume that [๐, ๐ก] โ= 0, which by (5.14) is equivalent to ๐ โ= ยฑ๐ ,
and sin(๐ฅ1๐1) โ= 0, and sin(๐ฅ2๐2) โ= 0. Then we have the generic case of a localrotary reflection with pointwise invariant line of direction
๐+ ๐ก = ๐โ๐๐ฅ1๐1 + ๐โ๐๐ฅ2๐2 , (5.36)
rotation axis in direction
๐โ ๐ก = ๐โ๐๐ฅ1๐1 โ ๐โ๐๐ฅ2๐2 , (5.37)
rotation plane with basis (5.18), and by (5.26) and (5.29) the general rotationangle
ฮ = ๐ โ arccosS(๐๐ก),
S(๐๐ก)= cos(๐ฅ1๐1) cos(๐ฅ2๐2)โ S(๐๐) sin(๐ฅ1๐1) sin(๐ฅ2๐2). (5.38)
Whenever ๐ = ยฑ๐ , or when sin(๐ฅ1๐1) = 0 (๐ฅ1๐1 = 0, ๐[mod 2๐], i.e., ๐ =ยฑ1), we get for the pointwise invariant line in direction ๐ + ๐ก the simpler unit
quaternion direction expression ๐โ12 (ยฑ๐ฅ1๐1+๐ฅ2๐2)๐ , because we can apply
๐๐ผ๐ + ๐๐ฝ๐ = ๐12 (๐ผ+๐ฝ)๐(๐
12 (๐ผโ๐ฝ)๐ + ๐
12 (๐ฝโ๐ผ)๐ ) = ๐
12 (๐ผ+๐ฝ)๐2 cos
๐ผโ ๐ฝ
2, (5.39)
and similarly for the rotation axis ๐โ ๐ก we obtain the direction expression
๐โ12 (ยฑ๐ฅ1๐1+๐ฅ2๐2+๐)๐ ,
whereas the rotation angle is by (5.31) simply
ฮ = ๐ ยฑ ๐ฅ1๐1 โ ๐ฅ2๐2. (5.40)
For sin(๐ฅ2๐2) = 0 (๐ฅ2๐2 = 0, ๐[mod2๐], i.e., ๐ก = ยฑ1), the pointwise invariantline in direction ๐+ ๐ก simplifies by (5.39) to ๐โ
12 (๐ฅ1๐1+๐ฅ2๐2)๐, and the rotation axis
with direction ๐โ ๐ก simplifies to ๐โ12 (๐ฅ1๐1+๐ฅ2๐2+๐)๐, whereas the angle of rotation
is by (5.31) simplyฮ = ๐ + ๐ฅ1๐1 โ ๐ฅ2๐2. (5.41)
5.4. Phase Angle QFT with Respect to ๐, ๐, Including Quaternion Conjugation
Even when quaternion conjugation is applied to the signal โ we can propose afurther new version of the QFT๐,๐
๐ , with a straightforward two phase angle inter-pretation. The following definition to some degree ignores the resulting local rotaryreflection effect of combining quaternion conjugation and left and right phase fac-tors of Section 5.3, but depending on the application context, it may neverthelessbe of interest in its own right.
38 E. Hitzer and S.J. Sangwine
Definition 5.4 (Phase angle QFTwith respect to ๐, ๐, including quaternion con-jugation). Let ๐, ๐ โ โ, ๐2 = ๐2 = โ1, be any two pure unit quaternions. Thephase angle quaternion Fourier transform with respect to ๐, ๐, involving quaternionconjugation, is
โฑ๐,๐๐๐ท {โ}(๐) = โฑ๐,๐
๐ท {โ}(โ๐1, ๐2) =
โซโ2
๐โ๐12 (๐ฅ1๐1+๐ฅ2๐2)โ(๐) ๐โ๐
12 (๐ฅ1๐1โ๐ฅ2๐2)๐2๐.
(5.42)where again โ โ ๐ฟ1(โ2,โ), ๐2๐ = ๐๐ฅ1๐๐ฅ2 and ๐,๐ โ โ2.
Based on Lemma 5.1, one possible geometric interpretation of the integrandof (5.42)) is a local phase rotation of โ+ by angle โ(๐ฅ1๐1 โ ๐ฅ2๐2)/2 + (๐ฅ1๐1 +
๐ฅ2๐2)/2 = +๐ฅ2๐2 in the ๐+ plane, and a second local phase rotation of โโ by angleโ(๐ฅ1๐1 โ ๐ฅ2๐2)/2โ (๐ฅ1๐1 + ๐ฅ2๐2)/2 = โ๐ฅ1๐1 in the ๐โ plane. This is expressedin the following corollary to Theorem 4.5.
Corollary 5.5 (Phase angle QFT of ๐ยฑ, involving quaternion conjugation). Thephase angle QFT with quaternion conjugation of Definition 5.4 applied to the โยฑOPS split parts, with respect to any two pure unit quaternions ๐, ๐, of a quaternionmodule function โ โ ๐ฟ1(โ2,โ) leads to the quasi-complex expressions
โฑ๐,๐๐๐ท {โ+}(๐) = โฑ๐,๐
๐ท {โ+}(โ๐1, ๐2) =
โซโ2
โ+๐+๐๐ฅ2๐2๐2๐ฅ =
โซโ2
๐โ๐๐ฅ2๐2โ+๐2๐ฅ ,
(5.43)
โฑ๐,๐๐๐ท {โโ}(๐) = โฑ๐,๐
๐ท {โโ}(โ๐1, ๐2) =
โซโ2
โโ๐โ๐๐ฅ1๐1๐2๐ฅ =
โซโ2
๐โ๐๐ฅ1๐1โโ๐2๐ฅ .
(5.44)
Note that based on Theorem 3.5 the two phase rotation planes (analysisplanes) are again freely steerable. Corollary 5.5 leads to discretized and fast ver-sions of the phase angle QFT with quaternion conjugation of Definition 5.4.
6. Conclusion
The involution maps ๐( )๐ and ๐( )๐ have led us to explore a range of similar quater-nionic maps ๐ ๏ฟฝโ ๐๐๐ and ๐ ๏ฟฝโ ๐๐๐, where ๐, ๐ are taken to be unit quaternions.Geometric interpretations of these maps as reflections, rotations, and rotary reflec-tions in 4D can mostly be found in [1]. We have further developed these geometricinterpretations to gain a complete local transformation geometric understandingof the integrands of the proposed new quaternion Fourier transformations (QFTs)applied to general quaternionic signals โ โ ๐ฟ1(โ2,โ). This new geometric under-standing is also valid for the special cases of the hitherto well-known left-sided,right-sided, and left- and right-sided (two-sided) QFTs of [2, 4, 8, 3] and numerousother references.
Our newly gained geometric understanding itself motivated us to proposenew types of QFTs with specific geometric properties. The investigation of these
2. Orthogonal 2D Planes Split 39
new types of QFTs with the generalized form of the orthogonal 2D planes splitof Definition 3.2 lead to important QFT split theorems, which allow the use ofdiscrete and (complex) Fourier transform software for efficient discretized and fastnumerical implementations.
Finally, we are convinced that our geometric interpretation of old and newQFTs paves the way for new applications, e.g., regarding steerable filter design forspecific tasks in image, colour image and signal processing, etc.
Acknowledgement. E.H. wishes to thank God for the joy of doing this research,his family, and S.J.S. for his great cooperation and hospitality.
References
[1] H.S.M. Coxeter. Quaternions and reflections. The American Mathematical Monthly,53(3):136โ146, Mar. 1946.
[2] T.A. Ell. Quaternion-Fourier transforms for analysis of 2-dimensional linear time-invariant partial-differential systems. In Proceedings of the 32nd Conference on De-cision and Control, pages 1830โ1841, San Antonio, Texas, USA, 15โ17 December1993. IEEE Control Systems Society.
[3] T.A. Ell and S.J. Sangwine. Hypercomplex Fourier transforms of color images. IEEETransactions on Image Processing, 16(1):22โ35, Jan. 2007.
[4] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[5] E. Hitzer. Directional uncertainty principle for quaternion Fourier transforms. Ad-vances in Applied Clifford Algebras, 20(2):271โ284, 2010.
[6] E. Hitzer. OPS-QFTs: A new type of quaternion Fourier transform based on theorthogonal planes split with one or two general pure quaternions. In InternationalConference on Numerical Analysis and Applied Mathematics, volume 1389 of AIPConference Proceedings, pages 280โ283, Halkidiki, Greece, 19โ25 September 2011.American Institute of Physics.
[7] E. Hitzer. Two-sided Clifford Fourier transform with two square roots of โ1 in ๐ถโ๐,๐.In Proceedings of the 5th Conference on Applied Geometric Algebras in ComputerScience and Engineering (AGACSE 2012), La Rochelle, France, 2โ4 July 2012.
[8] S.J. Sangwine. Fourier transforms of colour images using quaternion, or hypercom-plex, numbers. Electronics Letters, 32(21):1979โ1980, 10 Oct. 1996.
Eckhard HitzerCollege of Liberal Arts, Department of Material ScienceInternational Christian University, 181-8585 Tokyo, Japane-mail: [email protected]
Stephen J. SangwineSchool of Computer Science and Electronic EngineeringUniversity of Essex, Wivenhoe Park, Colchester, CO4 3SQ, UKe-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 41โ56cโ 2013 Springer Basel
3 Quaternionic Spectral Analysis ofNon-Stationary Improper Complex Signals
Nicolas Le Bihan and Stephen J. Sangwine
Abstract. We consider the problem of the spectral content of a complex im-proper signal and the time-varying behaviour of this spectral content. Thesignals considered are one-dimensional (1D), complex-valued, with possiblecorrelation between the real and imaginary parts, i.e., improper complex sig-nals. As a consequence, it is well known that the โclassicalโ (complex-valued)Fourier transform does not exhibit Hermitian symmetry and also that it isnecessary to consider simultaneously the spectrum and the pseudo-spectrumto completely characterize such signals. Hence, an โaugmentedโ representa-tion is necessary. However, this does not provide a geometric analysis of thecomplex improper signal.
We propose another approach for the analysis of improper complex sig-nals based on the use of a 1D Quaternion Fourier Transform (QFT). In thecase where complex signals are non-stationary, we investigate the extension ofthe well-known โanalytic signalโ and introduce the quaternion-valued โhyper-analytic signalโ. As with the hypercomplex two-dimensional (2D) extensionof the analytic signal proposed by Bulow in 2001, our extension of analyticsignals for complex-valued signals can be obtained by the inverse QFT of thequaternion-valued spectrum after suppressing negative frequencies.
Analysis of the hyperanalytic signal reveals the time-varying frequencycontent of the corresponding complex signal. Using two different represen-tations of quaternions, we show how modulus and quaternion angles of thehyperanalytic signal are linked to geometric features of the complex signal.This allows the definition of the angular velocity and the complex envelope ofa complex signal. These concepts are illustrated on synthetic signal examples.
The hyperanalytic signal can be seen as the exact counterpart of theclassical analytic signal, and should be thought of as the very first and sim-plest quaternionic time-frequency representation for improper non-stationarycomplex-valued signals.
Mathematics Subject Classification (2010). 65T50, 11R52.
Keywords. Quaternions, complex signals, Fourier transform, analytic signal.
42 N. Le Bihan and S.J. Sangwine
1. Introduction
The analytic signal has been known since 1948 from the work of Ville [21] andGabor [7]. It can be easily described, even though its theoretical ramifications aredeep. Its use in non-stationary signal analysis is routine and it has been used innumerous applications. Simply put, given a real-valued signal ๐(๐ก), its analyticsignal ๐(๐ก) is a complex signal with real part equal to ๐(๐ก), and imaginary partorthogonal to ๐(๐ก). The imaginary part is sometimes known as the quadraturesignal โ in the case where ๐(๐ก) is a sinusoid, the imaginary part of the analyticsignal is in quadrature, that is with a phase difference of โ๐/2. The orthogonalsignal is related to ๐(๐ก) by the Hilbert transform [9, 10]. The analytic signal hasthe interesting property that its modulus โฃ๐(๐ก)โฃ is an envelope of the signal ๐(๐ก).The envelope is also known as the instantaneous amplitude. Thus if ๐(๐ก) is anamplitude-modulated sinusoid, the envelope โฃ๐(๐ก)โฃ, subject to certain conditions,is the modulating signal. The argument of the analytic signal, โ ๐(๐ก) is known asthe instantaneous phase. The analytic signal has a third interesting property: ithas a one-sided Fourier transform. Thus a simple algorithm for constructing theanalytic signal (algebraically or numerically) is to compute the Fourier transformof ๐(๐ก), multiply the Fourier transform by a unit step which is zero for negativefrequencies, and then construct the inverse Fourier transform.
In this chapter we extend the concept of the analytic signal from the case ofa real signal ๐(๐ก) with a complex analytic signal ๐(๐ก), to a complex signal ๐ง(๐ก) witha hypercomplex analytic signal โ(๐ก), which we call the hyperanalytic signal. Just asthe classical complex analytic signal contains both the original real signal (in thereal part) and a real orthogonal signal (in the imaginary part), a hyperanalyticsignal contains two complex signals: the original signal and an orthogonal signal.We have previously published partial results on this topic [19, 15, 14]. Here wedevelop a clear idea of how to generalise the classic case of amplitude modulationto a complex signal, and show for the first time that this leads to a correctlyextended analytic signal concept in which the (complex ) envelope and phase haveclear interpretations.
The construction of an orthogonal signal alone would not constitute a fullgeneralisation of the classical analytic signal to the complex case: it is also neces-sary to generalise the envelope and phase concepts, and this can only be done byinterpreting the original and orthogonal complex signals as a pair. In this chapterwe have only one way to do this: by representing the pair of complex signals asa quaternion signal. This arises naturally from the method above for creating anorthogonal signal, but also from the CayleyโDickson construction of a quaternionas a complex number with complex real and imaginary parts (with different rootsof โ1 used in each of the two levels of complex number).
Extension of the analytic signal concept to 2D signals, that is images, withreal, complex or quaternion-valued pixels is of interest, but outside the scopeof this chapter. Some work has been done on this, notably by Bulow, Felsberg,Sommer and Hahn [1, 2, 6, 8]. The principal issue to be solved in the 2D case
3. โ-spectral analysis 43
is to generalise the concept of a single-sided spectrum. Hahn considered a singlequadrant or orthant spectrum, Sommer et al. considered a spectrum with supportlimited to half the complex plane, not necessarily confined to two quadrants, butstill with real sample or pixel values.
Recently, Lilly and Olhede [16] have published a paper on bivariate analyticsignal concepts without explicitly considering the complex signal case which wecover here. Their approach is linked to a specific signal model, the modulatedelliptical signal, which they illustrate with the example of a drifting oceanographicfloat. The approach taken in the present chapter is more general and withoutreference to a specific signal model.
2. 1D Quaternion Fourier Transform
In this section, we will be concerned with the definition and properties of thequaternion Fourier transform (QFT) of complex-valued signals. Before introducingthe main definitions, we give some prerequisites on quaternion-valued signals.
2.1. Preliminary Remarks
Quaternions were discovered by Sir W.R. Hamilton in 1843 [11]. Quaternions are4D hypercomplex numbers that form a noncommutative division ring denotedโ. A quaternion ๐ โ โ has a Cartesian form: ๐ = ๐0 + ๐1๐ + ๐2๐ + ๐3๐, with๐0, ๐1, ๐2, ๐3 โ โ and ๐, ๐,๐ roots of โ1 satisfying ๐2 = ๐2 = ๐2 = ๐๐๐ = โ1.The scalar part of ๐ is: S(๐) = ๐0. The vector part of ๐ is: V(๐) = ๐ โ S(๐).Quaternion multiplication is not commutative, so that in general ๐๐ โ= ๐๐ for๐, ๐ โ โ. The conjugate of ๐ is ๐ = ๐0 โ ๐1๐ โ ๐2๐ โ ๐3๐. The norm of ๐ is
โฅ๐โฅ = โฃ๐โฃ2 = (๐20 + ๐21 + ๐22 + ๐23) = ๐๐. A quaternion with ๐0 = 0 is called pure. Ifโฃ๐โฃ = 1, then ๐ is called a unit quaternion. The inverse of ๐ is ๐โ1 = ๐/ โฅ๐โฅ. Pureunit quaternions are special quaternions, among which are ๐, ๐ and ๐. Togetherwith the identity of โ, they form a quaternion basis : {1, ๐, ๐,๐}. In fact, givenany two unit pure quaternions, ๐ and ๐, which are orthogonal to each other (i.e.,S(๐๐) = 0), then {1, ๐, ๐, ๐๐} is a quaternion basis.
Quaternions can also be viewed as complex numbers with complex compo-nents, i.e., one can write ๐ = ๐ง1 + ๐๐ง2 in the basis {1, ๐, ๐,๐} with ๐ง1, ๐ง2 โ โ๐, i.e.,๐ง๐ผ = โ(๐ง๐ผ) + ๐โ(๐ง๐ผ) for ๐ผ = 1, 21. This is called the CayleyโDickson form.
Among the possible representations of ๐, two of them will be used in thischapter: the polar (Euler) form and the polar CayleyโDickson form [20].
Polar form. Any non-zero quaternion ๐ can be written:
๐ = โฃ๐โฃ ๐๐๐๐๐ = โฃ๐โฃ (cos๐๐ + ๐๐ sin๐๐) ,
1In the sequel, we denote by โ๐ the set of complex numbers with root of โ1 = ๐. Note thatthese are degenerate quaternions, where all vector parts point in the same direction.
44 N. Le Bihan and S.J. Sangwine
where ๐๐ is the axis of ๐ and ๐๐ is the angle of ๐. For future use in this chapter,we give here their explicit expressions:{
๐๐ = V(๐) / โฃV(๐)โฃ ,๐๐ = arctan (โฃV(๐)โฃ / S(๐)) . (2.1)
The axis and angle are used in interpreting the hyperanalytic signal in Section 4.
Polar CayleyโDickson form. Any quaternion ๐ also has a unique polar CayleyโDickson form [20] given by:
๐ = ๐ด๐ exp(๐ต๐๐) = (๐0 + ๐1๐) exp((๐0 + ๐1๐)๐), (2.2)
where ๐ด๐ = ๐0 + ๐1๐ is the complex modulus of ๐ and ๐ต๐ = ๐0 + ๐1๐ its complexphase. It is proven in [20] that given a quaternion ๐ = (๐0 + ๐1๐)๐ = ๐0๐ + ๐1๐, itsexponential is given as:
๐๐ = cos โฃ๐โฃ+ ๐
โฃ๐โฃ sin โฃ๐โฃ (2.3)
= cos(โ
๐20 + ๐21) + ๐๐0โ
๐20 + ๐21sin(โ
๐20 + ๐21)
+ ๐๐1โ
๐20 + ๐21sin(โ
๐20 + ๐21). (2.4)
As a consequence, using the right-hand side expression in (2.2), the Cartesiancoordinates of ๐ can be linked to the polar CayleyโDickson ones in the followingway:โงโจโฉ
๐ด๐ =๐0 + ๐๐1
cos(โ
๐22 + ๐23)
๐ต๐ = arctan(โ
๐22 + ๐23)โ
๐22 + ๐23
(๐0๐3 + ๐1๐2๐20 + ๐21
+ ๐๐0๐3 โ ๐1๐2๐20 + ๐21
).
(2.5)
In Section 4, we will make use of the complex modulus ๐ด๐ for interpretationof the hyperanalytic signal.
2.2. 1D Quaternion Fourier Transform
In this chapter, we use a 1D version of the (right-side) QFT first defined in discrete-time form in [17]. Thus, it is necessary to specify the axis (a pure unit quaternion)of the transform. So, we will denote by โฑ๐ [ ] a QFT with transformation axis ๐.For convenience below, we refer to this as a ๐๐น๐๐. In order to work with theclassical quaternion basis, in the sequel we will use ๐ as the transform axis. Theonly restriction on the transform axis is that it must be orthogonal to the originalbasis of the signal (here {1, ๐}). We now present the definition and some propertiesof the transform used here.
3. โ-spectral analysis 45
Definition 1. Given a complex-valued signal ๐ง(๐ก) = ๐ง๐(๐ก) + ๐๐ง๐(๐ก), its quaternionFourier transform with respect to axis ๐ is:
๐๐(๐) = โฑ๐ [๐ง(๐ก)] =
+โโซโโ
๐ง(๐ก)๐โ๐2๐๐๐กd๐ก, (2.6)
and the inverse transform is:
๐ง(๐ก) = โฑโ1๐ [๐๐(๐)] =
+โโซโโ
๐๐(๐)๐๐2๐๐๐กd๐. (2.7)
Property 1. Given a complex signal ๐ง(๐ก) = ๐ง๐(๐ก)+๐๐ง๐(๐ก) and its quaternion Fouriertransform denoted ๐๐(๐), then the following properties hold:
โ The even part of ๐ง๐(๐ก) is in โ(๐๐(๐)),โ The odd part of ๐ง๐(๐ก) is in โ๐(๐๐(๐)),โ The even part of ๐ง๐(๐ก) is in โ๐(๐๐(๐)),โ The odd part of ๐ง๐(๐ก) is in โ๐(๐๐(๐)).
Proof. Expand (2.6) into real and imaginary parts with respect to ๐, and expandthe quaternion exponential into cosine and sine components:
๐๐(๐) =
+โโซโโ
[๐ง๐(๐ก) + ๐๐ง๐(๐ก)] [cos(2๐๐๐ก)โ ๐ sin(2๐๐๐ก)] d๐ก
=
+โโซโโ
๐ง๐(๐ก) cos(2๐๐๐ก)d๐กโ ๐
+โโซโโ
๐ง๐(๐ก) sin(2๐๐๐ก)d๐ก
+ ๐
+โโซโโ
๐ง๐(๐ก) cos(2๐๐๐ก)d๐กโ ๐
+โโซโโ
๐ง๐(๐ก) sin(2๐๐๐ก)d๐ก,
from which the stated properties are evident. โก
These properties are central to the justification of the use of the QFT toanalyze a complex-valued signal carrying complementary but different informationin its real and imaginary parts. Using the QFT, it is possible to have the oddand even parts of the real and imaginary parts of the signal in four differentcomponents in the transform domain. This idea was also the initial motivation ofBulow, Sommer and Felsberg when they developed the monogenic signal for images[1, 2, 6]. Note that the use of hypercomplex Fourier transforms was originallyintroduced in 2D Nuclear Magnetic Resonance image analysis [5, 3].
We now turn to the link between a complex signal and the quaternion signalthat can be uniquely associated to it.
46 N. Le Bihan and S.J. Sangwine
Property 2. A one-sided ๐๐น๐ ๐(๐), obtained from a complex signal as in Defi-nition 1 by suppressing the negative frequencies, that is with ๐(๐) = 0 for ๐ < 0,corresponds to a full quaternion-valued signal in the time domain.
Proof. The proof is based on the symmetry properties, listed in Property 1, fulfilledby the ๐๐น๐๐ of complex-valued signals. From this, it is easily verified that, giventhe ๐๐น๐๐ of a complex signal ๐ข(๐ก) denoted by ๐๐(๐), then (1 + sign(๐))๐๐(๐)is right-sided (i.e., it vanishes for all ๐ < 0). Here sign(๐) is the classical signfunction, which is equal to 1 for ๐ > 0 and equal to โ1 for ๐ < 0. Second, theinverse transform, or ๐ผ๐๐น๐๐ of (1 + sign(๐))๐๐(๐) can be decomposed as follows:
๐ผ๐๐น๐๐ [(1 + sign(๐))๐๐(๐)] = ๐ผ๐๐น๐๐ [๐๐(๐)] + ๐ผ๐๐น๐๐ [sign(๐)๐๐(๐)]
By definition, ๐ผ๐๐น๐๐ [๐๐(๐)] is a complex-valued signal with non-zero real and๐-imaginary parts, and null ๐-imaginary and ๐-imaginary parts. To complete theproof we show that ๐ผ๐๐น๐๐ [sign(๐)๐๐(๐)] has null real and ๐-imaginary parts, andnon-zero ๐-imaginary and ๐-imaginary parts. Consider the original complex signal๐ข(๐ก) = ๐ข๐(๐ก)+๐ข๐(๐ก)๐, as composed of odd and even, real and imaginary parts (fourcomponents in total). Property 1 shows how these four components map to thefour Cartesian components of ๐๐(๐), namely:
โ The even part of ๐ข๐(๐ก) ๏ฟฝโ โ(๐๐(๐)), which is even,โ The even part of ๐ข๐(๐ก) ๏ฟฝโ โ๐(๐๐(๐)), which is even,โ The odd part of ๐ข๐(๐ก) ๏ฟฝโ โ๐(๐๐(๐)), which is odd.โ The odd part of ๐ข๐(๐ก) ๏ฟฝโ โ๐(๐๐(๐)), which is odd.
Multiplication by sign(๐) changes the parity to the following:
โ sign(๐)โ(๐๐(๐)) is odd,โ sign(๐)โ๐(๐๐(๐)), is odd,โ sign(๐)โ๐(๐๐(๐)), is even.โ sign(๐)โ๐(๐๐(๐)), is even.
and the inverse transform ๐ผ๐๐น๐๐ maps these components to:
โ sign(๐)โ(๐๐(๐)) ๏ฟฝโ ๐๐ข1(๐ก),โ sign(๐)โ๐(๐๐(๐)) ๏ฟฝโ ๐๐๐ข2(๐ก) = ๐๐ข2(๐ก),โ sign(๐)โ๐(๐๐(๐)) ๏ฟฝโ ๐๐ข3(๐ก),โ sign(๐)โ๐(๐๐(๐)) ๏ฟฝโ ๐๐ข4(๐ก),
where ๐ข๐ฅ(๐ก), ๐ฅ = 1, 2, 3, 4 are real functions of ๐ก. Hence ๐ผ๐๐น๐๐ [sign(๐)๐๐(๐)] hasnull real and null ๐-imaginary parts, but non-zero ๐-imaginary and ๐-imaginaryparts as previously stated. โก
Property 3. Given a complex signal ๐ฅ(๐ก), one can associate to it a unique canonicalpair corresponding to a modulus and phase. These modulus and phase are uniquelydefined through the hyperanalytic signal, which is quaternion valued.
Proof. Cancelling the negative frequencies of the QFT leads to a quaternion signalin the time domain. Then, any quaternion signal has a modulus and phase definedusing its CayleyโDickson polar form. โก
3. โ-spectral analysis 47
2.3. Convolution
We consider the special case of convolution of a complex signal by a real signal.Consider ๐ and ๐ such that: ๐ : โ+ โ โ and ๐ : โ+ โ โ. Now, consider theQFT๐ of their convolution:
โฑ๐ [(๐ โ ๐)(๐ก)] =+โโซโโ
+โโซโโ
๐(๐)๐(๐ก โ ๐)d๐๐โ2๐๐๐๐กd๐ก
=
+โโซโโ
+โโซโโ
๐(๐)๐โ๐2๐๐(๐กโฒ+๐)๐(๐กโฒ)๐๐d๐กโฒ
=
+โโซโโ
๐(๐)๐โ๐2๐๐๐d๐
+โโซโโ
๐(๐กโฒ)๐โ๐2๐๐๐กโฒd๐กโฒ
= โฑ๐ [๐(๐ก)]โฑ๐ [๐(๐ก)] .
(2.8)
Thus, the definition used for the QFT here verifies the convolution theorem in theconsidered case. This specific case will be of use in our definition of the hyperan-alytic signal.
2.4. The Quaternion Fourier Transform of the Hilbert Transform
It is straightforward to verify that the quaternion Fourier transform of a real signal๐ฅ(๐ก) = 1/๐๐ก is โ๐ sign(๐), where ๐ is the axis of the transform. Substituting ๐ฅ(๐ก)into (2.6), we get:
โฑ๐
[1
๐๐ก
]=
1
๐
+โโซโโ
๐โ๐2๐๐๐ก
๐กd๐ก,
and this is clearly isomorphic to the classical complex case. The solution in theclassical case is โ๐ sign(๐), and hence in the quaternion case must be as statedabove.
It is also straightforward to see that, given an arbitrary real signal ๐ฆ(๐ก),subject only to the constraint that its classical Hilbert transform โ [๐ฆ(๐ก)] exists,then one can easily show that the classical Hilbert transform of the signal may beobtained using a quaternion Fourier transform as follows:
โ [๐ฆ(๐ก)] = โฑโ1๐ [โ๐ sign(๐)๐๐(๐)] (2.9)
where ๐๐(๐) = โฑ๐ [๐ฆ(๐ก)]. This result follows from the isomorphism between thequaternion and complex Fourier transforms when operating on a real signal, andit may be seen to be the result of a convolution between the signal ๐ฆ(๐ก) and thequaternion Fourier transform of ๐ฅ(๐ก) = 1/๐๐ก. Note that ๐ and ๐๐(๐) commute asa consequence of ๐ฆ(๐ก) being real.
48 N. Le Bihan and S.J. Sangwine
3. The Hypercomplex Analytic Signal
We define the hyperanalytic signal ๐ง+(๐ก) by a similar approach to that originallydeveloped by Ville [21]. The following definitions give the details of the construc-tion of this signal. Note that the signal ๐ง(๐ก) is considered to be non-analytic, orimproper, in the classical (complex) sense, that is its real and imaginary parts arenot orthogonal. However, the following definitions are valid if ๐ง(๐ก) is analytic, asit can be considered as a degenerate case of the more general non-analytic case.
Definition 2. Consider a complex signal ๐ง(๐ก) = ๐ง๐(๐ก) + ๐๐ง๐(๐ก) and its quaternionFourier transform ๐๐(๐) as defined in Definition 1. Then, the hypercomplex ana-logue of the Hilbert transform of ๐ง(๐ก), is as follows:
โ๐ [๐ง(๐ก)] = โฑโ1๐ [โ๐ sign(๐)๐๐(๐)] , (3.1)
where the Hilbert transform is thought of as: โ๐ [๐ง(๐ก)] = ๐.๐ฃ. (๐ง โ (1/๐๐ก)), wherethe principal value (p.v.) is understood in its classical way. This result follows fromequation (2.9) and the linearity of the quaternion Fourier transform. To extractthe imaginary part, the vector part of the quaternion signal must be multiplied byโ๐. An alternative is to take the scalar or inner product of the vector part with๐. Note that ๐ and ๐๐(๐) anticommute because ๐ is orthogonal to ๐, the axis of๐๐(๐). Therefore the ordering is not arbitrary, but changing it simply conjugatesthe result.
Definition 3. Given a complex-valued signal ๐ง(๐ก) that can be expressed in the formof a quaternion as ๐ง(๐ก) = ๐ง๐(๐ก) + ๐๐ง๐(๐ก), then the hypercomplex analytic signal of๐ง(๐ก), denoted ๐ง+(๐ก) is given by:
๐ง+(๐ก) = ๐ง(๐ก) + ๐โ๐ [๐ง(๐ก)] , (3.2)
where โ๐ [๐ง(๐ก)] is the hypercomplex analogue of the Hilbert transform of ๐ง(๐ก)defined in the preceding definition. The quaternion Fourier transform of this hy-percomplex analytic signal is thus:
๐+(๐) = ๐๐(๐)โ ๐2 sign(๐)๐๐(๐)
= [1 + sign(๐)]๐๐(๐)
= 2๐(๐)๐๐(๐),
where ๐(๐) is the classical unit step function.
This result is unique to the quaternion Fourier transform representation ofthe hypercomplex analytic signal โ the hypercomplex analytic signal has a one-sided quaternion Fourier spectrum. This means that the hypercomplex analyticsignal may be constructed from a complex signal ๐ง(๐ก) in exactly the same waythat an analytic signal may be constructed from a real signal ๐ฅ(๐ก), by suppressionof negative frequencies in the Fourier spectrum. The only difference is that in thehypercomplex analytic case, a quaternion rather than a complex Fourier transformmust be used, and of course the complex signal must be put in the form ๐ง(๐ก) =
3. โ-spectral analysis 49
๐ง๐(๐ก) + ๐๐ง๐(๐ก) which, although a quaternion signal, is isomorphic to the originalcomplex signal.
A second important property of the hypercomplex analytic signal is that itmaintains a separation between the different even and odd parts of the originalsignal.
Property 4. The original signal ๐ง(๐ก) is the simplex part [4, Theorem 1], [18, ยง 13.1.3]of its corresponding hypercomplex analytic signal. The perplex part is the orthog-onal or โquadratureโ component, ๐(๐ก). They are obtained by:
๐ง(๐ก) =1
2(๐ง+(๐ก)โ ๐๐ง+(๐ก)๐) , (3.3)
๐(๐ก) =1
2(๐ง+(๐ก) + ๐๐ง+(๐ก)๐) . (3.4)
Proof. This follows from equation (3.2). Writing this in full by substituting theorthogonal signal for โ๐ [๐ง(๐ก)]:
๐ง+(๐ก) = ๐ง(๐ก) + ๐๐(๐ก) = ๐ง๐(๐ก) + ๐๐ง๐(๐ก) + ๐๐๐(๐ก)โ ๐๐๐(๐ก),
and substituting this into equation (3.3), we get:
๐ง(๐ก) =1
2
(๐ง๐(๐ก) + ๐๐ง๐(๐ก) + ๐๐๐(๐ก)โ ๐๐๐(๐ก)
โ๐ [๐ง๐(๐ก) + ๐๐ง๐(๐ก) + ๐๐๐(๐ก)โ ๐๐๐(๐ก)] ๐
),
and since ๐ and ๐ are orthogonal unit pure quaternions, ๐๐ = โ๐๐:
=1
2
(๐ง๐(๐ก) + ๐๐ง๐(๐ก) + ๐๐๐(๐ก)โ ๐๐๐(๐ก)
+ ๐ง๐(๐ก) + ๐๐ง๐(๐ก)โ ๐๐๐(๐ก) + ๐๐๐(๐ก)
),
from which the first part of the result follows. Equation (3.4) differs only in thesign of the second term, and it is straightforward to see that if ๐ง+(๐ก) is substituted,๐ง(๐ก) cancels out, leaving ๐(๐ก). โก
4. Geometric Instantaneous Amplitude and Phase
Using the hypercomplex analytic signal, we now present the geometric featuresthat can be obtained thanks to two representations of ๐ง+(๐ก). First, we notice thatthe hypercomplex analytic signal has a polar form given as:
๐ง+(๐ก) = ๐+R(๐ก)๐๐
+(๐ก)๐+(๐ก), (4.1)
where ๐+R = โฃ๐ง+(๐ก)โฃ is the real modulus of ๐ง+(๐ก), ๐+(๐ก) is its argument and ๐+(๐ก) its
axis. The real modulus, or real envelope, is not very informative, as it is real valuedand so does not provide a 2D description of the slowly varying part (envelope) of๐ง(๐ก). Nonetheless, a complex modulus can be defined using the modulus of thepolar CayleyโDickson representation that is more informative on the geometricfeatures of the original improper signal ๐ง(๐ก).
50 N. Le Bihan and S.J. Sangwine
4.1. Complex Envelope
The complex modulus of the hypercomplex analytic signal has properties verysimilar to the โclassicalโ case. It is defined as the modulus of the CayleyโDicksonpolar representation of ๐ง+(๐ก). Considering the CayleyโDickson polar form of thehypercomplex analytic signal given as:
๐ง+(๐ก) = ๐+C(๐ก)๐
ฮฆ+(๐ก)๐ , (4.2)
then the complex envelope of ๐ง(๐ก) is simply ๐+C(๐ก). An illustration of a complex
envelope of a complex improper signal is given in Figures 1 and 3. In Figure 1,an improper complex signal made of a low frequency envelope modulating a highfrequency linear frequency sweep is presented. The complex envelope obtainedthrough the hyperanalytic signal is displayed in red (and in blue a negated versionto show how the envelope encompass the signal). In Figure 3, a similar signal, butwith a quadratic sweep, is displayed. Again, the complex envelope fits the slowlyvarying pattern of the signal.
Note that we do not make use of the phase of the CayleyโDickson polar formin the sequel. It was demonstrated in [15] that this phase can reveal informationon the โmodulationโ part of the improper complex signal. Here we are looking forgeometric descriptors of the improper complex signal through the hyperanalyticsignal. For this purpose, we now show how the phase of the polar form of ๐ง+(๐ก)can be linked with the angular velocity concept.
4.2. Angular Velocity
The phase of a hyperanalytic signal is slightly different from the well-known con-cept of phase for a complex-valued signal. First, it is a three-dimensional quantitymade of an axis and a phase (i.e., a pure unit quaternion and a scalar). Using thepolar form of ๐ง+(๐ก), its normalized part, denoted ๐ง+(๐ก), is simply:
๐ง+(๐ก) =๐ง+(๐ก)
๐+R(๐ก)
, (4.3)
where it is assumed that ๐+R(๐ก) โ= 0 for all ๐ก. In the case where ๐+
R(๐ก) = 0, it simplymeans that the original signal ๐ง(๐ก) = 0 and so is ๐ง+(๐ก). If not, then ๐ง+(๐ก) is a unitquaternion, i.e., an element of ๐ฎ3. Without loss of generality, one can write ๐ง+(๐ก)as:
๐ง+(๐ก) = exp [๐(๐ก)v(๐ก)] , (4.4)
where ๐(๐ก) is scalar valued and v(๐ก) is a pure unit quaternion. Now, it is wellknown [13, 12] from quaternion formulations of kinematics that the instantaneousfrequency is given by:
๐(๐ก) =d๐(๐ก)
d๐ก= arg
(d๐ง+(๐ก)
d๐ก
), (4.5)
where the time derivative of ๐ง+(๐ก) is understood as:
d๐ง+(๐ก)
d๐ก= lim
ฮ๐กโ0
๐ง+(๐ก+ฮ๐ก)๐ง+(๐ก)โ1
ฮ๐ก. (4.6)
3. โ-spectral analysis 51
1
0.5
0
0.5
1
00.10.20.30.40.50.60.70.80.91
1.5
1
0.5
0
0.5
1
1.5
TimeReal
Imag
Figure 1. Improper complex signal consisting of a modulating lowfrequency envelope and a high frequency linear frequency sweep. Theenvelope ๐+
C(๐ก) is indicated by the arrow at left and โ๐+C(๐ก) by the arrow
at right.
This is in fact the multiplicative increment (a unit pure quaternion) between thehyperanalytic signal at time ๐ก and time ๐ก+ฮ๐ก. The argument in (4.5) is thus theamount of angle per unit of time by which the signal has been rotated (in rad/s).The axis of this increment gives the direction of this rotation. This is an angularvelocity, which corresponds, for a complex improper signal, to its instantaneousfrequency. This angular velocity is a geometrical and spectral local informationon the signal ๐ง(๐ก), as it gives locally the frequency content and the geometricalorientation and behaviour of the signal with time.
In Figures 1 to 4, we illustrate this concept of instantaneous frequency forimproper complex signals. In Figure 1, an improper complex signal made of alinear sweep (single frequency signal with frequency linearly changing with time)and a complex envelope is presented. Its complex envelope is presented in red (theinverse of the envelope is also plotted, in blue, for visualization purposes), whichfits the โlow frequencyโ behaviour of the signal. In Figure 2, the angular velocity ofthe hyperanalytic signal is compared to the frequency sweep used to generate the
52 N. Le Bihan and S.J. Sangwine
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
5
10
15
20
25
30
35
40
45
50
Time
Freq
uenc
y
Figure 2. Linear frequency sweep estimated as the angular velocity ofthe hyperanalytic signal (with ripples and discontinuity), and originallinear frequency sweep used to generate the original improper complexsignal (without ripples). The discontinuity in the estimate is due to thefact that the angular velocity obtained from the hyperanalytic signal isnot differentiable at ๐ก = 0.5.
signal, showing good agreement, except at a singularity point (where the phase ofthe hyperanalytic signal is not differentiable). A similar procedure is carried outin Figures 3 and 4, where the โhigh frequencyโ content is now a quadratic sweep(frequency varying quadratically with time). The same conclusions can be drawnabout the complex envelope and the angular velocity.
The complex envelope and the angular velocity are the extensions, for com-plex improper signals, of the well-known envelope and instantaneous frequencyfor real signals. They allow a local description of the geometrical and spectral be-haviour of the signal and thus consist in the simplest time-frequency representationfor improper complex signals. As in the classical case, these local descriptors havetheir limitations. Here, the number of frequency components must be kept to onefor the angular velocity to be able to recover the instantaneous frequency. Thislimitation makes impossible the identification of local spectral contents in the caseof mixtures of signals or wide-band signals. Such signals require more sophisticated
3. โ-spectral analysis 53
1
0.5
0
0.5
1
00.2
0.40.6
0.81
0.25
0.2
0.15
0.1
0.05
0
0.05
0.1
0.15
0.2
0.25
Time
Real
Imag
Figure 3. Improper complex signal consisting in a modulating lowfrequency envelope and a high frequency quadratic frequency sweep.
time-frequency representations that could be developed, based on the quaternionFT. The study of such representations is a natural step after the work presented.
5. Conclusions
We have shown in this chapter how the classical analytic signal concept may beextended to the case of the hyperanalytic signal of an original complex signal. Thequaternion based approach yields an interpretation of the hyperanalytic signal as aquaternion signal which leads naturally to the definition of the complex envelope.Also, the use of the polar form of the hyperanalytic signal allows the derivationof an angular velocity, which is indeed an instantaneous frequency. Both the enve-lope and the angular velocity allow a local and spectral description of a compleximproper signal, leading to the first geometrical time-frequency representation forcomplex improper signals.
Future work will consist in developing other time-frequency representationsfor improper complex signals with more diverse spectral behaviour.
54 N. Le Bihan and S.J. Sangwine
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
5
10
15
20
25
Time
Freq
uenc
y
Figure 4. Quadratic frequency sweep estimated as the angular veloc-ity of the hyperanalytic signal (with ripples), and original quadraticfrequency sweep used to generate the original improper complex sig-nal (without ripples). The discontinuity at ๐ก = 0.5 is due to the factthat the angular velocity obtained from the hyperanalytic signal is notdifferentiable at this point.
Acknowledgement
We thank Eckhard Hitzer for suggesting the proof of Property 2.
References
[1] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[2] T. Bulow and G. Sommer. Hypercomplex signals โ a novel extension of the ana-lytic signal to the multidimensional case. IEEE Transactions on Signal Processing,49(11):2844โ2852, Nov. 2001.
[3] M.A. Delsuc. Spectral representation of 2D NMR spectra by hypercomplex numbers.Journal of magnetic resonance, 77(1):119โ124, Mar. 1988.
3. โ-spectral analysis 55
[4] T.A. Ell and S.J. Sangwine. Hypercomplex Wiener-Khintchine theorem with ap-plication to color image correlation. In IEEE International Conference on ImageProcessing (ICIP 2000), volume II, pages 792โ795, Vancouver, Canada, 11โ14 Sept.2000. IEEE.
[5] R.R. Ernst, G. Bodenhausen, and A. Wokaun. Principles of Nuclear Magnetic Res-onance in One and Two Dimensions. International Series of Monographs on Chem-istry. Oxford University Press, 1987.
[6] M. Felsberg and G. Sommer. The monogenic signal. IEEE Transactions on SignalProcessing, 49(12):3136โ3144, Dec. 2001.
[7] D. Gabor. Theory of communication. Journal of the Institution of Electrical Engi-neers, 93(26):429โ457, 1946. Part III.
[8] S.L. Hahn. Multidimensional complex signals with single-orthant spectra. Proceed-ings of the IEEE, 80(8):1287โ1300, Aug. 1992.
[9] S.L. Hahn. Hilbert transforms. In A.D. Poularikas, editor, The transforms and appli-cations handbook, chapter 7, pages 463โ629. CRC Press, Boca Raton, 1996. A CRChandbook published in cooperation with IEEE press.
[10] S.L. Hahn. Hilbert transforms in signal processing. Artech House signal processinglibrary. Artech House, Boston, London, 1996.
[11] W.R. Hamilton. Lectures on Quaternions. Hodges and Smith, Dublin, 1853. Avail-able online at Cornell University Library: http://historical.library.cornell.edu/math/.
[12] A.J. Hanson. Visualizing Quaternions. The Morgan Kaufmann Series in Interactive3D Technology. Elsevier/Morgan Kaufmann, San Francisco, 2006.
[13] J.B. Kuipers. Quaternions and Rotation Sequences. Princeton University Press,Princeton, New Jersey, 1999.
[14] N. Le Bihan and S.J. Sangwine. About the extension of the 1D analytic signal toimproper complex valued signals. In Eighth International Conference on Mathemat-ics in Signal Processing, page 45, The Royal Agricultural College, Cirencester, UK,16โ18 December 2008.
[15] N. Le Bihan and S.J. Sangwine. The H-analytic signal. In Proceedings of EUSIPCO2008, 16th European Signal Processing Conference, page 5, Lausanne, Switzerland,25โ29 Aug. 2008. European Association for Signal Processing.
[16] J.M. Lilly and S.C. Olhede. Bivariate instantaneous frequency and bandwidth. IEEETransactions on Signal Processing, 58(2):591โ603, Feb. 2010.
[17] S.J. Sangwine and T.A. Ell. The discrete Fourier transform of a colour image. In J.M.Blackledge and M.J. Turner, editors, Image Processing II Mathematical Methods, Al-gorithms and Applications, pages 430โ441, Chichester, 2000. Horwood Publishing forInstitute of Mathematics and its Applications. Proceedings Second IMA Conferenceon Image Processing, De Montfort University, Leicester, UK, September 1998.
[18] S.J. Sangwine, T.A. Ell, and N. Le Bihan. Hypercomplex models and processing ofvector images. In C. Collet, J. Chanussot, and K. Chehdi, editors,Multivariate ImageProcessing, Digital Signal and Image Processing Series, chapter 13, pages 407โ436.ISTE Ltd, and John Wiley, London, and Hoboken, NJ, 2010.
[19] S.J. Sangwine and N. Le Bihan. Hypercomplex analytic signals: Extension of theanalytic signal concept to complex signals. In Proceedings of EUSIPCO 2007, 15th
56 N. Le Bihan and S.J. Sangwine
European Signal Processing Conference, pages 621โ4, Poznan, Poland, 3โ7 Sept.2007. European Association for Signal Processing.
[20] S.J. Sangwine and N. Le Bihan. Quaternion polar representation with a complexmodulus and complex argument inspired by the CayleyโDickson form. Advances inApplied Clifford Algebras, 20(1):111โ120, Mar. 2010. Published online 22 August2008.
[21] J. Ville. Theorie et applications de la notion de signal analytique. Cables et Trans-mission, 2A:61โ74, 1948.
Nicolas Le BihanGIPSA-Lab/CNRS11 Rue des mathematiquesF-38402 Saint Martin dโHeres, Francee-mail: [email protected]
Stephen J. SangwineSchool of Computer Science and Electronic EngineeringUniversity of EssexWivenhoe ParkColchester CO4 3SQ, UKe-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 57โ83cโ 2013 Springer Basel
4 Quaternionic Local Phase for Low-levelImage Processing Using Atomic Functions
E. Ulises Moya-Sanchez and E. Bayro-Corrochano
Abstract. In this work we address the topic of image processing using anatomic function (AF) in a representation of quaternionic algebra. Our ap-proach is based on the most important AF, the up(๐ฅ) function. The mainreason to use the atomic function up(๐ฅ) is that this function can expressanalytically multiple operations commonly used in image processing such aslow-pass filtering, derivatives, local phase, and multiscale and steering filters.Therefore, the modelling process in low level-processing becomes easy usingthis function. The quaternionic algebra can be used in image analysis becauselines (even), edges (odd) and the symmetry of some geometric objects in โ2
are enhanced. The applications show an example of how up(๐ฅ) can be ap-plied in some basic operations in image processing and for quaternionic phasecomputation.
Mathematics Subject Classification (2010). Primary 11R52; secondary 65D18.
Keywords. Quaternionic phase.
1. Introduction
The visual system is the most advanced of our senses. Therefore, it is easy tounderstand that the processing of images plays an important role in human per-ception and computer vision [3, 9]. In this chapter we address the topic of imageprocessing using geometric algebra (GA) for computer vision applications in low-level and mid-level (geometric feature extraction and analysis) processing, whichbelong to the first layers of a bottom-up computer vision system.
Complex and hypercomplex numbers play an important role in signal pro-cessing [9, 5], especially in order to obtain local features such as the frequency andphase. In 1D signal analysis, it is usual to compute via the FFT a local magni-tude and phase using the analytic signal. The entire potential of the local phase
This work has been supported by CINVESTAV and CONACYT.
58 E.U. Moya-Sanchez and E. Bayro-Corrochano
information of the images is shown when constraints or invariants are required tobe found. In other words, we use the quaternionic (local) phase because the localphase can be used to link the low-level image processing with the upper layers.The phase information is invariant to illumination changes and can be used todetect low-level geometric characteristics of lines or edges [9, 3, 13]. The phasecan also be used to measure the local decomposition of the image according to itssymmetries [6, 9].
This work has two basic goals: to present a new approach based on an atomicfunction (AF) up(๐ฅ) in a representation of geometric algebra (GA) and to applythe phase information to reduce the gap between low-level processing and imageanalysis. Our approach is based on the most important AF, the up(๐ฅ) function[12], quaternionic algebra, and multiscale and steering filters. The up(๐ฅ) functionhas good locality properties (compact support), the derivative of any order canbe expressed easily, the up(๐ฅ) and dup(๐ฅ) can mimic the simple cells of the mam-malian visual processing system [14], and the approximation to other functions(polynomial) is relatively simple [12]. As we show in this work, the atomic func-tion up(๐ฅ) can be used as a building block to build multiple operations commonlyused in image processing, such as low-pass filtering, ๐th-order derivatives, localphase, etc.
The applications presented show an example of how the AF can be applied toquaternionic phase analysis. It is based on line and edge detection and symmetrymeasurement using the phase. As a result, we can reduce the gap between thelow-level processing and the computer vision applications without abandoning thegeometric algebra framework.
2. Atomic Functions
The atomic functions were first developed in the 1970s, jointly by V.L. and V.A.Rvachev. By definition, the AF are compactly supported, infinitely differentiablesolutions of differential functional equations (see (2.1)) with a shifted argument[12], that is
๐ฟ๐(๐ฅ) = ๐
๐โ๐=1
๐(๐)๐(๐๐ฅโ ๐(๐)), โฃ๐โฃ > 1, ๐, ๐, ๐ โ ๐, (2.1)
where ๐ฟ = d๐
d๐ฅ๐+๐1
d๐โ1
d๐ฅ๐โ1 + โ โ โ +๐๐ is a linear differential operator with constant
coefficients. In the AF class, the function up(๐ฅ) is the simplest and, at the sametime, the most useful primitive function to generate other kinds of AFs [12].
2.1. Mother Atomic Function up(๐)
In general, the atomic function up(๐ฅ) is generated by infinite convolutions of rect-angular impulses. The function up(๐ฅ) has the following representation in terms of
4. Quaternionic Local Phase 59
150 100 50 0 50 100 1500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
space (x)150 100 50 0 50 100 150
0
0.2
0.4
0.6
0.8
1
frequency ( )
Figure 1. Left: atomic function up(๐ฅ); right: Fourier transform up(๐).
the Fourier transform [12, 8]:
up(๐ฅ) =1
2๐
โซ โ
โโ
โโ๐=1
sin(๐2โ๐)๐2โ๐
๐๐๐๐ฅd๐. (2.2)
=1
2๐
โซ โ
โโup(๐)๐๐๐๐ฅd๐ (2.3)
Figure 1 shows the up(๐ฅ) and its Fourier transform up(๐).
The AF windows were compared with classic windows by means of parameterssuch as the equivalent noise bandwidth, the 50% overlapping region correlation, theparasitic modulation amplitude, the maximum conversion losses (in decibels), themaximum side lobe level (in decibels), the asymptotic decay rate of the side lobes(in decibels per octave), the window width at the six- decibel level, the coherentgain, etc. All atomic windows exceed classic ones in terms of the asymptotic decayrate [12].
However the main reasons that we found to use the AF up(๐ฅ) instead ofother kernels such as Gauss, Gabor, and log-Gabor is that many operations com-monly used in image processing can be expressed analytically, in contrast usingthe Gauss or the log-Gabor this could be impossible. Additionally the followinguseful properties of the AF have been reported in [12, 10]
โ There are explicit equations for the values of the moments and Fourier trans-form (see (2.3)). The even moments of up(๐ฅ) are
๐๐ =
1โซโ1
๐ฅ๐ up(๐ฅ)d๐ฅ (2.4)
๐2๐ =(2๐)!
22๐ โ 1
๐โ๐=1
๐2๐โ2๐
(2๐โ 2๐)!(2๐ + 1)!(2.5)
60 E.U. Moya-Sanchez and E. Bayro-Corrochano
where ๐0 = 1 and ๐2๐+1 = 0. The odd-order moments are
๐๐ =
1โซ0
๐ฅ๐ up(๐ฅ)d๐ฅ (2.6)
๐2๐+1 =1
(๐+ 1)22๐+1
๐+1โ๐=1
๐2๐โ2๐+2
(2๐
2๐+ 2๐
)(2.7)
where ๐2๐ = ๐2๐. A similar relation for Gauss or Log-Gabor does not exist.โ The up(๐ฅ) function is a compactly supported function in the spatial do-main. Therefore, we can obtain good local characteristics. In addition tocompactly supported functions and integrable functions, functions that havea sufficiently rapid decay at infinity can also be convolved. The Gauss andLog-Gabor functions do not have such compact support.
โ Translations with smaller steps yield polynomials (๐ฅ๐) of any degree, i.e.,โโ
๐=โโ๐๐ up(๐ฅโ (๐2๐)) โก ๐ฅ๐ ๐๐, ๐ฅ โ โ. (2.8)
โ Since derivatives of any order can be represented in terms of simple shifts,we can easily represent any derivative operator or ๐th-order derivative:
d(๐) up(๐ฅ) = 2๐(๐+1)/22๐โ๐=1
๐ฟ๐ up(2๐๐ฅ+ 2๐ + 1โ 2๐), (2.9)
where ๐ฟ2๐ = โ๐ฟ๐, ๐ฟ2๐โ1 = ๐ฟ๐, ๐ฟ2๐ = 1.โ The AFs are infinitely differentiable (๐ถโ). As a result, the AFs and theirFourier transforms are rapidly decreasing functions. (Their Fourier trans-forms decrease on the real axis faster than any power function.)
A natural extension of the up(๐ฅ) function to the case of many variables is basedon the usual tensor product of 1D up(๐ฅ) [10]. As a result, we have
up(๐ฅ, ๐ฆ) = up(๐ฅ) up(๐ฆ). (2.10)
In Figure 2, we show a 2D atomic function in the spatial and frequency domains.
2.2. The dup(๐) Function
There are some mask operators, including the Sobel, Prewitt, and Kirsh, that areused to extract edges from images. A common drawback of these operators is thatit is impossible to ensure that they adapt to the intrinsic signal parameters over awide range of the working band, i.e., they are truncated and discrete versions of adifferential operator. [8]. This means that adaptation of the differential operator tothe behavior of the input signal by broadening or narrowing its band is desirable,in order to ensure a maximum signal-to-noise ratio [8].
This problem reduces to the synthesis of infinitely differentiable finite func-tions with a small wide bandwidth that are used for constructing the weightingwindows [8]. One of the most effective solutions is obtained with the help of the
4. Quaternionic Local Phase 61
Figure 2. Top: Two views of up(๐ฅ, ๐ฆ). Bottom: Fourier transform up(๐, ๐).
atomic functions [8]. The AF can be used in two ways: construction of a windowin the frequency region to obtain the required improvement in properties of thepulse characteristic; or direct synthesis based on (2.1) [8]. Therefore, the functionup(๐ฅ) satisfies (2.1) as follows:
dup(๐ฅ) = 2 up(2๐ฅ+ 1)โ 2 up(2๐ฅโ 1) (2.11)
Figure 3 shows the dup(๐ฅ) function and its Fourier transform ๐๐ข๐(๐). If we computethe Fourier transform of (2.11), we obtain
๐๐๐น [up(2๐ฅ)] =(๐๐๐ โ ๐โ๐๐
)๐น [up(2๐ฅ)]) (2.12)
๐น (dup(๐ฅ)) = 2๐ sin(๐)๐น (up(2๐ฅ)) (2.13)
By differentiating (2.1) term by term, we obtain [8]
d(๐) up(๐ฅ) = 2๐(๐+1)/22๐โ๐=1
๐ฟ๐ up(2๐๐ฅ+ 2๐ + 1โ 2๐), (2.14)
where ๐ฟ2๐ = โ๐ฟ๐, ๐ฟ2๐โ1 = ๐ฟ๐, and ๐ฟ2๐ = 1. The function dup(๐ฅ) provides a goodwindow in the spatial frequency regions because the side lobe has been completely
62 E.U. Moya-Sanchez and E. Bayro-Corrochano
150 100 50 0 50 100 150
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Space (x)150 100 50 0 50 100 150
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Frequency ( )
Figure 3. Left: derivative of atomic function dup(๐ฅ); Right: Fourier
transform dup(๐).
eliminated [8]. Similarly to (2.10), we can get a 2D expression of each derivative:
dup(๐ฅ, ๐ฆ)๐ฅ = dup(๐ฅ) up(๐ฆ) (2.15)
dup(๐ฅ, ๐ฆ)๐ฆ = up(๐ฅ) dup(๐ฆ) (2.16)
dup(๐ฅ, ๐ฆ)๐ฅ,๐ฆ = dup(๐ฅ) dup(๐ฆ) (2.17)
Figure 4 shows two graphics of dup(๐ฅ, ๐ฆ)๐ฆ, dup(๐ฅ, ๐ฆ)๐ฅ,๐ฆ in the spatial domain.
Figure 4. Left: dup(๐ฅ, ๐ฆ)๐ฅ,๐ฆ; right: dup(๐ฅ, ๐ฆ)๐ฆ.
3. Why the Use of the Atomic Function?
In this section we justify our introduction of the atomic function up(๐ฅ) as a kernel.In this regard, we discuss the role of the Gauss, LogGauss and the up(๐ฅ) kernelsin low-level image processing. First, we analyse their characteristics from a math-ematical perspective and then we carry out an experimental test to show theirstrengths in the processing of real images. We have to stress that we are interestedin the application of the up(๐ฅ) kernel not just for filtering in scale space. This
4. Quaternionic Local Phase 63
up(๐ฅ) kernel is promising because it has a range of multiple possible applicationswhich is wider than the possible applications of the Gauss or LogGauss kernels.
3.1. Dyadic Shifts and the Riesz Transform
As was explained in Section 2, atomic functions are compactly supported, in-finitely differentiable solutions of differential equations with a shifted argument.Consequently an atomic function can be seen as an appropriate building block oflinear shift invariant (LSI) operators to implement complex operators for imageprocessing. In contrast the Gauss and LogGauss functions require derivatives andconvolutions with a function of series of impulses for building complex operators.We believe that atomic functions are promising to develop operators based onlinear combination of shifted atomic functions. Section 2.2 illustrates the differ-entiator atomic function dup(๐ฅ). In 2002 Petermichl [16] showed that the Hilberttransform lies in the closed convex hull of dyadic singular operators, thus theHilbert transform can be represented as an average of dyadic shifts. Petermichl etal. [17] show that the same is true for โ๐, therefore the Riesz transforms can alsobe obtained as the results of averaging of dyadic shifts. Thus we claim that theRiesz transforms can be obtained by averaging dyadic shifts of atomic functions.
3.2. Monogenic Signals and the Atomic Function
The monogenic signal was introduced by Felsberg, Bulow, and Sommer [7]. Weoutline briefly the concept of the monogenic signal and then explain the use ofthe atomic-function based monogenic signal. If we embed โ3 into a subspace of โspanned by just {1, ๐, ๐} according to
๐ = (๐, ๐, 1)๐ = ๐ฅ3 + ๐ฅ1๐+ ๐ฅ2๐, (3.1)
and further embed the vector field ๐ as follows:
๐โ = (โ๐,โ๐, 1)๐ = ๐3 โ ๐1๐โ ๐2๐ (3.2)
then โร ๐(๐) = 0 and โ โ ๐(๐) = โจโ, ๐(๐)โฉ = 0 are equivalent to the generalizedCauchyโRiemann equations from Clifford analysis [4, 15]. All functions that fulfillthese equations are known as ๐๐๐๐๐๐๐๐๐ functions. Using the same embedding,the monogenic signal can be defined in the frequency domain as follows:
๐น๐ (๐ข) = ๐บ3(๐ข1, ๐ข2, 0)โ ๐๐บ1(๐ข1, ๐ข2, 0)โ ๐๐บ2(๐ข1, ๐ข2, 0)
= ๐น (๐)โ (๐, ๐)๐ญ๐ (๐) =โฃ๐โฃ+ (1,๐)๐
โฃ๐โฃ ๐น (๐), (3.3)
where the inverse Fourier transform of ๐น๐ (๐ข) is given by
๐๐ (๐) = ๐(๐)โ (๐, ๐)๐๐ (๐) = ๐(๐) + โ๐ โ ๐(๐), (3.4)
64 E.U. Moya-Sanchez and E. Bayro-Corrochano
where ๐๐ (๐) stands for the Riesz transform obtained by taking the inverse trans-form of ๐ญ๐ (๐) as follows:
๐ญ๐ (๐) =๐๐
โฃ๐โฃ๐น (๐) = ๐ฏ๐ (๐)๐น (๐)โโ ๐๐ (๐) = โ๐
2๐ โฃ๐โฃ2 โ ๐(๐)
= โ๐ (๐) โ ๐(๐), (3.5)
where โ stands for the convolution operation. Note that here the Riesz transformis the generalization of the 1D Hilbert transform. Using the fundamental solutionof the 3D Laplace equation restricted to the open half-space ๐ง > 0 with boundarycondition, the solution is defined as
๐๐ (๐ฅ, ๐ฆ, ๐ง) = โ๐ โ ๐(๐ฅ, ๐ฆ, ๐ง) + โ๐ โ โ๐ โ ๐(๐ฅ, ๐ฆ, ๐ง)
= โ๐ โ (1 + โ๐ โ )๐(๐ฅ, ๐ฆ, ๐ง), (3.6)
where โ๐ stands for the 2D Poisson kernel. Setting in ๐๐ (๐ฅ, ๐ฆ, ๐ง) the variable ๐งequal to zero, we obtain the so-called monogenic signal. Some authors have used theGauss kernel instead of the Poisson kernel, because the Poisson kernel establishesa linear scale space similar to the Gaussian scale space.
The atomic function is also an LSI operator; therefore, it appears that its useensures a computation in a linear scale space as well.
The monogenic functions are the solutions of the generalized CauchyโRie-mann equations or Laplace-type equations. The atomic function can be used tocompute compactly supported solutions of functional differential equations, forexample, (2.1). Conditions under which the type of equations (2.1) have solu-tions with compact support and an explicit form were obtained by Ravachev [12].Compactly supported solutions of equations of the type (2.1) are called atomicfunctions .
Now, for the case of 2D signal processing, we can apply the wavelet steerabil-ity and the Riesz transform. In this regard we will utilize the quaternion waveletatomic function, which will be discussed in more depth below. For the scale spacefiltering, authors use the Gauss or Poisson kernels for the Riesz transformation[6, 7]. It is known that the Poisson kernel is the fundamental solution of the 3DLaplace equation, however there are authors who use instead the Poisson, Gaussor LogGauss function. Felsberg [6] showed that the spatial extent of the Poissonkernel is greater than that of the Gaussian kernel. In addition the uncertainty ofthe Poisson kernel for 1D signals is slightly worse than that of the Gaussian kernel(by a factor of
โ2). Nowadays many authors are now using the LogGauss kernel
for implementing monogenic signals, because it has better properties than boththe Gauss and Poisson kernels. Thus, we can infer that, in contrast, the atomicfunction as a spatially-compact kernel guarantees an analytic and closed solutionof Laplace type equations. In addition, as we said above, the Riesz transformscan be obtained by averaging dyadic operators, thus the use of a compact atomic
4. Quaternionic Local Phase 65
function avoids increased truncation errors. However, by the case of the noncom-pact Poisson, Gauss and LogGauss kernels, in practice, their larger spatial extentsrequire either larger filter masks otherwise they cause increased truncation errors.
4. Quaternion Algebra โ
The even subalgebra ๐ข+3,0,0 (bivector basis) is isomorphic to the quaternion alge-
bra โ, which is an associative, non-commutative, four-dimensional algebra thatconsists of one real element and three imaginary elements.
๐ = ๐+ ๐๐+ ๐๐ + ๐๐, ๐, ๐, ๐, ๐ โ โ (4.1)
The units ๐, ๐ obey the relations ๐2 = ๐2 = โ1, ๐๐ = ๐. โ is geometrically inspired,and the imaginary components can be described in terms of the basis of โ3 space,๐โ ๐23, ๐ โ ๐12,๐โ ๐31. Another important property of โ is the phase concept.A polar representation of ๐ is
๐ = โฃ๐โฃ ๐๐๐๐๐๐๐๐๐, (4.2)
where โฃ๐โฃ = โ๐๐ where ๐ is a conjugate of ๐ = ๐ โ ๐๐ โ ๐๐ โ ๐๐ and the angles
(๐, ๐, ๐) represent the three quaternionic phases [5].
4.1. Quaternionic Atomic Function qup(๐, ๐)
Since a 2D signal can be split into even (e) and odd (o) parts [5],
๐(๐ฅ, ๐ฆ) = ๐ee(๐ฅ, ๐ฆ) + ๐oe(๐ฅ, ๐ฆ) + ๐eo(๐ฅ, ๐ฆ) + ๐oo(๐ฅ, ๐ฆ), (4.3)
we can separate the four components of up(๐ฅ, ๐ฆ) and represent it as a quaternionas follows [13, 14]:
qup(๐ฅ, ๐ฆ) = up(๐ฅ, ๐ฆ)[cos(๐ค๐ฅ) cos(๐ค๐ฆ)
+ ๐ sin(๐ค๐ฅ) cos(๐ค๐ฆ) + ๐ cos(๐ค๐ฅ) sin(๐ค๐ฆ) + ๐ sin(๐ค๐ฅ) sin(๐ค๐ฆ)]
= qupee(๐ฅ, ๐ฆ) + ๐ qupoe(๐ฅ, ๐ฆ) + ๐ qupeo(๐ฅ, ๐ฆ) + ๐ qupoo(๐ฅ, ๐ฆ). (4.4)
Figure 5 shows the quaternion atomic function qup(๐ฅ, ๐ฆ) in the spatial domain withits four components: the real part qupee(๐ฅ, ๐ฆ) and the imaginary parts qupeo(๐ฅ, ๐ฆ),qupoe(๐ฅ, ๐ฆ), and qupoo(๐ฅ, ๐ฆ). We can see even and odd symmetries in the horizon-tal, vertical, and diagonal axes.
Figure 5. Atomic function qup(๐ฅ, ๐ฆ). From left to right: qupee(๐ฅ, ๐ฆ),qupoe(๐ฅ, ๐ฆ), qupeo(๐ฅ, ๐ฆ), and qupoo(๐ฅ, ๐ฆ).
66 E.U. Moya-Sanchez and E. Bayro-Corrochano
4.2. Quaternionic Atomic Wavelet
The importance of wavelet transforms (real, complex, hypercomplex) has beendiscussed in many references [1, 18]. In this work, we use the quaternionic wavelettransform (๐๐๐ ). The ๐๐๐ can be seen as an extension of the complex wavelettransform (๐ถ๐๐ ) [1]. The multiresolution analysis applied in ๐ ๐๐ and ๐ถ๐๐can be straightforwardly extended to the ๐๐๐ [1].
Our approach is based on the quaternionic Fourier transform (๐๐น๐ ), (4.5),which is also the basis for the quaternionic analytic function defined by Bulow [5].The kernel of the 2D ๐๐น๐ is given by
๐โ๐2๐๐๐ฅ๐(๐ฅ, ๐ฆ)๐โ๐2๐๐๐ฆ, (4.5)
where the real part is cos(2๐๐๐ฅ) cos(2๐๐๐ฆ); the imaginary parts (๐, ๐,๐) are its par-tial Hilbert transform cos(2๐๐๐ฅ) sin(2๐๐๐ฆ), sin(2๐๐๐ฅ) cos(2๐๐๐ฆ) (horizontal andvertical); and total Hilbert transform (diagonal) sin(2๐๐๐ฅ) sin(2๐๐๐ฆ). Using the๐๐น๐ basis, we can obtain the 2D phase information that satisfies the definitionof the quaternionic analytic signal. To compute the multiscale approach, we use a2D separable implementation. We independently apply two sets of โ and ๐ waveletfilters.
โ = exp
(๐๐1๐ข1๐ฅ
๐1
)up(๐ฅ, ๐ฆ, ๐1) exp
(๐๐ 1๐๐ฃ1๐ฆ
๐1
)= โee + ๐โoe + ๐โeo + ๐โoo (4.6)
๐ = exp
(๐๐2๐ค2๐ฅ
๐2
)up(๐ฅ, ๐ฆ, ๐1) exp
(๐๐ 2๐๐ฃ2๐ฆ
๐2
)= ๐ee + ๐๐oe + ๐๐eo + ๐๐oo, (4.7)
where the extra ๐ parameter in the up function stands for the filter width. Theprocedure for quaternionic wavelet multiresolution analysis depicted partially inFigure 6 is as follows [1]:
1. Convolve the image (2D signal) at level ๐ with the scale and wave filtersโ and ๐ along the rows of the 2D signal. The latter filters are the discreteversions of those filters given in (4.6) and (4.7).
2. The โ and ๐ filters are convolved with the columns of the previous responsesof the filters โ and ๐.
3. Subsample the responses of these filters by a factor of 2 (โ 2).4. The real part of the approximation at level ๐ + 1 is taken as input at the
next level. This process continues through all levels, repeating the steps juststated.
As can be seen in Figure 6, the low level of the pyramid is the highest level ofresolution, and as you move up, the resolution decreases.
4. Quaternionic Local Phase 67
Image
Level 1
h1
g1
Rows
h1
g1
Columns
h1
g1
2
22
2
2
2
Approximation
Horizontal
Vertical
Diagonal2=subsampling
Image2
Level 2
h2
g2
Rows
h2
g2
Colums
h2
g2
2
22
2
2
2
Approximation
Horizontal
Vertical
Diagonal2=subsampling
Figure 6. Multiresolution approach.
4.3. Steerable Quaternionic Filter
Due to our approach not being invariant to rotations [6], we require a filter bankin order to get different line or edge orientations. Figure 7 shows different orien-tation filters and real and imaginary parts. We steer the mother atomic functionwavelet through a multiresolution pyramid in order to detect quaternionic phasechanges, which can be used for the feature detection of lines, edges, centroids, andorientation in geometric structures.
5. Quaternionic Local Phase Information
In this chapter, we refer to the phase information as the local phase in order toseparate the structure or geometric information and the amplitude in a certainpart of the signal. Moreover, the phase information permits us to obtain invariantor equivariant1 response. For instance, it has been shown that the phase has aninvariant response to changes in image brightness and the phase can be used tomeasure the symmetry or asymmetry of objects [11, 2, 9]. These invariant andequivariant responses are the key part to link the low-level processing with theimage analysis and the upper layers in computer vision applications.
The local phase means the computation of the phase at a certain positionin a real signal. In 1D signals, the analytic signal based on the Hilbert transform(๐๐ฏ(๐ฅ)) [5] is given by
๐๐ด(๐(๐ฅ)) = ๐(๐ฅ) + ๐๐๐ฏ(๐ฅ), (5.1)
๐๐ด(๐(๐ฅ)) = โฃ๐ดโฃ ๐๐๐, (5.2)
where โฃ๐ดโฃ =โ
๐(๐ฅ)2 + ๐๐ฏ(๐ฅ)2 and ๐ = arctan
(๐(๐ฅ)๐๐ฏ(๐ฅ)
)permits us to extract the
magnitude and phase independently. In 2D signals, the Hilbert transform is notenough to compute the magnitude and phase independently in any direction [6]. Inorder to solve this, the quaternionic analytic (see (5.3)) signal and the monogenic
1Equivariance: monotonic dependency of value or parameter under some transformation.
68 E.U. Moya-Sanchez and E. Bayro-Corrochano
Figure 7. From left to right: qup(15โ), qup(40โ), qup(90โ), qup(140โ),and qup(155โ)
signal have been proposed by Bulow [5] and Felsberg [6], respectively. Until now,we have used an approximation of the quaternionic analytic signal based on thebasis of ๐๐น๐ to extract some oriented axis symmetries.
Figure 8 contains, at the top, an image with lines and edges, at the centrean image profile, and at the bottom the profiles of the three quaternionic phases.In the phase profiles, we can distinguish between a line (even) and an edge (odd)using the phase (๐). These results are similar to the results reported by Granlund[9] using only a complex phase, because they used an image that changes in onedirection.
5.1. Quaternionic Analytic Signal
The quaternionic analytic signal in the spatial domain is defined as [5]:
๐ ๐๐ด(๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) + ๐๐๐ฏ๐(๐ฅ, ๐ฆ) + ๐๐๐ฏ๐(๐ฅ, ๐ฆ) + ๐๐๐ฏ๐(๐ฅ, ๐ฆ), (5.3)
where ๐๐ฏ๐(๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) โ (๐ฟ(๐ฆ)/๐๐ฅ) and ๐๐ฏ๐(๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) โ (๐ฟ(๐ฅ)/๐๐ฆ) are thepartial Hilbert transforms and ๐๐ฏ๐(๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) โ (๐ฟ(๐ฅ, ๐ฆ)/๐2๐ฅ๐ฆ
)is the total
Hilbert transform. Bulow has shown that the ๐๐น๐ kernel is expressed in terms ofthe Hilbert transforms. The phases can be computed easily using a 3D rotation ma-trix โณ, which can be factored into three rotations, ๐ = ๐ ๐ฅ(2๐), ๐ ๐ง(2๐), ๐ ๐ฆ(2๐),
4. Quaternionic Local Phase 69
0 20 40 60 80 100 120
20
40
60
80
100
120
0 20 40 60 80 100 120
0
0.2
0.4
0.6
0.8
1
20 40 60 80 100 120
1
0.5
0
0.5
1
1.5
Figure 8. Image (top left), image profile (top right), and three quater-nionic phases: profile, line, and edge (bottom).
in the coordinate axes [5], i.e.,
โณ(๐) =โณ(๐1)โณ(๐2)โณ(๐3) (5.4)
๐1 = ๐๐๐, ๐2 = ๐๐๐, ๐3 = ๐๐๐ (5.5)
โณ(๐) =
โโ๐2 + ๐2 โ ๐2 โ ๐2 2(๐๐โ ๐๐) 2(๐๐+ ๐๐)2(๐๐+ ๐๐) ๐2 โ ๐2 + ๐2 โ ๐2 2(๐๐โ ๐๐)2(๐๐โ ๐๐) 2(๐๐+ ๐๐) ๐2 โ ๐2 โ ๐2 + ๐2
โโ (5.6)
๐ =
โโโโโโโโโ
cos(2๐) cos(2๐) โ sin(2๐) cos(2๐) sin(2๐)
cos(2๐) sin(2๐) cos(2๐)+ sin(2๐) sin(2๐) cos(2๐) cos(2๐)
cos(2๐) sin(2๐) sin(2๐)โ sin(2๐) cos(2๐)
sin(2๐) sin(2๐) cos(2๐)โ cos(2๐) sin(2๐) sin(2๐) cos(2๐)
sin(2๐) sin(2๐) sin(2๐)+ cos(2๐) cos(2๐)
โโโโโโโโโ .
(5.7)
The quaternionic phases are expressed by the following rules:
๐ = โ1
2arcsin(2(๐๐โ ๐๐)). (5.8)
70 E.U. Moya-Sanchez and E. Bayro-Corrochano
โ If ๐ โ ]โ๐4 ,
๐4
[, then ๐ = 1
2 arg๐ (๐๐ฏ๐ (๐)) and ๐ = 12 arg๐ (๐ฏ๐ (๐) ๐).
โ If ๐ = ยฑ๐4 , then select either ๐ = 0 and ๐ = 1
2 arg๐ (๐ฏ๐ (๐) ๐) or ๐ = 0 and
๐ = 12 arg๐ (๐๐ฏ๐ (๐)).
โ If ๐๐๐๐๐๐๐๐๐ = โ๐ and ๐ โฅ 0, then ๐โ ๐โ ๐.โ If ๐๐๐๐๐๐๐๐๐ = โ๐ and ๐ < 0, then ๐โ ๐+ ๐.
The phase ranges are (๐, ๐, ๐) โ [โ๐, ๐[ร [โ๐2 ,
๐2
[ร [โ๐4 ,
๐4
]. The applications of
the quaternionic analytic signal in image processing have to be limited in a narrowband, and Bulow used a Gauss window with the ๐๐น๐ kernel (which can be seenas Gabor) to approximate the quaternionic analytic function. In our work, insteadof a Gauss window, we use a compact support window, the up(๐ฅ, ๐ฆ) function.
5.2. Quaternionic Phase Analysis
There are many points of view to see the phase information. The local phase canbe used as a measure of the symmetry in the 2D signals of the object [11, 9].The symmetry is related to middle-level properties if it remains invariant undersome transformation. A symmetric analysis related to the phase was proposedby P. Kovesi, using the even (๐๐(๐ฅ)) and odd (๐๐(๐ฅ)) responses of wavelets atscale ๐ [11]:
Sym(๐ฅ) =
โ๐
โฃ๐๐(๐ฅ)โฃ โ โฃ๐๐(๐ฅ)โฃโ๐
๐ด๐(๐ฅ)(5.9)
Asym(๐ฅ) =
โ๐
โฃ๐๐(๐ฅ)โฃ โ โฃ๐๐(๐ฅ)โฃโ๐
๐ด๐(๐ฅ)(5.10)
where ๐ด๐(๐ฅ) =โ
๐๐(๐ฅ)2 + ๐๐(๐ฅ)2. Since the phase information has the capabilityto decode the geometrical information into even or odd symmetries (see Figure 8),we use the phase information to do a geometric analysis of the image. Note thatthe quaternionic phase analysis helps to reduce the gap between the low-level andthe middle-level processing.
5.3. Hilbert Transform Using AF
The Hilbert transform and the derivative are closely related, and the Hilbert trans-form can actually be computed using a derivative and some convolution proper-ties [19]:
๐ โ (๐ โ โ) = (๐ โ ๐) โ โ (5.11)
โ (๐ โ ๐) = โ๐ โ ๐ = ๐ โ โ๐, (5.12)
4. Quaternionic Local Phase 71
where
๐(๐ฅ, ๐ฆ), ๐(๐ฅ, ๐ฆ), โ(๐ฅ, ๐ฆ) โ โ2 and โ = ๐1โ
โ๐ฅ+ ๐2
โ
โ๐ฆ.
If ๐(๐ฅ, ๐ฆ) = โ(1/๐) log โฃ๐ฅโฃ log โฃ๐ฆโฃ 2, and if we use the convolution distribution prop-erties, we can express the Hilbert transform and the partial Hilbert transform (see(5.3)) as
๐๐ฏ๐(๐ฅ, ๐ฆ) =โ๐(๐ฅ, ๐ฆ)
โ๐ฅโ โ 1
๐log โฃ๐ฅโฃ (5.13)
๐๐ฏ๐ (๐ฅ, ๐ฆ) =โ๐(๐ฅ, ๐ฆ)
โ๐ฆโ โ 1
๐log โฃ๐ฆโฃ (5.14)
๐๐ฏ๐(๐ฅ, ๐ฆ) =โ2๐(๐ฅ, ๐ฆ)
โ๐ฅโ๐ฆโ โ 1
๐2log โฃ๐ฅโฃ log โฃ๐ฆโฃ (5.15)
and we can use the convolution association property to get the equation of a certainpart of the signal in terms of dup(๐ฅ):
๐๐ฏ๐(๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) โ
(dup(๐ฅ, ๐ฆ)๐ฅ โ โ 1
๐log โฃ๐ฅโฃ
)(5.16)
๐๐ฏ๐ (๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) โ
(dup(๐ฅ, ๐ฆ)๐ฆ โ โ 1
๐log โฃ๐ฆโฃ
)(5.17)
๐๐ฏ๐(๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) โ
(dup(๐ฅ, ๐ฆ)๐ฅ๐ฆ โ โ 1
๐2log โฃ๐ฅโฃ log โฃ๐ฆโฃ
). (5.18)
Moreover, it has been shown by Petermichl et al. [17] that the Hilbert andRiesz transforms can be implemented with the average of dyadic shifts, and thedyadic shift operations appear naturally using the AF and its derivatives.
6. Applications
6.1. Convolution ๐ฐ โ qup(๐, ๐)
Figure 9 shows the convolution of the qup (real and imaginary parts) filter on ashadow chessboard image. The real part corresponds to a low-pass filtering of theimage. In addition, we can see a selective detection of lines in the horizontal and
2log โฃ๐ฅโฃ is the fundamental solution of the Laplace equation.
Figure 9. Convolution of qup(๐ฅ, ๐ฆ) with chessboard image. From leftto right: original image, real part, ๐-part, ๐-part, and ๐-part.
72 E.U. Moya-Sanchez and E. Bayro-Corrochano
Figure 10. Convolution of qup(๐ฅ, ๐ฆ) with image (letter H). From leftto right: original image, real part, ๐-part, ๐-part, and ๐-part.
Figure 11. Convolution of qup(๐ฅ, ๐ฆ) with image (letter L). From leftto right: original image, real part, ๐-part, ๐-part, and ๐-part are shown.
vertical orientations, particularly in the ๐-part and ๐-part. On the other hand, the๐-part can be used to detect the corners as a diagonal response. We can see howthe response in the light part is more intensive than that in the shadowed part.We show in Figures 10 and 11 how the qup(๐ฅ, ๐ฆ) filters respond to other images.We can see a similar behaviour in Figure 9. We can also notice that the directconvolution of qup(๐ฅ, ๐ฆ) with the image is sensitive to the contrast.
6.2. Derivative of up(๐)
Figure 12 illustrates the convolution of the first derivative, dup(๐ฅ, ๐ฆ), and a shadowchessboard. The convolution with dup can be used as an oriented change detec-tor with a simple rotation. Similarly to Figure 9, the convolved image has a lowresponse in the shadow part.
Figure 12. Convolution of dup(๐ฅ, ๐ฆ) with the chess image. (a) Testimage; (b) result of the convolution of the image with dup(๐ฅ, ๐ฆ, 0โ); (c)dup(๐ฅ, ๐ฆ, 45โ); (d) dup(๐ฅ, ๐ฆ, 135โ).
4. Quaternionic Local Phase 73
6.3. Image Processing Using Monogenic Signals
We use these three kernels to implement monogenic signals whereby the apertureof the filters are varied at three scales of a multiresolution pyramid. The samereference frequency and the same aperture rate variation are used for the threemonogenic signals. In Figures 13, 14 and 15, you can appreciate at these scales thebetter consistency of the up(๐ฅ) kernel particularly to detect via the phase concept
(a) Upper row: Gauss, Hx and Hy frequency; lower row: energy, ๐, ๐
(b) Upper row: LogGabor, Hx and Hy frequency; lower row: energy, ๐, ๐
Figure 13. Monogenic signal and response at the first level of filtering(๐ = 1). (a) Gabor monogenic; (b) LogGabor monogenic signal.
74 E.U. Moya-Sanchez and E. Bayro-Corrochano
(c) Upper row: up, Hx and Hy frequency; lower row: energy, ๐, ๐
Figure 13. Monogenic signal and response at the first level of filtering(๐ = 1). (c) up monogenic.
the corners and borders diminishing the expected blurring by a checkerboard im-age. We suspect that this an effect of the compactness in space of the up(๐ฅ) kernel,the Moire effects are milder especially if you observe the images at scales 2 and3, whereas in both the Gauss and LogGauss kernel-based monogenic phases thereare noticeable artifacts in the form of bubbles or circular spots. These are often aresult of Moire effects. In contrast the phases of the up(๐ฅ) based monogenic signalstill preserves the scales and edges of the checkerboard.
6.4. Quaternionic Local Phase
As we have mentioned before, the (local) phase information can be used to extractthe structure information (line and edges) independently of illumination changes,and we can measure the symmetry (or asymmetry) with the phase information asa middle-level property. The performance of qup to detect edges or lines using thequaternionic phases in different images is shown in this section.
In Figure 17, we present different geometric figures with lines that representthe profiles (top, centre, and bottom). Similarly to Figure 8, the profiles of theimages are shown as well as the quaternionic phase profile. This figure illustrateshow the three phases have a different symmetric or antisymmetric response on thetop, center, and bottom profile. These results motivate us to use the quaternionicphases as a measure of symmetry using the even or odd response of the quaternionicphases in a horizontal or vertical direction. In others words, the quaternionic phaseinformation decomposes the image into oriented edges or lines.
4. Quaternionic Local Phase 75
As an example of other possible applications of the phase information inimage analysis of geometric figures, we show Figure 18. The original image, themagnitude, and the three phases are shown. The first row of Figure 18 shows foursquares with a different illumination and area; one of the squares is rotated, andthis is the only square that can be seen in the ๐ phase. In the second row, thefour squares have been rotated 15โ, and the ๐ phase responds to the four squares,whereas in the third row, the four figures are rotated 45โ, and the ๐ phase onlydetects three of the four squares. In this example, the ๐ phase only responds tothe orientation of the square (a geometric transformation), independently of thesize, contrast, or position of the image.
The magnitude and the quaternionic phases for a circle with lines are shownin Figure 19. In this image, we tuned the filter parameters in order to obtain aresponse (positive and negative) to diagonal lines. Again, in this image, the shadowor the size of the lines in the image does not change the phase response. The blacklines correspond to a cone beam with a 45โ orientation (ยฑ15โ) and the white linesare oriented 135โ (ยฑ15โ).
In Figures 19 and 18, lines or geometric objects have been shown separately.Figure 20 illustrates an image with lines and geometric objects. In this case, we canuse the quaternionic phases to extract different oriented edges of geometric objectsor different oriented line textures. The ๐ and ๐ phases detect lines or edges in somevertical and horizontal directions, while the ๐ phase detects the diagonal responseand the corners in squares or geometric figures. Even if the geometric objects haveinternal lines, different illuminations, or positions, the edges of each square can
(a) Upper row: Gauss, Hx and Hy frequency; lower row: energy, ๐, ๐
Figure 14. Monogenic signal and response at the second level of fil-tering (๐ = 2). (a) Gabor filters.
76 E.U. Moya-Sanchez and E. Bayro-Corrochano
(b) Upper row: LogGabor, Hx and Hy frequency; lower row: energy, ๐, ๐
(c) Upper row: up, Hx and Hy frequency; lower row: energy, ๐, ๐
Figure 14. Monogenic signal and response at the second level of fil-tering (๐ = 2). (b) LogGabor and (c) up filters.
be detected in the ๐ and ๐ images. Furthermore, in the ๐ phase, the horizontallines are highlighted, whereas the vertical lines do not appear. The ๐ phases showa similar result, but in this case the vertical edges and lines are highlighted. In the๐ phase, the vertical lines or edges are highlighted.
4. Quaternionic Local Phase 77
(a) Upper row: Gauss, Hx and Hy frequency; lower row: energy, ๐, ๐
(b) Upper row: LogGabor, Hx and Hy frequency; lower row: energy, ๐, ๐
Figure 15. Monogenic signal and response at the third level of filtering(๐ = 3). (a) Gabor and (b) LogGabor filters.
6.5. Phase Symmetry
As an example of many possible applications, we show the computation of themain axis in one orientation using the ๐ phase. We use the information of Fig-ure 21 to determine a threshold. A reflection operation can be used to measurethe symmetry.
78 E.U. Moya-Sanchez and E. Bayro-Corrochano
(c) Upper row: up, Hx and Hy frequency; lower row: energy, ๐, ๐
Figure 15. Monogenic signal and response at the third level of filtering(๐ = 3). (c) up filters.
Figure 16. From left to right: circle; edges.
6.6. Multiresolution and Steerable Filters: qup(๐, ๐)
In this section, we present multiscale and steerable filter results. Figure 22 shows acircle and its multiscale processing. In the left column, we see a multiscale filtering.In this figure, the ๐ phase responds better to diagonal lines, and we can see coarseto fine details. In the right group of the figure, we can see the effect of differentorientations of our steerable filters.
7. Conclusions
In this work, we have presented image processing and analysis using an atomicfunction and the quaternionic phase concept. As we have shown, the atomic func-tion up(๐ฅ) has shown potential in image processing as a building block to build
4. Quaternionic Local Phase 79
100 200 300 400 500 600 7000.2
0
0.2
0.4
0.6
0.8
1
Top Center Bottom
20 40 60 80 100 1200
0.5
1
1.5
2
2.5
3
3.5Top Center Bottom
100 200 300 400 500 600 700
0
20 40 60 80 100 1200.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
100 200 300 400 500 600 700
1.5708
0
1.5708
20 40 60 80 100 1200.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
100 200 300 400 500 600 700
1.5708
0
1.5708
20 40 60 80 100 1200.3
0.2
0.1
0
0.1
0.2
0.3
0.4
Figure 17. Left: quaternionic phase profiles of a circle. Top, centreand bottom phase profiles are shown. Right: quaternionic phase profilesof edge images. Top, centre and bottom phase profiles are shown.
multiple operations that can be done analytically such as low-pass filtering, deriva-tives, local phase, and multiscale and steering filters. We have shown that thefunction qup(๐ฅ, ๐ฆ) is useful to detect lines or edges in a specific orientation usingthe quaternionic phase concept. Additionally, an oriented texture can be chosen
80 E.U. Moya-Sanchez and E. Bayro-Corrochano
Figure 18. From left to right: image, magnitude, and the quaternionicphases ๐, ๐ and ๐. We can see how the ๐ phase responds to the rotatedobject.
Figure 19. From left to right: image, magnitude, and the quaternionicphases. We can see how the ๐ phase responds to the rotated object.
using the quaternionic phases. As an initial step, we have shown how to do the im-age analysis of geometric objects in โ2 using the symmetry response of the phase.As in other applications of geometric algebra we can take advantage of the con-straints. Since the information from the three phases is independent of illuminationchanges, algorithms using the quaternionic atomic function can be less sensitivethan other methods based on the illumination changes. These results motivatedus to find other invariants such as rotation invariants using the Riesz transform.In future work, we expect to develop a complete computer vision approach basedon geometric algebra.
4. Quaternionic Local Phase 81
Figure 20. Image (left) and 2D quaternionic phases (left to right: ๐,๐, ๐). A texture based on lines can be detected or discriminated, andat the same time, the phase information can highlight the edges.
Figure 21. Circle image. The symmetry behaviour of the phase is shown.
References
[1] E. Bayro-Corrochano. The theory and use of the quaternion wavelet transform. Jour-nal of Mathematical Imaging and Vision, 24:19โ35, 2006.
[2] J. Bernd. Digital Image Processing. Springer-Verlag, New York, 1993.
[3] J. Bigun. Vision with Direction. Springer, 2006.
[4] F. Brackx, R. Delanghe, and F. Sommen. Clifford Analysis, volume 76. Pitman,Boston, 1982.
[5] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[6] M. Felsberg. Low-Level Image Processing with the Structure Multivector. PhD thesis,Christian-Albrechts-Universitat, Institut fur Informatik und Praktische Mathematik,Kiel, 2002.
[7] M. Felsberg and G. Sommer. The monogenic signal. IEEE Transactions on SignalProcessing, 49(12):3136โ3144, Dec. 2001.
[8] A.S. Gorshkov, V.F. Kravchenko, and V.A. Rvachev. Estimation of the discrete de-rivative of a signal on the basis of atomic functions. Izmeritelnaya Tekhnika, 1(8):10,1992.
[9] G.H. Granlund and H. Knutsson. Signal Processing for Computer Vision. Kluwer,Dordrecht, 1995.
82 E.U. Moya-Sanchez and E. Bayro-Corrochano
(a)
(b)
Figure 22. Multiresolution and steerable filters. Columns left to right:image, magnitude, quaternionic phases ๐, ๐, ๐. (a) multiscale approach.(b) steering filters top to bottom 0โ, 45โ, 90โ convolved with the image.
4. Quaternionic Local Phase 83
[10] V.M. Kolodyazhnya and V.A. Rvachov. Atomic functions: Generalization to themultivariable case and promising applications. Cybernetics and Systems Analysis,46(6), 2007.
[11] P. Kovesi. Invariant Measures of Image Features from Phase Information. PhD the-sis, University of Western Australia, Australia, 1996.
[12] V. Kravchenko, V. Ponomaryov, and H. Perez-Meana. Adaptive digital processing ofmultidimensional signals with applications. Moscow Fizmatlit, Moscow, 2010.
[13] E. Moya-Sanchez and E. Bayro-Corrochano. Quaternion atomic function waveletfor applications in image processing. In I. Bloch and R. Cesar, editors, Progress inPattern Recognition, Image Analysis, Computer Vision, and Applications, volume6419 of Lecture Notes in Computer Science, pages 346โ353. Springer, 2010.
[14] E. Moya-Sanchez and E. Vazquez-Santacruz. A geometric bio-inspired model forrecognition of low-level structures. In T. Honkela, W. Duch, M. Girolami, andS. Kaski, editors, Artificial Neural Networks and Machine Learning โ ICANN 2011,volume 6792 of Lecture Notes in Computer Science, pages 429โ436. Springer, 2011.
[15] M.N. Nabighian. Toward a three-dimensional automatic interpretation of potentialfield data via generalized Hilbert transforms: Fundamental relations. Geophysics,49(6):780โ786, June 1982.
[16] S. Petermichl. Dyadic shifts and logarithmic estimate for Hankel operators withmatrix symbol. Comptes Rendus de lโAcademie des Sciences, 330(6):455โ460, 2000.
[17] S. Petermichl, S. Treil, and A. Volberg. Why are the Riesz transforms averages of thedyadic shifts? Publicacions matematiques, 46(Extra 1):209โ228, 2002. Proceedingsof the 6th International Conference on Harmonic Analysis and Partial DifferentialEquations, El Escorial (Madrid), 2002.
[18] I.W. Selesnick, R.G. Baraniuk, and N.G. Kingsbury. The dual-tree complex wavelettransform. IEEE Signal Processing Magazine, 22(6):123โ151, Nov. 2005.
[19] B. Svensson. A Multidimensional Filtering Framework with Applications to Lo-cal Structure Analysis and Image Enhancement. PhD thesis, Linkoping University,Linkooping, Sweden, 2008.
