advances in structured operator theory and related areas: the leonid lerer anniversary volume
TRANSCRIPT
Operator Theory
Advances and Applications
237
Marinus A. Kaashoek
Leiba Rodman
Hugo J. Woerdeman
Editors
Advances in Structured Operator Theory and Related AreasThe Leonid Lerer Anniversary Volume
Joseph A. Ball (Blacksburg, VA, USA) Harry Dym (Rehovot, Israel) Marinus A. Kaashoek (Amsterdam, The Netherlands) Heinz Langer (Vienna, Austria) Christiane Tretter (Bern, Switzerland)
Vadim Adamyan (Odessa, Ukraine)
Albrecht Böttcher (Chemnitz, Germany) B. Malcolm Brown (Cardiff, UK) Raul Curto (Iowa, IA, USA) Fritz Gesztesy (Columbia, MO, USA) Pavel Kurasov (Stockholm, Sweden) Leonid E. Lerer (Haifa, Israel) Vern Paulsen (Houston, TX, USA) Mihai Putinar (Santa Barbara, CA, USA) Leiba Rodman (Williamsburg, VA, USA) Ilya M. Spitkovsky (Williamsburg, VA, USA)
Lewis A. Coburn (Buffalo, NY, USA) Ciprian Foias (College Station, TX, USA) J.William Helton (San Diego, CA, USA) Thomas Kailath (Stanford, CA, USA) Peter Lancaster (Calgary, Canada) Peter D. Lax (New York, NY, USA) Donald Sarason (Berkeley, CA, USA) Bernd Silbermann (Chemnitz, Germany) Harold Widom (Santa Cruz, CA, USA)
Associate Editors: Honorary and Advisory Editorial Board:
Operator Theory: Advances and Applications
Founded in 1979 by Israel Gohberg
Editors:
Volume 237
Wolfgang Arendt (Ulm, Germany)
Subseries Linear Operators and Linear Systems Subseries editors: Daniel Alpay (Beer Sheva, Israel) Birgit Jacob (Wuppertal, Germany) André C.M. Ran (Amsterdam, The Netherlands) Subseries Advances in Partial Differential Equations Subseries editors: Bert-Wolfgang Schulze (Potsdam, Germany) Michael Demuth (Clausthal, Germany) Jerome A. Goldstein (Memphis, TN, USA) Nobuyuki Tose (Yokohama, Japan) Ingo Witt (Göttingen, Germany)
Marinus A. Kaashoek Leiba RodmanHugo J. WoerdemanEditors
Advances in
The Leonid Lerer Anniversary Volume
Structured Operator Theoryand Related Areas
© 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 Publisherâs 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
ISBN 978-3-0348-0638-1 ISBN 978-3-0348-0639-8 (eBook)DOI 10.1007/978-3-0348-0639-8Springer Basel Heidelberg New York Dordrecht London
Library of Congress Control Number:
Springer is part of Springer Science+Business Media (www.birkhauser-science.com) Basel
Mathematics Subject Classification 47A68, 47B35, 39B42, 93B28, 15A15; 11C20, 47A13, 47A48 (2010):
2013948131
athematics
USA
EditorsMarinus A. Kaashoek
Netherlands
Leiba RodmanDepartment of MCollege of William and Mary
Hugo J. Woerdeman
Drexel UniversityPUSA
Amsterdam Williamsburg
hiladelphia
Department of Mathematics, FEWVU University
Department of Mathematics
ISSN - ISSN 2296-4878 (electronic) 0255 0156
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Portrait of Leonid Lerer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Leonid Lererâs Curriculum Vitae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Leonid Lererâs List of Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
M.A. KaashoekLeonia Lererâs Mathematical Work and Amsterdam Visits . . . . . . . . . . . 1
H. BartLeonia Lerer and My First Visit to Israel . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
I. KarelinThrough the Eyes of a Student . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
A.C.M. RanReminiscences on Visits to Haifa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
H.J. WoerdemanMy First Research Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
J.A. Ball and V. BolotnikovInterpolation in Sub-Bergman Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
H. Bart, T. Ehrhardt and B. SilbermannZero Sums of Idempotents and Banach Algebras Failingto be Spectrally Regular . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
T. Bella, Y. Eidelman, I. Gohberg, V. Olshevsky and E. TyrtyshnikovFast Inversion of Polynomial-Vandermonde Matrices forPolynomial Systems Related to Order OneQuasiseparable Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
xiii
vi Contents
H. Dym and M. PoratLong Proofs of Two CarlsonâSchneider Type Inertia Theorems . . . . . . 107
T. Ehrhardt and I.M. SpitkovskyOn the Kernel and Cokernel of Some Toeplitz Operators . . . . . . . . . . . . 127
A.E. Frazho, M.A. Kaashoek and A.C.M. RanRational Matrix Solutions of a Bezout Type Equationon the Half-plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
M.A. Kaashoek and F. van SchagenInverting Structured Operators Related to Toeplitz PlusHankel Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
P. Lancaster and I. ZaballaOn the Sign Characteristics of Selfadjoint Matrix Polynomials . . . . . . . 189
Yu.I. LyubichQuadratic Operators in Banach Spaces and NonassociativeBanach Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
L. RodmanStrong Stability of Invariant Subspaces of Quaternion Matrices . . . . . . 221
H.J. WoerdemanDeterminantal Representations of Stable Polynomials . . . . . . . . . . . . . . . 241
Preface
This volume is dedicated to Leonid Arie Lerer on the occasion of his seventiethbirthday (April 19, 2013). Leonia, as he is known to his friends, is an expert in thetheory of structured matrices and operators and related matrix-valued functions.He has been a great inspiration to many.
Leonid Lerer started his mathematical career in Kishinev, Ukraine, with AlekMarkus and Israel Gohberg as research supervisors. He defended his Ph.D. thesisin 1969 in Kharkov, Ukraine. In December 1973 he immigrated to Israel. Since1981 he has been a professor at the Technion in Haifa where at present he has thestatus of emeritus. He has educated six Ph.D. students and five masters students.His more than 80 papers cover a wide spectrum of topics, ranging from functionalanalysis and operator theory, linear and multilinear algebra, ordinary differentialequations, to systems and control theory.
This anniversary volume begins with a picture of Leonid Lerer, his Curricu-lum Vitae and List of Publications, and personal notes written by former students,mathematical friends and colleagues.
The main part of the book consists of a selection of peer-reviewed researcharticles presenting recent results in areas that are close to Lererâs mathematicalinterests. This includes articles on Toeplitz, WienerâHopf, and Toeplitz plus Han-kel operators, Bezout equations, inertia type results, matrix polynomials, in oneand severable variables, and related areas in matrix and operator theory.
We present this book to Leonid Lerer on behalf of all the authors as a tokenof our respect and gratitude. We wish him many more years of good health andhappiness.
March 2013 Rien KaashoekLeiba RodmanHugo Woerdeman
Leonid Arie Lerer in an undated photograph
Operator Theory:Advances and Applications, Vol. 237, ixâxiicâ 2013 Springer Basel
Leonid Lererâs Curriculum Vitae
Date and place of birth: April 19, 1943; USSRDate of immigration: December 1973Marital status: Married, 2 children
Academic degrees
1965 M.Sc. Mathematics, Magna Cum LaudeKishinev State University, Kishinev
1969 Ph.D. MathematicsPhysico-Technical Institute of Low Temperatures of theAcademy of Sciences of the Ukrainian SSR, Kharkov
Academic appointments
1969â73 Lecturer, Dept. of Physics and MathematicsKishinev State University
1974â75 Senior Researcher,Technion Research & Development FoundationTechnion, Haifa
1975â80 Senior Lecturer, Dept. of Mathematics, Technion, Haifa
1981â88 Assoc. Professor, Dept. of Mathematics, Technion, Haifa
1988 (Feb.)â Full Professor, Dept. of Mathematics, Technion, Haifa
Visiting academic appointments
Visiting Professor, Dept. of Mathematics and Computer Scienceat Vrije Universiteit, Amsterdam:
Feb. 1984âFeb. 1985, Mar.âSept. 1990, Apr.âNov. 2002,Feb.âApr. 2003, Jul.âAug. 2003
Invited Guest, Dept. of Mathematics and Computer Scienceat Vrije Universiteit, Amsterdam:
Jul.âAug. 1978, 1980, 1982, 1983, 1985â93, 1995, 1996â97,2000, 2004, Sept. 2005 (3â6 weeks each time)
x Leonid Lererâs Curriculum Vitae
Invited Guest, Institute for Mathematics and its Applications,University of Minnesota, Minneapolis:
June 1992 (2 weeks)
Invited Guest, The Thomas Stieltjes Institute of Mathematics,Vrije Universiteit, Amsterdam:
Jul.âAug. 1994 (5 weeks)
Research field
Operator Theory, Systems Theory, Linear Algebra, Integral Equations
Public professional activities
â Editorial board of the international journal âIntegral Equations and OperatorTheoryâ (since its foundation in 1977).
â Editorial board of the book series âOperator Theory: Advances and Appli-cationsâ, Birkhauser Verlag, BaselâBostonâBerlin.
â Special editor of âLinear Algebra and its Applicationsâ, 1988 and 2005.â Editor of âGohberg Memorial Volumesâ, Birkhauser, Basel, 2009â.â Editor of âConvolution equations and singular integral operators. Selectedpapersâ, Birkhauser, Basel, 2009â2010.
â Organizing committee for:1. The third, forth, fifth, sixth, seventh, ninth, tenth, eleventh, thirteenth,fourteens and fifteenth Haifa Matrix Theory Conferences, 1985â2007.
2. International Conference âOperator Theory: Advances and Applica-tionsâ, Calgary, Canada, 1988.
3. Conference of the International Linear Algebra Society (ILAS), 2001.4. Workshops on Operator Theory and Applications, Amsterdam, 2002and 2003.
â Organizer and Chairman of Invited Special Sessions at1. International Symposium on the Mathematical Theory of Networks andSystems, Amsterdam, The Netherlands, 1989.
2. Second SIAM Conference on Linear Algebra in Signals, Systems andControl, San Francisco, USA, 1991.
3. International Symposium on Mathematical Theory of Networks andSystems, St. Louis, Missouri, USA, 1991.
4. The 14th Matrix Theory Conference, Haifa, 2007.5. The 15th Matrix Theory Conference, Haifa, 2009.6. International Workshop on Operator Theory and Applications,Williamsburg, USA, 2008.
â Program Committees for the International Symposia of the MathematicalTheory of Networks and Systems:1. MTNS â 89, Amsterdam, The Netherlands;2. MTNS â 93, Regensburg, Germany
Leonid Lererâs Curriculum Vitae xi
Grants and awards
â Israel-U.S. Binational Science Foundation (BSF) Grant, 1988â1992.â Fellowship of the Netherlands Organization for Scientific Research (NWO),MarchâJune, 1990.
â Grants from the Technion V.P.R. Fund â annually since 1979.â Grants from the Fund for Promotion of Research at the Technion â annuallysince 1981.
â Israel-U.S. Binational Science Foundation (BSF) Grant, 1995â1999.â Fellowship of the Netherlands Organization for Scientific Research (NWO),AprilâSeptember 2002.
â Grant for Promotion of Funded Research ($ 2,500) (for the proposal submit-ted to BSF, graded âvery goodâ but not funded by this agency), 2005.
â ISF grant 121/09 for the period 2009â2012, $ 40,000 yearly (still in progress).
Graduate students
B.A. Kon M.Sc. 1976Thesis title: âAsymptotic distribution of the spectra ofmultidimensional WienerâHopf operators andmultidimensional singular integral operatorsâ.
M. Tismenetsky M.Sc. 1977Thesis title: âSpectral analysis of polynomial pencils via theroots of the corresponding matrix equationsâ.
B.A. Kon Ph.D. 1981Thesis title: âOperators of WienerâHopf types and resultantsof analytic functionsâ.
M.Tismenetsky Ph.D. 1981Thesis title: âBezoutians, Toeplitz and Hankel matrices in thespectral theory of matrix polynomialsâ.
H.J. Woerdeman M.Sc. 1985 (Vrije Univ., Amsterdam)Thesis title: âResultant operators for analytic matrix functionsin a finitely connected domainâ.
J. Haimovici Ph.D. 1991Thesis title: âOperator equations and Bezout operators foranalytic operator functionsâ.
I. Karelin M.Sc. 1993Thesis title: âThe algebraic Riccati equation and the spectraltheory of matrix polynomialsâ.
G. Gomez Ph.D. 1996Thesis title: âBezout operators for analytic operator functionsand inversion of structured operatorsâ.
xii Leonid Lererâs Curriculum Vitae
I. Karelin Ph.D. 2000Thesis title: âFactorization of rational matrix functions,generalized Bezoutians and matrix quadratic equationsâ.
I. Margulis M.Sc. 2008Thesis title: âInertia theorems based on operator equations ofLyapunov type and their applicationsâ.
I. Margulis Ph.D. in progress.
Memberships
â Israel Mathematical Societyâ Society for Industrial and Applied Mathematicsâ The New York Academy of Sciencesâ European Mathematical Society
Invited talks at conferences
â Invited participation at 80 conferences.â Plenary speaker at:
1. SIAM Conference on Linear Algebra in Signals, Systems and Control,Boston, Mass., USA, 1986.
2. Workshop on Matrix and Operator Theory, Rotterdam, The Nether-lands, 1989.
3. Workshop on Linear Algebra for Control Theory, Institute for Mathe-matics and its Applications, University of Minnesota, USA, June 1992.
4. Colloquium of the Royal Dutch Academy of Arts and Sciences on Chal-lenges of a Generalized System Theory, June 1992.
5. The Second Conference of ILAS, Lisbon, Portugal, August 1992.6. The Third Conference of ILAS, Rotterdam, The Netherlands, August1994.
7. International Workshop on Operator Theory and Applications(IWOTA-95), Regensburg, Germany, August 1995.
8. Workshop on Operator Theory and Analysis on the occasion of the 60thbirthday of M.A. Kaashoek, Vrije Universiteit, Amsterdam, November1997.
9. AMS-IMS-SIAM Summer Research Conference on Structured Matri-ces in Operator Theory, Numerical Analysis, Control, Signal and ImageProcessing, July 1999.
10. Workshop on Operator Theory, System Theory and Scattering Theory,Beer-Sheba, June 2005.
Leonid (Arie) LererDepartment of MathematicsTechnion-Israel Institute of TechnologyHaifa 32000, Israele-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, xiiiâxixcâ 2013 Springer Basel
Leonid Lererâs List of Publications
Theses
1. M.Sc. âLocalization of zeroes of polynomials and some properties of normalmatricesâ, Kishinev State University, 1965.
2. Ph.D. âSome problems in the theory of linear operators and in the theory ofbases in locally convex spacesâ, Physico-Technical Institute of Low Temper-atures of the Academy of Sciences of the Ukrainian SSR, Kharkov, 1969.
Original papers in professional journals
1. L. Lerer, On the diagonal elements of normal matrices, Mat. Issled. 2 (1967),153â163.
2. L. Lerer, About the spectral theory of bounded operators in a locally convexspace, Mat. Issled. 2 (1967), 206â214.
3. L. Lerer, On completeness of the system of root vectors of a Fredholm oper-ator in a locally convex space, Mat. Issled. 3 (1968), 31â60.
4. L. Lerer, Basic sequences in a Montel space, Mat. Zametki 6 (1969), 329â334.Announcement of results: Mat. Issled. 3 (1968), 235â236.
5. L. Lerer, Certain criteria for stability of bases in locally convex spaces, I,Mat. Issled. 4 (1969), 35â55.
6. L. Lerer, Certain criteria for stability of bases in locally convex spaces, II,Mat. Issled. 4 (1969), 42â57.
7. L. Lerer, The stability of bases in locally convex spaces, Dokl. Akad. NaukU.S.S.R. 184 (1969), 30â33; Soviet Math. Dokl. 10 (1969), 24â28.
8. L. Lerer, On a class of perturbation for operators that admit reduction, Mat.Issled. 6 (1971), 168â173.
9. S. Buzurniuk, L. Lerer and E. Shevchik, An operator modification of SeidelMethod, Tezisi Dokl. Nauchn. Konf. Kishinev, (1972), 22â23.
10. L. Lerer, The asymptotic distribution of the spectra of finite truncations ofWienerâHopf operators, Dokl. Akad. Nauk U.S.S.R. 207 (1972), 1035â1038;Soviet Math. Dokl. 13 (1972), 1651â1655.
11. L. Lerer, The asymptotic distribution of the spectra. I. General theorems andthe distribution of the spectrum of truncated WienerâHopf operators, Mat.Issled. 8 (1972), 141â164.
xiv Leonid Lererâs List of Publications
12. L. Lerer, On the asymptotic distribution of the spectra. II. The distributionof the spectrum of truncated dual operators, Mat. Issled. 8 (1973), 84â95.
13. I. Gohberg and L. Lerer, Resultants of matrix polynomials, Bull. Amer. Math.Soc. 82 (1976), 565â567.
14. I. Gohberg and L. Lerer, Singular integral operators as a generalization ofthe resultant matrix, Applicable Anal. 7 (1977/78), 191â205.
15. I. Gohberg, L. Lerer and L. Rodman, Factorization indices for matrix poly-nomials, Bull. Amer. Math. Soc. 84 (1978), 275â277.
16. L. Lerer, On approximating the spectrum of convolution type operators. I.WienerâHopf matricial integral operator, Israel J. Math. 30 (1978), 339â362.
17. I. Gohberg and L. Lerer, On resultant operators of a pair of analytic functions,Proc. Amer. Math. Soc. 72 (1978), 65â73.
18. I. Gohberg, L. Lerer and L. Rodman, On canonical factorization of opera-tor polynomials, spectral divisors and Toeplitz matrices, Integral EquationsOperator Theory 1 (1978), 176â214.
19. I. Gohberg and L. Lerer, Factorization indices and Kronecker indices, IntegralEquations Operator Theory 2 (1979), 199â243. Erratum: Integral EquationsOperator Theory 2 (1979), 600â601.
20. I. Gohberg, L. Lerer and L. Rodman, Stable factorization of operator poly-nomials, I. Spectral divisors simply behaved at infinity, J. Math. Anal. Appl.74 (1980), 401â431.
21. I. Gohberg, L. Lerer and L. Rodman, Stable factorization of operator polyno-mials, II. Main results and applications to Toeplitz operators, J. Math. Anal.Appl. 75 (1980), 1â40.
22. I. Gohberg, M.A. Kaashoek, L. Lerer and L. Rodman, Common multiples andcommon divisors of matrix polynomials, I. Spectral method, Indiana Univ.Math. J. 30 (1981), 321â356.
23. L. Lerer and M. Tismenetsky, The Bezoutian and the eigenvalue separationproblem, Integral Equations Operator Theory 5 (1982), 386â445.
24. I. Gohberg, M.A. Kaashoek, L. Lerer and L. Rodman, Common multiplesand common divisors of matrix polynomials, II. Vandermonde and resultant,Linear and Multilinear Algebra 12 (1982/83), 159â203.
25. I. Gohberg and L. Lerer, On non-square sections of WienerâHopf operators,Integral Equations Operator Theory 5 (1982), 518â532. Errata: Integral Equa-tions Operator Theory 6 (1983), 904.
26. I. Gohberg, L. Lerer and L. Rodman, WienerâHopf factorization of piecewisematrix polynomials, Linear Algebra Appl. 52/53 (1983), 315â350.
27. L. Lerer, L. Rodman and M. Tismenetsky, Bezoutian and SchurâCohn prob-lem for operator polynomials, J. Math. Anal. Appl. 103 (1984), 83â102.
Leonid Lererâs List of Publications xv
28. P. Lancaster, L. Lerer and M. Tismenetsky, Factored forms for solutions ofðŽð âððµ = ð¶ and ð â ðŽððµ = ð¶ in companion matrices, Linear AlgebraAppl. 62 (1984), 19â49.
29. I. Gohberg, M.A. Kaashoek, L. Lerer and L. Rodman, Minimal divisors ofrational matrix functions with prescribed zero and pole structure, in: Topicsin operator theory systems and networks (Rehovot, 1983), Oper. Theory Adv.Appl. 12, Birkhauser Verlag, Basel, 1984, pp. 241â275.
30. L. Lerer and M. Tismenetsky, On the location of spectrum of matrix poly-nomials, Contemp. Math. 47 (1985), 287â297.
31. B. Kon and L. Lerer, Resultant operators for analytic functions in a simpleconnected domain, Integral Equations Operator Theory 9 (1986), 106â120.
32. I. Gohberg, M.A. Kaashoek and L. Lerer, Minimality and irreducibility of thetime invariant linear boundary value systems, Internat. J. Control 44 (1986),363â379.
33. I. Gohberg, M.A. Kaashoek, L. Lerer and L. Rodman, On Toeplitz andWienerâHopf operators with contourwise rational matrix and operator sym-bols, in: Constructive methods of WienerâHopf factorization, Oper. TheoryAdv. Appl. 21, Birkhauser Verlag, Basel, 1986, pp. 75â126.
34. L. Lerer and M. Tismenetsky, Generalized Bezoutian and the inversion prob-lem for block matrices, I. General scheme, Integral Equations Operator Theory9 No. 6 (1986), 790â819.
35. L. Lerer and H.J. Woerdeman, Resultant operators and the Bezout equationfor analytic matrix functions, I. J. Math. Anal. Appl. 125 (1987), 531â552.
36. L. Lerer and H.J. Woerdeman, Resultant operators and the Bezout equationfor analytic matrix functions, II. J. Math. Anal. Appl. 125 (1987), 553â567.
37. I. Gohberg, M.A. Kaashoek and L. Lerer, On minimality in the partial real-ization problem, Systems Control Lett. 9 (1987), 97â104.
38. I. Gohberg and L. Lerer, Matrix generalizations of M.G. Kreın theorems onorthogonal polynomials, in: Orthogonal matrix-valued polynomials and appli-cations (Tel Aviv, 1987/88), Oper. Theory Adv. Appl. 34, Birkhauser Verlag,Basel, 1987/88, pp. 137â202.
39. L. Lerer and M. Tismenetsky, Generalized Bezoutian and matrix equations,Linear Algebra Appl. 99 (1988), 123â160.
40. I. Gohberg, M.A. Kaashoek and L. Lerer, Nodes and realizations of rationalmatrix functions: minimality theory and applications, in: Topics in operatortheory and interpolation, Oper. Theory Adv. Appl. 29, Birkhauser Verlag,Basel, 1988, pp. 181â232.
41. L. Lerer, The matrix quadratic equation and factorization of matrix polyno-mials, in: The Gohberg anniversary collection, Vol. I (Calgary, AB, 1988),Oper. Theory Adv. Appl. 40, Birkhauser Verlag, Basel, 1989, pp. 279â324.
xvi Leonid Lererâs List of Publications
42. L. Lerer, L. Rodman and M. Tismenetsky, Inertia theorems for matrix poly-nomials, Linear and Multilinear Algebra 30 (1991), 157â182.
43. I. Gohberg, M.A. Kaashoek and L. Lerer, A directional partial realizationproblem, Systems Control Lett. 17 (1991), 305â314.
44. I. Gohberg, M.A. Kaashoek and L. Lerer, Minimality and realization of dis-crete time-varying systems, in: Time-variant systems and interpolation, Oper.Theory Adv. Appl. 56, Birkhauser Verlag, Basel, 1992, pp. 261â296.
45. L. Lerer and L. Rodman, Spectrum separation and inertia for operator poly-nomials, J. Math. Anal. Appl. 169 (1992), 260â282.
46. I. Gohberg, M.A. Kaashoek and L. Lerer, Minimal rank completion problemsand partial realization, in: Recent Advances in Math. Theory of Systems,Control, Networks and Signal Processing, I, Mita Press, Tokyo, 1992, 65â70.
47. L. Lerer and L. Rodman, Sylvester and Lyapunov equations and some in-terpolation problems for rational matrix functions, Linear Algebra Appl. 185(1993), 83â117.
48. I. Koltracht, B.A. Kon and L. Lerer, Inversion of structured operators, Inte-gral Equations Operator Theory 20 (1994), 410â448.
49. G. Gomez and L. Lerer, Generalized Bezoutian for analytic operator functionsand inversion of structured operators, in: Systems and Networks: Mathemat-ical Theory and Applications (U. Helmke, R. Mennicken, J. Saures, eds.),Akademie Verlag, 1994, pp. 691â696.
50. L. Lerer and L. Rodman, Bezoutians and factorizations of rational matrixfunctions and matrix equations, in: Systems and Networks: MathematicalTheory and Applications (U. Helmke, R. Mennicken, J. Saures, eds.), Akade-mie Verlag, 1994, pp. 761â766.
51. I. Haimovici and L. Lerer, Bezout operators for analytic operator functions.I. A general concept of Bezout operator, Integral Equations Operator Theory21 (1995), 33â70.
52. L. Lerer and L. Rodman, Inertia of operator polynomials and stability ofdifferential equations, J. Math. Anal. Appl. 192 (1995), 579â606.
53. L. Lerer and L. Rodman, Common zero structure of rational matrix functions,J. Funct. Anal. 136 (1996), 1â38.
54. I. Gohberg, M.A. Kaashoek and L. Lerer, Factorization of banded lower tri-angular infinite matrices, Linear Algebra Appl. 247 (1996), 347â357.
55. L. Lerer and L. Rodman, Symmetric factorization and localization of zeroesof rational matrix functions, Linear and Multilinear Algebra 40 (1996), 259â281.
56. L. Lerer and L. Rodman, Bezoutian of rational matrix functions, J. Funct.Anal. 141 (1996), 1â36.
Leonid Lererâs List of Publications xvii
57. L. Lerer and A.C.M. Ran, ðœ-pseudo-spectral and ðœ-inner-pseudo-outer fac-torization for matrix polynomials, Integral Equations Operator Theory 29(1997), 23â51.
58. L. Lerer and L. Rodman, Inertia theorems for Hilbert space operators basedon Lyapunov and Stein equations, Math. Nachr. 198 (1999), 131â148.
59. L. Lerer and L. Rodman, Bezoutian of rational matrix functions, matrix equa-tions and factorizations of rational matrix functions, Linear Algebra Appl.302-303 (1999), 105â133.
60. I. Karelin and L. Lerer, Generalized Bezoutian, matrix quadratic equationsand factorization of rational matrix functions, in: Recent advances in operatortheory (Groningen, 1998), Oper. Theory Adv. Appl. 122, Birkhauser Verlag,Basel, 2001, pp. 303â321.
61. I. Karelin, L. Lerer and A.C.M. Ran, ðœ-symmetric factorization and the al-gebraic Riccati equation, in: Recent advances in operator theory (Groningen,1998), Oper. Theory Adv. Appl. 124, Birkhauser Verlag, Basel, 2001, pp.319â360.
62. I. Karelin and L. Lerer, Matrix quadratic equations and column/row factor-ization of matrix polynomials, Int. J. Appl. Math. Comput. Sci. 11 (2001),1285â1310.
63. L. Lerer and L. Rodman, Inertia bounds for matrix polynomials and appli-cations, in: Linear operators and matrices, Oper. Theory Adv. Appl. 130,Birkhauser Verlag, Basel, 2002, pp. 255â276.
64. L. Lerer and A. Ran, A new inertia theorem for Stein equations, inertia of in-vertible block Toeplitz matrices and matrix orthogonal polynomials, IntegralEquations Operator Theory 47 (2003), 339â360.
65. L. Lerer, M.A. Petersen and A.C.M. Ran, Existence of minimal nonsquareðœ-symmetric factorizations for self-adjoint rational matrix functions, LinearAlgebra Appl. 379 (2004), 159â178.
66. I. Gohberg, I. Haimovici, M.A. Kaashoek and L. Lerer, The Bezout inte-gral operator: main property and underlying abstract scheme, in: The statespace method generalizations and applications, Oper. Theory Adv. Appl. 161,Birkhauser Verlag, Basel, 2005, pp. 225â270.
67. I. Gohberg, M.A. Kaashoek and L. Lerer, Quasi-commutativity of entire ma-trix functions and the continuous analogue of the resultant, in:Modern opera-tor theory and applications, Oper. Theory Adv. Appl. 170, Birkhauser Verlag,Basel, 2007, pp. 101â106.
68. I. Gohberg, M.A. Kaashoek and L. Lerer, The continuous analogue of the re-sultant and related convolution operators, in: The Extended Field of OperatorTheory, Oper. Theory Adv. Appl. 171, Birkhauser, Basel, 2007, 107â127.
69. I. Gohberg, M.A. Kaashoek and L. Lerer, On a class of entire matrix functionequations, Linear Algebra Appl. 425 (2007), 434â442.
xviii Leonid Lererâs List of Publications
70. I. Gohberg, M.A. Kaashoek and L. Lerer, The inverse problem for Kreınorthogonal matrix functions, (Russian) Funktsional. Anal. i Prilozhen. 41(2007), 44â57; translation in Funct. Anal. Appl. 41 (2007), 115â125.
71. L. Lerer, I. Margulis and A.C.M. Ran, Inertia theorems based on operatorLypunov equations, Oper. Matrices 2 (2008), 153â166.
72. I. Gohberg, M.A. Kaashoek and L. Lerer, The resultant for matrix polynomi-als and quasi commutativity, Indiana Univ. Math. J. 57 (2008), 2793â2813.
73. D. Alpay, I. Gohberg, M.A. Kaashoek, L. Lerer and A. Sakhnovich, Kreınsystems, in: Modern analysis and applications. The Mark Kreın CentenaryConference. Vol. 2: Differential operators and mechanics, Oper. Theory Adv.Appl. 191, Birkhauser Verlag, Basel, 2009, pp. 19â36.
74. M.A. Kaashoek and L. Lerer, Quasi-commutativity of regular matrix poly-nomials: resultant and Bezoutian, in: Topics in operator theory. Volume 1.Operators, matrices and analytic functions, Oper. Theory Adv. Appl. 202,Birkhauser Verlag, Basel, 2010, pp. 297â314.
75. M.A. Kaashoek, L. Lerer and I. Margulis, Kreın orthogonal entire matrixfunctors and related Lypunov equations: A state space approach, IntegralEquations Operator Theory 65 (2009), 223â242.
76. D. Alpay, I. Gohberg, M.A. Kaashoek, L. Lerer and A.L. Sakhnovich, Kreınsystems and canonical systems on a finite interval: accelerants with a jumpdiscontinuities at the origin and continuous potentials, Integral EquationsOperator Theory 68 (2010), 115â150.
77. L. Lerer and A.C.M. Ran, The discrete algebraic Riccati equation and Her-mitian block-Toeplitz matrices, in: A panorama of modern operator theoryand related topics, Oper. Theory Adv. Appl. 218, Birkhauser/ Springer BaselAG, Basel, 2012, pp. 495â512.
78. M.A. Kaashoek and L. Lerer, The band method and inverse problems fororthogonal functions of SzegoâKreın type, Indag. Math. (N. S.) 23 (2012),900â920.
79. M.A. Kaashoek and L. Lerer, On a class of matrix polynomial equations,Linear Algebra Appl. 439 (2013), 613â620.
Edited books
1. Convolution equations and singular integral equations. Edited by L. Lerer,V. Olshevsky and I. Spitkovsky, Oper. Theory Adv. Appl. 206, BirkhauserVerlag, Basel, 2010.
2. A panorama of modern operator theory and related topics. The Israel Go-hberg memorial volume. Edited by Harry Dym, Marinus A. Kaashoek, PeterLancaster, Heinz Langer and Leonid Lerer. Oper. Theory Adv. Appl. 218,Birkhauser/Springer Basel AG, Basel, 2012.
Leonid Lererâs List of Publications xix
Other publications(Reports listed were not fully incorporated in papers published elsewhere)
1. I. Gohberg, L. Lerer and L. Rodman, On factorization, indices and com-pletely decomposable matrix polynomials, Technical report 80â47, Tel-AvivUniversity, 1980, 72 pages.
2. Feldman, I.A. WienerâHopf operator equations and their application to thetransport equation. Translated from the Russian by C.G. Lekkerker, L. Lererand R. Troelstra. Integral Equations Operator Theory 3 (1980), 43â61.
3. I. Gohberg, M.A. Kaashoek, L. Lerer and L. Rodman, Common multiplesand common divisors of matrix polynomials, II. Vandermonde and resultantmatrices, Technical report 80â53, Tel-Aviv University, 1981, 122 pages.
4. L. Lerer and H.J. Woerdeman, Resultant operators and the Bezout equationfor analytic matrix functions, Rapport No. 299, Vrije Universiteit, Amster-dam, 1985, 54 pages.
5. H. Bart, M.A. Kaashoek and L. Lerer, Review on âMatrix Polynomialsâ byI. Gohberg, P. Lancaster and L. Rodman, Linear Algebra Appl. 64 (1985),167â172.
6. L. Lerer and M. Tismenetsky, Toeplitz classification of matrices and inversionformulas, II. Block-Toeplitz and perturbed block-Toeplitz matrices, TechnicalReport, IBM SC, Haifa, 1986, 38 pages.
7. L. Lerer and A.C.M. Ran, On a new inertia theorem and orthogonal polyno-mials, Proceedings of the Sixteenth International Symposium on MathematicalTheory of Networks and Systems (MTNS), Leuven, Belgium, 2004.
8. A. Berman, L. Lerer and R. Loewy, Preface to the 2005 Haifa Matrix TheoryConference at the Technion, Linear Algebra Appl. 416 (2006), 15â16.
9. M.A. Kaashoek and L. Lerer, Gohbergâs mathematical work in the period1998â2008, in: Israel Gohberg and his Friends, eds. H. Bart, T. Hempfling,M.A. Kaashoek, Birkhauser Verlag, Basel, 2008, pp. 111â115.
10. L. Lerer, V. Olshevsky and I. Spitkovsky, Introduction, in: Convolution equa-tions and singular integral operators, Oper. Theory Adv. Appl. 206, Birkhau-ser Verlag, Basel, 2010, pp. ixâxxii.
Operator Theory:Advances and Applications, Vol. 237, 1â7câ 2013 Springer Basel
Leonia Lererâs Mathematical Workand Amsterdam Visits
M.A. Kaashoek
It is a great pleasure to congratulate professor Lerer on the occasion of his 70thbirthday and to wish him many happy returns.
I will refer to professor Lerer as Leonia, the name used by his Dutch friends.Arie would have been an alternative. The latter is a very common Dutch name,but few know about its Hebrew meaning. Leonid is out of the question; too manypolitical recollections.
Leonia Lerer, Bil Helton, Harm Bart, Israel Gohberg, Joe PincusVU University campus, 1979
Leonia and I met for the first time in October 1976 at the Mathemati-sches Forschungsinstitut Oberwolfach, a wonderful conference resort in the south-western part of Germany (at that time Bundesrepublik Deutschland), in the blackforest near Freiburg. It was a very special meeting, organized by Gohberg, Gramschand Neubauer, with a select group of 39 participants, including a strong US delega-tion consisting of Kevin Clancey, Lewis Coburn, Chandler Davis, Ron Douglas, Bill
2 M.A. Kaashoek
Helton, and Joe Pincus. From Israel, besides Gohberg, three other participants:Harry Dym, Paul Furhrman, and Leonia Lerer. Directly after this meeting RonDouglas and Leonia came to Amsterdam. It was Leoniaâs first visit. Three yearslater, again directly after an Oberwolfach meeting, we had a mini-conference inAmsterdam with the five persons on the picture at the previous page as the mainlecturers. If I remember it correctly, this was Leoniaâs third visit to Amsterdam.
Many other visits were to follow. There were short visits of about 3 to 5weeks, in total more than 20, often supported by the Dutch National ScienceFoundation. Apart from that, having a sabbatical at the Technion, Leonia heldvisiting professorships at the VU University, for four periods:
â February 1984âFebruary 1985,â MarchâSeptember 1990,â AprilâNovember 2002,â FebruaryâApril and JulyâAugust 2003.
In his first period as visiting professor, Leonia supervised the master thesis of HugoWoerdeman, now professor at Drexel University and co-editor of this volume. Thejoint work of Leonia and Hugo resulted in two papers on âResultant operators andthe Bezout equation for analytic matrix functions,â both appeared in 1987 in theJournal of Math. Anal. Appl. Later in an acknowledgment in his PhD thesis, Hugowrote: I am indebted to professor Lerer, who introduced me in a very stimulatingway to mathematical research.
When one lives in the Netherlands for so many years as Leonia did, one learnsthe local customs and uses the local means of transportation:
On the bicycle with Israel Gohberg.
Leonia Lererâs Mathematical Work and Amsterdam Visits 3
Leonia and I have 24 joint papers, of these 24 papers 19 were written jointlywith Israel Gohberg. Our many meetings, in the Netherlands as well as in Israel,and our joint work belong to the gratifying experiences of my life. Andre Ran isLeoniaâs second co-author at the VU Amsterdam; together Leonia and Andre wrote7 joint papers on topics involving ðœ-spectral factorization and inertia theorems.
What did Leonia talk about at Oberwolfach in 1976? Here I present theabstract of his talk as it appears in the Tagungsbericht :
The abstract has three elements. First element: from polynomials to analyticfunctions; second element: from resultant matrices to resultant operators, andthird element: singular integral operators as generalized resultant matrices. It wasa beautiful lecture.
Leoniaâs work has a wonderful mix of matrix and operator theory on the onehand and matrix function theory on the other hand. It reminds me of a statementPaul Halmos once made in an interview. He said: I still have this religion that ifyou know the answer to every matrix question, somehow you answer every operatorquestion. I do not believe in this statement, and I think Leonia does not either;there is much more two way traffic between the two fields. But Leoniaâs talk atOberwolfach certainly provided some support for the Halmos religion.
Resultants and Bezout operators form a main theme in Leoniaâs work. Aboutone third of his papers after 1976 have the word resultant or Bezout in the title.His talk in Oberwolfach was based on his first papers after immigration to Israel.They appeared in the Bulletin and Proceedings of the AMS, both in 1976, andboth co-authored by Gohberg.
His work in this area is partially motivated by mathematical system andcontrol theory with location of zeros and problems of stability as main themes.The famous Anderson-Jury paper in IEEE Transactions Automatic Control from1976 served as a source of inspiration for his later work on generalized Bezoutoperators. As a further illustration of Leoniaâs work I will discuss in more detail
4 M.A. Kaashoek
his 1994 paper, with Israel Koltracht and Ben Kon as co-authors. I consider thisarticle as one of Leoniaâs top papers. It has a short title:
The KKL-paper deals with bounded linear integro-differential operators ðŽon ð¿2[0, ð] of which the action is given by
(ðŽð)(ð¥) =ð
ðð¥
â« ð
0
(â
âð¡ÎŠ(ð¥, ð¡)
)ð(ð¡) ðð¡,
Ί(ð¥, ð¡) =1
2
ðâð=1
â« 2ðââ£ð¥âð¡â£
ð¥+ð¡
ðð
(ð + ð¥â ð¡
2
)ðð
(ð â ð¥+ ð¡
2
)ðð .
Here ðð , ðð â ð¿2[0, ð], ð = 1, 2, . . . , ð. If ðŽ is as above, we say that ðŽ belongs to theKKL class, and we shall refer to ðð , ðð â ð¿2[0, ð], ð = 1, 2, . . . , ð, as the definingdata.
The paper has two beautiful theorems. To state the first theorem we needthe Volterra integral operator on ð¿2[0, ð] which is defined by
(ð ð)(ð¥) = âðâ« ð¥
0
ð(ð¡) ðð¡, 0 †ð¥ †ð.
Furthermore, with the defining data ðð, ðð â ð¿2[0, ð], ð = 1, 2, . . . , ð, and theVolterra integral operator we associate the Lyapunov equation
ðð â ð âð =
ðâð=1
ððâšâ , ððâ©. (1)
Note that the right-hand side of the Lyapunov equation (1) is an operator of finiterank. If this operator is zero, then ð is also equal to zero. Thus, using terminol-ogy common for structured matrices, an operator ð satisfying (1) is an operatorwith a relatively small displacement. The authors proved that these operators ðwith a relative small displacement are precisely the integro-differential operatorsðŽ introduced above. This is the first beautiful theorem.
Theorem 1. [KKL-1994] An operator ðŽ on ð¿2[0, ð] belongs to the KKL class ifand only if for some ðð , ðð â ð¿2[0, ð], ð = 1, 2, . . . , ð the operator ð = ðŽ satisfiesthe identity (1).
Now assume that ðŽ satisfies the Lyapunov equation (1), and let ðŽ be invert-ible. Multiplying (1) from the left and from the right by the inverse of ðŽ yields
Leonia Lererâs Mathematical Work and Amsterdam Visits 5
another Lyapunov operator identity, which is analogous to (1):
ð ðŽâ1 â ðŽâ1ð â =ðâ
ð=1
ðŸðâšâ , ðœðâ©, where ðŽðœð = ðð and ðŽðŸð = ðð .
A variant of the first theorem now yields the second which can be viewed asan operator analogue of the finite-dimensional inversion theorems for Toeplitzmatrices due to GohbergâSemencul and GohbergâHeinig.
Theorem 2. [KKL-1994] Assume ðŽ belongs to the KKL class with defining dataðð , ðð â ð¿2[0, ð], ð = 1, 2, . . . , ð, and let ðŽ be a Fredholm operator. If the equations
ðŽðœð = ðð and ðŽðŸð = ðð (ð = 1, . . . , ð)
are solvable, then ðŽ has a bounded inverse and ðŽâ1 is given by
(ðŽâ1ð) =ð
ðð¥
â« ð
0
(â
âð¡ð(ð¥, ð¡)
)ð(ð¡) ðð¡,
where
ð(ð¥, ð¡) =
ðâð=1
â« min(ð¥,ð¡)
0
ðŸð(ð¥â ð )ðœð(ð¡â ð ) ðð .
The KKL paper present lots of examples, another characteristic of Leoniaâswork. Here we mention the following two, both are taken from the KKL paper.
Example 1. Let ð â ð¿1[âð, ð], and let ðŽ be the operator from the KKL classdefined by the following data:
ð1(ð¥) = 1, ð2(ð¥) = 1 +
â« ðâð¥
0
ð(ð ) ðð ,
ð1(ð¥) =
â« 0
ð¥âð
ð(ð ) ðð , ð2(ð¥) = 1, 0 †ð¥ †ð.
Then ðŽ is the convolution operator given by
(ðŽð)(ð¡) = ð(ð¡) +
â« ð
0
ð(ð¡â ð )ð(ð ) ðð .
Example 2. Let ð, ð â ð¿2[0, ð], and let ðŽ be the operator from the KKL classdefined by the following data:
ð1(ð¥) = ð(ð¥), ð2(ð¥) = ð(ð¥),
ð1(ð¥) = 1â ð(ð â ð¥) ð2(ð¥) = 1â ð(ð â ð¥), 0 †ð¥ †ð.Then ðŽ is the Bezout operator defined by the entire functions
ð¹ (ð§) = 1 + ðð§
â« ð
ð
ððð§ð¡ð(ð¡) ðð¡ and ðº(ð§) = 1 + ðð§
â« ð
ð
ððð§ð¡ð(ð¡) ðð¡.
Other themes in the work of Leonia are:
â Operators in locally convex spacesâ Asymptotic distribution of spectra and related limit theorems
6 M.A. Kaashoek
â Rational matrix functionsâ Spectral theory of matrix and operator polynomialsâ Inverse problems for SzegoâKrein orthogonal polynomials and their continu-ous analogs
â Partial realization problemsâ Minimality of partial realization problems in discrete time, including multi-variable systems
The majority of my joint papers with Leonia belong to the areas described by thelast four bullets.
Leonia and Israel Gohberg
Gohberg had a profound influence on the work of Leonia. He was Leoniaâsmathematical (grand-)father, and according to MatSciNet Gohberg is Leoniaâs topco-author. Gohbergâs passing away in October 2009 was a great personal loss forLeonia, as for many of us. For Leonia it meant a set back for a long period.
Leonia and Alek MarkusAmsterdam (November 29, 2002)
Leonia Lererâs Mathematical Work and Amsterdam Visits 7
Alek Markus is Leoniaâs doctor-father. Of course, the mathematical son fol-lowed his mathematical fathers in many ways, but not always. On one particularnon-mathematical point, it was the other way around: the mathematical son wasleading and his mathematical fathers were following. Of the three, Leonia was thefirst to immigrate from Kishinev to Israel. Gohberg followed later and Markusmany years later.
I conclude with best wishes: for Leonia personally, for Bertha his wife, andfor his two daughters Hannah and Safira, and for the new family member Gal, thehusband of Safira. It is my sincere hope that both of us will have the time andenergy to continue our joint work.
M.A. KaashoekDepartment of MathematicsFaculty of SciencesVU UniversityDe Boelelaan 1081aNL-1081 HV Amsterdam, The Netherlandse-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 9â10câ 2013 Springer Basel
Leonia Lerer and My First Visit to Israel
H. Bart
In 1981, my wife Greetje and I went to Israel for the first time. I was to attendthe Toeplitz Memorial Conference in Tel-Aviv organized by Israel Gohberg. Butbefore going there we went to Haifa. There, at the Technion, I gave my veryfirst lecture in Israel. The topic was âNew methods for solving integral equationsâ.Leonia Lerer was our host. We had already met him before during one of his visitsto Amsterdam.
For both Greetje and me, coming to Israel meant something special. Being childrenof traditionally protestant parents we were raised with the stories from the Bible.So we were excited to be at the places we had heard about so much. We madeseveral trips, some on our own, but at least one with Leonia. He took us to RoshHaNikra, a white chalk cliff face located on the coast of the Mediterranean Sea,which opens up into spectacular interconnected grottos. Leoniaâs daughter Hannawent with us, doing the driving. Bertha, Leoniaâs wife could not join us becauseshe was pregnant. A little later Saffira was born.
10 H. Bart
We had good contacts as families. Several times Leonia was our guest inThe Netherlands, sometimes on his own, sometimes accompanied by other familymembers. In preparing this note, my wife and I went through our old pictures inorder to find traces of such get togethers. The photographs we found brought backgood memories. The picture above, showing Leonia in our garden talking withRoland Duduchava, was taken in 2003 during a celebration of Israelâs Gohberg75th birthday.
Leonia: congratulations with your seventieth birthday and best wishes to youand your family for the future. Also thanks for everything you did in connectionwith the many Haifa matrix theory conferences.
H. BartEconometric InstituteErasmus University RotterdamP.O. Box 1738NL-3000 DR Rotterdam, The Netherlandse-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 11â12câ 2013 Springer Basel
Through the Eyes of a Student
Irina Karelin
I began my graduate studies in 1989 with Professor Lerer. I did my aliah in March1989. The main reason to come to Haifa was to study in Technion â Israel Instituteof Technology. I was a graduate of Kharkov University (Ukraine) in Mathematicsand wanted to continue my studies. My degree thesis was in Operator theoryarea and I of course expected to find somebody who is working in this area andwill agree to be my supervisor for master studies. I came to the Department ofMathematics in Technion to Professor Abraham Berman. He asked me about mystudies in Kharkov and referred me to Professor Lerer.
We talked about my previous studies and my final student work. Leonid toldme that the subject of my work is rather close to one of his areas of interest, thathe has some problems to be solved and proposed me to do my master degree underhis supervision. I was happy to accept his offer and we started to work together.To approach to the subject of your future research Leonid said you have to reada lot. He proposed me a long list of articles and some books for reading. Amongthese works were the books Matrix Polynomials by Gohberg, Lancaster, Rodman,The Theory of Matrices by Lancaster and Tismenetsky, articles by Lerer and hisjoint works with Rodman, Tismenetsky and some other authors. In the beginningall seemed to me unknown and new but I always could turn to Leonid with anyquestion. I understood it later that all that Leonid advised to do was well thoughtout, methodically correct. All chapters and articles he chose much contributed tomy entering into the subject.
At the same time during my first semester in Technion I attended Leonidâslectures for graduate students on realizations and factorizations of rational matrixfunctions. I liked him as a lecturer. It was not only my impression. I heard thisopinion from other students. His teaching manner was calm and pleasing. Hisexplanations always were very clear and accurate. Atmosphere he created wasalways good and friendly.
The subject of my research was chosen. It was related to factorizations ofcolumn and row reduced matrix polynomials. At the beginning our meetings werenot so frequent, later on we met much more often. But every meeting was very use-ful and productive for me. Sometimes my research progress was good and smooth,sometimes I felt frustrated from unsuccessful attempts to find a solution of some
12 I. Karelin
problem. Even in such crisis moments of my work I knew that great mathematicianand very good man, Professor Lerer always encourages and supports me, turns metowards right direction, gives useful advice.
Leonid himself worked very intensively and also had many unsolved problemsfor his students. When I decided to study for PhD degree and Leonid agreed tocontinue with me there was no question what to do. He proposed me a number ofopportunities for my research. The theme of my work was chosen. It was connectedto factorization of rational matrix functions and Riccati equations. Leonidâs styleof supervision was as it had to be. He didnât generally tell me, and I supposeto other students as well, what to do. He encouraged our independent research,didnât restrict our freedom, but he always might warn us of some things and triedto steer us towards others. So when I prepared my conference report his assistancewas very important and fruitful. He helped me to choose more profitable resultsfor the presentation, to build the report in the right form, to provide possiblequestions.
Many times I run tutorials in the courses where Leonid taught. Students ap-preciated Leonid as a lecturer very much. I heard from students that they liked hisexplanations, liked his logical and consistent presentation of material, illustratedwith many examples and solved problems. Working with Leonid as his teachingassistant I acquired wide experience for my future professional life.
Papers of Leonid always had sharp form and profound content. Statementsand proofs of results were exact, explanations at the same time were sufficient andnever redundant. Leonid also asked his students exact statements and logical, fullexplanations. When I prepared my PhD thesis Leonid read the written text andreturned it back to me many times until the version of the thesis was good.
Today I work in industry as an algorithm developer. While working with Pro-fessor Lerer I acquired many skills very useful in my professional activities. Leonidtaught me to do thorough literature review, helped me to develop my mathematicalthinking, my writing skills, taught me to be an independent researcher.
Dear Professor Lerer! I want to express my deep gratitude and admiration!Happy birthday!
Irina KarelinCmosaiX, Ltd.Kibutz Yagur 30065, Israele-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 13â13câ 2013 Springer Basel
Reminiscences on Visits to Haifa
Andre C.M. Ran
My first contacts with Leonid Lerer date from the middle of the eighties, when Iwas still a PhD student at the VU university. Leonia visited Amsterdam regularly,and we maintained warm contacts without actually working together. This changedwhen I visited Israel for a somewhat longer time in 1996.We started a collaborationwhich produced a first paper in 1997, and I became a frequent visitor to Haifa afterthis. All in all, we wrote six papers together, several with students of Leonia.
To visit the Technion and work with Leonia was always a great pleasure. Themathematics was wonderful, and the warm personality of Leonia made visitingthere a joy. Leonia was aware of the sensitivities of my family regarding travel toIsrael, and so he made sure I was met at the airport by a trusted driver. I was alsonot allowed to leave campus without a local to guide me.
In early 1998 I visited Haifa for the Tenth Haifa Matrix Meeting. Beforeleaving I talked to my father, who was terminally ill. A few days into my visit inHaifa my sister called me to inform me that my father had passed away quietlyon January 7. I will always be grateful for the support and warmth that I receivedfrom Leonia on that day and the following one. Obviously, I had to arrange a veryquick return to the Netherlands, and Leonia helped and comforted me as much aspossible.
Fortunately, I could return to Haifa for several more Haifa Matrix Meetingsand an ILAS Meeting over the course of the years, most of them combined withsome extra time to work together with Leonia.
I would like to close with a heartfelt thanks to Leonia for his inspiring leadin our joint work and for his warm friendship.
Andre C.M. RanAfdeling wiskunde, FEWVrije UniversiteitDe Boelelaan 1081 aNL-1081 HV AmsterdamThe Netherlandse-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 15â15câ 2013 Springer Basel
My First Research Experience
Hugo J. Woerdeman
My first exposure to mathematical research was under the supervision of Profes-sor Leonid (âLeoniaâ) Lerer during the academic year 1984â1985. From the verybeginning we had a very amicable relationship, and what will always stand outin my memory is his consistently calm and friendly demeanor. Leonia was alwaysencouraging, and he never lost his patience, even during the slow periods. He alsomanaged to enjoy life to the fullest â among other things, he savored the freshDutch raw herring during lunch. It was a great environment for me to enter thiscompletely new endeavor. I learned about resultants of analytic functions, singularintegral operators, and many other things I had never heard of, and was expectedto generalize the scalar case to the matrix case. So I learned about realizations,Jordan triples, matrix factorizations, and a lot of other good stuff. So, how did wemanage to make a breakthrough? At some point Leonia encouraged me to workthrough a specific example, and I remember generating pages and pages of com-putations with matrices that very quickly became quite large. And then at somepoint things clicked: I could see the big picture and focus in all those pages ofscribbles on the steps where something real had happened. And the rest is historyas they say, as after a year of hard work we were able to put two papers togetherresulting in my first two publications.
So on this occasion of your 70th birthday, Leonia, I would like to thankyou very much for the excellent guidance you gave me during my initial researchexperience, and for the many years of friendship that followed. I will never forgetthe way my first mathematical breakthrough came about, and it has been a guidefor me ever since. Thank you!
Hugo J. WoerdemanDepartment of MathematicsDrexel UniversityPhiladelphia, PA 19104, USAe-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 17â39câ 2013 Springer Basel
Interpolation in Sub-Bergman Spaces
Joseph A. Ball and Vladimir Bolotnikov
Dedicated to Leonia Lerer, a long-time friend and colleague
Abstract. A general interpolation problem with operator argument is studiedfor functions ð from sub-Bergman spaces associated with an analytic functionð mapping the open unit disk ð» into the closed unit disk.
Mathematics Subject Classification (2010). 30E05, 47A57, 46E22.
Keywords. De BrangesâRovnyak space, Schur-class function.
1. Introduction
Given a Hilbert space ðŽ and a positive integer ð, we denote by ðŽð(ðŽ) the standardweighted Bergman space of ðŽ-valued functions ð analytic on the open unit disk ð»
and with finite norm â¥ðâ¥ðð(ðŽ):
ðð(ðŽ) ={ð(ð§) =
âðâ¥0
ððð§ð : â¥ðâ¥2ðð(ðŽ) :=
âðâ¥0
ðð,ð â¥ððâ¥2ðŽ <â}
(1.1)
where the weights ðð,ðâs are defined by
ðð,ð :=1(
ð+ðâ1ð
) = ð!(ðâ 1)!(ð + ðâ 1)! . (1.2)
It is clear from (1.1) that the spacesð1(ðŽ) andð2(ðŽ) are respectively the standardvector-valued Hardy space ð»2(ðŽ) and the unweighted Bergman space ðŽ2
2(ðŽ) ofthe unit disk. It then follows that ðð(ðŽ) is the reproducing kernel Hilbert spacewith reproducing kernel
ðð(ð§, ð) =1
(1 â ð§ð)ð ðŒðŽand that for ð > 1, the ðð(ðŽ)-norm equals
â¥ðâ¥2ðŽð(ðŽ) =â«ð»
(ðâ 1) â¥ð(ð§)â¥2ðŽ(1 â â£ð§â£2)ðâ2ððŽ(ð§) <â
where ððŽ is the planar Lebesgue measure normalized so that ðŽ(ð») = 1.
18 J.A. Ball and V. Bolotnikov
For Hilbert spaces ð° and ðŽ, we denote by â(ð° ,ðŽ) the space of boundedlinear operators mapping ð° into ðŽ (abbreviated to â(ð°) in case ð° = ðŽ) and wedefine the operator-valued Schur class ð®(ð° ,ðŽ) to be the class of analytic functionsð on ð» whose values ð(ð§) are contraction operators in â(ð° ,ðŽ). Each Schur-classfunction ð â ð®(ð° ,ðŽ) induces a contractive multiplication operator ðð : ð ï¿œâ ððfrom ðð(ð°) into ðð(ðŽ) for every ð ⥠1; in fact the Schur-class is exactly the classof contractive multipliers onðð(ðŽ) for each ð = 1, 2, . . . (see, e.g., [14, Proposition4.1]). The latter property is equivalent to the positivity of the kernels
ðŸð,ð(ð§, ð) =ðŒðŽ â ð(ð§)ð(ð)â
(1â ð§ð)ð (ð ⥠1) (1.3)
on ð» à ð» for all ð ⥠1. Thus the positivity of ðŸð,ð for some ð ⥠1 alreadyguarantees the membership ð â ð®(ð° ,ðŽ) so that ðŸð,ð is positive for all ð ⥠1.
Thus, with any function ð â ð®(ð° ,ðŽ), one can associate a family of positivekernels (1.3) and thereby a family of reproducing kernel Hilbert space â(ðŸð,ð).The spaces â(ðŸð,1) were introduced by de Branges and Rovnyak [11, 12] as a con-venient and natural setting for canonical operator models. Further developmentson de BrangesâRovnyak spaces can be found in [18]. The study of â(ðŸð,2) wasinitiated in [20, 21] and there are very few publications concerning the case whereð > 2 (see, e.g., [19]).
The general complementation theory applied to the contractive operatorðð :ðð(ð°) â ðð(ðŽ) (see, e.g., [18]) provides the characterization of â(ðŸð,ð) as the
operator range â(ðŸð,ð) = Ran(ðŒ âðððâð)
12 â ðð(ðŽ) with the lifted norm
â¥(ðŒ âðððâð)
12 ðâ¥2â(ðŸð,ð)
= â¥(ðŒ â ð)ðâ¥ðð(ðŽ) (1.4)
where ð is the orthogonal projection onto Ker(ðŒ â ðððâð)
12 . It follows that
â¥ðâ¥â(ðŸð,ð) ⥠â¥ðâ¥ðð(ðŽ) for every ð â â(ðŸð,ð) and thus the spaces â(ðŸð,ð) arecontractively included in ðð(ðŽ). For this reason, the spaces â(ðŸð,ð) are termedsub-Hardy (in [18] for ð = 1) and sub-Bergman (in [20], [21] for ð = 2). Upon
setting ð = (ðŒ âðððâð)
12 â in (1.4) we get
â¥(ðŒ âðððâð)ââ¥â(ðŸð,ð) = âš(ðŒ âððð
âð)â, ââ©ðð(ðŽ). (1.5)
The purpose of this paper is to study an interpolation problem of NevanlinnaâPicktype in the space â(ðŸð,ð). To formulate the problem we need several definitions.
A pair (ðž, ð ) consisting of operators ð â â(ð³ ) and ðž â â(ð³ ,ðŽ) is calledan output pair. An output pair (ðž, ð ) is called ð-output-stable if the associatedð-observability operator
ðªð,ðž,ð : ð¥ ï¿œâ ðž(ðŒ â ð§ð )âðð¥ =
ââð=0
(ðâ1ð,ððžððð¥) ð§ð (1.6)
Interpolation in Sub-Bergman Spaces 19
maps ð³ into ðð(ðŽ) and is bounded. For an ð-output stable pair (ðž, ð ), we definethe tangential functional calculus ð ï¿œâ (ðžâð)â§ð¿(ð â) on ðð(ðŽ) by
(ðžâð)â§ð¿(ð â) =ââð=0
ð âððžâðð if ð(ð§) =ââð=0
ððð§ð â ðð(ðŽ). (1.7)
The computationâš ââð=0
ð âððžâðð , ð¥
â©ð³=
ââð=0
âšðð, ðžð
ðð¥â©ðŽ
=
ââð=0
ðð,ð â âšðð, ð
â1ð,ððžð
ðð¥â©ðŽ = âšð, ðªð,ðž,ðð¥â©ðð(ðŽ)
shows that the ð-output stability of (ðž, ð ) is exactly what is needed to verifythat the infinite series in the definition (1.7) of (ðžâð)â§ð¿(ð â) converges in theweak topology on ð³ . The same computation shows that tangential evaluation withoperator argument amounts to the adjoint of ðªð,ðž,ð (in the metric of ðð(ðŽ)):
(ðžâð)â§ð¿(ð â) = ðªâð,ðž,ð ð for ð â ðð(ðŽ). (1.8)
Since â(ðŸð,ð) is included in ðð(ðŽ), evaluation (1.7) applies to functions inâ(ðŸð,ð). In this paper we study the following interpolation problem.
Problem 1.1. Given a Schur-class function ð â ð®(ð° ,ðŽ), given an ð-output stablepair (ðž, ð ) â â(ð³ ,ðŽ) à â(ð³ ) and given x â ð³ ,
(i) Find all ð â â(ðŸð,ð) such that
(ðžâð)â§ð¿(ð â) := ðªâð,ðž,ð ð = x. (1.9)
(ii) (norm-constrained version): Find all ð â â(ðŸð,ð) satisfying (1.9) withâ¥ðâ¥â(ðŸð,ð) †1.
The Hardy-space special case of this problem (where ð = 1 and â(ðŸð) :=â(ðŸð,1)) is the classical de BrangesâRovnyak space and has been studied by theauthors and S. ter Horst in [5, 6]. The set of all functions in ð»2(ðŽ) satisfying acondition of the form ðªâð,ðž,ð ð = 0 (i.e., condition (1.9) with ð = 0 and x = 0) is
one way to describe a generic shift-invariant subspace ofð»2(ðŽ) and the descriptionof the set of all solutions amounts to a calculation of the BeurlingâLax representerÎ(ð§) for the space â³ in terms of the interpolation data {ðž, ð, ð = 0,x = 0}.Another special case is to allow a general ð but still insist that ð = 0 and x =0; this special case recovers the BeurlingâLax representation theorem for shift-invariant subspaces â³ in the weighted Bergman space obtained by the authorsin [4].
20 J.A. Ball and V. Bolotnikov
2. Interpolation in reproducing kernel Hilbert spaces:A brief survey
The following operator interpolation problem with norm constraint is well knownin the literature: Given Hilbert space operators ðŽ â â(ðŽ,ð³ ) and ðµ â â(ð° ,ð³ ),describe all ð â â(ð° ,ðŽ) that satisfy the conditions
ðŽð = ðµ and â¥ð⥠†1. (2.1)
The solvability criterion is known as the Douglas factorization lemma [13]: Thereis an ð â â(ð° ,ðŽ) satisfying (2.1) if and only if ðŽðŽâ ⥠ðµðµâ. If this is the case,then (see, e.g., [6]) ð â â(ð° ,ðŽ) satisfies conditions (2.1) if and only if the operatorâ¡â£ðŒð° ðµâ ðâ
ðµ ðŽðŽâ ðŽð ðŽâ ðŒðŽ
â€âŠ :â¡â£ð°ð³ðŽ
â€âŠââ¡â£ð°ð³ðŽ
â€âŠ is positive semidefinite. (2.2)
On the other hand, if ðŽðŽâ ⥠ðµðµâ, then there exist (unique) contractions ð1 ââ(ð° ,RanðŽ) and ð2 â â(ðŽ,RanðŽ) such that(ðŽðŽâ)
12ð1 = ðµ, (ðŽðŽâ)
12ð2 = ðŽ, Kerð1 = Kerðµ, Kerð2 = KerðŽ. (2.3)
Applying Schur complement arguments to the positive semidefinite operator in(2.2) leads us to the following more explicit description of the set of all solutionsto the problem (2.1) (see, e.g., [6] for the proof).
Lemma 2.1. Let ðŽðŽâ ⥠ðµðµâ. Then an operator ð satisfies condition (2.2) (andtherefore, also conditions (2.1)) if and only if it is of the form
ð = ðâ2ð1 + (ðŒ âðâ2ð2)12ð(ðŒ âðâ1ð1)
12 (2.4)
where ð1 and ð2 are defined in (2.3) and where the parameter ð is an arbitrarycontraction from Ran(ðŒ âðâ1ð1) into Ran(ðŒ âðâ2ð2).
Remark 2.2. It follows from (2.4) that there is a unique ð subject to conditions(2.1) if and only if ð1 is isometric on ð° or ð2 is isometric on ðŽ. Furthermore, foreach ð in (2.4) and each ð¢ â ð° , we have
â¥ðð¢â¥2 = â¥ðâ2ð1ð¢â¥2 + â¥(ðŒ âðâ2ð2)12ð(ðŒ âðâ1ð1)
12ð¢â¥2,
so thatðâ2ð1 is the minimal norm solution to the problem (2.1) (see [6, Section 2]).
The left tangential NevanlinnaâPick interpolation problem for the reproduc-ing kernel Hilbert space â(ðŸ) can be formulated as follows. We are given vectorsðŠ1, . . . , ðŠð â ðŽ and points ð1, . . . , ðð â ð» along with numbers ð¥1, . . . , ð¥ð â â andseek ð â â(ðŸ) (possibly also with â¥ðâ¥â(ðŸ) †1) satisfying the left tangentialNevanlinnaâPick interpolation conditions
âšð(ðð), ðŠðâ© = ð¥ð for ð = 1, . . . , ð. (2.5)
This problem can be reformulated more abstractly as follows. We introduce theâ(âð,ðŽ)-valued function
ð§ ï¿œâ ð¹ (ð§) :=[ðŸ(ð§, ð1)ðŠ1 â â â ðŸ(ð§, ðð)ðŠð
]. (2.6)
Interpolation in Sub-Bergman Spaces 21
Then ð¹ induces a multiplication operator ðð¹ : âð â â(ðŸ)
ðð¹ :
[ð1...ðð
]â ð¹ (ð§)
[ð1...ðð
]= ð1ðŸ(ð§, ð1)ðŠ1 + â â â + ðððŸ(ð§, ðð)ðŠð.
Then a standard reproducing-kernel computation gives us that ðâð¹ : â(ðŸ)â âð
is given by
ðâð¹ : ð ï¿œâ
â¡â¢â£âšð(ð1), ðŠ1â©ðŽ...âšð(ðð), ðŠðâ©ðŽ
â€â¥âŠand the NevanlinnaâPick problem with interpolation conditions (2.5) can be re-formulated as follows: for given ð¹ as in (2.6) and x â âð, find ð â â(ðŸ) (possiblyalso with â¥ðâ¥â(ðŸ) †1) so that ðâ
ð¹ ð = x.
We now formulate our abstract left tangential NevanlinnaâPick interpolationproblem as follows. Let ðŸ(ð§, ð) be an â(ðŽ)-valued positive kernel on a Cartesianproduct set ΩÃΩ and letâ(ðŸ) be the associated reproducing kernel Hilbert space,that is, the unique inner product space of ðŽ-valued functions on Ω that containsthe functions ð§ ï¿œâ ðŸð(ð§) := ðŸ(ð§, ð)ðŠ for all fixed ð â Ω and ðŠ â ðŽ which in turnhave the reproducing property for â(ðŸ):
âšð, ðŸððŠâ©â(ðŸ) = âšð(ð), ðŠâ©ðŽ for all ð â â(ðŸ).For ð³ an auxiliary Hilbert space, we let â³(ð³ ,â(ðŸ)) (the space of multipliersfrom ð³ into â(ðŸ)) denote the space of â(ð³ ,ðŽ)-valued functions such that thefunction ð§ ï¿œâ ð¹ (ð§)ð¥ is in â(ðŸ) for each ð¥ â ð³ . A consequence of the closed-graphtheorem is that the multiplication operatorðð¹ : ð¥ ï¿œâ ð¹ (â )ð¥ is then bounded as anoperator from ð³ into â(ðŸ). With this notation in hand we can pose the followinginterpolation problem:
Problem 2.3. Given a positive kernel ðŸ along with ð¹ ââ³(ð³ ,â(ðŸ)) and x â ð³ ,
(i) Find all functions ð â â(ðŸ) satisfying
ðâð¹ð = x. (2.7)
(ii) Find all functions ð â â(ðŸ) satisfying (2.7) with â¥ðâ¥â(ðŸ) †1.
As a straightforward application of the general Hilbert space results in Lemma2.1, Remark 2.2 and the preceding discussion, we have the following solution ofProblem 2.3.
Proposition 2.4. Problem 2.3 (ii) has a solution if and only if
ð ⥠xxâ, where ð :=ðâð¹ðð¹ . (2.8)
Problem 2.3 (i) has a solution if and only if x â Ranð 12 .
22 J.A. Ball and V. Bolotnikov
Proof. By specializing the Douglas lemma to the case where
ðŽ =ðâð¹ : â(ðŸ)â ð³ and ðµ = x â ð³ âŒ= â(â,ð³ ), (2.9)
we see that solutions ð : ââ â(ðŸ) to problem (2.1) necessarily have the form ofa multiplication operator ðð for some function ð â â(ðŸ). This observation leadsus to (2.8). On the other hand, by the second equality in (2.8), Ranð
12 = Ranðâ
ð¹ .
Thus, x belongs to Ranð12 if and only if it belongs to Ranðâ
ð¹ , that is, if and onlyif equality ðâ
ð¹ ð = x holds for some ð â â(ðŸ), which means that this ð solvesProblem 2.3. â¡
Let us assume that the operator ð = ðâð¹ðð¹ is strictly positive definite.
Then the operator ðð¹ðâ 1
2 is an isometry and the space
ð© = {ð¹ (ð§)ð¥ : ð¥ â ð³} with norm â¥ð¹ð¥â¥â(ð) = â¥ð 12ð¥â¥ð³ (2.10)
is isometrically included in â(ðŸ). Moreover, the orthogonal complement of ð© in
â(ðŸ) is the reproducing kernel Hilbert space â(ᅵᅵ) with reproducing kernelᅵᅵ(ð§, ð) = ðŸ(ð§, ð)â ð¹ (ð§)ðâ1ð¹ (ð)â. (2.11)
The following theorem is an adaptation of Lemma 2.1 to the special case (2.9).
Theorem 2.5. Assume that condition (2.8) holds and that ð is strictly positive
definite. Let ᅵᅵ be the kernel defined in (2.11).
1. A function ð : Ωâ ðŽ solves Problem 2.3 (ii) if and only if it is of the form
ð(ð§) = ð¹ (ð§)ðâ1x+ â(ð§) (2.12)
for some function â â â(ᅵᅵ) subject to â¥ââ¥â(ᅵᅵ) â€â1â â¥ðâ 1
2xâ¥2.2. The representation (2.12) is orthogonal (in the metric of â(ðŸ)) so thatð¹ (ð§)ðâ1x is the minimal-norm solution of Problem 2.3.
3. Problem 2.3 (ii) has a unique solution if and only if
â¥ðâ 12x⥠= 1 or ᅵᅵ(ð§, ð) â¡ 0.
Proof. It is readily seen that
ð1 = ðâ 1
2x â ð³ âŒ= â(â,ð³ ) and ð2 = ðâ 1
2ðâð¹ â â(â(ðŸð),ð³ )
are the operators ð1 and ð2 from (2.4) after specialization to the case (2.9).
Statements (2) and (3) now follow from Remark 2.2, since ðâ12x â â(â,ð³ ) being
isometric means that â¥ðâ 12x⥠= 1 and, on the other hand, the isometric property
for the operatorðð¹ðâ 1
2 means that the space ð© defined in (2.10) is equal to the
whole space â(ðŸ). Thus â(ᅵᅵ) = â(ðŸ)âð© = {0} or ᅵᅵ â¡ 0.
In the present framework, the parametrization formula (2.4) takes the form
ðð =ðð¹ðâ 1
2x+
â1â â¥ðâ 1
2xâ¥2 â (ðŒ âðð¹ðâ1ðâ
ð¹ )12ð (2.13)
Interpolation in Sub-Bergman Spaces 23
where ð is equal to the operator of multiplication ðð : ââ â(ᅵᅵ) by a functionð â â(ᅵᅵ) with â¥ð⥠†1. Since ðð¹ð
â 12 is an isometry, the second term on the
right-hand side of (2.13) is equal to the operatorðâ of multiplication by a function
â â â(ᅵᅵ) such that â¥ââ¥â(ᅵᅵ) = â¥ââ¥â(ðŸ) â€â1â â¥ðâ 1
2xâ¥2. â¡
Remark 2.6. The parametrization formula (2.12) can be obtained in a more an-alytic way (still originating from (2.2)) as follows. Specializing (2.2) to the case(2.9) we conclude that a function ð solves Problem 2.3 (ii) if and only if
P =
â¡â£ 1 xâ ðâð
x ð ðâð¹
ðð ðð¹ ðŒâ(ðŸ)
â€âŠ ⥠0.
The latter condition is equivalent to the positivity on ΩÃΩ of the following kernel:
K(ð§, ð) =
â¡â£ 1 xâ ð(ð)â
x ð ð¹ (ð)â
ð(ð§) ð¹ (ð§) ðŸ(ð§, ð)
â€âŠ ર 0. (2.14)
This equivalence is justified by the fact that the set of all vectors of the form
ð(ð§) =
ðâð=1
â¡â£ ððð¥ð
ðŸ(â , ð§ð)ðŠð
â€âŠ (ðð â â, ðŠð â ðŽ, ð¥ð â ð³ , ð§ð â Ω)
is dense in ââð³ ââ(ðŸ) and since for every such vector,
âšPð, ðâ©ââð³ââ(ðŸð)
=
ðâð,â=1
âšK(ð§ð , ð§â)
[ ðâð¥âðŠâ
],[ ððð¥ððŠð
]â©ââð³âðŽ
.
If in addition, ð is strictly positive definite, one can take its Schur complement inK to get the inequality[
1â â¥ðâ 12xâ¥2 ð(ð)â â xâðâ1ð¹ (ð)â
ð(ð§)â ð¹ (ð§)ðâ1x ðŸ(ð§, ð)â ð¹ (ð§)ðâ1ð¹ (ð)â]ર 0
which is equivalent to (2.14). By [9, Theorem 2.2], the latter positivity is equivalent
to the membership of the function â(ð§) := ð(ð§) â ð¹ (ð§)ðâ1x in the space â(ᅵᅵ)together with the norm constraint â¥ââ¥2â(ᅵᅵ)
†1â â¥ðâ 12xâ¥.
Remark 2.7. The function â on the right-hand side of (2.12) is in fact the gen-eral solution of the homogeneous interpolation problem ðâ
ð¹â = 0. Thus the firstpart of Theorem 2.5 states that the solution set of the corresponding homogeneous
interpolation problem coincides with the reproducing kernel Hilbert space â(ᅵᅵ).The results of this sort hold true even in a more general setting of Hilbert modules[1, 10]. The most interesting part, however, is to get a more detailed parametriza-
tion ofâ(ᅵᅵ). Although Problem 2.3 is too general to get such a parametrization, inthe context of NevanlinnaâPick type Problem 1.1 in sub-Bergman spaces â(ðŸð,ð)we shall see that something can be done in this direction (see Theorem 3.8 below).
24 J.A. Ball and V. Bolotnikov
3. The main result
To apply the general results from Section 2 to sub-Bergman spaces â(ðŸð,ð) wemust show that Problem 1.1 is a particular case of Problem 2.3. In other words(see formulas (1.9) and (2.7)), we need to find an â(ð³ ,ðŽ)-valued function ð¹ (ð§)in the multiplier space â³(ð³ ,â(ðŸ)) such that the adjoint of the multiplicationoperator ðð¹ (in metric of â(ðŸð,ð)) is equal to the adjoint of the observabilityoperator ðªð,ðž,ð (in metric of ðð(ðŽ)). This function ð¹ will be constructed in Sec-tion 3.2 below. Then we will immediately get the parametrization formula (2.12)for the solution set of Problem 1.1 and then Theorem 3.8 (the main result of thepaper) will give a more detailed description of the solution set of the associatedhomogeneous problem; the result involves the construction of multiplier functionsÎð â â³(ðð(ðŽ)) and Îð â â³(ðð(ð³ ),ðð(ðŽ)) for ð = 1, 2, . . . , ð â 1. The addi-tional ingredients needed to arrive at this result are introduced in Sections 3.1â3.4below.
Before starting this program, we make one last observation. For an operatorðŽ : ð³ â â(ðŸð,ð) â ðð(ðŽ), the adjoint operator can be taken in the metric ofðð(ðŽ) as well as in the metric of â(ðŸð,ð) and these two operations in general arenot the same. To avoid confusion, in what follows we use the notation ðŽâ for theadjoint of ðŽ in the metric of ðð(ðŽ) and ðŽ[â] for the adjoint of ðŽ in the metric ofâ(ðŸð,ð). The precise value of ð (1 †ð †ð) occurring here will be clear from thecontext.
3.1. Operators ðµ and ð·ð
Recall that if the pair (ðž, ð ) is ð-output stable, then the ð-observability gramian
ð¢ð,ðž,ð := (ðªð,ðž,ð )âðªð,ðž,ð =
ââð=0
ðâ1ð,ð â ð âððžâðžð ð (3.1)
is bounded on ð³ and the strong convergence of the power series in its represen-tation (3.1) follows from the definition of the inner product in ðð(ðŽ). One mayconclude that (ðž, ð ) is ð-output stable if and only if the power series in (3.1) con-verges weakly (and therefore strongly, since all terms are positive semidefinite).The power series representation (3.1) suggests that ð¢ð,ðž,ð be defined for ð = 0 bysimply letting ð¢0,ðž,ð := ðž
âðž.
Lemma 3.1. If (ðž, ð ) is ð-output stable, then it is ð-output stable for all ð =0, . . . , ðâ 1 and the observability gramians satisfy the Stein identity
ð¢ð,ðž,ð â ð âð¢ð,ðž,ðð = ð¢ðâ1,ðž,ð for all ð = 1, . . . , ð. (3.2)
Proof. Since ðð,ð †ððâ1,ð (see (1.2)), we conclude that if the power series in(3.1) converges for some integer ð, it also converges for any positive integer ð < ð.Identity (3.2) follows from power series representations for ð¢ð,ðž,ð and ð¢ðâ1,ðž,ð ,
due to the binomial-coefficient identity(ðð
)=(ðâ1ð
)+(ðâ1ðâ1). â¡
Interpolation in Sub-Bergman Spaces 25
The evaluation map (1.7) extends to Schur-class functions ð â ð®(ð° ,ðŽ) by(ðžâð)â§ð¿(ð â) = ðªâð,ðž,ðððâ£ð° .
Lemma 3.2. Let (ðž, ð ) be ð-output stable, let ð â ð®(ð° ,ðŽ) be a Schur-class functionand let ð â â(ð° ,ð³ ) be defined by
ðâ = (ðžâð)â§ð¿(ð â) := ðªâð,ðž,ðððâ£ð° . (3.3)
Then the pair (ð, ð ) is ð-output stable and the following equality holds:
ðªâð,ðž,ððð = ðªâð,ð,ð : ðð(ð°)â ð³ . (3.4)
Furthermore, the operators
ðð := ð¢ð,ðž,ð â ð¢ð,ð,ð : ð³ â ð³ (3.5)
satisfy the Stein identities
ðð â ð âððð = ððâ1 for ð = 1, . . . , ð, (3.6)
as well as the inequalities
ðð ⥠ððâ1 ⥠0 for ð = 2, . . . , ð. (3.7)
Proof. Making use of power series representations
ð(ð§) =âðâ¥0
ððð§ð and â(ð§) =
âðâ¥0
âðð§ð
of a given ð â ð®(ð° ,ðŽ) and of an arbitrary fixed function â â ðð(ð°) we have
(ððâ)(ð§) =
âââ=0
ââ ðâð=0
ððâââð
ââ ð§â which together with (1.8) impliesðªâð,ðž,ðððâ = (ðžâ(ðâ))â§ð¿ (ð â) =
âââ=0
ð ââðžâ
ââ ââð=0
ððâââð
ââ , (3.8)
Note that the latter series converges weakly since the pair (ðž, ð ) is ð-outputstable and ðâ â ðð(ðŽ). If we regularize the series by replacing ðð by ðððð and byreplacing âð by ð
ðâð, we even get that the double series in (3.8), after taking theinner product against a fixed vector ð¥ â ð³ , converges absolutely. We may thenrearrange the series to have the form
âšðªâð,ðž,ððððâð, ð¥â© =ââ
ð,ð=0
âšðð+ð(ð â)ð+ððžâððâð, ð¥â©.
We may then invoke Abelâs theorem to take the limit as ð â 1 (justified by thefacts that (ðž, ð ) is ð-output stable and that ðâ â ðð(ðŽ)) to get
ðªâð,ðž,ðððâ = (ðžâðâ)â§ð¿(ð â) =ââ
ð,ð=0
(ð â)ð+ððžâððâð. (3.9)
26 J.A. Ball and V. Bolotnikov
On the other hand, due to (3.3),
ðªâð,ð,ðâ = (ðââ)â§ð¿(ð â) =ââð=0
ð âððââð =ââð=0
ð âððªâð,ðž,ððâð
=ââð=0
ð âð
ââ ââð=0
ð âððžâðð
ââ âð = ââð,ð=0
(ð â)ð+ððžâððâð
where all the series converge weakly, since that in (3.9) does. Since â was pickedarbitrarily in ðð(ð°), the last equality and (3.9) imply (3.4). Therefore, the op-erator ðªâð,ð,ð : ðð(ð°) â ð³ is bounded and hence the pair (ð, ð ) is ð-outputstable.
By Lemma 3.1, the pairs (ðž, ð ) and (ð, ð ) are ð-output stable for all ð =1, . . . , ð. By using the definition (3.5) of the operators ðð along with the identities(3.2), we get
ðð â ð âððð = ð¢ð,ðž,ð â ð¢ð,ð,ð â ð â (ð¢ð,ðž,ð â ð¢ð,ð,ð )ð
= (ð¢ð,ðž,ð â ð âð¢ð,ðž,ðð )â (ð¢ð,ð,ð â ð âð¢ð,ð,ðð )
= ð¢ðâ1,ðž,ð â ð¢ðâ1,ð,ð =: ððâ1
and we arrive at (3.6). Since ðð is a contraction on ðð(ðŽ), we may replace ð byð in the preceding proof to conclude that
ðªâð,ðž,ððð = ðªâð,ð,ð : ðð(ð°)â ð³ for ð = 1, . . . , ð. (3.10)
Therefore, from (3.5) and (3.10) we see that
ðð = ðªâð,ðž,ððªð,ðž,ð âðªâð,ð,ððªð,ð,ð = ðªâð,ðž,ð (ðŒ âðððâð)ðªð,ðž,ð ⥠0.
Finally, we invoke (3.6) to conclude that ðð = ð âððð + ððâ1 ⥠ððâ1 whichcompletes the proof of (3.7). â¡
3.2. Functions ððºð
We now assume that we are given the data set (ðž, ð, ð(ð§),x) for an interpolationproblem as in Problem 1.1. By making use of the operator ð defined by (3.3), wenow introduce â(ð³ ,ðŽ)-valued functions
ð¹ðð (ð§) = (ðž â ð(ð§)ð)(ðŒ â ð§ð )âð, ð = 1, . . . , ð, (3.11)
which are therefore completely specified by the data set (ðž,ð, ð, ð(ð§),x). For themultiplication operator ðð¹ð
ðwe have
ðð¹ðð= ðªð,ðž,ð âðððªð,ð,ð = (ðŒ âððð
âð)ðªð,ðž,ð (3.12)
where the first equality follows from (3.11) and (1.6), while the second equality isa consequence of (3.10).
Lemma 3.3. Let (ðž, ð ) be an ð-output stable pair, let ð â ð®(ð° ,ðŽ) be a Schur-classfunction and let ð and ðð be defined as in (3.3) and (3.5) respectively. Then
Interpolation in Sub-Bergman Spaces 27
1. The function ð¹ðð given by (3.11) is in the space of multipliersâ³(ð³ ,â(ðŸð,ð)),
and moreover
ð[â]ð¹ð
ð
ðð¹ðð= ðð and ð
[â]ð¹ð
ð
= ðªâð,ðž,ð . (3.13)
2. The kernel ð ð(ð§, ð) =
[ðð ð¹ð
ð (ð)â
ð¹ðð (ð§) ðŸð,ð(ð§, ð)
]is positive on ð»Ã ð».
Proof. Formula (3.12) and the range characterization of â(ðŸð,ð) imply that ðð¹ðð
maps ð³ into â(ðŸð,ð). Furthermore, it follows from (3.11), (1.5), (3.4), and (3.5)that
â¥ð¹ðð ð¥â¥2â(ðŸð,ð)
= âš(ðŒ âðððâð)ðªð,ðž,ðð¥,ðªð,ðž,ðð¥â©ðð(ðŽ)
= âš(ðªâð,ðž,ððªð,ðž,ð âðªâð,ð,ððªð,ð,ð )ð¥, ð¥â©ð³ = âšððð¥, ð¥â©ð³for every ð¥ â ð³ which implies the first relation in (3.13). On the other hand, uponmaking subsequent use of (3.12) and (1.5), we see that
âšð¥,ð [â]ð¹ð
ð
ðâ©ð³ = âšð¹ðð ð¥, ðâ©â(ðŸð,ð) = âš(ðŒ âððð
âð)ðªð,ðž,ðð¥, ðâ©â(ðŸð,ð)
= âšðªð,ðž,ðð¥, ðâ©ðð(ðŽ) = âšð¥, ðªâð,ðž,ð ðâ©ð³for every ð â â(ðŸð,ð) and ð¥ â ð³ , which implies the second equality in (3.13). Bythe first equality in (3.13), the operator[
ðð ð[â]ð¹ð
ð
ðð¹ðð
ðŒ
]=
[ð
[â]ð¹ð
ð
ðð¹ðð
ð[â]ð¹ð
ð
ðð¹ðð
ðŒ
]=
[ð
[â]ð¹ð
ð
ðŒ
] [ðð¹ð
ððŒ] â â([ ð³
â(ðŸð,ð)
])is positive semidefinite; as in Remark 2.6, this condition in turn is equivalent topositivity of the kernel ð ð(ð§, ð) on ð»Ã ð». â¡
We now conclude from (3.13) that the interpolation condition (1.9) in Prob-lem 1.1 can be written in the form ðâ
ð¹ððð = x. Therefore Proposition 1.1 and
Theorem 2.5 apply leading us to the following result.
Lemma 3.4. Let ð and ðð be defined as in (3.3) and (3.5) respectively.
1. Problem 1.1 (i) has a solution if and only if x â Ranð 12ð .
2. If ðð is strictly positive definite, the function ðmin(ð§) = ð¹ðð (ð§)ð
â1ð x solves
Problem 1.1 (i) and has the minimal possible norm â¥ðminâ¥â(ðŸð,ð) = â¥ðâ12
ð xâ¥.3. A function â â â(ðŸð,ð) satisfies (ðžââ)â§ð¿(ð â) = 0 if and only if â is in the
reproducing kernel Hilbert space â(ᅵᅵð,ð) with reproducing kernel
ᅵᅵð,ð(ð§, ð) =ðŒðŽ â ð(ð§)ð(ð)â
(1â ð§ð)ð â ð¹ðð (ð§)ð
â1ð ð¹ð
ð (ð)â. (3.14)
28 J.A. Ball and V. Bolotnikov
3.3. ð±-inner function Î and the Schur-class function íThe assumption ðð > 0 allowed us to get a simple explicit formula for the minimal-
norm solution in Lemma 3.4. The parametrization of the space â(ᅵᅵð,ð) will beestablished in Theorem 3.8 below under the stronger assumption ð1 > 0. By (3.7)this condition implies that ðð > 0 for all ð = 1, . . . , ð.
