Operator Theory: Advances and Applications Vol. 200 Founded in 1979 by Israel Gohberg
Editors: Harry Dym (Rehovot, Isr...

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Operator Theory: Advances and Applications Vol. 200 Founded in 1979 by Israel Gohberg

Editors: Harry Dym (Rehovot, Israel) Joseph A. Ball (Blacksburg, VA, USA) Marinus A. Kaashoek (Amsterdam, The Netherlands) Heinz Langer (Vienna, Austria) Christiane Tretter (Bern, Switzerland)

Associate Editors: Vadim Adamyan (Odessa, Ukraine) Albrecht Böttcher (Chemnitz, Germany) B. Malcolm Brown (Cardiff, UK) Raul Curto (Iowa City, IA, USA) Fritz Gesztesy (Columbia, MO, USA) Pavel Kurasov (Lund, Sweden) Leonid E. Lerer (Haifa, Israel) Vern Paulsen (Houston, TX, USA) Mihai Putinar (Santa Barbara, CA, USA) Leiba Rodman (Williamsburg, VI, USA) Ilya M. Spitkovsky (Williamsburg, VI, USA)

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)

Honorary and Advisory Editorial Board: 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, AB, Canada) Peter D. Lax (New York, NY, USA) Donald Sarason (Berkeley, CA, USA) Bernd Silbermann (Chemnitz, Germany) Harold Widom (Santa Cruz, CA, USA)

A State Space Approach to Canonical Factorization with Applications

Harm Bart Israel Gohberg Marinus A. Kaashoek André C.M. Ran

Birkhäuser

L O L S

Linear Operators & Linear Systems

Authors: Harm Bart Econometrisch Instituut Erasumus Universiteit Rotterdam Postbus 1738 3000 DR Rotterdam The Netherlands e-mail: [email protected]

Marinus A. Kaashoek, André C. M. Ran Department of Mathematics, FEW Vrije Universiteit De Boelelaan 1081A 1081 HV Amsterdam The Netherlands e-mail: [email protected] [email protected]

Israel Gohberg (Z"L)

2010 Mathematics Subject Classiﬁcation: Primary 46C20; 47A48, 47A56, 47A68, 93B36; secondary: 42A85, 82D75 Library of Congress Control Number: 2010923703

Bibliographic information published by Die Deutsche Bibliothek. Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliograﬁe; detailed bibliographic data is available in the Internet at http://dnb.ddb.de

ISBN 978-3-7643-8752-5 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, speciﬁcally the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microﬁlms or in other ways, and storage in data banks. For any kind of use permission of the copyright owner must be obtained.

© 2010 Birkhäuser / Springer Basel AG P.O. Box 133, CH-4010 Basel, Switzerland Part of Springer Science+Business Media Printed on acid-free paper produced from chlorine-free pulp. TCF∞ Printed in Germany ISBN 978-3-7643-8752-5

e-ISBN 978-3-7643-8753-2

987654321

www.birkhauser.ch

Contents Preface

xi

0 Introduction

1

Part I Convolution equations, canonical factorization and the state space method

7

1 The 1.1 1.2 1.3

role of canonical factorization in solving convolution equations Wiener-Hopf integral equations and factorization . . . . . . . Block Toeplitz equations and factorization . . . . . . . . . . . Singular integral equations and factorization . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

9 9 13 15 17

2 The 2.1 2.2 2.3 2.4 2.5 2.6

state space method and factorization Preliminaries on realization . . . . . . . . Realization of rational matrix functions . Realization of analytic operator functions Inversion . . . . . . . . . . . . . . . . . . . Products . . . . . . . . . . . . . . . . . . . Factorization . . . . . . . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

19 19 22 23 26 27 30 33

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

Part II Convolution equations with rational matrix symbols 3 Explicit solutions using realizations 3.1 Canonical factorization of rational matrix functions in form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Wiener-Hopf integral operators . . . . . . . . . . . . . 3.3 Block Toeplitz operators . . . . . . . . . . . . . . . . . 3.4 Singular integral equations . . . . . . . . . . . . . . . . 3.5 The Riemann-Hilbert boundary value problem . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . .

35 37 state . . . . . . . . . . . . . . . . . .

space . . . . . . . . . . . . . . . . . . . . . . . .

37 42 46 50 51 56

vi

Contents

4 Factorization of non-proper rational matrix functions 4.1 Preliminaries about matrix pencils . . . . . . . . . . . . . . . 4.2 Realization of a non-proper rational matrix function . . . . . 4.3 Explicit canonical factorization . . . . . . . . . . . . . . . . . 4.4 Inversion of singular operators with a rational matrix symbol 4.5 The Riemann-Hilbert boundary value problem revisited (1) . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

Part III Equations with non-rational symbols

57 57 59 61 68 71 74

75

5 Factorization of matrix functions analytic in a strip 5.1 Exponentially dichotomous operators and bisemigroups . . 5.2 Spectral splitting and proof of Theorem 5.2 . . . . . . . . . 5.3 Realization triples . . . . . . . . . . . . . . . . . . . . . . . 5.4 Construction of realization triples . . . . . . . . . . . . . . . 5.5 Inverting matrix functions analytic in a strip . . . . . . . . 5.6 Inverting full line convolution operators . . . . . . . . . . . 5.7 Inverting Wiener-Hopf integral operators . . . . . . . . . . . 5.8 Explicit canonical factorization . . . . . . . . . . . . . . . . 5.9 The Riemann-Hilbert boundary value problem revisited (2) Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

77 78 81 87 91 93 98 100 106 111 113

6 Convolution equations and the transport equation 6.1 The transport equation . . . . . . . . . . . . . . . 6.2 The case of a ﬁnite number of scattering directions 6.3 Wiener-Hopf equations with operator-valued kernel 6.4 Construction of a canonical factorization . . . . . . 6.5 The matching of the subspaces . . . . . . . . . . . 6.6 Formulas for solutions . . . . . . . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . functions . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

115 116 118 122 124 135 138 142

7 Wiener-Hopf factorization and factorization indices 7.1 Canonical factorization of operator functions . . 7.2 Proof of Theorem 7.2 . . . . . . . . . . . . . . . . 7.3 Wiener-Hopf factorization and spectral invariants Notes . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

143 143 147 157 167

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

Contents

vii

Part IV Factorization of selfadjoint rational matrix functions 8 Preliminaries concerning minimal factorization 8.1 Minimal realizations . . . . . . . . . . . . 8.2 Minimal factorization . . . . . . . . . . . 8.3 Pseudo-canonical factorization . . . . . . Notes . . . . . . . . . . . . . . . . . . . .

169

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

171 171 174 176 178

9 Factorization of positive deﬁnite rational matrix functions 9.1 Preliminaries on selfadjoint rational matrix functions 9.2 Spectral factorization . . . . . . . . . . . . . . . . . . 9.3 Positive deﬁnite functions on the unit circle . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

181 181 185 189 195

. . . .

. . . .

. . . .

. . . .

. . . .

10 Pseudo-spectral factorizations of selfadjoint rational matrix functions 197 10.1 Nonnegative rational matrix functions . . . . . . . . . . . . . . . . 197 10.2 Selfadjoint rational matrix functions and further generalizations . . 205 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 11 Review of the theory of matrices in indeﬁnite inner product spaces 11.1 Subspaces of indeﬁnite inner product spaces . . . . . . . . . . . 11.2 H-selfadjoint matrices . . . . . . . . . . . . . . . . . . . . . . . 11.3 H-dissipative matrices . . . . . . . . . . . . . . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

Part V Riccati equations and factorization 12 Canonical factorization and Riccati equations 12.1 Preliminaries on spectral angular subspaces . 12.2 Angular operators and factorization . . . . . 12.3 Riccati equations and canonical factorization 12.4 Left versus right canonical factorization . . . Notes . . . . . . . . . . . . . . . . . . . . . .

211 211 212 215 216

217 . . . . .

. . . . .

13 The symmetric algebraic Riccati equation 13.1 Spectral factorization and Riccati equations . . . 13.2 Stabilizing solutions . . . . . . . . . . . . . . . . 13.3 Symmetric Riccati equations and pseudo-spectral Notes . . . . . . . . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

219 219 221 227 229 231

. . . . . . . . . . . . . . . . factorization . . . . . . . .

. . . .

. . . .

233 233 238 242 247

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

viii

Contents

14 J-spectral factorization 14.1 Deﬁnition of J-spectral factorization . . . . . . . . . . . . 14.2 J-spectral factorizations and invariant subspaces . . . . . 14.3 J-spectral factorizations and Riccati equations . . . . . . 14.4 Two special cases of J-spectral factorization . . . . . . . . 14.5 J-spectral factorization with respect to other contours . . 14.6 Left versus right J-spectral factorization . . . . . . . . . . 14.7 J-spectral factorization relative to the unit circle revisited Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

Part VI Factorizations and symmetries

249 249 251 256 259 262 273 276 288 289

15 Factorization of positive real rational matrix functions 15.1 Rational matrix functions with a positive deﬁnite real part 15.2 Canonical factorization of functions with a positive deﬁnite real part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.3 Generalization to pseudo-canonical factorization . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

291 . . . . 291 . . . . 294 . . . . 297 . . . . 300

16 Contractive rational matrix functions 16.1 State space analysis of contractive rational matrix functions 16.2 Strictly contractive rational matrix functions . . . . . . . . 16.3 An application to spectral factorization . . . . . . . . . . . 16.4 An application to canonical factorization . . . . . . . . . . . 16.5 A generalization to pseudo-canonical factorization . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

301 301 304 305 307 309 312

17 J-unitary rational matrix functions 17.1 Realizations of J-unitary rational matrix functions . 17.2 Factorization of J-unitary rational matrix functions 17.3 Factorization of unitary rational matrix functions . . 17.4 Intermezzo on the Redheﬀer transformation . . . . . 17.5 J-inner rational matrix functions . . . . . . . . . . . 17.6 Inner-outer factorization . . . . . . . . . . . . . . . . 17.7 Unitary completions of minimal degree . . . . . . . . 17.8 Bi-inner completions of inner functions . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

313 313 321 324 328 333 336 339 341 345

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

Contents

ix

Part VII Applications of J-spectral factorizations 18 Application to the rational Nehari problem 18.1 Problem statement and main result . . . 18.2 Intermezzo about linear fractional maps 18.3 The J-spectral factorization approach . 18.4 Proof of the main result . . . . . . . . . 18.5 The case of a non-stable given function . 18.6 The Nehari-Takagi problem . . . . . . . Notes . . . . . . . . . . . . . . . . . . .

347

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

349 349 352 359 361 366 368 370

19 Review of some control theory for linear systems 371 19.1 Stability and feedback . . . . . . . . . . . . . . . . . . . . . . . . . 371 19.2 Parametrization of internally stabilizing compensators . . . . . . . 374 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378 20 H-inﬁnity control applications 20.1 The standard problem and model matching . 20.2 The one-sided model matching problem . . . 20.3 The two-sided model matching problem . . . 20.4 State space solution of the standard problem Notes . . . . . . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

379 380 382 386 392 404

Bibliography

405

List of symbols

415

Index

419

Preface The present book deals with canonical factorization problems for diﬀerent classes of matrix and operator functions. Such problems appear in various areas of mathematics and its applications. The functions we consider have in common that they appear in the state space form or can be represented in such a form. The main results are all expressed in terms of the matrices or operators appearing in the state space representation. This includes necessary and suﬃcient conditions for canonical factorizations to exist and explicit formulas for the corresponding factors. Also, in the applications the entries in the state space representation play a crucial role. The theory developed in the book is based on a geometric approach which has its origins in diﬀerent ﬁelds. One of the initial steps can be found in mathematical systems theory and electrical network theory, where a cascade decomposition of an input-output system or a network is related to a factorization of the associated transfer function. Canonical factorization has a long and interesting history which starts in the theory of convolution equations. Solving Wiener-Hopf integral equations is closely related to canonical factorization. The problem of canonical factorization also appears in other branches of applied analysis and in mathematical systems theory, in H∞ -control theory in particular. The ﬁrst book devoted to the state space factorization theory was published in 1979 as the monograph “Minimal factorization of matrix and operator functions,” Operator Theory: Advances and Applications 1, Birkh¨ auser Verlag, written by the ﬁrst three authors. Some of the factorization results published in the 1979 book appeared there in print for the ﬁrst time. The present book is the second book written by the four of us in which the state space factorization method is systematically used and developed further. In the earlier book [20], published in 2008, the emphasis is on non-canonical factorizations and degree 1 factorizations, in particular. In the present book we concentrate on canonical factorizations. Together both books present a rich and far reaching update of the 1979 monograph [11]. In the present book the emphasis is on canonical factorization and symmetric factorization with applications to diﬀerent classes of convolution equations. For

xii

Preface

the latter we have in mind the transport equation, singular integral equations, equations with symbols analytic in a strip, and equations involving factorization of non-proper rational matrix functions. A large part of the book will deal with factorization of matrix functions satisfying various symmetries. A main theme will be the eﬀect of these symmetries on factorization and how the symmetries can be used in eﬀective ways to get state space formulas for the factors. Applications to H∞ -control theory, which have been developed in the 1980s and 1990s, will also be included. The text is largely self-contained, and will be of interest to experts and students in mathematics, sciences and engineering. The authors gratefully acknowledge a visitor fellowship for the second author from the Netherlands Organization for Scientiﬁc Research (NWO), and the ﬁnancial support from the School of Economics of the Erasmus University at Rotterdam, from the School of Mathematical Sciences of Tel-Aviv University and the Nathan and Lily Silver Family Foundation, and from the Mathematics Department of the Vrije Universiteit at Amsterdam. These funds allowed us to meet and to work together on the book for diﬀerent extended periods of time in Amsterdam and Tel-Aviv. The authors

Amsterdam – Rotterdam – Tel-Aviv, Summer 2009

Postscript On Monday October 12, 2009, Israel Gohberg, the second author of this book, passed away at the age of 81. At that time the preparation of the book was in a ﬁnal phase and only some minor work had to be done. Israel Gohberg was one of the initiators using state space methods in solving problems appearing in various branches of mathematical analysis and its applications. His fundamental insights and inspiring leadership have been driving forces in our joint work.

Chapter 0

Introduction This monograph presents a uniﬁed approach for solving canonical factorization problems for diﬀerent classes of matrix and operator functions. The notion of canonical factorization originates from the theory of convolution equations. For instance, canonical factorization, provided it exists, allows one to invert WienerHopf, Toeplitz and singular integral operators, and when the factors are known one can also build explicitly the inverses of these operators. The problem of canonical factorization also appears in various branches of applied analysis, in linear transport theory, in interpolation theory, in mathematical systems theory, in particular, in H∞ -control theory. The various matrix and operator functions that are considered in this book have in common that they appear in a natural way as functions of the form W (λ) = D + C(λI − A)−1 B

(1)

or (after a suitable transformation) can be represented in this form. In the above formula λ is a complex variable, and A, B, C, and D are matrices or linear operators acting between appropriate Banach or Hilbert spaces, which in this book often will be ﬁnite dimensional. When the underlying spaces are all ﬁnite dimensional, A, B, C, and D can be viewed as matrices and the function W is a rational matrix function which is analytic at inﬁnity. From mathematical systems theory it is known that, conversely, any rational matrix function which is analytic at inﬁnity admits a representation of the above form. In systems theory the right hand side of (1) is called a state space realization of the function W , and one refers to the space in which A is acting as the state space. The method of factorization employed in this book uses realizations as in (1), and for this reason it is referred to as the state space method. It allows one to deal with factorization from a geometric point of view. This state space factorization approach has its origins in diﬀerent ﬁelds, for instance, in the theory of non-selfadjoint operators [27], [141], in mathematical systems theory and electrical

2

Chapter 0. Introduction

network theory [23], [95], [94], and in the factorization theory of matrix polynomials [67], [131]. In all three areas a state space representation of the function to be factored is used, and the factors are also expressed in state space form. The ﬁrst book to deal with factorization problems in a systematic way using the state space approach is the monograph [11] of the ﬁrst three authors. This monograph appeared in 1979, very soon after the ﬁrst main results were obtained. In fact, some of the factorization results were published in [11] for the ﬁrst time. The present book is the second book written by the four of us in which the state space factorization method is systematically used and developed further. In our ﬁrst book [20], published in 2008, the emphasis is on non-canonical factorizations and degree 1 factorizations, in particular. In the present book we concentrate on canonical factorizations. As a result the overlap between the main parts of the two books is minor. Together both books present a rich and far reaching update of the 1979 monograph [11]. In the present book special attention is paid to various factorizations with additional symmetries such as spectral factorization, inner-outer factorization, and J-spectral factorization. The latter require elements of the theory of spaces with an indeﬁnite metric. Factorizations with symmetries appear in a natural way in H∞ -control problems and the related Nehari approximation problem. In fact, the latter problems are the main topic of the ﬁnal part of the book. We also deal with applications to problems in the theory of algebraic Riccati equations, to inversion problems for Wiener-Hopf, Toeplitz and singular integral operators, and to Riemann-Hilbert problems. The linear transport equation from mathematical physics is another important area of application in this book. It requires inﬁnite dimensional realizations of a special type. We have made an eﬀort to make the text reasonably self-contained. For that reason we included some known material about realizations, minimal factorizations of rational matrix functions, angular operators, and the theory of matrices in indeﬁnite inner product spaces. In the ﬁnal part we also brieﬂy review elements of control theory of linear systems. Not counting the present introduction, the book consists of 20 chapters grouped into 7 parts. We shall now give a short description of the contents of the book. Part I. The ﬁrst part has a preparatory character. In the ﬁrst chapter we review the role of canonical factorization in inverting Wiener-Hopf integral operators and block Toeplitz operators. Also the role of this factorization in solving singular integral equations is described. The second chapter presents in detail the elements of the state space method that are used in this book. Part II. This part starts with the canonical factorization theorem for rational matrix functions in state space form. This theorem is then used to invert explicitly Wiener-Hopf, Toeplitz and singular integral operators with a rational matrix symbol, with the inverses being presented explicitly in state space formulas. For

3 rational matrix symbols the solution to the homogeneous Riemann-Hilbert boundary value problem is also given in state space form. In the ﬁrst chapter of this part we consider proper rational matrix functions, that is, rational matrix functions that are analytic at inﬁnity. The case of non-proper rational symbols is treated in the second chapter of this part. In this case the realization (1) is replaced by W (λ) = I + C(λG − A)−1 B,

(2)

where I is an identity matrix, G and A are square matrices, and B and C are matrices of appropriate sizes. A square rational matrix function, proper or not, always admits such a realization. We develop this realization result, and prove a canonical factorization theorem for the realization (2). As an application we solve the homogeneous Riemann-Hilbert boundary value problem for an arbitrary rational matrix symbol. Part III. In this part we carry out a program analogous to that of the second part, but now for certain classes of non-rational matrix and operator functions. For instance, for matrix functions analytic on a strip but not at inﬁnity we develop a realization theory, prove a canonical factorization theorem in state space form, and develop its applications to Wiener-Hopf integral equations. A new feature is that the problems involved require us to employ realizations with an unbounded main operator A and deal with curves cutting through the spectrum of this main operator. In this part it is also shown that, after an appropriate modiﬁcation, the state space method can be used to solve the integro-diﬀerential equation appearing in linear transport theory, which forces us to use realizations of operator-valued functions. In the ﬁnal chapter of this part we make an excursion into non-canonical Wiener-Hopf factorization for analytic operator-valued functions on a curve, and identify the so-called factorization indices in state space terms. Part IV. The fourth part deals with factorization of rational matrix functions that have Hermitian values on the imaginary axis, the real line or the unit circle. In the analysis of such functions, minimal realizations play an important role. These are realizations of which the order of the state matrix in (1) is a small possible. Also the so-called state space similarity theorem, which tells us that a minimal realization is unique up to a basis transformation in the state space, enters into the analysis. These facts are reviewed in the ﬁrst chapter of this part. In this ﬁrst chapter, using the notion of local minimality, also the concept of a pseudocanonical factorization relative to a curve is introduced and studied for rational matrix functions with singularities on the given curve. The eﬀect on minimal realizations of the function having Hermitian values on the imaginary axis, the real line or the unit circle is described in the second chapter of this part. This then leads to the construction of special canonical and pseudo-canonical factorizations with additional relations between the factors. Included are spectral factorization for positive deﬁnite rational matrix functions and pseudo-spectral factorization for nonnegative rational matrix functions. In the ﬁnal chapter we present (without proofs) some background material on matrices in indeﬁnite inner product spaces,

4

Chapter 0. Introduction

and review the main results from this area that are used in this book. Part V. In this part the canonical factorization theorem is presented in a diﬀerent way using the notion of an angular subspace and Riccati equations. In this case one has to look for angular subspaces that are also spectral subspaces, and the solutions of the Riccati equation must have additional spectral properties. These results, which have a preliminary character, are presented in the ﬁrst chapter of this part. In the second chapter we introduce the symmetric algebraic Riccati equation, and describe spectral factorization as well as pseudo-spectral factorization in terms of Hermitian solutions of such a Riccati equation. In the ﬁnal chapter of this part we continue the study of rational matrix functions that take Hermitian values on certain curves. The emphasis will be on rational matrix functions that have Hermitian values for which the inertia is independent of the point on the curve. Such functions may still admit a symmetric canonical factorization, provided we allow for a constant Hermitian invertible matrix in the middle. Such a factorization is commonly known as a J-spectral factorization. Necessary and suﬃcient conditions for its existence are given, ﬁrst in terms of invariant subspaces and then in terms of solutions of a corresponding symmetric algebraic Riccati equation. We also study the question when a function which admits a left J-spectral factorization admits a right J-spectral factorization too. Part VI. In this part we study rational matrix functions that are unitary or of the form identity matrix plus contractions, and rational matrix functions that have a positive real part. Because of the state space similarity theorem, these additional symmetries can be restated in terms of special properties of the minimal realizations of the rational matrix functions considered. These reformulations involve an algebraic Riccati equation. The results are known in systems theory as the bounded real lemma and the positive real lemma, respectively. They allow us to solve related canonical and pseudo-canonical factorization problems in state space form. In the ﬁnal chapter of this part realizations are used to analyze rational matrix functions of which the values on the imaginary axis are J-unitary matrices. Solutions to various factorization problems are given. Special attention is paid to factorization of J-unitary rational matrix functions into J-unitary factors. In this chapter we also discuss problems of embedding a contractive rational matrix function into a unitary rational matrix function of larger size. Part VII. In this part the state space theory of J-spectral factorization, developed in the ﬁnal chapter of the ﬁfth part, is used to solve H∞ problems. The ﬁrst chapter of this part contains the solution of the Nehari interpolation problem for rational matrix interpolants. The second chapter presents a short review of elements of control theory that play an important role in the third (and ﬁnal) chapter of this part. This ﬁnal chapter is about H∞ -control. Here we use the Jspectral factorization theory to obtain the solutions of some of the main problems in this area, namely the standard problem, the one-sided problem, and the full model matching problem.

5 As the description of the contents given above shows, the emphasis in the book is mainly on rational matrix functions and ﬁnite dimensional realizations. An exception is Part III. The latter part deals with non-rational matrix functions and operator-valued functions, and it uses realizations that have an inﬁnite dimensional state space. Other exceptions are Chapter 2 in Part I and Chapter 12 in Part V. For the material in the other chapters of the book, in particular, in Parts IV–VII, often extensions to an inﬁnite dimensional setting exist; they require appropriate modiﬁcations. See, e.g., the books [5], [35], [42], [73], and the references therein. A few remarks about terminology and notation At the end of this book, after the bibliography, the reader will ﬁnd a List of Symbols and an Index. The latter contains in alphabetical order the various terms that are used in this book with references to the pages where they are introduced. In addition, we would like to mention the following. In the sequel, whenever convenient, a p × q matrix with complex entries will be identiﬁed with the (linear) operator from Cq into Cp deﬁned by the canonical action of the matrix on the standard orthogonal basis of Cq . Conversely, a linear operator from Cq into Cp is identiﬁed with its p × q matrix representation with respect to the standard orthogonal bases of Cq and Cp .

F−

F +

Γ

Throughout the word “operator” refers to a bounded linear transformation acting between Banach or Hilbert spaces (ﬁnite or inﬁnite dimensional). We assume the reader to be familiar with Sections I.1 and I.2 in [51] which contain the standard spectral theory of operators, including the notion of a Riesz projection and the corresponding functional calculus (see, also Chapter V in [144]). In particular, we shall often use the notions of a Cauchy domain and Cauchy contour which are deﬁned as follows. A Cauchy domain is an open set in the complex plane C consisting of a ﬁnite number of components such that its boundary is composed of a ﬁnite number of simple closed non-intersecting rectiﬁable curves. A Cauchy contour Γ is the positively oriented boundary of a bounded Cauchy domain. We write F+ for the interior domain of Γ, and F− for the exterior domain, i.e., the

6

Chapter 0. Introduction

complement of the closure F+ of F+ in the Riemann sphere C∞ = C ∪ {∞}. The picture on the previous page illustrates this notion. We shall also work with the extended real line and the extended imaginary axis as contours on the Riemann sphere C∞ . For the real line the orientation will be from left to right and for the imaginary axis from bottom to top. Thus for the extended real line the interior domain is the open upper half plane, which will be denoted by C+ ; for the extended imaginary axis it is the open left half plane, which is denoted by Cleft . We shall also freely use the Lesbesgue integral and related Lp spaces (see, e.g., Appendix 2 in [53]). Functions which are equal almost everywhere (shorthand: a.e.) are often identiﬁed, sometimes without explicitly mentioning this. Finally, when dealing with inner-outer factorization, we shall always assume that the outer factor is invertible outer (see Section 17.6). In the outer-co-inner factorizations considered in this book, the outer factor will be assumed to be invertible outer as well.

Part I Convolution equations, canonical factorization and the state space method This part has a preparatory character. It consists of two chapters. In the ﬁrst chapter we review the role of canonical factorization in inverting Wiener-Hopf integral operators and block Toeplitz operators. The role of this factorization in solving singular integral equations is described as well. The second chapter presents in detail the basic elements of the state space method that are used throughout this book. The central notion is that of a realization of a matrix or operator function. Three important operations on realizations are studied.

Chapter 1

The role of canonical factorization in solving convolution equations This chapter has a preparatory character. We review (without giving proofs) the role of canonical factorization in inverting systems of convolution equations. The chapter consists of three sections. Section 1.1 deals with Wiener-Hopf integral equations, Section 1.2 with block Toeplitz equations, and Section 1.3 with singular integral equations.

1.1 Wiener-Hopf integral equations and factorization In this section we outline the factorization method of [61] to solve systems of Wiener-Hopf integral equations. Such a system may be written as a single vectorvalued Wiener-Hopf equation ∞ φ(t) − k(t − s)φ(s) ds = f (t), t ≥ 0. (1.1) 0

(−∞, ∞), that is, Here φ and f are m-dimensional vector functions and k ∈ Lm×m 1 the kernel function k is an m×m matrix function whose entries are in L1 (−∞, ∞). We assume that the given vector function f has its component functions in the Lebesgue space Lp [0, ∞), and we express this property by writing f ∈ Lm p [0, ∞). Throughout this section p will be ﬁxed and 1 ≤ p < ∞. The problem we shall consider is to ﬁnd a solution φ of equation (1.1) that also belongs to the space Lm p [0, ∞). The usual method (see [61]) for solving equation (1.1) is as follows. First assume that (1.1) has a solution φ in Lm p [0, ∞). Extend φ and f to the full real

10

Chapter 1. The role of canonical factorization

line by putting f (t) = −

φ(t) = 0,

0

∞

k(t − s)φ(s) ds,

t < 0.

Then φ, f ∈ Lm p (−∞, ∞) and the full line convolution equation φ(t) −

∞

−∞

k(t − s)φ(s) ds = f (t),

−∞ < t < ∞

is satisﬁed. By applying the Fourier transformation and leaving the part of f that is given in the right-hand side, one gets W (λ)Φ+ (λ) − F− (λ) = F+ (λ),

λ ∈ R,

(1.2)

where W (λ) = Im − Φ+ (λ) =

∞

∞

e

iλt

k(t) dt,

−∞

eiλt φ(t) dt,

0

F+ (λ) = F− (λ) =

∞

0 0

eiλt f (t) dt,

(1.3)

eiλt f (t) dt.

(1.4)

−∞

Here Im is the m×m identity matrix. Note that the functions K and F+ are given, but the functions Φ+ and F− have to be found. In fact in this way the problem to solve (1.1) is reduced to that of ﬁnding two functions Φ+ and F− such that (1.2) holds, while furthermore Φ+ and F− must be as in (1.4) with φ ∈ Lm p [0, ∞) and f ∈ Lm (−∞, 0]. p To ﬁnd Φ+ and F− of the desired form such that (1.2) holds, one factorizes the m×m matrix function W appearing in (1.2). This function is called the symbol of the integral equation (1.1). Note that W is continuous on the real line, and by the Riemann-Lebesgue lemma limλ∈R, λ→∞ W (λ) exists and is equal to Im . Assume that the symbol admits a factorization of the following form: λ ∈ R, (1.5) W (λ) = Im + G− (λ) Im + G+ (λ) , where

G+ (λ) =

with g+ ∈ nants

∞

e 0

Lm×m [0, ∞) 1

iλt

g+ (t) dt,

and g− ∈

G− (λ) =

Lm×m (−∞, 0] 1

det Im + G+ (λ) ,

0

−∞

eiλt g− (t) dt,

while, in addition, the determi-

det Im + G− (λ)

do not vanish in the closed upper and lower half plane, respectively. We shall refer to the factorization (1.5) as a right canonical factorization of W with respect to

1.1. Wiener-Hopf integral equations and factorization

11

−1 the real line. Under the conditions stated above the functions Im + G+ (λ) and −1 admit representations as Fourier transforms: Im + G− (λ) ∞ −1 = Im + eiλt γ+ (t) dt, (1.6) Im + G+ (λ)

−1 Im + G− (λ)

=

Im +

0

0 −∞

eiλt γ− (t) dt,

(1.7)

with γ+ ∈ Lm×m [0, ∞) and γ− ∈ Lm×m (−∞, 0]. Using the factorization (1.5) and 1 1 omitting the variable λ, equation (1.2) can be rewritten as (Im + G+ )Φ+ − (Im + G− )−1 F− = (Im + G− )−1 F+ .

(1.8)

Let P be the projection acting on the Fourier transforms of Lm p (−∞, ∞)-functions according to the following rule: ∞ ∞ eiλt h(t) dt = eiλt h(t) dt. P −∞

0

Applying P to (1.8) one gets (Im + G+ )Φ+ = P (Im + G− )−1 F+ , and hence

Φ+ = (Im + G+ )−1 P (Im + G− )−1 F+ ,

(1.9)

which is the formula for the solution of equation (1.2). To obtain the solution φ of the original equation (1.1), i.e., to obtain the inverse Fourier transform of Φ+ , one can employ the formulas (1.6) and (1.7). In fact ∞ γ(t, s)f (s) ds, t ≥ 0, φ(t) = f (t) + 0

where the m × m matrix function γ(t, s) is given by γ(t, s) = γ+ (t − s) + γ− (t − s) +

0

min(t, s)

γ+ (t − r)γ− (r − s) dr.

We conclude the description of this factorization method by mentioning that the m equation (1.1) has a unique solution in Lm p [0, ∞) for each f in Lp [0, ∞) if and only if its symbol admits a factorization as in (1.5). For details, see [50], [61]. Let T be the Wiener-Hopf integral operator on Lm p [0, ∞) associated with equation (1.1), that is, T is the operator on Lm [0, ∞) given by p (T φ)(t) = φ(t) −

0

∞

k(t − s)φ(s) ds,

t ≥ 0.

12

Chapter 1. The role of canonical factorization

The function W in the left-hand side of (1.3) is also referred to as the symbol of T . Obviously the operator T is invertible if and only if the equation (1.1) has a m unique solution in Lm p [0, ∞) for each f in Lp [0, ∞). Thus the results reviewed above can be summarized as follows. Theorem 1.1. Let T be the Wiener-Hopf integral operator on Lm p [0, ∞) with symbol W . Then T is invertible if and only if W admits a right canonical factorization with respect to the real line. Furthermore, if (1.5) is such a factorization of W , then the inverse of T is the integral operator given by ∞ γ(t, s)f (s) ds, t ≥ 0, (T −1 f )(t) = f (t) + 0

where the kernel function γ is deﬁned by s ⎧ ⎪ ⎪ γ (t − s) + γ+ (t − r)γ− (r − s) dr, ⎨ + 0 γ(t, s) = t ⎪ ⎪ ⎩ γ− (t − s) + γ+ (t − r)γ− (r − s) dr, 0

0 ≤ s < t, (1.10) 0≤t<s

with γ− and γ+ as in (1.6) and (1.7), respectively. To illustrate the method, let us consider a special choice for the right-hand side f (cf., [61]). Take f (t) = e−iqt x0 , (1.11) where x0 is a ﬁxed vector in Cm and q is a complex number with q < 0. Then ∞ i x0 , F+ (λ) = ei(λ−q)t x0 dt = λ ≥ 0. λ−q 0 Now observe that −1 −1 i − Im + G− (q) x0 Im + G− (λ) λ−q is the Fourier transform of an Lm p (−∞, 0]-function and hence it vanishes when the projection P is applied. It follows that in the present case the formula for Φ+ may be written as Φ+ (λ) =

−1 −1 i Im + G+ (λ) Im + G− (q) x0 . λ−q

Recall that the solution φ is the inverse Fourier transform of Φ+ . So we have t −1 eiqs γ+ (s) ds Im + G− (q) x0 . φ(t) = e−iqt Im + 0

(1.12)

1.2. Block Toeplitz equations and factorization

13

1.2 Block Toeplitz equations and factorization In this section we consider the discrete analogue of a Wiener-Hopf integral equation, that is, a block Toeplitz equation . So we consider an equation of the type ∞

aj−k ξk = ηj ,

j = 0, 1, 2, . . . .

(1.13)

k=0

Throughout we assume that the coeﬃcients aj are given complex m × m matrices satisfying ∞

aj < ∞, (1.14) j=−∞

η = (ηj )∞ j=0 is a given (ξk )∞ ∈ m p such that k=0

m and vector from m p = p (C ). The problem is to ﬁnd ξ = (1.13) is satisﬁed. We shall restrict ourselves to the case 1 ≤ p ≤ 2; the ﬁnal results however are valid for 2 < p ≤ ∞ as well. Assume ξ ∈ m p is a solution of (1.13). Then one can write (1.13) in the form ∞

aj−k ξk = ηj ,

j = 0, ±1, ±2, . . . ,

(1.15)

k=−∞

where ξk = 0 for k < 0 and ηj is deﬁned by (1.15) for j < 0. Multiplying both sides of (1.15) by λj with |λ| = 1 and summing over j, one gets a(λ)ξ+ (λ) − η− (λ) = η+ (λ), where a(λ) =

∞

λj aj ,

η+ (λ) =

j=−∞

ξ+ (λ) =

∞

|λ| = 1, ∞

λj ηj ,

(1.16)

(1.17)

j=0

λj ξj ,

η− (λ) =

j=0

−1

λj ηj .

j=−∞

In this way the problem to solve (1.13) is reduced to that of ﬁnding two sequences ξ+ and η− such that (1.16) holds, while moreover, ξ+ and η− must be as in (1.2) ∞ m with (ξj )∞ j=0 and (η−j−1 )j=0 from p . The usual way (cf., [61] or the book [40]) of solving (1.16) is again by factorizing the symbol a(λ) of the given block Toeplitz equation. Assume that a(λ) admits a right canonical factorization with respect to the unit circle . By deﬁnition this means that a(λ) can be written as a(λ) = h+ (λ)

=

h− (λ)h+ (λ), ∞

j=0

λj h+ j ,

|λ| = 1, h− (λ) =

0

j=−∞

(1.18) λj h− j ,

14

Chapter 1. The role of canonical factorization

− ∞ m×m ∞ where (h+ of all absolutely convergent j )j=0 and (h−j )j=0 belong to the space 1 sequences of complex m× m matrices, det h+ (λ) = 0 for |λ| ≤ 1 and det h− (λ) = 0 −1 for |λ| ≥ 1 (including λ = ∞). Then h−1 + and h− also admit a representation of the form ∞ 0

−1 j + (λ) = λ γ , h (λ) = λj γj− , (1.19) h−1 + − j j=0

j=−∞

− ∞ m×m . Deﬁning the projection P by with (γj+ )∞ j=0 and (γ−j )j=0 from 1

P

∞

j

λ bj

j=−∞

=

∞

λj bj ,

j=0

one gets from (1.16) and (1.18) −1 ξ+ = h−1 + P h− η+ .

(1.20)

Here, for convenience, the variable λ is omitted. The solution of the original equation (1.13) can now be written as ∞

ξk =

γks ηs ,

k = 0, 1, . . . ,

(1.21)

s=0

where

γks =

⎧ s

⎪ ⎪ ⎪ γ+ γ− , ⎪ ⎪ ⎨ r=0 k−r r−s

s ≤ k,

k ⎪

⎪ ⎪ + − ⎪ γk−r γr−s , ⎪ ⎩

s ≥ k.

r=0

Note that for s = k both sums in the above formula deﬁne the same matrix. The assumption that a(λ) admits a right canonical factorization as in (1.18) m is equivalent to the requirement that for each η = (ηj )∞ j=0 in p the equation m (1.13) has a unique solution ξ = (ξk )∞ in . For details we refer to [61], [40]. p k=0 m Let T be the block Toeplitz operator on p associated with the Toeplitz equation (1.13), that is, T is the operator on m p given by Tξ = η

⇐⇒

∞

aj−k ξk = ηj ,

j = 0, 1, 2, . . . .

k=0

The function a appearing in the left-hand side of (1.17) is also referred to as the m symbol of T . Obviously T is invertible if and only if for each η = (ηj )∞ j=0 in p m the equation (1.13) has a unique solution ξ = (ξk )∞ in . This allows us to p k=0 summarize the results reviewed above as follows.

1.3. Singular integral equations and factorization

15

Theorem 1.2. Let T be the block Toeplitz operator on m p with symbol a(λ) satisfying (1.14). Then T is invertible if and only a(λ) admits a right canonical factorization with respect to the unit circle. Furthermore, if (1.18) is such a factorization of the function a(λ), then the inverse of T is given by ⎡ ⎤ γ11 γ12 γ13 · · · ⎢γ21 γ22 γ23 · · ·⎥ ⎢ ⎥ T −1 = ⎢γ31 γ32 γ33 · · ·⎥ , ⎣ ⎦ .. .. .. .. . . . . where the matrices γks are deﬁned by ⎧ s

⎪ ⎪ ⎪ γ+ γ− , ⎪ ⎪ ⎨ r=0 k−r r−s γks = k ⎪ ⎪ ⎪ + − ⎪ γk−r γr−s , ⎪ ⎩

s ≤ k, (1.22) s ≥ k,

r=0

with γj+ and γj− being determined by (1.19). By way of illustration, we consider the special case when ηj = q j η0 ,

j = 0, 1, . . . .

(1.23)

Here η0 is a ﬁxed vector in Cm and q is a complex number with |q| < 1. Then clearly 1 η0 , η+ (λ) = |λ| ≤ 1, 1 − λq and one checks without diﬃculty that formula (1.21) becomes ξk = q k

k

−1 q −s γs+ h−1 )η0 , − (q

k = 0, 1, . . . .

(1.24)

s=0

This is the analogue of formula (1.12) in the previous section.

1.3 Singular integral equations and factorization In this section we review the factorization method that is used to solve systems of singular integral equations [48]. Consider the singular integral equation 1 φ(τ ) dτ = f (t), t ∈ Γ, (1.25) a(t)φ(t) + b(t) πi Γ τ − t with integration taken over a Cauchy contour Γ. (For the deﬁnition of the latter notion see the ﬁnal paragraphs of Chapter 0 dealing with terminology and notation.) We write F+ for the interior domain of Γ, and F− for the exterior domain

16

Chapter 1. The role of canonical factorization

(i.e., the complement of F + in the Riemann sphere C ∪ {∞}). The functions a and b in (1.25) are given continuous m × m matrix functions deﬁned on Γ, and f is a given function from Lm p (Γ), p ﬁxed, 1 < p < ∞. As usual in the theory of singular integral equations, it is assumed that the interior domain F+ of Γ is connected and contains 0; the exterior domain F− of Γ contains ∞. The problem is to ﬁnd φ ∈ Lm p (Γ) such that(1.25) is satisﬁed. For φ a rational function without poles on Γ we put 1 φ(τ ) (Sφ)(t) = dτ = f (t), t ∈ Γ, (1.26) πi Γ τ − t where the integral is taken in the sense of the Cauchy principal value. The operator S deﬁned in this way can be extended by continuity to a bounded linear operator, again denoted by S, on all of Lm p (Γ). Equation (1.25) can now be written as aIφ + bSφ = f,

(1.27)

where I is the identity operator on Lm p (Γ). In other words, the study of the equation (1.25) reduces to that of the operator aI + bS. Here a and b are viewed as multiplication operators. Equation (1.25) has a unique solution φ ∈ Lm p (Γ) for each choice of f ∈ Lm p (Γ) if and only if the operator aI + bS is invertible as an operator on Lm p (Γ). In the remainder of this section we shall discuss a necessary and suﬃcient condition for this to happen, and we shall give formulas for the inverse (aI + bS)−1 . The operator S enjoys the property S 2 = I. Hence the operators PΓ =

1 (I + S), 2

QΓ =

1 (I − S) 2

are complementary projections on Lm p (Γ). The image of P Γ consists of all functions in Lm (Γ) that admit an analytic continuation into F+ . Similarly, the image of QΓ p is the set of all functions in Lm (Γ) that admit an analytic continuation into F− p vanishing at ∞. Putting c = a + b and d = a − b, one can write the equation (1.27) in the form cP Γ φ + dQΓ φ = f . The following is known (see [62] for the case when the coeﬃcients a and b are scalar functions and [48] for the matrix-valued case). The operator aI + bS = cP Γ + dQΓ is invertible if and only if the matrices c(λ) and d(λ) are invertible for each λ ∈ Γ and the function w given by w(λ) = d(λ)−1 c(λ) admits a right canonical factorization with respect to Γ . By this we mean a factorization w(λ) = w− (λ)w+ (λ),

λ ∈ Γ,

(1.28)

where w− and w+ are m×m matrix functions, analytic and taking invertible values on an open neighborhood of F − and F + , respectively. With the help of (1.28), the −1 operator aI +bS = cP Γ +dQΓ can be rewritten as aI +bS = dw− (w+ P Γ +w− QΓ ),

1.3. Singular integral equations and factorization

17

and its inverse is given by (aI + bS)−1

−1 −1 −1 = (w+ P Γ + w− QΓ )w− d −1 −1 −1 −1 −1 = w+ P Γ w− d + w− QΓ w− d .

(1.29)

Replacing P Γ and QΓ by 12 (I + S) and 12 (I − S), respectively, one gets (aI + bS)−1

1 −1 1 −1 −1 −1 (c + d−1 )I + (w+ − w− )Sw− d 2 2 1 1 −1 −1 [(a + b)−1 + (a − b)−1 ]I + (w+ = − w− )Sw− (a − b)−1 2 2 1 −1 −1 = (a + b)−1 a(a − b)−1 I + (w+ − w− )Sw− (a − b)−1 . 2 =

Summarizing we get the following theorem. Theorem 1.3. The singular integral operator T = aI + bS on Lm p (Γ) is invertible if and only if the matrices a(λ) + b(λ) and a(λ) − b(λ) are invertible for each λ ∈ Γ and the function w given by −1 w(λ) = a(λ) + b(λ) a(λ) + b(λ) admits a right canonical factorization with respect to Γ. Furthermore, if (1.28) is such a factorization of w, then the inverse of T is given by 1 −1 −1 − w− )Sw− (a − b)−1 . T −1 = (a + b)−1 a(a − b)−1 I + (w+ 2

(1.30)

Thus, as before for Wiener-Hopf and block Toeplitz operators, canonical factorization is a useful method for inverting singular integral operators too.

Notes The material in this chapter is standard, and can be found in much more detail and greater generality in various monographs and papers, for instance, see the books [29] and [50]. A ﬁrst introduction to the theory of Wiener-Hopf integral equations and the theory of (block) Toeplitz operators can be found in Chapters XII and XIII of [51] and Chapters XXIII–XXV of [52], respectively. More information can be found in the monographs [37], [62], [63], [64] and [24]. For an extensive review (with many additional references) of the factorization theory of matrix functions with respect to a curve and its applications to inversion of singular integral operators of diﬀerent types, including Wiener-Hopf and block Toeplitz operators, the reader is referred to the recent survey paper [59].

Chapter 2

The state space method and factorization This chapter describes in detail the elements of the state space method that are used throughout this book. The central notion is that of a realization of a matrix or operator function. The chapter consists of six sections. Section 2.1 presents preliminaries on realization, including the relevant deﬁnitions and the connection with systems theory. In the next two sections the realization problem is discussed. First for rational matrix functions in Section 2.2, and then for analytic operator functions in a possibly inﬁnite dimensional setting in Section 2.3. The last three sections are devoted to the main operations on realizations that are needed in this book: inversion (Section 2.4), taking products (Section 2.5), and factorization (Section 2.6).

2.1 Preliminaries on realization Let W be a rational matrix function which is also proper, that is, W has no pole at inﬁnity. As is well-known such a function can always be represented (see the next section for an explicit construction) in the form W (λ) = D + C(λI − A)−1 B.

(2.1)

Here λ is a complex variable, A is a square matrix, I is the identity matrix of the same size as A, and B and C are matrices of appropriate sizes. Since A, B, C and D are matrices, it is immediate from Cramer’s rule that the right-hand side of (2.1) is also a proper rational matrix function. We shall understand the equality in (2.1) as an equality between rational matrix functions, and we shall refer to (2.1) as a matrix-valued realization of W . Sometimes we simply say that the quadruple of matrices (A, B, C, D) is a realization of W . A rational matrix function has many

20

Chapter 2. The state space method and factorization

diﬀerent realizations. Of particular interest are those matrix-valued realizations of W of which the order of the matrix A is as small as possible. These realizations are called minimal ; we shall describe their properties in Chapter 8. For operator-valued functions W , expressions of the type (2.1) are important too but have to be considered with some care. Let W be an L(U, Y )-valued function on a subset Ω of C. Here U and Y are possibly inﬁnite dimensional complex Banach spaces. We say that W admits a realization on Ω whenever W can be written as W (λ) = D + C(λIX − A)−1 B,

λ ∈ Ω.

(2.2)

Here A is a bounded linear operator on a complex Banach space X such that Ω is a subset of ρ(A), the resolvent set of A. Furthermore, IX is the identity operator on X, and B ∈ L(U, X), C ∈ L(X, Y ), D ∈ L(U, Y ), that is B : U → X, C : X → Y, and D : U → Y, are bounded linear operators. The fact that Ω ⊂ ρ(A) implies that the right-hand side of (2.2) is a well-deﬁned bounded linear operator which maps U into Y for each λ ∈ Ω. Also, W (λ) is a bounded linear operator mapping U into Y for each λ ∈ Ω. Note that (2.2) requires these operators to be equal for each λ ∈ Ω. When Ω is open, an obvious necessary condition for W to admit a realization on Ω is that W be analytic on Ω. When Ω is a punctured open neighborhood of ∞, then (2.2) implies limλ→∞ W (λ) = D and so W is proper. Often the identity matrix I in (2.1) and the identity operator IX in (2.2) will be suppressed, and we simply write (λ − A)−1 in place of (λI − A)−1 or (λIX − A)−1 . When X and Y are both ﬁnite dimensional, then the realization (2.2) is called ﬁnite dimensional . In that case W (λ), A, B, C and D can be identiﬁed in the usual way with matrices. In the next two sections we shall address the realization problem, i.e., the question under what conditions a given matrix or operator function admits a realization. First however, we sketch a connection with systems theory which reﬂects itself in some terminology to be introduced at the end of the present section. A system Σ can be considered as a physical object which produces an output in response to an input. Schematically:

u

Σ

y

where u denotes the input and y denotes the output. Mathematically, the input u and the output y are vector-valued functions of a parameter t. The input can

2.1. Preliminaries on realization

21

be chosen freely (at least in principle), but the output is uniquely determined by the choice of the input. The relationship between the input and the output can be quite complicated. Here we consider the simplest model which means that the relationship in question is described by a causal linear time invariant system, i.e., a system of diﬀerential equations of the type ⎧ ⎪ ⎪ x (t) = Ax(t) + Bu(t), ⎨ y(t) = Cx(t) + Du(t), t ≥ 0, Σ (2.3) ⎪ ⎪ ⎩ x(0) = 0, where A, B, C and D are matrices of appropriate sizes, A and D square. Application of the Laplace transform (under appropriate conditions on the input and output functions) changes (2.3) into

λ x(s) =

A x(λ) + B u(λ),

y(λ)

Cx (λ) + D u(λ),

=

and from these expressions one can solve y(λ) in terms of u (λ), resulting in (λ). y(λ) = D + C(λ − A)−1 B u So in what is called the frequency domain, the input-output behavior of (2.3) is determined by the function D + C(λ− A)−1 B, which is called the transfer function of the system (2.3). Note that this function appears in the realized form. The connection with systems theory indicated above is reﬂected in the terminology which is customarily used in dealing with realizations. Returning to (2.2), the space X is usually called the state space of the realization, and the operator A is referred to as its state space operator or main operator . Further we call B the input operator , C the output operator , and D the external operator of (2.2). The realization is called strictly proper when D = 0 and biproper if D is an invertible operator. In the latter case, the operator A − BD−1 C is well-deﬁned. It is referred to by the term the associate state space operator or associate main operator and (by slight abuse of notation as A× does not depend only on A) denoted by A× . This operator will play a crucial role in the inversion and factorization results to be discussed later on. In the situation where U = Y and D is the identity operator, we say that (2.2) is a unital realization. The associate main operator then has the form A× = A − BC. In the case of a matrix-valued realization, the terms state space matrix , main matrix , input matrix , output matrix , external matrix , associate state space matrix , and associate main matrix will be used. Other elements of systems theory involving stability properties, feedback and stabilization, will be reviewed in Chapter 19. These will be of central importance in Chapter 20 (the ﬁnal chapter of the book) which is concerned with H∞ -control.

22

Chapter 2. The state space method and factorization

2.2 Realization of rational matrix functions In this section we construct a matrix-valued realization for a given proper rational (possibly non-square) matrix function. Theorem 2.1. Every proper rational matrix function has a matrix-valued realization. Moreover, the realization can be chosen in such a way that the set of eigenvalues of the main matrix coincides with the set of poles of W . Proof. Let W be a proper rational r × m matrix function, and let wij be the (i, j)-entry of W . Since W is rational, we have wij (λ) =

pij (λ) , qij (λ)

i = 1, . . . , r, j = 1, . . . , n,

where pij and qij are scalar polynomials. The polynomials qij are non-zero and can be taken to be monic. Without loss of generality we may assume that the polynomials pij and qij have no common zero. Taking the least common multiple of the polynomials qij , we obtain a monic polynomial q. Deﬁne ΩW to be the set of all complex λ for which q(λ) = 0. Notice that C \ ΩW is precisely the set of all points in C where W has a pole. One checks without diﬃculty that W has a representation of the form W (λ) = W (∞) +

1 H(λ), q(λ)

λ ∈ ΩW ,

where H is an r ×m matrix polynomial. Since W is proper, this matrix polynomial is either identically equal to zero or it has degree strictly smaller than k, the degree of the scalar polynomial q. Write q(λ) = λk +

k−1

λj qj ,

H(λ) =

j=0

k−1

λj Hj ,

j=0

and, with Ir the r × r identity matrix, ⎡

0 ⎢ I ⎢ A=⎢ ⎢ ⎣ 0

0 0 ..

... ...

0 0

.

...

I

−q0 Ir −q1 Ir .. . −qk−1 Ir

⎤ ⎥ ⎥ ⎥, ⎥ ⎦

⎡ ⎢ ⎢ B=⎢ ⎢ ⎣

H0 H1 .. . Hk−1

⎤ ⎥ ⎥ ⎥, ⎥ ⎦

C = 0 . . . 0 Ir .

Then the resolvent set ρ(A) of A coincides with ΩW , the subset of C on which q takes non-zero values. For λ ∈ ρ(A), deﬁne C1 (λ), . . . , Ck (λ) by [ C1 (λ) C2 (λ) . . . Ck (λ) ] = C(λ − A)−1 .

2.3. Realization of analytic operator functions

23

From the special form of the matrix A (second companion type) we see that Cj+1 (λ) = λCj (λ),

j = 0, . . . , k − 1,

and C1 (λ) = q(λ)−1 I. Hence C(λ − A)

−1

B =

k−1

j=0

Cj+1 (λ)Hj =

1 H(λ). q(λ)

It follows that W (λ) = W (∞) + C(λ − A)−1 B for each λ ∈ ΩW = ρ(A). Thus W has a matrix-valued realization such that the set of eigenvalues of the main matrix A is equal to C \ ΩW . In other words, the set of eigenvalues of A coincides with the set of poles of W, as desired. Let W be a proper rational matrix function. Elaborating on Theorem 2.1 and its proof, we note that W does not admit any realization involving a main matrix A whose spectrum σ(A) is strictly smaller than C \ ΩW , the set of poles of W . Indeed, we would then have a realization of W on an open subset of C strictly larger than ΩW and such a subset would contain a pole of W , contradicting the fact that W has to be analytic on it. It is not diﬃcult to construct realizations of W having a main matrix A with spectrum strictly larger than C \ ΩW and where certain eigenvalues of A (namely those belonging to ΩW ) do not correspond with poles of W . So the realization constructed in the proof of Theorem 2.1 enjoys a certain minimality property. However, it does this only in a weak sense. This one sees, for instance, by looking at the pole orders. If μ is a pole of W , its order as a pole of W is generally strictly smaller than the order of μ as a pole of the resolvent (λ − A)−1 . With the proper notion of minimality to be introduced in Section 8.1, this anomaly disappears so that the two pole orders are the same. The key point is that the state space dimension (which is equal to rk) of the realization of the proof of Theorem 2.1 is generally not the least possible.

2.3 Realization of analytic operator functions In this section we consider the realization problem for possibly non-rational operator functions. First we consider operator functions that are analytic on a bounded Cauchy domain in C. Recall from Chapter 0 that the boundary of such a Cauchy domain consists of a ﬁnite number of simple closed non-intersecting rectiﬁable curves. Theorem 2.2. Let Ω be a bounded Cauchy domain, and let W be an operator function with values in L(U, Y ), where U and Y are complex Banach spaces. Suppose W is analytic on Ω and continuous on the closure of Ω. Then, given a bounded linear operator D : U → Y , there exists a realization for W on Ω having D as its external operator. In particular, if U = Y , then W admits a unital realization on Ω.

24

Chapter 2. The state space method and factorization

Proof. Let Γ be the positively oriented boundary of Ω (so that Ω is the interior domain of Γ). With Γ and U we associate the space C(Γ; U ) of all U -valued continuous functions on Γ endowed with the supremum norm. This will become the state space of the realization to be constructed. Write B for the canonical embedding of U into C(Γ; U ), so (Bu)(z) = u for each u ∈ U and z ∈ Γ. Next, deﬁne C : C(Γ; U ) → Y by setting 1 Cf = D − W (z) f (z) dz, f ∈ C(Γ; U ). 2πi Γ Here D is the given operator from U into Y . Finally, the operator A from C(Γ; U ) into C(Γ; U ) is the multiplication operator given by (Af )(z) = zf (z),

f ∈ C(Γ; U ), z ∈ Γ.

All these operators are linear and bounded. We claim that W (λ) = D + C(λ − A)−1 B,

λ ∈ Ω ⊂ ρ(A).

Take λ ∈ Ω. Then λ − A is invertible with inverse given by (λ − A)−1 g (z) =

1 g(z), λ−z

g ∈ C(Γ; U ), z ∈ Γ.

It follows that (λ − A)−1 Bu (z) =

1 u, λ−z

u ∈ U, z ∈ Γ,

and hence C(λ − A)−1 Bu =

1 2πi

Γ

1 D − W (z) u dz, λ−z

u ∈ U.

By the Cauchy integral formula, the right-hand side of this identity is W (λ)u−Du, and the desired result is immediate. Theorem 2.2 remains true when the conditions on Ω and W are replaced by the simpler hypotheses that Ω is any bounded open set in C and W is just analytic on Ω. In that case the space C(Γ; U ) must be replaced by an appropriate Banach space deﬁned in terms of the behavior of W near the boundary of Ω. For details, cf., [113]; see also the next theorem. Theorem 2.3. Let Ω ⊂ C be an open punctured neighborhood of ∞ in the Riemann sphere C∞ , let U and Y be complex Banach spaces, and let W : Ω → L(U, Y ) be analytic and proper. Then W admits a realization on Ω with external operator D = limλ→∞ W (λ).

2.3. Realization of analytic operator functions

25

Proof. First assume Ω is the full complex plane. Then, by Liouville’s theorem, the function W has the constant value D = limλ→∞ W (λ). Now take for the state space X the zero space {0}, and the desired realization for W on C is obtained trivially. Next, consider the more interesting case where Ω is diﬀerent from C. For notational reasons we will assume that 0 ∈ / Ω. The general case can be reduced to this situation by a simple translation. Deﬁne X to be the space of all Y -valued functions, analytic on Ω ∪ {∞}, such that f (z) < ∞. f • = sup z∈Ω ∪{∞} max(1, W (z)) Taking · • for the norm, X is a Banach space. Introduce B : U → X by ⎧ z ∈ Ω, ⎨ z W (z)u − W (∞)u , (Bu)(z) = ⎩ lim z W (z)u − W (∞)u , z = ∞. z→∞

Further, let C : X → Y be given by Cf = f (∞). Finally, deﬁne A : X → X by ⎧ z ∈ Ω, ⎨ z f (z) − f (∞) , (Af )(z) = ⎩ lim z f (z) − f (∞) , z = ∞. z→∞

All these operators are linear and bounded. We claim that W (λ) = W (∞) + C(λ − A)−1 B,

λ ∈ Ω ⊂ ρ(A).

Take λ ∈ Ω. For g ∈ X, put ⎧ zg(λ) − λg(z) ⎪ , ⎪ ⎪ ⎨ z−λ h(z) = g(λ) − λg (λ), ⎪ ⎪ ⎪ ⎩ g(λ),

z ∈ Ω, z = λ, z = λ, z = ∞,

where g stands for the derivative of g. Then h ∈ X, and by direct computation one sees that (λ − A)h (z) = λg(z), z ∈ Ω ∪ {∞}. Now λ is non-zero (since Ω does not contain the origin), and it follows that λ − A is surjective. But λ − A is injective too. Indeed, if f ∈ X and Af = λf , then f (z) =

z f (∞), z−λ

z ∈ Ω, z = λ,

which, on account of the deﬁnition of the norm · • on X, implies f (∞) = 0 (cf., the behavior of f when z → λ), hence f = 0. It follows that λ ∈ ρ(A) and

26

Chapter 2. The state space method and factorization

(λ − A)−1 g = λ−1 h. We now apply this result tog = Bu with u ∈ U . With this g, we have h(∞) = (Bu)(λ) = λ W (λ)u − W (∞) u, and so (λ − A)−1 Bu (∞) = λ−1 h(∞) = W (λ)u − W (∞) u. In other words C(λ − A)−1 Bu = W (λ)u − W (∞) u. As u ∈ U was taken arbitrarily, we get W (λ) = W (∞) + C(λ − A)−1 B for each λ ∈ Ω.

2.4 Inversion We begin with some heuristics. Consider the realization W (λ) = D + C(λ − A)−1 B,

λ ∈ ρ(A),

(2.4)

and view W as the transfer function of the linear time invariant system ⎧ x (t) = Ax(t) + Bu(t), ⎪ ⎪ ⎨ y(t) = Cx(t) + Du(t), t ≥ 0, Σ ⎪ ⎪ ⎩ x(0) = 0. Assuming that we are in the biproper situation where D is invertible, we can solve u in terms of x and y: u(t) = −D−1 Cx(t) + D−1 y(t),

t ≥ 0.

Inserting this into Σ yields ⎧ x (t) = A× x(t) + BD−1 y(t), ⎪ ⎪ ⎨ u(t) = −D−1 Cx(t) + D−1 y(t), Σ× ⎪ ⎪ ⎩ x(0) = 0.

t ≥ 0,

Here A× = A − BD−1 C is the associate main operator of the given realization as introduced in the last paragraph of Section 2.1. The linear time invariant systems Σ and Σ× can be seen as each other’s inverse. The transfer function of Σ is given by (2.4), the transfer function of Σ× by W × (λ) = D−1 − D−1 C(λ − A× )−1 BD −1 ,

λ ∈ ρ(A× ).

So it is to be expected that W and W × are related by inversion. We shall now make this precise. Theorem 2.4. Consider the biproper realization W (λ) = D + C(λ − A)−1 B,

λ ∈ ρ(A).

2.5. Products

27

Put A× = A − BD−1 C, and take λ ∈ ρ(A). Then W (λ) is invertible if and only if λ belongs to ρ(A× ). In that case, for λ ∈ ρ(A) ∩ ρ(A× ), the following identities hold: W (λ)−1

=

D−1 − D−1 C(λ − A× )−1 BD−1 ,

(λ − A× )−1

=

(λ − A)−1 − (λ − A)−1 BW (λ)−1 C(λ − A)−1 .

Moreover, again for λ ∈ ρ(A) ∩ ρ(A× ), we have W (λ)D−1 C(λ − A× )−1

=

C(λ − A)−1 ,

(λ − A× )−1 BD−1 W (λ)

=

(λ − A)−1 B.

Proof. For λ ∈ ρ(A× ), put W × (λ) = D−1 − D−1 C(λ − A× )−1 BD−1 . Then, when λ ∈ ρ(A) ∩ ρ(A× ), one has W (λ)W × (λ) = D + C(λ − A)−1 B D −1 − D−1 C(λ − A× )−1 BD−1 =

IY + C(λ − A)−1 BD −1 − C(λ − A× )−1 BD −1 + −C(λ − A)−1 BD−1 C(λ − A× )−1 BD−1 .

Now use that BD −1 C = A−A× = (λ − A× )−(λ − A). It follows that W (λ)W × (λ) = IY . Analogously one has W × (λ)W (λ) = IU . The expression for (λ − A× )−1 as well as the last two identities in the theorem are obtained in a similar way. Instead of the previous proof one can also give an argument using Schur complements of the operator matrix A − λI B . C I For details, see the second proof of Theorem 2.1 in [20] or Sections 2 and 4 in [19].

2.5 Products Again we begin with some heuristical remarks. This time we start with two linear time invariant systems ⎧ ⎪ ⎪ x1 (t) = A1 x1 (t) + B1 u1 (t), ⎨ Σ1 t ≥ 0, y1 (t) = C1 x1 (t) + D1 u1 (t), ⎪ ⎪ ⎩ x1 (0) = 0,

28

Chapter 2. The state space method and factorization ⎧ x (t) = ⎪ ⎪ ⎨ 2 y2 (t) = Σ2 ⎪ ⎪ ⎩ x2 (0) =

A2 x2 (t) + B2 u2 (t), t ≥ 0,

C2 x2 (t) + D2 u2 (t), 0,

and we assume that the output y2 of Σ2 can be and is used as the input u1 = y2 for Σ1 , resulting in the cascade synthesis Σ of the systems Σ1 and Σ2 . The input for Σ is u = u2 and the output (modulo u1 = y2 ) is y = y1 . The equations governing the relationship between u and y then are ⎧ x1 (t) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ x (t) ⎪ ⎨ 2 y(t) ⎪ ⎪ ⎪ ⎪ x1 (0) ⎪ ⎪ ⎪ ⎩ x2 (0)

= A1 x1 (t) + B1 C2 x2 (t) + B1 D2 u(t), = A2 x2 (t) + B2 u(t), = C1 x1 (t) + D1 C2 x2 (t) + D1 D2 u(t),

t ≥ 0,

= 0, = 0,

and this is a linear time invariant system which can be rewritten as ⎧ ⎪ x1 A1 B1 C2 x1 B1 D2 ⎪ ⎪ ⎪ = + u, ⎪ ⎪ ⎪ x2 0 A2 x2 B2 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ x1 Σ: C1 D1 C2 y = + D1 D2 u, ⎪ x2 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ x1 0 ⎪ ⎪ ⎪ (0) = . ⎪ ⎩ 0 x2 The transfer functions of Σ1 and Σ2 are W1 (λ)

=

D1 + C1 (λ − A1 )−1 B1 ,

λ ∈ ρ(A1 ),

(2.5)

W2 (λ)

=

D2 + C2 (λ − A2 )−1 B2 ,

λ ∈ ρ(A2 ),

(2.6)

respectively, and the transfer function of Σ is the product W1 W2 of W1 and W2 , in other words W (λ) = W1 (λ)W2 (λ). So our considerations lead to a product formula for realizations. Here are the details. First we specify the spaces associated with the realizations (2.5) and (2.6), and the actions of the operators involved: A 1 : X 1 → X1 ,

B1 : U1 → X1 ,

C1 : X1 → Y1 ,

D1 : U1 → Y1 ,

2.5. Products

29

A2 : X2 → X2 ,

B2 : U2 → X2 ,

C2 : X2 → Y2 ,

D2 : U2 → Y2 .

Now assume Y1 = U2 . Put U = U1 , Y = Y2 , and introduce A1 B1 C2 ˙ 2 → X1 +X ˙ 2, A = : X1 +X 0 A2 B

B1 D2

=

˙ 2, : Y → X1 +X

B2

C1

D1 C2

C

=

D

= D1 D2 : U → Y.

˙ 2 → Y, : X1 +X

Then the following result holds true. Theorem 2.5. Let W1 and W2 be given by the realizations (2.5) and (2.6), respectively. Then, with A, B, C and D as above, W1 (λ)W2 (λ) = D + C(λ − A)−1 B,

λ ∈ ρ(A1 ) ∩ ρ(A2 ) ⊂ ρ(A).

Proof. Take λ ∈ ρ(A1 ) ∩ ρ(A2 ). Then λ ∈ ρ(A). Indeed, λ − A is invertible with inverse given by ⎤ ⎡ −1 (λ − A1 ) H(λ) ⎦ : X1 +X ˙ 2 → X1 +X ˙ 2, (λ − A)−1 = ⎣ −1 0 (λ − A2 ) −1

−1

where H(λ) = − (λ − A1 ) B1 C2 (λ − A2 ) . Employing this, and the expressions for B, C and D given prior to the theorem, D + C(λ − A)−1 B is seen to be equal to

D1 D 2 +

C1

D1 C2

⎡ ⎣

−1

(λ − A1 ) 0

= D 1 D2 + =

C1 (λ − A1 )

−1

D1 + C1 (λ − A1 )−1 B1

⎤⎡

H(λ) −1

⎦⎣

(λ − A2 )

B1 D2

⎤ ⎦

B2

C1 H(λ) + D1 C2 (λ − A2 )

−1

B1 D2

B2

D2 + C2 (λ − A2 )−1 B2 .

Thus D + C(λ − A)−1 B = W1 (λ)W2 (λ), as desired.

The product W1 (λ)W2 (λ) is deﬁned for λ ∈ ρ(A1 ) ∩ ρ(A2 ), a punctured neighborhood of ∞ in C ∪ {∞}. On the other hand D + C(λ − A)−1 B is deﬁned

30

Chapter 2. The state space method and factorization

for λ ∈ ρ(A). As we have seen above ρ(A1 ) ∩ ρ(A2 ) ⊂ ρ(A). In general, this inclusion is strict. Equality occurs, for instance, when the spectra σ(A1 ) and σ(A2 ) of the operators A1 and A2 are disjoint. Another case where one has the equality ρ(A) = ρ(A1 ) ∩ ρ(A1 ) is when ρ(A) is connected. In particular, the equality in question is valid when W1 and W2 are rational matrix functions, and (2.5) and (2.6) are matrix-valued realizations. The realization of Theorem 2.5 is called the product of the realizations (2.5) and (2.6), in that order. The counterpart of taking products is factorization. In the next section this topic will be discussed for functions given by a biproper realization. We close the present section with a remark preparing for this discussion. The main operator A in the product realization is given in the form of a 2 × 2 upper triangular operator matrix: A1 B1 C2 ˙ 2 → X1 +X ˙ 2. A= : X1 +X 0 A2 Analogously, assuming the external operators to be invertible, the associate main operator A× = A − BD−1 C is of 2 × 2 lower triangular type: 0 A× 1 ˙ 2 → X1 +X ˙ 2 : X1 +X A× = B2 D−1 C1 A× 2 −1 × −1 where A× 1 = A1 − B1 D1 C1 and A2 = A2 − B2 D2 C2 are the associate main ˙ {0} is an invariant operators of (2.5) and (2.6), respectively. Note that M = X1 + ˙ 2 is an invariant subspace for A× , and that M subspace for A, that M × = {0} +X and M × match in the sense that the state space of the product realization is the direct sum of M and M × . This state of aﬀairs turns out to be a key point in the discussion of factorization we now turn to.

2.6 Factorization The theorems in this section will serve as a basis for the more involved factorization results to be given in the sequel. Subspaces of Banach spaces are always assumed to be closed, otherwise we use the term linear manifold. For simplicity (and without loss of generality) we assume the external spaces U and Y to be equal. Theorem 2.6. Consider the biproper realization W (λ) = D + C(λIX − A)−1 B,

λ ∈ ρ(A),

(2.7)

and let A× = A − BD−1 C be its associate main operator. Let M and M × be invariant subspaces for A and A× , respectively, and suppose X = M M ×.

(2.8)

2.6. Factorization

31

Assume D = D1 D2 , where D1 and D2 are invertible operators on Y , and write A1 A+ ˙ × → M +M ˙ ×, : M +M A = 0 A2 B

=

C

=

B1 B2

C1

˙ ×, : Y → M +M C2

˙ × → Y. : M +M

Introduce the functions W1 and W2 via the biproper realizations W1 (λ)

= D1 + C1 (λIM − A1 )−1 B1 D2−1 ,

λ ∈ ρ(A1 ),

(2.9)

W2 (λ)

= D2 + D1−1 C2 (λIM × − A2 )−1 B2 ,

λ ∈ ρ(A2 ).

(2.10)

Then W admits the factorization W (λ) = W1 (λ)W2 (λ),

λ ∈ ρ(A1 ) ∩ ρ(A2 ) ⊂ ρ(A).

The function W is deﬁned and analytic on ρ(A), while the factors W1 and W2 are deﬁned and analytic on the sets ρ(A1 ) and ρ(A2 ), respectively. In particular, the factors may be deﬁned and analytic on domains where the left-hand side is not. This will turn out to be relevant in applications (cf., the remarks made at the end of this section). ˙ × in the usual manner, the product of the realProof. Identifying X and M +M izations (2.9) and (2.10) is precisely the realization (2.7). The desired result now follows from Theorem 2.5. We shall refer to (2.8) as the matching condition, and when this condition is satisﬁed we refer to M, M × as a pair of matching subspaces. A pair of matching subspaces M, M × satisfying A[M ] ⊂ M,

A× [M × ] ⊂ M ×

will be called a supporting pair of subspaces for the realization (2.7). Matching pairs of subspaces correspond to projections. So Theorem 2.6 has a counterpart in terms of projections. We say that a projection Π : X → X is a supporting projection for the realization (2.7) if A[Ker Π] ⊂ Ker Π,

A× [Im Π] ⊂ Im Π.

Here Ker T stands for the null space of an operator or matrix T , and Im T for its range.

32

Chapter 2. The state space method and factorization

Theorem 2.7. Let Π be a supporting projection for the biproper realization W (λ) = D + C(λIX − A)−1 B,

λ ∈ ρ(A).

Assume D = D1 D2 , where D1 and D2 are invertible operators on Y , and introduce the functions W1 and W2 via the biproper realizations W1 (λ)

= D1 + C(λIX − A)−1 (IX − Π)BD2−1 ,

λ ∈ ρ(A),

W2 (λ)

= D2 + D1−1 CΠ(λIX − A)−1 B,

λ ∈ ρ(A).

Then W (λ) = W1 (λ)W2 (λ) for all λ ∈ ρ(A). This factorization holds on the resolvent set ρ(A) of A. However, in many cases (relevant for applications), the factors in the right-hand side have an analytic extension to larger domains (see Theorem 2.6; cf., also the remarks made at the end of this section). Proof. The fact that Π is a supporting projection for the given biproper realization means nothing else than that the identities ΠA = ΠAΠ and A× Π = ΠA× Π are satisﬁed. Hence (I − Π)(A − A× )Π = AΠ − ΠA. Now take λ ∈ ρ(A). Then W1 (λ)W2 (λ)

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 (I − Π)BD−1 CΠ(λ − A)−1 B

as desired.

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 (I − Π)(A − A× )Π(λ − A)−1 B

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 (AΠ − ΠA)(λ − A)−1 B

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 Π(λ − A) − (λ − A)Π (λ − A)−1 B

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 ΠB − CΠ(λ − A)−1 B,

=

D + C(λ − A)−1 B = W (λ),

The material presented above contains two factorization results: Theorems 2.6 and 2.7. These theorems contain not only diﬀerent expressions for the factors, these factors also have diﬀerent domains. For rational matrix functions and matrix-valued realizations, the diﬀerences are not essential. In the case of an inﬁnite dimensional state space one has to be more careful, the reason being that

2.6. Factorization

33

ρ(A1 ) ∩ ρ(A2 ) can then be a proper subset of ρ(A). For an exhaustive discussion of the issues involved, see Section 2.5 in [20]. We shall meet the diﬀerences referred to above when the factorization results are applied, as will be done later on, for solving Wiener-Hopf, Toeplitz or singular integral equations. In that context, it is also necessary to have information on the sets where the factors take invertible values and to have expressions for the inverses. In other words, it is necessary to have a good understanding of the relationship between Theorems 2.6 and 2.7 on the one hand, and the inversion result Theorem 2.4 on the other. The point here is that, by taking inverses, the factorizations of the function W (λ) given in Theorems 2.6 and 2.7 directly induce factorizations of the point-wise inverse W −1 of W , that is the function given by W −1 (λ) = W (λ)−1 , while on the other hand factorizations of W −1 can also be obtained by applying Theorems 2.6 and 2.7 to the realization W −1 (λ) = D−1 − D−1 C(λ − A× )−1 BD−1 .

(2.11)

Note here that if M, M × is a supporting pair of subspaces for the realization (2.7), then M × , M is a supporting pair of subspaces for the realization (2.11), and, analogously, if Π is a supporting projection for (2.7), then I − Π is a supporting projection for (2.11). The analysis in [20], Section 2.5 also clariﬁes these matters; the upshot is that the two approaches lead to essentially the same result.

Notes The notion of a realization originates from the Kalman theory of linear timeinvariant systems [95]. The literature on the subject is rich; see, e.g., the text books [94], [33]. In a somewhat diﬀerent form the notion of realization also appears in the theory of characteristic operator functions [27], [141]. The realization problem has many diﬀerent faces, depending on the class of matrix or operator functions one is dealing with. The material of the ﬁrst two sections is standard. Theorem 2.1 is a variation on Theorem 4.20 in [10]. Other constructions of matrix-valued realizations, including realizations with smallest possible state space dimension, can be found in text books; see, e.g., [94], [33] or [85] and references given there. The realization theorems for analytic operator functions in Section 2.3 originate from [57]. The operations of inversion and taking products are standard in systems theory. Theorem 2.11 has a natural Schur complement interpretation; see Section 2.2 in [20] and the paper [19]. The factorization theorem in the ﬁnal section originates from [21]; see also the ﬁrst chapter of [11]. For a brief description of the history of the factorization principle presented here, we refer to the Editorial introduction in [54].

Part II Convolution equations with rational matrix symbols The canonical factorization theorem for rational matrix functions in state space form is the ﬁrst result presented and proved in this part. This theorem is then used to invert explicitly Wiener-Hopf, Toeplitz and singular integral operators with a rational matrix symbol, with the inverses being presented explicitly in state space formulas. For rational matrix symbols the solution to the homogeneous RiemannHilbert boundary value problem is also given in state space form. This part consists of two chapters. In the ﬁrst chapter (Chapter 3) we consider proper rational matrix functions, that is, rational matrix functions that are analytic at inﬁnity. The case of non-proper rational symbols is treated in the second chapter (Chapter 4). This requires a diﬀerent type of realization. This modiﬁed realization result is developed and a corresponding canonical factorization theorem is proved. As an application the homogeneous Riemann-Hilbert boundary value problem is solved for an arbitrary rational matrix symbol.

Chapter 3

Explicit solutions using realizations As we have seen in Chapter 1, canonical factorization serves as a tool to solve Wiener-Hopf integral equations, their discrete analogues, and the more general singular integral equations. In this chapter the state space factorization method developed in Chapter 2 is used to solve the problem of canonical factorization (necessary and suﬃcient conditions for its existence) and to derive explicit formulas for its factors. This is done in Section 3.1 for rational matrix functions and later in Section 7.1 for operator-valued transfer functions that are analytic on an open neighborhood of a curve. The results are applied to invert Wiener-Hopf integral equations with a rational matrix symbol (Section 3.2), block Toeplitz operators (Section 3.3) and singular integral equations (Section 3.4). The methods developed in this chapter also allow us to deal with the Riemann-Hilbert boundary value problem. This is done in the ﬁnal section which also contains material on the homogeneous Wiener-Hopf equation.

3.1 Canonical factorization of rational matrix functions in state space form In this section and the next one we shall consider the factorization theorems of Section 2.6 for the special case when the two factors satisfy additional spectral conditions. Recall from Chapter 0 that a Cauchy contour is the positively oriented boundary of a bounded Cauchy domain in C and that such a contour consists of a ﬁnite number of simple closed non-intersecting rectiﬁable curves. We say that a Cauchy contour Γ splits the spectrum σ(S) of a bounded linear operator S if Γ ∩ σ(S) = ∅. In that case σ(S) decomposes into two disjoint compact sets σ+ and σ− such that σ+ is in the interior domain of Γ and σ− is in the exterior domain

38

Chapter 3. Explicit solutions using realizations

of Γ. If Γ splits the spectrum of S, then we have a Riesz projection, also called spectral projection, associated with S and Γ, namely 1 P (S; Γ) = (λ − S)−1 dλ. 2πi Γ The subspace N = Im P (S; Γ) will be called the spectral subspace for S corresponding to the contour Γ (or to the spectral set σ+ ). Lemma 3.1. Let Y1 and Y2 be complex Banach spaces, and consider the operator S11 S12 ˙ Y2 → Y1 + ˙ Y2 . : Y1 + S= (3.1) 0 S22 ˙ Y2 such that Ker Π = Y1 . Then the compression Let Π be any projection of Y = Y1 + ΠS|Im Π : Im Π → Im Π and S22 : Y2 → Y2 are similar. Furthermore, Y1 is a spectral subspace for S if and only if σ(S11 ) ∩ σ(S22 ) = ∅, and in that case σ(S) = σ(S11 ) ∪ σ(S22 ) while, in addition, 1 −1 Y1 = Im P (S; Γ) = Im (λI − S) dλ , (3.2) 2πi Γ where Γ is a Cauchy contour around σ(S11 ) separating σ(S11 ) from σ(S22 ). ˙ Y2 along Y1 onto Y2 . As Ker P = Ker Π, Proof. Let P be the projection of Y = Y1 + we have P = P Π and the map E = P |Im Π : Im Π → Y2 is an invertible operator. Write S0 for the compression ΠS|Im Π : Im Π → Im Π of S to Im Π, and take x = Πy. Then ES0 x = P ΠSΠy = P SΠy = P SP Πy = S22 Ex, and hence S0 and S22 are similar. Now suppose σ(S11 ) ∩ σ(S22 ) = ∅. Then ρ(S11 ) ∪ ρ(S22 ) = C and hence ρ(S) ∩ ρ(S22 ) . ρ(S) = ρ(S) ∩ ρ(S11 ) The upper triangular form of S in (3.1) ensues ρ(S) ∩ ρ(S11 ) = ρ(S) ∩ ρ(S22 ) = ρ(S11 ) ∩ ρ(S22 ) and it follows that ρ(S11 ) ∪ ρ(S22 ) = ρ(S), an identity which can be rewritten as σ(S) = σ(S11 ) ∪ σ(S22 ). Still under the assumption that σ(S11 ) ∩ σ(S22 ) = ∅, let Γ be a Cauchy contour Γ around σ(S11 ) separating σ(S11 ) from σ(S22 ). Then Γ splits the spectrum of S. In fact, if λ ∈ Γ, then both λ − S11 and λ − S22 are invertible and ⎡ ⎤ (λ − S11 )−1 (λ − S11 )−1 S12 (λ − S22 )−1 ⎦ (λ − S)−1 = ⎣ −1 0 (λ − S22 )

3.1. Canonical factorization of rational matrix functions in state space form 39 which leads to an expression of the type P (S; Γ) =

I

∗

0

0

for the Riesz projection associated with S and Γ. In particular, it is clear that Y1 = Im P (S; Γ). So Y1 is a spectral subspace for S and (3.2) holds. Next assume that Y1 = Im Q, where Q is a Riesz projection for S. Put Π = I − Q, and let S0 be the restriction of S to Im Π. Then σ(S11 )∩ σ(S0 ) = ∅. By the ﬁrst part of the proof, the operators S0 and S22 are similar. So σ(S0 ) = σ(S22 ), and hence we have shown that σ(S11 ) ∩ σ(S22 ) = ∅. Let Γ be a Cauchy contour. As before (see the one but last paragraph in Chapter 0) we denote by F+ and F− the interior and exterior domain of Γ, respectively. Note that ∞ ∈ F− . Let W be a rational m × m matrix function, with W (∞) = I, analytic on an open neighborhood of Γ, whose values on Γ are invertible matrices. By a right canonical factorization of W with respect to Γ we mean a factorization λ ∈ Γ, (3.3) W (λ) = W− (λ)W+ (λ), where W− and W+ are rational m × m matrix functions, analytic and taking invertible values on (an open neighborhood of) F − and F + , respectively. If in (3.3) the factors W− and W+ are interchanged, we speak of a left canonical factorization. Theorem 3.2. Let Γ be a Cauchy contour and let W be a rational m × m matrix function, Suppose W admits the realization W (λ) = Im + C(λIn − A)−1 B such that the main matrix A has no eigenvalues on Γ. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (i) A× = A − BC has no eigenvalues on Γ, ˙ Ker P (A× ; Γ). (ii) Cn = Im P (A; Γ) + In that case, a right canonical factorization of W is given by W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

where the factors and their inverses can be written as W− (λ)

=

Im + C(λIn − A)−1 (I − Π)B,

W+ (λ)

=

Im + CΠ(λIn − A)−1 B,

W−−1 (λ)

=

Im − C(I − Π)(λIn − A× )−1 B,

W+−1 (λ)

=

Im − C(λIn − A× )−1 ΠB.

Here Π is the projection of Cn along Im P (A; Γ) onto Ker P (A× ; Γ).

40

Chapter 3. Explicit solutions using realizations

For left canonical factorizations an analogous theorem holds. In the result in ˙ Im P (A× ; Γ). question, (ii) is replaced by Cn = Ker P (A; Γ) + The expressions for the functions W− and W+ suggest that these functions are deﬁned on the resolvent set ρ(A) of A. Similarly, W−−1 and W+−1 seem to have ρ(A× ) as their domain. At ﬁrst sight this is at variance with the requirements for Wiener-Hopf factorization. We will address this point in the proof. Proof. From the deﬁnition given above it is clear that a necessary condition in order that W admits a right canonical factorization with respect to Γ is that W takes invertible values on Γ. By Theorem 2.4 this necessary condition is met if and only if (i) holds true. Assume that (i) is satisﬁed. The spectral projections P (A; Γ) and P (A× ; Γ) are then well-deﬁned. The image X− = Im P (A; Γ) of P (A; Γ) and the null space X+ = Ker P (A× ; Γ) of P (A× ; Γ) are invariant for A and A× , respectively. Suppose (ii) is fulﬁlled too, and write A− A0 B− A= , B= , C = C− C+ 0 A+ B+ for the matrix presentations of A, B and C with respect to the decomposition ˙ X+ . With Cn = X− + W− (λ) W+ (λ)

= IX− + C− (λ − A− )−1 B− , −1

= IX+ + C+ (λ − A+ )

B+ ,

λ ∈ ρ(A− ), λ ∈ ρ(A+ ),

we have (from Theorem 2.6) the factorization W (λ) = W− (λ)W+ (λ),

λ ∈ ρ(A− ) ∩ ρ(A+ ) ⊂ ρ(A).

As X− is a spectral subspace for A, we can apply Lemma 3.1 to show that σ(A− ) and σ(A+ ) are disjoint. But then ρ(A) = ρ(A− ) ∩ ρ(A+ ) and it follows that W (λ) = W− (λ)W+ (λ),

λ ∈ ρ(A− ) ∩ ρ(A+ ) = ρ(A).

(3.4)

Applying Lemma 3.1 once again we see that σ(A− ) = σ(A) ∩ F+ ,

σ(A+ ) = σ(A) ∩ F− ,

(3.5)

where F+ and F− are the interior and exterior domain of Γ, respectively. In a similar way one proves that × σ(A× − ) = σ(A ) ∩ F+ ,

× σ(A× + ) = σ(A ) ∩ F− .

(3.6)

Using the ﬁrst parts of (3.5) and (3.6), it now follows that W− is analytic and has invertible values on an open neighborhood of F − . Analogously, employing the second parts of (3.5) and (3.6), one gets that W+ is analytic and has invertible

3.1. Canonical factorization of rational matrix functions in state space form 41 values on an open neighborhood of F + . Thus (3.4) is a right canonical Wiener-Hopf factorization with respect to Γ. The projection Π of Cn along Im P (A; Γ) onto Ker P (A× ; Γ) is a supporting projection for the given realization of W . Also In −Π is a supporting projection for the realization W (λ)−1 = Im − C(λIn − A× )−1 B of W −1 . With this in mind, one checks without diﬃculty that W− , W+ , W−−1 and W+−1 can also be written as in the theorem. For an exhaustive discussion of the intricacies concerning inversion, factorization, and the combination of these operations (in fact: the relationship between Theorems 2.6, 2.7 and 2.4), see Section 2.5 in [20]. Note, however, that in the present case there is no ambiguity because we are working here with rational matrix functions. Next, suppose that W (λ) = W− (λ)W+ (λ) is a right canonical factorization ˙ Ker P (A× ; Γ). with respect to Γ. We only have to show that Cn = Im P (A; Γ) + × We ﬁrst prove that Im P (A; Γ) ∩ Ker P (A ; Γ) = {0}. Without loss of generality it may be assumed that the values of W− and W+ at inﬁnity are equal to Im . Suppose x ∈ Im P (A; Γ) ∩ Ker P (A× ; Γ), and consider (λ − A)−1 x. This function is analytic on an open neighborhood of F − . On the other hand the function (λ − A× )−1 x is analytic on an open neighborhood of F + . For λ in the intersection ρ(A) ∩ ρ(A× ), we have W (λ)C(λ − A× )−1

= C(λ − A× )−1 + C(λ − A)−1 BC(λ − A× )−1 = C(λ − A× )−1 + C(λ − A)−1 (A − A× )(λ − A× )−1 = C(λ − A)−1 ,

and it follows that W+ (λ)C(λ − A× )−1 = W− (λ)−1 C(λ − A)−1 . The analyticity properties of the factors W− , W+ and their inverses now imply that the function W+ (λ)C(λ − A× )−1 x = W (λ)−1 C(λ − A)−1 x is analytic on the Riemann sphere C∞ . By Liouville’s theorem it must be constant. As it takes the value zero at inﬁnity, it is identically zero. Hence both C(λ − A× )−1 x and C(λ − A)−1 x vanish. Next use the identity (λ − A× )−1 BC(λ − A)−1 = (λ − A)−1 − (λ − A× )−1 to obtain (λ − A× )−1 x = (λ − A)−1 x. But then this function is analytic on the Riemann sphere too. Using Liouville’s theorem again, we see that it must be identically zero. Thus x = 0. Observe that up to this point in the proof we have not used the ﬁnite dimensionality of the state space. It will play a role in the next paragraph. We now ﬁnish the proof by a duality argument. Let Γ∗ be the adjoint curve of Γ, i.e., the curve obtained from Γ by complex conjugation. Also introduce ¯ ∗ , V− (λ) = W− (λ) ¯ ∗ and the functions V, V+ and V− by putting V (λ) = W (λ) ¯ ∗ . Clearly V has the realization V (λ) = I + B ∗ (λ − A∗ )−1 C ∗ V+ (λ) = W+ (λ) and V (λ) = V+ (λ)V− (λ) is a left canonical factorization. Arguing as above, we may conclude that Ker P (A∗ , Γ∗ ) ∩ Im P (A× )∗ , Γ∗ = 0. It follows that

42

Chapter 3. Explicit solutions using realizations

Ker P (A∗ , Γ∗ ) + Im P (A× )∗ , Γ∗ = Cn . In the ﬁrst instance, this equality holds for the closure of Ker P (A∗ , Γ∗ ) + Im P (A× )∗ , Γ∗ , but in Cn all linear manifolds are closed. With minor modiﬁcations we could have worked in Theorem 3.2 with two curves, one splitting the spectrum of A and the other splitting the spectrum of A× (cf., [100]). Finally, let us mention that Theorem 3.2 remains true if the Cauchy contour Γ is replaced by the extended real line R∞ , i.e., the closure of the real line in the Riemann sphere C∞ . In that case F+ is the open upper half plane and F− is the open lower half plane. For details, see Theorem 4.5 at the end of Section 4.3 below which, by the way, deals with the situation where W is a not necessarily proper rational matrix function.

3.2 Wiener-Hopf integral operators In this section the general factorization result proved in the preceding sections is used to provide explicit formulas for solutions of ﬁnite systems of the Wiener-Hopf equation φ(t) −

∞

0

k(t − s)φ(s) ds = f (t),

t ≥ 0,

(3.7)

(−∞, ∞), i.e., where φ and f are m-dimensional vector functions and k ∈ Lm×m 1 the kernel function k is an m × m matrix function of which the entries are in L1 (−∞, ∞). We assume that the given vector function f has its component functions in Lp [0, ∞), and we express this property by writing f ∈ Lm p [0, ∞). Throughout this section, p will be ﬁxed and 1 ≤ p < ∞. The problem we shall consider is to ﬁnd a solution φ for equation (3.7) that also belongs to the space Lm p [0, ∞). As was explained in Section 1.1 the equation (3.7) has a unique solution in Lm p [0, ∞) for each f in Lm [0, ∞) if and only if its symbol I − K(λ) admits a factorization m p as in (1.5). Our aim is to apply the factorization theory developed in the previous sections to get the canonical factorization (1.5). Therefore, in the sequel we assume that the symbol is a rational m × m matrix function. As K(λ) is the Fourier (−∞, ∞)–function, the symbol is continuous on the real transform of an Lm×m 1 line. In particular, Im − K(λ) has no poles on the real line. Furthermore, by the Riemann-Lebesgue lemma, lim

λ∈R, |λ|→∞

K(λ) = 0,

which implies that the symbol Im − K(λ) has the value In at ∞. The fact that Im − K(λ) is rational is equivalent to the requirement that the kernel function k is in the linear space spanned by all functions of the form p(t)eiαt , t > 0, h(t) = t < 0, q(t)eiβt ,

3.2. Wiener-Hopf integral operators

43

where p(t) and q(t) are matrix polynomials in t with coeﬃcients in Cm×m , and α and β are complex numbers with α > 0 and β < 0. From Section 2.2 we know that the matrix function Im − K(λ) admits a realization Im − K(λ) = Im + C(λIn − A)−1 B such that the main matrix A has no real eigenvalues. In the next theorem we express the solvability of equation (3.7) in terms of such a realization and give explicit formulas for its solutions in the same terms. Theorem 3.3. Let Im − K(λ) = Im + C(λIn − A)−1 B be a realization for the symbol of equation (3.7), and suppose A has no real eigenvalues. In order that m (3.7) has a unique solution φ in Lm p [0, ∞) for each f in Lp [0, ∞), the following two conditions are necessary and suﬃcient: (i) A× = A − BC has no real eigenvalues; ˙ M × , where M is the spectral subspace of A corresponding to the (ii) Cn = M + eigenvalues of A in the upper half plane, and M × is the spectral subspace of A× corresponding to the eigenvalues of A× in the lower half plane. Assume conditions (i) and (ii) hold true, and let Π be the projection of Cn along M onto M × . Then Im − K(λ) admits a right canonical factorization with respect to the real line that has the form Im − K(λ) = Im + G− (λ) Im + G+ (λ) , λ ∈ R, where the factors and their inverses can be written as Im + G+ (λ)

=

Im + CΠ(λIn − A)−1 B,

Im + G− (λ) −1 Im + G+ (λ) −1 Im + G− (λ)

=

Im + C(λIn − A)−1 (In − Π)B,

=

Im − C(λIn − A× )−1 ΠB,

=

Im − C(In − Π)(λIn − A× )−1 B.

The functions γ+ and γ− in (1.6) and (1.7) are given by γ+ (t) = γ− (t) =

×

+iCe−itA ΠB,

t > 0, ×

−iC(In − Π)e−itA t B,

t < 0.

Finally, the solution φ to (3.7) can be written as ∞ φ(t) = f (t) + γ(t, s)f (s) ds, 0

⎧ × × ⎨ +iCe−itA ΠeisA B,

where γ(t, s) =

⎩

×

s < t, ×

−iCe−itA (In − Π)eisA B,

s > t.

44

Chapter 3. Explicit solutions using realizations

Proof. We have already mentioned that equation (3.7) has a unique solution in m Lm p [0, ∞) for each f in Lp [0, ∞) if and only if the symbol Im − K(λ) admits a right canonical factorization as in (1.5). So to prove the necessity and suﬃciency of the conditions (i) and (ii), it suﬃces to show that the conditions (i) and (ii) together are equivalent to the statement that Im − K(λ) admits a right canonical factorization as in (1.5). We ﬁrst observe that condition (i) is equivalent to the requirement that Im − K(λ) is invertible for all λ ∈ R (see Theorem 2.4). But then we can apply Theorem 3.2 in combination with the remark made at the end of Section 3.1 to prove the ﬁrst part of the theorem. Next assume that conditions (i) and (ii) hold true. Applying Theorem 3.2 once again, we get the desired formulas for Im + G+ (λ), Im + G− (λ) and their inverses. The formulas for γ+ and γ− are now obtained by noticing that

∞

×

eiλt e−itA Π dt

=

i(λ − A× )−1 Π,

=

−i(I − Π)(λ − A× )−1 ,

0 0

−∞

×

eiλt (I − Π)e−itA dt

λ ∈ ρ(A× ), λ ≥ 0, λ ∈ ρ(A× ), λ ≤ 0,

where I = In . The proof of the latter identity uses (the ﬁrst conclusion in) Lemma 3.1. It remains to prove the ﬁnal formula for γ(t, s). We use (1.10), and compute ﬁrst that ×

×

γ+ (t − r)γ− (r − s) = Ce−i(t−r)A ΠBC(I − Π)e−i(r−s)A B. Now Ker Π = M is A-invariant and Im Π = M × is A× -invariant. Thus ΠA(I−Π) = 0 and (I − Π)A× Π = 0, and it follows that ΠBC(I − Π) = Π(A − A× )(I − Π) = ΠA× − A× Π. But then γ+ (t − r)γ− (r − s) = =

×

×

Ce−i(t−r)A (A× Π − ΠA× )e−i(r−s)A B −i

× × d Ce−i(t−r)A Πe−i(r−s)A B. dr

Inserting this in (1.30) we obtain for s < t that γ(t, s) =

×

iCe−i(t−s)A ΠB − −i(t−s)A×

s

i 0

× × d Ce−i(t−r)A Πe−i(r−s)A B dr dr ×

×

ΠB − Ce−i(t−r)A Πe−i(r−s)A B|sr=0

=

iCe

=

iCe−itA ΠeisA B,

×

×

3.2. Wiener-Hopf integral operators

45

while for s > t we get ×

γ(t, s) = −iC(I − Π)e−i(t−s)A B + = −iC(I − Π)e

−i(t−s)A×

t

i 0

× × d Ce−i(t−r)A Πe−i(r−s)A B dr dr ×

×

B − Ce−i(t−r)A Πe−i(r−s)A B|tr=0

×

×

= −iCe−itA (I − Π)eisA B.

This completes the proof.

Corollary 3.4. Let Im − K(λ) = Im + C(λIn − A)−1 B be a realization for the symbol of equation (3.7). Assume that A and A× = A − BC have no spectrum on the real line, and that ˙ Cn = Im P +Ker P ×, (3.8) where P and P × are the Riesz projections of A and A× , respectively, corresponding to the spectra in the upper half plane. Fix x ∈ Ker P , and let the right-hand side of (3.7) be given by f (t) = Ce−itA x, t ≥ 0. Then the unique solution φ in Lm p [0, ∞) of equation (3.7) is given by ×

φ(t) = Ce−itA Πx,

t ≥ 0.

Here Π is the projection of Cn onto Ker P × along Im P . Proof. Since x ∈ Ker P , the vector e−itA x is exponentially decaying in norm when t → ∞, and thus the function f belongs to Lm p [0, ∞). Furthermore, the conditions (i) and (ii) in Theorem 3.3 are fulﬁlled, and hence equation (3.7) has a unique solution φ ∈ Lm p [0, ∞). Moreover from Theorem 3.3 we know that φ is given by φ(t) = f (t) + iCe−itA

×

t

0

−iCe

×

ΠeisA BCe−isA x ds

−itA×

t

∞

× (I − Π)eisA BCe−isA x ds .

Now use that ×

×

eisA BCe−isA = ieisA (iA× − iA)e−isA = i

d isA× −isA e e . ds

It follows that × × φ(t) = f (t) − Ce−itA ΠeisA e−isA x|t0 × × +Ce−itA (I − Π)eisA e−isA x|∞ . t

46

Chapter 3. Explicit solutions using realizations ×

×

Since (I − Π) = (I − Π)P × , the function (I − Π)eisA = (I − Π)P × eisA is exponentially decaying for s → ∞. As we have seen, the same holds true for e−isA x. Thus ×

×

×

φ(t) = f (t) − Ce−itA ΠeitA e−itA x + Ce−itA Πx ×

×

−Ce−itA (I − Π)eitA e−itA x ×

= f (t) + Ce−itA Πx − Ce−itA x ×

= Ce−itA Πx,

which completes the proof.

Finally, let us return to the special situation where the functionf is given by formula (1.11), and assume that the conditions (i) and (ii) in Theorem 3.3 are satisﬁed. Then the solution φ admits the representation t × φ(t) = e−iqt {Im + i Cei(q−A )s ΠB ds} (3.9) 0

·{Im − C(I − Π)(q − A× )−1 B}x0 ;

(3.10)

see formula (1.12).

3.3 Block Toeplitz operators In the previous section the factorization theory was applied to ﬁnite systems of Wiener-Hopf integral equations. In this section we carry out a similar program for their discrete analogues, block Toeplitz equations (cf., Section 1.2). So we consider an equation of the type ∞

aj−k ξk = ηj ,

j = 0, 1, 2, . . . .

(3.11)

k=0

Throughout we assume that the coeﬃcients aj are given complex m × m matrices satisfying ∞

aj < ∞, j=−∞

and η = (ηj )∞ j=0 m ξ = (ξk )∞ k=0 ∈ p

m is a given vector from m p = p (C ). The problem is to ﬁnd such that (3.11) is satisﬁed. As before, we shall apply our factorization theory. For that reason we assume j that the symbol a(λ) = ∞ j=−∞ λ aj is a rational m × m matrix function whose value at ∞ is Im . Note that a(λ) has no poles on the unit circle. Therefore the conditions on a(λ) are equivalent to the following assumptions:

3.3. Block Toeplitz operators

47

(j) the sequence (aj − δj0 Im )∞ j=0 is a linear combination of sequences of the form j r ∞ α j D j=0 , where |α| < 1, r is a nonnegative integer and D is a complex m × m matrix; (jj) the sequence (a−j )∞ j=1 is a linear combination of sequences of the form ∞ −j s ∞ δjk F j=1 , β j E j=1 , where |β| > 1, s and k are nonnegative integers and E and F are complex m × m matrices. From Section 2.2 we know that the matrix function a(λ) admits a realization a(λ) = Im + C(λIn − A)−1 B

(3.12)

such that the main matrix A has no eigenvalues on the unit circle. The next theorem is the analogue of Theorem 3.3. Theorem 3.5. Let (3.12) be a realization for the symbol a(λ) of the equation (3.11), and suppose A has no eigenvalues on the unit circle. Then (3.11) has a unique m ∞ m solution ξ = (ξk )∞ k=0 in p for each η = (ηj )j=0 in p if and only if the following two conditions are satisﬁed: (i) A× = A − BC has no eigenvalues on the unit circle, ˙ M × , where M is the spectral subspace of A corresponding to the (ii) Cn = M + eigenvalues of A inside the unit circle, and M × is the spectral subspace of A× corresponding to the eigenvalues of A× outside the unit circle. Assume conditions (i) and (ii) are satisﬁed, and let Π be the projection of Cn along M onto M × . Then the function a(λ) admits a right canonical factorization with respect to the unit circle that has the form a(λ) = h− (λ)h+ (λ),

|λ| = 1,

where the factors and their inverses can be written as h+ (λ)

= Im + CΠ(λIn − A)−1 B,

h− (λ)

= Im + C(λIn − A)−1 (In − Π)B,

h−1 + (λ)

= Im − C(λIn − A× )−1 ΠB,

h−1 − (λ)

= Im − C(In − Π)(λIn − A× )−1 B.

− ∞ The sequences (γj+ )∞ j=0 and (γ−j )j=0 in (1.19) are given by

γ0+

=

Im + C(A× )−1 ΠB,

γj+

=

C(A× )−(j+1) ΠB,

γ0−

=

Im ,

γj−

=

−C(In − Π)(A× )−(j+1) B,

j = 1, 2, . . . ,

j = −1, −2, . . . .

48

Chapter 3. Explicit solutions using realizations

Finally, the solution ξ to (3.11) can be written as ξk =

γks

∞

s=0

⎧ C(A× )−(k+1) Π(A× )s B, ⎪ ⎪ ⎪ ⎨ = Im + C(A× )−(s+1) Π(A× )s B, ⎪ ⎪ ⎪ ⎩ −C(A× )−(k+1) (In − Π)(A× )s B,

γks ηs where

s < k, s = k, s > k.

Proof. The proof of Theorem 3.5 is similar to that of Theorem 3.3. Here we only derive the ﬁnal formula for γks . With respect to the formulas for γj+ , we note that Im Π is A× -invariant and the restriction of A× to Im Π is invertible. So, with slight abuse of notation as far as inverses of A× are involved, h+ (λ)−1

=

Im − C(λ − A× )−1 ΠB −1 × −1 Im + C I − λ(A× )−1 (A ) ΠB

=

Im +

=

∞

λj C(A× )−(j+1) ΠB.

j=0

Now compare coeﬃcients with h+ (λ)−1 = γj− are obtained by comparing h− (λ)−1

∞

j=0

λj γj+ . Similarly, the formulas for

=

Im − C(I − Π)(λ − A× )−1 B

=

Im − C(I − Π)

=

Im −

−1

∞

1 × j−1 (A ) B λj j=1

λj C(I − Π)(A× )−(j+1) B

j=−∞

0 with h− (λ)−1 = j=−∞ λj γj− . Here I = In . To obtain the formulas for γks we employ (1.22). For s < k we must ﬁnd + γks = γk−s γ0− +

s−1

r=0

+ − γk−r γr−s ,

while for s > k we need to calculate γks =

− γ0+γk−s

+

k−1

r=0

+ − γk−r γr−s .

3.3. Block Toeplitz operators

49

Again by slight abuse of notation + − γk−r γr−s

=

−C(A× )−(k−r+1) ΠBC(I − Π)(A× )−(r−s+1) B

=

−C(A× )−(k−r+1) (A× Π − ΠA× )(A× )−(r−s+1) B

=

−C(A× )−(k−r) Π(A× )−(r−s+1) B + +C(A× )−(k−r+1) Π(A× )−(r−s) B.

Observe that if we replace r by r + 1 in the last one of the latter two terms we get the ﬁrst one. So the summation in the formula for γks is telescoping and collapses into just a few terms. We proceed as follows. For s < k we get + γks = γk−s γ0− − C(A× )−(k−s+1) ΠB + C(A× )−(k+1) Π(A× )s B. + Since γ0− = I and γk−s = C(A× )−(k−s+1) ΠB, this results in

γks = C(A× )−(k+1) Π(A× )s B. For s > k the computation is a little more involved as γ0+ = In + C(A× )−1 ΠB. Using that ΠBC I − Π) = A× Π − ΠA× , it goes this way: γks = − I + C(A× )−1 ΠB C I − Π)(A× )−(k−s+1) B +C(A× )−(k+1) Π(A× )s B − C(A× )−1 Π(A× )−(k−s) B =

−C I − Π)(A× )−(k−s+1) B +C(A× )−1 (ΠA× − A× Π)(A× )−(k−s+1) B +C(A× )−(k+1) Π(A× )s B − C(A× )−1 Π(A× )−(k−s) B

=

C(A× )−(k+1) Π(A× )s B − C(A× )−(k−s+1) B

=

−C(A× )−(k+1) (I − Π)(A× )s B.

It remains to consider the case k = s. Then we have γss = γ0+ γ0− +

k−1

+ − γs−r γr−s .

r=0

Following the line of argument as in the case s < k this yields γss

=

Im + C(A× )−1 ΠB − C(A× )−1 ΠB + C(A× )−(k+1) Π(A× )k B

=

Im + C(A× )−(k+1) Π(A× )k B,

which completes the proof.

50

Chapter 3. Explicit solutions using realizations

The main step in the factorization method for solving the equation (3.11) is to construct a right canonical factorization of the symbol a(λ) with respect to the unit circle. In Theorem 3.5 we obtained explicit formulas for the case when a(λ) is rational and has the value In at ∞. The latter condition is not essential. Indeed, by a suitable M¨ obius transformation one can transform the symbol α(λ) into a function which is invertible at inﬁnity (see Section 3.6). Next one makes the Wiener-Hopf factorization of the transformed symbol with respect to the image of the unit circle under the M¨ obius transformation. Here one can use the same formulas as in Theorem 3.5. Finally, using the inverse M¨obius transformation, one can obtain explicit formulas for the factorization with respect to the unit circle, and hence also for the solution of equation (3.11).

3.4 Singular integral equations In this section we apply Theorem 3.2 to solve the singular integral equation from Section 1.3: 1 φ(τ ) a(t)φ(t) + b(t) dτ = f (t), t ∈ Γ, (3.13) πi Γ τ − t where Γ is a Cauchy contour. The problem is to ﬁnd φ ∈ Lm p (Γ) such that (3.13) is satisﬁed. Recall that (3.13) can be rewritten in the form aIφ + bSφ = f , where S is the singular integral operator as in (1.26). Put c = a + b and d = a − b. Then we know from Section 1.3 that the operator aI + bS is invertible if and only if c(λ) and d(λ) are invertible for all λ ∈ Γ and the function w(λ) = d(λ)−1 c(λ) admits a right canonical factorization with respect to Γ. The next theorem deals with the case when w(λ) is rational and has the value Im at ∞. Theorem 3.6. Suppose det a(λ) + b(λ) and det a(λ) − b(λ) do not vanish on Γ, −1 a(λ) + b(λ) is a rational function which has and assume w(λ) = a(λ) − b(λ) the value Im at inﬁnity. Let w(λ) = Im + C(λIn − A)−1 B be a realization for w. Suppose A and A× = A − BC have no spectrum on Γ. ˙ × , where M is the spectral Then aI + bS is invertible if and only if Cn = M +M subspace corresponding to the eigenvalues of A inside Γ, and M × is the spectral subspace corresponding to the eigenvalues of A× outside Γ. In that case the func−1 −1 tions w+ , w+ , w− and w− appearing in the expressions for (aI + bS)−1 given in Section 1.3 can be speciﬁed as follows: w+ (λ)

= Im + CΠ(λIn − A)−1 B,

w− (λ)

= Im + C(λIn − A)−1 (In − Π)B,

−1 (λ) w+

= Im − C(λIn − A× )−1 ΠB,

−1 (λ) w−

= Im − C(In − Π)(λIn − A× )−1 B.

3.5. The Riemann-Hilbert boundary value problem

51

Here Π is the projection of Cn along M onto M × and I = In is the identity operator on Cn . By way of illustration, we consider the special case when 1 a(t) − b(t) η, f (t) = t−α where α is a complex number outside Γ and η ∈ Cm . Put g(t) =

1 η. t−α

−1 −1 −1 d = w− g. Then one can write f = dg, where as before d = a − b. Hence w− Observe now that the function 1 −1 −1 w− (t) − w− (α) η t−α

is analytic outside Γ and vanishes at ∞. So when we apply P Γ to it, we get zero. It follows that 1 −1 P Γ w− w−1 (α)η. g (t) = t−α − But then 1 −1 −1 −1 w− (t) − w− QΓ w− g (t) = (α) η, t−α and hence 1 1 −1 −1 −1 (t)w− (α)η + (α) η. w+ Im − w− (t)w− (aI + bS)−1 f (t) = t−α t−α In the situation of Theorem 3.6, the right-hand side of this equality becomes 1 1 η− C (t − A× )−1 Π + (t − A)−1 (I − Π) B t−α t−α · Im − C(I − Π)(α − A× )−1 B η. The case when w(λ) is rational, but does not have the value Im at ∞, can be treated by applying a suitable M¨ obius transformation. The argument is similar to that indicated at the end of Section 3.3.

3.5 The Riemann-Hilbert boundary value problem In this section we consider the (homogeneous) Riemann-Hilbert boundary value problem (on the real line): W (λ)Φ+ (λ) = Φ− (λ),

−∞ < λ < +∞.

(3.14)

52

Chapter 3. Explicit solutions using realizations

The precise formulation of this problem is as follows. Let W be a given m × m matrix function, with entries that are integrable on the real line. The problem is to describe all pairs Φ+ , Φ− of Cm -valued functions such that (3.14) is satisﬁed while, in addition, Φ+ and Φ− are the Fourier transforms of integrable Cm -valued functions with support in [0, ∞) and (−∞, 0], respectively. For such a pair of functions Φ+ , Φ− we have that Φ+ is continuous on the closed upper half plane, analytic in the open upper half plane and vanishes at inﬁnity, the same being true for Φ− with the understanding that the upper half plane is replaced by the lower. The functions W that we shall deal with are rational m × m matrix functions with the value Im at inﬁnity and given in the form of a realization. Theorem 3.7. Let W be a rational m × m matrix function, and suppose W admits the realization W (λ) = Im + C(λIn − A)−1 B. Suppose further that both A and A× = A − BC have no eigenvalues on the real line. Let M be the spectral subspace of A corresponding to the eigenvalues of A in the upper half plane, and let M × be the spectral subspace of A× corresponding to the eigenvalues of A× in the lower half plane. Then the pair of functions Φ+ , Φ− is a solution of the Riemann-Hilbert boundary value problem (3.14) if and only if there exists x ∈ M ∩ M × such that Φ+ (λ) = C(λIn − A× )−1 x,

Φ− (λ) = C(λIn − A)−1 x.

(3.15)

Moreover, the vector x in (3.15) is uniquely determined by the pair Φ+ , Φ− . Proof. Take x ∈ M ∩ M × and deﬁne Φ+ and Φ− by (3.15). From Theorem 2.4 we know that W (λ)C(λ−A× )−1 = C(λ−A)−1 . It follows that (3.14) is satisﬁed. Here × the speciﬁc choice of x does not even play a role. Put φ+ (t) = −iCe−itA x, t ≥ 0. Since x ∈ M × , the function φ+ is integrable on [0, ∞). Similarly, as x ∈ M , the function φ− given by φ− (t) = iCe−itA x, t ≤ 0 is integrable on (−∞, 0]. A straightforward computation shows that ∞ 0 Φ+ (λ) = eiλt φ+ (t)dt, Φ− (λ) = eiλt φ− (t)dt (3.16) 0

−∞

and the proof of the “if part” of the theorem is complete. The proof of the “only if part” is somewhat more involved. Let Φ+ , Φ− be a solution of (3.14) given in the form (3.16) with integrable φ+ and φ− . It will be convenient to extend φ+ and φ− to integrable functions on the full real line by stipulating that they vanish on [−∞, 0) and [0, ∞), respectively. For λ ∈ R put ρ(λ) = (λ − A)−1 BΦ+ (λ). Note that (λ − A)−1 appears as a Fourier transform of a matrix function with entries from L1 (R). In fact ∞ −1 eiλt (t)dt, λ ∈ R, (λ − A) = −∞

where (t) =

ie−itA P,

t < 0,

−itA

t < 0.

−ie

(In − P ),

3.5. The Riemann-Hilbert boundary value problem

53

Using inverse Fourier transforms and the fact that the support of φ+ is contained in [0, ∞), we have ∞ ∞ eiλt (t − s)Bφ+ (s) ds dt, λ ∈ R. ρ(λ) = −∞

Introduce

γ− (t) = γ+ (t) =

∞

0 ∞ 0

0

(t − s)Bφ+ (s) ds,

(t < 0),

γ− (t) = 0

(t > 0),

(t − s)Bφ+ (s) ds,

(t > 0),

γ+ (t) = 0

(t < 0),

and for each λ ∈ R set

ρ+ (λ) =

∞ −∞

ρ− (λ) =

∞ −∞

eiλt γ+ (t) dt = eiλt γ− (t) dt =

∞ 0

0

−∞

eiλt γ+ (t) dt, eiλt γ− (t) dt.

Obviously, ρ(λ) = ρ− (λ) + ρ+ (λ) for each λ ∈ R. From (3.14) and the deﬁnition of ρ it follows that Φ+ (λ) + Cρ+ (λ) = Φ− (λ) − Cρ− (λ),

λ ∈ R.

(3.17)

The left-hand side of (3.17) is continuous on the closed upper half plane, analytic in the open upper half plane and vanishes at inﬁnity. The same is true for the right-hand side of (3.17) provided the upper half plane is replaced by the lower half plane. But then we can apply Liouville’s theorem to show that both sides of (3.17) are identically zero. Hence 0 eiλt Cγ− (t) dt, λ ≤ 0, (3.18) Φ− (λ) = Cρ− (λ) = −∞

Φ+ (λ) = −Cρ+ (λ) = − For t < 0 we have γ− (t)

0

= where x =

∞

=

0

eiλt Cγ+ (t) dt,

λ ≥ 0.

(3.19)

(t − s)Bφ+ (s) ds

ie−itA

∞ 0

∞

∞

eisA P Bφ+ (s) ds = ie−itA x,

eisA P Bφ+ (s) ds. Clearly x ∈ Im P , and we conclude that 0 eiλt ie−itA x dt = (λ − A)−1 x, λ ≤ 0. ρ− (λ) = 0

−∞

(3.20)

54

Chapter 3. Explicit solutions using realizations

Next, ﬁx λ ∈ R. Since (λ − A)ρ(λ) = BΦ+ (λ) and (λ − A)ρ− (λ) = x, we can use the ﬁrst part of (3.19) to show that (λ − A)ρ+ (λ) + x = (λ − A)ρ(λ) = BΦ+ (λ) = −BCρ+ (λ). Recall that A× = A − BC. It follows that ρ+ (λ) = −(λ − A× )−1 x,

λ ∈ R.

(3.21)

The left-hand side of (3.21) is continuous on the closed upper half plane and analytic in the open upper half plane. Thus (3.21) implies that P × x = 0, where P × is the spectral projection of A× corresponding to the eigenvalues in the upper half plane. Since Im P = M and Ker P × = M × , we see that x ∈ M ∩ M × . From (3.19) and (3.21) it follows that the ﬁrst identity in (3.15) holds. Similarly, (3.18) and (3.20) yield the second identity in (3.15). It remains to prove the unicity of x. Take u ∈ M ∩ M × , and assume that C(λ−A)−1 u = 0. It suﬃces to show that u = 0. To do this, recall (see Theorem 2.4) that (λ − A× )−1 = (λ − A)−1 − (λ − A)−1 BW (λ)−1 C(λ − A)−1 ,

λ ∈ R.

Thus the assumption C(λ − A)−1 u = 0 yields (λ − A× )−1 u = (λ − A)−1 u,

λ ∈ R.

(3.22)

The fact that u ∈ M × implies that (λ − A× )−1 u is analytic on λ ≥ 0. On the other hand, u ∈ M gives that (λ − A)−1 is analytic on λ ≤ 0. Since both (λ − A× )−1 u and (λ − A)−1 u vanish at inﬁnity, Liouville’s theorem implies that (λ − A)−1 u is identically zero on R, hence u = 0. There is an intimate connection between the Riemann-Hilbert boundary value problem (on the real line) and the homogeneous Wiener-Hopf integral equation. This is already clear from the material presented in Section 1.1 by specializing to the situation where f = 0. The fact is further underlined by the above proof of Theorem 3.7. Indeed, notice that (3.19) implies that φ+ = −Cγ+ , and hence we see from the deﬁnition of γ+ that φ+ (t) −

0

∞

k(t − s)φ+ (s) ds = 0,

t > 0,

where k(t) = −C(t)B, and hence k(λ) = −C(λIn − A)−1 B. Thus φ+ is the solution of the homogeneous Wiener-Hopf integral equation with symbol given by Im + C(λIn − A)−1 B. A more detailed (but straightforward) analysis gives the following result, the formulation of which is in line with Theorem 3.3.

3.5. The Riemann-Hilbert boundary value problem

55

Theorem 3.8. Let Im −K(λ) = Im +C(λIn −A)−1 B be a realization for the symbol of the homogeneous Wiener-Hopf equation ∞ k(t − s)φ(s)ds = 0, t ≥ 0, (3.23) φ(t) − 0

×

and let A = A − BC. Assume that both A and A× have no real eigenvalues, in other words, det Im − K(λ) = 0, −∞ < λ < +∞. Let M be the spectral subspace of A corresponding to the eigenvalues of A in the upper half plane, and let M × be the spectral subspace of A× corresponding to the eigenvalues of A× in the lower half plane. Then φ is a solution of (3.23) if and only if there exists x ∈ M ∩ M × such that ×

φ(t) = Ce−itA x,

t ≥ 0.

(3.24)

Moreover, the vector x in (3.24) is uniquely determined by φ. Formula (3.24) has to be understood in the sense of equality in the solution m space Lm 1 [0, ∞) (or, more generally, Lp [0, ∞) with 1 ≤ p < ∞; cf., Section 1.1 and Theorem 3.3). As a direct consequence of Theorem 3.8, one sees that the dimension of the null space of the Wiener-Hopf integral operator T deﬁned by the left-hand side of (3.23) is equal to dim(M ∩ M × ). It can also be proved that the codimension of its range is equal to codim (M + M × ). In fact, under the conditions of Theorem 3.8, the operator T is a Fredholm operator (see Section XI.1 in [51] for the deﬁnition of this notion), and its Fredholm index, which is deﬁned as the diﬀerence of the codimension of its range and the dimension of its null space, is equal to ind T

=

codim (M + M × ) − dim(M ∩ M × )

=

dim

Cn − dim(M ∩ M × ) M + M×

=

dim

M + M× Cn − dim − dim(M ∩ M × ) M× M×

= = =

M Cn − dim − dim(M ∩ M × ) × M M ∩ M× Cn dim × − dim M M dim

rank P × − rank P.

Here P and P × are the spectral projections corresponding to the eigenvalues in the upper half plane of A and A× , respectively. (In the step from the third to the fourth equality in the above calculation we used Lemma 2 in [89].) More detailed

56

Chapter 3. Explicit solutions using realizations

information about the null space and range of the Wiener-Hopf integral operator T can be obtained in this way (see, e.g., Theorem XIII.8.1 in [51]). We shall return to this theme, in a more general context, in Chapter 7, where it will be shown that the factorization indices in a non-canonical Wiener-Hopf factorization can be expressed in terms of the spaces M and M × , and related subspaces deﬁned in terms of these spaces and the matrices appearing in the realization of the symbol.

Notes The ﬁrst section of this chapter originates from Section 1.2 in [11]. The basic facts about Cauchy domains (see also the ﬁnal paragraphs of Chapter 0), Riesz projections and spectral subspaces, used in this ﬁrst section, can be found in Sections I.1 – I.3 of [51]. The material in Sections 3.2, 3.3 and 3.4 goes back to Chapter 4 in [11]. For Section 3.5 we refer to [12]. We shall return to canonical factorization in a more general setting in Chapters 5 and 7; see Theorems 5.14 and 7.1. Other state space methods for solving convolution equations, also based on matrix-valued realizations but not employing factorization, are developed in [12] and [13].

Chapter 4

Factorization of non-proper rational matrix functions In this chapter we treat the problem of factorizing a non-proper rational matrix function. The realization used in the earlier chapters is replaced by W (λ) = I + C(λG − A)−1 B.

(4.1)

Here I = Im is the m × m identity matrix, A and G are square matrices of order n say, and the matrices C and B are of sizes m × n and n × m, respectively. Any rational m × m matrix function W , proper or non-proper, admits such a representation. The representation (4.1) allows us to extend the results obtained in the previous chapter to arbitrary rational matrix functions. As an application we treat the problem to invert a singular integral operator with a rational matrix symbol. This chapter consists of ﬁve sections. In Section 4.1 we review the spectral theory of matrix pencils. Section 4.2 presents the realization theorem for nonproper rational matrix functions referred to in the previous paragraph. The corresponding canonical factorization theorem is given in Section 4.3. The ﬁnal two sections deal with applications to inverting singular integral operators (Section 4.4) and solving Riemann-Hilbert problems (Section 4.5).

4.1 Preliminaries about matrix pencils Let A and G be complex n×n matrices. The linear matrix-valued function λG−A, where λ is a complex variable, is called a (linear matrix ) pencil . We say that the pencil λG − A is regular on Ω or Ω-regular if λG − A is invertible for each λ ∈ Ω. Here Ω is a subset of C. From now on Γ will be a Cauchy contour. Its interior domain is denoted by F+ and its exterior domain by F− . We shall assume that ∞ ∈ F− . Pencils that

58

Chapter 4. Factorization of non-proper rational matrix functions

are Γ-regular admit block matrix partitionings that are comparable to spectral decompositions of a single matrix. This fact is summarized by the following theorem, the proof of which can be found in [140] (see also Section IV.1 of [51]). Theorem 4.1. Let λG − A be a Γ-regular pencil, and let the matrices P and Q be deﬁned by 1 P = 2πi

−1

Γ

G(λG − A)

dλ,

1 Q= 2πi

Γ

(λG − A)−1 Gdλ.

(4.2)

Then P and Q are projections such that (i) P A = AQ and P G = GQ, (ii) (λG − A)−1 P = Q(λG − A)−1 on Γ and this function has an analytic continuation on F− which vanishes at ∞, (iii) (λG − A)−1 (I − P ) = (I − Q)(λG − A)−1 on Γ and this function has an analytic continuation on F+ . The properties (i)–(iii) in the above proposition determine P and Q uniquely, that is, if P and Q are projections such that (i)–(iii) hold, then P and Q are given by the integral formulas in (4.2). For a better understanding of the above result, let us write A and G as block matrices relative to the decompositions of Cm induced by the projections P and Q. Condition (i) in Theorem 4.1 implies that A and G have block diagonal representations: A = G =

A1

0

0

A2

G1

0

0

G2

˙ Ker Q → Im P + ˙ Ker P, : Im Q + ˙ Ker Q → Im P + ˙ Ker P. : Im Q +

Property (ii) is equivalent to saying that the pencil λG1 − A1 is regular on F− and G1 is invertible; property (iii) amounts to regularity of the pencil λG2 − A2 on F+ . In the particular case when G is the identity matrix I, the two projections P and Q coincide, and P is just the spectral (or Riesz) projection of A corresponding to the eigenvalues in F+ . The latter means (see Section 3.1 or Section I.2 in [51]) that P is a projection commuting with A, the eigenvalues of A|Im P are in F+ and the eigenvalues of A|Ker P are in F− . In that case, Im P is the spectral subspace of A corresponding to the eigenvalues of A in F+ , and Ker P is the spectral subspace of A corresponding to the eigenvalues of A in F− .

4.2. Realization of a non-proper rational matrix function

59

4.2 Realization of a non-proper rational matrix function In this section we derive the representation (4.1), and present some useful identities related to (4.1). Theorem 4.2. Let W be a rational m × m matrix function, and let Ω be the subset of C on which W is analytic. Then, given an m × m matrix D, the function W admits a representation W (λ) = D + C(λG − A)−1 B,

λ ∈ Ω,

(4.3)

where λG − A is an Ω-regular m × m matrix pencil, and B and C are matrices of sizes n × m and m × n, respectively. The set Ω is the complement in C of the set of ﬁnite poles of W (i.e., the poles of W in C). In later applications, D will be taken to be Im , the m × m identity matrix. Proof. Let us ﬁrst remark that W admits a decomposition W (λ) = K(λ) + L(λ),

λ ∈ Ω,

(4.4)

where L is an m × m matrix polynomial and K is a proper rational m × m matrix such that the subset of C on which K is analytic coincides with Ω. Such a decomposition is not unique. In fact, given (4.4) we can obtain another decomposition of F with the same properties by adding a constant matrix to K and subtracting the same matrix from L. This, however, is all the freedom one has. In other words the decomposition (4.4) will be unique if we ﬁx the value of K at inﬁnity. From now on we shall assume that K(∞) = D. The results obtained in Section 2.2 then imply that K admits a realization K(λ) = D + CK (λ − AK )−1 BK ,

λ ∈ Ω,

where AK , BK and CK are matrices of appropriate sizes and the resolvent set of the (square) matrix AK coincides with Ω. The latter can be reformulated by saying that the eigenvalues of AK are just the ﬁnite poles of W . Proceeding with the second term in the right-hand side of the identity (4.4), we write L(λ) = L0 + λL1 + · · · + λq Lq , and introduce ⎡ ⎢ ⎢ GL = ⎢ ⎢ ⎣

⎤

0 Im 0

.. ..

. .

Im 0

⎥ ⎥ ⎥, ⎥ ⎦

⎡ ⎢ ⎢ BL = ⎢ ⎢ ⎣

L0 L1 .. . Lq

⎤ ⎥ ⎥ ⎥, ⎥ ⎦

CL =

−Im

0

···

0

,

60

Chapter 4. Factorization of non-proper rational matrix functions

where the blanks in GL indicate zero entries. The matrix GL is square of size l = m(q + 1). Also GL is nilpotent (of order q + 1), and hence Il − λGL is invertible for each λ in C. The ﬁrst row in the block matrix representation of (Il − λGL )−1 is equal to [ Im λIm . . . λq Im ] and it follows that L(λ) = CL (λGL − Il )−1 BL on all of the (ﬁnite) complex plane. By combining the representation results for K and L we see that W can be written in the form (4.3) with A=

AK

0

0

Il

,

B=

BK BL

,

C=

CK

CL

,

G=

I

0

0

GL

.

Here I is the identity matrix of the same size as AK . The fact that GL is nilpotent, implies that the matrix λG − A is invertible if and only if λ is an eigenvalue of AK , that is if and only if λ is a ﬁnite pole of W . The following proposition, which describes some elementary operations on a rational matrix function in terms of a given realization, is the natural analogue of Theorem 2.4 for realizations of the form (4.1). Theorem 4.3. Let W (λ) = I + C(λG− A)−1 B, and put A× = A− BC. Then W (λ) is invertible if and only if λG − A× is invertible, and in that case the following identities hold: W (λ)−1 = I − C(λG − A× )−1 B,

(4.5)

W (λ)C(λG − A× )−1 = C(λG − A)−1 ,

(4.6)

(λG − A× )−1 BW (λ) = (λG − A)−1 B,

(4.7)

(λG − A× )−1 = (λG − A)−1 − (λG − A)−1 BW (λ)−1 C(λG − A)−1 .

(4.8)

Proof. Fix λ ∈ C such that λG − A is invertible. Then det W (λ)

=

det I + C(λG − A)−1 B = det I + (λG − A)−1 BC

=

det (λG − A)−1 det(λG − A + BC)

=

det(λG − A× ) . det(λG − A)

It follows that W (λ) is invertible if and only if λG − A× is invertible. Also, in that case, a straightforward computation yields

4.3. Explicit canonical factorization

61

W (λ)C(λG − A× )−1 − C(λG − A× )−1 = C(λG − A)−1 BC(λG − A× )−1 = C(λG − A)−1 A − A× (λG − A× )−1 = C(λG − A)−1 (λG − A× ) − (λG − A) (λG − A× )−1 = C(λG − A)−1 − C(λG − A× )−1 . Since W (λ) is invertible, this proves (4.6). The identity (4.7) is proved in a similar way. Using (4.6) a straightforward computation shows that W (λ) I − C(λG − A× )−1 B = I, and hence (4.5) holds. Finally, (4.8) follows by applying (4.6) and again using the identity BC = (λG − A× ) − (λG − A). Instead of the above argument one can also use an analogue of the second proof of Theorem 2.1 in [20], which uses Schur complements arguments (cf., the remark made in the ﬁnal paragraph of Section 2.4).

4.3 Explicit canonical factorization In this section we show how the realization (4.1) can be used to construct a canonical factorization of an arbitrary rational matrix function. Necessary and suﬃcient conditions for the existence of such a factorization and formulas for the factors are stated explicitly in terms of the data appearing in the realization. The next theorem, a counterpart of Theorem 3.2 for non-proper rational matrix functions, is the main result. Theorem 4.4. Let W be a rational m × m matrix function without poles on the curve Γ, and let W be given by the Γ-regular realization W (λ) = I + C(λG − A)−1 B,

λ ∈ Γ.

(4.9)

Put A× = A − BC. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (i) the pencil λG − A× is Γ-regular, ˙ Ker P × and Cn = Im Q + ˙ Ker Q× . (ii) Cn = Im P + Here n is the order of the matrices G and A, and 1 1 G(λG − A)−1 dλ, P× = G(λG − A× )−1 dλ, P = 2πi Γ 2πi Γ

62

Chapter 4. Factorization of non-proper rational matrix functions

1 1 −1 × Q= (λG − A) Gdλ, Q = (λG − A× )−1 Gdλ. 2πi Γ 2πi Γ If the conditions (i) and (ii) are satisﬁed, a right canonical factorization with respect to Γ is given by λ ∈ Γ,

W (λ) = W− (λ)W+ (λ),

where the factors and their inverses can be written as W− (λ)

=

I + C(λG − A)−1 (I − Δ)B, −1

(4.10)

W+ (λ)

=

I + CΛ(λG − A)

W−−1 (λ)

=

I − C(I − Λ)(λG − A× )−1 B,

(4.12)

W+−1 (λ)

=

I − C(λG − A× )−1 ΔB.

(4.13)

B,

(4.11)

Here Δ is the projection along Im P onto Ker P × , and Λ is the projection of Cn along Im Q onto Ker Q× . Finally, the ﬁrst equality in (ii) implies the second and conversely. Proof. We split the proof into four parts. The ﬁrst part concerns the condition (i). In the second part we prove that the ﬁrst equality in (ii) implies the second and conversely. In the third part we use (i) and (ii) to derive the canonical factorization and the formulas for its factors. The ﬁnal part concerns the necessity of the condition (ii). Part 1. From the deﬁnition given in Section 3.1 it is clear that a necessary condition in order that W admits a right canonical factorization with respect to Γ is that W takes invertible values on Γ. By Theorem 4.3 this necessary condition is fulﬁlled if and only if (i) holds true. In what follows we shall assume that (i) is satisﬁed. Part 2. In this part we prove the last statement of the theorem. Consider the operators P × |Im P : Im P → Im P × ,

Q× |Im Q : Im Q → Im Q× .

(4.14)

The ﬁrst equality in (ii) is equivalent to the invertibility of the ﬁrst operator in (4.14). To see this, note that Ker P × |Im P = Ker P × ∩ Im P , and thus P × |Im P is injective if and only if Ker P × ∩ Im P = {0}. Next, observe each that for y ∈ Im P we have y = (I − P × )y + P × |ImP y ∈ Ker P × + Im P × |Im P . Thus Ker P × + Im P ⊂ Ker P × + Im P × |Im P . The reverse inclusion is also true. × × × Indeed, for z ∈ Im P we have P z = ×(P z − z) + z ∈ Ker P × + Im P . It follows × × that Ker P + Im P |Im P = Ker P + Im P , and hence P |Im P considered as an operator into Im P × is surjective if and only if Cn = Ker P × + Im P . Thus, as claimed, the ﬁrst identity in (ii) amounts to the same as the invertibility of the ﬁrst operator in (4.14). Similarly, the second equality in (ii) is equivalent to the invertibility of the second operator in (4.14). Notice that GQ = P G,

GQ× = P × G,

(4.15)

4.3. Explicit canonical factorization

63

which is clear from the deﬁnitions of the projections Q, P and Q× , P × . Furthermore, from the material presented in Section 4.1, applied to λG − A as well as to λG − A× , we see that G maps Im Q and Im Q× in a one-one manner onto Im P and Im P × , respectively. Thus the operators E = G|Im Q : Im Q → Im P and E × = G|Im Q× : Im Q× → Im P × are invertible and, in addition, E × (Q× |Im Q ) = (P × |Im P )E. So the operators in (4.14) are equivalent, and hence the ﬁrst operator in (4.14) is invertible if and only if the same is true for the second operator in (4.14). This proves that the ﬁrst equality in (ii) implies the second and vice versa. Part 3. Next assume that (i) and the direct sum decompositions in (ii) hold true. Our aim is to obtain a canonical factorization of W . Write A, G, B, C as well as A× = A − BC in block form relative to the decompositions in (ii): A11 A12 ˙ Ker P × , ˙ Ker Q× → Im P + (4.16) A = : Im Q + 0 A22 G = B

=

C

=

×

A

=

G11

0

0

G22

B1 B2

˙ Ker Q× → Im P + ˙ Ker P × , : Im Q +

(4.17)

˙ Ker P × , : Cn → Im P +

C2

C1

A× 11

0

A× 21

A× 22

˙ Ker Q× → Cn , : Im Q +

(4.18) (4.19)

˙ Ker P × . ˙ Ker Q× → Im P + : Im Q +

(4.20)

From Theorem 4.1, applied to λG − A as well as to λG − A× , we know that AQ = P A,

A× Q× = P × A× .

(4.21)

The ﬁrst identity in (4.21) implies that A maps Im Q into Im P . This explains the zero entry in the left lower corner of the block matrix for A. From (4.15) we conclude that G has the desired block diagonal form. From the second identity in (4.21) it follows that A× maps Ker Q× into Ker P × , which justiﬁes the zero in the right upper corner of the block matrix for A× . Taking into account the identity A× = A − BC gives A12 = B1 C2 , A× 11 = A11 − B1 C1 ,

A× 21 = −B2 C1 , A× 22 = A22 − B2 C2 .

(4.22) (4.23)

64

Chapter 4. Factorization of non-proper rational matrix functions

Deﬁne the matrix functions W− and W+ by (4.10) and (4.11), respectively. Using the block matrix representations of A, G, B, and C we may rewrite W− and W+ in the form W− (λ) = I + C1 (λG1 − A11 )−1 B1 , λ ∈ Γ, (4.24) W+ (λ) = I + C2 (λG2 − A22 )−1 B2 ,

λ ∈ Γ.

(4.25)

From the block matrix representation of A and the ﬁrst identity in (4.22) we see that ⎡ ⎤−1 ⎡ ⎤ −B1 C2 B1 λG1 − A11 ⎦ ⎣ ⎦ W− (λ)W+ (λ) = I + C1 C2 ⎣ 0 λG2 − A22 B2 = I + C(λG − A)−1 B = W (λ), which gives the factorization W = W− W+ . Next, we check the analytic properties of the factors. Obviously, W− and W+ have no poles on Γ. Note that λG1 − A11 = (λG − A)|Im Q : Im Q → Im P. Thus we know from Section 4.1 that (λG1 − A11 )−1 has an analytic extension on F− which vanishes at inﬁnity. So W− is continuous on F− ∪ Γ and analytic on F− (including inﬁnity). To see that a similar statement holds true for W+ on F+ , we ﬁrst note that the linear maps J

=

(I − Q)|Ker Q× : Ker Q× → Ker Q,

H

=

(I − P )|Ker P × : Ker P × → Ker P,

are invertible. In fact, J −1 = Λ|Ker Q and H −1 = Δ|Ker P , where Λ is the projection along Im Q onto Ker Q× , and Δ is the projection along Im P onto Ker P × . Next, take x ∈ Ker Q× . Then (λG2 − A22 )x = Δ(λG − A)x = Δ(λG − A)(I − Q)x = Δ(λG − A)Jx, which shows that H(λG2 − A22 ) = (λG − A)|Ker Q J. But then we can use Theorem 4.1 and the invertibility of the operators H and J to show that the function (λG2 −A2 )−1 has an analytic extension on F+ . Hence W+ is continuous on F+ ∪ Γ and analytic on F+ . From the factorization W (λ) = W− (λ)W+ (λ) for λ ∈ Γ it follows that W− (λ) and W+ (λ) are both invertible for each λ ∈ Γ. So we can apply Theorem 4.3 to show that W−−1 (λ)

=

−1 I − C1 (λG1 − A× B1 , 11 )

(4.26)

W+−1 (λ)

=

−1 I − C2 (λG2 − A× B2 . 22 )

(4.27)

4.3. Explicit canonical factorization

65

Here we use the two identities in (4.23). Using the block matrix representations of A, G, B and C given above, it is clear that (4.26) and (4.27) yield the formulas (4.12) and (4.13), respectively. We proceed by checking the analyticity properties of the functions W−−1 and −1 W+ . First note that × × × λG2 − A× 22 = (λG − A )|Ker Q× : Ker Q → Ker P .

Thus by applying Theorem 4.1 with λG − A× in place of λG − A we see that −1 the function (λG2 − A× has an analytic extension on F+ . It follows that the 22 ) −1 function W+ is continuous on F+ ∪ Γ and analytic on F+ . To prove the analogous result for W−−1 with respect to F− we use that × × H × (λG1 − A× 11 ) = (λG − A )|Im Q× J , where J × = Q× |Im Q : Im Q → Im Q× and H × = P × |Im P : Im P → Im P × are invertible linear maps of which the inverses are given by (J × )−1 = (I − Λ)|Im Q× ,

(H × )−1 = (I − Δ)|Im P × .

−1 Since (λG − A× )|Im Q× is analytic on F− by virtue of Theorem 4.1 applied −1 . Hence the to λG − A× , we conclude that the same holds true for (λG1 − A× 11 ) function W− (λ)−1 is continuous on F− ∪ Γ and analytic on F− . Thus we have proved that W = W− W+ is a right canonical factorization with respect to the curve Γ. Part 4. In this part we prove the necessity of the equalities in (ii). So in what follows we assume that W = W− W+ is a canonical factorization of W with respect to Γ. Take x ∈ Im P ∩ Ker P × and, for λ ∈ Γ, put ϕ− (λ) = C(λG − A)−1 x,

ϕ+ (λ) = C(λG − A× )−1 x.

Since x ∈ Im P , the ﬁrst identity in (4.21) allows us to rewrite ϕ− as −1 ϕ− (λ) = (C|Im Q ) (λG − A)|Im Q x, and hence Theorem 4.1(ii) implies that ϕ− has an analytic continuation on F− which vanishes at inﬁnity. Similarly, since −1 x, ϕ+ (λ) = (C|Ker Q× ) (λG − A× )−1 |Ker Q× we conclude from Theorem 4.1(iii) applied to λG − A× that ϕ+ has an analytic continuation on F+ . Note that W (λ)−1 ϕ− (λ) = ϕ+ (λ) for each λ ∈ Γ, because of formula (4.6) in Theorem 4.3. It follows that W− (λ)−1 ϕ− (λ) = W+ (λ)ϕ+ (λ),

λ ∈ Γ.

66

Chapter 4. Factorization of non-proper rational matrix functions

Now use the analyticity properties of the factors W− and W+ . We conclude that W−−1 ϕ− has an analytic continuation on F− which vanishes at inﬁnity, and W+ ϕ+ has an analytic continuation on F+ . Liouville’s theorem implies that both functions are identically zero. It follows that ϕ− (λ) = 0 for each λ ∈ Γ. But then we can apply formula (4.8) to show that (λG − A× )−1 x = (λG − A)−1 x,

λ ∈ Γ.

Now, repeat part of the above reasoning. Note that (λG − A)−1 x has an analytic continuation on F− which vanishes at inﬁnity, and (λG − A× )−1 x has an analytic continuation on F+ . Again using Liouville’s theorem we conclude that both matrix functions (λG − A)−1 x and (λG − A× )−1 x are identically zero on Γ. This yields x = 0. We proved that Im P ∩ Ker P × = {0}. Recall that G maps Im Q in a one-one manner onto Im P . Thus (4.15) shows that G maps Im Q ∩ Ker Q× in a one-one manner into Im P ∩ Ker P × . Hence Im Q ∩ Ker Q× = {0} too. Next we show that Im Q + Ker Q× = Cn . Take y ∈ Cn such that y is orthogonal to Im Q + Ker Q× . Let y ∗ be the row vector of which the j-th entry is equal to the complex conjugate of the j-th entry of y (j = 1, . . . , m). For λ ∈ Γ, put ψ− (λ) = y ∗ (λG − A× )−1 B,

ψ+ (λ) = y ∗ (λG − A)−1 B.

Since y ∗ (I − Q)× = 0, Theorem 4.1 shows that ψ− (λ) = y ∗ (λG − A× )−1 P × B, and thus ψ− has an analytic continuation on F− which vanishes at inﬁnity. Similarly, y ∗ Q = 0 implies that ψ+ has an analytic continuation on F+ . Now, use the canonical factorization W = W− W+ and (4.7) to show that ψ+ (λ)W+ (λ)−1 = ψ− (λ)W− (λ),

λ ∈ Γ.

But then, as before, we can use Liouville’s theorem to show that both sides of the identity are equal to zero. It follows that ψ+ (λ) = 0 for each λ ∈ Γ, and we can use formula (4.8) to show that y ∗ (λG − A× )−1 = y ∗ (λG − A)−1 ,

λ ∈ Γ.

Recall that y ∗ Q and y ∗ (I − Q× ) are both zero. Thus Theorem 4.1 implies that y ∗ (λG − A× )−1 has an analytic continuation on F− which vanishes at inﬁnity, and the function y ∗ (λG − A)−1 has an analytic continuation on F+ . So, by Liouville’s theorem, y ∗ (λG−A)−1 = 0 on Γ, and thus y = 0. This gives Im Q+ Ker Q× = Cn . Combining this with with what we saw in the preceding paragraph, we obtain ˙ Ker Q× = Cn . But then the result of Part 2 yields the direct sum decomIm Q + position Im P˙+ Ker P × = Cn , and (ii) is proved. The fact that in Theorem 4.4 the curve Γ is bounded is not essential. We only use that Γ is a closed curve on the Riemann sphere C∞ and that W has no poles on Γ. Thus Γ may pass through inﬁnity. For instance, let us replace Γ by the

4.3. Explicit canonical factorization

67

extended real line R∞ which passes through inﬁnity. By the results of Section 2.2, the condition that the m × m rational matrix function W has no poles on R ∪ {∞} implies that W can be represented in the form W (λ) = D + C(λ − A)−1 B,

λ ∈ R,

(4.28)

where A is a square matrix with no real eigenvalues. The condition that W takes invertible values on R ∪ {∞} now amounts to the requirement that D is invertible and the matrix A − BD−1 C has no real eigenvalues. Also, in that case, W −1 (λ) = D −1 − D−1 C(λ − A× )−1 BD −1 ,

λ ∈ R,

where A× = A−BD−1 C. With these minor modiﬁcations the proof of Theorem 4.4 also applies to realizations of the form (4.28), and yields the following theorem. Theorem 4.5. Let W be a rational m × m matrix function without poles on the real line, and let W be given by the realization W (λ) = D + C(λIn − A)−1 B,

λ ∈ R,

(4.29)

where A is an n×n matrix with no real eigenvalues. Then W admits a right canonical factorization with respect to R ∪ {∞} if and only if the following conditions are satisﬁed: (i) D is invertible and A× = A − BD−1 C has no real eigenvalues, ˙ M ×. (ii) Cn = M + Here n is the order of the matrix A, the space M is the spectral subspace of A corresponding to its eigenvalues in the upper half plane, and M × is the spectral subspace of A× corresponding to its eigenvalues in the lower half plane. Furthermore, if the conditions (i) and (ii) are fulﬁlled, then a right canonical factorization with respect to R ∪ {∞} is given by W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

where the factors and their inverses can be written as W− (λ)

= D + C(λIn − A)−1 (I − Π)B,

W+ (λ)

= I + D−1 CΠ(λIn − A)−1 B,

W−−1 (λ)

= D−1 − D−1 C(I − Π)(λIn − A× )−1 BD −1 ,

W+−1 (λ)

= I − D−1 C(λIn − A× )−1 ΠB.

Here Π is the projection of Cn along M onto M × . Since there is no a priori assumption on the invertibility of (the external) operator D, Theorem 4.5 is a slight extension of Theorem 3.2 dealing with matrix functions too. The results can be generalized to the case of operator functions (cf., Section 7.1 below).

68

Chapter 4. Factorization of non-proper rational matrix functions

4.4 Inversion of singular operators with a rational matrix symbol In this section we apply the results of the previous sections to solve the problem of inverting the singular integral equation 1 ϕ(τ ) dτ = g(t), t ∈ Γ. (4.30) a(t)ϕ(t) + b(t) πi Γ τ − t Throughout we assume that a and b are rational m × m matrix functions which do not have poles on the Cauchy contour Γ. We shall analyze equation (4.30) under the additional condition that the diﬀerence a(λ) − b(λ) is invertible for each λ ∈ Γ. Since we are interested in invertibility, the latter condition is not an essential restriction (cf., Theorem 1.3). The fact that the matrix a(λ) − b(λ) is invertible for λ ∈ Γ allows us to introduce the operator T = MW PΓ + QΓ which we consider on Lm 2 (Γ). Here −1 a(λ) + b(λ) , W (λ) = a(λ) − b(λ) m and MW is the operator of multiplication by W on Lm 2 (Γ), that is, for ϕ ∈ L2 (Γ) we have (MW ϕ)(t) = W (t)ϕ(t) for almost all t ∈ Γ. Furthermore, P Γ and QΓ are the orthogonal projections on Lm 2 (Γ) associated with the singular integral operator introduced in Section 1.3. Thus, for ϕ ∈ Lm 2 (Γ), 1 ϕ(τ ) 1 (P Γ ϕ)(t) = ϕ(t) + dτ, (4.31) 2 2πi Γ τ − t

(QΓ ϕ)(t) =

1 1 ϕ(t) − 2 2πi

Γ

ϕ(τ ) dτ, τ −t

(4.32)

for almost all t ∈ Γ. The image of P Γ consists of all functions in Lm 2 (Γ) that admit an analytic continuation into F+ . Similarly, the image of QΓ is the subspace of all functions in Lm 2 (Γ) that admit an analytic continuation into F− and vanish at inﬁnity. Note that equation (4.30) is equivalent to −1 g(λ). (MW P Γ + QΓ )ϕ = g, where g(λ) = a(λ) − b(λ) Since W is a rational m × m matrix function without poles on Γ, we know from Theorem 4.2 that W admits a Γ-regular realization W (λ) = I + C(λG − A)−1 B,

λ ∈ Γ.

(4.33)

The main result of this section provides an explicit inversion formula for the operator MW P Γ + QΓ in terms of the realization (4.33). Theorem 4.6. Let the rational m × m matrix function W be given by the Γ-regular realization (4.33), and put A× = A − BC. Then MW P Γ + QΓ is an invertible operator on Lm 2 (Γ) if and only if the following two conditions are satisﬁed:

4.4. Inversion of singular operators with a rational matrix symbol

69

(1) the pencil λG − A× is Γ-regular, ˙ Ker P × , (2) Cn = Im P + where n is the order of the matrices A and G, and 1 P = 2πi

Γ

G(λG − A)

−1

dλ,

P

×

1 = 2πi

Γ

G(λG − A× )−1 dλ.

(4.34)

In that case

(MW P Γ + QΓ )−1 g (λ)

=

g(λ) − C(λG − A× )−1 B(P Γ g)(λ) + C(λG − A× )−1 − C(λG − A× )−1 (I − Π) 1 P × G(ζG − A× )−1 Bg(ζ)dζ , λ ∈ Γ. · 2πi Γ

Here Π is the projection of Cn along Im P onto Ker P × . With suitable changes, the theorem remains true when P and P × are replaced by the projections Q and Q× (also) appearing in Theorem 4.4. Proof. From the general theory of singular integral equations reviewed in Section 1.3 we know that the operator MW P Γ + QΓ is invertible if and only if W admits a right canonical factorization with respect to Γ. Since W is given by (4.33), the latter is the case if and only if conditions (i) and (ii) in Theorem 4.4 are fulﬁlled. By the ﬁnal statement in Theorem 4.4, conditions (i) and (ii) in Theorem 4.4 are equivalent to conditions (1) and (2) in the present theorem. Thus we have proved that MW P Γ + QΓ is invertible if and only if (1) and (2) are satisﬁed. To get the formula for the inverse of MW P Γ + QΓ we again use the general theory of singular integral equations, the inversion formula (1.29) in particular. Let W = W− W+ be a right canonical factorization of W with respect to Γ. For g ∈ Lm 2 (Γ) we then have, suppressing the variable λ, (MW P Γ + QΓ )−1 g = W+−1 P Γ (W−−1 g) + W− QΓ (W−−1 g) . Taking into account the form of P Γ and QΓ in (4.31) and (4.32), this identity can be rewritten as 1 1 (MW P Γ + QΓ )−1 g (λ) = g(λ) + W (λ)−1 g(λ) 2 2 1 1 −1 W+ (λ) − W− (λ) W− (τ )−1 g(τ ) dτ, + 2πi Γ τ − λ

λ ∈ Γ. (4.35)

70

Chapter 4. Factorization of non-proper rational matrix functions

Next, we use the formulas for W+ , W− and their inverses given in Theorem 4.4. This yields W+ (λ)−1 − W− (λ) W− (τ )−1 = −C(λG − A× )−1 ΔB − C(λG − A)−1 (I − Δ)B × −1

+ λG − A )

× −1

ΔBC(I − Λ)(τ G − A )

−1

+ C(λG − A)

(4.36)

B

(I − Δ)BC(I − Λ)(τ G − A× )−1 B.

Here Δ and Λ are the projections deﬁned in Theorem 4.4. Using these deﬁnitions, and the partitionings of A, G, and A× in (4.16), (4.17) and (4.20), respectively, we obtain ΔA(I − Λ) = 0,

(I − Δ)A× Λ = 0,

ΔG = GΛ.

Since BC = A − A× , it follows that ΔBC(I − Λ) = A× Λ − ΔA× = (A× − λG)Λ − Δ(A× − τ G) − (τ − λ)ΔG, and (I − Δ)BC(I − Λ) = A(I − Λ) − (I − Δ)A× = (A − λG)(I − Λ) − (I − Δ)(A× − τ G) − (τ − λ)(I − Δ)G. Inserting these expressions into (4.36) gives W+ (λ)−1 − W− (λ) W− (τ )−1 = −C(τ G − A× )−1 B −(τ − λ)C(λG − A× )−1 ΔG(τ G − A× )−1 B −(τ − λ)C(λG − A)−1 (I − Δ)G(τ G − A× )−1 B. Next we use that (τ − λ)C(λG − A× )−1 G(τ G − A× )−1 B can be written as C(λG − A× )−1 (τ G − A× ) − (λG − A× ) (τ G − A× )−1 B which in turn is equal to C(λG − A× )−1 B − C(τ G − A× )−1 B, and this leads to W+ (λ)−1 − W− (λ) W− (τ )−1 = −C(λG − A× )−1 B +(τ − λ) C(λG − A× )−1 − C(λG − A)−1 ·(I − Δ)G(τ G − A× )−1 B.

(4.37)

4.5. The Riemann-Hilbert boundary value problem revisited (1)

71

Using (4.37) and (4.5) in (4.35) we obtain 1 (MW P Γ + QΓ )−1 g (λ) = g(λ) − C(λG − A× )−1 Bg(λ) 2 1 1 × −1 −C(λG − A ) B g(τ ) dτ 2πi Γ τ − λ + C(λG − A× )−1 − C(λG − A)−1 (I − Δ) 1 × −1 · G(τ G − A ) Bg(τ ) dτ , 2πi Γ

λ ∈ Γ.

Finally, note that Δ = Π and (I − Π)P × = I − Π. Since P Γ is given by (4.31), we see that we have derived the desired expression for the inverse of the operator M W P Γ + QΓ .

4.5 The Riemann-Hilbert boundary value problem revisited (1) In this section we treat the (homogeneous) Riemann-Hilbert boundary value problem for non-proper rational matrix functions. As before Γ is a Cauchy contour. As usual, the interior domain of Γ is denoted by F+ , and its exterior domain, which contains the point inﬁnity, by F− . Throughout W is a rational m × m matrix function which does not have poles on Γ. We say that a pair of Cm -valued functions Φ+ , Φ− is a solution of the Riemann-Hilbert boundary problem of W with respect to Γ if Φ+ and Φ− are continuous on F+ ∪ Γ and F− ∪ Γ, respectively, Φ+ and Φ− are analytic in F+ and F− , respectively, Φ− vanishes at inﬁnity, and W (λ)Φ+ (λ) = Φ− (λ),

λ ∈ Γ.

(4.38)

Since W is assumed to be a rational m × m matrix function which has no poles on Γ, we may assume that W is given by a Γ-regular realization W (λ) = I + C(λG − A)−1 B,

λ ∈ Γ.

(4.39)

We shall also assume that W takes invertible values on Γ. This additional condition is equivalent to the requirement that the pencil λG−A× is Γ-regular. The following theorem is the natural analogue of Theorem 3.7. Theorem 4.7. Let W be given by (4.39), and assume that the pencil λG − A× is a Γ-regular. Put 1 1 G(λG − A)−1 dλ, P× = G(λG − A× )−1 dλ. P = 2πi Γ 2πi Γ

72

Chapter 4. Factorization of non-proper rational matrix functions

Then the pair of functions Φ+ and Φ− is a solution of the Riemann-Hilbert boundary value problem of W with respect to Γ if and only if there exists x belonging to Im P ∩ Ker P × such that Φ+ (λ) = C(λG − A× )−1 x,

Φ− (λ) = C(λG − A)−1 x.

(4.40)

Moreover the vector x in (4.40) is uniquely determined by Φ+ , Φ− With the appropriate modiﬁcations, the theorem remains true when P and P × are replaced by the projections Q and Q× (also) appearing in Theorem 4.4. Proof. Take x ∈ Im P ∩ Ker P × , and deﬁne Φ+ and Φ− by (4.40). Formula (4.6) implies that (4.38) is satisﬁed. Since x = P x, Theorem 4.1 (ii) shows that Φ− is continuous on F− ∪ Γ, analytic in F− , and vanishes at inﬁnity. Similarly, using x = (I −P × )x, Theorem 4.1 (iii), applied to λG−A× , yields that Φ+ is continuous on F+ ∪ Γ and analytic on F+ . Thus the functions Φ+ and Φ− have the desired properties, and the pair Φ+ , Φ− is a solution. To prove the converse, assume that the pair Φ+ , Φ− is a solution of the Riemann-Hilbert problem for W with respect to Γ. For λ ∈ Γ, introduce ρ(λ) = (λG−A)−1 BΦ+ (λ). The n×m matrix function ρ is continuous on Γ, thus it makes sense to put 1 ρ(τ ) 1 ρ+ (λ) = ρ(λ) + dτ, λ ∈ Γ, 2 2πi Γ τ − λ 1 1 ρ(τ ) ρ(λ) − dτ, λ ∈ Γ; ρ− (λ) = 2 2πi Γ τ − λ cf., the expressions (4.31) and (4.32). The function ρ+ is continuous on F+ ∪ Γ and analytic in F+ , and ρ− has the same properties with F− in place of F+ . Moreover, ρ− vanishes at inﬁnity. We ﬁrst show that Φ+ (λ)

= −Cρ+ (λ),

λ ∈ F+ ∪ Γ,

(4.41)

Φ− (λ)

=

λ ∈ F− ∪ Γ.

(4.42)

Cρ− (λ),

Since the pair Φ+ , Φ− satisﬁes (4.38), we have Φ− (λ) = Φ+ (λ) + C(λG − A)−1 BΦ+ (λ) = Φ+ (λ) + Cρ(λ),

λ ∈ Γ.

But ρ(λ) = ρ− (λ) + ρ+ (λ) on Γ, and therefore Φ− (λ) − Cρ− (λ) = Φ+ (λ) + Cρ+ (λ),

λ ∈ Γ.

(4.43)

The right-hand side of (4.43) is continuous on F+ ∪ Γ and analytic in F+ . On the other hand, the left-hand side of (4.43) is continuous on F− ∪ Γ, analytic in

4.5. The Riemann-Hilbert boundary value problem revisited (1)

73

F− and vanishes at inﬁnity. Thus, by Liouville’s theorem, both sides of (4.43) are identically zero on Γ, and the identities (4.41) and (4.42) hold. Next, we compute the function ρ− . From the deﬁnition of ρ(λ) we see that (λG − A)ρ(λ) = BΦ+ (λ) for λ ∈ Γ. Since Φ+ is continuous on F+ ∪ Γ and analytic in F+ , we conclude that for each λ ∈ Γ, 1 1 1 (λG − A)ρ(λ) = (τ G − A)ρ(τ ) dτ 2 2πi Γ τ − λ 1 1 (λG − A) + (τ − λ)G ρ(τ ) dτ = 2πi Γ τ − λ 1 ρ(τ ) = (λG − A) dτ + x, 2πi Γ τ − λ 1 Gρ(τ ) dτ. 2πi Γ Using the deﬁnition of ρ− , the above calculation shows that where

x=

ρ− (λ) = (λG − A)−1 x,

λ ∈ Γ.

(4.44)

To compute ρ+ , recall that ρ(λ) = ρ− (λ) + ρ+ (λ) on Γ. This, together with (4.41) and (4.42), yields (λG − A)ρ+ (λ)

= (λG − A)ρ(λ) − (λG − A)ρ− (λ) = BΦ+ (λ) − x = −BCρ+ (λ) − x,

λ ∈ Γ.

×

Since A = A − BC, we obtain ρ+ (λ) = −(λG − A× )−1 x,

λ ∈ Γ.

(4.45)

From (4.44) and the fact that ρ− is continuous on F− ∪ Γ, analytic in F− , and vanishes at inﬁnity, we conclude that x = P x. Similarly, we obtain from (4.45) that x = (I − P × )x. Thus x ∈ Im P ∩ Ker P × . Formulas (4.41), (4.42), (4.44) and (4.45) now show that the functions Φ+ and Φ− have the desired representation (4.40). It remains to prove the uniqueness of the vector x in (4.40). To do this assume that u ∈ Im P ∩ Ker P × , and let C(λG − A)−1 u be identically zero on Γ. It suﬃces to show that u = 0. For this purpose we use the identity (4.8). Applying this identity to the vector u, we see that (λG − A× )−1 u = (λG − A)−1 u, ×

λ ∈ Γ.

(4.46)

Since u ∈ Ker P , the left-hand side of (4.46) has an analytic continuation on F+ ; see Theorem 4.1 (iii). Similarly, u ∈ Im P implies that the right side of (4.46) has an analytic continuation on F− which vanishes at inﬁnity; see Theorem 4.1 (ii). But then we can apply Liouville’s theorem to show that these functions are identically zero on Γ, which yields u = 0.

74

Chapter 4. Factorization of non-proper rational matrix functions

Notes The extension of the Riesz spectral theory for operators to operator pencils, which is described in Section 4.1, is due to Stummel [140]; the results can also be found in Section IV.1 of [51]. Section 4.2 combines the classical realization theory for proper rational matrix functions with that of matrix polynomials; for the latter, see [65]. The main source for the material in Sections 4.2 and 4.3 is the paper [55]; Section 4.4 is based on [56]. Section 4.5 seems to be new. For realizations of the form considered in this chapter, non-canonical Wiener-Hopf factorization has been studied in [151]. Instead of (4.3) other realizations of W can be used; see for instance [79], where (4.3) is replaced by the realization W (λ) = D + (λ − α)C(λG − A)−1 B which can also be used for non-square matrix functions.

Part III Equations with non-rational symbols In this part we carry out a program analogous to that of the second part, but now for certain classes of non-rational matrix and operator functions. Included are matrix functions analytic in a strip but not at inﬁnity, an operator function appearing in linear transport theory, and operator functions analytic on a given curve. There are three chapters. The main topic of the ﬁrst chapter (Chapter 5) is a canonical factorization theorem for matrix functions analytic in a strip but not necessarily at inﬁnity. Its applications to diﬀerent classes of Wiener-Hopf equations are included too. The realizations of such matrix functions require that we consider systems with an inﬁnite dimensional state space and with a state operator that is unbounded and exponentially dichotomous. Thus the theory of strongly continuous semigroups plays an important role in this material. Chapter 6 is entirely dedicated to the solution of an integro-diﬀerential equation from mathematical physics describing stationary migration of particles in a medium. To illustrate the approach, the special case of a ﬁnite number of scattering directions is considered ﬁrst. This restriction makes it possible to reduce the problem to a canonical factorization problem for rational matrix functions. The general situation features an inﬁnite dimensional separable Hilbert space as state space. The ﬁnal chapter (Chapter 7) deals with canonical factorization and non-canonical Wiener-Hopf factorization for operator-valued functions that are analytic on a given curve. In this chapter the so-called factorization indices are described in state space terms.

Chapter 5

Factorization of matrix functions analytic in a strip This chapter deals with m × m matrix-valued functions of the form ∞ W (λ) = I − eiλt k(t) dt,

(5.1)

−∞

where k is an m× m matrix-valued function with the property that for some ω < 0 the entries of e−ω|t| k(t) are Lebesgue integrable on the real line. In other words, k is of the form k(t) = eω|t| h(t) with h ∈ Lm×m (R). (5.2) 1 It follows that the function W is analytic in the strip |λ| < τ , where τ = −ω. This strip contains the real line. The aim is to extend the canonical factorization theorem of Chapter 5 to functions of the type (5.1). In general, the function W in (5.1) is not a rational matrix function, and hence one cannot expect a representation of W in the form W (λ) = I + C(λ − A)−1 B

(5.3)

with A, B, C matrices. Also a realization with A, B and C bounded linear operators will not work. Indeed, in that case the function W would be analytic at inﬁnity, however in general it is not. Thus to get a representation of the type (5.3) one has to allow for unbounded linear operators. In fact, we shall have to allow for A and C to be unbounded while B can be taken to be bounded. This chapter consists of nine sections. In Sections 5.1 and 5.2 we present preliminary material on exponentially dichotomous operators and associated bisemigroups. These exponentially dichotomous operators appear as state operators in the realization triples deﬁned in Section 5.3. In Section 5.4 we construct realization triples for m × m matrix-valued functions W of the form (5.1) with k as in

78

Chapter 5. Factorization of matrix functions analytic in a strip

(5.2), and in Section 5.5 we use the realization triples to invert such a matrix function W . It turns out that inversion is only possible when the associate operator A× = A − BC is exponentially dichotomous too. The inversion formula of Section 5.5 is used in Section 5.6 to derive an explicit formula for the kernel function of the inverse of a full line convolution integral operator when the symbol W is given by (5.1) and (5.2). This section also contains some preliminary material about Hankel operators. The ﬁnal three sections concern applications. Sections 5.7 and 5.8 deal with inversion of a Wiener-Hopf integral equation with a kernel function k of the form (5.2) and with canonical factorization of the corresponding symbol. In Section 5.9 we revisit the Riemann-Hilbert boundary value problem.

5.1 Exponentially dichotomous operators and bisemigroups We begin with some preliminaries about strongly continuous semigroups of operators (also called C0 -semigroups). Free use will be made of the standard theory of these semigroups as explained, for instance, in Chapter XIX of [51]. Besides ordinary C0 -semigroups deﬁned on the positive half line [0, ∞), henceforth to be called right semigroups, we shall also consider semigroups deﬁned on the negative half line (−∞, 0]. The latter will be called left semigroups. Notice that T (t) is a left semigroup if and only if T (−t) is a right semigroup. Let T (t) be a strongly continuous right or left semigroup. As is well-known, there exist constants M and ω such that T (t) ≤ M eω|t| ,

t ∈ J.

Here J is the half line [0, ∞) or (−∞, 0] according to T (t) being a right or a left semigroup. If the above inequality is satisﬁed for a given real number ω and some positive constant M , we say that T (t) is of exponential type ω. Semigroups of negative exponential type will be called exponentially decaying. Next we introduce the concept of an exponentially dichotomous operator. Let X be a complex Banach space, and let A be a (possibly unbounded) linear operator with domain D(A) in X and with values in X, in short A(X → X). Further, let P : X → X be a (bounded linear) projection of X commuting with A. The latter means that P maps D(A) into itself and P Ax = AP x for each x ∈ D(A). Put X− = Im P and X+ = Ker P . Then ˙ X+ , X = X− +

(5.4)

and this decomposition reduces A, that is, ˙ [D(A) ∩ X+ ], D(A) = [D(A) ∩ X− ] +

(5.5)

5.1. Exponentially dichotomous operators and bisemigroups

79

with A mapping [D(A) ∩ X− ] into X− and [D(A) ∩ X+ ] into X+ . So with respect to the decompositions (5.4) and (5.5), the operator A has the matrix representation 0 A− . (5.6) A= 0 A+ Here A− (X− → X− ) is the restriction of A to X− , and A+ (X+ → X+ ) is the restriction of A to X+ . In particular, the domain D(A− ) of A− is D(A) ∩ X− and the domain D(A+ ) of A− + is D(A) ∩ X+ . Thus (5.5) can be rewritten as ˙ D(A+ ). D(A) = D(A− ) + The operator A is said to be exponentially dichotomous if the operators A− and A+ in (5.6) are generators of exponentially decaying strongly continuous left and right semigroups, respectively. In that case the projection P , which will turn out to be unique (see Proposition 5.1 below), is called the separating projection for A. We say that A is of exponential type ω (< 0) if this is true for the semigroups generated by A− and A+ . Suppose, for the moment, that A : X → X is a bounded linear operator. Then A is exponentially dichotomous if and only if the spectrum σ(A) of A does not meet the imaginary axis. In that situation the separating projection for A is simply the Riesz projection corresponding to the part of σ(A) lying in the open right half plane λ > 0. Next, observe that generators of exponentially decaying strongly continuous semigroups belong to the class of exponentially dichotomous operators, the left semigroup case corresponding to the separating projection being the identity operator and the right semigroup case corresponding to the separating projection being the zero operator on X. Returning to the general case, we note that the operators A− and A+ in the deﬁnition of an exponentially dichotomous operator are closed and densely deﬁned. Hence the same is true for their direct sum A. Furthermore, if A is of (negative) exponential type ω, then, by the Hille-Yosida-Phillips theorem (see, e.g., Theorem XIX.2.3 in [51]), the spectrum σ(A− ) of A− is contained in the closed half plane λ ≥ −ω, whereas σ(A+ ) is a subset of λ ≤ ω. In particular, the strip |λ| < −ω is contained in ρ(A), the resolvent set of A. This justiﬁes the use of the term “separating projection” for P . It is convenient to adopt the following notation and terminology. Suppose A(X → X) is an exponentially dichotomous operator with separating projection P , and let A− and A+ be as above. Thus A− and A+ are the restrictions of A to X− = Im P and X+ = Ker P , respectively. With A we associate a function E(·; A) with domain R \ {0} and with values in L(X), the space of all bounded operators on X. The deﬁnition is as follows: for x ∈ X, ⎧ ⎨ −etA− P x, t < 0, E(t; A)x = (5.7) ⎩ etA+ (I − P )x, t > 0,

80

Chapter 5. Factorization of matrix functions analytic in a strip

where, following standard conventions, etA− denotes the value at t(< 0) of the semigroup generated by A− and etA+ denotes the value at t(> 0) of the semigroup generated by A+ . We call E(· ; A) the bisemigroup generated by A. The operator A will be referred to as the bigenerator of E(· ; A). For each x ∈ X the function E(t; A)x is continuous on R \ {0}, and lim E(t; A)x = −P x, t↑0

lim E(t; A)x = (I − P )x. t↓0

(5.8)

We conclude that E(· ; A) is an exponentially decaying operator function which is strongly continuous on the real line, except at the origin where it has (at worst) ˙ D(A+ ), the function E(t; A)x is a jump discontinuity. For x ∈ D(A) = D(A− ) + even diﬀerentiable on R \ {0}. In fact, we have d E(t; A)x = AE(t; A)x = E(t; A)Ax, dt

t = 0.

Obviously the derivative of E(· ; A)x is continuous on R \ {0}, exponentially decaying (in both directions) and has (at worst) a jump discontinuity at the origin. From (5.7) it is clear that E(t, A)P = P E(t, A) = E(t; A),

t < 0,

E(t, A)(I − P ) = (I − P )E(t, A) = E(t; A),

t > 0.

Also the following semigroup properties hold: E(t + s, A) =

−E(t; A)E(s; A),

t, s < 0,

E(t + s, A) =

E(t, A)E(s; A),

t, s > 0.

One of the reasons for the diﬀerent signs to appear in the deﬁnition of E(t; A) is that in this way the following identity holds: ∞ −1 e−λt E(t; A)x dt, x ∈ X, |λ| < −ω. (5.9) (λ − A) x = −∞

Here ω is a negative constant such that A is of exponential type ω. The proof of (5.9) is based on standard semigroup theory (see, e.g., Theorem XIX.2.2 in [51]). With the help of (5.8) and (5.9) we now can prove the uniqueness of the separating projection. Proposition 5.1. Let A(X → X) be an exponentially dichotomous operator. Then A has precisely one separating projection. Proof. Let P be a separating projection for A, and let E(· ; A) be the associate bisemigroup. A priori E(· ; A) depends not only on A but also on P . However, (5.9) and the fact that E(· ; A) is strongly continuous on R \ {0} imply that E(· ; A) is uniquely determined by A. On the other hand the ﬁrst identity in (5.8) shows that P is uniquely determined by E(· ; A). So along with E(· ; A) the separating projection is uniquely determined by A.

5.2. Spectral splitting and proof of Theorem 5.2

81

From (5.9) it follows that on a strip around the imaginary axis, the resolvent (λ − A)−1 of A is the pointwise two-sided Laplace transform of an exponentially decaying operator function which is strongly continuous on R \ {0} and has (at worst) a jump discontinuity at zero. The following theorem shows that this property characterizes exponentially dichotomous operators. Theorem 5.2. Let A(X → X) be a densely deﬁned closed linear operator on the complex Banach space X. Then A is exponentially dichotomous if and only if the imaginary axis is contained in the resolvent set of A and ∞ −1 (λ − A) x = e−λt E(t)x dt, x ∈ X, λ = 0, (5.10) −∞

where E : R \ {0} → L(X) is exponentially decaying and strongly continuous, and E has (at worst) a jump discontinuity at zero. In that case the function E is the bisemigroup generated by A. The above theorem will play an important role in Section 5.5. For the sake of completeness its proof is given in the next section. The reader who is ready to accept Theorem 5.2 may proceed directly to Section 5.3.

5.2 Spectral splitting and proof of Theorem 5.2 In this section we prove Theorem 5.2. The proof will be based on the spectral splitting results proved in Section XV.3 of [51], which originate from [16]. It will be convenient ﬁrst to prove the following result which is the semigroup version of Theorem 5.2. Theorem 5.3. Let S(X → X) be a densely deﬁned closed linear operator on the complex Banach space X. Then S is the inﬁnitesimal generator of a strongly continuous right semigroup of negative exponential type if and only if the imaginary axis is contained in the resolvent set of S and ∞ (λ − S)−1 x = e−λt E(t)x dt, x ∈ X, λ = 0, (5.11) 0

where E : [0, ∞) → L(X) is exponentially decaying and strongly continuous. In that case the function E is the right semigroup generated by S. Proof. The “only if part” of Theorem 5.2 is immediate from standard semigroup theory. To prove the “if part” let ω be a negative real number and L a positive constant such that E(t) ≤ Leωt , t ≥ 0. (5.12) For λ > ω and x ∈ X, put

R(λ)x = 0

∞

e−λt E(t)x dt.

(5.13)

82

Chapter 5. Factorization of matrix functions analytic in a strip

Then R(λ) is a well-deﬁned bounded linear operator on X with norm not exceeding L. The function R is pointwise analytic on λ > ω, and hence it is analytic on λ > ω. We shall prove that λ > ω implies that λ ∈ ρ(S) and R(λ) = (λ − S)−1 . Let T = S −1 be the (bounded) inverse of S. For 0 = λ ∈ ρ(S), one has −1 λ ∈ ρ(T ) and (λ − S)−1 = −λ−1 T (λ−1 − T )−1 . Take λ on the imaginary axis, λ = 0. Combining (5.11) and (5.13) we get R(λ) = (λ − S)−1 = −λ−1 T (λ−1 − T )−1 , and hence R(λ) = λR(λ) − I T . But then the unicity theorem for analytic functions gives that these identities hold on all of λ > ω. A simple computation now shows that R(λ) = (λ − S)−1 for each λ with λ > ω. We have seen that the open half plane λ > ω is contained in ρ(S) and (λ − S)−1 x =

∞

e−λt E(t)x dt,

x ∈ X, λ > ω.

0

(5.14)

Diﬀerentiating the left-and right-hand side of (5.14) for the variable λ, one ﬁnds −n

(λ − S)

(−1)n x= (n − 1)!

∞

tn−1 e−λt E(t)x dt,

0

x ∈ X, λ > ω.

(5.15)

Here n is an arbitrary positive integer. Taking λ > ω and combining (5.12) and (5.15) we get the estimate (λ − S)−n x ≤ Observe that ∞ t 0

n−1 −(λ−ω)t

e

L (n − 1)!

1 dt = (λ − ω)n

∞

0

tn−1 e−(λ−ω)t dt x.

∞

sn−1 e−s ds =

0

(n − 1)! . (λ − ω)n

Thus (λ − S)−n ≤ L(λ − ω)−n for real λ > ω and n = 1, 2, . . . . The HilleYosida-Phillips theorem ([51], page 419) now guarantees that S is the generator of a strongly continuous right semigroup T (t) of exponential type ω < 0. But then (5.11) holds with E(t) replaced by T (t). As the operator-valued functions E) and T are both strongly continuous, they must coincide, and the proof is complete. Proof of Theorem 5.2. We split the proof into three parts. Throughout ω is a negative real number and L is a positive constant such that E(t) ≤ Leω|t| ,

0 = t ∈ R.

(5.16)

Part 1. In this part we show that (λ − A)−1 is well-deﬁned and uniformly bounded

5.2. Spectral splitting and proof of Theorem 5.2

83

on each closed strip |λ| ≤ h where 0 < h < −ω. To do this, let us consider the following expressions: ∞ e−λt E(t)x dt, λ > ω, Ψ+ (λ)x = Ψ− (λ)x

0

0

e−λt E(t)x dt,

= −∞

λ < −ω.

Here x ∈ X. Clearly Ψ+ (λ) is a well-deﬁned bounded linear operator on X which depends analytically on λ on the open half plane λ > ω, and an analogous statement holds of course for Ψ− (λ). Note that Ψ− (λ) + Ψ+ (λ) is analytic on the strip |λ| < −ω and coincides on the imaginary axis with (λ − A)−1 . Thus |λ| < −ω implies λ ∈ ρ(A) and (λ − A)−1 = Ψ− (λ) + Ψ+ (λ), i.e., ∞ e−λt E(t)x dt, x ∈ X, |λ| < −ω. (5.17) (λ − A)−1 x = −∞

A detailed argument can be given along the lines indicated in the second paragraph of the proof of Theorem 5.3. From (5.16) one easily deduces that Ψ+ (λ)

≤

L , λ − ω

λ > ω,

Ψ− (λ)

≤

−L , λ + ω

λ < −ω.

On the strip |λ| < −ω, the norm of (λ − A)−1 = Ψ− (λ) + Ψ+ (λ) can now be estimated as follows: (λ − A)−1 ≤

−2Lω , ω 2 − (λ)2

|λ| < −ω.

(5.18)

In particular (λ − A)−1 is uniformly bounded on each closed strip |λ| ≤ h with 0 < h < −ω. Part 2. Fix 0 < h < −ω. From what has been proved in the previous part, we know that sup (λ − A)−1 < ∞. (5.19) | λ|≤h

This allows us to use the spectral theory developed in Section XV.3 of [51]. First we introduce the operators −α+i∞ 1 Q− = λ−2 (λ − A)−1 dλ, 2πi −α−i∞ Q+ =

−1 2πi

α+i∞

α−i∞

λ−2 (λ − A)−1 dλ.

84

Chapter 5. Factorization of matrix functions analytic in a strip

Here 0 < α < h, and hence (5.19) implies that Q− and Q+ are well-deﬁned bounded linear operators on X. It can be proved that these operators do not depend on the particular choice of α; nevertheless, in what follows we keep α ﬁxed. (Notice that in Section XV.3 of [51] the operators Q− and Q+ are denoted by S− and S+ , respectively.) We deﬁne M− = Im Q− ,

M+ = Im Q+ .

Put T = A−1 . Then T is a bounded linear operator on X commuting with (λ−A)−1 for each λ in the strip |λ| ≤ h. It follows that T commutes with Q− and Q+ . Since T is bounded, this implies that T M− ⊂ M− and T M+ ⊂ M+ . We also know that Im T = D(A), and thus T M− and T M+ belong to D(A). This allows us to deﬁne operators A− (M− → M− ) and A+ (M+ → M+ ) by setting D(A− ) = T M−,

A− x = Ax,

x ∈ D(A− ),

D(A+ ) = T M+,

A+ x = Ax,

x ∈ D(A+ ).

In other words,

−1 A− = T |M− ,

−1 A+ = T |M+ .

The ﬁrst part of Lemma XV.3.3 in [51] shows that A− and A+ are closed and densely deﬁned linear operators, and their spectra satisfy the inclusion relations σ(A− ) ⊂

{λ ∈ C | λ ≤ −h},

σ(A+ ) ⊂

{λ ∈ C | λ ≥ h}.

We shall now prove that −1

(λ − A− )

−1

(λ − A+ )

x

x

∞

=

e−λt E(t)x dt,

x ∈ M− , Reλ > −h,

(5.20)

e−λt E(t)x dt,

x ∈ M+ , Reλ < h.

(5.21)

0 0

= −∞

Following Section XV.3, page 330, of [51], we introduce two auxiliary sets N− and N+ . By deﬁnition N− is the set of all vectors x ∈ X for which there exists an X-valued function ϕ− x , bounded and analytic on λ > −h, which takes its values in D(A) and satisﬁes (λ − A)ϕ− x (λ) = x,

λ > −h.

Roughly speaking, N− consists of all vectors x ∈ X such that (λ − A)−1 x has a bounded analytic continuation to the open half plane λ > −h. The function ϕ− x (assuming it exists) is uniquely determined by x. Analogously, we let N+ be the

5.2. Spectral splitting and proof of Theorem 5.2

85

set of all vectors x ∈ X for which there exists an X-valued function ϕ+ x , bounded and analytic on λ < h, which takes its values in D(A) and satisﬁes (λ − A)ϕ+ x (λ) = x,

λ < h.

Also ϕ+ x is unique, provided it exists. Obviously, the sets N− and N+ are (possibly non-closed) linear manifolds of X. The second part of Lemma XV.3.3 in [51] states that D(A2− ) ⊂ N− and 2 D(A+ ) ⊂ N+ . Now, ﬁx x ∈ D(A2− ). Then x ∈ N− , and hence (λ − A)−1 x extends to a bounded analytic function on λ > −h. Notice that Ψ+ (λ) is also bounded and analytic on λ > −h. Recall that Ψ− (λ) is equal to (λ − A)−1 − Ψ+ (λ) for each λ in the strip |λ| < ω. It follows that Ψ− (λ)x extends to a bounded analytic function on λ > −h. On the other hand Ψ− (λ)x is analytic on λ < −ω and bounded on λ ≤ h. Hence Ψ− (λ)x determines a bounded entire function. From the estimate given for Ψ− (λ) in the previous part, it is clear that lim

λ∈R, λ→−∞

Ψ− (λ)x = 0.

But then we can use Liouville’s theorem to show that Ψ− (λ)x vanishes identically. We conclude that 0 e−λt E(t)x dt = 0, λ < −ω. −∞

Since E(t)x is continuous on −∞ < t < 0, it follows that E(t)x = 0 for all negative real numbers t. −1 is densely Now recall that T |M− is one-to-one and that A− = T |M− deﬁned. This implies that 2 D(A2− ) = Im T |M− , and that D(A2− ) is dense in M− . Thus the result of the previous paragraph shows that E(t) vanishes on M− for −∞ < t < 0. For x ∈ M− and |λ| < h, we have −1 (T |M− )x = (λ − A− )−1 x. (λ − A)−1 x = −(I − λT )−1 T x = − I − λ(T |M− ) Hence, for x ∈ M− and |λ| < h, −1

(λ − A− )

−1

x = (λ − A)

∞

x=

e −∞

−λt

∞

E(t)x dt =

e−λt E(t)x dt.

0

By analytic continuation this proves (5.20). Formula (5.21) is proved in a similar manner. Part 3. In this part we complete the proof. First we show that for t > 0 the operator E(t) maps M− into M− . To see this, take x ∈ M− , and let f be a

86

Chapter 5. Factorization of matrix functions analytic in a strip

continuous linear functional on X annihilating M− . Then f (λ − A− )−1 x = 0 for λ > −h, and thus (5.20) yields ∞ e−λt f E(t)x dt = 0, λ > −h.

0

This implies that f E(t)x = 0 for t > 0, and so, by the Hahn-Banach theorem, E(t)x ∈ M− for t > 0. Thus E(t)M− ⊂ M− for t > 0. The result of the previous paragraph enables us to deﬁne an operator-valued function E− : (0, ∞) → L(M− ) by stipulating that E− (t) = E(t)|M− . Our assumptions on the behavior of E near the origin (together with the BanachSteinhaus theorem) imply that E− can be extended to a strongly continuous function, deﬁned on 0 ≤ t < ∞, by putting x ∈ M− .

E− (0)x = lim E(t)x, t↓0

The identity (5.20) can now be written as ∞ e−λt E− (t) dt, (λ − A− )−1 x =

x ∈ M− , λ > −h.

0

Since A− (M− → M− ) is closed and densely deﬁned, it follows from Theorem 5.3 that E− is a strongly continuous right semigroup, and that A− is its inﬁnitesimal generator. In the same way one proves that E(t)M+ ⊂ M+ for t < 0, and we deﬁne E+ : (−∞, 0] → L(M+ ) by setting E+ (t) = −E(t)|M+ ,

E+ (0)x = lim −E(t)x, t↑0

x ∈ M+ .

Then the analogue of Theorem 5.3 for left semigroups shows that E+ is a strongly continuous left semigroup which has A+ (M+ → M+) as its generator. Next, consider the operator P on X deﬁned by P x = lim −E(t)x, t↑0

x ∈ X.

By the Banach-Steinhaus theorem, P is a bounded linear operator on X. For t < 0 we have that E(t) vanishes on M− , and so P x = 0 for each x ∈ M− . For x ∈ M+ and t < 0 we have E(t)x = −E+ (t)x, and thus P x = x. These properties of P imply that M− ∩ M+ = {0} and M− + M+ is closed. (5.22) The ﬁrst part of (5.22) is obvious. To prove the second part, let x1 , x2 , . . . be a sequence in M− , let y1 , y2 , . . . be a sequence in M+ , and assume that xn + yn → z for n → ∞. It suﬃces to show that z ∈ M− + M+ . Since P is continuous on X and P is zero on M− , we have P z = lim P (xn + yn ) = lim P yn . n→∞

n→∞

5.3. Realization triples

87

But P yn = yn ∈ M+ and M+ is closed. Thus P z ∈ M+ . Moreover, yn = P yn converges to P z if n → ∞. Thus xn = (xn +yn)−yn converges to z −P z if n → ∞. Also, M− is closed. We conclude that z − P z ∈ M− , and hence z = z − P z + P z belongs to M− + M+ . So M− + M+ is closed. Finally, the ﬁrst part of the proof of Theorem XV.3.1 in [51] shows that ˙ + , and that P is the M− + M+ is dense in X. We conclude that X = M− +M projection of X along M− onto M+ . Recall that D(A) = Im T , where T = A−1 . It follows that ˙ M+ ) = T M−+ ˙ T M+ = D(A− )+ ˙ D(A+ ). D(A) = T X = T (M− + Hence P maps D(A) into itself, and P commutes with A. Thus relative to the decompositions ˙ M+ , X = M− +

˙ D(A+ ), D(A) = D(A− ) +

the operator A admits the partitioning A− A= 0

0 A+

.

Therefore A is an exponentially dichotomous operator, P is the separating projection for A, and E(·) = E(· ; A).

5.3 Realization triples In this section we introduce the realizations that will be used to obtain representations of the type (5.3). We begin with some additional notation. m m By Dm 1 (R) we denote the linear submanifold of L1 (R) = L1 (R, C ) consistm m ing of all f ∈ L1 (R) for which there exists g ∈ L1 (R) such that ⎧ t ⎪ ⎪ ⎪ g(s) ds, a.e. on (−∞, 0), ⎪ ⎨ −∞ f (t) = (5.23) ∞ ⎪ ⎪ ⎪ ⎪ g(s) ds, a.e. on (0, ∞). ⎩ t

Dm 1 (R),

then there is only one g ∈ Lm If f ∈ 1 (R) such that (5.23) holds. This g is called the derivative of f and is denoted by f . From (5.23) it follows that f (0+) = limt↓0 f (t) and f (0−) = limt↑0 f (t) exist; in fact, ∞ 0 f (0+) = g(s) ds, f (0−) = g(s) ds. 0

−∞

Let ω be a negative constant. A triple Θ = (A, B, C) of operators is called a realization triple of exponential type ω if the following conditions are satisﬁed:

88

Chapter 5. Factorization of matrix functions analytic in a strip

(C1) −iA is an exponentially dichotomous operator of exponential type ω with domain D(A) and range in a Banach space X; (C2) B : Cm → X is a linear operator; (C3) C is a possibly unbounded operator with domain D(C) in X and range in Cm such that D(A) ⊂ D(C) and C is A-bounded; (C4) there exists a linear operator ΛΘ from X into Lm 1 (R) such that ∞ e−ω|t| (ΛΘ x)(t) dt < ∞, (i) sup x ≤1

−∞

(ii) for every x ∈ D(A) we have (ΛΘ x)(t) = iCE(t; −iA)x, t ∈ R, and the function ΛΘ x belongs to Dm 1 (R). In (ii), the function E(t; −iA) is the bisemigroup generated by −iA. Note that B, being a linear operator from Cm into X, is automatically bounded. Observe also that (i) implies that ΛΘ is bounded and maps X into Lm 1,ω (R) where Lm 1,ω (R) =

−ω|·| f ∈ Lm f (·) ∈ Lm 1 (R) | e 1 (R) .

(5.24)

Taking into account (ii) and the fact that D(A) is dense in X, one sees that ΛΘ is uniquely determined. Since ω is negative, Lm 1,ω (R) given by (5.24) is a linear manifold in Dm 1 (R). The space X is called the state space and the space Cm the input/output space of the triple. We shall refer to A as the main operator of the triple. Suppose Θ is a realization triple of exponential type ω and ω ≤ ω1 < 0. Then Θ is a realization triple of exponential type ω1 too. To see this, note that (i) and (ii) are fulﬁlled with ω replaced by ω1 . When the actual value of ω is not relevant, we simply call Θ a realization triple. Thus Θ = (A, B, C) is a realization triple if Θ is a realization triple of exponential type ω for some ω < 0. The operator ΛΘ does not depend on the value of ω, and the same is true with regard to the separating projection for −iA. This projection will be denoted by PΘ , although it is deﬁned in terms of A alone. The case when C is a bounded linear operator from X into Cm is of special interest. In that case C is obviously A-bounded, and (C4) is fulﬁlled with ΛΘ x = iCE(·, −iA)x for each x ∈ X. Thus when C is bounded, then conditions (C3) and (C4) are automatically satisﬁed. Let Θ = (A, B, C) be a realization triple with state space X. Notice that item (i) in (C4) implies that ΛΘ : X → Lm 1 (R) is a bounded linear operator. Since (ii) prescribes ΛΘ on D(A), the boundedness of ΛΘ and the density of D(A) in X imply that ΛΘ is uniquely determined by the operators A and C. The operator ΛΘ plays the role of the observability operator in systems theory. For its dual analogue (the controllability operator) we refer to the following proposition.

5.3. Realization triples

89

Proposition 5.4. Suppose Θ = (A, B, C) is a realization triple of exponential type ω < 0, and let ΓΘ : Lm 1 (R) → X be deﬁned by ∞ E(−t; −iA)Bϕ(t) dt, ϕ ∈ Lm (5.25) ΓΘ ϕ = 1 (R). −∞

Then ΓΘ is a bounded linear operator, and ΓΘ maps Dm 1 (R) into D(A). Proof. The operator function E(· ; −iA) is strongly continuous. Now recall the following well-known fact: if a sequence of operators converges in the strong operator topology, then the convergence is uniform on compact subsets of the underlying space. Because of the ﬁnite dimensionality of Cm , the operator B is of ﬁnite rank, hence compact. It follows that the function E(· ; −iA)B is continuous on R \ {0} with a possible jump at the origin where continuity is taken with respect the operator norm. It follows that the integral in (5.25) is well-deﬁned for each ϕ ∈ Lm 1 (R), and that ΓΘ is a bounded linear operator. Now ﬁx ϕ ∈ Dm 1 (R). For simplicity we restrict ourselves to the case when ϕ vanishes almost everywhere on (−∞, 0). By our assumption on ϕ there exists ψ ∈ Lm 1 (R) such that ∞ ϕ(t) = − ψ(s) ds, t > 0. t

But then

ΓΘ ϕ

= − = − = −

∞ 0

0

E(−t; −iA)B

∞ ∞ t

∞ s

0

0

∞

ψ(s) ds dt

t

E(−t; −iA)Bψ(s) ds dt

E(−t; −iA)Bψ(s) dt ds.

The last equality follows by applying Fubini’s theorem. Since A is exponentially dichotomous, zero belongs to the resolvent set of A. So it makes sense to consider the operator iE(−t; −iA)A−1 B. This function is diﬀerentiable on [0, ∞), and its derivative is the continuous operator-valued function −E(−t; −iA)B. Here diﬀerentiation and continuity are taken with respect to the operator norm which we can use because of the compactness of B. Thus s E(−t; −iA)B dt = iE(−s; −iA)A−1B − iPΘ A−1 B, − 0

where PΘ is the separating projection of −iA. Hence ∞ iE(−s; −iA)A−1 B − iPΘ A−1 B ψ(s) ds ΓΘ ϕ = 0

=

A−1

0

∞

iE(−s; −iA)B − iPΘ B ψ(s) ds .

90

Chapter 5. Factorization of matrix functions analytic in a strip

This shows that ΓΘ ϕ belongs to Im A−1 = D(A).

Let Θ = (A, B, C) be a realization triple of exponential type ω < 0 and having input/output space Cm . With Θ we associate two m × m matrix functions. These functions will be denoted by kΘ and WΘ , and they are called the kernel function associated with Θ and the transfer function of Θ, respectively. The ﬁrst of these is deﬁned as follows. For every u in Cm , we have that ΛΘ Bu belongs to Lm 1,ω (R). Thus the expression u ∈ Cm , (5.26) kΘ (.)u = ΛΘ Bu (.), determines a unique element kΘ of Lm×m 1,ω (R), that is each column of kΘ belongs m m m to Lm (R). In fact k (.)u ∈ L (R) ⊂ Dm Θ 1,ω 1,ω 1 (R) ⊂ L1 (R) for each u ∈ C . Next let us turn to WΘ . This function is given by WΘ (λ) = I + C(λ − A)−1 B,

|λ| < −ω.