E. Ulises Moya-Sanchez and E. Bayro-CorrochanoCINVESTAVCampus Guadalajara, Av. del Bosque 1145Colonia El Bajio, CP 45019Zapopan, Jalisco, Mexicoe-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 85โ104cโ 2013 Springer Basel
5 Bochnerโs Theorems in the Frameworkof Quaternion Analysis
S. Georgiev and J. Morais
Abstract. Let ๐(๐ฅ) be a nondecreasing function, such that ๐(โโ) = 0,๐(โ) = 1 and let us denote by โฌ the class of functions which can be repre-sented by a FourierโStieltjes integral ๐(๐ก) =
โซโโโ ๐๐๐ก๐ฅ๐๐(๐ฅ). The purpose of
this chapter is to give a characterization of the class โฌ and to give a gener-alization of the classical theorem of Bochner in the framework of quaternionanalysis.
Mathematics Subject Classification (2010). Primary 30G35; secondary 42A38.
Keywords. Quaternion analysis, quaternion Fourier transform, quaternionFourierโStieltjes integral, Bochner theorem.
1. Introduction
In a recent paper [5], we discussed special properties of the asymptotic behaviourof the quaternion Fourier transform (QFT) and provided a straightforward gener-alization of the classical BochnerโMinlos theorem to the framework of quaternionanalysis. The main objective of the present chapter is to extend, using similar tech-niques, the theorem of Bochner on Fourier integral transforms of complex-valuedfunctions of positive type to functions with values in the Hamiltonian quater-nion algebra in which the exponential function is replaced by a (noncommutative)quaternion exponential product. The FourierโStieltjes transform (FST) is a well-known generalization of the standard Fourier transform, and is frequently appliedin certain areas of theoretical and applied probability and stochastic processes con-texts. The present study aims to develop further numerical integration methodsfor solving partial differential equations in the quaternion analysis setting.
The chapter is organized as follows. Section 2 recalls the classical Bochnertheorem on the Fourier integral transforms of functions of positive type and col-lects some basic concepts in quaternion analysis. Section 3 defines and analyzesdifferent types of quaternion FourierโStieltjes transforms (QFST) and establishes
86 S. Georgiev and J. Morais
a number of their important properties. The underlying signals are continuousfunctions of bounded variation defined in โ2 and taking values on the quaternionalgebra. We proceed by proving the uniform continuity on this transform. Then,we describe the interplay between uniform continuity and quaternion distribution.Section 4 contains the main result of the chapter โ the counterpart of the Bochnertheorem for the noncommutative structure of quaternion functions (see Theorem4.3 below). To prove our Bochner theorem, we introduce the notion of positive-type quaternion function and deduce some of its characteristics, and rely on theasymptotic behaviour and other general properties of the QFST. To the best ofour knowledge this is done here for the first time. In the interests of simplicity ofpresentation, we have not extended this work to its most general form. Furtherinvestigations are in progress and will be reported in full in a forthcoming paper.
2. Preliminaries
2.1. Bochner Theorem
In this subsection, we review the classical Bochner theorem on Fourier integraltransforms, which can be found, e.g., in [1, 2].
A complex-valued function ๐(๐ก) defined on the interval (โโ,โ) is said tobe positive definite if it satisfies the following conditions:
1. ๐ is bounded and continuous on (โโ,โ);
2. ๐(โ๐ก) = ๐(๐ก), for all ๐ก;3. for any set of real numbers ๐ก1, . . . , ๐ก๐ , complex numbers ๐1, . . . , ๐๐ , and any
positive integer ๐ , the inequality
๐ โ๐=1
๐ โ๐=1
๐๐๐๐๐(๐ก๐ โ ๐ก๐) โฅ 0 is satisfied.
The following theorem is due to Bochner [1].
Theorem 2.1 (Bochner). If ๐(๐ฅ) is a nondecreasing bounded function on the in-terval (โโ,โ), and if ๐(๐ก) is defined by the Stieltjes integral
๐(๐ก) =
โซ โ
โโ๐๐๐ก๐ฅ๐๐(๐ฅ), โโ < ๐ก <โ, (2.1)
then ๐(๐ก) is a continuous function of the positive type.
It is of interest to remark at this point that the FourierโStieltjes transformof nondecreasing bounded functions can be easily seen continuous functions ofpositive type. Conversely, if ๐(๐ก) is measurable on (โโ,โ), and ๐ is of the positivetype, then there exists a nondecreasing bounded function ๐(๐ฅ) such that ๐(๐ก) isgiven by (2.1) for almost all ๐ฅ, โโ < ๐ฅ < โ. In the converse part, Bochnerassumed ๐(๐ก) to be continuous, and showed that ๐(๐ฅ) is such that (2.1) is true forall ๐ก. Riesz, on the other hand, succeeded to prove that the measurability of ๐ wassufficient in the converse.
5. Bochnerโs Theorems 87
2.2. Quaternion Analysis
The present subsection collects some basic facts about quaternions and the (left-sided) QFT, which will be needed throughout the text.
In all that follows let
โ := {๐ง = ๐+ ๐๐+ ๐๐ + ๐๐ : ๐, ๐, ๐, ๐ โ โ} (2.2)
denote the Hamiltonian skew field, i.e., quaternions, where the imaginary units ๐,๐, and ๐ are subject to the multiplication rules:
๐2 = ๐2 = ๐2 = โ1,๐๐ = ๐ = โ๐๐, ๐๐ = ๐ = โ๐๐, ๐๐ = ๐ = โ๐๐. (2.3)
Like in the complex case, S(๐ง) = ๐ and V(๐ง) = ๐๐+ ๐๐ + ๐๐ define the scalar andvector parts of ๐ง. The conjugate of ๐ง is ๐ง = ๐โ ๐๐โ ๐๐ โ ๐๐, and the norm of ๐ง isdefined by
โฃ๐งโฃ = โ๐ง๐ง =โ๐ง๐ง =
โ๐2 + ๐2 + ๐2 + ๐2, (2.4)
which coincides with the corresponding Euclidean norm of ๐ง as a vector in โ4. Forx := (๐ฅ1, ๐ฅ2) โ โ2 we consider โ-valued functions defined in โ2, i.e., functions ofthe form
๐(x) := [๐(x)]0 + [๐(x)]1๐+ [๐(x)]2๐ + [๐(x)]3๐, (2.5)
[๐ ]๐ : โ2 โโ โ (๐ = 0, 1, 2, 3).
Properties (like integrability, continuity or differentiability) that are ascribed to ๐have to be fulfilled by all its components [๐ ]๐.
Let ๐ฟ1(โ2;โ) denote the linear space of integrableโ-valued functions definedin โ2. The left-sided QFT of ๐ โ ๐ฟ1(โ2;โ) is given by [6]
โฑ(๐) : โ2 โ โ, โฑ(๐)(๐) :=
โซโ2
e(๐,x) ๐(x) ๐2x, (2.6)
where the kernel function
e : โ2 ร โ2 โ โ, e(๐,x) := ๐โ๐๐2๐ฅ2๐โ๐๐1๐ฅ1 , (2.7)
is called the (left-sided) quaternion Fourier kernel. For ๐ = 1, 2, ๐ฅ๐ will denote thespace coordinates and ๐๐ the angular frequencies. The previous definition of theQFT varies from the original one only in the fact that we use 2D vectors insteadof scalars and that it is defined to be 2D. It is of interest to remark at this pointthat the product in (2.6) has to be performed in a fixed order since, in general,e(๐,x) does not commute with every element of the algebra.
Under suitable conditions, the original signal ๐ can be reconstructed fromโฑ(๐) by the inverse transform. The (left-sided) inverse QFT of ๐ โ ๐ฟ1(โ2;โ) isgiven by
โฑโ1(๐) : โ2 โ โ, โฑโ1(๐)(x) =1
(2๐)2
โซโ2
e(๐,x) ๐(๐) ๐2๐ , (2.8)
where e(๐,x) is called the inverse (left-sided) quaternion Fourier kernel.
88 S. Georgiev and J. Morais
3. Quaternion FourierโStieltjes Transform and its Properties
This section generalizes the classical FST to Hamiltonโs quaternion algebra. Usingthe 4D analogue of the (complex-valued) function ๐(๐ฅ) described before, we extendthe FST to the QFST. As we shall see later, some properties of the FST can beextended in this context.
3.1. Quaternion FourierโStieltjes Transform
Partially motivated by the results from Bulow [3], the idea behind the constructionof a quaternion counterpart of the Stieltjes integral is to replace the exponentialfunction in (2.1) by a suitable (noncommutative) quaternion exponential product.
In the sequel, we consider the functions ๐1, ๐2 : โ โ โ of the quaternionform (2.5), such that โฃ๐๐โฃ โค ๐ฟ๐ for real numbers ๐ฟ๐ <โ (๐ = 1, 2). Due to the non-commutativity of the quaternions, we shall define three different types of QFST.
Definition 3.1. The QFST โฑ๐ฎ(๐1, ๐2) : โ2 โ โ of ๐1(๐ฅ1) and ๐2(๐ฅ2) is definedas the Stieltjes integrals:
1. Right-sided QFST:
โฑ๐ฎ๐(๐1, ๐2)(๐1, ๐2) :=
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
๐๐1๐ฅ1๐๐๐2๐ฅ2 . (3.1)
2. Left-sided QFST:
โฑ๐ฎ๐(๐1, ๐2)(๐1, ๐2) :=
โซโ2
๐๐๐2๐ฅ2๐๐๐1๐ฅ1๐๐1(๐ฅ1)๐๐2(๐ฅ2). (3.2)
3. Two-sided QFST:
โฑ๐ฎ๐ (๐1, ๐2)(๐1, ๐2) :=
โซโ2
๐๐๐1๐ฅ1๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
๐๐2๐ฅ2 . (3.3)
Remark 3.2. We remind the reader that, the order of the exponentials in (3.1)โ(3.3) is fixed because of the noncommutativity of the quaternion product. It is ofinterest to remark at this point that in case that ๐๐1๐๐2 = ๐(x)๐2x, the formulaeabove reduce to the usual definitions for the right-, left- and two-sided QFT [4, 7, 6],which only differ by the signs in the exponential kernel terms.
It is significant to note that, in practice, the integrals (3.1)โ(3.3) will alwaysexist. For example, for any two real variables ๐1, ๐2, and for real constants ๐, ๐ itholds that
โฃโฑ๐ฎ๐(๐1, ๐2)โฃ โค
โฃโฃโฃโซ ๐โโ (โซ ๐โโ+โซโ๐
)๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐๐๐1๐ฅ1๐๐๐2๐ฅ2
โฃโฃโฃ+โฃโฃโฃโซโ๐ (โซ ๐
โโ+โซโ๐
)๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐๐๐1๐ฅ1๐๐๐2๐ฅ2
โฃโฃโฃโค (โฃ๐1(โ)โ ๐1(๐)โฃ+ โฃ๐1(๐)โ ๐1(โโ)โฃ)ร (โฃ๐2(โ)โ ๐2(๐)โฃ+ โฃ๐2(๐)โ ๐2(โโ)โฃ) . (3.4)
5. Bochnerโs Theorems 89
Remark 3.3. Throughout this text we will only investigate the integral (3.1), whichwe denote for simplicity by โฑ๐ฎ(๐1, ๐2). Nevertheless, all computations can beeasily adapted to (3.2) and (3.3). In view of (3.1) and (3.2), a straightforwardcalculation shows that
โฑ๐ฎ๐(๐1, ๐2)(๐1, ๐2) =
โซโ2
๐โ๐๐2๐ฅ2๐โ๐๐1๐ฅ1๐๐1(๐ฅ1)๐๐2(๐ฅ2)
= โฑ๐ฎ ๐(๐2, ๐1)(โ๐1,โ๐2). (3.5)
For the sake of further simplicity, in the considerations to follow we will often omitthe subscript and, additionally, write only โฑ๐ฎ(๐1,๐2) instead of โฑ๐ฎ(๐1,๐2)(๐1,๐2).We can restate the definition of right-sided QFST (3.1) in equivalent terms as fol-lows.
Lemma 3.4. The (right-sided ) QFST has the closed-form representation
โฑ๐ฎ(๐1, ๐2) := [ฮฆ(๐1, ๐2)]0 + [ฮฆ(๐1, ๐2)]1 + [ฮฆ(๐1, ๐2)]2 + [ฮฆ(๐1, ๐2)]3 , (3.6)
where we used the integrals
[ฮฆ(๐1, ๐2)]0 =
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2) cos(๐1๐ฅ1) cos(๐2๐ฅ2), (3.7)
[ฮฆ(๐1, ๐2)]1 =
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ sin(๐1๐ฅ1) cos(๐2๐ฅ2), (3.8)
[ฮฆ(๐1, ๐2)]2 =
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ cos(๐1๐ฅ1) sin(๐2๐ฅ2), (3.9)
[ฮฆ(๐1, ๐2)]3 =
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2). (3.10)
Corollary 3.5. The (right-sided) QFST satisfies the following identities:
โฑ๐ฎ(๐1, ๐2) + โฑ๐ฎ(๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2) + ฮฆ1(๐1, ๐2)) , (3.11)
โฑ๐ฎ(๐1, ๐2)โโฑ๐ฎ(๐1,โ๐2) = 2 (ฮฆ2(๐1, ๐2) + ฮฆ3(๐1, ๐2)) , (3.12)
โฑ๐ฎ(๐1, ๐2) + โฑ๐ฎ(โ๐1, ๐2) = 2 (ฮฆ0((๐1, ๐2) + ฮฆ2(๐1, ๐2)) , (3.13)
โฑ๐ฎ(๐1, ๐2)โโฑ๐ฎ(โ๐1, ๐2) = 2 (ฮฆ1(๐1, ๐2) + ฮฆ3(๐1, ๐2)) , (3.14)
โฑ๐ฎ(๐1, ๐2) + โฑ๐ฎ(โ๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2) + ฮฆ3(๐1, ๐2)) , (3.15)
โฑ๐ฎ(๐1, ๐2)โโฑ๐ฎ(โ๐1,โ๐2) = 2 (ฮฆ1(๐1, ๐2) + ฮฆ2(๐1, ๐2)) , (3.16)
โฑ๐ฎ(๐1,โ๐2) + โฑ๐ฎ(โ๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2)โ ฮฆ2(๐1, ๐2)) , (3.17)
โฑ๐ฎ(๐1,โ๐2)โโฑ๐ฎ(โ๐1,โ๐2) = 2 (ฮฆ1(๐1, ๐2)โ ฮฆ3(๐1, ๐2)) , (3.18)
โฑ๐ฎ(โ๐1, ๐2) + โฑ๐ฎ(โ๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2)โ ฮฆ1(๐1, ๐2)) , (3.19)
โฑ๐ฎ(โ๐1, ๐2)โโฑ๐ฎ(โ๐1,โ๐2) = 2 (ฮฆ2(๐1, ๐2)โ ฮฆ3(๐1, ๐2)) . (3.20)
3.2. Properties
This subsection presents certain properties of the asymptotic behaviour of theQFST, and establishes some of its basic properties.
90 S. Georgiev and J. Morais
We begin with the notions of monotonic increasing, bounded variation anddistribution in the context of quaternion analysis.
Definition 3.6. A function ๐ : โ โ โ is called monotonic increasing, if all of itscomponents [๐]๐ (๐ = 0, 1, 2, 3) are monotonic increasing functions.
Definition 3.7. A function ๐ : โ โ โ is called with bounded variation on โ, ifthere exists a real number ๐ <โ such that
โซโโฃ๐๐(๐ฅ)โฃ < ๐ .
Definition 3.8. A function ๐ : โ โ โ is said to be a quaternion distribution, if itis of bounded variation and monotonic increasing, and if the following limits exist
lim๐ฅโโ๐ฆ+
๐(๐ฅ) = ๐(๐ฆ + 0), and lim๐ฅโโ๐ฆโ๐(๐ฅ) = ๐(๐ฆ โ 0), (3.21)
(taken over all directions) for which
๐(๐ฆ) =1
2[๐(๐ฆ + 0) + ๐(๐ฆ โ 0)] (3.22)
holds almost everywhere on โ.
To proceed with, it is significant to note that, for every two functions ๐1, ๐2 :โโ โ of the quaternion form (2.5), the study of the properties of the distribution
โฑ๐ฎ(๐1, ๐2) :=
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
๐๐1๐ฅ1๐๐๐2๐ฅ2 (3.23)
is reduced to the separate study of each componentโซโ2
[๐๐1(๐ฅ1)
]๐
[๐๐2(๐ฅ2)
]๐
๐๐๐1๐ฅ1๐๐๐2๐ฅ2 (๐,๐ = 0, 1, 2, 3) . (3.24)
From now on, we denote the class of functions which can be represented as (3.23)by โฌ. Functions in โฌ are called (right) quaternion Bochner functions and โฌ willbe referred to as the (right) quaternion Bochner set. It follows that all membersof โฌ are entire functions of the real variables ๐1, ๐2.
Firstly, we shall note the following property of โฌ:Proposition 3.9. โฌ is a linear space.
Proof. Let ๐, ๐ โ โฌ and ๐ง1, ๐ง2, ๐ง3, ๐ง4 be quaternion numbers. It is easily seen that
๐ง1๐๐ง2 + ๐ง3๐๐ง4 โ โฌ. (3.25)
โกThe foregoing discussion suggests the computation of certain equalities for
quaternion Bochner functions. If ๐ โ โฌ is given, then
๐(0, ๐2) =(๐1(โ)โ ๐1(โโ)
) โซโ
๐๐2(๐ฅ2)๐๐๐2๐ฅ2 , (3.26)
๐(๐1, 0) =
โซโ
๐๐1(๐ฅ1)(๐2(โ)โ ๐2(โโ)
)๐๐๐1๐ฅ1 , (3.27)
๐(0, 0) =(๐1(โ)โ ๐1(โโ)
) (๐2(โ)โ ๐2(โโ)
). (3.28)
5. Bochnerโs Theorems 91
In particular, a simple argument gives
๐(โ๐1,โ๐2) =
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2
=
โซโ2
๐๐๐2๐ฅ2๐๐๐1๐ฅ1๐๐1(๐ฅ1)๐๐2(๐ฅ2)
= ๐(๐1, ๐2), (3.29)
where ๐ is any function which can be represented as โฑ๐ฎ ๐(๐2, ๐1)(๐1, ๐2). We nowhave the following property:
Proposition 3.10. Every element of โฌ is a continuous bounded function.
Proof. Let ๐ be any function in โฌ. The continuity of ๐ is obvious. For the bound-edness a direct computation shows that
โฃ๐(๐1, ๐2)โฃ โคโซโ2
โฃ๐๐1(๐ฅ1)๐๐2(๐ฅ2)โฃ โฃ๐๐๐1๐ฅ1๐๐๐2๐ฅ2 โฃ (3.30)
โค โฃ๐1(โ)โ ๐1(โโ)โฃ โฃ๐2(โ)โ ๐2(โโ)โฃ . โก
Suppose now that ๐1, ๐2 : โ โ โ are fixed, and ๐ โ โฌ does not vanishidentically. Question: Depending on the parity of ๐1(๐ฅ1) and ๐2(๐ฅ2), what can wesay about the representation of ๐? If there is any similarity or difference in theparity of these functions, does this lead to some particular functions ๐ โ โฌ? Theanswers are given by the following proposition:
Proposition 3.11. Let ๐1, ๐2 : โโ โ, and ๐ โ โฌ be given.
1. If ๐1(๐ฅ1) is an odd function then
๐(๐1, ๐2) = 2
โซ โโโ
โซ โ
0
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ sin(๐1๐ฅ1)๐
๐๐2๐ฅ2 ,
๐(0, ๐2) = 0. (3.31)
2. If ๐1(๐ฅ1) is an even function then
๐(๐1, ๐2) = 2
โซ โโโ
โซ โ
0
๐๐1(๐ฅ1)๐๐2(๐ฅ2) cos(๐1๐ฅ1)๐
๐๐2๐ฅ2 ,
๐(0, 0) = 2(๐1(โ)โ ๐1(0)
) (๐2(โ)โ ๐2(โโ)
). (3.32)
3. If ๐2(๐ฅ2) is an odd function then
๐(๐1, ๐2) = 2
โซ โ0
โซ โ
โโ๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐๐๐1๐ฅ1 sin(๐2๐ฅ2)๐,
๐(๐1, 0) = 0. (3.33)
92 S. Georgiev and J. Morais
4. If ๐2(๐ฅ2) is an even function then
๐(๐1, ๐2) = 2
โซ โ
0
โซ โ
โโ๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐๐๐1๐ฅ1 cos(๐2๐ฅ2),
๐(0, 0) = 2(๐1(โ) โ ๐1(โโ)
) (๐2(โ)โ ๐2(0)
). (3.34)
5. If ๐1(๐ฅ1) and ๐2(๐ฅ2) are both odd functions then
๐(๐1, ๐2) = 4
โซ โ0
โซ โ
0
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2),
๐(๐1, 0) = ๐(0, ๐2) = 0. (3.35)
6. If ๐1(๐ฅ1) is an odd function and ๐2(๐ฅ2) an even function then
๐(๐1, ๐2) = 4
โซ โ
0
โซ โ
0
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ sin(๐1๐ฅ1) cos(๐2๐ฅ2),
๐(0, ๐2) = 0. (3.36)
7. If ๐1(๐ฅ1) is an even function and ๐2(๐ฅ2) an odd function then
๐(๐1, ๐2) = 4
โซ โ0
โซ โ
0
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ cos(๐1๐ฅ1) sin(๐2๐ฅ2),
๐(๐1, 0) = 0. (3.37)
8. If ๐1(๐ฅ1) and ๐2(๐ฅ2) are both even functions then
๐(๐1, ๐2) = 4
โซ โ
0
โซ โ
0
๐๐1(๐ฅ1)๐๐2(๐ฅ2) cos(๐1๐ฅ1) cos(๐2๐ฅ2),
๐(0, 0) = 4(๐1(โ)โ ๐1(0)
) (๐2(โ)โ ๐2(0)
). (3.38)
Proof. For sake of brevity we prove only the first statement, the proof of theremaining statements is similar. Let ๐1(๐ฅ1) be an odd function. Using the trigono-metric representation of the function ๐๐๐1๐ฅ1 , a straightforward computation showsthat
๐(๐1, ๐2) =
โซ โ
โโ
โซ โ
โโ๐๐1(๐ฅ1)๐๐
2(๐ฅ2) ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
=
โซ โ
โโ
(โซ โ
โโ๐๐1(๐ฅ1) cos(๐1๐ฅ1)
)๐๐2(๐ฅ2) ๐
๐๐2๐ฅ2
+
โซ โ
โโ
(โซ โ
โโ๐๐1(๐ฅ1) sin(๐1๐ฅ1)
)๐๐2(๐ฅ2) ๐ ๐
๐๐2๐ฅ2 (3.39)
= 2
โซ โ
โโ
โซ โ
0
๐๐1(๐ฅ1)๐๐2(๐ฅ2) ๐ sin(๐1๐ฅ1) ๐
๐๐2๐ฅ2 . โก
3.3. Uniform Continuity
In this subsection we discuss uniform continuity and its relationship to the QFST.We begin by defining uniform continuity.
5. Bochnerโs Theorems 93
Definition 3.12. A quaternion function ๐ : ฮฉ โ โ2 โ โ is uniformly continuouson ฮฉ, if and only if for all ๐ > 0 there exists a ๐ฟ > 0, such that โฃ๐(๐1)โ ๐(๐2)โฃ < ๐for all ๐1, ๐2 โ ฮฉ whenever โฃ๐1 โ ๐2โฃ < ๐ฟ.
We now prove some results related to the asymptotic behaviour of the QFST.
Proposition 3.13. Let ๐ be an element of โฌ. For any natural number ๐, let ๐๐ :โร [โ๐, ๐]โ โ be the function given by
๐๐(๐1, ๐2) =
โซ ๐
โ๐
โซ โ
โโ๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐๐๐1๐ฅ1๐๐๐2๐ฅ2 .
Then ๐๐(๐1, ๐2) โโ๐โโโ ๐(๐1, ๐2) uniformly. Also, if the ๐๐ are uniformly con-tinuous functions, then ๐ is also a uniformly continuous function.
Proof. A first straightforward computation shows that
โฃ๐(๐1, ๐2)โ ๐๐(๐1, ๐2)โฃ
=
โฃโฃโฃโฃ โโซโโโโซโโ
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
๐๐1๐ฅ1๐๐๐2๐ฅ2 โ๐โซโ๐
โโซโโ
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
๐๐1๐ฅ1๐๐๐2๐ฅ2
โฃโฃโฃโฃ=
โฃโฃโฃโฃ โ๐โซโโโโซโโ
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
๐๐1๐ฅ1๐๐๐2๐ฅ2 +โโซ๐
โโซโโ
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
๐๐1๐ฅ1๐๐๐2๐ฅ2
โฃโฃโฃโฃโค โฃ๐1(โ)โ ๐1(โโ)โฃ โฃ๐2(โ๐)โ ๐2(โโ)โฃ+ โฃ๐1(โ)โ ๐1(โโ)โฃ โฃ๐2(โ)โ ๐2(๐)โฃ .
(3.40)
Moreover, bearing in mind that
๐2(โ๐)โ ๐2(โโ) โโ๐โโโ 0, ๐2(โ) โ ๐2(๐) โโ๐โโโ 0, (3.41)
and hence, ๐๐(๐1, ๐2) โโ๐โโโ ๐(๐1, ๐2) uniformly. In addition, we claim that ifall the ๐๐(๐1, ๐2) are uniformly continuous functions, it follows that ๐(๐1, ๐2) isalso a uniformly continuous function. โก
Likewise, we have the following analogous result.
Proposition 3.14. Let ๐ be an element of โฌ. For any natural number ๐, let ๐๐ :[โ๐, ๐]ร โโ โ be the function given by
๐๐(๐1, ๐2) =
โซ โ
โโ
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2) ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 .
Then ๐๐(๐1, ๐2) โโ๐โโโ ๐(๐1, ๐2) uniformly. Also, if all the ๐๐ are uniformlycontinuous functions then ๐ is also a uniformly continuous function.
94 S. Georgiev and J. Morais
Proposition 3.15. Let ๐ be an element of โฌ. For any natural number ๐, let ๐๐ :[โ๐, ๐]ร [โ๐, ๐]โ โ be the function given by
๐๐(๐1, ๐2) =
โซ ๐
โ๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2) ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 .
Then ๐๐(๐1, ๐2) โโ๐โโโ ๐(๐1, ๐2) uniformly. Also, if all the ๐๐ are uniformlycontinuous functions then ๐ is also a uniformly continuous function.
Proof. We set ๐ด := ๐๐1(๐ฅ1)๐๐2(๐ฅ2) ๐
๐๐1๐ฅ1 ๐๐๐2๐ฅ2 . The proof follows from a simpleobservation:โซ โ
โโ
โซ โ
โโ๐ดโ
โซ ๐
โ๐
โซ ๐
โ๐๐ด =
โซ โ๐
โโ
โซ โ๐
โโ๐ด+
โซ โ๐
โโ
โซ ๐
โ๐๐ด+
โซ โ๐
โโ
โซ โ
๐
๐ด
+
โซ ๐
โ๐
โซ โ๐
โโ๐ด+
โซ ๐
โ๐
โซ โ
๐
๐ด+
โซ โ
๐
โซ โ๐
โโ๐ด (3.42)
+
โซ โ
๐
โซ ๐
โ๐๐ด+
โซ โ
๐
โซ โ
๐
๐ด. โก
We come now to the main theorem of this section.
Theorem 3.16. Let ๐ โ โฌ be given, and ๐ : โ2 โโ โ be a continuous and absolutelyintegrable function. For any ๐1, ๐2 : โ โโ โ the following relations hold:
1.
โซโ2
๐(๐ก1, ๐ก2)๐(๐1 โ ๐ก1, ๐ โ ๐2)๐๐ก1๐๐ก2
=
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)
โซโ2
๐๐๐ก1๐ฅ1๐๐๐ก2๐ฅ2๐(๐1 โ ๐ก1, ๐ โ ๐2)๐๐ก1๐๐ก2; (3.43)
2.
โซโ2
๐(๐ก1, ๐ก2)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2
= (2๐)2โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)โฑโ1(๐)(๐ฅ1, ๐ฅ2). (3.44)
Proof. Assume ๐ : โ2 โโ โ to be a continuous and absolutely integrable function.For any real variables ๐ and ๐ we define the function
๐ (๐ฅ1, ๐ฅ2, ๐, ๐) :=
โซ ๐
โ๐
โซ ๐
โ๐๐๐๐ก1๐ฅ1๐๐๐ก2๐ฅ2๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2. (3.45)
We set ๐๐(๐ก1, ๐ก2) =โซ ๐โ๐โซ ๐โ๐ ๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐๐๐ก1๐ฅ1๐๐๐ก2๐ฅ2 . Using the fact that ๐ is an
absolutely integrable function, it follows thatโซ ๐
โ๐
โซ ๐
โ๐๐๐(๐ก1, ๐ก2)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2
=
โซ ๐
โ๐
โซ ๐
โ๐
(โซ ๐
โ๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐๐๐ก1๐ฅ1๐๐๐ก2๐ฅ2
)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2
5. Bochnerโs Theorems 95
=
โซ ๐
โ๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)
โซ ๐
โ๐
โซ ๐
๐
๐๐๐ก1๐ฅ1๐๐๐ก2๐ฅ2๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2
=
โซ ๐
โ๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐ (๐ฅ1, ๐ฅ2, ๐, ๐). (3.46)
From the last proposition we know that lim๐โโโ ๐๐(๐1, ๐2) = ๐(๐1, ๐2) converges
uniformly. Moreover, since ๐(๐ก1, ๐ก2) is an absolutely integrable function, it followsthat ๐ (๐ฅ1, ๐ฅ2, ๐, ๐) is also a uniformly continuous function. Hence
lim๐โโโ
๐โซโ๐
๐โซโ๐
๐๐(๐ก1, ๐ก2)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2 =
๐โซโ๐
๐โซโ๐
๐(๐ก1, ๐ก2)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2. (3.47)
With this argument in hand, and based on (3.46) we conclude that
lim๐โโโ
โซ ๐
โ๐
โซ ๐
โ๐๐๐(๐ก1, ๐ก2)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2 =
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ (๐ฅ1, ๐ฅ2, ๐, ๐).
(3.48)From the last equality and from (3.47) we obtainโซ
โ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ (๐ฅ1, ๐ฅ2, ๐, ๐) =
โซ ๐
โ๐
โซ ๐
๐
๐(๐ก1, ๐ก2)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2. (3.49)
In addition, we have
๐ (๐ฅ1, ๐ฅ2, ๐, ๐) โโ ๐ โโโ๐ โโโ
โซโ2
๐๐๐ก1๐ฅ1 ๐๐๐ก2๐ฅ2๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2, (3.50)
and hence, for any fixed ๐, ๐ > 0 it follows thatโซ ๐
โ๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐ (๐ฅ1, ๐ฅ2, ๐, ๐) (3.51)
โโ ๐ โโโ๐ โโโ
โซ ๐
โ๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)
โซโ2
๐๐๐ก1๐ฅ1 ๐๐๐ก2๐ฅ2๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2.
For the sake of brevity, in the considerations to follow we will often omit the argu-ment and write simply ๐ instead of ๐ (๐ฅ1, ๐ฅ2, ๐, ๐). Since ๐ is uniformly bounded,there exists a positive constant ๐ so that โฃ๐ โฃ โค ๐ for all ๐ฅ1, ๐ฅ2, ๐, ๐ โ โ. Wedefine
๐ผ :=
โฃโฃโฃโฃโฃโซ โ๐
โโ
โซ โ๐
โโ๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐ +
โซ ๐
โ๐
โซ โ๐
โโ๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐
+
โซ โ
๐
โซ โ๐
โโ๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐ +
โซ โ๐
โโ
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐
96 S. Georgiev and J. Morais
+
โซ โ
๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐ +
โซ โ๐
โโ
โซ โ
๐
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
+
โซ ๐
โ๐
โซ โ
๐
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐ +
โซ โ
๐
โซ โ
๐
๐๐1(๐ฅ1)๐๐2(๐ฅ2)๐
โฃโฃโฃโฃ . (3.52)
Therefore, we obtain
๐ผ โค๐ [โฃ๐1(โ๐)โ ๐1(โโ)โฃ โฃ๐2(โ๐)โ ๐2(โโ)โฃ+ โฃ๐1(โ๐)โ ๐1(โโ)โฃ โฃ๐2(๐)โ ๐2(โ๐)โฃ+ โฃ๐1(โ๐)โ ๐1(โโ)โฃ โฃ๐2(โ)โ ๐2(๐)โฃ+ โฃ๐1(๐)โ ๐1(โ๐)โฃ โฃ๐2(โ๐)โ ๐2(โโ)โฃ+ โฃ๐1(๐)โ ๐1(โ๐)โฃ โฃ๐2(โ)โ ๐2(๐)โฃ+ โฃ๐1(โ)โ ๐1(๐)โฃ โฃ๐2(โ๐)โ ๐2(โโ)โฃ+ โฃ๐1(โ)โ ๐1(๐)โฃ โฃ๐2(๐)โ ๐2(โ๐)โฃ+ โฃ๐1(โ)โ ๐1(๐)โฃ โฃ๐2(โ)โ ๐2(๐)โฃ] โโ๐,๐โโโ 0. (3.53)
Using the last inequality and (3.51) we get
lim๐,๐โโโ
โซ ๐
โ๐
โซ ๐
โ๐๐๐1(๐ฅ1)๐๐
2(๐ฅ2)๐ (๐ฅ1, ๐ฅ2, ๐, ๐) (3.54)
=
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)
โซโ2
๐๐๐ก1๐ฅ1 ๐๐๐ก2๐ฅ2๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2.
Based on this and with (3.48) we findโซโ2
๐(๐ก1, ๐ก2)๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2
=
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)
โซโ2
๐๐๐ก1๐ฅ1 ๐๐๐ก2๐ฅ2๐(๐ก1, ๐ก2)๐๐ก1๐๐ก2
= (2๐)2โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)โฑโ1(๐)(๐ฅ1, ๐ฅ2). (3.55)
Making the change of variables ๐ก1 โโ ๐1โ ๐ก1, ๐ก2 โโ ๐โ๐2 in the definition of ๐,we finally findโซ
โ2
๐(๐ก1, ๐ก2)๐(๐1 โ ๐ก1, ๐ โ ๐2)๐๐ก1๐๐ก2 (3.56)
=
โซโ2
๐๐1(๐ฅ1)๐๐2(๐ฅ2)
โซโ2
๐๐๐ก1๐ฅ1 ๐๐๐ก2๐ฅ2๐(๐1 โ ๐ก1, ๐ โ ๐2)๐๐ก1๐๐ก2. โก
Theorem 3.17. For any ๐1, ๐2 : โ โโ โ, consider the functions
๐(๐1) =
โซโ
๐๐1(๐ฅ1) ๐๐๐1๐ฅ1 , โ(๐2) =
โซโ
๐๐2(๐ฅ2) ๐๐๐2๐ฅ2 . (3.57)
5. Bochnerโs Theorems 97
Then for any real number ๐ the following equalities hold:
๐1(๐)โ ๐1(0) =1
2๐
โซโ
๐(๐1)๐โ๐๐๐1 โ 1
โ๐๐1๐๐1, (3.58)
๐2(๐)โ ๐2(0) =1
2๐
โซโ
โ(๐2)๐โ๐๐๐2 โ 1
โ๐๐2๐๐2. (3.59)
Proof. We begin with the following observation:
๐(๐1)(๐โ๐๐๐1 โ 1
)=
โซโ
๐๐1(๐ฅ1) ๐๐๐1๐ฅ1
(๐โ๐๐๐1 โ 1
)=
โซโ
๐๐1(๐ฅ1) ๐๐๐1(๐ฅ1โ๐) โ
โซโ
๐๐1(๐ฅ1) ๐๐๐1๐ฅ1
=
โซโ
๐๐1(๐ฅ1 + ๐) ๐๐๐1๐ฅ1 โโซโ
๐๐1(๐ฅ1) ๐๐๐1๐ฅ1
= lim๐โโโ
(โซ ๐
โ๐๐๐1(๐ฅ1 + ๐) ๐๐๐1๐ฅ1 โ
โซ ๐
โ๐๐๐1(๐ฅ1) ๐
๐๐1๐ฅ1
)= lim
๐โโโ
[๐1(๐+ ๐)๐๐๐๐1 โ ๐1(โ๐+ ๐)๐โ๐๐๐1 โ ๐1(๐)๐๐๐๐1
+ ๐1(โ๐)๐โ๐๐๐1 โโซ ๐
โ๐
(๐1(๐ฅ1 + ๐)โ ๐1(๐ฅ1)
)๐๐ฅ1 ๐
๐๐1๐ฅ1๐๐1
]= โ
โซโ
(๐1(๐ฅ1 + ๐)โ ๐1(๐ฅ1)
)๐๐ฅ1 ๐
๐๐1๐ฅ1๐๐1.
(3.60)
Therefore, it is easy to see that
๐(๐1)(๐โ๐๐๐1 โ 1
) 1
โ๐๐1=
โซโ
(๐1(๐ฅ1 + ๐)โ ๐1(๐ฅ1)
)๐๐ฅ1 ๐
๐๐1๐ฅ1 . (3.61)
From the last equality and from the inverse Fourier transform formula we find that
๐1(๐ฅ1 + ๐)โ ๐1(๐ฅ1) =1
2๐
โซโ
๐(๐1)๐โ๐๐๐1 โ 1
โ๐๐1๐โ๐๐ฅ1๐1๐๐1, (3.62)
and, in particular for ๐ฅ1 = 0 we find that
๐1(๐)โ ๐1(0) =1
2๐
โซโ
๐(๐1)๐โ๐๐๐1 โ 1
โ๐๐1๐๐1. (3.63)
In a similar way we may deduce that
๐2(๐)โ ๐2(0) =1
2๐
โซโ
โ(๐1)๐โ๐๐๐1 โ 1
โ๐๐1๐๐1 , (3.64)
which completes the proof. โก
98 S. Georgiev and J. Morais
3.4. Connecting Uniform Continuity to Quaternion Distributions
In this subsection we describe the connection between uniform continuity andquaternion distributions.
Theorem 3.18. Let the sequences of quaternion distributions
๐1(๐ฅ1), ๐2(๐ฅ1), . . . , ๐
๐(๐ฅ1), . . . , ๐1(๐ฅ2), ๐2(๐ฅ2), . . . , ๐
๐(๐ฅ2), . . . (3.65)
be convergent to the distributions ๐0(๐ฅ1) and ๐0(๐ฅ2), respectively. Assume that
lim๐โโโ๐๐(ยฑโ) = ๐0(ยฑโ), lim
๐โโโ ๐๐(ยฑโ) = ๐0(ยฑโ), (3.66)
where ๐๐(๐1, ๐2) โ โฌ corresponds to the distributions ๐๐(๐ฅ1) and ๐๐(๐ฅ2), and๐0(๐1, ๐2) corresponds to the distributions ๐0(๐ฅ1) and ๐0(๐ฅ2), respectively. Then
lim๐โโโ ๐๐(๐1, ๐2) = ๐0(๐1, ๐2).
Proof. To begin with, we choose ๐ > 0 so that ๐0(๐ฅ1) and ๐0(๐ฅ2) are continuousfunctions at ยฑ๐. Then
lim๐โโโ๐๐(ยฑ๐) = ๐0(ยฑ๐), lim
๐โโโ ๐๐(ยฑ๐) = ๐0(ยฑ๐). (3.67)
For brevity of exposition, in the sequel we omit the arguments and write simply๐๐๐ and ๐๐๐ instead of ๐๐๐(๐ฅ1) and ๐๐๐(๐ฅ2), respectively. Now, we make use ofthe following representations for ๐๐(๐1, ๐2) and ๐0(๐1, ๐2):
๐๐(๐1, ๐2) =
โซโ2
๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
=
โซ ๐
โ๐
โซ ๐
โ๐๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ โ๐
โโ
โซ โ๐
โโ๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
+
โซ โ๐
โโ
โซ ๐
โ๐๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ โ๐
โโ
โซ โ
๐
๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
+
โซ ๐
โ๐
โซ โ๐
โโ๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ ๐
โ๐
โซ โ
๐
๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
+
โซ โ
๐
โซ โ๐
โโ๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ โ
๐
โซ ๐
โ๐๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
+
โซ โ
๐
โซ โ
๐
๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.68)
and,
๐0(๐1, ๐2) =
โซโ2
๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
=
โซ ๐
โ๐
โซ ๐
โ๐๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ โ๐
โโ
โซ โ๐
โโ๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
+
โซ โ๐
โโ
โซ ๐
โ๐๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ โ๐
โโ
โซ โ
๐
๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
5. Bochnerโs Theorems 99
+
โซ ๐
โ๐
โซ โ๐
โโ๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ ๐
โ๐
โซ โ
๐
๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
+
โซ โ
๐
โซ โ๐
โโ๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 +
โซ โ
๐
โซ ๐
โ๐๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2
+
โซ โ
๐
โซ โ
๐
๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 . (3.69)
Furthermore, we set
๐ผ1(๐1, ๐2, ๐)
=
โซ ๐
โ๐
โซ ๐
โ๐๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โ
โซ ๐
โ๐
โซ ๐
โ๐๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.70)
๐ผ2(๐1, ๐2, ๐)
=
โซ โ๐
โโ
โซ โ๐
โโ๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โ
โซ โ๐
โโ
โซ โ๐
โโ๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.71)
๐ผ3(๐1, ๐2, ๐)
=
โซ โ๐
โโ
โซ ๐
โ๐๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โ
โซ โ๐
โโ
โซ ๐
โ๐๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.72)
๐ผ4(๐1, ๐2, ๐)
=
โซ โ๐
โโ
โซ โ
๐
๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โโซ โ๐
โโ
โซ โ
๐
๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.73)
๐ผ5(๐1, ๐2, ๐)
=
โซ ๐
โ๐
โซ โ๐
โโ๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โ
โซ ๐
โ๐
โซ โ๐
โโ๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.74)
๐ผ6(๐1, ๐2, ๐)
=
โซ ๐
โ๐
โซ โ
๐
๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โโซ ๐
โ๐
โซ โ
๐
๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.75)
๐ผ7(๐1, ๐2, ๐)
=
โซ โ
๐
โซ โ๐
โโ๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โ
โซ โ
๐
โซ โ๐
โโ๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.76)
๐ผ8(๐1, ๐2, ๐)
=
โซ โ
๐
โซ ๐
โ๐๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โ
โซ โ
๐
โซ ๐
โ๐๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 , (3.77)
๐ผ9(๐1, ๐2, ๐)
=
โซ โ
๐
โซ โ
๐
๐๐0๐๐0 ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 โโซ โ
๐
โซ โ
๐
๐๐๐๐๐๐ ๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2 . (3.78)
We further define ๐ด๐(๐1, ๐2, ๐) = lim๐โโโ โฃ๐ผ ๐(๐1, ๐)โฃ (๐ = 1, 2, 3, 4, 5, 6, 7, 8).
100 S. Georgiev and J. Morais
For ๐ผ1(๐1, ๐) a straightforward computation shows that
โฃ๐ผ1(๐1, ๐2, ๐)โฃ
=
โฃโฃโฃโฃโฃ(๐๐(๐)๐๐๐1๐ โ ๐๐(โ๐)๐โ๐๐1๐ โ
โซ ๐
โ๐๐๐(๐ฅ1)๐๐1๐
๐๐1๐ฅ1๐๐ฅ1
)(๐๐(๐)๐๐๐2๐ โ ๐๐(โ๐)๐โ๐๐2๐ โ
โซ ๐
โ๐๐๐(๐ฅ2)๐๐2๐
๐๐2๐ฅ2๐๐ฅ2
)โ(๐0(๐)๐๐๐1๐ โ ๐0(โ๐)๐โ๐๐1๐ โ
โซ ๐
โ๐๐0(๐ฅ1)๐๐1๐
๐๐1๐ฅ1๐๐ฅ1
)(๐0(๐)๐๐๐2๐ โ ๐๐(โ๐)๐โ๐๐2๐ โ
โซ ๐
โ๐๐0(๐ฅ2)๐๐2๐
๐๐2๐ฅ2๐๐ฅ2
)โฃโฃโฃโฃโฃ. (3.79)
Based on these results and on (3.67), and the convergence of the sequences ๐๐(๐ฅ1)and ๐๐(๐ฅ2), it follows that ๐ด1(๐1, ๐2, ๐) = 0. Hence
โฃ๐ผ2(๐1, ๐)โฃ โคโซ โ๐
โโ
โซ โ๐
โโ๐๐0(๐ฅ1)๐๐
0(๐ฅ2) โฃ๐๐๐1๐ฅ1๐๐๐2๐ฅ2 โฃ
+
โซ โ๐
โโ
โซ โ๐
โโ๐๐๐(๐ฅ1)๐๐
๐(๐ฅ2) โฃ๐๐๐1๐ฅ1๐๐๐2๐ฅ2 โฃ
โค โฃ๐0(โ๐)โ ๐0(โโ)โฃ โฃ๐0(โ๐)โ ๐0(โโ)โฃ+ โฃ๐๐(โ๐)โ ๐๐(โโ)โฃ โฃ๐๐(โ๐)โ ๐๐(โโ)โฃ . (3.80)
From the last inequality and from (3.66) it follows that for any ๐ > 0 we can choose๐ > 0 large enough so that ๐ด2(๐1, ๐2, ๐) < ๐. In a similar way, we have
๐ด๐(๐1, ๐2, ๐) < ๐, ๐ = 3, 4, 5, 6, 7, 8, 9. (3.81)
By combining these arguments, we finally obtain
lim๐โโโ ๐๐(๐1, ๐2) = ๐0(๐1, ๐2), (3.82)
which completes the proof. โก
4. The Main Theorem
In this section we shall extend Bochnerโs result to the noncommutative structureof quaternion functions. We first define a few general properties of quaternionpositive-type functions.
4.1. Positive Functions
Let us define the notion of positive definite measure in the context of quaternionanalysis.
Definition 4.1. A function ๐ : โ2 โโ โ is called positive definite, if it satisfies thefollowing conditions:
5. Bochnerโs Theorems 101
1. ๐ is bounded and continuous on โ2.
2. ๐(โ๐) = ๐(๐) for all ๐ โ โ2, where ๐ is any function which can be repre-
sented as โฑ๐ฎ ๐(๐2, ๐1)(๐1, ๐2).3. For any ๐๐ = (๐1๐, ๐2๐) โ โ2, quaternion numbers ๐ง๐ (๐ = 1, 2, . . . , ๐),
and any positive integer ๐ the following inequality is satisfied:๐โ
๐,๐=1๐โฅ๐
๐(๐๐ โ ๐๐) ๐ง๐๐ง๐ +๐โ
๐,๐=1๐>๐
๐ง๐๐ง๐๐(๐๐ โ ๐๐) โฅ 0. (4.1)
These parameters are measured such that:
(i) When ๐๐ = ๐๐ for ๐ โ= ๐ (๐, ๐ = 1, 2, . . . , ๐) it follows๐โ
๐,๐=1๐โค๐
๐(0, 0)๐ง๐๐ง๐ +
๐โ๐,๐=1๐>๐
๐ง๐๐ง๐ ๐(0, 0) =
๐โ๐,๐=1๐โค๐
(๐(0, 0)๐ง๐๐ง๐ + ๐ง๐๐ง๐๐(0, 0)
)
=โ๐,๐=1๐โค๐
(๐(0, 0)๐ง๐๐ง๐ + ๐(0, 0)๐ง๐๐ง๐
)
= 2
๐โ๐,๐=1๐โค๐
S(๐(0, 0)๐ง๐๐ง๐) โฅ 0. (4.2)
(ii) When ๐๐ = ๐๐, ๐ง๐ = ๐ง๐ for ๐ โ= ๐ (๐, ๐ = 1, 2, . . . , ๐), and by using theprevious observation we get
2 S(๐(0, 0))
๐โ๐=1
โฃ๐ง๐โฃ2 โฅ 0 โ S(๐(0, 0)) โฅ 0. (4.3)
(iii) When ๐ = 2, ๐1 = ๐, and ๐2 = 0, it follows that
๐(๐)๐ง1๐ง2 + ๐(0, 0)(โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2
)+ ๐ง2๐ง1๐(๐)
= 2 S(๐(๐)๐ง1๐ง2) +(โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2
)๐(0, 0). (4.4)
In particular, when ๐ง1 = ๐ง2 we obtain
2 โฃ๐ง1โฃ2 (S(๐(๐)) + ๐(0, 0)) โฅ 0 โ S(๐(๐)) + ๐(0, 0) โฅ 0. (4.5)
Remark 4.2. It should be noted that not all functions in โฌ are positive definite.To verify this claim take, for example, the function ๐(๐1, ๐2) = โ๐โ๐๐1 . It is easyto see that expression (4.1) is not satisfied.
4.2. Bochner Theorem
Before proving the generalization of Bochnerโs theorem to quaternion functions,we consider the following set
๐ :=
{๐ : โ2 โโ โ : ๐ =
โซโ2
๐โฃโฃ๐1(๐ฅ1)
โฃโฃ ๐ โฃโฃ๐2(๐ฅ2)โฃโฃ ๐๐๐1๐ฅ1๐๐๐2๐ฅ2
}โ โฌ. (4.6)
102 S. Georgiev and J. Morais
We are now ready for the classification theorem.
Theorem 4.3. If ๐ โ ๐ then ๐ is positive definite.
Proof. Statement 1 of Definition 4.1 is proved in Proposition 3.10, and Statement2 follows from (3.29). Let ๐๐ = (๐1๐, ๐2๐) โ โ2, ๐ง๐ โ โ (๐ = 1, 2, . . . , ๐). Forany ๐1, ๐2 : โโ โ, straightforward computations show that
๐โ๐,๐=1๐โค๐
๐(๐๐ โ ๐๐)๐ง๐๐ง๐ +
๐โ๐,๐=1๐>๐
๐ง๐๐ง๐๐(๐๐ โ ๐๐)
=๐โ
๐,๐=1๐โค๐
(โซโ2
๐โฃโฃ๐1(๐ฅ1)
โฃโฃ ๐ โฃโฃ๐2(๐ฅ2)โฃโฃ ๐๐(๐1๐โ๐1๐)๐ฅ1๐๐(๐2๐โ๐2๐)๐ฅ2
)๐ง๐๐ง๐
+
๐โ๐,๐=1๐>๐
๐ง๐๐ง๐
โซโ2
๐ โฃ๐1(๐ฅ1)โฃ ๐ โฃ๐2(๐ฅ2)โฃ ๐๐(๐1๐โ๐1๐)๐ฅ1๐๐(๐2๐โ๐2๐)๐ฅ2
=
๐โ๐,๐=1๐โค๐
(โซโ2
๐โฃโฃ๐1(๐ฅ1)
โฃโฃ ๐ โฃโฃ๐2(๐ฅ2)โฃโฃ ๐๐(๐1๐โ๐1๐)๐ฅ1๐๐(๐2๐โ๐2๐)๐ฅ2
)๐ง๐๐ง๐
+
๐โ๐,๐=1๐>๐
๐ง๐๐ง๐
โซโ2
๐โ๐(๐2๐โ๐2๐)๐ฅ2๐โ๐(๐1๐โ๐1๐)๐ฅ1๐โฃโฃ๐1(๐ฅ1)
โฃโฃ ๐ โฃโฃ๐2(๐ฅ2)โฃโฃ
=
๐โ๐=1
โฃ๐ง๐โฃ2(โฃโฃ๐1(โ)
โฃโฃโ โฃโฃ๐1(โโ)โฃโฃ) (โฃโฃ๐2(โ)
โฃโฃโ โฃโฃ๐2(โโ)โฃโฃ)
+
๐โ๐,๐=1๐<๐
โซโ2
๐โฃโฃ๐1(๐ฅ1)
โฃโฃ ๐ โฃโฃ๐2(๐ฅ2)โฃโฃโโ ๐๐(๐1๐โ๐1๐)๐ฅ1๐๐(๐2๐โ๐2๐)๐ฝ๐ง๐๐ง๐
+
๐ง๐๐ง๐๐โ๐(๐2๐โ๐2๐)๐ฅ2๐โ๐(๐1๐โ๐1๐)๐ฅ1
โโ =
๐โ๐=1
โฃ๐ง๐โฃ2(โฃโฃ๐1(โ)
โฃโฃโ โฃโฃ๐1(โโ)โฃโฃ) (โฃโฃ๐2(โ)
โฃโฃโ โฃโฃ๐2(โโ)โฃโฃ)
+ 2
๐โ๐,๐=1๐<๐
โซโ2
๐โฃโฃ๐1(๐ฅ1)
โฃโฃ ๐ โฃโฃ๐2(๐ฅ2)โฃโฃS(๐๐(๐1๐โ๐1๐)๐ฅ1๐๐(๐2๐โ๐2๐)๐ฅ2๐ง๐๐ง๐
)
โฅ๐โ
๐=1
โฃ๐ง๐โฃ2(โฃโฃ๐1(โ)
โฃโฃโ โฃโฃ๐1(โโ)โฃโฃ) (โฃโฃ๐2(โ)
โฃโฃ โ โฃโฃ๐2(โโ)โฃโฃ)
โ 2
๐โ๐,๐=1๐<๐
โซโ2
โฃ๐ง๐โฃโฃ๐ง๐โฃ ๐โฃโฃ๐1(๐ฅ1)
โฃโฃ ๐ โฃโฃ๐2(๐ฅ2)โฃโฃ
5. Bochnerโs Theorems 103
=
โโโ ๐โ๐=1
โฃ๐ง๐โฃ2 โ 2
๐โ๐,๐=1๐<๐
โฃ๐ง๐โฃโฃ๐ง๐โฃ
โโโ (โฃโฃ๐1(โ)โฃโฃโ โฃโฃ๐1(โโ)
โฃโฃ) (โฃโฃ๐2(โ)โฃโฃโ โฃโฃ๐2(โโ)
โฃโฃ)
=
(๐โ
๐=1
โฃ๐ง๐โฃ)2 (โฃโฃ๐1(โ)
โฃโฃโ โฃโฃ๐1(โโ)โฃโฃ) (โฃโฃ๐2(โ)
โฃโฃโ โฃโฃ๐2(โโ)โฃโฃ) โฅ 0. (4.7)
This proves that ๐ is positive definite. โก
The extension of Bochnerโs result to a much larger class of quaternion func-tions remains a challenge to future research.
Acknowledgement
Partial support from the Foundation for Science and Technology (FCT) via thegrant DD-VU-02/90, Bulgaria, is acknowledged by the first named author. Thesecond named author acknowledges financial support from the Foundation for Sci-ence and Technology (FCT) via the post-doctoral grant SFRH/ BPD/66342/2009.This work was supported by FEDER funds through the COMPETE โ Opera-tional Programme Factors of Competitiveness (โPrograma Operacional Factoresde Competitividadeโ) and by Portuguese funds through the Center for Researchand Development in Mathematics and Applications (University of Aveiro) andthe Portuguese Foundation for Science and Technology (โFCT โ Fundacao paraa Ciencia e a Tecnologiaโ), within project PEst-C/MAT/UI4106/2011 with theCOMPETE number FCOMP-01-0124-FEDER-022690.
References
[1] S. Bochner. Monotone funktionen, Stieltjessche integrate, und harmonische analyse.Mathematische Annalen, 108:378โ410, 1933.
[2] S. Bochner. Lectures on Fourier Integrals. Princeton University Press, Princeton, NewJersey, 1959.
[3] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[4] T. Bulow, M. Felsberg, and G. Sommer. Non-commutative hypercomplex Fouriertransforms of multidimensional signals. In G. Sommer, editor, Geometric computingwith Clifford Algebras: Theoretical Foundations and Applications in Computer Visionand Robotics, pages 187โ207, Berlin, 2001. Springer.
[5] S. Georgiev, J. Morais, K.I. Kou, and W. Sproรig. BochnerโMinlos theorem andquaternion Fourier transform. To appear.
[6] B. Mawardi, E. Hitzer, A. Hayashi, and R. Ashino. An uncertainty principle for quater-nion Fourier transform. Computers and Mathematics with Applications, 56(9):2411โ2417, 2008.
104 S. Georgiev and J. Morais
[7] B. Mawardi and E.M.S. Hitzer. Clifford Fourier transformation and uncertainty prin-ciple for the Clifford algebra ๐ถโ3,0. Advances in Applied Clifford Algebras, 16(1):41โ61,2006.
S. GeorgievDepartment of Differential EquationsUniversity of SofiaSofia, Bulgariae-mail: [email protected]
J. MoraisCentro de Investigacao e Desenvolvimento
em Matematica e Aplicacoes (CIDMA)Universidade de AveiroP3810-193 Aveiro, Portugale-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 105โ120cโ 2013 Springer Basel
6 BochnerโMinlos Theorem andQuaternion Fourier Transform
S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
Abstract. There have been several attempts in the literature to generalizethe classical Fourier transform by making use of the Hamiltonian quaternionalgebra. The first part of this chapter features certain properties of the as-ymptotic behaviour of the quaternion Fourier transform. In the second partwe introduce the quaternion Fourier transform of a probability measure, andwe establish some of its basic properties. In the final analysis, we introducethe notion of positive definite measure, and we set out to extend the classicalBochnerโMinlos theorem to the framework of quaternion analysis.
Mathematics Subject Classification (2010). Primary 30G35; secondary 42A38;tertiary 42A82.
Keywords. Quaternion analysis, quaternion Fourier transform, asymptotic be-haviour, positive definitely measure, BochnerโMinlos theorem.
1. Introduction and Statement of Results
As is well known, the classical Fourier transform (FT) has wide applications inengineering, computer sciences, physics and applied mathematics. For instance,the FT can be used to provide signal analysis techniques where the signal fromthe original time domain is transformed to the frequency domain. Therefore thereexists great interest and considerable effort to extend the FT to higher dimensions,and study its properties and interdependencies (see, e.g., [1โ7, 9โ11, 17, 18, 20โ23] and elsewhere). In view of numerous applications in physics and engineeringproblems, one is particularly interested in higher-dimensional analogues to โ๐, inparticular, for ๐ = 4. To this end, so far quaternion analysis offers the possibilityof generalizing the underlying function theory in 2D to 4D, with the advantageof meeting exactly these goals. To aid the reader, see [15, 16, 19, 24, 25] for morecomplete accounts of this subject and related topics.
The first part of the present work is devoted to the study of the asymp-totic behaviour of the quaternion Fourier transform (QFT). The QFT was first
106 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
introduced by Ell in [10]. He proposed a two-sided QFT and studied some of itsapplications and properties. Later, Bulow [7] (cf. [8] and [1,20]) has also conducteda generalization of the real and complex FT using the quaternion algebra basedon two complex variables, but, to the best of our knowledge, a detailed study ofits asymptotic behaviour has not been carried out yet. The main motivation of thepresent study is to develop further general numerical methods for partial differen-tial equations and to extend localization theorems for summation of Fourier seriesin the quaternion analysis setting. In a forthcoming article we shall describe theseconnections in more detail and illustrate them by some typical examples.
Due to the noncommutativity of the quaternions, there are three differenttypes of QFT: a right-sided QFT, a left-sided QFT, and a two-sided QFT [23].We will carry out the investigation of the following finite integral (defined fromthe time domain to the frequency domain)
โฑ๐(๐)(๐1, ๐2) :=
โซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2) ๐โ๐๐1๐ฅ1 ๐โ๐๐2๐ฅ2 ๐๐ฅ1๐๐ฅ2, (1.1)
where the signal ๐ : [๐, ๐]ร [๐, ๐] โ โ2 โโ โ will be taken to be
๐(๐ฅ1, ๐ฅ2) := [๐(๐ฅ1, ๐ฅ2)]0 + [๐(๐ฅ1, ๐ฅ2)]1๐+ [๐(๐ฅ1, ๐ฅ2)]2๐ + [๐(๐ฅ1, ๐ฅ2)]3๐,
[๐ ]๐ : [๐, ๐]ร [๐, ๐] โโ โ (๐ = 0, 1, 2, 3)
satisfying certain conditions, guaranteeing the convergence of the above integral.โฑ๐(๐)(๐1, ๐2) is the (finite) right-sided Fourier transform [21] of the quaternionfunction ๐(๐ฅ1, ๐ฅ2), and it may be interpreted as a quaternionic extension of theclassical FT; the exponential product ๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2 is called the (right-sided)quaternion Fourier kernel, and for ๐ = 1, 2; ๐ฅ๐ will denote the space and ๐๐ theangular frequency variables. The previous definition of the QFT varies from theoriginal one only in the fact that we use 2D vectors instead of scalars and that itis defined to be two-dimensional. Here ๐, ๐ and ๐ are unit pure quaternions (i.e.,the quaternions with unit magnitude having no scalar part) that are orthogonalto each other. We point out that the product in (1.1) has to be written in a fixedorder since, in general, ๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2 does not commute with every element ofthe algebra.
Remark 1.1. Throughout this text we investigate the integral (1.1) only that, forsimplicity, we denote by โฑ(๐). Nevertheless, all results can be easily performedfrom the left-hand side:
โฑ๐(๐)(๐1, ๐2) :=
โซ ๐
๐
โซ ๐
๐
๐โ๐๐2๐ฅ2 ๐โ๐๐1๐ฅ1๐(๐ฅ1, ๐ฅ2) ๐๐ฅ1๐๐ฅ2,
since
โฑ๐(๐)(โ๐1,โ๐2) =
โซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2)๐๐๐1๐ฅ1๐๐๐2๐ฅ2๐๐ฅ1๐๐ฅ2
6. BochnerโMinlos Theorem and Quaternion Fourier Transform 107
=
โซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2)๐โ๐๐2๐ฅ2๐โ๐๐1๐ฅ1๐๐ฅ1๐๐ฅ2
=
โซ ๐
๐
โซ ๐
๐
๐โ๐๐2๐ฅ2๐โ๐๐1๐ฅ1๐(๐ฅ1, ๐ฅ2)๐๐ฅ1๐๐ฅ2 = โฑ๐(๐)(๐1, ๐2).
Lemma 1.2. The QFT of a 2D signal ๐ โ ๐ฟ1([๐, ๐]ร [๐, ๐];โ) has the closed-formrepresentation:
โฑ(๐)(๐1, ๐2) := ฮฆ0(๐1, ๐2) + ฮฆ1(๐1, ๐2) + ฮฆ2(๐1, ๐2) + ฮฆ3(๐1, ๐2),
where the integrals are
ฮฆ0(๐1, ๐2) =
โซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2) cos(๐1๐ฅ1) cos(๐2๐ฅ2)๐๐ฅ1๐๐ฅ2,
ฮฆ1(๐1, ๐2) = โโซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2)๐ sin(๐1๐ฅ1) cos(๐2๐ฅ2)๐๐ฅ1๐๐ฅ2,
ฮฆ2(๐1, ๐2) = โโซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2)๐ cos(๐1๐ฅ1) sin(๐2๐ฅ2)๐๐ฅ1๐๐ฅ2,
ฮฆ3(๐1, ๐2) =
โซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2)๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2)๐๐ฅ1๐๐ฅ2.
For illustrative purposes, we have the following identities:
Corollary 1.3. The QFT of a 2D signal ๐ โ ๐ฟ1([๐, ๐] ร [๐, ๐];โ) satisfies the fol-lowing relations:
โฑ(๐)(๐1, ๐2) + โฑ(๐)(๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2) + ฮฆ1(๐1, ๐2)) ,
โฑ(๐)(๐1, ๐2)โโฑ(๐)(๐1,โ๐2) = 2 (ฮฆ2(๐1, ๐2) + ฮฆ3(๐1, ๐2)) ,
โฑ(๐)(๐1, ๐2) + โฑ(๐)(โ๐1, ๐2) = 2 (ฮฆ0(๐1, ๐2) + ฮฆ2(๐1, ๐2)) ,
โฑ(๐)(๐1, ๐2)โโฑ(๐)(โ๐1, ๐2) = 2 (ฮฆ1(๐1, ๐2) + ฮฆ3(๐1, ๐2)) ,
โฑ(๐)(๐1, ๐2) + โฑ(๐)(โ๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2) + ฮฆ3(๐1, ๐2)) ,
โฑ(๐)(๐1, ๐2)โโฑ(๐)(โ๐1,โ๐2) = 2 (ฮฆ1(๐1, ๐2) + ฮฆ2(๐1, ๐2)) ,
โฑ(๐)(๐1,โ๐2) + โฑ(๐)(โ๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2)โ ฮฆ2(๐1, ๐2)) ,
โฑ(๐)(๐1,โ๐2)โโฑ(๐)(โ๐1,โ๐2) = 2 (ฮฆ1(๐1, ๐2)โ ฮฆ3(๐1, ๐2)) ,
โฑ(๐)(โ๐1, ๐2) + โฑ(๐)(โ๐1,โ๐2) = 2 (ฮฆ0(๐1, ๐2)โ ฮฆ1(๐1, ๐2)) ,
โฑ(๐)(โ๐1, ๐2)โโฑ(๐)(โ๐1,โ๐2) = 2 (ฮฆ2(๐1, ๐2)โ ฮฆ3(๐1, ๐2)) .
Under suitable conditions, the original signal ๐ can be reconstructed fromโฑ(๐) by the inverse transform (frequency to time domains).
Definition 1.4. The (right-sided) inverse QFT of ๐ โ ๐ฟ1(โ2;โ) is given by
โฑโ1๐ (๐) : โ2 โโ โ,
โฑโ1๐ (๐)(๐ฅ1, ๐ฅ2) :=
1
(2๐)2
โซโ2
๐(๐1, ๐2) ๐๐๐2๐ฅ2 ๐๐๐1๐ฅ1 ๐๐1๐๐2.
108 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
It has the closed-form representation:
โฑโ1๐ (๐)(๐ฅ1, ๐ฅ2) := ฮฆ0(๐ฅ1, ๐ฅ2) + ฮฆ1(๐ฅ1, ๐ฅ2) + ฮฆ2(๐ฅ1, ๐ฅ2) + ฮฆ3(๐ฅ1, ๐ฅ2),
where the integrals are
ฮฆ0(๐ฅ1, ๐ฅ2) =1
(2๐)2
โซ ๐
๐
โซ ๐
๐
๐(๐1, ๐2) cos(๐1๐ฅ1) cos(๐2๐ฅ2)๐๐1๐๐2,
ฮฆ1(๐ฅ1, ๐ฅ2) =1
(2๐)2
โซ ๐
๐
โซ ๐
๐
๐(๐1, ๐2)๐ sin(๐1๐ฅ1) cos(๐2๐ฅ2)๐๐1๐๐2,
ฮฆ2(๐ฅ1, ๐ฅ2) =1
(2๐)2
โซ ๐
๐
โซ ๐
๐
๐(๐1, ๐2)๐ cos(๐1๐ฅ1) sin(๐2๐ฅ2)๐๐1๐๐2,
ฮฆ3(๐ฅ1, ๐ฅ2) = โ 1
(2๐)2
โซ ๐
๐
โซ ๐
๐
๐(๐1, ๐2)๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2)๐๐1๐๐2.
The quaternion exponential product ๐๐๐ฅ2๐2 ๐๐๐1๐ฅ1 is called the inverse (right-sided)quaternion Fourier kernel.
Remark 1.5. Again, all computations can easily be converted to other conventions,since
โฑโ1๐ (๐)(โ๐ฅ1,โ๐ฅ2) =
1
(2๐)2
โซ ๐
๐
โซ ๐
๐
๐(๐1, ๐2)๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2๐๐1๐๐2
=1
(2๐)2
โซ ๐
๐
โซ ๐
๐
๐๐๐1๐ฅ1 ๐๐๐2๐ฅ2๐(๐1, ๐2)๐๐1๐๐2
:= โฑโ1๐ (๐)(๐ฅ1, ๐ฅ2).
For convenience, below we will denote โฑโ1๐ as โฑโ1.
The present chapter has two main goals. The first consists in studying theasymptotic behaviour of the integral (1.1) under the assumption that ๐ belongsto ๐ฟ1((๐, ๐) ร (๐, ๐);โ) where ๐, ๐, ๐, ๐ can be both finite and infinite points. Weshall be interested in the connection between the function ๐(๐ฅ1, ๐ฅ2), and the be-haviour of its Fourier transform โฑ(๐)(๐1, ๐2) at infinity. These properties have aninterest on their own for further applications to number theory, combinatorics, sig-nal processing, imaging, computer vision and numerical analysis. The complexityof the underlying computations will need some attention. Central to this view-point are certain Fourier transform techniques, which as in the complex case,would be familiar to the reader. Our second goal consists in extending the clas-sical BochnerโMinlos theorem to a noncommutative structure as in the case ofquaternion functions. The resulting theorem guarantees the existence and unique-ness of the corresponding probability measure defined on a dual space. This willbe done using the concept of quaternion Fourier transform of a probability mea-sure. For the readerโs convenience and for the sake of easy reference, the chapteris motivated by the results presented in [12, 13].
6. BochnerโMinlos Theorem and Quaternion Fourier Transform 109
Although the presented results can be extended to generalized Clifford alge-bras as well, we will focus the discussion on the QFT for conciseness here. Theproof of its generalization to higher dimensions is possible but needs more com-plicated calculations, which exceed the scope of this manuscript. These results arestill under further investigation and will be reported in a forthcoming paper.
2. Preliminaries
At this stage we briefly recall basic algebraic facts about quaternions necessary forthe sequel. Let โ := {๐ = ๐+ ๐๐+ ๐๐ + ๐๐ : ๐, ๐, ๐, ๐ โ โ} be a four-dimensionalassociative and noncommutative algebra, where the imaginary units ๐, ๐, and ๐are subject to the Hamiltonian multiplication rules
๐2 = ๐2 = ๐2 = โ1;๐๐ = ๐ = โ๐๐, ๐๐ = ๐ = โ๐๐, ๐๐ = ๐ = โ๐๐.
The scalar and vector parts of ๐, S(๐) andV(๐), are defined as the ๐ and ๐๐+๐๐+๐๐terms, respectively. For the scalar part the cyclic product rule S(๐๐๐) = S(๐๐๐) isvalid. Further step is quaternion conjugation introduced similarly to that of thecomplex numbers ๐ = ๐โ๐๐โ๐๐โ๐๐. The quaternion conjugation is an anti-linearinvolution
๐ = ๐, ๐+ ๐ = ๐+ ๐, ๐๐ = ๐๐, ๐๐ = ๐๐ (โ๐ โ โ).
The norm of ๐ is defined by โฃ๐โฃ = โ๐๐ =
โ๐๐ =
โ๐2 + ๐2 + ๐2 + ๐2, and it
coincides with its corresponding Euclidean norm as a vector in โ4. For everytwo quaternions ๐ and ๐ the triangle inequalities hold โฃ๐+ ๐โฃ โค โฃ๐โฃ + โฃ๐โฃ, andโฃ โฃ๐โฃ โ โฃ๐โฃ โฃ โค โฃ๐ยฑ ๐โฃ. Also, we have โฃS(๐)โฃ โค โฃ๐โฃ and โฃV(๐)โฃ โค โฃ๐โฃ.
In the sequel, a quaternion sequence is a collection of real quaternions ๐0, ๐1,๐2, . . . โlabelledโ by nonnegative integers. We shall denote such a sequence by {๐๐}where ๐ = 0, 1, 2, . . . and ๐๐ = ๐0,๐ + ๐1,๐๐+ ๐2,๐๐ + ๐3,๐๐ are the elements of thesequence, ๐๐,๐ โ โ (๐ = 0, 1, 2, 3). To supplement our investigations, we recall thekey notion of convergence of a quaternion sequence.
Definition 2.1. The quaternion sequence {๐๐} is called convergent to the quater-nion ๐ = ๐+๐๐+๐๐+๐๐ if lim
๐โโโ โฃ๐๐โ๐โฃ = 0. We will use the traditional notation:
lim๐โโโ ๐๐ = ๐.
Lemma 2.2. Let {๐๐} and {๐๐} be two quaternion sequences for which lim๐โโโ ๐๐ = ๐
and lim๐โโโ ๐๐ = ๐, for ๐, ๐ โ โ. Then
1. lim๐โโโ(๐๐ ยฑ ๐๐) = ๐ยฑ ๐;
2. lim๐โโโ(๐๐๐๐) = ๐๐;
3. lim๐โโโ(๐ผ๐๐) = ๐ผ๐, ๐ผ โ โ.
110 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
Now, let ๐ be the space of sequences
๐ := {{๐๐} : lim๐โโโ๐๐ก๐๐ = 0, โ ๐ก โ โ0},
where ๐๐ โ โ, and let
๐ ๐ := {{๐๐} : โฅ๐โฅ2๐ :=โโ๐=0
(1 + ๐2)๐ โฃ๐๐โฃ2 < +โ}, ๐ โ โค.
Proposition 2.3. We have ๐ = โฉ๐โโค๐ ๐.
Proof. Let ๐ โ ๐ be arbitrarily chosen and fixed, and ๐ก โ โ0. From the definitionof the space ๐ we have that lim
๐โโโ๐๐ก๐๐ = 0. Then for ๐ถ > 0 a natural number
๐ = ๐(๐ถ) can be found so that for every ๐ > ๐ the following holds
๐๐กโฃ๐๐โฃ โค ๐ถ.
Hence, it followsโโ๐=0
(1 + ๐2
)[ ๐ก2 ] โฃ๐๐โฃ2 โค โฃ๐0โฃ2 + ๐ถ2โโ๐=1
(1 + ๐2
)[ ๐ก4 ] 1
๐2๐ก< +โ,
and therefore ๐ โ ๐ [ ๐ก4 ]. Since ๐ก โ โ0 is arbitrary we conclude that ๐ โ ๐ ๐ for every
๐ โ โ0, from where ๐ โ โฉ๐โโ๐ ๐. Since ๐ โ ๐ was arbitrarily chosen it followsthat ๐ โ โฉ๐โโ๐ ๐. Now, let ๐ โ โฉ๐โโค๐ ๐. Then for every ๐ โ โ0 we have that
โโ๐=0
(1 + ๐2
)๐ โฃ๐๐โฃ2 < +โ,
from where lim๐โโโ
(1 + ๐2
)๐ โฃ๐๐โฃ2 = 0. Therefore lim๐โโโ๐๐ก๐๐ = 0 for every ๐ก โค 2๐.
Since ๐ was arbitrarily chosen then lim๐โโโ๐๐ก๐๐ = 0 for every ๐ก โ โ0. Conse-
quently, ๐ โ ๐ and since ๐ โ โฉ๐โโ๐ ๐ was arbitrarily chosen we conclude thatโฉ๐โโ๐ ๐ โ ๐ . โก
In the sequel, consider the countable family of semi-norms on ๐
โฅ๐โฅ2๐ =
โโ๐=0
(1 + ๐2
)๐ โฃ๐๐โฃ2 .Lemma 2.4. ๐ is completely a Hausdorff space.
Proof. If ๐ โ ๐ and โฅ๐โฅ๐ = 0 we getโโ๐=0
(1 + ๐2)๐ โฃ๐๐โฃ2 = 0.
Whence, ๐๐ = 0 for every natural ๐ (including zero), and consequently ๐ = 0. It
follows that ๐ is a Hausdorff space. Now, let {๐๐๐} โโ๐โโโ {๐๐}, i.e., lim๐โโโ
๐๐๐ =
๐๐ for every ๐ โ โ0. Then โฅโฅ๐๐ โ ๐โฅโฅ๐โโ๐โโโ 0.
It follows that ๐ is completely a space. โก
6. BochnerโMinlos Theorem and Quaternion Fourier Transform 111
Let us define the metric
๐(๐, ๐) =
โโ๐=0
โฅ๐โ ๐โฅ๐1 + โฅ๐โ ๐โฅ๐
2โ๐
for ๐, ๐ โ ๐ . Evidently, the above metric has the translation property. In otherwords, from the countable family of seminorms we can define a metric with thetranslation property. From this follows the result that
Lemma 2.5. ๐ is a Frechet space.
Now, let ๐ โฒ denote the topological dual space to ๐ given by ๐ โฒ = โช๐โโค๐ ๐.We will denote the set of all sequences by โโ0 , and we equip the space ๐ โฒ withcylindrical topology. Every element ๐โฒ โ ๐ โฒ acts on each element ๐ โ ๐ as follows
โจ๐โฒ, ๐โฉ =โโ๐=0
๐โฒ๐๐๐; lim๐โโโ ๐โฒ๐ = ๐โฒ, lim
๐โโโ ๐๐ = ๐.
3. Asymptotic Behaviour of the QFT
There are numerous questions that remain untouched about the behaviour of theQFT. We will fill in some of these gaps and discuss some rudiments of the asymp-totic behaviour of (1.1).
We begin by proving the following result, which provides an interesting andefficient convergence characterization of the QFT.
Theorem 3.1. Let ๐, ๐, ๐, ๐ โ โ and ๐ โ ๐ฟ1((๐, ๐)ร(๐, ๐);โ), then โฑ(๐) is uniformlycontinuous and bounded. Moreover, it satisfies
lim๐1โโยฑโ
โฑ(๐)(๐1, ๐2) = 0
uniformly in ๐2, and
lim๐2โโยฑโ
โฑ(๐)(๐1, ๐2) = 0
uniformly in ๐1. The results hold with the same proof when the region (๐, ๐)ร(๐, ๐)
is replaced by its topological closure (๐, ๐)ร (๐, ๐) (๐, ๐, ๐, ๐ can be ยฑโ) or anyregion ๐บ (or ๐บ) in the space โ2.
Proof. It is easy to show that
โฃโฑ(๐)(๐1, ๐2)โฃ โคโซ ๐
๐
โซ ๐
๐
โฃ๐(๐ฅ1, ๐ฅ2)โฃ ๐๐ฅ1๐๐ฅ2
= โฅ๐โฅ๐ฟ1((๐,๐)ร(๐,๐);โ) . (3.1)
So, the transform โฑ(๐) is bounded.To prove the limit assertions, we use a density argument as in the classical
case (RiemannโLebesgue Lemma). We first assume that both ๐ and โ๐ฅ1๐ arecontinuous with compact support. Obviously, such functions form a dense subspace
112 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
in ๐ฟ1((๐, ๐)ร (๐, ๐);โ). By using integration by parts to the ๐ฅ1-variable, with theiterated integration, we have
โฑ(๐)(๐1, ๐2) =
โซ ๐
๐
(๐(๐, ๐ฅ2)
1
i๐1๐โi๐1๐ ๐โj๐2๐ฅ2 โ ๐(๐, ๐ฅ2)
1
i๐1๐โi๐1๐ ๐โj๐2๐ฅ2
)๐๐ฅ2
โโซ ๐
๐
โซ ๐
๐
โ๐ฅ1๐(๐ฅ1, ๐ฅ2)1
i๐1๐โi๐1๐ฅ1 ๐โj๐2๐ฅ2 ๐๐ฅ1๐๐ฅ2
= โโซ ๐
๐
โซ ๐
๐
โ๐ฅ1๐(๐ฅ1, ๐ฅ2)1
i๐1๐โi๐1๐ฅ1 ๐โj๐2๐ฅ2 ๐๐ฅ1๐๐ฅ2.
Therefore,
โฃโฑ(๐)(๐1, ๐2)โฃ โค 1
โฃ๐1โฃ โฅโ๐ฅ1๐โฅ๐ฟ1((๐,๐)ร(๐,๐);โ) . (3.2)
So,
lim๐1โโยฑโ
โฑ(๐)(๐1, ๐2) = 0
uniformly in ๐2. Similarly we can prove
lim๐2โโยฑโ
โฑ(๐)(๐1, ๐2) = 0
uniformly in ๐1.
Now assume ๐ โ ๐ฟ1((๐, ๐) ร (๐, ๐);โ). For any given ๐ > 0, there exists afunction ๐๐ in the above-mentioned dense class, such that
โฅ๐ โ ๐๐โฅ๐ฟ1((๐,๐)ร(๐,๐);โ) < ๐.
By (3.1), we have
โฃโฑ(๐)(๐1, ๐2)โฃ โค โฃโฑ(๐๐)(๐1, ๐2)โฃ+ โฅ๐ โ ๐๐โฅ๐ฟ1((๐,๐)ร(๐,๐);โ)
โค โฃโฑ(๐๐)(๐1, ๐2)โฃ+ ๐.
By using the result just proved for the density class, we have
lim๐1โยฑโ
โฃโฑ(๐)(๐1, ๐2)โฃ โค ๐,
uniformly in ๐2. Since ๐ is arbitrary, we have
lim๐1โยฑโ
โฑ(๐)(๐1, ๐2) = 0 (3.3)
uniformly in ๐2. Similarly,
lim๐2โยฑโ
โฑ(๐)(๐1, ๐2) = 0, (3.4)
uniformly in ๐1.
Now we show the uniform continuity of โฑ(๐)(๐1, ๐2). Given (3.3) and (3.4),since continuous functions are uniform continuous in compact sets, it suffices to
6. BochnerโMinlos Theorem and Quaternion Fourier Transform 113
show that โฑ(๐) is continuous at every point (๐1, ๐2). In fact,
โฑ(๐)(๐1 + ๐1, ๐2 + ๐2)โโฑ(๐)(๐1, ๐2)
=
โซ ๐
๐
โซ ๐
๐
๐(๐ฅ1, ๐ฅ2)[๐โ๐๐ฅ1(๐1+๐1)๐โ๐๐ฅ2(๐2+๐2) โ ๐โ๐๐ฅ2๐1๐โ๐๐ฅ2๐2
]๐๐ฅ1๐๐ฅ2.
For any ๐1, ๐2 > 0, the integrand is dominated by a constant multiple of โฃ๐(๐ฅ1, ๐ฅ2)โฃ.Since the factor in the straight brackets tends to zero, by the Lebesgue DominatedConvergence Theorem, we have
lim๐1,๐2โ0
โฑ(๐)(๐1 + ๐1, ๐2 + ๐2)โโฑ(๐)(๐1, ๐2) = 0,
proving the desired continuity. โก
4. BochnerโMinlos Theorem
In this section we extend the classical BochnerโMinlos theorem (named afterBochner and Adolโfovich Minlos) to the framework of quaternion analysis. Theresulting theorem guarantees the existence of the corresponding probability mea-sure defined on a dual space. In particular, some interesting properties of theunderlying measure are extended in this setting.
Definition 4.1. Let ๐ be a finite positive measure on โ2. The QFT of ๐ is thefunction โฑ๐๐(๐) : โ
2 โโ โ given by
โฑ๐๐(๐)(๐1, ๐2) :=
โซโ2
๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2๐๐(๐ฅ1, ๐ฅ2),
or the function โฑ๐๐(๐) : โ2 โโ โ given by
โฑ๐๐(๐)(๐1, ๐2) :=
โซโ2
๐โ๐๐2๐ฅ2๐โ๐๐1๐ฅ1๐๐(๐ฅ1, ๐ฅ2).
Proposition 4.2. Let ๐ be a finite positive measure on โ2. The functionals โฑ๐๐(๐)and โฑ๐๐(๐) satisfy the following basic properties:
1. โฑ๐๐(๐)(0, 0) = โฑ๐๐(๐)(0, 0) = 1;
2. โฑ๐๐(๐)(โ๐1,โ๐2) = โฑ๐๐(๐)(๐1, ๐2);
3. โฑ๐๐(๐)(โ๐1,โ๐2) = โฑ๐๐(๐)(๐1, ๐2);4. โฑ๐๐(๐)(โ๐1,โ๐2) + โฑ๐๐(๐)(๐1, ๐2)
= 2
โซโ2
(cos(๐1๐ฅ1) cos(๐2๐ฅ2) + ๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2)) ๐๐(๐ฅ1, ๐ฅ2);
5. โฑ๐๐(๐)(๐1, ๐2) + โฑ๐๐(๐)(โ๐1,โ๐2) = 2
โซโ2
cos(๐1๐ฅ1) cos(๐2๐ฅ2)๐๐(๐ฅ1, ๐ฅ2).
Proof. For the first statement, note that
โฑ๐๐(๐)(0, 0) =
โซโ2
๐๐(๐ฅ1, ๐ฅ2) = ๐(โ2).
114 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
For Statement 2 a straightforward computation shows that
โฑ๐๐(๐)(โ๐1,โ๐2) =
โซโ2
๐๐๐1๐ฅ1๐๐๐2๐ฅ2๐๐(๐ฅ1, ๐ฅ2) =
โซโ2
๐โ๐๐2๐ฅ2๐โ๐๐1๐ฅ1๐๐(๐ฅ1, ๐ฅ2)
=
โซโ2
๐โ๐๐2๐ฅ2๐โ๐๐1๐ฅ1๐๐(๐ฅ1, ๐ฅ2) = โฑ๐๐(๐)(๐1, ๐2).
The proof of Statement 3 will be omitted, being similar to the previous one. Now,taking into account that
๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2 = cos(๐1๐ฅ1) cos(๐2๐ฅ2)โ ๐ sin(๐1๐ฅ1) cos(๐2๐ฅ2)
โ ๐ cos(๐1๐ฅ1) sin(๐2๐ฅ2) + ๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2),
and
๐๐๐1๐ฅ1๐๐๐2๐ฅ2 = cos(๐1๐ฅ1) cos(๐2๐ฅ2) + ๐ sin(๐1๐ฅ1) cos(๐2๐ฅ2)
+ ๐ cos(๐1๐ฅ1) sin(๐2๐ฅ2) + ๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2),
we obtain
โฑ๐๐(๐)(๐1, ๐2) + โฑ๐๐(๐)(โ๐1,โ๐2)
=
โซโ2
(๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2 + ๐๐๐1๐ฅ1๐๐๐2๐ฅ2
)๐๐(๐ฅ1, ๐ฅ2)
= 2
โซโ2
(cos(๐1๐ฅ1) cos(๐2๐ฅ2) + ๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2)) ๐๐(๐ฅ1, ๐ฅ2).
For the last statement we use the relation
๐๐๐ค2๐ฅ2๐๐๐1๐ฅ1 = cos(๐1๐ฅ1) cos(๐2๐ฅ2) + ๐ sin(๐1๐ฅ1) cos(๐2๐ฅ2)
+ ๐ cos(๐1๐ฅ1) sin(๐2๐ฅ2)โ ๐ sin(๐1๐ฅ1) sin(๐2๐ฅ2).
Therefore, it follows
โฑ๐๐(๐)(๐1, ๐2) + โฑ๐๐(๐)(โ๐1,โ๐2)
=
โซโ2
(๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2 + ๐๐๐2๐ฅ2๐๐๐1๐ฅ1
)๐๐(๐ฅ1, ๐ฅ2)
= 2
โซโ2
cos(๐1๐ฅ1) cos(๐2๐ฅ2)๐๐(๐ฅ1, ๐ฅ2). โก
In the sequel, let us denote by ๐(โ2) the Schwartz space of smooth quaternionfunctions on โ2. We formulate a first result.
Proposition 4.3. Let ๐ โ ๐(โ2) and ๐ be a finite positive measure on โ2. Then
1.
โซโ2
โฑ๐๐(๐)(๐1, ๐2)๐(๐1, ๐2)๐๐1๐๐2 =
โซโ2
โฑ๐๐(๐)(๐ฅ1, ๐ฅ2)๐๐(๐ฅ1, ๐ฅ2);
2.
โซโ2
โฑ๐๐(๐)(๐1, ๐2)๐(๐1, ๐2)๐๐1๐๐2 =
โซโ2
โฑ๐๐(๐)(๐ฅ1, ๐ฅ2)๐๐(๐ฅ1, ๐ฅ2).
6. BochnerโMinlos Theorem and Quaternion Fourier Transform 115
Proof. For simplicity we just present the computations of the first equality. Theproof of the second one is similar. A direct computation shows thatโซ
โ2
โฑ๐๐(๐)(๐1, ๐2)๐(๐1, ๐2)๐๐1๐๐2
=
โซโ2
โซโ2
๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2๐๐(๐ฅ1, ๐ฅ2)๐(๐1, ๐2)๐๐1๐๐2
=
โซโ2
โซโ2
๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2๐(๐1, ๐2)๐๐1๐๐2๐๐(๐ฅ1, ๐ฅ2)
=
โซโ2
โฑ๐๐(๐)(๐ฅ1, ๐ฅ2)๐๐(๐ฅ1, ๐ฅ2). โก
We now analyze some key properties of the above-mentioned functionals.
Proposition 4.4. Let ๐ and ๐ be finite positive measures on โ2. The functionalsโฑ๐๐(๐) and โฑ๐๐(๐) are linear, i.e., for every ๐, ๐ โ โ it holds:
โฑ๐๐(๐๐+ ๐๐) = โฑ๐๐(๐)๐+ โฑ๐๐(๐)๐,
โฑ๐๐(๐๐+ ๐๐) = โฑ๐๐(๐)๐+ โฑ๐๐(๐)๐.
Proposition 4.5. Let ๐ be a finite positive measure on โ2. For any ๐๐ โ โ โ {0}(๐ = 1, 2) the following conditions hold:
1. โฑ๐๐(๐(๐1๐ฅ1, ๐2๐ฅ2)) = โฑ๐๐(๐(๐ฅ1, ๐ฅ2))(๐1๐1
, ๐2๐2
);
2. โฑ๐๐(๐(๐1๐ฅ1, ๐2๐ฅ2)) = โฑ๐๐(๐(๐ฅ1, ๐ฅ2))(๐1๐1
, ๐2๐2
).
Proof. For simplicity we just present the proof of the first condition. A straight-forward computation shows that
โฑ๐๐(๐(๐1๐ฅ1, ๐2๐ฅ2)) =
โซโ2
๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2๐๐(๐1๐ฅ1, ๐2๐ฅ2)
=
โซโ2
๐โ๐๐1๐1
(๐1๐ฅ1)๐โ๐ ๐2๐2
(๐2๐ฅ2)๐๐(๐1๐ฅ1, ๐2๐ฅ2)
=
โซโ2
๐โ๐๐1๐1
๐ฆ1๐โ๐ ๐2๐2
๐ฆ2๐๐(๐ฆ1, ๐ฆ2)
= โฑ๐๐(๐(๐ฅ1, ๐ฅ2))
(๐1
๐1,๐2
๐2
). โก
We proceed to define the notion of positive definitely function in the contextof quaternion analysis.
Definition 4.6. Let ๐ be a quaternion function on โ2 that is continuous andbounded. For every finite positive measure ๐ on โ2 the function ๐ is said tobe positive definite if
๐โ๐,๐=1,๐<๐
๐ง๐๐ง๐๐(๐๐ โ ๐๐) +๐โ
๐,๐=1,๐<๐
๐(๐๐ โ ๐๐)๐ง๐๐ง๐ +๐โ๐=1
โฃ๐ง๐โฃ2 ๐(โ2) โฅ 0
116 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
for every ๐1, ๐2, . . . , ๐๐ โ โ2, ๐ง1, ๐ง2, . . . , ๐ง๐ โ โ. These parameters are measuredsuch that:
1. When ๐1 = ๐2 = โ โ โ = ๐๐ , and ๐ง1 = ๐ง2 = โ โ โ = ๐ง๐ it follows
2๐(0, 0) + ๐(โ2) โฅ 0;
2. When ๐ = 2, ๐1 = (๐ฅ1, ๐ฅ2), and ๐2 = (0, 0), we have
๐ง1๐ง2๐(๐ฅ1, ๐ฅ2) + ๐(โ๐ฅ1,โ๐ฅ2)๐ง2๐ง1 +(โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2
)๐(โ2) โฅ 0,
which is valid if ๐(โ๐ฅ1,โ๐ฅ2) = ๐(๐ฅ1, ๐ฅ2).
Proposition 4.7. The functional โฑ๐๐(๐) is positive definite and bounded.
Proof. Let ๐1, ๐2, . . . , ๐๐ โ โ2 such that ๐๐ = (๐1๐, ๐2๐), and ๐ง1, ๐ง2, . . . , ๐ง๐ โ โ.Direct computations show that
๐โ๐,๐,๐<๐
๐ง๐๐ง๐โฑ๐๐(๐)(๐๐ โ ๐๐) +
๐โ๐,๐=1,๐<๐
โฑij(๐)(๐๐ โ ๐๐)๐ง๐๐ง๐ +
๐โ๐=1
โฃ๐ง๐โฃ2 ๐ (โ2)
=
โซโ2
๐โ๐,๐=1,๐<๐
๐ง๐๐ง๐๐โ๐(๐1๐โ๐1๐)๐ฅ1๐โ๐(๐2๐โ๐2๐)๐ฅ2๐๐(๐ฅ1, ๐ฅ2)
+
โซโ2
๐โ๐,๐=1,๐<๐
๐โ๐(๐2๐โ๐2๐)๐ฅ2๐โ๐(๐1๐โ๐1๐)๐ฅ1๐ง๐๐ง๐๐๐(๐ฅ1, ๐ฅ2)
+
โซโ2
๐โ๐=1
โฃ๐ง๐โฃ2 ๐๐(๐ฅ1, ๐ฅ2)
=
โซโ2
๐โ๐,๐=1,๐<๐
๐ง๐๐ง๐๐โ๐(๐1๐โ๐1๐)๐ฅ1๐โ๐(๐2๐โ๐2๐)๐ฅ2๐๐(๐ฅ1, ๐ฅ2)
+
โซโ2
๐โ๐,๐=1,๐<๐
๐ง๐๐ง๐๐โ๐(๐1๐โ๐1๐)๐ฅ1๐โ๐(๐2๐โ๐2๐)๐ฅ2๐๐(๐ฅ1, ๐ฅ2)
+
โซโ2
๐โ๐=1
โฃ๐ง๐โฃ2 ๐๐(๐ฅ1, ๐ฅ2)
=
โซโ2
โกโฃ ๐โ๐=1
โฃ๐ง๐โฃ2 + 2
๐โ๐,๐=1,๐<๐
S(๐ง๐๐ง๐๐
โ๐(๐1๐โ๐1๐)๐ฅ1๐โ๐(๐2๐โ๐2๐)๐ฅ2
)โคโฆ ๐๐(๐ฅ1, ๐ฅ2)
โฅโซโ2
โโ ๐โ๐=1
โฃ๐ง๐โฃ2 โ 2
๐โ๐,๐=1,๐<๐
โฃ๐ง๐โฃ โฃ๐ง๐โฃโโ ๐๐(๐ฅ1, ๐ฅ2) โฅ 0.
Furthermore, it follows that
โฃโฑ๐๐(๐)(๐1, ๐2)โฃ =โฃโฃโซ
โ2 ๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2๐๐(๐ฅ1, ๐ฅ2)
โฃโฃ
6. BochnerโMinlos Theorem and Quaternion Fourier Transform 117
โคโซโ2
โฃ๐โ๐๐1๐ฅ1๐โ๐๐2๐ฅ2 โฃ ๐๐(๐ฅ1, ๐ฅ2) = ๐(โ2)< +โ. โก
Likewise we can prove the following proposition.