Let us observe that if ð1 is boundedly invertible, then so is the observabilitygramian ð¢1,ðž,ð (see formula (3.5)) which in turn implies (see, e.g., [8]) that theoperator ð is strongly stable in the sense that ð ð converge to zero in the strongoperator topology. Let ðœ be the signature operator given by
ðœ =
[ðŒðŽ 00 âðŒð°
]and let Î(ð§) =
[ð(ð§) ð(ð§)ð(ð§) ð(ð§)
](3.15)
be a â(ðŽ â ð°)-valued function such that for all ð§, ð â ð»,
ðœ âÎ(ð§)ðœÎ(ð)â1â ð§ð =
[ðžð
](ðŒ â ð§ð )â1ðâ11 (ðŒ â ðð â)â1 [ðžâ ðâ
]. (3.16)
The function Î is determined by equality (3.16) uniquely up to a constant ðœ-unitary factor on the right. One possible choice of Î satisfying (3.16) is
Î(ð§) = ð· + ð§
[ðžð
](ðŒ â ð§ð )â1ðµ
where the operator [ ðµð· ] : ðŽâð° â[ ð³ðŽâð°
]is an injective solution to the ðœ-Cholesky
factorization problem[ðµð·
]ðœ[ðµâ ð·â
]=
[ðâ11 00 ðœ
]ââ¡â£ððžð
â€âŠðâ11
[ð â ðžâ ðâ
].
Such a solution exists due to identity (3.2) for ð = 1, that is,
ð1 â ð âð1ð = ðžâðž âðâð. (3.17)
If spec(ð ) â© ð â= ð (which is the case if, e.g., dimð³ < â), then a function Îsatisfying (3.16) can be taken in the form
Î(ð§) = ðŒ + (ð§ â ð)[ðžð
](ðŒð³ â ð§ð )â1ðâ11 (ððŒð³ â ð â)â1
[ðžâ âðâ ] (3.18)
where ð is an arbitrary point in ð â spec(ð ) (see [7]). For Î of the form (3.18),the verification of identity (3.16) is straightforward and relies on the Stein iden-tity (3.17) only. It follows from (3.16) that Î is ðœ-contractive on ð», i.e., thatÎ(ð§)ðœÎ(ð§)â †ðœ for all ð§ â ð». A less trivial fact is that due to the strong sta-bility of ð , the function Î is ðœ-inner; that is, the nontangential boundary valuesÎ(ð¡) exist for almost all ð¡ â ð and are ðœ-unitary: Î(ð¡)ðœÎ(ð¡)â = ðœ . Every ðœ-innerfunction Î =
[ð ðð ð
]gives rise to the one-to-one linear fractional transform
â° ï¿œâ TÎ[â° ] := (ðâ° + ð)(ðâ° + ð)â1 (3.19)
Interpolation in Sub-Bergman Spaces 29
mapping the Schur class ð®(ð° ,ðŽ) into itself. In case Î is a ðœ-inner functionsatisfying identity (3.16), the transform (3.19) establishes a one-to-one corre-spondence between ð®(ð° ,ðŽ) and the set of all Schur class functions ðº such that(ðžâðº)â§ð¿(ð â) = ðâ. Since the given function ð satisfies the latter condition bydefinition (3.3) of ð , it follows that ð = TÎ[â° ] for some (uniquely determined)function â° â ð®(ð° ,ðŽ) which is recovered from ð by
â° = (ðâ ðð)â1(ððâ ð). (3.20)
Since Î is ðœ-inner, it follows that ð(ð§) is boundedly invertible for all ð§ â ð» andalso â¥ðâ1(ð§)ð(ð§)⥠< 1 for all ð§ â ð».
The following result gives the construction of the multiplier Ίð needed for theMain Result (Theorem 3.8 below); the remaining multipliers Ίð (ð = 1, . . . , ðâ 1)are obtained in Lemma 3.7 below (formula (3.30)).
Lemma 3.5. Let us assume that Schur-class functions ð® and â° are related as in(3.20). Then the following identity holds for all ð§ â ð»:
ðŒðŽ â ð(ð§)ð(ð)â1â ð§ð â ð¹ð
1 (ð§)ðâ11 ð¹ð
1 (ð)â = Ίð(ð§)
ðŒðŽ â â°(ð§)â°(ð)â1â ð§ð Ίð(ð)
â, (3.21)
where ð¹ð1 (ð§) = (ðž â ð(ð§)ð)(ðŒ â ð§ð )â1 (according to formula (3.11)) and where
Ίð = (ðâ ððâ1ð)(ðŒ + â°ðâ1ð)â1. (3.22)
Proof. Substituting (3.20) into (3.22) gives
Ίð = (ðâ ððâ1ð)(ðŒ + (ðâ ðð)â1(ððâ ð)ðâ1ð)â1
= (ðâ ððâ1ð) (ðâ ðð+ (ððâ ð)ðâ1ð)â1 (ðâ ðð) = ðâ ðð.Recalling the matrix representation for Î in (3.15) and the expression (3.19) forâ° , we therefore have[
ðŒðŽ âð]Î =[ðâ ðð ðâ ðð] = Ίð
[ðŒðŽ ââ°] . (3.23)
Multiplying both sides of (3.16) by[ðŒðŽ âð(ð§)] on the left and by [ ðŒðŽ
âð(ð)â]on
the right we get[ðŒðŽ âð(ð§)] ðœ âÎ(ð§)ðœÎ(ð)â
1â ð§ð
[ðŒðŽ
âð(ð)â]= ð¹ð
1 (ð§)ðâ11 ð¹ð
1 (ð)â
which can be rearranged (see the formula (3.15) for ðœ) to
ðŒðŽ â ð(ð§)ð(ð)â1â ð§ð â ð¹ð
1 (ð§)ðâ11 ð¹ð
1 (ð)â =
[ðŒðŽ âð(ð§)] Î(ð§)ðœÎ(ð)â
1â ð§ð
[ðŒðŽ
âð(ð)â]
which together with (3.23) implies (3.21). â¡
In case Î =[ð ðð ð
]is taken in the form (3.18), the formula (3.22) takes the
form
Ίð(ð§) = (ðŒ + (ð§ â ð)ðžÎ(ð§)â1ðžâ)(ðŒ + (ð§ â ð)â°(ð§)ðÎ(ð§)â1ðžâ)â1 (3.24)
30 J.A. Ball and V. Bolotnikov
where
Î(ð§) = (ððŒ â ð â)ð1(ðŒ â ð§ð )â (ð§ â ð)ðâð.Indeed, let us first note that(
ðŒ â (ð§ â ð)ð(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðâ)â1= ðŒ + (ð§ â ð) (ðŒ â (ð§ â ð)ð(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðâ)â1Ãð(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðâ
= ðŒ + (ð§ â ð)ð (ðŒ â (ð§ â ð)(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðâð)â1à (ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðâ
= ðŒ + (ð§ â ð)ðÎ(ð§)â1ðâ.With this result in hand we see next that
ð(ð§)â1ð(ð§) = (ð§ â ð) (ðŒ â (ð§ â ð)ð(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðâ)â1Ãð(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðžâ
= (ð§ â ð) (ðŒ + (ð§ â ð)ðÎ(ð§)â1ðâ)Ãð(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðžâ
= (ð§ â ð)ðÎ(ð§)â1 (Î(ð§) + (ð§ â ð)ðâð)à (ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1ðžâ
= (ð§ â ð)ðÎ(ð§)â1ðžâ.Therefore we get
ð(ð§)â ð(ð§)ð(ð§)â1ð(ð§) = ðŒ + (ð§ â ð)ðž(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1à (ðžâ + (ð§ â ð)ðâðÎ(ð§)â1ðžâ)
= ðŒ + (ð§ â ð)ðž(ðŒ â ð§ð )â1ðâ11 (ððŒ â ð â)â1à (Î(ð§) + (ð§ â ð)ðâð)Î(ð§)â1ðžâ
= ðŒ + (ð§ â ð)ðžÎ(ð§)â1ðžâand (3.24) follows from the two latter equalities and (3.22).
3.4. Inner functions Κð
Our next construction is similar to that in the previous section. Due to identities(3.6) and (strict) inequalities (3.7), we can find operators Bð , Dð : ð³ â ð³ so that[
ð Bð
ð12ð Dð
][ðâ1ð+1 0
0 ðŒð³
][ð â ð
12ð
Bâð Dâð
]=
[ðâ1ð+1 0
0 ðŒð³
]and [
ð â ð12ð
Bâð Dâð
][ðð+1 00 ðŒð³
][ð Bð
ð12
ð Dð
]=
[ðð+1 00 ðŒð³
].
Interpolation in Sub-Bergman Spaces 31
In fact, the latter equalities determineBð andDð uniquely up to a common unitaryfactor on the right:
Bð = (ðâ1ð+1 âðŽðâ1ð+1ðŽâ)
12 , Dð = âðâ
12
ð ðŽâððBð .
We now define the functions
Κð(ð§) = Dð + ð§ð12
ð (ðŒ â ð§ð )â1Bð (3.25)
for ð = 1, . . . , ðâ 1, which are inner and satisfy the identitiesðŒð³ âΚð(ð§)Κð(ð)
â
1â ð§ð = ð12
ð (ðŒ â ð§ð )â1ðâ1ð+1(ðŒ â ðð â)â1ð12
ð . (3.26)
As was shown in [4, Proposition 7.2], the function Κð in (3.25) for ð = 1, . . . , ðâ 1can alternatively be given by
Κð(ð§) = ð12
ð (ðŒ â ð§ð )â1ðâ1ð+1(ð§ðŒ â ð â)Bâ1ð .
If spec(ð ) â© ð â= ð, then a function Κð satisfying (3.26) can be taken in the form
Κð(ð§) = ðŒ + (ð§ â ð)ð 12
ð (ðŒð³ â ð§ð )â1ðâ1ð+1(ððŒð³ â ð â)â1ð12
ð (3.27)
where ð is an arbitrary point in ðâspec(ð ). For this choice, an alternative formulafor Κð is the following
Κð(ð§) = ð12ð (ðŒ â ð§ð )â1ðâ1ð+1(ð§ðŒ â ð â)(ððŒ â ð â)â1ðð+1(ðŒ â ðð )â1ðâ
12
ð .
Lemma 3.6. Let Κ1, . . . ,Κðâ1 be the inner functions defined in (3.25). Then
ðâ1âð=1
(ðŒ â ð§ð )âð+1ðâ 1
2
ð
Κð(ð§)Κð(ð)â
(1â ð§ð)ðâððâ 1
2
ð (ðŒ â ðð â)âð+1
=ðâ11
(1â ð§ð)ðâ1 â (ðŒ â ð§ð )âð+1ðâ1ð (ðŒ â ðð â)âð+1.
(3.28)
Proof. Equality
(ðŒ â ð§ð )âð+1ðâ 1
2
ð
ðŒð³ âΚð(ð§)Κð(ð)â
1â ð§ð ðâ 1
2
ð (ðŒ â ðð â)âð+1
= (ðŒ â ð§ð )âððâ1ð+1(ðŒ â ðð â)âð
follows immediately from (3.26) and can be written equivalently as
(ðŒ â ð§ð )âð+1ðâ 1
2
ð
Κð(ð§)Κð(ð)â
(1â ð§ð)ðâððâ 1
2
ð (ðŒ â ðð â)âð+1
= (ðŒ â ð§ð )âð+1 ðâ1ð
(1 â ð§ð)ðâð(ðŒ â ðð â)âð+1
â (ðŒ â ð§ð )âððâ1ð+1
(1â ð§ð)ðâðâ1 (ðŒ â ðð â)âð.
Summing up the latter equalities for ð = 1, . . . , ðâ 1 we get (3.28). â¡
32 J.A. Ball and V. Bolotnikov
3.5. The main result
Now we are in a position to represent the kernel (3.14) as the sum of ð positivekernels.
Lemma 3.7. The kernel ᅵᅵð,ð(ð§, ð) defined in (3.14) can be represented as
ᅵᅵð,ð(ð§, ð) = Ίð(ð§)ðŸâ°,ð(ð§, ð)Ίð(ð)â +
ðâ1âð=1
Ίð(ð§)Ίð(ð)â
(1â ð§ð)ð (3.29)
where Ίð is defined in (3.24) and where
Ίð(ð§) = ð¹ððâð(ð§)ð
â 12
ðâðΚðâð(ð§) (ð = 1, . . . , ðâ 1). (3.30)
Proof. We first divide both sides in (3.21) by (1âð§ð)ðâ1 and combine the obtainedidentity with (3.14) to get
ᅵᅵð,ð(ð§, ð) = Ίð(ð§)ðŒðŽ â â°(ð§)â°(ð)â(1â ð§ð)ð Ίð(ð)
â
+ð¹ð1 (ð§)ð
â11 ð¹ð
1 (ð)â
(1â ð§ð)ðâ1 â ð¹ðð (ð§)ð
â1ð ð¹ð
ð (ð)â. (3.31)
Multiplying both sides in (3.28) by ð¹ð1 (ð§) = (ðž â ð(ð§)ð)(ðŒ â ð§ð )â1 on the left
and by ð¹ð1 (ð)
â on the right we getðâ1âð=1
ð¹ðð (ð§)ð
â 12
ð
Κð(ð§)Κð(ð)â
(1â ð§ð)ðâððâ 1
2
ð ð¹ðð (ð)
â
=ð¹ð1 (ð§)ð
â11 ð¹ð
1 (ð)â
(1 â ð§ð)ðâ1 â ð¹ðð (ð§)ð
â1ð ð¹ð
ð (ð)â
which being plugged into (3.31) gives the desired representation (3.29). â¡Identity (3.29) means that the map
ð : ᅵᅵð,ð(â , ð)ðŠ ï¿œâ
â¡â¢â¢â¢â¢â£ðŸâ°,ð(â , ð)Ίð(ð)
âðŠ1
(1ââ ð)ðâ1Ίðâ1(ð)âðŠ...
1(1ââ ð)1Ί1(ð)
âðŠ
â€â¥â¥â¥â¥âŠextends by linearity and continuity to an isometry from â(ᅵᅵð,ð) to â(ðŸâ°,ð) âððâ1(ð³ )ââ â â âð1(ð³ ). Furthermore, standard reproducing-kernel arguments showthat ðâ is given by multiplication by the function Ί :=
[Ίð â â â Ί1
]. We con-
clude that the meaning of the identity (3.29) is that the function Ί is an coisometric
multiplier from â(ðŸâ°,ð)âððâ1(ð³ )â â â â âð1(ð³ ) onto â(ᅵᅵð,ð). Combining thisobservation with statement (3) in Lemma 3.4 we arrive at the following result.
Theorem 3.8. Let us assume that the data set of Problem 1.1 is such that theoperator ð1 defined in (3.5) is strictly positive definite. Let Ίð be defined as in(3.22), (3.30). Then:
Interpolation in Sub-Bergman Spaces 33
1. A function ð is a solution of Problem 1.1 (i) if and only if it is of the form
ð = ðmin +Ίðâð +ðâ1âð=1
Ίðâð , ðmin(ð§) = ð¹ðð (ð§)ð
â1ð x (3.32)
for some choice of âð â â(ðŸâ°,ð) and âð â ðð(ð³ ) for ð = 1, . . . , ðâ 1.
2. A function ð is a solution of Problem 1.1 (ii) if and only if it is of the form(3.32) for some âð â â(ðŸâ°,ð) and âð â ðð(ð³ ) for ð = 1, . . . , ðâ 1 such that
â¥âðâ¥2â(ðŸð,ð)+
ðâ1âð=1
â¥âðâ¥2ðð(ð³ ) †1â â¥ðâ 12xâ¥2.
In particular, such solutions exist if and only if â¥ðâ 12xâ¥2 †1.
If ð â¡ 0, then the space â(ðŸð,ð) amounts to the Bergman space ðð(ðŽ).In this case, we get from (3.3) that ð = 0 and subsequently, ðð = ð¢ð,ðž,ð forð = 1, . . . , ð. We then have equality (3.23) with â° â¡ 0 and Ίð and inner â(ðŽ)-valued function subject to
ðŒðŽ âΚð(ð§)Κð(ð)â
1â ð§ð = ðž(ðŒ â ð§ð )â1ð¢â11,ðž,ð (ðŒ â ðð â)â1ðžâ.
In the present context, Theorem 3.8 parametrizes the solution set of the interpola-tion problem with operator argument in ðð(ðŽ). The homogeneous version of thisproblem can be interpreted as a Beurling-type theorem for ðð(ðŽ).
Another particular case where ð = 1 partly recovers results on interpolationin de BrangesâRovnyak spaces presented in [6].
4. Examples
Parametrization (3.32) is especially transparent in case dimð° = dimðŽ = 1 anddimð³ < â. Besides, in this case (as we will see below), the matrices ðð areall positive definite for ð = 2, . . . , ð. The matrix ð1 may be singular but in thescalar-valued case we are able to handle this option as well.
If dimð³ < â and dimð° = dimðŽ = 1, then with respect to an appropriatebasis of ð³ the output stable observable pair (ðž, ð ) has the following form: ð is ablock diagonal matrix ð = diag{ð1, . . . , ðð} with the diagonal block ðð equal tothe upper triangular ðð à ðð Jordan block with the number ð§ð â ð» on the maindiagonal and ðž is the row vector
ðž =[ðž1 . . . ðžð
], where ðžð =
[1 0 . . . 0
] â â1Ãðð .
It is not hard to show that for (ðž, ð ) as above and for every function ð analyticat ð§1, . . . , ð§ð, evaluation (1.7) amounts to
(ðžâð)â§ð¿(ð â) = Col1â€ðâ€ð Col0â€ð<ðð
ð (ð)(ð§ð)
ð!. (4.1)
34 J.A. Ball and V. Bolotnikov
If we specify the entries of the column xâ by letting
xâ = Col1â€ðâ€ð Col0â€ð<ðð ð¥ðð ,
then it is readily seen that Problem 1.1 amounts to the following LagrangeâSylvester interpolation problem:
LSP: Given a scalar Schur-class function ð â ð®, distinct points ð§1, . . . , ð§ð â ð»
and a collection {ð¥ðð} of complex numbers, find all functions ð â â(ðŸð,ð) suchthat
ð (ð)(ð§ð)/ð! = ð¥ðð for ð = 0, . . . , ðð â 1; ð = 1, . . . , ð.
The auxiliary column ðâ is now defined from the derivatives of the givenfunction ð via formula (4.1), and we define the matrix ð1 as the unique solutionof the Stein equation (3.17). This matrix ð1 turns out to be equal to the SchwarzâPick matrix
ð1 =
â¡â¢â£â¡â£ 1
â!ð!
ââ+ð
âð§ââðð1â ð(ð§)ð(ð)
1â ð§ð
â£â£â£â£â£ð§=ð§ðð=ð§ð
â€âŠð=0,...,ððâ1
â=0,...,ððâ1
â€â¥âŠð
ð,ð=1
,
which in turn is known to be positive definite unless ð is a Blaschke productof degree ð < n := ð1 + â â â + ðð, in which case ð1 is positive semidefinite andrankð1 = ð. It is not hard to show that the matrices ðð defined in (3.5) are equalto higher-order SchwarzâPick matrices
ðð =
â¡â¢â£â¡â£ 1
â!ð!
ââ+ð
âð§ââðð1â ð(ð§)ð(ð)(1â ð§ð)ð
â£â£â£â£â£ð§=ð§ðð=ð§ð
â€âŠð=0,...,ððâ1
â=0,...,ððâ1
â€â¥âŠð
ð,ð=1
.
Lemma 4.1. If a Schur-class function is not a unimodular constant, then the ma-trices ðð are positive definite for ð ⥠2.
Proof. By (3.7), it suffices to prove the statement for ð = 2. Observe that ifð¶ð â â1Ãðð is a row-vector with a non-zero left-most entry and ðð â âððÃðð isthe upper triangular Jordan block, then the pair (ð¶ð, ðð) is observable in the sensethat the gramian ð¢1,ð¶ð,ðð is positive definite. Consequently, if ð¶ =
[ð¶1 â â â ð¶ð
]is a block-row vector with all blocks ð¶ð having nonzero left-most entries and ifð = diag{ð1, . . . , ðð} is the conformally decomposed block-diagonal matrix asabove, then the pair (ð¶, ð ) is observable. In particular we may take ð¶ equal tothe top row in the matrix ð1. The left-most entries in its blocks are equal to1âð(ð§1)ð(ð§ð)
1âð§1ð§ðand are non-zero unless ð is a unimodular constant. Thus, for this
choice of ð¶ we have
0 < ð¢1,ð¶,ð =
ââð=0
ð âðð¶âð¶ð ð â€ââð=0
ð âðð1ð ð. (4.2)
Interpolation in Sub-Bergman Spaces 35
Since the spectral radius of ð is strictly less than one, it follows from the Steinidentity ð2 â ð âð2ð = ð1 that ð2 can be represented via converging series
ð2 =
ââð=0
ð âðð1ð ð
and now we conclude from (4.2) that ð2 > 0 regardless of whether ð1 is invertibleor not. â¡
We now proceed to three different cases.
Case 1: ð is not a finite Blaschke product or it is a finite Blaschke product ofdegree deg ð > n. In this case ð1 > 0 and all the solutions ð to the problem LSPare given by formula (3.32), where now all the ingredients are not only explicitbut also computable.
Case 2: ð be a Blaschke product of degree deg ð = n. Then the matrix ð1 isinvertible, but the associated function â° defined by (3.19) is a unimodular constant,so that the corresponding sub-Bergman space â(ðŸâ°,ð) is trivial. Observe, that inthis case, the formula (3.21) takes the form
ðŒðŽ â ð(ð§)ð(ð)â1â ð§ð = ð¹ð
1 (ð§)ðâ11 ð¹ð
1 (ð)â. (4.3)
Furthermore, the parametrizing formula (3.32) does not contain the second termon the right. In particular, if ð = 1, then Problem 1.1(i) has a unique solution.
Case 3: deg ð < n. Since the matrices ðð are invertible for ð = 2, . . . , ð, theformula (3.32) for ðmin as well as formulas (3.30) for Ίð for ð = 1, . . . , ð â 2 stillmake sense and by the preceding analysis,
ð¹ð2 (ð§)ð
â12 ð¹ð
2 (ð)â
(1â ð§ð)ðâ1 â ð¹ðð (ð§)ð
â1ð ð¹ð
ð (ð)â =
ðâ2âð=1
Ίð(ð§)Ίð(ð)â
(1 â ð§ð)ð . (4.4)
The only remaining question is to modify appropriately the function Ίðâ1.
Let us consider conformal block decompositions
ð1 =
[ð1 ððâ ð
], ð =
[ð ð10 ð2
], ðž =
[ðž ðž1
], ð =
[ð ð1
](4.5)
with ð1, ð â âðÃð and ðž, ð â â1Ãð where ð := deg ð. Making use of (4.5) andtaking the advantage of the upper triangular structure of ð , we also decomposethe function ð¹ð
1 as
ð¹ð1 (ð§) =
[ð¹ð1 (ð§) ð¹ð
1 (ð§)], where ð¹ð
1 (ð§) = (ðž â ð(ð§)ð)(ðŒ â ð§ð )â1. (4.6)
The block ð1 is the SchwarzâPick matrix of a Blaschke product of degree ð based
on ð points (counted with multiplicities) and therefore, ð1 is invertible. On the
36 J.A. Ball and V. Bolotnikov
other hand the subproblem of LSP based on these points is of the type consideredin Case 2 and thus, by virtue of (4.3) we have
ðŒðŽ â ð(ð§)ð(ð)â1â ð§ð = ð¹ð
1 (ð§)ðâ11 ð¹ð
1 (ð)â. (4.7)
Since rankð1 = rankð1 = ð it follows that ð = ðâðâ11 ð so that ð1 can berepresented as
ð1 =
[ð
121
ðâðâ12
1
] [ð
121 ð
â 12
1 ð
].
From this representation and from the fact that the kernel ð 1(ð§, ð) is positive onð»Ã ð» (see statement (2) in Lemma 3.3) we conclude that
ð¹ð1 (ð§)
[ðâ11 ðâðŒ
]â¡ 0
which implies that the entries ð¹ð1 and ð¹ð
1 in (4.6) are related by ð¹ð1 = ð¹ð
1 ðâ11 ð
so that ð¹ð1 (ð§) can be written as
ð¹ð1 (ð§) = ð¹
ð1 (ð§)
[ðŒ ðâ11 ð
]. (4.8)
We now define the function
Ίðâ1(ð§) := ð¹ð1 (ð§)ð
â 12
1 Κ1(ð§) (4.9)
where Κ1 is the inner âðÃð-valued function given by
Κ1(ð§) = ðŒ + (ð§ â ð)ð 121
[ðŒ ðâ11 ð
](ðŒ â ð§ð )â1ðâ12 (ððŒ â ð â)â1
[ðŒ
ðâðâ11
]ð
121
(compare with (3.27)) and satisfying the identity
ðŒð â Κ1(ð§)Κ1(ð)â
1â ð§ð = ð121
[ðŒ ðâ11 ð
](ðŒ â ð§ð )â1ðâ12 (ðŒ â ðð â)â1
[ðŒ
ðâðâ11
]ð
121
similar to that in (3.26). By (4.8),
ð¹ð2 (ð§) = ð¹
ð1 (ð§)(ðŒ â ð§ð )â1 = ð¹ð
1 (ð§)[ðŒ ðâ11 ð
](ðŒ â ð§ð )â1,
which together with (4.7), (4.9) and the previous identity implies
ð¹ð2 (ð§)ð
â12 ð¹ð
2 (ð)â = ð¹ð
1 (ð§)ðâ 1
21
ðŒð â Κ1(ð§)Κ1(ð)â
1â ð§ð ðâ 1
21 ð¹ð
1 (ð)â
=ðŒðŽ â ð(ð§)ð(ð)â
1â ð§ð â Ίðâ1(ð§)Ίðâ1(ð)â
1â ð§ð .
Therefore the kernel (3.14) can be written as
ᅵᅵð,ð(ð§, ð) =Ίðâ1(ð§)Ίðâ1(ð)â
(1â ð§ð)ð +ð¹ð2 (ð§)ð
â12 ð¹ð
2 (ð)â
(1â ð§ð)ðâ1 â ð¹ðð (ð§)ð
â1ð ð¹ð
ð (ð)â
Interpolation in Sub-Bergman Spaces 37
which together with (4.4) implies
ᅵᅵð,ð(ð§, ð) =
ðâ1âð=1
Ίð(ð§)Ίð(ð)â
(1â ð§ð)ð .
Thus in the present case we have the same parametrization of the solution set butthe parameter âðâ1 is taken in ððâ1(âð) rather than in ððâ1(ân).
5. Some open questions and directions for future work
5.1. Stein equations, inertia theorems, and associated orthogonal polynomials
In the classical setting, Stein equations and associated inertial theorems (wherethe solution of the Stein equation may be indefinite) are closely associated withorthogonal polynomials with respect to an appropriate weight on the unit circleand location of the zeros of these polynomials (inside or outside the unit circle).Here instead of a single Stein equation we have a nested family of Stein equations(3.6), (3.7) conceivably associated with a family of orthogonal polynomials withrespect to a weight on the unit disk rather than on the unit circle. It would be ofinterest to extend the classical theory to this nested/Bergman-space setting. Thiswould follow up on one of the interests of Leonia Lerer (see, e.g., [15, 16]) to whomthis paper is dedicated.
5.2. Overlapping spaces
In the classical theory of de BrangesâRovnyak spaces (see in particular Sarasonâsbook [18]), a prominent role is played by the overlapping space
â(1â ð(ð§)ð(ðâ(1â ð§ð)
)â©â
(ð(ð§)ð(ð)â
(1â ð§ð)).
It would be of interest to determine if something significant can be said about
the analogous overlapping spaces â(1âð(ð§)ð(ð)â
(1âð§ð)ð
)â©â
(ð(ð§)ð(ð)â
(1âð§ð)ð
)for ð > 1; some
results in this direction appear in [19].
5.3. Characterization of de BrangesâRovnyak spaces viabackward-shift operator identities
In the classical setting (ð = 1) the de BrangesâRovnyak reproducing kernel Hilbertspaces â(ðŸð,1), and more generally reproducing kernel Hilbert spaces with kernel
of the form ðœâÎ(ð§)ðœÎ(ð€)â
1âð§ð, can be characterized via invariance under the backward-
shift operator combined with an appropriate functional identity (see Section 5.3 of[2] for a very general formulation and additional references and history). A naturalquestion for future investigation is whether such a characterization can be given
for reproducing kernel Hilbert spaces with kernel of the form ðœâÎ(ð§)ðœÎ(ð€)â
(1âð§ð)ð.
38 J.A. Ball and V. Bolotnikov
Acknowledgement
We would like to thank our colleague Sanne ter Horst for useful comments on anearly draft of this manuscript.
References
[1] D. Alpay and V. Bolotnikov, On tangential interpolation in reproducing kernelHilbert modules and applications, in: Topics in interpolation theory (eds. H. Dym,B. Fritzsche, V. Katsnelson and B. Kirstein), pp. 37â68, OT 95, Birkhauser, Basel,1997.
[2] D. Arov and H. Dym, ðœ-Contractive Matrix Valued Functions and Related Topics,Encyclopedia of Mathematics and its Applications 116, Cambridge, 2008.
[3] J.A. Ball and V. Bolotnikov, Contractive multipliers from Hardy space to weightedHardy space, Preprint, arXiv:1209.3690
[4] J.A. Ball and V. Bolotnikov,Weighted Bergman spaces: shift-invariant subspaces andinput/state/output linear systems, Integral Equations Operator Theory 76 (2013)no. 3, 301â356.
[5] J.A. Ball, V. Bolotnikov and S. ter Horst Interpolation in de BrangesâRovnyakspaces, Proc. Amer. Math. Soc. 139 (2011), no. 2, 609â618.
[6] J.A. Ball, V. Bolotnikov and S. ter Horst, Abstract interpolation in vector-valuedde BrangesâRovnyak spaces, Integral Equations Operator Theory 70 (2011), no. 2,227â268.
[7] J.A. Ball, I. Gohberg, and L. Rodman. Interpolation of Rational Matrix Functions,OT45, Birkhauser Verlag, 1990.
[8] J.A. Ball and M.W. Raney, Discrete-time dichotomous well-posed linear systemsand generalized SchurâNevanlinnaâPick interpolation, Complex Anal. Oper. Theory1 (2007), 1â54.
[9] F. Beatrous and J. Burbea, Positive-definiteness and its applications to interpolationproblems for holomorphic functions, Trans. Amer. Math. Soc., 284 (1984), no. 1, 247â270.
[10] V. Bolotnikov and L. Rodman, Remarks on interpolation in reproducing kernelHilbert spaces, Houston J. Math. 30 (2004), no. 2, 559â576.
[11] L. de Branges and J. Rovnyak, Canonical models in quantum scattering theory, Per-turbation Theory and its Applications in Quantum Mechanics (ed. C. Wilcox), pp.295â392, Wiley, New York, 1966.
[12] L. de Branges and J. Rovnyak, Square Summable Power Series, Holt, Rinehart andWinston, NewâYork, 1966.
[13] R.G. Douglas, On majorization, factorization, and range inclusion of operators onHilbert space, Proc. Amer. Math. Soc. 17 (1966), 413â415.
[14] O. Giselsson and A. Olofsson, On some Bergman shift operators, Complex Anal.Oper. Theory 6 (2012), 829â842.
[15] L. Lerer and A.C.M. Ran, A new inertia theorem for Stein equations, inertia ofinvertible Hermitian block Toeplitz matrices and matrix orthogonal polynomials, In-tegral Equations and Operator Theory 47 (2003) no. 3, 339â360
Interpolation in Sub-Bergman Spaces 39
[16] L. Lerer, I. Margulis, and A.C.M. Ran, Inertia theorems based on operator Lyapunovequations, Oper. Matrices 2 (2008) no. 2, 153â166.
[17] M. Rosenblum and J. Rovnyak, Hardy Classes and Operator Theory, Oxford Uni-versity Press, 1985.
[18] D.E. Sarason, Sub-Hardy Hilbert Spaces in the Unit Disk, John Wiley & Sons, Inc.,New York, 1994.
[19] S. Sultanic, Sub-Bergman Hilbert spaces, J. Math. Anal. Appl. 324 (2006), no. 1,639â649.
[20] K. Zhu, Sub-Bergman Hilbert spaces on the unit disk. Indiana Univ. Math. J. 45(1996), no. 1, 165â176.
[21] K. Zhu, Sub-Bergman Hilbert spaces on the unit disk. II, J. Funct. Anal. 202 (2003),no. 2, 327â341
Joseph A. BallDepartment of MathematicsVirginia TechBlacksburg, VA 24061-0123, USAe-mail: [email protected]
Vladimir BolotnikovDepartment of MathematicsThe College of William and MaryWilliamsburg VA 23187-8795, USAe-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 41â78câ 2013 Springer Basel
Zero Sums of Idempotents and BanachAlgebras Failing to be Spectrally Regular
H. Bart, T. Ehrhardt and B. Silbermann
Dedicated to Leonia Lerer, in celebration of his seventieth birthday
Abstract. A large class of Banach algebras is identified allowing for non-trivialzero sums of idempotents, hence failing to be spectrally regular. Belonging toit are the ð¶â-algebras known under the name Cuntz algebras. Other Banachalgebras lying in the class are those of the form â(ð) with ð a (non-trivial)Banach space isomorphic to a (finite) direct sum of at least two copies of ð.There do exist (somewhat exotic) Banach spaces for which â(ð) is spectrallyregular.
Mathematics Subject Classification (2010). Primary 46H99, 47C15; Secondary30G30, 46E15.
Keywords. Logarithmic residue, spectral regularity, (zero) sum of idempotents,Cuntz algebra, space of (bounded) continuous functions, Cantor type set.
1. Introduction
All algebras considered in this paper are assumed to be associative. A logarithmicresidue is a contour integral of the type
1
2ðð
â«âÎ
ð â²(ð)ð(ð)â1ðð, (1)
where the analytic function ð has its values in a unital complex Banach algebra â¬and âÎ is a suitable contour in the complex plane â, in fact the positively orientedboundary of a Cauchy domain Î. In the scalar case ⬠= â, the expression (1) isequal to the number of zeros of ð in Î, multiplicities of course taken into account.Thus, in that situation, the integral (1) vanishes if and only if ð takes non-zerovalues, not only on âÎ (which has been implicitly assumed in order to let (1)make sense) but on all of Î. This state of affairs leads to the following question: iffor a Banach algebra-valued analytic function ð the integral (1) vanishes, can oneconclude that ð takes invertible values on all of Î?
42 H. Bart, T. Ehrhardt and B. Silbermann
There are many situations where the answer to this question is positive (see[6], [8], [9], [11], [12], [13] and [15]); in general it is negative, however. The Banachalgebra â(â2) of all bounded linear operators on â2 is a counterexample (see [5] and[6]). This comes about from the fact that â(â2) allows for non-trivial zero sumsof idempotents, i.e., a finite collection of non-zero idempotents adding up to zero.Indeed, for Banach algebras featuring that phenomenon the answer to the abovequestion is always negative.
For a long time, â(â2) was basically the only known counterexample in con-nection with the issue stated above. In this paper, the existence of non-trivial zerosums of idempotents will be established for a large class of Banach algebras. Be-longing to it are the ð¶â-algebras known under the name Cuntz algebras. OtherBanach algebras lying in the class are those of the form â(ð) with ð a non-trivialBanach space isomorphic to a (finite) direct sum of at least two copies of ð . Hereâ(ð) stands for the Banach algebra of bounded linear operators on ð .
This brings us to a description of the contents of the different sections to befound below. Apart from the introduction contained in Section 1 and the list ofreferences, the paper consists of seven sections. Section 2 is of a preliminary natureand introduces the main concepts. Among them is that of spectral regularity ofa Banach algebra, meaning basically that for the algebra in question the answerto the question formulated above is always positive. It is also recalled from [6]that a spectrally regular Banach algebra does not allow for non-trivial zero sumsof idempotents. Section 3 contains information about zero sums of idempotentsin general algebras. Special attention is paid to sums involving a small number ofterms. Section 4 gives a criterion for a Banach algebra to have a non-trivial zerosum of idempotents, hence failing to be spectrally regular. The pertinent conditionis phrased in terms reminiscent of the defining characteristic for the ð¶â-algebrascalled Cuntz algebras, but there is no restriction here to the ð¶â-context (as is thecase in [15]). Sections 5 and 6 deal with the situation where the Banach algebraunder consideration is of the form â(ð) with ð an infinite-dimensional Banachspace. The idempotents in â(ð) then are the projections in ð . There do existinfinite-dimensional Banach spaces ð for which â(ð) is spectrally regular so thatnon-trivial zero sums of projections in ð do not exist. This follows by combiningresults from [11] with the quite remarkable examples given in [2], [22] and [1].Section 5 is concerned with the situation where a finite number of projectionsadd up to an operator which is quasinilpotent or compact. For non-trivial sumsof that type, the conclusion is drawn that they involve at least five projectionswith infinite rank and co-rank. At the end of the section, the following embeddingissue comes up: for Banach spaces ð and ð , when can â(ð ) be viewed as acontinuously embedded subalgebra of â(ð)? One simple case in which there is apositive answer is identified. Generally speaking however, the issue seems to berather non-trivial. In Section 6, the criterion exhibited in Section 3 leads to theconclusion that â(ð) allows for non-trivial zero sums of projections, hence is notspectrally regular, whenever ð is non-trivial and isomorphic to a (finite) directsum of at least two copies of itself. In Sections 7 and 8, this is used to identify new
Sums of Idempotents and Spectral Regularity of Banach Algebras 43
examples of Banach algebras lacking the property of being spectrally regular. Someof these relate to deep problems in general topology and the geometry of Banachspaces. This is especially true for Banach algebras of the form â(ð) with ð takento be a Banach space of bounded continuous functions. For instance, using a trulyremarkable theorem of A.A. Miljutin [28], the following result is obtained: If ð isan uncountable compact metrizable topological space and ð(ð;â) stands for theBanach space of all complex continuous functions on ð (endowed with the max-norm), then the operator algebra â(ð(ð;â)) allows for non-trivial zero sums ofprojections, hence is not spectrally regular. In two examples given towards the endof the paper, a (generalizing) modification of the well-known Cantor constructionplays an important role.
One final remark to close the introduction. The expression (1) defines the leftlogarithmic residue of ð . There is also a right version, obtained by replacing the leftlogarithmic derivative ð â²(ð)ð(ð)â1 by the right logarithmic derivative ð(ð)â1ð â²(ð).For some special cases, the relationship between left logarithmic residues and rightlogarithmic residues has been investigated: see [7], [8], [9], and [11]. As far asthe issues considered in the present paper are concerned, the results that can beobtained for the left and the right version of the logarithmic residue are analogousto each other. Therefore the qualifiers left and right will be suppressed in whatfollows.
2. Preliminaries
A spectral configuration is a triple (â¬,Î, ð) where ⬠is a non-trivial unital complexBanach algebra, Î is a bounded Cauchy domain in â (see [34] or [20]) and ð is aâ¬-valued analytic function defined on an open neighborhood of the closure of Îand having invertible values on all of the boundary âÎ of Î. With such a spectralconfiguration, taking âÎ to be positively oriented, one can associate the contourintegral
ð¿ð (ð ; Î) =1
2ðð
â«âÎ
ð â²(ð)ð(ð)â1ðð.
We call it the logarithmic residue associated with (â¬,Î, ð); sometimes the termlogarithmic residue of ð with respect to Î is used as well.
The spectral configuration (â¬,Î, ð) is called winding free when the logarith-mic residue ð¿ð (ð ; Î) = 0, spectrally winding free if ð¿ð (ð ; Î) is quasinilpotent,and spectrally trivial in case ð takes invertible values on all of Î. This terminologyis taken from [13].
In [13] a unital Banach algebra ⬠is said to be spectrally regular if a spectralconfiguration having ⬠as the underlying Banach algebra is spectrally trivial when-ever it is spectrally winding free. Here we shall work with a (possibly) somewhatweaker notion. We shall call ⬠spectrally regular if a spectral configuration having⬠as the underlying Banach algebra is spectrally trivial whenever it is windingfree. Whether this notion is strictly weaker than the one employed (under the
44 H. Bart, T. Ehrhardt and B. Silbermann
same name) in [13] is not known. It makes sense to adopt the weaker form of spec-tral regularity here because in this paper we are (mainly) interested in ânegativeresultsâ, i.e., in results having as conclusion that the Banach algebra under con-sideration fails to be spectrally regular. In that case we call it spectrally irregular.So the unital Banach algebra ⬠is spectrally irregular if and only if there exists aspectral configuration having ⬠as the underlying Banach algebra which is windingfree but not spectrally trivial.
Closely connected to the issue of spectral (ir)regularity is that of zero sumsof idempotents. Let ð be an algebra (possibly without norm). A (finite) numberof idempotents ð1, . . . , ðð in ð are said to form a zero sum if they add up to thezero element in ð, i.e., if ð1 + â â â + ðð = 0. The zero sum is called trivial whenðð = 0, ð = 1, . . . , ð. Non-triviality of a zero sum of idempotents then means thatat least one among the idempotents involved does not vanish. By leaving out thezero terms, such a non-trivial zero sum can be transformed into a genuine zerosum, that is one where all terms are non-zero.
In line with what has been the case up to now, here too spectral irregularitywill be brought to light via the construction of non-trivial zero sums of idempo-tents. The background for this is the following basic result taken from [5].
Theorem 2.1. Let ⬠be a unital Banach and let ð1, . . . , ðð be idempotents in â¬. If⬠is spectrally regular and ð1 + â â â + ðð = 0, then ðð = 0, ð = 1, . . . , ð.
We say that Banach algebra ⬠has the non-trivial zero sum property if thereexist a positive integer ð and non-zero idempotents ð1, . . . , ðð in ⬠such thatð1 + â â â + ðð = 0. The above theorem can then be read as follows: if ⬠has thenon-trivial zero sum property, then ⬠is spectrally irregular ; schematically:
non-trivial zero sum property â spectral irregularity.
It is an open problem whether or not the reverse implication is valid too. Noexample is known of a spectrally irregular Banach algebra which fails to have thenon-trivial zero sum property, so of a spectrally irregular Banach algebra allowingfor trivial zeros sums of idempotents only.
We close this section by mentioning that the issue of spectral (ir)regularityis closely related to non-commutative Gelfand theory (cf. [31]). In this connectionfamilies of matrix representations play an important role: the existence of certainspecific families for a Banach algebra ⬠implies spectral regularity of ⬠(see [13]and [14]). The strongest result of this type is that ⬠is spectrally regular when⬠allows for a radical-separating family of matrix representations. Here a family{ðð : ⬠â âððÃðð}ðâΩ is called radical-separating if it separates the points of â¬modulo the radical of â¬. Thus spectral irregularity brings with it that ⬠does notallow for a radical-separating family of matrix representations; schematically:
non-trivial zero sum propertyâ spectral irregularity
â absence of radical-separating families.
Sums of Idempotents and Spectral Regularity of Banach Algebras 45
The absence of radical-separating families (so a-fortiori having the non-trivial zerosum property) implies the non-existence of the families of other types consideredin [13] and [14].
3. Sums of small numbers of idempotents
In this section we focus on (zero) sums of a small number of idempotents (notcounting the possibly repeated occurrence of the unit element). To put the (new)material to be presented into proper perspective, we first recall some known facts(see [5]).
If ð is any algebra (normable or not), then a zero sum of three idempotentsin ð is necessarily trivial. There is an example of an algebra which allows fora genuine zero sum of four idempotents. The algebra in question is, however,non-normable. It must be because, as Theorem 4.3 from [5] asserts, zero sumsof four idempotents in a normed algebra are always trivial. We shall return tothis (non-trivial) result, in a moment. In [30] it is shown that every boundedlinear operator on the separable Hilbert space â2 can be written as the sum of fiveidempotents in the Banach algebra â(â2) of all bounded linear operators on â2.An immediate consequence of this is that â(â2) allows for non-trivial zero sumsof six idempotents. In fact one can make do with one less: the five idempotentsconstructed in [5], Example 3.1 yield a genuine zero sum in â(â2).
We now return to Theorem 4.3 from [5]. As was already indicated, the the-orem says that a zero sum of four idempotents in a Banach algebra (or, moregenerally a normed algebra) is always trivial. Here is an extension of this result.
Theorem 3.1. Let ð1, ð2, ð3 and ð4 be idempotents in a non-trivial Banach algebra⬠with unit element ðâ¬, and let ð be a non-negative integer. If
ð1 + ð2 + ð3 + ð4 + ðð⬠= 0,
then ð = 0 and ð1 = ð2 = ð3 = ð4 = 0.
Theorem 4.3 in [5] corresponds to the case where the integer ð is a prioriassumed to be zero. As to the proof of Theorem 3.1, we shall first show thatð = 0 which brings us in the situation considered in [5]. Then we shall extend theargument to cover the case ð = 0 too. In this way, a new proof of Theorem 4.3in [5] is obtained which is more transparent than the original one. The argumentgiven in [5], although conceptually elementary, is technically quite complicated.The reasoning presented below is suggested by the material in [19], Section 3. Thespectrum of an element ð¥ â ⬠is denoted by ð(ð¥).
Proof. As already indicated we first show that ð = 0. Put ð¥1 = ð1 + ð2 â ð⬠andð¥2 = ð3 + ð4 â ðâ¬. Then, by Lemma 3 in [19],
ð â ð(ð¥ð) â {â1,+1} â âð â ð(ð¥ð) â {â1,+1}, ð = 1, 2.
46 H. Bart, T. Ehrhardt and B. Silbermann
As ð¥1 + ð¥2 + (ð + 2)ð⬠= 0, we also have
ð â ð(ð¥1) â â(ð+ ð + 2) â ð(ð¥2).
Introduce the set of (negative) integers
ð = {âð â 1,â2ð â 3,â3ð â 5, . . .} ⪠{âð â 3,â2ð â 5,â3ð â 7, . . .},
and assume ð(ð¥1) contains an element ð which is not inð . Then âðâðâ2 belongsto ð(ð¥2) and ð+ ð + 2 â ð(ð¥2) provided that âðâ ð â 2 /â {â1,+1}. The lattercomes down to ð â= âðâ 1,âðâ 3 which certainly holds because ð is not in ð . Soð+ ð + 2 â ð(ð¥2). But then âðâ 2ð â 4 â ð(ð¥1), and we get ð+ 2ð + 4 â ð(ð¥1)because âðâ 2ðâ 4 /â {â1,+1}. Proceeding in this way (formally by induction ofcourse), we obtain ð+2ð(ð+2) â ð(ð¥1), ð = 0, 1, 2, . . . . As ð+2 is positive, thisconflicts with the boundedness of ð(ð¥1) implied by the non-triviality of â¬. Theconclusion is that ð(ð¥1) â ð .
Next assume that there is an element ð â ð(ð¥1) which does not belong toâð ⪠{â1,+1}. As ð differs from â1 and +1, we have âð â ð(ð¥1), and thisimplies âð â ð . But then ð â âð and we have a contradiction. The conclusionis that ð(ð¥1) â âð ⪠{â1,+1}.
The upshot of these arguments is that ð(ð¥1) â ð â© [âð ⪠{â1,+1}]. Now ðconsists of negative integers, so ð and âð are disjoint. Hence ð(ð¥1) is containedin ð â© {â1,+1}, and it follows that the latter is non-empty. Thus either â1 or+1 must belong to ð , and this is the case only when ð = 0, as desired.
In this way we have arrived at the situation considered in [5], Theorem 4.3: wehave four idempotents ð1, ð2, ð3 and ð4 in a Banach algebra adding up to the zeroelement. The following argument provides a new (and more transparent) proof forthe conclusion of [5], namely that ð1 = ð2 = ð3 = ð4 = 0.
With ð¥1 = ð1 + ð2 â ð⬠and ð¥2 = ð3 + ð4 â ð⬠as before, we have ð¥1 +ð¥2 + 2ð⬠= 0, and so ð â ð(ð¥1) â â(ð + 2) â ð(ð¥2). Also with ð = 0, theset ð introduced above becomes ð = {â1,â3,â5, . . .}. As we have seen above,ð(ð¥1) â ð â© {â1,+1}, and it follows that ð(ð¥1) = {â1}. We shall now investigatethe behavior of the resolvent (ððâð¥1)â1 which is defined and analytic on â â {â1}.
Take ð â= â1,+1, and put ðŠ = ð1â ð2. As has been observed in [19], and canbe easily verified, ð¥21+ ðŠ
2 = ð⬠and ðŠð¥1 = âð¥1ðŠ. With the help of these identities,one easily obtains the identities ðŠ = (ðð⬠â ð¥1)ðŠ(ðð⬠+ ð¥1)â1 and
(ðð⬠â ð¥1)(ðð⬠+ ð¥1 â ðŠ(ðð⬠+ ð¥1)â1ðŠ)
= ð2ð⬠â ð¥21 â (ðð⬠â ð¥1)ðŠ(ðð⬠+ ð¥1)â1ðŠ
= ð2ð⬠â ð¥21 â ðŠ2
= (ð2 â 1)ðâ¬.
Sums of Idempotents and Spectral Regularity of Banach Algebras 47
Dividing by ð2 â 1 we get
(ðð⬠â ð¥1)(
1
ð2 â 1(ðð⬠+ ð¥1 â ðŠ(ðð⬠+ ð¥1)â1ðŠ)
)= ðâ¬. (2)
In the same way one proves that, again for ð â= â1,+1,(1
ð2 â 1(ðð⬠+ ð¥1 â ðŠ(ðð⬠+ ð¥1)â1ðŠ)
)(ðð⬠â ð¥1) = ðâ¬,
and, in combination with (2), this leads to
(ðð⬠â ð¥1)â1 =(
1
ð2 â 1(ðð⬠+ ð¥1 â ðŠ(ðð⬠+ ð¥1)â1ðŠ)
).
Now 1 /â ð(ð¥1), so (ðð⬠+ ð¥1)â1 is analytic in a neighborhood of â1, and itfollows that (ðð⬠â ð¥1)â1 has a simple pole at â1. But then
(ðð⬠â (ð⬠+ ð¥1)
)â1has a simple pole at the origin. From ð(ð⬠+ ð¥1) = {0}, we see that ð⬠+ ð¥1 isquasinilpotent. Hence, by standard spectral theory,(
ðð⬠â (ð⬠+ ð¥1))â1
=1
ððâ¬, ð â= 0.
Thus ðð⬠â (ð⬠+ ð¥1) = ð ð⬠for all complex ð, and so ð⬠+ ð¥1 = 0. In otherwords ð1 + ð2 = 0. From ð1 = ð
21 = (âð2)2 = ð22 = ð2 = âð1 it is now clear that
ð1 = ð2 = 0. In the same way one shows that ð3 and ð4 vanish too. â¡
Corollary 3.2. Let ð1, ð2, ð3 and ð4 be idempotents in a Banach algebra ⬠with unitelement ðâ¬, let ð be a non-negative integer, and let ð¥ be a proper closed two-sidedideal in â¬. If
ð1 + ð2 + ð3 + ð4 + ðð⬠â ð¥ ,then ð = 0 and ð1, ð2, ð3, ð4 â ð¥ .
Proof. Pass to the quotient algebra â¬/ð¥ (which is non-trivial because ð¥ is proper),and apply Theorem 3.1 with ⬠replaced by â¬/ð¥ . â¡
4. Banach algebras of Cuntz type
The material in the next paragraph is presented by way of motivation for a defi-nition that will given below. The symbol ð¿ð,ð stands for the Kronecker delta.
The Cuntz algebra ðªð is the universal unital ð¶â-algebra generated by ðisometries ð£1, . . . , ð£ð â ðªð satisfying the identities
ðâð=1
ð£ðð£âð = ð, ð£âðð£ð = ð¿ð,ðð, ð, ð = 1, . . . , ð, (3)
where ð is the unit element in ðªð. Here ð is an integer larger than or equal to2. The first to consider this algebra was J. Cuntz [17]. The Cuntz algebras areuniversal in the sense that for fixed ð, any two concrete realization generated byisometries ð£1, . . . , ð£ð and ð£1, . . . , ð£ð, respectively, are *-isomorphic to each other
48 H. Bart, T. Ehrhardt and B. Silbermann
(cf. [17], [18]). For completeness, and to make the proper connection with [17], wenote that the relations (3) come down to the same as
ðâð=1
ð£ðð£âð = ð, ð£âðð£ð = ð, ð = 1, . . . , ð.
To see this multiply the first part of (3) from the left with ð£âð and from the rightwith ð£ð, and recall that a sum of nonnegative elements in a ð¶â-algebra can onlyvanish when so do all its terms.
Returning to general Banach algebras (not necessarily ð¶â), we stipulate thata non-trivial unital Banach algebra ⬠will be said to have the Cuntz ð-property ifð is an integer larger than one and there exist elements ð£1, . . . , ð£ð, ð€1, . . . , ð€ð in⬠such that
ðâð=1
ð£ðð€ð = ðâ¬, ð€ðð£ð = ð¿ð,ððâ¬, ð, ð = 1, . . . , ð. (4)
We emphasize that this definition does not imply that the algebra is generated bythe elements ð£1, . . . , ð£ð, ð€1, . . . , ð€ð. The following statement can be easily verified.If, for ð = 1, 2, the Banach algebra ⬠has the Cuntz ðð -property, then ⬠has theCuntz (ð1 + ð2 â 1)-property; hence, if ⬠has the Cuntz ð-property, then ⬠hasthe Cuntz (ðð â ð + 1)-property for each positive integer ð. The argument is asfollows. Suppose ð£ð ,1, . . . , ð£ð ,ðð , ð€ð ,1, . . . , ð€ð ,ðð â ⬠satisfy the identities
ðð âð=1
ð£ð ,ðð€ð ,ð = ðâ¬, ð€ð ,ðð£ð ,ð = ð¿ð,ððâ¬, ð, ð = 1, . . . , ðð .
For ð = 1, . . . , ð1, write ð£ð = ð£2,1ð£1,ð , ð€ð = ð€1,ðð€2,1. Also, for the values of ðranging from ð1 + 1, up to ð1 + ð2 â 1, put ð£ð = ð£2,ðâð1+1 and ð€ð = ð€2,ðâð1+1.Then (4) holds with ð = ð1 + ð2 â 1.
A non-trivial unital Banach algebra ⬠is said to be of Cuntz type if ⬠has theCuntz ð-property for some integer ð larger than one. Such an algebra necessarilyis non-commutative.
Theorem 4.1. Let ð be an integer larger than or equal to five, let ⬠be a non-trivialunital Banach algebra, and suppose ⬠is of Cuntz type. Then ⬠allows for a zerosum of ð non-zero idempotents.
Thus Banach algebras of Cuntz type have the non-trivial zero sum prop-erty introduced in Section 2 and are (therefore) spectrally irregular. As has beenmentioned in Section 3, in a Banach algebra non-trivial zero sums of idempotentsinvolving less than five idempotents cannot exist.
Proof. We shall break up the reasoning into seven steps. In the first, some prepara-tory action is taken.
Step 1. If ðŽ1, . . . , ðŽð are matrices, their direct sum will be denoted by ðŽ1ââ â â âðŽ ð
orâ ðð=1ðŽð . Soâ ð
ð=1ðŽð is the block diagonal matrix with ðŽ1, . . . , ðŽ ð on the diagonal
Sums of Idempotents and Spectral Regularity of Banach Algebras 49
(in this order). In case all the matrices ðŽð coincide with a single matrix ðŽ, we useðŽâ ð for â ð
ð=1ðŽð = ðŽ â â â â â ðŽ (with ð terms in the direct sum). If ðŽ is the
sum of ð non-zero idempotents, then so is ðŽâð. Indeed, if ðŽ =âð
ð=1 ðð , then
ðŽâð =âð
ð=1 ðâðð .
Whenever convenient complex matrices will be identified with matrices hav-ing entries in â¬. The identification goes via âtensorizingâ with ðâ¬, i.e., via replacingeach scalar entry by the corresponding multiple of the unit element in â¬.Step 2. Consider the matrices
ð1 =1
2
[â5 35
â1 7
], ð2 =
1
2
[â5 â351 7
].
ð1 =1
6
â¡â¢â£ 5 15 â15 15 â170 210 â14
â€â¥âŠ, ð2 =1
6
â¡â¢â£ 5 â15 â1â5 15 1
70 â210 â14
â€â¥âŠ, ð3 =1
3
â¡â¢â£ 10 0 1
0 0 0
â70 0 â7
â€â¥âŠ.These are idempotents, regardless of whether they are considered as complex ma-trices (in which case they have rational entries and rank 1), or as matrices withentries in â¬. Also
ð1 +ð2 =
[â5 0
0 7
], ð1 +ð2 +ð3 =
â¡â¢â£5 0 0
0 5 0
0 0 â7
â€â¥âŠ .Thus [â5] â [7 ] is the sum of the two non-zero idempotents ð1 and ð2, and[5]â [5]â [â7 ] is the sum of the three non-zero idempotents ð1, ð2 and ð3.
Step 3. For simplification of notation, put ð = ðâ 1. Then [[â5]â [7 ]]âðis a sum
of two non-zero idempotents, and[[5] â [5]â [â7 ]]âð
is a sum of three non-zeroidempotents. In fact [
[â5]â [7 ]]âð=ðâð
1 +ðâð2 ,[
[5]â [5]â [â7 ]]âð= ðâð
1 +ðâð2 +ðâð
3 .
Now choose permutation similarities Î ð and Î ð (having the effect of interchang-ing rows and columns) such that Î â1ð = Î ð and Î â1ð = Î ð while, moreover,
Î ð
([[â5]â [7 ]]âð
)Î ð = [â5]âð â [7 ]âð,
Î ð
([[5]â [5]â [â7 ]]âð
)Î ð = [5]â(ð+1) â [5]â(ðâ1) â [â7 ]âð.
50 H. Bart, T. Ehrhardt and B. Silbermann
It follows that the right-hand sides of these expressions are the sum of two andthree idempotents, respectively:
[â5]âð â [7 ]âð =2â
ð=1
Î ð
(ðâð
ð
)Î ð ,
[5]â(ð+1) â [5]â(ðâ1) â [â7 ]âð =
3âð=1
Î ð
(ðâð
ð
)Î ð .
Step 4. Before proceeding with the main line of the argument, we make an auxiliaryremark. Let ð£ â â¬ðÃð, ð€ â â¬ðÃð and suppose ð€ð£ is the identity element in â¬ðÃð.Further assume ð â â¬ðÃð is a sum of ð non-zero idempotents in â¬ðÃð. Then theproduct ð£ðð€ â â¬ðÃð is a sum of ð non-zero idempotents in â¬ðÃð. To see this,write ð = ð1 + â â â + ðð with ð1, . . . , ðð non-zero idempotents in â¬ðÃð. Then wehave ð£ðð€ = ð£ð1ð€ + â â â + ð£ððð€ with (ð£ððð€)
2 = ð£ððð€ð£ððð€ = ð£ð2ðð€ = ð£ððð€. Fromð€ð£ððð€ð£ = ðð one sees that ð£ððð€ cannot be zero.
Step 5. From now on ð£1, . . . , ð£ð, ð€1, . . . , ð€ð will be elements in ⬠satisfying (4).Put ð = [ð£1 . . . ð£ð+1] and ð = [ð€1 . . . ð€ð+1]
â€, where †signals the operation oftaking the transpose. Then ð â â¬1Ã(ð+1),ð â â¬(ð+1)Ã1, and (4) can be rephrasedby saying that ðð is the identity element in ⬠and ðð is the identity elementin â¬(ð+1)Ã(ð+1). Write
ð1 = ð â [ðâ¬]â(2ðâ1) = ð â [ðâ¬]â(ðâ1) â [ðâ¬]âð,
ð1 =ð â [ðâ¬]â(2ðâ1) =ð â [ðâ¬]â(ðâ1) â [ðâ¬]âð.
These matrices belong to â¬2ðÃ3ð and â¬3ðÃ2ð, respectively. Also ð1ð1 is the iden-tity element in â¬3ðÃ3ð. In combination with what we obtained in Step 4, this givesthat ð1
([5]â(ð+1)â [5]â(ðâ1)â [â7]âð
)ð1 is a sum of three non-zero idempotents
in â¬2ðÃ2ð :
ð1([5]â(ð+1) â [5]â(ðâ1) â [â7 ]âð
)ð1 =
3âð=1
ð1Î ð
(ðâð
ð
)Î ðð1.
Using the defining expressions for ð1 and ð1, we can rewrite the left-hand side ofthis identity as 5ðð â [5]â(ðâ1) â [â7 ]âð and this, in view of the fact that ððis the identity element in â¬, is equal to [5]âðâ [â7 ]âð. So the latter is the sum ofthree non-zero idempotents in â¬2ðà 2ð :
[5]âð â [â7 ]âð =3â
ð=1
ð1Î ð
(ðâð
ð
)Î ðð1.
Again referring to Step 2, we recall that [â5]âðâ [7 ]âð is the sum of the two non-
zero idempotents in â¬2ðà 2ð, namely Î ð
(ðâð
1
)Î ð and Î ð
(ðâð
2
)Î ð . Thus
Sums of Idempotents and Spectral Regularity of Banach Algebras 51
the zero element in â¬2ðà 2ð appears as the sum of five non-zero idempotents:
2âð=1
Î ð
(ðâð
ð
)Î ð +
3âð=1
ð1Î ð
(ðâð
ð
)Î ðð1 = 0.
Step 6. Next we make a reduction from â¬2ðà 2ð to â¬. Putð2 =
[ð£1 â â â ð£ð ð£ð+1ð£1 â â â ð£ð+1ð£ð
],
ð2 =[ð€1 â â â ð€ð ð€1ð€ð+1 â â â ð€ðð€ð+1
]â€.
These matrices belong to â¬1Ã2ð and â¬2ðÃ1, respectively. Also ð2ð2 is the identityelement in â¬2ðÃ2ð, and so (again on account of Step 4) the zero element in â¬appears as the sum of five non-zero idempotents in â¬:
2âð=1
ð2Î ð
(ðâð
ð
)Î ðð2 +
3âð=1
ð2ð1Î ð
(ðâð
ð
)Î ðð1ð2 = 0. (5)
Step 7.We have proved the statement in the theorem now forð = 5: there exist fivenon-zero idempotents ð(5)1, . . . , ð(5)5 â ⬠with ð(5)1+ â â â + ð(5)5 = 0. To get theresult for arbitrary ð larger than or equal to five, it suffices to deal with the casesð = 6 up to and including ð = 9. Indeed, we then have that for ð = 5, 6, 7, 8, 9there are non-zero idempotents ð(ð)1, . . . , ð(ð)ð â ⬠such that ð(ð)1+ â â â +ð(ð)ð = 0.Given any integer ð larger than or equal to five, we write ð in the form 5ð + ðwith ð a non-negative integer and ð â {5, 6, 7, 8, 9}. Then
ðð1 + â â â + ðð5 + ð(ð)1 + â â â + ð(ð)ð = 0
is a zero sum involving ð non-zero idempotents in â¬.We finish the proof by establishing the existence of the idempotents
ð(ð)1, . . . , ð(ð)ð featuring in the previous paragraph. Take ð â {6, 7, 8, 9}, and put
ðð =
â§âšâ©
[ðð 0
0 0
], ð = 1, . . . , (ð â 5),
[ðð 0
0 ð 5+ðâð
], ð = (ð â 4), . . . , 5,
[0 0
0 ð 5+ðâð
], ð = 6, . . . , ð.
Then ð1, . . . , ðð are ð non-zero idempotents in â¬2Ã2 adding up to the zero ele-ment in â¬2Ã2. For ð(ð)ð we can now take [ð£1 ð£2]ðð [ð€1 ð€2]
â€, ð = 1, . . . , ð. Indeed,[ð€1 ð€2]
â€[ð£1 ð£2] is the unit element in â¬2Ã2. â¡Elaborating on the above proof, we rewrite the identity (5) as
2âð=1
ðð
(ðâ(ðâ1)ð
)ðð +
3âð=1
ðð
(ðâ(ðâ1)ð
)ðð = 0,
52 H. Bart, T. Ehrhardt and B. Silbermann
where ðð â â¬1à 2(ðâ1) and ðð â â¬2(ðâ1)Ã1 are given by
ðð =[ð£1 â â â ð£ðâ1 ð£ðð£1 â â â ð£ðð£ðâ1
]Î ð
ðð = Î ð
[ð€1 â â â ð€ðâ1 ð€1ð€ð â â â ð€ðâ1ð€ð
]â€,
and ðð â â¬1à 3(ðâ1) and ðð â â¬3(ðâ1)Ã1 by
ðð =[ð£21 ð£1ð£2 â â â ð£1ð£ð ð£2 â â â ð£ðâ1 ð£ðð£1 â â â ð£ðð£ðâ1
]Î ð
ðð = Î ð
[ð€21 ð€2ð€1 â â â ð€ðð€1 ð€2 â â â ð€ðâ1 ð€1ð€ð â â â ð€ðâ1ð€ð
]â€.
Anticipating on what we shall need in the proof of Theorem 6.3 below, wenote that ðððð and ðððð are the identity elements in â¬2(ðâ1)à 2(ðâ1) andâ¬3(ðâ1)à 3(ðâ1), respectively.
In our reasoning, the idempotents ð1, . . . , ðð meant in Theorem 4.1 come upas linear combinations involving rational coefficients of monomials in the elementsð£1, . . . , ð£ð, ð€1, . . . , ð€ð satisfying (4). It is possible to give additional information,for instance on the number of the monomials involved (maximally 9(ðâ 1)2) andtheir degree (at most 4), but we refrain from giving further details here. Insteadwe say something more about the cases ð = 2, ð = 5 and ð = 3, ð = 5.
When ð = 2, the matrix Î ð can be chosen to be the 2Ã 2 identity matrix,and one obtains
ðð =[ð£1 ð£2ð£1
], ðð =
[ð€1 ð€1ð€2
]â€.
Also, for Î ð one can take the 3Ã 3 identity matrix which leads toðð =
[ð£21 ð£1ð£2 ð£2ð£1
], ðð =
[ð€21 ð€2ð€1 ð€1ð€2
]â€.
The idempotents in ⬠associated with the 2 à 2 matrices ðð, ð = 1, 2, from
Step 2 in the proof of Theorem 4.1 are now[ð£1 ð£2ð£1
]ðð
[ð€1 ð€1ð€2
]â€. Similarly,
those corresponding to the 3 Ã 3 matrices ðð, ð = 1, 2, 3 (see again Step 2),
can be written as[ð£21 ð£1ð£2 ð£2ð£1
]ðð
[ð€1 ð€1ð€2 ð€2ð€1
]â€. These five non-zero
idempotents, involving degree four polynomials in the elements ð£1, ð£2, ð€1 and ð€2,add up to the zero element in â¬.
In case ð = 3, the matrix Î ð can be chosen to be the 4 Ã 4 permutationsimilarity corresponding to the exchange of the second and the third row (column),and one gets
ðð =[ð£1 ð£3ð£1 ð£2 ð£3ð£2
],
ðð =[ð€1 ð€1ð€3 ð€2 ð€2ð€3
]â€.
Also, for Î ð one can take the 6 Ã 6 permutation similarity corresponding to theexchange of the third and the fifth row (column), which leads to
ðð =[ð£21 ð£1ð£2 ð£3ð£1 ð£2 ð£1ð£3 ð£3ð£2
],
ðð =[ð€21 ð€2ð€1 ð€1ð€3 ð€2 ð€3ð€1 ð€2ð€3
]â€.
Sums of Idempotents and Spectral Regularity of Banach Algebras 53
The idempotents in ⬠associated with the 2Ã2 matricesðð, ð = 1, 2, from Step 2in the proof of Theorem 4.1 are now[
ð£1 ð£3ð£1 ð£2 ð£3ð£2] [ðð 0
0 ðð
] [ð€1 ð€1ð€3 ð€2 ð€2ð€3
]â€.
Similarly, those corresponding to the 3Ã 3 matrices ðð, ð = 1, 2, 3 (see once moreStep 2), can be written as
[ð£21 ð£1ð£2 ð£3ð£1 ð£2 ð£1ð£3 ð£3ð£2
] [ðð 0
0 ðð
]à [ð€2
1 ð€2ð€1 ð€1ð€3 ð€2 ð€3ð€1 ð€2ð€3
]â€.
These five non-zero idempotents, involving degree four polynomials in the elementsð£1, ð£2, ð£3, ð€1, ð€2 and ð€3, add up to the zero element in â¬.
For these small values of ð (2 and 3) and ð (= 5), the five polynomials thatcame up above can of course be computed explicitly. Once this is done, it alsopossible to prove directly that they constitute non-zero idempotents in ⬠whichadd up to the zero element of â¬. For other low values of ð and ð such an approachmight work too; for higher values it becomes practically unmanageable though.
A Cuntz algebra obviously is a Banach algebra (actually a ð¶â-algebra) ofCuntz type. Thus we have the following direct consequence of Theorem 4.1.
Corollary 4.2. Given an integer ð larger than or equal to five, a Cuntz algebraallows for a zero sum of ð non-zero idempotents.
In particular Cuntz algebras have the non-trivial zero sum property and are(consequently) spectrally irregular.
5. Sums of idempotents in Banach algebras of the form í(ð¿)
Throughout this section, ð will denote a (non-trivial) Banach space. Adoptingstandard terminology, by a projection in ð we mean an idempotent in the Banachalgebra â(ð) of bounded linear operators on ð . If ð is a projection in ð , the(possibly infinite) dimension of the range of ð will be called the rank of ð , writtenrankð . The following result generalizes Proposition 2.1 in [5] which states thatnon-trivial zero sums of finite rank projections cannot exist.
Proposition 5.1. Let ð1, . . . , ðð be finite rank projections in ð and assume theirsum ð1 + â â â + ðð is quasinilpotent. Then ðð = 0, ð = 1, . . . ,ð.
So, in fact, the sum ð1 + â â â + ðð vanishes.
Proof. Put ð = ð1+ â â â +ðð. Then ð is quasinilpotent and has finite rank. Henceð is nilpotent and the trace of ð, written traceð, vanishes. As ð1, . . . , ðð are
54 H. Bart, T. Ehrhardt and B. Silbermann
projections, we have traceðð = rankðð, ð = 1, . . .ð. It follows that
ðâð=1
rankðð =
ðâð=1
traceðð = trace
( ðâð=1
ðð
)= traceð = 0,
and we get ðð = 0, ð = 1, . . . , ð, as desired. â¡
If ð is a projection in ð , the (possibly infinite) dimension of the null spaceof ð is called the co-rank of ð , written co-rankð . Note that co-rankð coincideswith rank (ðŒð â ð ), where ðŒð is the identity operator on ð .
Proposition 5.2. Let ð be a positive integer and let ð1, . . . , ðð be projections in ð.Assume these projections have all finite co-rank and ð1+â â â +ðð is quasinilpotent.Then ð is finite dimensional and ðð = 0, ð = 1, . . . ,ð.
So, actually, the sum ð1 + â â â + ðð vanishes.
Proof. Taking into account Proposition 5.1, it suffices to show that ð is finitedimensional. Put ð = ð1 + â â â + ðð. Then ð is quasinilpotent and, consequently,ððŒð â ð is invertible. On the other hand
ððŒð â ð =ðâ
ð=1
(ðŒð â ðð)
is of finite rank. But then so is ðŒð = (ððŒð â ð)â1(ððŒð â ð), and the finitedimensionality of ð follows. â¡
Theorem 5.3. Let ð be a positive integer, let ð1, . . . , ðð be projections in ð, andsuppose the sum ð1 + â â â + ðð is compact. Then (precisely) one of the followingstatements holds:
(a) ð1, . . . , ðð are all of finite rank (hence so is their sum),(b) ð ⥠5 and at least five among the projections ð1, . . . , ðð have both infinite
rank and co-rank.
It is worthwhile to say a few words about the situation when all the projec-tions ð1, . . . , ðð have finite co-rank. If that is the case and, in addition,ð1 + â â â + ðð is compact, then (a) holds, i.e., ð1, . . . , ðð are all of finite rankas well. It follows that ðŒð = (ðŒð â ð1) + ð1 has finite rank, and we arrive atone of the conclusions also appearing in Proposition 5.2, namely that ð is finitedimensional. So a compact sum of finite co-rank projections can only occur whenthe underlying space is finite dimensional.
Proof. First assume that each of the projections ð1, . . . , ðð is either of finite rankor of finite co-rank. Write ð for the number of projections among ð1, . . . , ððthat are of finite co-rank. If ð = 0, we have (a). So suppose ð is at least one.Renumbering (if necessary), we can achieve the situation where ð1, . . . , ðð have
Sums of Idempotents and Spectral Regularity of Banach Algebras 55
finite co-rank and ðð+1, . . . , ðð are of finite rank. Now
ðâð=1
ðð =ðâ
ð=1
ðð âðâ
ð=ð+1
ðð,
where the first sum in the right-hand side is compact (by hypothesis) and the sec-ond of finite rank. So ð1+â â â +ðð is compact. The projections (ðŒðâð1), . . . , (ðŒðâðð) are of finite rank. Further
ðŒð =1
ð
( ðâð=1
ðð +
ðâð=1
(ðŒð â ðð)).
It follows that ðŒð is compact and, consequently, ð is finite dimensional. Underthese circumstances, the validity of (a) is a triviality.
Next consider the case when among ð1, . . . , ðð, there are idempotents whichare neither of finite rank nor of finite co-rank. Let there be ð of those. We mayassume (renumbering if necessary) that ð1, . . . , ðð are of this type and (hence)ðð+1, . . . , ðð are not, i.e., they are of finite rank or finite co-rank. Let ð bethe number of idempotents among ðð+1, . . . , ðð that have finite co-rank, andsuppose (without loss of generality) that ðð+1, . . . , ðð+ð are of that kind. Thenðð+ð+1, . . . , ðð have finite rank.
The same is true for the projections (ðŒð â ðð+1), . . . , (ðŒð â ðð+ð), and itfollows that ð1+â â â +ðð+ððŒð is compact. Now apply Corollary 3.2 with ⬠= â(ð)and taking for ð¥ the ideal of compact operators on ð . This gives that all theprojections ð1, . . . , ðð are compact. But then they are of finite rank, which isimpossible in view of how the number ð has been introduced. So ð (and a fortiorið) must be at least five, as claimed in (b). â¡
In the situation where the sum of idempotents in Theorem 5.3 is both compactand quasinilpotent (for instance because it vanishes), the conclusion of the theoremcan be sharpened.
Theorem 5.4. Let ð be a positive integer and let ð1, . . . , ðð be projections in ð.Suppose the sum ð1+ â â â +ðð is compact and quasinilpotent. Then (precisely) oneof the following statements holds:
(a) ðð = 0, ð = 1, . . . ,ð (so, in fact, the sum ð1 + â â â + ðð vanishes),(b) ð ⥠5 and at least five among the idempotents ð1, . . . , ðð have both infinite
rank and co-rank.
Proof. Combine Theorem 5.3 and Proposition 5.1. â¡As we have seen, in dealing with non-trivial zero sums of idempotents, the
number five plays a special role. This fact is underlined by the following result onzero sums of five projections.
Corollary 5.5. Let ð1, ð2, ð3, ð4 and ð5 be projections in ð, not all equal to thezero operator on ð, and assume ð1+ð2+ð3+ð4+ð5 = 0. Then all five projectionsð1, ð2, ð3, ð4 and ð5 have both infinite rank and co-rank.
56 H. Bart, T. Ehrhardt and B. Silbermann
Before we proceed with some additional observations, we make a connectionwith [15]. Notions like finite rank and compactness can be introduced in a meaning-ful way for elements in ð¶â-algebras (see [4] and [24]). It turns out that the resultsobtained so far in this section have analogues in this ð¶â-context. For details werefer to [15], Section 3 in particular. The proofs presented there (of Propositions3.9 and 3.10, of Theorems 3.12 and 3.13, and of Corollary 3.14) are modificationsof those given here.
Now let us return to zero sums of projections in the given Banach space ð .Suppose we have such a sum:
ð1 + â â â + ðð = 0, (6)
with ð a positive integer and ð1, . . . , ðð non-zero projections in ð . Then neces-sarilyð ⥠5, and at least five among the idempotents ð1, . . . , ðð have both infiniterank and co-rank (Theorem 5.4). Without loss of generality we may assume thatðð is of this type. Thus both the image and the null space of ðð have infinitedimension. Let ð be any positive integer larger than ð, and let ð1, . . . , ððâð bepositive integers too. Then a routine argument shows that ðð can be written asðð = ðð + ðð+1 + â â â + ðð with ðð a projection of both infinite rank andco-rank, and ðð a projection of finite rank ððâð, ð = ð+1, . . . , ð. In this way wearrive at the zero sum
ð1 + â â â + ððâ1 +ðð +ðð+1 + â â â +ðð = 0, (7)
involving a total of ð non-zero idempotents, featuring just as many projectionsof both infinite rank and co-rank as there are in the original zero sum (6), andcompared to that one having ðâð additional projections of prescribed finite rank.One may ask whether it is also possible to transform (6) into a zero sum (7) withð terms by writing ðð as a sum ðð = ðð + ðð+1 + â â â + ðð involving onlyprojections of both infinite rank and co-rank. This is problematic because, as hasbeen shown in [22], there are Banach spaces lacking complemented subspaces withboth infinite dimension and codimension; see however Theorem 6.3 below.
Suppose that besidesð another (non-trivial) Banach space ð is given. If thereexists an injective continuous Banach algebra homomorphism Ί : â(ð ) â â(ð)(possibly non-unital) and if â(ð ) has the non-trivial zero sum property, thenobviously so does â(ð). Indeed, if ð1 + â â â + ðð = 0 is a non-trivial zero sumof projections in ð , then Ί(ð1) + â â â + Ί(ðð) = 0 is a non-trivial zero sum ofprojections in ð . This straightforward observation lead to the following embeddingissue: when can â(ð ) be viewed as a continuously embedded subalgebra of â(ð)?This is a non-trivial question indeed. Even first attempts to deal with it touch ondeep problems in the geometry of Banach spaces. Here are some observations.
There is one very simple case in which the answer is positive.
Proposition 5.6. Assume ð is isomorphic to a complemented subspace of ð. Thenâ(ð ) can be continuously embedded into â(ð), i.e., there exists an injective con-tinuous Banach algebra homomorphism from â(ð ) into â(ð).
Sums of Idempotents and Spectral Regularity of Banach Algebras 57
Proof. We may assume that ð is a complemented subspace of ð . Let ðœ be thenatural embedding of the (closed) subspace ð intoð and let ð be the projection ofð onto ð , viewed as an operator from ð into ð . Then ðœ : ð â ð and ð : ð â ðare bounded linear operators, injective and surjective, respectively. Also ððœ is theidentity operator on ð . Define Ί : â(ð ) â â(ð) by Ί(ð ) = ðœðð, ð â â(ð ).Then Ί is an injective continuous Banach algebra homomorphism. â¡
The complementedness assumption in Proposition 5.6 is essential: when ðis a non-complemented closed subspace ð of ð it may happen that an injectivecontinuous Banach algebra homomorphism from â(ð ) into â(ð) does not exist.For an example we need to delve into the geometry of Banach spaces.
For quite some time it was an open problem whether there exists an infinite-dimensional Banach space ð that solves the so-called scalar-plus-compact problem.This means that ð has only very few operators in the sense that each boundedlinear operator on ð is the sum of a scalar multiple of ðŒð and a compact operatoron ð. It was in [2] that a first example was given. In the recent paper [1], anotherspace ðž with the property in question has been produced, this time having theadditional feature that ðž contains a closed subspace ð¹ which is isomorphic tothe separable Hilbert space â2. As is clear from Corollary 2.2 in [10], a Banachspace that solves the scalar-plus-compact problem does not allow for non-trivialzero sums of projections (see also [11], Corollary 3.4). So this is the case for ðž.However, as has been established in [5], Example 3.1, such a non-trivial zero sumof projections does exist for the Banach space â2. By isomorphy, this carries over toð¹ . Hence there is no injective continuous Banach algebra homomorphism (possiblynon-unital) of â(ð¹ ) into â(ðž). This conclusion is corroborated by the fact thatâ(ðž) is spectrally regular (see Corollary 4.3 in [11]) and â(ð¹ ) is not (cf. [13],Section 4 in particular).
6. Cuntz type Banach algebras of the form í(ð¿)
As was mentioned in Section 1 (Introduction), the first example found of a unitalBanach algebra failing to be spectrally regular was the operator algebra â(â2).Actually, this algebra allows for non-trivial zero sums of idempotents which impliesthat it lacks the property of being spectrally regular. Example 3.1 in [5], exhibitingthis, fits in the framework of Cuntz type algebras developed in Section 4. In thepresent section, the set up in question will be considered for Banach algebras ofthe form â(ð). But first we make some introductory remarks.
As we have seen already at the end of the previous section, it can happen thatâ(ð) is spectrally regular even when ð has infinite dimension. The examples from[2] and [1] mentioned there are rather spectacular and connected to deep problemsfrom Banach space geometry. Another similarly remarkable instance where â(ð)is spectrally regular can be found in [22]. Indeed, an example is given there of aninfinite-dimensional Banach space ð such that each bounded linear operator onð is the sum of a scalar multiple of ðŒð and a strictly singular operator on ð . In
58 H. Bart, T. Ehrhardt and B. Silbermann
that case the arguments given in the first part of [11], Section 4 apply upon slightmodification.
Next we turn to the investigation of the situation where â(ð) is of Cuntztype (hence spectrally irregular because of the occurrence of non-trivial zero sumsof idempotents). The following observation is straightforward.
Proposition 6.1. If the Banach algebra ⬠has the Cuntz ð-property, then so hasthe operator algebra â(â¬).Proof. Let ð£1, . . . , ð£ð, ð€1, . . . , ð€ð be elements in ⬠satisfying (4). Replacing theseelements by their left regular representations on ⬠(considered as a Banach space),we obtain operators ð1, . . . , ðð,ð1, . . . ,ðð in â(â¬) such that (4) holds with thelower case letters replaced by the corresponding upper case ones, and with ðŒðreplaced by the identity operator ðŒâ¬ on â¬. â¡
From Proposition 6.1 we see that with every Banach algebra of Cuntz typethere comes one of the form â(ð) with ð a Banach space. This in itself is alreadyreason enough to pay special attention to the case of such operator algebras. Alsoit is an elementary fact that a unital Banach algebra ⬠can be identified (forinstance via the use of left regular representations) with a Banach subalgebra ofâ(â¬). Proposition 6.1 and its proof now tell us that each Banach algebra havingthe Cuntz ð-property can be viewed as a Banach subalgebra of a Banach algebrahaving the Cuntz ð-property too and being of the type â(ð) for some Banachspace ð . This is an additional reason for now looking at Banach algebras of theform â(ð).Theorem 6.2. Let ð be a non-trivial Banach space, and let ð be an integer largerthan one. Then the operator algebra â(ð) has the Cuntz ð-property if and only ifð is isomorphic to ðð, where ðð denotes the direct sum of ð copies of ð.
For the arguments below to work, the norm on ðð needs to have an appro-priate relationship with the given norm on ð . What matters is that the followinglinear operators are continuous: the embeddings
ð â ð¥ ï¿œâ (0, . . . , 0, ð¥, 0, . . . , 0)†â ðð,
with ð¥ in the ðth position, and the projections
ðð â (ð¥1, . . . , ð¥ð)†ᅵâ ð¥ð â ð.Here ð is allowed to take the values 1, . . . , ð. Such norms are mutually equivalent,and we will settle here on the norm â£â£â£ â â£â£â£ : ðð â [0,â) given by
â£â£â£(ð¥1, . . . , ð¥ð)†â£â£â£ = maxð=1,...,ð
â¥ð¥ðâ¥, (ð¥1, . . . , ð¥ð)†â ðð,
where ⥠â ⥠is the norm on ð . One could also take any norm on ðð induced by amonotone norm on âð (see [16] for the definition). For ð such a monotone normon âð, the norm â£â£â£ â â£â£â£ð on ð induced by ð has the form â£â£â£(ð¥1, . . . , ð¥ð)†â£â£â£ð =
ð(â¥ð¥1â¥, . . . , â¥ð¥ðâ¥), (ð¥1, . . . , ð¥ð)†â ðð.
Sums of Idempotents and Spectral Regularity of Banach Algebras 59
Proof. First let us deal with the âonly if partâ of the theorem. So assume theexistence of ð1, . . . , ðð,ð1, . . . ,ðð in â(ð) satisfying
ðâð=1
ðððð = ðŒð , ðððð = ð¿ð,ððŒð , ð, ð = 1, . . . , ð, (8)
where ðŒð stands for the identity operator on ð . Now introduce the bounded linearoperators ð : ðð â ð and ð : ð â ðð by
ð (ð¥1, . . . , ð¥ð)†= ð1ð¥1 + â â â + ððð¥ð, (ð¥1, . . . , ð¥ð)
†â ðð, (9)
ðð¥ = (ð1ð¥, . . . ,ððð¥)â€, ð¥ â ð. (10)
By (8) these are each others inverse. Thus ð is isomorphic to ðð.Next we turn to the âif partâ of the theorem. Let the bounded linear operators
ð : ðð â ð and ð : ð â ðð be each others inverse. The choice made for thenorm on ðð implies that the expressions (9) and (10) determine bounded linearoperators ð1, . . . , ðð,ð1, . . . ,ðð on ð . The identities (8) follow from ðð = ðŒðand ðð = ðŒðð . â¡
Combining Proposition 6.1 and Theorem 6.2, one immediately gets that thefollowing is true. If the Banach algebra ⬠has the Cuntz ð-property, then â¬, viewedas Banach space, is isomorphic to â¬ð. The converse is not true. For an example,take ⬠= ââ with the coordinatewise product. Clearly ââ and â2â are isomorphic.However ââ, being commutative, is not of Cuntz type. Note also that ââ providesan example of a unital Banach algebra ⬠which (being commutative) is spectrallyregular while â(â¬), being of Cuntz type, lacks this property. Contrasting with thiswe have that ⬠is spectrally regular whenever â(â¬) is. Indeed, ⬠can be viewedas a Banach subalgebra of â(â¬) (for instance via left regular representations) andCorollary 4.1 in [13] applies.
As an immediate consequence of Theorems 6.2 and 4.1 we have the followingresult. If ð is a non-trivial Banach space, ð is an integer larger than one, and theBanach spaces ðð and ð are isomorphic, then the operator algebra â(ð) allowsfor non-trivial zero sums of idempotents (hence it is spectrally irregular). In factwe can say a bit more (cf. the second paragraph after Corollary 5.5).
Theorem 6.3. Let ð be a non-trivial Banach space, let ð be an integer larger thanone, and assume ð is isomorphic to ðð. Then, given an integer ð larger than orequal to five, there exist ð projections ð1, . . . , ðð in ð, all of infinite rank andinfinite co-rank, and such that ð1 + â â â + ðð = 0.
Proof. For this we return to the proof of Theorem 4.1 and the remark following it.The Banach algebra ⬠featuring there is now taken to be â(ð) and, whenever con-venient, complex matrices will be identified with operator matrices having entriesin â(ð), in this case via âtensorizingâ with ðŒð .