(5.27)

To see that WΘ is well-deﬁned, ﬁx λ in the resolvent set ρ(A) of A. Since the operator (λ − A)−1 maps X into the domain D(A) of A, and D(A) is contained in the domain of C, the product C(λ − A)−1 is well-deﬁned. Hence C(λ − A)−1 B is a well-deﬁned linear transformation on Cn . The fact that −iA is an exponentially dichotomous operator of exponential type ω implies that |λ| < −ω is contained in ρ(−iA), and thus |λ| < −ω is contained in ρ(A). We conclude that WΘ is a well-deﬁned analytic m × m matrix function on |λ| < −ω. The next proposition explains the relation between the two functions WΘ and kΘ . Proposition 5.5. Suppose Θ = (A, B, C) is a realization triple of exponential type ω < 0. Then ∞ WΘ (λ) = I − eiλt kΘ (t) dt, |λ| < −ω. (5.28) −∞

Proof. It suﬃces to show that for x ∈ X and |λ| < −ω we have ∞ −1 eiλt (ΛΘ x)(t) dt, C(λ − A) x = −

(5.29)

−∞

that is, −C(λ − A)−1 x is equal to the Fourier transform (Λ Θ x)(λ) of ΛΘ x. In what follows λ is ﬁxed subject to |λ| < −ω. We already know that C(λ − A)−1 is a well-deﬁned map from X into Cm . Obviously, this map is linear. To show that it is also bounded, take x ∈ X. Using the fact that C is A-bounded, there exists a constant M such that C(λ − A)−1 x ≤ M (λ − A)−1 x + A(λ − A)−1 x . Now A(λ − A)−1 x = −x + λ(λ − A)−1 x. Thus C(λ − A)−1 x ≤ M (λ − A)−1 + 1 + |λ|(λ − A)−1 x.

5.4. Construction of realization triples

91

It follows that C(λ − A)−1 is a bounded linear operator from X into Cm . m Now consider the map x → (Λ Θ x)(λ) from X into C . This map is linear and bounded too. Linearity is obvious. Boundedness follows from the estimate ∞ x)(λ) ≤ e−ω|t| ΛΘ x(t) dt, (Λ Θ −∞

together with condition (i) in the deﬁnition of a realization triple. We have now shown that, for λ ﬁxed, both sides of (5.29) are continuous in x. Hence it suﬃces to prove (5.29) for x ∈ D(A) because of D(A) = X. Take x ∈ D(A), and put y = Ax. Since −iA is an exponentially dichotomous operator of exponential type ω, we use (5.9) for −iA in place of A and −iλ in place of λ to show that ∞ eiλt E(t; −iA)y dt. (5.30) (λ − A)−1 y = −i −∞

Recall that CA−1 is a bounded linear operator. It follows that C(λ − A)−1 x = =

CA−1 (λ − A)−1 y ∞ −i eiλt CA−1 E(t; −iA)y dt −∞

=

−i

−∞

=

−i

∞

∞

−∞

eiλt CE(t; −iA)x dt eiλt (ΛΘ x)(t) dt,

the latter equality holding by virtue of condition (ii) in the deﬁnition of a realization triple. Thus (5.29) is proved. From (5.29) it follows that C(λ − A)−1 is analytic on |λ| < −ω. This result can also be proved directly using that C is A-bounded. In fact, employing the C-boundedness of A one can show that the function λ → C(λ − A)−1 is analytic on the resolvent set ρ(A).

5.4 Construction of realization triples In this section we construct a representation of the form (5.3) for the m × m matrix-valued function W in (5.1) with the kernel function k being given by (5.2). The following theorem is the main result. Theorem 5.6. An m×m matrix function W is the transfer function of a realization triple if and only if W is of the form ∞ eiλt k(t) dt, (5.31) W (λ) = I − −∞

92

Chapter 5. Factorization of matrix functions analytic in a strip

where k is an m × m matrix function with the property that there exist ω < 0 and h ∈ Lm×m (R) such that 1 (5.32) k(t) = eω|t| h(t). If W is given by (5.31) and (5.32) for some ω < 0 and h ∈ Lm×m (R), then 1 W = WΘ with Θ = (A, B, C) constructed in the following way: the state space X m of Θ is Lm 1 (R), the input/output space is C , D(A) = D(C) = Dm 1 (R), −iωf (t) + if (t), (Af )(t) = iωf (t) + if (t), (By)(t) = e−ω|t| k(t)y, ∞ f (s) ds. Cf = i

a.e. on −∞ < t < 0, a.e. on 0 < t < ∞,

a.e. on R,

−∞

Proof. Let Θ be a realization triple, and let W = WΘ be its transfer function. Then, by Proposition 5.5 in the preceding section, (5.31) holds with k = kΘ . Using the fact that the second operator in a realization triple is bounded, we see from (i) in the deﬁnition of a realization triple that ∞ e−ω|t| kΘ (t)y dt < ∞, sup y ≤1

−∞

for some ω < 0. Hence k = kΘ satisﬁes (5.32). This proves the “if part” of the theorem. Next, let W be given by (5.31) and (5.32) for some ω < 0 and h ∈ Lm×m (R), 1 and let Θ = (A, B, C) be the triple of operators deﬁned in the second part of the theorem. We need to show that this triple is a realization triple and that W = WΘ . As is well-known (cf., [51], page 420), the backward translation semigroup on Lm 1 [0, ∞) is strongly continuous. The inﬁnitesimal generator of this semigroup has Dm 1 [0, ∞) as its domain and its action amounts to taking the derivative. Here Dm [0, ∞) is the linear manifold consisting of all functions f ∈ Dm 1 1 (R) with the property that f (t) = 0 for t < 0, and hence the derivative f is well-deﬁned for each f ∈ Dm 1 [0, ∞). Using this, one sees that −iA an exponentially dichotomous operator of exponential type ω and that the bisemigroup associated with −iA acts as follows: for t < 0, −e−ωt f (t + s), a.e. on −∞ < s < 0, E(t; −iA)f (s) = 0, a.e. on 0 < s < ∞, and for t > 0, E(t; −iA)f (s) =

0,

a.e. on −∞ < s < 0,

eωt f (t + s),

a.e. on 0 < s < ∞.

5.5. Inverting matrix functions analytic in a strip

93

The separating projection for −iA is the projection of the state space X = Lm 1 (R) m onto Lm (−∞, 0] along L [0, ∞). 1 1 Condition (5.32) on k implies that the operator B from Cm into Lm 1 (R) is bounded. From the deﬁnition of C and A we see that Cf ≤ −ωf + Af ,

f ∈ D(A).

Thus C is A-bounded. Deﬁne Λ : X → Lm 1 (R) by (Λf )(t) = eω|t| f (t),

a.e. on R.

(5.33)

Then Λ satisﬁes the conditions (i) and (ii) in the deﬁnition of a realization triple with Λ in place of ΛΘ . For (i) this is obvious. To check the ﬁrst part of (ii), one uses the above description of the bisemigroup E(t; −iA) and the deﬁnition of C. As to the second part of (ii), observe that f ∈ Dm 1 (R) and ω < 0 imply that the function eωt f (t) belongs to Dm 1 (R) too. We have now proved that Θ = (A, B, C) is a realization triple We claim that the kernel function kΘ associated with Θ coincides with k. Indeed, for y ∈ Cm the following identities hold almost everywhere on R: kΘ (t)y = (ΛBy)(t) = (eω|t| By)(t) = k(t)y. Since Cm has a ﬁnite basis, it follows that kΘ (t) = k(t) almost everywhere on R. In other words, kΘ and k coincide as elements of Lm×m (R). 1

5.5 Inverting matrix functions analytic in a strip Let Θ = (A, B, C) be a realization triple with state space X. In this section we shall employ the operator A× (X → X). Here is the deﬁnition: the domain of A× is equal to the domain of A, and its action is deﬁned by A× = A− BC. We call A× the associate main operator of the triple Θ. As one may expect from Section 2.4, the operator A× plays an important role in inverting WΘ (λ). In fact, we have the following theorem. Theorem 5.7. Let the m × m matrix function W be given by W (λ) = I + C(λ − A)−1 B, with Θ = (A, B, C) being a realization triple. Let A× be the associate main operator of Θ. Then W (λ) is invertible for each λ ∈ R if and only if the spectrum of A× does not intersect the real line. In that case (A× , B, −C) is a realization triple, and W (λ)−1 = I − C(λ − A× )−1 B, λ ∈ R, (5.34) W (λ)C(λ − A× )−1 = C(λ − A)−1 , × −1

(λ − A ) × −1

(λ − A )

−1

= (λ − A)

−1

BW (λ) = (λ − A) −1

− (λ − A)

−1

BW (λ)

B,

λ ∈ R,

(5.35)

λ ∈ R,

(5.36)

−1

C(λ − A)

,

λ ∈ R. (5.37)

94

Chapter 5. Factorization of matrix functions analytic in a strip

Proof. We split the proof into four parts. In the ﬁrst part we show that W (λ) is invertible for each λ ∈ R if and only if the spectrum of A× does not intersect the real line, and we derive the expressions (5.34) – (5.37). The remaining three parts are concerned with the statement that (A× , B, −C) is a realization triple. Part 1. Suppose A× has no spectrum on the real line. This condition means that for each real λ the linear operator λ − A× maps D(A× ) = D(A) in a one-one way onto X, and hence the linear operator I − C(λ − A× )−1 B acting on Cm is well-deﬁned. We claim that it is the inverse of W (λ). To see this we ﬁrst prove (5.35). From BCx = (A − A× )x for each x ∈ D(A), it follows that BC(λ − A)−1 = (A − A× )(λ − A)−1 . Using the latter identity and ﬁxing λ ∈ R, we obtain the equality (5.35) from the following calculation: W (λ)C(λ − A× )−1 = C(λ − A× )−1 + C(λ − A)−1 BC(λ − A)−1 = C(λ − A× )−1 + C(λ − A)−1 (A − A× )(λ − A)−1 = C(λ − A× )−1 + C(λ − A)−1 (A − λ) + (λ − A× ) (λ − A)−1 = C(λ − A)−1 . From (5.35) we obtain that W (λ) I − C(λ − A× )−1 B = W (λ) − C(λ − A)−1 = I,

λ ∈ R.

Hence W (λ) is invertible for each λ ∈ R. Next assume W (λ) is invertible for each λ ∈ R. We claim that A× has no spectrum on the real line and that (5.37) holds. To prove this, ﬁx λ ∈ R and let R(λ) be the operator on X deﬁned by the right-hand side of (5.37). Since (λ − A× )(λ − A)−1 = I + BC(λ − A)−1 , we have (λ − A× )R(λ)

=

I + BC(λ − A)−1 + − I − BC(λ − A)−1 BW (λ)−1 C(λ − A)−1

=

I + BC(λ − A)−1 +B − I − C(λ − A)−1 B W (λ)−1 C(λ − A)−1

=

I + BC(λ − A)−1 − BC(λ − A)−1 ,

and so (λ − A× )R(λ) = I. Thus to prove (5.37) it remains to show that λ − A× is one-to-one. Let x ∈ D(A× ) = D(A) and suppose (λ−A× )x = 0. Since A× x = Ax−BCx, we have (λ − A)−1 BCx = −x, and hence W (λ)Cx = Cx + C(λ − A)−1 BCx = Cx − Cx = 0.

5.5. Inverting matrix functions analytic in a strip

95

By assumption W (λ) is invertible. Therefore Cx = 0 and, consequently, (λ−A)x = (λ−A× )x = 0. Now use the fact that A has no spectrum on the real line. It follows that x = 0, and hence λ − A× is one-to-one. Note that in passing we established (5.34), (5.35) and (5.37). The argument for (5.36) is analogous to that for (5.35). In the remaining three parts it is assumed that A× has no spectrum on the real line, or equivalently, that W (λ) is invertible for each λ ∈ R. Part 2. We show that A× is closed and that C is A× -bounded. Applying (5.37) with λ = 0 we see that (A× )−1 = A−1 + A−1 BW (0)−1 CA−1 .

(5.38)

Since C is A-bounded, the operator CA−1 is bounded. Thus in the right-hand side of (5.38) the operators B, A−1 and CA−1 are all bounded. It follows that (A× )−1 is bounded too. Hence A× is a closed operator. Recall that the operators A−1 and (A× )−1 map X into D(A) = D(A× ). Since the latter space is contained in D(C), we can apply C to both sides of (5.38). This yields C(A× )−1 = CA−1 + CA−1 BW (0)−1 CA−1 . But CA−1 is bounded. Hence C(A× )−1 is bounded, which implies that C is A× bounded. Part 3. In this part we show that −iA× is exponentially dichotomous. To do this we apply Theorem 5.2. First some preparations. Recall that ∞ W (λ) = I − eiλt kΘ (t) dt, λ ∈ R, −∞

(R). By the matrix-valued version of with kΘ belonging to the space eω|·| Lm×m 1 Wiener’s theorem (see, e.g., [52], page 830), the fact that W (λ) is invertible for each λ ∈ R implies that ∞ W (λ)−1 = I − eiλt k × (t) dt, λ ∈ R, (5.39) −∞

(R). In fact (see [47], Section 18), taking |ω| smaller if necfor some k × ∈ Lm×m 1 essary we may assume that k × also belongs to eω|·| Lm×m (R). Next note that for 1 each x ∈ X and each y ∈ Cm , ∞ C(λ − A)−1 x = −i eiλt (ΛΘ x)(t) dt, λ ∈ R, −∞

(λ − A)−1 x

= −i

(λ − A)−1 By

= −i

∞

−∞

∞

−∞

eiλt E(t; −iA)x dt,

λ ∈ R,

eiλt E(t; −iA)By dt,

λ ∈ R;

96

Chapter 5. Factorization of matrix functions analytic in a strip

cf., (5.29) and (5.30). Using these formulas in (5.37), and taking inverse Fourier transforms, we see that ∞ (λ − A× )−1 x = −i eiλt E(t; −iA) + E1 (t) + E2 (t) x dt, λ ∈ R, −∞

where for each x ∈ X we have E1 (t)x = i E2 (t)x = −i

∞

−∞

∞ −∞

E(t − s; −iA)B(ΛΘ x)(s) ds,

E(t − s; −iA)B

∞

−∞

k × (s − r)(ΛΘ x)(r) dr

(5.40) ds.

(5.41)

Recall that the function E(· ; −iA)B is exponentially decaying, that k × belongs to eω|·|Lm×m (R), and that for each x ∈ X the function ΛΘ x belongs to eω|·| Lm 1 (R). 1 These facts imply that E1 and E2 are exponentially decaying too. Moreover, a routine argument shows that these functions are strongly continuous, that is, for each x ∈ X the functions E1 (·)x and E2 (·)x are continuous in the norm of X. We conclude that the function E(· ; −iA) + E1 (·) + E2 (·) is exponential decaying, strongly continuous on R \ {0}, and that at zero it has (at worst) a jump discontinuity. But then we can apply Theorem 5.2 with A replaced by −iA× and λ replaced by −iλ to show that −iA× is exponentially dichotomous. Furthermore, the bisemigroup generated by −iA× is given by E(· ; −iA× ) = E(· ; −iA) + E1 (·) + E2 (·),

(5.42)

where E(· ; −iA) is the bisemigroup generated by −iA, and the functions E1 (·) and E2 (·) are given by (5.40) and (5.41), respectively. Part 4. In this part we complete the proof and show Θ× = (A× , B, −C) is a realization triple. The negative constant ω having been taken suﬃciently close to zero, one has that Θ is of exponential type ω and k× belongs to eω|·| Lm×m (R). A 1 standard reasoning now shows that the convolution product k× ∗ ΛΘ x , given by ∞ × k ∗ (ΛΘ x) (t) = k × (t − s)(ΛΘ x)(s) ds, a.e. on R, −∞

determines a bounded linear operator from X into Lm 1 (R) such that ∞ e−ω|t| k× ∗ (ΛΘ x)(t) dt < ∞. sup x ≤1

−∞

But then the expression Λ× x = −ΛΘ x + k × ∗ (ΛΘ x)

(5.43)

deﬁnes a bounded linear operator Λ× : X → Lm 1 (R) for which condition (i) in the deﬁnition of a realization triple (Section 5.3), with ΛΘ replaced by Λ× , is satisﬁed.

5.5. Inverting matrix functions analytic in a strip

97

Next, take x ∈ X, and consider the Fourier transform of Λ× x. Using formula (5.43) we see that for each λ ∈ R we have × x)λ (Λ

× = −(Λ Θ x)(λ) + k (λ)(ΛΘ x)(λ)

=

I − k × (λ) (Λ Θ x)(λ)

= −W (λ)−1 C(λ − A)−1 x. In this calculation the ﬁnal equality results from (5.29) and (5.39). Next, using (5.35) we see that × x)(λ) = C(λ − A× )−1 x, (Λ

λ ∈ R.

(5.44)

Note that this equality actually holds in a strip |λ| < −ω containing the real line. Now take x ∈ D(A× ) = D(A), and put z = A× x. Then C(λ − A× )−1 x = × −1 C(A ) (λ − A× )−1 z, and the operator C(A× )−1 is bounded by the result of the second part of the proof. Since −iA× is exponentially dichotomous, by the third part of the proof, we can use formula (5.9), with A replaced by −iA× and by −iλ, to show that ∞ × x)(λ) = −iC(A× )−1 eiλt E(t; −iA× )z dt (Λ =

−i −i

∞

−∞

=

−∞

∞

−∞

eiλt C(A× )−1 E(t; −iA× )z dt eiλt CE(t; −iA× )x dt,

λ ∈ R.

Thus we have proved that (Λ× x)(t) = −iCE(t; −iA× )x almost everywhere on R. It remains to show that Λ× x ∈ Dm 1 (R). In view of the properties of ΛΘ and the identity (5.43), it suﬃces to show m that k × ∗ (ΛΘ x) belongs to Dm 1 (R). Since ΛΘ x = D1 (R), we can consider its derivative g, that is the function given by ⎧ t ⎪ ⎪ g(s) ds, a.e. on (−∞, 0), ⎪ ⎨ (ΛΘ x)(t) =

−∞

⎪ ⎪ ⎪ ⎩ −

t

∞

g(s) ds,

a.e. on (0, −∞).

Now use that (k × ∗ f ) = k × ∗ f + k × (· ) f (0+) − f (0−) ,

f ∈ Dm 1 (R).

98

Chapter 5. Factorization of matrix functions analytic in a strip

If follows that

k × ∗ (ΛΘ x) (t) =

⎧ ⎪ ⎪ ⎪ ⎨

t

h(s) ds,

a.e. on (−∞, 0),

−∞

⎪ ⎪ ⎪ ⎩ −

∞

h(s) ds

0, s < 0.

It follows that, almost everywhere on [0, ∞), 0 (L+ ψ)(t) = −iC1 E(t; −iA) E(−s; −iA)B1 ψ(s) ds = −iCE(t; −iA)

−∞ 0

−∞

E(−s; −iA)Bψ(s) ds

= (−QΛΘ ΓΘ ψ)(t), and (5.49) has been obtained for the case when Im B ⊂ D(A). The general situation, where Im B need not be contained in D(A), can be treated with an approximation argument based on the fact that B can be approximated (in norm) by bounded linear operators from Cm into X with ranges inside D(A). This is true because D(A) is dense in X and Cm is ﬁnite dimensional.

5.7 Inverting Wiener-Hopf integral operators In this section we study inversion of the Wiener-Hopf integral operator T : ∞ k(t − s)f (s) ds, a.e. on [0, ∞). (5.51) T f (t) = f (t) − 0

5.7. Inverting Wiener-Hopf integral operators

101

It will be assumed that the m × m matrix kernel function k is the kernel function associated with some realization triple. This implies that T is a well-deﬁned bounded linear operator on Lm 1 (R). We shall prove the following theorem. Theorem 5.10. Let T be the Wiener-Hopf integral operator on Lm 1 (R) given by (5.51). Assume that k = kΘ for some realization triple Θ = (A, B, C). Then T is invertible if and only if the following two conditions are satisﬁed: (i) Θ× = (A× , B, −C) is a realization triple, ˙ Ker PΘ× . (ii) X = Im PΘ + Here X is the state space of both Θ and Θ× , and PΘ and PΘ× are the separating projections of −iA and −iA× , respectively. If (i) and (ii) hold, the inverse of T is given by ∞ kΘ× (t − s)g(s) ds (T −1 g)(t) = g(t) − 0 ∞ ΛΘ× (I − Π)E(−s, −iA× )Bg(s) ds(t), a.e. on [0, ∞). − 0

Here Π is the projection of X onto Ker PΘ× along Im PΘ . To facilitate the proof of Theorem 5.10 we ﬁrst establish two lemmas. If Θ is a realization triple with main operator A, the separating projection of the operator −iA will be denoted by PΘ . Lemma 5.11. Let Θ = (A, B, C) and Θ× = (A× , B, −C) be realization triples with state space X. Then the operator J × : Im PΘ → Im PΘ× ,

J × x = PΘ× x,

(5.52)

˙ Ker PΘ× , and in that case is invertible if and only if X = Im PΘ + (J × )−1 = (I − Π)|Im PΘ× ,

Π = I − (J × )−1 PΘ× ,

(5.53)

where Π is the projection of X along Im PΘ onto Ker PΘ× . Proof. Obviously Ker J × = Im PΘ ∩ Ker PΘ× . Thus J × is one-to-one if and only if Im PΘ ∩ Ker PΘ× = {0}. Next, assume J × is surjective. Take x ∈ X. Then PΘ× x = J × PΘ z = PΘ× PΘ z for some z ∈ X. This yields x

= PΘ× x + (I − PΘ× )x = PΘ× PΘ z + (I − PΘ× )x = PΘ z + (I − PΘ× )(x − PΘ z).

Hence x ∈ Im PΘ + Ker PΘ× , and we conclude that Im PΘ + Ker PΘ× = X. Thus ˙ PΘ× provided that J × is invertible. Moreover, the above calcuX = Im PΘ +Ker lations show that × −1 PΘ× x = PΘ z = (I − Π)x = (I − Π)PΘ× PΘ z = (I − Π)PΘ× x, J

102

Chapter 5. Factorization of matrix functions analytic in a strip

which proves the ﬁrst identity in (5.53). ˙ Ker PΘ× . Then J × is injective. To complete the proof, assume X = Im PΘ + × To prove that J is surjective, take y ∈ Im PΘ× . Since PΘ× y = y and PΘ× Π = 0, we have y = PΘ× y = PΘ× (I − Π)y + PΘ× Πy = PΘ× (I − Π)y. Put x = (I − Π)y. Then x ∈ Im PΘ and J × x = y. This shows that J × is surjective, and thus J × is invertible. Moreover, we see that (J × )−1 y = x = (I − Π)y, which proves the second identity in (5.53). Lemma 5.12. Assume that Θ = (A, B, C) and Θ× = (A× , B, −C) are realization triples, with Cm being the input/output space of both Θ and Θ× . Introduce the maps m K : Lm 1 [0, ∞) → L1 [0, ∞),

∞

(Kϕ)(t) = 0 m K × : Lm 1 [0, ∞) → L1 [0, ∞),

(K × ϕ)(t) =

kΘ (t − s)ϕ(s) ds,

∞

0

a.e. on [0, ∞),

× kΘ (t − s)ϕ(s) ds,

U : Im PΘ× → Lm 1 [0, ∞),

(U x)(t) = (ΛΘ x)(t),

U × : Im PΘ → Lm 1 [0, ∞),

(U × x)(t) = −(ΛΘ× x)(t),

R : Lm 1 [0, ∞) → Im PΘ ,

Rϕ =

R × : Lm 1 [0, ∞) → Im PΘ× ,

R× ϕ = −

J : Im PΘ× → Im PΘ ,

Jx = PΘ x,

J × : Im PΘ → Im PΘ× ,

J × x = PΘ× x.

∞

0

a.e. on (−∞, 0]),

a.e. on [0, ∞), a.e. on [0, ∞),

E(−t; −iA)Bϕ(t) dt,

∞

0

E(−t; −iA× )Bϕ(t) dt,

Then all these operators are well-deﬁned, linear and bounded. Moreover,

I −K

U

R

J

I − K× R

×

m ˙ ˙ : Lm 1 [0, ∞) + Im PΘ× → L1 [0, ∞) + Im PΘ ,

U× J

×

m ˙ ˙ : Lm 1 [0, ∞) + Im PΘ → L1 [0, ∞) + Im PΘ× ,

5.7. Inverting Wiener-Hopf integral operators are bounded linear operators, which are invertible, and −1 I −K U I − K× U × . = R× J× R J

103

(5.54)

Proof. As we have seen in Section 5.6 the operators K and K × are bounded. To see that the other operators are well-deﬁned and bounded too it suﬃces to m make a few observations. Let Q be the projection of Lm 1 (R) onto L1 [0, ∞) along m L1 (−∞, 0]. Then U = QΛΘ |Im PΘ× ,

U × = −QΛΘ× |Im PΘ ,

and hence these two operators are well-deﬁned and bounded. Next, viewing PΘ and PΘ× as operators from X onto Im PΘ and Im PΘ× , respectively, we have , R = PΘ ΓΘ |Lm 1 [0,∞)

R× = −PΘ× ΓΘ× |Lm . 1 [0,∞)

From these identities and Proposition 5.4 it follows that R and R× are also welldeﬁned and bounded. It remains to prove (5.54). This amounts to checking eight identities. Pairwise these identities have analogous proofs. So, actually only four identities have to be taken care of. This will be done in the remaining part of the proof which is divided into four steps. Step 1. First we prove that R(I − K × ) + JR× = 0. Take ϕ in Lm 1 [0, ∞). We need to show that RK × ϕ = PΘ R× ϕ + Rϕ. Whenever this is convenient, it may be assumed that ϕ is a continuous function with compact support in (0, ∞). By applying Fubini’s theorem, one gets ∞ ∞ RK × ϕ = E(−t; −iA)BkΘ× (t − s)ϕ(s) ds dt 0

0

∞

= 0

0

∞

E(−t; −iA)BkΘ× (t − s)ϕ(s) dt ds.

For s > 0 and x ∈ X, consider the identity ∞ E(−t; −iA)B(ΛΘ× x)(t − s) dt 0

= E(−s; −iA)x − PΘ E(−s; −iA× )x. To prove it, we ﬁrst take x ∈ D(A) = D(A× ). Then, for t = 0 and t = s, d E(−t; −iA)E(t − s; −iA× )x dt = iE(−t; −iA)BCE(t − s; −iA× )x = iE(−t; −iA)BC(A× )−1 E(t − s; −iA× )A× x.

(5.55)

104

Chapter 5. Factorization of matrix functions analytic in a strip

Because C(A× )−1 is bounded, the last expression is a continuous function of t on the intervals [0, s] and [s, ∞). It follows that (5.55) holds for x ∈ D(A). The validity of (5.55) for arbitrary x ∈ X can now be obtained by a standard approximation argument based on the fact that D(A) is dense in X and the continuity of the operators involved. Substituting (5.55) in the expression for RK × ϕ, one immediately gets the desired identity R(I − K × ) + JR× = 0. Step 2. Next we show that RU × + JJ × = IIm PΘ . Take x in Im PΘ . Then ∞ RU × x = − E(−t; −iA)B(ΛΘ× x)(t) dt. (5.56) 0

Apart from the minus sign, the right-hand side of (5.56) is exactly the same as the left-hand side of (5.55) for s = 0. It is easy to check that (5.55) also holds for s = 0, provided that the right-hand side is interpreted as −PΘ x + PΘ PΘ× x. Thus RU × x = PΘ× x = x − PΘ PΘ× x, and the desired identity RU × + JJ × = IIm PΘ is proved. Step 3. This step concerns the identity (I − K)U × + U J × = 0. Take x ∈ Im PΘ . m Then U × x = −QΛΘ× x, where Q is the projection of Lm 1 (R) onto L1 [0, ∞) along m L1 (−∞, 0]. Here the latter two spaces are considered as subspaces of Lm 1 (R). Observe now that QΛΘ× = ΛΘ× (I − PΘ× )x. For x ∈ D(A) = D(A× ) this is evident, and for arbitrary x one can use an approximation argument. Hence KU × x = Qh, where h = −kΘ ∗ ΛΘ× (I − PΘ× )x , that is, h is the (full line) convolution product of −kΘ and ΛΘ× (I − PΘ× )x. Taking Fourier transforms, one gets h(λ)

= C(λ − A)−1 BC(λ − A× )−1 (I − PΘ× )x = C(λ − A)−1 (I − PΘ× )x − C(λ − A× )−1 (I − PΘ× )x.

Put g = U × x + U PΘ× x. Since both U and U × map into Im Q = Lm 1 [0, ∞), we have g = Qg. Also g = −ΛΘ× (I − PΘ× )x + ΛΘ(I − PΘ )PΘ× x, and hence g(λ) = −C(λ − A× )−1 (I − PΘ× )x − C(λ − A)−1 (I − PΘ )PΘ× x. Since x ∈ Im PΘ , it follows that h(λ) − g(λ) = C(λ − A)−1 PΘ (I − PΘ× )x. So h(λ) − g(λ) is the Fourier transform of −ΛΘ PΘ (I − PΘ× )x. But then h − g = −ΛΘ PΘ (I − PΘ× )x = −(I − Q)ΛΘ (I − PΘ× )x. Applying Q to both sides of this identity, we get Qh = Qg = g. In other words, KU × x = U × x + PΘ× x for all x ∈ X, and this is nothing else than the identity (I − K)U × + U J × = 0. Step 4. Finally, we prove (I − K)(I − K × ) + U R× = I. Let L be the (full line) convolution integral operator associated with kΘ , featured in Theorem 5.8. Since Θ and Θ× are both realization triples, the operator I − L is invertible with inverse (I − L)−1 = I − L× , where L× is the convolution integral operator associated

5.7. Inverting Wiener-Hopf integral operators

105

m ˙ m with Θ× . With respect to the decomposition Lm 1 (R) = L1 [0, ∞) + L1 (−∞, 0), we write I − L and its inverse in the form I − K× ∗ I − K L+ × . (5.57) , I −L = I −L= ∗ ∗ ∗ L× −

Thus L+ is the right Hankel operator associated with kΘ , and the operator L× − is the left Hankel operator associated with kΘ× . But then Lemma 5.9 yields L+ ψ

=

L× −ϕ =

−QΛΘ ΓΘ ψ,

ψ ∈ Lm 1 (−∞, 0],

(5.58)

(I − Q)ΛΘ× ΓΘ× ϕ,

ϕ ∈ Lm 1 [0, ∞).

(5.59)

Since I − L× is the inverse of I − L, formula (5.57) shows that (I − K)(I − K × ) + L+ L× − = I. × So, in order to get the desired identity, it suﬃces to show that L+ L× − = UR . As was observed in the last paragraph of Step 2 of the present proof, (5.55) also holds for s = 0, that is ∞ E(−t; −iA)B(ΛΘ× x)(t) dt = PΘ (I − PΘ× )x, x ∈ X. 0

Analogously, one has 0 E(−t; −iA)B(ΛΘ× x)(t) dt = (I − PΘ )PΘ× x, −∞

x ∈ X.

Using the expressions for L+ and L× − given in (5.58) and (5.59) we obtain L+ L × −ϕ

= −QΛΘ ΓΘ (I − Q)ΛΘ× ΓΘ× ϕ = −QΛΘ (I − PΘ )PΘ× ΓΘ× ϕ = U RΘ ϕ.

Thus (I − K)(I − K × ) + U R× = I holds, and the lemma is proved.

Following [13] (see also Section III.4 in [51]) we summarize the result of the preceding lemma by saying that the operators I −K and J × are matricially coupled with (5.54) being the coupling relation. The coupling relation is very useful. For instance, this relation and Corollary III.4.3 in [51] immediately yield the following result. Corollary 5.13. Let the operators K, K × , U, U × , R, R× , J and J × be as in (5.54). Then I − K is invertible if and only if J × is invertible, and in that case (I − K)−1 = I − K × − U × (J × )−1 R× ,

(J × )−1 = J − R(I − K)−1 U.

(5.60)

106

Chapter 5. Factorization of matrix functions analytic in a strip

Proof of Theorem 5.10. We split the proof into two parts. In the ﬁrst part we show that the invertibility of T implies that Θ× is a realization triple. In the second part we assume that Θ× is a realization triple and complete the proof by using Lemma 5.11 and Corollary 5.13. Part 1. Since the kernel function k is equal to kΘ , we know from Proposition 5.5 that the symbol of T is equal to WΘ . Assume T is invertible. From the general theory of Wiener-Hopf operators we know that this assumption implies that WΘ (λ) is invertible for all real λ. But then we can use the ﬁnal part of Theorem 5.7 to show that Θ× is a realization triple. Part 2. In this part we assume that Θ× is a realization triple. From Corollary 5.13 we know that T = I −K is invertible if and only if J × is invertible. By Lemma 5.11 the latter happens if and only if condition (ii) is satisﬁed. Together with the result of the ﬁrst part we have now shown that T is invertible if and only if conditions (i) and (ii) are both fulﬁlled. Moreover, if these conditions are fulﬁlled we see from the ﬁrst parts of formulas (5.60) and (5.53) that (I − K)−1 = I − K × − U × (I − Π)R× , where K × , R× and U × are as in Lemma 5.12, and Π is the projection of X along Im PΘ onto Ker PΘ× . Using the deﬁnitions of the operators K × , R× and U × given in Lemma 5.12, the formula for T −1 presented in Theorem 5.10 is now clear.

5.8 Explicit canonical factorization In this section we use realization triples to construct a canonical factorization for an m × m matrix function W of the form (5.1) with k being given by (5.2). By Theorem 5.6 such a function is the transfer function of a realization triple Θ = (A, B, C). In what follows it is assumed that Θ is given. We present necessary and suﬃcient conditions for the existence of a canonical factorization in terms of the operators appearing in the realization triple, Also, supposing these conditions are fulﬁlled, we give formulas for the factors and their inverses in a canonical factorization of W . The main result (Theorem 5.14 below) is the natural analogue of Theorem 5.3 for the functions considered in this section. For the deﬁnition of a canonical factorization, see Section 1.1 (cf., also Section 3.1). Theorem 5.14. Let the m × m matrix function W be given by W (λ) = I + C(λ − A)−1 B, with Θ = (A, B, C) a realization triple, and let A× be the associate main operator of Θ. Then W admits a canonical factorization with respect to the real line if and only if the following two conditions are satisﬁed: (i) Θ× = (A× , B, −C) is a realization triple,

5.8. Explicit canonical factorization

107

˙ Ker PΘ× . (ii) X = Im PΘ + Here X is the state space of both Θ and Θ× , and PΘ and PΘ× are the separating projections of −iA and −iA× , respectively. If the conditions (i) and (ii) are satisﬁed, then the projection Π of X along Im PΘ onto Ker PΘ× maps D(A) = D(A× ) into itself, and a canonical factorization W = W− W+ with respect to the real line is given by W (λ) = W− (λ)W+ (λ), λ ∈ R, where the factors and their inverses can be written as W− (λ)

=

I + C(λ − A)−1 (I − Π)B,

W+ (λ)

=

I + CΠ(λ − A)−1 B,

W−−1 (λ)

=

I − C(I − Π)(λ − A× )−1 B,

W+−1 (λ)

=

I − C(λ − A× )−1 ΠB.

The projection Π maps D(A) = D(A× ) into itself and D(A) ⊂ D(C). Hence the right-hand sides of the ﬁrst two of the above four expressions are well-deﬁned on ρ(A), and those of the last two are well-deﬁned on ρ(A× ). In particular the formulas make sense for λ in a strip containing the real line. At ﬁrst sight this seems to be short of the requirements for Wiener-Hopf factorization. We will come back to and resolve this point at the end of the proof. Proof of Theorem 5.14. The proof will be divided into four parts. In the ﬁrst we show that the conditions (i) and (ii) are necessary and suﬃcient. In the remaining three parts we assume that (i) and (ii) are satisﬁed. Part 1. Let K be the Wiener-Hopf integral operator with kernel function kΘ . Then the function W is the symbol of the operator I − K, and hence we know from the general theory of Wiener-Hopf integral equations that W admits a canonical factorization with respect to the real line if and only if T = I − K is invertible. The ﬁrst part of Theorem 5.10 implies that the latter is satisﬁed if and only if (i) and (ii) are fulﬁlled. Thus (i) and (ii) are necessary and suﬃcient in order that W admits a canonical factorization with respect to the real line. In the remaining three parts of the proof we assume that conditions (i) and (ii) are satisﬁed; Π will be the projection of X along Im PΘ onto Ker PΘ× . Part 2. In this part we show that Π maps D(A) into itself. To do this we need the operator J × deﬁned by (5.52). Our hypotheses imply (see Lemma 5.11) that J × is invertible and that Π = I − (J × )−1 PΘ× . Recall that PΘ× maps D(A) into D(A) ∩ Im PΘ× . Thus in order to prove that D(A) is invariant under Π it suﬃces to show that (J × )−1 maps D(A) ∩ Im PΘ× into D(A). From the relation (5.54) and the invertibility of the operator I − K, it follows (see Corollary 5.13) that (J × )−1 = J − R(I − K)−1U,

108

Chapter 5. Factorization of matrix functions analytic in a strip

where U , R and J are as in Lemma 5.12. Take x ∈ D(A) ∩ Im PΘ× . Then U x = m m QΛΘ x ∈ Dm 1 [0, ∞), where Q is the projection of L1 (R) onto L1 [0, ∞) along m L1 (−∞, 0]. From the general theory of Wiener-Hopf operators we know that (I − K)−1 = (I + Γ1 )(I + Γ2 ),

(5.61)

where for j = 1, 2 the operator Γj is the integral operator given by ∞ (Γj ϕ)(t) = γj (t − s)ϕ(s) ds, t > 0, 0

Lm×m (R). 1

with γj ∈ In fact, γ1 has its support in [0, ∞) and γ2 in (−∞, 0]; see Section 1.5. From the representation (5.61) it follows that (I − K)−1 maps m×m Dm (R) and f ∈ Dm 1 [0, ∞) into itself. Note in this context that for h ∈ L1 1 (R), m we have h ∗ f ∈ D1 (R) and (h ∗ f ) = h ∗ f + h(·) f (0+) − f (0−) . Thus (I − K)−1 U x ∈ Dm 1 (R). But then the ﬁnal part of Proposition 5.4 tells us that we end up in D(A) by applying ΓΘ . We conclude that R(I − K)−1 U maps D(A) ∩ Im PΘ× into D(A). Since the separating projection PΘ maps D(A) into itself, we know that J maps D(A) ∩ Im PΘ× also into D(A). Thus (J × )−1 maps D(A) ∩ Im PΘ× into D(A), and hence Π maps D(A) into itself. Amplifying on the above, we note that J × maps D(A) ∩ Im PΘ in a one-to-one way onto D(A) ∩ Im PΘ× . Since J × is invertible, it suﬃces to show that (J × )−1 maps D(A) ∩ Im PΘ× into D(A) ∩ Im PΘ . We have already shown that (J × )−1 maps D(A) ∩ Im PΘ× into D(A), and the inclusion (J × )−1 Im PΘ× ⊂ Im PΘ is clear from the deﬁnition of J × . Thus J × has the desired property. Part 3. According to our hypotheses and the fact that Π maps D(A) into itself, we have the following direct sum decompositions:

Write

A=

X

=

D(A)

=

A1

Z

0

A2

˙ Ker PΘ× , Im PΘ + ˙ D(A) ∩ Ker PΘ× . D(A) ∩ Im PΘ +

,

B=

B1 B2

,

C=

C1

C2

(5.62) (5.63)

(5.64)

for the corresponding matrix representations of A, B, and C. We now show that Θ1 = (A1 , B1 , C1 ),

× Θ× 1 = (A1 , B1 , −C1 ),

Θ2 = (A2 , B2 , C2 ),

× Θ× 2 = (A2 , B2 , −C2 ),

are realization triples, and we analyze the spectral properties of their main oper× ators. Here A× 1 = A1 − B1 C1 and A2 = A2 − B2 C2 .

5.8. Explicit canonical factorization

109

We start with Θ1 . Note that A1 (Im PΘ → Im PΘ ) and C1 (Im PΘ → Cm ) are the restrictions of A and C to D(A) ∩ Im PΘ , respectively. Since PΘ is the separating projection of Θ, this implies that Θ1 is a realization triple. From the deﬁnition of A1 it follows that −iA1 is the inﬁnitesimal generator of a strongly continuous left semigroup of negative exponential type. Thus the kernel function k1 = kΘ1 has its support in (−∞, 0] and W1 (λ) = I − k1 (λ) = I + C1 (λ − A1 )−1 is deﬁned and analytic on an open half plane of the type Im λ < −ω with ω strictly negative. × : Im PΘ → Im PΘ× be the operator deﬁned by Next, we consider Θ× 1 . Let J × (5.52). We know that J is invertible, mapping D(A) ∩ Im PΘ onto D(A) ∩ Im PΘ× . It is easy to check that J × provides a similarity between the operator A× 1 and the restriction of A× to D(A× ) ∩ Im PΘ× . Hence iA× 1 is the inﬁnitesimal generator of a strongly continuous left semigroup of negative exponential type. But then Theorem 5.7 guarantees that Θ× 1 is a realization triple. Furthermore , the kernel function k1× associated with Θ× 1 has its support in (−∞, 0], and −1 W1 (λ)−1 = I − k1× (λ) = I − C1 (λ − A× B1 1)

for all λ with Im λ < −ω. Here it is assumed that the negative constant ω has been taken suﬃciently close to zero. × We proceed by considering Θ2 and Θ× 2 . Obviously Θ2 is a realization triple, and a similarity argument of the type presented above yields that the same is true for Θ2 . The operators −iA2 and −iA× 2 are inﬁnitesimal generators of strongly continuous right semigroups of negative exponential type. Hence the kernel functions k2 and k2× associated with Θ2 and Θ× 2 , respectively, have their support in [0, ∞). Finally, taking |ω| smaller if necessary, we have that W2 (λ) = I − k2 (λ) = I + C2 (λ − A2 )−1 B2 and

−1 W2 (λ)−1 = I − k2× (λ) = I − C2 (λ − A× B2 2)

are deﬁned and analytic on Im λ > −ω. We may assume that both Θ and Θ× are of exponential type ω. For values of λ with |λ| < −ω one then has W (λ) = I + C1 (λ − A1 )B1−1 + C2 (λ − A2 )−1 B2 +C1 (λ − A1 )−1 Z(λ − A2 )−1 B2 . Now Ker PΘ× is an invariant subspace for ⎡ ⎤ Z − B1 C2 A× 1 ⎦, A× = ⎣ × −B2 C1 A2

110

Chapter 5. Factorization of matrix functions analytic in a strip

and so Z = B1 C2 . Substituting this in the above expression for W (λ), we get W (λ) = W1 (λ)W2 (λ). Clearly this is a canonical Wiener-Hopf factorization. The expressions obtained for the factors and their inverses are not quite the same as those given in the theorem. One veriﬁes without diﬃculty, however, that for λ in the intersection of ρ(A) and ρ(A× ), they amount to the same. For further information on this point we refer again (see the proof of Theorem 3.2) to Section 2.5 in [20] where the case when all three operators A, B and C are bounded is analyzed in great detail. Inspired by the terminology used in [20] (see also [11], Section 1.1), we introduce some additional terminology and notation. Let Θ = (A, B, C) be a realization triple with state space X, and let Π be a projection of X which maps D(A) into itself. We then have X

=

D(A)

=

˙ Im Π, Ker Π + ˙ D(A) ∩ Im Π , D(A) ∩ Ker Π +

and with respect to these decompositions the operators A, B and C have the form A11 A12 B1 A= , B= , C = C1 C2 . A21 A22 B2 The triple (A22 , B2 , C2 ) will be called the projection of Θ = (A, B, C) associated with Π, and it is denoted by prΠ (Θ). Note that (A11 , B1 , C1 ) is the projection prI−Π (Θ) of (A, B, C) associated with the projection I − Π. A particularly interesting case for what follows is when Π is a supporting projection for Θ. This means that besides the Π-invariance of D(A) = D(A× ) the following inclusion relations are satisﬁed: A D(A) ∩ Ker Π ⊂ Ker Π, A× D(A× ) ∩ Im Π ⊂ Im Π. In that situation we have A12 = B1 C2 and A21 = 0. Also Π is a supporting projection for the realization triple Θ = (A, B, C) if and only I − Π is a supporting projection for Θ× = (A× , B, −C). Finally, if Π is supporting for Θ, the arguments used in Part 3 of the proof of Theorem 5.14 show that prΠ (Θ) and prI−Π (Θ) are again realization triples. With this notation and terminology we have the following alternative version of Theorem 5.10. Theorem 5.15. Let T be the Wiener-Hopf integral operator on Lm 1 (R) given by (5.51). Assume that k = kΘ for some realization triple Θ = (A, B, C). Then T is invertible if and only if the following two conditions are satisﬁed: (i) Θ× = (A× , B, −C) is a realization triple, ˙ Ker PΘ× . (ii) X = Im PΘ +

5.9. The Riemann-Hilbert boundary value problem revisited (2)

111

Here X is the state space of both Θ and Θ× , and PΘ and PΘ× are the separating projections of −iA and −iA× , respectively. If (i) and (ii) hold, then the projection Π of X onto Ker PΘ× along Im PΘ is a supporting projection for Θ, the complementary projection I − Π is a supporting projection for Θ× , and ∞ −1 γ(t, s)g(s) ds. T g (t) = g(t) − 0

Here γ is given by s ⎧ × × × ⎪ ⎪ k (t − s) − k+ (t − r)k− (r − s) dr, ⎪ ⎨ + γ(t, s) =

s < t,

0

t ⎪ ⎪ ⎪ × × × ⎩ k− (t − s) − k+ (t − r)k− (r − s) dr,

s > t,

0

× × and k− are the kernel functions associated with the realization triples where k+ × prΠ (Θ ) and prI−Π (Θ× ), respectively.

5.9 The Riemann-Hilbert boundary value problem revisited (2) In this section we deal with the Riemann-Hilbert boundary value problem on the real line for matrix functions W of the form ∞ eiλt k(t) dt, (5.65) W (λ) = I − −∞

where k is an m× m matrix-valued function with the property that for some ω < 0 the entries of e−ω|t| k(t) are Lebesgue integrable on the real line. In this case W is analytic in a strip around the real axis. For such a function the Riemann-Hilbert problem consists of ﬁnding pairs Φ+ , Φ− of Cm -valued functions on the real line such that W (λ)Φ+ (λ) = Φ− (λ), −∞ < λ < ∞ (5.66) while, in addition, Φ+ and Φ− are Fourier transforms of integrable Cm -valued functions with support in [0, ∞) and (−∞, 0], respectively. These requirements imply that Φ+ and Φ− both vanish at inﬁnity and that they are continuous on the closed upper and closed lower half plane, respectively. From the special form of k in (5.65) we know that W is the transfer function of some realization triple Θ = (A, B, C). The following theorem gives the solution of the Riemann-Hilbert problem for W in terms of the operators A, B and C appearing in the triple. Theorem 5.16. Let W be the transfer function of realization triple Θ = (A, B, C). Assume Θ× = (A× , B, C) is a realization triple too (or, equivalently, that W (λ)

112

Chapter 5. Factorization of matrix functions analytic in a strip

is invertible for all λ ∈ R). Write PΘ and PΘ× for the separating projections of −iA and −iA× , respectively. Then the pair of functions Φ+ , Φ− is a solution of the Riemann-Hilbert boundary value problem (5.66) if and only if there exists x ∈ Im PΘ ∩ Ker PΘ× such that ∞ × −1 eiλt (Λ× (5.67) Φ+ (λ) = C(λ − A ) x = Θ x)(t) dt, 0

Φ− (λ)

=

−1

C(λ − A)

x =−

0

−∞

eiλt (ΛΘ x)(t) dt.

(5.68)

Moreover the vector x is uniquely determined by the functions Φ+ , Φ− . Proof. Take x ∈ Im PΘ ∩ Ker PΘ× . Condition (C4) in the deﬁnition of a realization triple implies that (Λ× Θ x)(t) is zero almost everywhere on the half line −∞ < t ≤ 0, while (ΛΘ x)(t) is zero almost everywhere on 0 ≤ t < ∞. It follows that we can apply (5.29) to both Θ and Θ× in order to show that ∞ ∞ × −1 eiλt (Λ× x)(t) dt = eiλt (Λ× x, Θ Θ x)(t) dt = C(λ − A ) 0

−∞

0

∞

eiλt (ΛΘ x)(t) dt

∞

= −∞

eiλt (ΛΘ x)(t) dt = −C(λ − A)−1 x.

Thus the functions Φ+ and Φ− in (5.67) and (5.68) are well-deﬁned. Furthermore, these functions are Fourier transforms of integrable Cm -valued functions with support in [0, ∞) and (−∞, 0], respectively. From (5.35) we see that (5.66) is satisﬁed. Thus the pair Φ+ , Φ− is a solution of the Riemann-Hilbert problem. To prove the reverse implication, assume that the pair Φ+ , Φ− is a solution of the Riemann-Hilbert problem (5.66). Write Φ+ and Φ− in the form ∞ 0 iλt e φ+ (t) dt, Φ− (λ) = eiλt φ− (t) dt, Φ+ (λ) = 0

−∞

Lm 1 [0, ∞)

Lm 1 (−∞, 0].

and φ− ∈ Now, let kΘ be the kernel function where φ+ ∈ associated with Θ, and consider the integral operator on Lm 1 [0, ∞) deﬁned by ∞ (Kf )(t) = kΘ (t − s)f (s) ds, a.e. on [0, ∞). 0

Using (5.66) and the fact that W (λ) = Im − kΘ (λ), a routine argument yields that φ+ − Kφ+ = 0. In other words, φ+ ∈ Ker (I − K). Next, we use the coupling relation (5.54) together with Corollary 4.3 in Section III.4 of [51]. It follows that φ+ = U × x for some x in Ker J × , where U × : Im PΘ → Lm 1 [0, ∞), J × : Im PΘ → Im PΘ× ,

(U × x)(t) = −(ΛΘ× x)(t), J × x = PΘ× x.

a.e. on [0, ∞),

5.9. The Riemann-Hilbert boundary value problem revisited (2)

113

Obviously, Ker J × = Im PΘ ∩ Ker PΘ× . Thus there exists x ∈ Im PΘ ∩ Ker PΘ× such that φ+ (t) = −(ΛΘ× x)(t) a.e. on the half line 0 ≤ t < ∞. But then (5.67) is satisﬁed. By (5.66), the identity (5.29) applied to Θ× = (A× , B, −C), and (5.35) we have Φ− (λ) = W (λ)Φ+ (λ) = C(λ − A)−1 x. Hence (5.68) holds too. It remains to prove the uniqueness of the vector x in (5.67) and (5.68). Assume that x is a second vector with the same properties as the vector x. So x ∈ Im PΘ ∩ Ker PΘ× while (5.67) and (5.68) hold true with x in place of x. Let J × and U × be as in the previous paragraph. Since Ker J × = Im PΘ ∩ Ker PΘ× , we have x − x ∈ Ker J × . Furthermore, the fact that the left-hand side of (5.67) does × not depend on x nor on x yields that (Λ× Θ x)(t) = (ΛΘ x )(t) a.e. on [0, ∞). Thus × × × × U x = U x . It follows that both J (x − x ) and U (x − x ) are equal to zero. If x = x , this implies that the operator deﬁned by the right-hand side of (5.54) is not invertible, which is impossible by Lemma 5.12. We conclude that x = x , as desired.

Notes The material presented in this chapter is taken from the papers [16] and [15]. In [16] the reader will also ﬁnd a systematic treatment of realization triples (A, B, C) with C bounded and A unbounded. The notion of an exponentially dichotomous operator, which has been introduced in [16], has proved to be quite useful in other areas. See, e.g., the papers [22] and [93]. The theory of realization triples is also used in [14] and [92]. The papers [90] and [91] present an extension of the theory of realization triples to operator-valued functions by introducing two-sided Pritchard-Salomon realizations. In particular, the factorization theory of Section 5.8 is developed further in [91]. For more information on exponentially dichotomous operators, including various perturbation theorems and a wide variety of applications, we refer to the monograph [111]. See also the notes to Chapter 6.

Chapter 6

Convolution equations and the transport equation In this chapter the factorization theory developed in the previous chapters is applied to solve a linear transport equation. It is known that the transport equation may be transformed into a Wiener-Hopf integral equation with an operator-valued kernel function (see [40]). An equation of the latter type can be solved explicitly if a canonical factorization of its symbol is available (cf., Sections 1.1 and 3.2). In our case the symbol may be represented as a transfer function, and to make the factorization the general factorization theorem of the second chapter can be applied. This requires that one ﬁnds an appropriate pair of invariant subspaces. In the case of the transport equation the choice of the subspaces is evident, but to prove that their direct sum is the whole space takes some eﬀort. The latter is related to a new diﬃculty that appears here. Namely, in this case the curve cuts through the spectra of the main operator and the associate main operator. Nevertheless, due to the special structure of the operators involved, the factorization can be made and explicit formulas are obtained. Since our main purpose is to show how our method works, we restrict ourselves to the case when the kernel function describing the eﬀect of the scattering is of ﬁnite rank. In Section 6.1 we describe the transport equation that is considered in this chapter. To illustrate our approach we ﬁrst study (in Section 6.2) a simpliﬁed model, namely when the scattering appears only in a ﬁnite number of directions. In Section 6.3 the vector-valued Wiener-Hopf equation associated to the transport equation is introduced. In Section 6.4 it is shown that under appropriate conditions a canonical factorization of the symbol associated with the equation can be constructed, and the matching of corresponding invariant subspaces is established in Section 6.5. In Section 6.6, the ﬁnal section of the chapter, we present formulas for the solution.

116

Chapter 6. Convolution equations and the transport equation

6.1 The transport equation Transport theory is a branch of mathematical physics concerned with the mathematical analysis of equations that describe the migration of particles in a medium, for instance, a ﬂow of electrons through a metal strip or radiative transfer in a stellar atmosphere. For the plane symmetric case, a stationary transport problem through a homogeneous medium can be modelled by an integro-diﬀerential equation of the following form: 1 ∂ψ(t, μ) k(μ, μ )ψ(t, μ ) dμ , t ≥ 0. (6.1) μ + ψ(t, μ) = ∂t −1 This equation is a balance equation. The unknown function ψ is a density function related to the expected number of particles in an inﬁnitesimal volume element. The right-hand side describes the eﬀect of the collisions. The variable μ is equal to cos α where α is the scattering angle, and therefore −1 ≤ μ ≤ 1. The variable t is not a time variable but a position variable, sometimes referred to as the optical depth. The kernel function k in the right-hand side of (6.1), which is called the scattering function, is assumed to be a real symmetric L1 -function on [−1, 1] × [−1, 1]. We shall consider the so-called half range problem, that is, we assume the medium to be semi-inﬁnite, and hence the position variable runs over the interval 0 ≤ t < ∞. Since the density of the incoming particles is known, the values of ψ(0, μ) are known for 0 < μ ≤ 1. Thus the above equation will be considered together with the boundary condition ψ(0, μ) = f+ (μ),

0 < μ ≤ 1,

(6.2)

where f+ is a given function on (0,1]. In the sequel we shall consider f+ as a function on [−1, 1] by setting f+ (μ) = 0 for −1 ≤ μ < 0, and we assume that f+ ∈ L2 [−1, 1]. There is also a boundary condition at inﬁnity, which appears in diﬀerent forms. Here we take the condition at inﬁnity to be t lim ψ(t, μ) exp = 0, −1 ≤ μ < 0. (6.3) t→∞ μ Thus the problem is to solve (6.1) under the boundary conditions (6.2) and (6.3). In this chapter we shall assume (cf., [81], [82] and [108]) that the scattering function k is given by n

aj pj (μ)pj (μ ), (6.4) k(μ, μ ) = j=0

where pj (μ) is the j-th normalized Legendre polynomial (see [53], page 26) and −∞ < aj < 1,

j = 0, 1, . . . , n.

(6.5)

6.1. The transport equation

117

In particular, the integral operator deﬁned by the right-hand side of (6.1) has ﬁnite rank. By writing ψ(t)(μ) = ψ(t, μ), we may consider the unknown function ψ as a vector function on [0, ∞) with values in H = L2 [−1, 1]. In this way equation (6.1) can be written as an operator diﬀerential equation: T

dψ (t) + ψ(t) = Kψ(t), dt

t ≥ 0,

(6.6)

where the derivative is taken with respect to the norm in H. In (6.6) the operators T and K are deﬁned by

T f (μ) = μf (μ),

Kf =

n

aj f, pj pj .

(6.7)

j=0

Because of (6.5), the operator I−K is strictly positive, and hence (6.6) is equivalent to dψ (I − K)−1 T = −ψ. dt In [81], [82], [108] this equation is solved by diagonalizing the operator (I − K)−1 T . Equation (6.1) with boundary conditions (6.2) and (6.3) can also be written as a Wiener-Hopf integral equation with an operator-valued kernel function (cf., [40]). In order to do this, let us introduce some notation. Let H+ and H− be the subspaces of H = L2 [−1, 1] consisting of all functions that are zero almost everywhere on [−1, 0] and [0, 1], respectively. By P+ and P− we denote the orthogonal projections of H = L2 [−1, 1] onto the subspace H+ and H− , respectively. Furthermore, h will be the operator-valued function deﬁned by ⎧ 1 t ⎪ ⎪ t > 0, exp − (P+ Kf )(μ), ⎪ ⎪ μ ⎨ μ (6.8) h(t)f (μ) = ⎪ ⎪ 1 t ⎪ ⎪ ⎩ − exp − t < 0, (P− Kf )(μ), μ μ and F is the vector-valued function given by ⎧ t ⎪ ⎪ , f ⎨ + (μ)exp − μ F (t)(μ) = ⎪ ⎪ ⎩ 0,

0 < μ ≤ 1, (6.9) −1 ≤ μ ≤ 0.

The operator-valued function h is referred to as the propagator function associated with the half range problem (6.6).

118

Chapter 6. Convolution equations and the transport equation

Given these functions h and F , equation (6.1) with the boundary conditions (6.2) and (6.3) can be written as ∞ ψ(t) − h(t − s)ψ(s) ds = F (t), t ≥ 0. (6.10) 0

To see this, multiply equation(6.1) by μ−1 exp(t/μ) and integrate over (0, t) when μ > 0 or over (t, ∞) in case μ < 0. With the help of the boundary conditions one gets in this way the integral equation (6.10). In [40] the asymptotics of solutions of equation (6.10) are found and used to describe the asymptotics of solutions of the transport equation.

6.2 The case of a ﬁnite number of scattering directions To make the method used in this chapter more transparent we ﬁrst consider the case when scattering occurs in a ﬁnite number of directions only. This assumption reduces the equation (6.1) and the boundary condition (6.2) to μi

n

dψ k(μi , μj )ψ(t, μj ), (t, μi ) + ψ(t, μi ) = dt j=1

(6.11)

i = 1, . . . , n, t ≥ 0, where ψ(0, μi ) = ϕ+ (μi ),

μi > 0.

(6.12)

n

To treat this version of the problem, introduce the C -valued function ⎡ ⎤ ψ(t, μ1 ) ⎢ ⎥ .. ψ(t) = ⎣ ⎦, . ψ(t, μn ) and the matrices T = diag [μ1 , . . . , μn ],

K = [k(μi , μj )]ni,j =1 .

(6.13)

Observe that T and K are real symmetric (hence selfadjoint) n × n matrices. Using this notation, equation (6.11) taken with the boundary condition (6.12) can be rewritten as 0 ≤ t < ∞, T ψ (t) + ψ(t) = Kψ(t), (6.14) P+ ψ(0) = x+ , where P+ is the projection on ⎡ ⎤ ⎡ x1 ⎢ ⎥ ⎢ P+ ⎣ ... ⎦ = ⎣ xn

Cn deﬁned by ⎤ y1 0 .. ⎥ , y = ⎦ i . x1 yn

if μi ≤ 0, if μi > 0,

6.2. The case of a ﬁnite number of scattering directions

119

and x+ is a given vector in Im P+ . Observe that P+ is the spectral projection of T corresponding to the positive eigenvalues of T . In what follows we assume additionally that T is invertible, which is the generic case and corresponds to the requirement that all μi in (6.11) are diﬀerent from 0; cf., formula (6.12). We shall look for solutions ψ of (6.14) in the space Ln2 [0, ∞). The ﬁrst step in solving (6.14) is based on the observation that, for invertible T , equation (6.14) is equivalent to a Wiener-Hopf integral equation with a rational matrix symbol. In fact, the following theorem holds. Theorem 6.1. Suppose T in (6.13) is invertible and let ψ ∈ Ln2 [0, ∞). Then ψ is a solution of equation (6.14) if and only if ψ is a solution of the Wiener-Hopf integral equation with a special right-hand side, namely ψ(t) −

0

∞

h(t − s)Kψ(s) ds = e−tT

−1

x+ ,

t ≥ 0,

(6.15)

where h is the propagator function associated with problem (6.14), that is,

h(t) =

⎧ −1 ⎨ T −1 e−tT P+ ,

t > 0,

⎩

t < 0.

−T −1 e−tT

−1

(6.16) P− ,

Here P− = I − P+ . Proof. Assume ψ is a solution of (6.14). Applying T −1 to the ﬁrst identity in (6.14), and solving the resulting equation by using variation of constants, yields ψ(t) = e−tT

−1

ψ(0) + e−tT

−1

t

esT

−1

T −1 Kψ(s)ds,

t ≥ 0.

0 −1

(6.17)

−1

Next, apply etT P− to both sides of (6.17) and use that etT and P− commute. −1 −1 Since etT P− is exponentially decaying on [0, ∞), the function etT P− Kψ(t) is integrable on [0, ∞), and thus lim e

t→∞

tT −1

P− ψ(t) = P− ψ(0) +

∞ 0

esT

−1

P− T −1 Kψ(s)ds.

(6.18)

−1

Again using that the function etT P− ψ(t) is integrable on [0, ∞), we see that the left-hand side of (6.18) has to be equal to zero, which proves that P− ψ(0) = −

0

∞

esT

−1

P− T −1 Kψ(s)ds.

(6.19)

Now, replace ψ(0) in (6.17) by P+ ψ(0) + P− ψ(0), use the boundary condition in

120

Chapter 6. Convolution equations and the transport equation

(6.14), and apply (6.19). This gives ∞ −1 −tT −1 ψ(t) = e x+ − e−(t−s)T P− T −1 Kψ(s)ds t t −1 e−(t−s)T T −1 Kψ(s)ds + 0

= e−tT

−1

x+ +

∞

0

h(t − s)Kψ(s)ds,

t ≥ 0.

Thus ψ is a solution of (6.15). To prove the converse statement, assume that ψ is a solution of (6.15). Thus ψ(t)

=

e−tT

−1

x+ + e−tT

−e−tT

−1

−1

∞

t

esT

0

esT

−1

t

−1

P+ T −1Kψ(s)ds

P− T −1 Kψ(s)ds,

(6.20) t ≥ 0.

It follows that ψ is absolutely continuous on each compact subinterval of [0, ∞), and hence the integrands in the right-hand side of (6.20) are continuous functions of the variable s. But then ψ is diﬀerentiable on (0, ∞), and we see that ψ (t)

=

−T −1ψ(t) + P+ T −1 Kψ(t) + P− T −1Kψ(t)

=

−T −1ψ(t) + T −1 Kψ(t),

t ≥ 0,

and hence ψ satisﬁes the ﬁrst equation in (6.14). From (6.20) it also follows that ∞ −1 esT P− T −1 Kψ(s)ds, t ≥ 0, ψ(0) = x+ − 0

which implies that P+ ψ(0) = P+ x+ = x+ . We conclude that ψ is a solution of the problem (6.14). A direct computation yields that the symbol of the Wiener-Hopf operator associated with (6.15) is the n × n matrix function W given by W (λ) = In − iT −1(λ + iT −1)−1 K, where In is the n × n identity matrix. Thus the symbol W is not only a rational matrix function but it is already given in a concrete realized form, namely W (λ) = In + C(λ − A)−1 B, with

A = −iT −1,

B = K,

C = −iT −1.

(6.21)

6.2. The case of a ﬁnite number of scattering directions

121

Notice that A does not have eigenvalues on the real line. Thus in order to solve equation (6.15) we can apply Theorem 3.3. This requires us to analyze the spectral properties of the matrix A× = A − BC = −iT −1(I − K).

(6.22)

In view of (6.5) it is natural to assume I − K is positive deﬁnite. Lemma 6.2. Assume I − K is positive deﬁnite. Then the matrix A× in (6.22) has no real eigenvalues and ˙ M ×, (6.23) Cn = M + where M is the spectral subspace of the matrix A in (6.21) corresponding to the eigenvalues in the upper half plane, and M × is the spectral subspace of A× in (6.22) corresponding to the eigenvalues in the lower half plane. Proof. Let · , · be the standard inner product in Cm and put S = (I − K)−1 T . Since I − K is positive deﬁnite, S is well-deﬁned and the sesquilinear form [x, y] = (I − K)x, y

(6.24)

is an inner product on Cn . From [Sx, y] = (I − K)Sx, y = T x, y and the fact that T is selfadjoint, it follows that S is selfadjoint with respect to the inner product [· , ·]. But then the same holds true for iA× = S −1. Thus A× is invertible and its eigenvalues are on the imaginary axis. In particular, A× has no real eigenvalues. Recall that P+ is the spectral projection of T corresponding to the positive eigenvalues of T . Let P+× be the analogous projection for S. Since T and S are invertible, T |Ker P+ is negative deﬁnite and S|Im P × is positive deﬁnite. Thus +

0 = x ∈ Ker P+

=⇒ T x, x < 0,

0 = x ∈ Im P+×

=⇒

[Sx, x] > 0.

But [Sx, x] = T x, x for each x ∈ Cn . It follows that Ker P+ ∩ Im P+× = {0}. In particular, rank P+ ≥ rank P+× . By repeating the argument with Ker P+ replaced by Ker P+× and Im P+× by Im P+ , we see that rank P+× ≥ rank P+ . But then we ˙ Im P+× . Finally, from iA = T −1 we see that may conclude that Cn = Ker P+ + × −1 M = Ker P+ , and from iA = S we conclude that M × = Im P+× . We can now apply Theorem 3.3 to solve equation (6.15). Note however that the right-hand side of (6.15) is of a special form. In fact, in terms of the matrices appearing in (6.21) this right-hand side can be written as g(t) = iCe−itA x+ , where x+ ∈ Im P+ . Thus instead of Theorem 3.3 we can also directly apply Corollary 3.4. This yields the following result.

122

Chapter 6. Convolution equations and the transport equation

Theorem 6.3. Assume I − K is positive deﬁnite and T is invertible. Then the matrix (I − K)−1 T is selfadjoint with respect to the inner product (6.24) and the half range problem (6.14) has a unique solution ψ in Ln2 (0, ∞), namely ψ(t) = e−tT

−1

(I−K)

Πx+ ,

t ≥ 0,

(6.25)

where Π is the projection of Cn along Ker P+ onto the spectral subspace Im P+× of (I − K)−1 T corresponding to its positive eigenvalues.

6.3 Wiener-Hopf equations with operator-valued kernel functions It is well-known that the Wiener-Hopf integral equation ∞ ψ(t) − k(t − s)ψ(s) dy = F (t), t≥0

(6.26)

0

can be solved by constructing appropriate factorizations of its symbol (cf., Sections 1.1, 3.2, the papers [49], [71], or the survey article [59]). In this section we shall describe this method for the case when k is an L1 -kernel function the values of which are compact operators on a separable Hilbert space H. So we assume that k(t) is a compact operator for each real t, that k(·)f, g is measurable on the real line for each f and g in H, and that ∞ k(t) dt < 0, −∞

where ·, · is the inner product on H, and · is the operator norm for operators on H. Note that the kernel function h considered in the previous section falls into this category. Recall that the symbol of equation (6.26) is the operator-valued function I − K(λ), where K(λ) is the Fourier transform of the kernel function k, i.e., ∞ K(λ) = eiλt k(t) dt, −∞ < λ < ∞. −∞

By the Riemann-Lebesgue lemma, we have limλ∈R, λ→±∞ K(λ) = 0. Here we also need the concept of canonical factorization, this time in the present inﬁnite dimensional context. The symbol is said to admit a (right ) canonical factorization with respect to the real line if I − K(λ) = G− (λ)G+ (λ),

−∞ < λ < ∞,

where the factors G− and G+ meet the following requirements:

(6.27)

6.3. Wiener-Hopf equations with operator-valued kernel functions

123

(i) the operator function G− is analytic on the (open) lower half plane λ < 0 and continuous on the closure of the left half plane in the Riemann sphere (inﬁnity included); also for each λ in this closure (inﬁnity included), the operator G− (λ) is invertible; (ii) the operator function G+ is analytic on the (open) upper half plane λ > 0 and continuous on the closure of the right half plane in the Riemann sphere (inﬁnity included); also for each λ in this closure (inﬁnity included), the operator G+ (λ) is invertible. Note that the deﬁnition is analogous to that given earlier in the matrix-valued case (see Sections 1.1 and 3.1). According to [49], because of the fact that k is an L1 -kernel function the values of which are compact operators on H, the inverses of the factors in the right-hand side of (6.27) can be written as ∞ 0 −1 −1 iλt e γ+ (t), G− (λ) = I + eiλt γ− (t) dt, (6.28) G+ (λ) = I + 0

−∞

where, γ+ and γ− are L1 -functions on [0, ∞) and (−∞, 0], respectively, whose values are compact operators on H. Let L2 (R+ , H) denote the space of all L2 -integrable functions on [0, ∞) with values in H. The identities (6.28) are important, because they allow for explicit formulas for the solutions of (6.26). Indeed, by [18] equation (6.26) has a unique solution ψ in L2 (R+ , H) for each F ∈ L2 (R+ , H) if and only if a canonical factorization (6.27) exists, and in that case (just as in Section 1.1 for matrix-valued kernel functions) the Fourier transform ψ of the solution ψ is given by −1 (6.29) ψ(λ) = G−1 + (λ)P G− (λ)F (λ) , where F is the Fourier transform of the right-hand side of equation (6.26), and P is the projection deﬁned by ∞ ∞ P f (t)eitλ dt = f (t)eitλ dt. −∞

0

Taking inverse Fourier transforms in (6.29) one ﬁnds ∞ ψ(t) = F (t) + γ(t, s)F (s) ds, 0

where γ(t, s) is given by (1.10), i.e., s ⎧ ⎪ ⎪ (t − s) + γ+ (t − r)γ− (r − s) dr, γ + ⎪ ⎨ 0 γ(t, s) = t ⎪ ⎪ ⎪ ⎩ γ (t − s) + γ (t − r)γ (r − s) dr, −

0

+

−

0 ≤ s < t, 0 ≤ t < s.

124

Chapter 6. Convolution equations and the transport equation

As we observed already, in (6.10) the kernel function h(·) is an L1 -function on the real line whose values are compact (in fact ﬁnite rank) operators on L2 [−1, 1]. In the next section we shall prove that the corresponding symbol admits a canonical factorization, and we shall describe the factors explicitly.

6.4 Construction of a canonical factorization We now return to equation (6.10). Note that its symbol is given by I − H(λ), where ∞ eiλt h(t)K dt = (I − iλT )−1 K, −∞ < λ < ∞. H(λ) = −∞

Here h is given by (6.8), and the operators T and K are as in formula (6.7). The operator function H is analytic on the strip |λ| < 1. Note that σ(T ) is the closed interval [−1, 1], so that (I − iλT )−1 is deﬁned for all λ in the complement of the union of the subsets i[1, ∞) and i(−∞, −1] of the imaginary axis. In this section we show that I − H(λ) admits a canonical factorization with respect to the real line: I − H(λ) = G− (λ)G+ (λ), −∞ < λ < ∞, where the factors and their inverses can be written as G− (λ)

= I − (I − iλT )−1(I − P )K(I − P K)−1 ,

G+ (λ) G−1 − (λ)

= I − (I − Q∗ K)−1 (I − Q∗ )(I − iλT )−1K, −1 = I + (I − Q∗ K)−1 Q∗ I − iλ(T × )∗ K,

G−1 + (λ)

= I + (I − iλT × )−1 P K(I − P K)−1 .

Here T × = (I − K)−1 T , and P and Q are projections of which the deﬁnition will be given below. With regard to the domains of the factors and their inverses, the situation is similar to what we encountered earlier for Theorems 3.2 and 5.14. In order to make the factorization we transform the symbol of equation (6.10) into another function W which is deﬁned and continuous on the imaginary axis. This will be done as follows. Recall that T and K are both selfadjoint and that I −K, being a strictly positive operator because of (6.5), is invertible. Hence, for non-zero purely imaginary values of λ, ∗ I − H(i/λ)∗ = I − (I + λ−1 T )−1 K = I − K(I − λ−1 T )−1 = I − λK(λ − T )−1 = (I − K) I − K(I − K)−1 T (λ − T )−1 .

6.4. Construction of a canonical factorization

125

We now introduce W by writing W (λ) = I − (I − K)−1 KT (λ − T )−1 .

(6.30)

Note that this expression is a unital realization for W . The state space is the separable Hilbert space H = L2 [−1, 1]. The operator T is the main operator, and (I − K)−1 T is the associate main operator, denoted above by T × (in line with the notation adopted in Section 2.1). Via (6.30) the function W is deﬁned and analytic on the resolvent set of T , so on the complement of the interval [−1, 1]. In particular, W is deﬁned and continuous on the imaginary axis punctured at the origin. We shall now prove that by setting W (0) = (I −K)−1 the restriction of W (now deﬁned on the complement of the set [−1, 0) ∪ (0, 1]) to the imaginary axis is a continuous function. For this we need to show that lim

α→ 0, α ∈ R

W (iα) = (I − K)−1 .

(6.31)

It is convenient to establish the following lemma which will also play a role later on in this section. Lemma 6.4. Let S be a bounded selfadjoint operator on a given Hilbert space. Then S(iα − S)−1 ≤ 1,

0 = α ∈ R,

(6.32)

while, furthermore, lim

S(iα − S)−1 f = −f,

f ⊥ Ker S.

α→ 0, α ∈ R

(6.33)

Under the additional assumption that S is a nonnegative operator, the limit result (6.33) can be sharpened to lim

λ→ 0, λ ≤ 0

S(λ − S)−1 f = −f,

f ⊥ Ker S.

(6.34)

Proof. Let ES (t) be the spectral resolution of the identity for S, and let f be an element of the underlying Hilbert space. Then ∞ t2 S(iα − S)−1 f 2 ≤ d ES (t)f 2 2 2 α + t −∞ ∞ ≤ d ES (t)f 2 −∞

=

f 2.

This proves (6.32). Next, observe that f + S(iα − S)−1 f 2 ≤

∞ −∞

α2 d ES (t)f 2 . + t2

α2

126

Chapter 6. Convolution equations and the transport equation

So by Lebesgue’s dominated convergence theorem we get lim

α→ 0, α ∈ R

f + S(iα − S)−1 f ≤ =

ES (0+)f 2 − ES (0−)f 2 (ES (0+) − ES (0−))f 2 ,

which is zero if f ⊥ Ker S. Hence (6.33) is proved. The argument for (6.34), taking nonnegativity of the operator S for granted, is analogous. The proof of (6.31) is now as follows. As Ker T = {0}, we see from Lemma 6.4 that limα→ 0, α ∈ R (iα + T )−1 T f = f, f ∈ H. Since K is compact (actually even of ﬁnite rank), it follows that (iα + T )−1T K tends to K in the operator norm if α ∈ R, α → 0. Taking adjoints, we obtain that the same holds true for −KT (iα − T )−1 . But then we have (6.31), where the convergence is with respect to the operator norm. So with W (0) = (I − K)−1 , indeed W becomes a continuous function on the imaginary axis. It is this operator function for which we want a (right) canonical Wiener-Hopf factorization. This time not with respect to the real line (see the deﬁnition in Section 6.3) but for the analogous situation where the curve in the Riemann sphere is the imaginary axis with inﬁnity included. The theory concerning canonical factorization developed earlier suggests that we have to ﬁnd an invariant subspace M for T such that the spectrum of T restricted to M lies in the closed right half plane, and an invariant subspace M × for T × such that the spectrum of T × restricted to M × lies in the closed left half plane. Since T is selfadjoint the choice of M is clear: M = H+ , where H+ is the subspace of H = L2 [−1, 1] consisting of all functions that are zero almost everywhere on [−1, 0]. As we shall see below, after replacing the standard inner product on L2 [−1, 1] by a suitable equivalent one, the operator T × is selfadjoint too. So for M × we can take the spectral subspace of T × corresponding to the part of the spectrum of T × on (−∞, 0]. The ﬁrst diﬃculty is to prove the matching of the subspaces M and M × , i.e., to show ˙ M × . Taking for granted that this has been established a second that H = M + diﬃculty appears, because in the present case the imaginary axis does not split the spectra of T and T × . So we cannot apply directly the theory developed so far, but we have to prove, using the speciﬁcs of the situation, that the factors obtained have the desired boundary behavior. The purpose of this section is to show that this approach works indeed. We begin by considering the operator T × = (I − K)−1 T . As I − K is strictly positive, [f, g] = (I −K)f, g deﬁnes an inner product on H = L2 [−1, 1] equivalent with the standard one. Writing A[∗] for the adjoint of an operator A relative to the inner product [· , ·], we have A[∗] = (I − K)−1 A∗ (I − K). ×

(6.35)

In particular, we see that the operator T is selfadjoint with respect to the inner product [·, ·]. Let E × (·) be the corresponding spectral resolution and introduce Hm = Im E × (0),

Hp = Ker E × (0).

6.4. Construction of a canonical factorization Then Hm and Hp are both invariant under T × and σ T × |Hp ⊂ [0, ∞) ∩ σ(T × ). σ T × |Hm ⊂ (−∞, 0] ∩ σ(T × ),

127

(6.36)

For T the situation is more straightforward. Indeed, T is selfadjoint with respect to the original (standard) inner product on H and leaves invariant the spaces H− and H+ featured in Section 6.1. Further σ T |H+ = [0, 1]. (6.37) σ T |H− = [−1, 0], The subspaces M and M × mentioned above are H+ and Hm , respectively. ˙ Hm . So proving that these subspaces match amounts to showing that H = H+ + In fact, in the next section we shall show the following stronger result: ˙ Hp , H = H− +

˙ Hm . H = H+ +

(6.38)

Let P be the projection of H along H− onto Hp , and let Q be the projection of H along H+ onto Hm . Since the subspaces H− , H+ are invariant under T and Hm , Hp are invariant under T × , both P and Q are supporting projections for the realization (6.30). Associated with these projections are two factorizations: !− (λ), !+ (λ)W W (λ) = W

W (λ) = W− (λ)W+ (λ).

(6.39)

With the appropriate choice for the value of the factors at the origin, both these factorizations are canonical factorizations of W with respect to the imaginary axis; the ﬁrst a left and the second a right factorization. In the sequel we only need the second factorization in (6.39). First we give the expressions for the factors W− (λ) and W+ (λ): W− (λ)

=

−1 I − (I − K)−1 KT |H+ λ − T |H+ (I − Q),

(6.40)

W+ (λ)

=

−1 I − (I − K)−1 KT |Hm λ − QT |Hm Q.

(6.41)

Note that there is slight abuse of notation here. Indeed, the operator I − Q in the formula for W− (λ) should be interpreted as a mapping from H onto H+ , and Q in the expression for W+ (λ) must be seen as a mapping from from H onto Hm . In particular QT |Hm should be read as the compression of T to Hm (relative to the ˙ Hm ). The function W− is deﬁned and analytic on the decomposition H = H+ + resolvent set of T |H+ so, by the second part of (6.37) on the complement of the interval [0, 1]. Similarly, the function W+ is deﬁned and analytic on the resolvent ˙ H− = H+ + ˙ Hm , and set of the compression operator QT |Hm . Now H = H+ + Lemma 3.1 guarantees that QT |Hm is similar to T |H− . In particular the resolvent sets of QT |Hm and T |H− coincide. It follows from the ﬁrst part of (6.37) that function W+ is deﬁned and analytic on the complement of the interval [−1, 0]. The argument also indicates that the second factorization in (6.39) holds for all

128

Chapter 6. Convolution equations and the transport equation

λ outside the interval [−1, 1]. Indeed this interval is precisely the union of the spectra of T |H+ and QT |Hm (cf., Theorem 2.6). Next we deal with the invertibility of the factors W− (λ) and W+ (λ). The above realization of W− has (I − Q)T × |H+ : H+ → H+ as its associate main operator. From Section 2.4 we now know that W− (λ) is invertible for λ in the intersection of the resolvent sets of T |H+ and (I − Q)T × |H+ . The ˙ Hm = resolvent set of T |H+ is the complement of the interval [0, 1]. As H = H+ + × × ˙ Hp + Hm , the compression operator (I − Q)T |H+ is similar to T |Hp . Hence, by the second part of (6.36), the resolvent set of (I − Q)T × |H+ is the complement of the set [0, ∞) ∩ σ(T × ). It follows that W− (λ) is invertible for all non-zero λ with λ ≤ 0, its inverse (see Theorem 2.4) being given by W− (λ)−1

=

−1 I + (I − K)−1 KT |H+ λ − (I − Q)T × |H+ (I − Q)

=

−1 I + KT × |H+ λ − (I − Q)T × |H+ (I − Q).

In an analogous manner one proves that W+ (λ) is invertible for all non-zero λ with λ ≥ 0, its inverse having the representation −1 Q W+ (λ)−1 = I + (I − K)−1 KT |Hm λ − T |× Hm −1 Q. I + KT × |Hm λ − T |× Hm

=

The above formulas contain the precise description of the factors W− , W+ and their inverses W−−1 , W+−1 . Giving up some precision but gaining in conciseness, we can also write W− (λ)

=

I − (I − K)−1 KT (λ − T )−1(I − Q),

W+ (λ)

=

I − (I − K)−1 KT Q(λ − T )−1 ,

W−−1 (λ)

=

I + KT ×(I − Q)(λ − T × )−1 ,

W+−1 (λ)

=

I + KT ×(λ − T × )−1 Q;

see Section 2.4 and [20], Section 2.5 for details. We have come close to proving that the second factorization in (6.39) is a (right) canonical factorization of W with respect to the imaginary axis. To make the proof complete we need to check the behavior of the functions at inﬁnity and at the origin. As far as the behavior at inﬁnity is concerned the situation is simple. Indeed the functions W− , W+ , W−−1 and W+−1 are analytic there with value the identity operator on H. For the origin the situation is more complicated.

6.4. Construction of a canonical factorization

129

Earlier we completed the deﬁnition of the function W , initially introduced via the unital realization (6.30), by stipulating that W (0) = (I − K)−1 . Now we make a similar move with respect to W− and W+ , in the ﬁrst instance given by (6.40) and (6.41), respectively. Indeed, we stipulate that W− (0) = (I − K)−1 (I − KQ),

W+ (0) = (I − K)−1 (I − KP ∗ ).

In this manner the closed left half plane λ ≤ 0 is contained in the domain of W− , and the closed right half plane λ ≥ 0 is contained in the domain of W+ . Our task is now threefold: to verify the invertibility of W− (0) and W+ (0), to demonstrate the continuity of W− and W+ on the appropriate half planes, i.e., to show that lim

W− (λ)

= (I − K)−1 (I − KQ),

(6.42)

lim

W+ (λ)

= (I − K)−1 (I − KP ∗ ),

(6.43)

λ→ 0, λ ≤ 0

λ→ 0, λ ≥ 0

and to verify that the factorization W = W− W+ holds at the origin. As a ﬁrst step we present the following lemma (which will also be used in Section 6.6 below). Lemma 6.5. Let P, Q and K be as above. Then Q∗ (I − K)P = 0,

(I − Q∗ )(I − P ) = 0.

(6.44)

Proof. Note that Im P = Hp is orthogonal to Im Q = Hm with respect to the inner product [f, g] = (I − K)f, g. Thus (I − K)P f, Qg = [P f, Qg] = 0,

f, g ∈ L2 [−1, 1].

This yields the ﬁrst identity in (6.44). Next observe that, relative to the usual inner product on H = L2 [−1, 1], the space Im (I − Q) = H+ is orthogonal to Im (I − P ) = H− . It follows that (I − P )f, (I − Q)g = 0, which proves the second identity in (6.44).

f, g ∈ L2 [−1, 1],

Corollary 6.6. The operators I − KQ and (I − K)−1 (I − KP ∗ ) are invertible and each other’s inverse. Proof. As K is compact (actually even of ﬁnite rank), the operator I − KQ is Fredholm of index zero. In particular I − KQ is invertible if and only if I − KQ is left invertible. Thus it suﬃces to show that the operator (I − K)−1 (I − KP ∗ ) is a left inverse of I − KQ. Now the identities in Lemma 6.5 can be rewritten as

130

Chapter 6. Convolution equations and the transport equation

Q∗ KP = Q∗ P and Q∗ + P − Q∗ P = I. Combining these, one gets I −K

=

I − (Q∗ + P − Q∗ P )K

=

I − Q∗ K − P K + Q ∗ P K

=

I − Q∗ K − P K + Q∗ KP K

=

(I − Q∗ K)(I − P K).

Taking adjoints yields I − K = (I − KP ∗ )(I − KQ), and this identity can be rewritten as (I − K)−1 (I − KP ∗ )(I − KQ) = I. The corollary can be rephrased by saying that W+ (0) = (I − K)−1 (I − KP ∗ ) is invertible with inverse I−KQ. Likewise W− (0) = (I−K)−1 (I−KQ) is invertible with inverse (I − K)−1 (I − KP ∗ )(I − K). It remains to verify (6.42) and (6.43). We begin with (6.42). For λ ≤ 0, λ = 0, we have W− (λ)

=

−1 I − K(I − K)−1 T |H+ λ − T |H+ (I − Q),

=

−1 I − K+ T |H+ λ − T |H+ (I − Q).

Here K+ is the restriction of K(I − K)−1 to H+ considered as an operator from H+ into H and, as before, I − Q should be read as a mapping from H onto H+ . The restriction operator T |H+ : H+ → H+ is selfadjoint and nonnegative. It also has a trivial null space. So we can apply Lemma 6.4 to show that lim

λ → 0, λ ≤ 0

−1 T |H+ λ − T |H+ f+ = −f+ ,

f+ ∈ H+ .

∗ Along with K+ , the operator K+ : H → H+ is compact (actually even of ﬁnite rank), and it follows that

lim

λ → 0, λ ≤ 0

−1 ∗ ∗ T |H+ λ − T |H+ K+ = −K+ ,

with convergence in norm. Taking adjoints we get lim

λ → 0, λ ≤ 0

−1 K+ T |H+ λ − T |H+ = −K+ ,

and hence lim

λ → 0, λ ≤ 0

−1 K+ T |H+ λ − T |H+ (I − Q) = −K+ (I − Q).

A simple computation gives I + K+(I − Q) = (I − K)−1 (I − KQ), and (6.42) is immediate.

6.4. Construction of a canonical factorization

131

Next we turn to (6.43). By Corollary 6.6 and the continuity of the operation of taking the inverse, it suﬃces to show that lim

λ → 0, λ ≥ 0

For λ ≥ 0, λ = 0, we have W+ (λ)−1

W+ (λ)−1 = I − KQ.

=

−1 I + KT × |Hm λ − T |× Hm Q,

=

−1 I + Km T × |Hm λ − T × |Hm Q.

Here Km is the restriction of K to Hm considered as an operator from Hm into H and, as before, Q should be read as a mapping from H onto Hm . Because T × = (I − K)−1 T , the operator T × |Hm has a trivial null space. Further it is nonpositive with respect to the alternative inner product [· , ·], and Km is compact. Using Lemma 6.4 in an analogous way as in the previous paragraph, we see that −1 lim Km T × |Hm λ − T × |Hm = −Km , λ → 0, λ ≥ 0

and we get lim λ → 0, λ ≥ 0 W+ (λ)−1 = I − KQ, as desired. From what we have obtained so far and a continuity argument it is already clear that the second factorization in (6.39) holds at the origin too. The calculation W− (0)W+ (0)

= (I − K)−1 (I − KQ)(I − K)−1 (I − KP ∗ ) = (I − K)−1 (I − KP ∗ )−1 I − K) (I − K)−1(I − KP ∗ ) = (I − K)−1 = W (0),

based on Corollary 6.6, corroborates this fact. Our ultimate goal in this section is to produce a right canonical factorization with respect to the real line of the symbol I −H(λ) of equation (6.10). For non-zero real λ we have I − H(λ) = W (i/λ)∗ (I − K), and with the right interpretation this identity even holds on the extended real line. Indeed, as W is given by a unital realization, the value of W at ∞ is I, and this corresponds with the fact that H(0) = K. Also by the Riemann-Lebesgue lemma, H vanishes at ∞, and this is in accord with W (0) = (I − K)−1 . The right canonical Wiener-Hopf factorization W = W− W+ with respect to the imaginary axis that we obtained for W now induces a right canonical Wiener-Hopf factorization with respect to the real line for the symbol. The details are given in the next two paragraphs. We begin by deﬁning a function G− on the complement in C∞ of the interval i[1, ∞) which is situated on the imaginary axis. The determining expressions are G− (λ)

= W+ (i/λ)∗ (I − Q∗ K),

G− (0) = I − Q∗ K, G− (∞)

= I.

132

Chapter 6. Convolution equations and the transport equation

Note that G− is analytic on the complement of i[1, ∞) in the ﬁnite complex plane C. Also G− is continuous on the closed lower half plane λ ≤ 0, this time inﬁnity included. Indeed, lim

λ → ∞, λ ≤ 0

G− (λ)

= =

lim

μ → 0, μ ≥ 0

W+ (μ)∗ (I − Q∗ K)

(I − P K)(I − K)−1 (I − Q∗ K) = I = G− (0).

Here we used (6.43). Further deﬁne G+ on the complement in C∞ of the interval i(−∞, −1], again located on the imaginary axis, by G+ (λ)

= (I − Q∗ K)−1 W− (i/λ)∗ (I − K),

G+ (0) = I − P K, G+ (∞)

= I.

Then G+ is analytic on the complement of i(−∞, −1]) in C. Also G+ is continuous on the closed upper plane λ ≥ 0, inﬁnity included. Indeed, using (6.42) one gets lim

λ → ∞, λ ≥ 0

G+ (λ) =

lim

(I − Q∗ K)−1 W− (μ)∗ (I − K) = I = G+ (0).

μ → 0, μ ≤ 0

Observe that I − H(λ) = G− (λ)G+ (λ), λ ∈ R. For non-zero λ this is clear from the corresponding factorization for W ; for λ = 0 we have G− (0)G+ (0) = (I − Q∗ K)(I − P K) = I − K = I − H(0). From what we saw in the preceding paragraph it is now clear that we have arrived at a right canonical factorization with respect to the real line, of the symbol I − H(λ). Explicit formulas for the −1 factors G− , G+ and their inverses G−1 − , G+ can be obtained from the descriptions −1 −1 of W+ , W− , W+ and W− given earlier in this section. In fact the formulas in question coincide with the ones already presented in the ﬁrst paragraph of this section. For the veriﬁcation of this we need the following intertwining result. Lemma 6.7. Let P and Q be as above. Then (I − Q∗ )T = T P . Proof. It is suﬃcient to establish the identities (I − Q∗ )T (I − P ) = 0 and Q∗ T P = 0. For the ﬁrst of these we argue as follows. Clearly (I − Q∗ )T (I − P )f, g = T (I − P )f, (I − Q)g. Now (I − P )f ∈ H− and (I − Q)g ∈ H+ . As H− is T -invariant we also have T (I − P )f ∈ H− . But H− ⊥ H+ . So (I − Q∗ )T (I − P )f, g = 0 for all f and g in H. It follows that (I − Q∗ )T (I − P ) = 0, as desired. Next observe that Q∗ T P f, g = T P f, Qg = [(I − K)−1 T P f, Qg] = [T × P f, Qg]. As P f ∈ Hp and Hp is invariant under T × , we have T × P f ∈ Hp . Furthermore Qg ∈ Hm . But Hm ⊥ Hp is H endowed with the inner product [·, ·]. It follows that (Q∗ T P f, g) = 0 for all f and g. Hence Q∗ T P = 0, which is the second identity we wanted to establish.

6.4. Construction of a canonical factorization

133

We proceed by deriving the state space formulas for G− , G+ and their inverses −1 −1 , KT Q(λ − T )−1. Hence, for λ = 0, G−1 − G+ . Recall that W+ (λ) = I − (I − K) G− (λ)

= = =

W+ (i/λ)∗ (I − Q∗ K) ∗ I − (I − K)−1 KT Q(i/λ − T )−1 (I − Q∗ K) I − iλ(I − iλT )−1 Q∗ T K(I − K)−1 (I − Q∗ K).

On account of Lemma 6.7, we have Q∗ T = T (I − P ). Also (I − K)−1 (I − Q∗ K) = (I − P K)−1 , and we get G− (λ)

−1 I − iλ (I − iλT ) T (I − P )K(I − K)−1 (I − Q∗ K) = I − K − iλ (I − iλT )−1 T (I − P )K (I − P K)−1 .

=

But then, proceeding in a straightforward manner, G− (λ)

=

I − K − iλT (I − iλT )−1(I − P )K (I − P K)−1 I − K + (I − P )K − (I − iλT )−1(I − P )K (I − P K)−1 I − P K − (I − iλT )−1 (I − P )K (I − P K)−1

=

I − (I − iλT )−1 (I − P )K(I − P K)−1 .

= =

In this computation λ was of course taken to be non-zero. For λ = 0, the last expression in the above series of identities reduces to I − (I − P )K(I − P K)−1 and this is easily seen to be equal to (I − K)(I − P K)−1 . The latter can be rewritten as I − Q∗ K which was earlier identiﬁed as the value G− (0) of G− in the origin. So in the ﬁnal analysis the zero value of λ is admissible too. Next we turn to G+ which was deﬁned using W− . For the latter we have the expression W− (λ) = I − (I − K)−1 KT (λ − T )−1 (I − Q) and we can carry out a similar computation as the one presented above: G+ (λ)

(I − Q∗ K)−1 W− (i/λ)∗ (I − K) ∗ = (I − Q∗ K)−1 I − (I − K)−1 KT (i/λ − T )−1 (I − Q) (I − K) = (I − Q∗ K)−1 I − iλ(I − Q∗ )(I − iλT )−1 T K(I − K)−1 (I − K) −1 = (I − Q∗ K)−1 I − K − iλ(I − Q∗ )T (I − iλT ) K = (I − Q∗ K)−1 I − K + (I − Q∗ )(I − iλT − I)(I − iλT )−1K = (I − Q∗ K)−1 I − K + (I − Q∗ )K − (I − Q∗ )(I − iλT )−1 K =

134

Chapter 6. Convolution equations and the transport equation =

(I − Q∗ K)−1 I − Q∗ K − (I − Q∗ )(I − iλT )−1 K

=

I − (I − Q∗ K)−1 (I − Q∗ )(I − iλT )−1 K.

For λ = 0, the last expression comes down to I − (I − Q∗ K)−1 (I − Q∗ )K and this is easily seen to be equal to (I − QK ∗ )−1 (I − K), so to I − P K. The latter was earlier identiﬁed as the value G+ (0) of G+ in the origin. So here the zero value of λ is admissible too. −1 Let us now deal with G−1 − and G+ . The ﬁrst of these functions is tied to W+−1 for which we have the expression W+ (λ)−1 = I + KT × (λ − T × )−1 Q. From this we get −1 ∗ −1 G−1 W+ (i/λ)∗ − (λ) = (I − Q K) ∗ = (I − Q∗ K)−1 I + KT ×(i/λ − T × )−1 Q ∗ = (I − Q∗ K)−1 I + Q∗ (i/λ − T × )−1 (T × )∗ K =

−1 × ∗ (I − Q∗ K)−1 I + iλQ∗ I − iλ(T × )∗ (T ) K

=

−1 K (I − Q∗ K)−1 I − Q∗ K + Q∗ I − iλ(T × )∗

=

−1 I + (I − Q∗ K)−1 Q∗ I − iλ(T × )∗ K.

Putting λ = 0 in the last expression gives I + (I − Q∗ K)−1 Q∗ K which is obviously equal to (I − Q∗ K)−1 , the value of G−1 − at the origin. Finally we consider G−1 . For the appropriate values of λ, we have + −1 −1 W− (i/λ)∗ G−1 (I − Q∗ K) + (λ) = (I − K) = = =

∗ −1 (I − K)−1(I − Q∗ K) (I − K)−1 W− (i/λ) (I − K)

[∗] −1 W− (i/λ) (I − P K)−1

−1 [∗] W− (i/λ) (I − P K)−1 .

Here we have used (6.35) and the fact, already noted above, that I − P K and (I − K)−1 (I − Q∗ K) are each other’s inverse. Recall now that W− (λ)−1 = I + KT ×(I − Q)(λ − T × )−1 . Thus, as T × and K are [· , ·]-selfadjoint, [∗] G−1 I + KT ×(I − Q)(i/λ − T × )−1 (I − P K)−1 + (λ) = =

[∗]

I + (1/iλ − T × )−1 (I − Q)

T × K (I − P K)−1

6.5. The matching of the subspaces

135

= I + iλ(I − iλT × )−1 I − Q[∗] T × K (I − P K)−1 . As an intermediate step, we note that the identity in Lemma 6.7 can be rewritten as I − Q[∗] T × = T × P . Indeed, T ×P

= (I − K)−1 T P = (I − K)−1 (I − Q∗ )T = (I − K)−1 (I − Q)∗ (I − K)T × = (I − Q)[∗] T × = I − Q[∗] T × .

This makes it possible to proceed as follows: G−1 I + iλ(I − iλT × )−1 T × P K (I − P K)−1 + (λ) = = I − P K + (I − iλT × )−1 P K (I − P K)−1 = I + (I − iλT × )−1 P K(I − P K)−1 . The check for λ = 0 yields the desired result, namely I + P K(I − P K)−1 = (I − P K)−1 which is the value of G−1 + at the origin.

6.5 The matching of the subspaces In the canonical factorization carried out in the previous section, we used that ˙ Hp , H = H− +

˙ Hm . H = H+ +

(6.45)

In this section we shall prove that, indeed, the space H may be decomposed in these two ways. Let P− and P+ be the orthogonal projections of H onto H− and H+ , respectively. Also, put Pm = E × (0) and Pp = I − E × (0), where E × (t) is the spectral resolution of the identity for the operator T × = (I − K)−1 T with respect to the inner product [f, g] = (I − K)f, g). By deﬁnition H− = Im P− ,

H+ = Im P+ ,

Hm = Im Pm ,

Hp = Im Pp .

We claim that ˙ Hp H = H− +

⇐⇒ P+ |Hp : Hp → H+ is bijective,

(6.46)

˙ Hm H = H+ +

⇐⇒ P− |Hm : Hm → H− is bijective.

(6.47)

The argument for this is simple and in a diﬀerent context (involving a diﬀerent notation too) spelled out in the beginning of Part 2 of the proof of Theorem 4.4.

136

Chapter 6. Convolution equations and the transport equation

For the convenience of the reader we give it here too. Note that Ker P+ |Hp = H− ∩ Hp , and thus P+ |Hp is injective if and only if H− ∩ Hp = {0}. Next, observe that for each y ∈ Hp we have y = (I − P+ )y + P+ |Hp y ∈ H− + Im P+ |Hp . Thus H− + Hp ⊂ H− + Im P+ |Hp . The reverse inclusion also holds. Indeed, for z ∈ Hp wehave P+ z = (P+ z − z) + z ∈ Ker P+ + Hp = H− + Hp . It follows that H− + Im P+ |Hp = H− + Hp , and hence P+ |Hp is surjective if and only if H = ˙ Hp H− + Hp . This proves (6.46). The proof of (6.47) is similar. Now H = Hm + ˙ H+ . Combining this with (6.46) and (6.47), we see that (6.45) and H = H− + holds if and only if the operator V = P− Pm + P+ Pp is bijective. It is not diﬃcult to prove that V is injective. Indeed, take f ∈ H and assume V f = 0. Put fm = Pm f and fp = Pp f . Then P− fm + P+ fp = V f = 0, and hence P+ fp = 0 and P− fm = 0. The latter gives fm = P+ fm , and we get 0 ≥ [T × fm , fm ] = T fm , fm = T P+ fm , P+ fm ≥ 0. It follows that P+ fm ∈ Ker T . But T is injective. So P+ fm = 0. As P− fm = 0 too, we have fm = 0. In the same way one proves that fp = 0. Hence f = 0, as desired. To prove that V is surjective too, we use that I − V is compact. Indeed, as soon as we know that this is the case, the Fredholm alternative implies that V = I − (I − V ) is surjective if and only if V is injective. Lemma 6.8. The operator I − V is compact. Proof. The compact operators form an ideal and I −V

= P− + P+ − P− Pm − P+ Pp = P− + P+ Pm + P+ Pp − P− Pm − P+ Pp = P− + P+ Pm − P− Pm = (P+ − P− ) (Pm − P− ) .

Hence it suﬃces to prove that Pm − P− is compact. Now Pm = E × (0), where E × (t) is the spectral resolution of the identity for T × with respect to the inner product [· , ·]. Similarly, P− = E(0), where E(t) is the spectral resolution of the identity for T . As T and T × are injective, in both cases the spectral resolutions are continuous at zero. So, using a standard formula for the spectral resolution (see [99], Problem VI.5.7) we may write, for each f ∈ H, 1 (Pm − P− )f = lim h ↓ 0 2πi

Γh

(λ − T )−1 − (λ − T × )−1 f dλ.

(6.48)

Here h is a (suﬃciently small) positive number and Γh is the union of two nonclosed oriented curves as in the following picture:

6.5. The matching of the subspaces

137

+ih −a

0 −i h

The positive number a is chosen in such a way that the spectra of T and T × both are in the open half-line (−a, ∞). For the diﬀerence of the resolvents of T and T × appearing in (6.48) we have (λ − T )−1 − (λ − T × )−1

=

(λ − T )−1 I − (λ − T )(λ − T × )−1

=

(λ − T )−1 (T − T × )(λ − T × )−1

=

−(λ − T )−1 KT ×(λ − T × )−1 ,

and from the latter expression we see that it is a ﬁnite rank (hence compact) operator. Let Δ be the closed contour obtained from Γh by letting the positive number h go to zero. As T × is selfadjoint in H endowed with the inner product [· , ·], we know from (6.32) in Lemma 6.4 and the choice of a that T × (λ − T × )−1 is bounded in norm on Δ \ {0}. Next, let us investigate (λ − T )−1 K. First we shall prove that q0 (ic − T )−1 K ≤ " , |c|

0 = c ∈ R,

(6.49)

where q is some positive constant. To prove this, recall that K is the ﬁnite rank operator given by the right-hand side of (6.7), and hence (ic − T )−1 K ≤

n

|aj | pj (ic − T )−1 pj ,

0 = c ∈ R.

j=0

For each j the function pj is a normalized Legendre polynomial in t (and so the norm of pj appearing in the above expression is actually equal to 1). Also T is the multiplication operator given by the left-hand side of (6.7). So to ﬁnd an upper bound for (ic − T )−1 pj , we need to estimate #

1 −1

c2

t2k dt . + t2

(6.50)

138

Chapter 6. Convolution equations and the transport equation

As t2k+2 ≤ t2k for |t| ≤ 1, it suﬃces to ﬁnd an upper bound for (6.50) for the case k = 0. But # # 1 dt 2 1 = arctan , 0 = c ∈ R. 2 2 |c| |c| −1 c + t This proves (6.49) for an appropriate choice of q0 . Note that the function (λ − T )−1 KT ×(λ − T × )−1 is continuous on Δ \ {0}. Also, for some positive constant q, q 0= c ∈ R. (ic − T )−1 KT × (ic − T × )−1 ≤ " , |c| A straightforward Cauchy argument now gives that (λ − T )−1KT × (λ − T × )−1 dλ lim h↓0

Γh

exists in norm. But then the same is true for (λ − T )−1 − (λ − T × )−1 dλ. lim h↓0

Γh

As the integrand in this expression is a compact operator-valued function, we can use (6.48) to show that Pm − P− is compact too. Close inspection of the above proof shows that I − V is in fact a trace class operator (cf., Lemma 6.3 in [11]).

6.6 Formulas for solutions Let I − H(λ) be the symbol of the Wiener-Hopf integral equation (6.10). From the results of the previous sections we know that I − H(λ) admits a right canonical factorization with respect to the real line: I − H(λ) = G− (λ)G+ (λ),

−∞ < λ < ∞.

(6.51)

As we have seen in Section 6.3, this implies that equation (6.10) is uniquely solvable in L1 ([0, ∞), H), where H = L2 [−1, 1]. This fact and the equivalence (explained in the ﬁrst section of this chapter) of equations (6.1) and (6.10), allows us to prove the following result. Theorem 6.9. Consider equation (6.1) with the kernel function k being given by (6.4). Let T and K be the operators on L2 [−1, 1] deﬁned by (6.7), and assume that I − K is strictly positive. Then equation (6.1) has a unique solution ψ satisfying the initial condition (6.2) and ∞ 1 |ψ(t, μ)|2 dμ dt < ∞. (6.52) 0

−1

6.6. Formulas for solutions

139

This solution is given by −1

× ψ(t, ·) = e−t(Tp ) P f+ ,

t ≥ 0.

(6.53)

Here f+ is the given function appearing in the initial condition (6.2), the operator P is the projection of L2 [−1, 1] deﬁned directly after (6.38), and Tp× is the restriction of T × = (I − K)−1 T to Hp = Im P . Note that (6.53) is the natural analogue of (6.25) in Theorem 6.3. Formula (6.53) features the inverse of the injective operator Tp× = T × |Hp : Hp → Hp . This operator has dense range and with respect to the inner product is nonnegative [· , ·]. Hence its inverse (Tp× )−1 Hp → Hp is an unbounded operator which has Im Tp× as its (dense) domain and is nonnegative with regard to the inner product [· , ·]. Thus the expression × −1 (6.54) e−t(Tp ) is well-deﬁned via the operational calculus for selfadjoint operators based on the notion of the resolution of the identity. One can view (6.54) also as the operator semigroup generated by the unbounded inﬁnitesimal generator −(Tp× )−1 . Proof. Recall that I −H(λ) is the symbol of equation (6.10). Since I −H(λ) admits the canonical Wiener-Hopf factorization (6.51) we can use the general theory of Wiener-Hopf equations (see the one but last paragraph in Section 6.3) to show that equation (6.10) has a unique solution ψ in L1 ([0, ∞), H), where H = L2 [−1, 1]. Moreover, the Fourier transform ψ of ψ is given by −1 ψ(λ) = G−1 (6.55) + (λ)P G− (λ)F (λ) , where F is the Fourier transform of the right-hand side of equation (6.10), and P is the projection deﬁned by ∞ ∞ itλ e f (t) dt = eitλ f (t) dt. P −∞

0

Since ψ ∈ L1 ([0, ∞), H), condition (6.52) is fulﬁlled. To derive formula (6.53), we ﬁrst compute ψ using equation (6.55). Recall that F is given by (6.9). It follows that F(λ) = (I − iλT )−1 T f+ ,

λ ≥ 0.

(6.56)

As we know from Section 6.4 the inverses of the factors G− (λ) and G+ (λ) in (6.51) are given by −1 ∗ −1 ∗ G−1 Q I − iλ(T × )∗ K, − (λ) = I + (I − Q K) × −1 G−1 P K(I − P K)−1 . + (λ) = I + (I − iλT )

140

Chapter 6. Convolution equations and the transport equation

Here T × = (I − K)−1 T . Let us use these formulas to compute ψ(λ) from (6.55). −1 As a ﬁrst step we have G− (λ)F(λ) = F (λ) + X(λ)F(λ), where −1 K X(λ) = (I − Q∗ K)−1 Q∗ I − iλ(T × )∗ =

−1 (I − Q∗ K)−1 Q∗ I − iλT (I − K)−1 K

=

−1 K. (I − Q∗ K)−1 Q∗ (I − K) I − K − iλT

Thus X(λ)F (λ) = (I − Q∗ K)−1 Q∗ (I − K)R(λ)T f+ , where R(λ)

(I − K − iλT )−1 K(I − iλT )−1 = (I − K − iλT )−1 (I − iλT ) − (I − K − iλT ) (I − iλT )−1

=

=

(I − K − iλT )−1 − (I − iλT )−1 .

Hence X(λ)F (λ) = (I − Q∗ K)−1 Q∗ (I − K)(I − K − iλT )−1 T f+ −(I − Q∗ K)−1Q∗ (I − K)(I − iλT )−1 T f+ . We conclude that ∗ −1 ∗ G−1 Q (I − K)F (λ) − (λ)F (λ) = F (λ) − (I − Q K) −1 T f+ . + (I − Q∗ K)−1 Q∗ I − iλ(T × )∗

Now apply the projection P. Since f+ ∈ H+ and T |H+ is nonnegative, we have P(F) = F. Furthermore, using the spectral properties of T × and the deﬁnition of Q, we see that the function Q∗ (I − iλ(T × )∗ )−1 is annihilated by P. Therefore P G−1 = F(λ) − (I − Q∗ K)−1 Q∗ (I − K)F (λ) − (λ)F (λ) = (I − Q∗ K)−1 I − Q∗ K − Q∗ (I − K) F(λ) = (I − Q∗ K)−1 (I − Q∗ )F (λ). Put Z(λ) = (I − Q∗ K)−1 (I − Q∗ )F (λ). Recall from the previous section that I − P K is invertible with inverse (I − K)−1 (I − Q∗ K). Hence (I − P K)−1 (I − Q∗ K)−1 = (I − K)−1 , −1 and it follows that G−1 + (λ)P G− (λ)F (λ) = Z(λ) + H(λ), where

=

(I − iλT × )−1 P K(I − K)−1 (I − Q∗ )F (λ) (I − iλT × )−1 P I − (I − K) (I − K)−1 (I − Q∗ )F (λ)

=

A(λ) − B(λ),

H(λ) =

6.6. Formulas for solutions

141

with A(λ) = (I − iλT × )−1 P (I − K)−1 (I − Q∗ )T (I − iλT )−1f+ , B(λ) = (I − iλT × )−1 P (I − Q∗ )T (I − iλT )−1 f+. Using (I − Q∗ )T = T P and T × = (I − K)−1 T we get A(λ)

=

(I − iλT × )−1 P (I − K)−1 T P (I − iλT )−1 f+

=

(I − iλT × )−1 P T × P (I − iλT )−1 f+

=

(I − iλT × )−1 T × P (I − iλT )−1 f+ ,

and B(λ)

=

(I − iλT × )−1 P T P (I − iλT )−1 f+

=

(I − iλT × )−1 P T (I − iλT )−1 f+ .

Thus (with λ = 0 in the intermediate steps) H(λ)

= (I − iλT × )−1 (T × P − P T )(I − iλT )−1 f+ =

1 (I − iλT × )−1 P (I − iλT ) − (I − iλT × )P (I − iλT )−1 f+ iλ

=

1 1 (I − iλT × )−1 P f+ − P (I − iλT )−1 f+ iλ iλ

=

1 1 P f+ + T × (I − iλT × )−1 P f+ − P f+ − P T (I − iλT )−1 f+ iλ iλ

= T × (I − iλT × )−1 P f+ − P T (I − iλT )−1 f+ . Therefore −1 × × −1 G−1 P f+ + (λ)P G− (λ)F (λ) = T (I − iλT ) + (I − Q∗ K)−1 (I − Q∗ ) − P T (I − iλT )−1 f+ . From Lemma 6.5 we get (I − Q∗ K)−1 (I − Q∗ ) − P

= (I − Q∗ K)−1 (I − Q∗ − P + Q∗ KP ) = (I − Q∗ K)−1 (I − Q∗ − P + Q∗ P ) = (I − Q∗ K)−1 (I − Q∗ )(I − P ) = 0,

and we conclude that −1 × × −1 ψ(λ) = G−1 P f+ . + (λ)P G− (λ)F (λ) = T (I − iλT )

(6.57)

142

Chapter 6. Convolution equations and the transport equation

Now T × maps Hp into Hp , and so the operator Tp× = T × |Hp : Hp → Hp is well-deﬁned. Since Tp× is injective, the expression (6.57) can be rewritten as −1 P f+ . ψ(λ) = − iλ − (Tp× )−1

(6.58)

As was already observed, (Tp× )−1 Hp → Hp is an unbounded operator which has Im Tp× as its (dense) domain and is nonnegative with regard to the inner product [· , ·]. Hence we can take the inverse Fourier transform in (6.58), to get the desired formula (6.53). From (6.53) we see that ψ(0) = ψ(0, · ) = P f+ . Now let P+ be the orthogonal projection of L2 [−1, 1] onto H+ . Since Ker P = H− , we have P+ (I − P ) = 0, and thus P+ ψ(0) = P+ (P f+ ) = P+ (P f+ + (I − P )f+ ) = P+ f+ = f+ . Therefore ψ satisﬁes the initial condition (6.2). Finally, the uniqueness statement follows from the general theory of Wiener-Hopf equations.

Notes The theory of the linear transport equation has a long history. For this see the books [28] and [96] which also contain extensive lists of references. The material in Section 6.2 is taken from Section XIII.9 of [51] where the reader can also ﬁnd an illustrative example. The other sections in this chapter follow basically Chapter 6 in [11] which was inspired by the dissertation [80] and the papers [81], [82]. In [108] one can also ﬁnd an analytic description of the subspaces concerned. Later results based on [110] and [124] are also included here. Further developments using the method described in this chapter can be found in [110], where the case of non-degenerate kernel functions k(μ, μ ) is treated. See also the book [78], and the paper [124]. For an alternative proof of Theorem 6.9, not using Wiener-Hopf factorization, we refer to Section XIX.7 in [51]. The results presented in Sections 6.4 – 6.6 can also be understood from the point of view described in Chapter 5. Note, however, that in Sections 6.4–6.6 the symbol is an operator-valued function (and not a matrix-valued function as in Chapter 5). On the other hand, the operator (T × )−1 appearing in Theorem 6.9 is exponentially dichotomous. This has been proved in Section 5.2 of the recent monograph [111]. The latter book also contains many new additions related to the analysis of equation (6.6). See also the notes to Chapter 5.

Chapter 7

Wiener-Hopf factorization and factorization indices This chapter concerns canonical as well as non-canonical Wiener-Hopf factorization of an operator-valued function which is analytic on a Cauchy contour. Such an operator function is given by a realization with a possibly inﬁnite dimensional Banach space as state space, and with a bounded state operator and with bounded input-output operators. The ﬁrst main result is a generalization to operator-valued functions of the canonical factorization theorem for rational matrix functions presented earlier in Section 3.1. In terms of the given realization, necessary and sufﬁcient conditions are also presented in order that the operator function involved admits a (possibly non-canonical) Wiener-Hopf factorization. The corresponding factorization indices are described in terms of certain spectral invariants which are deﬁned in terms of the realization but do only depend on the operator function and not on the particular choice of the realization. The analysis of these spectral invariants is one of the main themes of this chapter. The chapter consists of three sections. Section 7.1 describes the main result for canonical factorization and introduces the spectral invariants involved. The proof that the spectral invariants do not depend on the particular realization is given in Section 7.2. The ﬁnal section of the chapter, Section 7.3, deals with noncanonical Wiener-Hopf factorization and the corresponding factorization indices.

7.1 Canonical factorization of operator functions Throughout this chapter, W is an operator function, analytic on an open neighborhood of a given Cauchy contour Γ, and with values that are operators on a possibly inﬁnite dimensional Banach space Y . Anticipating the results to be presented below, we note that in this situation W admits a realization on Γ involving

144

Chapter 7. Wiener-Hopf factorization and factorization indices

a possibly inﬁnite dimensional state space X and having IY as external operator: W (λ) = IY + C(λIX − A)−1 B,

(7.1)

where Γ splits the spectrum of A, that is Γ ⊂ ρ(A). This is immediate from Theorem 2.2. As before, we denote by F+ the interior domain of Γ, and by F− the complement of F + in the Riemann sphere C∞ . By a right canonical factorization of W with respect to Γ we mean a factorization W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

(7.2)

where W− and W+ are functions with values in L(Y ) satisfying (i) W− is analytic on F− and continuous on F − , (ii) W+ is analytic on F+ and continuous on F + , (iii) W− and W+ take invertible values on F − and F + , respectively. If in (7.2) the factors W− and W+ are interchanged, we speak of a left canonical factorization. A necessary condition for a right or left canonical factorization with respect to Γ to exist is that W takes invertible values on Γ. In terms of the realization (7.1) this means that Γ also splits the spectrum of the associate main operator A× = A − BC (see Theorem 2.4). We now extend Theorem 3.2 to a possibly inﬁnite dimensional context. Theorem 7.1. Let W be an operator function, analytic on an open neighborhood of a Cauchy contour Γ, and with values that are operators on a Banach space Y . Let (7.1) be a realization of W , i.e., W (λ) = IY + C(λIX − A)−1 B, and suppose Γ splits the spectrum of A. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (a) Γ splits the spectrum of A× = A − BC, ˙ Ker P (A× ; Γ). (b) X = Im P (A; Γ) + In that case, a right canonical factorization of W is given by W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

where the factors and their inverses can be written as W− (λ)

=

Im + C(λIX − A)−1 (IX − Π)B,

W+ (λ)

=

Im + CΠ(λIX − A)−1 B,

W−−1 (λ)

=

Im − C(IX − Π)(λIX − A× )−1 B,

W+−1 (λ)

=

Im − C(λIX − A× )−1 ΠB.

Here Π is the projection of Cn along Im P (A; Γ) onto Ker P (A× ; Γ).

7.1. Canonical factorization of operator functions

145

For left canonical factorizations an analogous theorem holds. In the result in ˙ Im P (A× ; Γ). The theorem also question, (b) is replaced by X = Ker P (A; Γ) + has an analogue for appropriate closed contours in the Riemann sphere C∞ like the extended real line or the extended imaginary axis. Proof. To establish the theorem, we can rely for a large part on the proof of Theorem 3.2. In fact, we only have to add an argument for the following assertion: if W admits a right canonical factorization with respect to Γ, then the decomposition in (b) holds. The ﬁrst step consists in showing that if W admits a right canonical factorization with respect to Γ, then there is a way of representing W in the form −1 B such that Γ splits the spectra of A and A × while, W (λ) = IY + C(λI − A) X × = Im P (A; Γ) + ; Γ). ˙ Ker P (A in addition, X Let W (λ) = W− (λ)W+ (λ), λ ∈ Γ, be a right canonical factorization of W . Recall that ∞ belongs to F− . Since W− (∞) is invertible we may assume without loss of generality that W− (∞) = IY . From the identity W− (λ) = W (λ)W+ (λ)−1 and the fact that W is analytic on a neighborhood of Γ, it follows that W− has an analytic extension, again denoted by W− , to some open neighborhood Ω− of the closed set F− ∪ Γ. Taking Ω− suﬃciently small, we have that W− assumes only invertible values on Ω− . But then Theorems 2.3 and 2.4 can be applied to show that W− admits a realization of the form W− (λ) = IY + C− (λIX− − A− )−1 B− ,

λ ∈ Ω− ,

(7.3)

× where σ(A− ) ⊂ F+ and σ(A× − ) ⊂ F+ . Here A− = A− − B− C− . A similar reasoning holds for W+ . This function has an analytic extension, again denoted by W+ , to some open neighborhood Ω+ of the closed set F+ ∪ Γ. Taking Ω+ suﬃciently small, we have that W+ (λ) is invertible for all λ ∈ Ω+ . But then Theorems 2.2 and 2.4 yield that W+ admits a realization

W+ (λ) = IY + C+ (λIX+ − A+ )−1 B+ ,

λ ∈ Ω+ ,

(7.4)

× such that σ(A+ ) ⊂ F− and σ(A× + ) ⊂ F− . Here A+ = A+ − B+ C+ . On Γ we have the factorization W (λ) = W− (λ)W+ (λ), and so we can apply λ ∈ Γ, −1 B, the product rule of Section 2.5 to show that W (λ) = IY + C(λI − A) X where X = X− X+ and A : X → X, B : Y → X and C : X → Y are given by the operator matrices

= A

A−

B− C+

0

A+

,

= B

B− B+

,

= C

C−

C+

.

−1 B has the desired properties. This can be seen The realization IY + C(λI − A) X as follows.

146

Chapter 7. Wiener-Hopf factorization and factorization indices

˜ and the corresponding one From the operator matrix representation for A, × for A = A − BC : X → X, namely A× 0 − × = , A −B+ C− A× + and A × . Furthermore, the it is immediate that Γ splits the spectra of both A × spectral projections P (A; Γ) and P (A , Γ) are of the form IX− 0 IX− × , Γ) = Γ) = , P (A . P (A; 0 0 0 Γ) = X− {0} and Ker P (A × ; Γ) = {0} X+ , and from this Hence Im P (A; = Im P (A; Γ) + × ; Γ) is immediate. ˙ Ker P (A X The proof can now be ﬁnished by verifying the following two identities: Γ) ∩ Ker P (A × ; Γ) = dim Im P (A; Γ) ∩ Ker P (A× ; Γ) , dim Im P (A; % $ X X = dim . dim Γ) + Ker P (A × ; Γ) Im P (A; Γ) + Ker P (A× ; Γ) Im P (A; In other words, we are ready once it has been shown that the right-hand side of these identities depend only on W and Γ and are independent of the realization (7.1) of W . This is indeed the case as is seen from Theorem 7.2 below which even exhibits several other spectral invariants. Theorem 7.2. Let W be an operator function, analytic on an open neighborhood of a Cauchy contour Γ, and with values that are operators on a Banach space Y . Let (7.1) be a realization of W , i.e., W (λ) = IY + C(λIX − A)−1 B, and suppose Γ splits the spectrum of A. In addition, assume that Γ also splits the spectrum of A× = A − BC. Introduce P = P (A; Γ),

M = Im P,

Then the quantities ×

dim(M ∩ M ), dim dim

dim

P × = P (A× ; Γ),

M × = Ker P × .

X , M + M×

M ∩ M × ∩ Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk−1 M ∩ M × ∩ Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk M + M × + Im B + Im AB + · · · + Im AB k M + M × + Im B + Im AB + · · · + Im AB k−1

,

k = 0, 1, 2 . . . ,

,

k = 0, 1, 2 . . . ,

depend on W only and do not depend on the realization (7.1) of W .

7.2. Proof of Theorem 7.2

147

The theorem has an analogue for appropriate closed contours in the Riemann sphere C∞ like the extended real line or the extended imaginary axis. To put Theorem 7.2 in context, consider a proper rational matrix function W having the value Im at inﬁnity. With a realization W (λ) = Im + C(λIn − A)−1 B of W , one can associate the numbers k = 0, 1, 2, . . . , (7.5) dim Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk−1 , codim Im B + Im AB + · · · + Im AB k−1 , k = 0, 1, 2, . . . . (7.6) Here the codimension is taken with respect to Cn . Now realizations of rational matrix functions are not unique and the above numbers, as well as their diﬀerences, generally vary with diﬀerent choices of A, B and C in the realization for W . The above theorem shows that this dependence on the speciﬁc form of (7.1) disappears when one combines the spaces appearing in (7.5) and (7.6) with certain spectral subspaces of A and A× . We will meet the subspaces featuring in (7.5) and (7.6) again in Section 8.1. The proof of Theorem 7.2 is rather complicated and we will devote a separate section to it.

7.2 Proof of Theorem 7.2 Let W and Γ be as in Theorem 7.2, and suppose we have the realizations −1 B, W (λ) = IY + C(λI − A) X

(7.7)

−1 B, W (λ) = IY + C(λI − A) X

(7.8)

and A × . In other and A × as well as those of A where Γ splits the spectra of A × × words Γ ⊂ ρ(A) ∩ ρ(A ) ∩ ρ(A) ∩ ρ(A ). Writing Γ), P = P (A;

! = Im P , M

× ; Γ), P× = P (A

!× = Ker P × , M

Γ), P = P (A;

M = Im P ,

× ; Γ), P × = P (A

M × = Ker P × ,

we need to show that !∩M !× ) = dim(M ∩ M × ), dim(M $ % $ % X X dim = dim , !+M !× M M + M× % $ ∩ Ker C A ∩ · · · ∩ Ker C A k−1 !∩M !× ∩ Ker C M dim ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k !∩M !× ∩ Ker C M $ % ∩ Ker C A ∩ · · · ∩ Ker C A k−1 M ∩ M × ∩ Ker C = dim , ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k M ∩ M × ∩ Ker C

148

Chapter 7. Wiener-Hopf factorization and factorization indices $

% !+M !× + Im B + Im A B + · · · + Im A B k−1 + Im A B k M dim + Im A B + · · · + Im A B k−1 !+M !× + Im B M % $ + Im A B + · · · + Im A B k−1 + Im A B k M + M × + Im B . = dim + Im A B + · · · + Im A B k−1 M + M × + Im B Here k = 0, 1, 2, . . . . It is convenient to ﬁrst present a series of auxiliary results. These concern and Ψ given by the integrals the operators Ψ = 1 × )−1 B C(λ − A) −1 dλ, Ψ (λ − A (7.9) 2πi Γ 1 × )−1 B C(λ − A) −1 dλ. (λ − A (7.10) Ψ = 2πi Γ :X →X and Ψ :X → X. Note that Ψ and Ψ also admit the representation: Lemma 7.3. The operators Ψ C(λ −A × )−1 dλ, = 1 −1 B Ψ (λ − A) 2πi Γ −1 B C(λ −A × )−1 dλ, = 1 (λ − A) Ψ 2πi Γ

(7.11) (7.12)

Proof. From Theorem 2.4 we know that − A) −1 , −A × )−1 = C(λ W (λ)C(λ

−A × )−1 = C(λ − A) −1 , W (λ)C(λ

× )−1 BW (λ) = (λ − A) −1 B, (λ − A

× )−1 BW (λ) = (λ − A) −1 B. (λ − A

Now make the appropriate substitutions. and Ψ the following identities hold: Lemma 7.4. For the products of Ψ Ψ = (P× − P )2 , Ψ

Ψ = (P× − P )2 . Ψ

∩ ρ(A). For λ ∈ Γ, we have Proof. It is assumed that Γ ⊂ ρ(A) −1 B −1 B. = C(λ − A) −1 (μ − A) − A) −1 (μ − A) C(λ Indeed, taking advantage of the resolvent identity, we get for λ ∈ Γ, − A) −1 (μ − A) −1 B (μ − λ)C(λ −1 B (λ − A) −1 − (μ − A) =C − C(μ − A) −1 B − A) −1 B = C(λ

(7.13)

7.2. Proof of Theorem 7.2

149 = W (λ) − I − W (μ) − I − A) −1 B − C(μ − A) −1 B = C(λ −1 B (λ − A) −1 − (μ − A) =C −1 B. − A) −1 (μ − A) = (μ − λ)C(λ

Now, when λ = μ, divide by μ − λ; for λ = μ, employ a continuity argument. Ψ, we use the expression (7.10) for Ψ, formula (7.11) for Ψ, To compute Ψ and the identity (7.13): 2 1 × )−1 B C(λ − A) −1 ΨΨ = (λ − A 2πi Γ Γ C(μ −A × )−1 dλ dμ −1 B ·(μ − A) =

1 2πi

2 Γ

Γ

× )−1 B C(λ − A) −1 (λ − A C(μ −A × )−1 dλ dμ −1 B ·(μ − A)

=

1 2πi

2 Γ

Γ

× )−1 (A −A × )(λ − A) −1 (λ − A −1 (A −A × )(μ − A × )−1 dλ dμ ·(μ − A)

=

=

=

1 2πi

1 2πi

2 Γ

Γ

Γ

−1 ) × )−1 − (λ − A) (λ − A × )−1 − (μ − A) −1 dλ dμ · (μ − A

2 × −1 −1 (λ − A ) − (λ − A) dλ

(P× − P )2 .

Ψ = (P× − P )2 , interchange the roles of the realizations (7.7) and (7.8). For Ψ and Ψ satisfy the following intertwining relations: Lemma 7.5. The operators Ψ P = (I − P× )Ψ, Ψ

P × = (I − P )Ψ, Ψ

(7.14)

P = (I − P × )Ψ, Ψ

P × = (I − P )Ψ. Ψ

(7.15)

Proof. Focussing on the ﬁrst identity in (7.14), note that the function × )−1 B C(λ − A) −1 P P × (λ − A

150

Chapter 7. Wiener-Hopf factorization and factorization indices

is analytic on an open neighborhood of F− ∪ Γ. Here F− is the exterior domain of Γ (including ∞). Furthermore, the expansion of this function at inﬁnity is of the C P plus lower order terms. Hence form λ−2 P× B 1 × )−1 B C(λ − A) −1 Pdλ = 0. P× (λ − A 2πi Γ C(λ − A) −1 (I − P) is analytic on an × )−1 B On the other hand (I − P × )(λ − A open neighborhood of F+ ∪ Γ, where F+ is the interior domain of Γ, and so 1 × )−1 B C(λ − A) −1 (I − P)dλ = 0. (I − P × )(λ − A 2πi Γ It follows that P Ψ

= = =

1 2πi 1 2πi 1 2πi

Γ

Γ

Γ

× )−1 B C(λ − A) −1 P dλ (λ − A × )−1 B C(λ − A) −1 P dλ (I − P × )(λ − A × )−1 B C(λ − A) −1 dλ (I − P × )(λ − A

= (I − P × )Ψ, as desired. This proves the ﬁrst identity in (7.14). The second identity in (7.14) is proved given by (7.11). The identities in (7.15) in a similar way using the formula for Ψ follow from those in (7.14) by interchanging the roles of the realizations (7.7) and (7.8). and Ψ satisfy the following Lyapunov equations: Lemma 7.6. The operators Ψ A −A × Ψ Ψ

=

C P − P× B C, B

(7.16)

Ψ A × − A Ψ

=

C P× − PB C, B

(7.17)

A −A × Ψ Ψ

=

C P − P× B C, B

(7.18)

Ψ A × − A Ψ

=

C P× − PB C. B

(7.19)

via (7.9), we have Proof. Using the deﬁnition of Ψ 1 A = × )−1 B C(λ − A) −1 A dλ Ψ (λ − A 2πi Γ

7.2. Proof of Theorem 7.2 1 2πi

=

Γ

151

× )−1 B − λI + λI) dλ C(λ − A) −1 (A (λ − A

1 × −1 −1 × )−1 B C dλ λ(λ − A ) B C(λ − A) dλ − (λ − A 2πi Γ Γ 1 × + A × )(λ − A × )−1 B C(λ − A) −1 dλ − P × B C (λI − A 2πi Γ 1 2πi

=

=

− P × B C. C P + A × Ψ B

=

This gives (7.16). The identity (7.17) can be proved similarly by using the alter of Lemma 7.3. For (7.18) and (7.19), use (7.16) and (7.17) native expression for Ψ and interchange the roles of of the realizations (7.7) and (7.8). Direct computations as the one above of course also work. Ψ, B, B, C and C are related as follows: Lemma 7.7. The operators Ψ, B = (P − P × )B, Ψ

Ψ = C( P − P × ), C

(7.20)

B = (P − P × )B, Ψ

Ψ = C( P − P × ). C

(7.21)

we have Proof. Using the expression (7.9) for Ψ, B Ψ

=

=

=

=

1 2πi

Γ

1 2πi 1 2πi 1 2πi

Γ

Γ

Γ

× )−1 B C(λ − A) −1 B dλ (λ − A × )−1 B W1 (λ) − I dλ (λ − A × )−1 B W2 (λ) − I dλ (λ − A

× )−1 BW 2 (λ) dλ − (λ − A

1 2πi

Γ

× )−1 B dλ. (λ − A

(λ) by (λ − A) −1 B. Hence × )−1 BW By Theorem 2.4, we may replace (λ − A B = Ψ

1 2πi

−1 B dλ − 1 (λ − A) 2πi Γ

Γ

× )−1 B dλ, (λ − A

and this can be rewritten as the ﬁrst part of (7.20). The second part can be proved via a similar computation. The identities in (7.21) follow by interchanging the roles of the realizations (7.7) and (7.8).

152

Chapter 7. Wiener-Hopf factorization and factorization indices

Proof of Theorem 7.2. The proof will be divided into three parts. The ﬁrst con!, M !× , M and M × , ending tains some preliminary observations about the spaces M up in an argument establishing the identities % $ % $ X X × × !∩M ! ) = dim(M ∩ M ), = dim . dim dim(M !+M !× M M + M× Part 1. We begin by noting that M !∩M !× ] ⊂ M ∩ M × , Ψ[

M !+M !× ] ⊂ M + M × , Ψ[

M ∩ M ×] ⊂ M !∩ M !× , Ψ[

M + M ×] ⊂ M !+M !× . Ψ[

To prove this, it suﬃces to show that M !× ⊂ M , Ψ

M ! ⊂ M ×, Ψ

M ⊂ M !× , Ψ

M× ⊂ M !. Ψ

These inclusions, however, are obvious from (7.14) and (7.15). Next observe that !∩M !× M !+M !× M

Ψ), ⊂ Ker (I − Ψ Ψ , ⊃ Im I − Ψ

Ψ), M ∩ M × ⊂ Ker (I − Ψ Ψ . M + M × ⊃ Im I − Ψ

(7.22) (7.23)

Ψ, follow from The formulas concerning the product Ψ Ψ = P P× + (I − P × )(I − P ), I −Ψ which, in turn, is immediate from Lemma 7.4. The two expressions involving the Ψ are obtained by interchanging the roles of the realizations (7.7) and product Ψ (7.8). Consider the restriction operators ! !× Ψ| M ∩M

!∩M !× → M ∩ M × , : M

Ψ| M ∩M ×

!∩M !× . : M ∩ M× → M

!∩ M !× From (7.22) it is clear that these operators are each others inverse. Hence M × and M ∩ M are linearly isomorphic and so they have the same (possibly inﬁnite) dimension. Next we turn to the operators : Φ

X !+M !× M

→

X M + M×

,

: Φ

X M + M×

→

X !+M !× M

,

and |Ψ, respectively. These are well-deﬁned because of the inclusions induced by Ψ M + M ×] ⊂ M !+ M !× . Also it follows from (7.23) M !+ M !× ] ⊂ M + M × and Ψ[ Ψ[

7.2. Proof of Theorem 7.2

153

and Φ are each other’s inverse. Thus the quotient spaces X/ M !+ M !× and that Φ M + M × are linearly isomorphic. In particular they have the same (possibly X/ inﬁnite) dimension. Part 2. In this part of the proof we shall verify that for all nonnegative integers k the following identities hold: $ % !∩M !× ∩ Ker C ∩ Ker C A ∩ · · · ∩ Ker C A k−1 M dim !∩M !× ∩ Ker C ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k M $ % ∩ Ker C A ∩ · · · ∩ Ker C A k−1 M ∩ M × ∩ Ker C = dim . ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k M ∩ M × ∩ Ker C This will be done by showing that the quotient spaces appearing in these identities are linearly isomorphic. To facilitate the discussion, we adopt the notation A) = Ker C ∩ Ker C A ∩ · · · ∩ Ker C A k−1 , Ker k (C| A) is read as X. Of course the where, following standard convention, Ker 0 (C| A) is deﬁned similarly. First we shall prove that the operator Ψ notation Ker k (C| × × ! ! maps M ∩ M ∩ Ker k (C|A) into M ∩ M ∩ Ker k (C|A). This has already been established for k = 0 (Part 1). For k = 1 it must be proved that M !∩M !× ∩ Ker C ⊂ M ∩ M × ∩ Ker C. Ψ M !∩M !× ⊂ M ∩ M × , and so it is enough to derive We know already that Ψ ⊂ Ker C or, what comes down to the same, M !∩M !× ∩ Ker C the inclusion Ψ !∩ M !× ∩ Ker C ⊂ Ker C Ψ. The latter, however, is immediate from the identity M × for which we refer to Lemma 7.7. C Ψ = −C(I − P ) − C P + C We proceed by induction. Let k be a nonnegative integer and suppose that maps M !∩ M !× ∩ Ker k (C| A) into M ∩ M × ∩ Ker k (C| A). We shall the operator Ψ show that the same is true with k replaced by k + 1. Clearly !∩M !× ∩ Ker k+1 (C| A) = M !∩M !× ∩ Ker k (C| A) ∩ Ker C A k , M and C replaced by M , M × , A and C, respectively. !, M !× , A and similarly with M maps Hence, in view of the induction hypothesis, it is suﬃcient to verify that Ψ × k ! ! the space M ∩ M ∩ Ker k+1 (C|A) into Ker C A . In other words, what we need is the inclusion ⊂ M !∩M !× ∩ Ker k+1 (C| A k Ψ A). Ker C (7.24) can be A k Ψ With the help of (the second identity in) Lemma 7.6, the operator C written as A k Ψ C

C P× + PB C +Ψ A × ) A k−1 (−B = C A k−1 − B C P × + (PB −Ψ B) C +Ψ A = C

154

Chapter 7. Wiener-Hopf factorization and factorization indices

and we may conclude that A k Ψ ∩ Ker C A k−1 Ψ ⊃ M !× ∩ Ker C A. Ker C

(7.25)

A). Employing the induction hypothesis A k−1 ⊃ M ∩ M × ∩ Ker k (C| Now Ker C k−1 M !∩ M !× ∩ Ker k (C| A) , i.e., A ⊃ Ψ once again gives Ker C ⊃ M !∩M !× ∩ Ker k (C| A k−1 Ψ A). Ker C But then A = A k−1 Ψ Ker C

−1 Ker C A k−1 Ψ A

⊃

!∩M !× ∩ Ker k (C| −1 M A) A

=

× A) ! ∩A −1 M ! ∩A −1 Ker k (C| −1 M A

⊃

× !∩ A +B C −1 [M !× ] ∩ A −1 Ker k (C| A) M

⊃ ⊃

A) ∩A −1 Ker k (C| !∩A ×−1 M × ∩ Ker C M !∩M !× ∩ Ker C ∩A −1 Ker k (C| A) , M

and hence, taking into account (7.25), ⊃ M !∩M !× ∩ Ker C ∩A −1 Ker k (C| A) . A k Ψ Ker C A)] = Ker k+1 (C| A), the inclusion (7.24) follows. ∩A −1 Ker k (C| As Ker C maps Fix the nonnegative integer k. As we have seen, the linear operator Ψ A) into M ∩ M × ∩ Ker k (C| A). Likewise Ψ maps the space !∩M !× ∩ Ker k (C| M A) into M !∩ M !× ∩ Ker k (C| A). The same is true with k replaced M ∩ M × ∩ Ker k (C| by k + 1. But then the linear operators k Θ

:

A) A) !∩M !× ∩ Ker k (C| M ∩ M × ∩ Ker k (C| M → , (7.26) A) A) !∩M !× ∩ Ker k+1 (C| M M ∩ M × ∩ Ker k+1 (C|

k Θ

:

A) A) !∩M !× ∩ Ker k (C| M ∩ M × ∩ Ker k (C| M → , (7.27) × × ! ! M ∩ M ∩ Ker k+1 (C|A) M ∩ M ∩ Ker k+1 (C|A)

and Ψ, respectively, are well-deﬁned. They are also each others induced by Ψ inverse. This can be deduced easily from !∩M !× ∩ Ker k (C| A) ⊂ M

Ψ), Ker (I − Ψ

A) ⊂ M ∩ M × ∩ Ker k (C|

Ψ), Ker (I − Ψ

7.2. Proof of Theorem 7.2

155

two inclusions which are immediate from (7.22). Thus the quotient spaces appearing in (7.26) and (7.27) are linearly isomorphic. In particular they have the same (possibly inﬁnite) dimension. Part 3. Finally we shall prove that the identities $ % !+M !× + Im B + Im A B + · · · + Im A B k−1 + Im A B k M dim + Im A B + · · · + Im A B k−1 !+M !× + Im B M % $ + Im A B + · · · + Im A B k−1 + Im A B k M + M × + Im B = dim + Im A B + · · · + Im A B k−1 M + M × + Im B are valid for all nonnegative integers k. This will be done by showing that the quotient spaces appearing in these identities are linearly isomorphic. To facilitate the discussion, we adopt the notation B) = Im B + Im A B + · · · + Im A B k−1 , Im k (A| B) is read as {0}. Of course the where, following standard convention, Im 0 (A| notation Im k (A|B) is deﬁned similarly. First we shall verify that the operator Ψ × × ! ! maps M + M + Im k (A|B) into M + M + Im k (A|B). This has already been established for k = 0 (Part 1). For k = 1 it must be proved that M !+M !× + Im B ⊂ M + M × + Im B. Ψ M !+M !× ⊂ M + M × , and so it is enough to derive We know already that Ψ or, what comes down to the same, Im B] ⊂ M + M × + Im B the inclusion Ψ × The latter, however, is immediate from the identity Im ΨB ⊂ M + M + Im B. −B for which we refer to Lemma 7.7. B = P B + (I − P × )B Ψ We proceed by induction. Let k be a positive integer and suppose that the maps the space M !+ M !× + Im k (A| B) into M + M × + Im k (A| B). We operator Ψ shall show that the same is true with k replaced by k + 1. Clearly !+M !× + Im k+1 (A| B) = M !+M !× + Im k (A| B) + Im A k B, M and B replaced by M , M × , A and B, respectively. !, M !× , A and similarly with M maps Im A k B Hence, in view of the induction hypothesis, it suﬃces to verify that Ψ B). In other words, what we need is the inclusion into M + M × + Im k+1 (A| A k B B). ⊂ M + M × + Im k (A| Im Ψ

(7.28)

can be A k B With the help of (the ﬁrst identity in) Lemma 7.6, the operator Ψ written as A k B Ψ

= =

C P − P × B C +A × Ψ) A k−1 B (B k−1 C + B( C P − C −C Ψ) +A Ψ A B, (I − P× )B

156

Chapter 7. Wiener-Hopf factorization and factorization indices

and we may conclude that A k B + Im A Ψ A k−1 B. ⊂ M × + Im B Im Ψ

(7.29)

B). Employing the induction hypothesis k−1 B ⊂ M !+M !× + Im k (A| Now Im A k−1 B), i.e., ⊂ M + M × + Im k (A| Im A B once again gives Ψ A k−1 B B). ⊂ M + M × + Im k (A| Im Ψ But then Ψ A k−1 B Im A

=

Im Ψ A k−1 B A

⊂

M + M × + Im k (A| B) A

=

Im k (A| M +A M× + A B) A

⊂

× C [M × ] + A Im k (A| B) +B M+ A +A Im k (A| × M × + Im B B) M +A

⊂

+A Im k (A| B) , M + M × + Im B

⊂

and hence, taking into account (7.29), +A Im k (A| B) . ⊂ M + M × + Im B A k B Im Ψ B) = Im k+1 (A| B), the inclusion (7.28) follows. +A Im k (A| As Im B maps Fix the nonnegative integer k. As we have seen, the linear operator Ψ × × ! ! M + M + Im k (A|B) into M + M + Im k (A|B). Likewise Ψ maps the space B) into M !+ M !× + Im k (A| B). The same is true with k replaced M + M × + Im k (A| by k + 1. But then the linear operators k Φ

:

B) B) !+M !× + Im k (A| M + M × + Im k (A| M → , B) B) !+M !× ∩ Im k+1 (A| M M + M × + Im k+1 (A|

(7.30)

k Φ

:

B) B) !+M !× + Im k (A| M + M × + Im k (A| M → × × ! ! M + M ∩ Im k+1 (A|B) M + M + Im k+1 (A|B)

(7.31)

and Ψ, respectively, are well-deﬁned. They are also each other’s induced by Ψ inverse. This can be deduced easily from A) ⊃ !+ M !× + Im k (C| M

Ψ), Im (I − Ψ

A), M + M × + Im k (C| ⊃

Ψ), Ker (I − Ψ

7.3. Wiener-Hopf factorization and spectral invariants

157

two inclusion relations which are immediate from (7.23). Thus the quotient spaces appearing in (7.30) and (7.31) are linearly isomorphic. In particular they have the same (possibly inﬁnite) dimension. The symmetry in the arguments employed in the above proof (Parts 2 and 3 especially) suggests the possible use of a duality reasoning. Working in a ﬁnite dimensional context this line of approach is indeed possible. In the inﬁnite dimensional situation, however, it does not work, an obstacle being that (sums of) operator ranges need not be closed.

7.3 Wiener-Hopf factorization and spectral invariants Let Y, W, Γ, F+ and F− be as in the preceding two sections, and let ε+ , ε− ∈ C be points in F+ and F− , respectively. By a right Wiener-Hopf factorization of W with respect to Γ (and the points ε+ and ε− ) we mean a factorization W (λ) = W− (λ)D(λ)W+ (λ),

λ ∈ Γ,

(7.32)

where the factors W− and W+ are operator-valued functions, the values being operators on Y , such that (i) W− is analytic on F− and continuous on F − , (ii) W+ is analytic on F+ and continuous on F + , (iii) W− and W+ take invertible values on F − and F + , respectively, (iv) the middle term D in (7.32) has the form D(λ) = Π0 +

κ r

λ − ε+ j j=1

λ − ε−

Πj ,

λ ∈ Γ,

(7.33)

where κ1 , . . . , κr are non-zero integers, κ1 ≤ κ2 ≤ · · · ≤ κr , the operators Π1 , . . . , Πr are mutually disjoint rank 1 projections on Y , and Π0 = IY − (Π1 + · · · + Πr ) so Π0 is a projection disjoint from Π1 , . . . , Πr . A necessary condition for such a factorization to exist is that W takes invertible values on Γ. In terms of a realization of W on Γ this means that Γ splits the spectrum of the associate main operator (see again Theorem 2.4). If in (7.32) the factors W− and W+ are interchanged, we speak of a left Wiener-Hopf factorization. We will focus on the right version; for the left variant analogous results hold. A few remarks are in order. Suppose W admits a right Wiener-Hopf factorization with respect to Γ and the points ε+ ∈ F+ and ε− ∈ F− . Then W also admits a right Wiener-Hopf factorization with respect to Γ and any other two

158

Chapter 7. Wiener-Hopf factorization and factorization indices

points γ+ ∈ F+ and γ− ∈ F− . For γ− in the ﬁnite complex plane this is clear from the simple identity λ − ε+ λ − γ+ λ − γ− λ − ε+ = . λ − ε− λ − γ+ λ − γ− λ − ε− For γ− = ∞, use

λ − ε+ λ − ε−

=

λ − ε+ λ − γ+

λ − γ+

1 λ − ε−

.

This brings the middle term D(λ) into the form D(λ) = Π0 +

r

λ − γ+

κj

λ ∈ Γ.

Πj ,

(7.34)

j=1

κj featured in the latter expression have Note that the scalar functions λ − γ+ their zeros and poles in γ+ and ∞. When the origin belongs to F+ , one can take γ+ = 0 and (7.34) becomes D(λ) = Π0 +

r

λκj Πj ,

λ ∈ Γ.

j=1

This type of middle term plays a role in the study of Toeplitz equations where Γ is taken to be the unit circle (see [52], Chapter XXIV). Although a right Wiener-Hopf factorization is (generally) not unique, the non-zero integers κ1 , . . . , κr are. They are called the right (Wiener-Hopf ) factorization indices of W with respect to Γ. Left factorization indices are deﬁned similarly. Sometimes the term partial indices is used instead of factorization indices. Finally, we mention that right (left) canonical factorization corresponds to the case when the right (left) factorization indices are all zero. For the convenience of the reader, we recall (from the previous section) that Ker k (C|A) and Im k (A|B) are deﬁned as Ker k (C|A)

= Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk−1 ,

Im k (A|B)

= Im B + Im AB + · · · + Im Ak−1 B.

Theorem 7.8. Let the function W be given by the realization (7.1), i.e, W (λ) = IY + C(λIX − A)−1 B, where Γ splits the spectrum of A. Then W admits a right Wiener-Hopf factorization with respect to Γ if and only if the following two conditions are satisﬁed: (a) Γ splits the spectrum of A× = A − BC,

7.3. Wiener-Hopf factorization and spectral invariants (b) dim M ∩ M × < ∞ and dim

X M + M×

159

< ∞,

where M = Im P (A; Γ) and M × = Ker P (A× ; Γ). In that case, the right factorization indices of W can be described in terms of the operators appearing in (7.1) as follows: (c) the number s of negative right factorization indices and the negative right factorization indices −α1 , . . . , −αs (in the ordinary order: −α1 ≤ · · · ≤ −αs ) themselves are given by M ∩ M× s = dim , M ∩ M × ∩ Ker C &

αj = k = 1, 2, . . . | dim

M ∩ M × ∩ Ker k−1 (C|A) M ∩ M × ∩ Ker k (C|A)

' ≥ j , j = 1, . . . , s,

(d) the number t of positive right factorization indices and the positive right factorization indices ω1 , . . . , ωt (in reversed order: ωt ≤ · · · ≤ ω1 ) themselves are given by M + M × + Im B t = dim , M + M× & ωj = k = 1, 2, . . . | dim

M + M × + Im k (A|B) M + M × + Im k−1 (A|B)

' ≥ j , j = 1, . . . , t.

As was already indicated above, for left Wiener-Hopf factorizations an analogous theorem holds. The theorem also has an analogue for appropriate closed contours in the Riemann sphere C∞ like the extended real line or the extended imaginary axis. Proof. For the (long and complicated) proof of the “if part” of Theorem 7.8 we refer to [17]. Here we shall concentrate on the “only if part” and the description of the right factorization indices. So we shall assume that W admits a Wiener-Hopf factorization (7.32) with respect to the contour Γ and, say, the points ε+ ∈ F+ and ε− ∈ F− . According to Theorem 7.2 it suﬃces to prove that there exists a special realization for W , for convenience also written as (7.1), such that Γ splits the spectra of A and A× and for which (b)–(d) hold. The argument consists of several steps. Step 1. Write the negative right factorization indices of W in the ordinary order (so from small to large) as −α1 , . . . , −αs , and the positive right factorization indices

160

Chapter 7. Wiener-Hopf factorization and factorization indices

in the reversed order (so from large to small) as ω1 , . . . , ωt : −α1 ≤ · · · ≤ −αs < 0 < ωt ≤ · · · ≤ ω1 . Then D can be written in the form α ω s 1

λ − ε− j λ − ε+ j D(λ) = P0 + P−j + Pj , λ − ε+ λ − ε− j=1 j=t

(7.35)

(7.36)

where P−1 , . . . , P−s , Pt , . . . , P1 are mutually disjoint rank 1 projections on Y , and P0 = IY − (P−1 + · · · + P−s + Pt + · · · + P1 ), so P0 is a projection disjoint from P−1 , . . . , P−s , Pt , . . . , P1 . For deﬁniteness, we shall assume that s and t are both positive. Step 2. Fix j among the integers 1, . . . , s, and let Dj− (λ) be the scalar function given by α λ − ε− j Dj− (λ) = , λ = ε+ . λ − ε+ Write Jj− for the lower triangular Jordan block with eigenvalue ε+ and order αj , so that σ(Jj− ) = {ε+ } . Further introduce ⎤ ⎡ αj αj ⎢ (ε+ − ε− ) αj ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ .. ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ − , Bj = ⎢ ⎥ ⎢ ⎥ ⎢ (ε − ε )2 αj ⎥ + − ⎢ ⎥ 2 ⎢ ⎥ ⎢ ⎥ ⎣ ⎦ αj (ε+ − ε− ) 1 Cj−

=

0

... 0 1

.

Then Dj− (λ) = 1 + Cj− (λ − Jj− )−1 Bj− is a (minimal) realization of Dj− . Now Jj−× − ε− Iαj is similar with the lower triangular nilpotent Jordan block of order αj and having eigenvalue ε+ , a similarity being given by the upper triangular matrix ( )αj ν−μ ν − 1 (ε+ − ε− ) , ν −μ μ,ν=1 ν−1 where μ−1 is read as zero for μ > ν. Thus σ Jj−× = {ε− }. Clearly P Jj− ; Γ = I and P Jj−× ; Γ = 0. Hence Im P Jj− ; Γ = Ker P Jj−× ; Γ = C αj ,

7.3. Wiener-Hopf factorization and spectral invariants and so, trivially,

Im k Jj− Bj− = C αj ,

k = 0, 1, . . . .

161

(7.37)

Furthermore, as is easily veriﬁed, Ker k Cj− |Jj− = C αj −k {0}k ,

k = 0, 1, . . . ,

(7.38)

where the right-hand side of the equality is read as {0}αj for k ≥ αj . Step 3. Take j among the integers 1, . . . , t, and let Dj+ (λ) be the scalar function given by ω λ − ε+ j + , λ = ε− . Dj (λ) = λ − ε− Write Jj+ for the lower triangular Jordan block with eigenvalue ε− and order ωj , so that σ(Jj− ) = {ε− } . Further introduce ⎡ Bj+

⎢ ⎢ ⎢ =⎢ ⎢ ⎣

1 0 .. .

⎤ ⎥ ⎥ ⎥ ⎥, ⎥ ⎦

0 Cj+

( =

ωj (ε− − ε+ ) 1

) ωj ωj ωj (ε− − ε+ ) . . . (ε− − ε+ ) . 2 ωj 2

Then Dj+ (λ) = 1+Cj+ (λ−Jj+ )−1 Bj+ is a (minimal) realization of Dj+ . Analogously to what we saw in the previous step for the matrix Jj−× − ε− Iαj , the matrix Jj+× − ε+ Iωj is similar with the lower triangular nilpotent Jordan block of order ωj and having ε+ as eigenvalue. Thus σ Jj+× = {ε+ }. Clearly P Jj+ ; Γ = 0 and P Jj+× ; Γ = I. Hence Im P Jj− ; Γ = Ker P Jj−× ; Γ = {0}ωj , and so, trivially, Ker k Cj+ |Jj+ = {0}ωj ,

k = 0, 1, . . . .

(7.39)

Furthermore, as is easily veriﬁed, Im k Jj+ |Bj+ = C k {0}ωj −k ,

k = 0, 1 . . . ,

where the right-hand side of the equality is read as Cωj for k ≥ ωj .

(7.40)

162

Chapter 7. Wiener-Hopf factorization and factorization indices

Step 4. Let D0 (λ) be the diagonal matrix given by ⎡ D1− (λ) ⎢ .. ⎢ . ⎢ ⎢ Ds− (λ) ⎢ D0 (λ) = ⎢ ⎢ Dt+ (λ) ⎢ ⎢ .. ⎣ .

⎤

D1+ (λ)

⎥ ⎥ ⎥ ⎥ ⎥ ⎥, ⎥ ⎥ ⎥ ⎦

(7.41)

i.e., D0 (λ) is the direct sum of the matrices D1− (λ), . . . , Ds− (λ), Dt+ (λ), . . . , D1+ (λ). Then D0 is a rational m × m matrix function, where m = s + t. To obtain a realization for D0 , we introduce n = α1 + · · · + αs + ωt + · · · + ω1 , and introduce an n × n matrix A0 , an n × m matrix B0 and an m × n matrix C0 as follows: A0 is the direct sum of the matrices J1− , . . . , Js− , Jt+ , . . . , J1+ , B0 is the direct sum of the matrices B1− , . . . , Bs− , Bt+ , . . . , B1+ , and C0 is the direct sum of the matrices C1− , . . . , Cs− , Ct+ , . . . , C1+ . Then, indeed, D0 (λ) = Im + C0 (λIn − A0 )−1 B0 is a (minimal) realization. Obviously, Γ splits the spectra of A0 and A× 0 = A0 − B0 C0 . In fact, these spectra coincide with {ε+, ε− }. (Without the assumption introduced in Step 1 that s and t are both positive, we would have that the spectra of A0 and A× 0 are subsets of {ε+ , ε− }, and these inclusions are both proper if and only if one of the integers s or t equals zero.) Put M0 = Im P (A0 ; Γ) and M0× = Ker P (A× 0 ; Γ). Then M0 = M0× = Cα1 · · · Cαs {0}ωt · · · {0}ω1 .

(7.42)

Further we have, for k = 1, 2, . . . , Ker k (C0 |A0 ) = Im k (A0 |B0 ) =

Ker k (C1− |J1− ) · · · Ker k (Cs− |Js− ) {0}ωt · · · {0}ω1 , Cα1 · · · Cαs Im k (Jt+ |Bt+ )) · · · Im k (J1+ |B1+ ),

and M0 ∩ M0× ∩ Ker k (C0 |A0 )

=

Ker k (C0 |A0 ),

(7.43)

M0 + M0× + Im k (A0 |B0 )

=

Im k (A0 |B0 ).

(7.44)

Here we used (7.39) and (7.37). It is clear from (7.42) that dim M0 ∩ M0× = α1 + · · · + αs . Combining (7.43) and (7.38), we get dim M0 ∩ M0× ∩ Ker k (C0 |A0 ) = max{0, α1 − k} + · · · + max{0, αs − k}.

7.3. Wiener-Hopf factorization and spectral invariants In particular

163

dim M0 ∩ M0× ∩ Ker C0 ) = (α1 − 1) + · · · + (αs − 1),

and it follows that

dim

M0 ∩ M0× M0 ∩ M0× ∩ Ker C0

= s.

Thus, with M, M × , C replaced by M0 , M0× , C0 , respectively, the ﬁrst identity in Theorem 7.8, item (c) is satisﬁed . We also have

M0 ∩ M0× ∩ Ker k−1 (C0 |A0 ) dim = max{αl − k + 1} − max{αl − k} × M0 ∩ M0 ∩ Ker k (C0 |A0 ) l∈{1,...,s} =

1 = {l = 1, . . . , s | αl ≥ k}.

l∈{1,...,s}, αl ≥k

Now, ﬁx j ∈ {1, . . . , s}. Then {l = 1, . . . , s | αl ≥ k} ≥ j and hence

⇔

k ∈ {1, . . . , αj },

* + k = 1, 2, . . . | {l = 1, . . . , s | αl ≥ k} ≥ j = αj .

Combining these elements we see that, with M, M × , A, B, C replaced by M0 , M0× , A0 , B0 , C0 , respectively, the second identity in Theorem 7.8, item (c) holds too. For the two identities in Theorem 7.8, item (d), the analogous observation is true. The arguments are basically the same as the ones presented for item (b). Step 5. Next we deal with the middle term D in the factorization (7.32), written in the form (7.36) with −α1 , . . . , −αs , ωt , . . . , ω1 satisfying (7.35), P−1 , . . . , P−s , Pt , . . . , P1 mutually disjoint rank 1 projections on Y and P0 = IY − (P−1 + · · · + P−s + Pt + · · · + P1 ) . Clearly P0 and P−1 + · · · + P−s + Pt + · · · + P1 are complementary projections. Put Y0 = Ker P0 . Then Y0 = Im P−1 · · · P−s Im Pt · · · P1 and so Y0 can be identiﬁed with Cm where, as before, m = s + t. Thus Y = Cm Im P0 and with respect to this decomposition D(λ) can be written as an operator matrix D0 (λ) 0 D(λ) = . 0 I Here D0 is given by (7.41) and I is the identity operator on Im P0 . Now let C0 , CD = BD = B0 0 , AD = A0 , 0

164

Chapter 7. Wiener-Hopf factorization and factorization indices

where A0 , B0 and C0 are as in Step 4. Then we have the realization D(λ) = IY + CD (λIn − AD )−1 BD , n = α1 +· · ·+αs +ωt +· · ·+ω1 , with Γ splitting the spectra of AD = A0 and A× D = × × A0 − B0 C0 = A× 0 . Write MD = Im P (AD ; Γ) and MD = Ker P (AD ; Γ). In × other words, MD = M0 and MD = M0× where, again, we use the notation of the previous step. For k = 1, 2, . . . , clearly, Ker k (CD |AD ) = Ker k (C0 |A0 ) and Im k (AD |BD ) = Im k (A0 |B0 . It follows that, with M, M × , A, B, C replaced by × M D , MD , AD , BD , CD , respectively, (b)–(d) in Theorem 7.8 are satisﬁed. Step 6. We begin this sixth and ﬁnal step by representing the factors W− and W+ in the Wiener-Hopf factorization (7.32) in the form W− (λ)

= IY + C− (λIX− − A− )−1 B− ,

λ ∈ Ω− ,

W+ (λ)

= IY + C+ (λI X+ − A+ )−1 B+ ,

λ ∈ Ω+ ,

with σ(A− ) ⊂ F+ ,

σ(A× − ) ⊂ F+ ,

σ(A× + ) ⊂ F− .

σ(A+ ) ⊂ F− ,

Why this can be done is explained in the proof of Theorem 7.1. On Γ we have the factorization (7.32), and so we can apply the product rule of Section 2.5 to show that W (λ) = IY + C(λIX − A)−1 B, λ ∈ Γ, where X = X − Cn X + , n = α1 + · · · + αs + ωt + · · · + ω1 , and A : X → are given by ⎡ ⎡ ⎤ A− B− CD B− C+ B− ⎢ ⎢ ⎥ ⎢ ⎢ ⎥ AD BD C+ ⎥ , A=⎢ 0 B = ⎢ BD ⎣ ⎣ ⎦ 0 0 A+ B+

(7.45)

X, B : Y → X and C : X → Y ⎤ ⎥ ⎥ ⎥, ⎦

C=

C−

CD

C+

.

Here the operator matrices are taken with respect to the decomposition (7.45). Now the realization obtained for W this way has the desired properties. This can be seen as follows. Obviously Γ splits the spectrum of A and the same is true for A× = A − BC which has the matrix representation ⎡ ⎤ A× 0 0 − ⎢ ⎥ ⎢ ⎥ A× = ⎢ −BD C− A× 0 ⎥. D ⎣ ⎦ × −B+ C− −B+CD A+

7.3. Wiener-Hopf factorization and spectral invariants

165

Let M = Im P (A; Γ) and M × = Ker P (A× ; Γ). Assume for the moment that we have established the identities × ∩Ker k (CD |AD ) {0+ }, (7.46) M ∩ M × ∩ Ker k (C|A) = {0− } MD ∩MD × M + M × + Im k (A|B) = X− MD + MD + Im k (AD |BD ) X+ ,

(7.47)

where 0− is the zero element in X− , 0+ is the zero element in X+ and k is allowed to take the values 0, 1, 2, . . . . Then it would be clear from the conclusions obtained in the previous step that (b)–(d) in Theorem 7.8 are met and we would be ready. So we need to concentrate on (7.46) and (7.47). Clearly P = P (A; Γ) has the form ⎡ ⎤ IX− P1 P2 ⎢ ⎥ ⎢ ⎥ P (AD ; Γ) P3 ⎥ . P =⎢ 0 ⎣ ⎦ 0 0 0 From the fact that P (A; Γ) is a projection one gets the relations P1 P (AD ; Γ) = 0,

P1 P3 = 0,

P (AD ; Γ)P3 = P3 ,

(where the two outer ones imply the middle). In turn ⎡ ⎤⎡ IX− −P1 −P2 IX− 0 0 ⎢ ⎥⎢ ⎢ ⎥⎢ In −P3 ⎥ ⎢ 0 P (AD ; Γ) 0 P =⎢ 0 ⎣ ⎦⎣ 0 0 IX+ 0 0 0

these give ⎤⎡ IX− P1 ⎥⎢ ⎥⎢ In ⎥⎢ 0 ⎦⎣ 0 0

P2

⎤

⎥ ⎥ P3 ⎥ , ⎦ IX+

with the ﬁrst and last factor in the right-hand side invertible and being each other’s inverse. Hence ⎤ ⎡ IX− −P1 −P2 ⎥ ⎢ ⎥ ⎢ In −P3 ⎥ X− MD + {0+ } = X− MD {0+ }. M =⎢ 0 ⎦ ⎣ 0 0 IX+ × X+ , and it follows that In the same way one gets M × = {0− } MD

M ∩ M× M + M×

× = {0− } MD ∩ MD {0+}, × = X− M D + M D X+ .

Thus (7.46) and (7.47) are valid for k = 0.

(7.48) (7.49)

166

Chapter 7. Wiener-Hopf factorization and factorization indices

To prove (7.46) for arbitrary k we argue as follows. A simple induction argument shows that CAl is of the form % $ l−1

l ν l l = 0, 1, . . . , (7.50) CA = ∗ Qν,l CD AD + CD AD ∗ , ν=0

where Q0,l , . . . Ql−1,l and the stars denote appropriate but here not explicitly speciﬁed operators. Together with (7.48) this gives that the right-hand side of (7.46) is contained in the left-hand side. The reverse inclusion can be proved by an induction argument in which (7.50) is employed once more. Finally let us turn to (7.47). For Al B there is an expression analogous to (7.50), namely ⎡

⎤

∗

⎢ l−1 ⎢ AB=⎢ AνD BD Rν,l + AlD BD ⎣ l

ν=0

⎥ ⎥ ⎥, ⎦

l = 0, 1, . . . ,

(7.51)

∗

where R0,l , . . . Rl−1,l and the stars stand for certain operators. Together with (7.49) this yields that the left-hand side of (7.47) is contained in the right-hand side. The reverse inclusion can be proved by an induction argument in which (7.51) is used once again. We close this section with a couple of observations on the dimension numbers featuring in Theorems 7.2 and 7.8. For shortness sake, introduce M ∩ M × ∩ Ker k−1 (C|A) α k = dim , M ∩ M × ∩ Ker k (C|A) ω k

=

dim

M + M × + Im k (A|B) . M + M × + Im k−1 (A|B)

Here k may run through the positive integers 1, 2, . . . . Recall that Ker 0 (C|A) is read as X and Im 0 (A|B) as {0}, so M ∩ M× , α 1 = dim M ∩ M × ∩ Ker C ω 1

=

dim

M + M × + Im B . M + M×

Using standard linear algebra arguments it can be shown that the sequences α 1 , α 2 , . . . and ω 1 , ω 2 , . . . are decreasing, i.e., k+1 , α k ≥ α

ω k ≥ ω k+1 ,

k = 1, 2, . . . .

7.3. Wiener-Hopf factorization and spectral invariants

167

In addition it can be proved that α k and ω k vanish for k suﬃciently large, provided that M ∩ M × and M + M × have ﬁnite dimension and codimension, respectively. In fact we then even have, M ∩ M × ∩ Ker k (C|A)

=

{0},

M + M × + Im k (A|B)

=

X,

again holding for k suﬃciently large. The considerations in Step 4 in the above proof corroborate these facts. 2 , . . . ; for ω 1 , ω 2 , . . . the situation Here are some details for the integers α 1 , α is analogous. The mapping M ∩ M × ∩ Ker k (C|A) M ∩ M × ∩ Ker k−1 (C|A) → × M ∩ M ∩ Ker k+1 (C|A) M ∩ M × ∩ Ker k (C|A) k . Assume now that induced by A is easily seen to be injective. Hence α k+1 ≤ α M ∩ M × has ﬁnite dimension. Then there exists a positive integer r such that M ∩ M × ∩ Ker k (C|A) = M ∩ M × ∩ Ker r (C|A),

k = r, r + 1, . . . .

Evidently M ∩M × ∩ Ker r (C|A) is invariant under both A and A× . Also A and A× coincide on M ∩ M × ∩ Ker r (C|A). As the restriction of A to M and that of A× to M × have no eigenvalue in common, it follows that M ∩ M × ∩ Ker r (C|A) = {0}.

Notes This chapter is based on the papers [17] and [18]. The material of these papers relevant for this book has been reorganized and several of the arguments have been improved. The details are as follows. The “if part” of Theorem 7.1 is a special case of Theorem 3.1 in [17]; it also has the ﬁrst part of Theorem 1.5 in [11] as a less general predecessor. Theorem 7.2 combines Theorems 5.1 and 6.1 of [18] in a more appropriate formulation. The proof of Theorem 7.2 given in Section 7.2 is a signiﬁcant improvement over the argument given in [18]. The results from [17] and [18] to be mentioned in connection with Theorem 7.8 are Theorem 3.1 and Corollary 3.2 in [17] and Theorem 1.2 in [18]. The spectral invariants appearing in Theorem 7.2 are closely related to the block similarity invariants of operator blocks of the ﬁrst or third kind; see [58], Section XI.5 in particular. For a review of the theory of possibly non-canonical Wiener-Hopf factorization of matrix-valued functions taking invertible values, we refer to the book [29] and the more recent survey article [59]. Wiener-Hopf factorization of operator-valued functions goes back to [71] and [72]; see also the recent book [73] . The fact that the Wiener-Hopf factorization indices depend on the given function only (and not on the particular Wiener-Hopf factorization) is wellknown for continuous matrix-valued functions (see [60]) and for certain classes of continuous operator-valued functions (see [49]). The latter do not cover the class of operator-valued functions considered in this chapter.

Part IV Factorization of selfadjoint rational matrix functions This part deals with factorization problems for rational matrix functions that have Hermitian values on the real line, the imaginary axis, or the unit circle. Included are problems of spectral factorization and pseudo-spectral factorization. The emphasis is on positive deﬁnite and nonnegative functions. In general, the factorizations considered are canonical or pseudo-canonical, and they are symmetric in the sense that they consist of two factors, where the ﬁrst factor is the adjoint of the second (relative to the given curve). This part consists of four chapters. Minimal realizations play an important role in the analysis of rational matrix functions that have Hermitian values on a curve. These are realization of which the order of the state matrix is equal to the MacMillan degree of the function. In the ﬁrst chapter (Chapter 8) we review the theory of such realizations. Included are the state space similarity theorem and the minimal factorization theorem. In this ﬁrst chapter we also introduce the notion of pseudo-canonical factorization and describe such factorizations in state space terms. In Chapter 9 we study in a state space setting spectral factorizations, that is, symmetric canonical factorizations for rational matrix functions that are positive deﬁnite on the unit circle, the real line or the imaginary axis. Chapter 10 carries out a similar program for nonnegative functions. In this case one has to consider symmetric pseudo-canonical factorization. In the ﬁnal chapter (Chapter 11) we present (without proofs) some background material on matrices in ﬁnite dimensional indeﬁnite inner product spaces, and review the main results from this area that are used in this part and the other remaining parts.

Chapter 8

Preliminaries concerning minimal factorization In this chapter we gather together several results concerning minimal realizations and minimal factorizations that will play an important role in the sequel. Most of these results can also be found in Part II of the book [20]. For the reader’s convenience we have chosen to summarize them here (without proofs). Special attention is given to the notion of pseudo-canonical factorization, which is a generalization of canonical factorization by allowing singularities on the curve. This chapter consists of three sections. Sections 8.1 and 8.2 deal with minimal realizations and minimal factorizations, respectively. Section 8.3 is devoted to pseudo-canonical factorization.

8.1 Minimal realizations Let W be a proper rational m × m matrix function, and let W (λ) = D + C(λIn − A)−1 B

(8.1)

be a realization of W . The realization is said to be minimal if the dimension n of the state space has the smallest possible value. This smallest possible value is equal to the McMillan degree of W (see Section 8.5 in [20] for details). The McMillan degree of W will be denoted by δ(W ). For a characterization of minimality in terms of the matrices A, B and C, we need some more terminology. Let A be an n × n matrix, let B be an n × m matrix, and let C be an m × n matrix. The pair (A, B) is called controllable if Im (A|B) = Im B + Im AB + · · · + Im AB n−1 = Cn .

172

Chapter 8. Preliminaries concerning minimal factorization

So (A, B) is controllable if and only if Cn is the unique A-invariant subspace containing Im B. The pair (C, A) is said to be observable if Ker (C|A) = Ker C ∩ Ker CA ∩ · · · ∩ Ker CAn−1 = {0}. Thus (C, A) is observable if and only if {0} is the unique A-invariant subspace contained in Ker C. In line with these deﬁnitions, the realization (8.1) is called controllable, respectively observable, if the pair (A, B) is controllable, respectively the pair (C, A) is observable. From Sections 7.1 and 7.3 in [20] we now recall the main results on minimal realizations in the following two theorems. Theorem 8.1. A realization of a proper rational matrix function is minimal if and only if it is controllable and observable. Theorem 8.2. Let W be a proper rational matrix function and suppose W (λ)

=

D1 + C1 (λIn − A1 )−1 B1 ,

(8.2)

W (λ)

=

D2 + C2 (λIn − A2 )−1 B2 ,

(8.3)

are minimal realizations of W . Then D1 = D2 and there exists a unique invertible n × n matrix S such that S −1 A1 S = A2 ,

S −1 B1 = B2 ,

C1 S = C2 .

(8.4)

This second theorem is known as the state space similarity theorem; the operator S is called a (state space) similarity between the realizations (8.2) and (8.3). In the situation where (8.1) is a minimal realization, there is a close connection between the poles of W and the eigenvalues of A. Obviously, whether or not the realization is minimal, the poles of W form a subset of σ(A). However, when the realization is minimal, the spectrum of A coincides with the set of poles of W . In addition, when W is a square matrix-valued function, and D is invertible so that A× = A − BD−1 C is well-deﬁned, σ(A× ) is precisely equal to the set of zeros of W . Here a zero of W is a pole of the inverse W −1 of W . For further details, including a more intrinsic deﬁnition of the notion of a zero of a rational matrix function, taking into account multiplicities and pole orders too, see Chapter 8 in [20]. From Chapter 7 in [20] we also recall that (8.1) is minimal when σ(A) ∩ σ(A× ) = ∅. Next we consider the concept of local minimality. Let λ0 be a point in the complex plane. The realization (8.1) is called locally minimal at λ0 if Im P B + Im P AB + · · · + Im P AB n−1

=

Im P,

(8.5)

Ker CP ∩ Ker CAP ∩ · · · ∩ Ker CAn−1 P

=

Ker P,

(8.6)

8.1. Minimal realizations

173

where P is the Riesz projection of A at λ0 . There is a local version of the observation given at the end of the previous paragraph: if λ0 is not a common eigenvalue of A and A× , then (8.1) is minimal at λ0 . For details see Section 8.4 in [20]) where it is also shown that the realization (8.1) is minimal if and only if it is minimal at each point in the complex plane. We ﬁnish this section by reviewing some results on Jordan chains and copole functions. Let W be a rational square matrix-valued function, and let ϕ be a Cm -valued function which is analytic at λ0 with ϕ(λ0 ) = 0. We call ϕ a co-pole function of W at λ0 if W (λ)ϕ(λ) is analytic at λ0 and limλ→λ0 W (λ)ϕ(λ) is nonzero. For this to happen, it is necessary that det W (λ) does not vanish identically. As before, let W −1 denote the pointwise inverse of W , i.e., the function determined by the expression W −1 (λ) = W (λ)−1 . Now, if ϕ is a co-pole function of W at λ0 , then the function ψ(λ) = W (λ)ϕ(λ) is a so-called root function of W −1 at λ0 , that is, ψ is analytic at λ0 with ψ(λ0 ) = 0 and limλ→λ0 W (λ)−1 ψ(λ) = 0. The converse is also true. A root function of W −1 at λ0 is also referred to as a pole function of W at λ0 (see [7], page 67). The next two results have been taken from [20], Section 8.4 (Proposition 8.21 and Corollary 8.22). Proposition 8.3. Let the rational square matrix-valued function W be given by the realization (8.1), and let λ0 be an eigenvalue of A. Assume the realization is minimal at λ0 . Let k ≥ 1, and let ϕ(λ) = (λ − λ0 )k ϕk + (λ − λ0 )k+1 ϕk+1 + · · · be a co-pole function of W at λ0 . Put xj =

∞

P (A − λ0 )ν−j−1 Bϕν ,

j = 0, . . . , k − 1,

(8.7)

ν =k

where P is the Riesz projection of A corresponding to λ0 . Then x0 , . . . , xk−1 is a Jordan chain of A at λ0 , that is, x0 = 0 and (A − λ0 )x0 = 0,

(A − λ0 )r xk−1 = xk−1−r ,

r = 0, . . . , k − 1.

(8.8)

Moreover, each Jordan chain of A at λ0 is obtained in this way. Finally, if the chain x0 , . . . , xk−1 given by (8.7) is maximal, that is, xk−1 ∈ Im (A − λ0 ), then ϕk = 0. With respect to (8.7) there is no convergence issue; actually only a ﬁnite number of terms in the sum are non-zero. Proposition 8.4. Let the rational square matrix-valued function W be given by the realization (8.1), and suppose det W (λ) ≡ 0. Let λ0 be an eigenvalue of A, and assume that (8.1) is minimal at λ0 . If x0 , . . . , xk−1 is a Jordan chain of A at λ0 , then Cx0 , . . . , Cxk−1 is a Jordan chain of W −1 at λ0 , and each Jordan chain of W −1 at λ0 is obtained in this way.

174

Chapter 8. Preliminaries concerning minimal factorization

For later use (see Section 10.1) we introduce the following terminology suggested by Proposition 8.3. Let W be given by the realization(8.1). If x0 , . . . , xk−1 ∞ is a Jordan chain of A at λ0 , any co-pole function ϕ(λ) = j=k (λ − λ0 )j ϕj satisfying (8.7) will be called a co-pole function corresponding to the Jordan chain x0 , . . . , xk−1 . In this case Cxj is precisely the coeﬃcient of (λ − λ0 )r in the Taylor expansion of W (λ)ϕ(λ) at λ0 . To see this, use (8.7) and the fact that the coeﬃcients in the principal part of the Laurent expansion of W at λ0 are given by the expression CP (A − λ0 )j−1 B, where P is the Riesz projection of A corresponding to the eigenvalue λ0 . These observations lie also behind Proposition 8.4 above.

8.2 Minimal factorization The McMillan degree features a sublogarithmic property. Indeed, if W1 and W2 are rational matrix functions and W = W1 W2 , that is W (λ) = W1 (λ)W2 (λ), then the McMillan degree of W is less than or equal to the sum of the McMillan degrees of W1 and W2 : δ(W1 W2 ) ≤ δ(W1 ) + δ(W2 ).

(8.9)

This is clear from Theorem 2.5 and the deﬁnition of the McMillan degree given in the beginning of the previous section. A factorization W = W1 W2 is called a minimal factorization (involving two factors) if equality occurs, that is, when δ(W ) = δ(W1 ) + δ(W2 ). Intuitively, this means that there is no pole-zero cancellation in the product W1 W2 ; this is made precise in Theorem 9.1 in [20]. Let W (λ) = D + C(λIn − A)−1 B be a realization of an m× m rational matrix function, assume that D is invertible, and let D = D1 D2 with D1 , D2 m × m matrices (automatically invertible). Put A× = A − BD−1 C. Suppose M, M × is a pair of subspaces of Cn satisfying A× M × ⊂ M × ,

AM ⊂ M,

˙ M × = Cn . M+

(8.10)

In that case we know (see Section 2.6) that W admits a factorization W = W1 W2 where the factors can be described using the projection Π onto M × along M as follows: W1 (λ)

=

D1 + C(λIn − A)−1 (I − Π)BD2−1 ,

(8.11)

W2 (λ)

=

D2 + D1−1 CΠ(λIn − A)−1 B.

(8.12)

The next theorem, which is a reformulation of the main result in [20], Section 9.1, shows that the above factorization principle yields all minimal factorizations of W whenever the given realization is minimal.

8.2. Minimal factorization

175

Theorem 8.5. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of the m × m rational matrix-valued function W , and assume D is invertible. (i) Let D = D1 D2 with D1 , D2 (invertible) m×m matrices. If a pair of subspaces M and M × of Cn satisﬁes (8.10), then the factorization W = W1 W2 , with the factors W1 and W2 given by (8.11) and (8.12), is a minimal factorization. (ii) If W = W1 W2 is a minimal factorization of W involving proper rational m × m matrix functions W1 and W2 , then there is a unique pair of subspaces M and M × satisfying (8.10) such that the factors W1 and W2 are given by (8.11) and (8.12) where D1 and D2 are the (invertible) values of W1 and W2 at ∞, respectively. The notion of minimal factorization can be extended to products involving an arbitrary number of factors. Indeed, a factorization W = W1 · · · Wk is called a minimal factorization if δ(W ) = δ(W1 ) + · · · + δ(Wk ).

(8.13)

In general all we can say is that the left-hand side of (8.13) does not exceed the right-hand side. The special case of complete factorization is of particular interest. Let W be a rational m × m matrix-valued function which is biproper , that is, W is analytic at inﬁnity and has an invertible value there. A minimal factorization of W into biproper rational m × m matrix functions, each having McMillan degree 1, is called a complete factorization of W . The number of factors in such a complete factorization is necessarily equal to the McMillan degree of W . If W (λ) = D + C(λIn − A)−1 B is a minimal realization of W , then W admits a complete factorization if and only if the matrices A and A× can be brought into complementary triangular form, i.e., there is a basis such that, with respect to this basis, A has upper triangular form and A× has lower triangular form. For further details, see Chapter 10 in [20]. We shall meet complete factorization later in Section 17.3. We conclude this section with some remarks on a local version of minimal factorization. First we introduce the local (McMillan) degree. Let W be a proper rational matrix function, let W (λ) = D + C(λIn − A)−1 B

(8.14)

be a minimal realization of W , and let μ ∈ C. The algebraic multiplicity af μ as an eigenvalue of A is called the local (McMillan) degree of W at μ, written δ(W ; μ). By the state space similarity theorem, this deﬁnition does not depend on the choice of the minimal realization (8.14). For an alternative deﬁnition of the local degree, we refer to Section 8.4 in [20] where the square case is considered. In that situation, when det W (λ) does not vanish identically, the local degree of W

176

Chapter 8. Preliminaries concerning minimal factorization

at μ coincides with the pole-multiplicity of W at μ in the sense of [20], Section 8.2. It is obvious, again from Theorem 2.5, that the global sublogarithmic property (8.9) has the following local counterpart: δ(W1 W2 ; μ) ≤ δ(W1 μ) + δ(W2 ; μ).

(8.15)

A factorization W = W1 W2 is said to be locally minimal at μ if equality occurs in (8.15), that is, when δ(W1 W2 ; μ) = δ(W1 μ) + δ(W2 ; μ). Intuitively, this means that in the product W1 W2 no pole-zero cancellation occurs at the point μ (see again Theorem 9.1 in [20]). For the case of proper rational matrix functions (as considered here), the minimality of a factorization comes down to local minimality at each point in the complex plane. Thus W = W1 W2 is a minimal factorization if and only if δ(W1 W2 ; λ) = δ(W1 λ) + δ(W2 ; λ),

λ ∈ C;

see Section 9.1 in [20].

8.3 Pseudo-canonical factorization Let Γ be a Cauchy contour in C. As before, the interior domain of Γ is denoted by F+ , and the exterior domain by F− . By deﬁnition (see Chapter 0), ∞ ∈ F− . Let W be an m × m rational matrix function, possibly having poles and zeros on Γ. By a right pseudo-canonical factorization of W with respect to Γ we mean a factorization W (λ) = W− (λ)W+ (λ),

λ ∈ Γ, λ not a pole of W,

(8.16)

where W− and W+ are rational m × m matrix functions such that W− is analytic and takes invertible values on F− (i.e., W− has neither poles nor zeros there), W+ is analytic and takes invertible values on F+ (i.e., W− has neither poles nor zeros there), and the factorization (8.16) is locally minimal at each point of Γ. If in (8.16) the factors W− and W+ are interchanged, we speak of a left pseudocanonical factorization. In passing we mention that the deﬁnition of pseudo-canonical factorization given in the second paragraph of [20], Section 9.2 is not quite correct. The point is that the function W is allowed to have poles and zeros on Γ. This is explicitly stated in the third paragraph of the section in question, but the formal deﬁnition referred to above in the second paragraph erroneously suggests otherwise. As for canonical factorization, the notion of pseudo-canonical factorization extends to factorization with respect to the real line and the imaginary axis. To be more speciﬁc, if Γ is the closure of the real line on the Riemann sphere, then F+ is the open upper half plane, and F− is the open lower half plane. Replacing R

8.3. Pseudo-canonical factorization

177

by iR means only replacing the open upper half plane by the open left half plane, and the open lower half plane by the open right half plane. A pseudo-canonical factorization is not only minimal at each point of Γ but also at all other points of C and at inﬁnity. This follows from the conditions on the poles and zeros of the factors W− and W+ in (8.16). Thus a pseudo-canonical factorization is a minimal factorization. In combination with Theorem 8.5 this fact makes it possible to describe all right pseudo-canonical factorizations of a biproper rational matrix function W in terms of a minimal realization of W . The resulting theorem (which is taken from Section 9.2 in [20]) is given below. In contrast to the main theorem on canonical factorization (Theorem 3.2) we are forced here to work with minimal realizations. Theorem 8.6. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a biproper rational matrix-valued function W , and put A× = A−BD−1 C. Let Γ be a Cauchy contour. Let D = D1 D2 , with D1 and D2 invertible square matrices. Then there is a one-to-one correspondence between the right pseudo-canonical factorizations W = W− W+ of W with respect to Γ with W− (∞) = D1 and W+ (∞) = D2 , and the pairs of subspaces M, M × of Cn with the following properties: (i) M is an A-invariant subspace such that the restriction A|M of A to M has no eigenvalues in F− , and M contains the span of all eigenvectors and generalized eigenvectors of A corresponding to eigenvalues in F+ , (ii) M × is an A× -invariant subspace such that the restriction A× |M × of A× to M × has no eigenvalues in F+ , and M × contains the span of all eigenvectors and generalized eigenvectors of A× corresponding to eigenvalues in F− , ˙ M ×. (iii) Cn = M + The correspondence is as follows: given a pair of subspaces M, M × of Cn with the properties (i), (ii) and (iii), a right pseudo-canonical factorization of W with respect to Γ is given by W (λ) = W− (λ)W+ (λ), where W− (λ)

= D1 + C(λIn − A)−1 (I − Π)BD2−1 ,

(8.17)

W+ (λ)

= D2 + D1−1 CΠ(λIn − A)−1 B,

(8.18)

where Π is the projection along M onto M × . Conversely, given a right pseudocanonical factorization of W with respect to Γ and with W− (∞) = D1 , W+ (∞) = D2 , there exists a unique pair of subspaces M, M × with the properties (i), (ii) and (iii) above, such that the factors W− and W+ are given by (8.17) and (8.18), respectively. The span of all eigenvectors and generalized eigenvectors of A corresponding to eigenvalues in F+ mentioned in (i) is just the spectral subspace of A associated with the part of the spectrum of A lying in F+ . Similarly, the span of all eigenvectors and generalized eigenvectors of A× featuring in (ii) corresponding to

178

Chapter 8. Preliminaries concerning minimal factorization

eigenvalues in F− is the spectral subspace of A× associated with the part of σ(A× ) lying in F− . A pair of subspaces M, M × for which (i), (ii) and (iii) hold need not be unique. In line with this, pseudo-canonical factorizations are generally not unique either. An example illustrating this is given in [133]; see also Section 9.2 in [20]. Note that for an m × m rational matrix function W , a canonical factorization of W with respect to the curve Γ is a pseudo-canonical factorization with the additional property that the factors have no poles or zeros on the curve. In that case, W has no poles or zeros on Γ also. Conversely, if W has no poles or zeros on Γ, then any pseudo-canonical factorization W = W1 W2 of W is automatically a canonical factorization. Indeed, if W has no poles or zeros on Γ, then the fact that the factorization W = W1 W2 is locally minimal at each point of Γ, implies that W1 and W2 have no poles or zeros on Γ, and thus the pseudo-canonical factorization W = W1 W2 is a canonical one. As a result we have the following special case of Theorem 8.6. Theorem 8.7. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a biproper rational matrix-valued function W , and put A× = A − BD−1 C. Let Γ be a Cauchy contour. Assume that A has no eigenvalues on Γ. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (i) A× has no eigenvalues on Γ, ˙ Ker P (A× ; Γ). (ii) Cn = Im P (A; Γ)+ In that case, the right canonical factorizations with respect to Γ are of the form W = W− W+ , with W− and W+ given by (8.17) and (8.18), where Π is the projection along Im P (A; Γ) onto Ker P (A× ; Γ), and where D = D1 D2 , with D1 and D2 invertible square matrices. This correspondence is a one-to-one correspondence between the right canonical factorizations of W and the factorizations of D into square factors. Observe that the above theorem is a modest reﬁnement of Theorem 3.2 in the sense that we allow the value of W at inﬁnity to be an arbitrary invertible matrix here. The result of the theorem also holds for non-minimal realizations. The argument for this consists of a straightforward modiﬁcation of the proof of Theorem 3.2. Theorem 8.7 allows for analogues in which the Cauchy contour Γ is replaced by the extended real or imaginary axis.

Notes The material in the ﬁrst section is standard and can be found in many textbooks; see, e.g., [94], or the more recent [33], [85]. The idea of minimal factorization originates from mathematical systems theory and has been developed systematically in Chapter 4 of [11] (see also [21]), and with further details in Part II of [20]. An

8.3. Pseudo-canonical factorization

179

extensive analysis of factorization into square degree 1 factors can be found in Part III of [20]. The analysis involves a connection with a problem of job scheduling from operations research. Minimal factorization into possibly non-square factors of McMillan degree 1 is always possible. This has been established in [143]. The notion of a pseudo-canonical factorization is introduced and developed in [132], [133].

Chapter 9

Factorization of positive deﬁnite rational matrix functions The central theme of this chapter is the state space analysis of rational matrix functions with Hermitian values either on the real line, on the imaginary axis, or on the unit circle. The main focus will be on rational matrix functions that take positive deﬁnite values on one of these contours. It will be shown that if W is such a function, then W admits a spectral factorization, i.e., a canonical factorization W = W− W+ with an additional symmetry between the corresponding factors, depending on the contour. This chapter consists of three sections. In Section 9.1 we analyze selfadjointness of a rational matrix function relative to the real line, the imaginary axis or the unit circle. The analysis is done in terms of (minimal) realizations of the functions involved. Elements of the theory of matrices that are selfadjoint with respect to an indeﬁnite inner product enter into the analysis in a natural way. Section 9.2 deals with rational matrix functions that are positive deﬁnite on the real line or on the imaginary axis. The results of Section 9.1 are used to show that such a function admits a spectral factorization and in terms of a given realization an explicit formula for the corresponding spectral factor is given. Section 9.3 presents an analogous result for rational matrix functions that are positive deﬁnite on the unit circle.

9.1 Preliminaries on selfadjoint rational matrix functions Let Γ be one of the following two contours in the complex plane: the real line R, or the imaginary axis iR. A rational m × m matrix function W is called selfadjoint on Γ or Hermitian on Γ if for each λ ∈ Γ, λ not a pole of W , the matrix W (λ)

182

Chapter 9. Factorization of positive deﬁnite rational matrix functions

is selfadjoint or, which is the same, Hermitian. By the uniqueness theorem for analytic functions, a rational matrix function W is selfadjoint on R if and only if ¯ ∗ for all λ ∈ C, λ not a pole of W . Similarly, W is selfadjoint on iR W (λ) = W (λ) ¯ ∗ , λ not a pole of W . From these characterizations if and only if W (λ) = W (−λ) it follows that if W is selfadjoint on Γ and det W (λ) does not vanish identically, then W −1 is also selfadjoint on Γ. This section is concerned with the problem how selfadjointness of a rational matrix function is reﬂected in properties of the matrices in a minimal realization of the function. For proper rational matrix functions this is described in the following theorem. Theorem 9.1. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function. Then the following statements hold: (i) W is Hermitian on the real line if and only if D = D∗ and there exists an n × n matrix H such that HA = A∗ H,

HB = C ∗ ,

H = H ∗;

(9.1)

(ii) W is Hermitian on the imaginary axis if and only if D = D∗ and there exists an n × n matrix H such that HA = −A∗ H,

HB = C ∗ ,

H = −H ∗ .

(9.2)

In both cases (because of the minimality of the realization), the matrix H is uniquely determined by the matrices in the given realization of W and invertible. A matrix H such that H = −H ∗ is called skew-Hermitian. For such a matrix iH is Hermitian. Proof. We ﬁrst prove (i). Assume the matrix function W is Hermitian on R, so ¯ ∗ coincide. Hence the rational matrix functions W (λ) and W (λ) W (λ) = D ∗ + B ∗ (λ − A∗ )−1 C ∗ is also a minimal realization for W . By the state space similarity theorem (Theorem 8.2) we obtain the existence of a unique (invertible) n × n matrix H such that HA = A∗ H, HB = C ∗ , B ∗ H = C. Taking adjoints one gets H ∗ A∗ = AH,

C = B∗H ∗,

H ∗B = C ∗.

Comparing these two sets of equations and employing the uniqueness of H, we see that H = H ∗ . Clearly D = D∗ as D = W (∞) must be selfadjoint.

9.1. Preliminaries on selfadjoint rational matrix functions

183

For the converse, suppose D = D ∗ and there exist an n×n matrix H for which (9.1) holds. From the ﬁrst equality in (9.1) we see that H(λ−A)−1 = (λ−A∗ )−1 H. Then, using the second equality in (9.1), one computes ¯ ∗ W (λ)

=

D∗ + B ∗ (λ − A∗ )−1 C ∗ = D + B ∗ (λ − A∗ )−1 HB

=

D + B ∗ H(λ − A)−1 B = D + C(λ − A)−1 B = W (λ).