Proposition 4.8. The functional โฑ๐๐(๐) is positive definite and bounded.
Below we shall assume that ๐ is a probably measure on ๐ โฒ. For every elements๐ โ ๐ and ๐โฒ โ ๐ โฒ we define the functionals ๐๐๐ and ๐๐๐ on ๐ as follows:
๐๐๐ : ๐ โโ ๐ , ๐ ๏ฟฝโโ ๐๐๐(๐) :=
โซ๐ โฒ๐๐โจ๐โฒ,๐โฉ๐๐โจ๐โฒ,๐โฉ๐๐(๐โฒ)
and
๐๐๐ : ๐ โโ ๐ , ๐ ๏ฟฝโโ ๐๐๐(๐) :=
โซ๐ โฒ๐๐โจ๐โฒ,๐โฉ๐๐โจ๐โฒ,๐โฉ๐๐(๐โฒ).
Next we present a generalization of the classical BochnerโMinlos theorem on pos-itive definite functions to the case of quaternion functions.
Theorem 4.9 (BochnerโMinlos theorem). The functional ๐๐๐ satisfies the followingthree conditions:
1. Normalization: ๐๐๐(0) = 1;2. Positivity:
๐โ๐,๐=1,๐<๐
๐ง๐๐ง๐๐๐๐(๐๐ โ ๐๐) +
๐โ๐,๐=1,๐<๐
๐ij(๐๐ โ ๐๐)๐ง๐๐ง๐ +โ๐
โฃ๐ง๐โฃ2 โฅ 0;
3. Continuity: ๐๐๐ is continuous in the sense of Frechet topology.
Proof. We begin the proof by noting that
๐๐๐(0) =
โซ๐ โฒ๐๐(๐โฒ) = ๐(๐ โฒ) = 1.
For simplicity we will prove Statement 2 in the case ๐ = 2 only, i.e., we will provethat
๐ง1๐ง2 ๐๐๐(๐1 โ ๐2) + ๐ij(๐1 โ ๐2)๐ง2๐ง1 + โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2 โฅ 0.
For the sake of convenience we set ๐ง = ๐1 โ ๐2. It follows that
๐ง1๐ง2 ๐๐๐(๐ง) + ๐ij(๐ง)๐ง2๐ง1 + โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2
=
โซ๐ โฒ
(๐ง1๐ง2๐
๐โจ๐โฒ,๐งโฉ๐๐โจ๐โฒ,๐งโฉ + ๐ง1๐ง2๐๐โจ๐โฒ,๐งโฉ๐๐โจ๐โฒ,๐งโฉ + โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2
)๐๐(๐โฒ)
=
โซ๐ โฒ
[2 S(๐ง1๐ง2๐
๐โจ๐โฒ,๐งโฉ๐๐โจ๐โฒ,๐งโฉ + โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2)]
๐๐(๐โฒ).(4.1)
Notice that the last equality follows from the relation
๐ง1๐ง2 ๐๐๐(๐ง) = ๐๐๐(โ๐ง)๐ง2๐ง1.
Now, let ๐ง๐ = โฃ๐ง๐โฃ ๐V(๐ง๐)
โฃV(๐ง๐)โฃ ๐๐ , with ๐๐ = arg(๐ง๐) (๐ = 1, 2). Then
S(๐ง1๐ง2 ๐๐๐(๐ง)) โฅ โ โฃ๐ง1โฃ โฃ๐ง2โฃ .
118 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
From here and (4.1) we obtain
๐ง1๐ง2 ๐๐๐(๐ง) + ๐๐๐(โ๐ง)๐ง2๐ง1 + โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2
โฅโซ๐ โฒ
(โฃ๐ง1โฃ2 + โฃ๐ง2โฃ2 โ 2 โฃ๐ง1โฃ โฃ๐ง2โฃ
)๐๐(๐โฒ) โฅ 0.
Using induction we may conclude that Statement 2 is valid for every natural ๐.For the proof of the remaining statement, let lim๐โโโ ๐๐ = ๐ be understood inthe sense of the topology of Frechet. Then lim๐โโโ ๐๐ = ๐ holds in the usualsense. From here and from the definition of the functional ๐๐๐ we conclude thatlim๐โโโ ๐๐๐(๐๐) = ๐๐๐(๐). Therefore for every ๐ > 0 a natural number ๐ =๐(๐) > 0 can be found so that for ๐ > ๐ the following holds
โฅ๐(๐๐)โ ๐(๐)โฅ๐ < ๐.
Consequently, it follows that
๐ (๐(๐๐), ๐(๐)) =โโ
๐+1
โฅ๐(๐๐)โ ๐(๐)โฅ๐1 + โฅ๐(๐๐)โ ๐(๐)โฅ๐
2โ๐ โค ๐โโ
๐+1
2โ๐. โก
Proposition 4.10. For every element ๐ โ ๐ the functionals ๐๐๐ and ๐๐๐ satisfy theadditional properties:
1. ๐๐๐(โ๐) = ๐๐๐(๐);
2. ๐๐๐(โ๐) = ๐๐๐(๐).
Though the significance of our approach to concrete applications, such as thecharacterization of measurement configurations for functional spaces, was the mainreason for restricting ourselves to the quaternionic case, doubtless, the reduction incalculations for proving the results played an important role too. As was alreadymentioned, it is possible to perform an analogous study to generalized Cliffordalgebras following the same ideas. Further investigations will be presented in aforthcoming paper.
Acknowledgement
Partial support from the Foundation for Science and Technology (FCT) via thegrant DD-VU-02/90, Bulgaria is acknowledged by the first named author. The sec-ond named author acknowledges financial support from the Foundation for Scienceand Technology (FCT) via the post-doctoral grant SFRH/ BPD/66342/2009. Thiswork was supported by FEDER funds through COM PETE โ Operational Pro-gramme Factors of Competitiveness (โPrograma Operacional Factores de Compet-itividadeโ) and by Portuguese funds through the Center for Research and Develop-ment in Mathematics and Applications (University of Aveiro) and the PortugueseFoundation for Science and Technology (โFCT โ Fundacao para a Ciencia e aTecnologiaโ), within project PEst-C/MAT/UI4106/2011 with COMPETE numberFCOMP-01-0124-FEDER-022690. The third named author acknowledges financialsupport from the research grant of the University of Macau No. MYRG142(Y1-L2)-FST11-KKI.
6. BochnerโMinlos Theorem and Quaternion Fourier Transform 119
References
[1] M. Bahri. Generalized Fourier transform in real clifford algebra ๐๐(0, ๐). Far EastJournal of Mathematical Sciences, 48(1):11โ24, Jan. 2011.
[2] P. Bas, N. Le Bihan, and J.M. Chassery. Color image watermarking using quaternionFourier transform. In Proceedings of the IEEE International Conference on AcousticsSpeech and Signal Processing, ICASSP, volume 3, pages 521โ524. Hong-Kong, 2003.
[3] E. Bayro-Corrochano, N. Trujillo, and M. Naranjo. Quaternion Fourier descriptorsfor preprocessing and recognition of spoken words using images of spatiotemporalrepresentations. Mathematical Imaging and Vision, 28(2):179โ190, 2007.
[4] F. Brackx, N. De Schepper, and F. Sommen. The CliffordโFourier transform. Journalof Fourier Analysis and Applications, 11(6):669โ681, 2005.
[5] F. Brackx, N. De Schepper, and F. Sommen. The Fourier transform in Cliffordanalysis. Advances in Imaging and Electron Physics, 156:55โ201, 2009.
[6] F. Brackx, R. Delanghe, and F. Sommen. Clifford Analysis, volume 76. Pitman,Boston, 1982.
[7] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[8] T. Bulow, M. Felsberg, and G. Sommer. Non-commutative hypercomplex Fouriertransforms of multidimensional signals. In G. Sommer, editor, Geometric computingwith Clifford Algebras: Theoretical Foundations and Applications in Computer Visionand Robotics, pages 187โ207, Berlin, 2001. Springer.
[9] J. Ebling and G. Scheuermann. Clifford Fourier transform on vector fields. IEEETransactions on Visualization and Computer Graphics, 11(4):469โ479, July/Aug.2005.
[10] T.A. Ell. Quaternion-Fourier transforms for analysis of 2-dimensional linear time-invariant partial-differential systems. In Proceedings of the 32nd Conference on De-cision and Control, pages 1830โ1841, San Antonio, Texas, USA, 15โ17 December1993. IEEE Control Systems Society.
[11] T.A. Ell and S.J. Sangwine. Hypercomplex Fourier transforms of color images. IEEETransactions on Image Processing, 16(1):22โ35, Jan. 2007.
[12] S. Georgiev. BochnerโMinlos theorem and quaternion Fourier transform. In Gurle-beck [14].
[13] S. Georgiev, J. Morais, and W. Sproรig. Trigonometric integrals in the framework ofquaternionic analysis. In Gurlebeck [14].
[14] K. Gurlebeck, editor. 9th International Conference on Clifford Algebras and theirApplications, Weimar, Germany, 15โ20 July 2011.
[15] K. Gurlebeck and W. Sproรig. Quaternionic Analysis and Elliptic Boundary ValueProblems. Berlin: Akademie-Verlag, Berlin, 1989.
[16] K. Gurlebeck and W. Sproรig. Quaternionic and Clifford Calculus for Physicists andEngineers. Wiley, Aug. 1997.
[17] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
120 S. Georgiev, J. Morais, K.I. Kou and W. Sproรig
[18] E.M.S. Hitzer and B. Mawardi. Clifford Fourier transform on multivector fields anduncertainty principles for dimensions ๐ = 2(mod 4) and ๐ = 3(mod 4). Advances inApplied Clifford Algebras, 18(3-4):715โ736, 2008.
[19] V.V. Kravchenko. Applied Quaternionic Analysis, volume 28 of Research and Expo-sition in Mathematics. Heldermann Verlag, Lemgo, Germany, 2003.
[20] B. Mawardi. Generalized Fourier transform in Clifford algebra ๐โ0,3. Far East Jour-nal of Mathematical Sciences, 44(2):143โ154, 2010.
[21] B. Mawardi, E. Hitzer, A. Hayashi, and R. Ashino. An uncertainty principlefor quaternion Fourier transform. Computers and Mathematics with Applications,56(9):2411โ2417, 2008.
[22] B. Mawardi and E.M.S. Hitzer. Clifford Fourier transformation and uncertainty prin-ciple for the Clifford algebra ๐ถโ3,0. Advances in Applied Clifford Algebras, 16(1):41โ61, 2006.
[23] S.-C. Pei, J.-J. Ding, and J.-H. Chang. Efficient implementation of quaternionFourier transform, convolution, and correlation by 2-D complex FFT. IEEE Trans-actions on Signal Processing, 49(11):2783โ2797, Nov. 2001.
[24] M. Shapiro and N.L. Vasilevski. Quaternionic ๐-hyperholomorphic functions, singu-lar integral operators and boundary value problems I. ๐-hyperholomorphic functiontheory. Complex Variables, Theory and Application, 27(1):17โ46, 1995.
[25] M. Shapiro and N.L. Vasilevski. Quaternionic ๐-hyperholomorphic functions, singu-lar integral operators and boundary value problems II. Algebras of singular integraloperators and Riemann type boundary value problems. Complex Variables, TheoryAppl., 27(1):67โ96, 1995.
S. GeorgievDepartment of Differential Equations, University of SofiaSofia, Bulgariae-mail: [email protected]
J. MoraisCentro de Investigacao e Desenvolvimento em Matematica e Aplicacoes (CIDMA)Universidade de Aveiro, 3810-193 Aveiro, Portugale-mail: [email protected]
K.I. KouDepartment of Mathematics, Faculty of Science and TechnologyUniversity of Macau, Macaue-mail: [email protected]
W. SproรigFreiberg University of Mining and TechnologyFreiberg, Germanye-mail: [email protected]
Part II
Clifford Algebra
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 123โ153cโ 2013 Springer Basel
7 Square Roots of โ1 in RealClifford Algebras
Eckhard Hitzer, Jacques Helmstetter and Rafal Ablamowicz
Abstract. It is well known that Clifford (geometric) algebra offers a geomet-ric interpretation for square roots of โ1 in the form of blades that squareto minus 1. This extends to a geometric interpretation of quaternions as theside face bivectors of a unit cube. Systematic research has been done [33] onthe biquaternion roots of โ1, abandoning the restriction to blades. Biquater-nions are isomorphic to the Clifford (geometric) algebra ๐ถโ3,0 of โ3. Furtherresearch on general algebras ๐ถโ๐,๐ has explicitly derived the geometric rootsof โ1 for ๐+ ๐ โค 4 [20]. The current research abandons this dimension limitand uses the Clifford algebra to matrix algebra isomorphisms in order to al-gebraically characterize the continuous manifolds of square roots of โ1 foundin the different types of Clifford algebras, depending on the type of associ-ated ring (โ, โ, โ2, โ2, or โ). At the end of the chapter explicit computergenerated tables of representative square roots of โ1 are given for all Cliffordalgebras with ๐ = 5, 7, and ๐ = 3 (mod 4) with the associated ring โ. Thisincludes, e.g., ๐ถโ0,5 important in Clifford analysis, and ๐ถโ4,1 which in appli-cations is at the foundation of conformal geometric algebra. All these rootsof โ1 are immediately useful in the construction of new types of geometricCliffordโFourier transformations.
Mathematics Subject Classification (2010). Primary 15A66; secondary 11E88,42A38, 30G35.
Keywords. Algebra automorphism, inner automorphism, center, centralizer,Clifford algebra, conjugacy class, determinant, primitive idempotent, trace.
1. Introduction
The young London Goldsmith professor of applied mathematics W.K. Cliffordcreated his geometric algebras1 in 1878 inspired by the works of Hamilton on
1In his original publication [11] Clifford first used the term geometric algebras. Subsequently inmathematics the new term Clifford algebras [28] has become the proper mathematical term. Foremphasizing the geometric nature of the algebra, some researchers continue [9, 16, 17] to use theoriginal term geometric algebra(s).
124 E. Hitzer, J. Helmstetter and R. Ablamowicz
quaternions and by Grassmannโs exterior algebra. Grassmann invented the anti-symmetric outer product of vectors, that regards the oriented parallelogram areaspanned by two vectors as a new type of number, commonly called bivector. Thebivector represents its own plane, because outer products with vectors in the planevanish. In three dimensions the outer product of three linearly independent vectorsdefines a so-called trivector with the magnitude of the volume of the parallelepipedspanned by the vectors. Its orientation (sign) depends on the handedness of thethree vectors.
In the Clifford algebra [16] of โ3 the three bivector side faces of a unit cube{๐1๐2, ๐2๐3, ๐3๐1} oriented along the three coordinate directions {๐1, ๐2, ๐3} cor-respond to the three quaternion units ๐, ๐, and ๐. Like quaternions, these threebivectors square to minus one and generate the rotations in their respective planes.
Beyond that Clifford algebra allows to extend complex numbers to higherdimensions [7, 17] and systematically generalize our knowledge of complex num-bers, holomorphic functions and quaternions into the realm of Clifford analysis.It has found rich applications in symbolic computation, physics, robotics, com-puter graphics, etc. [8, 9, 12, 14, 27]. Since bivectors and trivectors in the Cliffordalgebras of Euclidean vector spaces square to minus one, we can use them tocreate new geometric kernels for Fourier transformations. This leads to a largevariety of new Fourier transformations, which all deserve to be studied in theirown right [5, 6, 9, 13, 18, 19, 22,23,26, 29โ32].
In our current research we will treat square roots of โ1 in Clifford algebras๐ถโ๐,๐ of both Euclidean (positive definite metric) and non-Euclidean (indefinitemetric) non-degenerate vector spaces, โ๐ = โ๐,0 and โ๐,๐, respectively. We knowfrom Einsteinโs special theory of relativity that non-Euclidean vector spaces are offundamental importance in nature [15]. They are further, e.g., used in computervision and robotics [12] and for general algebraic solutions to contact problems [27].Therefore this chapter is about characterizing square roots of โ1 in all Cliffordalgebras ๐ถโ๐,๐, extending previous limited research on ๐ถโ3,0 in [33] and ๐ถโ๐,๐, ๐ =๐ + ๐ โค 4 in [20]. The manifolds of square roots of โ1 in ๐ถโ๐,๐, ๐ = ๐ + ๐ = 2,compare Table 1 of [20], are visualized in Figure 1.
First, we introduce necessary background knowledge of Clifford algebras andmatrix ring isomorphisms and explain in more detail how we will characterize andclassify the square roots of โ1 in Clifford algebras in Section 2. Next, we treatsection by section (in Sections 3 to 7) the square roots of โ1 in Clifford algebraswhich are isomorphic to matrix algebras with associated rings โ, โ, โ2, โ2, andโ, respectively. The term associated means that the isomorphic matrices will onlyhave matrix elements from the associated ring. The square roots of โ1 in Section 7with associated ring โ are of particular interest, because of the existence of classesof exceptional square roots of โ1, which all include a nontrivial term in the centralelement of the respective algebra different from the identity. Section 7 thereforeincludes a detailed discussion of all classes of square roots of โ1 in the algebras๐ถโ4,1, the isomorphic ๐ถโ0,5, and in ๐ถโ7,0. Finally, we add Appendix A with tablesof square roots of โ1 for all Clifford algebras with ๐ = 5, 7, and ๐ = 3 (mod 4).
7. Square Roots of โ1 in Real Clifford Algebras 125
Figure 1. Manifolds of square roots ๐ of โ1 in ๐ถโ2,0 (left), ๐ถโ1,1 (cen-ter), and ๐ถโ0,2 โผ= โ (right). The square roots are ๐ = ๐ผ+ ๐1๐1 + ๐2๐2 +๐ฝ๐12, with ๐ผ, ๐1, ๐2, ๐ฝ โ โ, ๐ผ = 0, and ๐ฝ2 = ๐21๐
22 + ๐22๐
21 + ๐21๐
22.
The square roots of โ1 in Section 7 and in Appendix A were all computed withthe Maple package CLIFFORD [2], as explained in Appendix B.
2. Background and Problem Formulation
Let ๐ถโ๐,๐ be the algebra (associative with unit 1) generated overโ by ๐+๐ elements๐๐ (with ๐ = 1, 2, . . . , ๐ + ๐) with the relations ๐2๐ = 1 if ๐ โค ๐, ๐2๐ = โ1 if ๐ > ๐and ๐โ๐๐ + ๐๐๐โ = 0 whenever โ โ= ๐, see [28]. We set the vector space dimension๐ = ๐+ ๐ and the signature ๐ = ๐โ ๐. This algebra has dimension 2๐, and its evensubalgebra ๐ถโ0(๐, ๐) has dimension 2๐โ1 (if ๐ > 0). We are concerned with squareroots of โ1 contained in ๐ถโ๐,๐ or ๐ถโ0(๐, ๐). If the dimension of ๐ถโ๐,๐ or, ๐ถโ0(๐, ๐)is โค 2, it is isomorphic to โ โผ= ๐ถโ0,0, โ
2 โผ= ๐ถโ1,0, or โ โผ= ๐ถโ0,1, and it is clear thatthere is no square root of โ1 in โ and โ2 = โรโ, and that there are two squaresroots ๐ and โ๐ in โ. Therefore we only consider algebras of dimension โฅ 4. Squareroots of โ1 have been computed explicitly in [33] for ๐ถโ3,0, and in [20] for algebrasof dimensions 2๐ โค 16.
An algebra ๐ถโ๐,๐ or ๐ถโ0(๐, ๐) of dimension โฅ 4 is isomorphic to one of thefive matrix algebras:โณ(2๐,โ),โณ(๐,โ),โณ(2๐,โ2),โณ(๐,โ2) orโณ(2๐,โ). Theinteger ๐ depends on ๐. According to the parity of ๐, it is either 2(๐โ2)/2 or 2(๐โ3)/2
for ๐ถโ๐,๐, and, either 2(๐โ4)/2 or 2(๐โ3)/2 for ๐ถโ0(๐, ๐). The associated ring (either
โ, โ, โ2, โ2, or โ) depends on ๐ in this way2:
2Compare Chapter 16 on matrix representations and periodicity of 8, as well as Table 1 on p.217 of [28].
126 E. Hitzer, J. Helmstetter and R. Ablamowicz
๐ mod 8 0 1 2 3 4 5 6 7
associated ring for ๐ถโ๐,๐ โ โ2 โ โ โ โ2 โ โ
associated ring for ๐ถโ0(๐, ๐) โ2 โ โ โ โ2 โ โ โ
Therefore we shall answer the following question: what can we say about the squareroots of โ1 in an algebra ๐ that is isomorphic toโณ(2๐,โ),โณ(๐,โ),โณ(2๐,โ2),โณ(๐,โ2), or, โณ(2๐,โ)? They constitute an algebraic submanifold in ๐; howmany connected components3 (for the usual topology) does it contain? Which aretheir dimensions? This submanifold is invariant by the action of the group Inn(๐)of inner automorphisms4 of ๐, i.e., for every ๐ โ ๐, ๐2 = โ1โ ๐(๐)2 = โ1 โ๐ โInn(๐). The orbits of Inn(๐) are called conjugacy classes5; how many conjugacyclasses are there in this submanifold? If the associated ring is โ2 or โ2 or โ, thegroup Aut(๐) of all automorphisms of ๐ is larger than Inn(๐), and the action ofAut(๐) in this submanifold shall also be described.
We recall some properties of ๐ that do not depend on the associated ring.The group Inn(๐) contains as many connected components as the group G(๐) ofinvertible elements in ๐. We recall that this assertion is true for โณ(2๐,โ) butnot for โณ(2๐ + 1,โ) which is not one of the relevant matrix algebras. If ๐ is anelement of ๐, let Cent(๐) be the centralizer of ๐ , that is, the subalgebra of all๐ โ ๐ such that ๐๐ = ๐๐ . The conjugacy class of ๐ contains as many connectedcomponents6 as G(๐) if (and only if) Cent(๐)
โฉG(๐) is contained in the neutral7
connected component of G(๐), and the dimension of its conjugacy class is
dim(๐)โ dim(Cent(๐)). (2.1)
Note that for invertible ๐ โ Cent(๐) we have ๐โ1๐๐ = ๐ .Besides, let Z(๐) be the center of ๐, and let [๐,๐] be the subspace spanned
by all [๐, ๐] = ๐๐ โ ๐๐ . In all cases ๐ is the direct sum of Z(๐) and [๐,๐]. For
3Two points are in the same connected component of a manifold, if they can be joined by acontinuous path inside the manifold under consideration. (This applies to all topological spacessatisfying the property that each neighborhood of any point contains a neighborhood in whichevery pair of points can always be joined by a continuous path.)4An inner automorphism ๐ of ๐ is defined as ๐ : ๐ โ ๐, ๐(๐ฅ) = ๐โ1๐ฅ๐, โ๐ฅ โ ๐, with givenfixed ๐ โ ๐. The composition of two inner automorphisms ๐(๐(๐ฅ)) = ๐โ1๐โ1๐ฅ๐๐ = (๐๐)โ1๐ฅ(๐๐)is again an inner automorphism. With this operation the inner automorphisms form the groupInn(๐), compare [35].5The conjugacy class (similarity class) of a given ๐ โ ๐, ๐2 = โ1 is {๐(๐) : ๐ โ Inn(๐)}, compare[34]. Conjugation is transitive, because the composition of inner automorphisms is again an innerautomorphism.6According to the general theory of groups acting on sets, the conjugacy class (as a topologicalspace) of a square root ๐ of โ1 is isomorphic to the quotient of G(๐) and Cent(๐) (the subgroupof stability of ๐). Quotient means here the set of left handed classes modulo the subgroup. Ifthe subgroup is contained in the neutral connected component of G(๐), then the number of
connected components is the same in the quotient as in G(๐). See also [10].7Neutral means to be connected to the identity element of ๐.
7. Square Roots of โ1 in Real Clifford Algebras 127
example,8 Z(โณ(2๐,โ)) = {๐1 โฃ ๐ โ โ} and Z(โณ(2๐,โ)) = {๐1 โฃ ๐ โ โ}. If theassociated ring is โ or โ (that is for even ๐), then Z(๐) is canonically isomorphicto โ, and from the projection ๐ โ Z(๐) we derive a linear form Scal : ๐ โ โ.When the associated ring9 is โ2 or โ2 or โ, then Z(๐) is spanned by 1 (the unitmatrix10) and some element ๐ such that ๐2 = ยฑ1. Thus, we get two linear formsScal and Spec such that Scal(๐)1+Spec(๐)๐ is the projection of ๐ in Z(๐) for every๐ โ ๐. Instead of ๐ we may use โ๐ and replace Spec with โSpec. The followingassertion holds for every ๐ โ ๐: The trace of each multiplication11 ๐ ๏ฟฝโ ๐๐ or๐ ๏ฟฝโ ๐๐ is equal to the product
tr(๐) = dim(๐) Scal(๐). (2.2)
The word โtraceโ (when nothing more is specified) means a matrix trace in โ,which is the sum of its diagonal elements. For example, the matrix ๐ โ โณ(2๐,โ)
with elements ๐๐๐ โ โ, 1 โค ๐, ๐ โค 2๐ has the trace tr(๐) =โ2๐
๐=1 ๐๐๐ [24].
We shall prove that in all cases Scal(๐) = 0 for every square root of โ1 in๐. Then, we may distinguish ordinary square roots of โ1, and exceptional ones.In all cases the ordinary square roots of โ1 constitute a unique12 conjugacy classof dimension dim(๐)/2 which has as many connected components as G(๐), andthey satisfy the equality Spec(๐) = 0 if the associated ring is โ2 or โ2 or โ. Theexceptional square roots of โ1 only exist13 if ๐ โผ= โณ(2๐,โ). In โณ(2๐,โ) thereare 2๐ conjugacy classes of exceptional square roots of โ1, each one characterizedby an equality Spec(๐) = ๐/๐ with ยฑ๐ โ {1, 2, . . . , ๐} [see Section 7], and theirdimensions are< dim(๐)/2 [see equation (7.5)]. For instance, ๐ (mentioned above)and โ๐ are central square roots of โ1 inโณ(2๐,โ) which constitute two conjugacyclasses of dimension 0. Obviously, Spec(๐) = 1.
For symbolic computer algebra systems (CAS), like MAPLE, there exist Clif-ford algebra packages, e.g., CLIFFORD [2], which can compute idempotents [3]and square roots of โ1. This will be of especial interest for the exceptional squareroots of โ1 in โณ(2๐,โ).
Regarding a square root ๐ of โ1, a Clifford algebra is the direct sum of thesubspaces Cent(๐) (all elements that commute with ๐) and the skew-centralizer
8A matrix algebra based proof is, e.g., given in [4].9This is the case for ๐ (and ๐ ) odd. Then the pseudoscalar ๐ โ ๐ถโ๐,๐ is also in Z(๐ถโ๐,๐).10The number 1 denotes the unit of the Clifford algebra ๐, whereas the bold face 1 denotes theunit of the isomorphic matrix algebra โณ.11These multiplications are bilinear over the center of ๐.12Let ๐ be an algebra โณ(๐,๐) where ๐ is a division ring. Thus two elements ๐ and ๐ of ๐induce ๐-linear endomorphisms ๐ โฒ and ๐โฒ on ๐๐; if ๐ is not commutative, ๐ operates on ๐๐
on the right side. The matrices ๐ and ๐ are conjugate (or similar) if and only if there are two๐-bases ๐ต1 and ๐ต2 of ๐๐ such that ๐ โฒ operates on ๐ต1 in the same way as ๐โฒ operates on ๐ต2.This theorem allows us to recognize that in all cases but the last one (with exceptional squareroots of โ1), two square roots of โ1 are always conjugate.13The pseudoscalars of Clifford algebras whose isomorphic matrix algebra has ring โ2 or โ2
square to ๐2 = +1.
128 E. Hitzer, J. Helmstetter and R. Ablamowicz
SCent(๐) (all elements that anticommute with ๐). Every Clifford algebra multivec-tor has a unique split by this Lemma.
Lemma 2.1. Every multivector ๐ด โ ๐ถโ๐,๐ has, with respect to a square root ๐ โ๐ถโ๐,๐ of โ1, i.e., ๐โ1 = โ๐, the unique decomposition
๐ดยฑ =1
2(๐ดยฑ ๐โ1๐ด๐), ๐ด = ๐ด+ +๐ดโ, ๐ด+๐ = ๐๐ด+, ๐ดโ๐ = โ๐๐ดโ. (2.3)
Proof. For ๐ด โ ๐ถโ๐,๐ and a square root ๐ โ ๐ถโ๐,๐ of โ1, we compute
๐ดยฑ๐ =1
2(๐ดยฑ ๐โ1๐ด๐)๐ =
1
2(๐ด๐ ยฑ ๐โ1๐ด(โ1)) ๐โ1=โ๐
=1
2(๐๐โ1๐ด๐ ยฑ ๐๐ด)
= ยฑ๐1
2(๐ดยฑ ๐โ1๐ด๐). โก
For example, in Clifford algebras ๐ถโ๐,0 [23] of dimensions ๐ = 2 mod 4,Cent(๐) is the even subalgebra ๐ถโ0(๐, 0) for the unit pseudoscalar ๐, and thesubspace ๐ถโ1(๐, 0) spanned by all ๐-vectors of odd degree ๐, is SCent(๐). Themost interesting case is โณ(2๐,โ), where a whole range of conjugacy classes be-comes available. These results will therefore be particularly relevant for construct-ing CliffordโFourier transformations using the square roots of โ1.
3. Square Roots of โ1 in ํ(2๐ ,โ)
Here ๐ = โณ(2๐,โ), whence dim(๐) = (2๐)2 = 4๐2. The group G(๐) has twoconnected components determined by the inequalities det(๐) > 0 and det(๐) < 0.
For the case ๐ = 1 we have, e.g., the algebra ๐ถโ2,0 isomorphic to โณ(2,โ).The basis {1, ๐1, ๐2, ๐12} of ๐ถโ2,0 is mapped to{(
1 00 1
),
(0 11 0
),
(1 00 โ1
),
(0 โ11 0
)}.
The general element ๐ผ+ ๐1๐1 + ๐2๐2 + ๐ฝ๐12 โ ๐ถโ2,0 is thus mapped to(๐ผ+ ๐2 โ๐ฝ + ๐1๐ฝ + ๐1 ๐ผโ ๐2
)(3.1)
in โณ(2,โ). Every element ๐ of ๐ =โณ(2๐,โ) is treated as an โ-linear endomor-phism of ๐ = โ2๐. Thus, its scalar component and its trace (2.2) are related asfollows: tr(๐) = 2๐Scal(๐). If ๐ is a square root of โ1, it turns ๐ into a vectorspace over โ (if the complex number ๐ operates like ๐ on ๐ ). If (๐1, ๐2, . . . , ๐๐) isa โ-basis of ๐ , then (๐1, ๐(๐1), ๐2, ๐(๐2), . . . , ๐๐, ๐(๐๐)) is an โ-basis of ๐ , and the2๐ร 2๐ matrix of ๐ in this basis is
diag
((0 โ11 0
), . . . ,
(0 โ11 0
)๏ธธ ๏ธท๏ธท ๏ธธ
๐
)(3.2)
Consequently all square roots of โ1 in ๐ are conjugate. The centralizer of asquare root ๐ of โ1 is the algebra of all โ-linear endomorphisms ๐ of ๐ (since
7. Square Roots of โ1 in Real Clifford Algebras 129
๐ operates like ๐ on ๐ ). Therefore, the โ-dimension of Cent(๐) is ๐2 and itsโ-dimension is 2๐2. Finally, the dimension (2.1) of the conjugacy class of ๐ isdim(๐) โ dim(Cent(๐)) = 4๐2 โ 2๐2 = 2๐2 = dim(๐)/2. The two connected com-ponents of G(๐) are determined by the sign of the determinant. Because of thenext lemma, the โ-determinant of every element of Cent(๐) is โฅ 0. Therefore, theintersection Cent(๐)
โฉG(๐) is contained in the neutral connected component of
G(๐) and, consequently, the conjugacy class of ๐ has two connected componentslike G(๐). Because of the next lemma, the โ-trace of ๐ vanishes (indeed its โ-traceis ๐๐, because ๐ is the multiplication by the scalar ๐: ๐(๐ฃ) = ๐๐ฃ for all ๐ฃ) whenceScal(๐) = 0. This equality is corroborated by the matrix written above.
We conclude that the square roots of โ1 constitute one conjugacy class withtwo connected components of dimension dim(๐)/2 contained in the hyperplanedefined by the equation
Scal(๐) = 0. (3.3)
Before stating the lemma that here is so helpful, we show what happens inthe easiest case ๐ = 1. The square roots of โ1 in โณ(2,โ) are the real matrices(
๐ ๐๐ โ๐
)with
(๐ ๐๐ โ๐
)(๐ ๐๐ โ๐
)= (๐2 + ๐๐)1 = โ1; (3.4)
hence ๐2 + ๐๐ = โ1, a relation between ๐, ๐, ๐ which is equivalent to (๐ โ ๐)2 =(๐ + ๐)2 + 4๐2 + 4 โ (๐ โ ๐)2 โฅ 4 โ ๐ โ ๐ โฅ 2 (one component) or ๐ โ ๐ โฅ 2(second component). Thus, we recognize the two connected components of squareroots of โ1: The inequality ๐ โฅ ๐+ 2 holds in one connected component, and theinequality ๐ โฅ ๐+ 2 in the other one, compare Figure 2.
Figure 2. Two components of square roots of โ1 in โณ(2,โ).
In terms of ๐ถโ2,0 coefficients (3.1) with ๐โ ๐ = ๐ฝ + ๐1 โ (โ๐ฝ + ๐1) = 2๐ฝ, weget the two component conditions simply as
๐ฝ โฅ 1 (one component), ๐ฝ โค โ1 (second component). (3.5)
Rotations (det(๐) = 1) leave the pseudoscalar ๐ฝ๐12 invariant (and thus preserve thetwo connected components of square roots of โ1), but reflections (det(๐โฒ) = โ1)change its sign ๐ฝ๐12 โ โ๐ฝ๐12 (thus interchanging the two components).
130 E. Hitzer, J. Helmstetter and R. Ablamowicz
Because of the previous argument involving a complex structure on the realspace ๐ , we conversely consider the complex space โ๐ with its structure of vectorspace over โ. If (๐1, ๐2, . . . , ๐๐) is a โ-basis of โ๐, then (๐1, ๐๐1, ๐2, ๐๐2, . . . , ๐๐, ๐๐๐)is an โ-basis. Let ๐ be a โ-linear endomorphism of โ๐ (i.e., a complex ๐ ร ๐matrix), let trโ(๐) and detโ(๐) be the trace and determinant of ๐ in โ, and trโ(๐)and detโ(๐) its trace and determinant for the real structure of โ๐.
Example. For ๐ = 1 an endomorphism of โ1 is given by a complex number ๐ =๐+ ๐๐, ๐, ๐ โ โ. Its matrix representation is according to (3.2)(
๐ โ๐๐ ๐
)with
(๐ โ๐๐ ๐
)2
= (๐2 โ ๐2)
(1 00 1
)+ 2๐๐
(0 โ11 0
). (3.6)
Then we have trโ(๐) = ๐ + ๐๐, trโ
(๐ โ๐๐ ๐
)= 2๐ = 2โ(trโ(๐)) and detโ(๐) =
๐+ ๐๐, detโ
(๐ โ๐๐ ๐
)= ๐2 + ๐2 = โฃ detโ(๐)โฃ2 โฅ 0.
Lemma 3.1. For every โ-linear endomorphism ๐ we can write trโ(๐) = 2โ(trโ(๐))and detโ(๐) = โฃ detโ(๐)โฃ2 โฅ 0.
Proof. There is a โ-basis in which the โ-matrix of ๐ is triangular [then detโ(๐)is the product of the entries of ๐ on the main diagonal]. We get the โ-matrixof ๐ in the derived โ-basis by replacing every entry ๐ + ๐๐ of the โ-matrix with
the elementary matrix
(๐ โ๐๐ ๐
). The conclusion soon follows. The fact that the
determinant of a block triangular matrix is the product of the determinants of theblocks on the main diagonal is used. โก
4. Square Roots of โ1 in ํ(2๐ ,โ2)
Here ๐ = โณ(2๐,โ2) = โณ(2๐,โ) ร โณ(2๐,โ), whence dim(๐) = 8๐2. Thegroup G(๐) has four14 connected components. Every element (๐, ๐ โฒ) โ ๐ (with๐, ๐ โฒ โ โณ(2๐,โ)) has a determinant in โ2 which is obviously (det(๐), det(๐ โฒ)),and the four connected components of G(๐) are determined by the signs of thetwo components of detโ2(๐, ๐ โฒ).
The lowest-dimensional example (๐ = 1) is ๐ถโ2,1 isomorphic to โณ(2,โ2).Here the pseudoscalar ๐ = ๐123 has square ๐2 = +1. The center of the algebra is{1, ๐} and includes the idempotents ๐ยฑ = (1ยฑ๐)/2, ๐2ยฑ = ๐ยฑ, ๐+๐โ = ๐โ๐+ = 0.The basis of the algebra can thus be written as {๐+, ๐1๐+, ๐2๐+, ๐12๐+, ๐โ, ๐1๐โ,๐2๐โ, ๐12๐โ}, where the first (and the last) four elements form a basis of the
14In general, the number of connected components of G(๐) is two if ๐ = โณ(๐,โ), and one if๐ = โณ(๐,โ) or ๐ = โณ(๐,โ), because in all cases every matrix can be joined by a continuouspath to a diagonal matrix with entries 1 or โ1. When an algebra ๐ is a direct product of two
algebras โฌ and ๐, then G(๐) is the direct product of G(โฌ) and G(๐), and the number of connectedcomponents of G(๐) is the product of the numbers of connected components of G(โฌ) and G(๐).
7. Square Roots of โ1 in Real Clifford Algebras 131
subalgebra ๐ถโ2,0 isomorphic toโณ(2,โ). In terms of matrices we have the identitymatrix (1,1) representing the scalar part, the idempotent matrices (1, 0), (0,1),and the ๐ matrix (1,โ1), with 1 the unit matrix of โณ(2,โ).
The square roots of (โ1,โ1) in ๐ are pairs of two square roots of โ1 inโณ(2๐,โ). Consequently they constitute a unique conjugacy class with four con-nected components of dimension 4๐2 = dim(๐)/2. This number can be obtained intwo ways. First, since every element (๐, ๐ โฒ) โ ๐ (with ๐, ๐ โฒ โโณ(2๐,โ)) has twicethe dimension of the components ๐ โโณ(2๐,โ) of Section 3, we get the componentdimension 2โ 2๐2 = 4๐2. Second, the centralizer Cent(๐, ๐ โฒ) has twice the dimensionof Cent(๐) ofโณ(2๐,โ), therefore dim(๐)โCent(๐, ๐ โฒ) = 8๐2 โ 4๐2 = 4๐2. In theabove example for ๐ = 1 the four components are characterized according to (3.5)by the values of the coefficients of ๐ฝ๐12๐+ and ๐ฝโฒ๐12๐โ as
๐1 : ๐ฝ โฅ 1, ๐ฝโฒ โฅ 1,
๐2 : ๐ฝ โฅ 1, ๐ฝโฒ โค โ1,๐3 : ๐ฝ โค โ1, ๐ฝโฒ โฅ 1,
๐4 : ๐ฝ โค โ1, ๐ฝโฒ โค โ1. (4.1)
For every (๐, ๐ โฒ) โ ๐ we can with (2.2) write tr(๐) + tr(๐ โฒ) = 2๐Scal(๐, ๐ โฒ) and
tr(๐)โ tr(๐ โฒ) = 2๐Spec(๐, ๐ โฒ) if ๐ = (1,โ1); (4.2)
whence Scal(๐, ๐ โฒ) = Spec(๐, ๐ โฒ) = 0 if (๐, ๐ โฒ) is a square root of (โ1,โ1), compare(3.3).
The group Aut(๐) is larger than Inn(๐), because it contains the swap auto-morphism (๐, ๐ โฒ) ๏ฟฝโ (๐ โฒ, ๐) which maps the central element ๐ to โ๐, and inter-changes the two idempotents ๐+ and ๐โ. The group Aut(๐) has eight connectedcomponents which permute the four connected components of the submanifold ofsquare roots of (โ1,โ1). The permutations induced by Inn(๐) are the permu-tations of the Klein group. For example for ๐ = 1 of (4.1) we get the followingInn(โณ(2,โ2)) permutations
det(๐) > 0, det(๐โฒ) > 0 : identity,
det(๐) > 0, det(๐โฒ) < 0 : (๐1, ๐2), (๐3, ๐4),
det(๐) < 0, det(๐โฒ) > 0 : (๐1, ๐3), (๐2, ๐4),
det(๐) < 0, det(๐โฒ) < 0 : (๐1, ๐4), (๐2, ๐3). (4.3)
Beside the identity permutation, Inn(๐) gives the three permutations that permutetwo elements and also the other two ones.
The automorphisms outside Inn(๐) are(๐, ๐ โฒ) ๏ฟฝโ (๐๐ โฒ๐โ1, ๐โฒ๐๐โฒโ1) for some (๐, ๐โฒ) โ G(๐). (4.4)
If det(๐) and det(๐โฒ) have opposite signs, it is easy to realize that this automor-phism induces a circular permutation on the four connected components of squareroots of (โ1,โ1): If det(๐) and det(๐โฒ) have the same sign, this automorphismleaves globally invariant two connected components, and permutes the other two
132 E. Hitzer, J. Helmstetter and R. Ablamowicz
ones. For example, for ๐ = 1 the automorphisms (4.4) outside Inn(๐) permute thecomponents (4.1) of square roots of (โ1,โ1) in โณ(2,โ2) as follows
det(๐) > 0, det(๐โฒ) > 0 : (๐1), (๐2, ๐3), (๐4),
det(๐) > 0, det(๐โฒ) < 0 : ๐1 โ ๐2 โ ๐4 โ ๐3 โ ๐1,
det(๐) < 0, det(๐โฒ) > 0 : ๐1 โ ๐3 โ ๐4 โ ๐2 โ ๐1,
det(๐) < 0, det(๐โฒ) < 0 : (๐1, ๐4), (๐2), (๐3). (4.5)
Consequently, the quotient of the group Aut(๐) by its neutral connected compo-nent is isomorphic to the group of isometries of a square in a Euclidean plane.
5. Square Roots of โ1 in ํ(๐ ,โ)
Let us first consider the easiest case ๐ = 1, when ๐ = โ, e.g., of ๐ถโ0,2. The squareroots of โ1 in โ are the quaternions ๐๐ + ๐๐ + ๐๐๐ with ๐2 + ๐2 + ๐2 = 1. Theyconstitute a compact and connected manifold of dimension 2. Every square root๐ of โ1 is conjugate with ๐, i.e., there exists ๐ฃ โ โ : ๐ฃโ1๐๐ฃ = ๐ โ ๐๐ฃ = ๐ฃ๐. If weset ๐ฃ = โ๐๐+ 1 = ๐+ ๐๐๐ โ ๐๐ + 1 we have
๐๐ฃ = โ๐2๐+ ๐ = ๐ + ๐ = (๐(โ๐) + 1)๐ = ๐ฃ๐.
๐ฃ is invertible, except when ๐ = โ๐. But ๐ is conjugate with โ๐ because ๐๐ = ๐(โ๐),hence, by transitivity ๐ is also conjugate with โ๐.
Here ๐ = โณ(๐,โ), whence dim(๐ด) = 4๐2. The ring โ is the algebra overโ generated by two elements ๐ and ๐ such that ๐2 = ๐2 = โ1 and ๐๐ = โ๐๐. Weidentify โ with the subalgebra generated by15 ๐ alone.
The group G(๐) has only one connected component. We shall soon prove thatevery square root of โ1 in ๐ is conjugate with ๐1. Therefore, the submanifold ofsquare roots of โ1 is a conjugacy class, and it is connected. The centralizer of๐1 in ๐ is the subalgebra of all matrices with entries in โ. The โ-dimension ofCent(๐1) is ๐2, its โ-dimension is 2๐2, and, consequently, the dimension (2.1) ofthe submanifold of square roots of โ1 is 4๐2 โ 2๐2 = 2๐2 = dim(๐)/2.
Here ๐ = โ๐ is treated as a (unitary) module over โ on the right side:The product of a line vector ๐ก๐ฃ = (๐ฅ1, ๐ฅ2, . . . , ๐ฅ๐) โ ๐ by ๐ฆ โ โ is ๐ก๐ฃ ๐ฆ =(๐ฅ1๐ฆ, ๐ฅ2๐ฆ, . . . , ๐ฅ๐๐ฆ). Thus, every ๐ โ ๐ determines an โ-linear endomorphism of๐ : The matrix ๐ multiplies the column vector ๐ฃ = ๐ก(๐ฅ1, ๐ฅ2, . . . , ๐ฅ๐) on the leftside ๐ฃ ๏ฟฝโ ๐๐ฃ. Since โ is a subring of โ, ๐ is also a vector space of dimension 2๐over โ. The scalar ๐ always operates on the right side (like every scalar in โ). If(๐1, ๐2, . . . , ๐๐) is an โ-basis of ๐ , then (๐1, ๐1๐, ๐2, ๐2๐, . . . , ๐๐, ๐๐๐) is a โ-basis of๐ . Let ๐ be a square root of โ1, then the eigenvalues of ๐ in โ are +๐ or โ๐.If we treat ๐ as a 2๐ vector space over โ, it is the direct (โ-linear) sum of theeigenspaces
๐ + = {๐ฃ โ ๐ โฃ ๐(๐ฃ) = ๐ฃ๐} and ๐ โ = {๐ฃ โ ๐ โฃ ๐(๐ฃ) = โ๐ฃ๐}, (5.1)
15This choice is usual and convenient.
7. Square Roots of โ1 in Real Clifford Algebras 133
representing ๐ as a 2๐ร 2๐ โ-matrix w.r.t. the โ-basis of ๐ , with โ-scalar eigen-values (multiplied from the right): ๐ยฑ = ยฑ๐.
Since ๐๐ = โ๐๐, the multiplication ๐ฃ ๏ฟฝโ ๐ฃ๐ permutes ๐ + and ๐ โ, as ๐(๐ฃ) =ยฑ๐ฃ๐ is mapped to ๐(๐ฃ)๐ = ยฑ๐ฃ๐๐ = โ(๐ฃ๐)๐. Therefore, if (๐1, ๐2, . . . , ๐๐) is a โ-basisof ๐ +, then (๐1๐, ๐2๐, . . . , ๐๐๐) is a โ-basis of ๐ โ, consequently (๐1, ๐1๐, ๐2, ๐2๐,. . ., ๐๐, ๐๐๐) is a โ-basis of ๐ , and (๐1, ๐2, . . . , ๐๐=๐) is an โ-basis of ๐ . Since ๐ by๐(๐๐) = ๐๐๐ for ๐ = 1, 2, . . . , ๐ operates on the โ-basis (๐1, ๐2, . . . , ๐๐) in the sameway as ๐1 on the natural โ-basis of ๐ , we conclude that ๐ and ๐1 are conjugate.
Besides, Scal(๐1) = 0 because 2๐1 = [๐1, ๐๐1] โ [๐,๐], thus ๐1 /โ Z(๐).Whence,16
Scal(๐) = 0 for every square root of โ 1. (5.2)
These results are easily verified in the above example of ๐ = 1 when ๐ = โ.
6. Square Roots of โ1 in ํ(๐ ,โ2)
Here, ๐ = โณ(๐,โ2) = โณ(๐,โ) รโณ(๐,โ), whence dim(๐ด) = 8๐2. The groupG(๐) has only one connected component (see Footnote 14).
The square roots of (โ1,โ1) in ๐ are pairs of two square roots of โ1 inโณ(๐,โ). Consequently, they constitute a unique conjugacy class which is con-nected and its dimension is 2ร 2๐2 = 4๐2 = dim(๐)/2.
For every (๐, ๐ โฒ) โ ๐ we can write Scal(๐) + Scal(๐ โฒ) = 2 Scal(๐, ๐ โฒ) and,similarly to (4.2),
Scal(๐)โ Scal(๐ โฒ) = 2 Spec(๐, ๐ โฒ) if ๐ = (1,โ1); (6.1)
whence Scal(๐, ๐ โฒ) = Spec(๐, ๐ โฒ) = 0 if (๐, ๐ โฒ) is a square root of (โ1,โ1), comparewith (5.2).
The group Aut(๐) has two17 connected components; the neutral componentis Inn(๐), and the other component contains the swap automorphism (๐, ๐ โฒ) ๏ฟฝโ(๐ โฒ, ๐).
The simplest example is ๐ = 1, ๐ = โ2, where we have the identity pair(1, 1) representing the scalar part, the idempotents (1, 0), (0, 1), and ๐ as the pair(1,โ1).
๐ = โ2 is isomorphic to ๐ถโ0,3. The pseudoscalar ๐ = ๐123 has the square๐2 = +1. The center of the algebra is {1, ๐}, and includes the idempotents ๐ยฑ =12 (1ยฑ๐), ๐2ยฑ = ๐ยฑ, ๐+๐โ = ๐โ๐+ = 0. The basis of the algebra can thus be writtenas {๐+, ๐1๐+, ๐2๐+, ๐12๐+, ๐โ, ๐1๐โ, ๐2๐โ, ๐12๐โ} where the first (and the last) fourelements form a basis of the subalgebra ๐ถโ0,2 isomorphic to โ.
16Compare the definition of Scal(๐) in Section 2, remembering that in the current section the
associated ring is โ.17Compare Footnote 14.
134 E. Hitzer, J. Helmstetter and R. Ablamowicz
7. Square Roots of โ1 in ํ(2๐ ,โ)
The lowest-dimensional example for ๐ = 1 is the Pauli matrix algebra๐ =โณ(2,โ)isomorphic to the geometric algebra ๐ถโ3,0 of the 3D Euclidean space and ๐ถโ1,2.The ๐ถโ3,0 vectors ๐1, ๐2, ๐3 correspond one-to-one to the Pauli matrices
๐1 =
(0 11 0
), ๐2 =
(0 โ๐๐ 0
), ๐3 =
(1 00 โ1
), (7.1)
with ๐1๐2 = ๐๐3 =
(๐ 00 โ๐
). The element ๐ = ๐1๐2๐3 = ๐1 represents the
central pseudoscalar ๐123 of ๐ถโ3,0 with square ๐2 = โ1. The Pauli algebra has thefollowing idempotents
๐1 = ๐21 = 1, ๐0 = (1/2)(1+ ๐3), ๐โ1 = 0 . (7.2)
The idempotents correspond via
๐ = ๐(2๐โ 1), (7.3)
to the square roots of โ1:๐1 = ๐1 =
(๐ 00 ๐
), ๐0 = ๐๐3 =
(๐ 00 โ๐
), ๐โ1 = โ๐1 =
(โ๐ 00 โ๐
), (7.4)
where by complex conjugation ๐โ1 = ๐1. Let the idempotent ๐โฒ0 = 12 (1โ๐3) corre-
spond to the matrix ๐ โฒ0 = โ๐๐3. We observe that ๐0 is conjugate to ๐ โฒ0 = ๐โ11 ๐0๐1 =
๐1๐2 = ๐0 using ๐โ11 = ๐1 but ๐1 is not conjugate to ๐โ1. Therefore, only ๐1, ๐0, ๐โ1
lead to three distinct conjugacy classes of square roots of โ1 inโณ(2,โ). CompareAppendix B for the corresponding computations with CLIFFORD for Maple.
In general, if ๐ = โณ(2๐,โ), then dim(๐) = 8๐2. The group G(๐) has oneconnected component. The square roots of โ1 in ๐ are in bijection with theidempotents ๐ [3] according to (7.3). According18 to (7.3) and its inverse ๐ =12 (1 โ ๐๐) the square root of โ1 with Spec(๐โ) = ๐/๐ = โ1, i.e., ๐ = โ๐ (seebelow), always corresponds to the trival idempotent ๐โ = 0, and the square rootof โ1 with Spec(๐+) = ๐/๐ = +1, ๐ = +๐, corresponds to the identity idempotent๐+ = 1.
If ๐ is a square root ofโ1, then ๐ = โ2๐ is the direct sum of the eigenspaces19
associated with the eigenvalues ๐ and โ๐. There is an integer ๐ such that thedimensions of the eigenspaces are respectively ๐ + ๐ and ๐ โ ๐. Moreover, โ๐ โค๐ โค ๐. Two square roots of โ1 are conjugate if and only if they give the same
18On the other hand it is clear that complex conjugation always leads to ๐โ = ๐+, wherethe overbar means complex conjugation in โณ(2๐,โ) and Clifford conjugation in the isomorphicClifford algebra ๐ถโ๐,๐. So either the trivial idempotent ๐โ = 0 is included in the bijection (7.3) ofidempotents and square roots of โ1, or alternatively the square root of โ1 with Spec(๐โ) = โ1is obtained from ๐โ = ๐+.19The following theorem is sufficient for a matrix ๐ in โณ(๐,๐), if ๐ is a (commutative) field.The matrix ๐ is diagonalizable if and only if ๐ (๐) = 0 for some polynomial ๐ that has only simpleroots, all of them in the field ๐. (This implies that ๐ is a multiple of the minimal polynomial,but we do not need to know whether ๐ is or is not the minimal polynomial.)
7. Square Roots of โ1 in Real Clifford Algebras 135
integer ๐. Then, all elements of Cent(๐) consist of diagonal block matrices with 2square blocks of (๐+๐)ร(๐+๐) matrices and (๐โ๐)ร(๐โ๐) matrices. Therefore,the โ-dimension of Cent(๐) is (๐+ ๐)2 + (๐โ ๐)2. Hence the โ-dimension (2.1) ofthe conjugacy class of ๐ :
8๐2 โ 2(๐+ ๐)2 โ 2(๐โ ๐)2 = 4(๐2 โ ๐2). (7.5)
Also, from the equality tr(๐) = (๐+๐)๐โ(๐โ๐)๐ = 2๐๐ we deduce that Scal(๐) = 0and that Spec(๐) = (2๐๐)/(2๐๐) = ๐/๐ if ๐ = ๐1 (whence tr(๐) = 2๐๐).
As announced on page 127, we consider that a square root of โ1 is ordinaryif the associated integer ๐ vanishes, and that it is exceptional if ๐ โ= 0. Thus thefollowing assertion is true in all cases: the ordinary square roots of โ1 in ๐ con-stitute one conjugacy class of dimension dim(๐)/2 which has as many connectedcomponents as G(๐), and the equality Spec(๐) = 0 holds for every ordinary squareroot of โ1 when the linear form Spec exists. All conjugacy classes of exceptionalsquare roots of โ1 have a dimension < dim(๐)/2.
All square roots of โ1 in โณ(2๐,โ) constitute (2๐ + 1) conjugacy classes20
which are also the connected components of the submanifold of square roots of โ1because of the equality Spec(๐) = ๐/๐, which is conjugacy class specific.
When ๐ = โณ(2๐,โ), the group Aut(๐) is larger than Inn(๐) since it con-tains the complex conjugation (that maps every entry of a matrix to the conjugatecomplex number). It is clear that the class of ordinary square roots of โ1 is invari-ant by complex conjugation. But the class associated with an integer ๐ other than0 is mapped by complex conjugation to the class associated with โ๐. In particularthe complex conjugation maps the class {๐} (associated with ๐ = ๐) to the class{โ๐} associated with ๐ = โ๐.
All these observations can easily verified for the above example of ๐ = 1 ofthe Pauli matrix algebra ๐ = โณ(2,โ). For ๐ = 2 we have the isomorphism of๐ = โณ(4,โ) with ๐ถโ0,5, ๐ถโ2,3 and ๐ถโ4,1. While ๐ถโ0,5 is important in Cliffordanalysis, ๐ถโ4,1 is both the geometric algebra of the Lorentz space โ4,1 and theconformal geometric algebra of 3D Euclidean geometry. Its set of square roots ofโ1 is therefore of particular practical interest.
Example. Let ๐ถโ4,1 โผ= ๐ where ๐ = โณ(4,โ) for ๐ = 2. The ๐ถโ4,1 1-vectors canbe represented21 by the following matrices:
๐1 =
โโโโ1 0 0 00 โ1 0 00 0 โ1 00 0 0 1
โโโโ , ๐2 =
โโโโ0 1 0 01 0 0 00 0 0 10 0 1 0
โโโโ , ๐3 =
โโโโ0 โ๐ 0 0๐ 0 0 00 0 0 โ๐0 0 ๐ 0
โโโโ ,
20Two conjugate (similar) matrices have the same eigenvalues and the same trace. This sufficesto recognize that 2๐+ 1 conjugacy classes are obtained.21For the computations of this example in the Maple package CLIFFORD we have used theidentification ๐ = ๐23. Yet the results obtained for the square roots of โ1 are independent of thissetting (we can alternatively use, e.g., ๐ = ๐12345 , or the imaginary unit ๐ โ โ), as can easilybe checked for ๐1 of (7.7), ๐0 of (7.8) and ๐โ1 of (7.9) by only assuming the standard Cliffordproduct rules for ๐1 to ๐5.
136 E. Hitzer, J. Helmstetter and R. Ablamowicz
๐4 =
โโโโ0 0 1 00 0 0 โ11 0 0 00 โ1 0 0
โโโโ , ๐5 =
โโโโ0 0 โ1 00 0 0 11 0 0 00 โ1 0 0
โโโโ . (7.6)
We find five conjugacy classes of roots ๐๐ of โ1 in ๐ถโ4,1 for ๐ โ {0,ยฑ1,ยฑ2}: fourexceptional and one ordinary. Since ๐๐ is a root of ๐(๐ก) = ๐ก2 +1 which factors overโ into (๐ก โ ๐)(๐ก + ๐), the minimal polynomial ๐๐(๐ก) of ๐๐ is one of the following:๐ก โ ๐, ๐ก+ ๐, or (๐ก โ ๐)(๐ก + ๐). Respectively, there are three classes of characteristicpolynomial ฮ๐(๐ก) of the matrix โฑ๐ in โณ(4,โ) which corresponds to ๐๐, namely,(๐กโ๐)4, (๐ก+๐)4, and (๐กโ๐)๐1(๐ก+๐)๐2 , where ๐1+๐2 = 2๐ = 4 and ๐1 = ๐+๐ = 2+๐,๐2 = ๐ โ ๐ = 2 โ ๐. As predicted by the above discussion, the ordinary rootcorresponds to ๐ = 0 whereas the exceptional roots correspond to ๐ โ= 0.
1. For ๐ = 2, we have ฮ2(๐ก) = (๐กโ ๐)4, ๐2(๐ก) = ๐กโ ๐, and so โฑ2 = diag(๐, ๐, ๐, ๐)which in the above representation (7.6) corresponds to the non-trivial centralelement ๐2 = ๐ = ๐12345. Clearly, Spec(๐2) = 1 = ๐
๐ ; Scal(๐2) = 0; theโ-dimension of the centralizer Cent(๐2) is 16; and the โ-dimension of theconjugacy class of ๐2 is zero as it contains only ๐2 since ๐2 โ Z(๐). Thus, theโ-dimension of the class is again zero in agreement with (7.5).
2. For ๐ = โ2, we have ฮโ2(๐ก) = (๐ก + ๐)4, ๐โ2(๐ก) = ๐ก + ๐, and โฑโ2 =diag(โ๐,โ๐,โ๐,โ๐) which corresponds to the central element ๐โ2 = โ๐ =โ๐12345. Again, Spec(๐โ2) = โ1 = ๐
๐ ; Scal(๐โ2) = 0; the โ-dimension ofthe centralizer Cent(๐โ2) is 16 and the conjugacy class of ๐โ2 contains only๐โ2 since ๐โ2 โ Z(๐). Thus, the โ-dimension of the class is again zero inagreement with (7.5).
3. For ๐ โ= ยฑ2, we consider three subcases when ๐ = 1, ๐ = 0, and ๐ = โ1.When ๐ = 1, then ฮ1(๐ก) = (๐กโ ๐)3(๐ก+ ๐) and ๐1(๐ก) = (๐กโ ๐)(๐ก+ ๐). Then theroot โฑ1 = diag(๐, ๐, ๐,โ๐) corresponds to
๐1 =1
2(๐23 + ๐123 โ ๐2345 + ๐12345). (7.7)
Note that Spec(๐1) =12 = ๐
๐ so ๐1 is an exceptional root of โ1.When ๐ = 0, then ฮ0(๐ก) = (๐กโ ๐)2(๐ก+ ๐)2 and ๐0(๐ก) = (๐กโ ๐)(๐ก+ ๐). Thus theroot of โ1 in this case is โฑ0 = diag(๐, ๐,โ๐,โ๐) which corresponds to just
๐0 = ๐123. (7.8)
Note that Spec(๐0) = 0 thus ๐0 = ๐123 is an ordinary root of โ1.When ๐ = โ1, then ฮโ1(๐ก) = (๐กโ ๐)(๐ก+ ๐)3 and ๐โ1(๐ก) = (๐กโ ๐)(๐ก+ ๐). Then,the root of โ1 in this case is โฑโ1 = diag(๐,โ๐,โ๐,โ๐) which corresponds to
๐โ1 =1
2(๐23 + ๐123 + ๐2345 โ ๐12345). (7.9)
Since Scal(๐โ1) = โ 12 = ๐
๐ , we gather that ๐โ1 is an exceptional root.
7. Square Roots of โ1 in Real Clifford Algebras 137
As expected, we can also see that the roots ๐ and โ๐ are related viathe grade involution whereas ๐1 = โ๐โ1 where ห denotes the reversion in๐ถโ4,1.
Example. Let ๐ถโ0,5 โผ= ๐ where ๐ = โณ(4,โ) for ๐ = 2. The ๐ถโ0,5 1-vectors canbe represented22 by the following matrices:
๐1 =
โโโโ0 โ1 0 01 0 0 00 0 0 โ10 0 1 0
โโโโ , ๐2 =
โโโโ0 โ๐ 0 0โ๐ 0 0 00 0 0 โ๐0 0 โ๐ 0
โโโโ , ๐3 =
โโโโโ๐ 0 0 00 ๐ 0 00 0 ๐ 00 0 0 โ๐
โโโโ ,
๐4 =
โโโโ0 0 โ1 00 0 0 11 0 0 00 โ1 0 0
โโโโ , ๐5 =
โโโโ0 0 โ๐ 00 0 0 ๐โ๐ 0 0 00 ๐ 0 0
โโโโ , (7.10)
Like for ๐ถโ4,1, we have five conjugacy classes of the roots ๐๐ of โ1 in ๐ถโ0,5 for๐ โ {0,ยฑ1,ยฑ2}: four exceptional and one ordinary. Using the same notation as inExample 7, we find the following representatives of the conjugacy classes.
1. For ๐ = 2, we have ฮ2(๐ก) = (๐ก โ ๐)4, ๐2(๐ก) = ๐ก โ ๐, and โฑ2 = diag(๐, ๐, ๐, ๐)which in the above representation (7.10) corresponds to the non-trivial cen-tral element ๐2 = ๐ = ๐12345. Then, Spec(๐2) = 1 = ๐
๐ ; Scal(๐2) = 0; theโ-dimension of the centralizer Cent(๐2) is 16; and the โ-dimension of theconjugacy class of ๐2 is zero as it contains only ๐2 since ๐2 โ Z(๐). Thus, theโ-dimension of the class is again zero in agreement with (7.5).
2. For ๐ = โ2, we have ฮโ2(๐ก) = (๐ก + ๐)4, ๐โ2(๐ก) = ๐ก + ๐, and โฑโ2 =diag(โ๐,โ๐,โ๐,โ๐) which corresponds to the central element ๐โ2= โ ๐ =โ๐12345. Again, Spec(๐โ2) = โ1 = ๐
๐ ; Scal(๐โ2) = 0; the โ-dimension ofthe centralizer Cent(๐โ2) is 16 and the conjugacy class of ๐โ2 contains only๐โ2 since ๐โ2 โ Z(๐). Thus, the โ-dimension of the class is again zero inagreement with (7.5).
3. For ๐ โ= ยฑ2, we consider three subcases when ๐ = 1, ๐ = 0, and ๐ = โ1.When ๐ = 1, then ฮ1(๐ก) = (๐กโ ๐)3(๐ก+ ๐) and ๐1(๐ก) = (๐กโ ๐)(๐ก+ ๐). Then theroot โฑ1 = diag(๐, ๐, ๐,โ๐) corresponds to
๐1 =1
2(๐3 + ๐12 + ๐45 + ๐12345). (7.11)
Since Spec(๐1) =12 = ๐
๐ , ๐1 is an exceptional root of โ1.
22For the computations of this example in the Maple package CLIFFORD we have used theidentification ๐ = ๐3. Yet the results obtained for the square roots of โ1 are independent of thissetting (we can alternatively use, e.g., ๐ = ๐12345, or the imaginary unit ๐ โ โ), as can easily be
checked for ๐1 of (7.11), ๐0 of (7.12) and ๐โ1 of (7.13) by only assuming the standard Cliffordproduct rules for ๐1 to ๐5.
138 E. Hitzer, J. Helmstetter and R. Ablamowicz
When ๐ = 0, then ฮ0(๐ก) = (๐ก โ ๐)2(๐ก + ๐)2 and ๐0(๐ก) = (๐ก โ ๐)(๐ก + ๐). Thusthe root of โ1 is this case is โฑ0 = diag(๐, ๐,โ๐,โ๐) which corresponds to just
๐0 = ๐45. (7.12)
Note that Spec(๐0) = 0 thus ๐0 = ๐45 is an ordinary root of โ1.When ๐ = โ1, then ฮโ1(๐ก) = (๐กโ ๐)(๐ก+ ๐)3 and ๐โ1(๐ก) = (๐กโ ๐)(๐ก+ ๐). Then,the root of โ1 in this case is โฑโ1 = diag(๐,โ๐,โ๐,โ๐) which corresponds to
๐โ1 =1
2(โ๐3 + ๐12 + ๐45 โ ๐12345). (7.13)
Since Scal(๐โ1) = โ 12 = ๐
๐ , we gather that ๐โ1 is an exceptional root.Again we can see that the roots ๐2 and ๐โ2 are related via the grade
involution whereas ๐1 = โ๐โ1 where ห denotes the reversion in ๐ถโ0,5.
Example. Let ๐ถโ7,0 โผ= ๐ where ๐ = โณ(8,โ) for ๐ = 4. We have nine conjugacyclasses of roots ๐๐ of โ1 for ๐ โ {0,ยฑ1,ยฑ2 ยฑ 3 ยฑ 4}. Since ๐๐ is a root of apolynomial ๐(๐ก) = ๐ก2 + 1 which factors over โ into (๐ก โ ๐)(๐ก + ๐), its minimalpolynomial ๐(๐ก) will be one of the following: ๐กโ ๐, ๐ก+ ๐, or (๐กโ ๐)(๐ก+ ๐) = ๐ก2 + 1.
Respectively, each conjugacy class is characterized by a characteristic poly-nomial ฮ๐(๐ก) of the matrix ๐๐ โ โณ(8,โ) which represents ๐๐. Namely, we have
ฮ๐(๐ก) = (๐กโ ๐)๐1(๐ก+ ๐)๐2 ,
where ๐1 + ๐2 = 2๐ = 8 and ๐1 = ๐ + ๐ = 4 + ๐ and ๐2 = ๐ โ ๐ = 4 โ ๐. Theordinary root of โ1 corresponds to ๐ = 0 whereas the exceptional roots correspondto ๐ โ= 0.
1. When ๐ = 4, we have ฮ4(๐ก) = (๐กโ ๐)8, ๐4(๐ก) = ๐กโ ๐, and โฑ4 = diag(
8๏ธท ๏ธธ๏ธธ ๏ธท๐, . . . , ๐)
which in the representation used by CLIFFORD [2] corresponds to the non-trivial central element ๐4 = ๐ = ๐1234567. Clearly, Spec(๐4) = 1 = ๐
๐ ;Scal(๐4) = 0; the โ-dimension of the centralizer Cent(๐4) is 64; and theโ-dimension of the conjugacy class of ๐4 is zero since ๐4 โ Z(๐). Thus, theโ-dimension of the class is again zero in agreement with (7.5).
2. When ๐ = โ4, we have ฮโ4(๐ก) = (๐ก + ๐)8, ๐โ4(๐ก) = ๐ก + ๐, and โฑโ4 =
diag(
8๏ธท ๏ธธ๏ธธ ๏ธทโ๐, . . . ,โ๐) which corresponds to ๐โ4 = โ๐ = โ๐1234567. Again,
Spec(๐โ4) = โ1 =๐
๐; Scal(๐โ4) = 0;
the โ-dimension of the centralizer Cent(๐) is 64 and the conjugacy class of๐โ4 contains only ๐โ4 since ๐โ4 โ Z(๐). Thus, the โ-dimension of the classis again zero in agreement with (7.5).
3. When ๐ โ= ยฑ4, we consider seven subcases when ๐ = ยฑ3, ๐ = ยฑ2, ๐ = ยฑ1,and ๐ = 0.
7. Square Roots of โ1 in Real Clifford Algebras 139
When ๐ = 3, then ฮ3(๐ก) = (๐กโ ๐)7(๐ก+ ๐) and ๐3(๐ก) = (๐กโ ๐)(๐ก+ ๐). Then the
root โฑ3 = diag(
7๏ธท ๏ธธ๏ธธ ๏ธท๐, . . . , ๐,โ๐) corresponds to
๐3 =1
4(๐23 โ ๐45 + ๐67 โ ๐123 + ๐145 โ ๐167 + ๐234567 + 3๐1234567). (7.14)
Since Spec(๐3) =34 = ๐
๐ , ๐3 is an exceptional root of โ1.When ๐ = 2, then ฮ2(๐ก) = (๐ก โ ๐)6(๐ก + ๐)2 and ๐2(๐ก) = (๐ก โ ๐)(๐ก + ๐). Then
the root โฑ2 = diag(
6๏ธท ๏ธธ๏ธธ ๏ธท๐, . . . , ๐,โ๐,โ๐) corresponds to
๐2 =1
2(๐67 โ ๐45 โ ๐123 + ๐1234567). (7.15)
Since Spec(๐2) =12 = ๐
๐ , ๐2 is also an exceptional root.
When ๐ = 1, then ฮ1(๐ก) = (๐ก โ ๐)5(๐ก + ๐)3 and ๐1(๐ก) = (๐ก โ ๐)(๐ก + ๐). Then
the root โฑ1 = diag(
5๏ธท ๏ธธ๏ธธ ๏ธท๐, . . . , ๐,โ๐,โ๐,โ๐) corresponds to
๐1 =1
4(๐23 โ ๐45 + 3๐67 โ ๐123 + ๐145 + ๐167 โ ๐234567 + ๐1234567). (7.16)
Since Spec(๐1) =14 = ๐
๐ , ๐1 is another exceptional root.
When ๐ = 0, then ฮ0(๐ก) = (๐ก โ ๐)4(๐ก + ๐)4 and ๐0(๐ก) = (๐ก โ ๐)(๐ก + ๐). Thenthe root โฑ0 = diag(๐, ๐, ๐, ๐,โ๐,โ๐,โ๐,โ๐) corresponds to
๐0 =1
2(๐23 โ ๐45 + ๐67 โ ๐234567). (7.17)
Since Spec(๐0) = 0 = ๐๐ , we see that ๐0 is an ordinary root of โ1.
When ๐ = โ1, then ฮโ1(๐ก) = (๐ก โ ๐)3(๐ก + ๐)5 and ๐โ1(๐ก) = (๐ก โ ๐)(๐ก + ๐).
Then the root โฑโ1 = diag(๐, ๐, ๐,
5๏ธท ๏ธธ๏ธธ ๏ธทโ๐, . . . ,โ๐) corresponds to
๐โ1 =1
4(๐23 โ ๐45 + 3๐67 + ๐123 โ ๐145 โ ๐167 โ ๐234567 โ ๐1234567). (7.18)
Thus, Spec(๐โ1) = โ 14 = ๐
๐ and so ๐โ1 is another exceptional root.
When ๐ = โ2, then ฮโ2(๐ก) = (๐ก โ ๐)2(๐ก + ๐)6 and ๐โ2(๐ก) = (๐ก โ ๐)(๐ก + ๐).
Then the root โฑโ2 = diag(๐, ๐,
6๏ธท ๏ธธ๏ธธ ๏ธทโ๐, . . . ,โ๐) corresponds to
๐โ2 =1
2(๐67 โ ๐45 + ๐123 โ ๐1234567). (7.19)
Since Spec(๐โ2) = โ 12 = ๐
๐ , we see that ๐โ2 is also an exceptional root.
When ๐ = โ3, then ฮโ3(๐ก) = (๐กโ ๐)(๐ก+ ๐)7 and ๐โ3(๐ก) = (๐กโ ๐)(๐ก+ ๐). Then
the root โฑโ3 = diag(๐,
7๏ธท ๏ธธ๏ธธ ๏ธทโ๐, . . . ,โ๐) corresponds to
๐โ3 =1
4(๐23 โ ๐45 + ๐67 + ๐123 โ ๐145 + ๐167 + ๐234567 โ 3๐1234567). (7.20)
140 E. Hitzer, J. Helmstetter and R. Ablamowicz
Again, Spec(๐โ3) = โ 34 = ๐
๐ and so ๐โ3 is another exceptional root of โ1.As expected, we can also see that the roots ๐ and โ๐ are related via
the reversion whereas ๐3 = โ๐โ3, ๐2 = โ๐โ2, ๐1 = โ๐โ1 where ยฏ denotesthe conjugation in ๐ถโ7,0.
8. Conclusions
We proved that in all cases Scal(๐) = 0 for every square root of โ1 in ๐ isomorphicto ๐ถโ๐,๐. We distinguished ordinary square roots of โ1, and exceptional ones.
In all cases the ordinary square roots ๐ of โ1 constitute a unique conjugacyclass of dimension dim(๐)/2 which has as many connected components as thegroup G(๐) of invertible elements in ๐. Furthermore, we have Spec(๐) = 0 (zeropseudoscalar part) if the associated ring is โ2, โ2, or โ. The exceptional squareroots of โ1 only exist if ๐ โผ=โณ(2๐,โ) (see Section 7).
For ๐ =โณ(2๐,โ) of Section 3, the centralizer and the conjugacy class of asquare root ๐ of โ1 both have โ-dimension 2๐2 with two connected components,pictured in Figure 2 for ๐ = 1.
For ๐ =โณ(2๐,โ2) =โณ(2๐,โ)รโณ(2๐,โ) of Section 4, the square roots of(โ1,โ1) are pairs of two square roots of โ1 inโณ(2๐,โ). They constitute a uniqueconjugacy class with four connected components, each of dimension 4๐2. Regardingthe four connected components, the group Inn(๐) induces the permutations of theKlein group whereas the quotient group Aut(๐)/Inn(๐) is isomorphic to the groupof isometries of a Euclidean square in 2D.
For ๐ =โณ(๐,โ) of Section 5, the submanifold of the square roots ๐ of โ1is a single connected conjugacy class of โ-dimension 2๐2 equal to the โ-dimensionof the centralizer of every ๐ . The easiest example is โ itself for ๐ = 1.
For ๐ =โณ(๐,โ2) =โณ(2๐,โ)รโณ(2๐,โ) of Section 6, the square roots of(โ1,โ1) are pairs of two square roots (๐, ๐ โฒ) of โ1 in โณ(2๐,โ) and constitute aunique connected conjugacy class of โ-dimension 4๐2. The group Aut(๐) has twoconnected components: the neutral component Inn(๐) connected to the identityand the second component containing the swap automorphism (๐, ๐ โฒ) ๏ฟฝโ (๐ โฒ, ๐).The simplest case for ๐ = 1 is โ2 isomorphic to ๐ถโ0,3.
For ๐ = โณ(2๐,โ) of Section 7, the square roots of โ1 are in bijection tothe idempotents. First, the ordinary square roots of โ1 (with ๐ = 0) constitutea conjugacy class of โ-dimension 4๐2 of a single connected component which isinvariant under Aut(๐). Second, there are 2๐ conjugacy classes of exceptionalsquare roots of โ1, each composed of a single connected component, characterizedby equality Spec(๐) = ๐/๐ (the pseudoscalar coefficient) with ยฑ๐ โ {1, 2, . . . , ๐},and their โ-dimensions are 4(๐2 โ ๐2). The group Aut(๐) includes conjugationof the pseudoscalar ๐ ๏ฟฝโ โ๐ which maps the conjugacy class associated with ๐to the class associated with โ๐. The simplest case for ๐ = 1 is the Pauli matrixalgebra isomorphic to the geometric algebra ๐ถโ3,0 of 3D Euclidean space โ3, andto complex biquaternions [33].
7. Square Roots of โ1 in Real Clifford Algebras 141
Section 7 includes explicit examples for ๐ = 2: ๐ถโ4,1 and ๐ถโ0,5, and for ๐ = 4:๐ถโ7,0. Appendix A summarizes the square roots of โ1 in all ๐ถโ๐,๐ โผ= โณ(2๐,โ)for ๐ = 1, 2, 4. Appendix B contains details on how square roots of โ1 can becomputed using the package CLIFFORD for Maple.
Among the many possible applications of this research, the possibility of newintegral transformations in Clifford analysis is very promising. This field thus ob-tains essential algebraic information, which can, e.g., be used to create steerabletransformations, which may be steerable within a connected component of a sub-manifold of square roots of โ1.
Appendix A. Summary of Roots of โ1 in ๐ชโ๐,๐ โผ=ํ(2๐ ,โ)for ๐ = 1, 2, 4
In this appendix we summarize roots of โ1 for Clifford algebras ๐ถโ๐,๐ โผ=โณ(2๐,โ)for ๐ = 1, 2, 4. These roots have been computed with CLIFFORD [2]. Maple [25]worksheets written to derive these roots are posted at [21].
Table 1. Square roots of โ1 in ๐ถโ3,0 โผ=โณ(2,โ), ๐ = 1
๐ ๐๐ ฮ๐(๐ก)
1 ๐ = ๐123 (๐กโ ๐)2
0 ๐23 (๐กโ ๐)(๐ก+ ๐)
โ1 โ๐ = โ๐123 (๐ก+ ๐)2
Table 2. Square roots of โ1 in ๐ถโ4,1 โผ=โณ(4,โ), ๐ = 2
๐ ๐๐ ฮ๐(๐ก)
2 ๐ = ๐12345 (๐กโ ๐)4
1 12 (๐23 + ๐123 โ ๐2345 + ๐12345) (๐กโ ๐)3(๐ก+ ๐)
0 ๐123 (๐กโ ๐)2(๐ก+ ๐)2
โ1 12 (๐23 + ๐123 + ๐2345 โ ๐12345) (๐กโ ๐)(๐ก+ ๐)3
โ2 โ๐ = โ๐12345 (๐ก+ ๐)4
142 E. Hitzer, J. Helmstetter and R. Ablamowicz
Table 3. Square roots of โ1 in ๐ถโ0,5 โผ=โณ(4,โ), ๐ = 2
๐ ๐๐ ฮ๐(๐ก)
2 ๐ = ๐12345 (๐กโ ๐)4
1 12 (๐3 + ๐12 + ๐45 + ๐12345) (๐กโ ๐)3(๐ก+ ๐)
0 ๐45 (๐กโ ๐)2(๐ก+ ๐)2
โ1 12 (โ๐3 + ๐12 + ๐45 โ ๐12345) (๐กโ ๐)(๐ก+ ๐)3
โ2 โ๐ = โ๐12345 (๐ก+ ๐)4
Table 4. Square roots of โ1 in ๐ถโ2,3 โผ=โณ(4,โ), ๐ = 2
๐ ๐๐ ฮ๐(๐ก)
2 ๐ = ๐12345 (๐กโ ๐)4
1 12 (๐3 + ๐134 + ๐235 + ๐) (๐กโ ๐)3(๐ก+ ๐)
0 ๐134 (๐กโ ๐)2(๐ก+ ๐)2
โ1 12 (โ๐3 + ๐134 + ๐235 โ ๐) (๐กโ ๐)(๐ก+ ๐)3
โ2 โ๐ = โ๐12345 (๐ก+ ๐)4
Table 5. Square roots of โ1 in ๐ถโ7,0 โผ=โณ(8,โ), ๐ = 4
๐ ๐๐ ฮ๐(๐ก)
4 ๐ = ๐1234567 (๐กโ ๐)8
3 14 (๐23 โ ๐45 + ๐67 โ ๐123 + ๐145
โ ๐167 + ๐234567 +3๐)(๐กโ ๐)7(๐ก+ ๐)
2 12 (๐67 โ ๐45 โ ๐123 + ๐) (๐กโ ๐)6(๐ก+ ๐)2
1 14 (๐23 โ ๐45 + 3๐67 โ ๐123 + ๐145
+ ๐167 โ ๐234567 + ๐)(๐กโ ๐)5(๐ก+ ๐)3
0 12 (๐23 โ ๐45 + ๐67 โ ๐234567) (๐กโ ๐)4(๐ก+ ๐)4
โ1 14 (๐23 โ ๐45 + 3๐67 + ๐123 โ ๐145
โ ๐167 โ ๐234567 โ ๐)(๐กโ ๐)3(๐ก+ ๐)5
โ2 12 (๐67 โ ๐45 + ๐123 โ ๐) (๐กโ ๐)2(๐ก+ ๐)6
โ3 14 (๐23 โ ๐45 + ๐67 + ๐123 โ ๐145
+ ๐167 + ๐234567โ 3๐)(๐กโ ๐)(๐ก+ ๐)7
โ4 โ๐ = โ๐1234567 (๐ก+ ๐)8
7. Square Roots of โ1 in Real Clifford Algebras 143
Table 6. Square roots of โ1 in ๐ถโ1,6 โผ=โณ(8,โ), ๐ = 4
๐ ๐๐ ฮ๐(๐ก)
4 ๐ = ๐1234567 (๐กโ ๐)8
3 14 (๐4 โ ๐23 โ ๐56 + ๐1237 + ๐147
+ ๐1567โ ๐23456 +3๐)(๐กโ ๐)7(๐ก+ ๐)
2 12 (โ๐23 โ ๐56 + ๐147 + ๐) (๐กโ ๐)6(๐ก+ ๐)2
1 14 (โ๐4 โ ๐23 โ 3๐56 โ ๐1237 + ๐147
+ ๐1567 โ ๐23456 + ๐)(๐กโ ๐)5(๐ก+ ๐)3
0 12 (๐4 + ๐23 + ๐56 + ๐23456) (๐กโ ๐)4(๐ก+ ๐)4
โ1 14 (โ๐4 โ ๐23 โ 3๐56 + ๐1237 โ ๐147
โ ๐1567 โ ๐23456 โ ๐)(๐กโ ๐)3(๐ก+ ๐)5
โ2 12 (โ๐23 โ ๐56 โ ๐147 โ ๐) (๐กโ ๐)2(๐ก+ ๐)6
โ3 14 (๐4 โ ๐23 โ ๐56 โ ๐1237 โ ๐147
โ ๐1567โ ๐23456โ 3๐)(๐กโ ๐)(๐ก+ ๐)7
โ4 โ๐ = โ๐1234567 (๐ก+ ๐)8
Table 7. Square roots of โ1 in ๐ถโ3,4 โผ=โณ(8,โ), ๐ = 4
๐ ๐๐ ฮ๐(๐ก)
4 ๐ = ๐1234567 (๐กโ ๐)8
3 14 (๐4 + ๐145 + ๐246 + ๐347 โ ๐12456
โ๐13457โ๐23467+3๐)(๐กโ ๐)7(๐ก+ ๐)
2 12 (๐145 โ ๐12456 โ ๐13457 + ๐) (๐กโ ๐)6(๐ก+ ๐)2
1 14 (โ๐4 + ๐145 + ๐246 โ ๐347 โ 3๐12456
โ ๐13457โ ๐23467 +๐)(๐กโ ๐)5(๐ก+ ๐)3
0 12 (๐4 + ๐12456 + ๐13457 + ๐23467) (๐กโ ๐)4(๐ก+ ๐)4
โ1 14 (โ๐4 โ ๐145 โ ๐246 + ๐347 โ 3๐12456
โ ๐13457โ ๐23467โ๐)(๐กโ ๐)3(๐ก+ ๐)5
โ2 12 (โ๐145 โ ๐12456 โ ๐13457 โ ๐) (๐กโ ๐)2(๐ก+ ๐)6
โ3 14 (๐4 โ ๐145 โ ๐246 โ ๐347 โ ๐12456
โ๐13457โ๐23467โ3๐)(๐กโ ๐)(๐ก+ ๐)7
โ4 โ๐ = โ๐1234567 (๐ก+ ๐)8
144 E. Hitzer, J. Helmstetter and R. Ablamowicz
Table 8. Square roots of โ1 in ๐ถโ5,2 โผ=โณ(8,โ), ๐ = 4
๐ ๐๐ ฮ๐(๐ก)
4 ๐ = ๐1234567 (๐กโ ๐)8
3 14 (โ๐23 + ๐123 + ๐2346 + ๐2357 โ ๐12346
โ๐12357+๐234567+3๐)(๐กโ ๐)7(๐ก+ ๐)
2 12 (๐123 โ ๐12346 โ ๐12357 + ๐) (๐กโ ๐)6(๐ก+ ๐)2
1 14 (โ๐23 + ๐123 โ ๐2346 + ๐2357 โ 3๐12346
โ๐12357โ๐234567+๐)(๐กโ ๐)5(๐ก+ ๐)3
0 12 (๐23 + ๐12346 + ๐12357 + ๐234567) (๐กโ ๐)4(๐ก+ ๐)4
โ1 14 (โ๐23 โ ๐123 + ๐2346 โ ๐2357 โ 3๐12346
โ๐12357โ๐234567โ๐)(๐กโ ๐)3(๐ก+ ๐)5
โ2 12 (โ๐123 โ ๐12346 โ ๐12357 โ ๐) (๐กโ ๐)2(๐ก+ ๐)6
โ3 14 (โ๐23 โ ๐123 โ ๐2346 โ ๐2357 โ ๐12346
โ๐12357+๐234567โ3๐)(๐กโ ๐)(๐ก+ ๐)7
โ4 โ๐ = โ๐1234567 (๐ก+ ๐)8
Appendix B. A Sample Maple Worksheet
In this appendix we show a computation of roots of โ1 in ๐ถโ3,0 in CLIFFORD.Although these computations certainly can be performed by hand, as shown inSection 7, they illustrate how CLIFFORD can be used instead especially whenextending these computations to higher dimensions.23 To see the actual Mapleworksheets where these computations have been performed, see [21].> restart:with(Clifford):with(linalg):with(asvd):> p,q:=3,0; ##<<-- selecting signature> B:=diag(1$p,-1$q): ##<<-- defining diagonal bilinear form> eval(makealiases(p+q)): ##<<-- defining aliases> clibas:=cbasis(p+q); ##assigning basis for Cl(3,0)
๐, ๐ := 3, 0
clibas := [Id , e1 , e2 , e3 , e12 , e13 , e23 , e123 ]
> data:=clidata(); ##<<-- displaying information about Cl(3,0)
data := [complex , 2, simple,Id
2+e1
2, [Id , e2 , e3 , e23 ], [Id , e23 ], [Id , e2 ]]
> MM:=matKrepr(); ##<<-- displaying default matrices to generators
Cliplus has been loaded. Definitions for type/climon andtype/clipolynom now include &C and &C[K]. Type ?cliprod forhelp.
23In showing Maple display we have edited Maple output to save space. Package asvd is a
supplementary package written by the third author and built into CLIFFORD. The primarypurpose of asvd is to compute Singular Value Decomposition in Clifford algebras [1].
7. Square Roots of โ1 in Real Clifford Algebras 145
MM := [e1 =
โกโฃ 1 0
0 โ1
โคโฆ , e2 =
โกโฃ 0 1
1 0
โคโฆ , e3 =
โกโฃ 0 โe23e23 0
โคโฆ]
Pauli algebra representation displayed in (7.1):> sigma[1]:=evalm(rhs(MM[1]));> sigma[2]:=evalm(rhs(MM[2]));> sigma[3]:=evalm(rhs(MM[3]));
๐1, ๐2, ๐3 :=
โกโฃ 0 1
1 0
โคโฆ ,
โกโฃ 0 โe23e23 0
โคโฆ ,
โกโฃ 1 0
0 โ1
โคโฆ
We show how we represent the imaginary unit ๐ in the field โ and the diagonalmatrix diag(๐, ๐) :> ii:=e23; ##<<-- complex imaginary unit> II:=diag(ii,ii); ##<<-- diagonal matrix diag(i,i)
ii := e23
II :=
โกโฃ e23 0
0 e23
โคโฆ
We compute matrices ๐1,๐2, . . . ,๐8 representing each basis element in ๐ถโ3,0isomorphic with โ(2). Note that in our representation element ๐23 in ๐ถโ3,0 is usedto represent the imaginary unit ๐.> for i from 1 to nops(clibas) do
lprint(โThe basis elementโ,clibas[i],โis represented by the followingmatrix:โ);M[i]:=subs(Id=1,matKrepr(clibas[i])) od;
โThe basis elementโ, Id, โis represented by the following matrix:โ
๐1 :=
โกโฃ 1 0
0 1
โคโฆ
โThe basis elementโ, e1, โis represented by the following matrix:โ
๐2 :=
โกโฃ 1 0
0 โ1
โคโฆ
โThe basis elementโ, e2, โis represented by the following matrix:โ
๐3 :=
โกโฃ 0 1
1 0
โคโฆ
โThe basis elementโ, e3, โis represented by the following matrix:โ
๐4 :=
โกโฃ 0 โe23e23 0
โคโฆ
146 E. Hitzer, J. Helmstetter and R. Ablamowicz
โThe basis elementโ, e12, โis represented by the following matrix:โ
๐5 :=
โกโฃ 0 1
โ1 0
โคโฆ
โThe basis elementโ, e13, โis represented by the following matrix:โ
๐6 :=
โกโฃ 0 โe23
โe23 0
โคโฆ
โThe basis elementโ, e23, โis represented by the following matrix:โ
๐7 :=
โกโฃ e23 0
0 โe23
โคโฆ
โThe basis elementโ, e123, โis represented by the following matrix:โ
๐8 :=
โกโฃ e23 0
0 e23
โคโฆ
We will use the procedure phi from the asvd package which gives an isomor-phism from โ(2) to ๐ถโ3,0. This way we can find the image in ๐ถโ3,0 of any complex2 ร 2 complex matrix ๐ด. Knowing the image of each matrix ๐1,๐2, . . . ,๐8 interms of the Clifford polynomials in ๐ถโ3,0, we can easily find the image of ๐ด inour default spinor representation of ๐ถโ3,0 which is built into CLIFFORD.
Procedure Centralizer computes a centralizer of ๐ with respect to the Clif-ford basis ๐ฟ:> Centralizer:=proc(f,L) local c,LL,m,vars,i,eq,sol;
m:=add(c[i]*L[i],i=1..nops(L));vars:=[seq(c[i],i=1..nops(L))];eq:=clicollect(cmul(f,m)-cmul(m,f));if eq=0 then return L end if:sol:=op(clisolve(eq,vars));m:=subs(sol,m);m:=collect(m,vars);return sort([coeffs(m,vars)],bygrade);end proc:
Procedures Scal and Spec compute the scalar and the pseudoscalar parts of ๐ .> Scal:=proc(f) local p: return scalarpart(f); end proc:> Spec:=proc(f) local N; global p,q;
N:=p+q:return coeff(vectorpart(f,N),op(cbasis(N,N)));end proc:
The matrix idempotents in โ(2) displayed in (7.2) are as follows:> d:=1:Eps[1]:=sigma[1] &cm sigma[1];
7. Square Roots of โ1 in Real Clifford Algebras 147
> Eps[0]:=evalm(1/2*(1+sigma[3]));> Eps[-1]:=diag(0,0);
Eps1, Eps0, Epsโ1 :=
โกโฃ 1 0
0 1
โคโฆ ,
โกโฃ 1 0
0 0
โคโฆ ,
โกโฃ 0 0
0 0
โคโฆ
This function ff computes matrix square root of โ1 corresponding to the matrixidempotent ๐๐๐ :> ff:=eps->evalm(II &cm (2*eps-1));
ff := eps โ evalm(II &cm (2 eps โ 1))
We compute matrix square roots of โ1 which correspond to the idempotents๐ธ๐๐ 1, ๐ธ๐๐ 0, ๐ธ๐๐ โ1, and their characteristic and minimal polynomials. Note thatin Maple the default imaginary unit is denoted by ๐ผ.> F[1]:=ff(Eps[1]); ##<<-- this square root of -1 corresponds to Eps[1]
Delta[1]:=charpoly(subs(e23=I,evalm(F[1])),t);Mu[1]:=minpoly(subs(e23=I,evalm(F[1])),t);
๐น1 :=
โกโฃ e23 0
0 e23
โคโฆ , ฮ1 := (๐กโ ๐ผ)2, ๐1 := ๐กโ ๐ผ
> F[0]:=ff(Eps[0]); ##<<-- this square root of -1 corresponds to Eps[0]Delta[0]:=charpoly(subs(e23=I,evalm(F[0])),t);Mu[0]:=minpoly(subs(e23=I,evalm(F[0])),t);
๐น0 :=
โกโฃ e23 0
0 โe23
โคโฆ , ฮ0 := (๐กโ ๐ผ) (๐ก+ ๐ผ), ๐0 := 1 + ๐ก2
> F[-1]:=ff(Eps[-1]); ##<<-- this square root of -1 corresponds to Eps[-1]Delta[-1]:=charpoly(subs(e23=I,evalm(F[-1])),t);Mu[-1]:=minpoly(subs(e23=I,evalm(F[-1])),t);
๐นโ1 :=
โกโฃ โe23 0
0 โe23
โคโฆ , ฮโ1 := (๐ก+ ๐ผ)2, ๐โ1 := ๐ก+ ๐ผ
Now, we can find square roots of โ1 in ๐ถโ3,0 which correspond to the matrixsquare roots ๐นโ1, ๐น0, ๐น1 via the isomorphism ๐ : ๐ถโ3,0 โ โ(2) realized with theprocedure phi.