With the 2 Ã 2 matrices ð1 and ð2 and the 3Ã 3 matrices ð1, ð2 and ð3
as is Step 2 of the proof of Theorem 4.1, we have a non-trivial zero sum of five
60 H. Bart, T. Ehrhardt and B. Silbermann
projections in ð , namely
2âð=1
ðð
(ðâ(ðâ1)ð
)ðð +
3âð=1
ðð
(ðâ(ðâ1)ð
)ðð = 0,
where the bounded linear operators ðð ,ðð , ðð and ðð act as follows
ðð : ð2(ðâ1) â ð, ðð : ð â ð2(ðâ1),
ðð : ð3(ðâ1) â ð, ðð : ð â ð3(ðâ1),
and the products ðððð and ðððð yield the identity operators on ð2(ðâ1) andð3(ðâ1), respectively. Recall that, as the scalar matrices, ð1,ð2, ð1, ð2 and ð3
are rank one idempotents. Also note that the operators ðð and ðð are injective,and that ðð and ðð are surjective. It now suffices to establish the followingauxiliary result (valid because ð is infinite dimensional): if ð is a rank one ð Ã ðidempotent matrix, ð : ðð â ð is an injective linear operator, and ð : ð â ðð
is a surjective linear operator, then both the dimension of the range space of theoperator ð ð ð : ð â ð and that of its null space are infinite.
To see this, we reason as follows. Modulo similarity it may be assumed thatð has one in the (1, 1)th position and zeros everywhere else. Viewed as an op-erator matrix, ð then has ðŒð in the (1, 1)th position and the zero operator onð everywhere else. From this (using the infinite dimensionality of ð) it fol-lows that the range of ð , written Imð , and the null space of ð , denoted byKerð , both have infinite dimension. Of course ð is viewed here as an operatoron ðð. As ð is surjective, Imð ð ð = ð [Imð ], and the injectivity of ð givesthat Imð ð ð and Imð have the same dimension. Again using the injectivityof ð , we get Kerð ð ð = ðâ1[Kerð ]. The surjectivity of ð now implies thatð [Kerð ð ð ] = Kerð and we see that the dimension of Kerð does not exceedthat of Kerð ð ð . Thus both the dimension of ImVRW and that of Kerð ð ðare infinite.
This covers the case ð = 5. For ð > 5, use a similar argument and theconstruction described in Step 7 of the proof of Theorem 4.1. â¡
The following corollary will be used at several places in the next section (forinstance in the proof of Theorems 7.2 and 7.4).
Corollary 6.4. Let ð and ð be non-trivial Banach spaces, let ð be isomorphic toa complemented subspace of ð, and suppose ð is isomorphic to ð ð where ð isan integer larger than one. Then, given an integer ð larger than or equal to five,there exist ð projections ð1, . . . , ðð in ð, all of infinite rank and infinite co-rank,and such that ð1 + â â â + ðð = 0.
In particular â(ð) is spectrally irregular.Proof. We may assume that ð is a complemented subspace of ð so that thesituation is as in the proof of Proposition 5.6. Let Ί : â(ð ) â â(ð) be theinjective continuous Banach algebra homomorphism constructed there. Clearly
Sums of Idempotents and Spectral Regularity of Banach Algebras 61
maps projections in ð of infinite rank into projections of infinite rank inð , and theanalogue of this for co-ranks is valid too. By Theorem 6.3 there existð projectionsð1, . . . , ðð in ð , all of infinite rank and infinite co-rank, and such that ð1+ â â â +ðð = 0. For ð = 1, . . . ,ð, put ðð = Ί(ðð). Then the projections ð1, . . . , ðð havethe desired properties. â¡
7. Applications to specific Banach spaces
We will now use the material presented in the previous section to identify certainBanach spaces ð for which the operator algebra â(ð) allows for non-trivial zerosums, hence is spectrally irregular. In most cases this will be done by showing thatâ(ð) is of Cuntz type so that Theorem 4.1 applies; occasionally we will need torefer to Corollary 6.4.
For Σ a non-empty set, 1 †ð †â, and ð a Banach space, let âð(Σ;ð) denotethe Banach space of all âð-functions from Σ into ð, i.e., the functions ð : Σ â ðfor which the following quantities are finite: in case ð =â,
â¥ðâ¥â = supðâΣ
â¥ð(ð)â¥ð ,
in case 1 †ð <â,
â¥ðâ¥ð =(
supð¹ finite subset of Σ
âðâð¹
â¥ð(ð)â¥ðð) 1
ð
.
Here ⥠â â¥ð stands for the (given) norm on ð. With the usual algebraic operationsand the norm given by the expressions above, âð(Σ;ð) is a Banach space.
Theorem 7.1. Let Σ be an infinite set, let 1 †ð †â, and let ð be a Banachspace. Then the Banach space âð(Σ;ð) is isomorphic to its square âð(Σ;ð)
2. Also,when ð is non-trivial, the operator algebra â(âð(Σ;ð)) has the Cuntz 2-property.
Proof. By Theorem 6.2 it is sufficient to prove the first part of the theorem. Write Σas the disjoint union of Σ1 and Σ2 of two sets both sets having the same cardinalityas Σ. This is possible by virtue of the basic set theoretical result saying that thesum of two infinite cardinalities is that same cardinality again. Let ð1 : Σâ Σ1 and
ð2 : Σ â Σ2 be bijections. Now define ð : âð(Σ;ð)â(âð(Σ;ð)
)2by stipulating
that ðð = (ð1ð,ð2ð)†where ðð : âð(Σ;ð)â âð(Σ;ð) is given by
ððð = ð â£Î£ð â ðð : Σâ ð, ð = 1, 2.
Also introduce ð :(âð(Σ;ð)
)2 â âð(Σ;ð))by ð (ð, ð)†= ð1ð + ð2ð with
ð1ð â£Î£1 = ð â ðâ11 , ð1ð â£Î£2 = 0, ð2ðâ£Î£2 = ð â ðâ12 , ð2ðâ£Î£1 = 0.
Then ð and ð are each others inverse. â¡
62 H. Bart, T. Ehrhardt and B. Silbermann
As a particular case of Theorem 7.1 we have that, for 1 †ð †â, the operatoralgebras â(âð) are of Cuntz type. Hence they allow for non-trivial zero sums ofidempotents and are spectrally irregular. The case ð = 2 was already covered in[6]; see also [5].
In the same vein, by pasting together two copies of the real line â, it becomesclear that ð¿ð(â) is isomorphic to a direct sum of two copies of itself. Thus â(ð¿ð(â)
)has the Cuntz 2-property. Possible generalizations of this result involve measurespaces (Σ, ð) such that two or more copies of (Σ, ð) can be combined into a measurespace which is equivalent to one copy of (Σ, ð), in the sense that there exists abijective measurable function whose inverse is measurable too.
As is usual, the subspaces of ââ consisting of those complex sequences havinga limit, respectively limit zero, are denote by ð, respectively ð0. Clearly ð0 is iso-morphic to the direct sum ð20 of two copies of itself. Thus Theorem 6.2 guaranteesthat â(ð0) has the Cuntz 2-property. But then â(ð0) allows for non-trivial zerosums, hence is spectrally irregular. The same conclusion holds for â(ð). This canbe derived from Corollary 6.4 upon noting that ð0 is a complemented subspace ofð (having codimension one).
The above observations concerning ð0 and â(ð0) can be brought into a moregeneral context. Let ð0(Σ;ð) be the (closed) subspace of ââ(Σ;ð) consisting of thefunctions ð from Σ into ð having the following property: for each ð > 0 there exista finite subset ð¹ â Σ (depending on ð) such that â¥ð(ð¡)â¥ð < ð for all ð¡ â Σ âð¹ .Here, as before, ⥠â â¥ð denotes the norm on the Banach space ð.
Theorem 7.2. Let Σ be an infinite set and let ð be a Banach space. Then theBanach space ð0(Σ;ð) isomorphic to its square ð0(Σ;ð)
2. Also, when ð is non-trivial, the operator algebra â(ð0(Σ;ð)) has the Cuntz 2-property.
The proof of Theorem 7.2 follows the same line of thought as the argumentgiven for Theorem 7.1.
We can also obtain an analogue of ð by stipulating that ð(Σ;ð) is the (closed)subspace of ââ(Σ;ð) consisting of the functions ð : Σ â ð having the followingproperty: there exists ð§ â ð (depending on ð) such that for each ð > 0 there existsa finite subset ð¹ â Σ (depending on ð) with â¥ð(ð) â ð§â¥ð < ð for all ð â Σ âð¹ .Clearly ð0(Σ;ð) is a closed subspace of ð(Σ;ð). Another closed subspace of ð(Σ;ð)
is the set ð of all constant functions on Σ (which can be viewed as a copy of ð).
As ð(Σ;ð) = ð0(Σ;ð) â ð we have that ð0(Σ;ð) is complemented in ð(Σ;ð).Thus by Corollary 6.4, for each integer ð larger than or equal to five, the operatoralgebra â(ð(Σ;ð)) allows for zero sums involving ð non-zero idempotents, henceis spectrally irregular.
To further demonstrate the applicability of Theorem 4.1 and Corollary 6.4,we present a couple of results involving functions that are continuous on the realline â with the possible exception of the points in a certain infinite subset of âwhere jumps occur.
Let Î be an infinite subset of â having no accumulation point in â. Byð¶Î(â;ð) we denote the subspace of ââ(â;ð) consisting of all functions ð : ââ ð
Sums of Idempotents and Spectral Regularity of Banach Algebras 63
that are continuous on â except possibly in the points of Î where jumps may occur.The latter means that for each ð â Î both limð¡âð ð(ð¡) and limð¡âð ð(ð¡) exist, one orboth of them possibly different from the value of ð at ð. It is not hard to see thatð¶Î(â;ð) is closed in ââ(â;ð).
A few preliminary remarks are in order.As Î has no accumulation point in â, every compact subset of â contains
only a finite number of points of Î. Suppose Î is bounded below. Then we canfind a monotonically increasing sequence ð1, ð2, ð3, . . . of real numbers such thatlimðââ
ðð = â and Î = {ð1, ð2, ð3, . . .}. It is now easy to see that ð¶Î(â;ð) is
isomorphic to ð¶â(â;ð), where â stands for the set of non-negative integers.Next assume that Î is bounded above. Then there exists a monotonically
decreasing sequence ð1, ð2, ð3, . . . of real numbers with limðââ
ðð = ââ and Î =
{. . . , ð3, ð2, ð1}, and again it follows that ð¶Î(â;ð) is isomorphic to ð¶â(â;ð).Finally, in case Î is neither bounded below nor bounded above, there are
a monotonically increasing sequence ð1, ð2, ð3, . . . and a monotonically decreasingsequence ð1, ð2, ð3, . . . of real numbers for which ð1 < ð1,
limðââ
ðð =â, limðââ
ðð = ââ,
and Î = {. . . , ð3, ð2, ð1, ð1, ð2, ð3, . . .}. The conclusion is now that ð¶Î(â;ð) isisomorphic to ð¶â€(â;ð), where †denotes the set of all integers.
Theorem 7.3. Let Î be an infinite subset of the real line â having no accumulationpoint in â. Suppose Î is neither bounded below nor above. Then the Banach spaceð¶Î(â;ð) is isomorphic to its square ð¶Î(â;ð)
2. Also, when the Banach space ðis non-trivial, the operator algebra â(ð¶Î(â;ð)
)has the Cuntz 2-property.
Proof. It is enough to show an isomorphy between ð¶Î(â;ð) and ð¶Î(â;ð)2. As
ð¶Î(â;ð) is isomorphic to ð¶â€(â;ð), it suffices to consider the Î = â€. Write â asthe disjoint union of Σ1 and Σ2:
Σ1 =âªðââ€
(2ð, 2ð + 1], Σ2 =âªðââ€
(2ð + 1, 2ð + 2].
Also define ð1 and ð2 on â by
ð1(ð¡) = ð + ð¡, ð2(ð¡) = ð + 1 + ð¡, ð¡ â (ð, ð + 1], ð â â€.
Then ð1 : ââ Σ1 and ð2 : ââ Σ2 are bijective with inverses given by
ðâ11 (ð ) = âð + ð , ð â (2ð, 2ð + 1], ð = 0, 1, 2, . . . ,
ðâ12 (ð ) = âð â 1 + ð , ð â (2ð + 1, 2ð + 2], ð = 0, 1, 2, . . . .
Take ð in ð¶â€(â;ð). Then, along with ð , the function ð â£Î£1 â ð1 : â â ð iscontinuous on â except maybe in the points of †where jumps occur. Thus itbelongs to ð¶â€(â;ð), and the same is true for ð â£Î£2 â ð2 : â â ð. Now intro-
duce ð : ð¶â€(â;ð) â(ð¶â€(â;ð)
)2by stipulating that ðð = (ð1ð,ð2ð)
†with
64 H. Bart, T. Ehrhardt and B. Silbermann
ððð = ð â£Î£ð â ðð . Then ð is bijective. For its inverse ð from(ð¶â€(â;ð)
)2into
ð¶â€(â;ð) we have ð (ð, ð)†= ð1ð + ð2ð with
ð1ð â£Î£1 = ð â ðâ11 , ð1ð â£Î£2 = 0, ð2ðâ£Î£2 = ð â ðâ12 , ð2ðâ£Î£1 = 0,
and with this the argument is complete. â¡
Theorem 7.4. Let Î be a an infinite subset of the real line â having no accu-mulation point in â. Suppose Î is bounded below or above. Also assume ð to benon-trivial. Then, given an integer ð larger than or equal to five, there exist ðprojections ð1, . . . , ðð in ð¶Î(â;ð), all of finite rank and co-rank, and such thatð1 + â â â + ðð = 0.
In particular â(ð¶Î(â;ð))is spectrally irregular.
Proof. As ð¶Î(â;ð) is isomorphic to ð¶â(â;ð), it suffices to consider the caseÎ = â. Write ð¶â for the (closed) subspace of ð¶â(â;ð) consisting of all functionsð â ð¶â(â;ð) for which ð vanishes on (0,â). Further let ð¶+ be the (closed)subspace of ð¶â(â;ð) having as elements the functions ð â ð¶â(â;ð) such that ðvanishes on (ââ, 0]. Then ð¶â(â;ð) = ð¶â â ð¶+. Also ð¶+ is isomorphic to ð¶2
+,as can be seen via an argument analogous to the proof of Theorem 7.3. Now applyCorollary 6.4. â¡
In the case where Î is bounded below, we have the possibility to introducethe subalgebra ð¶Î,â(â;ð) of ð¶Î(â;ð) consisting of all ð â ð¶Î(â;ð) for whichlim
ð¡âââ ð(ð¡) exists in ð. This subalgebra is again closed, so we have a Banach subal-
gebra of ð¶Î(â;ð) here. We do not know whether for Î bounded below the Banachspace ð¶Î(â;ð) is isomorphic to ð¶â€(â;ð), the space ð¶Î,â(â;ð) is however.
Lemma 7.5. Let Î be a an infinite subset of the real line â having no accumula-tion point in â. Suppose Î is bounded below. Then ð¶Î,â(â;ð) is isomorphic toð¶â€(â;ð).
Proof. As in the proof of Theorem 7.4 it suffices to consider the case Î = â. Definethe function ð on â by
ð(ð¡) =
â§âšâ©log ð¡, ð¡ â (0, 1],3ð + 1 + ð¡, ð¡ â (âð â 1,âð], ð = 0, 1, 2, . . . ,
ð + ð¡, ð¡ â (ð + 1, ð + 2], ð = 0, 1, 2, . . . .
Then ð : ââ â is bijective with inverse ðâ1 : ââ â given by
ðâ1(ð ) =
â§âšâ©ðð , ð â (ââ, 0],â3ð â 1 + ð , ð â (2ð, 2ð + 1], ð = 0, 1, 2, . . . ,
âð + ð , ð â (2ð + 1, 2ð + 2], ð = 0, 1, 2, . . . .
Take ð in ð¶â,â(â;ð). Then the function ð â ð : â â ð belongs to ð¶â€(â;ð).Introduce ð : ð¶â,â(â;ð)â ð¶â€(â;ð) by stipulating that ðð = ð â ð. Then ðis bijective. For its inverse ð : ð¶â€(â;ð)â ð¶â,â(â;ð) we have ð ð = ð â ðâ1.
Sums of Idempotents and Spectral Regularity of Banach Algebras 65
Take ð in ð¶â€(â;ð). Then the function ð â ð : ââ ð belongs to ð¶â,â(â;ð).Introduce ð : ð¶â€(â;ð)â ð¶â,â(â;ð) by stipulating that ðð = ð â ð. Then ðis bijective. For its inverse ð : ð¶â,â(â;ð)â ð¶â€(â;ð) we have ð ð = ð â ðâ1. â¡
The next result is now immediate from combining Lemma 7.5 and Theo-rem 7.3.
Theorem 7.6. Let Î be an infinite subset of the real line â having no accumulationpoint in â. Suppose Î is bounded below. Then the Banach space ð¶Î,â(â;ð) isisomorphic to its square ð¶Î,â(â;ð)2. Also, when the Banach space ð is non-trivial, the operator algebra â(ð¶Î,â(â;ð)
)has the Cuntz 2-property.
Under the assumption that Î is bounded above, Lemma 7.5 and Theorem 7.6hold with ð¶Î,â(â;ð) replaced by the Banach subalgebra ð¶Î,+(â;ð) of ð¶Î(â;ð)consisting of all ð â ð¶Î(â;ð) for which lim
ð¡â+â ð(ð¡) exists in ð.
Theorems 7.3, 7.4 and 7.6 were concerned with functions on â being con-tinuous with the possible exception of jump discontinuities only. Now we turn toBanach spaces consisting of continuous functions. For ð a topological space andð a (complex) Banach space, the expression ðâ(ð;ð) will denote the Banachspace of all bounded continuous functions from ð to ð, endowed with the usualalgebraic operations (defined pointwise) and the sup-norm. Clearly ðâ(ð;ð) is asubspace of ââ(ð;ð). In case ð is compact, ðâ(ð;ð) coincides with the Banachspace ð(ð;ð) of all continuous functions from ð to ð, again endowed with theusual algebraic operations (defined pointwise) and the max-norm.
We begin with some auxiliary observations.
Proposition 7.7. Let ð be a topological space, let ð be a Banach space, and let ð be apositive integer larger than or equal to two. Suppose ð is homeomorphic to the topo-logical direct sum of ð copies of itself. Then the Banach space ðâ(ð;ð) is isomor-phic to ðâ(ð;ð)ð. Also, when ð is non-trivial, the operator algebra â(ðâ(ð;ð))has the Cuntz ð-property.
The condition on ð is met if and only if ð can be written as the disjoint unionof ð clopen (which by definition means: open and closed) sets ð1, . . . , ðð, each ofwhich as a topological subspace of ð (i.e., provided with the relative topology withrespect to ð) is homeomorphic to ð. Note that we have a fractal type structurehere. Each ðð is again the disjoint union of ð clopen sets homeomorphic to ð, andfor these this is true again. And so on, indefinitely. We shall come back to thispoint after the proof. See also Examples A, B and C below.
Proof. Let ð1, . . . , ðð be clopen sets as above. For ð = 1, . . . , ð, let ðð be a home-omorphism of ð onto ðð . Define ð : ðâ(ð;ð)â ðâ(ð;ð)ð by
ðð = (ð1ð, . . . ,ððð)â€, ððð = ð â£ðð â ðð , ð = 1, . . . , ð.
Then ð is linear and bounded. In fact, with the choice made for the norm onðâ(ð;ð)ð in the paragraph directly following Theorem 6.2, the operator ð is
66 H. Bart, T. Ehrhardt and B. Silbermann
norm preserving. It is also bijective. This can be seen as follows. For ð = 1, . . . , ð,let ðð : ðâ(ð;ð)â ðâ(ð;ð) be given by
ððð â£ðð = ð â ðâ1ð , ððð â£ðð = 0, ð = 1, . . . , ð, ð â= ð.Note that ððð â ðâ(ð;ð) because the sets ð1, . . . , ðð are clopen in ð. The inverseð : ðâ(ð;ð)ð â ðâ(ð;ð) of the operator ð is now given by ð (ð1, . . . , ðð)
†=ð1ð1 + â â â + ðððð. This proves the first statement in the proposition; the secondis immediate from Theorem 6.2. â¡
Returning to the fractal type structure mentioned prior to the above proof,we recall that ðð is again the disjoint union of ð clopen sets homeomorphic to ð,and for these clopen sets this is true again. By induction, it follows that for anypositive integer ð, the space ð is the disjoint union of ðð clopen subsets of ð, eachhomeomorphic to ð. In particular, there exists a (countably) infinite collectionof clopen subsets of ð, each homeomorphic to ð, possibly not mutually disjointhowever.
In general it is not clear whether, under the conditions of Proposition 7.7,there is an infinite collection of mutually disjoint clopen subsets of ð. A fortiori, itis not clear whether ð can be written as the disjoint union of an infinite collectionof clopen sets homeomorphic to ð. When this is possible, we have the followingresult.
Proposition 7.8. Let ð be a topological space, let ð be a Banach space, and supposeð is the disjoint union of an infinite collection of clopen sets homeomorphic to ð.Then the Banach space ðâ(ð;ð) is isomorphic to its square ðâ(ð;ð)2. Also,when ð is non-trivial, the operator algebra â(ðâ(ð;ð)) has the Cuntz 2-property.
Proof. Let Σ be an infinite index set, and assume ð is the disjoint union of thefamily {ðð}ðâΣ of clopen subsets of ð, each homeomorphic to ð. In other words,ð is the topological direct sum of the family {ðð}ðâΣ. Using the type of argumentfeaturing in the proof of Proposition 7.7, one finds that ðâ(ð;ð) is isomorphic toââ(Σ; ðâ(ð;ð)
). By Theorem 7.1, the latter space is isomorphic to its square,
but then so is ðâ(ð;ð). This establishes the first statement in the proposition;the second now comes from Theorem 6.2. â¡
By ðŸ we denote the familiar ternary Cantor set, also called the Cantormiddle-third set. Recall that ðŸ is compact, so that ðâ(ðŸ;ð) coincides withð(ðŸ;ð).Corollary 7.9. Let ð be a Banach space. Then the Banach space ð(ðŸ;ð) is iso-morphic to its square ð(ðŸ;ð)2. Also, when ð is non-trivial, the operator algebraâ(ð(ðŸ;ð)) has the Cuntz 2-property.
Proof. Let ðŸâ and ðŸ+ be the intersection of ðŸ with the closed interval [0,1/3]and [2/3, 1], respectively. Then ðŸâ and ðŸ+ are open sets in ðŸ. Also ðŸ is thedisjoint union of ðŸâ and ðŸ+. So ðŸ is the topological direct sum of ðŸâ and ðŸ+.Now note that ðŸ = 3ðŸâ and ðŸ = â2 + 3ðŸ+. It follows that both ðŸâ and ðŸ+
Sums of Idempotents and Spectral Regularity of Banach Algebras 67
are homeomorphic to ðŸ. Now apply Proposition 7.7 to get the first part of thetheorem, and Theorem 6.2 to get the second. â¡
The next result is a mild generalization of a truly remarkable result ofA.A. Miljutin [28].
Theorem 7.10. Let ð be an uncountable compact metrizable topological space andsuppose ð is a finite-dimensional Banach space. Then the Banach spaces ð(ð;ð)and ð(ðŸ;ð) are isomorphic.
Proof. Consider ðð(ð;â), the Banach space of real-valued continuous functions onð, which is the real analogue of the complex Banach space ð(ð;â). By a celebratedresult of A.A. Miljutin [28], the real Banach spaces ðð(ð;â) and ðð(ðŸ;â) are iso-morphic (see also [29], the remark below Theorem 6.2.5). A simple complexificationargument now shows that ð(ð;â) and ð(ðŸ;â) are isomorphic too. To make thestep from complex-valued functions to those having values in the (complex) finite-dimensional Banach space ð, we argue as follows. Write ð for the dimension ofð. As each ð-dimensional complex Banach space is isomorphic to âð, we mayassume that ð actually coincides with âð. For ð = 1, . . . , ð, let ð ð : â
ð â â
be the ðth coordinate function. Take ð in ð(ð;âð) and put ðð = ð ð â ð . Thenðð â ð(ð;â). As we have already seen, there exists a bijective bounded linearoperator from ð(ð;â) onto ð(ðŸ;â). Let us denote it by ðµ. For ð â ðŸ, now write
ðð(ð) =(ðµð1(ð), . . . , ðµðð(ð)
)â€. Then ð (ð) belongs to ð(ðŸ;âð) and we have a
mapping ð : ð(ð;âð) â ð(ðŸ;âð). This mapping is easily seen to be a bijectivebounded linear operator. â¡
Theorem 7.11. Let ð be an uncountable compact metrizable topological space andsuppose ð is a non-trivial finite-dimensional Banach space. Then the operatoralgebra â(ð(ð;ð)) has the Cuntz 2-property.
Proof. The Banach spaces ð(ð;ð) and ð(ðŸ;ð) are isomorphic. Hence the sameis true for â(ð(ð;ð)) and â(ð(ðŸ;ð)). The desired result is now immediate fromCorollary 7.9. â¡
In the next result, certain notions from general topology play a role. Hereare the pertinent definitions taken from [29]. A non-empty topological space ð iscalled topologically complete if there exists a complete metric on ð which generatesthe topology of ð. The space is said to be nowhere locally compact if no point ofð has a neighborhood with compact closure. Finally, ð is called zero-dimensionalif ð has a base consisting of clopen sets.
Theorem 7.12. Let ð be a non-empty topological space which is topologically com-plete, nowhere locally compact and zero-dimensional. Further let ð be a Banachspace. Then the Banach space ðâ(ð;ð) is isomorphic to its square ðâ(ð;ð)2.Also, in case ð is non-trivial, the operator algebra â(ð(ð;ð)) has the Cuntz 2-property.
68 H. Bart, T. Ehrhardt and B. Silbermann
Proof. By a result of P. Alexandroff and P. Urysohn [3], cited as Theorem 1.9.8in [29], the space ð is homeomorphic to the space â of all irrational numbers. Letâ1 be the set of negative irrational numbers, and let â2 be the set of the positiveones. Then â1 and â2 are clopen sets in â and â is the disjoint union of â1 andâ2. Also both â1 and â2 are homeomorphic to â. This is clear from the fact thatthese spaces satisfy the conditions mentioned in the theorem, but it can also beseen in a more direct way by constructing concrete homeomorphisms ð1 from â
onto â1 and ð2 from â onto â2. For the latter on can take for instance ð2 : ââ â2given by
ð2(ð¡) =
â§âšâ©1
1â ð¡ , ð¡ â â, ð¡ < 0,
ð¡+ 1, ð¡ â â, ð¡ > 0.
Now apply Proposition 7.7. â¡Theorem 7.13. Let ð be a non-empty topological space with a countable number ofpoints and no isolated points. Further let ð be a Banach space. Then the Banachspaces ðâ(ð;ð) and ðâ(ð;ð)2 are isomorphic. Also, when ð is non-trivial, theoperator algebra â(ð(ð;ð)) has the Cuntz 2-property.
Proof. By a result of W. Sierpinski [33] (see Theorem 1.9.6 in [29]), the space ðis homeomorphic to â, the space of rational numbers. The argument for dealingwith â is similar to that given in the proof of the preceding result for â; use, forinstance,
â2 as a âdivision pointâ instead of 0. â¡
In each of the above specializations to concrete spaces, we have isomorphy ofa Banach space with its square (hence with all its powers). In order to prove this,we needed some rather non-trivial results from general topology. In some casesone can avoid this by settling for something less, for instance isomorphy with thecube instead of the square. An example is given below. More on Banach spacesisomorphic to their cubes can be found in the next section.
Example A. Let ð be the subset of the open interval (â1, 1) consisting of therational numbers ââ
ð=1
3ðð
(1
4
)ð
, (11)
where the ðð are allowed to take the numerical values â1, 0, 1 (and no others),while only a finite number among ð1, ð2, ð3, . . . may differ from zero, so that (11)is actually a finite sum. To get an idea of what is going on, let us look at a fewcases.
The first is where all ðð vanish. Then (11) only gives the number 0. Nextconsider the situation where ðð = 0 for all ð ⥠2. This leads to the three numbersâ 3
4 , 0 and34 . When ðð = 0 for ð ⥠3, the sum (11) reduces to 3
4ð1 +316 ð2 with
ð1, ð2 â {â1, 0, 1}, and so we arrive at the nine numbersâ1516, â3
4, â 9
16, â 3
16, 0,
3
16,
9
16,3
4,15
16.
Sums of Idempotents and Spectral Regularity of Banach Algebras 69
These include the three outcomes we already had in the previous stage. In caseðð vanishes for all ð ⥠4, the sum (11) becomes 3
4ð1 +316 ð2 +
364 ð3 with the
restrictions stipulated above on ð1, ð2 and ð3. Thus, besides the nine numbersindicated above, we get eighteen additional ones, making a total of twentyseven:
â6364, â15
16, â57
64, â51
64, â3
4, â45
64, â39
64, â 9
16, â33
64,
â1564, â 3
16, â 9
64, â 3
64, 0,
3
64,
9
64,
3
16,
15
64,
33
64,
9
16,
39
64,
45
64,
3
4,
51
64,
57
64,
15
16,
63
64.
And so on.
The number of points in ð is countably infinite and ð, being a subspace ofthe real line, has a countable base. Also it is a straightforward matter to provethat ð has no isolated points. For ð a non-trivial Banach space, Theorem 7.13now gives that the operator algebra â(ð(ð;ð)) has the Cuntz 2-property. Theproof given above employs Sierpinskyâs (rather non-trivial) characterization of âas the unique non-empty countable space without isolated points. Being contentwith the Cuntz 3-property, we can avoid the use of heavy machinery from generaltopology. Here is the argument.
Split ð in three parts ðâ, ð0 and ð+:
ðâ = ð â©(â1,â1
2
), ð0 = ð â©
(â14,1
4
), ð+ = ð â©
(1
2, 1
).
These parts correspond to ð1 = â1, ð1 = 0 and ð1 = 1, respectively. Clearly ðâ, ð0and ð0 are open subsets of ð which itself is the dsjoint union of these sets. But thenðâ, ð0 and ð0 are closed in ð too. Now note that ð = 3+4ðâ = 4ð0 = â3+4ð+.Hence ð is homeomorphic to the topological direct sum of three copies of itself.Applying Proposition 7.7 we get that the operator algebra â(ðâ(ð;ð)) has theCuntz 3-property. â¡
The material presented above, is concerned with non-trivial Banach spacesthat are isomorphic to their squares and, consequently, allow for non-trivial zerosums of projections. For such a Banach space ð there exist âCuntz operatorsâð1, ð2,ð1 and ð2 in â(ð) satisfying the identities
ð1ð1 + ð2ð2 = ðŒð , ðððð = ð¿ð,ððŒð , ð, ð = 1, 2.
For several of the above concrete instances of Banach spaces isomorphic to theirsquare or a higher power we gave (or would be able to give) explicit descriptionsof these Cuntz operators (cf. Theorems 7.1 and 7.6; see also Proposition 7.7 andExample A). In combination with the material presented in Section 4, such descrip-tions, when available, can be used to obtain explicit expressions for the projectionsforming a non-trivial zero sum. The expressions in question are complicated andnot very illuminating; we refrain from giving further details here.
70 H. Bart, T. Ehrhardt and B. Silbermann
8. Banach spaces isomorphic to their cubes
The applications we gave in the previous section were concerned with Banachspaces having the Cuntz 2-property. However, this does not cover all possiblecases. Indeed, Theorem 10 in [21] provides an example of a Banach space which isisomorphic to its cube while it is not isomorphic to its square. More generally itis shown in [23] that for every integer ð ⥠2, there is a Banach space ðž such thatðžð is isomorphic to ðžð if and only if ð = ð (mod ð). In particular the space ðžis then not isomorphic to ðžð, but it is isomorphic to ðžð+1, and the latter impliesthat â(ðž) is of Cuntz type, hence spectrally irregular because of the occurrenceof non-trivial zero sums of idempotents. The conclusion is that one needs the fullforce of Theorem 6.2.
The examples of the type meant above are complicated. In this section weembark on a somewhat less ambitious endeavor: to construct Banach spaces, evi-dently isomorphic to their cubes, but for which it is not clear whether or not theyare isomorphic to their square. For a given integer ð larger than 2, the constructioncan be modified such as to result in Banach spaces ð¹ with ð¹ð isomorphic to ð¹ð ifð = ð (mod ð) while it is unclear whether or not ð¹ð is isomorphic to ð¹ð in caseð â= ð (mod ð). We refrain from giving the details concerning this refinement.
We now begin with the construction which, as one will realize, is inspiredby that of the familiar Cantor set. The starting point is a non-empty topologicalspace ð , not necessarily compact or metrizable. Suppose ðâ, ð0 and ð+ are mutu-ally disjoint subspaces of ð , all three homeomorphic to ð , hence non-empty. Letðð : ð â ðð be a homeomorphism from ð onto ðð . Here ð â {â, 0,+}. Forð = 1, 2, 3, . . . and ð1, ð2, . . . , ðð â {â, 0,+}, write
ðð1,ð2,...,ðð = ðð1ðð2 . . . ððð [ð ]. (12)
Note that, as far as the expressions ðâ, ð0 and ð+ are concerned, no confusion ispossible. Indeed, the sets ðâ = ðâ[ð ], ð0 = ð0[ð ] and ð+ = ð+[ð ] coming from(12), coincide with the originally given ðâ, ð0 and ð+. Clearly, all the subspacesðð1,ð2,...,ðð are non-empty,
ðð1,ð2,...,ðð â ðð1,ð2,...,ððâ1 â â â â â ðð1,ð2 â ðð1 ââª
ðâ{â,0,+}ðð ,
and ðð1,ð2,...,ðð = ðð1 [ðð2,...,ðð ]. We also have
ðð1,ð2,...,ðð â© ðð =
{ðð,ð2,...,ðð , ð = ð1,
â , ð â= ð1.Note further that
ðð1,ð2,...,ðð â© ðð1,ð2,...,ðð â= â â ðð = ðð, ð = 1, . . . , ð,
and so ðð1,ð2,...,ðð and ðð1,ð2,...,ðð coincide if and only if they are not disjoint.
Sums of Idempotents and Spectral Regularity of Banach Algebras 71
For ð = 1, 2, 3 . . ., introduce
ðð =âª
ð1,ð2,...,ðð â{â,0,+}ðð1,ð2,...,ðð .
Then ðð is non-empty and ð1 = ðâ ⪠ð0 ⪠ð+. Regardless of whether ðð+1 isâ, 0 or +, the inclusion ðð1,ð2,...,ðð,ðð+1 â ðð1,ð2,...,ðð holds. Hence
ðð+1 ââª
ð1,ð2,...,ðð â{â,0,+}ðð1,ð2,...,ðð = ðð,
so ð1 â ð2 â ð3 â . . . . We also note that ðð+1 = ðâ[ðð] ⪠ð0[ðð] ⪠ð+[ðð],and this true for ð = 0 too when we interpret ð0 as ð . The identity
ðð [ðð] = ðð+1 â© ðð, ð â {â, 0,+}, ð = 1, 2, 3, . . . ,
needed later, is now immediate.Let ð =
â©âð=1 ðð be the intersection of the descending chain of sets ð1, ð2,
ð3, . . . . In order to directly relate ð to the sets ðð1,ð2,...,ðð we do the following.Write ð¥ for the collection of (infinite) sequences with entries from {â, 0,+}. Withan element {ð1, ð2, ð3 . . .} from ð¥ , we associate the intersection of the descendingsequence ðð1 â ðð1,ð2 â ðð1,ð2,ð3 â . . . , i.e., the set
ðð1,ð2,ð3,... =
ââ©ð=1
ðð1,ð2,...,ðð .
As ðð1 â ð1, ðð1,ð2 â ð2, ðð1,ð2,ð3 â ðð and so on, we have that ðð1,ð2,ð3,... â ð.So the union of all the sets ðð1,ð2,ð3,... is contained in ð. In fact there is equality:
ð =âª
{ð1,ð2,ð3,...} âð¥ðð1,ð2,ð3,... =
âª{ð1,ð2,ð3,...}âð¥
ââ©ð=1
ðð1,ð2,...,ðð .
For completeness we mention that, given ð â ð, there is precisely one sequence{ð1, ð2, ð3, . . .} â ð¥ such that ð â ðð1,ð2,ð3,... . As we shall see below in an example,it may happen that the set ðð1,ð2,ð3,... contains more than one point. It can beempty too.
Next introduce ðâ = ð â©ðâ, ð0 = ð â©ð0 and ð+ = ð â©ð+. Then ð, beinga subset of the disjoint union ðââªð0âªð+, is the disjoint union of ðâ, ð0 and ð+.Also, for ð â {â, 0,+} we have (using the injectivity of ðð in the second equalitybelow)
ðð [ð] = ðð
[ ââ©ð=1
ðð
]=
ââ©ð=1
ðð[ðð]=
ââ©ð=1
[ðð+1 â© ðð
]=
[ ââ©ð=1
ðð+1
]â© ðð = ð â© ðð = ðð.
Hence the restriction ðð of ðð to ð, viewed as a mapping ðð : ð â ðð , is ahomeomorphism from ð onto ðð . Consequently, for ð, ð â {â, 0,+}, the mappingððð
â1ð : ðð â ðð is a homeomorphism from ðð onto ðð . For completeness we
72 H. Bart, T. Ehrhardt and B. Silbermann
mention that ðâ1ð : ðð â ð is the restriction of ðâ1ð to ðð, considered as a mappingonto ð.
The statements in the previous paragraph are only of interest when the setð is non-empty. A relevant special case in which this necessarily holds is when theunderlying topological space ð is compact and the sets ðâ, ð0 and ðâ are closedin ð . In that case we can even deduce the non-emptiness of all sets ðð1,ð2,ð3,...
with the sequence {ð1, ð2, ð3, . . .} taken from ð¥ . Closedness of ðâ, ð0 and ðâ isguaranteed when ð , in addition to being compact, is also Hausdorff.
In order to conclude that ð is non-empty, sometimes fixed point theoremscan be employed too. Here is such a case, which applies, for instance, to non-emptyclosed subset ð of real or complex Banach spaces. Suppose ð is a complete metricspace (possibly non-compact) and let ð denote the metric on ð . Further assumethat the homeomorphisms ðâ : ð â ðâ, ð0 : ð â ð0 and ð+ : ð â ð+ arecontractions and there exists a constant ð â (0, 1) such that
ð(ðð(ð¥), ðð(ðŠ)
) †ðð(ð¥, ðŠ), ð¥, ðŠ â ð.Take ð â {â, 0,+}. Then by the Banach fixed point theorem, ðð has a (unique)fixed point. Now let ð¢â,â, ð¢0,â and ð¢+,â be the fixed points of ðâ, ð0 and ð+,respectively. Clearly these belong to ðâ = ðâ[ð ], ð0 = ð0[ð ] and ð+ = ð+[ð ],respectively. Note now that ð¢â,â â ðâ,â,â,... â ð, ð¢0,â â ð0, 0, 0, ... â ð andð¢+,â â ð+,+,+,... â ð. In particular ð is non-empty.
In what follows we shall assume that ð is non-empty. Along with ð, thehomeomorphic images ðâ, ð0 and ð+ of ð under, respectively, the homeomor-phisms ðâ = ðââ£ð : ð â ðâ, ð0 = ð0â£ð : ð â ð0 and ð+ = ð+â£ð : ð â ð+,are then non-empty as well. Recall now that ð is the disjoint union of ðâ, ð0 andð+. Thus, if the latter three sets happen to be clopen in ð, we have that ð ishomeomorphic to the topological direct sum of three copies of itself. But then,given a Banach space ð, we can conclude from Proposition 7.7 that the Banachspace ðâ(ð;ð) is isomorphic to its cube ðâ(ð;ð)3, hence the operator algebraâ(ðâ(ð;ð)) has the Cuntz 3-property and is (therefore) spectrally irregular. Itis not clear whether or not ðâ(ð;ð) is isomorphic to its square ðâ(ð;ð)2. Thismight among other things, depend on the choice of the Banach space ð.
The requirement, featuring in the above paragraph, that ðâ, ð0 and ð+ areclopen in ð, is met when the three sets ðâ, ð0 and ð+ are open in the underlyingspace ð . Here is the argument. Clearly under this assumption ðâ = ð â©ðâ, ð0 =ðâ©ð0 and ð+ = ðâ©ð+ are open in ð. But then ðâ = ð â [ð0âªð+] is closed in ð.Similarly ð0 and ð+ are closed in ð too. The same type of reasoning shows thatðâ, ð0 and ð+ are clopen in ð whenever ðâ, ð0 and ð+ are closed in ð .
By way of illustration, we now present an example in which the underlyingtopological space ð is not compact.
Example B. Let ð be the open interval (0, 1) and take for ðâ, ð0 and ð+ the openintervals
ðâ =(0,1
3
), ð0 =
(1
3,2
3
), ð+ =
(2
3, 1
).
Sums of Idempotents and Spectral Regularity of Banach Algebras 73
Further, define ðâ : ð â ðâ, ð0 : ð â ð0 and ð+ : ð â ð+ by
ðâ(ð¡) =1
3ð¡, ð+(ð¡) =
2
3+1
3ð¡, 0 < ð¡ < 1,
ð0(ð¡) =
â§âšâ©
1
3+1
5ð¡, 0 < ð¡ <
5
12,
ð¡,5
12†ð¡ †7
12,
7
15+1
5ð¡,
7
12< ð¡ < 1.
Obviously these mappings are homeomorphisms. Observe that ð0 acts as the iden-tity mapping on the closed interval
[512 ,
712
]. Hence this interval is contained in (and
actually equal) to the set ð0, 0, 0, ... (notation as above). But then the closed in-terval
[512 ,
712
]is a subset of every set ðð and so
[512 ,
712
] â ð. In particular ðis non-empty, in fact even uncountable. As the sets ðâ, ð0 and ð+ are open inð , we may conclude that, given a Banach space ð, the Banach space ðâ(ð;ð)is isomorphic to its cube ðâ(ð;ð)3, hence the operator algebra â
(ðâ(ð;ð)) hasthe Cuntz 3-property and is (therefore) spectrally irregular. It is not clear whetheror not ðâ(ð;ð) is isomorphic to its square ðâ(ð;ð)2. This is an open ques-tion, even for the case ð = â, so for the Banach space ðâ(ð;â). Note here thatTheorem 7.13 does not apply because ð is uncountable. Further Theorem 7.12cannot be used for the space ð, containing the closed interval
[512 ,
712
], is not
nowhere locally compact. For that matter, it is not zero-dimensional neither. Fi-nally, Theorem 7.11 cannot be employed to show that â(ðâ(ð;â)) has the Cunz2-property. The reason is that ð is not compact. Indeed, ð is not a closed subsetof the real line; zero is an accumulation point of ð which does not belong to ð.
As was said before, the construction presented above is inspired by that ofthe familiar Cantor set, earlier denoted by ðŸ. It is illuminating to observe that, infact, a slight modification of ðŸ is a subset of ð. The set in question, here denotedby ðŸâ, is obtained as follows. Start with the open interval (0, 1) and leave out theclosed middle third interval
[13 ,
23
]. What remains is the union of the two open
intervals(0, 13)and
(23 , 1). Next in each of them cut out the closed intervals
[19 ,
29
]and
[79 ,
89
], which leaves us with the four open intervals
(0, 19),(29 ,
13
),(23 ,
79
)and(
89 , 1). And so on (formal definition by induction), resulting in
ðŸâ = (0, 1) âââª
ð=1
3ðâ1âªð=1
[3ð â 23ð
,3ð â 13ð
].
Clearly ðŸâ is contained in the usual Cantor middle-third set ðŸ which admits therepresentation
ðŸ = [0, 1] âââª
ð=1
3ðâ1âªð=1
(3ð â 23ð
,3ð â 13ð
).
74 H. Bart, T. Ehrhardt and B. Silbermann
Along with ðŸ, the set ðŸâ is uncountable. Indeed,
ðŸ âðŸâ â {0, 1} âªââª
ð=1
3ðâ1âªð=1
{3ð â 23ð
,3ð â 13ð
},
and the set in the right-hand side of this inclusion is countable. (Actually, ðŸâcoincides with the complement in ðŸ of the countable subset of ðŸ consisting of theend points of the intervals left out.) One sees that ðŸâ â ð by ignoring the presenceof ð0 and ð0. More precisely, by looking at the sets ðð1,ð2,...,ðð with ð1, . . . , ðð takenfrom {â,+}, so avoiding the use of the index 0. In sharp contrast to ðŸ and ðŸâ,the Cantor type set ð contains (countably many) closed intervals.
The topological space ðŸâ is neither countable nor compact. So neitherTheorem 7.13 nor Theorem 7.11 applies. However, as ðŸâ is homeomorphic tothe topological direct sum of two copies of itself, we can conclude from Proposi-tion 7.7 that for every Banach space ð, the Banach space ðâ(ðŸâ;ð) is isomor-phic to its square ðâ(ðŸâ;ð)2. Hence, when ð is non-trivial, the operator algebraâ(ðâ(ðŸâ;ð)) has the Cuntz 2-property and is (therefore) spectrally irregular.Theorem 7.12 does apply, and its proof indicates that ðŸâ is homeomorphic to â,the space of irrational real numbers. â¡
We give another example, this one along lines suggested by Van Mill.
Example C. Start with the half closed, half open square ð = [0, 1]Ã (0, 1). Thenfor ðâ, ð0 and ð+, take
ðâ =[0,1
5
]Ã (0, 1), ð0 =
[2
5,3
5
]Ã (0, 1), ð+ =
[4
5, 1
]Ã(2
5,3
5
).
Further, for 0 †ð¥ †1, 0 < ðŠ < 1, put
ðâ(ð¥,ðŠ)=(1
5
âð¥,ðŠ
), ð0(ð¥,ðŠ)=
(2
5+1
5ð¥,ðŠ
), ð+(ð¥,ðŠ)=
(4
5+1
5ð¥,2
5+1
5ðŠ
).
Then ðâ : ð â ðâ, ð0 : ð â ð0 and ð+ : ð â ð+ are homeomorphisms.
ðâ
ᅵᅵ
ð0
ᅵᅵ
ð+
ᅵᅵ
ð ðâ ð0 ð+
Sums of Idempotents and Spectral Regularity of Banach Algebras 75
By repeated application of these homomorphisms, the sets ðâ, ð0, ð+ areâsqueezedâ into, respectively,
the half closed, half open rectangle
[0,
1
25
]Ã (0, 1), (= ðâ,â,â, ...),
the open line segment
{1
2
}Ã (0, 1), (= ð0, 0, 0, ...),
the singleton set
{(1,1
2
)}, (= ð+, +,+, ...).
The topological space ð resulting from the construction described above isuncountable but not compact, so Theorems 7.13 and 7.11 do not apply. Note thatTheorem 7.12 does not apply either. As ðâ, ð0 and ð+ are closed in ð , the spaceð is homeomorphic to the topological direct sum of three of its copies. So, if ð isa Banach space, then ðâ(ð;ð) is isomorphic to ðâ(ð;ð)3. It is not clear whetheror not ðâ(ð;ð) is isomorphic to ðâ(ð;ð)2. So on this basis, and assuming ð tobe non-trivial, we can conclude that â(ðâ(ð;ð)) has the Cuntz 3-property, butnot (yet) that it has the Cuntz 2-property. â¡
We close with one more remark. Both Examples B and C feature a topologicalspace ð such that (evidently!) ð is homeomorphic to the topological direct sumof three copies of ð while it is not clear whether or not ð is homeomorphic to thetopological direct sum of two copies of ð. There do exist topological spaces that arehomeomorphic to the direct sum of three copies of itself but not to that of two. Onesuch example is given in [26]. However, that example is not of primary interest tous because the conditions of Theorem 7.10 are satisfied. This is different for a spaceconstructed by W. Hanf for whose description (modulo Stone duality, i.e., in the
76 H. Bart, T. Ehrhardt and B. Silbermann
language of Boolean algebras) we refer to [27], Section 6.2 (see also [25], [32], [35]and [36]). That space, here for the moment denoted by ð , although uncountable,compact and Hausdorff, is not metrizable; hence Theorem 7.10 does not apply.The question arises whether or not ðâ(ð ;â) or, more generally, ðâ(ð ;ð) withð a Banach space, is isomorphic to its square. Evidently it is isomorphic to itscube. Recall here (cf. the first paragraph of this section) that a very sophisticatedexample of a Banach space homeomorphic to its cube but not to its square hasbeen given by W.T. Gowers [21].
Acknowledgement
The authors gratefully acknowledge stimulating contacts with Jan van Mill fromthe Free University in Amsterdam about the subject matter of Sections 7 and 8of the present paper.
References
[1] S.A. Argyros, D. Freeman, R.G. Haydon, E. Odell, Th. Raikoftsalis, Th. Schlump-recht, D. Zisimopoulou, Embedding uniformly convex spaces into spaces with veryfew operators. J. Funct. Anal. 262 (2012), 825â849.
[2] S.A. Argyros, R.G. Haydon, A hereditarily indecomposable ââ-space that solvesthe scalar-plus-compact problem, Acta Math. 206 (2011), 1â54.
[3] P. Alexandroff and P. Urysohn, Uber nulldimensionale Punktmengen, Math. Ann.98 (1928), 89â106.
[4] B.A. Barnes, G.J. Murphy, M.R.F. Smyth, T.T. West, Riesz and Fredholm Theoryin Banach Algebras, Research Notes in Mathematics, Vol. 67, Pitman (AdvancedPublishing Program), Boston, London, Melbourne 1982.
[5] H. Bart, T. Ehrhardt, B. Silbermann, Zero sums of idempotents in Banach algebras,Integral Equations and Operator Theory 19 (1994), 125â134.
[6] H. Bart, T. Ehrhardt, B. Silbermann, Logarithmic residues in Banach algebras,Integral Equations and Operator Theory 19 (1994), 135â152.
[7] H. Bart, T. Ehrhardt, B. Silbermann, Sums of idempotents and logarithmic residuesin matrix algebras, In: Operator Theory: Advances and Applications, Vol. 122,Birkhauser, Basel 2001, 139â168.
[8] H. Bart, T. Ehrhardt, B. Silbermann, Logarithmic residues of Fredholm operatorvalued functions and sums of finite rank projections, In: Operator Theory: Advancesand Applications, Vol. 130, Birkhauser, Basel 2001, 83â106.
[9] H. Bart, T. Ehrhardt, B. Silbermann, Logarithmic residues of analytic Banach alge-bra valued functions possessing a simply meromorphic inverse, Linear Algebra Appl.341 (2002), 327â344.
[10] H. Bart, T. Ehrhardt, B. Silbermann, Sums of idempotents in the Banach algebragenerated by the compact operators and the identity, In: Operator Theory: Advancesand Applications, Vol. 135, Birkhauser, Basel 2002, 39â60.
Sums of Idempotents and Spectral Regularity of Banach Algebras 77
[11] H. Bart, T. Ehrhardt, B. Silbermann, Logarithmic residues in the Banach algebragenerated by the compact operators and the identity, Mathematische Nachrichten268 (2004), 3â30.
[12] H. Bart, T. Ehrhardt, B. Silbermann, Trace conditions for regular spectral behaviorof vector-valued analytic functions, Linear Algebra Appl. 430 (2009), 1945â1965.
[13] H. Bart, T. Ehrhardt, B. Silbermann, Spectral regularity of Banach algebras and non-commutative Gelfand theory, In: H. Dym et al. (eds.): Operator Theory: Advancesand Applications. The Israel Gohberg Memorial Volume, Vol. 218, Birkhauser, Basel2012, 123â153.
[14] H. Bart, T. Ehrhardt, B. Silbermann, Families of homomorphisms in non-com-mutative Gelfand theory: comparisons and counterexamples, In: W. Arendt et al.(eds.), Spectral Theory, Mathematical System Theory, Evolution Equations, Differen-tial and Difference Equations, Operator Theory: Advances and Applications, OT 221,Birkhauser, Springer Basel AG, 2012, 131â160.
[15] H. Bart, T. Ehrhardt, B. Silbermann, Logarithmic residues, Roucheâs Theorem, spec-tral regularity, and zero sums of idempotents: the ð¶â-algebra case, Indag. Math. 23(2012), 816â847.
[16] D.S. Bernstein, Matrix Mathematics, Second Edition, Princeton University Press,Princeton and Oxford, 2009.
[17] J. Cuntz, Simple ð¶â-algebras generated by isometries, Commun. Math. Physics, 57(1977), 173â185.
[18] K.R. Davidson, ð¶â-algebras by Example, Fields Institute Monographs, 6. AmericanMathematical Society, Providence, Rhode Island, 1996.
[19] T. Ehrhardt, V. Rabanovich, Yu. Samoılenko, B. Silbermann, On the decompositionof the identity into a sum of idempotents, Methods Funct. Anal. Topology 7 (2001),1â6.
[20] I. Gohberg, S. Goldberg, M.A. Kaashoek, Classes of Linear Operators, Vol. I, Op-erator Theory: Advances and Applications, Vol. 49, Birkhauser, Basel 1990.
[21] W.T. Gowers, A solution to the Schroeder-Bernstein problem for Banach spaces,Bull. London Math. Soc. 28, No. 3 (1996), 297â304.
[22] W.T. Gowers, B. Maurey, The unconditional basic sequence problem, Journal A.M.S.6 (1993), 851â874.
[23] W.T. Gowers, B. Maurey, Banach spaces with small spaces of operators, Math. Ann.307 (1997), 543â568.
[24] R. Hagen, S. Roch, B. Silbermann, ð¶â-algebras and Numerical Analysis, MarcelDekker, New York, 2001.
[25] W. Hanf, On some fundamental problems concerning isomorphism of Boolean alge-bras, Math. Scand. 5 (1957),205â217.
[26] J. Ketonen, The structure of countable Boolean algebras, Ann.of Math. (2) 108,No. 1 (1978), 41â89.
[27] S. Koppelberg, Handbook of Boolean algebras, Vol. 1 (J.D. Monk and R. Bonnet,eds.), North-Holland Publishing Co., Amsterdam, 1989.
78 H. Bart, T. Ehrhardt and B. Silbermann
[28] A.A. Miljutin, Isomorphisms of the spaces of continuous functions over compact setsof the cardinality of the continuum, Teor. Funkciı Funkcional Anal. i Prilozen. Vyp. 2(1966), 150â156 (Russian).
[29] J. van Mill, The infinite-dimensional topology of function spaces, North-HollandMathematical Library, 64, North-Holland Publishing Co., Amsterdam, 2001.
[30] C. Pearcy, D. Topping, Sums of small numbers of idempotents, Michigan Math. J.14 (1967), 453â465.
[31] S. Roch, P.A. Santos, B. Silbermann, Non-commutative Gelfand Theories, SpringerVerlag, London Dordrecht, Heidelberg, New York 2011.
[32] B.M. Scott, On an example of Sundaresan, The Proceedings of the 1980 Topol-ogy Conference (Univ. Alabama, Birmingham, Ala., 1980), Topology Proc. 5 (1980),185â186.
[33] W. Sierpinski, Sur une propriete topologique des ensembles denombrables denses ensoi, Fund. Math. 1 (1920), 11â16.
[34] A.E. Taylor, D.C. Lay, Introduction to Functional Analysis, Second Edition, JohnWiley and Sons, New York 1980.
[35] V. Trnkova, Isomorphisms of sums of countable Boolean algebras, Proc. Amer. Math.Soc., 80, No. 3 (1980), 389â392.
[36] V. Trnkova, V. Koubek, Isomorphisms of sums of Boolean algebras, Proc. Amer.Math. Soc., 66, No. 2 (1977), 231â236.
H. BartEconometric InstituteErasmus University RotterdamP.O. Box 1738NL-3000 DR Rotterdam, The Netherlandse-mail: [email protected]
T. EhrhardtMathematics DepartmentUniversity of CaliforniaSanta Cruz, CA-95064, USAe-mail: [email protected]
B. SilbermannFakultat fur MathematikTechnische Universitat ChemnitzD-09107 Chemnitz, Germanye-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 79â106câ 2013 Springer Basel
Fast Inversion of Polynomial-VandermondeMatrices for Polynomial Systems Relatedto Order One Quasiseparable Matrices
T. Bella, Y. Eidelman, I. GohbergZâL, V. Olshevskyand E. Tyrtyshnikov
Dedicated to Leonia Lerer on the occasion of his seventieth birthday
Abstract. While Gaussian elimination is well known to require ðª(ð3) opera-tions to invert an arbitrary matrix, Vandermonde matrices may be invertedusing ðª(ð2) operations by a method of Traub [24]. While this original versionof the Traub algorithm was noticed to be unstable, it was shown in [12] thatwith a minor modification, the Traub algorithm can typically yield a very highaccuracy. This approach has been extended from classical Vandermonde ma-trices to polynomial-Vandermonde matrices involving real orthogonal poly-nomials [3], [10], and Szego polynomials [19]. In this paper we present analgorithm for inversion of a class of polynomial-Vandermonde matrices withspecial structure related to order one quasiseparable matrices, generalizingmonomials, real orthogonal polynomials, and Szego polynomials. We derive afast ðª(ð2) inversion algorithm applicable in this general setting, and exploreits reduction in the previous special cases. Some very preliminary numericalexperiments are presented, demonstrating that, as observed by our colleaguesin previous work, good forward accuracy is possible in some circumstances,which is consistent with previous work of this type.
Mathematics Subject Classification (2010). 15A09; 65F05.
Keywords. Inversion of vandermonde matrices; polynomial vandermonde ma-trices; quasiseparable matrices.
1. Introduction
Let ð = {ð0(ð¥), ð1(ð¥), . . . , ððâ1(ð¥)} be a sequence of polynomials satisfyingdeg(ðð) = ð, and ð¥1, . . . , ð¥ð a set of pairwise distinct values. Then the corre-
80 T. Bella et al.
sponding polynomial-Vandermonde matrix is given by
ðð (ð¥) =
â¡â¢â¢â¢â£ð0(ð¥1) ð1(ð¥1) â â â ððâ1(ð¥1)ð0(ð¥2) ð1(ð¥2) â â â ððâ1(ð¥2)...
......
ð0(ð¥ð) ð1(ð¥ð) â â â ððâ1(ð¥ð)
â€â¥â¥â¥âŠ . (1.1)
In this paper the problem of inversion of the matrix ðð (ð¥) for a given system ofpolynomials ð satisfying some special recurrence relations is considered. While thestructure-ignoring approach of Gaussian elimination for inversion of ðð (ð¥) requiresðª(ð3) operations, the special structure allows algorithms to be derived exploitingthat structure, resulting in fast algorithms that can compute the ð2 entries of theinverse in only ðª(ð2) operations.
In the simplest case where ð = {1, ð¥, ð¥2, . . . , ð¥ðâ1}, ðð (ð¥) reduces to a clas-sical Vandermonde matrix and the inversion algorithm is due to Traub [24]. Inaddition to the order of magnitude decrease in complexity, it was observed in [12]that a minor modification of the original Traub algorithm results in very goodaccuracy. The derivation of this fast and accurate algorithm attracted attention inthe community, and several results were published giving fast algorithms for inver-sion of ðð (ð¥) for various special cases of the polynomial system ð . This previouswork is listed in Table 1.
Matrix ðð (ð¥) Polynomials ð Fast inversion algorithm
Classical Vandermonde monomials Traub [24]
ChebyshevâVandermonde Chebyshev GohbergâOlshevsky [10]
Three-Term Vandermonde Real orthogonal CalvettiâReichel [3]
SzegoâVandermonde Szego Olshevsky [19]
Table 1. Fast ðª(ð2) inversion algorithms for polynomial-Vandermondematrices.
In this paper, we consider a more general class of polynomials that containsall of those listed in Table 1 as special cases. This more general class of polynomialsis related to a class of rank structured matrices called quasiseparable matrices, andhence we refer to them as quasiseparable polynomials. As quasiseparable polynomi-als generalize monomials, real orthogonal polynomials, and Szego polynomials, theresulting inversion algorithm for polynomial-Vandermonde matrices ðð (ð¥) whosedefining polynomials ð are quasiseparable polynomials generalizes the previouswork in Table 1 in addition to providing new results.
In addition to generalizing these results, it is also applicable to some interest-ing new classes of polynomials for which no fast inversion algorithm of this type iscurrently available. The algorithm that is derived in this paper relies on the use ofperturbed recurrence relations for associated polynomials for the computational
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 81
speedup, and thus the classes of polynomials for which it may be used are bestreferred to in terms of the recurrence relations that they satisfy. One such class ofpolynomials are those satisfying the three recurrence relations
ðð(ð¥) = (ðŒðð¥â ð¿ð) â ððâ1(ð¥)â (ðœðð¥+ ðŸð) â ððâ2(ð¥). (1.2)
As collected in Table 2, special cases of these recurrence relations are satisfied bymonomials, real orthogonal polynomials (including Chebychev polynomials), andSzego polynomials. It is of interest that although three-term recurrence relationsfor Szego polynomials (shown in Table 2) do exist in most cases [9], far more oftentwo-term recurrence relations are used for computations with Szego polynomials.
Two-term recurrence relations for the Szego polynomials{ð#ð
}in terms of the
reflection coefficients ðð and complimentary parameters ðð are[ð0(ð¥)
ð#0 (ð¥)
]=
1
ð0
[11
],
[ðð(ð¥)
ð#ð (ð¥)
]=
1
ðð
[1 âðâðâðð 1
] [ððâ1(ð¥)ð¥ð#ðâ1(ð¥)
],
(1.3)which involve an auxiliary sequence of polynomials {ðð}. In this paper, general-izations of these two-term recurrence relations of the form[
ðºð(ð¥)ðð(ð¥)
]=
[ðŒð ðœððŸð 1
] [ðºðâ1(ð¥)
(ð¿ðð¥+ ðð)ððâ1(ð¥)
], (1.4)
which we will refer to as Szego-type recurrence relations, are also considered. Fi-nally, motivated by the most generally applicable recurrence relations available forthe class of quasiseparable polynomials that we will consider [7], the [EGO05]-typerecurrence relations[
ðºð(ð¥)ðð(ð¥)
]=
[ðŒð ðœððŸð ð¿ðð¥+ ðð
] [ðºðâ1(ð¥)ððâ1(ð¥)
], (1.5)
are considered as well. Details about these classes and their corresponding recur-rence relations will be given later.
Polynomial System ð Recurrence relations
monomials ðð(ð¥) = ð¥ â ððâ1(ð¥)Chebyshev polynomials ðð(ð¥) = 2ð¥ â ððâ1(ð¥)â ððâ2(ð¥)Real orthogonal polynomials ðð(ð¥) = (ðŒðð¥â ð¿ð)ððâ1(ð¥)â ðŸð â ððâ2(ð¥)
Szego polynomialsðð(ð¥) =
(1ððð¥+ ðð
ððâ1
1ðð
)ððâ1(ð¥)
â ( ðð
ððâ1
ððâ1
ððâ ð¥)ððâ2(ð¥)
Table 2. Systems of polynomials and corresponding recurrence relations.
82 T. Bella et al.
1.1. Structure of the paper
In Section 2 an inversion formula valid for a general system of polynomials (al-though expensive in general) is presented. The formula presented there reducesthe problem of inversion of ðð (ð¥) to that of evaluating the so-called associated
polynomials ᅵᅵ corresponding to the polynomial system ð . A relation between the
polynomial systems ð and ᅵᅵ is presented in terms of their confederate matrices.
This relation suggests a procedure for evaluating the associated polynomials ᅵᅵ.In Section 3 quasiseparable matrices and polynomials are defined and shown togeneralize the confederate matrices of the motivating special cases. Conversionsare given between the polynomial language (i.e., polynomials satisfying recurrencerelations) and the matrix language (i.e., generators of a quasiseparable matrix),and the motivating recurrence relations are identified in terms of the generatorsof their quasiseparable confederate matrices. In Section 4, perturbed recurrence
relations are presented for the associated polynomials ᅵᅵ. Three different sets ofrecurrence relations are given, two generalizing known formulas for real orthogo-nal polynomials and Szego polynomials, and a third that produces new formulasfor these cases. We briefly describe in Section 5 a fast algorithm for computingthe coefficients of the master polynomial, which are required for computing theperturbations of the recurrence relations. In Section 6 the reduction of the de-scribed algorithms in the special cases of monomials, real orthogonal polynomials,and Szego polynomials are examined in detail as well. Section 7 consists of someresults of preliminary numerical experiments with the proposed algorithm, andconclusions are offered in the final section.
2. Confederate matrices and associated polynomials
In this section we present the formula that will be used to invert a polynomial-Vandermonde matrix. Such a matrix is completely determined by ð polynomi-als ð = {ð0(ð¥), . . . , ððâ1(ð¥)} and ð nodes ð¥ = (ð¥1, . . . , ð¥ð). The desired inverseðð (ð¥)
â1 is given by the formula
ðð (ð¥)â1 = ðŒ â ð ð
ᅵᅵ(ð¥) â diag(ð1, . . . , ðð), (2.1)
with
ðð =ðâ
ð=1ð â=ð
(ð¥ð â ð¥ð)â1
(see [18], [19]) where ðŒ is the antidiagonal matrix (with ones on the antidiagonal
and zeros elsewhere), and ᅵᅵ is the system of associated (generalized Horner) poly-nomials, defined as follows: if we define themaster polynomial ð (ð¥) by ð (ð¥) = (ð¥âð¥1) â â â (ð¥â ð¥ð), then for the polynomial sequence ð = {ð0(ð¥), . . . , ððâ1(ð¥), ð (ð¥)},the associated polynomials ᅵᅵ = {ð0(ð¥), . . . , ððâ1(ð¥), ð (ð¥)} are those satisfying the
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 83
relations
ð (ð¥) â ð (ðŠ)ð¥â ðŠ =
ðâ1âð=0
ðð(ð¥) â ððâðâ1(ðŠ), (2.2)
see [16]. It can be shown that for any polynomials ð , a corresponding sequence of
polynomials ᅵᅵ satisfying (2.2) exist, and can be understood as a generalization ofthe Horner polynomials associated with the monomials; see, for instance, [2].
This discussion gives a relation between the inverse ðð (ð¥)â1 and the poly-
nomial-Vandermonde matrix ðᅵᅵ(ð¥), where ᅵᅵ is the system of polynomials as-sociated with ð . The next definition from [17] provides a connection between
recurrence relations for ð with those for ᅵᅵ.
Definition 2.1. Let the sequence of polynomials ð = {ð0(ð¥), ð1(ð¥), . . . , ðð(ð¥)} withdeg(ðð) = ð satisfy the ð-term recurrence relations
ðð(ð¥) = (ðŒðð¥â ððâ1,ð) â ððâ1(ð¥)â ððâ2,ð â ððâ2(ð¥) â â â â â ð0,ð â ð0(ð¥) (2.3)
for ð = 1, . . . , ð, and let
ð (ð¥) = ð0 â ð0(ð¥) + ð1 â ð1(ð¥) + â â â + ððâ1 â ððâ1(ð¥) + ðð â ðð(ð¥) (2.4)
for ðð â= 0. Then the confederate matrix of ð (ð¥) with respect to ð is given by
ð¶ð (ð ) =
â¡â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â£
ð01
ðŒ1
ð02
ðŒ2
ð03
ðŒ3â â â ð0,ð
ðŒðâ â â â â â ð0,ð
ðŒðâ ð0
ðŒððð1ðŒ1
ð12
ðŒ2
ð13
ðŒ3â â â ð1,ð
ðŒðâ â â â â â ð1,ð
ðŒðâ ð1
ðŒððð
0 1ðŒ2
ð23
ðŒ3â â â ... â â â â â â ð2,ð
ðŒðâ ð2
ðŒððð
0 0 1ðŒ3
. . . ððâ2,ð
ðŒð
. . ....
......
. . .. . . ððâ1,ð
ðŒð
. . .. . .
......
.... . .
. . . 1ðŒð
. . .. . .
......
.... . .
. . .. . .
. . .. . .
...
0 0 â â â â â â â â â 0 1ðŒðâ1
ððâ1,ð
ðŒðâ ððâ1
ðŒððð
â€â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥âŠ
. (2.5)
We refer to [17] for many useful properties of the confederate matrix and onlyrecall here that det(ð¥ðŒ â ð¶ð (ð )) = ð (ð¥)/(ðŒ0 â ðŒ1 â . . . â ðŒð), and that similarly,the characteristic polynomial of the ð à ð leading submatrix of ð¶ð (ð ) is equal toðð(ð¥)/ðŒ0 â ðŒ1 â . . . â ðŒð.
The motivation for considering confederate matrices is that they will allow thecomputation of the polynomials associated with the given system of polynomials.
The confederate matrices of ð and ᅵᅵ are related by
ð¶ï¿œï¿œ(ð ) = ðŒ â ð¶ð (ð )ð â ðŒ. (2.6)
84 T. Bella et al.
Recurrence Relations Confederate matrix ð¶ð (ðð)
ðð(ð¥) = ð¥ððâ1(ð¥)
â¡â¢â¢â¢â¢â¢â¢â¢â£
0 0 â â â â â â 0
1. . .
. . . 0
0. . .
. . .. . .
......
. . .. . .
. . ....
0 â â â 0 1 0
â€â¥â¥â¥â¥â¥â¥â¥âŠMonomials Companion matrix
ðð(ð¥) = (ðŒðð¥â ð¿ð)ððâ1(ð¥)â ðŸðððâ2(ð¥)
â¡â¢â¢â¢â¢â¢â¢â¢â¢â£
ð¿1ðŒ1
ðŸ2
ðŒ20 â â â 0
1ðŒ1
ð¿2ðŒ2
. . .. . .
...
0 1ðŒ2
. . . ðŸðâ1
ðŒðâ10
.... . . ð¿ðâ1
ðŒðâ1
ðŸð
ðŒð
0 â â â 0 1ðŒðâ1
ð¿ððŒð
â€â¥â¥â¥â¥â¥â¥â¥â¥âŠReal orthogonalpolynomials Tridiagonal matrix
ðð(ð¥)=[1ððâ ð¥+ ðð
ððâ1
1ðð
]ððâ1(ð¥)
â ðð
ððâ1
ððâ1
ððâ ð¥ððâ2(ð¥)
â¡â¢â¢â¢â¢â¢â£
âðâ0ð1 âðâ0ð1ð2 â â â âðâ0ð1 â â â ððâ1ððð1 âðâ1ð2 â â â âðâ1ð2 â â â ððâ1ðð
ð2 â â â âðâ2ð3 â â â ððâ1ðð. . .
. . ....
ððâ1 âðâðâ1ðð
â€â¥â¥â¥â¥â¥âŠ
Szego polynomials Unitary Hessenberg matrix
Table 3. Polynomial systems and corresponding confederate matrices.
(see [18], [19]). The passage from ð¶ð (ð ) to ð¶ï¿œï¿œ(ð ) in (2.6) can be seen as atransposition across the antidiagonal, or a pertransposition.
In Table 3, the confederate matrices corresponding to the polynomials consid-ered in previous work are given, including monomials, real orthogonal polynomials,and Szego polynomials. These equivalences are all well known. We will show laterthat all of these confederate matrices in Table 3 are special cases of order one qua-siseparable matrices, and use properties of these confederate matrices to derivethe fast algorithm.
In accordance with (2.1), the main computational burden in inversion is tocompute ðᅵᅵ, which requires evaluating ð associated polynomials {ðð(ð¥)}ðâ1ð=0 at ðpoints {ð¥ð}ðð=1. Using (2.6) directly to accomplish this is expensive for arbitraryð , since it leads to the full ð-term recurrence relations. However, in special caseswhere sparse recurrence relations for ð may be found, this leads to a fast algorithm.For instance, in the monomial case, ð = {1, ð¥, ð¥2, . . . , ð¥ðâ1} satisfy the obvious
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 85
recurrence relations ð¥ð = ð¥ â ð¥ðâ1 and hence the confederate matrix (2.5) becomes
ð¶ð (ð ) =
â¡â¢â¢â¢â¢â¢â¢â£0 0 â â â 0 âð01 0 â â â 0 âð10 1
. . ....
......
. . .. . . 0
...0 â â â 0 1 âððâ1
â€â¥â¥â¥â¥â¥â¥âŠ (2.7)
which is the well-known companion matrix. Using the pertransposition rule (2.6),we obtain the confederate matrix
ð¶ï¿œï¿œ(ðð) =
â¡â¢â¢â¢â¢â¢â¢â£âððâ1 âððâ2 â â â âð1 âð01 0 â â â 0 0
0 1. . .
... 0...
. . .. . . 0
...0 â â â 0 1 0
â€â¥â¥â¥â¥â¥â¥âŠ . (2.8)
for the associated polynomials ðð(ð¥). Using the formula (2.3), we read from thematrix (2.8) the familiar Horner recurrence relations
ð0(ð¥) = 1, ðð(ð¥) = ð¥ â ððâ1(ð¥) + ððâð. (2.9)
Thus the use of the Horner recurrence relations provides the computationalspeedup in the original Traub algorithm. In the next sections we use the quasisep-arability of the confederate matrices to derive corresponding recurrence relationsthat accomplish the same speedup in a more general setting.
3. Quasiseparable matrices and polynomials
Definition 3.1 (Quasiseparable matrices and polynomials).
â A matrix ðŽ is called (ð»,ð)-quasiseparable if
(i) it is strongly upper Hessenberg (upper Hessenberg with a nonzero firstsubdiagonal),
and
(ii) max(rank ðŽ12) = ð where the maximum is taken over all symmetricpartitions of the form
ðŽ =
[ â ðŽ12
â â]; (3.1)
86 T. Bella et al.
for instance, the low-rank blocks of a 5Ã5 (ð»,ð)-quasiseparable matrixwould be those shaded below:â¡â¢â¢â¢â¢â£
â â â â â â â â â â 0 â â â â 0 0 â â â 0 0 0 â â
â€â¥â¥â¥â¥âŠâ¡â¢â¢â¢â¢â£â â â â â â â â â â 0 â â â â 0 0 â â â 0 0 0 â â
â€â¥â¥â¥â¥âŠâ¡â¢â¢â¢â¢â£â â â â â â â â â â 0 â â â â 0 0 â â â 0 0 0 â â
â€â¥â¥â¥â¥âŠâ¡â¢â¢â¢â¢â£â â â â â â â â â â 0 â â â â 0 0 â â â 0 0 0 â â
â€â¥â¥â¥â¥âŠ.
â Let ðŽ = [ððð ] be an (ð»,ð)-quasiseparable matrix. Then the system of poly-nomials {ðð(ð¥)} related to ðŽ via
ðð(ð¥) = ðŒ1 â â â ðŒð det (ð¥ðŒ âðŽ)(ðÃð) (where ðŒð = 1/ðð+1,ð)
is called a system of (ð»,ð)-quasiseparable polynomials. That is, (ð»,ð)-quasiseparable polynomials are those polynomials with an (ð»,ð)-quasisepar-able confederate matrix.
The low-rank property described in this definition means there is redundancyin the definition of the ð2 entries of an (ð»,ð)-quasiseparable matrix, and theseentries may be described by a smaller number ðª(ðð) of parameters. If ð is suf-ficiently small and independent of ð, then this provides a significant reduction.The following well-known result may be found, for instance, in [4], and it providesthis smaller set of parameters, called the generators of the (ð»,ð)-quasiseparablematrix. An ðÃð matrix ðŽ is (ð»,ð)-quasiseparable if and only if it may be writtenin the form
ðŽ =
ᅵᅵᅵᅵ
ᅵᅵᅵᅵᅵᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ï¿œð1
. . .
. . .
ðð
ᅵᅵᅵᅵ
ᅵᅵᅵ
ð2ð1
. . .
ððððâ10
ðððÃððâð
with
ðÃðð = (ðð+1) â â â (ððâ1), ðð,ð+1 = ðŒ. (3.2)
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 87
Here ðð, ðð, ðð are scalars, the elements ðð are row vectors of maximal size ð, âðare column vectors of maximal size ð, and ðð are matrices of maximal size ðÃðsuch that all products are defined. The elements {ðð, ðð, ðð, ðð, ðð, âð} are calledthe generators of the matrix ðŽ.
The elements in the upper part of the matrix ðððÃððâð are products of a row
vector, a (possibly empty) sequence of matrices possibly of different sizes, andfinally a column vector, as depicted here:
ðððÃððâð =
1Ãð¢ðïž· ïžžïžž ïž·ðð
. .
ð¢ðÃð¢ð+1ïž· ïžžïžž ïž·ðð+1
. .
ð¢ð+1Ãð¢ð+2ïž· ïžžïžž ïž·ðð+2
. .
â â â
ð¢ðâ2Ãð¢ðâ1ïž· ïžžïžž ïž·ððâ1
. .
ð¢ðâ1Ã1ïž·ïžžïžžïž·âð
(3.3)
with ð¢ð â©œ ð for each ð = 1, . . . , ðâ 1.The generator definition and formula (2.3) together give the following ð-term
recurrence relations for (ð», 1)-quasiseparable polynomials. These ð-term recur-rence relations are not useful for computations, as they would not provide fastalgorithms, but rather are used theoretically in the derivations.
Lemma 3.2. Let ð be a sequence of (ð», 1)-quasiseparable polynomials specified bythe generators of the corresponding (ð», 1)-quasiseparable confederate matrix. Thenð satisfies the ð-term recurrence relations
ðð(ð¥) =1
ðð+1ðð
â¡â£(ð¥ â ðð)ððâ1(ð¥) â ðâ2âð=0
(ðð+1ð
Ãð+1,ðâððð(ð¥)
)â€âŠ , (3.4)
for ð = 1, . . . , ðâ 1.It is easily verified that each of the confederate matrices in Table 3 is (ð», 1)-
quasiseparable. Therefore, the class of (ð», 1)-quasiseparable polynomials includesas special cases the important classical polynomial classes of real orthogonal poly-nomials and Szego polynomials. The next theorems show that, like these motivat-ing examples, the confederate matrices corresponding to the polynomials satisfyingthe recurrence relations (1.2), (1.4), and (1.5) are also (ð», 1)-quasiseparable. Fur-thermore, explicit expressions for the generators of the confederate matrices aregiven in terms of the recurrence relations coefficients. This is useful in case theinput to the algorithms is to be these recurrence relations coefficients, as the al-gorithms are given in terms of the quasiseparable generators. We omit the proofs,which follow from repeated use of the appropriate recurrence relations and givengenerators to produce the ð-term recurrence relations of Lemma 3.2.
Theorem 3.3. Let ð = {ð0(ð¥), . . . , ðð(ð¥)} be a sequence of polynomials satisfyingdeg(ðð) = ð and the recurrence relations (1.2). Then the confederate matrix of
88 T. Bella et al.
ðð(ð¥) with respect to ð is given by
ð¶ð (ðð) =
â¡â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â£
ð¿1ðŒ1
ð¿1ðŒ1
ðœ2+ðŸ2
ðŒ2â â â
ð¿1ðŒ1
ðœ2+ðŸ2
ðŒ2
(ðœ3
ðŒ3
)(ðœ4
ðŒ4
)â â â (
ðœð
ðŒð
)1ðŒ1
ð¿2ðŒ2+ ðœ2
ðŒ1ðŒ2â â â
(ð¿2ðŒ2
+ðœ2
ðŒ1ðŒ2
)ðœ3+ðŸ3
ðŒ3
(ðœ4
ðŒ4
)â â â (
ðœð
ðŒð
). . .
. . .
(ð¿ðâ1ðŒðâ1
+ðœðâ1
ðŒðâ2ðŒðâ1
)ðœð+ðŸð
ðŒð1
ðŒðâ1
ð¿ððŒð
+ ðœð
ðŒðâ1ðŒð
â€â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥âŠ.
Furthermore, ð¶ð (ðð) is an (ð», 1)-quasiseparable matrix with generators
ð1 =ð¿1ðŒ1, ðð =
ð¿ððŒð
+ðœð
ðŒðâ1ðŒð, ð = 2, . . . , ð
ðð+1ðð =1
ðŒð, ðð =
ðððœð+1 + ðŸð+1
ðŒð+1, ð = 1, . . . , ðâ 1
ðð =ðœð+1
ðŒð+1, ð = 2, . . . , ðâ 1 âð = 1, ð = 2, . . . , ð.
(3.5)
Theorem 3.4. Let ð = {ð0(ð¥), . . . , ðð(ð¥)} be a sequence of polynomials satisfyingdeg(ðð) = ð and the recurrence relations (1.4). Then the confederate matrix ofðð(ð¥) with respect to ð is given byâ¡â¢â¢â¢â¢â¢â¢â¢â£
â ð1+ðŸ1
ð¿1â(ðŒ1 â ðœ1ðŸ1)ðŸ2
ð¿2â â â â(ðŒ1 â ðœ1ðŸ1) â â â (ðŒðâ1 â ðœðâ1ðŸðâ1)ðŸð
ð¿ð1ð¿1
â ð2+ðŸ2ðœ1
ð¿2â â â âðœ1(ðŒ2 â ðœ2ðŸ2) â â â (ðŒðâ1 â ðœðâ1ðŸðâ1)ðŸð
ð¿ð
1ð¿2
. . ....
. . . âðœðâ1(ðŒðâ1 â ðœðâ1ðŸðâ1)ðŸð
ð¿ð1
ð¿ðâ1â ðð+ðŸððœðâ1
ð¿ð
â€â¥â¥â¥â¥â¥â¥â¥âŠ.
(3.6)Furthermore, ð¶ð (ðð) is an (ð», 1)-quasiseparable matrix with generators
ð1 = âð1 + ðŸ1ð¿1
, ðð = âðð + ðŸððœðâ1ð¿ð
, ð = 2, . . . , ð
ðð = 1, ð = 2, . . . , ð
ðð =1
ð¿ð, ð = 1, . . . , ðâ 1
ð1 = 1, ðð = ðœðâ1, ð = 2, . . . , ðâ 1ðð = ðŒðâ1 â ðœðâ1ðŸðâ1, ð = 2, . . . , ðâ 1âð = âðŸð
ð¿ð(ðŒðâ1 â ðœðâ1ðŸðâ1) , ð = 2, . . . , ð
We next give a detailed example of the specification of this result to theclassical Szego case.
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 89
Example 3.5 (Classical Szego polynomials). With the choices
ðŒð =1
ðð, ðœð = âðâð, ðŸð = â ðð
ðð, ð¿ð =
1
ðð, ðð = 0
from which it follows that
ðŒð â ðœððŸð = 1â â£ððâ£2ðð
= ðð,ðŸðð¿ð
= âðð
the two-term recurrence relations (1.4) become[ðð(ð¥)
ð#ð (ð¥)
]=
1
ðð
[1 âðâðâðð 1
] [ððâ1(ð¥)ð¥ð#ðâ1(ð¥)
](3.7)
and the matrix (3.6) reduces to the matrixâ¡â¢â¢â¢â¢â¢â£ð1 ð1ð2 ð1ð2ð3 â â â ð1 â â â ððâ1ððð1 âðâ1ð2 âðâ1ð2ð3 â â â âðâ1ð2 â â â ððâ1ðð0 ð2 âðâ2ð3 â â â âðâ2ð3 â â â ððâ1ðð...
. . .. . .
. . ....
0 â â â 0 ððâ1 âðâðâ1ðð
â€â¥â¥â¥â¥â¥âŠ .
Using the convention that ð0 := â1 to insert 1 = âðâ0 throughout the first row,this matrix becomes exactly the unitary Hessenberg matrix displayed in Table 3.This demonstrates that the Szego polynomials are a special case of polynomialssatisfying (1.4), and likewise the unitary Hessenberg matrix is a special case ofthose of the form (3.6).
We also note that the condition ðð â= 0 is not satisfied by the real orthogonalpolynomials, and hence the form (3.6) cannot be used for them.
Theorem 3.6. Let ð = {ð0(ð¥), . . . , ðð(ð¥)} be a sequence of polynomials satisfyingdeg(ðð) = ð and the recurrence relations (1.5). Then the confederate matrix ofðð(ð¥) with respect to ð is given by
â¡â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â£
â ð1ð¿1
âðœ1(ðŸ2ð¿2) âðœ1ðŒ2(
ðŸ3ð¿3) âðœ1ðŒ2ðŒ3(
ðŸ4ð¿4) â â â âðœ1ðŒ2ðŒ3ðŒ4 â â â ðŒðâ1(
ðŸðð¿ð
)1ð¿1
â ð2ð¿2
âðœ2(ðŸ3ð¿3) âðœ2ðŒ3(
ðŸ4ð¿4) â â â âðœ2ðŒ3ðŒ4 â â â ðŒðâ1(
ðŸðð¿ð
)
0 1ð¿2
â ð3ð¿3
âðœ3(ðŸ4ð¿4)
. . . âðœ3ðŒ4 â â â ðŒðâ1(ðŸðð¿ð
)
0 0 1ð¿3
â ð4ð¿4
. . ....
.... . .
. . .. . .
. . . âðœðâ1(ðŸðð¿ð
)
0 â â â 0 0 1ð¿ðâ1
â ððð¿ð
â€â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥âŠ
(3.8)
90 T. Bella et al.
Furthermore, ð¶ð (ðð) is an (ð», 1)-quasiseparable matrix with generators
ðð = âððð¿ð, ð = 1, . . . , ð, ðð = 1, ð = 2, . . . , ð
ðð =1
ð¿ð, ð = 1, . . . , ðâ 1, ðð = ðœð, ð = 1, . . . , ðâ 1
ðð = ðŒð, ð = 2, . . . , ðâ 1, âð = âðŸðð¿ð, ð = 2, . . . , ð
Example 3.7 (Szego polynomials). If we choose
ðŒð = ðð, ðœð = ðâðâ1ðð, ðŸð =
ðððð, ð¿ð =
1
ðð, ðð =
ðâðâ1ðððð
the two-term recurrence relations (1.5) do not reduce to the known two-termrecurrence relations for the Szego polynomials (1.3), but become instead the newrelations [
ðº0(ð¥)
ð#0 (ð¥)
]=
[01
][ðºð(ð¥)
ð#ð (ð¥)
]=
[ðð ðâðâ1ðððð
ðð
ð¥+ðâðâ1ðð
ðð
][ðºðâ1(ð¥)ð#ðâ1(ð¥)
]. (3.9)
The matrix (3.8) does in fact reduce to the classical unitary Hessenberg matrixdisplayed in (6.8).
Both the classical Szego formula (3.7) and the new formula (3.9) describe,
of course, the same Szego polynomials {ð#ð (ð¥)}. However, the auxilary polynomi-als {ðºð(ð¥)} differ from {ðð(ð¥)} used in (3.7). Indeed, it is well known that theauxiliary polynomials involved in (3.7) satisfy
ðð(ð¥) = ð¥ð â [ð#ð (
1
ð¥â)
]â,
and in particular, degðð(ð¥) = degð#ð (ð¥). At the same time, it is easy to seethat the auxiliary polynomials {ðð(ð¥)} of the new formula (3.9) are different; in
particular degðºð(ð¥) = degð#ð (ð¥) â 1.Example 3.8 (Real orthogonal polynomials). For systems with ðŒð = 0, the ma-trix (3.8) becomes tridiagonal, and the corresponding system of polynomials areorthogonal on a real interval. Indeed, ðŒð = 0 implies ðºðâ1 = ðœðâ1ððâ2(ð¥) andhence the relations (1.5) become just the familiar three-term recurrence relations
ðð(ð¥) = (ð¿ðð¥+ ðð)ððâ1(ð¥) + ðŸððœðâ1ððâ2(ð¥).