So W is selfadjoint on R. Next we show that (because of minimality) the identities in (9.1) imply that H is invertible. Indeed, assume Hx = 0 for some x ∈ Cn . Then the ﬁrst equality in (9.1) yields HAx = 0. Repeating the argument, using induction, we obtain HAk x = 0 for k = 0, 1, 2, . . .. Using the two other equalities in (9.1) we see that CAk x = B ∗ HAk x = 0 for k = 0, 1, 2, . . . . Since the given realization is minimal, the pair (C, A) is observable, and hence x = 0. Thus H is invertible. The proof of (ii) can be given using the same type of reasoning as for (i). On the other hand (ii) also follows directly from (i) by using the transformation ! (λ) = W (−iλ). Since W is assumed to be selfadjoint on λ → −iλ. Indeed, put W ! ! admits the minimal realization iR, the function W is selfadjoint on R. Moreover, W ! (λ) = D + C(λ − A) −1 B, W = iC and A = iA. By (i), there exists an (invertible) selfadjoint matrix where C we derive the desired and HB =C ∗ . Setting H = −iH ˜ H such that H A = A∗ H equalities in (9.2). In the proof of the “if parts” of (i) and (ii), minimality does not play a role. Thus, if (9.1) holds and D = D ∗ , then W (λ) = D + C(λ − A)−1 B is selfadjoint on R. Similarly, if D = D ∗ and (9.2) holds, then W is selfadjoint on iR. In the next proposition we consider the case when the rational matrix function in Theorem 9.1 is biproper, and we describe the eﬀect of the matrices H on the associate main operator A× = A − BD−1 C. Proposition 9.2. Let W (λ) = D + C(λIn − A)−1 B be a realization of an m × m rational matrix function, and let H be an n × n matrix. Assume D is invertible, and put A× = A − BD−1 C. Then the following statements hold: (i) If D = D∗ and (9.1) is satisﬁed, then HA× = (A× )∗ H; (ii) If D = D∗ and (9.2) is satisﬁed, then HA× = −(A× )∗ H. Proof. Assume D = D ∗ and the identities (9.1). Then HA× = HA − HBD−1 C = A∗ H − C ∗ D −∗ B ∗ H = (A× )∗ H, so (i) holds. Statement (ii) is proved analogously.

184

Chapter 9. Factorization of positive deﬁnite rational matrix functions

Next we analyze how the matrix H appearing in Theorem 9.1 behaves under a state space similarity transformation on the realization of W . Theorem 9.3. For i = 1, 2, let W (λ) = D + Ci (λn − Ai )−1 Bi be a minimal realization of the rational matrix function W , and let S be the (unique invertible) n × n matrix such that SA1 = A2 S,

C1 = C2 S,

B2 = SB1 .

Then the following statements hold: (i) Let W be selfadjoint on the real line. For i = 1, 2, write Hi for the (unique invertible) Hermitian n × n matrix such that Hi Ai = A∗i Hi and Hi Bi = Ci∗ . Then H1 = S ∗ H2 S; (ii) Let W be selfadjoint on iR. For i = 1, 2, write Hi for the (unique invertible) skew-Hermitian n × n matrix such that Hi Ai = −A∗i Hi and Hi Bi = Ci∗ . Then H1 = S ∗ H2 S. Proof. We shall only prove (i); statement (ii) can be veriﬁed analogously. One easily checks that S ∗ H2 S satisﬁes (9.1): S ∗ H2 SA1 = S ∗ H2 A2 S = S ∗ A∗2 H2 S = A∗1 S ∗ H2 S, S ∗ H2 SB1 = S ∗ H2 B2 = S ∗ C2∗ = C1∗ . By the uniqueness of H1 the assertion (i) follows.

We conclude this section with a comment on the theory of matrices acting in an indeﬁnite inner product space. Elements of this theory play an important role in the study of selfadjoint rational matrix functions. To see the connection, let H be an invertible Hermitian n × n matrix, and consider on Cn the sesquilinear form [x, y] = Hx, y. If HA = A∗ H, then [Ax, y] = [x, Ay], and hence A is selfadjoint in the indeﬁnite inner product [· , · ] on Cn induced by H. Thus the ﬁrst part and third identity in (9.1) imply that A is selfadjoint in an indeﬁnite inner product space. In the sequel we call A H-selfadjoint if H = H ∗ and HA = A∗ H. Notice that the third identity in (9.2) implies that iH is Hermitian, and hence the ﬁrst identity in (9.2) can be summarized by saying that iA is iH-selfadjoint. In Section 11.2 we review the results from the theory of matrices acting in an indeﬁnite inner product space insofar as they are useful to us in this and the next chapters.

9.2. Spectral factorization

185

9.2 Spectral factorization The ﬁrst factorization result to be presented in this section concerns an important class of rational matrix functions, namely those which are positive deﬁnite on the contour Γ under consideration (again, either R or iR). A rational m × m matrix function W is called positive deﬁnite on Γ if for each λ ∈ Γ, λ not a pole of W , the matrix W (λ) is positive deﬁnite. Suppose W is a rational m × m matrix function. A factorization ¯ ∗ L(λ) W (λ) = L(λ)

(9.3)

is called a right spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed upper half plane ¯ ∗ and its inverse are analytic on (inﬁnity included). In that case the function L(λ) the closed lower half plane (including inﬁnity). Thus a right spectral factorization with respect to R is a right canonical factorization with respect to the real line featuring an additional symmetry property between the factors. A factorization (9.3) is called a left spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed lower ¯ ∗ and its inverse are half plane (inﬁnity included), in which case the function L(λ) analytic on the closed upper half plane including inﬁnity). Such a factorization is a left canonical factorization with respect to R. A factorization ¯ ∗ L(λ) W (λ) = L(−λ) (9.4) is called a right spectral factorization with respect to the imaginary axis if L and L−1 are rational m × m matrix functions which are analytic on the closed left half plane (inﬁnity included). Such a factorization is, in particular, a right canonical factorization with respect to iR. Analogously, (9.4) is called a left spectral factorization with respect to the imaginary axis if L and L−1 are rational m × m matrix functions which are analytic on the closed right half plane (inﬁnity included). The factors in a spectral factorization are uniquely determined up to multi¯ ∗ L(λ) is plication with a constant unitary matrix. More precisely, if W (λ) = L(λ) a spectral factorization with respect to the real line, and E is an m × m unitary ¯ ∗ L(λ) ˜ λ) ˜ ˜ matrix, then W (λ) = L( with L(λ) = EL(λ) is again a spectral factorization of W , and this is all the freedom one has. To see the latter, assume that ¯ ∗ L(λ) and W (λ) = L( ¯ ∗ L(λ) ˜ λ) ˜ W (λ) = L(λ) are right spectral factorizations with respect to R, then −1 ¯ −∗ L(λ) ¯ ∗. ˜ ˜ λ) L(λ)L(λ) = L( The left-hand side of this identity is an m × m rational matrix function which is analytic on the closed upper half plane and the right-hand side is analytic on the closed lower half plane (in both cases the point inﬁnity included). By Liouville’s theorem neither side depends on λ, that is, there exists an m × m matrix E such ¯ −∗ L( ¯ ∗ . But this implies that E is invertible and ˜ −1 and L(λ) ˜ λ) that E = L(λ)L(λ) ˜ E ∗ = E −1 . Hence E is unitary and E = L(λ) = EL(λ), as desired.

186

Chapter 9. Factorization of positive deﬁnite rational matrix functions

If (9.4) is a right (respectively, left) spectral factorization of W with respect to the real line, we refer to L as the right (respectively, left ) spectral factor . Without further explanation a similar terminology will be used in comparable circumstances. Note that existence of a spectral factorization implies that W has no poles or zeros on the given contour and on the contour it is positive deﬁnite. The converse also holds: for positive deﬁnite rational matrix functions, both left and right spectral factorizations exist. This will now be proved for the case when W is a proper rational m × m matrix function. Moreover, explicit formulas for the factors will be given in terms of a realization of W . First we consider the situation where W is positive deﬁnite on the real line. Theorem 9.4. Let W (λ) = D + C(λIn − A)−1 B be a realization of the rational m × m matrix function W . Suppose A has no real eigenvalues, W is positive deﬁnite on the real line, and W (∞) = D is positive deﬁnite too. Further assume there exists an invertible Hermitian n × n matrix H for which HA = A∗ H and HB = C ∗ . Then, with respect to the real line, W admits right and left spectral factorization. Such factorizations can be obtained in the following way. Let M+ and M− be the spectral subspaces of A associated with the parts of σ(A) lying in × × the lower and upper half plane, respectively, and let M+ and M− be the spectral × × subspaces of A associated with the parts of σ(A ) lying in the lower and upper half plane, respectively. Then × ˙ M+ Cn = M− + ,

× ˙ M− Cn = M+ + .

(9.5)

× , Π− for the projection of Write Π+ for the projection of Cn along M− onto M+ × n C along M+ onto M− , and introduce

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

(9.6)

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

(9.7)

Then ¯ ∗ L+ (λ), W (λ) = L+ (λ)

¯ ∗ L− (λ), W (λ) = L− (λ)

are right and left spectral factorizations with respect to the real line, respectively. These spectral factorizations are uniquely determined by the fact that they have the value D 1/2 at inﬁnity. The conditions of the theorem are satisﬁed in case W has no poles on the real line, W (λ) is positive deﬁnite for all real λ, and the given (biproper) realization of W is a minimal one. Proof. The invertibility of W (λ) for real λ combined with the fact that A has no real eigenvalues implies that A× does not have real eigenvalues either (see Theorem × × 2.4). So the subspaces M+ , M− , M+ and M− are well-deﬁned. Let P and P × be

9.2. Spectral factorization

187

the Riesz projections of A, and A× , respectively, with respect to the upper half plane. From HA = A∗ H and HA× = (A× )∗ H one easily computes that HP = (I − P ∗ )H,

HP × = (I − P × )∗ H.

× × and M− satisfy It follows that the spaces M+ , M− , M+ ⊥ HM+ = M+ ,

⊥ HM− = M− ,

× ×⊥ HM+ = M+ ,

× ×⊥ HM− = M− . (9.8)

× × = {0}. Suppose x ∈ M+ ∩ M− . As M+ First it will be shown that M+ ∩ M− is invariant under A, we have Ax ∈ M+ . But then the ﬁrst identity in (9.8) shows × that HAx, x = 0. The space M− is invariant under A× . Thus A× x belongs to × M− , and the last identity in (9.8) yields HA× x, x = 0. Hence

0 = H(A − A× )x, x = HBD−1 Cx, x = D −1 Cx, Cx = D−1/2 Cx2 . As D > 0, it follows that Cx = 0. Thus A× x = (A−BD−1 C)x = Ax. We conclude × that M+ ∩ M− is invariant under both A and A× , and we have A|M+ ∩M × = −

A× |M+ ∩M × . However, −

σ(A|M+ ∩M × ) ⊂ σ(A|M+ ) ⊂ {λ | λ > 0}, −

σ(A× |M+ ∩M × ) ⊂ σ(A× |M × ) ⊂ {λ | λ < 0}. −

−

× = {0}. Thus A|M+ ∩M × = A |M+ ∩M × implies that M+ ∩ M− − − Proving (9.5) is now done via a dimension argument. Since H is invertible, ⊥ have the same dimension. In the ﬁrst identity in (9.8) shows that M+ and M+ × particular, dim M+ = n/2. Similarly, the last identity in (9.8) yields dim M− = n/2. Hence the ﬁrst identity in (9.5) holds. Let Π− be the projection along M+ × onto M− . The second identity in (9.5) is established in a similar way. × Let Π− be the projection along M+ onto M− . Then Π− is a supporting projection, and by Theorem 3.2 the corresponding factorization is a left canonical factorization given by W (λ) = K− (λ)L− (λ), ×

where L− is given by (9.7), and K− (λ) = D1/2 + C(λ − A)−1 (I − Π− )BD−1/2 . ¯ ∗ . Using (9.7) and (9.1) we have It remains to prove that K− (λ) = L− (λ) ¯ ∗ L− (λ)

=

D1/2 + B ∗ (λ − A∗ )−1 Π∗− C ∗ D−1/2

=

D1/2 + C(λ − A)−1 H −1 Π∗− HBD−1/2 .

¯ ∗ , it suﬃces to show that H(I − Π− ) = Π∗ H. Thus in order to get K− (λ) = L− (λ) −

188

Chapter 9. Factorization of positive deﬁnite rational matrix functions

Using the deﬁnition of Π− , together with the ﬁrst and the last identity in (9.8), we see that H(I − Π− )x, (I − Π− )y = 0 and HΠ− x, Π− y = 0 for all x and y in Cn . Hence for all x, y, H(I − Π− )x, y = H(I − Π− )x, Π− y = Hx, Π− y, which yields the desired identity H(I − Π− ) = Π∗− H. As for the last statement in the theorem, recall that the factors in a spectral factorization are uniquely determined up to multiplication with a constant unitary matrix. This settles the theorem as far as left spectral factorization is concerned. For right spectral factorizations the reasoning is similar. With minor modiﬁcations one proves the following theorem concerning left and right spectral factorizations with respect to the imaginary axis. Theorem 9.5. Let W (λ) = D + C(λIn − A)−1 B be a realization of the rational m × m matrix function W . Suppose A has no pure imaginary eigenvalues, W is positive deﬁnite on the imaginary axis, and W (∞) = D is positive deﬁnite too. Further assume there exists an invertible skew-Hermitian n × n matrix H for which HA = −A∗ H and HB = C ∗ . Then, with respect to the imaginary axis, W admits right and left spectral factorization. Such factorizations can be obtained in the following way. Let M+ and M− be the spectral subspaces of A associated with × the parts of σ(A) lying in the right and left half plane, respectively, and let M+ × × × and M− be the spectral subspaces of A associated with the parts of σ(A ) lying in the right and left half plane, respectively. Then × ˙ M+ Cn = M − + ,

× ˙ M− Cn = M+ + .

× Write Π+ for the projection of Cn along M− onto M+ , Π− for the projection of × Cn along M+ onto M− , and introduce

Then

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

¯ ∗ L+ (λ), W (λ) = L+ (−λ)

¯ ∗ L− (λ), W (λ) = L− (−λ)

are right and left spectral factorizations with respect to the imaginary axis, respectively. These spectral factorizations are uniquely determined by the fact that they have the value D 1/2 at inﬁnity. The conditions of the theorem are satisﬁed in case W has no poles on the imaginary axis, W (λ) is positive deﬁnite for λ ∈ iR, and the given (biproper) realization of W is a minimal one. In terms of the theory of spaces with an indeﬁnite metric (see the appendix at the end of this chapter), the identities in (9.8) say × × and M− are Lagrangian subspaces in that the spectral subspaces M+ , M− , M+ the indeﬁnite inner product induced by H.

9.3. Positive deﬁnite functions on the unit circle

189

9.3 Positive deﬁnite functions on the unit circle In this section we shall discuss rational matrix functions that take positive deﬁnite values on the unit circle T and their spectral factorizations. This class of functions is more complicated than the ones discussed in the previous sections, the main reason being that inﬁnity is not on the contour, and so the value at inﬁnity is not necessarily a selfadjoint matrix. A rational m × m matrix function W is called selfadjoint on the unit circle or Hermitian on the unit circle if for each λ ∈ T, λ not a pole of W , the matrix W (λ) is selfadjoint or, which is the same, Hermitian. By the uniqueness theorem for analytic functions, a rational matrix function W is selfadjoint on T if and ¯ −1 )∗ , for all λ ∈ C, λ not a pole of W . It follows that if W only if W (λ) = W (λ is selfadjoint on T and det W (λ) does not vanish identically, then W −1 is also selfadjoint on T. We ﬁrst discuss how selfadjointness of W is reﬂected in properties of the matrices in a minimal realization of the function. For proper rational matrix functions this is described in the following theorem, a counterpart of Theorem 9.1 for the unit circle. Theorem 9.6. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function. Then W is Hermitian on T if and only if A is invertible, D∗ = D − CA−1 B, and there exists an n × n matrix H such that HA = A−∗ H,

HB = A−∗ C ∗ ,

H = −H ∗ .

(9.9)

The matrix H is uniquely determined by the matrices in the given realization of W and invertible. Recall that A−∗ stands for (A∗ )−1 or, which amounts to the same, (A−1 )∗ . The ﬁrst part of (9.9) means that A is iH-unitary, that is, A is unitary with respect to the indeﬁnite inner product induced by the selfadjoint matrix iH (cf., Chapter 11 and Section 17.1). The ﬁrst part of (9.9) can be rewritten as A∗ HA = H. Note that, given the invertibility of H, the identity A∗ HA = H implies the invertibility of A. Proof. First observe that if W is Hermitian on T, then W has no pole at 0, as W (∞) = D and W (0) = W (∞)∗ . Because of minimality, this shows that A is invertible and D∗ = D − CA−1 B. But then ¯ −1 )∗ W (λ

=

D∗ + B ∗ (λ−1 − A∗ )−1 C ∗

=

D∗ − B ∗ A−∗ (λ − A−∗ )−1 λC ∗

=

D∗ − B ∗ A−∗ C ∗ − B ∗ A−∗ (λ − A−∗ )−1 A−∗ C ∗ .

¯ −1 )∗ coincide. Thus, again by Now the rational matrix functions W (λ) and W (λ the state space similarity theorem (Theorem 8.2), there exists a unique invertible

190

Chapter 9. Factorization of positive deﬁnite rational matrix functions

matrix H such that HA = A−∗ H,

HB = A−∗ C ∗ ,

−B ∗ A−∗ H = C.

Taking adjoints and employing the uniqueness of H, one ﬁnds H = −H ∗ . This settles the “only if part” of the theorem; the “if part” is obtained via a straightforward computation (not using minimality). Because of minimality, the identities in (9.9) imply that H is invertible. The argument is similar to that given in the third paragraph of the proof of Theorem 9.1. Next, we consider the associate main operator. Proposition 9.7. Let W (λ) = D + C(λIn − A)−1 B be a realization of an m × m rational matrix function and assume D is invertible. Suppose A is invertible too, D∗ = D − CA−1 B, and there exists an n × n matrix H such that (9.9) holds. Then A× = A − BD−1 C is invertible and HA× = (A× )−∗ H. Proof. From the invertibility of A and D, and the assumption D∗ = D − CA−1 B, we get A 0 A B I 0 I A−1 B = 0 D∗ 0 I C D CA−1 I =

I

BD −1

0

I

A×

0

0

D

I

0

D−1 C

I

.

As both A and D∗ are invertible, A× must be invertible too. Furthermore, by (9.9), we have (A× )∗ HA×

=

(A∗ − C ∗ D−∗ B ∗ )H(A − BD −1 C)

=

H − C ∗ D−∗ B ∗ HA − A∗ HBD−1 C + C ∗ D−∗ B ∗ HBD−1 C

=

H + C ∗ D−∗ (D − D∗ + B ∗ HB)D−1 C

=

H + C ∗ D−∗ (CA−1 B + B ∗ A−∗ C ∗ )D−1 C.

However, as D − D∗ = CA−1 B, we have B ∗ A−∗ C ∗ = −CA−1 B. Therefore, (A× )∗ HA× = H. Next we analyze how the matrix H appearing in Theorem 9.6 behaves under a state space similarity transformation on the realization of W . The proof of the next theorem is analogous to the proof of Theorem 9.3. Theorem 9.8. For i = 1, 2, let W (λ) = D + Ci (λIn − Ai )−1 Bi be minimal realizations of the rational m× m matrix function W , and let S be the (unique invertible) n × n matrix such that SA1 = A2 S,

C1 = C2 S,

B2 = SB1 .

9.3. Positive deﬁnite functions on the unit circle

191

Suppose W is Hermitian on the unit circle. For i = 1, 2, write Hi for the (unique invertible) skew-Hermitian n × n matrix such that A∗i Hi Ai = Hi and Hi Bi = ∗ ∗ A−∗ i Ci . Then H1 = S H2 S. The above results can also be obtained by reduction to the real line results of Section 9.1. To illustrate this, let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function, and let α ∈ T be a regular point for A, that is, α is not an eigenvalue of A. Consider the M¨obius transformation φ(λ) = α(λ − i)(λ + i)−1 .

(9.10)

Note that φ maps the upper half plane in a one-to-one way onto the open unit disc D, and the extended real line is mapped in a one-to-one way onto the unit circle ! (λ) = W (φ(λ)). Then W ! is again an m × m rational T, with φ(∞) = α. Put W ! admits the realization matrix function and (see Section 3.6 in [20]) the function W −1 ! W (λ) = D + C(λIn − A) B, where = (iα + iA)(α − A)−1 = φ−1 (A), A = 2iαC(α − A)−1 , C

= (α − A)−1 B, B

= W (α) = D + C(α − A)−1 B. D

Moreover this realization is again minimal. Now assume that W is selfadjoint on ! is selfadjoint on R and by Theorem 9.1 there exists an invertible n × n T, then W such that matrix H A =A ∗ H, H

B =C ∗, H

=H ∗. H

Observe that A =A ∗ H H

α − A∗ )−1 (−iα ¯ − iA∗ )H ⇐⇒ H(iα + iA)(α − A)−1 = (¯ − A) ⇐⇒ (¯ α − A∗ )H(iα + iA) = (−iα ¯ − iA∗ )H(α + iα − iαA∗ H − iA∗ HA ⇐⇒ iH ¯ HA + iα − iαA∗ H + iA∗ HA = −iH ¯ HA = 2iA∗ HA. ⇐⇒ 2iH

We already know (see the ﬁrst paragraph of the proof of Theorem 9.6) that the Using this = A−∗ H. operator A is invertible, and thus we may conclude that HA and the invertibility of H, one gets B H

=

−1 − AH −1 )−1 B − A)−1 B = (αH H(α

=

−1 A−∗ )−1 B = (α − A−∗ )−1 HB −1 − H (αH

=

= −α ¯ (¯ α − A∗ )−1 A∗ HB. (αA∗ − In )−1 A∗ HB

192

Chapter 9. Factorization of positive deﬁnite rational matrix functions

we know that C ∗ = −2iα From the deﬁnition of C ¯ (¯ α − A∗ )−1 C ∗ , and hence = 2iC ∗ . B =C ∗ ⇐⇒ A∗ HB H Then H has the properties listed in (9.9). Now deﬁne H by 2iH = H. In a similar way it can be shown that Proposition 9.7 and Theorem 9.8 follow from the analogous results in Section 9.1. We now turn to spectral factorization. Suppose W is a rational m× m matrix function. A factorization ¯ −1 )∗ L(λ) W (λ) = L(λ (9.11) is called a right spectral factorization with respect to the unit circle if L and L−1 are rational matrix functions which are analytic on the closure of the (open) unit ¯ −1 )∗ and its inverse are analytic on the closure disc D. In that case the function L(λ of Dext , the exterior domain of the unit circle in C (inﬁnity included). Thus, in particular, a right spectral factorization with respect to the unit circle is a right canonical factorization with respect to T. Analogously, (9.11) is called a left spectral factorization with respect to the unit circle if L and L−1 are analytic on the closure ¯ −1 )∗ and its inverse are of Dext (inﬁnity included), in which case the function L(λ analytic on the closed unit disc. Such a factorization is a left canonical factorization with respect to T. Observe that the existence of a spectral factorization implies that W has positive deﬁnite values on the unit circle. As we will see in the next theorem, the converse is also true. A rational m × m matrix function W is called positive deﬁnite on the unit circle if for each λ ∈ T, λ not a pole of W , the matrix W (λ) is positive deﬁnite. Left and right spectral factorization of functions which are positive deﬁnite on the unit circle is slightly more complicated than spectral factorization of functions which are positive deﬁnite on either the real line or the imaginary axis. This is mainly caused by the fact that the value at inﬁnity generally is no longer positive deﬁnite. Theorem 9.9. Let W (λ) = D + C(λIn − A)−1 B be a realization of a rational m × m matrix function such that W (λ) is positive deﬁnite for |λ| = 1. Suppose D is invertible, A is invertible, and A has no eigenvalues on the unit circle. Furthermore, assume there exists an invertible skew-Hermitian n × n matrix H such that HA = A−∗ H and HB = A−∗ C ∗ . Then, with respect to the unit circle, W admits right and left spectral factorization. Such factorizations can be obtained in the following way. Let M+ and M− be the spectral subspaces of A associated with × × the parts of σ(A) lying in Dext and D, respectively, and let M+ and M− be the spectral subspaces of A× associated with the parts of σ(A× ) lying in Dext and D, respectively. Then × ˙ M+ Cn = M − + ,

× ˙ M− Cn = M+ + .

(9.12)

9.3. Positive deﬁnite functions on the unit circle

193

× Write Π+ for the projection of Cn along M− onto M+ , Π− for the projection of Cn × −1 along M+ onto M− . Then D+ = D − CA (I − Π+ )B and D− = D − CA−1 (I − Π− )B are selfadjoint. Further there are unique rational matrix functions L+ and L− such that

¯ −1 )L+ (λ), W (λ) = L+ (λ

¯ −1 )L− (λ) W (λ) = L− (λ

are right and left spectral factorizations with respect to the unit circle, respectively, 1/2 1/2 and such that L+ (∞) = D+ , L− (∞) = D− . These functions are given by 1/2

1/2

1/2

1/2

L+ (λ)

=

D+ + D+ D−1 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D− + D− D−1 CΠ− (λIn − A)−1 B.

(9.13) (9.14)

The conditions of the theorem are satisﬁed in case W has no poles on the unit circle, takes positive deﬁnite values there, and the given (biproper) realization of W is a minimal one. Proof. Our hypotheses imply that A and A× do not have eigenvalues on the unit circle. Let P and P × be the Riesz projections of A and A× , respectively, corresponding to the eigenvalues in Dext . As in the proof of Theorem 9.4, one ﬁrst shows that HP = (I − P ∗ )H and HP × = (I − P × )∗ H, using A∗ HA = H and × × (A× )∗ HA× = H. Hence for the subspaces M+ , M− , M+ and M− we again have the identities ⊥ , HM+ = M+

⊥ HM− = M− ,

× ×⊥ HM+ = M+ ,

× ×⊥ HM− = M− .

Now introduce ϕ(λ) = −i(λ − i)(λ + i)−1 (i.e., (9.10) with α = −i). Observe that ϕ−1 = ϕ, and ϕ maps the circle to the real line, D to the open upper half plane and Dext to the open lower half plane. Consider V (λ) = W (ϕ(λ)). Then V (λ) is positive deﬁnite on the real line and − A) −1 B V (λ) = W (−i) + C(λ = (I + iA)(−A − iI)−1 . Since W (−i) is invertible, we can use Proposition with A 3.4 in [20] to show that the associate main matrix in the above realization of V is × = (I + iA× )(−A× − iI)−1 . Using A∗ HA = H and (A× )∗ HA× = H given by A with and A × are H-selfadjoint. The spectral subspaces of A one computes that A respect to the upper and lower half planes are M− and M+ , respectively, while × with respect to the upper and lower half planes are the spectral subspaces of A × × M− and M+ , respectively. From the proof of Theorem 9.4, it now follows that (9.12) holds. So the projections Π+ and Π− are well-deﬁned, and they are supporting projections giving rise to right and left canonical factorizations, respectively. Moreover H(I − Π+ ) = Π∗+ H, and H(I − Π− ) = Π∗− H.

194

Chapter 9. Factorization of positive deﬁnite rational matrix functions A canonical factorization corresponding to Π+ is given by W = W− W+ where W− (λ)

= D + C(λ − A)−1 (I − Π+ )B,

W+ (λ)

= I + D−1 CΠ+ (λ − A)−1 B.

For later use, recall that the factors W− and W+ in a canonical factorization are uniquely determined by their values at inﬁnity. It remains to show that from 1/2 −1/2 ¯ −1 )∗ . We shall in fact = L+ (λ L+ (λ) = D+ W+ (λ) it follows that W− (λ)D+ −1 ∗ ¯ prove that W (λ) = W+ (λ ) D+W+ (λ). Observe that D+ = W− (0). To see that D+ is selfadjoint, just carry out the calculation ∗ D+

= D∗ − B ∗ (I − Π∗+ )A−∗ C ∗ = D∗ − B ∗ A−∗ C ∗ + B ∗ Π∗+ A−∗ C ∗ = D − CH −1 A∗ Π∗+ HB = D − CA−1 H −1 Π∗+ HB = D − CA−1 (I − Π+ )B = D+ .

Then write W (λ) = K(λ)D+ W+ (λ) with −1 −1 K(λ) = DD+ + C(λ − A)−1 (I − Π+ )BD+ .

¯ −1 )∗ : Now compute W+ (λ ¯ −1 )∗ W+ (λ

=

I + B ∗ (λ−1 − A∗ )−1 Π∗+ C ∗ D−∗

=

I − B ∗ A−∗ Π∗+ C ∗ D −∗ −B ∗ A−∗ (λ − A−∗ )−1 A−∗ Π∗+ C ∗ D −∗

=

(D∗ − B ∗ A−∗ Π∗+ C ∗ )D −∗ +C(λ − A)−1 AH −1 Π∗+ HA−1 BD−∗ .

We claim that (D ∗ − B ∗ A−∗ Π∗+ C ∗ )D −∗ AH −1 Π∗+ HA−1 BD−∗

−1 = DD+ ,

(9.15)

−1 = (I − Π+ )BD+ .

(9.16)

Indeed, for (9.15), observe that D+ is invertible because W (0) = D∗ is invertible, and D ∗ = D+ W+ (0) = D+ (I − D −1 CΠ+ A−1 B). −1 ∗ So D+ D = D−1 (D − CΠ+ A−1 B). Taking adjoints yields (9.15).

9.3. Positive deﬁnite functions on the unit circle

195

−1 ∗ To verify (9.16), compute (I −Π+ )BD+ D , using what we just have proved: −1 ∗ D (I − Π+ )BD+

= (I − Π+ )B(I − D−1 CΠ+ A−1 B) = (I − Π+ )(A − BD−1 CΠ+ )A−1 B = (I − Π+ )(A − (A − A× )Π+ )A−1 B = (I − Π+ ){A(I − Π+ ) + A× Π+ }A−1 B.

Now Im Π+ is A× -invariant, so (I − Π+ )A× Π+ = 0. Hence (I − Π+ )BD+ D∗

=

(I − Π+ )A(I − Π+ )A−1 B

=

A(I − Π+ )A−1 B

=

AH −1 Π∗+ HA−1 B,

as Ker Π is A-invariant. Thus (9.16) holds. Now using (9.15) and (9.16) we see ¯ −1 )∗ = DD−1 + C(λ − A)−1 (I − Π+ )BD−1 = K(λ). W+ (λ + + ¯ −1 )∗ D+ W+ (λ) we see that D+ must be positive deﬁnite. Since As W (λ) = W+ (λ the factors W+ and W− in a canonical factorization are uniquely determined by their values at inﬁnity, it follows that the factor L+ in a right spectral factorization is also uniquely determined by its value at inﬁnity. Thus the part of the theorem concerned with right spectral factorization follows. For the other part dealing with left spectral factorization the reasoning is similar.

Notes The results of Section 9.1 can be found in several sources, e.g., [26] and [45]. The factorization results of Sections 9.2 and 9.3 are based on [119] (see also Chapter 1 in [120]). Spectral factorizations play an important role in mathematical systems theory, see e.g., [4]. In [4], [41] and [147] spectral factorizations of a selfadjoint rational matrix function W are studied in state space form, starting from diﬀerent representations of W . Part IV of [20] is devoted to stability of minimal factorizations of rational matrix functions. The issue of stability of factorizations within the class of spectral factorizations has also been studied. This requires the analysis of perturbations of H-selfadjoint matrices and stability of their invariant Lagrangian subspaces. For instance, from Theorem 14.12 in [20] it follows straightforwardly that canonical factorizations are Lipschitz stable under small perturbations of the matrices in the realization. Restricting attention to spectral factorizations of positive deﬁnite rational matrix functions, and to perturbations of the matrices in the realizations

196

Chapter 9. Factorization of positive deﬁnite rational matrix functions

that make the perturbed rational matrix function also positive deﬁnite, it still holds that spectral factorization is Lipschitz stable in this sense. For these and related results we refer to [123], see also [127].

Chapter 10

Pseudo-spectral factorizations of selfadjoint rational matrix functions In this chapter we consider rational matrix functions on a contour having values that are selfadjoint matrices, but not necessarily positive deﬁnite ones. Whereas in the previous chapter we studied spectral factorization, in the present chapter the focus will be on functions that have poles or zeros on the contour, and so we will consider pseudo-spectral factorization here. This chapter consists of two sections. Section 10.1 develops the notion of pseudo-spectral factorization for nonnegative rational matrix functions. The contours considered are the real line, the imaginary axis and the unit circle. In Section 10.2 the main result of the ﬁrst section is generalized to the case of arbitrary selfadjoint rational matrix functions with positive deﬁnite value at inﬁnity.

10.1 Nonnegative rational matrix functions In this section we consider rational matrix functions W having nonnegative values on either the real line, the imaginary axis or the unit circle. The section may be viewed as a continuation of the discussion in Chapter 9. However, in contrast to the situation there, in this section we consider cases where W may have poles or zeros on the contour. A rational m × m matrix function W is called nonnegative on the real line if for each λ ∈ R, λ not a pole of W , the matrix W (λ) is nonnegative. Without further explanation, the analogous terminology will be used for rational matrix functions having nonnegative values on the imaginary axis or on the unit circle, respectively.

198

Chapter 10. Pseudo-spectral factorizations

As in Section 9.2 we shall start by considering the case of nonnegative rational matrix functions W on the real line, and continue with the situation where W is nonnegative on the imaginary axis. However, it is the latter case that we shall use frequently in the subsequent chapters. Therefore only for this case shall we provide a detailed proof. The real line situation can then be dealt with by using the M¨obius transformation λ → −iλ. The section is concluded by presenting the results for the case of the unit circle. Again, the proof may be obtained by using a M¨ obius transformation (cf., the proofs of Theorems 9.4 and 9.9). A factorization ¯ ∗ L(λ) W (λ) = L(λ) (10.1) is called a right pseudo-spectral factorization with respect to the real line if L has no poles or zeros in the open upper half plane and the factorization is locally minimal at each point of the real line. Analogously, (10.1) is called a left pseudospectral factorization with respect to the real line if L has no poles or zeros in the open lower half plane and the factorization is locally minimal at each point of the real line. Such right or left pseudo-spectral factorizations are pseudo-canonical factorizations with respect to iR in the sense of Section 8.3. Although a nonnegative rational matrix function generally does not allow for a left or right spectral factorization, it does admit left and right pseudo-spectral factorization. Theorem 10.1. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a rational m × m matrix function which is nonnegative on the real line, and assume D is positive deﬁnite. Then, with respect to the real line, W admits left and right pseudo-spectral factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) Hermitian n × n matrix with HA = A∗ H and HB = C ∗ . Then there are unique A-invariant subspaces M+ and M− , and × × unique A× -invariant subspaces M+ and M− , such that (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane, and σ(A|M+ ) ⊂ {λ | λ ≤ 0}, (ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane, and σ(A|M− ) ⊂ {λ | λ ≥ 0}, × (iii) M+ contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open lower half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}, +

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open upper half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0}, −

(v) H[M+ ] =

⊥ M+ ,

H[M− ] =

⊥ M− ,

× H[M+ ]

×⊥ = M+ ,

× ×⊥ H[M− ] = M− .

The subspaces in question also satisfy the matching conditions × ˙ M+ Cn = M − + ,

× ˙ M− Cn = M+ + .

(10.2)

10.1. Nonnegative rational matrix functions

199

× Let Π+ be the projection along M− onto M+ , let Π− be the projection along M+ × onto M− , and introduce

Then

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

(10.3)

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

(10.4)

¯ ∗ L+ (λ), W (λ) = L+ (λ)

¯ ∗ L− (λ), W (λ) = L− (λ)

are right and left pseudo-spectral factorizations with respect to the real line, respectively. These pseudo-spectral factorizations are uniquely determined by the fact that they have the value D 1/2 at inﬁnity. All possible right pseudo-spectral factors can be obtained from L+ as given in (10.3) by multiplying on the left with a unitary matrix, and likewise, all possible left pseudo-spectral factors are obtained from L− as given in (10.4) by multipli¯ ∗L − (λ) − (λ) cation on the left with a unitary matrix. Indeed, suppose W (λ) = L −∗ ¯ ¯ ∗= is another left pseudo-spectral factorization of W . Put E(λ) = L− (λ) L− (λ) −1 − (λ)L− (λ) . Then E(λ) is analytic outside the real line, and on the real line it L is unitary, except for possible poles. So for all values of λ concerned, the norm of E(λ) is 1. But then E cannot have poles. Indeed, in the vicinity of a pole the norm of E(λ) cannot be bounded (cf., [134], Chapter 10, page 211). It follows that E is analytic on the whole complex plane. But then it must be a constant function by Liouville’s theorem. As it is unitary for real λ, we conclude that the sole value of E is a unitary matrix. Let W be a rational m × m, and suppose W is nonnegative on the real line. A factorization ¯ ∗ L(λ) W (λ) = L(−λ) is called a right pseudo-spectral factorization with respect to the imaginary axis if L has no poles or zeros in the open left half plane and the factorization is locally minimal at each point of the imaginary axis. Left pseudo-spectral factorizations with respect to the imaginary axis are deﬁned by replacing the upper half plane by the lower half plane. Theorem 10.2. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is nonnegative on the imaginary axis, and assume D is positive deﬁnite. Put A× = A − BD −1 C. Then, with respect to the imaginary axis, W admits left and right pseudo-spectral factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) skew-Hermitian n × n matrix with HA = −A∗ H and HB = C ∗ . Then there are unique A-invariant subspaces M+ and M− , and unique A× -invariant subspaces × × M+ and M− , such that (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open right half plane, and σ(A|M+ ) ⊂ {λ | λ ≥ 0},

200

Chapter 10. Pseudo-spectral factorizations

(ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open left half plane, and σ(A|M− ) ⊂ {λ | λ ≤ 0}, × contains the spectral subspace of A× associated with the part of σ(A× ) (iii) M+ lying in the open right half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0},, +

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open left half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}, −

(v) H[M+ ] =

⊥ M+ ,

H[M− ] =

⊥ M− ,

× H[M+ ]

×⊥ = M+ ,

× ×⊥ H[M− ] = M− .

The subspaces in question also satisfy the matching conditions × ˙ M+ , Cn = M − +

× ˙ M− Cn = M+ + .

× Let Π+ be the projection of Cn along M− onto M+ , let Π− be the projection of × n C along M+ onto M− , and deﬁne L+ and L− by (10.3) and (10.4), that is

Then

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

¯ ∗ L+ (λ), W (λ) = L+ (−λ)

¯ ∗ L− (λ), W (λ) = L− (−λ)

(10.5)

are right and left pseudo-spectral factorizations with respect to the imaginary axis, respectively. These pseudo-spectral factorizations are the unique ones for which L+ (∞) = D1/2 and L− (∞) = D1/2 . As was noted before, Theorem 10.1 can be derived from Theorem 10.2 via the transformation λ → −iλ. Conversely, Theorem 10.2 obtained from Theorem 10.1 by the transformation λ → iλ. Before we prove Theorem 10.2 we need some preparations concerning the spectral properties of nonnegative rational matrix functions. First we discuss the partial pole-multiplicities and partial zero-multiplicities of W . These notions have been deﬁned in Sections 8.2 and 8.1 of [20], respectively. We start with a minimal realization W (λ) = D + C(λ − A)−1 B. (10.6) Assume that W is biproper, i.e., D is invertible. Then the eigenvalues of A coincide with the poles of W and the eigenvalues of A× coincide with the zeros of W . More precisely, the partial multiplicities of λ as an eigenvalue of A coincide with the partial pole-multiplicities of λ as a pole of W , and the multiplicities of λ as an eigenvalue of A× coincide with the partial zero-multiplicities of λ as a zero of W (cf., [20], Section 8.4, in particular Proposition 8.23). We also need the connection between the Jordan chains of A at an eigenvalue λ0 and the co-pole functions of W at λ0 described in Proposition 8.3. For a nonnegative rational matrix function, we have the following addition to that proposition.

10.1. Nonnegative rational matrix functions

201

Proposition 10.3. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization for a rational m × m matrix function which is selfadjoint on the imaginary axis, and let H be the (unique) invertible skew-Hermitian n × n matrix such that HA = −A∗ H,

HB = C ∗ .

Let λ0 ∈ iR be an eigenvalue of A, let x0 , . . . , xk−1 be a Jordan chain for A at λ0 , and let ϕ be a co-pole function of W at λ0 corresponding to the Jordan chain ¯ has a zero of order at least k x0 , . . . , xk−1 . Then the function W (λ)ϕ(λ), ϕ(−λ) at λ0 and its Taylor expansion at λ0 has the following form: ¯ = (−1)k x0 , Hxk−1 (λ − λ0 )k + · · · W (λ)ϕ(λ), ϕ(−λ) · · · + (−1)k xk−1 , Hxk−1 (λ − λ0 )2k−1 + h.o.t. , where h.o.t. stands for higher order terms. Proof. The fact that ϕ is a co-pole function of W at λ0 implies that W (λ)ϕ(λ) is analytic at λ0 . This together with the fact that ϕ has a zero of order at least k at ¯ has a zero of order at least k at λ0 λ0 shows that the function W (λ)ϕ(λ), ϕ(−λ) too. The property that ϕ is a co-pole function of W at λ0 corresponding to the Jordan chain x0 , . . . , xk−1 means that xj =

∞

P0 (A − λ0 )ν−j−1 Bϕν ,

j = 0, . . . , k − 1

(10.7)

ν=k

(where the sum in the right-hand side of the identity is actually ﬁnite so that there is no convergence issue). Here P0 is the Riesz projection of A corresponding to the eigenvalue λ0 , and ϕν is the coeﬃcient of (λ − λ0 )ν in the Taylor expansion of ϕ at λ0 . We use this connection to compute Hxi , xk−1 . The fact that λ0 is in iR yields HP0 = P0∗ H. Indeed, since HAH −1 = −A∗ , we have that HP H −1 is the Riesz projection of −A∗ for the eigenvalue λ0 . Thus, using Proposition I.2.5 in [51], we get HP0 H −1 = P (−A∗ ; {λ0 }) = P (A∗ ; {−λ0 }) = P (A∗ ; {λ0 }) = P (A; {λ0 })∗ = P0∗ . Also, note that the vectors x0 , . . . , xk−1 belong to Im P0 . In particular, P0 xk−1 = xk−1 . Now use (10.7) and the identities HA = −A∗ H and HB = C ∗ . This gives, for i = 0, . . . , k − 1, Hxi , xk−1 = = =

∞

ν=k ∞

ν=k ∞

HP0 (A − λ0 )ν−i−1 Bϕν , xk−1 H(A − λ0 )ν−i−1 Bϕν , P0 xk−1 (−1)ν−i−1 ϕν , C(A − λ0 )ν−i−1 xk−1

ν=k

=

k+i

ν=k

(−1)ν−i−1 ϕν , Cxk−ν+i .

202

Chapter 10. Pseudo-spectral factorizations

From the ﬁnal paragraph of Section 8.1 we know that the vector Cxk−ν+1 is given by Cxk−ν+1 = (W ϕ)k−ν+i , where (W ϕ)j is the coeﬃcient of (λ − λ0 )j in the Taylor expansion of W (λ)ϕ(λ) at λ0 . So Hxi , xk−1 =

k+i

(−1)ν−i−1 ϕν , (W ϕ)k−ν+i ,

i = 0, . . . , k − 1.

(10.8)

ν=k

On the other hand we have ¯ = W (λ)ϕ(λ), ϕ(−λ)

∞

=k

(−1)ν (W ϕ)−ν , ϕν (λ − λ0 ) .

(10.9)

ν=k

Comparing formulas (10.8) and (10.9), we see that for i = 0, . . . , k − 1 the coeﬃcient of (λ − λ0 )k+i in the Taylor expansion of W (λ)ϕ(λ) at λ0 is given by (−1)i+1 xk−1 , Hxi . Now note that Hxi , xk−1 =

H(A − λ0 )k−i−1 xk−1 , xk−1

=

(−1)k−1−i Hxk−1 , (A − λ0 )k−1−i xk−1

=

(−1)k−1−i Hxk−1 , xi = (−1)k−1−i xk−1 , Hxi .

We conclude that (−1)i+1 xk−1 , Hxi = (−1)k Hxi , xk−1 , which completes the proof. Specializing to the case when W is nonnegative on iR we obtain the following result. Proposition 10.4. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization for a rational m × m matrix function which is nonnegative on iR. Assume D is positive deﬁnite, and let H be the (unique invertible) skew-Hermitian n × n matrix such that HA = −A∗ H and HB = C ∗ . Then the partial multiplicities corresponding to pure imaginary eigenvalues of A and A× are all even, the sign characteristic of (iA, iH) consists of the integers +1 only, and the sign characteristic of the pair (iA× , iH) consists of the integers −1 only. For the deﬁnition of the notion of sign characteristic the reader is referred to Section 11.2 below. Proof. Let us ﬁrst prove the proposition for the matrix A. Let λ0 = iμ0 be a pure imaginary eigenvalue of A, and let x0 , . . . , xk−1 be a maximal Jordan chain for A at λ0 . Then x0 , −ix1 , (−i)2 x2 , . . . , (−i)k−1 xk−1 is a Jordan chain of iA for its eigenvalue −μ0 . In fact, all Jordan chains of iA for −μ0 can be obtained in this way. Choose a Jordan basis for A such that relative to it the pair (iA, iH) is in canonical form (see Section 11.2). This means, in particular, that if x0 , . . . , xk−1 is a maximal Jordan chain of A for λ0 , which is part of this basis, then iHx0 , (−i)k−1 xk−1 =

10.1. Nonnegative rational matrix functions

203

ik Hx0 , xk−1 is either +1 or −1. The sequence of +1’s and −1’s, obtained in this manner, is the sign characteristic of the pair (iA, iH). Let x0 , . . . , xk−1 be as in the previous paragraph, and let ϕ(λ) = (λ − λ0 )k ϕk + (λ − λ0 )k+1 ϕk+1 + · · · be a corresponding co-pole function for W at λ0 . From Proposition 10.3 we know that on a neighborhood of λ0 ¯ = (λ − λ0 )k h(λ), W (λ)ϕ(λ), ϕ(−λ) where the scalar function h is analytic at λ0 and h(λ0 ) = (−1)k Hx0 , xk−1 . Consider the pure imaginary λ = iμ in this neighborhood. Rewriting the expression above in terms of μ − μ0 , and using the fact that W is nonnegative, one sees that k is even and (−i)k Hx0 , xk−1 > 0. This proves that the partial multiplicities corresponding to pure imaginary eigenvalues of A are even, and that the sign characteristic of the pair (iA, iH) consists of +1’s only. To prove the part of the proposition concerning A× , note that the function W (λ)−1 = D−1 − D−1 C(λ − A× )−1 BD−1 is nonnegative on iR too. Moreover, for this realization we have (−H)A× = −(A× )∗ (−H) and (−H)BD−1 = (−D −1 C)∗ . So, the corresponding indeﬁnite inner product is given by −H rather than H. The desired result now follows by basically repeating the argument given above. We now have all the equipment necessary for the proof of Theorem 10.2. Proof of Theorem 10.2. Based on Proposition 10.4 the existence and uniqueness × × , M− such that of A-invariant subspaces M+ , M− and A× -invariant subspaces M+ (i), (ii) and (iii) hold follow from Theorem 11.5 in Section 11.2 below. × To prove the ﬁrst equality in (10.2) one establishes M+ ∩ M− ⊂ Ker C as in × the proof of Theorem 9.4: use (9.2) instead of (9.1). Hence M+ ∩ M− is invariant for both A and A× . However, as the realization is minimal, an A-invariant subspace × contained in Ker C must be the zero space. Thus M+ ∩ M− = {0}. To show × × n ˙ M− it remains to note that dim M+ = dim M− = n/2. In a similar C = M+ + × ˙ M+ manner one gets Cn = M− + . × Denote by Π+ the projection along M− onto M+ , then Π+ is a supporting projection, and by Theorem 8.5 the factorization W (λ) = K(λ)L+ (λ), with L+ given by (10.3) and K(λ) = D1/2 + C(λ − A)−1 (I − Π+ )BD−1/2 , is minimal. Moreover, L+ has no poles in the open left half plane because Π+ A = Π+ AΠ+ . So L+ (λ) = D 1/2 + D−1/2 CΠ+ (λ − Π+ AΠ+ )−1 Π+ B.

204 Also

Chapter 10. Pseudo-spectral factorizations

−1/2 L−1 − CΠ+ (λ − Π+ A× Π+ )−1 Π+ BD−1/2 , + (λ) = D

¯ ∗ . Indeed, thus L+ has no zeros in the open left half plane. Finally, K(λ) = L+ (−λ) ¯ ∗ L+ (−λ)

= D 1/2 − B ∗ (λ + A∗ )−1 Π∗+ C ∗ D−1/2 = D 1/2 + C(λ − A)−1 H −1 Π∗+ HBD−1/2 .

As H[Ker Π+ ] = (Ker Π+ )⊥ and H[Im Π+ ] = (Im Π+ )⊥ , we have H −1 (Π+ )∗ H = I − Π+ . But then the factorization corresponding to Π+ is a right pseudo-spectral factorization. One proves in a similar way that Π− gives rise to a left pseudospectral factorization. Next, we introduce the notion of left and right pseudo-spectral factorizations with respect to the unit circle, Let W be a rational matrix function having nonnegative values on T. A factorization ¯ −1 )∗ L(λ) W (λ) = L(λ is called a right pseudo-spectral factorization with respect to the unit circle if L has no poles or zeros in the open unit disc and the factorization is locally minimal at each point of the unit circle. Left pseudo-spectral factorizations with respect to the unit circle are deﬁned by replacing the open unit disc D by Dext . In dealing with pseudo-spectral factorizations with respect to the unit circle, we discuss only a restricted class of rational matrix functions that are nonnegative on the unit circle, namely those which are biproper. Because of symmetry, this forces the function to have an invertible value at zero too. The restriction is induced by our methods, rather than by the problem itself. The following theorem can be obtained from using an appropriate M¨ obius transformation (cf., the proof of Theorem 9.9). Theorem 10.5. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a rational m × m matrix function which is nonnegative on the unit circle, and assume D and A are invertible. Then, with respect to the unit circle, W admits left and right pseudo-spectral factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) skew-Hermitian n × n matrix satisfying A∗ HA = H and A∗ HB = C ∗ . Then there are unique A-invariant subspaces × × M+ , M− and unique A× -invariant subspaces M+ , M− , such that (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open exterior of the unit disc, and σ(A|M+ ) ⊂ {λ | |λ| ≥ 1}, (ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open unit disc, and σ(A|M− ) ⊂ {λ | |λ| ≤ 1}, × contains the spectral subspace of A× associated with the part of σ(A× ) (iii) M+ lying in the open exterior of the unit disc, and σ(A× |M × ) ⊂ {λ | |λ| ≥ 1}, +

10.2. Selfadjoint rational matrix functions and further generalizations

205

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open unit disc, and σ(A× |M × ) ⊂ {λ | |λ| ≤ 1}, −

⊥ , (v) H[M+ ] = M+

× ×⊥ H[M+ ] = M+ ,

⊥ H[M− ] = M− ,

× ×⊥ H[M− ] = M− .

The subspaces in question also satisfy (10.2), i.e., × ˙ M− Cn = M+ + ,

× ˙ M+ Cn = M− + .

× , and let Π− be the projection Let Π+ be the projection of Cn along M− onto M+ × n of C along M+ onto M− , and deﬁne L+ and L− by (9.13) and (9.14), so 1/2

1/2

1/2

1/2

L+ (λ)

=

D+ + D+ D−1 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D− + D− D−1 CΠ− (λIn − A)−1 B,

where D+ = D − CA−1 (I − Π+ )B and D− = D − CA−1 (I − Π− )B. Then ¯ −1 )∗ L+ (λ), W (λ) = L+ (λ

¯ −1 )∗ L− (λ), W (λ) = L− (λ

are right and left pseudo-spectral factorizations with respect to the unit circle, respectively. The functions L+ and L− are the unique right and left pseudo-spectral 1/2 1/2 factors, respectively, such that L+ (∞) = D+ and L− (∞) = D− .

10.2 Selfadjoint rational matrix functions and further generalizations The main result of Section 10.1 will be generalized here to the case of an arbitrary selfadjoint rational matrix function with positive deﬁnite value at inﬁnity. We start with the case of selfadjoint functions on the real line. Theorem 10.6. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is selfadjoint on the real line, and assume D is positive deﬁnite. Then, with respect to the real line, W admits right and left pseudo-canonical factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) Hermitian n×n matrix such that HA = A∗ H and HB = C ∗ . Then there exist A-invariant subspaces M+ and M− , and A× × × and M− such that invariant subspaces M+ (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane, and σ(A|M+ ) ⊂ {λ | λ ≤ 0}, (ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane, and σ(A|M− ) ⊂ {λ | λ ≥ 0}, × (iii) M+ contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open lower half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}, +

206

Chapter 10. Pseudo-spectral factorizations

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open upper half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0}, −

× × and M− are maximal (v) M+ and M− are maximal H-nonnegative, and M+ H-nonpositive.

The subspaces in question also satisfy × ˙ M− , Cn = M + +

× ˙ M+ Cn = M− + .

× Let Π+ be the projection of |BC n onto M+ along M− , and let Π− be the projection × n of C onto M− along M+ , and introduce

L− (λ)

=

D1/2 + C(λIn − A)−1 (I − Π+ )BD−1/2 ,

(10.10)

L+ (λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

(10.11)

K+ (λ)

=

D1/2 + C(λIn − A)−1 (I − Π− )BD−1/2 ,

(10.12)

K− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

(10.13)

Then W (λ) = L− (λ)L+ (λ),

W (λ) = K+ (λ)K− (λ),

(10.14)

are right and left pseudo-canonical factorizations with respect to the real line, respectively. × × The subspaces M+ , M− , M+ and M− are not unique. In line with this, the uniqueness of the factorizations that we had at earlier occasions is lacking here. Also, not all pseudo-canonical factorizations for selfadjoint rational matrix functions are obtained in the way described in Theorem 10.6 . The theorem will be obtained from the more general result stated below.

Theorem 10.7. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is selfadjoint on the real line, and assume D is positive deﬁnite. Suppose D = D+ D− with D+ and D− m × m matrices (automatically invertible). Let H be the (unique invertible) Hermitian n× n matrix for which HA = A∗ H and HB = C ∗ . Let M+ be an A-invariant maximal Hnonnegative subspace, and let M− be an A× -invariant maximal H-nonpositive subspace. Then ˙ M− . Cn = M+ + (10.15) In that case, the projection Π of Cn along M+ onto M− is a supporting projection, and (hence) W admits a minimal factorization W (λ) = W+ (λ)W− (λ) with W+ and W− given by W+ (λ)

−1 = D+ + C(λIn − A)−1 (I − Π)BD− ,

W− (λ)

−1 = D− + D+ CΠ(λIn − A)−1 B.

10.2. Selfadjoint rational matrix functions and further generalizations

207

For the existence of A-invariant maximal H-nonnegative and maximal Hnonpositive subspaces, see Section 11.2 below. Proof. First we show that M+ ∩ M− = {0}. Choose x ∈ M+ ∩ M− . As M+ is nonnegative and M− is nonpositive, we have Hx, x = 0. On M+ the Schwartz inequality holds for the H-inner product. Since x ∈ M+ and Ax ∈ M+ , we get |HAx, x|2 ≤ HAx, Ax · Hx, x = 0. So for all x ∈ M+ ∩ M− we have HAx, x = 0. In the same way one shows that for all x ∈ M+ ∩ M− we have HA× x, x = 0. It follows that 0 = H(A − A× )x, x = HBD−1 Cx, x = C ∗ D −1 Cx, x = D −1/2 Cx2 , and hence M+ ∩ M− ⊂ Ker C. But then A× x = Ax − BCx = Ax for all x belonging to M+ ∩ M− , and so M+ ∩ M− is A-invariant. Hence CAn x = 0 for all x ∈ M+ ∩ M− and n = 0, 1, 2, . . . . So M+ ∩ M − ⊂

∞ ,

Ker CAj = {0}.

j=0

Now (see Section 11.2) every maximal nonnegative subspace has the same dimension as M+ . Also, for a maximal H-nonpositive subspace M− , the subspace ⊥ H −1 [M− ] is maximal H-nonnegative. Hence ⊥ ⊥ ] = dim M− = n − dim M− , dim M+ = dim H −1 [M−

and from this we get (10.15), i.e, the ﬁrst part of the theorem. To obtain the second part, apply Theorem 8.5. Proof of Theorem 10.6. For the existence of A-invariant subspaces M+ , M− and × × A× -invariant subspaces M+ , M− such that (i), (ii) and (iii) hold we refer to Section 11.2. The matching of the appropriate subspaces is an immediate consequence of Theorem 10.7. The factorizations (10.14), where the factors are given by (10.10)– (10.13) are minimal by Theorem 8.5. As in the proof of Theorem 10.2 one shows that L+ and K+ have no zeros or poles in the open upper half plane. In the same vein, L− and K− have no zeros or poles in the open lower half plane. Hence the factorizations in (10.14) are right and left pseudo-canonical factorizations, respectively. Analogues of Theorems 10.6 and 10.7 concerning rational matrix functions which are selfadjoint on the unit circle or imaginary axis can be derived too. An analogue of Theorem 10.7 also holds true if one takes M+ to be A-invariant maximal H-nonpositive (instead of maximal H-nonnegative) and M− to be A× invariant maximal H-nonnegative (instead of maximal H-nonpositive). A similar remark can be made concerning Theorem 10.6.

208

Chapter 10. Pseudo-spectral factorizations

We ﬁnish this section with a theorem concerning symmetric factorization of rational matrix functions which are nonnegative. Here we shall present only the case involving the imaginary axis. Theorem 10.8. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is nonnegative on iR. Assume D is positive deﬁnite, and let H be the (unique invertible) skew-Hermitian n × n matrix such that HA = −A∗ H and HB = C ∗ . Suppose M and M × are subspaces of CN for which A[M ] ⊂ M,

A× [M × ] ⊂ M × ,

H[M ] = M ⊥ ,

H[M × ] = M ×⊥ .

(10.16)

˙ M × . Let Π be the projection of Cn along M onto M × , and Then Cn = M + introduce (10.17) L(λ) = D 1/2 + D−1/2 CΠ(λIn − A)−1 B. Then ¯ ∗ L(λ) W (λ) = L(−λ)

(10.18)

is a minimal factorization. Conversely, given a minimal factorization (10.18), with L(∞) = D1/2 , the factor L is as in (10.17) for a supporting projection Π such that M = Ker Π and M × = Im Π satisfy (10.16). ˙ M ×. Proof. Let M and M × be as in the theorem. We shall show that Cn = M + The argument follows a (by now) familiar pattern. One ﬁrst shows that the intersection M ∩ M × is contained in Ker C (see, e.g., the proof of Theorem 9.4, or the proof of Theorem 10.7). Then M ∩ M × is both A-invariant and A× -invariant and contained in Ker C. By minimality (in fact observability) it follows that M ∩ M × = {0}. Since dim M = dim M × = n/2, we have the desired matching. Denote by Π the projection along M onto M × . Then Π is a supporting projection. Write the factorization of W corresponding to Π and the factorization D = D 1/2 D1/2 as W (λ) = K(λ)L(λ), where K(λ) =

D 1/2 + C(λ − A)−1 (I − Π)BD−1/2 ,

L(λ) =

D 1/2 + D−1/2 CΠ(λ − A)−1 B.

Arguing as in the proof of Theorem 9.4 we have Π∗ H = H(I − Π). Using also (9.2) ¯ ∗ = K(λ). it then follows easily that L(−λ) ¯ ∗ L(λ) is a minimal factorization with Conversely, suppose W (λ) = L(−λ) 1/2 L(∞) = D . Let Π be the corresponding supporting projection (which exists by ¯ ∗ , where L(λ) Theorem 8.5). From the fact that the left-hand factor K(λ) is L(−λ) ∗ is the right-hand factor, and using (9.2), we have Π H = H(I − Π). Thus both M = Ker Π and M × = Im Π satisfy (i) and (ii).

10.2. Selfadjoint rational matrix functions and further generalizations

209

Notes This chapter originates from [119] which deals with rational matrix functions that are selfadjoint on the real line. The term pseudo-canonical is from a later date, and is taken from [132]. The results presented here for nonnegative rational matrix functions on the unit circle are based on Section 3 of [104]. In this case, the restriction to W being invertible at inﬁnity and at zero may be lifted by considering a diﬀerent type of realization, namely, realizations of the type discussed in [79]. In mathematical systems theory also the following problem is of interest: given is a nonnegative rational matrix function W as in Theorem 10.8, without poles on the imaginary axis. One is looking for all possible factorizations W (λ) = ¯ ∗ L(λ), where L has all its poles in the open left half plane, but there is L(−λ) no condition on the zeros of L. This problem too sometimes goes by the name of “spectral factorization problem” and such factors L are sometimes also called “spectral factors”. The problem of parametrizing such factors is considered in many papers and books, see, e.g., [116] and [46] and the references given there. The papers [30], [31], provide a discussion involving computational aspects. For matrix polynomials a similar problem is considered in the literature, see e.g., [88] and [66]. For later developments on factorization of selfadjoint matrix polynomials, see [103], [125]. In [20] stability of factorizations of rational matrix functions under small perturbations of the matrices in a realization is studied. For the particular case where the function is positive semideﬁnite on the real line, and the factorizations are of the type (10.1), stability under small perturbations is treated in [123]. This involves stability of invariant Lagrangian subspaces for matrices that are selfadjoint in a space with an indeﬁnite inner product. It turns out that the left and right pseudo-spectral factorizations are stable (see Theorem 2.5 in [123]).

Chapter 11

Review of the theory of matrices in indeﬁnite inner product spaces In this chapter we present some background material on matrices in indeﬁnite inner product spaces, and review the main results from this area that are used in this book. No proofs will be provided; we refer to the literature for more information. Good sources are [68] and [70]. The material is not only useful for understanding of the results of the preceding two chapters, but is also intended for use in subsequent chapters. This chapter consists of three sections. Section 11.1 considers subspaces that are negative, positive or neutral relative to an indeﬁnite inner product and various generalizations of such subspaces. Section 11.2 deals with matrices that are selfadjoint relative to an indeﬁnite inner product, and Section 11.3 with matrices that are dissipative relative to an indeﬁnite inner product.

11.1 Subspaces of indeﬁnite inner product spaces Let H be an invertible Hermitian n × n matrix. On Cn we denote the usual inner product with ·, ·. The indeﬁnite inner product given by H is deﬁned as follows: [x, y] = Hx, y. A vector x ∈ Cn is called H-positive, H-negative, or H-neutral, respectively, if [x, x] > 0, [x, x] < 0, or [x, x] = 0, respectively. A subspace M of Cn is called H-nonnegative, H-nonpositve, or H-neutral, respectively, if [x, x] ≥ 0, [x, x] ≤ 0, or [x, x] = 0, respectively, for all x ∈ M . Observe that an H-neutral subspace is at the same time H-nonnegative and H-nonpositive.

212

Chapter 11. Matrices in indeﬁnite inner product spaces

Although the Cauchy-Schwarz inequality does not hold for just any two vectors x, y in an indeﬁnite inner product space, it does hold for vectors x, y which are both in an H-nonnegative subspace, or both in an H-nonpositive subspace. Note that it follows from this that M is H-neutral if and only if H[M ] ⊂ M ⊥ . A subspace M of Cn will be called maximal H-nonnegative whenever it is H-nonnegative and not properly contained in a larger H-nonnegative subspace. Similarly, M will be called a maximal H-nonpositive subspace if it is H-nonpositive and not properly contained in a larger H-nonpositive subspace. The ﬁrst part of the following proposition can be found in Theorem 2.3.2 in [70], the second part is Lemma 6.3 in [25]. Proposition 11.1. The dimension of any maximal H-nonnegative subspace coincides with the number of positive eigenvalues of H, while the dimension of any maximal H-nonpositive subspace coincides with the number of negative eigenvalues of H. Also, if M is maximal H-nonpositive then H −1 [M ⊥ ] is maximal Hnonnegative. A subspace M of Cn is said to be H-Lagrangian if H[M ] = M ⊥ . Such a subspace is both maximal H-nonnegative and maximal H-nonpositive, and hence such a subspace can exist only if H has as many positive eigenvalues as it has negative ones. As an example, suppose n is even, n = 2k say, and let 0 Ik H=i . −Ik 0 Then any subspace of the form M = Im [ P I ]∗ with P Hermitian will be a Lagrangian subspace. The concepts involving ordinary orthogonality have straightforward analogues for H-orthogonality. For instance, vectors x and y in Cn are H-orthogonal if [x, y] = 0. A subspace M is called H-nondegenerate in case there is no non-zero vector x ∈ M that is H-orthogonal to all vectors in M . An equivalent requirement is that M ∩ H[M ]⊥ = {0}. It follows that for H-nondegenerate subspaces M , one has ˙ H[M ]⊥ . Cn = M + Conversely, each subspace M of Cn with this property is H-nondegenerate.

11.2 H-selfadjoint matrices Let the indeﬁnite inner product on Cn be given by the invertible Hermitian matrix H. An n × n matrix A has an H-adjoint A[∗] deﬁned by [Ax, y] = [x, A[∗] y].

11.2. H-selfadjoint matrices

213

Thus A[∗] = H −1 A∗ H. The matrix A is called H-selfadjoint if A = A[∗] or which amounts to the same, HA = A∗ H. As an example, let A = Jn (λ) be the n × n upper triangular Jordan block with a real eigenvalue λ, and let H = εPn , where ε is +1 or −1, and Pn is the standard n × n involutary matrix (also called the n × n reversed identity matrix). Thus Pn is the n × n matrix with 1s on the diagonal running from the lower left corner to the upper right corner, and 0s elsewhere. Clearly H is invertible and selfadjoint while, moreover, HA = A∗ H. Hence A is H-selfadjoint. As a second example, suppose n is even, n = 2k say, let λ be non-real, and let A = diag Jk (λ), Jk (λ) be the block diagonal sum of two Jordan blocks of size k with eigenvalues λ and λ, respectively. Further, let H = P2k . Then again HA = A∗ H, so A is H-selfadjoint. It turns out that these two examples can serve as the building blocks for any pair (A, H), where A is H-selfadjoint. To state this more precisely, ﬁrst observe that if A is H-selfadjoint, and if S is an invertible matrix, then S −1 AS is S ∗ HSselfadjoint. The map (A, H) → (S −1 AS, S ∗ HS) deﬁnes an equivalence relation on the set of pairs (A, H) with A being H-selfadjoint. The following result, which can be found in [70], Theorem 5.1.1, describes a canonical form for pairs of matrices of this type. Theorem 11.2. Let A be an H-selfadjoint matrix. Then there exists an invertible matrix S such that S −1 AS is equal to the block-diagonal matrix diag Jk1 (λ1 ), . . . , Jkm (λm ), Jkm+1 (λm+1 ), Jkm+1 (λm+1 ), . . . , Jkl (λl ), Jkl (λl ) , while S ∗ HS = diag ε1 Pk1 , . . . , εm Pkm , P2km+1 , . . . , P2kl . Here λ1 , . . . , λm are the real eigenvalues of A, geometric multiplicities counted, λm+1 , λm+1 , . . . , λl , λl are the non-real eigenvalues of A, geometric multiplicities counted too, and the numbers ε1 , . . . , εm take the values +1 and −1. Behind the theorem is the fact that if A is H-selfadjoint, then the spectrum of A is closed under complex conjugation, taking (partial) multiplicities into account. By slight abuse of terminology, the ordered m-tuple (ε1 , . . . , εm ) is called the sign characteristic of the pair (A, H). It is uniquely determined by the pair (A, H) up to permutations of signs corresponding to equal Jordan blocks. Next, we consider invariant maximal H-nonnegative and invariant maximal H-nonpositive subspaces. We start again with examples. Let A be a single Jordan block of size n × n with a real eigenvalue, and take H = εPn . Denote the standard

214

Chapter 11. Matrices in indeﬁnite inner product spaces

basis of Cn by e1 , . . . , en . Introduce ⎧ span {e1 , . . . , en/2 } in case n is even, ⎪ ⎪ ⎨ M + = span {e1 , . . . , e(n+1)/2 } in case n is odd and ε = +1, ⎪ ⎪ ⎩ span {e1 , . . . , e(n−1)/2 } in case n is odd and ε = −1,

M−

⎧ span {e1 , . . . , en/2 } in case n is even, ⎪ ⎪ ⎨ = span {e1 , . . . , e(n+1)/2 } in case n is odd and ε = −1, ⎪ ⎪ ⎩ span {e1 , . . . , e(n−1)/2 } in case n is odd and ε = +1.

Then M + is A-invariant and maximal H-nonnegative, while M − is A-invariant and maximal H-nonpositive. As a second example, suppose n is even, n = 2k say, let A = Jk (λ) ⊕ Jk (λ) with λ non-real, let H = P2k , and write e1 , . . . , e2k for the standard basis of C2k . Then, for l = 0, . . . , k, we have that M = span {e1 , . . . , el , ek+1 , . . . , e2k−l } is an A-invariant H-Lagrangian subspace. If A is H-selfadjoint, and λ is a real eigenvalue of A, then the spectral invariant subspace of A corresponding to λ is H-orthogonal to the spectral invariant subspace of A corresponding to all other eigenvalues. A similar statement holds for a pair of complex conjugate non-real eigenvalues λ, λ. This allows one to build up A-invariant maximal H-nonnegative subspaces by taking direct sums of subspaces constructed “locally” as in the previous two examples. In particular the following holds, see Theorem 5.12.1 in [70]. Theorem 11.3. Let A be H-selfadjoint. The following statements hold: (i) There exists an A-invariant maximal H-nonnegative subspace Mu+ such that σ(A|Mu+ ) is in the closed upper half plane. Furthermore, any such Mu+ contains the spectral invariant subspace of A corresponding to the open upper half plane. (ii) There exists an A-invariant maximal H-nonpositive subspace Mu− such that σ(A|Mu− ) is in the closed upper half plane. Furthermore, any such Mu− contains the spectral invariant subspace of A corresponding to the open upper half plane. (iii) There exists an A-invariant maximal H-nonnegative subspace Ml+ such that σ(A|M + ) is in the closed lower half plane. Furthermore, any such Ml+ conl tains the spectral invariant subspace of A corresponding to the open lower half plane. (iv) There exists an A-invariant maximal H-nonpositive subspace Ml− such that σ(A|M − ) is in the closed lower half plane. Furthermore, any such Ml− conl tains the spectral invariant subspace of A corresponding to the open lower half plane.

11.3. H-dissipative matrices

215

Our next concern is the existence of A-invariant H-Lagrangian subspaces. These do not always exist. The next theorem gives a necessary and suﬃcient condition. Theorem 11.4. Let A be H-selfadjoint. There exists an A-invariant H-Lagrangian subspace if and only if for each real eigenvalue μ of A the following two conditions hold: (i) the number of odd partial multiplicities associated with μ is even, (ii) exactly half of those odd partial multiplicities associated with μ have sign +1 corresponding to them in the sign characteristic of (A, H), the other half have sign −1 corresponding to them. In particular, if all the partial multiplicities associated with the real eigenvalues of A are even, there does exist an A-invariant H-Lagrangian subspace. To elucidate what is said in Theorem 11.4, let us return to Theorem 11.2. With the notation employed there, write s(1), . . . , s(t) for the positive integers such that λs(j) = μ, j = 1, . . . , t. Then the numbers ks(1) , . . . , ks(t) are the partial multiplicities associated with μ, and the corresponding signs in the sign characteristic of (A, H) are εs(1) , . . . , εs(t) . Item (i) of the above theorem declares that the number of j for which ks(j) is odd is even, 2p say. Suppose ks(r1 ) , . . . , ks(r2p ) are odd. Then item (ii) of the theorem says that among the signs εs(r1 ) , . . . , εs(r2p ) there are p having the value +1 and p with the value −1. We now state a result on the uniqueness of A-invariant H-Lagrangian subspaces. In one direction, this result can be found in Theorem 5.12.4 in [70], the other direction is proved in [122]. Theorem 11.5. Assume that A is H-selfadjoint. The following two statements are equivalent: (i) There exist unique A-invariant H-Lagrangian subspaces Mu and Ml such that σ(A|Mu ) is in the closed upper half plane and σ(A|Ml ) is in the closed lower half plane; (ii) The real eigenvalues of A have even partial multiplicities, and for each real eigenvalue μ of A the signs in the sign characteristic of the pair (A, H) corresponding to the partial multiplicities associated with μ are all the same. In particular, the existence of subspaces Mu and Ml with the properties mentioned in (i) is guaranteed when A has no real eigenvalues. In this case Mu and Ml are the spectral subspaces of A associated with the part of σ(A) lying in the open upper and open lower half plane, respectively.

11.3 H-dissipative matrices Next, we turn to another class of matrices. An n × n matrix is H-dissipative if 1 ∗ 2i (HA − A H) is nonnegative. It can be shown that the spectral subspace of an

216

Chapter 11. Matrices in indeﬁnite inner product spaces

H-dissipative matrix A associated with the part of σ(A) lying in the open upper half plane is H-nonnegative, while the spectral subspace corresponding to the part of σ(A) lying in the open lower half plan is H-nonpositive. Theorem 11.6. Let A be H-dissipative. Then the following statements hold: (i) There exists an A-invariant maximal H-nonnegative subspace M+ such that σ(A|M+ ) is in the closed upper half plane. Furthermore, any such M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane. (ii) There exists an A-invariant maximal H-nonpositive subspace M− such that σ(A|M− ) is in the closed lower half plane. Furthermore, any such M− contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane. The usual proof of this result is quite involved, uses a ﬁxed point argument, and holds in an inﬁnite dimensional setting as well, see [6], [87]. A constructive argument for the ﬁnite dimensional case can be found in [129], [137]. 1 (HA − A∗ H) is posThe matrix A is said to be strictly H-dissipative if 2i itive deﬁnite. In that case A cannot have real eigenvalues. Hence, for a strictly H-dissipative matrix A, the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane is maximal H-positive, and, similarly, the spectral subspace of A corresponding to the part of σ(A) contained in the open lower half plane is maximal H-negative.

Notes The material in this chapter is taken from the books [68] and [70]. For other books in this area, with an emphasis on inﬁnite dimensional spaces, see [87], [25], and [6].

Part V Riccati equations and factorization In this part the canonical factorization theorem is presented in a diﬀerent way using the notion of an angular subspace and Riccati equations. In this case one has to look for angular subspaces that are also spectral subspaces, and the solutions of the Riccati equation must have additional spectral properties. Spectral factorization as well as pseudo-spectral factorization are described in terms of Hermitian solutions of such a Riccati equation. The study of rational matrix functions that take Hermitian values on certain curves, started in the previous part, is continued with an analysis of rational matrix functions that have Hermitian values for which the inertia is independent of the point on the curve. Such functions may still admit a symmetric canonical factorization, provided one allows for a constant Hermitian invertible matrix as a middle factor. A factorization of this type is commonly known as a J-spectral factorization. This part consists of three chapters. The ﬁrst chapter (Chapter 12), which has a preliminary character, introduces the (non-symmetric) algebraic Riccati equation and presents the state space canonical factorization theorem in terms of solutions of such an equation. Pseudo-canonical factorization is treated in an analogous way. In the second chapter (Chapter 13) the symmetric algebraic Riccati equation is introduced, and spectral factorization as well as pseudo-spectral factorization are described using such Riccati equations. In the third chapter (Chapter 14) the notion of a J-spectral factorization of a rational matrix function is introduced. Necessary and suﬃcient conditions for the existence of a such factorization are given, ﬁrst in terms of invariant subspaces and then in terms of solutions of a corresponding symmetric algebraic Riccati equation. The connection between left and right J-spectral factorization is also studied.

Chapter 12

Canonical factorization and Riccati equations In this chapter the canonical factorization theorem from Section 7.1 is presented in a diﬀerent way using the notion of an angular subspace and Riccati equations. In this case one has to look for solutions of the Riccati equation that have additional spectral properties. Section 12.1, which has a preliminary character, deals with angular subspaces, and in particular those that are also spectral subspaces. Section 12.2 deals with the connection between factorization and Riccati equations in general, while Section 12.3 contains the main result. It speciﬁes further the main theorem of the second section for the case of canonical factorization. In Section 12.4, as an application, we solve in state space form the problem of obtaining a right canonical factorization when a left one is given (or reversely).

12.1 Preliminaries on spectral angular subspaces Let X be a complex Banach space, let X1 and X2 be closed subspaces of X, and suppose ˙ X2 . (12.1) X = X1 + A closed subspace N of X is said to be angular relative to the decomposition ˙ N . In that case there is a unique operator R : X2 → X1 , called (12.1) if X = X1 + the angular operator for N , such that ) ( R N = {Rx + x | x ∈ X2 } = Im , I where I, as always in this section, stands for the identity operator on the appropriate space which can be easily identiﬁed from the context (in this case X2 ).

220

Chapter 12. Canonical factorization and Riccati equations Let N be an angular subspace of X relative to (12.1), and let T11 T12 ˙ 2 → X1 +X ˙ 2 T = : X1 +X T21 T22

(12.2)

be an operator on X. We consider the question when N is invariant under T . For this purpose, set I R ˙ 2 → X1 +X ˙ 2. E= : X1 +X 0 I This operator is invertible, and maps X2 in a one-to-one way onto N . It follows that T leaves N invariant if and only if E −1 T E leaves X2 invariant. A direct computation yields ⎤ ⎡ T11 − RT21 −RT21 R − RT22 + T11 R + T12 ⎦. E −1 T E = ⎣ (12.3) T21 T22 + T21 R This formula shows that E −1 T E leaves X2 invariant if and only if the angular operator R for N satisﬁes the algebraic Riccati equation RT21 R + RT22 − T11 R − T12 = 0.

(12.4)

More precisely, this equation is usually referred to as a nonsymmetric algebraic Riccati equation. In the next chapter we shall encounter symmetric algebraic Riccati equations. The 2 × 2 operator matrix (12.2) is often referred to as the Hamiltonian corresponding to the algebraic Riccati equation (12.4). Next, let E2 be the restriction of E to X2 considered as an operator from X2 into N . Then E2 is invertible. In fact, E2−1 is the restriction of E −1 to N viewed as an operator from N into X2 . Using (12.3) we see that E2−1 (T |N )E2 = T22 + T21 R, and hence T |N and T22 + T21 R are similar. In this section we want additionally that N is a spectral subspace of T . The next proposition shows in terms of the angular operator when this happens. Proposition 12.1. Let N be an angular subspace of X relative to the decomposition (12.1), and let T be the operator on X given by (12.2). Then N is a spectral subspace for T if and only if the angular operator R for N satisﬁes the algebraic Riccati equation (12.4) and σ(T11 − RT21) ∩ σ(T22 + T21 R) = ∅. More precisely the following holds. If N = Im P (T ; Γ), where Γ is a Cauchy contour that splits σ(T ), then σ(T22 + T21 R) is inside Γ and σ(T11 − RT21) is outside Γ. Conversely, if Γ is a Cauchy contour such that σ(T22 + T21 R) is inside Γ and σ(T11 − RT21 ) is outside Γ, then the spectrum of T does not intersect with Γ and N = Im P (T ; Γ).

12.2. Angular operators and factorization

221

Proof. We use the operator E introduced before. The operator E is invertible and maps X2 in a one-to-one way onto N . Since a spectral subspace of T is invariant under T , we may assume without loss of generality that the angular operator R for N satisﬁes the Riccati equation (12.4). Then formula (12.3) shows that E

−1

TE =

T11 − RT21

0

T21

T22 + T21 R

.

(12.5)

Since E maps X2 in a one-to-one way onto N , the space N is a spectral subspace for T if and only if X2 is a spectral subspace for E −1 T E, and we can apply Lemma 3.1 to get the desired result.

12.2 Angular operators and factorization In this section we use the concepts introduced in the previous section to bring the factorization theorem (see Section 2.6) for realizations in a diﬀerent form. The ˙ X2 main point is that throughout we work with a ﬁxed decomposition X = X1 + of the state space X of the realization that has to be factorized and the factors are described with respect to this decomposition. In the ﬁnite dimensional case this corresponds to working with a ﬁxed coordinate system. Theorem 12.2. Let W (λ) = D + C(λIX − A)−1 B be a biproper realization with state space X and input-output space Y . Let X1 and X2 be closed subspaces of ˙ X2 , let N be a closed subspace of X X such that (12.1) holds, i.e., X = X1 + ˙ N , and denote the which is angular relative to this decomposition, so X = X1 + corresponding angular operator by R. Assume A[X1 ] ⊂ X1 ,

A× [N ] ⊂ N,

(12.6)

and let D = D1 D2 with D1 and D2 invertible operators on Y . Write A11 A12 ˙ X2 → X1 + ˙ X2 , A = : X1 + 0 A22 B

=

C

=

B1 B2 C1

˙ X2 , : Y → X1 + C2

˙ X2 → Y. : X1 +

Then R satisﬁes the algebraic Riccati equation RB2 D −1 C1 R − R(A22 − B2 D−1 C2 ) + (A11 − B1 D−1 C1 )R + (A12 − B1 D−1 C2 ) = 0.

(12.7)

222

Chapter 12. Canonical factorization and Riccati equations

Introduce the functions W1 and W2 via the biproper realizations W1 (λ)

= D1 + C1 (λIX1 − A11 )−1 B1 D2−1 ,

W2 (λ)

= D2 + D1−1 C2 (λIX2 − A22 )−1 B2 .

Then W admits the factorization λ ∈ ρ(A11 ) ∩ ρ(A22 ) ⊂ ρ(A).

W (λ) = W1 (λ)W2 (λ), Also put −1 C1 , A× 11 = A11 − (B1 − RB2 )D

−1 A× (C1 R + C2 ). (12.8) 22 = A22 − B2 D

× Then, for λ ∈ ρ(A× 11 ) ∩ ρ(A22 ) ∩ ρ(A11 ) ∩ ρ(A22 ), the operators W (λ), W1 (λ) and W2 (λ) are invertible, and

W (λ)−1 = W2 (λ)−1 W1 (λ)−1 , where W1−1 (λ)

=

−1 D1−1 − D1−1 C1 (λIX1 − A× (B1 − RB2 )D−1 , 11 )

W2−1 (λ)

=

−1 D2−1 − D−1 (C1 R + C2 )(λIX2 − A× B2 D2−1 . 22 )

Proof. The ﬁrst part of the theorem is a direct consequence of the observations presented before Proposition 12.1, applied to A× . Indeed, let E be the invertible operator I R E= , 0 I = E −1 B, C = CE. Then = E −1 AE, B and write A A11 A12 − RA22 + A11 R , A = 0 A22 B

=

B1 − RB2

,

B2 C

=

C1

C1 R + C2

and it follows that ×

A =E

−1

×

A E=

A× 11

H

−B2 D−1 C1

A× 22

,

12.2. Angular operators and factorization

223

× where A× 11 and A22 are deﬁned by (12.8), and where H is equal to the left-hand side of (12.7). Now E maps X1 onto X1 and X2 onto N . Thus (12.6) implies that

1 ] ⊂ X1 , A[X

× [X2 ] ⊂ X2 . A

Hence (12.7) is satisﬁed. It remains to prove the factorization W = W1 W2 and to establish the formu On the −1 B. las for W1 , W2 and their inverses. We have W (λ) = D + C(λI − A) other hand, by the product rule for realizations, −1 B, W1 (λ)W2 (λ) = D + C(λI − A) where

= A

A11

(B1 − RB2 )D−1 (C1 R + C2 )

0

A22

.

It remains to observe that by (12.7) (B1 − RB2 )D−1 (C1 R + C2 ) = A12 − RA22 + A11 R. So W = W1 W2 . The formulas for the inverses are immediate.

The next theorem is a symmetric version of Theorem 12.2. Theorem 12.3. Let W (λ) = D + C(λIX − A)−1 B be a biproper realization with state space X and input-output space Y . Let X1 and X2 be closed subspaces of X ˙ X2 . Further, let N1 and N2 be closed subspaces of X for which with X = X1 + ˙ N2 , X = X1 +

˙ X2 , X = N1 +

˙ X2 while N1 is that is, N2 is angular relative to the decomposition X = X1 + ˙ X1 . Let R12 : X2 → X1 and R21 : X1 → X2 be the angular relative to X = X2 + corresponding angular operators. Assume A[N1 ] ⊂ N1 ,

˙ N2 , X = N1 +

A× [N2 ] ⊂ N2 ,

and let D = D1 D2 with D1 and D2 invertible operators on Y . Write A11 A12 ˙ X2 → X1 + ˙ X2 , : X1 + A = A21 A22 B

=

C

=

B1 B2 C1

˙ X2 , : Y → X1 + C2

˙ X2 → Y, : X1 +

(12.9)

224

Chapter 12. Canonical factorization and Riccati equations

and put R1 = IX1 − R12 R21 and R2 = IX2 − R21 R12 . Then R1 : X1 → X1 and R2 : X2 → X2 are invertible. Introduce the functions W1 and W2 via the biproper realizations −1 −1 R1 (B1 − R12 B2 )D2−1 , W1 (λ) = D1 + (C1 + C2 R21 ) λIX1 − (A11 + A12 R21 ) −1 W2 (λ) = D2 + D1−1 (C1 R12 + C2 )R2−1 λIX2 − (A22 − R21 A12 ) (B2 − R21 B1 ). Then W admits the factorization W (λ) = W1 (λ)W2 (λ),

λ ∈ ρ(A11 + A12 R21 ) ∩ ρ(A22 − R21 A12 ) ⊂ ρ(A).

Also put −1 A× C1 − R12 A21 + R12 B2 D−1 C1 , 11 = A11 − B1 D −1 C2 + A21 R12 − B2 D−1 C1 R12 . A× 22 = A22 − B2 D × Then, for λ ∈ ρ(A11 +A12 R21 ) ∩ ρ(A22 −R21 A12 ) ∩ ρ(A× 11 ) ∩ ρ(A22 ), the operators W (λ), W1 (λ) and W2 (λ) are invertible, and

W (λ)−1 = W2 (λ)−1 W1 (λ)−1 , where W1−1 (λ)

=

−1 D1−1 − D1−1 (C1 + C2 R21 )R1−1 (λIX1 − A× (B1 − R12 B2 )D−1 , 11 )

W2−1 (λ)

=

−1 −1 D2−1 − D−1 (C1 R12 + C2 )(λIX2 − A× R2 (B2 − R21 B1 )D2−1 . 22 )

We prepare for the proof of the theorem with a lemma. Lemma 12.4. Let X be a Banach space, and let X1 and X2 be closed subspaces of ˙ X2 . Further, let N1 and N2 be closed subspaces of X for which X with X = X1 + ˙ N2 , X = X1 +

˙ X2 , X = N1 +

˙ X2 while N1 is ani.e., N2 is angular relative to the decomposition X = X1 + ˙ gular relative to X = X2 + X1 . Let R12 : X2 → X1 and R21 : X1 → X2 be the corresponding angular operators. Then the following statements are equivalent: ˙ N2 ; (i) X = N1 + (ii) I − R21 R12 is invertible; (iii) I − R12 R21 is invertible; I R12 ˙ X2 → X1 + ˙ X2 is invertible. (iv) F = : X1 + R21 I

12.2. Angular operators and factorization

225

In case the equivalent conditions (i)–(iv) hold, the projection PN of X along N1 onto N2 is given by R12 (I − R21 R12 )−1 −R21 I , PN = I while the complementary projection I − PN can be written as I I − PN = (I − R12 R21 )−1 I −R12 . R21 Proof. The equivalence of (ii), (iii) and (iv) is straightforward. Observe that F maps X1 and X2 in a one-to-one manner onto N1 and N2 , respectively. Since ˙ X2 , it is clear that X = N1 + ˙ N2 if and only if F is invertible. So (i) X = X1 + and (iv) are equivalent. To complete the proof it remains to prove the formula for PN . Observe that the expression in the right-hand side of the claimed identity for PN does deﬁne a projection. Its image and kernel are given by I R12 , , Im Im I R21 respectively, so it is indeed equal to the projection PN .

Proof of Theorem 12.3. From Lemma 12.4 we know that the operator I R12 ˙ X2 → X1 + ˙ X2 : X1 + F = R21 I = F −1 AF, B = F −1 B and C = CF . Then W (λ) = is invertible. Introduce A −1 ˆ × ˆ ˆ D + C(λI − A) B. Note that A[X1 ] ⊂ X1 and A [X2 ] ⊂ X2 , where, following − BD −1 C, × = A and so A × = F −1 A× F . Write standard convention A . 1 12 11 A B A = = C = 1 C 2 , , B , C A 22 2 0 A B and put W1 (λ)

=

1 (λ − A 11 )−1 B 1 D−1 , D1 + C 2

W2 (λ)

=

22 )−1 B 2 (λ − A 2 . D2 + D1−1 C

11 ) ∩ ρ(A 22 ) ⊂ ρ(A) = ρ(A), the function W is the product of W1 Then on ρ(A and W2 .

226

Chapter 12. Canonical factorization and Riccati equations The inverse of F is given by −R1−1 R12 R1−1 −1 ˙ X 2 → X1 + ˙ X2 . : X1 + F = −R21 R1−1 I + R21 R1−1 R12

Using this and the expression for F , one easily sees that 11 A 1 D−1 B 2 1 C

= R1−1 (A11 + A12 R21 − R12 A21 − R12 A22 R21 ), = R1−1 (B1 − R12 B2 )D2−1 , = C1 + C2 R21 .

Now R21 satisﬁes the algebraic Riccati equation R21 A12 R21 + R21 A11 − A22 R21 − A21 = 0, 11 = A11 + A12 R21 . Thus, for the function W1 , we have and it follows that A 1 D−1 1 (λ − A 11 )−1 B W1 (λ) = D1 + C 2 −1 −1 = D1 + (C1 + C2 R21 ) λ − (A11 + A12 R21 ) R1 (B1 − R12 B2 )D2−1 , as desired. Next we compute the function W2 . Using the alternative formula I + R12 R2−1 R21 −R12 R2−1 −1 ˙ X2 → X1 + ˙ X2 F = : X1 + −R2−1 R21 R2−1 for the inverse of F , we obtain 22 A

= R2−1 (A22 − R21 A12 )R2−1 ,

2 B

= R2−1 (B2 − R21 B1 ),

1 D1−1 C

= D1−1 (C1 R12 + C2 ).

Hence, for the function W2 we get 22 )−1 B 2 (λ − A 2 W2 (λ) = D2 + D1−1 C −1 = D2 + D1−1 (C1 R12 + C2 )R2−1 λ − (A22 − R21 A12 ) (B1 − R12 B2 )D2−1 , again as desired. This proves that the factorization claimed in the theorem holds on ρ(A11 + A12 R21 ) ∩ ρ(A22 − R21 A12 )

12.3. Riccati equations and canonical factorization

227

which is a subset of ρ(A). What remains to be done is to deduce the formulas for the inverses. But this amounts to repeating the work with W replaced by W −1 . In doing so, one employs the Riccati equation R12 (A21 − B2 D−1 C1 )R12 + R12 (A22 − B2 D−1 C2 ) −(A11 − B1 D −1 C1 )R12 − (A12 − B1 D−1 C2 ) = 0

for R12 instead of the one for R21 used above. The details are omitted.

12.3 Riccati equations and canonical factorization In this section Theorem 12.2 is speciﬁed further for the case of canonical factorization. As usual, Γ is a Cauchy contour in the complex plane, F+ is its interior domain, and F− its exterior domain (inﬁnity included). Theorem 12.5. Let W (λ) = D + C(λIX − A)−1 B be a biproper realization with state space X and input-output space Y . Assume that the spectrum of A does not intersect Γ. Put X1 = Im P (A; Γ) and let X2 be a closed subspace of X such that ˙ X2 , so X = X1 + ˙ X2 . X = Im P (A; Γ) + Let D = D1 D2 with D1 and D2 invertible operators on Y , and write A11 A12 ˙ X2 → X1 + ˙ X2 , : X1 + A = 0 A22 B

=

C

=

B1 B2 C1

˙ X2 , : Y → X1 + C2

˙ X2 → Y. : X1 +

Then W admits a right canonical factorization with respect to Γ if and only if the Riccati equation RB2 D−1 C1 R − R(A22 − B2 D−1 C2 ) + (A11 − B1 D−1 C1 )R + (A12 − B1 D

−1

(12.10)

C2 ) = 0

has a (unique) solution R satisfying the constraints ⊂ σ A11 − (B1 − RB2 )D −1 C1 σ A22 − B2 D −1 (C1 R + C2 ) ⊂

F+ ,

(12.11)

F− .

(12.12)

228

Chapter 12. Canonical factorization and Riccati equations

In that case a right canonical factorization W (λ) = W− (λ)W+ (λ) of W with respect to Γ is obtained by taking W− (λ)

= D1 + C1 (λ − A11 )−1 (B1 − RB2 )D2−1 ,

W+ (λ)

= D2 + D1−1 (C1 R + C2 )(λ − A22 )−1 B2 .

Moreover, the inverses of W− and W+ are given by W−−1 (λ)

=

−1 D1−1 − D1−1 C1 (λ − A× (B1 − RB2 )D −1 , 11 )

W+−1 (λ)

=

−1 D2−1 − D−1 (C1 R + C2 )(λ − A× B2 D2−1 , 22 )

where −1 C1 , A× 11 = A11 − (B1 − RB2 )D

−1 A× (C1 R + C2 ). 22 = A22 − B2 D

With the appropriate modiﬁcations, the theorem also holds for certain contours in the Riemann sphere. For instance, if for Γ one takes the (extended) imaginary axis, one has to take for F+ the open left half plane and for F− the open right half plane. For left canonical factorizations analogous results hold: just interchange the roles of inner and outer domains (see the comment after Theorem 3.2). Proof. The subspace X1 = Im P (A; Γ) is invariant under A, and hence the zero entry in the left lower corner of the operator matrix representation of A is justiﬁed. Furthermore σ(A11 ) ⊂ F+ and σ(A22 ) ⊂ F− . ˙ X2 we have Next note that relative to the decomposition X = X1 + ⎡ ⎤ A11 − B1 D−1 C1 A12 − B1 D−1 C2 ⎦. A× = A − BD−1 C = ⎣ −B2 D−1 C1 A22 − B2 D−1 C2 Thus −A× is precisely the Hamiltonian of the Riccati equation (12.10). Assume that W admits a right canonical factorization with respect to Γ. Then, in particular, W (λ) is invertible for each λ ∈ Γ; hence, by Theorem 2.4, the spectrum of the operator A× does not intersect Γ. Thus we can use Theorem 7.1 to show that N = Ker P (A× ; Γ) is an angular subspace for the decomposition ˙ X2 . Let R be the corresponding angular operator. Since A× leaves N X = X1 + invariant, we know that R satisﬁes the Riccati equation −RB2 D−1 C1 R + R(A22 − B2 D−1 C2 ) − (A11 − B1 D−1 C1 )R

(12.13)

−(A12 − B1 D−1 C2 ) = 0, which is equivalent to (12.10). Now Proposition 12.1, applied to A× and with the roles of the interior and exterior domain of the contour Γ being reversed, shows that (12.11) and (12.12) are fulﬁlled.

12.4. Left versus right canonical factorization

229

Conversely, let R be a solution of the Riccati equation (12.10) for which (12.11) and (12.12) are satisﬁed. Thus R satisﬁes the Riccati equation (12.13) which has A× as its Hamiltonian. Hence the corresponding angular subspace N is invariant under A× . Next we again use Proposition 12.1 with T = A× and with the roles of the interior and exterior domain of the contour Γ being reversed. This yields that the spectrum of A× does not intersect Γ and that N = Ker P (A× ; Γ). Since ˙ X2 , the latter implies that N is an angular subspace of X relative to X = X1 + ˙ Ker P (A× ; Γ). But then Theorem 3.2 implies that W admits a X = Im P (A; Γ) + right canonical factorization with respect to the contour Γ. To show uniqueness of the solution R of (12.10) for which the spectral inclusions (12.11) and (12.12) are satisﬁed, it suﬃces to note that these spectral inclusions imply that N = Ker P (A× ; Γ). Indeed, in that case the angular opera˙ 2 is uniquely determined. tor R for N relative to X = X1 +X It remains to get the formulas for the factors. First note that Theorem 12.2 shows that W (λ) = W− (λ)W+ (λ) with the factors W− (λ), W+ (λ) and their inverses being of the desired form. The spectral properties of A11 and A22 , together × with those of A× 11 and A22 , show that the factorization W (λ) = W− (λ)W+ (λ) is a right canonical factorization with respect to Γ.

12.4 Left versus right canonical factorization In this section we answer the following question: if a rational matrix function W admits a left canonical factorization, under what conditions does it also have a right canonical factorization? And, if so, how can the right factorization be obtained from the left one? Our starting point is a given biproper operator function W , a Cauchy contour Γ, and a left canonical factorization W (λ) = Y+ (λ)Y− (λ),

λ ∈ Γ.

(12.14)

The biproper factors Y+ and Y− are given in terms of realizations, that is, Y+ (λ)

= D+ + C+ (λIX+ − A+ )−1 B+ ,

(12.15)

Y− (λ)

= D− + C− (λIX− − A− )−1 B− .

(12.16)

We are looking for a right canonical factorization W (λ) = W− (λ)W+ (λ). The key idea for solving this problem is the following: combine the realizations of Y+ and Y− into a realization for W using the product rule for realizations, then apply the canonical factorization theorem (Theorem 7.1) to see if a right canonical factorization exists and, if so, produce formulas for the factors. As before the interior of Γ will be denoted by F+ , the exterior by F− . We (may and) shall assume that the operators in the realizations are chosen in such a way that the operators D+ and D− are invertible, the spectra of the operators

230

Chapter 12. Canonical factorization and Riccati equations

−1 × A+ and A× + = A+ − B+ D+ C+ are contained in F− , and those of A− and A− = −1 A− − B− D− C− in F+ . Then, in particular, the spectra of A− and A+ are disjoint and the Lyapunov equation

A+ Z − ZA− = −B+ C−

(12.17)

has a unique solution Z : X− → X+ (see Section I.4 in [51]). Similarly, the Lyapunov equation × −1 −1 A× (12.18) − Z − ZA+ = B− D− D+ C+ has a unique solution Z : X+ → X− . These facts are used in the following theorem and its proof. Theorem 12.6. Let W (λ) = Y+ (λ)Y− (λ) be a left canonical factorization of W with respect to the Cauchy contour Γ, and let the factors be given by (12.15) and (12.16). Let Q : X− → X+ and P : X+ → X− be the unique solutions of the Lyapunov equations (12.17) and (12.18), respectively, that is, A+ Q − QA− = −B+ C− ,

× −1 −1 A× − P − P A+ = B− D− D+ C+ .

(12.19)

Then W has a right canonical factorization W (λ) = W− (λ)W+ (λ) with respect to Γ if and only if IX+ − QP is invertible, or, which amounts to the same, IX− − P Q is invertible. In that case, on the appropriate domains, the factors W− and W+ , and their inverses W−−1 and W+−1 , are given by W− (λ)

=

D+ + (D+ C− + C+ Q)(λIX− − A− )−1 −1 · (IX− − P Q)−1 (B− D− − P B+ ),

W+ (λ)

=

−1 D− + (D+ C+ + C− P )(IX+ − QP )−1

· (λIX+ − A+ )−1 (B+ D− − QB− ), W−−1 (λ)

=

−1 −1 D+ − D+ (D+ C− + C+ Q)(IX− − P Q)−1 −1 −1 −1 · (λIX− − A× (B− D− − P B+ )D+ , −)

W+−1 (λ)

=

−1 −1 −1 −1 D− − D− (D+ C+ + C− P )(λIX+ − A× +) −1 · (IX+ − QP )−1 (B+ D− − QB− )D− .

Proof. First we use (12.15) and (12.16) to obtain a realization for W given in the ˙ X+ and deﬁne A : X → X by form (12.14). So we write X = X− + A− 0 ˙ X+ → X − + ˙ X+ . : X− + A= B+ C− A+ Then, by the product rule (see Section 2.5), W (λ) = D+ D− +

D+ C−

C+

−1 λIX − A

B− B+ D−

.

12.4. Left versus right canonical factorization

231

The associate main operator of this realization is × −1 −1 D+ C+ A− −B− D− × ˙ X+ → X− + ˙ X+ . : X− + A = 0 A× + The spectra of A and A× do not intersect Γ. Put M = Im P (A; Γ),

M × = Ker P (A× ; Γ).

In order that W admits a right canonical factorization with respect to Γ it is ˙ M ×. necessary and suﬃcient (see Theorem 7.1) that X = M + From the matrix representation of A given above we see that Ker P (A; Γ) ˙ X+ , and hence for some Z : X− → X+ we have coincides with X+ . So X = M + I M = Im . Z The fact that M is invariant under A now amounts to (12.17). But then the operator Z must be equal to Q. In a similar way one shows that P × M = Im , I where P : X+ → X− is the unique solution of (12.18). From Lemma 12.4 we know ˙ M is equivalent to the invertibility of the matrix that the condition X = M × + I P , Q I which, in turn, is equivalent to the invertibility of I − QP or, which amounts to the same, the invertibility of I − P Q. This proves the ﬁrst part of the theorem. The formulas for the factors follow by applying Theorem 12.3 with X− , X+ , M , M × , Q and P in the role of X1 , X2 , N1 , N2 , R21 and R12 , respectively. With the obvious modiﬁcations, Theorem 12.6 holds true for canonical factorizations with respect to the usual contours in the Riemann sphere (real line and imaginary axis).

Notes This chapter is a rewritten and enriched version of Chapter 5 in [11]. Theorem 12.5 in Section 12.3 seems to be new. The material in the ﬁnal section can be found in [8]. The notion of an angular operator is standard in operator theory and goes back to [101]. The theory of Riccati equations is important in system theory; see, e.g., the text books [94], [33]. For more details on this subject we also refer to the monograph [106] and to Section 1.6 in [69].

Chapter 13

The symmetric algebraic Riccati equation As we know from the previous part there is an intimate connection between canonical factorization and Riccati equations. In this chapter this connection is developed further for the case when the rational matrix functions involved have Hermitian values on the imaginary axis. In this case the corresponding Riccati equation has additional symmetry properties too. The chapter consists of three sections. In Section 13.1 we discuss two special cases, which both lead to symmetric algebraic Riccati equations of a special type. In a somewhat more general form, this symmetric version of the algebraic Riccati equation is studied in Section 13.2, with special attention for stabilizing solutions. The study is completed in Section 13.3 where we consider Hermitian solutions of the symmetric algebraic Riccati equation and related pseudo-spectral factorizations.

13.1 Spectral factorization and Riccati equations In this section we present two illustrative special cases of spectral factorization. In both cases the corresponding Riccati equations are symmetric. For our ﬁrst case, the starting point is a rational m × m matrix function G given in realized form G(λ) = Im + C(λIn − A)−1 B, with σ(A) in the open ¯ ∗ G(λ). Clearly W is left half plane, and we consider the product W (λ) = G(−λ) a nonnegative rational m × m matrix function on the imaginary axis. We shall assume additionally that G(λ) is invertible for each λ ∈ iR, which in the present situation is equivalent to the requirement that A× = A − BC has no eigenvalue on iR. The fact that G(λ) is invertible for each λ ∈ iR means that W is positive deﬁnite on R and, as we shall see, Theorem 9.5 can be applied to show that the function W admits a left spectral factorization with respect to iR. We shall use

234

Chapter 13. The symmetric algebraic Riccati equation

Theorem 12.5 to obtain such a factorization explicitly in terms of the matrices A, B and C appearing in the realization of G. Theorem 13.1. Let G(λ) = Im + C(λIn − A)−1 B be a realization of a rational m × m matrix function G such that A has all its eigenvalues in the open left half plane. Put A× = A − BC, and assume that A× has no eigenvalue on iR. Then the Riccati equation −P BB ∗ P + P A× + (A× )∗ P = 0

(13.1)

has a unique Hermitian solution P such that A× − BB ∗ P has all its eigenvalues in ¯ ∗ G(λ) the left half plane. Furthermore, the rational matrix function W (λ) = G(−λ) admits a left spectral factorization of W with respect to the imaginary axis. In fact, ¯ ∗ L− (λ) with W (λ) = L− (−λ) L− (λ) = Im + (C + B ∗ P )(λIn − A)−1 B, is such a factorization. By Theorem 2.4, the inverse L−1 − of the spectral factor L− in the above theorem is given by ∗ × ∗ −1 B. L−1 − (λ) = Im − (C + B P )(λIn − A + BB P )

In comparable situations later on in the book, where obtaining descriptions of inverses of factors would involve only a routine application of Theorem 2.4, we will refrain from giving the expressions. Proof. We split the proof into two parts. In the ﬁrst part we show that equation (13.1) has a unique Hermitian solution P such that A× − BB ∗ P has all its eigenvalues in the left half plane. ¯ ∗ = Im − B ∗ (λIn + A∗ )−1 C ∗ . Part 1. From the given realization of G we get G(−λ) Now apply the product rule from Section 2.5). This gives W (λ) = I +

−B

∗

C

$ λ−

−A∗

C ∗C

0

A

%−1

C∗ B

.

(13.2)

It is easy to check that the hypotheses of Theorem 9.5 are satisﬁed with the skewHermitian matrix H given by 0 −In H= . (13.3) In 0 Hence W admits both a left and a right spectral factorization with respect to iR. In particular W admits both a left and a right canonical factorization with respect to the imaginary axis.

13.1. Spectral factorization and Riccati equations

235

Put F− = Cleft and F+ = Cright , where Cleft and Cright are the open left and right half planes, respectively. By hypothesis σ(A) ⊂ Cleft . So σ(−A∗ ) ⊂ Cright . Thus the realization of W in (13.2) is of the form required in Theorem 12.5, and the Riccati equation (12.10) in the theorem reduces here to −RBB ∗ R − RA× − (A× )∗ R = 0,

(13.4)

where, as usual, A× = A − BC. Since W admits a left canonical factorization with respect to the imaginary axis, (the appropriate version of) Theorem 12.5 (see the remark made between the theorem and its proof) shows that (13.4) has a unique solution R satisfying σ (A× )∗ + RBB ∗ ⊂ Cleft . (13.5) σ A× + BB ∗ R ⊂ Cleft , Here we used that σ − (A× )∗ − RBB ∗ ⊂ Cright is equivalent to the second inclusion in (13.5). Taking adjoints in (13.4) and (13.5) we see that (13.4) and (13.5) remain true if R is replaced by R∗ . But then the uniqueness of the solution implies R = R∗ . Note that for R = R∗ the two inclusions in (13.5) are equivalent. Thus we see that (13.4) has a unique Hermitian solution R satisfying the ﬁrst inclusion in (13.5). When R is replaced −P , equation (13.4) transforms into equation (13.1). Thus (13.1) has a unique Hermitian solution P satisfying σ(A× − BB ∗ P ) ⊂ Cleft . Part 2. Theorem 12.5 also yields a canonical factorization of the rational matrix function given by (13.2). In fact, such a factorization is given by W (λ) = W− (λ)W+ (λ) where the factors and their inverses are given by W− (λ)

=

I − B ∗ (λ + A∗ )−1 (C ∗ + P B),

W+ (λ)

=

W−−1 (λ)

=

I + (B ∗ P + C)(λ − A)−1 B, −1 ∗ I + B ∗ λ + (A× )∗ − P BB ∗ (C + P B),

W+−1 (λ)

=

−1 I − (B ∗ P + C) λ − A× + BB ∗ P B.

¯ ∗ , and hence Comparing the ﬁrst two expressions we see that W− (λ) = W+ (−λ) the factorization W (λ) = W− (λ)W+ (λ) is a left spectral factorization with respect to iR. Now put L− = W+ to arrive at the desired result. For our second special case, we assume that W is proper, Hermitian on the imaginary axis, and has no poles there. This implies that W can be written in the form W (λ) = D + C(λIn − A)−1 B − B ∗ (λIn + A∗ )−1 C ∗ , (13.6) where D is Hermitian and A has all its eigenvalues in the open left half plane. On the basis of this representation we shall prove the following theorem. Theorem 13.2. Let the rational m × m function W be given by (13.6), where D is positive deﬁnite and A has all its eigenvalues in the open left half plane. Assume

236

Chapter 13. The symmetric algebraic Riccati equation

additionally that W has no zeros on the imaginary axis, and put A× = A−BD−1 C. Then the Riccati equations P BD−1 B ∗ P − P A× − (A× )∗ P + C ∗ D −1 C = 0,

(13.7)

QC ∗ D−1 CQ − Q(A× )∗ − A× Q + BD−1 B ∗ = 0

(13.8)

have unique Hermitian solutions P and Q that satisfy σ A× − BD−1 B ∗ P ⊂ Cleft , σ (A× )∗ − C ∗ D −1 CQ ⊂ Cleft .

(13.9)

Furthermore, with respect to the imaginary axis, W admits left and right spectral factorizations, ¯ ∗ L− (λ), W (λ) = L− (−λ)

¯ ∗ L+ (λ), W (λ) = L+ (−λ)

(13.10)

respectively, with the factors L− and L+ being given by L− (λ) = D 1/2 + D−1/2 (C + B ∗ P )(λIn − A)−1 B,

(13.11)

L+ (λ) = D1/2 − D−1/2 (CQ + B ∗ )(λIn + A∗ )−1 C ∗ .

(13.12)

Proof. We split the proof into four parts. In the ﬁrst three parts the attention is focussed on equation (13.7) and the ﬁrst parts of (13.9) and (13.10). Part 1. From (13.6) we get %−1 $ −A∗ 0 −C ∗ ∗ C . W (λ) = D + B λ− B 0 A The main matrix of this realization has no pure imaginary eigenvalues. This follows from the assumption on the eigenvalues of A. Clearly W is selfadjoint on the imaginary axis and takes invertible values there. It follows that for λ ∈ iR the signature of the matrix W (λ), that is, the diﬀerence between the number of positive and negative eigenvalues of W (λ), does not depend on λ. As W (∞) = D is positive deﬁnite, we obtain that W (λ) is positive deﬁnite for λ ∈ iR. So the hypotheses of Theorem 9.5 are satisﬁed with the skew-Hermitian matrix H given by (13.3). Hence W admits both a left and a right spectral factorization with respect to iR. To get the formulas for the factors we will apply (the appropriate version of) Theorem 12.5 (see the remark made between the theorem and its proof) Part 2. For the case considered here the Riccati equation (12.10) in Theorem 12.5 has the form RBD −1 B ∗ R − RA× − (A× )∗ R + C ∗ D −1 C = 0. This is precisely equation (13.7) with R in place of P . Since W admits a left canonical factorization with respect to the imaginary axis, Theorem 12.5 tells us that equation (13.7) has a unique solution P satisfying σ A× − BD−1 B ∗ P ⊂ Cleft . (13.13) σ − (A× )∗ + P BD −1 B ∗ ⊂ Cright ,

13.1. Spectral factorization and Riccati equations

237

Using the symmetry properties in (13.7) and (13.13), we see that P ∗ is also a solution of (13.7) satisfying (13.13). Because of the uniqueness of P , we have P = P ∗ , and hence P is a Hermitian solution of (13.7) satisfying the ﬁrst inclusion in (13.9). On the other hand, if P is a Hermitian solution of (13.7) satisfying the ﬁrst inclusion in (13.9), then P actually satisﬁes both inclusions in (13.13), and hence P = P. Part 3. Next, we derive the ﬁrst factorization in (13.10). By Theorem 12.5 the matrix function W admits a right canonical factorization, W (λ) = W− (λ)W+ (λ), with respect to iR. The factors in this factorization are given by W− (λ)

=

D1/2 + B ∗ (λ + A∗ )−1 (−C ∗ − P B)D−1/2 ,

W+ (λ)

=

D1/2 + D−1/2 (B ∗ P + C)(λ − A)−1 B.

¯ ∗ = W− (λ), and hence the ﬁrst identity in Put L− (λ) = W+ (λ). Then L− (−λ) (13.10) holds. Moreover, the function L− (λ) is given by (13.11). Since the factorization W (λ) = W− (λ)W+ (λ) is a canonical one, we also know that the factoriza¯ ∗ L− (λ) is a left spectral factorization of W with respect to tion W (λ) = L− (−λ) iR. Part 4. Finally, to get the corresponding result for the Riccati equation (13.8) and the second factorization in (13.10), we apply the results obtained in the preceding paragraphs to V (λ) = W (−λ), that is, to V (λ) = D + B ∗ (λ − A∗ )−1 C ∗ − C(λ + A)−1 B. Note that A∗ has all its eigenvalues in Cleft . Furthermore, if the function V admits a left spectral factorization with respect to the imaginary axis, V (λ) = ¯ ∗ K− (λ) say, then W (λ) = K− (λ) ¯ ∗ K− (−λ) is a right spectral factorization K− (−λ) of W with respect to iR. We conclude this section with a few remarks about the Hermitian solutions of the Riccati equations appearing in Theorem 13.2. Let W be given by (13.6) with D positive deﬁnite. First we show that any Hermitian solution P of (13.7) is invertible whenever the pair (C, A) is observable. Suppose P x = 0. Since P is Hermitian, we also have x∗ P = 0. Then (13.7) yields x∗ C ∗ D−1 Cx = 0. As D is positive deﬁnite, this gives Cx = 0. But then, again using (13.7), we get P A× = 0, and hence P Ax = P A× x+ P BD−1 Cx = 0. So Ker P is A-invariant and is contained in Ker C. Hence Ker P is contained in Ker (C|A), and thus Ker P = {0} when Ker (C|A) = {0}. In a similar way one shows that controllability of the pair (A, B) implies that every Hermitian solution Q of (13.8) is invertible. Thus, if the realization C(λ − A)−1 B is minimal, then the Hermitian solutions of the Riccati equations (13.7) and (13.8) are automatically invertible. Now let P be an invertible Hermitian solution of (13.7). Multiplying (13.7) from both sides by P −1 shows that Q = P −1 is an invertible Hermitian solution of

238

Chapter 13. The symmetric algebraic Riccati equation

(13.8). The converse is also true, that is, if Q is an invertible Hermitian solution of (13.8), then P = Q−1 is an invertible Hermitian solution of (13.7). Thus the map P → Q = P −1 provides a one-to-one correspondence between the invertible Hermitian solutions P of (13.7) and the invertible Hermitian solutions Q of (13.8). Furthermore, in this case (with Q = P −1 ) we have σ A× − BD−1 B ∗ P = σ − (A× )∗ + C ∗ D−1 CQ . Indeed, by (13.7) we have P A× − P BD−1 B ∗ P = −(A× )∗ P + C ∗ D−1 C, and so A× − BD−1 B ∗ P

= P −1 (P A× − P BD−1 B ∗ P ) = P −1 − (A× )∗ P + C ∗ D−1 C = P −1 − (A× )∗ + C ∗ D−1 CP −1 P = P −1 − (A× )∗ + C ∗ D−1 CQ P.

In particular, if the eigenvalues of A× − BD−1 B ∗ are in the open left half plane, then those of (A× )∗ − C ∗ D−1 CQ are in the open right half plane. Comparing this with (13.9), we see that in Theorem 13.2 the matrix Q is not the inverse of the matrix P .

13.2 Stabilizing solutions The equations (13.1) and (13.7) are special cases of the general symmetric algebraic Riccati equation −P BR−1 B ∗ P + P A + A∗ P + Q = 0, (13.14) with R and Q selfadjoint, R invertible. Note that the Hamiltonian (see Section 12.1) corresponding to equation (13.14) is the 2 × 2 block matrix −Q −A∗ T = . (13.15) −BR−1B ∗ A We shall assume throughout this section that A is an n × n matrix, B an n × m matrix, Q a selfadjoint n × n matrix, and R a positive deﬁnite m × m matrix. Thus the Hamiltonian T can be viewed as an operator on C2n = Cn ⊕ Cn . We shall also assume that the pair (A, B) is stabilizable. The latter means that there there exists an m × n matrix F such that A − BF has all its eigenvalues in the open left half plane. Equation (13.14) plays an important role in optimal control theory, where one is mainly interested in stabilizing solutions P . A solution P of (13.14) is said to be iR-stabilizing, or simply stabilizing when no confusion is possible, if the matrix A − BR−1 B ∗ P has all its eigenvalues in the open left half plane. In order that such a solution exists the pair (A, B) has to be stabilizable. In general, however,

13.2. Stabilizing solutions

239

this condition is not suﬃcient. An additional condition on the eigenvalues of the Hamiltonian T is required. Theorem 13.3. Consider the symmetric algebraic Riccati equation (13.14) with R positive deﬁnite and Q selfadjoint. Then the following two statements are equivalent: (i) There exists an iR-stabilizing solution of (13.14); (ii) The pair (A, B) is stabilizable and the Hamiltonian T given by (13.15) does not have pure imaginary eigenvalues. Moreover, if (13.14) has an iR-stabilizing solution, then it is unique and Hermitian. The proof of the implication (i) ⇒ (ii) and of the ﬁnal statement of the theorem concerning the uniqueness of the iR-stabilizing solution do not require R to be positive deﬁnite; selfadjointness and invertibility of R are enough. It will be convenient ﬁrst to prove a lemma using a somewhat more general setting. For this purpose we return to the general algebraic Riccati equation which was studied in Chapter 12: XT21 X + XT22 − T11X − T12 = 0.

(13.16)

Taking T21 = −BR−1 B ∗ ,

T22 = A,

T11 = −A∗ ,

T12 = −Q,

(13.17)

and setting X = P , we see that we arrive at (13.14). Note that in this case ∗ T22 = −T11 ,

∗ T12 = T12 ,

∗ T21 = T21 .

(13.18)

In this symmetric case the coeﬃcients Tij , 1 ≤ i, j ≤ 2, are square matrices, all of the same order, n say. In what follows H will denote the Hamiltonian of (13.16), that is, H = 2 Tij i,j=1 . Note that the identities in (13.18) hold if and only if

∗

JH = −H J,

0 where J = −In

In . 0

(13.19)

We are now ready to state the lemma. Lemma 13.4. Let X be a solution of (13.16) such that σ(T22 + T21 X) ⊂ Cleft. If, in addition, the coeﬃcients of (13.16) satisfy the identities in (13.18), then the Hamiltonian H has no pure imaginary eigenvalues and σ(T11 − XT21 ) ⊂ Cright .

240

Chapter 13. The symmetric algebraic Riccati equation

Proof. We shall use freely the results of Section 12.1. Let N be the angular subspace determined by X. Then N is invariant under the Hamiltonian H and the restriction H|N is similar to the matrix T22 + T21 X. Since the identities in (13.18) are satisﬁed, (13.19) holds. The symmetry relation JH = −H ∗ J implies that the eigenvalues of H are placed symmetrically with respect to the imaginary axis (multiplicities included). Note that the dimension of the angular subspace N is equal to n, where n is the size of the matrices Tij , 1 ≤ i, j ≤ 2. Since N is invariant under H and H|N is similar to T22 + T21 X, the condition on the spectrum of T22 + T21X, implies that σ(H|N ) ⊂ Cleft . It follows that H has at least n eigenvalues (multiplicities taken into account) in Cleft. The symmetry referred to above then gives that H also has at least n eigenvalues in Cright . But the order of H is 2n. So H has precisely n eigenvalues in Cleft , and also precisely n eigenvalues in Cright . In particular, H has no eigenvalue on the imaginary axis. Next, recall formula (12.5) for the present setting, that is, E

−1

HE =

T11 − XT21

0

T21

T22 + T21 X

, where E =

In 0

X In

.

(13.20)

As H and E −1 HE have the same set of eigenvalues (multiplicities taken into account) and σ(T22 + T21 X) ⊂ Cleft , the result of the previous paragraph implies that σ(T11 − XT21 ) ⊂ Cright , which completes the proof. Corollary 13.5. Assume the coeﬃcients of the Riccati equation (13.16) satisfy the symmetry conditions in (13.18). Then equation (13.16) has at most one solution X such that σ(T22 + T21 X) ⊂ Cleft . Moreover, this solution, if it exists, is Hermitian. Proof. Assume X is a solution of (13.16) such that σ(T22 + T21 X) is a subset of Cleft. Then, by Lemma 13.4, the Hamiltonian H has no pure imaginary eigenvalues and σ(T11 − XT21 ) ⊂ Cright . But then we can apply Proposition 12.1 to show that the angular subspace N determined by X is the spectral subspace of H corresponding to the eigenvalues of H in the open left half plane. In particular, N is uniquely determined and does not depend on the particular choice of the solution X. This implies that X is also uniquely determined. Again assume that X is a solution of (13.16) such that σ(T22 + T21 X) is a subset of Cleft . Then σ(T11 − XT21) ⊂ Cright . By taking adjoints and using the identities in (13.18) we see that the latter inclusion implies that σ(T22 + T21 X ∗ ) is a subset of Cleft . Furthermore, from the identities in (13.18) it also follows that X ∗ is a solution of (13.16). But then, by the uniqueness result of the previous paragraph, X ∗ = X. Hence X is Hermitian, as desired. Proof of Theorem 13.3. The implication (i) ⇒ (ii) and the ﬁnal statements of the theorem follow directly by applying Lemma 13.4 and Corollary 13.5 with the coeﬃcients Tij , 1 ≤ i, j ≤ 2, being taken as in (13.17). It remains to prove the implication (ii) ⇒ (i). Let F be an m × n matrix such that A− BF has all its eigenvalues in the open left half plane. Such a matrix exists

13.2. Stabilizing solutions

241

because (A, B) is stabilizable. Introduce the rational m × m matrix function ∗ F R ∗ V (λ) = R + B − RF (λ − G)−1 , (13.21) B

where G=

−A∗ + F ∗ B ∗

−Q − F ∗ RF

0

A − BF

.

The fact that R is invertible implies that the realization (13.21) is biproper, and one veriﬁes easily that the associate main operator is precisely the Hamiltonian T . Thus F∗ −1 ∗ −1 −1 −1 V (λ) = R − R B − F (λ − T ) . BR−1 Since A−BF has all its eigenvalues in the open left half plane, G has no eigenvalue on the imaginary axis. By assumption the same holds true for T . Thus V has no poles or zeros on iR. In particular, V (λ) is invertible for each λ ∈ iR. With J as in (13.19) we have ∗ F R ∗ = B ∗ − RF . JG = −G∗ J, J B So, by the remark made after the proof of Theorem 9.1, the values of V on iR are selfadjoint matrices. Since V (λ) is invertible for each λ ∈ iR, it follows that the signature of the matrices V (λ) for λ ∈ iR, i.e., the diﬀerence between the number of positive and negative eigenvalues of the selfadjoint matrix V (λ), is constant. As V (∞) = R is positive deﬁnite, we obtain that V (λ) is positive deﬁnite for λ ∈ iR. Hence we know from Theorem 9.5 that V admits a left spectral factorization with respect to iR. To ﬁnish the proof of (ii) ⇒ (i), we apply (the appropriate version of) Theorem 12.5 (see the remark made between the theorem and its proof) with A11 = −A∗ + F ∗ B ∗ , B1 = F ∗ R,

B2 = B,

A12 = −Q − F ∗ RF, C1 = B ∗ ,

C2 = −RF,

A22 = A − BF, D = R.

Via these choices, equation (12.10) transforms into (13.14) with P as the unknown. Furthermore, the inclusions (12.11) and (12.12) change into σ(−A∗ + P BR−1 B ∗ ) ⊂ Cright ,

σ(A − BR−1 B ∗ P ) ⊂ Cleft .

(13.22)

The conclusion is that equation (13.14) has a unique solution P satisfying the inclusions in (13.22). The second of these shows that P is a stabilizing solution of (13.14). Thus (i) is proved.

242

Chapter 13. The symmetric algebraic Riccati equation

Let P be an iR-stabilizing solution of (13.14). Then by deﬁnition, the spectral inclusion σ(A − BR−1 B ∗ P ) ⊂ Cleft holds. Furthermore, since P is Hermitian, also σ(−A∗ + P BR−1B ∗ P ) ⊂ Cright ; see also Lemma 13.4. So one of the spectral inclusions in (13.22) implies the other one automatically; cf., the two spectral inclusions (12.11), (12.12).

13.3 Symmetric Riccati equations and pseudo-spectral factorization We now continue the discussion of Section 13.2. The object of study will be the algebraic Riccati equation A∗ P + P A + Q − (P B + S ∗ )R−1 (B ∗ P + S) = 0.

(13.23)

Observe that compared to (13.14) there are some additional terms. On the other hand, (13.23) can be rewritten in the more familiar form (13.14) as (A∗ − S ∗ R−1 B ∗ )P + P (A − BR−1 S) + (Q − S ∗ R−1 S) − P BR−1 B ∗ P = 0. The Hamiltonian of this equation is given by −A∗ + S ∗ R−1B ∗ −Q + S ∗ R−1 S T = . −BR−1 B ∗ A − BR−1 S

(13.24)

Also of importance is the rational matrix function (λ − A)−1 B Q S∗ ∗ ∗ −1 I . W (λ) = −B (λ + A ) S R I

(13.25)

Note that W is selfadjoint on the imaginary axis, and admits the realization $ %−1 −A∗ −Q −S ∗ ∗ S λ− . (13.26) W (λ) = R + B B 0 A For the inverse of W , one computes that W (λ)−1 = R−1 − R

−1

B∗

S

(λ − T )−1

−S ∗ B

R−1 .

Letting n be the order of the matrix A and the skew-Hermitian 2n × 2n matrix J as in (13.19), we have ∗ ) ( −A∗ −Q −A∗ −Q ∗ −S ∗ J = B∗ S =− J, J , B 0 A 0 A

13.3. Symmetric Riccati equations and pseudo-spectral factorization

243

and hence also JT = −T ∗ J. The hypotheses we shall have in eﬀect in this section are more stringent than those in Section 13.2. In fact, we shall assume A is an n × n matrix and B an n × m matrix such that (A, B) is a controllable pair (as opposed to the weaker condition of stabilizability). As in Section 13.2 we take R positive deﬁnite and Q selfadjoint. In the next theorem we characterize when the function W introduced above is nonnegative on the imaginary axis. The characterization is given in terms of the existence of Hermitian solutions of the Riccati equation (13.23). Also we specify further the pseudo-spectral factorization result in Theorem 10.2, again in terms of Hermitian solutions of (13.23). Theorem 13.6. Consider the Riccati equation (13.23) with (A, B) a controllable pair, R positive deﬁnite and Q selfadjoint. Let T be the matrix given by (13.24) and let W be the rational matrix function deﬁned by (13.25). Then the following statements are equivalent: (i) Equation (13.23) has a Hermitian solution P ; (ii) The rational matrix function W is nonnegative on the imaginary axis; (iii) The partial multiplicities of T at its pure imaginary eigenvalues are all even; (iv) There exists a T -invariant subspace M such that J[M ] = M ⊥ . In that case, so if the equivalent conditions (i)−(iv) hold, then, given a Hermitian solution P of (13.23), the rational matrix function W (λ) factors as ¯ ∗ L(λ), W (λ) = L(−λ) where

(13.27)

L(λ) = R1/2 + R−1/2 (B ∗ P + S)(λIn − A)−1 B.

(13.28)

⊥

Moreover, if M is a T -invariant subspace such that J[M ] = M , then M is of the form ( ) P M = Im In for a Hermitian solution P of (13.23). In addition, if both A and T |M have all their eigenvalues in the closed left half plane, then the factorization (13.27) is a pseudo-spectral factorization with respect to the imaginary axis. Proof. (i) ⇒ (ii) Suppose (13.23) has a Hermitian solution P . With this P , deﬁne L(λ) by (13.28). We then have ¯ ∗ L(λ) = L(−λ)

R − B ∗ (λ + A∗ )−1 (S ∗ + P B) + (B ∗ P + S)(λ − A)−1 B −B ∗ (λ + A∗ )−1 (S ∗ + P B)R−1 (B ∗ P + S)(λ − A)−1 B.

Using (13.23), one rewrites the last term as B ∗ (λ + A∗ )−1 Q + (λ − A∗ )P + P (A − λ) (λ − A)−1 B

244

Chapter 13. The symmetric algebraic Riccati equation

which, in turn, can be transformed into B ∗ (λ + A∗ )−1 Q(λ − A)−1 B + B ∗ P (λ − A)−1 B − B ∗ (λ + A∗ )−1 P B. ¯ ∗ L(λ) = W (λ), and (ii) holds. Moreover, the identity (13.27) is proved. Thus L(−λ) (ii) ⇒ (iii) To prove that (ii) implies (iii) a couple of preparatory remarks are needed. Let A be an n × n matrix and B be an n × m matrix. For any m × n matrix F introduce AF = A − BF, Then,

QF

SF∗

SF

R

=

I

−F ∗

0

I

Thus W (λ)

=

QF = Q − S ∗ F − F ∗ S + F ∗ RF.

SF = S − RF,

−B ∗ (λ + A∗ )−1

I

S I

F∗

0

I ·

=

−B(λ + A∗ )−1

WF (λ) =

∗

−B (λ +

R

QF

I

0

−F

I

SF∗

.

I

SF R 0 (λ − A)−1 B

F

I

I

I − B ∗ (λ + A∗ )−1 F ∗ QF SF∗ (λ − A)−1 B · . I + F (λ − A)−1 B SF R

Now introduce

Q S∗

A∗F )−1

I

QF

SF∗

SF

R

(λ − AF )−1 B I

,

and Φ(λ) = I + F (λ − A)−1 B. Then Φ(λ)−1 = I − F (λ − AF )−1 B. Using the fourth identity in Theorem 2.4 one sees that (λ − A)−1 BΦ(λ)−1 = (λ − AF )−1 B. ¯ ∗ WF (λ)Φ(λ). So W (λ) is nonnegative for λ ∈ iR if and only Thus W (λ) = Φ(−λ) if WF (λ) is nonnegative for λ ∈ iR, provided λ is not a pole of the functions involved. Next, notice that WF (λ) has the realization %−1 $ −SF∗ −A∗F −QF ∗ SF WF (λ) = R + B . λ− 0 AF B One readily computes that −A∗F −QF 0

AF

−

−SF∗ B

R−1

B∗

SF

= T,

13.3. Symmetric Riccati equations and pseudo-spectral factorization

245

where T is given by (13.24). So WF (λ)

−1

= R

−1

−R

−1

B

∗

SF

(λ − T )

−1

−SF∗ B

R−1 .

Since the pair (A, B) is controllable, we can use the pole placement theorem from mathematical systems theory (see Theorem 19.3 in Chapter 20 below), to conclude that there exists an m × n matrix F such that all the eigenvalues of AF are in the open left half plane. Using such an F , we see that the matrix −A∗F −QF 0

AF

has no imaginary eigenvalues. This allows us to show (see formula (4.7) in Section 4.3 in [20]) that the matrix functions WF (λ) 0 λI2n − T 0 and 0 I2n 0 Im are analytically equivalent on an open set containing the imaginary axis. It follows that for each λ ∈ iR the partial multiplicities of λ as an eigenvalue of T are equal to the partial multiplicities of λ as a zero of WF . Since WF is nonnegative on iR, we know from Proposition 10.4 that the partial multiplicities of λ ∈ iR as a zero of WF are even. Hence the partial multiplicities at the pure imaginary eigenvalues of T are even. Thus (ii) implies (iii). (iii) ⇒ (iv) This implication can be seen from Theorem 11.4 in Chapter 11 applied to A = iT and H = iJ. Indeed, since there are no odd partial multiplicities corresponding to pure imaginary eigenvalues of T , the condition of Theorem 11.4 is satisﬁed. Hence there exists an A-invariant subspace M such that H[M ] = M ⊥ . This subspace then is also T -invariant and satisﬁes J[M ] = M ⊥ . (iv) ⇒ (i) Let M be T -invariant subspace such that J[M ] = M ⊥ , and write X1 , M = Im X2 for appropriate n × n matrices X1 and X2 . It will be shown that X2 is invertible. Once this is done, we can take P = X1 X2−1. From T [M ] ⊂ M one obtains that P solves (13.23), while from J[M ] = M ⊥ one has P = P ∗ . Hence (i) holds. We have also shown that any T -invariant J-Lagrangian subspace M is the graph of a Hermitian ∗solution P of the Riccati equation, that is, M is of the form M = Im P I for a matrix P = P ∗ that solves (13.23). It remains to verify that Ker X2 = {0}. As dim M = n, the null spaces Ker X1 and Ker X2 have a trivial intersection. So it is suﬃcient to establish that X2 x = 0

246

Chapter 13. The symmetric algebraic Riccati equation

implies X1 x = 0. Let X2 x = 0. Then X1 x ∈ M, 0 and hence

T

X1 x

0

=

−A∗ X1 x + S ∗ R−1B ∗ X1 x −BR−1 B ∗ X1 x

∈ M.

Now M is iJ-Lagrangian, i.e., J[M ] = M ⊥ . So / 0=

T

X1 x 0

,J

X1 x 0

0 = −R−1 B ∗ X1 x, B ∗ X1 x.

As R is positive deﬁnite, we obtain B ∗ X1 x = 0. Hence −A∗ X1 x X1 x T = . 0 0 But this vector is in M , so it must be of the form X1 y . X2 y Thus X2 y = 0 and X1 y = −A∗ X1 x. As X2 y = 0, we have B ∗ X1 y = 0 by the argument given above. So B ∗ A∗ X1 x = 0. Now consider ∗2 A X1 x X1 x −A∗ X1 x 2 T =T = . 0 0 0 Repeating the argument we get B ∗ A∗2 X1 x = 0. Continuing in this way we arrive at X1 x ∈ Ker B ∗ A∗j for all j. As (A, B) is a controllable pair, the pair (B ∗ , A∗ ) is observable, and thus we see that X1 x = 0, as desired. It is easily seen that the eigenvalues of T |M coincide with those of the matrix A − BR−1 S − BR−1 B ∗ P . Thus, if both A and T |M have their eigenvalues in the closed left half plane, then the factorization (13.27) with L given by (13.28) is a pseudo-spectral factorization. Notice that the full force of the controllability condition on the pair (A, B) was only used in the last part of the proof. More precisely, the implications (i) ⇒ (ii) and (iii) ⇒ (iv) are true without any condition on (A, B), and for the implication (ii) ⇒ (iii) only stabilizability of (A, B) was used.

13.3. Symmetric Riccati equations and pseudo-spectral factorization

247

Notes The connection between Riccati equations and factorizations as discussed in Section 13.1 goes back to [147] and [41]. The main result of Section 13.2 originates from [102], see also [106], Section 9.3. The results of Section 13.1 and 13.2, and similar results for the discrete time algebraic Riccati equation, play an important role in several problems in mathematical systems theory, notably, LQ-optimal control, Kalman ﬁltering and stochastic realization (see, e.g., [84], [85], [33]). The main result of Section 13.3 appeared for the ﬁrst time in [105] and [34]. See also Chapter 7 in [106]. The parametrization of solutions of the algebraic Riccati equation in terms of invariant subspaces of the matrix T , as described in Theorem 13.6, also plays a role in [135], [136].

Chapter 14

J-spectral factorization In this chapter we continue the study of rational matrix functions that take Hermitian values on certain contours. In contrast to the previous chapters, the emphasis will not be on positive deﬁnite or nonnegative rational matrix functions, but rather on ones that have values for which the inertia is independent of the point on the contour. Such functions may still admit a symmetric canonical factorization, provided we allow for a constant Hermitian invertible matrix as a middle factor. Such a factorization is commonly known as a J-spectral factorization. We shall give necessary and suﬃcient conditions for its existence, and study the question when a function which admits a left J-spectral factorization also admits a right J-spectral factorization. This chapter consists of seven sections. The ﬁrst four sections and the one but last deal with J-spectral factorization with respect to the imaginary axis. Section 14.1 introduces the notion of J-spectral factorization. The next two sections provide necessary and suﬃcient conditions for the existence of such factorizations; in Section 14.2 these conditions are stated in terms of certain invariant subspaces and in Section 14.3 they are given in terms of Riccati equations. Two special cases are discussed in detail in Section 14.4. The ﬁfth section (Section 14.5) deals with Jspectral factorization with respect to the unit circle and the real line. Section 14.6 concerns the topic of left versus right J-spectral factorization. In Section 14.7 an alternative approach is used to derive J -spectral factorizations with respect to the unit circle. The main result of this ﬁnal section extends to a more general setting the ﬁrst main result of Section 14.5.

14.1 Deﬁnition of J-spectral factorization Throughout this chapter J is an invertible Hermitian m × m matrix. Often we shall assume additionally that J −1 = J. Thus in that case we have J = J ∗ = J −1 .

(14.1)

250

Chapter 14. J-spectral factorization

Such a matrix is called a signature matrix. Up to a congruence transformation any selfadjoint invertible matrix is a signature matrix. Suppose W is a rational m × m matrix function. A factorization ¯ ∗ JL(λ) W (λ) = L(−λ)

(14.2)

is called a right J-spectral factorization with respect to the imaginary axis if L and L−1 are rational m × m matrix functions which are analytic on the closed left ¯ ∗ and its inverse are half plane (inﬁnity included). In that case the function L(−λ) analytic on the closed right half plane (including inﬁnity). Thus a right J-spectral factorization with respect to the imaginary axis is a right canonical factorization with respect to iR featuring an additional symmetry property between the factors. A factorization (9.3) is called a left J-spectral factorization with respect to the imaginary axis if L and L−1 are rational m×m matrix functions which are analytic ¯ ∗ on the closed right half plane (inﬁnity included), in which case the function L(λ) and its inverse are analytic on the closed left half plane (inﬁnity included). Such a factorization is a left canonical factorization with respect to iR. The existence of a right or left J-spectral factorization implies that W admits a canonical factorization with respect to the imaginary axis. In particular, in order that a right or left J-spectral factorization of W exists it is necessary that W is biproper and has no poles or zeros on the imaginary axis. Furthermore, the identity (14.2) gives that W is selfadjoint on the imaginary axis. Contrary to spectral factorizations for positive deﬁnite rational matrix functions, J-spectral factorizations do not always exist for biproper rational matrix functions that satisfy the obvious necessary conditions mentioned in the previous paragraph. Since a J-spectral factorization is a canonical factorization, we can use Theorem 3.2 to prepare for an example of this phenomenon. Let ⎡ ⎤ λ−1 0 ⎢ λ + 1⎥ ⎥. W (λ) = ⎢ (14.3) ⎣λ + 1 ⎦ 0 λ−1 Obviously, W is biproper and its values on the imaginary axis are selfadjoint. Furthermore, W has no pole or zero on the imaginary axis. The function W has the minimal realization W (λ) = D + C(λ − A)−1 B, with 0 1 1 0 1 0 0 1 D= , A= , B= , C= . (14.4) 1 0 0 −1 0 −2 2 0 The associate main operator is given by A× = A − BD−1 C =

−1 0 0

1

= −A.

14.2. J-spectral factorizations and invariant subspaces

251

Now for a right canonical factorization with respect to the imaginary axis to exist, × ˙ M+ we must have C2 = M− + , where M− is the spectral subspace of A associated × with the part of σ(A) lying in the left half plane, and M+ is the spectral subspace of × × A associated with the part of σ(A ) lying in the right half plane. However, since × . Hence a right canonical factorization in this case A× = −A, we have M− = M+ of W with respect to iR does not exist. Analogously, a left canonical factorization does not exist either. Hence neither left nor right J-spectral factorizations of W with respect to the imaginary axis exist for any choice of J = J ∗ = J −1 . To further clarify the connection between J-spectral factorization and canonical factorization we present the following proposition. Proposition 14.1. Let W be a biproper rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Then W (∞) is congruent to a signature matrix J, and for such a matrix J the function W admits a right (respectively, left) J-spectral factorization with respect to the imaginary axis if and only if it admits a right (respectively, left) canonical factorization with respect to the imaginary axis. Proof. Since W is selfadjoint on the imaginary axis and proper, we see that D = W (∞) is well-deﬁned and selfadjoint. The fact that W is biproper means that D is invertible. Thus D is an invertible selfadjoint matrix, and hence congruent to a signature matrix, J say: D = E ∗ JE for some invertible matrix E. Let W (λ) = W− (λ)W+ (λ) be a right canonical factorization of W with respect to the imaginary axis. Since W , W− and W+ are biproper we have D = D− D+ , where D− = W− (∞) and D+ = W+ (∞). It follows that the factorization W (λ) = W− (λ)W+ (λ) can be rewritten as W (λ) = V− (λ)DV+ (λ), where −1 , V− (λ) = W− (λ)D−

−1 V+ (λ) = D+ W− (λ).

In particular, the values of V− and V+ at inﬁnity are equal to the m × m identity matrix. Since V+ and V+−1 are analytic on the closed right half plane (inﬁnity included) and the functions V− and V−−1 are analytic on the closed left half plane (inﬁnity included), the factorization is unique. Now we use that D is selfadjoint and that W is selfadjoint on the imaginary axis. It follows that ¯ ∗ DV− (−λ) ¯ ∗, W (λ) = V+ (−λ) and in this factorization the factors have the same analyticity properties as those in W (λ) = V− (λ)DV+ (λ). Because of the uniqueness of the latter factorization, ¯ ∗ . Recall that D = E ∗ JE. Put L(λ) = EV+ (λ). we conclude that V− (λ) = V+ (−λ) ∗ ¯ Then W (λ) = L(−λ) JL(λ), and this factorization is a left J-spectral factorization with respect to the imaginary axis. The reverse implication is trivial.

14.2 J-spectral factorizations and invariant subspaces In this section necessary and suﬃcient conditions for existence of a right or left J-spectral factorization with respect to the imaginary axis will be derived in terms

252

Chapter 14. J-spectral factorization

of invariant subspaces. It will be assumed that the obvious necessary conditions for the existence of a J-spectral factorization are satisﬁed, that is, the rational m × m matrix function W for which we wish to ﬁnd J-spectral factorizations with respect to iR is assumed to be biproper, to have no poles or zeros on iR, and to be selfadjoint on iR. We begin with two lemmas which can be viewed as further reﬁnements of Theorem 9.1(ii). Lemma 14.2. Let W be a biproper rational m×m matrix function that is selfadjoint on the imaginary axis and has no pole there. Then W admits a minimal realization W (λ) = D + B ∗ H ∗ (λI2n − A)−1 B,

(14.5)

such that D = D∗ is invertible, H is invertible, HA = −A∗ H, and the matrices A and H partition as A11 A12 , A= 0 A22

H ∗ = −H, H=

(14.6)

0

∗ −H21

H21

H22

,

(14.7)

where A11 and A22 are n × n matrices which have all their eigenvalues in the right open half plane and left open half plane, respectively. Proof. Since W is biproper, D = W (∞) is invertible. The fact that D is selfadjoint is covered by item (ii) in Theorem 9.1. p − A) −1 B be a minimal realization of W . The Next, let W (λ) = D + C(λI fact that W has no poles on iR and the minimality of the realization imply that has no eigenvalue on iR. Furthermore, using item (ii) of Theorem 9.1 again, we A know that there exists a unique invertible p × p matrix T for which we have = −A ∗ T, TA

=C ∗ , TB

T = −T∗ .

(14.8)

corresponding to the eigenvalues in the Let N+ be the spectral subspace of A ⊥ ∗ T yields T[N+ ] = N+ open right half plane. The identity T A = −A . But then ⊥ the invertibility of T implies that dim N+ = dim N+ . The latter can only happen when p is even, that is, p = 2n for some nonnegative integer n. In particular, dim N+ = n. Now let f1 , . . . , fn be an orthogonal basis of N+ , and let fn+1 , . . . , f2n ⊥ ⊥ be an orthogonal basis of N+ . Since Cn = N+ ⊕ N+ , the vectors f1 , . . . , f2n form an orthogonal basis of C2n , and we can consider the unitary matrix U that transforms the basis f1 , . . . , f2n into the standard basis e1 , . . . , e2n of C2n . Deﬁne −1 , A = U AU

B = U B,

−1 , C = CU

H = U TU ∗ .

Then W (λ) = D + C(λI2n − A)−1 B is a minimal realization of W . The fact that U −1 = U ∗ together with (14.8) shows that HA = −A∗ H,

HB = C ∗ ,

H = −H ∗ .

14.2. J-spectral factorizations and invariant subspaces

253

Thus W is of the form (14.5) and (14.6) holds. The spectral subspace M+ of A corresponding to the eigenvalues in the open right half plane is given by M+ = span {e1 , . . . , en }.

(14.9)

The ﬁrst identity in (14.8) yields ⊥ = span {en+1 , . . . , e2n }. H [span {e1 , . . . , en } = H[M+ ] = M+

(14.10)

It follows that the matrices A, and H can be partitioned as in (14.7). All blocks in these representations of A and H are n × n matrices. The zero entry in A follows from the A-invariance of M+ and the fact that this space is given by (14.9), while the zero entry in H follows from (14.10). The deﬁnition of M+ and the identity (14.9) also imply that all the eigenvalues of A11 are in the open right half plane and those of A22 are in the open left half plane. Lemma 14.3. Let W be a biproper rational m×m matrix function that is selfadjoint on the imaginary axis and has no pole there. Then W admits a minimal realization W (λ) = D + C(λI2n − A)−1 B,

(14.11)

such that D = D∗ is invertible and the matrices A, B and C can be partitioned as B1 −A∗22 A12 , B= , C = −B2∗ B1∗ , (14.12) A= 0 A22 B2 where A12 is a selfadjoint n × n matrix, A22 is a n × n matrix which has all its eigenvalues in the open left half plane, and both B1 and B2 are n × m matrices. Proof. From the preceding lemma we know that W admits a minimal realization −1 B, ∗H ∗ (λI2n − A) W (λ) = D + B is invertible, where D = D∗ is invertible, H A = −A ∗ H, H

∗ = −H, H

and H partition as and the matrices A = A

11 A

12 A

0

22 A

,

= H

0

∗ 21 −H

21 H

22 H

,

(14.13)

11 are in the open right half plane and those of A 22 such that the eigenvalues of A are in the open left half plane.

254

Chapter 14. J-spectral factorization is invertible, it follows that H 21 is invertible, and hence we can deﬁne Since H ⎤ ⎡ 1 −1 −1 ⎢ H21 − 2 H21 H22 ⎥ S=⎣ ⎦. 0 In

=B ∗H ∗ , and consider the matrices The matrix S is invertible. Put C A = S −1 AS,

B = S −1 B,

−1 , C = CS

H = S ∗ HS.

Obviously, W (λ) = D + C(λIn − A)−1 B is a minimal realization of W . It remains to prove that A, B, C can be partitioned in the desired way. A straightforward calculation shows that 0 −In ∗ ∗ . (14.14) HA = −A H, HB = C , H= In 0 S, and S −1 are all block upper triangular, the same holds true Since the matrices A, for A. The ﬁrst identity in (14.14) together with the third identity in (14.14) shows that A is of the form given in (14.12) with A12 being selfadjoint. Furthermore, since the entry in the right lower corner of S and S −1 is the n × n identity matrix we 22 , and hence A22 is an n × n matrix which has all its eigenvalues see that A22 = A in the open left half plane. The second and third identities in (14.14) show that B and C are as in (14.12). Obviously, B1 and B2 are matrices of size n × m. The external matrix D in the realizations (14.5) and (14.11) is congruent to a signature matrix J, that is, D = E ∗ JE for some invertible matrix E. Replacing W (λ) by (E ∗ )−1 W (λ)E −1 we may assume that the external matrix is actually equal to J. In the next theorem we shall make this assumption. Theorem 14.4. Let W be a rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Suppose W is given by W (λ) = J + C(λI2n − A)−1 B, where J is a signature matrix and −A∗22 A12 B1 A= , B= , 0 A22 B2

C=

−B2∗

B1∗

,

such that A12 is a selfadjoint n × n matrix, A22 is an n × n matrix which has all its eigenvalues in the open left half plane, and both B1 and B2 are n × m matrices. Then W admits a left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ),

14.2. J-spectral factorizations and invariant subspaces if and only if

⎡ A× = ⎣

−A∗22 + B1 B2∗

A12 − B1 B1∗

B2 B2∗

A22 − B2 B1∗

255 ⎤ ⎦

has no eigenvalues on the imaginary axis, and the spectral subspace of A× corresponding to its eigenvalues in the open left half plane is of the form X Im In for some Hermitian matrix X. In that case the unique left J-spectral factor L− for which L− (∞) = Im is given by L− (λ) = Im + J −1 (B1∗ − B2∗ X)(λIn − A22 )−1 B2 . In this expression (as well as in other comparable formulas below) the matrix J −1 can be replaced by J. Proof. In order to prove the ﬁrst part of the theorem, we have only to check when W admits a left canonical factorization with respect to the imaginary axis (see Proposition 14.1). Let M be a spectral subspace of A corresponding to its eigenvalues in the open right half plane. Then M = Im [I 0]∗ . Writing M × for the spectral subspace of A× corresponding to its eigenvalues in the open left half plane, the matching condition ˙ M× (14.15) Cn = M + is satisﬁed if and only if M × = Im [X ∗ I]∗ for some matrix X. With H as in (14.14), the subspace M × is iH-Lagrangian (see Section 11.1). Thus −I Im = H[M × ] = (M × )⊥ = Ker X ∗ I , X which implies X = X ∗ . Applying the left-version of Theorem 3.2 the ﬁrst part of the theorem follows. Next let us deal with the second part. So suppose (14.15) is satisﬁed and write the projection Π of Cn along M onto M × in the form 0 X Π= . 0 I Then the unique right hand factor L− in a left canonical factorization with respect to the imaginary axis of W , satisfying the additional condition that L(∞) = Im ,

256

Chapter 14. J-spectral factorization

is given by L− (λ)

=

I + J −1 CΠ(λ − ΠAΠ)−1 ΠB −1

0

B1∗

−

B2∗ X

$

λ−

0 XA22

=

I +J

=

I + J −1 (B1∗ − B2∗ X)(λ − A22 )−1 B2 ,

0

%−1

XB2

A22

B2

as was claimed.

In Section 14.5 below we shall consider J-spectral factorization for selfadjoint rational matrix functions on the real line or on the unit circle.

14.3 J-spectral factorizations and Riccati equations In this section, necessary and suﬃcient conditions for existence of a right or left Jspectral factorization with respect to the imaginary axis will be derived in terms of Riccati equations. It will be assumed that the obvious necessary conditions for the existence of a J-spectral factorization are satisﬁed, that is, the rational m × m matrix function W for which we wish to ﬁnd J-spectral factorizations with respect to iR is assumed to be biproper, to have no poles or zeros on iR, and to be selfadjoint on iR. As in Theorem 14.4 we assume that the external matrix (that is, the value at inﬁnity) is a signature matrix. Theorem 14.5. Let W be a rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Suppose W is given by W (λ) = J + C(λI2n − A)−1 B, where J is a signature matrix and B1 −A∗22 A12 , B= , A= 0 A22 B2

C=

−B2∗

B1∗

,

such that A12 is a selfadjoint n × n matrix, A22 is an n × n matrix which has all its eigenvalues in the open left half plane, and both B1 and B2 are n × m matrices. Then W admits a left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ), if and only if the algebraic Riccati equation XB2 J −1 B2∗ X + X(A22 − B2 J −1 B1∗ ) + (A∗22 − B1 J −1 B2∗ )X −A12 + B1 J

−1

B1∗

=0

(14.16)

14.3. J-spectral factorizations and Riccati equations

257

has a (unique) iR-stabilizing Hermitian solution X. In that case the unique left J-spectral factor L− for which L− (∞) = Im is given by L− (λ) = Im + J −1 (B1∗ − B2∗ X)(λIn − A22 )−1 B2 .

(14.17)

In line with the deﬁnition given in the paragraph preceding Theorem 13.14, a solution of (14.16) is said to be iR-stabilizing (or simply stabilizing) if the matrix A22 − B2 J −1 B1∗ + B2 J −1 B2∗ X has its eigenvalues in the open left half plane. Proof. In order to prove the ﬁrst part of the theorem, we have only to check when W admits a left canonical factorization with respect to the imaginary axis (see Proposition 14.1). A straightforward application of Theorem 12.5, with F+ equal to Cleft and F− equal to Cright , tells us that W admits a left canonical factorization with respect to the imaginary axis if and only if the Riccati equation (14.16) has a unique solution X satisfying the additional spectral constraints (14.18) σ − A∗22 + (B1 − XB2 )J −1 B2∗ ⊂ Cright , σ A22 − B2 J −1 (B1∗ − B2∗ X) ⊂ Cleft . (14.19) Next, note that X satisﬁes (14.16) and the spectral constraints (14.18) and (14.19) if and only if the same holds true for X ∗ . Because of uniqueness it follows that X = X ∗ . The second spectral constraint (14.19) means that X is a stabilizing solution of (14.16). This completes the proof of the ﬁrst part of the theorem. To prove the second part one applies the second part of Theorem 12.5 with D1 = J and D2 = Im . Theorem 14.6. Let W be a rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Suppose W is given by W (λ) = J + B ∗ H ∗ (λI2n − A)−1 B, where J is a signature matrix, H is invertible, HA = −A∗ H and H ∗ = −H, and the matrices A and H partition as ∗ 0 −H21 A11 A12 , H= , A= 0 A22 H21 H22 where A11 and A22 are n × n matrices which have all their eigenvalues in the open right half plane and open left half plane, respectively. Put 12 A 1 B

1 ∗ 1 A22 H22 + H22 A22 + H21 A12 , 2 2 1 = H21 B1 + H22 B2 . 2 =

(14.20) (14.21)

258

Chapter 14. J-spectral factorization

Then W admits a left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ), if and only if the algebraic Riccati equation 1 J −1 B ∗ )X (14.22) ∗ ) + (A∗ − B XB2 J −1 B2∗ X + X(A22 − B2 J −1 B 1 22 2 −1 ∗ −A12 + B1 J B = 0. 1

has a (unique) iR-stabilizing Hermitian solution X. In that case the unique left J-spectral factor L− for which L− (∞) = Im is given by ∗ − B ∗ X)(λIn − A22 )−1 B2 . L− (λ) = Im + J −1 (B 1 2

(14.23)

Recall that an iR-stabilizing solution X of (14.22) is one for which the matrix 1∗ + B2 J −1 B2∗ X has its eigenvalues in the open left half plane. A22 − B2 J −1 B = S −1 AS, B = S −1 B and Proof. Put C = B ∗ H ∗ , and consider the matrices A = CS, where C ⎡ ⎤ 1 −1 −1 H H − H 22 ⎢ 21 ⎥ 2 21 S=⎣ ⎦. 0 I −1 B, 2n − A) and from the proof of Lemma 14.3 we know Then W (λ) = J + C(λI that A, B and C partition as ⎡ ⎤ ⎡ ⎤ ∗ A 12 1 . −A B 22 =⎣ =⎣ = −B ⎦, ⎦, 1∗ , 2∗ B A B C 22 2 0 A B 2 = B2 . Since 22 = A22 and B where A ⎡ S −1

⎢H21 = ⎣ 0

⎤ 1 H22 ⎥ 2 ⎦, I

12 and B 2 are given by (14.20) and (14.21), respecone readily computes that A satisﬁes the −1 B 2n − A) tively. It follows that the realization W (λ) = J + C(λI conditions of Theorem 14.5. Note that the Riccati equation (14.16) transforms 1 and the matrix A12 by A 12 . Furinto equation (14.22) when B1 is replaced by B 1 , formula (14.17) transforms into (14.23). thermore, when passing from B1 to B But then we can apply Theorem 14.5 to ﬁnish the proof.

14.4. Two special cases of J-spectral factorization

259

Note that the procedure to ﬁnd the J-spectral factor, if it exists, now consists of two main steps. The ﬁrst is to ﬁnd a realization as in Theorem 14.6, which can be done by using an orthogonal basis transformation (see the proof of Lemma 14.2), and then to ﬁnd the stabilizing solution X of (14.22) in case it exists. With this in mind, let us return to the counterexample given in Section 14.1. Let W be the rational 2 × 2 matrix function given by (14.3). The realization of this function given in Section 14.1, involving the matrices featured in (14.4), can be rewritten as W (λ) = J + B ∗ H ∗ (λI2 − A)−1 B, where 0 1 1 0 1 0 0 −1 J= , A= , B= , H= . 1 0 0 −1 0 −2 1 0 This realization satisﬁes the conditions required in the ﬁrst part of Theorem 14.6. So it makes sense to check the situation with respect to the Riccati equation 1 = 1 0 . Since 12 = 0 and B featured in the theorem. Note that in this case A B2 = 0 −2 , it follows that in the algebraic Riccati equation (14.22) both the quadratic and the constant term vanish. Hence (14.22) reduces to a linear equation, namely 2x = 0. So x = 0 is the unique solution, and this solution is not stabilizing. Hence, W does not admit a J-spectral factorization with respect to the imaginary axis, which corroborates what was already observed in the paragraph preceding Proposition 14.1.

14.4 Two special cases of J-spectral factorization In this section we consider two special cases. The ﬁrst concerns the situation where the rational matrix function appears already as a product ¯ ∗ J V (λ) W (λ) = V (−λ)

(14.24)

where J is a signature matrix and V has all its poles in the open left half plane. This situation is encountered in several problems in mathematical systems theory, notably in the theory of H∞ -control (see Chapter 20 below). Let W be the rational m × m matrix function given by the product (14.24), where V (λ) = D+C(λIn −A)−1 B. Observe that W is selfadjoint on the imaginary axis. We assume that A has all its eigenvalues in the open left half plane and that the (possibly non-square) matrix D is of full column rank (that is, Ker D = {0}). The latter implies that D∗ J D is selfadjoint and invertible, and hence D ∗ J D is congruent to some signature matrix, J say. We are looking for a J-spectral factorization of D. Theorem 14.7. Let V (λ) = D + C(λIn − A)−1 B be a given rational p × m matrix function. Assume A has all its eigenvalues in the open left half plane and the p×m matrix D has full column rank. Let J be a p × p signature matrix, and let E be an invertible m × m matrix such that J = E ∗ D∗ J DE is an m × m signature

260

Chapter 14. J-spectral factorization

¯ ∗ J V (λ) has a matrix. Then the rational m × m matrix function W (λ) = V (−λ) left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ), if and only if the algebraic Riccati equation XBJ −1 B ∗ X + X(A − BJ −1 D∗ J C) + (A∗ − C ∗ J DJ −1 B ∗ )X (14.25) + C ∗ J DJD∗ J C − C ∗ J C = 0 has a (unique) iR-stabilizing Hermitian solution X. In that case, the corresponding left J-spectral factor of W is given by L− (λ) = E −1 + JE ∗ (D ∗ J C − B ∗ X)(λIn − A)−1 B. Recall that an iR-stabilizing solution X of (14.25) is one such that the matrix A − BJ −1 D∗ J C + BJ −1 B ∗ X has its eigenvalues in the open left half plane. = DE, B = BE, and consider the rational m × m matrix function Proof. Put D ¯ ∗ J V (λ), W (λ) = E ∗ W (λ)E = V (−λ) where V (λ) = V (λ)E = DE + C(λIn − A)−1 BE. Using the product rule for −1 B, 2n − A) realizations, we see that W admits the realization W (λ) = J + C(λI where ∗ −A∗ C ∗ J C C J DE = = = −E ∗ B ∗ E ∗ D∗ J C . A , B , C 0 A BE Obviously, W is selfadjoint on the imaginary axis. Furthermore, W is biproper. has Since A has all its eigenvalues in the open left half plane, we know that A no eigenvalue on iR, and hence W has no pole on iR. We conclude that the 2n − A) −1 B meets all the requirements of the ﬁrst realization W (λ) = J + C(λI part of Theorem 14.5. It follows that W admits a left J-spectral factorization with respect to the imaginary axis if and only if the Riccati equation (14.25) has a unique stabilizing Hermitian solution X. Moreover, in that case a left J-spectral ¯ ∗ JK− (λ) of W with respect to the imaginary axis factorization W (λ) = K− (−λ) is obtained by taking K− (λ) = Im + J −1 E ∗ (D ∗ J C − B ∗ X)(λIn − A)−1 BE. Recall that W (λ) = E −∗ W (λ)E −1 . It follows that W admits a left J-spectral ! . Thus factorization with respect to the imaginary axis if and only if so does W the result of the preceding paragraph shows that W admits a left J-spectral factorization with respect to the imaginary axis if and only if the Riccati equation (14.25) has a unique stabilizing Hermitian solution X. Moreover, in that case a ¯ ∗ JL− (λ) of W with respect to the left J-spectral factorization W (λ) = L− (−λ) imaginary axis is obtained by taking L− (λ) = K− (λ)E −1 .

14.4. Two special cases of J-spectral factorization

261

In our second example we assume that the rational m × m matrix function is given in the following manner (cf., the paragraph preceding Theorem 13.2): W (λ) = J + C(λIn − A)−1 B − B ∗ (λIn + A∗ )−1 C ∗ ,

(14.26)

where A has only eigenvalues in the open left plane and J is a signature matrix. The function W admits a realization $ %−1 ∗ −A∗ 0 C ∗ C W (λ) = J + −B λI2n − . (14.27) B 0 A This realization satisﬁes all the requirements of the ﬁrst part of Theorem 14.5, which yields immediately the following result. Theorem 14.8. Let the rational m × m matrix function W be given by (14.26), where J is a signature matrix and A has its eigenvalues in the open left half plane. Then W admits a left J-spectral factorization with respect to the imaginary axis, ¯ ∗ JL− (λ), W (λ) = L− (−λ) if and only if the algebraic Riccati equation XBJB ∗ X + X(A − BJC) + (A∗ − C ∗ JB ∗ )X + C ∗ JC = 0 has a (unique) Hermitian solution X such that the matrix A − BJC + BJB ∗ X has all its eigenvalues in the open left half plane (so X is iR-stabilizing). In that case the unique left J-spectral factor L− for which L− (∞) = Im and its inverse L−1 − are given by L− (λ) = Im + J(C − B ∗ X)(λIn − A)−1 B, So far we have mainly concentrated on left J-spectral factorizations. The analogous results for right J-spectral factorization of W can be obtained by simply applying the left factorization results to V (λ) = W (−λ). Indeed, a left J-spectral factorization, ¯ ∗ JK− (λ), V (λ) = K− (−λ) of V with respect to iR yields a right J-spectral factorization, ¯ ∗ JL+(λ), W (λ) = L+ (−λ) of W with respect to iR by taking L+ (λ) = K− (−λ). Let us apply this observation to W given by the realization (14.27). Note that %−1 $ −A 0 B ∗ V (λ) = W (−λ) = J + −C B . λI2n − 0 A∗ C∗ Since A has all its eigenvalues in the open left half plane, the same holds true for A∗ . Thus we can apply Theorem 14.8 together with the above scheme to get the following right J-spectral factorization result.

262

Chapter 14. J-spectral factorization

Theorem 14.9. Let the rational matrix function W be given by (14.26), where J is a signature matrix and A has its eigenvalues in the open left half plane. Then W admits a right J-spectral factorization with respect to the imaginary axis, ¯ ∗ JL+(λ), W (λ) = L+ (−λ) if and only if the algebraic Riccati equation Y C ∗ JCY + Y (A∗ − C ∗ JB ∗ ) + (A − BJC)Y + BJB ∗ = 0

(14.28)

has a (unique) Hermitian solution Y such that A∗ − C ∗ JB + C ∗ JCY has all its eigenvalues in the open left half plane (so X is iR-stabilizing). In that case the unique right J-spectral factor L+ for which L+ (∞) = Im and its inverse L−1 + are given by L+ (λ) = Im + J(CY − B ∗ )(λIn + A∗ )−1 C ∗ .

14.5 J-spectral factorization with respect to other contours In this section we consider J-spectral factorizations with respect to the real line R and to the unit circle T featuring an additional symmetry property between the factors. Here, as before, J is is an invertible Hermitian m × m matrix. We begin by considering the case of the unit circle. Suppose W is a rational m × m matrix function. A factorization ¯ −1 )∗ JL(λ) W (λ) = L(λ

(14.29)

is called a right J-spectral factorization with respect to the unit circle if L and L−1 are rational m × m matrix functions which are analytic on the closed unit disc. In ¯ −1 )∗ and its inverse are analytic on the closure of Dext that case the function L(λ (inﬁnity included). Thus a right J-spectral factorization with respect to the unit circle is a right canonical factorization with respect to T featuring an additional symmetry property between the factors. A factorization (14.29) is called a left J-spectral spectral factorization with respect to the unit circle if L and L−1 are rational m × m matrix functions which are analytic on the closure of Dext (inﬁnity ¯ −1 )∗ and its inverse are analytic on the included), in which case the function L(λ closed unit disc. Such a factorization is a left canonical factorization with respect to T. The case of J-spectral factorization with respect to the unit circle is somewhat more complicated than that of J-spectral factorization with respect to the imaginary axis. The ﬁrst result is an analogue of Proposition 14.1. Proposition 14.10. Let W be a rational m×m matrix function that is selfadjoint on the unit circle and has neither poles nor zeros there. Then there exists a signature matrix J such for each λ ∈ T the matrix W (λ) is congruent to J. For such a matrix

14.5. J-spectral factorization with respect to other contours

263

J, the function W admits a right (respectively, left) J-spectral factorization with respect to the unit circle if and only if it admits a right (respectively, left) canonical factorization with respect to the unit circle. We can use a M¨obius transform to reduce the case of the unit circle to the case of the imaginary axis. To be precise, let V (λ) = W (λ − i)/(λ + i) . Then V is a rational m × m matrix function that has neither poles nor zeros on the imaginary axis, and has selfadjoint values there. Moreover, V (∞) = W (1), and thus V is biproper. Also, right and left J-spectral factorizations of W , and right and left canonical factorization of W can easily be obtained from the corresponding factorizations of V . Thus the proposition above actually follows from Proposition 14.1. For the sake of completeness we shall give a direct proof. Proof. By assumption, W (λ) is invertible and selfadjoint for each λ ∈ T. Thus the number of eigenvalues of W (λ) in the open unit disc does not depend on the particular choice of λ ∈ T. In other words W (λ) has constant signature on T. Now let J be a signature matrix the signature of which is equal to this constant signature. Then for each λ ∈ T the matrix W (λ) is congruent to J. Let W (λ) = W− (λ)W+ (λ) be a right canonical factorization of W with respect to T. Consider ¯ −1 )∗ , !+ (λ) = W+ (λ W

¯ −1 )∗ . !− (λ) = W− (λ W

!+ (λ)W !− (λ) is again a right canonical factorization of W with reThen W (λ) = W !+ (λ)−1 W− (λ) is a constant matrix, F say. This shows spect to T. It follows that W −1 ∗ ¯ that W (λ) = W+ (λ ) F W+ (λ). Since W (λ) is selfadjoint for λ ∈ T, it follows that F is congruent to the signature matrix J introduced in the ﬁrst paragraph of the proof. Thus F = E ∗ JE for some invertible matrix E. Put L+ (λ) = EW+ (λ). ¯ −1 )∗ JL+ (λ) is a left J-spectral factorization of W with respect Then W (λ) = L+ (λ to the unit circle. The reverse implication is trivial. In what follows we assume that W is a biproper rational m × m matrix function which is selfadjoint on the unit circle and has no pole there. Such a function can be represented in the form W (λ) = D0 + C(λIn − A)−1 B + B ∗ (λ−1 In − A∗ )−1 C ∗ , where A has all its eigenvalues in the open unit disc. The fact that W is proper implies that W is analytic at zero. We shall assume additionally that A is invertible. Note that the invertibility of A follows from the analyticity at zero whenever the realization C(λ − A)−1 B is minimal. The invertibility assumption on A allows us to write W (λ) = D0 − B ∗ A−∗ C ∗ + C(λ − A)−1 B − B ∗ A−∗ (λ − A−∗ )−1 A−∗ C ∗ . Since W (∞) = D0 − B ∗ A−∗ C ∗ = W (0)∗ one has D0 − B ∗ A−∗ C ∗ = (D0 − CA−1 B)∗ .

264

Chapter 14. J-spectral factorization

Hence D0 is selfadjoint. We shall assume additionally that D0 = J0 for some signature matrix J0 . Thus W is of the form W (λ) = J0 − B ∗ A−∗ C ∗ + C(λ − A)−1 B − B ∗ A−∗ (λ − A−∗ )−1 A−∗ C ∗ . (14.30) We shall prove the following factorization result. Theorem 14.11. Let W be a biproper rational m × m matrix function given by (14.30), where J0 is a signature matrix and A is an invertible n × n matrix having all its eigenvalues in the open unit disc. In order that, for some signature matrix J the function W admits a left J-spectral factorization with respect to the unit circle, it is necessary and suﬃcient that there exists a Hermitian n × n matrix Y such that J0 + B ∗ Y B is invertible and Y is a solution of the equation Y = A∗ Y A − (C ∗ + A∗ Y B)(J0 + B ∗ Y B)−1 (C + B ∗ Y A)

(14.31)

with A − B(J0 + B ∗ Y B)−1 (C + B ∗ Y A) having all its eigenvalues in the open unit disc. In that case Y is unique and for J one can take any signature matrix J determined by (14.32) J0 + B ∗ Y B = E ∗ JE, where E is some invertible matrix. Furthermore, if Y is a Hermitian matrix with the properties mentioned above, then for a signature matrix J determined by the ¯ −1 )∗ JL− (λ) of W expression (14.32), a left J-spectral factorization W (λ) = L− (λ with respect to the unit circle is obtained by taking L− (λ) = E + E(J0 + B ∗ Y B)−1 (C + B ∗ Y A)(λIn − A)−1 B.

(14.33)

Equation (14.31) is a particular case of the so-called discrete algebraic Riccati equation. A solution Y of equation (14.31) is called T-stabilizing, or simply stabilizing when no confusion can arise, if J0 + B ∗ Y B is invertible and the matrix A − B(J0 + B ∗ Y B)−1 (C + B ∗ Y A) has all its eigenvalues in the open unit disc. In the above theorem, the existence of such a solution is required. Proof. We split the proof into six parts. Part 1. Since W is biproper and given by (14.30), we can write a realization for − A) −1 B, where D = W (∞) = J0 − B ∗ A−∗ C ∗ and W . In fact W (λ) = D + C(λ −∗ A 0 −A−∗ C ∗ = B ∗ A−∗ C . (14.34) , C A= , B= B 0 A Recall that the matrix A is invertible and has all its eigenvalues in the open unit disc D. Hence A−∗ has all its eigenvalues in Dext. This allows us to apply Theorem 12.5 with F− = D and F+ = Dext . It follows that W admits a left canonical factorization with respect to T if and only if the equation Y BD−1 B ∗ A−∗ Y − Y (A − BD−1 C) + (A−∗ + A−∗ C ∗ D−1 B ∗ A−∗ )Y + A−∗ C ∗ D−1 C = 0

(14.35)

14.5. J-spectral factorization with respect to other contours

265

has a unique solution Y satisfying the following additional spectral constraints: ⊂ Dext , σ A−∗ + (A−∗ C ∗ + Y B)D−1 B ∗ A−∗ (14.36) (14.37) σ A − BD−1 (B ∗ A−∗ Y + C) ⊂ D. Furthermore, if Y is such a solution of (14.35), then a left canonical factorization W (λ) = W1 (λ)W2 (λ) of W with respect to T is obtained by taking W1 (λ) = D − B ∗ A−∗ (λ − A−∗ )−1 (A−∗ C ∗ + Y B),

(14.38)

W2 (λ) = I + D−1 (B ∗ A−∗ Y + C)(λ − A)−1 B.

(14.39)

Let Y be the solution of (14.35) satisfying (14.36) and (14.37). We claim that J0 + B ∗ Y B is invertible. To prove this it will be convenient to rewrite W1 as a function of λ−1 . This can be done as follows: W1 (λ)

= D − B ∗ (λA∗ − I)−1 (A−∗ C ∗ + Y B) = D + B ∗ λ−1 (λ−1 − A∗ )−1 (A−∗ C ∗ + Y B) = D + B ∗ (λ−1 − A∗ + A∗ )(λ−1 − A∗ )−1 (A−∗ C ∗ + Y B) = D + B ∗ A−∗ C ∗ + B ∗ Y B + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B).

Recall that D = J0 − B ∗ A−∗ C ∗ . Thus W1 (λ) = J0 + B ∗ Y B + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B).

(14.40)

Since A is invertible, both W and W2 are analytic at zero. From the above formula for W1 we see that W1 is also analytic at zero. Hence W (0) = W1 (0)W2 (0). But W (0) is invertible. Thus W1 (0) = J0 + B ∗ Y B is invertible too. Part 2. In this part Y stands for a solution of (14.35) such that J0 + B ∗ Y B is invertible. We prove that in this case Y is also a solution of (14.31). Furthermore, we show that D −1 (C + B ∗ A−∗ Y ) = (J0 + B ∗ Y B)−1 (C + B ∗ Y A).

(14.41)

Multiplying (14.35) on the left by A∗ and regrouping terms one obtains A∗ Y A − Y − (A∗ Y B + C ∗ )D−1 (C + B ∗ A−∗ Y ) = 0.

(14.42)

So Y = A∗ Y A − (A∗ Y B + C ∗ )D −1 (C + B ∗ A−∗ Y ). Multiplying the latter identity on the left with B ∗ A−∗ and adding C to both sides gives C + B ∗ A−∗ Y = C + B ∗ Y A − (B ∗ Y B + B ∗ A−∗ C ∗ )D −1 (C + B ∗ A−∗ Y ).

266

Chapter 14. J-spectral factorization

It follows that

=

I + (B ∗ Y B + B ∗ A−∗ C ∗ )D −1 (C + B ∗ A−∗ Y ) D + B ∗ A−∗ C ∗ + B ∗ Y B D−1 (C + B ∗ A−∗ Y )

=

(J0 + B ∗ Y B)D −1 (C + B ∗ A−∗ Y ).

C + B∗Y A =

Since J0 + B ∗ Y B is invertible, we see that (14.41) holds. Using (14.41) in (14.42) gives that Y is a solution of (14.31). Part 3. In this part we show that Y ∗ is a solution of (14.35) whenever so is Y . For this purpose we consider the Hamiltonian T of (14.35), that is, ⎡ ⎤ −A−∗ − A−∗ C ∗ D−1 B ∗ A−∗ −A−∗ C ∗ D−1 C ⎦. T =⎣ −1 ∗ −∗ −1 BD B A −(A − BD C) − BD −1 C), where A, B and C are given by (14.34). Put Note that T = −(A ) ( 0 I . H= −I 0 !∗ and H = −H ∗ . Next we carry out the following =A −∗ H, H B =A −∗ C Then H A computation: −A−∗ C ∗ 0 A∗ ∗ −∗ −1 C D − CA B = D − B A B 0 A−1 ∗ −A−∗C ∗ −1 CA = D− B B =

D + B ∗ A−∗ C ∗ − CA−1 B = J0 − CA−1 B = D ∗ .

A −1 B = D∗ and we can apply item (iii) in Proposition 9.2 to show Thus D − C that T is invertible and HT = T ∗ H. Taking adjoints in (14.35) we obtain the equation (14.43) Y ∗ A−1 BD −∗ B ∗ Y ∗ + Y ∗ (A−1 + A−1 BD−∗ CA−1 ) ∗ ∗ −∗ ∗ ∗ ∗ −∗ −1 −(A − C D B )Y + C D CA = 0, where Y ∗ is the unknown. The Hamiltonian T∗ of this equation is given by ⎤ ⎡ ∗ A − C ∗ D−∗ B ∗ −C ∗ D−∗ CA−1 ⎦. T∗ = ⎣ −1 −∗ ∗ −1 −1 −∗ −1 A BD B A + A BD CA

14.5. J-spectral factorization with respect to other contours

267

It follows that T∗ = HT ∗ H. This together with the result of the previous paragraph shows that T∗ = T −1 . Now let Y be a solution of (14.35). It follows that Y ∗ is a solution of (14.43). Using the general theory of Riccati equations (see Section 12.1), this implies that the space ( ∗ ) Y N∗ = Im I is invariant under T∗ . But T∗ = T −1 . Thus the ﬁnite dimensional space N∗ is invariant under the Hamiltonian T of (14.35). But then (again see Section 12.1) we may conclude that Y ∗ is a solution of (14.35) too. Part 4. Let Y be a solution of (14.35) satisfying the additional spectral constraints (14.36) and (14.37). In this part we show that Y must be Hermitian. Now Y is uniquely determined by the given properties. Since, by the result of the previous part of the proof, Y ∗ a solution of (14.35), it thus suﬃces to show that the conditions (14.36) and (14.37) hold with Y ∗ in place of the matrix Y . From the ﬁrst part of the proof we know that J0 + B ∗ Y B is invertible. Hence the identity (14.41) holds. Using this identity, we can rewrite (14.37) as σ A − B(J0 + B ∗ Y B)−1 (C + B ∗ Y A) ⊂ D. Taking adjoints, we arrive at σ (A∗ − (A∗ Y ∗ B + C ∗ )(J0 + B ∗ Y B)−1 B ∗ ⊂ D. Next, note that ∗ −1 A − (A∗ Y ∗ B + C ∗ )(J0 + B ∗ Y B)−1 B ∗ −1 −∗ = I − (Y ∗ B + A−∗ C ∗ )(J0 + B ∗ Y B)−1 B ∗ A = I + (Y ∗ B + A−∗ C ∗ ) J0 + B ∗ Y B −1 ∗ −∗ B A −B ∗ (Y ∗ B + A−∗ C ∗ ) = A−∗ + (Y ∗ B + A−∗ C ∗ )D −1 B ∗ A−∗ . Here we used that D = J0 − B ∗ A−∗ C ∗ . We conclude that σ(A−∗ + (Y ∗ B + A−∗ C ∗ )D −1 B ∗ A−∗ ) ⊂ Dext, which is (14.36) with Y ∗ in place of Y . In Part 3 of the proof we saw that Y ∗ is a solution of (14.35). Furthermore, J0 + B ∗ Y ∗ B = (J0 + B ∗ Y ∗ B)∗ is invertible. Thus we know that (14.41) holds with Y ∗ in place of Y , that is,

268

Chapter 14. J-spectral factorization D −1 (C + B ∗ A−∗ Y ∗ ) = (J0 + B ∗ Y ∗ B)−1 (C + B ∗ Y ∗ A).

(14.44)

Using this we show that (14.37) holds with Y ∗ in place of Y . Indeed, taking adjoints in (14.36) we get σ(A−1 + A−1 BD−∗ (B ∗ Y ∗ + CA−1 ) ⊂ Dext. Now

A−1 + A−1 BD−∗ (B ∗ Y ∗ + CA−1

−1

−1 = I + BD−∗ (B ∗ Y ∗ + CA−1 ) A −1 A = I − B(D ∗ + B ∗ Y ∗ B + CA−1 B)−1 (B ∗ Y ∗ + CA−1 ) = A − B(J0 + B ∗ Y ∗ B)−1 (B ∗ Y ∗ A + C). Here we used that D∗ = J0 − CA−1 B. Now apply the identity (14.44). It follows that σ(A − BD −1 (C + B ∗ A−∗ Y ∗ )) ⊂ D, which is (14.37) with Y ∗ in place of Y . Part 5. Let Y be a Hermitian matrix such that J0 + B ∗ Y B is invertible and Y is a stabilizing solution of (14.31). In this part we show that in that case Y is a solution of (14.35) and that Y satisﬁes the spectral constraints (14.36) and (14.37). As a ﬁrst step let us prove that under the above conditions on Y again (14.41) holds. Indeed, multiplying (14.31) from the left by B ∗ A−∗ and adding C to both sides we get C + B ∗ A−∗ Y

=

=

C + B ∗ Y A − (B ∗ A−∗ C ∗ + B ∗ Y B) ·(J0 + B ∗ Y B)−1 (C + B ∗ Y A) (J0 + B ∗ Y B) − (B ∗ A−∗ C ∗ + B ∗ Y B) ·(J0 + B ∗ Y B)−1 (C + B ∗ Y A)

=

(J0 − B ∗ Y B)(J0 + B ∗ Y B)−1 (C + B ∗ Y A)

=

D(J0 + B ∗ Y B)−1 (C + B ∗ Y A).

Hence (14.41) holds indeed. Using this we can rewrite (14.31) as A∗ Y A − Y − (A∗ Y B + C ∗ )D−1 (C + B ∗ A−∗ Y ) = 0. Multiplying the latter on the left by A−∗ and regrouping terms we see that Y satisﬁes (14.35). Since Y is a stabilizing solution of (14.31) and (14.41) holds, the spectral constraint (14.37) is satisﬁed too. It remains to prove (14.36) To do this we ﬁrst

14.5. J-spectral factorization with respect to other contours

269

note that

A−1 + A−1 BD −∗ (B ∗ Y + CA−1

−1

−1 = I + BD−∗ (B ∗ Y + CA−1 ) A −1 A = I − B(D∗ + B ∗ Y ∗ B + CA−1 B)−1 (B ∗ Y ∗ + CA−1 ) = A − B(J0 + B ∗ Y ∗ B)−1 (B ∗ Y ∗ A + C) = A − BD−1 (C + B ∗ A−∗ Y ). Thus, since Y is Hermitian, we see that (14.36) follows from (14.37) by taking adjoints and an inverse. Because of the uniqueness of the solution Y in the ﬁrst part of the proof, the result of the present part also shows that the Hermitian stabilizing solution of (14.31), if it exists, is unique Part 6. In this ﬁnal part we complete the argument. Assume that for some J the function W admits a left J-spectral factorization with respect to the unit circle. Then by the ﬁrst part of the proof, equation (14.35) has a solution Y satisfying (14.36) and (14.37). Moreover for this Y we have that J0 + B ∗ Y B is invertible. Part 4 of the proof tells us that Y is Hermitian. From Part 2 we know that Y is a solution of (14.31) which, according to (14.37) and (14.41), is stabilizing. Conversely, if Y is a Hermitian matrix such that J0 + B ∗ Y B is invertible and Y is a stabilizing solution of (14.31), then Y is a solution of (14.35) and Y satisﬁes (14.36) and (14.37). Hence W admits a left canonical factorization with respect to the unit circle, and thus, by Proposition 14.10, also a left J-spectral factorization with respect to the unit circle. Finally, take a signature matrix J such that (14.32) holds. It remains to establish the formula for the left spectral factor L− . To do this we use the left canonical factorization W (λ) = W1 (λ)W2 (λ) obtained in Part 1. Combining (14.39) and (14.41) we get W2 (λ) = I + (J0 + B ∗ Y B)−1 (C + B ∗ Y A)(λ − A)−1 B. Thus, using the expression (14.40) for W1 (λ), ¯ −1 )∗ W2 (λ

=

I + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B)(J0 + B ∗ Y B)−1 J0 + B ∗ Y B + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B) (J0 + B ∗ Y B)−1

=

W1 (λ)(J0 + B ∗ Y B)−1 ,

=

¯ −1 )∗ (J0 +B ∗ Y B)W2 (λ). Now let J be a signature and it follows that W (λ) = W2 (λ ¯ −1 )∗ JL− (λ), with matrix such that (14.32) holds. Then we see that W (λ) = L− (λ L− given by (14.33), is a left J-spectral factorization with respect to the unit circle.

270

Chapter 14. J-spectral factorization

We now turn to a situation arising from linear-quadratic optimal control theory. It concerns the following version of the discrete algebraic Riccati equation X = A∗ XA + Q − A∗ XB(R + B ∗ XB)−1 B ∗ XA.

(14.45)

Here A, B, Q and R are given matrices of sizes n × n, n × m, n × n and m × m, respectively. We will consider the case when A has all its eigenvalues in the open unit circle, R and Q are Hermitian, and R is invertible. Of special interest are the stabilizing solutions of (14.45). A solution X of (14.45) is said to be T-stabilizing, or simply stabilizing when there is no danger of confusion, if R+B ∗ XB is invertible and A − B(R + B ∗ XB)−1 B ∗ XA has all its eigenvalues in the open unit disc. In connection with (14.45) we consider the rational matrix function W (λ) = R + B ∗ (λ−1 In − A∗ )−1 Q(λIn − A)−1 B.

(14.46)

Note that this function is Hermitian on the unit circle. Proposition 14.12. Let A, B, Q and R be as above, so A is an n × n matrix having its eigenvalues in the open unit disc, B is an m × m matrix, R is an invertible Hermitian m × m matrix, and Q is a Hermitian n × n matrix. Assume in addition that A is invertible. The following two statements are equivalent: (i) The Riccati equation (14.45) has a (unique) Hermitian T-stabilizing solution; (ii) For some Hermitian matrix J, the rational matrix function (14.46) admits a left J-spectral factorization with respect to the unit circle. In that case J is congruent to R+B ∗ XB. Also, if X is the Hermitian T-stabilizing solution of (14.45), then ¯ −1 )∗ (R + B ∗ XB)L− (λ), W (λ) = L− (λ with

L− (λ) = Im + (R + B ∗ XB)−1 B ∗ XA(λIn − A)−1 B,

is a left (R + B ∗ XB)-spectral factorization with respect to the unit disc. The function L− is the unique left (R + B ∗ XB)-spectral factor with L− (∞) = Im . The additional assumption that A is invertible plays an essential role in the proof as we give it below. Indeed, the argument involves a reduction to earlier results, in particular to Theorem 14.11. However, instead of Theorem 14.11 one can employ Theorem 14.15 below which does not feature the hypothesis that A is invertible. Before we prove the proposition, let us remark that in the case of the linear quadratic optimal control problem of mathematical systems theory, one has that R is positive deﬁnite and Q is positive semideﬁnite. Hence the function (14.46) is positive deﬁnite on the unit circle, and as A has is eigenvalues in the open unit disc, it has no poles on the unit circle. Thus, in that case, the function does admit

14.5. J-spectral factorization with respect to other contours

271

a right spectral factorization with J = I, and hence there is a stabilizing solution X to the discrete algebraic Riccati equation. In addition, for that solution the matrix R + B ∗ XB is positive deﬁnite. Proof. We shall deduce Proposition 14.12 from Theorem 14.11. First, a realization for (14.46) is given as $ A W (λ) = R + −B ∗ A−∗ Q B ∗ A−∗ λ − −A−∗ Q

0

%−1 B

A−∗

0

.

Since A is has all its eigenvalues in the open unit disc, there is a unique solution to the equation X0 − A∗ X0 A = Q. (14.47) Taking as a similarity transformation the matrix I 0 , X0 I and using Q − X0 = −A∗ X0 A, the realization above can be rewritten as: W (λ)

= R + −B ∗ A−∗ (Q − X0 ) B ∗ A

−∗

$ λ−

A

0

0

A−∗

%−1

B

X0 B

= R + B ∗ X0 A(λ − A)−1 + B ∗ A−∗ (λ − A−∗ )−1 X0 B. The latter expression is of the form (14.30), with C = B ∗ X0 A and with J0 = R + B ∗ X0 B. So, we can apply Theorem 14.11, with (14.31) suitably modiﬁed, to conclude that W admits a left J-spectral factorization if and only if there is a solution Y , satisfying additional constraints, of the equation Y = A∗ Y A − (A∗ X0 B + A∗ Y B)(R + B ∗ X0 B + B ∗ Y B)−1 (B ∗ Y A + B ∗ X0 A). Putting X = X0 + Y and taking into account (14.47), we see that the above equation becomes (14.45) for X. The additional constraints referred to above are: in the ﬁrst place, invertibility of R + B ∗ X0 B + B ∗ Y B = R + B ∗ XB, which we also required for the solution of (14.45), and, secondly, the condition that the eigenvalues of A − B(R + B ∗ (X0 + Y )B)−1 B ∗ (X0 + Y )A = A − B(R + B ∗ XB)−1 B ∗ XA are in the open unit disc. But this is exactly what is required for the stabilizing solution of the equation (14.45). The expressions for the factorization also follow directly from the formulas in Theorem 14.11.

272

Chapter 14. J-spectral factorization

We conclude this section by considering J-spectral factorization of a selfadjoint function on the real line. As before J is an invertible Hermitian m × m matrix. Suppose W is a rational m × m matrix function. A factorization ¯ ∗ JL(λ) W (λ) = L(λ)

(14.48)

is called a right J-spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed upper half plane ¯ ∗ and its inverse are analytic on (inﬁnity included). In that case the function L(λ) the closed lower half plane (inﬁnity included). Thus a right J-spectral factorization with respect to the real line is a right canonical factorization with respect to R featuring an additional symmetry property between the factors. A factorization (14.48) is called a left J-spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed lower ¯ ∗ and its inverse are half plane (inﬁnity included), in which case the function L(λ) analytic on the closed upper half plane (inﬁnity included). Such a factorization is a left canonical factorization with respect to R. Results for this type of factorization can be derived in a straightforward manner from J-spectral factorization theorems with respect to the imaginary axis. Indeed, if W is selfadjoint on the real line, then V given by V (λ) = W (−iλ) is self¯ adjoint on the imaginary axis. Also W (λ) = L+ (λ)JL + (λ) is a right J-spectral fac¯ torization of W with respect to the real line if and only V (λ) = K+(−λ)JK + (λ), with K+ (λ) = L+ (−iλ), is a right J-spectral factorization of V with respect to the imaginary axis. As an illustration we show how one can derive the following result as a corollary from Theorem 14.9. Theorem 14.13. Let the rational m × m matrix function W be given by W (λ) = J + C(λIn − A)−1 B + B ∗ (λIn − A∗ )−1 C ∗ , where J is an m × m signature matrix and A is an n × n matrix having all its eigenvalues in the open upper half plane. Then W admits a right J-spectral factorization with respect to the real line, ¯ ∗ JL+ (λ), W (λ) = L+ (λ) if and only if the algebraic Riccati equation Y C ∗ JCY − Y (A∗ − C ∗ JB ∗ ) + (A − BJC)Y − BJB ∗ = 0

(14.49)

has a (unique) skew-Hermitian solution Y such that A∗ − C ∗ JB ∗ − C ∗ JCY has all its eigenvalues in the open lower half plane. In that case, the unique right J-spectral factor L+ for which L+ (∞) = Im is given by L+ (λ) = Im + J(CY + B ∗ )(λIn − A∗ )−1 C ∗ .

14.6. Left versus right J-spectral factorization

273

A solution Y of the Riccati equation (14.49) is called R-stabilizing, or simply stabilizing when confusion is not possible, if A∗ − C ∗ JB ∗ − C ∗ JCY has all its eigenvalues in the open lower half plane. In the above theorem, the existence of such a solution is required. Proof. Write V (λ) = W (−iλ). Then V (λ)

= =

J + C(−iλ − A)−1 B + B ∗ (−iλ − A∗ )−1 C −1 −1 J + (iC) λ − (iA) B + B ∗ λ + (iA)∗ (iC).

Notice that iA has all its eigenvalues in the open left half plane. By Theorem 14.9 the function V admits a right J-spectral factorization with respect to the imaginary axis if and only if the equation (14.50) X(iC)∗ J(iC)X + X (iA)∗ − (iC)∗ JB ∗ ∗ + iA − BJ(iC) X + BJB = 0 has a Hermitian solution X such that the matrix (iA)∗ −(iC)∗ JB ∗ +(iC)∗ J(iC)X has all its eigenvalues in the open left half plane. In that case, a right J-spectral ¯ ∗ JK+ (λ) of V with respect to the imaginary axis is factorization V (λ) = K+ (−λ) −1 (iC)∗ . Next we replace obtained by taking K+ (λ) = I +J(iCX −B ∗ ) λ+(iA)∗ X by iY and multiply equation (14.50) by −1. In this way (14.50) is shown to be equivalent to (14.49). Furthermore Y isskew-Hermitian if and only if X is Hermitian, and A∗ − C ∗ JB ∗ − C ∗ JCY = i (iA)∗ − (iC)∗ JB ∗ + (iC)∗ J(iC)X . Finally, put L+ (λ) = K+ (iλ). Then L+ (λ)

−1 = I + J(iCX − B ∗ ) λ + (iA)∗ (iC)∗ = I + J(−CY − B ∗ )(iλ − iA∗ )−1 (−i)C ∗ = I + J(CY + B ∗ )(λ − A∗ )−1 C ∗ .

Using these formulas it is now straightforward to complete the argument.

14.6 Left versus right J-spectral factorization The existence of a left canonical factorization does not always imply the existence of a right canonical factorization. The same is true for J-spectral factorization. In this section we answer the following question: if a rational matrix function W admits a left J-spectral factorization, under what conditions does it also have a right J-spectral factorization? And, if so, how can the right factorization be obtained from the left one? The main result can be viewed as a symmetric version of Theorem 12.6. We restrict our attention to factorization with respect to the imaginary axis.

274

Chapter 14. J-spectral factorization

For later purposes it will be convenient to only assume that J is an invertible Hermitian matrix. We do not stipulate it to be a signature matrix here. Theorem 14.14. Let J be an invertible Hermitian m × m matrix, and let W be a rational m × m matrix function. Suppose ¯ ∗ JL− (λ) W (λ) = L− (−λ) is a left J-spectral factorization with respect to the imaginary axis, and L− admits the realization L− (λ) = Im + C(λIn − A)−1 B (14.51) with A and A× = A − BC having their eigenvalues in the open left half plane. Let Q and P be the unique (Hermitian) solutions of the Lyapunov equations QA + A∗ Q = A× P + P (A× )∗

=

C ∗ JC.

(14.52)

−BJ −1 B ∗ .