First, we let reprI denote element in ๐ถโ3,0 which represents the diagonal(2๐) ร (2๐) with ๐ผ = ๐ on the diagonal where ๐2 = โ1. This element will replacethe imaginary unit ๐ผ in the minimal polynomials.> reprI:=phi(diag(I$(2*d)),M);
reprI := e123
148 E. Hitzer, J. Helmstetter and R. Ablamowicz
Now, we compute the corresponding square roots ๐1, ๐0, ๐โ1 in ๐ถโ3,0.> f[1]:=phi(F[1],M); ##<<-- element in Cl(3,0) corresponding to F[1]
cmul(f[1],f[1]); ##<<-- checking that this element is a root of -1Mu[1]; ##<<-- recalling minpoly of matrix F[1]subs(e23=I,evalm(subs(t=evalm(F[1]),Mu[1]))); ##<<-- F[1] in Mu[1]mu[1]:=subs(I=reprI,Mu[1]); ##<<-- defining minpoly of f[1]cmul(f[1]-reprI,Id); ##<<-- verifying that f[1] satisfies mu[1]
๐1 := e123
โId , ๐กโ ๐ผ,
โกโฃ 0 0
0 0
โคโฆ
๐1 := ๐กโ e123 , 0
> f[0]:=phi(F[0],M); ##<<-- element in Cl(3,0) corresponding to F[0]cmul(f[0],f[0]); ##<<-- checking that this element is a root of -1Mu[0]; ##<<-- recalling minpoly of matrix F[0]subs(e23=I,evalm(subs(t=evalm(F[0]),Mu[0]))); ##<<-- F[0] in Mu[0]mu[0]:=subs(I=reprI,Mu[0]); ##<<-- defining minpoly of f[0]cmul(f[0]-reprI,f[0]+reprI); ##<<-- f[0] satisfies mu[0]
๐0 := e23
โId , 1 + ๐ก2,
โกโฃ 0 0
0 0
โคโฆ
๐0 := 1 + ๐ก2, 0
> f[-1]:=phi(F[-1],M); ##<<-- element in Cl(3,0) corresponding to F[-1]cmul(f[-1],f[-1]); ##<<-- checking that this element is a root of -1Mu[-1]; ##<<-- recalling minpoly of matrix F[-1]subs(e23=I,evalm(subs(t=evalm(F[-1]),Mu[-1]))); ##<<-- F[-1] in Mu[-1]mu[-1]:=subs(I=reprI,Mu[-1]); ##<<-- defining minpoly of f[-1]cmul(f[-1]+reprI,Id); ##<<-- f[-1] satisfies mu[-1]
๐โ1 := โe123
โId , ๐ก+ ๐ผ,
โกโฃ 0 0
0 0
โคโฆ
๐โ1 := ๐ก+ e123 , 0
Functions RdimCentralizer and RdimConjugClass of ๐ and ๐ compute thereal dimension of the centralizer Cent(๐) and the conjugacy class of ๐ (see (7.4)).> RdimCentralizer:=(d,k)->2*((d+k)ห2+(d-k)ห2); ##<<-- from the theory> RdimConjugClass:=(d,k)->4*(dห2-kห2); ##<<-- from the theory
RdimCentralizer := (๐, ๐) โ 2 (๐+ ๐)2 + 2 (๐โ ๐)2
RdimConjugClass := (๐, ๐) โ 4 ๐2 โ 4 ๐2
7. Square Roots of โ1 in Real Clifford Algebras 149
Now, we compute the centralizers of the roots and use notation ๐, ๐, ๐1, ๐2 dis-played in Examples.Case ๐ = 1 :> d:=1:k:=1:n1:=d+k;n2:=d-k;
A1:=diag(I$n1,-I$n2); ##<<-- this is the first matrix root of -1
n1 := 2, n2 := 0, A1 :=
โกโฃ ๐ผ 0
0 ๐ผ
โคโฆ
> f[1]:=phi(A1,M); cmul(f[1],f[1]); Scal(f[1]), Spec(f[1]);
๐1 := e123 , โId , 0, 1
> LL1:=Centralizer(f[1],clibas); ##<<-- centralizer of f[1]dimCentralizer:=nops(LL1); ##<<-- real dimension of centralizer of f[1]RdimCentralizer(d,k); ##<<-- dimension of centralizer of f[1] from theoryevalb(dimCentralizer=RdimCentralizer(d,k)); ##<<-- checkingequality
LL1 := [Id , e1 , e2 , e3 , e12 , e13 , e23 , e123 ]dimCentralizer := 8, 8, true
Case ๐ = 0 :> d:=1:k:=0:n1:=d+k;n2:=d-k;
A0:=diag(I$n1,-I$n2); ##<<-- this is the second matrix root of -1
n1 := 1, n2 := 1, A0 :=
โกโฃ ๐ผ 0
0 โ๐ผ
โคโฆ
> f[0]:=phi(A0,M); cmul(f[0],f[0]); Scal(f[0]), Spec(f[0]);
๐0 := e23 , โId , 0, 0
> LL0:=Centralizer(f[0],clibas); ##<<-- centralizer of f[0]dimCentralizer:=nops(LL0); ##<<-- real dimension of centralizer of f[0]RdimCentralizer(d,k); ##<<-- dimension of centralizer of f[0] from theoryevalb(dimCentralizer=RdimCentralizer(d,k)); ##<<-- checking equality
LL0 := [Id , e1 , e23 , e123 ]dimCentralizer := 4, 4, true
Case ๐ = โ1 :> d:=1:k:=-1:n1:=d+k;n2:=d-k;
Am1:=diag(I$n1,-I$n2); ##<<-- this is the third matrix root of -1
n1 := 0, n2 := 2, Am1 :=
โกโฃ โ๐ผ 0
0 โ๐ผ
โคโฆ
> f[-1]:=phi(Am1,M); cmul(f[-1],f[-1]); Scal(f[-1]), Spec(f[-1]);
๐โ1 := โe123 , โId , 0, โ1
150 E. Hitzer, J. Helmstetter and R. Ablamowicz
> LLm1:=Centralizer(f[-1],clibas); ##<<-- centralizer of f[-1]dimCentralizer:=nops(LLm1); ##<<-- real dimension of centralizer of f[-1]RdimCentralizer(d,k); ##<<--dimension of centralizer of f[-1] from theoryevalb(dimCentralizer=RdimCentralizer(d,k)); ##<<-- checking equality
LLm1 := [Id , e1 , e2 , e3 , e12 , e13 , e23 , e123 ]
dimCentralizer := 8, 8, true
We summarize roots of โ1 in ๐ถโ3,0:> โF[1]โ=evalm(F[1]); ##<<-- square root of -1 in C(2)
Mu[1]; ##<<-- minpoly of matrix F[1]โf[1]โ=f[1]; ##<<-- square root of -1 in Cl(3,0)mu[1]; ##<<-- minpoly of element f[1]
๐น1 =
โกโฃ e23 0
0 e23
โคโฆ , ๐กโ ๐ผ
๐1 = e123 , ๐กโ e123
> โF[0]โ=evalm(F[0]); ##<<-- square root of -1 in C(2)Mu[0]; ##<<-- minpoly of matrix F[0]โf[0]โ=f[0]; ##<<-- square root of -1 in Cl(3,0)mu[0]; ##<<-- minpoly of element f[0]
๐น0 =
โกโฃ e23 0
0 โe23
โคโฆ , 1 + ๐ก2
๐0 = e23 , 1 + ๐ก2
> โF[-1]โ=evalm(F[-1]); ##<<-- square root of -1 in C(2)Mu[-1]; ##<<-- minpoly of matrix F[-1]โf[-1]โ=f[-1]; ##<<-- square root of -1 in Cl(3,0)mu[-1]; ##<<-- minpoly of element f[-1]
๐นโ1 =
โกโฃ โe23 0
0 โe23
โคโฆ , ๐ก+ ๐ผ
๐โ1 = โe123 , ๐ก+ e123
Finaly, we verify that roots ๐1 and ๐โ1 are related via the reversion:> reversion(f[1])=f[-1]; evalb(%);
โe123 = โe123 , true
References
[1] R. Ablamowicz. Computations with Clifford and Grassmann algebras. Advances inApplied Clifford Algebras, 19(3โ4):499โ545, 2009.
7. Square Roots of โ1 in Real Clifford Algebras 151
[2] R. Ablamowicz and B. Fauser. CLIFFORD with bigebra โ a Maple package forcomputations with Clifford and Grassmann algebras. Available at http://math.
tntech.edu/rafal/, cโ1996โ2012.
[3] R. Ablamowicz, B. Fauser, K. Podlaski, and J. Rembielinski. Idempotents of Cliffordalgebras. Czechoslovak Journal of Physics, 53(11):949โ954, 2003.
[4] J. Armstrong. The center of an algebra. Weblog, available athttp://unapologetic.wordpress.com/2010/10/06/the-center-of-an-algebra/,accessed 22 March 2011.
[5] M. Bahri, E. Hitzer, R. Ashino, and R. Vaillancourt. Windowed Fourier transformof two-dimensional quaternionic signals. Applied Mathematics and Computation,216(8):2366โ2379, June 2010.
[6] M. Bahri, E.M.S. Hitzer, and S. Adji. Two-dimensional Clifford windowed Fouriertransform. In E.J. Bayro-Corrochano and G. Scheuermann, editors, Geometric Al-gebra Computing in Engineering and Computer Science, pages 93โ106. Springer,London, 2010.
[7] F. Brackx, R. Delanghe, and F. Sommen. Clifford Analysis, volume 76. Pitman,Boston, 1982.
[8] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[9] T. Bulow, M. Felsberg, and G. Sommer. Non-commutative hypercomplex Fouriertransforms of multidimensional signals. In G. Sommer, editor, Geometric computingwith Clifford Algebras: Theoretical Foundations and Applications in Computer Visionand Robotics, pages 187โ207, Berlin, 2001. Springer.
[10] C. Chevalley. The Theory of Lie Groups. Princeton University Press, Princeton,1957.
[11] W.K. Clifford. Applications of Grassmannโs extensive algebra. American Journal ofMathematics, 1(4):350โ358, 1878.
[12] L. Dorst and J. Lasenby, editors. Guide to Geometric Algebra in Practice. Springer,Berlin, 2011.
[13] J. Ebling and G. Scheuermann. Clifford Fourier transform on vector fields. IEEETransactions on Visualization and Computer Graphics, 11(4):469โ479, July 2005.
[14] M. Felsberg. Low-Level Image Processing with the Structure Multivector. PhD thesis,Christian-Albrechts-Universitat, Institut fur Informatik und Praktische Mathematik,Kiel, 2002.
[15] D. Hestenes. Space-Time Algebra. Gordon and Breach, London, 1966.
[16] D. Hestenes. New Foundations for Classical Mechanics. Kluwer, Dordrecht, 1999.
[17] D. Hestenes and G. Sobczyk. Clifford Algebra to Geometric Calculus. D. ReidelPublishing Group, Dordrecht, Netherlands, 1984.
[18] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[19] E. Hitzer. OPS-QFTs: A new type of quaternion Fourier transform based on theorthogonal planes split with one or two general pure quaternions. In InternationalConference on Numerical Analysis and Applied Mathematics, volume 1389 of AIP
152 E. Hitzer, J. Helmstetter and R. Ablamowicz
Conference Proceedings, pages 280โ283, Halkidiki, Greece, 19โ25 September 2011.American Institute of Physics.
[20] E. Hitzer and R. Ablamowicz. Geometric roots of โ1 in Clifford algebras ๐ถโ๐,๐ with๐ + ๐ โค 4. Advances in Applied Clifford Algebras, 21(1):121โ144, 2010. Publishedonline 13 July 2010.
[21] E. Hitzer, J. Helmstetter, and R. Ablamowicz. Maple worksheets created withCLIFFORD for a verification of results presented in this chapter. Available at: http://math.tntech.edu/rafal/publications.html, cโ2012.
[22] E. Hitzer and B. Mawardi. Uncertainty principle for Clifford geometric algebras๐ถโ๐,0, ๐ = 3(mod 4) based on Clifford Fourier transform. In T. Qian, M.I. Vai, andY. Xu, editors,Wavelet Analysis and Applications, Applied and Numerical HarmonicAnalysis, pages 47โ56. Birkhauser Basel, 2007.
[23] E.M.S. Hitzer and B. Mawardi. Clifford Fourier transform on multivector fields anduncertainty principles for dimensions ๐ = 2(mod 4) and ๐ = 3(mod 4). Advances inApplied Clifford Algebras, 18(3-4):715โ736, 2008.
[24] R.A. Horn and C.R. Johnson. Matrix Analysis. Cambridge University Press, Cam-bridge, 1985.
[25] W.M. Incorporated. Maple, a general purpose computer algebra system. http://www.maplesoft.com, cโ 2012.
[26] C. Li, A. McIntosh, and T. Qian. Clifford algebras, Fourier transform and singu-lar convolution operators on Lipschitz surfaces. Revista Matematica Iberoamericana,10(3):665โ695, 1994.
[27] H. Li. Invariant Algebras and Geometric Reasoning. World Scientific, Singapore,2009.
[28] P. Lounesto. Clifford Algebras and Spinors, volume 286 of London MathematicalSociety Lecture Notes. Cambridge University Press, 1997.
[29] B. Mawardi and E.M.S. Hitzer. Clifford Fourier transformation and uncertainty prin-ciple for the Clifford algebra ๐ถโ3,0. Advances in Applied Clifford Algebras, 16(1):41โ61, 2006.
[30] A. McIntosh. Clifford algebras, Fourier theory, singular integrals, and harmonic func-tions on Lipschitz domains. In J. Ryan, editor, Clifford Algebras in Analysis andRelated Topics, chapter 1. CRC Press, Boca Raton, 1996.
[31] T. Qian. Paley-Wiener theorems and Shannon sampling in the Clifford analysis set-ting. In R. Ablamowicz, editor, Clifford Algebras โ Applications to Mathematics,Physics, and Engineering, pages 115โ124. Birkauser, Basel, 2004.
[32] S. Said, N. Le Bihan, and S.J. Sangwine. Fast complexified quaternion Fourier trans-form. IEEE Transactions on Signal Processing, 56(4):1522โ1531, Apr. 2008.
[33] S.J. Sangwine. Biquaternion (complexified quaternion) roots of โ1. Advances inApplied Clifford Algebras, 16(1):63โ68, June 2006.
[34] Wikipedia article. Conjugacy class. Available at http://en.wikipedia.org/wiki/
Conjugacy_class, accessed 19 March 2011.
[35] Wikipedia article. Inner automorphism. Available at http://en.wikipedia.org/
wiki/Inner_automorphism, accessed 19 March 2011.
7. Square Roots of โ1 in Real Clifford Algebras 153
Eckhard HitzerCollege of Liberal Arts, Department of Material ScienceInternational Christian University181-8585 Tokyo, Japane-mail: [email protected]
Jacques HelmstetterUnivesite Grenoble IInstitut Fourier (Mathematiques)B.P. 74F-38402 Saint-Martin dโHeres, Francee-mail: [email protected]
Rafal AblamowiczDepartment of Mathematics, Box 5054Tennessee Technological UniversityCookeville, TN 38505, USAe-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 155โ176cโ 2013 Springer Basel
8 A General Geometric Fourier Transform
Roxana Bujack, Gerik Scheuermann and Eckhard Hitzer
Abstract. The increasing demand for Fourier transforms on geometric alge-bras has resulted in a large variety. Here we introduce one single straight-forward definition of a general geometric Fourier transform covering mostversions in the literature. We show which constraints are additionally neces-sary to obtain certain features such as linearity or a shift theorem. As a result,we provide guidelines for the target-oriented design of yet unconsidered trans-forms that fulfill requirements in a specific application context. Furthermore,the standard theorems do not need to be shown in a slightly different formevery time a new geometric Fourier transform is developed since they areproved here once and for all.
Mathematics Subject Classification (2010). Primary 15A66, 11E88; secondary42A38, 30G35.
Keywords. Fourier transform, geometric algebra, Clifford algebra, image pro-cessing, linearity, scaling, shift.
1. Introduction
The Fourier transform by Jean Baptiste Joseph Fourier is an indispensable toolin many fields of mathematics, physics, computer science and engineering, espe-cially for the analysis and solution of differential equations, or in signal and imageprocessing, fields which cannot be imagined without it. The kernel of the Fouriertransform consists of the complex exponential function. With the square root ofminus one, the imaginary unit ๐, as part of the argument it is periodic and thereforesuitable for the analysis of oscillating systems.
William Kingdon Clifford created the geometric algebras in 1878, [8]. Theyusually contain continuous submanifolds of geometric square roots of โ1 [16, 17].Each multivector has a natural geometric interpretation so the generalization ofthe Fourier transform to multivector-valued functions in the geometric algebras isvery reasonable. It helps to interpret the transform, apply it in a target-oriented
156 R. Bujack, G. Scheuermann and E. Hitzer
way to the specific underlying problem and it allows a new point of view on fluidmechanics.
Many different application-oriented definitions of Fourier transforms in geo-metric algebras have been developed. For example the CliffordโFourier transformintroduced by Jancewicz [19] and expanded by Ebling and Scheuermann [10]and Hitzer and Mawardi [18] or the one established by Sommen in [21] and re-established by Bulow [7]. Further we have the quaternionic Fourier transform byEll [11] and later by Bulow [7], the spacetime Fourier transform by Hitzer [15], theCliffordโFourier transform for colour images by Batard et al. [1], the CylindricalFourier transform by Brackx et al. [6], the transforms by Felsberg [13] or Ell andSangwine [20, 12]. All these transforms have different interesting properties anddeserve to be studied independently from one another. But the analysis of theirsimilarities reveals a lot about their qualities, too. We concentrate on this matterand summarize all of them in one general definition.
Recently there have been very successful approaches by De Bie, Brackx, DeSchepper and Sommen to construct CliffordโFourier transforms from operator ex-ponentials and differential equations [3, 4, 9, 5]. The definition presented in thischapter does not cover all of them, partly because their closed integral form is notalways known or is highly complicated, and partly because they can be producedby combinations and functions of our transforms.
We focus on continuous geometric Fourier transforms over flat spaces โ๐,๐ intheir integral representation. That way their finite, regular discrete equivalents asused in computational signal and image processing can be intuitively constructedand direct applicability to the existing practical issues and easy numerical man-ageability are ensured.
2. Definition of the GFT
We examine geometric algebras ๐ถโ๐,๐, ๐+ ๐ = ๐ โ โ over โ๐+๐ [14] generated bythe associative, bilinear geometric product with neutral element 1 satisfying
๐๐๐๐ + ๐๐๐๐ = ๐๐๐ฟ๐๐, (2.1)
for all ๐, ๐ โ {1, . . . , ๐} with the Kronecker symbol ๐ฟ and
๐๐ =
{1 โ๐ = 1, . . . , ๐,
โ1 โ๐ = ๐+ 1, . . . , ๐.(2.2)
For the sake of brevity we want to refer to arbitrary multivectors
๐จ =
๐โ๐=0
โ1โค๐1<โ โ โ <๐๐โค๐
๐๐1...๐๐๐๐1 . . .๐๐๐ โ ๐ถโ๐,๐, (2.3)
where ๐๐1...๐๐ โ โ, as
๐จ =โ๐
๐๐๐๐ . (2.4)
8. A General Geometric Fourier Transform 157
where each of the 2๐ multi-indices ๐ โ {1, . . . , ๐} indicates a basis vector of ๐ถโ๐,๐by ๐๐ = ๐๐1 . . .๐๐๐ , 1 โค ๐1 < โ โ โ < ๐๐ โค ๐, ๐โ = ๐0 = 1 and its associatedcoefficient ๐๐ = ๐๐1...๐๐ โ โ.
Definition 2.1. The exponential function of a multivector ๐จ โ ๐ถโ๐,๐ is defined bythe power series
๐๐จ :=โโ
๐=0
๐จ๐
๐!. (2.5)
Lemma 2.2. For two multivectors ๐จ๐ฉ = ๐ฉ๐จ that commute we have
๐๐จ+๐ฉ = ๐๐จ๐๐ฉ. (2.6)
Proof. Analogous to the exponent rule of real matrices. โกNotation 2.3. For each geometric algebra๐ถโ๐,๐ we will write I ๐,๐ = {๐ โ ๐ถโ๐,๐, ๐
2 โโโ} to denote the real multiples of all geometric square roots of โ1 , compare [16]and [17]. We choose the symbol I to be reminiscent of the imaginary numbers.
Definition 2.4. Let ๐ถโ๐,๐ be a geometric algebra, ๐จ : โ๐ โ ๐ถโ๐,๐ be a mul-tivector field and ๐,๐ โ โ๐ vectors. A Geometric Fourier Transform (GFT)โฑ๐น1,๐น2(๐จ) is defined by two ordered finite sets ๐น1 = {๐1(๐,๐), . . . , ๐๐(๐,๐)}, ๐น2 ={๐๐+1(๐,๐), . . . , ๐๐(๐,๐)} of mappings ๐๐(๐,๐) : โ
๐รโ๐ โ I ๐,๐, โ๐ = 1, . . . , ๐and the calculation rule
โฑ๐น1,๐น2(๐จ)(๐) :=
โซโ๐
โ๐โ๐น1
๐โ๐(๐,๐)๐จ(๐)โ๐โ๐น2
๐โ๐(๐,๐) d๐๐. (2.7)
This definition combines many Fourier transforms into a single general one. Itenables us to prove the well-known theorems which depend only on the propertiesof the chosen mappings.
Example. Depending on the choice of ๐น1 and ๐น2 we obtain previously publishedtransforms.
1. In the case of ๐จ : โ๐ โ ๐ข๐,0, ๐ = 2 (mod 4) or ๐ = 3 (mod 4), we canreproduce the CliffordโFourier transform introduced by Jancewicz [19] for๐ = 3 and expanded by Ebling and Scheuermann [10] for ๐ = 2 and HitzerandMawardi [18] for ๐ = 2 (mod 4) or ๐ = 3 (mod 4) using the configuration
๐น1 = โ ,๐น2 = {๐1},
๐1(๐,๐) = 2๐๐๐๐ โ ๐,(2.8)
with ๐๐ being the pseudoscalar of ๐บ๐,0.2. Choosing multivector fields โ๐ โ ๐ข0,๐,
๐น1 = โ ,๐น2 = {๐1, . . . , ๐๐},
๐๐(๐,๐) = 2๐๐๐๐ฅ๐๐ข๐, โ๐ = 1, . . . , ๐
(2.9)
we have the SommenโBulowโCliffordโFourier transform from [21, 7].
158 R. Bujack, G. Scheuermann and E. Hitzer
3. For๐จ : โ2 โ ๐ข0,2 โ โ the quaternionic Fourier transform [11, 7] is generatedby
๐น1 = {๐1},๐น2 = {๐2},
๐1(๐,๐) = 2๐๐๐ฅ1๐ข1,
๐2(๐,๐) = 2๐๐๐ฅ2๐ข2.
(2.10)
4. Using ๐ข3,1 we can build the spacetime, respectively the volume-time, Fouriertransform from [15]1 with the ๐ข3,1-pseudoscalar ๐4 as follows
๐น1 = {๐1},๐น2 = {๐2},
๐1(๐,๐) = ๐4๐ฅ4๐ข4,
๐2(๐,๐) = ๐4๐4๐4(๐ฅ1๐ข1 + ๐ฅ2๐ข2 + ๐ฅ3๐ข3).
(2.11)
5. The CliffordโFourier transform for colour images by Batard, Berthier andSaint-Jean [1] for ๐ = 2, ๐ = 4,๐จ : โ2 โ ๐ข4,0, a fixed bivector ๐ฉ, and thepseudoscalar ๐ can intuitively be written as
๐น1 = {๐1},๐น2 = {๐2},
๐1(๐,๐) =1
2(๐ฅ1๐ข1 + ๐ฅ2๐ข2)(๐ฉ + ๐๐ฉ),
๐2(๐,๐) = โ1
2(๐ฅ1๐ข1 + ๐ฅ2๐ข2)(๐ฉ + ๐๐ฉ),
(2.12)
but (๐ฉ+ ๐๐ฉ) does not square to a negative real number, see [16]. The specialproperty that ๐ฉ and ๐๐ฉ commute allows us to express the formula using
๐น1 = {๐1, ๐2},๐น2 = {๐3, ๐4},
๐1(๐,๐) =1
2(๐ฅ1๐ข1 + ๐ฅ2๐ข2)๐ฉ,
๐2(๐,๐) =1
2(๐ฅ1๐ข1 + ๐ฅ2๐ข2)๐๐ฉ,
๐3(๐,๐) = โ1
2(๐ฅ1๐ข1 + ๐ฅ2๐ข2)๐ฉ,
๐4(๐,๐) = โ1
2(๐ฅ1๐ข1 + ๐ฅ2๐ข2)๐๐ฉ,
(2.13)
which fulfills the conditions of Definition 2.4.
1Please note that Hitzer uses a different notation in [15]. His ๐ = ๐ก๐0 + ๐ฅ1๐1 + ๐ฅ2๐2 + ๐ฅ3๐3corresponds to our ๐ = ๐ฅ1๐1 + ๐ฅ2๐2 + ๐ฅ3๐3 + ๐ฅ4๐4, with ๐0๐0 = ๐0 = โ1 being equivalent toour ๐4๐4 = ๐4 = โ1.
8. A General Geometric Fourier Transform 159
6. Using ๐ข0,๐ and
๐น1 = {๐1},๐น2 = โ ,
๐1(๐,๐) = โ๐ โง ๐
(2.14)
produces the cylindrical Fourier transform as introduced by Brackx, de Schep-per and Sommen in [6].
3. General Properties
First we prove general properties valid for arbitrary sets ๐น1, ๐น2.
Theorem 3.1 (Existence). The geometric Fourier transform exists for all integrablemultivector fields ๐จ โ ๐ฟ1(โ
๐).
Proof. The property
๐2๐ (๐,๐) โ โโ (3.1)
of the mappings ๐๐ for ๐ = 1, . . . , ๐ leads to
๐2๐ (๐,๐)
โฃ๐2๐ (๐,๐)โฃ
= โ1 (3.2)
for all ๐๐(๐,๐) โ= 0. So using the decomposition
๐๐(๐,๐) =๐๐(๐,๐)
โฃ๐๐(๐,๐)โฃ โฃ๐๐(๐,๐)โฃ (3.3)
we can write โ๐ โ โ
๐ ๐๐(๐,๐) =
โงโจโฉ(โ1)๐โฃ๐๐(๐,๐)โฃ๐ for ๐ = 2๐, ๐ โ โ0
(โ1)๐ ๐๐(๐,๐)โฃ๐๐(๐,๐)โฃ โฃ๐๐(๐,๐)โฃ
๐ for ๐ = 2๐ + 1, ๐ โ โ0(3.4)
which results in
๐โ๐๐(๐,๐) =
โโ๐=0
(โ๐๐(๐,๐))๐
๐!
=
โโ๐=0
(โ1)๐โฃ๐๐(๐,๐)โฃ2๐(2๐)!
โ ๐๐(๐,๐)
โฃ๐๐(๐,๐)โฃโโ๐=0
(โ1)๐ โฃ๐๐(๐,๐)โฃ2๐+1
(2๐ + 1)!
= cos (โฃ๐๐(๐,๐)โฃ)โ ๐๐(๐,๐)
โฃ๐๐(๐,๐)โฃ sin (โฃ๐๐(๐,๐)โฃ) .
(3.5)
160 R. Bujack, G. Scheuermann and E. Hitzer
Because of
โฃ๐โ๐๐(๐,๐)โฃ =โฃโฃโฃโฃcos (โฃ๐๐(๐,๐)โฃ)โ ๐๐(๐,๐)
โฃ๐๐(๐,๐)โฃ sin (โฃ๐๐(๐,๐)โฃ)โฃโฃโฃโฃ
โค โฃcos (โฃ๐๐(๐,๐)โฃ)โฃ+โฃโฃโฃโฃ ๐๐(๐,๐)โฃ๐๐(๐,๐)โฃ
โฃโฃโฃโฃ โฃsin (โฃ๐๐(๐,๐)โฃ)โฃโค 2
(3.6)
the magnitude of the improper integral
โฃโฑ๐น1,๐น2(๐จ)(๐)โฃ =โฃโฃโฃโฃโฃโฃโซโ๐
โ๐โ๐น1
๐โ๐(๐,๐)๐จ(๐)โ๐โ๐น2
๐โ๐(๐,๐) d๐๐
โฃโฃโฃโฃโฃโฃโคโซโ๐
โ๐โ๐น1
โฃโฃโฃ๐โ๐(๐,๐)โฃโฃโฃ โฃ๐จ(๐)โฃ
โ๐โ๐น2
โฃโฃโฃ๐โ๐(๐,๐)โฃโฃโฃ d๐๐
โคโซโ๐
โ๐โ๐น1
2 โฃ๐จ(๐)โฃโ๐โ๐น2
2 d๐๐
= 2๐โซโ๐
โฃ๐จ(๐)โฃ d๐๐
(3.7)
is finite and therefore the geometric Fourier transform exists. โก
Theorem 3.2 (Scalar linearity). The geometric Fourier transform is linear with re-spect to scalar factors. Let ๐, ๐ โ โ and ๐จ,๐ฉ,๐ช : โ๐ โ ๐ถโ๐,๐ be three multivectorfields that satisfy ๐จ(๐) = ๐๐ฉ(๐) + ๐๐ช(๐), then
โฑ๐น1,๐น2(๐จ)(๐) = ๐โฑ๐น1,๐น2(๐ฉ)(๐) + ๐โฑ๐น1,๐น2(๐ช)(๐). (3.8)
Proof. The assertion is an easy consequence of the distributivity of the geomet-ric product over addition, the commutativity of scalars and the linearity of theintegral. โก
4. Bilinearity
All geometric Fourier transforms from the introductory example can also be ex-pressed in terms of a stronger claim. The mappings ๐1, . . . , ๐๐, with the first ๐terms to the left of the argument function and the ๐ โ ๐ others on the right of it,are all bilinear and therefore take the form
๐๐(๐,๐) = ๐๐
(๐โ๐=1
๐ฅ๐๐๐ ,๐โ๐=1
๐ข๐๐๐
)
=๐โ
๐,๐=1
๐ฅ๐๐๐(๐๐ , ๐๐)๐ข๐ = ๐๐๐๐๐,
(4.1)
โ๐ = 1, . . . , ๐, where ๐๐ โ (I ๐,๐)๐ร๐, (๐๐)๐๐ = ๐๐(๐๐ , ๐๐) according to Nota-tion 2.3.
8. A General Geometric Fourier Transform 161
Example. Ordered in the same way as in the previous example, the geometricFourier transforms expressed in the way of (4.1) take the following shapes:
1. In the CliffordโFourier transform ๐1 can be written with
๐1 = 2๐๐๐ Id . (4.2)
2. The ๐ = ๐ = ๐ mappings ๐๐, ๐ = 1, . . . , ๐ of the BulowโCliffordโFouriertransform can be expressed using
(๐๐)๐๐ =
{2๐๐๐ for ๐ = ๐ = ๐,
0 otherwise.(4.3)
3. Similarly the quaternionic Fourier transform is generated using
(๐1)๐๐ =
{2๐๐ for ๐ = ๐ = 1,
0 otherwise,
(๐2)๐๐ =
{2๐๐ for ๐ = ๐ = 2,
0 otherwise.
(4.4)
4. We can build the spacetime Fourier transform with
(๐1)๐๐ =
{๐4 for ๐ = ๐ = 1,
0 otherwise,
(๐2)๐๐ =
{๐4๐4๐4 for ๐ = ๐ โ {2, 3, 4},0 otherwise.
(4.5)
5. The CliffordโFourier transform for colour images can be described by
๐1 =1
2๐ฉ Id,
๐2 =1
2๐๐ฉ Id,
๐3 = โ1
2๐ฉ Id,
๐4 = โ1
2๐๐ฉ Id .
(4.6)
6. The cylindrical Fourier transform can also be reproduced with mappingssatisfying (4.1) because we can write
๐ โง ๐ = ๐1๐2๐ฅ1๐ข2 โ ๐1๐2๐ฅ2๐ข1
+ โ โ โ + ๐๐โ1๐๐๐ฅ๐โ1๐ข๐ โ ๐๐โ1๐๐๐ฅ๐๐ข๐โ1
(4.7)
and set
(๐1)๐๐ =
{0 for ๐ = ๐,
๐๐๐๐ otherwise.(4.8)
162 R. Bujack, G. Scheuermann and E. Hitzer
Theorem 4.1 (Scaling). Let 0 โ= ๐ โ โ be a real number, ๐จ(๐) = ๐ฉ(๐๐) twomultivector fields and all ๐น1, ๐น2 be bilinear mappings then the geometric Fouriertransform satisfies
โฑ๐น1,๐น2(๐จ)(๐) = โฃ๐โฃโ๐โฑ๐น1,๐น2(๐ฉ)(๐๐
). (4.9)
Proof. A change of coordinates together with the bilinearity proves the assertion by
โฑ๐น1,๐น2(๐จ)(๐) =
โซโ๐
โ๐โ๐น
๐โ๐(๐,๐)๐ฉ(๐๐)โ๐โ๐ต
๐โ๐(๐,๐) d๐๐
๐๐=๐=
โซโ๐
โ๐โ๐น
๐โ๐(๐๐ ,๐)๐ฉ(๐)
โ๐โ๐ต
๐โ๐(๐๐ ,๐)โฃ๐โฃโ๐ d๐๐
๐ bilin.= โฃ๐โฃโ๐
โซโ๐
โ๐โ๐น
๐โ๐(๐,๐๐ )๐ฉ(๐)
โ๐โ๐ต
๐โ๐(๐,๐๐ ) d๐๐ (4.10)
= โฃ๐โฃโ๐โฑ๐น1,๐น2(๐ฉ)(๐๐
). โก
5. Products with Invertible Factors
To obtain properties of the GFT like linearity with respect to arbitrary multi-vectors or a shift theorem we will have to change the order of multivectors andproducts of exponentials. Since the geometric product usually is neither commu-tative nor anticommutative this is not trivial. In this section we provide usefulLemmata that allow a swap if at least one of the factors is invertible. For moreinformation see [14] and [17].
Remark 5.1. Every multiple of a square root of โ1, ๐ โ I ๐,๐ is invertible, sincefrom ๐2 = โ๐, ๐ โ โ โ {0} follows ๐โ1 = โ๐/๐. Because of that, for all ๐,๐ โ โ๐ afunction ๐๐(๐,๐) : โ
๐ ร โ๐ โ I ๐,๐ is pointwise invertible.
Definition 5.2. For an invertible multivector ๐ฉ โ ๐ถโ๐,๐ and an arbitrary multivec-tor ๐จ โ ๐ถโ๐,๐ we define
๐จ๐0(๐ฉ) =1
2(๐จ+๐ฉโ1๐จ๐ฉ),
๐จ๐1(๐ฉ) =1
2(๐จโ๐ฉโ1๐จ๐ฉ).
(5.1)
Lemma 5.3. Let ๐ฉ โ ๐ถโ๐,๐ be invertible with the unique inverse ๐ฉโ1 = ๏ฟฝ๏ฟฝ/๐ฉ2,
๐ฉ2 โ โ โ {0}. Every multivector ๐จ โ ๐ถโ๐,๐ can be expressed unambiguously by thesum of ๐จ๐0(๐ฉ) โ ๐ถโ๐,๐ that commutes and ๐จ๐1(๐ฉ) โ ๐ถโ๐,๐ that anticommutes withrespect to ๐ฉ. That means
๐จ = ๐จ๐0(๐ฉ) +๐จ๐1(๐ฉ),
๐จ๐0(๐ฉ)๐ฉ = ๐ฉ๐จ๐0(๐ฉ),
๐จ๐1(๐ฉ)๐ฉ = โ๐ฉ๐จ๐1(๐ฉ).
(5.2)
8. A General Geometric Fourier Transform 163
Proof. We will only prove the assertion for ๐จ๐0(๐ฉ).
Existence: With Definition 5.2 we get
๐จ๐0(๐ฉ) +๐จ๐1(๐ฉ) =1
2(๐จ+๐ฉโ1๐จ๐ฉ +๐จโ๐ฉโ1๐จ๐ฉ)
= ๐จ(5.3)
and considering
๐ฉโ1๐จ๐ฉ =๏ฟฝ๏ฟฝ๐จ๐ฉ
๐ฉ2 = ๐ฉ๐จ๐ฉโ1 (5.4)
we also get
๐จ๐0(๐ฉ)๐ฉ =1
2(๐จ+๐ฉโ1๐จ๐ฉ)๐ฉ
=1
2(๐จ+๐ฉ๐จ๐ฉโ1)๐ฉ
=1
2(๐จ๐ฉ +๐ฉ๐จ)
= ๐ฉ1
2(๐ฉโ1๐จ๐ฉ +๐จ)
= ๐ฉ๐จ๐0(๐ฉ)
(5.5)
Uniqueness: From the first claim in (5.2) we get
๐จ๐1(๐ฉ) = ๐จโ๐จ๐0(๐ฉ), (5.6)
together with the third one this leads to
(๐จโ๐จ๐0(๐ฉ))๐ฉ = โ๐ฉ(๐จโ๐จ๐0(๐ฉ))
๐จ๐ฉ โ๐จ๐0(๐ฉ)๐ฉ = โ๐ฉ๐จ+๐ฉ๐จ๐0(๐ฉ)
๐จ๐ฉ +๐ฉ๐จ = ๐จ๐0(๐ฉ)๐ฉ +๐ฉ๐จ๐0(๐ฉ)
(5.7)
and from the second claim finally follows
๐จ๐ฉ +๐ฉ๐จ = 2๐ฉ๐จ๐0(๐ฉ)
1
2(๐ฉโ1๐จ๐ฉ +๐จ) = ๐จ๐0(๐ฉ).
(5.8)
The derivation of the expression for ๐จ๐1(๐ฉ) works analogously. โกCorollary 5.4 (Decomposition w.r.t. commutativity). Let ๐ฉ โ ๐ถโ๐,๐ be invertible,then โ๐จ โ ๐ถโ๐,๐
๐ฉ๐จ = (๐จ๐0(๐ฉ) โ๐จ๐1(๐ฉ))๐ฉ. (5.9)
Definition 5.5. For ๐ โ โ,๐จ โ ๐ถโ๐,๐, the ordered set ๐ต = {๐ฉ1, . . . ,๐ฉ๐} of invert-ible multivectors and any multi-index ๐ โ {0, 1}๐ we define
๐จ๐๐(โโ๐ต )
: = ((๐จ๐๐1 (๐ฉ1))๐๐2 (๐ฉ2) . . .)๐๐๐ (๐ฉ๐),
๐จ๐๐(โโ๐ต )
: = ((๐จ๐๐๐ (๐ฉ๐))๐๐๐โ1 (๐ฉ๐โ1). . .)๐๐1 (๐ฉ1)
(5.10)
recursively with ๐0, ๐1 as in Definition 5.2.
164 R. Bujack, G. Scheuermann and E. Hitzer
Example. Let ๐จ = ๐0 + ๐1๐1 + ๐2๐2 + ๐12๐12 โ ๐ข2,0 then, for example
๐จ๐0(๐1) =1
2(๐จ+ ๐โ1
1 ๐จ๐1)
=1
2(๐จ+ ๐0 + ๐1๐1 โ ๐2๐2 โ ๐12๐12)
= ๐0 + ๐1๐1
(5.11)
and further
๐จ๐0,0(โโโโ๐1,๐2) = (๐จ๐0(๐1))๐0(๐2)
= (๐0 + ๐1๐1)๐0(๐2) = ๐0.(5.12)
The computation of the other multi-indices with ๐ = 2 works analogously andtherefore
๐จ =โ
๐โ{0,1}๐๐จ๐๐(๐1,๐2)
= ๐จ๐00(โโโโ๐1,๐2) +๐จ๐01(โโโโ๐1,๐2) +๐จ๐10(โโโโ๐1,๐2) +๐จ๐11(โโโโ๐1,๐2)
= ๐0 + ๐1๐1 + ๐2๐2 + ๐12๐12.
(5.13)
Lemma 5.6. Let ๐ โ โ, ๐ต = {๐ฉ1, . . . ,๐ฉ๐} be invertible multivectors and for ๐ โ{0, 1}๐ let โฃ๐โฃ :=โ๐
๐=1 ๐๐, then โ๐จ โ ๐ถโ๐,๐
๐จ =โ
๐โ{0,1}๐๐จ
๐๐(โโ๐ต )
,
๐จ๐ฉ1 . . .๐ฉ๐ = ๐ฉ1 . . .๐ฉ๐
โ๐โ{0,1}๐
(โ1)โฃ๐โฃ๐จ๐๐(โโ๐ต )
,
๐ฉ1 . . .๐ฉ๐๐จ =โ
๐โ{0,1}๐(โ1)โฃ๐โฃ๐จ
๐๐(โโ๐ต )
๐ฉ1 . . .๐ฉ๐.
(5.14)
Proof. Apply Lemma 5.3 repeatedly. โก
Remark 5.7. The distinction of the two directions can be omitted using the equality
๐จ๐๐(โโโโโโโ๐ฉ1,...,๐ฉ๐)
= ๐จ๐๐(โโโโโโโ๐ฉ๐,...,๐ฉ1)
. (5.15)
We established it for the sake of notational brevity and will not formulate norprove every assertion for both directions.
Lemma 5.8. Let ๐น = {๐1(๐,๐), . . . , ๐๐(๐,๐)} be a set of pointwise invertible func-tions then the ordered product of their exponentials and an arbitrary multivector๐จ โ ๐ถโ๐,๐ satisfies
๐โ๐=1
๐โ๐๐(๐,๐)๐จ =โ
๐โ{0,1}๐๐จ
๐๐(โโ๐น )
(๐,๐)
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐), (5.16)
where ๐ด๐๐(โโ๐น )
(๐,๐) := ๐ด๐๐(โโโโโ๐น (๐,๐))
is a multivector-valued function โ๐ ร โ๐ โ๐ถโ๐,๐.
8. A General Geometric Fourier Transform 165
Proof. For all ๐,๐ โ โ๐ the commutation properties of ๐๐(๐,๐) dictate the prop-erties of ๐โ๐๐(๐,๐) by
๐โ๐๐(๐,๐)๐จDef. 2.1=
โโ๐=0
(โ๐๐(๐,๐))๐
๐!๐จ
Lem. 5.3=
โโ๐=0
(โ๐๐(๐,๐))๐
๐!
(๐จ๐0(๐๐(๐,๐)) +๐จ๐1(๐๐(๐,๐))
).
(5.17)
The shape of this decomposition of ๐จ may depend on ๐ and ๐. To stress this factwe will interpret ๐จ๐0(๐๐(๐,๐)) as a multivector function and write ๐จ๐0(๐๐)(๐,๐).According to Lemma 5.3 we can move ๐จ๐0(๐๐)(๐,๐) through all factors, because itcommutes. Analogously swapping ๐จ๐1(๐๐)(๐,๐) will change the sign of each factorbecause it anticommutes. Hence we get
= ๐จ๐0(๐๐)(๐,๐)
โโ๐=0
(โ๐๐(๐,๐))๐
๐!+๐จ๐1(๐๐)(๐,๐)
โโ๐=0
(๐๐(๐,๐))๐
๐!
= ๐จ๐0(๐๐)(๐,๐)๐โ๐๐(๐,๐) +๐จ๐1(๐๐)(๐,๐)๐
๐๐(๐,๐).
(5.18)
Applying this repeatedly to the product we can deduce
๐โ๐=1
๐โ๐๐(๐,๐)๐จ =
๐โ1โ๐=1
๐โ๐๐(๐,๐)
(๐จ๐0(๐๐)(๐,๐)๐
โ๐๐(๐,๐)
+๐จ๐1(๐๐)(๐,๐)๐๐๐(๐,๐)
)
=
๐โ2โ๐=1
๐โ๐๐(๐,๐)
โโโโโโโโ๐จ
๐0,0(โโโโโโ๐๐โ1,๐๐)
(๐,๐)๐โ๐๐โ1(๐,๐)๐โ๐๐(๐,๐)
+๐จ๐1,0(
โโโโโโ๐๐โ1,๐๐)
(๐,๐)๐๐๐โ1(๐,๐)๐โ๐๐(๐,๐)
+๐จ๐0,1(
โโโโโโ๐๐โ1,๐๐)
(๐,๐)๐โ๐๐โ1(๐,๐)๐๐๐(๐,๐)
+๐จ๐1,1(
โโโโโโ๐๐โ1,๐๐)
(๐,๐)๐๐๐โ1(๐,๐)๐๐๐(๐,๐)
โโโโโโโโ ...
...... (5.19)
=โ
๐โ{0,1}๐๐จ
๐๐(โโ๐น )
(๐,๐)๐โ
๐=1
๐โ(โ1)๐๐๐๐(๐,๐). โก
6. Separable GFT
From now on we want to restrict ourselves to an important group of geometricFourier transforms whose square roots of โ1 are independent from the first argu-ment.
Definition 6.1. We call a GFT left (right) separable, if
๐๐ = โฃ๐๐ (๐,๐)โฃ ๐๐(๐), (6.1)
โ๐ = 1, . . . , ๐, (๐ = ๐+1, . . . , ๐), where โฃ๐๐(๐,๐)โฃ : โ๐รโ๐ โ โ is a real functionand ๐๐ : โ
๐ โ I ๐,๐ a function that does not depend on ๐.
166 R. Bujack, G. Scheuermann and E. Hitzer
Example. The first five transforms from the introductory example are separable,while the cylindrical transform (vi) can not be expressed as in (6.1) except for thetwo-dimensional case.
We have seen in the proof of Lemma 5.8 that the decomposition of a con-stant multivector ๐จ with respect to a product of exponentials generally resultsin multivector-valued functions ๐จ๐๐(๐น )(๐,๐) of ๐ and ๐. Separability guaranteesindependence from ๐ and therefore allows separation from the integral.
Corollary 6.2 (Decomposition independent from ๐). Consider a set of functions๐น = {๐1(๐,๐), . . . , ๐๐(๐,๐)} satisfying condition (6.1) then the ordered product oftheir exponentials and an arbitrary multivector ๐จ โ ๐ถโ๐,๐ satisfies
๐โ๐=1
๐โ๐๐(๐,๐)๐จ =โ
๐โ{0,1}๐๐จ
๐๐(โโ๐น )
(๐)
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐). (6.2)
Remark 6.3. If a GFT can be expressed as in 6.1 but with multiples of squareroots of โ1, ๐๐ โ I ๐,๐, which are independent from ๐ and ๐, the parts ๐จ
๐๐(โโ๐น )
of ๐จ will be constants. Note that the first five GFTs from the reference examplesatisfy this stronger condition, too.
Definition 6.4. For a set of functions ๐น = {๐1(๐,๐), . . . , ๐๐(๐,๐)} and a multi-index ๐ โ {0, 1}๐, we define the set of functions ๐น (๐) by
๐น (๐) :={(โ1)๐1๐1(๐,๐), . . . , (โ1)๐๐๐๐(๐,๐)
}. (6.3)
Theorem 6.5 (Left and right products). Let ๐ช โ ๐ถโ๐,๐ and ๐จ,๐ฉ : โ๐ โ ๐ถโ๐,๐ betwo multivector fields with ๐จ(๐) = ๐ช๐ฉ(๐) then a left separable geometric Fouriertransform obeys
โฑ๐น1,๐น2(๐จ)(๐) =โ
๐โ{0,1}๐๐ช
๐๐(โโ๐น1)
(๐)โฑ๐น1(๐),๐น2(๐ฉ)(๐). (6.4)
If ๐จ(๐) = ๐ฉ(๐)๐ช we analogously get
โฑ๐น1,๐น2(๐จ)(๐) =โ
๐โ{0,1}(๐โ๐)
โฑ๐น1,๐น2(๐)(๐ฉ)(๐)๐ช๐๐(โโ๐น2)
(๐)(6.5)
for a right separable GFT.
Proof. We restrict ourselves to the proof of the first assertion.
โฑ๐น1,๐น2(๐จ)(๐) =
โซโ๐
โ๐โ๐น1
๐โ๐(๐,๐)๐ช๐ฉ(๐)โ๐โ๐น2
๐โ๐(๐,๐) d๐๐
Lem. 5.8=
โซโ๐
โโ โ๐โ{0,1}๐
๐ช๐๐(โโ๐น1)
(๐)
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐)
โโ ๐ฉ(๐)
โ๐โ๐น2
๐โ๐(๐,๐) d๐๐
8. A General Geometric Fourier Transform 167
=โ
๐โ{0,1}๐๐ช
๐๐(โโ๐น1)
(๐)
โซโ๐
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐)
๐ฉ(๐)โ๐โ๐น2
๐โ๐(๐,๐) d๐๐
=โ
๐โ{0,1}๐๐ช
๐๐(โโ๐น1)
(๐)โฑ๐น1(๐),๐น2(๐ฉ)(๐).
The second one follows in the same way. โก
Corollary 6.6 (Uniform constants). Let the claims from Theorem 6.5 hold. If theconstant ๐ช satisfies ๐ช = ๐ช
๐๐(โโ๐น1)
(๐) for a multi-index ๐ โ {0, 1}๐ then the theorem
simplifies to
โฑ๐น1,๐น2(๐จ)(๐) = ๐ชโฑ๐น1(๐),๐น2(๐ฉ)(๐) (6.6)
for ๐จ(๐) = ๐ช๐ฉ(๐) respectively
โฑ๐น1,๐น2(๐จ)(๐) = โฑ๐น1,๐น2(๐)(๐ฉ)(๐)๐ช (6.7)
for ๐จ(๐) = ๐ฉ(๐)๐ช and ๐ช = ๐ช๐๐(โโ๐น2)
(๐) for a multi-index ๐ โ {0, 1}(๐โ๐).2
Corollary 6.7 (Left and right linearity). The geometric Fourier transform is left(respectively right) linear if ๐น1 (respectively ๐น2) consists only of functions ๐๐ withvalues in the center of ๐ถโ๐,๐, that means โ๐,๐ โ โ๐, โ๐จ โ ๐ถโ๐,๐ : ๐จ๐๐(๐,๐) =๐๐(๐,๐)๐จ.
Remark 6.8. Note that for empty sets ๐น1 (or ๐น2) necessarily all elements satisfycommutativity and therefore the condition in Corollary 6.7.
The different appearances of Theorem 6.5 are summarized in Table 1 andTable 2.
We have seen how to change the order of a multivector and a product ofexponentials in the previous section. To get a shift theorem we will have to separatesums appearing in the exponent and sort the resulting exponentials with respect tothe summands. Note that Corollary 6.2 can be applied in two ways here, becauseexponentials appear on both sides.
Not every factor will need to be swapped with every other. So, to keep thingsshort, we will make use of the notation ๐(๐ฝ)๐(๐1, . . . , ๐๐, 0, . . . , 0) for ๐ โ {1, . . . , ๐}instead of distinguishing between differently sized multi-indices for every ๐ thatappears. The zeros at the end substitutionally indicate real numbers. They com-mute with every multivector. That implies, that for the last ๐โ ๐ factors no swapand therefore no separation needs to be made. It would also be possible to use thenotation ๐(๐ฝ)๐(๐1, . . . , ๐๐โ1, 0, . . . , 0) for ๐ โ {1, . . . , ๐}, because every function com-mutes with itself. The choice we have made means that no exceptional treatment
2Corollary 6.6 follows directly from (๐ช๐๐(
โโ๐น1)
)๐๐(
โโ๐น1)
= 0 for all ๐ โ= ๐ because no non-zero
component of ๐ช can commute and anticommute with respect to a function in ๐น1.
168 R. Bujack, G. Scheuermann and E. Hitzer
Table 1. Theorem 6.5 (Left products) applied to the GFTs of the firstexample enumerated in the same order.Notations: on the LHS โฑ๐น1,๐น2 = โฑ๐น1,๐น2(๐จ)(๐), on the RHS โฑ๐น โฒ
1,๐นโฒ2=
โฑ๐น โฒ1,๐น
โฒ2(๐ฉ)(๐)
GFT ๐จ(๐) = ๐ช๐ฉ(๐)
1. Clifford โฑ๐1 = ๐ชโฑ๐1
2. Bulow โฑ๐1,...,๐๐ = ๐ชโฑ๐1,...,๐๐
3. Quaternionic โฑ๐1,๐2 = ๐ช๐0(๐)โฑ๐1,๐2 +๐ช๐1(๐)โฑโ๐1,๐24. Spacetime โฑ๐1,๐2 = ๐ช๐0(๐4)โฑ๐1,๐2 +๐ช๐1(๐4)โฑโ๐1,๐25. Colour Image โฑ๐1,๐2,๐3,๐4 = ๐ช
๐00(โโโโ๐ฉ,๐๐ฉ)
โฑ๐1,๐2,๐3,๐4
+๐ช๐10(
โโโโ๐ฉ,๐๐ฉ)
โฑโ๐1,๐2,๐3,๐4+๐ช
๐01(โโโโ๐ฉ,๐๐ฉ)
โฑ๐1,โ๐2,๐3,๐4
+๐ช๐11(
โโโโ๐ฉ,๐๐ฉ)
โฑโ๐1,โ๐2,๐3,๐46. Cylindrical ๐ = 2 โฑ๐1 = ๐ช๐0(๐12)โฑ๐1 +๐ช๐1(๐12)โฑโ๐1
Cylindrical ๐ โ= 2 -
of ๐1 is necessary. But please note that the multivectors (๐ฝ)๐ indicating the com-mutative and anticommutative parts will all have zeros from ๐ to ๐ and thereforeform a strictly triangular matrix.
Lemma 6.9. Let a set of functions ๐น = {๐1(๐,๐), . . . , ๐๐(๐,๐)} fulfil (6.1) and belinear with respect to ๐. Further let ๐ฝ โ {0, 1}๐ร๐ be a strictly lower triangularmatrix, that is associated column by column with a multi-index ๐ โ {0, 1}๐ by โ๐ =
1, . . . , ๐ : (โ๐
๐=1 ๐ฝ๐,๐) mod 2 = ๐๐, with (๐ฝ)๐ being its ๐th row, then
๐โ๐=1
๐โ๐๐(๐+๐,๐) =โ
๐โ{0,1}๐
โ๐ฝโ{0,1}๐ร๐,โ
๐๐=1(๐ฝ)๐ mod 2=๐
๐โ๐=1
๐โ๐๐(๐,๐)
๐(๐ฝ)๐ (โโโโโโโโโโโ๐1,...,๐๐,0,...,0)
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐)
(6.8)or alternatively with strictly upper triangular matrices ๐ฝ :
๐โ๐=1
๐โ๐๐(๐+๐,๐) =โ
๐โ{0,1}๐
โ๐ฝโ{0,1}๐ร๐,โ๐
๐=1(๐ฝ)๐ mod 2=๐
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐)๐โ
๐=1
๐โ๐๐(๐,๐)
๐(๐ฝ)๐ (โโโโโโโโโโโ0,...,0,๐๐,...,๐๐)
.
(6.9)
We do not explicitly indicate the dependence of the partition on ๐ as in Corol-lary 6.2, because the functions in the exponents already contain this dependence.Please note that the decomposition is pointwise.
8. A General Geometric Fourier Transform 169
Table 2. Theorem 6.5 (Right products) applied to the GFTs of thefirst example, enumerated in the same order.Notations: on the LHS โฑ๐น1,๐น2 = โฑ๐น1,๐น2(๐จ)(๐), on the RHS โฑ๐น โฒ
1,๐นโฒ2=
โฑ๐น โฒ1,๐น
โฒ2(๐ฉ)(๐)
GFT ๐จ(๐) = ๐ฉ(๐)๐ช
1. Clif. ๐ = 2 (mod 4) โฑ๐1 = โฑ๐1๐ช๐0(๐) + โฑโ๐1๐ช๐1(๐)
Clif. ๐ = 3 (mod 4) โฑ๐1 = โฑ๐1๐ช
2. Bulow โฑ๐1,...,๐๐
=โ
๐โ{0,1}๐ โฑ(โ1)๐1๐1,...,(โ1)๐๐๐๐๐ช๐๐(โโโโโโ๐1,...,๐๐)
3. Quaternionic โฑ๐1,๐2 = โฑ๐1,๐2๐ช๐0(๐) + โฑ๐1,โ๐2๐ช๐1(๐)
4. Spacetime โฑ๐1,๐2 = โฑ๐1,๐2๐ช๐0(๐4๐4) + โฑ๐1,โ๐2๐ช๐1(๐4๐4)
5. Colour Image โฑ๐1,๐2,๐3,๐4 = โฑ๐1,๐2,๐3,๐4๐ช๐00(โโโโ๐ฉ,๐๐ฉ)
+โฑ๐1,๐2,โ๐3,๐4๐ช๐10(โโโโ๐ฉ,๐๐ฉ)
+โฑ๐1,๐2,๐3,โ๐4๐ช๐01(โโโโ๐ฉ,๐๐ฉ)
+โฑ๐1,๐2,โ๐3,โ๐4๐ช๐11(โโโโ๐ฉ,๐๐ฉ)
6. Cylindrical โฑ๐1 = โฑ๐1๐ช
Proof. We will only prove the first assertion. The second one follows analogouslyby applying Corollary 6.2 the other way around.
๐โ๐=1
๐โ๐๐(๐+๐,๐) ๐น lin.=
๐โ๐=1
๐โ๐๐(๐,๐)โ๐๐(๐,๐) (6.10)
Lem. 2.2=
๐โ๐=1
๐โ๐๐(๐,๐)๐โ๐๐(๐,๐) (6.11)
= ๐โ๐1(๐,๐)๐โ๐1(๐,๐)๐โ
๐=2
๐โ๐๐(๐,๐)๐โ๐๐(๐,๐) (6.12)
Cor. 6.2= ๐โ๐1(๐,๐)(๐
โ๐2(๐,๐)๐0(๐1)
๐โ๐1(๐,๐)๐โ๐2(๐,๐)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐๐1(๐,๐)๐โ๐2(๐,๐))
๐โ๐=3
๐โ๐๐(๐,๐)๐โ๐๐(๐,๐). (6.13)
Now we use Corollary 6.2 to step by step rearrange the order of the product.
Cor. 6.2= ๐โ๐1(๐,๐)
(๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐00(โโโโ๐1,๐2)
๐โ๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
170 R. Bujack, G. Scheuermann and E. Hitzer
+ ๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐01(โโโโ๐1,๐2)
๐โ๐1(๐,๐)๐๐2(๐,๐)๐โ๐3(๐,๐)
+ ๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐10(โโโโ๐1,๐2)
๐๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
+ ๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐11(โโโโ๐1,๐2)
๐๐1(๐,๐)๐๐2(๐,๐)๐โ๐3(๐,๐)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐00(โโโโ๐1,๐2)
๐๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐01(โโโโ๐1,๐2)
๐๐1(๐,๐)๐๐2(๐,๐)๐โ๐3(๐,๐)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐10(โโโโ๐1,๐2)
๐โ๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐11(โโโโ๐1,๐2)
๐โ๐1(๐,๐)๐๐2(๐,๐)๐โ๐3(๐,๐)
)๐โ
๐=4
๐โ๐๐(๐,๐)๐โ๐๐(๐,๐). (6.14)
There are only 2๐ฟ ways of distributing the signs of ๐ฟ exponents, so some of thesummands can be combined.
= ๐โ๐1(๐,๐)
((๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐00(โโโโ๐1,๐2)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐10(โโโโ๐1,๐2)
)๐โ๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
+(๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐01(โโโโ๐1,๐2)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐11(โโโโ๐1,๐2)
)๐โ๐1(๐,๐)๐๐2(๐,๐)๐โ๐3(๐,๐)
+(๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐10(โโโโ๐1,๐2)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐00(โโโโ๐1,๐2)
)๐๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
+(๐โ๐2(๐,๐)๐0(๐1)
๐โ๐3(๐,๐)
๐11(โโโโ๐1,๐2)
+ ๐โ๐2(๐,๐)๐1(๐1)
๐โ๐3(๐,๐)
๐01(โโโโ๐1,๐2)
)๐๐1(๐,๐)
๐๐2(๐,๐)๐โ๐3(๐,๐)
) ๐โ๐=4
๐โ๐๐(๐,๐)๐โ๐๐(๐,๐). (6.15)
To get a compact notation we expand all multi-indices by adding zeros untilthey have the same length. Note that the last non-zero argument in terms like
๐000(โโโโโ๐1, 0, 0) always coincides with the exponent of the corresponding factor. Be-
cause of that it will always commute and could also be replaced by a zero.
= ๐โ๐1(๐,๐)
๐000(โโโโ๐1,0,0)((
๐โ๐2(๐,๐)
๐000(โโโโโ๐1,๐2,0)
๐โ๐3(๐,๐)
๐000(โโโโโโ๐1,๐2,๐3)
+ ๐โ๐2(๐,๐)
๐100(โโโโโ๐1,๐2,0)
๐โ๐3(๐,๐)
๐100(โโโโโโ๐1,๐2,๐3)
)๐โ๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
8. A General Geometric Fourier Transform 171
+(๐โ๐2(๐,๐)
๐000(โโโโโ๐1,๐2,0)
๐โ๐3(๐,๐)
๐010(โโโโโโ๐1,๐2,๐3)
+ ๐โ๐2(๐,๐)
๐100(โโโโโ๐1,๐2,0)
๐โ๐3(๐,๐)
๐110(โโโโโโ๐1,๐2,๐3)
)๐โ๐1(๐,๐)๐๐2(๐,๐)๐โ๐3(๐,๐)
+(๐โ๐2(๐,๐)
๐000(โโโโโ๐1,๐2,0)
๐โ๐3(๐,๐)
๐100(โโโโโโ๐1,๐2,๐3)
+ ๐โ๐2(๐,๐)
๐100(โโโโโ๐1,๐2,0)
๐โ๐3(๐,๐)
๐000(โโโโโโ๐1,๐2,๐3)
)๐๐1(๐,๐)๐โ๐2(๐,๐)๐โ๐3(๐,๐)
+(๐โ๐2(๐,๐)
๐000(โโโโโ๐1,๐2,0)
๐โ๐3(๐,๐)
๐110(โโโโโโ๐1,๐2,๐3)
+ ๐โ๐2(๐,๐)
๐100(โโโโโ๐1,๐2,0)
๐๐3(๐,๐)
๐010(โโโโโโ๐1,๐2,๐3)
)๐๐1(๐,๐)๐๐2(๐,๐)๐โ๐3(๐,๐)
) ๐โ๐=4
๐โ๐๐(๐,๐)๐โ๐๐(๐,๐) (6.16)
For ๐ฟ = 3 we look at all strictly lower triangular matrices ๐ฝ โ {0, 1}๐ฟร๐ฟ with theproperty
โ๐ = 1, . . . , ๐ฟ :
(๐ฟโ
๐=1
(๐ฝ)๐,๐
)mod 2 = ๐๐. (6.17)
That means the ๐th row (๐ฝ)๐ of ๐ฝ contains a multi-index (๐ฝ)๐ โ {0, 1}๐ฟ, with thelast ๐ฟ โ ๐ โ 1 entries being zero and the ๐th column sum being even when ๐๐ = 0and odd when ๐๐ = 1. For example, the first multi-index is ๐ = (0, 0, 0). There areonly two different strictly lower triangular matrices that have columns summingup to even numbers:
๐ฝ =
โโ0 0 00 0 00 0 0
โโ and ๐ฝ =
โโ0 0 01 0 01 0 0
โโ . (6.18)
The first row of each contains the multi-index that belongs to ๐โ๐1(๐,๐), the sec-ond one belongs to ๐โ๐2(๐,๐) and so on. So the summands with exactly thesemulti-indices are the ones assigned to the product of exponentials whose signs areinvariant during the reordering. With this notation and all ๐ฝ โ {0, 1}3ร3 thatsatisfy the property (6.17) we can write
๐โ๐=1
๐โ๐๐(๐+๐,๐) =โ
๐โ{0,1}3
โ๐ฝ
3โ๐=1
๐โ๐๐(๐,๐)
๐(๐ฝ)๐ (โโโโโโโโโโโ๐1,...,๐๐,0,...,0)
3โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐)
๐โ๐=4
๐โ๐๐(๐,๐)๐โ๐๐(๐,๐).
(6.19)
Using mathematical induction with matrices ๐ฝ โ {0, 1}๐ฟร๐ฟ as introduced abovefor growing ๐ฟ and Corollary 6.2 repeatedly until we reach ๐ฟ = ๐ we get
=โ
๐โ{0,1}๐
โ๐ฝ
๐โ๐=1
๐โ๐๐(๐,๐)
๐(๐ฝ)๐ (โโโโโโโโโโโ๐1,...,๐๐,0,...,0)
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐). (6.20)
โก
172 R. Bujack, G. Scheuermann and E. Hitzer
Remark 6.10. The number of summands actually appearing is usually much smallerthan in Theorem 6.11. It is determined by the number of distinct strictly lower(upper) triangular matrices ๐ฝ with entries being either zero or one, namely:
2๐(๐โ1)
2 . (6.21)
Theorem 6.11 (Shift). Let ๐จ(๐) = ๐ฉ(๐ โ ๐0) be multivector fields, ๐น1, ๐น2, lin-ear with respect to ๐, and let ๐ โ {0, 1}๐,๐ โ {0, 1}(๐โ๐) be multi-indices, and๐น1(๐), ๐น2(๐) be as introduced in Definition 6.4, then a separable GFT suffices
โฑ๐น1,๐น2(๐จ)(๐) =โ๐,๐
โ๐ฝ,๐พ
๐โ๐=1
๐โ๐๐(๐0,๐)
๐(๐ฝ)๐ (โโโโโโโโโโโ๐1,...,๐๐,0,...,0)
โฑ๐น1(๐),๐น2(๐)(๐ฉ)(๐)
๐โ๐=๐+1
๐โ๐๐(๐0,๐)
๐(๐พ)๐โ๐ (โโโโโโโโโโโ0,...,0,๐๐,...,๐๐)
,
(6.22)
where ๐ฝ โ {0, 1}๐ร๐ and ๐พ โ {0, 1}(๐โ๐)ร(๐โ๐) are the strictly lower, respectivelyupper, triangular matrices with rows (๐ฝ)๐, (๐พ)๐โ๐ summing up to (
โ๐๐=1(๐ฝ)๐) mod
2 = ๐ respectively(โ๐
๐=๐+1(๐พ)๐โ๐)mod 2 = ๐ as in Lemma 6.9.
Proof. First we rewrite the transformed function in terms of ๐ฉ(๐) using a changeof coordinates.
โฑ๐น1,๐น2(๐จ)(๐) =
โซโ๐
๐โ๐=1
๐โ๐๐(๐,๐)๐จ(๐)๐โ
๐=๐+1
๐โ๐๐(๐,๐) d๐๐
=
โซโ๐
๐โ๐=1
๐โ๐๐(๐,๐)๐ฉ(๐โ ๐0)
๐โ๐=๐+1
๐โ๐๐(๐,๐) d๐๐
๐=๐โ๐0=
โซโ๐
๐โ๐=1
๐โ๐๐(๐+๐0,๐)๐ฉ(๐)
๐โ๐=๐+1
๐โ๐๐(๐+๐0,๐) d๐๐
(6.23)
Now we separate and sort the factors using Lemma 6.9.
Lem. 6.9=
โซโ๐
โ๐โ{0,1}๐
โ๐ฝโ{0,1}๐ร๐
โ(๐ฝ)๐ mod 2=๐
๐โ๐=1
๐โ๐๐(๐0,๐)
๐(๐ฝ)๐ (โโโโโโโโโโโ๐1,...,๐๐,0,...,0)
๐โ๐=1
๐โ(โ1)๐๐๐๐(๐,๐)๐ฉ(๐)โ๐โ{0,1}(๐โ๐)
โ๐พโ{0,1}(๐โ๐)ร(๐โ๐)โ
(๐พ)๐ mod 2=๐
๐โ๐=๐+1
๐โ(โ1)๐๐โ๐๐๐(๐,๐)๐โ
๐=๐+1
๐โ๐๐(๐0,๐)
๐(๐พ)๐โ๐ (โโโโโโโโโโโ0,...,0,๐๐,...,๐๐)
d๐๐
8. A General Geometric Fourier Transform 173
=โ๐,๐
โ๐ฝ,๐พ
๐โ๐=1
๐โ๐๐(๐0,๐)
๐(๐ฝ)๐ (โโโโโโโโโโโ๐1,...,๐๐,0,...,0)
(6.24)
โฑ๐น1(๐),๐น2(๐)(๐ฉ)(๐)
๐โ๐=๐+1
๐โ๐๐(๐0,๐)
๐(๐พ)๐โ๐ (โโโโโโโโโโโ0,...,0,๐๐,...,๐๐)
โก
Corollary 6.12 (Shift). Let ๐จ(๐) = ๐ฉ(๐ โ ๐0) be multivector fields, ๐น1 and ๐น2
each consisting of mutually commutative functions3 being linear with respect to ๐,then the GFT obeys
โฑ๐น1,๐น2(๐จ)(๐) =
๐โ๐=1
๐โ๐๐(๐0,๐)โฑ๐น1,๐น2(๐ฉ)(๐)
๐โ๐=๐+1
๐โ๐๐(๐0,๐). (6.25)
Remark 6.13. For sets ๐น1, ๐น2 that each consist of less than two functions thecondition of Corollary 6.12 is necessarily satisfied, compare, e.g., the CliffordโFourier transform, the quaternionic transform or the spacetime Fourier transformlisted in the preceeding examples.
The specific forms taken by our standard examples are summarized in Table 3.As expected they are often shorter than what could be expected from Remark 6.10.
7. Conclusions and Outlook
For multivector fields over โ๐,๐ with values in any geometric algebra ๐บ๐,๐ we havesuccessfully defined a general geometric Fourier transform. It covers all popularFourier transforms from current literature in the introductory example. Its exis-tence, independent of the specific choice of functions ๐น1, ๐น2, can be proved forall integrable multivector fields, see Theorem 3.1. Theorem 3.2 shows that ourgeometric Fourier transform is generally linear over the field of real numbers. Alltransforms from the reference example consist of bilinear ๐น1 and ๐น2. We provedthat this property is sufficient to ensure the scaling property of Theorem 4.1.
If a general geometric Fourier transform is separable as introduced in Defi-nition 6.1, then Theorem 6.5 (Left and right products) guarantees that constantfactors can be separated from the vector field to be transformed. As a consequencegeneral linearity is achieved by choosing ๐น1, ๐น2 with values in the centre of thegeometric algebra ๐ถโ๐,๐, compare Corollary 6.7. All examples except for the cylin-drical Fourier transform [6] satisfy this claim.
Under the condition of linearity with respect to the first argument of thefunctions of the sets ๐น1 and ๐น2 additionally to the separability property just men-tioned, we have also proved a shift property (Theorem 6.11).
In future publications we are going to state the necessary constraints fora generalized convolution theorem, invertibility, derivation theorem and we willexamine how simplifications can be achieved based on symmetry properties of the
3Cross commutativity between ๐น1 and ๐น2 is not necessary.
174 R. Bujack, G. Scheuermann and E. Hitzer
Table 3. Theorem 6.11 (Shift) applied to the GFTs of the first exam-ple, enumerated in the same order.Notations: on the LHS โฑ๐น1,๐น2 = โฑ๐น1,๐น2(๐จ)(๐), on the RHS โฑ๐น โฒ
1,๐นโฒ2=
โฑ๐น โฒ1,๐น
โฒ2(๐ฉ)(๐). In the second row ๐พ represents all strictly upper
triangular matrices โ {0, 1}๐ร๐ with rows (๐พ)๐โ๐ summing up to(โ๐๐=๐+1(๐พ)๐โ๐
)mod 2 = ๐. The simplified shape of the colour im-
age FT results from the commutativity of ๐ฉ and ๐๐ฉ and application ofLemma 2.2.
GFT ๐จ(๐) = ๐ฉ(๐โ ๐0)
1. Clifford โฑ๐1 = โฑ๐1๐โ2๐๐๐0โ ๐
2. Bulow โฑ๐1,...,๐๐ =โ
๐โ{0,1}๐โ
๐พ โฑ(โ1)๐1๐1,...,(โ1)๐๐๐๐โ๐๐=1 ๐
โ2๐๐ฅ0๐๐ข๐
๐(๐พ)๐ (โโโโโโโโโโโ0,...,0,๐๐,...,๐๐)
3. Quaternionic โฑ๐1,๐2 = ๐โ2๐๐๐ฅ01๐ข1โฑ๐1,๐2๐โ2๐๐๐ฅ02๐ข2
4. Spacetime โฑ๐1,๐2 = ๐โ๐4๐ฅ04๐ข4โฑ๐1,๐2๐โ๐4๐4๐4(๐ฅ1๐ข1+๐ฅ2๐ข2+๐ฅ3๐ข3)
5. Colour Image โฑ๐1,๐2,๐3,๐4 = ๐โ12 (๐ฅ01๐ข1+๐ฅ02๐ข2)(๐ฉ+๐๐ฉ)โฑ๐1,๐2,๐3,๐4
๐12 (๐ฅ01๐ข1+๐ฅ02๐ข2)(๐ฉ+๐๐ฉ)
6. Cyl. ๐ = 2 โฑ๐1 = ๐๐0โง๐โฑ๐1
Cyl. ๐ โ= 2 -
multivector fields to be transformed. We will also construct generalized geometricFourier transforms in a broad sense from combinations of the ones introduced inthis chapter and from decomposition into their sine and cosine parts which willalso cover the vector and bivector Fourier transforms of [9]. It would further beof interest to extend our approach to Fourier transforms defined on spheres orother non-Euclidean manifolds, to functions in the Schwartz space and to square-integrable functions.
References
[1] T. Batard, M. Berthier, and C. Saint-Jean. Clifford Fourier transform for color imageprocessing. In Bayro-Corrochano and Scheuermann [2], pages 135โ162.
[2] E.J. Bayro-Corrochano and G. Scheuermann, editors. Geometric Algebra Computingin Engineering and Computer Science. Springer, London, 2010.
[3] F. Brackx, N. De Schepper, and F. Sommen. The CliffordโFourier transform. Journalof Fourier Analysis and Applications, 11(6):669โ681, 2005.
[4] F. Brackx, N. De Schepper, and F. Sommen. The two-dimensional CliffordโFouriertransform. Journal of Mathematical Imaging and Vision, 26(1):5โ18, 2006.
8. A General Geometric Fourier Transform 175
[5] F. Brackx, N. De Schepper, and F. Sommen. The CliffordโFourier integral kernel ineven dimensional Euclidean space. Journal of Mathematical Analysis and Applica-tions, 365(2):718โ728, 2010.
[6] F. Brackx, N. De Schepper, and F. Sommen. The Cylindrical Fourier Transform. InBayro-Corrochano and Scheuermann [2], pages 107โ119.
[7] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[8] W.K. Clifford. Applications of Grassmannโs extensive algebra. American Journal ofMathematics, 1(4):350โ358, 1878.
[9] H. De Bie and F. Sommen. Vector and bivector Fourier transforms in Clifford anal-ysis. In K. Guerlebeck and C. Koenke, editors, 18th International Conference on theApplication of Computer Science and Mathematics in Architecture and Civil Engi-neering, page 11, 2009.
[10] J. Ebling. Visualization and Analysis of Flow Fields using Clifford Convolution. PhDthesis, University of Leipzig, Germany, 2006.
[11] T.A. Ell. Quaternion-Fourier transforms for analysis of 2-dimensional linear time-invariant partial-differential systems. In Proceedings of the 32nd Conference on De-cision and Control, pages 1830โ1841, San Antonio, Texas, USA, 15โ17 December1993. IEEE Control Systems Society.
[12] T.A. Ell and S.J. Sangwine. Hypercomplex Fourier transforms of color images. IEEETransactions on Image Processing, 16(1):22โ35, Jan. 2007.
[13] M. Felsberg. Low-Level Image Processing with the Structure Multivector. PhD thesis,Christian-Albrechts-Universitat, Institut fur Informatik und Praktische Mathematik,Kiel, 2002.
[14] D. Hestenes and G. Sobczyk. Clifford Algebra to Geometric Calculus. D. ReidelPublishing Group, Dordrecht, Netherlands, 1984.
[15] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[16] E. Hitzer and R. Ablamowicz. Geometric roots of โ1 in Clifford algebras ๐ถโ๐,๐ with๐ + ๐ โค 4. Advances in Applied Clifford Algebras, 21(1):121โ144, 2010. Publishedonline 13 July 2010.
[17] E. Hitzer, J. Helmstetter, and R. Ablamowicz. Square roots of โ1 in real Cliffordalgebras. In K. Gurlebeck, editor, 9th International Conference on Clifford Algebrasand their Applications, Weimar, Germany, 15โ20 July 2011. 12 pp.
[18] E.M.S. Hitzer and B. Mawardi. Clifford Fourier transform on multivector fields anduncertainty principles for dimensions ๐ = 2(mod 4) and ๐ = 3(mod 4). Advances inApplied Clifford Algebras, 18(3-4):715โ736, 2008.
[19] B. Jancewicz. Trivector Fourier transformation and electromagnetic field. Journalof Mathematical Physics, 31(8):1847โ1852, 1990.
[20] S.J. Sangwine and T.A. Ell. The discrete Fourier transform of a colour image. In J.M.Blackledge and M.J. Turner, editors, Image Processing II Mathematical Methods, Al-gorithms and Applications, pages 430โ441, Chichester, 2000. Horwood Publishing forInstitute of Mathematics and its Applications. Proceedings Second IMA Conferenceon Image Processing, De Montfort University, Leicester, UK, September 1998.
176 R. Bujack, G. Scheuermann and E. Hitzer
[21] F. Sommen. Hypercomplex Fourier and Laplace transforms I. Illinois Journal ofMathematics, 26(2):332โ352, 1982.
Roxana BujackUniversitat LeipzigInstitut fur InformatikJohannisgasse 26D-04103 Leipzig, Germanye-mail: [email protected]
Gerik ScheuermannUniversitat LeipzigInstitut fur InformatikJohannisgasse 26D-04103 Leipzig, Germanye-mail: [email protected]
Eckhard HitzerCollege of Liberal Arts, Department of Material Science,International Christian University,181-8585 Tokyo, Japane-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 177โ195cโ 2013 Springer Basel
9 CliffordโFourier Transform andSpinor Representation of Images
Thomas Batard and Michel Berthier
Abstract. We propose in this chapter to introduce a spinor representation forimages based on the work of T. Friedrich. This spinor representation gener-alizes the usual Weierstrass representation of minimal surfaces (i.e., surfaceswith constant mean curvature equal to zero) to arbitrary surfaces (immersedin โ3). We investigate applications to image processing focusing on segmen-tation and CliffordโFourier analysis. All these applications involve sections ofthe spinor bundle of image graphs, that is spinor fields, satisfying the so-calledDirac equation.
Mathematics Subject Classification (2010). Primary 68U10, 53C27; secondary53A05, 43A32.
Keywords. Image processing, spin geometry, CliffordโFourier transform.
1. Introduction
The idea of this chapter is to perform grey-level image processing using the geo-metric information given by the Gauss map variations of image graphs. While itis well known that one can parameterize the Gauss map of a minimal surface bya meromorphic function (see below), it is a much more recent result (see [5]) thatsuch a parametrization can be extended to arbitrary surfaces of โ3 when dealingwith spin geometry.
Let us first recall that a minimal surface ฮฃ immersed in โ3, that is a surfacewith constant mean curvature equal to zero, can be described with one holomor-phic function ๐ and one meromorphic function ๐ such that the product ๐๐2 isholomorphic. This is the so-called Weierstrass representation of ฮฃ (see [6] or [8]for details). The function ๐ is nothing else but the composition of the Gauss mapof ฮฃ with the stereographic projection from the unit sphere to the complex plane.
This work was partially supported by ONR Grant N00014-09-1-0493.
178 T. Batard and M. Berthier
The main result of T. Friedrich in [5] states that there is a one-to-one cor-respondance between spinor fields ๐โ of constant length on a Riemannian surface(ฮฃ, ๐) and satisfying
๐ท๐โ = ๐ป๐โ (1.1)
where ๐ท is a Dirac operator in one hand, and isometric immersions of ฮฃ in โ3
with mean curvature equal to๐ป , on the other hand. The Weierstrass representationappears to be the particular case corresponding to ๐ป โก 0.
Let us describe now the method introduced in the following. Let
๐ : ฮฉ โ โ2 โโ โ3
(๐ฅ, ๐ฆ) ๏ฟฝโโ (๐ฅ, ๐ฆ, ๐ผ(๐ฅ, ๐ฆ))(1.2)
be the immersion in the three-dimensional Euclidean space of a grey-level image๐ผ defined on a domain ฮฉ of โ2. The first step (see ยง 2) consists in computing thespinor field ๐โ that describes the image surface ฮฃ. We follow here the paper of T.Friedrich [5]: ๐โ is obtained from the restriction to the surface ฮฃ of a parallel spinor๐ on โ3. The computation of ๐โ requires us to deal with irreducible representationsof the complex Clifford algebra ๐ถโ3,0 โ โ and with the generalized Weierstrassrepresentation of ฮฃ based on period forms. In practice, ๐โ is given by a field ofelements of โ2.
As said before, the spinor field ๐โ characterizes the geometry of the surfaceฮฃ immersed in โ3 by the parametrization (1.2). In the same way that the normalof a minimal surface is parameterized by the meromorphic function ๐, the normalof the surface ฮฃ is parameterized by the spinor field ๐โ. The latter explains howthe tangent plane to ฮฃ varies in the ambient space.
There are many reasons to believe that such a generalized Weierstrass para-metrization may reveal itself to be an efficient tool in the context of image pro-cessing:
1. The field ๐โ of elements of โ2 (see (2.26)) encodes the Riemannian structureof the surface ฮฃ in a very tractable way (although the definition of ๐โ mayappear quite complicated).
2. The geometrical methods based on the study of the so-called structure tensorinvolve only the eigenvalues of the structure tensor, that means in some sensethe values of the first fundamental form of the surface. The spinor field ๐โ
contains both intrinsic and extrinsic information. Studying the variations of๐โ allows us to get not only information about the variations (derivative) ofthe first fundamental form, but also about the geometric embedding of thesurface ฮฃ and in particular about the mean curvature.
3. We are dealing here with first-order instead of zero-order geometric variationsof ฮฃ. As shown later, this appears to be more relevant by taking into accountboth edges and textures.
4. As will be detailed in the sequel, the spinor field ๐โ can be decomposed as aseries of basic spinor fields using a suitable CliffordโFourier transform. Thisseries corresponds to a harmonic decomposition of the surface ฮฃ adapted to
9. CliffordโFourier Transform 179
the Riemannian geometry. This is in fact the main novelty of this chaptersince the usual techniques of Fourier analysis do not involve geometric data.
5. One can envisage the possibility of performing diffusion in this context. Theusual Laplace Beltrami operator can be replaced by the squared Atiyah SingerDirac operator [7] (the Atiyah Singer Dirac operator acting as an ellipticoperator of order one on spinor fields).
To illustrate some of these ideas, we investigate rapidly in ยง 3 applications tosegmentation and more precisely to edge and texture detection. As stated before,the basic idea is to replace the usual order-one structure tensor by an order-twostructure tensor called the spinor tensor obtained from the derivative of the spinorfield ๐โ. This spinor tensor measures the variations of the unit normal of theimage surface. Experiments show that this approach is particularly well adaptedto texture detection.
We define in ยง 4 the CliffordโFourier transform of a spinor field. For this,we follow the approach of [3] that relies on a spin generalization of the usualnotion of group character. We are led to compute the group morphisms fromโค/๐โครโค/๐โค to Spin(3). Since this last group acts on the sections of the spinorbundle, a CliffordโFourier transform can be defined by averaging this action. Oneof the key ideas here is to split the spinor bundle of the surface according to theClifford multiplication by the bivector coding the tangent plane to the surface.This has two advantages: the first one is to involve the geometry in the process,the second one is to reduce the computation of the CliffordโFourier transform totwo usual complex Fourier transforms. It is important to notice that although theFourier transform we propose is, as usual, a global transformation on the image, theway it is computed takes into account local geometric data. We finally introducethe harmonic decomposition mentioned above and show some results of filteringon standard images.
The reader will find in Appendix A the mathematical definitions and resultsused throughout the text.
2. Spinor Representation of Images
This section is devoted to the explicit computation of the spinor field ๐โ of a givensurface immersed in Euclidean space. It is obtained as the restriction of a constantspinor field of โ3 the components of which are determined using period forms.
2.1. Spinors and Graphs
Let ๐ผ : ฮฉ โโ โ be a differentiable function defined on a domain ฮฉ of โ2. Weconsider the surface ฮฃ immersed in โ3 by the parametrization:
๐(๐ฅ, ๐ฆ) = (๐ฅ, ๐ฆ, ๐ผ(๐ฅ, ๐ฆ)). (2.1)
Also, let ๐ be the metric on ฮฃ induced by the Euclidean metric of โ3. The cou-ple (ฮฃ, ๐) is a Riemannian surface of global chart (ฮฉ, ๐). We denote by ๐ theRiemannian manifold (โ3, โฅ โฅ2) and by (๐ง1, ๐ง2, ๐) an orthonormal frame field of
180 T. Batard and M. Berthier
๐ with (๐ง1, ๐ง2) an orthonormal frame field on ฮฃ, and by ๐ the global unit fieldnormal to ฮฃ. One can choose (๐ง1, ๐ง2, ๐) with the following matrix representationโโโโโโโโโโโโโ
๐ผ๐ฅโ(๐ผ2
๐ฅ + ๐ผ2๐ฆ )(๐ผ
2๐ฅ + ๐ผ2
๐ฆ + 1)
โ๐ผ๐ฆโ๐ผ2๐ฅ + ๐ผ2
๐ฆ
โ๐ผ๐ฅโ๐ผ2๐ฅ + ๐ผ2
๐ฆ + 1
๐ผ๐ฆโ(๐ผ2
๐ฅ + ๐ผ2๐ฆ )(๐ผ
2๐ฅ + ๐ผ2
๐ฆ + 1)
๐ผ๐ฅโ๐ผ2๐ฅ + ๐ผ2
๐ฆ
โ๐ผ๐ฆโ๐ผ2๐ฅ + ๐ผ2
๐ฆ + 1
๐ผ2๐ฅ + ๐ผ2
๐ฆโ(๐ผ2
๐ฅ + ๐ผ2๐ฆ )(๐ผ
2๐ฅ + ๐ผ2
๐ฆ + 1)0
1โ๐ผ2๐ฅ + ๐ผ2
๐ฆ + 1
โโโโโโโโโโโโโ . (2.2)
Note that ๐ง1 and ๐ง2 are not defined when ๐ผ๐ฅ = ๐ผ๐ฆ = 0. This has no consequencein the sequel since we deal only with the normal ๐.
Following [5] the surface ฮฃ can be represented by a spinor field ๐โ withconstant length satisfying the Dirac equation:
๐ท๐โ = ๐ป๐โ (2.3)
where ๐ป denotes the mean curvature of ฮฃ. We recall here the basic idea (seeAppendix A for notations and definitions). Let ๐ be a parallel spinor field of ๐ ,i.e., satisfying
โ๐๐ ๐ = 0 (2.4)
for all vector fields ๐ on ๐ . Let also ๐ be the restriction ๐โฃฮฃ of ๐ to ฮฃ. The spinorfield ๐ decomposes into
๐ = ๐+ + ๐โ (2.5)
with
๐+ =1
2(๐+ ๐๐ โ ๐) ๐โ =
1
2(๐โ ๐๐ โ ๐) (2.6)
and satisfies
๐ท๐ = โ๐ป โ ๐ โ ๐. (2.7)
This last equation reads
๐ท(๐+ + ๐โ) = โ๐ป โ ๐ โ (๐+ + ๐โ) (2.8)
and implies
๐ท๐+ = โ๐๐ป๐โ ๐ท๐โ = ๐๐ป๐+. (2.9)
If we set ๐โ = ๐+ โ ๐๐โ then ๐ท๐โ = ๐ป๐โ and ๐โ is of constant length.
9. CliffordโFourier Transform 181
Proposition 2.1. The spinor fields ๐+, ๐โ and ๐โ are given by
๐+ =1
2
โโโโโโโโโ
โโ1โ ๐ผ๐ฆโ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ ๐ข+
โโ ๐ผ๐ฅ โ ๐โ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ ๐ฃ
โโ1 +๐ผ๐ฆโ
1 + ๐ผ2๐ฅ + ๐ผ2
๐ฆ
โโ ๐ฃ +
โโ ๐ผ๐ฅ + ๐โ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ ๐ข
โโโโโโโโโ (2.10)
๐โ =1
2
โโโโโโโโโ
โโ1 +๐ผ๐ฆโ
1 + ๐ผ2๐ฅ + ๐ผ2
๐ฆ
โโ ๐ขโโโ ๐ผ๐ฅ โ ๐โ
1 + ๐ผ2๐ฅ + ๐ผ2
๐ฆ
โโ ๐ฃ
โโ1โ ๐ผ๐ฆโ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ ๐ฃ โโโ ๐ผ๐ฅ + ๐โ
1 + ๐ผ2๐ฅ + ๐ผ2
๐ฆ
โโ ๐ข
โโโโโโโโโ (2.11)
and
๐โ =1
2(1โ ๐)
โโโโโโโโโ
โโ1โ ๐๐ผ๐ฆโ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ ๐ข+
โโ 1 + ๐๐ผ๐ฅโ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ ๐ฃ
โโ1 +๐๐ผ๐ฆโ
1 + ๐ผ2๐ฅ + ๐ผ2
๐ฆ
โโ ๐ฃ +
โโ ๐๐ผ๐ฅ โ 1โ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ ๐ข
โโโโโโโโโ (2.12)
where ๐ข and ๐ฃ are (constant) complex numbers.
Proof. Since ๐ is a parallel spinor field on ๐ , ๐ = (๐ข, ๐ฃ) where ๐ข and ๐ฃ are two(constant) complex numbers. Let ๐2 be the irreducible complex representation ofโ๐(3) described in Appendix A.1. Recall that
๐ =1
ฮ(โ๐ผ๐ฅ๐1 โ ๐ผ๐ฆ๐2 + ๐3) (2.13)
where ฮ =โ
๐ผ2๐ฅ + ๐ผ2
๐ฆ + 1, so that
๐2(๐) = โ๐ผ๐ฅฮ
(0 ๐๐ 0
)โ ๐ผ๐ฆ
ฮ
( โ๐ 00 ๐
)+
1
ฮ
(0 โ11 0
). (2.14)
By definition:
๐ โ ๐ = ๐2(๐)
(๐ข๐ฃ
). (2.15)
Simple computations lead now to the result. โก
The next step consists in computing the components (๐ข, ๐ฃ) of the constant field ๐.This is done by considering a quaternionic structure on the spinor bundle ๐(ฮฃ) ofthe surface ฮฃ and period forms.
182 T. Batard and M. Berthier
2.2. Quaternionic Structure and Period Forms
Let ๐ผ be the complex structure on ๐(ฮฃ) given by the multiplication by ๐. A quater-nionic structure on ๐(ฮฃ) is a linear map ๐ฝ that satisfies ๐ฝ2 = โ๐ผ๐ and ๐ผ๐ฝ = โ๐ฝ๐ผ.In the sequel ๐ฝ is given by
๐ฝ
(๐1
๐2
)=
( โ๐2
๐1
). (2.16)
If we write ๐1 = ๐ผ1 + ๐๐ฝ1 and ๐2 = ๐ผ2 + ๐๐ฝ2, the corresponding quaternion isgiven by
๐1 + ๐2๐ = (๐ผ1 + ๐๐ฝ1) + (๐ผ2 + ๐๐ฝ2)๐ = ๐ผ1 + ๐๐ฝ1 + ๐ผ2๐ + ๐ฝ2๐ (2.17)
and
๐(๐1 + ๐2๐) = โ๐2 + ๐1๐, (2.18)
i.e., ๐ฝ is the left multiplication by ๐. Since
๐+(ฮฃ) =
{(๐1
๐2
), ๐1 =
๐ผ๐ฅ โ ๐
๐ผ๐ฆ +ฮ๐2
}(2.19)
and
๐+(ฮฃ) =
{(๐1
๐2
), ๐1 =
๐ผ๐ฅ โ ๐
๐ผ๐ฆ โฮ๐2
}(2.20)
then ๐ฝ๐+(ฮฃ) โ ๐โ(ฮฃ) and ๐ฝ๐โ(ฮฃ) โ ๐+(ฮฃ). We also denote by ๐ฝ the quater-nionic structure (obtained in the same way) on ๐(๐).
Let us consider ๐ = (๐ข, ๐ฃ) a constant spinor field on ๐ and ๐โ its restrictionon ฮฃ. Let also ๐ : โ3 โโ โ and ๐ : โ3 โโ โ be the functions defined by
๐(๐) = โโ(๐ โ ๐, ๐) (2.21)
and
๐(๐) = ๐(๐ โ ๐, ๐ฝ(๐)) (2.22)
where ( , ) denotes the Hermitian product. Using the representation ๐2, one cancheck that
๐ โ ๐ =
(โ๐๐2๐ข+ (๐๐1 โ๐3)๐ฃ
(๐๐1 +๐3)๐ข+ ๐๐2๐ฃ
)(2.23)
for ๐ = (๐1,๐2,๐3). The equations ๐(๐) = ๐1 and ๐(๐) = ๐2 + ๐๐3 areequivalent to:
โฃ๐ขโฃ2 = โฃ๐ฃโฃ2 , ๐ข๐ฃ = โ1
2(2.24)
and
๐ข๐ฃ = โ1
2, ๐ข2 + ๐ฃ2 = 1, ๐ข2 = ๐ฃ2. (2.25)
This implies ๐ข = ยฑ1/โ2 and ๐ฃ = โ๐ข.
9. CliffordโFourier Transform 183
Definition 2.2. The spinor representation of the image given by the parametrization(2.1) is defined by
๐โ =1
2โ2(1โ ๐)
โโโโโโโโโ
โโ1โ 1 + ๐( ๐ผ๐ฅ + ๐ผ๐ฆ)โ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ โโโ1 +
1 + ๐(โ๐ผ๐ฅ + ๐ผ๐ฆ)โ1 + ๐ผ2
๐ฅ + ๐ผ2๐ฆ
โโ
โโโโโโโโโ . (2.26)
This means that ๐ข = 1/โ2 and ๐ฃ = โ1/โ2 in the expression (2.12).