4. Sparse recurrence relations for associated polynomials
At this point we can give an overview of the procedure for inversion of a polynomial-Vandermonde matrix whose polynomials are quasiseparable polynomials. Giventhe generators of an (ð», 1)-quasiseparable matrix (or using the theorems of the
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 91
previous section to obtain these generators given the recurrence relations coef-ficients) which is the confederate matrix with respect to the master polynomialð (ð¥) =
â(ð¥â ð¥ð) defined by the nodes ð¥ð, ð = 1, . . . , ð, with (2.4), we have
ᅵᅵᅵᅵᅵ
ᅵᅵᅵᅵᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ï¿œð1
. . .
. . .
ðð
ᅵᅵᅵᅵᅵ
ᅵᅵ
ð2ð1
. . .
ððððâ10
ðððÃððâð
ð¶ð (ð ) = â 1ðð
0
ð0
...
ððâ1(4.1)
Applying (2.6) gives us the confederate matrix for the associated polynomials as
ᅵᅵᅵᅵᅵ
ᅵᅵᅵᅵᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵᅵ
ᅵ
. . .
. . .
ᅵᅵᅵᅵᅵ
ᅵᅵ
. . .
0
ðð
ð1
ᅵᅵᅵᅵᅵ
ᅵᅵ
ððððâ1
ð2ð1
ððâððÃðâð,ðâðâðâð
ð¶ï¿œï¿œ(ð ) = â 1ðð
0
ððâ1 â â â ð0
(4.2)
From this last equation we can see that the ð-term recurrence relations satisfied
by the associated polynomials ᅵᅵ are given by
ðð(ð¥) =1
ðð+1ðð
[(ð¥â ðð)ððâ1(ð¥) â
ðâ2âð=0
(ðð+1ᅵᅵ
Ãð+1,ðâððð(ð¥)
)ïžž ïž·ïž· ïžž
typical term as in (3.4)
â ððâð
ððð0(ð¥)ïžž ïž·ïž· ïžž
perturbation
]
(4.3)where, in order to simplify the formulas, we have introduced the notation
ðð = ððâð+1, ðð = ððâð+1, ðð = ððâð+1,
ðð = âðâð+1, ᅵᅵð = ððâð+1, âð = ððâð+1.(4.4)
Notice that the nonzero top row of the second matrix in (4.2) introduces pertur-bation terms into the recurrence relations for all of the associated polynomials.
92 T. Bella et al.
Having found explicit ð-term recurrence relations for the sequence of poly-nomials associated with the given polynomials, the next goal is to find sparserecurrence relations. The motivation is that the ð-term recurrence relations areslow; they lead to ðª(ð3) algorithms, while two- and three-term recurrence rela-tions lead to ðª(ð2) algorithms.
Sparse recurrence relations are not, of course, available for all polynomialsequences ð ; this is a special property. In this section we consider the case whereð is a system of (ð», 1)-quasiseparable polynomials, and we derive sparse recurrencerelations for the associated system of polynomials.
For certain polynomial systems whose confederate matrix is not Hessenberg,such recurrence relations are derived in [7]. Obtaining similar formulas for theleading minors of ð¶ï¿œï¿œ(ð ) of the form shown in (4.2) is not simple, as the secondterm in (4.2) now affects each column in such a way that the resulting leadingsubmatrices become only (ð», 2)-quasiseparable, as opposed to the submatrices of(4.1), which are (ð», 1)-quasiseparable.
A summary of the results obtained in this section is presented in the nextTable 4. It is worth noting that in this paper we generalize all of the previousalgorithms, and not just the most widely applicable one. As a result, some ofthe results that are derived have some restrictions on their use. This leads togeneralizations of the classical algorithms as well as some new ones below.
Generators of ð â Perturbed Type of recurrence Restrictions
recurrence relations for ᅵᅵ relations derived on applicability
Theorem 4.1 Perturbed 3-term ðð â= 0
Theorem 4.2 Perturbed Szego-type ðð â= 0
Theorem 4.3 Perturbed [EGO05]-type none
Table 4. Perturbed recurrence relations for the system of associated
polynomials ᅵᅵ.
Theorem 4.1 (Perturbed three-term recurrence relations). Let ð = {ð0(ð¥), . . .,ððâ1(ð¥), ð (ð¥)} be a system of (ð», 1)-quasiseparable polynomials corresponding toan (ð», 1)-quasiseparable matrix of size ðÃð with generators {ðð, ðð, ðð, ðð, ðð, âð},with the convention that ðð = 1, ðð = 0. Suppose further that ðð â= 0 for ð =
1, . . . , ð â 1. Then the system of polynomials ᅵᅵ associated with ð satisfies therecurrence relations
ð0(ð¥) = ðð, ð1(ð¥) =1
ð2ð1
(ð¥â ð1
)ð0(ð¥) +
1
ð2ð1ððâ1
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 93
ðð(ð¥) = (ᅵᅵðð¥â ð¿ð) â ððâ1(ð¥)â (ðœðð¥+ ðŸð) â ððâ2(ð¥)ïžž ïž·ïž· ïžžtypical three-term recurrence relation terms
+ ᅵᅵðððâð â ðœðððâð+1ïžž ïž·ïž· ïžžperturbation term
,
(4.5)for ð = 2, . . . , ðâ 1, where
ᅵᅵð =1
ðð+1ðð, ð¿ð =
1
ðð+1ðð
(ðð â ððððâ1âð ᅵᅵðâ1
âðâ1
)(4.6)
ðœð =1
ðð+1ðð
âðᅵᅵðâ1âðâ1
, ðŸð =1
ðð+1ðð
âð
âðâ1
(âðâ1ððâ1 â ððâ1ᅵᅵðâ1
), (4.7)
and the coefficients ðð, ð = 0, . . . , ð are as defined in (2.4).
Proof. Let ð = {ð 0(ð¥), ð 1(ð¥), . . . , ð ðâ1(ð¥)} be the system of polynomials corre-sponding to the (ð», 2)-quasiseparable matrix ð¶ï¿œï¿œ(ð ) of the form in (4.2). Thenfrom (2.3) and (4.2), we have for ð = 1, 2, . . . , ðâ 1ð ð(ð¥) =
1
ðð+1ðð
[(ð¥â ðð)ð ðâ1(ð¥) â ððâ1âðð ðâ2(ð¥)â ððâ2ᅵᅵðâ1âðð ðâ3(ð¥)
â â â â â ð2ᅵᅵ3 â â â ᅵᅵðâ1âðð 1(ð¥)â ð1ᅵᅵ2 â â â ᅵᅵðâ1âðð 0(ð¥) + ððâð
]. (4.8)
It suffices to show that the system of polynomials {ð0(ð¥), ð1(ð¥), . . ., ððâ1(ð¥)}defined by the recurrence relations in (4.5)â(4.7) coincide with the system ð; thatis, that ðð(ð¥) = ð ð(ð¥) for each ð. We present this proof by induction. By directconfirmation, it is seen that ð0(ð¥) = ð 0(ð¥) and ð1(ð¥) = ð 1(ð¥).
Next suppose that the conclusion is true for each index less than or equal toð â 1 for some 2 â©œ ð â©œ ðâ 1. Then (4.8) for ð â 1 yields
ð¥ððâ2(ð¥) = ððððâ1ððâ1(ð¥) + ððâ1ððâ2(ð¥) (4.9)
+ ððâ2âðâ1ððâ3(ð¥) + ððâ3ᅵᅵðâ2âðâ1ððâ4(ð¥)
+ â â â + ð2ᅵᅵ3 â â â ᅵᅵðâ2âðâ1ð1(ð¥)+ ð1ᅵᅵ2 â â â ᅵᅵðâ2âðâ1ð0(ð¥) â ððâð+1.
Next, the polynomial ðð(ð¥) satisfies the recurrence relations (4.5), noting that by
hypothesis, âð = ððâð+1 â= 0 for each ð. Inserting (4.9) into (4.5) and using theinductive hypothesis, we arrive at exactly (4.8) for ðð(ð¥), which completes theproof. â¡
Theorem 4.2 (Perturbed Szego-type recurrence relations). Let ð = {ð0(ð¥), . . .,ððâ1(ð¥), ð (ð¥)} be a system of (ð», 1)-quasiseparable polynomials corresponding toan (ð», 1)-quasiseparable matrix of size ðÃð with generators {ðð, ðð, ðð, ðð, ðð, âð},with the convention that ðð = 0, ðð = 1. Suppose further that ðð â= 0 for ð =
2, . . . , ð â 1. Then the system of polynomials ᅵᅵ associated with ð satisfy the re-currence relations [
ðº0(ð¥)ð0(ð¥)
]=
[ âð1ðððð
], (4.10)
94 T. Bella et al.
[ðºð(ð¥)ðð(ð¥)
]= 1
ðð+1ðð
[ð£ð âðð+1
âð/ᅵᅵð 1
] [ðºðâ1(ð¥)
ð¢ð(ð¥)ððâ1(ð¥) +ððâð
]ïžž ïž·ïž· ïžž
perturbation term
(4.11)for ð = 1, . . . , ðâ1, with auxiliary polynomials ðºð(ð¥), and the coefficients ðð, ð =0, . . . , ð are as defined in (2.4), with the notations
ð¢ð(ð¥) = (ð¥â ðð) + ððâðᅵᅵð
, ð£ð = ðð+1ᅵᅵð+1ðð â ðð+1âð
ᅵᅵð. (4.12)
Proof. Suppose first that the generators are such that ðð â= 0 for each ð. The proofin this case will be given by showing that the system of polynomials generated bythe perturbed two-term recurrence relations (4.10)â(4.11) coincide with those givenby Theorem 4.1. From (4.11),(ð£ð +
ðð+1âð
ᅵᅵð
)[ðºðâ1(ð¥)
ð¢ð(ð¥)ððâ1(ð¥) + ððâð
]= ðð+1ðð
[1 ðð+1
â âð
ᅵᅵðð£ð
][ðºð(ð¥)ðð(ð¥)
],
(4.13)and using (4.11) for ð + 1, we have
ðºð(ð¥) =
(ᅵᅵð+1
âð+1
)(ðð+2ðð+1ðð+1(ð¥) â ð¢ð+1ðð(ð¥)â ððâðâ1)
Together with this, (4.13) produces (4.5) as desired. As by assumption ðð â= 0, foreach ð, Theorem 4.1 implies the result.
For the case of a polynomial system ð where ðð = 0 for some ð, note thatthe coefficients of the polynomials generated by the two-term recurrence relationsdepend continuously on the entries of the 2 Ã 2 transfer matrix. Let {ðð} be asequence tending to zero with ðð â= 0 for each ð, and consider a sequence ð ð withðð = ðð for each ð such that ðð = 0 in the original polynomial system ð , and allother generators the same as in ð . Then the result of the theorem holds for thesystem ð ð for every ð by above, and ð ð â ð , so by continuity, the result musthold for ð as well. This completes the proof. â¡
The formulas of the previous two theorems generalize the classical formulasfor monomials, real-orthogonal polynomials, and Szego polynomials (demonstratedbelow). We emphasize at this point that these formulas have limitations in thegeneral case: Theorem 4.1 requires nonzero ðð for each ð, and Theorem 4.2 requiresnonzero ðð for each ð. The next theorem is more general, and does not have anysuch limitations.
Theorem 4.3 (Perturbed [EGO05]-type recurrence relations). Let ð = {ð0(ð¥), . . .,ððâ1(ð¥), ð (ð¥)} be a system of (ð», 1)-quasiseparable polynomials corresponding toan (ð», 1)-quasiseparable matrix of size ðÃð with generators {ðð, ðð, ðð, ðð, ðð, âð},
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 95
with the convention that ðð = 0, ðð = 0. Then the system of polynomials ᅵᅵ asso-ciated with ð satisfy the recurrence relations[
ð¹0(ð¥)ð0(ð¥)
]=
[0ðð
], (4.14)[
ð¹ð(ð¥)ðð(ð¥)
]=
1
ðð+1ðð
[ððððᅵᅵð âððððððâð ð¥â ðð
] [ð¹ðâ1(ð¥)ððâ1(ð¥)
]ïžž ïž·ïž· ïžž
typical terms
+1
ðð+1ðð
[0
ððâð
]ïžž ïž·ïž· ïžžperturbation term
(4.15)
with auxiliary polynomials ð¹ð(ð¥), and the coefficients ðð, ð = 0, . . . , ð are as de-fined in (2.4).
Proof. The recurrence relations (4.15) define a system of polynomials which satisfythe ð-term recurrence relations
ðð(ð¥) = (ðŒðð¥â ððâ1,ð) â ððâ1(ð¥)â ððâ2,ð â ððâ2(ð¥)â â â â â ð0,ð â ð0(ð¥) (4.16)
for some coefficients ðŒð, ððâ1,ð, . . . , ð0,ð. The proof is presented by showing thatthese ð-term recurrence relations coincide exactly with (4.3). Using relations for
ðð(ð¥) and ð¹ðâ1(ð¥) from (4.15), we have
ðð(ð¥) =1
ðð+1ðð
[(ð¥â ðð)ððâ1(ð¥)â ððâ1âðððâ2(ð¥)
+ âðððâ1ᅵᅵðâ1ð¹ðâ2(ð¥) +ððâð
ððð0(ð¥)
].
(4.17)
Notice that again using (4.15) to eliminate ð¹ðâ2(ð¥) from the equation (4.17) will
result in an expression for ðð(ð¥) in terms of ððâ1(ð¥), ððâ2(ð¥), ððâ3(ð¥), ð¹ðâ3(ð¥),and ð0(ð¥) without modifying the coefficients of ððâ1(ð¥), ððâ2(ð¥), or ð0(ð¥). Againapplying (4.15) to eliminate ð¹ðâ3(ð¥) results in an expression in terms of ððâ1(ð¥),ððâ2(ð¥), ððâ3(ð¥), ððâ4(ð¥), ð¹ðâ4(ð¥), and ð0(ð¥) without modifying the coefficients ofððâ1(ð¥), ððâ2(ð¥), ððâ3(ð¥), or ð0(ð¥). Continuing in this way, the ð-term recurrencerelations of the form (4.16) are obtained without modifying the coefficients of theprevious ones. Suppose that for some 0 < ð < ð â 1 the expression for ðð(ð¥) is ofthe form
ðð(ð¥) =1
ðð+1ðð
[(ð¥â ðð)ððâ1(ð¥)â ððâ1âðððâ2(ð¥)â â â â
â â â â ðð+1ᅵᅵÃð+1,ðâððð(ð¥) + ðð+1ᅵᅵ
Ãð,ðâðð¹ð(ð¥) +
ððâð
ððð0(ð¥)
].
(4.18)
Using (4.15) for ð¹ð(ð¥) gives the relation
ðð(ð¥) =1
ðð+1ðð
[(ð¥â ðð)ððâ1(ð¥)â ððâ1âðððâ2(ð¥) â â â â
â â â â ðð ᅵᅵÃð,ðâðððâ1(ð¥) + ðð ᅵᅵÃðâ1,ðâðð¹ðâ1(ð¥) +ððâð
ððð0(ð¥)
].
(4.19)
96 T. Bella et al.
Therefore since (4.17) is the case of (4.18) for ð = ð â 2, (4.18) is true for each
ð = ð â 2, ð â 3, . . . , 0, and for ð = 0, using the fact that ð¹0 = 0 we have exactly(4.3) as desired. â¡
5. Computing the coefficients ð·ð of the master polynomial ð· (ð)
Note that in order to use the recurrence relations of the previous section it isnecessary to decompose the master polynomial ð (ð¥) into the ð basis; that is, thecoefficients ðð as in (2.4) must be computed. To this end, an efficient method ofcalculating these coefficients follows.
It is easily seen that the last polynomial ðð(ð¥) in the system ð does not affectthe resulting confederate matrix ð¶ð (ð ). Thus, if
ᅵᅵ = {ð0(ð¥), . . . , ððâ1(ð¥), ð¥ððâ1(ð¥)},we have ð¶ð (ð ) = ð¶ï¿œï¿œ(ð ). Decomposing the polynomial ð (ð¥) into the ᅵᅵ basis can
be done recursively by setting ð(0)ð (ð¥) = 1 and then for ð = 0, . . . , ðâ 1 updating
ð(ð+1)ð (ð¥) = (ð¥â ð¥ð+1) â ð(ð)ð (ð¥).
Lemma 5.1. Let ð = {ð0(ð¥), . . . , ðð(ð¥)} be given by (2.3), and ð(ð¥) =âð
ð=1 ðð â ðð(ð¥), where ð < ðâ 1. Then the coefficients of ð¥ â ð(ð¥) = âð+1
ð=1 ðð â ðð(ð¥) can becomputed by â¡â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â£
ð0...ðððð+1
0...0
â€â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥âŠ=
[ð¶ð (ðð) 0
0 â â â 0 1ðŒð
0
]â¡â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â£
ð0...ðð00...0
â€â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥âŠ. (5.1)
Proof. It can be easily checked that
ð¥ â [ ð0(ð¥) â â â ðð(ð¥)]â [ ð0(ð¥) â â â ðð(ð¥)
] â [ ð¶ð (ðð) 00 â â â 0 1
ðŒð0
]=[0 â â â 0 ð¥ â ðð(ð¥)
].
Multiplying the latter equation by the column of the coefficients, we obtain (5.1).â¡
This lemma suggests the following algorithm for computing coefficients{ð0, ð1, . . ., ððâ1, ðð} in (2.4) of the master polynomial.Algorithm 5.2 (Coefficients of the master polynomial in the ð basis). Cost: ðª(ðÃð(ð)), where ð(ð) is the cost of multiplication of an ðà ð quasiseparable matrixby a vector.
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 97
Input: A quasiseparable confederate matrix ð¶ð (ðð) and ð nodes ð¥=(ð¥1,ð¥2,. . .,ð¥ð).
1. Set[ð(0)0 â â â ð
(0)ðâ1 ð
(0)ð
]=[1 0 â â â 0
]2. For ð = 1 : ð,â¡â¢â¢â¢â¢â£
ð(ð)0...
ð(ð)ðâ1ð(ð)ð
â€â¥â¥â¥â¥âŠ =([
ð¶ï¿œï¿œ(ð¥ððâ1(ð¥)) 00 â â â 0 1 0
]â ð¥ððŒ
)â¡â¢â¢â¢â¢â£ð(ðâ1)0...
ð(ðâ1)ðâ1ð(ðâ1)ð
â€â¥â¥â¥â¥âŠwhere ᅵᅵ = {ð0(ð¥), . . . , ððâ1(ð¥), ð¥ððâ1(ð¥)}.
3. Take[ð0 â â â ððâ1 ðð
]=[ð(ð)0 â â â ð
(ð)ðâ1 ð
(ð)ð
]Output: Coefficients {ð0, ð1, . . . , ððâ1, ðð} such that (2.4) is satisfied.
It is clear that the computational burden in implementing this algorithm isin multiplication of the matrix ð¶ï¿œï¿œ(ðð) by the vector of coefficients. The cost ofeach such step is ðª(ð(ð)), where ð(ð) is the cost of multiplication of an ð à ðquasiseparable matrix by a vector, thus the cost of computing the ð coefficients isðª(ð Ãð(ð)). Using a fast ðª(ð) algorithm for multiplication of a quasiseparablematrix by a vector from [5], the cost of this algorithm is ðª(ð2).
6. Special cases of these new inversion algorithms
In what follows we show how these algorithms (as the previous section containsa choice of three perturbed recurrence relations, each leading to an inversion al-gorithm for the corresponding polynomial-Vandermonde matrix) generalizes theprevious work in the important special cases of monomials, real orthogonal poly-nomials, and Szego polynomials. The reductions in all three special cases are sum-marized in Table 5.
6.1. First special case. Monomials and the classical Traub algorithm
As shown earlier, the well-known companion matrix (2.7) results when the poly-nomial system ð is simply a system of monomials. By choosing the generatorsðð = 1, ðð = 1, ðð = 0, ðð = 1, ðð = 1, and âð = 0, the matrix (4.1) reducesto (2.7), and also (4.2) reduces to the confederate matrix for the Horner polyno-mials (2.8). In this special case, the perturbed three-term recurrence relations ofTheorem 4.1 become
ð0(ð¥) = ðð, ðð(ð¥) = ð¥ððâ1(ð¥) + ððâð, (6.1)
coinciding with the known recurrence relations for the Horner polynomials, usedin the evaluation of the polynomial
ð (ð¥) = ð0 + ð1ð¥+ â â â + ððâ1ð¥ðâ1 + ððð¥ð. (6.2)
98 T. Bella et al.
Special Case R.R. Type Resulting R.R.
Theorem 4.1 â 3-term r.r. (6.1)
Monomials Theorem 4.2 â Szego-type r.r. (6.1)
Theorem 4.3 â [EGO05]-type r.r. (6.1)
Theorem 4.1 â 3-term r.r. (6.6)
Real orthogonal Theorem 4.2 â Szego-type r.r. N/A, ðð = 0.
Theorem 4.3 â [EGO05]-type r.r. (6.6)
Theorem 4.1 â 3-term r.r. (6.13)
Szego polynomials Theorem 4.2 â Szego-type r.r. (6.11)
Theorem 4.3 â [EGO05]-type r.r. (6.12)
Table 5. Reduction of derived recurrence relations in special cases.
In fact, after eliminating the auxiliary polynomials present in Theorems 4.2 and4.3, these recurrence relations also reduce to (6.1). Thus all of the presented re-currence relations generalize those used in the classical Traub algorithm.
6.2. Second special case. Real orthogonal polynomials andthe CalvettiâReichel algorithm
Consider the almost tridiagonal confederate matrix
ð¶ð (ð ) =
â¡â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â£
ð1 â2 0 â â â 0 âð0/ððð1 ð2 â3
. . .... âð1/ðð
0 ð2 ð3 â4 0...
0 0 ð3 ð4. . . âððâ3/ðð
.... . .
. . .. . .
. . . âð â ððâ2/ðð0 â â â 0 0 ððâ1 ðð â ððâ1/ðð
â€â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥âŠ. (6.3)
The corresponding system of polynomialsð satisfy three-term recurrence relations.Such confederate matrices can be seen as special cases of our general class bychoosing the generators ðð = 1, ðð = 0, and ðð = 1, and in this case the matrix(4.1) reduces to (6.3).
To invert the corresponding polynomial-Vandermonde matrix by our algo-
rithm, we first find the confederate matrix ð¶ï¿œï¿œ(ð ) of the polynomial system ᅵᅵassociated with ð . That is, we must evaluate the polynomials corresponding to
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 99
the confederate matrix ð¶ð (ð ) given byâ¡â¢â¢â¢â¢â¢â¢â¢â¢â¢â¢â£
ðð â ððâ1/ðð âð â ððâ2/ðð âððâ3/ðð â â â âð1/ðð âð0/ððððâ1 ððâ1 âðâ1
. . ....
0 ððâ2 ððâ2 âðâ2 0...
0 0 ððâ3 ððâ3. . .
.... . .
. . .. . .
. . . â20 â â â 0 0 ð1 ð1
â€â¥â¥â¥â¥â¥â¥â¥â¥â¥â¥âŠ.
(6.4)Note that the highlighted column corresponds to the full recurrence relation
ð3(ð¥) =1
ððâ3(ð¥ â ððâ2)ð2(ð¥)â âðâ1
ððâ3ð1(ð¥) +
1
ððâ3ððâ3ðð
ð0(ð¥) (6.5)
In this case the perturbed three-term recurrence relations from Theorem 4.1 aswell as the two-term recurrence relations from Theorem 4.3 both become
ðð(ð¥) =1
ððâð(ð¥â ððâð)ððâ1(ð¥)â ððâð+1
ððâðâðâð+1ððâ2(ð¥) +
1
ððâðððâð (6.6)
which coincides with the Clenshaw rule for evaluating
ð (ð¥) = ð0ð0(ð¥) + ð1ð1(ð¥) + â â â + ððâ1ððâ1(ð¥) + ðððð(ð¥). (6.7)
Thus our formula generalizes both the Clenshaw rule and the algorithms designedfor inversion of three-term-Vandermonde matrices in [3] and [10].
Notice that the Szego-like two-term recurrence relations of Theorem 4.2 areinapplicable as ðð = 0 is a necessary choice of generators.
6.3. Third special case. Szego polynomials and the algorithm of [18]
Next consider the important special case of the almost unitary Hessenberg matrixof Table 2,
ð¶ð (ð ) =
â¡â¢â¢â¢â¢â¢â£âðâ0ð1 âðâ0ð1ð2 âðâ0ð1ð2ð3 â â â âðâ0ð1 â â â ððâ1ððð1 âðâ1ð2 âðâ1ð2ð3 â â â âðâ1ð2 â â â ððâ1ðð0 ð2 âðâ2ð3 â â â âðâ2ð3 â â â ððâ1ðð...
. . .. . .
. . ....
0 â â â 0 ððâ1 âðâðâ1ðð
â€â¥â¥â¥â¥â¥âŠ (6.8)
that corresponds to the Szego polynomials (represented by the reflection coeffi-cients ðð and complimentary parameters ðð), and polynomial ð (ð¥). The Szegopolynomials are known to satisfy the two-term recurrence relations (1.3) as wellas the three-term recurrence relations
ð#0 (ð¥) = 1, ð#1 (ð¥) =1
ð1â ð¥ð#0 (ð¥) â
ð1ð1ð#0 (ð¥)
ð#ð (ð¥) =
[1
ððâ ð¥+ ðð
ððâ11
ðð
]ð#ðâ1(ð¥) â
ððððâ1
ððâ1ðð
â ð¥ â ð#ðâ2(ð¥)(6.9)
100 T. Bella et al.
(see [13], [9]). As above, the polynomials associated with the system of Szegopolynomials are determined by the confederate matrix ð¶ï¿œï¿œ(ð ) given byâ¡â¢â¢â¢â¢â¢â¢â£
âðððâðâ1 â ððâ1
ððâððððâ1ðâðâ2 â ððâ2
ððâ â â âððððâ1 . . . ð1ðâ0 â ð0
ðð
ððâ1 âððâ1ðâðâ2 â â â âððâ1ððâ2 . . . ð1ðâ00
. . .. . .
......
. . .. . .
...0 â â â ð1 âð1ðâ0
â€â¥â¥â¥â¥â¥â¥âŠ .(6.10)
For this special case, let ðð = 1, ðð = ðð, ðð = âðððâðâ1, ðð = ðâðâ1, ðð = ððâ1,and âð = âððâ1ðð (alternatively ðð = ðâðâ1ðð, ðð = ðð, âð = âðð). This choiceof generators reduces (4.1) to the matrix (6.8) as well as (4.2) to (6.10), and inthis case the perturbed two-term recurrence relations of Theorem 4.2 become[
ð0(ð¥)
ð#0 (ð¥)
]=
1
ðð
[ âðð1
],[
ðð(ð¥)
ð#ð (ð¥)
]=
1
ððâð
[1 âðâðâð
âððâð 1
][ððâ1(ð¥)ð¥ð#ðâ1(ð¥) + ððâð
],
(6.11)
coinciding with those recurrence relations derived in [18]. The recurrence relationsfrom Theorem 4.3 reduce to new two-term recurrence relations; that is, relationsthat do not generalize those derived in [18]. They become[
ð¹ð(ð¥)
ð#ð (ð¥)
]=
1
ððâð
[ððâðððâð+1 âððâðð
âðâð+1
âððâð+1ððâð ð¥+ ððâððâðâð+1
][ð¹ðâ1(ð¥)ð#ðâ1(ð¥)
]
+
[0ððâð
].
(6.12)
Also, the perturbed three-term recurrence relations of Theorem 4.1 reduce to
ð0(ð¥) =1
ðð, ð1(ð¥) =
{1
ððâ1â ð¥ð0(ð¥)â ððâ1ð
âð
ððâ1ð0(ð¥)
}+ððâ1ððâ1
.
ðð(ð¥) =
[1
ððâðâ ð¥+ ððâð
ððâð+1
1
ððâð
]ððâ1(ð¥) â ððâð
ððâð+1
ððâð+1
ððâðâ ð¥ â ððâ2(ð¥)
+ððâð â ððâð+1ððâð+1
ððâð
ððâð+1
ððâð(6.13)
in this case, also coinciding with the perturbed three-term recurrence relations in[18]. Thus both of these theorems generalize the recurrence relations derived in[18] as well.
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 101
7. Numerical experiments
The numerical properties of the Traub algorithm and its generalizations (that arethe special cases of our general algorithm) were studied by many different authors.It was noticed in [12] that a version of the Traub algorithm can yield high accuracyin certain cases if the algorithm is preceded with the Leja ordering of the nodes;that is, ordering such that
â£ð¥1⣠= max1â©œðâ©œð
â£ð¥ðâ£,ðâ1âð=1
â£ð¥ð â ð¥ð ⣠= maxðâ©œðâ©œð
ðâ1âð=1
â£ð¥ð â ð¥ð â£, ð = 2, . . . , ðâ 1
(see [22], [15], [20]). It was noticed in [12] that the same is true for ChebyshevâVandermonde matrices.
No error analysis was done, but the conclusions of the above authors wasthat in many cases the Traub algorithm and its extensions can yield much betteraccuracy than Gaussian elimination, even for very ill-conditioned matrices.
We made our preliminary experiments with the general algorithm, and ourconclusions are consistent with the experience of our colleagues. In all cases westudied the proposed algorithm yields better accuracy than Gaussian elimination,e.g., in the new special cases of SzegoâVandermonde and (ð», 1)-quasiseparable-Vandermonde matrices. However, our experiments need to be done for differentspecial cases and also the numerical properties of different recurrence relations areworth analyzing. This is a topic for future study.
The algorithm has been implemented in MATLAB version 7. The results ofthe algorithm using standard MATLAB code, and hence double precision arith-metic, were compared with exact solutions calculated using the MATLAB Sym-bolic Toolbox command vpa(), which allows software-implemented precision ofarbitrary numbers of digits. The number of digits was set to 64, however in caseswhere the condition number of the coefficient matrix exceeded 1030, this was raisedto 100 digits to maintain accuracy.
We compare the forward accuracy of the inverse computed by the algorithmwith respect to the inverse computed in high precision, defined by
ð =â¥ðð (ð¥)â1 â ðð (ð¥)â1â¥2
â¥ðð (ð¥)â1â¥2 (7.1)
where ðð (ð¥)â1 is the solution computed by each algorithm in MATLAB in double
precision, and ðð (ð¥)â1 is the exact solution. In the tables, TraubQS denotes the
proposed Traub-like algorithm, and inv() indicates MATLABâs inversion com-mand. Finally, cond(ð ) denotes the condition number of the matrix ð computedvia the MATLAB command cond().
Experiment 1. In this experiment, the problem was chosen by choosing the gen-erators that define the recurrence relations of the polynomial system randomly
102 T. Bella et al.
in (â1, 1), and the nodes ð¥ð were selected equidistant on (â1, 1) via the formula
ð¥ð = â1 + 2
(ð
ðâ 1), ð = 0, 1, . . . , ðâ 1.
We test the accuracy of the inversion algorithm for various sizes ð of matricesgenerated in this way. Some results are tabulated in Table 6, and shown graphicallyin Figure 1.
ð cond(V) inv() TraubQS
10 4.2e04 4.1e-14 3.4e-152.2e05 2.5e-14 6.3e-153.7e08 1.0e-13 8.9e-14
15 1.1e10 3.5e-11 3.5e-111.1e11 1.5e-12 4.8e-134.7e11 1.3e-13 7.7e-14
20 7.6e14 1.1e-10 3.4e-121.2e15 4.2e-11 1.1e-117.8e17 1.2e-09 1.7e-15
25 4.8e19 1.2e-09 1.7e-131.1e24 5.9e-07 1.3e-111.5e27 8.4e-08 2.4e-09
30 3.3e24 7.2e-07 1.1e-135.0e27 2.8e-06 1.7e-111.8e30 1.3e-03 9.5e-10
35 7.3e23 2.4e-04 6.9e-108.3e26 2.6e-03 1.2e-061.4e27 2.9e-05 1.4e-08
40 1.1e31 8.2e-02 2.4e-132.4e32 3.4e+00 9.9e-121.7e33 1.2e-01 1.0e-08
45 4.3e30 1.7e-01 1.7e-051.7e31 5.9e-01 1.0e-083.9e35 1.0e-02 2.4e-08
50 2.1e42 1.0e+00 4.7e-063.9e44 1.0e+00 7.0e-066.6e45 1.0e+00 6.3e-06
Table 6. Equidistant nodes on (â1, 1).
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 103
1010
1020
1030
1040
1016
1014
1012
1010
108
106
104
102
100
cond(V)
forw
ard
erro
r
Equidistant Nodes on ( 1,1)
inv()Traub QS
Figure 1. Equidistant nodes on (â1, 1).
Notice that the performance of the proposed inversion algorithm is an im-provement over that of MATLABâs standard inversion command inv() in thisspecific case.
Experiment 2. Next, the values for the generators and the nodes were chosenrandomly on the unit disc. We test the accuracy for various 30 Ã 30 matricesgenerated in this way, and present some results in Table 7 and Figure 2.
1025
1030
1016
1014
1012
1010
108
106
104
102
100
cond(V)
forw
ard
erro
r
Random Parameters on the Unit Disc
inv()Traub QS
Figure 2. Random parameters on the unit disc.
104 T. Bella et al.
cond(V) inv() TraubQS
1.7e21 1.3e-07 3.5e-143.9e23 1.2e-05 9.6e-154.3e23 2.7e-03 1.2e-142.8e24 2.4e-05 8.4e-132.9e24 3.9e-03 4.3e-12
1.8e25 6.8e-07 2.6e-122.2e25 8.9e-03 3.4e-143.1e25 1.3e-03 3.6e-143.5e25 2.9e-03 7.9e-146.8e25 1.0e+00 2.2e-11
2.2e27 1.0e-02 2.9e-114.9e27 3.6e+00 2.3e-136.6e27 9.9e+00 7.6e-137.6e27 4.6e-04 2.0e-122.0e28 1.9e-03 5.7e-14
2.4e28 6.9e-04 9.6e-152.6e28 2.5e-02 1.2e-135.2e28 2.4e-05 1.7e-126.9e30 1.2e-03 2.5e-141.4e33 1.0e+00 2.9e-13
Table 7. Random parameters on the unit disc.
8. Conclusions
In this paper we used properties of confederate matrices to extend the classicalTraub algorithm for inversion of Vandermonde matrices to the general polynomial-Vandermonde case. The relation between polynomial systems satisfying some re-currence relations and quasiseparable matrices allowed an order of magnitudecomputational savings in this case, resulting in an ðª(ð2) algorithm as opposedto Gaussian elimination, which requires ðª(ð3) operations. Finally, some numeri-cal experiments were presented that indicate that, under some circumstances, theresulting algorithm can give better performance than Gaussian elimination.
Fast Inversion of (ð», 1)-quasiseparable-Vandermonde Matrices 105
References
[1] M. Bakonyi and T. Constantinescu, Schurâs algorithm and several applications, inPitman Research Notes in Mathematics Series, vol. 61, Longman Scientific and Tech-nical, Harlow, 1992.
[2] T. Bella, Y. Eidelman, I. Gohberg, I. Koltracht and V. Olshevsky, A BjorckâPereyra-type algorithm for SzegoâVandermonde matrices based on properties of unitary Hes-senberg matrices, Linear Algebra and Applications, Volume 420, Issues 2-3 pp. 634â647, 2007.
[3] D. Calvetti and L. Reichel, Fast inversion of Vandermonde-like matrices involvingorthogonal polynomials, BIT, 1993.
[4] Y. Eidelman and I. Gohberg, On a new class of structured matrices, Integral Equa-tions and Operator Theory, 34 (1999), 293â324.
[5] Y. Eidelman and I. Gohberg, Linear complexity inversion algorithms for a class ofstructured matrices, Integral Equations and Operator Theory, 35 (1999), 28â52.
[6] Y. Eidelman and I. Gohberg, A modification of the Dewildeâvan der Veen methodfor inversion of finitestructured matrices, Linear Algebra Appl., 343-344 (2002), 419â450.
[7] Y. Eidelman, I. Gohberg and V. Olshevsky, Eigenstructure of Order-One-Quasi-separable Matrices. Three-term and Two-term Recurrence Relations, Linear Algebraand its Applications, Volume 405, 1 August 2005, pages 1â40.
[8] G. Forney, Concatenated codes, The M.I.T. Press, 1966, Cambridge.
[9] L.Y. Geronimus, Polynomials orthogonal on a circle and their applications, Amer.Math. Translations, 3 pp. 1â78, 1954 (Russian original 1948).
[10] I. Gohberg and V. Olshevsky, Fast inversion of ChebyshevâVandermonde matrices,Numerische Mathematik, 67, No. 1 (1994), 71â92.
[11] I. Gohberg and V. Olshevsky, A fast generalized ParkerâTraub algorithm for in-version of Vandermonde and related matrices, Journal of Complexity, 13(2) (1997),208â234.A short version in pp. in Communications, Computation, Control and Signal Process-ing: A tribute to Thomas Kailath, eds. A. Paulraj, V. Roychowdhury and C. Shaper,Kluwer Academic Publishing, 1996, 205â221.
[12] I. Gohberg and V. Olshevsky, The fast generalized ParkerâTraub algorithm for inver-sion of Vandermonde and related matrices, J. of Complexity, 13(2) (1997), 208â234.
[13] U. Grenader and G. Szego, Toeplitz forms and Applications, University of CaliforniaPress, 1958.
[14] W.G. Horner, A new method of solving numerical equations of all orders by contin-uous approximation, Philos. Trans. Roy. Soc. London, (1819), 308â335.
[15] N. Higham, Stability analysis of algorithms for solving confluent Vandermonde-likesystems, SIAM J. Matrix Anal. Appl., 11(1) (1990), 23â41.
[16] T. Kailath and V. Olshevsky, Displacement structure approach to polynomial Van-dermonde and related matrices, Linear Algebra and Its Applications, 261 (1997),49â90.
[17] J. Maroulas and S. Barnett, Polynomials with respect to a general basis. I. Theory,J. of Math. Analysis and Appl., 72 (1979), 177â194.
106 T. Bella et al.
[18] V. Olshevsky, Eigenvector computation for almost unitary Hessenberg matrices andinversion of SzegoâVandermonde matrices via Discrete Transmission lines. LinearAlgebra and Its Applications, 285 (1998), 37â67.
[19] V. Olshevsky, Associated polynomials, unitary Hessenberg matrices and fast general-ized ParkerâTraub and BjorckâPereyra algorithms for SzegoâVandermonde matricesinvited chapter in the book âStructured Matrices: Recent Developments in The-ory and Computation,â 67â78, (D. Bini, E. Tyrtyshnikov, P. Yalamov, eds.), 2001,NOVA Science Publ., USA.
[20] V. Olshevsky, Pivoting for structured matrices and rational tangential interpolation,in Fast Algorithms for Structured Matrices: Theory and Applications, CONM/323,pp. 1â75, AMS publications, May 2003.
[21] F. Parker, Inverses of Vandermonde matrices, Amer. Math. Monthly, 71 (1964), 410â411.
[22] L. Reichel and G. Opfer, ChebyshevâVandermonde systems, Math. of Comp., 57(1991), 703â721.
[23] J. Stoer, R. Bulirsch, Introduction to Numerical Analysis, Springer-Verlag, 1992,277â301.
[24] J. Traub, Associated polynomials and uniform methods for the solution of linearproblems, SIAM Review, 8, No. 3 (1966), 277â301.
T. BellaDepartment of MathematicsUniversity of Rhode IslandKingston, RI 02881, USAe-mail: [email protected]
Y. Eidelman and I. GohbergZâL
School of Mathematical SciencesRaymond and Beverly Sackler Faculty of Exact SciencesTel Aviv UniversityRamat-Aviv 69978, Israele-mail: [email protected]
V. OlshevskyDepartment of MathematicsUniversity of ConnecticutStorrs, CT 06269, USAe-mail: [email protected]
E. TyrtyshnikovInstitute of Numerical MathematicsRussian Academy of SciencesGubkina Street, 8Moscow, 119991, Russiae-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 107â125câ 2013 Springer Basel
Long Proofs of Two CarlsonâSchneiderType Inertia Theorems
Harry Dym and Motke Porat
To Leonid Lerer on the occasion of his retirement from the faculty of the Department ofMathematics at the Technion, with affection and respect.
Abstract. This expository note is devoted to a discussion of the equivalenceof inertia theorems of the CarlsonâSchneider type with the existence of finite-dimensional reproducing kernel Krein spaces of the de Branges type. The firstfive sections focus on an inertia theorem connected with a Lyapunov equation.A sixth supplementary section sketches an analogous treatment of the Steinequation. The topic was motivated by a question raised by Leonid Lerer.
Mathematics Subject Classification (2010). 46C20, 46E22, 47B32, 47B50,93B20.
Keywords. Inertia theorems, realization theory, reproducing kernel spaces,finite-dimensional de BrangesâKrein spaces, factorization of rational matrix-valued functions, LyapunovâStein equations.
1. Introduction
The papers [9] by Lerer and Tismenetsky and [5] by Dym and Volok both studythe zero distribution of matrix polynomials, but by very different methods. Inparticular, [9] rests heavily on the spectral theory of matrix polynomials thatis conveniently summarized in the monograph [6] and on the CarlsonâSchneiderinertia theorem, whereas [5] uses reproducing kernel space methods. Some yearsago, in an exchange of E-mails with the first listed author of this note, Leonidwondered how the analysis in [5] managed to avoid invoking the CarlsonâSchneiderinertia theorem. The purpose of the first five sections is to suggest an answer tothis question by showing how to deduce CarlsonâSchneider type theorems fromreproducing kernel formulas. To be more precise, the theorem we consider is weakerthan the CarlsonâSchneider theorem because of the observability assumption; itcorresponds to Corollary 1 on p. 449 of [8], which is credited to C.T. Chen [2]
108 H. Dym and M. Porat
and H.K. Wimmer [10]. The proof is a bit on the long side, but is completelyself-contained and uses only elementary ideas from linear system theory. A sixthsection deals with an analogous inertia theorem for the disc.
The notation Î + (resp., Î â) for the open right (resp., left) half-plane andâ°â(ðŽ), â°0(ðŽ) and â°+(ðŽ) for the sum of the algebraic multiplicities of the eigen-values of a matrix ðŽ â âðÃð in Î â, ðâ and Î +, respectively, will be needed toformulate the version of the CarlsonâSchneider inertia theorem under considera-tion:
Theorem 1.1. If ðŽ â âðÃð, ð¶ â âðÃð, (ð¶,ðŽ) is an observable pair and ð = ð â ââðÃð is a solution of the Lyapunov equation
ðŽâð + ððŽ+ ð¶âð¶ = 0, (1.1)
then
(1) ð(ðŽ) â© ðâ = â and ð is invertible, i.e., â°0(ðŽ) = â°0(ð ) = 0,(2) â°+(ðŽ) = â°â(ð ) and â°â(ðŽ) = â°+(ð ).
This note is organized as follows: In Section 2 we present some preliminaryfacts and notation for subsequent use. A short well-known proof of Theorem 1.1based on elementary tools of linear system theory is presented in Section 3. InSection 4 Theorem 1.1 is used to establish a finite-dimensional reproducing kernelKrein space (RKKS) with a reproducing kernel (RK) of the special form
ðŸð(ð) =ðŒð âÎ(ð)Î(ð)â
ðð(ð)(1.2)
where ðð(ð) = ð + ðâ and Î admits a factorization Î = Îâ11 Î2, with Î1 andÎ2 both inner rmvfâs (rational matrix-valued functions) with respect to Î +. Aproof of Theorem 1.1 that is based on methods and formulas of reproducing kernelspaces is presented in Section 5. An analogous treatment of an inertia theorem forthe Stein equation is sketched in the sixth section.
2. Preliminaries
We begin with a lemma that verifies assertion (1) of Theorem 1.1 and reduces theproblem of establishing (2) to the case where the matrix ðŽ is of particular form.Subsequently, we present some notation and formulas that will be used frequentlyin the later sections.
Lemma 2.1. If the assumptions in Theorem 1.1 are in force, then ð(ðŽ) â© ðâ = â and ð is invertible and without loss of generality ðŽ may be assumed to be of theform
ðŽ = ðŽ+, or ðŽ = ðŽâ, or ðŽ =
[ðŽ+ 00 ðŽâ
], with ð(ðŽÂ±) â Π±. (2.1)
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 109
Proof. The proof is separated into parts:
1. If ðð¢ = 0 for some ð¢ â âð, then, by (1.1),
0 = ð¢âðŽâðð¢+ ð¢âððŽð¢+ ð¢âð¶âð¶ð¢ = ð¢âð¶âð¶ð¢ =â ð¶ð¢ = 0.
Therefore, (1.1) implies that ððŽð¢ = 0 and hence, by repeating the previ-ous argument, that ð¶ðŽð¢ = 0 and ððŽ2ð¢ = 0. Thus, ð¶ðŽðð¢ = 0 for everynonnegative integer ð and hence, as (ð¶,ðŽ) is observable, ð¢ = 0, i.e., ð isinvertible.
2. If ðŽð¢ = ðð¢ for some ð¢ â âð and ð â â, then ð¢âðŽâ = ðâð¢â and hence (1.1)implies that
0 = ð¢âðŽâðð¢+ ð¢âððŽð¢+ ð¢âð¶âð¶ð¢ = (ð + ðâ)ð¢âðð¢+ ð¢âð¶âð¶ð¢
= ð¢âð¶âð¶ð¢ if ð â ðâThus, ð¶ðŽðð¢ = ððð¶ð¢ = 0 in this case, which implies that ð¢ = 0, since thepair (ð¶,ðŽ) is observable. Therefore, ð(ðŽ) â© ðâ = â .
3. As ð(ðŽ)â© ðâ = â , ðŽ = ððœðâ1, where ðœ is a Jordan matrix with ð(ðœ) â Î +
or ð(ðŽ) â Î + or is of the form
ðœ =
[ðŽ+ 00 ðŽâ
]where ð(ðŽÂ±) â Π±.
Then ð is an invertible Hermitian solution of equation (1.1) if and only ifðâðð a Hermitian invertible solution of the equation
ðœâ(ðâðð) + (ðâðð)ðœ + (ð¶ð)âð¶ð = 0.
Moreover, ð(ðœ) â© ðâ = â , â°Â±(ðœ) = â°Â±(ðŽ), â°Â±(ðâðð) = â°Â±(ð ) and the pair(ð¶ð, ðœ) is observable, since
rank
[ððŒð âðŽð¶
]= rank
[ð 00 ðŒð
] [ððŒð â ðœð¶ð
]ðâ1 = rank
[ððŒð â ðœð¶ð
]. â¡
In view of the preceding lemma, we can without loss of generality assumethat ðŽ is of the form (2.1) and ð := â°+(ðŽ). In this note we shall always assumethat 1 †ð †ð â 1 and shall leave the cases ð = 0 and ð = ð to the reader.Correspondingly, let ð¶ =
[ð¶1 ð¶2
]with ð¶1 â âðÃð, ð¶2 â âðÃ(ðâð) and ð =[
ð11 ð12ð21 ð22
]with ð11 â âðÃð, ð22 â â(ðâð)Ã(ðâð) be the decompositions of the
matrices ð¶ and ð that are conformal with that of ðŽ. Then the Lyapunov equation(1.1) is equivalent to the first, second and fourth of the following four equations:
ðŽâ+ð11 + ð11ðŽ+ + ð¶â1ð¶1 = 0, (2.2)
ðŽââð21 + ð21ðŽ+ + ð¶â2ð¶1 = 0, (2.3)
ðŽâ+ð12 + ð12ðŽâ + ð¶â1ð¶2 = 0 (2.4)
and
ðŽââð22 + ð22ðŽâ + ð¶â2ð¶2 = 0. (2.5)
110 H. Dym and M. Porat
Finally, let ð®ðÃðð (Î +) denote the generalized Schur class of mvfâs (matrix-valued
functions) ð (ð), that are meromorphic in Î + and for which the kernel
Îð ð(ð) =
ðŒð â ð (ð)ð (ð)âðð(ð)
has ð negative squares on ð¥+ð à ð¥+ð , and ð¥+ð denotes the domain of analyticity of ð in Î +.
3. A quick proof of Theorem 1.1
In this section we present a variant of the well-known short proof of Theorem 1.1(see, e.g., [8] or [3]).
Proof. Assertion (1) of the theorem is verified in Lemma 2.1. Next, as ð(ðŽÂ±) â Π±,(2.2) and (2.5) both have unique solutions:
ð11 = ââ« â0
ðâð¡ðŽâ+ð¶â1ð¶1ð
âð¡ðŽ+ðð¡ †0 and ð22 =
â« â
0
ðð¡ðŽââð¶â2ð¶2ð
ð¡ðŽâðð¡ ⥠0,
respectively. The observability of the pair (ð¶,ðŽ) implies that the pairs (ð¶1, ðŽ+)and (ð¶2, ðŽâ) are observable, as is perhaps verified most easily by the PopovâBelevitchâHautus test. Then Lemma 2.1 implies that ð11 and ð22 are invertiblematrices and hence, by Schur complements, that
ð =
[ðŒð ðâ111 ð120 ðŒðâð
]â [ð11 00 ð22 â ð21ðâ111 ð12
] [ðŒð ðâ111 ð120 ðŒðâð
].
Thus, by the Sylvester law of inertia,
â°+(ð ) = â°+(ð11) + â°+(ð22 â ð21ðâ111 ð12)
= â°+(ð22 â ð21ðâ111 ð12)
= ðâ ð = â°â(ðŽ)as ð22 > 0, ð11 < 0 and ð12 = ð
â21, and
â°â(ð ) = â°â(ð11) + â°â(ð22 â ð21ðâ111 ð12)
= â°â(ð11)= ð = â°+(ðŽ). â¡
4. From the inertia theorem to a RKKS
In this section we will establish a finite-dimensional reproducing kernel Krein spaceâ³ with a RK of the form (1.2) and then use the inertia theorem to obtain a coprimefactorization formula for the rmvf Î â ð®ðÃð
â°â(ð )(Î +) of the form
Î = Îâ11 Î2,
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 111
where Î1 and Î2 are both inner rational matrix-valued functions (rmvfâs) with re-spect to Î +, and hence BlaschkeâPotapov products. Moreover, we will give explicitrealization formulas for Î1 and Î2 that are minimal.
Theorem 4.1. If ðŽ â âðÃð, ð¶ â âðÃð, the pair (ð¶,ðŽ) is observable and ð =ð â â âðÃð is a solution of the Lyapunov equation (1.1), then:
(1) ð(ðŽ) â© ðâ = â and ð is invertible, i.e., â°0(ðŽ) = â°0(ð ) = 0,
(2) The space
â³ = {ð¹ (ð)ð¢ : ð¢ â âð} (4.1)
with
ð¹ (ð) = ð¶(ððŒð âðŽ)â1, ð â â â ð(ðŽ) (4.2)
and indefinite inner product
âšð¹ (ð)ð¢, ð¹ (ð)ð£â©â³ = ð£âðð¢, (4.3)
is an ð-dimensional RKKS with RK
ðŸð(ð) = ð¹ (ð)ðâ1ð¹ (ð)â, ð, ð â â â ð(ðŽ). (4.4)
(3) The RK may be expressed in the form
ðŸð(ð) =ðŒð âÎ(ð)Î(ð)â
ðð(ð), ð, ð â â â ð(ðŽ), (4.5)
where the ðÃð rmvf
Î(ð) = ðŒð â ð¶(ððŒð âðŽ)â1ðâ1ð¶â (4.6)
admits the factorization
Î(ð) = Î1(ð)â1Î2(ð), (4.7)
with
Î1(ð) = ðŒð + ð¶1(ððŒð âðŽ1)â1ðâ111 ð¶
â1 , (4.8)
Î2(ð) = ðŒð + ð¶(ððŒðâð âðŽâ)â1ð ð¶â, (4.9)
ðŽ1 = âðâ111 ðŽâ+ð11, ð¶ = ð¶2 â ð¶1ð
â111 ð12
and
ð = â(ð22 â ð21ðâ111 ð12)â1.
(4) The rmvfâs Î1 and Î2 are finite BlaschkeâPotapov products that are bothinner with respect to Î + and are left coprime.
(5) The realizations (4.8) and (4.9) are minimal.
Proof. Assertion (1) is verified in Lemma 2.1. Next, sinceâ³ is a RKKS with RKthat is given by formula (4.4) and (ð¶,ðŽ) is an observable pair, the columns of thermvf ð¹ are linearly independent. Therefore, (2) holds.
The formula (4.5) in (3) may be verified by a straightforward calculationthat uses the Lyapunov equation (1.1). Another direct calculation serves to verify
112 H. Dym and M. Porat
the factorization formula and it is easily seen that ð11 is invertible (just as in theproof in Section 3) and hence that ð22âð12ðâ111 ð21 is also invertible (by the Schurcomplements formula for ð ) and ð is well defined. Thus, (3) holds.
The observability of the pairs (ð¶1, ðŽ+) and (ð¶2, ðŽâ) is inherited from theobservability of the pair (ð¶,ðŽ) (as is most easily seen by the PopovâBelevitchâHautus test). Therefore, successive applications of Theorem 1.1 to equation (2.2)for the ð à ð matrix ð11 and to equation (2.5) for the (ð â ð) à (ð â ð) matrixð22, yields the implications
â°Â±(ðŽ+) = â°â(ð11) =â â°â(ð11) = ð =â ð11 < 0.
and
â°Â±(ðŽâ) = â°â(ð22) =â â°+(ð22) = ðâ ð =â ð22 > 0.
Thus, as ð11 < 0 and ð12 = ðâ21,
ð = â(ð22 â ð21ðâ111 ð12)â1 < 0.
A direct calculation shows that
Î1(ð)âÎ1(ð)â ðŒð = (ð + ðâ)ð¶1(ððŒð +ðŽ+)
â1ð11(ðâðŒð +ðŽâ+)â1ð¶â1
and hence
Î1(ð)âÎ1(ð) = ðŒð on ðâ and Î1(ð)
âÎ1(ð) †ðŒð in Î +.
A similar calculation shows that, if ð = âð â1, thenÎ2(ð)
âÎ2(ð) â ðŒð = â (ð+ ðâ)ð¶ðâ1(ðâðŒðâð âðŽââ)â1ð(ððŒðâð âðŽâ)â1ðâ1ð¶â
+ ð¶ðâ1(ðâðŒðâð âðŽââ)â1ð(ððŒðâð âðŽâ)â1ðâ1ð¶â
with ð = ðŽââð + ððŽâ + ð¶âð¶. Now, multiplying equation (2.3) on the right byðâ111 ð12 and subtracting it from equation (2.5), yields the equation
0 = ðŽââ(ð22 â ð21ðâ111 ð12) + ð22ðŽâ â ð21ðŽ+ðâ111 ð12 + ð¶
â2ð¶2 â ð¶â2ð¶1ð
â111 ð12
= ðŽââ(âð â1) + (ð22 â ð21ðâ111 ð12 + ð21ðâ111 ð12ðŽâ â ð21ðŽ+ð
â)ðŽâ111 ð12 + ð¶
â2ð¶
= ðŽââð + ððŽâ + ð21ðâ111 ð12ðŽâ â ð21ðŽ+ð
â111 ð12 + ð¶
â2ð¶ =M.
Next, multiplying equation (2.4) on the right by ð21ðâ111 and subtracting it fromM,
0 = ðŽââð + ððŽâ â ð21ðŽ+ðâ111 ð12 + ð¶
â2ð¶ â ð21ðâ111 ðŽ
â+ð12 â ð21ðâ111 ð¶
â1ð¶2
= ðŽââð + ððŽâ â ð21ðâ111 (ð11ðŽ+ +ðŽâ+ð11)ð
â111 ð12 + ð¶
â2ð¶ + (ð¶â â ð¶â2 )ð¶2
= ðŽââð + ððŽâ + ð21ðâ111 ð¶
â1ð¶1ð
â111 ð12 â (ð¶ â ð¶2)
â(ð¶ â ð¶2) + ð¶âð¶
= ðŽââð + ððŽâ + ð¶âð¶ + ð21ð
â111 ð¶
â1ð¶1ð
â111 ð12 â (ð¶1ð
â111 ð12)
â(ð¶1ðâ111 ð12)
= ðŽââð + ððŽâ + ð¶âð¶.
Therefore,
Î2(ð)âÎ2(ð)â ðŒð = â(ð+ ðâ)ð¶ðâ1(ðâðŒðâð âðŽââ)â1ð(ððŒðâð âðŽâ)â1ðâ1ð¶â
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 113
and hence
Î2(ð)âÎ2(ð) = ðŒð on ðâ and Î2(ð)
âÎ2(ð) †ðŒð in Î +.
As ð(ðŽ1), ð(ðŽâ) â Î â, the rmvfâs Î1 and Î2 are holomorphic in Î + and henceÎ1 and Î2 are inner with respect to Î +. Therefore, as all inner rmvfâs are finiteBlaschkeâPotapov products, it follows that Î1 and Î2 are finite BlaschkeâPotapovproducts.
Next, (5) is verified by establishing the observability and controllability offour pairs of matrices. This will be done in four steps.
5.1 The pair (ð¶1, ðŽ1) is observable.
If there exist ð â â and ð¢ â âð, such that ð¶1ð¢ = 0 and ðŽ1ð¢ = ðð¢, thenâðâ111 ðŽ
â+ð11ð¢ = ðð¢ and equation (2.2) implies that
0 = ðŽâ+ð11ð¢+ ð11ðŽ+ð¢+ ð¶â1ð¶1ð¢
= âðð11ð¢+ ð11ðŽ+ð¢ = ð11(ðŽ+ð¢â ðð¢).Thus, ðŽ+ð¢ = ðð¢, as ð11 is invertible, and hence the observability of the pair(ð¶1, ðŽ+) implies ð¢ = 0.
5.2 The pair (ðŽ1, ðâ111 ð¶
â1 ) is controllable.
If there exist ð â â and 0 â= ð¢ â âð, such that ð¶1ðâ111 ð¢ = 0 and ðŽâ1ð¢ = ðð¢,
i.e., âð11ðŽ+ðâ111 ð¢ = ðð¢, then equation (2.2) implies that
0 = ðŽâ+ð¢+ ð11ðŽ+ðâ111 ð¢+ ð¶
â1ð¶1ð
â111 ð¢ = ðŽ
â+ð¢â ðð¢.
Thus, ð â ð(ðŽâ+)â©ð(âð11ðŽ+ðâ111 ) = ð(ðŽ
â+)â©ð(âðŽ+) = â , as ð(ðŽ+) â Î +,
a contradiction. Thus ð¢ = 0, i.e., the pair (ð¶1ðâ111 , ðŽ
â1) is observable.
5.3 The pair (ð¶,ðŽâ) is observable.
If there exist ð â â and a vector ð¢ â âðâð, such that ð¶ð¢ = 0 and ðŽâð¢ = ðð¢,then ð¶2ð¢ = ð¶1ð
â111 ð12ð¢ and equations (2.2) and (2.4) imply that
0 = ðð12ð¢+ðŽâ+ð12ð¢+ ð¶
â1ð¶1ð
â111 ð12ð¢
= (ððŒðâð +ðŽâ+ â (ðŽâ+ð11 + ð11ðŽ+)ð
â111 )ð12ð¢
= (ððŒðâð â ð11ðŽ+ðâ111 )ð12ð¢ = 0.
If ð12ð¢ = 0, then ð¶2ð¢ = 0 and thus the observability of the pair (ð¶2, ðŽâ)implies ð¢ = 0. If ð12ð¢ â= 0, then ð â ð(ð11ðŽ+ð
â111 ) = ð(ðŽ+). Then, ð â
ð(ðŽ+) â© ð(ðŽâ) â Î + â© Î â = â , a contradiction and so ð¢ = 0.
5.4 The pair (ðŽâ, ð ð¶â) is controllable.
The equation ðŽââð + ððŽâ + ð¶âð¶ = 0, with ð = âð â1, implies
âð ðŽââ âðŽâð + ð ð¶âð¶ð = 0. (4.10)
114 H. Dym and M. Porat
If there exist ð â â and ð¢ â âðâð, such that ð¶ð ð¢ = 0 and ðŽââð¢ = ðð¢, then,as follows from (4.10),
0 = âð ðŽââð¢âðŽâð ð¢+ ð ð¶âð¶ð ð¢= âðð ð¢âðŽâð ð¢ = â(ðŽâ + ððŒðâð)ð ð¢ .
Thus, if ð ð¢ â= 0, then ð â ð(âðŽâ). Therefore ð â ð(âðŽâ) â© ð(ðŽââ) = â , asð(ðŽâ) â Î â, and so ð¢ = 0 and the pair (ð¶ð,ðŽââ) is observable.
Finally, since the realization of Î is also minimal and
degÎ = degÎâ11 + degÎ2,
the factorization (4.7) is left coprime. â¡
Remark 4.2. The factorization (4.7) is a special case of a general theorem byKrein and Langer (see, e.g., [7]) which states that every Î â ð®ðÃð
ð (Î +), admitsa factorization
Î = Îâ11 Î2,
where Î1 is a BlaschkeâPotapov product of degree ð, Î2 is in the Schur classð®ðÃð(Î +) and kerÎ2(ð)
â â© kerÎ1(ð)â = {0}.
5. From the RKKS to the inertia theorem
Theorem 5.1. Let ðŽ â âðÃð and ð¶ â âðÃð. If ð(ðŽ) â© ðâ = â (i.e., â°0(ðŽ) = 0),the pair (ð¶,ðŽ) is observable, ð¹ (ð) is the ð à ð rmvf defined by the realizationformula
ð¹ (ð) = ð¶(ððŒð âðŽ)â1, (5.1)
ð = ð â â âðÃð is invertible (i.e., â°0(ð ) = 0) and the space
â³ = {ð¹ (ð)ð¢ : ð¢ â âð}endowed with the indefinite inner product
âšð¹ (ð)ð¢, ð¹ (ð)ð£â©â³ = ð£âðð¢ (5.2)
is an ð-dimensional RKKS, with RK
ðŸð(ð) =ðŒð âÎ(ð)Î(ð)â
ðð(ð), ð, ð â â â ð(ðŽ)
where
Î(ð) = ðŒð â ð¶(ððŒð âðŽ)â1ðâ1ð¶â, (5.3)
then ð is a solution of the Lyapunov equation (1.1) and â°Â±(ðŽ) = â°â(ð ).
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 115
Proof. It is readily checked that
ðŸð(ð) = ð¹ (ð)ðâ1ð¹ (ð)â, ð, ð â â â ð(ðŽ)
is a RK for the spaceâ³. Therefore, by the uniqueness of the RK,
ð¹ (ð)ðâ1ð¹ (ð)â =ðŒð âÎ(ð)Î(ð)â
ðð(ð), ð, ð â â â ð(ðŽ), (5.4)
which implies, by a straightforward calculation, that the matrix ð is a solution ofthe Lyapunov equation (1.1).
Let ð¶ =[ð¶1 ð¶2
]with ð¶1 â âðÃð and ð =
[ð11 ð12ð21 ð22
]with ð11 â âðÃð,
be conformal with the decomposition (2.1) of ðŽ with ðŽ+ â âðÃð and ð(ðŽÂ±) â Π±.The rest of the proof is divided into steps:
1. ð11 and ð22 â ð21ðâ111 ð12 are invertible matrices.
The verification is the same as the proof in Section 3. âŒ2. If Î1 and Î2 are given by formulas (4.8) and (4.9), then
Î(ð) = Î1(ð)â1Î2(ð).
Moreover, the realization formulas (4.8) and (4.9) are minimal.
The verification is the same as in steps 3 and 5 in the proof of Theorem 4.1. âŒ3. There exist ð1, . . . , ðð â â â ð(ðŽ) and ð£1, . . . , ð£ð â âð such that
ð =[ð¹ (ð1)
âÎ1(ð1)ð£1 . . . ð¹ (ðð)âÎ1(ðð)
âð£ð]
is invertible.
This follows from the fact that
â³1 = {ð¹ (ð)âð¢ : ð â Î + â ð(ðŽ), ð¢ â âð} = âð. (5.5)
To verify the equality in (5.5), note that if there is a vector ð£ â âð such thatð£ â¥â³1, then âšð¹ (ð)âð¢, ð£â© = âšð¢, ð¹ (ð)ð£â© = 0 for every ð¢ â âð and ð â â+ â ð(ðŽ),and hence ð¹ (ð)ð£ = 0 for every ð â â+ â ð(ðŽ). But this implies that ð£ = 0, sincethe pair (ð¶,ðŽ) is observable. Therefore, ð£ ⥠â³1 if and only if ð£ = 0 and thusequality holds in (5.5).
Therefore, as the matrices Î1(ð) and Î2(ð) are all invertible matrices, exceptfor a finite number of points ð in â, there exist ð1, . . . , ðð â â+ â ð(ðŽ) andð£1, . . . , ð£ð â âð such that the vectors
ð€1 = ð¹ (ð1)âÎ1(ð1)ð£1, . . . , ð€ð = ð¹ (ðð)
âÎ1(ðð)âð£ð
are linearly independent and so ð is invertible. âŒ4. Let
ðŸÎð ð (ð) =
ðŒð âÎð (ð)Îð (ð)â
ðð(ð)
116 H. Dym and M. Porat
be the RK of the finite-dimensional RKKS â(Îð ) = ð»ð2 âÎð ð»
ð2 , for ð = 1, 2. If
ð1 and ð2 are the ðà ð matrices defined by the formulas
(ð1)ðð = ð£âððŸ
Î1ðð(ðð)ð£ð ððð (ð2)ðð = ð£
âððŸ
Î2ðð(ðð)ð£ð , (5.6)
then ð1 ⥠0, ð2 ⥠0 and
ð âðâ1ð =ð2 âð1. (5.7)
For every ð¥ â âð and ð = 1, 2,
âšðð ð¥, ð¥â© =ðâ
ð,ð=1
ð¥ðð¥âð (ðð )ðð =
ðâð,ð=1
ð¥âð ð£âððŸ
Îð ðð(ðð)ð£ðð¥ð
=
âšðâ
ð=1
ðŸÎð ððð£ðð¥ð,
ðâð=1
ðŸÎð ððð£ðð¥ð
â©â(Îð )
⥠0,
i.e., ðð ⥠0 for ð = 1, 2. Next, equation (5.4) implies, for every ð, ð â â â ð(ðŽ),
ð¹ (ð)ðâ1ð¹ (ð)â =ðŒð âÎ1(ð)
â1Î2(ð)Î2(ð)âÎ1(ð)
ââ
ðð(ð),
i.e.,
Î1(ð)ð¹ (ð)ðâ1ð¹ (ð)âÎ1(ð)
â =Î1(ð)Î1(ð)
â âÎ2(ð)Î2(ð)â
ðð(ð)(5.8)
for every ð, ð â â â ð(ðŽ). Therefore,(ð âðâ1ð )ðð = ð£
âðÎ1(ðð)ð¹ (ðð)ð
â1ð¹ (ðð)âÎ1(ðð)âð£ð
= ð£âð
[Î1(ðð)Î1(ðð)
â âÎ2(ðð)Î2(ðð)â
ððð (ðð)
]ð£ð
= ð£âð
[ðŒð âÎ2(ðð)Î2(ðð)
â
ððð (ðð)
]ð£ð â ð£âð
[ðŒð âÎ1(ðð)Î1(ðð)
â
ððð (ðð)
]ð£ð
= ð£âððŸÎ2ðð(ðð)ð£ð â ð£âððŸÎ1
ðð(ðð)ð£ð
= (ð2)ðð â (ð1)ðð
and thus (5.7) holds. âŒ
5. â°+(ð ) †rankð2 and â°â(ð ) †rankð1.
Formula (5.7) can be reexpressed as
ð âðâ1ð =[ðŒð ðŒð
] [âð1 00 ð2
] [ðŒððŒð
]and since ð is invertible, the Sylvester law of inertia implies
â°Â±(ð ) = â°Â±(ðâ1) = â°Â±(ð âðâ1ð ) †â°Â±([âð1 0
0 ð2
]).
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 117
Therefore, since ð1 ⥠0 and ð2 ⥠0,
â°+([âð1 0
0 ð2
])= rankð2 and â°â
([âð1 00 ð2
])= rankð1,
as claimed. âŒ6. rankð1 †degÎ1 and rankð2 †degÎ2.
Since ðŸÎð ð (ð) = ð¹ð (ð)ð ð ð¹ð (ð)
â for ð = 1, 2, with
ð¹1(ð) = ð¶1(ððŒð âðŽ1)â1, ð¹2(ð) = ð¶(ððŒðâð âðŽâ)â1 and ð ð > 0
and the exhibited realizations are minimal, the matrices
(ðð )ðð = ð£âððŸ
Îð ðð(ðð)ð£ð = ð£
âð ð¹ð (ðð)ð ð ð¹ð (ðð)
âð£ð
are of the form ðð = ðâð ð ð ðð , where ðð =
[ð¹ð (ð1)
âð£1 . . . ð¹ð (ðð)âð£ð.
]There-
fore,rankðð = rank(ðâð ð ð ðð ) †rankð ð = degÎð . âŒ
7. â°Â±(ð ) = â°â(ðŽ).It is well known (see, e.g., [3]) that if the realization
Î(ð) = ᅵᅵ + ð¶(ððŒð âðŽ)â1ðµis minimal, then deg Î is equal to the size of the matrix ðŽ. Thus, Step 2 guaranteesthat degÎ1 = â°+(ðŽ) and degÎ2 = â°â(ðŽ), and hence, Steps 5 and 6 imply thatâ°+(ð ) †â°â(ðŽ) and â°â(ð ) †â°+(ðŽ). Therefore, as â°0(ð ) = â°0(ðŽ) = 0,
ð = â°+(ð ) + â°â(ð ) †â°â(ðŽ) + â°+(ðŽ) = ð,which means that â°+(ð ) = â°â(ðŽ) and â°â(ð ) = â°+(ðŽ). â¡
6. A disc analog
In this section we shall sketch an analog of the preceding analysis for the Steinequation
ð âðŽâððŽ = ð¶âð¶, (6.1)
The notations ð» and ðŒ for the open unit disc and the exterior of the closedunit disc and ðâ(ðŽ), ð0(ðŽ) and ð+(ðŽ) for the sum of the algebraic multiplicitiesof the eigenvalues of a matrix ðŽ â âðÃð in ðŒ,ð and ð», respectively, will beneeded to formulate the version of the CarlsonâSchneider inertia theorem underconsideration. It is equivalent to Theorem 4 on p. 453 of [8].
Theorem 6.1. If ðŽ â âðÃð, ð¶ â âðÃð, (ð¶,ðŽ) is an observable pair and ð = ð â ââðÃð is a solution of the Stein equation (6.1), then
(1) ð(ðŽ) â© ð = â and ð is invertible, i.e., ð0(ðŽ) = â°0(ð ) = 0,(2) ð+(ðŽ) = â°+(ð ) and ðâ(ðŽ) = â°â(ð ).
This section is organized as follows: In subsection 6.1 we present some pre-liminary facts and notation for subsequent use. A short well-known proof of Theo-
118 H. Dym and M. Porat
rem 6.1 based on elementary tools of linear system theory is presented in Subsec-tion 6.2. In Subsection 6.3, Theorem 6.1 is used to establish a finite-dimensionalreproducing kernel Krein space with a reproducing kernel of the special form
ðŸð(ð) =ðŒð âÎ(ð)Î(ð)â
ðð(ð)(6.2)
where ðð(ð) = 1 â ððâ and Î admits a factorization Î = Îâ11 Î2, with Î1 andÎ2 both inner rmvfâs with respect to ð». Finally, a proof of Theorem 6.1 thatis based on methods and formulas of reproducing kernel spaces is presented inSubsection 6.4.
6.1. Preliminaries
Lemma 6.2. If the assumptions in Theorem 6.1 are in force, then ð(ðŽ) â© ð = â and without loss of generality ðŽ may be assumed to be of the form ðœ = ðŽÂ± or
ðŽ =
[ðŽâ 00 ðŽ+
], with ð(ðŽâ) â ðŒ and ð(ðŽ+) â ð». (6.3)
Proof. The proof is separated into parts:
1. If ðŽð¢ = ðð¢ for some ð¢ â âð and ð â ð, then ð¢âðŽâ = ðâð¢â and hence (6.1)implies that
0 = ð¢âðð¢â ð¢âðŽâððŽð¢â ð¢âð¶âð¶ð¢ = âð¢âð¶âð¶ð¢.Thus, ð¶ðŽðð¢ = ððð¶ð¢ = 0, which implies that ð¢ = 0, since the pair (ð¶,ðŽ) isobservable. Therefore, ð(ðŽ) â© ð = â .
2. As ð(ðŽ) â© ð = â , ðŽ = ððœðâ1, where ðœ is a Jordan matrix of the formðœ = ðŽÂ± or
ðœ =
[ðŽâ 00 ðŽ+
]with ð(ðŽâ) â ðŒ and ð(ðŽ+) â ð».
Then ð is an invertible Hermitian solution of equation (6.1) if and only ifðâðð is an invertible Hermitian solution of the equation
ðâðð â ðœâ(ðâðð)ðœ = (ð¶ð)âð¶ð.
Moreover, ð(ðœ) â© ð = â , ð±(ðœ) = ð±(ðŽ), â°Â±(ðâðð) = â°Â±(ð ) and the pair(ð¶ð, ðœ) is observable, since
rank
[ððŒð â ðŽð¶
]= rank
[ð 00 ðŒð
] [(ððŒð â ðœ)ð¶ð
]ðâ1 = rank
[(ððŒð â ðœ)ð¶ð
]. â¡
In view of the preceding lemma, we can without loss of generality assumethat ðŽ is of the form (6.3) and let ð := ðâ(ðŽ). Just as in the half-plane case, weshall assume that 1 †ð †ð â 1. Correspondingly, let ð¶ =
[ð¶1 ð¶2
]with ð¶1 â
âðÃð, ð¶2 â âðÃ(ðâð) and ð =
[ð11 ð12ð21 ð22
]with ð11 â âðÃð, ð22 â â(ðâð)Ã(ðâð)
be the decompositions of the matrices ð¶ and ð that are conformal with that ofðŽ. Then the Stein equation (6.1) is equivalent to the first, second and fourth of
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 119
the following four equations:
ð11 âðŽââð11ðŽâ = ð¶â1ð¶1, (6.4)
ð21 âðŽâ+ð21ðŽâ = ð¶â2ð¶1, (6.5)
ð12 âðŽââð12ðŽ+ = ð¶â1ð¶2 (6.6)
and
ð22 âðŽâ+ð22ðŽ+ = ð¶â2ð¶2. (6.7)
Finally, let ð®ðÃðð (ð») denote the generalized Schur class of ðÃð mvfâs ð (ð), that
are meromorphic in ð» and for which the kernel
Îð ð(ð) =
ðŒð â ð (ð)ð (ð)âðð(ð)
has ð negative squares on ð¥+ð à ð¥+ð , and ð¥+ð denotes the domain of analyticity of ð in ð».
6.2. A quick proof of Theorem 6.1
In this section we present a variant of the well-known short proof of Theorem 6.1
Proof. It is verified in Lemma 6.2 that ð(ðŽ) â© ð = â . Next, as ð(ðŽâ) â ðŒ andð(ðŽ+) â ð», (6.4) and (6.7) both have unique solutions:
ð11 = âââ
ð=0(ðŽââ)
âðâ1ð¶â1ð¶1ðŽâðâ1â †0, and ð22 =
ââð=0
(ðŽâ+)ðð¶â2ð¶2ðŽ
ð+ ⥠0
respectively. The observability of the pair (ð¶,ðŽ) implies that the pairs (ð¶1, ðŽâ)and (ð¶2, ðŽ+) are observable, as is perhaps verified most easily by the PopovâBelevitchâHautus test. Thus,
ð11ð¢ = 0 =â 0 ⥠âââ
ð=0(ð¶1ðŽ
âðâ1â ð¢)â(ð¶1ðŽ
âðâ1â ð¢) = 0
=â ð¶1ðŽâðâ1â ð¢ = 0, ð = 0, 1, . . . =â ð¢ = 0,
i.e., ð11 is invertible and by similar calculations it can be proved that ð22 is alsoinvertible. Therefore, ð11 < 0, ð21 = ð â12 and ð22 â ð â12ðâ111 ð12 ⥠ð22 > 0. BySchur complements,
ð =
[ðŒð ðâ111 ð120 ðŒðâð
]â [ð11 00 ð22 â ð â12ðâ111 ð12
] [ðŒð ðâ111 ð120 ðŒðâð
]is invertible and thus, by the Sylvester law of inertia,
â°â(ð ) = â°â(ð11) + â°â(ð22 â ð21ðâ111 ð12)
= â°â(ð11) = ð = ðâ(ðŽ)and
â°+(ð ) = â°+(ð11) + â°+(ð22 â ð21ðâ111 ð12)
= â°+(ð22 â ð21ðâ111 ð12) = ðâ ð = ð+(ðŽ). â¡
120 H. Dym and M. Porat
6.3. From the inertia theorem to a RKKS
In this section we will establish a finite-dimensional reproducing kernel Krein spaceâ³ with a RK of the form (6.2) and then use the inertia theorem to obtain a coprimefactorization formula for the rmvf Î â ð®ðÃð
â°â(ð )(ð») of the form
Î = Îâ11 Î2,
with Î1 and Î2 both inner rmvfâs with respect to ð» and hence BlaschkeâPotapovproducts. Moreover, we will give explicit realization formulas for Î1 and Î2 andprove that they are minimal.
Theorem 6.3. If ðŽ â âðÃð, ð¶ â âðÃð, the pair (ð¶,ðŽ) is observable and ð =ð â â âðÃð is a solution of the Stein equation (6.1), then:
(1) ð(ðŽ) â© ð = â and ð is invertible, i.e., ð0(ðŽ) = â°0(ð ) = 0,(2) The space
â³ = {ð¹ (ð)ð¢ : ð¢ â âð} (6.8)
with
ð¹ (ð) = ð¶(ðŒð â ððŽ)â1, ðâ1 â â â ð(ðŽ) (6.9)
and indefinite inner product
âšð¹ (ð)ð¢, ð¹ (ð)ð£â©â³ = ð£âðð¢, (6.10)
is an ð-dimensional RKKS with RK
ðŸð(ð) = ð¹ (ð)ðâ1ð¹ (ð)â, ðâ1, ðâ1 â â â ð(ðŽ). (6.11)
(3) The RK may be expressed in the form
ðŸð(ð) =ðŒð âÎ(ð)Î(ð)â
ðð(ð), ðâ1, ðâ1 â â â ð(ðŽ), (6.12)
where ð â ð and
Î(ð) = ðŒð â ðð(ð)ð¶(ðŒð â ððŽ)â1ðâ1(ðŒð â ðâðŽâ)â1ð¶â. (6.13)
The matrices ð11 and ð22 â ð21ðâ111 ð12 are invertible and the rmvf Î(ð)admits the factorization
Î(ð) = Î1(ð)â1Î2(ð), (6.14)
with ðŽ of the form (6.3),
Î1(ð) = ðŒð â (ðâ ð)ð¶1(ðŒð â ððŽâ)â1ðâ111 (ððŒð âðŽââ)â1ð¶â1 (6.15)
Î2(ð) = ðŒð â ðð(ð)ð¹ (ð)ð1(ðŒðâð â ððŽ+)(ðŒðâð â ððŽ+)â1ð2ð
â1ð¹ (ð)
â, (6.16)
ð1 =
[âðâ111 ð12ðŒðâð
]ððð ð2 = (ð22 â ð21ðâ111 ð12)
â1.
(4) The rmvfâs Î1 and Î2 are finite BlaschkeâPotapov products that are bothinner with respect to ð» and are left coprime.
(5) The realizations (6.15) and (6.16) are minimal.
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 121
Proof. Assertion (1) is verified in Lemma 6.2. Next, sinceâ³ is a RKKS with RKthat is given by formula (6.11) and (ð¶,ðŽ) is an observable pair, the columns ofthe rmvf ð¹ are linearly independent. Therefore, (2) holds.
Formula (6.12) in (3) may be verified by a straightforward calculation thatuses the Stein equation (6.1). Moreover, it is easily seen that ð11 is invertible (justas in the proof in Section 6.2 and hence that ð22 â ð12ðâ111 ð21 is also invertible(by the Schur complements formula for ð ) and ð1 and ð2 are well defined. Next,a lengthy calculation that takes advantage of the formulas (6.13), (6.15) and
(ððŒð âðŽâ)â1ð¶âð¶(ðŒð â ððŽ)â1 = 1
ðð (ðŒð â ððŽ)â1 + 1
ððŽâ(ððŒð âðŽâ)â1ð,
serves to verify that Î1(ð)Î(ð) = Î2(ð) and hence that (6.14) holds.The observability of the pairs (ð¶1, ðŽâ) and (ð¶2, ðŽ+) is inherited from the
observability of the pair (ð¶,ðŽ) (as is most easily seen by the PopovâBelevitchâHautus test). Therefore, successive applications of Theorem 6.1 to equation (6.4)for the ð à ð matrix ð11 and to equation (6.7) for the (ð â ð) à (ð â ð) matrixð22, yields the implications
ð±(ðŽâ) = â°Â±(ð11) =â â°â(ð11) = ð =â ð11 < 0
and
ð±(ðŽ+) = â°Â±(ð22) =â â°+(ð22) = ðâ ð =â ð22 > 0.
Thus, as ð22 > 0 and ð12 = ðâ21,
ð2 = (ð22 â ð21ðâ111 ð12)â1 > 0.
A direct calculation shows that
ðŒð âÎ1(ð)Î1(ð)â = â(1â â£ðâ£2)Ί1ð11Ί
â1, (6.17)
where Ί1 = ð¶1(ðŒð â ððŽâ)â1ðâ111 (ððŒð âðŽââ)â1(ðŒð â ðâðŽââ). A similar calculation,based on the fact that
ðŸ := ð â ð[ðâ111 00 0
]ð =
[0 00 ðâ12
](6.18)
is a solution of the Riccati equation
ðŸ âðŽâðŸðŽ = (ðŒð â ðâðŽâ)ðŸðâ1ð¹ (ð)âð¹ (ð)ðâ1ðŸ(ðŒð â ððŽ), (6.19)
shows that
ðŒð âÎ2(ð)âÎ2(ð) = (1â â£ðâ£2)Ί2ðŸÎŠ
â2, (6.20)
where Ί2 = ð¹ (ð)ðâ1(ðŒðâðâðŽâ)â1(ðŒðâðâðŽâ). Since ð(ðŽ+) â ð» and ð(ðŽââ) â ðŒ,
the rmvfâs Î1 and Î2 are holomorphic in ð» and as ð11 < 0 and ðŸ ⥠0, it followsfrom formulas (6.17) and (6.20) that the rmvfâs Î1 and Î2 are inner with respectto ð». Therefore, as all inner rmvfâs are finite BlaschkeâPotapov products, Î1 andÎ2 are finite BlaschkeâPotapov products.
122 H. Dym and M. Porat
Next, since observability and controllability for realizations of the form (6.13),and in particular for the realizations (6.15) and (6.16), may also be verified byapplying the PopovâBelevitchâHautus test to the pairs (ð¶,ðŽ) and (ðŽ,ðµ), withðµ = ðâ1(ðŒð â ðâðŽâ)â1ð¶â, it suffices to check that
rank
[ððŒð âðŽð¶
]= ð and rank
[ððŒð âðŽ ðµ
]= ð
for all points ð â â (see, e.g., Theorem 3.5 in [1], where realizations of this formare discussed). This will be done in four steps.
6.1 The pair (ð¶1(ðŒð â ððŽâ)â1ðâ111 , ðŽââ) is observable.
If there exist ð¢ â= 0 in âð and ð â â such that ð¶1(ðŒð â ððŽâ)â1ðâ111 ð¢ = 0and ðŽââð¢ = ðð¢, then
0 = ð¶â1ð¶1(ðŒð â ððŽâ)â1ðâ111 ð¢
= [ð11(ðŒð â ððŽâ) + (ððŒð âðŽââ)ð11ðŽâ](ðŒð â ððŽâ)â1ðâ111 ð¢
= (ððŒð âðŽââ)[(ðâ ð)â1ðŒð + ð11ðŽâ(ðŒð â ððŽâ)â1ðâ111 ]ð¢
= (ððŒð âðŽââ)[ð11(ðŒð â ððŽâ) + (ðâ ð)ð11ðŽâ](ðŒð â ððŽâ)â1ðâ111 ð¢
= (ððŒð âðŽââ)ð11(ðŒð â ððŽâ)(ðŒð â ððŽâ)â1ðâ111 ð¢
which, as (ððŒð âðŽââ)ð11 is invertible and ð¢ â= 0, implies that ð â ð». On theother hand, ð â ð(ðŽââ) =â ð â ðŒ, which is a contradiction.
6.2 The pair (ðŽââ, ð¶â1 ) is controllable.
This follows immediately from the observability of the pair (ð¶1, ðŽâ).
6.3 The pair (ð¹ (ð)ð1(ðŒðâð â ððŽ+), ðŽ+) is observable.
If there exist ð â= 0 in âðâð and ð â â such that ð¹ (ð)ð1(ðŒðâð â ððŽ+)ð = 0and ðŽ+ð = ðð, then
0 = ð¹ (ð)
[âðâ111 ð12ðŒðâð
](1 â ðð)ð
= (1 â ðð)[âð¶1(ðŒð â ððŽâ)â1ðâ111 ð12 + ð¶2(ðŒðâð â ððŽ+)â1]ð
= (1 â ðð)[âð¶â1ð¶1(ðŒð â ððŽâ)â1ðâ111 ð12 + ð¶â1ð¶2(1â ðð)â1]ð
= (ðð â 1)ð¶â1ð¶1(ðŒð â ððŽâ)â1ðâ111 ð12ð + ð¶â1ð¶2ð.
The equations (6.4) and (6.6) imply that
ð¶â1ð¶1 = (ðŒð â ðâðŽââ)ð11 + ðâðŽââð11(ðŒð â ððŽâ)and
ð¶â1ð¶2ð = (ðŒð â ððŽââ)ð12ð
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 123
and hence
0 = [(ððâ 1)(ðŒð â ðâðŽââ)ð11(ðŒð â ððŽâ)â1ðâ111
+ (ððâ 1)ðâðŽââ + (ðŒð â ððŽââ)]ð12ð= [(ððâ 1)(ðŒð â ðâðŽââ)ð11(ðŒð â ððŽâ)â1ðâ111 + (ðŒð â ðâðŽââ)]ð12ð= (ðŒð â ðâðŽââ)ð11[(ððâ 1)ðŒð + ðŒð â ððŽâ](ðŒð â ððŽâ)â1ðâ111 ð12ð
= ð(ðŒð â ðâðŽââ)ð11(ððŒð âðŽâ)(ðŒð â ððŽâ)â1ðâ111 ð12ð.
Therefore, since ð â ð, ð â ð» and ð11 is invertible, ð12ð = 0 and then bythe preceding calculation
ð¶2(ðŒðâð â ððŽ+)â1ð = 0 =â ð¶2(1â ðð)â1ð = 0 =â ð¶2ð = 0 .
Finally, ðŽ+ð = ðð and ð¶2ð = 0, together with the observability of (ð¶2, ðŽ+)lead to contradiction.