(14.53)

Then W admits a right J-spectral factorization with respect to the imaginary axis if and only if I − QP is invertible, or, which amounts to the same, I − P Q is invertible. In that case, a right J-spectral factorization of W with respect to the imaginary axis is given by ¯ ∗ JL+ (λ), W (λ) = L+ (−λ)

(14.54)

where L+ (λ) and its inverse are given by (14.55) L+ (λ) = Im + (CP − J −1 B ∗ )(I − QP )−1 ∗ −1 ∗ ·(λIn + A ) (C J − QB), −1 −1 ∗ B ) λIn + (A× )∗ L−1 + (λ) = Im − (CP − J ·(I − QP )

−1

(14.56) ∗

(C J − QB).

Proof. We bring ourselves in the situation of Section 12.4 by introducing ¯ ∗ = Im − B ∗ (λIn + A∗ )−1 C ∗ , Y+ (λ) = L− (−λ) Y− (λ) = JL− (λ) = J + JC(λIn − A)−1 B. Then W (λ) = Y+ (λ)Y− (λ) is a left canonical factorization, here taken with respect to the imaginary axis (cf., the remark made after the proof of Theorem 12.6). In terms of the notation employed in Section 12.4, Y+ (λ)

= D+ + C+ (λ − A+ )−1 B+ ,

Y− (λ)

= D− + C− (λ − A− )−1 B− ,

14.6. Left versus right J-spectral factorization with D+ = Im , D− = J,

A+ = −A∗ , A− = A,

275

B+ = C ∗ , B− = B,

C+ = −B ∗ , C− = JC.

× × ∗ × For the associate main matrices we have A× + = −(A ) and A− = A . Thus the Lyapunov equations (12.19) reduce to the equations (14.53) and (14.52). Application of Theorem 12.6 now shows that W admits a right canonical factorization with respect to the imaginary axis if and only if I − QP is invertible, or, which amounts to the same, I − P Q is invertible. Assume this is the case. Then, again by virtue of Theorem 12.6, we have the right canonical factorization W (λ) = W− (λ)W+ (λ), where

W− (λ)

=

D+ + (D+ C− + C+ Q)(λIX− − A− )−1 −1 · (IX− − P Q)−1 (B− D− − P B+ ),

W+ (λ)

=

−1 D− + (D+ C+ + C− P )(IX+ − QP )−1

· (λIX+ − A+ )−1 (B+ D− − QB− ). Making the appropriate substitutions, we get W− (λ)

=

I + (JC − B ∗ Q)(λ − A)−1 (I − P Q)−1 (BJ −1 − P C ∗ ),

W+ (λ)

=

J + (JCP − B ∗ )(I − QP )−1(λ + A∗ )−1 (C ∗ J − QB).

Put L+ (λ) = J −1 W+ (λ). Then L+(λ) is given by (14.55). Taking into account the ¯ ∗ is precisely W− (λ). It follows selfadjointness of Q and P , one sees that L+ (−λ) ∗ ¯ that W (λ) = L+ (−λ) JL+(λ), and this factorization is a right J-spectral factor−1 ization of W with respect to the imaginary axis. Finally, L−1 + (λ) = W+ (λ)J, and according to Theorem 12.6, W+−1 (λ)

=

−1 −1 −1 −1 D− − D− (D+ C+ + C− P )(λIX+ − A× +) −1 · (IX+ − QP )−1 (B+ D− − QB− )D− .

Via the appropriate substitutions this becomes −1 W+−1 (λ) = J −1 − (CP − J −1 B ∗ ) λ + (A× )∗ (I − QP )−1 (C ∗ J − QB)J −1 . Multiplying the latter identity from the right by J gives (14.56).

For the case when J is a signature matrix (that is, J = J ∗ = J −1 ) it is also possible to derive the previous result from Theorem 14.9. Indeed, let Q be the solution of (14.52), and introduce ) ( I 0 . T = Q I

276

Chapter 14. J-spectral factorization

Then one has (via the product rule for realizations) ¯ ∗ JL− (λ) W (λ) = L− (−λ) =J+

JC

−B ∗

$ T

λ−T

−1

A

0

C ∗ JC

−A∗

%−1 T

T

−1

B

C∗J

= J + (JC − B ∗ Q)(λ − A)−1 B − B ∗ (λ + A∗ )−1 (C ∗ J − QB). Clearly, one can now apply Theorem 14.9. The stabilizing solution of equation (14.28), taken for this particular situation, and the solution P of (14.53) are related as follows: if Y is the stabilizing solution, then I +QY is invertible, the matrix P = Y (I + QY )−1 solves (14.53), and I − QP = (I + QY )−1 is invertible. Conversely, if P is the solution of (14.53) and I − QP is invertible, then Y = P (I − QP )−1 is Hermitian and it is the desired stabilizing solution. Finally, for the case where J = I, and so W is positive deﬁnite on the imaginary line, the condition that I − QP is invertible should be automatically fulﬁlled on account of Theorem 9.4. That this is indeed the case can be seen as follows. First recall that A has all its eigenvalues in the open left half plane. This implies that P is positive semideﬁnite and Q is negative semideﬁnite. Since J = I we get from (14.53) that Ker P is invariant under A∗ . Now write P , Q, A and C ˙ Im P as with respect to the decomposition Cn = Ker P + Q11 Q12 A11 0 0 0 , Q= , A= , C = C1 C2 . P = Q21 Q22 A21 A22 0 P22 Then Q22 is negative semideﬁnite and P22 is positive deﬁnite. Finally, I − QP is invertible if and only if I − Q22 P22 is invertible as a map from Im P to itself. Since 1/2 1/2 I − Q22 P22 is similar to I − P22 Q22 P22 , and the latter is positive deﬁnite, we see that invertibility of I − QP is indeed automatically satisﬁed.

14.7 J-spectral factorization relative to the unit circle revisited In this section we present a somewhat more general form of Theorem 14.11, using an alternative approach. As in the ﬁrst part of Section 14.5, the function W is a rational m × m matrix function which is selfadjoint on the unit circle and has no pole there. Such a function can be represented in the form W (λ) = D0 + C(λIn − A)−1 B + B ∗ (λ−1 In − A∗ )−1 C ∗ ,

(14.57)

where D0 is a Hermitian m × m matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. In contrast to the situation considered in

14.7. J-spectral factorization relative to the unit circle revisited

277

Section 14.5 we do not assume that A is invertible, and hence the representation (14.30) is not available in the present context. Similar to what was done in Theorem 14.11, we associate with the representation (14.57) the Riccati equation Y = A∗ Y A − (C ∗ + A∗ Y B)(D0 + B ∗ Y B)−1 (C + B ∗ Y A).

(14.58)

Recall from the paragraph directly following Theorem 14.11 that a solution Y to this Riccati equation is called T-stabilizing (or simply stabilizing) if D0 + B ∗ Y B is invertible and the matrix A − B(D0 + B ∗ Y B)−1 (C + B ∗ Y A)

(14.59)

has all its eigenvalues in the open unit disc. The following theorem is the main result of this section. Theorem 14.15. Let W be a rational m×m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n×n matrix having all its eigenvalues in the open unit disc. In order that, for some signature matrix J the function W admits a left J-spectral factorization with respect to the unit circle, it is necessary and suﬃcient that the Riccati equation (14.58) has a Hermitian T-stabilizing solution Y . In that case Y is unique, and for J one can take any signature matrix J determined by D0 + B ∗ Y B = E ∗ JE, (14.60) where E is some invertible matrix. Furthermore, if Y is the Hermitian T-stabilizing solution to (14.58), then for a signature matrix J determined by (14.60), a left J¯ −1 )∗ JL− (λ) of W with respect to the unit spectral factorization W (λ) = L− (λ circle can be obtained by taking L− (λ) = E + E(D0 + B ∗ Y B)−1 (C + B ∗ Y A)(λIn − A)−1 B.

(14.61)

To prove the above theorem we cannot use the method employed in Section 14.5. Instead we shall use the connection between canonical factorization and invertibility of Toeplitz operators described in Section 1.2. For this purpose we need the block Toeplitz operator T on m 2 deﬁned by the rational m × m mam trix function W (λ−1 ). Recall (see Section 1.2) that m 2 = 2 (C ) stands for the Hilbert space of all square summable sequences (x0 , x1 , x2 , . . .) with entries in Cm . Furthermore, by deﬁnition, T is the operator on m 2 given by the block matrix representation ⎡ ⎤ R0 R−1 R−2 · · · ⎢ ⎥ ⎢R1 R0 R−1 · · ·⎥ ⎢ ⎥ (14.62) T =⎢ ⎥, ⎢R2 R1 R0 · · ·⎥ ⎣ ⎦ .. .. .. .. . . . .

278

Chapter 14. J-spectral factorization

where . . . , R−1 , R0 , R1 , . . . are the coeﬃcients in the Laurent expansion W (λ−1 ) =

∞

λj Rj

j=−∞

of the function W (λ−1 ) on the unit circle. When W is given by (14.57), we have R0 = D0 ,

∗ Rj = R−j = CAj−1 B,

j = 1, 2, . . . .

(14.63)

The following lemma provides one of the main steps in the proof of Theorem 14.15. As always in this section, J stands for a signature matrix. Lemma 14.16. Let W be a rational m × m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. Assume W admits a left J-spectral factorization with respect to the unit circle. Then the block Toeplitz operator T on m 2 deﬁned by the rational m × m matrix function W (λ−1 ) is invertible, and the n × n matrix Y given by ⎡ ⎤ C ⎢ ⎥ ⎢ CA ⎥ ∗ ⎢ ⎥ A∗ C ∗ A∗ 2 C ∗ · · · T −1 ⎢ Y =− C (14.64) ⎥ ⎢CA2 ⎥ ⎣ ⎦ .. . is a Hermitian stabilizing solution to the Riccati equation (14.58). Proof. A left J-spectral factorization with respect to the unit circle is, in particular, a left canonical factorization with respect to the unit circle. But then the function W (λ−1 ) admits a right canonical factorization with respect to the unit circle, and Theorem 1.2 tells us that the block Toeplitz operator T is invertible. This, together with the fact that A has all its eigenvalues in the open unit disc, gives that the matrix Y is well-deﬁned by (14.64). Note that T is selfadjoint because W (λ−1 ) has Hermitian values on the unit circle. But then T −1 is selfadjoint too, and (14.64) shows that Y is Hermitian m m Note that m 2 can be identiﬁed with the Hilbert space direct sum C ⊕ 2 . Via this identiﬁcation the operator T partitions as ⎡ ⎤ R1 ⎢ ⎥ ⎢ R2 ⎥ R0 Λ∗ ⎢ ⎥ , where Λ = ⎢ ⎥ : Cm → m T = (14.65) 2 . ⎢ R3 ⎥ Λ T ⎣ ⎦ .. . Put Δ = R0 − Λ∗ T −1 Λ. Since the 2 × 2 operator matrix in (14.65) and the operator in its right lower corner are both invertible, a standard Schur complement

14.7. J-spectral factorization relative to the unit circle revisited

279

argument (see [19] or the second proof of Theorem 2.1 in [20]) tells us that Δ is invertible as well. Furthermore, relative to the Hilbert space direct sum decomposition Cm ⊕ m 2 the inverse of T admits the block matrix representation ⎡ ⎤ −Δ−1 Λ∗ T −1 Δ−1 ⎦. (14.66) T −1 = ⎣ −1 −1 −1 −1 −1 ∗ −1 −T ΛΔ T + T ΛΔ Λ T Recall from (14.63) that R0 = D0 . Combining the second part of (14.63) with (14.64) we obtain that B ∗ Y B = −Λ∗ T −1Λ. It follows that D0 + B ∗ Y B = Δ, and hence D0 + B ∗ Y B is invertible, as desired. To prove that Y satisﬁes the Riccati equation (14.58) we ﬁrst consider the operator T −1 − ST −1S ∗ , where S is the (block) forward shift on m 2 . Thus the actions of S and S ∗ on m are given by 2 S ∗ (x0 , x1 , x2 , . . .) = (x1 , x2 , x3 , . . .).

S(x0 , x1 , x2 , . . .) = (0, x0 , x1 , . . .),

A straightforward computation shows that the partitioning of ST −1 S ∗ relative to the Hilbert space direct sum Cm ⊕ m 2 is given by 0 0 −1 ∗ ST S = . 0 T −1 This identity, together with the identity (14.66), yields −Δ−1 Λ∗ T −1 Δ−1 −1 −1 ∗ = T − ST S −T −1 ΛΔ−1 T −1ΛΔ−1 Λ∗ T −1 =

I −T

−1

Λ

Δ−1 I

−Λ∗ T −1 .

Next, let Γ be the operator from Cn to m 2 given by ⎡ ⎤ C ⎢ ⎥ ⎢ CA ⎥ ⎢ ⎥ Γ=⎢ ⎥. ⎢CA2 ⎥ ⎣ ⎦ .. .

(14.67)

(14.68)

Note that this operator Γ is well-deﬁned because the matrix A has all its eigenvalues in the open unit disc. As is easily checked ΓA = S ∗ Γ,

ΓB = Λ,

Y = −Γ∗ T −1Γ.

(14.69)

280

Chapter 14. J-spectral factorization

From these identities and (14.67) it follows that Y − A∗ Y A =

−Γ∗ T −1Γ + A∗ Γ∗ T −1 ΓA

=

−Γ∗ T −1Γ + Γ∗ ST −1 S ∗ Γ

=

−Γ∗ (T −1 − ST −1S ∗ )Γ I ∗ −Γ Δ−1 I −Λ∗ T −1 Γ. −1 −T Λ

= Furthermore

[ I − Λ∗ T −1 ]Γ = C − Λ∗ T −1 S ∗ Γ = C − B ∗ Γ∗ T −1 Γ = C + B ∗ Y A.

(14.70)

Summarizing (and using that Y is Hermitian) we have Y − A∗ Y A = −(C + B ∗ Y A)∗ Δ−1 (C + B ∗ Y A). Since Δ = D0 + B ∗ Y B, this identity shows that Y satisﬁes the Riccati equation (14.58). Write A× for the matrix (14.59). We need to show that for Y given by (14.64), all eigenvalues of A× are in the open unit disc. Using (14.67), the fact that S ∗ S is the identity operator on m 2 , and the identities in (14.69) and (14.70), we see that I Δ−1 I −Λ∗ T −1 Γ S ∗ T −1 Γ = T −1 S ∗ Γ + S ∗ −1 −T Λ = T −1 ΓA − T −1 ΛΔ−1 (C + B ∗ Y A) = T −1 Γ A − BΔ−1 (C + B ∗ Y A) = T −1ΓA× . Thus S ∗ T −1 Γ = T −1 ΓA× . It follows that (S ∗ )k T −1 Γ = T −1 Γ(A× )k ,

k = 1, 2, . . . .

But then the fact that S ∗n converges to zero in the strong operator topology yields lim T −1 Γ(A× )k x = lim (S ∗ )k T −1Γx = 0,

k→∞

k→∞

x ∈ Cn .

(14.71)

We shall use (14.71) to prove that A× has all its eigenvalues in the open unit disc. To do this we ﬁrst decompose Cn as Cn = X1 ⊕ X2 , where X2 = Ker Γ and X1 = (Ker Γ)⊥ . Notice that X2 is an invariant subspace for A, and C[X2 ] = {0}. We also have Y A[X2 ] = {0}. Indeed Y [AX2 ] ⊂ Y [X2 ] = −Γ∗ T −1 Γ[X2 ] = {0}.

14.7. J-spectral factorization relative to the unit circle revisited

281

Using C[X2 ] = {0} and Y A[X2 ] = {0} in (14.59), we see that A× |X2 = A|X2 , and X2 is an invariant subspace for A× too. In other words, A× admits a matrix representation of the form × 0 A11 × : X1 ⊕ X2 → X1 ⊕ X2 , A = (14.72) × A× 21 A22 where A× 22 = A|X2 : X2 → X2 . Since X2 is an invariant subspace for A and A has all its eigenvalues in the open unit disc, A22 has all its eigenvalues in the open unit disc too. Hence, in order to prove that A× has all its eigenvalues in the open unit disc, it now suﬃces to prove that A× 11 has this property. Let τ1 be the canonical embedding of X1 into Cn = X1 ⊕ X2 , and let Γ1 be the one-to-one operator from X1 into m 2 deﬁned by Γ1 = Γτ1 . Take x ∈ X1 . Since Γ is equal to k zero on X2 , we see from (14.72) that T −1 Γ(A× )k x = T −1 Γ1 (A× 11 ) x. But then × k −1 −1 (14.71) tells us that limk→∞ T Γ1 (A11 ) x = 0. Observe that T Γ1 is one-toone and has a closed (ﬁnite dimensional) range, that is, T −1 Γ1 is left invertible. × k k Hence limk→∞ T −1 Γ1 (A× 11 ) x = 0 implies that limk→∞ (A11 ) x = 0. Since x is an arbitrary element of X1 , the latter holds if and only if the eigenvalues of A× 11 are in the open unit disc. Lemma 14.16 proves the necessity part of Theorem 14.15. The suﬃciency part, the formula for the J-spectral factorization, and the uniqueness statement are covered by the next two lemmas. Lemma 14.17. Let W be a rational m × m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. Assume Y is a Hermitian stabilizing solution of the Riccati equation (14.58). Then W admits a left J-spectral factorization with respect to the unit circle. Such a factorization can be obtained as follows. Choose an m × m signature matrix J such that D0 + B ∗ Y B = E ∗ JE, where E is some invertible matrix, and deﬁne L− by (14.61), i.e., L− (λ) = E + E(D0 + B ∗ Y B)−1 (C + B ∗ Y A)(λIn − A)−1 B. ¯ −1 )∗ JL− (λ) is a left J-spectral factorization of W with respect Then W (λ) = L− (λ to the unit circle. Proof. Put Δ = D0 + B ∗ Y B, C0 = C + B ∗ Y A, and set Ψ(λ) = Δ + C0 (λ − A)−1 B.

(14.73)

Note that A − BΔ−1 C0 is equal to the matrix A× deﬁned by (14.59). Thus Ψ(λ)−1 = Δ−1 − Δ−1 C0 (λ − A× )−1 BΔ−1 .

(14.74)

The fact that A and A× have all their eigenvalues in the open unit disc implies that Ψ(λ) and Ψ(λ)−1 are both analytic on the closure of the exterior of the unit

282

Chapter 14. J-spectral factorization

disc, inﬁnity included. Since L− (λ) = EΔ−1 Ψ(λ), the same holds true for L− (λ) ¯ −1 )∗ JL− (λ) is a left spectral factorization with and L− (λ)−1 . It follows that L− (λ respect to the unit circle. It remains to show that ¯ −1 )∗ JL− (λ). W (λ) = L− (λ

(14.75)

From L− (λ) = EΔ−1 Ψ(λ) and Δ = D0 + B ∗ Y B = E ∗ JE we see that ¯ −1 )∗ JL− (λ) = Ψ(λ ¯ −1 )∗ Δ−1 Ψ(λ). L− (λ Using the deﬁnitions of Δ and C0 , the Riccati equation (14.58) can be rewritten as Y − A∗ Y A = −C0∗ Δ−1 C0 . It follows that λC0∗ Δ−1 C0 = −Y (λ − A) + (I − λA∗ )Y (λ − A) − λ(I − λA∗ )Y. Using this identity we obtain B ∗ (I − λA∗ )−1 (λC0∗ Δ−1 C0 )(λ − A)−1 B = −B ∗ (I − λA∗ )−1 Y B + B ∗ Y B − λB ∗ Y (λ − A)−1 B = −λB ∗ (I − λA∗ )−1 A∗ Y B − B ∗ Y B − B ∗ Y A(λ − A)−1 B. Hence ¯ −1 )∗ Δ−1 Ψ(λ) Ψ(λ = Δ + λB ∗ (I − λA∗ )−1 C0∗ Δ−1 Δ + C0 (λ − A)−1 B = Δ + λB ∗ (I − λA∗ )−1 C0∗ + C0 (λ − A)−1 B +B ∗ (I − λA∗ )−1 (λC0∗ Δ−1 C0 )(λ − A)−1 B. From the deﬁnitions of Δ and C0 given in the beginning of the proof we see that Δ − B ∗ Y B = D0 and C0 − B ∗ Y A = C. Thus the calculations above yield ¯ −1 )∗ Δ−1 Ψ(λ) = D0 + λC(I − λA)−1 + B ∗ (λ − A∗ )−1 C ∗ . Ψ(λ According to (14.57) the right-hand side in the previous identity is equal to W (λ). ¯ −1 )∗ Δ−1 Ψ(λ) = W (λ), as desired. Thus Ψ(λ Lemma 14.18. Let W be a rational m × m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. Assume Y is a Hermitian stabilizing solution of the Riccati equation (14.58). Then the block Toeplitz operator T on m 2 deﬁned by the rational m × m matrix function W (λ−1 ) is invertible and Y is uniquely determined by the expression (14.64).

14.7. J-spectral factorization relative to the unit circle revisited

283

Proof. As in the proof of the preceding lemma, we set Δ = D0 + B ∗ Y B and C0 = C + B ∗ Y A. Furthermore, Ψ(λ) is the rational m× m matrix function deﬁned by (14.73). Put Θ(λ) = Ψ(λ−1 ). The proof of the preceding lemma tells us that ¯ −1 )∗ Δ−1 Θ(λ). W (λ−1 ) = Θ(λ −1 Hence the block Toeplitz operator T on m ) admits the factor2 deﬁned by W (λ ∗ ization T = (TΘ ) Ξ TΘ , where TΘ is the block Toeplitz operator on m 2 deﬁned by Θ, and Ξ is the block diagonal operator on m 2 given by

Ξ = diag (Δ−1 , Δ−1 , Δ−1 , . . .). From (14.73), (14.74) and Θ(λ) = Ψ(λ−1 ) we know that Θ(λ)

=

Δ + λC0 (I − λA)−1 B,

(14.76)

Θ(λ)−1

=

Δ−1 − λΔ−1 C0 (I − λA× )−1 BΔ−1 ,

(14.77)

where A× is given by (14.59). From (14.76), (14.77), and the fact that both A and A× have all their eigenvalues in the open unit disc we see that TΘ is invertible and TΘ−1 is given by ⎤ ⎡ × Θ0 0 0 ··· ⎥ ⎢ × ⎢ Θ1 Θ × 0 · · ·⎥ 0 ⎥ ⎢ (14.78) TΘ−1 = ⎢ × ⎥, × ⎥ ⎢Θ2 Θ× Θ0 1 ⎦ ⎣ .. .. .. . . . × × −1 at zero. Furthermore, where Θ× 0 , Θ1 , Θ2 , . . . are the Taylor coeﬃcients of Θ(λ) (14.76) yields −1 Θ× , 0 = Δ

−1 Θ× C0 (A× )j−1 BΔ−1 , j = −Δ

j = 1, 2, . . . .

(14.79)

Let Γ be the operator from Cn into m 2 deﬁned by (14.68). Using the identities in (14.78) and (14.79) we compute that . (14.80) Γ∗ TΘ−1 = β A∗ β (A∗ )2 β · · · , with β given by β = C ∗ Δ−1 − A∗

∞ ∗ j ∗ −1 × j (A ) C Δ C0 (A ) BΔ−1 .

(14.81)

j=0

As T = (TΘ )∗ ΞTΘ and TΘ is invertible, we conclude that T is invertible. Moreover, using (14.80), we have Γ∗ T −1Γ = (Γ∗ TΘ−1 ) Ξ−1 (Γ∗ TΘ−1)∗ =

∞

j=0

β∗ Aj . (A∗ )j βΔ

(14.82)

284

Chapter 14. J-spectral factorization

We proceed by showing that β = (C ∗ + A∗ Y B)Δ−1 , where A× is given by (14.59). To prove this we use the fact that Y satisﬁes the Riccati equation (14.58). A straightforward computation gives Y

A∗ Y A − (C ∗ + A∗ Y B)Δ−1 (C + B ∗ Y A) = A∗ Y A − BΔ−1 (C + B ∗ Y A) − C ∗ Δ−1 (C + B ∗ Y A)

=

=

A∗ Y A× − C ∗ Δ−1 C0 .

We conclude that Y − A∗ Y A× = −C ∗ Δ−1 C0 . Since both A and A× have all their eigenvalues in the open unit disc, we obtain Y =−

∞

(A∗ )j C ∗ Δ−1 C0 (A× )j .

j=0

Using the latter identity in (14.81) we arrive at β = C ∗ Δ−1 + A∗ Y BΔ−1 = (C ∗ + A∗ Y B)Δ−1 . Finally, the identity β = (C ∗ + A∗ Y B)Δ−1 and the fact that Y satisﬁes the Riccati equation yield β∗ . Y − A∗ Y A = −(C ∗ + A∗ Y B)Δ−1 (C + B ∗ Y A) = −βΔ

(14.83)

∞ β∗ Aj because A has all its eigenvalues in the But then Y = − j=0 (A∗ )j βΔ open unit disc. Comparing the latter expression for Y with (14.82) we see that Y = −Γ∗ T −1 Γ. Thus Y is given by (14.64), as desired. In Theorem 14.15 we restricted the attention to stabilizing solutions of the Riccati equation (14.58) that are required to be Hermitian. This requirement is not essential: Theorem 14.15 remains true if Y is just an arbitrary stabilizing solution of (14.58). The reason is that a stabilizing solution of (14.58) is always Hermitian. This result is the contents of the following proposition. Proposition 14.19. If Y is a T-stabilizing solution of the Riccati equation (14.58), then Y is Hermitian. Proof. Let Y be a stabilizing solution of (14.58), and put Δ = D0 + B ∗ Y B. Then Δ is invertible, and (14.84) σ A − BΔ−1 (C + B ∗ Y A) ⊂ D. Consider the m × m rational matrix functions W− (λ)

=

Im + Δ−1 (C + B ∗ Y A)(λIn − A)−1 B,

(14.85)

W+ (λ)

=

Δ + B ∗ (λ−1 In − A∗ )−1 (C ∗ + A∗ Y B).

(14.86)

14.7. J-spectral factorization relative to the unit circle revisited

285

The ﬁrst part of the proof consists of showing that W (λ) = W+ (λ)W− (λ) and that this factorization is a left canonical one with respect to the unit circle. Part 1. To prove that W (λ) = W+ (λ)W− (λ), we use a modiﬁcation of the argument used to prove (14.75). Put C0 = C + B ∗ Y A,

B0 = C ∗ + A∗ Y B.

(14.87)

Then equation (14.58) can be rewritten as Y − A∗ Y A = −C0 Δ−1 B0 , and hence λB0 Δ−1 C0 = −Y (λ − A) + (I − λA∗ )Y (λ − A) − λ(I − λA∗ )Y. It then follows that B ∗ (λ−1 − A∗ )−1 B0 Δ−1 C0 (λIn − A)−1 B = −B ∗ (λ−1 − A∗ )−1 A∗ Y B − B ∗ Y B − B ∗ Y A)(λIn − A)−1 B. This yields W+ (λ)W− (λ)

=

Δ + B ∗ (λ−1 − A∗ )−1 B0 + C0 (λIn − A)−1 B +B ∗ (λ−1 − A∗ )−1 B0 Δ−1 (λIn − A)−1 B

=

D0 + B ∗ (λ−1 − A∗ )−1 C ∗ + C(λIn − A)−1 B = W (λ).

Next we prove that W (λ) = W+ (λ)W− (λ) is a left canonical factorization with respect to the unit circle. To do this, using (14.85), we ﬁrst note that W− (λ)−1 = Im − Δ−1 (C + B ∗ Y A)(λIn − A× )−1 B,

(14.88)

where A× = A − BΔ−1 (C + B ∗ Y A). From (14.84) we know that A× has all its eigenvalues in D. By assumption the same holds true for the matrix A. Thus (14.85) and (14.88) tell us that both W− and W−−1 are analytic on the complement of D, inﬁnity included. Thus the factor W− has the desired properties. As A has all its eigenvalues in D, the same holds true for A∗ . Thus (14.86) tells us that W+ is analytic on the closed unit disc D. We have to show that W+−1 also is analytic on D. To do this, put ¯ −1 )∗ , V+ (λ) = W− (λ

¯ −1 )∗ . V− (λ) = W+ (λ

Using the properties of W− derived in the previous paragraph, we see that V+ and V+−1 are analytic on D. Furthermore, V− is analytic on |λ| ≥ 1, inﬁnity included. ¯ −1 )∗ , and Now, recall that W is selfadjoint on the unit circle. Hence W (λ) = W (λ thus W (λ) = W+ (λ)W− (λ) = V+ (λ)V− (λ). But then V− (λ)W− (λ)−1 = V+ (λ)−1 W+ (λ).

(14.89)

286

Chapter 14. J-spectral factorization

The left-hand side of (14.89) is analytic on |λ| ≥ 1 with inﬁnity included, and the right-hand side of (14.89) is analytic on D. By Liouville’s theorem, there exists a constant matrix K such that V− (λ) = KW− (λ),

W+ (λ) = V+ (λ)K.

(14.90)

As det W+ (λ) does not vanish identically, K is invertible. Hence the second identity in (14.90) tells us that W+ (λ) = K −1 V+ (λ)−1 is analytic on D. Thus W+ and W+−1 are analytic on D, as desired. We conclude that W (λ) = W+ (λ)W− (λ) is a left canonical factorization with respect to the unit circle. Part 2. In this part we establish the inclusion (14.91) σ A∗ − (C ∗ + A∗ Y B)Δ−1 B ∗ ⊂ D. Put Φ(λ) = W+ (λ−1 ). Then, with Ω = A∗ and Ω× = A∗ − (C ∗ + A∗ Y B)Δ−1 B ∗ , Φ(λ)

=

Δ + B ∗ (λIn − Ω)−1 (C ∗ + A∗ Y B),

(14.92)

Φ−1 (λ)

=

Δ−1 − Δ−1 B ∗ (λIn − Ω× )−1 (C ∗ + A∗ Y B)Δ−1 .

(14.93)

We want to prove that σ(Ω× ) ⊂ D. Take |λ0 | ≥ 1. As σ(Ω) ⊂ D, we have λ0 ∈ σ(Ω), and hence λ0 ∈ σ(Ω) ∩ σ(Ω× ). From (14.92) and (14.93) we see that Ω× is the associate main matrix of the realization (14.92). But then λ0 ∈ σ(Ω) ∩ σ(Ω× ) implies that the realization in (14.92) is locally minimal at λ0 . Since W+ and W+−1 are analytic on D, the rational matrix function Φ has no poles or zeros on |λ| ≥ 1. But then the local minimality at λ0 implies that λ0 is not an eigenvalue of Ω× . Recall that λ0 is an arbitary complex number with |λ0 | ≥ 1. We conclude that σ(Ω× ) is contained in D, that is, (14.91) is proved. −1 ). Since Part 3. Let T be the block Toeplitz operator on m 2 determined by W (λ W admits a left canonical factorization with respect to the unit circle, the function W (λ−1 ) admits a right canonical factorization with respect to the unit circle, and hence T is invertible. We claim that ⎤ ⎡ C ⎥ ⎢ ⎢ CA ⎥ ∗ ⎥ ⎢ −1 A∗ C ∗ A∗2 C ∗ · · · T ⎢ (14.94) Y =− C ⎥. ⎢CA2 ⎥ ⎦ ⎣ .. . Since the values of W (λ−1 ) on the unit circle are Hermitian, the operator T is selfadjoint, and hence the same holds true for T −1 . But then the identity (14.94) shows that Y is Hermitian. Thus it remains to prove (14.94). To prove (14.94) we follow the same line of reasoning as in the proof of Lemma 14.18. Put Θ(λ) = ΔW− (λ−1 ),

Φ(λ) = W+ (λ−1 ).

(14.95)

14.7. J-spectral factorization relative to the unit circle revisited

287

Here W+ and W− are as in Part 1 of the proof; see (14.85) and (14.86). By the result of Part 1 we have that W− (λ−1 ) = Φ(λ)Δ−1 Θ(λ). Moreover, Θ and Θ−1 are analytic on D, and Φ and Φ−1 are analytic on |λ| ≥ 1 with inﬁnity included. Let TΘ and TΦ be the block Toeplitz operators on m 2 determined by Θ and Φ, respectively. By the results mentioned in the previous paragraph, the operators TΘ and TΦ are invertible, TΘ−1 = TΘ−1 and TΦ−1 = TΦ−1 . Furthermore, T −1 = TΘ−1 Ξ−1 TΦ−1 , where, as in the proof of Lemma 14.18, the operator Ξ is the block diagonal operator on m 2 given by Ξ = diag (Δ−1 , Δ−1 , Δ−1 , . . .). Note that Θ−1 (λ)

=

Δ−1 − Δ−1 (C + B ∗ Y A)(λ=1 In − A× )−1 BΔ−1 ,

Φ−1 (λ)

=

Δ−1 − Δ−1 B ∗ (λIn − Ω× )−1 (C ∗ + A∗ Y B)Δ−1 .

Here A× = A − BΔ−1 (C + B ∗ Y A),

Ω× = A∗ − (C ∗ + A∗ Y B)Δ−1 B ∗ ,

and the eigenvalues of these two matrices are all in the open unit disc. Let Γ be the operator deﬁned by (14.68). We now repeat the arguments used in the proof of Lemma 14.18, more speciﬁcally appearing in the paragraphs after (14.79). This together with a duality argument yields Γ∗ T −1 Γ = (Γ∗ TΘ−1)Ξ−1 (TΦ−1 Γ) =

∞

−1 γ (A∗ )j βΔ Aj .

(14.96)

j=0

Here β =

∗

C Δ

−1

∗

−A

∞

(A∗ )j C ∗ Δ−1 (C + B ∗ Y A)(A× )j BΔ−1 ,

j=0

γ =

Δ−1 C − Δ−1 B ∗

(Ω× )j (C ∗ + A∗ Y B)Δ−1 CAJ A.

j=0

Note that the Riccati equation (14.58) can be rewritten in the following two equivalent forms Y − A∗ Y A×

=

−C ∗ Δ−1 (C + B ∗ Y A),

Y − Ω× Y A =

−(C ∗ + A∗ Y B)Δ−1 C.

Since the eigenvalues of the matrices A, A∗ , A× and Ω× are all in the open unit disc, we see that the formulas for β and γ can be transformed into ∗ ∗ −1 β = (C + A Y B)Δ , γ = Δ−1 (C + B ∗ Y A). γ , and we see from (14.96) This allows us to rewrite (14.58) as Y − A∗ Y A = −βΔ that (14.94) holds.

288

Chapter 14. J-spectral factorization

Notes As noted J-spectral factorization is a special form of canonical factorization, reﬂecting the symmetry condition on the given function. This chapter develops this theme in a systematic way for rational matrix functions. Sections 14.2 and 14.3 are based on [121]. For Section 14.4 we refer to [76], see also [112] and [83]. A good source for Section 14.5 is [98], see also [97]. The linear quadratic optimal control problem for discrete time systems, mentioned in Section 14.5 in the paragraph before Proposition 14.12, can be found in many books on mathematical systems theory, see, e.g., [85]. The connection with the algebraic Riccati equation of the form (14.45) is also shown in the latter book. Much more information on this equation, including its connection to factorization in more general setting than the one exhibited in Proposition 14.12, can be found in Part III of [106]. Section 14.6 is based on [9], see also [8]. The ﬁnal section is inspired by [44]. In fact, Theorem 14.15 is just the symmetric version of Theorem 1.1 in [44]. The notion of J-spectral factorization plays an important role in control theory; see, e.g., the books [43], [85], [150], the papers [76], [145] and the references in these papers. The ﬁnal part of this book is devoted to this connection, with an emphasis on H∞ -problems.

Part VI Factorizations and symmetries In this part we study rational matrix functions that are unitary or of the form identity matrix plus contractions, and rational matrix functions that have a positive real part. Because of the state space similarity theorem, these additional symmetries can be restated in terms of special properties of the minimal realizations of the rational matrix functions considered. These reformulations involve an algebraic Riccati equation. The results are known in systems theory as the bounded real lemma and the positive real lemma, respectively. This part consists of three chapters. In the ﬁrst chapter (Chapter 15) we study rational matrix functions that have a positive deﬁnite real part or a nonnegative real part on the real line, and we present canonical and pseudo-canonical factorization theorems for such functions in state space form. In the second chapter (Chapter 16) realizations are used to study rational matrix functions of which the values on the imaginary axis (or on the real line) are contractive matrices. Included are solutions to spectral and canonical factorization problems for functions V of the form ¯ ∗ W (λ), V (λ) = I − W (−λ) V (λ) = I + W (λ), where W has contractive values on the imaginary axis (or on the real line) and is strictly contractive at inﬁnity. In the third chapter (Chapter 17) realizations are used to study rational matrix functions of which the values on the imaginary axis are J-unitary matrices. Solutions to various factorization problems are given. Special attention is paid to factorization of J-unitary rational matrix functions into J-unitary factors. In this chapter we also discuss problems of embedding a contractive rational matrix function into a unitary rational matrix function of larger size.

Chapter 15

Factorization of positive real rational matrix functions This chapter is concerned with canonical factorization (with respect to the real line) of rational matrix functions with a positive deﬁnite real part on the real line. Also the generalization to pseudo-canonical factorization for functions that have a nonnegative real part is developed. All factorizations are obtained explicitly using state space realizations of the functions involved. In Section 15.1 rational matrix functions that have a positive deﬁnite real part or a nonnegative real part on the real line are characterized in terms of realizations. Section 15.2 deals with canonical factorization, and Section 15.3, the ﬁnal section of the chapter, with pseudo-canonical factorization.

15.1 Rational matrix functions with a positive deﬁnite real part In this section we consider rational m × m matrix functions W which have the property that W (λ) + W (λ)∗ ≥ 0,

λ ∈ R, λ not a pole of W .

(15.1)

In this case we say that W has a nonnegative real part on the real line. If in (15.1) the inequality is strict, that is, W (λ) + W (λ)∗ > 0,

λ ∈ R, λ not a pole of W .

(15.2)

we say that W has a positive deﬁnite real part on the real line. The following two theorems characterize these properties in terms of realizations of W .

292

Chapter 15. Factorization of positive real rational matrix functions

Theorem 15.1. Let W (λ) = D + C(λIn − A)−1 B be a rational m × m matrix function, and let (A, B) be controllable. Write G = D + D∗ and assume G is positive deﬁnite. Then W has a nonnegative real part on the real line if and only if there is a Hermitian solution X of the equation −iA∗ X + iXA − (XB − iC ∗ ) G−1 (B ∗ X + iC) = 0.

(15.3)

Furthermore, for any Hermitian solution X of (15.3) one has

where

¯ ∗ = K(λ) ¯ ∗ K(λ), W (λ) + W (λ)

(15.4)

K(λ) = G1/2 + G−1/2 (C − iB ∗ X)(λIn − A)−1 B.

(15.5)

Finally, if, in addition, the pair (C, A) is observable, then each solution X of (15.3) is invertible. For later use we note that equation (15.3) can be rewritten as −(iA∗ −iC ∗ G−1 B ∗ )X +X(iA−iBG−1C)−C ∗ G−1 C −XBG−1B ∗ X = 0. (15.6) ¯ ∗ . Then W has a nonnegative real part on R Proof. Put V (λ) = W (−iλ) + W (iλ) if and only if V is nonnegative on the imaginary axis. Using the given realization of W we have V (λ)

= = =

D + C(−iλIn − A)−1 B + D∗ + B ∗ (−iλIn − A∗ )−1 C ∗ −1 −1 G + (iC) λIn − (iA) B − B ∗ λIn + (iA)∗ (iC)∗ . (λ − iA)−1 B 0 (iC)∗ −1 . I −B ∗ λ + (iA)∗ iC G I

Thus we can apply Theorem 13.6, with R = G, Q = 0, S = iC and iA instead of A, to show that W has a nonnegative real part on R if and only equation (15.3) has a Hermitian solution. Next, let X be a Hermitian solution of (15.3). By the second part of Theo¯ ∗ L(λ), where rem 13.6, the function V admits a factorization V (λ) = L(−λ) L(λ) = G1/2 + G−1/2 (B ∗ X + iC)(λ − iA)−1 B. ¯ ∗ = V (iλ), we see that (15.4) holds with K being given by (15.5) As W (λ) + W (λ) To prove the ﬁnal part, assume additionally that the pair (C, A) is observable, and let X be a Hermitian solution of (15.3). We have to show that X is invertible. Since X is square it suﬃces to prove that Ker X = {0}. Assume Xx = 0. Then x∗ X = 0 because X is Hermitian, and by (15.3) we have 0 = −C ∗ G−1 Cx, x. As G > 0, this gives Cx = 0. Multiplying (15.3) on the right by x we then obtain iXAx = 0. So Ker X is A-invariant and contained in Ker C. Therefore Ker X = {0} and X is invertible.

15.1. Rational matrix functions with a positive deﬁnite real part

293

Theorem 15.2. Let W (λ) = D + C(λIn − A)−1 B be a rational m × m matrix function, and let (A, B) be controllable. Write G = D + D∗ and assume G is positive deﬁnite. If, in addition, A has no real eigenvalues, then the following statements are equivalent: (i) The function W has a positive deﬁnite real part on the real line; (ii) Equation (15.3) has a Hermitian solution X such that the matrix A − BG−1 C + iBG−1 B ∗ X

(15.7)

has no real eigenvalues; (iii) The matrix

⎡ H =⎣

iA∗ − iC ∗ G−1 B ∗

C ∗ G−1 C

−BG−1 B ∗

iA − iBG−1 C

⎤ ⎦

has no pure imaginary eigenvalues. Moreover, in that case equation (15.3) has a unique Hermitian solution X such that the matrix (15.7) has its eigenvalues in the open upper half plane. Proof. As in the proof of the previous theorem, we consider the rational m × m ¯ ∗ . Using the given realization of W we matrix function V (λ) = W (−iλ) + W (iλ) see (see (13.6) and the second part of the proof of Theorem 13.2) that V admits −1 B, where 2n − A) the realization V (λ) = G + C(λI iC ∗ iA∗ 0 = = = B ∗ iC . , B A , C 0 iA B − BG −1 C is precisely equal to the block matrix H appear× = A It follows that A has no pure imaginary ing in item (c). Since A has no real eigenvalue, the matrix A eigenvalue. Thus V has no pole on the imaginary axis. Hence (cf., Section 8.1) the × )−1 BG 2n − A −1 is minimal at each point realization V −1 (λ) = G−1 − G−1 C(λI of the imaginary axis. But then V −1 has no pole on the imaginary axis if and only × has no pure imaginary eigenvalue. As A × = H, we conclude that condition if A (iii) is equivalent to the requirement that V (λ) is invertible for each λ ∈ iR. (i) ⇒ (iii) If (i) is satisﬁed, then V (λ) is positive deﬁnite for each λ ∈ iR. In particular, V (λ) is invertible for each λ ∈ iR, and hence, by the result of the previous paragraph, (iii) holds. (iii) ⇒ (i) Conversely, assume (iii) is satisﬁed. Recall that V has no pole on the imaginary axis. Furthermore, V (λ) is selfadjoint for λ ∈ iR. Since V (λ) is invertible for each λ ∈ iR, it follows that for imaginary λ the signature of the matrix V (λ) does not depend on λ. Next, observe that the rational matrix function V is biproper and that its value at inﬁnity is equal to G. Hence the value of V

294

Chapter 15. Factorization of positive real rational matrix functions

at inﬁnity is positive deﬁnite. We obtain that V (λ) is positive deﬁnite for each λ ∈ iR. Thus (i) holds. (i) ⇒ (ii) Assume W has a positive deﬁnite real part on R. Theorem 15.1 implies that equation (15.3) has a Hermitian solution X. Hence we have the fac¯ ∗ K(λ) with K(λ) being given by (15.5). Since ¯ ∗ = K(λ) torization W (λ) + W (λ) A has no eigenvalue on R, the functions W and K have no pole on R. The fact that W has a positive deﬁnite real part on R and the fact that W has no pole on ¯ ∗ is invertible for each λ ∈ R. Hence K(λ) is R together imply that W (λ) + W (λ) also invertible for each λ ∈ R. Thus K(λ)−1 has no pole on R. Notice that K(λ)−1 = G−1/2 − G−1 (C − iB ∗ X)(λ − Z)−1 B, −1

(15.8)

∗

where Z = A − BG (C − iB ). Let λ0 ∈ R. Then λ0 is not a common eigenvalue of A and Z. Thus we can apply the material presented in Section 8.1 to show that the realization given by the right-hand side of (15.8) is minimal at λ0 . But then the fact that K(λ)−1 has no pole on R implies that λ0 is not an eigenvalue of Z. Thus Z = A − BG−1 C + iBG−1 B ∗ X has no real eigenvalue. This proves (ii). ¯ ∗ = K(λ) ¯ ∗ K(λ) with K(λ) (ii) ⇒ (i) Let X be as in (ii). Then W (λ) + W (λ) −1 being given by (15.5). Observe that K(λ) is given by (15.8), where Z is as ¯ ∗ K(λ) above. According to our hypothesis Z has no real eigenvalue. Hence K(λ) is positive deﬁnite for each λ ∈ R. Thus (i) holds. To prove the second part of the theorem, we apply Theorem 13.3. Recall that equation (15.3) can be rewritten into the algebraic Riccati equation (15.6). The Hamiltonian of this Riccati equation is precisely the block matrix H deﬁned in item (iii). According to our hypotheses (A, B) is controllable. This implies that the pair (iA − iBG−1 C, B) is also controllable. But controllability implies stabilizability. Thus the pair (iA − iBG−1 C, B) is stabilizable. But then Theorem 13.3 tells us that condition (iii) implies that equation (15.3) has a unique Hermitian solution X such that the eigenvalues of iA − iBG−1 C − BG−1 B ∗ X are in the open left half plane. Multiplication by −i then gives the desired result.

15.2 Canonical factorization of functions with a positive deﬁnite real part In this section we consider canonical factorization of functions with a positive deﬁnite real part on the real line. Using state space realizations we shall prove the following result. Theorem 15.3. Let W be a proper rational matrix function having no real poles and such that D = W (∞) satisﬁes D + D ∗ > 0. Assume that W has a positive deﬁnite real part on the real line. Then W admits both a right and a left canonical factorization with respect to the real line. We start with some preparations that are of independent interest and will be useful in the next section too. Let T be a square matrix. If the real part of T

15.2. Canonical factorization of functions with a positive deﬁnite real part

295

is positive deﬁnite, then T is injective, hence invertible. Indeed, for non-zero x we have 2(T x, x) = (T + T ∗ )x, x > 0. Also, if T is invertible, then T −1 has a positive deﬁnite real part if and only if this is the case for T . This is immediate from either of the identities T −1 + T −∗ = T −1(T + T ∗ ) T −∗ ,

T −1 + T −∗ = T −∗ (T + T ∗ ) T −1 .

Now let W (λ) = D + C(λIn − A)−1 B be a rational m × m matrix function with G = D + D ∗ positive deﬁnite, and assume W has a nonnegative real part on R. Then D is invertible, G× deﬁned by G× = D−1 + D−∗ is positive deﬁnite, G× = D−1 G D−∗ , and W −1 has a nonnegative real part on R. For W −1 we have the realization W −1 (λ) = D−1 − D−1 C(λIn − A× )−1 BD−1 ,

(15.9)

where, as usual, A× = A − BD−1 C. This gives rise to the following analogue of equation (15.3): −i(A× )∗ X + iXA× − (XBD−1 + iC ∗ D−∗ )(G× )−1 (D −∗ B ∗ X − iD −1 C) = 0, (15.10) which can also be written as an algebraic Riccati equation − i(A× )∗ + iC ∗ D−∗ (G× )−1 D−∗ B ∗ X (15.11) +X iA× + iBD−1 (G× )−1 D −1 C −C ∗ D −∗ (G× )−1 D−1 C − XBD−1 (G× )−1 D−∗ B ∗ X = 0. Now let us look at the right coeﬃcient of X in this expression. Using the identity (G× )−1 = DG−1 D∗ , we get iA× + iBD−1 (G× )−1 D−1 C

=

iA − iBD−1 C + iBD−1 (DG−1 D∗ )D −1 C

=

iA − iBD−1 C + iBG−1 D ∗ D−1 C

=

iA − iBG−1 (G − D∗ )D−1 C

=

iA − iBG−1 DD−1 C = iA − iBG−1 C.

Thus the right coeﬃcient of X in (15.11) is equal to the right coeﬃcient of X in (15.6). The left coeﬃcient of X in (15.11) is the adjoint of the right coeﬃcient of X in (15.11), and the same is true with (15.11) replaced by (15.6). Hence the left coeﬃcient of X in (15.11) is equal to the left coeﬃcient of X in (15.6). For the constant term in (15.11), we have −C ∗ D−∗ (G× )−1 D−1 C = −C ∗ D−∗ (D ∗ G−1 D)D−1 C = −C ∗ G−1 C, and the latter is the constant term in (15.11). Finally, the identities −BD −1 (G× )−1 D −∗ B ∗ = −BD−1 (DG−1 D∗ )D−∗ B ∗ = −BG−1 B ∗

296

Chapter 15. Factorization of positive real rational matrix functions

show that the coeﬃcients of the quadratic terms in (15.11) and (15.6) coincide too. We conclude that the equations (15.3), (15.6), (15.10) and (15.11) all amount to the same. Lemma 15.4. Let W (λ) = D +C(λIn −A)−1 B be a rational m×m matrix function such that G = D + D∗ > 0. Assume X is an invertible Hermitian matrix satisfying (15.3). Then 1 1 (XA − A∗ X) = − (B ∗ X + iC)∗ G−1 (B ∗ X + iC), 2i 2 1 1 = − (DD−∗ B ∗ X − iC)∗ G−1 (DD−∗ B ∗ X − iC). XA× − (A× )∗ X 2i 2 In particular both A and A× are (−X)-dissipative. Proof. The ﬁrst identity is just a restatement of (15.3). Recall that (15.3) and (15.10) amount to the same. Hence X also satisﬁes (15.10). Now note that the second identity in the lemma is just another way of writing (15.10). Here we use that (G× )−1 = D∗ G−1 D. Before turning to the proof of Theorem 15.3 we present another lemma. Lemma 15.5. Let W (λ) = D +C(λIn −A)−1 B be a rational m×m matrix function such that G = D + D ∗ > 0 and the pair (C, A) is observable. Assume X is an invertible Hermitian matrix satisfying (15.3). Let N1 , N1× be maximal X-nonpositive subspaces and N2 , N2× be maximal X-nonnegative subspaces such that N1 , N2 are invariant under A and N1× , N2× are invariant under A× . Then ˙ N2× , Cn = N1 +

˙ N1× . C n = N2 +

(15.12)

Proof. Applying Proposition 11.1 we obtain dim N1 + dim N2× = n,

dim N2 + dim N1× = n.

Therefore in order to prove that (15.12) holds, it suﬃces to show that the intersections N1 ∩ N2× and N2 ∩ N1× are both trivial. Take x ∈ N1 ∩ N2× . Then Xx, x = 0. Now the Cauchy-Schwartz inequality holds on N1 . Thus |XAx, x|2 ≤ XAx, AxXx, x = 0, |Xx, Ax|2 ≤ XAx, AxXx, x = 0. Using this together with the ﬁrst identity in Lemma 15.4, we get 1 0 = XAx, x = − G−1/2 (B ∗ X + iC)x2 . 2 Similarly, employing the Cauchy-Schwartz inequality on N2× and the second identity in Lemma 15.4, we get 1 0 = XA× x, x = − G−1/2 (DD−∗ B ∗ X − iC)x2 . 2

15.3. Generalization to pseudo-canonical factorization

297

Thus (B ∗ X + iC)x = 0 and (DD−∗ B ∗ X − iC)x = 0. Adding these two identities we arrive at 0 = (I + DD −∗ )B ∗ Xx = GD −∗ B ∗ Xx. Hence B ∗ Xx = 0, and it also follows that Cx = 0. Thus Ax = A× x for x ∈ N1 ∩ N2× . Hence N1 ∩ N2× is an A-invariant subspace contained in Ker C. Given the observability of the pair (C, A), this yields N1 ∩ N2× = {0}. The proof of N2 ∩N1× = {0} is analogous. It can also be obtained by applying the result of the previous paragraph to the rational matrix function W −1 . Proof of Theorem 15.3. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of W . Since W has no poles on R, the minimality of the realization guarantees that A has no eigenvalues on R. As we have seen (in the ﬁrst paragraph after Theorem 15.3), the positive deﬁniteness of D + D∗ implies that D is invertible. Similarly we conclude that W takes invertible values on R. Hence we know from Theorem 2.4 that A× = A − BD−1 C has no real eigenvalues either. Since W has a positive deﬁnite real part, we can use Theorem 15.1 to deduce that equation (15.3) has an invertible Hermitian solution X, say. Lemma 15.4 now gives that both A and A× are (−X)-dissipative. × be the spectral subspaces of A and A× , respectively, corLet M+ and M+ × responding to the open upper half plane, and let M− and M− , be the spectral subspaces of A and A× , respectively, corresponding to the open lower half plane. × As A and A× are (−X)-dissipative, we have that M+ and M+ are maximal X× nonpositive. Similarly, the spaces M− and M− are maximal X-nonnegative. Using × × ˙ M− ˙ M+ Lemma 15.5 we may conclude that Cn = M+ + and Cn = M− + . But then Theorem 3.2 guarantees that W admits the desired canonical factorizations.

15.3 Generalization to pseudo-canonical factorization In this section the results of the previous section concerning canonical factorizations will be generalized to pseudo-canonical factorizations. Theorem 15.6. Let W be a proper rational m × m matrix function having no real poles such that D = W (∞) satisﬁes D+D ∗ > 0. Assume that W has a nonnegative real part on the real line. Then, with respect to the real line, W admits both right and left pseudo-canonical factorization. Such factorizations can be obtained in the following manner. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization, and put G = D + D∗ . Then there exists an invertible Hermitian matrix X satisfying −iA∗ X + iXA − (XB − iC ∗ ) G−1 (B ∗ X + iC) = 0.

(15.13)

Also there are A-invariant subspaces M+ and M− , and A× -invariant subspaces × × M+ and M− , such that (i) M+ is maximal X-nonpositive, M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane, and σ(A|M+ ) ⊂ {λ | λ ≥ 0},

298

Chapter 15. Factorization of positive real rational matrix functions

(ii) M− is maximal X-nonnegative, M− contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane, and σ(A|M− ) ⊂ {λ | λ ≤ 0}, × × is maximal X-nonpositive, M+ contains the spectral subspace of A× (iii) M+ × associated with the part of σ(A ) lying in the open upper half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0}, +

(iv)

× M−

× is maximal X-nonnegative, M− contains the spectral subspace of A× × associated with the part of σ(A ) lying in the open lower half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}. −

For such subspaces the matching conditions × ˙ M− Cn = M+ + ,

× ˙ M+ Cn = M− +

(15.14)

× and Πl for the are satisﬁed. Write Πr for the projection along M+ onto M− × projection along M− onto M+ . Further put

!− (λ) W

=

D + C(λIn − A)−1 (In − Πr )B,

!+ (λ) W

=

In + D−1 CΠr (λIn − A)−1 B,

W+ (λ)

=

D + C(λIn − A)−1 (In − Πl )B,

W− (λ)

=

In + D−1 CΠl (λIn − A)−1 B.

!− (λ)W !+ (λ) and W (λ) = W+ (λ)W− (λ) are a right and a left Then W (λ) = W pseudo-canonical factorization with respect to the real line, respectively. Proof. In view of the minimality of the given realization we can employ Theorem 15.1 to show that there is an invertible Hermitian matrix X such that (15.13), which is identical to (15.3), holds. By Lemma 15.4 the matrices A and × × and M− with A× are (−X)-dissipative. The existence of subspaces M+ , M− , M+ the properties mentioned above is now guaranteed by Theorem 11.6. Lemma 15.5 gives the direct sums (15.14), and the conclusion of the theorem is straightforward by Theorem 8.6. As a further application of Lemma 15.5 we prove the following result on skew selfadjoint matrix functions. A rational m × m matrix function W is called skewHermitian on the real line if W (λ) is skew-Hermitian for all λ in R, λ not a pole of W . Proposition 15.7. Let W (λ) = D + V (λ), where V is a strictly proper rational m × m matrix function that has no real poles, is skew-Hermitian on the real line and vanishes at inﬁnity. Assume D + D ∗ > 0. The following statements are true.

15.3. Generalization to pseudo-canonical factorization

299

!1 (λ)W !2 (λ) where W !1 has all (i) W admits a minimal factorization W (λ) = W its poles, respectively zeros, in the open upper, respectively lower, half plane, !2 has all its poles, respectively zeros, in the open lower, respectively and W upper, half plane. (ii) W admits a minimal factorization W (λ) = W1 (λ)W2 (λ) where W1 has all its poles, respectively zeros, in the open lower, respectively upper, half plane, and W2 has all its poles, respectively zeros, in the open upper, respectively lower, half plane. Proof. Recall that D + D ∗ > 0 implies that D is invertible. Since V is skewHermitian on the real line, we see that W (λ) + W (λ)∗ = D + D ∗ > 0 for λ ∈ R. From the latter it follows that W (λ) is invertible for each λ ∈ R. Now let W (λ) = D + C(λIn − A)−1 B be a minimal realization of W . Then both A and A× have no eigenvalues on the real line. From C(λIn − A)−1 B = −(C(λIn − A)−1 B)∗ for λ ∈ R and the minimality of the realization we may conclude (by the state space similarity theorem) that there is a unique invertible matrix Y such that Y A = A∗ Y,

Y B = C∗,

C = −B ∗ Y.

Taking adjoints in the above equations, and using the uniqueness of Y , one deduces that Y = −Y ∗ . Put X = −iY . Then X is selfadjoint. As XA = A∗ X, the matrix A is X-selfadjoint. Furthermore, from −iA∗ X + iXA = 0 and XB − iC ∗ = 0, we see that X is an invertible Hermitian solution of (15.13). But then we can use Lemma 15.4 to show that A× is (−X)-dissipative. Let Mu and Ml be the spectral subspaces of A associated with the part of σ(A) lying in the open upper and open lower half plane, respectively. Also let Mu× and Ml× be the spectral subspaces of A× associated with the part of σ(A× ) lying in the open upper and open lower half plane, respectively. Since the matrix A is X-selfadjoint and has no real eigenvalues, we know (see Theorem 11.5) that the spaces Mu and Ml are X-Lagrangian. In particular, these spaces are both maximal X-nonpositive and maximal X-nonnegative. The fact that A× is (−X)-dissipative and has no real eigenvalues either, gives that the same conclusion holds for Mu× ˙ Mu× as well as Cn = Ml + ˙ Ml× . and Ml× . But then Lemma 15.5 gives Cn = Mu + Let Π bethe projection of Cn along Mu onto Mu× . Then Π is a supporting projection of the minimal realization W (λ) = D + C(λI − A)−1 B. Hence W !2 (λ) such that (see Chapter 8) !1 (λ)W admits a minimal factorization W (λ) = W !1 and W !2 coincide with the eigenvalues of A|Mu the following holds: the poles of W !1 and W !2 coincide with the eigenvalues and A|Mu× , respectively, and the zeros of W × × of A |Mu and A |Mu× , respectively. Since A and A× have no real eigenvalues ˙ Mu× = Cn , we have σ(A|M × ) = σ(A|Ml ) and σ(A× |Mu ) = σ(A× |M × ). and Mu + u l !2 (λ) has the !1 (λ)W From these remarks it is clear that the factorization W (λ) = W desired properties. The factorization W (λ) = W1 (λ)W2 (λ) is obtained in a similar ˙ Ml× . way using the other direct sum decomposition Cn = Ml +

300

Chapter 15. Factorization of positive real rational matrix functions

Notes This chapter is based on [126], see also [129] and [128]. Rational matrix functions with a positive deﬁnite real part play a role in circuit and systems theory. In particular, Theorem 15.2 is a version of what is known as the positive real lemma. There are several variants of this result, see, for instance, Section 5.2 in [4], where also the connection with spectral factorization and Riccati equations is discussed. Another version in terms of Riccati inequalities is given in Section 12.6.3 in [83]. An inﬁnite dimensional version may be found as Exercise 6.28 in [35].

Chapter 16

Contractive rational matrix functions In this chapter rational matrix functions are studied of which the values on the imaginary axis or on the real line are contractive matrices. Included are solutions to spectral or canonical factorization problems for functions V of the form ¯ ∗ W (λ) V (λ) = I − W (−λ)

or V (λ) = I + W (λ),

where W is a rational matrix function which has contractive values on the imaginary axis or on the real line and, in addition, has a strictly contractive value at inﬁnity. This chapter consists of ﬁve sections. Sections 16.1 and 16.2 present a state space analysis (involving algebraic Riccati equations) of rational matrix functions that are contractive or strictly contractive on the imaginary axis. In Section 16.3 a state space formula is derived for the spectral factor in a spectral factorization of a ¯ ∗ W (λ), where W is strictly rational matrix function of the form V (λ) = I −W (−λ) proper and strictly contractive on the imaginary axis. The ﬁnal two sections of the chapter deal with canonical and pseudo-canonical factorization, respectively, for functions of the form V (λ) = I + W (λ), where W (λ) is strictly proper and strictly contractive for real λ (Section 16.4) or just contractive (Section 16.5).

16.1 State space analysis of contractive rational matrix functions A rational p × m matrix function W is called contractive on the imaginary axis if the values that W takes on the imaginary axis are contractive matrices. Such a function does not have a pole on the imaginary axis. Moreover, it is proper and the value at inﬁnity is again contractive. Of special interest is the subclass consisting

302

Chapter 16. Contractive rational matrix functions

of the contractive rational matrix functions W that are strictly contractive at inﬁnity, i.e., the value of W at ∞ has norm smaller than 1. The ﬁrst main result of this section is a characterization of this subclass in terms of realizations. Theorem 16.1. Let W (λ) = D + C(λIn − A)−1 B be a realization of a p × m rational matrix function, and assume D is a strict contraction. Then the following two assertions hold: (i) Assume (C, A) is an observable pair. Then W is contractive on the imaginary axis if and only if the algebraic Riccati equation −AP − P A∗ − BB ∗ − (P C ∗ + BD∗ )(I − DD ∗ )−1 (CP + DB ∗ ) = 0 (16.1) has a Hermitian solution P . (ii) Assume (A, B) is a controllable pair. Then W is contractive on the imaginary axis if and only if the algebraic Riccati equation A∗ P + P A − C ∗ C − (P B − C ∗ D)(I − D∗ D)−1 (B ∗ P − D∗ C) = 0 (16.2) has a Hermitian solution P . ¯ ∗ . Since W is proper, the same holds true for Proof. Put V (λ) = I − W (λ)W (−λ) ∗ V . Moreover, V (∞) = I − DD , and hence V (∞) is positive deﬁnite, because D is assumed to be a strict contraction. Note that W is contractive on iR if and only if V is nonnegative on iR. Using the given realization for W we have V (λ)

∗

= I − DD +

=

C

DB

−1

−C(λ − A)

I

∗

$

λ−

A BB ∗ 0

%−1

−A∗

−BB ∗

BD∗

DB ∗

I − DD∗

−BD ∗

C∗

(λ + A∗ )−1 C ∗ I

.

The latter expression is of the form (13.25) and we see that (i) is an immediate consequence of the equivalence of statements (i) and (ii) in Theorem 13.6. To prove assertion (ii) we use a duality argument. First note that a matrix X is a (strict) contraction if and only if X ∗ is a (strict) contraction. So W is ¯ ∗ . The latter contractive on iR if and only this is the case for the function W (−λ) ∗ ∗ ∗ ∗ −1 ∗ ¯ has the realization W (−λ) = D − B (λ + A ) C . Also, the controllability of the pair (A, B) implies the observability of (B ∗ , −A∗ ). Finally, D∗ is a strict contraction. Thus assertion (ii) follows from part (i) by taking adjoints. Suppose D is a strict contraction. If the pair (C, A) is observable, then each Hermitian solution P of (16.2) is invertible. To see this, we argue as follows. Assume P x = 0. Multiplying (16.2) from the left by x∗ and from the right by x yields x∗ C ∗ Cx + x∗ C ∗ D(I − D∗ D)−1 DC ∗ x = 0. Now C ∗ C and I − D∗ D are

16.1. State space analysis of contractive rational matrix functions

303

nonnegative (in fact even I − DD∗ > 0), and it follows that x∗ C ∗ Cx = 0. Hence Cx = 0. But then, multiplying (16.2) on the right by x, we get P Ax = 0. So Ker P is A-invariant and contained in Ker C. As (C, A) is an observable pair, it follows that Ker P = {0}. Since P is a square matrix, this yields the invertibility of P . In a similar fashion one proves that each solution of (16.1) is invertible provided that the pair (A, B) is controllable, or, which amounts to the same, the pair B ∗ , A∗ ) is observable. Finally we note that P is an invertible solution of (16.2) if and only if −P 1 is an invertible solution of (16.1). Indeed, replacing P by −P −1 in (16.1) and multiplying from the left and the right with P , one gets (16.2). In working out the details, identities of the type D(I − D∗ D)−1 = (I − DD∗ )−1 D and I + D(I − D ∗ D)−1 D∗ = (I − DD ∗ )−1 play a role. Theorem 16.2. Let W (λ) = D+C(λIn −A)−1 B be a realization of a p×m rational matrix function, and let D be a strict contraction. Assume, in addition, that the pair (C, A) is observable. Then W is contractive on the imaginary axis if and only if the matrix A + BD∗ (I − DD∗ )−1 C B(I − D ∗ D)−1 B ∗ T = (16.3) −C ∗ (I − DD∗ )−1 C −A∗ − C ∗ (I − DD∗ )−1 DB ∗ has only even partial multiplicities at its pure imaginary eigenvalues. Proof. Let V be as in the proof of Theorem 16.1, and recall that W is contractive on iR if and only if V is nonnegative on iR. The desired result is now immediate from the equivalence of statements (ii) and (iii) in Theorem 13.6 combined with the fact that (16.3) is the Hamiltonian of the equation (16.1). Theorem 16.2 has a counterpart in which (16.3) is replaced by the Hamiltonian of (16.2). As a special case of Theorem 16.1 let us consider rational matrix functions which are contractive not only on the imaginary axis but on the full closed right half plane. Theorem 16.3. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a rational p× m matrix function. Assume that W is contractive on the imaginary axis, and let D be a strict contraction. Then the following statements are equivalent: (i) For each λ in the closed right half plane, λ not a pole of W , the matrix W (λ) is a contraction; (ii) The matrix A has all its eigenvalues in the open left half plane; (iii) There is a positive deﬁnite solution of (16.1). Proof. Suppose A has all its eigenvalues in the open left half plane. Then W (λ) is analytic in the closed right half plane. As W (λ) is a contraction for each λ ∈ iR and at inﬁnity, the maximum modulus theorem implies W (λ) is contractive for all

304

Chapter 16. Contractive rational matrix functions

λ in the open right half plane as well. Thus (ii) implies (i). Conversely, suppose (i) holds. Then W must be analytic in the closed right half plane (inﬁnity included), and by minimality of the realization the matrix A has all its eigenvalues in the open left half plane. The equivalence of (ii) and (iii) follows by rewriting (16.1) as AP + P A∗ = RR∗ , where Q = −P and R = B (P C ∗ + BD ∗ )(I − DD ∗ )−1/2 . Since the realization is minimal, (A, B) is a controllable pair, and hence the same holds true for the pair (A, R). But then we can apply a well-known inertia theorem (see, e.g., Theorem 13.1.4 in [107]) to show that (ii) and (iii) are equivalent.

16.2 Strictly contractive rational matrix functions In this section we specify further the results of the previous section for the case of rational matrix functions W that are strictly contractive on the imaginary axis. By this we mean that W (λ) < 1 for λ ∈ iR. Such a function does not have a pole on iR and is proper. Theorem 16.4. Let W (λ) = D +C(λIn −A)−1 B be a realization of a p×m rational matrix function W . Assume A has no pure imaginary eigenvalues, D is a strict contraction, and the pair (C, A) is observable. Then the following statements are equivalent: (i) The function W is strictly contractive on the imaginary axis; (ii) Equation (16.1) has an iR-stabilizing solution P , that is, it has a solution P such that −A∗ − C ∗ (I − DD∗ )−1 DB ∗ − C ∗ (I − DD∗ )−1 CP has all its eigenvalues in the open left half plane; (iii) The matrix T given by (16.3) has no pure imaginary eigenvalues. Moreover, if one of the above conditions is satisﬁed, then the iR-stabilizing solution P in (ii) is unique and Hermitian. Proof. Suppose W is strictly contractive on iR. Then the rational m × m matrix ¯ ∗ is positive deﬁnite on iR and V (∞) = I − DD ∗ function V (λ) = I − W (λ)W (−λ) is positive deﬁnite too. In particular, V is biproper and V has no pole or zero on iR. Recall (see the proof of Theorem 16.1) that %−1 $ −BD∗ A BB ∗ ∗ ∗ V (λ) = I − DD + C DB . λ− 0 −A∗ C∗ The associate main matrix of this realization is T given by (16.3). It follows that T has no eigenvalues on iR. So (iii) holds. Conversely, if T has no pure imaginary eigenvalues, then V has no poles or zeros on iR. As V (∞) is positive deﬁnite, it follows that V (λ) is positive deﬁnite for λ ∈ iR. Hence W (∞) is strictly contractive

16.3. An application to spectral factorization

305

for λ ∈ iR. We have now proved the equivalence of (i) and (iii). The equivalence of (ii) and (iii) is a direct consequence of Theorem 13.3. Note here that the observability of the pair (C, A) is equivalent to the controllability of (A∗ , C ∗ ), and the latter implies the stabilizabilty of (A∗ , C ∗ ). The ﬁnal statement of the theorem is covered by Theorem 13.3 as well. Corollary 16.5. Let (C, A) be an observable pair, and assume that A has no pure imaginary eigenvalue. Then the Riccati equation Y C ∗ CY − Y A∗ − AY = 0 has a unique Hermitian solution Y such that A − Y C ∗ C has all its eigenvalues in the open right half plane. Proof. Apply Theorem 16.4 with D = 0 and B = 0. Then D is a strict contraction and W is identically equal to zero. In particular, (i) in Theorem 16.4 is satisﬁed. Next, note that with D = 0 and B = 0 equation (16.1) reduces to −AP − P A∗ − P C ∗ CP = 0, and by Theorem 16.4, with D = 0 and B = 0, this equation has a unique Hermitian solution P such that −A∗ −C ∗ CP has all its eigenvalues in the open left half plane. But then A − Y C ∗ C has all its eigenvalues in the open right half plane. Now put Y = −P , then we see that Y has all the desired properties.

16.3 An application to spectral factorization In this section we consider functions of the form ¯ ∗ W (λ), V (λ) = I − W (−λ)

(16.4)

where W is a proper rational p × m matrix function which is strictly contractive on the imaginary axis. In fact we shall assume that W is strictly proper , that is W vanishes at inﬁnity. Thus V is positive deﬁnite on the imaginary axis and has a positive deﬁnite value at inﬁnity (namely Im ). Hence W admits a right spectral factorization. Using a minimal realization of W , such a factorization is constructed in the following theorem. Theorem 16.6. Let W (λ) = C(λIn − A)−1 B be a minimal realization of the p × m rational matrix function W which is strictly contractive on the imaginary axis. Then the Riccati equations XBB ∗ X − XA − A∗ X + C ∗ C

=

0,

(16.5)

Y C ∗ CY − Y A∗ − AY

=

0,

(16.6)

have Hermitian solutions X and Y , respectively, such that the matrices A−BB ∗ X and A−Y C ∗ C have all their eigenvalues in the open right half plane, and In −XY

306

Chapter 16. Contractive rational matrix functions

is invertible (or, which amounts to the same, In − Y X is invertible). Furthermore, ¯ ∗ W (λ) admits with respect to the imaginary axis, the function V (λ) = Im − W (−λ) ∗ ¯ the right spectral factorization V (λ) = L+ (−λ) L+ (λ) with L+ and its inverse L−1 + being given by L+ (λ)

=

I + B ∗ X(In − Y X)−1 (λIn − A + Y C ∗ C)−1 B,

(16.7)

L−1 + (λ)

=

I − B ∗ X(λIn − A + BB ∗ X)−1(In − Y X)−1B.

(16.8)

Proof. By Corollary 16.5, the equation (16.6) has a Hermitian solution Y such that A − Y C ∗ C has all its eigenvalues in the open right half plane. Next, we apply Theorem 16.4 to ¯ ∗ = −B ∗ (λ + A∗ )−1 C ∗ . ! (λ) = W (−λ) W ! (λ) = −B ∗ (λ + A∗ )−1 C ∗ satisﬁes the general hypotheses of TheoNotice that W ! is strictly contractive on iR and its value at inﬁnity is rem 16.4. Furthermore, W zero. In particular, item (i) in Theorem 16.4 is satisﬁed. Hence item (iii) is satisﬁed as well, i.e., the matrix T of (16.3) has no pure-imaginary eigenvalues. ¯ ∗ W (λ) which has the realization Now consider the function V (λ) = I−W (−λ) %−1 $ 0 −A∗ C ∗ C ∗ 0 . (16.9) V (λ) = I + B λ− B 0 A

Put = A

−A∗

C ∗C

0

A

,

×

A =

−A∗

C ∗C

−B ∗ B

A

.

Since W is contractive on the imaginary axis, the function W has no pure imaginary poles. According to our assumptions the given realization of W is minimal. This implies that A has no eigenvalue on iR. But then we can use the triangular to show that the same holds true for the matrix A. Since V is positive form of A deﬁnite on the imaginary axis, we know that V (λ) is invertible for each λ ∈ iR. × has no pure imaginary eigenvalues either. But then Theorem 2.4 gives that A × are similar.) (Alternatively, this may be seen from the fact that T and A with respect to the open Let M− be the spectral subspace of the matrix A × × with respect to the left half plane, and let M+ be the spectral subspace of A open right half plane. Observe that V is positive deﬁnite on the imaginary axis and has a positive deﬁnite value at inﬁnity, namely Im . This suggests the use of × ˙ M+ Theorem 9.4 to show that C2n = M− + . For this a skew-Hermitian H must be identiﬁed with the properties required in Theorem 9.4. This can be done along × ˙ M+ the lines indicated in the proof of Theorem 13.1. So indeed C2n = M− + . The fact that Y is Hermitian and the eigenvalues of A∗ − C ∗ CY are in the open right half plane implies that σ(A − Y C ∗ C) ∩ σ(−A∗ + C ∗ CY ) = ∅. Hence Proposition ∗ 12.1 gives that the spectral subspace M− is given by M− = Im I Y .

16.4. An application to canonical factorization

307

× × by Theorem 11.5. Now M+ is an H-Lagrangian invariant subspace for A From Theorem 13.6 we see that there is a Hermitian solution X of (16.5) such ∗ × × to M × that M+ = Im X I . Moreover, A − BB ∗ X and the restriction of A + ∗ have the same eigenvalues. Thus, the eigenvalues of A − BB X are in the open × ˙ M+ right half plane. As C2n = M− + , the invertibility of I − XY follows from Lemma 12.4. Finally, we apply Theorem 12.3 to show that V admits the factorization V (λ) = V1 (λ)V2 (λ), where

V1 (λ)

=

I − B ∗ (λ + A∗ − C ∗ CY )−1 (I − XY )−1 XB ∗ ,

V2 (λ)

=

I + B ∗ X(I − Y X)−1 (λ − A + Y C ∗ C)−1 B,

V1−1 (λ)

=

I + B ∗ (I − XY )−1 (λ + A∗ − XBB ∗ )−1 XB,

V2−1 (λ)

=

I − B ∗ X(λ − A + BB ∗ X)−1 (I − Y X)−1 B.

−1 Clearly, V2 = L+ and V2−1 = L−1 + with L+ and L+ being given by (16.7) and (16.7), respectively. Furthermore, taking into account that X and Y are Hermitian,

¯ ∗ L+ (−λ)

¯ ∗ = V2 (−λ) = I + B(−λ − A∗ + CC ∗ Y )−1 (I − XY )−1 XB = V1 (λ).

¯ ∗ L+ (λ), and from the location of the eigenvalues of Thus we have V (λ) = L+ (−λ) ∗ ∗ A − Y C C and A − BB X we see that this is a right spectral factorization.

16.4 An application to canonical factorization Consider a function of the form V (λ) = Im + C(λIn − A)−1 B,

(16.10)

where W (λ) = C(λIn − A)−1 B is strictly contractive on the real line. By this we mean that the values of W on R are strict contractions, and this implies that W has no pole on the real line. Hence the latter holds true for V too. It follows also that V takes invertible values on the real line, i.e., V has no zero there. Now assume for the moment that (16.10) is a minimal realization for W . Since V has neither a pole nor a zero on the real line, the minimality of the realization implies that the matrices A and A× = A−BC have no real eigenvalues. ! (λ) = C(iλIn − A)−1 B is strictly contractive Furthermore, since the function W on the imaginary axis, we can apply Theorem 16.1(ii) to establish the existence of a Hermitian matrix X for which iXA − iA∗ X + XBB ∗ X + C ∗ C = 0.

(16.11)

308

Chapter 16. Contractive rational matrix functions

Finally, because of the minimality (see the remark in the paragraph after the proof of Theorem 16.1), such a matrix X is invertible. Summarizing, if (16.10) is a minimal realization and the matrix function W (λ) = C(λIn − A)−1 B is strictly contractive for real λ, then both A and A× have no real eigenvalues and there exists a Hermitian invertible matrix X solving (16.11). The next theorem describes canonical factorizations of a function of the form (16.10) in terms of a realization having the properties just described. Theorem 16.7. Let V (λ) = Im + C(λIn − A)−1 B be a realization of an m × m rational matrix function such that A and A× = A − BC have no real eigenvalues, and assume that there exists a Hermitian invertible X satisfying (16.11), i.e., iXA − iA∗ X + XBB ∗ X + C ∗ C = 0. Let M− and M+ be the spectral subspaces of A associated with the parts of σ(A) × lying in the open lower and open upper half plane, respectively, and let M− and × × × M+ be the spectral subspaces of A associated with the parts of σ(A ) lying in the open lower and open upper half plane, respectively. Then × ˙ M+ , Cn = M − +

× ˙ M− Cn = M+ + .

(16.12)

Moreover, V admits both a left and a right canonical factorization with respect to the real line, V (λ) = V+ (λ)V− (λ), V (λ) = V− (λ)V+ (λ), with the factors being given by V+ (λ)

= Im + C(λIn − A)−1 (In − Πl )B,

V− (λ)

= Im + CΠl (λIn − A)−1 B,

V− (λ)

= In + C(λIn − A)−1 (In − Πr )B,

V+ (λ)

= Im + CΠr (λIn − A)−1 B.

× Here Πl is the projection along M− onto M+ , and Πr is the projection along M+ × onto M− .

Proof. In view of Theorem 3.2, only (16.12) needs to be proved. We begin the veriﬁcation of (16.12) by observing that (16.11) implies 1 (XA − A∗ X) = 2i 1 XA× − (A× )∗ X = 2i

1 (XBB ∗ X + C ∗ C), 2

(16.13)

1 (iXB + C ∗ )(C − iB ∗ X). 2

(16.14)

These two identities imply that XAx, x and XA× x, x are nonnegative for all x ∈ Cn . In other words, both A and A× are X-dissipative, that is, they are

16.5. A generalization to pseudo-canonical factorization

309

dissipative in the indeﬁnite inner product given by X (cf., Section 11.3). Because of × this property, it follows that M+ and M+ are maximal X-nonnegative, while M− × and M− are maximal X-nonpositive (see Section 11.3). Using Proposition 11.1 it × × follows that dim M+ + dim M− = n and dim M− + dim M+ = n. Thus (16.12) is × = obtained via a dimension argument as soon as we have shown that M+ ∩ M− × M− ∩ M+ = {0}. × Take x ∈ M+ ∩ M− . Then Xx, x = 0, as x belongs to both an Xnonnegative subspace and an X-nonpositive subspace. Now the Cauchy-Schwartz inequality holds on M+ . Thus |XAx, x|2 ≤ XAx, AxXx, x = 0, and

|Xx, Ax|2 ≤ XAx, AxXx, x = 0.

From (16.13) we get 0 =

1 1 (XA − A∗ X)x, x = (Cx2 + B ∗ Xx2 ). 2i 2

× we have A× x = (A − BC)x = Ax. Hence Cx = 0, and so for x ∈ M+ ∩ M− × Consequently M+ ∩ M− is both A-invariant and A× -invariant. As

σ(A|M+ ∩M × ) ⊂ σ(A|M+ ) ⊂ {λ | λ > 0}, −

σ(A× |M+ ∩M × ) ⊂ σ(A× |M × ) ⊂ {λ | λ < 0}, −

−

× and A|M+ ∩M × = A |M+ ∩M × , we have that M+ ∩ M− = {0}. In a similar way one ×

−

−

× = {0}. shows that M− ∩ M+

Note that the above theorem together with the arguments given in the ﬁrst two paragraphs of this section yield the following corollary. Corollary 16.8. Let V (λ) = Im + W (λ), where W is a strictly proper rational matrix function which is strictly contractive on the real line. Then V admits both a right and a left canonical factorization with respect to the real line.

16.5 A generalization to pseudo-canonical factorization In this section the result of the previous section is generalized to pseudo-canonical factorizations. As a preparation we recall from Theorem 11.6 the following facts. Let X be an n × n invertible Hermitian matrix and let A be an n × n matrix which is X-dissipative. Then there exist A-invariant subspaces M+ and M− such that M+ is maximal X-nonnegative and M− is maximal X-nonpositive, σ(A|M+ ) ⊂ {λ | λ ≥ 0},

σ(A|M− ) ⊂ {λ | λ ≤ 0},

310

Chapter 16. Contractive rational matrix functions

M+ contains the spectral subspace of A corresponding to the eigenvalues of A in the open upper half plane, and M− contains the spectral subspace of A corresponding to the eigenvalues of A in the open lower half plane. These facts allow us to deal with rational matrix functions that are contractive on the real line. A rational matrix function W is called contractive on the real line if the values that W takes on R are contractive matrices. Such a function does not have a pole on the real line. Theorem 16.9. Let W be a strictly proper rational m × m matrix function which is contractive on the real line. Then V (λ) = Im + W (λ) admits both a right and a left pseudo-canonical factorization with respect to the real line. Such factorizations can be obtained as follows. Let W (λ) = C(λIn − A)−1 B be a minimal realization. Then there exists an invertible Hermitian matrix X satisfying iXA − iA∗ X + XBB ∗ X + C ∗ C = 0.

(16.15)

× Let M− and M− be maximal X-nonpositive subspaces that are invariant under A and A× , respectively, such that

σ(A|M− ) ⊂ {λ | λ ≤ 0},

σ(A× |M × ) ⊂ {λ | λ ≤ 0}, −

× and let M+ and M+ be maximal X-nonnegative subspaces that are invariant under × A and A , respectively, such that

σ(A|M+ ) ⊂ {λ | λ ≥ 0},

σ(A× |M × ) ⊂ {λ | λ ≥ 0}. +

× × ˙ M+ ˙ M− Then (16.12) holds, that is Cn = M− + and Cn = M+ + . Let Πl be the × projection along M− onto M+ , and put

V+ (λ)

= Im + C(λIn − A)−1 (In − Πl )B,

V− (λ)

= Im + CΠl (λIn − A)−1 B.

Then V (λ) = V+ (λ)V− (λ) is a left pseudo-canonical factorization with respect to × the real line. Write Πr for the projection along M+ onto M− , and set V− (λ)

=

Im + C(λ − A)−1 (In − Πr )B,

V+ (λ)

=

Im + CΠr (λIn − A)−1 B.

Then V (λ) = V− (λ)V+ (λ) is a right pseudo-canonical factorization with respect to the real line . Proof. By applying Theorem 16.1 (ii) to W (λ) = C(iλIn − A)−1 B we see that there is an invertible Hermitian X such that (16.15) holds. Once (16.12) is proved the rest of the theorem is a consequence of Theorem 8.5. Of the two equalities in (16.12) only the ﬁrst will be proved, the second can be established analogously.

16.5. A generalization to pseudo-canonical factorization

311

× As M+ is maximal X-nonnegative and M− is maximal X-nonpositive we × have dim M+ + dim M− = n, by Proposition 11.1. So it remains to show that × × M + ∩ M− = {0}. Take x ∈ M+ ∩ M− . As in the proof of Theorem 16.7, one shows × × that Cx = 0, and thus Ax = A x. Obviously, it follows from this that M+ ∩ M− is A-invariant and contained in Ker C. Because of the minimality, we can conclude × that M+ ∩ M− = {0}.

Note that the location of the spectra of the operators A|M− , A× |M × , A|M+ −

and A× |M × do not play a role in the proof of the identities in (16.12). Thus we + also have the following result. Proposition 16.10. Let V (λ) = Im + C(λIn − A)−1B be a minimal realization, and let X be an invertible Hermitian solution of (16.15). Let M be any A-invariant maximal X-nonnegative subspace, and let M × be any A× -invariant maximal X˙ M × . Let Π be the projection along M onto nonpositive subspace. Then Cn = M + × M , and write V1 (λ)

= Im + C(λIn − A)−1 (I − Π)B,

V2 (λ)

= Im + CΠ(λIn − A)−1 B.

Then V (λ) = V1 (λ)V2 (λ) is a minimal factorization. A similar result holds for any A-invariant maximal X-nonpositive subspace M and any A× -invariant maximal X-nonnegative subspace M × . Notice that there are various similarities between the proofs of Theorems 16.7 and 16.9 on the one hand and those of Theorems 15.3 and 15.6 on the other hand. These similarities are not surprising. In fact, the main results of the previous two sections are closely related to those in Sections 15.2 and 15.3 of the previous chapter. To see this we use the Cayley transformation −1 F (λ) = I − W (λ) I + W (λ) .

(16.16)

Here are the details. Let W be a strictly proper rational m × m matrix function, and let F be the rational m × m matrix function given by (16.16). Since W is strictly proper, I + W (λ) is biproper, and hence F is well-deﬁned. Furthermore, F is biproper and its value at inﬁnity is equal to Im . The identity ¯ ∗ −1 Im − W (λ) ¯ ∗ W (λ) Im + W (λ) −1 ¯ ∗ = 2 Im + W (λ) F (λ) + F (λ) shows that F has a nonnegative real part on R if and only if W is contractive on R. Moreover, F has a positive deﬁnite real part on R if and only if W is strictly contractive on R.

312

Chapter 16. Contractive rational matrix functions

Assume now that W is given by the realization W (λ) = C(λIn − A)−1 B. −1 −1 = 2(Im + W (λ) − Im , we Since F (λ) = 2Im − (Im + W (λ) (Im + W (λ) see that F admits the realization F (λ) = Im − 2C(λIn − A× )−1 B,

(16.17)

where, as usual, A× = A−BC. Now apply Theorem 15.1 to F using the realization (16.17). For this case equation (15.3), with X replaced by Y , has the form 1 −i(A× )∗ Y + iY A× − (Y B + i2C ∗ )(B ∗ Y − 2iC) = 0. 2

(16.18)

Using A× = A − BC and setting Y = −2X, a straightforward computation shows that (16.18) is equivalent to iXA − iA∗ X + XBB ∗ X + C ∗ C = 0,

(16.19)

and the latter equation is precisely (16.11). By applying Theorem 15.1 to F and using the equivalence between (16.18) and (16.19) we obtain the following result. Proposition 16.11. Let W (λ) = C(λIn − A)−1 B, and assume that the pair (A, B) is controllable. Then W is contractive on the real line if and only if the equation (16.19) has a Hermitian solution. Moreover, if the given realization is minimal, then any Hermitian solution of (16.19) is invertible. The above proposition provides an alternative proof of Theorem 16.1(ii) for the case when W is square and D = 0. The details involve a transformation λ → iλ (cf., the beginning of the proof of Theorem 16.9).

Notes The state space characterizations of contractive and strictly contractive rational matrix functions given in Theorems 16.1, 16.3 and 16.4 are versions of what is known as the bounded real lemma in mathematical systems theory. These results play an important role in robust and optimal control theory, see, e.g., the text books [77] and [150]. The bounded real lemma may also be found in [4] in another form. The application to spectral factorization (Section 16.3) is classical and can be found in Chapter 7 in [4]. The result that a function of the form identity plus a strict contraction admits canonical factorization (Section 16.4) is well-known; see e.g., [29] and the references given there. A surprising fact is that this property actually characterizes the circle or the line; for this see [109]. The state space results given in Sections 16.4 and 16.5 are based on [74].

Chapter 17

J-unitary rational matrix functions In this chapter realizations are used to study rational matrix functions of which the values on the imaginary axis are J-unitary matrices. Solutions to various factorization problems are given. Special attention is paid to factorization of J-unitary rational matrix functions into J-unitary factors. We also discuss the problem of embedding a contractive rational matrix function as the (1, 2) block in a unitary rational matrix function. The latter problem is related to the Darlington synthesis problem from network theory. This chapter consists of eight sections. Realization and minimal factorization of J-unitary rational matrix functions are the main topics of Sections 17.1 and 17.2. In Section 17.3 the factorization results are speciﬁed further for unitary rational matrix functions. The Redheﬀer transform, which allows one to relate J-unitary rational matrix functions to certain classes of unitary rational matrix functions, is introduced in Section 17.4. This transform is used in Section 17.5 in the study of Jinner rational matrix functions. A state space analysis of inner-outer factorization is the main topic of Section 17.6. The ﬁnal two sections deal with completion problems. Section 17.7 presents state space formulas for unitary completions of minimal degree, and Section 17.8 presents such formulas for bi-inner completions of non-square inner rational matrix functions.

17.1 Realizations of J-unitary rational matrix functions Throughout this section, J stands for an m × m signature matrix, that is, J is an invertible Hermitian matrix such that J = J −1 . An m × m matrix M is said to be J-unitary if M ∗ JM = J. Since all matrices in the latter identity are square and J is invertible, it follows that a J-unitary matrix M is invertible and M −1 = JM ∗ J.

314

Chapter 17. J-unitary rational matrix functions

If M is a J-unitary matrix, then M ∗ and M −1 are both J-unitary too. Indeed, M JM ∗ = (M ∗ J)−1 M ∗ = J −1 (M ∗ )−1 M ∗ = J −1 = J, M −1)∗ JM −1 = (JM ∗ J)∗ J(JM ∗ J) = J(M JM ∗ )J = J. In this chapter we deal with rational matrix functions of which the values on the imaginary axis are J-unitary matrices. A rational m × m matrix function W is called J-unitary on the imaginary axis if it takes J-unitary values on the imaginary axis. In other words, W is J-unitary with respect to the imaginary axis whenever W (λ)∗ JW (λ) = J, λ ∈ iR, λ not a pole of W . (17.1) Equivalently, W is J-unitary with respect to the imaginary axis if and only if ¯ ∗ JW (λ) = J, W (−λ)

¯ not a pole of W . λ, −λ

(17.2)

In the sequel we shall only consider matrix functions that are J-unitary with respect to the imaginary axis and not with respect to other contours. Therefore we shall feel free to omit the phrase “with respect to the imaginary axis.” ¯ ∗ Observe that if W is J-unitary, then both the functions W (λ)−1 and W (−λ) are J-unitary as well. Furthermore, if W1 and W2 are two J-unitary rational matrix functions, their product W1 W2 will also be J-unitary. First we shall characterize the property of being a J-unitary rational matrix function in terms of realizations. We shall assume throughout that the rational matrix functions are proper. Theorem 17.1. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a proper rational m × m matrix function. The following statements are equivalent: (i) W is J-unitary; (ii) D is J-unitary and there exists an n × n matrix H such that AH + HA∗ = BJB ∗ ,

CH = DJB ∗ ,

H = H ∗;

(17.3)

(iii) D is J-unitary and there exists an n × n matrix G such that GA + A∗ G = C ∗ JC,

GB = C ∗ JD,

G = G∗ .

(17.4)

In this case the matrices H and G are uniquely determined by the given realization, they are invertible and G = H −1 . Proof. Assume that W is J-unitary. Taking the limit in (17.1) for λ → ∞ we see that D∗ JD = J. Thus D is a J-unitary matrix. In particular, D is invertible, ¯ ∗ )−1 J = W (λ) for all λ for and hence W is biproper. By (17.2) we have J(W (−λ) ¯ which λ is not a pole of W and −λ is not a zero of W . Now one computes that ¯ −∗ J = JD−∗ J + JD−∗ B ∗ λIn − (−A∗ + C ∗ D−∗ B ∗ ) −1 C ∗ D−∗ J. JW (−λ)

17.1. Realizations of J-unitary rational matrix functions

315

The fact that the realization is minimal yields, by the state space similarity theorem, the existence of a unique (invertible) n × n matrix H such that AH = −HA∗ + HC ∗ D−∗ B ∗ ,

B = HC ∗ D−∗ J,

JD −∗ B ∗ = CH. (17.5)

Next, take adjoints and use D∗ JD = J to see that (17.5) also holds with H ∗ in place of H. By uniqueness it follows that H = H ∗ . Hence (17.3) holds, and so (i) implies (ii), even with the additional condition that H is invertible. Next assume D is J-unitary and there exists an n × n matrix H such that (17.3) holds. A straightforward computation gives ¯ ∗ = I + C(λ − A)−1 BD−1 DJD∗ I − D−∗ B ∗ (λ + A∗ )−1 C ∗ W (λ)JW (−λ) =

J + C(λ − A)−1 BJD ∗ − DJB ∗ (λ + A∗ )−1 C ∗ −C(λ − A)−1 BJB ∗ (λ + A∗ )−1 C ∗

=

J + C(λ − A)−1 HC ∗ − CH(λ + A∗ )−1 C ∗ −C(λ − A)−1 (H(λ + A∗ ) − (λ − A)H)(λ + A∗ )−1 C ∗

=

J + C(λ − A)−1 HC ∗ − CH(λ + A∗ )−1 C ∗ −C(λ − A)−1 HC ∗ + CH(λ + A∗ )−1 C ∗ = J.

Thus the function W (λ)∗ is J-unitary. But then so is W . We have now proved that (i) and (ii) are equivalent. The equivalence of (i) and (iii) can be established in the same way. Actually the implication (iii) ⇒ (i) can be obtained directly from (17.4) without having to take recourse to the function W (λ)∗ . As above, (i) implies the stronger version of (iii) with the extra requirement that G is invertible. The uniqueness and invertibility of H and G follow from the minimality. The invertibility can also be proved directly, and in fact from slightly weaker conditions. Assume (17.3) holds and that the pair (A, B) is controllable. Then H is invertible. Indeed, assume Hx = 0. Then DJB ∗ x = CHx = 0, so B ∗ x = 0. Hence (AH + HA∗ )x = 0 too. With Hx = 0, this gives HA∗ x = −AHx = 0. So Ker H ⊂ Ker B ∗ and A∗ [Ker H] ⊂ Ker H. Thus Ker H ⊂ Ker (B ∗ |A∗ ) = {0}. So H is invertible. Likewise, one shows that if (17.4) is satisﬁed and the pair (C, A) is observable, then G is invertible. Finally, let H be as in (17.3), then (17.4) holds with H −1 in place of G. By uniqueness it follows that G = H −1 . In the argument for the implication (ii) ⇒ (i) given above, the minimality of the given realization does not play a role. Similarly the minimality condition is irrelevant for the implication (iii) ⇒ (i). This is also reﬂected by the following proposition.

316

Chapter 17. J-unitary rational matrix functions

Proposition 17.2. Let W (λ) = D + C(λIn − A)−1 B be a realization of a rational m × m matrix function. Assume D is J-unitary, and let H and G be given Hermitian n × n matrices. Consider the following four statements: (i) AH + HA∗ = BJB ∗ ,

CH = DJB ∗ ;

(ii) AH + HA∗ = HC ∗ JCH,

CH = DJB ∗ ;

(iii) GA + A∗ G = C ∗ JC,

GB = C ∗ JD;

(iv) GA + A∗ G = GBJB ∗ G,

GB = C ∗ JD.

Then (i) and (ii) are equivalent, and so are (iii) and (iv). Each of (i)–(iv) implies that W is J-unitary. Moreover, if (A, B) is controllable and (i) holds, then all four statements are equivalent and the realization is minimal. Likewise, if (C, A) is observable and (iii) holds, then again all four statements are equivalent and the realization is minimal. Proof. To see the equivalence of (i) and (ii), use D ∗ JD = J to see that BJB ∗ = HC ∗ JCH. In an analogous manner one sees that (iii) and (iv) are equivalent. For the case when the realization is minimal the fact that (i) and (iii) imply that W is J-unitary is covered by Theorem 17.1. The general case is proved using the type of arguments occurring the proof of Theorem 17.1. Now suppose that (A, B) is controllable, and that (i) holds. In the proof of Theorem 17.1 we have already shown that this implies that H is invertible. Taking G = H −1 it follows that (iii) is satisﬁed, and hence also (iv). Next, we show that in this case (C, A) is observable. Indeed, by induction one shows that H −1 Ker (C|A) ⊂ Ker (B ∗ |A∗ ) = {0}. Hence the realization is minimal. The equivalence of all four statements now follows from Theorem 17.1. The reasoning for ﬁnal statement of the theorem is similar. The next proposition shows that under certain additional conditions the ﬁrst identity in (i) of Proposition 17.2 implies the second identity in (i), and analogously for (i) replaced by (iii). Proposition 17.3. Let W (λ) = D + C(λIn − A)−1 B be a realization of a J-unitary rational m × m matrix function, and let H and G be n × n Hermitian matrices. The following two statements are true: (i) If the pair (A, B) is controllable and GA + A∗ G = C ∗ JC, then GB = C ∗ JD. (ii) If the pair (C, A) is observable and AH +HA∗ = BJB ∗ , then CH = DJB ∗ . Proof. We only prove the ﬁrst part of the proposition, the second part can be ¯ ∗ JW (λ) established analogously. Assume that W is J-unitary. Computing W (−λ) one sees that this is equivalent to

−B ∗

D∗ JC

$ λ−

−A∗

C ∗ JC

0

A

%−1

C ∗ JD B

= 0.

(17.6)

17.1. Realizations of J-unitary rational matrix functions

317

Now assume that GA + A∗ G = C ∗ JC. Using I G S= 0 I as a similarity transformation in the realization (17.6), we see that (17.6) is equivalent to $ %−1 ∗ −A∗ 0 C JD − GB −B ∗ D ∗ JC − B ∗ G λ− = 0. 0 A B But this identity, in turn, is equivalent to (D ∗ JC−B ∗ G)(λ−A)−1 B = 0, λ ∈ ρ(A). The fact that (A, B) is controllable now implies that GB = C ∗ JD. The Hermitian matrix H in Theorem 17.1(ii), which is uniquely determined by the conditions stated there, will be called the Hermitian matrix associated with the minimal realization W (λ) = D + C(λIn − A)−1 B. Our next concern is how the associated Hermitian matrix behaves under similarity transformation on the realization. Proposition 17.4. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function. Write H for the Hermitian matrix associated with this realization, and let S be an invertible n × n matrix. Then the Hermitian matrix associated with the minimal realization W (λ) = D + CS −1 (λIn − SAS −1 )−1 SB

(17.7)

is given by SHS ∗ . Proof. For the (minimal) realization (17.7), the matrix SHS ∗ satisﬁes the requirements of condition (ii) in Theorem 17.1. As a consequence of the above proposition the number of positive and the number of negative eigenvalues of the matrix H do not depend on the particular choice of the minimal realization of the function W . The number of positive eigenvalues of H will be denoted by π+ (W ). At the end of this section, in Proposition 17.10, it will be seen how to express π+ (W ) completely in terms of W itself rather than in terms of the associated Hermitian matrix H. The next two propositions describe how the associated Hermitian matrix behaves under the operations of inversion, taking adjoints, and multiplication. Proposition 17.5. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function. Write H for the Hermitian matrix associated with this realization and, as usual, A× for the matrix A−BD −1 C. Then the Hermitian matrices associated with the minimal realizations W (λ)−1

=

D−1 − D−1 C(λIn − A× )−1 BD−1 ,

¯ ∗ W (−λ)

=

D∗ − B ∗ (λIn + A∗ )−1 C ∗ ,

318

Chapter 17. J-unitary rational matrix functions

are −H and −H −1 , respectively. Proof. For the ﬁrst realization, use (17.3) and observe that A× (−H) + (−H)(A× )∗

= −AH − HA∗ + BD−1 CH + HC ∗ D−∗ B ∗ = −BJB ∗ + BD−1 DJB ∗ + BJD∗ D−∗ B ∗ = BJB ∗ = (BD−1 )J(D −∗ B ∗ ),

and −D −1 C(−H) = D−1 CH = JB ∗ = D−1 J(BD−1 )∗ . The claim for the second realization is straightforward from the fact that in Theorem 17.1, the matrix G is the inverse of H. Proposition 17.6. For j = 1, 2, let Wj (λ) = Dj + Cj (λInj − Aj )−1 Bj be a minimal realization of a J-unitary rational m×m matrix function Wj having as the Hermitian matrix associated to it Hj . Suppose W = W1 , W2 is a minimal factorization. Then W is a J-unitary rational matrix function, W (λ) = D1 D2 +

D1 C2

C1

$

λIn1 +n2 −

A1

B1 C2

0

A2

%−1

B1 D 2

B2

is a minimal realization of W , and the associated Hermitian matrix is the block diagonal matrix diag (H1 , H2 ). Proof. Applying (17.3) to both realizations, using also D2 JD2∗ = J, one sees that ∗ H1 0 A1 A1 B1 C2 H1 0 0 + 0 A2 0 H2 C2∗ B1∗ A∗2 0 H2 =

B1 JB1∗

B1 C2 H2

H2 C2∗ B1∗

B2 JB2∗

B1 D2 = J D2∗ B1∗ B2

B2∗ .

So the ﬁrst equality in (17.3) is satisﬁed for the product realization. Also, H1 0 C1 D1 C2 = C1 H1 D1 C2 H2 0 H2 = D1 JB1∗

D1 D2 JB2∗ = (D1 D2 )J D2∗ B1∗

B2∗ ,

and this proves the second equality of (17.3) for the product realization.

Next, we present a few examples. As before, J stands for an m × m signature matrix.

17.1. Realizations of J-unitary rational matrix functions

319

Example 17.7. Let R be an m × m matrix such that R∗ JR = JR, and let ω ∈ / iR. Then the rational m × m matrix function W given by λ−ω R λ+ω ¯ is J-unitary. To be more speciﬁc, let u be a vector in Cm such that u∗ Ju = Ju, u = 0, and take for R the rank 1 matrix W (λ) = Im − R +

R=

1 Juu∗ . u∗ Ju

Then R∗ JR = JR = R∗ J = (u∗ Ju)−1 uu∗ . (Note here that uu∗ is a rank 1 matrix, while u∗ Ju is just a scalar.) A minimal realization for W for this particular choice of R may be obtained by setting (ω + ω ¯) . ∗ u Ju The associated Hermitian matrix satisﬁes AH + HA∗ = BJB ∗ , which in this case becomes −(ω + ω ¯ )H = u∗ Ju. So H = −(u∗ Ju)(2ω)−1 . Example 17.8. Let α ∈ iR, n ∈ N, and let x ∈ Cm be a J-neutral vector, i.e., x∗ Jx = 0. Then i W (λ) = Im + Jxx∗ (λ − α)2n is J-unitary. A minimal realization for W can be obtained by setting A = J2n (α), the Jordan block of size 2n with eigenvalue α, and ⎡ ⎤ 0 ⎢ .. ⎥ ⎢ ⎥ C = i Jx 0 · · · 0 , B = ⎢ . ⎥, ⎣0⎦ x∗ A = −¯ ω,

B = u∗ ,

C=−

where C is an m × 2n matrix and B is a 2n × m matrix. The associated Hermitian matrix can be computed to be the following matrix: 0 if p + q = 2n + 1, 2n H = [hp q ]p,q=1 , hp q = (−1)q i if p + q = 2n + 1. We conclude this section with a few remarks on matrix-valued kernel functions and their state space representations. Introduce the functions KW (λ, μ)

=

J − W (λ)JW (μ)∗ , λ+μ ¯

K∗,W (μ, λ)

=

J − W (μ)∗ JW (λ) . λ+μ ¯

Here W is a rational m × m matrix function. Furthermore, λ and μ are complex numbers, not poles of W , λ = −μ.

320

Chapter 17. J-unitary rational matrix functions

Lemma 17.9. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational matrix function having H as its associated Hermitian matrix. Then the following two identities hold: KW (λ, μ)

=

−C(λ − A)−1 H −1 (¯ μ − A∗ )−1 C ∗ ,

(17.8)

K∗,W (μ, λ)

=

−B ∗ (¯ μ − A∗ )−1 H −1 (λ − A)−1 B.

(17.9)

Proof. We shall only prove (17.9); identity (17.8) can be obtained in an analogous fashion. First note that W (μ)∗ JW (λ) = D∗ + B ∗ (¯ μ − A∗ )−1 C ∗ J D + C(λ − A)−1 B μ − A∗ )−1 C ∗ JD + D ∗ JC(λ − A)−1 B = D∗ JD + B ∗ (¯ μ − A∗ )−1 C ∗ JC(λ − A)−1 B. + B ∗ (¯ Now use the identities D∗ JD = J, C ∗ JD = HB and C ∗ JC = H −1 A + A∗ H −1 which hold by Theorem 17.1. Then one sees that ¯)B ∗ (¯ μ − A∗ )−1 H −1 (λ − A)−1 B. W (μ)∗ JW (λ) = J + (λ + μ From this (17.9) is immediate.

The kernel function KW (λ, μ) is said to have κ negative squares if for each r ∈ N and any collection of points ω1 , . . . , ωr in the complex plane, not poles of W , and any collection of vectors u1 , . . . , ur in Cm the r × r Hermitian matrix ∗ r (17.10) uj KW (ωj , ωi )ui i,j=1 has at most κ negative eigenvalues, and it has exactly κ negative eigenvalues for at least one choice of r, ω1 , . . . , ωr and u1 , . . . , ur . For K∗,W (μ, λ), the deﬁnition is of course similar. Proposition 17.10. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Then the number of negative squares of each of the functions KW and K∗,W is equal to π+ (W ), the number of positive eigenvalues of the matrix H. This result corroborates the already established fact that the integer π+ (W ) is independent of the particular minimal realization of W (cf., the paragraph after the proof of Proposition 17.4). Proof. It follows from the previous lemma that K∗,W has at most π+ (W ) negative squares. Indeed, if ω1 , . . . , ωr is a collection of points in the complex plane, not poles of W , and u1 , . . . , ur is a collection of vectors in Cm , then the r×r Hermitian matrix (17.10) can be written in the form −E ∗ H −1 E, where H is the Hermitian matrix associated with the given realization of W .

17.2. Factorization of J-unitary rational matrix functions

321

Next, consider M = span {(λ − A)−1 Bu | u ∈ Cm , λ ∈ C not an eigenvalue of A}. Clearly, for u ∈ Cm and λ not an eigenvalue of A, the vector λ(λ−A)−1 Bu belongs to M . Since M is closed in Cn , this implies that Bu = lim λ(λ − A)−1 Bu ∈ M, λ→∞

u ∈ Cm .

Thus Im B ⊂ M . Next, note that A(λ − A)−1 Bu = −Bu + λ(λ − A)−1 Bu ∈ M . Hence M is invariant under A. But then Im (A|B) ⊂ M . By hypothesis, the given realization of W is minimal. This implies that Im (A|B) = Cn . We conclude that M = Cn . The latter implies that Cn has a basis x1 , . . . , xn such that for each j the vector xj is of the form xj = (λi − A)−1 Buj for some vector uj ∈ Cm and some ωj ∈ C. Consider the n × n matrix X = [ x1 · · · xn ]. We obtain that for these uj and ωi we have n ∗ uj K∗,W (ωj , ωi )ui i,j=1 = −X ∗ H −1 X. As X is invertible, this matrix has exactly π+ (W ) negative eigenvalues. This settles the matter for K∗,W ; for KW the argument is similar.

17.2 Factorization of J-unitary rational matrix functions In this section minimal factorizations of J-unitary rational matrix functions into a product of two J-unitary rational matrix functions will be studied. Here, as in the previous section, J is an m × m signature matrix. To state the main theorem we need to recall a notion introduced in Section 11.1. Let H = H ∗ be an invertible n×n matrix. A subspace M ⊂ Cn is called H-nondegenerate if M ∩ [HM ]⊥ = {0}. ˙ [HM ]⊥ = Cn , as a simple dimension count shows. For such a subspace one has M + Also note that (HM )⊥ = H −1 [M ⊥ ]. Theorem 17.11. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Let M be an A-invariant H −1 -nondegenerate subspace, and denote by Π the projection of Cn onto H[M ⊥ ] along M . Let D = D1 D2 be a factorization of D into two J-unitary constant matrices, and put W1 (λ)

=

D1 + C(λIn − A)−1 (I − Π)BD2−1 ,

W2 (λ)

=

D2 + D1−1 CΠ(λIn − A)−1 B.

Then W = W1 W2 , this factorization is minimal, and the factors W1 and W2 are J-unitary. Conversely, any minimal factorization W = W1 W2 with J-unitary

322

Chapter 17. J-unitary rational matrix functions

factors W1 and W2 is obtained in this way. Moreover, given a ﬁxed factorization D = D1 D2 , the correspondence between minimal factorizations of W with two J-unitary factors and H-nondegenerate invariant subspaces of A is one-to-one. Proof. From (17.3) we know that A× = −HA∗ H −1 , where A× = A − BD−1 C. It follows that H[M ⊥ [ is A× -invariant because M is A-invariant. Since the subspace M is H −1 -nondegenerate, the projection Π is a supporting projection. Hence the factorization W = W1 W2 is a minimal one. To complete the proof of the ﬁrst part of the theorem it remains to show that the factors W1 and W2 are J-unitary rational matrix functions. In fact, it suﬃces to show that one of them is J-unitary, the J-unitarity of the other one then follows automatically. Since Π is a supporting projection we know that a minimal realization, W1 (λ) = D1 + C1 (λ − A1 )−1 B1 , of W1 is obtained by taking A1

=

τM ∗ AτM : M → M,

B1

=

τM ∗ (I − Π)BD2−1 : Cm → M,

C1

=

CτM : M → Cm .

Here τM is the canonical embedding of M into Cn , and hence τM ∗ τM is the orthogonal projection of Cn onto M . Put G1 = τM ∗ H −1 τM . Then G1 is invertible. Indeed, suppose G1 x = 0 for some x ∈ M . Then H −1 x ∈ Ker τM ∗ = M ⊥ , i.e., x ∈ H(M ⊥ ). So x ∈ M ∩ H(M ⊥ ) = {0}. Next, we shall show that the conditions of Theorem 17.1 (iii) are satisﬁed. First, note that (G1 A1 + A∗1 G1 ) =

τM ∗ H −1 τM τM ∗ AτM + τM ∗ AτM τM ∗ H −1 τM

=

τM ∗ (H −1 A + A∗ H −1 )τM

=

τM ∗ C ∗ JCτM ∗ = C1∗ JC1 .

Furthermore, we have G1 B1 = τM ∗ H −1 τM τM ∗ (I − Π)BD2−1 = τM ∗ H −1 (I − Π)BD2−1 . Now, as M is H −1 -nondegenerate, Im H −1 Π = M ⊥ and H −1 [M ]⊥ = H[M ⊥ ] = Im Π. This yields H −1 Πx, y = H −1 Πx, Πy = Πx, H −1 Πy = x, H −1 Πy. Hence Π∗ H −1 = H −1 Π, that is, the projection Π is H −1 -selfadjoint. Therefore H −1 (I − Π) = (I − Π∗ )H −1 . Moreover, as (I − Π)τM = τM we have the identity τM ∗ (I − Π∗ ) = τM ∗ . Thus G1 B1

=

τM ∗ H −1 (I − Π)BD2−1

=

τM ∗ (I − Π∗ )H −1 BD2−1

=

τM ∗ H −1 BD2−1 = τM ∗ C ∗ JD1 = C1∗ JD1 .

17.2. Factorization of J-unitary rational matrix functions

323

Hence the conditions of Theorem 17.1 (iii) are satisﬁed, and thus W1 is J-unitary. The converse statement is a direct consequence of Proposition 17.6 and Theorem 8.5. As a special case of the preceding theorem we state the following proposition concerning the case where one of the factors is of degree 1. Proposition 17.12. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Suppose x is an eigenvector of A corresponding to the eigenvalue ω of A, and assume H −1 x, x = 0. Then W admits a minimal factorization W = W1 W2 into two J-unitary factors where the factor W1 is given by 1 W1 (λ) = Im + Cxx∗ C ∗ J . (17.11) (λ − ω)H −1 x, x Furthermore, in case ω ∈ / iR the scalar x∗ C ∗ JCx is non-zero and λ+ω ¯ 1 1− Cxx∗ C ∗ J. W1 (λ) = Im − ∗ ∗ x C JCx λ−ω

(17.12)

Observe that the factor W1 is of the form as given in Example 17.7 Proof. As H −1 x, x = 0, the subspace M = span {x} is H −1 -nondegenerate. Therefore we can apply the previous theorem. The projection I − Π is given by (I − Π)v =

H −1 v, x x∗ H −1 v x = x. H −1 x, x x∗ H −1 x

Taking D1 = I and D2 = D one obtains W1 (λ)

= I + C(λ − A|M )−1 (I − Π)BD−1 = I+

Cxx∗ H −1 BD−1 (λ − ω)H −1 x, x

= I+

Cxx∗ C ∗ J . (λ − ω)H −1 x, x

This proves (17.11). Next we apply (17.4) in the present setting. Recall that G = H −1 . It follows that x∗ C ∗ JCx = (ω + ω ¯ )x∗ H −1 x. Thus, when ω ∈ / iR or, equivalently, ω + ω ¯ = 0, x∗ C ∗ JCx = 0,

H −1 x, x =

x∗ C ∗ JCx . ω+ω ¯

Employing this in (17.11) immediately yields (17.12).

324

Chapter 17. J-unitary rational matrix functions

17.3 Factorization of unitary rational matrix functions In this section we shall consider the special case of rational matrix functions that are unitary on the imaginary axis, that is, we continue the theme of the previous section with J = I. For simplicity, we call such functions unitary rational matrix functions and omit the additional qualiﬁer “on the imaginary axis.” Let W be a unitary rational matrix function. Then W is bounded by 1 on the imaginary axis, and hence W cannot have pure imaginary poles. Since W −1 is also a unitary rational matrix function, W cannot have pure imaginary zeros either. Replacing λ by λ−1 one also sees that W has to be biproper. Lemma 17.13. Let W (λ) = D + C(λ− A)−1 B be a minimal realization of a unitary rational m×m matrix function, and let H be the Hermitian matrix associated with this realization. Then A has no pure imaginary eigenvalues. Let P be the spectral projection of A corresponding to the part of σ(A) lying in the open right half plane. Then Im P is maximal H −1 -positive and Ker P is maximal H −1 -negative. Proof. Since the realization is minimal and W has no poles on the imaginary axis, the matrix A has no pure imaginary eigenvalues. By Theorem 17.1 with G = H −1 we have GA + A∗ G = C ∗ C. Because of the minimality of the realization we also know that the pair (C, A) is observable. Let us denote by ν(G) the number of negative eigenvalues of G, and by π(G) the number of positive eigenvalues of G. By a well-known inertia theorem (see Theorem 13.1.4 in [107]) we have ν(G) = dim Ker P and π(G) = dim Im P . Now put M = Im P , let τM be the canonical embedding of M into Cn , and introduce AM = τM ∗ AτM , GM = τM ∗ GτM and CM = CτM . Then GM is Hermitian, and (using the fact that M is invariant under A) we have GM AM + A∗M GM

=

τM ∗ GτM τM ∗ AτM + τM ∗ A∗ τM τM ∗ GτM

=

∗ τM ∗ (GA + A∗ G)τM = τM ∗ C ∗ CτM = CM CM .

The invariance of M under A also implies that Ker (CM |AM ) ⊂ Ker (C|A), and hence (CM , AM ) is an observable pair too. Moreover, AM has only eigenvalues in the open right half plane. The inertia theorem referred to above then gives that GM is positive deﬁnite. But this is equivalent to saying that Im P is H −1 -positive. As π(H −1 ) = dim Im P , it is actually maximal H −1 -positive. The other part of the proposition is proved in a similar way. Observe that an H −1 -positive subspace is in particular H −1 -nondegenerate . Likewise, an H −1 -negative subspace is H −1 -nondegenerate. So we are in a position to apply Theorem 17.11. This yields the following two results of which we shall only prove the second. Theorem 17.14. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a unitary rational m × m matrix function (so, in particular, D is invertible), and

17.3. Factorization of unitary rational matrix functions

325

let A× = A − BD−1 C be the associate main operator. Then W admits a minimal factorization W = W1 W2 having the following additional properties: (i) W1 has its poles in the left half plane and its zeros in the right half plane, (ii) W2 has its poles in the right half plane and its zeros in the left half plane, (iii) δ(W1 ) = n − π+ (W ) and δ(W2 ) = π+ (W ). Such a factorization can be obtained as follows. Let P denote the spectral projection corresponding to the part of σ(A) lying in the open left half plane, and write P × for the spectral projection of A× corresponding to the part of σ(A× ) lying in the ˙ Ker P × and the functions open right half plane. Then Cn = Im P + W1 (λ)

=

Im + C(λIn − A)−1 (In − Π)BD−1 ,

W2 (λ)

=

D + CΠ(λIn − A)−1 B,

(17.13)

meet the requirements. Here Π is the projection of Cn along Im P onto Ker P × . Theorem 17.15. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a unitary rational m × m matrix function (so, in particular, D is invertible), and let A× = A − BD−1 C be the associate main operator. Then W admits a minimal factorization W = W1 W2 having the following additional properties: (i) W1 has its poles in the right half plane and its zeros in the left half plane, (ii) W2 has its poles in the left half plane and its zeros in the right half plane, (iii) δ(W1 ) = π+ (W ) and δ(W2 ) = n − π+ (W ). Such a factorization can be obtained as follows. Let P denote the spectral projection corresponding to the part of σ(A) lying in the open right half plane, and write P × for the spectral projection of A× corresponding to the part of σ(A× ) lying in the ˙ Ker P × and the functions open left half plane. Then Cn = Im P + W1 (λ)

=

Im + C(λIn − A)−1 (In − Π)BD−1 ,

W2 (λ)

=

D + CΠ(λIn − A)−1 B,

(17.14)

meet the requirements. Here Π is the projection of Cn along Im P onto Ker P × . Proof. With Lemma 17.13 in mind, the idea is to apply Theorem 17.11 taking M = Im P . We need to ﬁnd H[Im P ]⊥ = Ker (P ∗ H −1 ). From (17.3) we know that A× = −HA∗ H −1 , hence P × = −HP ∗ H −1 . It follows that H[Im P ]⊥ = Im P × . Let Π be the projection along Im P onto Im P × . Then, by Theorem 17.11, the function W admits the factorization W = W1 W2 , where W1 and W2 are given by (17.14), and these factors are unitary. Moreover, the factorization is minimal. Finally, the poles of W1 are the eigenvalues of A|Im P (counting multiplicities), its zeros are the eigenvalues of A× |Im P × (counting multiplicities too). Similarly, the poles of W2 are the eigenvalues of A|Ker P , while the

326

Chapter 17. J-unitary rational matrix functions

zeros of W2 are the eigenvalues of A× |Ker P × . So the position of poles and zeros of W1 and W2 is as required. It also follows that δ(W1 ) = dim Ker Π = dim Im P = π+ (W ), and hence by minimality also δ(W2 ) = n − π+ (W ).

Our next theorem is on complete factorization of a unitary rational matrix function into unitary factors (cf., Part III in [20]). Theorem 17.16. Let W be a unitary rational m × m matrix function of McMillan degree n. Then W admits a minimal factorization into n factors of McMillan degree 1. Moreover, each of these factors can be taken to be unitary. In order to prove this theorem we ﬁrst show that a unitary rational matrix function allows for a realization with very special properties. Lemma 17.17. Let W be a unitary rational m × m matrix function with W (∞) = Im . Then W admits a minimal realization W (λ) = Im + C(λIn −A)−1 B such that A11 A12 A= , (17.15) 0 A22 where A11 and A22 are upper triangular, A11 has all its eigenvalues in the open right half plane, A22 has all its eigenvalues in the open left half plane, and the Hermitian matrix associated with the realization is given by In 0 H= . (17.16) 0 −In Proof. Take an arbitrary minimal realization W (λ) = I +C(λ−A)−1 B. By Schur’s theorem there is an orthogonal change of basis such that A is upper triangular. In fact, we may take the eigenvalues of A on the diagonal in any order we like. This is known as the ordered Schur form of A. We apply this to construct a similarity transformation such that A is of the form A11 A12 A= 0 A22 where A11 is upper triangular having all its eigenvalues in the open right half plane, and A22 is upper triangular having all its eigenvalues in the open left half plane. The spectral projection of A corresponding to its eigenvalues in the open right half plane is given by I 0 P = . 0 0

17.3. Factorization of unitary rational matrix functions

327

Let H be the Hermitian matrix associated with this realization, and let G be its inverse. Decompose G in the same way as A, and write G11 G12 . G= G∗12 G22 Because of Lemma 17.13 we have that Im P is maximal G-positive, and so G11 is positive deﬁnite. Likewise, since Ker P is maximal G-negative, G22 is negative deﬁnite. Next, we employ the Schur complement of G11 in G. So we factorize G as 0 I 0 G11 I G−1 11 G12 . G= G∗12 I 0 I 0 G22 − G∗12 G−1 11 G12 Since G11 is positive deﬁnite and G22 is negative deﬁnite, the Schur complement G22 − G∗12 G−1 11 G12 is negative deﬁnite too. ∗ Now take the Cholesky decomposition of G11 , that is, write G11 = C11 C11 with C11 upper triangular. Likewise, take the Cholesky decomposition of the Schur ∗ complement. Thus G22 − G∗12 G−1 11 G12 = −C22 C22 with C22 upper triangular. Put −1 −1 C11 −G−1 11 G12 C22 . S= −1 0 C22 Then, using Proposition 17.4, one checks that the realization W (λ) = I + CS(λ − S −1 AS)−1 S −1 B has all the desired properties.

Proof of Theorem 17.16. Without loss of generality we may assume that W has the value Im at inﬁnity. Let W (λ) = Im + C(λIn − A)−1 B be a minimal realization as in the previous lemma, and let H be the Hermitian matrix associated with this realization. In particular, A is upper triangular. For this realization we have by (17.3) that A× = −HA∗ H −1 . This is clearly a lower triangular matrix. Now let e1 , . . . , en be the standard basis of Cn . For k = 1, . . . , n, deﬁne Πk to be the orthogonal projection of Cn onto span {ek }. Then for j = 1, . . . , n − 1 the projection Πj+1 + · · · + Πn is a supporting projection for the minimal realization W (λ) = Im + C(λIn − A)−1 B. It then follows from Theorem 10.5 in [20] that W admits a factorization into n factors of degree 1. It remains to prove that each of the factors is unitary. Clearly, for each integer j = 1, . . . , n − 1 the image and kernel of Πj+1 + · · · + Πn are both H −1 -nondegenerate and are each other’s H-orthogonal complements. From Theorem 17.11 it then follows that for each j the products W1 · · · Wj and Wj+1 · · · Wn are unitary. From this one concludes that each Wj separately is unitary.

328

Chapter 17. J-unitary rational matrix functions

17.4 Intermezzo on the Redheﬀer transformation In this section we study the Redheﬀer transform of a J-unitary rational matrix function. This will allow us to relate J-unitary rational matrix functions to certain classes of unitary rational matrix functions. The results obtained will be used in the next section. All the time, J will be a signature matrix. The starting point of our considerations is a 2 × 2 block matrix M11 M12 M= , (17.17) M21 M22 with M11 a p×p matrix and M22 a q ×q matrix. When M22 is an invertible matrix, the Redheﬀer transform Λ of M is deﬁned as follows: −1 −1 M21 M12 M22 Λ11 Λ12 M11 − M12 M22 Λ= = . (17.18) −1 −1 Λ21 Λ22 −M22 M21 M22 We refer to the map M → Λ as the Redheﬀer transformation. Let J = diag (Ip , −Iq ). The matrix M in (17.17) is said to be J-contractive if M ∗ JM ≤ J. The next lemma shows that for such a matrix the requirement that M22 is invertible is automatically fulﬁlled. Hence the Redheﬀer transform of a J-contractive matrix M with J = diag (Ip , −Iq ) is well-deﬁned. Lemma 17.18. Let J = diag (Ip , −Iq ). If the matrix M in (17.17) is J-contractive, then M22 is invertible, the (well-deﬁned) Redheﬀer transform Λ of M is a con−1 traction, and M22 M21 < 1. Conversely, if M22 is invertible and the Redheﬀer transform Λ of M is a contraction, then M is J-contractive. Proof. Assume that the matrix M is J-contractive. By considering the (2, 2)-entry of M ∗ JM and using M ∗ JM ≤ J, we see that ∗ ∗ M22 M22 ≥ Iq + M12 M12 .

(17.19)

∗ M22 is positive deﬁnite, and hence, because M22 is square, the matrix Thus M22 −∗ M22 is invertible. Multiplying the inequality (17.19) from the left by M22 and from −1 −∗ −1 −∗ −1 −∗ −1 ∗ the right by M22 , we get Iq − M22 M12 M12 M22 ≥ M22 M22 . Since M22 M22 −∗ −1 ∗ is positive deﬁnite, we may conclude that so is Iq − M22 M12 M12 M22 . But this is −1 equivalent to M22 M21 < 1. Next assume that M22 is invertible and consider the equations x u M11 M12 = . (17.20) y v M21 M22

Then, as M22 is invertible, these equations are equivalent to Λ11 Λ12 x u = . Λ21 Λ22 v y

(17.21)

17.4. Intermezzo on the Redheﬀer transformation

329

Indeed, rewrite (17.20) as M11 x + M12 y = u and M21 x + M22 y = v. Solving for y in the second of these equations, one gets −1 −1 M21 x + M22 v. y = −M22

(17.22)

Inserting this in the ﬁrst of the two equations above, we obtain −1 −1 u = (M11 − M12 M22 M21 )x + M12 M22 v.

(17.23)

Together, (17.22) and (17.23) prove the desired equivalence between (17.20) and (17.21). Notice that the condition that the matrix M is J-contractive is equivalent to the inequality u2 − v2 ≤ x2 − y2 . Indeed, M ∗ JM ≤ J is equivalent to ( ) ( ) ( ) ( ) x x u u 2 2 u − v = J (17.24) ,M = JM , y y v v = M ∗ JM

( ) ( ) ( ) ( ) x x x x = x2 − y2 . , ≤ J , y y y y

Similarly, the condition that the Redheﬀer transform Λ is a contraction is equivalent to u2 + y2 ≤ x2 + v2 . But u2 − v2 ≤ x2 − y2 ⇐⇒ u2 + y2 ≤ x2 + v2 . Thus, as desired, M is J-contractive amounts to the same as M22 is invertible and Λ is a contraction. Corollary 17.19. Let J = diag (Ip , −Iq ), and assume that the matrix M in (17.17) is J-contractive. Then M ∗ is J-contractive too. Proof. By Lemma 17.18, the fact that M is J-contractive implies that M22 is invertible and the Redheﬀer transform Λ of M is a contraction. Since M22 is ∗ invertible, so is M22 . Thus the Redheﬀer transform of M ∗ is well-deﬁned. Moreover, the Redheﬀer transform of M ∗ is equal to Λ∗ . As Λ is a contraction, the same holds true for Λ∗ . But then the converse part of Lemma 17.18 shows that M ∗ is J-contractive too. Proposition 17.20. Let J = diag (Ip , −Iq ). The matrix M in (17.17) is J-unitary if and only if M22 is invertible and the Redheﬀer transform of M is unitary. Proof. Since a J-unitary matrix is J-contractive and a unitary matrix is a contraction, we see from Lemma 17.18 that without loss of generality we may assume that the matrix M22 is invertible. This allows us to use the equivalence of the equations (17.20) and (17.21). Next, using a calculation as in (17.24), one sees that M is J-contractive if and only if the equality x2 − y2 = u2 − v2 holds. Furthermore, the condition that Λ is unitary is equivalent to x2 + v2 = u2 + y2 . But x2 − y2 = u2 − v2 ⇐⇒ x2 + v2 = u2 + y2 .

330

Chapter 17. J-unitary rational matrix functions

Hence M is J-unitary if and only if Λ is unitary.

Next we pass from matrices to matrix functions. Consider a rational matrix function W, W11 (λ) W12 (λ) W (λ) = , (17.25) W21 (λ) W22 (λ) with W11 a p×p rational matrix function and W22 a q ×q rational matrix function. −1 Assume W22 to be regular, i.e., det W22 (λ) ≡ 0. Thus W22 is a well-deﬁned rational matrix function. Under these assumptions the Redheﬀer transform of W is deﬁned to be the rational matrix function Σ given by Σ11 (λ) Σ12 (λ) Σ(λ) = (17.26) Σ21 (λ) Σ22 (λ) =

W11 (λ) − W12 (λ)W22 (λ)−1 W21 (λ)

W12 (λ)W22 (λ)−1

−W22 (λ)−1 W21 (λ)

W22 (λ)−1

.

As before, let J = diag (Ip , −Iq ). If the rational matrix function W is J-unitary with respect to the imaginary axis, then we know from Proposition 17.20 that the Redheﬀer transform Σ is unitary. In particular, it has no pure imaginary poles and zeros (see the second paragraph of Section 17.3). The following theorem is the main result of this section. Theorem 17.21. Let W be a rational matrix function, and let C1 D1 0 + (λIn − A)−1 B1 B2 W (λ) = C2 0 D2

(17.27)

−1 be a realization of W . Assume D2 is invertible, and put A× 2 = A − B2 D2 C2 . Then the Redheﬀer transform Σ of W has the realization C1 0 D1 −1 + (λIn − A× Σ(λ) = B1 B2 D2−1 , (17.28) 2) −1 −1 −D2 C2 0 D2

and this realization is minimal if and only if so is the realization (17.28). Moreover, assuming both realizations (17.27) and (17.28) to be minimal, the following holds. Let J = diag (Ip , −Iq ) and suppose W is J-unitary on the imaginary axis. If HW and HΣ denote the Hermitian matrices associated with the realizations (17.27) and (17.28), respectively, then HW = HΣ . Proof. Write W in the form (17.25). From Theorem 2.4 we have −1 B2 D2−1 , W22 (λ)−1 = D2−1 − D2−1 C2 (λ − A× 2)

17.4. Intermezzo on the Redheﬀer transformation

331

and with the help of this expression one computes −1 B2 D2−1 W12 (λ)W22 (λ)−1 = C1 (λ − A)−1 B2 D2−1 − D2−1 C2 (λ − A× 2 ) = W22 (λ)−1 W21 (λ)

=

−1 C1 (λ − A× B2 D2−1 , 2)

−1 D2−1 − D2−1 C2 (λ − A× B2 D2−1 C2 (λ − A)−1 B1 2)

−1 D2−1 C2 (λ − A× B1 . 2) Now W12 (λ)W22 (λ)−1 W21 (λ) = C1 (λ − A)−1 B2 W22 (λ)−1 W21 (λ), and hence

=

W11 (λ) − W12 (λ)W22 (λ)−1 W21 (λ)

−1 = D1 + C1 (λ − A)−1 B1 − C1 (λ − A)−1 B2 D2−1 C2 (λ − A× B1 2)

× −1 = D1 + C1 (λ − A)−1 B1 − C1 (λ − A)−1 (A − A× B1 2 )(λ − A ) −1 = D1 + C1 (λ − A× B1 . 2)

This proves (17.28). Next we deal with minimality. Assume the realization (17.27) is minimal. To prove the minimality of the realization (17.28), assume the realization (17.28) is not observable. Then ( ) C1 × Ker , A = {0}. 2 −D2−1 C2 Observe that the subspace on the left-hand side is invariant under A× 2 . Hence there exists an eigenvalue λ0 of A× and there is a non-zero vector x such that 2 −1 × A× x = λ x, and C x = 0, −D C x = 0. By the deﬁnition of A this implies 0 1 2 2 2 2 −1 × that Ax = A× 2 x − B2 D2 C2 x = A2 x = λ0 x. So ( ) C1 , A = {0}. Ker C2 Hence the realization (17.27) is not observable, which is a contradiction. It follows that the realization (17.28) is observable. A similar argument proves that the realization (17.28) is controllable. The reverse implication, minimality of (17.28) implies minimality of (17.27), is proved in an analogous way. Now assume both realizations are minimal. It remains to prove the equality of the corresponding Hermitian matrices. This is seen as follows. According to Theorem 17.1 the matrix HW is uniquely determined by the four expressions D1∗ D1 = Ip , D2∗ D2 = Iq and ∗ B C D 1 1 1 HW = . AHW + HW A∗ = B1 B1∗ − B2 B2∗ , C2 −D2 B2∗

332

Chapter 17. J-unitary rational matrix functions

Next, using the same theorem with Ip+q as the signature matrix, we know that HΣ is uniquely determined by the identities D1∗ D1 = Ip , D2∗ D2 = Iq and × ∗ A× 2 HΣ + HΣ (A2 )

C1 −D2−1 C2

=

B1 B1∗ + B2 D2−∗ D2−1 B2∗ ,

HΣ

=

(17.29)

D1 B1∗

.

D2−1 D2−∗ B2∗

∗ −1 Since D2∗ D2 = Iq and A× 2 = A − B2 C2 = A + B2 B2 HW , we obtain that the formulas for HΣ are satisﬁed by HW . Uniqueness of the associated Hermitian matrix proves then that HW = HΣ .

We ﬁnish this section by returning to the examples of Section 17.1. Consider, ∗ for J = diag (Ip , −Iq ), the function W of Example 17.7. So, taking u = u∗1 u∗2 , W (λ) = Ip+q

( ) −u1 + (λ + ω ¯ )−1 u∗1 u2

u∗2

2ω . u∗ Ju

Using Theorem 17.21 one ﬁnds, for Redheﬀer transform Σ of W , Σ(λ) = Ip+q −

1 2ω uu∗ , λ − α u∗ Ju

where

−¯ ωu1 2 − ωu2 2 2ω 2 = u . 2 u∗ Ju u∗ Ju For the Example 17.8, things are somewhat ∗ more complicated. We use the realization presented there, writing x = x∗1 x∗2 . The Redheﬀer transform of α = −¯ ω−

W (λ) = Ip+q

x1 i x∗1 + (λ − α)2n −x2

x∗2

then becomes ⎡

Σ(λ) = Ip+q + i x

where

0 ···

⎡ ⎢ ⎢ A× 2 = J2n (α) + ⎢ ⎣

0 .. . 0 ix2 2

⎤ 0 ⎢.⎥ −1 ⎢ .. ⎥ 0 (λ − A× ) ⎢ ⎥, 2 ⎣0⎦ x∗ 0

···

0

···

⎤ 0 .. ⎥ .⎥ ⎥. 0⎦ 0

17.5. J-inner rational matrix functions

333

−1 Since Σ only involves the entry of (λ − A× in the upper right corner, this 2) can be computed further. The entry in question is just 1 over the characteristic polynomial of A× 2 , and so

Σ(λ) = Ip+q +

i (λ −

α)2n

− ix2 2

xx∗ .

17.5 J-inner rational matrix functions A matrix M is called a J-contraction if M ∗ JM ≤ J. A rational matrix function W is called J-inner if W is J-unitary on the imaginary axis and, in addition, W (λ) is a J-contraction for λ in the open right half plane, λ not a pole of W . Note that we restrict the attention here to functions that are J-inner relative to the imaginary axis. If W is J-inner with J = I, then W is called bi-inner or two-sided inner (cf., Section 17.6 below). Clearly, if W is bi-inner it cannot have poles in the right open half plane. Also, if a unitary rational matrix W is analytic on the right half plane, then by the maximum modulus theorem W (λ) ≤ 1 for λ > 0, i.e., W is bi-inner. Thus a unitary rational matrix function W is bi-inner if and only if it is analytic on the right half plane. Recall from the second paragraph of Section 17.3 that a unitary rational matrix function has no pure imaginary poles or zeros, and that it is biproper. The next theorem characterizes the property of being J-inner in terms of a minimal realization. Theorem 17.22. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Then W is J-inner if and only if H is negative deﬁnite. First we state a result that is of independent interest, and which proves one direction of Theorem 17.22. Proposition 17.23. If W is a J-inner rational matrix function, where the signature matrix J has the form J = diag (Ip , −Iq ), then its Redheﬀer transform Σ is biinner. If, in addition, W is given by the minimal realization (17.27), then A× 2 = A − B2 D2−1 C2 has all its eigenvalues in the open left half plane, and the Hermitian matrix H associated with (17.27) is negative deﬁnite. Proof. The ﬁrst part of the proposition can be derived from Proposition 17.20 and Lemma 17.18. For the second part, consider a minimal realization of W written in the form (17.27) with the partitioning induced by J = diag (Ip , −Iq ). Then we also have a minimal realization (17.28) of Σ. Since Σ is bi-inner, it is analytic in the right half plane, and by minimality of the realization this shows that A× 2 has all its eigenvalues in the left half plane.

334

Chapter 17. J-unitary rational matrix functions

It follows from the fact that H satisﬁes the Lyapunov equation (17.29) and from minimality that H is negative deﬁnite (see Corollary 1 in Section 13.1 in [107]). Proof of Theorem 17.22. Assume H is negative deﬁnite. For λ > 0 we then have ¯ − A∗ )−1 C ∗ J D + C(λ − A)−1 B J − W (λ)∗ JW (λ) = J − D ∗ + B ∗ (λ ¯ − A∗ )−1 C ∗ JD − D ∗ JC(λ − A)−1 B = J − D∗ JD − B ∗ (λ ¯ − A∗ )−1 C ∗ JC(λ − A)−1 B. −B ∗ (λ Using the identities D∗ JD = J, C ∗ JD = HB and C ∗ JC = H −1 A + A∗ H −1 , which hold by Theorem 17.1, one sees that ¯ − A∗ )−1 H −1 (λ − A)−1 B ≥ 0. J − W (λ)∗ JW (λ) = −2(Re λ)B ∗ (λ Hence W is J-inner. Conversely, if W is J-inner, where J = diag (Ip , −Iq ), then H is negative deﬁnite by Proposition 17.23. So, it remains to show that we can reduce the general case to the situation where J is of the form J = diag (Ip , −Iq ). To this end, let T be an invertible matrix such that T ∗ JT = J1 = diag (Ip , −Iq ) for some nonnegative integers p and q. Such a T does exist. Observe that J = T −∗ J1 T −1 , and since J = J −1 , we obtain that J = T J1 T ∗ . Consider the matrix function W1 = T −1 W T . Then W1 is J1 -inner, and has a minimal realization W1 (λ) = T −1DT + T −1 C(λ − A)−1 BT. We claim that H is the Hermitian matrix associated with this minimal realization. Indeed, using J = T J1 T ∗ we have AH + HA∗ = BJB ∗ = BT J1 T ∗ B ∗ , T −1 CH = T −1 DJB ∗ = (T −1 DT )J1 T ∗ B ∗ . By Theorem 17.1, the matrix H is the Hermitian matrix associated with the given minimal realization of W1 . So we can apply Proposition 17.23 to W1 in order to conclude that H is negative deﬁnite. In the next theorem we analyze J-inner functions in terms of a realization which is not necessarily minimal. As always in this chapter, J stands for a signature matrix. Theorem 17.24. Let W (λ) = D + C(λIn − A)−1 B be a (possibly non-minimal) realization of a rational m × m matrix function. Suppose D ∗ JD = J, and assume there exists a Hermitian matrix X such that XA + A∗ X = C ∗ JC,

XB = C ∗ JD,

Ker (C|A) ⊂ Ker X.

Then W is J-unitary. In that case W is J-inner if and only if X is nonpositive.

17.5. J-inner rational matrix functions

335

Proof. With respect to the orthogonal decomposition Cn = Im X ⊕ Ker X write ) ( G 0 . X= 0 0 Note that G is invertible and Hermitian. Also, with respect to the decomposition Cn = Im X ⊕ Ker X, write B1 A11 A12 , B= , C = C1 C2 . A= A21 A22 B2 Then XB = C ∗ JD yields XB =

GB1 0

=

C1∗ C2∗

JD.

Since D∗ JD = J, we know that D is invertible. Hence JD is invertible, and so C2 = 0. Now XA + A∗ X = C ∗ JC gives ∗ ∗ C + A G GA JC 0 GA 11 12 1 11 1 = . XA + A∗ X = A∗12 G 0 0 0 As G is invertible, one obtains A12 = 0. Therefore W (λ) = D + C1 (λ − A11 )−1 B1 ,

(17.30)

and for this realization of W we have GA11 + A∗11 G = C1∗ JC1 and GB1 = C1∗ JD. It is now suﬃcient to show that (17.30) is minimal. Indeed, the proof can then be completed by applying Theorems 17.1 and 17.22. One checks that Ker CAj = Ker C1 Aj11 ⊕ Ker X,

j = 0, 1, 2, . . . .

As Ker (C|A) ⊂ Ker X by assumption, we obtain Ker (C1 |A11 ) = {0}. Thus (C1 , A11 ) is an observable pair. It remains to show that (A11 , B1 ) is controllable. For this it suﬃces to prove that (A× 11 , B1 ) is a controllable pair, where −1 −1 ∗ A× C1 . Now A× A11 G, while B1 = G−1 C1∗ JD. So it is 11 = A11 − B1 D 11 = −G enough to show that (−A∗11 , C1∗ JD) is a controllable pair. But this is equivalent to (D∗ JC1 , −A11 ) being an observable pair. Now D∗ J is invertible, and hence Ker (D∗ JC1 | − A11 ) = Ker (C1 |A11 ) = {0}, which completes the proof.

We ﬁnish this section with a theorem on the multiplicative structure of J-inner rational matrix functions. It states that a J-inner rational matrix function admits a complete factorization into J-inner factors of McMillan degree 1.

336

Chapter 17. J-unitary rational matrix functions

Theorem 17.25. Let W be a J-inner rational matrix function of McMillan degree n. Then there are J-inner rational matrix functions W1 , . . . , Wn of McMillan degree 1 such that W = W1 · · · Wn . Proof. Employing a similar argument as in the proof of Lemma 17.17, taking into account Theorem 17.22, one can prove that the J-inner rational matrix function W admits a realization with upper triangular main matrix and having −I as its associated Hermitian matrix. Following the line of argument of the proof of Theorem 17.16 one then proves that a J-inner rational matrix function admits a minimal factorization into n factors of degree 1, and that these factors can be taken to be J-unitary. It remains to show that the factors are actually J-inner. Let us consider for each of the factors a minimal realization of the form Wj (λ) = Dj +

1 Dj JBj∗ h−1 j Bj . λ − aj

The Hermitian matrix associated with this realization is denoted by hj ; it is just a real number in this case (compare Example 17.7). Consider the minimal realization for W resulting from taking the product realization of the above minimal realizations of the Wi ’s. According to Proposition 17.6, the Hermitian matrix H associated with this product realization is the diagonal matrix with the numbers h1 , . . . , hn on the diagonal. According to Proposition 17.4 and the state space similarity theorem, there is an invertible matrix S such that SHS ∗ is the Hermitian matrix associated with the minimal realization of W mentioned in the ﬁrst paragraph of this proof. That is, SHS ∗ = −I. But this is only possible if all numbers hi are negative. Then we can apply Theorem 17.22 to conclude that each of the factors is J-inner.

17.6 Inner-outer factorization In this section we consider inner-outer factorization of a possibly non-square p × q rational matrix function L. First we introduce the necessary terminology. A p×q rational matrix function V is called inner if V is analytic on the closed right half plane (including the imaginary axis and inﬁnity) and the values of V on the imaginary axis are isometries. The latter means that V (λ)∗ V (λ) = Ip for each λ ∈ iR. Since V is assumed to be proper, this identity also holds at inﬁnity. By the maximum modulus principle, an inner function V satisﬁes V (λ) ≤ 1,

λ ≥ 0.

Note that for V to be inner, we must have q ≤ p. If q = p , then V is inner if and only if V is bi-inner (cf., the ﬁrst two paragraphs of Section 17.5). A rational square matrix-valued function X is said to be an invertible outer function if X is analytic on the closed right half plane (inﬁnity included) and

17.6. Inner-outer factorization

337

det X(λ) = 0 for λ ≥ 0 (again with inﬁnity included). Finally, given a p × q rational matrix function L, we say that a factorization L(λ) = V (λ)X(λ), is an inner-outer factorization if V is a p × q inner rational function and X is a q × q invertible outer rational matrix function.1 Clearly, for such a factorization to exist L must be analytic in the closed right half plane (inﬁnity included) and the values of L on iR ∪ {∞} have to be left invertible matrices. As we shall see (Theorem 17.26 below), these two conditions are not only necessary for L to have an inner-outer factorization but also suﬃcient. ¯ ∗ L(λ). Obviously, if L has an inner-outer factorization Put Φ(λ) = L(−λ) L = V X (suppressing the variable λ), then, since V takes isometric values on the imaginary axis and at inﬁnity, we have ¯ ∗ X(λ), Φ(λ) = X(−λ) and this factorization is a left spectral factorization (with respect to iR) of the rational q × q matrix function Φ. This gives a hint about how to construct an inner-outer factorization. Indeed, assume L is analytic in the closed right half plane, inﬁnity included, ¯ ∗ X(λ) be a left spectral factorization of Φ with respect to and let Φ(λ) = X(−λ) iR. Put V (λ) = L(λ)X(λ)−1 . Then V is analytic in the closed right half plane (inﬁnity included) because both L and X −1 are analytic there. In addition, V takes isometric values on the imaginary axis. Hence L = V X is an inner-outer factorization. This leads to the following theorem. Theorem 17.26. Let L(λ) = D +C(λIn −A)−1 B be a realization of a p×q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L takes left invertible values on the imaginary axis, and D∗ D = Iq . Then L admits an inner-outer factorization L(λ) = V (λ)X(λ) with the inner factor V and the invertible outer factor X given by −1 V (λ) = D + (I − DD ∗ )C + DB ∗ P λIn − (A − BD ∗ C + BB ∗ P ) B, X(λ) = Iq + (D∗ C − B ∗ P )(λIn − A)−1 B. Here P is the (unique) Hermitian iR-stabilizing solution of P BB ∗ P + P (A − BD∗ C) + (A∗ − C ∗ DB ∗ )P − C ∗ (I − DD∗ )C = 0, that is, the solution P = P ∗ for which A − BD∗ C + BB ∗ P has all its eigenvalues in the open left half plane. 1 Note that in our deﬁnition of inner-outer factorization, the outer factor is required to be invertible outer. This restricted version of inner-outer factorization is used throughout the book.

338

Chapter 17. J-unitary rational matrix functions

¯ ∗ L(λ). Using D∗ D = I and the given realization for L Proof. Put Φ(λ) = L(−λ) − A) −1 B, where we compute that Φ is given by the realization Φ(λ) = I + C(λ ∗ C D −A∗ C ∗ C = −B ∗ D ∗ C . (17.31) = , B= , C A 0 A B Since L has left invertible values on the imaginary axis (that is, has full column rank there), Φ takes positive deﬁnite values on the imaginary axis. Thus we know from Section 9.2 that Φ admits a left spectral factorization with respect to iR. It follows that an inner-outer factorization does exist under the assumptions of the theorem. To ﬁnd the spectral factorization in concrete form, we proceed as in the proof of Theorem 13.1. In other words we apply Theorem 12.5 with the data given by (17.31). The same argument as in the proof of Theorem 13.1 gives that the Riccati equation featured in the theorem has a Hermitian stabilizing solution P . Now use P to deﬁne X(λ) by the expression given in the theorem which is the analogue of the expression for L− (λ) in Theorem 13.1. With the function X obtained this ¯ ∗ X(λ). way, we have the left spectral factorization Φ(λ) = X(−λ) −1 It remains to compute V (λ) = L(λ)X(λ) . Note that −1 X −1 (λ) = I − (D∗ C − B ∗ P ) λ − (A − BD∗ C + BB ∗ P ) B. From A − BD∗ C + BB ∗ P = A − B(D∗ C − B ∗ P ), we now obtain −1 (λ − A)−1 BX(λ)−1 = λ − (A − BD∗ C + BB ∗ P ) B. Using the latter identity it is straightforward to deduce the formula for V given in the theorem. The following corollary will be useful in the ﬁnal chapter of the book. Corollary 17.27. Let L(λ) = D +C(λIn −A)−1 B be a realization of a p×q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L takes left invertible values on the imaginary axis, and D ∗ D = Iq . Then there is a q × p rational matrix function L (λ) which has no poles on the imaginary line including inﬁnity, such that L (iω)L(iω) = Iq , ω ∈ R. Proof. Let L(λ) = V (λ)X(λ) be an inner-outer factorization of L and take L (λ) = X(λ)−1 V (−λ)∗ . Next we consider the dual problem of outer-co-inner factorization. A possibly non-square rational matrix function V is called co-inner if V is analytic on the closed right half plane (including inﬁnity), and takes co-isometric values on the ¯ ∗. imaginary axis. In other words, V is co-inner if V is inner, where V (λ) = V (λ) Note that for V to be co-inner, we must have p ≤ q.

17.7. Unitary completions of minimal degree

339

A factorization L(λ) = X(λ)V (λ), where X is invertible outer and V is co-inner is called an outer-co-inner factor¯ ∗ = X(λ)X(−λ) ¯ ∗ is a right ization.2 Obviously, in that case Φ(λ) = L(λ)L(−λ) spectral factorization with respect to iR, and conversely. Using a duality argument we obtain the following counterpart to Theorem 17.26. Theorem 17.28. Let L(λ) = D + C(λ − A)−1 B be a realization of a p × q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L(iω) is right invertible for each ω ∈ R, and DD ∗ = I. Then L admits an outer-co-inner factorization L(λ) = X(λ)V (λ), with the co-inner factor and the invertible outer factor being given by −1 V (λ) = D + C λIn − (A − BD∗ C + QC ∗ C) B(I − D ∗ D) + QC ∗ D , X(λ) = Ip + C(λIn − A)−1 (BD∗ − QC ∗ ). Here Q is the (unique) Hermitian iR-stabilizing solution QC ∗ CQ + (A − BD∗ C)Q + Q(A∗ − C ∗ DB ∗ ) − B(I − D∗ D)B ∗ = 0, that is, the solution Q = Q∗ for which A − BD∗ C + QC ∗ C has all its eigenvalues in the open left half plane. Note that Proof. Let L(λ) = D ∗ +B ∗ (λ−A∗ )−1 C ∗ , and apply Theorem 17.26 to L. ∗ A also has all its eigenvalues in the open left half plane. So, applying Theorem yields a factorization L(λ) is 17.26 to L = V (λ)X(λ), where V is inner and X ∗ ∗ ¯ is co-inner and X(λ) = X( ¯ is invertible λ) invertible outer. Then V (λ) = V (λ) outer. So L(λ) = X(λ)V (λ) is an outer-co-inner factorization of L. Theorem 17.26 Those for V and X are now obtained also gives formulas for the factors V and X. ∗ ¯ ¯ ∗. λ) from the expressions V (λ) = V (λ) and X(λ) = X(

17.7 Unitary completions of minimal degree In this section we deal with the following completion problem. Given a strictly proper rational m × p matrix function W , having contractive values on the imaginary axis, ﬁnd an (m + p) × (m + p) rational matrix function U having unitary values on the imaginary axis, such that U11 (λ) W (λ) . (17.32) U (λ) = U21 (λ) U22 (λ) 2 Note that in our deﬁnition of outer-co-inner factorization, the outer factor is required to be invertible outer (cf., footnote 1).

340

Chapter 17. J-unitary rational matrix functions

In other words, we want to ﬁnd a unitary rational matrix function U such that W is embedded as a (right upper) corner in U . Moreover, we wish to ﬁnd such a U which has the same McMillan degree as W . We shall normalize U so that U (∞) = Im+p . This problem can be treated for the more general case of a proper W (see [75]). However, for sake of simplicity we shall conﬁne ourselves to the strictly proper case. The following theorem describes all possible solutions. Theorem 17.29. Let W (λ) = C(λIn − A)−1 B be a minimal realization of an m × p strictly proper rational matrix function W which is contractive on the imaginary axis. Then the set of all unitary rational (m + p) × (m + p) matrix functions U of the form (17.32) with U (∞) = Im+p and δ(U ) = δ(W ) is in one-to-one correspondence with the set of Hermitian solutions of the algebraic Riccati equation XC ∗ CX − AX − XA∗ + BB ∗ = 0.

(17.33)

Moreover, these Hermitian solutions X are invertible, and the one-to-one correspondence referred to above is given by Im 0 C (17.34) U (λ) = + (λIn − A)−1 XC ∗ B . ∗ −1 B X 0 Ip Proof. Suppose U is a unitary rational matrix function with W as its right upper corner block entry, U (∞) = Im+p and δ(U ) = δ(W ). The McMillan degree of W is n, the size of the main matrix in the given minimal realization W (λ) = C(λIn − A)−1 B of W . Hence δ(U ) = n, and U has a realization of the type U (λ) =

Im

0

0

Ip

+

1 C 2 C

−1 B 1 (λIn − A)

. 2 . B

−1 B 1 (λIn − A) 2 is realization of W . Comparing this realization Clearly W (λ) = C with the given one, and using the state space similarity theorem for minimal real = S −1 AS, B 2 = izations, we see that there exists an invertible n×n matrix with A −1 −1 S B and C1 = CS. Introducing C2 = C2 S and B1 = SB1 , we get Im 0 C (17.35) U (λ) = + (λIn − A)−1 B1 B , 0 Ip C2 and this realization of U is a minimal one. Since U is unitary, there is a Hermitian X such that ∗ ∗ C B1 B 1 , X= . AX + XA∗ = B1 B B∗ B∗ C2

(17.36)

17.8. Bi-inner completions of inner functions

341

In particular, we have B1∗ = CX. Inserting B1∗ = CX into the ﬁrst part of (17.36) we obtain (17.33). Moreover, X is invertible by minimality of the realization of U (see Theorem 17.1), and so C2 = B ∗ X −1 , which yields (17.34). Conversely, suppose that X is a Hermitian solution of (17.33). By minimality of the realization of W we have that X is invertible. The argument is as follows. Suppose Xx = 0. Then (17.33) gives x∗ BB ∗ x = 0, hence B ∗ x = 0. Again using (17.33) we get XA∗ x = 0, and we see that Ker X ⊂ Ker (B ∗ |A∗ ) = {0}. Let U be given by (17.34). Then, by Theorem 17.1, the rational matrix function U is unitary. Obviously, W is the right upper corner block entry of U and δ(U ) = δ(W ), and U (∞) = Im+p . To show that the correspondence between Hermitian solutions X of (17.33) and the set of all unitary rational matrix functions U of the form (17.32) with U (∞) = I and δ(U ) = δ(W ) is one-to-one we argue as follows. We have seen in the previous part of the theorem that any such U is necessarily of the form (17.34) for some Hermitian solution of (17.33). Assume that for two solutions X1 and X2 the functions U1 and U2 given by (17.34) with these solutions in place of X coincide. Then, from (17.36) it is seen that C ∗ (X1 − X2 ) = 0. A(X1 − X2 ) + (X1 − X2 )A = 0, C2 Hence Im (X1 − X2 ) is A-invariant, and it is also contained in Ker C. This implies that Im (X1 − X2 ) ⊂ Ker (C|A) = {0}. Thus X1 = X2 .

17.8 Bi-inner completions of inner functions Our aim in this section is to complete a possibly non-square inner function to a (square) bi-inner one. It is convenient to begin with two propositions. With the notation used in the ﬁrst proposition we anticipate Theorem 17.32 below. + C(λIn − A)−1 B be a realization of a p × q Proposition 17.30. Let V (λ) = D rational matrix function, and assume ∗D = Iq , D

σ(A) ⊂ Cleft,

= C ∗ D, YB

(17.37)

where Y is the unique (Hermitian) solution of the Lyapunov equation Y A + A∗ Y = C ∗ C.

(17.38)

Then V is inner. Conversely, if V is inner, the given realization of V is minimal, and Y is the unique (Hermitian) solution of the Lyapunov equation (17.38), then (17.37) is satisﬁed. Since A has all its eigenvalues in the open left half plane, equation (17.38) has a unique solution Y , and this solution is given by ∞ ∗ etA C ∗ CetA dt. (17.39) Y =− 0

342

Chapter 17. J-unitary rational matrix functions

From this representation one sees that the matrix Y is generally negative semidefinite, and that it has the stronger property of being negative deﬁnite when the is minimal (or even just observable). Thus + C(λIn − A)−1 B realization V (λ) = D the above result can be viewed as a special case of Theorem 17.24. It is illustrative to give a direct proof. is Proof. Assume that (17.37) holds with Y as indicated in the theorem. Then D an isometry by the ﬁrst condition in (17.37). Thus p ≥ q. For pure imaginary λ, a straightforward computation, using (17.37) and (17.38), gives ∗ ¯ − A∗ )−1 C ∗ D +B ∗ (λ + C(λ − A)−1 B V (λ)∗ V (λ) = D ∗ (λ + A∗ )−1 Y B +B ∗ Y (λ − A−1 B = Iq − B ∗ (λ + A∗ )−1 (Y A + A∗ Y )(λ − A)−1 B −B ∗ (λ + A∗ )−1 Y B +B ∗ Y (λ − A)−1 B = Iq − B ∗ (λ + A∗ )−1 Y (A − λ) + (A∗ + λ)Y )(λ − A −1 B = Iq . −B Hence V has isometric values on iR. Since V is analytic in the open right half plane by the second condition in (17.37), we may conclude that V is inner. + C(λIn − A)−1 B be Next, let V be inner and let the realization V (λ) = D minimal. Clearly, since V is inner, the ﬁrst two conditions in (17.37) are satisﬁed. = C ∗ D. This Let Y be the unique solution of (17.38). It remains to show that Y B is done by using the same arguments as used in the proof of Proposition 17.3. Proposition 17.31. Let U (λ) = D + C(λIn − A)−1 B be a realization of a p × p rational matrix function. Assume D∗ D = Iq ,

σ(A) ⊂ Cleft,

Y B = C ∗ D,

(17.40)

where Y is the unique (Hermitian) solution of the Lyapunov equation Y A + A∗ Y = C ∗ C.

(17.41)

Then U is bi-inner and the McMillan degree of U is equal to the rank of Y which, in turn, is equal to dim Ker (C|A)⊥ . Proof. The fact that U is bi-inner follows from Proposition 17.30. Since σ(A) is contained in Cleft , the unique solution Y of (17.41) is given by the integral representation (17.39), from which we easily obtain Ker Y = Ker (C|A). Now consider the decomposition Cn = X1 ⊕ X2 , where X1 = Ker (C|A) and X2 is the orthogonal compliment of X1 in Cn . Thus X1 = Ker Y and X2 = Im Y . In particular rank Y = dim X2 . Write A, B, C and Y as block matrices according to the decomposition Cn = X1 ⊕ X2 . Then B1 A1 0 0 , B= , C = 0 C2 , , (17.42) A= Y = B2 0 A2 0 Y2

17.8. Bi-inner completions of inner functions

343

and U (λ) = D + C2 (λIn − A2 )−1 B2 . Since rank Y = rank Y2 = dim X2 , it suﬃces to prove that this second realization of U is minimal. From (17.41), the third identity in (17.40) and the partitioning of A, B, C and Y in (17.42), we see that Y2 A2 + A∗2 Y2 = C2∗ C2 ,

Y2 B2 = C2∗ D.

−1 C2 , the associate main matrix of the realization U (λ) = For A× 2 = A2 − B2 D D + C2 (λIn − A2 )−1 B2 , this gives −1 Y2 A× C2 = −A∗2 Y2 + C2∗ C2 − C2∗ DD−1 C2 = −A∗2 Y2 . 2 = Y2 A2 − Y2 B2 D ∗ Now Y2 is invertible. Thus A× 2 and −A2 are similar. From the second part of (17.40) and the partitioning of A in (17.42), we see that σ(A2 ) ⊂ Cleft . Taking × ∗ into account the similarity of A× 2 and −A2 , it follows that σ(A2 ) ⊂ Cright . In × particular, σ(A2 ) and σ(A2 ) are disjoint. But then, by a remark made after the proof of Theorem 7.6 in [20], the realization U (λ) = D + C2 (λIn − A2 )−1 B2 is minimal.

Let V be as in Proposition 17.30, so in particular V is inner. Returning to the aim of this section, we shall now complete V to a p × p bi-inner rational matrix function. Before turning to the theorem in question, we make some preparations. is an isometry. Thus According to the ﬁrst condition in (17.37) the matrix D p ≥ q. When p = q, there is nothing to do. Therefore in what follows we take D ∗ is an orthogonal is an isometry, implies that Ip − D p > q. The fact that D projection of rank p − q. Thus we can choose a p × (p − q) isometry E such that D ∗ = EE ∗ . Now note that there exists an n × (p − q) matrix B such that Ip − D Y B = C ∗ E.

(17.43)

Since Y is Hermitian, to prove that equation (17.43) has a solution of the desired form, it suﬃces to show that Ker Y ⊂ Ker E ∗ C. In fact, we have Ker Y ⊂ Ker C. Indeed, assume that Y x = 0, then we see from (17.38) that x∗ C ∗ Cx = 0, which is equivalent to Cx = 0. be a realization of a p × q + C(λIn − A)−1 B Theorem 17.32. Let V (λ) = D rational matrix function satisfying the conditions (17.37), where Y is the unique (Hermitian) solution of the Lyapunov equation (17.38). Let E be a p × (p − q) D ∗ = EE ∗ , and let B be an n × (p − q) matrix solution isometry such that Ip − D of (17.43). Put U (λ) = D + C(λIn − A)−1 B, where B and D are the p × p ] and D = [ E D ]. Then U is a p × p bi-inner matrices given by B = [ B B completion of V , that is, U is a bi-inner rational p × p matrix function of the form [ V (λ) V (λ) ], and the McMillan degree of U is equal to the rank of Y . The rational p × (p − q) matrix function V can be described explicitly; it is actually given by the realization V (λ) = E + C(λIn − A)−1 B .

344

Chapter 17. J-unitary rational matrix functions

Proof. To prove that U is bi-inner, apply Proposition 17.30 to U with its given realization. Since (17.38) holds, it suﬃces to show that Y B = C ∗ D and D∗ D = Ip . These facts follow from the third identity in (17.37) and the deﬁnitions of E and B . Indeed, we have = C∗E C ∗D = C ∗ D, Y B = Y B Y B ∗ E D ∗ = Ip , = EE ∗ + D DD∗ = E D ∗ D

and, since D is a square matrix, DD ∗ = Ip amounts to the same as D∗ D = Ip . The ﬁnal statement is an immediate corollary of Proposition 17.31. Next, we return to the inner-outer factorization discussed in Section 17.6. The point we focus on here is the completion of the inner factor to a bi-inner function. Let L(λ) = D + C(λIn − A)−1 B be a realization of a p × q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L(iω) is left invertible for each ω ∈ R, and D ∗ D = Iq . Let L(λ) = V (λ)X(λ) be the inner-outer factorization constructed in Theorem 17.26, in particular, −1 V (λ) = D + (I − DD∗ )C + DB ∗ Y λIn − (A − BD ∗ C + BB ∗ Y ) B, where Y = Y ∗ satisﬁes the algebraic Riccati equation Y BB ∗ Y + Y (A − BD∗ C) + (A∗ − C ∗ DB ∗ )Y − C ∗ (I − DD∗ )C = 0, and A − BD∗ C + BB ∗ Y has all its eigenvalues in the open left half plane. Choose a p × (p − q) isometry E such that I − DD ∗ = EE ∗ , and let B be any n × (p − q) matrix such that Y B = C ∗ E. Corollary 17.33. In the situation described in the previous paragraph, introduce U (λ) = V (λ) V (λ) , where the rational p × (p − q) matrix function V is given by −1 V (λ) = E + (I − DD∗ )C + DB ∗ Y λIn − (A − BD ∗ C + BB ∗ Y ) B . Then U is bi-inner. Proof. All we need to show is that Theorem 17.32 may be applied with the matrices A − BD ∗ C + BB ∗ Y and (I − DD∗ )C + DB ∗ Y in place of A and ∗ C, respectively. For this we need to verify the identities (I − DD∗ )C + DB ∗ Y D = Y B and Y (A − BD∗ C + BB ∗ Y ) + (A − BD∗ C + BB ∗ Y )∗ Y ∗ = (I − DD ∗ )C + DB ∗ Y (I − DD∗ )C + DB ∗ Y . This involves nothing more than a routine computation using that D∗ D = I and that Y = Y ∗ is a solution of the Riccati equation featured in the paragraph preceding the corollary.

17.8. Bi-inner completions of inner functions

345

Notes The ﬁrst three sections are largely based on [3]. The Redheﬀer transformation of Section 17.4, which is a standard tool in the analysis of 2 × 2 block matrix functions, originates from [130]. Theorem 17.22 in Section 17.5 also implies that if W is a J-inner rational matrix function, then the function K∗,W (μ, λ) has no negative squares, that is, it is a positive deﬁnite kernel, see also Theorem 2.5 in [39]. Theorem 17.25 in Section 17.5 is a simple case of a more far-reaching theory concerning the multiplicative structure of general matrix-valued J-inner functions, which originates from [118]; see also Chapter 4 in [39]. Factorizations in degree 1 factors, of which Theorem 17.25 provides an example, are the main topic of Part III in [20]. Section 17.6 originates from Section 7.4 in [43]; for the corresponding state space formulas, see [146]. Section 17.7 is related to the problem of Darlington synthesis. The latter problem can be found in [4]. The presentation given here is based on [75]. For further results in this direction, including Darlington embedding for time-variant systems, see [36] and Chapter 6 in [117]. The result presented in Section 17.8 may be found in, e.g., Chapter 12 (page 249) in [149].

Part VII Applications of J-spectral factorizations In this part, the state space theory of J-spectral factorization, developed in the preceding two parts, is used to solve H∞ -problems. There are three chapters. The ﬁrst chapter (Chapter 18) presents the solution of the Nehari interpolation problem for rational matrix functions. The second chapter (Chapter 19) reviews elements from control and mathematical systems theory that play an essential role in the ﬁnal chapter. The third and ﬁnal chapter (Chapter 20) treats H∞ -control problems. Here we use the J-spectral factorization theory to obtain the solutions of some of the main problems in this area, namely the standard problem, the one-sided problem, and the full model matching problem.

Chapter 18

Application to the rational Nehari problem In this chapter the rational matrix version of the Nehari problem (relative to the imaginary axis) is solved using a J-spectral factorization approach. The data of the problem are given in realized form. This together with the state space results on J-spectral factorization derived in Chapter 14 allows us to solve the problem and to obtain an explicit linear fractional representation of all its solutions, again in realized form. The main attention is given to the so-called suboptimal case. The more general Nehari-Takagi problem is also solved using the J-spectral factorization method. This chapter consists of six sections. Section 18.1 presents the problem statement and the main theorem. Section 18.2 deals with the theory of linear fractional maps. Such maps will play an important role in this and the ﬁnal chapter. In Section 18.3 the rational matrix Nehari problem is reduced to a J-spectral factorization of a special kind, and all solutions are described in terms of the coeﬃcients of the J-spectral factor. This result is used in Section 18.4 to prove the main theorem of Section 18.1. Section 18.5 deals with the Nehari problem for the non-stable case, when the given function does not necessarily have all its poles in the open left half plane. Section 18.6, the ﬁnal section of the chapter, gives the solution of the rational matrix Nehari-Takagi problem.

18.1 Problem statement and main result Let R be a rational p × q matrix function which does not have a pole on the imaginary axis and at inﬁnity. In particular, R is proper. In this section we study the problem of ﬁnding all proper rational p × q matrix functions K such that K

350

Chapter 18. Application to the rational Nehari problem

has all its poles in the open right half plane and K − R∞ = sup K(s) − R(s) < γ, s∈iR

(18.1)

where γ is a pre-speciﬁed positive number. Note that both R and K are proper and have no pole on the imaginary axis, and hence the so-called inﬁnity norm K − R ∞ is well-deﬁned. We shall refer to this problem as the (suboptimal) rational Nehari problem for R relative to the imaginary axis with tolerance γ. The latter qualiﬁer will be omitted when γ = 1. The word “suboptimal” refers to the fact that we use in (18.1) a strict inequality. We ﬁrst deal with the case when R is stable. A rational matrix function is called iR-stable, or simply stable when no confusion is possible (as will be the case in this chapter), if all its poles are in the open left half plane. Note that such a function is proper and has no pole on iR. We shall assume additionally that R is strictly proper. To state the main result we start with a realization of R. Since R is stable and strictly proper, we can choose a realization of R of the form R(λ) = C(λIn − A)−1 B,

(18.2)

with the property that A has all its eigenvalues in the open left half plane. Let P and Q be the unique solutions of the Lyapunov equations AP + P A∗ = −BB ∗ ,

A∗ Q + QA = −C ∗ C,

respectively. Note that P and Q are given by ∞ τA ∗ τ A∗ e BB e dτ, Q= P = 0

∞

(18.3)

∗

eτ A C ∗ Ceτ A dτ.

0

Hence P and Q are nonnegative Hermitian matrices. One usually refers to P as the controllability gramian, and to Q as the observability gramian, corresponding to the realization (18.2). We shall prove the following theorem. Theorem 18.1. Let R(λ) = C(λIn − A)−1 B be a realization of the p × q rational matrix function R, and assume that A has all its eigenvalues in the open left half plane. Then the rational Nehari problem for R relative to the imaginary axis with tolerance γ is solvable if and only if the matrix γ 2 In − P 1/2 QP 1/2 is positive deﬁnite. In that case all solutions of the Nehari problem for R can be obtained in the following way. Introduce the rational matrix functions X11 (λ)

= Ip + CP (λIn + A∗ )−1 Z −1 C ∗ ,

(18.4)

X12 (λ)

= CP (λIn + A∗ )−1 Z −1 QB,

(18.5)

X21 (λ)

= −B ∗ (λIn + A∗ )−1 Z −1 C ∗ ,

(18.6)

X22 (λ)

= Iq − B ∗ (λIn + A∗ )−1 Z −1 QB,

(18.7)

18.1. Problem statement and main result

351

where Z = γ 2 In − QP . Then all solutions K of the rational Nehari problem for R relative to the imaginary axis are given by −1 K(λ) = − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) ,

(18.8)

where H is any rational p × q matrix function which has all its poles in the open right half plane and satisﬁes H∞ < γ. Moreover, there is a one-to-one correspondence between the solution K and the free parameter H. Before we prove the above theorem (in Section 18.4 below) it will be convenient ﬁrst to make some preparations. The following lemma restates the necessary and suﬃcient condition appearing in Theorem 18.1 in operator language. Lemma 18.2. Let R(λ) = C(λIn − A)−1 B be a realization of the p × q rational matrix function R, and assume that A has all its eigenvalues in the open left half plane. Consider the Hankel operator HR generated by R, that is the ﬁnite rank integral operator from Lq2 [0, ∞) into Lp2 [0, ∞) given by ∞ (HR f )(t) = CeA(t+τ ) Bf (τ ) dτ. 0

Then HR < γ if and only if the matrix γ 2 In − P 1/2 QP 1/2 is positive deﬁnite. Proof. We need the controllability operator Ξ and the observability operator Ω associated with the realization (18.2). Thus ∞ Ξf = eτ A Bf (τ ) dτ, Ξ : Lq2 [0, ∞) → Cq , 0

Ω : Cn → Lp2 [0, ∞),

(Ωx)(t) = CetA x,

t > 0.

Clearly P = ΞΞ∗ , Q = Ω∗ Ω and HR = ΩΞ. Now let λ1 (X) denote the largest eigenvalue of an operator X all of whose non-zero spectrum consists of positive eigenvalues. Then HR 2

=

∗ λ1 (HR HR ) = λ1 (Ξ∗ Ω∗ ΩΞ)

=

λ1 (ΞΞ∗ Ω∗ Ω) = λ1 (P Q) = λ1 (P 1/2 QP 1/2 ).

Hence HR < γ if and only if all the eigenvalues of P 1/2 QP 1/2 are strictly less than γ 2 . Thus HR < γ if and only if γ 2 I − P 1/2 QP 1/2 is positive deﬁnite. We close the section by showing that, without loss of generality, we may assume that in Theorem 18.1 the tolerance γ = 1. Indeed, consider for the original problem R(λ) = γ −1 R(λ), and K(λ) = γ −1 K(λ). Then we have R − K∞ < γ − K ∞ < 1. Moreover, if R is given by the realization (18.2), if and only if R = γ −1 C. One easily admits the realization R = C(λ − A)−1 B, where C then R

352

Chapter 18. Application to the rational Nehari problem

of the corresponding Lyapunov equations (18.3), sees that, for solutions P and Q −2 = γ Q. Hence Z = I − PQ = γ −2 Z. For the functions Xij (λ) one has P = P , Q appearing in Theorem 18.1 we have the following: 11 (λ) X

=

(λ + A∗ )−1 Z −1 C ∗ = X11 (λ), Ip + CP

12 (λ) X

=

(λ + A∗ )−1 Z −1 QB = γ −1 X21 (λ), CP

21 (λ) X

=

−1 C ∗ = γX21 (λ), −B ∗ (λ + A∗ )−1 Z

22 (λ) X

=

−1 QB = X22 (λ). Iq − B ∗ (λ + A∗ )−1 Z

Suppose that K(λ) is a solution to the problem with γ = 1, given by 11 (λ)H(λ) 12 (λ))(X 21 (λ)H(λ) 22 (λ))−1 , K(λ) = −(X +X +X satisfying H ∞ < 1. Now taking H(λ) = γ H(λ) for some H we have H∞ < γ, and with K(λ) = γ K(λ), we obtain that (18.8) holds.

18.2 Intermezzo about linear fractional maps The expression (18.8), which assigns to the rational matrix function H a rational matrix function K, is usually called a linear fractional map. Such maps will play an important role in this and the ﬁnal chapter. Therefore, we review some of the main properties of linear fractional maps in this section. It will be convenient ﬁrst to introduce some notation and terminology. Given a p × q rational matrix function F , we write F ∗ for the adjoint of F relative to the ¯ ∗ . (In engineering literature, including [76], imaginary axis, that is, F ∗ (λ) = F (−λ) [43]), this function is often denoted by F ∼ .) By Rat we shall denote the set of all rational matrix functions that are proper and have no pole on the imaginary axis iR, and Ratp×q will stand for the set of all F in Rat that are of size p × q. If F belongs to Ratp×q , then F ∗ belongs to Ratq×p . Note that Ratp×q is closed under the usual addition of matrix functions as well as under scalar multiplication. Also for F ∈ Ratp×q and G ∈ Ratq×r , we have F G ∈ Ratp×r . In particular Ratp×p is an algebra. The unit element in this algebra is Ep , the p × p matrix function which is identically equal to the p × p identity matrix Ip . A function F ∈ Ratp×p is said to be invertible in Ratp×p if F has an inverse G in Ratp×p , that is, G ∈ Ratp×p and F G = GF = Ep . For a rational p × p matrix function F such that det F (λ) ≡ 0, the pointwise inverse F −1 , deﬁned by F −1 (λ) = F (λ)−1 , is again a rational matrix function. If F ∈ Ratp×p and det F (λ) ≡ 0, then F −1 need not be an element of Ratp×p . Indeed, F −1 might have a pole on the imaginary axis or fail to be proper. In fact, F −1 ∈ Ratp×p if and only if F is biproper and det F (λ) has no zero on iR, and in that case F −1 is the inverse of F in the algebra Ratp×p .

18.2. Intermezzo about linear fractional maps

353

A function F in Ratp×q is analytic on the imaginary axis and at inﬁnity. Hence we can consider the norm F ∞ = sup F (s).

(18.9)

s ∈ iR

This is the usual L∞ -norm for bounded matrix functions on iR which we already p×q used in (18.1). We write F ∈ Ratp×q and its B , whenever F belongs to Rat p×q inﬁnity-norm F ∞ is strictly less than 1. Thus RatB is the open unit ball in Ratp×q with respect to the norm deﬁned by (18.9). Note that F ∞ < 1 is ¯ ∗ F (λ) being positive deﬁnite on iR ∪ {∞}. For the latter equivalent to Ip − F (−λ) property we use the notation Ep − F ∗ F > 0. Now let Θ ∈ Rat(p+q)×(p+q) , and let us partition Θ as a 2 × 2 block matrix function in the following way: Θ11 (λ) Θ12 (λ) (18.10) Θ(λ) = Θ21 (λ) Θ22 (λ) with Θ11 (λ) a p × p matrix and Θ22 (λ) a q × q matrix. With this partitioning of Θ we associate the linear fractional map −1 (FΘ H)(λ) = Θ11 (λ)H(λ) + Θ12 (λ) Θ21 (λ)H(λ) + Θ22 (λ) .

(18.11)

Here H is assumed to be in Ratp×q . In general, it is not clear for which H the map is well-deﬁned. However for a J-unitary Θ, with J = diag (Ip , −Iq ), we have the following result. Theorem 18.3. Let Θ ∈ Rat (p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ). Then Θ is invertible in Rat(p+q)×(p+q) , the maps FΘ and FΘ−1 are well-deﬁned on Ratp×q and map Ratp×q into itself. Moreover B B H ∈ Ratp×q B .

H = FΘ−1 FΘ H = FΘ FΘ−1 H,

(18.12)

Proof. We divide the proof into three parts. In the ﬁrst part it is shown that Θ−1 is in Rat(p+q)×(p+q) and is J-unitary, and also that the maps FΘ and FΘ−1 are p×q well-deﬁned on Ratp×q into B . In the second part we prove that FΘ maps RatB itself. In the ﬁnal part the identities in (18.12) will be established. Part 1. Since Θ is proper and has no pole on iR, the fact that Θ is J-unitary implies that for each λ ∈ iR ∪ {∞} the matrix Θ(λ) is J-unitary and hence invertible. It follows that Θ is invertible in Rat(p+q)×(p+q) and that Θ−1 is J-unitary. The fact that the matrix Θ(λ) is J-unitary for λ ∈ iR ∪ {∞} implies that Θ22 (λ) is invertible and Θ22 (λ)−1 Θ21 (λ) < 1 for λ ∈ iR ∪ {∞} . It follows that Θ22 is invertible in Ratq×q and that −1 Θ−1 Θ21 (λ) = 22 Θ21 ∞ = sup Θ22 (λ) λ∈R

max

λ ∈ iR ∪{∞}

Θ22 (λ)−1 Θ21 (λ) < 1.

354

Chapter 18. Application to the rational Nehari problem

p×q −1 −1 Next, take H ∈ Rat 21 ∞ H∞ < 1. Thus B−1 . Then Θ22 Θ21 H∞ ≤ Θ22 Θ Θ21 H + Θ22 = Θ22 Θ22 Θ21 H + Eq is invertible in Ratq×q . It follows that FΘ H −1 is also J-unitary, FΘ−1 is well-deﬁned is well-deﬁned for H ∈ Ratp×q B . Since Θ p×q on RatB too.

into itself. Take H in Ratp×q Part 2. In this part we show that FΘ maps Ratp×q B B , and write F = FΘ H. First note that F (Θ11 H + Θ12 )(Θ21 H + Θ22 )−1 H (18.13) = = Θ X −1 , −1 Eq (Θ21 H + Θ22 )(Θ21 H + Θ22 ) Eq

where X = Θ21 H + Θ22 . The fact that Θ is J-unitary, with J = diag (Ip , −Iq ) is equivalent to the identity 0 0 E E p p Θ∗ Θ = . (18.14) 0 −Eq 0 −Eq Hence, using (18.13), we obtain ∗

Eq − F F

=

−

F

∗

Eq

0

0

−Eq

−X

−∗

=

−X

−∗

=

X −∗ Eq − H ∗ H X −1 .

=

H

∗

H

∗

Eq

Eq

Ep

Θ

∗

F

Eq

Ep

0

0

−Eq

Ep

0

0

−Eq

Θ

H Eq

H

Eq

X −1

X −1

¯ ∗ F (λ) = X(−λ) ¯ −∗ Iq − H(−λ) ¯ ∗ H(λ) X(λ)−1 . Now It follows that Iq − F (−λ) ¯ ∗ H(λ) is positive deﬁnite on iR ∪ {∞}. H∞ < 1. This means that Ip − H(−λ) ∗ ¯ But then Iq − F (−λ) F (λ) is also positive deﬁnite on iR ∪ {∞}. The latter is equivalent to F ∞ < 1. Thus F ∈ Ratp×q B , as desired. From what has been proved so far, we conclude that the result of the previous steps also hold with Θ−1 instead of Θ. Thus FΘ−1 maps Ratp×q into itself. B Therefore, to complete the proof, it remains to prove the identities in (18.12). In fact, by interchanging the roles of Θ and Θ−1, it suﬃces to prove the ﬁrst identity in (18.12). This will be done in the next part. Part 3. Take H ∈ Ratp×q B , and put F = FΘ H, G = FΘ−1 F . From (18.14) we see that ∗ ∗ E Θ 0 0 −Θ E p p 11 21 Θ−1 = Θ∗ = . (18.15) −Θ∗12 Θ∗22 0 −Eq 0 −Eq

18.2. Intermezzo about linear fractional maps

355

By using (18.13) for Θ as well as for Θ−1 , we have

F

= Θ−1

Eq

(Θ21 H + Θ22 )−1 ,

Eq

G

H

=Θ

Eq

F Eq

(−Θ∗12 F + Θ∗22 )−1 .

Now observe that −Θ∗12 F

+

Θ∗22

=

=

=

0

Eq

0

Eq

0

Eq

Θ∗11 F − Θ∗21

=

−Θ∗12F + Θ∗22

Θ

−1

Θ

H

0

Eq

−1

Θ

F

Eq

(Θ21 H + Θ22 )−1

Eq

H Eq

(Θ21 H + Θ22 )−1 = (Θ21 H + Θ22 )−1 .

In particular, (Θ21 H + Θ22 )−1 (−Θ∗12 F + Θ∗22)−1 = Eq . But then G =

=

=

Ep

Ep

Ep

0

0

0

G

=

Eq

Θ

−1

Θ

H Eq

H

Ep

0

Θ

−1

F Eq

(−Θ∗12 F + Θ∗22 )−1

Eq

(Θ21 H + Θ22 )−1 (−Θ∗12 F + Θ∗22 )−1

= H,

which proves the ﬁrst identity in (18.12).

We are particulary interested in proper rational p × q matrix functions that are analytic on the closed left half plane with inﬁnity included. The class of these p×q functions will be denoted by Ratp×q have no pole + . Since the functions in Rat+ p×q p×q on iR and are proper, Rat+ is a linear subspace of Rat . We write Ratp×q +, B for the set of all F Ratp×q such that (18.9) holds. Thus + p×q ∩ Ratp×q Ratp×q +, B = Rat+ B .

356

Chapter 18. Application to the rational Nehari problem

Now, as in Theorem 18.3, let Θ ∈ Rat(p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ). Fix R ∈ Ratp×q , and consider Θ11 − RΘ21 Θ12 − RΘ22 Ep −R V11 V12 = = Θ. (18.16) V = 0 Eq V21 V22 Θ21 Θ22 Since Θ is invertible in Rat(p+q)×(p+q) by Theorem 18.3, it follows that the same holds true for V . Let FV be the linear fractional map deﬁned by V . Since V21 = Θ21 and V22 = Θ22 , we know from Theorem 18.3 that for each function H in Ratp×q the B function V21 H + V22 is invertible in Ratq×q . Thus FV is well-deﬁned on Ratp×q B . Moreover, since V11 H + V12

= (Θ11 − RΘ21 )H + (Θ12 − RΘ22 ) = (Θ11 H + Θ12 ) − R(Θ21 H + Θ22 ),

we see that

H ∈ Ratp×q B .

FV H = FΘ H − R,

The fact V22 = Θ22 implies that V22 is invertible in Rat is the second main result of this section.

q×q

(18.17)

. The following theorem

Theorem 18.4. Let Θ ∈ Rat(p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ), and let V be given by (18.16), where R ∈ Ratp×q . Then V is invertible in Rat (p+q)×(p+q) , and V22 is invertible in Rat q×q . Assume additionally that (p+q)×(p+q)

(a) V and V −1 belong to Rat+

,

−1 (b) V22 and V22 belong to Ratq×q + .

Then FV is well-deﬁned and one-to-one on Ratp×q +, B . Also p×q FV Ratp×q | R + K∞ < 1 . +, B = K ∈ Rat+

(18.18)

Note that conditions (a) and (b) in the above theorem are not independent. Indeed, the property that V22 belongs to Ratq×q follows from the fact that V + (p+q)×(p+q) . belongs to Rat+ Proof. The fact that V is invertible in Rat(p+q)×(p+q) and V22 in Ratq×q has already been proved in the two paragraphs preceding Theorem 18.4. From Theorem 18.3 we know that FΘ is well-deﬁned and one-to-one on Ratp×q B . But then we see p×q from (18.17) that the same holds true for FV . Now recall that Ratp×q +, B ⊂ RatB . p×q This allows us to conclude that FV is well-deﬁned and one-to-one on Rat+, B . It remains to show the identity (18.18). This will be done in two parts. The ﬁrst part covers the inclusion p×q | R + K∞ < 1 . (18.19) FV Ratp×q +, B ⊂ K ∈ Rat+

18.2. Intermezzo about linear fractional maps

357

The reverse inclusion is proved in the second step. p×q Part 1. Take H in Ratp×q +, B . We ﬁrst show that FV H belongs to Rat+ . From condition (a) we know that V is analytic on the closed left half plane. Hence the same holds true for the entries Vij , i, j = 1, 2. Now V22 = Θ22 is invertible in −1 Ratq×q + , and so V22 V21 H is analytic on the closed left half plane. Moreover, −1 −1 V21 H∞ ≤ V22 V21 ∞ H∞ ≤ Θ−1 V22 22 Θ21 ∞ H∞ < 1. −1 By the maximum modulus principle, this gives V22 (λ)V21 (λ)H(λ) < 1 for λ in −1 + Iq is invertible for each λ the closure of Cleft . It follows that V22 (λ)V21 (λ)H(λ) −1 in the closed left half plane, and that the function V22 (λ)V21 (λ)H(λ) + Iq )−1 is −1 again analytic on the closed left half plane. Thus V22 V21 H + Eq is invertible in q×q Ratq×q too. Combining these facts + . By assumption, V22 is invertible in Rat+ q×q we obtain that V21 H + V22 belongs to Rat+ and is invertible in Ratq×q + . But p×q then FV H belongs to Rat+ , as desired. Next, consider K = FV H. Using (18.17), we see that R + K = FΘ H. Since p×q Ratp×q and FΘ maps Ratp×q into itself (by Theorem 18.3), +, B is a subset of RatB B we know that R + K belongs to Ratp×q , that is, R + K∞ < 1. Thus (18.19) is B proved

Part 2. Take K ∈ Ratp×q + , and suppose R + K∞ < 1. Since R + K belongs p×q to RatB , we know from Theorem 18.3 that there exists a unique H in Ratp×q B such that FΘ H = R + K. In fact, by (18.12), the function in question is H = FΘ−1 (R + K). Furthermore, according to (18.17), the equality FΘ H = R + K yields FV H = K. Note that H has no poles on iR ∪ {∞}. The main diﬃculty is to show that H is analytic on the open left half plane Cleft. From (18.15) and H = FΘ−1 (R + K) we know that −1 H = FΘ−1 (R + K) = Θ∗11 (R + K) − Θ∗21 − Θ∗12 (R + K) + Θ∗22 . Put H1 = Θ∗11 (R + K) − Θ∗21 ,

H2 = −Θ∗12 (R + K) + Θ∗22 .

Then H2 is invertible in Ratq×q and H = H1 H2−1 . Moreover, R+K K H1 −1 −1 = Θ = V . H2 Eq Eq (p+q)×(p+q)

(18.20)

Since V −1 and K belong to Rat+ and Ratp×q + , respectively, we see and H2 belongs from the second equality in (18.20) that H1 belongs to Ratp×q + to Ratq×q . In other words, H and H are analytic in the open left half plane. 1 2 + Hence, in order to prove that H is analytic on the open left half plane Cleft , it remains to show H2−1 is analytic in Cleft .

358

Chapter 18. Application to the rational Nehari problem Multiplying (18.20) from the left by V we get V21 H1 + V22 H2 = Eq , hence

−1 −1 −1 −1 V22 = V22 (V21 H1 + V22 H2 ) = V22 (V21 H + V22 )H2 = (V22 V21 H + Eq )H2 .

Now introduce the scalar rational functions f (λ)

=

g(λ)

=

det V22 (λ)−1 , det V22 (λ)−1 V21 (λ)H(λ) + Iq ,

h(λ)

=

det H2 (λ).

Then f = gh. Also f, g and h have no poles or zeros on iR ∪{∞}. This allows us to use winding number arguments (see Section IV.5 in [32]; also [53], pages 143 and 152). For simplicity we write wn◦ (f ) for the winding number around the origin of f , and we use the analogous notation for g and h. Note that wn◦ (f ) is just equal to the diﬀerence of the number of zeros and number of poles (multiplicities taken into account) of f in Cleft , and similarly for wn◦ (g) and wn◦ (h). First observe −1 that, by condition (b) in our theorem, both V22 and V22 are analytic in the closed left half plane. Thus f has no zeros or poles in the closed left half plane, which implies that wn◦ (f ) = 0. Since −1 −1 V22 V21 H∞ ≤ V22 V21 ∞ H∞ < 1,

it follows that g is analytic on the closed left half plane and has no zeros in the closed left half plane. Thus wn◦ (g) is also zero. The fact that f = gh implies that wn◦ (f ) is the sum of wn◦ (g) and wn◦ (h). Hence wn◦ (h) = 0. We already know that h is analytic on the closed left half plane. Thus wn◦ (h) = 0 tells us that h has no zeros on the closed left half plane. This implies H2 is analytic on Cleft , and hence the same holds true for H. Next we present a more general version of Theorem 18.4. In this more general version K ∈ Ratp×q is not supposed to be analytic on the open left half plane Cleft but K is required to have a prescribed number of poles in Cleft . To state the result we need the following terminology. Let F ∈ Ratp×q . By the number of poles of F in the open left half plane, multiplicities taken into account, we mean the nonnegative integer

δ(F ; λ). (18.21) λ ∈ Cleft

Here δ(F ; λ) is the local degree of F at λ deﬁned in the one but last paragraph of Section 8.2. Since δ(F ; λ) is non-zero if and only if λ is a pole of F , the sum in (18.21) is ﬁnite. Theorem 18.5. Let Θ ∈ Rat(p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ), and let V be given by (18.16), where R ∈ Ratp×q . Then V is invertible in Rat(p+q)×(p+q) , and V22 is invertible in Ratq×q . Assume additionally that

18.3. The J-spectral factorization approach (p+q)×(p+q)

(α) V and V −1 belong to Rat+

359

,

−1 (β) V22 belongs to Ratq×q and V22 has precisely κ poles, multiplicities taken + into account, in Cleft.

Then FV is well-deﬁned and one-to-one on Ratp×q +, B . Also p×q | R + K∞ < 1 and K FV Ratp×q +, B = K ∈ Rat

(18.22) has κ poles in Cleft , multiplicities taken into account .

For κ = 0 the above theorem is just Theorem 18.4. To prove Theorem 18.5 one can use the same line of reasoning as in the proof of Theorem 18.4 above. However, the winding number argument employed in the ﬁnal paragraph of the proof of Theorem 18.4 has to be used in a more sophisticated way. For the details we refer to the literature; see, e.g., [86] and the references therein.

18.3 The J-spectral factorization approach In this section we shall exhibit the connection between the rational Nehari problem and J-spectral factorization. From the ﬁnal paragraph of Section 18.1 we know that without loss of generality the tolerance γ can be assumed to be equal to 1. Therefore, in what follows we take γ = 1. Let R be a stable rational p × q matrix function. With R we associate the (p + q) × (p + q) matrix function W given by ¯ ∗ JG(λ), W (λ) = G(−λ) (

where J=

Ip 0

0 −Iq

)

( ,

G(λ) =

Ip 0

(18.23) R(λ) Iq

) .