The two 1-forms
๐๐ (๐) = 2โ(๐ โ (๐โ)+, (๐โ)โ) = โโ(๐ โ ๐, ๐) (2.27)
๐๐(๐) = ๐(๐ โ (๐โ)+, ๐ฝ((๐โ)+)) + ๐(๐ โ (๐โ)โ, ๐ฝ((๐โ)โ))= ๐(๐ โ ๐, ๐ฝ(๐)) (2.28)
are exact and verify ๐(๐โฃฮฃ) = ๐๐ , ๐(๐โฃฮฃ) = ๐๐. The generalized Weierstrass para-metrization is actually given by the isometric immersion:โซ
(๐๐ , ๐๐) : ฮฃ โโ๐. (2.29)
2.3. Dirac Equation and Mean Curvature
We only mention here some results that can be used when dealing with diffusion.We do not go into further details since we will not treat this problem in the presentchapter. Let (ฮฃ, ๐) be an oriented two-dimensional Riemannian manifold and ๐ aspinor field without zeros solution of the Dirac equation ๐ท๐ = ๐๐. Then ๐ definesan isometric immersion
(ฮฃ, โฃ๐โฃ4๐) โโ โ3 (2.30)
with mean curvature ๐ป = ๐/โฃ๐โฃ2 (see [5]).
3. Spinors and Segmentation
The aim of this section is to introduce the spinor tensor corresponding to thevariations of the unit normal and to show its capability to detect both edges andtextures.
3.1. The Spinor Tensor
We propose here to deal with a second-order version of the classical approachof edge detection based on the so-called structure tensor (see [10]). Instead ofmeasuring edges from eigenvalues of the Riemannian metric, we focus here on the
184 T. Batard and M. Berthier
eigenvalues of the tensor obtained from the derivative of the spinor field ๐โ. Moreprecisely let
๐ =
(๐1
๐2
)(3.1)
be a section of the spinor bundle ๐(ฮฃ) given in an orthonormal frame, i.e., โฃ๐โฃ2 =
โฃ๐1โฃ2 + โฃ๐2โฃ2 and let ๐ = (๐1, ๐2) be a section of the tangent bundle ๐ (ฮฃ). Weconsider the connection โ on ๐(ฮฃ) given by the connection 1-form ๐ = 0. Thus
โ๐๐ =
โโโโ ๐1โ๐1
โ๐ฅ+๐2
โ๐1
โ๐ฆ
๐1โ๐2
โ๐ฅ+๐2
โ๐2
โ๐ฆ
โโโโ (3.2)
and
โฃโ๐๐โฃ2 = ๐21
โฃโฃโฃโฃโ๐1
โ๐ฅ
โฃโฃโฃโฃ2 + 2๐1๐2โ(โ๐1
โ๐ฅ
โ๐1
โ๐ฆ
)+๐2
2
โฃโฃโฃโฃโ๐1
โ๐ฆ
โฃโฃโฃโฃ2+๐2
1
โฃโฃโฃโฃโ๐2
โ๐ฅ
โฃโฃโฃโฃ2 + 2๐1๐2โ(โ๐2
โ๐ฅ
โ๐2
โ๐ฆ
)+๐2
2
โฃโฃโฃโฃโ๐2
โ๐ฆ
โฃโฃโฃโฃ2 . (3.3)
If we denote
๐บ๐ =
โโโโโโฃโฃโฃโฃโ๐1
โ๐ฅ
โฃโฃโฃโฃ2+ โฃโฃโฃโฃโ๐2
โ๐ฅ
โฃโฃโฃโฃ2 โ(โ๐1
โ๐ฅ
โ๐1
โ๐ฆ+
โ๐2
โ๐ฅ
โ๐2
โ๐ฆ
)โ(โ๐1
โ๐ฅ
โ๐1
โ๐ฆ+
โ๐2
โ๐ฅ
โ๐2
โ๐ฆ
) โฃโฃโฃโฃโ๐1
โ๐ฆ
โฃโฃโฃโฃ2+ โฃโฃโฃโฃโ๐2
โ๐ฆ
โฃโฃโฃโฃ2โโโโโ (3.4)
then
(๐1 ๐2)๐บ๐(๐1 ๐2)๐ = โฃโ๐๐โฃ2 . (3.5)
๐บ๐ is a field of real symmetric matrices.As in the case of the usual structure tensor (i.e., Di Zenzo tensor, see [10])
the optima of โฃโ๐๐โฃ2 under the constraint โฅ๐โฅ = 1 (for the Euclidean norm) aregiven by the field of eigenvalues of ๐บ๐. Applying the above formula to the spinor๐โ of Definition 2.2 leads to
๐บ๐โ =1
2(1 + ๐ผ2๐ฅ + ๐ผ2
๐ฆ )2
(๐บ11
๐โ ๐บ12๐โ
๐บ21๐โ ๐บ22
๐โ
)(3.6)
with
๐บ11๐โ = ๐ผ2
๐ฅ๐ฅ + ๐ผ2๐ฅ๐ฆ + ๐ผ2
๐ฅ๐ฅ๐ผ2๐ฆ + ๐ผ2
๐ฅ๐ฆ๐ผ2๐ฅ โ 2๐ผ๐ฅ๐ฅ๐ผ๐ฅ๐ฆ๐ผ๐ฅ๐ผ๐ฆ
๐บ22๐โ = ๐ผ2
๐ฆ๐ฆ + ๐ผ2๐ฅ๐ฆ + ๐ผ2
๐ฆ๐ฆ๐ผ2๐ฅ + ๐ผ2
๐ฅ๐ฆ๐ผ2๐ฆ โ 2๐ผ๐ฆ๐ฆ๐ผ๐ฅ๐ฆ๐ผ๐ฅ๐ผ๐ฆ
๐บ12๐โ = ๐ผ๐ฅ๐ฅ๐ผ๐ฅ๐ฆ + ๐ผ๐ฅ๐ฆ๐ผ๐ฆ๐ฆ + ๐ผ๐ฅ๐ฅ๐ผ๐ฅ๐ฆ๐ผ
2๐ฆ + ๐ผ๐ฅ๐ฆ๐ผ๐ฆ๐ฆ๐ผ
2๐ฅ โ ๐ผ2
๐ฅ๐ฆ๐ผ๐ฅ๐ผ๐ฆ โ ๐ผ๐ฅ๐ฅ๐ผ๐ฆ๐ฆ๐ผ๐ฅ๐ผ๐ฆ
๐บ21๐โ = ๐บ12
๐โ .
(3.7)
9. CliffordโFourier Transform 185
Definition 3.1. The tensor ๐บ๐โ is called the spinor tensor of the surface ฮฃ.
Note that as already mentioned this last tensor corresponds to the tensor involvedin the measure of the variations of the unit normal ๐ introduced in ยง 2.1. Indeed,we have
(๐1 ๐2)๐บ๐โ(๐1 ๐2)๐ = โฅ๐๐๐โฅ2. (3.8)
3.2. Experiments
We compare in Figure 1 the edge and texture detection methods based on theusual structure tensor (Figure 1(b) and 1(d)) and on the spinor tensor (Figure 1(e)and 1(f)).
The structure tensor only takes into account the first-order derivatives of thefunction ๐ผ. The subsequent segmentation method detects the strongest grey-levelvariations of the image. As a consequence, this method provides thick edges, ascan be observed.
The spinor tensor takes into account the second-order derivatives of the func-tion ๐ผ too. By definition, it measures the strongest variations of the unit normalto the surface parametrized by the graph of ๐ผ. We observe that this new approachprovides thinner edges than the first one. It appears also to be more relevant todetect textures.
4. Spinors and CliffordโFourier Transform
We first define a CliffordโFourier transform using spin characters that is groupmorphisms from โ2 to Spin(3). Then, we introduce a harmonic decomposition ofspinor fields and show some results of filtering applied to images.
4.1. CliffordโFourier Transform with Spin Characters
Let us recall the idea of the construction of the CliffordโFourier transform forcolour image processing introduced in [3]. From the mathematical viewpoint, aFourier transform is defined through group actions and more precisely throughirreducible and unitary representations of the involved group. This is closely relatedto the well-known shift theorem stating that:
โฑ๐๐ผ(๐ข) = ๐๐๐ผ๐ขโฑ๐(๐ข) (4.1)
where ๐๐ผ(๐ข) = ๐(๐ผ+ ๐ข). The group morphism
๐ผ ๏ฟฝโโ ๐๐๐ผ๐ข (4.2)
is a so-called character of the additive group (โ,+), that is an irreducible unitaryrepresentation of dimension 1.
The definition proposed in [3] relies on a Clifford generalization of this notionby introducing spin characters. It can be shown that the group morphisms fromโค/๐โคร โค/๐โค to Spin(3) are given by
๐๐ข,๐ฃ,๐ต : (๐,๐) ๏ฟฝโโ ๐2๐(๐ข๐/๐+๐ฃ๐/๐)๐ต (4.3)
186 T. Batard and M. Berthier
9. CliffordโFourier Transform 187
where
๐2๐(๐ข๐๐ + ๐ฃ๐
๐ )๐ต = cos 2๐(๐ข๐๐ + ๐ฃ๐
๐
)+ sin 2๐
(๐ข๐๐ + ๐ฃ๐
๐
)๐ต (4.4)
(๐ข, ๐ฃ) โ โค/๐โคร โค/๐โค, and
๐ต = ๐พ1๐1๐2 + ๐พ2๐1๐3 + ๐พ3๐2๐3 (4.5)
is a unit bivector, i.e., ๐พ21 + ๐พ2
2 + ๐พ23 = 1. The map ๐๐ข,๐ฃ,๐ต is called a spin character
of the group โค/๐โค ร โค/๐โค. Recalling that Spin(3) acts on the sections of thespinor bundle, we are led to propose the following definition.
Definition 4.1. The CliffordโFourier transform of a spinor ๐ of ๐(ฮฃ) is given by
โฑ(๐)(๐ข, ๐ฃ) =โ
๐โโค/๐โค
๐โโค/๐โค
๐๐ข,๐ฃ,๐ง1โง๐ง2 (๐,๐)(โ๐,โ๐) โ ๐(๐,๐) (4.6)
where (๐ง1, ๐ง2) is an orthonormal frame of ๐ (ฮฃ).
Since the spinor bundle of ฮฃ splits into
๐(ฮฃ) = ๐+๐ง1โง๐ง2(ฮฃ)โ ๐โ๐ง1โง๐ง2(ฮฃ) (4.7)
we have
๐๐ข,๐ฃ,๐ง1โง๐ง2(๐,๐)(โ๐,โ๐) โ ๐(๐,๐) = ๐2๐๐(๐ข๐๐ + ๐ฃ๐
๐
)๐+(๐,๐) ๐ฃโ๐(๐,๐)
+ ๐โ2๐๐(๐ข๐๐ + ๐ฃ๐
๐
)๐โ(๐,๐) ๐ฃ๐(๐,๐) (4.8)
where ๐ฃโ๐, respectively ๐ฃ๐, is the unit eigenspinor field of eigenvalueโ๐, respectively๐, relatively to the operator ๐ง1 โง ๐ง2 โ (here โ denotes the Clifford multiplication).Consequently
โฑ(๐)(๐ข, ๐ฃ) =(๐+
โ1(๐ข, ๐ฃ), ๐โ(๐ข, ๐ฃ)
)(4.9)
in the frame (๐ฃโ๐, ๐ฃ๐), where ห and ห โ1 denote the Fourier transform on
๐ฟ2(โค/๐โคร โค/๐โค,โ),
also called discrete Fourier transform, and its inverse.
4.2. Spinor Field Decomposition
The inverse CliffordโFourier transform of ๐ is
โฑโ1(๐)(๐ข, ๐ฃ) =โ
๐โโค/๐โค
๐โโค/๐โค
๐๐ข,๐ฃ,๐ง1โง๐ง2(๐,๐)(๐,๐) โ ๐(๐,๐) (4.10)
This means that every spinor field ๐ may be written as a superposition of basicspinor fields, i.e.,
๐ =โ
๐๐,๐ (4.11)
where
๐๐,๐ : (๐ข, ๐ฃ) ๏ฟฝโโ ๐๐ข,๐ฃ,๐ง1โง๐ง2(๐,๐)(๐,๐) โ โฑ(๐)(๐,๐) (4.12)
188 T. Batard and M. Berthier
Following the splitting ๐(ฮฃ) = ๐+๐ง1โง๐ง2(ฮฃ)โ ๐โ๐ง1โง๐ง2(ฮฃ), we have
๐๐,๐ =(๐+๐,๐, ๐
โ๐,๐
)in the frame (๐ฃโ๐, ๐ฃ๐), with
๐+๐,๐ : (๐ข, ๐ฃ) ๏ฟฝโโ ๐โ2๐๐(๐ข๐/๐+๐ฃ๐/๐)๐+
โ1(๐,๐)
and
๐โ๐,๐ : (๐ข, ๐ฃ) ๏ฟฝโโ ๐2๐๐(๐ข๐/๐+๐ฃ๐/๐)๐โ(๐,๐)
Moreover,
โฃ๐๐,๐โฃ2 = โฃ๐+๐,๐โฃ2 + โฃ๐โ๐,๐โฃ2
since ๐+๐ง1โง๐ง2(ฮฃ) and ๐โ๐ง1โง๐ง2(ฮฃ) are orthogonal.
4.3. Experiments
Let us now give an example of applications of the CliffordโFourier transform onspinor fields to image processing. In order to perform filtering with the decompo-sition (4.11), we proceed as follows. Let ๐ผ be a grey-level image, and ๐โ be thecorresponding spinor representation given in Definition 2.2. We apply a Gaussianmask ๐๐ of variance ๐ in the spectrum โฑ๐โ of ๐โ. Then, we consider the norm ofits inverse Fourier transform, i.e., โฃโฑโ1๐๐โฑ๐โโฃ and the function โฃโฑโ1๐๐โฑ๐โโฃ ๐ผ.
Figures 2 and 3 show results of this process for different values of ๐ (leftcolumn โฃโฑโ1๐๐โฑ๐โโฃ and right column โฃโฑโ1๐๐โฑ๐โโฃ ๐ผ). It is clear that for ๐ suffi-ciently high, we have โฃโฑโ1๐๐โฑ๐โโฃ ๐ผ โ ๐ผ and โฃโฑโ1๐๐โฑ๐โโฃ โ 1 since โฃ๐โโฃ = 1. Thisexplains why the two left lower images are almost white and the two right lowerimages are almost the same as the originals.
We can see in the left columns of Figures 2 and 3 that the filtering actsthrough ๐โ as a smoothing of the geometry of the image. More precisely, when ๐is small, the modulus โฃโฑโ1๐๐โฑ๐โโฃ is small at points corresponding to nearly allthe geometric variations of the image. When ๐ increases the modulus is affectedonly at points corresponding to the strongest geometric variations, i.e., to bothedges and textures (and also where the noise is high).
The right columns of Figures 2 and 3 show that the filtering acts throughโฃโฑโ1๐๐โฑ๐โโฃ ๐ผ as a diffusion that leaves the geometric data untouched (the higherthe value of ๐, the more important is the diffusion). This appears clearly in Figure 4(compare the plumes of the hat) or in Figure 5 (compare the hair).
These experiments show that our approach is relevant to deal with harmonicanalysis together with Riemannian geometry.
Conclusion
Spin geometry is a powerful mathematical tool to deal with many theoreticaland applied geometric problems. In this chapter we have shown how to take ad-vantage of the generalized Weierstrass representation to perform grey-level imageprocessing, in particular edge and texture detection. Our main contribution is the
9. CliffordโFourier Transform 189
Figure 2. Left: โฃโฑโ1(๐๐ โฑ๐โ)โฃ for ๐ = 100, 1000, 10000, 100000 (fromtop to bottom). Right: โฃโฑโ1(๐๐ โฑ๐โ)โฃ๐ผ
definition of a CliffordโFourier transform for spinor fields that relies on a general-ization of the usual notion of character (the spin character). One important factis that this new transform takes into account the Riemannian geometry of the
190 T. Batard and M. Berthier
Figure 3. Left: โฃโฑโ1(๐๐ โฑ๐โ)โฃ for ๐ = 100, 1000, 10000, 100000 (fromtop to bottom). Right: โฃโฑโ1(๐๐ โฑ๐โ)โฃ๐ผ
image surface by involving the spinor field that parameterizes the normal and thebivector field coding the tangent plane. We have also introduced what appears to
9. CliffordโFourier Transform 191
Figure 4. Left: original. Right: โฃโฑโ1๐๐โฑ๐โโฃ ๐ผ with ๐ = 100
Figure 5. Left: original. Right: โฃโฑโ1๐๐โฑ๐โโฃ ๐ผ with ๐ = 100
be a harmonic decomposition of the parametrization and investigated applicationsto filtering.
Note that there are only two cases where the Grassmannian ๐บ๐,2 of 2-planesin โ๐ admits a rational parametrization. In fact, one can show that ๐บ3,2 โ โ๐ 1
and ๐บ4,2 โ โ๐ 1 รโ๐ 1 (see [9]). The case treated here corresponds to ๐บ3,2. As aconsequence the generalization to colour images is not straightforward. Neverthe-less, a quite different approach is possible to tackle this problem and will be thesubject of a forthcoming paper.
Let us also mention that one may envisage performing diffusion on grey-levelimages through the heat equation given by the Dirac operator. The latter is wellknown be a square root of the Laplacian. Preliminary results are discussed in [2]that show that this diffusion better preserves edges and textures than the usualRiemannian approaches.
192 T. Batard and M. Berthier
Appendix A. Mathematical Background
We recall here some definitions and results concerning spin geometry. The readermay refer to [7] for details and conventions. We focus on the particular case of anoriented surface immersed in โ3.
A.1. Complex Representations of ๐ชโ3,0 โ โ
Let (๐1, ๐2, ๐3) be an orthonormal basis of โ3. The Clifford algebra ๐ถโ3,0 is thequotient of the tensor algebra of the vectorial space โ3 by the ideal generatedby the elements ๐ขโ ๐ข+๐(๐ข) where ๐ is the Euclidean quadratic form. It can beshown that ๐ถโ3,0 is isomorphic to the product โรโ of two copies of the quaternionalgebra. The complex Clifford algebra ๐ถโ3,0โโ is isomorphic to โ(2)โโ(2) whereโ(2) denotes the algebra of 2ร2-matrices with complex entries. This decompositionis given by
๐ถโ3,0 โ โ โ (๐ถโ3,0 โ โ)+ โ (๐ถโ3,0 โ โ)โ (A.1)
where(๐ถโ3,0 โ โ)ยฑ = (1ยฑ ๐3)๐ถโ3,0 โ โ (A.2)
and ๐3 is the pseudoscalar ๐1๐2๐3. More precisely, the subalgebra (๐ถโ3,0 โโ)+ isgenerated by the elements
๐ผ1 =1 + ๐1๐2๐3
2, ๐ผ2 =
๐2๐3 โ ๐1
2, ๐ผ3 =
๐2 + ๐1๐3
2, ๐ผ4 =
๐3 โ ๐1๐2
2(A.3)
and an isomorphism with โ(2) is given by sending these elements to the matrices
๐ด1 =
(1 00 1
), ๐ด2 =
(0 ๐๐ 0
), ๐ด3 =
(๐ 00 โ๐
), ๐ด4 =
(0 1
โ1 0
). (A.4)
In the same way, (๐ถโ3,0 โ โ)โ is generated by
๐ฝ1 =1โ ๐1๐2๐3
2, ๐ฝ2 =
๐2๐3 + ๐1
2, ๐ฝ3 =
๐1๐3 โ ๐2
2, ๐ฝ4 =
โ๐3 โ ๐1๐2
2(A.5)
and an isomorphism is given by sending these elements to the above matrices ๐ด1,๐ด2, ๐ด3 and ๐ด4.
Let us denote by ๐ the natural representation of โ(2) on โ2. The two equiv-alent classes ๐1 and ๐2 of irreducible complex representations of ๐ถโ3,0 โ โ aregiven by
๐1(๐1 + ๐2) = ๐(๐1) ๐2(๐1 + ๐2) = ๐(๐2). (A.6)
They are characterized by
๐1(๐3) = ๐ผ๐ and ๐2(๐3) = โ๐ผ๐ (A.7)
For the sake of completeness, let us list these representations explicitly:
๐1(1) = ๐(๐ผ1) = ๐ด1, ๐1(๐1) = ๐(โ๐ผ2) = โ๐ด2
๐1(๐2) = ๐(๐ผ3) = ๐ด3, ๐1(๐3) = ๐(๐ผ4) = ๐ด4
๐1(๐1๐2) = ๐(โ๐ผ4) = โ๐ด4, ๐1(๐1๐3) = ๐(๐ผ3) = ๐ด3
๐1(๐2๐3) = ๐(๐ผ2) = ๐ด2, ๐1(๐3) = ๐(๐ผ1) = ๐ด1
(A.8)
9. CliffordโFourier Transform 193
and
๐2(1) = ๐(๐ฝ1) = ๐ด1, ๐2(๐1) = ๐(๐ฝ2) = ๐ด2
๐2(๐2) = ๐(โ๐ฝ3) = โ๐ด3, ๐2(๐3) = ๐(โ๐ฝ4) = โ๐ด4
๐2(๐1๐2) = ๐(โ๐ฝ4) = โ๐ด4, ๐2(๐1๐3) = ๐(๐ฝ3) = ๐ด3
๐2(๐2๐3) = ๐(๐ฝ2) = ๐ด2, ๐2(๐3) = ๐(โ๐ฝ1) = โ๐ด1.
(A.9)
The complex spin representation of Spin(3) is the homomorphism
ฮ3 : Spin(3) โโ โ(2) (A.10)
given by restricting an irreducible complex representation of ๐ถโ3,0โโ to the spinorgroup Spin(3) โ (๐ถโ3,0 โ โ)0 (see for example [4] for the definition of the Spingroup). Note that ฮ3 is independant of the chosen representation.
A.2. Spin Structures and Spinor Bundles
Let us denote by ๐ the Riemannian manifold โ3 and ๐๐๐(๐) the principal๐๐(3)-bundle of oriented orthonormal frames of ๐ . A spin structure on ๐ is aprincipal Spin(3)-bundle ๐Spin(๐) together with a 2-sheeted covering
๐Spin(๐) โโ ๐๐๐(๐) (A.11)
that is compatible with ๐๐(3) and Spin(3) actions. The Spinor bundle ๐(๐)is the bundle associated to the spin structure ๐Spin(๐) and the complex spinrepresentation ฮ3. More precisely, it is the quotient of the product ๐Spin(๐)รโ2
by the action
Spin(3)ร ๐Spin(๐)ร โ2 โโ ๐Spin(๐)ร โ2 (A.12)
that sends (๐, ๐, ๐ง) to (๐๐โ1,ฮ3(๐)๐ง). We will write
๐(๐) = ๐Spin(๐)รฮ3 โ2. (A.13)
It appears that the fiber bundle ๐(๐) is a bundle of complex left modules overthe Clifford bundle ๐ถ๐(๐) = ๐Spin(๐)ร๐ด๐ โ๐(3) of ๐ . In the sequel
(๐ข, ๐) ๏ฟฝโโ ๐ข โ ๐ (A.14)
denotes the corresponding multiplication for ๐ข โ ๐ (๐) and ๐ a section of ๐(๐).We consider now an oriented surface ฮฃ embedded in ๐ . Let us denote by
(๐ง1, ๐ง2) an orthonormal frame of ๐ (ฮฃ) and ๐ the global unit field normal to ฮฃ.Using the map
(๐ง1, ๐ง2) ๏ฟฝโโ (๐ง1, ๐ง2, ๐) (A.15)
it is possible to pull back the bundle ๐Spin(๐)โฃฮฃ to obtain a spin structure ๐Spin(ฮฃ)
on ฮฃ. Since ๐ถโ2,0 โ โ is isomorphic to (๐ถโ3,0 โ โ)0 under the map ๐ผ defined by
๐ผ(๐0 + ๐1) = ๐0 + ๐1๐ (A.16)
the algebra ๐ถโ2,0 โโ acts on โ2 via ๐2. This representation leads to the complexspinor representation ฮ2 of Spin(2). It can be shown that the induced bundle
๐(ฮฃ) = ๐Spin(ฮฃ)รฮ3โ๐ผ โ2 (A.17)
194 T. Batard and M. Berthier
coincides with the spinor bundle of the induced spin structure on ฮฃ. Once again๐(ฮฃ) is a bundle of complex left modules over the Clifford bundle ๐ถ๐(ฮฃ) of ฮฃ: theClifford multiplication is given by the map
(๐ฃ, ๐) ๏ฟฝโโ ๐ฃ โ ๐ โ ๐ (A.18)
for ๐ฃ โ ๐ (ฮฃ) and ๐ a section of ๐ (ฮฃ).The Spinor bundle ๐(ฮฃ) decomposes into
๐(ฮฃ) = ๐+(ฮฃ)โ ๐โ(ฮฃ) (A.19)
where
๐ยฑ(ฮฃ) = {๐ โ ๐(ฮฃ), ๐ โ ๐ง1 โ ๐ง2 โ ๐ = ยฑ๐} (A.20)
(compare [5]). Since ๐2(๐ง1๐ง2๐) is minus the identity, this is equivalent to
๐ยฑ(ฮฃ) = {๐ โ ๐(ฮฃ), ๐๐ โ ๐ = ยฑ๐}. (A.21)
A.3. Spinor Connections and Dirac Operators
Let โ๐ and โฮฃ be the LeviโCivita connections on the tangent bundles ๐ (๐)and ๐ (ฮฃ) respectively. The classical Gauss formula asserts that
โ๐๐ ๐ = โฮฃ
๐๐ โ โจโ๐๐ ๐, ๐ โฉ๐ (A.22)
where ๐ and ๐ are vector fields on ฮฃ. A similar formula exists when dealing withspinor fields. Let us first recall that one may construct on ๐(๐) and ๐(ฮฃ) somespinor LeviโCivita connections compatible with the Clifford multiplication, thatis connections which we continue to denote by โ๐ and โฮฃ verifying
โ๐๐ (๐ โ ๐) = (โ๐
๐ ๐ ) โ ๐+ ๐ โ โ๐๐ ๐ (A.23)
when ๐ and ๐ are vector fields on ๐ and ๐ is a section of ๐(๐) and a similarformula for โฮฃ. The analog of the Gauss formula reads
โ๐๐ ๐ = โฮฃ
๐๐โ 1
2(โ๐
๐ ๐) โ ๐ โ ๐ (A.24)
for ๐ a section of ๐(ฮฃ) and ๐ a vector field on ฮฃ (see [1] for a proof). If (๐ง1, ๐ง2) isan orthonormal frame of ๐ (ฮฃ), following [5], the Dirac operator on ๐(ฮฃ) is definedby
๐ท = ๐ง1 โ โฮฃ๐ง1 + ๐ง2 โ โฮฃ
๐ง2 (A.25)
and it can be verified that ๐ท๐ยฑ(ฮฃ) โ ๐โ(ฮฃ).Let now ๐ and ๐ be respectively a section of ๐(๐) and the section of ๐(ฮฃ)
given by the restriction ๐โฃฮฃ. We obtain from the Gauss spinor formula
๐ง1 โ โ๐๐ง1๐+ ๐ง2 โ โ๐
๐ง2๐ = ๐ท๐โ 1
2(๐ง1 โ (โ๐
๐ง1 ๐) โ ๐ โ ๐+ ๐ง2 โ (โ๐๐ง2 ๐) โ ๐ โ ๐). (A.26)
Since
๐ง1 โ (โ๐๐ง1 ๐) + ๐ง2 โ (โ๐
๐ง2 ๐) = โ2๐ป (A.27)
where ๐ป is the mean curvature of ฮฃ, it follows that
๐ท๐ = ๐ง1 โ โ๐๐ง1๐+ ๐ง2 โ โ๐
๐ง2๐โ๐ป โ ๐ โ ๐. (A.28)
9. CliffordโFourier Transform 195
References
[1] C. Bar. Metrics with harmonic spinors.Geometric and Functional Analysis, 6(6):899โ942, 1996.
[2] T. Batard and M. Berthier. The spinor representation of images. In K. Gurlebeck,editor, 9th International Conference on Clifford Algebras and their Applications,Weimar, Germany, 15โ20 July 2011.
[3] T. Batard, M. Berthier, and C. Saint-Jean. Clifford Fourier transform for color im-age processing. In E.J. Bayro-Corrochano and G. Scheuermann, editors, GeometricAlgebra Computing in Engineering and Computer Science, pages 135โ162. Springer,London, 2010.
[4] T. Batard, C.S. Jean, and M. Berthier. A metric approach to nD images edge detec-tion with Clifford algebras. Journal of Mathematical Imaging and Vision, 33:296โ312,2009.
[5] T. Friedrich. On the spinor representation of surfaces in Euclidean 3-space. Journalof Geometry and Physics, 28:143โ157, 1998.
[6] B. Lawson. Lectures on minimal manifolds, volume I, volume 9 of Mathematics Lec-ture Series. Publish or Perish Inc., Wilmington, Del., second edition, 1980.
[7] B. Lawson and M.-L. Michelson. Spin Geometry. Princeton University Press, Prince-ton, New Jersey, 1989.
[8] R. Osserman. A survey of minimal surfaces. Dover Publications, Inc., New York,second edition, 1986.
[9] I.A. Taimanov. Two-dimensional Dirac operator and surface theory. Russian Math-ematical Surveys, 61(1):79โ159, 2006.
[10] S.D. Zenzo. A note on the gradient of a multi-image. Computer Vision, Graphics,and Image Processing, 33:116โ125, 1986.
Thomas BatardDepartment of Applied MathematicsTel Aviv UniversityRamat AvivTel Aviv 69978, Israele-mail: [email protected]
Michel BerthierMIA LaboratoryLa Rochelle UniversityAvenue Michel CrepeauF-17042 La Rochelle, Francee-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 197โ219cโ 2013 Springer Basel
10 Analytic Video (2D + ๐) SignalsUsing CliffordโFourier Transformsin Multiquaternion GrassmannโHamiltonโClifford Algebras
P.R. Girard, R. Pujol, P. Clarysse, A. Marion,R. Goutte and P. Delachartre
Abstract. We present an algebraic framework for (2D+ ๐ก) video analytic sig-nals and a numerical implementation thereof using Clifford biquaternions andCliffordโFourier transforms. Though the basic concepts of CliffordโFouriertransforms are well known, an implementation of analytic video sequences us-ing multiquaternion algebras does not seem to have been realized so far. Aftera short presentation of multiquaternion Clifford algebras and CliffordโFouriertransforms, a brief pedagogical review of 1D and 2D quaternion analytic sig-nals using right quaternion Fourier transforms is given. Then, the biquater-nion algebraic framework is developed to express CliffordโFourier transformsand (2D + ๐ก) video analytic signals in standard and polar form constitutedby a scalar, a pseudoscalar and six phases. The phase extraction procedure isfully detailed. Finally, a numerical implementation using discrete fast Fouriertransforms of an analytic multiquaternion video signal is provided.
Mathematics Subject Classification (2010). Primary 15A66; secondary 42A38.
Keywords. Multiquaternion Clifford algebras, Clifford biquaternions, CliffordโFourier transforms, 2D + ๐ก analytic video signals.
1. Introduction
In recent decades, new algebraic structures based on Clifford algebras have beendeveloped [19,31]. Many of these developments use a geometric approach whereaswe propose an algebraic multiquaternion approach [10โ12]. This chapter focuses ona (2D+ ๐ก) Clifford biquaternion analytic signal using CliffordโFourier transforms[1, 29]. After a short pedagogical review of 1D complex and 2D quaternion ana-lytic signals using right quaternion Fourier transforms, we shall provide a concrete
198 P.R. Girard et al.
algebraic Clifford biquaternion framework for the CliffordโFourier transform, theanalytic CliffordโFourier transform and the analytic signal both in standard andpolar form. Finally, we shall give a numerical implementation, using discrete fastFourier transforms, of a (2D + ๐ก) video analytic signal.
2. Multiquaternion GrassmannโHamiltonโClifford Algebras
In 1844, in his Ausdehnungslehre, Hermann Gunther Grassmann (1809โ1877) laidthe foundations of an ๐-dimensional associative multivector calculus [14โ16,27]. Ayear before, in 1843, William Rowan Hamilton (1805โ1865) discovered the quater-nions and later on, was to introduce biquaternions [6, 22โ24]. In 1878, WilliamKingdon Clifford (1845โ1879) was to give a concise definition of Clifford algebrasand to demonstrate a theorem relating Grassmannโs system to Hamiltonโs quater-nions [4],[5, pp. 266โ276]. Due to this close connection, we shall refer to Cliffordalgebras also as GrassmannโHamiltonโClifford algebras or multiquaternion alge-bras [13, 17, 18].
Definition 2.1. A Clifford algebra ๐ถ๐ is an algebra composed of ๐ generators๐1, ๐2, . . . , ๐๐ multiplying according to the rule ๐๐๐๐ = โ๐๐๐๐ (๐ โ= ๐) and suchthat ๐2
๐ = ยฑ1. The algebra ๐ถ๐ contains 2๐ elements constituted by the ๐ genera-tors, the various products ๐๐๐๐ , ๐๐๐๐๐๐, . . . and the unit element 1.
Theorem 2.2. If ๐ = 2๐ (๐ : integer), the Clifford algebra ๐ถ2๐ is the tensorproduct of ๐ quaternion algebras. If ๐ = 2๐โ 1, the Clifford algebra ๐ถ2๐โ1 is thetensor product of ๐ โ 1 quaternion algebras and the algebra (1, ๐) where ๐ is theproduct of the 2๐ generators (๐ = ๐1๐2 . . .๐2๐) of the algebra ๐ถ2๐.
Examples of Clifford algebras are the complex numbers โ (with ๐1 = ๐),quaternionsโ (๐1 = ๐, ๐2 = ๐), biquaternions (๐1 = ๐ผ๐, ๐2 = ๐ผ๐, ๐3 = ๐ผ๐, ๐ผ2 = โ1,๐ผ commuting with ๐, ๐,๐, ๐2
๐ = 1) and tetraquaternions โ โ โ (๐0 = ๐, ๐1 =๐๐ฐ, ๐2 = ๐๐ฑ , ๐3 = ๐๐ฒ, where the small ๐, ๐,๐ commute with the capital ๐ฐ,๐ฑ ,๐ฒ).These examples prove the Clifford theorem up to dimension ๐ = 4.
3. Analytic Signal in ๐ต Dimensions
Consider a GrassmannโHamiltonโClifford algebra having ๐ generators ๐๐ (with๐2๐ = โ1) and let ๐ด be a general element of this algebra [7, 21]. Call ๐พ[๐ด] the
conjugate of ๐ด such that
๐พ(๐ด๐ต) = ๐พ(๐ต)๐พ(๐ด),๐พ(๐๐) = โ๐๐. (3.1)
Given a function ๐ (๐) having its value in the Clifford algebra with ๐ = (๐ฅ1, ๐ฅ2,. . ., ๐ฅ๐), let ๐น (๐) with ๐ = (๐ข1, ๐ข2, . . . , ๐ข๐) denote the CliffordโFourier transform
๐น (๐) =
โซโ๐
๐(๐)๐โ
๐=1
๐โ๐๐2๐๐ข๐๐ฅ๐๐๐๐. (3.2)
10. Analytic Video (2D + ๐ก) Signals 199
The inverse CliffordโFourier transform is given by
๐(๐) =
โซโ๐
๐น (๐)
๐โ1โ๐=0
๐๐๐โ๐2๐๐ข๐โ๐๐ฅ๐โ๐๐๐๐. (3.3)
Proof. Inserting ๐น (๐) in the equation above, one obtains after integration respec-tively over ๐๐๐ and ๐๐๐โฒ [29]โซ
โ2๐
๐(๐โฒ)๐โ
๐=1
๐โ๐๐2๐๐ข๐๐ฅโฒ๐๐๐๐โฒ
๐โ1โ๐=0
๐๐๐โ๐2๐๐ข๐โ๐๐ฅ๐โ๐๐๐๐ (3.4)
=
โซโ๐
๐(๐โฒ)๐ฟ๐ (๐โ ๐โฒ) ๐๐๐โฒ = ๐(๐). โก
Given two functions ๐ (๐) , ๐ (๐) taking their values in the Clifford algebra,define the following product
โจ๐ (๐) , ๐ (๐)โฉ = 1
2
โซโ๐
[๐๐พ (๐) + ๐๐พ (๐)] ๐๐๐ (3.5)
and similarly โจ๐น (๐) , ๐บ (๐)โฉ. These products satisfy Plancherelโs theorem
โจ๐ (๐) , ๐ (๐)โฉ = โจ๐น (๐) , ๐บ (๐)โฉ (3.6)
Proof.
๐(๐) =
โซโ๐
๐น (๐)๐๐๐2๐๐ข๐๐ฅ๐ . . . ๐๐12๐๐ข1๐ฅ1๐๐๐ (3.7)
๐พ๐(๐) =
โซโ๐
๐โ๐12๐๐ขโฒ1๐ฅ1 . . . ๐โ๐๐2๐๐ขโฒ
๐๐ฅ๐๐พ [๐บ(๐โฒ)] ๐๐๐โฒ. (3.8)
After integration over ๐๐๐, one has
1
2
โซโ๐
๐ (๐)๐พ [๐ (๐)] ๐๐๐ =1
2
โซโ2๐
๐น (๐) ๐ฟ๐ (๐โ ๐โฒ)๐พ [๐บ (๐โฒ)] ๐๐๐๐๐๐โฒ
=1
2
โซโ๐
๐น (๐)๐พ [๐บ (๐)] ๐๐๐; (3.9)
similarly,
1
2
โซโ๐
๐ (๐)๐พ [๐ (๐)] ๐๐๐ =1
2
โซโ2๐
๐บ (๐) ๐ฟ๐ (๐โ ๐โฒ)๐พ [๐น (๐โฒ)] ๐๐๐๐๐๐โฒ
=1
2
โซโ๐
๐บ (๐)๐พ [๐น (๐)] ๐๐๐. (3.10)
Hence, adding the last two equations, one obtains equation (3.5). โก
200 P.R. Girard et al.
Next consider a real scalar function ๐ (๐) and its CliffordโFourier transform๐น (๐), the analytic CliffordโFourier transform ๐น๐ด(๐) and the analytic signal ๐๐ด(๐)are respectively defined by
๐น๐ด(๐) =
๐โ๐=1
[1 + sign (๐ข๐)]๐น (๐) (3.11)
๐๐ด(๐) =
โซโ๐
๐น๐ด(๐)
๐โ1โ๐=0
๐๐๐โ๐2๐๐ข๐โ๐๐ฅ๐โ๐๐๐๐ (3.12)
with sign(๐ข๐) yielding โ1, 0, or 1 depending on whether ๐ข๐ is negative, zero orpositive. In particular, the analytic signal satisfies the Parseval relation
โจ๐๐ด (๐) , ๐๐ด (๐)โฉ = โจ๐น๐ด (๐) , ๐น๐ด (๐)โฉ . (3.13)
4. Analytic Signals in 1 and 2 Dimensions
4.1. Complex 1D Analytic Signal
The one-dimensional analytic signal notion was introduced by D. Gabor [9] andJ. Ville [30] in 1948. Within the Clifford algebra framework it can be presentedas follows. Using the complex numbers as Clifford algebra (๐1 = ๐) for modeling1D physics, the CliffordโFourier transform of a scalar function ๐(๐ฅ1) is simply thestandard Fourier transform
๐น (๐ข1) =
โซ โ
โโ๐(๐ฅ1)๐
โ๐2๐๐ข1๐ฅ1๐๐ฅ1 (4.1)
with the inverse transformation
๐(๐ฅ1) =
โซ โ
โโ๐น (๐ข1)๐
๐2๐๐ข1๐ฅ1๐๐ข1. (4.2)
Furthermore, one defines the scalar product of two functions ๐(๐ฅ1), ๐(๐ฅ1) as
โจ๐(๐ฅ1), ๐(๐ฅ1)โฉ = 1
2
โซ โ
โโ(๐๐โ + ๐๐โ) ๐๐ฅ1 (4.3)
where ๐โ(๐ฅ1), ๐โ(๐ฅ1) are the complex conjugate functions. A few properties of the
Fourier transform are recalled in Table 1.Under an orthogonal symmetry transforming the basis vector ๐1 into โ๐1 one
has ๐น (โ๐ข1) = ๐น (๐ข1)โ where ๐น (๐ข1)
โ is the complex conjugate of ๐น (๐ข1). Hence,only one orthant (๐ข1 โฅ 0) of the Fourier domain is necessary to obtain the signal.The analytic Fourier transform is defined by
๐น๐ด(๐ข1) = [1 + sign(๐ข1)]๐น (๐ข1) (4.4)
and the analytic signal by
๐๐ด(๐ฅ1) =
โซ โ
โโ๐น๐ด(๐ข1)๐
๐2๐๐ข1๐ฅ1๐๐ข1. (4.5)
10. Analytic Video (2D + ๐ก) Signals 201
Table 1. Properties of the complex Fourier transform (๐, ๐ โ โ)
Property Complex function Fourier transform
Linearity ๐๐ (๐ฅ1) + ๐๐ (๐ฅ1) ๐๐น (๐ข1) + ๐๐บ (๐ข1)Translation ๐ (๐ฅ1 โ ๐ฅ0) ๐โ๐2๐๐ข1๐ฅ0๐น (๐ข1)
Scaling ๐ (๐๐ฅ1)1โฃ๐โฃ๐น
(๐ข1๐)
Partial derivativeโ๐๐ (๐ฅ1)
โ๐ฅ๐1(๐2๐๐ข1)
๐ ๐น (๐ข1)
Plancherel โจ๐, ๐โฉ = โจ๐น,๐บโฉParseval โจ๐, ๐โฉ = โจ๐น, ๐น โฉ
The analytic signal satisfies Plancherelโs theoremโซ โ
โโโฃ๐๐ด(๐ฅ1)โฃ2 ๐๐ฅ1 =
โซ โ
โโโฃ๐น๐ด(๐ข1)โฃ2 ๐๐ข1. (4.6)
As example, consider the function ๐(๐ฅ) = ๐ cos ๐๐ก which yields the analytic signal๐๐ด = ๐๐๐๐๐ก. The instantaneous amplitude and the instantaneous phase of a realsignal ๐ at a given position ๐ฅ can be defined as the modulus and the angularargument of the complex-valued analytic signal. Therefore, the analytic signalplays a key role in one-dimensional signal processing. It is also important to notethat the analytic signal concept is global, i.e., ๐๐ด(๐ฅ) depends on the whole of theoriginal signal and not only on values at positions near ๐ฅ. This concept can besummarized into three main properties:
1. The analytic signal has a one-sided spectrum;2. The original signal can be reconstructed from its analytic signal (in particular,
the real part of the analytic signal is equal to the original signal);3. The local amplitude (envelope) and the local phase of the original signal
can be derived as the modulus and angular argument of the analytic signalrespectively.
4.2. Quaternion 2D Analytic Signal
4.2.1. Clifford Algebra. The analytic signal notion can be extended to the 2D casein several ways [2, 3, 8, 20, 25]. Taking the quaternions โ as a Clifford algebra, itcontains the elements
[1, ๐1 = ๐, ๐2 = ๐, ๐1๐2 = ๐] . (4.7)
Hence, a general element ๐ can be expressed as
๐ = ๐0 + ๐๐1 + ๐๐2 + ๐๐3
= [๐0, ๐1, ๐2, ๐3] (4.8)
the conjugate ๐๐ being
๐๐ = [๐0,โ๐1,โ๐2,โ๐3] . (4.9)
202 P.R. Girard et al.
Figure 1. Orthogonal symmetry with respect to a straight line.
A 2D vector is expressed by ๐ = ๐๐1+๐๐2 with ๐๐๐ = โ๐2 = (๐1)2+(๐2)
2,a bivector by ๐ต = ๐๐3 and a scalar by ๐0.
Under an orthogonal symmetry with respect to a straight line passing throughthe origin and perpendicular to the unit vector ๐ = ๐๐1+๐๐2, (๐๐๐ = 1), a vector istransformed into the vector ๐ โฒ = ๐๐๐
๐๐๐= ๐๐๐ [10, p. 76], a bivector ๐ต = ๐ โง๐ =
โ 12 (๐๐ โ ๐ ๐) into ๐ตโฒ = โ๐๐ต๐, whereas a scalar remains invariant (Figure 1).
Hence, the entire quaternion transforms as
๐โฒ = ๐โ๐๐โ๐ (4.10)
where ๐โ = ๐๐ is the dual of ๐ (๐โ๐ = ๐๐๐๐ = ๐๐). One verifies the conservation ofthe square of the norm of the quaternion ๐โฒ๐โฒ๐ = ๐๐๐ under an orthogonal symme-try. Equation (4.10) allows us to deduce the involution formulas developed belowwhich are useful for the comprehension of the properties of the analytic Cliffordquaternion Fourier transform and the extraction of phases. In particular, if ๐ = ๐(orthogonal symmetry with respect to the ๐ฆ-axis), one has
๐โฒ = ๐พ1(๐) = ๐โ๐๐โ๐ = ๐๐๐๐๐ = โ๐๐๐ = ๐0 โ ๐๐1 + ๐๐2 โ ๐๐3 (4.11)
where ๐พ1(๐) is an involution of ๐. Similarly, if ๐ = ๐ (orthogonal symmetry withrespect to the ๐ฅ-axis)
๐โฒ = ๐พ2(๐) = ๐โ๐๐โ๐ = ๐๐๐๐๐ = โ๐๐๐ = ๐0 + ๐๐1 โ ๐๐2 โ ๐๐3. (4.12)
A combination of the two above orthogonal symmetries (๐ โ โ๐, ๐ โ โ๐) yieldsa rotation of ๐ around the origin
๐โฒ = ๐พ12(๐) = ๐๐๐๐๐ = โ๐๐๐ = ๐0 โ ๐๐1 โ ๐๐2 + ๐๐3. (4.13)
10. Analytic Video (2D + ๐ก) Signals 203
Table 2. Properties of the right quaternion Fourier transform (๐, ๐ โ โ)
Property scalar functions RQFT
Linearity ๐๐ (๐ฅ1, ๐ฅ2) + ๐๐ (๐ฅ1, ๐ฅ2) ๐๐น (๐ข1, ๐ข2) + ๐๐บ (๐ข1, ๐ข2)Translation ๐ (๐ฅ1 โ ๐1, ๐ฅ2 โ ๐2) ๐โ๐2๐๐ข1๐1๐น (๐ข1, ๐ข2)๐
โ๐2๐๐ข2๐2
Scaling ๐ (๐1๐ฅ1, ๐2๐ฅ2)1โฃ๐1โฃ
1โฃ๐2โฃ๐น
(๐ข1๐1
, ๐ข2๐2
)Partial derivative
โ๐๐โ๐ฅ๐1
(๐ฅ1, ๐ฅ2) ๐น (๐ข1, ๐ข2) (๐2๐๐ข1)๐
(๐ = 4๐) (2๐๐ข2)๐๐น (๐ข1, ๐ข2)
(๐ = 4๐+ 1)โ๐๐โ๐ฅ๐2
(๐ฅ1, ๐ฅ2) (2๐๐ข2)๐๐๐พ1[๐น (๐ข1, ๐ข2)]
(๐ = 4๐+ 2) โ (2๐๐ข2)๐๐น (๐ข1, ๐ข2)
(๐ = 4๐+ 3) โ (2๐๐ข2)๐ ๐๐พ1[๐น (๐ข1, ๐ข2)]
Plancherel โจ๐, ๐โฉ = โจ๐น,๐บโฉParseval โจ๐, ๐โฉ = โจ๐น, ๐น โฉ
4.2.2. Right Quaternion Fourier Transform (RQFT). As CliffordโFourier trans-form, we shall take the right quaternion Fourier transform [1, 26, 28]
๐น (๐ข1, ๐ข2) =
โซโ2
๐(๐ฅ1, ๐ฅ2)๐โ๐2๐๐ข1๐ฅ1๐โ๐2๐๐ข2๐ฅ2๐๐ฅ1๐๐ฅ2 (4.14)
and its inverse transform
๐(๐ฅ1, ๐ฅ2) =
โซโ2
๐น (๐ข1, ๐ข2)๐๐2๐๐ข2๐ฅ2๐๐2๐๐ข1๐ฅ1๐๐ข1๐๐ข2. (4.15)
The scalar product of two quaternion functions ๐(๐), ๐(๐) with (๐) = (๐ฅ1, ๐ฅ2) isdefined by
โจ๐(๐), ๐(๐)โฉ = 1
2
โซโ2
(๐๐๐ + ๐๐๐) ๐๐ฅ1๐๐ฅ2 (4.16)
where ๐๐, ๐๐ are the quaternion conjugates. Table 2 lists a few properties of theright Fourier transform.
From the definition of the QFT and the above orthogonal symmetry proper-ties, it follows that with (๐) = (๐ข1, ๐ข2)
๐น (โ๐ข1, ๐ข2) = ๐พ1 [๐น (๐)] = โ๐๐น (๐)๐ (4.17)
๐น (๐ข1,โ๐ข2) = ๐พ2 [๐น (๐)] = โ๐๐น (๐)๐ (4.18)
๐น (โ๐ข1,โ๐ข2) = ๐พ12 [๐น (๐)] = โ๐๐น (๐)๐ (4.19)
Hence, only one orthant of the Fourier space is necessary to represent the entireFourier space (Figure 2).
4.2.3. Analytic Signal. The analytic quaternion Fourier transform (Figure 3) isdefined as
๐น๐ด(๐) = [1 + sign(๐ข1)] [1 + sign(๐ข2)]๐น (๐) (4.20)
204 P.R. Girard et al.
๏ฟฝ๐ข1
๏ฟฝ๐ข2
๐น (๐ข1, ๐ข2)๐น (โ๐ข1, ๐ข2)= ๐พ1[๐น (๐ข1, ๐ข2)]= โ๐๐น (๐ข1, ๐ข2)๐
๐น (โ๐ข1,โ๐ข2)= ๐พ12[๐น (๐ข1, ๐ข2)]= โ๐๐น (๐ข1, ๐ข2)๐
๐น (๐ข1,โ๐ข2)= ๐พ2[๐น (๐ข1, ๐ข2)]= โ๐๐น (๐ข1, ๐ข2)๐
Figure 2. The quaternion Fourier spectrum of a real signal can bereconstructed from only one quadrant of the Fourier plane.
๏ฟฝ๐ข1
๏ฟฝ๐ข2
4๐น (๐ข1, ๐ข2)
Figure 3. The spectrum of the analytic quaternion Fourier transformis obtained from only one quadrant.
and the analytic signal by
๐๐ด(๐) =
โซโ2
๐น๐ด(๐)๐๐2๐๐ข2๐ฅ2๐๐2๐๐ข1๐ฅ1๐๐ข1๐๐ข2. (4.21)
The analytic signal satisfies Plancherelโs theoremโซโ2
๐๐ด(๐) [๐๐ด(๐)]๐ ๐๐ฅ1๐๐ฅ2 =
โซโ2
๐น๐ด(๐) [๐น๐ด(๐)]๐ ๐๐ข1๐๐ข2. (4.22)
10. Analytic Video (2D + ๐ก) Signals 205
Table 3. Examples of analytic signals.
Function Analytic signal
cos๐1๐ฅ1 cos๐2๐ฅ2 ๐๐๐2๐ฅ2๐๐๐1๐ฅ1
sin๐1๐ฅ1 sin๐2๐ฅ2 ๐๐๐๐2๐ฅ2๐๐๐1๐ฅ1
sin๐1๐ฅ1 cos๐2๐ฅ2 โ๐๐๐๐2๐ฅ2๐๐๐1๐ฅ1
cos๐1๐ฅ1 sin๐2๐ฅ2 โ๐๐๐๐2๐ฅ2๐๐๐1๐ฅ1
cos (๐1๐ฅ1 ยฑ ๐2๐ฅ2 + ๐) (1โ ๐) ๐๐๐๐๐๐2๐ฅ2๐๐๐1๐ฅ1
Examples of analytic signals are given in Table 3. The analytic signal being aquaternion, it can be represented as
๐ = โฃ๐โฃ ๐๐๐3๐๐๐2๐๐๐1 = โฃ๐โฃ ๐ (4.23)
with โฃ๐โฃ = โ๐๐๐ and where ๐ = ๐๐๐3๐๐๐2๐๐๐1 is a unit quaternion (๐๐๐ = 1). The
procedure of extracting the triplet of phases (๐1, ๐2, ๐3) is recalled below [29].Using the involutions ๐พ1,๐พ2,๐พ12 defined above one has (with โ representing thequaternion multiplication)
๐2 = ๐ โ [๐พ2 (๐)]๐
= [cos 2๐2 cos 2๐3, 0, sin 2๐2, cos 2๐2 sin 2๐3] (4.24)
๐12 = [๐พ12 (๐)]๐ โ ๐= [cos 2๐1 cos 2๐2, cos 2๐2 sin 2๐1, sin 2๐2, 0] . (4.25)
From ๐2 one obtains ๐2 = 12 arcsin๐2(3) where ๐2(3) means the third component
of ๐2. If cos 2๐2 โ= 0, one has
๐3 =Arg [๐2 (1) + ๐๐2(4)]
2, ๐1 =
Arg [๐12(1) + ๐๐12(2)]
2. (4.26)
If cos 2๐2 = 0 (๐2 = ยฑ๐/4), one has an indeterminacy and only (๐1 ยฑ ๐3) can bedetermined; adopting the choice ๐3 = 0, one has
๐1 = ๐ โ [๐พ1 (๐)]๐ = [cos 2๐1, 0, 0,โ sin 2๐1] (4.27)
and thus
๐1 = โ๐ด๐๐ [๐1(1) + ๐๐1(4)]
2. (4.28)
Finally, having determined the phases (๐1, ๐2, ๐3) one computes ๐๐๐3๐๐๐2๐๐๐1 ; if itis equal to โ๐ and ๐3 โฅ 0, one takes ๐3 โ ๐; if one has โ๐ and ๐3 < 0 then onetakes ๐3 + ๐. Hence, the domain of the phases is
(๐1, ๐2, ๐3) โ[โ๐
2,๐
2
]โช[โ๐
2,๐
2
]โช [โ๐, ๐[ . (4.29)
206 P.R. Girard et al.
5. Clifford Biquaternion 2D + ๐ Analytic Signal: Implementation
5.1. Clifford Biquaternion Algebra
In 2D + ๐ก dimensions, we shall take as Clifford algebra, Clifford biquaternionshaving as generators (๐1 = ๐๐, ๐2 = ๐๐, ๐3 = ๐๐, ๐ = ๐โฒ๐ผ, ๐2 = 1, ๐2
๐ = โ1) with ๐โฒ
designating the usual complex imaginary (๐โฒ2 = โ1) and where the tensor product๐ผ = 1โ ๐ (๐ผ2 = โ1) commutes with ๐ = ๐โ 1, ๐ = ๐ โ 1, ๐ = ๐โ 1; ๐1 correspondsto the time axis ๐ฅ1, ๐2 and ๐3 correspond respectively to the ๐ฅ2 and ๐ฅ3 axes. Thefull algebra contains the elements[
1 ๐ = ๐2๐3 ๐ = ๐3๐1 ๐ = ๐1๐2
๐ = ๐โฒ๐ผ = โ๐1๐2๐3 ๐1 = ๐๐ ๐2 = ๐๐ ๐3 = ๐๐
]. (5.1)
Hence, a general element of the algebra can be expressed as a Clifford biquaternion(a clifbquat for short)
๐ด = ๐+ ๐๐
= (๐0 + ๐๐1 + ๐๐2 + ๐๐3) + ๐(๐0 + ๐๐1 + ๐๐2 + ๐๐3)
= [๐0, ๐1, ๐2, ๐3] + ๐ [๐0, ๐1, ๐2, ๐3] (5.2)
where ๐, ๐ are quaternions. The product of two clifbquats ๐ด and ๐ต = ๐โฒ + ๐๐โฒ isgiven by
๐ด๐ต = (๐+ ๐๐) (๐โฒ + ๐๐โฒ)
= (๐๐โฒ + ๐๐โฒ) + ๐ (๐๐โฒ + ๐๐โฒ) . (5.3)
The conjugate of ๐ด is
๐ด๐ = ๐๐ + ๐๐๐
= [๐0,โ๐1,โ๐2,โ๐3] + ๐ [๐0,โ๐1,โ๐2,โ๐3] (5.4)
where ๐๐, ๐๐ are respectively the quaternion conjugates of ๐ and ๐. The complexconjugate of ๐ด is
๐ด = ๐โ ๐๐. (5.5)
A vector is expressed by
๐ = ๐(๐๐1 + ๐๐2 + ๐๐3) (5.6)
with ๐๐๐ = ๐21 + ๐2
2 + ๐23, a bivector by ๐ต = ๐ โง ๐ = โ 1
2 (๐๐โ ๐๐) , a trivector by
๐ = ๐ โง๐ต = 12 (๐๐ต +๐ต๐). A unit vector ๐ is defined by ๐๐๐ = 1.
Under an orthogonal symmetry with respect to a hyperplane (space of dimen-sion ๐ โ 1) perpendicular to a unit vector ๐, a vector ๐ is transformed into thevector ๐ โฒ = ๐๐๐, a bivector ๐ต into ๐ตโฒ = โ๐๐ต๐ and a trivector ๐ into ๐ โฒ = ๐๐๐,whereas a scalar remains invariant. These formulas allow us to derive the trans-formation of a clifbquat under an arbitrary orthogonal symmetry (Figure 4). Inparticular, if ๐ = ๐๐ (orthogonal symmetry with respect to the hyperplane ๐๐ฅ2๐ฅ3),one has the involution
๐ดโฒ = ๐พ1 (๐ด) = [๐0, ๐1,โ๐2,โ๐3] + ๐ [โ๐0,โ๐1, ๐2, ๐3] . (5.7)
10. Analytic Video (2D + ๐ก) Signals 207
Figure 4. Orthogonal symmetry with respect to a hyperplane.
Similarly, if ๐ = ๐๐ (orthogonal symmetry with respect to the hyperplane ๐๐ฅ1๐ฅ3),one has
๐ดโฒ = ๐พ2(๐ด) = [๐0,โ๐1, ๐2,โ๐3] + ๐ [โ๐0, ๐1,โ๐2, ๐3] . (5.8)
If ๐ = ๐๐, one has
๐ดโฒ = ๐พ3(๐ด) = [๐0,โ๐1,โ๐2, ๐3] + ๐ [โ๐0, ๐1, ๐2,โ๐3] . (5.9)
A combination of ๐พ1 followed by ๐พ2 leads to a rotation of ๐ around the origin
๐ดโฒ = ๐พ12(๐ด) = ๐๐ด๐๐ = [๐0,โ๐1,โ๐2, ๐3] + ๐ [๐0,โ๐1,โ๐2, ๐3] (5.10)
with ๐ = ๐๐ = โ๐. Similarly,
๐ดโฒ = ๐พ13(๐ด) = [๐0,โ๐1, ๐2,โ๐3] + ๐ [๐0,โ๐1, ๐2,โ๐3] (5.11)
๐ดโฒ = ๐พ23(๐ด) = [๐0, ๐1,โ๐2,โ๐3] + ๐ [๐0, ๐1,โ๐2,โ๐3] . (5.12)
A combination of three symmetries leads to
๐ดโฒ = ๐พ123(๐ด) = [๐0, ๐1, ๐2, ๐3] + ๐ [โ๐0,โ๐1,โ๐2,โ๐3] . (5.13)
5.2. CliffordโFourier Transform
The CliffordโFourier 2D+ ๐ก transform and its inverse are respectively defined with๐ = (๐ฅ1, ๐ฅ2, ๐ฅ3), and ๐ = (๐ข1, ๐ข2, ๐ข3) by
๐น (๐) =
โซ๐ 3
๐(๐)๐โ๐๐2๐๐ข1๐ฅ1๐โ๐๐2๐๐ข2๐ฅ2๐โ๐๐2๐๐ข3๐ฅ3๐๐ฅ1๐๐ฅ2๐๐ฅ3 (5.14)
๐(๐) =
โซ๐ 3
๐น (๐)๐๐๐2๐๐ข3๐ฅ3๐๐๐2๐๐ข2๐ฅ2๐๐2๐๐ข1๐ฅ1๐๐ข1๐๐ข2๐๐ข3. (5.15)
208 P.R. Girard et al.
Table 4. Computation principle of the CliffordโFourier transform(CFT) with a clifbquat entry (for the inverse CFT, the formulas arethe same except that the FFT is replaced by an IFFT and the integra-tion order is reversed: ๐ฅ3, ๐ฅ2, ๐ฅ1).
CFT of ๐(x) = [๐1, ๐2, ๐3, ๐4] + ๐ [๐5, ๐6, ๐7, ๐8]
real component: ๐๐ (x) = ๐ (x)FFT on ๐ฅ1 : ๐1 + ๐๐๐2
FFT on ๐ฅ2 : (๐1๐ผ + ๐๐๐1๐ฝ) + ๐๐ (๐2๐ผ + ๐๐๐2๐ฝ)FFT on ๐ฅ3 : (๐1๐ผ๐พ + ๐๐๐1๐ผ๐ฟ) + ๐๐ (๐1๐ฝ๐พ + ๐๐๐1๐ฝ๐ฟ)
+๐๐ [(๐2๐ผ๐พ + ๐๐๐2๐ผ๐ฟ) + ๐๐ (๐2๐ฝ๐พ + ๐๐๐2๐ฝ๐ฟ)]๐น๐ = ๐ถ๐น๐ (๐๐) = [๐1๐ผ๐พ , ๐1๐ฝ๐ฟ,โ๐2๐ผ๐ฟ, ๐2๐ฝ๐พ ] +๐ [โ๐2๐ฝ๐ฟ, ๐2๐ผ๐พ , ๐1๐ฝ๐พ , ๐1๐ผ๐ฟ]
CFT(๐)=(๐น1 + ๐๐น2 + ๐๐น3 + ๐๐น4) + ๐ (๐น5 + ๐๐น6 + ๐๐น7 + ๐๐น8)
Considering two clifbquat functions ๐(๐) and ๐(๐), one defines the product
โจ๐, ๐โฉ = 1
2
โซโ3
(๐๐๐ + ๐๐๐) ๐๐ฅ1๐๐ฅ2๐๐ฅ3 (5.16)
which satisfies Plancherelโs theorem
โจ๐ (๐) , ๐ (๐)โฉ = โจ๐น (๐) , ๐บ (๐)โฉ . (5.17)
To compute the direct CliffordโFourier transform (CFT), one proceeds in cascadeintegrating first with respect to ๐ฅ1 using a standard FFT. A second FFT (inte-gration with respect to ๐ฅ2) is then applied to each real component of the previouscomplex number. Then, a third FFT (integration on ๐ฅ3) is applied on each of theresulting real components. Finally, all the components are properly displayed asa clifbquat. For the inverse CliffordโFourier transform, one proceeds in the sameway on each real component of the clifbquat using an IFFT and reversing theorder of integration (see Table 4).
The above involutions lead to the following symmetries of the CliffordโFouriertransform of a scalar function ๐(๐ฅ1, ๐ฅ2, ๐ฅ3)
๐น (โ๐ข1, ๐ข2, ๐ข3) = ๐พ1 [๐น (๐)] , ๐น (๐ข1,โ๐ข2, ๐ข3) = ๐พ2 [๐น (๐)] (5.18)
๐น (๐ข1, ๐ข2,โ๐ข3) = ๐พ3 [๐น (๐)] , ๐น (โ๐ข1,โ๐ข2, ๐ข3) = ๐พ12 [๐น (๐)] (5.19)
๐น (โ๐ข1, ๐ข2,โ๐ข3) = ๐พ13 [๐น (๐)] , ๐น (๐ข1,โ๐ข2,โ๐ข3) = ๐พ23 [๐น (๐)] (5.20)
๐น (โ๐ข1,โ๐ข2,โ๐ข3) = ๐พ123 [๐น (๐)] . (5.21)
Hence, only one orthant of the Fourier space is necessary to represent the en-tire Fourier space. A few properties of the CliffordโFourier transform of a scalarfunction are given in Table 5.
5.3. Analytic Signal
The analytic CliffordโFourier transform is defined by
๐น๐ด(๐) = [1 + sign (๐ข1)] [1 + sign (๐ข2)] [1 + sign (๐ข3)]๐น (๐) (5.22)
10. Analytic Video (2D + ๐ก) Signals 209
Table 5. Properties of the CliffordโFourier transform of a scalar func-tion (๐, ๐ โ โ).
Property Functions CFT
Linearity ๐๐ (๐) + ๐๐ (๐) ๐๐น (๐) + ๐๐บ (๐)Translation ๐ (๐ฅ1 โ ๐1, ๐ฅ2, ๐ฅ3 โ ๐3) ๐โ๐๐2๐๐ข1๐1๐น (๐)๐โ๐๐2๐๐ข3๐3
Scaling ๐ (๐1๐ฅ1, ๐2๐ฅ2, ๐3๐ฅ3)1โฃ๐1โฃ
1โฃ๐2โฃ
1โฃ๐3โฃ๐น
(๐ข1๐1
, ๐ข2๐2
, ๐ข3๐3
)Partial derivative
โ๐๐โ๐ฅ๐1
(๐ฅ1, ๐ฅ2, ๐ฅ3) ๐น (๐) (๐๐2๐๐ข1)๐
(๐ = 4๐) (2๐๐ข2)๐๐น (๐)
(๐ = 4๐+ 1)โ๐๐โ๐ฅ๐2
(๐ฅ1, ๐ฅ2, ๐ฅ3) (2๐๐ข2)๐๐๐๐พ1[๐น (๐)]
(๐ = 4๐+ 2) โ (2๐๐ข2)๐๐น (๐)
(๐ = 4๐+ 3) โ (2๐๐ข2)๐ ๐๐๐พ1[๐น (๐)]
โ๐๐โ๐ฅ๐3
(๐ฅ1, ๐ฅ2, ๐ฅ3) ๐น (๐) (๐๐2๐๐ข3)๐
Plancherel โจ๐, ๐โฉ = โจ๐น,๐บโฉParseval โจ๐, ๐โฉ = โจ๐น, ๐น โฉ
Table 6. Examples of analytic signals.
Function: ๐(๐) Analytic signal: ๐๐ด(๐)
cos๐1๐ฅ1 cos๐2๐ฅ2 cos๐3๐ฅ3 ๐๐๐๐3๐ฅ3๐๐๐๐2๐ฅ2๐๐๐๐1๐ฅ1
cos๐1๐ฅ1 sin๐2๐ฅ2 sin๐3๐ฅ3 ๐๐๐๐๐3๐ฅ3๐๐๐๐2๐ฅ2๐๐๐๐1๐ฅ1
sin๐1๐ฅ1 sin๐2๐ฅ2 sin๐3๐ฅ3 ๐๐๐๐๐3๐ฅ3๐๐๐๐2๐ฅ2๐๐๐๐1๐ฅ1
cos๐1๐ฅ1 cos๐2๐ฅ2 sin๐3๐ฅ3 โ๐๐๐๐๐๐3๐ฅ3๐๐๐๐2๐ฅ2๐๐๐๐1๐ฅ1
cos (๐1๐ฅ1 + ๐2๐ฅ2 + ๐3๐ฅ3) (1โ ๐+ ๐ โ ๐) ๐๐๐๐3๐ฅ3๐๐๐๐2๐ฅ2๐๐๐๐1๐ฅ1
= 2๐๐๐/4๐โ๐๐/4๐๐๐๐3๐ฅ3๐๐๐๐2๐ฅ2๐๐๐๐1๐ฅ1
and the analytic signal by
๐๐ด(๐) =
โซ๐ 3
๐น๐ด(๐)๐๐๐2๐๐ข3๐ฅ3๐๐๐2๐๐ข2๐ฅ2๐๐๐2๐๐ข1๐ฅ1๐๐ข1๐๐ข2๐๐ข3. (5.23)
Examples of analytic signals are given in Table 6. The analytic signal being aClifford biquaternion, it can be represented as ๐ด = ๐๐ where ๐ = ๐ผ+๐๐ฝ consists ofa scalar and a pseudo-scalar and where ๐ is a unit Clifford biquaternion (๐๐๐ = 1).Writing, ๐ด๐ด๐ = ๐2 = ๐1 + ๐๐2, one finds
๐ผ =
โ๐1 +
โ๐21 โ ๐2
2
2, ๐ฝ =
โ๐1 โ
โ๐21 โ ๐2
2
2(5.24)
(if ๐ผ = ๐ฝ = 0, one adopts the choice ๐ = 1, in order to have a unit Cliffordbiquaternion ๐ in all cases). The unit Clifford biquaternion ๐ is obtained via the
210 P.R. Girard et al.
relation
๐ = ๐โ1๐ด = (๐ผ+ ๐๐ฝ)โ1 ๐ด =
(๐ผโ ๐๐ฝ
๐ผ2 โ ๐ฝ2
)๐ด, (5.25)
and is decomposed according to ๐ = ๐๐, with ๐ = ๐1 + ๐ (๐๐2 + ๐๐3 + ๐๐4) being aunit Clifford biquaternion such that ๐๐ = ๐ and where ๐ = ๐1+๐๐2+๐๐3+๐๐4 with๐๐๐ = 1. Using a procedure similar to that used in special relativity [11, p. 82], onehas
๐ =1 + ๐โ
2 + ๐+ ๐๐(5.26)
with ๐ = ๐ (๐๐). If ๐ = โ1, one has a particular case which has to be treated so asto yield a unit Clifford biquaternion ๐ in all cases. For example, if ๐ = ยฑ๐, (with๐ = โ1) one takes ๐ = ๐๐; if ๐ = ๐ (๐๐1 + ๐๐2 + ๐๐3), one adopts
๐ = ๐(๐๐1 + ๐๐2 + ๐๐3)โ
๐21 + ๐2
2 + ๐23
. (5.27)
The unit Clifford biquaternion ๐ is then obtained as ๐ = ๐๐๐. Both, ๐ and ๐ can beput into a polar form, respectively
๐ = ๐๐๐๐2[๐๐๐๐3
(๐๐๐๐1
)๐๐๐๐3
]๐๐๐๐2 (5.28)
= cos๐1 cos 2๐2 cos 2๐3
+ ๐(๐ sin๐1 + ๐ cos๐1 cos 2๐3 sin 2๐2 + ๐ cos๐1 sin 2๐3) (5.29)
with ๐๐ = ๐ and ๐ = ๐๐๐1๐๐๐3๐๐๐2 . The phases of ๐ are extracted according to therules:
๐1 = arcsin ๐ (2, 2) (5.30)
where ๐ (2, 2) means the second component of the second quaternion of ๐. Ifcos๐1 โ= 0,
๐3 =1
2arcsin
(๐ (2, 4)
cos๐1
); (5.31)
if cos๐1 โ= 0 and cos 2๐3 โ= 0
๐2 =1
2arcsin
(๐ (2, 3)
cos๐1 cos 2๐3
). (5.32)
The particular cases are treated as follows. If ๐1 = ยฑ๐2 , the choice ๐2 = ๐3 = 0
is adopted; if ๐1 โ= ยฑ๐2 and ๐3 = ยฑ๐
4 , one chooses ๐2 = 0. The sign of thereconstructed ๐ is then compared to that of the initial ๐; if it is opposed, ๐1
is replaced by [๐1 โ ๐ sign (๐1)]. The phases of ๐ = ๐๐๐1๐๐๐3๐๐๐2 are extractedaccording to the procedure presented for the 2D analytic signal by Sommer [29,p. 194] (except that eventual phase shifts are reported on ๐3 rather than on ๐1).
10. Analytic Video (2D + ๐ก) Signals 211
Within the 2D + ๐ก Clifford biquaternion framework, the procedure of extractingthe triplet of phases (๐1, ๐2, ๐3) is as follows. From the relations
๐1 = ๐ โ [๐พ2 (๐)]๐
= [cos 2๐1 cos 2๐3, sin 2๐1 cos 2๐3, 0, sin 2๐3] (5.33)
๐2 = [๐พ1 (๐)]๐ โ ๐= [cos 2๐2 cos 2๐3, 0, cos 2๐3 sin 2๐2, sin 2๐3] (5.34)
one obtains ๐3 = 12 arcsin๐1(4) where ๐1(4) means the fourth component of ๐1. If
cos 2๐3 โ= 0, one has
๐1 =1
2Arg [๐1 (1) + ๐๐1(2)] , ๐2 =
1
2Arg [๐2(1) + ๐๐2(3)] . (5.35)
If cos 2๐3 = 0, one has an indeterminacy and only (๐1 ยฑ ๐2) can be determinedfrom the relation
๐3 = [๐พ12 (๐)]๐ โ ๐ = [cos 2 (๐1 โ ๐2) , 0,โ sin 2 (๐1 โ ๐2) , 0] ; (5.36)
adopting the choice ๐1 = 0, one has
๐2 =1
2Arg [๐3(1) + ๐๐3(3)] . (5.37)
Finally, having determined the phases (๐1, ๐2, ๐3) one computes ๐๐๐1๐๐๐3๐๐๐2 ; if itis equal to โ๐ and ๐3 โฅ 0, one takes ๐3 โ ๐; if one has โ๐ and ๐3 < 0 then onetakes ๐3 + ๐. Hence, the phase domains for ๐ and for ๐ are respectively
(๐1, ๐2, ๐3) โ[โ๐
2,๐
2
]โช[โ๐
2,๐
2
]โช [โ๐, ๐[ (5.38)
(๐1, ๐2,๐3) โ [โ๐, ๐[ โช[โ๐
2,๐
2
]โช[โ๐
2,๐
2
]. (5.39)
Finally, the analytic signal is characterized by a scalar, a pseudo-scalar and thesix phases above.
6. Results
6.1. Introductory Example
One considers the function
๐(๐ก, ๐ฅ, ๐ฆ) = cos(๐1๐ก+ ๐1๐ฅ+ ๐1๐ฆ) + cos(๐2๐ก+ ๐2๐ฅ+ ๐2๐ฆ) (6.1)
with ๐๐ = ฮฉยฑ ๐2 , ๐๐ = ๐พ ยฑ ๐
2 , ๐๐ = ๐ ยฑ ๐2 and ๐2 > ๐1, ๐2 > ๐1, ๐2 > ๐1.
212 P.R. Girard et al.
The analytic signal is given by
๐๐ด(๐ก, ๐ฅ, ๐ฆ) =
โกโขโขโฃ2 cos(๐๐ก+ ๐๐ฅ+ ๐๐ฆ) cos(ฮฉ๐ก+๐พ๐ฅ+ ๐๐ฆ),
โ2 cos(๐๐กโ ๐๐ฅ+ ๐๐ฆ) cos(ฮฉ๐กโ๐พ๐ฅ+ ๐๐ฆ),2 cos(๐๐ก+ ๐๐ฅโ ๐๐ฆ) cos(ฮฉ๐กโ๐พ๐ฅโ ๐๐ฆ),
โ2 cos(๐๐กโ ๐๐ฅโ ๐๐ฆ) cos(ฮฉ๐กโ๐พ๐ฅโ ๐๐ฆ)
โคโฅโฅโฆ
+๐
โกโขโขโฃ2 cos(๐๐กโ ๐๐ฅ+ ๐๐ฆ) sin(ฮฉ๐กโ๐พ๐ฅ+ ๐๐ฆ),2 cos(๐๐ก+ ๐๐ฅ+ ๐๐ฆ) sin(ฮฉ๐ก+๐พ๐ฅ+ ๐๐ฆ),
โ2 cos(๐๐กโ ๐๐ฅโ ๐๐ฆ) sin(ฮฉ๐กโ๐พ๐ฅโ ๐๐ฆ),โ2 cos(๐๐ก+ ๐๐ฅโ ๐๐ฆ) sin(ฮฉ๐ก+๐พ๐ฅโ ๐๐ฆ)
โคโฅโฅโฆ . (6.2)
One then computes
๐4 = [๐๐ด(1)]2+ [๐๐ด(6)]
2= 4 cos2(๐๐ก+ ๐๐ฅ+ ๐๐ฆ) (6.3)
๐5 = ๐๐ด๐๐ด๐ =[8(1 + cos 2๐๐ก cos 2๐๐ฅ cos 2๐๐ฆ), 0, 0, 0]+๐ [8 sin 2๐๐ก sin 2๐พ๐ฅ cos 2๐๐ฆ, 0, 0, 0]
. (6.4)
The phase and group velocities are obtained as follows. Write
ฮฆ1 = ฮฉ๐ก+๐พ๐ฅ+ ๐๐ฆ,ฮฆ2 = ๐๐ก+ ๐๐ฅ+ ๐๐ฆ; (6.5)
one has
ฮฆ1 = arctan๐๐ด(6)
๐๐ด(1)(6.6)
ฮฆ2 =1
2arccos
(๐4
2โ 1
). (6.7)
Hence the phase and group velocities are
๐ฃ๐๐ฅ =
โฮฆ1โ๐กโฮฆ1โ๐ฅ
, ๐ฃ๐๐ฆ =
โฮฆ1โ๐กโฮฆ1โ๐ฆ
(6.8)
๐ฃ๐๐ฅ =
โฮฆ2โ๐กโฮฆ2โ๐ฅ
, ๐ฃ๐๐ฆ =
โฮฆ2โ๐กโฮฆ2โ๐ฆ
. (6.9)
Finally, one might notice that the scalar part of ๐5 contains only the modulationof the signal.
6.2. Analytic Video Signal 2D + ๐ Implementation
The implementation considers the scalar function
๐(๐ฅ1, ๐ฅ2, ๐ฅ3) = cos๐ฅ1 cos๐ฅ2 cos๐ฅ3 (6.10)
10. Analytic Video (2D + ๐ก) Signals 213
(๐ฅ1 being the time and ๐ฅ2, ๐ฅ3 respectively ๐ฅ, ๐ฆ) where (๐ฅ1, ๐ฅ2, ๐ฅ3) vary between[0, 2๐]. The explicit form of the analytic video signal is
๐ด =
โกโขโขโฃ(
cos๐ฅ1 cos๐ฅ2 cos๐ฅ3,โ cos๐ฅ1 sin๐ฅ2 sin๐ฅ3,sin๐ฅ1 cos๐ฅ2 sin๐ฅ3,โ sin๐ฅ1 sin๐ฅ2 cos๐ฅ3
)+๐
(sin๐ฅ1 sin๐ฅ2 sin๐ฅ3, sin๐ฅ1 cos๐ฅ2 cos๐ฅ3,cos๐ฅ1 sin๐ฅ2 cos๐ฅ3, cos๐ฅ1 cos๐ฅ2 sin๐ฅ3
)โคโฅโฅโฆ (6.11)
with ๐ด๐ด๐ = 1, hence ๐ = 1. The eight components of ๐ด are represented at thetime ๐ฅ1 = 2๐
8 s in Figure 5
๐ด = (๐, ๐, ๐, ๐) + ๐ (๐, ๐, ๐, โ) . (6.12)
The eight components in polar form (๐ผ, ๐1, ๐2, ๐3, ๐ฝ, ๐1, ๐2, ๐3) are represented inFigure 6.
7. Conclusion
This chapter has presented a concrete algebraic framework, i.e., Cliffordโs biquater-nions for the expression of CliffordโFourier transforms and 2D+ ๐ก analytic signals.Then, we have shown how to put the analytic signal into a polar form constitutedby a scalar, a pseudoscalar and six phases. Finally, using discrete fast Fouriertransforms we have implemented numerically the Clifford biquaternion Fouriertransform, the analytic Fourier transform and the analytic signal both in standardand polar form. Our next objective will be to extract physical information fromthe phases in medical images.
Appendix A. Examples of (2D + ๐) CliffordโFourier Transforms
A.1. Example 1
Signal
๐(๐ฅ1, ๐ฅ2, ๐ฅ3) = cos 2๐๐1๐ฅ1 cos 2๐๐2๐ฅ2 cos 2๐๐3๐ฅ3.
CliffordโFourier transform
๐น (๐ข1, ๐ข2, ๐ข3) =1
8[๐ฟ (๐1 โ ๐ข1) + ๐ฟ (๐1 + ๐ข1)] [๐ฟ (๐2 โ ๐ข2) + ๐ฟ (๐2 + ๐ข2)]
ร [๐ฟ (๐3 โ ๐ข3) + ๐ฟ (๐3 + ๐ข3)] .
Analytic CliffordโFourier transform
๐น๐ด(๐ข1, ๐ข2, ๐ข3) = ๐ฟ (๐1 โ ๐ข1) ๐ฟ (๐2 โ ๐ข2) ๐ฟ (๐3 โ ๐ข3) .
Analytic signal
๐๐ด(๐ฅ1, ๐ฅ2, ๐ฅ3) = ๐๐๐2๐๐3๐ฅ3๐๐๐2๐๐2๐ฅ2๐๐๐2๐๐1๐ฅ1 .
214 P.R. Girard et al.
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
0.6
0.4
0.2
0
0.2
0.4
0.6
Figure 5. The eight components of the analytic video signal ๐ด =(๐, ๐, ๐, ๐) + ๐(๐, ๐, ๐, โ) at the time ๐ฅ1 = 2๐
8 s.
10. Analytic Video (2D + ๐ก) Signals 215
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
1
1
1
1
1
[pixels]
[pix
els]
[Rad
ians
]
10 20 30 40 50 60
10
20
30
40
50
60
1.5
1
0.5
0
0.5
1
1.5
[pixels]
[pix
els]
[Radians]
10 20 30 40 50 60
10
20
30
40
50
60
1
0.5
0
0.5
1
1.5
[pixels]
[pix
els]
[Radians]
10 20 30 40 50 60
10
20
30
40
50
60
3
2
1
0
1
2
3
[pixels]
[pix
els]
10 20 30 40 50 60
10
20
30
40
50
60
1
0.5
0
0.5
1
[pixels]
[pix
els]
[Radians]
10 20 30 40 50 60
10
20
30
40
50
60
0.7
0.75
0.8
0.85
[pixels]
[pix
els]
[Radians]
10 20 30 40 50 60
10
20
30
40
50
60
0.05
0
0.05
[pixels]
[pix
els]
[Radians]
10 20 30 40 50 60
10
20
30
40
50
60
0.05
0
0.05
Figure 6. Polar form of the analytic video signal at the time ๐ฅ1 = 2๐8
s, (๐) is the scalar, (๐, ๐, ๐) are the three phases (๐1, ๐2, ๐3); (e) is thepseudo-scalar and (๐, ๐, โ) are the three phases (๐1, ๐2, ๐3).
216 P.R. Girard et al.
A.2. Example 2
Signal
๐(๐ฅ1, ๐ฅ2, ๐ฅ3) = cos (2๐๐1๐ฅ1 + 2๐๐2๐ฅ2 + 2๐๐3๐ฅ3) = ๐1 + ๐2 + ๐3 + ๐4
with
๐1 = cos 2๐๐1๐ฅ1 cos 2๐๐2๐๐ฅ2 cos 2๐๐3๐ฅ3
๐2 = โ sin 2๐๐1๐ฅ1 sin 2๐๐2๐๐ฅ2 cos 2๐๐3๐ฅ3
๐3 = โ sin 2๐๐1๐ฅ1 cos 2๐๐2๐๐ฅ2 sin 2๐๐3๐ฅ3
๐4 = โ cos 2๐๐1๐ฅ1 sin 2๐๐2๐๐ฅ2 sin 2๐๐3๐ฅ3.
CliffordโFourier transform
๐น (๐ข1, ๐ข2, ๐ข3) = ๐น1 + ๐น2 + ๐น3 + ๐น4, ๐น๐ = ๐ถ๐น๐ (๐๐)
where
๐น1 =1
8[๐ฟ (๐1 โ ๐ข1) + ๐ฟ (๐1 + ๐ข1)] [๐ฟ (๐2 โ ๐ข2) + ๐ฟ (๐2 + ๐ข2)]
ร [๐ฟ (๐3 โ ๐ข3) + ๐ฟ (๐3 + ๐ข3)]
and
๐น2 =โ18
[โ๐๐๐ฟ (๐1 โ ๐ข1) + ๐๐๐ฟ (๐1 + ๐ข1)] [โ๐๐๐ฟ (๐2 โ ๐ข2) + ๐๐๐ฟ (๐2 + ๐ข2)]
ร [โ๐ฟ (๐3 โ ๐ข3) + ๐ฟ (๐3 + ๐ข3)]
=โ๐
8[โ๐ฟ (๐1 โ ๐ข1) + ๐ฟ (๐1 + ๐ข1)] [โ๐ฟ (๐2 โ ๐ข2) + ๐ฟ (๐2 + ๐ข2)]
ร [โ๐ฟ (๐3 โ ๐ข3) + ๐ฟ (๐3 + ๐ข3)]
and
๐น3 =โ18
[โ๐๐๐ฟ (๐1 โ ๐ข1) + ๐๐๐ฟ (๐1 + ๐ข1)] [๐ฟ (๐2 โ ๐ข2) + ๐ฟ (๐2 + ๐ข2)]
ร [โ๐๐๐ฟ (๐3 โ ๐ข3) + ๐๐๐ฟ (๐3 + ๐ข3)]
=๐
8[โ๐ฟ (๐1 โ ๐ข1) + ๐ฟ (๐1 + ๐ข1)] [๐ฟ (๐2 โ ๐ข2) + ๐ฟ (๐2 + ๐ข2)]
ร [โ๐ฟ (๐3 โ ๐ข3) + ๐ฟ (๐3 + ๐ข3)]
and
๐น4 =โ18
[๐ฟ (๐1 โ ๐ข1) + ๐ฟ (๐1 + ๐ข1)] [โ๐๐๐ฟ (๐2 โ ๐ข2) + ๐๐๐ฟ (๐2 + ๐ข2)]
ร [โ๐๐๐ฟ (๐3 โ ๐ข3) + ๐๐๐ฟ (๐3 + ๐ข3)]
=โ๐
8[๐ฟ (๐1 โ ๐ข1) + ๐ฟ (๐1 + ๐ข1)] [โ๐ฟ (๐2 โ ๐ข2) + ๐ฟ (๐2 + ๐ข2)]
ร [โ๐ฟ (๐3 โ ๐ข3) + ๐ฟ (๐3 + ๐ข3)] .
10. Analytic Video (2D + ๐ก) Signals 217
Analytic CliffordโFourier transform
๐น๐ด(๐ข1, ๐ข2, ๐ข3) = (1โ ๐+ ๐ โ ๐) ๐ฟ (๐1 โ ๐ข1) ๐ฟ (๐2 โ ๐ข2) ๐ฟ (๐3 โ ๐ข3)
Analytic signal
๐๐ด(๐ฅ1, ๐ฅ2, ๐ฅ3) = (1โ ๐+ ๐ โ ๐)๐๐๐2๐๐3๐ฅ3๐๐๐2๐๐2๐ฅ2๐๐๐2๐๐1๐ฅ1 = ๐+ ๐๐
where
๐ =
[cos (2๐๐1๐ฅ1 + 2๐๐2๐ฅ2 + 2๐๐3๐ฅ3) ,โ cos (2๐๐1๐ฅ1 โ 2๐๐2๐ฅ2 + 2๐๐3๐ฅ3) ,cos (2๐๐1๐ฅ1 + 2๐๐2๐ฅ2 โ 2๐๐3๐ฅ3) ,โ cos (2๐๐1๐ฅ1 โ 2๐๐2๐ฅ2 โ 2๐๐3๐ฅ3)
]๐ =
[sin (2๐๐1๐ฅ1 โ 2๐๐2๐ฅ2 + 2๐๐3๐ฅ3) , sin (2๐๐1๐ฅ1 + 2๐๐2๐ฅ2 + 2๐๐3๐ฅ3) ,
โ sin (2๐๐1๐ฅ1 โ 2๐๐2๐ฅ2 โ 2๐๐3๐ฅ3) ,โ sin (2๐๐1๐ฅ1 + 2๐๐2๐ฅ2 โ 2๐๐3๐ฅ3)
]
References
[1] F. Brackx, N. De Schepper, and F. Sommen. The two-dimensional CliffordโFouriertransform. Journal of Mathematical Imaging and Vision, 26(1):5โ18, 2006.
[2] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[3] T. Bulow and G. Sommer. A novel approach to the 2D analytic signal. In F. Solinaand A. Leonardis, editors, Computer Analysis of Images and Patterns, volume 1689of Lecture Notes in Computer Science, pages 25โ32. Springer, Berlin/Heidelberg,1999.
[4] W.K. Clifford. Applications of Grassmannโs extensive algebra. American Journal ofMathematics, 1(4):350โ358, 1878.
[5] W.K. Clifford. Mathematical Papers. Chelsea Publishing Company, New York, 1968.First published 1882, edited by R. Tucker.
[6] M.J. Crowe. A History of Vector analysis: The Evolution of the Idea of a VectorialSystem. University of Notre Dame, Notre Dame, London, 1967.
[7] P. Delachartre, P. Clarysse, R. Goutte, and P.R. Girard. Mise en oeuvre du signalanalytique dans les algebres de Clifford. In GRETSI, page 4, Bordeaux, France, 2011.
[8] T.A. Ell. Hypercomplex Spectral Transformations. PhD thesis, University of Min-nesota, June 1992.
[9] D. Gabor. Theory of communication. Journal of the Institution of Electrical Engi-neers, 93(26):429โ457, 1946. Part III.
[10] P.R. Girard. Quaternions, Algebre de Clifford et Physique Relativiste. PPUR, Lau-sanne, 2004.
[11] P.R. Girard. Quaternions, Clifford Algebras and Relativistic Physics. Birkhauser,Basel, 2007. Translation of [10].
[12] P.R. Girard. Quaternion Grassmann-Hamilton-Clifford-algebras: new mathematicaltools for classical and relativistic modeling. In O. Dossel and W.C. Schlegel, editors,World Congress on Medical Physics and Biomedical Engineering, September 7โ12,2009, volume 25/IV of IFMBE Proceedings, pages 65โ68. Springer, 2010.
218 P.R. Girard et al.
[13] P.R. Girard. Multiquaternion Grassmann-Hamilton-Clifford algebras in physics andengineering: a short historical perspective. In K. Gurlebeck, editor, 9th InternationalConference on Clifford Algebras and their Applications, page 9. Weimar, Germany,15โ20 July 2011.
[14] H. Grassmann. Die lineale Ausdehungslehre: ein neuer Zweig der Mathematik,dargestellt und durch Anwendungen auf die ubrigen Zweige der Mathematik, wie auchdie Statik, Mechanik, die Lehre von Magnetismus und der Krystallonomie erlautert.Wigand, Leipzig, second, 1878 edition, 1844.
[15] H. Grassmann. Der Ort der Hamiltonโschen Quaternionen in der Ausdehnungslehre.Mathematische Annalen, 12:375โ386, 1877.
[16] H. Grassmann. Gesammelte mathematische und physikalische Werke. B.G. Teubner,Leipzig, 1894โ1911. 3 volumes in 6 parts.
[17] A. Gsponer and J.P. Hurni. Quaternions in mathematical physics (1): Alphabeti-cal bibliography. Preprint, available at: http://www.arxiv.org/abs/arXiv:mathph/0510059, 2008. 1430 references.
[18] A. Gsponer and J.P. Hurni. Quaternions in mathematical physics (2): Analyticalbibliography. Preprint, available at: http://www.arxiv.org/abs/arXiv:mathphy/
0511092, 2008. 1100 references.
[19] K. Gurlebeck and W. Sproรig. Quaternionic and Clifford Calculus for Physicists andEngineers. Wiley, Aug. 1997.
[20] S.L. Hahn. Multidimensional complex signals with single-orthant spectra. Proceed-ings of the IEEE, 80(8):1287โ1300, Aug. 1992.
[21] S.L. Hahn and K.M. Snopek. The unified theory of ๐-dimensional complex andhypercomplex analytic signals. Bulletin of the Polish Academy of Sciences TechnicalSciences, 59(2):167โ181, 2011.
[22] H. Halberstam and R.E. Ingram, editors. The Mathematical Papers of Sir WilliamRowan Hamilton, volume III Algebra. Cambridge University Press, Cambridge, 1967.
[23] W.R. Hamilton. Elements of Quaternions. Chelsea Publishing Company, New York,reprinted 1969 edition, 1969. 2 volumes (1899โ1901).
[24] T.L. Hankins. Sir William Rowan Hamilton. Johns Hopkins University Press, Bal-timore, London, 1980.
[25] J.P. Havlicek, J.W. Havlicek, and A.C. Bovik. The analytic image. In Proceedings1997 International Conference on Image Processing (ICIP โฒ97), volume 2, pages446โ449, Washington, DC, USA, October 26โ29 1997.
[26] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[27] H.-J. Petsche. Grassmann. Birkhauser, Basel, 2006.
[28] S.J. Sangwine and T.A. Ell. The discrete Fourier transform of a colour image. In J.M.Blackledge and M.J. Turner, editors, Image Processing II Mathematical Methods, Al-gorithms and Applications, pages 430โ441, Chichester, 2000. Horwood Publishing forInstitute of Mathematics and its Applications. Proceedings Second IMA Conferenceon Image Processing, De Montfort University, Leicester, UK, September 1998.
[29] G. Sommer, editor. Geometric computing with Clifford Algebras: Theoretical Foun-dations and Applications in Computer Vision and Robotics. Springer, Berlin, 2001.
10. Analytic Video (2D + ๐ก) Signals 219
[30] J. Ville. Theorie et applications de la notion de signal analytique. Cables et Trans-mission, 2A:61โ74, 1948.
[31] J. Vince. Geometric Algebra for Computer Graphics. Springer, London, 2008.
P.R. Girard, P. Clarysse, A. Marion, R. Goutte and P. DelachartreUniversite de Lyon, CREATIS; CNRS UMR 5220Inserm U1044; INSA-Lyon;Universite Lyon 1, FranceBat. Blaise Pascal7 avenue Jean CapelleF-69621 Villeurbanne, France
e-mail: [email protected]@creatis.insa-lyon.fr
R. PujolUniversite de LyonPole de Mathematiques, INSA-LyonBat. Leonard de Vinci21 avenue Jean CapelleF-69621 Villeurbanne, France
e-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 221โ246cโ 2013 Springer Basel
11 Generalized Analytic Signalsin Image Processing: Comparison,Theory and Applications
Swanhild Bernstein, Jean-Luc Bouchot, Martin Reinhardtand Bettina Heise
Abstract. This article is intended as a mathematical overview of the general-izations of analytic signals to higher-dimensional problems, as well as of theirapplications to and of their comparison on artificial and real-world imagesamples.
We first start by reviewing the basic concepts behind analytic signaltheory and derive its mathematical background based on boundary valueproblems of one-dimensional analytic functions. Following that, two gener-alizations are motivated by means of higher-dimensional complex analysis orClifford analysis. Both approaches are proven to be valid generalizations ofthe known analytic signal concept.
In the last part we experimentally motivate the choice of such higher-dimensional analytic or monogenic signal representations in the context ofimage analysis. We see how one can take advantage of one or the other rep-resentation depending on the application.