6.4 The pair (ðŽ+, ð2ðâ1ð¹ (ð)
â) is controllable.
We shall prove that the pair (ð¹ (ð)ð1ð2, ðŽâ+) is observable: If there exist
nonzero ð â âðâð and ð â â such that ðŽâ+ð = ðð and ðð¹ (ð)ð1ð2ð = 0,then, by (6.19),
ð1ð¹ (ð)âð¹ (ð)ð1
=[0 ðŒðâð
]ðŸðâ1ð¹ (ð)âð¹ (ð)ðâ1ðŸ
[0ðŒðâð
]=[0 ðŒðâð
](ðŒð â ðâðŽâ)â1(ðŸ âðŽâðŸðŽ)(ðŒð â ððŽ)â1
[0ðŒðâð
]= (ðŒðâð â ðâðŽâ+)â1(ðâ12 âðŽâ+ðâ12 ðŽ+)(ðŒðâð â ððŽ+)
â1
= ðâ12 (ðŒðâð â ððŽ+)â1 + (ðŒðâð â ðâðŽâ+)â1ðâðŽâ+ðâ12 ,
and hence
0 = ðâð2ðâ1ð¹ (ð)
âð¹ (ð)ð1ð2ð
= ðâ(ðŒðâð â ððŽ+)â1ð2ð + ð
âð2(ðŒðâð â ðâðŽâ+)â1ðâðŽâ+ð= ðâ(1â ððâ)â1ð2ð + ð
âð2(1â ðâð)â1ðâðð
=
((1â ðâð) + ðâð(1 â ððâ)
â£1â ððââ£2)ðâð2ð =
(1â â£ðâ£2â£1â ððââ£2
)ðâð2ð.
Therefore, as ð /â ð and ð2 > 0, we have ð = 0, which is a contradiction.
Finally, since the realization of Î is also minimal and
degÎ = degÎâ11 + degÎ2,
the factorization (6.14) is left coprime. â¡
124 H. Dym and M. Porat
6.4. From the RKKS to the inertia theorem
Theorem 6.4. Let ðŽ â âðÃð and ð¶ â âðÃð. If ð(ðŽ) â© ð = â (i.e., ð0(ðŽ) = 0),the pair (ð¶,ðŽ) is observable, ð¹ (ð) is the ð à ð rmvf defined by the realizationformula
ð¹ (ð) = ð¶(ðŒð â ððŽ)â1, (6.21)
ð = ð â â âðÃð is invertible (i.e., â°0(ð ) = 0) and the space
â³ = {ð¹ (ð)ð¢ : ð¢ â âð}endowed with the indefinite inner product
âšð¹ (ð)ð¢, ð¹ (ð)ð£â©â³ = ð£âðð¢ (6.22)
is an ð-dimensional RKKS, with RK
ðŸð(ð) =ðŒð âÎ(ð)Î(ð)â
ðð(ð), ðâ1, ðâ1 â â â ð(ðŽ)
where
Î(ð) = ðŒð â ðð(ð)ð¶(ðŒð â ððŽ)â1ðâ1(ðŒð â ðâðŽâ)â1ð¶â, (6.23)
then ð is a solution of the Stein equation (6.1) and ð±(ðŽ) = â°Â±(ð ).Proof. It is readily checked that
ðŸð(ð) = ð¹ (ð)ðâ1ð¹ (ð)â, ðâ1, ðâ1 â â â ð(ðŽ)
is a RK for the spaceâ³. Therefore, by uniqueness of the RK,
ð¹ (ð)ðâ1ð¹ (ð)â =ðŒð âÎ(ð)Î(ð)â
ðð(ð), ðâ1, ðâ1 â â â ð(ðŽ), (6.24)
which implies, by a straightforward calculation, that the matrix ð is a solution ofthe Stein equation (6.1).
Let ð¶ =[ð¶1 ð¶2
]with ð¶1 â âðÃð and ð =
[ð11 ð12ð21 ð22
]with ð11 â âðÃð,
be conformal with the decomposition (6.3) of ðŽ with ðŽâ â âðÃð.
The rest of the proof is divided into steps:
1. ð11 and ð22 â ð21ðâ111 ð12 are invertible matrices.
The verification is the same as the proof in Section 6.2. âŒ2. If Î1 and Î2 are given by formulas (6.15) and (6.16), then
Î(ð) = Î1(ð)â1Î2(ð).
Moreover, the realization formulas (6.15) and (6.16) are minimal.
The verification is the same as in steps 3 and 5 in the proof of Theorem 6.3. âŒ3. Repeat steps 3 to 6 of Theorem 5.1. They are applicable to this setting, with onlyminor changes in the proof.
4. â°Â±(ð ) = ð±(ðŽ).
Long Proofs of Two CarlsonâSchneider Type Inertia Theorems 125
It is well known (see, e.g., [1]) that if the realization
Î(ð) = ᅵᅵ + ðð(ð)ð¶(ððŒ âðŽ)â1ðµ (or Î(ð) = ᅵᅵ + ðð(ð)ð¶(ðŒ â ððŽ)â1ðµ)is minimal, then degÎ is equal to the size of the matrix ðŽ. Thus, Step 2 guar-antees that degÎ1 = ðâ(ðŽ) and degÎ2 = ð+(ðŽ), and hence, Steps 5 and 6of Theorem 5.1 imply that â°+(ð ) †ð+(ðŽ) and â°â(ð ) †ðâ(ðŽ). Therefore, asâ°0(ð ) = ð0(ðŽ) = 0,
ð = â°+(ð ) + â°â(ð ) †ð+(ðŽ) + ðâ(ðŽ) = ð,which means that â°+(ð ) = ð+(ðŽ) and â°â(ð ) = ðâ(ðŽ). â¡
References
[1] D. Alpay and H. Dym, On a New Class of Realization Formulas and Their Applica-tion, Linear Algebra and Its Applications, 241-243 (1996), 3â84.
[2] C.T. Chen, A generalization of the inertia theorem, SIAM J. Appl. Math. 25 (1973),158â161.
[3] H. Dym, Linear Algebra in Action, Graduate Studies in Mathematics, 78 AmericanMathematical Society, Providence, 2007.
[4] V. Derkach and H. Dym, On Linear Fractional Transformations Associated withGeneralized ðœ-Inner Matrix Functions, Integral Equations and Operator Theory, 65(2009), 1â50.
[5] H. Dym and D. Volok, Zero Distribution of matrix polynomials, Linear Algebra andits Applications, 425 (2007), 714â738.
[6] I. Gohberg, P. Lancaster and L. Rodman, Matrix Polynomials, Academic Press, NewYork-London, 1982.
[7] M.G. Krein and H. Langer, Uber die verallgemeinerten Resolventen und die cha-rakteristische Funktion eines isometrischen Operators im Raume Î ð, Hilbert spaceOperators and Operator Algebras (Proc. Intern. Conf., Tihany, 1970); Colloq. Math.Soc. Janos Bolyai, 5 (1972), 353â399.
[8] P. Lancaster and M. Tismenetsky, The Theory of Matrices, second edition withapplications, Computer Science and Applied Mathematics, Academic Press, NewYork, 1985.
[9] L. Lerer and M. Tismenetsky, The Bezoutian and The Eigenvalue-Separation Prob-lem for Matrix Polynomials, Integral Equations and Operator Theory, 5 (1982),386â445.
[10] H.K. Wimmer, Inertia theorems for matrices, controllability, and linear vibrations,Linear Algebra and its Applications 8 (1974), 337â343
Harry Dym and Motke PoratDepartment of MathematicsThe Weizmann Institute of ScienceRehovot 76100, Israele-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 127â144câ 2013 Springer Basel
On the Kernel and Cokernel ofSome Toeplitz Operators
Torsten Ehrhardt and Ilya M. Spitkovsky
To Professor Leonia Lerer, in celebration of his seventieth birthday.
Abstract. We show that the kernel and/or cokernel of a block Toeplitz op-erator ð (ðº) are trivial if its matrix-valued symbol ðº satisfies the condi-tion ðº(ð¡â1)ðº(ð¡)â = ðŒð . As a consequence, the WienerâHopf factorizationof ðº (provided it exists) must be canonical. Our setting is that of weightedHardy spaces on the unit circle. We extend our result to Toeplitz operatorson weighted Hardy spaces on the real line, and also Toeplitz operators onweighted sequence spaces.
Mathematics Subject Classification (2010). Primary 47B35. Secondary 47A68,47B30.
Keywords. Toeplitz operators, WienerâHopf factorization, partial indices, dis-crete convolution operators.
1. Introduction
The Wiener algebra ð by definition consists of functions ð defined on the unitcircle ð = { ð§ â â : â£ð§â£ = 1 } and having absolutely convergent Fourier series:
ð(ð¡) =
ââð=ââ
ððð¡ð , where
ââð=ââ
â£ðð ⣠<â. (1.1)
The functions ð â ð with ðð = 0 for all ð †0 (resp., ð ⥠0) form the subalgebrað± ofð the elements of which admit analytic continuations to the interior (resp.,exterior, including the point of infinity) of ð. It is a classical result by Gohbergand Krein [9] (see also the monographs [7, 12] and a survey [8] for more detailedbibliographical information and far reaching generalizations) that any invertible
128 T. Ehrhardt and I.M. Spitkovsky
matrix function1 ðº âððÃð admits a representation
ðº(ð¡) = ðºâ(ð¡)Î(ð¡)ðº+(ð¡), ð¡ â ð, (1.2)
where ðºÂ±1+ âððÃð+ , ðºÂ±1â âððÃð
â ,
Î(ð¡) = diag[ð¡Ï°1 , . . . , ð¡Ï°ð ] (1.3)
and Ï°1, . . . ,Ï°ð â â€. Representation (1.2) plays a central role in a variety ofapplications, including systems of convolution type equations on the half-line andToeplitz operators ð (ðº) with matrix symbols ðº. In particular, the defect numbers(i.e., dimensions of the kernel and cokernel) of these operators are expressed interms of the partial indices Ï°ð . Namely,
dimkerð (ðº) = ââÏ°ðâ€0
Ï°ð , dim kerð (ðº)â =âÏ°ðâ¥0
Ï°ð . (1.4)
Note that the partial indices are defined by ðº uniquely, up to their order.However, for ð > 1 the partial indices, and the factorization itself, are genericallynot stable: a necessary and sufficient stability criterion (also going back to [9])reads
max{Ï°ð â Ï°ð : ð, ð = 1, . . . , ð} †1, (1.5)
and thus requires the a priori knowledge of the partial indices. The latter is avail-able in some particular cases, e.g., for rational, triangular, or sectorial matrixfunction, see, e.g., [8], but in general the problem remains open.
One recent result in this direction, obtained by Voronin2 [13], claims that allthe partial indices are equal to zero (and thus (1.5) holds) for matrix functions ðºsatisfying
ðº(ð¡â1)ðº(ð¡)â = ðŒð , ð¡ â ð. (1.6)
The proof, published in [14], makes use of the description of all factorizations(1.2) of a given matrix function ðº âððÃð . In this paper, we propose a differentapproach, which provides information about the defect numbers of ð (ðº) with ðºsatisfying (1.6) but not necessarily lying in ððÃð and, moreover, not necessarilyfactorable. Namely, we will show that (under certain mild additional conditionson the spaces where the operators act) at least one of the defect numbers is zero.Note that according to Coburnâs lemma this property holds for general Toeplitzoperators with scalar non-zero symbols, but fails starting with ð = 2. Additionalconditions on the matrix symbol ðº under which the property persists are of greatinterest. Some such conditions, analytic in nature and thus very different from(1.6), were established in [6].
In Section 3 we consider ð (ðº) with measurable bounded symbols ðº actingon weighted Hardy spaces on the unit circle or the real line. Section 4 deals withoperators acting on weighted discrete âð. Dealing with weighted spaces presents an
1Here and below we are using the standard notational convention: given any set ð, ððÃð
stands for the set of all ðÃð matrices with the entries in ð, and ððÃ1 is abbreviated to ðð .2Voronin works with matrix functions defined on the real line â but the transition between â
and ð is obvious via an appropriate linear fractional transformation.
Toeplitz Operators 129
additional difficulty, and certain nested properties had to be established in orderto overcome it. These properties are tackled in Section 2, along with the relatedresults on the kernels of homogeneous RiemannâHilbert problems.
We end this introduction with some basic observations about condition (1.6).Firstly, the set of matrix functions satisfying (1.6) forms a group under multiplica-tion. Furthermore, such matrix functions can be defined arbitrarily on a half-circle(without loss of generality, say for Im(ð¡) > 0); the values on the complementaryhalf-circle are then determined by (1.6) uniquely. Note also that (1.6) holds if ðºis even (that is, ðº(ð¡â1) = ðº(ð¡)) and unitary valued.
2. Homogeneous RiemannâHilbert problems on the unit circle
For 1 †ð < â and a given positive weight ð on the unit circle ð, let ð¿ð(ð; ð)denote the space of all measurable functions ð defined on ð and such that
â¥ðâ¥ð,ð :=(â« 2ð
0
â£ð(ððð¥)ð(ððð¥)â£ð ðð¥)1/ð
<â. (2.1)
In case ð â¡ 1 we simply write ð¿ð(ð). Throughout the paper we will assume that
1 < ð <â, 1
ð+1
ð= 1, (2.2)
and require that the weight ð satisfies
ð â ð¿ð(ð) and ðâ1 â ð¿ð(ð). (2.3)
The conditions (2.3) imply that ð¿â(ð) â ð¿ð(ð; ð) â ð¿1(ð). Therefore we candefine the Fourier coefficients of a function ð â ð¿ð(ð; ð),
ðð =1
2ð
â« 2ð
0
ð(ððð¥) ðâððð¥ ðð¥. (2.4)
We also introduce the weighted Hardy spaces
ð¿ð+(ð; ð) =
{ð â ð¿ð(ð; ð) : ðð = 0 for all ð < 0
}, (2.5)
ð¿ðâ(ð; ð) =
{ð â ð¿ð(ð; ð) : ðð = 0 for all ð > 0
}, (2.6)
as well as
ð¿ðâ,0(ð; ð) =
{ð â ð¿ð(ð; ð) : ðð = 0 for all ð ⥠0
}. (2.7)
In order to present our result, the following notation will be handy. For amatrix or vector function ð defined on ð, we introduce the âtildeâ operation,
ð(ð¡) = ð(ð¡â1), ð¡ â ð,
as well as the complex adjoint function ðâ,
ðâ(ð¡) = ð(ð¡)ð, ð¡ â ð,
130 T. Ehrhardt and I.M. Spitkovsky
which is the function obtained by taking the transpose and the complex conjugatepointwise on ð. The complex adjoint and the tilde operation commute with eachother. Therefore, the notation ðâ is unambiguous. We also note that
ð â ð¿ð±(ð; ð) =â ð â ð¿ð
â(ð; ð)and
ð â ð¿ð±(ð; ð) =â ðâ â ð¿ð
â(ð; ð).
The main results of this section are based on sufficient conditions whichstate that certain spaces are nested. These conditions will be analyzed first in thefollowing lemma. Notice that we only need the much simpler âifâ parts. The proofof the âonly ifâ parts is provided for completenessâ sake.
Given (2.2), we also introduce ð â (1,â] by1
ð+
1
max{ð, ð} =1
min{ð, ð} .Equivalently,
ð =
{ððâ£ðâð⣠if ð â= ð,+â if ð = ð (= 2).
(2.8)
Lemma 2.1. Let (2.2) and (2.8) hold. Then:
(a) ð¿ð(ð, ð) â ð¿ð(ð; ðâ1) if and only if ð ⥠ð and ðâ1ðâ1 â ð¿ð(ð).
(b) ð¿ð(ð; ðâ1) â ð¿ð(ð; ð) if and only if ð ⥠ð and ðð â ð¿ð(ð).
Proof. The âifâ parts follow easily from Holderâs inequality using (2.8). Therefore,we restrict ourselves to the âonly ifâ parts.
(a): The inclusion means that
â£ððâ£ð â ð¿1(ð) â â£ððâ1â£ð â ð¿1(ð).By the substitution ð = â£ðð⣠this is equivalent to
ðð â ð¿1(ð) â (ððâ1ðâ1)ð â ð¿1(ð).In case ð ⥠ð make the substitution â = ðð to conclude that
â â ð¿ð/ð(ð) â âðâððâð â ð¿1(ð).By the closed graph theorem, the corresponding linear operator is bounded, andthis map gives rise to a bounded linear functional on ð¿ð/ð(ð). It follows thatðâððâð â ð¿ð/ð(ð)â² = ð¿ð/(ðâð)(ð), which implies the above.
In case ð < ð, we make a substitution â = ðð to conclude that
â â ð¿1(ð) â âðâððâð â ð¿ð/ð(ð).
Again by the closed graph theorem, the corresponding linear operator must bebounded. There exists ð > 0 such that ðžð = {ð¡ â ð : ð(ð¡)âðð(ð¡)âð ⥠ð} haspositive measure. Then for each measurable subset ðž â ðžð, when taking thecharacteristic function â = ððž ,
ðð(ðž)ð/ð = â¥ðððžâ¥ð/ð †â¥ððžðâððâðâ¥ð/ð †ð¶â¥ððžâ¥1 = ð¶ð(ðž).
Toeplitz Operators 131
Thus, 0 < ð/𶠆ð(ðž)1âð/ð . Since we can find a sequence of measurable setsðž(ð) â ðžð with ð(ðž
(ð)) > 0 but ð(ðž(ð))â 0, a contradiction follows.
Part (b) can be proved similarly, by replacing ð with ð and ð with ðâ1. â¡
Based on Lemma 2.1 we will now show that a certain homogeneous RiemannâHilbert problem has only a trivial solution. A corresponding result holds also forthe âadjointâ RiemannâHilbert problem.
Proposition 2.2. Let ðº be an ð à ð matrix-valued measurable function on ð
satisfying ᅵᅵâ(ð¡)ðº(ð¡) = ðŒð , and let ð satisfy (2.3). Then:
(a) If ð ⥠ð and ðâ1ðâ1 â ð¿ð(ð), then the equation
ðº(ð¡)ð+(ð¡) = ðâ(ð¡) (2.9)
with ð+ â ð¿ð+(ð; ð)
ð and ðâ â ð¿ðâ,0(ð; ð)
ð has only the trivial solutionð+ = ðâ = 0.
(b) If ð ⥠ð and ðð â ð¿ð(ð), then the equation
ðºâ(ð¡)â+(ð¡) = ââ(ð¡) (2.10)
with â+ â ð¿ð+(ð; ð
â1)ð and ââ â ð¿ðâ,0(ð; ð
â1)ð has only the trivial solutionâ+ = ââ = 0.
Proof. (a): First, notice that the conditions of Lemma 2.1(a) hold. Thus,
ð¿ð(ð, ð) â ð¿ð(ð; ðâ1), (2.11)
which is what we are going to use below. Now assume that (2.9) holds. Passing tothe complex adjoint and applying the tilde operation yields
ðâ+(ð¡)ᅵᅵâ(ð¡) = ðââ(ð¡).
Multiplying the equations together and using ᅵᅵâ(ð¡)ðº(ð¡) = ðŒð we obtain
ðâ+(ð¡)ð+(ð¡) = ðââ(ð¡)ðâ(ð¡),
which is a scalar function since we are multiplying a row with a column vectorfunction. Indeed, using the components of
ð±(ð¡) =(ð±,1(ð¡), . . . , ð±,ð(ð¡)
)ðthe previous equation reads
ðâð=1
ðâ+,ð(ð¡)ð+,ð(ð¡) =
ðâð=1
ðââ,ð(ð¡)ðâ,ð(ð¡). (2.12)
Because of (2.11) and Holderâs inequality, each of the occuring products is in ð¿1(ð).
Furthermore, since ðâ+,ð â ð¿ð+(ð; ð) â ð¿ð
+(ð; ðâ1), again by (2.11), it follows that
each product ðâ+,ðð+,ð and thus the left-hand side of (2.12) is in ð¿1+(ð). For similar
132 T. Ehrhardt and I.M. Spitkovsky
reasons the right-hand belongs to ð¿1â,0(ð). Since ð¿1+(ð) and ð¿
1â,0(ð) have a trivial
intersection, we conclude that
ðâ+(ð¡)ð+(ð¡) =ðâ
ð=1
ðâ+,ð(ð¡)ð+,ð(ð¡) = 0. (2.13)
There are now two possibilities to finish the proof, both instructive in theirown ways. Firstly, one could use the Fourier coefficients [ð+]ð â âð of ð+(ð¡),
ð+(ð¡) =
ââð=0
[ð+]ðð¡ð, â£ð¡â£ = 1,
noting that
ðâ+(ð¡) =ââð=0
[ð+]ððð¡ð, â£ð¡â£ = 1, (2.14)
involves the complex conjugates and the transpose. Thus
0 = (ðâ+ð+)(ð¡) =ââð=0
ð¡ððâ
ð=0
[ð+]ðð[ð+]ðâð,
and it follows that for each ð ⥠0,
ðâð=0
[ð+]ðð[ð+]ðâð = 0.
Consider ð = 0 to obtain [ð+]0ð[ð+]0 = 0 and thus [ð+]0 = 0. Next consider ð = 2
to obtain [ð+]1 = 0, then ð = 4 to get [ð+]2 = 0, and so on. Hence all Fouriercoefficients are zero. This implies ð+(ð¡) = 0 and thus ðâ(ð¡) = 0.
An alternative way of reasoning is to realize that both ð+(ð§) and ðâ+(ð§) =
ð+(ð§)ðadmit analytic continuations into the unit disk { ð§ â â : â£ð§â£ < 1 }. In the
special case of ð§ being real (and â£ð§â£ < 1) we have ðâ+(ð§) = ð+(ð§)ð, and (2.13)
becomes
0 = ð+(ð§)ðð+(ð§), for ð§ real, â£ð§â£ < 1.
Since the values of ð+ are vectors, this is a zero sum of non-negative real numbers.Therefore, ð+(ð§) = 0 for ð§ real, and by analytic continuation we obtain the samefor all â£ð§â£ < 1. As for a function in ð¿1+(ð) there is a one-to-one correpondencebetween its analytic continuation into the unit disk and the boundary values (a.e.)on ð, we obtain that ð+(ð¡) = 0 for â£ð¡â£ = 1 and consequently ðâ(ð¡) = 0.
(b): We remark that the assumptions and Lemma 2.1 imply that
ð¿ð(ð, ðâ1) â ð¿ð(ð; ð). (2.15)
Thus the result can be obtained from part (a) by interchanging ð with ð and ð
with ðâ1. For ðºâ the corresponding condition ᅵᅵ(ð¡)ðºâ(ð¡) = ðŒð holds. â¡
Toeplitz Operators 133
3. Toeplitz operators on the unit circle and real line
Given a Banach space ð , we denote by â(ð) the space of all bounded linearoperators acting on ð . Also, we use the notation
ðŒ(ðŽ) = dimkerðŽ and ðœ(ðŽ) = dimð/ImðŽ (= dimkerðŽâ)
for ðŽ â â(ð). Therein ðŽâ â â(ð â²) stands for the adjoint of ðŽ.Toeplitz operators on ð³ð
+(ð; ð). If we assume, in addition to (2.3), that the weight
ð satisfies the HuntâMuckenhouptâWheeden (or ðŽð) condition,
supðŒ
(1
â£ðŒâ£â«ðŒ
ð(ð¡)ð â£ðð¡â£)1/ð(
1
â£ðŒâ£â«ðŒ
ð(ð¡)âð â£ðð¡â£)1/ð
<â, (3.1)
where the supremum is taken over all subarcs ðŒ of ð, then it is possible to considerToeplitz operators on ð¿ð
+(ð). In fact, the HuntâMuckenhouptâWheeden conditionis necessary and sufficient for the boundedness of Riesz projection
ð :
ââð=ââ
ððð¡ð ï¿œâ
ââð=0
ððð¡ð (3.2)
on the space ð¿ð(ð; ð). Under these conditions the image of ð equals the Hardyspace ð¿ð
+(ð; ð), and the kernel of ð equals ð¿ðâ,0(ð; ð), see [11] or [1] and references
therein.For ðº â ð¿â(ð)ðÃð the block Toeplitz operator acting on ð¿ð
+(ð; ð)ð is de-
fined by
ð (ðº) : ð ï¿œâ ð (ðºð), (3.3)
where ðºð stands for the pointwise product on ð of the matrix-valued function ðºwith the vector-valued function ð . Under the above assumptions ð (ðº) is a boundedlinear operator with norm
â¥ð (ðº)â¥â(ð¿ð+(ð;ð)
ð ) †ð¶ð,ð â¥ðºâ¥ð¿â(ð)ðÃð , (3.4)
where ð¶ð,ð = â¥ðâ¥â(ð¿ð(ð;ð)).
Theorem 3.1. Let ðº â ð¿â(ð)ðÃð , and assume that ᅵᅵâ(ð¡)ðº(ð¡) = ðŒð . Considerð (ðº) on ð¿ð
+(ð; ð)ð , with ð satisfying (3.1).
(a) If ð ⥠ð and ðâ1ðâ1 â ð¿ð(ð), then ðŒ(ð (ðº)) = 0.(b) If ð ⥠ð and ðð â ð¿ð(ð), then ðœ(ð (ðº)) = 0.
Proof. (a): If the Toeplitz operator ð (ðº) has a non-trivial kernel, then there existsa nonzero ð+ â ð¿ð
+(ð; ð) such that
0 = ð (ðº)ð+ = ð (ðºð+),
which is equivalent to
ðâ = ðºð+ â ð¿ðâ,0(ð; ð).
Now Proposition 2.2 implies ð+ = 0, which is a contradiction.
134 T. Ehrhardt and I.M. Spitkovsky
(b): The dual space of ð¿ð+(ð; ð)
ð can be identified with ð¿ð+(ð; ð
â1)ð via
Î : ð¿ð+(ð; ð
â1)ð â (ð¿ð+(ð; ð)
ð )â²,
(Îð)(ð) =
â« 2ð
0
ð(ððð¥)ðð(ððð¥) ðð¥.
Under this identification it is easy to see that the adjoint operator of ð (ðº) (actingon ð¿ð
+(ð; ð)ð ) is the operator ð (ðºâ) acting on ð¿ð
+(ð; ðâ1)ð . Therefore, one can
rely on part (a) and just interchange ð with ð and ð with ðâ1 to obtain the desiredresult. â¡
Factorization. An ð Ãð matrix-valued measurable function ðº defined on ð pos-sesses a factorization in ð¿ð(ð; ð) if it can be written in the form
ðº(ð¡) = ðºâ(ð¡)Î(ð¡)ðº+(ð¡), ð¡ â ð,
such that
ðº+ â ð¿ð+(ð; ð
â1)ðÃð , ðºâ â ð¿ðâ(ð; ð)ðÃð ,
ðºâ1+ â ð¿ð+(ð; ð)
ðÃð , ðºâ1â â ð¿ðâ(ð; ð
â1)ðÃð ,(3.5)
where
Î(ð¡) = diag[ð¡Ï°1 , . . . , ð¡Ï°ð ],
and Ï°1, . . . ,Ï°ð â †are called the partial indices. Notice that if ðº possesses afactorization, then ðº and ðºâ1 belong to ð¿1(ð).
This is a weighted version of the ð¿ð-factorization as defined in [12]. Its exis-tence guarantees that the respective homogeneous RiemannâHilbert problem hasfinitely many linearly independent solutions, while the closure ð of the rangeð ðº = {ðâ+ðºð+} has finite codimension and, moreover, ð ðº is ârationally closedâ,that is, contains all rational vector functions belonging to ð. If ðº â ð¿â(ð)ðÃð ,then the Toeplitz operator ð (ðº) is Fredholm on ð¿ð
+(ð; ð)ð if and only if ðº is
ð¿ð(ð; ð) factorable and in addition the operator ðºâð+ðºâ1â is bounded in the
metric of ð¿ð(ð; ð)ð .In the following theorem we assume (2.2), (2.3) and (2.8); however, (3.1) need
not hold.
Theorem 3.2. Let ðº â ð¿1(ð)ðÃð satisfy ᅵᅵâ(ð¡)ðº(ð¡) = ðŒð , and assume in additionthat ðº possesses a factorization in ð¿ð(ð; ð).
(a) If ð ⥠ð and ðâ1ðâ1 â ð¿ð(ð), then the partial indices Ï°ð ⥠0.(b) If ð ⥠ð and ðð â ð¿ð(ð), then the partial indices Ï°ð †0.(c) If ð = 2 and both ðð and ðâ1ðâ1 belong to ð¿â(ð), then the factorization is
canonical, i.e., Ï°ð = 0 and
ðº(ð¡) = ðºâ(ð¡)ðº+(ð¡). (3.6)
(d) If the factorization is canonical and if the assumptions of (a) or (b) hold,
then one can choose the factors to satisfy ᅵᅵââðºâ = ᅵᅵâ+ðº+ = ðŒð .
Toeplitz Operators 135
Proof. (a): Assume that for some ð, Ï°ð < 0. Then define
ð+(ð¡) = ðºâ1+ (ð¡)ðð
where ðð is the ðth unit vector in âð . Applying ðº(ð¡) = ðºâ(ð¡)Î(ð¡)ðº+(ð¡) we obtain
ðº(ð¡)ð+(ð¡) = ðºâ(ð¡) â ð¡Ï°ððð =: ðâ(ð¡).
It is now easy to see that 0 ââ¡ ð+ â ð¿ð+(ð; ð)
ð and ðâ â ð¿ðâ,0(ð; ð)
ð , whence we
get a contradiction to Proposition 2.2(a).(b): Passing to the complex conjugates in the factorization we see that ðºâ
possesses a factorization in ð¿ð(ð; ðâ1),
ðºâ(ð¡) = ðºâ+(ð¡) diag[ð¡âÏ°1 , . . . , ð¡âÏ°ð ]ðºââ(ð¡),
with partial indices âÏ°1, . . . ,âÏ°ð . Now we can apply the results of part (a),interchanging ð with ð and replacing ð by ðâ1, or argue similar as in (a) and applyProposition 2.2(b).
(c): This follows directly from (a) and (b).
(d): From the factorization ðº = ðºâðº+ we obtain ᅵᅵâ = ᅵᅵâ+ᅵᅵââ and through
inversion and using ðº = (ᅵᅵâ)â1,
ðºâðº+ = ðº = (ᅵᅵâ)â1 = (ᅵᅵââ)â1(ᅵᅵâ+)
â1.
Under the assumptions of (a) we have ð¿ð(ð; ð) â ð¿ð(ð; ðâ1). Consequently,
ᅵᅵââðºâ = (ᅵᅵâ+)â1ðºâ1+ ,
with the left and right-hand sides lying in ð¿1â(ð)ðÃð and ð¿1+(ð)ðÃð , respectively
(compare (3.5)). Thus, each of the products (ᅵᅵâ+)â1ðºâ1+ and ᅵᅵââðºâ is identically
equal to some ð¶ â âðÃð , which must be nonsingular. In particular, using the
analyticity of ðº+(ð§) and ᅵᅵâ+(ð§) = ðº+(ð§)
ðfor â£ð§â£ < 1, we get
ð¶ = (ðº+(0)â)â1ðº+(0)
â1,
which is therefore positive definite. Letting
ð»â(ð¡) = ðºâ(ð¡)ðº+(0), ð»+(ð¡) = ðº+(0)â1ðº+(ð¡),
we see that ðº = ð»âð»+ is also a factorization, while ᅵᅵââð»â = ᅵᅵâ+ð»+ = ðŒð holds.
If the assumptions stated in (b) hold, then we have ð¿ð(ð; ðâ1) â ð¿ð(ð; ð) and weconsider
ðºâ1â (ᅵᅵââ)â1 = ðº+ᅵᅵ
â+,
and proceed analogously. â¡Examples. We now provide some simple examples illustrating Theorems 3.2 and3.1, as well as the essentiality of the nesting conditions in order for these theoremsto hold.
Both examples feature a scalar piecewise continuous function. For such func-tions a Fredholmness criterion and index formula were established by Gohbergand Krupnik, see [10] or more recent [4, 5]. Moreover, it is shown in [12] thatpiecewise continuous functions not satisfying the Fredholmness criterion, do not
136 T. Ehrhardt and I.M. Spitkovsky
admit a factorization in the above sense. With these considerations in mind, thestatements below are easily established.
Example 3.3. We consider 1 < ð < â, ð â¡ 1, and the following function whichhas two jump discontinuities at ð¡ = ±ð. Let ðœ â â and define
ð(ððð¥) =
{ððððœ if â ð/2 < ð¥ < ð/2ðâðððœ if ð/2 < ð¥ < 3ð/2.
This (scalar) function satisfies the condition ðâ(ð¡)ð(ð¡) = 1. The âsizeâ of the jumpdiscontinuities is given by
1
2ðarg
ð(ð â 0)ð(ð + 0)
= ðœ and1
2ðarg
ð(âð â 0)ð(âð + 0)
= âðœ.
The function can be written as
ð(ð¡) = (âð¡/ð)ðœ(ð¡/ð)âðœ,
and its factorization in ð¿ð(ð) can be easily obtained from there. Its specific formwill depend on the relation between ðœ and ð.
Case 1: If â£ðœâ£ < min{ 1ð , 1ð }, then a canonical factorization is given byð(ð¡) =
[(1â ð/ð¡)âðœ(1 + ð/ð¡)ðœ
] â [(1â ð¡/ð)ðœ(1 + ð¡/ð)âðœ].
Moreover, ð (ð) is invertible on ð¿ð(ð).
Case 2: If 1/ð < ðœ < 1/ð, then a factorization with Ï°1 = 1 is given by
ð(ð¡) =[ð(1â ð/ð¡)âðœ+1(1 + ð/ð¡)ðœ
] â ð¡ â [(1â ð¡/ð)ðœâ1(1 + ð¡/ð)âðœ].
The operator ð (ð) is Fredholm on ð¿ð(ð) and has a trivial kernel and a cokernelof dimension one.
Case 3: If 1/ð < ðœ < 1/ð, then a factorization with Ï°1 = â1 is given byð(ð¡) =
[ð(1â ð/ð¡)âðœ(1 + ð/ð¡)ðœâ1
] â ð¡â1 â [(1â ð¡/ð)ðœ(1 + ð¡/ð)âðœ+1].
The operator ð (ð) is Fredholm on ð¿ð(ð) and has kernel dimension one and atrivial cokernel.
These results are in agreement with Theorems 3.1 and 3.2.
Case 4: If ðœ â †+ {1/ð, 1/ð}, then ð possesses no factorization and ð (ð) is notFredholm. However, by Theorem 3.1, ð (ð) has a trivial kernel in case ð ⥠ð and atrivial cokernel in case ð ⥠ð.
These four cases essentially cover all values for ðœ since adding an integer toðœ changes the function ð by at most a sign.
Example 3.4. Let ð = 2, fix two distinct points ð1, ð2 â ð with Im(ðð) > 0, andconsider the weight
ð(ð¡) = â£1â ð1â£ðŒ1 â£1â ð1â£ðŒ1 â£1â ð2â£ðŒ2 â£1â ð2â£ðŒ2
Toeplitz Operators 137
with â£ðŒð⣠< 1/2, which guarantees that ð satisfies (2.3) and in fact the ðŽ2 condi-tion. Introduce the function
ð(ð¡) = (âð¡/ð1)ðœ1(âð¡/ð1)âðœ1(âð¡/ð2)ðœ2(âð¡/ð2)âðœ2
with ðœ1, ðœ2 â â. This function satisfies ðâð = 1. We can factor ð = ðâð+ with
ðâ(ð¡) = (1â ð¡/ð1)âðœ1+1(1â ð¡/ð1)ðœ1(1 â ð¡/ð2)âðœ2(1â ð¡/ð2)ðœ2â1ð1/ð2,
ð+(ð¡) = (1â ð¡/ð1)ðœ1â1(1â ð¡/ð1)âðœ1(1 â ð¡/ð2)ðœ2(1â ð¡/ð2)âðœ2+1.
This is a canonical factorization in ð¿ð(ð; ð) if and only if
â1/2 + ðŒ1 < ðœ1 â 1 < 1/2 + ðŒ1, â1/2 + ðŒ2 < ðœ2 < 1/2 + ðŒ2,
â1/2 + ðŒ1 < âðœ1 < 1/2 + ðŒ1, â1/2 + ðŒ2 < âðœ2 + 1 < 1/2 + ðŒ2.
For instance, we can choose the values ðŒ1 = â1/4, ðŒ2 = 1/4, and ðœð â (1/4, 3/4).It is easy to verify that ðâ±ð± â= 1, and that this also holds if we modify the factorsby some constant. This contrasts the statement of Theorem 3.2(d). Clearly, neitherthe assumptions of (a) nor (b) hold.
Toeplitz operators on ð³ð+(â;ð). For a weight ð on the real line â, let the weighted
space ð¿ð(â;ð) consist of all measurable functions ð defined on â for which
â¥ðâ¥ð,ð :=(â« â
âââ£ð(ð¥)ð(ð¥)â£ð ðð¥
)1/ð<â.
We will assume ð â ð¿ð(â; (1 + â£ð¥â£)â1) and ðâ1 â ð¿ð(â; (1 + â£ð¥â£)â1), and further-more the HuntâMuckenhouptâWheeden (or ðŽð) condition on â,
supðŒ
(1
â£ðŒâ£â«ðŒ
ð(ð¥)ð ðð¥
)1/ð (1
â£ðŒâ£â«ðŒ
ð(ð¥)âð ðð¥
)1/ð
<â
with the supremum taken over all finite intervals. This condition is equivalent tothe boundedness of the projection ð = (ðŒ + ð)/2 on ð¿ð(â;ð), where ð is thesingular integral operator on â. In fact, the image and the kernel of ð equalð¿ð+(â;ð) and ð¿
ðâ(â;ð), respectively, the Hardy spaces of all functions in ð¿
ð(â;ð)which admit an analytic continuation into the upper/lower complex half-plane (fordetails, see, e.g., [2, p. 302]).
For ð» â ð¿â(â)ðÃð the Toeplitz operator on ð¿ð+(â;ð)
ð is defined by
ð (ð»)ð = ð (ð»ð).
We now have the following result. Therein the notation ᅵᅵ(ð¥) = ð(âð¥), ᅵᅵ(ð¥) =ð»(âð¥), and ð»â(ð¥) = ð»(ð¥)ð is used, along with by now standard (2.2), (2.8).
Theorem 3.5. Let ð» â ð¿â(â)ðÃð , and assume that ᅵᅵâ(ð¥)ð»(ð¥) = ðŒð . Considerð (ð») on ð¿ð
+(â;ð)ð .
(a) If ð ⥠ð and ðâ1ᅵᅵâ1 â ð¿ð(â), then ðŒ(ð (ð»)) = 0.(b) If ð ⥠ð and ðᅵᅵ â ð¿ð(â), then ðœ(ð (ð»)) = 0.
138 T. Ehrhardt and I.M. Spitkovsky
Proof. (a): The result can be proved either in analogy to Theorem 3.1, or by notingthat the map ð defined by
(ð ð)(ð¥) =2
ð+ ð¥ð
(ðâ ð¥ð+ ð¥
)is an isometric isomorphism from ð¿ð(ð; ð) onto ð¿ð(â;ð), where
ð(ð¥) =21/ðâ1
(1 + ð¥2)1/ðâ1/2ð
(ðâ ð¥ð+ ð¥
).
Furthermore, ð maps ð¿ð+(ð; ð) and ð¿
ðâ,0(ð; ð) onto ð¿
ð+(â;ð) and ð¿
ðâ(â;ð), re-
spectively. Moreover, ð ð (ðº)ð â1 = ð (ð») with ð»(ð¥) = ðº((ð â ð¥)/(ð + ð¥)). Hencethe statements about defect numbers of ð (ð») can be reduced to the correspondingstatements for ð (ðº) on the space ð¿ð
+(ð; ð). We also remark that ð satisfies the ðŽð
condition on ð if and only if ð satisfies the ðŽð condition on â. Finally, ᅵᅵâð» = ðŒðimplies that ᅵᅵâðº = ðŒð . We now compute that (since 1/ðâ 1/2 = 1/2(1/ðâ 1/ð))
ð(ð¥)â1ᅵᅵ(ð¥)â1 = 2â2/ð(1 + ð¥2)1/ðâ1/ðð( ðâð¥ð+ð¥ )
â1ð( ðâð¥ð+ð¥ )
â1.
In view of (2.8), it follows that ðâ1ᅵᅵâ1 â ð¿ð(â) if and only if ðâ1ðâ1 â ð¿ð(ð).This proves part (a).
(b): This can be proved analogously by passing to the adjoint of ð (ð»), whichcan be identified with ð (ð»â) acting of ð¿ð(â;ðâ1). â¡
4. Toeplitz operators on âð spaces
In this section, we are going to establish analogous statements for Toeplitz opera-tors on weighted âð spaces. Toeplitz operators on such spaces have been analyzed,e.g., in [3], although the focus there was to develop a Fredholm theory for piecewisecontinuous symbols. We also refer to [5] and the references therein.
For 1 < ð < â and a weight function ð : †â â+ define âð(ð) as the spaceof all sequences ð¥ = {ð¥ð}âð=ââ such that
â¥ð¥â¥ð,ð :=( ââ
ð=âââ£ð¥ðð(ð)â£ð
)1/ð
<â.
By ð00 we denote the set of all sequences with finite support, i.e., sequences {ð¥ð}for which only finitely many of the ð¥ðâs are nonzero. The set ð00 is dense in thespaces âð(ð).
Given two sequences, ð¥ = {ð¥ð}âð=ââ and ðŠ = {ðŠð}âð=ââ, one can define theconvolution
ð§ = ð¥ â ðŠ, ð§ð =ââ
ð=ââð¥ðâððŠð
provided that for each ð the series defining ð§ð converges.
Toeplitz Operators 139
Multiplier algebras. Let ðð,ð stand for the set of all sequences ð = {ðð}âð=ââsuch that
(i) ð â ð¥ â âð(ð) for each ð¥ â ð00, and(ii) â¥ðâ¥ðð,ð := sup
ð¥âð00
â¥ð â ð¥â¥ð,ðâ¥ð¥â¥ð,ð <â.
In this case, the mapð¿(ð) : ð¥ ï¿œâ ð â ð¥
extends via continuity to a bounded linear operator acting on âð(ð), called theconvolution operator with symbol ð. Obviously, â¥ð¿(ð)â¥â(âð(ð)) = â¥ðâ¥ðð,ð.
It will be convenient to use the following notation. First, let
ðð : {ð¥ð}âð=ââ ï¿œâ {ð¥ðâð}âð=ââ (4.1)
be the shift operators acting on appropriate spaces of sequences. Denote by ðð thesequence
ðð = {ð¿ð,ð}âð=ââ, (4.2)
where ð¿ð,ð stands for the Kronecker symbol. Note that ððð = ð â ðð. Given aweight ð, define the weight ð by
ð(ð) = ð(âð). (4.3)
We are now able to state the following basic properties concerning convolutionsand convolution operators. Notice in particular that (c) implies thatðð,ð is indeedan algebra, and thus a Banach algebra with appropriate norm. The unit elementin ðð,ð is ð0.
Proposition 4.1. Let (2.2) hold. Then:
(a) We have
ðð,ð =ðð,ðâ1 ââ©ðââ€
ðð
(âð(ð) â© âð(ðâ1)) .
(b) If ð â ðð,ð and ð¥ â âð(ð), then ð¿(ð)ð¥ = ð â ð¥ and the series defining theconvolution converges absolutely.
(c) If ð, ð âðð,ð, then ð â ð âðð,ð and ð¿(ð â ð) = ð¿(ð)ð¿(ð). In particular,
(ð â ð) â ð¥ = ð â (ð â ð¥), ð¥ â âð(ð).(d) Suppose that
supðââ€
(ð(ð+ 1)
ð(ð)+ð(ðâ 1)ð(ð)
)<â. (4.4)
Then(ðŠ â ð) â ð¥ = ðŠ â (ð â ð¥)
whenever ð¥ â âð(ð), ðŠ â âð(ðâ1), and ð âðð,ð =ðð,ðâ1 . In particular, ðŠ âð¥is well defined.
We will need statements (c) and (d) in the proof of the Theorem 4.3 below.Notice that condition (4.4) is equivalent to the boundedness of the shift operatorsðð on â
ð(ð). In other words, it means that ð00 âðð,ð.
140 T. Ehrhardt and I.M. Spitkovsky
Proof. (a): From duality, i.e., the identification of the dual space of âð(ð) withâð(ðâ1) it follows that ð âðð,ð if and only if
supðŠâð00
supð¥âð00
1
â¥ðŠâ¥ð,ðâ1â¥ð¥â¥ð,ð
â£â£â£â£â£â£âð
ðŠâð
âð
ððâðð¥ð
â£â£â£â£â£â£ <â,which involves only finite sums and is thus equal to
supð¥âð00
supðŠâð00
1
â¥ðŠâ¥ð,ðâ1â¥ð¥â¥ð,ð
â£â£â£â£â£â£âð
ð¥âð
âð
ððâððŠð
â£â£â£â£â£â£ <â.The latter is equivalent to ð âðð,ðâ1 . Hence ðð,ð =ðð,ðâ1 .
If ð â ðð,ð, then by definition ð â ðð = ððð belongs to âð(ð). Since also
ð âðð,ðâ1 , it follows that ð â ðð = ððð belongs to âð(ðâ1) as well. This concludes
the proof of (a).(b): Let ð âðð,ð and ð¥ â âð(ð). Since ðâðð â âð(ðâ1), the series
ââð=ââ
â£ððâðð¥ð⣠â€( ââ
ð=âââ£ððâðð
â1(ð)â£ð)1/ð ( ââ
ð=âââ£ð¥ðð(ð)â£ð
)1/ð
is finite. Consequently, the convolution product ðâð¥ is well defined and its definingseries are absolutely convergent. Moreover,
â£[ð â ð¥]ð⣠†â¥ðâððâ¥ð,ðâ1â¥ð¥â¥ð,ð. (4.5)
Since ð¿(ð)ð¥ = ð â ð¥ for ð¥ â ð00, the estimate (4.5) implies that the equality holdsfor all ð¥ â âð(ð).
(c): Notice first that ðâ (ðâð¥) is well defined. Also, ðâ ð is well defined due to(b) since ð âðð,ð and ð âðð,ð â âð(ð). It is easy to see that (ðâð)âð¥ = ðâ(ðâð¥)holds for ð¥ = ðð and thus for all ð¥ â ð00. This is enough to conclude that ð â ð isa multiplier. Indeed,
â¥(ð â ð) â ð¥â¥ð,ð = â¥ð â (ð â ð¥)â¥ð,ð †â¥ðâ¥ðð,ðâ¥ð â ð¥â¥ð,ð †â¥ðâ¥ðð,ðâ¥ðâ¥ðð,ðâ¥ð¥â¥ð,ðwhenever ð¥ â ð00.
It follows that ðâ ð âðð,ð and ð¿(ðâ ð)ð¥ = ð¿(ð)ð¿(ð)ð¥ for ð¥ â ð00. Due to thedensity of ð00 this holds for all ð¥ â âð(ð). Finally, from this and (b) we concludethat (ð â ð) â ð¥ = ð â (ð â ð¥) for all ð¥ â âð(ð).
(d): It is easily seen that
ðð â (ð â ðð) = (ðð â ð) â ðð = ðð+ðð.
Thus ðŠ â (ð â ð¥) = (ðŠ â ð) â ð¥ for ð¥, ðŠ â ð00. Due to assumption (4.4), the shiftoperator ðð is bounded on both â
ð(ð) and âð(ðâ1) for each ð. If we consider the ðthcomponent in the two products under consideration, we can estimate in analogyto (4.5) as follows:
â£[ðŠ â (ð â ð¥)]ð⣠†â¥ðâððŠâ¥ð,ðâ1â¥ðâ¥ðð,ðâ¥ð¥â¥ð,ð
Toeplitz Operators 141
and
â£[(ðŠ â ð) â ð¥]ð⣠†â¥ðŠâ¥ð,ðâ1â¥ðâ¥ðð,ðâ1 â¥ðâðð¥â¥ð,ðfor all ð¥ â âð(ð) and ðŠ â âð(ðâ1). Using these estimates and density, the desiredequality follows. Finally, if we take ð = ð0 â ðð,ð, we obtain that ðŠ â ð¥ is welldefined. â¡
The discrete analogue of Lemma 2.1 is the following. Notice the slight differ-ence in the conditions.
Lemma 4.2. Let (2.2) and (2.8) hold. Then:
(a) âð(ð) â âð(ðâ1) if and only if
(ð ⥠ð and ðâ1ðâ1 â âð) or (ð < ð and ðâ1ðâ1 â ââ). (4.6)
(b) âð(ðâ1) â âð(ð) if and only if
(ð ⥠ð and ðð â âð) or (ð < ð and ðð â ââ). (4.7)
Proof. (a) â âifâ part: The case ð ⥠ð follows from Holderâs inequality. The caseð > ð can be reduced to the case ð = ð = 2 due to the inclusions âð(ð) â â2(ð)and â2(ðâ1) â âð(ðâ1).
(a) â âonly ifâ part: The case ð ⥠ð can be settled as in Lemma 2.1. Assumeð < ð. The assumption implies, after making the same kind of substitutions as inLemma 2.1, that
â = {âð} â â1 =â {âðð(ð)âðð(ð)âð} â âð/ð.By the closed graph theorem, this must be a bounded linear map. Now considerâ = ðð in order to conclude that
ð(ð)âðð(ð)âð = â¥{âðð(ð)âðð(ð)âð}â¥ð/ð †ð¶â¥{âð}â¥1 = ð¶.This implies that ð(ð)â1ð(ð)â1 is bounded and thus a sequence in ââ.
(b): This can be proved by interchanging ð with ð and ð with ðâ1. â¡
Toeplitz operators. Let us first introduce the projection
ð : {ð¥ð}âð=ââ â âð(ð) ï¿œâ {ðŠð}âð=ââ â âð(ð), ðŠð =
{ð¥ð if ð ⥠00 if ð < 0
and the spaces
âð+(ð) ={ð¥ â âð(ð) : ð¥ð = 0 for all ð < 0
}and
âðâ(ð) ={ð¥ â âð(ð) : ð¥ð = 0 for all ð > 0
}.
The image of ð equals âð+(ð), and the kernel of ð equals ðâ1(âðâ(ð)).
For ð âððÃðð,ð we define the Toeplitz operator acting on âð+(ð)
ð by
ð (ð)ð¥ = ð (ð â ð¥),
142 T. Ehrhardt and I.M. Spitkovsky
i.e., ð (ð)ð¥ = ðð¿(ð)ð¥. Given ð¥ = {ð¥ð} we defineð¥ð = {ð¥ðð}, ᅵᅵ = {ð¥ð}.
The unit element in the matrix version of the multiplier algebra is ð0 â ðŒð .Theorem 4.3. Let ð âððÃð
ð,ð , let condition (4.4) on ð hold, and assume that
ᅵᅵð â ð = ð0 â ðŒð .Consider ð (ð) on âð+(ð)
ð .
(a) If condition (4.6) holds, then ðŒ(ð (ð)) = 0.(b) If condition (4.7) holds, then ðœ(ð (ð)) = 0.
Remark. Due to the assumption ð â ððÃðð,ð , the convolution product ᅵᅵð â ð is
always well defined. The relation ᅵᅵð â ð = ð0 â ðŒð can be rewritten asââ
ð=ââ(ᅵᅵð)
ð ððâð = ð¿0,ððŒð for each ð â â€.
Hence, at least formally, this corresponds to(âðᅵᅵðð ð¡
ð)(â
ðððð¡
ð)= ðŒð , ð¡ â ð,
and thus to the condition ᅵᅵâ(ð¡)ðº(ð¡) = ðŒð of the previous section.
Proof. Assume that the kernel of ð (ð) is non-trivial. Then there exist ð¥ â âð+(ð)ð ,ð¥ â= 0, and ðŠ â ðâ1(âðâ(ð))ð such that
ð â ð¥ = ðŠ.Since ð âððÃð
ð,ð we obviously have ᅵᅵð âððÃðð,ð as well. Multiplying with ᅵᅵð we
obtain
ᅵᅵð â ðŠ = ᅵᅵð â (ð â ð¥) = (ᅵᅵð â ð) â ð¥ = ð¥due to Proposition 4.1 (c) noting that ð¥ â âð(ð)ð and ᅵᅵð â ð = ð0 â ðŒð . Nowmultiply with ᅵᅵð to obtain
ᅵᅵð â (ᅵᅵð â ðŠ) = ᅵᅵð â ð¥.By Proposition 4.1 (d), this convolution product is well defined and we have
ᅵᅵð â (ᅵᅵð â ðŠ) = (ᅵᅵð â ᅵᅵð ) â ðŠsince ᅵᅵð â ððÃð
ð,ð = ððÃðð,ðâ1 and ðŠ, ᅵᅵ â âð(ð)ð â âð(ðâ1)ð , making now use of
the assumption (4.6). From ᅵᅵð â ᅵᅵð = ðŠð , which follows from ð â ð¥ = ðŠ, we obtainð§ := ᅵᅵð â ð¥ = ðŠð â ðŠ.
The sequence ð§ = 0 because ð§ð = 0 for ð ⥠0 due to the right-hand side andð§ð = 0 for ð < 0 due to the left-hand side. In particular, for ð ⥠0 we obtain
ðâð=0
ð¥ððð¥ðâð = 0.
Toeplitz Operators 143
We recursively consider ð = 0, 2, . . . to conclude ð¥ð = 0 for all ð. Thus ð¥ = 0 andðŠ = 0.
The proof of part (b) is analogous, by interchanging ð with ð and ð withðâ1. Notice that the adjoint of ð (ð) on âð+(ð)ð can be identified with the operator
ð (ð) acting on âð+(ðâ1)ð with ðð = ðâð
ð . Clearly, ᅵᅵð â ð = ð0 â ðŒð holds as well.
Furthermore, ð âððÃðð,ðâ1 =ððÃð
ð,ð . â¡
References
[1] A. Bottcher and Yu.I. Karlovich, Carleson curves, Muckenhoupt weights, andToeplitz operators, Progress in Math., vol. 154, Birkhauser Verlag, Basel and Boston,1997.
[2] A. Bottcher, Yu.I. Karlovich, and I.M. Spitkovsky, Convolution operators and fac-torization of almost periodic matrix functions, Operator Theory: Advances and Ap-plications, vol. 131, Birkhauser Verlag, Basel and Boston, 2002.
[3] A. Bottcher and M. Seybold, Discrete WienerâHopf operators on spaces with Muck-enhoupt weight, Studia Math. 143 (2000), no. 2, 121â144.
[4] A. Bottcher and B. Silbermann, Introduction to large truncated Toeplitz matrices,Springer-Verlag, New York, 1999.
[5] , Analysis of Toeplitz operators, second ed., Springer Monographs in Mathe-matics, Springer-Verlag, Berlin, 2006, prepared jointly with A. Karlovich.
[6] M.C. Camara, L. Rodman, and I.M. Spitkovsky, One sided invertibility of matri-ces over commutative rings, corona problems, and Toeplitz operators with matrixsymbols, submitted.
[7] K.F. Clancey and I. Gohberg, Factorization of matrix functions and singular integraloperators, Operator Theory: Advances and Applications, vol. 3, Birkhauser, Baseland Boston, 1981.
[8] I. Gohberg, M.A. Kaashoek, and I.M. Spitkovsky, An overview of matrix factoriza-tion theory and operator applications, in: Factorization and integrable systems (Faro,2000), Operator Theory: Advances and Applications, vol. 141, Birkhauser Verlag,Basel and Boston, 2003, pp. 1â102.
[9] I. Gohberg and M.G. Krein, Systems of integral equations on a half-line with kerneldepending upon the difference of the arguments, Uspekhi Mat. Nauk 13 (1958), no. 2,3â72 (in Russian), English translation: Amer. Math. Soc. Transl. 14 (1960), no. 2,217â287.
[10] I. Gohberg and N. Krupnik, One-dimensional linear singular integral equations. I,Operator Theory: Advances and Applications, vol. 53, Birkhauser Verlag, Basel,1992, Introduction, translated from the 1979 German translation by B. Luderer andS. Roch and revised by the authors.
[11] R. Hunt, B. Muckenhoupt, and R. Wheeden, Weighted norm inequalities for theconjugate function and Hilbert transform, Trans. Amer. Math. Soc. 176 (1973), 227â251.
144 T. Ehrhardt and I.M. Spitkovsky
[12] G.S. Litvinchuk and I.M. Spitkovskii, Factorization of measurable matrix functions,Operator Theory: Advances and Applications, vol. 25, Birkhauser Verlag, Basel,1987, translated from the Russian by B. Luderer, with a foreword by B. Silbermann.
[13] A.F. Voronin, On the well-posedness of the Riemann boundary value problem with amatrix coefficient, Dokl. Akad. Nauk 414 (2007), no. 2, 156â158 (in Russian), Englishtranslation: Dokl. Math. 75 (2007), no. 3, 358â360.
[14] , Partial indices of unitary and Hermitian matrix functions, Sibirsk. Mat.Zh. 51 (2010), no. 5, 1010â1016 (in Russian), English translation: Sib. Math. J. 51(2010), no. 5, 805â809.
Torsten EhrhardtMathematics DepartmentUniversity of CaliforniaSanta Cruz, CA-95064, USAe-mail: [email protected]
Ilya M. SpitkovskyMathematics DepartmentThe College of William and MaryWilliamsburg, VA 23187-8795, USAe-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 145â160câ 2013 Springer Basel
Rational Matrix Solutions of aBezout Type Equation on the Half-plane
A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
Dedicated to Leonia Lerer on the occasion of his 70th birthday, in friendship
Abstract. A state space description is given of all stable rational matrix so-lutions of a general rational Bezout type equation on the right half-plane.Included are a state space formula for a particular solution satisfying a cer-tain ð»2 minimality condition, a state space formula for the inner functiondescribing the null space of the multiplication operator corresponding to theBezout equation, and a parameterization of all solutions using the particularsolution and this inner function. A state space version of the related Tolokon-nikov lemma is also presented.
Mathematics Subject Classification (2010). Primary 47B35, 39B42; Secondary47A68, 93B28.
Keywords. Bezout equation; stable rational matrix functions; state space rep-resentation; algebraic Riccati equation; stabilizing solution, right invertiblemultiplication operator; WienerâHopf operators.
1. Introduction
In this paper ðº is a stable rational ðà ð matrix function. Here stable means thatðº is proper, that is, the limit of ðº(ð ) as ð â â exists, and ðº has all its poles inthe open left half-plane {ð â â ⣠â(ð ) < 0}. In other words, ðº is a rational matrix-valued ð»â function, where the latter means that ðº is analytic and bounded onthe open right half-plane. In this paper ð will be larger than ð, and thus ðº willbe a âfatâ non-square matrix function. We shall be interested in stable rationalðÃð matrix-valued solutions ð of the Bezout type equation
ðº(ð )ð(ð ) = ðŒð, âð ⥠0. (1.1)
The symbol ðŒð on the right-hand side denotes the ðÃð identity matrix.
146 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
Throughout we shall assume that ðº admits a state space realization of theform
ðº(ð ) = ð¶(ð ðŒð âðŽ)â1ðµ +ð·. (1.2)
Here ðŽ is an ðÃð matrix which is assumed to be stable, that is, all the eigenvaluesof ðŽ are contained in the open left half-plane. Moreover, ðµ, ð¶ and ð· are matricesof appropriate sizes. Our aim is to give necessary and sufficient conditions for thesolvability of (1.1), and to give a full description of all stable rational matrix-valuedsolutions, in terms of the matrices appearing in the realization (1.2). The resultswe present are the continuous analogs of the main theorems in [6] and [8].
To state the main results we need some additional notation. By ð we denotethe controllability Gramian associated with the realization (1.2), that is, ð is the(unique) solution of the Lyapunov equation
ðŽð + ððŽâ +ðµðµâ = 0. (1.3)
Consider the algebraic Riccati equation
ðŽâð+ððŽ+ (ð¶ â Îâð)â(ð·ð·â)â1(ð¶ â Îâð) = 0, (1.4)
where Î is defined by
Î = ðµð·â + ðð¶â. (1.5)
Here it is assumed that ð· is right invertible, which is a natural condition. Indeed,if (1.1) has a stable rational matrix solution ð , then using (1.2) and the fact thatð is proper, we see that ð·ð(â) = limð ââðº(ð )ð(ð ) = ðŒð. Hence ð(â) is aright inverse of ð·, and thus ð· is right invertible. A solution ð of (1.4) is calledthe stabilizing solution of the algebraic Riccati equation (1.4) if ð is Hermitianand the ðà ð matrix ðŽ0 given by
ðŽ0 = ðŽâ Î(ð·ð·â)â1(ð¶ â Îâð) (1.6)
is stable. If it exists, a stabilizing solution is unique (cf. formula (2.11)). Thefollowing is the first main result of this paper.
Theorem 1.1. There is a stable rational ð à ð matrix function ð satisfying theequation ðº(ð )ð(ð ) = ðŒð if and only if the following three conditions hold
1. The matrix ð· is right invertible,2. there exists a stabilizing solution ð of the Riccati equation (1.4), and3. the matrix ðŒð â ðð is invertible.
In that case a particular solution of (1.1) is given by
Î(ð ) =(ðŒð â ð¶1(ð ðŒð âðŽ0)
â1(ðŒð â ðð)â1ðµ)ð·â(ð·ð·â)â1, (1.7)
where ðŽ0 is the stable ðà ð matrix given by (1.6) and
ð¶1 = ð·â(ð·ð·â)â1(ð¶ â Îâð) +ðµâð. (1.8)
Matrix Solutions of a Bezout Type Equation on the Half-plane 147
The matrix ð·â(ð·ð·â)â1 appearing in (1.7) is MooreâPenrose right inverseof ð·. In what follows we shall often denote ð·â(ð·ð·â)â1 by ð·+. Note thatdimKerð· = ðâð.
The rational ðà ð matrix function appearing in the right-hand side of (1.7)between the brackets will be denoted by ð , that is,
ð (ð ) = ðŒð â ð¶1(ð ðŒð âðŽ0)â1(ðŒð â ðð)â1ðµ. (1.9)
Note that the value of ð at infinity is invertible. Hence ð (ð )â1 is a well-definedrational matrix function. We shall see that ð (ð )â1 is again stable. Thus both ð (ð )and ð (ð )â1 are stable rational matrix functions. In other words the entries of bothð (ð ) and ð (ð )â1 are ð»â functions. In this case we say that ð is invertible outer.
Among other things the following theorem describes the set of all stablerational solutions to ðº(ð )ð(ð ) = ðŒð.
Theorem 1.2. There exists a stable rational ð à ð matrix function ð satisfyingðº(ð )ð(ð ) = ðŒð if and only if ð· is right invertible and there exists a stable ra-tional ðà ð matrix function ð which is invertible outer and satisfies the equationðº(ð )ð (ð ) = ð·. In this case one such ð is given by (1.9) and the inverse of thisð is given by
ð (ð )â1 = ðŒð + ð¶1(ðŒð â ðð)â1(ð ðŒð âðŽ)â1ðµ. (1.10)
Moreover, using this function ð the following holds.
(i) Let ðž be any isometry mapping âðâð into âð such that Imðž = Kerð·. Thenthe function
Î(ð ) = ð (ð )ðž =(ðŒð â ð¶1(ð ðŒð âðŽ0)
â1(ðŒð â ðð)â1ðµ)ðž (1.11)
is a stable rational ðà (ðâð) matrix function satisfying ðº(ð )Î(ð ) = 0, andÎ is inner, that is, Î(âð )âÎ(ð ) = ðŒðâð.
(ii) If â is any âð-valued ð»2 function satisfying ðº(ð )â(ð ) = 0, then there existsa unique â(ðâð)-valued ð»2 function ð such that â(ð ) = Î(ð )ð(ð ). In fact,ð(ð ) = Î(âð )ââ(ð ).
(iii) The set of all stable rational ðÃð matrix functions ð satisfying ðº(ð )ð(ð ) =ðŒð is given by
ð(ð ) =(ðŒð â ð¶1(ð ðŒð âðŽ0)
â1(ðŒð â ðð)â1ðµ)Ã
Ã(ð·â(ð·ð·â)â1 + ðžð(ð )
), (1.12)
where ð is an arbitrary stable rational (ð âð) Ãð matrix function. More-over, if ð satisfies ðº(ð )ð(ð ) = ðŒð, then ð in (1.12) is given by ð(ð ) =ðžâð (ð )â1ð(ð ).
(iv) The rational ðà ð matrix function
ðºðð¥ð¡(ð ) =
[ðº(ð )
ðžâð (ð )â1
], âð ⥠0, (1.13)
148 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
is invertible outer and its inverse is given by
ðºðð¥ð¡(ð )â1 =
[Î(ð ) Î(ð )
], âð ⥠0. (1.14)
Note that item (ii) tells us that the null space Kerððº of the multiplicationoperator ððº defined by ðº, mapping ð»2
ð into ð»2ð, is given by Kerððº = Îð»2
ðâð.Thus Î plays the role of the inner function in the BeurlingâLax theorem specifiedfor Kerððº. Furthermore, (1.12) in item (iii) can be rewritten in the followingequivalent form ð(ð ) = Î(ð ) + Î(ð )ð(ð ). Using this form of (1.12) we expect ourstate space formulas also to be useful in deriving rational ð»â solutions of (1.1)that satisfy an additional ð»â norm constraint, by reducing the norm constraintproblem to a generalized Sarason problem (cf. Section I.7 in [7]). Finally, item (iv)is inspired by Tolokonnikovâs lemma (see [18] and [16, Appendix 3]).
The formulas in Theorems 1.1 and 1.2 can be easily converted into a Matlabprogram to compute Î in (1.7), the function ð in (1.9), and Î in (1.11).
We see Theorems 1.1 and 1.2 as the closed right half-plane analogues ofTheorem 1.1 in [6] and Theorem 1.1 in [8], which deal with equation (1.1) in thesetting of rational matrix functions analytic in the closed unit disc. Obviously,a way to obtain the set of all stable rational matrix solutions to equation (1.1)is to use the Cayley transform to derive the right half-plane solutions from theiranalogues in the disc case as given in [8]. However, note that in the present half-plane case there is an additional difficulty: The constant functions are not in ð¿2,whereas in the disc case they are in ð¿2. Furthermore, the particular solution Î inTheorem 1.1 is not the analogue of the least squares solution in [6]. On the otherhand, as we shall show in Section 4, the function Î has an interpretation in termsof solutions to a somewhat different minimization problem (see Theorem 4.1).
We take this occasion to mention that Theorem 1.1 in [6] and Theorem 1.1 in [8]have predecessors in the papers [14] and [13]. In particular, see Lemma 4.1 andTheorem 4.2 in [13]. We are grateful to Dr. Sander Wahls for mentioning to usthese and several related other references. It is interesting to see the role the Bezoutequation plays in solving the engineering problems considered in [14] and [13]. Theproofs in [6] and [8] are quite different from those in [14] and [13]; also differentRiccati equations are used and different state space formulas are obtained.
There is an extensive literature on the Bezout equation and the related coronaequation, see, e.g., the classical papers [4], [9], [18], the books [16], [15], [1], themore recent papers [19], [20], [21], [22], and the references therein. Also, findingrational matrix solutions in state space form for Bezout equations is a classicaltopic in mathematical system theory; see, e.g., the book [23], and the papers[11], [10]. However, as far as we know the formulas we present here are new andcannot easily be obtained using the methods presented in the classical sources. Theinterpretation of the special solution (1.7) as a limit of solutions of minimizationproblems also seems to be new. Moreover, the approach we follow in the presentpaper and the earlier papers [6, 8] can be extended to a Wiener space setting. Infact, in a Wiener space setting the function ð given by (1.9) appears in a very
Matrix Solutions of a Bezout Type Equation on the Half-plane 149
natural way; see also the comment at the end of Section 2. We plan to return tothis in a future paper, also for the discrete case.
The paper consists of four sections, including this introduction. In the secondsection we present the preliminaries from operator theory used in the proofs, andwe explain the role of the Riccati equation (1.4), and prove the necessity of con-ditions 1, 2, 3 in Theorem 1.1. The third section contains the proofs of Theorem1.1 and 1.2. In the final section we consider an optimization problem, which helpsin identifying Î in as a solution with a special minimality property.
2. Operator theory and Riccati equation
In this section we prove the necessity of conditions 1, 2, 3 in Theorem 1.1. Our proofrequires some preliminaries from operator theory and uses the Riccati equation(1.4).
Let Ω be any proper rational ð à ð matrix function with no pole on theimaginary axis ðâ. With Ω we associate the WienerâHopf operator ðΩ and theHankel operator ð»Î©, both mapping ð¿
2ð(â
+) into ð¿2ð(â+). These operators are the
integral operators defined by
(ðΩð)(ð¡) = Ω(â)ð(ð¡) +â« ð¡
0
ð(ð¡â ð)ð(ð)ðð, ð¡ ⥠0, ð â ð¿2ð(â+), (2.1)
(ð»Î©ð)(ð¡) =
â« â
0
ð(ð¡+ ð)ð(ð)ðð, ð¡ ⥠0 ð â ð¿2ð(â+). (2.2)
Here ð is the Lebesque integrable (continuous) matrix function on the imaginaryaxis determined by Ω via the Fourier transform:
Ω(ð ) = Ω(â) +â« â
ââðâð ðð(ð) ðð, ð â ðâ.
In the sequel we shall freely use the basic theory of WienerâHopf and Hankeloperators which can be found in Chapters XII and XIII of [12]. Note that in [12]the Fourier transform is taken with respect to the real line instead of the imaginaryaxis as is done here.
Now let ðº be the stable rational ðÃð function given by (1.2). Then
ðº(ð ) = ð· +
â« â
0
ðâð ðð¶ðððŽðµ ðð, ð â ðâ.
Hence the WienerâHopf operator ððº and the Hankel operator ð»ðº are given by
(ððºð)(ð¡) = ð·ð(ð¡) +
â« ð¡
0
ð¶ð(ð¡âð)ðŽðµð(ð)ðð, ð¡ ⥠0, (2.3)
(ð»ðºð)(ð¡) =
â« â
0
ð¶ð(ð¡+ð)ðŽðµð(ð)ðð, ð¡ ⥠0. (2.4)
With ðº we also associate the rational ðÃð matrix function ð given by ð (ð ) =ðº(ð )ðº(âð )â. Note that ð is a proper rational ðÃð matrix function with no pole
150 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
on the imaginary axis. By ðð we denote the corresponding WienerâHopf operatoracting on ð¿2ð(â
+). It is well known (see, e.g., formula (24) in Section XII.2 of [12])that
ðð = ððºðâðº +ð»ðºð»
âðº. (2.5)
Next assume that the equation ðº(ð )ð(ð ) = ðŒð has a stable rational matrixsolution ð . The fact that ð is stable implies that ð is proper and has no poleson the imaginary axis, and thus ðð is well defined. Furthermore, ððºðð = ððºð ;see [12, Proposition XIII.1.2]. Since ðºð is identically equal to the ðÃð identitymatrix ððºð is the identity operator on ð¿2ð(â
+), and hence ðð is a right inverseof ððº. The fact ððº that is right invertible, implies that ððºð
âðº is invertible and
hence strictly positive. The identity (2.5) then shows that ðð is strictly positivetoo, and hence is invertible.
In the following proposition we use the algebraic Riccati equation (1.4) toobtain necessary and sufficient conditions for ðð to be invertible in terms of thematrices ðŽ, ðµ, and ð¶ appearing in the realization (1.2). As in Section 1, we denoteby ð the controllability Gramian associated with the realization (1.2), that is, ð isthe solution of the Lyapunov equation (1.3). Finally, Î is the ðÃð matrix definedby (1.5).
Proposition 2.1. Let ð (ð ) = ðº(ð )ðº(âð )â. Then the operator ðð is invertible ifand only if the algebraic Riccati equation
ðŽâð+ððŽ+ (ð¶ â Îâð)â (ð·ð·â)â1 (ð¶ â Îâð) = 0 (2.6)
has a stabilizing solution ð, that is, ð is a Hermitian solution of (2.6) and theoperator ðŽ0, defined by
ðŽ0 = ðŽâ Îð¶0, where ð¶0 = (ð·ð·â)â1 (ð¶ â Îâð) , (2.7)
is stable.
Proof. The proposition is an immediate consequence of Theorem 14.8 in [3]. Tosee this, we first show that
ð (ð ) = ð·ð·â + ð¶(ð ðŒð âðŽ)â1Îâ Îâ(ð ðŒð +ðŽâ)â1ð¶â. (2.8)
This partial fraction expansion for ð follows from the Lyapunov equation (1.3),and its immediate consequence
â(ð ðŒð âðŽ)â1ðµðµâ(ð ðŒð +ðŽâ)â1 = (ð ðŒð âðŽ)â1ð â ð (ð ðŒð +ðŽâ)â1.By employing ðº(ð ) = ð· + ð¶(ð ðŒð âðŽ)â1ðµ the identity (2.8) then follows from
ð (ð ) = ðº(ð )ðº(âð )â = (ð· + ð¶(ð ðŒð âðŽ)â1ðµ)(ð·â âðµâ(ð ðŒð +ðŽâ)ð¶â)= ð·ð·â + ð¶(ð ðŒð âðŽ)â1ðµð·â âð·ðµâ(ð ðŒð +ðŽâ)ð¶â+ ð¶(ð ðŒð âðŽ)â1ðð¶â â ð¶ð (ð ðŒð + ðŽâ)â1ð¶â.
Using Î = ðµð·â + ðð¶â, this yields (2.8). Given (2.8) we can apply Theorem 14.8in [3], replacing ðœ by ð·ð·â and ðµ by Î, and rewriting the corresponding algebraicRiccati equation in the form (2.6). â¡
Matrix Solutions of a Bezout Type Equation on the Half-plane 151
From the partial fraction expansion (2.8) it follows that the action of theWienerâHopf operator ðð on ð¿2ð(â
+) is given by
(ðð ð)(ð¡) = ð·ð·âð(ð¡) +
â« ð¡
0
ð¶ð(ð¡âð)ðŽÎð(ð) ðð
+
â« â
ð¡
Îâðâ(ð¡âð)ðŽâð¶âð(ð) ðð. ð¡ ⥠0.
(2.9)
Byðobs andð0, obs we denote the observability operators mapping the state spaceâð into ð¿2ð(â
+) defined by
(ðobsð¥)(ð¡) = ð¶ðð¡ðŽð¥ and (ð0, obsð¥)(ð¡) = ð¶0ð
ð¡ðŽ0ð¥, where ð¥ â âð. (2.10)
Proposition 2.2. Assume that ðð is invertible, or equivalently, there exists a sta-bilizing solution ð to the algebraic Riccati equation (2.6). Then this stabilizingsolution is uniquely determined by
ð =ð âobsð
â1ð ðobs. (2.11)
Proof. To establish this, let us first show that ð satisfies the following Lyapunovequation
ðŽâð+ððŽ0 + ð¶âð¶0 = 0. (2.12)
Recall that ðŽ0 = ðŽâ Îð¶0. Then (2.12) follows from the Riccati equation
0 = ðŽâð+ððŽ+ (ð¶ â Îâð)â (ð·ð·â)â1 (ð¶ â Îâð)= ðŽâð+ð (ðŽ0 + Îð¶0) + (ð¶ â Îâð)â ð¶0
= ðŽâð+ððŽ0 + ð¶âð¶0.
Thus (2.12) holds. Because ðŽ and ðŽ0 are both stable, the stabilizing solution ðcan also be written as
ð =
â« â
0
ðð¡ðŽâð¶âð¶0ð
ð¡ðŽ0ðð¡ =ð âobsð0, obs. (2.13)
Next we prove that
ðâ1ð ðobs =ð0, obs. (2.14)
This essentially follows from [2], Corollary 6.3. For completeness we provide aproof. It suffices to compute ðð ð0, obs. To do this, we use (2.9). Fix ð¥ â âð. Fromthe second identity in (2.10) and the first identiy in (2.7) it follows thatâ« ð¡
0
ð¶ð(ð¡âð)ðŽÎ(ð0, obsð¥)(ð) ðð =
â« ð¡
0
ð¶ð(ð¡âð)ðŽÎð¶0ðððŽ0ð¥ ðð
=
â« ð¡
0
ð¶ð(ð¡âð)ðŽ(ðŽâðŽ0)ðððŽ0ð¥ ðð = ð¶ðð¡ðŽ
( â« ð¡
0
ð¶ðâððŽ(ðŽâðŽ0)ðððŽ0 ðð
)ð¥
= âð¶ðð¡ðŽ( â« ð¡
0
ð
ðð(ðâððŽðððŽ0) ðð
)ð¥ = âð¶ðð¡ðŽ0ð¥+ ð¶ðð¡ðŽð¥.
152 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
Furthermore, using the Lyapunov identity (2.12) we obtainâ« â
ð¡
Îâðâ(ð¡âð)ðŽâð¶â(ð0, obsð¥)(ð) ðð =
â« â
ð¡
Îâðâ(ð¡âð)ðŽâð¶âð¶0ð
ððŽ0ð¥ ðð
= ââ« â
ð¡
Îâðâ(ð¡âð)ðŽâ(ðŽâð+ððŽ0)ð
ððŽ0ð¥ ðð
= âÎâðâð¡ðŽâ( â« â
ð¡
ðððŽâ(ðŽâð +ððŽ0)ð
ððŽ0 ðð)ð¥
= âÎâðâð¡ðŽâ( â« â
ð¡
ð
ðð(ðððŽ
âððððŽ0) ðð
)ð¥ = âÎâðâð¡ðŽâ(â ðð¡ðŽâ
ððð¡ðŽ0
)ð¥
= Îâððð¡ðŽ0ð¥.
Using (2.9) and the second identity in (2.7) we conclude that
(ðð ð0, obs)(ð¡) = ð·ð·âð¶0ð
ð¡ðŽ0 + (âð¶ðð¡ðŽ0 + ð¶ðð¡ðŽ) + Îâððð¡ðŽ0
= (ð·ð·âð¶0 + Îâð)ðð¡ðŽ0 â ð¶ðð¡ðŽ0 + ð¶ðð¡ðŽ = ð¶ðð¡ðŽ.
This proves ðð ð0, obs =ðobs, and hence (2.14) holds. Together (2.13) and (2.14)
show thatð âobsð
â1ð ðobs =ð
âobsð0, obs = ð. In particular, the stabilizing solution
is uniquely determined by (2.11). â¡
Lemma 2.3. Assume ðð is invertible. Then ðŒâð»âðºðâ1ð ð»ðº is positive. Furthermore,the following are equivalent:
(i) ððº is right invertible,(ii) ðŒ âð»âðºðâ1ð ð»ðº is strictly positive,
(iii) ðŒ âð»âðºðâ1ð ð»ðº is invertible.
Proof. Rewriting (2.5) as ððºðâðº = ðð âð»ðºð»
âðº, and multiplying the latter identity
from the left and from the right by ðâ1/2ð shows that
ðâ1/2ð ððºð
âðºð
â1/2ð = ðŒ â ðâ1/2ð ð»ðºð»
âðºð
â1/2ð . (2.15)
Hence ðŒ â ðâ1/2ð ð»ðºð»âðºð
â1/2ð is positive which shows that ð»âðºð
â1/2ð is a contrac-
tion. But then ð»âðºðâ1ð ð»ðº =
(ð»âðºð
â1/2ð
)(ð»âðºð
â1/2ð
)âis also a contraction, and
thus the operator ðŒ âð»âðºðâ1ð ð»ðº is positive.
Since ðŒ â ð»âðºðâ1ð ð»ðº is positive, the equivalence of items (ii) and (iii) is
trivial. Assume (ii) holds. Then ðâ1/2ð ð»ðº is a strict contraction, and hence the
same holds true for ðâ1/2ð ð»ðºð»
âðºð
â1/2ð . But then ðŒâðâ1/2ð ð»ðºð»
âðºð
â1/2ð is strictly
positive, and (2.15) shows that ððº is right invertible. The converse implication isproved in a similar way. â¡
Corollary 2.4. Assume that ðð is invertible, or equivalently, there exists a stabi-lizing solution ð to the algebraic Riccati equation (2.6). Then the spectral radiusof ðð is at most one.
Matrix Solutions of a Bezout Type Equation on the Half-plane 153
Furthermore, the following are equivalent:
(i) ððº is right invertible,(ii) ðspec(ðð ) < 1,(iii) ðŒð âðð is invertible.
Proof. Let ðcon be the controllability operator mapping ð¿ð2(â
+) into âð de-fined by
ðconâ =
â« â
0
ðð¡ðŽðµâ(ð¡)ðð¡, â â ð¿2ð(â+).
Then ð =ðconðâcon and ð»ðº =ðobsðcon. Using these two identities and (2.11),
we obtain for the spectral radius of ð»âðºðâ1ð ð»ðº that
ðspec(ð»âðºð
â1ð ð»ðº) = ðspec(ð
âconð
âobsð
â1ð ðobsðcon)
= ðspec(ðâconððcon)
= ðspec(ððconðâcon) = ðspec(ðð ).
(2.16)
By Lemma 2.3 the operator ðŒâð»âðºðâ1ð ð»ðº is positive. Hence the spectral radius of
ð»âðºðâ1ð ð»ðº is at most one, and the preceding calculation shows that ðspec(ðð ) †1
Since ðspec(ðð ) †1, the equivalence of items (ii) and (iii) is trivial. Assume
ðspec(ðð ) < 1. Then (2.16) shows that ðŒâð»âðºðâ1ð ð»ðº is invertible, and Lemma 2.3tells us that ððº is right invertible. To prove the converse implication, assume thatððº is right invertible. Then, by Lemma 2.3, the operator ðŒâð»âðºðâ1ð ð»ðº is strictly
positive. Hence ðspec(ðŒâð»âðºðâ1ð ð»ðº)<1, and (2.16) shows that ðspec(ðð )<1. â¡
Necessity of the conditions 1, 2, 3 in Theorem 1.1. Assume that the equationðº(ð )ð(ð ) = ðŒð has a stable rational matrix solution ð . As was shown in theparagraph preceding Theorem 1.1, this implies that ð· is right invertible. Thuscondition 1 is necessary. Furthermore, in the paragraph directly after (2.5) it wasshown that ðº(ð )ð(ð ) = ðŒð has a stable rational matrix solution also implies thatððº is right invertible and ðð is invertible. Given the latter we can apply Proposition2.1 to show that condition 2 is necessary. Finally, using Corollary 2.4, we see thatððº is right invertible and ðð is invertible imply that ðŒð â ðð is invertible, whichshows that condition 3 is necessary. â¡
Comment. The identities appearing in this section can also be used to give analternative formula for the function ð in (1.9), namely
ð (ð ) = ðŒð ââ« â
0
ðâð ð¡ðŠ(ð¡)ðð¡, âð ⥠0, where ðŠ = ð âðº(ððºðâðº)â1ð. (2.17)
This formula also makes sense in a Wiener space setting. From formula (2.17) forð it follows that ððºðŠ = ð, which immediately implies that ðº(ð )ð (ð ) = ð·. Thelatter identity will be derived in the next section (see the second paragraph of theproof of Theorem 1.1) using state space computations. We plan to prove (2.17) ina future paper.
154 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
3. Proof of the two main theorems
It will be convenient first to prove the two identities given in the following lemma.
Lemma 3.1. Assume conditions 1, 2, 3 in Theorem 1.1 are satisfied. Then
ðµð¶1 = ðŽ(ðŒð â ðð)â (ðŒð â ðð)ðŽ0, (3.1)
ð·ð¶1 = ð¶(ðŒð â ðð). (3.2)
Proof. Recall that ð¶1 and ð¶0 are respectively defined in (1.8) and (2.7). Thisimplies that
ð¶1 = ð·âð¶0 +ðµ
âð. (3.3)
To prove the first identity, we use the Lyapunov equation (2.12) with Î defined in(1.5) to compute
ðµð¶1 = ðµð·âð¶0 +ðµðµ
âð = (Îâ ðð¶â)ð¶0 â (ðŽð + ððŽâ)ð
= Îð¶0 + ððŽâð+ ðððŽ0 âðŽððâ ððŽâð
= Îð¶0 + ðððŽ0 âðŽðð = ðŽâðŽ0 + ðððŽ0 âðŽðð= ðŽ(ðŒð â ðð)â (ðŒð â ðð)ðŽ0.
The second identity follows from
ð·ð¶1 = ð¶ â Îâð+ð·ðµâð = ð¶ âð·ðµâðâ ð¶ðð +ð·ðµâð
= ð¶(ðŒð â ðð).Thus both identities are proved. â¡Proof of Theorem 1.1. In the previous section we have seen that the conditions 1,2, 3 in Theorem 1.1 are necessary. Therefore in what follows we assume these threeconditions are fullfilled. The latter allows us to introduce the ðÃð rational matrixfunction
ð (ð ) = ðŒð â ð¶1(ð ðŒð âðŽ0)â1(ðŒð â ðð)â1ðµ. (3.4)
Note that ð is stable, because the matrix ðŽ0 which is given by (1.6) is stable. Thelatter follows from the fact that condition 2 is satisfied. We claim that
ð (ð )â1 = ðŒð + ð¶1(ðŒð â ðð)â1(ð ðŒð âðŽ)â1ðµ. (3.5)
Since ðŽ is stable, we see that ð is invertible outer. To prove (3.5), we use (3.1).Indeed, using (3.1), we obtain
ðŽ0 + (ðŒð â ðð)â1ðµð¶1 = ðŽ0 + (ðŒð â ðð)â1ðŽ(ðŒð â ðð)âðŽ0
= (ðŒð â ðð)â1ðŽ(ðŒð â ðð). (3.6)
Recall that the inverse of ðŒð â ðŸ(ð ðŒð â ðŒ)â1ðœ is the state space realization givenby ðŒð + ðŸ
(ð ðŒð â (ðŒ + ðœðŸ)
)â1ðœ. Using this for the state space realization for ð in
(3.4) with (3.6), we obtain
ð (ð )â1 = ðŒð + ð¶1(ð ðŒð â (ðŒð â ðð)â1ðŽ(ðŒð â ðð))â1(ðŒð â ðð)â1ðµ= ðŒð + ð¶1(ðŒð â ðð)â1(ð ðŒð âðŽ)â1ðµ.
Matrix Solutions of a Bezout Type Equation on the Half-plane 155
Hence the inverse of ð (ð ) is given by (3.5). In particular, ð is an invertible outerfunction.
Next we show that ðº(ð )ð (ð ) = ð·. To do this we use (3.2) together with thestate space formula for ð (ð )â1 in (3.5), to obtain
ð·ð (ð )â1 = ð· +ð·ð¶1(ðŒð â ðð)â1(ð ðŒð âðŽ)â1ðµ= ð· + ð¶(ð ðŒð âðŽ)â1ðµ = ðº(ð ).
In other words, ðº(ð ) = ð·ð (ð )â1. By multiplying the latter identity from the leftby ð (ð ) we obtain ðº(ð )ð (ð ) = ð·.
Finally, by comparing (1.7) and (3.4), we see that Î(ð ) = ð (ð )ð·â(ð·ð·â)â1.It follows that
ðº(ð )Î(ð ) = ðº(ð )ð (ð )ð·â(ð·ð·â)â1 = ð·ð·â(ð·ð·â)â1 = ðŒð.