(18.24)

Note that J is a (p + q) × (p + q) signature matrix. The fact that R is stable implies that G and G−1 are analytic on the closed right half plane (inﬁnity included), and hence the right-hand side of (18.23) is a left J-spectral factorization of W relative to iR. In this section we shall show that the rational Nehari problem for R relative to the imaginary axis is solvable if and only if W admits a right J-spectral factorization of W relative to iR with an additional condition on the inverse of the spectral factor. The ﬁrst step is given by the next proposition. This proposition, which does not involve realizations and does not require R to be stable, will also provide one of the main steps in the proof of Theorem 18.1 which will be given in the next section. Proposition 18.6. Let R be a proper rational p × q matrix function, and consider ¯ ∗ JG(λ), where J and G are deﬁned by (18.24). the factorization W (λ) = G(−λ)

360

Chapter 18. Application to the rational Nehari problem

Assume that W admits a right J-spectral factorization with respect to the imagi¯ ∗ JL+ (λ), with the additional property that the rational nary axis, W (λ) = L+ (−λ) q ×q matrix function in the right lower corner of L−1 + (λ) is biproper and its inverse is analytic on the closed left half plane. Then the rational Nehari problem for R relative to the imaginary axis is solvable. Moreover, all solutions can be obtained in the following way. Partition L−1 + (λ) as a 2 × 2 block matrix function, L−1 + (λ)

=

Y11 (λ)

Y12 (λ)

Y21 (λ)

Y22 (λ)

,

(18.25)

where Y22 (λ) has size q × q. Then all solutions K of the rational Nehari problem for R relative to the imaginary axis are given by −1 K(λ) = − Y11 (λ)H(λ) + Y12 (λ) Y21 (λ)H(λ) + Y22 (λ) ,

(18.26)

where H is any rational p × q matrix function which has all its poles in the open right half plane and satisﬁes H∞ < 1. Finally, there is a one-to-one correspondence between the solution K and the free parameter H. Proof. We shall apply the results of the previous section. Put Ip R(λ) Θ(λ) = L(λ)−1 . 0 Iq Then Θ ∈ Rat(p+q)×(p+q) and Θ is J-unitary on the imaginary axis. Introduce V (λ) = L−1 + (λ). Then Ep −R V = Θ, 0 Eq −1 and thus (18.16) is satisﬁed. From V = L−1 + and the properties of L+ and L+ we see that V satisﬁes all conditions necessary to apply Theorem 18.4. Thus − K ∈ Ratp×q FV Ratp×q | R − K∞ < 1 . + +, B =

This proves that (18.26) indeed describes the set of all solutions of the rational Nehari problem for R relative to the imaginary axis. Since FV is one-to-one on Ratp×q +, B , by Theorem 18.3, we also obtain the one-to-one correspondence between the solutions K and the free parameter H. In Proposition 18.6 we have that W admits a J-spectral factorization W (λ) = ¯ ∗ JL+ (λ) with the additional property that the q × q matrix function in L+ (−λ) the right lower corner of L−1 + is biproper and has an analytic inverse on the closed left half plane. This property, which involves an inverse of a block of the inverse of L+ , can be replaced by the following more simple condition: the p × p matrix

18.4. Proof of the main result

361

function in the left upper corner of L+ is biproper and its inverse is analytic in the closed left half plane. To see this, write ⎤ ⎤ ⎡ ⎡ L11 (λ) L12 (λ) X11 (λ) X12 (λ) ⎦, ⎦. ⎣ L−1 L+ (λ) = ⎣ + (λ) = L21 (λ) L22 (λ) X21 (λ) X22 (λ) A straightforward Schur complement argument gives that L−1 11 is analytic in the −1 closed left half plane if and only if X22 is analytic in the closed left half plane. Indeed, from Section 2.2 in [20] we have that −1 (λ) X22

=

L22 (λ) − L21 (λ)L−1 11 (λ)L12 (λ),

L−1 11 (λ)

=

−1 X11 (λ) − X12 (λ)X22 (λ)X21 (λ).

This observation will be used in the ﬁnal chapter to smoothen the phrasing of several theorems.

18.4 Proof of the main result Proof of Theorem 18.1. We split the proof into ﬁve parts. Throughout this section we take γ = 1. As has been explained in the ﬁnal paragraph of Section 18.1, this can be done without loss of generality. Furthermore, in what follows R is the strictly proper p × q rational matrix function given by formula (18.2). Part 1. Let K be a solution of the rational Nehari problem for R relative to the imaginary axis. Deﬁne F to be the p × q rational matrix function on iR given by F (iλ) = K(iλ) − R(iλ). Note that F is continuous on the imaginary axis, limλ∈R, |λ|→∞ F (iλ) exists and is equal to a p × q matrix D, say. Furthermore, F ∞ = sup F (iλ) < 1. λ∈R

Now, since K is analytic, the Hankel operator generated by F is equal to the Hankel operator generated by −R, that is, HF = HK−R = HR and HF < 1 (see, e.g., Section XII.2 in [51]). So HR < 1, and hence, by Lemma 18.2, the matrix I − P 1/2 QP 1/2 is positive deﬁnite. In the remaining Parts 2–5 of the proof it is assumed that I − P 1/2 QP 1/2 is positive deﬁnite. We show that under this condition the Nehari problem is solvable and we derive all its solutions. The main work is done in Parts 3 and 4. Part 2 has a preliminary character, and in Part 5 we ﬁnish the proof by applying Proposition 18.6. Part 2. As a ﬁrst step we show that I − P 1/2 QP 1/2 is positive deﬁnite implies that I − Q1/2 P Q1/2 is positive deﬁnite too. To see this, we argue as follows. Introduce T = Q1/2 P 1/2 . Clearly I − T ∗ T is positive deﬁnite, and hence T is a

362

Chapter 18. Application to the rational Nehari problem

strict contraction (i.e., T < 1). But then so is T ∗ = P 1/2 Q1/2 . Thus, as desired, I − Q1/2 P Q1/2 is positive deﬁnite. Next, put K = Z −1 Q, where Z = I − QP while Q and P are the unique solutions to the Lyapunov equations (18.3). Note that Z is invertible, because the matrix I − P 1/2 QP 1/2 is positive deﬁnite. We claim that K is nonnegative and that the following identity holds: KA + A∗ K = KBB ∗ K − Z −1 C ∗ CZ −∗ .

(18.27)

To prove that K is nonnegative , we use ZQ1/2 = (I − QP )Q1/2 = Q1/2 (I − Q1/2 P Q1/2 ). This yields Z −1 Q1/2 = Q1/2 (I − Q1/2 P Q1/2 )−1 , and hence K = Z −1 Q = Q1/2 (I − Q1/2 P Q1/2 )−1 Q1/2 ≥ 0.

(18.28)

To prove (18.27) we ﬁrst multiply the second identity in (18.3) from the left by Z −1 and from the right by Z −∗ . Using K = Z −1 Q = QZ −∗ , this yields KAZ −∗ + Z −1 A∗ K = −Z −1 C ∗ CZ −∗ . Now observe that KAZ −∗

= KA(I − P Q)−1 = KA I + P (I − QP )−1 Q = KA + KAP Z −1 Q = KA + KAP K.

But then, taking advantage of the ﬁrst identity in (18.3) , we obtain KAZ −∗ + Z −1 A∗ K

= KA + A∗ K + K(AP + A∗ P )K = KA + A∗ K − KBB ∗ K.

Thus KA + A∗ K

=

KAZ −∗ + Z −1A∗ K + KBB ∗ K

=

KBB ∗ K − Z −1 C ∗ CZ −∗ ,

which proves (18.27). ¯ ∗ JG(λ), where J and G are deﬁned by (18.24). It Part 3. Put W (λ) = G(−λ) was already observed that this factorization is a left J-spectral factorization with respect to iR. In this part we prove that W also admits a right J-spectral factorization with respect to iR. To do this we use that I − P 1/2 QP 1/2 is positive deﬁnite and apply Theorem 14.14 with L− (λ) = G(λ).

18.4. Proof of the main result

363

Employing the realization (18.2) of R, one gets Ip R(λ) Ip 0 C L− (λ) = = + (λ − A)−1 0 B . 0 0 Iq 0 Iq So, with = A, A

= B

0 B

,

= C

C

,

0

and the associate main matrix of this − A) −1 B, we have L− (λ) = Ip+q + C(λ × = A −B C obviously coincides with the main matrix A = A. realization A the solutions of For the realization considered here we denote by P and Q are the the equations (14.53) and (14.52), respectively. In other words, P and Q unique solutions of AP + PA∗ = BB ∗ ,

+ QA = C ∗ C. A∗ Q

= −Q. It follows that I − PQ = I − P Q, and therefore (ii) So P = −P and Q 1/2 1/2 implies that I − P Q = (I −P Q) = P (I −P QP 1/2 )P −1/2 is invertible. Hence P is invertible too. I −Q Thus by Theorem 14.14 the rational (p + q) × (p + q) matrix function W ¯ ∗ JL+ (λ), with respect to admits a right J-spectral factorization, W (λ) = L+ (−λ) iR. In fact, for L+ one can take −CP Ip 0 (18.29) + Z −1 (λ + A∗ )−1 C ∗ QB , L+ (λ) = ∗ B 0 Iq where Z = I − QP . Theorem 14.14 also tells us that for this choice of the right J-spectral factor L+ we have −CP Ip 0 −1 − (λ + A∗ )−1 Z −1 C ∗ QB , L+ (λ) = (18.30) B∗ 0 Iq where, as before, Z = I − QP . Now partition L−1 + (λ) as V (λ) =

L−1 + (λ)

=

X11 (λ)

X12 (λ)

X21 (λ)

X22 (λ)

,

(18.31)

where the block in the right lower corner has size q × q. Comparing (18.30) and (18.31) we see that the rational matrix functions Xij , i, j = 1, 2, are precisely the functions given by (18.4)– (18.7). Part 4. In this part, again assuming I − P 1/2 QP 1/2 to be positive deﬁnite, we show that the q × q rational matrix function X22 (λ) in the right lower corner of

364

Chapter 18. Application to the rational Nehari problem

the block matrix in (18.31) has precisely the properties which will allow us to apply Proposition 18.6. Obviously, X22 is biproper. Since the eigenvalues of A are in the open left half plane, those of −A∗ are in the open right half plane as well, and hence X22 is −1 is also analytic analytic on the closed left half plane. It remains to show that X22 on the closed left half plane. From the expression for X22 (λ) we see that −1 X22 (λ) = I + B ∗ (λ − A0 )−1 Z −1 QB,

where A0 = −A∗ + Z −1 QBB ∗ = −A∗ + KBB ∗ , with K as in Part 2 of the present −1 proof. Thus, in order to show that X22 is analytic on the closed left half plane, it suﬃces to show that A0 has all its eigenvalues in the open right half plane. To determine the location of the eigenvalues of A0 we ﬁrst prove that A0 K + KA∗0 = KBB ∗ K + Z −1 C ∗ CZ −∗ .

(18.32)

This identity follows from (18.27). Indeed, using the deﬁnition of A0 , we have A0 K = (−A∗ + KBB ∗ )K = −A∗ K + KBB ∗ K. But then, using (18.27), we see that A0 K + KA∗0 = −A∗ K − KA + 2KBB ∗ K = KBB ∗ K + Z −1 C ∗ CZ −∗ , which proves (18.32). The identity (18.32) implies that A0 does not have pure imaginary eigenvalues. Indeed, suppose A0 has a pure imaginary eigenvalue. Then the same holds true for A∗0 , that is, there is a pure imaginary λ0 and a non-zero vector x such that A∗0 x = λ0 x. This implies x∗ A0 = −λ0 x∗ , and hence x∗ (A0 K + KA∗0 )x = 0. From (18.32) it then follows that x∗ KBB ∗ Kx = 0. In other words, x∗ KB = 0. Using the deﬁnition of A0 , we see that −λ0 x∗ = x∗ A0 = −x∗ A∗ + x∗ KBB ∗ = −x∗ A∗ . We conclude that A∗ has a pure imaginary eigenvalue which is impossible because by assumption A (and hence A∗ too) has all its eigenvalues in the open left half plane. Thus a contradiction has been obtained, and we conclude that A0 has no pure imaginary eigenvalue. It remains to show that A0 has no eigenvalues in the open left half plane. If K would be invertible, then K would be positive deﬁnite, and the statement that A0 has no eigenvalues in the open left half plane would now follow immediately from A0 K + KA∗0 ≥ 0 and the classical Carlson-Schneider inertia theorem (see Theorem 13.1.3 in [107]). However since K may not be invertible an additional argument is required, which will be presented in the next two paragraphs. Let n be the order of the square matrix A. Note that K, Q, and Z are also square matrices of order n. Put X1 = Im K and X2 = Ker K. Since K is selfadjoint,

18.4. Proof of the main result

365

we have the orthogonal direct sum decomposition Cn = X1 ⊕ X2 . The identity K = Z −1 Q implies that Ker Q = Ker K. Hence, by selfadjointness, Im Q = Im K. It follows that relative to the decomposition Cn = X1 ⊕ X2 the matrices K and Q admit the following 2 × 2 block matrix representation: Q1 0 K1 0 , Q= , K= 0 0 0 0 where both K1 and Q1 are positive deﬁnite. Next, we partition A, B, and C relative to the decomposition Cn = X1 ⊕ X2 . This yields B1 A11 0 , B= , C = [ C1 0 ]. A= B2 A21 A22 Here we used that X2 = Ker Q = Ker (C|A), which implies that X2 is A-invariant and that C is zero on X2 . From ZK = Q, we see that Z[Im K] = Im Q, and hence Z[X1 ] = X1 . Thus. relative to Cn = X1 ⊕ X2 , the matrix Z partitions as Z11 Z12 , Z= 0 Z22 where both Z11 and Z22 are invertible. Employing the block matrix representations for K, A and B we compute A0 . We have ∗ K1 0 B1 B1∗ B1 B2∗ A11 A∗21 + . A0 = − 0 A∗22 0 0 B2 B1∗ B2 B2∗ Thus A0 has the form

A0 =

A0,11

0

A0,22

,

where A0,11 = −A∗11 + K1 B1 B1∗ and A0,22 = A∗22 . Since A has all its eigenvalues in the open left half plane, the same holds true for A22 . Hence A0,22 has all its eigenvalues in the open right half plane. Thus, in order to prove that A0 has all its eigenvalues in the open right half plane, it suﬃces to show that A0,11 has this property. This will be done in the next paragraph. Since A0 has no pure imaginary eigenvalue, the same holds true for A0,11 . From (18.32), using the block matrices in the previous paragraph, we see that A0,11 K1 + K1 A∗0,11 ≥ 0. As K1 is positive deﬁnite we can now apply the CarlsonSchneider inertia theorem (i.e., Theorem 13.1.3 in [107]) to show that the inertia of A0,11 is equal to the inertia of K1 . Using again that K1 is positive deﬁnite, it follows that all the eigenvalues of A0,11 are in the open right half plane, as desired. Part 5. We are now ready to complete the proof. Assume I −P 1/2 QP 1/2 is positive deﬁnite. By the previous two parts of the proof, the rational matrix function

366

Chapter 18. Application to the rational Nehari problem

¯ ∗ JG(λ) admits a right J-spectral factorization with respect to the W (λ) = G(−λ) ¯ ∗ JL+ (λ), with the additional property imaginary axis, written W (λ) = L+ (−λ) that the q × q matrix function in the right lower corner of L−1 + (λ) is biproper and its inverse is analytic on the closed left half plane. It was also shown that L−1 + (λ) partitions as X11 (λ) X12 (λ) −1 L+ (λ) = , X21 (λ) X22 (λ) where the rational matrix functions Xij , i, j = 1, 2, are precisely the functions given by (18.4)– (18.7). But then we can apply Proposition 18.6 to get the desired description of all solutions. .

18.5 The case of a non-stable given function In this section we return to the general case, where the rational p × q matrix function R is not necessarily stable, i.e., does not necessarily have all its poles in the open left half plane. Throughout we assume R to be proper and to have no poles on the imaginary axis. Write R = R− + R+ , where R− is a stable rational p × q matrix function which is strictly proper, and R+ is a proper rational p × q matrix function which has all its poles in the open right half plane. The required location of the poles determines R− and R+ uniquely. Recall that we seek proper rational p × q matrix functions K such that K has all its poles in the open right half plane and (18.33) R − K∞ = R− − (K − R+ )∞ < γ. The second term in (18.33) gives us a hint of how to solve the Nehari problem for R. In fact, from (18.33) we see that K is a solution to the Nehari problem with tolerance γ for R if and only if K − R+ is a solution to the Nehari problem with tolerance γ for R− . This remark allows us to extend Theorem 18.1 to the case when the given function R is non-stable. To describe the resulting theorem, we shall assume that R− and R+ are given in the form R− (λ) = C− (λIn − A− )−1 B− ,

R+ (λ) = D + C+ (λIn − A+ )−1 B+ , (18.34)

where A− has all its eigenvalues in the open left half plane, and A+ has all its eigenvalues in the open right half plane. In the situation where the realizations in (18.34) are minimal, these conditions on the location of the spectra of A− and A+ are automatically fulﬁlled. Put ∞ ∞ ∗ τ A− ∗ τ A∗ ∗ − e B− B− e dτ, Q− = eτ A− C− C− eτ A− dτ. (18.35) P− = 0

0

Note that P− and Q− are well-deﬁned because all the eigenvalues of A− are in the open left half plane. The following theorem is the main result of this section.

18.5. The case of a non-stable given function

367

Theorem 18.7. Let R = R− + R+ with R− and R+ being given by (18.34). Assume A− and A+ have all their eigenvalues in the open left and open right half plane, respectively, and let P− and Q− be given by (18.35). Then the rational Nehari problem for R relative to the imaginary axis with tolerance γ is solvable if and 1/2 1/2 only if the matrix γ 2 In − P− Q− P− is positive deﬁnite. In this case the matrix 2 Z− = γ In − P− Q− is invertible and all solutions of the Nehari problem under consideration can be obtained in the following way. Introduce rational matrix functions Yij , i, j = 1, 2, by setting ∗ −C+ C− P− + DB− Ip −D Y11 (λ) Y12 (λ) + (18.36) = ∗ Y21 (λ) Y22 (λ) 0 Iq 0 −B− %−1 $ ∗ 0 B+ A+ −B+ B− . . λI2n − −1 ∗ −1 0 −A∗− Z− C− Z− Q− B− Then all solutions K to the rational Nehari problem for R relative to the imaginary axis with tolerance γ are given by −1 K(λ) = − Y11 (λ)H(λ) + Y12 (λ) Y21 (λ)H(λ) + Y22 (λ) , (18.37) where H is any rational p × q matrix function which has all its poles in the open right half plane and satisﬁes H∞ < γ. Moreover, there is a one-to-one correspondence between the solution K and the free parameter H. Proof. From Theorem 18.1 we know that the Nehari problem with tolerance γ for 1/2 1/2 R− is solvable if and only if the matrix γ 2 I − P− Q− P− is positive deﬁnite. On the other hand, we also know (see the second paragraph of this section) that the Nehari problem with tolerance γ for R is solvable if and only if the Nehari problem with tolerance γ for R+ is solvable. These two “if and only if” statements together yield the ﬁrst part of the theorem. 1/2 1/2 Next, assume that the matrix γ 2 I − P− Q− P− is positive deﬁnite. As we have already seen, K is a solution to the Nehari problem with tolerance γ for R + R+ , where K is an arbitrary solution to the if and only if K is of the form K Nehari problem with tolerance γ for R− . By Theorem 18.1, applied to R− in place of R, the latter solutions are given by −1 , K(λ) = − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) with the coeﬃcients in this linear fractional representation given by X11 (λ)

=

−1 ∗ Ip + C− P− (λ + A∗− )−1 Z− C− ,

X12 (λ)

=

−1 C− P− (λ + A∗− )−1 Z− Q− B− ,

X21 (λ)

=

−1 ∗ ∗ −B− (λ + A∗− )−1 Z− C− ,

X22 (λ)

=

−1 ∗ Iq − B− (λ + A∗− )−1 Z− Q− B − ,

368

Chapter 18. Application to the rational Nehari problem

where Z− = γ 2 I − P− Q− , which is invertible. It follows that K(λ)

= = =

=

=

R+ (λ) + K(λ) −1 R+ (λ) − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) −1 R+ (λ) X21 (λ)H(λ) + X22 (λ) X21 (λ)H(λ) + X22 (λ) −1 − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) − X11 (λ) − R+ (λ)X21 (λ) H(λ) + X12 (λ) − R+ (λ)X22 (λ) −1 . X21 (λ)H(λ) + X22 (λ) −1 − Y11 (λ)H(λ) + Y12 (λ) Y21 (λ)H(λ) + Y22 (λ) ,

where H is any rational p × q matrix function having all its poles in the open right half plane and satisﬁes H∞ < γ. Moreover, the coeﬃcient matrix Y11 (λ) Y12 (λ) Y (λ) = Y21 (λ) Y22 (λ) is given by

Y (λ) =

Ip

−R+ (λ)

0

Iq

X11 (λ)

X12 (λ)

X21 (λ)

X22 (λ)

.

Now, using the formulas for Xij , i, j = 1, 2, one gets Ip 0 C − P− X11 (λ) X12 (λ) −1 ∗ = + (λ + A∗− )−1 Z− C− ∗ −B− X21 (λ) X22 (λ) 0 Iq

−1 Z− Q− B − .

Furthermore, employing the realization of R+ , Ip −R+ (λ) Ip −D −C+ (λ − A+ )−1 0 B+ . = + 0 Iq 0 Iq 0 Taking the product of these realizations (see Theorem 2.5) we reach the conclusion that the coeﬃcient matrix Y (λ) admits the desired realization (18.36). The fact that there is one-to-one correspondence between the solution K and the free parameter H in (18.37) follows directly from the corresponding result in Theorem 18.1.

18.6 The Nehari-Takagi problem In the Nehari-Takagi problem the given function R is the same as in the Nehari problem. However the solutions K are allowed to come from a wider class. To

18.6. The Nehari-Takagi problem

369

be more more speciﬁc, let the rational p × q matrix function R be as in the ﬁrst paragraph of Section 18.1. Thus R is proper and does not have a pole on the imaginary axis. Let κ be a non-negative integer. Then the (rational) NehariTakagi problem (relative to the imaginary axis) is the problem of ﬁnding all proper rational p × q matrix functions K such that K has no pole on the imaginary axis and at most κ poles in the open left half plane (multiplicities taken into account), and K − R∞ = sup K(s) − R(s) < γ, (18.38) s∈iR

where γ is a pre-speciﬁed positive number. When κ = 0, the conditions on K reduce to the requirement that K has all its poles in the open right half plane. Thus with κ = 0 the Nehari-Takagi problem is just the Nehari problem considered in the preceding sections. In this section we take γ = 1, which can be done without loss of generality (cf., the last paragraph of Section 18.1), and we assume that R is strictly proper and stable. Thus R admits a realization R(λ) = C(λIn − A)−1 B where A has all its eigenvalues in the open left half plane. The following result is the analogue of Theorem 18.1 for the Nehari-Takagi problem. Theorem 18.8. Let (λ) = C(λIn − A)−1 B be a realization of the rational p × q matrix function R, assume A has all its eigenvalues in the open left half plane, and let ∞ ∞ ∗ sA ∗ sA∗ P = e BB e ds, Q= esA C ∗ CesA ds 0

0

(i.e., P and Q are the controllability and observability gramians corresponding to the given realization). Suppose In − P Q is invertible. Then the rational NehariTakagi problem for R relative to the imaginary axis with γ = 1 is solvable if and only if the matrix P Q has at most κ eigenvalues (multiplicities taken into account) larger than 1. Moreover, if κ0 is the number of eigenvalues of P Q larger than 1, then all solutions K of the Nehari-Takagi problem for R relative to the imaginary axis with γ = 1 such that K has precisely κ0 poles in the open left half plane are given by the linear fractional formula −1 . (18.39) K(λ) = − Θ11 (λ)G(λ) + Θ12 (λ) Θ21 (λ)G(λ) + Θ22 (λ) Here the free parameter G is an arbitrary rational p × q matrix function which has all its poles in the open right half plane and G∞ < 1. Furthermore, the coeﬃcients Θij , i, j = 1, 2, are given by Θ11 (λ)

= Ip + CP (λIn + A∗ )−1 (In − QP )−1 C ∗ ,

Θ12 (λ)

= CP (λIn + A∗ )−1 (In − QP )−1 QB,

Θ21 (λ)

= −B ∗ (λIn + A∗ )−1 (In − QP )−1 C ∗ ,

Θ22 (λ)

= Iq − B ∗ (λIn + A∗ )−1 (In − QP )−1 QB.

370

Chapter 18. Application to the rational Nehari problem

To prove the above theorem one can follow the same line of reasoning as used in this chapter to prove Theorem 18.1. The role of Theorem 18.4 has to be taken over by Theorem 18.5. For further details we refer to the literature; see for example [86] and the references therein.

Notes The Nehari problem has its roots in the classical papers of Nehari [114] and Adamjan-Arov-Krein [1], [2]. The rational matrix version played an important role in the early development of H-inﬁnity control theory; see, e.g., the lecture notes [43]. Here one already ﬁnds the J-spectral factorization approach. For an overview of the various methods to deal with the matrix Nehari problem we refer to the notes to Chapter 20 in [7]. The Takagi version of the Nehari problem has its roots in [142]. The result with a full proof can also be found in Section 20.5 of [7]. For an abstract approach to the Nehari-Takagi problem, covering applications to time-invariant inﬁnite-dimensional systems and time-varying ﬁnite-dimensional linear systems, we refer to [86].

Chapter 19

Review of some control theory for linear systems In this chapter a brief survey is given of a number of basic elements of control and mathematical systems theory. The main aim is to give the reader some understanding for the type of problems that will be treated in the ﬁnal chapter. The chapter consists of two sections. Section 19.1 introduces the concepts of stability of systems and the method of feedback to stabilize a system. Section 19.2 deals with the notion of internal stability of a closed loop system. In particular the Youla-Jabr-Bongiorno parametrization of all stabilizing compensators is presented.

19.1 Stability and feedback In this section we consider a causal input-output system Σ as in the ﬁgure below:

u

Σ

y

As usual (cf., Section 2.1) the symbol u denotes the input and y the output. Mathematically input and output are vector-valued functions of a (time) parameter t. Such an input-output system is called externally stable or bounded-input boundedoutput stable (BIBO-stable) if a bounded input u produces a bounded output y, that is, supt≥0 u(t) < ∞ implies supt≥0 y(t) < ∞.

372

Chapter 19. Review of some control theory for linear systems

Now let us assume that Σ is a causal linear time invariant system given by the following ﬁnite dimensional state space representation: x (t) = Ax(t) + Bu(t), (19.1) y(t) = Cx(t) + Du(t), t ≥ 0. Here A, B, C, D are matrices of appropriate sizes, and A is a square matrix. We refer to (19.1) as a realization of the system. The realization (19.1) is called stable if for any initial value x(0), with zero input u, the state x(t) will go to zero if t → ∞. It is easily seen that stability of the realization (19.1) is equivalent to the requirement that the matrix A has all its eigenvalues in the open left half plane. If the latter holds, A is said to be a stable matrix . Given (19.1) the eﬀect of inputs on outputs can be described in the time domain by a lower triangular integral operator y(t) = CetA x(0) +

t 0

k(t − s)u(s) ds + Du(t),

(19.2)

where k(t) is the so-called impulse response function. As we have already seen in Section 2.1, in the frequency domain with x(0) = 0 the connection between input and output is given by y(λ) = W (λ) u(λ), where W is the transfer function of the system, and u and y denote the Laplace transforms of the input u and the output y, respectively. In terms of (19.1) we have k(t) = CetA B,

W (λ) = D + C(λ − A)−1 B.

(19.3)

From (19.2) and the ﬁrst identity in (19.3) it is clear that stability of the realization (19.1) implies external stability of the corresponding system. The converse is also true when the realization is minimal, that is, when the pair (A, B) is controllable and the pair (C, A) is observable. We summarize this and related results in the following theorems. Theorem 19.1. Let (19.1) be a minimal realization, then the corresponding system is externally stable if and only if the realization is stable. Theorem 19.2. Let k be the impulse response function and let W be the transfer function of the linear time invariant system given by (19.1). The following statements are equivalent: 1. The system given by (19.1) is externally stable; ∞ 2. k(t) dt < ∞; 0

3. The rational matrix function W is iR-stable, that is, W has all its poles in the open left half plane.

19.1. Stability and feedback

373

An important issue is stabilizing an unstable system. The simplest method is that of static state feedback. To explain this method consider the system given by the state space representation: x (t) = Ax(t) + Bu(t), y(t)

=

t ≥ 0.

x(t),

Note that the output is equal to the state. This case is sometimes referred to as the full information case. The problem is to ﬁnd a static feedback control law u(t) = F x(t) + v(t) that will make the system sending v to x stable. That is, to ﬁnd a matrix F of appropriate size such that x (t) = (A + BF )x(t) + Bv(t) is stable. This amounts to requiring that the matrix A + BF is stable, i.e., all its eigenvalues are in the open left half plane. For such a matrix F to exist the pair (A, B) should be stabilizable in the sense of Section 13.2. Two questions appear: ﬁrst, when is a pair of matrices (A, B) stabilizable, and second, how to construct a stabilizing matrix F ? We start with an observation concerning the so-called single input case. In that situation, the matrix B is an n × 1 vector, and one may assume without loss of generality that A and B have the form ⎡ ⎤ ⎤ ⎡ 0 0 1 ⎢ ⎥ ⎥ ⎢ .. ⎢ ... ⎥ ⎥ ⎢ . ⎢ ⎥ ⎢ B=⎢ ⎥ A=⎢ ⎥, ⎥. ⎣0⎦ ⎣ 1 ⎦ −an

···

Consider F = [ fn · · · f1 ]. Then ⎡ ⎢ ⎢ A + BF = ⎢ ⎢ ⎣

···

0

−a1

1

1

⎤ ..

fn − an

···

.

···

1 f1 − a1

⎥ ⎥ ⎥. ⎥ ⎦

So, in this case, any polynomial can be obtained as the characteristic polynomial of A + BF by an appropriate choice of F . Next we make a second observation. Let A be an n × n matrix, let B be an n × m matrix, and write Cn = Im (A|B) X0 . With respect to this direct sum decomposition, the matrices A and B can be written as A11 A10 B1 , A= , B= 0 A00 0

374

Chapter 19. Review of some control theory for linear systems

with (A11 , B1 ) controllable. Thus, for any m × n matrix F = F1 A11 + B1 F1 A10 + B1 F0 , A + BF = 0 A00

F0 , one has

and hence σ(A+BF ) = σ(A11 +B1 F1 ) ∪ σ(A00 ). Note that σ(A00 ), the second part in the right-hand side of the preceding identity, is independent of the particular choice of X0 and also of the choice of F . Therefore the eigenvalues of A00 are called the uncontrollable eigenvalues of A relative to the matrix B. Clearly, A has no uncontrollable eigenvalues relative to B if and only if the pair (A, B) is controllable. From the discussion in the previous paragraph we conclude that, in order for (A, B) to be stabilizable, it is necessary that the uncontrollable eigenvalues of A relative to B are in the open left half plane. The converse of this observation would follow if any controllable pair is stabilizable. This is the case for single input as we have already seen. That it is true in general appears from the next result which is actually quite a bit stronger, and is known as the pole placement theorem. Theorem 19.3. Let A be an n × n matrix, and let B be an n × m matrix. The following two statements are equivalent: (i) The pair (A, B) is controllable; (ii) For any scalar polynomial p(λ) = λn + p1 λn−1 + · · · + pn−1 λ + pn , there is an m × n matrix F such that the characteristic polynomial of A + BF coincides with p. Corollary 19.4. Let A be an n × n matrix and let B be an n × m matrix. The pair (A, B) is stabilizable if and only if the uncontrollable eigenvalues of A relative to the matrix B are in the open left half plane. Let A be an n × n matrix and let C be an m × n matrix. The pair (C, A) is called detectable when there exists an n × m matrix R such that A − RC is stable. In other words the pair (C, A) is detectable if and only if the pair (A∗ , C ∗ ) is stabilizable. By deﬁnition the unobservable eigenvalues of A relative to C are the uncontrollable eigenvalues of A∗ relative to C ∗ . It is also possible to give a direct deﬁnition of the latter notion, involving a decomposition of the type Cn = Ker (C|A) X0 . From the above deﬁnitions and Corollary 19.4 it is clear that the pair (C, A) is detectable if and only if the unobservable eigenvalues of A relative to the matrix C are in the open left half plane.

19.2 Parametrization of internally stabilizing compensators In this section G is the transfer function of a system Σ with two inputs u and w, and two outputs y and z. Here u is the control input, w a disturbance, y is the

19.2. Parametrization of internally stabilizing compensators

375

output which can be measured and z is the output to be controlled. Throughout, we shall assume that the system Σ is given in state space form as follows: ⎧ x (t) = Ax(t) + B1 w(t) + B2 u(t), ⎪ ⎪ ⎨ z(t) = C1 x(t) + D1 u(t), (19.4) ⎪ ⎪ ⎩ t ≥ 0. y(t) = C2 x(t) + D2 w(t), It will be convenient to rewrite the realization (19.4) in the form ⎧ ⎪ w(t) ⎪ ⎪ , x (t) = Ax(t) + B1 B2 ⎪ ⎪ ⎪ u(t) ⎨ ⎪ ⎪ z(t) ⎪ ⎪ ⎪ = ⎪ ⎩ y(t)

C1 C2

x(t) +

0

D1

D2

0

w(t) u(t)

.

From the latter representation we see that the transfer function of (19.4) is given by G11 (λ) G12 (λ) 0 D1 C1 G(λ) = = + (λ − A)−1 B1 B2 . D2 0 C2 G21 (λ) G22 (λ) In particular, the transfer function G22 is strictly proper. Let C be a causal ﬁnite dimensional linear time invariant system of the type considered in the previous section, and let K be its transfer function. Thus K is a proper rational matrix function To deﬁne what it means that C is an internally stabilizing compensator for Σ we introduce two additional inputs v1 and v2 as in the following ﬁgure:

These two additional inputs are regarded as disturbances: v1 is a disturbance on the control input u, while v2 is a disturbance on the measured output. Then the system

376

Chapter 19. Review of some control theory for linear systems

C with transfer function K is said to be an internally stabilizing compensator for the system Σ if the nine transfer functions from the disturbances w, v1 , v2 to z, u and y are all stable rational matrix functions. In this case, by slight abuse of terminology, we shall also say that K is an internally stabilizing compensator for the transfer function G of Σ. After Laplace transform, the nine transfer functions from the disturbances w, v1 , v2 to z, u and y are given by ⎡

zˆ

⎤

⎡

I

⎢ ⎥ ⎢ ⎢ u ⎥ ⎢ ⎣ ˆ ⎦=⎣ 0 0 yˆ

−G12 I −G22

0

⎤−1 ⎡

⎥ −K ⎥ ⎦ I

⎢ ⎢ ⎣

G11

0

0

I

G21

0

0

⎤⎡

w ˆ

⎤

⎥ ⎥⎢ ⎥ ⎢ 0 ⎥ ⎦ ⎣ vˆ1 ⎦ . I vˆ2

(19.5)

Now G22 is strictly proper and K is proper. Hence the rational matrix functions I − G22 (λ)K(λ) and I − K(λ)G22 (λ) are biproper with the value I at inﬁnity. It follows that the inverses I − G22 K and I − KG22 are well-deﬁned. Using these facts, the product of the ﬁrst two matrices in the right-hand side of the identity (19.5)can be computed as ⎡ ⎢ ⎢ ⎣

G11 + G12 K(I − G22 K)−1 G21

G12 (I − KG22 )−1

G12 K(I − G22 K)−1

K(I − G22 K)−1 G21

(I − KG22 )−1

K(I − G22 K)−1

(I − G22 K)−1 G21

G22 (I − KG22 )−1

(I − G22 K)−1

⎤ ⎥ ⎥. ⎦

Theorem 19.5. Let G be the transfer function of the system Σ given by (19.4), and let C be a causal ﬁnite dimensional linear time invariant system whose transfer function K is a proper rational matrix function. Then C is an internally stabilizing compensator for Σ if and only if K stabilizes G22 in the sense that the transfer functions from v1 and v2 to u and y are stable rational matrix functions, that is, the four functions (I − KG22 )−1 ,

K(I − G22 K)−1 ,

G22 (I − KG22 )−1 ,

(I − G22 K)−1 ,

are stable. There is a beautiful parametrization of all internally stabilizing compensators, known as the Youla-Jabr-Bongiorno parametrization. In order to state the parametrization we need a doubly coprime factorization of G22 , that is, a factorization !(λ)−1 N (λ), G22 (λ) = N (λ)M (λ)−1 = M (19.6) and M ! are iR-stable rational matrix functions of appropriate sizes, where N, M, N with the additional property that there exist iR-stable rational matrix functions

19.2. Parametrization of internally stabilizing compensators and Y such that X, Y, X X(λ) −Y (λ) M (λ) (λ) −N

!(λ) M =

N (λ) M (λ)

Y (λ)

N (λ)

X(λ)

Y (λ)

377

X(λ)

X(λ)

−Y (λ)

(λ) −N

!(λ) M

(19.7)

= I.

Such a factorization always exists, in fact we can readily give formulas for all matrix functions involved in terms of the realization of G22 . To do this we assume that the realization G22 (λ) = C2 (λI − A)−1 B2 , has two additional properties, namely (C2 , A) is detectable, and (A, B2 ) is stabilizable. That is, there exist matrices F and H such that the matrices AF = A + B2 F and AH = A + HC2 are both stable. Then, one choice of a doubly coprime factorization is given by the functions ⎧ −1 N (λ) = C2 (λ − AF )−1 B2 , ⎪ ⎪M (λ) = I + F (λ − AF ) B2 , ⎪ ⎪ ⎪ ⎪ !(λ) = I + C2 (λ − AH )−1 H, (λ) = C2 (λ − AH )−1 B2 , ⎨M N (19.8) ⎪ ⎪ Y (λ) = −F (λ − AF )−1 H, X(λ) = I − C2 (λ − AF )−1 H, ⎪ ⎪ ⎪ ⎪ ⎩ Y (λ) = −F (λ − AH )−1 H. X(λ) = I − F (λ − AH )−1 B2 , Next, we give the Youla-Jabr-Bongiorno parametrization, which describes all internally stabilizing compensators of Σ in terms of iR- stable, proper rational matrix functions in a one-to-one way. Theorem 19.6. Let G be the transfer function of the system Σ given by (19.4), and let M , N , X, Y be the iR-stable rational matrix functions related to the doubly coprime factorization of G22 . Let C be a causal ﬁnite dimensional linear time invariant system whose transfer function K is a proper rational matrix function. Then C is an internally stabilizing compensator of Σ if and only if K has the form K(λ)

=

−1 Y (λ) − M (λ)Q(λ) X(λ) − N (λ)Q(λ) ,

(19.9)

where Q is an iR-stable rational matrix function. Moreover, the map from Q to K is one-to-one. !, N , X, Y we have the following alternative Replacing M , N , X, Y by M expression for the transfer function K of the compensator: !(λ) . (λ) −1 Y (λ) − Q(λ)M K(λ) = X(λ) − Q(λ)N

378

Chapter 19. Review of some control theory for linear systems

Notes The results of the ﬁrst section are standard results in mathematical systems theory, see, e.g., [94] or the more recent [33], [84]. For analogous results in the discrete time case we refer to [94], Chapter 21 of [150], and to [85]. A proof of Theorem 19.5 can be found in Chapter 4 of [43]. The formulas (19.8) giving the doubly coprime factorization in state space terms were derived in [115], see also Section 4.5 in [43]. Theorem 19.6 presents a result of [148].

Chapter 20

H-inﬁnity control applications The focus of the chapter is on a part of control theory called H-inﬁnity control. The problem involved is the general H-inﬁnity control problem, the so-called standard problem. It concerns the construction of a stabilizing controller with additional constraints on the maximum of the norm of the closed loop transfer function, taken over the values of the argument on the imaginary line. In its simplest form the problem is equivalent to the rational matrix Nehari problem considered in Chapter 18. The label H-inﬁnity is related to the fact that a proper rational matrix function is stable if and only if it is analytic and uniformly bounded in the open right half plane. A function with the latter properties is usually referred to as an H∞ -function (on the right half plane). The chapter consists of four sections. Section 20.1 introduces the standard problem mentioned above, and shows how this problem can be reduced to a model matching problem. In the next two sections we discuss a one-sided model matching problem (Section 20.2) and the two-sided model matching problem (Section 20.3). In particular, it will be shown how these two problems reduce to J-spectral factorization problems involving certain rational matrix functions. All of this will be done in general terms, without any state space formulas as yet. In the ﬁnal section (Section 20.4) we use results from Chapter 14 and present the solution to the model matching problem in state space terms. This leads to the solution of the standard problem in these terms too. In this chapter, as in Section 18.2, we use the following notation: if R is a rational matrix function, then R∗ denotes the rational matrix function given by R∗ (λ) = ¯ ∗ . (In engineering literature, including [76] and [43], this function is often R(−λ) denoted by R∼ .) Recall also from Section 18.2 that Rat denotes the set of all proper rational matrix functions that are analytic on the imaginary axis. Furthermore, Ratp×q stands for the set of all F in Rat that are of size p × q, and Ratp×q + denotes the set of all F in Ratp×q that are analytic on the closed left half plane. In the present chapter we shall also use the notation Ratp×q (Rat− ) which will −

380

Chapter 20. H-inﬁnity control applications

denote the set of all F in Ratp×q (in Rat) that are analytic in the closed right half plane. In other words, F belongs to Ratp×q if and only if F is an iR stable p × q − if and only if F ∗ ∈ Ratq×p rational matrix function. Note also that F ∈ Ratp×q − + .

20.1 The standard problem and model matching Throughout this chapter G is the transfer function of a system Σ with two inputs u and w, and two outputs y and z. The input u is the control input, w is a disturbance, y is the output we can measure, and z is the output to be controlled. As in Section 19.2 we assume that the system is given by the state space representation ⎧ ⎪ ⎪ x (t) = Ax(t) + B1 w(t) + B2 u(t), ⎨ z(t) = C1 x(t) + D1 u(t), (20.1) ⎪ ⎪ ⎩ y(t) = C2 x(t) + D2 w(t), t ≥ 0. In particular, the function G is of the form G(λ)

G11 (λ)

G12 (λ)

G21 (λ)

G22 (λ)

= =

0

D1

D2

0

+

C1 C2

(λ − A)−1 B1

B2 .

(20.2)

Taking Laplace transforms and assuming the system to be at rest at t = 0 we have

z(λ) y(λ)

G11 (λ)

G12 (λ)

G21 (λ)

G22 (λ)

=

w(λ) u (λ)

.

(20.3)

Our goal is to ﬁnd a proper rational matrix function K such that: (1) K is the transfer function of an internally stabilizing compensator C of Σ (see Section 19.2), and (2) the inﬂuence of w on z is kept small in a sense we shall explain presently. Inserting u (λ) = K(λ) y (λ) into (20.3), one sees that ⎧ ⎨z(λ) = G11 (λ)w(λ) + G12 (λ)K(λ) y (λ), (20.4) ⎩y(λ) = G (λ)w(λ) + G22 (λ)K(λ) y (λ). 21 Since G22 is strictly proper, so is G22 K, and hence the determinant of the matrix I − G22 (λ)K(λ) does not vanish identically. By the second equation in (20.4) we −1 have y(λ) = I − G22 (λ)K(λ) G21 (λ)w(λ). Inserting this into the ﬁrst equation

20.1. The standard problem and model matching

381

of (20.4), we obtain that the closed loop transfer function from w to z is given by the Redheﬀer representation (20.5) z(λ) = RG (K)(λ) w(λ) =

−1 G11 (λ) + G12 (λ)K(λ) I − G22 (λ)K(λ) G21 (λ) w(λ).

The second requirement on K is that, given a tolerance γ, we want RG (K) to be in Rat and to satisfy the bound RG (K)∞ = max RG (K)(λ) < γ. λ∈iR

(20.6)

This problem is known in control theory as the standard problem of H-inﬁnity control . The approach to solving this problem using J-spectral factorization techniques starts from the Youla parametrization of internally stabilizing compensators which we reviewed in Section 19.2. This leads, as we shall see in the ﬁnal two paragraphs of this section, to an equivalent and easier to handle problem. Indeed, from the given rational matrix function G one constructs three rational matrix functions, T1 , T2 and T3 such that internally stabilizing compensators for which (20.6) holds are in one-to-one correspondence with iR-stable rational matrix functions Q for which T1 − T2 QT3 ∞ < γ. (20.7) The latter problem is called the model matching problem . It turns out that under mild assumptions (see Section 20.4 below) the rational matrix functions T1 , T2 and T3 are iR stable. In particular, these functions have no poles on the imaginary axis and at inﬁnity, and hence they are all in Rat. Furthermore, we shall see that T2 has a left inverse in Rat and T3 has a right inverse in Rat. A particular case (see the next section) of the model matching problem, when T2 is square and T3 = I, is a variation on the Nehari problem as discussed in Chapter 18. Next, we present the reduction of the standard problem to a model matching problem. All necessary calculations take place in Rat, i.e., in the set of rational matrix functions that are analytic on iR and at inﬁnity. As before, we partition the transfer function G as in the ﬁrst part of (20.3). Also we shall employ the same notation as in Section 19.2 insofar as it concerns the doubly coprime factorization of G22 in (19.6) and the parametrization of the transfer functions of the internally stabilizing compensators of the system Σ in Theorem 19.6. We can then introduce three new functions, namely T1 (λ)

=

G11 (λ) + G12 (λ)M (λ)Y (λ)G21 (λ),

(20.8)

T2 (λ)

=

G12 (λ)M (λ),

(20.9)

T3 (λ)

=

!(λ)G21 (λ). M

(20.10)

382

Chapter 20. H-inﬁnity control applications

Recall that the problem we wish to solve is to ﬁnd, if possible, internally stabilizing compensators C of the system Σ with a proper transfer function K such that RG (K) belongs to Rat and (20.6) is satisﬁed, i.e., RG (K)∞ = max RG (K)(λ) < γ. λ∈iR

Here RG (K) is given by −1 G21 (λ); RG (K)(λ) = G11 (λ) + G12 (λ)K(λ) I − G22 (λ)K(λ)

(20.11)

see (20.5). In case K is given by (19.9) involving the function Q featured there, we can rewrite RG (K) as follows. Theorem 20.1. With K as in (19.9), the closed loop transfer function is given by RG (K)(λ) = T1 (λ) − T2 (λ)Q(λ)T3 (λ), where T1 , T2 and T3 are given by (20.8), (20.9) and (20.10), respectively !(λ)−1 N (λ) and (19.9) into I − G22 (λ)K(λ) −1 , and Proof. Inserting G22 (λ) = M suppressing the variable λ for notational convenience, we get !(X − N Q) − N (Y − M Q) −1 M ! (I − G22 K)−1 = (X − N Q) M =

!. (X − N Q)M

In the actual derivation of these identities, the doubly coprime factorization in (19.6) and the deﬁning properties given by (19.7) are employed. Again using (19.9), !. Substituting this in the formula for we arrive at K(I − G22 K)−1 = (Y − M Q)M the closed loop transfer function (20.11) yields RG (K) = =

!G21 ) − G12 M QM !G21 (G11 + G12 Y M !G21 ) − T2 QT3 . (G11 + G12 Y M

Now from the deﬁning properties of a doubly coprime factorization (19.6) one !. Inserting this in the formula above we obtain that T1 = sees that M Y = Y M ! G11 + G12 Y M G21 . This completes the proof.

20.2 The one-sided model matching problem In this section we consider the model matching problem (20.7) with T1 ∈ Ratl×p , T2 ∈ Ratl×q and T3 = Ip . Furthermore, we assume that T2 has a left inverse in − Ratq×l . In particular, T1 is analytic on the imaginary axis (with inﬁnity included) and T2 is iR-stable. Note that the left invertibility of T2 implies that l ≥ q, that is, T2 is a “tall” matrix.

20.2. The one-sided model matching problem

383

Given T1 and T2 as in the previous paragraph, the problem is to ﬁnd necessary and suﬃcient conditions for the existence of an iR-stable rational q × p matrix function Q, i.e., Q ∈ Ratq×p − , such that T1 − T2 Q∞ < γ, and to give a full parametrization of all such Q. We refer to this problem as the one-sided model matching problem corresponding to T1 and T2 . We shall explain how this problem reduces to the Nehari problem, and we shall present a necessary and suﬃcient condition for its solution in terms of a J-spectral factorization. The following theorem is the main result of this section. Theorem 20.2. Let T1 ∈ Ratl×p and T2 ∈ Ratl×q be given, and assume T2 has a − left inverse in Rat. Let γ > 0, and put ∗ Il T2 (λ) T1 (λ) T2 (λ) 0 0 , Υ(λ) = T1∗ (λ) Ip 0 −γ 2 Ip 0 Ip J

=

Iq

0

0

−Ip

.

such the norm constraint T1 − T2 Q∞ < γ is Then there exists Q ∈ Ratq×p − satisﬁed if and only if Υ admits a left J-spectral factorization Υ(λ) = W ∗ (λ)JW (λ),

(20.12)

with respect to the imaginary axis having the additional property that the q × q block in the left upper corner of W (λ) has an inverse in Ratq×q − . Moreover, writing 2 −1 W (λ) = ωij (λ) i,j=1 , where ω11 (λ) and ω22 (λ) are of sizes q × q and p × p, respectively, all solutions Q of the one-sided model matching problem corresponding to T1 and T2 are given by −1 , Q(λ) = − ω11 (λ)U (λ) + ω12 (λ) ω21 (λ)U (λ) + ω22 (λ)

(20.13)

with U ∞ < 1. where U is a rational matrix function in Ratq×p − and has a left inverse in Rat, we know from Proof. Since T2 belongs to Ratl×q − Theorem 17.26 that T2 admits an inner-outer factorization with an invertible outer factor. Thus T2 = V X, where V is inner, and both X and X −1 are analytic in the closed right half plane. If T2 happened to be square, the reduction to the Nehari problem would now be easy. Indeed, in that case V is bi-inner, and hence ∞, T1 − T2 Q∞ = T1 − V XQ∞ = V ∗ T1 − XQ∞ = R − Q = XQ. Actually, since both X and Q are in Rat− , also Q where R = V ∗ T1 and Q is in Rat− . Thus, this is not quite the Nehari problem as presented in Chapter 18, ∗ . Also note that R∗ is not but applying the results of Section 18.3 to R∗ yields Q

384

Chapter 20. H-inﬁnity control applications

stable, but it is just in Rat. At this point we use the fact that Proposition 18.6, when applied to R∗ , does not require R∗ to be stable. Recall that this point was made explicitly in the paragraph preceding the statement of Proposition 18.6. However, in general, T2 is only left invertible and not square, in which case V is only inner and not bi-inner. To deal with this more general case, we proceed = V V is bi-inner. We as follows (see Section 17.8): take V such that U has the same McMillan degree as V , that is, in the way choose V such that U outlined in Section 17.8. Then V∗ XQ T1 − T2 Q∞ = ∞ . T1 − (V )∗ 0 It follows that T1 − T2 Q∞ < γ if and only if for each for λ ∈ iR ∪ {∞} the following two conditions hold: (a) Φ(λ) = γ 2 Ip − T1∗ (λ)V (λ)(V )∗ (λ)T1 (λ) > 0, (b) γ 2 Ip − T1∗ (λ)V (λ)(V )∗ (λ)T1 (λ) −(V ∗ (λ)T1 (λ) − X(λ)Q(λ))∗ (V ∗ (λ)T1 (λ) − X(λ)Q(λ)) > 0. Using (a), the inequality (b) can be reduced to Φ−(V ∗ T1 −XQ)∗ (V ∗ T1 −XQ) > 0 where, for notational convenience, the variable λ being suppressed. Now, let Φ(λ) = N ∗ (λ)N (λ) be a left canonical factorization of Φ relative to the imaginary axis. Then condition (b) above is equivalent to Ip − N −∗ (V ∗ T1 − XQ)∗ (V ∗ T1 − XQ)N −1 > 0, i.e., to V ∗ T1 N −1 − XQN −1∞ < 1. Observe that this, in turn, is precisely an = XQN −1 . instance of Nehari’s problem, with R = V ∗ T1 N −1 and Q We apply the Nehari problem to R. Applying the result of Section 18.3, in particular Proposition 18.6 (which we apply with left half plane and right half plane interchanged) one sees that this Nehari problem is solvable if and only if the function Ψ(λ), deﬁned by Iq Iq V ∗ (λ)T1 (λ)N −1 (λ) Iq 0 0 , Ψ(λ) = N −∗ (λ)T1∗ (λ)V (λ) Ip 0 −Ip 0 Ip has a left J-spectral factorization of the form ) ( 0 Iq ∗ L− (λ), Ψ(λ) = L− (λ) 0 −Ip

(20.14)

with the additional property that the p × p block entry in the right lower corner p×p −1 of L−1 − has an inverse in Rat− . Moreover, in that case, if we partition L− (λ)

20.2. The one-sided model matching problem

385

2 Lij (λ) i,j=1 , with L11 a q × q rational matrix function, then all as L−1 − (λ) = solutions to this Nehari problem are given by Q(λ) = −(L11 (λ)U (λ) + L12 (λ))(L21 (λ)U (λ) + L22 (λ))−1 , where U runs over all functions in Ratq×p for which U ∞ < 1. Finally, recall − (see the ﬁnal paragraph of Section 18.3) that the additional property of the p × p block entry in the right lower corner of L−1 − is equivalent to the q × q block entry in the left upper corner of L− having an inverse in Ratp×p − . Put Q(λ) = X −1 (λ)Q(λ)N (λ). From the results of the previous paragraph, we get that all solutions to the one-sided model matching problem are given by Q(λ) = − X −1 (λ)L11 (λ)U (λ) + X −1(λ)L12 (λ) −1 , · N −1 (λ)L21 (λ)U (λ) + N −1 (λ)L22 (λ) where U runs over all functions in Ratq×p for which U ∞ < 1. − Next, introduce W (λ) = L− (λ)

X(λ)

0

0

N (λ)

.

(20.15)

Note that the q × q block entry in the left upper corner of W has an inverse in Ratq×q − . Furthermore, W

−1

(λ) =

X −1 (λ)L11 (λ)

X −1 (λ)L12 (λ)

N −1 (λ)L21 (λ)

N −1 (λ)L22 (λ)

.

So all solutions are parametrized by the function W −1 . It remains to establish the identity (20.12), that is, once more suppressing the variable λ, Υ = W ∗ JW. Let us denote the right side of the previous identity by Ξ. Thus Ξ = W ∗ JW . Using the deﬁnition of W in (20.15) together with formula (20.14), we see that Ξ=

X∗

0

0

N∗

Ψ

X

0

0

N

.

386

Chapter 20. H-inﬁnity control applications

It follows that Ξ= = = = = = = =

X∗

0

T1∗ V

N∗

X∗

0

T1∗ V

N∗

T2∗

0

T1∗ V V ∗

I

T2∗

0

T1∗ V V ∗

I

I

0

0

−I

V ∗V

0

0

−I

X

V ∗ T1

0

N

X

V ∗ T1

0

N

I

0

0

−N ∗ N

T2

V V ∗ T1

0

I

I

0

0

−γ 2 I + T1∗ V (V )∗ T1 T2∗ V V ∗ T1

T1∗ V V ∗ T2

−γ 2 I + T1∗ (V V ∗ + V (V )∗ )T1 X ∗ V ∗ T1

T1∗ V X

−γ 2 I + T1∗ T1 T2∗ T1

T1∗ T2

−γ 2 I + T1∗ T1

T1∗

I

I

0

0

−γ 2 I

0

I

T2∗ T2

V V ∗ T1

T2∗ T2

0

T2

T2∗ T2

T2∗

T2

T1

0

I

= Υ.

Thus we conclude that we may obtain W from a J-spectral factorization of a function that is easily described in terms of T1 and T2 , as desired. Note also that the positivity of γ 2 − T1∗ V (V )∗ T1 on iR ∪ {∞} is implied by the J-spectral factorization.

20.3 The two-sided model matching problem In this section we extend the analysis of the previous section to the two-sided model matching problem. It will turn out that in this case we need two J-spectral factorizations. l×q Theorem 20.3. Let T1 ∈ Ratl×p and T3 ∈ Ratm×p . Assume that − , T2 ∈ Rat− − T2 has a left inverse in Rat, and T3 has a right inverse in Rat. Let γ > 0, and

20.3. The two-sided model matching problem put Ω(λ) =

T3 (λ)

0

T1 (λ)

Il

387

Ip

0

0

−γ 2 Il

T3∗ (λ)

T1∗ (λ)

0

Il

.

(20.16)

Then there exists Q ∈ Ratq×m such that T1 − T2 QT3 < γ if and only if two − conditions (i) and (ii) hold. The ﬁrst condition (i) is as follows: (i) With respect to the imaginary axis, Ω admits a right J-spectral factorization 0 Im ∗ ¯ , where J = Ω(λ) = V (λ)JV (−λ) , (20.17) 0 −Il having the additional property that the m × m block in the upper left-hand . corner of V has an inverse in Ratm×m − With V as in (20.17), deﬁne 0 −T2∗(λ) −Im Ω(λ) = V −∗ (λ) I 0 0

0 Il

V −1 (λ)

0

I

−T2 (λ)

0

.

(20.18)

Then the second condition (ii) is: admits a left J-spectral factorization (ii) With respect to the imaginary axis, Ω of the form 0 I q ¯ ∗ JW (λ), where J = Ω(λ) = W (−λ) , (20.19) 0 −Im having the additional property that the q × q block in the upper left-hand corner of W has an inverse in Ratq×q − . Moreover, when (i) and (ii) are satisﬁed, (all ) the solutions Q to the two-sided model matching problem corresponding to T1 , T2 and T3 can be obtained as follows. 2 Partition W −1 = Xij i,j=1 , with X11 a q × q rational matrix function. Then Q = −(X11 U + X12 )(X21 U + X22 )−1 ,

(20.20)

where U is an iR-stable rational q × m matrix function with U ∞ < 1. Proof. The idea of the proof is to reduce the two-sided model matching problem to the one-sided model matching problem discussed in the previous section. The proof is divided into several steps. Part 1. We ﬁrst show that condition (i) in the theorem is a necessary condition. ¯ ∗ and T (λ) = T3 (λ) ¯ ∗ . Note the crucial To this end, introduce T1 (λ) = T1 (λ) 3 diﬀerence with the functions T1∗ and T3∗ : the functions T1 and T3 are analytic in the closed right half plane, inﬁnity included. With the help of these functions,

388

Chapter 20. H-inﬁnity control applications

∞ < γ, where rewrite T1 − T2 QT3 ∞ < γ in the following way: T1 − T3 Q = Q T2 (with the obvious interpretations for these functions). Taking into Q account Theorem 20.2, this gives that the ﬁrst condition is necessary. Indeed, with T3 T1 L= 0 I and V = W , we obtain L

∗

I

0

0

−γ 2 I

L= W

∗

I

0

0

−γ 2 I

W.

Part 2. The next step is to rewrite the two-sided model matching problem in an equivalent way. Use Theorem 17.28 to write T3 (λ) = Y (λ)V1 (λ) where Y is an m × m invertible outer function and V1 is an m × p co-inner function. Let V1 be such that V = V1∗ (V1 )∗ is bi-inner (see Corollary 17.33). Write R = T1 V = T1 V1∗ T1 (V1 )∗ = R1 R2 , where R1 is an l × m and R2 is an l × (p − m) matrix function. As V is bi-inner, T3 V = Y 0 . Thus we have T1 − T2 QT3 ∞ < γ if and only if

R1

R2 − T2 QY

0 ∞ < γ.

In turn, this can be rewritten as ∗ γ 2 Il > R1 (λ) − T2 (λ)Q(λ)Y (λ) R1 (λ) − T2 (λ)Q(λ)Y (λ) + R2 (λ)R2∗ (λ), for all λ ∈ iR ∪ {∞}, or equivalently, suppressing the variable λ again, as γ 2 Il − R2 R2∗ > (R1 − T2 QY )(R1 − T2 QY )∗ . This implies that γ 2 Il − R2 R2∗ > 0, and if we write γ 2 Il − R2 R2∗ = M M ∗ with M and M −1 in Ratl×l − , then we can rewrite the inequality above as Il > M −1 (R1 − T2 QY )(R1 − T2 QY )∗ M −∗ . Thus T1 − T2 QT3 ∞ < γ if and only if the following two conditions hold: γ 2 Il − R2 R2∗ > 0,

M −1 R1 − M −1 T2 QY ∞ < 1.

(20.21)

Note that the last of these two conditions is a one-sided model matching prob. Observe also that M −1 R1 = lem for QY , as both Y and Y −1 are in Ratm×m − l×m −1 ∗ ∗ M T1 V1 is in Rat , because V1 is inner and hence analytic in the closed left half plane, inﬁnity included. Also M −1 T2 is in Ratl×q − . Although we do not know

20.3. The two-sided model matching problem

389

that M −1 R1 is in Ratl×m (that is, we do not know that it is analytic in the closed − right half plane), still all conditions of Theorem 20.2 are met. Thus we may apply Theorem 20.2, to see that solvability of the one-sided model matching problem, which is the second condition in (20.21), is equivalent to a J-spectral factorization problem in the following way. Put −1 M T2 M −1 R1 K= . 0 Im Then, by Theorem 20.2, solvability of the one-sided model matching problem, which (as just noted) is the second part of (20.21), is equivalent to existence of a (m+q)×(m+q) , matrix function P such that P and P −1 are in Rat− ( ) ( ) 0 0 I I K∗ l K = P∗ q P, 0 −Im 0 −Im and, in addition, the q × q-block of P in the upper left corner has an inverse in . Recall that the last condition is equivalent to the requirement that the Ratm×m − m × m-block in the right lower corner of P −1 is in Ratm×m . Moreover, (all) the − solutions Q to the one-sided model matching problem corresponding to M −1 T2 2 and M −1 R1 are generated by P −1 as follows: if P −1 = Pij i,j=1 , with P11 of size q × q, then Q

= −(P11 U + P12 )(P21 U + P22 )−1 Y −1 = −(P11 U + P12 )(Y P21 U + Y P22 )−1 .

Introduce W = P

Iq 0

0 Y −1

.

Then W and W −1 are analytic in the right half plane and the m × m block in the right lower corner of W −1 is equal to Y P22 , which is also in Ratm×m . Finally, W − generates all solutions Q. Let 0 Iq K=K . 0 Y −1 We conclude that solvability of the one-sided model matching problem, which is the second part of (20.21), is equivalent to existence of a J-spectral factorization of the form Iq Il 0 0 ∗ ∗ W =K K, (20.22) W 0 −Im 0 −Im with the additional property that the m × m block in the right lower corner of W −1 is in Ratm×m . −

390

Chapter 20. H-inﬁnity control applications

Part 3. Continuing with the considerations above, we compute −1 M T2 M −1 R1 Iq 0 = K 0 Im 0 Y −1 −M −1

=

M −1R1 Y −1 Y −1

0 −M

=

=

R1

0

Y

0

Y

−M

R1

−T2

∗ K

0

−T2

0

0

Im

0

Im

−T2

0

−1

It follows that

−1

0

Im

.

Il

0

0

−Im

Il

0

0

−Im

K

is equal to

0

−T2∗

Im

0

0

−M ∗

Y∗

R1∗

−1

0

Y

−M

R1

−1

0

Im

−T2

0

,

which, in turn, can be written as,

0

−T2∗

Im

0

$

0

Y

−M

R1

Il 0

0 −Im

0

−M ∗

Y∗

R1∗

%−1

0

Im

−T2

0

.

Now the product of the middle three terms is easily seen to be equal to Y R1∗ YY∗ . R1 Y R1 R1∗ − M M ∗ Observe also that Y Y ∗ = T3 T3∗ and Y R1∗ = Y V1 T1∗ = T3 T1∗ . Furthermore, R1 R1∗ − M M ∗

= R1 R1∗ − γ 2 Il + R2 R2∗ 2

= −γ Il + R1

R2

R1∗ R2∗

= −γ 2 Il + T1 V ∗ V T1∗ = −γ 2 Il + T1 T1∗ .

20.3. The two-sided model matching problem Hence

YY∗

Y R1∗

R1 Y

R1 R1∗ − M M ∗

= =

391

T3 T3∗

T3 T1∗

T1 T3∗

−γ 2 Il + T1 T1∗

T3

0

Ip

T1

I

0

∗ T3

0 −γ 2 IL

T1∗

0

I

= Ω.

Part 4. After these preliminaries we can now complete the proof in one direction. Indeed, to show that both the conditions (i) and (ii) need to be satisﬁed, note that we already saw at the beginning of the proof that (i) is necessary. Assuming that (i) holds, we continue the computation above, with V as in (20.17), and see that ∗

K

Il

0

0

−Im

K

= =

0

−T2∗

Im

0

0

−T2∗

Im

0

$ V V

−∗

Im

0

0

−Il

Im 0

0 −Il

%−1 V

0

Im

−T2

0

∗

V

−1

0

Im

−T2

0

.

Thus, by (20.22), the second condition (ii) is necessary as well. Part 5. For the converse, assume that both (i) and (ii) are satisﬁed. As in the proof of Theorem 20.2, applied to T1 and T3 , in place of T1 and T2 , we see that (i) implies that the ﬁrst condition in (20.21) holds. Now follow the arguments in Parts 3 and 4 backwards to see that also the second condition in (20.21) is met. As we have already seen that these two conditions taken together are equivalent to the two-sided model matching problem, the proof is complete. Note that for the factorization (20.17) we need the analogue of Theorem 14.7 for right J-spectral factorization, applied to the function Ω given by (20.16). This analogue can be obtained by applying the left factorization result of Theorem 14.7 to the function Ω(−λ); cf., the paragraphs immediately following Theorem 14.8. In addition, the analogue of Theorem 14.7 for right J-spectral factorization provides us with a formula for the right J-spectral factor V , satisfying I 0 m ¯ ∗ V (λ). Ω(λ) = V (−λ) 0 −Il ¯ ∗ . We state the result of carrying The function we need will then be V (λ) = V (−λ) out all this in state space form as a lemma, which will be useful in the next section. Lemma 20.4. Let H(λ) = D+C(λIn −A)−1 B be a realization of an (m+l)×(p+l) rational matrix function H. Write J = diag (Ip , −Il ), J = diag (Im , −Il ), and

392

Chapter 20. H-inﬁnity control applications

assume that DJ D∗ = J. Also assume that A has all its eigenvalues in the open ¯ ∗ . Then Ω admits a right J-spectral left half plane. Put Ω(λ) = H(λ)J H(−λ) factorization with respect to the imaginary axis if and only if the algebraic Riccati equation XC ∗ JCX + X(A∗ − C ∗ J −1 DJ B ∗ ) + (A − BJ D ∗ J −1 C)X +BJ D∗ JDJ B ∗ − BJB ∗ = 0 has a Hermitian solution X such that A∗ − C ∗ J −1 (DJ B ∗ − CX) has its eigenvalues in the open left half plane. If X is such a solution (necessarily unique), and V (λ) = Im+l + C(λIn − A)−1 (BJ D∗ − XC ∗ )J −1 , then Ω(λ) = V (λ)JV (−λ)∗ is a right J-spectral factorization of Ω with respect to the imaginary axis.

20.4 State space solution of the standard problem In this section we return to the standard problem. We recall the basic facts about the problem. The starting point is a system in state space form ⎧ x (t) = Ax(t) + B1 w(t) + B2 u(t), ⎪ ⎪ ⎨ z(t) = C1 x(t) + D1 u(t), (20.23) ⎪ ⎪ ⎩ t ≥ 0. y(t) = C2 x(t) + D2 w(t), The input vector u(t) belongs to Cq , the noise vector w(t) belongs to Cp , the state vector x(t) belongs to Cn , the measured output y(t) belongs to Cm , and ﬁnally, the output z(t) to be controlled belongs to Cl . Thus the sizes of the matrices featured in (20.23) are as follows: A is n × n, B1 is n × p, B2 is n × q, C1 is l × n, C2 is m × n, D1 is l × q, and D2 is m × p. Throughout the section we assume that the following simplifying assumptions hold: A1. (A, B1 ) is controllable and (C1 , A) is observable, A2. (A, B2 ) is stabilizable and (C2 , A) is detectable, that is, there are matrices F and H so that both A + B2 F and A + HC2 have all their eigenvalues in the open left half plane. A3. D1∗ C1 = 0, D1∗ D1 = Iq , D2 B1∗ = 0, D2 D2∗ = Im . Given is also γ > 0. The problem we consider is to ﬁnd an internally stabilizing compensator K from y to u such that (20.6) holds. As we have explained in Section 20.1 this problem can be transformed into a model matching problem, using the rational matrix functions T1 , T2 , and T3

20.4. State space solution of the standard problem

393

appearing in (20.8)–(20.10). First we shall use (20.23) to derive state space realizations for T1 , T2 , and T3 . For this purpose we ﬁx matrices H and F such that AF = A+B2 F and AH = A+HC2 are stable matrices. Recall that assumption A2 guarantees the existence of matrices H and F with these properties. It is a matter of straightforward calculations to check that the following proposition holds. Proposition 20.5. Write G22 (λ) = C2 (λIn − A)−1 B2 , and assume assumption A2 is satisﬁed. Let F and H be matrices such that AF = A+ B2 F and AH = A+ HC2 are stable matrices. Suppose a doubly coprime factorization of G22 (λ) is given by the functions in (19.8). Then T1 (λ) =

C1 + D1 F

−D1 F

$ λI2n −

AF

−B2 F

0

AH

%−1

B1 B1 + HD2

,

T2 (λ) = D1 + (C1 + D1 F )(λIn − AF )−1 B2 , T3 (λ) = D2 + C2 (λIn − AH )−1 (B1 + HD2 ). Observe that T1 , T2 and T3 are in Rat− . Next, we show that T2 has a left inverse, while T3 has a right inverse, both in Rat. Lemma 20.6. Under the assumptions A1, A2, A3, the matrix function T2 has a left inverse in Rat and T3 has a right inverse in Rat. Proof. By Corollary 17.27, it suﬃces to show that T2 (λ) is left invertible for all λ ∈ iR and that T3 (λ) is right invertible for all λ ∈ iR. First we show that A − λIn B2 (20.24) C1 D1 is left invertible for all λ ∈ iR if and only if T2 (λ) is left invertible for all λ ∈ iR. To see that this is the case, we ﬁrst establish that T2 (λ) is left invertible for all λ ∈ iR if and only if AF − λIn B2 (20.25) C1 + D1 F D1 is left invertible for all λ ∈ iR. Indeed, assume that T2 (λ) is left invertible for all pure imaginary λ, and that for some λ0 ∈ iR and some vectors u and x we have 0 AF − λ0 In B2 x = . (20.26) 0 C1 + D1 F D1 u Then, since λ0 is not an eigenvalue of AF , it follows that x = (λ0 − AF )−1 B2 u. Inserting this in (C1 + D1 F )x + D1 u = 0, gives T2 (λ0 )u = 0. Since T2 (λ0 ) is left invertible u = 0, and hence also x = 0.

394

Chapter 20. H-inﬁnity control applications

Conversely, assume T2 (λ0 )u = 0 for some u and some pure imaginary λ0 . Suppose that (20.25) is left invertible for all λ ∈ iR. Put x = (λ0 − AF )−1 B2 u, then (20.26) holds, hence x = 0 and u = 0. Now (20.25) can be written as

AF − λ0 In

B2

C1 + D1 F

D1

=

A − λIn

B2

In

0

C1

D1

F

Iq

.

Thus we see that (20.25) is left invertible if and only if (20.24) is left invertible. Next we show that A − λIn B2 C1

D1

is left invertible for all λ ∈ iR. Indeed, assume that for some λ0 ∈ iR and some vectors u and x we have A − λ0 In B2 x 0 = . C1 D1 u 0 Then, in particular, C1 x + D1 u = 0. Using D1∗ D1 = Iq and D1∗ C1 = 0, this implies that u = 0. But then (A − λ0 In )x = 0 and C1 x = 0. Since the pair (C1 , A1 ) is observable by assumption, it follows that x = 0. For sake of convenience, and without loss of generality, we shall assume from now on that γ = 1. The ﬁrst main result in this section is the following theorem. Theorem 20.7. Suppose the system (20.23) satisﬁes the assumptions A1, A2 and A3, and let γ = 1. Then there is an internally stabilizing compensator K for the system (20.23) satisfying (20.6) if and only if the following two conditions hold: (i) there is a Hermitian solution Y of the Riccati equation Y (C1∗ C1 − C2∗ C2 )Y + AY + Y A∗ + B1 B1∗ = 0

(20.27)

with the additional properties that A∗ +(C1∗ C1 −C2∗ C2 )Y is stable and Y > 0, (ii) with the unique Y from (i) there is a Hermitian solution Z of the Riccati equation Z(Y C2∗ C2 Y − B2 B2∗ )Z + Z(A + Y C1∗ C1 ) + (A∗ + C1∗ C1 Y )Z + C1∗ C1 = 0 (20.28) with the additional properties that A+Y C1∗ C1 −B2 B2∗ Z +Y C2∗ C2 Y Z is stable and Z > 0.

20.4. State space solution of the standard problem

395

Moreover, when (i) and (ii) are satisﬁed, (all ) the internally stabilizing compensators can be obtained as follows. Introduce Ψ11 (λ) Ψ12 (λ) Ψ(λ) = Ψ21 (λ) Ψ22 (λ) Iq 0 −B2∗ Z −1 B2 Y C2∗ , (20.29) = (λIn − A) + 0 Im C2 (In + Y Z) = A+Y C1∗ C1 −B2 B2∗ Z +Y C2∗ C2 Y Z. Then (all ) the internally stabilizing where A compensators satisfying (20.6) are given by −1 K(λ) = Ψ11 (λ)U (λ) + Ψ12 (λ) Ψ21 (λ)U (λ) + Ψ22 (λ) , where U is an iR-stable rational q × m matrix function with U ∞ < 1. Note that condition (i) requires the Riccati equation (20.27) to have a positive deﬁnite iR-stabilizing solution. From Theorem 13.3 we know that the iR-stabilizing solution is unique. Similarly, condition (ii) requires (20.28) to have a positive deﬁnite iR-stabilizing solution, which is unique for the same reason. It will be convenient to split the proof in a number of lemmas. Lemma 20.8. The existence of a right J-spectral factorization (20.17) in condition (i) of Theorem 20.3 is equivalent to the existence of an iR-stabilizing Hermitian solution Y to the Riccati equation (20.27). Moreover, the additional property that the m × m block in the left upper corner of V has an inverse in Ratm×m is − equivalent to Y > 0. Proof. We split the proof in two parts. Part 1. Starting from the formulas for T1 and T3 given in Proposition 20.5 we form T3 0 . L= T1 Il −1 B, where n − A) This matrix function has the realization L(λ) = D + C(λI 0 B1 AF −B2 F = = , B , A 0 AH B1 + HD2 0 = C

0

C2

C1 + D1 F

−D1 F

,

D=

D2

0

0

Il

.

It will be more convenient however to work with a similar realization. Put In In . (20.30) S= 0 In

396

Chapter 20. H-inﬁnity control applications

Note that =S A

AF 0

= C

−HC2 AH

S

0

C2

C1 + D1 F

−C1

−1

,

=S B

−HD2

0

B1 + HD2

0

,

S −1 .

Also put J = diag (Ip , −Il ) and J = diag (Im , −Il ). Using the factorization principle from Section 2.6 one sees that L can be factored as L(λ) = L1 (λ)L2 (λ), where Im 0 0 L1 (λ) = (λ − AF )−1 −H 0 , + C1 + D1 F 0 Il

L2 (λ) =

D2 0

0

+

Il

C2 C1

Because L1 is of the form

(λ − AH )−1 B1 + HD2 (

Im L1 (λ) = Ξ(λ) where

0 .

) 0 , Il

Ξ(λ) = −(C1 + D1 F )(λIn − AH )−1 H,

we have that L1 and its inverse are in Rat− . Thus Ω admits a right J-spectral factorization if and only if Ω2 = L2 J L∗2 admits a right J-spectral factorization. Moreover, Ω = V JV ∗ with V and its inverse in Rat− if and only if Ω2 = V2 JV2∗ , where V2 = L−1 1 V , and V2 and its inverse are in Rat− . Now applying Lemma 20.4 to Ω2 , and using that D2 D2∗ = Im and D2 B1∗ = 0, we obtain that a right J-spectral factorization of Ω2 exists if and only if the algebraic Riccati equation X(C2∗ C2 − C1∗ C1 )X + XA∗ + AX − B1 B1∗ = 0 has a Hermitian solution X for which A∗ + (C2∗ C2 − C1∗ C1 )X has all its eigenvalues in the open left half plane. Comparing with (20.27) we see that this is equivalent to taking X = −Y . Observe also that this solution Y is unique since X is unique. In addition V (λ) = L1 (λ)−1 V2 (λ), where C2 Im 0 + (λIn − AH )−1 H − XC2∗ XC1∗ V2 (λ) = C1 0 Il

=

Im

0

0

Il

+

C2 C1

(λIn − AH )−1 H + Y C2∗

−Y C1∗ .

20.4. State space solution of the standard problem

397

Part 2. Next, we show that the property that the m × m block in the left upper corner of V has an inverse in Rat− , is equivalent to Y being positive deﬁnite. Because of the special form of H1 , we have that the m × m block in the left upper corner of V is equal to the m × m block in the left upper corner of V2 . Let us denote this block by V11 . Then −1 V11 (λ)−1 = Im − C2 λIn − (A − Y C2∗ C2 ) (H + Y C2∗ ). Now using (20.27) we have that (A − Y C2∗ C2 )Y + Y (A∗ − C2∗ C2 Y ) = −B1 B1∗ − Y (C1∗ C1 + C2∗ C2 )Y ≤ −B1 B1∗ ≤ 0.

(20.31)

Since the pair (A, B1 ) is controllable it follows from standard arguments concerning Lyapunov equations (see, e.g., Theorem 4 in Section 13.1 in [107]) that A−Y C2∗ C2 has its spectrum in the open left half plane if and only if Y is positive deﬁnite. This ﬁnishes the ﬁrst part of the proof of Theorem 20.7. Next we consider the second condition in Theorem 20.3 and its equivalence to the remaining parts of Theorem 20.7. Lemma 20.9. The existence of a left J-spectral factorization as in (20.19) in condition (ii) of Theorem 20.3 is equivalent to the existence of an iR-stabilizing solution Z of (20.28). Moreover, the additional property that the q × q block in the upper left corner of W is in Ratq×q is equivalent to Z being positive deﬁnite. − Proof. Again we shall split the argument into several parts. Part 1. For the ﬁrst step we start by computing the function from condition (ii) of Theorem 20.3 as follows. Using the notation of the proof of Lemma 20.8, deﬁne 0 Im 0 Im −1 −1 −1 L(λ) = V (λ) = V2 (λ) L1 (λ) . −T2 (λ) 0 −T2 (λ) 0 in condition (ii) of Theorem 20.3 is given by Observe that the function Ω −Im 0 ∗ Ω(λ) = L (λ)J L(λ), where J = . 0 Il First we show that the existence of a left J-spectral factorization 0 Iq ∗ = W JW, where J = Ω 0 −Il amounts to the existence of a left J-spectral factorization of the matrix function 1 arises from a certain factorization of L. In fact, the argument 1 , where L ∗J L L 1 will be similar to the one used in the proof of the previous lemma.

398

Chapter 20. H-inﬁnity control applications Using the product rule and then simplifying, we get $ 0 I 0 m = Im+l + (λIn − AF )−1 H L1 (λ)−1 −T2 (λ) 0 C1 + D1 F $ · $ =

0

Im

−D1

0

0

Im

−D1

0

+

+

0

C1 + D1 F

−1

(λIn − AF )

0

−1

C1 + D1 F

(λIn − AF )

−B2

H

0

−B2

%

%

0

%

.

Thus, again applying the multiplication rule, we obtain a formula for L(λ), by −1 ∗ pre-multiplying the above expression with V2 (λ) . Using also C1 D1 = 0, this + C(λ − A) −1 B, where yields L(λ) =D A − Y C2∗ C2 + Y C1∗ C1 −Y C1∗ C1 0 H + Y C2∗ A= , B= , 0 AF −B2 H −C2 0 0 Im = = , C . D −D1 0 −C1 C1 + D1 F It is convenient to consider another realization. With S as in (20.30) and writing AY = A − Y C2∗ C2 + Y C1∗ C1 , we have AY −B2 F − Y C2∗ C2 −1 B2 Y C2∗ =S =S A S , B , 0 AF −B2 H = C

−C2

−C2

−C1

D1 F

S −1 .

=L 1 L 2 , where It is now easily checked that L 0 Im −C2 L1 (λ) = + (λIn − AY )−1 B2 −D1 0 −C1 2 (λ) L

=

Iq

0

0

Im

+

−F −C2

(λIn − AF )−1 −B2

Y C2∗ , H .

2 is in Rat− , and as Since AF is stable, L −F AF − −B2 H = AF − B2 F + HC2 = AH −C2

20.4. State space solution of the standard problem

399

−1 is in Rat− too. has all its eigenvalues in the open left half plane, L 2 From the considerations in the previous paragraph it follows that the rational admits a left J-spectral factorization if and only if the =L ∗J L matrix function Ω function L1 J L1 admits a left J-spectral factorization.. In that case, if W1 is a left ∗ J L. 1 , then W = W1 L 2 is a J-spectral factor of L ∗1 J L J-spectral factor of L Part 2 . In this part we continue to use the notation of the previous part. We 1 . This yields that there exists a left J-spectral now apply Theorem 14.7 to L ∗ factorization of Ω = L J L if and only if there is a Hermitian solution X of the algebraic Riccati equation X(B2 B2∗ − Y C2∗ C2 Y )X + X(A + Y C1∗ C1 ) + (A∗ + C1∗ C1 Y )X − C1∗ C1 = 0 having the additional property σ(A + Y C1∗ C1 + B2 B2∗ X − Y C2∗ C2 Y X) ⊂ Cleft . This solution X is unique. Taking Z = −X we see that Z satisﬁes the algebraic Riccati equation (20.28) and is the iR-stabilizing solution of that equation. Thus the left J-spectral factor ∗ J L 1 is given by W1 of L 1 B2∗ Z (λIn − AY )−1 B2 Y C2∗ , (20.32) W1 (λ) = Iq+m + −C2 − C2 Y Z 2 (λ) becomes and the product W (λ) = W1 (λ)L %−1 ( )$ B2 AY −B2 F − Y C2∗ C2 B2∗ Z −F Iq+m + λ− −C2 − C2 Y Z −C2 0 AF −B2

Y C2∗ H

.

Part 3. We now consider the additional property that the q × q block in the upper left corner of W has an inverse in Ratq×q − , and prove that this is equivalent to Z being positive deﬁnite. Let us denote the q × q block in the upper left corner of W by W11 . Then $ %−1 AY −B2 F − Y C2∗ C2 B2 ∗ λI2n − . W11 (λ) = Iq + B2 Z −F 0 AF −B2 −1 is Thus the main operator in the realization of W11 AY −B2 F − Y C2∗ C2 B2 = B2∗ Z A − 0 AF −B2

=

AY − B2 B2∗ Z

−Y C2∗ C2

B2 B2∗ Z

A

−F

.

We have to show that this matrix has all its eigenvalues in the open left half plane if and only if Z is positive deﬁnite.

400

Chapter 20. H-inﬁnity control applications In order to do this, it is helpful to consider a similar matrix. Take In 0 S= , −In In

and put = = S −1 AS A

A + Y C1 C1∗ − B2 B2∗ Z

−Y C2∗ C2

Y C1∗ C1

A − Y C2∗ C2

.

has all its eigenvalues in the left half We shall show that Z > 0 ifand onlyif A Z 0 Z 0 ∗ +A A plane. To this end, consider 0 Y −1 0 Y −1 =

−C1∗ C1 − ZB2 B2∗ − ZY C2∗ C2 Y Z

C1∗ C1 − ZY C2∗ C2

C1∗ C1 − C2∗ C2 Y Z

Λ

,

where, because of (20.31), Λ = =

Y −1 (A − Y C2∗ C2 ) + (A∗ − C2∗ C2 Y )Y −1 −Y −1 B1 B1∗ Y −1 − C1∗ C1 − C2∗ C2 .

Substituting the latter expression for Λ in the right lower corner of the matrix above, we obtain Z 0 Z 0 ∗ +A A 0 Y −1 0 Y −1 C1∗ C1 − ZY C2∗ C2 −C1∗ C1 − ZB2 B2∗ − ZY C2∗ C2 Y Z = C1∗ C1 − C2∗ C2 Y Z −Y −1 B1 B1∗ Y −1 − C1∗ C1 − C2∗ C2 ⎤ ⎡ −C1 C1 ⎥ ⎢ ∗ ∗ ZB2 0 ZY C2∗ ⎢ 0 ⎥ C1 ⎥ ⎢ B2 Z =− ⎥. ⎢ ∗ −1 ∗ ∗ −1 ⎥ ⎢ B1 Y ⎦ −C1 0 Y B1 C2 ⎣ 0 C2 Y Z

C2

as shorthand for the latter factor, this reduces to With the notation C Z 0 Z 0 +A ∗ ∗ C. A = −C 0 Y −1 0 Y −1

(20.33)

20.4. State space solution of the standard problem

401

A is observable. Suppose Next, we show that the pair C, A

x y

= λ0

x y

,

C

x y

= 0,

or, which comes down to the same, (A + Y C1∗ C1 − B2 B2∗ Z)x − Y C2 C2∗ y

= λ0 x,

Y C1∗ C1 x + (A − Y C2∗ C2 )y

= λ0 y,

and C1 x = C1 y,

B2∗ Zx = 0,

B1∗ Y −1 y = 0,

C2 Y Zx = −C2 y.

Using C1 x = C1 y, it follows that (A − Y C2∗ C2 + Y C1∗ C1 )y = λ0 y. Combining this with B1∗ Y −1 y = 0, and putting w = Y −1 y, we obtain (AY − Y C2∗ C2 Y + Y C1 C1∗ Y + B1 B1∗ )w = λ0 Y w,

B1∗ w = 0.

Now use (20.27) to see that this implies Y A∗ w = λ0 Y w. As Y is invertible we have A∗ w = λ0 w and B1∗ w = 0. Since (A, B1 ) is controllable, it follows that w = 0. Hence y = 0 too. From (A + Y C1∗ C1 − B2 B2∗ Z)x − Y C2 C2∗ y = λ0 x, combined with y = 0, C1 x = C1 y = 0 and B2∗ Zx = 0 we then have Ax = λ0 x. The observability of the pair (C1 , A) ﬁnally gives x = 0. We ﬁnish by applying the result of Theorem 4 in Section 13.1 in [107] to the equation (20.33). Combined with the fact that Y > 0, this gives that Z > 0 if and has all its eigenvalues in the open left half plane. only if A This concludes the proof of the equivalence of (i) and (ii) in Theorem 20.7. We bring the argument to a close as follows. Proof of Theorem 20.7. In view of the two preceding lemmas, it remains to prove the formulas for the parametrization of the internally stabilizing compensators satisfying (20.6). Recall from Theorem 19.6, in particular from formula (19.9), that K = (Y − M Q)(X − N Q)−1 . Also we have formula (20.20), that is the expression Q = −(X11 U + X12 )(X21 U + X22 )−1 , where W −1 =

X11

X12

X21

X22

.

402

Chapter 20. H-inﬁnity control applications

Here W is obtained from Part 1 of the proof of Lemma 20.9. Combining the expressions, we see that K = (Y X21 + M X11 )U + (Y X22 + M X12 ) −1 · (XX21 + N X11 )U + (XX22 + N X12) = (Ψ11 U + Ψ12 )(Ψ21 U + Ψ22 )−1 , with Ψ given by Ψ=

Ψ11

Ψ12

Ψ21

Ψ22

=

M

Y

X11

X12

N

X

X21

X22

.

The formulas in (19.8) now give F M (λ) Y (λ) Iq 0 + (λIn − AF )−1 B2 = C2 0 Im N (λ) X(λ)

−H .

2 (λ) from Part 1 of the proof of the Fortuitously, this is equal to the function L 2 W −1 = W −1 , where W1 is previous lemma. Since W = W1 L2 , we get Ψ = L 1 given by (20.32). Hence Ψ is given by (20.29), as desired. We conclude with the second main result of this chapter. Theorem 20.10. Suppose the system (20.23) satisﬁes the assumptions A1, A2 and A3, and let γ be an arbitrary positive number. Then there is an internally stabilizing compensator K for the system (20.23) satisfying (20.6) if and only if the following three conditions hold: (i) there is a positive deﬁnite iR-stabilizing solution X of the Riccati equation X(γ −2 B1 B1∗ − B2 B2∗ )X + A∗ X + XA + C1∗ C1 = 0,

(20.34)

(ii) there is a positive deﬁnite iR-stabilizing solution Y of the Riccati equation Y (γ −2 C1∗ C1 − C2∗ C2 )Y + AY + Y A∗ + B1 B1∗ = 0,

(20.35)

(iii) X < γ −2 Y −1 or, equivalently, all eigenvalues of XY are in the open disc {z | |z| < γ −2}. In that case (all ) the internally stabilizing compensators K can be obtained as follows. Introduce Φ11 (λ) Φ12 (λ) Φ(λ) = Φ21 (λ) Φ22 (λ) ) ( ∗ ) ( B2 X 0 Iq −1 Y C2∗ B2 , − (I − γ −2 Y X)−1 (λ − A) = C2 Im 0

20.4. State space solution of the standard problem

403

= A − Y (C2∗ C2 − γ −2 C1∗ C1 ) − B2 B2∗ X(In − γ −2 Y X)−1 . Then (all ) the where A internally stabilizing compensators satisfying (20.6) are given by K(λ) = Φ11 (λ) + Φ12 (λ)U (λ)(Im − Φ22 (λ)U (λ))−1 Φ21 (λ), where U is an iR-stable rational q × m matrix function satisfying U ∞ < γ. Proof. The theorem may be derived from the previous one upon giving the connections between X and Z. Again we assume γ = 1 without loss of generality. Under this assumption, condition (i) in Theorem 20.7 is exactly the same as the second condition in the present theorem. Henceforth we suppose it is satisﬁed. Thus, throughout the proof, Y will be a positive deﬁnite iR-stabilizing solution of (20.27), or, equivalently, of (20.35) with γ = 1. The argument below is divided into four parts. Part 1. Introduce the block matrices −A∗ −C1∗ C1 −C1∗ C1 −A∗ − C1∗ C1 Y = H= . , H B1 B1∗ − B2 B2∗ A Y C2∗ C2 Y − B2 B2∗ A + Y C1∗ C1 In the terminology of Section 12.1 the matrix H is the Hamiltonian of the Riccati is the Hamiltonian of the Riccati equation equation (20.34) with γ = 1, while H (20.28). Introduce also ) ( In 0 . S= Y In Since Y is a solution of the Riccati equation (20.35) and γ = 1, a direct computa tion gives S −1 HS = H. Part 2. Here we assume that Z is the (unique) Hermitian iR-stabilizing solution of equation (20.28), and in addition that Z is positive deﬁnite. That is, it is assumed that condition (ii) in Theorem 20.7 is met. Since Z is iR-stabilizing, the space ∗ corresponding to the open left half Im Z ∗ In is the spectral subspace of H plane. It follows that Z Z = Im SIm In + Y Z In is the spectral subspace of H corresponding to the open left half plane. Our next concern is the invertibility of In + Y Z. Since Z is positive deﬁnite, In + Y Z = Z −1/2 (In + Z 1/2 Y Z 1/2 )Z 1/2 is similar to a positive deﬁnite matrix. Consequently, In + Y Z is invertible. Next, put X = Z(In + Y Z)−1 . We shall show that X is positive deﬁnite, X is the iR-stabilizing solution of (20.34) (with γ = 1), and that X < Y −1 . For this, note that X = Z(In + Y Z)−1 = (Z −1 + Y )−1 , so that X is positive deﬁnite. Furthermore, Z X Im = Im . In + Y Z In

404

Chapter 20. H-inﬁnity control applications

Hence X is the Hermitian iR-stabilizing solution of (20.34). In addition, since Z is positive deﬁnite also X −1 > Y , and as both X and Y are positive deﬁnite this yields X < Y −1 . We conclude that all conditions of Theorem 20.10 are satisﬁed. Part 3. This part deals with the reverse implication. So, we start with the positive deﬁnite iR-stabilizing solution X of (20.34) with γ = 1 such that X < Y −1 . We show that Z = (In − Y X)−1 is well-deﬁned and positive deﬁnite, and that Z is the iR-stabilizing solution of (20.28). Since X < Y −1 , the matrix I − Y X is invertible, hence Z is well-deﬁned. In addition, Z = X(In −Y X)−1 = (X −1 −Y )−1 is positive deﬁnite because X 0 A1/2 A−1 A−∗ λ−A ρ(A) σ(A) P (A; Γ) AM A[M ] A−1 [M ] A|M A(X1 → X2 ) D(A)

quotient space of M over N symbol for orthogonality in Hilbert space orthogonal complement of subspace M in Hilbert space orthogonality of sets V and W orthogonal direct sum (of subspaces) of Hilbert spaces algebraic (possibly non-orthogonal) direct sum of linear manifolds or (sub)spaces

List of symbols L(Y ) L(U, Y ) C(Γ, U )

Lp (Ω) Lm p (Ω) Lm×r (Ω) p Lm 1,ω (R) Dm 1 (R) Dm 1 [0, ∞) L2 (R+ , H) · , · [· , ·] A[ ] D + C(λI − A)−1 B A× Ker (C|A) Im (A|B) E(·; A) etS etS PΘ prΠ (Θ) W −1 δ(W ) δ(W ; λ0 )

417 Banach algebra of all bounded linear operators on Banach space Y Banach space of all bounded linear operators from Banach space U into Banach space Y Banach space of all U -valued continuous functions on Γ endowed with the supremum norm Lebesgue space of p-integrable functions on a measurable set Ω space of Cm -valued functions of which the entries are in Lp (Ω) space of m × r matrix functions of which the columns are in Lm p (Ω) a weighted Lm 1 -space; see Section 5.3 a certain linear submanifold of Lm 1 (R); see Section 5.3 linear manifold of all functions f ∈ Dm 1 (R) with f (t) = 0 for t < 0 the space of all square integrable functions on [0, ∞) with values in Hilbert space H standard inner product in Cm or L2 [−1, 1] alternative inner product in Cm or L2 [−1, 1] adjoint of an operator with respect to alternative inner product (in L2 [−1, 1]) realization associate state space operator (or matrix), associate main operator (or matrix) corresponding to a realization stands for Ker C ∩ Ker CA ∩ Ker CA2 ∩ · · · stands for Im B + Im AB + Im A2 B + · · · bisemigroup generated by exponentially dichotomous operator A value at t(< 0) of the left semigroup generated by S value at t(> 0) of the right semigroup generated by S separating projection for −iA where A is the main operator of the spectral triple Θ projection of realization triple Θ = (A, B, C) associated with a projection Π pointwise inverse of rational matrix function W , deﬁned by W −1 (λ) = W (λ)−1 ) McMillan degree of a rational matrix function W local degree of W at λ0

418

List of symbols

π+ (W )

number of positive eigenvalues of the Hermitian matrix associated with a minimal realization of J-unitary rational matrix function W

F∗

adjoint of the rational matrix function F relative to the ¯ ∗ imaginary axis, deﬁned by F ∗ (λ) = F (−λ) the set of all rational matrix functions that are proper and have no pole at the imaginary axis the set of all p × q matrix functions in Rat the set of all F in Ratp×q such that sups∈iR F (s) ≤ 1 the set of all matrix functions in Ratp×q that are analytic on the closed left half plane, inﬁnity included the set Ratp×q ∩ Ratp×q + B the set of all rational matrix functions that are analytic on the closed right half plane, inﬁnity included, that is, the set of all iR-stable rational matrix functions the set of all p × q matrix functions in Rat− the unit element in the algebra Ratp×p

Rat Ratp×q Ratp×q B Ratp×q + Ratp×q +, B Rat− Ratp×q − Ep

Index H-adjoint of matrix, 212 H-Lagrangian subspace, 212 H-dissipative, 215 H-negative vector, 211 H-neutral subspace, 211 H-neutral vector, 211 H-nondegenerate, 212 H-nonnegative subspace, 211 H-nonpositive subspace, 211 H-orthogonal vectors, 212 H-orthogonality, 212 H-positive vector, 211 H-selfadjoint matrix, 184, 213 J-contraction, 333 J-contractive, 328 J-inner rational matrix function, 333 J-unitary matrix, 313 J-unitary rational matrix function on the imaginary axis, 314 R-stabilizing solution of Riccati equation, 273 T-stabilizing solution of discrete Riccati equation, 264, 270 Ω-regular (linear matrix) pencil, 57 iR-stabilizing solution of Riccati equation, 238 iR-stable rational matrix function, 350 angular operator, 219 angular subspace, 219 associate main matrix of matrix realization, 21

associate main operator of realization, 21 realization triple, 93 associate state space matrix of matrix realization, 21 associate state space operator of realization, 21 bi-inner rational matrix function, 333 BIBO-stable, 371 bigenerator of bisemigroup, 80 biproper rational matrix function, 175 biproper realization, 21 bisemigroup generated by exponentially dichotomous operator, 80 block Toeplitz equation, 13 bounded-input bounded-output stable, 371 Cauchy contour, 5 Cauchy domain, 5 co-inner rational matrix, 339 co-pole function, 173 corresponding to Jordan chain, 174 complete factorization, 175 contractive rational matrix function on imaginary axis, 301 on real line, 310 controllability gramian, 350 controllable pair of matrices, 171 controllable realization, 172

420 coupling relation (between operators), 106 derivative of f ∈ Dm 1 (R), 87 detectable pair of matrices, 374 doubly comprime factorization, 376 exponential type of exponentially dichotomous operator, 79 of realization triple, 87 of semigroup, 78 exponentially decaying semigroup, 78 exponentially dichotomous operator, 79 exterior domain of Cauchy contour, 5 external matrix of matrix realization, 21 external operator of realization, 21 externally stable system, 371 ﬁnite dimensional realization, 20 gramian, 350 half range problem, 116 Hamiltonian, 220 Hermitian matrix associated with minimal realization, 317 Hermitian rational matrix function on imaginary axis, 181 on real line, 181 on unit circle, 189 indeﬁnite inner product given by Hermitian matrix, 211 inner rational matrix function, 336 inner-outer factorization (with invertible outer factor), 337 input matrix of matrix realization, 21 input operator of realization, 21 input space of realization triple, 88 interior domain of Cauchy contour, 5 internal stability, 376

Index internally stabilizing compensator for system, 376 invertible in Ratp×p , 352 invertible outer rational matrix function, 337 kernel function associated with realization triple, 90 of Wiener-Hopf equation, 9 left J-spectral factorization with respect to the imaginary axis, 250 with respect to the real line, 272 with respect to the unit circle, 262 left (C0 -)semigroup, 78 left canonical factorization of operator function (with respect to Cauchy contour), 144 of rational matrix function, 39 left Hankel operator, 100 left pseudo-canonical factorization, 176 left pseudo-spectral factorization with respect to imaginary axis, 199 with respect to real line, 198 with respect to unit circle, 204 left spectral factor, 185 left spectral factorization with respect to the imaginary axis, 185 with respect to the real line, 185 with respect to the unit circle, 192 left Wiener-Hopf factorization (with respect to Cauchy contour), 158 linear fractional map, 352 linear manifold, 30 linear matrix pencil, 57

Index local minimality at a given point, 172 main matrix of matrix realization, 21 main operator of realization, 21 realization triple, 88 manifold, 30 matching condition, 31 matricially coupled operators, 105 matrix-valued realization of rational matrix function, 19 maximal H-nonnegative subspace, 212 maximal H-nonpositive subspace, 212 McMillan degree, 171 minimal factorization (involving arbitrary number of factors), 175 minimal realization, 20 model matching problem, 381 one-sided, 383 two-sided, 381, 386 negative squares, 320 Nehari problem, 350 Nehari-Takagi problem (relative to the imaginary axis), 369 nonnegative rational matrix function on imaginary axis, 197 on real line, 197 on unit circle, 197 nonnegative real part on the real line, 291 number of poles of function in Ratp×q in the open left half plane, multiplicities taken into account, 358 observability gramian, 350 observable pair of matrices, 172 observable realization, 172 one-sided model matching problem , 383

421 outer rational matrix function (invertible), 337 outer-co-inner factorization (with invertible outer factor), 339 output matrix of matrix realization, 21 output operator of realization, 21 output space of realization triple, 88 pair of matching subspaces, 31 partial indices, 159 pencil, 57 pole placement theorem, 374 positive deﬁnite rational matrix function on imaginary axis, 185 on real line, 185 on unit circle, 192 positive deﬁnite real part on the real line, 291 product of realizations, 30 projection of realization triple associated with a projection, 110 propagator function, 117 proper rational matrix function, 19 rational Nehari problem (relative to the imaginary axis with given tolerance), 350 rational Nehari-Takagi problem (relative to the imaginary axis), 369 realization of a system, 372 of operator function on given set, 20 of rational matrix function, 19 realization triple, 88 of given exponential type, 87 Redheﬀer transform of 2 × 2 block matrix, 328 of rational matrix function, 330 Redheﬀer transformation, 328 regular (linear matrix) pencil, 57

422 resolvent set of operator, 20 Riccati equation algebraic, 220 discrete algebraic, 264, 270 symmetric algebraic, 238 Riemann-Hilbert boundary value problem, 52 Riesz projection, 38 right J-spectral factorization with respect to the imaginary axis, 250 with respect to the real line, 272 with respect to the unit circle, 262 right (C0 -)semigroup, 78 right (Wiener-Hopf) factorization indices (with respect to Cauchy contour), 159 right canonical factorization (of symbol) with respect to real line, 10 (of symbol) with respect to the unit circle, 13 of operator function (with respect to Cauchy contour), 144 of rational matrix function, 39 of Wiener-Hopf equation with integrable operator-valued kernel function, 122 with respect to Cauchy contour, 16 right Hankel operator, 100 right pseudo-canonical factorization, 176 right pseudo-spectral factorization with respect to imaginary axis, 199 with respect to real line, 198 with respect to unit circle, 204 right spectral factor, 185 right spectral factorization with respect to the imaginary axis, 185

Index with respect to the real line, 185 with respect to the unit circle, 192 right Wiener-Hopf factorization (with respect to Cauchy contour), 157 scattering function, 116 selfadjoint rational matrix function on imaginary axis, 181 on real line, 181 on the unit circle, 189 separating projection for exponentially dichotomous operator, 79 sign characteristic of pair of matrices, 213 signature matrix, 250 similarity between realizations, 172 singular integral equation, 15 skew-Hermitian matrix, 182 skew-Hermitian rational matrix function on real line, 298 solution of Riemann-Hilbert boundary problem, 71 spectral projection, 38 spectral subspace, 38 splitting of spectrum, 37 stabilizable pair of matrices, 238 stabilizing solution of discrete Riccati equation, 264, 270 of Riccati equation, 238, 273 stable matrix, 372 stable rational matrix function, 350 stable realization of system, 372 standard problem of H-inﬁnity control, 381 state space matrix of matrix realization, 21 state space of realization, 21 triple, 88 state space operator of realization, 21

Index state space similarity between realizations, 172 theorem, 172 strictly H-dissipative matrix, 216 strictly contractive at inﬁnity, 302 strictly contractive rational matrix function on imaginary axis, 304 on the real line, 307 strictly proper rational matrix function, 305 strictly proper realization, 21 suboptimal rational Nehari problem relative to the imaginary axis with given tolerance, 350 supporting pair of subspaces, 31 supporting projection for realization, 31 realization triple, 110 symbol of (block) Toeplitz equation, 13 (block) Toeplitz operator, 14 Wiener-Hopf equation with integrable operator-valued kernel, 122 Wiener-Hopf integral equation, 10 Wiener-Hopf integral operator, 12 transfer function of realization triple, 90 of system, 21 two-sided inner rational matrix function, 333 two-sided model matching problem, 381, 386 uncontrollable eigenvalues, 374 unital realization, 21 unitary rational matrix functions, 324 unobservable eigenvalues of, 374

423 Wiener-Hopf equation, 9 Wiener-Hopf integral operator, 11 Wiener-Hopf operator with kernel function k, 99 zero of rational matrix function, 172

Editors: Harry Dym (Rehovot, Israel) Joseph A. Ball (Blacksburg, VA, USA) Marinus A. Kaashoek (Amsterdam, The Netherlands) Heinz Langer (Vienna, Austria) Christiane Tretter (Bern, Switzerland)

Associate Editors: Vadim Adamyan (Odessa, Ukraine) Albrecht Böttcher (Chemnitz, Germany) B. Malcolm Brown (Cardiff, UK) Raul Curto (Iowa City, IA, USA) Fritz Gesztesy (Columbia, MO, USA) Pavel Kurasov (Lund, Sweden) Leonid E. Lerer (Haifa, Israel) Vern Paulsen (Houston, TX, USA) Mihai Putinar (Santa Barbara, CA, USA) Leiba Rodman (Williamsburg, VI, USA) Ilya M. Spitkovsky (Williamsburg, VI, USA)

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)

Honorary and Advisory Editorial Board: 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, AB, Canada) Peter D. Lax (New York, NY, USA) Donald Sarason (Berkeley, CA, USA) Bernd Silbermann (Chemnitz, Germany) Harold Widom (Santa Cruz, CA, USA)

A State Space Approach to Canonical Factorization with Applications

Harm Bart Israel Gohberg Marinus A. Kaashoek André C.M. Ran

Birkhäuser

L O L S

Linear Operators & Linear Systems

Authors: Harm Bart Econometrisch Instituut Erasumus Universiteit Rotterdam Postbus 1738 3000 DR Rotterdam The Netherlands e-mail: [email protected]

Marinus A. Kaashoek, André C. M. Ran Department of Mathematics, FEW Vrije Universiteit De Boelelaan 1081A 1081 HV Amsterdam The Netherlands e-mail: [email protected] [email protected]

Israel Gohberg (Z"L)

2010 Mathematics Subject Classiﬁcation: Primary 46C20; 47A48, 47A56, 47A68, 93B36; secondary: 42A85, 82D75 Library of Congress Control Number: 2010923703

Bibliographic information published by Die Deutsche Bibliothek. Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliograﬁe; detailed bibliographic data is available in the Internet at http://dnb.ddb.de

ISBN 978-3-7643-8752-5 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, speciﬁcally the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microﬁlms or in other ways, and storage in data banks. For any kind of use permission of the copyright owner must be obtained.

© 2010 Birkhäuser / Springer Basel AG P.O. Box 133, CH-4010 Basel, Switzerland Part of Springer Science+Business Media Printed on acid-free paper produced from chlorine-free pulp. TCF∞ Printed in Germany ISBN 978-3-7643-8752-5

e-ISBN 978-3-7643-8753-2

987654321

www.birkhauser.ch

Contents Preface

xi

0 Introduction

1

Part I Convolution equations, canonical factorization and the state space method

7

1 The 1.1 1.2 1.3

role of canonical factorization in solving convolution equations Wiener-Hopf integral equations and factorization . . . . . . . Block Toeplitz equations and factorization . . . . . . . . . . . Singular integral equations and factorization . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

9 9 13 15 17

2 The 2.1 2.2 2.3 2.4 2.5 2.6

state space method and factorization Preliminaries on realization . . . . . . . . Realization of rational matrix functions . Realization of analytic operator functions Inversion . . . . . . . . . . . . . . . . . . . Products . . . . . . . . . . . . . . . . . . . Factorization . . . . . . . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

19 19 22 23 26 27 30 33

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

Part II Convolution equations with rational matrix symbols 3 Explicit solutions using realizations 3.1 Canonical factorization of rational matrix functions in form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Wiener-Hopf integral operators . . . . . . . . . . . . . 3.3 Block Toeplitz operators . . . . . . . . . . . . . . . . . 3.4 Singular integral equations . . . . . . . . . . . . . . . . 3.5 The Riemann-Hilbert boundary value problem . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . .

35 37 state . . . . . . . . . . . . . . . . . .

space . . . . . . . . . . . . . . . . . . . . . . . .

37 42 46 50 51 56

vi

Contents

4 Factorization of non-proper rational matrix functions 4.1 Preliminaries about matrix pencils . . . . . . . . . . . . . . . 4.2 Realization of a non-proper rational matrix function . . . . . 4.3 Explicit canonical factorization . . . . . . . . . . . . . . . . . 4.4 Inversion of singular operators with a rational matrix symbol 4.5 The Riemann-Hilbert boundary value problem revisited (1) . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

Part III Equations with non-rational symbols

57 57 59 61 68 71 74

75

5 Factorization of matrix functions analytic in a strip 5.1 Exponentially dichotomous operators and bisemigroups . . 5.2 Spectral splitting and proof of Theorem 5.2 . . . . . . . . . 5.3 Realization triples . . . . . . . . . . . . . . . . . . . . . . . 5.4 Construction of realization triples . . . . . . . . . . . . . . . 5.5 Inverting matrix functions analytic in a strip . . . . . . . . 5.6 Inverting full line convolution operators . . . . . . . . . . . 5.7 Inverting Wiener-Hopf integral operators . . . . . . . . . . . 5.8 Explicit canonical factorization . . . . . . . . . . . . . . . . 5.9 The Riemann-Hilbert boundary value problem revisited (2) Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

77 78 81 87 91 93 98 100 106 111 113

6 Convolution equations and the transport equation 6.1 The transport equation . . . . . . . . . . . . . . . 6.2 The case of a ﬁnite number of scattering directions 6.3 Wiener-Hopf equations with operator-valued kernel 6.4 Construction of a canonical factorization . . . . . . 6.5 The matching of the subspaces . . . . . . . . . . . 6.6 Formulas for solutions . . . . . . . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . functions . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

115 116 118 122 124 135 138 142

7 Wiener-Hopf factorization and factorization indices 7.1 Canonical factorization of operator functions . . 7.2 Proof of Theorem 7.2 . . . . . . . . . . . . . . . . 7.3 Wiener-Hopf factorization and spectral invariants Notes . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

143 143 147 157 167

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

Contents

vii

Part IV Factorization of selfadjoint rational matrix functions 8 Preliminaries concerning minimal factorization 8.1 Minimal realizations . . . . . . . . . . . . 8.2 Minimal factorization . . . . . . . . . . . 8.3 Pseudo-canonical factorization . . . . . . Notes . . . . . . . . . . . . . . . . . . . .

169

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

171 171 174 176 178

9 Factorization of positive deﬁnite rational matrix functions 9.1 Preliminaries on selfadjoint rational matrix functions 9.2 Spectral factorization . . . . . . . . . . . . . . . . . . 9.3 Positive deﬁnite functions on the unit circle . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

181 181 185 189 195

. . . .

. . . .

. . . .

. . . .

. . . .

10 Pseudo-spectral factorizations of selfadjoint rational matrix functions 197 10.1 Nonnegative rational matrix functions . . . . . . . . . . . . . . . . 197 10.2 Selfadjoint rational matrix functions and further generalizations . . 205 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 11 Review of the theory of matrices in indeﬁnite inner product spaces 11.1 Subspaces of indeﬁnite inner product spaces . . . . . . . . . . . 11.2 H-selfadjoint matrices . . . . . . . . . . . . . . . . . . . . . . . 11.3 H-dissipative matrices . . . . . . . . . . . . . . . . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

Part V Riccati equations and factorization 12 Canonical factorization and Riccati equations 12.1 Preliminaries on spectral angular subspaces . 12.2 Angular operators and factorization . . . . . 12.3 Riccati equations and canonical factorization 12.4 Left versus right canonical factorization . . . Notes . . . . . . . . . . . . . . . . . . . . . .

211 211 212 215 216

217 . . . . .

. . . . .

13 The symmetric algebraic Riccati equation 13.1 Spectral factorization and Riccati equations . . . 13.2 Stabilizing solutions . . . . . . . . . . . . . . . . 13.3 Symmetric Riccati equations and pseudo-spectral Notes . . . . . . . . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

219 219 221 227 229 231

. . . . . . . . . . . . . . . . factorization . . . . . . . .

. . . .

. . . .

233 233 238 242 247

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

viii

Contents

14 J-spectral factorization 14.1 Deﬁnition of J-spectral factorization . . . . . . . . . . . . 14.2 J-spectral factorizations and invariant subspaces . . . . . 14.3 J-spectral factorizations and Riccati equations . . . . . . 14.4 Two special cases of J-spectral factorization . . . . . . . . 14.5 J-spectral factorization with respect to other contours . . 14.6 Left versus right J-spectral factorization . . . . . . . . . . 14.7 J-spectral factorization relative to the unit circle revisited Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

Part VI Factorizations and symmetries

249 249 251 256 259 262 273 276 288 289

15 Factorization of positive real rational matrix functions 15.1 Rational matrix functions with a positive deﬁnite real part 15.2 Canonical factorization of functions with a positive deﬁnite real part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.3 Generalization to pseudo-canonical factorization . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

291 . . . . 291 . . . . 294 . . . . 297 . . . . 300

16 Contractive rational matrix functions 16.1 State space analysis of contractive rational matrix functions 16.2 Strictly contractive rational matrix functions . . . . . . . . 16.3 An application to spectral factorization . . . . . . . . . . . 16.4 An application to canonical factorization . . . . . . . . . . . 16.5 A generalization to pseudo-canonical factorization . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

301 301 304 305 307 309 312

17 J-unitary rational matrix functions 17.1 Realizations of J-unitary rational matrix functions . 17.2 Factorization of J-unitary rational matrix functions 17.3 Factorization of unitary rational matrix functions . . 17.4 Intermezzo on the Redheﬀer transformation . . . . . 17.5 J-inner rational matrix functions . . . . . . . . . . . 17.6 Inner-outer factorization . . . . . . . . . . . . . . . . 17.7 Unitary completions of minimal degree . . . . . . . . 17.8 Bi-inner completions of inner functions . . . . . . . . Notes . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

313 313 321 324 328 333 336 339 341 345

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

Contents

ix

Part VII Applications of J-spectral factorizations 18 Application to the rational Nehari problem 18.1 Problem statement and main result . . . 18.2 Intermezzo about linear fractional maps 18.3 The J-spectral factorization approach . 18.4 Proof of the main result . . . . . . . . . 18.5 The case of a non-stable given function . 18.6 The Nehari-Takagi problem . . . . . . . Notes . . . . . . . . . . . . . . . . . . .

347

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

349 349 352 359 361 366 368 370

19 Review of some control theory for linear systems 371 19.1 Stability and feedback . . . . . . . . . . . . . . . . . . . . . . . . . 371 19.2 Parametrization of internally stabilizing compensators . . . . . . . 374 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378 20 H-inﬁnity control applications 20.1 The standard problem and model matching . 20.2 The one-sided model matching problem . . . 20.3 The two-sided model matching problem . . . 20.4 State space solution of the standard problem Notes . . . . . . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

379 380 382 386 392 404

Bibliography

405

List of symbols

415

Index

419

Preface The present book deals with canonical factorization problems for diﬀerent classes of matrix and operator functions. Such problems appear in various areas of mathematics and its applications. The functions we consider have in common that they appear in the state space form or can be represented in such a form. The main results are all expressed in terms of the matrices or operators appearing in the state space representation. This includes necessary and suﬃcient conditions for canonical factorizations to exist and explicit formulas for the corresponding factors. Also, in the applications the entries in the state space representation play a crucial role. The theory developed in the book is based on a geometric approach which has its origins in diﬀerent ﬁelds. One of the initial steps can be found in mathematical systems theory and electrical network theory, where a cascade decomposition of an input-output system or a network is related to a factorization of the associated transfer function. Canonical factorization has a long and interesting history which starts in the theory of convolution equations. Solving Wiener-Hopf integral equations is closely related to canonical factorization. The problem of canonical factorization also appears in other branches of applied analysis and in mathematical systems theory, in H∞ -control theory in particular. The ﬁrst book devoted to the state space factorization theory was published in 1979 as the monograph “Minimal factorization of matrix and operator functions,” Operator Theory: Advances and Applications 1, Birkh¨ auser Verlag, written by the ﬁrst three authors. Some of the factorization results published in the 1979 book appeared there in print for the ﬁrst time. The present book is the second book written by the four of us in which the state space factorization method is systematically used and developed further. In the earlier book [20], published in 2008, the emphasis is on non-canonical factorizations and degree 1 factorizations, in particular. In the present book we concentrate on canonical factorizations. Together both books present a rich and far reaching update of the 1979 monograph [11]. In the present book the emphasis is on canonical factorization and symmetric factorization with applications to diﬀerent classes of convolution equations. For

xii

Preface

the latter we have in mind the transport equation, singular integral equations, equations with symbols analytic in a strip, and equations involving factorization of non-proper rational matrix functions. A large part of the book will deal with factorization of matrix functions satisfying various symmetries. A main theme will be the eﬀect of these symmetries on factorization and how the symmetries can be used in eﬀective ways to get state space formulas for the factors. Applications to H∞ -control theory, which have been developed in the 1980s and 1990s, will also be included. The text is largely self-contained, and will be of interest to experts and students in mathematics, sciences and engineering. The authors gratefully acknowledge a visitor fellowship for the second author from the Netherlands Organization for Scientiﬁc Research (NWO), and the ﬁnancial support from the School of Economics of the Erasmus University at Rotterdam, from the School of Mathematical Sciences of Tel-Aviv University and the Nathan and Lily Silver Family Foundation, and from the Mathematics Department of the Vrije Universiteit at Amsterdam. These funds allowed us to meet and to work together on the book for diﬀerent extended periods of time in Amsterdam and Tel-Aviv. The authors

Amsterdam – Rotterdam – Tel-Aviv, Summer 2009

Postscript On Monday October 12, 2009, Israel Gohberg, the second author of this book, passed away at the age of 81. At that time the preparation of the book was in a ﬁnal phase and only some minor work had to be done. Israel Gohberg was one of the initiators using state space methods in solving problems appearing in various branches of mathematical analysis and its applications. His fundamental insights and inspiring leadership have been driving forces in our joint work.

Chapter 0

Introduction This monograph presents a uniﬁed approach for solving canonical factorization problems for diﬀerent classes of matrix and operator functions. The notion of canonical factorization originates from the theory of convolution equations. For instance, canonical factorization, provided it exists, allows one to invert WienerHopf, Toeplitz and singular integral operators, and when the factors are known one can also build explicitly the inverses of these operators. The problem of canonical factorization also appears in various branches of applied analysis, in linear transport theory, in interpolation theory, in mathematical systems theory, in particular, in H∞ -control theory. The various matrix and operator functions that are considered in this book have in common that they appear in a natural way as functions of the form W (λ) = D + C(λI − A)−1 B

(1)

or (after a suitable transformation) can be represented in this form. In the above formula λ is a complex variable, and A, B, C, and D are matrices or linear operators acting between appropriate Banach or Hilbert spaces, which in this book often will be ﬁnite dimensional. When the underlying spaces are all ﬁnite dimensional, A, B, C, and D can be viewed as matrices and the function W is a rational matrix function which is analytic at inﬁnity. From mathematical systems theory it is known that, conversely, any rational matrix function which is analytic at inﬁnity admits a representation of the above form. In systems theory the right hand side of (1) is called a state space realization of the function W , and one refers to the space in which A is acting as the state space. The method of factorization employed in this book uses realizations as in (1), and for this reason it is referred to as the state space method. It allows one to deal with factorization from a geometric point of view. This state space factorization approach has its origins in diﬀerent ﬁelds, for instance, in the theory of non-selfadjoint operators [27], [141], in mathematical systems theory and electrical

2

Chapter 0. Introduction

network theory [23], [95], [94], and in the factorization theory of matrix polynomials [67], [131]. In all three areas a state space representation of the function to be factored is used, and the factors are also expressed in state space form. The ﬁrst book to deal with factorization problems in a systematic way using the state space approach is the monograph [11] of the ﬁrst three authors. This monograph appeared in 1979, very soon after the ﬁrst main results were obtained. In fact, some of the factorization results were published in [11] for the ﬁrst time. The present book is the second book written by the four of us in which the state space factorization method is systematically used and developed further. In our ﬁrst book [20], published in 2008, the emphasis is on non-canonical factorizations and degree 1 factorizations, in particular. In the present book we concentrate on canonical factorizations. As a result the overlap between the main parts of the two books is minor. Together both books present a rich and far reaching update of the 1979 monograph [11]. In the present book special attention is paid to various factorizations with additional symmetries such as spectral factorization, inner-outer factorization, and J-spectral factorization. The latter require elements of the theory of spaces with an indeﬁnite metric. Factorizations with symmetries appear in a natural way in H∞ -control problems and the related Nehari approximation problem. In fact, the latter problems are the main topic of the ﬁnal part of the book. We also deal with applications to problems in the theory of algebraic Riccati equations, to inversion problems for Wiener-Hopf, Toeplitz and singular integral operators, and to Riemann-Hilbert problems. The linear transport equation from mathematical physics is another important area of application in this book. It requires inﬁnite dimensional realizations of a special type. We have made an eﬀort to make the text reasonably self-contained. For that reason we included some known material about realizations, minimal factorizations of rational matrix functions, angular operators, and the theory of matrices in indeﬁnite inner product spaces. In the ﬁnal part we also brieﬂy review elements of control theory of linear systems. Not counting the present introduction, the book consists of 20 chapters grouped into 7 parts. We shall now give a short description of the contents of the book. Part I. The ﬁrst part has a preparatory character. In the ﬁrst chapter we review the role of canonical factorization in inverting Wiener-Hopf integral operators and block Toeplitz operators. Also the role of this factorization in solving singular integral equations is described. The second chapter presents in detail the elements of the state space method that are used in this book. Part II. This part starts with the canonical factorization theorem for rational matrix functions in state space form. This theorem is then used to invert explicitly Wiener-Hopf, Toeplitz and singular integral operators with a rational matrix symbol, with the inverses being presented explicitly in state space formulas. For

3 rational matrix symbols the solution to the homogeneous Riemann-Hilbert boundary value problem is also given in state space form. In the ﬁrst chapter of this part we consider proper rational matrix functions, that is, rational matrix functions that are analytic at inﬁnity. The case of non-proper rational symbols is treated in the second chapter of this part. In this case the realization (1) is replaced by W (λ) = I + C(λG − A)−1 B,

(2)

where I is an identity matrix, G and A are square matrices, and B and C are matrices of appropriate sizes. A square rational matrix function, proper or not, always admits such a realization. We develop this realization result, and prove a canonical factorization theorem for the realization (2). As an application we solve the homogeneous Riemann-Hilbert boundary value problem for an arbitrary rational matrix symbol. Part III. In this part we carry out a program analogous to that of the second part, but now for certain classes of non-rational matrix and operator functions. For instance, for matrix functions analytic on a strip but not at inﬁnity we develop a realization theory, prove a canonical factorization theorem in state space form, and develop its applications to Wiener-Hopf integral equations. A new feature is that the problems involved require us to employ realizations with an unbounded main operator A and deal with curves cutting through the spectrum of this main operator. In this part it is also shown that, after an appropriate modiﬁcation, the state space method can be used to solve the integro-diﬀerential equation appearing in linear transport theory, which forces us to use realizations of operator-valued functions. In the ﬁnal chapter of this part we make an excursion into non-canonical Wiener-Hopf factorization for analytic operator-valued functions on a curve, and identify the so-called factorization indices in state space terms. Part IV. The fourth part deals with factorization of rational matrix functions that have Hermitian values on the imaginary axis, the real line or the unit circle. In the analysis of such functions, minimal realizations play an important role. These are realizations of which the order of the state matrix in (1) is a small possible. Also the so-called state space similarity theorem, which tells us that a minimal realization is unique up to a basis transformation in the state space, enters into the analysis. These facts are reviewed in the ﬁrst chapter of this part. In this ﬁrst chapter, using the notion of local minimality, also the concept of a pseudocanonical factorization relative to a curve is introduced and studied for rational matrix functions with singularities on the given curve. The eﬀect on minimal realizations of the function having Hermitian values on the imaginary axis, the real line or the unit circle is described in the second chapter of this part. This then leads to the construction of special canonical and pseudo-canonical factorizations with additional relations between the factors. Included are spectral factorization for positive deﬁnite rational matrix functions and pseudo-spectral factorization for nonnegative rational matrix functions. In the ﬁnal chapter we present (without proofs) some background material on matrices in indeﬁnite inner product spaces,

4

Chapter 0. Introduction

and review the main results from this area that are used in this book. Part V. In this part the canonical factorization theorem is presented in a diﬀerent way using the notion of an angular subspace and Riccati equations. In this case one has to look for angular subspaces that are also spectral subspaces, and the solutions of the Riccati equation must have additional spectral properties. These results, which have a preliminary character, are presented in the ﬁrst chapter of this part. In the second chapter we introduce the symmetric algebraic Riccati equation, and describe spectral factorization as well as pseudo-spectral factorization in terms of Hermitian solutions of such a Riccati equation. In the ﬁnal chapter of this part we continue the study of rational matrix functions that take Hermitian values on certain curves. The emphasis will be on rational matrix functions that have Hermitian values for which the inertia is independent of the point on the curve. Such functions may still admit a symmetric canonical factorization, provided we allow for a constant Hermitian invertible matrix in the middle. Such a factorization is commonly known as a J-spectral factorization. Necessary and suﬃcient conditions for its existence are given, ﬁrst in terms of invariant subspaces and then in terms of solutions of a corresponding symmetric algebraic Riccati equation. We also study the question when a function which admits a left J-spectral factorization admits a right J-spectral factorization too. Part VI. In this part we study rational matrix functions that are unitary or of the form identity matrix plus contractions, and rational matrix functions that have a positive real part. Because of the state space similarity theorem, these additional symmetries can be restated in terms of special properties of the minimal realizations of the rational matrix functions considered. These reformulations involve an algebraic Riccati equation. The results are known in systems theory as the bounded real lemma and the positive real lemma, respectively. They allow us to solve related canonical and pseudo-canonical factorization problems in state space form. In the ﬁnal chapter of this part realizations are used to analyze rational matrix functions of which the values on the imaginary axis are J-unitary matrices. Solutions to various factorization problems are given. Special attention is paid to factorization of J-unitary rational matrix functions into J-unitary factors. In this chapter we also discuss problems of embedding a contractive rational matrix function into a unitary rational matrix function of larger size. Part VII. In this part the state space theory of J-spectral factorization, developed in the ﬁnal chapter of the ﬁfth part, is used to solve H∞ problems. The ﬁrst chapter of this part contains the solution of the Nehari interpolation problem for rational matrix interpolants. The second chapter presents a short review of elements of control theory that play an important role in the third (and ﬁnal) chapter of this part. This ﬁnal chapter is about H∞ -control. Here we use the Jspectral factorization theory to obtain the solutions of some of the main problems in this area, namely the standard problem, the one-sided problem, and the full model matching problem.

5 As the description of the contents given above shows, the emphasis in the book is mainly on rational matrix functions and ﬁnite dimensional realizations. An exception is Part III. The latter part deals with non-rational matrix functions and operator-valued functions, and it uses realizations that have an inﬁnite dimensional state space. Other exceptions are Chapter 2 in Part I and Chapter 12 in Part V. For the material in the other chapters of the book, in particular, in Parts IV–VII, often extensions to an inﬁnite dimensional setting exist; they require appropriate modiﬁcations. See, e.g., the books [5], [35], [42], [73], and the references therein. A few remarks about terminology and notation At the end of this book, after the bibliography, the reader will ﬁnd a List of Symbols and an Index. The latter contains in alphabetical order the various terms that are used in this book with references to the pages where they are introduced. In addition, we would like to mention the following. In the sequel, whenever convenient, a p × q matrix with complex entries will be identiﬁed with the (linear) operator from Cq into Cp deﬁned by the canonical action of the matrix on the standard orthogonal basis of Cq . Conversely, a linear operator from Cq into Cp is identiﬁed with its p × q matrix representation with respect to the standard orthogonal bases of Cq and Cp .

F−

F +

Γ

Throughout the word “operator” refers to a bounded linear transformation acting between Banach or Hilbert spaces (ﬁnite or inﬁnite dimensional). We assume the reader to be familiar with Sections I.1 and I.2 in [51] which contain the standard spectral theory of operators, including the notion of a Riesz projection and the corresponding functional calculus (see, also Chapter V in [144]). In particular, we shall often use the notions of a Cauchy domain and Cauchy contour which are deﬁned as follows. A Cauchy domain is an open set in the complex plane C consisting of a ﬁnite number of components such that its boundary is composed of a ﬁnite number of simple closed non-intersecting rectiﬁable curves. A Cauchy contour Γ is the positively oriented boundary of a bounded Cauchy domain. We write F+ for the interior domain of Γ, and F− for the exterior domain, i.e., the

6

Chapter 0. Introduction

complement of the closure F+ of F+ in the Riemann sphere C∞ = C ∪ {∞}. The picture on the previous page illustrates this notion. We shall also work with the extended real line and the extended imaginary axis as contours on the Riemann sphere C∞ . For the real line the orientation will be from left to right and for the imaginary axis from bottom to top. Thus for the extended real line the interior domain is the open upper half plane, which will be denoted by C+ ; for the extended imaginary axis it is the open left half plane, which is denoted by Cleft . We shall also freely use the Lesbesgue integral and related Lp spaces (see, e.g., Appendix 2 in [53]). Functions which are equal almost everywhere (shorthand: a.e.) are often identiﬁed, sometimes without explicitly mentioning this. Finally, when dealing with inner-outer factorization, we shall always assume that the outer factor is invertible outer (see Section 17.6). In the outer-co-inner factorizations considered in this book, the outer factor will be assumed to be invertible outer as well.

Part I Convolution equations, canonical factorization and the state space method This part has a preparatory character. It consists of two chapters. In the ﬁrst chapter we review the role of canonical factorization in inverting Wiener-Hopf integral operators and block Toeplitz operators. The role of this factorization in solving singular integral equations is described as well. The second chapter presents in detail the basic elements of the state space method that are used throughout this book. The central notion is that of a realization of a matrix or operator function. Three important operations on realizations are studied.

Chapter 1

The role of canonical factorization in solving convolution equations This chapter has a preparatory character. We review (without giving proofs) the role of canonical factorization in inverting systems of convolution equations. The chapter consists of three sections. Section 1.1 deals with Wiener-Hopf integral equations, Section 1.2 with block Toeplitz equations, and Section 1.3 with singular integral equations.

1.1 Wiener-Hopf integral equations and factorization In this section we outline the factorization method of [61] to solve systems of Wiener-Hopf integral equations. Such a system may be written as a single vectorvalued Wiener-Hopf equation ∞ φ(t) − k(t − s)φ(s) ds = f (t), t ≥ 0. (1.1) 0

(−∞, ∞), that is, Here φ and f are m-dimensional vector functions and k ∈ Lm×m 1 the kernel function k is an m×m matrix function whose entries are in L1 (−∞, ∞). We assume that the given vector function f has its component functions in the Lebesgue space Lp [0, ∞), and we express this property by writing f ∈ Lm p [0, ∞). Throughout this section p will be ﬁxed and 1 ≤ p < ∞. The problem we shall consider is to ﬁnd a solution φ of equation (1.1) that also belongs to the space Lm p [0, ∞). The usual method (see [61]) for solving equation (1.1) is as follows. First assume that (1.1) has a solution φ in Lm p [0, ∞). Extend φ and f to the full real

10

Chapter 1. The role of canonical factorization

line by putting f (t) = −

φ(t) = 0,

0

∞

k(t − s)φ(s) ds,

t < 0.

Then φ, f ∈ Lm p (−∞, ∞) and the full line convolution equation φ(t) −

∞

−∞

k(t − s)φ(s) ds = f (t),

−∞ < t < ∞

is satisﬁed. By applying the Fourier transformation and leaving the part of f that is given in the right-hand side, one gets W (λ)Φ+ (λ) − F− (λ) = F+ (λ),

λ ∈ R,

(1.2)

where W (λ) = Im − Φ+ (λ) =

∞

∞

e

iλt

k(t) dt,

−∞

eiλt φ(t) dt,

0

F+ (λ) = F− (λ) =

∞

0 0

eiλt f (t) dt,

(1.3)

eiλt f (t) dt.

(1.4)

−∞

Here Im is the m×m identity matrix. Note that the functions K and F+ are given, but the functions Φ+ and F− have to be found. In fact in this way the problem to solve (1.1) is reduced to that of ﬁnding two functions Φ+ and F− such that (1.2) holds, while furthermore Φ+ and F− must be as in (1.4) with φ ∈ Lm p [0, ∞) and f ∈ Lm (−∞, 0]. p To ﬁnd Φ+ and F− of the desired form such that (1.2) holds, one factorizes the m×m matrix function W appearing in (1.2). This function is called the symbol of the integral equation (1.1). Note that W is continuous on the real line, and by the Riemann-Lebesgue lemma limλ∈R, λ→∞ W (λ) exists and is equal to Im . Assume that the symbol admits a factorization of the following form: λ ∈ R, (1.5) W (λ) = Im + G− (λ) Im + G+ (λ) , where

G+ (λ) =

with g+ ∈ nants

∞

e 0

Lm×m [0, ∞) 1

iλt

g+ (t) dt,

and g− ∈

G− (λ) =

Lm×m (−∞, 0] 1

det Im + G+ (λ) ,

0

−∞

eiλt g− (t) dt,

while, in addition, the determi-

det Im + G− (λ)

do not vanish in the closed upper and lower half plane, respectively. We shall refer to the factorization (1.5) as a right canonical factorization of W with respect to

1.1. Wiener-Hopf integral equations and factorization

11

−1 the real line. Under the conditions stated above the functions Im + G+ (λ) and −1 admit representations as Fourier transforms: Im + G− (λ) ∞ −1 = Im + eiλt γ+ (t) dt, (1.6) Im + G+ (λ)

−1 Im + G− (λ)

=

Im +

0

0 −∞

eiλt γ− (t) dt,

(1.7)

with γ+ ∈ Lm×m [0, ∞) and γ− ∈ Lm×m (−∞, 0]. Using the factorization (1.5) and 1 1 omitting the variable λ, equation (1.2) can be rewritten as (Im + G+ )Φ+ − (Im + G− )−1 F− = (Im + G− )−1 F+ .

(1.8)

Let P be the projection acting on the Fourier transforms of Lm p (−∞, ∞)-functions according to the following rule: ∞ ∞ eiλt h(t) dt = eiλt h(t) dt. P −∞

0

Applying P to (1.8) one gets (Im + G+ )Φ+ = P (Im + G− )−1 F+ , and hence

Φ+ = (Im + G+ )−1 P (Im + G− )−1 F+ ,

(1.9)

which is the formula for the solution of equation (1.2). To obtain the solution φ of the original equation (1.1), i.e., to obtain the inverse Fourier transform of Φ+ , one can employ the formulas (1.6) and (1.7). In fact ∞ γ(t, s)f (s) ds, t ≥ 0, φ(t) = f (t) + 0

where the m × m matrix function γ(t, s) is given by γ(t, s) = γ+ (t − s) + γ− (t − s) +

0

min(t, s)

γ+ (t − r)γ− (r − s) dr.

We conclude the description of this factorization method by mentioning that the m equation (1.1) has a unique solution in Lm p [0, ∞) for each f in Lp [0, ∞) if and only if its symbol admits a factorization as in (1.5). For details, see [50], [61]. Let T be the Wiener-Hopf integral operator on Lm p [0, ∞) associated with equation (1.1), that is, T is the operator on Lm [0, ∞) given by p (T φ)(t) = φ(t) −

0

∞

k(t − s)φ(s) ds,

t ≥ 0.

12

Chapter 1. The role of canonical factorization

The function W in the left-hand side of (1.3) is also referred to as the symbol of T . Obviously the operator T is invertible if and only if the equation (1.1) has a m unique solution in Lm p [0, ∞) for each f in Lp [0, ∞). Thus the results reviewed above can be summarized as follows. Theorem 1.1. Let T be the Wiener-Hopf integral operator on Lm p [0, ∞) with symbol W . Then T is invertible if and only if W admits a right canonical factorization with respect to the real line. Furthermore, if (1.5) is such a factorization of W , then the inverse of T is the integral operator given by ∞ γ(t, s)f (s) ds, t ≥ 0, (T −1 f )(t) = f (t) + 0

where the kernel function γ is deﬁned by s ⎧ ⎪ ⎪ γ (t − s) + γ+ (t − r)γ− (r − s) dr, ⎨ + 0 γ(t, s) = t ⎪ ⎪ ⎩ γ− (t − s) + γ+ (t − r)γ− (r − s) dr, 0

0 ≤ s < t, (1.10) 0≤t<s

with γ− and γ+ as in (1.6) and (1.7), respectively. To illustrate the method, let us consider a special choice for the right-hand side f (cf., [61]). Take f (t) = e−iqt x0 , (1.11) where x0 is a ﬁxed vector in Cm and q is a complex number with q < 0. Then ∞ i x0 , F+ (λ) = ei(λ−q)t x0 dt = λ ≥ 0. λ−q 0 Now observe that −1 −1 i − Im + G− (q) x0 Im + G− (λ) λ−q is the Fourier transform of an Lm p (−∞, 0]-function and hence it vanishes when the projection P is applied. It follows that in the present case the formula for Φ+ may be written as Φ+ (λ) =

−1 −1 i Im + G+ (λ) Im + G− (q) x0 . λ−q

Recall that the solution φ is the inverse Fourier transform of Φ+ . So we have t −1 eiqs γ+ (s) ds Im + G− (q) x0 . φ(t) = e−iqt Im + 0

(1.12)

1.2. Block Toeplitz equations and factorization

13

1.2 Block Toeplitz equations and factorization In this section we consider the discrete analogue of a Wiener-Hopf integral equation, that is, a block Toeplitz equation . So we consider an equation of the type ∞

aj−k ξk = ηj ,

j = 0, 1, 2, . . . .

(1.13)

k=0

Throughout we assume that the coeﬃcients aj are given complex m × m matrices satisfying ∞

aj < ∞, (1.14) j=−∞

η = (ηj )∞ j=0 is a given (ξk )∞ ∈ m p such that k=0

m and vector from m p = p (C ). The problem is to ﬁnd ξ = (1.13) is satisﬁed. We shall restrict ourselves to the case 1 ≤ p ≤ 2; the ﬁnal results however are valid for 2 < p ≤ ∞ as well. Assume ξ ∈ m p is a solution of (1.13). Then one can write (1.13) in the form ∞

aj−k ξk = ηj ,

j = 0, ±1, ±2, . . . ,

(1.15)

k=−∞

where ξk = 0 for k < 0 and ηj is deﬁned by (1.15) for j < 0. Multiplying both sides of (1.15) by λj with |λ| = 1 and summing over j, one gets a(λ)ξ+ (λ) − η− (λ) = η+ (λ), where a(λ) =

∞

λj aj ,

η+ (λ) =

j=−∞

ξ+ (λ) =

∞

|λ| = 1, ∞

λj ηj ,

(1.16)

(1.17)

j=0

λj ξj ,

η− (λ) =

j=0

−1

λj ηj .

j=−∞

In this way the problem to solve (1.13) is reduced to that of ﬁnding two sequences ξ+ and η− such that (1.16) holds, while moreover, ξ+ and η− must be as in (1.2) ∞ m with (ξj )∞ j=0 and (η−j−1 )j=0 from p . The usual way (cf., [61] or the book [40]) of solving (1.16) is again by factorizing the symbol a(λ) of the given block Toeplitz equation. Assume that a(λ) admits a right canonical factorization with respect to the unit circle . By deﬁnition this means that a(λ) can be written as a(λ) = h+ (λ)

=

h− (λ)h+ (λ), ∞

j=0

λj h+ j ,

|λ| = 1, h− (λ) =

0

j=−∞

(1.18) λj h− j ,

14

Chapter 1. The role of canonical factorization

− ∞ m×m ∞ where (h+ of all absolutely convergent j )j=0 and (h−j )j=0 belong to the space 1 sequences of complex m× m matrices, det h+ (λ) = 0 for |λ| ≤ 1 and det h− (λ) = 0 −1 for |λ| ≥ 1 (including λ = ∞). Then h−1 + and h− also admit a representation of the form ∞ 0

−1 j + (λ) = λ γ , h (λ) = λj γj− , (1.19) h−1 + − j j=0

j=−∞

− ∞ m×m . Deﬁning the projection P by with (γj+ )∞ j=0 and (γ−j )j=0 from 1

P

∞

j

λ bj

j=−∞

=

∞

λj bj ,

j=0

one gets from (1.16) and (1.18) −1 ξ+ = h−1 + P h− η+ .

(1.20)

Here, for convenience, the variable λ is omitted. The solution of the original equation (1.13) can now be written as ∞

ξk =

γks ηs ,

k = 0, 1, . . . ,

(1.21)

s=0

where

γks =

⎧ s

⎪ ⎪ ⎪ γ+ γ− , ⎪ ⎪ ⎨ r=0 k−r r−s

s ≤ k,

k ⎪

⎪ ⎪ + − ⎪ γk−r γr−s , ⎪ ⎩

s ≥ k.

r=0

Note that for s = k both sums in the above formula deﬁne the same matrix. The assumption that a(λ) admits a right canonical factorization as in (1.18) m is equivalent to the requirement that for each η = (ηj )∞ j=0 in p the equation m (1.13) has a unique solution ξ = (ξk )∞ in . For details we refer to [61], [40]. p k=0 m Let T be the block Toeplitz operator on p associated with the Toeplitz equation (1.13), that is, T is the operator on m p given by Tξ = η

⇐⇒

∞

aj−k ξk = ηj ,

j = 0, 1, 2, . . . .

k=0

The function a appearing in the left-hand side of (1.17) is also referred to as the m symbol of T . Obviously T is invertible if and only if for each η = (ηj )∞ j=0 in p m the equation (1.13) has a unique solution ξ = (ξk )∞ in . This allows us to p k=0 summarize the results reviewed above as follows.

1.3. Singular integral equations and factorization

15

Theorem 1.2. Let T be the block Toeplitz operator on m p with symbol a(λ) satisfying (1.14). Then T is invertible if and only a(λ) admits a right canonical factorization with respect to the unit circle. Furthermore, if (1.18) is such a factorization of the function a(λ), then the inverse of T is given by ⎡ ⎤ γ11 γ12 γ13 · · · ⎢γ21 γ22 γ23 · · ·⎥ ⎢ ⎥ T −1 = ⎢γ31 γ32 γ33 · · ·⎥ , ⎣ ⎦ .. .. .. .. . . . . where the matrices γks are deﬁned by ⎧ s

⎪ ⎪ ⎪ γ+ γ− , ⎪ ⎪ ⎨ r=0 k−r r−s γks = k ⎪ ⎪ ⎪ + − ⎪ γk−r γr−s , ⎪ ⎩

s ≤ k, (1.22) s ≥ k,

r=0

with γj+ and γj− being determined by (1.19). By way of illustration, we consider the special case when ηj = q j η0 ,

j = 0, 1, . . . .

(1.23)

Here η0 is a ﬁxed vector in Cm and q is a complex number with |q| < 1. Then clearly 1 η0 , η+ (λ) = |λ| ≤ 1, 1 − λq and one checks without diﬃculty that formula (1.21) becomes ξk = q k

k

−1 q −s γs+ h−1 )η0 , − (q

k = 0, 1, . . . .

(1.24)

s=0

This is the analogue of formula (1.12) in the previous section.

1.3 Singular integral equations and factorization In this section we review the factorization method that is used to solve systems of singular integral equations [48]. Consider the singular integral equation 1 φ(τ ) dτ = f (t), t ∈ Γ, (1.25) a(t)φ(t) + b(t) πi Γ τ − t with integration taken over a Cauchy contour Γ. (For the deﬁnition of the latter notion see the ﬁnal paragraphs of Chapter 0 dealing with terminology and notation.) We write F+ for the interior domain of Γ, and F− for the exterior domain

16

Chapter 1. The role of canonical factorization

(i.e., the complement of F + in the Riemann sphere C ∪ {∞}). The functions a and b in (1.25) are given continuous m × m matrix functions deﬁned on Γ, and f is a given function from Lm p (Γ), p ﬁxed, 1 < p < ∞. As usual in the theory of singular integral equations, it is assumed that the interior domain F+ of Γ is connected and contains 0; the exterior domain F− of Γ contains ∞. The problem is to ﬁnd φ ∈ Lm p (Γ) such that(1.25) is satisﬁed. For φ a rational function without poles on Γ we put 1 φ(τ ) (Sφ)(t) = dτ = f (t), t ∈ Γ, (1.26) πi Γ τ − t where the integral is taken in the sense of the Cauchy principal value. The operator S deﬁned in this way can be extended by continuity to a bounded linear operator, again denoted by S, on all of Lm p (Γ). Equation (1.25) can now be written as aIφ + bSφ = f,

(1.27)

where I is the identity operator on Lm p (Γ). In other words, the study of the equation (1.25) reduces to that of the operator aI + bS. Here a and b are viewed as multiplication operators. Equation (1.25) has a unique solution φ ∈ Lm p (Γ) for each choice of f ∈ Lm p (Γ) if and only if the operator aI + bS is invertible as an operator on Lm p (Γ). In the remainder of this section we shall discuss a necessary and suﬃcient condition for this to happen, and we shall give formulas for the inverse (aI + bS)−1 . The operator S enjoys the property S 2 = I. Hence the operators PΓ =

1 (I + S), 2

QΓ =

1 (I − S) 2

are complementary projections on Lm p (Γ). The image of P Γ consists of all functions in Lm (Γ) that admit an analytic continuation into F+ . Similarly, the image of QΓ p is the set of all functions in Lm (Γ) that admit an analytic continuation into F− p vanishing at ∞. Putting c = a + b and d = a − b, one can write the equation (1.27) in the form cP Γ φ + dQΓ φ = f . The following is known (see [62] for the case when the coeﬃcients a and b are scalar functions and [48] for the matrix-valued case). The operator aI + bS = cP Γ + dQΓ is invertible if and only if the matrices c(λ) and d(λ) are invertible for each λ ∈ Γ and the function w given by w(λ) = d(λ)−1 c(λ) admits a right canonical factorization with respect to Γ . By this we mean a factorization w(λ) = w− (λ)w+ (λ),

λ ∈ Γ,

(1.28)

where w− and w+ are m×m matrix functions, analytic and taking invertible values on an open neighborhood of F − and F + , respectively. With the help of (1.28), the −1 operator aI +bS = cP Γ +dQΓ can be rewritten as aI +bS = dw− (w+ P Γ +w− QΓ ),

1.3. Singular integral equations and factorization

17

and its inverse is given by (aI + bS)−1

−1 −1 −1 = (w+ P Γ + w− QΓ )w− d −1 −1 −1 −1 −1 = w+ P Γ w− d + w− QΓ w− d .

(1.29)

Replacing P Γ and QΓ by 12 (I + S) and 12 (I − S), respectively, one gets (aI + bS)−1

1 −1 1 −1 −1 −1 (c + d−1 )I + (w+ − w− )Sw− d 2 2 1 1 −1 −1 [(a + b)−1 + (a − b)−1 ]I + (w+ = − w− )Sw− (a − b)−1 2 2 1 −1 −1 = (a + b)−1 a(a − b)−1 I + (w+ − w− )Sw− (a − b)−1 . 2 =

Summarizing we get the following theorem. Theorem 1.3. The singular integral operator T = aI + bS on Lm p (Γ) is invertible if and only if the matrices a(λ) + b(λ) and a(λ) − b(λ) are invertible for each λ ∈ Γ and the function w given by −1 w(λ) = a(λ) + b(λ) a(λ) + b(λ) admits a right canonical factorization with respect to Γ. Furthermore, if (1.28) is such a factorization of w, then the inverse of T is given by 1 −1 −1 − w− )Sw− (a − b)−1 . T −1 = (a + b)−1 a(a − b)−1 I + (w+ 2

(1.30)

Thus, as before for Wiener-Hopf and block Toeplitz operators, canonical factorization is a useful method for inverting singular integral operators too.

Notes The material in this chapter is standard, and can be found in much more detail and greater generality in various monographs and papers, for instance, see the books [29] and [50]. A ﬁrst introduction to the theory of Wiener-Hopf integral equations and the theory of (block) Toeplitz operators can be found in Chapters XII and XIII of [51] and Chapters XXIII–XXV of [52], respectively. More information can be found in the monographs [37], [62], [63], [64] and [24]. For an extensive review (with many additional references) of the factorization theory of matrix functions with respect to a curve and its applications to inversion of singular integral operators of diﬀerent types, including Wiener-Hopf and block Toeplitz operators, the reader is referred to the recent survey paper [59].

Chapter 2

The state space method and factorization This chapter describes in detail the elements of the state space method that are used throughout this book. The central notion is that of a realization of a matrix or operator function. The chapter consists of six sections. Section 2.1 presents preliminaries on realization, including the relevant deﬁnitions and the connection with systems theory. In the next two sections the realization problem is discussed. First for rational matrix functions in Section 2.2, and then for analytic operator functions in a possibly inﬁnite dimensional setting in Section 2.3. The last three sections are devoted to the main operations on realizations that are needed in this book: inversion (Section 2.4), taking products (Section 2.5), and factorization (Section 2.6).

2.1 Preliminaries on realization Let W be a rational matrix function which is also proper, that is, W has no pole at inﬁnity. As is well-known such a function can always be represented (see the next section for an explicit construction) in the form W (λ) = D + C(λI − A)−1 B.

(2.1)

Here λ is a complex variable, A is a square matrix, I is the identity matrix of the same size as A, and B and C are matrices of appropriate sizes. Since A, B, C and D are matrices, it is immediate from Cramer’s rule that the right-hand side of (2.1) is also a proper rational matrix function. We shall understand the equality in (2.1) as an equality between rational matrix functions, and we shall refer to (2.1) as a matrix-valued realization of W . Sometimes we simply say that the quadruple of matrices (A, B, C, D) is a realization of W . A rational matrix function has many

20

Chapter 2. The state space method and factorization

diﬀerent realizations. Of particular interest are those matrix-valued realizations of W of which the order of the matrix A is as small as possible. These realizations are called minimal ; we shall describe their properties in Chapter 8. For operator-valued functions W , expressions of the type (2.1) are important too but have to be considered with some care. Let W be an L(U, Y )-valued function on a subset Ω of C. Here U and Y are possibly inﬁnite dimensional complex Banach spaces. We say that W admits a realization on Ω whenever W can be written as W (λ) = D + C(λIX − A)−1 B,

λ ∈ Ω.

(2.2)

Here A is a bounded linear operator on a complex Banach space X such that Ω is a subset of ρ(A), the resolvent set of A. Furthermore, IX is the identity operator on X, and B ∈ L(U, X), C ∈ L(X, Y ), D ∈ L(U, Y ), that is B : U → X, C : X → Y, and D : U → Y, are bounded linear operators. The fact that Ω ⊂ ρ(A) implies that the right-hand side of (2.2) is a well-deﬁned bounded linear operator which maps U into Y for each λ ∈ Ω. Also, W (λ) is a bounded linear operator mapping U into Y for each λ ∈ Ω. Note that (2.2) requires these operators to be equal for each λ ∈ Ω. When Ω is open, an obvious necessary condition for W to admit a realization on Ω is that W be analytic on Ω. When Ω is a punctured open neighborhood of ∞, then (2.2) implies limλ→∞ W (λ) = D and so W is proper. Often the identity matrix I in (2.1) and the identity operator IX in (2.2) will be suppressed, and we simply write (λ − A)−1 in place of (λI − A)−1 or (λIX − A)−1 . When X and Y are both ﬁnite dimensional, then the realization (2.2) is called ﬁnite dimensional . In that case W (λ), A, B, C and D can be identiﬁed in the usual way with matrices. In the next two sections we shall address the realization problem, i.e., the question under what conditions a given matrix or operator function admits a realization. First however, we sketch a connection with systems theory which reﬂects itself in some terminology to be introduced at the end of the present section. A system Σ can be considered as a physical object which produces an output in response to an input. Schematically:

u

Σ

y

where u denotes the input and y denotes the output. Mathematically, the input u and the output y are vector-valued functions of a parameter t. The input can

2.1. Preliminaries on realization

21

be chosen freely (at least in principle), but the output is uniquely determined by the choice of the input. The relationship between the input and the output can be quite complicated. Here we consider the simplest model which means that the relationship in question is described by a causal linear time invariant system, i.e., a system of diﬀerential equations of the type ⎧ ⎪ ⎪ x (t) = Ax(t) + Bu(t), ⎨ y(t) = Cx(t) + Du(t), t ≥ 0, Σ (2.3) ⎪ ⎪ ⎩ x(0) = 0, where A, B, C and D are matrices of appropriate sizes, A and D square. Application of the Laplace transform (under appropriate conditions on the input and output functions) changes (2.3) into

λ x(s) =

A x(λ) + B u(λ),

y(λ)

Cx (λ) + D u(λ),

=

and from these expressions one can solve y(λ) in terms of u (λ), resulting in (λ). y(λ) = D + C(λ − A)−1 B u So in what is called the frequency domain, the input-output behavior of (2.3) is determined by the function D + C(λ− A)−1 B, which is called the transfer function of the system (2.3). Note that this function appears in the realized form. The connection with systems theory indicated above is reﬂected in the terminology which is customarily used in dealing with realizations. Returning to (2.2), the space X is usually called the state space of the realization, and the operator A is referred to as its state space operator or main operator . Further we call B the input operator , C the output operator , and D the external operator of (2.2). The realization is called strictly proper when D = 0 and biproper if D is an invertible operator. In the latter case, the operator A − BD−1 C is well-deﬁned. It is referred to by the term the associate state space operator or associate main operator and (by slight abuse of notation as A× does not depend only on A) denoted by A× . This operator will play a crucial role in the inversion and factorization results to be discussed later on. In the situation where U = Y and D is the identity operator, we say that (2.2) is a unital realization. The associate main operator then has the form A× = A − BC. In the case of a matrix-valued realization, the terms state space matrix , main matrix , input matrix , output matrix , external matrix , associate state space matrix , and associate main matrix will be used. Other elements of systems theory involving stability properties, feedback and stabilization, will be reviewed in Chapter 19. These will be of central importance in Chapter 20 (the ﬁnal chapter of the book) which is concerned with H∞ -control.

22

Chapter 2. The state space method and factorization

2.2 Realization of rational matrix functions In this section we construct a matrix-valued realization for a given proper rational (possibly non-square) matrix function. Theorem 2.1. Every proper rational matrix function has a matrix-valued realization. Moreover, the realization can be chosen in such a way that the set of eigenvalues of the main matrix coincides with the set of poles of W . Proof. Let W be a proper rational r × m matrix function, and let wij be the (i, j)-entry of W . Since W is rational, we have wij (λ) =

pij (λ) , qij (λ)

i = 1, . . . , r, j = 1, . . . , n,

where pij and qij are scalar polynomials. The polynomials qij are non-zero and can be taken to be monic. Without loss of generality we may assume that the polynomials pij and qij have no common zero. Taking the least common multiple of the polynomials qij , we obtain a monic polynomial q. Deﬁne ΩW to be the set of all complex λ for which q(λ) = 0. Notice that C \ ΩW is precisely the set of all points in C where W has a pole. One checks without diﬃculty that W has a representation of the form W (λ) = W (∞) +

1 H(λ), q(λ)

λ ∈ ΩW ,

where H is an r ×m matrix polynomial. Since W is proper, this matrix polynomial is either identically equal to zero or it has degree strictly smaller than k, the degree of the scalar polynomial q. Write q(λ) = λk +

k−1

λj qj ,

H(λ) =

j=0

k−1

λj Hj ,

j=0

and, with Ir the r × r identity matrix, ⎡

0 ⎢ I ⎢ A=⎢ ⎢ ⎣ 0

0 0 ..

... ...

0 0

.

...

I

−q0 Ir −q1 Ir .. . −qk−1 Ir

⎤ ⎥ ⎥ ⎥, ⎥ ⎦

⎡ ⎢ ⎢ B=⎢ ⎢ ⎣

H0 H1 .. . Hk−1

⎤ ⎥ ⎥ ⎥, ⎥ ⎦

C = 0 . . . 0 Ir .

Then the resolvent set ρ(A) of A coincides with ΩW , the subset of C on which q takes non-zero values. For λ ∈ ρ(A), deﬁne C1 (λ), . . . , Ck (λ) by [ C1 (λ) C2 (λ) . . . Ck (λ) ] = C(λ − A)−1 .

2.3. Realization of analytic operator functions

23

From the special form of the matrix A (second companion type) we see that Cj+1 (λ) = λCj (λ),

j = 0, . . . , k − 1,

and C1 (λ) = q(λ)−1 I. Hence C(λ − A)

−1

B =

k−1

j=0

Cj+1 (λ)Hj =

1 H(λ). q(λ)

It follows that W (λ) = W (∞) + C(λ − A)−1 B for each λ ∈ ΩW = ρ(A). Thus W has a matrix-valued realization such that the set of eigenvalues of the main matrix A is equal to C \ ΩW . In other words, the set of eigenvalues of A coincides with the set of poles of W, as desired. Let W be a proper rational matrix function. Elaborating on Theorem 2.1 and its proof, we note that W does not admit any realization involving a main matrix A whose spectrum σ(A) is strictly smaller than C \ ΩW , the set of poles of W . Indeed, we would then have a realization of W on an open subset of C strictly larger than ΩW and such a subset would contain a pole of W , contradicting the fact that W has to be analytic on it. It is not diﬃcult to construct realizations of W having a main matrix A with spectrum strictly larger than C \ ΩW and where certain eigenvalues of A (namely those belonging to ΩW ) do not correspond with poles of W . So the realization constructed in the proof of Theorem 2.1 enjoys a certain minimality property. However, it does this only in a weak sense. This one sees, for instance, by looking at the pole orders. If μ is a pole of W , its order as a pole of W is generally strictly smaller than the order of μ as a pole of the resolvent (λ − A)−1 . With the proper notion of minimality to be introduced in Section 8.1, this anomaly disappears so that the two pole orders are the same. The key point is that the state space dimension (which is equal to rk) of the realization of the proof of Theorem 2.1 is generally not the least possible.

2.3 Realization of analytic operator functions In this section we consider the realization problem for possibly non-rational operator functions. First we consider operator functions that are analytic on a bounded Cauchy domain in C. Recall from Chapter 0 that the boundary of such a Cauchy domain consists of a ﬁnite number of simple closed non-intersecting rectiﬁable curves. Theorem 2.2. Let Ω be a bounded Cauchy domain, and let W be an operator function with values in L(U, Y ), where U and Y are complex Banach spaces. Suppose W is analytic on Ω and continuous on the closure of Ω. Then, given a bounded linear operator D : U → Y , there exists a realization for W on Ω having D as its external operator. In particular, if U = Y , then W admits a unital realization on Ω.

24

Chapter 2. The state space method and factorization

Proof. Let Γ be the positively oriented boundary of Ω (so that Ω is the interior domain of Γ). With Γ and U we associate the space C(Γ; U ) of all U -valued continuous functions on Γ endowed with the supremum norm. This will become the state space of the realization to be constructed. Write B for the canonical embedding of U into C(Γ; U ), so (Bu)(z) = u for each u ∈ U and z ∈ Γ. Next, deﬁne C : C(Γ; U ) → Y by setting 1 Cf = D − W (z) f (z) dz, f ∈ C(Γ; U ). 2πi Γ Here D is the given operator from U into Y . Finally, the operator A from C(Γ; U ) into C(Γ; U ) is the multiplication operator given by (Af )(z) = zf (z),

f ∈ C(Γ; U ), z ∈ Γ.

All these operators are linear and bounded. We claim that W (λ) = D + C(λ − A)−1 B,

λ ∈ Ω ⊂ ρ(A).

Take λ ∈ Ω. Then λ − A is invertible with inverse given by (λ − A)−1 g (z) =

1 g(z), λ−z

g ∈ C(Γ; U ), z ∈ Γ.

It follows that (λ − A)−1 Bu (z) =

1 u, λ−z

u ∈ U, z ∈ Γ,

and hence C(λ − A)−1 Bu =

1 2πi

Γ

1 D − W (z) u dz, λ−z

u ∈ U.

By the Cauchy integral formula, the right-hand side of this identity is W (λ)u−Du, and the desired result is immediate. Theorem 2.2 remains true when the conditions on Ω and W are replaced by the simpler hypotheses that Ω is any bounded open set in C and W is just analytic on Ω. In that case the space C(Γ; U ) must be replaced by an appropriate Banach space deﬁned in terms of the behavior of W near the boundary of Ω. For details, cf., [113]; see also the next theorem. Theorem 2.3. Let Ω ⊂ C be an open punctured neighborhood of ∞ in the Riemann sphere C∞ , let U and Y be complex Banach spaces, and let W : Ω → L(U, Y ) be analytic and proper. Then W admits a realization on Ω with external operator D = limλ→∞ W (λ).

2.3. Realization of analytic operator functions

25

Proof. First assume Ω is the full complex plane. Then, by Liouville’s theorem, the function W has the constant value D = limλ→∞ W (λ). Now take for the state space X the zero space {0}, and the desired realization for W on C is obtained trivially. Next, consider the more interesting case where Ω is diﬀerent from C. For notational reasons we will assume that 0 ∈ / Ω. The general case can be reduced to this situation by a simple translation. Deﬁne X to be the space of all Y -valued functions, analytic on Ω ∪ {∞}, such that f (z) < ∞. f • = sup z∈Ω ∪{∞} max(1, W (z)) Taking · • for the norm, X is a Banach space. Introduce B : U → X by ⎧ z ∈ Ω, ⎨ z W (z)u − W (∞)u , (Bu)(z) = ⎩ lim z W (z)u − W (∞)u , z = ∞. z→∞

Further, let C : X → Y be given by Cf = f (∞). Finally, deﬁne A : X → X by ⎧ z ∈ Ω, ⎨ z f (z) − f (∞) , (Af )(z) = ⎩ lim z f (z) − f (∞) , z = ∞. z→∞

All these operators are linear and bounded. We claim that W (λ) = W (∞) + C(λ − A)−1 B,

λ ∈ Ω ⊂ ρ(A).

Take λ ∈ Ω. For g ∈ X, put ⎧ zg(λ) − λg(z) ⎪ , ⎪ ⎪ ⎨ z−λ h(z) = g(λ) − λg (λ), ⎪ ⎪ ⎪ ⎩ g(λ),

z ∈ Ω, z = λ, z = λ, z = ∞,

where g stands for the derivative of g. Then h ∈ X, and by direct computation one sees that (λ − A)h (z) = λg(z), z ∈ Ω ∪ {∞}. Now λ is non-zero (since Ω does not contain the origin), and it follows that λ − A is surjective. But λ − A is injective too. Indeed, if f ∈ X and Af = λf , then f (z) =

z f (∞), z−λ

z ∈ Ω, z = λ,

which, on account of the deﬁnition of the norm · • on X, implies f (∞) = 0 (cf., the behavior of f when z → λ), hence f = 0. It follows that λ ∈ ρ(A) and

26

Chapter 2. The state space method and factorization

(λ − A)−1 g = λ−1 h. We now apply this result tog = Bu with u ∈ U . With this g, we have h(∞) = (Bu)(λ) = λ W (λ)u − W (∞) u, and so (λ − A)−1 Bu (∞) = λ−1 h(∞) = W (λ)u − W (∞) u. In other words C(λ − A)−1 Bu = W (λ)u − W (∞) u. As u ∈ U was taken arbitrarily, we get W (λ) = W (∞) + C(λ − A)−1 B for each λ ∈ Ω.

2.4 Inversion We begin with some heuristics. Consider the realization W (λ) = D + C(λ − A)−1 B,

λ ∈ ρ(A),

(2.4)

and view W as the transfer function of the linear time invariant system ⎧ x (t) = Ax(t) + Bu(t), ⎪ ⎪ ⎨ y(t) = Cx(t) + Du(t), t ≥ 0, Σ ⎪ ⎪ ⎩ x(0) = 0. Assuming that we are in the biproper situation where D is invertible, we can solve u in terms of x and y: u(t) = −D−1 Cx(t) + D−1 y(t),

t ≥ 0.

Inserting this into Σ yields ⎧ x (t) = A× x(t) + BD−1 y(t), ⎪ ⎪ ⎨ u(t) = −D−1 Cx(t) + D−1 y(t), Σ× ⎪ ⎪ ⎩ x(0) = 0.

t ≥ 0,

Here A× = A − BD−1 C is the associate main operator of the given realization as introduced in the last paragraph of Section 2.1. The linear time invariant systems Σ and Σ× can be seen as each other’s inverse. The transfer function of Σ is given by (2.4), the transfer function of Σ× by W × (λ) = D−1 − D−1 C(λ − A× )−1 BD −1 ,

λ ∈ ρ(A× ).

So it is to be expected that W and W × are related by inversion. We shall now make this precise. Theorem 2.4. Consider the biproper realization W (λ) = D + C(λ − A)−1 B,

λ ∈ ρ(A).

2.5. Products

27

Put A× = A − BD−1 C, and take λ ∈ ρ(A). Then W (λ) is invertible if and only if λ belongs to ρ(A× ). In that case, for λ ∈ ρ(A) ∩ ρ(A× ), the following identities hold: W (λ)−1

=

D−1 − D−1 C(λ − A× )−1 BD−1 ,

(λ − A× )−1

=

(λ − A)−1 − (λ − A)−1 BW (λ)−1 C(λ − A)−1 .

Moreover, again for λ ∈ ρ(A) ∩ ρ(A× ), we have W (λ)D−1 C(λ − A× )−1

=

C(λ − A)−1 ,

(λ − A× )−1 BD−1 W (λ)

=

(λ − A)−1 B.

Proof. For λ ∈ ρ(A× ), put W × (λ) = D−1 − D−1 C(λ − A× )−1 BD−1 . Then, when λ ∈ ρ(A) ∩ ρ(A× ), one has W (λ)W × (λ) = D + C(λ − A)−1 B D −1 − D−1 C(λ − A× )−1 BD−1 =

IY + C(λ − A)−1 BD −1 − C(λ − A× )−1 BD −1 + −C(λ − A)−1 BD−1 C(λ − A× )−1 BD−1 .

Now use that BD −1 C = A−A× = (λ − A× )−(λ − A). It follows that W (λ)W × (λ) = IY . Analogously one has W × (λ)W (λ) = IU . The expression for (λ − A× )−1 as well as the last two identities in the theorem are obtained in a similar way. Instead of the previous proof one can also give an argument using Schur complements of the operator matrix A − λI B . C I For details, see the second proof of Theorem 2.1 in [20] or Sections 2 and 4 in [19].

2.5 Products Again we begin with some heuristical remarks. This time we start with two linear time invariant systems ⎧ ⎪ ⎪ x1 (t) = A1 x1 (t) + B1 u1 (t), ⎨ Σ1 t ≥ 0, y1 (t) = C1 x1 (t) + D1 u1 (t), ⎪ ⎪ ⎩ x1 (0) = 0,

28

Chapter 2. The state space method and factorization ⎧ x (t) = ⎪ ⎪ ⎨ 2 y2 (t) = Σ2 ⎪ ⎪ ⎩ x2 (0) =

A2 x2 (t) + B2 u2 (t), t ≥ 0,

C2 x2 (t) + D2 u2 (t), 0,

and we assume that the output y2 of Σ2 can be and is used as the input u1 = y2 for Σ1 , resulting in the cascade synthesis Σ of the systems Σ1 and Σ2 . The input for Σ is u = u2 and the output (modulo u1 = y2 ) is y = y1 . The equations governing the relationship between u and y then are ⎧ x1 (t) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ x (t) ⎪ ⎨ 2 y(t) ⎪ ⎪ ⎪ ⎪ x1 (0) ⎪ ⎪ ⎪ ⎩ x2 (0)

= A1 x1 (t) + B1 C2 x2 (t) + B1 D2 u(t), = A2 x2 (t) + B2 u(t), = C1 x1 (t) + D1 C2 x2 (t) + D1 D2 u(t),

t ≥ 0,

= 0, = 0,

and this is a linear time invariant system which can be rewritten as ⎧ ⎪ x1 A1 B1 C2 x1 B1 D2 ⎪ ⎪ ⎪ = + u, ⎪ ⎪ ⎪ x2 0 A2 x2 B2 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ x1 Σ: C1 D1 C2 y = + D1 D2 u, ⎪ x2 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ x1 0 ⎪ ⎪ ⎪ (0) = . ⎪ ⎩ 0 x2 The transfer functions of Σ1 and Σ2 are W1 (λ)

=

D1 + C1 (λ − A1 )−1 B1 ,

λ ∈ ρ(A1 ),

(2.5)

W2 (λ)

=

D2 + C2 (λ − A2 )−1 B2 ,

λ ∈ ρ(A2 ),

(2.6)

respectively, and the transfer function of Σ is the product W1 W2 of W1 and W2 , in other words W (λ) = W1 (λ)W2 (λ). So our considerations lead to a product formula for realizations. Here are the details. First we specify the spaces associated with the realizations (2.5) and (2.6), and the actions of the operators involved: A 1 : X 1 → X1 ,

B1 : U1 → X1 ,

C1 : X1 → Y1 ,

D1 : U1 → Y1 ,

2.5. Products

29

A2 : X2 → X2 ,

B2 : U2 → X2 ,

C2 : X2 → Y2 ,

D2 : U2 → Y2 .

Now assume Y1 = U2 . Put U = U1 , Y = Y2 , and introduce A1 B1 C2 ˙ 2 → X1 +X ˙ 2, A = : X1 +X 0 A2 B

B1 D2

=

˙ 2, : Y → X1 +X

B2

C1

D1 C2

C

=

D

= D1 D2 : U → Y.

˙ 2 → Y, : X1 +X

Then the following result holds true. Theorem 2.5. Let W1 and W2 be given by the realizations (2.5) and (2.6), respectively. Then, with A, B, C and D as above, W1 (λ)W2 (λ) = D + C(λ − A)−1 B,

λ ∈ ρ(A1 ) ∩ ρ(A2 ) ⊂ ρ(A).

Proof. Take λ ∈ ρ(A1 ) ∩ ρ(A2 ). Then λ ∈ ρ(A). Indeed, λ − A is invertible with inverse given by ⎤ ⎡ −1 (λ − A1 ) H(λ) ⎦ : X1 +X ˙ 2 → X1 +X ˙ 2, (λ − A)−1 = ⎣ −1 0 (λ − A2 ) −1

−1

where H(λ) = − (λ − A1 ) B1 C2 (λ − A2 ) . Employing this, and the expressions for B, C and D given prior to the theorem, D + C(λ − A)−1 B is seen to be equal to

D1 D 2 +

C1

D1 C2

⎡ ⎣

−1

(λ − A1 ) 0

= D 1 D2 + =

C1 (λ − A1 )

−1

D1 + C1 (λ − A1 )−1 B1

⎤⎡

H(λ) −1

⎦⎣

(λ − A2 )

B1 D2

⎤ ⎦

B2

C1 H(λ) + D1 C2 (λ − A2 )

−1

B1 D2

B2

D2 + C2 (λ − A2 )−1 B2 .

Thus D + C(λ − A)−1 B = W1 (λ)W2 (λ), as desired.

The product W1 (λ)W2 (λ) is deﬁned for λ ∈ ρ(A1 ) ∩ ρ(A2 ), a punctured neighborhood of ∞ in C ∪ {∞}. On the other hand D + C(λ − A)−1 B is deﬁned

30

Chapter 2. The state space method and factorization

for λ ∈ ρ(A). As we have seen above ρ(A1 ) ∩ ρ(A2 ) ⊂ ρ(A). In general, this inclusion is strict. Equality occurs, for instance, when the spectra σ(A1 ) and σ(A2 ) of the operators A1 and A2 are disjoint. Another case where one has the equality ρ(A) = ρ(A1 ) ∩ ρ(A1 ) is when ρ(A) is connected. In particular, the equality in question is valid when W1 and W2 are rational matrix functions, and (2.5) and (2.6) are matrix-valued realizations. The realization of Theorem 2.5 is called the product of the realizations (2.5) and (2.6), in that order. The counterpart of taking products is factorization. In the next section this topic will be discussed for functions given by a biproper realization. We close the present section with a remark preparing for this discussion. The main operator A in the product realization is given in the form of a 2 × 2 upper triangular operator matrix: A1 B1 C2 ˙ 2 → X1 +X ˙ 2. A= : X1 +X 0 A2 Analogously, assuming the external operators to be invertible, the associate main operator A× = A − BD−1 C is of 2 × 2 lower triangular type: 0 A× 1 ˙ 2 → X1 +X ˙ 2 : X1 +X A× = B2 D−1 C1 A× 2 −1 × −1 where A× 1 = A1 − B1 D1 C1 and A2 = A2 − B2 D2 C2 are the associate main ˙ {0} is an invariant operators of (2.5) and (2.6), respectively. Note that M = X1 + ˙ 2 is an invariant subspace for A× , and that M subspace for A, that M × = {0} +X and M × match in the sense that the state space of the product realization is the direct sum of M and M × . This state of aﬀairs turns out to be a key point in the discussion of factorization we now turn to.

2.6 Factorization The theorems in this section will serve as a basis for the more involved factorization results to be given in the sequel. Subspaces of Banach spaces are always assumed to be closed, otherwise we use the term linear manifold. For simplicity (and without loss of generality) we assume the external spaces U and Y to be equal. Theorem 2.6. Consider the biproper realization W (λ) = D + C(λIX − A)−1 B,

λ ∈ ρ(A),

(2.7)

and let A× = A − BD−1 C be its associate main operator. Let M and M × be invariant subspaces for A and A× , respectively, and suppose X = M M ×.

(2.8)

2.6. Factorization

31

Assume D = D1 D2 , where D1 and D2 are invertible operators on Y , and write A1 A+ ˙ × → M +M ˙ ×, : M +M A = 0 A2 B

=

C

=

B1 B2

C1

˙ ×, : Y → M +M C2

˙ × → Y. : M +M

Introduce the functions W1 and W2 via the biproper realizations W1 (λ)

= D1 + C1 (λIM − A1 )−1 B1 D2−1 ,

λ ∈ ρ(A1 ),

(2.9)

W2 (λ)

= D2 + D1−1 C2 (λIM × − A2 )−1 B2 ,

λ ∈ ρ(A2 ).

(2.10)

Then W admits the factorization W (λ) = W1 (λ)W2 (λ),

λ ∈ ρ(A1 ) ∩ ρ(A2 ) ⊂ ρ(A).

The function W is deﬁned and analytic on ρ(A), while the factors W1 and W2 are deﬁned and analytic on the sets ρ(A1 ) and ρ(A2 ), respectively. In particular, the factors may be deﬁned and analytic on domains where the left-hand side is not. This will turn out to be relevant in applications (cf., the remarks made at the end of this section). ˙ × in the usual manner, the product of the realProof. Identifying X and M +M izations (2.9) and (2.10) is precisely the realization (2.7). The desired result now follows from Theorem 2.5. We shall refer to (2.8) as the matching condition, and when this condition is satisﬁed we refer to M, M × as a pair of matching subspaces. A pair of matching subspaces M, M × satisfying A[M ] ⊂ M,

A× [M × ] ⊂ M ×

will be called a supporting pair of subspaces for the realization (2.7). Matching pairs of subspaces correspond to projections. So Theorem 2.6 has a counterpart in terms of projections. We say that a projection Π : X → X is a supporting projection for the realization (2.7) if A[Ker Π] ⊂ Ker Π,

A× [Im Π] ⊂ Im Π.

Here Ker T stands for the null space of an operator or matrix T , and Im T for its range.

32

Chapter 2. The state space method and factorization

Theorem 2.7. Let Π be a supporting projection for the biproper realization W (λ) = D + C(λIX − A)−1 B,

λ ∈ ρ(A).

Assume D = D1 D2 , where D1 and D2 are invertible operators on Y , and introduce the functions W1 and W2 via the biproper realizations W1 (λ)

= D1 + C(λIX − A)−1 (IX − Π)BD2−1 ,

λ ∈ ρ(A),

W2 (λ)

= D2 + D1−1 CΠ(λIX − A)−1 B,

λ ∈ ρ(A).

Then W (λ) = W1 (λ)W2 (λ) for all λ ∈ ρ(A). This factorization holds on the resolvent set ρ(A) of A. However, in many cases (relevant for applications), the factors in the right-hand side have an analytic extension to larger domains (see Theorem 2.6; cf., also the remarks made at the end of this section). Proof. The fact that Π is a supporting projection for the given biproper realization means nothing else than that the identities ΠA = ΠAΠ and A× Π = ΠA× Π are satisﬁed. Hence (I − Π)(A − A× )Π = AΠ − ΠA. Now take λ ∈ ρ(A). Then W1 (λ)W2 (λ)

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 (I − Π)BD−1 CΠ(λ − A)−1 B

as desired.

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 (I − Π)(A − A× )Π(λ − A)−1 B

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 (AΠ − ΠA)(λ − A)−1 B

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 Π(λ − A) − (λ − A)Π (λ − A)−1 B

=

D + C(λ − A)−1 (I − Π)B + CΠ(λ − A)−1 B +C(λ − A)−1 ΠB − CΠ(λ − A)−1 B,

=

D + C(λ − A)−1 B = W (λ),

The material presented above contains two factorization results: Theorems 2.6 and 2.7. These theorems contain not only diﬀerent expressions for the factors, these factors also have diﬀerent domains. For rational matrix functions and matrix-valued realizations, the diﬀerences are not essential. In the case of an inﬁnite dimensional state space one has to be more careful, the reason being that

2.6. Factorization

33

ρ(A1 ) ∩ ρ(A2 ) can then be a proper subset of ρ(A). For an exhaustive discussion of the issues involved, see Section 2.5 in [20]. We shall meet the diﬀerences referred to above when the factorization results are applied, as will be done later on, for solving Wiener-Hopf, Toeplitz or singular integral equations. In that context, it is also necessary to have information on the sets where the factors take invertible values and to have expressions for the inverses. In other words, it is necessary to have a good understanding of the relationship between Theorems 2.6 and 2.7 on the one hand, and the inversion result Theorem 2.4 on the other. The point here is that, by taking inverses, the factorizations of the function W (λ) given in Theorems 2.6 and 2.7 directly induce factorizations of the point-wise inverse W −1 of W , that is the function given by W −1 (λ) = W (λ)−1 , while on the other hand factorizations of W −1 can also be obtained by applying Theorems 2.6 and 2.7 to the realization W −1 (λ) = D−1 − D−1 C(λ − A× )−1 BD−1 .

(2.11)

Note here that if M, M × is a supporting pair of subspaces for the realization (2.7), then M × , M is a supporting pair of subspaces for the realization (2.11), and, analogously, if Π is a supporting projection for (2.7), then I − Π is a supporting projection for (2.11). The analysis in [20], Section 2.5 also clariﬁes these matters; the upshot is that the two approaches lead to essentially the same result.

Notes The notion of a realization originates from the Kalman theory of linear timeinvariant systems [95]. The literature on the subject is rich; see, e.g., the text books [94], [33]. In a somewhat diﬀerent form the notion of realization also appears in the theory of characteristic operator functions [27], [141]. The realization problem has many diﬀerent faces, depending on the class of matrix or operator functions one is dealing with. The material of the ﬁrst two sections is standard. Theorem 2.1 is a variation on Theorem 4.20 in [10]. Other constructions of matrix-valued realizations, including realizations with smallest possible state space dimension, can be found in text books; see, e.g., [94], [33] or [85] and references given there. The realization theorems for analytic operator functions in Section 2.3 originate from [57]. The operations of inversion and taking products are standard in systems theory. Theorem 2.11 has a natural Schur complement interpretation; see Section 2.2 in [20] and the paper [19]. The factorization theorem in the ﬁnal section originates from [21]; see also the ﬁrst chapter of [11]. For a brief description of the history of the factorization principle presented here, we refer to the Editorial introduction in [54].

Part II Convolution equations with rational matrix symbols The canonical factorization theorem for rational matrix functions in state space form is the ﬁrst result presented and proved in this part. This theorem is then used to invert explicitly Wiener-Hopf, Toeplitz and singular integral operators with a rational matrix symbol, with the inverses being presented explicitly in state space formulas. For rational matrix symbols the solution to the homogeneous RiemannHilbert boundary value problem is also given in state space form. This part consists of two chapters. In the ﬁrst chapter (Chapter 3) we consider proper rational matrix functions, that is, rational matrix functions that are analytic at inﬁnity. The case of non-proper rational symbols is treated in the second chapter (Chapter 4). This requires a diﬀerent type of realization. This modiﬁed realization result is developed and a corresponding canonical factorization theorem is proved. As an application the homogeneous Riemann-Hilbert boundary value problem is solved for an arbitrary rational matrix symbol.

Chapter 3

Explicit solutions using realizations As we have seen in Chapter 1, canonical factorization serves as a tool to solve Wiener-Hopf integral equations, their discrete analogues, and the more general singular integral equations. In this chapter the state space factorization method developed in Chapter 2 is used to solve the problem of canonical factorization (necessary and suﬃcient conditions for its existence) and to derive explicit formulas for its factors. This is done in Section 3.1 for rational matrix functions and later in Section 7.1 for operator-valued transfer functions that are analytic on an open neighborhood of a curve. The results are applied to invert Wiener-Hopf integral equations with a rational matrix symbol (Section 3.2), block Toeplitz operators (Section 3.3) and singular integral equations (Section 3.4). The methods developed in this chapter also allow us to deal with the Riemann-Hilbert boundary value problem. This is done in the ﬁnal section which also contains material on the homogeneous Wiener-Hopf equation.

3.1 Canonical factorization of rational matrix functions in state space form In this section and the next one we shall consider the factorization theorems of Section 2.6 for the special case when the two factors satisfy additional spectral conditions. Recall from Chapter 0 that a Cauchy contour is the positively oriented boundary of a bounded Cauchy domain in C and that such a contour consists of a ﬁnite number of simple closed non-intersecting rectiﬁable curves. We say that a Cauchy contour Γ splits the spectrum σ(S) of a bounded linear operator S if Γ ∩ σ(S) = ∅. In that case σ(S) decomposes into two disjoint compact sets σ+ and σ− such that σ+ is in the interior domain of Γ and σ− is in the exterior domain

38

Chapter 3. Explicit solutions using realizations

of Γ. If Γ splits the spectrum of S, then we have a Riesz projection, also called spectral projection, associated with S and Γ, namely 1 P (S; Γ) = (λ − S)−1 dλ. 2πi Γ The subspace N = Im P (S; Γ) will be called the spectral subspace for S corresponding to the contour Γ (or to the spectral set σ+ ). Lemma 3.1. Let Y1 and Y2 be complex Banach spaces, and consider the operator S11 S12 ˙ Y2 → Y1 + ˙ Y2 . : Y1 + S= (3.1) 0 S22 ˙ Y2 such that Ker Π = Y1 . Then the compression Let Π be any projection of Y = Y1 + ΠS|Im Π : Im Π → Im Π and S22 : Y2 → Y2 are similar. Furthermore, Y1 is a spectral subspace for S if and only if σ(S11 ) ∩ σ(S22 ) = ∅, and in that case σ(S) = σ(S11 ) ∪ σ(S22 ) while, in addition, 1 −1 Y1 = Im P (S; Γ) = Im (λI − S) dλ , (3.2) 2πi Γ where Γ is a Cauchy contour around σ(S11 ) separating σ(S11 ) from σ(S22 ). ˙ Y2 along Y1 onto Y2 . As Ker P = Ker Π, Proof. Let P be the projection of Y = Y1 + we have P = P Π and the map E = P |Im Π : Im Π → Y2 is an invertible operator. Write S0 for the compression ΠS|Im Π : Im Π → Im Π of S to Im Π, and take x = Πy. Then ES0 x = P ΠSΠy = P SΠy = P SP Πy = S22 Ex, and hence S0 and S22 are similar. Now suppose σ(S11 ) ∩ σ(S22 ) = ∅. Then ρ(S11 ) ∪ ρ(S22 ) = C and hence ρ(S) ∩ ρ(S22 ) . ρ(S) = ρ(S) ∩ ρ(S11 ) The upper triangular form of S in (3.1) ensues ρ(S) ∩ ρ(S11 ) = ρ(S) ∩ ρ(S22 ) = ρ(S11 ) ∩ ρ(S22 ) and it follows that ρ(S11 ) ∪ ρ(S22 ) = ρ(S), an identity which can be rewritten as σ(S) = σ(S11 ) ∪ σ(S22 ). Still under the assumption that σ(S11 ) ∩ σ(S22 ) = ∅, let Γ be a Cauchy contour Γ around σ(S11 ) separating σ(S11 ) from σ(S22 ). Then Γ splits the spectrum of S. In fact, if λ ∈ Γ, then both λ − S11 and λ − S22 are invertible and ⎡ ⎤ (λ − S11 )−1 (λ − S11 )−1 S12 (λ − S22 )−1 ⎦ (λ − S)−1 = ⎣ −1 0 (λ − S22 )

3.1. Canonical factorization of rational matrix functions in state space form 39 which leads to an expression of the type P (S; Γ) =

I

∗

0

0

for the Riesz projection associated with S and Γ. In particular, it is clear that Y1 = Im P (S; Γ). So Y1 is a spectral subspace for S and (3.2) holds. Next assume that Y1 = Im Q, where Q is a Riesz projection for S. Put Π = I − Q, and let S0 be the restriction of S to Im Π. Then σ(S11 )∩ σ(S0 ) = ∅. By the ﬁrst part of the proof, the operators S0 and S22 are similar. So σ(S0 ) = σ(S22 ), and hence we have shown that σ(S11 ) ∩ σ(S22 ) = ∅. Let Γ be a Cauchy contour. As before (see the one but last paragraph in Chapter 0) we denote by F+ and F− the interior and exterior domain of Γ, respectively. Note that ∞ ∈ F− . Let W be a rational m × m matrix function, with W (∞) = I, analytic on an open neighborhood of Γ, whose values on Γ are invertible matrices. By a right canonical factorization of W with respect to Γ we mean a factorization λ ∈ Γ, (3.3) W (λ) = W− (λ)W+ (λ), where W− and W+ are rational m × m matrix functions, analytic and taking invertible values on (an open neighborhood of) F − and F + , respectively. If in (3.3) the factors W− and W+ are interchanged, we speak of a left canonical factorization. Theorem 3.2. Let Γ be a Cauchy contour and let W be a rational m × m matrix function, Suppose W admits the realization W (λ) = Im + C(λIn − A)−1 B such that the main matrix A has no eigenvalues on Γ. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (i) A× = A − BC has no eigenvalues on Γ, ˙ Ker P (A× ; Γ). (ii) Cn = Im P (A; Γ) + In that case, a right canonical factorization of W is given by W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

where the factors and their inverses can be written as W− (λ)

=

Im + C(λIn − A)−1 (I − Π)B,

W+ (λ)

=

Im + CΠ(λIn − A)−1 B,

W−−1 (λ)

=

Im − C(I − Π)(λIn − A× )−1 B,

W+−1 (λ)

=

Im − C(λIn − A× )−1 ΠB.

Here Π is the projection of Cn along Im P (A; Γ) onto Ker P (A× ; Γ).

40

Chapter 3. Explicit solutions using realizations

For left canonical factorizations an analogous theorem holds. In the result in ˙ Im P (A× ; Γ). question, (ii) is replaced by Cn = Ker P (A; Γ) + The expressions for the functions W− and W+ suggest that these functions are deﬁned on the resolvent set ρ(A) of A. Similarly, W−−1 and W+−1 seem to have ρ(A× ) as their domain. At ﬁrst sight this is at variance with the requirements for Wiener-Hopf factorization. We will address this point in the proof. Proof. From the deﬁnition given above it is clear that a necessary condition in order that W admits a right canonical factorization with respect to Γ is that W takes invertible values on Γ. By Theorem 2.4 this necessary condition is met if and only if (i) holds true. Assume that (i) is satisﬁed. The spectral projections P (A; Γ) and P (A× ; Γ) are then well-deﬁned. The image X− = Im P (A; Γ) of P (A; Γ) and the null space X+ = Ker P (A× ; Γ) of P (A× ; Γ) are invariant for A and A× , respectively. Suppose (ii) is fulﬁlled too, and write A− A0 B− A= , B= , C = C− C+ 0 A+ B+ for the matrix presentations of A, B and C with respect to the decomposition ˙ X+ . With Cn = X− + W− (λ) W+ (λ)

= IX− + C− (λ − A− )−1 B− , −1

= IX+ + C+ (λ − A+ )

B+ ,

λ ∈ ρ(A− ), λ ∈ ρ(A+ ),

we have (from Theorem 2.6) the factorization W (λ) = W− (λ)W+ (λ),

λ ∈ ρ(A− ) ∩ ρ(A+ ) ⊂ ρ(A).

As X− is a spectral subspace for A, we can apply Lemma 3.1 to show that σ(A− ) and σ(A+ ) are disjoint. But then ρ(A) = ρ(A− ) ∩ ρ(A+ ) and it follows that W (λ) = W− (λ)W+ (λ),

λ ∈ ρ(A− ) ∩ ρ(A+ ) = ρ(A).

(3.4)

Applying Lemma 3.1 once again we see that σ(A− ) = σ(A) ∩ F+ ,

σ(A+ ) = σ(A) ∩ F− ,

(3.5)

where F+ and F− are the interior and exterior domain of Γ, respectively. In a similar way one proves that × σ(A× − ) = σ(A ) ∩ F+ ,

× σ(A× + ) = σ(A ) ∩ F− .

(3.6)

Using the ﬁrst parts of (3.5) and (3.6), it now follows that W− is analytic and has invertible values on an open neighborhood of F − . Analogously, employing the second parts of (3.5) and (3.6), one gets that W+ is analytic and has invertible

3.1. Canonical factorization of rational matrix functions in state space form 41 values on an open neighborhood of F + . Thus (3.4) is a right canonical Wiener-Hopf factorization with respect to Γ. The projection Π of Cn along Im P (A; Γ) onto Ker P (A× ; Γ) is a supporting projection for the given realization of W . Also In −Π is a supporting projection for the realization W (λ)−1 = Im − C(λIn − A× )−1 B of W −1 . With this in mind, one checks without diﬃculty that W− , W+ , W−−1 and W+−1 can also be written as in the theorem. For an exhaustive discussion of the intricacies concerning inversion, factorization, and the combination of these operations (in fact: the relationship between Theorems 2.6, 2.7 and 2.4), see Section 2.5 in [20]. Note, however, that in the present case there is no ambiguity because we are working here with rational matrix functions. Next, suppose that W (λ) = W− (λ)W+ (λ) is a right canonical factorization ˙ Ker P (A× ; Γ). with respect to Γ. We only have to show that Cn = Im P (A; Γ) + × We ﬁrst prove that Im P (A; Γ) ∩ Ker P (A ; Γ) = {0}. Without loss of generality it may be assumed that the values of W− and W+ at inﬁnity are equal to Im . Suppose x ∈ Im P (A; Γ) ∩ Ker P (A× ; Γ), and consider (λ − A)−1 x. This function is analytic on an open neighborhood of F − . On the other hand the function (λ − A× )−1 x is analytic on an open neighborhood of F + . For λ in the intersection ρ(A) ∩ ρ(A× ), we have W (λ)C(λ − A× )−1

= C(λ − A× )−1 + C(λ − A)−1 BC(λ − A× )−1 = C(λ − A× )−1 + C(λ − A)−1 (A − A× )(λ − A× )−1 = C(λ − A)−1 ,

and it follows that W+ (λ)C(λ − A× )−1 = W− (λ)−1 C(λ − A)−1 . The analyticity properties of the factors W− , W+ and their inverses now imply that the function W+ (λ)C(λ − A× )−1 x = W (λ)−1 C(λ − A)−1 x is analytic on the Riemann sphere C∞ . By Liouville’s theorem it must be constant. As it takes the value zero at inﬁnity, it is identically zero. Hence both C(λ − A× )−1 x and C(λ − A)−1 x vanish. Next use the identity (λ − A× )−1 BC(λ − A)−1 = (λ − A)−1 − (λ − A× )−1 to obtain (λ − A× )−1 x = (λ − A)−1 x. But then this function is analytic on the Riemann sphere too. Using Liouville’s theorem again, we see that it must be identically zero. Thus x = 0. Observe that up to this point in the proof we have not used the ﬁnite dimensionality of the state space. It will play a role in the next paragraph. We now ﬁnish the proof by a duality argument. Let Γ∗ be the adjoint curve of Γ, i.e., the curve obtained from Γ by complex conjugation. Also introduce ¯ ∗ , V− (λ) = W− (λ) ¯ ∗ and the functions V, V+ and V− by putting V (λ) = W (λ) ¯ ∗ . Clearly V has the realization V (λ) = I + B ∗ (λ − A∗ )−1 C ∗ V+ (λ) = W+ (λ) and V (λ) = V+ (λ)V− (λ) is a left canonical factorization. Arguing as above, we may conclude that Ker P (A∗ , Γ∗ ) ∩ Im P (A× )∗ , Γ∗ = 0. It follows that

42

Chapter 3. Explicit solutions using realizations

Ker P (A∗ , Γ∗ ) + Im P (A× )∗ , Γ∗ = Cn . In the ﬁrst instance, this equality holds for the closure of Ker P (A∗ , Γ∗ ) + Im P (A× )∗ , Γ∗ , but in Cn all linear manifolds are closed. With minor modiﬁcations we could have worked in Theorem 3.2 with two curves, one splitting the spectrum of A and the other splitting the spectrum of A× (cf., [100]). Finally, let us mention that Theorem 3.2 remains true if the Cauchy contour Γ is replaced by the extended real line R∞ , i.e., the closure of the real line in the Riemann sphere C∞ . In that case F+ is the open upper half plane and F− is the open lower half plane. For details, see Theorem 4.5 at the end of Section 4.3 below which, by the way, deals with the situation where W is a not necessarily proper rational matrix function.