Mathematics Subject Classification (2010). Primary 94A12; secondary 44A12,30G35.
Keywords. Monogenic signal, monogenic functional theory, image processing,texture, Riesz transform.
1. Introduction
In the past years and since the pioneer work of Gabor [11], the analytic signalhas attracted much interest in signal processing and information theory. Due toan orthogonal decomposition of oscillating signals into envelope and instantaneousphase or respectively into energetic and structural components, this concept hasbecome very suitable for analyzing signals. In this context such a property is calleda split of identity and allows to separate the different characteristics of a signalinto useful components.
222 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
While this approach has given rise to many one-dimensional signal process-ing methods, other developments have been directed towards higher-dimensionalgeneralizations. Of particular interest is the two-dimensional case, i.e., how to dealwith images in an analytic way. As it will be demonstrated in our chapter, twomain directions have been taken, one based on multidimensional complex analysisand another one based on Clifford analysis.
This article is intended as an overview of the mathematical concepts behindanalytic signals based on the Hilbert transform (Section 2). Then, the mathemat-ical generalizations are detailed in Section 3. The end of that section is dedicatedto illustrative examples of the detailed differences between the two generaliza-tions. Section 4 describes the use of spinors for image analysis tasks. The lastsection of this article (Section 5) illustrates their applications like demodulationof two-dimensional AM-FM signals as provided, e.g., in interferometry and someapplications to the processing of natural images.
2. Analytic Signal Theory and Signal Decomposition
Analytic signals were introduced for signal processing in the context of communi-cation theory in the late 40๐ [11]. Since then, there has been a growing interestin the analytic signal as a useful tool for representing real-valued signals [25]. Westart here by first reviewing the basics of analytic signal theory and the Hilberttransform and see why the so-called split of identity is an interesting property. Inthe last part we review the mathematical basics and see how we can derive theanalytic signal from a boundary value problem in complex analysis.
2.1. Basic Analytic Signal Theory and the Hilbert Transform
Definition 2.1 (One-dimensional Fourier Transform). In the following, we use asFourier transform โฑ :
โฑ(๐)(๐ข) = ๐(๐ข) =1โ2๐
โซโ
๐(๐ก)๐โ๐๐ก๐ขd๐ก (2.1)
for ๐ก โ โ, ๐ข โ โ and ๐ โ ๐ฟ2(โ)
Definition 2.2 (Hilbert Transform). The Hilbert transform of a signal ๐ โ ๐ฟ2(โ)(or more generally ๐ โ ๐ฟ๐(โ), 1 < ๐ <โ) is defined, either in the spatial domainas a convolution with the Hilbert kernel (2.2), or as a Fourier multiplier (2.3):
โ๐ = โ โ ๐ (2.2)
โฑ(โ๐)(๐ข) = โ๐ sign(๐ข)โฑ(๐)(๐ข) (2.3)
where we have made use of two functions:
โ The Hilbert kernel โ(๐ก) = 1๐๐ก .
โ The operator sign(๐ข) =
โงโจโฉ1 ๐ข > 0
0 ๐ข = 0
โ1 ๐ข < 0.
11. Generalized Analytic Signals in Image Processing 223
Following its definition, we notice that the Hilbert transform acts as an asym-metric phase shift: if we write ยฑ๐ = ๐ยฑ๐๐/2, the phase of the Fourier spectrum ofthe Hilbert transform is obtained by a rotation of ยฑ90โ.Proposition 2.3 (Properties of the Hilbert Transform). Given a signal ๐ the fol-lowing hold true:
โ โ๐ข โ= 0, โฃโ๐(๐ข)โฃ = โฃโฑ(๐)(๐ข)โฃ,โ โโ๐ = โ๐ โ โโ1 = โโ.
Note that a constant function being not in ๐ฟ2 can not be reconstructed in this way.
The analytic signal is computed as a complex combination of both the originalsignal and its Hilbert transform:
Definition 2.4 (Analytic Signal).
๐๐ด = ๐ + ๐โ๐. (2.4)
Due to its definition, an analytic signal has a one-sided Fourier spectrum.Moreover, its values are doubled on the positive side. We also remark that it ispossible to recover the original signal from its analytic description by taking thereal part.
The following proposition holds:
Proposition 2.5.
โจ๐,โ๐โฉ๐ฟ2= 0 Orthogonality, (2.5)
โฅ๐โฅ22 = โฅโ๐โฅ22 Energy conservation. (2.6)
The energy equality is valid only if the DC, or zero-frequency component ofthe signal is neglected [10].
Note that it is possible to write the complex analytic signal in polar coor-dinates. In this case we have: โ๐ก โ โ, ๐๐ด(๐ก) = ๐ด(๐ก)๐๐๐(๐ก). ๐ด is called the localamplitude and ๐ is called the local phase. These local features are defined asfollows [11]:
Definition 2.6 (Local features).
๐ด(๐ก) =โ
๐(๐ก)2 +โ๐(๐ก)2 (2.7)
๐(๐ก) = arctan
(โ๐(๐ก)
๐(๐ก)
)= arctan
(โ (๐๐ด(๐ก))
โ (๐๐ด(๐ก))
). (2.8)
Proposition 2.7 (InvarianceโEquivariance, Split of identity [10]). The local phase,together with the local amplitude, fulfil the properties of invariance and equivari-ance:
โ The local phase depends only on the local structure.โ The local amplitude depends only on the local energy.
If moreover these features constitute a complete description of the signal, they aresaid to perform a split of identity.
224 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
However as stated in [10], a split of identity is strictly valid only for band-limited signals with a local zero mean property.
If these conditions are fulfilled the analytic signal representation relies on anorthogonal decomposition of the structural information (the local phase), and theenergetic information (the local amplitude).
This split of identity is illustrated in Figure 1. The first plot represents threesignals. They are sine waves generated from a mother sine wave (the red one). Theblue curve corresponds to a modification in terms of amplitude of the red one,while the green curve has half the frequency of the red one. Figures 1(b) and 1(c)are respectively the local amplitudes and phases of these three signals. Note thata small phase shift has been added to the blue curve for better readability. We canclearly see that due to the split of identity, modifying one local characteristic ofthe signal does not affect the other one and vice versa.
Figure 1. Illustration of the split of identity. (Explanation in the text.)
2.2. From Analytic Function to Analytic Signal
While the analytic signal is a very common concept in the field of signal theory, itsbasic mathematics can be derived from the theory of analytic functions. The closeconnection can be understood when considering the following RiemannโHilbert
11. Generalized Analytic Signals in Image Processing 225
problem with respect to the complex parameter ๐ง = ๐ฅ+ ๐๐ฆ:
โ๐น
โ๐ง= 0, ๐ง โ โ, ๐ฆ โฅ 0, (2.9)
โ (๐น (๐ฅ)) = ๐(๐ฅ), ๐ฅ โ โ. (2.10)
One solution of this problem is given by the Cauchy integral
๐น (๐ง) = ๐นฮ๐(๐ง) :=1
2๐๐
โซโ
1
๐ โ ๐ง๐(๐)d๐. (2.11)
Of course this solution is unique only up to a constant. Normally, this constantwill be fixed by the condition โ (๐น (๐ง0)) = ๐, i.e., the imaginary part of ๐น given atan interior point.
When we now consider the trace of ๐นฮ, i.e., the boundary value, we arriveat the so-called PlemeljโSokhotzki formula:
tr๐นฮ๐ =1
2(๐ผ + ๐โ)๐ =
1
2๐ +
1
2๐โ๐ =: ๐ฮ๐. (2.12)
Up to the factor 1/2 this corresponds to our above definition of an analytic signal.
In this way an analytic signal represents the boundary values of an ana-lytic function in the upper half-plane (or for periodic functions in the unit disc).Starting from this concept we are now going to take a look at higher-dimensionalgeneralizations.
3. Higher-dimensional Generalizations
Different approaches have been studied in past years to extend the definition of ananalytic signal to higher-dimensional spaces. Two of them have gained the great-est interest based respectively on multidimensional complex analysis and Cliffordanalysis.
3.1. Using Multiple Complex Variables
3.1.1. Mathematics. In 1998 Bulow proposed a definition of a hypercomplex signalbased on the so-called partial and total Hilbert transform [6]. To adapt this to ourpoint of view, that analytic signals are functions in a Hardy space, we consider thefollowing RiemannโHilbert problem in โ2:
โ๐น
โ ๐ง1= 0, (๐ง1, ๐ง2) โ โ2, ๐ฆ1, ๐ฆ2 โฅ 0, (3.1)
โ๐น
โ ๐ง2= 0, (๐ง1, ๐ง2) โ โ2, ๐ฆ1, ๐ฆ2 โฅ 0, (3.2)
โ (๐น (๐ฅ1, ๐ฅ2)) = ๐(๐ฅ1, ๐ฅ2), ๐ฅ1, ๐ฅ2 โ โ2. (3.3)
For the solution, (see, e.g., [8] or [21]), we need to point out that the domain isa poly-domain in the sense of โ๐, so that we can give it in the form of a Cauchy
226 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
integral:
๐น (๐ง1, ๐ง2) =1
4๐2
โซโ2
1
(๐1 โ ๐ง1)(๐2 โ ๐ง2)๐(๐1, ๐2)d๐1d๐2. (3.4)
Now again looking at the corresponding PlemeljโSokhotzki formula we get
tr๐น (๐ฅ1, ๐ฅ2) =1
4๐(๐ฅ1, ๐ฅ2)โ 1
4
โซโ2
1
(๐1 โ ๐ฅ1)(๐2 โ ๐ฅ2)๐(๐1, ๐2)d๐1d๐2
+ ๐1
4
(โซโ
1
๐1 โ ๐ฅ1๐(๐1, ๐ฅ2)d๐1 +
โซโ
1
๐2 โ ๐ฅ2๐(๐ฅ1, ๐2)d๐2
), (3.5)
which up to the factor 1/4 corresponds to the definition of an analytic signal givenby Hahn [13]. Here
โ1๐(๐ฅ1, ๐ฅ2) =
โซโ
1
๐1 โ ๐ฅ1๐(๐1, ๐ฅ2)d๐1 (3.6)
โ2๐(๐ฅ1, ๐ฅ2) =
โซโ
1
๐2 โ ๐ฅ2๐(๐ฅ1, ๐2)d๐2 (3.7)
is called a partial Hilbert transform, and
โ๐ ๐ =1
4
โซโ2
1
(๐1 โ ๐ฅ1)(๐2 โ ๐ฅ2)๐(๐1, ๐2)d๐1d๐2 (3.8)
a total Hilbert transform. On the level of Fourier symbols we get
โฑ(tr๐น )(๐ข1, ๐ข2) = (1 + sign๐ข1)(1 + sign๐ข2)โฑ๐(๐ข1, ๐ข2). (3.9)
Let us now take a look at the definition of Bulow. To this end we consider๐น to be a function of two variables ๐ง1 and ๐ท2 with two different imaginary units๐ and ๐ (with ๐2 = ๐2 = โ1), i.e., ๐ง1 = ๐ฅ1 + ๐๐ฆ1 and ๐ท2 = ๐ฅ2 + ๐๐ฆ2. We remarkthat both imaginary units can be understood as elements of the quaternionic basiswith multiplication rules ๐๐ = โ๐๐ = ๐. In this way the above RiemannโHilbertproblem can be rewritten as
โ
โ๐ง1๐น = 0, (๐ง1, ๐ท2) โ โ2, ๐ฆ1, ๐ฆ2 โฅ 0, (3.10)
๐นโ
โ๐ท2= 0, (๐ง1, ๐ท2) โ โ2, ๐ฆ1, ๐ฆ2 โฅ 0, (3.11)
โ (๐น (๐ฅ1, ๐ฅ2)) = ๐(๐ฅ1, ๐ฅ2), ๐ฅ1, ๐ฅ2 โ โ2, (3.12)
where the second equation should be understood as โ ๐ท2 being applied from theright due to the non-commutativity of the complex units ๐ and ๐.
The solution is given by
๐น (๐ง1, ๐ท2) =1
4๐2
โซโ2
1
(๐1 โ ๐ง1)(๐2 โ ๐ท2)๐(๐1, ๐2)d๐1d๐2, (3.13)
11. Generalized Analytic Signals in Image Processing 227
so that we get from the PlemeljโSokhotzki formulae
tr๐น (๐ฅ1, ๐ฅ2) =1
4(๐ผ + ๐๐ป1)(๐ผ + ๐๐ป2)๐(๐ฅ1, ๐ฅ2) (3.14)
=1
4(๐ + ๐โ1๐ + ๐โ2๐ + ๐โ๐ ๐)(๐ฅ1, ๐ฅ2). (3.15)
While (3.15) is a quaternionic-valued function, it still corresponds to a boundaryvalue of a function holomorphic in two variables. For the representation in theFourier domain one has to keep in mind that one has to apply one Fourier transformwith respect to the complex plane in ๐, and one Fourier transform with respect tothe complex plane generated by ๐. Taking into account that ๐๐ = โ๐๐ one arrivesat the so-called quaternionic Fourier transform [16, 6]:
๐ฌโฑ๐ =
โซโ2
๐๐๐ฅ1๐1๐(๐ฅ1, ๐ฅ2)๐๐๐ฅ2๐2d๐ฅ1d๐ฅ2, (3.16)
and the following representation in Fourier symbols
๐ฌโฑ(tr๐น )(๐ข1, ๐ข2) = (1 + sign๐ข1)(1 + sign๐ข2)๐ฌโฑ๐(๐ข1, ๐ข2). (3.17)
3.1.2. Image Analysis. In image analysis problems, we can introduce the followingfeatures according to [13]
Amplitude. The local amplitude of a multidimensional analytic signal is definedin a similar way as in the one-dimensional case:
๐ด๐ด(๐ฅ, ๐ฆ) =โโฃ๐(๐ฅ, ๐ฆ)โฃ2 + โฃโ1๐(๐ฅ, ๐ฆ)โฃ2 + โฃโ2๐(๐ฅ, ๐ฆ)โฃ2 + โฃโ๐ ๐(๐ฅ, ๐ฆ)โฃ2 (3.18)
This is also called energetic information.
Phase. The phase is a feature describing how much a vector or quaternion numberdiverge from the real axis. It is defined in a manner similar to the classical complexplane.
๐๐ด = arctan
(โโ1๐2 +โ2๐2 +โ๐ ๐2
๐
). (3.19)
This angle ๐๐ด is what is called phase or structural information.
Orientation. Because we are currently considering 2D signals (that is, images), wecan also describe orientation information, as the principal direction carrying thephase information. The imaginary plane, spanned by {๐, ๐}, is two dimensionaland therefore we can also define an angle ๐๐ด in this plane:
๐๐ด = arctan
(โ2๐
โ1๐
). (3.20)
This new angle is called the orientation of the signal or geometric information.
228 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
3.2. Using Clifford Analysis
Another approach to higher dimensions is Clifford analysis.
3.2.1. Mathematics. Here we use Clifford algebra ๐ถโ0,๐ [4]. This is the free algebraconstructed over โ๐ generated modulo the relation
๐ฅ2 = โโฃ๐ฅโฃ2๐0, ๐ฅ โ โ๐ (3.21)
where ๐0 is the identity of ๐ถโ0,๐.For the algebra ๐ถโ0,๐ we have the anti-commutation relationship
๐๐๐๐ + ๐๐๐๐ = โ2๐ฟ๐๐๐0, (3.22)
where ๐ฟ๐๐ is the Kronecker symbol. Each element ๐ฅ of โ๐ may be represented by
๐ฅ =
๐โ๐=1
๐ฅ๐๐๐. (3.23)
A first-order differential operator which factorizes the Laplacian is given bythe Dirac operator
๐ท๐(๐ฅ) =๐โ
๐=1
โ๐
โ๐ฅ๐. (3.24)
The RiemannโHilbert problem for the Dirac operator in โ3 can be stated inthe form
๐ท๐น (๐ฅ) = 0, ๐ฅ โ โ3, ๐ฅ3 > 0, (3.25)
โ (๐น (๐ฅ1, ๐ฅ2)) = ๐(๐ฅ1, ๐ฅ2), ๐ฅ1, ๐ฅ2 โ โ2. (3.26)
To solve this problem we follow the same idea as above.
๐นฮ๐ =
โซโ2
๐ฅโ ๐ฆ
โฃ๐ฅโ ๐ฆโฃ2 ๐3๐(๐ฅ1, ๐ฅ2)d๐ฅ1d๐ฅ2 (3.27)
tr๐นฮ๐ =1
2(๐ผ + ๐ฮ)๐
=1
2๐(๐ฆ1, ๐ฆ2) +
1
2
โซโ2
๐1(๐ฅ1 โ ๐ฆ1) + ๐2(๐ฅ2 โ ๐ฆ2)
โฃ๐ฅโ ๐ฆโฃ2 ๐3๐(๐ฅ1, ๐ฅ2)d๐ฅ1d๐ฅ2. (3.28)
Because quaternions โ are isomorphic to the even subalgebra ๐ถโ+0,3, i.e., allelements of the form
๐0 + ๐1๐1๐2 + ๐2๐1๐3 + ๐3๐2๐3, ๐0, ๐1, ๐2, ๐3 โ โ (3.29)
we can set ๐ = ๐1๐2 and ๐ = ๐2๐3 so that
tr๐นฮ๐ =1
2(๐ผ + ๐ฮ)๐ (3.30)
=1
2๐(๐ฆ1, ๐ฆ2) +
1
2
โซโ2
๐(๐ฅ1 โ ๐ฆ1) + ๐(๐ฅ2 โ ๐ฆ2)
โฃ๐ฅโ ๐ฆโฃ2 ๐(๐ฅ1, ๐ฅ2)d๐ฅ1d๐ฅ2. (3.31)
Up to the factor 1/2 this is the monogenic signal ๐๐ = ๐ + ๐โ1๐ + ๐โ2๐ :=๐+(๐, ๐)โ๐ of Sommer and Felsberg [10]. Here โ1, โ2 and โ denote respectively
11. Generalized Analytic Signals in Image Processing 229
the first and second component of the Riesz transform, and the Riesz transformitself [23]. Defined as Fourier multipliers, it holds:
โ๐(๐ข1, ๐ข2) =๐(๐ข1, ๐ข2)
โฅ(๐ข1, ๐ข2)โฅ2 ๐(๐ข1, ๐ข2), (3.32)
โ1๐(๐ข1, ๐ข2) =๐๐ข1
โฅ(๐ข1, ๐ข2)โฅ2 ๐(๐ข1, ๐ข2), (3.33)
โ2๐(๐ข1, ๐ข2) =๐๐ข2
โฅ(๐ข1, ๐ข2)โฅ2 ๐(๐ข1, ๐ข2), (3.34)
where โฅ(๐ข1, ๐ข2)โฅ2 =โ
๐ข21 + ๐ข2
2. An equivalent definition in the spatial domaincan be obtained by convolution with the two-dimensional Riesz kernel, i.e., for๐ = 1, 2
โ๐๐ = ๐๐ฅ๐โฅ๐ฅโฅ32
โ ๐, (3.35)
with ๐ being a constant.
3.2.2. Image Analysis. Following [10], three features can be computed, and willalso be denoted as energetic, structural and geometrical information, as alreadyintroduced for the multidimensional analytic signal.
Amplitude. The local amplitude of a monogenic signal is defined in a similarmanner as for the analytic signal:
๐ด๐ (๐ฅ, ๐ฆ) =โโฃ๐(๐ฅ, ๐ฆ)โฃ2 + โฃโ๐(๐ฅ, ๐ฆ)โฃ2 =
โ๐๐ (๐ฅ, ๐ฆ)๐๐ (๐ฅ, ๐ฆ), (3.36)
where the overbar denotes the conjugation of a quaternion.
Phase.
๐๐ (๐ฅ, ๐ฆ) = arctanโฃโ๐(๐ฅ, ๐ฆ)โฃ๐(๐ฅ, ๐ฆ)
, (3.37)
and we also have that ๐๐ denotes the angle between ๐ด(๐ฅ, ๐ฆ) and ๐๐ (in the planespanned by the two complex vectors). This yields values ๐๐ โ [โ๐/2;๐/2].
An alternative but equivalent definition is using the arccosine:
๐๐ = arccos๐
โฃ๐๐ โฃ . (3.38)
In (3.38), we have ๐๐ โ [0;๐].
Orientation. Once again, we can derive an orientation ๐๐ โ [โ๐, ๐] based on themonogenic signal which represents the direction of the phase information.
๐๐ = arctanโ2๐
โ1๐. (3.39)
We note that this definition actually only provides an orientation modulo๐. To determine the orientation respectively the direction modulo 2๐, a furtherorientation unwrapping step or sign estimation is needed [18, 5].
230 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
Representation. In the rest of this chapter the different features of images arerepresented by using different colourmaps: a grey colour map is used for the am-plitude representation, which is normalized between [0, 1], a jet colormap for thephase (running between blue and red and mapping the interval [0, ๐)) [24], and aHSV colormap for the orientation (running between red, blue and red in a peri-odic way and mapping the interval [โ๐/2, ๐/2)), as depicted in Figure 2 (left, resp.from top to bottom) [24]. As a last representation, we use a colormap accordingto the Middlebury representation [1] where the colour encodes the orientation andthe intensity is computed according to the phase (or structure) information. Thisrepresentation can be seen on the right-hand side of Figure 2.
Figure 2. Scales used for the different colour coding. (See explanationin text.)
3.3. Illustrations
We want here to illustrate the differences between the generalizations proposed.We will visually assess the characteristics of both approaches first applied to aSiemens star1 then to a checkerboard image. Both examples are interesting fortheir regularity (point symmetry for the star and many horizontal and verticalline symmetries for the checkerboard).
An example of such star is depicted in Figure 3(a). The two other imagesin the first row of Figure 3 illustrate the two components of the Riesz transform.As we can see, and we will come back to this property later, the partial Riesztransforms show in some directions behaviour similar to steered derivatives. The
1The Siemens star is a test image used to characterize the resolution of different optical and
graphical devices such as printers or computer projectors. The image is interesting as it showsmuch regularity, as well as many intrinsic one-dimensional and two-dimensional parts.
11. Generalized Analytic Signals in Image Processing 231
Figure 3. The Siemens star together with the different Riesz andHilbert transforms presented in this section. (Additional explanationin the text.)
first component tends to emphasize horizontal edges while the second one tendsto respond more to vertical ones.
The second row shows the results of applying the different Hilbert transformsto the Siemens star. The two first images represent the results of the two partialHilbert transforms and the last one depicts the result of the total Hilbert transform.We can notice the high anisotropy of these transforms at, for instance, the strongvertical respectively horizontal delineation through the centres of the images. Wecan also notice the patchy response of the total Hilbert transform.
As the Riesz kernel in polar coordinate [๐, ๐ผ] of the spatial domain reads
๐ (๐, ๐ผ) โผ 1
๐2๐๐๐ผ, (3.40)
it exhibits an isotropic behavior with respect to its magnitude. In comparison,the partial and the total Hilbert transforms both induce a strict relationship tothe orthogonal coordinate system, and therefore also the two-dimensional analyticsignal inherits this characteristic.
Next we consider local features computed according to the formulas intro-duced above. The results are depicted in Figure 4. The first row corresponds to
232 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
Figure 4. Local features computed with the monogenic signal repre-sentation (first row), and the multidimensional analytic signal (secondrow). The images are depicted in a pseudo-colour representation withamplitude: grey, phase: jet, orientation: HSV-Middlebury.
monogenic features, while the second one corresponds to analytic features. Thephase is displayed in a jet colormap, the orientation in a hue-saturation-valueHSV colormap. The last column shows the orientations with intensities weightedproportionally to the cosine of the phase. It is shown according to the Middleburyrepresentation2: strength (cosine of the phase) is encoded as an intensity value ofthe colour and the colour itself corresponds to the orientation. The main differ-ences between these two sets of features lie in the shape and in the boundaries.While monogenic features yield rather smooth boundaries, the analytic represen-tation creates abrupt changes due to its anisotropy. We remark that the phasegives reasonable insights into the structure in the images.
In comparison to the Siemens star, the checkerboard example (see Figure 5(a))shows many orthogonal features. In this case, we see that the partial Hilberttransforms give some good insights into the closeness of an edge and preservethe checkerboard structure (Figure 5(d) and 5(e)), while the Riesz transform gives
2The Middlebury benchmark for optical flow is a web resource for comparing results on optical
flow computations. The colour error representation is well suited for encoding our orientation.More information can be found in [1].
11. Generalized Analytic Signals in Image Processing 233
Figure 5. The checkerboard together with the different Riesz andHilbert transforms presented in this section.
more local responses. The total Hilbert transform acts as an accurate corner de-tector, as can be seen from its response in Figure 5(f).
When discussing the analytic and monogenic features (Figure 6) we remarkthat this effect is preserved. The Riesz transform, being well localized at the edges,does not yield many differences inside any of the squares and seems to jumpfrom one extreme to another across the edges. See in particular Figure 6(b) foran illustrative example of the phase. On the other hand, the Hilbert transformcontains more neighbourhood information and yields a smoother transition in thephase from one square to the next. These features have to be considered carefullybased on the application problem one wishes to solve.
4. The Geometric Approach
For a better understanding of signals a geometric interpretation can help. The fol-lowing considerations about complex numbers, quaternions, rotations, the unitarygroup, special unitary and special orthogonal groups, as well as the spin group, arewell known and can be found in numerous papers. We would like to suggest thebook by Lounesto [19], which provides a comprehensive knowledge of these topics.
234 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
Figure 6. Local features computed with the monogenic signal repre-sentation (first row), and the multidimensional analytic signal (secondrow). The images are depicted in a pseudo-colour representation withamplitude: grey, phase: jet, orientation: HSV-Middlebury.
The analytic signal ๐๐ด(๐ก) = ๐ด(๐ก)๐๐๐(๐ก) consists of boundary values of an analyticfunction, but the analytic signal can also be seen as a complex-valued function,where ๐๐๐(๐ก) = cos๐(๐ก) + ๐ sin๐(๐ก) has unit modulus and hence can be identifiedwith the unit circle ๐1. But there is even more. The set of unit complex numbers isa group with the complex multiplication as group operation, which is the unitarygroup ๐(1) = {๐ง โ โ : ๐ง๐ง = 1}. On the other hand a unit complex number canalso be seen as a rotation in โ2 if we identify the unit complex number with thematrix
๐ ๐ =
(cos๐ โ sin๐sin๐ cos๐
)โ ๐๐(2), (4.1)
i.e., the group of all counter-clockwise rotations in โ2. Now all this can also bedescribed inside Clifford algebras. Let us consider the Clifford algebra ๐ถโ0,2 withgenerators ๐1, ๐2. The complex numbers can be identified with all elements ๐ฅ +๐ฆ๐12, ๐ฅ, ๐ฆ โ โ, i.e., the even subalgebra ๐ถโ+0,2 of the Clifford algebra ๐ถโ0,2. The
rotation (4.1) can also be described by a Clifford multiplication. To see that, weidentify (๐ฅ, ๐ฆ) โ โ2 with ๐ฅ๐1 + ๐ฆ๐2 โ ๐ถโ0,2, and set
๐ ๐(๐ฅ, ๐ฆ)๐ = (cos ๐
2 + ๐12 sin๐2 )โ1(๐ฅ๐1 + ๐ฆ๐2)(cos
๐2 + ๐12 sin
๐2 ), (4.2)
11. Generalized Analytic Signals in Image Processing 235
where cos ๐2 + ๐12 sin
๐2 โ Spin(2) = {๐ โ ๐ถโ+0,2 : ๐ ๐ = 1}, the spin group of even
products of Clifford vectors. It is easily seen that ๐ and โ๐ in Spin(2) represent thesame rotation, which means that Spin(2) is a two-fold cover of ๐๐(2). The basis forall these interpretations is the description of complex numbers in a trigonometricway, which is possible by using a logarithm function, which is well known forcomplex numbers. All of that can be generalized into higher dimensions and hasbeen used for monogenic signals. We will start with quaternions because they arethe even subalgebra of the Clifford algebra ๐ถโ0,3.
4.1. Quaternions and Rotations
A quaternion ๐ โ โ can be written as
๐ = ๐0 + ๐ = S(๐) +V(๐) = โฃ๐โฃ ๐
โฃ๐โฃ , (4.3)
where โฃ๐โฃ is the absolute value or norm of ๐ in โ4 and๐โฃ๐โฃ โ โ1 is a unit quaternion.
Because of โฃโฃ ๐โฃ๐โฃโฃโฃ2 =
3โ๐=0
๐2๐
โฃ๐โฃ2 = 1, (4.4)
the set of unit quaternions โ1 can be identified with ๐3, the three-dimensionalsphere in โ4.
On the other hand the Clifford algebra ๐ถโ0,3 is generated by the elements๐1, ๐2 and ๐3 with ๐2
1 = ๐22 = ๐2
3 = โ1 and ๐๐๐๐+๐๐๐๐ = โ2๐ฟ๐,๐. Its even subalgebra
๐ถโ+0,3, as defined in (3.29), can be identified with quaternions by ๐1๐2 โผ ๐, ๐1๐3 โผ ๐and ๐2๐3 โผ ๐.
Furthermore,
Spin(3) = {๐ข โ ๐ถโ+0,3 : ๐ข๏ฟฝ๏ฟฝ = 1} = โ1. (4.5)
That means a unit quaternion can be considered as a spinor. Because Spin(3)is a double cover of the group ๐๐(3), rotations can be described by unit quater-nions. The monogenic signal is interpreted as a spinor in [26] and lately in [2].
4.2. Quaternions in Trigonometric Form
In this section we represent the monogenic signal in a similar manner to the ana-lytic signal. The analytic signal is a holomorphic and analytic function and there-fore connected to complex numbers. Complex numbers can be written in algebraicor trigonometric form as:
๐ง = ๐ฅ+ ๐๐ฆ = ๐๐๐๐.
The analytic signal is given by
๐ด(๐ก)๐๐๐(๐ก),
with amplitude ๐ด(๐ก) and (local) phase ๐(๐ก). We want to obtain a similar represen-tation of the monogenic signal using quaternions. A simple computation leads to
๐ = โฃ๐โฃ(
๐0โฃ๐โฃ +
๐
โฃ๐โฃโฃ๐โฃโฃ๐โฃ)
= โฃ๐โฃ (cos๐+ ๐ข sin๐),
236 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
where ๐ = arccos๐0โฃ๐โฃ and ๐ข =
๐
โฃ๐โฃ โ ๐2. (Alternatively, the argument ๐ can be
defined by the arctan.)We have to mention that this representation is different from all previous ones,specifically the vector ๐ข is a unit vector in โ3 and not a unit quaternion.We can represent the quaternion ๐ by its amplitude โฃ๐โฃ, the phase ๐ and theorientation ๐ข. Moreover,
๐ = โฃ๐โฃ ๐๐ข๐,where ๐ is the usual exponential function.
With the aid of an appropriate logarithm we can compute ๐ข๐ from๐โฃ๐โฃ = ๐๐ข๐.
Next, we want to explain the orientation ๐ข. We have already obtained that
๐ = โฃ๐โฃ (cos๐+ ๐ข sin๐),
where ๐ข =๐
โฃ๐โฃ โ ๐2 and ๐ข2 = โ1, i.e., ๐ข behaves like a complex unit. But because
of ๐ข โ ๐2 we can express ๐ข in spherical coordinates. We have
๐ข =๐1๐+ ๐2๐ + ๐3๐
โฃ๐1๐+ ๐2๐ + ๐3๐โฃ = ๐
(๐1โฃ๐โฃ +
(๐2(โ๐๐) + ๐3๐)
โฃ๐โฃ), (4.6)
and if we set cos ๐ =๐1โฃ๐โฃ we get
๐ข = ๐(cos ๐ + ๐ข sin ๐), ๐ข =๐
โฃ๐โฃ and ๐ = ๐๐3 โ ๐๐๐2. (4.7)
Because of
๐ = ๐๐3 โ ๐๐๐2 = ๐(๐3 + ๐๐2) (4.8)
and with cos ๐ =๐3โฃ๐โฃ we get that
๐ข = ๐(cos ๐ + ๐ sin ๐). (4.9)
Finally, we put everything together we obtain
๐ = ๐0 + ๐1๐+ ๐2๐ + ๐3๐ (4.10)
= โฃ๐โฃ (cos๐+ ๐ข sin๐) = โฃ๐โฃ (cos๐+ ๐(cos ๐ + ๐ข sin ๐
)sin๐
)(4.11)
= โฃ๐โฃ (cos๐+ ๐ (cos ๐ + ๐ (cos ๐ + ๐ sin ๐) sin ๐) sin๐) (4.12)
= โฃ๐โฃ (cos๐+ ๐ sin๐ cos ๐ + ๐ sin๐ sin ๐ sin ๐ + ๐ sin๐ sin ๐ cos ๐) , (4.13)
where ๐, ๐ โ [0, ๐] and ๐ โ [0, 2๐]. In case of a reduced quaternion, i.e., ๐3 = 0, asimilar computation leads to
๐ = ๐0 + ๐1๐+ ๐2๐ (4.14)
= โฃ๐โฃ (cos๐+ ๐ข sin๐) = โฃ๐โฃ (cos๐+ ๐ (cos ๐ โ ๐ sin ๐) sin๐) (4.15)
= โฃ๐โฃ (cos๐+ ๐ sin๐ cos ๐ + ๐ sin๐ sin ๐) , (4.16)
where ๐ โ [0, ๐] and ๐ โ [0, 2๐].
11. Generalized Analytic Signals in Image Processing 237
It is easily seen that ๐ can be computed by
tan ๐ =๐2๐1
โโ ๐ = arctan๐2๐1
.
If we compare that with the monogenic signal
๐๐ (๐ฅ, ๐ฆ) = ๐(๐ฅ, ๐ฆ) + ๐(โ1๐)(๐ฅ, ๐ฆ) + ๐(โ2๐)(๐ฅ, ๐ฆ)
we see that (compare with (3.39))
๐ = arctan(โ2๐)(๐ฅ, ๐ฆ)
(โ1๐)(๐ฅ, ๐ฆ)= ๐๐ (๐ฅ, ๐ฆ). (4.17)
Therefore the vector ๐ข = ๐ cos ๐+ ๐ sin ๐ can also be considered as the orientation.
4.3. Exponential Function and Logarithm for Quaternionic Arguments
The exponential function for quaternions and para-vectors in a Clifford algebra isdefined in [12] and many other papers.
Definition 4.1. For ๐ โ โ the exponential function is defined as
๐๐ :=
โโ๐=0
๐๐
๐!. (4.18)
Lemma 4.2. With ๐ข =๐
โฃ๐โฃ the exponential function can be written as
๐๐ = ๐๐0 (cos โฃ๐โฃ+ ๐ข sin โฃ๐โฃ) = ๐๐0๐๐ขโฃ๐โฃ. (4.19)
Remark 4.3. The formula
๐๐ขโฃ๐โฃ = cos โฃ๐โฃ+ ๐ข sin โฃ๐โฃ (4.20)
can be considered as a generalized Euler formula.
It is always a challenge to define a logarithm. We will use the followingdefinition.
Definition 4.4. Let ๐ข =๐
โฃ๐โฃ . Then the logarithm is defined as
ln ๐ :=
{ln โฃ๐โฃ+ ๐ข arccos
๐0โฃ๐โฃ , โฃ๐โฃ โ= 0, or โฃ๐โฃ = 0 and ๐0 > 0,
undefined, โฃ๐โฃ = 0 and ๐0 โค 0.(4.21)
Remark 4.5. A logarithm cannot be uniquely defined for โ1 because
๐๐ข๐ = cos๐ + ๐ข sin๐ = โ1, (4.22)
for all ๐ข โ ๐2.
Remark 4.6. More precisely, we can define the ๐th branch, ๐ โ โค, of the logarithmbecause cos ๐ก is a 2๐ periodic function.
238 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
Theorem 4.7 (Exponential and logarithm function).
1. For โฃ๐โฃ โ= 0 or โฃ๐โฃ = 0 and ๐0 > 0,
๐ln ๐ = ๐. (4.23)
2. For โฃ๐โฃ โ= ๐๐, ๐ โ โคโ{0} the following holds true
ln ๐๐ = ๐. (4.24)
Lemma 4.8. For ๐ โ โ1 and ๐ โ= โ1 both relations are true:
๐ln ๐ = ln ๐๐ = ๐. (4.25)
5. Applications to Image Analysis
5.1. Motivations
In several imaging applications only intensity-based images (encoded mostly ingray-scale representation) are provided. Apart from monochromatic camera im-ages, we can cite, e.g., computerized tomography images which encodes local ab-sorption inside a body, or optical coherence tomography images which representsthe back-scattering at an interface. These kinds of images directly describe naturalscenes or physical quantities. In other types of images information is encoded in-directly, e.g., in varying amplitude or frequency of fringe patterns. They are calledamplitude modulated (AM) or frequency modulated (FM) signals. Textures canbe interpreted as a trade-off between both ideas: they depict natural scenes andcan be described as generalized AM-FM signals.
To enrich the information content of a pure intensity image (i.e., imagesencoded with a single value at each pixel), we test the concept of analytic signalsin image processing.
5.2. Application to AM-FM Image Demodulation
Here we study the applicability of the monogenic signal representation to AM-FM signal demodulation, as needed for instance in interferometric imaging [18].A certain given two-dimensional signal (= an image, Figure 7(a)) exhibits bothamplitude modulations (Figure 7(b)) and frequency modulations (Figure 7(c)).The aim is to separate each component of the signal by means of monogenic signalanalysis.
The three features described in the previous section are computed and theirresults are depicted as local orientation in Figure 7(d), local amplitude in Fig-ure 7(e) and local phase in Figure 7(f).
It appears that for such AM-FM signals, the orientation is able to describethe direction of the phase modulation, while the local amplitude gives a goodapproximation of the amplitude modulation (corresponding to the energy of thetwo-dimensional signal) and the phase encodes information about the frequencymodulation (understood as the structural information).
11. Generalized Analytic Signals in Image Processing 239
Figure 7. Example of a two-dimensional AM-FM signal. The first rowshows the input ground truth image together with its amplitude and fre-quency modulations. The second row depicts the recovered orientation,amplitude and phases. The images are displayed using the conventionaljet colormap.
The next example, in Figure 8(a), shows a fringe pattern as an example of areal-world interferometric AM-FM image. The following images show the mono-genic analysis of this image. The local amplitude is depicted beneath the fringepattern (Figure 8(d)). This image gives us a coarse idea of how much structure is tobe found within a given neighbourhood. The second column illustrates the phasecalculation either on the whole image (Figure 8(b)) or only where the local ampli-tude is above a given threshold (Figure 8(e)). The two images in the last columnrepresent the monogenic orientation encoded in HSV without or with the previousmask. As we would expect, illumination changes are appearing in the amplitude,while local structures are contained in both phase and orientation features.
5.3. Application to Texture Analysis
A task of particular interest in artificial vision, is the characterization or descrip-tion of textures. The problem here is to find interesting features to describe agiven texture the best we can in order to classify it for instance [14]. The use of
240 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
Figure 8. Example of a fringe pattern and its monogenic decomposi-tion. Phase (second column) is encoded as a jet colormap and orientationas HSV. The two last images show phase and orientation masked witha binary filter set to one when the local amplitude gets over a certainthreshold.
steerable filters could both optimize feature computations and affect the classifi-cation. In other words, if we can compute good descriptive features, we can bettercharacterize a texture.
Considering textures from a more general viewpoint as approximate AM-FM signals, we examine here the use of monogenic representation for the localcharacterization of a textured object as depicted in Figure 9(a).
When looking at the local monogenic signal description (amplitude in Fig-ure 9(b), phase in Figure 9(c) and orientation in Figure 9(d)), we indeed see theserepetitive features along the textured object. Moreover, these estimated valuesseem to be robust against small imperfections in the periodicity.
5.4. Applications to Natural Image Scenes
In this part of our chapter, we want to highlight the interest of the monogenicsignal for natural images. Such images have completely different characteristicsfrom those introduced above. For instance, natural images are often embeddedin a fully cluttered background, encoded with several colour channels, and have
11. Generalized Analytic Signals in Image Processing 241
Figure 9. Example of a textured image superposed with a reliabil-ity mask together with its monogenic analysis. Regions with too littleamplitude are masked out as unreliable.
information at many different scales, etc. In practical applications one needs toapply band-pass filters before analyzing such images [10]. Note that this workconsiders only grey-scale images, but there is literature dealing with multichannelimages [3].
In the following we will describe two tasks useful for image processing. Thefirst part deals with edge detection. We will see how the Riesz transform can beused as an edge detector in images. Then we will see how the orientation estimationis useful, for instance, in computer vision tasks, and how monogenic signal analysiscan help with this, as has already been done for structure interpretation [22, 15].
5.4.1. Edge Detection. The Riesz transform acts as an edge detector for severalreasons. This becomes clear when one has a closer look at its definition as a Fouriermultiplier. Indeed, let us recall the ๐๐กโ Riesz multiplier (๐ = 1, 2, see (3.32)):
โ๐๐ = ๐๐ข๐โฃ๐ขโฃ๐, (5.1)
and we have
โ๐๐ = ๐1
โฃ๐ขโฃ โ๐๐, (5.2)
so that the Riesz transform acts as a normalized derivative operator.Another (eventually better) way to see this derivative effect is to consider the
Fourier multipliers in polar coordinates [17], as given in (3.40). Figure 10 illustratesthis behavior. The first column shows examples of natural grey-level images. Thesecond and third columns show the first and second Riesz components respectively.It appears that they act as edge detectors steered in the ๐ฅ and ๐ฆ directions. If wecompare the two Riesz components, we can see the response to different kinds ofedges.
5.4.2. Orientation Estimation of Edges. An important task in image processingand higher level computer vision is to estimate the orientation of edges. As thisis often the first step towards feature description and image interpretation (we
242 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
Figure 10. First and second components of the Riesz transform onsome natural images. Notice for instance the table leg appearing inFigure 10(f) but not in Figure 10(e), showing the directions of the com-ponents.
refer the reader to [7, 20] for some non-exhaustive surveys), one wants to have anorientation estimator which is as reliable as possible.
As stated in earlier sections, an orientation can be computed from an analyticor a monogenic signal analysis. For simplicity, let us consider the case of images,where the input function is defined on a set ๐ท โ โ2. Using polar coordinates inthe Fourier domain (๐, ๐ฝ), it holds that
โ๐ = ๐(cos๐ฝ, sin๐ฝ)๐ ๐, (5.3)
but on the other hand, we also have
โ๐ = ๐๐(cos๐ฝ, sin๐ฝ)๐ ๐, (5.4)
so that both the gradient and the Riesz operator have similar effects on the anglesin the Fourier domain.
It has been shown in [9] that using monogenic orientation estimation in-creases the robustness compared to the traditional Sobel operator. Moreover in
11. Generalized Analytic Signals in Image Processing 243
Figure 11. Local features computed by means of monogenic signal analysis.
their work Felsberg and Sommer introduced an improved version based on localneighbourhood considerations and by using the phase as a confidence value.
Figure 11 illustrates the monogenic analysis of our two test images. The firstcolumn represents the local amplitude of the image whereas the second columnshows the local monogenic phases. The last two columns illustrate the computationof the monogenic orientation. The colours are encoded on a linear periodic basisaccording to the Middlebury colour coding. The last column shows exactly thesame orientation but with the phase in the important role of intensity information.The basic idea is to keep relevant orientation only where the structural information(i.e., the phase) is high.
Note that we do not discuss here the local-zero mean property in naturalimage scenes. So, for example, background and illumination effects may influencethe procedure and will be discussed elsewhere.
6. Conclusion
In this chapter the specificity and analysis of two generalizations of the analyticsignal to higher dimensions have been detailed mathematically, based respectivelyon multiple complex analysis and Clifford analysis. It is shown that they are bothvalid extensions of the one-dimensional concept of the analytic signal. The maindifference between the two approaches is with regard to rotation invariance due
244 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
to the point symmetric definition of the sign function in the case of the mono-genic approach compared to the single orthant definition of the multidimensionalanalytic signal.
In a second part we have illustrated the analytic or monogenic analysis of im-ages on both artificial samples and real-world examples in terms of fringe analysisand texture analysis. In the context of AM-FM signal demodulation the monogenicsignal analysis yields a robust decomposition into energectic, structural and geo-metric information. Finally some ideas for the use of generalized analytic signalsin higher-level image processing and computer vision tasks were given showing thehigh potential for further research.
Acknowledgment
We thank the ERASMUS program, and gratefully acknowledge financial supportfrom the Federal Ministry of Economy, Family and Youth, and from the NationalFoundation for Research, Technology and Development. This work was furthersupported in part by the Austrian Science Fund under grant number P21496 N23.
References
[1] S. Baker, D. Scharstein, J. Lewis, S. Roth, M. Black, and R. Szeliski. A database andevaluation methodology for optical flow. International Journal of Computer Vision,92:1โ31, 2011. See also website: http://vision.middlebury.edu/flow/.
[2] T. Batard and M. Berthier. The spinor representation of images. In K. Gurlebeck,editor, 9th International Conference on Clifford Algebras and their Applications,Weimar, Germany, 15โ20 July 2011.
[3] T. Batard, M. Berthier, and C. Saint-Jean. Clifford Fourier transform for color im-age processing. In E.J. Bayro-Corrochano and G. Scheuermann, editors, GeometricAlgebra Computing in Engineering and Computer Science, pages 135โ162. Springer,London, 2010.
[4] F. Brackx, R. Delanghe, and F. Sommen. Clifford Analysis, volume 76. Pitman,Boston, 1982.
[5] T. Bulow, D. Pallek, and G. Sommer. Riesz transform for the isotropic estimation ofthe local phase of Moire interferograms. In G. Sommer, N. Kruger, and C. Perwass,editors, DAGM-Symposium, Informatik Aktuell, pages 333โ340. Springer, 2000.
[6] T. Bulow and G. Sommer. Hypercomplex signals โ a novel extension of the ana-lytic signal to the multidimensional case. IEEE Transactions on Signal Processing,49(11):2844โ2852, Nov. 2001.
[7] V. Chandrasekhar, D.M. Chen, A. Lin, G. Takacs, S.S. Tsai, N.M. Cheung,Y. Reznik, R. Grzeszczuk, and B. Girod. Comparison of local feature descriptorsfor mobile visual search. In Image Processing (ICIP), 2010 17th IEEE InternationalConference on, pages 3885โ3888. IEEE, 2010.
[8] A. Dzhuraev. On RiemannโHilbert boundary problem in several complex variables.Complex Variables and Elliptic Equations, 29(4):287โ303, 1996.
11. Generalized Analytic Signals in Image Processing 245
[9] M. Felsberg and G. Sommer. A new extension of linear signal processing for estimat-ing local properties and detecting features. Proceedings of the DAGM 2000, pages195โ202, 2000.
[10] M. Felsberg and G. Sommer. The monogenic signal. IEEE Transactions on SignalProcessing, 49(12):3136โ3144, Dec. 2001.
[11] D. Gabor. Theory of communication. Journal of the Institution of Electrical Engi-neers, 93(26):429โ457, 1946. Part III.
[12] K. Gurlebeck, K. Habetha, and W. Sprossig. Holomorphic Functions in the Planeand ๐-dimensional Space. Birkhauser, 2008.
[13] S.L. Hahn. Multidimensional complex signals with single-orthant spectra. Proceed-ings of the IEEE, 80(8):1287โ1300, Aug. 1992.
[14] R.M. Haralick, K. Shanmugam, and I.H. Dinstein. Textural features for image classi-fication. IEEE Transactions on Systems, Man and Cybernetics, 3(6):610โ621, 1973.
[15] B. Heise, S.E. Schausberger, C. Maurer, M. Ritsch-Marte, S. Bernet, and D. Stifter.Enhancing of structures in coherence probe microscopy imaging. In Proceedings ofSPIE, pages 83350Gโ83350Gโ7, 2012.
[16] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[17] U. Kothe and M. Felsberg. Riesz-transforms versus derivatives: On the relationshipbetween the boundary tensor and the energy tensor. Scale Space and PDE Methodsin Computer Vision, pages 179โ191, 2005.
[18] K.G. Larkin, D.J. Bone, and M.A. Oldfield. Natural demodulation of two-dimension-al fringe patterns. I. general background of the spiral phase quadrature transform.Journal of the Optical Society of America A, 18(8):1862โ1870, 2001.
[19] P. Lounesto. Clifford Algebras and Spinors, volume 286 of London MathematicalSociety Lecture Notes. Cambridge University Press, 1997.
[20] K. Mikolajczyk and C. Schmid. A performance evaluation of local descriptors. IEEETransactions on Pattern Analysis and Machine Intelligence, 27(10):1615โ1630, 2005.
[21] W. Rudin. Function Theory in the Unit Ball of โ๐. Springer, 1980.
[22] V. Schlager, S. Schausberger, D. Stifter, and B. Heise. Coherence probe microscopyimaging and analysis for fiber-reinforced polymers. Image Analysis, pages 424โ434,2011.
[23] E.M. Stein. Singular Integrals and Differentiability Properties of Functions, vol-ume 30 of Princeton Mathematical Series. Princeton University Press, 1970.
[24] The Mathworks, Inc. MATLAB Rโ R2012b documentation: colormap. Software doc-umentation available at: http://www.mathworks.de/help/matlab/ref/colormap.
html, 1994โ2012.
[25] J. Ville. Theorie et applications de la notion de signal analytique. Cables et Trans-mission, 2A:61โ74, 1948.
[26] D. Zang and G. Sommer. Signal modeling for two-dimensional image structures.Journal of Visual Communication and Image Representation, 18(1):81โ99, 2007.
246 S. Bernstein, J.-L. Bouchot, M. Reinhardt and B. Heise
Swanhild Bernstein and Martin ReinhardtTechnische Universitat Bergakademie FreibergFakultat fur Mathematik und InformatikInstitut fur Angewandte AnalysisD-09596 Freiberg, Germany
e-mail: [email protected]@googlemail.com
Jean-Luc BouchotJohannes Kepler UniversityDepartment of Knowledge-Based Mathematical Systems, FLLLAltenbergerstr., 69A-4040 Linz, Austria
e-mail: [email protected]
Bettina HeiseJohannes Kepler UniversityDepartment of Knowledge-Based Mathematical Systems, FLLL
and
Christian Doppler Laboratory MS-MACHCenter for Surface- and Nanoanalytics, ZONAAltenbergerstr., 69A-4040 Linz, Austria
e-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 247โ268cโ 2013 Springer Basel
12 Colour Extension of MonogenicWavelets with Geometric Algebra:Application to Color Image Denoising
Raphael Soulard and Philippe Carre
Abstract. We define a colour monogenic wavelet transform. This is basedon recent greyscale monogenic wavelet transforms and an extension to coloursignals aimed at defining non-marginal tools. Wavelet based colour image pro-cessing schemes have mostly been made by separately using a greyscale toolon every colour channel. This may have some unexpected effects on coloursbecause those marginal schemes are not necessarily justified. Here we proposea definition that considers a colour (vector) image right at the beginning ofthe mathematical definition so that we can expect to create an actual colourwavelet transform โ which has not been done so far to our knowledge. Thisprovides a promising multiresolution colour geometric analysis of images. Weshow an application of this transform through the definition of a full denoisingscheme based on statistical modelling of coefficients.
Mathematics Subject Classification (2010). Primary 68U10; secondary 15A66,42C40.
Keywords. Colour wavelets, analytic, monogenic, wavelet transforms, imageanalysis, denoising.
1. Introduction
Wavelets have been widely used for handling images for more than 20 years. Itseems that the human visual system sees images through different channels relatedto particular frequency bands and directions; and wavelets provide such decom-positions. Since 2001, the analytic signal and its 2D generalizations have broughta great improvement to wavelets [1, 8, 9] by a natural embedding of an AM/FManalysis in the subband coding framework. This yields an efficient representation of
This work is part of the French ANR project VERSO-CAIMAN.
248 R. Soulard and P. Carre
geometric structures in greyscale images thanks to a local phase carrying geomet-ric information complementary to an amplitude envelope having good invarianceproperties. So it codes the signal in a more coherent way than standard wavelets.The last and seemingly most appropriate proposition [9] of analytic wavelets forimage analysis is based on the monogenic signal [5] defined with geometric algebra.
In parallel a colour monogenic signal was proposed [3] as a mathematicalextension of the monogenic signal; paving the way to non-marginal colour toolsespecially by using geometric algebra and above all by considering a colour signalright at the foundation of the mathematical construction.
We define here a colour monogenic wavelet transform that extends the mono-genic wavelets of [9] to colour. These analytic wavelets are defined for colour 2Dsignals (images) and avoid the classical pitfall of marginal processing (greyscaletools used separately on colour channels) by relying on a sound mathematicaldefinition. We can therefore expect to handle coherent information of multireso-lution colour geometric structure, which would ease wavelet based colour imageprocessing. To our knowledge colour wavelets have not been proposed so far.
We first give a technical study of analytic signals and wavelets with the intentto popularize them since they rely on non-trivial concepts of geometric algebra,complex and harmonic analysis, as well as non-separable wavelet frames. Thenwe describe our colour monogenic wavelet transform, and finally an application tocolour denoising will be presented.
Notation:2D vector coordinates: ๐=(๐ฅ, ๐ฆ) , ๐=(๐1, ๐2) โ โ2; ๐ โ โค2
Euclidean norm: โฅ๐โฅ =โ๐ฅ2 + ๐ฆ2
Complex imaginary number: ๐ โ โ
Argument of a complex number: argConvolution symbol: โFourier transform: โฑ
2. Analytic Signal and 2D Generalization
2.1. Analytic Signal (1D)
An analytic signal ๐ ๐ด is a multi-component signal associated to a real signal ๐ tobe analyzed. The definition is well known in the 1D case where ๐ ๐ด(๐ก) = ๐ (๐ก)+๐ (โโ๐ )(๐ก) is the complex signal made of ๐ and its Hilbert transform (with โ(๐ก) = 1/๐๐ก).
The polar form of the 1D analytic signal provides an AM/FM representationof ๐ with โฃ๐ ๐ดโฃ being the amplitude envelope and ๐ = arg (๐ ๐ด) the instantaneousphase. This classical tool can be found in many signal processing books and is forexample used in communications.
Interestingly, we can also interpret the phase in terms of the signal shape,i.e., there is a direct link between the angle ๐ and the local structure of ๐ . Sucha link between a 2D phase and the local geometric structures of an image isvery attractive in image processing. That is why there were several attempts to
12. Colour Monogenic Wavelets 249
generalize it to 2D signals; and among them the monogenic signal introduced byFelsberg [5] seems the most advanced since it is rotation invariant.
2.2. Monogenic Signal (2D)
Without going beyond the strictly necessary details, we review here the key pointsof the fundamental construction of the monogenic signal, which will be necessaryto understand the colour extension.
The definition of the 1D case given above can be interpreted in terms ofsignal processing: the Hilbert transform makes a โpure ๐
2 -phase-shiftโ. But such aphase-shift is not straightforward to define in 2D (similar to many other 1D signaltools) so let us look at the equivalent complex analysis definition of the 1D analyticsignal. It says that ๐ ๐ด is the holomorphic extension of ๐ restricted to the real line.But complex algebra is limited regarding generalizations to higher dimensions. Tobypass this limitation we can see a holomorphic function as a 2D harmonic fieldthat is an equivalent harmonic analysis concept involving the 2D Laplace equationฮ๐=0. It then can be generalized within the framework of 3D harmonic fields by
using the 3D Laplace operator ฮ3=(
๐ฟ2
๐ฟ๐ฅ2 + ๐ฟ2
๐ฟ๐ฆ2 + ๐ฟ2
๐ฟ๐ง2
). The whole generalization
relies on this natural choice and the remaining points are analogous to the 1Dcase (see [5] for more details). Note that in Felsbergโs thesis [5] this construction isexpressed in terms of geometric algebra but here we have avoided this for the sakeof simplicity. Finally the 2D monogenic signal ๐ ๐ด associated to ๐ is the 3-vector-valued signal:
๐ ๐ด(๐) =
โกโขโขโขโขโฃ๐ (๐)
๐ ๐1(๐) =๐ฅ
2๐ โฅ๐โฅ3 โ ๐ (๐)
๐ ๐2(๐) =๐ฆ
2๐ โฅ๐โฅ3 โ ๐ (๐)
โคโฅโฅโฅโฅโฆ . (2.1)
Where ๐ ๐1 and ๐ ๐2 are analogous to the imaginary part of the complex 1D ana-lytic signal. Interestingly, this construction reveals the two components of a Riesztransform:
โ{๐ } = (๐ ๐1(๐), ๐ ๐2(๐)) =
(๐ฅ
2๐ โฅ๐โฅ3 โ ๐ (๐),๐ฆ
2๐ โฅ๐โฅ3 โ ๐ (๐))
, (2.2)
in the same way that the 1D case exhibits a Hilbert transform. Note that we getback to a signal processing interpretation since the Riesz transform can also beviewed as a pure 2D phase-shift. In the end, by focusing on the complex analysisdefinition of the analytic signal we end up with a convincing generalization of theHilbert transform.
Now recall that the motivation to build 2D analytic signals arises from thestrong link existing between the phase and the geometric structure. To define the2D phase related to the Riesz transform the actual monogenic signal must beexpressed in spherical coordinates, which yields the following amplitude envelope
250 R. Soulard and P. Carre
๐ ๐ด ๐ ๐(๐)
โ1 0 1 0 max 0 ๐ โ๐2 0 ๐
2
Figure 1. Felsbergโs monogenic signal associated to a narrow-bandsignal ๐ . The orientation ๐ is shown modulo ๐ for visual convenience.Phase values of small coefficients have no meaning so they are replacedby black pixels.
and 2-angle phase:
Amplitude: ๐ด =โ
๐ 2 + ๐ 2๐1 + ๐ 2๐2 , ๐ = ๐ด cos๐ ,Orientation: ๐ = arg (๐ ๐1 + ๐๐ ๐2) , ๐ ๐1 = ๐ด sin๐ cos ๐ ,1D Phase: ๐ = arccos (๐ /๐ด) , ๐ ๐2 = ๐ด sin๐ sin ๐ .
(2.3)
A monogenic signal analysis is illustrated in Figure 1.
2.3. Physical Interpretation
Felsberg shows a direct link between the angles ๐ and ๐ and the geometric localstructure of ๐ . The signal is thus expressed as an โ๐ด-strongโ (๐ด = amplitude) 1Dstructure with orientation ๐. ๐ is analogous to the 1D local phase and indicateswhether the structure is a line or an edge. A direct drawback is that intrinsically2D structures are not handled. Yet this tool has found many applications in im-age analysis, from contour detection to motion estimation (see [9] and referencestherein p. 1).
From a signal processing viewpoint the AM/FM representation provided byan analytic signal is well suited for narrowband signals. That is why it seems natu-ral to embed it within a wavelet transform that performs subband decomposition.We now present the monogenic wavelet analysis proposed by Unser in [9].
3. Monogenic Wavelets
So far there is one proposition of computable monogenic wavelets in the literature[9]. It provides 3D vector-valued monogenic subbands consisting of a rotation-covariant magnitude and a new 2D phase. This representation โ specially definedfor 2D signals โ is a great theoretical improvement over complex and quaternionwavelets [8, 1], similar to the way that the monogenic signal itself is an improve-ment over its complex and quaternion counterparts.
12. Colour Monogenic Wavelets 251
The proposition of [9] consists of one real-valued โprimaryโ wavelet transformin parallel with an associated complex-valued wavelet transform. Both transformsare linked to each other by a Riesz transform so they carry out a monogenic mul-tiresolution analysis. We end up with three vector coefficients forming subbandsthat are monogenic.
3.1. Primary Transform
The primary transform is real-valued and relies on a dyadic pyramid decompo-sition tied to a wavelet frame. Only one 2D wavelet is needed and the dyadicdownsampling is done only on the low frequency branch; leading to a redundancyof 4 : 3. The scaling function ๐๐พ and the mother wavelet ๐ are defined in theFourier domain as
๐๐พโฑโโ(4(sin2 1
2๐1 + sin2 12๐2
)โ 83 sin
2 12๐1 sin
2 12๐2
) ๐พ2
โฅ๐โฅ๐พ , (3.1)
๐(๐) = (โฮ)๐พ2 ๐2๐พ(2๐) . (3.2)
Note that ๐๐พ is a cardinal polyharmonic spline of order ๐พ and spans the space ofthose splines with its integer shifts. It also generates โ as a scaling function โ avalid multiresolution analysis.
This particular construction is made by an extension of a wavelet basis (non-redundant) related to a critically-sampled filterbank. This extension to a waveletframe (redundant) adds some degrees of freedom used by the authors to tune theinvolved functions. In addition a specific subband regression algorithm is used onthe synthesis side. The construction is fully described in [10].
3.2. The Monogenic Transform
The second โRiesz partโ transform is a complex-valued extension of the primaryone. We define the associated complex-valued wavelet by including the Riesz com-ponents
๐โฒ = โ(
๐ฅ
2๐ โฅ๐โฅ3 โ ๐(๐))
+ ๐
(๐ฆ
2๐ โฅ๐โฅ3 โ ๐(๐))
. (3.3)
It can be shown that this generates a valid wavelet basis and that it can be extendedto the pyramid described above. The joint consideration of both transforms leadsto monogenic subbands from which the amplitude and the phase can be extractedwith an overall redundancy of 4 : 1. The monogenic wavelet transform by Unseret al. is illustrated in Figure 2.
So far no applications of the monogenic wavelets have been proposed. In [9]a demonstration of AM/FM analysis is done with fine orientation estimation andgives very good results in terms of coherency and accuracy. Accordingly this toolmay be used for analysis tasks rather than for processing.
Motivated by the powerful analysis provided by the monogenic wavelet trans-form we propose now to extend it to colour images.
252 R. Soulard and P. Carre
๐ ๐ด ๐ ๐(๐)
0 1 0 max 0 ๐ โ๐2 0 ๐
2
Figure 2. Unserโs MWT of the image โfaceโ. Same graphic chart asFigure 1. We used ๐พ = 3 and the scales are ๐ โ {โ1,โ2,โ3}.
4. Colour Monogenic Wavelets
We define here our proposition that combines a fundamental generalization ofthe monogenic signal for colour signals with the monogenic wavelets describedabove. The challenge is to avoid the classical marginal definition that would applya greyscale monogenic transform to each of the three colour channels of a colourimage. We believe that the monogenic signal has a favorable theoretical frameworkfor a colour extension and this is why we propose to start from this particularwavelet transform rather than from a more classical one.
The colour generalization of the monogenic signal can be expressed withinthe geometric algebra framework. This algebra is very general and embeds thecomplex numbers and quaternions as subalgebras. Its elements are โmultivectorsโ,naturally linked with various geometric entities. The use of this fundamental toolis gaining popularity in the literature because it allows rewriting sophisticatedconcepts with simpler algebraic expressions and so paves the way to innovativeideas and generalizations in many fields.
For simplicityโs sake and since we would not have enough space to present thefundamentals of geometric algebra, we express the construction here in classicalterms, as we did above in Section 2.2. We may sometimes point out some necessaryspecific mechanisms but we refer the reader to [3, 5] for further details.
4.1. The Colour Monogenic Signal
Starting from Felsbergโs approach that is originally expressed in the geometricalgebra of โ3; the extension proposed in [3] is written in the geometric algebra ofโ5 for 3-vector-valued 2D signals of the form (๐ ๐ , ๐ ๐บ, ๐ ๐ต). By simply increasingthe dimensions we can embed each colour channel along a different axis and the
12. Colour Monogenic Wavelets 253
original equation from Felsberg involving a 3D Laplace operator can be generalizedin 5D with
ฮ5 =
(๐ฟ2
๐ฟ๐ฅ21
+๐ฟ2
๐ฟ๐ฅ22
+๐ฟ2
๐ฟ๐ฅ23
+๐ฟ2
๐ฟ๐ฅ24
+๐ฟ2
๐ฟ๐ฅ25
).
Then the system can be simplified by splitting it into three systems each witha 3D Laplace equation, reduced to be able to apply Felsbergโs condition to eachcolour channel. At this stage the importance of geometric algebra appears since analgebraic simplification between vectors leads to a 5-vector colour monogenic signalthat is non-marginal. Instead of naively applying the Riesz transform to each colourchannel, this fundamental generalization leads to the following colour monogenicsignal: ๐ ๐ด=(๐ ๐ , ๐ ๐บ, ๐ ๐ต, ๐ ๐1, ๐ ๐2) where ๐ ๐1 and ๐ ๐2 are the Riesz transform appliedto ๐ ๐ +๐ ๐บ+๐ ๐ต.
Now that the colour extension of Felsbergโs monogenic signal has been de-fined, let us construct the colour extension of the monogenic wavelets.
4.2. The Colour Monogenic Wavelet Transform
We can now define a wavelet transform whose subbands are colour monogenicsignals. The goal is to obtain vector coefficients of the form
(๐๐ , ๐๐บ, ๐๐ต, ๐๐1, ๐๐2) (4.1)
such that
๐๐1 =๐ฅ
2๐ โฅ๐โฅ3 โ (๐๐ + ๐๐บ + ๐๐ต),
๐๐2 =๐ฆ
2๐ โฅ๐โฅ3 โ (๐๐ + ๐๐บ + ๐๐ต).
It turns out that we can very simply use the transforms presented above by ap-plying the primary one on each colour channel and the Riesz part on the sum ofthe three. The five related colour wavelets illustrated in Figure 3 and forming onecolour monogenic wavelet ๐๐ด are:
๐๐ =
โโ๐00
โโ , ๐๐บ =
โโ0๐0
โโ , ๐๐ต =
โโ00๐
โโ (4.2)
๐๐1 =
โโโ๐ฅ
2๐โฅ๐โฅ3 โ ๐๐ฅ
2๐โฅ๐โฅ3 โ ๐๐ฅ
2๐โฅ๐โฅ3 โ ๐
โโโ ๐๐2 =
โโโ๐ฆ
2๐โฅ๐โฅ3 โ ๐๐ฆ
2๐โฅ๐โฅ3 โ ๐๐ฆ
2๐โฅ๐โฅ3 โ ๐
โโโ (4.3)
๐๐ด =(๐๐ , ๐๐บ, ๐๐ต, ๐๐1, ๐๐2
)(4.4)
We then get five vector coefficients verifying our conditions, and forming a colourmonogenic wavelet transform. The associated decomposition is described by thediagram of Figure 4. This provides a multiresolution colour monogenic analysismade of a 5-vector-valued pyramid transform. The five decompositions of two im-ages are shown in Figure 5 from left to right. Each one consists of four juxtaposed
254 R. Soulard and P. Carre
๐๐ ๐๐บ ๐๐ต ๐๐1 ๐๐2
Figure 3. Space representation of the 5 colour wavelets.
Figure 4. Colour MWT scheme. Each colour channel is analyzed withthe primary wavelet transform symbolized by a ๐ block and the sumโ๐ + ๐บ + ๐ตโ is analyzed with the โRiesz partโ wavelet transform (๐๐1
and ๐๐2 blocks).
image-like subbands resulting from a 3-level decomposition. We fixed ๐พ = 3, be-cause it gave good experimental results.
4.3. Interpretation
Let us look at the first three greymaps. These are the three primary transforms๐๐ , ๐๐บ and ๐๐ต where white (respectively black) pixels are high positive (respec-tively negative) values. Note that our transform is non-separable and so providesat each scale only one subband related to all orientations. We are not subjectto the arbitrarily separated horizontal, vertical and diagonal analyses usual withwavelets. This advantage is even greater in colour. Whereas marginal separabletransforms show three arbitrary orientations within each colour channel โ which isnot easily interpretable โ the colour monogenic wavelet transform provides a morecompact energy representation of the colour image content regardless of the localorientation. The colour information is well separated through ๐๐ , ๐๐บ and ๐๐ต: see,e.g., that the blue contours of the first image are present only in ๐๐ต , and in each ofthe three decompositions it is clear that every orientation is equally represented allalong the round contours. This is different from separable transforms that preferparticular directions. The multiresolution framework causes the horizontal blue
12. Colour Monogenic Wavelets 255
Images ๐๐ ๐๐บ ๐๐ต Riesz part
Figure 5. Colour MWT of images. The two components of the Rieszpart are displayed in the same figure part with the magnitude of ๐๐1+๐๐๐2encoded in the intensity, and its argument (local orientation) encodedin the hue.
low-frequency structure of the second image to be coded mainly in the third scaleof ๐๐ต.
But the directional analysis is not lost thanks to the Riesz part that completesthis representation. Now look at the โ2-in-1โ last decomposition forming the Riesz
part. It is displayed in one colour map where the geometric energyโ
๐2๐1 + ๐2๐2 isencoded into the intensity (with respect to the well-known HSV colour space) andthe orientation arg(๐๐1 + ๐๐๐2)(๐) is encoded in the hue (e.g., red is for {0, ๐} andcyan is for ยฑ๐
2 ). This way of displaying the Riesz part reveals well the providedgeometric analysis of the image.
The Riesz part gives a precise analysis that is local both in space and scale.If there is a local colour geometric structure in the image at a certain scale theRiesz part exhibits a high intensity in the corresponding position and subband.This is completed with an orientation analysis (hue) of the underlying structure.For instance a horizontal (respectively vertical) structure in the image will becoded by an intense cyan (respectively red) point in the corresponding subband.The orientation analysis is strikingly coherent and accurate. See, for example, thatcolour structures with constant orientation (second image) exhibit a constant huein the Riesz part over the whole structure.
256 R. Soulard and P. Carre
Note that low intensity corresponds to โno structureโ, i.e., where the imagehas no geometric information. It is sensible not to display the orientation (lowintensity makes the hue invisible) for these coefficients since the values have nomeaning in these cases.
In short, the colour and geometric information of the image are well separatedfrom each other, and the orientation analysis is very accurate. In addition theinvariance properties of the primary and Riesz wavelet transforms are kept in thecolour extension for a slight overall redundancy of 20 : 9 โ 2.2. This transform isnon-marginal because the RGB components are considered as well as the intensity(๐ +๐บ+๐ต) โ which involves two different colour spaces.
4.4. Reconstruction Issue
Image processing tasks such as denoising need the synthesis part of filterbanks. Inthe case of redundant representations, there are often several ways to perfectly re-construct the transformed image. However, when wavelet coefficients are processed,the reconstruction method affects the way in which the wavelet domain processingwill modify the image. In other words, we have to combine the redundant data sothat the retrieved graphical elements are consistent with the modifications thathave been done to the wavelet coefficients.
This issue occurs in the scalar case, since the pyramids we use have a redun-dancy of 4 : 3. The associated reconstruction algorithm has been well defined bythe authors in [10] and consists of using the spatial redundancy of each subbandat the synthesis stage, by using the so-called subband regression algorithm.
In our case, we have to face another kind of redundancy, which stems fromthe monogenic model. Apart from the wavelet decomposition, the monogenic rep-resentation (as well as the analytic representation) is already basically redundantsince additional signals are processed (the Riesz part). In our case, the followingdistinct reconstructions are possible:
โ We can reconstruct the whole colour image (๐ ,๐บ,๐ต) solely from the primarypart (๐๐ , ๐๐บ, ๐๐ต).
โ The Riesz part ๐๐1 + ๐๐๐2 can be used to reconstruct ๐ +๐บ+๐ตโ which is only a partial reconstruction.
โ One can also combine both reconstructions with a specific application drivenmethod.
In every case, the reconstruction is perfect. What is unknown is the meaning ofwavelet domain processing with respect to the chosen reconstruction method.
Let us now study the new colour wavelet transform from an experimentalpoint of view.
12. Colour Monogenic Wavelets 257
5. Wavelet Coefficients Study for Denoising
We propose in this section a data restoration algorithm based on the colour mono-genic wavelet transform defined above. Classical denoising methods consist in per-forming non-linear processing in the wavelet domain by thresholding the coef-ficients. However in our case, information carried by the wavelet coefficients isricher than in the usual orthogonal case. The colour monogenic wavelet decompo-sition is not orthogonal, so we have to further study modelling of this new kind ofcoefficient.
In addition, we saw that the colour monogenic decomposition is composed oftwo kinds of data, i.e.:
โ The โprimary partโ: A set of coefficients associated to colorimetric informa-tion, forming three pyramids linked to the colour channels.
โ The โRiesz partโ: A geometric measure composed of a norm and an angle ateach point, processed by the Riesz transform and giving some informationabout shape and structure.
In order to carry out an image restoration algorithm, a coefficient selection processhas to be defined.
We first experimentally characterize noise-related coefficients in the differentsubbands in order to identify their distribution and correlation. Then both thecolorimetric information and the geometric information are merged into a singlethresholding to perform a colour monogenic wavelet based denoising.
5.1. Modeling of Noise-related Coefficients
Classical wavelet based denoising techniques rely on the assumption that the distri-bution of noise-related wavelet coefficients can be modelled efficiently by a centredGaussian law. This usually implies a constant threshold over the whole transform(see [4]). Here we have a non-orthogonal transform so we need to observe the be-haviour of the noise term of a noisy image through the colour monogenic waveletanalysis. Despite the singularity of our transform, we retain the classical Gaussianmodel, a method which will be experimentally validated.
5.1.1. Primary Coefficients. As shown in Figure 6 we observe that the decompo-sition of the centred Gaussian noise with variance ๐2 = 1 remains centered andGaussian even after the decomposition, but with different variances.
Experimental values for the standard deviations are given in Table 1. Thecontinuous theoretical distributions have been plotted above the histograms (leftside of Figure 6) so as to confirm the Gaussian model:
๐ =1
๐๐ โ2๐
exp
(โ๐ฅ2
2๐2๐
)(5.1)
where ๐๐ is the estimated variance of the coefficients. The decomposition is per-formed through a set of band-pass filters, with very limited correlation between thebasis functions, which implies that we retain Gaussian signals with some degreeof independence.
258 R. Soulard and P. Carre
0.00.10.20.30.4
-5 -4 -3 -2 -1 0 1 2 3 4 5
0.0000.1170.2330.350
-5 -4 -3 -2 -1 0 1 2 3 4 5 0.0000.2330.4670.700
-5 -4 -3 -2 -1 0 1 2 3 4 5
0.00.10.20.3
-5 -4 -3 -2 -1 0 1 2 3 4 5 0.00.20.40.6
-5 -4 -3 -2 -1 0 1 2 3 4 5
0.0000.1170.2330.350
-5 -4 -3 -2 -1 0 1 2 3 4 5 0.00.20.40.6
-5 -4 -3 -2 -1 0 1 2 3 4 5
0.0000.1170.2330.350
-5 -4 -3 -2 -1 0 1 2 3 4 5 0.00.20.40.6
-5 -4 -3 -2 -1 0 1 2 3 4 5
Original noise
Primary part 1st scale
Primary part 2nd scale
Primary part 3rd scale
Primary part 4th scale
Riesz part modulus 1st scale
Riesz part modulus 2nd scale
Riesz part modulus 3rd scale
Riesz part modulus 4th scale)
Figure 6. Histograms (bars) of primary subbands and modulus ofRiesz subbands, and Probability Density Function models (line).
Table 1. Standard deviations ๐๐ of subbands after decomposition ofGaussian noise with variance 1 (with ๐พ = 3).
Scale ๐ 1(High freq.) 2 3 4Standard deviation ๐๐ 1.339 1.502 1.512 1.560
The values of ๐๐ can also be derived analytically from the definitions of thefilters. Recall that linear filtering of a stationary random signal ๐ฅ by a filter ๐ปwith output ๐ฆ can be studied with power spectral densities ฮจ๐ฅ = โฃโฑ [๐ฅ]โฃ2 (PSD)
and autocorrelations ๐ ๐ฅ(๐) = โฑโ1ฮจ๐ฅ. In particular we have ฮจ๐ฆ = ฮจ๐ฅ โฃ๐ปโฃ2. Theoutput variance ๐2
๐ฆ is equal to ๐ ๐ฆ(0) which reduces to
๐2๐ฆ =
๐2๐ฅ
4๐2
(2๐,2๐)โซโซ(0,0)
โฃ๐ป(๐)โฃ2 ๐๐.
For example, the first scale output of our filterbank is directly linked to thefirst stage high-pass filter:
๐21 =
๐2
4๐2
(2๐,2๐)โซโซ(0,0)
โฃโฃโฃโฃโฃ(4(sin2 1
2๐1 + sin2 12๐2)โ 8
3 sin2 1
2๐1 sin2 1
2๐2
)๐พ2 โฅ๐โฅ๐พ
โฃโฃโฃโฃโฃ2
๐๐ (5.2)
The remaining coefficients are tied to equivalent filters of each filterbank output.To have more realistic data, we introduce the image peppers altered by
additive Gaussian white noise (SNR = 83dB) in Figure 7.
12. Colour Monogenic Wavelets 259
Figure 7. Image peppers altered by additive whiteGaussian noise.
We illustrate in Figure 8 a histogram of ๐๐ (primary part, red channel) on thefirst scale. As in classical denoising, we assume that the first scale mainly containsnoise related to coefficients since natural images should not have substantial highfrequency content. Again, this histogram visually suggests that the Gaussian modelis justified.
To finally confirm the Gaussian model, we computed a Kolmogorov-Smirnovtest comparing experimental coefficients with a true Gaussian distribution. Thetest was positive for the image peppers with a ๐ -value of 0.9. Let us now studythe modelling of the Riesz part coefficients.
5.1.2. Riesz Part Coefficients. Recall that structure information is both carriedby an angle and a modulus. Handling and thresholding of circular data such as anangle is a difficult issue. In this exploratory work, and as a natural first step, we willconcentrate on the modulus. Moreover, the modulus is tied to certain amplitudeinformation related to geometrical structures, and for which thresholding will berelevant. Thresholding an angle is less intuitive.
The Riesz part coefficients can also be viewed as outputs of filtering pro-cesses, so we again propose to make a Gaussian assumption for their distribution.Real and imaginary parts of subbands follow centered Gaussian distributions withvariance ๐2
๐ .
Studying the modulus requires us to find the distribution law of the random
variable ๐ = (๐2 + ๐ 2)12 where (๐,๐ ) are two independent Gaussian random
variables with zero-mean and standard variation ๐๐ . In this case, it is well known
260 R. Soulard and P. Carre
Figure 8. Distribution of first scale coefficients of red primary decom-position ๐๐ .
that ๐ follows a Rayleigh distribution
๐๐(๐) =๐
๐2๐
exp(โ๐2/2๐2
๐
)1๐>0
with 1๐>0 being the Heaviside step. The moments of ๐ are then ๐ธ(๐) = ๐๐ โ
๐/2and ๐ (๐) = ๐2
๐ (2โ๐/2). The sequence of operations in the experiment of Figure 6were:
โ Generate a test colour image made of Gaussian noise only,โ analyze it through the colour monogenic wavelet transform,โ process histograms of the modulus of the primary subbands and the Rieszsubbands,
โ compute standard deviations of the primary subbands to get experimentalvalues for ๐๐ ,
โ compute theoretical distributions ๐ and ๐๐ with the measured ๐๐ ,โ plot histograms and theoretical curves in the same diagram for each scale.
We can see that the modeled Riesz-part curves โ on the right side of Figure 6 โcorrespond well to the measured histograms. Histograms of higher scales are lessregular because of the small number of coefficients in those subbands.
Note that the Rayleigh distribution would theoretically be fully justified ifstatistical independence were guaranteed between ๐ and ๐ . In our case, the goal
12. Colour Monogenic Wavelets 261
Figure 9. Experimental distribution of the Riesz-part modulus on thefirst scale.
of the modelling is simply to define a threshold so that we can keep the Rayleighmodel for the modulus of the Riesz part.
In the case of the noisy natural image peppers introduced above, we observethe first scale histogram of the Riesz part modulus in Figure 9. According to thehypothesis that the first scale contains mainly noise, the related histogram is stillRayleigh-shaped.
To further illustrate this choice, Figure 10 shows the Riesz part modulushistograms for several scales of the decomposed noisy image peppers. We cansee that the dispersion increases when useful information becomes โ with scaleโ more important than noise-related coefficients. Therefore, it is a good idea toestimate the threshold on the first scale, which will discard all noise coefficientswhile preserving structural information within other scales.
Finally, a Gaussian model according to the primary subbands and a Rayleighmodel for the modulus of Riesz subbands will be chosen. We can now apply auto-matic thresholding to perform colour denoising.
5.2. Thresholding
We propose to first focus on thresholding the primary part โ which is the mostintuitive. Thresholding will be done in the following way:
โ Estimation of noise level from the first scale (by assumption it is the solescale containing only noise).
262 R. Soulard and P. Carre
Figure 10. Experimental distribution of the Riesz-part modulus on thefirst three scales.
โ Soft thresholding of the different scales with the threshold equal to threetimes the estimated noise level.
Using the first scale to estimate the threshold is classical (see [7]). By assumingthat the first scale contains mainly noise, one can use the experimental median ofthe coefficients to estimate the standard deviation of the Gaussian noise distribu-tion well. So we have the estimate ๏ฟฝ๏ฟฝ = median(โฃ๐1โฃ)/0.6745 with ๐1 being thecoefficients of the first scale.
Recall that the soft thresholding is defined as follows:
thresholding(๐ง, ๐) = sign(๐ง)โ โฃ๐งโฃ โ ๐
โ+, (5.3)
where ๐ง is the coefficient, ๐ the threshold and โ.โ+ is max(., 0).
As already discussed in Section 4.4, about the synthesis part of the colourmonogenic filterbank, there are several ways to reconstruct the denoised image.Within our work we must ensure that the Riesz part controls the denoising process.Otherwise, this would be equivalent to marginal wavelet denoising without anyuse of the monogenic analysis. That means, it would consider the colorimetricinformation (primary part) without taking into account the geometric information(Riesz part). But how to combine both the primary and the Riesz decomposition ina unified reconstruction is an open issue. In this work we propose to use the Rieszpart jointly with the primary part in the thresholding process rather than in thefilterbank synthesis. So the filterbank synthesis will be done for the thresholdedprimary part only. Finally, the denoising scheme actually fully uses the colourmonogenic analysis.
12. Colour Monogenic Wavelets 263
Thresholding from the modulus of the Riesz coefficients is not straightforwardbecause of the Rayleigh distribution for which classical schemes do not hold. Asimilar issue is dealt with in [6] with the modulus of Gabor wavelets, where settingthe following threshold is suggested:
๐ธ(๐) + ๐ โโ
๐ (๐), (5.4)
where ๐ธ(๐) and ๐ (๐) are the moments of the Rayleigh distribution already givenabove, and estimated from the modulus of the first scale Riesz part. We proposeto use ๐ = 3 which gave satisfactory experimental results. In the context of thehistograms plotted in Figure 10, this corresponds to a threshold of 193. We cansee that it actually discards most of the noise from the first scale, while preservinghigher coefficients on the following scales.
The combination of both thresholding processes is done through a simple log-ical AND operator, i.e., at a given position, we keep the primary coefficient only ifboth the primary part and the Riesz part lie above their respective thresholds. Thisstep may be easily improved through a wider study about how to combine colori-metric information with geometric information for this kind of colour monogenicanalysis.
5.3. Experimental Results
We finally show different results of colour image restoration in Figures 11, 12,13 and 14. Note that our test images are strongly altered so as to easily revealthe particularities of different denoising methods. With less substantial levels ofnoise, the visual comparison is too difficult. We compare our approach to the twoclassical techniques that have become the reference in wavelet based denoising [2]:
โ Soft thresholding of the decimated orthogonal wavelet transform(filter Daubechies 4),
โ Hard thresholding of the undecimated wavelet transform (Daubechies 4).
Thresholds are established by the estimation of the noise variance as describedabove, whereas the low frequency subband is not modified.
We can see that the colour monogenic wavelet transform performance is in-termediate between those of decimated wavelets and of undecimated wavelets. Thelatter has a large redundancy (๐ฟ scales on an image of ๐ pixels will give (3๐ฟ+1)๐coefficients) while preserving orthogonality properties between interlaced decom-positions, which makes it one of the most efficient denoising tools at the moment.On the other hand, decimated wavelets are known to produce unwanted oscilla-tions โ the so-called pseudo-Gibbs phenomenon โ as the price to pay for their veryfast implementation.
We observe that the colour monogenic approach allows us to preserve themain colour information of a colour image with limited redundancy. However,these experimental results hint that this is not fully satisfactory, since unwantedoscillation artifacts appear due to the thresholding process. Contrary to classicalwavelets, the artifacts are smoother and more rounded, which may be thought of
264 R. Soulard and P. Carre
Figure 11. Restoration of image house. Upper row: Noisy image anddecimated wavelet based denoising. Lower row: Undecimated waveletbased denoising and proposed approach.
as visually less annoying. The โartifact shapeโ is due to the reconstruction func-tions, which is independent of the monogenic analysis. All the information is wellselected and preserved. First we can see that image discontinuities are retainedwell. Textures are substantially smoothed โ see, e.g., the mandrill image shownin Figure 14 โ which is usual whatever the denoising method. Note that no falsecolour is introduced, contrary to the decimated wavelet method. See for examplethe parrot image shown in Figure 13, where the decimated wavelet method in-troduces many coloured artifacts: green artifacts around the black area of the redparrotโs beak, yellow artifacts on the right of the beak, green artifacts on the largeyellow area on the right side of the parrot.
Given the results, it is clear that one of the possible future directions ofthis work is about the numerical scheme used to perform the monogenic wavelet
12. Colour Monogenic Wavelets 265
Figure 12. Restoration of image peppers. Upper row: Noisy image anddecimated wavelet based denoising. Lower row: Undecimated waveletbased denoising and proposed approach.
analysis, since the current version is very sensitive to coefficient modification. Theshape of the basis functions would be explored as well.
We think that the main way to improve this work is to focus on the physicalinterpretation when designing the key monogenic colour concept. According to us,this transform suffers from a lack of unity around its different components, as wellas a poor link between them and the visual features. Although the generalizationis not strictly marginal, it has a marginal style since it reduces to apply the Riesztransform to the intensity of the image. We are currently working on a new def-inition of colour monogenic wavelets, where the physical interpretation is takenmore into account, and the local geometry is studied more deeply from a vectordifferential geometry viewpoint.
266 R. Soulard and P. Carre
Figure 13. Restoration of image parrot. Upper row: Noisy image anddecimated wavelet based denoising. Lower row: Undecimated waveletbased denoising and proposed approach.
6. Conclusion
We have defined a colour extension of the recent monogenic wavelet transformproposed in [9]. This extension is non-marginal since we have taken care to con-sider a vector signal at the very beginning of the fundamental construction and itleads to a definition fundamentally different from the marginal approach. The useof non-separable wavelets together with the monogenic framework permits a goodorientation analysis, well separated from the colour information. This colour trans-form can be a great colour image analysis tool thanks to its good separation ofinformation into various data. A statistical modelling of the coefficients for thresh-olding as well as a full denoising scheme is given and compared to state-of-the artwavelet based denoising methods.
12. Colour Monogenic Wavelets 267
Figure 14. Restoration of image mandrill. Upper row: Noisy im-age and decimated wavelet based denoising. Lower row: Undecimatedwavelet based denoising and proposed approach.
Although it is not marginal the colour generalization has a marginal style,since it reduces to applying the Riesz transform to the intensity of the image. Sothe geometric analysis is done without considering the colour information and itwould be much more attractive to have a complete representation of the colourmonogenic signal in terms of magnitude and phase(s) with both colour and geo-metric interpretation. The numerical scheme used is fragile with respect to thevisual impact of modifying wavelet coefficients. Our future work includes a newdefinition of monogenic filterbanks by focusing on reconstruction.
268 R. Soulard and P. Carre
References
[1] W.L. Chan, H.H. Choi, and R.G. Baraniuk. Coherent multiscale image process-ing using dual-tree quaternion wavelets. IEEE Transactions on Image Processing,17(7):1069โ1082, July 2008.
[2] I. Daubechies. Ten Lectures on Wavelets. Society for Industrial and Applied Math-ematics, 1992.
[3] G. Demarcq, L. Mascarilla, and P. Courtellemont. The color monogenic signal: Anew framework for color image processing. application to color optical flow. In 16thIEEE International Conference on Image Processing (ICIP), pages 481โ484, 2009.
[4] D.L. Donoho. De-noising by soft-thresholding. IEEE Transactions on InformationTheory, 41(3):613โ627, 1995.
[5] M. Felsberg. Low-Level Image Processing with the Structure Multivector. PhD thesis,Christian-Albrechts-Universitat, Institut fur Informatik und Praktische Mathematik,Kiel, 2002.
[6] P. Kovesi. Image features from phase congruency. VIDERE: Journal of ComputerVision Research, 1(3):2โ26, 1999.
[7] S. Mallat. A Wavelet Tour of Signal Processing. Academic Press, third edition, 2008.First edition published 1998.
[8] I.W. Selesnick, R.G. Baraniuk, and N.G. Kingsbury. The dual-tree complex wavelettransform. IEEE Signal Processing Magazine, 22(6):123โ151, Nov. 2005.
[9] M. Unser, D. Sage, and D. Van De Ville. Multiresolution monogenic signal analysisusing the Riesz-Laplace wavelet transform. IEEE Transactions on Image Processing,18(11):2402โ2418, Nov. 2009.
[10] M. Unser and D. Van De Ville. The pairing of a wavelet basis with a mildly re-dundant analysis via subband regression. IEEE Transactions on Image Processing,17(11):2040โ2052, Nov. 2008.
Raphael Soulard and Philippe CarreXlim-SIC laboratoryUniversity of Poitiers, Francee-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 269โ284cโ 2013 Springer Basel
13 Seeing the Invisible andMaxwellโs Equations
Swanhild Bernstein
Abstract. In this chapter we study inverse scattering for Dirac operators withscalar, vector and quaternionic potentials. For that we consider factorizationsof the Helmholtz equation and related fundamental solutions; the standardGreenโs function and Faddeevโs Green function. This chapter is motivated byoptical coherence tomography.
Mathematics Subject Classification (2010). Primary 30G35; secondary 45B05.
Keywords. Optical coherence tomography, Dirac operator, inverse scattering,Faddeevโs Green function.
1. Preliminaries
Let โ be the algebra of real quaternions and ๐น the complex quaternions or bi-quaternions. The vectors ๐1, ๐2, ๐3 are the generating vectors with
๐๐๐๐ + ๐๐๐๐ = โ2๐ฟ๐๐and ๐0 the unit element. An arbitrary element ๐ โ โ is given by
๐ =3โ
๐=0
๐๐๐๐ , ๐๐ โ โ
and an arbitrary element ๐ โ ๐น by
๐ =3โ
๐=0
๐๐๐๐ , ๐๐ โ โ.
We denote by S(๐) the scalar part ๐0, by V(๐) = q =โ3
๐=1 ๐๐๐๐ the vector partand define the conjugated quaternion by
๐ = S(๐)โV(๐)
270 S. Bernstein
for a quaternion ๐ โ โ or a biquaternion ๐ โ ๐น. The algebra of quaternions is freeof zero divisors, i.e., if ๐1๐2 = 0 for ๐1, ๐2 โ โ then ๐1 = 0 or ๐2 = 0. This is nottrue for biquaternions, for example
(1 + ๐๐3)(1โ ๐๐3) = 1โ ๐2๐23 = 1โ 1 = 0 and
(๐1 + ๐๐3)(๐1 + ๐๐3) = ๐21 + ๐(๐1๐3 + ๐3๐1) + ๐2๐2
3 = โ1 + 1 = 0.
A zero divisor ๐ โ ๐น is an element ๐ โ= 0 such that there exists a ๐ โ ๐น with ๐ โ= 0and ๐๐ = 0.
Let ๐บ โ โ3 be a domain and ๐บ๐ = {๐ฅ โ โ3 : ๐ฅ โโ ๐บ} = โ3โ๐บ. A(bi)quaternion-valued function ๐ข belongs to ๐ถ(๐บ), ๐ฟ๐(๐บ), ๐ป1(๐บ) etc. when eachreal respectively complex-valued component ๐ข๐ belongs to that function space. For๐ โ โ, ๐ฟ2,๐ denotes the set of (scalar-valued) functions ๐ข such that
โฅ๐ขโฅ๐ =โฅโฅโฅ(1 + โฃ๐ฅโฃ2)๐ /2๐ขโฅโฅโฅ
๐ฟ2(โ3)<โ.
By ๐ป๐ผ, ๐ผ โฅ 0, we denote the Sobolev space of (scalar-valued) functions ๐ข suchthat
โฅ๐ขโฅ๐ป๐ผ =โฅโฅโฅ(1 + โฃ๐โฃ๐ผ/2)๏ฟฝ๏ฟฝโฅโฅโฅ
๐ฟ2(โ3)<โ,
and the weighted Sobolev spaces ๐ป๐ผ,๐ , ๐ โ โโฅโฅโฅ(1 + โฃ๐ฅโฃ๐ /2 ๐ข)โฅโฅโฅ๐ป๐ผ
<โ.
Furthermore, we identify a vector k โ โ3 with the bi-quaternion
๐1๐1 + ๐2๐2 + ๐3๐3.
We will denote the vector and the element in ๐น or โ with k. By โ we denote theinner product in โ3
k โ k =
3โ๐=1
๐2๐ ,
whereas
kk = k2 = โ3โ
๐=1
๐2๐
is the product of the quaternion k with itself. By ๐ท we denote the Dirac operator
๐ท =
3โ๐=1
๐๐โ
โ๐ฅ๐=
3โ๐=1
๐๐โ๐ .
In particular, we have for a vector field u = (๐ข1, ๐ข2, ๐ข3) โผ ๐ข1๐1 + ๐ข2๐2 + ๐ข3๐3:
๐ทu =
(โ divucurlu
),
13. Seeing the Invisible 271
where
divu = โ โ u =
3โ๐=1
โ๐๐ข๐ and curlu = โร u =
โโ โ2๐ข3 โ โ3๐ข2
โ3๐ข1 โ โ1๐ข3
โ1๐ข2 โ โ2๐ข1
โโ .
Because the multiplication in the algebra of quaternions is not commutative weintroduce two different multiplication operators with vectors k โ โ or k โ ๐น:
k๐๐ = k ๐ but ๐k๐ = ๐ k, ๐ โ ๐น.
2. Motivation
2.1. Optical Coherence Tomography
This chapter is motivated by optical coherence tomography (OCT). Methods intomography usually use diffraction of beams to reconstruct an image. OCT isdifferent because it uses the interference of light waves. Therefore the mathematicalmodel is given by scattering of waves, see for example [7] and [8].