This completes the proof of Theorem 1.1. â¡
Proof of Theorem 1.2. Given the above proof of Theorem 1.1 it remains to proveitems (i)â(iv) in Theorem 1.2. We do this in four steps.
Step 1. First we show that Î is inner. To do this, recall that
Î(ð ) = ðž + ð¶1(ð ðŒð âðŽ0)â1ðµð, where ðµð = â(ðŒð â ðð)â1ðµðž. (3.7)
We shall make use of the following Lyapunov equation
ðŽâ0(ðâððð) + (ð âððð)ðŽ0 + ð¶â1ð¶1 = 0. (3.8)
To see this, notice, that (3.1), (3.2) and (3.3) with (2.7) and (2.12) yield
ð¶â1ð¶1 = (ð¶â0ð· +ððµ)ð¶1
= ð¶â0ð¶(ðŒð âðð ) +ððŽ(ðŒð â ðð)âð(ðŒð â ðð)ðŽ0
= â(ððŽ+ðŽâ0ð)(ðŒð âðð ) +ððŽ(ðŒð â ðð)âð(ðŒð â ðð)ðŽ0
= âðŽâ0(ð âððð)â (ð âððð)ðŽ0.
Therefore (3.8) holds. The Lyapunov equation in (3.8) also yields
â(ð ðŒð +ðŽâ0)â1ð¶â1ð¶1(ð ðŒð âðŽ0)â1
= (ðâððð)(ð ðŒð âðŽ0)â1 â (ð ðŒð +ðŽâ0)â1(ð âððð).
(3.9)
To see this, simply multiply the previous equation by ð ðŒð + ðŽâ0 on the left and
ð ðŒð âðŽ0 on the right.
156 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
To show that Î is an inner function, notice that (3.9) gives
Î(âð )âÎ(ð ) = (ðžâ âðµâð (ð ðŒð +ðŽâ0)â1ð¶â1 ) (ðž + ð¶1(ð ðŒð âðŽ0)â1ðµð
)= ðŒðâð + ðžâð¶1(ð ðŒð âðŽ0)
â1ðµð âðµâð (ð ðŒð +ðŽâ0)â1ð¶â1ðžâðµâð (ð ðŒð +ðŽâ0)â1ð¶â1ð¶1(ð ðŒð âðŽ0)
â1ðµð
= ðŒðâð + ðžâð¶1(ð ðŒð âðŽ0)â1ðµð âðµâð (ð ðŒð +ðŽâ0)â1ð¶â1ðž
+ðµâð (ðâððð)(ð ðŒð âðŽ0)â1ðµð
âðµâð (ð ðŒð +ðŽâ0)â1(ðâððð)ðµð
= ðŒðâð + (ðµâð (ðâððð) + ðžâð¶1) (ð ðŒð âðŽ0)â1ðµð
âðµâð (ð ðŒð +ðŽâ0)â1 (ð¶â1ðž + (ðâððð)ðµð) = ðŒðâð.
The last equality follows from the fact that
ðµâð (ð âððð) + ðžâð¶1 = 0. (3.10)
To verify this, observe that
ðµâð (ðâððð) = âðžâðµâ(ðŒð âðð )â1(ðâððð) = âðžâðµâððžâð¶1 = ðž
â(ðµâð+ð·âð¶0) = ðžâðµâð.
Hence ðµâð (ðâððð) + ðžâð¶1 = 0. Therefore Î(ð ) is an inner function.
Step 2. It will be convenient first to prove item (iv). Take ð in the right half-plane,i.e., âð ⥠0. Using the definition of ð in (1.9), and the identities (1.7) and (1.11),we see that Î(ð ) = ð (ð )ð·+ and Î(ð ) = ð (ð )ðž, and hence[
Î(ð ) Î(ð )]= ð (ð )
[ð·+ ðž
].
Next observe that the ðà ð matrix [ð·+ ðž]is invertible, and[
ð·+ ðž] [ð·ðžâ
]= ðŒð.
Thus[Î(ð ) Î(ð )
]is invertible, and[
Î(ð ) Î(ð )]â1
=
[ð·ðžâ
]ð (ð )â1 =
[ðº(ð )
ðžâð (ð )â1
].
This proves (1.14). Since the function[Î Î
]is a stable rational ð Ã ð matrix
function, we see that the function defined by (1.13) is invertible outer.
Step 3. In this part we prove item (iii). Let ð be given by (1.12). In otherwords ð(ð ) = ð (ð )(ð·â(ð·ð·â)â1+ðžð(ð )), where ð is an arbitrary stable rationalmatrix function of size (ð â ð) à ð. Since ðº(ð )ð (ð ) = ð·, we see ðº(ð )ð(ð ) =ð·ð·â(ð·ð·â)â1 + ð·ðžð(ð ). But ð·ð·â(ð·ð·â)â1 = ðŒð and ð·ðž = 0. We concludethat ðº(ð )ð(ð ) = ðŒð, as desired.
Next we deal with the reverse implication. Let ð be any stable rationalð Ãð matrix function satsfying the equation ðº(ð )ð(ð ) = ðŒð. Put ð» = ð â Î.Then ð» is a rational matrix-valued function, and ðº(ð )ð»(ð ) = 0. Notice that
Matrix Solutions of a Bezout Type Equation on the Half-plane 157
ðžâð (ð )â1Î(ð ) = 0. Using item (iv) we obtain
ð»(ð ) =[Î(ð ) Î(ð )
] [ ðº(ð )ðžâð (ð )â1
]ð»(ð )
=[Î(ð ) Î(ð )
] [ 0ðžâð (ð )â1ð»(ð )
]= Î(ð )ðžâð (ð )â1ð»(ð ) = Î(ð )ðžâð (ð )â1ð(ð ).
(3.11)
Thus ð»(ð ) = Î(ð )ð(ð ), where ð(ð ) = ðžâð (ð )â1ð(ð ). Since ð is invertible outer,the inverse ð (â )â1 is a rational ðà ð matrix function. Thus ð is a rational matrixfunction of size (ðâð)Ãð, and ð has the desired representation (1.12).
Step 4. We prove item (ii). Let â be any âð-valued ð»2 function such thatðº(ð )â(ð ) = 0 for âð > 0. Repeating the first three identities in (3.11) with âin place of ð» , we see that â(ð ) = Î(ð )ð(ð ), where ð(ð ) = ðžâð (ð )â1â(ð ). Sinceð is invertible outer, the entries of ð (â )â1 are ð»â functions. Hence the entriesof ð are ð»2 functions. Furthermore, using the fact that Î is inner, we see thatð(ð ) = Î(âð )ââ(ð ) for âð > 0. â¡
To complete this section, let us establish the following useful (see the nextsection) identity
Î(âð )âÎ(ð ) = âðµâð (ð ðŒð +ðŽâ0)â1ð¶â0 . (3.12)
For convenience, let us set ðµ1 = â(ðŒð â ðð)â1ðµð·+. Then (3.12) follows from(3.3), (3.9) and (3.10), that is,
Î(âð )âÎ(ð ) = (ðžâ âðµâð (ð ðŒð +ðŽâ0)â1ð¶â1)(ð·+ + ð¶1(ð ðŒð âðŽ0)â1ðµ1
)= ðžâð¶1(ð ðŒð âðŽ0)
â1ðµ1 âðµâð (ð ðŒð +ðŽâ0)â1ð¶â1ð·+
âðµâð (ð ðŒð +ðŽâ0)â1ð¶â1ð¶1(ð ðŒð âðŽ0)â1ðµ1
=((ðžâð¶1 +ðµ
âð (ð âððð)
)(ð ðŒð âðŽ0)
â1ðµ1
âðµâð (ð ðŒð +ðŽâ0)â1(ð¶â1ð·
+ + (ð âððð)ðµ1
)= âðµâð (ð ðŒð +ðŽâ0)â1
(ð¶â1ð·
+ + (ð âððð)ðµ1
)= âðµâð (ð ðŒð +ðŽâ0)â1 (ð¶â1 âððµ)ð·â(ð·ð·â)â1= âðµâð (ð ðŒð +ðŽâ0)â1ð¶â0 .
This establishes (3.12).
4. The minimization problem
Throughout this section ðº is a stable rational ð à ð matrix function, and weassume that ðº is given by the stable state space representation (1.2). We alsoassume that ððº is right invertible.
For each ðŸ > 0 let ð€ðŸ be the scalar weight function given by ð€ðŸ(ð ) = (ð +ðŸ)â1.Note that for each ð â ð»âðÃð the function ð€ðŸð belongs to ð»2
ðÃð. With ðº and
158 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
the weight function ð€ðŸ we associate the following minimization problem:
inf{â¥ð€ðŸðâ¥2 ⣠ðº(ð )ð(ð ) = ðŒð (âð > 0) and ð â ð»âðÃð
}. (4.1)
The problem is to check whether or not the infimum is a minimum, and if so, tofind a minimizing function. We shall show in this section that such a minimizingð exists and is unique. It what follows this minimizing function will be denotedby ÎðŸ . The next theorem shows that ÎðŸ is a stable rational matrix function andprovides an explicit formula for ÎðŸ .
Theorem 4.1. For each ðŸ > 0 there is a unique solution to the optimization problem(4.1), and this solution is given by
ÎðŸ(ð ) = Î(ð )âÎ(ð )ðµâð (ðŸðŒ âðŽâ0)â1ð¶â0 , âð > 0. (4.2)
Here Î and Î are the rational matrix functions given by (1.7) and (1.11), respec-tively, the matrix ðŽ0 is defined by (1.6) and ð¶0 by (2.7). In particular, we haveÎ(ð ) = limðŸââ ÎðŸ(ð ).
Proof. Fix ðŸ > 0. Since for each ð â ð»âðÃð the function ð€ðŸð belongs to ð»2ðÃð,
we have
â¥ð€ðŸÎðŸâ¥2 = inf{â¥ð€ðŸðâ¥2 ⣠ð€ðŸðºð = ð€ðŸðŒð and ð â ð»âðÃð
}(4.3)
⥠inf{â¥ðâ¥2 ⣠ðºð = ð€ðŸðŒð and ð â ð»2
ðÃð
}= â¥ððŸâ¥2. (4.4)
The last optimization problem is a least squares optimization problem. So the op-timal solution ððŸ for the problem (4.4) is unique. We first derive a formula for ððŸ .
From item (ii) in Theorem 1.2 we know that Kerððº = ImðÎ. By taking theFourier transform, we see that ððŸ is the unique matrix function in ð»
2ðÃð such that
ðºððŸ = ð€ðŸðŒð and ððŸ is orthogonal to Îð»2(ðâð)Ãð. Using ðºÎ = ðŒð, we obtain
that all ð»2 solutions to ðºð = ð€ðŸðŒð are given by
ð = ð€ðŸÎ +Îð»2(ðâð)Ãð.
So we are looking for a ð»2 function ððŸ such that
ððŸ = ð€ðŸÎ +Îð¹ and ððŸ ⥠Îð»2(ðâð)Ãð,
where ð¹ is a matrix function inð»2(ðâð)Ãð. By exploiting that Î is inner, we obtain
ð€ðŸÎâÎ+ ð¹ ⥠ð»2
(ðâð)Ãð.
But then (3.12) tells us that the latter is equivalent to
âð€ðŸðµâð (ð ðŒð +ðŽ
â0)â1ð¶â0 + ð¹ ⥠ð»2
(ðâð)Ãð. (4.5)
However,âð€ðŸ(ð )ðµâð (ð ðŒð+ðŽ
â0)â1ð¶â0 admits a partial fraction expansion of the form
âð€ðŸ(ð )ðµâð (ð ðŒð +ðŽ
â0)â1ð¶â0 = (ð + ðŸ)â1ðµâð (ðŸðŒð âðŽâ0)â1ð¶â0 +Ωâ(ð ),
where Ω is in ð»2(ðâð)Ãð. Using this in the orthogonality relation (4.5), we see that
ð¹ (ð ) = âð€ðŸ(ð )ðµâð (ðŸðŒð âðŽâ0)â1ð¶â0 . In other words,
ððŸ(ð ) = ð€ðŸ(ð )Î(ð ) â ð€ðŸ(ð )Î(ð )ðµâð (ðŸðŒð âðŽâ0)â1ð¶â0 . (4.6)
Matrix Solutions of a Bezout Type Equation on the Half-plane 159
Next put ÎðŸ(ð ) = (ð + ðŸ)ððŸ(ð ). Then (4.6) implies that ÎðŸ is given by (4.2).Hence ÎðŸ is a stable rational matrix function. In particular, ÎðŸ belongs to ð»
âðÃð.
Furthermore, it follows that the inequality on the left-hand side of (4.4) is an equal-ity. We conclude that ÎðŸ given by (4.2) is the unique solution to the minimizationproblem (4.1). â¡
References
[1] M. Bakonyi and H.J. Woerdeman, Matrix completions, moments, and sums of Her-mitian squares, Princeton University Press, 2010.
[2] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran, Factorization of matrix andoperator functions: the state space method, OT 178, Birkhauser Verlag, Basel, 2008.
[3] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran, A state space approach tocanonical factorization: convolution equations and mathematical systems, OT 200,Birkhauser Verlag, Basel, 2010.
[4] L. Carlson, Interpolation by bounded analytic functions and the corona problem,Ann. Math. 76 (1962), 547â559.
[5] A.E. Frazho, M.A. Kaashoek, and A.C.M. Ran, The non-symmetric discrete alge-braic Riccati equation and canonical factorization of rational matrix functions onthe unit circle, Integral Equations and Operator Theory 66 (2010), 215â229.
[6] A.E. Frazho, M.A. Kaashoek and A.C.M. Ran, Right invertible multiplication op-erators and stable rational matrix solutions to an associate Bezout equation, I. theleast squares solution. Integral Equations and Operator Theory 70 (2011), 395â418.
[7] C. Foias, A. Frazho, I. Gohberg, and M.A. Kaashoek, Metric constrained interpola-tion, commutant lifting and systems, Birkhauser Verlag, Basel, 1998;
[8] A.E. Frazho, M.A. Kaashoek and A.C.M. Ran, Right invertible multiplication op-erators and stable rational matrix solutions to an associate Bezout equation, II:Description of all solutions, Operators and Matrices 6 (2012), 833â857.
[9] P. Fuhrmann, On the corona theorem and its applications to spectral problems inHilbert space, Trans. Amer. Math. Soc. 132 (1968), 55â66.
[10] P.A. Fuhrmann and R. Ober, On coprime factorizations, in: Operator Theory and itsApplications. The Tsuyoshi Ando Anniversary Volume. OT 62, Birkhauser Verlag,Basel, 1993, pp. 39â75.
[11] T. Georgiou and M.C. Smith, Optimal robustness in the gap metric, IEEE Trans.Automatic Control 35 (1990), 673â686.
[12] I. Gohberg, S. Goldberg, and M.A. Kaashoek, Classes of Linear Operators, VolumeI, OT 49, Birkhauser Verlag, Basel, 1990.
[13] G. Gu and E.F. Badran, Optimal design for channel equalization via the filterbankapproach, IEEE Trans.Signal Proc. 52 (2004), 536â545.
[14] G. Gu and L. Li, Worst-case design for optimal channel equalization in filterbanktransceivers, IEEE Trans. Signal Proc. 51 (2003), 2424â2435.
[15] J.W. Helton, Operator Theory, analytic functions, matrices and electrical engineer-ing, Regional Conference Series in Mathematics 68, Amer. Math. Soc. ProvidenceRI, 1987.
160 A.E. Frazho, M.A. Kaashoek and A.C.M. Ran
[16] N.K. Nikolâskii, Treatise on the shift operator, Grundlehren 273, Springer Verlag,Berlin 1986.
[17] V.V. Peller, Hankel Operators and their Applications, Springer Monographs in Math-ematics, Springer 2003.
[18] V.A. Tolokonnikov, Estimates in Carlsonâs corona theorem. Ideals of the algebrað»â, the problem of Szekefalvi-Nagy, Zap. Naucn. Sem. Leningrad. Otdel. Mat. Inst.Steklov. (LOMI) 113 (1981), 178â1981 (Russian).
[19] S. Treil, Lower bounds in the matrix corona theorem and the codimension one con-jecture, GAFA 14 (2004), 1118â1133.
[20] S. Treil and B.D. Wick, The matrix-valued ð»ð corona problem in the disk andpolydisk, J. Funct. Anal. 226 (2005), 138â172.
[21] T.T. Trent and X. Zhang, A matricial corona theorem, Proc. Amer. Math. Soc. 134(2006), 2549â2558.
[22] T.T. Trent, An algorithm for the corona solutions on ð»â(ð·), Integral Equationsand Operator Theory 59 (2007), 421â435.
[23] M. Vidyasagar, Control system synthesis: a factorization approach, The MIT Press,Cambridge, MA, 1985.
A.E. FrazhoDepartment of Aeronautics and AstronauticsPurdue UniversityWest Lafayette, IN 47907, USAe-mail: [email protected]
M.A. KaashoekDepartment of MathematicsFaculty of SciencesVU UniversityDe Boelelaan 1081aNL-1081 HV Amsterdam, The Netherlandse-mail: [email protected]
A.C.M. RanDepartment of MathematicsFaculty of SciencesVU UniversityDe Boelelaan 1081aNL-1081 HV Amsterdam, The Netherlands
and
Unit for BMI, North-West UniversityPotchefstroom, South Africae-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 161â187câ 2013 Springer Basel
Inverting Structured Operators Related toToeplitz Plus Hankel Operators
M.A. Kaashoek and F. van Schagen
Dedicated to our dear friend Leonia Lerer, on the occasion of his 70th birthday.
Abstract. In this paper the EllisâGohbergâLay theorem on inversion of certainToeplitz plus Hankel operators is derived as a corollary of an abstract inversiontheorem for a certain class of structured operators. The main results also coverthe inversion theorems considered in [6].
Mathematics Subject Classification (2010). Primary 47A62, 47B35; Secondary47A50, 15A09, 65F05.
Keywords. Inversion, structured operators, structured matrices, Toeplitz op-erators, Hankel operators, Stein equation.
1. Introduction
To introduce the main results of this paper we first present an example. Let ðand â be matrix functions of sizes ð à ð and ð à ð, respectively, with entries inthe Wiener algebra on the unit circle (see Section XXIX.2 in [7]). With ð andâ we associate Hankel operators ðº and ð» , where ðº : â2+(â
ð) â â2+(âð) and
ð» : â2+(âð)â â2+(â
ð), as follows:
ðº =
â¡â¢â¢â¢â£ð0 ð1 ð2 â â â ð1 ð2 ð3 â â â ð2 ð3 ð4...
.... . .
â€â¥â¥â¥âŠ and ð» =
â¡â¢â¢â¢â£â0 â1 â2 â â â â1 â2 â3 â â â â2 â3 â4...
.... . .
â€â¥â¥â¥âŠ . (1.1)
Here ðð and âð , ð = 0, 1, 2, . . ., are the Fourier coefficients corresponding to theanalytic parts of ð and â, respectively. Furthermore, we have two invertible blockToeplitz operators ð and ð acting on â2+(â
ð) and â2+(âð), respectively. As for ðº
and ð» the entries of the matrix functions defining ð and ð belong to the Wiener
162 M.A. Kaashoek and F. van Schagen
algebra on the unit circle. We are interested in inverting the operator ð âðºð â1ð» .For this purpose we need linear maps
ð1, ð2 : âð â â1+(â
ð), ð1, ð2 : âð â â1+(â
ð) (1.2)
satisfying the equations
(ð âðºð â1ð»)ð1 = ðð, (ð âð»ð â1ðº)ð1 = ðð, (1.3)
ðâ2(ð âðºð â1ð») = ðâð, ðâ2(ð âð»ð â1ðº) = ðâð . (1.4)
Here for each positive integer ð the symbol ðð denotes the canonical embedding ofâð into â2+(â
ð), that is, ððð¢ = col [ð¿ð,0ð¢]âð=0, where ð¢ â âð and ð¿ð,ð is the Kronecker
delta. Note that ðððâð = ðŒ â ðððâð , where ðð is the forward shift on â2+(âð). Since
â1+(âð) and â1+(â
ð) are contained in â2+(âð) and â2+(â
ð), respectively, the productsðð0 = ð
âððð and ðð0 = ð
âððð , ð = 1, 2, are well defined.
For any linear map ð¥ from âð into â1+(âð) we denote by ðð¥ the block lower
triangular Toeplitz operator acting on â2+(âð) whose first column is equal to ð¥,
and for an ð Ãð matrix ð¢ the symbol Îð¢ denotes the block diagonal operatoron â2+(â
ð) with all diagonal entries equal to ð¢.
Theorem 1.1. Assume there exist linear maps ð1, ð2 and ð1, ð2 as in (1.2) satisfy-ing equations (1.3) and (1.4). Then ð10 = ð
â20 and ð10 = ð
â20. Furthermore, assume
that at least one of the matrices ð10 and ð10 is invertible. Then both the matricesð10 and ð10 are invertible, and the operators ð â ðºð â1ð» and ð â ð»ð â1ðº areinvertible. Moreover,
(ð âðºð â1ð»)â1 = ðð1Îðâ110ð âð2
â ðððð1Îðâ110ð âð2ð
âð , (1.5)
(ð âð»ð â1ðº)â1 = ðð1Îðâ110ð âð2
â ðððð1Îðâ110ð âð2ð
âð . (1.6)
Here
ð1 = âð â1ðºð1, ð1 = âð â1ð»ð1, ðâ2 = âðâ2ð»ð â1, ðâ2 = âðâ2ðºð â1. (1.7)
Note that invertibility of ð â ðºð â1ð» (or ð â ð»ð â1ðº) is equivalent toinvertibility of the operator T given by
ð =
[ð ðºð» ð
]:
[â2+(â
ð)â2+(â
ð)
]â[â2+(â
ð)â2+(â
ð)
]. (1.8)
Moreover, in that case
ðâ1 =[
(ð âðºð â1ð»)â1 âð â1ðº(ð âð»ð â1ðº)â1âð â1ð»(ð âðºð â1ð»)â1 (ð âð»ð â1ðº)â1
].
Using this connection one sees that for the selfadjoint case the above theorem isequivalent to Theorem 3.1 in [5] which is an infinite-dimensional generalization ofthe GohbergâHeinig inversion theorem from [8]. In a somewhat less explicit formthe above theorem also appears in Section 5 of [6].
With the term âToeplitz plus Hankel operatorâ used in the title we have inmind operators of the type (1.8).
Inverting Structured Operators 163
In the present paper we put Theorem 1.1 in a more general setting. Moreprecisely we derive Theorem 1.1 as a corollary of the two abstract inversion the-orems presented below, the alternative versions of these theorems arising fromRemark 1.4, and the auxiliary result Lemma 2.2.
To state our main results we need some additional notation. Consider theHilbert space operators
ðŽ1 : ð³1 â ð³1, ðµ1 : ð° â ð³1, ð¶1 : ð³1 â ðŽ, (1.9)
ðŽ2 : ð³2 â ð³2, ðµ2 : ðŽ â ð³2, ð¶2 : ð³2 â ð° . (1.10)
Throughout we assume that ð : ð³2 â ð³1 andð : ð³1 â ð³2 are operators satisfyingthe following Stein equations
ð âðŽ1ððŽ2 = ðµ1ð¶2, ðâðŽ2ððŽ1 = ðµ2ð¶1. (1.11)
If the identities in (1.11) are satisfied, we shall refer to the set of operators(ðŽ1, ðµ1, ð¶1;ðŽ2, ðµ2, ð¶2) as a data set associated with the pair (ð,ð). We sum-marize the structure of the data set associated with (ð,ð) with the followingdiagram:
ðµ1 ðŽ1 ð¶1ð° ââ ð³1 ââ ð³1 ââ ðŽð â ð â
ð° ââ ð³2 ââ ð³2 ââ ðŽ.ð¶2 ðŽ2 ðµ2
(1.12)
We shall be interested in inverting the operators ðŒð³1 âðð using solutions ofthe following four equations
(ðŒð³1 â ðð)ð = ðµ1, (ðŒð³2 âðð )ð = ðµ2, (1.13)
ð (ðŒð³1 â ðð) = ð¶1, ð(ðŒð³2 âðð ) = ð¶2. (1.14)
Here the unknowns are operators
ð : ð° â ð³1, ð : ð³1 â ðŽ, ð : ðŽ â ð³2, ð : ð³2 â ð° . (1.15)
Theorem 1.2. Let the operators ð, ð , ð, and ð in (1.15) be solutions of theequations (1.13) and (1.14). Then
ðŒðŽ + ð ððµ2 = ðŒðŽ + ð¶1ðð, ðŒð° + ð¶2ðð = ðŒð° + ðððµ1. (1.16)
Assume in addition that ðŒð³1 âðð is invertible. Then ðŒð³1 âðŽ1ððŽ2ð is invertibleif and only if at least one of the two operators ðŒð° + ð¶2ðð and ðŒðŽ + ð ððµ2 isinvertible. In that case both ðŒð° + ð¶2ðð and ðŒðŽ + ð ððµ2 are invertible and
ð(ðŒð³1 â ðð)â1 âðŽ2ð(ðŒð³1 â ðð)â1ðŽ1
=ð (ðŒðŽ + ð ððµ2)â1ð âðŽ2ðð(ðŒð° + ð¶2ðð)
â1ðððŽ1.(1.17)
Theorem 1.3. Assume that there exist operators ð and ð as in (1.15) such that theidentities in (1.14) are satisfied. If the associate operator ðŒðŽ +ð ððµ2 is invertible,then Ker (ðŒð³1 â ðð) â
â©âð=0Kerð¶2ðŽ
ð2ð. Moreover, if the operator ðŒð³1 â ðð is
Fredholm of index zero andâ©â
ð=0Kerð¶2ðŽð2 = {0}, then ðŒð³1 â ðð is invertible.
164 M.A. Kaashoek and F. van Schagen
Remark 1.4. Notice that there is a lot of symmetry in the diagram (1.12) withrespect to the roles of ð and ð, the indices 1 and 2, and the spaces ð° and ðŽ.Therefore, in Theorems 1.2 and 1.3 one may simultaneously interchange thesepairs and obtain analoguous results. Also one can use duality which, for instance,interchanges the roles of ðµð and ð¶ð , ð = 1, 2. To be more precise, let (ðŽ1, ðµ1, ð¶1;ðŽ2, ðµ2, ð¶2) be a data set associated with (ð,ð), then (ðŽ2, ðµ2, ð¶2;ðŽ1, ðµ1, ð¶1) is adata set associated with (ð,ð ) and (ðŽâ2, ð¶
â2 , ðµ
â2 ;ðŽ
â1, ð¶
â1 , ðµ
â1) is a data set associated
with (ð â, ðâ). These invariances under symmetry and duality yield alternativeversions of Theorem 1.2 and Theorem 1.3; see Theorem 2.3 and Theorems 3.6,3.7 and 3.8. The latter three theorems also play a role in the proof of Theorem 1.1.
To illustrate the preceding Theorems 1.2 and 1.3 we briefly sketch how Theo-rem 1.1 can be obtained as a corollary of these two theorems. For simplicity we takehere ð = ðŒ and ð = ðŒ. First we make a special choice of the data in (1.9)â(1.11),as follows:
ð = ðºð» on ð³1 = â2+(â
ð) and ð is the identity operator on ð³2 = â2+(â
ð),
ðŽ1 = ðâð , ðµ1 = ðºðð : â
ð â â2+(âð), ð¶1 = ð
âð : â
2+(â
ð)â âð,
ðŽ2 = ðð, ðµ2 = ðð : âð â â2+(â
ð), ð¶2 = ðâðð» : â2+(â
ð)â âð.
Using that ðº and ð» are Hankel operators, we see that
ð âðŽ1ððŽ2 = ðºð» â ðâððºð»ðð = ðºð» âðºðððâðð»= ðº(ðŒ â ðððâð )ð» = ðºððð
âðð» = ðµ1ð¶2.
Furthermore, ð â ðŽ2ððŽ1 = ðŒ â ðððâð = ðððâð = ðµ2ð¶1. Thus the corresponding
Stein equations (1.11) are satisfied. Furthermore, since ð = ðºð» and ð = ðŒ, wehave ðŒ â ðð = ðŒ âðºð» and the left-hand side of (1.17) becomes
(ðŒ âðºð»)â1 â ðð(ðŒ âðºð»)â1ðâð .Next one shows that in this case, the equations (1.13) and (1.14) are equivalentto the equations (1.3) and (1.4). Hence for this case (1.16) yields ð10 = ð
â20 and
ð10 = ðâ20. It is then a matter of direct checking to prove that the invertibility of
ðŒ âðºð» and ðŒ âð»ðº follows from Theorem 1.3, and that formulas (1.5) and (1.6)can be obtained from (1.17). For further details see the full proof of Theorem 1.1in Section 4.
The setting described by the formulas (1.9)â(1.14) is of particular interestwhen the data are matrices. To illustrate this, assume that ðº and ð» in (1.1) areof finite rank and ð = ðŒ and ð = ðŒ. In that case we know from mathematicalsystems theory (see, e.g., [3, Chapter 7]) that the defining functions ð and â arerational matrix functions and admit stable finite-dimensional realizations:
ð(ð§) = ð¶1(ðŒð1 â ð§ðŽ1)â1ðµ1 and â(ð§) = ð¶2(ðŒð2 â ð§ðŽ2)
â1ðµ2. (1.18)
Here stable means that ðŽ1 and ðŽ2 have their eigenvalues in the open unit disc.From (1.18) it follows that the entries ðð and âð in ðº and ð» are given by ðð =
Inverting Structured Operators 165
ð¶1ðŽð1ðµ1 and âð = ð¶2ðŽ
ð2ðµ2 for ð = 0, 1, 2, . . .. Hence ðº = Î1Î1 and ð» = Î2Î2
with
Îð =[ðµð ðŽððµð ðŽ2
ððµð â â â ] , Îð =
â¡â¢â¢â¢â£ð¶ð
ð¶ððŽð
ð¶ððŽ2ð
...
â€â¥â¥â¥âŠ . (1.19)
Furthermore, (1.11) holds with
ð =
ââð=0
ðŽð1ðµ1ð¶2ðŽ
ð2 = Î1Î2 and ð =
ââð=0
ðŽð2ðµ2ð¶1ðŽ
ð1 = Î2Î1. (1.20)
In particular, one sees that ðŒ â ðºð» = ðŒ â Î1Î1Î2Î2 is invertible if and only ifðŒ â ðð = ðŒ â Î1Î2Î2Î1 is invertible. Moreover, the inverse of ðŒ â ðºð» can beexpressed in terms of the inverse of ðŒâðð and vice versa (cf. the remark precedingLemma 2.2).
Now assume, as in Theorem 1.1, that there exist linear maps ð1, ð2, ð1, ð2as in (1.2) satisfying equations (1.3) and (1.4) (with ð and ð identity operators)such that ð10 or ð10 is invertible. As we have seen, this implies that ðŒ â ðºð» isinvertible. To compute (ðŒ â ðºð»)â1 we apply Theorem 1.2 with ð , ð, and theassociate data {ðŽ1, ðµ1, ð¶1;ðŽ2, ðµ2, ð¶2} as in (1.20). In this case equations (1.13)and (1.14) are just matrix equations, and identity (1.17) leads to the followinginversion formula:
(ðŒ âðºð»)â1 = ðŒ +ð»ð1Îðâ110ð»âð2 â ðâðð»ð1Îðâ1
10ð»âð2
ðð. (1.21)
Here ð1 and ð2 are defined by the first and third identity in (1.7), and ð»ð¥ denotesthe Hankel operator with first column equal to ð¥. Moreover, the Hankel operatorsappearing in (1.21) are all of finite rank. The inversion formula (1.21) is the discreteanalogue of the inversion formula in [10, Theorem 0.1]. The full proof of (1.21)will be presented in Subsection 4.2.
We shall also show that the inversion formulas for ð â ðºð â1ð» and ð âð»ð â1ðº in Theorem 1.1 can be replaced by formulas analogous to (1.21), that is,with the Toeplitz operators in (1.5) and (1.6) being replaced by Hankel operators.See Theorem 4.3 for the precise formulation.
Theorems 1.2 and 1.3 also apply to other inversion problems than the onesrelated to Theorem 1.1. As an illustration we derive Theorem 2.1 in [6] as a corol-lary of Theorems 1.2 and 1.3. This implies that all examples in [6] are also coveredby Theorems 1.2 and 1.3 above. Whether or not the main inversion theorems inSection 2 of [6] also imply Theorems 1.2 and 1.3 remains an open question.
The operators considered in this paper belong to the area of structured ma-trices and operators which include Toeplitz and Hankel operators, Vandermonde,Cauchy and Pick matrices, resultants and Bezoutians, controllability and observ-ability operators, and many other classes of matrices and operators. The literatureon the subject is vast. Here we only mention [14], the review paper [12], and the
166 M.A. Kaashoek and F. van Schagen
books [11], [13], [15], [16]. We see our Theorems 1.2 and 1.3 as a contribution tothis rich field of research.
This paper consists of six sections including the present introduction. Theproofs of Theorems 1.2 and 1.3 are given in Sections 2 and 3, respectively. InSection 4 we return to Theorem 1.1, and complete the sketch of the proof givenabove. In this section we also prove (1.21) and derive formulas for (ð âðºð â1ð»)â1and (ð â ð»ð â1ðº)â1 analogous to (1.21). In Section 5 we derive Theorem 2.1in [6] as a corollary of Theorems 1.2 and 1.3 above. In general, with ð and ðgiven, one can find different sets of operators (ðŽ1, ðµ1, ð¶1;ðŽ2, ðµ2, ð¶2) such theequations (1.9), (1.10), and (1.11) are satisfied. These different choices of the dataset (ðŽ1, ðµ1, ð¶1;ðŽ2, ðµ2, ð¶2) often lead to different versions of formula (1.17), and itcan happen that for some choice of the data set formula (1.17) leads to a formula forð(ðŒ âðð)â1 while for another choice it does not. This phenomenon is illustratedon finite block Toeplitz matrices in the final section.
2. Proof of Theorem 1.2
In this section we prove Theorem 1.2. For simplicity, in this section, any identityoperator will be denoted by ðŒ, that is, in what follows we shall omit the subscriptindicating on which space the identity operator acts. Furthermore, we shall freelyuse the operators appearing in (1.9), (1.10), and (1.11). We begin with two lemmas.
Lemma 2.1. If the second identity of (1.13) and the first of (1.14) are satisfied,then ðŒ + ð ððµ2 = ðŒ + ð¶1ðð . Analogously, if the first identity of (1.13) and thesecond of (1.14) are satisfied, then ðŒ + ð¶2ðð = ðŒ + ðððµ1.
Proof. From the second identity of (1.13) and the first of (1.14) we get
ðŒ + ð ððµ2 = ðŒ + ð ð (ðŒ âðð )ð = ðŒ + ð (ðŒ â ðð)ðð = ðŒ + ð¶1ðð.
The other two identities give
ðŒ + ð¶2ðð = ðŒ + ð(ðŒð³ âðð )ðð = ðŒ + ðð(ðŒ â ðð)ð = ðŒ + ðððµ1. â¡
In the proof of the following lemma we use a few times the classical resultthat given two operators ð¹1 and ð¹2 the invertibility of ðŒ + ð¹1ð¹2 is equivalent tothe invertibility of ðŒ+ð¹2ð¹1. Moreover, in that case the inverse of ðŒ+ð¹1ð¹2 is givenby (ðŒ + ð¹1ð¹2)
â1 = ðŒ â ð¹1(ðŒ + ð¹2ð¹1)â1ð¹2 (see [1], first paragraph on page 30).Lemma 2.2. Assume that ðŒ â ðð is invertible, and let ð and ð be the opera-tors defined by the first equations of (1.13) and (1.14). Then the following areequivalent:
(i) ðŒ + ð¶2ðð is invertible,(ii) ðŒ + ð ððµ2 is invertible,(iii) ðŒ âðŽ1ððŽ2ð is invertible.
Inverting Structured Operators 167
Furthermore, in that case ðŒ â ððŽ2ððŽ1 is also invertible and
(ðŒ + ð¶2ðð)â1 = ðŒ â ð¶2ð(ðŒ âðŽ1ððŽ2ð)
â1ðµ1, (2.1)
(ðŒ + ð ððµ2)â1 = ðŒ â ð¶1(ðŒ â ððŽ2ððŽ1)
â1ððµ2. (2.2)
Proof. Since ð = (ðŒ â ðð)â1ðµ1, we have
ðŒ + ð¶2ðð = ðŒ + ð¶2ð(ðŒ â ðð)â1ðµ1.
It follows that ðŒ + ð¶2ðð is invertible if and only if
ðŒ +ðµ1ð¶2ð(ðŒ â ðð)â1 =(ðŒ â (ððâðµ1ð¶2ð)
)(ðŒ â ðð)â1
is invertible. By the first identity in (1.11), we have ðð â ðµ1ð¶2ð = ðŽ1ððŽ2ð.This proves the equivalence of (i) and (iii). Moreover in that case
(ðŒ + ð¶2ðð)â1 =
(ðŒ + ð¶2ð(ðŒ â ðð)â1ðµ1
)â1= ðŒ â ð¶2ð(ðŒ â ðð)â1
(ðŒ +ðµ1ð¶2ð(ðŒ â ðð)â1
)â1ðµ1
= ðŒ â ð¶2ð(ðŒ âðŽ1ððŽ2ð
)â1ðµ1.
This proves identity (2.1).Since ðŒ â ðŽ1ððŽ2ð is invertible if and only if ðŒ â ððŽ2ððŽ1 is invertible,
the equivalence of (ii) and (iii), and the identity (2.2) can be proved using anappropriate modification of the arguments employed in the previous paragraph.
â¡
Proof of Theorem 1.2. Assuming that ðŒ âðð is invertible and given Lemmas 2.1and 2.2, it remains to prove the identity (1.17). We will show that
ð (ðŒ + ð ððµ2)â1ð = ð(ðŒ â ðð)â1 â (ðŒ âðŽ2ððŽ1ð )
â1ðŽ2ððŽ1, (2.3)
ðŽ2ðð(ðŒ + ð¶2ðð)â1ðððŽ1 = ðŽ2ð(ðŒ â ðð)â1ðŽ1 â (ðŒ âðŽ2ððŽ1ð )
â1ðŽ2ððŽ1.(2.4)
Subtracting these two identities (1.17) appears. In deriving (2.3) and (2.4) we shalluse a few times that
ðµ1ð¶2ð = (ðŒ âðŽ1ððŽ2ð)â (ðŒ â ðð), (2.5)
ðµ2ð¶1ð = (ðŒ âðŽ2ððŽ1ð )â (ðŒ âðð ). (2.6)
These identities follow from the ones in (1.11).Let us now prove (2.3). Using (2.2) and (2.6), a standard computation (cf.
the state space formulas in Theorem 2.4 in [2]) yields
ð (ðŒ + ð ððµ2)â1 = (ðŒ âðð )â1ðµ2
(ðŒ â ð¶1(ðŒ â ððŽ2ððŽ1)
â1ððµ2
)= (ðŒ âðð )â1ðµ2
(ðŒ â ð¶1ð (ðŒ âðŽ2ððŽ1ð )
â1ðµ2
)= (ðŒ âðŽ2ððŽ1ð )
â1ðµ2.
168 M.A. Kaashoek and F. van Schagen
We proceed with
ð (ðŒ + ð ððµ2)â1ð = (ðŒ âðŽ2ððŽ1ð )
â1ðµ2ð¶1(ðŒ â ðð)â1
= (ðŒ âðŽ2ððŽ1ð )â1ðµ2ð¶1
(ðŒ + ðð(ðŒ â ðð)â1
)= (ðŒ âðŽ2ððŽ1ð )
â1ðµ2ð¶1
+ (ðŒ âðŽ2ððŽ1ð )â1(ðµ2ð¶1ð )(ðŒ âðð )â1ð.
Again using (2.6), it follows that
ð (ðŒ + ð ððµ2)â1ð = (ðŒ âðŽ2ððŽ1ð )
â1ðµ2ð¶1
+ (ðŒ âðð )â1ð â (ðŒ âðŽ2ððŽ1ð )â1ð.
According to the second identity in (1.11) we have ðµ2ð¶1 â ð = âðŽ2ððŽ1, andhence the above calculations yield (2.3).
Next we prove (2.4). Using (2.1) and (2.5) (cf. the state space formulas inTheorem 2.4 in [2]) we have
ðŽ2ðð(ðŒ + ð¶2ðð)â1 = ðŽ2ð(ðŒ â ðð)â1ðµ1
(ðŒ â ð¶2ð(ðŒ âðŽ1ððŽ2ð)
â1ðµ1
)= ðŽ2ð(ðŒ âðŽ1ððŽ2ð)
â1ðµ1.
Hence
ðŽ2ðð(ðŒ + ð¶2ðð)â1ð = ðŽ2ð(ðŒ âðŽ1ððŽ2ð)
â1ðµ1ð¶2(ðŒ âðð )â1= ðŽ2ð(ðŒ âðŽ1ððŽ2ð)
â1ðµ1ð¶2
+ðŽ2ð(ðŒ âðŽ1ððŽ2ð)â1ðµ1ð¶2ðð (ðŒ âðð )â1
= ðŽ2ð(ðŒ âðŽ1ððŽ2ð)â1ðµ1ð¶2
+ðŽ2ð(ðŒ âðŽ1ððŽ2ð)â1ðµ1ð¶2ð(ðŒ â ðð)â1ð.
Using (2.5) we see that
ðŽ2ðð(ðŒ + ð¶2ðð)â1ð = ðŽ2ð(ðŒ âðŽ1ððŽ2ð)
â1ðµ1ð¶2
âðŽ2ð(ðŒ âðŽ1ððŽ2ð)â1ð +ðŽ2ð(ðŒ â ðð)â1ð
= âðŽ2ð(ðŒ âðŽ1ððŽ2ð)â1ðŽ1ððŽ2 +ðŽ2ðð (ðŒ âðð )â1.
Now rewrite
ðŽ2ð(ðŒ âðŽ1ððŽ2ð)â1ðŽ1ððŽ2 = ðŽ2ððŽ1ð (ðŒ âðŽ2ððŽ1ð )
â1ðŽ2
=(ðŒ â (ðŒ âðŽ2ððŽ1ð )
)(ðŒ âðŽ2ððŽ1ð )
â1ðŽ2
= (ðŒ âðŽ2ððŽ1ð )â1ðŽ2 âðŽ2.
Similarly
ðŽ2ð(ðŒ â ðð)â1ð = ðŽ2ðð (ðŒ âðð )â1
= ðŽ2
(ðŒ â (ðŒ âðð )
)(ðŒ âðð )â1 = ðŽ2(ðŒ âðð )â1 âðŽ2.
Inverting Structured Operators 169
It follows that
ðŽ2ðð(ðŒ + ð¶2ðð)â1ð = â(ðŒ âðŽ2ððŽ1ð )
â1ðŽ2 +ðŽ2(ðŒ âðð )â1.Hence
ðŽ2ðð(ðŒ + ð¶2ðð)â1ðððŽ1 = â(ðŒ âðŽ2ððŽ1ð )
â1ðŽ2ððŽ1 +ðŽ2(ðŒ âðð )â1ððŽ1
= â(ðŒ âðŽ2ððŽ1ð )â1ðŽ2ððŽ1 +ðŽ2ð(ðŒ â ðð)â1ðŽ1.
Thus (2.4) holds, and the proof is complete. â¡
Using Remark 1.4 about interchanging the roles of ð and ð one obtains thefollowing alternative of Theorem 1.2.
Theorem 2.3. Let the operators ð, ð , ð, and ð in (1.15) be solutions of theequations (1.13) and (1.14). Then
ðŒðŽ + ð ððµ2 = ðŒðŽ + ð¶1ðð, ðŒð° + ð¶2ðð = ðŒð° + ðððµ1.
Assume in addition that ðŒð³2 âðð is invertible. Then ðŒð³2 âðŽ2ððŽ1ð is invertibleif and only if at least one of the two operators ðŒðŽ + ð¶1ðð and ðŒð° + ðððµ1 isinvertible. In that case both ðŒðŽ + ð¶1ðð and ðŒð° + ðððµ1 are invertible and
ð (ðŒð³2 âðð )â1 âðŽ1ð (ðŒð³2 âðð )â1ðŽ2
= ð(ðŒð° + ðððµ1)â1ð âðŽ1ðð (ðŒðŽ + ð¶1ðð )â1ð ððŽ2.
3. Conditions of invertibility, proof of Theorem 1.3
The next four results extend and sharpen Theorem 1.3. Hence in order to proveTheorem 1.3 it suffices to prove the four results presented below.
Proposition 3.1. Assume that there exist operators ð and ð as in (1.15) such that(1.14) is satisfied, and let the associated operator ðŒðŽ + ð ððµ2 be invertible. Then
(ðŒð³1 â ðð)ð¥ = 0 â ð¶2ðŽð2ðð¥ = 0, (ð = 0, 1, 2, . . .).
Proposition 3.2. Assume that there exist operators ð and ð as in (1.15) suchthat (1.13) is satisfied, and let the associated operator ðŒð° + ð¶2ðð be invertible.Then
ðŠ(ðŒð³1 â ðð) = 0 â ðŠððŽð2ðµ2 = 0 (ð = 0, 1, 2, . . .).
Corollary 3.3. Assume that the operator ðŒð³1 â ðð is Fredholm of index zero andthat
â©âð=0Kerð¶2ðŽ
ð2 = {0}. Furthermore, assume that there exist operators ð and
ð as in (1.15) such that (1.14) is satisfied. Then ðŒðŽ + ð ððµ2 is invertible impliesthat ðŒð³1 â ðð is invertible.
Corollary 3.4. Assume that the operator ðŒð³1 â ðð is Fredholm of index zero andthat
â©âð=0Kerðµ
â2ðŽ
âð2 = {0}. Furthermore, assume that there exist operators ð
and ð as in (1.15) such that (1.13) is satisfied. Then ðŒð° + ð¶2ðð is invertibleimplies that ðŒð³1 â ðð is invertible.
170 M.A. Kaashoek and F. van Schagen
Using Remark 1.4 we see that it suffices to prove Proposition 3.1 and Corol-lary 3.3.
Proof of Proposition 3.1. Assume that the operator ðŒðŽ + ð ððµ2 is invertible, andlet (ðŒð³1 â ðð)ð¥ = 0. Then ð¶1ð¥ = ð (ðŒð³1 â ðð)ð¥ = 0 and
ð¶2ðð¥ = ð(ðŒð³1 âðð )ðð¥ = ðð(ðŒð³1 â ðð)ð¥ = 0.
Also
(ðŒð³1 âðŽ1ððŽ2ð)ð¥ =(ðŒð³1 â (ð âðµ1ð¶2)ð
)ð¥ = (ðŒð³1 â ðð+ðµ1ð¶2ð)ð¥
= (ðŒð³1 â ðð)ð¥+ ðµ1ð¶2ðð¥ = 0,
and
(ðŒð³1 â ððŽ2ððŽ1)ð¥ =(ðŒð³1 â ð (ðâðµ2ð¶1)
)ð¥ = (ðŒð³1 â ðð+ ððµ2ð¶1)ð¥
= (ðŒð³1 â ðð)ð¥+ ððµ2ð¶1ð¥ = 0.
Next observe that
(ðŒðŽ + ð ððµ2)ð¶1 = ð¶1 + ð ððµ2ð¶1 = ð¶1 + ð ð (ðâðŽ2ððŽ1)
= ð¶1 + ð ððâ ð ððŽ2ððŽ1 = ð â ð ððŽ2ððŽ1.
We see that
(ðŒðŽ + ð ððµ2)ð¶1ððŽ2ðð¥ = ð ððŽ2ðð¥â ð ððŽ2ððŽ1ððŽ2ðð¥
= ð ððŽ2ð(ðŒð³1 âðŽ1ððŽ2ð)ð¥ = 0.
By assumption ðŒðŽ + ð ððµ2 is invertible. Hence ð¶1ððŽ2ðð¥ = 0. Therefore
0 = ðµ2ð¶1ððŽ2ðð¥ = (ðâ ðŽ2ððŽ1)ððŽ2ðð¥ = ðððŽ2ðð¥â (ðŽ2ð)ðŽ1ððŽ2ðð¥
= ðððŽ2ðð¥âðŽ2ðð¥ = â(ðŒð³2 âðð )ðŽ2ðð¥.
We conclude that
(ðŒð³1 â ðð)ð¥ = 0 â (ðŒð³2 âðð )ðŽ2ðð¥ = 0. (3.1)
Next we prove that
(ðŒð³1 â ðð)ð¥ = 0 â (ðŒð³2 âðð )ðŽð2ðð¥ = 0, ð = 0, 1, 2, . . . . (3.2)
For ð = 0 we have (ðŒð³2 âðð )ðð¥ = ð(ðŒð³2 âðð)ð¥ = 0. We proceed by induction.Assume that the right-hand side of (3.2) holds for ð = ð ⥠0. Then
(ðŒð³1 â ðð)ððŽð2ðð¥ = ð (ðŒð³2 âðð )ðŽð
2ðð¥ = 0.
Thus (ðŒð³1 âðð)ᅵᅵ = 0 where ð¥ = ððŽð2ðð¥. Now apply (3.1) with ᅵᅵ replacing ð¥. It
follows that
0 = (ðŒð³2 âðð )ðŽ2ðᅵᅵ = (ðŒð³2 âðð )ðŽ2ðððŽð2ðð¥
= â(ðŒð³2 âðð )ðŽ2(ðŒð³2 âðð )ðŽð2ðð¥+ (ðŒð³2 âðð )ðŽ2ðŽ
ð2ðð¥
= (ðŒð³2 âðð )ðŽð+12 ðð¥.
Inverting Structured Operators 171
Thus by induction (3.2) holds. Using (3.2) we see that
ð¶2ðŽð2ðð¥ = ð(ðŒð³2 âðð )ðŽð
2ðð¥ = 0, ð = 0, 1, 2, . . . .
We proved the proposition. â¡
Proof of Corollary 3.3. Let the operator ðŒðŽ + ð ððµ2 be invertible, and assumethat (ðŒð³1 â ðð)ð¥ = 0. According to Proposition 3.1 this implies that the vectorðð¥ belongs to
â©âð=0Kerð¶2ðŽ
ð2 . By our hypotheses the latter space consists of the
zero vector only. So ðð¥ = 0. But then ð¥ = ððð¥ = 0. Since ðŒð³1 â ðð is Fredholmof index zero and Ker (ðŒð³1 â ðð) = 0, it follows that the operator ðŒð³1 â ðð isinvertible. â¡
Remark 3.5. Assume thatâ©â
ð=0Kerð¶2ðŽð2 = {0} and ðŒð³1 â ðð is Fredholm of
index zero. Assume also that ð and ð exist satisfying (1.14). Then Proposition 3.1and Lemma 2.2 imply that the operator ðŒðŽ + ð ððµ2 is invertible if and only if theoperators ðŒð³1 â ðð and ðŒð³1 âðŽ1ððŽ2ð are both invertible.
We conclude this section with three alternatives of Theorem 1.3. They followdirectly from Theorem 1.3 by using the symmetry and duality arguments men-tioned in Remark 1.4.
Theorem 3.6. Assume that there exist operators ð and ð as in (1.15) such that theidentities in (1.14) are satisfied. If the associated operator ðŒð°+ðððµ1 is invertible,then Ker (ðŒð³2 â ðð ) â
â©âð=0Kerð¶1ðŽ
ð1ð . Moreover, if the operator ðŒð³2 â ðð is
Fredholm of index zero andâ©â
ð=0Kerð¶1ðŽð1 = {0}, then ðŒð³2 âðð is invertible.
Theorem 3.7. Assume that there exist operators ð andð as in (1.15) such that theidentities in (1.13) are satisfied. If the associate operator ðŒð° +ð¶2ðð is invertible,then ImððŽð
2ðµ2 â Im (ðŒð³1 â ðð) for ð = 0, 1, 2, . . .. Moreover, if ðŒð³1 â ðð isFredholm of index zero and span {ImðŽð
2ðµ2 ⣠ð = 0, 1, 2, . . .} is dense in ð³2, thenðŒð³1 â ðð is invertible.
Theorem 3.8. Assume that there exist operators ð and ð as in (1.15) such that theidentities in (1.14) are satisfied. If the associate operator ðŒðŽ+ð¶1ðð is invertible,then ImððŽð
1ðµ1 â Im (ðŒð³2 â ðð ) for ð = 0, 1, 2, . . .. Moreover, if ðŒð³2 â ðð isFredholm of index zero and span {ImðŽð
1ðµ1 ⣠ð = 0, 1, 2, . . .} is dense in ð³1, thenðŒð³2 âðð is invertible.
4. Toeplitz plus Hankel operators
This section consists of three subsections. In the first we prove Theorem 1.1. Inthe second subsection we derive the identity (1.21), and in the third we deriveformulas for (ð âðºð â1ð») and (ð âð»ð â1ðº)â1 analogous to the one in (1.21).We begin with a general remark.
Remark. Since the entries of the matrix functions defining the operators ðº, ð», ð ,ð all belong to the Wiener algebra on the unit circle, it follows that for arbitrary
172 M.A. Kaashoek and F. van Schagen
linear maps ð¥ : âð â â1+(âð) and ðŠ : âð â â1+(â
ð) one has
ð»ð¥ : âð â â1+(âð), ð ð¥ : âð â â1+(â
ð), ð â1ð¥ : âð â â1+(âð),
ðºðŠ : âð â â1+(âð), ð ðŠ : âð â â1+(â
ð), ð â1ðŠ : âð â â1+(âð).
In particular, the linear maps ð1, ð2 and ð1, ð2 defined by (1.7) have their valuesin â1+(â
ð) and â1+(âð), respectively.
4.1. Proof of Theorem 1.1
In this subsection we prove Theorem 1.1. In order to do this we apply Theorems 1.2and 1.3 with a special choice of ð and ð and the associated data set, namely
ð = ðŒð³1 â (ð âðºð â1ð»), ð = ðŒð³2 , (4.1)
ðŽ1 = ðâð , ðµ1 = ðºð
â1ðð(ðâððâ1ðð)â1, ð¶1 = ð
âð, (4.2)
ðŽ2 = ðð, ðµ2 = ðð, ð¶2 = ðâðð
â1ð». (4.3)
Here ð³1 = ð³2 = â2+(â
ð), and ðð is the forward shift on â2+(â
ð). Note that
ðŒ â ðð = ð âðºð â1ð».Since ð is the invertible and ðº (and ð») are compact operators, we see that ðŒâððis a Fredholm operator of index zero.
To see that for the above operators the identities in (1.11) are valid we firstrecall a useful equality. Note that ðâðð â1ðð is the entry in the left upper corner ofthe block matrix representing ð â1. Since ð is an invertible Toeplitz operator, thematrix ðâðð
â1ðð is invertible and
ð â1 â ððð â1ðâð = (ð â1ðð)(ðâððâ1ðð)â1(ðâðð
â1). (4.4)
This result is well known and follows using a standard Schur complement argument(see, e.g., [6, Section 4]).
Next we deal with the Stein equations (1.11). Using that ðº and ð» are Hankeloperators, that ð is a Toeplitz operator, and that ð â1 satisfies (4.4) we see that
ð âðŽ1ððŽ2 = ðŒ â (ð âðºð â1ð»)â ðâððð + ðâð(ð âðºð â1ð»)ðð= âð + ðâðð ðð +ðºð â1ð» â ðâððºð â1ð»ðð= ðº(ð â1 â ððð â1ðâð )ð»= ðº(ð â1ðð)(ðâðð
â1ðð)â1(ðâððâ1)ð» = ðµ1ð¶2.
Also
ðâðŽ2ððŽ1 = ðŒ â ðððâð = ðððâð = ðµ2ð¶1.
Thus equations (1.11) are satisfied. Furthermore, since ðŒ â ðð = ð â ðºð â1ð» ,the left-hand side of (1.17) becomes
(ð âðºð â1ð»)â1 â ðð(ð âðºð â1ð»)â1ðâð .
Inverting Structured Operators 173
Finally we see that
ðŒ âðŽ1ððŽ2ð = ðŒ â ðâð(ðŒ â (ð âðºð â1ð»))ðð
= ðŒ â ðâððð + (ðâðð ðð â ðâððºð â1ð»ðð)= ð âðº1ð
â1ð»1,
(4.5)
where ðº1 = ðâððº and ð»1 = ð
âðð» .
As a next step towards the proof of Theorem 1.1 it will be convenient first toprove the following proposition. In what follows we freely use the terminology andnotation introduced in Theorem 1.1 and in the paragraphs preceding Theorem 1.1.
Proposition 4.1. The following five conditions are equivalent.
(1) Equation (1.3) has solutions ð1 and ð1 and at least one of the matrices ð10and ð10 is invertible.
(2) Equation (1.3) has solutions ð1 and ð1 and both matrices ð10 and ð10 areinvertible.
(3) Equation (1.4) has solutions ðâ2 and ðâ2 and at least one of the matrices ðâ20and ðâ20 is invertible.
(4) Equation (1.4) has solutions ðâ2 and ðâ2 and both matrices ðâ20 and ðâ20 areinvertible.
(5) The operators ð âðºð â1ð» and ð âðº1ðâ1ð»1, where ðº1 = ð
âððº and ð»1 =
ðâðð», are invertible.
Moreover in that case
(ð âðºð â1ð»)â1 â ðð(ð âðºð â1ð»)â1ðâð= ð1ð
â110 ð
â2 â ððð â1ðºð1ðâ110 ð
â2ð»ð
â1ðâð ,(4.6)
(ð âð»ð â1ðº)â1 â ðð(ð âð»ð â1ðº)â1ðâð= ð1ð
â110 ð
â2 â ððð â1ð»ð1ðâ110 ð
â2ðºð
â1ðâð .(4.7)
Proof. We split the proof into seven parts. The first part has an auxiliary character.The equivalence of the five conditions is proved in Parts 2â6. In the final part weprove formulas (4.6) and (4.7). Throughout we use the operators defined by (4.1),(4.2), and (4.3).
Part 1. We shall present operators satisfying (1.13) and (1.14). When ð1 and ð1are linear maps satisfying equation (1.3), we put
ð = ð â1ðºð1(ðâððâ1ðð)â1, ð = ð1. (4.8)
We claim that with ð and ð defined as in the first paragraph of this subsection,the operators ð and ð in (4.8) satisfy (1.13). Indeed, using (1.3), we have
(ðŒ â ðð)ð = (ð âðºð â1ð»)ð â1ðºð1(ðâðð â1ðð)â1= ðºð â1(ð âð»ð â1ðº)ð1(ðâðð â1ðð)â1= ðºð â1ðð(ðâðð
â1ðð)â1 = ðµ1,
(ðŒ âðð )ð = (ð âðºð â1ð»)ð1 = ðð = ðµ2.
174 M.A. Kaashoek and F. van Schagen
Hence (1.13) is satisfied. Furthermore, (1.3) gives
ðŒ + ð¶1ðð = ðŒ + ðâððð1 = ðŒ + ðâð(ðŒ â (ð âðºð â1ð»))ð1
= ðâðð1 + ðŒ â ðâððð = ð10,ðŒ + ð¶2ðð = ðŒ + ðâðð
â1ð»ð â1ðºð1(ðâððâ1ðð)â1
= ðŒ â ðâðð â1(ð âð»ð â1ðº)ð1(ðâðð â1ðð)â1 + ðâðð1(ðâðð â1ðð)â1= ðŒ â ðâðð â1ðð(ðâðð â1ðð)â1 + ðâðð1(ðâðð â1ðð)â1= ð10(ð
âðð
â1ðð)â1.
Similar results hold for ð2 in place of ð1 and ð2 in place of ð1. Indeed, whenð2 and ð2 are linear maps satisfying equation (1.4), we put
ð = ðâ2, ð = ðâ2ð»ð â1. (4.9)
These operators satisfy (1.14). Indeed, using (1.4), we have
ð(ðŒ âðð ) = ðâ2ð»ð â1(ð âðºð â1ð»)= ðâ2(ð âð»ð â1ðº)ð â1ð» = ðâðð
â1ð» = ð¶2,
ð (ðŒ â ðð) = ðâ2(ð âðºð â1ð») = ðâð = ð¶1.
Hence (1.14) is satisfied. Furthermore, (1.4) gives
ðŒ + ð ððµ2 = ðŒ + ðâ2(ðŒ â (ð âðºð â1ð»))ðð
= ðŒ â ðâ2(ð âðºð â1ð»)ðð + ðâ2ðð = ðâ2ðð = ðâ20,ðŒ + ðððµ1 = ðŒ + ð
â2ð»ð
â1ðºð â1ðð(ðâððâ1ðð)â1
= ðŒ + ðâ2ðð(ðâðð
â1ðð)â1 â ðâ2(ð âð»ð â1ðº)ð â1ðð(ðâðð â1ðð)â1= ðŒ + ðâ2ðð(ð
âðð
â1ðð)â1 â ðâðð â1ðð(ðâðð â1ðð)â1 = ðâ20(ðâðð â1ðð)â1.Part 2. In this part we show that condition (5) implies conditions (1)â(4). Soassume (5) is satisfied. Then there exists linear maps as in (1.2) satisfying equations(1.3) and (1.4). Also ðŒ â ðð is invertible, and using (4.5) we see that the sameholds true for ðŒ â ðŽ1ððŽ2ð. From Lemma 2.2 we conclude that ðŒ + ð¶2ðð andðŒ + ð ððµ2 are invertible. Hence ð10 and ð
â20 are invertible. Now use Theorem 1.2
and notice that the equalities in (1.16) imply that ðŒ+ðððµ1 and ðŒ+ð¶1ðð also areinvertible. We conclude that ðâ20 and ð10 are invertible, ð10 = ð
â20 and ð10 = ð
â20.
So indeed condition (5) implies conditions (1)â(4).
Next we make a remark that will help to simplify the remaining parts of theproof. Assume that ðŒ â ðð is invertible. Then the equations (1.13) and (1.14)are uniquely solvable, and with these solutions the equalities in (1.16) hold true.Moreover Lemma 2.2 shows that if one of the four operators in (1.16) is invertible,then ðŒ âðŽ1ððŽ2ð is invertible and hence condition (5) is satisfied. Conclusion: inorder to finish the proof of the equivalence of the five conditions (1)â(5) we onlyhave to show that (1) and (3) each separately imply that ðŒ â ðð is invertible.(Trivially, condition (2) implies (1) and (4) implies (3).)
Inverting Structured Operators 175
Part 3. In this part we show that condition (1) with ð10 invertible implies thatðŒ â ðð is invertible. Define ð and ð by (4.8). As we have seen in the firstpart of the proof, ðŒ + ð¶1ðð = ð10, and hence ðŒ + ð¶1ðð is invertible. Assumethat ðŠ(ðŒ â ðð ) = 0. According to Theorem 3.8 it follows that ðŠððŽð
1ðµ1 = 0 forð = 0, 1, 2, . . .. So for all nonnegative integers ð we obtain that ðŠ(ðâð)
ððºð â1ðð = 0.
Here we use that ðâðð â1ðð is invertible. Since ð is invertible, ð factors as ð =
ðâð+, where ð+ and ðâ are invertible Toeplitz operators, ð+ and ð â1+ are both
lower triangular, and ðâ and ð â1â are both upper triangular (see [9] or [2, Theorem1.2]). Then
0 = ðŠ(ðâð)ððºð â1ðð = (ðŠðº)ðð
ð ðâ1ðð = (ðŠðº)ðð
ð ðâ1+ ð â1â ðð = (ðŠðº)ð â1+ ðð
ð ððð£â0,
where ð£â0 is the invertible (1, 1)-entry of ð â1â . Hence (ðŠðºð â1+ )ððð ðð = 0 for ð =
0, 1, 2, . . .. So we obtain ðŠðºð â1+ = 0, and ðŠðº = 0. Since ðŒâðð = ð âðºð â1ð» wehave ðŠ(ð â ðºð â1ð») = 0, and ðŠðº = 0 implies that ðŠð = 0. But ð is invertible,and therefore ðŠ = 0. So ðŒ â ðð has a dense range. Since ðŒ â ðð is a Fredholmoperator of index zero, it follows that ðŒ â ðð is invertible.
Part 4. In this part we show that condition (1) with ð10 invertible implies thatðŒâðð is invertible. Define ð and ð by (4.8). As we have seen in the first part ofthe proof, ðŒ+ð¶2ðð = ð10, and hence the operator ðŒ+ð¶2ðð is invertible. Assumethat ðŠ(ðŒ â ðð) = 0. From Theorem 3.7 we see that 0 = ðŠððŽð
2ðµ2 = ðŠðððð ðð for
ð = 0, 1, 2, . . .. It follows that ðŠð = 0, and therefore ðŠ = ðŠâðŠðð = ðŠ(ðŒâðð) = 0.So ðŒ â ðð has a dense range. Since ðŒ âðð is a Fredholm operator of index zero,it follows that ðŒ â ðð is invertible.
Part 5. In this part we show that condition (3) with ðâ20 invertible implies thatðŒ âðð is invertible. Define ð and ð by (4.9). As we have seen in the first part ofthe proof, ðŒ + ð ððµ2 = ð
â20, and hence ðŒ + ð ððµ2 is invertible. From Theorem 1.3
we conclude that
Ker (ðŒ â ðð) âââ©ð=0
Kerð¶2ðŽð2ð.
Assume that (ðŒâðð)ð¥ = 0. Then, by the previous identity ð¶2ðŽð2ðð¥ = ð¶2ðŽ
ð2ð¥ = 0.
Using the definitions of ðŽ2 and ð¶2, we obtain ðâðð
â1ð»ððð ð¥ = 0 for ð = 0, 1, 2, . . ..
As above in Part 3 write ð = ðâð+. Then
0 = ðâððâ1+ ð â1â ð»ðð
ð ð¥ = ðâðð
â1+ ð â1â
(ðâð)ðð»ð¥ = ð£+0ð
âð
(ðâð)ðð â1â ð»ð¥.
We see that ð»ð¥ = 0. But then ð ð¥ = 0 and ð¥ = 0. As above we conclude thatð âðºð â1ð» is invertible. So ðŒ â ðð is invertible.
Part 6. In this part we show that condition (3) with ðâ20 invertible that ðŒâðð isinvertible. Define ð and ð by (4.9). As we have seen in the first part of the proofðŒ + ðððµ1 = ð
â20, and hence ðŒ + ðððµ1 is invertible. Assume that (ðŒ â ðð)ð¥ = 0.
According to Theorem 3.6 we have that ð¶1ðŽð1ðð¥ = 0 for ð = 0, 1, 2, . . .. So
ððððððð¥ = 0 for all ð. But then ðð¥ = 0, and we conclude that ð¥ = 0. Hence
ðŒ â ðð is invertible.
176 M.A. Kaashoek and F. van Schagen
Part 7. Finally we apply Theorem 1.2 to show that the inverse of ð âðºð â1ð» isgiven by (1.17) whenever one of the conditions (1)â(5) is satisfied. Formula (1.17)translates to (4.6), and hence
(ð âðºð â1ð»)â1 â ðð(ð âðºð â1ð»)â1ðâð= ð1ð
â110 ð
â2 â ððð â1ðºð1ðâ110 ð
â2ð»ð
â1ðâð .
The identity (4.7) one obtains by just switching the roles of ð and ð , ðº and ð» ,ð and ð, ðð0 and ðð0. â¡
We now are ready to prove Theorem 1.1. Recall that for a linear map ðŒ :âð â â2+(â
ð) we denote by ððŒ theðÃð block lower triangular Toeplitz operatorwith first column equal to ðŒ. For an ðÃð matrix ð¢ the symbol Îð¢ denotes theblock diagonal operator on â2+(â
ð) with all diagonal entries equal to ð¢.
Proof of Theorem 1.1. First we prove (1.5). Use Proposition 4.1 to derive the iden-tity (4.6). By multiplying this identity ðâ 1 times from the left by ðð and ðâ 1times from the right by ðâð , and adding the resulting identities one gets (also using(1.7))
(ð âðºð â1ð»)â1 â ððð (ð âðºð â1ð»)â1(ðâð )ð
=
ðâ1âð=0
ðððð1ð
â110 ð
â2(ð
âð)
ð â ðð(
ðâ1âð=0
ððð ð1ð
â110 ð
â2(ð
âð)
ð
)ðâð .
Since for any â we have limðââ(ðâð)ðâ = 0, the left-hand side converges pointwise
to (ð âðºð â1ð»)â1. Notice thatðâ1âð=0
ðððð1ð
â110 ð
â2(ð
âð)
ð = ðð1Îðâ110Î ðð
âð2,
where Î ð = ðŒ â ððð (ð
âð)
ð, and limðââ ðð1Îðâ110Î ðð
âð2= ðð1Îðâ1
10ð âð2
. Similarly,
using that ð1, ð2 : âð â â1+(â
ð), one finds
ââð=0
ððð ð1ð
â110 ð
â2(ð
âð)
ð = ðð1Îðâ110ð âð2 .
So we get
(ð âðºð â1ð»)â1 = ðð1Îðâ110ð âð2
â ðððð1Îðâ110ð âð2ð
âð .
We proved (1.5). Formula (1.6) one obtains in a similar way from (4.7). â¡
It is interesting to specify Theorem 1.1 for the case when ðº = 0, ð» = 0 andð = ðŒ. Then ð â ðºð â1ð» = ð . Recall that ð is assumed to be invertible. Thehypotheses ðº = 0, ð» = 0 and ð = ðŒ imply that (1.3) and (1.4) reduce to
ð1 = ð â1ðð, ð1 = ðð, ðâ2 = ð
âðð â1, ðâ2 = ð
âð .
In particular, ð1 is the first column of ð â1, ðâ2 is the first row of ð
â1, and ð10 = ðâ20is the (1, 1) entry of ð â1. Since we assume that the entries of the matrix function
Inverting Structured Operators 177
defining ð belong to the Wiener algebra on the circle, the classical result from [9]then tells us that ð10 is invertible and
ð â1 = ðð1Îðâ110ð âð2. (4.10)
The above formula for ð â1 is precisely (1.5) for the case when ðº and ð» are zero.Indeed, when ðº and ð» are zero, then (1.7) tells us that the matrices ð1 and ð2 arezero. But in that case (1.5) reduces to (4.10).
4.2. Finite rank Hankel operators
In this subsection we return to the case when ðº and ð» are of finite rank andð = ðŒð and ð = ðŒð. We shall derive the identity (1.21). To do this we use the dataappearing in (1.18), (1.19) and (1.20).
Proof of (1.21). Assume there exist linear maps ð1, ð2, ð1, ð2 as in (1.2) satisfyingequations (1.3) and (1.4) (with ð and ð identity operators) such that ð10 or ð10is invertible. As Theorem 1.1 tells us, this implies that ðŒ â ðºð» is invertible. Weintend to apply Theorem 1.2 with the data as in (1.18), (1.19) and (1.20). Firstone checks that the operators
ð = Î1ð1, ð = ðâ2Î1, ð = Î2ð1, ð = ðâ2Î2 (4.11)
satisfy the identities (1.13) and (1.14). This allows us to prove the following iden-tities:
ðŒ + ð¶2ðð = ð10, ðŒ + ð¶1ðð = ð10, (4.12)
ðŒ + ðððµ1 = ðâ20, ðŒ + ð ððµ2 = ð
â20. (4.13)
To see this let us establish the first identity in (4.12):
ðŒ + ð¶2ðð = ðŒ + ðâðÎ2Î2Î1Î1ð1 = ðŒ + ðâððºð»ð1
= ðŒ + ðâð(ð1 â (ðŒ âðºð»)ð1) = ð10.The other identities in (4.12) and (4.13) are proved in a similar way. It follows(see Lemma 2.1) that ð10 = ð
â20 and ð10 = ð
â20. Since we assume that one of the
matrices ð10 and ð10 is invertible, both are invertible, and we get from (1.17),(4.11), (4.12), and (4.13) that
ð(ðŒ â ðð)â1 âðŽ2ð(ðŒ â ðð)â1ðŽ1
= Î2ð1ðâ110 ð
â2Î1 âðŽ2Î2Î1Î1ð1ð
â110 ð
â2Î2Î2Î1ðŽ1
= Î2ð1ðâ110 ð
â2Î1 âðŽ2Î2ð»ð1ð
â110 ð
â2ðºÎ1ðŽ1.
The fact that both ðŽ1 and ðŽ2 are stable then yields:
ð(ðŒ â ðð)â1 =ââð=0
ðŽð2Î2ð1ð
â110 ð
â2Î1ðŽ
ð1 â
ââð=0
ðŽð+12 Î2ð»ð1ð
â110 ð
â2ðºÎ1ðŽ
ð+11 .
178 M.A. Kaashoek and F. van Schagen
Using (ðŒ âðºð»)â1 = ðŒ + Î2ð(ðŒ â ðð)â1Î1, we find
(ðŒ âðºð»)â1 = ðŒ +ââð=0
Î2ðŽð2Î2ð1ð
â110 ð
â2Î1ðŽ
ð1Î1
âââð=0
Î2ðŽð+12 Î2ð»ð1ð
â110 ð
â2ðºÎ1ðŽ
ð+11 Î1.
Next put ð1 = âðºð1 and ðâ2 = âðâ2ð» ; cf. the identities in (1.7). ThenÎ2ðŽ
ð2Î2ð1 = (ðâð)
ð(âð1), ðâ2Î1ðŽð1Î1 = âðâ2ðð
ð (ð ⥠0).
Also use
Î2ðŽð+12 Î2ð»ð1 = (ðâð)
ð+1ðºð»ð1 = (ðâð)ð+1(ð1 â ðð) = (ðâð)
ð+1ð1,
ðâ2ðºÎ1ðŽð+11 Î1 = ð
â2ðºð»ð
ð+1ð = (ðâ2 â ðâð)ðð+1
ð = ðâ2ðð+1ð .
In this way we obtain
(ðŒ âðºð»)â1 = ðŒ +ââð=0
(ðâð)ðð1ð
â110 ð
â2ð
ðð â
ââð=0
(ðâð)ð+1ð1ð
â110 ð
â2ð
ð+1ð
= ðŒ +ð»ð1Îðâ110ð»âð2 â ðâðð»ð1Îðâ1
10ð»âð2
ðð.
(4.14)
This proves (1.21). â¡
Remark. Notice that all the Hankel operators in (4.14) are of finite rank. Forexample, the identity ð»ð1 = âÎ2
[ð ðŽ2ð ðŽ2
2ð â â â ] implies that the rankof ð»ð1 is at most the order of ðŽ2.
Formula (1.21) remains true in the more general case when ðº and ð» are offinite rank. To see this, assume there exist linear maps ð1, ð2, ð1, ð2 as in (1.2)satisfying equations (1.3) and (1.4) (with ð and ð identity operators) such thatð10 or ð10 is invertible. Hence, by Theorem 1.1, the operator ðŒ âðºð» is invertibleTo derive the analogue of the formula (1.21) for this case we apply Theorem 1.2.Put
ð = ð», ð = ðº,
ðŽ1 = ðâð , ðµ1 = ð»ðð, ð¶1 = ð
âð,
ðŽ2 = ðâð , ðµ2 = ðºðð, ð¶2 = ð
âð .
The corresponding Stein equations (1.11) are satisfied, and
ð = ð»ð1, ð = ðºð1, ð = ðâ2, and ð = ðâ2solve the equations (1.13) and (1.14). Notice that the equalities (4.12), (4.13) holdtrue. Furthermore, ð10 = ð
â20 and ð10 = ð
â20. Theorem 1.2 yields
ðº(ðŒ âð»ðº)â1 â ðâððº(ðŒ âð»ðº)â1ðâð = ðºð1ðâ110 ðâ2 â ðâððºð»ð1ðâ110 ð
â2ðºð
âð .
Since ðâððºð»ð1 = ðâð(ð1 â ðð) = ðâðð1 and ðºð1 = âð1 and ðâ2ðº = âðâ2, we get
ðº(ðŒ âð»ðº)â1 â ðâððº(ðŒ âð»ðº)â1ðâð = âð1ðâ110 ðâ2 + ð
âðð1ð
â110 ð
â2ðâð .
Inverting Structured Operators 179
Using the same reasoning as in the proof of Theorem 1.1 we obtain
ðº(ðŒ âð»ðº)â1 = âð»ð1Îâ1ð10ð âð2
+ ðâðð»ð1Îâ1ð10ð âð2ð
âð . (4.15)
To derive from (4.15) a formula for (ðŒ âðºð»)â1, we use the identity(ðŒ âðºð»)â1 = ðŒ +ðº(ðŒ âð»ðº)â1ð». (4.16)
Thus we have to multiply (4.15) from the right by ð» . For this purpose we use thefollowing lemma.
Lemma 4.2. Let ð¥ : âð â â1+(âð), and let ðŸ : â1+(â
ð) â â1+(âð ) be a Hankel
operator of which the defining matrix function has entries in the Wiener algebra.Then
ðŸðð¥ = ð»ðŸð¥ : â1+(â
ð)â â1+(âð ).
Here ðð¥ is the lower triangular Toeplitz operator with first column ð¥ and ð»ðŸð¥ isthe Hankel operator with with first column ðŸð¥.
Proof. Note that ðŸð¥ : âð â â1+(âð ). The ðth column of ðŸðð¥ is ðŸð
ððð¥ and the
ðth column of ð»ðŸð¥ is (ðâð )
ððŸð¥. These are equal since ðŸððð = (ðâð )
ððŸ. â¡
From this lemma we see that ð âð2ð» = ð»â
ð»âð2. With as before ðâ2 = âðâ2ð» it
follows that ð âð2ð» = âð»âð2 . Next we consider ð âð2ðâðð» . The dual of this operator
is ð»âðððð2 = ð»âðððð2 = ð»ð»âððð2 . Note that
ðâ2ðâðð» = âðâ2ðºð»ðð = (âðâ2 + ðâð)ðð = âðâ2ðð,
to see that ð»âðððð2 = âð»ðâðð2 = âðâðð»ð2 . We get ð
âð2ð
âðð» = âð»âð2
ðð, and
therefore (4.15) and (4.16) together yield
(ðŒ âðºð»)â1 = ðŒ +ð»ð1Îâ1ð10ð»âð2 â ðâðð»ð1Î
â1ð10ð»âð2
ðð, (4.17)
which is (1.21) for the present case.
Remark. Using (4.16) the identity (4.17) also follows from (1.6) with ð = ðŒ andð = ðŒ . Indeed, using (4.16) and (1.6) we obtain
(ðŒ âðºð»)â1 = ðŒ +ðºðð1Îâ1ð10ð âð2ð» âðºðððð1Îâ1ð10
ð âð2ðâðð».
Applying Lemma 4.2 yields the identities:
ðºðð1 = âð»ð1 , ðºðððð1 = âðâðð»ð1 , ðâð2ð» = âð»âð2 , ð âð2ðâðð» = âð»âð2
ðð.
Using these identities one obtains (4.17).
In the next section we shall prove the analogue of formula (4.17) in the generalsetting of Theorem 1.1.
180 M.A. Kaashoek and F. van Schagen
4.3. Hankel type formulas for the inverse of Toeplitz plus Hankel
In this section we derive formulas for the inverses of ð âðºð â1ð» and ð âð»ð â1ðºthat generalize formula (1.21) (and (4.17)). We begin with some notation.
Let ð be the defining function of ð , and ð the one of ð . Since R and V areinvertible, detð and detð have no zero on the unit circle, and hence ðâ1 and ðâ1
are well defined matrix functions. Moreover, the entries of ðâ1 and ðâ1 belong tothe Wiener algebra on the unit circle. We denote by ð Ã and ð Ã the block Toeplitzoperators defined by ðâ1 and ðâ1, respectively.
Theorem 4.3. Assume there exist linear maps ð1, ð2 and ð1, ð2 as in (1.2) satisfy-ing equations (1.3) and (1.4). Then ð10 = ð
â20 and ð10 = ð
â20. Furthermore, assume
that at least one of the matrices ð10 and ð10 is invertible. Then both the matricesð10 and ð10 are invertible, and the operators ð â ðºð â1ð» and ð â ð»ð â1ðº areinvertible. Moreover,
(ð âðºð â1ð»)â1 = ð à +ð»ð1Îâ1ð10ð»âð2 â ðâðð»ð1Î
â1ð10ð»âð2
ðð, (4.18)
(ð âð»ð â1ðº)â1 = ð à +ð»ð1Îâ1ð10ð»âð2 â ðâðð»ð1Î
â1ð10ð»âð2
ðð. (4.19)
Here
ð1 = âð â1ðºð1, ð1 = âð â1ð»ð1, ðâ2 = âðâ2ð»ð â1, ðâ2 = âðâ2ðºð â1.Proof. As in the proof of Theorem 1.1, we conclude from Proposition 4.1 that theidentity (4.6) holds. Multiply this identity from the left by ðâð and from the rightby âðð. This yields (cf. the first identity in (5.16) of [6]):
(ð âðºð â1ð»)â1 â ðâð(ð âðºð â1ð»)â1ðð = ð1ðâ110 ðâ2 â ðâðð1ðâ110 ð
â2ðð. (4.20)
Next, by ð â 1 times repeatedly multiplying (4.20) from the left by ðâð and fromthe right by ðð, and adding the resulting identities, we get
(ð âðºð â1ð»)â1 â (ðâð)ð(ð âðºð â1ð»)â1ðð
ð
=
ðâ1âð=0
(ðâð)ðð1ð
â110 ð
â2ð
ðð â
ðâ1âð=0
(ðâð)ððâðð1ð
â110 ð
â2ððð
ðð .
(4.21)
In order to determine for the terms in (4.21) the limits for ð going to infinity, wewill deal with each of these terms separately.
We begin with the first term on the right-hand side. Let Î ð be the projectionof â2+(â
ð) mapping (ð¥0, ð¥1, ð¥2, . . .) onto (ð¥0, . . . , ð¥ðâ1, 0, 0, . . .). Then
ðâ1âð=0
(ðâð)ðð1ð
â110 ð
â2ð
ðð = ð»ð1Î
â1ð10Î ðð»
âð2 .
Since ð»âð2 is compact, limðââÎ ðð»âð2= ð»âð2 , with convergence in the operator
norm, and henceââð=0
(ðâð)ðð1ð
â110 ð
â2ð
ðð = ð»ð1Îðâ1
10ð»âð2 . (4.22)
Inverting Structured Operators 181
Similarly one derives that
ââð=0
(ðâð)ððâðð1ð
â110 ð
â2ððð
ðð = ð
âðð»ð1Î
â1ð10ð»âð2
ðð. (4.23)
Next we proceed with the left-hand side of the identity (4.21). First re-mark that the convergence of the right-hand side yields the existence of the limitlimðââ(ðâð)ð(ð âðºð â1ð»)â1ðð
ð . We will prove
limðââ(ð
âð)
ð(ð âðºð â1ð»)â1ððð = ð
Ã. (4.24)
The proof of (4.24) will be based on the following two observations: (a) the operator(ð âðºð â1ð»)â1 â ð â1 is compact, and (b) the operator ð â1 â ð à is compact.Note that (a) follows from
(ð âðºð â1ð»)â1 = ð â1 +ð â1ðºð â1ð»(ð âðºð â1ð»)â1.Indeed, using the latter identity and ðº (or ð») is compact, we get item (a). Toget item (b), we use that ð is invertible, and hence the defining function ð of ð admits a canonical factorizations, ð = ðâð+. Recall that ð à is the block Toeplitzoperator defined by ðâ1 = ðâ1+ ð
â1â and ð â1 = ð Ã+ð
Ãâ, where ð
Ã+ and ð Ãâ are the
block Toeplitz operators defined by ðâ1+ and ðâ1â , respectively. But then, using astandard identity (see formula (4) in [7, Section XXIII]), we see that ð Ãâð â1 isthe product of two compact Hankel operators, which proves item (b).
Given items (a) and (b) we see that (ð â ðºð â1ð»)â1 = ð à +ðŸ, where ðŸis a compact operator. Since (ðâð)
ð â 0 pointwise, the compactness of ðŸ impliesthat (ðâð)ððŸ â 0 in operator norm, and thus (ðâð)ððŸðð
ð â 0 in operator norm as
ðââ. The fact that ð Ã is a block Toeplitz operator is equivalent to ðâðð Ãðð =
ð Ã, and hence (ðâð)ðð Ãðð
ð = ð Ã for each ð. We conclude that the left-hand side
of (4.24) is equal to
limðââ(ð
âð)
ðð Ãððð + lim
ðââ(ðâð )
ððŸððð = ð
Ã.
This proves (4.24). Combining the limits (4.22), (4.23), and (4.24) we obtain (4.18).
The proof of (4.19) can be done in exactly the same manner. â¡
5. New proof of the main inversion theorem in [6]
In this section we show how Theorem 2.1 in [6] can be obtained from Theorem 1.2.
We begin with some preliminaries. Let ð³ be a Hilbert space with two or-thogonal direct sum decompositions:
ð³ = ð°1 â ðŽ1 = ð°2 â ðŽ2. (5.1)
182 M.A. Kaashoek and F. van Schagen
On ð³ we have two operators ðŽ and ðŸ such that relative to these decompositions
ðŽ =
[ðŒ1 00 ðŒ2
]:
[ð°1ðŽ1
]â[ð°2ðŽ2
], where ðŒ2 is invertible; (5.2)
ðŸ =
[ð 1 00 ð 2
]:
[ð°2ðŽ2
]â[ð°1ðŽ1
], where ð 2 is invertible and ð 2 = ðŒ
â12 . (5.3)
In what follows we denote by ðâ the orthogonal projection of ð³ onto the subspaceâ, viewed as an operator from ð³ to â. Furthermore, ðâ denotes the canonicalembedding of â into ð³ , that is, ðâ = ðââ. The following result is Theorem 2.1in [6].
Theorem 5.1. Let ð be an invertible operator on ð³ and let ðŽ and ðŸ be as in (5.2)and (5.3), respectively. Assume that
ððŽ1(ð âðŸððŽ)ððŽ1 = 0. (5.4)
Consider the operators defined by
Î = ðâ1ðð°2 : ð°2 â ð³ , Κ = ðâ1ðð°1 : ð°1 â ð³ ,Î¥ = ðð°2ð
â1 : ð³ â ð°2, Ω = ðð°1ðâ1 : ð³ â ð°1.
Furthermore, put ð0 = ðð°2Î and ð0 = ðð°1Κ. Then ð0 is invertible if and only ifð0 is invertible, and in this case the inverse of ð satisfies the identity
ðâ1 âðŽðâ1ðŸ = Îðâ10 Î¥âðŽÎšðâ10 ΩðŸ. (5.5)
First we state and prove a preliminary lemma.
Lemma 5.2. Let ð be an invertible operator on ð³ and let ðŽ and ðŸ be as in (5.2)and (5.3), respectively. Let
ðŽ0 =
[0 00 ðŒ2
]:
[ð°1ðŽ1
]â[ð°2ðŽ2
], ðŸ0 =
[0 00 ð 2
]:
[ð°2ðŽ2
]â[ð°1ðŽ1
].
Then (5.5) holds true if and only if
ðâ1 âðŽ0ðâ1ðŸ0 = Îðâ10 Î¥âðŽ0Κð
â10 ΩðŸ0.