3.2 Wiener-Hopf integral operators In this section the general factorization result proved in the preceding sections is used to provide explicit formulas for solutions of ﬁnite systems of the Wiener-Hopf equation φ(t) −

∞

0

k(t − s)φ(s) ds = f (t),

t ≥ 0,

(3.7)

(−∞, ∞), i.e., where φ and f are m-dimensional vector functions and k ∈ Lm×m 1 the kernel function k is an m × m matrix function of which the entries are in L1 (−∞, ∞). We assume that the given vector function f has its component functions in Lp [0, ∞), and we express this property by writing f ∈ Lm p [0, ∞). Throughout this section, p will be ﬁxed and 1 ≤ p < ∞. The problem we shall consider is to ﬁnd a solution φ for equation (3.7) that also belongs to the space Lm p [0, ∞). As was explained in Section 1.1 the equation (3.7) has a unique solution in Lm p [0, ∞) for each f in Lm [0, ∞) if and only if its symbol I − K(λ) admits a factorization m p as in (1.5). Our aim is to apply the factorization theory developed in the previous sections to get the canonical factorization (1.5). Therefore, in the sequel we assume that the symbol is a rational m × m matrix function. As K(λ) is the Fourier (−∞, ∞)–function, the symbol is continuous on the real transform of an Lm×m 1 line. In particular, Im − K(λ) has no poles on the real line. Furthermore, by the Riemann-Lebesgue lemma, lim

λ∈R, |λ|→∞

K(λ) = 0,

which implies that the symbol Im − K(λ) has the value In at ∞. The fact that Im − K(λ) is rational is equivalent to the requirement that the kernel function k is in the linear space spanned by all functions of the form p(t)eiαt , t > 0, h(t) = t < 0, q(t)eiβt ,

3.2. Wiener-Hopf integral operators

43

where p(t) and q(t) are matrix polynomials in t with coeﬃcients in Cm×m , and α and β are complex numbers with α > 0 and β < 0. From Section 2.2 we know that the matrix function Im − K(λ) admits a realization Im − K(λ) = Im + C(λIn − A)−1 B such that the main matrix A has no real eigenvalues. In the next theorem we express the solvability of equation (3.7) in terms of such a realization and give explicit formulas for its solutions in the same terms. Theorem 3.3. Let Im − K(λ) = Im + C(λIn − A)−1 B be a realization for the symbol of equation (3.7), and suppose A has no real eigenvalues. In order that m (3.7) has a unique solution φ in Lm p [0, ∞) for each f in Lp [0, ∞), the following two conditions are necessary and suﬃcient: (i) A× = A − BC has no real eigenvalues; ˙ M × , where M is the spectral subspace of A corresponding to the (ii) Cn = M + eigenvalues of A in the upper half plane, and M × is the spectral subspace of A× corresponding to the eigenvalues of A× in the lower half plane. Assume conditions (i) and (ii) hold true, and let Π be the projection of Cn along M onto M × . Then Im − K(λ) admits a right canonical factorization with respect to the real line that has the form Im − K(λ) = Im + G− (λ) Im + G+ (λ) , λ ∈ R, where the factors and their inverses can be written as Im + G+ (λ)

=

Im + CΠ(λIn − A)−1 B,

Im + G− (λ) −1 Im + G+ (λ) −1 Im + G− (λ)

=

Im + C(λIn − A)−1 (In − Π)B,

=

Im − C(λIn − A× )−1 ΠB,

=

Im − C(In − Π)(λIn − A× )−1 B.

The functions γ+ and γ− in (1.6) and (1.7) are given by γ+ (t) = γ− (t) =

×

+iCe−itA ΠB,

t > 0, ×

−iC(In − Π)e−itA t B,

t < 0.

Finally, the solution φ to (3.7) can be written as ∞ φ(t) = f (t) + γ(t, s)f (s) ds, 0

⎧ × × ⎨ +iCe−itA ΠeisA B,

where γ(t, s) =

⎩

×

s < t, ×

−iCe−itA (In − Π)eisA B,

s > t.

44

Chapter 3. Explicit solutions using realizations

Proof. We have already mentioned that equation (3.7) has a unique solution in m Lm p [0, ∞) for each f in Lp [0, ∞) if and only if the symbol Im − K(λ) admits a right canonical factorization as in (1.5). So to prove the necessity and suﬃciency of the conditions (i) and (ii), it suﬃces to show that the conditions (i) and (ii) together are equivalent to the statement that Im − K(λ) admits a right canonical factorization as in (1.5). We ﬁrst observe that condition (i) is equivalent to the requirement that Im − K(λ) is invertible for all λ ∈ R (see Theorem 2.4). But then we can apply Theorem 3.2 in combination with the remark made at the end of Section 3.1 to prove the ﬁrst part of the theorem. Next assume that conditions (i) and (ii) hold true. Applying Theorem 3.2 once again, we get the desired formulas for Im + G+ (λ), Im + G− (λ) and their inverses. The formulas for γ+ and γ− are now obtained by noticing that

∞

×

eiλt e−itA Π dt

=

i(λ − A× )−1 Π,

=

−i(I − Π)(λ − A× )−1 ,

0 0

−∞

×

eiλt (I − Π)e−itA dt

λ ∈ ρ(A× ), λ ≥ 0, λ ∈ ρ(A× ), λ ≤ 0,

where I = In . The proof of the latter identity uses (the ﬁrst conclusion in) Lemma 3.1. It remains to prove the ﬁnal formula for γ(t, s). We use (1.10), and compute ﬁrst that ×

×

γ+ (t − r)γ− (r − s) = Ce−i(t−r)A ΠBC(I − Π)e−i(r−s)A B. Now Ker Π = M is A-invariant and Im Π = M × is A× -invariant. Thus ΠA(I−Π) = 0 and (I − Π)A× Π = 0, and it follows that ΠBC(I − Π) = Π(A − A× )(I − Π) = ΠA× − A× Π. But then γ+ (t − r)γ− (r − s) = =

×

×

Ce−i(t−r)A (A× Π − ΠA× )e−i(r−s)A B −i

× × d Ce−i(t−r)A Πe−i(r−s)A B. dr

Inserting this in (1.30) we obtain for s < t that γ(t, s) =

×

iCe−i(t−s)A ΠB − −i(t−s)A×

s

i 0

× × d Ce−i(t−r)A Πe−i(r−s)A B dr dr ×

×

ΠB − Ce−i(t−r)A Πe−i(r−s)A B|sr=0

=

iCe

=

iCe−itA ΠeisA B,

×

×

3.2. Wiener-Hopf integral operators

45

while for s > t we get ×

γ(t, s) = −iC(I − Π)e−i(t−s)A B + = −iC(I − Π)e

−i(t−s)A×

t

i 0

× × d Ce−i(t−r)A Πe−i(r−s)A B dr dr ×

×

B − Ce−i(t−r)A Πe−i(r−s)A B|tr=0

×

×

= −iCe−itA (I − Π)eisA B.

This completes the proof.

Corollary 3.4. Let Im − K(λ) = Im + C(λIn − A)−1 B be a realization for the symbol of equation (3.7). Assume that A and A× = A − BC have no spectrum on the real line, and that ˙ Cn = Im P +Ker P ×, (3.8) where P and P × are the Riesz projections of A and A× , respectively, corresponding to the spectra in the upper half plane. Fix x ∈ Ker P , and let the right-hand side of (3.7) be given by f (t) = Ce−itA x, t ≥ 0. Then the unique solution φ in Lm p [0, ∞) of equation (3.7) is given by ×

φ(t) = Ce−itA Πx,

t ≥ 0.

Here Π is the projection of Cn onto Ker P × along Im P . Proof. Since x ∈ Ker P , the vector e−itA x is exponentially decaying in norm when t → ∞, and thus the function f belongs to Lm p [0, ∞). Furthermore, the conditions (i) and (ii) in Theorem 3.3 are fulﬁlled, and hence equation (3.7) has a unique solution φ ∈ Lm p [0, ∞). Moreover from Theorem 3.3 we know that φ is given by φ(t) = f (t) + iCe−itA

×

t

0

−iCe

×

ΠeisA BCe−isA x ds

−itA×

t

∞

× (I − Π)eisA BCe−isA x ds .

Now use that ×

×

eisA BCe−isA = ieisA (iA× − iA)e−isA = i

d isA× −isA e e . ds

It follows that × × φ(t) = f (t) − Ce−itA ΠeisA e−isA x|t0 × × +Ce−itA (I − Π)eisA e−isA x|∞ . t

46

Chapter 3. Explicit solutions using realizations ×

×

Since (I − Π) = (I − Π)P × , the function (I − Π)eisA = (I − Π)P × eisA is exponentially decaying for s → ∞. As we have seen, the same holds true for e−isA x. Thus ×

×

×

φ(t) = f (t) − Ce−itA ΠeitA e−itA x + Ce−itA Πx ×

×

−Ce−itA (I − Π)eitA e−itA x ×

= f (t) + Ce−itA Πx − Ce−itA x ×

= Ce−itA Πx,

which completes the proof.

Finally, let us return to the special situation where the functionf is given by formula (1.11), and assume that the conditions (i) and (ii) in Theorem 3.3 are satisﬁed. Then the solution φ admits the representation t × φ(t) = e−iqt {Im + i Cei(q−A )s ΠB ds} (3.9) 0

·{Im − C(I − Π)(q − A× )−1 B}x0 ;

(3.10)

see formula (1.12).

3.3 Block Toeplitz operators In the previous section the factorization theory was applied to ﬁnite systems of Wiener-Hopf integral equations. In this section we carry out a similar program for their discrete analogues, block Toeplitz equations (cf., Section 1.2). So we consider an equation of the type ∞

aj−k ξk = ηj ,

j = 0, 1, 2, . . . .

(3.11)

k=0

Throughout we assume that the coeﬃcients aj are given complex m × m matrices satisfying ∞

aj < ∞, j=−∞

and η = (ηj )∞ j=0 m ξ = (ξk )∞ k=0 ∈ p

m is a given vector from m p = p (C ). The problem is to ﬁnd such that (3.11) is satisﬁed. As before, we shall apply our factorization theory. For that reason we assume j that the symbol a(λ) = ∞ j=−∞ λ aj is a rational m × m matrix function whose value at ∞ is Im . Note that a(λ) has no poles on the unit circle. Therefore the conditions on a(λ) are equivalent to the following assumptions:

3.3. Block Toeplitz operators

47

(j) the sequence (aj − δj0 Im )∞ j=0 is a linear combination of sequences of the form j r ∞ α j D j=0 , where |α| < 1, r is a nonnegative integer and D is a complex m × m matrix; (jj) the sequence (a−j )∞ j=1 is a linear combination of sequences of the form ∞ −j s ∞ δjk F j=1 , β j E j=1 , where |β| > 1, s and k are nonnegative integers and E and F are complex m × m matrices. From Section 2.2 we know that the matrix function a(λ) admits a realization a(λ) = Im + C(λIn − A)−1 B

(3.12)

such that the main matrix A has no eigenvalues on the unit circle. The next theorem is the analogue of Theorem 3.3. Theorem 3.5. Let (3.12) be a realization for the symbol a(λ) of the equation (3.11), and suppose A has no eigenvalues on the unit circle. Then (3.11) has a unique m ∞ m solution ξ = (ξk )∞ k=0 in p for each η = (ηj )j=0 in p if and only if the following two conditions are satisﬁed: (i) A× = A − BC has no eigenvalues on the unit circle, ˙ M × , where M is the spectral subspace of A corresponding to the (ii) Cn = M + eigenvalues of A inside the unit circle, and M × is the spectral subspace of A× corresponding to the eigenvalues of A× outside the unit circle. Assume conditions (i) and (ii) are satisﬁed, and let Π be the projection of Cn along M onto M × . Then the function a(λ) admits a right canonical factorization with respect to the unit circle that has the form a(λ) = h− (λ)h+ (λ),

|λ| = 1,

where the factors and their inverses can be written as h+ (λ)

= Im + CΠ(λIn − A)−1 B,

h− (λ)

= Im + C(λIn − A)−1 (In − Π)B,

h−1 + (λ)

= Im − C(λIn − A× )−1 ΠB,

h−1 − (λ)

= Im − C(In − Π)(λIn − A× )−1 B.

− ∞ The sequences (γj+ )∞ j=0 and (γ−j )j=0 in (1.19) are given by

γ0+

=

Im + C(A× )−1 ΠB,

γj+

=

C(A× )−(j+1) ΠB,

γ0−

=

Im ,

γj−

=

−C(In − Π)(A× )−(j+1) B,

j = 1, 2, . . . ,

j = −1, −2, . . . .

48

Chapter 3. Explicit solutions using realizations

Finally, the solution ξ to (3.11) can be written as ξk =

γks

∞

s=0

⎧ C(A× )−(k+1) Π(A× )s B, ⎪ ⎪ ⎪ ⎨ = Im + C(A× )−(s+1) Π(A× )s B, ⎪ ⎪ ⎪ ⎩ −C(A× )−(k+1) (In − Π)(A× )s B,

γks ηs where

s < k, s = k, s > k.

Proof. The proof of Theorem 3.5 is similar to that of Theorem 3.3. Here we only derive the ﬁnal formula for γks . With respect to the formulas for γj+ , we note that Im Π is A× -invariant and the restriction of A× to Im Π is invertible. So, with slight abuse of notation as far as inverses of A× are involved, h+ (λ)−1

=

Im − C(λ − A× )−1 ΠB −1 × −1 Im + C I − λ(A× )−1 (A ) ΠB

=

Im +

=

∞

λj C(A× )−(j+1) ΠB.

j=0

Now compare coeﬃcients with h+ (λ)−1 = γj− are obtained by comparing h− (λ)−1

∞

j=0

λj γj+ . Similarly, the formulas for

=

Im − C(I − Π)(λ − A× )−1 B

=

Im − C(I − Π)

=

Im −

−1

∞

1 × j−1 (A ) B λj j=1

λj C(I − Π)(A× )−(j+1) B

j=−∞

0 with h− (λ)−1 = j=−∞ λj γj− . Here I = In . To obtain the formulas for γks we employ (1.22). For s < k we must ﬁnd + γks = γk−s γ0− +

s−1

r=0

+ − γk−r γr−s ,

while for s > k we need to calculate γks =

− γ0+γk−s

+

k−1

r=0

+ − γk−r γr−s .

3.3. Block Toeplitz operators

49

Again by slight abuse of notation + − γk−r γr−s

=

−C(A× )−(k−r+1) ΠBC(I − Π)(A× )−(r−s+1) B

=

−C(A× )−(k−r+1) (A× Π − ΠA× )(A× )−(r−s+1) B

=

−C(A× )−(k−r) Π(A× )−(r−s+1) B + +C(A× )−(k−r+1) Π(A× )−(r−s) B.

Observe that if we replace r by r + 1 in the last one of the latter two terms we get the ﬁrst one. So the summation in the formula for γks is telescoping and collapses into just a few terms. We proceed as follows. For s < k we get + γks = γk−s γ0− − C(A× )−(k−s+1) ΠB + C(A× )−(k+1) Π(A× )s B. + Since γ0− = I and γk−s = C(A× )−(k−s+1) ΠB, this results in

γks = C(A× )−(k+1) Π(A× )s B. For s > k the computation is a little more involved as γ0+ = In + C(A× )−1 ΠB. Using that ΠBC I − Π) = A× Π − ΠA× , it goes this way: γks = − I + C(A× )−1 ΠB C I − Π)(A× )−(k−s+1) B +C(A× )−(k+1) Π(A× )s B − C(A× )−1 Π(A× )−(k−s) B =

−C I − Π)(A× )−(k−s+1) B +C(A× )−1 (ΠA× − A× Π)(A× )−(k−s+1) B +C(A× )−(k+1) Π(A× )s B − C(A× )−1 Π(A× )−(k−s) B

=

C(A× )−(k+1) Π(A× )s B − C(A× )−(k−s+1) B

=

−C(A× )−(k+1) (I − Π)(A× )s B.

It remains to consider the case k = s. Then we have γss = γ0+ γ0− +

k−1

+ − γs−r γr−s .

r=0

Following the line of argument as in the case s < k this yields γss

=

Im + C(A× )−1 ΠB − C(A× )−1 ΠB + C(A× )−(k+1) Π(A× )k B

=

Im + C(A× )−(k+1) Π(A× )k B,

which completes the proof.

50

Chapter 3. Explicit solutions using realizations

The main step in the factorization method for solving the equation (3.11) is to construct a right canonical factorization of the symbol a(λ) with respect to the unit circle. In Theorem 3.5 we obtained explicit formulas for the case when a(λ) is rational and has the value In at ∞. The latter condition is not essential. Indeed, by a suitable M¨ obius transformation one can transform the symbol α(λ) into a function which is invertible at inﬁnity (see Section 3.6). Next one makes the Wiener-Hopf factorization of the transformed symbol with respect to the image of the unit circle under the M¨ obius transformation. Here one can use the same formulas as in Theorem 3.5. Finally, using the inverse M¨obius transformation, one can obtain explicit formulas for the factorization with respect to the unit circle, and hence also for the solution of equation (3.11).

3.4 Singular integral equations In this section we apply Theorem 3.2 to solve the singular integral equation from Section 1.3: 1 φ(τ ) a(t)φ(t) + b(t) dτ = f (t), t ∈ Γ, (3.13) πi Γ τ − t where Γ is a Cauchy contour. The problem is to ﬁnd φ ∈ Lm p (Γ) such that (3.13) is satisﬁed. Recall that (3.13) can be rewritten in the form aIφ + bSφ = f , where S is the singular integral operator as in (1.26). Put c = a + b and d = a − b. Then we know from Section 1.3 that the operator aI + bS is invertible if and only if c(λ) and d(λ) are invertible for all λ ∈ Γ and the function w(λ) = d(λ)−1 c(λ) admits a right canonical factorization with respect to Γ. The next theorem deals with the case when w(λ) is rational and has the value Im at ∞. Theorem 3.6. Suppose det a(λ) + b(λ) and det a(λ) − b(λ) do not vanish on Γ, −1 a(λ) + b(λ) is a rational function which has and assume w(λ) = a(λ) − b(λ) the value Im at inﬁnity. Let w(λ) = Im + C(λIn − A)−1 B be a realization for w. Suppose A and A× = A − BC have no spectrum on Γ. ˙ × , where M is the spectral Then aI + bS is invertible if and only if Cn = M +M subspace corresponding to the eigenvalues of A inside Γ, and M × is the spectral subspace corresponding to the eigenvalues of A× outside Γ. In that case the func−1 −1 tions w+ , w+ , w− and w− appearing in the expressions for (aI + bS)−1 given in Section 1.3 can be speciﬁed as follows: w+ (λ)

= Im + CΠ(λIn − A)−1 B,

w− (λ)

= Im + C(λIn − A)−1 (In − Π)B,

−1 (λ) w+

= Im − C(λIn − A× )−1 ΠB,

−1 (λ) w−

= Im − C(In − Π)(λIn − A× )−1 B.

3.5. The Riemann-Hilbert boundary value problem

51

Here Π is the projection of Cn along M onto M × and I = In is the identity operator on Cn . By way of illustration, we consider the special case when 1 a(t) − b(t) η, f (t) = t−α where α is a complex number outside Γ and η ∈ Cm . Put g(t) =

1 η. t−α

−1 −1 −1 d = w− g. Then one can write f = dg, where as before d = a − b. Hence w− Observe now that the function 1 −1 −1 w− (t) − w− (α) η t−α

is analytic outside Γ and vanishes at ∞. So when we apply P Γ to it, we get zero. It follows that 1 −1 P Γ w− w−1 (α)η. g (t) = t−α − But then 1 −1 −1 −1 w− (t) − w− QΓ w− g (t) = (α) η, t−α and hence 1 1 −1 −1 −1 (t)w− (α)η + (α) η. w+ Im − w− (t)w− (aI + bS)−1 f (t) = t−α t−α In the situation of Theorem 3.6, the right-hand side of this equality becomes 1 1 η− C (t − A× )−1 Π + (t − A)−1 (I − Π) B t−α t−α · Im − C(I − Π)(α − A× )−1 B η. The case when w(λ) is rational, but does not have the value Im at ∞, can be treated by applying a suitable M¨ obius transformation. The argument is similar to that indicated at the end of Section 3.3.

3.5 The Riemann-Hilbert boundary value problem In this section we consider the (homogeneous) Riemann-Hilbert boundary value problem (on the real line): W (λ)Φ+ (λ) = Φ− (λ),

−∞ < λ < +∞.

(3.14)

52

Chapter 3. Explicit solutions using realizations

The precise formulation of this problem is as follows. Let W be a given m × m matrix function, with entries that are integrable on the real line. The problem is to describe all pairs Φ+ , Φ− of Cm -valued functions such that (3.14) is satisﬁed while, in addition, Φ+ and Φ− are the Fourier transforms of integrable Cm -valued functions with support in [0, ∞) and (−∞, 0], respectively. For such a pair of functions Φ+ , Φ− we have that Φ+ is continuous on the closed upper half plane, analytic in the open upper half plane and vanishes at inﬁnity, the same being true for Φ− with the understanding that the upper half plane is replaced by the lower. The functions W that we shall deal with are rational m × m matrix functions with the value Im at inﬁnity and given in the form of a realization. Theorem 3.7. Let W be a rational m × m matrix function, and suppose W admits the realization W (λ) = Im + C(λIn − A)−1 B. Suppose further that both A and A× = A − BC have no eigenvalues on the real line. Let M be the spectral subspace of A corresponding to the eigenvalues of A in the upper half plane, and let M × be the spectral subspace of A× corresponding to the eigenvalues of A× in the lower half plane. Then the pair of functions Φ+ , Φ− is a solution of the Riemann-Hilbert boundary value problem (3.14) if and only if there exists x ∈ M ∩ M × such that Φ+ (λ) = C(λIn − A× )−1 x,

Φ− (λ) = C(λIn − A)−1 x.

(3.15)

Moreover, the vector x in (3.15) is uniquely determined by the pair Φ+ , Φ− . Proof. Take x ∈ M ∩ M × and deﬁne Φ+ and Φ− by (3.15). From Theorem 2.4 we know that W (λ)C(λ−A× )−1 = C(λ−A)−1 . It follows that (3.14) is satisﬁed. Here × the speciﬁc choice of x does not even play a role. Put φ+ (t) = −iCe−itA x, t ≥ 0. Since x ∈ M × , the function φ+ is integrable on [0, ∞). Similarly, as x ∈ M , the function φ− given by φ− (t) = iCe−itA x, t ≤ 0 is integrable on (−∞, 0]. A straightforward computation shows that ∞ 0 Φ+ (λ) = eiλt φ+ (t)dt, Φ− (λ) = eiλt φ− (t)dt (3.16) 0

−∞

and the proof of the “if part” of the theorem is complete. The proof of the “only if part” is somewhat more involved. Let Φ+ , Φ− be a solution of (3.14) given in the form (3.16) with integrable φ+ and φ− . It will be convenient to extend φ+ and φ− to integrable functions on the full real line by stipulating that they vanish on [−∞, 0) and [0, ∞), respectively. For λ ∈ R put ρ(λ) = (λ − A)−1 BΦ+ (λ). Note that (λ − A)−1 appears as a Fourier transform of a matrix function with entries from L1 (R). In fact ∞ −1 eiλt (t)dt, λ ∈ R, (λ − A) = −∞

where (t) =

ie−itA P,

t < 0,

−itA

t < 0.

−ie

(In − P ),

3.5. The Riemann-Hilbert boundary value problem

53

Using inverse Fourier transforms and the fact that the support of φ+ is contained in [0, ∞), we have ∞ ∞ eiλt (t − s)Bφ+ (s) ds dt, λ ∈ R. ρ(λ) = −∞

Introduce

γ− (t) = γ+ (t) =

∞

0 ∞ 0

0

(t − s)Bφ+ (s) ds,

(t < 0),

γ− (t) = 0

(t > 0),

(t − s)Bφ+ (s) ds,

(t > 0),

γ+ (t) = 0

(t < 0),

and for each λ ∈ R set

ρ+ (λ) =

∞ −∞

ρ− (λ) =

∞ −∞

eiλt γ+ (t) dt = eiλt γ− (t) dt =

∞ 0

0

−∞

eiλt γ+ (t) dt, eiλt γ− (t) dt.

Obviously, ρ(λ) = ρ− (λ) + ρ+ (λ) for each λ ∈ R. From (3.14) and the deﬁnition of ρ it follows that Φ+ (λ) + Cρ+ (λ) = Φ− (λ) − Cρ− (λ),

λ ∈ R.

(3.17)

The left-hand side of (3.17) is continuous on the closed upper half plane, analytic in the open upper half plane and vanishes at inﬁnity. The same is true for the right-hand side of (3.17) provided the upper half plane is replaced by the lower half plane. But then we can apply Liouville’s theorem to show that both sides of (3.17) are identically zero. Hence 0 eiλt Cγ− (t) dt, λ ≤ 0, (3.18) Φ− (λ) = Cρ− (λ) = −∞

Φ+ (λ) = −Cρ+ (λ) = − For t < 0 we have γ− (t)

0

= where x =

∞

=

0

eiλt Cγ+ (t) dt,

λ ≥ 0.

(3.19)

(t − s)Bφ+ (s) ds

ie−itA

∞ 0

∞

∞

eisA P Bφ+ (s) ds = ie−itA x,

eisA P Bφ+ (s) ds. Clearly x ∈ Im P , and we conclude that 0 eiλt ie−itA x dt = (λ − A)−1 x, λ ≤ 0. ρ− (λ) = 0

−∞

(3.20)

54

Chapter 3. Explicit solutions using realizations

Next, ﬁx λ ∈ R. Since (λ − A)ρ(λ) = BΦ+ (λ) and (λ − A)ρ− (λ) = x, we can use the ﬁrst part of (3.19) to show that (λ − A)ρ+ (λ) + x = (λ − A)ρ(λ) = BΦ+ (λ) = −BCρ+ (λ). Recall that A× = A − BC. It follows that ρ+ (λ) = −(λ − A× )−1 x,

λ ∈ R.

(3.21)

The left-hand side of (3.21) is continuous on the closed upper half plane and analytic in the open upper half plane. Thus (3.21) implies that P × x = 0, where P × is the spectral projection of A× corresponding to the eigenvalues in the upper half plane. Since Im P = M and Ker P × = M × , we see that x ∈ M ∩ M × . From (3.19) and (3.21) it follows that the ﬁrst identity in (3.15) holds. Similarly, (3.18) and (3.20) yield the second identity in (3.15). It remains to prove the unicity of x. Take u ∈ M ∩ M × , and assume that C(λ−A)−1 u = 0. It suﬃces to show that u = 0. To do this, recall (see Theorem 2.4) that (λ − A× )−1 = (λ − A)−1 − (λ − A)−1 BW (λ)−1 C(λ − A)−1 ,

λ ∈ R.

Thus the assumption C(λ − A)−1 u = 0 yields (λ − A× )−1 u = (λ − A)−1 u,

λ ∈ R.

(3.22)

The fact that u ∈ M × implies that (λ − A× )−1 u is analytic on λ ≥ 0. On the other hand, u ∈ M gives that (λ − A)−1 is analytic on λ ≤ 0. Since both (λ − A× )−1 u and (λ − A)−1 u vanish at inﬁnity, Liouville’s theorem implies that (λ − A)−1 u is identically zero on R, hence u = 0. There is an intimate connection between the Riemann-Hilbert boundary value problem (on the real line) and the homogeneous Wiener-Hopf integral equation. This is already clear from the material presented in Section 1.1 by specializing to the situation where f = 0. The fact is further underlined by the above proof of Theorem 3.7. Indeed, notice that (3.19) implies that φ+ = −Cγ+ , and hence we see from the deﬁnition of γ+ that φ+ (t) −

0

∞

k(t − s)φ+ (s) ds = 0,

t > 0,

where k(t) = −C(t)B, and hence k(λ) = −C(λIn − A)−1 B. Thus φ+ is the solution of the homogeneous Wiener-Hopf integral equation with symbol given by Im + C(λIn − A)−1 B. A more detailed (but straightforward) analysis gives the following result, the formulation of which is in line with Theorem 3.3.

3.5. The Riemann-Hilbert boundary value problem

55

Theorem 3.8. Let Im −K(λ) = Im +C(λIn −A)−1 B be a realization for the symbol of the homogeneous Wiener-Hopf equation ∞ k(t − s)φ(s)ds = 0, t ≥ 0, (3.23) φ(t) − 0

×

and let A = A − BC. Assume that both A and A× have no real eigenvalues, in other words, det Im − K(λ) = 0, −∞ < λ < +∞. Let M be the spectral subspace of A corresponding to the eigenvalues of A in the upper half plane, and let M × be the spectral subspace of A× corresponding to the eigenvalues of A× in the lower half plane. Then φ is a solution of (3.23) if and only if there exists x ∈ M ∩ M × such that ×

φ(t) = Ce−itA x,

t ≥ 0.

(3.24)

Moreover, the vector x in (3.24) is uniquely determined by φ. Formula (3.24) has to be understood in the sense of equality in the solution m space Lm 1 [0, ∞) (or, more generally, Lp [0, ∞) with 1 ≤ p < ∞; cf., Section 1.1 and Theorem 3.3). As a direct consequence of Theorem 3.8, one sees that the dimension of the null space of the Wiener-Hopf integral operator T deﬁned by the left-hand side of (3.23) is equal to dim(M ∩ M × ). It can also be proved that the codimension of its range is equal to codim (M + M × ). In fact, under the conditions of Theorem 3.8, the operator T is a Fredholm operator (see Section XI.1 in [51] for the deﬁnition of this notion), and its Fredholm index, which is deﬁned as the diﬀerence of the codimension of its range and the dimension of its null space, is equal to ind T

=

codim (M + M × ) − dim(M ∩ M × )

=

dim

Cn − dim(M ∩ M × ) M + M×

=

dim

M + M× Cn − dim − dim(M ∩ M × ) M× M×

= = =

M Cn − dim − dim(M ∩ M × ) × M M ∩ M× Cn dim × − dim M M dim

rank P × − rank P.

Here P and P × are the spectral projections corresponding to the eigenvalues in the upper half plane of A and A× , respectively. (In the step from the third to the fourth equality in the above calculation we used Lemma 2 in [89].) More detailed

56

Chapter 3. Explicit solutions using realizations

information about the null space and range of the Wiener-Hopf integral operator T can be obtained in this way (see, e.g., Theorem XIII.8.1 in [51]). We shall return to this theme, in a more general context, in Chapter 7, where it will be shown that the factorization indices in a non-canonical Wiener-Hopf factorization can be expressed in terms of the spaces M and M × , and related subspaces deﬁned in terms of these spaces and the matrices appearing in the realization of the symbol.

Notes The ﬁrst section of this chapter originates from Section 1.2 in [11]. The basic facts about Cauchy domains (see also the ﬁnal paragraphs of Chapter 0), Riesz projections and spectral subspaces, used in this ﬁrst section, can be found in Sections I.1 – I.3 of [51]. The material in Sections 3.2, 3.3 and 3.4 goes back to Chapter 4 in [11]. For Section 3.5 we refer to [12]. We shall return to canonical factorization in a more general setting in Chapters 5 and 7; see Theorems 5.14 and 7.1. Other state space methods for solving convolution equations, also based on matrix-valued realizations but not employing factorization, are developed in [12] and [13].

Chapter 4

Factorization of non-proper rational matrix functions In this chapter we treat the problem of factorizing a non-proper rational matrix function. The realization used in the earlier chapters is replaced by W (λ) = I + C(λG − A)−1 B.

(4.1)

Here I = Im is the m × m identity matrix, A and G are square matrices of order n say, and the matrices C and B are of sizes m × n and n × m, respectively. Any rational m × m matrix function W , proper or non-proper, admits such a representation. The representation (4.1) allows us to extend the results obtained in the previous chapter to arbitrary rational matrix functions. As an application we treat the problem to invert a singular integral operator with a rational matrix symbol. This chapter consists of ﬁve sections. In Section 4.1 we review the spectral theory of matrix pencils. Section 4.2 presents the realization theorem for nonproper rational matrix functions referred to in the previous paragraph. The corresponding canonical factorization theorem is given in Section 4.3. The ﬁnal two sections deal with applications to inverting singular integral operators (Section 4.4) and solving Riemann-Hilbert problems (Section 4.5).

4.1 Preliminaries about matrix pencils Let A and G be complex n×n matrices. The linear matrix-valued function λG−A, where λ is a complex variable, is called a (linear matrix ) pencil . We say that the pencil λG − A is regular on Ω or Ω-regular if λG − A is invertible for each λ ∈ Ω. Here Ω is a subset of C. From now on Γ will be a Cauchy contour. Its interior domain is denoted by F+ and its exterior domain by F− . We shall assume that ∞ ∈ F− . Pencils that

58

Chapter 4. Factorization of non-proper rational matrix functions

are Γ-regular admit block matrix partitionings that are comparable to spectral decompositions of a single matrix. This fact is summarized by the following theorem, the proof of which can be found in [140] (see also Section IV.1 of [51]). Theorem 4.1. Let λG − A be a Γ-regular pencil, and let the matrices P and Q be deﬁned by 1 P = 2πi

−1

Γ

G(λG − A)

dλ,

1 Q= 2πi

Γ

(λG − A)−1 Gdλ.

(4.2)

Then P and Q are projections such that (i) P A = AQ and P G = GQ, (ii) (λG − A)−1 P = Q(λG − A)−1 on Γ and this function has an analytic continuation on F− which vanishes at ∞, (iii) (λG − A)−1 (I − P ) = (I − Q)(λG − A)−1 on Γ and this function has an analytic continuation on F+ . The properties (i)–(iii) in the above proposition determine P and Q uniquely, that is, if P and Q are projections such that (i)–(iii) hold, then P and Q are given by the integral formulas in (4.2). For a better understanding of the above result, let us write A and G as block matrices relative to the decompositions of Cm induced by the projections P and Q. Condition (i) in Theorem 4.1 implies that A and G have block diagonal representations: A = G =

A1

0

0

A2

G1

0

0

G2

˙ Ker Q → Im P + ˙ Ker P, : Im Q + ˙ Ker Q → Im P + ˙ Ker P. : Im Q +

Property (ii) is equivalent to saying that the pencil λG1 − A1 is regular on F− and G1 is invertible; property (iii) amounts to regularity of the pencil λG2 − A2 on F+ . In the particular case when G is the identity matrix I, the two projections P and Q coincide, and P is just the spectral (or Riesz) projection of A corresponding to the eigenvalues in F+ . The latter means (see Section 3.1 or Section I.2 in [51]) that P is a projection commuting with A, the eigenvalues of A|Im P are in F+ and the eigenvalues of A|Ker P are in F− . In that case, Im P is the spectral subspace of A corresponding to the eigenvalues of A in F+ , and Ker P is the spectral subspace of A corresponding to the eigenvalues of A in F− .

4.2. Realization of a non-proper rational matrix function

59

4.2 Realization of a non-proper rational matrix function In this section we derive the representation (4.1), and present some useful identities related to (4.1). Theorem 4.2. Let W be a rational m × m matrix function, and let Ω be the subset of C on which W is analytic. Then, given an m × m matrix D, the function W admits a representation W (λ) = D + C(λG − A)−1 B,

λ ∈ Ω,

(4.3)

where λG − A is an Ω-regular m × m matrix pencil, and B and C are matrices of sizes n × m and m × n, respectively. The set Ω is the complement in C of the set of ﬁnite poles of W (i.e., the poles of W in C). In later applications, D will be taken to be Im , the m × m identity matrix. Proof. Let us ﬁrst remark that W admits a decomposition W (λ) = K(λ) + L(λ),

λ ∈ Ω,

(4.4)

where L is an m × m matrix polynomial and K is a proper rational m × m matrix such that the subset of C on which K is analytic coincides with Ω. Such a decomposition is not unique. In fact, given (4.4) we can obtain another decomposition of F with the same properties by adding a constant matrix to K and subtracting the same matrix from L. This, however, is all the freedom one has. In other words the decomposition (4.4) will be unique if we ﬁx the value of K at inﬁnity. From now on we shall assume that K(∞) = D. The results obtained in Section 2.2 then imply that K admits a realization K(λ) = D + CK (λ − AK )−1 BK ,

λ ∈ Ω,

where AK , BK and CK are matrices of appropriate sizes and the resolvent set of the (square) matrix AK coincides with Ω. The latter can be reformulated by saying that the eigenvalues of AK are just the ﬁnite poles of W . Proceeding with the second term in the right-hand side of the identity (4.4), we write L(λ) = L0 + λL1 + · · · + λq Lq , and introduce ⎡ ⎢ ⎢ GL = ⎢ ⎢ ⎣

⎤

0 Im 0

.. ..

. .

Im 0

⎥ ⎥ ⎥, ⎥ ⎦

⎡ ⎢ ⎢ BL = ⎢ ⎢ ⎣

L0 L1 .. . Lq

⎤ ⎥ ⎥ ⎥, ⎥ ⎦

CL =

−Im

0

···

0

,

60

Chapter 4. Factorization of non-proper rational matrix functions

where the blanks in GL indicate zero entries. The matrix GL is square of size l = m(q + 1). Also GL is nilpotent (of order q + 1), and hence Il − λGL is invertible for each λ in C. The ﬁrst row in the block matrix representation of (Il − λGL )−1 is equal to [ Im λIm . . . λq Im ] and it follows that L(λ) = CL (λGL − Il )−1 BL on all of the (ﬁnite) complex plane. By combining the representation results for K and L we see that W can be written in the form (4.3) with A=

AK

0

0

Il

,

B=

BK BL

,

C=

CK

CL

,

G=

I

0

0

GL

.

Here I is the identity matrix of the same size as AK . The fact that GL is nilpotent, implies that the matrix λG − A is invertible if and only if λ is an eigenvalue of AK , that is if and only if λ is a ﬁnite pole of W . The following proposition, which describes some elementary operations on a rational matrix function in terms of a given realization, is the natural analogue of Theorem 2.4 for realizations of the form (4.1). Theorem 4.3. Let W (λ) = I + C(λG− A)−1 B, and put A× = A− BC. Then W (λ) is invertible if and only if λG − A× is invertible, and in that case the following identities hold: W (λ)−1 = I − C(λG − A× )−1 B,

(4.5)

W (λ)C(λG − A× )−1 = C(λG − A)−1 ,

(4.6)

(λG − A× )−1 BW (λ) = (λG − A)−1 B,

(4.7)

(λG − A× )−1 = (λG − A)−1 − (λG − A)−1 BW (λ)−1 C(λG − A)−1 .

(4.8)

Proof. Fix λ ∈ C such that λG − A is invertible. Then det W (λ)

=

det I + C(λG − A)−1 B = det I + (λG − A)−1 BC

=

det (λG − A)−1 det(λG − A + BC)

=

det(λG − A× ) . det(λG − A)

It follows that W (λ) is invertible if and only if λG − A× is invertible. Also, in that case, a straightforward computation yields

4.3. Explicit canonical factorization

61

W (λ)C(λG − A× )−1 − C(λG − A× )−1 = C(λG − A)−1 BC(λG − A× )−1 = C(λG − A)−1 A − A× (λG − A× )−1 = C(λG − A)−1 (λG − A× ) − (λG − A) (λG − A× )−1 = C(λG − A)−1 − C(λG − A× )−1 . Since W (λ) is invertible, this proves (4.6). The identity (4.7) is proved in a similar way. Using (4.6) a straightforward computation shows that W (λ) I − C(λG − A× )−1 B = I, and hence (4.5) holds. Finally, (4.8) follows by applying (4.6) and again using the identity BC = (λG − A× ) − (λG − A). Instead of the above argument one can also use an analogue of the second proof of Theorem 2.1 in [20], which uses Schur complements arguments (cf., the remark made in the ﬁnal paragraph of Section 2.4).

4.3 Explicit canonical factorization In this section we show how the realization (4.1) can be used to construct a canonical factorization of an arbitrary rational matrix function. Necessary and suﬃcient conditions for the existence of such a factorization and formulas for the factors are stated explicitly in terms of the data appearing in the realization. The next theorem, a counterpart of Theorem 3.2 for non-proper rational matrix functions, is the main result. Theorem 4.4. Let W be a rational m × m matrix function without poles on the curve Γ, and let W be given by the Γ-regular realization W (λ) = I + C(λG − A)−1 B,

λ ∈ Γ.

(4.9)

Put A× = A − BC. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (i) the pencil λG − A× is Γ-regular, ˙ Ker P × and Cn = Im Q + ˙ Ker Q× . (ii) Cn = Im P + Here n is the order of the matrices G and A, and 1 1 G(λG − A)−1 dλ, P× = G(λG − A× )−1 dλ, P = 2πi Γ 2πi Γ

62

Chapter 4. Factorization of non-proper rational matrix functions

1 1 −1 × Q= (λG − A) Gdλ, Q = (λG − A× )−1 Gdλ. 2πi Γ 2πi Γ If the conditions (i) and (ii) are satisﬁed, a right canonical factorization with respect to Γ is given by λ ∈ Γ,

W (λ) = W− (λ)W+ (λ),

where the factors and their inverses can be written as W− (λ)

=

I + C(λG − A)−1 (I − Δ)B, −1

(4.10)

W+ (λ)

=

I + CΛ(λG − A)

W−−1 (λ)

=

I − C(I − Λ)(λG − A× )−1 B,

(4.12)

W+−1 (λ)

=

I − C(λG − A× )−1 ΔB.

(4.13)

B,

(4.11)

Here Δ is the projection along Im P onto Ker P × , and Λ is the projection of Cn along Im Q onto Ker Q× . Finally, the ﬁrst equality in (ii) implies the second and conversely. Proof. We split the proof into four parts. The ﬁrst part concerns the condition (i). In the second part we prove that the ﬁrst equality in (ii) implies the second and conversely. In the third part we use (i) and (ii) to derive the canonical factorization and the formulas for its factors. The ﬁnal part concerns the necessity of the condition (ii). Part 1. From the deﬁnition given in Section 3.1 it is clear that a necessary condition in order that W admits a right canonical factorization with respect to Γ is that W takes invertible values on Γ. By Theorem 4.3 this necessary condition is fulﬁlled if and only if (i) holds true. In what follows we shall assume that (i) is satisﬁed. Part 2. In this part we prove the last statement of the theorem. Consider the operators P × |Im P : Im P → Im P × ,

Q× |Im Q : Im Q → Im Q× .

(4.14)

The ﬁrst equality in (ii) is equivalent to the invertibility of the ﬁrst operator in (4.14). To see this, note that Ker P × |Im P = Ker P × ∩ Im P , and thus P × |Im P is injective if and only if Ker P × ∩ Im P = {0}. Next, observe each that for y ∈ Im P we have y = (I − P × )y + P × |ImP y ∈ Ker P × + Im P × |Im P . Thus Ker P × + Im P ⊂ Ker P × + Im P × |Im P . The reverse inclusion is also true. × × × Indeed, for z ∈ Im P we have P z = ×(P z − z) + z ∈ Ker P × + Im P . It follows × × that Ker P + Im P |Im P = Ker P + Im P , and hence P |Im P considered as an operator into Im P × is surjective if and only if Cn = Ker P × + Im P . Thus, as claimed, the ﬁrst identity in (ii) amounts to the same as the invertibility of the ﬁrst operator in (4.14). Similarly, the second equality in (ii) is equivalent to the invertibility of the second operator in (4.14). Notice that GQ = P G,

GQ× = P × G,

(4.15)

4.3. Explicit canonical factorization

63

which is clear from the deﬁnitions of the projections Q, P and Q× , P × . Furthermore, from the material presented in Section 4.1, applied to λG − A as well as to λG − A× , we see that G maps Im Q and Im Q× in a one-one manner onto Im P and Im P × , respectively. Thus the operators E = G|Im Q : Im Q → Im P and E × = G|Im Q× : Im Q× → Im P × are invertible and, in addition, E × (Q× |Im Q ) = (P × |Im P )E. So the operators in (4.14) are equivalent, and hence the ﬁrst operator in (4.14) is invertible if and only if the same is true for the second operator in (4.14). This proves that the ﬁrst equality in (ii) implies the second and vice versa. Part 3. Next assume that (i) and the direct sum decompositions in (ii) hold true. Our aim is to obtain a canonical factorization of W . Write A, G, B, C as well as A× = A − BC in block form relative to the decompositions in (ii): A11 A12 ˙ Ker P × , ˙ Ker Q× → Im P + (4.16) A = : Im Q + 0 A22 G = B

=

C

=

×

A

=

G11

0

0

G22

B1 B2

˙ Ker Q× → Im P + ˙ Ker P × , : Im Q +

(4.17)

˙ Ker P × , : Cn → Im P +

C2

C1

A× 11

0

A× 21

A× 22

˙ Ker Q× → Cn , : Im Q +

(4.18) (4.19)

˙ Ker P × . ˙ Ker Q× → Im P + : Im Q +

(4.20)

From Theorem 4.1, applied to λG − A as well as to λG − A× , we know that AQ = P A,

A× Q× = P × A× .

(4.21)

The ﬁrst identity in (4.21) implies that A maps Im Q into Im P . This explains the zero entry in the left lower corner of the block matrix for A. From (4.15) we conclude that G has the desired block diagonal form. From the second identity in (4.21) it follows that A× maps Ker Q× into Ker P × , which justiﬁes the zero in the right upper corner of the block matrix for A× . Taking into account the identity A× = A − BC gives A12 = B1 C2 , A× 11 = A11 − B1 C1 ,

A× 21 = −B2 C1 , A× 22 = A22 − B2 C2 .

(4.22) (4.23)

64

Chapter 4. Factorization of non-proper rational matrix functions

Deﬁne the matrix functions W− and W+ by (4.10) and (4.11), respectively. Using the block matrix representations of A, G, B, and C we may rewrite W− and W+ in the form W− (λ) = I + C1 (λG1 − A11 )−1 B1 , λ ∈ Γ, (4.24) W+ (λ) = I + C2 (λG2 − A22 )−1 B2 ,

λ ∈ Γ.

(4.25)

From the block matrix representation of A and the ﬁrst identity in (4.22) we see that ⎡ ⎤−1 ⎡ ⎤ −B1 C2 B1 λG1 − A11 ⎦ ⎣ ⎦ W− (λ)W+ (λ) = I + C1 C2 ⎣ 0 λG2 − A22 B2 = I + C(λG − A)−1 B = W (λ), which gives the factorization W = W− W+ . Next, we check the analytic properties of the factors. Obviously, W− and W+ have no poles on Γ. Note that λG1 − A11 = (λG − A)|Im Q : Im Q → Im P. Thus we know from Section 4.1 that (λG1 − A11 )−1 has an analytic extension on F− which vanishes at inﬁnity. So W− is continuous on F− ∪ Γ and analytic on F− (including inﬁnity). To see that a similar statement holds true for W+ on F+ , we ﬁrst note that the linear maps J

=

(I − Q)|Ker Q× : Ker Q× → Ker Q,

H

=

(I − P )|Ker P × : Ker P × → Ker P,

are invertible. In fact, J −1 = Λ|Ker Q and H −1 = Δ|Ker P , where Λ is the projection along Im Q onto Ker Q× , and Δ is the projection along Im P onto Ker P × . Next, take x ∈ Ker Q× . Then (λG2 − A22 )x = Δ(λG − A)x = Δ(λG − A)(I − Q)x = Δ(λG − A)Jx, which shows that H(λG2 − A22 ) = (λG − A)|Ker Q J. But then we can use Theorem 4.1 and the invertibility of the operators H and J to show that the function (λG2 −A2 )−1 has an analytic extension on F+ . Hence W+ is continuous on F+ ∪ Γ and analytic on F+ . From the factorization W (λ) = W− (λ)W+ (λ) for λ ∈ Γ it follows that W− (λ) and W+ (λ) are both invertible for each λ ∈ Γ. So we can apply Theorem 4.3 to show that W−−1 (λ)

=

−1 I − C1 (λG1 − A× B1 , 11 )

(4.26)

W+−1 (λ)

=

−1 I − C2 (λG2 − A× B2 . 22 )

(4.27)

4.3. Explicit canonical factorization

65

Here we use the two identities in (4.23). Using the block matrix representations of A, G, B and C given above, it is clear that (4.26) and (4.27) yield the formulas (4.12) and (4.13), respectively. We proceed by checking the analyticity properties of the functions W−−1 and −1 W+ . First note that × × × λG2 − A× 22 = (λG − A )|Ker Q× : Ker Q → Ker P .

Thus by applying Theorem 4.1 with λG − A× in place of λG − A we see that −1 the function (λG2 − A× has an analytic extension on F+ . It follows that the 22 ) −1 function W+ is continuous on F+ ∪ Γ and analytic on F+ . To prove the analogous result for W−−1 with respect to F− we use that × × H × (λG1 − A× 11 ) = (λG − A )|Im Q× J , where J × = Q× |Im Q : Im Q → Im Q× and H × = P × |Im P : Im P → Im P × are invertible linear maps of which the inverses are given by (J × )−1 = (I − Λ)|Im Q× ,

(H × )−1 = (I − Δ)|Im P × .

−1 Since (λG − A× )|Im Q× is analytic on F− by virtue of Theorem 4.1 applied −1 . Hence the to λG − A× , we conclude that the same holds true for (λG1 − A× 11 ) function W− (λ)−1 is continuous on F− ∪ Γ and analytic on F− . Thus we have proved that W = W− W+ is a right canonical factorization with respect to the curve Γ. Part 4. In this part we prove the necessity of the equalities in (ii). So in what follows we assume that W = W− W+ is a canonical factorization of W with respect to Γ. Take x ∈ Im P ∩ Ker P × and, for λ ∈ Γ, put ϕ− (λ) = C(λG − A)−1 x,

ϕ+ (λ) = C(λG − A× )−1 x.

Since x ∈ Im P , the ﬁrst identity in (4.21) allows us to rewrite ϕ− as −1 ϕ− (λ) = (C|Im Q ) (λG − A)|Im Q x, and hence Theorem 4.1(ii) implies that ϕ− has an analytic continuation on F− which vanishes at inﬁnity. Similarly, since −1 x, ϕ+ (λ) = (C|Ker Q× ) (λG − A× )−1 |Ker Q× we conclude from Theorem 4.1(iii) applied to λG − A× that ϕ+ has an analytic continuation on F+ . Note that W (λ)−1 ϕ− (λ) = ϕ+ (λ) for each λ ∈ Γ, because of formula (4.6) in Theorem 4.3. It follows that W− (λ)−1 ϕ− (λ) = W+ (λ)ϕ+ (λ),

λ ∈ Γ.

66

Chapter 4. Factorization of non-proper rational matrix functions

Now use the analyticity properties of the factors W− and W+ . We conclude that W−−1 ϕ− has an analytic continuation on F− which vanishes at inﬁnity, and W+ ϕ+ has an analytic continuation on F+ . Liouville’s theorem implies that both functions are identically zero. It follows that ϕ− (λ) = 0 for each λ ∈ Γ. But then we can apply formula (4.8) to show that (λG − A× )−1 x = (λG − A)−1 x,

λ ∈ Γ.

Now, repeat part of the above reasoning. Note that (λG − A)−1 x has an analytic continuation on F− which vanishes at inﬁnity, and (λG − A× )−1 x has an analytic continuation on F+ . Again using Liouville’s theorem we conclude that both matrix functions (λG − A)−1 x and (λG − A× )−1 x are identically zero on Γ. This yields x = 0. We proved that Im P ∩ Ker P × = {0}. Recall that G maps Im Q in a one-one manner onto Im P . Thus (4.15) shows that G maps Im Q ∩ Ker Q× in a one-one manner into Im P ∩ Ker P × . Hence Im Q ∩ Ker Q× = {0} too. Next we show that Im Q + Ker Q× = Cn . Take y ∈ Cn such that y is orthogonal to Im Q + Ker Q× . Let y ∗ be the row vector of which the j-th entry is equal to the complex conjugate of the j-th entry of y (j = 1, . . . , m). For λ ∈ Γ, put ψ− (λ) = y ∗ (λG − A× )−1 B,

ψ+ (λ) = y ∗ (λG − A)−1 B.

Since y ∗ (I − Q)× = 0, Theorem 4.1 shows that ψ− (λ) = y ∗ (λG − A× )−1 P × B, and thus ψ− has an analytic continuation on F− which vanishes at inﬁnity. Similarly, y ∗ Q = 0 implies that ψ+ has an analytic continuation on F+ . Now, use the canonical factorization W = W− W+ and (4.7) to show that ψ+ (λ)W+ (λ)−1 = ψ− (λ)W− (λ),

λ ∈ Γ.

But then, as before, we can use Liouville’s theorem to show that both sides of the identity are equal to zero. It follows that ψ+ (λ) = 0 for each λ ∈ Γ, and we can use formula (4.8) to show that y ∗ (λG − A× )−1 = y ∗ (λG − A)−1 ,

λ ∈ Γ.

Recall that y ∗ Q and y ∗ (I − Q× ) are both zero. Thus Theorem 4.1 implies that y ∗ (λG − A× )−1 has an analytic continuation on F− which vanishes at inﬁnity, and the function y ∗ (λG − A)−1 has an analytic continuation on F+ . So, by Liouville’s theorem, y ∗ (λG−A)−1 = 0 on Γ, and thus y = 0. This gives Im Q+ Ker Q× = Cn . Combining this with with what we saw in the preceding paragraph, we obtain ˙ Ker Q× = Cn . But then the result of Part 2 yields the direct sum decomIm Q + position Im P˙+ Ker P × = Cn , and (ii) is proved. The fact that in Theorem 4.4 the curve Γ is bounded is not essential. We only use that Γ is a closed curve on the Riemann sphere C∞ and that W has no poles on Γ. Thus Γ may pass through inﬁnity. For instance, let us replace Γ by the

4.3. Explicit canonical factorization

67

extended real line R∞ which passes through inﬁnity. By the results of Section 2.2, the condition that the m × m rational matrix function W has no poles on R ∪ {∞} implies that W can be represented in the form W (λ) = D + C(λ − A)−1 B,

λ ∈ R,

(4.28)

where A is a square matrix with no real eigenvalues. The condition that W takes invertible values on R ∪ {∞} now amounts to the requirement that D is invertible and the matrix A − BD−1 C has no real eigenvalues. Also, in that case, W −1 (λ) = D −1 − D−1 C(λ − A× )−1 BD −1 ,

λ ∈ R,

where A× = A−BD−1 C. With these minor modiﬁcations the proof of Theorem 4.4 also applies to realizations of the form (4.28), and yields the following theorem. Theorem 4.5. Let W be a rational m × m matrix function without poles on the real line, and let W be given by the realization W (λ) = D + C(λIn − A)−1 B,

λ ∈ R,

(4.29)

where A is an n×n matrix with no real eigenvalues. Then W admits a right canonical factorization with respect to R ∪ {∞} if and only if the following conditions are satisﬁed: (i) D is invertible and A× = A − BD−1 C has no real eigenvalues, ˙ M ×. (ii) Cn = M + Here n is the order of the matrix A, the space M is the spectral subspace of A corresponding to its eigenvalues in the upper half plane, and M × is the spectral subspace of A× corresponding to its eigenvalues in the lower half plane. Furthermore, if the conditions (i) and (ii) are fulﬁlled, then a right canonical factorization with respect to R ∪ {∞} is given by W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

where the factors and their inverses can be written as W− (λ)

= D + C(λIn − A)−1 (I − Π)B,

W+ (λ)

= I + D−1 CΠ(λIn − A)−1 B,

W−−1 (λ)

= D−1 − D−1 C(I − Π)(λIn − A× )−1 BD −1 ,

W+−1 (λ)

= I − D−1 C(λIn − A× )−1 ΠB.

Here Π is the projection of Cn along M onto M × . Since there is no a priori assumption on the invertibility of (the external) operator D, Theorem 4.5 is a slight extension of Theorem 3.2 dealing with matrix functions too. The results can be generalized to the case of operator functions (cf., Section 7.1 below).

68

Chapter 4. Factorization of non-proper rational matrix functions

4.4 Inversion of singular operators with a rational matrix symbol In this section we apply the results of the previous sections to solve the problem of inverting the singular integral equation 1 ϕ(τ ) dτ = g(t), t ∈ Γ. (4.30) a(t)ϕ(t) + b(t) πi Γ τ − t Throughout we assume that a and b are rational m × m matrix functions which do not have poles on the Cauchy contour Γ. We shall analyze equation (4.30) under the additional condition that the diﬀerence a(λ) − b(λ) is invertible for each λ ∈ Γ. Since we are interested in invertibility, the latter condition is not an essential restriction (cf., Theorem 1.3). The fact that the matrix a(λ) − b(λ) is invertible for λ ∈ Γ allows us to introduce the operator T = MW PΓ + QΓ which we consider on Lm 2 (Γ). Here −1 a(λ) + b(λ) , W (λ) = a(λ) − b(λ) m and MW is the operator of multiplication by W on Lm 2 (Γ), that is, for ϕ ∈ L2 (Γ) we have (MW ϕ)(t) = W (t)ϕ(t) for almost all t ∈ Γ. Furthermore, P Γ and QΓ are the orthogonal projections on Lm 2 (Γ) associated with the singular integral operator introduced in Section 1.3. Thus, for ϕ ∈ Lm 2 (Γ), 1 ϕ(τ ) 1 (P Γ ϕ)(t) = ϕ(t) + dτ, (4.31) 2 2πi Γ τ − t

(QΓ ϕ)(t) =

1 1 ϕ(t) − 2 2πi

Γ

ϕ(τ ) dτ, τ −t

(4.32)

for almost all t ∈ Γ. The image of P Γ consists of all functions in Lm 2 (Γ) that admit an analytic continuation into F+ . Similarly, the image of QΓ is the subspace of all functions in Lm 2 (Γ) that admit an analytic continuation into F− and vanish at inﬁnity. Note that equation (4.30) is equivalent to −1 g(λ). (MW P Γ + QΓ )ϕ = g, where g(λ) = a(λ) − b(λ) Since W is a rational m × m matrix function without poles on Γ, we know from Theorem 4.2 that W admits a Γ-regular realization W (λ) = I + C(λG − A)−1 B,

λ ∈ Γ.

(4.33)

The main result of this section provides an explicit inversion formula for the operator MW P Γ + QΓ in terms of the realization (4.33). Theorem 4.6. Let the rational m × m matrix function W be given by the Γ-regular realization (4.33), and put A× = A − BC. Then MW P Γ + QΓ is an invertible operator on Lm 2 (Γ) if and only if the following two conditions are satisﬁed:

4.4. Inversion of singular operators with a rational matrix symbol

69

(1) the pencil λG − A× is Γ-regular, ˙ Ker P × , (2) Cn = Im P + where n is the order of the matrices A and G, and 1 P = 2πi

Γ

G(λG − A)

−1

dλ,

P

×

1 = 2πi

Γ

G(λG − A× )−1 dλ.

(4.34)

In that case

(MW P Γ + QΓ )−1 g (λ)

=

g(λ) − C(λG − A× )−1 B(P Γ g)(λ) + C(λG − A× )−1 − C(λG − A× )−1 (I − Π) 1 P × G(ζG − A× )−1 Bg(ζ)dζ , λ ∈ Γ. · 2πi Γ

Here Π is the projection of Cn along Im P onto Ker P × . With suitable changes, the theorem remains true when P and P × are replaced by the projections Q and Q× (also) appearing in Theorem 4.4. Proof. From the general theory of singular integral equations reviewed in Section 1.3 we know that the operator MW P Γ + QΓ is invertible if and only if W admits a right canonical factorization with respect to Γ. Since W is given by (4.33), the latter is the case if and only if conditions (i) and (ii) in Theorem 4.4 are fulﬁlled. By the ﬁnal statement in Theorem 4.4, conditions (i) and (ii) in Theorem 4.4 are equivalent to conditions (1) and (2) in the present theorem. Thus we have proved that MW P Γ + QΓ is invertible if and only if (1) and (2) are satisﬁed. To get the formula for the inverse of MW P Γ + QΓ we again use the general theory of singular integral equations, the inversion formula (1.29) in particular. Let W = W− W+ be a right canonical factorization of W with respect to Γ. For g ∈ Lm 2 (Γ) we then have, suppressing the variable λ, (MW P Γ + QΓ )−1 g = W+−1 P Γ (W−−1 g) + W− QΓ (W−−1 g) . Taking into account the form of P Γ and QΓ in (4.31) and (4.32), this identity can be rewritten as 1 1 (MW P Γ + QΓ )−1 g (λ) = g(λ) + W (λ)−1 g(λ) 2 2 1 1 −1 W+ (λ) − W− (λ) W− (τ )−1 g(τ ) dτ, + 2πi Γ τ − λ

λ ∈ Γ. (4.35)

70

Chapter 4. Factorization of non-proper rational matrix functions

Next, we use the formulas for W+ , W− and their inverses given in Theorem 4.4. This yields W+ (λ)−1 − W− (λ) W− (τ )−1 = −C(λG − A× )−1 ΔB − C(λG − A)−1 (I − Δ)B × −1

+ λG − A )

× −1

ΔBC(I − Λ)(τ G − A )

−1

+ C(λG − A)

(4.36)

B

(I − Δ)BC(I − Λ)(τ G − A× )−1 B.

Here Δ and Λ are the projections deﬁned in Theorem 4.4. Using these deﬁnitions, and the partitionings of A, G, and A× in (4.16), (4.17) and (4.20), respectively, we obtain ΔA(I − Λ) = 0,

(I − Δ)A× Λ = 0,

ΔG = GΛ.

Since BC = A − A× , it follows that ΔBC(I − Λ) = A× Λ − ΔA× = (A× − λG)Λ − Δ(A× − τ G) − (τ − λ)ΔG, and (I − Δ)BC(I − Λ) = A(I − Λ) − (I − Δ)A× = (A − λG)(I − Λ) − (I − Δ)(A× − τ G) − (τ − λ)(I − Δ)G. Inserting these expressions into (4.36) gives W+ (λ)−1 − W− (λ) W− (τ )−1 = −C(τ G − A× )−1 B −(τ − λ)C(λG − A× )−1 ΔG(τ G − A× )−1 B −(τ − λ)C(λG − A)−1 (I − Δ)G(τ G − A× )−1 B. Next we use that (τ − λ)C(λG − A× )−1 G(τ G − A× )−1 B can be written as C(λG − A× )−1 (τ G − A× ) − (λG − A× ) (τ G − A× )−1 B which in turn is equal to C(λG − A× )−1 B − C(τ G − A× )−1 B, and this leads to W+ (λ)−1 − W− (λ) W− (τ )−1 = −C(λG − A× )−1 B +(τ − λ) C(λG − A× )−1 − C(λG − A)−1 ·(I − Δ)G(τ G − A× )−1 B.

(4.37)

4.5. The Riemann-Hilbert boundary value problem revisited (1)

71

Using (4.37) and (4.5) in (4.35) we obtain 1 (MW P Γ + QΓ )−1 g (λ) = g(λ) − C(λG − A× )−1 Bg(λ) 2 1 1 × −1 −C(λG − A ) B g(τ ) dτ 2πi Γ τ − λ + C(λG − A× )−1 − C(λG − A)−1 (I − Δ) 1 × −1 · G(τ G − A ) Bg(τ ) dτ , 2πi Γ

λ ∈ Γ.

Finally, note that Δ = Π and (I − Π)P × = I − Π. Since P Γ is given by (4.31), we see that we have derived the desired expression for the inverse of the operator M W P Γ + QΓ .

4.5 The Riemann-Hilbert boundary value problem revisited (1) In this section we treat the (homogeneous) Riemann-Hilbert boundary value problem for non-proper rational matrix functions. As before Γ is a Cauchy contour. As usual, the interior domain of Γ is denoted by F+ , and its exterior domain, which contains the point inﬁnity, by F− . Throughout W is a rational m × m matrix function which does not have poles on Γ. We say that a pair of Cm -valued functions Φ+ , Φ− is a solution of the Riemann-Hilbert boundary problem of W with respect to Γ if Φ+ and Φ− are continuous on F+ ∪ Γ and F− ∪ Γ, respectively, Φ+ and Φ− are analytic in F+ and F− , respectively, Φ− vanishes at inﬁnity, and W (λ)Φ+ (λ) = Φ− (λ),

λ ∈ Γ.

(4.38)

Since W is assumed to be a rational m × m matrix function which has no poles on Γ, we may assume that W is given by a Γ-regular realization W (λ) = I + C(λG − A)−1 B,

λ ∈ Γ.

(4.39)

We shall also assume that W takes invertible values on Γ. This additional condition is equivalent to the requirement that the pencil λG−A× is Γ-regular. The following theorem is the natural analogue of Theorem 3.7. Theorem 4.7. Let W be given by (4.39), and assume that the pencil λG − A× is a Γ-regular. Put 1 1 G(λG − A)−1 dλ, P× = G(λG − A× )−1 dλ. P = 2πi Γ 2πi Γ

72

Chapter 4. Factorization of non-proper rational matrix functions

Then the pair of functions Φ+ and Φ− is a solution of the Riemann-Hilbert boundary value problem of W with respect to Γ if and only if there exists x belonging to Im P ∩ Ker P × such that Φ+ (λ) = C(λG − A× )−1 x,

Φ− (λ) = C(λG − A)−1 x.

(4.40)

Moreover the vector x in (4.40) is uniquely determined by Φ+ , Φ− With the appropriate modiﬁcations, the theorem remains true when P and P × are replaced by the projections Q and Q× (also) appearing in Theorem 4.4. Proof. Take x ∈ Im P ∩ Ker P × , and deﬁne Φ+ and Φ− by (4.40). Formula (4.6) implies that (4.38) is satisﬁed. Since x = P x, Theorem 4.1 (ii) shows that Φ− is continuous on F− ∪ Γ, analytic in F− , and vanishes at inﬁnity. Similarly, using x = (I −P × )x, Theorem 4.1 (iii), applied to λG−A× , yields that Φ+ is continuous on F+ ∪ Γ and analytic on F+ . Thus the functions Φ+ and Φ− have the desired properties, and the pair Φ+ , Φ− is a solution. To prove the converse, assume that the pair Φ+ , Φ− is a solution of the Riemann-Hilbert problem for W with respect to Γ. For λ ∈ Γ, introduce ρ(λ) = (λG−A)−1 BΦ+ (λ). The n×m matrix function ρ is continuous on Γ, thus it makes sense to put 1 ρ(τ ) 1 ρ+ (λ) = ρ(λ) + dτ, λ ∈ Γ, 2 2πi Γ τ − λ 1 1 ρ(τ ) ρ(λ) − dτ, λ ∈ Γ; ρ− (λ) = 2 2πi Γ τ − λ cf., the expressions (4.31) and (4.32). The function ρ+ is continuous on F+ ∪ Γ and analytic in F+ , and ρ− has the same properties with F− in place of F+ . Moreover, ρ− vanishes at inﬁnity. We ﬁrst show that Φ+ (λ)

= −Cρ+ (λ),

λ ∈ F+ ∪ Γ,

(4.41)

Φ− (λ)

=

λ ∈ F− ∪ Γ.

(4.42)

Cρ− (λ),

Since the pair Φ+ , Φ− satisﬁes (4.38), we have Φ− (λ) = Φ+ (λ) + C(λG − A)−1 BΦ+ (λ) = Φ+ (λ) + Cρ(λ),

λ ∈ Γ.

But ρ(λ) = ρ− (λ) + ρ+ (λ) on Γ, and therefore Φ− (λ) − Cρ− (λ) = Φ+ (λ) + Cρ+ (λ),

λ ∈ Γ.

(4.43)

The right-hand side of (4.43) is continuous on F+ ∪ Γ and analytic in F+ . On the other hand, the left-hand side of (4.43) is continuous on F− ∪ Γ, analytic in

4.5. The Riemann-Hilbert boundary value problem revisited (1)

73

F− and vanishes at inﬁnity. Thus, by Liouville’s theorem, both sides of (4.43) are identically zero on Γ, and the identities (4.41) and (4.42) hold. Next, we compute the function ρ− . From the deﬁnition of ρ(λ) we see that (λG − A)ρ(λ) = BΦ+ (λ) for λ ∈ Γ. Since Φ+ is continuous on F+ ∪ Γ and analytic in F+ , we conclude that for each λ ∈ Γ, 1 1 1 (λG − A)ρ(λ) = (τ G − A)ρ(τ ) dτ 2 2πi Γ τ − λ 1 1 (λG − A) + (τ − λ)G ρ(τ ) dτ = 2πi Γ τ − λ 1 ρ(τ ) = (λG − A) dτ + x, 2πi Γ τ − λ 1 Gρ(τ ) dτ. 2πi Γ Using the deﬁnition of ρ− , the above calculation shows that where

x=

ρ− (λ) = (λG − A)−1 x,

λ ∈ Γ.

(4.44)

To compute ρ+ , recall that ρ(λ) = ρ− (λ) + ρ+ (λ) on Γ. This, together with (4.41) and (4.42), yields (λG − A)ρ+ (λ)

= (λG − A)ρ(λ) − (λG − A)ρ− (λ) = BΦ+ (λ) − x = −BCρ+ (λ) − x,

λ ∈ Γ.

×

Since A = A − BC, we obtain ρ+ (λ) = −(λG − A× )−1 x,

λ ∈ Γ.

(4.45)

From (4.44) and the fact that ρ− is continuous on F− ∪ Γ, analytic in F− , and vanishes at inﬁnity, we conclude that x = P x. Similarly, we obtain from (4.45) that x = (I − P × )x. Thus x ∈ Im P ∩ Ker P × . Formulas (4.41), (4.42), (4.44) and (4.45) now show that the functions Φ+ and Φ− have the desired representation (4.40). It remains to prove the uniqueness of the vector x in (4.40). To do this assume that u ∈ Im P ∩ Ker P × , and let C(λG − A)−1 u be identically zero on Γ. It suﬃces to show that u = 0. For this purpose we use the identity (4.8). Applying this identity to the vector u, we see that (λG − A× )−1 u = (λG − A)−1 u, ×

λ ∈ Γ.

(4.46)

Since u ∈ Ker P , the left-hand side of (4.46) has an analytic continuation on F+ ; see Theorem 4.1 (iii). Similarly, u ∈ Im P implies that the right side of (4.46) has an analytic continuation on F− which vanishes at inﬁnity; see Theorem 4.1 (ii). But then we can apply Liouville’s theorem to show that these functions are identically zero on Γ, which yields u = 0.

74

Chapter 4. Factorization of non-proper rational matrix functions

Notes The extension of the Riesz spectral theory for operators to operator pencils, which is described in Section 4.1, is due to Stummel [140]; the results can also be found in Section IV.1 of [51]. Section 4.2 combines the classical realization theory for proper rational matrix functions with that of matrix polynomials; for the latter, see [65]. The main source for the material in Sections 4.2 and 4.3 is the paper [55]; Section 4.4 is based on [56]. Section 4.5 seems to be new. For realizations of the form considered in this chapter, non-canonical Wiener-Hopf factorization has been studied in [151]. Instead of (4.3) other realizations of W can be used; see for instance [79], where (4.3) is replaced by the realization W (λ) = D + (λ − α)C(λG − A)−1 B which can also be used for non-square matrix functions.

Part III Equations with non-rational symbols In this part we carry out a program analogous to that of the second part, but now for certain classes of non-rational matrix and operator functions. Included are matrix functions analytic in a strip but not at inﬁnity, an operator function appearing in linear transport theory, and operator functions analytic on a given curve. There are three chapters. The main topic of the ﬁrst chapter (Chapter 5) is a canonical factorization theorem for matrix functions analytic in a strip but not necessarily at inﬁnity. Its applications to diﬀerent classes of Wiener-Hopf equations are included too. The realizations of such matrix functions require that we consider systems with an inﬁnite dimensional state space and with a state operator that is unbounded and exponentially dichotomous. Thus the theory of strongly continuous semigroups plays an important role in this material. Chapter 6 is entirely dedicated to the solution of an integro-diﬀerential equation from mathematical physics describing stationary migration of particles in a medium. To illustrate the approach, the special case of a ﬁnite number of scattering directions is considered ﬁrst. This restriction makes it possible to reduce the problem to a canonical factorization problem for rational matrix functions. The general situation features an inﬁnite dimensional separable Hilbert space as state space. The ﬁnal chapter (Chapter 7) deals with canonical factorization and non-canonical Wiener-Hopf factorization for operator-valued functions that are analytic on a given curve. In this chapter the so-called factorization indices are described in state space terms.

Chapter 5

Factorization of matrix functions analytic in a strip This chapter deals with m × m matrix-valued functions of the form ∞ W (λ) = I − eiλt k(t) dt,

(5.1)

−∞

where k is an m× m matrix-valued function with the property that for some ω < 0 the entries of e−ω|t| k(t) are Lebesgue integrable on the real line. In other words, k is of the form k(t) = eω|t| h(t) with h ∈ Lm×m (R). (5.2) 1 It follows that the function W is analytic in the strip |λ| < τ , where τ = −ω. This strip contains the real line. The aim is to extend the canonical factorization theorem of Chapter 5 to functions of the type (5.1). In general, the function W in (5.1) is not a rational matrix function, and hence one cannot expect a representation of W in the form W (λ) = I + C(λ − A)−1 B

(5.3)

with A, B, C matrices. Also a realization with A, B and C bounded linear operators will not work. Indeed, in that case the function W would be analytic at inﬁnity, however in general it is not. Thus to get a representation of the type (5.3) one has to allow for unbounded linear operators. In fact, we shall have to allow for A and C to be unbounded while B can be taken to be bounded. This chapter consists of nine sections. In Sections 5.1 and 5.2 we present preliminary material on exponentially dichotomous operators and associated bisemigroups. These exponentially dichotomous operators appear as state operators in the realization triples deﬁned in Section 5.3. In Section 5.4 we construct realization triples for m × m matrix-valued functions W of the form (5.1) with k as in

78

Chapter 5. Factorization of matrix functions analytic in a strip

(5.2), and in Section 5.5 we use the realization triples to invert such a matrix function W . It turns out that inversion is only possible when the associate operator A× = A − BC is exponentially dichotomous too. The inversion formula of Section 5.5 is used in Section 5.6 to derive an explicit formula for the kernel function of the inverse of a full line convolution integral operator when the symbol W is given by (5.1) and (5.2). This section also contains some preliminary material about Hankel operators. The ﬁnal three sections concern applications. Sections 5.7 and 5.8 deal with inversion of a Wiener-Hopf integral equation with a kernel function k of the form (5.2) and with canonical factorization of the corresponding symbol. In Section 5.9 we revisit the Riemann-Hilbert boundary value problem.

5.1 Exponentially dichotomous operators and bisemigroups We begin with some preliminaries about strongly continuous semigroups of operators (also called C0 -semigroups). Free use will be made of the standard theory of these semigroups as explained, for instance, in Chapter XIX of [51]. Besides ordinary C0 -semigroups deﬁned on the positive half line [0, ∞), henceforth to be called right semigroups, we shall also consider semigroups deﬁned on the negative half line (−∞, 0]. The latter will be called left semigroups. Notice that T (t) is a left semigroup if and only if T (−t) is a right semigroup. Let T (t) be a strongly continuous right or left semigroup. As is well-known, there exist constants M and ω such that T (t) ≤ M eω|t| ,

t ∈ J.

Here J is the half line [0, ∞) or (−∞, 0] according to T (t) being a right or a left semigroup. If the above inequality is satisﬁed for a given real number ω and some positive constant M , we say that T (t) is of exponential type ω. Semigroups of negative exponential type will be called exponentially decaying. Next we introduce the concept of an exponentially dichotomous operator. Let X be a complex Banach space, and let A be a (possibly unbounded) linear operator with domain D(A) in X and with values in X, in short A(X → X). Further, let P : X → X be a (bounded linear) projection of X commuting with A. The latter means that P maps D(A) into itself and P Ax = AP x for each x ∈ D(A). Put X− = Im P and X+ = Ker P . Then ˙ X+ , X = X− +

(5.4)

and this decomposition reduces A, that is, ˙ [D(A) ∩ X+ ], D(A) = [D(A) ∩ X− ] +

(5.5)

5.1. Exponentially dichotomous operators and bisemigroups

79

with A mapping [D(A) ∩ X− ] into X− and [D(A) ∩ X+ ] into X+ . So with respect to the decompositions (5.4) and (5.5), the operator A has the matrix representation 0 A− . (5.6) A= 0 A+ Here A− (X− → X− ) is the restriction of A to X− , and A+ (X+ → X+ ) is the restriction of A to X+ . In particular, the domain D(A− ) of A− is D(A) ∩ X− and the domain D(A+ ) of A− + is D(A) ∩ X+ . Thus (5.5) can be rewritten as ˙ D(A+ ). D(A) = D(A− ) + The operator A is said to be exponentially dichotomous if the operators A− and A+ in (5.6) are generators of exponentially decaying strongly continuous left and right semigroups, respectively. In that case the projection P , which will turn out to be unique (see Proposition 5.1 below), is called the separating projection for A. We say that A is of exponential type ω (< 0) if this is true for the semigroups generated by A− and A+ . Suppose, for the moment, that A : X → X is a bounded linear operator. Then A is exponentially dichotomous if and only if the spectrum σ(A) of A does not meet the imaginary axis. In that situation the separating projection for A is simply the Riesz projection corresponding to the part of σ(A) lying in the open right half plane λ > 0. Next, observe that generators of exponentially decaying strongly continuous semigroups belong to the class of exponentially dichotomous operators, the left semigroup case corresponding to the separating projection being the identity operator and the right semigroup case corresponding to the separating projection being the zero operator on X. Returning to the general case, we note that the operators A− and A+ in the deﬁnition of an exponentially dichotomous operator are closed and densely deﬁned. Hence the same is true for their direct sum A. Furthermore, if A is of (negative) exponential type ω, then, by the Hille-Yosida-Phillips theorem (see, e.g., Theorem XIX.2.3 in [51]), the spectrum σ(A− ) of A− is contained in the closed half plane λ ≥ −ω, whereas σ(A+ ) is a subset of λ ≤ ω. In particular, the strip |λ| < −ω is contained in ρ(A), the resolvent set of A. This justiﬁes the use of the term “separating projection” for P . It is convenient to adopt the following notation and terminology. Suppose A(X → X) is an exponentially dichotomous operator with separating projection P , and let A− and A+ be as above. Thus A− and A+ are the restrictions of A to X− = Im P and X+ = Ker P , respectively. With A we associate a function E(·; A) with domain R \ {0} and with values in L(X), the space of all bounded operators on X. The deﬁnition is as follows: for x ∈ X, ⎧ ⎨ −etA− P x, t < 0, E(t; A)x = (5.7) ⎩ etA+ (I − P )x, t > 0,

80

Chapter 5. Factorization of matrix functions analytic in a strip

where, following standard conventions, etA− denotes the value at t(< 0) of the semigroup generated by A− and etA+ denotes the value at t(> 0) of the semigroup generated by A+ . We call E(· ; A) the bisemigroup generated by A. The operator A will be referred to as the bigenerator of E(· ; A). For each x ∈ X the function E(t; A)x is continuous on R \ {0}, and lim E(t; A)x = −P x, t↑0

lim E(t; A)x = (I − P )x. t↓0

(5.8)

We conclude that E(· ; A) is an exponentially decaying operator function which is strongly continuous on the real line, except at the origin where it has (at worst) ˙ D(A+ ), the function E(t; A)x is a jump discontinuity. For x ∈ D(A) = D(A− ) + even diﬀerentiable on R \ {0}. In fact, we have d E(t; A)x = AE(t; A)x = E(t; A)Ax, dt

t = 0.

Obviously the derivative of E(· ; A)x is continuous on R \ {0}, exponentially decaying (in both directions) and has (at worst) a jump discontinuity at the origin. From (5.7) it is clear that E(t, A)P = P E(t, A) = E(t; A),

t < 0,

E(t, A)(I − P ) = (I − P )E(t, A) = E(t; A),

t > 0.

Also the following semigroup properties hold: E(t + s, A) =

−E(t; A)E(s; A),

t, s < 0,

E(t + s, A) =

E(t, A)E(s; A),

t, s > 0.

One of the reasons for the diﬀerent signs to appear in the deﬁnition of E(t; A) is that in this way the following identity holds: ∞ −1 e−λt E(t; A)x dt, x ∈ X, |λ| < −ω. (5.9) (λ − A) x = −∞

Here ω is a negative constant such that A is of exponential type ω. The proof of (5.9) is based on standard semigroup theory (see, e.g., Theorem XIX.2.2 in [51]). With the help of (5.8) and (5.9) we now can prove the uniqueness of the separating projection. Proposition 5.1. Let A(X → X) be an exponentially dichotomous operator. Then A has precisely one separating projection. Proof. Let P be a separating projection for A, and let E(· ; A) be the associate bisemigroup. A priori E(· ; A) depends not only on A but also on P . However, (5.9) and the fact that E(· ; A) is strongly continuous on R \ {0} imply that E(· ; A) is uniquely determined by A. On the other hand the ﬁrst identity in (5.8) shows that P is uniquely determined by E(· ; A). So along with E(· ; A) the separating projection is uniquely determined by A.

5.2. Spectral splitting and proof of Theorem 5.2

81

From (5.9) it follows that on a strip around the imaginary axis, the resolvent (λ − A)−1 of A is the pointwise two-sided Laplace transform of an exponentially decaying operator function which is strongly continuous on R \ {0} and has (at worst) a jump discontinuity at zero. The following theorem shows that this property characterizes exponentially dichotomous operators. Theorem 5.2. Let A(X → X) be a densely deﬁned closed linear operator on the complex Banach space X. Then A is exponentially dichotomous if and only if the imaginary axis is contained in the resolvent set of A and ∞ −1 (λ − A) x = e−λt E(t)x dt, x ∈ X, λ = 0, (5.10) −∞

where E : R \ {0} → L(X) is exponentially decaying and strongly continuous, and E has (at worst) a jump discontinuity at zero. In that case the function E is the bisemigroup generated by A. The above theorem will play an important role in Section 5.5. For the sake of completeness its proof is given in the next section. The reader who is ready to accept Theorem 5.2 may proceed directly to Section 5.3.

5.2 Spectral splitting and proof of Theorem 5.2 In this section we prove Theorem 5.2. The proof will be based on the spectral splitting results proved in Section XV.3 of [51], which originate from [16]. It will be convenient ﬁrst to prove the following result which is the semigroup version of Theorem 5.2. Theorem 5.3. Let S(X → X) be a densely deﬁned closed linear operator on the complex Banach space X. Then S is the inﬁnitesimal generator of a strongly continuous right semigroup of negative exponential type if and only if the imaginary axis is contained in the resolvent set of S and ∞ (λ − S)−1 x = e−λt E(t)x dt, x ∈ X, λ = 0, (5.11) 0

where E : [0, ∞) → L(X) is exponentially decaying and strongly continuous. In that case the function E is the right semigroup generated by S. Proof. The “only if part” of Theorem 5.2 is immediate from standard semigroup theory. To prove the “if part” let ω be a negative real number and L a positive constant such that E(t) ≤ Leωt , t ≥ 0. (5.12) For λ > ω and x ∈ X, put

R(λ)x = 0

∞

e−λt E(t)x dt.

(5.13)

82

Chapter 5. Factorization of matrix functions analytic in a strip

Then R(λ) is a well-deﬁned bounded linear operator on X with norm not exceeding L. The function R is pointwise analytic on λ > ω, and hence it is analytic on λ > ω. We shall prove that λ > ω implies that λ ∈ ρ(S) and R(λ) = (λ − S)−1 . Let T = S −1 be the (bounded) inverse of S. For 0 = λ ∈ ρ(S), one has −1 λ ∈ ρ(T ) and (λ − S)−1 = −λ−1 T (λ−1 − T )−1 . Take λ on the imaginary axis, λ = 0. Combining (5.11) and (5.13) we get R(λ) = (λ − S)−1 = −λ−1 T (λ−1 − T )−1 , and hence R(λ) = λR(λ) − I T . But then the unicity theorem for analytic functions gives that these identities hold on all of λ > ω. A simple computation now shows that R(λ) = (λ − S)−1 for each λ with λ > ω. We have seen that the open half plane λ > ω is contained in ρ(S) and (λ − S)−1 x =

∞

e−λt E(t)x dt,

x ∈ X, λ > ω.

0

(5.14)

Diﬀerentiating the left-and right-hand side of (5.14) for the variable λ, one ﬁnds −n

(λ − S)

(−1)n x= (n − 1)!

∞

tn−1 e−λt E(t)x dt,

0

x ∈ X, λ > ω.

(5.15)

Here n is an arbitrary positive integer. Taking λ > ω and combining (5.12) and (5.15) we get the estimate (λ − S)−n x ≤ Observe that ∞ t 0

n−1 −(λ−ω)t

e

L (n − 1)!

1 dt = (λ − ω)n

∞

0

tn−1 e−(λ−ω)t dt x.

∞

sn−1 e−s ds =

0

(n − 1)! . (λ − ω)n

Thus (λ − S)−n ≤ L(λ − ω)−n for real λ > ω and n = 1, 2, . . . . The HilleYosida-Phillips theorem ([51], page 419) now guarantees that S is the generator of a strongly continuous right semigroup T (t) of exponential type ω < 0. But then (5.11) holds with E(t) replaced by T (t). As the operator-valued functions E) and T are both strongly continuous, they must coincide, and the proof is complete. Proof of Theorem 5.2. We split the proof into three parts. Throughout ω is a negative real number and L is a positive constant such that E(t) ≤ Leω|t| ,

0 = t ∈ R.

(5.16)

Part 1. In this part we show that (λ − A)−1 is well-deﬁned and uniformly bounded

5.2. Spectral splitting and proof of Theorem 5.2

83

on each closed strip |λ| ≤ h where 0 < h < −ω. To do this, let us consider the following expressions: ∞ e−λt E(t)x dt, λ > ω, Ψ+ (λ)x = Ψ− (λ)x

0

0

e−λt E(t)x dt,

= −∞

λ < −ω.

Here x ∈ X. Clearly Ψ+ (λ) is a well-deﬁned bounded linear operator on X which depends analytically on λ on the open half plane λ > ω, and an analogous statement holds of course for Ψ− (λ). Note that Ψ− (λ) + Ψ+ (λ) is analytic on the strip |λ| < −ω and coincides on the imaginary axis with (λ − A)−1 . Thus |λ| < −ω implies λ ∈ ρ(A) and (λ − A)−1 = Ψ− (λ) + Ψ+ (λ), i.e., ∞ e−λt E(t)x dt, x ∈ X, |λ| < −ω. (5.17) (λ − A)−1 x = −∞

A detailed argument can be given along the lines indicated in the second paragraph of the proof of Theorem 5.3. From (5.16) one easily deduces that Ψ+ (λ)

≤

L , λ − ω

λ > ω,

Ψ− (λ)

≤

−L , λ + ω

λ < −ω.

On the strip |λ| < −ω, the norm of (λ − A)−1 = Ψ− (λ) + Ψ+ (λ) can now be estimated as follows: (λ − A)−1 ≤

−2Lω , ω 2 − (λ)2

|λ| < −ω.

(5.18)

In particular (λ − A)−1 is uniformly bounded on each closed strip |λ| ≤ h with 0 < h < −ω. Part 2. Fix 0 < h < −ω. From what has been proved in the previous part, we know that sup (λ − A)−1 < ∞. (5.19) | λ|≤h

This allows us to use the spectral theory developed in Section XV.3 of [51]. First we introduce the operators −α+i∞ 1 Q− = λ−2 (λ − A)−1 dλ, 2πi −α−i∞ Q+ =

−1 2πi

α+i∞

α−i∞

λ−2 (λ − A)−1 dλ.

84

Chapter 5. Factorization of matrix functions analytic in a strip

Here 0 < α < h, and hence (5.19) implies that Q− and Q+ are well-deﬁned bounded linear operators on X. It can be proved that these operators do not depend on the particular choice of α; nevertheless, in what follows we keep α ﬁxed. (Notice that in Section XV.3 of [51] the operators Q− and Q+ are denoted by S− and S+ , respectively.) We deﬁne M− = Im Q− ,

M+ = Im Q+ .

Put T = A−1 . Then T is a bounded linear operator on X commuting with (λ−A)−1 for each λ in the strip |λ| ≤ h. It follows that T commutes with Q− and Q+ . Since T is bounded, this implies that T M− ⊂ M− and T M+ ⊂ M+ . We also know that Im T = D(A), and thus T M− and T M+ belong to D(A). This allows us to deﬁne operators A− (M− → M− ) and A+ (M+ → M+ ) by setting D(A− ) = T M−,

A− x = Ax,

x ∈ D(A− ),

D(A+ ) = T M+,

A+ x = Ax,

x ∈ D(A+ ).

In other words,

−1 A− = T |M− ,

−1 A+ = T |M+ .

The ﬁrst part of Lemma XV.3.3 in [51] shows that A− and A+ are closed and densely deﬁned linear operators, and their spectra satisfy the inclusion relations σ(A− ) ⊂

{λ ∈ C | λ ≤ −h},

σ(A+ ) ⊂

{λ ∈ C | λ ≥ h}.

We shall now prove that −1

(λ − A− )

−1

(λ − A+ )

x

x

∞

=

e−λt E(t)x dt,

x ∈ M− , Reλ > −h,

(5.20)

e−λt E(t)x dt,

x ∈ M+ , Reλ < h.

(5.21)

0 0

= −∞

Following Section XV.3, page 330, of [51], we introduce two auxiliary sets N− and N+ . By deﬁnition N− is the set of all vectors x ∈ X for which there exists an X-valued function ϕ− x , bounded and analytic on λ > −h, which takes its values in D(A) and satisﬁes (λ − A)ϕ− x (λ) = x,

λ > −h.

Roughly speaking, N− consists of all vectors x ∈ X such that (λ − A)−1 x has a bounded analytic continuation to the open half plane λ > −h. The function ϕ− x (assuming it exists) is uniquely determined by x. Analogously, we let N+ be the

5.2. Spectral splitting and proof of Theorem 5.2

85

set of all vectors x ∈ X for which there exists an X-valued function ϕ+ x , bounded and analytic on λ < h, which takes its values in D(A) and satisﬁes (λ − A)ϕ+ x (λ) = x,

λ < h.

Also ϕ+ x is unique, provided it exists. Obviously, the sets N− and N+ are (possibly non-closed) linear manifolds of X. The second part of Lemma XV.3.3 in [51] states that D(A2− ) ⊂ N− and 2 D(A+ ) ⊂ N+ . Now, ﬁx x ∈ D(A2− ). Then x ∈ N− , and hence (λ − A)−1 x extends to a bounded analytic function on λ > −h. Notice that Ψ+ (λ) is also bounded and analytic on λ > −h. Recall that Ψ− (λ) is equal to (λ − A)−1 − Ψ+ (λ) for each λ in the strip |λ| < ω. It follows that Ψ− (λ)x extends to a bounded analytic function on λ > −h. On the other hand Ψ− (λ)x is analytic on λ < −ω and bounded on λ ≤ h. Hence Ψ− (λ)x determines a bounded entire function. From the estimate given for Ψ− (λ) in the previous part, it is clear that lim

λ∈R, λ→−∞

Ψ− (λ)x = 0.

But then we can use Liouville’s theorem to show that Ψ− (λ)x vanishes identically. We conclude that 0 e−λt E(t)x dt = 0, λ < −ω. −∞

Since E(t)x is continuous on −∞ < t < 0, it follows that E(t)x = 0 for all negative real numbers t. −1 is densely Now recall that T |M− is one-to-one and that A− = T |M− deﬁned. This implies that 2 D(A2− ) = Im T |M− , and that D(A2− ) is dense in M− . Thus the result of the previous paragraph shows that E(t) vanishes on M− for −∞ < t < 0. For x ∈ M− and |λ| < h, we have −1 (T |M− )x = (λ − A− )−1 x. (λ − A)−1 x = −(I − λT )−1 T x = − I − λ(T |M− ) Hence, for x ∈ M− and |λ| < h, −1

(λ − A− )

−1

x = (λ − A)

∞

x=

e −∞

−λt

∞

E(t)x dt =

e−λt E(t)x dt.

0

By analytic continuation this proves (5.20). Formula (5.21) is proved in a similar manner. Part 3. In this part we complete the proof. First we show that for t > 0 the operator E(t) maps M− into M− . To see this, take x ∈ M− , and let f be a

86

Chapter 5. Factorization of matrix functions analytic in a strip

continuous linear functional on X annihilating M− . Then f (λ − A− )−1 x = 0 for λ > −h, and thus (5.20) yields ∞ e−λt f E(t)x dt = 0, λ > −h.

0

This implies that f E(t)x = 0 for t > 0, and so, by the Hahn-Banach theorem, E(t)x ∈ M− for t > 0. Thus E(t)M− ⊂ M− for t > 0. The result of the previous paragraph enables us to deﬁne an operator-valued function E− : (0, ∞) → L(M− ) by stipulating that E− (t) = E(t)|M− . Our assumptions on the behavior of E near the origin (together with the BanachSteinhaus theorem) imply that E− can be extended to a strongly continuous function, deﬁned on 0 ≤ t < ∞, by putting x ∈ M− .

E− (0)x = lim E(t)x, t↓0

The identity (5.20) can now be written as ∞ e−λt E− (t) dt, (λ − A− )−1 x =

x ∈ M− , λ > −h.

0

Since A− (M− → M− ) is closed and densely deﬁned, it follows from Theorem 5.3 that E− is a strongly continuous right semigroup, and that A− is its inﬁnitesimal generator. In the same way one proves that E(t)M+ ⊂ M+ for t < 0, and we deﬁne E+ : (−∞, 0] → L(M+ ) by setting E+ (t) = −E(t)|M+ ,

E+ (0)x = lim −E(t)x, t↑0

x ∈ M+ .

Then the analogue of Theorem 5.3 for left semigroups shows that E+ is a strongly continuous left semigroup which has A+ (M+ → M+) as its generator. Next, consider the operator P on X deﬁned by P x = lim −E(t)x, t↑0

x ∈ X.

By the Banach-Steinhaus theorem, P is a bounded linear operator on X. For t < 0 we have that E(t) vanishes on M− , and so P x = 0 for each x ∈ M− . For x ∈ M+ and t < 0 we have E(t)x = −E+ (t)x, and thus P x = x. These properties of P imply that M− ∩ M+ = {0} and M− + M+ is closed. (5.22) The ﬁrst part of (5.22) is obvious. To prove the second part, let x1 , x2 , . . . be a sequence in M− , let y1 , y2 , . . . be a sequence in M+ , and assume that xn + yn → z for n → ∞. It suﬃces to show that z ∈ M− + M+ . Since P is continuous on X and P is zero on M− , we have P z = lim P (xn + yn ) = lim P yn . n→∞

n→∞

5.3. Realization triples

87

But P yn = yn ∈ M+ and M+ is closed. Thus P z ∈ M+ . Moreover, yn = P yn converges to P z if n → ∞. Thus xn = (xn +yn)−yn converges to z −P z if n → ∞. Also, M− is closed. We conclude that z − P z ∈ M− , and hence z = z − P z + P z belongs to M− + M+ . So M− + M+ is closed. Finally, the ﬁrst part of the proof of Theorem XV.3.1 in [51] shows that ˙ + , and that P is the M− + M+ is dense in X. We conclude that X = M− +M projection of X along M− onto M+ . Recall that D(A) = Im T , where T = A−1 . It follows that ˙ M+ ) = T M−+ ˙ T M+ = D(A− )+ ˙ D(A+ ). D(A) = T X = T (M− + Hence P maps D(A) into itself, and P commutes with A. Thus relative to the decompositions ˙ M+ , X = M− +

˙ D(A+ ), D(A) = D(A− ) +

the operator A admits the partitioning A− A= 0

0 A+

.

Therefore A is an exponentially dichotomous operator, P is the separating projection for A, and E(·) = E(· ; A).

5.3 Realization triples In this section we introduce the realizations that will be used to obtain representations of the type (5.3). We begin with some additional notation. m m By Dm 1 (R) we denote the linear submanifold of L1 (R) = L1 (R, C ) consistm m ing of all f ∈ L1 (R) for which there exists g ∈ L1 (R) such that ⎧ t ⎪ ⎪ ⎪ g(s) ds, a.e. on (−∞, 0), ⎪ ⎨ −∞ f (t) = (5.23) ∞ ⎪ ⎪ ⎪ ⎪ g(s) ds, a.e. on (0, ∞). ⎩ t

Dm 1 (R),

then there is only one g ∈ Lm If f ∈ 1 (R) such that (5.23) holds. This g is called the derivative of f and is denoted by f . From (5.23) it follows that f (0+) = limt↓0 f (t) and f (0−) = limt↑0 f (t) exist; in fact, ∞ 0 f (0+) = g(s) ds, f (0−) = g(s) ds. 0

−∞

Let ω be a negative constant. A triple Θ = (A, B, C) of operators is called a realization triple of exponential type ω if the following conditions are satisﬁed:

88

Chapter 5. Factorization of matrix functions analytic in a strip

(C1) −iA is an exponentially dichotomous operator of exponential type ω with domain D(A) and range in a Banach space X; (C2) B : Cm → X is a linear operator; (C3) C is a possibly unbounded operator with domain D(C) in X and range in Cm such that D(A) ⊂ D(C) and C is A-bounded; (C4) there exists a linear operator ΛΘ from X into Lm 1 (R) such that ∞ e−ω|t| (ΛΘ x)(t) dt < ∞, (i) sup x ≤1

−∞

(ii) for every x ∈ D(A) we have (ΛΘ x)(t) = iCE(t; −iA)x, t ∈ R, and the function ΛΘ x belongs to Dm 1 (R). In (ii), the function E(t; −iA) is the bisemigroup generated by −iA. Note that B, being a linear operator from Cm into X, is automatically bounded. Observe also that (i) implies that ΛΘ is bounded and maps X into Lm 1,ω (R) where Lm 1,ω (R) =

−ω|·| f ∈ Lm f (·) ∈ Lm 1 (R) | e 1 (R) .

(5.24)

Taking into account (ii) and the fact that D(A) is dense in X, one sees that ΛΘ is uniquely determined. Since ω is negative, Lm 1,ω (R) given by (5.24) is a linear manifold in Dm 1 (R). The space X is called the state space and the space Cm the input/output space of the triple. We shall refer to A as the main operator of the triple. Suppose Θ is a realization triple of exponential type ω and ω ≤ ω1 < 0. Then Θ is a realization triple of exponential type ω1 too. To see this, note that (i) and (ii) are fulﬁlled with ω replaced by ω1 . When the actual value of ω is not relevant, we simply call Θ a realization triple. Thus Θ = (A, B, C) is a realization triple if Θ is a realization triple of exponential type ω for some ω < 0. The operator ΛΘ does not depend on the value of ω, and the same is true with regard to the separating projection for −iA. This projection will be denoted by PΘ , although it is deﬁned in terms of A alone. The case when C is a bounded linear operator from X into Cm is of special interest. In that case C is obviously A-bounded, and (C4) is fulﬁlled with ΛΘ x = iCE(·, −iA)x for each x ∈ X. Thus when C is bounded, then conditions (C3) and (C4) are automatically satisﬁed. Let Θ = (A, B, C) be a realization triple with state space X. Notice that item (i) in (C4) implies that ΛΘ : X → Lm 1 (R) is a bounded linear operator. Since (ii) prescribes ΛΘ on D(A), the boundedness of ΛΘ and the density of D(A) in X imply that ΛΘ is uniquely determined by the operators A and C. The operator ΛΘ plays the role of the observability operator in systems theory. For its dual analogue (the controllability operator) we refer to the following proposition.

5.3. Realization triples

89

Proposition 5.4. Suppose Θ = (A, B, C) is a realization triple of exponential type ω < 0, and let ΓΘ : Lm 1 (R) → X be deﬁned by ∞ E(−t; −iA)Bϕ(t) dt, ϕ ∈ Lm (5.25) ΓΘ ϕ = 1 (R). −∞

Then ΓΘ is a bounded linear operator, and ΓΘ maps Dm 1 (R) into D(A). Proof. The operator function E(· ; −iA) is strongly continuous. Now recall the following well-known fact: if a sequence of operators converges in the strong operator topology, then the convergence is uniform on compact subsets of the underlying space. Because of the ﬁnite dimensionality of Cm , the operator B is of ﬁnite rank, hence compact. It follows that the function E(· ; −iA)B is continuous on R \ {0} with a possible jump at the origin where continuity is taken with respect the operator norm. It follows that the integral in (5.25) is well-deﬁned for each ϕ ∈ Lm 1 (R), and that ΓΘ is a bounded linear operator. Now ﬁx ϕ ∈ Dm 1 (R). For simplicity we restrict ourselves to the case when ϕ vanishes almost everywhere on (−∞, 0). By our assumption on ϕ there exists ψ ∈ Lm 1 (R) such that ∞ ϕ(t) = − ψ(s) ds, t > 0. t

But then

ΓΘ ϕ

= − = − = −

∞ 0

0

E(−t; −iA)B

∞ ∞ t

∞ s

0

0

∞

ψ(s) ds dt

t

E(−t; −iA)Bψ(s) ds dt

E(−t; −iA)Bψ(s) dt ds.

The last equality follows by applying Fubini’s theorem. Since A is exponentially dichotomous, zero belongs to the resolvent set of A. So it makes sense to consider the operator iE(−t; −iA)A−1 B. This function is diﬀerentiable on [0, ∞), and its derivative is the continuous operator-valued function −E(−t; −iA)B. Here diﬀerentiation and continuity are taken with respect to the operator norm which we can use because of the compactness of B. Thus s E(−t; −iA)B dt = iE(−s; −iA)A−1B − iPΘ A−1 B, − 0

where PΘ is the separating projection of −iA. Hence ∞ iE(−s; −iA)A−1 B − iPΘ A−1 B ψ(s) ds ΓΘ ϕ = 0

=

A−1

0

∞

iE(−s; −iA)B − iPΘ B ψ(s) ds .

90

Chapter 5. Factorization of matrix functions analytic in a strip

This shows that ΓΘ ϕ belongs to Im A−1 = D(A).

Let Θ = (A, B, C) be a realization triple of exponential type ω < 0 and having input/output space Cm . With Θ we associate two m × m matrix functions. These functions will be denoted by kΘ and WΘ , and they are called the kernel function associated with Θ and the transfer function of Θ, respectively. The ﬁrst of these is deﬁned as follows. For every u in Cm , we have that ΛΘ Bu belongs to Lm 1,ω (R). Thus the expression u ∈ Cm , (5.26) kΘ (.)u = ΛΘ Bu (.), determines a unique element kΘ of Lm×m 1,ω (R), that is each column of kΘ belongs m m m to Lm (R). In fact k (.)u ∈ L (R) ⊂ Dm Θ 1,ω 1,ω 1 (R) ⊂ L1 (R) for each u ∈ C . Next let us turn to WΘ . This function is given by WΘ (λ) = I + C(λ − A)−1 B,

|λ| < −ω.

(5.27)

To see that WΘ is well-deﬁned, ﬁx λ in the resolvent set ρ(A) of A. Since the operator (λ − A)−1 maps X into the domain D(A) of A, and D(A) is contained in the domain of C, the product C(λ − A)−1 is well-deﬁned. Hence C(λ − A)−1 B is a well-deﬁned linear transformation on Cn . The fact that −iA is an exponentially dichotomous operator of exponential type ω implies that |λ| < −ω is contained in ρ(−iA), and thus |λ| < −ω is contained in ρ(A). We conclude that WΘ is a well-deﬁned analytic m × m matrix function on |λ| < −ω. The next proposition explains the relation between the two functions WΘ and kΘ . Proposition 5.5. Suppose Θ = (A, B, C) is a realization triple of exponential type ω < 0. Then ∞ WΘ (λ) = I − eiλt kΘ (t) dt, |λ| < −ω. (5.28) −∞

Proof. It suﬃces to show that for x ∈ X and |λ| < −ω we have ∞ −1 eiλt (ΛΘ x)(t) dt, C(λ − A) x = −

(5.29)

−∞

that is, −C(λ − A)−1 x is equal to the Fourier transform (Λ Θ x)(λ) of ΛΘ x. In what follows λ is ﬁxed subject to |λ| < −ω. We already know that C(λ − A)−1 is a well-deﬁned map from X into Cm . Obviously, this map is linear. To show that it is also bounded, take x ∈ X. Using the fact that C is A-bounded, there exists a constant M such that C(λ − A)−1 x ≤ M (λ − A)−1 x + A(λ − A)−1 x . Now A(λ − A)−1 x = −x + λ(λ − A)−1 x. Thus C(λ − A)−1 x ≤ M (λ − A)−1 + 1 + |λ|(λ − A)−1 x.

5.4. Construction of realization triples

91

It follows that C(λ − A)−1 is a bounded linear operator from X into Cm . m Now consider the map x → (Λ Θ x)(λ) from X into C . This map is linear and bounded too. Linearity is obvious. Boundedness follows from the estimate ∞ x)(λ) ≤ e−ω|t| ΛΘ x(t) dt, (Λ Θ −∞

together with condition (i) in the deﬁnition of a realization triple. We have now shown that, for λ ﬁxed, both sides of (5.29) are continuous in x. Hence it suﬃces to prove (5.29) for x ∈ D(A) because of D(A) = X. Take x ∈ D(A), and put y = Ax. Since −iA is an exponentially dichotomous operator of exponential type ω, we use (5.9) for −iA in place of A and −iλ in place of λ to show that ∞ eiλt E(t; −iA)y dt. (5.30) (λ − A)−1 y = −i −∞

Recall that CA−1 is a bounded linear operator. It follows that C(λ − A)−1 x = =

CA−1 (λ − A)−1 y ∞ −i eiλt CA−1 E(t; −iA)y dt −∞

=

−i

−∞

=

−i

∞

∞

−∞

eiλt CE(t; −iA)x dt eiλt (ΛΘ x)(t) dt,

the latter equality holding by virtue of condition (ii) in the deﬁnition of a realization triple. Thus (5.29) is proved. From (5.29) it follows that C(λ − A)−1 is analytic on |λ| < −ω. This result can also be proved directly using that C is A-bounded. In fact, employing the C-boundedness of A one can show that the function λ → C(λ − A)−1 is analytic on the resolvent set ρ(A).

5.4 Construction of realization triples In this section we construct a representation of the form (5.3) for the m × m matrix-valued function W in (5.1) with the kernel function k being given by (5.2). The following theorem is the main result. Theorem 5.6. An m×m matrix function W is the transfer function of a realization triple if and only if W is of the form ∞ eiλt k(t) dt, (5.31) W (λ) = I − −∞

92

Chapter 5. Factorization of matrix functions analytic in a strip

where k is an m × m matrix function with the property that there exist ω < 0 and h ∈ Lm×m (R) such that 1 (5.32) k(t) = eω|t| h(t). If W is given by (5.31) and (5.32) for some ω < 0 and h ∈ Lm×m (R), then 1 W = WΘ with Θ = (A, B, C) constructed in the following way: the state space X m of Θ is Lm 1 (R), the input/output space is C , D(A) = D(C) = Dm 1 (R), −iωf (t) + if (t), (Af )(t) = iωf (t) + if (t), (By)(t) = e−ω|t| k(t)y, ∞ f (s) ds. Cf = i

a.e. on −∞ < t < 0, a.e. on 0 < t < ∞,

a.e. on R,

−∞

Proof. Let Θ be a realization triple, and let W = WΘ be its transfer function. Then, by Proposition 5.5 in the preceding section, (5.31) holds with k = kΘ . Using the fact that the second operator in a realization triple is bounded, we see from (i) in the deﬁnition of a realization triple that ∞ e−ω|t| kΘ (t)y dt < ∞, sup y ≤1

−∞

for some ω < 0. Hence k = kΘ satisﬁes (5.32). This proves the “if part” of the theorem. Next, let W be given by (5.31) and (5.32) for some ω < 0 and h ∈ Lm×m (R), 1 and let Θ = (A, B, C) be the triple of operators deﬁned in the second part of the theorem. We need to show that this triple is a realization triple and that W = WΘ . As is well-known (cf., [51], page 420), the backward translation semigroup on Lm 1 [0, ∞) is strongly continuous. The inﬁnitesimal generator of this semigroup has Dm 1 [0, ∞) as its domain and its action amounts to taking the derivative. Here Dm [0, ∞) is the linear manifold consisting of all functions f ∈ Dm 1 1 (R) with the property that f (t) = 0 for t < 0, and hence the derivative f is well-deﬁned for each f ∈ Dm 1 [0, ∞). Using this, one sees that −iA an exponentially dichotomous operator of exponential type ω and that the bisemigroup associated with −iA acts as follows: for t < 0, −e−ωt f (t + s), a.e. on −∞ < s < 0, E(t; −iA)f (s) = 0, a.e. on 0 < s < ∞, and for t > 0, E(t; −iA)f (s) =

0,

a.e. on −∞ < s < 0,

eωt f (t + s),

a.e. on 0 < s < ∞.

5.5. Inverting matrix functions analytic in a strip

93

The separating projection for −iA is the projection of the state space X = Lm 1 (R) m onto Lm (−∞, 0] along L [0, ∞). 1 1 Condition (5.32) on k implies that the operator B from Cm into Lm 1 (R) is bounded. From the deﬁnition of C and A we see that Cf ≤ −ωf + Af ,

f ∈ D(A).

Thus C is A-bounded. Deﬁne Λ : X → Lm 1 (R) by (Λf )(t) = eω|t| f (t),

a.e. on R.

(5.33)

Then Λ satisﬁes the conditions (i) and (ii) in the deﬁnition of a realization triple with Λ in place of ΛΘ . For (i) this is obvious. To check the ﬁrst part of (ii), one uses the above description of the bisemigroup E(t; −iA) and the deﬁnition of C. As to the second part of (ii), observe that f ∈ Dm 1 (R) and ω < 0 imply that the function eωt f (t) belongs to Dm 1 (R) too. We have now proved that Θ = (A, B, C) is a realization triple We claim that the kernel function kΘ associated with Θ coincides with k. Indeed, for y ∈ Cm the following identities hold almost everywhere on R: kΘ (t)y = (ΛBy)(t) = (eω|t| By)(t) = k(t)y. Since Cm has a ﬁnite basis, it follows that kΘ (t) = k(t) almost everywhere on R. In other words, kΘ and k coincide as elements of Lm×m (R). 1

5.5 Inverting matrix functions analytic in a strip Let Θ = (A, B, C) be a realization triple with state space X. In this section we shall employ the operator A× (X → X). Here is the deﬁnition: the domain of A× is equal to the domain of A, and its action is deﬁned by A× = A− BC. We call A× the associate main operator of the triple Θ. As one may expect from Section 2.4, the operator A× plays an important role in inverting WΘ (λ). In fact, we have the following theorem. Theorem 5.7. Let the m × m matrix function W be given by W (λ) = I + C(λ − A)−1 B, with Θ = (A, B, C) being a realization triple. Let A× be the associate main operator of Θ. Then W (λ) is invertible for each λ ∈ R if and only if the spectrum of A× does not intersect the real line. In that case (A× , B, −C) is a realization triple, and W (λ)−1 = I − C(λ − A× )−1 B, λ ∈ R, (5.34) W (λ)C(λ − A× )−1 = C(λ − A)−1 , × −1

(λ − A ) × −1

(λ − A )

−1

= (λ − A)

−1

BW (λ) = (λ − A) −1

− (λ − A)

−1

BW (λ)

B,

λ ∈ R,

(5.35)

λ ∈ R,

(5.36)

−1

C(λ − A)

,

λ ∈ R. (5.37)

94

Chapter 5. Factorization of matrix functions analytic in a strip

Proof. We split the proof into four parts. In the ﬁrst part we show that W (λ) is invertible for each λ ∈ R if and only if the spectrum of A× does not intersect the real line, and we derive the expressions (5.34) – (5.37). The remaining three parts are concerned with the statement that (A× , B, −C) is a realization triple. Part 1. Suppose A× has no spectrum on the real line. This condition means that for each real λ the linear operator λ − A× maps D(A× ) = D(A) in a one-one way onto X, and hence the linear operator I − C(λ − A× )−1 B acting on Cm is well-deﬁned. We claim that it is the inverse of W (λ). To see this we ﬁrst prove (5.35). From BCx = (A − A× )x for each x ∈ D(A), it follows that BC(λ − A)−1 = (A − A× )(λ − A)−1 . Using the latter identity and ﬁxing λ ∈ R, we obtain the equality (5.35) from the following calculation: W (λ)C(λ − A× )−1 = C(λ − A× )−1 + C(λ − A)−1 BC(λ − A)−1 = C(λ − A× )−1 + C(λ − A)−1 (A − A× )(λ − A)−1 = C(λ − A× )−1 + C(λ − A)−1 (A − λ) + (λ − A× ) (λ − A)−1 = C(λ − A)−1 . From (5.35) we obtain that W (λ) I − C(λ − A× )−1 B = W (λ) − C(λ − A)−1 = I,

λ ∈ R.

Hence W (λ) is invertible for each λ ∈ R. Next assume W (λ) is invertible for each λ ∈ R. We claim that A× has no spectrum on the real line and that (5.37) holds. To prove this, ﬁx λ ∈ R and let R(λ) be the operator on X deﬁned by the right-hand side of (5.37). Since (λ − A× )(λ − A)−1 = I + BC(λ − A)−1 , we have (λ − A× )R(λ)

=

I + BC(λ − A)−1 + − I − BC(λ − A)−1 BW (λ)−1 C(λ − A)−1

=

I + BC(λ − A)−1 +B − I − C(λ − A)−1 B W (λ)−1 C(λ − A)−1

=

I + BC(λ − A)−1 − BC(λ − A)−1 ,

and so (λ − A× )R(λ) = I. Thus to prove (5.37) it remains to show that λ − A× is one-to-one. Let x ∈ D(A× ) = D(A) and suppose (λ−A× )x = 0. Since A× x = Ax−BCx, we have (λ − A)−1 BCx = −x, and hence W (λ)Cx = Cx + C(λ − A)−1 BCx = Cx − Cx = 0.

5.5. Inverting matrix functions analytic in a strip

95

By assumption W (λ) is invertible. Therefore Cx = 0 and, consequently, (λ−A)x = (λ−A× )x = 0. Now use the fact that A has no spectrum on the real line. It follows that x = 0, and hence λ − A× is one-to-one. Note that in passing we established (5.34), (5.35) and (5.37). The argument for (5.36) is analogous to that for (5.35). In the remaining three parts it is assumed that A× has no spectrum on the real line, or equivalently, that W (λ) is invertible for each λ ∈ R. Part 2. We show that A× is closed and that C is A× -bounded. Applying (5.37) with λ = 0 we see that (A× )−1 = A−1 + A−1 BW (0)−1 CA−1 .

(5.38)

Since C is A-bounded, the operator CA−1 is bounded. Thus in the right-hand side of (5.38) the operators B, A−1 and CA−1 are all bounded. It follows that (A× )−1 is bounded too. Hence A× is a closed operator. Recall that the operators A−1 and (A× )−1 map X into D(A) = D(A× ). Since the latter space is contained in D(C), we can apply C to both sides of (5.38). This yields C(A× )−1 = CA−1 + CA−1 BW (0)−1 CA−1 . But CA−1 is bounded. Hence C(A× )−1 is bounded, which implies that C is A× bounded. Part 3. In this part we show that −iA× is exponentially dichotomous. To do this we apply Theorem 5.2. First some preparations. Recall that ∞ W (λ) = I − eiλt kΘ (t) dt, λ ∈ R, −∞

(R). By the matrix-valued version of with kΘ belonging to the space eω|·| Lm×m 1 Wiener’s theorem (see, e.g., [52], page 830), the fact that W (λ) is invertible for each λ ∈ R implies that ∞ W (λ)−1 = I − eiλt k × (t) dt, λ ∈ R, (5.39) −∞

(R). In fact (see [47], Section 18), taking |ω| smaller if necfor some k × ∈ Lm×m 1 essary we may assume that k × also belongs to eω|·| Lm×m (R). Next note that for 1 each x ∈ X and each y ∈ Cm , ∞ C(λ − A)−1 x = −i eiλt (ΛΘ x)(t) dt, λ ∈ R, −∞

(λ − A)−1 x

= −i

(λ − A)−1 By

= −i

∞

−∞

∞

−∞

eiλt E(t; −iA)x dt,

λ ∈ R,

eiλt E(t; −iA)By dt,

λ ∈ R;

96

Chapter 5. Factorization of matrix functions analytic in a strip

cf., (5.29) and (5.30). Using these formulas in (5.37), and taking inverse Fourier transforms, we see that ∞ (λ − A× )−1 x = −i eiλt E(t; −iA) + E1 (t) + E2 (t) x dt, λ ∈ R, −∞

where for each x ∈ X we have E1 (t)x = i E2 (t)x = −i

∞

−∞

∞ −∞

E(t − s; −iA)B(ΛΘ x)(s) ds,

E(t − s; −iA)B

∞

−∞

k × (s − r)(ΛΘ x)(r) dr

(5.40) ds.

(5.41)

Recall that the function E(· ; −iA)B is exponentially decaying, that k × belongs to eω|·|Lm×m (R), and that for each x ∈ X the function ΛΘ x belongs to eω|·| Lm 1 (R). 1 These facts imply that E1 and E2 are exponentially decaying too. Moreover, a routine argument shows that these functions are strongly continuous, that is, for each x ∈ X the functions E1 (·)x and E2 (·)x are continuous in the norm of X. We conclude that the function E(· ; −iA) + E1 (·) + E2 (·) is exponential decaying, strongly continuous on R \ {0}, and that at zero it has (at worst) a jump discontinuity. But then we can apply Theorem 5.2 with A replaced by −iA× and λ replaced by −iλ to show that −iA× is exponentially dichotomous. Furthermore, the bisemigroup generated by −iA× is given by E(· ; −iA× ) = E(· ; −iA) + E1 (·) + E2 (·),

(5.42)

where E(· ; −iA) is the bisemigroup generated by −iA, and the functions E1 (·) and E2 (·) are given by (5.40) and (5.41), respectively. Part 4. In this part we complete the proof and show Θ× = (A× , B, −C) is a realization triple. The negative constant ω having been taken suﬃciently close to zero, one has that Θ is of exponential type ω and k× belongs to eω|·| Lm×m (R). A 1 standard reasoning now shows that the convolution product k× ∗ ΛΘ x , given by ∞ × k ∗ (ΛΘ x) (t) = k × (t − s)(ΛΘ x)(s) ds, a.e. on R, −∞

determines a bounded linear operator from X into Lm 1 (R) such that ∞ e−ω|t| k× ∗ (ΛΘ x)(t) dt < ∞. sup x ≤1

−∞

But then the expression Λ× x = −ΛΘ x + k × ∗ (ΛΘ x)

(5.43)

deﬁnes a bounded linear operator Λ× : X → Lm 1 (R) for which condition (i) in the deﬁnition of a realization triple (Section 5.3), with ΛΘ replaced by Λ× , is satisﬁed.

5.5. Inverting matrix functions analytic in a strip

97

Next, take x ∈ X, and consider the Fourier transform of Λ× x. Using formula (5.43) we see that for each λ ∈ R we have × x)λ (Λ

× = −(Λ Θ x)(λ) + k (λ)(ΛΘ x)(λ)

=

I − k × (λ) (Λ Θ x)(λ)

= −W (λ)−1 C(λ − A)−1 x. In this calculation the ﬁnal equality results from (5.29) and (5.39). Next, using (5.35) we see that × x)(λ) = C(λ − A× )−1 x, (Λ

λ ∈ R.

(5.44)

Note that this equality actually holds in a strip |λ| < −ω containing the real line. Now take x ∈ D(A× ) = D(A), and put z = A× x. Then C(λ − A× )−1 x = × −1 C(A ) (λ − A× )−1 z, and the operator C(A× )−1 is bounded by the result of the second part of the proof. Since −iA× is exponentially dichotomous, by the third part of the proof, we can use formula (5.9), with A replaced by −iA× and by −iλ, to show that ∞ × x)(λ) = −iC(A× )−1 eiλt E(t; −iA× )z dt (Λ =

−i −i

∞

−∞

=

−∞

∞

−∞

eiλt C(A× )−1 E(t; −iA× )z dt eiλt CE(t; −iA× )x dt,

λ ∈ R.

Thus we have proved that (Λ× x)(t) = −iCE(t; −iA× )x almost everywhere on R. It remains to show that Λ× x ∈ Dm 1 (R). In view of the properties of ΛΘ and the identity (5.43), it suﬃces to show m that k × ∗ (ΛΘ x) belongs to Dm 1 (R). Since ΛΘ x = D1 (R), we can consider its derivative g, that is the function given by ⎧ t ⎪ ⎪ g(s) ds, a.e. on (−∞, 0), ⎪ ⎨ (ΛΘ x)(t) =

−∞

⎪ ⎪ ⎪ ⎩ −

t

∞

g(s) ds,

a.e. on (0, −∞).

Now use that (k × ∗ f ) = k × ∗ f + k × (· ) f (0+) − f (0−) ,

f ∈ Dm 1 (R).

98

Chapter 5. Factorization of matrix functions analytic in a strip

If follows that

k × ∗ (ΛΘ x) (t) =

⎧ ⎪ ⎪ ⎪ ⎨

t

h(s) ds,

a.e. on (−∞, 0),

−∞

⎪ ⎪ ⎪ ⎩ −

∞

h(s) ds

0, s < 0.

It follows that, almost everywhere on [0, ∞), 0 (L+ ψ)(t) = −iC1 E(t; −iA) E(−s; −iA)B1 ψ(s) ds = −iCE(t; −iA)

−∞ 0

−∞

E(−s; −iA)Bψ(s) ds

= (−QΛΘ ΓΘ ψ)(t), and (5.49) has been obtained for the case when Im B ⊂ D(A). The general situation, where Im B need not be contained in D(A), can be treated with an approximation argument based on the fact that B can be approximated (in norm) by bounded linear operators from Cm into X with ranges inside D(A). This is true because D(A) is dense in X and Cm is ﬁnite dimensional.

5.7 Inverting Wiener-Hopf integral operators In this section we study inversion of the Wiener-Hopf integral operator T : ∞ k(t − s)f (s) ds, a.e. on [0, ∞). (5.51) T f (t) = f (t) − 0

5.7. Inverting Wiener-Hopf integral operators

101

It will be assumed that the m × m matrix kernel function k is the kernel function associated with some realization triple. This implies that T is a well-deﬁned bounded linear operator on Lm 1 (R). We shall prove the following theorem. Theorem 5.10. Let T be the Wiener-Hopf integral operator on Lm 1 (R) given by (5.51). Assume that k = kΘ for some realization triple Θ = (A, B, C). Then T is invertible if and only if the following two conditions are satisﬁed: (i) Θ× = (A× , B, −C) is a realization triple, ˙ Ker PΘ× . (ii) X = Im PΘ + Here X is the state space of both Θ and Θ× , and PΘ and PΘ× are the separating projections of −iA and −iA× , respectively. If (i) and (ii) hold, the inverse of T is given by ∞ kΘ× (t − s)g(s) ds (T −1 g)(t) = g(t) − 0 ∞ ΛΘ× (I − Π)E(−s, −iA× )Bg(s) ds(t), a.e. on [0, ∞). − 0

Here Π is the projection of X onto Ker PΘ× along Im PΘ . To facilitate the proof of Theorem 5.10 we ﬁrst establish two lemmas. If Θ is a realization triple with main operator A, the separating projection of the operator −iA will be denoted by PΘ . Lemma 5.11. Let Θ = (A, B, C) and Θ× = (A× , B, −C) be realization triples with state space X. Then the operator J × : Im PΘ → Im PΘ× ,

J × x = PΘ× x,

(5.52)

˙ Ker PΘ× , and in that case is invertible if and only if X = Im PΘ + (J × )−1 = (I − Π)|Im PΘ× ,

Π = I − (J × )−1 PΘ× ,

(5.53)

where Π is the projection of X along Im PΘ onto Ker PΘ× . Proof. Obviously Ker J × = Im PΘ ∩ Ker PΘ× . Thus J × is one-to-one if and only if Im PΘ ∩ Ker PΘ× = {0}. Next, assume J × is surjective. Take x ∈ X. Then PΘ× x = J × PΘ z = PΘ× PΘ z for some z ∈ X. This yields x

= PΘ× x + (I − PΘ× )x = PΘ× PΘ z + (I − PΘ× )x = PΘ z + (I − PΘ× )(x − PΘ z).

Hence x ∈ Im PΘ + Ker PΘ× , and we conclude that Im PΘ + Ker PΘ× = X. Thus ˙ PΘ× provided that J × is invertible. Moreover, the above calcuX = Im PΘ +Ker lations show that × −1 PΘ× x = PΘ z = (I − Π)x = (I − Π)PΘ× PΘ z = (I − Π)PΘ× x, J

102

Chapter 5. Factorization of matrix functions analytic in a strip

which proves the ﬁrst identity in (5.53). ˙ Ker PΘ× . Then J × is injective. To complete the proof, assume X = Im PΘ + × To prove that J is surjective, take y ∈ Im PΘ× . Since PΘ× y = y and PΘ× Π = 0, we have y = PΘ× y = PΘ× (I − Π)y + PΘ× Πy = PΘ× (I − Π)y. Put x = (I − Π)y. Then x ∈ Im PΘ and J × x = y. This shows that J × is surjective, and thus J × is invertible. Moreover, we see that (J × )−1 y = x = (I − Π)y, which proves the second identity in (5.53). Lemma 5.12. Assume that Θ = (A, B, C) and Θ× = (A× , B, −C) are realization triples, with Cm being the input/output space of both Θ and Θ× . Introduce the maps m K : Lm 1 [0, ∞) → L1 [0, ∞),

∞

(Kϕ)(t) = 0 m K × : Lm 1 [0, ∞) → L1 [0, ∞),

(K × ϕ)(t) =

kΘ (t − s)ϕ(s) ds,

∞

0

a.e. on [0, ∞),

× kΘ (t − s)ϕ(s) ds,

U : Im PΘ× → Lm 1 [0, ∞),

(U x)(t) = (ΛΘ x)(t),

U × : Im PΘ → Lm 1 [0, ∞),

(U × x)(t) = −(ΛΘ× x)(t),

R : Lm 1 [0, ∞) → Im PΘ ,

Rϕ =

R × : Lm 1 [0, ∞) → Im PΘ× ,

R× ϕ = −

J : Im PΘ× → Im PΘ ,

Jx = PΘ x,

J × : Im PΘ → Im PΘ× ,

J × x = PΘ× x.

∞

0

a.e. on (−∞, 0]),

a.e. on [0, ∞), a.e. on [0, ∞),

E(−t; −iA)Bϕ(t) dt,

∞

0

E(−t; −iA× )Bϕ(t) dt,

Then all these operators are well-deﬁned, linear and bounded. Moreover,

I −K

U

R

J

I − K× R

×

m ˙ ˙ : Lm 1 [0, ∞) + Im PΘ× → L1 [0, ∞) + Im PΘ ,

U× J

×

m ˙ ˙ : Lm 1 [0, ∞) + Im PΘ → L1 [0, ∞) + Im PΘ× ,

5.7. Inverting Wiener-Hopf integral operators are bounded linear operators, which are invertible, and −1 I −K U I − K× U × . = R× J× R J

103

(5.54)

Proof. As we have seen in Section 5.6 the operators K and K × are bounded. To see that the other operators are well-deﬁned and bounded too it suﬃces to m make a few observations. Let Q be the projection of Lm 1 (R) onto L1 [0, ∞) along m L1 (−∞, 0]. Then U = QΛΘ |Im PΘ× ,

U × = −QΛΘ× |Im PΘ ,

and hence these two operators are well-deﬁned and bounded. Next, viewing PΘ and PΘ× as operators from X onto Im PΘ and Im PΘ× , respectively, we have , R = PΘ ΓΘ |Lm 1 [0,∞)

R× = −PΘ× ΓΘ× |Lm . 1 [0,∞)

From these identities and Proposition 5.4 it follows that R and R× are also welldeﬁned and bounded. It remains to prove (5.54). This amounts to checking eight identities. Pairwise these identities have analogous proofs. So, actually only four identities have to be taken care of. This will be done in the remaining part of the proof which is divided into four steps. Step 1. First we prove that R(I − K × ) + JR× = 0. Take ϕ in Lm 1 [0, ∞). We need to show that RK × ϕ = PΘ R× ϕ + Rϕ. Whenever this is convenient, it may be assumed that ϕ is a continuous function with compact support in (0, ∞). By applying Fubini’s theorem, one gets ∞ ∞ RK × ϕ = E(−t; −iA)BkΘ× (t − s)ϕ(s) ds dt 0

0

∞

= 0

0

∞

E(−t; −iA)BkΘ× (t − s)ϕ(s) dt ds.

For s > 0 and x ∈ X, consider the identity ∞ E(−t; −iA)B(ΛΘ× x)(t − s) dt 0

= E(−s; −iA)x − PΘ E(−s; −iA× )x. To prove it, we ﬁrst take x ∈ D(A) = D(A× ). Then, for t = 0 and t = s, d E(−t; −iA)E(t − s; −iA× )x dt = iE(−t; −iA)BCE(t − s; −iA× )x = iE(−t; −iA)BC(A× )−1 E(t − s; −iA× )A× x.

(5.55)

104

Chapter 5. Factorization of matrix functions analytic in a strip

Because C(A× )−1 is bounded, the last expression is a continuous function of t on the intervals [0, s] and [s, ∞). It follows that (5.55) holds for x ∈ D(A). The validity of (5.55) for arbitrary x ∈ X can now be obtained by a standard approximation argument based on the fact that D(A) is dense in X and the continuity of the operators involved. Substituting (5.55) in the expression for RK × ϕ, one immediately gets the desired identity R(I − K × ) + JR× = 0. Step 2. Next we show that RU × + JJ × = IIm PΘ . Take x in Im PΘ . Then ∞ RU × x = − E(−t; −iA)B(ΛΘ× x)(t) dt. (5.56) 0

Apart from the minus sign, the right-hand side of (5.56) is exactly the same as the left-hand side of (5.55) for s = 0. It is easy to check that (5.55) also holds for s = 0, provided that the right-hand side is interpreted as −PΘ x + PΘ PΘ× x. Thus RU × x = PΘ× x = x − PΘ PΘ× x, and the desired identity RU × + JJ × = IIm PΘ is proved. Step 3. This step concerns the identity (I − K)U × + U J × = 0. Take x ∈ Im PΘ . m Then U × x = −QΛΘ× x, where Q is the projection of Lm 1 (R) onto L1 [0, ∞) along m L1 (−∞, 0]. Here the latter two spaces are considered as subspaces of Lm 1 (R). Observe now that QΛΘ× = ΛΘ× (I − PΘ× )x. For x ∈ D(A) = D(A× ) this is evident, and for arbitrary x one can use an approximation argument. Hence KU × x = Qh, where h = −kΘ ∗ ΛΘ× (I − PΘ× )x , that is, h is the (full line) convolution product of −kΘ and ΛΘ× (I − PΘ× )x. Taking Fourier transforms, one gets h(λ)

= C(λ − A)−1 BC(λ − A× )−1 (I − PΘ× )x = C(λ − A)−1 (I − PΘ× )x − C(λ − A× )−1 (I − PΘ× )x.

Put g = U × x + U PΘ× x. Since both U and U × map into Im Q = Lm 1 [0, ∞), we have g = Qg. Also g = −ΛΘ× (I − PΘ× )x + ΛΘ(I − PΘ )PΘ× x, and hence g(λ) = −C(λ − A× )−1 (I − PΘ× )x − C(λ − A)−1 (I − PΘ )PΘ× x. Since x ∈ Im PΘ , it follows that h(λ) − g(λ) = C(λ − A)−1 PΘ (I − PΘ× )x. So h(λ) − g(λ) is the Fourier transform of −ΛΘ PΘ (I − PΘ× )x. But then h − g = −ΛΘ PΘ (I − PΘ× )x = −(I − Q)ΛΘ (I − PΘ× )x. Applying Q to both sides of this identity, we get Qh = Qg = g. In other words, KU × x = U × x + PΘ× x for all x ∈ X, and this is nothing else than the identity (I − K)U × + U J × = 0. Step 4. Finally, we prove (I − K)(I − K × ) + U R× = I. Let L be the (full line) convolution integral operator associated with kΘ , featured in Theorem 5.8. Since Θ and Θ× are both realization triples, the operator I − L is invertible with inverse (I − L)−1 = I − L× , where L× is the convolution integral operator associated

5.7. Inverting Wiener-Hopf integral operators

105

m ˙ m with Θ× . With respect to the decomposition Lm 1 (R) = L1 [0, ∞) + L1 (−∞, 0), we write I − L and its inverse in the form I − K× ∗ I − K L+ × . (5.57) , I −L = I −L= ∗ ∗ ∗ L× −

Thus L+ is the right Hankel operator associated with kΘ , and the operator L× − is the left Hankel operator associated with kΘ× . But then Lemma 5.9 yields L+ ψ

=

L× −ϕ =

−QΛΘ ΓΘ ψ,

ψ ∈ Lm 1 (−∞, 0],

(5.58)

(I − Q)ΛΘ× ΓΘ× ϕ,

ϕ ∈ Lm 1 [0, ∞).

(5.59)

Since I − L× is the inverse of I − L, formula (5.57) shows that (I − K)(I − K × ) + L+ L× − = I. × So, in order to get the desired identity, it suﬃces to show that L+ L× − = UR . As was observed in the last paragraph of Step 2 of the present proof, (5.55) also holds for s = 0, that is ∞ E(−t; −iA)B(ΛΘ× x)(t) dt = PΘ (I − PΘ× )x, x ∈ X. 0

Analogously, one has 0 E(−t; −iA)B(ΛΘ× x)(t) dt = (I − PΘ )PΘ× x, −∞

x ∈ X.

Using the expressions for L+ and L× − given in (5.58) and (5.59) we obtain L+ L × −ϕ

= −QΛΘ ΓΘ (I − Q)ΛΘ× ΓΘ× ϕ = −QΛΘ (I − PΘ )PΘ× ΓΘ× ϕ = U RΘ ϕ.

Thus (I − K)(I − K × ) + U R× = I holds, and the lemma is proved.

Following [13] (see also Section III.4 in [51]) we summarize the result of the preceding lemma by saying that the operators I −K and J × are matricially coupled with (5.54) being the coupling relation. The coupling relation is very useful. For instance, this relation and Corollary III.4.3 in [51] immediately yield the following result. Corollary 5.13. Let the operators K, K × , U, U × , R, R× , J and J × be as in (5.54). Then I − K is invertible if and only if J × is invertible, and in that case (I − K)−1 = I − K × − U × (J × )−1 R× ,

(J × )−1 = J − R(I − K)−1 U.

(5.60)

106

Chapter 5. Factorization of matrix functions analytic in a strip

Proof of Theorem 5.10. We split the proof into two parts. In the ﬁrst part we show that the invertibility of T implies that Θ× is a realization triple. In the second part we assume that Θ× is a realization triple and complete the proof by using Lemma 5.11 and Corollary 5.13. Part 1. Since the kernel function k is equal to kΘ , we know from Proposition 5.5 that the symbol of T is equal to WΘ . Assume T is invertible. From the general theory of Wiener-Hopf operators we know that this assumption implies that WΘ (λ) is invertible for all real λ. But then we can use the ﬁnal part of Theorem 5.7 to show that Θ× is a realization triple. Part 2. In this part we assume that Θ× is a realization triple. From Corollary 5.13 we know that T = I −K is invertible if and only if J × is invertible. By Lemma 5.11 the latter happens if and only if condition (ii) is satisﬁed. Together with the result of the ﬁrst part we have now shown that T is invertible if and only if conditions (i) and (ii) are both fulﬁlled. Moreover, if these conditions are fulﬁlled we see from the ﬁrst parts of formulas (5.60) and (5.53) that (I − K)−1 = I − K × − U × (I − Π)R× , where K × , R× and U × are as in Lemma 5.12, and Π is the projection of X along Im PΘ onto Ker PΘ× . Using the deﬁnitions of the operators K × , R× and U × given in Lemma 5.12, the formula for T −1 presented in Theorem 5.10 is now clear.

5.8 Explicit canonical factorization In this section we use realization triples to construct a canonical factorization for an m × m matrix function W of the form (5.1) with k being given by (5.2). By Theorem 5.6 such a function is the transfer function of a realization triple Θ = (A, B, C). In what follows it is assumed that Θ is given. We present necessary and suﬃcient conditions for the existence of a canonical factorization in terms of the operators appearing in the realization triple, Also, supposing these conditions are fulﬁlled, we give formulas for the factors and their inverses in a canonical factorization of W . The main result (Theorem 5.14 below) is the natural analogue of Theorem 5.3 for the functions considered in this section. For the deﬁnition of a canonical factorization, see Section 1.1 (cf., also Section 3.1). Theorem 5.14. Let the m × m matrix function W be given by W (λ) = I + C(λ − A)−1 B, with Θ = (A, B, C) a realization triple, and let A× be the associate main operator of Θ. Then W admits a canonical factorization with respect to the real line if and only if the following two conditions are satisﬁed: (i) Θ× = (A× , B, −C) is a realization triple,

5.8. Explicit canonical factorization

107

˙ Ker PΘ× . (ii) X = Im PΘ + Here X is the state space of both Θ and Θ× , and PΘ and PΘ× are the separating projections of −iA and −iA× , respectively. If the conditions (i) and (ii) are satisﬁed, then the projection Π of X along Im PΘ onto Ker PΘ× maps D(A) = D(A× ) into itself, and a canonical factorization W = W− W+ with respect to the real line is given by W (λ) = W− (λ)W+ (λ), λ ∈ R, where the factors and their inverses can be written as W− (λ)

=

I + C(λ − A)−1 (I − Π)B,

W+ (λ)

=

I + CΠ(λ − A)−1 B,

W−−1 (λ)

=

I − C(I − Π)(λ − A× )−1 B,

W+−1 (λ)

=

I − C(λ − A× )−1 ΠB.

The projection Π maps D(A) = D(A× ) into itself and D(A) ⊂ D(C). Hence the right-hand sides of the ﬁrst two of the above four expressions are well-deﬁned on ρ(A), and those of the last two are well-deﬁned on ρ(A× ). In particular the formulas make sense for λ in a strip containing the real line. At ﬁrst sight this seems to be short of the requirements for Wiener-Hopf factorization. We will come back to and resolve this point at the end of the proof. Proof of Theorem 5.14. The proof will be divided into four parts. In the ﬁrst we show that the conditions (i) and (ii) are necessary and suﬃcient. In the remaining three parts we assume that (i) and (ii) are satisﬁed. Part 1. Let K be the Wiener-Hopf integral operator with kernel function kΘ . Then the function W is the symbol of the operator I − K, and hence we know from the general theory of Wiener-Hopf integral equations that W admits a canonical factorization with respect to the real line if and only if T = I − K is invertible. The ﬁrst part of Theorem 5.10 implies that the latter is satisﬁed if and only if (i) and (ii) are fulﬁlled. Thus (i) and (ii) are necessary and suﬃcient in order that W admits a canonical factorization with respect to the real line. In the remaining three parts of the proof we assume that conditions (i) and (ii) are satisﬁed; Π will be the projection of X along Im PΘ onto Ker PΘ× . Part 2. In this part we show that Π maps D(A) into itself. To do this we need the operator J × deﬁned by (5.52). Our hypotheses imply (see Lemma 5.11) that J × is invertible and that Π = I − (J × )−1 PΘ× . Recall that PΘ× maps D(A) into D(A) ∩ Im PΘ× . Thus in order to prove that D(A) is invariant under Π it suﬃces to show that (J × )−1 maps D(A) ∩ Im PΘ× into D(A). From the relation (5.54) and the invertibility of the operator I − K, it follows (see Corollary 5.13) that (J × )−1 = J − R(I − K)−1U,

108

Chapter 5. Factorization of matrix functions analytic in a strip

where U , R and J are as in Lemma 5.12. Take x ∈ D(A) ∩ Im PΘ× . Then U x = m m QΛΘ x ∈ Dm 1 [0, ∞), where Q is the projection of L1 (R) onto L1 [0, ∞) along m L1 (−∞, 0]. From the general theory of Wiener-Hopf operators we know that (I − K)−1 = (I + Γ1 )(I + Γ2 ),

(5.61)

where for j = 1, 2 the operator Γj is the integral operator given by ∞ (Γj ϕ)(t) = γj (t − s)ϕ(s) ds, t > 0, 0

Lm×m (R). 1

with γj ∈ In fact, γ1 has its support in [0, ∞) and γ2 in (−∞, 0]; see Section 1.5. From the representation (5.61) it follows that (I − K)−1 maps m×m Dm (R) and f ∈ Dm 1 [0, ∞) into itself. Note in this context that for h ∈ L1 1 (R), m we have h ∗ f ∈ D1 (R) and (h ∗ f ) = h ∗ f + h(·) f (0+) − f (0−) . Thus (I − K)−1 U x ∈ Dm 1 (R). But then the ﬁnal part of Proposition 5.4 tells us that we end up in D(A) by applying ΓΘ . We conclude that R(I − K)−1 U maps D(A) ∩ Im PΘ× into D(A). Since the separating projection PΘ maps D(A) into itself, we know that J maps D(A) ∩ Im PΘ× also into D(A). Thus (J × )−1 maps D(A) ∩ Im PΘ× into D(A), and hence Π maps D(A) into itself. Amplifying on the above, we note that J × maps D(A) ∩ Im PΘ in a one-to-one way onto D(A) ∩ Im PΘ× . Since J × is invertible, it suﬃces to show that (J × )−1 maps D(A) ∩ Im PΘ× into D(A) ∩ Im PΘ . We have already shown that (J × )−1 maps D(A) ∩ Im PΘ× into D(A), and the inclusion (J × )−1 Im PΘ× ⊂ Im PΘ is clear from the deﬁnition of J × . Thus J × has the desired property. Part 3. According to our hypotheses and the fact that Π maps D(A) into itself, we have the following direct sum decompositions:

Write

A=

X

=

D(A)

=

A1

Z

0

A2

˙ Ker PΘ× , Im PΘ + ˙ D(A) ∩ Ker PΘ× . D(A) ∩ Im PΘ +

,

B=

B1 B2

,

C=

C1

C2

(5.62) (5.63)

(5.64)

for the corresponding matrix representations of A, B, and C. We now show that Θ1 = (A1 , B1 , C1 ),

× Θ× 1 = (A1 , B1 , −C1 ),

Θ2 = (A2 , B2 , C2 ),

× Θ× 2 = (A2 , B2 , −C2 ),

are realization triples, and we analyze the spectral properties of their main oper× ators. Here A× 1 = A1 − B1 C1 and A2 = A2 − B2 C2 .

5.8. Explicit canonical factorization

109

We start with Θ1 . Note that A1 (Im PΘ → Im PΘ ) and C1 (Im PΘ → Cm ) are the restrictions of A and C to D(A) ∩ Im PΘ , respectively. Since PΘ is the separating projection of Θ, this implies that Θ1 is a realization triple. From the deﬁnition of A1 it follows that −iA1 is the inﬁnitesimal generator of a strongly continuous left semigroup of negative exponential type. Thus the kernel function k1 = kΘ1 has its support in (−∞, 0] and W1 (λ) = I − k1 (λ) = I + C1 (λ − A1 )−1 is deﬁned and analytic on an open half plane of the type Im λ < −ω with ω strictly negative. × : Im PΘ → Im PΘ× be the operator deﬁned by Next, we consider Θ× 1 . Let J × (5.52). We know that J is invertible, mapping D(A) ∩ Im PΘ onto D(A) ∩ Im PΘ× . It is easy to check that J × provides a similarity between the operator A× 1 and the restriction of A× to D(A× ) ∩ Im PΘ× . Hence iA× 1 is the inﬁnitesimal generator of a strongly continuous left semigroup of negative exponential type. But then Theorem 5.7 guarantees that Θ× 1 is a realization triple. Furthermore , the kernel function k1× associated with Θ× 1 has its support in (−∞, 0], and −1 W1 (λ)−1 = I − k1× (λ) = I − C1 (λ − A× B1 1)

for all λ with Im λ < −ω. Here it is assumed that the negative constant ω has been taken suﬃciently close to zero. × We proceed by considering Θ2 and Θ× 2 . Obviously Θ2 is a realization triple, and a similarity argument of the type presented above yields that the same is true for Θ2 . The operators −iA2 and −iA× 2 are inﬁnitesimal generators of strongly continuous right semigroups of negative exponential type. Hence the kernel functions k2 and k2× associated with Θ2 and Θ× 2 , respectively, have their support in [0, ∞). Finally, taking |ω| smaller if necessary, we have that W2 (λ) = I − k2 (λ) = I + C2 (λ − A2 )−1 B2 and

−1 W2 (λ)−1 = I − k2× (λ) = I − C2 (λ − A× B2 2)

are deﬁned and analytic on Im λ > −ω. We may assume that both Θ and Θ× are of exponential type ω. For values of λ with |λ| < −ω one then has W (λ) = I + C1 (λ − A1 )B1−1 + C2 (λ − A2 )−1 B2 +C1 (λ − A1 )−1 Z(λ − A2 )−1 B2 . Now Ker PΘ× is an invariant subspace for ⎡ ⎤ Z − B1 C2 A× 1 ⎦, A× = ⎣ × −B2 C1 A2

110

Chapter 5. Factorization of matrix functions analytic in a strip

and so Z = B1 C2 . Substituting this in the above expression for W (λ), we get W (λ) = W1 (λ)W2 (λ). Clearly this is a canonical Wiener-Hopf factorization. The expressions obtained for the factors and their inverses are not quite the same as those given in the theorem. One veriﬁes without diﬃculty, however, that for λ in the intersection of ρ(A) and ρ(A× ), they amount to the same. For further information on this point we refer again (see the proof of Theorem 3.2) to Section 2.5 in [20] where the case when all three operators A, B and C are bounded is analyzed in great detail. Inspired by the terminology used in [20] (see also [11], Section 1.1), we introduce some additional terminology and notation. Let Θ = (A, B, C) be a realization triple with state space X, and let Π be a projection of X which maps D(A) into itself. We then have X

=

D(A)

=

˙ Im Π, Ker Π + ˙ D(A) ∩ Im Π , D(A) ∩ Ker Π +

and with respect to these decompositions the operators A, B and C have the form A11 A12 B1 A= , B= , C = C1 C2 . A21 A22 B2 The triple (A22 , B2 , C2 ) will be called the projection of Θ = (A, B, C) associated with Π, and it is denoted by prΠ (Θ). Note that (A11 , B1 , C1 ) is the projection prI−Π (Θ) of (A, B, C) associated with the projection I − Π. A particularly interesting case for what follows is when Π is a supporting projection for Θ. This means that besides the Π-invariance of D(A) = D(A× ) the following inclusion relations are satisﬁed: A D(A) ∩ Ker Π ⊂ Ker Π, A× D(A× ) ∩ Im Π ⊂ Im Π. In that situation we have A12 = B1 C2 and A21 = 0. Also Π is a supporting projection for the realization triple Θ = (A, B, C) if and only I − Π is a supporting projection for Θ× = (A× , B, −C). Finally, if Π is supporting for Θ, the arguments used in Part 3 of the proof of Theorem 5.14 show that prΠ (Θ) and prI−Π (Θ) are again realization triples. With this notation and terminology we have the following alternative version of Theorem 5.10. Theorem 5.15. Let T be the Wiener-Hopf integral operator on Lm 1 (R) given by (5.51). Assume that k = kΘ for some realization triple Θ = (A, B, C). Then T is invertible if and only if the following two conditions are satisﬁed: (i) Θ× = (A× , B, −C) is a realization triple, ˙ Ker PΘ× . (ii) X = Im PΘ +

5.9. The Riemann-Hilbert boundary value problem revisited (2)

111

Here X is the state space of both Θ and Θ× , and PΘ and PΘ× are the separating projections of −iA and −iA× , respectively. If (i) and (ii) hold, then the projection Π of X onto Ker PΘ× along Im PΘ is a supporting projection for Θ, the complementary projection I − Π is a supporting projection for Θ× , and ∞ −1 γ(t, s)g(s) ds. T g (t) = g(t) − 0

Here γ is given by s ⎧ × × × ⎪ ⎪ k (t − s) − k+ (t − r)k− (r − s) dr, ⎪ ⎨ + γ(t, s) =

s < t,

0

t ⎪ ⎪ ⎪ × × × ⎩ k− (t − s) − k+ (t − r)k− (r − s) dr,

s > t,

0

× × and k− are the kernel functions associated with the realization triples where k+ × prΠ (Θ ) and prI−Π (Θ× ), respectively.

5.9 The Riemann-Hilbert boundary value problem revisited (2) In this section we deal with the Riemann-Hilbert boundary value problem on the real line for matrix functions W of the form ∞ eiλt k(t) dt, (5.65) W (λ) = I − −∞

where k is an m× m matrix-valued function with the property that for some ω < 0 the entries of e−ω|t| k(t) are Lebesgue integrable on the real line. In this case W is analytic in a strip around the real axis. For such a function the Riemann-Hilbert problem consists of ﬁnding pairs Φ+ , Φ− of Cm -valued functions on the real line such that W (λ)Φ+ (λ) = Φ− (λ), −∞ < λ < ∞ (5.66) while, in addition, Φ+ and Φ− are Fourier transforms of integrable Cm -valued functions with support in [0, ∞) and (−∞, 0], respectively. These requirements imply that Φ+ and Φ− both vanish at inﬁnity and that they are continuous on the closed upper and closed lower half plane, respectively. From the special form of k in (5.65) we know that W is the transfer function of some realization triple Θ = (A, B, C). The following theorem gives the solution of the Riemann-Hilbert problem for W in terms of the operators A, B and C appearing in the triple. Theorem 5.16. Let W be the transfer function of realization triple Θ = (A, B, C). Assume Θ× = (A× , B, C) is a realization triple too (or, equivalently, that W (λ)

112

Chapter 5. Factorization of matrix functions analytic in a strip

is invertible for all λ ∈ R). Write PΘ and PΘ× for the separating projections of −iA and −iA× , respectively. Then the pair of functions Φ+ , Φ− is a solution of the Riemann-Hilbert boundary value problem (5.66) if and only if there exists x ∈ Im PΘ ∩ Ker PΘ× such that ∞ × −1 eiλt (Λ× (5.67) Φ+ (λ) = C(λ − A ) x = Θ x)(t) dt, 0

Φ− (λ)

=

−1

C(λ − A)

x =−

0

−∞

eiλt (ΛΘ x)(t) dt.

(5.68)

Moreover the vector x is uniquely determined by the functions Φ+ , Φ− . Proof. Take x ∈ Im PΘ ∩ Ker PΘ× . Condition (C4) in the deﬁnition of a realization triple implies that (Λ× Θ x)(t) is zero almost everywhere on the half line −∞ < t ≤ 0, while (ΛΘ x)(t) is zero almost everywhere on 0 ≤ t < ∞. It follows that we can apply (5.29) to both Θ and Θ× in order to show that ∞ ∞ × −1 eiλt (Λ× x)(t) dt = eiλt (Λ× x, Θ Θ x)(t) dt = C(λ − A ) 0

−∞

0

∞

eiλt (ΛΘ x)(t) dt

∞

= −∞

eiλt (ΛΘ x)(t) dt = −C(λ − A)−1 x.

Thus the functions Φ+ and Φ− in (5.67) and (5.68) are well-deﬁned. Furthermore, these functions are Fourier transforms of integrable Cm -valued functions with support in [0, ∞) and (−∞, 0], respectively. From (5.35) we see that (5.66) is satisﬁed. Thus the pair Φ+ , Φ− is a solution of the Riemann-Hilbert problem. To prove the reverse implication, assume that the pair Φ+ , Φ− is a solution of the Riemann-Hilbert problem (5.66). Write Φ+ and Φ− in the form ∞ 0 iλt e φ+ (t) dt, Φ− (λ) = eiλt φ− (t) dt, Φ+ (λ) = 0

−∞

Lm 1 [0, ∞)

Lm 1 (−∞, 0].

and φ− ∈ Now, let kΘ be the kernel function where φ+ ∈ associated with Θ, and consider the integral operator on Lm 1 [0, ∞) deﬁned by ∞ (Kf )(t) = kΘ (t − s)f (s) ds, a.e. on [0, ∞). 0

Using (5.66) and the fact that W (λ) = Im − kΘ (λ), a routine argument yields that φ+ − Kφ+ = 0. In other words, φ+ ∈ Ker (I − K). Next, we use the coupling relation (5.54) together with Corollary 4.3 in Section III.4 of [51]. It follows that φ+ = U × x for some x in Ker J × , where U × : Im PΘ → Lm 1 [0, ∞), J × : Im PΘ → Im PΘ× ,

(U × x)(t) = −(ΛΘ× x)(t), J × x = PΘ× x.

a.e. on [0, ∞),

5.9. The Riemann-Hilbert boundary value problem revisited (2)

113

Obviously, Ker J × = Im PΘ ∩ Ker PΘ× . Thus there exists x ∈ Im PΘ ∩ Ker PΘ× such that φ+ (t) = −(ΛΘ× x)(t) a.e. on the half line 0 ≤ t < ∞. But then (5.67) is satisﬁed. By (5.66), the identity (5.29) applied to Θ× = (A× , B, −C), and (5.35) we have Φ− (λ) = W (λ)Φ+ (λ) = C(λ − A)−1 x. Hence (5.68) holds too. It remains to prove the uniqueness of the vector x in (5.67) and (5.68). Assume that x is a second vector with the same properties as the vector x. So x ∈ Im PΘ ∩ Ker PΘ× while (5.67) and (5.68) hold true with x in place of x. Let J × and U × be as in the previous paragraph. Since Ker J × = Im PΘ ∩ Ker PΘ× , we have x − x ∈ Ker J × . Furthermore, the fact that the left-hand side of (5.67) does × not depend on x nor on x yields that (Λ× Θ x)(t) = (ΛΘ x )(t) a.e. on [0, ∞). Thus × × × × U x = U x . It follows that both J (x − x ) and U (x − x ) are equal to zero. If x = x , this implies that the operator deﬁned by the right-hand side of (5.54) is not invertible, which is impossible by Lemma 5.12. We conclude that x = x , as desired.

Notes The material presented in this chapter is taken from the papers [16] and [15]. In [16] the reader will also ﬁnd a systematic treatment of realization triples (A, B, C) with C bounded and A unbounded. The notion of an exponentially dichotomous operator, which has been introduced in [16], has proved to be quite useful in other areas. See, e.g., the papers [22] and [93]. The theory of realization triples is also used in [14] and [92]. The papers [90] and [91] present an extension of the theory of realization triples to operator-valued functions by introducing two-sided Pritchard-Salomon realizations. In particular, the factorization theory of Section 5.8 is developed further in [91]. For more information on exponentially dichotomous operators, including various perturbation theorems and a wide variety of applications, we refer to the monograph [111]. See also the notes to Chapter 6.

Chapter 6

Convolution equations and the transport equation In this chapter the factorization theory developed in the previous chapters is applied to solve a linear transport equation. It is known that the transport equation may be transformed into a Wiener-Hopf integral equation with an operator-valued kernel function (see [40]). An equation of the latter type can be solved explicitly if a canonical factorization of its symbol is available (cf., Sections 1.1 and 3.2). In our case the symbol may be represented as a transfer function, and to make the factorization the general factorization theorem of the second chapter can be applied. This requires that one ﬁnds an appropriate pair of invariant subspaces. In the case of the transport equation the choice of the subspaces is evident, but to prove that their direct sum is the whole space takes some eﬀort. The latter is related to a new diﬃculty that appears here. Namely, in this case the curve cuts through the spectra of the main operator and the associate main operator. Nevertheless, due to the special structure of the operators involved, the factorization can be made and explicit formulas are obtained. Since our main purpose is to show how our method works, we restrict ourselves to the case when the kernel function describing the eﬀect of the scattering is of ﬁnite rank. In Section 6.1 we describe the transport equation that is considered in this chapter. To illustrate our approach we ﬁrst study (in Section 6.2) a simpliﬁed model, namely when the scattering appears only in a ﬁnite number of directions. In Section 6.3 the vector-valued Wiener-Hopf equation associated to the transport equation is introduced. In Section 6.4 it is shown that under appropriate conditions a canonical factorization of the symbol associated with the equation can be constructed, and the matching of corresponding invariant subspaces is established in Section 6.5. In Section 6.6, the ﬁnal section of the chapter, we present formulas for the solution.

116

Chapter 6. Convolution equations and the transport equation

6.1 The transport equation Transport theory is a branch of mathematical physics concerned with the mathematical analysis of equations that describe the migration of particles in a medium, for instance, a ﬂow of electrons through a metal strip or radiative transfer in a stellar atmosphere. For the plane symmetric case, a stationary transport problem through a homogeneous medium can be modelled by an integro-diﬀerential equation of the following form: 1 ∂ψ(t, μ) k(μ, μ )ψ(t, μ ) dμ , t ≥ 0. (6.1) μ + ψ(t, μ) = ∂t −1 This equation is a balance equation. The unknown function ψ is a density function related to the expected number of particles in an inﬁnitesimal volume element. The right-hand side describes the eﬀect of the collisions. The variable μ is equal to cos α where α is the scattering angle, and therefore −1 ≤ μ ≤ 1. The variable t is not a time variable but a position variable, sometimes referred to as the optical depth. The kernel function k in the right-hand side of (6.1), which is called the scattering function, is assumed to be a real symmetric L1 -function on [−1, 1] × [−1, 1]. We shall consider the so-called half range problem, that is, we assume the medium to be semi-inﬁnite, and hence the position variable runs over the interval 0 ≤ t < ∞. Since the density of the incoming particles is known, the values of ψ(0, μ) are known for 0 < μ ≤ 1. Thus the above equation will be considered together with the boundary condition ψ(0, μ) = f+ (μ),

0 < μ ≤ 1,

(6.2)

where f+ is a given function on (0,1]. In the sequel we shall consider f+ as a function on [−1, 1] by setting f+ (μ) = 0 for −1 ≤ μ < 0, and we assume that f+ ∈ L2 [−1, 1]. There is also a boundary condition at inﬁnity, which appears in diﬀerent forms. Here we take the condition at inﬁnity to be t lim ψ(t, μ) exp = 0, −1 ≤ μ < 0. (6.3) t→∞ μ Thus the problem is to solve (6.1) under the boundary conditions (6.2) and (6.3). In this chapter we shall assume (cf., [81], [82] and [108]) that the scattering function k is given by n

aj pj (μ)pj (μ ), (6.4) k(μ, μ ) = j=0

where pj (μ) is the j-th normalized Legendre polynomial (see [53], page 26) and −∞ < aj < 1,

j = 0, 1, . . . , n.

(6.5)

6.1. The transport equation

117

In particular, the integral operator deﬁned by the right-hand side of (6.1) has ﬁnite rank. By writing ψ(t)(μ) = ψ(t, μ), we may consider the unknown function ψ as a vector function on [0, ∞) with values in H = L2 [−1, 1]. In this way equation (6.1) can be written as an operator diﬀerential equation: T

dψ (t) + ψ(t) = Kψ(t), dt

t ≥ 0,

(6.6)

where the derivative is taken with respect to the norm in H. In (6.6) the operators T and K are deﬁned by

T f (μ) = μf (μ),

Kf =

n

aj f, pj pj .

(6.7)

j=0

Because of (6.5), the operator I−K is strictly positive, and hence (6.6) is equivalent to dψ (I − K)−1 T = −ψ. dt In [81], [82], [108] this equation is solved by diagonalizing the operator (I − K)−1 T . Equation (6.1) with boundary conditions (6.2) and (6.3) can also be written as a Wiener-Hopf integral equation with an operator-valued kernel function (cf., [40]). In order to do this, let us introduce some notation. Let H+ and H− be the subspaces of H = L2 [−1, 1] consisting of all functions that are zero almost everywhere on [−1, 0] and [0, 1], respectively. By P+ and P− we denote the orthogonal projections of H = L2 [−1, 1] onto the subspace H+ and H− , respectively. Furthermore, h will be the operator-valued function deﬁned by ⎧ 1 t ⎪ ⎪ t > 0, exp − (P+ Kf )(μ), ⎪ ⎪ μ ⎨ μ (6.8) h(t)f (μ) = ⎪ ⎪ 1 t ⎪ ⎪ ⎩ − exp − t < 0, (P− Kf )(μ), μ μ and F is the vector-valued function given by ⎧ t ⎪ ⎪ , f ⎨ + (μ)exp − μ F (t)(μ) = ⎪ ⎪ ⎩ 0,

0 < μ ≤ 1, (6.9) −1 ≤ μ ≤ 0.

The operator-valued function h is referred to as the propagator function associated with the half range problem (6.6).

118

Chapter 6. Convolution equations and the transport equation

Given these functions h and F , equation (6.1) with the boundary conditions (6.2) and (6.3) can be written as ∞ ψ(t) − h(t − s)ψ(s) ds = F (t), t ≥ 0. (6.10) 0

To see this, multiply equation(6.1) by μ−1 exp(t/μ) and integrate over (0, t) when μ > 0 or over (t, ∞) in case μ < 0. With the help of the boundary conditions one gets in this way the integral equation (6.10). In [40] the asymptotics of solutions of equation (6.10) are found and used to describe the asymptotics of solutions of the transport equation.

6.2 The case of a ﬁnite number of scattering directions To make the method used in this chapter more transparent we ﬁrst consider the case when scattering occurs in a ﬁnite number of directions only. This assumption reduces the equation (6.1) and the boundary condition (6.2) to μi

n

dψ k(μi , μj )ψ(t, μj ), (t, μi ) + ψ(t, μi ) = dt j=1

(6.11)

i = 1, . . . , n, t ≥ 0, where ψ(0, μi ) = ϕ+ (μi ),

μi > 0.

(6.12)

n

To treat this version of the problem, introduce the C -valued function ⎡ ⎤ ψ(t, μ1 ) ⎢ ⎥ .. ψ(t) = ⎣ ⎦, . ψ(t, μn ) and the matrices T = diag [μ1 , . . . , μn ],

K = [k(μi , μj )]ni,j =1 .

(6.13)

Observe that T and K are real symmetric (hence selfadjoint) n × n matrices. Using this notation, equation (6.11) taken with the boundary condition (6.12) can be rewritten as 0 ≤ t < ∞, T ψ (t) + ψ(t) = Kψ(t), (6.14) P+ ψ(0) = x+ , where P+ is the projection on ⎡ ⎤ ⎡ x1 ⎢ ⎥ ⎢ P+ ⎣ ... ⎦ = ⎣ xn

Cn deﬁned by ⎤ y1 0 .. ⎥ , y = ⎦ i . x1 yn

if μi ≤ 0, if μi > 0,

6.2. The case of a ﬁnite number of scattering directions

119

and x+ is a given vector in Im P+ . Observe that P+ is the spectral projection of T corresponding to the positive eigenvalues of T . In what follows we assume additionally that T is invertible, which is the generic case and corresponds to the requirement that all μi in (6.11) are diﬀerent from 0; cf., formula (6.12). We shall look for solutions ψ of (6.14) in the space Ln2 [0, ∞). The ﬁrst step in solving (6.14) is based on the observation that, for invertible T , equation (6.14) is equivalent to a Wiener-Hopf integral equation with a rational matrix symbol. In fact, the following theorem holds. Theorem 6.1. Suppose T in (6.13) is invertible and let ψ ∈ Ln2 [0, ∞). Then ψ is a solution of equation (6.14) if and only if ψ is a solution of the Wiener-Hopf integral equation with a special right-hand side, namely ψ(t) −

0

∞

h(t − s)Kψ(s) ds = e−tT

−1

x+ ,

t ≥ 0,

(6.15)

where h is the propagator function associated with problem (6.14), that is,

h(t) =

⎧ −1 ⎨ T −1 e−tT P+ ,

t > 0,

⎩

t < 0.

−T −1 e−tT

−1

(6.16) P− ,

Here P− = I − P+ . Proof. Assume ψ is a solution of (6.14). Applying T −1 to the ﬁrst identity in (6.14), and solving the resulting equation by using variation of constants, yields ψ(t) = e−tT

−1

ψ(0) + e−tT

−1

t

esT

−1

T −1 Kψ(s)ds,

t ≥ 0.

0 −1

(6.17)

−1

Next, apply etT P− to both sides of (6.17) and use that etT and P− commute. −1 −1 Since etT P− is exponentially decaying on [0, ∞), the function etT P− Kψ(t) is integrable on [0, ∞), and thus lim e

t→∞

tT −1

P− ψ(t) = P− ψ(0) +

∞ 0

esT

−1

P− T −1 Kψ(s)ds.

(6.18)

−1

Again using that the function etT P− ψ(t) is integrable on [0, ∞), we see that the left-hand side of (6.18) has to be equal to zero, which proves that P− ψ(0) = −

0

∞

esT

−1

P− T −1 Kψ(s)ds.

(6.19)

Now, replace ψ(0) in (6.17) by P+ ψ(0) + P− ψ(0), use the boundary condition in

120

Chapter 6. Convolution equations and the transport equation

(6.14), and apply (6.19). This gives ∞ −1 −tT −1 ψ(t) = e x+ − e−(t−s)T P− T −1 Kψ(s)ds t t −1 e−(t−s)T T −1 Kψ(s)ds + 0

= e−tT

−1

x+ +

∞

0

h(t − s)Kψ(s)ds,

t ≥ 0.

Thus ψ is a solution of (6.15). To prove the converse statement, assume that ψ is a solution of (6.15). Thus ψ(t)

=

e−tT

−1

x+ + e−tT

−e−tT

−1

−1

∞

t

esT

0

esT

−1

t

−1

P+ T −1Kψ(s)ds

P− T −1 Kψ(s)ds,

(6.20) t ≥ 0.

It follows that ψ is absolutely continuous on each compact subinterval of [0, ∞), and hence the integrands in the right-hand side of (6.20) are continuous functions of the variable s. But then ψ is diﬀerentiable on (0, ∞), and we see that ψ (t)

=

−T −1ψ(t) + P+ T −1 Kψ(t) + P− T −1Kψ(t)

=

−T −1ψ(t) + T −1 Kψ(t),

t ≥ 0,

and hence ψ satisﬁes the ﬁrst equation in (6.14). From (6.20) it also follows that ∞ −1 esT P− T −1 Kψ(s)ds, t ≥ 0, ψ(0) = x+ − 0

which implies that P+ ψ(0) = P+ x+ = x+ . We conclude that ψ is a solution of the problem (6.14). A direct computation yields that the symbol of the Wiener-Hopf operator associated with (6.15) is the n × n matrix function W given by W (λ) = In − iT −1(λ + iT −1)−1 K, where In is the n × n identity matrix. Thus the symbol W is not only a rational matrix function but it is already given in a concrete realized form, namely W (λ) = In + C(λ − A)−1 B, with

A = −iT −1,

B = K,

C = −iT −1.

(6.21)

6.2. The case of a ﬁnite number of scattering directions

121

Notice that A does not have eigenvalues on the real line. Thus in order to solve equation (6.15) we can apply Theorem 3.3. This requires us to analyze the spectral properties of the matrix A× = A − BC = −iT −1(I − K).

(6.22)

In view of (6.5) it is natural to assume I − K is positive deﬁnite. Lemma 6.2. Assume I − K is positive deﬁnite. Then the matrix A× in (6.22) has no real eigenvalues and ˙ M ×, (6.23) Cn = M + where M is the spectral subspace of the matrix A in (6.21) corresponding to the eigenvalues in the upper half plane, and M × is the spectral subspace of A× in (6.22) corresponding to the eigenvalues in the lower half plane. Proof. Let · , · be the standard inner product in Cm and put S = (I − K)−1 T . Since I − K is positive deﬁnite, S is well-deﬁned and the sesquilinear form [x, y] = (I − K)x, y

(6.24)

is an inner product on Cn . From [Sx, y] = (I − K)Sx, y = T x, y and the fact that T is selfadjoint, it follows that S is selfadjoint with respect to the inner product [· , ·]. But then the same holds true for iA× = S −1. Thus A× is invertible and its eigenvalues are on the imaginary axis. In particular, A× has no real eigenvalues. Recall that P+ is the spectral projection of T corresponding to the positive eigenvalues of T . Let P+× be the analogous projection for S. Since T and S are invertible, T |Ker P+ is negative deﬁnite and S|Im P × is positive deﬁnite. Thus +

0 = x ∈ Ker P+

=⇒ T x, x < 0,

0 = x ∈ Im P+×

=⇒

[Sx, x] > 0.

But [Sx, x] = T x, x for each x ∈ Cn . It follows that Ker P+ ∩ Im P+× = {0}. In particular, rank P+ ≥ rank P+× . By repeating the argument with Ker P+ replaced by Ker P+× and Im P+× by Im P+ , we see that rank P+× ≥ rank P+ . But then we ˙ Im P+× . Finally, from iA = T −1 we see that may conclude that Cn = Ker P+ + × −1 M = Ker P+ , and from iA = S we conclude that M × = Im P+× . We can now apply Theorem 3.3 to solve equation (6.15). Note however that the right-hand side of (6.15) is of a special form. In fact, in terms of the matrices appearing in (6.21) this right-hand side can be written as g(t) = iCe−itA x+ , where x+ ∈ Im P+ . Thus instead of Theorem 3.3 we can also directly apply Corollary 3.4. This yields the following result.

122

Chapter 6. Convolution equations and the transport equation

Theorem 6.3. Assume I − K is positive deﬁnite and T is invertible. Then the matrix (I − K)−1 T is selfadjoint with respect to the inner product (6.24) and the half range problem (6.14) has a unique solution ψ in Ln2 (0, ∞), namely ψ(t) = e−tT

−1

(I−K)

Πx+ ,

t ≥ 0,

(6.25)

where Π is the projection of Cn along Ker P+ onto the spectral subspace Im P+× of (I − K)−1 T corresponding to its positive eigenvalues.

6.3 Wiener-Hopf equations with operator-valued kernel functions It is well-known that the Wiener-Hopf integral equation ∞ ψ(t) − k(t − s)ψ(s) dy = F (t), t≥0

(6.26)

0

can be solved by constructing appropriate factorizations of its symbol (cf., Sections 1.1, 3.2, the papers [49], [71], or the survey article [59]). In this section we shall describe this method for the case when k is an L1 -kernel function the values of which are compact operators on a separable Hilbert space H. So we assume that k(t) is a compact operator for each real t, that k(·)f, g is measurable on the real line for each f and g in H, and that ∞ k(t) dt < 0, −∞

where ·, · is the inner product on H, and · is the operator norm for operators on H. Note that the kernel function h considered in the previous section falls into this category. Recall that the symbol of equation (6.26) is the operator-valued function I − K(λ), where K(λ) is the Fourier transform of the kernel function k, i.e., ∞ K(λ) = eiλt k(t) dt, −∞ < λ < ∞. −∞

By the Riemann-Lebesgue lemma, we have limλ∈R, λ→±∞ K(λ) = 0. Here we also need the concept of canonical factorization, this time in the present inﬁnite dimensional context. The symbol is said to admit a (right ) canonical factorization with respect to the real line if I − K(λ) = G− (λ)G+ (λ),

−∞ < λ < ∞,

where the factors G− and G+ meet the following requirements:

(6.27)

6.3. Wiener-Hopf equations with operator-valued kernel functions

123

(i) the operator function G− is analytic on the (open) lower half plane λ < 0 and continuous on the closure of the left half plane in the Riemann sphere (inﬁnity included); also for each λ in this closure (inﬁnity included), the operator G− (λ) is invertible; (ii) the operator function G+ is analytic on the (open) upper half plane λ > 0 and continuous on the closure of the right half plane in the Riemann sphere (inﬁnity included); also for each λ in this closure (inﬁnity included), the operator G+ (λ) is invertible. Note that the deﬁnition is analogous to that given earlier in the matrix-valued case (see Sections 1.1 and 3.1). According to [49], because of the fact that k is an L1 -kernel function the values of which are compact operators on H, the inverses of the factors in the right-hand side of (6.27) can be written as ∞ 0 −1 −1 iλt e γ+ (t), G− (λ) = I + eiλt γ− (t) dt, (6.28) G+ (λ) = I + 0

−∞

where, γ+ and γ− are L1 -functions on [0, ∞) and (−∞, 0], respectively, whose values are compact operators on H. Let L2 (R+ , H) denote the space of all L2 -integrable functions on [0, ∞) with values in H. The identities (6.28) are important, because they allow for explicit formulas for the solutions of (6.26). Indeed, by [18] equation (6.26) has a unique solution ψ in L2 (R+ , H) for each F ∈ L2 (R+ , H) if and only if a canonical factorization (6.27) exists, and in that case (just as in Section 1.1 for matrix-valued kernel functions) the Fourier transform ψ of the solution ψ is given by −1 (6.29) ψ(λ) = G−1 + (λ)P G− (λ)F (λ) , where F is the Fourier transform of the right-hand side of equation (6.26), and P is the projection deﬁned by ∞ ∞ P f (t)eitλ dt = f (t)eitλ dt. −∞

0

Taking inverse Fourier transforms in (6.29) one ﬁnds ∞ ψ(t) = F (t) + γ(t, s)F (s) ds, 0

where γ(t, s) is given by (1.10), i.e., s ⎧ ⎪ ⎪ (t − s) + γ+ (t − r)γ− (r − s) dr, γ + ⎪ ⎨ 0 γ(t, s) = t ⎪ ⎪ ⎪ ⎩ γ (t − s) + γ (t − r)γ (r − s) dr, −

0

+

−

0 ≤ s < t, 0 ≤ t < s.

124

Chapter 6. Convolution equations and the transport equation

As we observed already, in (6.10) the kernel function h(·) is an L1 -function on the real line whose values are compact (in fact ﬁnite rank) operators on L2 [−1, 1]. In the next section we shall prove that the corresponding symbol admits a canonical factorization, and we shall describe the factors explicitly.

6.4 Construction of a canonical factorization We now return to equation (6.10). Note that its symbol is given by I − H(λ), where ∞ eiλt h(t)K dt = (I − iλT )−1 K, −∞ < λ < ∞. H(λ) = −∞

Here h is given by (6.8), and the operators T and K are as in formula (6.7). The operator function H is analytic on the strip |λ| < 1. Note that σ(T ) is the closed interval [−1, 1], so that (I − iλT )−1 is deﬁned for all λ in the complement of the union of the subsets i[1, ∞) and i(−∞, −1] of the imaginary axis. In this section we show that I − H(λ) admits a canonical factorization with respect to the real line: I − H(λ) = G− (λ)G+ (λ), −∞ < λ < ∞, where the factors and their inverses can be written as G− (λ)

= I − (I − iλT )−1(I − P )K(I − P K)−1 ,

G+ (λ) G−1 − (λ)

= I − (I − Q∗ K)−1 (I − Q∗ )(I − iλT )−1K, −1 = I + (I − Q∗ K)−1 Q∗ I − iλ(T × )∗ K,

G−1 + (λ)

= I + (I − iλT × )−1 P K(I − P K)−1 .

Here T × = (I − K)−1 T , and P and Q are projections of which the deﬁnition will be given below. With regard to the domains of the factors and their inverses, the situation is similar to what we encountered earlier for Theorems 3.2 and 5.14. In order to make the factorization we transform the symbol of equation (6.10) into another function W which is deﬁned and continuous on the imaginary axis. This will be done as follows. Recall that T and K are both selfadjoint and that I −K, being a strictly positive operator because of (6.5), is invertible. Hence, for non-zero purely imaginary values of λ, ∗ I − H(i/λ)∗ = I − (I + λ−1 T )−1 K = I − K(I − λ−1 T )−1 = I − λK(λ − T )−1 = (I − K) I − K(I − K)−1 T (λ − T )−1 .

6.4. Construction of a canonical factorization

125

We now introduce W by writing W (λ) = I − (I − K)−1 KT (λ − T )−1 .

(6.30)

Note that this expression is a unital realization for W . The state space is the separable Hilbert space H = L2 [−1, 1]. The operator T is the main operator, and (I − K)−1 T is the associate main operator, denoted above by T × (in line with the notation adopted in Section 2.1). Via (6.30) the function W is deﬁned and analytic on the resolvent set of T , so on the complement of the interval [−1, 1]. In particular, W is deﬁned and continuous on the imaginary axis punctured at the origin. We shall now prove that by setting W (0) = (I −K)−1 the restriction of W (now deﬁned on the complement of the set [−1, 0) ∪ (0, 1]) to the imaginary axis is a continuous function. For this we need to show that lim

α→ 0, α ∈ R

W (iα) = (I − K)−1 .

(6.31)

It is convenient to establish the following lemma which will also play a role later on in this section. Lemma 6.4. Let S be a bounded selfadjoint operator on a given Hilbert space. Then S(iα − S)−1 ≤ 1,

0 = α ∈ R,

(6.32)

while, furthermore, lim

S(iα − S)−1 f = −f,

f ⊥ Ker S.

α→ 0, α ∈ R

(6.33)

Under the additional assumption that S is a nonnegative operator, the limit result (6.33) can be sharpened to lim

λ→ 0, λ ≤ 0

S(λ − S)−1 f = −f,

f ⊥ Ker S.

(6.34)

Proof. Let ES (t) be the spectral resolution of the identity for S, and let f be an element of the underlying Hilbert space. Then ∞ t2 S(iα − S)−1 f 2 ≤ d ES (t)f 2 2 2 α + t −∞ ∞ ≤ d ES (t)f 2 −∞

=

f 2.

This proves (6.32). Next, observe that f + S(iα − S)−1 f 2 ≤

∞ −∞

α2 d ES (t)f 2 . + t2

α2

126

Chapter 6. Convolution equations and the transport equation

So by Lebesgue’s dominated convergence theorem we get lim

α→ 0, α ∈ R

f + S(iα − S)−1 f ≤ =

ES (0+)f 2 − ES (0−)f 2 (ES (0+) − ES (0−))f 2 ,

which is zero if f ⊥ Ker S. Hence (6.33) is proved. The argument for (6.34), taking nonnegativity of the operator S for granted, is analogous. The proof of (6.31) is now as follows. As Ker T = {0}, we see from Lemma 6.4 that limα→ 0, α ∈ R (iα + T )−1 T f = f, f ∈ H. Since K is compact (actually even of ﬁnite rank), it follows that (iα + T )−1T K tends to K in the operator norm if α ∈ R, α → 0. Taking adjoints, we obtain that the same holds true for −KT (iα − T )−1 . But then we have (6.31), where the convergence is with respect to the operator norm. So with W (0) = (I − K)−1 , indeed W becomes a continuous function on the imaginary axis. It is this operator function for which we want a (right) canonical Wiener-Hopf factorization. This time not with respect to the real line (see the deﬁnition in Section 6.3) but for the analogous situation where the curve in the Riemann sphere is the imaginary axis with inﬁnity included. The theory concerning canonical factorization developed earlier suggests that we have to ﬁnd an invariant subspace M for T such that the spectrum of T restricted to M lies in the closed right half plane, and an invariant subspace M × for T × such that the spectrum of T × restricted to M × lies in the closed left half plane. Since T is selfadjoint the choice of M is clear: M = H+ , where H+ is the subspace of H = L2 [−1, 1] consisting of all functions that are zero almost everywhere on [−1, 0]. As we shall see below, after replacing the standard inner product on L2 [−1, 1] by a suitable equivalent one, the operator T × is selfadjoint too. So for M × we can take the spectral subspace of T × corresponding to the part of the spectrum of T × on (−∞, 0]. The ﬁrst diﬃculty is to prove the matching of the subspaces M and M × , i.e., to show ˙ M × . Taking for granted that this has been established a second that H = M + diﬃculty appears, because in the present case the imaginary axis does not split the spectra of T and T × . So we cannot apply directly the theory developed so far, but we have to prove, using the speciﬁcs of the situation, that the factors obtained have the desired boundary behavior. The purpose of this section is to show that this approach works indeed. We begin by considering the operator T × = (I − K)−1 T . As I − K is strictly positive, [f, g] = (I −K)f, g deﬁnes an inner product on H = L2 [−1, 1] equivalent with the standard one. Writing A[∗] for the adjoint of an operator A relative to the inner product [· , ·], we have A[∗] = (I − K)−1 A∗ (I − K). ×

(6.35)

In particular, we see that the operator T is selfadjoint with respect to the inner product [·, ·]. Let E × (·) be the corresponding spectral resolution and introduce Hm = Im E × (0),

Hp = Ker E × (0).

6.4. Construction of a canonical factorization Then Hm and Hp are both invariant under T × and σ T × |Hp ⊂ [0, ∞) ∩ σ(T × ). σ T × |Hm ⊂ (−∞, 0] ∩ σ(T × ),

127

(6.36)

For T the situation is more straightforward. Indeed, T is selfadjoint with respect to the original (standard) inner product on H and leaves invariant the spaces H− and H+ featured in Section 6.1. Further σ T |H+ = [0, 1]. (6.37) σ T |H− = [−1, 0], The subspaces M and M × mentioned above are H+ and Hm , respectively. ˙ Hm . So proving that these subspaces match amounts to showing that H = H+ + In fact, in the next section we shall show the following stronger result: ˙ Hp , H = H− +

˙ Hm . H = H+ +

(6.38)

Let P be the projection of H along H− onto Hp , and let Q be the projection of H along H+ onto Hm . Since the subspaces H− , H+ are invariant under T and Hm , Hp are invariant under T × , both P and Q are supporting projections for the realization (6.30). Associated with these projections are two factorizations: !− (λ), !+ (λ)W W (λ) = W

W (λ) = W− (λ)W+ (λ).

(6.39)

With the appropriate choice for the value of the factors at the origin, both these factorizations are canonical factorizations of W with respect to the imaginary axis; the ﬁrst a left and the second a right factorization. In the sequel we only need the second factorization in (6.39). First we give the expressions for the factors W− (λ) and W+ (λ): W− (λ)

=

−1 I − (I − K)−1 KT |H+ λ − T |H+ (I − Q),

(6.40)

W+ (λ)

=

−1 I − (I − K)−1 KT |Hm λ − QT |Hm Q.

(6.41)

Note that there is slight abuse of notation here. Indeed, the operator I − Q in the formula for W− (λ) should be interpreted as a mapping from H onto H+ , and Q in the expression for W+ (λ) must be seen as a mapping from from H onto Hm . In particular QT |Hm should be read as the compression of T to Hm (relative to the ˙ Hm ). The function W− is deﬁned and analytic on the decomposition H = H+ + resolvent set of T |H+ so, by the second part of (6.37) on the complement of the interval [0, 1]. Similarly, the function W+ is deﬁned and analytic on the resolvent ˙ H− = H+ + ˙ Hm , and set of the compression operator QT |Hm . Now H = H+ + Lemma 3.1 guarantees that QT |Hm is similar to T |H− . In particular the resolvent sets of QT |Hm and T |H− coincide. It follows from the ﬁrst part of (6.37) that function W+ is deﬁned and analytic on the complement of the interval [−1, 0]. The argument also indicates that the second factorization in (6.39) holds for all

128

Chapter 6. Convolution equations and the transport equation

λ outside the interval [−1, 1]. Indeed this interval is precisely the union of the spectra of T |H+ and QT |Hm (cf., Theorem 2.6). Next we deal with the invertibility of the factors W− (λ) and W+ (λ). The above realization of W− has (I − Q)T × |H+ : H+ → H+ as its associate main operator. From Section 2.4 we now know that W− (λ) is invertible for λ in the intersection of the resolvent sets of T |H+ and (I − Q)T × |H+ . The ˙ Hm = resolvent set of T |H+ is the complement of the interval [0, 1]. As H = H+ + × × ˙ Hp + Hm , the compression operator (I − Q)T |H+ is similar to T |Hp . Hence, by the second part of (6.36), the resolvent set of (I − Q)T × |H+ is the complement of the set [0, ∞) ∩ σ(T × ). It follows that W− (λ) is invertible for all non-zero λ with λ ≤ 0, its inverse (see Theorem 2.4) being given by W− (λ)−1

=

−1 I + (I − K)−1 KT |H+ λ − (I − Q)T × |H+ (I − Q)

=

−1 I + KT × |H+ λ − (I − Q)T × |H+ (I − Q).

In an analogous manner one proves that W+ (λ) is invertible for all non-zero λ with λ ≥ 0, its inverse having the representation −1 Q W+ (λ)−1 = I + (I − K)−1 KT |Hm λ − T |× Hm −1 Q. I + KT × |Hm λ − T |× Hm

=

The above formulas contain the precise description of the factors W− , W+ and their inverses W−−1 , W+−1 . Giving up some precision but gaining in conciseness, we can also write W− (λ)

=

I − (I − K)−1 KT (λ − T )−1(I − Q),

W+ (λ)

=

I − (I − K)−1 KT Q(λ − T )−1 ,

W−−1 (λ)

=

I + KT ×(I − Q)(λ − T × )−1 ,

W+−1 (λ)

=

I + KT ×(λ − T × )−1 Q;

see Section 2.4 and [20], Section 2.5 for details. We have come close to proving that the second factorization in (6.39) is a (right) canonical factorization of W with respect to the imaginary axis. To make the proof complete we need to check the behavior of the functions at inﬁnity and at the origin. As far as the behavior at inﬁnity is concerned the situation is simple. Indeed the functions W− , W+ , W−−1 and W+−1 are analytic there with value the identity operator on H. For the origin the situation is more complicated.

6.4. Construction of a canonical factorization

129

Earlier we completed the deﬁnition of the function W , initially introduced via the unital realization (6.30), by stipulating that W (0) = (I − K)−1 . Now we make a similar move with respect to W− and W+ , in the ﬁrst instance given by (6.40) and (6.41), respectively. Indeed, we stipulate that W− (0) = (I − K)−1 (I − KQ),

W+ (0) = (I − K)−1 (I − KP ∗ ).

In this manner the closed left half plane λ ≤ 0 is contained in the domain of W− , and the closed right half plane λ ≥ 0 is contained in the domain of W+ . Our task is now threefold: to verify the invertibility of W− (0) and W+ (0), to demonstrate the continuity of W− and W+ on the appropriate half planes, i.e., to show that lim

W− (λ)

= (I − K)−1 (I − KQ),

(6.42)

lim

W+ (λ)

= (I − K)−1 (I − KP ∗ ),

(6.43)

λ→ 0, λ ≤ 0

λ→ 0, λ ≥ 0

and to verify that the factorization W = W− W+ holds at the origin. As a ﬁrst step we present the following lemma (which will also be used in Section 6.6 below). Lemma 6.5. Let P, Q and K be as above. Then Q∗ (I − K)P = 0,

(I − Q∗ )(I − P ) = 0.

(6.44)

Proof. Note that Im P = Hp is orthogonal to Im Q = Hm with respect to the inner product [f, g] = (I − K)f, g. Thus (I − K)P f, Qg = [P f, Qg] = 0,

f, g ∈ L2 [−1, 1].

This yields the ﬁrst identity in (6.44). Next observe that, relative to the usual inner product on H = L2 [−1, 1], the space Im (I − Q) = H+ is orthogonal to Im (I − P ) = H− . It follows that (I − P )f, (I − Q)g = 0, which proves the second identity in (6.44).

f, g ∈ L2 [−1, 1],

Corollary 6.6. The operators I − KQ and (I − K)−1 (I − KP ∗ ) are invertible and each other’s inverse. Proof. As K is compact (actually even of ﬁnite rank), the operator I − KQ is Fredholm of index zero. In particular I − KQ is invertible if and only if I − KQ is left invertible. Thus it suﬃces to show that the operator (I − K)−1 (I − KP ∗ ) is a left inverse of I − KQ. Now the identities in Lemma 6.5 can be rewritten as

130

Chapter 6. Convolution equations and the transport equation

Q∗ KP = Q∗ P and Q∗ + P − Q∗ P = I. Combining these, one gets I −K

=

I − (Q∗ + P − Q∗ P )K

=

I − Q∗ K − P K + Q ∗ P K

=

I − Q∗ K − P K + Q∗ KP K

=

(I − Q∗ K)(I − P K).

Taking adjoints yields I − K = (I − KP ∗ )(I − KQ), and this identity can be rewritten as (I − K)−1 (I − KP ∗ )(I − KQ) = I. The corollary can be rephrased by saying that W+ (0) = (I − K)−1 (I − KP ∗ ) is invertible with inverse I−KQ. Likewise W− (0) = (I−K)−1 (I−KQ) is invertible with inverse (I − K)−1 (I − KP ∗ )(I − K). It remains to verify (6.42) and (6.43). We begin with (6.42). For λ ≤ 0, λ = 0, we have W− (λ)

=

−1 I − K(I − K)−1 T |H+ λ − T |H+ (I − Q),

=

−1 I − K+ T |H+ λ − T |H+ (I − Q).

Here K+ is the restriction of K(I − K)−1 to H+ considered as an operator from H+ into H and, as before, I − Q should be read as a mapping from H onto H+ . The restriction operator T |H+ : H+ → H+ is selfadjoint and nonnegative. It also has a trivial null space. So we can apply Lemma 6.4 to show that lim

λ → 0, λ ≤ 0

−1 T |H+ λ − T |H+ f+ = −f+ ,

f+ ∈ H+ .

∗ Along with K+ , the operator K+ : H → H+ is compact (actually even of ﬁnite rank), and it follows that

lim

λ → 0, λ ≤ 0

−1 ∗ ∗ T |H+ λ − T |H+ K+ = −K+ ,

with convergence in norm. Taking adjoints we get lim

λ → 0, λ ≤ 0

−1 K+ T |H+ λ − T |H+ = −K+ ,

and hence lim

λ → 0, λ ≤ 0

−1 K+ T |H+ λ − T |H+ (I − Q) = −K+ (I − Q).

A simple computation gives I + K+(I − Q) = (I − K)−1 (I − KQ), and (6.42) is immediate.

6.4. Construction of a canonical factorization

131

Next we turn to (6.43). By Corollary 6.6 and the continuity of the operation of taking the inverse, it suﬃces to show that lim

λ → 0, λ ≥ 0

For λ ≥ 0, λ = 0, we have W+ (λ)−1

W+ (λ)−1 = I − KQ.

=

−1 I + KT × |Hm λ − T |× Hm Q,

=

−1 I + Km T × |Hm λ − T × |Hm Q.

Here Km is the restriction of K to Hm considered as an operator from Hm into H and, as before, Q should be read as a mapping from H onto Hm . Because T × = (I − K)−1 T , the operator T × |Hm has a trivial null space. Further it is nonpositive with respect to the alternative inner product [· , ·], and Km is compact. Using Lemma 6.4 in an analogous way as in the previous paragraph, we see that −1 lim Km T × |Hm λ − T × |Hm = −Km , λ → 0, λ ≥ 0

and we get lim λ → 0, λ ≥ 0 W+ (λ)−1 = I − KQ, as desired. From what we have obtained so far and a continuity argument it is already clear that the second factorization in (6.39) holds at the origin too. The calculation W− (0)W+ (0)

= (I − K)−1 (I − KQ)(I − K)−1 (I − KP ∗ ) = (I − K)−1 (I − KP ∗ )−1 I − K) (I − K)−1(I − KP ∗ ) = (I − K)−1 = W (0),

based on Corollary 6.6, corroborates this fact. Our ultimate goal in this section is to produce a right canonical factorization with respect to the real line of the symbol I −H(λ) of equation (6.10). For non-zero real λ we have I − H(λ) = W (i/λ)∗ (I − K), and with the right interpretation this identity even holds on the extended real line. Indeed, as W is given by a unital realization, the value of W at ∞ is I, and this corresponds with the fact that H(0) = K. Also by the Riemann-Lebesgue lemma, H vanishes at ∞, and this is in accord with W (0) = (I − K)−1 . The right canonical Wiener-Hopf factorization W = W− W+ with respect to the imaginary axis that we obtained for W now induces a right canonical Wiener-Hopf factorization with respect to the real line for the symbol. The details are given in the next two paragraphs. We begin by deﬁning a function G− on the complement in C∞ of the interval i[1, ∞) which is situated on the imaginary axis. The determining expressions are G− (λ)

= W+ (i/λ)∗ (I − Q∗ K),

G− (0) = I − Q∗ K, G− (∞)

= I.

132

Chapter 6. Convolution equations and the transport equation

Note that G− is analytic on the complement of i[1, ∞) in the ﬁnite complex plane C. Also G− is continuous on the closed lower half plane λ ≤ 0, this time inﬁnity included. Indeed, lim

λ → ∞, λ ≤ 0

G− (λ)

= =

lim

μ → 0, μ ≥ 0

W+ (μ)∗ (I − Q∗ K)

(I − P K)(I − K)−1 (I − Q∗ K) = I = G− (0).

Here we used (6.43). Further deﬁne G+ on the complement in C∞ of the interval i(−∞, −1], again located on the imaginary axis, by G+ (λ)

= (I − Q∗ K)−1 W− (i/λ)∗ (I − K),

G+ (0) = I − P K, G+ (∞)

= I.