In this section we present the mathematical treatment of single OCT fol-lowing [7]. Unscattered photons like x-rays and ๐พ-rays have been used to obtaintomographic ray projections for a long time. The mathematical problem of recon-structing a function from its straight ray projections has already been presentedby Radon in 1917 [31]. Its solution, the Fourier slice theorem, shows that someof the three-dimensional Fourier data of the object can be obtained from two-dimensional Fourier transforms of its projections.Optical tomography techniques, in particular, OCT, deviate in several respectsfrom the better known coherence tomography (CT) concept.
โ Diffraction optical tomography (DOT) uses highly diffracted and scatteredradiation, straight ray propagation can only be assumed for a fraction of thephotons.
โ OCT images are synthesized from a series of adjacent interferometric depth-scans performed by a straight propagating low-coherence probing beam. Thisleads to an advantageous decoupling of transversal resolution from depthresolution.
โ OCT uses backscattering, i.e., light propagates twice through the same ob-ject.
Let us consider a weakly inhomogeneous sample illuminated by the Rayleigh lengtharound the beam waist where we can assume plane-wave illumination with incidentwaves:
๐ (๐)(r, k, ๐ก) = ๐ด(๐)๐๐k(๐)โ rโ๐๐๐ก,
k(๐) is the wave vector of the illumination wave, โฃk(๐)โฃ = ๐ = 2๐/๐ the wavenumber. Using the outgoing free-space Greenโs function
๐น๐(r, rโฒ) =
๐๐๐(rโrโฒ)
4๐ โฃrโ rโฒโฃ
272 S. Bernstein
of the Helmholtz operator, the first-order Born approximation yields the scatteredwave as an approximate solution of the Helmholtz equation
๐๐ (r,k(๐ ), ๐ก) = ๐ (๐)(r,k(๐), ๐ก) +
โซ๐ (โฒ)
๐ (๐)(r,k(๐), ๐ก) โ ๐น๐ (rโฒ, ๐) โ ๐น๐(r, r
โฒ) ๐rโฒ.
k(๐ ) is the wave vector of the scattered wave, โฃk(๐ )โฃ = ๐. This integral is extendedover wavelets originating from the illuminated sample volume ๐ (r(โฒ)). The relativeamplitudes of these wavelets are determined by the scattering potential of thesample
๐น๐ (r, ๐) = ๐2[๐2(r, ๐)โ 1],
where ๐ is the complex refractive index distribution of the sample structure:
๐(r) = ๐(r)[1 + ๐๐ (r)],
with ๐(r) being the phase refractive index, and ๐ (r) the attenuation index. In OCT,backscattered light originating from the coherently illuminated sample volume isdetected at a distance ๐ much larger than the linear dimensions of that volume.Therefore
๐๐ (r,k, ๐ก) =๐ด(๐)
4๐ ๐๐๐k
(๐ )โ rโ๐๐๐กโซ๐ (โฒ)
๐น๐ (rโฒ)๐โkโ rโฒ ๐rโฒ = ๐ด๐ (r,k
(๐ ), ๐ก)๐๐k(๐ )โ rโ๐๐๐ก,
where the amplitude ๐ด(๐) of the illuminating wave has been assumed constantwithin the coherent probe volume.
3. Helmholtz Equation and Maxwellโs Equation
The propagation of waves can be described by Maxwellโs equations which consti-tute a first-order differential system. That system can be reduced to second-orderdifferential equations. Due to their importance, Maxwellโs equations under dif-ferent boundary conditions, radiation conditions and other properties are widelystudied.
In this chapter we want to discuss the scattering problem for Maxwellโs equa-tions in terms of the Dirac operator. The description of Maxwellโs equations withthe Dirac operator and Clifford algebras as well as quaternions has been studiedin several papers ([12, 16, 15, 23, 3, 11, 10, 27]). Usually, time harmonic equationsin homogeneous media were investigated. Inhomogeneous media and in particularchiral media are considered in [15] and [19].
The Helmholtz equation and its application in acoustics and electromagneticshave been widely investigated (see for example [30, 32, 25]). Also some kind ofDirac equation and scattering has been considered [13].
We will study some scattering problems in connection with the Dirac-typeoperator ๐ท +๐ ๐k, where k is a vector, described with the aid of quaternions.
This problem is also motivated by so-called Lax pairs and the Abiowitz, Kaup,Newell and Segur (AKNS) method (see [1]). In very simple terms this methodmeans that a nonlinear partial differential equation can be written as a pair of
13. Seeing the Invisible 273
linear equations where one operator describes the spectral problem and the otheroperator is the operator governing the associated time evolution. Most knownmethods use the complex โ-operator and can be reduced to a scalar or matrixRiemannโHilbert problem [9]. But it can also be done with a Dirac operator. Thefactorization of the Helmholtz equation is also an important tool, because we haveto consider the operator ๐ท +๐ ๐k and the scattering data.
4. Maxwellโs Equations
The well-known Maxwellโs equations are usually used to describe electromagneticphenomena. In optics the interaction of light with a medium is characterized bythese equations. Maxwellโs equations in optics can be found in [28]. Under theassumption of an isotropic material that obeys Ohmโs law for electric conductionand can act in a para- or diamagnetic manner, Maxwellโs equationsare as follows:
curlE = โโB
โ๐ก, divD = ๐,
curlH = ๐E+ ๐โE
โ๐ก, divB = 0,
where E is the electric field, H the magnetic field, D the electric induction, B themagnetic induction, ๐ the charge density not due to polarization of the medium, ยตthe permeability, ๐ describes the electric conductivity and ๐ the permittivity of themedium. The permittivity, permeability and electric conductivity are parametersthat are related to material properties of the medium and do not change withtime. In optics ยต โ 1 can be used. If we consider a perfectly transparent insulatorthen its electric conductivity and its external charge density can be taken to bezero.
Maxwellโs equations are considered together with so-called constitutive rela-tions, which describe the relation between induction vectors and field vectors:
D = ๐E and B = ยตH.
Interference occurs when radiation follows more than one path from its source tothe point of detection. The simplest example of interference is that between planewaves. We restrict our consideration to plane waves, which means that we canconsider time-harmonic Maxwellโs equations, i.e.,
E(๐ฅ, ๐ก) = Re (E(๐ฅ)๐โ๐๐๐ก), H(๐ฅ, ๐ก) = Re (H(๐ฅ)๐โ๐๐๐ก).
If we take everything into account we obtain the system
curlE = ๐๐ยตH, div(๐E) = 0
curlH = โ๐๐๐E, div(ยตH) = 0.
274 S. Bernstein
5. Maxwellโs Equations Written as Dirac Equations
5.1. Quaternionic Formulation
Let ๐0 be the free space permittivity and ยต0 be the free space permeability, then1โยต0๐0
= ๐, the speed of light, and ๐ = ๐โยต๐ is the wave number. We further
introduce the complex refractive index ๐ by
๐2 = ยต๐๐๐ =
(ยต
ยต0
)(๐
๐0
),
where ยต๐ and ๐๐ are called relative permittivity and permeability.
๐ = ๐+ ๐๐ ,
where the real part ๐ is the conventional refractive index, whereas ๐ is called theextinction coefficient, which describes the attenuation of the electric field in themedium. Maxwellโs equations, medium properties in optics, and their mathemat-ical description are studied in [28].
Remark 5.1. Force-free magnetic fields are an important special solution of non-linear equations of magnetohydrodynamics. They are characterized by
divB = 0 and curlB+ ๐ผ(๐ฅ)B = 0,
where ๐ผ(๐ฅ) is a scalar-valued function. This system is equivalent to
๐ท๐ผ๐ต = 0. (5.1)
This relation between force-free fields and (5.1) can be found in [17].
Remark 5.2. In case of static electric and magnetic fields in an inhomogeneousmedium we obtain the decoupled system
๐ท๐E = โ ๐โ๐, ๐ทยตH =
โยตj.
A very elegant quaternionic reformulation of the Maxwellโs equations (see[14]) in inhomogeneous media can now be obtained by the substitution
โฐ :=โ๐E and โ :=
โยตH,
which leads to the system {๐ท๐๐๐โฐ = โ๐๐โโ ๐โ
๐ ,
๐ท๐โ = ๐๐โฐ +โยต๐,
where ๐๐๐ =grad
โ๐โ
๐ (๐ =grad
โยตโ
ยต) and ๐ท๐ = ๐ท + ๐ (๐ทยต = ๐ท + ยต). This system
can also be written using matrices:(๐ท๐๐๐ โ๐๐๐๐ ๐ท๐
)( โฐโ)
=
[(๐ท โ๐๐๐๐ ๐ท
)+
(๐๐๐ 00 ๐
)]( โฐโ)
=
(โ ๐โ
๐โยตj
).
13. Seeing the Invisible 275
6. Factorization of the Wave Equation
Using vector and scalar potentials, Maxwellโs equations can be rewritten as second-order differential equations, that, in turn can be factorized with the help of Diracoperators:
1
๐2โ2
โ๐ก2โฮ =
(1
๐
โ
โ๐ก+ ๐๐ท
)(1
๐
โ
โ๐ก+ ๐๐ท
),
where ๐ โ โ is the complex unit. If we just take the time-harmonic case, i.e., thetime dependence is given by the expontial ๐๐๐๐ก, differentiation by ๐ก is simply givenby multiplication with ๐๐ and we can consider the operator
โ๐2
๐2โฮ =
(๐ท +
๐
๐
)(๐ท โ ๐
๐
).
These factorizations show the deep connection between the Helmholtz equationand Maxwellโs equations. But there is yet another factorization. Let k be a vector,then
(๐ท +๐k)(๐ท โ๐k) = (๐ท โ๐k)(๐ท +๐k) = โฮโ k2 = โฮ+ โฃkโฃ2 ,and a factorization of the Helmholtz equation arises from
(๐ท +๐ ๐k)(๐ท โ๐ ๐k) = โฮโ ๐2k2 = โฮโ โฃkโฃ2 .
7. Scattering
We will need some properties of the fundamental solution of the Helmholtz equa-tion.
Proposition 7.1 ([5]). Let
๐น๐(๐ฅโ ๐ฆ) := โ ๐๐๐โฃ๐ฅโ ๐ฆโฃ
4๐ โฃ๐ฅโ ๐ฆโฃ , ๐ฅ, ๐ฆ โ โ3, ๐ฅ โ= ๐ฆ.
๐น๐(โ , ๐ฆ) solves the Helmholtz equation ฮ๐ข + ๐2๐ข = 0 in โ3โ{๐ฆ}, and ๐ฆ โ ๐บ forevery bounded subset ๐บ โ โ3.
๐น๐(๐ฅ โ ๐ฆ) = โ ๐๐๐โฃ๐ฅโฃ
4๐ โฃ๐ฅโฃ๐โ๐๐๏ฟฝ๏ฟฝโ ๐ฆ +O(โฃ๐ฅโฃโ2),
๐ท๐ฅ๐น๐(๐ฅ โ ๐ฆ) = (๐ท๐ฅ๐น๐(๐ฅ))๐โ๐๐๏ฟฝ๏ฟฝโ ๐ฆ +O(โฃ๐ฅโฃโ2),
uniform in ๏ฟฝ๏ฟฝ = ๐ฅโฃ๐ฅโฃ โ ๐2 and ๐ฆ โ ๐บ for every bounded subset ๐บ โ โ3.
Let ๐ = const and โ (๐) โฅ 0. Then application of โ๐ทโ๐ to the fundamentalsolution of the Helmholtz operator gives
โ๐ทโ๐๐น๐(๐ฅ) = ๐ถ๐(๐ฅ) =
(๐ +
๐ฅ
โฃ๐ฅโฃ2 โ ๐๐๐ฅ
โฃ๐ฅโฃ
)๐๐๐โฃ๐ฅโฃ
4๐.
This fundamental solution was obtained in [20].
276 S. Bernstein
Proposition 7.2 ([4]). Let ๐บ be a bounded Lipschitz domain with boundary โ๐บ andoutward pointing unit normal n. Let u โ ๐ป1(๐บ). If u โ ker๐ท๐(๐บ), โ (๐) โฅ 0, then
u(๐ฅ) = ๐๐[u](๐ฅ) = โโซโ๐บ
๐ถ๐(๐ฅ โ ๐ฆ)n(๐ฆ)u(๐ฆ)๐๐ฆ.
If u โ ker๐ทk = ker(๐ท +๐k), then
u(๐ฅ) = ๐k[u](๐ฅ) = โโซโ๐บ
(โ๐ท๐ฅ๐น๐)(๐ฅโ ๐ฆ)n(๐ฆ)u(๐ฆ) + n(๐ฆ)u(๐ฆ)๐น๐(๐ฅโ ๐ฆ)k ๐๐ฆ,
= โโซโ๐บ
(๐ฅโ ๐ฆ
โฃ๐ฅโ ๐ฆโฃ2 โ ๐๐(๐ฅโ ๐ฆ)
โฃ๐ฅโ ๐ฆโฃ
)๐๐๐โฃ๐ฅโ ๐ฆโฃ
4๐ โฃ๐ฅโ ๐ฆโฃn(๐ฆ)u(๐ฆ)
โ ๐๐๐โฃ๐ฅโ ๐ฆโฃ
4๐ โฃ๐ฅโ ๐ฆโฃn(๐ฆ)u(๐ฆ)k ๐๐ฆ,
where ๐ =โk2 โ โ is chosen such that โ (๐) โฅ 0.
Proposition 7.3 (Radiation condition [18]). Let ๐ โ ๐ป1loc(๐บ
๐), ๐ โ ker๐ท๐(๐บ๐) and
๐ satisfy the radiation condition(๐ โ ๐ฅ
โฃ๐ฅโฃ2 + ๐๐๐ฅ
โฃ๐ฅโฃ
)๐(๐ฅ) = o(โฃ๐ฅโฃโ1) as โฃ๐ฅโฃ โ โ,
then
๐(๐ฅ) = โ๐๐[๐ ](๐ฅ) โ๐ฅ โ ๐บ๐.
If ๐ satisfies the radiation condition
๐๐(๐ฅ) +๐๐ฅ
โฃ๐ฅโฃ๐(๐ฅ)k = o(โฃ๐ฅโฃโ1) as โฃ๐ฅโฃ โ โ,
where ๐ =โk2 and โ (๐) โฅ 0, then
๐(๐ฅ) = โ๐k[๐ ](๐ฅ), for all ๐ฅ โ ๐บ๐.
If k is a zero divisor we suppose additionally ๐(๐ฅ) = o(โฃ๐ฅโฃโ1).
Remark 7.4. It looks as if the condition for a scalar ๐ can be written without theterm ๐ฅ/ โฃ๐ฅโฃ2 which apparently gives a faster decay. But this is not true because(1 + ๐๐ฅ/ โฃ๐ฅโฃ) is a zero divisor. See also [18].
Remark 7.5. This proposition is true for a quaternion-valued function, but wewill restrict ourselves to vector fields u, i.e., quaternion-valued functions with zeroscalar part.
8. The Scattering Problem
In this section we want to consider the scattering problem for the operator๐ท+๐k.The case ๐ท + ๐ can be treated similarly.
13. Seeing the Invisible 277
Statement of the problem: Let m(๐ฅ) be a quaternion-valued potential with com-pact support = ๐บ. Let k be a vector which will be identified with ๐1๐1 + ๐2๐2 +
๐3๐3 โ โ and ๐ =โk2 with โ (๐) โฅ 0.
Further, let u๐(๐ฅ) be a solution of
๐ทu๐(๐ฅ) + u๐(๐ฅ)k = 0 in โ3. (8.1)
The scattering problem then consists in determining a scattering solution u๐ (๐ฅ)such that
๐ทu+ uk = um(๐ฅ)k in โ3, (8.2)
u = u๐ + u๐ , (8.3)
and u๐ fullfils the radiation condition
๐u๐ (๐ฅ) +๐๐ฅ
โฃ๐ฅโฃu๐ (๐ฅ)k = o(โฃ๐ฅโฃโ1
), as โฃ๐ฅโฃ โ โ. (8.4)
If k is a zero divisor we additionally assume u๐ (๐ฅ) = o(1).
Remark 8.1. The unknown vector field m with compact support could also bequaternion-valued, i.e., it could have a non-zero scalar part. We writem to indicatethat we consider a scalar- or vector- or quaternion-valued function and not only ascalar-valued function.
Theorem 8.2. If u is a solution of the scattering problem (8.1)โ(8.4) then u โฃ๐บsolves the LippmannโSchwinger integral equation
u(๐ฅ)=u๐(๐ฅ)โโซ๐บ
(โ๐ท๐ฅ๐น๐(๐ฅโ ๐ฆ)u(๐ฆ))m(๐ฆ)k+u(๐ฆ)m(๐ฆ)k๐น๐(๐ฅโ ๐ฆ)k ๐๐ฆ. (8.5)
Conversely, if u is a solution of the LippmannโSchwinger equation then u is asolution of the scattering problem.
Proof. Let u be a solution of the scattering problem and v the integral on theright-hand side of (8.5). Because
๐ทv โ vk = um(๐ฅ)k = ๐ทu+ u(๐ฅ)k,
w = uโ v satisfies๐ทw +wk = 0 in โ3.
Furthermore,
w(๐ฅ) = (u๐(๐ฅ)โ u(๐ฅ))
โโซ๐บ
(โ๐ท๐ฅ๐น๐(๐ฅโ ๐ฆ))u(๐ฆ)m(๐ฆ)k+ u(๐ฆ)m(๐ฆ)k๐น๐(๐ฅโ ๐ฆ)k ๐๐ฆ,
๐ฅ โ โ3, and it satisfies the radiation condition which will be seen as follows:u๐(๐ฅ) โ u(๐ฅ) = u๐ (๐ฅ), which satisfies the radiation condition by assumption. Letus now consider the integral. First we look at the kernels
๐น๐(๐ฅโ ๐ฆ) = ๐น๐(๐ฅ)๐โ๐๐๏ฟฝ๏ฟฝโ ๐ฆ +O(โฃ๐ฅโฃโ2) and
(๐ท๐ฅ๐น๐)(๐ฅ โ ๐ฆ) = (๐ท๐ฅ๐น๐)(๐ฅ)๐โ๐๐๏ฟฝ๏ฟฝโ ๐ฆ +O(โฃ๐ฅโฃโ2
).
278 S. Bernstein
We conclude that
(โ๐ท๐ฅ๐น๐(๐ฅโ ๐ฆ)u(๐ฆ))m(๐ฆ)k+ u(๐ฆ)m(๐ฆ)k๐น๐(๐ฅโ ๐ฆ)k
=((โ๐ท๐ฅ๐น๐(๐ฅ)u(๐ฆ))m(๐ฆ)k + ๐2u(๐ฆ)m(๐ฆ)๐น๐(๐ฅโ ๐ฆ)
)๐โ๐๐๏ฟฝ๏ฟฝโ ๐ฆ +O(โฃ๐ฅโฃโ2
).
Therefore it is enough to consider
(โ๐ท๐ฅ๐น๐(๐ฅ)u(๐ฆ))m(๐ฆ)k + ๐2u(๐ฆ)m(๐ฆ)๐น๐(๐ฅโ ๐ฆ).
We obtain
๐((โ๐ท๐ฅ๐น๐(๐ฅ)u(๐ฆ))m(๐ฆ)k + ๐2u(๐ฆ)m(๐ฆ)๐น๐(๐ฅโ ๐ฆ)
)+
๐๐ฅ
โฃ๐ฅโฃ((โ๐ท๐ฅ๐น๐(๐ฅ)u(๐ฆ))m(๐ฆ)k + ๐2u(๐ฆ)m(๐ฆ)๐น๐(๐ฅโ ๐ฆ)
)k
+ ๐((โ๐ท๐ฅ๐น๐(๐ฅ)u(๐ฆ))m(๐ฆ)k + ๐2u(๐ฆ)m(๐ฆ)๐น๐(๐ฅโ ๐ฆ)
)+
๐๐ฅ
โฃ๐ฅโฃ((โ๐ท๐ฅ๐น๐(๐ฅ)u(๐ฆ))m(๐ฆ)k2 +
๐๐ฅ
โฃ๐ฅโฃ๐2u(๐ฆ)m(๐ฆ)๐น๐(๐ฅโ ๐ฆ)k
)=
(๐(โ๐ท๐ฅ๐น๐)(๐ฅ) +
๐๐2๐ฅ
โฃ๐ฅโฃ ๐น๐(๐ฅ)
)u(๐ฆ)m(๐ฆ)k
+ ๐2
(๐๐ฅ
โฃ๐ฅโฃ (โ๐ท๐ฅ๐น๐)(๐ฅ) + ๐๐น๐(๐ฅ)
)u(๐ฆ)m(๐ฆ)
= โ(
๐๐ฅ
โฃ๐ฅโฃ3 โ ๐๐2 ๐ฅ
โฃ๐ฅโฃ2 + ๐๐2 ๐ฅ
โฃ๐ฅโฃ2)
๐๐๐โฃ๐ฅโฃ
4๐u(๐ฆ)m(๐ฆ)k
โ ๐2
(๐๐ฅ2
โฃ๐ฅโฃ4 + ๐๐ฅ2
โฃ๐ฅโฃ3 + ๐1
โฃ๐ฅโฃ
)๐๐๐โฃ๐ฅโฃ
4๐u(๐ฆ)m(๐ฆ)
= โ๐๐ฅ๐๐๐โฃ๐ฅโฃ
4๐ โฃ๐ฅโฃ3 u(๐ฆ)m(๐ฆ)kโ ๐2 ๐๐ฅ2๐๐๐โฃ๐ฅโฃ
4๐ โฃ๐ฅโฃ4 u(๐ฆ)m(๐ฆ)
= O(โฃ๐ฅโฃโ2). โก
Remark 8.3. We can replace the region of integration by any domain ๐บ such thatthe support of m is contained in ๐บ
Using the unique continuation principle (8.5) we will prove that the homo-geneous equation has only the trivial solution and by the Fredholm theory theexistence of a solution for functions in ๐ถ(๐บ) or ๐ฟ๐(๐บ), 1 < ๐ < โ. Then thesolution of the LippmannโSchwinger equation can be computed and we obtain thesolution u from
u(๐ฅ) = u๐(๐ฅ) โโซ๐บ
(โ๐ท๐ฅ๐น๐(๐ฅ โ ๐ฆ))u(๐ฆ)m(๐ฆ)k+ u(๐ฆ)m(๐ฆ)k๐น๐(๐ฅโ ๐ฆ)k ๐๐ฆ.
Lemma 8.4. We have
โฃ๐น๐(๐ฅโ ๐ฆ)โฃ โค ๐
โฃ๐ฅโ ๐ฆโฃ and โฃ๐ท๐น๐(๐ฅโ ๐ฆ)โฃ โค ๐1
โฃ๐ฅโ ๐ฆโฃ2 +๐2
โฃ๐ฅโ ๐ฆโฃ .
13. Seeing the Invisible 279
Because the kernels of the LippmannโSchwinger equation are weakly singular,the integral operators are compact as mappings in spaces of continuous functionsas well as ๐ฟ๐, 1 < ๐ <โ, but only for bounded domains.
To get information about the solution of the LippmannโSchwinger equationwe would like to apply the Fredholm theory for compact operators. So far we knowthat the integral operator is a compact operator. Now we need to prove that thehomogeneous equation has only the trivial solution. For that we need the uniquecontinuation principle. The proof of this principle goes back to [24]. For a proofwe refer to [5, Lemma 8.5], which also uses ideas from [29] and [21].
Lemma 8.5 (Unique continuation principle, [5]). Let ๐บ be a domain in โ3 and let๐ข1, . . . , ๐ข๐ โ ๐ถ2(๐บ), be real-valued functions satisfying
โฃฮ๐ข๐โฃ โค ๐
๐โ๐=1
{โฃ๐ข๐โฃ+ โฃgrad๐ข๐โฃ} in ๐บ,
for ๐ = 1, . . . , ๐ and some constant ๐. Assume that ๐ข๐ vanishes in a neighborhoodof some ๐ฅ0 โ ๐บ for ๐ = 1, . . . , ๐. Then ๐ข๐ is identically zero in ๐บ for ๐ = 1, . . . , ๐.
To be able to apply the unique continuation principle to
(๐ท +๐k)uโm(๐ฅ)uk = 0
we apply the operator ๐ท again to obtain
๐ท((๐ท +๐k)uโm(๐ฅ)uk) = โฮu+ (๐ทu)kโ๐ท(m(๐ฅ)uk).
and we can use the equation
ฮu = (๐ทu)kโ๐ท(m(๐ฅ)uk).
The last equation equation can be rewritten as the following system
ฮu = โ(divu)k+ (curlu)ร k+ (divm)(uร k)
โ(curlm)(u โ k)โ 2
(3โ
๐=1
๐๐โ๐u
)k.
If there are constants ๐ถ1, ๐ถ2 > 0 such that
max๐ฅโ๐บ
โฃ๐๐(๐ฅ)โฃ โค ๐ถ1 and max๐ฅโ๐บ
โฃgrad๐๐(๐ฅ)โฃ โค ๐ถ2,
then we can apply the unique continuation principle to conclude that the homoge-neous LippmannโSchwinger equation over a bounded domain has only the trivialsolution.
Theorem 8.6. Let k โ โ3โ{0}. There exists a unique solution to the inverse scat-tering problem (8.1)โ(8.4) and the solution u depends continuously, with respectto the maximum norm, on the incident field u๐.
Proof. Due to the Fredholm theory it is enough to prove that the homogeneousequation has only the trivial solution. If that is proven, the integral equationis a bounded invertible operator in ๐ถ(๏ฟฝ๏ฟฝ๐ ). From this it follows that u dependscontinuously on the incident field u๐ with respect to the maximum norm. Let ๐ต๐ :=
280 S. Bernstein
{๐ฅ โ โ3 : โฃ๐ฅโฃ โค ๐ }. Due to Proposition 7.2 the function u can be represented asan integral over โ๐ต๐ , and due to the properties of ๐น๐(๐ฅโ ๐ฆ) we have u(๐ฅ) = o(1)as โฃ๐ฅโฃ โ โ. Now,
u(๐ฅ) = ๐k[๐ ](๐ฅ) =โซโ๐ต๐
(โ๐ท๐ฅ๐น๐)(๐ฅโ ๐ฆ)n(๐ฆ)u(๐ฆ) + n(๐ฆ)u(๐ฆ)๐น๐(๐ฅโ ๐ฆ)k ๐๐ฆ
โผโซโ๐ต๐
(๐ท๐ฆ๐น๐)(๐ฅโ ๐ฆ)๐ฆ
โฃ๐ฆโฃu(๐ฆ) +๐ฆ
โฃ๐ฆโฃu(๐ฆ)๐น๐(๐ฅโ ๐ฆ)k ๐๐ฆ
โผโซโ๐ต๐
๐น๐(๐ฆ)
{(1
โฃ๐ฆโฃ โ ๐๐
)u(๐ฆ) +
๐ฆ
โฃ๐ฆโฃu(๐ฆ)k}
๐๐ฆ
โผโซโ๐ต๐
๐น๐(๐ฆ)
{1
โฃ๐ฆโฃu(๐ฆ)โ ๐
(๐๐(๐ฆ) + ๐
๐ฆ
โฃ๐ฆโฃu(๐ฆ)k)}
๐๐ฆ
โผโซโ๐ต๐
๐น๐(๐ฆ)
{1
โฃ๐ฆโฃu(๐ฆ) + o(โฃ๐ฆโฃโ1)
}๐๐ฆ โ 0 as ๐ฆ โโ,
because u fulfils the radiation condition and u(๐ฆ) = o(1). We have proven thatu = 0 for โฃ๐ฅโฃ โฅ ๐ . The unique continuation principle (Lemma 8.5) implies thatu = 0 in โ3. โก
Up to now we have used the usual Greenโs function for the Helmholtz equation๐น๐(๐ฅ) and the Greenโs functions ๐ถ๐(๐ฅ) and ๐ถk(๐ฅ) for the Dirac operators ๐ท + ๐and ๐ท + ๐k. Another type of Greenโs functions are the exponentially growingGreenโs functions. These Greenโs functions were introduced by Faddeev [6] andlater on used by Nachman and Ablowitz [25], Beals and Coifman [2], Sylvesterand Uhlmann [32], Paivarinta [30] and Isozaki [13] in inverse scattering. Someinverse scattering problems for the Dirac operator are also discussed in [26] and[22].
The main idea is to consider plane waves
๐๐kโ ๐ฅ, k โ โ3, ๐ฅ โ โ3, k โ k = ๐2.
Therefore we analyze how the operators change when u is replaced by ๐๐kโ ๐ฅv. Weobtain
(ฮ + ๐2)u = (ฮ + ๐2)(๐๐kโ ๐ฅv) = ๐๐kโ ๐ฅ(ฮ + 2๐k โ โ)v,
(๐ท โ ๐k)u = (๐ท โ ๐k)(๐๐kโ ๐ฅv) = ๐๐kโ ๐ฅ (๐kv +๐ทv โ ๐kv) = ๐๐kโ ๐ฅ๐ทv
(๐ท +๐ ๐k)u = (๐ท +๐ ๐k)(๐๐kโ ๐ฅv) = ๐๐kโ ๐ฅ (๐kv +๐ทv + v(๐k)) .
In the last equation it does matter whether v is truly quaternion-valued or just avector. In the latter case we see that kv+vk = โkโ v+kรvโvโ k+vรk = โ2kโ v.Hence
(๐ท +๐ ๐k)u = (๐ท +๐ ๐k)(๐๐kโ ๐ฅv) = ๐๐kโ ๐ฅ(๐ทv โ 2๐k โ v).We will relate these new operators to the operator
โฮโ 2๐๐พ๐พ๐พ โ โ โ ๐2,
and Faddeevโs Green function.
13. Seeing the Invisible 281
8.1. Faddeevโs Green Function
The idea of Faddeev to obtain a nice Green function starts with decomposing thevector k = ๐๐๐+ ๐ก๐พ๐พ๐พ, where ๐พ๐พ๐พ โ ๐2 is an arbitrary direction and ๐๐๐ โ ๐พ๐พ๐พ = 0. If we applyฮ + ๐2 to ๐๐๐ก๐พ๐พ๐พโ ๐ฅw(๐ฅ) we obtain
(ฮ + ๐2)(๐๐๐ก๐พ๐พ๐พโ ๐ฅw(๐ฅ)) = ๐๐๐ก๐พ๐พ๐พโ ๐ฅ(โ๐ก2 + ๐2 + 2๐๐ก๐พ๐พ๐พ โ โ+ฮ)w(๐ฅ).
Let ๐2 = ๐2 โ ๐ก2, take the Fourier transform of the differential operator โฮ โ2๐๐ก๐พ๐พ๐พ โ โ โ ๐2. We then get Faddeevโs Green operator defined as
(๐๐พ๐พ๐พ(๐, ๐ง)๐)(๐ฅ) =1
(2๐)3
โซ๐๐๐ฅโ ๐
๐2 + 2๐ง๐พ๐พ๐พ โ ๐ โ ๐2๐(๐) ๐๐,
where ๐พ๐พ๐พ โ ๐2, ๐ โฅ 0, and ๐ง โ โ+ = {๐ง โ โ : โ (๐ง) > 0}. If โ (๐ง) โ= 0, (๐2 + 2๐ง๐พ๐พ๐พ โ ๐ โ ๐2)โ1 โ ๐ฟ1
loc(โ3). Therefore the integral is absolutely convergent for ๐ โ ๐ฎ.
For ๐ก โ โ, ๐๐พ๐พ๐พ(๐, ๐ก) is defined as the boundary value ๐๐พ๐พ๐พ(๐, ๐ก+ ๐0).
Proposition 8.7 ([13]). Let ๐ > 12 . Then
1. ๐๐พ๐พ๐พ(๐, ๐ง) is continuous with respect to ๐ โฅ 0, ๐พ๐พ๐พ โ ๐2, ๐ง โ โ+ except for (๐, ๐ง) =(0, 0).
2. ๐๐พ๐พ๐พ(๐, ๐ง) is analytic in ๐ง โ โ+.3. For any ๐ฟ0 > 0, there exists a constant ๐ถ > 0 such that
โฅ๐๐พ๐พ๐พ(๐, ๐ง)โฅ(๐ฟ2,๐ ,๐ป๐ผ,๐ ) โค๐ถ
(๐ + โฃ๐งโฃ)1โ๐ผ ,
with ๐+ โฃ๐งโฃ โฅ ๐ฟ0, and 0 โค ๐ผ โค 2.
We would like to have a similar Faddeevโs Green function for the operators๐ท+ ๐k = ๐ท+ ๐k๐ and ๐ท+๐ ๐k. There are differences between both cases due tothe fact that
(๐ท + ๐k)(๐ท + ๐k)u = โฮu+ ๐k๐ทu+ ๐๐ท(ku)โ k2u
= โฮu+ ๐k๐ทuโ ๐k๐ทuโ 2๐3โ
๐=1
๐๐โ๐u+ k โ ku
= (โฮโ 2๐k โ โ+ ๐2)u.
With ๐k = ๐๐ง๐พ๐พ๐พ we have
(๐ท + ๐๐ง๐พ๐พ๐พ)(๐ท + ๐๐ง๐พ๐พ๐พ)u = (โฮโ 2๐๐ง๐พ๐พ๐พ โ โ+ ๐ง2)u,
and hence
(๐ท + ๐๐ง๐พ๐พ๐พ โ ๐)(๐ท + ๐๐ง๐พ๐พ๐พ + ๐)u
= (๐ท + ๐๐ง๐พ๐พ๐พ)(๐ท + ๐๐ง๐พ๐พ๐พ)u+ (๐ท + ๐๐ง๐พ๐พ๐พ)(๐u) + ๐(๐ท + ๐๐ง๐พ๐พ๐พ)uโ ๐2u
= (โฮโ 2๐๐ง๐พ๐พ๐พ โ โ โ ๐2 + ๐ง2)u
= (โฮโ 2๐๐ง๐พ๐พ๐พ โ โ โ ๐2)u.
282 S. Bernstein
This is the way that Isozaki got a Faddeevโs Green function. We obtain
๐บ๐พ๐พ๐พ(๐, ๐ง) = (๐ท + ๐๐ง๐พ๐พ๐พ โ ๐)๐(๐, ๐ง).
The multiplication from the other side leads to a different Faddeevโs Greenfunction. In this case we have
(๐ท +๐๐ง๐พ๐พ๐พ ๐ +๐ ๐(๐ง๐พ๐พ๐พ+๐๐๐))(๐ท +๐๐ง๐พ๐พ๐พ ๐ โ๐ ๐(๐ง๐พ๐พ๐พ+๐๐๐))u = (โฮโ 2๐๐ง๐พ๐พ๐พ โ โ โ ๐2)u,
and thus with ๐k = ๐๐ง๐พ๐พ๐พ + ๐๐๐๐ we obtain
๐บ๐พ๐พ๐พ(๐, ๐ง) = (๐ท +๐๐ง๐พ๐พ๐พ ๐ โ๐ ๐k)๐(๐, ๐ง).
This easily shows that for both operators ๐บ๐พ๐พ๐พ(๐, ๐ง) the following is true.
Theorem 8.8. Let ๐ > 12 . Then
1. ๐๐พ๐พ๐พ(๐, ๐ง) is continuous with respect to ๐ โฅ 0, ๐พ๐พ๐พ โ ๐2, ๐ง โ โ+ except for (๐, ๐ง) =(0, 0).
2. ๐๐พ๐พ๐พ(๐, ๐ง) is analytic in ๐ง โ โ+.3. For any ๐ฟ0 > 0, and 0 โค ๐ผ โค 1, there exists a constant ๐ถ > 0 such that
โฅ๐๐พ๐พ๐พ(๐, ๐ง)โฅ(๐ฟ2,๐ ,๐ป๐ผ,๐ ) โค ๐ถ(๐ + โฃ๐งโฃ)๐ผ,with ๐+ โฃ๐งโฃ โฅ ๐ฟ0.
We conclude that we can also use Faddeevโs Green function to solve theinverse scattering problem and we can assume the solution u to have the structureu = ๐๐kโ ๐ฅv, which is the appropriate setting for optical coherence tomography.
Acknowledgment
I would like to thank Dr. B. Heise and the Christian Doppler Laboratory for Micro-scopic and Spectroscopic Material Characterization, Johannes Kepler UniversityLinz, Austria, for making me realize the importance of Maxwellโs equations andinverse scattering for OCT and for the opportunity to learn the physics behindthe mathematical formulae.
References
[1] M.J. Ablowitz and A.P. Clarkson. Solitons, Nonlinear Evolution Equations and In-verse Scattering, volume 149 of London Mathematical Society Lecture Note Series.Cambridge University Press, 1991.
[2] R. Beals and R.R. Coifman. Multidimensional inverse scattering and nonlinear par-tial differential equations. In F. Treves, editor, Pseudodifferential Operators and Ap-plications, volume 43 of Proceedings of Symposia in Pure Mathematics, pages 45โ70.American Mathematical Society, 1984.
[3] S. Bernstein. Factorization of the Schrodinger operator. In W. Sproรig, editor, Pro-ceedings of the Symposium Analytical and Numerical Methods in Quaternionic andClifford analysis, pages 1โ6, 1996.
[4] S. Bernstein. LippmannโSchwingerโs integral equation for quaternionic Dirac oper-ators. unpublished, available at http://euklid.bauing.uni-weimar.de/ikm2003/
papers/46/M_46.pdf, 2003.
13. Seeing the Invisible 283
[5] D. Colton and R. Kress. Inverse Acoustic and Electromagnetic Scattering Theory,volume 93 of Applied Mathematical Sciences. Springer-Verlag, Berlin, Heidelberg,New York, 1992.
[6] L.D. Faddeev. Increasing solutions of the Schrodinger equation. Doklady AkademiiNauk SSSR, 165:514โ517, 1965.
[7] A.F. Fercher. Optical coherence tomography. Journal of Biomedical Optics, 1(2):153โ173, 1996.
[8] A.F. Fercher, W. Drexler, C.K. Hitzenberger, and T. Lasser. Optical coherence to-mography โ principles and applications. Reports on Progress in Physics, 66:239โ303,2003.
[9] A.S. Fokas. A unified transform method for solving linear and certain nonlinearPDEs. Proceedings of the Royal Society A, Mathematical, Physical & EngineeringSciences, 453:1411โ1443, 1997.
[10] K. Gurlebeck and W. Sproรig. Quaternionic and Clifford Calculus for Physicists andEngineers. Wiley, Aug. 1997.
[11] K. Gurlebeck. Hypercomplex factorization of the Helmholtz equation. Zeitschrift furAnalysis und ihre Anwendungen, 5(2):125โ131, 1986.
[12] K. Imaeda. A new formulation of classical electrodynamics. Nuovo Cimento,32(1):138โ162, 1976.
[13] H. Isozaki. Inverse scattering theory for Dirac operators. Annales de lโI.H.P., sectionA, 66(2):237โ270, 1997.
[14] V.V. Kravchenko. On a new approach for solving Dirac equations with some poten-tials and Maxwellโs system in inhomogeneous media. In Operator theory: Advancesand Applications, volume 121 of Operator Theory: Advances and Applications, pages278โ306. Birkhauser Verlag, 2001.
[15] V.V. Kravchenko. Quaternionic Reformulation of Maxwellโs Equations for Inhomo-geneous Media and New Solutions. Zeitschrift fur Analysis und ihre Anwendungen,21(1):21โ26, 2002.
[16] V.V. Kravchenko. Applied Quaternionic Analysis, volume 28 of Research and Expo-sition in Mathematics. Heldermann Verlag, 2003.
[17] V.V. Kravchenko. On force-free magnetic fields: Quaternionic approach. Mathemat-ical Methods in the Applied Sciences, 28:379โ386, 2005.
[18] V.V. Kravchenko and R.P. Castillo. An analogue of the Sommerfeld radiation condi-tion for the Dirac operator. Mathematical Methods in the Applied Sciences, 25:1383โ1394, 2002.
[19] V.V. Kravchenko and M.P. Ramirez. New exact solutions of the massive Dirac equa-tion with electric or scalar potential. Mathematical Methods in the Applied Sciences,23:769โ776, 2000.
[20] V.V. Kravchenko and M.V. Shapiro. On a generalized system of Cauchy-Riemannequations with a quaternionic parameter. Russian Academy of Sciences, Doklady.,47:315โ319, 1993.
[21] R. Leis. Initial Boundary Value Problems in Mathematical Physics. John Wiley, NewYork, 1986.
284 S. Bernstein
[22] X. Li. On the inverse problem for the Dirac operator. Inverse Problems, 23:919โ932,2007.
[23] A. McIntosh and M. Mitrea. Clifford algebras and Maxwellโs equations in Lipschitzdomains. Mathematical Methods in the Applied Sciences, 22(18):1599โ1999, 1999.
[24] C. Muller. On the behavior of solutions of the differential equation ๐ฟ๐ข = ๐(๐ฅ, ๐ข)in a neighborhood of a point. Communications on Pure and Applied Mathematics,7:505โ515, 1954.
[25] A.I. Nachman and M.J. Ablowitz. A multidimensional inverse scattering method.Studies in Applied Mathematics, 71:243โ250, 1984.
[26] G. Nakamura and T. Tsuchida. Uniqueness for an inverse boundary value problem forDirac operators. Communications in Partial Differential Equations, 25:7-8:557โ577,2000.
[27] E.I. Obolashvili. Partial Differential Equations in Clifford Analysis, volume 96 ofPitman Monographs and Surveys in Pure and Applied Mathematics. Harlow: AddisonWesley Longman Ltd., 1998.
[28] K.-E. Peiponen, E.M. Vartiainen, and T. Asakura. Dispersion, Complex Analysis andOptical Spectroscopy. Springer Tracts in Modern Physics. Springer Verlag, Berlin,Heidelberg, 1999.
[29] M.H. Protter. Unique continuation principle for elliptic equations. Transactions ofthe American Mathematical Society, 95:81โ90, 1960.
[30] L. Paivarinta. Analytic methods for inverse scattering theory. In Y.K.K. Binghamand E. Somersalo, editors, New Analytic and Geometric Methods in Inverse Prob-lems, pages 165โ185. Springer, Berlin, Heidelberg, New York, 2003.
[31] J. Radon. Uber die Bestimmung von Funktionen durch ihre Integralwerte langsgewisser Mannigfaltigkeiten. Berichte uber die Verhandlungen der SachsischenAkademie der Wissenschaften (Reports on the proceedings of the Saxony Academyof Science), 69:262โ277, 1917.
[32] J. Sylvester and G. Uhlmann. A global uniqueness theorem for an inverse boundaryvalue problem. Annals of Mathematics, 125(1):153โ169, Jan. 1987.
Swanhild BernsteinTU Bergakademie FreibergInstitute of Applied AnalysisPruferstr. 9D-09596 Freiberg, Germanye-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 285โ298cโ 2013 Springer Basel
14A Generalized Windowed Fourier Trans-form in Real Clifford Algebra ๐ชโ0,๐
Mawardi Bahri
Abstract. The CliffordโFourier transform in ๐ถโ0,๐ (CFT) can be regarded as ageneralization of the two-dimensional quaternionic Fourier transform (QFT),which was first introduced from the mathematical aspect by Brackx. In thischapter, we propose the Clifford windowed Fourier transform using the kernelof the CFT. Some important properties of the transform are investigated.
Mathematics Subject Classification (2010). Primary 15A66; secondary 42B10.
Keywords. Multivector-valued function, Clifford algebra, CliffordโFouriertransform.
1. Introduction
Recently researchers have paid much attention to the generalization of the classi-cal windowed Fourier transform (WFT) using the quaternion algebra and Cliffordalgebra. The first attempt to extend the WFT to the quaternion algebra was byBulow and Sommer [6, 7]. They introduced a special case of the quaternionicwindowed Fourier transform (QWFT) known as quaternionic Gabor filters. Theyapplied these filters to obtain a local two-dimensional quaternionic phase. Theirgeneralization was obtained using the inverse (two-sided) quaternion Fourier ker-nel. A further extension of the quaternionic Gabor filter to quaternionic Gaborwavelets was introduced by Bayro-Corrochano [3] and Xi et al.. [20]. Hahn [12]constructed a Fourier-Wigner distribution of 2D quaternionic signals, which is infact closely related to the QWFT.
The WFT has been also studied in the quaternion algebra framework. Thegeneralization uses the kernel of the (right-sided) quaternion Fourier transform[18, 13] which was introduced in [2] and the kernel of the (two-sided) quaternionFourier transform recently proposed by Fu et al. in [9]. In [15], we applied the ๐ถโ๐,0Clifford windowed Fourier transform to linear time-varying systems. In this chap-ter, we continue the generalization of the quaternionic windowed Fourier transform
This work was partially supported by Bantuan Seminar Luar Negeri oleh DP2M DIKTI 2011,Indonesia.
286 M. Bahri
to the real Clifford algebra ๐ถโ0,๐ called the ๐ถโ0,๐ Clifford windowed Fourier trans-form (CWFT) which differs appreciably1 from the ๐ถโ๐,0 Clifford windowed Fouriertransform (see [17, 16]).
This chapter is organized as follows. A brief review of real Clifford algebrais given in ยง 2. ยง 3 introduces the CliffordโFourier transform in ๐ถโ0,๐ (CFT) andderives its important properties. ยง 4 discusses the basic ideas for constructing theClifford windowed Fourier transform in ๐ถโ0,๐ (CWFT) using the kernel of theCliffordโFourier transform in ๐ถโ0,๐. We show that some properties of the two-dimensional quaternionic windowed Fourier transform (see [2]) are not valid in theCWFT such as the shift property and the Heisenberg uncertainty principle.
2. Preliminaries
We will be working with real Clifford algebras. Let {๐1, ๐2, ๐3, . . . , ๐๐} be an or-thonormal vector basis of the real ๐-dimensional Euclidean vector space โ๐. Thereal Clifford algebra over โ๐ denoted by ๐ถโ0,๐ then has the graded 2๐-dimensionalbasis
{1, ๐1, ๐2, . . . , ๐๐, ๐12, ๐31, ๐23, . . . , ๐๐ = ๐1๐2 โ โ โ ๐๐}. (2.1)
Obviously, for ๐ = 2 (mod 4) the pseudoscalar ๐๐ = ๐1๐2 โ โ โ ๐๐ anti-commuteswith each basis of the Clifford algebra while ๐2๐ = โ1. The associative geometricmultiplication of the basis vectors is governed by the rules:
๐๐ ๐๐ = โ๐๐ ๐๐ for ๐ โ= ๐, 1 โค ๐, ๐ โค ๐,
๐2๐ = 1 for 1 โค ๐ โค ๐. (2.2)
An element of a Clifford algebra is called a multivector and has the following form
๐ =โ๐ด
๐๐ด๐๐ด, (2.3)
where ๐๐ด โ โ, ๐๐ด = ๐๐ผ1๐ผ2โ โ โ ๐ผ๐= ๐๐ผ1๐๐ผ2 โ โ โ ๐๐ผ๐
, and 1 โค ๐ผ1 โค ๐ผ2 โค โ โ โ โค ๐ผ๐ โค ๐with ๐ผ๐ โ {1, 2, . . . ๐}. For convenience, we introduce โจ๐โฉ๐ =
โโฃ๐ดโฃ=๐ ๐๐ด๐๐ด to
denote the ๐-vector part of ๐ (๐ = 0, 1, 2, . . . , ๐), then
๐ =
๐=๐โ๐=0
โจ๐โฉ๐ = โจ๐โฉ+ โจ๐โฉ1 + โจ๐โฉ2 + โ โ โ + โจ๐โฉ๐, (2.4)
where โจ. . .โฉ0 = โจ. . .โฉ.The Clifford conjugate ๐ of a multivector ๐ is defined as the anti-auto-
morphism for which
๐ =โ
๐ด๐๐ด๐๐ด, ๐๐ด = (โ1)๐(๐+1)/2, ๐๐ = โ๐๐, ๐ = 1, 2, 3, . . . , ๐. (2.5)
1The CWFT presented here is constructed using the kernel of the ๐ถโ0,๐ CliffordโFourier trans-form whose properties are quite different from the kernel of the ๐ถโ๐,0 CliffordโFourier transform
(see [1, 14]).
14. Clifford Windowed Fourier Transform 287
and hence
๐๐ = ๐๐ for arbitrary ๐, ๐ โ ๐ถโ0,๐. (2.6)
The scalar product of multivectors ๐ and ๐ and its associated norm is definedby, respectively,
โจ๐ ๐โฉ = ๐ โ ๐ =โ
๐ด๐๐ด๐๐ต, and โฃ๐ โฃ2 =
โ๐ด๐2๐ด. (2.7)
The product of two Clifford multivectors ๐ =โ๐
๐=1 ๐ฅ๐๐๐ and ๐ =โ๐
๐=1 ๐ฆ๐๐๐ splitsinto a scalar part and a 2-vector or so-called bivector part
๐๐ = ๐ โ ๐ + ๐ โง ๐ (2.8)
where
๐ โ ๐ =
๐โ๐=1
๐ฅ๐๐ฆ๐, and ๐ โง ๐ =
๐โ๐=1
๐โ๐=๐+1
๐๐๐๐(๐ฅ๐๐ฆ๐ โ ๐ฅ๐๐ฆ๐). (2.9)
In the following we give the definition of a ๐ถโ0,๐-valued function.
Definition 2.1. Let ฮฉ โ โ๐ be an open connected set. Functions ๐ defined in ฮฉwith values in ๐ถโ0,๐ can be expressed as:
๐ : ฮฉ โโ ๐ถโ0,๐.
They are of the form:
๐(๐) =โ
๐ด๐๐ด๐๐ด(๐), (2.10)
where ๐๐ด are real-valued functions in ฮฉ.
More specifically, we let ๐ฟ๐(ฮฉ;๐ถโ0,๐), 1 โค ๐ < โ and ๐ฟโ(ฮฉ;๐ถโ0,๐) denotethe usual Lebesgue space of integrable or essentially bound ๐ถโ0,๐-valued functionon ฮฉ. Notice that ๐ฟ๐(ฮฉ;๐ถโ0,๐) is a ๐ถโ0,๐-bimodule. Moreover, it can be provedto be a Banach module. The norms on the space are denoted โฅโ โฅ๐ฟ๐(โ๐;๐ถโ0,๐) and
โฅโ โฅ๐ฟโ๐(โ๐;๐ถโ0,๐), respectively. The set of ๐ถ๐-functions in ฮฉ with values in ๐ถโ0,๐ is
denoted by
๐ถ๐(ฮฉ;๐ถโ0,๐) = {๐ โฃ๐ : ฮฉ โโ ๐ถโ0,๐, ๐(๐) =โ๐ด
๐๐ด๐๐ด(๐)} (2.11)
Notice that if ๐๐ด โ ๐ถ๐(ฮฉ) then we say ๐ โ ๐ถ๐(ฮฉ;๐ถโ0,๐).Let us consider ๐ฟ2(โ๐;๐ถโ0,๐) as a left module. For two multivector functions
๐, ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐), an inner product is defined by
(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) =
โซโ๐
๐(๐)๐(๐) ๐๐๐ =โ
๐ด,๐ต๐๐ด๐๐ต
โซโ๐
๐๐ด(๐)๐๐ต(๐) ๐๐๐ (2.12)
In particular, if ๐ = ๐, then the scalar part of the above inner product gives the๐ฟ2-norm
โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ0,๐) =
โซโ๐
โ๐ด๐2๐ด(๐) ๐
๐๐ (2.13)
288 M. Bahri
3. CliffordโFourier Transform (CFT)
In the following, we introduce the CliffordโFourier transform (CFT) (see [8, 5, 19]).We can regard the CFT as an alternative representation of the classical tensorialFourier transform, i.e., we apply a one-dimensional Fourier transform ๐ times(each time with a different imaginary unit).
3.1. Definition of CFT
Definition 3.1. The CFT of a multivector function ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐) is the func-tion โฑ{๐}: โ๐ โ ๐ถโ0,๐ given by
โฑ{๐}(๐) =โซโ๐
๐(๐)
๐โ๐=1
๐โ๐๐๐๐๐ฅ๐๐๐๐, (3.1)
with ๐,๐ โ โ๐.
Note that
๐๐๐ =๐๐1 โง ๐๐2 โง โ โ โ โง ๐๐๐
๐๐(3.2)
and is scalar valued (๐๐๐ = ๐๐ฅ๐๐๐, ๐ = 1, 2, 3, . . . , ๐, no summation). Notice alsothat the CliffordโFourier kernel
โ๐๐=1 ๐
โ๐๐๐๐๐ฅ๐ in general does not commute withelements of ๐ถโ0,๐. Furthermore, the product has to be performed in a fixed order.
The existence of the inverse CFT is given by the following theorem. For moredetail and for proofs see [5, 7].
Theorem 3.2. Suppose that ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐) and โฑ{๐} โ ๐ฟ1(โ๐;๐ถโ0,๐). Thenthe CFT is invertible and its inverse is calculated by
๐(๐) =1
(2๐)๐
โซโ๐
โฑ{๐}(๐)๐โ1โ๐=0
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐ ๐๐๐. (3.3)
The CFT is a generalization of the quaternionic Fourier transform (QFT),so most of the properties of the QFT such as the shift property, convolution, andPlancherelโs theorem, have their corresponding CFT generalizations. Observe thatthis Fourier transform can be extended to the whole of ๐ฟ2(โ๐;๐ถโ0,๐) in the usualway by considering it in the weak or distributional sense.
The following subsection investigates some important properties of the CFT,which will be necessary to establish the Clifford windowed Fourier transform in๐ถโ0,๐ (CWFT).
3.2. Main Properties of CFT
We first establish a scalar Plancherel theorem. It states that for every ๐ โ๐ฟ2(โ๐;๐ถโ0,๐)
โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ0,๐) =1
(2๐)๐โฅโฑ{๐}โฅ2๐ฟ2(โ๐;๐ถโ0,๐) . (3.4)
This shows that the total signal energy computed in the spatial domain is equal tothe total signal energy computed in the Clifford domain. The Plancherel theorem
14. Clifford Windowed Fourier Transform 289
allows the energy of a Clifford-valued signal to be considered in either the spatialdomain or the Clifford domain and the change of domains for convenience ofcomputation.
Let us now formulate the CliffordโParseval theorem, which is needed to provethe orthogonality relation of the CWFT.
Theorem 3.3 (CFT Parseval). The inner product (2.12) of two Clifford functions๐, ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐) and their CFTs are related by
(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) =1
(2๐)๐(โฑ{๐},โฑ{๐})๐ฟ2(โ๐;๐ถโ0,๐). (3.5)
Proof. We have
(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐)
=
โซโ๐
๐(๐)๐(๐) ๐๐๐
(3.3)=
1
(2๐)๐
โซโ๐
[โซโ๐
โฑ{๐}(๐)๐โ1โ๐=0
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐ ๐๐๐
]๐(๐)๐๐๐
(2.6)=
1
(2๐)๐
โซโ๐
โฑ{๐}(๐)[โซ
โ๐
๐(๐)
๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐
]๐๐๐
=1
(2๐)๐
โซโ๐
โฑ{๐}(๐)โฑ{๐}(๐) ๐๐๐
=1
(2๐)๐(โฑ{๐},โฑ{๐})๐ฟ2(โ๐;๐ถโ0,๐).
This proves the theorem. โก
Notice that equation (3.5) is multivector valued. In particular, with ๐ = ๐,we get the multivector version of the Plancherel theorem, i.e.,
(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) =1
(2๐)๐(โฑ{๐},โฑ{๐})๐ฟ2(โ๐;๐ถโ0,๐). (3.6)
Due to the non-commutativity of the Clifford exponential product factors we havea left linearity property for general linear combinations with Clifford constantsand a scaling property.
Theorem 3.4 (Left linearity property). The CFT of two functions in the Cliffordmodule ๐, ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐) is a left linear operator 2, i.e.,
โฑ{๐ผ๐ + ๐ฝ๐}(๐) = ๐ผโฑ{๐}(๐) + ๐ฝโฑ{๐}(๐), (3.7)
where ๐ผ and ๐ฝ โ ๐ถโ0,๐ are Clifford constants.
2The CFT is also right linear for real constants ๐, ๐ โ โ.
290 M. Bahri
Theorem 3.5 (Scaling property). Suppose ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐). Let ๐ be a positivereal constant. Then the CFT of the function ๐๐(๐) = ๐(๐๐) is
โฑ{๐๐}(๐) = 1
๐๐โฑ{๐}
(๐๐
). (3.8)
Remark 3.6. The usual form of the shift and modulation properties of the complexFT does not hold for the CFT because of the non-commutativity of the CliffordโFourier kernel
๐โ๐=1
๐๐๐๐๐๐ฅ๐
๐โ1โ๐โ0
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐ โ=๐โ1โ๐=0
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐
๐โ๐=1
๐๐๐๐๐๐ฅ๐ . (3.9)
The following properties are extensions of the QFT, which are very usefulin solving partial differential equations in Clifford algebra. First, let us give anexplicit proof of the derivative properties stated in Table 1.
Table 1. Properties of the CFT of Clifford functions๐, ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐).The constants are ๐ผ, ๐ฝ โ ๐ถโ0,๐, ๐ โ โ โ {0}, and ๐ โ โ.
Property Clifford Function CliffordโFourier Transform
Left linearity ๐ผ๐(๐)+๐ฝ๐(๐) ๐ผโฑ{๐}(๐)+ ๐ฝโฑ{๐} (๐)
Scaling ๐(๐๐) 1
๐๐โฑ{๐}(๐๐ )
Partial derivative โ๐
โ๐ฅ๐1๐(๐) ๐โ๐1 ๐๐
1โฑ{๐}(๐)โ๐
โ๐ฅ๐1๐(๐) (๐1๐1)
๐โฑ{๐}(๐),๐ = ๐0 + ๐1๐1 + ๐๐๐123โ โ โ ๐if ๐ = 3 (mod 4)
โ๐๐โ๐ฅ๐
๐(๐๐๐๐)
๐โฑ{๐}(๐),๐ = 2, . . . , ๐โ 1,๐ = 2๐ , ๐ โ โ
โ๐๐โ๐ฅ๐
๐โฑ{๐}(๐)(๐๐๐๐)
๐
Power ๐(๐)(๐ฅ1๐1)๐ (โ1)๐ โ๐
โ๐๐1โฑ{๐}(๐)
๐(๐)(๐ฅ๐๐๐)๐ โ๐
โ๐๐๐โฑ{๐}(๐),
๐ = 2, . . . , ๐โ 1,๐ = 2๐ , ๐ โ โ
๐(๐)๐ฅ๐๐โ๐
โ๐๐๐โฑ{๐}(๐)๐๐๐
Plancherel (๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) =1
(2๐)๐ (โฑ{๐},โฑ{๐})๐ฟ2(โ๐;๐ถโ0,๐)
Scalar Parseval 1(2๐)๐ โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ0,๐) =
1(2๐)๐ โฅโฑ{๐}โฅ2๐ฟ2(โ๐;๐ถโ0,๐)
14. Clifford Windowed Fourier Transform 291
Theorem 3.7. Suppose ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐). Then the CFT of the ๐th partial deriv-ative of ๐ฅ1๐ โ ๐ฟ1(โ๐;๐ถโ0,๐) with respect to the variable ๐ฅ1 is given by
โฑ{โ๐๐
โ๐ฅ๐1๐โ๐1
}(๐) = ๐๐
1โฑ{๐}(๐), โ๐ โ โ (3.10)
and if ๐ฅ๐๐ โ ๐ฟ1(โ๐;๐ถโ0,๐), then for ๐ = 2, 3, . . . ,โโ 1 we have
โฑ{โ๐๐
โ๐ฅ๐๐
}(๐) =
โซโ๐
๐(๐) ๐โ๐1๐1๐ฅ1๐โ๐2๐2๐ฅ2 โ โ โ (๐๐ ๐๐)๐
ร๐โ
๐=๐
๐โ๐๐๐๐๐ฅ๐ ๐๐๐, ๐ = 2๐ + 1, ๐ โ โ, (3.11)
and
โฑ{โ๐๐
โ๐ฅ๐๐
}(๐) = (๐๐๐๐)
๐โฑ{๐}(๐), ๐ = 2๐ , ๐ โ โ. (3.12)
Proof. We only prove (3.10) of Theorem 3.7, the others being similar. In this proof,we first prove the theorem for ๐ = 1. Applying integration by parts and using thefact that ๐ tends to zero for ๐ฅ1 โโ we immediately obtain
โฑ{
โ
โ๐ฅ1๐ ๐โ1
1
}(๐)
=
โซโ๐
(โ
โ๐ฅ1๐(๐) ๐โ1
1
) ๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐
=
โซโ๐โ1
[โซโ
(โ
โ๐ฅ1๐(๐) ๐โ1
1
)๐โ๐1๐1๐ฅ1 ๐๐ฅ1
] ๐โ๐=2
๐โ๐๐๐๐๐ฅ๐ ๐๐โ1๐
=
โซโ๐โ1
[๐(๐) ๐โ1
1 ๐โ๐1๐1๐ฅ1 โฃ๐ฅ1=โ๐ฅ1=โโ
โโซโ
๐(๐) ๐โ11
โ
โ๐ฅ1๐โ๐1๐1๐ฅ1๐๐ฅ1
] ๐โ๐=2
๐โ๐๐๐๐๐ฅ๐ ๐๐โ1๐
=
โซโ๐
๐(๐)๐1
๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐
= ๐1โฑ{๐}(๐). (3.13)
In a similar way we get
โฑ{
โ
โ๐ฅ1๐
}(๐) = ๐1โฑ{๐๐1}(๐). (3.14)
292 M. Bahri
For ๐ = 2 we obtain
โฑ{
โ2
โ๐ฅ21
๐ ๐โ21
}(๐)
=
โซโ๐
(โ
โ๐ฅ1(
โ
โ๐ฅ1๐(๐)) ๐โ2
1
) ๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐
=
โซโ๐โ1
[โซโ
(โ
โ๐ฅ1(
โ
โ๐ฅ1๐(๐)) ๐โ2
1
)๐โ๐1๐1๐ฅ1 ๐๐ฅ1
] ๐โ๐=2
๐โ๐๐๐๐๐ฅ๐ ๐๐โ1๐
=
โซโ๐โ1
[(
โ
โ๐ฅ1๐(๐)) ๐โ2
1 ๐โ๐1๐1๐ฅ1 โฃ๐ฅ1=โ๐ฅ1=โโ
โ๐1
โซโ
โ
โ๐ฅ1๐(๐) ๐โ1
1 ๐โ๐1๐1๐ฅ1๐๐ฅ1
] ๐โ๐=2
๐โ๐๐๐๐๐ฅ๐ ๐๐โ1๐
= ๐21
โซโ๐
๐(๐)๐โ
๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐
= ๐21โฑ{๐}(๐). (3.15)
By repeating this process ๐โ2 additional times we finish the proof of Theorem 3.7.โก
Remark 3.8. Observe that when we assume that ๐ = ๐0 + ๐1๐1 + ๐๐๐123โ โ โ ๐, ๐ = 3(mod 4), then equation (3.14) takes the form
โฑ{โ๐๐
โ๐ฅ๐1
}(๐) = (๐1๐1)
๐โฑ{๐}(๐), ๐ โ โ. (3.16)
Theorem 3.9. Let ๐ฅ๐๐ โ ๐ฟ1(โ๐;๐ถโ0,๐). Then the CFT of the ๐th partial derivativeof a Clifford-valued function ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐) with respect to the variable ๐ฅ๐ isgiven by
โฑ{โ๐๐
โ๐ฅ๐๐
}(๐) = โฑ{๐}(๐)(๐๐๐๐)
๐, ๐ โ โ. (3.17)
Proof. For ๐ = 1 direct calculation gives
โ๐(๐)
โ๐ฅ๐=
โ
โ๐ฅ๐
1
(2๐)๐
โซโ๐
โฑ{๐}(๐)๐โ1โ๐=0
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐ ๐๐๐
=1
(2๐)๐
โซโ๐
โฑ{๐}(๐)(
โ
โ๐ฅ๐๐๐๐๐๐๐ฅ๐
) ๐โ1โ๐=1
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐ ๐๐๐
=1
(2๐)๐
โซโ๐
[โฑ{๐}(๐) ๐๐๐๐]๐โ1โ๐=0
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐ ๐๐๐
= โฑโ1 [โฑ{๐}(๐) ๐๐๐๐] . (3.18)
14. Clifford Windowed Fourier Transform 293
We therefore get
โฑ{
โ๐
โ๐ฅ๐
}(๐) = โฑ{๐}(๐)๐๐๐๐. (3.19)
By successive differentiation with respect to the variable ๐ฅ๐ and by induction weeasily obtain
โฑ{โ๐๐
โ๐ฅ๐๐
}(๐) = โฑ{๐}(๐)(๐๐๐๐)
๐, โ๐ โ โ. (3.20)
This ends the proof of (3.17). โก
Next we derive the power properties of the CFT stated in Table 1.
Theorem 3.10. If we assume that ๐ฅ1๐ โ ๐ฟ1(โ๐;๐ถโ0,๐). Then the CFT of the ๐thpartial derivative of ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐) with respect to the variable ๐ฅ1 is given by
โฑ{๐(๐)(๐ฅ1๐1)๐}(๐) = (โ1)๐ โ๐
โ๐๐1
โฑ{๐}(๐), โ๐ โ โ. (3.21)
If ๐ฅ๐๐ ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐), then for ๐ = 2, 3, . . . , ๐โ 1 we have
โ๐
โ๐๐๐
โฑ{๐}(๐) =โซโ๐
๐(๐) ๐โ๐1๐1๐ฅ1๐โ๐2๐2๐ฅ2 โ โ โ (โ๐๐๐ฅ๐)๐
ร๐โ
๐=๐
๐โ๐๐๐๐๐ฅ๐ ๐๐๐,๐ = 2๐ + 1, ๐ โ โ, (3.22)
and for ๐ = 2๐ , ๐ โ โ
โฑ{(๐๐๐ฅ๐)๐๐}(๐) = โ๐
โ๐๐๐
โฑ{๐}(๐). (3.23)
If ๐ฅ๐๐ โ ๐ฟ1(โ๐;๐ถโ0,๐), then
โฑ{๐ฅ๐๐ ๐}(๐) = โ๐
โ๐๐๐
โฑ{๐}(๐) ๐๐๐ , ๐ = 1, 2, 3, . . . , ๐. (3.24)
Proof. We only prove (3.21) of Theorem 3.10. It is not difficult to check that
๐(๐)(๐ฅ1๐1)๐
๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ = (โ1)๐ โ๐
โ๐๐1
๐(๐)๐โ
๐=1
๐โ๐๐๐๐๐ฅ๐ . (3.25)
We immediately obtainโซโ๐
๐(๐)(๐ฅ1๐1)๐
๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐ = (โ1)๐ โ๐
โ๐๐1
โซโ๐
๐(๐)
๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐,
(3.26)which gives the desired result. โก
294 M. Bahri
4. Clifford Windowed Fourier Transform (CWFT)
In this section, we introduce the Clifford windowed Fourier transform as a gener-alization of the two-dimensional quaternionic Fourier transform to higher dimen-sions. For this let us define the CliffordโGabor filter, which is a special case of theClifford atom operator [9].
4.1. Two-dimensional CliffordโGabor Filters
The two-dimensional CliffordโGabor filter3 is the extension of the complex Gaborfilter to the two-dimensional Clifford algebra. It takes the form
๐บ๐(๐, ๐1, ๐2) = ๐๐2๐ข0๐ฅ1๐๐1๐ฃ0๐ฅ2๐(๐,๐)
= ๐๐2๐ข0๐ฅ1๐๐1๐ฃ0๐ฅ2๐โ[(๐ฅ1/๐1)2+(๐ฅ2/๐2)
2]/2. (4.1)
Equation (4.1) is often called thequaternionic Gabor filter. Bulow and Sommer[6, 7] have applied it to get the local quaternionic phase of a two-dimensional realsignal. From this, we get the following facts:
โ It is generated using the kernel of the ๐ถ๐(0, 2) CFT.โ If the Gaussian function ๐(๐,๐) is replaced by the Clifford window function
๐(๐โ ๐), then it becomes the Clifford atom operator, i.e.,
๐๐,๐(๐) = ๐๐2๐ฃ0๐ฅ2๐๐1๐ข0๐ฅ1๐(๐โ ๐), ๐, ๐ โ โ2. (4.2)
โ Since the modulation property does not hold for the CFT, (4.2) can not beexpressed in terms of the CFT.
4.2. Definition of CWFT
Definition 4.1. The CWFT of a multivector function ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐) withrespect to the non-zero Clifford window function ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐) such that
โฃ๐โฃ1/2 ๐(๐) โ ๐ฟ2(โ๐;๐ถโ0,๐) is given by
๐บ๐๐(๐, ๐) = (๐, ๐๐,๐)๐ฟ2(โ๐;๐ถโ0,๐)
=
โซโ๐
๐(๐)๐๐,๐(๐) ๐๐๐
=
โซโ๐
๐(๐) ๐(๐โ ๐)
๐โ๐=1
๐โ๐๐๐๐๐ฅ๐ ๐๐๐. (4.3)
We then call
๐๐,๐(๐) =
๐โ1โ๐=0
๐๐๐โ๐๐๐โ๐๐ฅ๐โ๐๐(๐โ ๐), (4.4)
the atom operator as the kernel of the CWFT in (4.3). Notice that for ๐ = 2 theCWFT above is identical to the two-dimensional quaternionic windowed Fourier
3Here we start with the CliffordโGabor filter to obtain the Clifford atom operator, which isneeded to construct the Clifford windowed Fourier transform.
14. Clifford Windowed Fourier Transform 295
transform (see [2]) and for ๐ = 1 is the classical windowed Fourier transform (see[10, 11]).
Example. Consider the Clifford Gaussian window ๐ โ ๐ฟ2(โ2;๐ถโ0,2) given by:
๐(๐) = (2 + ๐1 + ๐2 โ ๐12)๐โ(๐ฅ2
1+๐ฅ22). (4.5)
Thus we obtain the Clifford window daughter functions (4.4) of the form
๐๐,๐(๐) = {๐๐2๐2๐ฅ2๐๐1๐1๐ฅ1(2 + ๐1 + ๐2 โ ๐12)} ๐โ((๐ฅ1โ๐1)2+(๐ฅ2โ๐2)2)
= {(2๐๐2๐2๐ฅ2๐๐1๐1๐ฅ1 + ๐1๐โ๐2๐2๐ฅ2๐๐1๐1๐ฅ1 + ๐2๐
๐2๐2๐ฅ2๐โ๐1๐1๐ฅ1
โ ๐12๐โ๐2๐2๐ฅ2๐โ๐1๐1๐ฅ1)}๐โ((๐ฅ1โ๐1)2+(๐ฅ2โ๐2)2). (4.6)
We first notice that, for fixed ๐,
๐บ๐๐(๐, ๐) = โฑ{๐ ๐(โ โ ๐)}(๐) = โฑ{๐ ๐๐๐}(๐), (4.7)
where ๐๐ is the translation operator defined by ๐๐๐ = ๐(๐โ๐). It thus means thatthe CWFT can be regarded as the CFT of the product of a multivector-valuedfunction ๐ and a shifted and Clifford reversion of the Clifford atom operator (4.4).
4.3. Properties of the CWFT
The following proposition describes the elementary properties of the CWFT. Itsproof can be easily obtained.
Proposition 4.2. Let ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐) be a Clifford window function.
Left linearity:
[๐บ๐(๐๐ + ๐๐)](๐, ๐) = ๐๐บ๐๐(๐, ๐) + ๐๐บ๐๐(๐, ๐), (4.8)
for arbitrary Clifford constants ๐, ๐ โ ๐ถโ0,๐.
Parity:๐บ๐๐(๐๐)(๐, ๐) = ๐บ๐๐(๐,โ๐), (4.9)
where ๐ is the parity operator defined by ๐๐(๐ฅ) = ๐(โ๐ฅ).
Theorem 4.3 (Orthogonality relation). Let ๐, ๐ be Clifford window functions and๐, ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐) be arbitrary. Then we haveโซ
โ๐
โซโ๐
๐บ๐๐(๐, ๐)๐บ๐๐(๐, ๐) ๐๐๐ ๐๐๐
= (2๐)๐(๐(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐), ๐)๐ฟ2(โ๐;๐ถโ0,๐). (4.10)
Proof. Applying (4.7) we haveโซโ๐
๐บ๐๐(๐, ๐)๐บ๐๐(๐, ๐) ๐๐๐ =
โซโ๐
โฑ{๐๐๐๐}โฑ{๐๐๐๐} ๐๐๐
=(โฑ{๐๐๐๐},โฑ{๐๐๐๐}
)๐ฟ2(โ๐;๐ถโ0,๐)
. (4.11)
We assume that Clifford windows ๐, ๐ โ ๐ฟ1(โ๐;๐ถโ0,๐)โฉ
๐ฟโ(โ๐;๐ถโ0,๐) so that
๐๐๐๐, ๐๐๐๐ โ ๐ฟ2(โ๐;๐ถโ0,๐). We know [6, 13] that Parsevalโs theorem is valid for
296 M. Bahri
the CFT. So, applying it to the right-hand side of (4.11) we easily get (compareto Grochenig [10])โซ
โ๐
๐บ๐๐(๐, ๐)๐บ๐๐(๐, ๐) ๐๐๐ = (โฑ{๐๐๐๐},โฑ{๐๐๐๐})๐ฟ2(โ๐;๐ถโ0,๐)
= (2๐)๐(๐๐๐๐, ๐๐๐๐
)๐ฟ2(โ๐;๐ถโ0,๐)
= (2๐)๐โซโ๐
๐(๐)๐(๐ โ ๐)๐(๐โ ๐)๐(๐) ๐๐๐. (4.12)
Observe that ๐๐ and ๐๐ are in ๐ฟ1(โ๐;๐ถโ0,๐). Then integrating (4.12) with respectto ๐๐๐ we immediately getโซ
โ๐
โซโ๐
๐บ๐๐(๐, ๐)๐บ๐๐(๐, ๐) ๐๐๐ ๐๐๐
= (2๐)๐โซโ๐
โซโ๐
๐(๐)๐(๐โ ๐)๐(๐โ ๐)๐(๐) ๐๐๐ ๐๐๐
= (2๐)๐โซโ๐
โซโ๐
๐(๐)๐(๐โ ๐)๐(๐โ ๐) ๐๐๐ ๐(๐) ๐๐๐
= (2๐)๐โซโ๐
๐(๐)
โซโ๐
๐(๐โ ๐)๐(๐โ ๐) ๐๐๐ ๐(๐) ๐๐๐, (4.13)
where from the second to the third line of (4.13) we applied Fubiniโs theorem tointerchange the order of integration. Using a standard density argument we cannow extend the result to the ๐ฟ2(โ๐;๐ถโ0,๐)-case. This proves the theorem. โก
From the above theorem, we obtain the following consequences.
(i) If ๐ = ๐ and ๐ถ๐ = (๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) is a multivector constant, thenโซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐บ๐๐(๐, ๐) ๐๐๐ ๐๐๐ = (2๐)๐โจ๐ถ๐โฉ(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐)
+ (2๐)๐(๐โจ๐ถ๐โฉ1, ๐)๐ฟ2(โ๐;๐ถโ0,๐) + โ โ โ + (2๐)๐(๐โจ๐ถ๐โฉ๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐). (4.14)
(ii) If ๐ = ๐ is a paravector, then (4.14) reduces toโซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐บ๐๐(๐, ๐) ๐๐๐ ๐๐๐ = (2๐)๐โจ๐ถ๐โฉ โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ0,๐)
+ (2๐)๐(๐โจ๐ถ๐โฉ1, ๐)๐ฟ2(โ๐;๐ถโ0,๐) + โ โ โ + (2๐)๐(๐โจ๐ถ๐โฉ๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐). (4.15)
Theorem 4.4 (Reconstruction formula). Let ๐, ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐) be two Clifford
window functions with (๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) โ= 0. Then every ๐-D Clifford signal ๐ โ๐ฟ2(โ๐;๐ถโ0,๐) can be fully reconstructed by
๐(๐) =1
(2๐)๐
โซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐๐,๐(๐) (๐, ๐)โ1๐ฟ2(โ๐;๐ถโ0,๐
)๐๐๐ ๐๐๐. (4.16)
14. Clifford Windowed Fourier Transform 297
Proof. By direct calculation, we obtain for every ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐)โซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐บ๐๐(๐, ๐) ๐๐๐ ๐๐๐
=
โซโ๐
โซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐๐,๐(๐)๐(๐) ๐๐๐ ๐๐๐ ๐๐๐
=
(โซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐๐,๐ ๐๐๐ ๐๐๐, ๐
)๐ฟ2(โ๐;๐ถโ0,๐)
. (4.17)
Applying equation (4.10) of Theorem 4.3 to the left-hand side of (4.17) gives forevery ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐)
(2๐)๐(๐(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐), ๐)๐ฟ2(โ๐;๐ถโ0,๐)
=
(โซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐๐,๐ ๐๐๐ ๐๐๐, ๐
)๐ฟ2(โ๐;๐ถโ0,๐)
. (4.18)
Because the inner product identity (4.18) holds for every ๐ โ ๐ฟ2(โ๐;๐ถโ0,๐) weconclude that
(2๐)๐๐(๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) =
โซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐๐,๐ ๐๐๐ ๐๐๐. (4.19)
If it is assumed that the inner product (๐, ๐)๐ฟ2(โ๐;๐ถโ0,๐) is invertible. Then multi-
plying both sides of (4.19) from the right side by (๐, ๐)โ1๐ฟ2(โ๐;๐ถโ0,๐) we immediately
obtain
(2๐)๐๐ =
โซโ๐
โซโ๐
๐บ๐๐(๐, ๐)๐๐,๐ (๐, ๐)โ1๐ฟ2(โ๐;๐ถโ0,๐) ๐
๐๐ ๐๐๐, (4.20)
which was to be proved. โก
Acknowledgment
The author would like to thank the reviewer whose deep and extensive commentsgreatly contributed to improve this chapter. He thanks Ass. Prof. Eckhard Hitzerfor his helpful guidance. He also wants to thank ICCA9 organizer Professor KlausGurlebeck.
References
[1] M. Bahri. Generalized Fourier transform in real clifford algebra ๐๐(0, ๐). Far EastJournal of Mathematical Sciences, 48(1):11โ24, Jan. 2011.
[2] M. Bahri, E. Hitzer, R. Ashino, and R. Vaillancourt. Windowed Fourier transformof two-dimensional quaternionic signals. Applied Mathematics and Computation,216(8):2366โ2379, June 2010.
[3] E. Bayro-Corrochano. The theory and use of the quaternion wavelet transform. Jour-nal of Mathematical Imaging and Vision, 24(1):19โ36, 2006.
[4] E. Bayro-Corrochano and G. Scheuermann, editors. Applied Geometric Algebras inComputer Science and Engineering. Springer, London, 2010.
298 M. Bahri
[5] F. Brackx, R. Delanghe, and F. Sommen. Clifford Analysis, volume 76. Pitman,Boston, 1982.
[6] T. Bulow. Hypercomplex Spectral Signal Representations for the Processing and Anal-ysis of Images. PhD thesis, University of Kiel, Germany, Institut fur Informatik undPraktische Mathematik, Aug. 1999.
[7] T. Bulow, M. Felsberg, and G. Sommer. Non-commutative hypercomplex Fouriertransforms of multidimensional signals. In G. Sommer, editor, Geometric computingwith Clifford Algebras: Theoretical Foundations and Applications in Computer Visionand Robotics, pages 187โ207, Berlin, 2001. Springer.
[8] T. Bulow and G. Sommer. Hypercomplex signals โ a novel extension of the ana-lytic signal to the multidimensional case. IEEE Transactions on Signal Processing,49(11):2844โ2852, Nov. 2001.
[9] Y. Fu, U. Kahler, and P. Cerejeiras. The BalianโLow theorem for the windowedquaternionic Fourier transform. Advances in Applied Clifford Algebras, page 16, 2012.published online 3 February 2012.
[10] K. Grochenig. Foundations of Time-Frequency Analysis. Applied and Numerical Har-monic Analysis. Birkhauser, Boston, 2001.
[11] K. Grochenig and G. Zimmermann. Hardyโs theorem and the short-time Fouriertransform of Schwartz functions. Journal of the London Mathematical Society, 2:205โ214, 2001.
[12] S.L. Hahn. Wigner distributions and ambiguity functions of 2-D quaternionic andmonogenic signals. IEEE Transactions on Signal Processing, 53:3111โ3128, 2005.
[13] E. Hitzer. Quaternion Fourier transform on quaternion fields and generalizations.Advances in Applied Clifford Algebras, 17(3):497โ517, May 2007.
[14] E.M.S. Hitzer and B. Mawardi. Clifford Fourier transform on multivector fields anduncertainty principles for dimensions ๐ = 2(mod 4) and ๐ = 3(mod 4). Advances inApplied Clifford Algebras, 18(3-4):715โ736, 2008.
[15] B. Mawardi. Clifford windowed Fourier transform applied to linear time-varyingsystems. Applied Mathematical Sciences, 6:2857โ2864, 2012.
[16] B. Mawardi, S. Adji, and J. Zhao. Real Clifford windowed Fourier transform. ActaMathematica Sinica, 27:505โ518, 2011.
[17] B. Mawardi, E. Hitzer, and S. Adji. Two-dimensional Clifford windowed Fouriertransform. In Bayro-Corrochano and Scheuermann [4], pages 93โ106.
[18] B. Mawardi, E. Hitzer, A. Hayashi, and R. Ashino. An uncertainty principlefor quaternion Fourier transform. Computers and Mathematics with Applications,56(9):2411โ2417, 2008.
[19] F. Sommen. A product and an exponential function in hypercomplex function theory.Applicable Analysis, 12:13โ26, 1981.
[20] Y. Xi, X. Yang, L. Song, L. Traversoni, and W. Lu. QWT: Retrospective and newapplication. In Bayro-Corrochano and Scheuermann [4], pages 249โ273.
Mawardi BahriDepartment of MathematicsUniversitas HasanuddinTamalanrea Makassar 90245, Indonesiae-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 299โ319cโ 2013 Springer Basel
15 The BalianโLow Theorem for theWindowed CliffordโFourier Transform
Yingxiong Fu, Uwe Kahler and Paula Cerejeiras
Abstract. In this chapter, we provide the definition of the CliffordโZak trans-form associated with the discrete version of the kernel of a windowed CliffordโFourier transform. We proceed with deriving several important properties ofsuch a transform. Finally, we establish the BalianโLow theorem for a Cliffordframe under certain natural assumptions on the window function.
Mathematics Subject Classification (2010). 15A66; 30G35.
Keywords. CliffordโZak transform, Clifford frame, BalianโLow theorem, Clif-fordโFourier transform, windowed CliffordโFourier transform.
1. Introduction
The last decade has seen a growing interest in generalizations of the Fourier trans-form, motivated by applications to higher-dimensional signal processing. Firststeps in that direction where already made in Brackx et al. [10], where a Fouriertransform for multivector-valued distributions in ๐ถโ0,๐ with compact support waspresented. Also worth mention is the alternative definition of the Fourier trans-form of Sommen [27], based on a generalization of the exponential function toโ๐รโ๐+1. A quaternionic Fourier transform was given in Bulow et al. [12] in thecontext of quaternionic-valued two-dimensional signals (the so-called hypercom-plex signals). Shortly after, motivated by spectral analysis of colour images Sang-wine et al. proposed in [26] a quaternionic Fourier transform in which the imaginaryunit ๐ was replaced by a unit quaternion. Almost in parallel, Felsberg defined in[17] his CliffordโFourier transform (CFT) for the low-dimensional Clifford alge-bras ๐ถโ2,0 and ๐ถโ3,0, using the pseudoscalar ๐๐ as imaginary unit. Following thisapproach, several authors have extended this transform to a three-dimensional set-ting and successfully detected vector-valued patterns in the frequency domain (cf.[15, 16, 24]). However, a major problem did remain: as the classical Fourier trans-form, the CliffordโFourier transform is ineffective for representing and computing
300 Y. Fu, U. Kahler and P. Cerejeiras
local information of signals. In fact, the harmonic analysis version of Heisenbergโsuncertainty principle states that it is impossible to localize simultaneously a func-tion and its Fourier transform.
One way to overcome this difficulty is by means of CliffordโGabor filters.They were initially proposed in [11], and later on also in [8], which extended theapplications of the complex Gabor filters. In general, they correspond to modula-tions of Gaussians. A good account on this subject can be found in [9].
A more general approach to this problem is by means of the windowed Fouriertransform (WFT), also called continuous Gabor transform or short-time Fouriertransform. Given ๐ โ ๐ฟ2(โ) and a fixed non-zero window function ๐ โ ๐ฟ2(โ), onedefines the WFT ๐๐{๐} โ ๐ฟ2(โ2) as
๐๐{๐}(๐ก, ๐) =โซโ
๐(๐ฅ)๐(๐ฅโ ๐ก)๐โ2๐๐๐ฅ๐๐๐ฅ. (1.1)
Mawardi et al. extended the theory of WFT to the Clifford case [1โ4, 23]. In [3]the definition of a windowed CliffordโFourier transform (WCFT) was establishedand several important properties were obtained such as shift, modulation, recon-struction formulae, etc. Furthermore, a Heisenberg type uncertainty principle forthe WCFT was derived.
Practical applications require a discrete version of this continuous transform.One can establish a discrete form of the WFT ๐๐ linked to a given countable setฮ โ โ2 = โ ร โ as the transformation that assigns to every function ๐ โ ๐ฟ2(โ)the number sequence
๐ ๏ฟฝโ { โจ๐, ๐๐,๐โฉ : (๐,๐) โ ฮ}, (1.2)
where โจโ , โ โฉ denotes the ordinary inner product in ๐ฟ2(โ) and the functions ๐๐,๐
are shifts and modulations of the window ๐ given by
๐๐,๐(๐ฅ) = ๐2๐๐๐๐ฅ๐(๐ฅโ ๐) (1.3)
for all (๐,๐) โ ฮ โ โ2. Such a collection {๐๐,๐ : (๐,๐) โ ฮ} is called Gaborsystem, or Weyl-Heisenberg system, generated by ๐ and ฮ. To recover the originalfunction ๐ from the number sequence {โจ๐, ๐๐,๐โฉ : (๐,๐) โ ฮ} it is necessary thatthe system forms an orthonormal basis or at least a frame.
Gabor systems are related to the classical uncertainty principle by the BalianโLow theorem (a stronger version of the said principle), as it expresses the fact thattime-frequency concentration and non-redundancy are incompatible properties ofa Gabor system if such a system is a frame for ๐ฟ2(โ) [7,13,14,18,21]. Specifically,if the window ๐ is such that the Gabor system {๐๐,๐ : ๐,๐ โ โค} constitutes anexact frame for ๐ฟ2(โ), i.e., if there exist constants 0 < ๐ต โค ๐ถ <โ such that
๐ตโฅ๐โฅ2 โคโ
๐,๐โโคโฃโจ๐, ๐๐,๐โฉโฃ2 โค ๐ถ โฅ๐โฅ2 , โ๐ โ ๐ฟ2(โ), (1.4)
15. The BalianโLow Theorem for the WCFT 301
and the system ceases to be a frame when any of its elements is removed then, itholds (โซ +โ
โโ๐ก2โฃ๐(๐ก)โฃ2๐๐ก
)(โซ +โ
โโ๐2โฃ๐(๐)2๐๐
)=โ. (1.5)
In other words, the window function ๐ maximizes the uncertainty principle insome sense. This result has been extended to higher dimensions and to a moregeneral set of time-frequency shifts in the standard coordinate system [6, 19]. Itwas also proved to be valid in a multi-window setting [25, 28] and in the case ofsuperframes [5].
Regarding the important question of discretizing the WCFT it is naturalto ask if the BalianโLow theorem holds for Gabor systems generated by certainClifford-valued window functions ๐ and countable sets ฮ โ โ2๐. Our goal in thischapter is to obtain it for Gabor systems which form a Clifford frame arising fromthe discrete version of the kernel of the WCFT. To this end it is necessary tointroduce a CliffordโZak transform and study some of its properties.
This chapter is organized as follows: Section 2 is devoted to the review ofthe necessary results on Clifford algebra and the definitions of both the CFT andthe WCFT. In Section 3 we provide the definition of a Clifford frame associatedwith the discrete version of the kernel of the WCFT, establish the definition of aCliffordโZak transform and derive some properties of it, which will play a key rolein the proof of the BalianโLow theorem. In Section 4 we demonstrate the BalianโLow theorem for Gabor systems which form a Clifford frame. Some conclusionsare drawn in Section 5.
2. Preliminaries
Let ๐ถโ๐,0 be the 2๐-dimensional universal real Clifford algebra over โ๐ constructed
from the basis {๐1, ๐2, . . . , ๐๐} under the usual relations
๐๐๐๐ + ๐๐๐๐ = 2๐ฟ๐๐, 1 โค ๐, ๐ โค ๐, (2.1)
where ๐ฟ๐๐ is the Kronecker delta function. An element ๐ โ ๐ถโ๐,0 can be representedas ๐ =
โ๐ด ๐๐ด๐๐ด, ๐๐ด โ โ, where ๐๐ด = ๐๐1๐2โ โ โ ๐๐ = ๐๐1๐๐2 โ โ โ ๐๐๐ , ๐ด = {๐1, ๐2, . . . ๐๐}
with 1 โค ๐1 โค ๐2 โค โ โ โ โค ๐๐ โค ๐, and ๐0 = ๐โ = 1 is the identity element of๐ถโ๐,0. The elements of the algebra ๐ถโ๐,0 for which โฃ๐ดโฃ = ๐ are called k-vectors.We denote the space of all k-vectors by
๐ถโ๐๐,0 := spanโ{๐๐ด : โฃ๐ดโฃ = ๐}. (2.2)
It is clear that the spaces โ and โ๐ can be identified with ๐ถโ0๐,0 and ๐ถโ1๐,0, re-spectively.
Of interest for this work is the (unit oriented) pseudoscalar element ๐๐ =๐1๐2 โ โ โ ๐๐. Observe that ๐2๐ = โ1 for ๐ = 2, 3(mod 4). For the sake of simplicity,if not otherwise stated, ๐ is always assumed to be ๐ = 2, 3(mod 4) for the remainingof this chapter.
302 Y. Fu, U. Kahler and P. Cerejeiras
We define the anti-automorphism reversionห : ๐ถโ๐,0 โ ๐ถโ๐,0 by its action
on the basis elements ๐๐ด = (โ1) ๐(๐โ1)2 ๐๐ด, for โฃ๐ดโฃ = ๐, and its reversion property
๐๐ = ๐๐ for every ๐, ๐ โ ๐ถโ๐,0. In particular, we remark that ๐๐ = โ๐๐.In what follows, we will require two types of scalar products. First, we in-
troduce the (real-valued) scalar product of ๐, ๐ โ ๐ถโ๐,0 as the scalar part of theirgeometric product
๐ โ ๐ := [๐ ๐ ]0 =โ๐ด
๐๐ด๐๐ด. (2.3)
As usual, when we set ๐ = ๐ we obtain the square of the modulus (or magnitude)of the multivector ๐ โ ๐ถโ๐,0,
โฃ๐ โฃ2 =[๐๐]0=โ๐ด
๐2๐ด. (2.4)
Also, an additional useful property of the scalar part [ ]0 is the cyclic sym-metric product
[๐๐๐]0 = [๐๐๐]0, โ๐, ๐, ๐ โ ๐ถโ๐,0. (2.5)
Second, we require an inner product in the function space under considera-tion. We denote by ๐ฟ๐(โ๐;๐ถโ๐,0) the left module of all Clifford-valued functions๐ : โ๐ โ ๐ถโ๐,0 with finite norm
โฅ๐โฅ๐ =
{ (โซโ๐ โฃ๐(x)โฃ๐๐๐x
) 1๐ , 1 โค ๐ <โ
ess supxโโ๐ โฃ๐(x)โฃ, ๐ =โ , (2.6)
where ๐๐x = ๐๐ฅ1๐๐ฅ2 โ โ โ ๐๐ฅ๐ represents the usual Lebesgue measure in โ๐. In theparticular case of ๐ = 2, we shall denote this norm by โฅ๐โฅ.
Given two functions ๐, ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0), we define a Clifford-valued bilinearform
๐, ๐ โ (๐, ๐) :=
โซโ๐
๐(x)๐(x)๐๐x, (2.7)
from which we construct the scalar inner product
โจ๐, ๐โฉ : = [(๐, ๐)]0 =
โซโ๐
[๐(x)๐(x)
]0๐๐x. (2.8)
We remark that (2.8) satisfies the (Clifford) CauchyโSchwarz inequality
โฃโจ๐, ๐โฉโฃ โค โฅ๐โฅ โฅ๐โฅ , โ๐, ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0). (2.9)
In the following, we recall the CFT, originally introduced by M. Felsberg(see [17]).
Definition 2.1. Let ๐ โ ๐ฟ1(โ๐;๐ถโ๐,0). The CFT of ๐ at the point ๐ โ โ๐ is definedas the ๐ถโ๐,0-valued (Lebesgue) integral
โฑ{๐}(๐) =โซโ๐
๐(x)๐โ2๐๐๐๐โ x๐๐x. (2.10)
The function ๐ โ โฑ{๐}(๐) s called the CFT of ๐.
15. The BalianโLow Theorem for the WCFT 303
Lemma 2.2 (Parsevalโs equality for CFT). If ๐ โ ๐ฟ1(โ๐;๐ถโ๐,0) โฉ ๐ฟ2 (โ๐;๐ถโ๐,0),then
โฅ๐โฅ = โฅโฑ{๐}โฅ. (2.11)
Lemma 2.2 asserts that the CFT is a bounded linear operator on ๐ฟ1(โ๐;๐ถโ๐,0) โฉ ๐ฟ2(โ๐;๐ถโ๐,0). Hence, standard density arguments allow us to extendthe CFT in an unique way to the whole of ๐ฟ2(โ๐;๐ถโ๐,0). In what follows wealways consider the properties of the CFT as an operator from ๐ฟ2(โ๐;๐ถโ๐,0) into๐ฟ2(โ๐;๐ถโ๐,0).
Definition 2.3. Let ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0) be a non-zero window function such that
โฃxโฃ1/2๐(x) is in ๐ฟ2(โ๐;๐ถโ๐,0). Then, the WCFT of ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0) with respectto ๐ is defined by
๐๐๐(๐,b) : =
โซโ๐
๐(x) ห๐(xโ b)๐โ2๐๐๐๐โ x๐๐x
= (๐, ๐๐,b), (2.12)
where ๐๐,b(x) := ๐2๐๐๐๐โ x๐(xโ b) denotes the kernel of the WCFT.