Proof. It is sufficient to prove that
ðŽðâ1ðŸ â ðŽ0ðâ1ðŸ0 = ðŽÎšð
â10 ΩðŸ âðŽ0Κð
â10 ΩðŸ0.
Write
ðŽðâ1ðŸ âðŽ0ðâ1ðŸ0 = (ðŽâðŽ0)ð
â1ðŸ +ðŽ0ðâ1(ðŸ âðŸ0),
ðŽÎšðâ10 ΩðŸ âðŽ0Κðâ10 ΩðŸ0 = (ðŽâðŽ0)Κð
â10 ΩðŸ +ðŽ0Κð
â10 Ω(ðŸ âðŸ0).
So it suffices to prove
(ðŽâðŽ0)ðâ1ðŸ = (ðŽâðŽ0)Κð
â10 ΩðŸ (5.6)
and
ðŽ0ðâ1(ðŸ âðŸ0) = ðŽ0Κð
â10 Ω(ðŸ âðŸ0). (5.7)
Inverting Structured Operators 183
Write
ðâ1 =[ð0 ð12ð21 ð22
]:
[ð°1ðŽ1]â[ð°1ðŽ1].
Then
Κ = ðâ1ðð°1 =
[ð0ð21
], and Ω = ðð°1ð
â1 =[ð0 ð12
].
A straightforward computation with these matrix representations reveals that (5.6)and (5.7) hold true. â¡
Proof of Theorem 5.1. In view of Lemma 5.2 we may assume that ðŒ1 = 0 andð 1 = 0. Let ð³1 = ð³2 = ð³ , ðŽ = ð°2, ð° = ð°1 â ð°1 and
ð = ðŒð³ â ð, ð = ðŒð³ ,
ðŽ1 = ðŸ, ðµ1 =[ðð°1 âððð°1 + ðð°1(ðŒð°1 + ðð°1ððð°1)
], ð¶1 = ðð°2 ,
ðŽ2 = ðŽ, ðµ2 = ðð°2 , ð¶2 =
[âðð°1ððð°1
].
First we present a few simple auxiliary identities that will play a role in thesequel:
ðð°1ðŽ1 = ðð°1ðŸ = 0, ðŽ2ðð°1 = ðŽðð°1 = 0,
ðŒð³ âðŽ2ðŽ1 = ðŒð³ âðŽðŸ = ðð°2ðð°2 , ðŒð³ = ððŽ1ððŽ1 + ðð°1ðð°1 .
Using these equalities and (5.4), we obtain
ð âðŸððŽ = (ððŽ1ððŽ1 + ðð°1ðð°1)(ð âðŸððŽ)(ððŽ1ððŽ1 + ðð°1ðð°1)
= ðð°1ðð°1ð + ððð°1ðð°1 â ðð°1ðð°1ððð°1ðð°1 .
So it follows that
ð âðŽ1ððŽ2 = ðŒð³ âðŸðŽâ (ð âðŸððŽ)= ðð°1ðð°1 â ðð°1ðð°1ð â ððð°1ðð°1 + ðð°1ðð°1ððð°1ðð°1 = ðµ1ð¶2.
Furthermore,
ðâðŽ2ððŽ1 = ðŒð³ âðŽðŸ = ðð°2ðð°2 = ðµ2ð¶1.
Next let us define
ð =[Κ âðð°1 +Κ(ðŒð°1 + ðð°1ððð°1)
], ð = Î¥
ð = Î, ð =
[âðð°1
Ω
].
Then
ðð =[ðð°1 âððð°1 + ðð°1(ðŒð°1 + ðð°1ððð°1)
]= ðµ1,
ð ð = ð¶1, ðð = ðµ2, ðð =
[âðð°1ððð°1
]= ð¶2.
184 M.A. Kaashoek and F. van Schagen
We proceed with ðŒð° + ð¶2ðð and ðŒðŽ + ð ððµ2. First
ðŒðŽ + ð ððµ2 = ðŒð°2 + ð (ðŒð³ â ð )ðð°2 = ðŒð°2 + ð ðð°2 â ð ððð°2
= ðŒð°2 + ðð°2ðâ1ðð°2 â ðð°2ðð°2 = ð0.
Next
ðŒð° + ð¶2ðð =
[ðŒð°1 00 ðŒð°1
]+
[âðð°1ððð°1
] [Κ âðð°1 +Κ(ðŒð°1 + ðð°1ððð°1)
].
We compute the four entries of this 2 à 2 matrix separately. The (1, 1)-entry isðŒð°1 â ðð°1ðΚ = ðŒð°1 â ðð°1ðð°1 = 0. The (1, 2)-entry is
ðð°1ððð°1 â ðð°1ðΚ(ðŒð°1 + ðð°1ððð°1) = âðŒð°1 .
For the (2, 1)-entry ðð°1Κ = ðð°1ðâ1ðð°1 = ð0. Finally the (2, 2)-entry is given by
ðŒð°1 â ðð°1ðð°1 + ðð°1Κ(ðŒð°1 + ðð°1ððð°1) = ðð°1Κ(ðŒð°1 + ðð°1ððð°1).
We have
ðŒð° + ð¶2ðð =
[0 ðŒð°1
ð0 ð0(ðŒð°1 + ðð°1ððð°1)
],
which is invertible if and only if ð0 is invertible. Our assumption was that ð0 orð0 is invertible. So we have that ðŒð° + ð¶2ðð or ðŒð°2 + ð ððµ2 is invertible andhence, according to Theorem 1.2 both are invertible and formula (1.17) holdstrue. The left-hand side of this formula is ðâ1 â ðŽðâ1ðŸ since we have thatð(ðŒð³ â ðð)â1 = ðâ1. To finish the proof we have to check that the right-handside of (1.17) gives the right-hand side of (5.5). For the first term this is immediatefrom the above-established equalities ð = Î, ðŒð°2 + ð ððµ2 = ð0 and ð = Î¥. Forthe second term first notice that
ðŽ2ðð(ðŒð° + ð¶2ðð)â1ðððŽ1
= ðŽ[Κ âðð°1 +Κ(ðŒð°1 + ðð°1ððð°1)
] [â(ðŒð°1 + ðð°1ððð°1) ðâ10
ðŒð°1 0
] [âðð°1
Ω
]ðŸ.
Use ðŽðð°1 = 0 and ðð°1ðŸ = 0 to see that this is equal to ðŽÎšðâ10 ΩðŸ. We provedformula (5.5). â¡
6. Examples
In this section ð is an invertible ð à ð block Toeplitz matrix with blocks of sizeðà ð, and we take ð = ðŒð³ â ð and ð = ðŒð³ , where ð³ = (âð)ð. With these ð andð we shall associate two different sets of matrices {ðŽ1, ðµ1, ð¶1;ðŽ2, ðµ2, ð¶2} suchthe equations (1.9), (1.10), and (1.11) are satisfied. For both choices we compute(1.17). We shall see that for the first choice (1.17) cannot be used to obtain aformula for ðâ1, while for the second choice (1.17) leads to the GohbergâHeinigformula [8] for ðâ1.First example. To introduce our first choice for the set (ðŽ1, ðµ1, ð¶1;ðŽ2, ðµ2, ð¶2)associated with ð and ð we need some auxiliary operators. We define the forward
Inverting Structured Operators 185
block shift ð : (âð)ð â (âð)ð and the two embedding operators ð : âð â (âð)ð
and ð : âð â (âð)ð by
ð
â¡â¢â¢â¢â£ð¥0ð¥1...
ð¥ðâ1
â€â¥â¥â¥âŠ =â¡â¢â¢â¢â£0 0 â â â 0ðŒð 0 0
. . .. . .
...ðŒð 0
â€â¥â¥â¥âŠâ¡â¢â¢â¢â£ð¥0ð¥1...
ð¥ðâ1
â€â¥â¥â¥âŠ , ðð¥ =
â¡â¢â¢â¢â£ð¥0...0
â€â¥â¥â¥âŠ , ðð¥ =
â¡â¢â¢â¢â£0...0ð¥
â€â¥â¥â¥âŠ ,for each ð¥ â âð. Let ð³ = ð³1 = ð³2 = (âð)ð, ð° = (âð)2 and ðŽ = âð. PutðŽ = ð + ððâ and
ðŽ1 = ðŽ, ðµ1 =[âðð+ ðŽ1ðð âð] , ð¶1 = 0,
ðŽ2 = ðŽâ = ðŽâ1, ðµ2 = 0, ð¶2 =
[ðâ
ðâð â ðâððŽ2
].
Then ð â ðŽ2ððŽ1 = ðµ2ð¶1 and ð â ðŽ1ððŽ2 = ðµ1ð¶2. The latter equality one cancheck as follows. First note that ð âðððâ = ðððâ + ððâð â ððâðððâ. It followsthat
â(ð âðŽ1ððŽ2) = ð âðŽ1ððŽ2 = ð â (ð + ððâ)ð (ðâ + ððâ)
= ðððâ + ððâð â ððâðððâ â ððâððŽ2 âððððâ= ðððâ + ððâð â ððâðððâ â ððâððŽ2 âððððâ= ðððâ + ððâð âðŽ1ððð
â â ððâððŽ2
= âðµ1ð¶2.
Recall that ð is assumed to be invertible. Define ð¥, ð€, ð§â and ðŠâ by
ðð¥ = ð, ðð€ = ðŽ1ðð, ð§âð = ðâ, ðŠâð = ðâððŽ2.
Then
ð =[âð+ ð€ âð¥] , ð = 0, ð = 0, ð =
[ð§â
ðâ â ðŠâ]
solve (1.13) and (1.14) and
ðŒð° + ð¶2ð =
[ðâð€ âðâð¥
(ðâ â ðŠâ)(ðð€ â ðð) ðŠâð
]Next notice that ðŒðŽ + ð ðð = ðŒðŽ is invertible. Thus Theorem 1.2 tells us thatðŒð° + ð¶2ð is invertible and
ðâ1 âðŽ2ðâ1ðŽ1 = ðŽ2ð(ðŒ + ð¶2ð)
â1ððŽ1. (6.8)
Note that in this case ðŽ1 and ðŽ2 are not stable, and it is not clear how one canuse (6.8) to derive a formula for ðâ1.
186 M.A. Kaashoek and F. van Schagen
Second example. The spaces ð³ , ð³1, ð³2, ð° , and ðŽ are as in the previous example.We choose
ðŽ1 = ðâ, ðµ1 =
[ð âðð + ðð¡0
], ð¶1 = ð
â,
ðŽ2 = ð, ðµ2 = ð, ð¶2 =
[ðâ â ðâððâ
],
where ð¡0 is the left upper entry of ð . Then the equations in (1.11) are satisfied.Using ð is invertible, define ð¥, ðŠ, ð€ and ð§ by
ðð¥ = ð, ð ð§ = ð, ðŠâð = ðâ, ð€âð = ðâ.
To satisfy (1.13) and (1.14) put
ð = ð¥, ð = ðŠâ, ð =[ð§ âð + ð§ð¡0
], ð =
[ð€â â ðâð€â
].
Then ðŒðŽ + ð ððµ2 = ðŠâðð and
ðŒð° + ð¶2ðð =
[ðâð§ âðŒ + ðâð§ð¡0ðâð§ ðâð§ð¡0
].
Now write for short ðŠ0 = ðŠâðð and ð§ð = ðâð§. So ðŒð° + ð¶2ðð is invertible if and
only if ð§ð is. A simple computation gives in this case that
ðâ1 âððâ1ðâ = ð¥ðŠâ10 ðŠâ âðð§ð§â1ð ð€âðâ. (6.9)
This is a well-known formula for the inverse of ð from [8]. In this case ðð = 0 and(ðâ)ð = 0. Hence one can easily derive from (6.9) a closed expression for ðâ1.
References
[1] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran, Factorization of matrix andoperator functions: the state space approach, OT 178, Birkhauser Verlag, Basel, 2008.
[2] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran, A state space approach tocanonical factorization with applications, OT 200, Birkhauser Verlag, Basel, 2010.
[3] M.J. Corless and A.E. Frazho, Linear systems and control, Marcel Dekker, Inc., NewYork, 2003.
[4] R.L. Ellis and I. Gohberg, Orthogonal systems and convolution operators, OT 140,Birkhauser Verlag, Basel, 2003.
[5] R.L. Ellis, I. Gohberg, and D.C. Lay, Infinite analogues of block Toeplitz matricesand related orthogonal functions, Integral Equations and Operator Theory 22 (1995),375â419.
[6] A.E. Frazho and M.A. Kaashoek, A contractive operator view on an inversion for-mula of GohbergâHeinig, in: Topics in Operator Theory I. Operators, matrices andanalytic functions, OT 202, Birkhauser Verlag, Basel, 2010, pp. 223â252.
[7] I. Gohberg, S. Goldberg, and M.A. Kaashoek, Classes of Linear Operators, VolumeII, OT 63, Birkhauser Verlag, Basel, 1993.
Inverting Structured Operators 187
[8] I. Gohberg, G. Heinig, The inversion of finite Toeplitz matrices consisting of elementsof a non-commutative algebra, Rev. Roum. Math. Pures et Appl. 20 (1974), 623â663 (in Russian); English transl. in: Convolution Equations and Singular IntegralOperators, (eds. L. Lerer, V. Olshevsky, I.M. Spitkovsky), OT 206, Birkhauser Verlag,Basel, 2010, pp. 7â46.
[9] I.C. Gohberg and M.G. Krein, Systems of integral equations with kernels dependingon the difference of arguments, Uspekhi Math. Nauk 13 2(80) (1958), 3â72 (Russian);English Transl., Amer. Math. Soc. Transl. (Series 2) 14 (1960), 217â287.
[10] G.J. Groenewald and M.A. Kaashoek, A GohbergâHeinig type inversion formulainvolving Hankel operators,in: Interpolation, Schur functions and moment problems,OT 165, Birkhauser Verlag, Basel, 2005, pp. 291â302.
[11] G. Heinig and K. Rost, Algebraic methods for Toeplitz-like matrices and operators,Akademie-Verlag, Berlin, 1984.
[12] T. Kailath and A.H. Sayed, Displacement structure: Theory and applications, SIAMRev. 37 (1995), 297â386.
[13] T. Kailath and A.H. Sayed (editors), Fast Reliable Algorithms for Matrices withStructure, SIAM, Philadelphia, 1999.
[14] I. Koltracht, B.A. Kon, and L. Lerer, Inversion of structured operators, Integralequations and Operator Theory 20 (1994), 410â448.
[15] V. Olshevsky (editor), Structured matrices in mathematics, Computer Science, andEngineering, Contempary Math. Series 280, 281, Amer. Math. Soc. 2001.
[16] V.Y. Pan, Structured matrices and polynomials, Birkhauser Boston, 2001.
M.A. Kaashoek and F. van SchagenDepartment of MathematicsFaculty of SciencesVU UniversityDe Boelelaan 1081aNL-1081 HV Amsterdam, The Netherlandse-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 189â196câ 2013 Springer Basel
On the Sign Characteristics ofSelfadjoint Matrix Polynomials
Peter Lancaster and Ion Zaballa
Dedicated to Leonid Lerer on the occasion of his seventieth birthday.
Abstract. An important role is played in the spectral analysis of selfadjointmatrix polynomials by the so-called âsign characteristicsâ associated with realeigenvalues. In this paper the ordering of the real eigenvalues by their signcharacteristics is clarified. In particular, the roles played by the signature ofthe leading and trailing polynomial coefficients are discussed.
Mathematics Subject Classification (2010). 15A21, 47B15.
Keywords. Matrix polynomial. Sign characteristics.
1. Introduction
Let ð¿0, ð¿1, . . . , ð¿â â âðÃð. We consider matrix polynomials:
ð¿(ð) := ð¿âðâ + ð¿ââ1ðââ1 + â â â + ð¿0, ð â â, detð¿â â= 0. (1)
Such a polynomial is said to be selfadjoint if the coefficients are either complexHermitian or, in particular, real and symmetric. The eigenvalues of ð¿ are the zerosof detð¿(ð), and the eigenfunctions are the real analytic functions on â formed bythe zeros of detð¿(ð); say ð1(ð), ð2(ð), . . . , ðð(ð) (in some order to be decided). Inthis way the eigenvalues can also be characterized as the zeros of the eigenfunctionsand, if ð0 is an eigenvalue of ð¿(ð), then dimKerð¿(ð0) is exactly the number ofeigenfunctions that annihilate at ð0.
The notion of âsign characteristicâ associated with a real eigenvalue playsan important role in the spectral analysis and perturbation theory of selfadjointmatrix polynomials; see [1], [2], and [3], for example. In particular, it should be
The first author was supported in part by the Natural Sciences and Engineering Research Councilof Canada.
The second author was supported in part by MICINN MTM2010-19356-C02-01, EJ GIC10/169-IT-361-10 and UPV/EHU UFI11/52.
190 P. Lancaster and I. Zaballa
noted that, for convenience, and because of many applications, it was assumedin [3] that ð¿â > 0. Here, the more general case of nonsingular ð¿â prevails as in(the more comprehensive) references [1] and [2]. In particular, we will need thefollowing fundamental result (Theorem 3.7 of [2] and Theorem 6.10 of [1]).
Theorem 1.1. Let ð¿(ð) be an ðÃð selfadjoint matrix polynomial with nonsingularleading coefficient and let ð1(ð), . . . , ðð(ð) be real analytic functions of real ð forwhich det(ðððŒð â ð¿(ð)) = 0, ð = 1, . . . , ð. Let ð1 < â â â < ðð be the different realeigenvalues of ð¿(ð). For each ð write
ðð(ð) = (ðâ ðð)ðððððð(ð), ð = 1, . . . , ð,
where ððð(ðð) â= 0 is real. Then the non-zero numbers among ðð1,. . . , ððð are thepartial multiplicities of ð¿(ð) associated with ðð.
The sign of ððð(ðð) (for ððð â= 0) is the sign characteristic attached to theelementary divisors (ðâ ðð)ððð of ð¿(ð).
Note that this statement provides definitions of partial multiplicities and signcharacteristics and these are associated with each elementary divisor. (The readeris referred to [1] and [2] for more comprehensive dicussion.) In particular, if ð¿(ð)is semisimple (that is, ððð = 1 for all ð and ð) then ðð is said to be of positive ornegative type according as the sign characteristic attached to the correspondingelementary divisor (ðâ ðð) is positive or negative, respectively.
2. Admissible sign characteristics
Our first objectives are to provide characterizations of admissible sign character-istics for polynomials ð¿(ð) with either positive definite leading coefficient ð¿â, orpositive definite trailing coefficient ð¿0.
The first result relates the inertia of the leading coefficient ð¿â to the asymp-totic behaviour of the eigenfunctions.
Theorem 2.1. Let (ð, ðâ ð, 0) be the inertia of ð¿â, and let ðmax be the largest realeigenvalue of ð¿(ð). Then there are ð indices {ð1, . . . , ðð} â {1, . . . , ð} such thatfor all ð > ðmax,
ðð(ð) > 0 ðð ð â {ð1, . . . , ðð} ððð ðð(ð) < 0 ðð ð /â {ð1, . . . , ðð}.Proof. For ð = 1, . . . , ð the zeros of ðð(ð) are real eigenvalues of the polynomialmatrix ð¿(ð) and, since this matrix has at most ðâ real eigenvalues (counting withmultiplicities), the number of real zeros of ðð(ð) is finite. Then for ð > ðmax eitherðð(ð) > 0 or ðð(ð) < 0 for ð = 1, . . . , ð.
On the other hand, for any fixed real ð, the real number ð(ð) is an eigenvalueof the selfadjoint (real or complex) matrix ð¿(ð). Let ð1(ð¿â) ⥠â â â ⥠ðð(ð¿â) denotethe eigenvalues of ð¿â with
ð1(ð¿â) ⥠â â â ⥠ðð(ð¿â) > 0 > ðð+1(ð¿â) ⥠â â â ⥠ðð(ð¿â)
On the Sign Characteristics of Selfadjoint Matrix Polynomials 191
We will use the âWeyl inequalitiesâ for the eigenvalues of the sum of two sym-metric or Hermitian matrices (see [4, Th. 4.3.1], for example). If ð»1, ð»2 are ðà ðHermitian or symmetric matrices then
ðð(ð»1) + ðð(ð»2) †ðð(ð»1 +ð»2) †ðð(ð»1) + ð1(ð»2)
where the eigenvalues of ð»1, ð»2 and ð»1+ð»2 are arranged in non-increasing order.
Let ð0 be a real number. Applying the left-hand Weyl inequality repeatedlyto the symmetric matrix ð¿(ð0) = (ð¿âð
â0 + â â â + ð¿1ð0) + (ð¿0) we have
ðð(ð¿(ð0)) ⥠ðð(ð¿âðâ0 + â â â + ð¿1ð0) + ðð(ð¿0)
⥠ðð(ð¿âðâ0 + â â â + ð¿2ð20) + ð0ðð(ð¿1) + ðð(ð¿0)
⥠â â â ⥠ðâ0ðð(ð¿â) + ð
ââ10 ðð(ð¿ðâ1) + â â â + ðð(ð¿0)
Thus, if ð = min{ðð(ð¿ðâ1), . . . , ðð(ð¿0)}, then for ð0 > 1,
ðð(ð¿(ð0)) ⥠ðâ0ðð(ð¿â) + (ðââ10 + â â â + ð0 + 1)ð
⥠ðâ0ðð(ð¿â) +ðâ0 â 1ð0 â 1ð.
(2)
Assume now that ðð(ð¿â) > 0. If ð ⥠0 then ðð(ð¿(ð0)) > 0 for ð0 > 1. Also,if ð < 0 then for ð0 > 1â ð
ðð(ð¿â)we have
ðâ0(ð0 â 1) > âð
ðð(ð¿â)ðâ0 > â
ð
ðð(ð¿â)ðâ0 +
ð
ðð(ð¿â),
whence
ðâ0(ð0 â 1) > âð
ðð(ð¿â)(ðâ0 â 1),
and so
ðð(ð¿â)ðâ0 > âð
ðâ0 â 1ð0 â 1 .
Using this in (2) we find that ðð(ð¿(ð0)) > 0 for ð0 > 1 + â£ðâ£ðð(ð¿â)
.
Next, we use the right-handWeyl inequality (ðð(ð»1+ð»2) †ðð(ð»1)+ð1(ð»2))to show in a similar way that
ðð(ð¿(ð0)) †ðâ0ðð(ð¿â) + ðââ10 ð1(ð¿ðâ1) + â â â + ð1(ð¿0),
†ðâ0ðð(ð¿â) +ðâ0 â 1ð0 â 1ð,
with ð = max{ð1(ð¿0), . . . , ð1(ð¿ââ1)}.If we assume that ðð(ð¿â) < 0 then, as above, ð < 0 implies ðð(ð¿(ð0)) < 0
for ð0 > 1. Similarly,ð > 0 implies that ðð(ð¿(ð0)) < 0 for ð0 > 1â ððð(ð¿â)
. Hence,
for ð = ð + 1, . . . , ð, ðð(ð¿(ð0)) < 0 for ð0 > 1â â£ðâ£ðð(ð¿â)
.
192 P. Lancaster and I. Zaballa
Bearing in mind that ðð(ð¿â) ⥠ðð(ð¿â), ð = 1, . . . , ð and ðð(ð¿â) †ðð+1(ð¿â) forð = ð+1, . . . , ð, we conclude that ðð(ð¿(ð0)) > 0 for ð = 1, . . . , ð and ðð(ð¿(ð0)) < 0for ð = ð + 1, . . . , ð whenever
ð0 > max
{1 +
â£ðâ£ðð(ð¿â)
, 1â â£ð â£ðð+1(ð¿â)
}. (3)
But the eigenvalues of ð¿(ð0) are ð1(ð0),. . . , ðð(ð0). Then, for ð0 satisfying (3),there are ð indices {ð1, . . . , ðð} â {1, . . . , ð} such that, if
{ð1, . . . , ððâð} = {1, . . . , ð}â{ð1, . . . , ðð},
then ððð(ð0) > 0, ð = 1, . . . , ð, and ððð(ð0) < 0, ð = 1, . . . , ð â ð. The theoremfollows using the fact that ðð(ð) is either positive or negative for ð > ðmax. â¡
In a similar way, the behaviour of the eigenfunctions of ð¿(ð) near zero isclosely related to the inertia of the trailing coefficient, ð¿0.
Theorem 2.2. Let ð¿(ð) = ð¿âðâ+ð¿ââ1ðââ1+ â â â +ð¿0 be an ðÃð selfadjoint matrix
polynomial with detð¿â â= 0. Let ð1(ð), . . . , ðð(ð) be the eigenfunctions of ð¿(ð)and let (ð, ð, ð¿) be the inertia of ð¿0. Let ðð§ be the positive real eigenvalue of ð¿(ð)closest to zero. Then there are ð indices {ð1, . . . , ðð} â {1, . . . , ð} and ð indices{ð1, . . . ðð} â {1, . . . , ð} â {{ð1, . . . , ðð} such that for 0 < ð < ðð§, ðð(ð) > 0 ifð â {ð1, . . . , ðð} and ðð(ð) < 0 if ð â {ð1, . . . , ðð}.
Proof. The proof follows the same lines as that of Theorem 2.1. First, for ð =1, . . . , ð, ðð(ð) is either positive or negative for ð between any two consecutivereal eigenvalues of ð¿(ð). In particular, each eigenfunction has constant sign in(0, ðð§). Let the eigenvalues of ð¿0 be
ð1(ð¿0) ⥠â â â ⥠ðð(ð¿0) > 0 > ðð+ð¿+1(ð¿0) ⥠â â â ⥠ðð(ð¿0),
and let ð0 be a positive real number. Then ðð(ð¿0) > 0 for ð = 1, . . . , ð and,applying successively the left-hand Weyl inequalities to ð¿(ð0), we obtain
ðð(ð¿(ð0)) ⥠ðð(ð¿âðâ0 + â â â + ð¿1ð0) + ðð(ð¿0),
⥠â â â ⥠(ðâ0ðð(ð¿â) + â â â + ð0ðð(ð¿1)) + ðð(ð¿0),⥠ð(ð0 + â â â + ðâ0) + ðð(ð¿0),
= ðð01â ðâ01â ð0 + ðð(ð¿0),
(4)
where ð = min{ðð(ð¿1), . . . , ðð(ð¿â)}.
On the Sign Characteristics of Selfadjoint Matrix Polynomials 193
Now, if 0 < ð0 < 1 â â£ðâ£ðð(ð¿0)+â£ð⣠then ðð(ð¿(ð0)) > 0. In fact, if ð ⥠0 then
0 < ð0 < 1 and it is plain that ðð(ð¿(ð0)) ⥠ðð(ð¿0) > 0. And if ð < 0 then
1 +ð
ðð(ð¿0)âð > ð0 â ðð(ð¿0)
ðð(ð¿0)âð > ð0 â âðð(ð¿0)/ð1â ðð(ð¿0)/ð > ð0
â âðð(ð¿0)ð
> ð0
(1â ðð(ð¿0)
ð
)â (1â ð0)
(âðð(ð¿0)
ð
)> ð0
â âðð(ð¿0)ð
> ð01
1â ð0 > ð01â ðâ01â ð0 â ðð(ð¿0) > âðð0 1â ð
â0
1â ð0 .
It follows from (4) that ðð(ð¿(ð0)) ⥠ðð0 1âðâ0
1âð0+ ðð(ð¿0) > 0.
Similarly, if ð = ð + ð¿ + 1, . . . , ð then ðð(ð¿0) < 0 and we can apply theright-hand Weyl inequalities to show that
ðð(ð¿(ð0)) â€ðð0 1â ðâ0
1â ð0 + ðð(ð¿0),where ð = max{ð1(ð¿1), . . . , ð1(ð¿ð)}. As in the previous case, if 0 †ð0 †1 +
â£ðâ£ðð(ð¿0)ââ£ð⣠then ðð(ð¿(ð0)) < 0.
Therefore, for ð0 > 0 close enough to zero, ðð(ð¿(ð0)) > 0 or ðð(ð¿(ð0)) < 0according as ðð(ð¿0) > 0 or ðð(ð¿0) < 0, respectively. Since ðð(ð0) is an eigenvalueof ð¿(ð0) and ðð(ð) does not change sign in (0, ðð§) there must be ð eigenfunctionstaking positive values in (0, ðð§) and ð eigenfunctions taking negative values in thesame open interval. â¡
3. The semisimple case
With the help of Theorems 2.1 and 2.2 we can establish a necessary conditionon the sign characteristics of the real eigenvalues of semisimple selfadjoint matrixpolynomials with positive definite leading and/or trailing coefficient. This leads toa proof of the following result which was stated in [5] â without proof.
Theorem 3.1. Let ð¿(ð) be an ð à ð semisimple selfadjoint matrix polynomialwith ð¿â > 0 and maximal and minimal real eigenvalues ðmax and ðmin, respec-tively. For any ðŒ †ðmax, let ð(ðŒ) denote the number of real eigenvalues (countingmultiplicities) of ð¿(ð) of positive type in (ðŒ,+â) and ð(ðŒ) denote the numberof real eigenvalues (counting multiplicites) of ð¿(ð) of negative type in [ðŒ,+â).Then
ð(ðŒ) †ð(ðŒ) for all ðŒ â [ðmin, ðmax]. (5)
In particular, this theorem says that, in the semisimple case, if ð¿â > 0 thenfor each real eigenvalue of negative type there is at least one larger real eigenvalueof positive type.
Proof. Let ðŒ â â be such that ðmin †ðŒ †ðmax and let ð0 â [ðŒ, ðmax] bean eigenvalue of ð¿(ð). Since ð¿(ð) is semisimple the algebraic multiplicity of ð0
194 P. Lancaster and I. Zaballa
coincides with its geometric multiplicity. Bearing in mind that dimKerð¿(ð0) is thenumber of eigenfunctions that have ð0 as a zero, we can associate an eigenfunction(perhaps in more than one way) with each eigenvalue.
Let ð0(ð) be the eigenfunction associated with ð0. Then, according to The-orem 1.1, we can write ð0(ð) = (ð â ð0)ð0(ð) with ð0(ð0) â= 0 and the sign to beassociated with ð0 is positive or negative according as ð0(ð0) > 0 or ð0(ð0) < 0. Ifthe negative sign applies, then ð0(ð) is decreasing at ð0 and so ð0(ð0+) < 0. Butð¿â > 0 and by Theorem 2.1 ð0(ð) > 0 for ð > ðmax. It follows then that there
is ᅵᅵ0 > ð0 such that ð0(ᅵᅵ0) = 0 and ðâ²0(ᅵᅵ0) > 0. Thus, the eigenvalue ᅵᅵ0 has anassociated positive sign.
This means first that ð0 < ðmax (i.e., ð(ðmax) = 0 and so ðmax is of positivetype) and, also, that for each eigenvalue of ð¿(ð) in the interval [ðŒ, ðmax) of negativetype there is a larger eigenvalue in (ðŒ, ðmax] of positive type. Noting that ðmax
is necessarily of positive type, and ð¿(ð) has no eigenvalues in (ðmax,+â), weconclude that ð(ðŒ) †ð(ðŒ) for all ðŒ â [ðmin, ðmax], as desired. â¡
When ð¿(ð) is semisimple and of even degree, the number of eigenvalues ofpositive type equals the number of negative type ([1, Prop. 4.2]). In this caseð(ðmin) = ð(ðmin) implying that ðmin is of negative type. On the other hand, wehave seen in the proof of the theorem that ðmax is of positive type.
Theorem 3.1 should be compared with (the more general) Theorem 1.11 andExample 1.5 of [2].
Theorem 3.2. Let ð¿(ð) be an ðÃð semisimple selfadjoint matrix polynomial withð¿0 > 0 and maximal and minimal real eigenvalues ðmax and ðmin, respectively.
â For ðŒ < 0 let ðâ(ðŒ) denote the number of real eigenvalues (counting multi-plicities) of ð¿(ð) of positive type in (ðŒ, 0] and ðâ(ðŒ) the number of realeigenvalues (counting multiplicites) of ð¿(ð) of negative type in [ðŒ, 0).
â For ðŒ > 0 let ð+(ðŒ) denote the number of real eigenvalues (counting multi-plicities) of ð¿(ð) of positive type in (0, ðŒ] and ð+(ðŒ) the number of realeigenvalues (counting multiplicites) of ð¿(ð) of negative type in [0, ðŒ).
Then ðâ(ðŒ) †ðâ(ðŒ) for all ðŒ â [ðmin, 0) and ð+(ðŒ) ⥠ð+(ðŒ) for all ðŒ â (0, ðmax).
Proof. The line of proof is similar to that of the previous theorem but using The-orem 2.2. Let ðð§ be the eigenvalue of ð¿(ð) closest to zero. If ðŒ â [ðmin, 0) andð0 â [ðŒ, 0) is an eigenvalue of ð¿(ð) of negative type, the corresponding eigenfunc-tion is negative to the right of ð0 (but close enough to ð0). By Theorem 2.2, that
eigenfunction is positive in (0, ðð§) so that there must be an eigenvalue 0 > ᅵᅵ0 < ð0of ð¿(ð) of positive type. Hence ðâ(ðŒ) †ðâ(ðŒ) for ðŒ â [ðmin, 0).
Similarly, if ðŒ â (0, ðmax] and ð0 â (0, ðŒ] is an eigenvalue of ð¿(ð) of positivetype, the corresponding eigenfunction is negative to the left (but near) ð0. ByTheorem 2.2, that eigenfunction is positive in (0, ðð§) so that ð0 > ðð§, ðð§ is of
negative type and there must be an eigenvalue 0 < ð𧠆ᅵᅵ0 < ð0 of ð¿(ð) ofnegative type. â¡
On the Sign Characteristics of Selfadjoint Matrix Polynomials 195
We remark again that, for the matrix polynomials of Theorem 3.2, ðð§ isnecessarily of negative type.
Putting together the previous results we can provide an additional neces-sary condition that the sign characteristics of all semisimple selfadjoint matrixpolynomials with positive definite leading and trailing coefficients must satisfy.
Theorem 3.3. Let ð¿(ð) be an ðÃð semisimple selfadjoint matrix polynomial withð¿â > 0 and ð¿0 > 0. With the notation of the previous theorem, the followingcondition holds:
ð+(ðmax) = ð+(ðmax). (6)
And if â is even then
ðâ(ðmin) = ðâ(ðmin) (7)
Proof. If ð¿(ð) has positive definite leading coefficient and ðð§ is the real positiveeigenvalue of ð¿(ð) closest to zero, then, by (5), ð(ðð§) ⥠ð(ðð§). That is to say, thenumber of eigenvalues of ð¿(ð) of positive type in (ðð§ ,+â) is not smaller than thenumber of eigenvalues of negative type in [ðð§ ,+â).
Now, ðð§ is of negative type because the trailing coefficient is positive definite,and ðmax is of positive type because the leading coefficient is positive definite.Thus, ð(ðð§) is also the number of eigenvalues of positive type in (0, ðmax] and ð(ðð§)is the number of eigenvalues of negative type in [0, ðmax). Hence, ð+(ðmax) = ð(ðð§)and ð+(ðmax) = ð(ðð§). Since ð(ðð§) ⥠ð(ðð§) and by Theorem 3.2, ð+(ðmax) â¥ð+(ðmax), we conclude that this is indeed an equality.
As mentioned above, if ð¿(ð) is of even degree and semisimple, then the num-ber of eigenvalues of positive type and negative type is the same. It follows fromð+(ðmax) = ð+(ðmax) that the number of eigenvalues of ð¿(ð) of positive type in[ðmin, 0] equals the number of eigenvalues of negative type in that interval. Takinginto account that 0 is not an eigenvalue of ð¿(ð) and that ðmin is of negative type,the number of eigenvalues of ð¿(ð) of negative type in [ðmin, 0) is ðâ(ðmin) and thatthe number of eigenvalues of positive type in [ðmin, 0] is ðâ(ðmin). In conclusionðâ(ðmin) = ðâ(ðmin) as claimed. â¡
4. Conclusions
Using the notion of the âsign characteristicâ of real eigenvalues of selfadjoint ma-trix polynomials, Theorems 2 and 3 establish new results on the ordering of realeigenvalues with respect to these signs. The results take into account the inertiaof the (invertible) leading coefficient, and the trailing coefficient, respectively.
In Section 3, these results have been applied to semisimple matrix polynomi-als with positive definite leading coefficient (Theorem 4), and with positive definitetrailing coefficient (Theorem 5).
These results will be used in [6] to provide solutions for the inverse realsymmetric quadratic eigenvalue problem, in the semisimple case, when the leadingand trailing coefficients are prescribed to be hold some definiteness constraints.
196 P. Lancaster and I. Zaballa
References
[1] Gohberg I., Lancaster P., and Rodman L., Spectral analysis of self-adjoint matrixpolynomials. Research paper 419 (1979) Dept. Mathematics and Statistics, Universityof Calgary, Canada.
[2] Gohberg I., Lancaster P., and Rodman L., Spectral analysis of self-adjoint matrixpolynomials. Annals of Math., 112, (1980), 33â71.
[3] Gohberg I., Lancaster P., and Rodman L., Matrix Polynomials, Academic Press,New York, 1982 and SIAM, Philadelphia, 2009.
[4] Horn R.A., Johnson Ch.R.,Matrix Analysis, Cambridge University Press, New York,1985.
[5] Lancaster P., Prells U., Zaballa I., An Orthogonality property for real symmet-ric matrix polynomials, Operators and Matrices, to appear (avaliable on linehttp://files.ele-math.com/preprints/oam-07-21-pre.pdf).
[6] Lancaster P., Zaballa I., On the Inverse Symmetric Quadratic Eigenvalue Problem.In preparation.
Peter LancasterDept. of Mathematics and StatisticsUniversity of CalgaryCalgaryAlberta, T2N 1N4, Canadae-mail: [email protected]
Ion ZaballaDepartamento de Matematica Aplicada y EIOEuskal Herriko Unibertsitatea (UPV/EHU)Apdo 644E-48080 Bilbao, Spaine-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 197â219câ 2013 Springer Basel
Quadratic Operators in Banach Spaces andNonassociative Banach Algebras
Yu.I. Lyubich
Dedicated to Leonia Lerer on the occasion of his 70th birthday
Abstract. A survey of a general theory of quadratic operators in Banachspaces with close relations to the nonassociative Banach algebras is presented.Some applications to matrix and integral quadratic operators in classical Ba-nach spaces are given.
Mathematics Subject Classification (2010). 46H70, 45G10.
Keywords. Cubic matrices, integral operators, Bernstein algebras.
1. Introduction
Let Ί be a field, and let ð be a linear space over Ί, finite- or infinite-dimensional.If, in addition, a bilinear mapping ð from the Cartesian square ð Ãð into ð isgiven then ð is called an algebra over Ί. In this setting the vectorð (ð¥, ðŠ) is calledthe product of ð¥ and ðŠ and usually denoted by ð¥ðŠ. Accordingly, the mapping ðis called a multiplication. Its bilinearity means the distributive laws
(ð¥ + ðŠ)ð§ = ð¥ð§ + ðŠð§, ð§(ð¥+ ðŠ) = ð§ð¥+ ð§ðŠ (ð¥, ðŠ, ð§ â ð)and the homogeneity of degree 1, i.e.,
(ðŒð¥)ðŠ = ð¥(ðŒðŠ) = ðŒ(ð¥ðŠ) â¡ ðŒð¥ðŠ, (ð¥, ðŠ â ð, ðŒ â Ί).Neither associativity nor unitality are assumed in this definition. For the generaltheory of nonassociative algebras the basic reference is the book [20].
Given an algebra ð , the diagonal restriction of the multiplication, i.e., themapping ð : ð â ð defined as ð ð¥ = ð¥ð¥ â¡ ð¥2 is called a quadratic mapping orquadratic operator. Its simplest properties are
ð (ðŒð¥) = ðŒ2ð ð¥, ð (âð¥) = ð ð¥, ð (0) = 0. (1.1)
198 Yu.I. Lyubich
If the algebra is commutative then we have the elementary identities
ð¥2 â ðŠ2 = (ð¥â ðŠ)(ð¥+ ðŠ) (1.2)
and
(ð¥Â± ðŠ)2 = ð¥2 + ðŠ2 ± 2ð¥ðŠ. (1.3)
Later on char(Ί) â= 2. From (1.3) it follows that
ð¥ðŠ =1
2(ð (ð¥ + ðŠ)â ð ð¥â ð ðŠ) , (1.4)
or in a more elegant form
ð¥ðŠ =1
4(ð (ð¥ + ðŠ)â ð (ð¥â ðŠ)) . (1.5)
As a result, with a fixed underlying linear space ð the correspondence between com-mutative algebras and quadratic operators is one-to-one. Furthermore, a subspaceð â ð is invariant for ð if and only if this is a subalgebra of the correspondingcommutative algebra. The quadratic operator corresponding to this subalgebra isthe restriction ð â£ð . In particular, a one-dimensional subspace is a subalgebra ifand only if its basis vector is an eigenvector of ð . Note that if ð ð¥ = ðð¥, ð â Ί,and ð â= 0 then ð§ = ð¥/ð is a fixed point of ð that is the same as an idempotent ofthe corresponding algebra: ð§2 = ð§.
Even without commutativity the mapping ð¥ ï¿œâ ð¥2 is quadratic by definition.However, ð¥2 = ð¥ â ð¥ where the new multiplication
ð¥ â ðŠ = ð¥ðŠ + ðŠð¥
2. (1.6)
is commutative. But for a given quadratic mapping a noncommutative multiplica-tion is not unique. The general form of such multiplications is
ð¥ðŠ = ð¥ â ðŠ + [ð¥, ðŠ] (1.7)
where the second summand is anticommutative, i.e., [ð¥, ðŠ] = â[ðŠ, ð¥] for all ð¥, ðŠ.Indeed, the anticommutativity is equivalent to the identity [ð§, ð§] = 0. Note that(1.7) is a unique decomposition of a given algebra into the sum of a commutativealgebra and an anticommutative one. If the algebra is associative then the com-mutative summand is an Jordan algebra, while the anticommutative one is a Liealgebra, and both are not associative, in general.
The quadratic operators come from the classical analysis, calculus of vari-ations and differential geometry, etc., as the second differentials of mappings ofclass ð¶2. On the other hand, they appear as the generators of nonlinear dynamicalsystems in form of cubic matrices (finite or infinite), the integral operators, etc.In this setting the corresponding algebras are very useful for studying of the dy-namics. A bright example is the Mendelian algebra in genetics connected with theHardyâWeinberg quadratic mapping, see, e.g., [10, Sections 3 and 8]. This algebrais three-dimensional but, in general, the genetic situations are multidimensional.In this way many interesting classes of nonassociative algebras appear after the
Quadratic Operators 199
pioneering works of Serebrovskii [19], Glivenko [7] and Etherington [6]. Much ear-lier Bernstein [1] suggested the quadratic operators in âð as an adequate languagefor a fundamental problem in the population dynamics. A synthesis of operatorand algebra approaches turned out to be extraordinary fruitful, see book [14] andthe references therein, especially, [18]. For purely algebraic aspects we refer to thebook [21].
In [10] the author investigated the Bernstein quadratic operators, and afterthat the corresponding algebras were introduced in [11] as a powerful tool for theBernstein problem. Eventually, in this way the problem has been completely solved[10, 11, 12, 8]. Note that the name Bernstein algebras appeared in [13] for the firsttime. The main structure theorem for these ones was first obtained in the operatorform [10, Theorem 4.1]. In [9] this result was reproduced in an explicit algebraicform as a base for a further investigation. Nowadays the Bernstein algebra theoryis a well-developed area of the modern algebra.
In the present paper we consider the quadratic operators in infinite-dimen-sional spaces with a special attention to the operators of finite rank, i.e., withfinite-dimensional images (Section 2). If a quadratic operator in a Banach spaceis continuous then the corresponding algebra is a Banach one, and vice versa. Westudy this relation in Sections 3 and then proceed to a compact situation (Section6) focusing on integral quadratic operators. These considerations culminate inSection 7 devoted mainly to the Bernstein quadratic operators and algebras. Inthe compact case they are of finite rank [15]. Moreover, the Bernstein integraloperators in ð¶[0, 1] with positive kernels are of rank 1. This remarkable resultannounced by Bernstein [1] was never proved until [16] where a proof was givenwith help of the corresponding Bernstein algebra introduced in [15].
2. Algebraic preliminaries
Obviously, the setðð of all quadratic operators in a linear spaceð is a linear spaceisomorphic to the space of all commutative multiplications in ð . Some quadraticoperators come naturally from the linear ones. For example, if ð is a linear operatorand ð is a linear functional then
ð ð¥ = ð[ð¥]ðð¥ (2.1)
is a quadratic operator corresponding to the multiplication
ð¥ðŠ =1
2(ð[ð¥]ððŠ + ð[ðŠ]ðð¥). (2.2)
Also note that if ð is a linear operator then the superposition ðð and ð ð arequadratic operators for every quadratic ð . Indeed, ð (ð¥ðŠ) and (ðð¥)(ððŠ) are bilinearfor any algebra. A quadratic operator ð is called elementary if
ð ð¥ = ð£[ð¥]ð, ð¥ â ð, (2.3)
where ð â ð and ð£[ð¥] is a quadratic functional, i.e., ð£[ð¥] = ð€[ð¥, ð¥] where ð€[ð¥, ðŠ] isa (unique) symmetric bilinear functional, the polar of ð£.
200 Yu.I. Lyubich
Accordingly,
ð¥ðŠ = ð€[ð¥, ðŠ]ð, ð¥ â ð, (2.4)
The simplest example is
ð¥ðŠ = ð[ð¥]ð[ðŠ]ð (2.5)
where ð is a linear functional. This algebra is associative.For any quadratic operator ð we consider the invariant subspace ð(ð ) =
Span(Imð ) of the space ð . In more detail,
ð(ð ) ={âð
ð=1ðŒðð ð¥ð : ð¥ð â ð, ðŒð â Ί, ð ⥠1
}.
The corresponding subalgebra (even an ideal) of the algebra ð is
ð â² ={âð
ð=1ðŒðð¥ððŠð : ð¥ð , ðŠð â ð, ðŒð â Ί, ð ⥠1
}.
Indeed, ð(ð ) â ð â², obviously. On the other hand ð â² â ð(ð ) by polarization (1.4)or (1.5). Hence, ð(ð ) = ð â². Up to the end of this section we keep all in terms ofð but, of course, everything can be immediately polarized.
The dimension ð(ð ), 0 †ð(ð ) †â, of the space ð(ð ) is called the rank ofthe quadratic operator ð . Obviously, ð(ð ) = 0 if and only if ð = 0. Denote byð¹ðð the set of all quadratic operators of finite rank.
Proposition 2.1. The set ð¹ðð is a subspace of ðð .
Proof. For every ðŒ â Ί/{0} we have ð(ðŒð ) = ð(ð ). Furthermore, ð(ð1 + ð2) âð(ð1) + ð(ð2). Thus,
ð(ðŒð ) = ð(ð ) (ðŒ â= 0), ð(ð1 + ð2) †ð(ð1) + ð(ð2). (2.6)â¡
In addition, ð(ðð ) †ð (ð ) and ð(ð ð ) †ð(ð ) for any linear operator ð .Hence, the subspace ð¹ðð is invariant for the mappings ð ï¿œâ ðð and ð ï¿œâ ð ð .
Proposition 2.2. Each ð â ð¹ðð is the sum of ð(ð ) elementary operators.
Proof. Let (ðð)ðð=1 be a basis of ð(ð ), so ð = ð(ð ), and let (ðâð )
ðð=1 be the dual
basis in ð(ð )â. Since ð ð¥ â ð(ð ) for every ð¥ â ð , we haveð ð¥ =
âð
ð=1ð£ð [ð¥]ðð , ð¥ â ð, (2.7)
where ð£ð [ð¥] = ðâð [ð ð¥] are quadratic functionals. â¡
Corollary 2.3. An operator ð â ðð â{0} is of rank 1 if and only if it is elementary.
Corollary 2.4. For ð â ð¹ðð the rank ð(ð ) is the minimal number of elementaryquadratic operators sum of which is equal to ð .
Proof. This follows from Proposition 2.2 because of the inequality (2.6). â¡
We call minimal any decomposition of ð â ðð â {0} into a sum of ð(ð ) ele-mentary quadratic operators. To investigate this case we use the following general
Quadratic Operators 201
Lemma 2.5. If (ðð(ð¡))ðð=1 is a linearly independent system of scalar functions on a
set ð then there are ð points ð¡1, . . . , ð¡ð in ð such that
det(ðð(ð¡ð))ðð,ð=1 â= 0. (2.8)
Proof. Let us consider the mapping Ω : ð â Ίð defined as Ω(ð¡) = (ðð(ð¡))ðð=1,
ð¡ â ð. The subspace ð¿ = Span(ImΩ) coincides with Ίð. Indeed, otherwise, ð¿would lie in a hyperplane, so there is (ðŒð)
ðð=1 â Ίð â {0} such that
ðâð=1
ðŒððð(ð¡) = 0, ð¡ â ð,
that contradicts the linear independence of the system (ðð(ð¡))ðð=1. As a result, there
are ð linearly independent vectors in ImΩ. They are Ω(ð¡ð) with some ð¡ð â ð. Thus,the columns of the matrix (ðð(ð¡ð))
ðð,ð=1 are linearly independent. â¡
Now let
ð ð¥ =
ð âð=1
ð£ð [ð¥]ðð , ð¥ â ð, (2.9)
where ðð are vectors and ð£ð [ð¥] are quadratic functionals.
Lemma 2.6. If in (2.9) the ð£ðâs are linearly independent then all ðð â ð(ð ).Proof. By Lemma 2.5 there are ð vectors ð¥1, . . . , ð¥ð such that
det(ð£ð [ð¥ð])ð ð,ð=1 â= 0.
Therefore, the system of linear equationsð â
ð=1
ð£ð [ð¥ð]ðð = ð ð¥ð, 1 †ð †ð ,
with unknown ðð âs is solvable. By Cramerâs rule all ðð â ð(ð ). â¡Now we are able to characterize the minimal decompositions.
Theorem 2.7. The following statements are equivalent:
1) The decomposition (2.9) is minimal, i.e., ð = ð(ð ).2) The systems (ðð)
ð ð=1 and (ð£ð)
ð ð=1 are both linearly independent.
3) The system (ðð)ð ð=1 is a basis in ð(ð ).
Proof. 1)â2). Suppose to the contrary. For definiteness, let the system (ðð)ð ð=1
be linearly dependent. Then one of ððâs is a linear combination of others. Thesubstitution of this expression into (2.9) reduces the number of summands to ð â1that contradicts the minimality.
2)â3) since all ðð â ð(ð ) by Lemma 2.6 and Span(ðð)ð ð=1 = ð(ð ) by (2.9).
3)â1). ð = ð(ð ) since ð(ð ) = dimð(ð ) and (ðð)ð ð=1 is a basis in ð(ð ). â¡
Corollary 2.8. If the decomposition (2.9) is minimal then the system (ðð)ð ð=1 is a
basis in ð(ð ) and ð£ð [ð¥] = ðâð [ð ð¥], 1 †ð †ð , where (ðâð )ð ð=1 is the dual basis in
ð(ð )â.
202 Yu.I. Lyubich
3. Continuous quadratic operators and Banach algebras
From now on the ground field Ί is â or â, until otherwise stated. Recall that areal or complex Banach space ð is said to be a Banach algebra if ð is an algebrawith continuous multiplication, i.e., the product ð¥ðŠ is a continuous function of(ð¥, ðŠ) â ð Ãð .Proposition 3.1. Let ð be an algebra and a Banach space with a norm â¥.â¥. Thenthe following statements are equivalent:
(1) ð is a Banach algebra.
(2) The product ð¥ðŠ is continuous at the point (0, 0).
(3) The inequality
â¥ð¥ðŠâ¥ †ð¶ â¥ð¥â¥ â¥ðŠâ¥ (3.1)
holds with a constant ð¶ > 0.
Proof. (1)â(2) trivially.(2)â(3). Suppose to the contrary. Then there is a sequence (ð¥ð, ðŠð)âð=1 â
ð Ãð such that â¥ð¥ððŠð⥠> ð â¥ð¥ð⥠â¥ðŠðâ¥. Obviously, ð¥ð â= 0 and ðŠð â= 0. By letting
ð¢ð = ð¥ð/âð â¥ð¥ð⥠, ð£ð = ðŠð/
âð â¥ðŠð⥠,
we obtain â¥ð¢ðð£ð⥠> 1, while â¥ð¢ð⥠= â¥ð£ð⥠= 1/âðâ 0, a contradiction.
(3)â(1). From (3.1) it follows that
â¥(ð¥+ ð¢)(ðŠ + ð£)â ð¥ðŠâ¥ = â¥ð¥ð£ + ð¢ðŠ + ð¢ð£â¥ †ð¶(â¥ð¥â¥ â¥ð£â¥ + â¥ð¢â¥ â¥ðŠâ¥+ â¥ð¢â¥ â¥ð£â¥).Hence, (ð¥+ ð¢)(ðŠ + ð£)â ð¥ðŠ when (ð¢, ð£)â 0 in ð Ãð . â¡
Sometimes, it can be reasonable to change the norm in a Banach algebra ðto an equivalent one. By definition, the ratio of two equivalent norms lies in asegment [ð, ð] of the semiaxis (0,â). When a norm runs over an equivalence class,the topology remains the same, while the constant ð¶ in (3.1) takes all positivevalues. This is true even if we only consider the norms proportional to a fixed one:{â¥.â¥â² = ð â¥.⥠: ð > 0}. Indeed, from (3.1) it follows that
â¥ð¥ðŠâ¥â² †ð¶ðâ1 â¥ð¥â¥â² â¥ðŠâ¥â² . (3.2)
By the way, with ð = ð¶ we have
â¥ð¥ðŠâ¥â² †â¥ð¥â¥â² â¥ðŠâ¥â² . (3.3)
It is useful to add a geometrical criterion to the Proposition 3.1. Let usconsider the balls ðµð = {ð§ â ð : â¥ð§â¥ †ð}, ð > 0, in a Banach space ð . Let ð bean algebra, as before.
Proposition 3.2. For ð in order to be a Banach algebra it is necessary and suf-ficient that the product ð¥ðŠ is bounded on every ðµð à ðµð and sufficient that thisproduct is bounded on a ðµð Ãðµð.
Quadratic Operators 203
Proof. Let ð be a Banach algebra. Then for (ð¥, ðŠ) â ðµð Ãðµð the inequality (3.1)yields â¥ð¥ðŠâ¥ †ð¶ð2. Conversely, let â¥ð¥ðŠâ¥ †ð for a ð > 0 and (ð¥, ðŠ) â ðµð à ðµð
with an ð. For any (ð¥, ðŠ) â ð Ãð the vectors ðð¥/ â¥ð¥â¥ and ððŠ/ â¥ðŠâ¥ belong to ðµð.Hence, â¥ð¥ðŠâ¥ â€ððâ2 â¥ð¥â¥ â¥ðŠâ¥. â¡
Now we proceed to the quadratic operators in a Banach spaceð and establishsome criteria for their continuity.
Proposition 3.3. Let ð : ð â ð be a quadratic operator. The following statementsare equivalent:
(1) ð is continuous.(2) The corresponding commutative algebra is a Banach algebra.(3) ð is bounded in the sense that the inequality
â¥ð ð¥â¥ †ð¶ â¥ð¥â¥2 (3.4)
holds with a constant ð¶ > 0.(4) The image of every ball ðµð is contained in a ball ðµð (ð).(5) The image of a ball ðµð is contained in a ball ðµð .(6) ð is continuous at the point ð¥ = 0.
Proof. (1)â(2) by polarization.(2)â(3) by taking ðŠ = ð¥ in the inequality (3.1).(3)â(4) since if â¥ð¥â¥ †ð then â¥ð ð¥â¥ †ð¶ð2 by (3.4).(4)â(5) trivially.(5)â(6) since if ð ðµð â ðµð with some ð and ð then by homogeneity we have
ð ðµð¿ â ðµð where ð¿ = ðâð/ð .
(6)â(2) since ð¥ðŠ is continuous at (0, 0) by polarization, and then one can referto (2)â(1) from Proposition 3.1.
(2)â(1) trivially. â¡We denote the linear space of continuous quadratic operators (a subspace of
ðð) by ðµðð in accordance with the equivalence (1) â (3). The latter allows usto introduce the norm
â¥ð ⥠= supð¥ â=0
â¥ð ð¥â¥â¥ð¥â¥2 = sup
â¥ð¥â¥=1
â¥ð ð¥â¥ , ð â ðµðð . (3.5)
This definition is a counterpart of that which is the standard for a linear continuous(â bounded) operator ð . Obviously,
â¥ðð ⥠†â¥ð ⥠â¥ð ⥠, â¥ð ð ⥠†â¥ð â¥2 â¥ð ⥠.At the end of Section 4 we show that the normed space ðµðð is Banach. Its(nonclosed, in general) subspace ðµðð â© ð¹ðð we denote by ðµð¹ðð .
Theorem 3.4. An operator ð â ð¹ðð belongs to ðµð¹ðð if in a decomposition (2.9)the quadratic functionals ð£ð [ð¥] are continuous. This condition is necessary if thedecomposition is minimal.
204 Yu.I. Lyubich
Proof. The sufficiency is obvious. The necessity follows from Corollary 2.8 sinceall linear functionals on the finite-dimensional space ð(ð ) are continuous. â¡
In conclusion of this section we prove the following
Proposition 3.5. A quadratic operator ð â= 0 of form (2.1) is continuous if andonly if the linear functional ð and the linear operator ð are both continuous.
Proof. The âifâ part is obvious. By the implication 1) â 2) in Proposition 3.3 itsuffices to prove the âonly ifâ part in terms of the multiplication (2.2). Let thelatter be continuous, and let ð ⣠kerð â= 0. Then there is ðŠ such that ð[ðŠ] = 0,ððŠ â= 0, and then there is a continuous linear functional ð such that ð[ððŠ] = 2.From (2.2) it follows that ð[ð¥] = ð[ð¥ðŠ], hence ð is continuous.
Now let ð ⣠kerð = 0. Since ð â= 0, we have ð â= 0 and ð â= 0. Then thereis a vector ð such that ð[ð] = 1. This yields ð¥ â ð[ð¥]ð â kerð for all ð¥ â ð .Hence, ðð¥ = ð[ð¥]ðð, so ðð â= 0. Taking a continuous linear functional ð such thatð[ðð] = 1 we get ð[ð¥] = ð[ð¥ð], hence ð is continuous again.
Now to complete the proof we return to (2.2) taking any ðŠ such that ð[ðŠ] = 1.Then we obtain ðð¥ = 2ð¥ðŠ â ð[ð¥]ððŠ. Thus, the operator ð is continuous. â¡
4. Intrinsic characterization of quadratic operators
Here we prove the following
Theorem 4.1. Let ð be a Banach space. A continuous mapping ð : ð â ðis a quadratic operator if and only if for every two vectors ð¥, ðŠ â ð and everycontinuous linear functional ð on ð the function ð[ð (ðŒð¥ + ðœðŠ)] is a quadraticform of the scalar variables ðŒ, ðœ.
Proof. âOnly if.â In the corresponding algebra we have
ð[ð (ðŒð¥ + ðœðŠ)] = ð[(ðŒð¥+ ðœðŠ)2] = ðŒ2ð[ð¥2] + ðœ2ð[ðŠ2] + 2ðŒðœð[ð¥ðŠ].
âIf.â Now we have
ð[ð (ðŒð¥+ ðœðŠ)] = ðŒ2ð(ð¥, ðŠ;ð) + ðœ2ð(ð¥, ðŠ;ð) + 2ðŒðœð(ð¥, ðŠ;ð) (4.1)
where ð, ð, ð are some scalar functions of the triple (ð¥, ðŠ;ð). By setting ðŒ = 1, ðœ = 0and ðŒ = 0, ðœ = 1 we obtain
ð(ð¥, ðŠ;ð) = ð[ð ð¥], ð(ð¥, ðŠ;ð) = ð[ð ðŠ], (4.2)
and then with ðŒ = 1, ðœ = ±1 we getð[ð (ð¥Â± ðŠ)] = ð[ð ð¥] + ð[ð ðŠ]± 2ð(ð¥, ðŠ;ð).
Hence,
ð(ð¥, ðŠ;ð) = ð [ð (ð¥, ðŠ)] (4.3)
where
ð (ð¥, ðŠ) =ð (ð¥+ ðŠ)â ð (ð¥â ðŠ)
4. (4.4)
Quadratic Operators 205
By substitution from (4.3) and (4.2) into (4.1) and linearity of the functional ðwe obtain
ð[ð (ðŒð¥ + ðœðŠ)] = ð[ðŒ2ð ð¥+ ðœ2ð ðŠ + 2ðŒðœð (ð¥, ðŠ)]. (4.5)
By the HahnâBanach theorem the continuous linear functionals on the Ba-nach space ð separate the vectors. Therefore, from (4.5) it follows that
ð (ðŒð¥ + ðœðŠ) = ðŒ2ð ð¥+ ðœ2ð ðŠ + 2ðŒðœð (ð¥, ðŠ). (4.6)
In particular,
ð (ðŒð¥) = ðŒ2ð ð¥, ð (0) = 0.
Now from (4.4) it follows that ð (ð¥, ðŠ) is continuous and ð (ð¥, ð¥) = ð ð¥. Itremains to prove that ð (ð¥, ðŠ) is a multiplication. This multiplication will turnout to be commutative automatically since ð (ð¥, ðŠ) = ð (ðŠ, ð¥) by (4.4). Also dueto this symmetry the bilinearity of ð (ð¥, ðŠ) reduces to the linearity of ð (ð¥, .).
According to (4.4) we have
4ð (ðŒð¥1 + ðœð¥2, ðŠ) = ð (ðŒð¥1 + ðœð¥2 + ðŠ)â ð (ðŒð¥1 + ðœð¥2 â ðŠ). (4.7)
In particular,
4ð (ðŒð¥,ðŠ) = ð (ðŒð¥+ ðŠ)â ð (ðŒð¥ â ðŠ) = 4ðŒð (ð¥, ðŠ)
by (4.6) with ðœ = ±1. Thus ð (ð¥, .) is homogeneous of degree 1.
To prove the additivity of ð (ð¥, .) we note that for every triple ð¥, ðŠ, ð§ â ðžthe formula (4.6) implies
ð (ð¥+ 𧠱 ðŠ) = ð (ð¥+ ð§) + ð (ðŠ)± 2ð (ð¥+ ð§, ðŠ)
whence
ð (ð¥+ ð§ + ðŠ) + ð (ð¥+ ð§ â ðŠ) = 2{ð (ð¥+ ð§) + ð (ðŠ)}and then
ð (ð¥â ð§ + ðŠ) + ð (ð¥â ð§ â ðŠ) = 2{ð (ð¥â ð§) + ð (ðŠ)}.By subtraction we get
{ð (ð¥+ð§+ðŠ)âð (ð¥âð§+ðŠ)}+{ð (ð¥+ð§âðŠ)âð (ð¥âð§âðŠ)} = 2{ð (ð¥+ð§)âð (ð¥âð§)}that can be rewritten as
ð (ð¥+ ðŠ, ð§) +ð (ð¥â ðŠ, ð§) = 2ð (ð¥, ð§) =ð (2ð¥, ð§).
By substitution ð¥ = (ð¢+ ð£)/2, ðŠ = (ð¢â ð£)/2 we finally obtainð (ð¢, ð§) +ð (ð£, ð§) =ð (ð¢+ ð£, ð§). â¡
In essence, the proof of the additivity above is a version of an argument whichshows that in a normed space the parallelogram identity implies the Euclideanstructure, see [4, Ch. 7, Section 3]. On the other hand, the topological aspect ofthe proof can be ignored that yields the following general result, cf. [2], no3, Ex. 8.
206 Yu.I. Lyubich
Theorem 4.2. Let ð and ð be some linear spaces over a field Ί, char(Ί) â= 2. Amapping ð : ð â ð is quadratic (i.e., generated by a bilinear mapping ð Ãð âð ) if and only if such are the restrictions of ð to all two-dimensional subspacesof ð.
Taking ð = Ί we obtain
Corollary 4.3. A scalar function on a linear space ð over Ί is a quadratic func-tional if and only if such are its restrictions to all two-dimensional subspaces.
As an application of Theorem 4.1 (with some elements of its proof) we prove
Proposition 4.4. With the norm (3.5) the space ðµðð of continuous quadraticoperators in a Banach space ð is a Banach space.
Proof. Let (ðð)âð=1 be a Cauchy sequence in ðµðð . Then (ððð¥)
âð=1 is a Cauchy
sequence in ð for every ð¥ â ð . Since ð is a Banach space, the self-mappingð ð¥ = limðââ ððð¥ is well defined on ð . This is continuous since the convergenceis uniform on every ball. This is quadratic since one can pass to the limit in
ðð(ðŒð¥ + ðœðŠ) = ðŒ2ððð¥+ ðœ
2ðððŠ + 2ðŒðœðð(ð¥, ðŠ)
where
ðð(ð¥, ðŠ) =1
4(ðð(ð¥+ ðŠ)â ðð(ð¥ â ðŠ)) . â¡
5. Cubic matrices and quadratic integral operators
Assume that a Banach space ð has a Schauder basis (ðð)ðð=1, 1 †ð †â. By defi-
nition, every vector ð¥ â ð can be uniquely represented as the sum of a convergentseries (or as a finite sum if ð <â) of the form
ð¥ =
ðâð=1
ðððð (5.1)
with scalar coefficients ðð . The latter are the coordinates of ð¥ at the given basis.The functionals ðð = ð
âð [ð¥] are linear and continuous.
If ð is an algebra then we have the table of multiplication
ðððð =
ðâð=1
ðððððð . (5.2)
The scalar coefficients ðððð are called the structural constants of the algebra regard-ing the basis (ðð)
ðð=1. The structural constants constitute a (ðÃðÃð)- cubic matrix
ð which also can be treated as a sequence of length ð of square (ð Ã ð)-matricesðð = [ðððð]
ðð,ð=1. The cubic matrix ð is called the structural matrix of the algebra at
the given basis.
Quadratic Operators 207
Example 5.1. The structural coefficients of the algebra (2.2) are
ðððð =1
2(ð[ðð]ððð + ð[ðð]ððð) (5.3)
where [ððð]ðð,ð=1 is the matrix of the operator ð at the same basis.
Theorem 5.2. Let ð be a Banach algebra with a Schauder basis (ðð)ðð=1, and let
ðð and ðð be the coordinates of vectors ð¥ and ðŠ, respectively. Then the coordinatesof the product ð¥ðŠ are the bilinear forms
ðð =
ðâð,ð=1
ðððððððð (5.4)
if ð <â and
ðð = limðââ
ðâð,ð=1
ðððððððð (5.5)
if ð =â.
Proof. For ð < â the formula (5.4) follows immediately by bilinearity of themultiplication and definition of the structural constants. Now let ð =â, and let
ð¥ð =
ðâð=1
ðððð, ðŠð =
ðâð=1
ðððð.
Then ð¥ð â ð¥ and ðŠð â ðŠ as ð â â. Since the multiplication is continuous, wehave ð¥ððŠð â ð¥ðŠ. Since the coordinate functionals ðâð are continuous, we get
ðð = ðâð [ð¥ðŠ] = lim
ðââ ðâð [ð¥ððŠð],
and the case reduces to the previous one. â¡
As an immediate consequence we get
Corollary 5.3. A Banach algebra with a Schauder basis is commutative if and onlyif the corresponding structural constants ðððð are symmetric with respect ð,ð, i.e.,
all matrices ðð = [ðððð]ðð,ð=1 are symmetric.
Note that the transformation (1.6) of an algebra to a commutative one cor-
responds to the standard symmetrization [ðððð] ï¿œâ 12 ([ð
ððð] + [ðððð]).
Corollary 5.4. Let ð be a continuous quadratic operator in a Banach space ð witha Schauder basis (ðð)
ðð=1, 1 †ð †â, and let ðððð be the structural constants of
the corresponding commutative Banach algebra. Then if ðð are the coordinates ofa vector ð¥ â ð then the ðth coordinate of the vector ð ð¥ is the quadratic form
ðð =ðâ
ð,ð=1
ðððððððð (5.6)
208 Yu.I. Lyubich
if ð is finite dimensional or
ðð = limðââ
ðâð,ð=1
ðððððððð (5.7)
if ð is infinite dimensional.
In the finite-dimensional space ð the formulas (5.4) are obviously valid forany multiplication, irrespective to a topology. On the other hand, a linear topologyin ð is unique and can be defined by any norm. Using the ð1-norm
â¥ð¥â¥ =ðâ
ð=1
â£ðð⣠(5.8)
we see that (5.4) implies the inequality (3.1) with
ð¶ =
ðâð=1
maxð,ð
â£â£â£ððððâ£â£â£ .This results in the following
Proposition 5.5. With respect to the linear topology all finite-dimensional algebrasare Banach, and all quadratic operators in finite-dimensional spaces are continu-ous.
This fails in any infinite-dimensional Banach space ð .
Example 5.6. The multiplication (2.5) is not continuous if such is ð and ð â= 0.Such a ð can be obtained as the linear extension of any unbounded scalar functionon an algebraic basis (a Hamel basis) Î such that â¥ð⥠= 1 for ð â Î.
With finite ð any cubic matrix [ðððð]ðð,ð,ð=1 is the structural matrix of the
multiplication defined by (5.4) at an arbitrary basis (ðð)ðð=1. Indeed, in this setting
ðððð =
ðâð=1
( ðâð,ð=1
ððððð¿ððð¿ðð
)ðð =
ðâð=1
ðððððð (5.9)
where ð¿ is the Kronecker delta.In the infinite-dimensional case a cubic matrix must satisfy some conditions
in order to be the structural matrix of a continuous multiplication. Of course,these conditions depend on the underlying Banach space. Let us consider themultiplications in the classical spaces ðð, 1 †ð †â. As usual, we set ð = ð/(ðâ1)for 1 < ð < â, ð = â for ð = 1 and ð = 1 for ð = â. We denote the norm ofð¥ â ðð by â¥ð¥â¥ð and say that a square matrix ð = [ððð]
âð,ð=1 is of class ðð if
â¥ðâ¥ð â¡( ââ
ð,ð=1
â£ðððâ£ð)1/ð
<â
for ð <â andâ¥ðâ¥â â¡ sup
ð,ðâ¥1â£ððð⣠<â.
Quadratic Operators 209
Now, given a cubic matrix ð = [ðððð]âð,ð,ð=1, we say that it is of class ðð,ð if all square
matrices ðð = [ðððð]âð,ð=1 are of class ðð and the sequence (
â¥â¥ððâ¥â¥ð)âð=1 belongs to ðð.
In this case we set
â¥ðâ¥ð,ð =â¥â¥â¥â¥(â¥â¥ððâ¥â¥ð)âð=1
â¥â¥â¥â¥ð
. (5.10)
Theorem 5.7. Let 1 †ð < â. In the space ðð every cubic matrix ð = [ðððð]âð,ð,ð=1
of class ðð,ð being assigned to the Schauder basis ð¿ð = (ð¿ðð)âð=1, 1 †ð < â, is the
structural matrix of a continuous multiplication.
Proof. For every pair ð¥ = (ðð)âð=1, ðŠ = (ðð)
âð=1 â ðð the values
ðð =ââ
ð,ð=1
ðððððððð, ð ⥠1, (5.11)
are determined since these series converge, even absolutely. Indeed,ââ
ð,ð=1
â£â£â£ððððððððâ£â£â£ †â¥â¥ððâ¥â¥ð â¥ð¥â¥ð â¥ðŠâ¥ðby the Holder inequality for ð > 1 and trivially for ð = 1. Moreover, the sequenceð§ = (ðð)
âð=1 belongs to ðð since
ââð=1
â£ðð â£ð â€ââð=1
ââð,ð=1
â£â£â£ððððððððâ£â£â£ð †(â¥ðâ¥ð,ð â¥ð¥â¥ð â¥ðŠâ¥ð)ð
by (5.10). The relation ð§ = ð¥ðŠ defines a multiplication in ðð, and the last inequalitycan be rewritten as
â¥ð¥ðŠâ¥ð †â¥ðâ¥ð,ð â¥ð¥â¥ð â¥ðŠâ¥ð . (5.12)
By Propositions 3.1 the multiplication is continuous. It remains to note that,according to (5.11), we have
ð¿ðð¿ð =
ââð=1
ððððð¿ð (5.13)
similarly to (5.9). â¡
Remark 5.8. In fact, the proof of Theorem 5.7 does not refer to the Schauder basisin ðð. The only important is that this a sequence space, so the multiplication canbe just introduced by formula (5.11), so that
ð¿ðð¿ð = (ðððð)âð=1,
instead of (5.13). For this reason Theorem 5.7 extends to ð =â in the same wayas before. In this case we have the inequality (5.12) with
â¥ðâ¥â,1 = supðâ¥1
ââð,ð=1
â£â£â£ððððâ£â£â£ . (5.14)
210 Yu.I. Lyubich
By the way, there is no Schauder basis in ðâ since this Banach space is not sepa-rable.
Corollary 5.9. Let ð» be a separable Hilbert space. Under conditionââ
ð,ð,ð=1
â£â£â£ððððâ£â£â£2 <â (5.15)
a cubic matrix ð = [ðððð]âð,ð,ð=1 assigned to any orthonormal basis (ðð)
âð=1 in ð» is
the structural matrix of a continuous multiplication.
Proof. One can assume that ð» = ð2 and ðð = ð¿ð. Then the sum in (5.15) turnsinto the square of â¥ðâ¥2,2. â¡
Remark 5.10. By the Parseval equality we have
â¥ððððâ¥2 =ââð=1
â£â£â£ððððâ£â£â£2 .Hence, (5.15) can be rewritten as
ââð,ð=1
â¥ððððâ¥2 <â. (5.16)
Let us omit the obvious reformulations of the last series of statements insetting of quadratic operators, cf. Corollary 5.4.
The condition ð â ðð,ð is not necessary for the structural matrices ð of contin-uous multiplications in ðð. We show this for ð = ð = 2, i.e., in the case of Corollary5.9. To this end it suffices to consider the algebras of form (2.2) with the structuralcoefficients (5.3). Recall that a linear operator ð in a Hilbert space is said to be aHilbertâSchmidt operator if
â¥ð â¥ð»ð â¡ (
ââð=1
â¥ðððâ¥2)1/2 <â
for an (and then for every) orthonormal basis (ðð)âð=1. All HilbertâSchmidt opera-
tors are continuous since â¥ð ⥠†â¥ð â¥ð»ð . However, for example, the unit operatoris not HilbertâSchmidt. In addition to Proposition 3.5 we have
Proposition 5.11. Let ð be a linear operator in a Hilbert space ð» with an orthonor-mal basis (ðð)
âð=1. For the algebra (2.2) the condition (5.15) is fulfilled for all linear
continuous functionals ð if and only if ð is a HilbertâSchmidt operator.
Proof. Let us deal with the equivalent condition (5.16). The functional ð[ð¥] canbe represented as the inner product âšð¥, ââ© with an â â ð» . Accordingly,4 â¥ððððâ¥2 = â¥âšðð, ââ©ððð + âšðð, ââ©ðððâ¥2
†â£âšðð, ââ©â£2 â¥ðððâ¥2 + â£âšðð, ââ©â£2 â¥ðððâ¥2 + 2â£âšðð, ââ©âšâ, ððâ©â£ â¥ððð⥠â¥ððð⥠,
Quadratic Operators 211
whence
2
ðâð,ð=1
â¥ððððâ¥2 â€ðâ
ð=1
â£âšðð, ââ©â£2ðâ
ð=1
â¥ðððâ¥2 +( ðâ
ð=1
â£âšðð, ââ©â£ â¥ðððâ¥)2
. (5.17)
Applying the CauchyâBunyakovski inequality to the last term in (5.17) we obtainðâ
ð,ð=1
â¥ððððâ¥2 â€ðâ
ð=1
â£âšðð, ââ©â£2ðâ
ð=1
â¥ðððâ¥2 †â¥ââ¥2ðâ
ð=1
â¥ðððâ¥2 .
Therefore,ââ
ð,ð=1
â¥ððððâ¥2 †â¥ð â¥2ð»ð â¥ââ¥2 <â
if ð is a HilbertâSchmidt operator. Conversely, if ð is not a HilbertâSchmidtoperator then starting with â = ð1 we get
2
ðâð,ð=1
â¥ððððâ¥2 =ðâ
ð=1
â¥ðððâ¥2 + â¥ðð1â¥2 ââ
as ðââ. â¡
Now we consider an integral counterpart of the structural matrix. This is thekernel ðŸ(ð , ð¡;ð¢) of the multiplication
(ð â ð)(ð¢) =â«â
â«â
ðŸ(ð , ð¡;ð¢)ð(ð )ð(ð¡) dð dð¡, ð¢ â â. (5.18)
In order to realize this formal construction we assume that, at least, the kernel ismeasurable and the measurable functions ð , ð run over a linear spaceðž such that in(5.18) the integrand belongs to ð¿1(â
2), and the integral belongs to ðž. If ðŸ(ð , ð¡;ð¢)is symmetric in ð , ð¡ ( âsymmetricâ for brevity) then the multiplication (5.18) iscommutative. For ð = ð we get the quadratic integral operator ððŸ : ðž â ðž:
(ððŸð)(ð¢) =
â«â
â«â
ðŸ(ð , ð¡;ð¢)ð(ð )ð(ð¡) dð dð¡, ð¢ â â, (5.19)
and here the kernel ðŸ can be changed for a symmetric one: ððŸ = ðᅵᅵ where
ᅵᅵ(ð , ð¡;ð¢) =1
2(ðŸ(ð , ð¡;ð¢) +ðŸ(ð¡, ð ;ð¢)) .
In role of ðž one can consider ð¿ð(â), 1 †ð †â.Theorem 5.12. If the function ðŸð¢(ð , ð¡) = ðŸ(ð , ð¡;ð¢) belongs to ð¿ð(â
2) for almostevery ð¢ and the function ð ð¢ = â¥ðŸð¢â¥ð of ð¢ belongs to ð¿ð(â) then the multiplication
(5.18) is defined and continuous in ð¿ð(â). Moreover,
â¥ð â ðâ¥ð †ð ð,ð â¥ðâ¥ð â¥ðâ¥ð (5.20)
where ð ð,ð is the ð¿ð-norm of ð ð¢.
Proof. The same as for Theorem 5.7 but with integrals instead of sums. â¡
212 Yu.I. Lyubich
Note that the conditions of Theorem 5.12 are symmetric since such isðŸð¢(ð , ð¡).
Corollary 5.13. Under conditions of Theorem 5.12 the quadratic operator (5.19)is defined and continuous in ð¿ð(â). Moreover,
â¥ððŸðâ¥ð †ð ð,ð â¥ðâ¥2ð . (5.21)
Corollary 5.14. If ðŸ(ð , ð¡;ð¢) is a HilbertâSchmidt kernel, i.e.,
ð 22,2 â¡â«â
â«â
â«â
â£ðŸ(ð , ð¡;ð¢)â£2 dð dð¡ dð¢ <â,
then in ð¿2(â) the quadratic operator (5.19) is defined and continuous. Moreover,
â¥ððŸðâ¥2 †ð 2,2 â¥ðâ¥22 . (5.22)
It is interesting to compare this result to Corollary 5.9. Let (ðð(ð¢))âð=1 be an
orthonormal basis in ð¿2(â). Then (ðð(ð )ðð(ð¡)ðð(ð¢))âð,ð,ð=1 is an orthonormal basis
in ð¿2(â3). The HilbertâSchmidt kernelðŸ(ð , ð¡;ð¢) belongs to ð¿2(â
3) by assumption.
If its coordinates are ðððð then for all ð, ð â ð¿2(â) the coordinates of ð â ð are
ðð =ââ
ð,ð=1
ðððððððŸð
where ðð and ðŸð are the coordinates of ð and ð. Hence, [ðððð]âð,ð,ð=1 is the structural
matrix of this algebra. Moreover, ð2,2 = ð 2,2 by Parsevalâs equality. Thus, (5.15)is equivalent to that ðŸ(ð , ð¡;ð¢) is the HilbertâSchmidt kernel.
An important case is a kernelðŸ(ð , ð¡;ð¢) with finite support, say,ðŸ(ð , ð¡;ð¢) = 0outside the cube 0 †ð , ð¡, 𢠆1, so, accordingly,
(ð â ð)(ð¢) =â« 1
0
â« 1
0
ðŸ(ð , ð¡;ð¢)ð(ð )ð(ð¡) dð dð¡, 0 †ð¢ †1. (5.23)
This yields the following important Banach algebra [15].
Proposition 5.15. With a continuous kernel ðŸ(ð , ð¡;ð¢), 0 †ð , ð¡, 𢠆1, the multi-plication (5.23) is defined and continuous in ð¶[0, 1]. Moreover,
â¥ð â ð⥠†ð(ðŸ) â¥ð⥠â¥ð⥠, ð(ðŸ) = maxð¢â[0,1]
â« 1
0
â« 1
0
â£ðŸ(ð , ð¡;ð¢)⣠dð dð¡. (5.24)
Proof. For ð , ð â ð¶[0, 1] we have ð â ð â ð¶[0, 1] since ðŸ(ð , ð¡;ð¢) is uniformlycontinuous. The rest is obvious. â¡
The symmetric continuous kernels ðŸ(ð , ð¡;ð¢) form a subspace ð¿ â ð¶([0, 1]3).We endow it with the norm ð(ðŸ) weaker than the standard one.
Corollary 5.16. The linear mapping ðœ : ðŸ ï¿œâ ððŸ is a continuous embedding of theð -normed space ð¿ into the space of continuous quadratic operators in ð¶[0, 1].
Proof. ðœ is injective since if ððŸ = 0 then ð â ð = 0 for all ð , ð â ð¶[0, 1], andthen ðŸ = 0 because of the completeness of the set of products ðð in ð¶([0, 1]2).Furthermore, â¥ððŸâ¥ †ð(ðŸ) because of (3.5) and (5.24), so ðœ is continuous. â¡
Quadratic Operators 213
6. Compact quadratic operators and algebras
The compactness of some linear and nonlinear operators is a powerful tool in thefunctional analysis, especially in the theory of integral and differential equations,see, e.g., [17] and the references therein. According to [15] a commutative Banachalgebra is called compact if such is the corresponding quadratic operator. In turn,a quadratic operator ð in a Banach space ð is called compact if the image ofa ball ðµð is relatively compact. In this case the image of every bounded set isrelatively compact. From Proposition 3.3 it follows that every compact quadraticoperator is continuous. Every continuous quadratic operator ð of finite rank iscompact. Therefore, every commutative Banach algebra ð with finite-dimensionalð â² is compact.
We denote by ð¶ðð the linear space of all compact quadratic operators, soðµð¹ðð â ð¶ðð â ðµðð . The subspace ð¶ðð is closed. Furthermore, we have
Theorem 6.1. Let a Banach space ð has a Schauder basis (ðð)ðð=1. Then the sub-
space ðµð¹ðð is dense in ð¶ðð .
Proof. If ð <â then ðµð¹ðð = ð¶ðð . Let ð =â, let ð â ð¶ðð , and let
ððð¥ =
ðâð=1
ðâð [ð¥]ðð , ð¥ â ð, ð ⥠1.
Then
(ððð )ð¥ =
ðâð=1
ðâð [ð ð¥]ðð ,
so, ððð is a continuous quadratic operator at most of rank ð. For ðââ we haveððð¥â ð¥ for all ð¥ â ð . Hence, (ððð )ð¥â ð ð¥ for all ð¥. This convergence is uniformon the ball ðµ1 since ð ðµ1 is relatively compact. Thus, â¥ððð â ð ⥠â 0. â¡
Remark 6.2. A similar theorem for the compact linear operators is well known,and the proof is the same. However, there exists a Banach space in which not everycompact linear operator can be approximated by continuous linear operators offinite rank [5]. The question arises: are the approximation properties for the linearand for the quadratic operators equivalent?
Now let us turn to the quadratic integral operators. For such an operator ofrank ð its symmetric kernel is called of rank ð as well.
Theorem 6.3. In the space ð¶[0, 1] general form of a symmetric continuous kernelof a finite rank ð is
ðŸ(ð , ð¡;ð¢) =ðâ
ð=1
ðŸð(ð , ð¡)ðð(ð¢) (6.1)
where all functions ðð(ð¢) are continuous, all partial kernels ðŸð(ð , ð¡) are symmetricand continuous, and the systems (ðð)
ðð=1 and (ðŸð)
ðð=1 are both linearly independent.
214 Yu.I. Lyubich
Proof. Every functionðŸ(ð , ð¡;ð¢) of form (6.1) with continuous ðð(ð¢) and symmetriccontinuous ðŸð(ð , ð¡) is the symmetric continuous kernel of the operator
(ððŸð)(ð¢) =
ðâð=1
ðð(ð¢)
â« 1
0
ðŸð(ð , ð¡)ð(ð )ð(ð¡) dð dð¡ (6.2)
in the space ð = ð¶[0, 1]. By Theorem 2.7 this is a minimal decomposition of ððŸinto a sum of elementary quadratic operators. Thus, ð(ððŸ) = ð.
Conversely, let a symmetric continuous kernel ðŸ(ð , ð¡;ð¢) be such that
(ððŸð)(ð¢) =ðâ
ð=1
ð£ð [ð ]ðð(ð¢)
where ð = ð(ððŸ ), ðð â ð¶[0, 1] and ð£ð are quadratic functionals on ð¶[0, 1]. ByCorollary 2.8 and the general form of continuous linear functionals on ð¶[0, 1],
ð£ð [ð ] = ðâð [ððŸð ] =
â« 1
0
dðð(ð¢)
â« 1
0
â« 1
0
ðŸ(ð , ð¡;ð¢)ð(ð )ð(ð¡) dð dð¡
where ðð(ð¢) are some functions of bounded variation. This yields (6.1) with
ðŸð(ð , ð¡) =
â« 1
0
ðŸ(ð , ð¡;ð¢) dðð(ð¢), 1 †ð †ð.
The partial kernels ðŸð(ð , ð¡) are symmetric and continuous. They are linearly in-dependent, as well as ðð(ð¢), by Theorem 2.7. â¡
Theorem 6.4. In the space ð¶[0, 1] any quadratic integral operator
(ððŸð)(ð¢) =
â« 1
0
â« 1
0
ðŸ(ð , ð¡;ð¢)ð(ð )ð(ð¡) dð dð¡, 0 †ð¢ †1, (6.3)
with continuous kernel ðŸ(ð , ð¡;ð¢) is compact.
Proof. Let ð â ð¶[0, 1], â¥ð⥠†1, and let ð = ð ð . Then
â£ð(ð¢+ ð)â ð(ð¢)⣠â€â« 1
0
â« 1
0
â£ðŸ(ð , ð¡;ð¢+ ð) âðŸ(ð , ð¡;ð¢)⣠dð dð¡, ð â â,
hence, the set ð ðµ1 is uniformly continuous. Furthermore, ð ðµ1 is bounded since
â£ð(ð¢)⣠â€â« 1
0
â« 1
0
â£ðŸ(ð , ð¡;ð¢)⣠dð dð¡, ð â â.
Thus, the set ð ðµ1 is relatively compact by the ArzelaâAscoli theorem. â¡
Further all kernels of integral operators in ð¶[0, 1] are assumed continuous.