Then G+ is analytic on the complement of i(−∞, −1]) in C. Also G+ is continuous on the closed upper plane λ ≥ 0, inﬁnity included. Indeed, using (6.42) one gets lim

λ → ∞, λ ≥ 0

G+ (λ) =

lim

(I − Q∗ K)−1 W− (μ)∗ (I − K) = I = G+ (0).

μ → 0, μ ≤ 0

Observe that I − H(λ) = G− (λ)G+ (λ), λ ∈ R. For non-zero λ this is clear from the corresponding factorization for W ; for λ = 0 we have G− (0)G+ (0) = (I − Q∗ K)(I − P K) = I − K = I − H(0). From what we saw in the preceding paragraph it is now clear that we have arrived at a right canonical factorization with respect to the real line, of the symbol I − H(λ). Explicit formulas for the −1 factors G− , G+ and their inverses G−1 − , G+ can be obtained from the descriptions −1 −1 of W+ , W− , W+ and W− given earlier in this section. In fact the formulas in question coincide with the ones already presented in the ﬁrst paragraph of this section. For the veriﬁcation of this we need the following intertwining result. Lemma 6.7. Let P and Q be as above. Then (I − Q∗ )T = T P . Proof. It is suﬃcient to establish the identities (I − Q∗ )T (I − P ) = 0 and Q∗ T P = 0. For the ﬁrst of these we argue as follows. Clearly (I − Q∗ )T (I − P )f, g = T (I − P )f, (I − Q)g. Now (I − P )f ∈ H− and (I − Q)g ∈ H+ . As H− is T -invariant we also have T (I − P )f ∈ H− . But H− ⊥ H+ . So (I − Q∗ )T (I − P )f, g = 0 for all f and g in H. It follows that (I − Q∗ )T (I − P ) = 0, as desired. Next observe that Q∗ T P f, g = T P f, Qg = [(I − K)−1 T P f, Qg] = [T × P f, Qg]. As P f ∈ Hp and Hp is invariant under T × , we have T × P f ∈ Hp . Furthermore Qg ∈ Hm . But Hm ⊥ Hp is H endowed with the inner product [·, ·]. It follows that (Q∗ T P f, g) = 0 for all f and g. Hence Q∗ T P = 0, which is the second identity we wanted to establish.

6.4. Construction of a canonical factorization

133

We proceed by deriving the state space formulas for G− , G+ and their inverses −1 −1 , KT Q(λ − T )−1. Hence, for λ = 0, G−1 − G+ . Recall that W+ (λ) = I − (I − K) G− (λ)

= = =

W+ (i/λ)∗ (I − Q∗ K) ∗ I − (I − K)−1 KT Q(i/λ − T )−1 (I − Q∗ K) I − iλ(I − iλT )−1 Q∗ T K(I − K)−1 (I − Q∗ K).

On account of Lemma 6.7, we have Q∗ T = T (I − P ). Also (I − K)−1 (I − Q∗ K) = (I − P K)−1 , and we get G− (λ)

−1 I − iλ (I − iλT ) T (I − P )K(I − K)−1 (I − Q∗ K) = I − K − iλ (I − iλT )−1 T (I − P )K (I − P K)−1 .

=

But then, proceeding in a straightforward manner, G− (λ)

=

I − K − iλT (I − iλT )−1(I − P )K (I − P K)−1 I − K + (I − P )K − (I − iλT )−1(I − P )K (I − P K)−1 I − P K − (I − iλT )−1 (I − P )K (I − P K)−1

=

I − (I − iλT )−1 (I − P )K(I − P K)−1 .

= =

In this computation λ was of course taken to be non-zero. For λ = 0, the last expression in the above series of identities reduces to I − (I − P )K(I − P K)−1 and this is easily seen to be equal to (I − K)(I − P K)−1 . The latter can be rewritten as I − Q∗ K which was earlier identiﬁed as the value G− (0) of G− in the origin. So in the ﬁnal analysis the zero value of λ is admissible too. Next we turn to G+ which was deﬁned using W− . For the latter we have the expression W− (λ) = I − (I − K)−1 KT (λ − T )−1 (I − Q) and we can carry out a similar computation as the one presented above: G+ (λ)

(I − Q∗ K)−1 W− (i/λ)∗ (I − K) ∗ = (I − Q∗ K)−1 I − (I − K)−1 KT (i/λ − T )−1 (I − Q) (I − K) = (I − Q∗ K)−1 I − iλ(I − Q∗ )(I − iλT )−1 T K(I − K)−1 (I − K) −1 = (I − Q∗ K)−1 I − K − iλ(I − Q∗ )T (I − iλT ) K = (I − Q∗ K)−1 I − K + (I − Q∗ )(I − iλT − I)(I − iλT )−1K = (I − Q∗ K)−1 I − K + (I − Q∗ )K − (I − Q∗ )(I − iλT )−1 K =

134

Chapter 6. Convolution equations and the transport equation =

(I − Q∗ K)−1 I − Q∗ K − (I − Q∗ )(I − iλT )−1 K

=

I − (I − Q∗ K)−1 (I − Q∗ )(I − iλT )−1 K.

For λ = 0, the last expression comes down to I − (I − Q∗ K)−1 (I − Q∗ )K and this is easily seen to be equal to (I − QK ∗ )−1 (I − K), so to I − P K. The latter was earlier identiﬁed as the value G+ (0) of G+ in the origin. So here the zero value of λ is admissible too. −1 Let us now deal with G−1 − and G+ . The ﬁrst of these functions is tied to W+−1 for which we have the expression W+ (λ)−1 = I + KT × (λ − T × )−1 Q. From this we get −1 ∗ −1 G−1 W+ (i/λ)∗ − (λ) = (I − Q K) ∗ = (I − Q∗ K)−1 I + KT ×(i/λ − T × )−1 Q ∗ = (I − Q∗ K)−1 I + Q∗ (i/λ − T × )−1 (T × )∗ K =

−1 × ∗ (I − Q∗ K)−1 I + iλQ∗ I − iλ(T × )∗ (T ) K

=

−1 K (I − Q∗ K)−1 I − Q∗ K + Q∗ I − iλ(T × )∗

=

−1 I + (I − Q∗ K)−1 Q∗ I − iλ(T × )∗ K.

Putting λ = 0 in the last expression gives I + (I − Q∗ K)−1 Q∗ K which is obviously equal to (I − Q∗ K)−1 , the value of G−1 − at the origin. Finally we consider G−1 . For the appropriate values of λ, we have + −1 −1 W− (i/λ)∗ G−1 (I − Q∗ K) + (λ) = (I − K) = = =

∗ −1 (I − K)−1(I − Q∗ K) (I − K)−1 W− (i/λ) (I − K)

[∗] −1 W− (i/λ) (I − P K)−1

−1 [∗] W− (i/λ) (I − P K)−1 .

Here we have used (6.35) and the fact, already noted above, that I − P K and (I − K)−1 (I − Q∗ K) are each other’s inverse. Recall now that W− (λ)−1 = I + KT ×(I − Q)(λ − T × )−1 . Thus, as T × and K are [· , ·]-selfadjoint, [∗] G−1 I + KT ×(I − Q)(i/λ − T × )−1 (I − P K)−1 + (λ) = =

[∗]

I + (1/iλ − T × )−1 (I − Q)

T × K (I − P K)−1

6.5. The matching of the subspaces

135

= I + iλ(I − iλT × )−1 I − Q[∗] T × K (I − P K)−1 . As an intermediate step, we note that the identity in Lemma 6.7 can be rewritten as I − Q[∗] T × = T × P . Indeed, T ×P

= (I − K)−1 T P = (I − K)−1 (I − Q∗ )T = (I − K)−1 (I − Q)∗ (I − K)T × = (I − Q)[∗] T × = I − Q[∗] T × .

This makes it possible to proceed as follows: G−1 I + iλ(I − iλT × )−1 T × P K (I − P K)−1 + (λ) = = I − P K + (I − iλT × )−1 P K (I − P K)−1 = I + (I − iλT × )−1 P K(I − P K)−1 . The check for λ = 0 yields the desired result, namely I + P K(I − P K)−1 = (I − P K)−1 which is the value of G−1 + at the origin.

6.5 The matching of the subspaces In the canonical factorization carried out in the previous section, we used that ˙ Hp , H = H− +

˙ Hm . H = H+ +

(6.45)

In this section we shall prove that, indeed, the space H may be decomposed in these two ways. Let P− and P+ be the orthogonal projections of H onto H− and H+ , respectively. Also, put Pm = E × (0) and Pp = I − E × (0), where E × (t) is the spectral resolution of the identity for the operator T × = (I − K)−1 T with respect to the inner product [f, g] = (I − K)f, g). By deﬁnition H− = Im P− ,

H+ = Im P+ ,

Hm = Im Pm ,

Hp = Im Pp .

We claim that ˙ Hp H = H− +

⇐⇒ P+ |Hp : Hp → H+ is bijective,

(6.46)

˙ Hm H = H+ +

⇐⇒ P− |Hm : Hm → H− is bijective.

(6.47)

The argument for this is simple and in a diﬀerent context (involving a diﬀerent notation too) spelled out in the beginning of Part 2 of the proof of Theorem 4.4.

136

Chapter 6. Convolution equations and the transport equation

For the convenience of the reader we give it here too. Note that Ker P+ |Hp = H− ∩ Hp , and thus P+ |Hp is injective if and only if H− ∩ Hp = {0}. Next, observe that for each y ∈ Hp we have y = (I − P+ )y + P+ |Hp y ∈ H− + Im P+ |Hp . Thus H− + Hp ⊂ H− + Im P+ |Hp . The reverse inclusion also holds. Indeed, for z ∈ Hp wehave P+ z = (P+ z − z) + z ∈ Ker P+ + Hp = H− + Hp . It follows that H− + Im P+ |Hp = H− + Hp , and hence P+ |Hp is surjective if and only if H = ˙ Hp H− + Hp . This proves (6.46). The proof of (6.47) is similar. Now H = Hm + ˙ H+ . Combining this with (6.46) and (6.47), we see that (6.45) and H = H− + holds if and only if the operator V = P− Pm + P+ Pp is bijective. It is not diﬃcult to prove that V is injective. Indeed, take f ∈ H and assume V f = 0. Put fm = Pm f and fp = Pp f . Then P− fm + P+ fp = V f = 0, and hence P+ fp = 0 and P− fm = 0. The latter gives fm = P+ fm , and we get 0 ≥ [T × fm , fm ] = T fm , fm = T P+ fm , P+ fm ≥ 0. It follows that P+ fm ∈ Ker T . But T is injective. So P+ fm = 0. As P− fm = 0 too, we have fm = 0. In the same way one proves that fp = 0. Hence f = 0, as desired. To prove that V is surjective too, we use that I − V is compact. Indeed, as soon as we know that this is the case, the Fredholm alternative implies that V = I − (I − V ) is surjective if and only if V is injective. Lemma 6.8. The operator I − V is compact. Proof. The compact operators form an ideal and I −V

= P− + P+ − P− Pm − P+ Pp = P− + P+ Pm + P+ Pp − P− Pm − P+ Pp = P− + P+ Pm − P− Pm = (P+ − P− ) (Pm − P− ) .

Hence it suﬃces to prove that Pm − P− is compact. Now Pm = E × (0), where E × (t) is the spectral resolution of the identity for T × with respect to the inner product [· , ·]. Similarly, P− = E(0), where E(t) is the spectral resolution of the identity for T . As T and T × are injective, in both cases the spectral resolutions are continuous at zero. So, using a standard formula for the spectral resolution (see [99], Problem VI.5.7) we may write, for each f ∈ H, 1 (Pm − P− )f = lim h ↓ 0 2πi

Γh

(λ − T )−1 − (λ − T × )−1 f dλ.

(6.48)

Here h is a (suﬃciently small) positive number and Γh is the union of two nonclosed oriented curves as in the following picture:

6.5. The matching of the subspaces

137

+ih −a

0 −i h

The positive number a is chosen in such a way that the spectra of T and T × both are in the open half-line (−a, ∞). For the diﬀerence of the resolvents of T and T × appearing in (6.48) we have (λ − T )−1 − (λ − T × )−1

=

(λ − T )−1 I − (λ − T )(λ − T × )−1

=

(λ − T )−1 (T − T × )(λ − T × )−1

=

−(λ − T )−1 KT ×(λ − T × )−1 ,

and from the latter expression we see that it is a ﬁnite rank (hence compact) operator. Let Δ be the closed contour obtained from Γh by letting the positive number h go to zero. As T × is selfadjoint in H endowed with the inner product [· , ·], we know from (6.32) in Lemma 6.4 and the choice of a that T × (λ − T × )−1 is bounded in norm on Δ \ {0}. Next, let us investigate (λ − T )−1 K. First we shall prove that q0 (ic − T )−1 K ≤ " , |c|

0 = c ∈ R,

(6.49)

where q is some positive constant. To prove this, recall that K is the ﬁnite rank operator given by the right-hand side of (6.7), and hence (ic − T )−1 K ≤

n

|aj | pj (ic − T )−1 pj ,

0 = c ∈ R.

j=0

For each j the function pj is a normalized Legendre polynomial in t (and so the norm of pj appearing in the above expression is actually equal to 1). Also T is the multiplication operator given by the left-hand side of (6.7). So to ﬁnd an upper bound for (ic − T )−1 pj , we need to estimate #

1 −1

c2

t2k dt . + t2

(6.50)

138

Chapter 6. Convolution equations and the transport equation

As t2k+2 ≤ t2k for |t| ≤ 1, it suﬃces to ﬁnd an upper bound for (6.50) for the case k = 0. But # # 1 dt 2 1 = arctan , 0 = c ∈ R. 2 2 |c| |c| −1 c + t This proves (6.49) for an appropriate choice of q0 . Note that the function (λ − T )−1 KT ×(λ − T × )−1 is continuous on Δ \ {0}. Also, for some positive constant q, q 0= c ∈ R. (ic − T )−1 KT × (ic − T × )−1 ≤ " , |c| A straightforward Cauchy argument now gives that (λ − T )−1KT × (λ − T × )−1 dλ lim h↓0

Γh

exists in norm. But then the same is true for (λ − T )−1 − (λ − T × )−1 dλ. lim h↓0

Γh

As the integrand in this expression is a compact operator-valued function, we can use (6.48) to show that Pm − P− is compact too. Close inspection of the above proof shows that I − V is in fact a trace class operator (cf., Lemma 6.3 in [11]).

6.6 Formulas for solutions Let I − H(λ) be the symbol of the Wiener-Hopf integral equation (6.10). From the results of the previous sections we know that I − H(λ) admits a right canonical factorization with respect to the real line: I − H(λ) = G− (λ)G+ (λ),

−∞ < λ < ∞.

(6.51)

As we have seen in Section 6.3, this implies that equation (6.10) is uniquely solvable in L1 ([0, ∞), H), where H = L2 [−1, 1]. This fact and the equivalence (explained in the ﬁrst section of this chapter) of equations (6.1) and (6.10), allows us to prove the following result. Theorem 6.9. Consider equation (6.1) with the kernel function k being given by (6.4). Let T and K be the operators on L2 [−1, 1] deﬁned by (6.7), and assume that I − K is strictly positive. Then equation (6.1) has a unique solution ψ satisfying the initial condition (6.2) and ∞ 1 |ψ(t, μ)|2 dμ dt < ∞. (6.52) 0

−1

6.6. Formulas for solutions

139

This solution is given by −1

× ψ(t, ·) = e−t(Tp ) P f+ ,

t ≥ 0.

(6.53)

Here f+ is the given function appearing in the initial condition (6.2), the operator P is the projection of L2 [−1, 1] deﬁned directly after (6.38), and Tp× is the restriction of T × = (I − K)−1 T to Hp = Im P . Note that (6.53) is the natural analogue of (6.25) in Theorem 6.3. Formula (6.53) features the inverse of the injective operator Tp× = T × |Hp : Hp → Hp . This operator has dense range and with respect to the inner product is nonnegative [· , ·]. Hence its inverse (Tp× )−1 Hp → Hp is an unbounded operator which has Im Tp× as its (dense) domain and is nonnegative with regard to the inner product [· , ·]. Thus the expression × −1 (6.54) e−t(Tp ) is well-deﬁned via the operational calculus for selfadjoint operators based on the notion of the resolution of the identity. One can view (6.54) also as the operator semigroup generated by the unbounded inﬁnitesimal generator −(Tp× )−1 . Proof. Recall that I −H(λ) is the symbol of equation (6.10). Since I −H(λ) admits the canonical Wiener-Hopf factorization (6.51) we can use the general theory of Wiener-Hopf equations (see the one but last paragraph in Section 6.3) to show that equation (6.10) has a unique solution ψ in L1 ([0, ∞), H), where H = L2 [−1, 1]. Moreover, the Fourier transform ψ of ψ is given by −1 ψ(λ) = G−1 (6.55) + (λ)P G− (λ)F (λ) , where F is the Fourier transform of the right-hand side of equation (6.10), and P is the projection deﬁned by ∞ ∞ itλ e f (t) dt = eitλ f (t) dt. P −∞

0

Since ψ ∈ L1 ([0, ∞), H), condition (6.52) is fulﬁlled. To derive formula (6.53), we ﬁrst compute ψ using equation (6.55). Recall that F is given by (6.9). It follows that F(λ) = (I − iλT )−1 T f+ ,

λ ≥ 0.

(6.56)

As we know from Section 6.4 the inverses of the factors G− (λ) and G+ (λ) in (6.51) are given by −1 ∗ −1 ∗ G−1 Q I − iλ(T × )∗ K, − (λ) = I + (I − Q K) × −1 G−1 P K(I − P K)−1 . + (λ) = I + (I − iλT )

140

Chapter 6. Convolution equations and the transport equation

Here T × = (I − K)−1 T . Let us use these formulas to compute ψ(λ) from (6.55). −1 As a ﬁrst step we have G− (λ)F(λ) = F (λ) + X(λ)F(λ), where −1 K X(λ) = (I − Q∗ K)−1 Q∗ I − iλ(T × )∗ =

−1 (I − Q∗ K)−1 Q∗ I − iλT (I − K)−1 K

=

−1 K. (I − Q∗ K)−1 Q∗ (I − K) I − K − iλT

Thus X(λ)F (λ) = (I − Q∗ K)−1 Q∗ (I − K)R(λ)T f+ , where R(λ)

(I − K − iλT )−1 K(I − iλT )−1 = (I − K − iλT )−1 (I − iλT ) − (I − K − iλT ) (I − iλT )−1

=

=

(I − K − iλT )−1 − (I − iλT )−1 .

Hence X(λ)F (λ) = (I − Q∗ K)−1 Q∗ (I − K)(I − K − iλT )−1 T f+ −(I − Q∗ K)−1Q∗ (I − K)(I − iλT )−1 T f+ . We conclude that ∗ −1 ∗ G−1 Q (I − K)F (λ) − (λ)F (λ) = F (λ) − (I − Q K) −1 T f+ . + (I − Q∗ K)−1 Q∗ I − iλ(T × )∗

Now apply the projection P. Since f+ ∈ H+ and T |H+ is nonnegative, we have P(F) = F. Furthermore, using the spectral properties of T × and the deﬁnition of Q, we see that the function Q∗ (I − iλ(T × )∗ )−1 is annihilated by P. Therefore P G−1 = F(λ) − (I − Q∗ K)−1 Q∗ (I − K)F (λ) − (λ)F (λ) = (I − Q∗ K)−1 I − Q∗ K − Q∗ (I − K) F(λ) = (I − Q∗ K)−1 (I − Q∗ )F (λ). Put Z(λ) = (I − Q∗ K)−1 (I − Q∗ )F (λ). Recall from the previous section that I − P K is invertible with inverse (I − K)−1 (I − Q∗ K). Hence (I − P K)−1 (I − Q∗ K)−1 = (I − K)−1 , −1 and it follows that G−1 + (λ)P G− (λ)F (λ) = Z(λ) + H(λ), where

=

(I − iλT × )−1 P K(I − K)−1 (I − Q∗ )F (λ) (I − iλT × )−1 P I − (I − K) (I − K)−1 (I − Q∗ )F (λ)

=

A(λ) − B(λ),

H(λ) =

6.6. Formulas for solutions

141

with A(λ) = (I − iλT × )−1 P (I − K)−1 (I − Q∗ )T (I − iλT )−1f+ , B(λ) = (I − iλT × )−1 P (I − Q∗ )T (I − iλT )−1 f+. Using (I − Q∗ )T = T P and T × = (I − K)−1 T we get A(λ)

=

(I − iλT × )−1 P (I − K)−1 T P (I − iλT )−1 f+

=

(I − iλT × )−1 P T × P (I − iλT )−1 f+

=

(I − iλT × )−1 T × P (I − iλT )−1 f+ ,

and B(λ)

=

(I − iλT × )−1 P T P (I − iλT )−1 f+

=

(I − iλT × )−1 P T (I − iλT )−1 f+ .

Thus (with λ = 0 in the intermediate steps) H(λ)

= (I − iλT × )−1 (T × P − P T )(I − iλT )−1 f+ =

1 (I − iλT × )−1 P (I − iλT ) − (I − iλT × )P (I − iλT )−1 f+ iλ

=

1 1 (I − iλT × )−1 P f+ − P (I − iλT )−1 f+ iλ iλ

=

1 1 P f+ + T × (I − iλT × )−1 P f+ − P f+ − P T (I − iλT )−1 f+ iλ iλ

= T × (I − iλT × )−1 P f+ − P T (I − iλT )−1 f+ . Therefore −1 × × −1 G−1 P f+ + (λ)P G− (λ)F (λ) = T (I − iλT ) + (I − Q∗ K)−1 (I − Q∗ ) − P T (I − iλT )−1 f+ . From Lemma 6.5 we get (I − Q∗ K)−1 (I − Q∗ ) − P

= (I − Q∗ K)−1 (I − Q∗ − P + Q∗ KP ) = (I − Q∗ K)−1 (I − Q∗ − P + Q∗ P ) = (I − Q∗ K)−1 (I − Q∗ )(I − P ) = 0,

and we conclude that −1 × × −1 ψ(λ) = G−1 P f+ . + (λ)P G− (λ)F (λ) = T (I − iλT )

(6.57)

142

Chapter 6. Convolution equations and the transport equation

Now T × maps Hp into Hp , and so the operator Tp× = T × |Hp : Hp → Hp is well-deﬁned. Since Tp× is injective, the expression (6.57) can be rewritten as −1 P f+ . ψ(λ) = − iλ − (Tp× )−1

(6.58)

As was already observed, (Tp× )−1 Hp → Hp is an unbounded operator which has Im Tp× as its (dense) domain and is nonnegative with regard to the inner product [· , ·]. Hence we can take the inverse Fourier transform in (6.58), to get the desired formula (6.53). From (6.53) we see that ψ(0) = ψ(0, · ) = P f+ . Now let P+ be the orthogonal projection of L2 [−1, 1] onto H+ . Since Ker P = H− , we have P+ (I − P ) = 0, and thus P+ ψ(0) = P+ (P f+ ) = P+ (P f+ + (I − P )f+ ) = P+ f+ = f+ . Therefore ψ satisﬁes the initial condition (6.2). Finally, the uniqueness statement follows from the general theory of Wiener-Hopf equations.

Notes The theory of the linear transport equation has a long history. For this see the books [28] and [96] which also contain extensive lists of references. The material in Section 6.2 is taken from Section XIII.9 of [51] where the reader can also ﬁnd an illustrative example. The other sections in this chapter follow basically Chapter 6 in [11] which was inspired by the dissertation [80] and the papers [81], [82]. In [108] one can also ﬁnd an analytic description of the subspaces concerned. Later results based on [110] and [124] are also included here. Further developments using the method described in this chapter can be found in [110], where the case of non-degenerate kernel functions k(μ, μ ) is treated. See also the book [78], and the paper [124]. For an alternative proof of Theorem 6.9, not using Wiener-Hopf factorization, we refer to Section XIX.7 in [51]. The results presented in Sections 6.4 – 6.6 can also be understood from the point of view described in Chapter 5. Note, however, that in Sections 6.4–6.6 the symbol is an operator-valued function (and not a matrix-valued function as in Chapter 5). On the other hand, the operator (T × )−1 appearing in Theorem 6.9 is exponentially dichotomous. This has been proved in Section 5.2 of the recent monograph [111]. The latter book also contains many new additions related to the analysis of equation (6.6). See also the notes to Chapter 5.

Chapter 7

Wiener-Hopf factorization and factorization indices This chapter concerns canonical as well as non-canonical Wiener-Hopf factorization of an operator-valued function which is analytic on a Cauchy contour. Such an operator function is given by a realization with a possibly inﬁnite dimensional Banach space as state space, and with a bounded state operator and with bounded input-output operators. The ﬁrst main result is a generalization to operator-valued functions of the canonical factorization theorem for rational matrix functions presented earlier in Section 3.1. In terms of the given realization, necessary and sufﬁcient conditions are also presented in order that the operator function involved admits a (possibly non-canonical) Wiener-Hopf factorization. The corresponding factorization indices are described in terms of certain spectral invariants which are deﬁned in terms of the realization but do only depend on the operator function and not on the particular choice of the realization. The analysis of these spectral invariants is one of the main themes of this chapter. The chapter consists of three sections. Section 7.1 describes the main result for canonical factorization and introduces the spectral invariants involved. The proof that the spectral invariants do not depend on the particular realization is given in Section 7.2. The ﬁnal section of the chapter, Section 7.3, deals with noncanonical Wiener-Hopf factorization and the corresponding factorization indices.

7.1 Canonical factorization of operator functions Throughout this chapter, W is an operator function, analytic on an open neighborhood of a given Cauchy contour Γ, and with values that are operators on a possibly inﬁnite dimensional Banach space Y . Anticipating the results to be presented below, we note that in this situation W admits a realization on Γ involving

144

Chapter 7. Wiener-Hopf factorization and factorization indices

a possibly inﬁnite dimensional state space X and having IY as external operator: W (λ) = IY + C(λIX − A)−1 B,

(7.1)

where Γ splits the spectrum of A, that is Γ ⊂ ρ(A). This is immediate from Theorem 2.2. As before, we denote by F+ the interior domain of Γ, and by F− the complement of F + in the Riemann sphere C∞ . By a right canonical factorization of W with respect to Γ we mean a factorization W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

(7.2)

where W− and W+ are functions with values in L(Y ) satisfying (i) W− is analytic on F− and continuous on F − , (ii) W+ is analytic on F+ and continuous on F + , (iii) W− and W+ take invertible values on F − and F + , respectively. If in (7.2) the factors W− and W+ are interchanged, we speak of a left canonical factorization. A necessary condition for a right or left canonical factorization with respect to Γ to exist is that W takes invertible values on Γ. In terms of the realization (7.1) this means that Γ also splits the spectrum of the associate main operator A× = A − BC (see Theorem 2.4). We now extend Theorem 3.2 to a possibly inﬁnite dimensional context. Theorem 7.1. Let W be an operator function, analytic on an open neighborhood of a Cauchy contour Γ, and with values that are operators on a Banach space Y . Let (7.1) be a realization of W , i.e., W (λ) = IY + C(λIX − A)−1 B, and suppose Γ splits the spectrum of A. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (a) Γ splits the spectrum of A× = A − BC, ˙ Ker P (A× ; Γ). (b) X = Im P (A; Γ) + In that case, a right canonical factorization of W is given by W (λ) = W− (λ)W+ (λ),

λ ∈ Γ,

where the factors and their inverses can be written as W− (λ)

=

Im + C(λIX − A)−1 (IX − Π)B,

W+ (λ)

=

Im + CΠ(λIX − A)−1 B,

W−−1 (λ)

=

Im − C(IX − Π)(λIX − A× )−1 B,

W+−1 (λ)

=

Im − C(λIX − A× )−1 ΠB.

Here Π is the projection of Cn along Im P (A; Γ) onto Ker P (A× ; Γ).

7.1. Canonical factorization of operator functions

145

For left canonical factorizations an analogous theorem holds. In the result in ˙ Im P (A× ; Γ). The theorem also question, (b) is replaced by X = Ker P (A; Γ) + has an analogue for appropriate closed contours in the Riemann sphere C∞ like the extended real line or the extended imaginary axis. Proof. To establish the theorem, we can rely for a large part on the proof of Theorem 3.2. In fact, we only have to add an argument for the following assertion: if W admits a right canonical factorization with respect to Γ, then the decomposition in (b) holds. The ﬁrst step consists in showing that if W admits a right canonical factorization with respect to Γ, then there is a way of representing W in the form −1 B such that Γ splits the spectra of A and A × while, W (λ) = IY + C(λI − A) X × = Im P (A; Γ) + ; Γ). ˙ Ker P (A in addition, X Let W (λ) = W− (λ)W+ (λ), λ ∈ Γ, be a right canonical factorization of W . Recall that ∞ belongs to F− . Since W− (∞) is invertible we may assume without loss of generality that W− (∞) = IY . From the identity W− (λ) = W (λ)W+ (λ)−1 and the fact that W is analytic on a neighborhood of Γ, it follows that W− has an analytic extension, again denoted by W− , to some open neighborhood Ω− of the closed set F− ∪ Γ. Taking Ω− suﬃciently small, we have that W− assumes only invertible values on Ω− . But then Theorems 2.3 and 2.4 can be applied to show that W− admits a realization of the form W− (λ) = IY + C− (λIX− − A− )−1 B− ,

λ ∈ Ω− ,

(7.3)

× where σ(A− ) ⊂ F+ and σ(A× − ) ⊂ F+ . Here A− = A− − B− C− . A similar reasoning holds for W+ . This function has an analytic extension, again denoted by W+ , to some open neighborhood Ω+ of the closed set F+ ∪ Γ. Taking Ω+ suﬃciently small, we have that W+ (λ) is invertible for all λ ∈ Ω+ . But then Theorems 2.2 and 2.4 yield that W+ admits a realization

W+ (λ) = IY + C+ (λIX+ − A+ )−1 B+ ,

λ ∈ Ω+ ,

(7.4)

× such that σ(A+ ) ⊂ F− and σ(A× + ) ⊂ F− . Here A+ = A+ − B+ C+ . On Γ we have the factorization W (λ) = W− (λ)W+ (λ), and so we can apply λ ∈ Γ, −1 B, the product rule of Section 2.5 to show that W (λ) = IY + C(λI − A) X where X = X− X+ and A : X → X, B : Y → X and C : X → Y are given by the operator matrices

= A

A−

B− C+

0

A+

,

= B

B− B+

,

= C

C−

C+

.

−1 B has the desired properties. This can be seen The realization IY + C(λI − A) X as follows.

146

Chapter 7. Wiener-Hopf factorization and factorization indices

˜ and the corresponding one From the operator matrix representation for A, × for A = A − BC : X → X, namely A× 0 − × = , A −B+ C− A× + and A × . Furthermore, the it is immediate that Γ splits the spectra of both A × spectral projections P (A; Γ) and P (A , Γ) are of the form IX− 0 IX− × , Γ) = Γ) = , P (A . P (A; 0 0 0 Γ) = X− {0} and Ker P (A × ; Γ) = {0} X+ , and from this Hence Im P (A; = Im P (A; Γ) + × ; Γ) is immediate. ˙ Ker P (A X The proof can now be ﬁnished by verifying the following two identities: Γ) ∩ Ker P (A × ; Γ) = dim Im P (A; Γ) ∩ Ker P (A× ; Γ) , dim Im P (A; % $ X X = dim . dim Γ) + Ker P (A × ; Γ) Im P (A; Γ) + Ker P (A× ; Γ) Im P (A; In other words, we are ready once it has been shown that the right-hand side of these identities depend only on W and Γ and are independent of the realization (7.1) of W . This is indeed the case as is seen from Theorem 7.2 below which even exhibits several other spectral invariants. Theorem 7.2. Let W be an operator function, analytic on an open neighborhood of a Cauchy contour Γ, and with values that are operators on a Banach space Y . Let (7.1) be a realization of W , i.e., W (λ) = IY + C(λIX − A)−1 B, and suppose Γ splits the spectrum of A. In addition, assume that Γ also splits the spectrum of A× = A − BC. Introduce P = P (A; Γ),

M = Im P,

Then the quantities ×

dim(M ∩ M ), dim dim

dim

P × = P (A× ; Γ),

M × = Ker P × .

X , M + M×

M ∩ M × ∩ Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk−1 M ∩ M × ∩ Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk M + M × + Im B + Im AB + · · · + Im AB k M + M × + Im B + Im AB + · · · + Im AB k−1

,

k = 0, 1, 2 . . . ,

,

k = 0, 1, 2 . . . ,

depend on W only and do not depend on the realization (7.1) of W .

7.2. Proof of Theorem 7.2

147

The theorem has an analogue for appropriate closed contours in the Riemann sphere C∞ like the extended real line or the extended imaginary axis. To put Theorem 7.2 in context, consider a proper rational matrix function W having the value Im at inﬁnity. With a realization W (λ) = Im + C(λIn − A)−1 B of W , one can associate the numbers k = 0, 1, 2, . . . , (7.5) dim Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk−1 , codim Im B + Im AB + · · · + Im AB k−1 , k = 0, 1, 2, . . . . (7.6) Here the codimension is taken with respect to Cn . Now realizations of rational matrix functions are not unique and the above numbers, as well as their diﬀerences, generally vary with diﬀerent choices of A, B and C in the realization for W . The above theorem shows that this dependence on the speciﬁc form of (7.1) disappears when one combines the spaces appearing in (7.5) and (7.6) with certain spectral subspaces of A and A× . We will meet the subspaces featuring in (7.5) and (7.6) again in Section 8.1. The proof of Theorem 7.2 is rather complicated and we will devote a separate section to it.

7.2 Proof of Theorem 7.2 Let W and Γ be as in Theorem 7.2, and suppose we have the realizations −1 B, W (λ) = IY + C(λI − A) X

(7.7)

−1 B, W (λ) = IY + C(λI − A) X

(7.8)

and A × . In other and A × as well as those of A where Γ splits the spectra of A × × words Γ ⊂ ρ(A) ∩ ρ(A ) ∩ ρ(A) ∩ ρ(A ). Writing Γ), P = P (A;

! = Im P , M

× ; Γ), P× = P (A

!× = Ker P × , M

Γ), P = P (A;

M = Im P ,

× ; Γ), P × = P (A

M × = Ker P × ,

we need to show that !∩M !× ) = dim(M ∩ M × ), dim(M $ % $ % X X dim = dim , !+M !× M M + M× % $ ∩ Ker C A ∩ · · · ∩ Ker C A k−1 !∩M !× ∩ Ker C M dim ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k !∩M !× ∩ Ker C M $ % ∩ Ker C A ∩ · · · ∩ Ker C A k−1 M ∩ M × ∩ Ker C = dim , ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k M ∩ M × ∩ Ker C

148

Chapter 7. Wiener-Hopf factorization and factorization indices $

% !+M !× + Im B + Im A B + · · · + Im A B k−1 + Im A B k M dim + Im A B + · · · + Im A B k−1 !+M !× + Im B M % $ + Im A B + · · · + Im A B k−1 + Im A B k M + M × + Im B . = dim + Im A B + · · · + Im A B k−1 M + M × + Im B Here k = 0, 1, 2, . . . . It is convenient to ﬁrst present a series of auxiliary results. These concern and Ψ given by the integrals the operators Ψ = 1 × )−1 B C(λ − A) −1 dλ, Ψ (λ − A (7.9) 2πi Γ 1 × )−1 B C(λ − A) −1 dλ. (λ − A (7.10) Ψ = 2πi Γ :X →X and Ψ :X → X. Note that Ψ and Ψ also admit the representation: Lemma 7.3. The operators Ψ C(λ −A × )−1 dλ, = 1 −1 B Ψ (λ − A) 2πi Γ −1 B C(λ −A × )−1 dλ, = 1 (λ − A) Ψ 2πi Γ

(7.11) (7.12)

Proof. From Theorem 2.4 we know that − A) −1 , −A × )−1 = C(λ W (λ)C(λ

−A × )−1 = C(λ − A) −1 , W (λ)C(λ

× )−1 BW (λ) = (λ − A) −1 B, (λ − A

× )−1 BW (λ) = (λ − A) −1 B. (λ − A

Now make the appropriate substitutions. and Ψ the following identities hold: Lemma 7.4. For the products of Ψ Ψ = (P× − P )2 , Ψ

Ψ = (P× − P )2 . Ψ

∩ ρ(A). For λ ∈ Γ, we have Proof. It is assumed that Γ ⊂ ρ(A) −1 B −1 B. = C(λ − A) −1 (μ − A) − A) −1 (μ − A) C(λ Indeed, taking advantage of the resolvent identity, we get for λ ∈ Γ, − A) −1 (μ − A) −1 B (μ − λ)C(λ −1 B (λ − A) −1 − (μ − A) =C − C(μ − A) −1 B − A) −1 B = C(λ

(7.13)

7.2. Proof of Theorem 7.2

149 = W (λ) − I − W (μ) − I − A) −1 B − C(μ − A) −1 B = C(λ −1 B (λ − A) −1 − (μ − A) =C −1 B. − A) −1 (μ − A) = (μ − λ)C(λ

Now, when λ = μ, divide by μ − λ; for λ = μ, employ a continuity argument. Ψ, we use the expression (7.10) for Ψ, formula (7.11) for Ψ, To compute Ψ and the identity (7.13): 2 1 × )−1 B C(λ − A) −1 ΨΨ = (λ − A 2πi Γ Γ C(μ −A × )−1 dλ dμ −1 B ·(μ − A) =

1 2πi

2 Γ

Γ

× )−1 B C(λ − A) −1 (λ − A C(μ −A × )−1 dλ dμ −1 B ·(μ − A)

=

1 2πi

2 Γ

Γ

× )−1 (A −A × )(λ − A) −1 (λ − A −1 (A −A × )(μ − A × )−1 dλ dμ ·(μ − A)

=

=

=

1 2πi

1 2πi

2 Γ

Γ

Γ

−1 ) × )−1 − (λ − A) (λ − A × )−1 − (μ − A) −1 dλ dμ · (μ − A

2 × −1 −1 (λ − A ) − (λ − A) dλ

(P× − P )2 .

Ψ = (P× − P )2 , interchange the roles of the realizations (7.7) and (7.8). For Ψ and Ψ satisfy the following intertwining relations: Lemma 7.5. The operators Ψ P = (I − P× )Ψ, Ψ

P × = (I − P )Ψ, Ψ

(7.14)

P = (I − P × )Ψ, Ψ

P × = (I − P )Ψ. Ψ

(7.15)

Proof. Focussing on the ﬁrst identity in (7.14), note that the function × )−1 B C(λ − A) −1 P P × (λ − A

150

Chapter 7. Wiener-Hopf factorization and factorization indices

is analytic on an open neighborhood of F− ∪ Γ. Here F− is the exterior domain of Γ (including ∞). Furthermore, the expansion of this function at inﬁnity is of the C P plus lower order terms. Hence form λ−2 P× B 1 × )−1 B C(λ − A) −1 Pdλ = 0. P× (λ − A 2πi Γ C(λ − A) −1 (I − P) is analytic on an × )−1 B On the other hand (I − P × )(λ − A open neighborhood of F+ ∪ Γ, where F+ is the interior domain of Γ, and so 1 × )−1 B C(λ − A) −1 (I − P)dλ = 0. (I − P × )(λ − A 2πi Γ It follows that P Ψ

= = =

1 2πi 1 2πi 1 2πi

Γ

Γ

Γ

× )−1 B C(λ − A) −1 P dλ (λ − A × )−1 B C(λ − A) −1 P dλ (I − P × )(λ − A × )−1 B C(λ − A) −1 dλ (I − P × )(λ − A

= (I − P × )Ψ, as desired. This proves the ﬁrst identity in (7.14). The second identity in (7.14) is proved given by (7.11). The identities in (7.15) in a similar way using the formula for Ψ follow from those in (7.14) by interchanging the roles of the realizations (7.7) and (7.8). and Ψ satisfy the following Lyapunov equations: Lemma 7.6. The operators Ψ A −A × Ψ Ψ

=

C P − P× B C, B

(7.16)

Ψ A × − A Ψ

=

C P× − PB C, B

(7.17)

A −A × Ψ Ψ

=

C P − P× B C, B

(7.18)

Ψ A × − A Ψ

=

C P× − PB C. B

(7.19)

via (7.9), we have Proof. Using the deﬁnition of Ψ 1 A = × )−1 B C(λ − A) −1 A dλ Ψ (λ − A 2πi Γ

7.2. Proof of Theorem 7.2 1 2πi

=

Γ

151

× )−1 B − λI + λI) dλ C(λ − A) −1 (A (λ − A

1 × −1 −1 × )−1 B C dλ λ(λ − A ) B C(λ − A) dλ − (λ − A 2πi Γ Γ 1 × + A × )(λ − A × )−1 B C(λ − A) −1 dλ − P × B C (λI − A 2πi Γ 1 2πi

=

=

− P × B C. C P + A × Ψ B

=

This gives (7.16). The identity (7.17) can be proved similarly by using the alter of Lemma 7.3. For (7.18) and (7.19), use (7.16) and (7.17) native expression for Ψ and interchange the roles of of the realizations (7.7) and (7.8). Direct computations as the one above of course also work. Ψ, B, B, C and C are related as follows: Lemma 7.7. The operators Ψ, B = (P − P × )B, Ψ

Ψ = C( P − P × ), C

(7.20)

B = (P − P × )B, Ψ

Ψ = C( P − P × ). C

(7.21)

we have Proof. Using the expression (7.9) for Ψ, B Ψ

=

=

=

=

1 2πi

Γ

1 2πi 1 2πi 1 2πi

Γ

Γ

Γ

× )−1 B C(λ − A) −1 B dλ (λ − A × )−1 B W1 (λ) − I dλ (λ − A × )−1 B W2 (λ) − I dλ (λ − A

× )−1 BW 2 (λ) dλ − (λ − A

1 2πi

Γ

× )−1 B dλ. (λ − A

(λ) by (λ − A) −1 B. Hence × )−1 BW By Theorem 2.4, we may replace (λ − A B = Ψ

1 2πi

−1 B dλ − 1 (λ − A) 2πi Γ

Γ

× )−1 B dλ, (λ − A

and this can be rewritten as the ﬁrst part of (7.20). The second part can be proved via a similar computation. The identities in (7.21) follow by interchanging the roles of the realizations (7.7) and (7.8).

152

Chapter 7. Wiener-Hopf factorization and factorization indices

Proof of Theorem 7.2. The proof will be divided into three parts. The ﬁrst con!, M !× , M and M × , ending tains some preliminary observations about the spaces M up in an argument establishing the identities % $ % $ X X × × !∩M ! ) = dim(M ∩ M ), = dim . dim dim(M !+M !× M M + M× Part 1. We begin by noting that M !∩M !× ] ⊂ M ∩ M × , Ψ[

M !+M !× ] ⊂ M + M × , Ψ[

M ∩ M ×] ⊂ M !∩ M !× , Ψ[

M + M ×] ⊂ M !+M !× . Ψ[

To prove this, it suﬃces to show that M !× ⊂ M , Ψ

M ! ⊂ M ×, Ψ

M ⊂ M !× , Ψ

M× ⊂ M !. Ψ

These inclusions, however, are obvious from (7.14) and (7.15). Next observe that !∩M !× M !+M !× M

Ψ), ⊂ Ker (I − Ψ Ψ , ⊃ Im I − Ψ

Ψ), M ∩ M × ⊂ Ker (I − Ψ Ψ . M + M × ⊃ Im I − Ψ

(7.22) (7.23)

Ψ, follow from The formulas concerning the product Ψ Ψ = P P× + (I − P × )(I − P ), I −Ψ which, in turn, is immediate from Lemma 7.4. The two expressions involving the Ψ are obtained by interchanging the roles of the realizations (7.7) and product Ψ (7.8). Consider the restriction operators ! !× Ψ| M ∩M

!∩M !× → M ∩ M × , : M

Ψ| M ∩M ×

!∩M !× . : M ∩ M× → M

!∩ M !× From (7.22) it is clear that these operators are each others inverse. Hence M × and M ∩ M are linearly isomorphic and so they have the same (possibly inﬁnite) dimension. Next we turn to the operators : Φ

X !+M !× M

→

X M + M×

,

: Φ

X M + M×

→

X !+M !× M

,

and |Ψ, respectively. These are well-deﬁned because of the inclusions induced by Ψ M + M ×] ⊂ M !+ M !× . Also it follows from (7.23) M !+ M !× ] ⊂ M + M × and Ψ[ Ψ[

7.2. Proof of Theorem 7.2

153

and Φ are each other’s inverse. Thus the quotient spaces X/ M !+ M !× and that Φ M + M × are linearly isomorphic. In particular they have the same (possibly X/ inﬁnite) dimension. Part 2. In this part of the proof we shall verify that for all nonnegative integers k the following identities hold: $ % !∩M !× ∩ Ker C ∩ Ker C A ∩ · · · ∩ Ker C A k−1 M dim !∩M !× ∩ Ker C ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k M $ % ∩ Ker C A ∩ · · · ∩ Ker C A k−1 M ∩ M × ∩ Ker C = dim . ∩ Ker C A ∩ · · · ∩ Ker C A k−1 ∩ Ker C A k M ∩ M × ∩ Ker C This will be done by showing that the quotient spaces appearing in these identities are linearly isomorphic. To facilitate the discussion, we adopt the notation A) = Ker C ∩ Ker C A ∩ · · · ∩ Ker C A k−1 , Ker k (C| A) is read as X. Of course the where, following standard convention, Ker 0 (C| A) is deﬁned similarly. First we shall prove that the operator Ψ notation Ker k (C| × × ! ! maps M ∩ M ∩ Ker k (C|A) into M ∩ M ∩ Ker k (C|A). This has already been established for k = 0 (Part 1). For k = 1 it must be proved that M !∩M !× ∩ Ker C ⊂ M ∩ M × ∩ Ker C. Ψ M !∩M !× ⊂ M ∩ M × , and so it is enough to derive We know already that Ψ ⊂ Ker C or, what comes down to the same, M !∩M !× ∩ Ker C the inclusion Ψ !∩ M !× ∩ Ker C ⊂ Ker C Ψ. The latter, however, is immediate from the identity M × for which we refer to Lemma 7.7. C Ψ = −C(I − P ) − C P + C We proceed by induction. Let k be a nonnegative integer and suppose that maps M !∩ M !× ∩ Ker k (C| A) into M ∩ M × ∩ Ker k (C| A). We shall the operator Ψ show that the same is true with k replaced by k + 1. Clearly !∩M !× ∩ Ker k+1 (C| A) = M !∩M !× ∩ Ker k (C| A) ∩ Ker C A k , M and C replaced by M , M × , A and C, respectively. !, M !× , A and similarly with M maps Hence, in view of the induction hypothesis, it is suﬃcient to verify that Ψ × k ! ! the space M ∩ M ∩ Ker k+1 (C|A) into Ker C A . In other words, what we need is the inclusion ⊂ M !∩M !× ∩ Ker k+1 (C| A k Ψ A). Ker C (7.24) can be A k Ψ With the help of (the second identity in) Lemma 7.6, the operator C written as A k Ψ C

C P× + PB C +Ψ A × ) A k−1 (−B = C A k−1 − B C P × + (PB −Ψ B) C +Ψ A = C

154

Chapter 7. Wiener-Hopf factorization and factorization indices

and we may conclude that A k Ψ ∩ Ker C A k−1 Ψ ⊃ M !× ∩ Ker C A. Ker C

(7.25)

A). Employing the induction hypothesis A k−1 ⊃ M ∩ M × ∩ Ker k (C| Now Ker C k−1 M !∩ M !× ∩ Ker k (C| A) , i.e., A ⊃ Ψ once again gives Ker C ⊃ M !∩M !× ∩ Ker k (C| A k−1 Ψ A). Ker C But then A = A k−1 Ψ Ker C

−1 Ker C A k−1 Ψ A

⊃

!∩M !× ∩ Ker k (C| −1 M A) A

=

× A) ! ∩A −1 M ! ∩A −1 Ker k (C| −1 M A

⊃

× !∩ A +B C −1 [M !× ] ∩ A −1 Ker k (C| A) M

⊃ ⊃

A) ∩A −1 Ker k (C| !∩A ×−1 M × ∩ Ker C M !∩M !× ∩ Ker C ∩A −1 Ker k (C| A) , M

and hence, taking into account (7.25), ⊃ M !∩M !× ∩ Ker C ∩A −1 Ker k (C| A) . A k Ψ Ker C A)] = Ker k+1 (C| A), the inclusion (7.24) follows. ∩A −1 Ker k (C| As Ker C maps Fix the nonnegative integer k. As we have seen, the linear operator Ψ A) into M ∩ M × ∩ Ker k (C| A). Likewise Ψ maps the space !∩M !× ∩ Ker k (C| M A) into M !∩ M !× ∩ Ker k (C| A). The same is true with k replaced M ∩ M × ∩ Ker k (C| by k + 1. But then the linear operators k Θ

:

A) A) !∩M !× ∩ Ker k (C| M ∩ M × ∩ Ker k (C| M → , (7.26) A) A) !∩M !× ∩ Ker k+1 (C| M M ∩ M × ∩ Ker k+1 (C|

k Θ

:

A) A) !∩M !× ∩ Ker k (C| M ∩ M × ∩ Ker k (C| M → , (7.27) × × ! ! M ∩ M ∩ Ker k+1 (C|A) M ∩ M ∩ Ker k+1 (C|A)

and Ψ, respectively, are well-deﬁned. They are also each others induced by Ψ inverse. This can be deduced easily from !∩M !× ∩ Ker k (C| A) ⊂ M

Ψ), Ker (I − Ψ

A) ⊂ M ∩ M × ∩ Ker k (C|

Ψ), Ker (I − Ψ

7.2. Proof of Theorem 7.2

155

two inclusions which are immediate from (7.22). Thus the quotient spaces appearing in (7.26) and (7.27) are linearly isomorphic. In particular they have the same (possibly inﬁnite) dimension. Part 3. Finally we shall prove that the identities $ % !+M !× + Im B + Im A B + · · · + Im A B k−1 + Im A B k M dim + Im A B + · · · + Im A B k−1 !+M !× + Im B M % $ + Im A B + · · · + Im A B k−1 + Im A B k M + M × + Im B = dim + Im A B + · · · + Im A B k−1 M + M × + Im B are valid for all nonnegative integers k. This will be done by showing that the quotient spaces appearing in these identities are linearly isomorphic. To facilitate the discussion, we adopt the notation B) = Im B + Im A B + · · · + Im A B k−1 , Im k (A| B) is read as {0}. Of course the where, following standard convention, Im 0 (A| notation Im k (A|B) is deﬁned similarly. First we shall verify that the operator Ψ × × ! ! maps M + M + Im k (A|B) into M + M + Im k (A|B). This has already been established for k = 0 (Part 1). For k = 1 it must be proved that M !+M !× + Im B ⊂ M + M × + Im B. Ψ M !+M !× ⊂ M + M × , and so it is enough to derive We know already that Ψ or, what comes down to the same, Im B] ⊂ M + M × + Im B the inclusion Ψ × The latter, however, is immediate from the identity Im ΨB ⊂ M + M + Im B. −B for which we refer to Lemma 7.7. B = P B + (I − P × )B Ψ We proceed by induction. Let k be a positive integer and suppose that the maps the space M !+ M !× + Im k (A| B) into M + M × + Im k (A| B). We operator Ψ shall show that the same is true with k replaced by k + 1. Clearly !+M !× + Im k+1 (A| B) = M !+M !× + Im k (A| B) + Im A k B, M and B replaced by M , M × , A and B, respectively. !, M !× , A and similarly with M maps Im A k B Hence, in view of the induction hypothesis, it suﬃces to verify that Ψ B). In other words, what we need is the inclusion into M + M × + Im k+1 (A| A k B B). ⊂ M + M × + Im k (A| Im Ψ

(7.28)

can be A k B With the help of (the ﬁrst identity in) Lemma 7.6, the operator Ψ written as A k B Ψ

= =

C P − P × B C +A × Ψ) A k−1 B (B k−1 C + B( C P − C −C Ψ) +A Ψ A B, (I − P× )B

156

Chapter 7. Wiener-Hopf factorization and factorization indices

and we may conclude that A k B + Im A Ψ A k−1 B. ⊂ M × + Im B Im Ψ

(7.29)

B). Employing the induction hypothesis k−1 B ⊂ M !+M !× + Im k (A| Now Im A k−1 B), i.e., ⊂ M + M × + Im k (A| Im A B once again gives Ψ A k−1 B B). ⊂ M + M × + Im k (A| Im Ψ But then Ψ A k−1 B Im A

=

Im Ψ A k−1 B A

⊂

M + M × + Im k (A| B) A

=

Im k (A| M +A M× + A B) A

⊂

× C [M × ] + A Im k (A| B) +B M+ A +A Im k (A| × M × + Im B B) M +A

⊂

+A Im k (A| B) , M + M × + Im B

⊂

and hence, taking into account (7.29), +A Im k (A| B) . ⊂ M + M × + Im B A k B Im Ψ B) = Im k+1 (A| B), the inclusion (7.28) follows. +A Im k (A| As Im B maps Fix the nonnegative integer k. As we have seen, the linear operator Ψ × × ! ! M + M + Im k (A|B) into M + M + Im k (A|B). Likewise Ψ maps the space B) into M !+ M !× + Im k (A| B). The same is true with k replaced M + M × + Im k (A| by k + 1. But then the linear operators k Φ

:

B) B) !+M !× + Im k (A| M + M × + Im k (A| M → , B) B) !+M !× ∩ Im k+1 (A| M M + M × + Im k+1 (A|

(7.30)

k Φ

:

B) B) !+M !× + Im k (A| M + M × + Im k (A| M → × × ! ! M + M ∩ Im k+1 (A|B) M + M + Im k+1 (A|B)

(7.31)

and Ψ, respectively, are well-deﬁned. They are also each other’s induced by Ψ inverse. This can be deduced easily from A) ⊃ !+ M !× + Im k (C| M

Ψ), Im (I − Ψ

A), M + M × + Im k (C| ⊃

Ψ), Ker (I − Ψ

7.3. Wiener-Hopf factorization and spectral invariants

157

two inclusion relations which are immediate from (7.23). Thus the quotient spaces appearing in (7.30) and (7.31) are linearly isomorphic. In particular they have the same (possibly inﬁnite) dimension. The symmetry in the arguments employed in the above proof (Parts 2 and 3 especially) suggests the possible use of a duality reasoning. Working in a ﬁnite dimensional context this line of approach is indeed possible. In the inﬁnite dimensional situation, however, it does not work, an obstacle being that (sums of) operator ranges need not be closed.

7.3 Wiener-Hopf factorization and spectral invariants Let Y, W, Γ, F+ and F− be as in the preceding two sections, and let ε+ , ε− ∈ C be points in F+ and F− , respectively. By a right Wiener-Hopf factorization of W with respect to Γ (and the points ε+ and ε− ) we mean a factorization W (λ) = W− (λ)D(λ)W+ (λ),

λ ∈ Γ,

(7.32)

where the factors W− and W+ are operator-valued functions, the values being operators on Y , such that (i) W− is analytic on F− and continuous on F − , (ii) W+ is analytic on F+ and continuous on F + , (iii) W− and W+ take invertible values on F − and F + , respectively, (iv) the middle term D in (7.32) has the form D(λ) = Π0 +

κ r

λ − ε+ j j=1

λ − ε−

Πj ,

λ ∈ Γ,

(7.33)

where κ1 , . . . , κr are non-zero integers, κ1 ≤ κ2 ≤ · · · ≤ κr , the operators Π1 , . . . , Πr are mutually disjoint rank 1 projections on Y , and Π0 = IY − (Π1 + · · · + Πr ) so Π0 is a projection disjoint from Π1 , . . . , Πr . A necessary condition for such a factorization to exist is that W takes invertible values on Γ. In terms of a realization of W on Γ this means that Γ splits the spectrum of the associate main operator (see again Theorem 2.4). If in (7.32) the factors W− and W+ are interchanged, we speak of a left Wiener-Hopf factorization. We will focus on the right version; for the left variant analogous results hold. A few remarks are in order. Suppose W admits a right Wiener-Hopf factorization with respect to Γ and the points ε+ ∈ F+ and ε− ∈ F− . Then W also admits a right Wiener-Hopf factorization with respect to Γ and any other two

158

Chapter 7. Wiener-Hopf factorization and factorization indices

points γ+ ∈ F+ and γ− ∈ F− . For γ− in the ﬁnite complex plane this is clear from the simple identity λ − ε+ λ − γ+ λ − γ− λ − ε+ = . λ − ε− λ − γ+ λ − γ− λ − ε− For γ− = ∞, use

λ − ε+ λ − ε−

=

λ − ε+ λ − γ+

λ − γ+

1 λ − ε−

.

This brings the middle term D(λ) into the form D(λ) = Π0 +

r

λ − γ+

κj

λ ∈ Γ.

Πj ,

(7.34)

j=1

κj featured in the latter expression have Note that the scalar functions λ − γ+ their zeros and poles in γ+ and ∞. When the origin belongs to F+ , one can take γ+ = 0 and (7.34) becomes D(λ) = Π0 +

r

λκj Πj ,

λ ∈ Γ.

j=1

This type of middle term plays a role in the study of Toeplitz equations where Γ is taken to be the unit circle (see [52], Chapter XXIV). Although a right Wiener-Hopf factorization is (generally) not unique, the non-zero integers κ1 , . . . , κr are. They are called the right (Wiener-Hopf ) factorization indices of W with respect to Γ. Left factorization indices are deﬁned similarly. Sometimes the term partial indices is used instead of factorization indices. Finally, we mention that right (left) canonical factorization corresponds to the case when the right (left) factorization indices are all zero. For the convenience of the reader, we recall (from the previous section) that Ker k (C|A) and Im k (A|B) are deﬁned as Ker k (C|A)

= Ker C ∩ Ker CA ∩ · · · ∩ Ker CAk−1 ,

Im k (A|B)

= Im B + Im AB + · · · + Im Ak−1 B.

Theorem 7.8. Let the function W be given by the realization (7.1), i.e, W (λ) = IY + C(λIX − A)−1 B, where Γ splits the spectrum of A. Then W admits a right Wiener-Hopf factorization with respect to Γ if and only if the following two conditions are satisﬁed: (a) Γ splits the spectrum of A× = A − BC,

7.3. Wiener-Hopf factorization and spectral invariants (b) dim M ∩ M × < ∞ and dim

X M + M×

159

< ∞,

where M = Im P (A; Γ) and M × = Ker P (A× ; Γ). In that case, the right factorization indices of W can be described in terms of the operators appearing in (7.1) as follows: (c) the number s of negative right factorization indices and the negative right factorization indices −α1 , . . . , −αs (in the ordinary order: −α1 ≤ · · · ≤ −αs ) themselves are given by M ∩ M× s = dim , M ∩ M × ∩ Ker C &

αj = k = 1, 2, . . . | dim

M ∩ M × ∩ Ker k−1 (C|A) M ∩ M × ∩ Ker k (C|A)

' ≥ j , j = 1, . . . , s,

(d) the number t of positive right factorization indices and the positive right factorization indices ω1 , . . . , ωt (in reversed order: ωt ≤ · · · ≤ ω1 ) themselves are given by M + M × + Im B t = dim , M + M× & ωj = k = 1, 2, . . . | dim

M + M × + Im k (A|B) M + M × + Im k−1 (A|B)

' ≥ j , j = 1, . . . , t.

As was already indicated above, for left Wiener-Hopf factorizations an analogous theorem holds. The theorem also has an analogue for appropriate closed contours in the Riemann sphere C∞ like the extended real line or the extended imaginary axis. Proof. For the (long and complicated) proof of the “if part” of Theorem 7.8 we refer to [17]. Here we shall concentrate on the “only if part” and the description of the right factorization indices. So we shall assume that W admits a Wiener-Hopf factorization (7.32) with respect to the contour Γ and, say, the points ε+ ∈ F+ and ε− ∈ F− . According to Theorem 7.2 it suﬃces to prove that there exists a special realization for W , for convenience also written as (7.1), such that Γ splits the spectra of A and A× and for which (b)–(d) hold. The argument consists of several steps. Step 1. Write the negative right factorization indices of W in the ordinary order (so from small to large) as −α1 , . . . , −αs , and the positive right factorization indices

160

Chapter 7. Wiener-Hopf factorization and factorization indices

in the reversed order (so from large to small) as ω1 , . . . , ωt : −α1 ≤ · · · ≤ −αs < 0 < ωt ≤ · · · ≤ ω1 . Then D can be written in the form α ω s 1

λ − ε− j λ − ε+ j D(λ) = P0 + P−j + Pj , λ − ε+ λ − ε− j=1 j=t

(7.35)

(7.36)

where P−1 , . . . , P−s , Pt , . . . , P1 are mutually disjoint rank 1 projections on Y , and P0 = IY − (P−1 + · · · + P−s + Pt + · · · + P1 ), so P0 is a projection disjoint from P−1 , . . . , P−s , Pt , . . . , P1 . For deﬁniteness, we shall assume that s and t are both positive. Step 2. Fix j among the integers 1, . . . , s, and let Dj− (λ) be the scalar function given by α λ − ε− j Dj− (λ) = , λ = ε+ . λ − ε+ Write Jj− for the lower triangular Jordan block with eigenvalue ε+ and order αj , so that σ(Jj− ) = {ε+ } . Further introduce ⎤ ⎡ αj αj ⎢ (ε+ − ε− ) αj ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ .. ⎢ ⎥ ⎢ ⎥ . ⎢ ⎥ − , Bj = ⎢ ⎥ ⎢ ⎥ ⎢ (ε − ε )2 αj ⎥ + − ⎢ ⎥ 2 ⎢ ⎥ ⎢ ⎥ ⎣ ⎦ αj (ε+ − ε− ) 1 Cj−

=

0

... 0 1

.

Then Dj− (λ) = 1 + Cj− (λ − Jj− )−1 Bj− is a (minimal) realization of Dj− . Now Jj−× − ε− Iαj is similar with the lower triangular nilpotent Jordan block of order αj and having eigenvalue ε+ , a similarity being given by the upper triangular matrix ( )αj ν−μ ν − 1 (ε+ − ε− ) , ν −μ μ,ν=1 ν−1 where μ−1 is read as zero for μ > ν. Thus σ Jj−× = {ε− }. Clearly P Jj− ; Γ = I and P Jj−× ; Γ = 0. Hence Im P Jj− ; Γ = Ker P Jj−× ; Γ = C αj ,

7.3. Wiener-Hopf factorization and spectral invariants and so, trivially,

Im k Jj− Bj− = C αj ,

k = 0, 1, . . . .

161

(7.37)

Furthermore, as is easily veriﬁed, Ker k Cj− |Jj− = C αj −k {0}k ,

k = 0, 1, . . . ,

(7.38)

where the right-hand side of the equality is read as {0}αj for k ≥ αj . Step 3. Take j among the integers 1, . . . , t, and let Dj+ (λ) be the scalar function given by ω λ − ε+ j + , λ = ε− . Dj (λ) = λ − ε− Write Jj+ for the lower triangular Jordan block with eigenvalue ε− and order ωj , so that σ(Jj− ) = {ε− } . Further introduce ⎡ Bj+

⎢ ⎢ ⎢ =⎢ ⎢ ⎣

1 0 .. .

⎤ ⎥ ⎥ ⎥ ⎥, ⎥ ⎦

0 Cj+

( =

ωj (ε− − ε+ ) 1

) ωj ωj ωj (ε− − ε+ ) . . . (ε− − ε+ ) . 2 ωj 2

Then Dj+ (λ) = 1+Cj+ (λ−Jj+ )−1 Bj+ is a (minimal) realization of Dj+ . Analogously to what we saw in the previous step for the matrix Jj−× − ε− Iαj , the matrix Jj+× − ε+ Iωj is similar with the lower triangular nilpotent Jordan block of order ωj and having ε+ as eigenvalue. Thus σ Jj+× = {ε+ }. Clearly P Jj+ ; Γ = 0 and P Jj+× ; Γ = I. Hence Im P Jj− ; Γ = Ker P Jj−× ; Γ = {0}ωj , and so, trivially, Ker k Cj+ |Jj+ = {0}ωj ,

k = 0, 1, . . . .

(7.39)

Furthermore, as is easily veriﬁed, Im k Jj+ |Bj+ = C k {0}ωj −k ,

k = 0, 1 . . . ,

where the right-hand side of the equality is read as Cωj for k ≥ ωj .

(7.40)

162

Chapter 7. Wiener-Hopf factorization and factorization indices

Step 4. Let D0 (λ) be the diagonal matrix given by ⎡ D1− (λ) ⎢ .. ⎢ . ⎢ ⎢ Ds− (λ) ⎢ D0 (λ) = ⎢ ⎢ Dt+ (λ) ⎢ ⎢ .. ⎣ .

⎤

D1+ (λ)

⎥ ⎥ ⎥ ⎥ ⎥ ⎥, ⎥ ⎥ ⎥ ⎦

(7.41)

i.e., D0 (λ) is the direct sum of the matrices D1− (λ), . . . , Ds− (λ), Dt+ (λ), . . . , D1+ (λ). Then D0 is a rational m × m matrix function, where m = s + t. To obtain a realization for D0 , we introduce n = α1 + · · · + αs + ωt + · · · + ω1 , and introduce an n × n matrix A0 , an n × m matrix B0 and an m × n matrix C0 as follows: A0 is the direct sum of the matrices J1− , . . . , Js− , Jt+ , . . . , J1+ , B0 is the direct sum of the matrices B1− , . . . , Bs− , Bt+ , . . . , B1+ , and C0 is the direct sum of the matrices C1− , . . . , Cs− , Ct+ , . . . , C1+ . Then, indeed, D0 (λ) = Im + C0 (λIn − A0 )−1 B0 is a (minimal) realization. Obviously, Γ splits the spectra of A0 and A× 0 = A0 − B0 C0 . In fact, these spectra coincide with {ε+, ε− }. (Without the assumption introduced in Step 1 that s and t are both positive, we would have that the spectra of A0 and A× 0 are subsets of {ε+ , ε− }, and these inclusions are both proper if and only if one of the integers s or t equals zero.) Put M0 = Im P (A0 ; Γ) and M0× = Ker P (A× 0 ; Γ). Then M0 = M0× = Cα1 · · · Cαs {0}ωt · · · {0}ω1 .

(7.42)

Further we have, for k = 1, 2, . . . , Ker k (C0 |A0 ) = Im k (A0 |B0 ) =

Ker k (C1− |J1− ) · · · Ker k (Cs− |Js− ) {0}ωt · · · {0}ω1 , Cα1 · · · Cαs Im k (Jt+ |Bt+ )) · · · Im k (J1+ |B1+ ),

and M0 ∩ M0× ∩ Ker k (C0 |A0 )

=

Ker k (C0 |A0 ),

(7.43)

M0 + M0× + Im k (A0 |B0 )

=

Im k (A0 |B0 ).

(7.44)

Here we used (7.39) and (7.37). It is clear from (7.42) that dim M0 ∩ M0× = α1 + · · · + αs . Combining (7.43) and (7.38), we get dim M0 ∩ M0× ∩ Ker k (C0 |A0 ) = max{0, α1 − k} + · · · + max{0, αs − k}.

7.3. Wiener-Hopf factorization and spectral invariants In particular

163

dim M0 ∩ M0× ∩ Ker C0 ) = (α1 − 1) + · · · + (αs − 1),

and it follows that

dim

M0 ∩ M0× M0 ∩ M0× ∩ Ker C0

= s.

Thus, with M, M × , C replaced by M0 , M0× , C0 , respectively, the ﬁrst identity in Theorem 7.8, item (c) is satisﬁed . We also have

M0 ∩ M0× ∩ Ker k−1 (C0 |A0 ) dim = max{αl − k + 1} − max{αl − k} × M0 ∩ M0 ∩ Ker k (C0 |A0 ) l∈{1,...,s} =

1 = {l = 1, . . . , s | αl ≥ k}.

l∈{1,...,s}, αl ≥k

Now, ﬁx j ∈ {1, . . . , s}. Then {l = 1, . . . , s | αl ≥ k} ≥ j and hence

⇔

k ∈ {1, . . . , αj },

* + k = 1, 2, . . . | {l = 1, . . . , s | αl ≥ k} ≥ j = αj .

Combining these elements we see that, with M, M × , A, B, C replaced by M0 , M0× , A0 , B0 , C0 , respectively, the second identity in Theorem 7.8, item (c) holds too. For the two identities in Theorem 7.8, item (d), the analogous observation is true. The arguments are basically the same as the ones presented for item (b). Step 5. Next we deal with the middle term D in the factorization (7.32), written in the form (7.36) with −α1 , . . . , −αs , ωt , . . . , ω1 satisfying (7.35), P−1 , . . . , P−s , Pt , . . . , P1 mutually disjoint rank 1 projections on Y and P0 = IY − (P−1 + · · · + P−s + Pt + · · · + P1 ) . Clearly P0 and P−1 + · · · + P−s + Pt + · · · + P1 are complementary projections. Put Y0 = Ker P0 . Then Y0 = Im P−1 · · · P−s Im Pt · · · P1 and so Y0 can be identiﬁed with Cm where, as before, m = s + t. Thus Y = Cm Im P0 and with respect to this decomposition D(λ) can be written as an operator matrix D0 (λ) 0 D(λ) = . 0 I Here D0 is given by (7.41) and I is the identity operator on Im P0 . Now let C0 , CD = BD = B0 0 , AD = A0 , 0

164

Chapter 7. Wiener-Hopf factorization and factorization indices

where A0 , B0 and C0 are as in Step 4. Then we have the realization D(λ) = IY + CD (λIn − AD )−1 BD , n = α1 +· · ·+αs +ωt +· · ·+ω1 , with Γ splitting the spectra of AD = A0 and A× D = × × A0 − B0 C0 = A× 0 . Write MD = Im P (AD ; Γ) and MD = Ker P (AD ; Γ). In × other words, MD = M0 and MD = M0× where, again, we use the notation of the previous step. For k = 1, 2, . . . , clearly, Ker k (CD |AD ) = Ker k (C0 |A0 ) and Im k (AD |BD ) = Im k (A0 |B0 . It follows that, with M, M × , A, B, C replaced by × M D , MD , AD , BD , CD , respectively, (b)–(d) in Theorem 7.8 are satisﬁed. Step 6. We begin this sixth and ﬁnal step by representing the factors W− and W+ in the Wiener-Hopf factorization (7.32) in the form W− (λ)

= IY + C− (λIX− − A− )−1 B− ,

λ ∈ Ω− ,

W+ (λ)

= IY + C+ (λI X+ − A+ )−1 B+ ,

λ ∈ Ω+ ,

with σ(A− ) ⊂ F+ ,

σ(A× − ) ⊂ F+ ,

σ(A× + ) ⊂ F− .

σ(A+ ) ⊂ F− ,

Why this can be done is explained in the proof of Theorem 7.1. On Γ we have the factorization (7.32), and so we can apply the product rule of Section 2.5 to show that W (λ) = IY + C(λIX − A)−1 B, λ ∈ Γ, where X = X − Cn X + , n = α1 + · · · + αs + ωt + · · · + ω1 , and A : X → are given by ⎡ ⎡ ⎤ A− B− CD B− C+ B− ⎢ ⎢ ⎥ ⎢ ⎢ ⎥ AD BD C+ ⎥ , A=⎢ 0 B = ⎢ BD ⎣ ⎣ ⎦ 0 0 A+ B+

(7.45)

X, B : Y → X and C : X → Y ⎤ ⎥ ⎥ ⎥, ⎦

C=

C−

CD

C+

.

Here the operator matrices are taken with respect to the decomposition (7.45). Now the realization obtained for W this way has the desired properties. This can be seen as follows. Obviously Γ splits the spectrum of A and the same is true for A× = A − BC which has the matrix representation ⎡ ⎤ A× 0 0 − ⎢ ⎥ ⎢ ⎥ A× = ⎢ −BD C− A× 0 ⎥. D ⎣ ⎦ × −B+ C− −B+CD A+

7.3. Wiener-Hopf factorization and spectral invariants

165

Let M = Im P (A; Γ) and M × = Ker P (A× ; Γ). Assume for the moment that we have established the identities × ∩Ker k (CD |AD ) {0+ }, (7.46) M ∩ M × ∩ Ker k (C|A) = {0− } MD ∩MD × M + M × + Im k (A|B) = X− MD + MD + Im k (AD |BD ) X+ ,

(7.47)

where 0− is the zero element in X− , 0+ is the zero element in X+ and k is allowed to take the values 0, 1, 2, . . . . Then it would be clear from the conclusions obtained in the previous step that (b)–(d) in Theorem 7.8 are met and we would be ready. So we need to concentrate on (7.46) and (7.47). Clearly P = P (A; Γ) has the form ⎡ ⎤ IX− P1 P2 ⎢ ⎥ ⎢ ⎥ P (AD ; Γ) P3 ⎥ . P =⎢ 0 ⎣ ⎦ 0 0 0 From the fact that P (A; Γ) is a projection one gets the relations P1 P (AD ; Γ) = 0,

P1 P3 = 0,

P (AD ; Γ)P3 = P3 ,

(where the two outer ones imply the middle). In turn ⎡ ⎤⎡ IX− −P1 −P2 IX− 0 0 ⎢ ⎥⎢ ⎢ ⎥⎢ In −P3 ⎥ ⎢ 0 P (AD ; Γ) 0 P =⎢ 0 ⎣ ⎦⎣ 0 0 IX+ 0 0 0

these give ⎤⎡ IX− P1 ⎥⎢ ⎥⎢ In ⎥⎢ 0 ⎦⎣ 0 0

P2

⎤

⎥ ⎥ P3 ⎥ , ⎦ IX+

with the ﬁrst and last factor in the right-hand side invertible and being each other’s inverse. Hence ⎤ ⎡ IX− −P1 −P2 ⎥ ⎢ ⎥ ⎢ In −P3 ⎥ X− MD + {0+ } = X− MD {0+ }. M =⎢ 0 ⎦ ⎣ 0 0 IX+ × X+ , and it follows that In the same way one gets M × = {0− } MD

M ∩ M× M + M×

× = {0− } MD ∩ MD {0+}, × = X− M D + M D X+ .

Thus (7.46) and (7.47) are valid for k = 0.

(7.48) (7.49)

166

Chapter 7. Wiener-Hopf factorization and factorization indices

To prove (7.46) for arbitrary k we argue as follows. A simple induction argument shows that CAl is of the form % $ l−1

l ν l l = 0, 1, . . . , (7.50) CA = ∗ Qν,l CD AD + CD AD ∗ , ν=0

where Q0,l , . . . Ql−1,l and the stars denote appropriate but here not explicitly speciﬁed operators. Together with (7.48) this gives that the right-hand side of (7.46) is contained in the left-hand side. The reverse inclusion can be proved by an induction argument in which (7.50) is employed once more. Finally let us turn to (7.47). For Al B there is an expression analogous to (7.50), namely ⎡

⎤

∗

⎢ l−1 ⎢ AB=⎢ AνD BD Rν,l + AlD BD ⎣ l

ν=0

⎥ ⎥ ⎥, ⎦

l = 0, 1, . . . ,

(7.51)

∗

where R0,l , . . . Rl−1,l and the stars stand for certain operators. Together with (7.49) this yields that the left-hand side of (7.47) is contained in the right-hand side. The reverse inclusion can be proved by an induction argument in which (7.51) is used once again. We close this section with a couple of observations on the dimension numbers featuring in Theorems 7.2 and 7.8. For shortness sake, introduce M ∩ M × ∩ Ker k−1 (C|A) α k = dim , M ∩ M × ∩ Ker k (C|A) ω k

=

dim

M + M × + Im k (A|B) . M + M × + Im k−1 (A|B)

Here k may run through the positive integers 1, 2, . . . . Recall that Ker 0 (C|A) is read as X and Im 0 (A|B) as {0}, so M ∩ M× , α 1 = dim M ∩ M × ∩ Ker C ω 1

=

dim

M + M × + Im B . M + M×

Using standard linear algebra arguments it can be shown that the sequences α 1 , α 2 , . . . and ω 1 , ω 2 , . . . are decreasing, i.e., k+1 , α k ≥ α

ω k ≥ ω k+1 ,

k = 1, 2, . . . .

7.3. Wiener-Hopf factorization and spectral invariants

167

In addition it can be proved that α k and ω k vanish for k suﬃciently large, provided that M ∩ M × and M + M × have ﬁnite dimension and codimension, respectively. In fact we then even have, M ∩ M × ∩ Ker k (C|A)

=

{0},

M + M × + Im k (A|B)

=

X,

again holding for k suﬃciently large. The considerations in Step 4 in the above proof corroborate these facts. 2 , . . . ; for ω 1 , ω 2 , . . . the situation Here are some details for the integers α 1 , α is analogous. The mapping M ∩ M × ∩ Ker k (C|A) M ∩ M × ∩ Ker k−1 (C|A) → × M ∩ M ∩ Ker k+1 (C|A) M ∩ M × ∩ Ker k (C|A) k . Assume now that induced by A is easily seen to be injective. Hence α k+1 ≤ α M ∩ M × has ﬁnite dimension. Then there exists a positive integer r such that M ∩ M × ∩ Ker k (C|A) = M ∩ M × ∩ Ker r (C|A),

k = r, r + 1, . . . .

Evidently M ∩M × ∩ Ker r (C|A) is invariant under both A and A× . Also A and A× coincide on M ∩ M × ∩ Ker r (C|A). As the restriction of A to M and that of A× to M × have no eigenvalue in common, it follows that M ∩ M × ∩ Ker r (C|A) = {0}.

Notes This chapter is based on the papers [17] and [18]. The material of these papers relevant for this book has been reorganized and several of the arguments have been improved. The details are as follows. The “if part” of Theorem 7.1 is a special case of Theorem 3.1 in [17]; it also has the ﬁrst part of Theorem 1.5 in [11] as a less general predecessor. Theorem 7.2 combines Theorems 5.1 and 6.1 of [18] in a more appropriate formulation. The proof of Theorem 7.2 given in Section 7.2 is a signiﬁcant improvement over the argument given in [18]. The results from [17] and [18] to be mentioned in connection with Theorem 7.8 are Theorem 3.1 and Corollary 3.2 in [17] and Theorem 1.2 in [18]. The spectral invariants appearing in Theorem 7.2 are closely related to the block similarity invariants of operator blocks of the ﬁrst or third kind; see [58], Section XI.5 in particular. For a review of the theory of possibly non-canonical Wiener-Hopf factorization of matrix-valued functions taking invertible values, we refer to the book [29] and the more recent survey article [59]. Wiener-Hopf factorization of operator-valued functions goes back to [71] and [72]; see also the recent book [73] . The fact that the Wiener-Hopf factorization indices depend on the given function only (and not on the particular Wiener-Hopf factorization) is wellknown for continuous matrix-valued functions (see [60]) and for certain classes of continuous operator-valued functions (see [49]). The latter do not cover the class of operator-valued functions considered in this chapter.

Part IV Factorization of selfadjoint rational matrix functions This part deals with factorization problems for rational matrix functions that have Hermitian values on the real line, the imaginary axis, or the unit circle. Included are problems of spectral factorization and pseudo-spectral factorization. The emphasis is on positive deﬁnite and nonnegative functions. In general, the factorizations considered are canonical or pseudo-canonical, and they are symmetric in the sense that they consist of two factors, where the ﬁrst factor is the adjoint of the second (relative to the given curve). This part consists of four chapters. Minimal realizations play an important role in the analysis of rational matrix functions that have Hermitian values on a curve. These are realization of which the order of the state matrix is equal to the MacMillan degree of the function. In the ﬁrst chapter (Chapter 8) we review the theory of such realizations. Included are the state space similarity theorem and the minimal factorization theorem. In this ﬁrst chapter we also introduce the notion of pseudo-canonical factorization and describe such factorizations in state space terms. In Chapter 9 we study in a state space setting spectral factorizations, that is, symmetric canonical factorizations for rational matrix functions that are positive deﬁnite on the unit circle, the real line or the imaginary axis. Chapter 10 carries out a similar program for nonnegative functions. In this case one has to consider symmetric pseudo-canonical factorization. In the ﬁnal chapter (Chapter 11) we present (without proofs) some background material on matrices in ﬁnite dimensional indeﬁnite inner product spaces, and review the main results from this area that are used in this part and the other remaining parts.

Chapter 8

Preliminaries concerning minimal factorization In this chapter we gather together several results concerning minimal realizations and minimal factorizations that will play an important role in the sequel. Most of these results can also be found in Part II of the book [20]. For the reader’s convenience we have chosen to summarize them here (without proofs). Special attention is given to the notion of pseudo-canonical factorization, which is a generalization of canonical factorization by allowing singularities on the curve. This chapter consists of three sections. Sections 8.1 and 8.2 deal with minimal realizations and minimal factorizations, respectively. Section 8.3 is devoted to pseudo-canonical factorization.

8.1 Minimal realizations Let W be a proper rational m × m matrix function, and let W (λ) = D + C(λIn − A)−1 B

(8.1)

be a realization of W . The realization is said to be minimal if the dimension n of the state space has the smallest possible value. This smallest possible value is equal to the McMillan degree of W (see Section 8.5 in [20] for details). The McMillan degree of W will be denoted by δ(W ). For a characterization of minimality in terms of the matrices A, B and C, we need some more terminology. Let A be an n × n matrix, let B be an n × m matrix, and let C be an m × n matrix. The pair (A, B) is called controllable if Im (A|B) = Im B + Im AB + · · · + Im AB n−1 = Cn .

172

Chapter 8. Preliminaries concerning minimal factorization

So (A, B) is controllable if and only if Cn is the unique A-invariant subspace containing Im B. The pair (C, A) is said to be observable if Ker (C|A) = Ker C ∩ Ker CA ∩ · · · ∩ Ker CAn−1 = {0}. Thus (C, A) is observable if and only if {0} is the unique A-invariant subspace contained in Ker C. In line with these deﬁnitions, the realization (8.1) is called controllable, respectively observable, if the pair (A, B) is controllable, respectively the pair (C, A) is observable. From Sections 7.1 and 7.3 in [20] we now recall the main results on minimal realizations in the following two theorems. Theorem 8.1. A realization of a proper rational matrix function is minimal if and only if it is controllable and observable. Theorem 8.2. Let W be a proper rational matrix function and suppose W (λ)

=

D1 + C1 (λIn − A1 )−1 B1 ,

(8.2)

W (λ)

=

D2 + C2 (λIn − A2 )−1 B2 ,

(8.3)

are minimal realizations of W . Then D1 = D2 and there exists a unique invertible n × n matrix S such that S −1 A1 S = A2 ,

S −1 B1 = B2 ,

C1 S = C2 .

(8.4)

This second theorem is known as the state space similarity theorem; the operator S is called a (state space) similarity between the realizations (8.2) and (8.3). In the situation where (8.1) is a minimal realization, there is a close connection between the poles of W and the eigenvalues of A. Obviously, whether or not the realization is minimal, the poles of W form a subset of σ(A). However, when the realization is minimal, the spectrum of A coincides with the set of poles of W . In addition, when W is a square matrix-valued function, and D is invertible so that A× = A − BD−1 C is well-deﬁned, σ(A× ) is precisely equal to the set of zeros of W . Here a zero of W is a pole of the inverse W −1 of W . For further details, including a more intrinsic deﬁnition of the notion of a zero of a rational matrix function, taking into account multiplicities and pole orders too, see Chapter 8 in [20]. From Chapter 7 in [20] we also recall that (8.1) is minimal when σ(A) ∩ σ(A× ) = ∅. Next we consider the concept of local minimality. Let λ0 be a point in the complex plane. The realization (8.1) is called locally minimal at λ0 if Im P B + Im P AB + · · · + Im P AB n−1

=

Im P,

(8.5)

Ker CP ∩ Ker CAP ∩ · · · ∩ Ker CAn−1 P

=

Ker P,

(8.6)

8.1. Minimal realizations

173

where P is the Riesz projection of A at λ0 . There is a local version of the observation given at the end of the previous paragraph: if λ0 is not a common eigenvalue of A and A× , then (8.1) is minimal at λ0 . For details see Section 8.4 in [20]) where it is also shown that the realization (8.1) is minimal if and only if it is minimal at each point in the complex plane. We ﬁnish this section by reviewing some results on Jordan chains and copole functions. Let W be a rational square matrix-valued function, and let ϕ be a Cm -valued function which is analytic at λ0 with ϕ(λ0 ) = 0. We call ϕ a co-pole function of W at λ0 if W (λ)ϕ(λ) is analytic at λ0 and limλ→λ0 W (λ)ϕ(λ) is nonzero. For this to happen, it is necessary that det W (λ) does not vanish identically. As before, let W −1 denote the pointwise inverse of W , i.e., the function determined by the expression W −1 (λ) = W (λ)−1 . Now, if ϕ is a co-pole function of W at λ0 , then the function ψ(λ) = W (λ)ϕ(λ) is a so-called root function of W −1 at λ0 , that is, ψ is analytic at λ0 with ψ(λ0 ) = 0 and limλ→λ0 W (λ)−1 ψ(λ) = 0. The converse is also true. A root function of W −1 at λ0 is also referred to as a pole function of W at λ0 (see [7], page 67). The next two results have been taken from [20], Section 8.4 (Proposition 8.21 and Corollary 8.22). Proposition 8.3. Let the rational square matrix-valued function W be given by the realization (8.1), and let λ0 be an eigenvalue of A. Assume the realization is minimal at λ0 . Let k ≥ 1, and let ϕ(λ) = (λ − λ0 )k ϕk + (λ − λ0 )k+1 ϕk+1 + · · · be a co-pole function of W at λ0 . Put xj =

∞

P (A − λ0 )ν−j−1 Bϕν ,

j = 0, . . . , k − 1,

(8.7)

ν =k

where P is the Riesz projection of A corresponding to λ0 . Then x0 , . . . , xk−1 is a Jordan chain of A at λ0 , that is, x0 = 0 and (A − λ0 )x0 = 0,

(A − λ0 )r xk−1 = xk−1−r ,

r = 0, . . . , k − 1.

(8.8)

Moreover, each Jordan chain of A at λ0 is obtained in this way. Finally, if the chain x0 , . . . , xk−1 given by (8.7) is maximal, that is, xk−1 ∈ Im (A − λ0 ), then ϕk = 0. With respect to (8.7) there is no convergence issue; actually only a ﬁnite number of terms in the sum are non-zero. Proposition 8.4. Let the rational square matrix-valued function W be given by the realization (8.1), and suppose det W (λ) ≡ 0. Let λ0 be an eigenvalue of A, and assume that (8.1) is minimal at λ0 . If x0 , . . . , xk−1 is a Jordan chain of A at λ0 , then Cx0 , . . . , Cxk−1 is a Jordan chain of W −1 at λ0 , and each Jordan chain of W −1 at λ0 is obtained in this way.

174

Chapter 8. Preliminaries concerning minimal factorization

For later use (see Section 10.1) we introduce the following terminology suggested by Proposition 8.3. Let W be given by the realization(8.1). If x0 , . . . , xk−1 ∞ is a Jordan chain of A at λ0 , any co-pole function ϕ(λ) = j=k (λ − λ0 )j ϕj satisfying (8.7) will be called a co-pole function corresponding to the Jordan chain x0 , . . . , xk−1 . In this case Cxj is precisely the coeﬃcient of (λ − λ0 )r in the Taylor expansion of W (λ)ϕ(λ) at λ0 . To see this, use (8.7) and the fact that the coeﬃcients in the principal part of the Laurent expansion of W at λ0 are given by the expression CP (A − λ0 )j−1 B, where P is the Riesz projection of A corresponding to the eigenvalue λ0 . These observations lie also behind Proposition 8.4 above.

8.2 Minimal factorization The McMillan degree features a sublogarithmic property. Indeed, if W1 and W2 are rational matrix functions and W = W1 W2 , that is W (λ) = W1 (λ)W2 (λ), then the McMillan degree of W is less than or equal to the sum of the McMillan degrees of W1 and W2 : δ(W1 W2 ) ≤ δ(W1 ) + δ(W2 ).

(8.9)

This is clear from Theorem 2.5 and the deﬁnition of the McMillan degree given in the beginning of the previous section. A factorization W = W1 W2 is called a minimal factorization (involving two factors) if equality occurs, that is, when δ(W ) = δ(W1 ) + δ(W2 ). Intuitively, this means that there is no pole-zero cancellation in the product W1 W2 ; this is made precise in Theorem 9.1 in [20]. Let W (λ) = D + C(λIn − A)−1 B be a realization of an m× m rational matrix function, assume that D is invertible, and let D = D1 D2 with D1 , D2 m × m matrices (automatically invertible). Put A× = A − BD−1 C. Suppose M, M × is a pair of subspaces of Cn satisfying A× M × ⊂ M × ,

AM ⊂ M,

˙ M × = Cn . M+

(8.10)

In that case we know (see Section 2.6) that W admits a factorization W = W1 W2 where the factors can be described using the projection Π onto M × along M as follows: W1 (λ)

=

D1 + C(λIn − A)−1 (I − Π)BD2−1 ,

(8.11)

W2 (λ)

=

D2 + D1−1 CΠ(λIn − A)−1 B.

(8.12)

The next theorem, which is a reformulation of the main result in [20], Section 9.1, shows that the above factorization principle yields all minimal factorizations of W whenever the given realization is minimal.

8.2. Minimal factorization

175

Theorem 8.5. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of the m × m rational matrix-valued function W , and assume D is invertible. (i) Let D = D1 D2 with D1 , D2 (invertible) m×m matrices. If a pair of subspaces M and M × of Cn satisﬁes (8.10), then the factorization W = W1 W2 , with the factors W1 and W2 given by (8.11) and (8.12), is a minimal factorization. (ii) If W = W1 W2 is a minimal factorization of W involving proper rational m × m matrix functions W1 and W2 , then there is a unique pair of subspaces M and M × satisfying (8.10) such that the factors W1 and W2 are given by (8.11) and (8.12) where D1 and D2 are the (invertible) values of W1 and W2 at ∞, respectively. The notion of minimal factorization can be extended to products involving an arbitrary number of factors. Indeed, a factorization W = W1 · · · Wk is called a minimal factorization if δ(W ) = δ(W1 ) + · · · + δ(Wk ).

(8.13)

In general all we can say is that the left-hand side of (8.13) does not exceed the right-hand side. The special case of complete factorization is of particular interest. Let W be a rational m × m matrix-valued function which is biproper , that is, W is analytic at inﬁnity and has an invertible value there. A minimal factorization of W into biproper rational m × m matrix functions, each having McMillan degree 1, is called a complete factorization of W . The number of factors in such a complete factorization is necessarily equal to the McMillan degree of W . If W (λ) = D + C(λIn − A)−1 B is a minimal realization of W , then W admits a complete factorization if and only if the matrices A and A× can be brought into complementary triangular form, i.e., there is a basis such that, with respect to this basis, A has upper triangular form and A× has lower triangular form. For further details, see Chapter 10 in [20]. We shall meet complete factorization later in Section 17.3. We conclude this section with some remarks on a local version of minimal factorization. First we introduce the local (McMillan) degree. Let W be a proper rational matrix function, let W (λ) = D + C(λIn − A)−1 B

(8.14)

be a minimal realization of W , and let μ ∈ C. The algebraic multiplicity af μ as an eigenvalue of A is called the local (McMillan) degree of W at μ, written δ(W ; μ). By the state space similarity theorem, this deﬁnition does not depend on the choice of the minimal realization (8.14). For an alternative deﬁnition of the local degree, we refer to Section 8.4 in [20] where the square case is considered. In that situation, when det W (λ) does not vanish identically, the local degree of W

176

Chapter 8. Preliminaries concerning minimal factorization

at μ coincides with the pole-multiplicity of W at μ in the sense of [20], Section 8.2. It is obvious, again from Theorem 2.5, that the global sublogarithmic property (8.9) has the following local counterpart: δ(W1 W2 ; μ) ≤ δ(W1 μ) + δ(W2 ; μ).

(8.15)

A factorization W = W1 W2 is said to be locally minimal at μ if equality occurs in (8.15), that is, when δ(W1 W2 ; μ) = δ(W1 μ) + δ(W2 ; μ). Intuitively, this means that in the product W1 W2 no pole-zero cancellation occurs at the point μ (see again Theorem 9.1 in [20]). For the case of proper rational matrix functions (as considered here), the minimality of a factorization comes down to local minimality at each point in the complex plane. Thus W = W1 W2 is a minimal factorization if and only if δ(W1 W2 ; λ) = δ(W1 λ) + δ(W2 ; λ),

λ ∈ C;

see Section 9.1 in [20].

8.3 Pseudo-canonical factorization Let Γ be a Cauchy contour in C. As before, the interior domain of Γ is denoted by F+ , and the exterior domain by F− . By deﬁnition (see Chapter 0), ∞ ∈ F− . Let W be an m × m rational matrix function, possibly having poles and zeros on Γ. By a right pseudo-canonical factorization of W with respect to Γ we mean a factorization W (λ) = W− (λ)W+ (λ),

λ ∈ Γ, λ not a pole of W,

(8.16)

where W− and W+ are rational m × m matrix functions such that W− is analytic and takes invertible values on F− (i.e., W− has neither poles nor zeros there), W+ is analytic and takes invertible values on F+ (i.e., W− has neither poles nor zeros there), and the factorization (8.16) is locally minimal at each point of Γ. If in (8.16) the factors W− and W+ are interchanged, we speak of a left pseudocanonical factorization. In passing we mention that the deﬁnition of pseudo-canonical factorization given in the second paragraph of [20], Section 9.2 is not quite correct. The point is that the function W is allowed to have poles and zeros on Γ. This is explicitly stated in the third paragraph of the section in question, but the formal deﬁnition referred to above in the second paragraph erroneously suggests otherwise. As for canonical factorization, the notion of pseudo-canonical factorization extends to factorization with respect to the real line and the imaginary axis. To be more speciﬁc, if Γ is the closure of the real line on the Riemann sphere, then F+ is the open upper half plane, and F− is the open lower half plane. Replacing R

8.3. Pseudo-canonical factorization

177

by iR means only replacing the open upper half plane by the open left half plane, and the open lower half plane by the open right half plane. A pseudo-canonical factorization is not only minimal at each point of Γ but also at all other points of C and at inﬁnity. This follows from the conditions on the poles and zeros of the factors W− and W+ in (8.16). Thus a pseudo-canonical factorization is a minimal factorization. In combination with Theorem 8.5 this fact makes it possible to describe all right pseudo-canonical factorizations of a biproper rational matrix function W in terms of a minimal realization of W . The resulting theorem (which is taken from Section 9.2 in [20]) is given below. In contrast to the main theorem on canonical factorization (Theorem 3.2) we are forced here to work with minimal realizations. Theorem 8.6. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a biproper rational matrix-valued function W , and put A× = A−BD−1 C. Let Γ be a Cauchy contour. Let D = D1 D2 , with D1 and D2 invertible square matrices. Then there is a one-to-one correspondence between the right pseudo-canonical factorizations W = W− W+ of W with respect to Γ with W− (∞) = D1 and W+ (∞) = D2 , and the pairs of subspaces M, M × of Cn with the following properties: (i) M is an A-invariant subspace such that the restriction A|M of A to M has no eigenvalues in F− , and M contains the span of all eigenvectors and generalized eigenvectors of A corresponding to eigenvalues in F+ , (ii) M × is an A× -invariant subspace such that the restriction A× |M × of A× to M × has no eigenvalues in F+ , and M × contains the span of all eigenvectors and generalized eigenvectors of A× corresponding to eigenvalues in F− , ˙ M ×. (iii) Cn = M + The correspondence is as follows: given a pair of subspaces M, M × of Cn with the properties (i), (ii) and (iii), a right pseudo-canonical factorization of W with respect to Γ is given by W (λ) = W− (λ)W+ (λ), where W− (λ)

= D1 + C(λIn − A)−1 (I − Π)BD2−1 ,

(8.17)

W+ (λ)

= D2 + D1−1 CΠ(λIn − A)−1 B,

(8.18)

where Π is the projection along M onto M × . Conversely, given a right pseudocanonical factorization of W with respect to Γ and with W− (∞) = D1 , W+ (∞) = D2 , there exists a unique pair of subspaces M, M × with the properties (i), (ii) and (iii) above, such that the factors W− and W+ are given by (8.17) and (8.18), respectively. The span of all eigenvectors and generalized eigenvectors of A corresponding to eigenvalues in F+ mentioned in (i) is just the spectral subspace of A associated with the part of the spectrum of A lying in F+ . Similarly, the span of all eigenvectors and generalized eigenvectors of A× featuring in (ii) corresponding to

178

Chapter 8. Preliminaries concerning minimal factorization

eigenvalues in F− is the spectral subspace of A× associated with the part of σ(A× ) lying in F− . A pair of subspaces M, M × for which (i), (ii) and (iii) hold need not be unique. In line with this, pseudo-canonical factorizations are generally not unique either. An example illustrating this is given in [133]; see also Section 9.2 in [20]. Note that for an m × m rational matrix function W , a canonical factorization of W with respect to the curve Γ is a pseudo-canonical factorization with the additional property that the factors have no poles or zeros on the curve. In that case, W has no poles or zeros on Γ also. Conversely, if W has no poles or zeros on Γ, then any pseudo-canonical factorization W = W1 W2 of W is automatically a canonical factorization. Indeed, if W has no poles or zeros on Γ, then the fact that the factorization W = W1 W2 is locally minimal at each point of Γ, implies that W1 and W2 have no poles or zeros on Γ, and thus the pseudo-canonical factorization W = W1 W2 is a canonical one. As a result we have the following special case of Theorem 8.6. Theorem 8.7. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a biproper rational matrix-valued function W , and put A× = A − BD−1 C. Let Γ be a Cauchy contour. Assume that A has no eigenvalues on Γ. Then W admits a right canonical factorization with respect to Γ if and only if the following two conditions are satisﬁed: (i) A× has no eigenvalues on Γ, ˙ Ker P (A× ; Γ). (ii) Cn = Im P (A; Γ)+ In that case, the right canonical factorizations with respect to Γ are of the form W = W− W+ , with W− and W+ given by (8.17) and (8.18), where Π is the projection along Im P (A; Γ) onto Ker P (A× ; Γ), and where D = D1 D2 , with D1 and D2 invertible square matrices. This correspondence is a one-to-one correspondence between the right canonical factorizations of W and the factorizations of D into square factors. Observe that the above theorem is a modest reﬁnement of Theorem 3.2 in the sense that we allow the value of W at inﬁnity to be an arbitrary invertible matrix here. The result of the theorem also holds for non-minimal realizations. The argument for this consists of a straightforward modiﬁcation of the proof of Theorem 3.2. Theorem 8.7 allows for analogues in which the Cauchy contour Γ is replaced by the extended real or imaginary axis.

Notes The material in the ﬁrst section is standard and can be found in many textbooks; see, e.g., [94], or the more recent [33], [85]. The idea of minimal factorization originates from mathematical systems theory and has been developed systematically in Chapter 4 of [11] (see also [21]), and with further details in Part II of [20]. An

8.3. Pseudo-canonical factorization

179

extensive analysis of factorization into square degree 1 factors can be found in Part III of [20]. The analysis involves a connection with a problem of job scheduling from operations research. Minimal factorization into possibly non-square factors of McMillan degree 1 is always possible. This has been established in [143]. The notion of a pseudo-canonical factorization is introduced and developed in [132], [133].

Chapter 9

Factorization of positive deﬁnite rational matrix functions The central theme of this chapter is the state space analysis of rational matrix functions with Hermitian values either on the real line, on the imaginary axis, or on the unit circle. The main focus will be on rational matrix functions that take positive deﬁnite values on one of these contours. It will be shown that if W is such a function, then W admits a spectral factorization, i.e., a canonical factorization W = W− W+ with an additional symmetry between the corresponding factors, depending on the contour. This chapter consists of three sections. In Section 9.1 we analyze selfadjointness of a rational matrix function relative to the real line, the imaginary axis or the unit circle. The analysis is done in terms of (minimal) realizations of the functions involved. Elements of the theory of matrices that are selfadjoint with respect to an indeﬁnite inner product enter into the analysis in a natural way. Section 9.2 deals with rational matrix functions that are positive deﬁnite on the real line or on the imaginary axis. The results of Section 9.1 are used to show that such a function admits a spectral factorization and in terms of a given realization an explicit formula for the corresponding spectral factor is given. Section 9.3 presents an analogous result for rational matrix functions that are positive deﬁnite on the unit circle.

9.1 Preliminaries on selfadjoint rational matrix functions Let Γ be one of the following two contours in the complex plane: the real line R, or the imaginary axis iR. A rational m × m matrix function W is called selfadjoint on Γ or Hermitian on Γ if for each λ ∈ Γ, λ not a pole of W , the matrix W (λ)

182

Chapter 9. Factorization of positive deﬁnite rational matrix functions

is selfadjoint or, which is the same, Hermitian. By the uniqueness theorem for analytic functions, a rational matrix function W is selfadjoint on R if and only if ¯ ∗ for all λ ∈ C, λ not a pole of W . Similarly, W is selfadjoint on iR W (λ) = W (λ) ¯ ∗ , λ not a pole of W . From these characterizations if and only if W (λ) = W (−λ) it follows that if W is selfadjoint on Γ and det W (λ) does not vanish identically, then W −1 is also selfadjoint on Γ. This section is concerned with the problem how selfadjointness of a rational matrix function is reﬂected in properties of the matrices in a minimal realization of the function. For proper rational matrix functions this is described in the following theorem. Theorem 9.1. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function. Then the following statements hold: (i) W is Hermitian on the real line if and only if D = D∗ and there exists an n × n matrix H such that HA = A∗ H,

HB = C ∗ ,

H = H ∗;

(9.1)

(ii) W is Hermitian on the imaginary axis if and only if D = D∗ and there exists an n × n matrix H such that HA = −A∗ H,

HB = C ∗ ,

H = −H ∗ .

(9.2)

In both cases (because of the minimality of the realization), the matrix H is uniquely determined by the matrices in the given realization of W and invertible. A matrix H such that H = −H ∗ is called skew-Hermitian. For such a matrix iH is Hermitian. Proof. We ﬁrst prove (i). Assume the matrix function W is Hermitian on R, so ¯ ∗ coincide. Hence the rational matrix functions W (λ) and W (λ) W (λ) = D ∗ + B ∗ (λ − A∗ )−1 C ∗ is also a minimal realization for W . By the state space similarity theorem (Theorem 8.2) we obtain the existence of a unique (invertible) n × n matrix H such that HA = A∗ H, HB = C ∗ , B ∗ H = C. Taking adjoints one gets H ∗ A∗ = AH,

C = B∗H ∗,

H ∗B = C ∗.

Comparing these two sets of equations and employing the uniqueness of H, we see that H = H ∗ . Clearly D = D∗ as D = W (∞) must be selfadjoint.

9.1. Preliminaries on selfadjoint rational matrix functions

183

For the converse, suppose D = D ∗ and there exist an n×n matrix H for which (9.1) holds. From the ﬁrst equality in (9.1) we see that H(λ−A)−1 = (λ−A∗ )−1 H. Then, using the second equality in (9.1), one computes ¯ ∗ W (λ)

=

D∗ + B ∗ (λ − A∗ )−1 C ∗ = D + B ∗ (λ − A∗ )−1 HB

=

D + B ∗ H(λ − A)−1 B = D + C(λ − A)−1 B = W (λ).

So W is selfadjoint on R. Next we show that (because of minimality) the identities in (9.1) imply that H is invertible. Indeed, assume Hx = 0 for some x ∈ Cn . Then the ﬁrst equality in (9.1) yields HAx = 0. Repeating the argument, using induction, we obtain HAk x = 0 for k = 0, 1, 2, . . .. Using the two other equalities in (9.1) we see that CAk x = B ∗ HAk x = 0 for k = 0, 1, 2, . . . . Since the given realization is minimal, the pair (C, A) is observable, and hence x = 0. Thus H is invertible. The proof of (ii) can be given using the same type of reasoning as for (i). On the other hand (ii) also follows directly from (i) by using the transformation ! (λ) = W (−iλ). Since W is assumed to be selfadjoint on λ → −iλ. Indeed, put W ! ! admits the minimal realization iR, the function W is selfadjoint on R. Moreover, W ! (λ) = D + C(λ − A) −1 B, W = iC and A = iA. By (i), there exists an (invertible) selfadjoint matrix where C we derive the desired and HB =C ∗ . Setting H = −iH ˜ H such that H A = A∗ H equalities in (9.2). In the proof of the “if parts” of (i) and (ii), minimality does not play a role. Thus, if (9.1) holds and D = D ∗ , then W (λ) = D + C(λ − A)−1 B is selfadjoint on R. Similarly, if D = D ∗ and (9.2) holds, then W is selfadjoint on iR. In the next proposition we consider the case when the rational matrix function in Theorem 9.1 is biproper, and we describe the eﬀect of the matrices H on the associate main operator A× = A − BD−1 C. Proposition 9.2. Let W (λ) = D + C(λIn − A)−1 B be a realization of an m × m rational matrix function, and let H be an n × n matrix. Assume D is invertible, and put A× = A − BD−1 C. Then the following statements hold: (i) If D = D∗ and (9.1) is satisﬁed, then HA× = (A× )∗ H; (ii) If D = D∗ and (9.2) is satisﬁed, then HA× = −(A× )∗ H. Proof. Assume D = D ∗ and the identities (9.1). Then HA× = HA − HBD−1 C = A∗ H − C ∗ D −∗ B ∗ H = (A× )∗ H, so (i) holds. Statement (ii) is proved analogously.

184

Chapter 9. Factorization of positive deﬁnite rational matrix functions

Next we analyze how the matrix H appearing in Theorem 9.1 behaves under a state space similarity transformation on the realization of W . Theorem 9.3. For i = 1, 2, let W (λ) = D + Ci (λn − Ai )−1 Bi be a minimal realization of the rational matrix function W , and let S be the (unique invertible) n × n matrix such that SA1 = A2 S,

C1 = C2 S,

B2 = SB1 .

Then the following statements hold: (i) Let W be selfadjoint on the real line. For i = 1, 2, write Hi for the (unique invertible) Hermitian n × n matrix such that Hi Ai = A∗i Hi and Hi Bi = Ci∗ . Then H1 = S ∗ H2 S; (ii) Let W be selfadjoint on iR. For i = 1, 2, write Hi for the (unique invertible) skew-Hermitian n × n matrix such that Hi Ai = −A∗i Hi and Hi Bi = Ci∗ . Then H1 = S ∗ H2 S. Proof. We shall only prove (i); statement (ii) can be veriﬁed analogously. One easily checks that S ∗ H2 S satisﬁes (9.1): S ∗ H2 SA1 = S ∗ H2 A2 S = S ∗ A∗2 H2 S = A∗1 S ∗ H2 S, S ∗ H2 SB1 = S ∗ H2 B2 = S ∗ C2∗ = C1∗ . By the uniqueness of H1 the assertion (i) follows.

We conclude this section with a comment on the theory of matrices acting in an indeﬁnite inner product space. Elements of this theory play an important role in the study of selfadjoint rational matrix functions. To see the connection, let H be an invertible Hermitian n × n matrix, and consider on Cn the sesquilinear form [x, y] = Hx, y. If HA = A∗ H, then [Ax, y] = [x, Ay], and hence A is selfadjoint in the indeﬁnite inner product [· , · ] on Cn induced by H. Thus the ﬁrst part and third identity in (9.1) imply that A is selfadjoint in an indeﬁnite inner product space. In the sequel we call A H-selfadjoint if H = H ∗ and HA = A∗ H. Notice that the third identity in (9.2) implies that iH is Hermitian, and hence the ﬁrst identity in (9.2) can be summarized by saying that iA is iH-selfadjoint. In Section 11.2 we review the results from the theory of matrices acting in an indeﬁnite inner product space insofar as they are useful to us in this and the next chapters.

9.2. Spectral factorization

185

9.2 Spectral factorization The ﬁrst factorization result to be presented in this section concerns an important class of rational matrix functions, namely those which are positive deﬁnite on the contour Γ under consideration (again, either R or iR). A rational m × m matrix function W is called positive deﬁnite on Γ if for each λ ∈ Γ, λ not a pole of W , the matrix W (λ) is positive deﬁnite. Suppose W is a rational m × m matrix function. A factorization ¯ ∗ L(λ) W (λ) = L(λ)

(9.3)

is called a right spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed upper half plane ¯ ∗ and its inverse are analytic on (inﬁnity included). In that case the function L(λ) the closed lower half plane (including inﬁnity). Thus a right spectral factorization with respect to R is a right canonical factorization with respect to the real line featuring an additional symmetry property between the factors. A factorization (9.3) is called a left spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed lower ¯ ∗ and its inverse are half plane (inﬁnity included), in which case the function L(λ) analytic on the closed upper half plane including inﬁnity). Such a factorization is a left canonical factorization with respect to R. A factorization ¯ ∗ L(λ) W (λ) = L(−λ) (9.4) is called a right spectral factorization with respect to the imaginary axis if L and L−1 are rational m × m matrix functions which are analytic on the closed left half plane (inﬁnity included). Such a factorization is, in particular, a right canonical factorization with respect to iR. Analogously, (9.4) is called a left spectral factorization with respect to the imaginary axis if L and L−1 are rational m × m matrix functions which are analytic on the closed right half plane (inﬁnity included). The factors in a spectral factorization are uniquely determined up to multi¯ ∗ L(λ) is plication with a constant unitary matrix. More precisely, if W (λ) = L(λ) a spectral factorization with respect to the real line, and E is an m × m unitary ¯ ∗ L(λ) ˜ λ) ˜ ˜ matrix, then W (λ) = L( with L(λ) = EL(λ) is again a spectral factorization of W , and this is all the freedom one has. To see the latter, assume that ¯ ∗ L(λ) and W (λ) = L( ¯ ∗ L(λ) ˜ λ) ˜ W (λ) = L(λ) are right spectral factorizations with respect to R, then −1 ¯ −∗ L(λ) ¯ ∗. ˜ ˜ λ) L(λ)L(λ) = L( The left-hand side of this identity is an m × m rational matrix function which is analytic on the closed upper half plane and the right-hand side is analytic on the closed lower half plane (in both cases the point inﬁnity included). By Liouville’s theorem neither side depends on λ, that is, there exists an m × m matrix E such ¯ −∗ L( ¯ ∗ . But this implies that E is invertible and ˜ −1 and L(λ) ˜ λ) that E = L(λ)L(λ) ˜ E ∗ = E −1 . Hence E is unitary and E = L(λ) = EL(λ), as desired.

186

Chapter 9. Factorization of positive deﬁnite rational matrix functions

If (9.4) is a right (respectively, left) spectral factorization of W with respect to the real line, we refer to L as the right (respectively, left ) spectral factor . Without further explanation a similar terminology will be used in comparable circumstances. Note that existence of a spectral factorization implies that W has no poles or zeros on the given contour and on the contour it is positive deﬁnite. The converse also holds: for positive deﬁnite rational matrix functions, both left and right spectral factorizations exist. This will now be proved for the case when W is a proper rational m × m matrix function. Moreover, explicit formulas for the factors will be given in terms of a realization of W . First we consider the situation where W is positive deﬁnite on the real line. Theorem 9.4. Let W (λ) = D + C(λIn − A)−1 B be a realization of the rational m × m matrix function W . Suppose A has no real eigenvalues, W is positive deﬁnite on the real line, and W (∞) = D is positive deﬁnite too. Further assume there exists an invertible Hermitian n × n matrix H for which HA = A∗ H and HB = C ∗ . Then, with respect to the real line, W admits right and left spectral factorization. Such factorizations can be obtained in the following way. Let M+ and M− be the spectral subspaces of A associated with the parts of σ(A) lying in × × the lower and upper half plane, respectively, and let M+ and M− be the spectral × × subspaces of A associated with the parts of σ(A ) lying in the lower and upper half plane, respectively. Then × ˙ M+ Cn = M− + ,

× ˙ M− Cn = M+ + .

(9.5)

× , Π− for the projection of Write Π+ for the projection of Cn along M− onto M+ × n C along M+ onto M− , and introduce

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

(9.6)

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

(9.7)

Then ¯ ∗ L+ (λ), W (λ) = L+ (λ)

¯ ∗ L− (λ), W (λ) = L− (λ)

are right and left spectral factorizations with respect to the real line, respectively. These spectral factorizations are uniquely determined by the fact that they have the value D 1/2 at inﬁnity. The conditions of the theorem are satisﬁed in case W has no poles on the real line, W (λ) is positive deﬁnite for all real λ, and the given (biproper) realization of W is a minimal one. Proof. The invertibility of W (λ) for real λ combined with the fact that A has no real eigenvalues implies that A× does not have real eigenvalues either (see Theorem × × 2.4). So the subspaces M+ , M− , M+ and M− are well-deﬁned. Let P and P × be

9.2. Spectral factorization

187

the Riesz projections of A, and A× , respectively, with respect to the upper half plane. From HA = A∗ H and HA× = (A× )∗ H one easily computes that HP = (I − P ∗ )H,

HP × = (I − P × )∗ H.

× × and M− satisfy It follows that the spaces M+ , M− , M+ ⊥ HM+ = M+ ,

⊥ HM− = M− ,

× ×⊥ HM+ = M+ ,

× ×⊥ HM− = M− . (9.8)

× × = {0}. Suppose x ∈ M+ ∩ M− . As M+ First it will be shown that M+ ∩ M− is invariant under A, we have Ax ∈ M+ . But then the ﬁrst identity in (9.8) shows × that HAx, x = 0. The space M− is invariant under A× . Thus A× x belongs to × M− , and the last identity in (9.8) yields HA× x, x = 0. Hence

0 = H(A − A× )x, x = HBD−1 Cx, x = D −1 Cx, Cx = D−1/2 Cx2 . As D > 0, it follows that Cx = 0. Thus A× x = (A−BD−1 C)x = Ax. We conclude × that M+ ∩ M− is invariant under both A and A× , and we have A|M+ ∩M × = −

A× |M+ ∩M × . However, −

σ(A|M+ ∩M × ) ⊂ σ(A|M+ ) ⊂ {λ | λ > 0}, −

σ(A× |M+ ∩M × ) ⊂ σ(A× |M × ) ⊂ {λ | λ < 0}. −

−

× = {0}. Thus A|M+ ∩M × = A |M+ ∩M × implies that M+ ∩ M− − − Proving (9.5) is now done via a dimension argument. Since H is invertible, ⊥ have the same dimension. In the ﬁrst identity in (9.8) shows that M+ and M+ × particular, dim M+ = n/2. Similarly, the last identity in (9.8) yields dim M− = n/2. Hence the ﬁrst identity in (9.5) holds. Let Π− be the projection along M+ × onto M− . The second identity in (9.5) is established in a similar way. × Let Π− be the projection along M+ onto M− . Then Π− is a supporting projection, and by Theorem 3.2 the corresponding factorization is a left canonical factorization given by W (λ) = K− (λ)L− (λ), ×

where L− is given by (9.7), and K− (λ) = D1/2 + C(λ − A)−1 (I − Π− )BD−1/2 . ¯ ∗ . Using (9.7) and (9.1) we have It remains to prove that K− (λ) = L− (λ) ¯ ∗ L− (λ)

=

D1/2 + B ∗ (λ − A∗ )−1 Π∗− C ∗ D−1/2

=

D1/2 + C(λ − A)−1 H −1 Π∗− HBD−1/2 .

¯ ∗ , it suﬃces to show that H(I − Π− ) = Π∗ H. Thus in order to get K− (λ) = L− (λ) −

188

Chapter 9. Factorization of positive deﬁnite rational matrix functions

Using the deﬁnition of Π− , together with the ﬁrst and the last identity in (9.8), we see that H(I − Π− )x, (I − Π− )y = 0 and HΠ− x, Π− y = 0 for all x and y in Cn . Hence for all x, y, H(I − Π− )x, y = H(I − Π− )x, Π− y = Hx, Π− y, which yields the desired identity H(I − Π− ) = Π∗− H. As for the last statement in the theorem, recall that the factors in a spectral factorization are uniquely determined up to multiplication with a constant unitary matrix. This settles the theorem as far as left spectral factorization is concerned. For right spectral factorizations the reasoning is similar. With minor modiﬁcations one proves the following theorem concerning left and right spectral factorizations with respect to the imaginary axis. Theorem 9.5. Let W (λ) = D + C(λIn − A)−1 B be a realization of the rational m × m matrix function W . Suppose A has no pure imaginary eigenvalues, W is positive deﬁnite on the imaginary axis, and W (∞) = D is positive deﬁnite too. Further assume there exists an invertible skew-Hermitian n × n matrix H for which HA = −A∗ H and HB = C ∗ . Then, with respect to the imaginary axis, W admits right and left spectral factorization. Such factorizations can be obtained in the following way. Let M+ and M− be the spectral subspaces of A associated with × the parts of σ(A) lying in the right and left half plane, respectively, and let M+ × × × and M− be the spectral subspaces of A associated with the parts of σ(A ) lying in the right and left half plane, respectively. Then × ˙ M+ Cn = M − + ,

× ˙ M− Cn = M+ + .

× Write Π+ for the projection of Cn along M− onto M+ , Π− for the projection of × Cn along M+ onto M− , and introduce

Then

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

¯ ∗ L+ (λ), W (λ) = L+ (−λ)

¯ ∗ L− (λ), W (λ) = L− (−λ)

are right and left spectral factorizations with respect to the imaginary axis, respectively. These spectral factorizations are uniquely determined by the fact that they have the value D 1/2 at inﬁnity. The conditions of the theorem are satisﬁed in case W has no poles on the imaginary axis, W (λ) is positive deﬁnite for λ ∈ iR, and the given (biproper) realization of W is a minimal one. In terms of the theory of spaces with an indeﬁnite metric (see the appendix at the end of this chapter), the identities in (9.8) say × × and M− are Lagrangian subspaces in that the spectral subspaces M+ , M− , M+ the indeﬁnite inner product induced by H.

9.3. Positive deﬁnite functions on the unit circle

189

9.3 Positive deﬁnite functions on the unit circle In this section we shall discuss rational matrix functions that take positive deﬁnite values on the unit circle T and their spectral factorizations. This class of functions is more complicated than the ones discussed in the previous sections, the main reason being that inﬁnity is not on the contour, and so the value at inﬁnity is not necessarily a selfadjoint matrix. A rational m × m matrix function W is called selfadjoint on the unit circle or Hermitian on the unit circle if for each λ ∈ T, λ not a pole of W , the matrix W (λ) is selfadjoint or, which is the same, Hermitian. By the uniqueness theorem for analytic functions, a rational matrix function W is selfadjoint on T if and ¯ −1 )∗ , for all λ ∈ C, λ not a pole of W . It follows that if W only if W (λ) = W (λ is selfadjoint on T and det W (λ) does not vanish identically, then W −1 is also selfadjoint on T. We ﬁrst discuss how selfadjointness of W is reﬂected in properties of the matrices in a minimal realization of the function. For proper rational matrix functions this is described in the following theorem, a counterpart of Theorem 9.1 for the unit circle. Theorem 9.6. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function. Then W is Hermitian on T if and only if A is invertible, D∗ = D − CA−1 B, and there exists an n × n matrix H such that HA = A−∗ H,

HB = A−∗ C ∗ ,

H = −H ∗ .

(9.9)

The matrix H is uniquely determined by the matrices in the given realization of W and invertible. Recall that A−∗ stands for (A∗ )−1 or, which amounts to the same, (A−1 )∗ . The ﬁrst part of (9.9) means that A is iH-unitary, that is, A is unitary with respect to the indeﬁnite inner product induced by the selfadjoint matrix iH (cf., Chapter 11 and Section 17.1). The ﬁrst part of (9.9) can be rewritten as A∗ HA = H. Note that, given the invertibility of H, the identity A∗ HA = H implies the invertibility of A. Proof. First observe that if W is Hermitian on T, then W has no pole at 0, as W (∞) = D and W (0) = W (∞)∗ . Because of minimality, this shows that A is invertible and D∗ = D − CA−1 B. But then ¯ −1 )∗ W (λ

=

D∗ + B ∗ (λ−1 − A∗ )−1 C ∗

=

D∗ − B ∗ A−∗ (λ − A−∗ )−1 λC ∗

=

D∗ − B ∗ A−∗ C ∗ − B ∗ A−∗ (λ − A−∗ )−1 A−∗ C ∗ .

¯ −1 )∗ coincide. Thus, again by Now the rational matrix functions W (λ) and W (λ the state space similarity theorem (Theorem 8.2), there exists a unique invertible

190

Chapter 9. Factorization of positive deﬁnite rational matrix functions

matrix H such that HA = A−∗ H,

HB = A−∗ C ∗ ,

−B ∗ A−∗ H = C.

Taking adjoints and employing the uniqueness of H, one ﬁnds H = −H ∗ . This settles the “only if part” of the theorem; the “if part” is obtained via a straightforward computation (not using minimality). Because of minimality, the identities in (9.9) imply that H is invertible. The argument is similar to that given in the third paragraph of the proof of Theorem 9.1. Next, we consider the associate main operator. Proposition 9.7. Let W (λ) = D + C(λIn − A)−1 B be a realization of an m × m rational matrix function and assume D is invertible. Suppose A is invertible too, D∗ = D − CA−1 B, and there exists an n × n matrix H such that (9.9) holds. Then A× = A − BD−1 C is invertible and HA× = (A× )−∗ H. Proof. From the invertibility of A and D, and the assumption D∗ = D − CA−1 B, we get A 0 A B I 0 I A−1 B = 0 D∗ 0 I C D CA−1 I =

I

BD −1

0

I

A×

0

0

D

I

0

D−1 C

I

.

As both A and D∗ are invertible, A× must be invertible too. Furthermore, by (9.9), we have (A× )∗ HA×

=

(A∗ − C ∗ D−∗ B ∗ )H(A − BD −1 C)

=

H − C ∗ D−∗ B ∗ HA − A∗ HBD−1 C + C ∗ D−∗ B ∗ HBD−1 C

=

H + C ∗ D−∗ (D − D∗ + B ∗ HB)D−1 C

=

H + C ∗ D−∗ (CA−1 B + B ∗ A−∗ C ∗ )D−1 C.

However, as D − D∗ = CA−1 B, we have B ∗ A−∗ C ∗ = −CA−1 B. Therefore, (A× )∗ HA× = H. Next we analyze how the matrix H appearing in Theorem 9.6 behaves under a state space similarity transformation on the realization of W . The proof of the next theorem is analogous to the proof of Theorem 9.3. Theorem 9.8. For i = 1, 2, let W (λ) = D + Ci (λIn − Ai )−1 Bi be minimal realizations of the rational m× m matrix function W , and let S be the (unique invertible) n × n matrix such that SA1 = A2 S,

C1 = C2 S,

B2 = SB1 .

9.3. Positive deﬁnite functions on the unit circle

191

Suppose W is Hermitian on the unit circle. For i = 1, 2, write Hi for the (unique invertible) skew-Hermitian n × n matrix such that A∗i Hi Ai = Hi and Hi Bi = ∗ ∗ A−∗ i Ci . Then H1 = S H2 S. The above results can also be obtained by reduction to the real line results of Section 9.1. To illustrate this, let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function, and let α ∈ T be a regular point for A, that is, α is not an eigenvalue of A. Consider the M¨obius transformation φ(λ) = α(λ − i)(λ + i)−1 .

(9.10)

Note that φ maps the upper half plane in a one-to-one way onto the open unit disc D, and the extended real line is mapped in a one-to-one way onto the unit circle ! (λ) = W (φ(λ)). Then W ! is again an m × m rational T, with φ(∞) = α. Put W ! admits the realization matrix function and (see Section 3.6 in [20]) the function W −1 ! W (λ) = D + C(λIn − A) B, where = (iα + iA)(α − A)−1 = φ−1 (A), A = 2iαC(α − A)−1 , C

= (α − A)−1 B, B

= W (α) = D + C(α − A)−1 B. D

Moreover this realization is again minimal. Now assume that W is selfadjoint on ! is selfadjoint on R and by Theorem 9.1 there exists an invertible n × n T, then W such that matrix H A =A ∗ H, H

B =C ∗, H

=H ∗. H

Observe that A =A ∗ H H

α − A∗ )−1 (−iα ¯ − iA∗ )H ⇐⇒ H(iα + iA)(α − A)−1 = (¯ − A) ⇐⇒ (¯ α − A∗ )H(iα + iA) = (−iα ¯ − iA∗ )H(α + iα − iαA∗ H − iA∗ HA ⇐⇒ iH ¯ HA + iα − iαA∗ H + iA∗ HA = −iH ¯ HA = 2iA∗ HA. ⇐⇒ 2iH

We already know (see the ﬁrst paragraph of the proof of Theorem 9.6) that the Using this = A−∗ H. operator A is invertible, and thus we may conclude that HA and the invertibility of H, one gets B H

=

−1 − AH −1 )−1 B − A)−1 B = (αH H(α

=

−1 A−∗ )−1 B = (α − A−∗ )−1 HB −1 − H (αH

=

= −α ¯ (¯ α − A∗ )−1 A∗ HB. (αA∗ − In )−1 A∗ HB

192

Chapter 9. Factorization of positive deﬁnite rational matrix functions

we know that C ∗ = −2iα From the deﬁnition of C ¯ (¯ α − A∗ )−1 C ∗ , and hence = 2iC ∗ . B =C ∗ ⇐⇒ A∗ HB H Then H has the properties listed in (9.9). Now deﬁne H by 2iH = H. In a similar way it can be shown that Proposition 9.7 and Theorem 9.8 follow from the analogous results in Section 9.1. We now turn to spectral factorization. Suppose W is a rational m× m matrix function. A factorization ¯ −1 )∗ L(λ) W (λ) = L(λ (9.11) is called a right spectral factorization with respect to the unit circle if L and L−1 are rational matrix functions which are analytic on the closure of the (open) unit ¯ −1 )∗ and its inverse are analytic on the closure disc D. In that case the function L(λ of Dext , the exterior domain of the unit circle in C (inﬁnity included). Thus, in particular, a right spectral factorization with respect to the unit circle is a right canonical factorization with respect to T. Analogously, (9.11) is called a left spectral factorization with respect to the unit circle if L and L−1 are analytic on the closure ¯ −1 )∗ and its inverse are of Dext (inﬁnity included), in which case the function L(λ analytic on the closed unit disc. Such a factorization is a left canonical factorization with respect to T. Observe that the existence of a spectral factorization implies that W has positive deﬁnite values on the unit circle. As we will see in the next theorem, the converse is also true. A rational m × m matrix function W is called positive deﬁnite on the unit circle if for each λ ∈ T, λ not a pole of W , the matrix W (λ) is positive deﬁnite. Left and right spectral factorization of functions which are positive deﬁnite on the unit circle is slightly more complicated than spectral factorization of functions which are positive deﬁnite on either the real line or the imaginary axis. This is mainly caused by the fact that the value at inﬁnity generally is no longer positive deﬁnite. Theorem 9.9. Let W (λ) = D + C(λIn − A)−1 B be a realization of a rational m × m matrix function such that W (λ) is positive deﬁnite for |λ| = 1. Suppose D is invertible, A is invertible, and A has no eigenvalues on the unit circle. Furthermore, assume there exists an invertible skew-Hermitian n × n matrix H such that HA = A−∗ H and HB = A−∗ C ∗ . Then, with respect to the unit circle, W admits right and left spectral factorization. Such factorizations can be obtained in the following way. Let M+ and M− be the spectral subspaces of A associated with × × the parts of σ(A) lying in Dext and D, respectively, and let M+ and M− be the spectral subspaces of A× associated with the parts of σ(A× ) lying in Dext and D, respectively. Then × ˙ M+ Cn = M − + ,

× ˙ M− Cn = M+ + .

(9.12)

9.3. Positive deﬁnite functions on the unit circle

193

× Write Π+ for the projection of Cn along M− onto M+ , Π− for the projection of Cn × −1 along M+ onto M− . Then D+ = D − CA (I − Π+ )B and D− = D − CA−1 (I − Π− )B are selfadjoint. Further there are unique rational matrix functions L+ and L− such that

¯ −1 )L+ (λ), W (λ) = L+ (λ

¯ −1 )L− (λ) W (λ) = L− (λ

are right and left spectral factorizations with respect to the unit circle, respectively, 1/2 1/2 and such that L+ (∞) = D+ , L− (∞) = D− . These functions are given by 1/2

1/2

1/2

1/2

L+ (λ)

=

D+ + D+ D−1 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D− + D− D−1 CΠ− (λIn − A)−1 B.

(9.13) (9.14)

The conditions of the theorem are satisﬁed in case W has no poles on the unit circle, takes positive deﬁnite values there, and the given (biproper) realization of W is a minimal one. Proof. Our hypotheses imply that A and A× do not have eigenvalues on the unit circle. Let P and P × be the Riesz projections of A and A× , respectively, corresponding to the eigenvalues in Dext . As in the proof of Theorem 9.4, one ﬁrst shows that HP = (I − P ∗ )H and HP × = (I − P × )∗ H, using A∗ HA = H and × × (A× )∗ HA× = H. Hence for the subspaces M+ , M− , M+ and M− we again have the identities ⊥ , HM+ = M+

⊥ HM− = M− ,

× ×⊥ HM+ = M+ ,

× ×⊥ HM− = M− .

Now introduce ϕ(λ) = −i(λ − i)(λ + i)−1 (i.e., (9.10) with α = −i). Observe that ϕ−1 = ϕ, and ϕ maps the circle to the real line, D to the open upper half plane and Dext to the open lower half plane. Consider V (λ) = W (ϕ(λ)). Then V (λ) is positive deﬁnite on the real line and − A) −1 B V (λ) = W (−i) + C(λ = (I + iA)(−A − iI)−1 . Since W (−i) is invertible, we can use Proposition with A 3.4 in [20] to show that the associate main matrix in the above realization of V is × = (I + iA× )(−A× − iI)−1 . Using A∗ HA = H and (A× )∗ HA× = H given by A with and A × are H-selfadjoint. The spectral subspaces of A one computes that A respect to the upper and lower half planes are M− and M+ , respectively, while × with respect to the upper and lower half planes are the spectral subspaces of A × × M− and M+ , respectively. From the proof of Theorem 9.4, it now follows that (9.12) holds. So the projections Π+ and Π− are well-deﬁned, and they are supporting projections giving rise to right and left canonical factorizations, respectively. Moreover H(I − Π+ ) = Π∗+ H, and H(I − Π− ) = Π∗− H.

194

Chapter 9. Factorization of positive deﬁnite rational matrix functions A canonical factorization corresponding to Π+ is given by W = W− W+ where W− (λ)

= D + C(λ − A)−1 (I − Π+ )B,

W+ (λ)

= I + D−1 CΠ+ (λ − A)−1 B.

For later use, recall that the factors W− and W+ in a canonical factorization are uniquely determined by their values at inﬁnity. It remains to show that from 1/2 −1/2 ¯ −1 )∗ . We shall in fact = L+ (λ L+ (λ) = D+ W+ (λ) it follows that W− (λ)D+ −1 ∗ ¯ prove that W (λ) = W+ (λ ) D+W+ (λ). Observe that D+ = W− (0). To see that D+ is selfadjoint, just carry out the calculation ∗ D+

= D∗ − B ∗ (I − Π∗+ )A−∗ C ∗ = D∗ − B ∗ A−∗ C ∗ + B ∗ Π∗+ A−∗ C ∗ = D − CH −1 A∗ Π∗+ HB = D − CA−1 H −1 Π∗+ HB = D − CA−1 (I − Π+ )B = D+ .

Then write W (λ) = K(λ)D+ W+ (λ) with −1 −1 K(λ) = DD+ + C(λ − A)−1 (I − Π+ )BD+ .

¯ −1 )∗ : Now compute W+ (λ ¯ −1 )∗ W+ (λ

=

I + B ∗ (λ−1 − A∗ )−1 Π∗+ C ∗ D−∗

=

I − B ∗ A−∗ Π∗+ C ∗ D −∗ −B ∗ A−∗ (λ − A−∗ )−1 A−∗ Π∗+ C ∗ D −∗

=

(D∗ − B ∗ A−∗ Π∗+ C ∗ )D −∗ +C(λ − A)−1 AH −1 Π∗+ HA−1 BD−∗ .

We claim that (D ∗ − B ∗ A−∗ Π∗+ C ∗ )D −∗ AH −1 Π∗+ HA−1 BD−∗

−1 = DD+ ,

(9.15)

−1 = (I − Π+ )BD+ .

(9.16)

Indeed, for (9.15), observe that D+ is invertible because W (0) = D∗ is invertible, and D ∗ = D+ W+ (0) = D+ (I − D −1 CΠ+ A−1 B). −1 ∗ So D+ D = D−1 (D − CΠ+ A−1 B). Taking adjoints yields (9.15).

9.3. Positive deﬁnite functions on the unit circle

195

−1 ∗ To verify (9.16), compute (I −Π+ )BD+ D , using what we just have proved: −1 ∗ D (I − Π+ )BD+

= (I − Π+ )B(I − D−1 CΠ+ A−1 B) = (I − Π+ )(A − BD−1 CΠ+ )A−1 B = (I − Π+ )(A − (A − A× )Π+ )A−1 B = (I − Π+ ){A(I − Π+ ) + A× Π+ }A−1 B.

Now Im Π+ is A× -invariant, so (I − Π+ )A× Π+ = 0. Hence (I − Π+ )BD+ D∗

=

(I − Π+ )A(I − Π+ )A−1 B

=

A(I − Π+ )A−1 B

=

AH −1 Π∗+ HA−1 B,

as Ker Π is A-invariant. Thus (9.16) holds. Now using (9.15) and (9.16) we see ¯ −1 )∗ = DD−1 + C(λ − A)−1 (I − Π+ )BD−1 = K(λ). W+ (λ + + ¯ −1 )∗ D+ W+ (λ) we see that D+ must be positive deﬁnite. Since As W (λ) = W+ (λ the factors W+ and W− in a canonical factorization are uniquely determined by their values at inﬁnity, it follows that the factor L+ in a right spectral factorization is also uniquely determined by its value at inﬁnity. Thus the part of the theorem concerned with right spectral factorization follows. For the other part dealing with left spectral factorization the reasoning is similar.

Notes The results of Section 9.1 can be found in several sources, e.g., [26] and [45]. The factorization results of Sections 9.2 and 9.3 are based on [119] (see also Chapter 1 in [120]). Spectral factorizations play an important role in mathematical systems theory, see e.g., [4]. In [4], [41] and [147] spectral factorizations of a selfadjoint rational matrix function W are studied in state space form, starting from diﬀerent representations of W . Part IV of [20] is devoted to stability of minimal factorizations of rational matrix functions. The issue of stability of factorizations within the class of spectral factorizations has also been studied. This requires the analysis of perturbations of H-selfadjoint matrices and stability of their invariant Lagrangian subspaces. For instance, from Theorem 14.12 in [20] it follows straightforwardly that canonical factorizations are Lipschitz stable under small perturbations of the matrices in the realization. Restricting attention to spectral factorizations of positive deﬁnite rational matrix functions, and to perturbations of the matrices in the realizations

196

Chapter 9. Factorization of positive deﬁnite rational matrix functions

that make the perturbed rational matrix function also positive deﬁnite, it still holds that spectral factorization is Lipschitz stable in this sense. For these and related results we refer to [123], see also [127].

Chapter 10

Pseudo-spectral factorizations of selfadjoint rational matrix functions In this chapter we consider rational matrix functions on a contour having values that are selfadjoint matrices, but not necessarily positive deﬁnite ones. Whereas in the previous chapter we studied spectral factorization, in the present chapter the focus will be on functions that have poles or zeros on the contour, and so we will consider pseudo-spectral factorization here. This chapter consists of two sections. Section 10.1 develops the notion of pseudo-spectral factorization for nonnegative rational matrix functions. The contours considered are the real line, the imaginary axis and the unit circle. In Section 10.2 the main result of the ﬁrst section is generalized to the case of arbitrary selfadjoint rational matrix functions with positive deﬁnite value at inﬁnity.

10.1 Nonnegative rational matrix functions In this section we consider rational matrix functions W having nonnegative values on either the real line, the imaginary axis or the unit circle. The section may be viewed as a continuation of the discussion in Chapter 9. However, in contrast to the situation there, in this section we consider cases where W may have poles or zeros on the contour. A rational m × m matrix function W is called nonnegative on the real line if for each λ ∈ R, λ not a pole of W , the matrix W (λ) is nonnegative. Without further explanation, the analogous terminology will be used for rational matrix functions having nonnegative values on the imaginary axis or on the unit circle, respectively.

198

Chapter 10. Pseudo-spectral factorizations

As in Section 9.2 we shall start by considering the case of nonnegative rational matrix functions W on the real line, and continue with the situation where W is nonnegative on the imaginary axis. However, it is the latter case that we shall use frequently in the subsequent chapters. Therefore only for this case shall we provide a detailed proof. The real line situation can then be dealt with by using the M¨obius transformation λ → −iλ. The section is concluded by presenting the results for the case of the unit circle. Again, the proof may be obtained by using a M¨ obius transformation (cf., the proofs of Theorems 9.4 and 9.9). A factorization ¯ ∗ L(λ) W (λ) = L(λ) (10.1) is called a right pseudo-spectral factorization with respect to the real line if L has no poles or zeros in the open upper half plane and the factorization is locally minimal at each point of the real line. Analogously, (10.1) is called a left pseudospectral factorization with respect to the real line if L has no poles or zeros in the open lower half plane and the factorization is locally minimal at each point of the real line. Such right or left pseudo-spectral factorizations are pseudo-canonical factorizations with respect to iR in the sense of Section 8.3. Although a nonnegative rational matrix function generally does not allow for a left or right spectral factorization, it does admit left and right pseudo-spectral factorization. Theorem 10.1. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a rational m × m matrix function which is nonnegative on the real line, and assume D is positive deﬁnite. Then, with respect to the real line, W admits left and right pseudo-spectral factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) Hermitian n × n matrix with HA = A∗ H and HB = C ∗ . Then there are unique A-invariant subspaces M+ and M− , and × × unique A× -invariant subspaces M+ and M− , such that (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane, and σ(A|M+ ) ⊂ {λ | λ ≤ 0}, (ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane, and σ(A|M− ) ⊂ {λ | λ ≥ 0}, × (iii) M+ contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open lower half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}, +

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open upper half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0}, −

(v) H[M+ ] =

⊥ M+ ,

H[M− ] =

⊥ M− ,

× H[M+ ]

×⊥ = M+ ,

× ×⊥ H[M− ] = M− .

The subspaces in question also satisfy the matching conditions × ˙ M+ Cn = M − + ,

× ˙ M− Cn = M+ + .

(10.2)

10.1. Nonnegative rational matrix functions

199

× Let Π+ be the projection along M− onto M+ , let Π− be the projection along M+ × onto M− , and introduce

Then

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

(10.3)

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

(10.4)

¯ ∗ L+ (λ), W (λ) = L+ (λ)

¯ ∗ L− (λ), W (λ) = L− (λ)

are right and left pseudo-spectral factorizations with respect to the real line, respectively. These pseudo-spectral factorizations are uniquely determined by the fact that they have the value D 1/2 at inﬁnity. All possible right pseudo-spectral factors can be obtained from L+ as given in (10.3) by multiplying on the left with a unitary matrix, and likewise, all possible left pseudo-spectral factors are obtained from L− as given in (10.4) by multipli¯ ∗L − (λ) − (λ) cation on the left with a unitary matrix. Indeed, suppose W (λ) = L −∗ ¯ ¯ ∗= is another left pseudo-spectral factorization of W . Put E(λ) = L− (λ) L− (λ) −1 − (λ)L− (λ) . Then E(λ) is analytic outside the real line, and on the real line it L is unitary, except for possible poles. So for all values of λ concerned, the norm of E(λ) is 1. But then E cannot have poles. Indeed, in the vicinity of a pole the norm of E(λ) cannot be bounded (cf., [134], Chapter 10, page 211). It follows that E is analytic on the whole complex plane. But then it must be a constant function by Liouville’s theorem. As it is unitary for real λ, we conclude that the sole value of E is a unitary matrix. Let W be a rational m × m, and suppose W is nonnegative on the real line. A factorization ¯ ∗ L(λ) W (λ) = L(−λ) is called a right pseudo-spectral factorization with respect to the imaginary axis if L has no poles or zeros in the open left half plane and the factorization is locally minimal at each point of the imaginary axis. Left pseudo-spectral factorizations with respect to the imaginary axis are deﬁned by replacing the upper half plane by the lower half plane. Theorem 10.2. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is nonnegative on the imaginary axis, and assume D is positive deﬁnite. Put A× = A − BD −1 C. Then, with respect to the imaginary axis, W admits left and right pseudo-spectral factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) skew-Hermitian n × n matrix with HA = −A∗ H and HB = C ∗ . Then there are unique A-invariant subspaces M+ and M− , and unique A× -invariant subspaces × × M+ and M− , such that (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open right half plane, and σ(A|M+ ) ⊂ {λ | λ ≥ 0},

200

Chapter 10. Pseudo-spectral factorizations

(ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open left half plane, and σ(A|M− ) ⊂ {λ | λ ≤ 0}, × contains the spectral subspace of A× associated with the part of σ(A× ) (iii) M+ lying in the open right half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0},, +

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open left half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}, −

(v) H[M+ ] =

⊥ M+ ,

H[M− ] =

⊥ M− ,

× H[M+ ]

×⊥ = M+ ,

× ×⊥ H[M− ] = M− .

The subspaces in question also satisfy the matching conditions × ˙ M+ , Cn = M − +

× ˙ M− Cn = M+ + .

× Let Π+ be the projection of Cn along M− onto M+ , let Π− be the projection of × n C along M+ onto M− , and deﬁne L+ and L− by (10.3) and (10.4), that is

Then

L+(λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

¯ ∗ L+ (λ), W (λ) = L+ (−λ)

¯ ∗ L− (λ), W (λ) = L− (−λ)

(10.5)

are right and left pseudo-spectral factorizations with respect to the imaginary axis, respectively. These pseudo-spectral factorizations are the unique ones for which L+ (∞) = D1/2 and L− (∞) = D1/2 . As was noted before, Theorem 10.1 can be derived from Theorem 10.2 via the transformation λ → −iλ. Conversely, Theorem 10.2 obtained from Theorem 10.1 by the transformation λ → iλ. Before we prove Theorem 10.2 we need some preparations concerning the spectral properties of nonnegative rational matrix functions. First we discuss the partial pole-multiplicities and partial zero-multiplicities of W . These notions have been deﬁned in Sections 8.2 and 8.1 of [20], respectively. We start with a minimal realization W (λ) = D + C(λ − A)−1 B. (10.6) Assume that W is biproper, i.e., D is invertible. Then the eigenvalues of A coincide with the poles of W and the eigenvalues of A× coincide with the zeros of W . More precisely, the partial multiplicities of λ as an eigenvalue of A coincide with the partial pole-multiplicities of λ as a pole of W , and the multiplicities of λ as an eigenvalue of A× coincide with the partial zero-multiplicities of λ as a zero of W (cf., [20], Section 8.4, in particular Proposition 8.23). We also need the connection between the Jordan chains of A at an eigenvalue λ0 and the co-pole functions of W at λ0 described in Proposition 8.3. For a nonnegative rational matrix function, we have the following addition to that proposition.

10.1. Nonnegative rational matrix functions

201

Proposition 10.3. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization for a rational m × m matrix function which is selfadjoint on the imaginary axis, and let H be the (unique) invertible skew-Hermitian n × n matrix such that HA = −A∗ H,

HB = C ∗ .

Let λ0 ∈ iR be an eigenvalue of A, let x0 , . . . , xk−1 be a Jordan chain for A at λ0 , and let ϕ be a co-pole function of W at λ0 corresponding to the Jordan chain ¯ has a zero of order at least k x0 , . . . , xk−1 . Then the function W (λ)ϕ(λ), ϕ(−λ) at λ0 and its Taylor expansion at λ0 has the following form: ¯ = (−1)k x0 , Hxk−1 (λ − λ0 )k + · · · W (λ)ϕ(λ), ϕ(−λ) · · · + (−1)k xk−1 , Hxk−1 (λ − λ0 )2k−1 + h.o.t. , where h.o.t. stands for higher order terms. Proof. The fact that ϕ is a co-pole function of W at λ0 implies that W (λ)ϕ(λ) is analytic at λ0 . This together with the fact that ϕ has a zero of order at least k at ¯ has a zero of order at least k at λ0 λ0 shows that the function W (λ)ϕ(λ), ϕ(−λ) too. The property that ϕ is a co-pole function of W at λ0 corresponding to the Jordan chain x0 , . . . , xk−1 means that xj =

∞

P0 (A − λ0 )ν−j−1 Bϕν ,

j = 0, . . . , k − 1

(10.7)

ν=k

(where the sum in the right-hand side of the identity is actually ﬁnite so that there is no convergence issue). Here P0 is the Riesz projection of A corresponding to the eigenvalue λ0 , and ϕν is the coeﬃcient of (λ − λ0 )ν in the Taylor expansion of ϕ at λ0 . We use this connection to compute Hxi , xk−1 . The fact that λ0 is in iR yields HP0 = P0∗ H. Indeed, since HAH −1 = −A∗ , we have that HP H −1 is the Riesz projection of −A∗ for the eigenvalue λ0 . Thus, using Proposition I.2.5 in [51], we get HP0 H −1 = P (−A∗ ; {λ0 }) = P (A∗ ; {−λ0 }) = P (A∗ ; {λ0 }) = P (A; {λ0 })∗ = P0∗ . Also, note that the vectors x0 , . . . , xk−1 belong to Im P0 . In particular, P0 xk−1 = xk−1 . Now use (10.7) and the identities HA = −A∗ H and HB = C ∗ . This gives, for i = 0, . . . , k − 1, Hxi , xk−1 = = =

∞

ν=k ∞

ν=k ∞

HP0 (A − λ0 )ν−i−1 Bϕν , xk−1 H(A − λ0 )ν−i−1 Bϕν , P0 xk−1 (−1)ν−i−1 ϕν , C(A − λ0 )ν−i−1 xk−1

ν=k

=

k+i

ν=k

(−1)ν−i−1 ϕν , Cxk−ν+i .

202

Chapter 10. Pseudo-spectral factorizations

From the ﬁnal paragraph of Section 8.1 we know that the vector Cxk−ν+1 is given by Cxk−ν+1 = (W ϕ)k−ν+i , where (W ϕ)j is the coeﬃcient of (λ − λ0 )j in the Taylor expansion of W (λ)ϕ(λ) at λ0 . So Hxi , xk−1 =

k+i

(−1)ν−i−1 ϕν , (W ϕ)k−ν+i ,

i = 0, . . . , k − 1.

(10.8)

ν=k

On the other hand we have ¯ = W (λ)ϕ(λ), ϕ(−λ)

∞

=k

(−1)ν (W ϕ)−ν , ϕν (λ − λ0 ) .

(10.9)

ν=k

Comparing formulas (10.8) and (10.9), we see that for i = 0, . . . , k − 1 the coeﬃcient of (λ − λ0 )k+i in the Taylor expansion of W (λ)ϕ(λ) at λ0 is given by (−1)i+1 xk−1 , Hxi . Now note that Hxi , xk−1 =

H(A − λ0 )k−i−1 xk−1 , xk−1

=

(−1)k−1−i Hxk−1 , (A − λ0 )k−1−i xk−1

=

(−1)k−1−i Hxk−1 , xi = (−1)k−1−i xk−1 , Hxi .

We conclude that (−1)i+1 xk−1 , Hxi = (−1)k Hxi , xk−1 , which completes the proof. Specializing to the case when W is nonnegative on iR we obtain the following result. Proposition 10.4. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization for a rational m × m matrix function which is nonnegative on iR. Assume D is positive deﬁnite, and let H be the (unique invertible) skew-Hermitian n × n matrix such that HA = −A∗ H and HB = C ∗ . Then the partial multiplicities corresponding to pure imaginary eigenvalues of A and A× are all even, the sign characteristic of (iA, iH) consists of the integers +1 only, and the sign characteristic of the pair (iA× , iH) consists of the integers −1 only. For the deﬁnition of the notion of sign characteristic the reader is referred to Section 11.2 below. Proof. Let us ﬁrst prove the proposition for the matrix A. Let λ0 = iμ0 be a pure imaginary eigenvalue of A, and let x0 , . . . , xk−1 be a maximal Jordan chain for A at λ0 . Then x0 , −ix1 , (−i)2 x2 , . . . , (−i)k−1 xk−1 is a Jordan chain of iA for its eigenvalue −μ0 . In fact, all Jordan chains of iA for −μ0 can be obtained in this way. Choose a Jordan basis for A such that relative to it the pair (iA, iH) is in canonical form (see Section 11.2). This means, in particular, that if x0 , . . . , xk−1 is a maximal Jordan chain of A for λ0 , which is part of this basis, then iHx0 , (−i)k−1 xk−1 =

10.1. Nonnegative rational matrix functions

203

ik Hx0 , xk−1 is either +1 or −1. The sequence of +1’s and −1’s, obtained in this manner, is the sign characteristic of the pair (iA, iH). Let x0 , . . . , xk−1 be as in the previous paragraph, and let ϕ(λ) = (λ − λ0 )k ϕk + (λ − λ0 )k+1 ϕk+1 + · · · be a corresponding co-pole function for W at λ0 . From Proposition 10.3 we know that on a neighborhood of λ0 ¯ = (λ − λ0 )k h(λ), W (λ)ϕ(λ), ϕ(−λ) where the scalar function h is analytic at λ0 and h(λ0 ) = (−1)k Hx0 , xk−1 . Consider the pure imaginary λ = iμ in this neighborhood. Rewriting the expression above in terms of μ − μ0 , and using the fact that W is nonnegative, one sees that k is even and (−i)k Hx0 , xk−1 > 0. This proves that the partial multiplicities corresponding to pure imaginary eigenvalues of A are even, and that the sign characteristic of the pair (iA, iH) consists of +1’s only. To prove the part of the proposition concerning A× , note that the function W (λ)−1 = D−1 − D−1 C(λ − A× )−1 BD−1 is nonnegative on iR too. Moreover, for this realization we have (−H)A× = −(A× )∗ (−H) and (−H)BD−1 = (−D −1 C)∗ . So, the corresponding indeﬁnite inner product is given by −H rather than H. The desired result now follows by basically repeating the argument given above. We now have all the equipment necessary for the proof of Theorem 10.2. Proof of Theorem 10.2. Based on Proposition 10.4 the existence and uniqueness × × , M− such that of A-invariant subspaces M+ , M− and A× -invariant subspaces M+ (i), (ii) and (iii) hold follow from Theorem 11.5 in Section 11.2 below. × To prove the ﬁrst equality in (10.2) one establishes M+ ∩ M− ⊂ Ker C as in × the proof of Theorem 9.4: use (9.2) instead of (9.1). Hence M+ ∩ M− is invariant for both A and A× . However, as the realization is minimal, an A-invariant subspace × contained in Ker C must be the zero space. Thus M+ ∩ M− = {0}. To show × × n ˙ M− it remains to note that dim M+ = dim M− = n/2. In a similar C = M+ + × ˙ M+ manner one gets Cn = M− + . × Denote by Π+ the projection along M− onto M+ , then Π+ is a supporting projection, and by Theorem 8.5 the factorization W (λ) = K(λ)L+ (λ), with L+ given by (10.3) and K(λ) = D1/2 + C(λ − A)−1 (I − Π+ )BD−1/2 , is minimal. Moreover, L+ has no poles in the open left half plane because Π+ A = Π+ AΠ+ . So L+ (λ) = D 1/2 + D−1/2 CΠ+ (λ − Π+ AΠ+ )−1 Π+ B.

204 Also

Chapter 10. Pseudo-spectral factorizations

−1/2 L−1 − CΠ+ (λ − Π+ A× Π+ )−1 Π+ BD−1/2 , + (λ) = D

¯ ∗ . Indeed, thus L+ has no zeros in the open left half plane. Finally, K(λ) = L+ (−λ) ¯ ∗ L+ (−λ)

= D 1/2 − B ∗ (λ + A∗ )−1 Π∗+ C ∗ D−1/2 = D 1/2 + C(λ − A)−1 H −1 Π∗+ HBD−1/2 .

As H[Ker Π+ ] = (Ker Π+ )⊥ and H[Im Π+ ] = (Im Π+ )⊥ , we have H −1 (Π+ )∗ H = I − Π+ . But then the factorization corresponding to Π+ is a right pseudo-spectral factorization. One proves in a similar way that Π− gives rise to a left pseudospectral factorization. Next, we introduce the notion of left and right pseudo-spectral factorizations with respect to the unit circle, Let W be a rational matrix function having nonnegative values on T. A factorization ¯ −1 )∗ L(λ) W (λ) = L(λ is called a right pseudo-spectral factorization with respect to the unit circle if L has no poles or zeros in the open unit disc and the factorization is locally minimal at each point of the unit circle. Left pseudo-spectral factorizations with respect to the unit circle are deﬁned by replacing the open unit disc D by Dext . In dealing with pseudo-spectral factorizations with respect to the unit circle, we discuss only a restricted class of rational matrix functions that are nonnegative on the unit circle, namely those which are biproper. Because of symmetry, this forces the function to have an invertible value at zero too. The restriction is induced by our methods, rather than by the problem itself. The following theorem can be obtained from using an appropriate M¨ obius transformation (cf., the proof of Theorem 9.9). Theorem 10.5. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a rational m × m matrix function which is nonnegative on the unit circle, and assume D and A are invertible. Then, with respect to the unit circle, W admits left and right pseudo-spectral factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) skew-Hermitian n × n matrix satisfying A∗ HA = H and A∗ HB = C ∗ . Then there are unique A-invariant subspaces × × M+ , M− and unique A× -invariant subspaces M+ , M− , such that (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open exterior of the unit disc, and σ(A|M+ ) ⊂ {λ | |λ| ≥ 1}, (ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open unit disc, and σ(A|M− ) ⊂ {λ | |λ| ≤ 1}, × contains the spectral subspace of A× associated with the part of σ(A× ) (iii) M+ lying in the open exterior of the unit disc, and σ(A× |M × ) ⊂ {λ | |λ| ≥ 1}, +

10.2. Selfadjoint rational matrix functions and further generalizations

205

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open unit disc, and σ(A× |M × ) ⊂ {λ | |λ| ≤ 1}, −

⊥ , (v) H[M+ ] = M+

× ×⊥ H[M+ ] = M+ ,

⊥ H[M− ] = M− ,

× ×⊥ H[M− ] = M− .

The subspaces in question also satisfy (10.2), i.e., × ˙ M− Cn = M+ + ,

× ˙ M+ Cn = M− + .

× , and let Π− be the projection Let Π+ be the projection of Cn along M− onto M+ × n of C along M+ onto M− , and deﬁne L+ and L− by (9.13) and (9.14), so 1/2

1/2

1/2

1/2

L+ (λ)

=

D+ + D+ D−1 CΠ+ (λIn − A)−1 B,

L− (λ)

=

D− + D− D−1 CΠ− (λIn − A)−1 B,

where D+ = D − CA−1 (I − Π+ )B and D− = D − CA−1 (I − Π− )B. Then ¯ −1 )∗ L+ (λ), W (λ) = L+ (λ

¯ −1 )∗ L− (λ), W (λ) = L− (λ

are right and left pseudo-spectral factorizations with respect to the unit circle, respectively. The functions L+ and L− are the unique right and left pseudo-spectral 1/2 1/2 factors, respectively, such that L+ (∞) = D+ and L− (∞) = D− .

10.2 Selfadjoint rational matrix functions and further generalizations The main result of Section 10.1 will be generalized here to the case of an arbitrary selfadjoint rational matrix function with positive deﬁnite value at inﬁnity. We start with the case of selfadjoint functions on the real line. Theorem 10.6. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is selfadjoint on the real line, and assume D is positive deﬁnite. Then, with respect to the real line, W admits right and left pseudo-canonical factorization. Such factorizations can be obtained in the following way. Let H be the (unique invertible) Hermitian n×n matrix such that HA = A∗ H and HB = C ∗ . Then there exist A-invariant subspaces M+ and M− , and A× × × and M− such that invariant subspaces M+ (i) M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane, and σ(A|M+ ) ⊂ {λ | λ ≤ 0}, (ii) M− contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane, and σ(A|M− ) ⊂ {λ | λ ≥ 0}, × (iii) M+ contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open lower half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}, +

206

Chapter 10. Pseudo-spectral factorizations

× (iv) M− contains the spectral subspace of A× associated with the part of σ(A× ) lying in the open upper half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0}, −

× × and M− are maximal (v) M+ and M− are maximal H-nonnegative, and M+ H-nonpositive.

The subspaces in question also satisfy × ˙ M− , Cn = M + +

× ˙ M+ Cn = M− + .

× Let Π+ be the projection of |BC n onto M+ along M− , and let Π− be the projection × n of C onto M− along M+ , and introduce

L− (λ)

=

D1/2 + C(λIn − A)−1 (I − Π+ )BD−1/2 ,

(10.10)

L+ (λ)

=

D1/2 + D−1/2 CΠ+ (λIn − A)−1 B,

(10.11)

K+ (λ)

=

D1/2 + C(λIn − A)−1 (I − Π− )BD−1/2 ,

(10.12)

K− (λ)

=

D1/2 + D−1/2 CΠ− (λIn − A)−1 B.

(10.13)

Then W (λ) = L− (λ)L+ (λ),

W (λ) = K+ (λ)K− (λ),

(10.14)

are right and left pseudo-canonical factorizations with respect to the real line, respectively. × × The subspaces M+ , M− , M+ and M− are not unique. In line with this, the uniqueness of the factorizations that we had at earlier occasions is lacking here. Also, not all pseudo-canonical factorizations for selfadjoint rational matrix functions are obtained in the way described in Theorem 10.6 . The theorem will be obtained from the more general result stated below.

Theorem 10.7. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is selfadjoint on the real line, and assume D is positive deﬁnite. Suppose D = D+ D− with D+ and D− m × m matrices (automatically invertible). Let H be the (unique invertible) Hermitian n× n matrix for which HA = A∗ H and HB = C ∗ . Let M+ be an A-invariant maximal Hnonnegative subspace, and let M− be an A× -invariant maximal H-nonpositive subspace. Then ˙ M− . Cn = M+ + (10.15) In that case, the projection Π of Cn along M+ onto M− is a supporting projection, and (hence) W admits a minimal factorization W (λ) = W+ (λ)W− (λ) with W+ and W− given by W+ (λ)

−1 = D+ + C(λIn − A)−1 (I − Π)BD− ,

W− (λ)

−1 = D− + D+ CΠ(λIn − A)−1 B.

10.2. Selfadjoint rational matrix functions and further generalizations

207

For the existence of A-invariant maximal H-nonnegative and maximal Hnonpositive subspaces, see Section 11.2 below. Proof. First we show that M+ ∩ M− = {0}. Choose x ∈ M+ ∩ M− . As M+ is nonnegative and M− is nonpositive, we have Hx, x = 0. On M+ the Schwartz inequality holds for the H-inner product. Since x ∈ M+ and Ax ∈ M+ , we get |HAx, x|2 ≤ HAx, Ax · Hx, x = 0. So for all x ∈ M+ ∩ M− we have HAx, x = 0. In the same way one shows that for all x ∈ M+ ∩ M− we have HA× x, x = 0. It follows that 0 = H(A − A× )x, x = HBD−1 Cx, x = C ∗ D −1 Cx, x = D −1/2 Cx2 , and hence M+ ∩ M− ⊂ Ker C. But then A× x = Ax − BCx = Ax for all x belonging to M+ ∩ M− , and so M+ ∩ M− is A-invariant. Hence CAn x = 0 for all x ∈ M+ ∩ M− and n = 0, 1, 2, . . . . So M+ ∩ M − ⊂

∞ ,

Ker CAj = {0}.

j=0

Now (see Section 11.2) every maximal nonnegative subspace has the same dimension as M+ . Also, for a maximal H-nonpositive subspace M− , the subspace ⊥ H −1 [M− ] is maximal H-nonnegative. Hence ⊥ ⊥ ] = dim M− = n − dim M− , dim M+ = dim H −1 [M−

and from this we get (10.15), i.e, the ﬁrst part of the theorem. To obtain the second part, apply Theorem 8.5. Proof of Theorem 10.6. For the existence of A-invariant subspaces M+ , M− and × × A× -invariant subspaces M+ , M− such that (i), (ii) and (iii) hold we refer to Section 11.2. The matching of the appropriate subspaces is an immediate consequence of Theorem 10.7. The factorizations (10.14), where the factors are given by (10.10)– (10.13) are minimal by Theorem 8.5. As in the proof of Theorem 10.2 one shows that L+ and K+ have no zeros or poles in the open upper half plane. In the same vein, L− and K− have no zeros or poles in the open lower half plane. Hence the factorizations in (10.14) are right and left pseudo-canonical factorizations, respectively. Analogues of Theorems 10.6 and 10.7 concerning rational matrix functions which are selfadjoint on the unit circle or imaginary axis can be derived too. An analogue of Theorem 10.7 also holds true if one takes M+ to be A-invariant maximal H-nonpositive (instead of maximal H-nonnegative) and M− to be A× invariant maximal H-nonnegative (instead of maximal H-nonpositive). A similar remark can be made concerning Theorem 10.6.

208

Chapter 10. Pseudo-spectral factorizations

We ﬁnish this section with a theorem concerning symmetric factorization of rational matrix functions which are nonnegative. Here we shall present only the case involving the imaginary axis. Theorem 10.8. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of an m × m rational matrix function which is nonnegative on iR. Assume D is positive deﬁnite, and let H be the (unique invertible) skew-Hermitian n × n matrix such that HA = −A∗ H and HB = C ∗ . Suppose M and M × are subspaces of CN for which A[M ] ⊂ M,

A× [M × ] ⊂ M × ,

H[M ] = M ⊥ ,

H[M × ] = M ×⊥ .

(10.16)

˙ M × . Let Π be the projection of Cn along M onto M × , and Then Cn = M + introduce (10.17) L(λ) = D 1/2 + D−1/2 CΠ(λIn − A)−1 B. Then ¯ ∗ L(λ) W (λ) = L(−λ)

(10.18)

is a minimal factorization. Conversely, given a minimal factorization (10.18), with L(∞) = D1/2 , the factor L is as in (10.17) for a supporting projection Π such that M = Ker Π and M × = Im Π satisfy (10.16). ˙ M ×. Proof. Let M and M × be as in the theorem. We shall show that Cn = M + The argument follows a (by now) familiar pattern. One ﬁrst shows that the intersection M ∩ M × is contained in Ker C (see, e.g., the proof of Theorem 9.4, or the proof of Theorem 10.7). Then M ∩ M × is both A-invariant and A× -invariant and contained in Ker C. By minimality (in fact observability) it follows that M ∩ M × = {0}. Since dim M = dim M × = n/2, we have the desired matching. Denote by Π the projection along M onto M × . Then Π is a supporting projection. Write the factorization of W corresponding to Π and the factorization D = D 1/2 D1/2 as W (λ) = K(λ)L(λ), where K(λ) =

D 1/2 + C(λ − A)−1 (I − Π)BD−1/2 ,

L(λ) =

D 1/2 + D−1/2 CΠ(λ − A)−1 B.

Arguing as in the proof of Theorem 9.4 we have Π∗ H = H(I − Π). Using also (9.2) ¯ ∗ = K(λ). it then follows easily that L(−λ) ¯ ∗ L(λ) is a minimal factorization with Conversely, suppose W (λ) = L(−λ) 1/2 L(∞) = D . Let Π be the corresponding supporting projection (which exists by ¯ ∗ , where L(λ) Theorem 8.5). From the fact that the left-hand factor K(λ) is L(−λ) ∗ is the right-hand factor, and using (9.2), we have Π H = H(I − Π). Thus both M = Ker Π and M × = Im Π satisfy (i) and (ii).

10.2. Selfadjoint rational matrix functions and further generalizations

209

Notes This chapter originates from [119] which deals with rational matrix functions that are selfadjoint on the real line. The term pseudo-canonical is from a later date, and is taken from [132]. The results presented here for nonnegative rational matrix functions on the unit circle are based on Section 3 of [104]. In this case, the restriction to W being invertible at inﬁnity and at zero may be lifted by considering a diﬀerent type of realization, namely, realizations of the type discussed in [79]. In mathematical systems theory also the following problem is of interest: given is a nonnegative rational matrix function W as in Theorem 10.8, without poles on the imaginary axis. One is looking for all possible factorizations W (λ) = ¯ ∗ L(λ), where L has all its poles in the open left half plane, but there is L(−λ) no condition on the zeros of L. This problem too sometimes goes by the name of “spectral factorization problem” and such factors L are sometimes also called “spectral factors”. The problem of parametrizing such factors is considered in many papers and books, see, e.g., [116] and [46] and the references given there. The papers [30], [31], provide a discussion involving computational aspects. For matrix polynomials a similar problem is considered in the literature, see e.g., [88] and [66]. For later developments on factorization of selfadjoint matrix polynomials, see [103], [125]. In [20] stability of factorizations of rational matrix functions under small perturbations of the matrices in a realization is studied. For the particular case where the function is positive semideﬁnite on the real line, and the factorizations are of the type (10.1), stability under small perturbations is treated in [123]. This involves stability of invariant Lagrangian subspaces for matrices that are selfadjoint in a space with an indeﬁnite inner product. It turns out that the left and right pseudo-spectral factorizations are stable (see Theorem 2.5 in [123]).

Chapter 11

Review of the theory of matrices in indeﬁnite inner product spaces In this chapter we present some background material on matrices in indeﬁnite inner product spaces, and review the main results from this area that are used in this book. No proofs will be provided; we refer to the literature for more information. Good sources are [68] and [70]. The material is not only useful for understanding of the results of the preceding two chapters, but is also intended for use in subsequent chapters. This chapter consists of three sections. Section 11.1 considers subspaces that are negative, positive or neutral relative to an indeﬁnite inner product and various generalizations of such subspaces. Section 11.2 deals with matrices that are selfadjoint relative to an indeﬁnite inner product, and Section 11.3 with matrices that are dissipative relative to an indeﬁnite inner product.

11.1 Subspaces of indeﬁnite inner product spaces Let H be an invertible Hermitian n × n matrix. On Cn we denote the usual inner product with ·, ·. The indeﬁnite inner product given by H is deﬁned as follows: [x, y] = Hx, y. A vector x ∈ Cn is called H-positive, H-negative, or H-neutral, respectively, if [x, x] > 0, [x, x] < 0, or [x, x] = 0, respectively. A subspace M of Cn is called H-nonnegative, H-nonpositve, or H-neutral, respectively, if [x, x] ≥ 0, [x, x] ≤ 0, or [x, x] = 0, respectively, for all x ∈ M . Observe that an H-neutral subspace is at the same time H-nonnegative and H-nonpositive.

212

Chapter 11. Matrices in indeﬁnite inner product spaces

Although the Cauchy-Schwarz inequality does not hold for just any two vectors x, y in an indeﬁnite inner product space, it does hold for vectors x, y which are both in an H-nonnegative subspace, or both in an H-nonpositive subspace. Note that it follows from this that M is H-neutral if and only if H[M ] ⊂ M ⊥ . A subspace M of Cn will be called maximal H-nonnegative whenever it is H-nonnegative and not properly contained in a larger H-nonnegative subspace. Similarly, M will be called a maximal H-nonpositive subspace if it is H-nonpositive and not properly contained in a larger H-nonpositive subspace. The ﬁrst part of the following proposition can be found in Theorem 2.3.2 in [70], the second part is Lemma 6.3 in [25]. Proposition 11.1. The dimension of any maximal H-nonnegative subspace coincides with the number of positive eigenvalues of H, while the dimension of any maximal H-nonpositive subspace coincides with the number of negative eigenvalues of H. Also, if M is maximal H-nonpositive then H −1 [M ⊥ ] is maximal Hnonnegative. A subspace M of Cn is said to be H-Lagrangian if H[M ] = M ⊥ . Such a subspace is both maximal H-nonnegative and maximal H-nonpositive, and hence such a subspace can exist only if H has as many positive eigenvalues as it has negative ones. As an example, suppose n is even, n = 2k say, and let 0 Ik H=i . −Ik 0 Then any subspace of the form M = Im [ P I ]∗ with P Hermitian will be a Lagrangian subspace. The concepts involving ordinary orthogonality have straightforward analogues for H-orthogonality. For instance, vectors x and y in Cn are H-orthogonal if [x, y] = 0. A subspace M is called H-nondegenerate in case there is no non-zero vector x ∈ M that is H-orthogonal to all vectors in M . An equivalent requirement is that M ∩ H[M ]⊥ = {0}. It follows that for H-nondegenerate subspaces M , one has ˙ H[M ]⊥ . Cn = M + Conversely, each subspace M of Cn with this property is H-nondegenerate.

11.2 H-selfadjoint matrices Let the indeﬁnite inner product on Cn be given by the invertible Hermitian matrix H. An n × n matrix A has an H-adjoint A[∗] deﬁned by [Ax, y] = [x, A[∗] y].

11.2. H-selfadjoint matrices

213

Thus A[∗] = H −1 A∗ H. The matrix A is called H-selfadjoint if A = A[∗] or which amounts to the same, HA = A∗ H. As an example, let A = Jn (λ) be the n × n upper triangular Jordan block with a real eigenvalue λ, and let H = εPn , where ε is +1 or −1, and Pn is the standard n × n involutary matrix (also called the n × n reversed identity matrix). Thus Pn is the n × n matrix with 1s on the diagonal running from the lower left corner to the upper right corner, and 0s elsewhere. Clearly H is invertible and selfadjoint while, moreover, HA = A∗ H. Hence A is H-selfadjoint. As a second example, suppose n is even, n = 2k say, let λ be non-real, and let A = diag Jk (λ), Jk (λ) be the block diagonal sum of two Jordan blocks of size k with eigenvalues λ and λ, respectively. Further, let H = P2k . Then again HA = A∗ H, so A is H-selfadjoint. It turns out that these two examples can serve as the building blocks for any pair (A, H), where A is H-selfadjoint. To state this more precisely, ﬁrst observe that if A is H-selfadjoint, and if S is an invertible matrix, then S −1 AS is S ∗ HSselfadjoint. The map (A, H) → (S −1 AS, S ∗ HS) deﬁnes an equivalence relation on the set of pairs (A, H) with A being H-selfadjoint. The following result, which can be found in [70], Theorem 5.1.1, describes a canonical form for pairs of matrices of this type. Theorem 11.2. Let A be an H-selfadjoint matrix. Then there exists an invertible matrix S such that S −1 AS is equal to the block-diagonal matrix diag Jk1 (λ1 ), . . . , Jkm (λm ), Jkm+1 (λm+1 ), Jkm+1 (λm+1 ), . . . , Jkl (λl ), Jkl (λl ) , while S ∗ HS = diag ε1 Pk1 , . . . , εm Pkm , P2km+1 , . . . , P2kl . Here λ1 , . . . , λm are the real eigenvalues of A, geometric multiplicities counted, λm+1 , λm+1 , . . . , λl , λl are the non-real eigenvalues of A, geometric multiplicities counted too, and the numbers ε1 , . . . , εm take the values +1 and −1. Behind the theorem is the fact that if A is H-selfadjoint, then the spectrum of A is closed under complex conjugation, taking (partial) multiplicities into account. By slight abuse of terminology, the ordered m-tuple (ε1 , . . . , εm ) is called the sign characteristic of the pair (A, H). It is uniquely determined by the pair (A, H) up to permutations of signs corresponding to equal Jordan blocks. Next, we consider invariant maximal H-nonnegative and invariant maximal H-nonpositive subspaces. We start again with examples. Let A be a single Jordan block of size n × n with a real eigenvalue, and take H = εPn . Denote the standard

214

Chapter 11. Matrices in indeﬁnite inner product spaces

basis of Cn by e1 , . . . , en . Introduce ⎧ span {e1 , . . . , en/2 } in case n is even, ⎪ ⎪ ⎨ M + = span {e1 , . . . , e(n+1)/2 } in case n is odd and ε = +1, ⎪ ⎪ ⎩ span {e1 , . . . , e(n−1)/2 } in case n is odd and ε = −1,

M−

⎧ span {e1 , . . . , en/2 } in case n is even, ⎪ ⎪ ⎨ = span {e1 , . . . , e(n+1)/2 } in case n is odd and ε = −1, ⎪ ⎪ ⎩ span {e1 , . . . , e(n−1)/2 } in case n is odd and ε = +1.

Then M + is A-invariant and maximal H-nonnegative, while M − is A-invariant and maximal H-nonpositive. As a second example, suppose n is even, n = 2k say, let A = Jk (λ) ⊕ Jk (λ) with λ non-real, let H = P2k , and write e1 , . . . , e2k for the standard basis of C2k . Then, for l = 0, . . . , k, we have that M = span {e1 , . . . , el , ek+1 , . . . , e2k−l } is an A-invariant H-Lagrangian subspace. If A is H-selfadjoint, and λ is a real eigenvalue of A, then the spectral invariant subspace of A corresponding to λ is H-orthogonal to the spectral invariant subspace of A corresponding to all other eigenvalues. A similar statement holds for a pair of complex conjugate non-real eigenvalues λ, λ. This allows one to build up A-invariant maximal H-nonnegative subspaces by taking direct sums of subspaces constructed “locally” as in the previous two examples. In particular the following holds, see Theorem 5.12.1 in [70]. Theorem 11.3. Let A be H-selfadjoint. The following statements hold: (i) There exists an A-invariant maximal H-nonnegative subspace Mu+ such that σ(A|Mu+ ) is in the closed upper half plane. Furthermore, any such Mu+ contains the spectral invariant subspace of A corresponding to the open upper half plane. (ii) There exists an A-invariant maximal H-nonpositive subspace Mu− such that σ(A|Mu− ) is in the closed upper half plane. Furthermore, any such Mu− contains the spectral invariant subspace of A corresponding to the open upper half plane. (iii) There exists an A-invariant maximal H-nonnegative subspace Ml+ such that σ(A|M + ) is in the closed lower half plane. Furthermore, any such Ml+ conl tains the spectral invariant subspace of A corresponding to the open lower half plane. (iv) There exists an A-invariant maximal H-nonpositive subspace Ml− such that σ(A|M − ) is in the closed lower half plane. Furthermore, any such Ml− conl tains the spectral invariant subspace of A corresponding to the open lower half plane.

11.3. H-dissipative matrices

215

Our next concern is the existence of A-invariant H-Lagrangian subspaces. These do not always exist. The next theorem gives a necessary and suﬃcient condition. Theorem 11.4. Let A be H-selfadjoint. There exists an A-invariant H-Lagrangian subspace if and only if for each real eigenvalue μ of A the following two conditions hold: (i) the number of odd partial multiplicities associated with μ is even, (ii) exactly half of those odd partial multiplicities associated with μ have sign +1 corresponding to them in the sign characteristic of (A, H), the other half have sign −1 corresponding to them. In particular, if all the partial multiplicities associated with the real eigenvalues of A are even, there does exist an A-invariant H-Lagrangian subspace. To elucidate what is said in Theorem 11.4, let us return to Theorem 11.2. With the notation employed there, write s(1), . . . , s(t) for the positive integers such that λs(j) = μ, j = 1, . . . , t. Then the numbers ks(1) , . . . , ks(t) are the partial multiplicities associated with μ, and the corresponding signs in the sign characteristic of (A, H) are εs(1) , . . . , εs(t) . Item (i) of the above theorem declares that the number of j for which ks(j) is odd is even, 2p say. Suppose ks(r1 ) , . . . , ks(r2p ) are odd. Then item (ii) of the theorem says that among the signs εs(r1 ) , . . . , εs(r2p ) there are p having the value +1 and p with the value −1. We now state a result on the uniqueness of A-invariant H-Lagrangian subspaces. In one direction, this result can be found in Theorem 5.12.4 in [70], the other direction is proved in [122]. Theorem 11.5. Assume that A is H-selfadjoint. The following two statements are equivalent: (i) There exist unique A-invariant H-Lagrangian subspaces Mu and Ml such that σ(A|Mu ) is in the closed upper half plane and σ(A|Ml ) is in the closed lower half plane; (ii) The real eigenvalues of A have even partial multiplicities, and for each real eigenvalue μ of A the signs in the sign characteristic of the pair (A, H) corresponding to the partial multiplicities associated with μ are all the same. In particular, the existence of subspaces Mu and Ml with the properties mentioned in (i) is guaranteed when A has no real eigenvalues. In this case Mu and Ml are the spectral subspaces of A associated with the part of σ(A) lying in the open upper and open lower half plane, respectively.

11.3 H-dissipative matrices Next, we turn to another class of matrices. An n × n matrix is H-dissipative if 1 ∗ 2i (HA − A H) is nonnegative. It can be shown that the spectral subspace of an

216

Chapter 11. Matrices in indeﬁnite inner product spaces

H-dissipative matrix A associated with the part of σ(A) lying in the open upper half plane is H-nonnegative, while the spectral subspace corresponding to the part of σ(A) lying in the open lower half plan is H-nonpositive. Theorem 11.6. Let A be H-dissipative. Then the following statements hold: (i) There exists an A-invariant maximal H-nonnegative subspace M+ such that σ(A|M+ ) is in the closed upper half plane. Furthermore, any such M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane. (ii) There exists an A-invariant maximal H-nonpositive subspace M− such that σ(A|M− ) is in the closed lower half plane. Furthermore, any such M− contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane. The usual proof of this result is quite involved, uses a ﬁxed point argument, and holds in an inﬁnite dimensional setting as well, see [6], [87]. A constructive argument for the ﬁnite dimensional case can be found in [129], [137]. 1 (HA − A∗ H) is posThe matrix A is said to be strictly H-dissipative if 2i itive deﬁnite. In that case A cannot have real eigenvalues. Hence, for a strictly H-dissipative matrix A, the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane is maximal H-positive, and, similarly, the spectral subspace of A corresponding to the part of σ(A) contained in the open lower half plane is maximal H-negative.

Notes The material in this chapter is taken from the books [68] and [70]. For other books in this area, with an emphasis on inﬁnite dimensional spaces, see [87], [25], and [6].

Part V Riccati equations and factorization In this part the canonical factorization theorem is presented in a diﬀerent way using the notion of an angular subspace and Riccati equations. In this case one has to look for angular subspaces that are also spectral subspaces, and the solutions of the Riccati equation must have additional spectral properties. Spectral factorization as well as pseudo-spectral factorization are described in terms of Hermitian solutions of such a Riccati equation. The study of rational matrix functions that take Hermitian values on certain curves, started in the previous part, is continued with an analysis of rational matrix functions that have Hermitian values for which the inertia is independent of the point on the curve. Such functions may still admit a symmetric canonical factorization, provided one allows for a constant Hermitian invertible matrix as a middle factor. A factorization of this type is commonly known as a J-spectral factorization. This part consists of three chapters. The ﬁrst chapter (Chapter 12), which has a preliminary character, introduces the (non-symmetric) algebraic Riccati equation and presents the state space canonical factorization theorem in terms of solutions of such an equation. Pseudo-canonical factorization is treated in an analogous way. In the second chapter (Chapter 13) the symmetric algebraic Riccati equation is introduced, and spectral factorization as well as pseudo-spectral factorization are described using such Riccati equations. In the third chapter (Chapter 14) the notion of a J-spectral factorization of a rational matrix function is introduced. Necessary and suﬃcient conditions for the existence of a such factorization are given, ﬁrst in terms of invariant subspaces and then in terms of solutions of a corresponding symmetric algebraic Riccati equation. The connection between left and right J-spectral factorization is also studied.

Chapter 12

Canonical factorization and Riccati equations In this chapter the canonical factorization theorem from Section 7.1 is presented in a diﬀerent way using the notion of an angular subspace and Riccati equations. In this case one has to look for solutions of the Riccati equation that have additional spectral properties. Section 12.1, which has a preliminary character, deals with angular subspaces, and in particular those that are also spectral subspaces. Section 12.2 deals with the connection between factorization and Riccati equations in general, while Section 12.3 contains the main result. It speciﬁes further the main theorem of the second section for the case of canonical factorization. In Section 12.4, as an application, we solve in state space form the problem of obtaining a right canonical factorization when a left one is given (or reversely).

12.1 Preliminaries on spectral angular subspaces Let X be a complex Banach space, let X1 and X2 be closed subspaces of X, and suppose ˙ X2 . (12.1) X = X1 + A closed subspace N of X is said to be angular relative to the decomposition ˙ N . In that case there is a unique operator R : X2 → X1 , called (12.1) if X = X1 + the angular operator for N , such that ) ( R N = {Rx + x | x ∈ X2 } = Im , I where I, as always in this section, stands for the identity operator on the appropriate space which can be easily identiﬁed from the context (in this case X2 ).

220

Chapter 12. Canonical factorization and Riccati equations Let N be an angular subspace of X relative to (12.1), and let T11 T12 ˙ 2 → X1 +X ˙ 2 T = : X1 +X T21 T22

(12.2)

be an operator on X. We consider the question when N is invariant under T . For this purpose, set I R ˙ 2 → X1 +X ˙ 2. E= : X1 +X 0 I This operator is invertible, and maps X2 in a one-to-one way onto N . It follows that T leaves N invariant if and only if E −1 T E leaves X2 invariant. A direct computation yields ⎤ ⎡ T11 − RT21 −RT21 R − RT22 + T11 R + T12 ⎦. E −1 T E = ⎣ (12.3) T21 T22 + T21 R This formula shows that E −1 T E leaves X2 invariant if and only if the angular operator R for N satisﬁes the algebraic Riccati equation RT21 R + RT22 − T11 R − T12 = 0.

(12.4)

More precisely, this equation is usually referred to as a nonsymmetric algebraic Riccati equation. In the next chapter we shall encounter symmetric algebraic Riccati equations. The 2 × 2 operator matrix (12.2) is often referred to as the Hamiltonian corresponding to the algebraic Riccati equation (12.4). Next, let E2 be the restriction of E to X2 considered as an operator from X2 into N . Then E2 is invertible. In fact, E2−1 is the restriction of E −1 to N viewed as an operator from N into X2 . Using (12.3) we see that E2−1 (T |N )E2 = T22 + T21 R, and hence T |N and T22 + T21 R are similar. In this section we want additionally that N is a spectral subspace of T . The next proposition shows in terms of the angular operator when this happens. Proposition 12.1. Let N be an angular subspace of X relative to the decomposition (12.1), and let T be the operator on X given by (12.2). Then N is a spectral subspace for T if and only if the angular operator R for N satisﬁes the algebraic Riccati equation (12.4) and σ(T11 − RT21) ∩ σ(T22 + T21 R) = ∅. More precisely the following holds. If N = Im P (T ; Γ), where Γ is a Cauchy contour that splits σ(T ), then σ(T22 + T21 R) is inside Γ and σ(T11 − RT21) is outside Γ. Conversely, if Γ is a Cauchy contour such that σ(T22 + T21 R) is inside Γ and σ(T11 − RT21 ) is outside Γ, then the spectrum of T does not intersect with Γ and N = Im P (T ; Γ).

12.2. Angular operators and factorization

221

Proof. We use the operator E introduced before. The operator E is invertible and maps X2 in a one-to-one way onto N . Since a spectral subspace of T is invariant under T , we may assume without loss of generality that the angular operator R for N satisﬁes the Riccati equation (12.4). Then formula (12.3) shows that E

−1

TE =

T11 − RT21

0

T21

T22 + T21 R

.

(12.5)

Since E maps X2 in a one-to-one way onto N , the space N is a spectral subspace for T if and only if X2 is a spectral subspace for E −1 T E, and we can apply Lemma 3.1 to get the desired result.

12.2 Angular operators and factorization In this section we use the concepts introduced in the previous section to bring the factorization theorem (see Section 2.6) for realizations in a diﬀerent form. The ˙ X2 main point is that throughout we work with a ﬁxed decomposition X = X1 + of the state space X of the realization that has to be factorized and the factors are described with respect to this decomposition. In the ﬁnite dimensional case this corresponds to working with a ﬁxed coordinate system. Theorem 12.2. Let W (λ) = D + C(λIX − A)−1 B be a biproper realization with state space X and input-output space Y . Let X1 and X2 be closed subspaces of ˙ X2 , let N be a closed subspace of X X such that (12.1) holds, i.e., X = X1 + ˙ N , and denote the which is angular relative to this decomposition, so X = X1 + corresponding angular operator by R. Assume A[X1 ] ⊂ X1 ,

A× [N ] ⊂ N,

(12.6)

and let D = D1 D2 with D1 and D2 invertible operators on Y . Write A11 A12 ˙ X2 → X1 + ˙ X2 , A = : X1 + 0 A22 B

=

C

=

B1 B2 C1

˙ X2 , : Y → X1 + C2

˙ X2 → Y. : X1 +

Then R satisﬁes the algebraic Riccati equation RB2 D −1 C1 R − R(A22 − B2 D−1 C2 ) + (A11 − B1 D−1 C1 )R + (A12 − B1 D−1 C2 ) = 0.

(12.7)

222

Chapter 12. Canonical factorization and Riccati equations

Introduce the functions W1 and W2 via the biproper realizations W1 (λ)

= D1 + C1 (λIX1 − A11 )−1 B1 D2−1 ,

W2 (λ)

= D2 + D1−1 C2 (λIX2 − A22 )−1 B2 .

Then W admits the factorization λ ∈ ρ(A11 ) ∩ ρ(A22 ) ⊂ ρ(A).

W (λ) = W1 (λ)W2 (λ), Also put −1 C1 , A× 11 = A11 − (B1 − RB2 )D

−1 A× (C1 R + C2 ). (12.8) 22 = A22 − B2 D

× Then, for λ ∈ ρ(A× 11 ) ∩ ρ(A22 ) ∩ ρ(A11 ) ∩ ρ(A22 ), the operators W (λ), W1 (λ) and W2 (λ) are invertible, and

W (λ)−1 = W2 (λ)−1 W1 (λ)−1 , where W1−1 (λ)

=

−1 D1−1 − D1−1 C1 (λIX1 − A× (B1 − RB2 )D−1 , 11 )

W2−1 (λ)

=

−1 D2−1 − D−1 (C1 R + C2 )(λIX2 − A× B2 D2−1 . 22 )

Proof. The ﬁrst part of the theorem is a direct consequence of the observations presented before Proposition 12.1, applied to A× . Indeed, let E be the invertible operator I R E= , 0 I = E −1 B, C = CE. Then = E −1 AE, B and write A A11 A12 − RA22 + A11 R , A = 0 A22 B

=

B1 − RB2

,

B2 C

=

C1

C1 R + C2

and it follows that ×

A =E

−1

×

A E=

A× 11

H

−B2 D−1 C1

A× 22

,

12.2. Angular operators and factorization

223

× where A× 11 and A22 are deﬁned by (12.8), and where H is equal to the left-hand side of (12.7). Now E maps X1 onto X1 and X2 onto N . Thus (12.6) implies that

1 ] ⊂ X1 , A[X

× [X2 ] ⊂ X2 . A

Hence (12.7) is satisﬁed. It remains to prove the factorization W = W1 W2 and to establish the formu On the −1 B. las for W1 , W2 and their inverses. We have W (λ) = D + C(λI − A) other hand, by the product rule for realizations, −1 B, W1 (λ)W2 (λ) = D + C(λI − A) where

= A

A11

(B1 − RB2 )D−1 (C1 R + C2 )

0

A22

.

It remains to observe that by (12.7) (B1 − RB2 )D−1 (C1 R + C2 ) = A12 − RA22 + A11 R. So W = W1 W2 . The formulas for the inverses are immediate.

The next theorem is a symmetric version of Theorem 12.2. Theorem 12.3. Let W (λ) = D + C(λIX − A)−1 B be a biproper realization with state space X and input-output space Y . Let X1 and X2 be closed subspaces of X ˙ X2 . Further, let N1 and N2 be closed subspaces of X for which with X = X1 + ˙ N2 , X = X1 +

˙ X2 , X = N1 +

˙ X2 while N1 is that is, N2 is angular relative to the decomposition X = X1 + ˙ X1 . Let R12 : X2 → X1 and R21 : X1 → X2 be the angular relative to X = X2 + corresponding angular operators. Assume A[N1 ] ⊂ N1 ,

˙ N2 , X = N1 +

A× [N2 ] ⊂ N2 ,

and let D = D1 D2 with D1 and D2 invertible operators on Y . Write A11 A12 ˙ X2 → X1 + ˙ X2 , : X1 + A = A21 A22 B

=

C

=

B1 B2 C1

˙ X2 , : Y → X1 + C2

˙ X2 → Y, : X1 +

(12.9)

224

Chapter 12. Canonical factorization and Riccati equations

and put R1 = IX1 − R12 R21 and R2 = IX2 − R21 R12 . Then R1 : X1 → X1 and R2 : X2 → X2 are invertible. Introduce the functions W1 and W2 via the biproper realizations −1 −1 R1 (B1 − R12 B2 )D2−1 , W1 (λ) = D1 + (C1 + C2 R21 ) λIX1 − (A11 + A12 R21 ) −1 W2 (λ) = D2 + D1−1 (C1 R12 + C2 )R2−1 λIX2 − (A22 − R21 A12 ) (B2 − R21 B1 ). Then W admits the factorization W (λ) = W1 (λ)W2 (λ),

λ ∈ ρ(A11 + A12 R21 ) ∩ ρ(A22 − R21 A12 ) ⊂ ρ(A).

Also put −1 A× C1 − R12 A21 + R12 B2 D−1 C1 , 11 = A11 − B1 D −1 C2 + A21 R12 − B2 D−1 C1 R12 . A× 22 = A22 − B2 D × Then, for λ ∈ ρ(A11 +A12 R21 ) ∩ ρ(A22 −R21 A12 ) ∩ ρ(A× 11 ) ∩ ρ(A22 ), the operators W (λ), W1 (λ) and W2 (λ) are invertible, and

W (λ)−1 = W2 (λ)−1 W1 (λ)−1 , where W1−1 (λ)

=

−1 D1−1 − D1−1 (C1 + C2 R21 )R1−1 (λIX1 − A× (B1 − R12 B2 )D−1 , 11 )

W2−1 (λ)

=

−1 −1 D2−1 − D−1 (C1 R12 + C2 )(λIX2 − A× R2 (B2 − R21 B1 )D2−1 . 22 )

We prepare for the proof of the theorem with a lemma. Lemma 12.4. Let X be a Banach space, and let X1 and X2 be closed subspaces of ˙ X2 . Further, let N1 and N2 be closed subspaces of X for which X with X = X1 + ˙ N2 , X = X1 +

˙ X2 , X = N1 +

˙ X2 while N1 is ani.e., N2 is angular relative to the decomposition X = X1 + ˙ gular relative to X = X2 + X1 . Let R12 : X2 → X1 and R21 : X1 → X2 be the corresponding angular operators. Then the following statements are equivalent: ˙ N2 ; (i) X = N1 + (ii) I − R21 R12 is invertible; (iii) I − R12 R21 is invertible; I R12 ˙ X2 → X1 + ˙ X2 is invertible. (iv) F = : X1 + R21 I

12.2. Angular operators and factorization

225

In case the equivalent conditions (i)–(iv) hold, the projection PN of X along N1 onto N2 is given by R12 (I − R21 R12 )−1 −R21 I , PN = I while the complementary projection I − PN can be written as I I − PN = (I − R12 R21 )−1 I −R12 . R21 Proof. The equivalence of (ii), (iii) and (iv) is straightforward. Observe that F maps X1 and X2 in a one-to-one manner onto N1 and N2 , respectively. Since ˙ X2 , it is clear that X = N1 + ˙ N2 if and only if F is invertible. So (i) X = X1 + and (iv) are equivalent. To complete the proof it remains to prove the formula for PN . Observe that the expression in the right-hand side of the claimed identity for PN does deﬁne a projection. Its image and kernel are given by I R12 , , Im Im I R21 respectively, so it is indeed equal to the projection PN .

Proof of Theorem 12.3. From Lemma 12.4 we know that the operator I R12 ˙ X2 → X1 + ˙ X2 : X1 + F = R21 I = F −1 AF, B = F −1 B and C = CF . Then W (λ) = is invertible. Introduce A −1 ˆ × ˆ ˆ D + C(λI − A) B. Note that A[X1 ] ⊂ X1 and A [X2 ] ⊂ X2 , where, following − BD −1 C, × = A and so A × = F −1 A× F . Write standard convention A . 1 12 11 A B A = = C = 1 C 2 , , B , C A 22 2 0 A B and put W1 (λ)

=

1 (λ − A 11 )−1 B 1 D−1 , D1 + C 2

W2 (λ)

=

22 )−1 B 2 (λ − A 2 . D2 + D1−1 C

11 ) ∩ ρ(A 22 ) ⊂ ρ(A) = ρ(A), the function W is the product of W1 Then on ρ(A and W2 .