3. Clifford Frame and CliffordโZak Transform
The BalianโLow theorem is regarded as a strong version of the uncertainty princi-ple for the Gabor system associated with the discrete version of the kernel of theclassical WFT. To establish the BalianโLow theorem for a Gabor system in theClifford algebra module ๐ฟ2(โ๐;๐ถโ๐,0) we consider the following discrete versionof the kernel of the WCFT
๐m,n(x) := ๐2๐๐๐mโ x๐(xโ n), x โ โ๐, m,n โ โค๐. (3.1)
A frame for a vector space equipped with an inner product allows each el-ement in the space to be written as a linear combination of the elements in theframe. In general frame elements are neither orthogonal to each other nor linearlyindependent. We now introduce the definition and properties of a Clifford framein ๐ฟ2(โ๐;๐ถโ๐,0) as follows.
Definition 3.1. {๐m,n : m,n โ โค๐} is a Clifford frame for ๐ฟ2(โ๐;๐ถโ๐,0) if thereexist real constants 0 < ๐ต โค ๐ถ <โ such that
๐ต โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ๐,0)โค
โm,nโโค๐
โฃโจ๐, ๐m,nโฉโฃ2 โค ๐ถ โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ๐,0),
โ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0), (3.2)
where ๐m,n is defined by (3.1) and the scalar inner product โจโ , โ โฉ is defined by (2.8).
Any two constants ๐ต,๐ถ satisfying condition (3.1) are called frame bounds.If ๐ต = ๐ถ, then {๐m,n : m,n โ โค๐} is called a tight frame.
To understand frames and reconstruction methods better, we study someimportant associated operators.
304 Y. Fu, U. Kahler and P. Cerejeiras
Definition 3.2. For any subset {๐m,n : m,n โ โค๐} โ ๐ฟ2(โ๐;๐ถโ๐,0), the coefficientoperator ๐น is defined by
๐น๐ = {โจ๐, ๐m,nโฉ : m,n โ โค๐}. (3.3)
The reconstruction operator ๐ for a sequence ๐ = (๐m,n)m,nโโค๐ is given by
๐ ๐ =โ
m,nโโค๐
๐m,n๐m,n โ ๐ฟ2(โ๐;๐ถโ๐,0). (3.4)
Finally, the frame operator ๐ in ๐ฟ2(โ๐;๐ถโ๐,0) is defined by
๐๐ =โ
m,nโโค๐
โจ๐, ๐m,nโฉ ๐m,n. (3.5)
Based on the classic frame theory [18], under the assumption that {๐m,n :m,n โ โค๐} is a frame defined by (3.1) for ๐ฟ2(โ๐;๐ถโ๐,0), we know that the frameoperator ๐ maps ๐ฟ2(โ๐;๐ถโ๐,0) to ๐ฟ2(โ๐;๐ถโ๐,0) and it is a self-adjoint, positiveand invertible operator satisfying
๐ต๐ผ โค ๐ โค ๐ถ๐ผ, ๐ถโ1๐ผ โค ๐โ1 โค ๐ตโ1๐ผ (3.6)
with ๐ผ being the identity operator. Moreover, observe that
โจ๐๐, ๐โฉ =โ
m,nโโค๐
โฃโจ๐, ๐m,nโฉโฃ2 (3.7)
and โm,nโโค๐
โฃโฃโจ๐, ๐โ1๐m,n
โฉโฃโฃ2 =โ
m,nโโค๐
โฃโฃโจ๐โ1๐, ๐m,n
โฉโฃโฃ2=โจ๐(๐โ1๐), ๐โ1๐
โฉ=โจ๐โ1๐, ๐
โฉ. (3.8)
Therefore, we have
๐ถโ1 โฅ๐โฅ2 โค โจ๐โ1๐, ๐โฉ=
โm,nโโค๐
โฃโฃโจ๐, ๐โ1๐m,n
โฉโฃโฃ2 โค ๐ตโ1 โฅ๐โฅ2 . (3.9)
Thus the collection {๐โ1๐m,n : m,n โ โค๐} is a so-called dual frame with framebounds ๐ถโ1 and ๐ตโ1. Using the factorizations ๐ผ = ๐โ1๐ = ๐๐โ1, we obtain theseries expansions
๐ = ๐(๐โ1๐) =โ
m,nโโค๐
โจ๐โ1๐, ๐m,n
โฉ๐m,n
=โ
m,nโโค๐
โจ๐, ๐โ1๐m,n
โฉ๐m,n (3.10)
and
๐ = ๐โ1๐๐ =โ
m,nโโค๐
โจ๐, ๐m,nโฉ ๐โ1๐m,n. (3.11)
15. The BalianโLow Theorem for the WCFT 305
Furthermore, it is well known that the classic Zak transform is a very usefultool to analyze Gabor systems
{๐2๐๐๐๐ก๐(๐กโ ๐) : ๐,๐ โ โค
}in ๐ฟ2(โ). The classic
Zak transform, ๐ : ๐ฟ2(โ)โ ๐ฟ2([0, 1)2), is defined by
๐ = ๐(๐ฅ), ๐ฅ โ โ ๏ฟฝโ ๐๐ = ๐๐(๐ก, ๐)
=โ๐โโค
๐2๐๐ ๐๐๐(๐กโ ๐), (๐ก, ๐) โ โ2. (3.12)
Moreover, the Zak transform is a unitary transformation from ๐ฟ2(โ) to ๐ฟ2([0, 1)2).Interest in this transform has been revived in recent years due to its relationshipto different types of coherent states, one of which is the affine coherent state,commonly known as wavelet [14, 20].
To analyse CliffordโGabor systems {๐2๐๐๐mโ x๐(x โ n) : m,n โ โค๐} in๐ฟ2(โ๐;๐ถโ๐,0), we need to establish a definition of the CliffordโZak transform andto show some of its properties.
First, we define the CliffordโZak transform pointwisely. The CliffordโZaktransform of a Clifford-valued function ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0), with t ๏ฟฝโ ๐(t), at thepoint (x, ๐) โ โ๐ ร โ๐, is given as
๐๐๐(x, ๐) :=โkโโค๐
๐2๐๐๐kโ ๐๐(xโ k). (3.13)
As in the one-dimensional case, periodicity properties allow us to consider asmaller domain for the variables (x, ๐) in this transform. In fact, the CliffordโZaktransform ๐๐ satisfies the following relations
๐๐๐(x, ๐ + n) = ๐๐๐(x, ๐), n โ โค๐, (3.14)
๐๐๐(x+ n, ๐) = ๐2๐๐๐nโ ๐๐๐๐(x, ๐), n โ โค๐. (3.15)
Thus, ๐๐๐ is uniquely determined by its values on [0, 1)๐ ร [0, 1)๐ := ๐2๐ โโ๐รโ๐. Henceforward, we consider ๐๐๐ as a function on ๐2๐, where its extensionto โ๐ ร โ๐ is trivially obtained by the quasiperiodic properties (3.14) and (3.15)of the transform.
Finally, we prove that the series defining ๐๐๐ converges in ๐ฟ2(๐2๐;๐ถโ๐,0).This will be achieved by showing that ๐๐ is a unitary map from ๐ฟ2(โ๐;๐ถโ๐,0)onto ๐ฟ2(๐2๐;๐ถโ๐,0).
Theorem 3.3. The CliffordโZak transform ๐๐ is a unitary map of ๐ฟ2(โ๐;๐ถโ๐,0)onto ๐ฟ2(๐2๐;๐ถโ๐,0).
Proof. Let ๐, ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0). In order to show that ๐๐๐ is a unitary mappingwe consider the auxiliary functions in ๐ฟ2(๐2๐;๐ถโ๐,0)
(x, ๐) ๏ฟฝโ ๐นk(x, ๐) := ๐2๐๐๐kโ ๐๐(xโ k), (3.16)
and
(x, ๐) ๏ฟฝโ ๐บk(x, ๐) := ๐2๐๐๐kโ ๐๐(xโ k), (3.17)
306 Y. Fu, U. Kahler and P. Cerejeiras
each obtained from the original ๐, ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0) by a specific modulation, anda translation, dependent on the parameter k โ โค๐. Note that we have
๐๐๐(x, ๐) =โkโโค๐
๐นk(x, ๐)
๐๐๐(x, ๐) =โkโโค๐
๐บk(x, ๐),(3.18)
for all (x, ๐) โ [0, 1)๐ ร [0, 1)๐. It is easy to see that ๐นk, ๐บk โ ๐ฟ2(๐2๐;๐ถโ๐,0), forall k โ โค๐.
We have
โจ๐นk, ๐บmโฉ๐ฟ2(๐2๐;๐ถโ๐,0)
=
[โซ๐2๐
๐2๐๐๐kโ ๐๐(xโ k) ห๐(xโm)๐โ2๐๐๐mโ ๐๐๐x๐๐๐]0
=
[โซ[0,1)๐ร[0,1)๐
๐(xโ k) ห๐(xโm)๐โ2๐๐๐(mโk)โ ๐๐๐x๐๐๐
]0
=
[(โซ[0,1)๐
๐(xโ k) ห๐(xโm)๐๐x
)(โซ[0,1)๐
๐โ2๐๐๐(mโk)โ ๐๐๐๐
)]0
, (3.19)
due to the cyclic property (2.5) and the relation ๐๐ = โ๐๐. Hence
โจ๐นk, ๐บmโฉ๐ฟ2(๐2๐;๐ถโ๐,0)=
{โจ๐(โ โ k), ๐(โ โ k)โฉ๐ฟ2([0,1)๐;๐ถโ๐,0)
, m = k
0 , m โ= k.(3.20)
Hence,
โจ๐๐๐, ๐๐๐โฉ๐ฟ2(๐2๐;๐ถโ๐,0)=
โk,mโโค๐
โจ๐นk, ๐บmโฉ๐ฟ2(๐2๐;๐ถโ๐,0)
=โkโโค๐
โจ๐นk, ๐บkโฉ๐ฟ2(๐2๐;๐ถโ๐,0)
=โkโโค๐
โจ๐(โ โ k), ๐(โ โ k)โฉ๐ฟ2([0,1)๐;๐ถโ๐,0)
=โkโโค๐
[โซ[0,1)๐
๐(xโ k) ห๐(x โ k)๐๐x
]0
=
โซโ๐
๐(y)๐(y)๐๐y = โจ๐, ๐โฉ๐ฟ2(โ๐;๐ถโ๐,0), (3.21)
therefore, completing the proof of the unitary property of the CliffordโZak trans-form. โก
Consequently, based on the above theorem, we get the following corollary.
15. The BalianโLow Theorem for the WCFT 307
Corollary 3.4. In particular, it holds
โฅ๐๐๐โฅ2๐ฟ2(๐2๐;๐ถโ๐,0)= โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ๐,0)
, (3.22)
where โฅ๐๐๐โฅ2๐ฟ2(๐2๐;๐ถโ๐,0)=โซ๐2๐ โฃ๐๐๐(x, ๐)โฃ2 ๐๐x๐๐๐.
The unitary nature of the CliffordโZak transform allows us to translate con-ditions on Clifford frames for ๐ฟ2(โ๐;๐ถโ๐,0) into those for ๐ฟ2(๐2๐;๐ถโ๐,0), wherethings are frequently easier to deal with.
Let us now define the space ๐ต as the set of all ๐น : โ๐ โ ๐ถโ๐,0 such that
๐น (x+ n, ๐) = ๐2๐๐๐nโ ๐๐น (x, ๐),
๐น (x, ๐ + n) = ๐น (x, ๐),
โฅ๐นโฅ2๐ฟ2(๐2๐;๐ถโ๐,0)=
โซ๐2๐
โฃ๐น (x, ๐)โฃ2 ๐๐x๐๐๐ <โ. (3.23)
In consequence, as ๐ต is a subset of ๐ฟ2(๐2๐;๐ถโ๐,0) the CliffordโZak transform ๐๐
is a unitary mapping between ๐ฟ2(โ๐;๐ถโ๐,0) and ๐ต.The following theorem provides some inversion formulas.
Theorem 3.5. If ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0) โฉ ๐ฟ1(โ๐;๐ถโ๐,0), then the following relations,
๐(x) =
โซ[0,1)๐
๐๐๐(x, ๐)๐๐๐, x โ โ๐, (3.24)
and
โฑ{๐ }(โ๐) =
โซ[0,1)๐
ห๐๐๐(x, ๐)๐2๐๐๐๐โ x๐๐x, ๐ โ โ๐, (3.25)
hold true, where โฑ denotes the CFT operator given by (2.10).
Before proceeding with the proof, we remark that the Zak transform can beextended in both arguments to the whole of โ๐ by relations (3.14) and (3.15). In
consequence, the above identities state that both the signal ๐, and the CFT โฑ{๐ },can be reconstructed on the whole of โ๐ via the CliffordโZak transform. Standarddensity arguments allow us to extend this result in an unique way to the whole of๐ฟ2(โ๐;๐ถโ๐,0).
Proof. By definition, there exist unique y โ [0, 1)๐ and n โ โค๐ such that x = y+n,andโซ
[0,1)๐๐๐๐(x, ๐)๐
๐๐ =
โซ[0,1)๐
๐๐๐(y + n, ๐)๐๐๐
=
โซ[0,1)๐
๐2๐๐๐nโ ๐๐๐๐(y, ๐)๐๐๐
=
โซ[0,1)๐
โkโโค๐
๐2๐๐๐(k+n)โ ๐๐(y โ k)๐๐๐
308 Y. Fu, U. Kahler and P. Cerejeiras
=
โซ[0,1)๐
โmโโค๐
๐2๐๐๐mโ ๐๐(y + nโm)๐๐๐
=
โซ[0,1)๐
โmโโค๐
๐2๐๐๐mโ ๐๐(xโm)๐๐๐
=
โซ[0,1)๐
๐(x)๐๐๐ +
โซ[0,1)๐
โmโ=0
๐2๐๐๐mโ ๐๐(xโm)๐๐๐
= ๐(x) +
โซ[0,1)๐
โmโ=0
๐2๐๐๐mโ ๐๐(xโm)๐๐๐. (3.26)
To calculate the remaining integral, we use Fubiniโs Theorem to validate the in-terchange between integration and summation so thatโซ
[0,1)๐
โmโ=0
๐2๐๐๐mโ ๐๐(xโm)๐๐๐
=โmโ=0
(โซ[0,1)๐
๐2๐๐๐mโ ๐๐๐๐
)๐(xโm) = 0. (3.27)
This completes the proof of the first identity. For the second, a direct calculationleads toโซ
[0,1)๐
ห๐๐๐(x, ๐)๐2๐๐๐๐โ x๐๐x =
โซ[0,1)๐
โkโโค๐
ห๐(xโ k)๐โ2๐๐๐kโ ๐๐2๐๐๐๐โ x๐๐x
=โkโโค๐
โซ[0,1)๐
ห๐(xโ k)๐2๐๐๐(xโk)โ ๐๐๐x (3.28)
=
โซโ๐
๐(y)๐2๐๐๐yโ ๐๐๐y = โฑ{๐}(โ๐). โก
In particular, we obtain the following reconstruction formula: for any ๐น โ ๐ต,we have (
๐โ1๐ ๐น
)(x) =
โซ[0,1)๐
๐น (x, ๐)๐๐๐, x โ โ๐. (3.29)
Lemma 3.6. If ๐m,n is defined by (3.1), then
๐๐๐m,n(x, ๐) = ๐โ2๐๐๐nโ ๐๐2๐๐๐mโ x๐๐๐(x, ๐), (x, ๐) โ ๐2๐. (3.30)
Proof. From the definitions of the CliffordโZak transform and of ๐m,n we obtain
๐๐๐m,n(x, ๐) =โkโโค๐
๐2๐๐๐kโ ๐๐2๐๐๐mโ (xโk)๐(xโ kโ n)
=โkโโค๐
๐2๐๐๐kโ ๐๐2๐๐๐mโ x๐(xโ kโ n)
= ๐2๐๐๐mโ xโkโโค๐
๐2๐๐๐kโ ๐๐(xโ kโ n)
15. The BalianโLow Theorem for the WCFT 309
= ๐2๐๐๐mโ xโkโโค๐
๐2๐๐๐(kโn)โ ๐๐(xโ k)
= ๐โ2๐๐๐nโ ๐๐2๐๐๐mโ xโkโโค๐
๐2๐๐๐kโ ๐๐(xโ k) (3.31)
= ๐โ2๐๐๐nโ ๐๐2๐๐๐mโ x๐๐๐(x, ๐). โก
Theorem 3.7. If ๐, ๐ โ ๐ฟ2(โ๐;โ โ ๐๐โ) and ๐m,n is defined as in (3.1) then wehave โ
m,nโโค๐
โฃโฃโจ๐, ๐m,nโฉ๐ฟ2(โ๐;๐ถโ๐,0)
โฃโฃ2 =โฅโฅโฅ๐๐๐๐๐๐
โฅโฅโฅ2๐ฟ2(๐2๐;๐ถโ๐,0)
. (3.32)
Proof. We remark that ๐ โ ๐ฟ2(โ๐;โ โ ๐๐โ) โ ๐ฟ2(โ๐;๐ถโ๐,0) so that we haveโฅ๐โฅ๐ฟ2(โ๐;โโ๐๐โ) = โฅ๐โฅ๐ฟ2(โ๐;๐ถโ๐,0)
. Then, based on Theorem 3.3 and Lemma 3.6,
we obtain thatโm,nโโค๐
โฃโฃโจ๐, ๐m,nโฉ๐ฟ2(โ๐;๐ถโ๐,0)
โฃโฃ2=
โm,nโโค๐
โฃโฃโจ๐๐๐, ๐๐๐m,nโฉ๐ฟ2(๐2๐;๐ถโ๐,0)
โฃโฃ2=
โm,nโโค๐
โฃโฃโฃ[โซ๐2๐ ๐๐๐๐๐๐m,n๐๐x๐๐๐
]0
โฃโฃโฃ2=
โm,nโโค๐
โฃโฃโฃ[โซ๐2๐ ๐๐๐๐๐๐๐2๐๐๐nโ ๐๐โ2๐๐๐mโ x๐๐x๐๐๐
]0
โฃโฃโฃ2 (3.33)
=โ
m,nโโค๐
โฃโฃโฃโฃโจ๐๐๐๐๐๐, ๐ธm,n
โฉ๐ฟ2(๐2๐;๐ถโ๐,0)
โฃโฃโฃโฃ2=โฅโฅโฅ๐๐๐๐๐๐
โฅโฅโฅ2๐ฟ2(๐2๐;๐ถโ๐,0)
,
where the set of all ๐ธm,n := ๐โ2๐๐๐nโ ๐๐2๐๐๐mโ x constitutes an orthonormal basisfor ๐ฟ2(๐2๐;โโ ๐๐โ). โก
4. BalianโLow Theorem for WCFT
Before proceeding with the BalianโLow theorem for a CliffordโGabor frame, letus recall some basic facts on the modulus of a Clifford number. In general, themodulus of arbitrary Clifford numbers is not multiplicative. In fact, for any twoelements ๐, ๐ โ ๐ถโ๐,0 we have โฃ๐๐โฃ โค 2
๐2 โฃ๐โฃ โฃ๐โฃ. However, in some special cases, the
multiplicative property does hold. An easy calculation leading to such a case isdescribed in the following lemma.
Lemma 4.1. Let ๐ โ ๐ถโ๐,0 be such that ๐๏ฟฝ๏ฟฝ = โฃ๐โฃ2. Thenโฃ๐๐โฃ = โฃ๐โฃ โฃ๐โฃ , โ๐ โ ๐ถโ๐,0. (4.1)
310 Y. Fu, U. Kahler and P. Cerejeiras
In this section, and since we need the CliffordโZak transform ๐๐ of the windowfunction ๐ to satisfy โฃ๐๐๐๐โฃ = โฃ๐ โฃ โฃ๐๐๐โฃ for any multivector ๐ โ ๐ถโ๐,0, it is necessaryto impose some restrictions on the non-zero window function ๐. In what followswe will require
๐ โ ๐ฟ2(โ๐;โโ ๐๐โ). (4.2)
Some simple examples of possible real-valued window functions are the two-dimen-sional Gaussian function and the two-dimensional first-order B-spline [23], whichcan be generalized to ๐ dimensions and to cases of linear combinations of real-valued window functions and pseudoscalars.
Lemma 4.2. Suppose that the non-zero window function ๐ satisfies (4.2). Then wehave
๐๐๐๐๐๐ = โฃ๐๐๐โฃ2 , (4.3)
and for any multivector ๐ โ ๐ถโ๐,0
โฃ๐๐๐๐โฃ = โฃ๐โฃ โฃ๐๐๐โฃ . (4.4)
Proof. Since ๐ โ ๐ฟ2(โ๐;โ โ ๐๐โ), an easy computation shows that there exist๐บ,๐ป โ ๐ฟ2(๐2๐;โ) such that
๐๐๐ = ๐บ+ ๐๐๐ป. (4.5)
Thus, based on the fact that ๐๐ = โ๐๐ and ๏ฟฝ๏ฟฝ = ๐บ, ๏ฟฝ๏ฟฝ = ๐ป for ๐บ,๐ป โ ๐ฟ2(๐2๐;โ),it follows that
๐๐๐๐๐๐ = (๐บ+ ๐๐๐ป)(๏ฟฝ๏ฟฝ+ ๏ฟฝ๏ฟฝ๐๐) = (๐บ+ ๐๐๐ป)(๐บโ๐ป๐๐)
= ๐บ๐บโ๐บ๐ป๐๐ + ๐๐๐ป๐บโ ๐๐๐ป๐ป๐๐
= ๐บ2 โ๐ป๐บ๐๐ + ๐๐๐ป๐บโ ๐๐๐ป2๐๐. (4.6)
Note that the pseudoscalar ๐๐ commutes with the scalar elements of the algebra.Thus,
๐ป๐บ๐๐ = ๐๐๐ป๐บ, ๐บ2 = โฃ๐บโฃ2 , ๐ป2 = โฃ๐ป โฃ2 . (4.7)
Substituting (4.7) into (4.6) leads to
๐๐๐ ๐๐๐ = โฃ๐บโฃ2 + โฃ๐ป โฃ2 = โฃ๐๐๐โฃ2 . (4.8)
Moreover, by Lemma 4.1 we see that for any ๐ โ ๐ถโ๐,0
โฃ๐๐๐๐โฃ = โฃ๐โฃ โฃ๐๐๐โฃ , (4.9)
which completes the proof. โก
Based on the properties of the CliffordโZak transform discussed in the pre-vious section, we are going to study the time-frequency localization property ofa CliffordโGabor system {๐m,n : m,n โ โค๐} for ๐ฟ2(โ๐;โ โ ๐๐โ), which is thecontent of the following theorem. This theorem will enable us later on to derivethe Clifford version of the BalianโLow theorem.
15. The BalianโLow Theorem for the WCFT 311
Theorem 4.3. Suppose that {๐m,n(x) : m,n โ โค๐} constitutes a frame for๐ฟ2(โ๐;โ โ ๐๐โ) with a non-zero window function ๐ satisfying (4.2). Then wehave
โณ๐ฅ๐โณ๐๐ =โ, ๐ = 1, 2, . . . , ๐, (4.10)
where
โณ๐ฅ๐ =
โซโ๐
๐ฅ2๐ โฃ๐(x)โฃ2 ๐๐x, โณ๐๐ =
โซโ๐
๐2๐ โฃโฑ๐(๐)โฃ2 ๐๐๐ (4.11)
and the CFT of ๐, โฑ๐(๐), is defined by (2.10).
We will divide the proof by demonstrating a sequence of lemmas. Denote by๐ท๐ and ๐๐ the following partial derivative and multiplication operators
๐ท๐๐(x) := โ๐ฅ๐๐(x)
1
2๐๐๐, ๐๐๐(x) := ๐ฅ๐๐(x), ๐ = 1, 2, . . . , ๐, (4.12)
where โ๐ฅ๐:= โ
โ๐ฅ๐. One can easily see that the product of these operators depends
on their order. In fact, they satisfy the commutation relation
[๐๐, ๐ท๐]๐(x) := (๐๐๐ท๐ โ๐ท๐๐๐) ๐(x) = โ๐(x)1
2๐๐๐. (4.13)
This is traditionally expressed by saying that the time and frequency variables arecanonically conjugate. In the first place, we review the following lemma which wasshown in [22, 24].
Lemma 4.4 (CFT partial derivative). The CFT of โ๐ฅ๐๐(x) โ ๐ฟ2(โ๐;๐ถโ๐,0) is
given by
โฑ{โ๐ฅ๐๐(x)}(๐) = 2๐๐๐โฑ{๐}(๐)๐๐, (4.14)
that is,
โฑ{๐ท๐๐(x)}(๐) = ๐๐โฑ{๐}(๐), ๐ = 1, 2, . . . , ๐. (4.15)
Lemma 4.5. Let ๐, ๐ โ ๐ฟ2(โ๐;โโ ๐๐โ). If we have
๐ท๐๐,๐ท๐๐,๐๐๐,๐๐๐ โ ๐ฟ2(โ๐;โโ ๐๐โ), ๐ = 1, 2, . . . , ๐, (4.16)
then the following relation holds
โจ๐๐๐,๐ท๐๐โฉ โ โจ๐ท๐๐,๐๐๐โฉ = ยฑ 1
2๐๐๐โจ๐, ๐โฉ , (4.17)
where the scalar inner product โจโ , โ โฉ is given by (2.8).
Proof. Let us choose ๐๐ , ๐๐ โ ๐ฎ(โ๐;โโ ๐๐โ), where ๐ฎ(โ๐;โโ ๐๐โ) denotes theSchwartz class. We recall that ๐ฎ(โ๐;โโ๐๐โ) is a dense subspace in ๐ฟ2(โ๐;โโ๐๐โ)and it is defined as the set of all smooth functions from โ๐ to โโ ๐๐โ such thatall of its partial derivatives are rapidly decreasing. Therefore, we have that if๐๐ โ ๐, ๐๐ โ ๐, then ๐ท๐๐๐ โ ๐ท๐๐,๐๐๐๐ โ ๐๐๐ and ๐ท๐๐๐ โ ๐ท๐๐,๐๐๐๐ โ๐๐๐, ๐ = 1, 2, . . . , ๐. Moreover, the convergence is in the ๐ฟ2-sense. Since it is easyto check that for fixed ๐ the operators ๐ท๐ and ๐๐ are self-adjoint, we obtain
โจ๐๐๐๐ , ๐ท๐๐๐โฉ โ โจ๐ท๐๐๐ ,๐๐๐๐โฉ = โจ๐ท๐๐๐๐๐ , ๐๐โฉ โ โจ๐๐๐ท๐๐๐ , ๐๐โฉ
312 Y. Fu, U. Kahler and P. Cerejeiras
= โโจ[๐๐, ๐ท๐]๐๐ , ๐๐โฉ =โจ๐๐
1
2๐๐๐, ๐๐
โฉ= ยฑ 1
2๐๐๐โจ๐๐ , ๐๐โฉ , (4.18)
where we use the fact that the pseudoscalar ๐๐ commutes with the scalar elementsof the algebra. Moreover, since the scalar inner product is continuous, the desiredresult holds in the limit. โก
The utility of the CliffordโZak transform ๐๐ for constructing Gabor basesstems from the following result.
Lemma 4.6. Suppose that the non-zero window function ๐ satisfies (4.2). Then wehave
1. {๐m,n : m,n โ โค๐} is a frame for ๐ฟ2(โ๐;โโ ๐๐โ) if and only if
0 < ๐ต โค โฃ๐๐๐(x, ๐)โฃ2 โค ๐ถ <โ ๐.๐. (x, ๐) โ ๐2๐. (4.19)
2. {๐m,n : m,n โ โค๐} is an orthonormal basis for ๐ฟ2(โ๐;โโ ๐๐โ) if and only
if โฃ๐๐๐(x, ๐)โฃ2 = 1 for almost all (x, ๐) โ ๐2๐.
Proof. Let us first prove statement 1. Assume that {๐m,n : m,n โ โค๐} is a framefor ๐ฟ2(โ๐;โโ ๐๐โ). According to the definition, there exist 0 < ๐ต โค ๐ถ <โ suchthat for all ๐ โ ๐ฟ2(โ๐;โโ ๐๐โ), it holds
๐ต โฅ๐โฅ2๐ฟ2(โ๐;โโ๐๐โ) โคโ
m,nโโค๐
โฃโจ๐, ๐m,nโฉโฃ2 โค ๐ถ โฅ๐โฅ2๐ฟ2(โ๐;โโ๐๐โ) . (4.20)
Then ๐น = ๐๐๐ โ ๐ฟ2(๐2๐;โโ ๐๐โ) and by Corollary 3.4 and Theorem 3.7 we get
๐ต โฅ๐นโฅ2๐ฟ2(๐2๐;โโ๐๐โ) โคโฅโฅโฅ๐น๐๐๐
โฅโฅโฅ2๐ฟ2(๐2๐;โโ๐๐โ)
โค ๐ถ โฅ๐นโฅ2๐ฟ2(๐2๐;โโ๐๐โ) , (4.21)
which implies that ๐ต โค โฃ๐๐๐โฃ2 โค ๐ถ ๐.๐. due to Lemma 4.2. Conversely, if ๐ต โคโฃ๐๐๐โฃ2 โค ๐ถ ๐.๐., then by Theorem 3.7 and Corollary 3.4, we conclude that for all๐ โ ๐ฟ2(โ๐;โโ ๐๐โ)
๐ต โฅ๐โฅ2๐ฟ2(โ๐;โโ๐๐โ) โคโ
m,nโโค๐
โฃโจ๐, ๐m,nโฉโฃ2 =โฅโฅโฅ๐๐๐๐๐๐
โฅโฅโฅ2๐ฟ2(๐2๐;โโ๐๐โ)
โค ๐ถ โฅ๐โฅ2๐ฟ2(โ๐;โโ๐๐โ) . (4.22)
Now, let us take a look at statement 2. If {๐m,n : m,n โ โค๐} is an orthonor-mal basis for ๐ฟ2(โ๐;โโ ๐๐โ) then, for all ๐ โ ๐ฟ2(โ๐;โโ ๐๐โ), it holdsโ
m,nโโค๐
โฃโจ๐, ๐m,nโฉโฃ2 = โฅ๐โฅ2๐ฟ2(โ๐;โโ๐๐โ) = โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ๐,0)
= โฅ๐๐๐โฅ2๐ฟ2(๐2๐;๐ถโ๐,0). (4.23)
Thus, by Theorem 3.7 we getโฅโฅโฅ๐๐๐๐๐๐โฅโฅโฅ2๐ฟ2(๐2๐;๐ถโ๐,0)
= โฅ๐๐๐โฅ2๐ฟ2(๐2๐;๐ถโ๐,0), (4.24)
15. The BalianโLow Theorem for the WCFT 313
which implies that โฃ๐๐๐(x, ๐)โฃ2 = 1 for almost all (x, ๐) โ ๐2๐ due to Lemma 4.2.
Conversely, if โฃ๐๐๐(x, ๐)โฃ2 = 1 for almost all (x, ๐) โ ๐2๐, then by Theorem 3.7,{๐m,n : m,n โ โค๐} is a tight frame for ๐ฟ2(โ๐;โโ ๐๐โ). Moreover, Corollary 3.4
yields โฅ๐โฅ2๐ฟ2(โ๐;๐ถโ๐,0)=โฅ๐๐๐โฅ2๐ฟ2(๐2๐;๐ถโ๐,0)
=1, which leads to โฅ๐m,nโฅ2๐ฟ2(โ๐;๐ถโ๐,0)=1.
Consequently, {๐m,n : m,n โ โค๐} is an orthonormal basis for ๐ฟ2(โ๐;โโ๐๐โ). โก
Combining Corollary 3.4, Theorem 3.7 and (3.7), we can assert that
โจ๐๐๐๐, ๐๐๐โฉ =โฅโฅโฅ๐๐๐๐๐๐
โฅโฅโฅ2๐ฟ2(๐2๐;๐ถโ๐,0)
, (4.25)
where ๐ is the frame operator defined by (3.2). Thus, we get that ๐๐๐๐โ1๐ corre-
sponds to a multiplication by โฃ๐๐๐โฃ2 on the space ๐ต defined by (3.23).The following lemma shows that โ๐ฅ1๐๐๐ โ ๐ฟ2(๐2๐;โ โ ๐๐โ) implies that
โ๐ฅ1 โฃ๐๐๐โฃ โ ๐ฟ2(๐2๐;โ) under a certain condition. In a similar way, we can get thecorresponding results for โ๐ฅ2๐๐๐, โ๐1๐๐๐ and โ๐2๐๐๐ in ๐ฟ2(๐2๐;โโ ๐๐โ).
Lemma 4.7. Under the hypotheses of Theorem 4.3 we see that if โ๐ฅ1๐๐๐ โ๐ฟ2(๐2๐;โโ ๐๐โ) then โ๐ฅ1 โฃ๐๐๐โฃ โ ๐ฟ2(๐2๐;โ) โ ๐ฟ2(๐2๐;โโ ๐๐โ).
Proof. Since ๐ โ ๐ฟ2(โ๐;โ โ ๐๐โ), an easy computation shows that there exist๐บ,๐ป โ ๐ฟ2(๐2๐;โโ ๐๐โ) such that
๐๐๐ = ๐บ+ ๐๐๐ป, ๐บ,๐ป โ โ. (4.26)
Thus โ๐ฅ1๐๐๐ = โ๐ฅ1๐บ+ โ๐ฅ1๐ป๐๐ and
โฅโ๐ฅ1๐๐๐โฅ2 =
โซ๐2๐
โฃโ๐ฅ1๐๐๐โฃ2 ๐๐x๐๐๐
=
โซ๐2๐
((โ๐ฅ1๐บ)2 + (โ๐ฅ1๐ป)2
)๐๐x๐๐๐ <โ, (4.27)
which means that
โฅโ๐ฅ1๐บโฅ <โ, โฅโ๐ฅ1๐ปโฅ <โ. (4.28)
Moreover, since {๐m,n : m,n โ โค๐} is a frame for ๐ฟ2(โ๐;โโ๐๐โ) by Lemma4.6 we have
0 < ๐ต โค โฃ๐๐๐(x, ๐)โฃ2 โค ๐ถ <โ, ๐.๐. (x, ๐) โ ๐2๐, (4.29)
which tells us that
๐บ2 โค ๐ถ <โ, ๐ป2 โค ๐ถ <โ. (4.30)
A simple calculation leads to
โ๐ฅ1 โฃ๐๐๐โฃ = โฃ๐๐๐โฃโ1(๐บโ๐ฅ1๐บ+๐ปโ๐ฅ1๐ป). (4.31)
Now, based on (4.28), (4.30), and the Minkowski inequality in ๐ฟ2(๐2๐;โ), it followsthat
โฅโ๐ฅ1 โฃ๐๐๐โฃโฅ2 =
โซ๐2๐
โฃ๐๐๐โฃโ2 โฃ(๐บโ๐ฅ1๐บ+๐ปโ๐ฅ1๐ป)โฃ2 ๐๐x๐๐๐
314 Y. Fu, U. Kahler and P. Cerejeiras
โค ๐ตโ1
โซ๐2๐
โฃ(๐บโ๐ฅ1๐บ+๐ปโ๐ฅ1๐ป)โฃ2 ๐๐x๐๐๐
โค ๐ตโ1(โฅ๐บโ๐ฅ1๐บโฅ+ โฅ๐ปโ๐ฅ1๐ปโฅ)2 (4.32)
โค ๐ตโ1๐ถ(โฅโ๐ฅ1๐บโฅ+ โฅโ๐ฅ1๐ปโฅ)2 <โ. โก
Now, let us consider the dual frame ๐m,n := ๐โ1๐m,n given by (3.9). Since๐๐๐๐
โ1๐ corresponds to a multiplication by โฃ ๐๐๐ โฃ2 on the space ๐ต, by Lemma 4.6
it follows that
๐๐๐m,n = ๐๐๐โ1๐โ1
๐ ๐๐๐m,n = โฃ๐๐๐โฃโ2๐๐๐m,n (4.33)
or
๐๐๐m,n = โฃ๐๐๐โฃโ2๐โ2๐๐๐nโ ๐๐2๐๐๐mโ x๐๐๐
= ๐โ2๐๐๐nโ ๐๐2๐๐๐mโ x๐๐๐, (4.34)
which belongs to the space ๐ต with ๐๐๐ = โฃ๐๐๐โฃโ2 ๐๐๐. In particular, (4.34) impliesthat
๐m,n(x) := ๐2๐๐๐mโ x๐(xโ n), (4.35)
which proves the following lemma associated with Lemma 4.2.
Lemma 4.8. Under the hypotheses of Theorem 4.3 the dual window function ๐
satisfies ๐๐๐ = โฃ๐๐๐โฃโ2๐๐๐ and ๐๐๐ ๐๐๐ = 1.
Analogous to the Minkowski inequality in ๐ฟ2(โ๐;โ), by the definition ofthe norm (2.6) and the CliffordโCauchyโSchwarz inequality (2.9), the followingCliffordโMinkowski inequality holds true in ๐ฟ2(๐2๐;๐ถโ๐,0). This will be necessaryin the proof of Theorem 4.3.
Lemma 4.9. For ๐, ๐ โ ๐ฟ2(๐2๐;๐ถโ๐,0), we have ๐+ ๐ โ ๐ฟ2(๐2๐;๐ถโ๐,0) and
โฅ๐+ ๐โฅ โค โฅ๐โฅ + โฅ๐โฅ . (4.36)
Now, we are in position to demonstrate Theorem 4.3.
Proof. We prove the theorem by contradiction. Suppose that {๐m,n : m,n โ โค๐}is a frame for ๐ฟ2(โ๐;โโ ๐๐โ) with a non-zero window function ๐ satisfying (4.2),and, furthermore, suppose that
โณ๐ฅ๐ =
โซโ๐
๐ฅ2๐ โฃ๐(x)โฃ2 ๐๐x <โ, (4.37)
and
โณ๐๐ =
โซโ๐
๐2๐ โฃโฑ๐(๐)โฃ2 ๐๐๐ <โ, ๐ = 1, 2, . . . , ๐. (4.38)
Now, based on Lemmas 2.2 and 4.4, we get ๐๐๐,๐ท๐๐ โ ๐ฟ2(โ๐;โ โ ๐๐โ), wherethe multiplication and partial derivative operators ๐๐, ๐ท๐ are defined by (4.12).
One has that
[๐๐(๐๐๐)] (x, ๐) = ๐ฅ๐๐๐๐ โ 1
2๐๐๐โ๐๐
๐๐๐, (4.39)
15. The BalianโLow Theorem for the WCFT 315
which means that ๐๐๐ โ ๐ฟ2(โ๐;โโ๐๐โ) if and only if โ๐๐๐๐๐ โ ๐ฟ2(๐2๐;โโ๐๐โ)
by Theorem 3.4 and Lemma 4.9. In a similar way, we find that ๐ท๐๐ โ ๐ฟ2(โ๐;โโ๐๐โ) if and only if โ๐ฅ๐
๐๐๐ โ ๐ฟ2(๐2๐;โโ ๐๐โ). By Lemma 4.8 we know that the
dual window function ๐ satisfies ๐๐๐ = โฃ๐๐๐โฃโ2 ๐๐๐. Consequently, we see that
โ๐ฅ๐(๐๐๐) = โ๐ฅ๐
(โฃ๐๐๐โฃโ2)๐๐๐ + โฃ๐๐๐โฃโ2
โ๐ฅ๐๐๐๐
= โ2 โฃ๐๐๐โฃโ3๐๐๐ โ๐ฅ๐
โฃ๐๐๐โฃ+ โฃ๐๐๐โฃโ2โ๐ฅ๐
๐๐๐ (4.40)
andโ๐๐
(๐๐๐) = โ2 โฃ๐๐๐โฃโ3๐๐๐ โ๐๐
โฃ๐๐๐โฃ+ โฃ๐๐๐โฃโ2โ๐๐
๐๐๐ (4.41)
are in ๐ฟ2(๐2๐;โ โ ๐๐โ) by Lemmas 4.9, 4.6, and 4.7. Hence, we conclude that๐๐๐,๐ท๐๐ โ ๐ฟ2(โ๐;โ โ ๐๐โ). For the functions ๐ and ๐, we shall next prove thefact
โจ๐๐๐,๐ท๐๐โฉ = โจ๐ท๐๐,๐๐๐โฉ , (4.42)
from which we derive the contradiction. In fact, in the first place, by Corollary 3.4and Lemma 4.8 we find that
โจ๐, ๐m,nโฉ = โจ๐๐๐, ๐๐๐m,nโฉ =[โซ
๐2๐
๐๐๐๐๐๐๐2๐๐๐nโ ๐๐โ2๐๐๐mโ x๐๐x๐๐๐
]0
=
[โซ๐2๐
๐2๐๐๐nโ ๐๐โ2๐๐๐mโ x๐๐x๐๐๐]0
= ๐ฟm,0๐ฟn,0, (4.43)
where ๐ฟm,k denotes the Kronecker delta function. Similarly, we can check that
โจ๐, ๐m,nโฉ = ๐ฟm,0๐ฟn,0. (4.44)
Now, since ๐๐๐,๐ท๐๐ โ ๐ฟ2(โ๐;โโ ๐๐โ) and {๐m,n}, {๐m,n} constitute dualframes for ๐ฟ2(โ๐;โโ ๐๐โ), we obtain
โจ๐๐๐,๐ท๐๐โฉ =โ
m,nโโค๐
โจ๐๐๐, ๐m,nโฉ โจ๐m,n, ๐ท๐๐โฉ . (4.45)
Based on (4.35), (4.44), and (2.5), a simple calculation leads to
โจ๐โm,โn,๐๐๐โฉ =[โซ
โ๐
๐โ2๐๐๐mโ x๐(x+ n)๐ฅ๐๐(x)๐๐x
]0
=
[โซโ๐
๐โ2๐๐๐mโ x๐(x+ n)๐ฅ๐ ห๐(x)๐๐x]0
=
[โซโ๐
๐ฅ๐๐(x+ n) ห๐(x)๐โ2๐๐๐mโ x๐๐x]0
=
[โซโ๐
(๐ฅ๐ โ ๐๐)๐(x) ห๐(xโ n)๐โ2๐๐๐mโ (xโn)๐๐x
]0
= โจ๐๐๐, ๐m,nโฉ โ ๐๐ โจ๐, ๐m,nโฉ = โจ๐๐๐, ๐m,nโฉ . (4.46)
On the other hand, by (4.43) and (2.5) we obtain
โจ๐ท๐๐, ๐โm,โnโฉ =[โซ
โ๐
โ๐ฅ๐๐(x)
1
2๐๐๐ห๐(x+ n)๐2๐๐๐mโ x๐๐x
]0
316 Y. Fu, U. Kahler and P. Cerejeiras
= 0โ[โซ
โ๐
๐(x)1
2๐๐๐โ๐ฅ๐
(ห๐(x+ n)๐2๐๐๐mโ x
)๐๐x
]0
= โ
โกโขโขโฃโซโ๐
๐(x)1
2๐๐๐
โโโโโ๐ฅ๐
ห๐(x+ n)๐2๐๐๐mโ x
+
ห๐(x+ n)๐๐2๐๐๐๐2๐๐๐mโ x
โโโโ ๐๐x
โคโฅโฅโฆ0
=
[โซโ๐
๐2๐๐๐mโ x๐(xโ n)๐ท๐๐(x)๐๐x
]0
โ๐๐
[โซโ๐
๐(x)1
2๐๐๐ห๐(x+ n)2๐๐๐๐
2๐๐๐mโ x๐๐x]0
= โจ๐m,n, ๐ท๐๐โฉ ยฑ๐๐ โจ๐m,n, ๐โฉ = โจ๐m,n, ๐ท๐๐โฉ . (4.47)
Consequently, substituting (4.46) and (4.47) into (4.45), we get
โจ๐๐๐,๐ท๐๐โฉ =โ
m,nโโค๐
โจ๐โm,โn,๐๐๐โฉ โจ๐ท๐๐, ๐โm,โnโฉ
=โ
m,nโโค๐
โจ๐ท๐๐, ๐โm,โnโฉ โจ๐โm,โn,๐๐๐โฉ = โจ๐ท๐๐,๐๐๐โฉ , (4.48)
which means that
โจ๐, ๐โฉ = 0 (4.49)
by Lemma 4.5. However, by (4.44), taking m = n = 0, we have
โจ๐, ๐โฉ = 1, (4.50)
which leads to the contradiction. Thus, we get
โณ๐ฅ๐โณ๐๐ =โ, ๐ = 1, 2, . . . , ๐. (4.51)
โก
We are now in a position to extend Theorem 4.3 to the case of an arbitraryClifford frame.
Theorem 4.10 (BalianโLow theorem). Suppose that {๐m,n(x) : m,n โ โค๐} con-stitutes a frame for ๐ฟ2(โ๐;๐ถโ๐,0) associated to a non-zero window function ๐ in๐ฟ2(โ๐;๐ถโ๐,0). Then we have
โณ๐ฅ๐โณ๐๐ =โ, ๐ = 1, 2, . . . , ๐, (4.52)
where
โณ๐ฅ๐ =
โซโ๐
๐ฅ2๐ โฃ๐(x)โฃ2 ๐๐x, โณ๐๐ =
โซโ๐
๐2๐ โฃโฑ๐(๐)โฃ2 ๐๐๐ (4.53)
and the CFT of ๐, โฑ๐(๐), is defined by (2.10).
Proof. If {๐m,n : m,n โ โค๐} constitutes a frame for ๐ฟ2(โ๐;๐ถโ๐,0) then the system{โm,n = [๐m,n]0 + ๐๐[๐m,n]๐ : m,n โ โค๐} is a frame for ๐ฟ2(โ๐;โ โ ๐๐โ). ByTheorem 4.3 the result follows. โก
15. The BalianโLow Theorem for the WCFT 317
5. Conclusions
The classical BalianโLow theorem is a strong form of the uncertainty principle forGabor systems which can be obtained by discretization of the kernel of the WFT.The WCFT is a generalization of the WFT in the framework of Clifford analysis.In this chapter, associated with the discretization of the kernel of the WCFT, weestablished a new kind of Gabor system
๐m,n(x) := ๐2๐๐๐mโ x๐(xโ n), m,n โ โค๐, (5.1)
which constitutes a Clifford frame satisfying the frame condition
๐ต โฅ๐โฅ2 โคโ
m,nโโค๐
โฃโจ๐, ๐m,nโฉโฃ2 โค ๐ถ โฅ๐โฅ2 , ๐ต, ๐ถ > 0, โ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0).
(5.2)To analyse Gabor systems {๐m,n(x) : m,n โ โค๐} in ๐ฟ2(โ๐, ๐ถโ๐,0), we establishedthe definition of the CliffordโZak transform for ๐ โ ๐ฟ2(โ๐;๐ถโ๐,0) by
๐๐๐(x, ๐) :=โkโโค๐
๐2๐๐๐kโ ๐๐(xโ k), (x, ๐) โ ๐2๐, (5.3)
and showed some properties of it. Furthermore, we proved the correspondingBalianโLow theorem for such Gabor systems which form a Clifford frame.
Acknowledgment
The first author is the recipient of a postdoctoral grant from Fundacao para aCiencia e a Tecnologia, ref. SFRH/BPD/46250/2008. This work was supportedby FEDER funds through COMPETE โ Operational Programme Factors of Com-petitiveness (โPrograma Operacional Factores de Competitividadeโ) and by Por-tuguese funds through the Center for Research and Development in Mathematicsand Applications (University of Aveiro) and the Portuguese Foundation for Scienceand Technology (โFCT โ Fundacao para a Ciencia e a Tecnologiaโ), within projectPEst-C/MAT/UI4106/2011 with COMPETE number FCOMP-01-0124-FEDER-022690. The work was also supported by the Foundation of Hubei EducationalCommittee (No. Q20091004) and the NSFC (No. 11026056).
References
[1] M. Bahri. Generalized Fourier transform in Clifford algebra ๐ถ๐(0, 3). Far East Jour-nal of Mathematical Science, 44(2):143โ154, Sept. 2010.
[2] M. Bahri. Generalized Fourier transform in real clifford algebra ๐๐(0, ๐). Far EastJournal of Mathematical Sciences, 48(1):11โ24, Jan. 2011.
[3] M. Bahri, S. Adji, and J.M. Zhao. Real Clifford windowed Fourier transform. ActaMathematica Sinica, 27(3):505โ518, 2011.
[4] M. Bahri, E. Hitzer, R. Ashino, and R. Vaillancourt. Windowed Fourier transformof two-dimensional quaternionic signals. Applied Mathematics and Computation,216(8):2366โ2379, June 2010.
318 Y. Fu, U. Kahler and P. Cerejeiras
[5] R. Balan. Extensions of no-go theorems to many signal systems. In A. Aldroubi andE.B. Lin, editors, Wavelets, Multiwavelets, and Their Applications, volume 216 ofContemporary Mathematics, pages 3โ14. American Mathematical Society, 1998.
[6] I.J. Benedetto, W. Czaja, and A.Y. Maltsev. The BalianโLow theorem for the sym-plectic form on โ2๐. Journal of Mathematical Physics, 44(4):1735โ1750, 2003.
[7] J.J. Benedetto, C. Heil, and D.F. Walnut. Differentiation and the BalianโLow the-orem. Journal of Fourier Analysis and Applications, 1(4):355โ402, 1994.
[8] F. Brackx, N. De Schepper, and F. Sommen. The two-dimensional CliffordโFouriertransform. Journal of Mathematical Imaging and Vision, 26(1):5โ18, 2006.
[9] F. Brackx, N. De Schepper, and F. Sommen. The Fourier transform in Cliffordanalysis. Advances in Imaging and Electron Physics, 156:55โ201, 2009.
[10] F. Brackx, R. Delanghe, and F. Sommen. Clifford Analysis, volume 76. Pitman,Boston, 1982.
[11] T. Bulow, M. Felsberg, and G. Sommer. Non-commutative hypercomplex Fouriertransforms of multidimensional signals. In G. Sommer, editor, Geometric computingwith Clifford Algebras: Theoretical Foundations and Applications in Computer Visionand Robotics, pages 187โ207, Berlin, 2001. Springer.
[12] T. Bulow and G. Sommer. Quaternionic Gabor filters for local structure classifi-cation. In Proceedings Fourteenth International Conference on Pattern Recognition,volume 1, pages 808โ810, 16โ20 Aug 1998.
[13] W. Czaja and A.M. Powell. Recent developments in the BalianโLow theorem. InC. Heil, editor, Harmonic Analysis and Applications, pages 79โ100. Birkhauser,Boston, 2006.
[14] I. Daubechies. Ten Lectures on Wavelets. Society for Industrial and Applied Math-ematics, 1992.
[15] J. Ebling and G. Scheuermann. Clifford Fourier transform on vector fields. IEEETransactions on Visualization and Computer Graphics, 11(4):469โ479, JulyAug.2005.
[16] J. Ebling and J. Scheuermann. Clifford convolution and pattern matching on vectorfields. In Proceedings IEEE Visualization, volume 3, pages 193โ200, Los Alamitos,CA, 2003. IEEE Computer Society.
[17] M. Felsberg. Low-Level Image Processing with the Structure Multivector. PhD thesis,Christian-Albrechts-Universitat, Institut fur Informatik und Praktische Mathematik,Kiel, 2002.
[18] K. Grochenig. Foundations of Time-Frequency Analysis. Applied and Numerical Har-monic Analysis. Birkhauser, Boston, 2001.
[19] K. Grochenig, D. Han, and G. Kutyniok. The BalianโLow theorem for the symplecticlattice in higher dimensions. Applied and Computational Harmonic Analysis, 13:169โ176, 2002.
[20] C. Heil and D.F. Walnut. Continuous and discrete wavelet transforms. SIAM Review,31:628โ666, 1989.
[21] S. Held, M. Storath, P. Massopust, and B. Forster. Steerable wavelet frames based onthe Riesz transform. IEEE Transactions on Image Processing, 19(3):653โ667, 2010.
15. The BalianโLow Theorem for the WCFT 319
[22] E.M.S. Hitzer and B. Mawardi. Clifford Fourier transform on multivector fields anduncertainty principles for dimensions ๐ = 2(mod 4) and ๐ = 3(mod 4). Advances inApplied Clifford Algebras, 18(3-4):715โ736, 2008.
[23] B. Mawardi, E. Hitzer, and S. Adji. Two-dimensional Clifford windowed Fouriertransform. In E. Bayro-Corrochano and G. Scheuermann, editors, Applied GeometricAlgebras in Computer Science and Engineering, pages 93โ106. Springer, London,2010.
[24] B. Mawardi and E.M.S. Hitzer. Clifford Fourier transformation and uncertainty prin-ciple for the Clifford algebra ๐ถโ3,0. Advances in Applied Clifford Algebras, 16(1):41โ61, 2006.
[25] M. Porat, Y.Y. Zeevi, and M. Zibulski. Multi-window Gabor schemes in signal andimage representations. In H.G. Feichtinger and T. Strohmer, editors, Gabor Analysisand Algorithms, pages 381โ408. Birkhauser, 1998.
[26] S.J. Sangwine and T.A. Ell. Hypercomplex Fourier transforms of color images. InIEEE International Conference on Image Processing (ICIP 2001), volume I, pages137โ140, Thessaloniki, Greece, 7โ10 Oct. 2001. IEEE.
[27] F. Sommen. Hypercomplex Fourier and Laplace transforms I. Illinois Journal ofMathematics, 26(2):332โ352, 1982.
[28] Y.Y. Zeevi and M. Zibulski. Analysis of multiwindow Gabor-type schemes by framemethods. Applied and Computational Harmonic Analysis, 4:188โ221, 1997.
Yingxiong FuHubei Key Laboratory of Applied MathematicsHubei University
and
Faculty of Mathematics and Computer ScienceHubei University, Hubei, China
and
Department of MathematicsUniversity of Aveiro, Portugal
e-mail: [email protected]
Uwe Kahler and Paula CerejeirasDepartment of MathematicsUniversity of Aveiro, Portugal
e-mail: [email protected]@ua.pt
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 321โ332cโ 2013 Springer Basel
16 Sparse Representation of Signalsin Hardy Space
Shuang Li and Tao Qian
Abstract. Mathematically, signals can be seen as functions in certain spaces.And processing is more efficient in a sparse representation where few coeffi-cients reveal the information. Such representations are constructed by decom-posing signals into elementary waveforms. A set of all elementary waveformsis called a dictionary. In this chapter, we introduce a new kind of sparse rep-resentation of signals in Hardy space ๐ป2(๐ป) via the compressed sensing (CS)technique with the dictionary
D = {๐๐ : ๐๐(๐ง) =
โ1โ โฃ๐โฃ21โ ๐๐ง
, ๐ โ ๐ป}.where ๐ป denotes the unit disk. In addition, we give examples exhibiting thealgorithm.
Mathematics Subject Classification (2010). 30H05, 42A50, 42A38.
Keywords. Hardy space, compressed sensing, analytic signals, reproducingkernels, sparse representation, redundant dictionary, ๐1 minimization.
1. Introduction
A basis gives unique representations for signals in some certain space. However,it does not always give sparse expressions. One of the problems of approximationtheory is to approximate functions with elements from a large candidate set calleda dictionary. Let ๐ป be a Hilbert space. Using terminology introduced by Mal-lat and Zhang [18], a dictionary is defined as a family of parameterized vectors
D = {๐๐พ}๐พโฮ in ๐ป such that โฅ๐๐พโฅ = 1 and span(๐๐พ) = ๐ป . The ๐๐พ are usuallycalled atoms. For the discrete-time situation, the approximation problem can bewritten as
๐ = D๐ฅ (1.1)
where ๐ is the discrete signal, matrix D represents the dictionary with atoms ascolumns and ๐ฅ is the vector of coefficients. Notice that we adopt the vector inner
322 S. Li and T. Qian
product instead of the Hilbert inner product. In general, D has more columnsthan rows because of redundancy. A natural question is: can we find the best๐ -term approximation in a redundant dictionary for a given signal? That is anoptimization problem:
min โฅ๐ โD๐ฅโฅ๐2 subject to โฅ๐ฅโฅ๐0 โค๐ (1.2)
where โฅ๐ฅโฅ๐0 is the number of nonzero coefficients of ๐ฅ. Unfortunately, findingan optimal ๐ -term approximation in redundant dictionaries is computationallyintractable because it is NP-hard [8, 9, 17]. Thus, it is necessary to rely on goodbut not optimal approximations with computational algorithms. Until now, threemain strategies have been investigated, they are matching pursuit , basis pursuitand compressed sensing.
1.1. Matching Pursuit and Basis Pursuit
Matching pursuit (MP) introduced by Mallat and Zhang computes signal approx-imations from a redundant dictionary by iteratively selecting one vector at a time.It is an example of a greedy algorithm. For a detailed description, please referto [17, 18]. A recent development is the Adaptive Fourier Decomposition (AFD),which is a variation of the greedy algorithm in the particular context of Hardyspaces [20]. An intrinsic feature of the algorithm is that when stopped after afew steps, it yields an approximation using only a few atoms. If the dictionary isorthogonal, the method works perfectly. If the dictionary is not orthogonal, thesituation is less clear [6]. The MP algorithm often yields locally optimal solutionsdepending on initial values. In contrast, basis pursuit (BP) performs a more globalsearch. It finds signal representations by solving the following problem
min โฅ๐ฅโฅ๐1 subject to ๐ = D๐ฅ. (1.3)
Given ๐ and D , we find ๐ฅ with minimal ๐1 norm. Notice that (1.3) is a convexoptimization program which is not NP-hard. Actually, the use of ๐1 minimizationto promote sparsity has a long history, dating back at least to the work of Beurling[1] on Fourier transform extrapolation from partial observations. Basis pursuit isan optimization principle, not an algorithm. Empirical evidence suggests that BPis more powerful than MP [6]. And the stability of BP has been proved in thepresence of noise for sufficiently sparse representations [11]. BP is closely connectedwith convex programming. The interior-point method and the homotopy methodcan be applied to BP in nearly linear time [6, 13].
1.2. Compressed Sensing
If the original signal is sparse in some sense, compressed sensing (CS) gives anexcellent recovery of the signal. CS is a new concept in signal processing. The ideashave their origins in certain abstract results by Kashin [7, 15] but were broughtinto the forefront by the work of Candes, Romberg and Tao [2โ5] and Donoho [10].The core idea behind CS is that a signal or image, unknown but supposed to becompressible by a known transform, can be subjected to fewer measurements than
16. Sparse Representation of Signals in Hardy Space 323
the nominal number of pixels, and yet be accurately reconstructed [12]. Basically,CS relies on random projection and BP. Suppose we have
๐ฆ = ฮฆ๐ฅ, (1.4)
where ๐ฅ is a finite vector, ฮฆ is observation matrix and ๐ฆ is the vector of availablemeasurements. Then the BP solution ๐ฅโ of
min โฅ๐ฅโฅ๐1 subject to ๐ฆ = ฮฆ๐ฅ (1.5)
recovers ๐ฅ exactly provided that ๐ฅ is sufficiently sparse and the matrix obeys theRestricted Isometry Property (RIP) [3,4]. However the RIP of a fixed matrix is veryhard to check, thus in practice we use random matrices instead. A Gaussian matrixฮฆ โ ๐ ๐ร๐ whose entries ฮฆ๐,๐ are independent and follow a normal distributionwith expectation 0 and variance 1/๐ is often adopted because we have [21] that:
Theorem 1.1. Let ฮฆ โ โ๐ร๐ be a Gaussian random matrix. Let ๐, ๐ฟ โ (0, 1) andassume
๐ โฅ ๐ถ๐ฟโ2(๐ log(๐/๐ ) + log ๐โ1
)for a universal constant ๐ถ > 0. Then with probability at least 1 โ ๐ the restrictedisometry constant of ฮฆ satisfies ๐ฟ๐ โค ๐ฟ.
The theorem tells us that we can raise the probability by increasing thenumber of rows of the matrix ฮฆ. We will not discuss details of CS, for more aboutthis technique, please see [2โ5, 10, 12, 13, 21].
1.3. Hardy Space ๐ฏ2(๐ป)
Hardy space is an important class of spaces connected to analytic signals. Controltheory and rational approximation theory are also bound with the research in thisfield. In this chapter, we discuss signal decomposition in ๐ป2(๐ป). ๐ป2(๐ป) containsanalytic signals with finite energy. Denote ๐ป = {๐ง โ โ : โฃ๐งโฃ < 1}, and let Hol(๐ป)be the space of analytic functions on ๐ป. For ๐ > 0, Hardy space ๐ป๐(๐ป) is definedas follows:
๐ป๐(๐ป) =
โงโจโฉ๐ โ Hol(๐ป) : โฅ๐โฅ๐๐ป๐ = sup0โค๐<1
2๐โซ0
โฃ๐(๐๐๐๐ฅ)โฃ๐ d๐ฅ/2๐ <โโซโฌโญ , (1.6)
and
๐ปโ(๐ป) =
{๐ โ Hol(๐ป) : โฅ๐โฅ๐ปโ = sup
๐งโ๐ปโฃ๐(๐ง)โฃ <โ
}. (1.7)
Indeed, ๐ป2(๐ป) is a Hilbert space consisting of all functions ๐(๐ง) =โโ
๐=0 ๐๐๐ง๐
analytic in the unit disc ๐ป such that โฅ๐โฅ2 =โโ
๐=0 โฃ๐๐โฃ2 <โ. It has reproducingkernels ๐๐(๐ง) =
11โ๐๐ง, ๐ โ ๐ป. Besides the Fourier basis {1, ๐ง, ๐ง2, . . . , ๐ง๐, . . .}, ๐ป2(๐ป)
324 S. Li and T. Qian
has another orthonormal basis {๐๐}โ๐=1 named after Takenaka and Malmquist(TM):
๐1(๐ง) =
โ1โ โฃ๐ง1โฃ21โ ๐ง1๐ง
(1.8)
and
๐๐(๐ง) =
โ1โ โฃ๐ง๐โฃ21โ ๐ง๐๐ง
๐โ1โ๐=1
๐ง โ ๐ง๐1โ ๐ง๐๐ง
(1.9)
for ๐ โฅ 2, where {๐ง๐}โ๐=1 must satisfy
โโ๐=1
1โ โฃ๐ง๐โฃ =โ. (1.10)
It is clear that the TM basis can be obtained by applying the GramโSchmidtprocedure to the reproducing kernels ๐๐ง๐ = 1/ (1โ ๐ง๐๐ง) under condition (1.10).TM systems have long been associated with fruitful results in applied mathematicssuch as control theory, signal processing and system identification. Qian et al.recently proposed an adaptive Fourier decomposition algorithm (AFD) that resultsin a TM system (not necessarily a basis) with selected poles according to thegiven signal [20]. However AFD is valid only for one-dimensional signals whichnaturally raises the question: what happens in higher dimensions? Inspired by theaforementioned works, the adaptive decomposition of any function of two or threevariables is obtained in [19] by means of quaternionic analysis. Unfortunately, theresult can not be directly generalized into (๐+1)-dimensional space in the Cliffordsetting because a Clifford number is not invertible in general.
It is clear that the normalized reproducing kernel
D = {๐๐ : ๐๐(๐ง) =โ1โ โฃ๐โฃ21โ ๐๐ง
, ๐ โ ๐ป} (1.11)
forms a dictionary of ๐ป2(๐ป), and this redundant dictionary does give a sparse rep-resentation of signals in Hardy space. In [19] Qian et al. construct a dictionary fordecomposition of quaternionic-valued signals of finite energy. It is a generalizationof (1.11). This chapter will focus on the method based on CS to decompose signalsin ๐ป2(๐ป). Later we will generalize it into quaternionic Hardy space.
The chapter is concisely arranged as follows. We discuss the main results inSection 2. In Section 3, examples are given to illustrate our algorithm.
2. Main Results
The singular value decomposition (SVD) is an effective tool for the analysis oflinear operators. It is also useful in this chapter. We recall that:
16. Sparse Representation of Signals in Hardy Space 325
Theorem 2.1 (Singular Value Decomposition [16]). Let ๐ด denote an arbitrary ma-trix with elements in โ๐ร๐ and let {๐ ๐}๐๐=1 be the nonzero singular values of ๐ด.Then ๐ด can be represented in the form
๐ด = ๐๐ท๐ โ, (2.1)
where ๐ โ โ๐ร๐, ๐ โ โ๐ร๐ are unitary and the ๐ร ๐ matrix ๐ท has elements ๐ ๐in the ๐, ๐ position (1 โค ๐ โค ๐) and zeros elsewhere. The ๐ ๐ are singular values of
๐ด, ๐ ๐ =โ
๐๐(๐ดโ๐ด). In fact, the diagonal matrix ๐ท can be written as(ฮฃ๐ 00 0
)(2.2)
where ฮฃ๐ = diag(๐ 1, ๐ 2, . . . , ๐ ๐) with ๐ 1 โฅ ๐ 2 โฅ โ โ โ , ๐ ๐.In the model (1.1), (1.11), we select ๐ points {๐๐}๐๐=1 in ๐ป, and sample ๐
points for each ๐๐๐equally spaced in [0, 2๐]. So we get a matrix D โ โ๐ร๐ . For
a signal in ๐ป2(๐ป), we also sample ๐ points on its boundary value equally spacedto form a vector ๐ โ โ๐ . Thus, the representation problem is
๐ = D๐ฅ. (2.3)
Suppose that D = ๐๐ท๐ โ, then ๐โ๐ = ๐ท๐ โ๐ฅ. Note that ๐ท is a diagonal matrix
and โฅ๐โ๐ โฅ2 = โฅ๐ โฅ2 , โฅ๐ โ๐ฅโฅ2 = โฅ๐ฅโฅ2. If the singular values of D decay fast, we canexpect a relatively sparse representation ๐ฅ with a small energy error. The mostimportant thing is that the assertion should be established in the sense of ๐ โโbecause one must explain what the situation would be when the number of atomsis large.
First, we give two lemmas, they are also the simple cases of our theorem.
Lemma 2.2. Suppose ๐ points {๐๐}๐โ1๐=0 are selected equally spaced on the circle
of radius ๐. Let ๐1 โฅ ๐2 โฅ โ โ โ โฅ ๐๐ be eigenvalues of the matrix DโD . Then wehave
lim๐โโ
1
๐
๐โ๐=1
๐๐ โฅ 1โ ๐2๐. (2.4)
Proof. (Sketch of proof.) Let
D =(๐๐0 ๐๐1 . . . ๐๐๐โ1
)(2.5)
then the Hermitian matrix ๐ป can be written as
๐ป = DโD =
โโโโโ๐๐0
๐๐1
...๐๐๐โ1
โโโโโ (๐๐0 ๐๐1 . . . ๐๐๐โ1
). (2.6)
with elements ๐ป๐๐ =โจ๐๐๐ , ๐๐๐
โฉ=โจ๐๐0 , ๐๐๐โ๐
โฉ= ๐๐โ๐. Notice that
๐๐ = โจ๐๐0 , ๐๐๐โฉ = 1โ ๐2
1โ ๐2๐๐๐๐=
1โ ๐2
1โ ๐2๐๐๐๐, (๐ = 0, 1, . . . , ๐ โ 1) (2.7)
326 S. Li and T. Qian
0 10 20 30 40 500
2
4
6
8
10
12
14
16
18
20
0 10 20 30 40 500
1
2
3
4
5
6
7
8
9
10
Figure 1. Selecting ๐ points on the circle of radius ๐ (2.9) gives that๐๐ โ (1 โ ๐2)๐2(๐โ1)๐ when ๐ is large. The small red circles are thecorresponding estimation whereas the blue points represent the realeigenvalues given by numerical calculation. They fit amazingly well eventhough ๐ is not very large.
where ๐๐ = ๐๐ is the argument of ๐๐. Recall Ky Fanโs maximum principle: let ๐ดbe any Hermitian operator, then for ๐ = 1, 2, . . . , ๐, we have
๐โ๐=1
๐๐(๐ด) = max
๐โ๐=1
โจ๐ด๐ฅ๐ , ๐ฅ๐โฉ (2.8)
where the eigenvalues ๐1(๐ด) โฅ ๐2(๐ด) โฅ โ โ โ โฅ ๐๐(๐ด), and the maximum are takenover all orthonormal ๐-tuples {๐ฅ1, . . . , ๐ฅ๐}.
Sampling ๐ points equally spaced on the orthonormal functions{๐โ๐๐๐ก}๐โ1
๐=0 , we prove that
lim๐โโ
๐๐+1
๐= (1โ ๐2)๐2๐ (2.9)
where ๐1 โฅ ๐2 โฅ โ โ โ โฅ ๐๐. Hence
lim๐โโ
1
๐
๐โ๐=1
๐๐ โฅ๐โ๐=1
(1 โ ๐2)๐2(๐โ1) = 1โ ๐2๐. (2.10)
โก
Remark 2.3. Note that trace(๐ป) = ๐ , thus (2.9) shows that the eigenvalues decayas a geometric series with common ratio ๐2 as ๐ โโ. See Figure 1.
Lemma 2.4. Suppose ๐ points {๐๐}๐๐=1 are selected equally spaced on the segment
[0, 1), and ๐1 is the largest eigenvalue of the matrix DโD . Then we have
lim๐โโ
๐1
๐โฅ
1โซ0
1โซ0
โ1โ ๐ 2
โ1โ ๐2
1โ ๐ ๐d๐ d๐ โ 0.815784. (2.11)
16. Sparse Representation of Signals in Hardy Space 327
0 10 20 30 40 50
40
45
35
30
25
20
15
10
5
0
โ50 10 20 30 40 50 60
โ10
0
10
20
30
40
50
Figure 2. Eigenvalue distribution
Sketch of Proof. The proof is similar to the previous lemma. In this situation, theentries of the Hermitian matrix ๐ป in (2.6) are
๐ป๐๐ =
โ1โ ๐2๐
โ1โ ๐2๐
1โ ๐๐๐๐, (2.12)
where ๐๐ is the ๐th point on [0, 1). Let ๐ฅ =(
1โ๐, 1โ
๐, . . . , 1โ
๐
)โ โ๐ with โฅ๐ฅโฅ2 =
1. We use Ky Fanโs principle to estimate the maximum eigenvalue. We find thatโจ๐ป๐ฅ, ๐ฅโฉ is a double Riemann sum of
1โซ0
1โซ0
โ1โ ๐ 2
โ1โ ๐2
1โ ๐ ๐d๐ d๐ . (2.13)
Hence,๐1(๐ป)
๐โฅ โจ๐ป๐ฅ, ๐ฅโฉ
๐โ 0.815784. (2.14)
โกRemark 2.5. Select ๐ points on the interval [0, 1). The maximum eigenvalues are41.1816 and 49.5133 respectively. ๐1(๐ป)/๐ satisfy (2.11). See Figure 2.
Generally, we select {๐๐} in the whole disc as follows. Divide [0, 2๐] and [0, 1)into ๐1 and ๐2 parts respectively. Hence, the number of ๐๐s is ๐ช(๐1๐2). Weobtain the main theorem as follows.
Theorem 2.6. Let ๐1 โฅ ๐2 โฅ ๐3 โฅ โ โ โ be eigenvalues of ๐ป = DโD . Then we have
lim๐1โโ๐2โโ
๐โ๐=1
๐๐
๐1๐2โฅ 1โ 1
2๐+ 1. (2.15)
328 S. Li and T. Qian
Sketch of Proof. ๐ป is actually a block matrix in this situation
๐ป =
โโโโโโโ๐ตโ1๐ต1 ๐ตโ1๐ต2 ๐ตโ1๐ต3 . . . ๐ตโ1๐ต๐2
๐ตโ2๐ต1 ๐ตโ2๐ต2 ๐ตโ2๐ต3 . . . ๐ตโ2๐ต๐2
๐ตโ3๐ต1 ๐ตโ3๐ต2 ๐ตโ3๐ต3 . . . ๐ตโ3๐ต๐2
......
.... . .
...๐ตโ๐2
๐ต1 ๐ตโ๐2๐ต2 ๐ตโ๐2
๐ต3 . . . ๐ตโ๐2๐ต๐2
โโโโโโโ (2.16)
with each block ๐ตโ๐ ๐ต๐ โ โ๐1ร๐1 . Let
๐๐(๐) = ๐๐โ1โ ๐2 (2.17)
and
๐๐(๐) =๐๐(๐)
โฅ๐๐(๐)โฅ๐ฟ2(0,1)
(2.18)
where ๐ โ (0, 1), ๐ โฅ 0. Denote ๐๐(๐) = ๐โ๐๐๐, ๐ โฅ 0.
Sample ๐2 points on ๐๐(๐) and get the vector ๐บ๐ โ โ๐2 . Sample ๐1 pointson ๐๐ to get ๐ธ๐ โ โ๐1 . Consider
โจ๐ป(๐บ๐
โ๐ธ๐), ๐บ๐
โ๐ธ๐โฉ
๐1๐2(2.19)
where the tensor product ๐บ๐
โ๐ธ๐ is a vector in โ๐1๐2 . We prove that
lim๐1โโ๐2โโ
โจ๐ป(๐บ๐
โ๐ธ๐), ๐บ๐
โ๐ธ๐โฉ
๐1๐2=
1
2๐+ 1โ 1
2๐+ 3, ๐ โฅ 0. (2.20)
Hence, Ky Fanโs maximum principle gives
lim๐1โโ๐2โโ
๐โ๐=1
๐๐
๐1๐2โฅ
๐โ๐=1
(1
2๐ โ 1โ 1
2๐ + 1
)= 1โ 1
2๐+ 1. (2.21)
โก
Remark 2.7. Note that trace(๐ป) = ๐1๐2, the theorem shows that the eigenvaluesdecay rapidly. Figure 3 shows the numerical calculation when ๐1 = ๐2 = 40. Wehave that ๐1/(๐1๐2) = 0.6544178 โ 2/(1ร 3), ๐2/(๐1๐2) = 0.1312327 โ 2/(3ร5), ๐3/(๐1๐2) = 0.057843 โ 2/(5ร7), ๐4/(๐1๐2) = 0.03319499โ 2/(7ร9), andso on. The several largest eigenvalues contribute a large part of the sum of all eigen-values. Since singular value ๐ ๐ =
โ๐๐(๐ป), we assert that ๐ ๐ tends to zero rapidly.
3. Numerical Examples
We give two numerical examples. In our examples, D โ โ900ร3000. This means thedictionary has 3000 atoms and each atom vector is of size 900. We embed D into
16. Sparse Representation of Signals in Hardy Space 329
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
1 2 3 4 5 6 7 8 9 10
Figure 3. Decay of eigenvalues (see Remark 2.7).
โ1800ร6000 for programming convenience by
D =
(โ (D) โโ (D)โ (D) โ (D)
). (3.1)
Therefore,
๐ = D๐ฅโโ ๐ =
(โ (๐ )โ (๐ )
)= D
(โ (๐ฅ)โ (๐ฅ)
)= D ๐ฅ. (3.2)
where โ (๐ ) and โ (๐ ) are the real and imaginary parts respectively.
Consider the minimization problem
min โฅ๐ฅโฅ1 subject to ๐ฆ = ฮฆD ๐ฅ (3.3)
where ฮฆ โ โ๐ร๐ is a Gaussian random matrix satisfying ฮฆ๐๐ โผ ๐ฉ (0, 1/๐). We
recover the signal by D๐ฅโ when the solution ๐ฅโ of (3.3) is obtained by convexprogramming. As we mentioned above, ๐ belongs to the vector space of 900ร 2 =1800 dimensions with a 900-dimensional real part and 900-dimensional imaginary
part. D has 3000ร 2 = 6000 columns and ๐ฅ is a 6000-element coefficient vector.
3.1. Example 1
๐ (๐ง) =0.247๐ง4 + 0.0355๐ง3
(1โ 0.9048๐ง)(1โ 0.3679๐ง)(3.4)
Set the Gaussian randommatrix ฮฆ with 160 rows and its entries ฮฆ๐๐ โผ ๐ฉ (0, 1/160).Runtime is 16.437 seconds and relative error is 3.7188ร10โ4. See Figures 4 and 5.
3.2. Example 2
๐ (๐ง) = ๐ง20 + ๐ง10 + ๐ง5. (3.5)
Choose the Gaussian randommatrix ฮฆ with 160 rows and entries ฮฆ๐,๐โผ๐ฉ (0,1/160).Runtime is 34.891 seconds and relative error is 1.7ร 10โ3. See Figures 6 and 7.
330 S. Li and T. Qian
0 200 400 600 800 1000 1200 1400 1600 1800 0 200 400 600 800 1000 1200 1400 1600 1800โ0.8
โ0.6
โ0.4
โ0.2
0
0.2
0.4
0.6
0.8
1
1.2
โ0.8
โ0.6
โ0.4
โ0.2
0
0.2
0.4
0.6
0.8
1
1.2
recovered signaloriginal signal
Figure 4. Example 1. The recovered signal almost coincides with theoriginal signal. See Section 3.1.
0 200 400 600 800 1000 1200 1400 1600 1800โ2.5
โ2
โ1.5
โ1
โ0.5
0
0.5
1
1.5
2
2.5
0 1000 2000 3000 4000 5000 6000โ0.05
0
0.05
0.1
0.2
0.3
0.15
0.25
Figure 5. Example 1. See Section 3.1.
References
[1] A. Beurling. Sur les integrales de Fourier absolument convergentes et leur applica-tion a une transformation functionelle. In Proceedings Scandinavian MathematicalCongress, Helsinki, Finland, 1938.
[2] E. Candes, J. Romberg, and T. Tao. Robust uncertainty principles: Exact signalreconstruction from highly incomplete frequency information. IEEE Transactionson Information Theory, 52(2):489โ509, 2006.
[3] E. Candes, J. Romberg, and T. Tao. Stable signal recovery from incomplete and in-accurate measurements. Communications on Pure Applied Mathematics, 59(8):1207โ1223, 2006.
[4] E. Candes and T. Tao. Decoding by linear programming. IEEE Transactions onInformation Theory, 51(12):4203โ4215, 2005.
16. Sparse Representation of Signals in Hardy Space 331
0 200 400 600 800 1000 1200 1400 1600 1800 0 200 400 600 800 1000 1200 1400 1600 1800โ3
โ2
โ1
0
1
2
3recovered signaloriginal signal
โ3
โ2
โ1
0
1
2
3
Figure 6. Example 2. The recovered signal almost coincides with theoriginal signal. See Section 3.2.
0 200 400 600 800 1000 1200 1400 1600 1800 0 1000 2000 3000 4000 5000 6000โ40
โ30
โ20
โ10
0
10
20
30
โ0.4
โ0.3
โ0.2
โ0.1
0
0.1
0.2
0.3
0.4
0.5
Figure 7. Example 2. See Section 3.2.
[5] E. Candes and T. Tao. Near optimal signal recovery from random projections: Uni-versal encoding stategies? IEEE Transactions on Information Theory, 52(12):5406โ5425, 2006.
[6] S. Chen, D. Donoho, and M. Saunders. Atomic decomposition by basis pursuit. SIAMJournal on Scientific Computing, 20(1):33โ61, 1999.
[7] A. Cohen, W. Dahmen, and R. DeVore. Compressed sensing and best ๐-term ap-proximation. Journal of the American Mathematical Society, 22:211โ231, 2009.
[8] G. Davis. Adaptive Nonlinear Approximations. PhD thesis, New York University,Courant Institute, 1994.
[9] G. Davis and S. Mallat. Adaptive greedy approximations. Constructive Approxima-tion, 13(1):57โ98, 1997.
332 S. Li and T. Qian
[10] D. Donoho. Compressed sensing. IEEE Transactions on Information Theory, 52(4):1289โ1306, 2006.
[11] D. Donoho and M. Elad. On the stability of the basis pursuit in the presence ofnoise. Signal Processing, 86(3):511โ532, 2006.
[12] D. Donoho and Y. Tsaig. Extensions of compressed sensing. Signal Processing,86(3):533โ548, 2006.
[13] M. Fornasier. Numerical methods for sparse recovery. In Theoretical Foundationsand Numerical Methods for Sparse Recovery [14], pages 93โ200.
[14] M. Fornasier, editor. Theoretical Foundations and Numerical Methods for SparseRecovery, volume 9 of Radon Series on Computational and Applied Mathematics.De Gruyter, Germany, 2010.
[15] B. Kashin. The widths of certain finite dimensional sets and classes of smooth func-tions. Izvestia, 41:334โ351, 1977.
[16] P. Lancaster and M. Tismenetsky. The Theory of Matrices with Applications. Aca-demic Press, second edition, 1985.
[17] S. Mallat. A Wavelet Tour of Signal Processing. Academic Press, third edition, 2008.First edition published 1998.
[18] S. Mallat and Z. Zhang. Matching pursuit with time-frequency dictionaries. IEEETransactions on Signal Processing, 41(12):3397โ3415, Dec. 1993.
[19] T. Qian, W. Sproรig, and J. Wang. Adaptive Fourier decomposition of functions inquaternionic Hardy spaces. Mathematical Methods in the Applied Sciences, 35(1):43โ64, 2012.
[20] T. Qian and Y.-B. Wang. Adaptive Fourier series โ a variation of greedy algorithm.Advances in Computational Mathematics, 34(3):279โ293, 2011.
[21] H. Rauhut. Compressive sensing and structured random matrices. In Fornasier [14],pages 1โ92.
Shuang Li and Tao QianDepartment of MathematicsUniversity of Macau, Macaue-mail: [email protected]
Quaternion and CliffordโFourier Transforms and Wavelets
Trends in Mathematics, 333โ338cโ 2013 Springer Basel
Index
analytic signal, 42, 67, 222, 223, 247โ250
โ Clifford, 208
โ biquaternion, 197
โ complex, 197
โ examples, 209
โ hypercomplex, 48
โ local amplitude, 223
โ local phase, 223
โ ๐-dimensionnal, 198
โ one-dimensional, 200
โ phases, 210
โ properties, 201
โ quaternion, 197
โ quaternion 2D, 201
โ quaternionic, 68
โ two-dimensional, 200, 226
โ local amplitude, 227
โ local orientation, 227
โ local phase, 227
โ video, 198, 212
angular velocity, 50
anticommutative part, 162
atomic function, 58, 64
โ 2D, 60
โ quaternionic, 65
โ up(๐ฅ), 58
BalianโLow theorem, 303, 316, 317
Banach module, 287
basis pursuit, 322
bivector, geometric interpretation, 124
Bochner theorem, 85
BochnerโMinlos theorem, 85, 113, 117
CauchyโRiemann equations, 63
CayleyโDickson form, 43
โ polar, 44
checkerboard, 233
chrominance, 22
โ plane, 22
clifbquat, see Clifford biquaternion
Clifford algebra, 198, 286
โ ๐ถโ0,3, basis, 133
โ ๐ถโ1,2, 134
โ ๐ถโ3,0, 134
โ automorphism group, 126
โ basis element matrices, 145
โ center, 126
โ central pseudoscalar, 134
โ characteristic polynomial, 136
โ complex, 192
โ complex conjugation, 135
โ conformal geometric algebra, 135
โ connected components
โ Klein group, 131
โ square isometries, 132
โ even subalgebra, 125
โ Fourier transform
โ steerable, 141
โ geometric algebras, 123
โ grade involution, 137
โ group of invertible elements, 126
โ idempotents, 133, 134
โ isomorphism
โ CLIFFORD package, 146
โ Lorentz space, 135
โ matrix idempotents, 147
โ matrix ring isomorphisms, 124, 125
โ dimension index ๐, 125
โ signature, 125
334 Index
โ minimal polynomial, 134
โ Pauli matrix algebra, 134
โ pseudoscalar, 127
โ relation to quaternions, 124
โ reversion, 137
โ scalar part, 127
โ signature, 125
โ software package, 127
โ spinor representation, 146
โ swap automorphism, 131
โ symbolic computer algebra, 127
โ trace, 127
โ wavelet transform, steerable, 141
Clifford analysis, applications, 124
Clifford biquaternion, 206
Clifford conjugate, 286
Clifford functions, 289
Clifford module, 289
Clifford polynomials, 146
Clifford, William Kingdon, 123, 198
CliffordโFourier transform, see Fouriertransform
colour
โ denoising, 257, 259, 261โ264
โ image
โ Fourier transform, 158
โ grey line, 22
โ processing, 15, 22
commutative part, 162
complex
โ Clifford algebra, 192
โ degenerate, 10
โ envelope, 50
โ representation, 181, 192
compressed sensing, 322
convolution, 12
correlation, 11
cylindrical Fourier transform, 159
decomposition
โ luminance and chrominance, 22
โ singular value, 324
demodulation, 238
Dirac equation, 180, 183
Dirac operator, 194, 228
dup(๐ฅ), 61
dyadic shifts, 63
edge detection, 183
eigen-angle, 7
eigen-axis, 7
envelope, 42
โ complex, 50
Euler formula for quaternions, 237
even-odd form, 6
exponential function, 157
Faddeevโs Green function, 281
โ Dirac operators, 281
filter design
โ steerable, 39
filter, steerable quaternionic, 67
filtering, 188
force-free fields, 274
Fourier transform, 222
โ Cliffordโ, 157, 185, 197, 200, 207,288, 299, 302
โ examples, 213
โ properties (scalar function), 209
โ windowed, 294, 300, 303, 317
โ cylindrical, 159
โ geometric, see geometric Fouriertransform
โ properties, 201
โ quaternion, 43, 44
โ properties, 203
โ quaternionic, see quaternionicFourier transform
โ Sommenโ, 157
FourierโStieltjes transform, 85
โ quaternionic, 88
frame
โ bound, 303
โ Clifford, 303, 307, 317
โ dual, 304, 314
โ tight, 303
frequency modulation, 11
Fubiniโs theorem, 296
Gabor filter
โ complex, 294
โ quaternionic, 294
Index 335
Gabor system, 300
โ Cliffordโ, 303, 305, 310
Gabor, D., 200
Gauss spinor formula, 194
generalized Weierstrass parametrization,183
geometric Fourier transform
โ definition, 157
โ existence, 159
โ linearity, 160
โ product theorem, 166
โ scaling theorem, 162
โ shift theorem, 172
geometric product, 156
GFT, 157
GramโSchmidt procedure, 324
Grassmann, Hermann Gunther, 198
group velocity, 212
Hamilton, William Rowan, 43, 198
Hardy space, 323
Hausdorff space, 110
Helmholtz equation, 275
โ factorization, 275
Hilbert space, 321, 323
Hilbert transform, 42, 47, 63, 67, 70, 222
โ partial, 66, 226
โ quaternion Fourier transform of, 47
โ total, 66, 226
image processing
โ grey line, 22
impulse response, 10
instantaneous amplitude, 42, 201
โ geometric, 49
instantaneous phase, 42, 201
โ geometric, 49
interferometry, 238
kernel
โ Gauss, 64
โ Poisson, 64
linearity, 10
LippmannโSchwinger integral equation,277
local phase, 67
luminance, 22
matching pursuit, 322
Maxwellโs equations, 273
โ in inhomogeneous media, 274
mean curvature, 183
monogenic
โ coefficients, modeling of, 257
โ colour signal, 248, 252, 253
โ colour wavelet, 252
โ colour wavelet transform, 248, 253,256
โ signal, 63, 228, 248โ250
โ local amplitude, 229
โ local orientation, 229
โ local phase, 229
โ wavelet transform, 248, 250, 251
multivector, 286
โ scalar product, 287
operator
โ coefficient, 304
โ frame, 304
โ reconstruction, 304
optical coherence tomography (OCT),271
optimization, 322
orthogonal 2D planes split, 15
โ determination from given planes, 26,27
โ exponential factor
โ identities, 21, 22
โ general, 18
โ geometric interpretation, 26
โ rotation, 26
โ orthogonality of OPS planes, 19
โ single pure unit quaternion, 21
โ subspace bases, 22, 25
parallel spinor field, 180
paravector, 296
Parseval relation, 200
Parseval theorem
โ Cliffordโ, 289
Parsevalโs equality, 303
period form, 181, 183
336 Index
phase, 248, 250, 251, 267
โ information, 67
โ quaternionic, 69, 294
โ velocity, 212
Plancherelโs theorem, 204, 288
PlemeljโSokhotzki formula, 225
probability measure, 108, 113
pseudoscalar, 286
pursuit
โ basis, 322
โ matching, 322
quaternion
โ algebra, 16, 65
โ definition, 5
โ over โ, 132
โ algebraic identites, 20
โ chrominance, 22
โ plane, 22
โ colour image
โ processing, 15, 22
โ complex sub-field, 6
โ conjugate, 5
โ decomposition, luminance andchrominance, 22
โ Euler formula, 237
โ exponential, 237
โ exponential factors
โ identity, 25, 36
โ grey line, 22
โ half-turn, 22, 25
โ Coxeter, 20
โ rotation, 18, 20
โ involution, 18
โ quaternion conjugation, 32
โ line reflection
โ pointwise invariant, 33
โ real line, 32
โ logarithm, 237
โ luminance, 22
โ modulus identity, 17
โ norm, 5
โ orientation, 236
โ orthogonal 2D planes split
โ general, 25
โ real coefficients, 20
โ orthogonal basis, 18, 19
โ orthogonality, 17
โ orthonormal basis, 22, 25
โ perplex part, 6
โ plane subspace bases, 18
โ polar CayleyโDickson form, 44
โ polar form, 43
โ properties, 43
โ pure, 5
โ quaternion maps, 38
โ reflection
โ hyperplane, 32
โ invariant hyperplane, 32
โ rotary axis, 33
โ rotary invariant line, 33
โ relation to Clifford algebra, 124
โ rotary reflection, 33
โ rotation angle, 35
โ rotation plane basis, 34
โ rotation, 33
โ double, 33
โ four-dimensional, 33
โ reflection, 33
โ rotary, 33
โ scalar part
โ symmetries, 17
โ scalar-part, 5
โ simplex and perplex parts, 24
โ simplex and perplex split, 21
โ simplex part, 6
โ split
โ orthogonality, 18
โ steerable, 18
โ subspace bases, 19
โ vector-part, 5
quaternionic
โ analytic function, 66
โ analytic signal, 68
โ filter, steerable, 67
โ phases, 69, 294
โ structure, 182
โ wavelet multiresolution, 66
quaternionic Fourier transform, 105,158, 202, 227, 299
โ analysis planes, 31
Index 337
โ asymptotic behaviour, 111
โ discrete, 30
โ dual-axis form, 8
โ factored form, 8
โ fast, 30
โ filter design
โ steerable, 39
โ forward transform, 8
โ generalization, 28
โ geometric interpretation, 31
โ geometric understanding
โ local, 38
โ inverse transform, 9
โ local phase rotations, 31
โ new forms, 28
โ new types, 39
โ of the Hilbert transform, 47
โ operator pairs, 8
โ phase angle, 31
โ phase angle transformation, 31
โ discrete, 32
โ fast, 32
โ split parts, 32
โ steerable, 32
โ phase rotation planes, 31
โ quasi-complex, 30
โ quaternion conjugation, 32, 36
โ local geometric interpretation, 37
โ local invariant line, 37
โ local phase rotation, 38
โ local rotation angle, 37
โ local rotation axis, 37
โ phase angle transformation, 38
โ phase angle transformation,discrete, 38
โ phase angle transformation, fast,38
โ phase angle transformation,interpretation, 38
โ phase angle transformation,quasi-complex, 38
โ phase angle transformation, splitparts, 38
โ phase angle transformation,steerable, 38
โ quasi-complex, 36
โ split parts, 36
โ reverse transform, 9
โ sandwich form, 8
โ single-axis definition, 9
โ split parts transformation, 30
โ split theorems, 39
โ steerable, 15
โ two pure unit quaternions, 28
โ windowed, 286
radiation condition, 276
RiemannโHilbert problem, 225
Riesz transform, 64, 229, 249, 251, 253,255, 257, 259, 262
โ wavelet, 256
scattering problem, 276
Schwartz space, 114
separability, 165
Siemens star, 231
singular value decomposition, 324
Sobel operator, 242
SommenโFourier transform, 157
spacetime Fourier transform, 158
spin
โ character, 185
โ group, 235
โ structure, 193
spinor, 235
โ bundle, 187, 193
โ connection, 194
โ field, 179
โ field, parallel, 180
โ formula, Gauss, 194
โ representation, 146, 183
โ tensor, 185
split
โ form, 7
โ of identity, 222, 223
โ w.r.t. commutativity, 162
square roots of โ1, 157
โ ๐ถโ0,2, 132
โ ๐ถโ0,3, 133
โ ๐ถโ0,5, 135, 137
โ ๐ถโ1,2, 134
โ ๐ถโ2,0, 128
338 Index
โ ๐ถโ2,1, 130
โ ๐ถโ2,3, 135
โ ๐ถโ3,0, 134
โ ๐ถโ4,1, 135
โ ๐ถโ7,0, 138
โ โ, 125
โ โ, โ2, 125
โ โ2, 133
โ โณ(2,โ), 135
โ โณ(2๐,โ), 134
โ โณ(2๐,โ), 128
โ โณ(2๐,โ2), 130
โ โณ(๐,โ2), 133
โ โณ(๐,โ), 132
โ ๐ โค 4, 125
โ algebraic submanifold, 126
โ inner automorphism, 126
โ bijection with idempotents, 134
โ biquaternions, ๐ถโ3,0, 125
โ central pseudoscalar, 134
โ centralizer, 126
โ centralizer computation, 149
โ compact manifold, 132
โ computation
โ CLIFFORD package, 141
โ conjugacy class, 126
โ dimension, 126
โ conjugate square root of โ1, 132
โ connected component, 126
โ dimension, 126
โ exceptional, 124, 127, 135
โ Fourier transformations, 124
โ idempotents, 134
โ Klein group, 131
โ Maple worksheets, 141, 144
โ matrix square root, 147
โ multivector split, 128
โ ordinary, 127, 135
โ Pauli matrix algebra, 134
โ quaternions, 132
โ scalar part zero, 127
โ skew-centralizer, 128
โ square isometries, 132
โ stability subgroup, 126
โ table โณ(2๐,โ), ๐ = 1, 2, 4, 141
โ visualization, 124
swap rule, 6
symmetry, 70
symplectic form, 6
texture, 238
โ detection, 185
tomography, optical coherence, 271
trivector, geometric interpretation, 124
uncertainty principle, 300, 303, 317
unique continuation principle, 279
up(๐ฅ), 58
vector space, non-Euclidean, 124
Ville, J., 200
wavelet
โ quaternionic multiresolution, 66
โ transform, analytic, 248
Weierstrass
โ generalized parametrization, 183
โ representation, 177
Zak transform, 305
โ Cliffordโ, 305, 307, 308, 312, 317