Theorem 6.5. If ðŸ(ð , ð¡;ð¢) is a HilbertâSchmidt kernel then in ð¿2(â) the quadraticoperator (5.19) is compact.
Quadratic Operators 215
Proof. According to Theorem 20 from [3, Chapter 4], we have to check the followingthree properties of the set ðº = {ð ð : â¥ð⥠†1}.
(1) ðº is bounded. This is obvious since ð is bounded.(2) If ð > 0, âð(ð¢) = 0 for â£ð¢â£ < ð and âð(ð¢) = 1 for â£ð¢â£ ⥠ð then
supðâðº â¥ðâð⥠tends to zero as ðââ. This follows from the inequality
â¥ðâðâ¥2 â€â«â
â«â
â«â
â£ðŸ(ð , ð¡, ð¢)â£2 âð(ð¢) dð dð¡ dð¢
since the integrand pointwise tends to zero under the ð¿1-majorant â£ðŸ(ð , ð¡, ð¢)â£2.(3) If ð â â and ðð (ð¢) = ð(ð¢ + ð) then supðâðº â¥ðð â ð⥠tends to zero as
ð â 0. This follows from the inequality
â¥ðð â ðâ¥2 â€â«â
â«â
â«â
â£ðŸ(ð , ð¡;ð¢+ ð)âðŸ(ð , ð¡;ð¢)â£2 dð dð¡ dð¢
the right-hand side of which can be rewritten asâ«â
â«â
â«â
â£â£â£ï¿œï¿œ(ð, ð, ð)(ðð<ð,ð> â 1)â£â£â£2 dð dð dð
by Parsevalâs equality for the FourierâPlansherel transform ðŸ ï¿œâ ᅵᅵ. â¡
Remark 6.6. Theorems 6.4 and 6.5 can be extended to the operators of form
(ððŸð)(ð¢) =
â«ð
â«ð
ðŸ(ð , ð¡, ð¢)ð(ð )ð(ð¡) dð dð¡, ð¢ â ð, (6.4)
in the space ð¶(ð) on a compact topological space ð with a Radon measure dð or,respectively, in ð¿2(ð) on a locally compact Abelian group ð with a Haar measure.
7. Baric and Bernstein algebras and quadratic operators
We start with a purely algebraic theory originated from the population genetics.A scalar function ð â= 0 on an algebra ð is called a weight if it is a linear multi-plicative functional or, the same, a homomorphism of ð into the ground field. Ifa weight exists and fixed then ð is called a baric algebra [6]. More rigorously, abaric algebra is a pair (ð,ð) where ð is an algebra and ð is a weight on ð . Thehyperplane ð»0
ð = kerð = {ð¥ â ð : ð[ð¥] = 0} is an ideal (the barideal) of ð .With ð[ð] = 1 the algebra (2.5) is baric with the weight ð. With ð â= 0 the
algebra (2.2) is baric with the weight ð if and only if ð[ðð¥] = ð[ð¥] for all ð¥ â ð .A linear functional ð â= 0 is called a weight of a quadratic operator ð if
ð[ð ð¥] = ð2[ð¥], ð¥ â ð. (7.1)
Lemma 7.1. For a quadratic operator ð and for the corresponding commutativealgebra ð the sets Ωð and Ωð of weights coincide.
Proof. Obviously, Ωð â Ωð . The converse inclusion follows by polarization. â¡
The following is a refinement of Proposition 2.2.
216 Yu.I. Lyubich
Proposition 7.2. If ð = ð(ð ) <â and ð â Ωð then
ð ð¥ = ð2[ð¥]ð1 +
ðâð=2
ð£ð [ð¥]ðð , ð¥ â ð, (7.2)
where ð1 â Imð , ð[ð1] = 1, (ðð)ðð=2 is a basis in ð(ð ) â©ð»0
ð, ð£ð [ð¥] are quadratic
functionals linearly independent along with ð2[ð¥].
Proof. In the proof of Proposition 2.2 one can take any ð1 = ð ð0 with ð[ð0] = 1.Then ð[ð1] = ð[ð ð0] = ð2[ð0] = 1. Let (ðð)
ðð=2 be a basis in ð(ð ) â© ð»0
ð, and
let (ðâð )ðð=2 be its dual basis extended to a basis in ð(ð )
â by setting ðâð [ð1] = 0,2 †ð †ð, and by joining ðâ1 = ð. In the corresponding decomposition of ð wehave ð£1[ð¥] = ð
â1[ð ð¥] = ð[ð ð¥] = ð
2[ð¥]. â¡
An important example is the operator (6.3) in the space ð¶[0, 1] under condi-tion â« 1
0
ðŸ(ð , ð¡;ð¢) dð¢ = 1, 0 †ð , ð¡ †1. (7.3)
In this case the functional
ð[ð ] =
â« 1
0
ð(ð¢) dð¢ (7.4)
is a weight. Then
ðŸ(ð , ð¡;ð¢) = ð1(ð¢) +
ðâð=2
ðŸð(ð , ð¡)ðð(ð¢) (7.5)
where
ð1(ð¢) = (ððŸ1)(ð¢) =
â« 1
0
â« 1
0
ðŸ(ð , ð¡;ð¢) dð dð¡,
â« 1
0
ðð(ð¢) dð¢ = 0, 2 †ð †ð,(7.6)
the sets {ðð(ð¢)}ðð=2 and {ðŸð(ð , ð¡)}ðð=2 ⪠{1(ð , ð¡)} are both linearly independent.A commutative baric algebra (ð,ð) is called a Bernstein algebra if
(ð¥2)2 = ð2[ð¥]ð¥2, ð¥ â ð. (7.7)
For example, the baric algebra (2.5) is Bernstein. The baric algebra (2.2) is Bern-stein if and only if ð 2 = ð , i.e., ð is a projection. (In both cases ð = ð.)
In terms of the Bernstein quadratic operators the identity (7.7) is
ð 2ð¥ = ð2[ð¥]ð ð¥, ð¥ â ð, (7.8)
or, equivalently, if ð1 = ð â£ð»1ð where ð»1
ð = {ð¥ â ð : ð[ð¥] = 1} thenð 21 = ð1. (7.9)
Lemma 7.3. Every Bernstein quadratic operator (or Bernstein algebra) with aweight ð has a fixed point (an idempotent, respectively) ð such that ð[ð] = 1.
Proof. Take ð = ð ð¥ with ð[ð¥] = 1. Then ð[ð] = 1 and ð ð = ð 2ð¥ = ð ð¥ = ð. â¡
Quadratic Operators 217
Below ð is an idempotent, ð[ð] = 1. Every vector ð¥ â ð can be uniquelyrepresented as
ð¥ = ð[ð¥]ð + ð§, ð§ â ð»0ð. (7.10)
In the barideal ð»0ð we have the linear operator ððð¥ = 2ðð¥.
Lemma 7.4. The operator ðð is a projection.
Proof. For ð[ð¥] = 0 the line ð+ð¡ð¥, ð¡ â â, belongs to ð»1ð. Hence, ((ð+ð¡ð¥)
2)2 = (ð+ð¡ð¥)2, i.e., ð+4ð¡ð(ðð¥)+ â â â = ð+2ð¡ðð¥+ â â â . Thus, 4ð(ðð¥) = 2ðð¥, i.e., ð 2
ð = ðð. â¡
As a consequence, we have the direct decomposition
ð»0ð = ð âð (ð = Imðð, ð = kerðð). (7.11)
As a result,
ð = Span{ð} â ð âð. (7.12)
The following lemma is an immediate corollary of Theorem 3.4.8 from [14].
Lemma 7.5. For every ð¥ â ð the ð -component of ð¥2 is ð¢2 with an ð¢ â ð .
Now let ð be a Banach space, and let ð â= 0 be a continuous linear functional.With the weight ð we consider a continuous Bernstein quadratic operator and thecorresponding BanachâBernstein algebra [15].
Theorem 7.6. A BanachâBernstein algebra ð is compact if and only if the subal-gebra ð â² is finite dimensional.
Proof. âIfâ is obvious. For the âonly ifâ note that the linear operator ðð is compactsince such is ð and
ððð¥ =1
2(ð (ð¥+ ð)â ð (ð¥â ð))
Since ðð is a projection and ððâ£ð = 1, the subspace ð is finite dimensional. Let(ð¢ð)
ðð=1 be its basis. By decomposition (7.12) and Lemma 7.5 we have
ð â² â Span{ð, ð¢1, . . . , ð¢ð, ð¢21, . . . , ð¢2ð , ð¢1ð¢2, . . . , ð¢ðâ1ð¢ð}. â¡
The operator form of Theorem 7.6 is
Theorem 7.7. A continuous Bernstein quadratic operator is compact if and only ifit is of finite rank.
Combining this result with Lemma 7.3 and Proposition 7.2 we obtain
Corollary 7.8. Every compact Bernstein quadratic operator is of form
ð ð¥ = ð2[ð¥]ð1 +
ðâð=2
ð£ð [ð¥]ðð , ð¥ â ð, (7.13)
where ð = ð(ð ) <â, ð ð1 = ð1, ð[ð1] = 1, (ðð)ðð=2 is a basis in ð(ð ) â©ð»0
ð, ð£ð [ð¥]
are continuous quadratic functionals linearly independent along with ð2[ð¥].
218 Yu.I. Lyubich
In addition,
ð£ð [ð ð¥] = ð2[ð¥]ð£ð [ð¥], 2 †ð †ð. (7.14)
This follows by ð¥ ï¿œâ ð ð¥ in (7.13) and by applying (7.8) to the result.Theorem 7.7 is applicable to the Bernstein quadratic integral operators in
ð¶[0, 1] since they are compact by Theorem 6.4. The standard weight in this caseis (7.4). Thus, all Bernstein kernels are of form (7.5) where the functions ðð andðŸð(ð , ð¡) are such as aforesaid.
The following theorem announced in [1] was proved for the first time in [16]applying Theorem 7.7. The latter (as well as Theorem 7.6) was proved in [15].
Theorem 7.9. If ðŸ(ð , ð¡;ð¢) > 0 is a Bernstein kernel then ðŸ only depends on ð¢.
In other words,ðŸ is of rank 1. Actually, in [16] the assumption is ðŸ(ð , ð¡;ð¢) â¥0 with ðŸ(ð , ð ;ð¢) > 0. Also note that all our results on continuous kernels remaintrue for ð¶(ð, ð) where ð is a compact and ð is a Radon measure, suppð = ð.
References
[1] S.N. Bernstein. Solution of a mathematical problem related to the theory of inheri-tance, Uchenye Zapiski n.-i. kafedr Ukrainy 1 (1924), 83â115 (in Russian).
[2] N. Bourbaki. Elements de mathematique, Algebre, Ch. 9: Formes sesquilineaires etformes quadratiques, Hermann, Paris, 1959 (in French).
[3] N. Danford and J.T. Schwartz. Linear operators, Part 1: General theory, Intersci.Publ., 1958.
[4] M. Day. Normed linear spaces, Springer-Verlag, 1958.
[5] P. Enflo. A counterexample to the approximation problem in Banach spaces, ActaMath. 130 (1973), 309â317.
[6] I.M.H. Etherington. Genetic algebras, Proc. Roy. Soc. Edinburgh, A 59 (1939), 242â258.
[7] V.I. Glivenko. Mendelian algebra, Doklady Akad. Nauk SSSR 8, No.4 (1936), 371â372 (in Russian).
[8] J.C. Gutierrez. Solution of the Bernstein problem in the non-regular case, J. Algebra223, No.1 (2000), 226â132.
[9] P. Holgate. Genetic algebras satisfying Bernsteinâs stationarity principle. J. LondonMath. Soc. (2) 9 (1975), 612â624.
[10] Yu.I. Lyubich. Basis concepts and theorems of evolutionary genetics for free popu-lations, Russian Math. Surveys 26, No. 5 (1971), 51â123.
[11] Yu.I. Lyubich, Two-level Bernstein populations, Math. USSR Sb. 24, No. 1 (1974),593â615.
[12] Yu.I. Lyubich. Proper Bernstein populations, Probl. Inform. Transmiss. (1978), 228â235.
[13] Yu.I. Lyubich. Bernstein algebras, Uspekhi Mat. Nauk 32, No. 6 (1977), 261â263 (inRussian).
Quadratic Operators 219
[14] Yu.I. Lyubich. Mathematical structures in population genetics, Springer-Verlag,1992. (Translated from the Russian original, 1983, Naukova Dumka.)
[15] Yu.I. Lyubich. BanachâBernstein algebras and their applications. Nonassociativealgebra and its applications, Lecture Notes in Pure and Appl. Math., 211, Dekker,2000, pp. 205â210.
[16] Yu.I. Lyubich. A theorem on Bernstein quadratic integral operators, Nonassociativealgebra and its applications, Lect. Notes Pure and Appl. Math., 246, ChapmanâHall,2006, pp. 245â251.
[17] L. Nirenberg. Topics in nonlinear functional analysis, Courant Inst, 1974.
[18] O. Reiersol. Genetic algebras studied recursively and by means of differential oper-ators, Math. Scand., 10 (1962), 25â44.
[19] A.S. Serebrovskii. Properties of Mendelian equalities, Doklady Akad. Nauk SSSR 2,No. 1 (1934), 33â36.
[20] R.D. Schafer. An introduction to nonassociative algebras, Acad. Press, 1966.
[21] A. Worz-Busekros. Algebras in Genetics, Springer-Verlag, 1980.
Yu.I. LyubichDepartment of MathematicsTechnion, 32000Haifa, Israele-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 221â239câ 2013 Springer Basel
Strong Stability of Invariant Subspacesof Quaternion Matrices
Leiba Rodman
Dedicated to Leonid Lerer on occasion of his 70th birthday
Abstract. Classes of matrices and their invariant subspaces with various ro-bustness properties are described, in the context of matrices over real quater-nions and quaternionic subspaces. Robustness is understood in the sense ofbeing close to the original invariant subspace under small perturbation of thematrix.
Mathematics Subject Classification (2010). 15B33.
Keywords. Real quaternions, matrices over quaternions, invariant subspaces,robustness, stability.
1. Introduction
In the present paper we study certain classes of invariant subspaces of matricesthat change little (in various senses) after small perturbations of the matrix. Suchclasses have been much studied in recent 35 years, for complex and real matrices,starting with the grounbreaking papers [3, 5]. The literature on the subject isextensive, and we mention here only chapters and parts in books [2, Chapter 8],[6, Chapter I.5], [7, Chapter 15], [8, Chapter II.4], [4, Part 4].
In [13], the results on stable and ðŒ-stable invariant subspaces for complexmatrices have been extended to matrices over the real quaternions and quater-nionic invariant subspaces. An ðŽ-invariant subspace is said to be stable if, looselyspeaking, every matrix ðµ close to ðŽ has an invariant subspace ð© close to â³;ðŒ-stability is characterized by the stronger property that the gap between ð© andâ³ is a Holder function of â¥ðµ â ðŽâ¥ of exponent ðŒ. (See the precise definitions inSection 2.)
Here, we continue this line of investigation. We extend to quaternionic ma-trices the characterizations of strongly stable and strongly ðŒ-stable invariant sub-spaces of complex matrices obtained in [12]. Strong stability means that stability
222 L. Rodman
holds for every ðµ-invariant subspace ð© (in the above notation), provided the ob-vious dimensionality conditions are satisfied, and similarly for strong ðŒ-stability.From the standpoint of basic backward error analysis, the notion of strong sta-bility is perhaps more appropriate than the notion of stability as a measure ofrobustness of invariant subspaces of matrices.
Besides the introduction, the paper consists of five sections. In Section 2 wepresent some known results on matrix analysis for quaternionic matrices, for thereadersâ benefit. Our main Theorems 3.3 and 6.1 are stated in Sections 3 and 6respectively. We prove Theorem 3.3 in Sections 4 and 5.
2. Quaternionic linear algebra
In this section we review briefly some known facts in linear algebra over the skewfield of real quaternions H, and introduce notation to be used throughout thepaper.
Denote by i, j, k the standard quaternion imaginary units. For ð¥ â H, ð¥ = ð0+ð1i+ð2j+ð3k with ðð â R (the real field), let â(ð¥) = ð0 and ð(ð) = ð1i+ð2j+ð3kbe the real and the vector part of ð¥, respectively, ð¥â = ð0 â ð1iâ ð2jâ ð3k be theconjugate quaternion, and â£ð¥â£ =âð20 + ð21 + ð22 + ð23 be the length of ð¥.
Let HðÃð (abbreviated to Hð if ð = 1) be the set of ð Ãð matrices withentries in H. The quaternionic conjugation naturally leads to the concept of con-jugate transposed matrices; we denote by ðŽâ â HðÃð the conjugate transposeof a matrix ðŽ â HðÃð. Hð is considered as a right quaternionic vector space,endowed with the quaternion-valued inner product âšð¢, ð£â© = ð£âð¢, ð¢, ð£ â Hð and
norm â¥ð£â¥ =ââšð£, ð£â©, ð£ â Hð. Matrices ðŽ â HðÃð will be also considered as lineartransformations Hð â Hð by way of matrix-vector multiplication ð¥ ï¿œâ ðŽð¥. Weuse the operator norm for matrices:
â¥ðŽâ¥ = max{â¥ðŽð¢â¥ : ð¢ â Hð, â¥ð¢â¥ = 1},where ðŽ â HðÃð. Note that â¥ðŽâ¥ coincides with the largest singular value of ðŽ.
Consider now the complex matrix representation of quaternions. For ð¥ =ð0 + ið1 + jð2 + kð3 â H, ð0, ð1, ð2, ð3 â R, define
ð(ð¥) =
[ð0 + ið1 ð2 + ið3âð2 + ið3 ð0 â ið1
]â C2Ã2,
(CðÃð stands for the set of ðà ð complex matrices ) and extend ð entrywise toa map
ðð,ð : HðÃð â C2ðÃ2ð, ðð,ð
([ð¥ð,ð ]
ð,ðð,ð=1
)= [ð(ð¥ð,ð)]
ð,ðð,ð=1, ð¥ð,ð â H.
We have:
(i) ðð,ð is a unital homomorphism of real algebras;(ii) if ð â HðÃð, ð â HðÃð, then ðð,ð(ðð ) = ðð,ð(ð)ðð,ð(ð );(iii) ðð,ð(ð
â) = (ðð,ð(ð))â, â ð â HðÃð;
Strong Stability of Invariant Subspaces 223
(iv) there exist positive constants ðð,ð, ð¶ð,ð such that
ðð,ðâ¥ðð,ð(ð)⥠†â¥ð⥠†ð¶ð,ðâ¥ðð,ð(ð)⥠(2.1)
for every ð â HðÃð.
Often, we will abbreviate ðð,ð to ð (with ð,ð understood from context).Let ðŽ â HðÃð. A vector ð¢ â Hð â {0} is said to be an eigenvector of ðŽ
corresponding to the eigenvalue ðŒ â H if
ðŽð£ = ð£ðŒ. (2.2)
The set of all eigenvalues of ðŽ is denoted ð(ðŽ), the spectrum of ðŽ. Note that ð(ðŽ)is closed under similarity of quaternions: If (2.2) holds, then ðŽ(ð£ð) = (ð£ð)(ðâ1ðŒð),ð â H â {0}, so ð£ð is an eigenvector of ðŽ corresponding to the eigenvalue ðâ1ðŒð.Note also that the similarity orbit {ðâ1ðŒð : ð â H â {0}} of ðŒ consists exactly ofthose quaternions ð for which â(ð) = â(ðŒ) and â£ð(ð)⣠= â£ð(ðŒ)â£. In particular, ifðŒ â ð(ðŽ), then also ðŒâ â ð(ðŽ).
The Jordan form is valid for quaternionic matrices:
Theorem 2.1. Let ðŽ â HðÃð. Then there exists an invertible ð â HðÃð such thatðâ1ðŽð has the form
ðâ1ðŽð = ðœð1(ð1)â â â â â ðœðð(ðð), ð1, . . . , ðð â H, (2.3)
where ðœð(ð) is the ðÃð (upper triangular) Jordan block having eigenvalue ð. Theform (2.3) is uniquely determined by ðŽ up to an arbitrary permutation of blocksand up to a replacement of ð1, . . . , ðð with ðŒâ11 ð1ðŒ1, . . . , ðŒ
â1ð ðððŒð, respectively,
where ðŒð â H â {0}, ð = 1, 2, . . . , ð.
A proof is given in [16]; the result goes back to [15].We will need the following transformation properties of bases and Jordan
forms under ð:
Proposition 2.2. Let ð¢1, . . . , ð¢ð be a basis (resp., orthogonal basis, orthonormalbasis, spanning set) of a subspace ð° â Hð. Then:
(1) The columns of ð (ð¢1), . . . , ð (ð¢ð) form a basis (resp., orthogonal basis, or-thonormal basis) of the subspace ð (ð°) â C2ð.
(2) The subspace ð (ð°) is independent of the choice of the basis (resp., orthogonalbasis, orthonormal basis) of ð° .
Proof. Part (1) is [13, Proposition 2.1]. If ð¢1, . . . , ð¢ð and ð¢â²1, . . . , ð¢
â²ð are two bases
for ð° , then [ð¢1 . . . ð¢ð ]ð = [ð¢â²1 . . . ð¢â²ð ] for some invertible matrix ð. Applying ð
to this equality, (2) follows. â¡
Proposition 2.3. Let ðŽ â HðÃð, and let ð© of Hð be an ðŽ-invariant subspace.Denote by ðœ the Jordan form of ðŽâ£ð© specialized so that ðœ â CðÃð. Then theJordan form (over the complexes) of ð(ðŽâ£ð© ) is ðœ â ðœ , where the overline standsfor the complex conjugation.
224 L. Rodman
Proof. Let ð¢1, . . . , ð¢ð be a Jordan basis for ðŽâ£ð© in ð© , such thatðŽ [ð¢1 ð¢2 . . . ð¢ð] = [ð¢1 ð¢2 . . . ð¢ð] ðœ. (2.4)
By Proposition 2.2, the columns of [ð(ð¢1), ð(ð¢2), . . . , ð(ð¢ð)] form a basis forð(ð© ) (cf. [13, Proposition 2.4]). It remains to apply the map ð to (2.4). â¡
The distance between two subspaces in Hð will be measured by the gap.Define the gap between two subspacesâ³ and ð© of Hð by
ð(â³,ð© ) = â¥ðâ³ â ðð© â¥,where ðð³ is the orthogonal projection on a subspace ð³ . If {ð¢1, . . . , ð¢ð} is anorthonormal basis for ð³ , then
ðð³ = [ð¢1 ð¢2 . . . ð¢ð][ð¢1 ð¢2 . . . ð¢ð]â. (2.5)
The gap is a metric on the set of all subspaces of Hð that turns the set intoa compact complete metric space; this is well known in the context of complexsubspaces, and can be proved for quaternion subspaces in the same way (see, e.g.,the proof of [7, Theorem 13.4.1]). Many basic properties of subspaces in Hð as theyrelate to the gap metric and are familiar in the setting of subspaces of the realvector space Rð or of Cð, remain valid in the setting of quaternion subspaces, withessentially the same proofs, for example, [7, Theorems 13.1.1, 13.1.2, 13.1.3, 13.4.1,13.4.2, 13.4.3]. Some are proved with complete details in [13, Theorem 2.11]. Wewill use these properties in the sequel as necessary, and present here only a fewof them (Theorem 2.4 below; parts (1) and (3) are standard, (2) is proved in [13],and a short proof of (4) is supplied). We denote by ð(ð¥, ð) = infð¡âð â¥ð¥ â ð¡â¥ thedistance from ð¥ â Hð to a set ð â Hð.
Theorem 2.4.(1) If ð(â³,ð© ) < 1, then dimâ³ = dimð© . (The dimensions of subspaces in Hð
are understood in the quaternionic sense.)(2) Assume â³â²,â³ are subspaces of Hð such that the sum â³â²+â³ is direct.
Then there exists ð¿ > 0 (which depends on â³ and â³â² only) such that, ifð© ,ð© â² are subspaces of Hð and
max{ð(â³,ð© ), ð(â³â²,ð© â²)} †ð¿, (2.6)
then the sum ð© â²+ð© is also direct.Assume in addition that â³â²+â³ = Hð. Then there exists ð¿1 > 0 such
that if (2.6) holds (with ð¿ replaced by ð¿1) for subspaces ð© ,ð© â² of Hð, thenð© â²+ð© = Hð and
max{ð(â³,ð© ), ð(â³â²,ð© â²)} †â¥ðâ³,â³â² â ðð© ,ð© â²â¥ (2.7)
â€(4(1 + 2â¥ðâ³,â³â²â¥) max
ð¥ââ³â², â¥ð¥â¥=1
{ð(ð¥,â³)â1
})à (ð(â³,ð© ) + ð(â³â²,ð© â²)) , (2.8)
where the matrix ðâ³,â³â² projects Hð onto â³ along â³â², whereas ðð© ,ð© â²
projects Hð onto ð© along ð© â².
Strong Stability of Invariant Subspaces 225
(3) For all subspaces ð©1,ð©2 â Hð and all invertible matrices ð â HðÃð, theinqualities(â¥ð⥠â¥ðâ1â¥)â1 ð (ð©1,ð©2) †ð (ðð©1, ðð©2) †â¥ð⥠â¥ðâ1â¥ð (ð©1,ð©2)
hold.(4) Let ð¬1+ð¬2 = Hð, a direct sum of subspaces. Then there exists a constant
ð¶ > 0 that depends on ð¬1,ð¬2 only such that
ð(â³1,â³2) †ð¶ð(â³1+ð¬2,â³2+ð¬2)
for every pair of subspaces â³1,â³2 â ð¬1.
Proof of (4). Let ð â HðÃð be an invertible matrix such that ðð¬1 and ðð¬2 areorthogonal. Then for subspacesâ³1,â³2 â ð¬1 the equality
ð(ðâ³1, ðâ³2) = ð(ðâ³1+ðð¬2, ðâ³2+ðð¬2) (2.9)
is obvious. Now (3) gives
ð(â³1,â³2) †â¥ð⥠â¥ðâ1⥠ð (ðâ³1, ðâ³2) †by (2.9)
†(â¥ð⥠â¥ðâ1â¥)2 ð (â³1+ð¬2,â³2+ð¬2). â¡
The complex representation ð keeps the gap between subspaces within uni-versal bounds (for fixed ð):
ðð,ðð (ð (ð°), ð (ð±)) †ð (ð° ,ð±) †ð¶ð,ðð (ð (ð°), ð (ð±)) (2.10)
for all subspaces ð° ,ð± â Hð, where the positive constants ðð,ð, ð¶ð,ð are taken from(2.1). Indeed, letting ð¢1, . . . , ð¢ð and ð£1, . . . , ð£â be orthonormal bases for ð° and ð± ,respectively, we have (see Proposition 2.2)
ð(ð° ,ð±) = â¥[ð¢1 . . . ð¢ð ]â[ð¢1 . . . ð¢ð ]â [ ð£1 . . . ð£â ]â[ ð£1 . . . ð£â ]⥠,ð(ð (ð±), ð (ð±)) = â¥[ð (ð¢1) . . . ð (ð¢ð) ]â[ð (ð¢1) . . . ð (ð¢ð) ]
â [ð (ð£1) . . . ð (ð£â) ]â[ð (ð£1) . . . ð (ð£â) ]â¥,and so (2.10) follows from (2.1).
The rank of ðŽ â HðÃð is, by definition, the dimension of the range of ðŽ, orequivalently, the dimension of the column space (understood as a right quaternionvector space) of ðŽ.
For ðŽ â HðÃð, the pseudoinverse, or MooreâPenrose inverse, is defined as amatrix ðŽ+ â HðÃð that is the unique solution of the following system of equations:
ðŽðŽ+ðŽ = ðŽ, ðŽ+ðŽðŽ+ = ðŽ+, (ðŽðŽ+)â = ðŽðŽ+, (ðŽ+ðŽ)â = ðŽ+ðŽ.
Let
ð¯ð,ð,ð := {ðŽ â HðÃð : rankðŽ = ð}, ð = 1, 2, . . . ,min{ð,ð}be sets of quaternion matrices of fixed rank.
226 L. Rodman
Theorem 2.5. The pseudoinverse is a (local) Lipschitz function on each of the setsð¯ð,ð,ð, namely: Given ðŽ â HðÃð, there exist positive constants ð¿, ðŸ dependingon ðŽ only such that
â¥ðµ+ âðŽ+⥠†ðŸâ¥ðµ âðŽâ¥for all ðµ â ð¯ð,ð,ð with â¥ðµ âðŽâ¥ †ð¿.
For a proof see [16, 14] for the complex matrices; it can be extended easilyto quaternion matrices using the complex representation ð.
Formulas for orthogonal projections on intersections and sums of subspacescan be given in terms of the relevant pseudoinverses:
Proposition 2.6. Let â³,ð© be subspaces in Hð. Then
ðâ³â©ð© = 2ðð© (ðð© +ðâ³)+ðâ³ and ðâ³+ð© = (ðâ³+ðð© )(ðâ³+ðð© )+. (2.11)
This result is proved in [9] for complex matrices; extension to quaternionmatrices is immediate (same proofs apply). For the second formula in (2.11) seealso [1].
Next, we discuss briefly root subspaces of quaternionic matrices. Let ðŽ âHðÃð, and let ð(ðŽ)(ð¥) be the minimal polynomial with real coefficients and leadingcoefficient 1 for ðŽ; in other words, ð(ðŽ)(ð¥) is the monic real polynomial of minimaldegree such that ð(ðŽ)(ðŽ) = 0. One easily verifies that the (real and complex) rootsof ð(ðŽ)(ð¥) are exactly the eigenvalues of ðŽ that belong to C. Write
ð(ðŽ)(ð¥) = ð1(ð¥)ð1 â â â ðð(ð¥)ðð ,
where the ðð(ð¥)âs are distinct monic irreducible real polynomials (i.e., of the formð¥â ð, ð real, or of the form ð¥2+ ðð¥+ ð, ð, ð â R, with no real roots), and the ðð âsare positive integers. The subspace
â³ð := {ð¢ : ð¢ â Hð, ðð(ðŽ)ððð¢ = 0}, ð = 1, 2, . . . , ð,
is called the root subspace of ðŽ corresponding to the roots of ðð(ð¥). Obviously,the root subspaces of ðŽ are ðŽ-invariant. We refer the reader to [13] for someelementary properties of the minimal polynomials and subspaces. In particular,
â³ =âð
ð=1(â³â©â³ð) for every ðŽ-invariant subspace â³. Also, root subspaces,and more generally, their sums, are Lipschitz functions of a matrix:
Proposition 2.7.(a) The roots of ð(ðŽ)(ð¥) depend continuously on ðŽ: Fix ðŽ â HðÃð, and let
ð1, . . . , ðð be all the distinct roots of ð(ðŽ)(ð¥) in the closed upper complexhalf-plane C+. Then for every ð > 0 there exists ð¿ > 0 such that if ðµ â HðÃð
satisfies â¥ðµ â ðŽâ¥ < ð¿, then the roots of ð(ðµ)(ð¥) in C+ are contained in theunion
âªð ð=1{ð§ â C+ : â£ð§ â ðð ⣠< ð}.
(b) Sums of root subspaces of ðŽ are Lipschitz functions of ðŽ: Given ðŽ andð1, . . . , ðð as in part (a), there exist ð¿0,ðŸ0 > 0 such that for every ðµ â HðÃð
satisfying â¥ðµâðŽâ¥ < ð¿0, it holds that if ð is any nonempty subset of ð1, . . . , ðð
Strong Stability of Invariant Subspaces 227
and ð â² is the set of all eigenvalues of ðµ contained in âªðâð {ð§ â C+ : â£ð§âðð ⣠<ð¿0}, then the sum of root subspaces â³â² of ðµ corresponding to ð â²âªð â² and thesum of root subspaces â³ of ðŽ corresponding to ð ⪠ð satisfy the inequality
ð(â,ââ²) †ðŸ0 â¥ðµ âðŽâ¥. (2.12)
For the readerâs convenience, we quote the main result (Theorem 2.8 below)on stability of invariant subspaces from [13], where a complete and detailed proofis given.
Let ðŽ â HðÃð. An ðŽ-invariant subspaceâ³â Hð is called stable if for everyð > 0 there exists ð¿ > 0 such that every ðµ â HðÃð satisfying â¥ðµ â ðŽâ¥ < ð¿ has aðµ-invariant subspace ð© â Hð with the property that ð(â³,ð© ) < ð. For a fixedðŒ ⥠1, an ðŽ-invariant subspace â³ is called ðŒ-stable, if there exist ð¿0,ðŸ0 > 0such that for every ðµ â HðÃð satisfying â¥ðµ â ðŽâ¥ †ð¿0 there exists a ðµ-invariantsubspace ð© with the property that
ð(â³,ð© ) †ðŸ0â¥ðµ âðŽâ¥1/ðŒ. (2.13)
Noting that ð(â³,ð© ) †1 for all subspaces â³,ð© â Hð, an equivalent definitionof ðŒ-stable ðŽ-invariant subspace â³ is obtained by requiring that there existsðŸ â²0 > 0 such that for every ðµ â HðÃð there is a ðµ-invariant subspace ð© satisfyinginequality
ð(â³,ð© ) †ðŸ â²0â¥ðµ âðŽâ¥1/ðŒ. (2.14)
Indeed, if (2.13) holds for all ðµ â HðÃð with â¥ðµâðŽâ¥ †ð¿0, then (2.14) holds withðŸ â²0 = max{ðŸ0, ð¿
â1/ðŒ0 } for all ðµ â HðÃð where in the case â¥ðµ â ðŽâ¥ > ð¿0 we take
any ðµ-invariant subspace for ð© .1-stable ðŽ-invariant subspaces are called Lipschitz stable.For two positive integers ð < ð, define
ð (ð, ð) =
{ð if no ð distinct ðth roots of 1 sum up to zero;ðâ 1 if there are ð distinct ðth roots of 1 that sum up to zero.
(2.15)In the following theorem as well as in later statements, the following property
of a matrix ðŽ â HðÃð and its invariant subspaceâ³ will be used:
(âµ(ðŽ,â³)) the intersection ofâ³ with any root subspace
â³ð of ðŽ corresponding to a real eigenvalue satisfies
dim (â³â©â³ð) †1 or dim (â³â©â³ð) ⥠dimâ³ð â 1. (2.16)
A root subspaceâ³ð of ðŽ â HðÃð is said to be of geometric multiplicity oneif there is only one eigenvector (up to scaling) of ðŽ inâ³ð.
Theorem 2.8. Let ðŽ â HðÃð, and let â³ â= {0} be an ðŽ-invariant subspace. Then:
(a) â³ is Lipschitz stable if and only if â³ is a sum of root subspaces.(b) â³ is stable if and only if for every root subspace â³ð of ðŽ that contains at
least two linearly independent eigenvectors of ðŽ, we have â³â©â³ð = {0} orâ³ð ââ³.
228 L. Rodman
(c) if â³ is ðŒ-stable, then â³ is stable and for every root subspace â³ð of ðŽ ofgeometric multiplicity one and such that {0} â= â³ â©â³ð â= â³ð, we haveðŒ ⥠ð (dim (â³â©â³ð), dimâ³ð).
Conversely, assume (âµ(ðŽ,â³)) holds. Then, if â³ is stable, and if for every rootsubspace â³ð of ðŽ of geometric multiplicity one and such that {0} â=â³â©â³ð â=â³ð, we have ðŒ ⥠ð (dim (â³â©â³ð), dimâ³ð), then â³ is ðŒ-stable.
Remark 2.9. The result of Theorem 2.8 holds for complex matrices and complexinvariant subspaces without the hypothesis (âµ(ðŽ,â³)); see [3, 5, 12].
Open Problem 2.10. Is the converse statement in Theorem 2.8(c) valid under thefollowing hypothesis (âµ0(ðŽ,â³)) which is weaker than (âµ(ðŽ,â³))?
(âµ0(ðŽ,â³)) the intersection ofâ³ with any root subspace
â³ð of geometric multiplicity one of ðŽ corresponding to a real eigenvaluesatisfies
dim (â³â©â³ð) †1 or dim (â³â©â³ð) ⥠dimâ³ð â 1.Analogous question with respect to Theorem 3.3(a) and Theorem 6.1.
3. Main results
Let ðŽ â HðÃð, and let let ð1, . . . , ðð be all the distinct roots of ð(ðŽ)(ð¥) in the
closed upper complex half-plane C+. Select ð0 > 0 so that the disks
ð·ð0(ðð) := {ð§ â C : â£ð§ â ðð ⣠< ð0}, ð = 1, 2, . . . , ð ,
do not intersect. By Proposition 2.7, there exists ð¿0 > 0 such that if ðµ â HðÃð,â¥ðµâðŽâ¥ < ð¿0, then all roots of ð(ðµ)(ð¥) in the closed upper half-plane are containedin âªð
ð=1ð·ð0(ðð).Fix ðŒ ⥠1. An ðŽ-invariant subspaceâ³â Hð is said to be strongly ðŒ-stable
if there exist ð †ð¿0, ðŸ > 0 (that depend on ðŽ and â³ only) with the followingproperty: For every ðµ â HðÃð such that â¥ðµ â ðŽâ¥ < ð, and every ðµ-invariantsubspace ð© such that
dim(ð© â©âð(ðµ)) = dim(â³â©âð(ðŽ)), ð = 1, 2, . . . , ð , (3.1)
the inequality
ð(ð© ,â³) †ðŸâ¥ðµ âðŽâ¥1/ðŒholds; here we denote by âð(ð) the sum of the root subspaces of the matrix ð
corresponding to the roots of its minimal polynomial in the set ð·ð(ðð) âªð·ð(ðð),ð = 1, 2, . . . , ð . Note that ðµ-invariant subspaces ð© such that (3.1) holds do exist;indeed, it follows from the condition â¥ðµ â ðŽâ¥ < ð †ð¿0 and the properties of ð¿0(see the first paragraph of this section) that
dimâð(ðµ) = dimâð(ðŽ), ð = 1, 2, . . . , ð ;
then existence of such ð© becomes obvious.
Strong Stability of Invariant Subspaces 229
When ðŒ = 1, we say that strongly 1-stable ðŽ-invariant subspace is stronglyLipschitz stable. Finally, an ðŽ-invariant subspace â³ â Hð is said to be stronglystable if for every ð > 0 there exists ð¿ > 0 (which can be taken †ð¿0) such thatfor every ðµ â HðÃð satisfying â¥ðµ â ðŽâ¥ < ð¿ and every ðµ-invariant subspace ð©satisfying (3.1) the inequality ð(â³,ð© ) < ð holds. Again, ðµ-invariant subspacesð© satisfying (3.1) do exist (if ð¿ is taken sufficiently small).
Clearly, strong ðŒ-stability (for any ðŒ ⥠1) implies strong stability; the con-verse is generally false (see Theorem 3.3 below). The concepts of strong ðŒ-stabilityand strong stability for invariant subspaces of complex matrices were introducedand studied in [12] and for a particular situation in [10].
Proposition 3.1. If an ðŽ-invariant subspace is strongly ðŒ-stable, then it is alsoðŒ-stable. If an ðŽ-invariant subspace is strongly stable, then it is stable.
Proof. The result follows from the already mentioned fact that the set of ðµ-invariant subspaces ð© for which (3.1) holds is nonempty provided ðµ is sufficientlyclose to ðŽ (cf. Proposition 2.7(b) and Theorem 2.4(1)). â¡
The following characterization of strong ðŒ-stability is proved in [12] for com-plex matrices. The definition of strongly ðŒ-stable invariant subspaces in the contextof complex matrices is analogous to the above definition, with all considerationsrestricted to complex matrices and invariant subspaces in Cð.
Theorem 3.2. Let ðŽ â CðÃð. An ðŽ-invariant subspaceâ³â Cð is strongly ðŒ-stablein the context of complex matrices if and only if for every (complex) eigenvalue ðof ðŽ that has at least two linearly independent associated (complex) eigenvectors,we have
â³â© {ð£ â Cð : (ðŽâ ððŒ)ðð£ = 0} = {0}or
â³â {ð£ â Cð : (ðŽâ ððŒ)ðð£ = 0},and for every (complex) eigenvalue ð of ðŽ that has a unique up to scaling (complex)eigenvector and such that
{0} â=â³â©{ð£ â Cð : (ðŽâ ððŒ)ðð£ = 0} â= {ð£ â Cð : (ðŽâ ððŒ)ðð£ = 0}we have
ðŒ ⥠dim {ð£ â Cð : (ðŽâ ððŒ)ðð£ = 0}.We prove analogous result in the context of quaternion matrices, imposing
the additional condition (âµ(ðŽ,â³)), as necessary:
Theorem 3.3. Let ðŽ â HðÃð, and let â³ be an ðŽ-invariant subspace. Then:
(a) ifâ³ is strongly ðŒ-stable, thenâ³ is strongly stable and for every root subspaceâ³ð of ðŽ of geometric multiplicity one and such that {0} â=â³â©â³ð â=â³ð,we have ðŒ ⥠dimâ³ð.
230 L. Rodman
Conversely, assume property (âµ(ðŽ,â³)) holds. Then, if â³ is strongly stable, andfor every root subspace â³ð of ðŽ of geometric multiplicity one and such that {0} â=â³â©â³ð â=â³ð, we have ðŒ ⥠dimâ³ð, then â³ is strongly ðŒ-stable.
(b) â³ is strongly Lipschitz stable if and only ifâ³ is a sum of root subspaces of ðŽ.(c) â³ is strongly stable if and only if â³ is stable.
Remark 3.4. We do not know whether or not the property (âµ(ðŽ,â³)) is essentialin the converse part of Theorem 3.3(a).
The proof of Theorem 3.3 will be given in the next two sections.
4. Preliminaries for the proof of Theorem 3.3
The following fact is the key; it allows us to reduce the proof to the case of justone root subspace.
Fact 4.1. Let ðŽ â HðÃð, and let â³ be an ðŽ-invariant subspace. Fix ðŒ ⥠1.Then â³ is strongly stable or strongly ðŒ-stable if and only if for every sum of rootsubspaces â â Hð for ðŽ the intersection â³ â© â is strongly stable or stronglyðŒ-stable, respectively, as an ðŽâ£â-invariant subspace.
We provide details of the proof for the case of strong ðŒ-stability only (thecase of strong stability can be proved analogously). The proof will be accomplishedby proving Steps 1, 2, 3, 4 below (often parallel to the proof of [13, Fact 4.2]).
Let ð1, . . . , ðð be all the distinct roots of ð(ðŽ)(ð¥) in C+, the closed upper
complex half-plane. By Proposition 2.7, there exists ð¿0 > 0 such that if ðµ â HðÃð,â¥ðµâðŽâ¥ < ð¿0, then all roots of ð(ðµ)(ð¥) in the closed upper half-plane are containedin âªð
ð=1ð·ð0(ðð), where ð0 > 0 is selected so that the disks ð·ð0(ðð), ð = 1, 2, . . . , ð ,do not intersect.
Assuming ð â HðÃð satisfies â¥ð â ðŽâ¥ < ð¿0, we let âð(ð) be the sum ofroot subspaces of ð corresponding to the roots of the minimal polynomial of ðin ð·ð0(ðð) âªð·ð0(ðð), ð = 1, 2, . . . , ð .
Fix â, a sum of root subspaces for ðŽ, and let â1, . . . ,âð be all the rootsubspaces of ðŽ contained in â. We denote by ðŸ0,ðŸ1, . . . positive constants thatdepend on ðŽ and ð© only.
Step 1. If an ðŽ-invariant subspace ð© â â is strongly ðŒ-stable as a ðŽâ£â-invariantsubspace, then ð© is also strongly ðŒ-stable as an ðŽ-invariant subspace.
Proof of Step 1. Suppose not. Then there exists a sequence {ðð}âð=1, ðð â HðÃð,such that â¥ðð â ðŽâ¥ < ðâ1, ð = 1, 2, . . . , and for some ðð-invariant subspaceð©ð such that
dim (ð©ð â©âð(ðð)) = dim (ð© â©âð(ðŽ)), ð = 1, 2, . . . , ð , (4.1)
we have
ð(ð©ð,ð© ) ⥠ðâ¥ðð âðŽâ¥1/ðŒ, ð = 1, 2, . . . . (4.2)
Strong Stability of Invariant Subspaces 231
Denote
âð =
ðâð=1
âð(ðð), ð = 1, 2, . . . .
For sufficiently large ð, the subspace âð is a direct complement of ââ¥, theorthogonal complement of â in Hð ((2.12) and Theorem 2.4(2)). For such ð, wedefine the linear transformation ðð : Hð â Hð by
ðð =
[ðŒ (âðâð,â⥠+ ðâ)â£â0 ðŒ
],
with respect to the orthogonal decomposition Hð = â⥠â â, where ðâð,â⥠is
the projection on âð along ââ¥. ((âðâð,â⥠+ ðâ)â£â is known as the angular
operator, cf. [2].) Clearly, ððâ⥠= ââ¥. Also, ððâð = â. Nowâ¥ðð â ðŒâ¥ = â¥(âðâð,â⥠+ ðâ)â£â⥠†â¥(âðâð,â⥠+ ðâ)â¥
(by Theorem2.4(2)) †ðŸ0ð (âð,â)(by Proposition 2.7(b)) †ðŸ1â¥ðð âðŽâ¥.
(4.3)
Next, let ðð = ðððððâ1ð ; then ððð©ð and â are ðð-invariant. It is easy to see
(in view of (4.3)) thatâ¥ðð âðŽâ¥ †ðŸ2â¥ðð âðŽâ¥. (4.4)
Now
ð(ððð©ð,ð©ð) = max
{sup
ð¥âð©ð, â¥ð¥â¥=1
ð(ð¥, ððð©ð), supðŠâððð©ð, â¥ðŠâ¥=1
ð(ðŠ,ð©ð)
}†max{â¥ðð â ðŒâ¥, â¥ðâ1ð â ðŒâ¥} †ðŸ3â¥ðð âðŽâ¥,
andð(ððð©ð,ð© ) ⥠ð(ð©ð,ð© )â ð(ððð©ð,ð©ð)
⥠ðâ¥ðð â ðŽâ¥1/ðŒ âðŸ3â¥ðð âðŽâ¥â¥ (ðâðŸ3)â¥ðð âðŽâ¥1/ðŒ
⥠(ðâðŸ3)ðŸâ1/ðŒ2 â¥ðð âðŽâ¥1/ðŒ.
(4.5)
Restricting ð so that âð â â, in view of (4.1) and (4.5), a contradiction withstrong ðŒ-stability of ð© as an ðŽâ£â-invariant subspace is obtained. â¡
Step 2. If an ðŽ-invariant subspace ð© â â is strongly ðŒ-stable as an ðŽ-invariantsubspace, then ð© is also strongly ðŒ-stable as an ðŽâ£â-invariant subspace.
Proof of Step 2. Suppose not. Then there exists a sequence {ðð,â}âð=1, ðð,â alinear transformation on â, and there exists an ðð,â-invariant subspace ð©ð, withthe following properties:
(1) â¥ðð,â âðŽâ£â⥠< ðâ1, ð = 1, 2, . . .;(2) dim (ð©ð â©âð(ðð,â)) = dim (ð© â©âð(ðŽ)) for ð = 1, 2, . . . , ð;
(3) ð(ð©ð,ð© ) ⥠ðâ¥ðð,â âðŽâ£ââ¥1/ðŒ for ð = 1, 2, . . ..
232 L. Rodman
As in the proof of Step 2 of [13, Fact 4.2], let
ðð =
[ðŽâ£âð 00 ðð,â
], ð = 1, 2, . . . ,
with respect to the decomposition Hð = âð+â, where âð is the sum of rootsubspaces of ðŽ not contained in â. Then â¥ðð â ðŽâ¥ = ðŸ4â¥ðð,â â ðŽâ£ââ¥, andðð â ðŽ as ðââ. By the strong ðŒ-stability of ð© , we have, for sufficiently largeð, and every ðð-invariant subspace ð© â²ð:
dim (ð© â²ð â©âð(ðð)) = dim (ð© â©âð(ðŽ)), ð = 1, 2, . . . , ð
=â ð(ð© â²ð,ð© ) †ðŸ5â¥ðð âðŽâ¥1/ðŒ
†ðŸ5ðŸ1/ðŒ4 â¥ðð,â âðŽâ£ââ¥1/ðŒ.
(4.6)
Applying (4.6) with ð© â²ð = ð©ð, we obtain a contradiction with item (3). â¡
Step 3. If an ðŽ-invariant subspace ð© is strongly ðŒ-stable, then ð© â©â is stronglyðŒ-stable as an ðŽâ£â-invariant subspace.
Proof of Step 3. Let ð : â â â be a linear transformation sufficiently close toðŽâ£â, and let ð© â² be an ð-invariant subspace such that
dim (ð© â² â©âð(ð)) = dim ((ð© â©â) â©âð(ðŽ)), ð = 1, 2, . . . , ð. (4.7)
Define ðð¥ = ðð¥ if ð¥ â â, and ðð¥ = ðŽð¥ if ð¥ â âð. By the strong ðŒ-stability of ð©we have
ð(ð© ,ð© â²+(ð© â©âð)) †ðŸ6â¥ð âðŽâ¥1/ðŒ.It is easy to see that
â¥ð âðŽâ¥ †ðŸ7â¥ð âðŽâ£ââ¥.Now
ð(ð© â©â,ð© â²) †by Theorem 2.4(4), withð¬1 = â,ð¬2 = ð© â©âð,
†ðŸ8 ð(ð© ,ð© â²+(ð© â©âð)) †ðŸ8ðŸ6ðŸ1/ðŒ7 â¥ð âðŽââ¥1/ðŒ,
and strong ðŒ-stability of ð© â©â follows. â¡
Step 4. Let â1, . . . ,âð be all the distinct root subspaces of ðŽ, and assume that anðŽ-invariant subspace ð© is such that every intersection ð© â©âð is strongly ðŒ-stableas an ðŽâ£âð -invariant subspace, ð = 1, 2, . . . , ð . Then ð© is strongly ðŒ-stable as anðŽ-invariant subspace.
Proof of Step 4. By Step 1,ð©â©âð is strongly ðŒ-stable as an ðŽ-invariant subspace,for ð = 1, 2, . . . , ð . Now, repeat the arguments in the proof of [13, Step 4 of Fact4.2], using [13, Theorem 2.12]. â¡
Strong Stability of Invariant Subspaces 233
5. Proof of Theorem 3.3
Part (b) of Theorem 3.3 follows easily from (a). Indeed, ifâ³ is strongly Lipschitzstable, then by part (a) we have that â³ is a sum of root subspaces. Conversely,assume â³ is a sum of root subspaces. By Fact 4.1 it follows that â³ is stronglyLipschitz stable.
Consider now part (a). In view of Fact 4.1 we may (and do) assume that Hð
is a root subspace for ðŽ. If Hð contains two linearly independent eigenvectors ofðŽ, then by Theorem 2.8 a nontrivial (i.e., not equal to {0} or to Hð) ðŽ-invariantsubspace cannot be stable, hence it cannot be strongly stable by Proposition 3.1,and we are done in this case: There are no nontrivial strongly ðŒ-stable ðŽ-invariantsubspaces.
Thus, suppose that ðŽ has only one eigenvector (up to scaling). Using Theorem2.1, we may assume without loss of generality that ðŽ = ðœð(ð), where ð â C hasnonnegative imaginary part. Then part (a) amounts to the following two state-ments; here and in the sequel we denote by ðð the vector having 1 in the ðthposition and zeros elsewhere:
Statement 5.1. Assume (âµ(ðŽ,â³)) holds. Then there exist ð¿,ðŸ > 0 such that forevery ðµ â HðÃð with â¥ðµ âðŽâ¥ < ð¿ and every ð-dimensional ðµ-invariant subspaceð© â Hð we have
ð(ð© , Span {ð1, . . . , ðð}) †ðŸâ¥ðµ âðŽâ¥1/ð. (5.1)
Statement 5.2. For every ð = 1, 2, . . . , ð â 1, and for every ðŒ < ð, there exists asequence {ðµð}âð=1, ðµð â HðÃð, and there exists a ð-dimensional ðµð-invariantsubspace â³ð â Hð, ð = 1, 2, . . ., such that limðââðµð = ðŽ and
ð(â³ð, Span {ð1, . . . , ðð}) ⥠ðâ¥ðµ âðŽâ¥1/ðŒ.Proof of Statement 5.1. Consider first the case when ð is nonreal. We use thecomplex representation ð of quaternion matrices; note that
ðð(ðŽ)ðâ1 =[ðœð(ð) 0
0 ðœð(ð)
]â C2ðÃ2ð,
for a suitable permutation matrix ð .By Theorem 3.2 the following claim holds true: there exist ð¿â²,ðŸ â² > 0 such
that for every ðµâ² â C2ðÃ2ð with â¥ðµâ²âðð(ðŽ)ðâ1⥠< ð¿â² and every 2ð-dimensionalðµâ²-invariant subspace â³â² â C2ð we have
ð(â³â², Span {ð1, . . . , ðâ, ðð+1, . . . ðð+2ðââ}) †ðŸ â²â¥ðµâ² â ðð(ðŽ)ðâ1â¥1/ð (5.2)
for some â, max{2ðâð, 0} †â †min{ð, 2ð}, which may depend on â³â². Note thefact that, for ðµ â HðÃð and its invariant subspace ð© , the subspace ð(ð© ) â C2ð
(see Proposition 2.2 for the definition) is ð(ðµ)-invariant. We apply the claim tomatrices ðµâ² of the form ðµâ² = ðð(ðµ)ðâ1 for some ðµ â HðÃð and their invariantsubspaces of the form
â³â² = ð (Col (ð(ð£1), . . . , ð(ð£ð))) ,
234 L. Rodman
where Col (ð(ð£1), . . . , ð(ð£ð)) is subspace spanned by the columns of {ð(ð£1), . . .,ð(ð£ð)}, and {ð£1, . . . , ð£ð} is a basis of a ðµ-invariant subspace fð© . In view of Propo-sition 2.3, the Jordan form of such matrices ðµâ² restricted toâ³â² is symmetric withrespect to the real axis, so in fact it must be â = ð in (5.2). We now have, for aðµ-invariant ð-dimensional subspace ð© â Hð:
ð(ð© , Span {ð1, . . . , ðð}) (by Proposition 2.2 and 2.1)
†ð¶ð,ðð(ð (ð© ), ð (Span {ð1, . . . , ðð})= ð¶ð,ðð(ð (ð (ð© )), ð (ð (Span {ð1, . . . , ðð}))†(by (5.2)) ð¶ð,ððŸ
â²â¥ð (ðµ)â ð (ðŽ)â¥1/ð†(by (2.1)) ðâ1ð,ðð¶ð,ððŸ
â²â¥ðµ âðŽâ¥1/ð,and Statement 5.1 follows for the case of a nonreal ð. Consider the case when ð isreal; then the result follows from Theorem 2.8. â¡
Proof of Statement 5.2. In view of Theorem 3.2, there are sequences of matrices{ðµð}âð=1, ðµð â CðÃð and of complex subspaces {â³ð}âð=1,â³ð â Cð with therequired properties. â¡
Finally, consider (c). Assume thatâ³ is stable. We need to prove thatâ³ isstrongly stable. By Fact 4.1, we may assume that Hð is the (sole) root subspace ofðŽ. Ignoring the trivial casesâ³ = {0} andâ³ = Hð, in view of description of non-trivial stable ðŽ-invariant subspaces (Theorem 2.8(b)) we may further assume thatðŽ = ðœ(ð) for some ð â C. Arguing by contradiction, assume â³ is not stronglystable. Letting ð = dimâ³, there exists ð0 > 0 such that for some sequence{ðµð}âð=1, ðµð â HðÃð, we have â¥ðµð â ðŽâ¥ < ðâ1 and ð(â³ð,â³) ⥠ð0 forsome ðµð-invariant subspace â³ð of dimension ð. Passing to a subsequence, wemay assume that {â³ð} converges in the gap norm: limðââ â³ð = ð© for somesubspace ð© . We then have ð(â³,ð© ) ⥠ð0 and dimð© = ð.
Moreover, ð© is ðŽ-invariant. Indeed, let ð¢1, . . . , ð¢ð be a basis for ð© . Thenthere exist ð¢ð,1, . . . , ð¢ð,ð such that ð¢ð,â ââ³ð, ð = 1, 2, . . ., and
limðââð¢ð,â = ð¢â, â = 1, 2, . . . , ð.
Clearly, for sufficiently large ð (which will be assumed), the vectorsð¢ð,1, . . . , ð¢ð,ð form a basis forâ³ð. Sinceâ³ð is ðµð-invariant, we have
ðµð [ð¢ð,1 . . . ð¢ð,ð] = [ð¢ð,1 . . . ð¢ð,ð] ðð (5.3)
for some matrix ðð. In fact,
ðð = (ðâððð)â1ðµððð,
where ðð = [ð¢ð,1 . . . ð¢ð,ð], and the invertibility of ðâððð follows from thelinear independence of the columns of ðð. Passing to the limit when ð â â in(5.3), the ðŽ-invariance of ð© follows.
But now we obtain a contradiction, because ðŽ has only one invariant subspaceof the fixed dimension ð (cf. [13, Proposition 2.10]).
Strong Stability of Invariant Subspaces 235
6. Strong ð¶-stability: alternative formulation
In this section we re-cast strong ðŒ-stability property of invariant subspaces in adifferent form that does not involve equalities (3.1), and is more in the spirit of thedefinition of ðŒ-stability. The next theorem provides the alternative formulation forthe strong stability property.
Theorem 6.1. Fix ðŒ ⥠1. In the following statements, (1) implies (2) for ðŽ â HðÃð
and an ðŽ-invariant subspace â³â Hð:
(1) â³ is strongly ðŒ-stable;(2) there are positive constants ð¿1, ð¿
â²1,ðŸ1 that depend on ðŽ andâ³ only such that
the set of all ðµ-invariant subspaces ð© for which the inequality ð(ð© ,â³) †ð¿â²1holds is non-empty for every ðµ â HðÃð satisfyingâ¥ðµ âðŽâ¥ †ð¿1, and
ð(ð© ,â³) †ðŸ1â¥ðµ âðŽâ¥1/ðŒ (6.1)
holds for every such ð© .
Assume in addition that the property (âµ(ðŽ,â³)) is satisfied. Then the condi-tions (1) and (2) are equivalent.
We do not state a parallel version for strong stability, because by Theorem 3.3strong stability is equivalent to stability, and therefore the definition of stabilitycan be thought of as an alternative version of strong stability.
The rest of this section is devoted to the proof of Theorem 6.1. We need twolemmas.
Lemma 6.2. Let ð¬1,ð¬2,ðµ1,ðµ2 be subspaces in Hð such that
ð¬1+ð¬2 = Hð, ðµ1 â ð¬1, ðµ2 â ð¬2.
Then there exists ð¿3 > 0 depending on ð¬1,ð¬2,ðµ1,ðµ2 only with the following prop-erty: If ðŽ,ð¬â²1,ð¬â²2 â Hð are subspaces such that
ð(ðŽ,ðµ1+ðµ2) < ð¿3, ð(ð¬â²1,ð¬1) < ð¿3, ð(ð¬â²2,ð¬2) < ð¿3,
and
ðŽ = (ðŽ â© ð¬â²1) + (ðŽ â© ð¬â²2), (6.2)
then
dim (ðŽ â© ð¬â²ð) = dimðµð , ð = 1, 2.
Proof. By Theorem 2.4 (2) we have ð¬â²1+ð¬â²2 = Hð (if ð¿3 is taken sufficiently small),hence (ðŽ â©ð¬â²1)+(ðŽ â© ð¬â²2) is a direct sum, and
dimðµ1 + dimðµ2 = dimðŽ = dim (ðŽ â© ð¬â²1) + dim (ðŽ â© ð¬â²2).Thus, it suffices to prove that
dim (ðŽ â© ð¬â²ð) †dimðµð , ð = 1, 2.
236 L. Rodman
We prove this for ð = 1; the case ð = 2 is analogous. Assume not. Then there existsequences of subspaces {ðŽð,ð¬â²1,ð,ð¬â²2,ð}âð=1 such that
ðŽð â ðµ1+ðµ2, ð¬â²ð,ð â ð¬ð , ð = 1, 2, (6.3)
as ð â â,ðŽð = (ðŽð â© ð¬â²1,ð) + (ðŽð â© ð¬â²2,ð), ð = 1, 2, . . . ,
but
dim (ðŽð â©ð¬â²1,ð) > dimðµ1, ð = 1, 2, . . . . (6.4)
Using compactness of the metric space of subspaces in Hð, we may assume that thesequence {ðŽð â© ð¬â²1,ð}âð=1 converges to a subspace ð² . Take ð¥ â ð² . Then thereis a sequence {ð¥ð}âð=1, ð¥ð â ðŽð â©ð¬â²1,ð, such that limðââ ð¥ð = ð¥. By (6.3) we
also have ð¥ â ð¬1, ð¥ â ðµ1+ðµ2, hence (because ðµ1 â ð¬1) ð¥ â ðµ1. Thus ð² â ðµ1, acontradiction with our assumption (6.4) (note that dim (ðŽð â©ð¬â²1,ð) = dimð² forlarge ð by Theorem 2.4(1)). â¡
Lemma 6.3. Given ðŽ andâ³ as in Theorem 6.1, for every ð4 > 0 there exists ð¿4 > 0with the property that for each ðµ â HðÃð and for each ðµ-invariant subspace ð©such that
â¥ðµ âðŽâ¥ < ð¿4, ð(ð© ,â³) < ð¿4, (6.5)
the inequality
maxð=1,2,...,ð
{ð(ð© â©âð(ðµ),â³â©âð(ðŽ))} †ð4holds, where the maximum is taken over all sums of root subspaces â1(ðŽ), . . .,âð (ðŽ) for ðŽ, and where the sums of root subspaces â1(ðµ), . . . ,âð (ðµ) for ðµ aresuch that the eigenvalues of ðµ to which âð(ðµ) corresponds are in a neighborhoodof the eigenvalues of ðŽ to which âð(ðŽ) corresponds, for ð = 1, 2, . . . , ð .
Proof. Denote by âð(ðŽ)ð, resp. âð(ðµ)
ð, the sum of root subspaces for ðŽ, resp. ðµ,
which is a direct complement to âð(ðŽ), resp. âð(ðµ), in Hð, and let ðâð(ðŽ),âð(ðŽ)ð ,
resp. ðâð(ðµ),âð(ðµ)ð , be the projection on âð(ðŽ) along âð(ðŽ)ð, resp. on âð(ðµ)
along âð(ðµ)ð. In what follows, we denote by ðŸ0,ðŸ1, . . . positive constants that
depend only on ðŽ andâ³. By Theorem 2.4(2)â¥â¥â¥ðâð(ðµ),âð(ðµ)ð â ðâð(ðŽ),âð(ðŽ)ð
â¥â¥â¥â€ ðŸ0(ð(âð(ðµ),âð(ðŽ)) + ð(âð(ðµ)
ð,âð(ðŽ)ð)
†ðŸ0ðŸ1â¥ðµ âðŽâ¥,(6.6)
where the second inequality follows from Proposition 2.7(b). Proposition 2.6 gives:
ð(ð© â©âð(ðµ),â³â©âð(ðŽ))
= 2â¥â¥ðð© (ðð© + ðâð(ðµ))
+ðâð(ðµ) â ðâ³(ðâ³ + ðâð(ðŽ))+ðâð(ðŽ)
â¥â¥ ; (6.7)
here ð© is a ðµ-invariant subspace. On the other hand, taking ð¿4 sufficiently smalland assuming (6.5) holds, Lemma 6.2 (applied with ð¬1 = âð(ðŽ), ð¬2 = âð(ðŽ)
ð,
Strong Stability of Invariant Subspaces 237
ðµ1 = â³ â© âð(ðŽ), ðµ2 = â³ â© âð(ðŽ)ð, ð¬â²1 = âð(ðµ), ð¬â²2 = âð(ðµ)
ð, ðŽ = ð© ),together with (6.6) and Theorem 2.4(1), yields that
dim (ð© â©âð(ðµ)) = dim (â³â©âð(ðŽ)), ð = 1, 2, . . . , ð ,
and hence
dim (ð© +âð(ðµ)) = dim (â³ +âð(ðŽ)), ð = 1, 2, . . . , ð .
Since the range of ðð© + ðâð(ðµ) is equal to ð© + âð(ðµ) (see [9, Corollary 2], forexample), and similarly for ðŽ, we now have
rank (ðð© + ðâð(ðµ)) = rank (ðâ³ + ðâð(ðŽ)), ð = 1, 2, . . . , ð .
Using Theorem 2.5, formula (6.7) gives the result of Lemma 6.3. â¡
Proof of Theorem 6.1, (1)â (2). Assume (1) holds. Then by Lemma 6.3 (takingð4 < 1 and using Theorem 2.4(1)) equalities (3.1) are guaranteed, and we have
ð(ð© ,â³) †ðŸ1â¥ðµ âðŽâ¥1/ðŒfor every ðµ-invariant subspace ð© provided the inequality
max{ð(ð© ,â³), â¥ðµ âðŽâ¥} †ð¿4/2holds. It remains to prove that for some ð¿1, 0 < ð¿1 †ð¿4/2, the set of all ðµ-invariantsubspacesð© for which the inequality ð(ð© ,â³) †ð¿4/2 holds is non-empty for everyðµ â HðÃð satisfying â¥ðµâðŽâ¥ †ð¿1. To this end we take advantage of the fact thatunder (1)â³ is ðŒ-stable. Thus, there exist ð¿â²,ðŸ â² > 0 such that
â¥ðµ âðŽâ¥ †ð¿â² â â ðµ-invariant ð© such that ð(ð© ,â³) †ðŸ â²â¥ðµ âðŽâ¥1/ðŒ.Now take
ð¿1 = min{ð¿â², ð¿4/2, (ð¿4/(2ðŸ â²))ðŒ}. â¡
Proof of Theorem 6.1 in the complex case. The implication (1) â (2) is verifiedas in the quaternionic case. Assume now that (2) holds. Thenâ³ is in particularðŒ-stable, and using the description of ðŒ-stability in the complex case (Theorem2.8, Remark 2.9), we easily reduce the proof to the case when ðŽ = ðœð(0). Theorem3.2 shows that â³ is strongly ð-stable. Thus, if ðŒ ⥠ð, we are done. However, ifðŒ < ð, then (6.1) cannot hold (unlessâ³ is trivial:â³ = {0} orâ³ = Hð).
Indeed, let ð, 1 †ð †ðâ 1, be the (complex) dimension of
â³ := Range
[ðŒð
0ðâð,ð
],
and let ð1, . . . , ðð be a set of distinct ðth roots of unity that do not sum upto zero. For ð > 0, let ðµð â CðÃð be the matrix obtained from ðŽ = ðœð(0)by adding ð in the lower left corner. Clearly, â¥ðµð â ðŽâ¥ = ð. Letting ð©ð to bethe ðµð-invariant subspace generated by the eigenvectors ðµð corresponding to the
eigenvalues ð1ð1/ð, . . . , ððð
1/ð, one verifies that
ð©ð = Range
[ðŒððð
],
238 L. Rodman
where ðð â C(ðâð)Ãð has ð1ð1/ð+ â â â + ððð1/ð in its top left corner (see the proof
of [7, Lemma 16.5.2], also [13, Lemma 4.16]). Now [13, Lemma 5.2(a3)] guaranteesexistence of a constant ðŸ0,1 > 0 which depends on ð only such that
ð (â³,ð©ð) ⥠ðŸ0,1 â£ð1 + â â â + ðð⣠ð1/ð (6.8)
as long as ð †1. Letting ð â 0, a contradiction with (6.1) is obtained. â¡Proof of Theorem 6.1, (2)â (1). The proof of (2) â (1), under the additionalhypothesis that (âµ(ðŽ,â³)) holds, is essentially the same as in the complex case(use Theorem 3.3(a) instead of Theorem 3.2). â¡
References
[1] W.N. Anderson, Jr., and R.J. Duffin. Series and parallel addition of matrices, J.Math. Anal. Appl., 26 (1969), 576â594.
[2] H. Bart, I. Gohberg, and M.A. Kaashoek. Minimal factorization of matrix and op-erator functions. Operator Theory: Advances and Applications, Vol 1, Birkhauser,1979.
[3] H. Bart, I. Gohberg, and M.A. Kaashoek. Stable factorizations of monic matrixpolynomials and stable invariant subspaces, Integral Equations Operator Theory, 1(1978), no. 4, 496â517.
[4] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran. Factorization of matrixand operator functions: the state space method, Operator Theory: Advances andApplications, Vol 178, Birkhauser, 2008.
[5] S. Campbell and J. Daughtry. The stable solutions of quadratic matrix equations,Proc. Amer. Math. Soc., 74 (1979), no. 1, 19â23.
[6] I. Gohberg, P. Lancaster, and L. Rodman. Matrix Polynomials, Academic Press,1982; republication, SIAM, 2009.
[7] I. Gohberg, P. Lancaster, and L. Rodman. Invariant Subspaces of Matrices withApplications, John Wiley, 1986; republication, SIAM, 2006.
[8] I. Gohberg, P. Lancaster, and L. Rodman. Matrices and Indefinite Scalar Products,Operator Theory: Advances and Applications, Vol. 8, Birkhauser, Basel and Boston,1983.
[9] R. Piziak, P.L. Odell, and R. Hahn. Constructing projections for sums and intersec-tions, Computers and Mathematics with Applications, 17 (1999), 67â74.
[10] A.C. M. Ran and L. Rodman. Stability of invariant Lagrangian subspaces I, OperatorTheory: Advances and Applications, 32 (1988), 181â218.
[11] A.C. M. Ran and L. Rodman. The rate of convergence of real invariant subspaces,Linear Algebra Appl., 207 (1994), 197â224.
[12] A.C.M. Ran and L. Roozemond. On strong ðŒ-stability of invariant subspaces ofmatrices, Operator Theory: Advances and Applications, 40 (1989), 427â435.
[13] L. Rodman. Stability of invariant subspaces of quaternion matrices, Complex Anal-ysis and Operator Theory, 6 (2012), 1069â1119.
[14] G.W. Stewart. On the continuity of the generalized inverse, SIAM J. Appl. Math.,17 (1969), 33â45.
Strong Stability of Invariant Subspaces 239
[15] N.A. Wiegmann. Some theorems on matrices with real quaternion entries, CanadianJ. of Math., 7 (1955), 191â201.
[16] F. Zhang. Quaternions and matrices of quaternions, Linear Algebra Appl., 251 (1997),21â57.
Leiba RodmanDepartment of MathematicsCollege of William and MaryWilliamsburg, VA 23187-8795, USAe-mail: [email protected]
Operator Theory:Advances and Applications, Vol. 237, 241â246câ 2013 Springer Basel
Determinantal Representations ofStable Polynomials
Hugo J. Woerdeman
Dedicated to Leonia Lerer on the occasion of his seventieth birthday
Abstract. For every stable multivariable polynomial ð, with ð(0) = 1, weconstruct a determinantal representation
ð(ð§) = det(ðŒ âð(ð§)),
where ð(ð§) is a matrix-valued rational function with â¥ð(ð§)⥠†1 andâ¥ð(ð§)ð⥠< 1 for ð§ â ðð and ð(ðð§) = ðð(ð§) for all ð â â â {0}.Mathematics Subject Classification (2010). 15A15; 11C20, 47A13, 47A48.
Keywords. Determinantal representation; multivariable polynomial; stablepolynomial.
1. Introduction
A polynomial ð(ð§) = ð(ð§1, . . . , ð§ð) is called stable if ð(ð§) â= 0 for ð§ â ð»ð, where
ð» is the open unit disk in â and ð» is its closure. We shall also use the notationð for the unit circle. The polynomial ð(ð§) is called semi-stable when ð(ð§) â= 0 forð§ â ð»ð. It is an open question whether every multivariable stable polynomial ð(ð§)with ð(0) = 1 can be written as
ð(ð§) = det(ðŒ âðŸð(ð§)),where ð(ð§) is a diagonal matrix with coordinate variables on the diagonal and ðŸis a strict contraction. For one and two variable polynomials such a representationalways exists; in one variable it is an easy consequence of the fundamental theoremof algebra, while in two variables the result follows from [8, Theorem 1] and [7].The question of existence of representations of the form det(ðŒ âðŸð(ð§)) was thetopic of the paper [6], where such representation was shown to lead to rational
The author was partially supported by NSF grant DMS-0901628.
242 H.J. Woerdeman
inner functions in the SchurâAgler class. As a consequence of these results, it canbe seen that for 5/6 < ð < 1, the stable polynomial
ð(ð§1, ð§2, ð§3) = 1 +ð
5ð§1ð§2ð§3
(ð§21ð§
22 + ð§
22ð§
23 + ð§
23ð§
21 â 2ð§1ð§2ð§23 â 2ð§1ð§22ð§3 â 2ð§21ð§2ð§3
)can not be represented as det(ðŒ â ðŸð(ð§)) where ðŸ is a 9 à 9 (or smaller size)contraction. It is an open question whether such a representation exists with alarger size contraction ðŸ. Our main result shows, however, that we can find a rep-resentation ð(ð§) = det(ðŒ âð(ð§)), with ð(ð§) a rational matrix function satisfyingâ¥ð(ð§)⥠†1 and â¥ð(ð§)7⥠< 1, ð§ â ð3, and ð(ðð§) = ðð(ð§). In fact, for thisparticular polynomial ð, one may choose ð(ð§) to be the 7 à 7 rational matrixfunction
ð(ð§) =
ââââââââââ
0 ð§1 00 ð§1 0
0 ð§1 00 ð§1 0
0 ð§1 00 0 ð§1
ð(ð§)â1ð§61
0 0
ââââââââââ .
Our main result is the following determinantal characterization of stable poly-nomials. Recall that the total degree tdegð of a polynomial ð is the maximum amongthe total degrees of all its terms, where the total degree of the monomial ð§ð1
1 â â â ð§ðð
ð
is ð1 + â â â + ðð.Theorem 1.1. Let ð be a polynomial in ð variables. Put ð = tdegð. Then ð isstable with ð(0) = 1 if and only if for some ð â â there exists a ð à ð rationalmatrix-valued function ð(ð§) which has no singularities on the set âªð>0ðð
ð so that
(i) ð(ð§) = det(ðŒð âð(ð§)),(ii) ð(ð§) is contractive and ð(ð§)ð is strictly contractive for all ð§ â ðð,(iii) ð(ðð§) = ðð(ð§) for all ð â â â {0} and ð§ â ðð.
In [13] real zero polynomials were studied, for which a desirable determinantalrepresentation is det(ðŒ + ð¥1ðŽ1 + â â â + ð¥ððŽð) with ðŽ1, . . . , ðŽð symmetric. Notall real zero polynomials have such a representation; see [12]. In [13, Theorem3.1] it was shown, however, that every square-free real zero polynomial can bewritten as det(ðŒ âð(ð¥)), where ð(ð¥) is a symmetric rational matrix functionand ð(ðð¥) = ðð(ð¥), ð â â, ð¥ â âð. Our Theorem 1.1 can be seen as an analogof [13, Theorem 3.1] to the setting of stable polynomials.
2. Proof of main result
We first need a couple of lemmas.
Lemma 2.1. Let ð(ð§) be a positive definite ðÃð matrix-valued trigonometric poly-nomial on ðð, so that in the Laurent expansion the (ð, ð)th entry is homogeneousof degree ð â ð. Then there exists a factorization ð(ð§) = ð (ð§)ð (ð§)â, with ð (ð§) a
Determinantal Representations of Stable Polynomials 243
rational matrix function of size ð Ã ð, say, so that in the Laurent expansion theðth row of the ð is homogeneous of degree ð â 1. The rational matrix function ðmay be chosen to be polynomial in at least ðâ 1 variables.
Proof. By [3, Corollary 5.2] a polynomial matrix function ð exists so that ð =
ð ð â. Write now ð =âð
ð=0 ð ð , where ð ð is homogeneous of degree ð. Next, write
ð ð = row(ððð)ðð=1. Observe that
ððð =
( ðâð=0
ððð
)( ðâð=0
ðâðð
)=
âðâð=ðâð
ððððâðð +
âðâð â=ðâð
ððððâðð,
but as the last term equals 0 due to ððð being homogeneous of degree ð â ð, weactually have
ððð =â
ðâð=ðâð
ððððâðð.
Define now
ð =
âââââð1,ðð§
âð1 â â â ð1,0ð§
01 0
0 ð2,ðð§âð+11 â â â ð2,0ð§
11
. . .. . .
. . .
0 ðð,ðð§âð+ðâ11 â â â ðð,0ð§
ðâ11
âââââ .Then the ðth row of ð is homogeneous of degree ðâ 1 and ð = ðð â. â¡
Lemma 2.2. Let ð(ð§) = ð0 + ð1ð§ + â â â + ððð§ð with ð0 = 1 be a one variable stablepolynomial, and let ð := ðŽðŽâ âðµâðµ, where
ðŽ =
âââ ð0...
. . .
ððâ1 â â â ð0
âââ , ðµ =
âââðð â â â ð1. . .
...ðð
âââ .In addition, let
ð¶ =
ââââââð1 1 0...
. . .
âððâ1 0 1âðð 0 â â â 0
âââââ .Then ð > 0, (
ðâ1 ð¶âðâ1
ðâ1ð¶ ðâ1
)⥠0,
(ðâ1 ð¶ðâðâ1
ðâ1ð¶ð ðâ1
)> 0. (2.1)
Proof. Let ð(ð§) = 1â£ð(ð§)â£2 , and write ð(ð§) =
ââð=ââ ððð§
ð, â£ð§â£ = 1. Introduce
ð = (ððâð)ðâ1ð,ð=0, ð = (ððâð+1)
ðâ1ð,ð=0, ð = (ððâð+ð)
ðð,ð=0.
244 H.J. Woerdeman
By the SchurâCohn criterion [9, Section 13.5] we have that ð > 0 and by theGohbergâSemencul formula [5] we have that ðâ1 = ð . In addition, it is easy tocheck that ðð¶ = ð . Next, observe that(
ðâ1 ð¶âðâ1
ðâ1ð¶ ðâ1
)=
(ð ð â
ð ð
),
has many identical rows and columns; indeed, columns ð and ð + ð¡ â 1 are equalfor ð = 2, . . . , ð, and by selfadjointness the same holds for the rows. Removingcolumns and rows ð + 1, . . . , 2ð â 1, we remain with the positive definite matrix(ððâð)
ðð,ð=0. But then the first inequality in (2.1) follows.
In addition, one may check that ðð¶ð = ð (this was observed in [4, Proof ofTheorem 2.1]; see also [1, Equation (2.3.14)]). It remains to observe that(
ðâ1 ð¶ðâðâ1
ðâ1ð¶ð ðâ1
)=
(ð ð â
ð ð
)= (ððâð)
2ðâ1ð,ð=0 > 0. â¡
Proof of Theorem 1.1. âIfâ: Suppose that ð(ð§) as described exists, and thatð(ð§) = 0 for some ð§ â ðð. Then 1 is an eigenvalue of ð(ð§). But as â¥ð(ð§)ð⥠< 1,this can not happen. Thus ð(ð§) â= 0 for ð§ â ðð. Now using that ð(ðð§) = ðð(ð§),we get that â¥ð(ð§)⥠< 1 for any ð§ â ððð where 0 < ð < 1. Thus ð(ð§) â= 0 for anyð§ â ððð where 0 < ð †1. In addition, for ð§ â ðð one has
ð(0) = limðâ0+
ð(ðð§) = limðâ0+
det(ðŒ âð(ðð§)) = limðâ0+
det(ðŒ â ðð(ð§)) = 1.
But now the stability of ð follows from Theorem 1â in [2].
âOnly ifâ: Let ð(ð§) = ð(ð§1, . . . , ð§ð) be a stable multivariable polynomial withð(0) = 1. As ð = tdegð, we may write ð(ð§) = 1 + ð1(ð§) + â â â + ðð(ð§), where ðð isa homogeneous multivariable polynomial of degree ð; i.e., ðð(ðð§) = ð
ððð(ð§), whereð â â. Introduce now,
ð¶(ð§) =
ââââââð1(ð§) 1 0
.... . .
âððâ1(ð§) 0 1âðð(ð§) 0 â â â 0
âââââ .Then det(ðŒð â ð¶(ð§)) = ð(ð§). In addition, if we let ð·ð = diag(ðð)ðâ1ð=0 , then
ð¶(ðð§) = ðð·ðð¶(ð§)ð·â1ð , ð â â â {0}, ð§ â âð. (2.2)
For fixed ð§ â ð»ðput
ðð§(ð) = ð(ðð§) = 1 + ðð1(ð§) + â â â + ðððð(ð§).Then ðð§ is stable, so by the SchurâCohn criterion (see, e.g., [11], [9, Section 13.5])we have that
ð(ð§) := ðŽ(ð§)ðŽ(1/ð§)â âðµ(1/ð§)âðµ(ð§)
Determinantal Representations of Stable Polynomials 245
is positive definite for ð§ â ðð. Here
ðŽ(ð§) =
âââ ð0(ð§)...
. . .
ððâ1(ð§) â â â ð0(ð§)
âââ , ðµ(ð§) =
âââðð(ð§) â â â ð1(ð§). . .
...ðð(ð§)
âââ ,and ð0(ð§) = 1. The matrix ð is also called the Bezoutian corresponding to ðð§and its reverse ââðð§(ð) = ðððð§(1/ð); see, for instance, [10] and [11]. It is easy tosee that if we write ð(ð§) = (ððð(ð§))
ðð,ð=1, then ððð(ðð§) = ððâðððð(ð§). But then
ð(ðð§) = ð·ðð(ð§)ð·â1ð follows. Next, by Lemma 2.2, we have that(
ð(ð§)â1 ð¶(ð§)âð(ð§)â1
ð(ð§)â1ð¶(ð§) ð(ð§)â1
)⥠0, ð§ â ðð. (2.3)
Multiplying all rows and columns on both sides with ð(ð§) we obtain(ð(ð§) ð(ð§)ð¶(ð§)â
ð¶(ð§)ð(ð§) ð(ð§)
)⥠0, ð§ â ðð. (2.4)
As ð(ð§) satisfies the conditions of Lemma 2.1 we may write ð(ð§) = ð (ð§)ð (ð§)â,ð§ â ðð, with the ðth row of ð being homogeneous of degree ð â 1. Thus ð (ðð§) =ð·ðð (ð§). Let now
ð(ð§) = ð (1/ð§)âð(ð§)â1ð¶(ð§)ð (ð§).
Then det(ðŒ âð(ð§)) = det(ðŒ â ð (ð§)ð (1/ð§)âð(ð§)â1ð¶(ð§)) = det(ðŒ âð¶(ð§)) = ð(ð§).Next,
ð(ðð§) = ð (1/ðð§)âð(ðð§)â1ð¶(ðð§)ð (ðð§)
= ð (1/ð§)âð·â1ð ð·ðð(ð§)â1ð·â1ð ðð·ðð¶(ð§)ð·
â1ð ð·ðð (ð§) = ðð(ð§).
Finally, for ð§ â ðð, we have that
ð (ð§)(ðŒ âð(ð§)âð(ð§))ð (ð§)â = ð(ð§)âð(ð§)ð¶(ð§)âð(ð§)â1ð¶(ð§)ð(ð§) ⥠0,
which follows from (2.4).Thus, as Ranð(ð§)â â Ranð (ð§)â, it follows that â¥ð(ð§)⥠†1. Using the
second inequality in (2.1) one can show in a similar way that â¥ð(ð§)ð⥠< 1. â¡
Remark 2.3. If ð(ð§) in Theorem 1.1 can be chosen to be analytic in ð»ðthen
one easily sees from (iii) that ð(ð§) =âð
ð=1 ð§ððð for some constant matricesð1, . . . ,ðð. When in addition,ð(ð§) satisfies that â¥ð1âð1+ â â â +ððâðð⥠†1for all contractions ð1, . . . , ðð (i.e., ð(ð§) is in SchurâAgler class), then it follows
from [6, Corollary 3.3] that det(ðŒ+âð
ð=1 ð§ððð) may be written as det(ðŒâðŸð(ð§))with ðŸ a contraction. It is an open problem what happens when ð(ð§) is not inthe SchurâAgler class.
Acknowledgments
The author wishes to thank Anatolii Grinshpan and Dmitry S. Kaliuzhnyi-Verbo-vetskyi for useful discussions and their input on earlier drafts of this paper.
246 H.J. Woerdeman
References
[1] M. Bakonyi and H.J. Woerdeman. Matrix completions, moments, and sums of Her-mitian squares. Princeton University Press, Princeton, NJ, 2011. xii+518 pp.
[2] Ph. Delsarte, Y.V. Genin, and Y.G. Kamp, A simple proof of Rudinâs multivariablestability theorem. IEEE Trans. Acoust. Speech Signal Process. 28 (1980), no. 6, 701â705.
[3] M.A. Dritschel. On factorization of trigonometric polynomials. Integral EquationsOperator Theory 49 (2004), no. 1, 11â42.
[4] I. Gohberg and L. Lerer, Matrix generalizations of M.G. Krein theorems on orthog-onal polynomials. Orthogonal matrix-valued polynomials and applications, 137â202,Oper. Theory Adv. Appl., 34, Birkhauser, Basel, 1988.
[5] I.C. Gohberg and A.A. Semencul, The inversion of finite Toeplitz matrices and theircontinual analogues. (Russian) Mat. Issled. 7 (1972), no. 2(24), 201â223, 290.
[6] A. Grinshpan, D.S. Kaliuzhnyi-Verbovetskyi, and H.J. Woerdeman. Norm-con-strained determinantal representations of multivariable polynomials. Complex Anal.Oper. Theory 7 (2013), no. 3, 635â654.
[7] A. Grinshpan, D.S. Kaliuzhnyi-Verbovetskyi, V. Vinnikov and H.J. Woerdeman.Stable polynomials and real zero polynomials in two variables. Preprint.
[8] A. Kummert. 2-D stable polynomials with parameter-dependent coefficients: gener-alizations and new results. IEEE Trans. Circuits Systems I: Fund. Theory Appl. 49(2002), 725â731.
[9] P. Lancaster and M. Tismenetsky. The theory of matrices. Second edition. ComputerScience and Applied Mathematics. Academic Press, Inc., Orlando, FL, 1985.
[10] P. Lancaster, L. Lerer, and M. Tismenetsky, Factored forms for solutions of ðŽð âððµ = ð¶ and ðâðŽððµ = ð¶ in companion matrices. Linear Algebra Appl. 62 (1984),19â49.
[11] L. Lerer, L. Rodman, and M. Tismenetsky, Bezoutian and SchurâCohn problem foroperator polynomials. J. Math. Anal. Appl. 103 (1984), no. 1, 83â102.
[12] T. Netzer and A. Thom. Polynomials with and without determinantal representa-tions. Linear Algebra Appl. 437 (2012), 1579â1595.
[13] T. Netzer, D. Plaumann and A. Thom. Determinantal representations and the Her-mite matrix. Michigan Math. J. 62 (2013), 407â420.
Hugo J. WoerdemanDepartment of MathematicsDrexel University3141 Chestnut St.Philadelphia, PA, 19104, USAe-mail: [email protected]