226

Chapter 12. Canonical factorization and Riccati equations The inverse of F is given by −R1−1 R12 R1−1 −1 ˙ X 2 → X1 + ˙ X2 . : X1 + F = −R21 R1−1 I + R21 R1−1 R12

Using this and the expression for F , one easily sees that 11 A 1 D−1 B 2 1 C

= R1−1 (A11 + A12 R21 − R12 A21 − R12 A22 R21 ), = R1−1 (B1 − R12 B2 )D2−1 , = C1 + C2 R21 .

Now R21 satisﬁes the algebraic Riccati equation R21 A12 R21 + R21 A11 − A22 R21 − A21 = 0, 11 = A11 + A12 R21 . Thus, for the function W1 , we have and it follows that A 1 D−1 1 (λ − A 11 )−1 B W1 (λ) = D1 + C 2 −1 −1 = D1 + (C1 + C2 R21 ) λ − (A11 + A12 R21 ) R1 (B1 − R12 B2 )D2−1 , as desired. Next we compute the function W2 . Using the alternative formula I + R12 R2−1 R21 −R12 R2−1 −1 ˙ X2 → X1 + ˙ X2 F = : X1 + −R2−1 R21 R2−1 for the inverse of F , we obtain 22 A

= R2−1 (A22 − R21 A12 )R2−1 ,

2 B

= R2−1 (B2 − R21 B1 ),

1 D1−1 C

= D1−1 (C1 R12 + C2 ).

Hence, for the function W2 we get 22 )−1 B 2 (λ − A 2 W2 (λ) = D2 + D1−1 C −1 = D2 + D1−1 (C1 R12 + C2 )R2−1 λ − (A22 − R21 A12 ) (B1 − R12 B2 )D2−1 , again as desired. This proves that the factorization claimed in the theorem holds on ρ(A11 + A12 R21 ) ∩ ρ(A22 − R21 A12 )

12.3. Riccati equations and canonical factorization

227

which is a subset of ρ(A). What remains to be done is to deduce the formulas for the inverses. But this amounts to repeating the work with W replaced by W −1 . In doing so, one employs the Riccati equation R12 (A21 − B2 D−1 C1 )R12 + R12 (A22 − B2 D−1 C2 ) −(A11 − B1 D −1 C1 )R12 − (A12 − B1 D−1 C2 ) = 0

for R12 instead of the one for R21 used above. The details are omitted.

12.3 Riccati equations and canonical factorization In this section Theorem 12.2 is speciﬁed further for the case of canonical factorization. As usual, Γ is a Cauchy contour in the complex plane, F+ is its interior domain, and F− its exterior domain (inﬁnity included). Theorem 12.5. Let W (λ) = D + C(λIX − A)−1 B be a biproper realization with state space X and input-output space Y . Assume that the spectrum of A does not intersect Γ. Put X1 = Im P (A; Γ) and let X2 be a closed subspace of X such that ˙ X2 , so X = X1 + ˙ X2 . X = Im P (A; Γ) + Let D = D1 D2 with D1 and D2 invertible operators on Y , and write A11 A12 ˙ X2 → X1 + ˙ X2 , : X1 + A = 0 A22 B

=

C

=

B1 B2 C1

˙ X2 , : Y → X1 + C2

˙ X2 → Y. : X1 +

Then W admits a right canonical factorization with respect to Γ if and only if the Riccati equation RB2 D−1 C1 R − R(A22 − B2 D−1 C2 ) + (A11 − B1 D−1 C1 )R + (A12 − B1 D

−1

(12.10)

C2 ) = 0

has a (unique) solution R satisfying the constraints ⊂ σ A11 − (B1 − RB2 )D −1 C1 σ A22 − B2 D −1 (C1 R + C2 ) ⊂

F+ ,

(12.11)

F− .

(12.12)

228

Chapter 12. Canonical factorization and Riccati equations

In that case a right canonical factorization W (λ) = W− (λ)W+ (λ) of W with respect to Γ is obtained by taking W− (λ)

= D1 + C1 (λ − A11 )−1 (B1 − RB2 )D2−1 ,

W+ (λ)

= D2 + D1−1 (C1 R + C2 )(λ − A22 )−1 B2 .

Moreover, the inverses of W− and W+ are given by W−−1 (λ)

=

−1 D1−1 − D1−1 C1 (λ − A× (B1 − RB2 )D −1 , 11 )

W+−1 (λ)

=

−1 D2−1 − D−1 (C1 R + C2 )(λ − A× B2 D2−1 , 22 )

where −1 C1 , A× 11 = A11 − (B1 − RB2 )D

−1 A× (C1 R + C2 ). 22 = A22 − B2 D

With the appropriate modiﬁcations, the theorem also holds for certain contours in the Riemann sphere. For instance, if for Γ one takes the (extended) imaginary axis, one has to take for F+ the open left half plane and for F− the open right half plane. For left canonical factorizations analogous results hold: just interchange the roles of inner and outer domains (see the comment after Theorem 3.2). Proof. The subspace X1 = Im P (A; Γ) is invariant under A, and hence the zero entry in the left lower corner of the operator matrix representation of A is justiﬁed. Furthermore σ(A11 ) ⊂ F+ and σ(A22 ) ⊂ F− . ˙ X2 we have Next note that relative to the decomposition X = X1 + ⎡ ⎤ A11 − B1 D−1 C1 A12 − B1 D−1 C2 ⎦. A× = A − BD−1 C = ⎣ −B2 D−1 C1 A22 − B2 D−1 C2 Thus −A× is precisely the Hamiltonian of the Riccati equation (12.10). Assume that W admits a right canonical factorization with respect to Γ. Then, in particular, W (λ) is invertible for each λ ∈ Γ; hence, by Theorem 2.4, the spectrum of the operator A× does not intersect Γ. Thus we can use Theorem 7.1 to show that N = Ker P (A× ; Γ) is an angular subspace for the decomposition ˙ X2 . Let R be the corresponding angular operator. Since A× leaves N X = X1 + invariant, we know that R satisﬁes the Riccati equation −RB2 D−1 C1 R + R(A22 − B2 D−1 C2 ) − (A11 − B1 D−1 C1 )R

(12.13)

−(A12 − B1 D−1 C2 ) = 0, which is equivalent to (12.10). Now Proposition 12.1, applied to A× and with the roles of the interior and exterior domain of the contour Γ being reversed, shows that (12.11) and (12.12) are fulﬁlled.

12.4. Left versus right canonical factorization

229

Conversely, let R be a solution of the Riccati equation (12.10) for which (12.11) and (12.12) are satisﬁed. Thus R satisﬁes the Riccati equation (12.13) which has A× as its Hamiltonian. Hence the corresponding angular subspace N is invariant under A× . Next we again use Proposition 12.1 with T = A× and with the roles of the interior and exterior domain of the contour Γ being reversed. This yields that the spectrum of A× does not intersect Γ and that N = Ker P (A× ; Γ). Since ˙ X2 , the latter implies that N is an angular subspace of X relative to X = X1 + ˙ Ker P (A× ; Γ). But then Theorem 3.2 implies that W admits a X = Im P (A; Γ) + right canonical factorization with respect to the contour Γ. To show uniqueness of the solution R of (12.10) for which the spectral inclusions (12.11) and (12.12) are satisﬁed, it suﬃces to note that these spectral inclusions imply that N = Ker P (A× ; Γ). Indeed, in that case the angular opera˙ 2 is uniquely determined. tor R for N relative to X = X1 +X It remains to get the formulas for the factors. First note that Theorem 12.2 shows that W (λ) = W− (λ)W+ (λ) with the factors W− (λ), W+ (λ) and their inverses being of the desired form. The spectral properties of A11 and A22 , together × with those of A× 11 and A22 , show that the factorization W (λ) = W− (λ)W+ (λ) is a right canonical factorization with respect to Γ.

12.4 Left versus right canonical factorization In this section we answer the following question: if a rational matrix function W admits a left canonical factorization, under what conditions does it also have a right canonical factorization? And, if so, how can the right factorization be obtained from the left one? Our starting point is a given biproper operator function W , a Cauchy contour Γ, and a left canonical factorization W (λ) = Y+ (λ)Y− (λ),

λ ∈ Γ.

(12.14)

The biproper factors Y+ and Y− are given in terms of realizations, that is, Y+ (λ)

= D+ + C+ (λIX+ − A+ )−1 B+ ,

(12.15)

Y− (λ)

= D− + C− (λIX− − A− )−1 B− .

(12.16)

We are looking for a right canonical factorization W (λ) = W− (λ)W+ (λ). The key idea for solving this problem is the following: combine the realizations of Y+ and Y− into a realization for W using the product rule for realizations, then apply the canonical factorization theorem (Theorem 7.1) to see if a right canonical factorization exists and, if so, produce formulas for the factors. As before the interior of Γ will be denoted by F+ , the exterior by F− . We (may and) shall assume that the operators in the realizations are chosen in such a way that the operators D+ and D− are invertible, the spectra of the operators

230

Chapter 12. Canonical factorization and Riccati equations

−1 × A+ and A× + = A+ − B+ D+ C+ are contained in F− , and those of A− and A− = −1 A− − B− D− C− in F+ . Then, in particular, the spectra of A− and A+ are disjoint and the Lyapunov equation

A+ Z − ZA− = −B+ C−

(12.17)

has a unique solution Z : X− → X+ (see Section I.4 in [51]). Similarly, the Lyapunov equation × −1 −1 A× (12.18) − Z − ZA+ = B− D− D+ C+ has a unique solution Z : X+ → X− . These facts are used in the following theorem and its proof. Theorem 12.6. Let W (λ) = Y+ (λ)Y− (λ) be a left canonical factorization of W with respect to the Cauchy contour Γ, and let the factors be given by (12.15) and (12.16). Let Q : X− → X+ and P : X+ → X− be the unique solutions of the Lyapunov equations (12.17) and (12.18), respectively, that is, A+ Q − QA− = −B+ C− ,

× −1 −1 A× − P − P A+ = B− D− D+ C+ .

(12.19)

Then W has a right canonical factorization W (λ) = W− (λ)W+ (λ) with respect to Γ if and only if IX+ − QP is invertible, or, which amounts to the same, IX− − P Q is invertible. In that case, on the appropriate domains, the factors W− and W+ , and their inverses W−−1 and W+−1 , are given by W− (λ)

=

D+ + (D+ C− + C+ Q)(λIX− − A− )−1 −1 · (IX− − P Q)−1 (B− D− − P B+ ),

W+ (λ)

=

−1 D− + (D+ C+ + C− P )(IX+ − QP )−1

· (λIX+ − A+ )−1 (B+ D− − QB− ), W−−1 (λ)

=

−1 −1 D+ − D+ (D+ C− + C+ Q)(IX− − P Q)−1 −1 −1 −1 · (λIX− − A× (B− D− − P B+ )D+ , −)

W+−1 (λ)

=

−1 −1 −1 −1 D− − D− (D+ C+ + C− P )(λIX+ − A× +) −1 · (IX+ − QP )−1 (B+ D− − QB− )D− .

Proof. First we use (12.15) and (12.16) to obtain a realization for W given in the ˙ X+ and deﬁne A : X → X by form (12.14). So we write X = X− + A− 0 ˙ X+ → X − + ˙ X+ . : X− + A= B+ C− A+ Then, by the product rule (see Section 2.5), W (λ) = D+ D− +

D+ C−

C+

−1 λIX − A

B− B+ D−

.

12.4. Left versus right canonical factorization

231

The associate main operator of this realization is × −1 −1 D+ C+ A− −B− D− × ˙ X+ → X− + ˙ X+ . : X− + A = 0 A× + The spectra of A and A× do not intersect Γ. Put M = Im P (A; Γ),

M × = Ker P (A× ; Γ).

In order that W admits a right canonical factorization with respect to Γ it is ˙ M ×. necessary and suﬃcient (see Theorem 7.1) that X = M + From the matrix representation of A given above we see that Ker P (A; Γ) ˙ X+ , and hence for some Z : X− → X+ we have coincides with X+ . So X = M + I M = Im . Z The fact that M is invariant under A now amounts to (12.17). But then the operator Z must be equal to Q. In a similar way one shows that P × M = Im , I where P : X+ → X− is the unique solution of (12.18). From Lemma 12.4 we know ˙ M is equivalent to the invertibility of the matrix that the condition X = M × + I P , Q I which, in turn, is equivalent to the invertibility of I − QP or, which amounts to the same, the invertibility of I − P Q. This proves the ﬁrst part of the theorem. The formulas for the factors follow by applying Theorem 12.3 with X− , X+ , M , M × , Q and P in the role of X1 , X2 , N1 , N2 , R21 and R12 , respectively. With the obvious modiﬁcations, Theorem 12.6 holds true for canonical factorizations with respect to the usual contours in the Riemann sphere (real line and imaginary axis).

Notes This chapter is a rewritten and enriched version of Chapter 5 in [11]. Theorem 12.5 in Section 12.3 seems to be new. The material in the ﬁnal section can be found in [8]. The notion of an angular operator is standard in operator theory and goes back to [101]. The theory of Riccati equations is important in system theory; see, e.g., the text books [94], [33]. For more details on this subject we also refer to the monograph [106] and to Section 1.6 in [69].

Chapter 13

The symmetric algebraic Riccati equation As we know from the previous part there is an intimate connection between canonical factorization and Riccati equations. In this chapter this connection is developed further for the case when the rational matrix functions involved have Hermitian values on the imaginary axis. In this case the corresponding Riccati equation has additional symmetry properties too. The chapter consists of three sections. In Section 13.1 we discuss two special cases, which both lead to symmetric algebraic Riccati equations of a special type. In a somewhat more general form, this symmetric version of the algebraic Riccati equation is studied in Section 13.2, with special attention for stabilizing solutions. The study is completed in Section 13.3 where we consider Hermitian solutions of the symmetric algebraic Riccati equation and related pseudo-spectral factorizations.

13.1 Spectral factorization and Riccati equations In this section we present two illustrative special cases of spectral factorization. In both cases the corresponding Riccati equations are symmetric. For our ﬁrst case, the starting point is a rational m × m matrix function G given in realized form G(λ) = Im + C(λIn − A)−1 B, with σ(A) in the open ¯ ∗ G(λ). Clearly W is left half plane, and we consider the product W (λ) = G(−λ) a nonnegative rational m × m matrix function on the imaginary axis. We shall assume additionally that G(λ) is invertible for each λ ∈ iR, which in the present situation is equivalent to the requirement that A× = A − BC has no eigenvalue on iR. The fact that G(λ) is invertible for each λ ∈ iR means that W is positive deﬁnite on R and, as we shall see, Theorem 9.5 can be applied to show that the function W admits a left spectral factorization with respect to iR. We shall use

234

Chapter 13. The symmetric algebraic Riccati equation

Theorem 12.5 to obtain such a factorization explicitly in terms of the matrices A, B and C appearing in the realization of G. Theorem 13.1. Let G(λ) = Im + C(λIn − A)−1 B be a realization of a rational m × m matrix function G such that A has all its eigenvalues in the open left half plane. Put A× = A − BC, and assume that A× has no eigenvalue on iR. Then the Riccati equation −P BB ∗ P + P A× + (A× )∗ P = 0

(13.1)

has a unique Hermitian solution P such that A× − BB ∗ P has all its eigenvalues in ¯ ∗ G(λ) the left half plane. Furthermore, the rational matrix function W (λ) = G(−λ) admits a left spectral factorization of W with respect to the imaginary axis. In fact, ¯ ∗ L− (λ) with W (λ) = L− (−λ) L− (λ) = Im + (C + B ∗ P )(λIn − A)−1 B, is such a factorization. By Theorem 2.4, the inverse L−1 − of the spectral factor L− in the above theorem is given by ∗ × ∗ −1 B. L−1 − (λ) = Im − (C + B P )(λIn − A + BB P )

In comparable situations later on in the book, where obtaining descriptions of inverses of factors would involve only a routine application of Theorem 2.4, we will refrain from giving the expressions. Proof. We split the proof into two parts. In the ﬁrst part we show that equation (13.1) has a unique Hermitian solution P such that A× − BB ∗ P has all its eigenvalues in the left half plane. ¯ ∗ = Im − B ∗ (λIn + A∗ )−1 C ∗ . Part 1. From the given realization of G we get G(−λ) Now apply the product rule from Section 2.5). This gives W (λ) = I +

−B

∗

C

$ λ−

−A∗

C ∗C

0

A

%−1

C∗ B

.

(13.2)

It is easy to check that the hypotheses of Theorem 9.5 are satisﬁed with the skewHermitian matrix H given by 0 −In H= . (13.3) In 0 Hence W admits both a left and a right spectral factorization with respect to iR. In particular W admits both a left and a right canonical factorization with respect to the imaginary axis.

13.1. Spectral factorization and Riccati equations

235

Put F− = Cleft and F+ = Cright , where Cleft and Cright are the open left and right half planes, respectively. By hypothesis σ(A) ⊂ Cleft . So σ(−A∗ ) ⊂ Cright . Thus the realization of W in (13.2) is of the form required in Theorem 12.5, and the Riccati equation (12.10) in the theorem reduces here to −RBB ∗ R − RA× − (A× )∗ R = 0,

(13.4)

where, as usual, A× = A − BC. Since W admits a left canonical factorization with respect to the imaginary axis, (the appropriate version of) Theorem 12.5 (see the remark made between the theorem and its proof) shows that (13.4) has a unique solution R satisfying σ (A× )∗ + RBB ∗ ⊂ Cleft . (13.5) σ A× + BB ∗ R ⊂ Cleft , Here we used that σ − (A× )∗ − RBB ∗ ⊂ Cright is equivalent to the second inclusion in (13.5). Taking adjoints in (13.4) and (13.5) we see that (13.4) and (13.5) remain true if R is replaced by R∗ . But then the uniqueness of the solution implies R = R∗ . Note that for R = R∗ the two inclusions in (13.5) are equivalent. Thus we see that (13.4) has a unique Hermitian solution R satisfying the ﬁrst inclusion in (13.5). When R is replaced −P , equation (13.4) transforms into equation (13.1). Thus (13.1) has a unique Hermitian solution P satisfying σ(A× − BB ∗ P ) ⊂ Cleft . Part 2. Theorem 12.5 also yields a canonical factorization of the rational matrix function given by (13.2). In fact, such a factorization is given by W (λ) = W− (λ)W+ (λ) where the factors and their inverses are given by W− (λ)

=

I − B ∗ (λ + A∗ )−1 (C ∗ + P B),

W+ (λ)

=

W−−1 (λ)

=

I + (B ∗ P + C)(λ − A)−1 B, −1 ∗ I + B ∗ λ + (A× )∗ − P BB ∗ (C + P B),

W+−1 (λ)

=

−1 I − (B ∗ P + C) λ − A× + BB ∗ P B.

¯ ∗ , and hence Comparing the ﬁrst two expressions we see that W− (λ) = W+ (−λ) the factorization W (λ) = W− (λ)W+ (λ) is a left spectral factorization with respect to iR. Now put L− = W+ to arrive at the desired result. For our second special case, we assume that W is proper, Hermitian on the imaginary axis, and has no poles there. This implies that W can be written in the form W (λ) = D + C(λIn − A)−1 B − B ∗ (λIn + A∗ )−1 C ∗ , (13.6) where D is Hermitian and A has all its eigenvalues in the open left half plane. On the basis of this representation we shall prove the following theorem. Theorem 13.2. Let the rational m × m function W be given by (13.6), where D is positive deﬁnite and A has all its eigenvalues in the open left half plane. Assume

236

Chapter 13. The symmetric algebraic Riccati equation

additionally that W has no zeros on the imaginary axis, and put A× = A−BD−1 C. Then the Riccati equations P BD−1 B ∗ P − P A× − (A× )∗ P + C ∗ D −1 C = 0,

(13.7)

QC ∗ D−1 CQ − Q(A× )∗ − A× Q + BD−1 B ∗ = 0

(13.8)

have unique Hermitian solutions P and Q that satisfy σ A× − BD−1 B ∗ P ⊂ Cleft , σ (A× )∗ − C ∗ D −1 CQ ⊂ Cleft .

(13.9)

Furthermore, with respect to the imaginary axis, W admits left and right spectral factorizations, ¯ ∗ L− (λ), W (λ) = L− (−λ)

¯ ∗ L+ (λ), W (λ) = L+ (−λ)

(13.10)

respectively, with the factors L− and L+ being given by L− (λ) = D 1/2 + D−1/2 (C + B ∗ P )(λIn − A)−1 B,

(13.11)

L+ (λ) = D1/2 − D−1/2 (CQ + B ∗ )(λIn + A∗ )−1 C ∗ .

(13.12)

Proof. We split the proof into four parts. In the ﬁrst three parts the attention is focussed on equation (13.7) and the ﬁrst parts of (13.9) and (13.10). Part 1. From (13.6) we get %−1 $ −A∗ 0 −C ∗ ∗ C . W (λ) = D + B λ− B 0 A The main matrix of this realization has no pure imaginary eigenvalues. This follows from the assumption on the eigenvalues of A. Clearly W is selfadjoint on the imaginary axis and takes invertible values there. It follows that for λ ∈ iR the signature of the matrix W (λ), that is, the diﬀerence between the number of positive and negative eigenvalues of W (λ), does not depend on λ. As W (∞) = D is positive deﬁnite, we obtain that W (λ) is positive deﬁnite for λ ∈ iR. So the hypotheses of Theorem 9.5 are satisﬁed with the skew-Hermitian matrix H given by (13.3). Hence W admits both a left and a right spectral factorization with respect to iR. To get the formulas for the factors we will apply (the appropriate version of) Theorem 12.5 (see the remark made between the theorem and its proof) Part 2. For the case considered here the Riccati equation (12.10) in Theorem 12.5 has the form RBD −1 B ∗ R − RA× − (A× )∗ R + C ∗ D −1 C = 0. This is precisely equation (13.7) with R in place of P . Since W admits a left canonical factorization with respect to the imaginary axis, Theorem 12.5 tells us that equation (13.7) has a unique solution P satisfying σ A× − BD−1 B ∗ P ⊂ Cleft . (13.13) σ − (A× )∗ + P BD −1 B ∗ ⊂ Cright ,

13.1. Spectral factorization and Riccati equations

237

Using the symmetry properties in (13.7) and (13.13), we see that P ∗ is also a solution of (13.7) satisfying (13.13). Because of the uniqueness of P , we have P = P ∗ , and hence P is a Hermitian solution of (13.7) satisfying the ﬁrst inclusion in (13.9). On the other hand, if P is a Hermitian solution of (13.7) satisfying the ﬁrst inclusion in (13.9), then P actually satisﬁes both inclusions in (13.13), and hence P = P. Part 3. Next, we derive the ﬁrst factorization in (13.10). By Theorem 12.5 the matrix function W admits a right canonical factorization, W (λ) = W− (λ)W+ (λ), with respect to iR. The factors in this factorization are given by W− (λ)

=

D1/2 + B ∗ (λ + A∗ )−1 (−C ∗ − P B)D−1/2 ,

W+ (λ)

=

D1/2 + D−1/2 (B ∗ P + C)(λ − A)−1 B.

¯ ∗ = W− (λ), and hence the ﬁrst identity in Put L− (λ) = W+ (λ). Then L− (−λ) (13.10) holds. Moreover, the function L− (λ) is given by (13.11). Since the factorization W (λ) = W− (λ)W+ (λ) is a canonical one, we also know that the factoriza¯ ∗ L− (λ) is a left spectral factorization of W with respect to tion W (λ) = L− (−λ) iR. Part 4. Finally, to get the corresponding result for the Riccati equation (13.8) and the second factorization in (13.10), we apply the results obtained in the preceding paragraphs to V (λ) = W (−λ), that is, to V (λ) = D + B ∗ (λ − A∗ )−1 C ∗ − C(λ + A)−1 B. Note that A∗ has all its eigenvalues in Cleft . Furthermore, if the function V admits a left spectral factorization with respect to the imaginary axis, V (λ) = ¯ ∗ K− (λ) say, then W (λ) = K− (λ) ¯ ∗ K− (−λ) is a right spectral factorization K− (−λ) of W with respect to iR. We conclude this section with a few remarks about the Hermitian solutions of the Riccati equations appearing in Theorem 13.2. Let W be given by (13.6) with D positive deﬁnite. First we show that any Hermitian solution P of (13.7) is invertible whenever the pair (C, A) is observable. Suppose P x = 0. Since P is Hermitian, we also have x∗ P = 0. Then (13.7) yields x∗ C ∗ D−1 Cx = 0. As D is positive deﬁnite, this gives Cx = 0. But then, again using (13.7), we get P A× = 0, and hence P Ax = P A× x+ P BD−1 Cx = 0. So Ker P is A-invariant and is contained in Ker C. Hence Ker P is contained in Ker (C|A), and thus Ker P = {0} when Ker (C|A) = {0}. In a similar way one shows that controllability of the pair (A, B) implies that every Hermitian solution Q of (13.8) is invertible. Thus, if the realization C(λ − A)−1 B is minimal, then the Hermitian solutions of the Riccati equations (13.7) and (13.8) are automatically invertible. Now let P be an invertible Hermitian solution of (13.7). Multiplying (13.7) from both sides by P −1 shows that Q = P −1 is an invertible Hermitian solution of

238

Chapter 13. The symmetric algebraic Riccati equation

(13.8). The converse is also true, that is, if Q is an invertible Hermitian solution of (13.8), then P = Q−1 is an invertible Hermitian solution of (13.7). Thus the map P → Q = P −1 provides a one-to-one correspondence between the invertible Hermitian solutions P of (13.7) and the invertible Hermitian solutions Q of (13.8). Furthermore, in this case (with Q = P −1 ) we have σ A× − BD−1 B ∗ P = σ − (A× )∗ + C ∗ D−1 CQ . Indeed, by (13.7) we have P A× − P BD−1 B ∗ P = −(A× )∗ P + C ∗ D−1 C, and so A× − BD−1 B ∗ P

= P −1 (P A× − P BD−1 B ∗ P ) = P −1 − (A× )∗ P + C ∗ D−1 C = P −1 − (A× )∗ + C ∗ D−1 CP −1 P = P −1 − (A× )∗ + C ∗ D−1 CQ P.

In particular, if the eigenvalues of A× − BD−1 B ∗ are in the open left half plane, then those of (A× )∗ − C ∗ D−1 CQ are in the open right half plane. Comparing this with (13.9), we see that in Theorem 13.2 the matrix Q is not the inverse of the matrix P .

13.2 Stabilizing solutions The equations (13.1) and (13.7) are special cases of the general symmetric algebraic Riccati equation −P BR−1 B ∗ P + P A + A∗ P + Q = 0, (13.14) with R and Q selfadjoint, R invertible. Note that the Hamiltonian (see Section 12.1) corresponding to equation (13.14) is the 2 × 2 block matrix −Q −A∗ T = . (13.15) −BR−1B ∗ A We shall assume throughout this section that A is an n × n matrix, B an n × m matrix, Q a selfadjoint n × n matrix, and R a positive deﬁnite m × m matrix. Thus the Hamiltonian T can be viewed as an operator on C2n = Cn ⊕ Cn . We shall also assume that the pair (A, B) is stabilizable. The latter means that there there exists an m × n matrix F such that A − BF has all its eigenvalues in the open left half plane. Equation (13.14) plays an important role in optimal control theory, where one is mainly interested in stabilizing solutions P . A solution P of (13.14) is said to be iR-stabilizing, or simply stabilizing when no confusion is possible, if the matrix A − BR−1 B ∗ P has all its eigenvalues in the open left half plane. In order that such a solution exists the pair (A, B) has to be stabilizable. In general, however,

13.2. Stabilizing solutions

239

this condition is not suﬃcient. An additional condition on the eigenvalues of the Hamiltonian T is required. Theorem 13.3. Consider the symmetric algebraic Riccati equation (13.14) with R positive deﬁnite and Q selfadjoint. Then the following two statements are equivalent: (i) There exists an iR-stabilizing solution of (13.14); (ii) The pair (A, B) is stabilizable and the Hamiltonian T given by (13.15) does not have pure imaginary eigenvalues. Moreover, if (13.14) has an iR-stabilizing solution, then it is unique and Hermitian. The proof of the implication (i) ⇒ (ii) and of the ﬁnal statement of the theorem concerning the uniqueness of the iR-stabilizing solution do not require R to be positive deﬁnite; selfadjointness and invertibility of R are enough. It will be convenient ﬁrst to prove a lemma using a somewhat more general setting. For this purpose we return to the general algebraic Riccati equation which was studied in Chapter 12: XT21 X + XT22 − T11X − T12 = 0.

(13.16)

Taking T21 = −BR−1 B ∗ ,

T22 = A,

T11 = −A∗ ,

T12 = −Q,

(13.17)

and setting X = P , we see that we arrive at (13.14). Note that in this case ∗ T22 = −T11 ,

∗ T12 = T12 ,

∗ T21 = T21 .

(13.18)

In this symmetric case the coeﬃcients Tij , 1 ≤ i, j ≤ 2, are square matrices, all of the same order, n say. In what follows H will denote the Hamiltonian of (13.16), that is, H = 2 Tij i,j=1 . Note that the identities in (13.18) hold if and only if

∗

JH = −H J,

0 where J = −In

In . 0

(13.19)

We are now ready to state the lemma. Lemma 13.4. Let X be a solution of (13.16) such that σ(T22 + T21 X) ⊂ Cleft. If, in addition, the coeﬃcients of (13.16) satisfy the identities in (13.18), then the Hamiltonian H has no pure imaginary eigenvalues and σ(T11 − XT21 ) ⊂ Cright .

240

Chapter 13. The symmetric algebraic Riccati equation

Proof. We shall use freely the results of Section 12.1. Let N be the angular subspace determined by X. Then N is invariant under the Hamiltonian H and the restriction H|N is similar to the matrix T22 + T21 X. Since the identities in (13.18) are satisﬁed, (13.19) holds. The symmetry relation JH = −H ∗ J implies that the eigenvalues of H are placed symmetrically with respect to the imaginary axis (multiplicities included). Note that the dimension of the angular subspace N is equal to n, where n is the size of the matrices Tij , 1 ≤ i, j ≤ 2. Since N is invariant under H and H|N is similar to T22 + T21 X, the condition on the spectrum of T22 + T21X, implies that σ(H|N ) ⊂ Cleft . It follows that H has at least n eigenvalues (multiplicities taken into account) in Cleft. The symmetry referred to above then gives that H also has at least n eigenvalues in Cright . But the order of H is 2n. So H has precisely n eigenvalues in Cleft , and also precisely n eigenvalues in Cright . In particular, H has no eigenvalue on the imaginary axis. Next, recall formula (12.5) for the present setting, that is, E

−1

HE =

T11 − XT21

0

T21

T22 + T21 X

, where E =

In 0

X In

.

(13.20)

As H and E −1 HE have the same set of eigenvalues (multiplicities taken into account) and σ(T22 + T21 X) ⊂ Cleft , the result of the previous paragraph implies that σ(T11 − XT21 ) ⊂ Cright , which completes the proof. Corollary 13.5. Assume the coeﬃcients of the Riccati equation (13.16) satisfy the symmetry conditions in (13.18). Then equation (13.16) has at most one solution X such that σ(T22 + T21 X) ⊂ Cleft . Moreover, this solution, if it exists, is Hermitian. Proof. Assume X is a solution of (13.16) such that σ(T22 + T21 X) is a subset of Cleft. Then, by Lemma 13.4, the Hamiltonian H has no pure imaginary eigenvalues and σ(T11 − XT21 ) ⊂ Cright . But then we can apply Proposition 12.1 to show that the angular subspace N determined by X is the spectral subspace of H corresponding to the eigenvalues of H in the open left half plane. In particular, N is uniquely determined and does not depend on the particular choice of the solution X. This implies that X is also uniquely determined. Again assume that X is a solution of (13.16) such that σ(T22 + T21 X) is a subset of Cleft . Then σ(T11 − XT21) ⊂ Cright . By taking adjoints and using the identities in (13.18) we see that the latter inclusion implies that σ(T22 + T21 X ∗ ) is a subset of Cleft . Furthermore, from the identities in (13.18) it also follows that X ∗ is a solution of (13.16). But then, by the uniqueness result of the previous paragraph, X ∗ = X. Hence X is Hermitian, as desired. Proof of Theorem 13.3. The implication (i) ⇒ (ii) and the ﬁnal statements of the theorem follow directly by applying Lemma 13.4 and Corollary 13.5 with the coeﬃcients Tij , 1 ≤ i, j ≤ 2, being taken as in (13.17). It remains to prove the implication (ii) ⇒ (i). Let F be an m × n matrix such that A− BF has all its eigenvalues in the open left half plane. Such a matrix exists

13.2. Stabilizing solutions

241

because (A, B) is stabilizable. Introduce the rational m × m matrix function ∗ F R ∗ V (λ) = R + B − RF (λ − G)−1 , (13.21) B

where G=

−A∗ + F ∗ B ∗

−Q − F ∗ RF

0

A − BF

.

The fact that R is invertible implies that the realization (13.21) is biproper, and one veriﬁes easily that the associate main operator is precisely the Hamiltonian T . Thus F∗ −1 ∗ −1 −1 −1 V (λ) = R − R B − F (λ − T ) . BR−1 Since A−BF has all its eigenvalues in the open left half plane, G has no eigenvalue on the imaginary axis. By assumption the same holds true for T . Thus V has no poles or zeros on iR. In particular, V (λ) is invertible for each λ ∈ iR. With J as in (13.19) we have ∗ F R ∗ = B ∗ − RF . JG = −G∗ J, J B So, by the remark made after the proof of Theorem 9.1, the values of V on iR are selfadjoint matrices. Since V (λ) is invertible for each λ ∈ iR, it follows that the signature of the matrices V (λ) for λ ∈ iR, i.e., the diﬀerence between the number of positive and negative eigenvalues of the selfadjoint matrix V (λ), is constant. As V (∞) = R is positive deﬁnite, we obtain that V (λ) is positive deﬁnite for λ ∈ iR. Hence we know from Theorem 9.5 that V admits a left spectral factorization with respect to iR. To ﬁnish the proof of (ii) ⇒ (i), we apply (the appropriate version of) Theorem 12.5 (see the remark made between the theorem and its proof) with A11 = −A∗ + F ∗ B ∗ , B1 = F ∗ R,

B2 = B,

A12 = −Q − F ∗ RF, C1 = B ∗ ,

C2 = −RF,

A22 = A − BF, D = R.

Via these choices, equation (12.10) transforms into (13.14) with P as the unknown. Furthermore, the inclusions (12.11) and (12.12) change into σ(−A∗ + P BR−1 B ∗ ) ⊂ Cright ,

σ(A − BR−1 B ∗ P ) ⊂ Cleft .

(13.22)

The conclusion is that equation (13.14) has a unique solution P satisfying the inclusions in (13.22). The second of these shows that P is a stabilizing solution of (13.14). Thus (i) is proved.

242

Chapter 13. The symmetric algebraic Riccati equation

Let P be an iR-stabilizing solution of (13.14). Then by deﬁnition, the spectral inclusion σ(A − BR−1 B ∗ P ) ⊂ Cleft holds. Furthermore, since P is Hermitian, also σ(−A∗ + P BR−1B ∗ P ) ⊂ Cright ; see also Lemma 13.4. So one of the spectral inclusions in (13.22) implies the other one automatically; cf., the two spectral inclusions (12.11), (12.12).

13.3 Symmetric Riccati equations and pseudo-spectral factorization We now continue the discussion of Section 13.2. The object of study will be the algebraic Riccati equation A∗ P + P A + Q − (P B + S ∗ )R−1 (B ∗ P + S) = 0.

(13.23)

Observe that compared to (13.14) there are some additional terms. On the other hand, (13.23) can be rewritten in the more familiar form (13.14) as (A∗ − S ∗ R−1 B ∗ )P + P (A − BR−1 S) + (Q − S ∗ R−1 S) − P BR−1 B ∗ P = 0. The Hamiltonian of this equation is given by −A∗ + S ∗ R−1B ∗ −Q + S ∗ R−1 S T = . −BR−1 B ∗ A − BR−1 S

(13.24)

Also of importance is the rational matrix function (λ − A)−1 B Q S∗ ∗ ∗ −1 I . W (λ) = −B (λ + A ) S R I

(13.25)

Note that W is selfadjoint on the imaginary axis, and admits the realization $ %−1 −A∗ −Q −S ∗ ∗ S λ− . (13.26) W (λ) = R + B B 0 A For the inverse of W , one computes that W (λ)−1 = R−1 − R

−1

B∗

S

(λ − T )−1

−S ∗ B

R−1 .

Letting n be the order of the matrix A and the skew-Hermitian 2n × 2n matrix J as in (13.19), we have ∗ ) ( −A∗ −Q −A∗ −Q ∗ −S ∗ J = B∗ S =− J, J , B 0 A 0 A

13.3. Symmetric Riccati equations and pseudo-spectral factorization

243

and hence also JT = −T ∗ J. The hypotheses we shall have in eﬀect in this section are more stringent than those in Section 13.2. In fact, we shall assume A is an n × n matrix and B an n × m matrix such that (A, B) is a controllable pair (as opposed to the weaker condition of stabilizability). As in Section 13.2 we take R positive deﬁnite and Q selfadjoint. In the next theorem we characterize when the function W introduced above is nonnegative on the imaginary axis. The characterization is given in terms of the existence of Hermitian solutions of the Riccati equation (13.23). Also we specify further the pseudo-spectral factorization result in Theorem 10.2, again in terms of Hermitian solutions of (13.23). Theorem 13.6. Consider the Riccati equation (13.23) with (A, B) a controllable pair, R positive deﬁnite and Q selfadjoint. Let T be the matrix given by (13.24) and let W be the rational matrix function deﬁned by (13.25). Then the following statements are equivalent: (i) Equation (13.23) has a Hermitian solution P ; (ii) The rational matrix function W is nonnegative on the imaginary axis; (iii) The partial multiplicities of T at its pure imaginary eigenvalues are all even; (iv) There exists a T -invariant subspace M such that J[M ] = M ⊥ . In that case, so if the equivalent conditions (i)−(iv) hold, then, given a Hermitian solution P of (13.23), the rational matrix function W (λ) factors as ¯ ∗ L(λ), W (λ) = L(−λ) where

(13.27)

L(λ) = R1/2 + R−1/2 (B ∗ P + S)(λIn − A)−1 B.

(13.28)

⊥

Moreover, if M is a T -invariant subspace such that J[M ] = M , then M is of the form ( ) P M = Im In for a Hermitian solution P of (13.23). In addition, if both A and T |M have all their eigenvalues in the closed left half plane, then the factorization (13.27) is a pseudo-spectral factorization with respect to the imaginary axis. Proof. (i) ⇒ (ii) Suppose (13.23) has a Hermitian solution P . With this P , deﬁne L(λ) by (13.28). We then have ¯ ∗ L(λ) = L(−λ)

R − B ∗ (λ + A∗ )−1 (S ∗ + P B) + (B ∗ P + S)(λ − A)−1 B −B ∗ (λ + A∗ )−1 (S ∗ + P B)R−1 (B ∗ P + S)(λ − A)−1 B.

Using (13.23), one rewrites the last term as B ∗ (λ + A∗ )−1 Q + (λ − A∗ )P + P (A − λ) (λ − A)−1 B

244

Chapter 13. The symmetric algebraic Riccati equation

which, in turn, can be transformed into B ∗ (λ + A∗ )−1 Q(λ − A)−1 B + B ∗ P (λ − A)−1 B − B ∗ (λ + A∗ )−1 P B. ¯ ∗ L(λ) = W (λ), and (ii) holds. Moreover, the identity (13.27) is proved. Thus L(−λ) (ii) ⇒ (iii) To prove that (ii) implies (iii) a couple of preparatory remarks are needed. Let A be an n × n matrix and B be an n × m matrix. For any m × n matrix F introduce AF = A − BF, Then,

QF

SF∗

SF

R

=

I

−F ∗

0

I

Thus W (λ)

=

QF = Q − S ∗ F − F ∗ S + F ∗ RF.

SF = S − RF,

−B ∗ (λ + A∗ )−1

I

S I

F∗

0

I ·

=

−B(λ + A∗ )−1

WF (λ) =

∗

−B (λ +

R

QF

I

0

−F

I

SF∗

.

I

SF R 0 (λ − A)−1 B

F

I

I

I − B ∗ (λ + A∗ )−1 F ∗ QF SF∗ (λ − A)−1 B · . I + F (λ − A)−1 B SF R

Now introduce

Q S∗

A∗F )−1

I

QF

SF∗

SF

R

(λ − AF )−1 B I

,

and Φ(λ) = I + F (λ − A)−1 B. Then Φ(λ)−1 = I − F (λ − AF )−1 B. Using the fourth identity in Theorem 2.4 one sees that (λ − A)−1 BΦ(λ)−1 = (λ − AF )−1 B. ¯ ∗ WF (λ)Φ(λ). So W (λ) is nonnegative for λ ∈ iR if and only Thus W (λ) = Φ(−λ) if WF (λ) is nonnegative for λ ∈ iR, provided λ is not a pole of the functions involved. Next, notice that WF (λ) has the realization %−1 $ −SF∗ −A∗F −QF ∗ SF WF (λ) = R + B . λ− 0 AF B One readily computes that −A∗F −QF 0

AF

−

−SF∗ B

R−1

B∗

SF

= T,

13.3. Symmetric Riccati equations and pseudo-spectral factorization

245

where T is given by (13.24). So WF (λ)

−1

= R

−1

−R

−1

B

∗

SF

(λ − T )

−1

−SF∗ B

R−1 .

Since the pair (A, B) is controllable, we can use the pole placement theorem from mathematical systems theory (see Theorem 19.3 in Chapter 20 below), to conclude that there exists an m × n matrix F such that all the eigenvalues of AF are in the open left half plane. Using such an F , we see that the matrix −A∗F −QF 0

AF

has no imaginary eigenvalues. This allows us to show (see formula (4.7) in Section 4.3 in [20]) that the matrix functions WF (λ) 0 λI2n − T 0 and 0 I2n 0 Im are analytically equivalent on an open set containing the imaginary axis. It follows that for each λ ∈ iR the partial multiplicities of λ as an eigenvalue of T are equal to the partial multiplicities of λ as a zero of WF . Since WF is nonnegative on iR, we know from Proposition 10.4 that the partial multiplicities of λ ∈ iR as a zero of WF are even. Hence the partial multiplicities at the pure imaginary eigenvalues of T are even. Thus (ii) implies (iii). (iii) ⇒ (iv) This implication can be seen from Theorem 11.4 in Chapter 11 applied to A = iT and H = iJ. Indeed, since there are no odd partial multiplicities corresponding to pure imaginary eigenvalues of T , the condition of Theorem 11.4 is satisﬁed. Hence there exists an A-invariant subspace M such that H[M ] = M ⊥ . This subspace then is also T -invariant and satisﬁes J[M ] = M ⊥ . (iv) ⇒ (i) Let M be T -invariant subspace such that J[M ] = M ⊥ , and write X1 , M = Im X2 for appropriate n × n matrices X1 and X2 . It will be shown that X2 is invertible. Once this is done, we can take P = X1 X2−1. From T [M ] ⊂ M one obtains that P solves (13.23), while from J[M ] = M ⊥ one has P = P ∗ . Hence (i) holds. We have also shown that any T -invariant J-Lagrangian subspace M is the graph of a Hermitian ∗solution P of the Riccati equation, that is, M is of the form M = Im P I for a matrix P = P ∗ that solves (13.23). It remains to verify that Ker X2 = {0}. As dim M = n, the null spaces Ker X1 and Ker X2 have a trivial intersection. So it is suﬃcient to establish that X2 x = 0

246

Chapter 13. The symmetric algebraic Riccati equation

implies X1 x = 0. Let X2 x = 0. Then X1 x ∈ M, 0 and hence

T

X1 x

0

=

−A∗ X1 x + S ∗ R−1B ∗ X1 x −BR−1 B ∗ X1 x

∈ M.

Now M is iJ-Lagrangian, i.e., J[M ] = M ⊥ . So / 0=

T

X1 x 0

,J

X1 x 0

0 = −R−1 B ∗ X1 x, B ∗ X1 x.

As R is positive deﬁnite, we obtain B ∗ X1 x = 0. Hence −A∗ X1 x X1 x T = . 0 0 But this vector is in M , so it must be of the form X1 y . X2 y Thus X2 y = 0 and X1 y = −A∗ X1 x. As X2 y = 0, we have B ∗ X1 y = 0 by the argument given above. So B ∗ A∗ X1 x = 0. Now consider ∗2 A X1 x X1 x −A∗ X1 x 2 T =T = . 0 0 0 Repeating the argument we get B ∗ A∗2 X1 x = 0. Continuing in this way we arrive at X1 x ∈ Ker B ∗ A∗j for all j. As (A, B) is a controllable pair, the pair (B ∗ , A∗ ) is observable, and thus we see that X1 x = 0, as desired. It is easily seen that the eigenvalues of T |M coincide with those of the matrix A − BR−1 S − BR−1 B ∗ P . Thus, if both A and T |M have their eigenvalues in the closed left half plane, then the factorization (13.27) with L given by (13.28) is a pseudo-spectral factorization. Notice that the full force of the controllability condition on the pair (A, B) was only used in the last part of the proof. More precisely, the implications (i) ⇒ (ii) and (iii) ⇒ (iv) are true without any condition on (A, B), and for the implication (ii) ⇒ (iii) only stabilizability of (A, B) was used.

13.3. Symmetric Riccati equations and pseudo-spectral factorization

247

Notes The connection between Riccati equations and factorizations as discussed in Section 13.1 goes back to [147] and [41]. The main result of Section 13.2 originates from [102], see also [106], Section 9.3. The results of Section 13.1 and 13.2, and similar results for the discrete time algebraic Riccati equation, play an important role in several problems in mathematical systems theory, notably, LQ-optimal control, Kalman ﬁltering and stochastic realization (see, e.g., [84], [85], [33]). The main result of Section 13.3 appeared for the ﬁrst time in [105] and [34]. See also Chapter 7 in [106]. The parametrization of solutions of the algebraic Riccati equation in terms of invariant subspaces of the matrix T , as described in Theorem 13.6, also plays a role in [135], [136].

Chapter 14

J-spectral factorization In this chapter we continue the study of rational matrix functions that take Hermitian values on certain contours. In contrast to the previous chapters, the emphasis will not be on positive deﬁnite or nonnegative rational matrix functions, but rather on ones that have values for which the inertia is independent of the point on the contour. Such functions may still admit a symmetric canonical factorization, provided we allow for a constant Hermitian invertible matrix as a middle factor. Such a factorization is commonly known as a J-spectral factorization. We shall give necessary and suﬃcient conditions for its existence, and study the question when a function which admits a left J-spectral factorization also admits a right J-spectral factorization. This chapter consists of seven sections. The ﬁrst four sections and the one but last deal with J-spectral factorization with respect to the imaginary axis. Section 14.1 introduces the notion of J-spectral factorization. The next two sections provide necessary and suﬃcient conditions for the existence of such factorizations; in Section 14.2 these conditions are stated in terms of certain invariant subspaces and in Section 14.3 they are given in terms of Riccati equations. Two special cases are discussed in detail in Section 14.4. The ﬁfth section (Section 14.5) deals with Jspectral factorization with respect to the unit circle and the real line. Section 14.6 concerns the topic of left versus right J-spectral factorization. In Section 14.7 an alternative approach is used to derive J -spectral factorizations with respect to the unit circle. The main result of this ﬁnal section extends to a more general setting the ﬁrst main result of Section 14.5.

14.1 Deﬁnition of J-spectral factorization Throughout this chapter J is an invertible Hermitian m × m matrix. Often we shall assume additionally that J −1 = J. Thus in that case we have J = J ∗ = J −1 .

(14.1)

250

Chapter 14. J-spectral factorization

Such a matrix is called a signature matrix. Up to a congruence transformation any selfadjoint invertible matrix is a signature matrix. Suppose W is a rational m × m matrix function. A factorization ¯ ∗ JL(λ) W (λ) = L(−λ)

(14.2)

is called a right J-spectral factorization with respect to the imaginary axis if L and L−1 are rational m × m matrix functions which are analytic on the closed left ¯ ∗ and its inverse are half plane (inﬁnity included). In that case the function L(−λ) analytic on the closed right half plane (including inﬁnity). Thus a right J-spectral factorization with respect to the imaginary axis is a right canonical factorization with respect to iR featuring an additional symmetry property between the factors. A factorization (9.3) is called a left J-spectral factorization with respect to the imaginary axis if L and L−1 are rational m×m matrix functions which are analytic ¯ ∗ on the closed right half plane (inﬁnity included), in which case the function L(λ) and its inverse are analytic on the closed left half plane (inﬁnity included). Such a factorization is a left canonical factorization with respect to iR. The existence of a right or left J-spectral factorization implies that W admits a canonical factorization with respect to the imaginary axis. In particular, in order that a right or left J-spectral factorization of W exists it is necessary that W is biproper and has no poles or zeros on the imaginary axis. Furthermore, the identity (14.2) gives that W is selfadjoint on the imaginary axis. Contrary to spectral factorizations for positive deﬁnite rational matrix functions, J-spectral factorizations do not always exist for biproper rational matrix functions that satisfy the obvious necessary conditions mentioned in the previous paragraph. Since a J-spectral factorization is a canonical factorization, we can use Theorem 3.2 to prepare for an example of this phenomenon. Let ⎡ ⎤ λ−1 0 ⎢ λ + 1⎥ ⎥. W (λ) = ⎢ (14.3) ⎣λ + 1 ⎦ 0 λ−1 Obviously, W is biproper and its values on the imaginary axis are selfadjoint. Furthermore, W has no pole or zero on the imaginary axis. The function W has the minimal realization W (λ) = D + C(λ − A)−1 B, with 0 1 1 0 1 0 0 1 D= , A= , B= , C= . (14.4) 1 0 0 −1 0 −2 2 0 The associate main operator is given by A× = A − BD−1 C =

−1 0 0

1

= −A.

14.2. J-spectral factorizations and invariant subspaces

251

Now for a right canonical factorization with respect to the imaginary axis to exist, × ˙ M+ we must have C2 = M− + , where M− is the spectral subspace of A associated × with the part of σ(A) lying in the left half plane, and M+ is the spectral subspace of × × A associated with the part of σ(A ) lying in the right half plane. However, since × . Hence a right canonical factorization in this case A× = −A, we have M− = M+ of W with respect to iR does not exist. Analogously, a left canonical factorization does not exist either. Hence neither left nor right J-spectral factorizations of W with respect to the imaginary axis exist for any choice of J = J ∗ = J −1 . To further clarify the connection between J-spectral factorization and canonical factorization we present the following proposition. Proposition 14.1. Let W be a biproper rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Then W (∞) is congruent to a signature matrix J, and for such a matrix J the function W admits a right (respectively, left) J-spectral factorization with respect to the imaginary axis if and only if it admits a right (respectively, left) canonical factorization with respect to the imaginary axis. Proof. Since W is selfadjoint on the imaginary axis and proper, we see that D = W (∞) is well-deﬁned and selfadjoint. The fact that W is biproper means that D is invertible. Thus D is an invertible selfadjoint matrix, and hence congruent to a signature matrix, J say: D = E ∗ JE for some invertible matrix E. Let W (λ) = W− (λ)W+ (λ) be a right canonical factorization of W with respect to the imaginary axis. Since W , W− and W+ are biproper we have D = D− D+ , where D− = W− (∞) and D+ = W+ (∞). It follows that the factorization W (λ) = W− (λ)W+ (λ) can be rewritten as W (λ) = V− (λ)DV+ (λ), where −1 , V− (λ) = W− (λ)D−

−1 V+ (λ) = D+ W− (λ).

In particular, the values of V− and V+ at inﬁnity are equal to the m × m identity matrix. Since V+ and V+−1 are analytic on the closed right half plane (inﬁnity included) and the functions V− and V−−1 are analytic on the closed left half plane (inﬁnity included), the factorization is unique. Now we use that D is selfadjoint and that W is selfadjoint on the imaginary axis. It follows that ¯ ∗ DV− (−λ) ¯ ∗, W (λ) = V+ (−λ) and in this factorization the factors have the same analyticity properties as those in W (λ) = V− (λ)DV+ (λ). Because of the uniqueness of the latter factorization, ¯ ∗ . Recall that D = E ∗ JE. Put L(λ) = EV+ (λ). we conclude that V− (λ) = V+ (−λ) ∗ ¯ Then W (λ) = L(−λ) JL(λ), and this factorization is a left J-spectral factorization with respect to the imaginary axis. The reverse implication is trivial.

14.2 J-spectral factorizations and invariant subspaces In this section necessary and suﬃcient conditions for existence of a right or left J-spectral factorization with respect to the imaginary axis will be derived in terms

252

Chapter 14. J-spectral factorization

of invariant subspaces. It will be assumed that the obvious necessary conditions for the existence of a J-spectral factorization are satisﬁed, that is, the rational m × m matrix function W for which we wish to ﬁnd J-spectral factorizations with respect to iR is assumed to be biproper, to have no poles or zeros on iR, and to be selfadjoint on iR. We begin with two lemmas which can be viewed as further reﬁnements of Theorem 9.1(ii). Lemma 14.2. Let W be a biproper rational m×m matrix function that is selfadjoint on the imaginary axis and has no pole there. Then W admits a minimal realization W (λ) = D + B ∗ H ∗ (λI2n − A)−1 B,

(14.5)

such that D = D∗ is invertible, H is invertible, HA = −A∗ H, and the matrices A and H partition as A11 A12 , A= 0 A22

H ∗ = −H, H=

(14.6)

0

∗ −H21

H21

H22

,

(14.7)

where A11 and A22 are n × n matrices which have all their eigenvalues in the right open half plane and left open half plane, respectively. Proof. Since W is biproper, D = W (∞) is invertible. The fact that D is selfadjoint is covered by item (ii) in Theorem 9.1. p − A) −1 B be a minimal realization of W . The Next, let W (λ) = D + C(λI fact that W has no poles on iR and the minimality of the realization imply that has no eigenvalue on iR. Furthermore, using item (ii) of Theorem 9.1 again, we A know that there exists a unique invertible p × p matrix T for which we have = −A ∗ T, TA

=C ∗ , TB

T = −T∗ .

(14.8)

corresponding to the eigenvalues in the Let N+ be the spectral subspace of A ⊥ ∗ T yields T[N+ ] = N+ open right half plane. The identity T A = −A . But then ⊥ the invertibility of T implies that dim N+ = dim N+ . The latter can only happen when p is even, that is, p = 2n for some nonnegative integer n. In particular, dim N+ = n. Now let f1 , . . . , fn be an orthogonal basis of N+ , and let fn+1 , . . . , f2n ⊥ ⊥ be an orthogonal basis of N+ . Since Cn = N+ ⊕ N+ , the vectors f1 , . . . , f2n form an orthogonal basis of C2n , and we can consider the unitary matrix U that transforms the basis f1 , . . . , f2n into the standard basis e1 , . . . , e2n of C2n . Deﬁne −1 , A = U AU

B = U B,

−1 , C = CU

H = U TU ∗ .

Then W (λ) = D + C(λI2n − A)−1 B is a minimal realization of W . The fact that U −1 = U ∗ together with (14.8) shows that HA = −A∗ H,

HB = C ∗ ,

H = −H ∗ .

14.2. J-spectral factorizations and invariant subspaces

253

Thus W is of the form (14.5) and (14.6) holds. The spectral subspace M+ of A corresponding to the eigenvalues in the open right half plane is given by M+ = span {e1 , . . . , en }.

(14.9)

The ﬁrst identity in (14.8) yields ⊥ = span {en+1 , . . . , e2n }. H [span {e1 , . . . , en } = H[M+ ] = M+

(14.10)

It follows that the matrices A, and H can be partitioned as in (14.7). All blocks in these representations of A and H are n × n matrices. The zero entry in A follows from the A-invariance of M+ and the fact that this space is given by (14.9), while the zero entry in H follows from (14.10). The deﬁnition of M+ and the identity (14.9) also imply that all the eigenvalues of A11 are in the open right half plane and those of A22 are in the open left half plane. Lemma 14.3. Let W be a biproper rational m×m matrix function that is selfadjoint on the imaginary axis and has no pole there. Then W admits a minimal realization W (λ) = D + C(λI2n − A)−1 B,

(14.11)

such that D = D∗ is invertible and the matrices A, B and C can be partitioned as B1 −A∗22 A12 , B= , C = −B2∗ B1∗ , (14.12) A= 0 A22 B2 where A12 is a selfadjoint n × n matrix, A22 is a n × n matrix which has all its eigenvalues in the open left half plane, and both B1 and B2 are n × m matrices. Proof. From the preceding lemma we know that W admits a minimal realization −1 B, ∗H ∗ (λI2n − A) W (λ) = D + B is invertible, where D = D∗ is invertible, H A = −A ∗ H, H

∗ = −H, H

and H partition as and the matrices A = A

11 A

12 A

0

22 A

,

= H

0

∗ 21 −H

21 H

22 H

,

(14.13)

11 are in the open right half plane and those of A 22 such that the eigenvalues of A are in the open left half plane.

254

Chapter 14. J-spectral factorization is invertible, it follows that H 21 is invertible, and hence we can deﬁne Since H ⎤ ⎡ 1 −1 −1 ⎢ H21 − 2 H21 H22 ⎥ S=⎣ ⎦. 0 In

=B ∗H ∗ , and consider the matrices The matrix S is invertible. Put C A = S −1 AS,

B = S −1 B,

−1 , C = CS

H = S ∗ HS.

Obviously, W (λ) = D + C(λIn − A)−1 B is a minimal realization of W . It remains to prove that A, B, C can be partitioned in the desired way. A straightforward calculation shows that 0 −In ∗ ∗ . (14.14) HA = −A H, HB = C , H= In 0 S, and S −1 are all block upper triangular, the same holds true Since the matrices A, for A. The ﬁrst identity in (14.14) together with the third identity in (14.14) shows that A is of the form given in (14.12) with A12 being selfadjoint. Furthermore, since the entry in the right lower corner of S and S −1 is the n × n identity matrix we 22 , and hence A22 is an n × n matrix which has all its eigenvalues see that A22 = A in the open left half plane. The second and third identities in (14.14) show that B and C are as in (14.12). Obviously, B1 and B2 are matrices of size n × m. The external matrix D in the realizations (14.5) and (14.11) is congruent to a signature matrix J, that is, D = E ∗ JE for some invertible matrix E. Replacing W (λ) by (E ∗ )−1 W (λ)E −1 we may assume that the external matrix is actually equal to J. In the next theorem we shall make this assumption. Theorem 14.4. Let W be a rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Suppose W is given by W (λ) = J + C(λI2n − A)−1 B, where J is a signature matrix and −A∗22 A12 B1 A= , B= , 0 A22 B2

C=

−B2∗

B1∗

,

such that A12 is a selfadjoint n × n matrix, A22 is an n × n matrix which has all its eigenvalues in the open left half plane, and both B1 and B2 are n × m matrices. Then W admits a left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ),

14.2. J-spectral factorizations and invariant subspaces if and only if

⎡ A× = ⎣

−A∗22 + B1 B2∗

A12 − B1 B1∗

B2 B2∗

A22 − B2 B1∗

255 ⎤ ⎦

has no eigenvalues on the imaginary axis, and the spectral subspace of A× corresponding to its eigenvalues in the open left half plane is of the form X Im In for some Hermitian matrix X. In that case the unique left J-spectral factor L− for which L− (∞) = Im is given by L− (λ) = Im + J −1 (B1∗ − B2∗ X)(λIn − A22 )−1 B2 . In this expression (as well as in other comparable formulas below) the matrix J −1 can be replaced by J. Proof. In order to prove the ﬁrst part of the theorem, we have only to check when W admits a left canonical factorization with respect to the imaginary axis (see Proposition 14.1). Let M be a spectral subspace of A corresponding to its eigenvalues in the open right half plane. Then M = Im [I 0]∗ . Writing M × for the spectral subspace of A× corresponding to its eigenvalues in the open left half plane, the matching condition ˙ M× (14.15) Cn = M + is satisﬁed if and only if M × = Im [X ∗ I]∗ for some matrix X. With H as in (14.14), the subspace M × is iH-Lagrangian (see Section 11.1). Thus −I Im = H[M × ] = (M × )⊥ = Ker X ∗ I , X which implies X = X ∗ . Applying the left-version of Theorem 3.2 the ﬁrst part of the theorem follows. Next let us deal with the second part. So suppose (14.15) is satisﬁed and write the projection Π of Cn along M onto M × in the form 0 X Π= . 0 I Then the unique right hand factor L− in a left canonical factorization with respect to the imaginary axis of W , satisfying the additional condition that L(∞) = Im ,

256

Chapter 14. J-spectral factorization

is given by L− (λ)

=

I + J −1 CΠ(λ − ΠAΠ)−1 ΠB −1

0

B1∗

−

B2∗ X

$

λ−

0 XA22

=

I +J

=

I + J −1 (B1∗ − B2∗ X)(λ − A22 )−1 B2 ,

0

%−1

XB2

A22

B2

as was claimed.

In Section 14.5 below we shall consider J-spectral factorization for selfadjoint rational matrix functions on the real line or on the unit circle.

14.3 J-spectral factorizations and Riccati equations In this section, necessary and suﬃcient conditions for existence of a right or left Jspectral factorization with respect to the imaginary axis will be derived in terms of Riccati equations. It will be assumed that the obvious necessary conditions for the existence of a J-spectral factorization are satisﬁed, that is, the rational m × m matrix function W for which we wish to ﬁnd J-spectral factorizations with respect to iR is assumed to be biproper, to have no poles or zeros on iR, and to be selfadjoint on iR. As in Theorem 14.4 we assume that the external matrix (that is, the value at inﬁnity) is a signature matrix. Theorem 14.5. Let W be a rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Suppose W is given by W (λ) = J + C(λI2n − A)−1 B, where J is a signature matrix and B1 −A∗22 A12 , B= , A= 0 A22 B2

C=

−B2∗

B1∗

,

such that A12 is a selfadjoint n × n matrix, A22 is an n × n matrix which has all its eigenvalues in the open left half plane, and both B1 and B2 are n × m matrices. Then W admits a left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ), if and only if the algebraic Riccati equation XB2 J −1 B2∗ X + X(A22 − B2 J −1 B1∗ ) + (A∗22 − B1 J −1 B2∗ )X −A12 + B1 J

−1

B1∗

=0

(14.16)

14.3. J-spectral factorizations and Riccati equations

257

has a (unique) iR-stabilizing Hermitian solution X. In that case the unique left J-spectral factor L− for which L− (∞) = Im is given by L− (λ) = Im + J −1 (B1∗ − B2∗ X)(λIn − A22 )−1 B2 .

(14.17)

In line with the deﬁnition given in the paragraph preceding Theorem 13.14, a solution of (14.16) is said to be iR-stabilizing (or simply stabilizing) if the matrix A22 − B2 J −1 B1∗ + B2 J −1 B2∗ X has its eigenvalues in the open left half plane. Proof. In order to prove the ﬁrst part of the theorem, we have only to check when W admits a left canonical factorization with respect to the imaginary axis (see Proposition 14.1). A straightforward application of Theorem 12.5, with F+ equal to Cleft and F− equal to Cright , tells us that W admits a left canonical factorization with respect to the imaginary axis if and only if the Riccati equation (14.16) has a unique solution X satisfying the additional spectral constraints (14.18) σ − A∗22 + (B1 − XB2 )J −1 B2∗ ⊂ Cright , σ A22 − B2 J −1 (B1∗ − B2∗ X) ⊂ Cleft . (14.19) Next, note that X satisﬁes (14.16) and the spectral constraints (14.18) and (14.19) if and only if the same holds true for X ∗ . Because of uniqueness it follows that X = X ∗ . The second spectral constraint (14.19) means that X is a stabilizing solution of (14.16). This completes the proof of the ﬁrst part of the theorem. To prove the second part one applies the second part of Theorem 12.5 with D1 = J and D2 = Im . Theorem 14.6. Let W be a rational m × m matrix function that is selfadjoint on the imaginary axis and has no pole there. Suppose W is given by W (λ) = J + B ∗ H ∗ (λI2n − A)−1 B, where J is a signature matrix, H is invertible, HA = −A∗ H and H ∗ = −H, and the matrices A and H partition as ∗ 0 −H21 A11 A12 , H= , A= 0 A22 H21 H22 where A11 and A22 are n × n matrices which have all their eigenvalues in the open right half plane and open left half plane, respectively. Put 12 A 1 B

1 ∗ 1 A22 H22 + H22 A22 + H21 A12 , 2 2 1 = H21 B1 + H22 B2 . 2 =

(14.20) (14.21)

258

Chapter 14. J-spectral factorization

Then W admits a left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ), if and only if the algebraic Riccati equation 1 J −1 B ∗ )X (14.22) ∗ ) + (A∗ − B XB2 J −1 B2∗ X + X(A22 − B2 J −1 B 1 22 2 −1 ∗ −A12 + B1 J B = 0. 1

has a (unique) iR-stabilizing Hermitian solution X. In that case the unique left J-spectral factor L− for which L− (∞) = Im is given by ∗ − B ∗ X)(λIn − A22 )−1 B2 . L− (λ) = Im + J −1 (B 1 2

(14.23)

Recall that an iR-stabilizing solution X of (14.22) is one for which the matrix 1∗ + B2 J −1 B2∗ X has its eigenvalues in the open left half plane. A22 − B2 J −1 B = S −1 AS, B = S −1 B and Proof. Put C = B ∗ H ∗ , and consider the matrices A = CS, where C ⎡ ⎤ 1 −1 −1 H H − H 22 ⎢ 21 ⎥ 2 21 S=⎣ ⎦. 0 I −1 B, 2n − A) and from the proof of Lemma 14.3 we know Then W (λ) = J + C(λI that A, B and C partition as ⎡ ⎤ ⎡ ⎤ ∗ A 12 1 . −A B 22 =⎣ =⎣ = −B ⎦, ⎦, 1∗ , 2∗ B A B C 22 2 0 A B 2 = B2 . Since 22 = A22 and B where A ⎡ S −1

⎢H21 = ⎣ 0

⎤ 1 H22 ⎥ 2 ⎦, I

12 and B 2 are given by (14.20) and (14.21), respecone readily computes that A satisﬁes the −1 B 2n − A) tively. It follows that the realization W (λ) = J + C(λI conditions of Theorem 14.5. Note that the Riccati equation (14.16) transforms 1 and the matrix A12 by A 12 . Furinto equation (14.22) when B1 is replaced by B 1 , formula (14.17) transforms into (14.23). thermore, when passing from B1 to B But then we can apply Theorem 14.5 to ﬁnish the proof.

14.4. Two special cases of J-spectral factorization

259

Note that the procedure to ﬁnd the J-spectral factor, if it exists, now consists of two main steps. The ﬁrst is to ﬁnd a realization as in Theorem 14.6, which can be done by using an orthogonal basis transformation (see the proof of Lemma 14.2), and then to ﬁnd the stabilizing solution X of (14.22) in case it exists. With this in mind, let us return to the counterexample given in Section 14.1. Let W be the rational 2 × 2 matrix function given by (14.3). The realization of this function given in Section 14.1, involving the matrices featured in (14.4), can be rewritten as W (λ) = J + B ∗ H ∗ (λI2 − A)−1 B, where 0 1 1 0 1 0 0 −1 J= , A= , B= , H= . 1 0 0 −1 0 −2 1 0 This realization satisﬁes the conditions required in the ﬁrst part of Theorem 14.6. So it makes sense to check the situation with respect to the Riccati equation 1 = 1 0 . Since 12 = 0 and B featured in the theorem. Note that in this case A B2 = 0 −2 , it follows that in the algebraic Riccati equation (14.22) both the quadratic and the constant term vanish. Hence (14.22) reduces to a linear equation, namely 2x = 0. So x = 0 is the unique solution, and this solution is not stabilizing. Hence, W does not admit a J-spectral factorization with respect to the imaginary axis, which corroborates what was already observed in the paragraph preceding Proposition 14.1.

14.4 Two special cases of J-spectral factorization In this section we consider two special cases. The ﬁrst concerns the situation where the rational matrix function appears already as a product ¯ ∗ J V (λ) W (λ) = V (−λ)

(14.24)

where J is a signature matrix and V has all its poles in the open left half plane. This situation is encountered in several problems in mathematical systems theory, notably in the theory of H∞ -control (see Chapter 20 below). Let W be the rational m × m matrix function given by the product (14.24), where V (λ) = D+C(λIn −A)−1 B. Observe that W is selfadjoint on the imaginary axis. We assume that A has all its eigenvalues in the open left half plane and that the (possibly non-square) matrix D is of full column rank (that is, Ker D = {0}). The latter implies that D∗ J D is selfadjoint and invertible, and hence D ∗ J D is congruent to some signature matrix, J say. We are looking for a J-spectral factorization of D. Theorem 14.7. Let V (λ) = D + C(λIn − A)−1 B be a given rational p × m matrix function. Assume A has all its eigenvalues in the open left half plane and the p×m matrix D has full column rank. Let J be a p × p signature matrix, and let E be an invertible m × m matrix such that J = E ∗ D∗ J DE is an m × m signature

260

Chapter 14. J-spectral factorization

¯ ∗ J V (λ) has a matrix. Then the rational m × m matrix function W (λ) = V (−λ) left J-spectral factorization with respect to the imaginary axis, W (λ) = L− (−λ)∗ JL− (λ), if and only if the algebraic Riccati equation XBJ −1 B ∗ X + X(A − BJ −1 D∗ J C) + (A∗ − C ∗ J DJ −1 B ∗ )X (14.25) + C ∗ J DJD∗ J C − C ∗ J C = 0 has a (unique) iR-stabilizing Hermitian solution X. In that case, the corresponding left J-spectral factor of W is given by L− (λ) = E −1 + JE ∗ (D ∗ J C − B ∗ X)(λIn − A)−1 B. Recall that an iR-stabilizing solution X of (14.25) is one such that the matrix A − BJ −1 D∗ J C + BJ −1 B ∗ X has its eigenvalues in the open left half plane. = DE, B = BE, and consider the rational m × m matrix function Proof. Put D ¯ ∗ J V (λ), W (λ) = E ∗ W (λ)E = V (−λ) where V (λ) = V (λ)E = DE + C(λIn − A)−1 BE. Using the product rule for −1 B, 2n − A) realizations, we see that W admits the realization W (λ) = J + C(λI where ∗ −A∗ C ∗ J C C J DE = = = −E ∗ B ∗ E ∗ D∗ J C . A , B , C 0 A BE Obviously, W is selfadjoint on the imaginary axis. Furthermore, W is biproper. has Since A has all its eigenvalues in the open left half plane, we know that A no eigenvalue on iR, and hence W has no pole on iR. We conclude that the 2n − A) −1 B meets all the requirements of the ﬁrst realization W (λ) = J + C(λI part of Theorem 14.5. It follows that W admits a left J-spectral factorization with respect to the imaginary axis if and only if the Riccati equation (14.25) has a unique stabilizing Hermitian solution X. Moreover, in that case a left J-spectral ¯ ∗ JK− (λ) of W with respect to the imaginary axis factorization W (λ) = K− (−λ) is obtained by taking K− (λ) = Im + J −1 E ∗ (D ∗ J C − B ∗ X)(λIn − A)−1 BE. Recall that W (λ) = E −∗ W (λ)E −1 . It follows that W admits a left J-spectral ! . Thus factorization with respect to the imaginary axis if and only if so does W the result of the preceding paragraph shows that W admits a left J-spectral factorization with respect to the imaginary axis if and only if the Riccati equation (14.25) has a unique stabilizing Hermitian solution X. Moreover, in that case a ¯ ∗ JL− (λ) of W with respect to the left J-spectral factorization W (λ) = L− (−λ) imaginary axis is obtained by taking L− (λ) = K− (λ)E −1 .

14.4. Two special cases of J-spectral factorization

261

In our second example we assume that the rational m × m matrix function is given in the following manner (cf., the paragraph preceding Theorem 13.2): W (λ) = J + C(λIn − A)−1 B − B ∗ (λIn + A∗ )−1 C ∗ ,

(14.26)

where A has only eigenvalues in the open left plane and J is a signature matrix. The function W admits a realization $ %−1 ∗ −A∗ 0 C ∗ C W (λ) = J + −B λI2n − . (14.27) B 0 A This realization satisﬁes all the requirements of the ﬁrst part of Theorem 14.5, which yields immediately the following result. Theorem 14.8. Let the rational m × m matrix function W be given by (14.26), where J is a signature matrix and A has its eigenvalues in the open left half plane. Then W admits a left J-spectral factorization with respect to the imaginary axis, ¯ ∗ JL− (λ), W (λ) = L− (−λ) if and only if the algebraic Riccati equation XBJB ∗ X + X(A − BJC) + (A∗ − C ∗ JB ∗ )X + C ∗ JC = 0 has a (unique) Hermitian solution X such that the matrix A − BJC + BJB ∗ X has all its eigenvalues in the open left half plane (so X is iR-stabilizing). In that case the unique left J-spectral factor L− for which L− (∞) = Im and its inverse L−1 − are given by L− (λ) = Im + J(C − B ∗ X)(λIn − A)−1 B, So far we have mainly concentrated on left J-spectral factorizations. The analogous results for right J-spectral factorization of W can be obtained by simply applying the left factorization results to V (λ) = W (−λ). Indeed, a left J-spectral factorization, ¯ ∗ JK− (λ), V (λ) = K− (−λ) of V with respect to iR yields a right J-spectral factorization, ¯ ∗ JL+(λ), W (λ) = L+ (−λ) of W with respect to iR by taking L+ (λ) = K− (−λ). Let us apply this observation to W given by the realization (14.27). Note that %−1 $ −A 0 B ∗ V (λ) = W (−λ) = J + −C B . λI2n − 0 A∗ C∗ Since A has all its eigenvalues in the open left half plane, the same holds true for A∗ . Thus we can apply Theorem 14.8 together with the above scheme to get the following right J-spectral factorization result.

262

Chapter 14. J-spectral factorization

Theorem 14.9. Let the rational matrix function W be given by (14.26), where J is a signature matrix and A has its eigenvalues in the open left half plane. Then W admits a right J-spectral factorization with respect to the imaginary axis, ¯ ∗ JL+(λ), W (λ) = L+ (−λ) if and only if the algebraic Riccati equation Y C ∗ JCY + Y (A∗ − C ∗ JB ∗ ) + (A − BJC)Y + BJB ∗ = 0

(14.28)

has a (unique) Hermitian solution Y such that A∗ − C ∗ JB + C ∗ JCY has all its eigenvalues in the open left half plane (so X is iR-stabilizing). In that case the unique right J-spectral factor L+ for which L+ (∞) = Im and its inverse L−1 + are given by L+ (λ) = Im + J(CY − B ∗ )(λIn + A∗ )−1 C ∗ .

14.5 J-spectral factorization with respect to other contours In this section we consider J-spectral factorizations with respect to the real line R and to the unit circle T featuring an additional symmetry property between the factors. Here, as before, J is is an invertible Hermitian m × m matrix. We begin by considering the case of the unit circle. Suppose W is a rational m × m matrix function. A factorization ¯ −1 )∗ JL(λ) W (λ) = L(λ

(14.29)

is called a right J-spectral factorization with respect to the unit circle if L and L−1 are rational m × m matrix functions which are analytic on the closed unit disc. In ¯ −1 )∗ and its inverse are analytic on the closure of Dext that case the function L(λ (inﬁnity included). Thus a right J-spectral factorization with respect to the unit circle is a right canonical factorization with respect to T featuring an additional symmetry property between the factors. A factorization (14.29) is called a left J-spectral spectral factorization with respect to the unit circle if L and L−1 are rational m × m matrix functions which are analytic on the closure of Dext (inﬁnity ¯ −1 )∗ and its inverse are analytic on the included), in which case the function L(λ closed unit disc. Such a factorization is a left canonical factorization with respect to T. The case of J-spectral factorization with respect to the unit circle is somewhat more complicated than that of J-spectral factorization with respect to the imaginary axis. The ﬁrst result is an analogue of Proposition 14.1. Proposition 14.10. Let W be a rational m×m matrix function that is selfadjoint on the unit circle and has neither poles nor zeros there. Then there exists a signature matrix J such for each λ ∈ T the matrix W (λ) is congruent to J. For such a matrix

14.5. J-spectral factorization with respect to other contours

263

J, the function W admits a right (respectively, left) J-spectral factorization with respect to the unit circle if and only if it admits a right (respectively, left) canonical factorization with respect to the unit circle. We can use a M¨obius transform to reduce the case of the unit circle to the case of the imaginary axis. To be precise, let V (λ) = W (λ − i)/(λ + i) . Then V is a rational m × m matrix function that has neither poles nor zeros on the imaginary axis, and has selfadjoint values there. Moreover, V (∞) = W (1), and thus V is biproper. Also, right and left J-spectral factorizations of W , and right and left canonical factorization of W can easily be obtained from the corresponding factorizations of V . Thus the proposition above actually follows from Proposition 14.1. For the sake of completeness we shall give a direct proof. Proof. By assumption, W (λ) is invertible and selfadjoint for each λ ∈ T. Thus the number of eigenvalues of W (λ) in the open unit disc does not depend on the particular choice of λ ∈ T. In other words W (λ) has constant signature on T. Now let J be a signature matrix the signature of which is equal to this constant signature. Then for each λ ∈ T the matrix W (λ) is congruent to J. Let W (λ) = W− (λ)W+ (λ) be a right canonical factorization of W with respect to T. Consider ¯ −1 )∗ , !+ (λ) = W+ (λ W

¯ −1 )∗ . !− (λ) = W− (λ W

!+ (λ)W !− (λ) is again a right canonical factorization of W with reThen W (λ) = W !+ (λ)−1 W− (λ) is a constant matrix, F say. This shows spect to T. It follows that W −1 ∗ ¯ that W (λ) = W+ (λ ) F W+ (λ). Since W (λ) is selfadjoint for λ ∈ T, it follows that F is congruent to the signature matrix J introduced in the ﬁrst paragraph of the proof. Thus F = E ∗ JE for some invertible matrix E. Put L+ (λ) = EW+ (λ). ¯ −1 )∗ JL+ (λ) is a left J-spectral factorization of W with respect Then W (λ) = L+ (λ to the unit circle. The reverse implication is trivial. In what follows we assume that W is a biproper rational m × m matrix function which is selfadjoint on the unit circle and has no pole there. Such a function can be represented in the form W (λ) = D0 + C(λIn − A)−1 B + B ∗ (λ−1 In − A∗ )−1 C ∗ , where A has all its eigenvalues in the open unit disc. The fact that W is proper implies that W is analytic at zero. We shall assume additionally that A is invertible. Note that the invertibility of A follows from the analyticity at zero whenever the realization C(λ − A)−1 B is minimal. The invertibility assumption on A allows us to write W (λ) = D0 − B ∗ A−∗ C ∗ + C(λ − A)−1 B − B ∗ A−∗ (λ − A−∗ )−1 A−∗ C ∗ . Since W (∞) = D0 − B ∗ A−∗ C ∗ = W (0)∗ one has D0 − B ∗ A−∗ C ∗ = (D0 − CA−1 B)∗ .

264

Chapter 14. J-spectral factorization

Hence D0 is selfadjoint. We shall assume additionally that D0 = J0 for some signature matrix J0 . Thus W is of the form W (λ) = J0 − B ∗ A−∗ C ∗ + C(λ − A)−1 B − B ∗ A−∗ (λ − A−∗ )−1 A−∗ C ∗ . (14.30) We shall prove the following factorization result. Theorem 14.11. Let W be a biproper rational m × m matrix function given by (14.30), where J0 is a signature matrix and A is an invertible n × n matrix having all its eigenvalues in the open unit disc. In order that, for some signature matrix J the function W admits a left J-spectral factorization with respect to the unit circle, it is necessary and suﬃcient that there exists a Hermitian n × n matrix Y such that J0 + B ∗ Y B is invertible and Y is a solution of the equation Y = A∗ Y A − (C ∗ + A∗ Y B)(J0 + B ∗ Y B)−1 (C + B ∗ Y A)

(14.31)

with A − B(J0 + B ∗ Y B)−1 (C + B ∗ Y A) having all its eigenvalues in the open unit disc. In that case Y is unique and for J one can take any signature matrix J determined by (14.32) J0 + B ∗ Y B = E ∗ JE, where E is some invertible matrix. Furthermore, if Y is a Hermitian matrix with the properties mentioned above, then for a signature matrix J determined by the ¯ −1 )∗ JL− (λ) of W expression (14.32), a left J-spectral factorization W (λ) = L− (λ with respect to the unit circle is obtained by taking L− (λ) = E + E(J0 + B ∗ Y B)−1 (C + B ∗ Y A)(λIn − A)−1 B.

(14.33)

Equation (14.31) is a particular case of the so-called discrete algebraic Riccati equation. A solution Y of equation (14.31) is called T-stabilizing, or simply stabilizing when no confusion can arise, if J0 + B ∗ Y B is invertible and the matrix A − B(J0 + B ∗ Y B)−1 (C + B ∗ Y A) has all its eigenvalues in the open unit disc. In the above theorem, the existence of such a solution is required. Proof. We split the proof into six parts. Part 1. Since W is biproper and given by (14.30), we can write a realization for − A) −1 B, where D = W (∞) = J0 − B ∗ A−∗ C ∗ and W . In fact W (λ) = D + C(λ −∗ A 0 −A−∗ C ∗ = B ∗ A−∗ C . (14.34) , C A= , B= B 0 A Recall that the matrix A is invertible and has all its eigenvalues in the open unit disc D. Hence A−∗ has all its eigenvalues in Dext. This allows us to apply Theorem 12.5 with F− = D and F+ = Dext . It follows that W admits a left canonical factorization with respect to T if and only if the equation Y BD−1 B ∗ A−∗ Y − Y (A − BD−1 C) + (A−∗ + A−∗ C ∗ D−1 B ∗ A−∗ )Y + A−∗ C ∗ D−1 C = 0

(14.35)

14.5. J-spectral factorization with respect to other contours

265

has a unique solution Y satisfying the following additional spectral constraints: ⊂ Dext , σ A−∗ + (A−∗ C ∗ + Y B)D−1 B ∗ A−∗ (14.36) (14.37) σ A − BD−1 (B ∗ A−∗ Y + C) ⊂ D. Furthermore, if Y is such a solution of (14.35), then a left canonical factorization W (λ) = W1 (λ)W2 (λ) of W with respect to T is obtained by taking W1 (λ) = D − B ∗ A−∗ (λ − A−∗ )−1 (A−∗ C ∗ + Y B),

(14.38)

W2 (λ) = I + D−1 (B ∗ A−∗ Y + C)(λ − A)−1 B.

(14.39)

Let Y be the solution of (14.35) satisfying (14.36) and (14.37). We claim that J0 + B ∗ Y B is invertible. To prove this it will be convenient to rewrite W1 as a function of λ−1 . This can be done as follows: W1 (λ)

= D − B ∗ (λA∗ − I)−1 (A−∗ C ∗ + Y B) = D + B ∗ λ−1 (λ−1 − A∗ )−1 (A−∗ C ∗ + Y B) = D + B ∗ (λ−1 − A∗ + A∗ )(λ−1 − A∗ )−1 (A−∗ C ∗ + Y B) = D + B ∗ A−∗ C ∗ + B ∗ Y B + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B).

Recall that D = J0 − B ∗ A−∗ C ∗ . Thus W1 (λ) = J0 + B ∗ Y B + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B).

(14.40)

Since A is invertible, both W and W2 are analytic at zero. From the above formula for W1 we see that W1 is also analytic at zero. Hence W (0) = W1 (0)W2 (0). But W (0) is invertible. Thus W1 (0) = J0 + B ∗ Y B is invertible too. Part 2. In this part Y stands for a solution of (14.35) such that J0 + B ∗ Y B is invertible. We prove that in this case Y is also a solution of (14.31). Furthermore, we show that D −1 (C + B ∗ A−∗ Y ) = (J0 + B ∗ Y B)−1 (C + B ∗ Y A).

(14.41)

Multiplying (14.35) on the left by A∗ and regrouping terms one obtains A∗ Y A − Y − (A∗ Y B + C ∗ )D−1 (C + B ∗ A−∗ Y ) = 0.

(14.42)

So Y = A∗ Y A − (A∗ Y B + C ∗ )D −1 (C + B ∗ A−∗ Y ). Multiplying the latter identity on the left with B ∗ A−∗ and adding C to both sides gives C + B ∗ A−∗ Y = C + B ∗ Y A − (B ∗ Y B + B ∗ A−∗ C ∗ )D −1 (C + B ∗ A−∗ Y ).

266

Chapter 14. J-spectral factorization

It follows that

=

I + (B ∗ Y B + B ∗ A−∗ C ∗ )D −1 (C + B ∗ A−∗ Y ) D + B ∗ A−∗ C ∗ + B ∗ Y B D−1 (C + B ∗ A−∗ Y )

=

(J0 + B ∗ Y B)D −1 (C + B ∗ A−∗ Y ).

C + B∗Y A =

Since J0 + B ∗ Y B is invertible, we see that (14.41) holds. Using (14.41) in (14.42) gives that Y is a solution of (14.31). Part 3. In this part we show that Y ∗ is a solution of (14.35) whenever so is Y . For this purpose we consider the Hamiltonian T of (14.35), that is, ⎡ ⎤ −A−∗ − A−∗ C ∗ D−1 B ∗ A−∗ −A−∗ C ∗ D−1 C ⎦. T =⎣ −1 ∗ −∗ −1 BD B A −(A − BD C) − BD −1 C), where A, B and C are given by (14.34). Put Note that T = −(A ) ( 0 I . H= −I 0 !∗ and H = −H ∗ . Next we carry out the following =A −∗ H, H B =A −∗ C Then H A computation: −A−∗ C ∗ 0 A∗ ∗ −∗ −1 C D − CA B = D − B A B 0 A−1 ∗ −A−∗C ∗ −1 CA = D− B B =

D + B ∗ A−∗ C ∗ − CA−1 B = J0 − CA−1 B = D ∗ .

A −1 B = D∗ and we can apply item (iii) in Proposition 9.2 to show Thus D − C that T is invertible and HT = T ∗ H. Taking adjoints in (14.35) we obtain the equation (14.43) Y ∗ A−1 BD −∗ B ∗ Y ∗ + Y ∗ (A−1 + A−1 BD−∗ CA−1 ) ∗ ∗ −∗ ∗ ∗ ∗ −∗ −1 −(A − C D B )Y + C D CA = 0, where Y ∗ is the unknown. The Hamiltonian T∗ of this equation is given by ⎤ ⎡ ∗ A − C ∗ D−∗ B ∗ −C ∗ D−∗ CA−1 ⎦. T∗ = ⎣ −1 −∗ ∗ −1 −1 −∗ −1 A BD B A + A BD CA

14.5. J-spectral factorization with respect to other contours

267

It follows that T∗ = HT ∗ H. This together with the result of the previous paragraph shows that T∗ = T −1 . Now let Y be a solution of (14.35). It follows that Y ∗ is a solution of (14.43). Using the general theory of Riccati equations (see Section 12.1), this implies that the space ( ∗ ) Y N∗ = Im I is invariant under T∗ . But T∗ = T −1 . Thus the ﬁnite dimensional space N∗ is invariant under the Hamiltonian T of (14.35). But then (again see Section 12.1) we may conclude that Y ∗ is a solution of (14.35) too. Part 4. Let Y be a solution of (14.35) satisfying the additional spectral constraints (14.36) and (14.37). In this part we show that Y must be Hermitian. Now Y is uniquely determined by the given properties. Since, by the result of the previous part of the proof, Y ∗ a solution of (14.35), it thus suﬃces to show that the conditions (14.36) and (14.37) hold with Y ∗ in place of the matrix Y . From the ﬁrst part of the proof we know that J0 + B ∗ Y B is invertible. Hence the identity (14.41) holds. Using this identity, we can rewrite (14.37) as σ A − B(J0 + B ∗ Y B)−1 (C + B ∗ Y A) ⊂ D. Taking adjoints, we arrive at σ (A∗ − (A∗ Y ∗ B + C ∗ )(J0 + B ∗ Y B)−1 B ∗ ⊂ D. Next, note that ∗ −1 A − (A∗ Y ∗ B + C ∗ )(J0 + B ∗ Y B)−1 B ∗ −1 −∗ = I − (Y ∗ B + A−∗ C ∗ )(J0 + B ∗ Y B)−1 B ∗ A = I + (Y ∗ B + A−∗ C ∗ ) J0 + B ∗ Y B −1 ∗ −∗ B A −B ∗ (Y ∗ B + A−∗ C ∗ ) = A−∗ + (Y ∗ B + A−∗ C ∗ )D −1 B ∗ A−∗ . Here we used that D = J0 − B ∗ A−∗ C ∗ . We conclude that σ(A−∗ + (Y ∗ B + A−∗ C ∗ )D −1 B ∗ A−∗ ) ⊂ Dext, which is (14.36) with Y ∗ in place of Y . In Part 3 of the proof we saw that Y ∗ is a solution of (14.35). Furthermore, J0 + B ∗ Y ∗ B = (J0 + B ∗ Y ∗ B)∗ is invertible. Thus we know that (14.41) holds with Y ∗ in place of Y , that is,

268

Chapter 14. J-spectral factorization D −1 (C + B ∗ A−∗ Y ∗ ) = (J0 + B ∗ Y ∗ B)−1 (C + B ∗ Y ∗ A).

(14.44)

Using this we show that (14.37) holds with Y ∗ in place of Y . Indeed, taking adjoints in (14.36) we get σ(A−1 + A−1 BD−∗ (B ∗ Y ∗ + CA−1 ) ⊂ Dext. Now

A−1 + A−1 BD−∗ (B ∗ Y ∗ + CA−1

−1

−1 = I + BD−∗ (B ∗ Y ∗ + CA−1 ) A −1 A = I − B(D ∗ + B ∗ Y ∗ B + CA−1 B)−1 (B ∗ Y ∗ + CA−1 ) = A − B(J0 + B ∗ Y ∗ B)−1 (B ∗ Y ∗ A + C). Here we used that D∗ = J0 − CA−1 B. Now apply the identity (14.44). It follows that σ(A − BD −1 (C + B ∗ A−∗ Y ∗ )) ⊂ D, which is (14.37) with Y ∗ in place of Y . Part 5. Let Y be a Hermitian matrix such that J0 + B ∗ Y B is invertible and Y is a stabilizing solution of (14.31). In this part we show that in that case Y is a solution of (14.35) and that Y satisﬁes the spectral constraints (14.36) and (14.37). As a ﬁrst step let us prove that under the above conditions on Y again (14.41) holds. Indeed, multiplying (14.31) from the left by B ∗ A−∗ and adding C to both sides we get C + B ∗ A−∗ Y

=

=

C + B ∗ Y A − (B ∗ A−∗ C ∗ + B ∗ Y B) ·(J0 + B ∗ Y B)−1 (C + B ∗ Y A) (J0 + B ∗ Y B) − (B ∗ A−∗ C ∗ + B ∗ Y B) ·(J0 + B ∗ Y B)−1 (C + B ∗ Y A)

=

(J0 − B ∗ Y B)(J0 + B ∗ Y B)−1 (C + B ∗ Y A)

=

D(J0 + B ∗ Y B)−1 (C + B ∗ Y A).

Hence (14.41) holds indeed. Using this we can rewrite (14.31) as A∗ Y A − Y − (A∗ Y B + C ∗ )D−1 (C + B ∗ A−∗ Y ) = 0. Multiplying the latter on the left by A−∗ and regrouping terms we see that Y satisﬁes (14.35). Since Y is a stabilizing solution of (14.31) and (14.41) holds, the spectral constraint (14.37) is satisﬁed too. It remains to prove (14.36) To do this we ﬁrst

14.5. J-spectral factorization with respect to other contours

269

note that

A−1 + A−1 BD −∗ (B ∗ Y + CA−1

−1

−1 = I + BD−∗ (B ∗ Y + CA−1 ) A −1 A = I − B(D∗ + B ∗ Y ∗ B + CA−1 B)−1 (B ∗ Y ∗ + CA−1 ) = A − B(J0 + B ∗ Y ∗ B)−1 (B ∗ Y ∗ A + C) = A − BD−1 (C + B ∗ A−∗ Y ). Thus, since Y is Hermitian, we see that (14.36) follows from (14.37) by taking adjoints and an inverse. Because of the uniqueness of the solution Y in the ﬁrst part of the proof, the result of the present part also shows that the Hermitian stabilizing solution of (14.31), if it exists, is unique Part 6. In this ﬁnal part we complete the argument. Assume that for some J the function W admits a left J-spectral factorization with respect to the unit circle. Then by the ﬁrst part of the proof, equation (14.35) has a solution Y satisfying (14.36) and (14.37). Moreover for this Y we have that J0 + B ∗ Y B is invertible. Part 4 of the proof tells us that Y is Hermitian. From Part 2 we know that Y is a solution of (14.31) which, according to (14.37) and (14.41), is stabilizing. Conversely, if Y is a Hermitian matrix such that J0 + B ∗ Y B is invertible and Y is a stabilizing solution of (14.31), then Y is a solution of (14.35) and Y satisﬁes (14.36) and (14.37). Hence W admits a left canonical factorization with respect to the unit circle, and thus, by Proposition 14.10, also a left J-spectral factorization with respect to the unit circle. Finally, take a signature matrix J such that (14.32) holds. It remains to establish the formula for the left spectral factor L− . To do this we use the left canonical factorization W (λ) = W1 (λ)W2 (λ) obtained in Part 1. Combining (14.39) and (14.41) we get W2 (λ) = I + (J0 + B ∗ Y B)−1 (C + B ∗ Y A)(λ − A)−1 B. Thus, using the expression (14.40) for W1 (λ), ¯ −1 )∗ W2 (λ

=

I + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B)(J0 + B ∗ Y B)−1 J0 + B ∗ Y B + B ∗ (λ−1 − A∗ )−1 (C ∗ + A∗ Y B) (J0 + B ∗ Y B)−1

=

W1 (λ)(J0 + B ∗ Y B)−1 ,

=

¯ −1 )∗ (J0 +B ∗ Y B)W2 (λ). Now let J be a signature and it follows that W (λ) = W2 (λ ¯ −1 )∗ JL− (λ), with matrix such that (14.32) holds. Then we see that W (λ) = L− (λ L− given by (14.33), is a left J-spectral factorization with respect to the unit circle.

270

Chapter 14. J-spectral factorization

We now turn to a situation arising from linear-quadratic optimal control theory. It concerns the following version of the discrete algebraic Riccati equation X = A∗ XA + Q − A∗ XB(R + B ∗ XB)−1 B ∗ XA.

(14.45)

Here A, B, Q and R are given matrices of sizes n × n, n × m, n × n and m × m, respectively. We will consider the case when A has all its eigenvalues in the open unit circle, R and Q are Hermitian, and R is invertible. Of special interest are the stabilizing solutions of (14.45). A solution X of (14.45) is said to be T-stabilizing, or simply stabilizing when there is no danger of confusion, if R+B ∗ XB is invertible and A − B(R + B ∗ XB)−1 B ∗ XA has all its eigenvalues in the open unit disc. In connection with (14.45) we consider the rational matrix function W (λ) = R + B ∗ (λ−1 In − A∗ )−1 Q(λIn − A)−1 B.

(14.46)

Note that this function is Hermitian on the unit circle. Proposition 14.12. Let A, B, Q and R be as above, so A is an n × n matrix having its eigenvalues in the open unit disc, B is an m × m matrix, R is an invertible Hermitian m × m matrix, and Q is a Hermitian n × n matrix. Assume in addition that A is invertible. The following two statements are equivalent: (i) The Riccati equation (14.45) has a (unique) Hermitian T-stabilizing solution; (ii) For some Hermitian matrix J, the rational matrix function (14.46) admits a left J-spectral factorization with respect to the unit circle. In that case J is congruent to R+B ∗ XB. Also, if X is the Hermitian T-stabilizing solution of (14.45), then ¯ −1 )∗ (R + B ∗ XB)L− (λ), W (λ) = L− (λ with

L− (λ) = Im + (R + B ∗ XB)−1 B ∗ XA(λIn − A)−1 B,

is a left (R + B ∗ XB)-spectral factorization with respect to the unit disc. The function L− is the unique left (R + B ∗ XB)-spectral factor with L− (∞) = Im . The additional assumption that A is invertible plays an essential role in the proof as we give it below. Indeed, the argument involves a reduction to earlier results, in particular to Theorem 14.11. However, instead of Theorem 14.11 one can employ Theorem 14.15 below which does not feature the hypothesis that A is invertible. Before we prove the proposition, let us remark that in the case of the linear quadratic optimal control problem of mathematical systems theory, one has that R is positive deﬁnite and Q is positive semideﬁnite. Hence the function (14.46) is positive deﬁnite on the unit circle, and as A has is eigenvalues in the open unit disc, it has no poles on the unit circle. Thus, in that case, the function does admit

14.5. J-spectral factorization with respect to other contours

271

a right spectral factorization with J = I, and hence there is a stabilizing solution X to the discrete algebraic Riccati equation. In addition, for that solution the matrix R + B ∗ XB is positive deﬁnite. Proof. We shall deduce Proposition 14.12 from Theorem 14.11. First, a realization for (14.46) is given as $ A W (λ) = R + −B ∗ A−∗ Q B ∗ A−∗ λ − −A−∗ Q

0

%−1 B

A−∗

0

.

Since A is has all its eigenvalues in the open unit disc, there is a unique solution to the equation X0 − A∗ X0 A = Q. (14.47) Taking as a similarity transformation the matrix I 0 , X0 I and using Q − X0 = −A∗ X0 A, the realization above can be rewritten as: W (λ)

= R + −B ∗ A−∗ (Q − X0 ) B ∗ A

−∗

$ λ−

A

0

0

A−∗

%−1

B

X0 B

= R + B ∗ X0 A(λ − A)−1 + B ∗ A−∗ (λ − A−∗ )−1 X0 B. The latter expression is of the form (14.30), with C = B ∗ X0 A and with J0 = R + B ∗ X0 B. So, we can apply Theorem 14.11, with (14.31) suitably modiﬁed, to conclude that W admits a left J-spectral factorization if and only if there is a solution Y , satisfying additional constraints, of the equation Y = A∗ Y A − (A∗ X0 B + A∗ Y B)(R + B ∗ X0 B + B ∗ Y B)−1 (B ∗ Y A + B ∗ X0 A). Putting X = X0 + Y and taking into account (14.47), we see that the above equation becomes (14.45) for X. The additional constraints referred to above are: in the ﬁrst place, invertibility of R + B ∗ X0 B + B ∗ Y B = R + B ∗ XB, which we also required for the solution of (14.45), and, secondly, the condition that the eigenvalues of A − B(R + B ∗ (X0 + Y )B)−1 B ∗ (X0 + Y )A = A − B(R + B ∗ XB)−1 B ∗ XA are in the open unit disc. But this is exactly what is required for the stabilizing solution of the equation (14.45). The expressions for the factorization also follow directly from the formulas in Theorem 14.11.

272

Chapter 14. J-spectral factorization

We conclude this section by considering J-spectral factorization of a selfadjoint function on the real line. As before J is an invertible Hermitian m × m matrix. Suppose W is a rational m × m matrix function. A factorization ¯ ∗ JL(λ) W (λ) = L(λ)

(14.48)

is called a right J-spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed upper half plane ¯ ∗ and its inverse are analytic on (inﬁnity included). In that case the function L(λ) the closed lower half plane (inﬁnity included). Thus a right J-spectral factorization with respect to the real line is a right canonical factorization with respect to R featuring an additional symmetry property between the factors. A factorization (14.48) is called a left J-spectral factorization with respect to the real line if L and L−1 are rational m × m matrix functions which are analytic on the closed lower ¯ ∗ and its inverse are half plane (inﬁnity included), in which case the function L(λ) analytic on the closed upper half plane (inﬁnity included). Such a factorization is a left canonical factorization with respect to R. Results for this type of factorization can be derived in a straightforward manner from J-spectral factorization theorems with respect to the imaginary axis. Indeed, if W is selfadjoint on the real line, then V given by V (λ) = W (−iλ) is self¯ adjoint on the imaginary axis. Also W (λ) = L+ (λ)JL + (λ) is a right J-spectral fac¯ torization of W with respect to the real line if and only V (λ) = K+(−λ)JK + (λ), with K+ (λ) = L+ (−iλ), is a right J-spectral factorization of V with respect to the imaginary axis. As an illustration we show how one can derive the following result as a corollary from Theorem 14.9. Theorem 14.13. Let the rational m × m matrix function W be given by W (λ) = J + C(λIn − A)−1 B + B ∗ (λIn − A∗ )−1 C ∗ , where J is an m × m signature matrix and A is an n × n matrix having all its eigenvalues in the open upper half plane. Then W admits a right J-spectral factorization with respect to the real line, ¯ ∗ JL+ (λ), W (λ) = L+ (λ) if and only if the algebraic Riccati equation Y C ∗ JCY − Y (A∗ − C ∗ JB ∗ ) + (A − BJC)Y − BJB ∗ = 0

(14.49)

has a (unique) skew-Hermitian solution Y such that A∗ − C ∗ JB ∗ − C ∗ JCY has all its eigenvalues in the open lower half plane. In that case, the unique right J-spectral factor L+ for which L+ (∞) = Im is given by L+ (λ) = Im + J(CY + B ∗ )(λIn − A∗ )−1 C ∗ .

14.6. Left versus right J-spectral factorization

273

A solution Y of the Riccati equation (14.49) is called R-stabilizing, or simply stabilizing when confusion is not possible, if A∗ − C ∗ JB ∗ − C ∗ JCY has all its eigenvalues in the open lower half plane. In the above theorem, the existence of such a solution is required. Proof. Write V (λ) = W (−iλ). Then V (λ)

= =

J + C(−iλ − A)−1 B + B ∗ (−iλ − A∗ )−1 C −1 −1 J + (iC) λ − (iA) B + B ∗ λ + (iA)∗ (iC).

Notice that iA has all its eigenvalues in the open left half plane. By Theorem 14.9 the function V admits a right J-spectral factorization with respect to the imaginary axis if and only if the equation (14.50) X(iC)∗ J(iC)X + X (iA)∗ − (iC)∗ JB ∗ ∗ + iA − BJ(iC) X + BJB = 0 has a Hermitian solution X such that the matrix (iA)∗ −(iC)∗ JB ∗ +(iC)∗ J(iC)X has all its eigenvalues in the open left half plane. In that case, a right J-spectral ¯ ∗ JK+ (λ) of V with respect to the imaginary axis is factorization V (λ) = K+ (−λ) −1 (iC)∗ . Next we replace obtained by taking K+ (λ) = I +J(iCX −B ∗ ) λ+(iA)∗ X by iY and multiply equation (14.50) by −1. In this way (14.50) is shown to be equivalent to (14.49). Furthermore Y isskew-Hermitian if and only if X is Hermitian, and A∗ − C ∗ JB ∗ − C ∗ JCY = i (iA)∗ − (iC)∗ JB ∗ + (iC)∗ J(iC)X . Finally, put L+ (λ) = K+ (iλ). Then L+ (λ)

−1 = I + J(iCX − B ∗ ) λ + (iA)∗ (iC)∗ = I + J(−CY − B ∗ )(iλ − iA∗ )−1 (−i)C ∗ = I + J(CY + B ∗ )(λ − A∗ )−1 C ∗ .

Using these formulas it is now straightforward to complete the argument.

14.6 Left versus right J-spectral factorization The existence of a left canonical factorization does not always imply the existence of a right canonical factorization. The same is true for J-spectral factorization. In this section we answer the following question: if a rational matrix function W admits a left J-spectral factorization, under what conditions does it also have a right J-spectral factorization? And, if so, how can the right factorization be obtained from the left one? The main result can be viewed as a symmetric version of Theorem 12.6. We restrict our attention to factorization with respect to the imaginary axis.

274

Chapter 14. J-spectral factorization

For later purposes it will be convenient to only assume that J is an invertible Hermitian matrix. We do not stipulate it to be a signature matrix here. Theorem 14.14. Let J be an invertible Hermitian m × m matrix, and let W be a rational m × m matrix function. Suppose ¯ ∗ JL− (λ) W (λ) = L− (−λ) is a left J-spectral factorization with respect to the imaginary axis, and L− admits the realization L− (λ) = Im + C(λIn − A)−1 B (14.51) with A and A× = A − BC having their eigenvalues in the open left half plane. Let Q and P be the unique (Hermitian) solutions of the Lyapunov equations QA + A∗ Q = A× P + P (A× )∗

=

C ∗ JC.

(14.52)

−BJ −1 B ∗ .

(14.53)

Then W admits a right J-spectral factorization with respect to the imaginary axis if and only if I − QP is invertible, or, which amounts to the same, I − P Q is invertible. In that case, a right J-spectral factorization of W with respect to the imaginary axis is given by ¯ ∗ JL+ (λ), W (λ) = L+ (−λ)

(14.54)

where L+ (λ) and its inverse are given by (14.55) L+ (λ) = Im + (CP − J −1 B ∗ )(I − QP )−1 ∗ −1 ∗ ·(λIn + A ) (C J − QB), −1 −1 ∗ B ) λIn + (A× )∗ L−1 + (λ) = Im − (CP − J ·(I − QP )

−1

(14.56) ∗

(C J − QB).

Proof. We bring ourselves in the situation of Section 12.4 by introducing ¯ ∗ = Im − B ∗ (λIn + A∗ )−1 C ∗ , Y+ (λ) = L− (−λ) Y− (λ) = JL− (λ) = J + JC(λIn − A)−1 B. Then W (λ) = Y+ (λ)Y− (λ) is a left canonical factorization, here taken with respect to the imaginary axis (cf., the remark made after the proof of Theorem 12.6). In terms of the notation employed in Section 12.4, Y+ (λ)

= D+ + C+ (λ − A+ )−1 B+ ,

Y− (λ)

= D− + C− (λ − A− )−1 B− ,

14.6. Left versus right J-spectral factorization with D+ = Im , D− = J,

A+ = −A∗ , A− = A,

275

B+ = C ∗ , B− = B,

C+ = −B ∗ , C− = JC.

× × ∗ × For the associate main matrices we have A× + = −(A ) and A− = A . Thus the Lyapunov equations (12.19) reduce to the equations (14.53) and (14.52). Application of Theorem 12.6 now shows that W admits a right canonical factorization with respect to the imaginary axis if and only if I − QP is invertible, or, which amounts to the same, I − P Q is invertible. Assume this is the case. Then, again by virtue of Theorem 12.6, we have the right canonical factorization W (λ) = W− (λ)W+ (λ), where

W− (λ)

=

D+ + (D+ C− + C+ Q)(λIX− − A− )−1 −1 · (IX− − P Q)−1 (B− D− − P B+ ),

W+ (λ)

=

−1 D− + (D+ C+ + C− P )(IX+ − QP )−1

· (λIX+ − A+ )−1 (B+ D− − QB− ). Making the appropriate substitutions, we get W− (λ)

=

I + (JC − B ∗ Q)(λ − A)−1 (I − P Q)−1 (BJ −1 − P C ∗ ),

W+ (λ)

=

J + (JCP − B ∗ )(I − QP )−1(λ + A∗ )−1 (C ∗ J − QB).

Put L+ (λ) = J −1 W+ (λ). Then L+(λ) is given by (14.55). Taking into account the ¯ ∗ is precisely W− (λ). It follows selfadjointness of Q and P , one sees that L+ (−λ) ∗ ¯ that W (λ) = L+ (−λ) JL+(λ), and this factorization is a right J-spectral factor−1 ization of W with respect to the imaginary axis. Finally, L−1 + (λ) = W+ (λ)J, and according to Theorem 12.6, W+−1 (λ)

=

−1 −1 −1 −1 D− − D− (D+ C+ + C− P )(λIX+ − A× +) −1 · (IX+ − QP )−1 (B+ D− − QB− )D− .

Via the appropriate substitutions this becomes −1 W+−1 (λ) = J −1 − (CP − J −1 B ∗ ) λ + (A× )∗ (I − QP )−1 (C ∗ J − QB)J −1 . Multiplying the latter identity from the right by J gives (14.56).

For the case when J is a signature matrix (that is, J = J ∗ = J −1 ) it is also possible to derive the previous result from Theorem 14.9. Indeed, let Q be the solution of (14.52), and introduce ) ( I 0 . T = Q I

276

Chapter 14. J-spectral factorization

Then one has (via the product rule for realizations) ¯ ∗ JL− (λ) W (λ) = L− (−λ) =J+

JC

−B ∗

$ T

λ−T

−1

A

0

C ∗ JC

−A∗

%−1 T

T

−1

B

C∗J

= J + (JC − B ∗ Q)(λ − A)−1 B − B ∗ (λ + A∗ )−1 (C ∗ J − QB). Clearly, one can now apply Theorem 14.9. The stabilizing solution of equation (14.28), taken for this particular situation, and the solution P of (14.53) are related as follows: if Y is the stabilizing solution, then I +QY is invertible, the matrix P = Y (I + QY )−1 solves (14.53), and I − QP = (I + QY )−1 is invertible. Conversely, if P is the solution of (14.53) and I − QP is invertible, then Y = P (I − QP )−1 is Hermitian and it is the desired stabilizing solution. Finally, for the case where J = I, and so W is positive deﬁnite on the imaginary line, the condition that I − QP is invertible should be automatically fulﬁlled on account of Theorem 9.4. That this is indeed the case can be seen as follows. First recall that A has all its eigenvalues in the open left half plane. This implies that P is positive semideﬁnite and Q is negative semideﬁnite. Since J = I we get from (14.53) that Ker P is invariant under A∗ . Now write P , Q, A and C ˙ Im P as with respect to the decomposition Cn = Ker P + Q11 Q12 A11 0 0 0 , Q= , A= , C = C1 C2 . P = Q21 Q22 A21 A22 0 P22 Then Q22 is negative semideﬁnite and P22 is positive deﬁnite. Finally, I − QP is invertible if and only if I − Q22 P22 is invertible as a map from Im P to itself. Since 1/2 1/2 I − Q22 P22 is similar to I − P22 Q22 P22 , and the latter is positive deﬁnite, we see that invertibility of I − QP is indeed automatically satisﬁed.

14.7 J-spectral factorization relative to the unit circle revisited In this section we present a somewhat more general form of Theorem 14.11, using an alternative approach. As in the ﬁrst part of Section 14.5, the function W is a rational m × m matrix function which is selfadjoint on the unit circle and has no pole there. Such a function can be represented in the form W (λ) = D0 + C(λIn − A)−1 B + B ∗ (λ−1 In − A∗ )−1 C ∗ ,

(14.57)

where D0 is a Hermitian m × m matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. In contrast to the situation considered in

14.7. J-spectral factorization relative to the unit circle revisited

277

Section 14.5 we do not assume that A is invertible, and hence the representation (14.30) is not available in the present context. Similar to what was done in Theorem 14.11, we associate with the representation (14.57) the Riccati equation Y = A∗ Y A − (C ∗ + A∗ Y B)(D0 + B ∗ Y B)−1 (C + B ∗ Y A).

(14.58)

Recall from the paragraph directly following Theorem 14.11 that a solution Y to this Riccati equation is called T-stabilizing (or simply stabilizing) if D0 + B ∗ Y B is invertible and the matrix A − B(D0 + B ∗ Y B)−1 (C + B ∗ Y A)

(14.59)

has all its eigenvalues in the open unit disc. The following theorem is the main result of this section. Theorem 14.15. Let W be a rational m×m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n×n matrix having all its eigenvalues in the open unit disc. In order that, for some signature matrix J the function W admits a left J-spectral factorization with respect to the unit circle, it is necessary and suﬃcient that the Riccati equation (14.58) has a Hermitian T-stabilizing solution Y . In that case Y is unique, and for J one can take any signature matrix J determined by D0 + B ∗ Y B = E ∗ JE, (14.60) where E is some invertible matrix. Furthermore, if Y is the Hermitian T-stabilizing solution to (14.58), then for a signature matrix J determined by (14.60), a left J¯ −1 )∗ JL− (λ) of W with respect to the unit spectral factorization W (λ) = L− (λ circle can be obtained by taking L− (λ) = E + E(D0 + B ∗ Y B)−1 (C + B ∗ Y A)(λIn − A)−1 B.

(14.61)

To prove the above theorem we cannot use the method employed in Section 14.5. Instead we shall use the connection between canonical factorization and invertibility of Toeplitz operators described in Section 1.2. For this purpose we need the block Toeplitz operator T on m 2 deﬁned by the rational m × m mam trix function W (λ−1 ). Recall (see Section 1.2) that m 2 = 2 (C ) stands for the Hilbert space of all square summable sequences (x0 , x1 , x2 , . . .) with entries in Cm . Furthermore, by deﬁnition, T is the operator on m 2 given by the block matrix representation ⎡ ⎤ R0 R−1 R−2 · · · ⎢ ⎥ ⎢R1 R0 R−1 · · ·⎥ ⎢ ⎥ (14.62) T =⎢ ⎥, ⎢R2 R1 R0 · · ·⎥ ⎣ ⎦ .. .. .. .. . . . .

278

Chapter 14. J-spectral factorization

where . . . , R−1 , R0 , R1 , . . . are the coeﬃcients in the Laurent expansion W (λ−1 ) =

∞

λj Rj

j=−∞

of the function W (λ−1 ) on the unit circle. When W is given by (14.57), we have R0 = D0 ,

∗ Rj = R−j = CAj−1 B,

j = 1, 2, . . . .

(14.63)

The following lemma provides one of the main steps in the proof of Theorem 14.15. As always in this section, J stands for a signature matrix. Lemma 14.16. Let W be a rational m × m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. Assume W admits a left J-spectral factorization with respect to the unit circle. Then the block Toeplitz operator T on m 2 deﬁned by the rational m × m matrix function W (λ−1 ) is invertible, and the n × n matrix Y given by ⎡ ⎤ C ⎢ ⎥ ⎢ CA ⎥ ∗ ⎢ ⎥ A∗ C ∗ A∗ 2 C ∗ · · · T −1 ⎢ Y =− C (14.64) ⎥ ⎢CA2 ⎥ ⎣ ⎦ .. . is a Hermitian stabilizing solution to the Riccati equation (14.58). Proof. A left J-spectral factorization with respect to the unit circle is, in particular, a left canonical factorization with respect to the unit circle. But then the function W (λ−1 ) admits a right canonical factorization with respect to the unit circle, and Theorem 1.2 tells us that the block Toeplitz operator T is invertible. This, together with the fact that A has all its eigenvalues in the open unit disc, gives that the matrix Y is well-deﬁned by (14.64). Note that T is selfadjoint because W (λ−1 ) has Hermitian values on the unit circle. But then T −1 is selfadjoint too, and (14.64) shows that Y is Hermitian m m Note that m 2 can be identiﬁed with the Hilbert space direct sum C ⊕ 2 . Via this identiﬁcation the operator T partitions as ⎡ ⎤ R1 ⎢ ⎥ ⎢ R2 ⎥ R0 Λ∗ ⎢ ⎥ , where Λ = ⎢ ⎥ : Cm → m T = (14.65) 2 . ⎢ R3 ⎥ Λ T ⎣ ⎦ .. . Put Δ = R0 − Λ∗ T −1 Λ. Since the 2 × 2 operator matrix in (14.65) and the operator in its right lower corner are both invertible, a standard Schur complement

14.7. J-spectral factorization relative to the unit circle revisited

279

argument (see [19] or the second proof of Theorem 2.1 in [20]) tells us that Δ is invertible as well. Furthermore, relative to the Hilbert space direct sum decomposition Cm ⊕ m 2 the inverse of T admits the block matrix representation ⎡ ⎤ −Δ−1 Λ∗ T −1 Δ−1 ⎦. (14.66) T −1 = ⎣ −1 −1 −1 −1 −1 ∗ −1 −T ΛΔ T + T ΛΔ Λ T Recall from (14.63) that R0 = D0 . Combining the second part of (14.63) with (14.64) we obtain that B ∗ Y B = −Λ∗ T −1Λ. It follows that D0 + B ∗ Y B = Δ, and hence D0 + B ∗ Y B is invertible, as desired. To prove that Y satisﬁes the Riccati equation (14.58) we ﬁrst consider the operator T −1 − ST −1S ∗ , where S is the (block) forward shift on m 2 . Thus the actions of S and S ∗ on m are given by 2 S ∗ (x0 , x1 , x2 , . . .) = (x1 , x2 , x3 , . . .).

S(x0 , x1 , x2 , . . .) = (0, x0 , x1 , . . .),

A straightforward computation shows that the partitioning of ST −1 S ∗ relative to the Hilbert space direct sum Cm ⊕ m 2 is given by 0 0 −1 ∗ ST S = . 0 T −1 This identity, together with the identity (14.66), yields −Δ−1 Λ∗ T −1 Δ−1 −1 −1 ∗ = T − ST S −T −1 ΛΔ−1 T −1ΛΔ−1 Λ∗ T −1 =

I −T

−1

Λ

Δ−1 I

−Λ∗ T −1 .

Next, let Γ be the operator from Cn to m 2 given by ⎡ ⎤ C ⎢ ⎥ ⎢ CA ⎥ ⎢ ⎥ Γ=⎢ ⎥. ⎢CA2 ⎥ ⎣ ⎦ .. .

(14.67)

(14.68)

Note that this operator Γ is well-deﬁned because the matrix A has all its eigenvalues in the open unit disc. As is easily checked ΓA = S ∗ Γ,

ΓB = Λ,

Y = −Γ∗ T −1Γ.

(14.69)

280

Chapter 14. J-spectral factorization

From these identities and (14.67) it follows that Y − A∗ Y A =

−Γ∗ T −1Γ + A∗ Γ∗ T −1 ΓA

=

−Γ∗ T −1Γ + Γ∗ ST −1 S ∗ Γ

=

−Γ∗ (T −1 − ST −1S ∗ )Γ I ∗ −Γ Δ−1 I −Λ∗ T −1 Γ. −1 −T Λ

= Furthermore

[ I − Λ∗ T −1 ]Γ = C − Λ∗ T −1 S ∗ Γ = C − B ∗ Γ∗ T −1 Γ = C + B ∗ Y A.

(14.70)

Summarizing (and using that Y is Hermitian) we have Y − A∗ Y A = −(C + B ∗ Y A)∗ Δ−1 (C + B ∗ Y A). Since Δ = D0 + B ∗ Y B, this identity shows that Y satisﬁes the Riccati equation (14.58). Write A× for the matrix (14.59). We need to show that for Y given by (14.64), all eigenvalues of A× are in the open unit disc. Using (14.67), the fact that S ∗ S is the identity operator on m 2 , and the identities in (14.69) and (14.70), we see that I Δ−1 I −Λ∗ T −1 Γ S ∗ T −1 Γ = T −1 S ∗ Γ + S ∗ −1 −T Λ = T −1 ΓA − T −1 ΛΔ−1 (C + B ∗ Y A) = T −1 Γ A − BΔ−1 (C + B ∗ Y A) = T −1ΓA× . Thus S ∗ T −1 Γ = T −1 ΓA× . It follows that (S ∗ )k T −1 Γ = T −1 Γ(A× )k ,

k = 1, 2, . . . .

But then the fact that S ∗n converges to zero in the strong operator topology yields lim T −1 Γ(A× )k x = lim (S ∗ )k T −1Γx = 0,

k→∞

k→∞

x ∈ Cn .

(14.71)

We shall use (14.71) to prove that A× has all its eigenvalues in the open unit disc. To do this we ﬁrst decompose Cn as Cn = X1 ⊕ X2 , where X2 = Ker Γ and X1 = (Ker Γ)⊥ . Notice that X2 is an invariant subspace for A, and C[X2 ] = {0}. We also have Y A[X2 ] = {0}. Indeed Y [AX2 ] ⊂ Y [X2 ] = −Γ∗ T −1 Γ[X2 ] = {0}.

14.7. J-spectral factorization relative to the unit circle revisited

281

Using C[X2 ] = {0} and Y A[X2 ] = {0} in (14.59), we see that A× |X2 = A|X2 , and X2 is an invariant subspace for A× too. In other words, A× admits a matrix representation of the form × 0 A11 × : X1 ⊕ X2 → X1 ⊕ X2 , A = (14.72) × A× 21 A22 where A× 22 = A|X2 : X2 → X2 . Since X2 is an invariant subspace for A and A has all its eigenvalues in the open unit disc, A22 has all its eigenvalues in the open unit disc too. Hence, in order to prove that A× has all its eigenvalues in the open unit disc, it now suﬃces to prove that A× 11 has this property. Let τ1 be the canonical embedding of X1 into Cn = X1 ⊕ X2 , and let Γ1 be the one-to-one operator from X1 into m 2 deﬁned by Γ1 = Γτ1 . Take x ∈ X1 . Since Γ is equal to k zero on X2 , we see from (14.72) that T −1 Γ(A× )k x = T −1 Γ1 (A× 11 ) x. But then × k −1 −1 (14.71) tells us that limk→∞ T Γ1 (A11 ) x = 0. Observe that T Γ1 is one-toone and has a closed (ﬁnite dimensional) range, that is, T −1 Γ1 is left invertible. × k k Hence limk→∞ T −1 Γ1 (A× 11 ) x = 0 implies that limk→∞ (A11 ) x = 0. Since x is an arbitrary element of X1 , the latter holds if and only if the eigenvalues of A× 11 are in the open unit disc. Lemma 14.16 proves the necessity part of Theorem 14.15. The suﬃciency part, the formula for the J-spectral factorization, and the uniqueness statement are covered by the next two lemmas. Lemma 14.17. Let W be a rational m × m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. Assume Y is a Hermitian stabilizing solution of the Riccati equation (14.58). Then W admits a left J-spectral factorization with respect to the unit circle. Such a factorization can be obtained as follows. Choose an m × m signature matrix J such that D0 + B ∗ Y B = E ∗ JE, where E is some invertible matrix, and deﬁne L− by (14.61), i.e., L− (λ) = E + E(D0 + B ∗ Y B)−1 (C + B ∗ Y A)(λIn − A)−1 B. ¯ −1 )∗ JL− (λ) is a left J-spectral factorization of W with respect Then W (λ) = L− (λ to the unit circle. Proof. Put Δ = D0 + B ∗ Y B, C0 = C + B ∗ Y A, and set Ψ(λ) = Δ + C0 (λ − A)−1 B.

(14.73)

Note that A − BΔ−1 C0 is equal to the matrix A× deﬁned by (14.59). Thus Ψ(λ)−1 = Δ−1 − Δ−1 C0 (λ − A× )−1 BΔ−1 .

(14.74)

The fact that A and A× have all their eigenvalues in the open unit disc implies that Ψ(λ) and Ψ(λ)−1 are both analytic on the closure of the exterior of the unit

282

Chapter 14. J-spectral factorization

disc, inﬁnity included. Since L− (λ) = EΔ−1 Ψ(λ), the same holds true for L− (λ) ¯ −1 )∗ JL− (λ) is a left spectral factorization with and L− (λ)−1 . It follows that L− (λ respect to the unit circle. It remains to show that ¯ −1 )∗ JL− (λ). W (λ) = L− (λ

(14.75)

From L− (λ) = EΔ−1 Ψ(λ) and Δ = D0 + B ∗ Y B = E ∗ JE we see that ¯ −1 )∗ JL− (λ) = Ψ(λ ¯ −1 )∗ Δ−1 Ψ(λ). L− (λ Using the deﬁnitions of Δ and C0 , the Riccati equation (14.58) can be rewritten as Y − A∗ Y A = −C0∗ Δ−1 C0 . It follows that λC0∗ Δ−1 C0 = −Y (λ − A) + (I − λA∗ )Y (λ − A) − λ(I − λA∗ )Y. Using this identity we obtain B ∗ (I − λA∗ )−1 (λC0∗ Δ−1 C0 )(λ − A)−1 B = −B ∗ (I − λA∗ )−1 Y B + B ∗ Y B − λB ∗ Y (λ − A)−1 B = −λB ∗ (I − λA∗ )−1 A∗ Y B − B ∗ Y B − B ∗ Y A(λ − A)−1 B. Hence ¯ −1 )∗ Δ−1 Ψ(λ) Ψ(λ = Δ + λB ∗ (I − λA∗ )−1 C0∗ Δ−1 Δ + C0 (λ − A)−1 B = Δ + λB ∗ (I − λA∗ )−1 C0∗ + C0 (λ − A)−1 B +B ∗ (I − λA∗ )−1 (λC0∗ Δ−1 C0 )(λ − A)−1 B. From the deﬁnitions of Δ and C0 given in the beginning of the proof we see that Δ − B ∗ Y B = D0 and C0 − B ∗ Y A = C. Thus the calculations above yield ¯ −1 )∗ Δ−1 Ψ(λ) = D0 + λC(I − λA)−1 + B ∗ (λ − A∗ )−1 C ∗ . Ψ(λ According to (14.57) the right-hand side in the previous identity is equal to W (λ). ¯ −1 )∗ Δ−1 Ψ(λ) = W (λ), as desired. Thus Ψ(λ Lemma 14.18. Let W be a rational m × m matrix function given by (14.57), where D0 is a Hermitian matrix and A is an n × n matrix having all its eigenvalues in the open unit disc. Assume Y is a Hermitian stabilizing solution of the Riccati equation (14.58). Then the block Toeplitz operator T on m 2 deﬁned by the rational m × m matrix function W (λ−1 ) is invertible and Y is uniquely determined by the expression (14.64).

14.7. J-spectral factorization relative to the unit circle revisited

283

Proof. As in the proof of the preceding lemma, we set Δ = D0 + B ∗ Y B and C0 = C + B ∗ Y A. Furthermore, Ψ(λ) is the rational m× m matrix function deﬁned by (14.73). Put Θ(λ) = Ψ(λ−1 ). The proof of the preceding lemma tells us that ¯ −1 )∗ Δ−1 Θ(λ). W (λ−1 ) = Θ(λ −1 Hence the block Toeplitz operator T on m ) admits the factor2 deﬁned by W (λ ∗ ization T = (TΘ ) Ξ TΘ , where TΘ is the block Toeplitz operator on m 2 deﬁned by Θ, and Ξ is the block diagonal operator on m 2 given by

Ξ = diag (Δ−1 , Δ−1 , Δ−1 , . . .). From (14.73), (14.74) and Θ(λ) = Ψ(λ−1 ) we know that Θ(λ)

=

Δ + λC0 (I − λA)−1 B,

(14.76)

Θ(λ)−1

=

Δ−1 − λΔ−1 C0 (I − λA× )−1 BΔ−1 ,

(14.77)

where A× is given by (14.59). From (14.76), (14.77), and the fact that both A and A× have all their eigenvalues in the open unit disc we see that TΘ is invertible and TΘ−1 is given by ⎤ ⎡ × Θ0 0 0 ··· ⎥ ⎢ × ⎢ Θ1 Θ × 0 · · ·⎥ 0 ⎥ ⎢ (14.78) TΘ−1 = ⎢ × ⎥, × ⎥ ⎢Θ2 Θ× Θ0 1 ⎦ ⎣ .. .. .. . . . × × −1 at zero. Furthermore, where Θ× 0 , Θ1 , Θ2 , . . . are the Taylor coeﬃcients of Θ(λ) (14.76) yields −1 Θ× , 0 = Δ

−1 Θ× C0 (A× )j−1 BΔ−1 , j = −Δ

j = 1, 2, . . . .

(14.79)

Let Γ be the operator from Cn into m 2 deﬁned by (14.68). Using the identities in (14.78) and (14.79) we compute that . (14.80) Γ∗ TΘ−1 = β A∗ β (A∗ )2 β · · · , with β given by β = C ∗ Δ−1 − A∗

∞ ∗ j ∗ −1 × j (A ) C Δ C0 (A ) BΔ−1 .

(14.81)

j=0

As T = (TΘ )∗ ΞTΘ and TΘ is invertible, we conclude that T is invertible. Moreover, using (14.80), we have Γ∗ T −1Γ = (Γ∗ TΘ−1 ) Ξ−1 (Γ∗ TΘ−1)∗ =

∞

j=0

β∗ Aj . (A∗ )j βΔ

(14.82)

284

Chapter 14. J-spectral factorization

We proceed by showing that β = (C ∗ + A∗ Y B)Δ−1 , where A× is given by (14.59). To prove this we use the fact that Y satisﬁes the Riccati equation (14.58). A straightforward computation gives Y

A∗ Y A − (C ∗ + A∗ Y B)Δ−1 (C + B ∗ Y A) = A∗ Y A − BΔ−1 (C + B ∗ Y A) − C ∗ Δ−1 (C + B ∗ Y A)

=

=

A∗ Y A× − C ∗ Δ−1 C0 .

We conclude that Y − A∗ Y A× = −C ∗ Δ−1 C0 . Since both A and A× have all their eigenvalues in the open unit disc, we obtain Y =−

∞

(A∗ )j C ∗ Δ−1 C0 (A× )j .

j=0

Using the latter identity in (14.81) we arrive at β = C ∗ Δ−1 + A∗ Y BΔ−1 = (C ∗ + A∗ Y B)Δ−1 . Finally, the identity β = (C ∗ + A∗ Y B)Δ−1 and the fact that Y satisﬁes the Riccati equation yield β∗ . Y − A∗ Y A = −(C ∗ + A∗ Y B)Δ−1 (C + B ∗ Y A) = −βΔ

(14.83)

∞ β∗ Aj because A has all its eigenvalues in the But then Y = − j=0 (A∗ )j βΔ open unit disc. Comparing the latter expression for Y with (14.82) we see that Y = −Γ∗ T −1 Γ. Thus Y is given by (14.64), as desired. In Theorem 14.15 we restricted the attention to stabilizing solutions of the Riccati equation (14.58) that are required to be Hermitian. This requirement is not essential: Theorem 14.15 remains true if Y is just an arbitrary stabilizing solution of (14.58). The reason is that a stabilizing solution of (14.58) is always Hermitian. This result is the contents of the following proposition. Proposition 14.19. If Y is a T-stabilizing solution of the Riccati equation (14.58), then Y is Hermitian. Proof. Let Y be a stabilizing solution of (14.58), and put Δ = D0 + B ∗ Y B. Then Δ is invertible, and (14.84) σ A − BΔ−1 (C + B ∗ Y A) ⊂ D. Consider the m × m rational matrix functions W− (λ)

=

Im + Δ−1 (C + B ∗ Y A)(λIn − A)−1 B,

(14.85)

W+ (λ)

=

Δ + B ∗ (λ−1 In − A∗ )−1 (C ∗ + A∗ Y B).

(14.86)

14.7. J-spectral factorization relative to the unit circle revisited

285

The ﬁrst part of the proof consists of showing that W (λ) = W+ (λ)W− (λ) and that this factorization is a left canonical one with respect to the unit circle. Part 1. To prove that W (λ) = W+ (λ)W− (λ), we use a modiﬁcation of the argument used to prove (14.75). Put C0 = C + B ∗ Y A,

B0 = C ∗ + A∗ Y B.

(14.87)

Then equation (14.58) can be rewritten as Y − A∗ Y A = −C0 Δ−1 B0 , and hence λB0 Δ−1 C0 = −Y (λ − A) + (I − λA∗ )Y (λ − A) − λ(I − λA∗ )Y. It then follows that B ∗ (λ−1 − A∗ )−1 B0 Δ−1 C0 (λIn − A)−1 B = −B ∗ (λ−1 − A∗ )−1 A∗ Y B − B ∗ Y B − B ∗ Y A)(λIn − A)−1 B. This yields W+ (λ)W− (λ)

=

Δ + B ∗ (λ−1 − A∗ )−1 B0 + C0 (λIn − A)−1 B +B ∗ (λ−1 − A∗ )−1 B0 Δ−1 (λIn − A)−1 B

=

D0 + B ∗ (λ−1 − A∗ )−1 C ∗ + C(λIn − A)−1 B = W (λ).

Next we prove that W (λ) = W+ (λ)W− (λ) is a left canonical factorization with respect to the unit circle. To do this, using (14.85), we ﬁrst note that W− (λ)−1 = Im − Δ−1 (C + B ∗ Y A)(λIn − A× )−1 B,

(14.88)

where A× = A − BΔ−1 (C + B ∗ Y A). From (14.84) we know that A× has all its eigenvalues in D. By assumption the same holds true for the matrix A. Thus (14.85) and (14.88) tell us that both W− and W−−1 are analytic on the complement of D, inﬁnity included. Thus the factor W− has the desired properties. As A has all its eigenvalues in D, the same holds true for A∗ . Thus (14.86) tells us that W+ is analytic on the closed unit disc D. We have to show that W+−1 also is analytic on D. To do this, put ¯ −1 )∗ , V+ (λ) = W− (λ

¯ −1 )∗ . V− (λ) = W+ (λ

Using the properties of W− derived in the previous paragraph, we see that V+ and V+−1 are analytic on D. Furthermore, V− is analytic on |λ| ≥ 1, inﬁnity included. ¯ −1 )∗ , and Now, recall that W is selfadjoint on the unit circle. Hence W (λ) = W (λ thus W (λ) = W+ (λ)W− (λ) = V+ (λ)V− (λ). But then V− (λ)W− (λ)−1 = V+ (λ)−1 W+ (λ).

(14.89)

286

Chapter 14. J-spectral factorization

The left-hand side of (14.89) is analytic on |λ| ≥ 1 with inﬁnity included, and the right-hand side of (14.89) is analytic on D. By Liouville’s theorem, there exists a constant matrix K such that V− (λ) = KW− (λ),

W+ (λ) = V+ (λ)K.

(14.90)

As det W+ (λ) does not vanish identically, K is invertible. Hence the second identity in (14.90) tells us that W+ (λ) = K −1 V+ (λ)−1 is analytic on D. Thus W+ and W+−1 are analytic on D, as desired. We conclude that W (λ) = W+ (λ)W− (λ) is a left canonical factorization with respect to the unit circle. Part 2. In this part we establish the inclusion (14.91) σ A∗ − (C ∗ + A∗ Y B)Δ−1 B ∗ ⊂ D. Put Φ(λ) = W+ (λ−1 ). Then, with Ω = A∗ and Ω× = A∗ − (C ∗ + A∗ Y B)Δ−1 B ∗ , Φ(λ)

=

Δ + B ∗ (λIn − Ω)−1 (C ∗ + A∗ Y B),

(14.92)

Φ−1 (λ)

=

Δ−1 − Δ−1 B ∗ (λIn − Ω× )−1 (C ∗ + A∗ Y B)Δ−1 .

(14.93)

We want to prove that σ(Ω× ) ⊂ D. Take |λ0 | ≥ 1. As σ(Ω) ⊂ D, we have λ0 ∈ σ(Ω), and hence λ0 ∈ σ(Ω) ∩ σ(Ω× ). From (14.92) and (14.93) we see that Ω× is the associate main matrix of the realization (14.92). But then λ0 ∈ σ(Ω) ∩ σ(Ω× ) implies that the realization in (14.92) is locally minimal at λ0 . Since W+ and W+−1 are analytic on D, the rational matrix function Φ has no poles or zeros on |λ| ≥ 1. But then the local minimality at λ0 implies that λ0 is not an eigenvalue of Ω× . Recall that λ0 is an arbitary complex number with |λ0 | ≥ 1. We conclude that σ(Ω× ) is contained in D, that is, (14.91) is proved. −1 ). Since Part 3. Let T be the block Toeplitz operator on m 2 determined by W (λ W admits a left canonical factorization with respect to the unit circle, the function W (λ−1 ) admits a right canonical factorization with respect to the unit circle, and hence T is invertible. We claim that ⎤ ⎡ C ⎥ ⎢ ⎢ CA ⎥ ∗ ⎥ ⎢ −1 A∗ C ∗ A∗2 C ∗ · · · T ⎢ (14.94) Y =− C ⎥. ⎢CA2 ⎥ ⎦ ⎣ .. . Since the values of W (λ−1 ) on the unit circle are Hermitian, the operator T is selfadjoint, and hence the same holds true for T −1 . But then the identity (14.94) shows that Y is Hermitian. Thus it remains to prove (14.94). To prove (14.94) we follow the same line of reasoning as in the proof of Lemma 14.18. Put Θ(λ) = ΔW− (λ−1 ),

Φ(λ) = W+ (λ−1 ).

(14.95)

14.7. J-spectral factorization relative to the unit circle revisited

287

Here W+ and W− are as in Part 1 of the proof; see (14.85) and (14.86). By the result of Part 1 we have that W− (λ−1 ) = Φ(λ)Δ−1 Θ(λ). Moreover, Θ and Θ−1 are analytic on D, and Φ and Φ−1 are analytic on |λ| ≥ 1 with inﬁnity included. Let TΘ and TΦ be the block Toeplitz operators on m 2 determined by Θ and Φ, respectively. By the results mentioned in the previous paragraph, the operators TΘ and TΦ are invertible, TΘ−1 = TΘ−1 and TΦ−1 = TΦ−1 . Furthermore, T −1 = TΘ−1 Ξ−1 TΦ−1 , where, as in the proof of Lemma 14.18, the operator Ξ is the block diagonal operator on m 2 given by Ξ = diag (Δ−1 , Δ−1 , Δ−1 , . . .). Note that Θ−1 (λ)

=

Δ−1 − Δ−1 (C + B ∗ Y A)(λ=1 In − A× )−1 BΔ−1 ,

Φ−1 (λ)

=

Δ−1 − Δ−1 B ∗ (λIn − Ω× )−1 (C ∗ + A∗ Y B)Δ−1 .

Here A× = A − BΔ−1 (C + B ∗ Y A),

Ω× = A∗ − (C ∗ + A∗ Y B)Δ−1 B ∗ ,

and the eigenvalues of these two matrices are all in the open unit disc. Let Γ be the operator deﬁned by (14.68). We now repeat the arguments used in the proof of Lemma 14.18, more speciﬁcally appearing in the paragraphs after (14.79). This together with a duality argument yields Γ∗ T −1 Γ = (Γ∗ TΘ−1)Ξ−1 (TΦ−1 Γ) =

∞

−1 γ (A∗ )j βΔ Aj .

(14.96)

j=0

Here β =

∗

C Δ

−1

∗

−A

∞

(A∗ )j C ∗ Δ−1 (C + B ∗ Y A)(A× )j BΔ−1 ,

j=0

γ =

Δ−1 C − Δ−1 B ∗

(Ω× )j (C ∗ + A∗ Y B)Δ−1 CAJ A.

j=0

Note that the Riccati equation (14.58) can be rewritten in the following two equivalent forms Y − A∗ Y A×

=

−C ∗ Δ−1 (C + B ∗ Y A),

Y − Ω× Y A =

−(C ∗ + A∗ Y B)Δ−1 C.

Since the eigenvalues of the matrices A, A∗ , A× and Ω× are all in the open unit disc, we see that the formulas for β and γ can be transformed into ∗ ∗ −1 β = (C + A Y B)Δ , γ = Δ−1 (C + B ∗ Y A). γ , and we see from (14.96) This allows us to rewrite (14.58) as Y − A∗ Y A = −βΔ that (14.94) holds.

288

Chapter 14. J-spectral factorization

Notes As noted J-spectral factorization is a special form of canonical factorization, reﬂecting the symmetry condition on the given function. This chapter develops this theme in a systematic way for rational matrix functions. Sections 14.2 and 14.3 are based on [121]. For Section 14.4 we refer to [76], see also [112] and [83]. A good source for Section 14.5 is [98], see also [97]. The linear quadratic optimal control problem for discrete time systems, mentioned in Section 14.5 in the paragraph before Proposition 14.12, can be found in many books on mathematical systems theory, see, e.g., [85]. The connection with the algebraic Riccati equation of the form (14.45) is also shown in the latter book. Much more information on this equation, including its connection to factorization in more general setting than the one exhibited in Proposition 14.12, can be found in Part III of [106]. Section 14.6 is based on [9], see also [8]. The ﬁnal section is inspired by [44]. In fact, Theorem 14.15 is just the symmetric version of Theorem 1.1 in [44]. The notion of J-spectral factorization plays an important role in control theory; see, e.g., the books [43], [85], [150], the papers [76], [145] and the references in these papers. The ﬁnal part of this book is devoted to this connection, with an emphasis on H∞ -problems.

Part VI Factorizations and symmetries In this part we study rational matrix functions that are unitary or of the form identity matrix plus contractions, and rational matrix functions that have a positive real part. Because of the state space similarity theorem, these additional symmetries can be restated in terms of special properties of the minimal realizations of the rational matrix functions considered. These reformulations involve an algebraic Riccati equation. The results are known in systems theory as the bounded real lemma and the positive real lemma, respectively. This part consists of three chapters. In the ﬁrst chapter (Chapter 15) we study rational matrix functions that have a positive deﬁnite real part or a nonnegative real part on the real line, and we present canonical and pseudo-canonical factorization theorems for such functions in state space form. In the second chapter (Chapter 16) realizations are used to study rational matrix functions of which the values on the imaginary axis (or on the real line) are contractive matrices. Included are solutions to spectral and canonical factorization problems for functions V of the form ¯ ∗ W (λ), V (λ) = I − W (−λ) V (λ) = I + W (λ), where W has contractive values on the imaginary axis (or on the real line) and is strictly contractive at inﬁnity. In the third chapter (Chapter 17) realizations are used to study rational matrix functions of which the values on the imaginary axis are J-unitary matrices. Solutions to various factorization problems are given. Special attention is paid to factorization of J-unitary rational matrix functions into J-unitary factors. In this chapter we also discuss problems of embedding a contractive rational matrix function into a unitary rational matrix function of larger size.

Chapter 15

Factorization of positive real rational matrix functions This chapter is concerned with canonical factorization (with respect to the real line) of rational matrix functions with a positive deﬁnite real part on the real line. Also the generalization to pseudo-canonical factorization for functions that have a nonnegative real part is developed. All factorizations are obtained explicitly using state space realizations of the functions involved. In Section 15.1 rational matrix functions that have a positive deﬁnite real part or a nonnegative real part on the real line are characterized in terms of realizations. Section 15.2 deals with canonical factorization, and Section 15.3, the ﬁnal section of the chapter, with pseudo-canonical factorization.

15.1 Rational matrix functions with a positive deﬁnite real part In this section we consider rational m × m matrix functions W which have the property that W (λ) + W (λ)∗ ≥ 0,

λ ∈ R, λ not a pole of W .

(15.1)

In this case we say that W has a nonnegative real part on the real line. If in (15.1) the inequality is strict, that is, W (λ) + W (λ)∗ > 0,

λ ∈ R, λ not a pole of W .

(15.2)

we say that W has a positive deﬁnite real part on the real line. The following two theorems characterize these properties in terms of realizations of W .

292

Chapter 15. Factorization of positive real rational matrix functions

Theorem 15.1. Let W (λ) = D + C(λIn − A)−1 B be a rational m × m matrix function, and let (A, B) be controllable. Write G = D + D∗ and assume G is positive deﬁnite. Then W has a nonnegative real part on the real line if and only if there is a Hermitian solution X of the equation −iA∗ X + iXA − (XB − iC ∗ ) G−1 (B ∗ X + iC) = 0.

(15.3)

Furthermore, for any Hermitian solution X of (15.3) one has

where

¯ ∗ = K(λ) ¯ ∗ K(λ), W (λ) + W (λ)

(15.4)

K(λ) = G1/2 + G−1/2 (C − iB ∗ X)(λIn − A)−1 B.

(15.5)

Finally, if, in addition, the pair (C, A) is observable, then each solution X of (15.3) is invertible. For later use we note that equation (15.3) can be rewritten as −(iA∗ −iC ∗ G−1 B ∗ )X +X(iA−iBG−1C)−C ∗ G−1 C −XBG−1B ∗ X = 0. (15.6) ¯ ∗ . Then W has a nonnegative real part on R Proof. Put V (λ) = W (−iλ) + W (iλ) if and only if V is nonnegative on the imaginary axis. Using the given realization of W we have V (λ)

= = =

D + C(−iλIn − A)−1 B + D∗ + B ∗ (−iλIn − A∗ )−1 C ∗ −1 −1 G + (iC) λIn − (iA) B − B ∗ λIn + (iA)∗ (iC)∗ . (λ − iA)−1 B 0 (iC)∗ −1 . I −B ∗ λ + (iA)∗ iC G I

Thus we can apply Theorem 13.6, with R = G, Q = 0, S = iC and iA instead of A, to show that W has a nonnegative real part on R if and only equation (15.3) has a Hermitian solution. Next, let X be a Hermitian solution of (15.3). By the second part of Theo¯ ∗ L(λ), where rem 13.6, the function V admits a factorization V (λ) = L(−λ) L(λ) = G1/2 + G−1/2 (B ∗ X + iC)(λ − iA)−1 B. ¯ ∗ = V (iλ), we see that (15.4) holds with K being given by (15.5) As W (λ) + W (λ) To prove the ﬁnal part, assume additionally that the pair (C, A) is observable, and let X be a Hermitian solution of (15.3). We have to show that X is invertible. Since X is square it suﬃces to prove that Ker X = {0}. Assume Xx = 0. Then x∗ X = 0 because X is Hermitian, and by (15.3) we have 0 = −C ∗ G−1 Cx, x. As G > 0, this gives Cx = 0. Multiplying (15.3) on the right by x we then obtain iXAx = 0. So Ker X is A-invariant and contained in Ker C. Therefore Ker X = {0} and X is invertible.

15.1. Rational matrix functions with a positive deﬁnite real part

293

Theorem 15.2. Let W (λ) = D + C(λIn − A)−1 B be a rational m × m matrix function, and let (A, B) be controllable. Write G = D + D∗ and assume G is positive deﬁnite. If, in addition, A has no real eigenvalues, then the following statements are equivalent: (i) The function W has a positive deﬁnite real part on the real line; (ii) Equation (15.3) has a Hermitian solution X such that the matrix A − BG−1 C + iBG−1 B ∗ X

(15.7)

has no real eigenvalues; (iii) The matrix

⎡ H =⎣

iA∗ − iC ∗ G−1 B ∗

C ∗ G−1 C

−BG−1 B ∗

iA − iBG−1 C

⎤ ⎦

has no pure imaginary eigenvalues. Moreover, in that case equation (15.3) has a unique Hermitian solution X such that the matrix (15.7) has its eigenvalues in the open upper half plane. Proof. As in the proof of the previous theorem, we consider the rational m × m ¯ ∗ . Using the given realization of W we matrix function V (λ) = W (−iλ) + W (iλ) see (see (13.6) and the second part of the proof of Theorem 13.2) that V admits −1 B, where 2n − A) the realization V (λ) = G + C(λI iC ∗ iA∗ 0 = = = B ∗ iC . , B A , C 0 iA B − BG −1 C is precisely equal to the block matrix H appear× = A It follows that A has no pure imaginary ing in item (c). Since A has no real eigenvalue, the matrix A eigenvalue. Thus V has no pole on the imaginary axis. Hence (cf., Section 8.1) the × )−1 BG 2n − A −1 is minimal at each point realization V −1 (λ) = G−1 − G−1 C(λI of the imaginary axis. But then V −1 has no pole on the imaginary axis if and only × has no pure imaginary eigenvalue. As A × = H, we conclude that condition if A (iii) is equivalent to the requirement that V (λ) is invertible for each λ ∈ iR. (i) ⇒ (iii) If (i) is satisﬁed, then V (λ) is positive deﬁnite for each λ ∈ iR. In particular, V (λ) is invertible for each λ ∈ iR, and hence, by the result of the previous paragraph, (iii) holds. (iii) ⇒ (i) Conversely, assume (iii) is satisﬁed. Recall that V has no pole on the imaginary axis. Furthermore, V (λ) is selfadjoint for λ ∈ iR. Since V (λ) is invertible for each λ ∈ iR, it follows that for imaginary λ the signature of the matrix V (λ) does not depend on λ. Next, observe that the rational matrix function V is biproper and that its value at inﬁnity is equal to G. Hence the value of V

294

Chapter 15. Factorization of positive real rational matrix functions

at inﬁnity is positive deﬁnite. We obtain that V (λ) is positive deﬁnite for each λ ∈ iR. Thus (i) holds. (i) ⇒ (ii) Assume W has a positive deﬁnite real part on R. Theorem 15.1 implies that equation (15.3) has a Hermitian solution X. Hence we have the fac¯ ∗ K(λ) with K(λ) being given by (15.5). Since ¯ ∗ = K(λ) torization W (λ) + W (λ) A has no eigenvalue on R, the functions W and K have no pole on R. The fact that W has a positive deﬁnite real part on R and the fact that W has no pole on ¯ ∗ is invertible for each λ ∈ R. Hence K(λ) is R together imply that W (λ) + W (λ) also invertible for each λ ∈ R. Thus K(λ)−1 has no pole on R. Notice that K(λ)−1 = G−1/2 − G−1 (C − iB ∗ X)(λ − Z)−1 B, −1

(15.8)

∗

where Z = A − BG (C − iB ). Let λ0 ∈ R. Then λ0 is not a common eigenvalue of A and Z. Thus we can apply the material presented in Section 8.1 to show that the realization given by the right-hand side of (15.8) is minimal at λ0 . But then the fact that K(λ)−1 has no pole on R implies that λ0 is not an eigenvalue of Z. Thus Z = A − BG−1 C + iBG−1 B ∗ X has no real eigenvalue. This proves (ii). ¯ ∗ = K(λ) ¯ ∗ K(λ) with K(λ) (ii) ⇒ (i) Let X be as in (ii). Then W (λ) + W (λ) −1 being given by (15.5). Observe that K(λ) is given by (15.8), where Z is as ¯ ∗ K(λ) above. According to our hypothesis Z has no real eigenvalue. Hence K(λ) is positive deﬁnite for each λ ∈ R. Thus (i) holds. To prove the second part of the theorem, we apply Theorem 13.3. Recall that equation (15.3) can be rewritten into the algebraic Riccati equation (15.6). The Hamiltonian of this Riccati equation is precisely the block matrix H deﬁned in item (iii). According to our hypotheses (A, B) is controllable. This implies that the pair (iA − iBG−1 C, B) is also controllable. But controllability implies stabilizability. Thus the pair (iA − iBG−1 C, B) is stabilizable. But then Theorem 13.3 tells us that condition (iii) implies that equation (15.3) has a unique Hermitian solution X such that the eigenvalues of iA − iBG−1 C − BG−1 B ∗ X are in the open left half plane. Multiplication by −i then gives the desired result.

15.2 Canonical factorization of functions with a positive deﬁnite real part In this section we consider canonical factorization of functions with a positive deﬁnite real part on the real line. Using state space realizations we shall prove the following result. Theorem 15.3. Let W be a proper rational matrix function having no real poles and such that D = W (∞) satisﬁes D + D ∗ > 0. Assume that W has a positive deﬁnite real part on the real line. Then W admits both a right and a left canonical factorization with respect to the real line. We start with some preparations that are of independent interest and will be useful in the next section too. Let T be a square matrix. If the real part of T

15.2. Canonical factorization of functions with a positive deﬁnite real part

295

is positive deﬁnite, then T is injective, hence invertible. Indeed, for non-zero x we have 2(T x, x) = (T + T ∗ )x, x > 0. Also, if T is invertible, then T −1 has a positive deﬁnite real part if and only if this is the case for T . This is immediate from either of the identities T −1 + T −∗ = T −1(T + T ∗ ) T −∗ ,

T −1 + T −∗ = T −∗ (T + T ∗ ) T −1 .

Now let W (λ) = D + C(λIn − A)−1 B be a rational m × m matrix function with G = D + D ∗ positive deﬁnite, and assume W has a nonnegative real part on R. Then D is invertible, G× deﬁned by G× = D−1 + D−∗ is positive deﬁnite, G× = D−1 G D−∗ , and W −1 has a nonnegative real part on R. For W −1 we have the realization W −1 (λ) = D−1 − D−1 C(λIn − A× )−1 BD−1 ,

(15.9)

where, as usual, A× = A − BD−1 C. This gives rise to the following analogue of equation (15.3): −i(A× )∗ X + iXA× − (XBD−1 + iC ∗ D−∗ )(G× )−1 (D −∗ B ∗ X − iD −1 C) = 0, (15.10) which can also be written as an algebraic Riccati equation − i(A× )∗ + iC ∗ D−∗ (G× )−1 D−∗ B ∗ X (15.11) +X iA× + iBD−1 (G× )−1 D −1 C −C ∗ D −∗ (G× )−1 D−1 C − XBD−1 (G× )−1 D−∗ B ∗ X = 0. Now let us look at the right coeﬃcient of X in this expression. Using the identity (G× )−1 = DG−1 D∗ , we get iA× + iBD−1 (G× )−1 D−1 C

=

iA − iBD−1 C + iBD−1 (DG−1 D∗ )D −1 C

=

iA − iBD−1 C + iBG−1 D ∗ D−1 C

=

iA − iBG−1 (G − D∗ )D−1 C

=

iA − iBG−1 DD−1 C = iA − iBG−1 C.

Thus the right coeﬃcient of X in (15.11) is equal to the right coeﬃcient of X in (15.6). The left coeﬃcient of X in (15.11) is the adjoint of the right coeﬃcient of X in (15.11), and the same is true with (15.11) replaced by (15.6). Hence the left coeﬃcient of X in (15.11) is equal to the left coeﬃcient of X in (15.6). For the constant term in (15.11), we have −C ∗ D−∗ (G× )−1 D−1 C = −C ∗ D−∗ (D ∗ G−1 D)D−1 C = −C ∗ G−1 C, and the latter is the constant term in (15.11). Finally, the identities −BD −1 (G× )−1 D −∗ B ∗ = −BD−1 (DG−1 D∗ )D−∗ B ∗ = −BG−1 B ∗

296

Chapter 15. Factorization of positive real rational matrix functions

show that the coeﬃcients of the quadratic terms in (15.11) and (15.6) coincide too. We conclude that the equations (15.3), (15.6), (15.10) and (15.11) all amount to the same. Lemma 15.4. Let W (λ) = D +C(λIn −A)−1 B be a rational m×m matrix function such that G = D + D∗ > 0. Assume X is an invertible Hermitian matrix satisfying (15.3). Then 1 1 (XA − A∗ X) = − (B ∗ X + iC)∗ G−1 (B ∗ X + iC), 2i 2 1 1 = − (DD−∗ B ∗ X − iC)∗ G−1 (DD−∗ B ∗ X − iC). XA× − (A× )∗ X 2i 2 In particular both A and A× are (−X)-dissipative. Proof. The ﬁrst identity is just a restatement of (15.3). Recall that (15.3) and (15.10) amount to the same. Hence X also satisﬁes (15.10). Now note that the second identity in the lemma is just another way of writing (15.10). Here we use that (G× )−1 = D∗ G−1 D. Before turning to the proof of Theorem 15.3 we present another lemma. Lemma 15.5. Let W (λ) = D +C(λIn −A)−1 B be a rational m×m matrix function such that G = D + D ∗ > 0 and the pair (C, A) is observable. Assume X is an invertible Hermitian matrix satisfying (15.3). Let N1 , N1× be maximal X-nonpositive subspaces and N2 , N2× be maximal X-nonnegative subspaces such that N1 , N2 are invariant under A and N1× , N2× are invariant under A× . Then ˙ N2× , Cn = N1 +

˙ N1× . C n = N2 +

(15.12)

Proof. Applying Proposition 11.1 we obtain dim N1 + dim N2× = n,

dim N2 + dim N1× = n.

Therefore in order to prove that (15.12) holds, it suﬃces to show that the intersections N1 ∩ N2× and N2 ∩ N1× are both trivial. Take x ∈ N1 ∩ N2× . Then Xx, x = 0. Now the Cauchy-Schwartz inequality holds on N1 . Thus |XAx, x|2 ≤ XAx, AxXx, x = 0, |Xx, Ax|2 ≤ XAx, AxXx, x = 0. Using this together with the ﬁrst identity in Lemma 15.4, we get 1 0 = XAx, x = − G−1/2 (B ∗ X + iC)x2 . 2 Similarly, employing the Cauchy-Schwartz inequality on N2× and the second identity in Lemma 15.4, we get 1 0 = XA× x, x = − G−1/2 (DD−∗ B ∗ X − iC)x2 . 2

15.3. Generalization to pseudo-canonical factorization

297

Thus (B ∗ X + iC)x = 0 and (DD−∗ B ∗ X − iC)x = 0. Adding these two identities we arrive at 0 = (I + DD −∗ )B ∗ Xx = GD −∗ B ∗ Xx. Hence B ∗ Xx = 0, and it also follows that Cx = 0. Thus Ax = A× x for x ∈ N1 ∩ N2× . Hence N1 ∩ N2× is an A-invariant subspace contained in Ker C. Given the observability of the pair (C, A), this yields N1 ∩ N2× = {0}. The proof of N2 ∩N1× = {0} is analogous. It can also be obtained by applying the result of the previous paragraph to the rational matrix function W −1 . Proof of Theorem 15.3. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of W . Since W has no poles on R, the minimality of the realization guarantees that A has no eigenvalues on R. As we have seen (in the ﬁrst paragraph after Theorem 15.3), the positive deﬁniteness of D + D∗ implies that D is invertible. Similarly we conclude that W takes invertible values on R. Hence we know from Theorem 2.4 that A× = A − BD−1 C has no real eigenvalues either. Since W has a positive deﬁnite real part, we can use Theorem 15.1 to deduce that equation (15.3) has an invertible Hermitian solution X, say. Lemma 15.4 now gives that both A and A× are (−X)-dissipative. × be the spectral subspaces of A and A× , respectively, corLet M+ and M+ × responding to the open upper half plane, and let M− and M− , be the spectral subspaces of A and A× , respectively, corresponding to the open lower half plane. × As A and A× are (−X)-dissipative, we have that M+ and M+ are maximal X× nonpositive. Similarly, the spaces M− and M− are maximal X-nonnegative. Using × × ˙ M− ˙ M+ Lemma 15.5 we may conclude that Cn = M+ + and Cn = M− + . But then Theorem 3.2 guarantees that W admits the desired canonical factorizations.

15.3 Generalization to pseudo-canonical factorization In this section the results of the previous section concerning canonical factorizations will be generalized to pseudo-canonical factorizations. Theorem 15.6. Let W be a proper rational m × m matrix function having no real poles such that D = W (∞) satisﬁes D+D ∗ > 0. Assume that W has a nonnegative real part on the real line. Then, with respect to the real line, W admits both right and left pseudo-canonical factorization. Such factorizations can be obtained in the following manner. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization, and put G = D + D∗ . Then there exists an invertible Hermitian matrix X satisfying −iA∗ X + iXA − (XB − iC ∗ ) G−1 (B ∗ X + iC) = 0.

(15.13)

Also there are A-invariant subspaces M+ and M− , and A× -invariant subspaces × × M+ and M− , such that (i) M+ is maximal X-nonpositive, M+ contains the spectral subspace of A associated with the part of σ(A) lying in the open upper half plane, and σ(A|M+ ) ⊂ {λ | λ ≥ 0},

298

Chapter 15. Factorization of positive real rational matrix functions

(ii) M− is maximal X-nonnegative, M− contains the spectral subspace of A associated with the part of σ(A) lying in the open lower half plane, and σ(A|M− ) ⊂ {λ | λ ≤ 0}, × × is maximal X-nonpositive, M+ contains the spectral subspace of A× (iii) M+ × associated with the part of σ(A ) lying in the open upper half plane, and σ(A× |M × ) ⊂ {λ | λ ≥ 0}, +

(iv)

× M−

× is maximal X-nonnegative, M− contains the spectral subspace of A× × associated with the part of σ(A ) lying in the open lower half plane, and σ(A× |M × ) ⊂ {λ | λ ≤ 0}. −

For such subspaces the matching conditions × ˙ M− Cn = M+ + ,

× ˙ M+ Cn = M− +

(15.14)

× and Πl for the are satisﬁed. Write Πr for the projection along M+ onto M− × projection along M− onto M+ . Further put

!− (λ) W

=

D + C(λIn − A)−1 (In − Πr )B,

!+ (λ) W

=

In + D−1 CΠr (λIn − A)−1 B,

W+ (λ)

=

D + C(λIn − A)−1 (In − Πl )B,

W− (λ)

=

In + D−1 CΠl (λIn − A)−1 B.

!− (λ)W !+ (λ) and W (λ) = W+ (λ)W− (λ) are a right and a left Then W (λ) = W pseudo-canonical factorization with respect to the real line, respectively. Proof. In view of the minimality of the given realization we can employ Theorem 15.1 to show that there is an invertible Hermitian matrix X such that (15.13), which is identical to (15.3), holds. By Lemma 15.4 the matrices A and × × and M− with A× are (−X)-dissipative. The existence of subspaces M+ , M− , M+ the properties mentioned above is now guaranteed by Theorem 11.6. Lemma 15.5 gives the direct sums (15.14), and the conclusion of the theorem is straightforward by Theorem 8.6. As a further application of Lemma 15.5 we prove the following result on skew selfadjoint matrix functions. A rational m × m matrix function W is called skewHermitian on the real line if W (λ) is skew-Hermitian for all λ in R, λ not a pole of W . Proposition 15.7. Let W (λ) = D + V (λ), where V is a strictly proper rational m × m matrix function that has no real poles, is skew-Hermitian on the real line and vanishes at inﬁnity. Assume D + D ∗ > 0. The following statements are true.

15.3. Generalization to pseudo-canonical factorization

299

!1 (λ)W !2 (λ) where W !1 has all (i) W admits a minimal factorization W (λ) = W its poles, respectively zeros, in the open upper, respectively lower, half plane, !2 has all its poles, respectively zeros, in the open lower, respectively and W upper, half plane. (ii) W admits a minimal factorization W (λ) = W1 (λ)W2 (λ) where W1 has all its poles, respectively zeros, in the open lower, respectively upper, half plane, and W2 has all its poles, respectively zeros, in the open upper, respectively lower, half plane. Proof. Recall that D + D ∗ > 0 implies that D is invertible. Since V is skewHermitian on the real line, we see that W (λ) + W (λ)∗ = D + D ∗ > 0 for λ ∈ R. From the latter it follows that W (λ) is invertible for each λ ∈ R. Now let W (λ) = D + C(λIn − A)−1 B be a minimal realization of W . Then both A and A× have no eigenvalues on the real line. From C(λIn − A)−1 B = −(C(λIn − A)−1 B)∗ for λ ∈ R and the minimality of the realization we may conclude (by the state space similarity theorem) that there is a unique invertible matrix Y such that Y A = A∗ Y,

Y B = C∗,

C = −B ∗ Y.

Taking adjoints in the above equations, and using the uniqueness of Y , one deduces that Y = −Y ∗ . Put X = −iY . Then X is selfadjoint. As XA = A∗ X, the matrix A is X-selfadjoint. Furthermore, from −iA∗ X + iXA = 0 and XB − iC ∗ = 0, we see that X is an invertible Hermitian solution of (15.13). But then we can use Lemma 15.4 to show that A× is (−X)-dissipative. Let Mu and Ml be the spectral subspaces of A associated with the part of σ(A) lying in the open upper and open lower half plane, respectively. Also let Mu× and Ml× be the spectral subspaces of A× associated with the part of σ(A× ) lying in the open upper and open lower half plane, respectively. Since the matrix A is X-selfadjoint and has no real eigenvalues, we know (see Theorem 11.5) that the spaces Mu and Ml are X-Lagrangian. In particular, these spaces are both maximal X-nonpositive and maximal X-nonnegative. The fact that A× is (−X)-dissipative and has no real eigenvalues either, gives that the same conclusion holds for Mu× ˙ Mu× as well as Cn = Ml + ˙ Ml× . and Ml× . But then Lemma 15.5 gives Cn = Mu + Let Π bethe projection of Cn along Mu onto Mu× . Then Π is a supporting projection of the minimal realization W (λ) = D + C(λI − A)−1 B. Hence W !2 (λ) such that (see Chapter 8) !1 (λ)W admits a minimal factorization W (λ) = W !1 and W !2 coincide with the eigenvalues of A|Mu the following holds: the poles of W !1 and W !2 coincide with the eigenvalues and A|Mu× , respectively, and the zeros of W × × of A |Mu and A |Mu× , respectively. Since A and A× have no real eigenvalues ˙ Mu× = Cn , we have σ(A|M × ) = σ(A|Ml ) and σ(A× |Mu ) = σ(A× |M × ). and Mu + u l !2 (λ) has the !1 (λ)W From these remarks it is clear that the factorization W (λ) = W desired properties. The factorization W (λ) = W1 (λ)W2 (λ) is obtained in a similar ˙ Ml× . way using the other direct sum decomposition Cn = Ml +

300

Chapter 15. Factorization of positive real rational matrix functions

Notes This chapter is based on [126], see also [129] and [128]. Rational matrix functions with a positive deﬁnite real part play a role in circuit and systems theory. In particular, Theorem 15.2 is a version of what is known as the positive real lemma. There are several variants of this result, see, for instance, Section 5.2 in [4], where also the connection with spectral factorization and Riccati equations is discussed. Another version in terms of Riccati inequalities is given in Section 12.6.3 in [83]. An inﬁnite dimensional version may be found as Exercise 6.28 in [35].

Chapter 16

Contractive rational matrix functions In this chapter rational matrix functions are studied of which the values on the imaginary axis or on the real line are contractive matrices. Included are solutions to spectral or canonical factorization problems for functions V of the form ¯ ∗ W (λ) V (λ) = I − W (−λ)

or V (λ) = I + W (λ),

where W is a rational matrix function which has contractive values on the imaginary axis or on the real line and, in addition, has a strictly contractive value at inﬁnity. This chapter consists of ﬁve sections. Sections 16.1 and 16.2 present a state space analysis (involving algebraic Riccati equations) of rational matrix functions that are contractive or strictly contractive on the imaginary axis. In Section 16.3 a state space formula is derived for the spectral factor in a spectral factorization of a ¯ ∗ W (λ), where W is strictly rational matrix function of the form V (λ) = I −W (−λ) proper and strictly contractive on the imaginary axis. The ﬁnal two sections of the chapter deal with canonical and pseudo-canonical factorization, respectively, for functions of the form V (λ) = I + W (λ), where W (λ) is strictly proper and strictly contractive for real λ (Section 16.4) or just contractive (Section 16.5).

16.1 State space analysis of contractive rational matrix functions A rational p × m matrix function W is called contractive on the imaginary axis if the values that W takes on the imaginary axis are contractive matrices. Such a function does not have a pole on the imaginary axis. Moreover, it is proper and the value at inﬁnity is again contractive. Of special interest is the subclass consisting

302

Chapter 16. Contractive rational matrix functions

of the contractive rational matrix functions W that are strictly contractive at inﬁnity, i.e., the value of W at ∞ has norm smaller than 1. The ﬁrst main result of this section is a characterization of this subclass in terms of realizations. Theorem 16.1. Let W (λ) = D + C(λIn − A)−1 B be a realization of a p × m rational matrix function, and assume D is a strict contraction. Then the following two assertions hold: (i) Assume (C, A) is an observable pair. Then W is contractive on the imaginary axis if and only if the algebraic Riccati equation −AP − P A∗ − BB ∗ − (P C ∗ + BD∗ )(I − DD ∗ )−1 (CP + DB ∗ ) = 0 (16.1) has a Hermitian solution P . (ii) Assume (A, B) is a controllable pair. Then W is contractive on the imaginary axis if and only if the algebraic Riccati equation A∗ P + P A − C ∗ C − (P B − C ∗ D)(I − D∗ D)−1 (B ∗ P − D∗ C) = 0 (16.2) has a Hermitian solution P . ¯ ∗ . Since W is proper, the same holds true for Proof. Put V (λ) = I − W (λ)W (−λ) ∗ V . Moreover, V (∞) = I − DD , and hence V (∞) is positive deﬁnite, because D is assumed to be a strict contraction. Note that W is contractive on iR if and only if V is nonnegative on iR. Using the given realization for W we have V (λ)

∗

= I − DD +

=

C

DB

−1

−C(λ − A)

I

∗

$

λ−

A BB ∗ 0

%−1

−A∗

−BB ∗

BD∗

DB ∗

I − DD∗

−BD ∗

C∗

(λ + A∗ )−1 C ∗ I

.

The latter expression is of the form (13.25) and we see that (i) is an immediate consequence of the equivalence of statements (i) and (ii) in Theorem 13.6. To prove assertion (ii) we use a duality argument. First note that a matrix X is a (strict) contraction if and only if X ∗ is a (strict) contraction. So W is ¯ ∗ . The latter contractive on iR if and only this is the case for the function W (−λ) ∗ ∗ ∗ ∗ −1 ∗ ¯ has the realization W (−λ) = D − B (λ + A ) C . Also, the controllability of the pair (A, B) implies the observability of (B ∗ , −A∗ ). Finally, D∗ is a strict contraction. Thus assertion (ii) follows from part (i) by taking adjoints. Suppose D is a strict contraction. If the pair (C, A) is observable, then each Hermitian solution P of (16.2) is invertible. To see this, we argue as follows. Assume P x = 0. Multiplying (16.2) from the left by x∗ and from the right by x yields x∗ C ∗ Cx + x∗ C ∗ D(I − D∗ D)−1 DC ∗ x = 0. Now C ∗ C and I − D∗ D are

16.1. State space analysis of contractive rational matrix functions

303

nonnegative (in fact even I − DD∗ > 0), and it follows that x∗ C ∗ Cx = 0. Hence Cx = 0. But then, multiplying (16.2) on the right by x, we get P Ax = 0. So Ker P is A-invariant and contained in Ker C. As (C, A) is an observable pair, it follows that Ker P = {0}. Since P is a square matrix, this yields the invertibility of P . In a similar fashion one proves that each solution of (16.1) is invertible provided that the pair (A, B) is controllable, or, which amounts to the same, the pair B ∗ , A∗ ) is observable. Finally we note that P is an invertible solution of (16.2) if and only if −P 1 is an invertible solution of (16.1). Indeed, replacing P by −P −1 in (16.1) and multiplying from the left and the right with P , one gets (16.2). In working out the details, identities of the type D(I − D∗ D)−1 = (I − DD∗ )−1 D and I + D(I − D ∗ D)−1 D∗ = (I − DD ∗ )−1 play a role. Theorem 16.2. Let W (λ) = D+C(λIn −A)−1 B be a realization of a p×m rational matrix function, and let D be a strict contraction. Assume, in addition, that the pair (C, A) is observable. Then W is contractive on the imaginary axis if and only if the matrix A + BD∗ (I − DD∗ )−1 C B(I − D ∗ D)−1 B ∗ T = (16.3) −C ∗ (I − DD∗ )−1 C −A∗ − C ∗ (I − DD∗ )−1 DB ∗ has only even partial multiplicities at its pure imaginary eigenvalues. Proof. Let V be as in the proof of Theorem 16.1, and recall that W is contractive on iR if and only if V is nonnegative on iR. The desired result is now immediate from the equivalence of statements (ii) and (iii) in Theorem 13.6 combined with the fact that (16.3) is the Hamiltonian of the equation (16.1). Theorem 16.2 has a counterpart in which (16.3) is replaced by the Hamiltonian of (16.2). As a special case of Theorem 16.1 let us consider rational matrix functions which are contractive not only on the imaginary axis but on the full closed right half plane. Theorem 16.3. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a rational p× m matrix function. Assume that W is contractive on the imaginary axis, and let D be a strict contraction. Then the following statements are equivalent: (i) For each λ in the closed right half plane, λ not a pole of W , the matrix W (λ) is a contraction; (ii) The matrix A has all its eigenvalues in the open left half plane; (iii) There is a positive deﬁnite solution of (16.1). Proof. Suppose A has all its eigenvalues in the open left half plane. Then W (λ) is analytic in the closed right half plane. As W (λ) is a contraction for each λ ∈ iR and at inﬁnity, the maximum modulus theorem implies W (λ) is contractive for all

304

Chapter 16. Contractive rational matrix functions

λ in the open right half plane as well. Thus (ii) implies (i). Conversely, suppose (i) holds. Then W must be analytic in the closed right half plane (inﬁnity included), and by minimality of the realization the matrix A has all its eigenvalues in the open left half plane. The equivalence of (ii) and (iii) follows by rewriting (16.1) as AP + P A∗ = RR∗ , where Q = −P and R = B (P C ∗ + BD ∗ )(I − DD ∗ )−1/2 . Since the realization is minimal, (A, B) is a controllable pair, and hence the same holds true for the pair (A, R). But then we can apply a well-known inertia theorem (see, e.g., Theorem 13.1.4 in [107]) to show that (ii) and (iii) are equivalent.

16.2 Strictly contractive rational matrix functions In this section we specify further the results of the previous section for the case of rational matrix functions W that are strictly contractive on the imaginary axis. By this we mean that W (λ) < 1 for λ ∈ iR. Such a function does not have a pole on iR and is proper. Theorem 16.4. Let W (λ) = D +C(λIn −A)−1 B be a realization of a p×m rational matrix function W . Assume A has no pure imaginary eigenvalues, D is a strict contraction, and the pair (C, A) is observable. Then the following statements are equivalent: (i) The function W is strictly contractive on the imaginary axis; (ii) Equation (16.1) has an iR-stabilizing solution P , that is, it has a solution P such that −A∗ − C ∗ (I − DD∗ )−1 DB ∗ − C ∗ (I − DD∗ )−1 CP has all its eigenvalues in the open left half plane; (iii) The matrix T given by (16.3) has no pure imaginary eigenvalues. Moreover, if one of the above conditions is satisﬁed, then the iR-stabilizing solution P in (ii) is unique and Hermitian. Proof. Suppose W is strictly contractive on iR. Then the rational m × m matrix ¯ ∗ is positive deﬁnite on iR and V (∞) = I − DD ∗ function V (λ) = I − W (λ)W (−λ) is positive deﬁnite too. In particular, V is biproper and V has no pole or zero on iR. Recall (see the proof of Theorem 16.1) that %−1 $ −BD∗ A BB ∗ ∗ ∗ V (λ) = I − DD + C DB . λ− 0 −A∗ C∗ The associate main matrix of this realization is T given by (16.3). It follows that T has no eigenvalues on iR. So (iii) holds. Conversely, if T has no pure imaginary eigenvalues, then V has no poles or zeros on iR. As V (∞) is positive deﬁnite, it follows that V (λ) is positive deﬁnite for λ ∈ iR. Hence W (∞) is strictly contractive

16.3. An application to spectral factorization

305

for λ ∈ iR. We have now proved the equivalence of (i) and (iii). The equivalence of (ii) and (iii) is a direct consequence of Theorem 13.3. Note here that the observability of the pair (C, A) is equivalent to the controllability of (A∗ , C ∗ ), and the latter implies the stabilizabilty of (A∗ , C ∗ ). The ﬁnal statement of the theorem is covered by Theorem 13.3 as well. Corollary 16.5. Let (C, A) be an observable pair, and assume that A has no pure imaginary eigenvalue. Then the Riccati equation Y C ∗ CY − Y A∗ − AY = 0 has a unique Hermitian solution Y such that A − Y C ∗ C has all its eigenvalues in the open right half plane. Proof. Apply Theorem 16.4 with D = 0 and B = 0. Then D is a strict contraction and W is identically equal to zero. In particular, (i) in Theorem 16.4 is satisﬁed. Next, note that with D = 0 and B = 0 equation (16.1) reduces to −AP − P A∗ − P C ∗ CP = 0, and by Theorem 16.4, with D = 0 and B = 0, this equation has a unique Hermitian solution P such that −A∗ −C ∗ CP has all its eigenvalues in the open left half plane. But then A − Y C ∗ C has all its eigenvalues in the open right half plane. Now put Y = −P , then we see that Y has all the desired properties.

16.3 An application to spectral factorization In this section we consider functions of the form ¯ ∗ W (λ), V (λ) = I − W (−λ)

(16.4)

where W is a proper rational p × m matrix function which is strictly contractive on the imaginary axis. In fact we shall assume that W is strictly proper , that is W vanishes at inﬁnity. Thus V is positive deﬁnite on the imaginary axis and has a positive deﬁnite value at inﬁnity (namely Im ). Hence W admits a right spectral factorization. Using a minimal realization of W , such a factorization is constructed in the following theorem. Theorem 16.6. Let W (λ) = C(λIn − A)−1 B be a minimal realization of the p × m rational matrix function W which is strictly contractive on the imaginary axis. Then the Riccati equations XBB ∗ X − XA − A∗ X + C ∗ C

=

0,

(16.5)

Y C ∗ CY − Y A∗ − AY

=

0,

(16.6)

have Hermitian solutions X and Y , respectively, such that the matrices A−BB ∗ X and A−Y C ∗ C have all their eigenvalues in the open right half plane, and In −XY

306

Chapter 16. Contractive rational matrix functions

is invertible (or, which amounts to the same, In − Y X is invertible). Furthermore, ¯ ∗ W (λ) admits with respect to the imaginary axis, the function V (λ) = Im − W (−λ) ∗ ¯ the right spectral factorization V (λ) = L+ (−λ) L+ (λ) with L+ and its inverse L−1 + being given by L+ (λ)

=

I + B ∗ X(In − Y X)−1 (λIn − A + Y C ∗ C)−1 B,

(16.7)

L−1 + (λ)

=

I − B ∗ X(λIn − A + BB ∗ X)−1(In − Y X)−1B.

(16.8)

Proof. By Corollary 16.5, the equation (16.6) has a Hermitian solution Y such that A − Y C ∗ C has all its eigenvalues in the open right half plane. Next, we apply Theorem 16.4 to ¯ ∗ = −B ∗ (λ + A∗ )−1 C ∗ . ! (λ) = W (−λ) W ! (λ) = −B ∗ (λ + A∗ )−1 C ∗ satisﬁes the general hypotheses of TheoNotice that W ! is strictly contractive on iR and its value at inﬁnity is rem 16.4. Furthermore, W zero. In particular, item (i) in Theorem 16.4 is satisﬁed. Hence item (iii) is satisﬁed as well, i.e., the matrix T of (16.3) has no pure-imaginary eigenvalues. ¯ ∗ W (λ) which has the realization Now consider the function V (λ) = I−W (−λ) %−1 $ 0 −A∗ C ∗ C ∗ 0 . (16.9) V (λ) = I + B λ− B 0 A

Put = A

−A∗

C ∗C

0

A

,

×

A =

−A∗

C ∗C

−B ∗ B

A

.

Since W is contractive on the imaginary axis, the function W has no pure imaginary poles. According to our assumptions the given realization of W is minimal. This implies that A has no eigenvalue on iR. But then we can use the triangular to show that the same holds true for the matrix A. Since V is positive form of A deﬁnite on the imaginary axis, we know that V (λ) is invertible for each λ ∈ iR. × has no pure imaginary eigenvalues either. But then Theorem 2.4 gives that A × are similar.) (Alternatively, this may be seen from the fact that T and A with respect to the open Let M− be the spectral subspace of the matrix A × × with respect to the left half plane, and let M+ be the spectral subspace of A open right half plane. Observe that V is positive deﬁnite on the imaginary axis and has a positive deﬁnite value at inﬁnity, namely Im . This suggests the use of × ˙ M+ Theorem 9.4 to show that C2n = M− + . For this a skew-Hermitian H must be identiﬁed with the properties required in Theorem 9.4. This can be done along × ˙ M+ the lines indicated in the proof of Theorem 13.1. So indeed C2n = M− + . The fact that Y is Hermitian and the eigenvalues of A∗ − C ∗ CY are in the open right half plane implies that σ(A − Y C ∗ C) ∩ σ(−A∗ + C ∗ CY ) = ∅. Hence Proposition ∗ 12.1 gives that the spectral subspace M− is given by M− = Im I Y .

16.4. An application to canonical factorization

307

× × by Theorem 11.5. Now M+ is an H-Lagrangian invariant subspace for A From Theorem 13.6 we see that there is a Hermitian solution X of (16.5) such ∗ × × to M × that M+ = Im X I . Moreover, A − BB ∗ X and the restriction of A + ∗ have the same eigenvalues. Thus, the eigenvalues of A − BB X are in the open × ˙ M+ right half plane. As C2n = M− + , the invertibility of I − XY follows from Lemma 12.4. Finally, we apply Theorem 12.3 to show that V admits the factorization V (λ) = V1 (λ)V2 (λ), where

V1 (λ)

=

I − B ∗ (λ + A∗ − C ∗ CY )−1 (I − XY )−1 XB ∗ ,

V2 (λ)

=

I + B ∗ X(I − Y X)−1 (λ − A + Y C ∗ C)−1 B,

V1−1 (λ)

=

I + B ∗ (I − XY )−1 (λ + A∗ − XBB ∗ )−1 XB,

V2−1 (λ)

=

I − B ∗ X(λ − A + BB ∗ X)−1 (I − Y X)−1 B.

−1 Clearly, V2 = L+ and V2−1 = L−1 + with L+ and L+ being given by (16.7) and (16.7), respectively. Furthermore, taking into account that X and Y are Hermitian,

¯ ∗ L+ (−λ)

¯ ∗ = V2 (−λ) = I + B(−λ − A∗ + CC ∗ Y )−1 (I − XY )−1 XB = V1 (λ).

¯ ∗ L+ (λ), and from the location of the eigenvalues of Thus we have V (λ) = L+ (−λ) ∗ ∗ A − Y C C and A − BB X we see that this is a right spectral factorization.

16.4 An application to canonical factorization Consider a function of the form V (λ) = Im + C(λIn − A)−1 B,

(16.10)

where W (λ) = C(λIn − A)−1 B is strictly contractive on the real line. By this we mean that the values of W on R are strict contractions, and this implies that W has no pole on the real line. Hence the latter holds true for V too. It follows also that V takes invertible values on the real line, i.e., V has no zero there. Now assume for the moment that (16.10) is a minimal realization for W . Since V has neither a pole nor a zero on the real line, the minimality of the realization implies that the matrices A and A× = A−BC have no real eigenvalues. ! (λ) = C(iλIn − A)−1 B is strictly contractive Furthermore, since the function W on the imaginary axis, we can apply Theorem 16.1(ii) to establish the existence of a Hermitian matrix X for which iXA − iA∗ X + XBB ∗ X + C ∗ C = 0.

(16.11)

308

Chapter 16. Contractive rational matrix functions

Finally, because of the minimality (see the remark in the paragraph after the proof of Theorem 16.1), such a matrix X is invertible. Summarizing, if (16.10) is a minimal realization and the matrix function W (λ) = C(λIn − A)−1 B is strictly contractive for real λ, then both A and A× have no real eigenvalues and there exists a Hermitian invertible matrix X solving (16.11). The next theorem describes canonical factorizations of a function of the form (16.10) in terms of a realization having the properties just described. Theorem 16.7. Let V (λ) = Im + C(λIn − A)−1 B be a realization of an m × m rational matrix function such that A and A× = A − BC have no real eigenvalues, and assume that there exists a Hermitian invertible X satisfying (16.11), i.e., iXA − iA∗ X + XBB ∗ X + C ∗ C = 0. Let M− and M+ be the spectral subspaces of A associated with the parts of σ(A) × lying in the open lower and open upper half plane, respectively, and let M− and × × × M+ be the spectral subspaces of A associated with the parts of σ(A ) lying in the open lower and open upper half plane, respectively. Then × ˙ M+ , Cn = M − +

× ˙ M− Cn = M+ + .

(16.12)

Moreover, V admits both a left and a right canonical factorization with respect to the real line, V (λ) = V+ (λ)V− (λ), V (λ) = V− (λ)V+ (λ), with the factors being given by V+ (λ)

= Im + C(λIn − A)−1 (In − Πl )B,

V− (λ)

= Im + CΠl (λIn − A)−1 B,

V− (λ)

= In + C(λIn − A)−1 (In − Πr )B,

V+ (λ)

= Im + CΠr (λIn − A)−1 B.

× Here Πl is the projection along M− onto M+ , and Πr is the projection along M+ × onto M− .

Proof. In view of Theorem 3.2, only (16.12) needs to be proved. We begin the veriﬁcation of (16.12) by observing that (16.11) implies 1 (XA − A∗ X) = 2i 1 XA× − (A× )∗ X = 2i

1 (XBB ∗ X + C ∗ C), 2

(16.13)

1 (iXB + C ∗ )(C − iB ∗ X). 2

(16.14)

These two identities imply that XAx, x and XA× x, x are nonnegative for all x ∈ Cn . In other words, both A and A× are X-dissipative, that is, they are

16.5. A generalization to pseudo-canonical factorization

309

dissipative in the indeﬁnite inner product given by X (cf., Section 11.3). Because of × this property, it follows that M+ and M+ are maximal X-nonnegative, while M− × and M− are maximal X-nonpositive (see Section 11.3). Using Proposition 11.1 it × × follows that dim M+ + dim M− = n and dim M− + dim M+ = n. Thus (16.12) is × = obtained via a dimension argument as soon as we have shown that M+ ∩ M− × M− ∩ M+ = {0}. × Take x ∈ M+ ∩ M− . Then Xx, x = 0, as x belongs to both an Xnonnegative subspace and an X-nonpositive subspace. Now the Cauchy-Schwartz inequality holds on M+ . Thus |XAx, x|2 ≤ XAx, AxXx, x = 0, and

|Xx, Ax|2 ≤ XAx, AxXx, x = 0.

From (16.13) we get 0 =

1 1 (XA − A∗ X)x, x = (Cx2 + B ∗ Xx2 ). 2i 2

× we have A× x = (A − BC)x = Ax. Hence Cx = 0, and so for x ∈ M+ ∩ M− × Consequently M+ ∩ M− is both A-invariant and A× -invariant. As

σ(A|M+ ∩M × ) ⊂ σ(A|M+ ) ⊂ {λ | λ > 0}, −

σ(A× |M+ ∩M × ) ⊂ σ(A× |M × ) ⊂ {λ | λ < 0}, −

−

× and A|M+ ∩M × = A |M+ ∩M × , we have that M+ ∩ M− = {0}. In a similar way one ×

−

−

× = {0}. shows that M− ∩ M+

Note that the above theorem together with the arguments given in the ﬁrst two paragraphs of this section yield the following corollary. Corollary 16.8. Let V (λ) = Im + W (λ), where W is a strictly proper rational matrix function which is strictly contractive on the real line. Then V admits both a right and a left canonical factorization with respect to the real line.

16.5 A generalization to pseudo-canonical factorization In this section the result of the previous section is generalized to pseudo-canonical factorizations. As a preparation we recall from Theorem 11.6 the following facts. Let X be an n × n invertible Hermitian matrix and let A be an n × n matrix which is X-dissipative. Then there exist A-invariant subspaces M+ and M− such that M+ is maximal X-nonnegative and M− is maximal X-nonpositive, σ(A|M+ ) ⊂ {λ | λ ≥ 0},

σ(A|M− ) ⊂ {λ | λ ≤ 0},

310

Chapter 16. Contractive rational matrix functions

M+ contains the spectral subspace of A corresponding to the eigenvalues of A in the open upper half plane, and M− contains the spectral subspace of A corresponding to the eigenvalues of A in the open lower half plane. These facts allow us to deal with rational matrix functions that are contractive on the real line. A rational matrix function W is called contractive on the real line if the values that W takes on R are contractive matrices. Such a function does not have a pole on the real line. Theorem 16.9. Let W be a strictly proper rational m × m matrix function which is contractive on the real line. Then V (λ) = Im + W (λ) admits both a right and a left pseudo-canonical factorization with respect to the real line. Such factorizations can be obtained as follows. Let W (λ) = C(λIn − A)−1 B be a minimal realization. Then there exists an invertible Hermitian matrix X satisfying iXA − iA∗ X + XBB ∗ X + C ∗ C = 0.

(16.15)

× Let M− and M− be maximal X-nonpositive subspaces that are invariant under A and A× , respectively, such that

σ(A|M− ) ⊂ {λ | λ ≤ 0},

σ(A× |M × ) ⊂ {λ | λ ≤ 0}, −

× and let M+ and M+ be maximal X-nonnegative subspaces that are invariant under × A and A , respectively, such that

σ(A|M+ ) ⊂ {λ | λ ≥ 0},

σ(A× |M × ) ⊂ {λ | λ ≥ 0}. +

× × ˙ M+ ˙ M− Then (16.12) holds, that is Cn = M− + and Cn = M+ + . Let Πl be the × projection along M− onto M+ , and put

V+ (λ)

= Im + C(λIn − A)−1 (In − Πl )B,

V− (λ)

= Im + CΠl (λIn − A)−1 B.

Then V (λ) = V+ (λ)V− (λ) is a left pseudo-canonical factorization with respect to × the real line. Write Πr for the projection along M+ onto M− , and set V− (λ)

=

Im + C(λ − A)−1 (In − Πr )B,

V+ (λ)

=

Im + CΠr (λIn − A)−1 B.

Then V (λ) = V− (λ)V+ (λ) is a right pseudo-canonical factorization with respect to the real line . Proof. By applying Theorem 16.1 (ii) to W (λ) = C(iλIn − A)−1 B we see that there is an invertible Hermitian X such that (16.15) holds. Once (16.12) is proved the rest of the theorem is a consequence of Theorem 8.5. Of the two equalities in (16.12) only the ﬁrst will be proved, the second can be established analogously.

16.5. A generalization to pseudo-canonical factorization

311

× As M+ is maximal X-nonnegative and M− is maximal X-nonpositive we × have dim M+ + dim M− = n, by Proposition 11.1. So it remains to show that × × M + ∩ M− = {0}. Take x ∈ M+ ∩ M− . As in the proof of Theorem 16.7, one shows × × that Cx = 0, and thus Ax = A x. Obviously, it follows from this that M+ ∩ M− is A-invariant and contained in Ker C. Because of the minimality, we can conclude × that M+ ∩ M− = {0}.

Note that the location of the spectra of the operators A|M− , A× |M × , A|M+ −

and A× |M × do not play a role in the proof of the identities in (16.12). Thus we + also have the following result. Proposition 16.10. Let V (λ) = Im + C(λIn − A)−1B be a minimal realization, and let X be an invertible Hermitian solution of (16.15). Let M be any A-invariant maximal X-nonnegative subspace, and let M × be any A× -invariant maximal X˙ M × . Let Π be the projection along M onto nonpositive subspace. Then Cn = M + × M , and write V1 (λ)

= Im + C(λIn − A)−1 (I − Π)B,

V2 (λ)

= Im + CΠ(λIn − A)−1 B.

Then V (λ) = V1 (λ)V2 (λ) is a minimal factorization. A similar result holds for any A-invariant maximal X-nonpositive subspace M and any A× -invariant maximal X-nonnegative subspace M × . Notice that there are various similarities between the proofs of Theorems 16.7 and 16.9 on the one hand and those of Theorems 15.3 and 15.6 on the other hand. These similarities are not surprising. In fact, the main results of the previous two sections are closely related to those in Sections 15.2 and 15.3 of the previous chapter. To see this we use the Cayley transformation −1 F (λ) = I − W (λ) I + W (λ) .

(16.16)

Here are the details. Let W be a strictly proper rational m × m matrix function, and let F be the rational m × m matrix function given by (16.16). Since W is strictly proper, I + W (λ) is biproper, and hence F is well-deﬁned. Furthermore, F is biproper and its value at inﬁnity is equal to Im . The identity ¯ ∗ −1 Im − W (λ) ¯ ∗ W (λ) Im + W (λ) −1 ¯ ∗ = 2 Im + W (λ) F (λ) + F (λ) shows that F has a nonnegative real part on R if and only if W is contractive on R. Moreover, F has a positive deﬁnite real part on R if and only if W is strictly contractive on R.

312

Chapter 16. Contractive rational matrix functions

Assume now that W is given by the realization W (λ) = C(λIn − A)−1 B. −1 −1 = 2(Im + W (λ) − Im , we Since F (λ) = 2Im − (Im + W (λ) (Im + W (λ) see that F admits the realization F (λ) = Im − 2C(λIn − A× )−1 B,

(16.17)

where, as usual, A× = A−BC. Now apply Theorem 15.1 to F using the realization (16.17). For this case equation (15.3), with X replaced by Y , has the form 1 −i(A× )∗ Y + iY A× − (Y B + i2C ∗ )(B ∗ Y − 2iC) = 0. 2

(16.18)

Using A× = A − BC and setting Y = −2X, a straightforward computation shows that (16.18) is equivalent to iXA − iA∗ X + XBB ∗ X + C ∗ C = 0,

(16.19)

and the latter equation is precisely (16.11). By applying Theorem 15.1 to F and using the equivalence between (16.18) and (16.19) we obtain the following result. Proposition 16.11. Let W (λ) = C(λIn − A)−1 B, and assume that the pair (A, B) is controllable. Then W is contractive on the real line if and only if the equation (16.19) has a Hermitian solution. Moreover, if the given realization is minimal, then any Hermitian solution of (16.19) is invertible. The above proposition provides an alternative proof of Theorem 16.1(ii) for the case when W is square and D = 0. The details involve a transformation λ → iλ (cf., the beginning of the proof of Theorem 16.9).

Notes The state space characterizations of contractive and strictly contractive rational matrix functions given in Theorems 16.1, 16.3 and 16.4 are versions of what is known as the bounded real lemma in mathematical systems theory. These results play an important role in robust and optimal control theory, see, e.g., the text books [77] and [150]. The bounded real lemma may also be found in [4] in another form. The application to spectral factorization (Section 16.3) is classical and can be found in Chapter 7 in [4]. The result that a function of the form identity plus a strict contraction admits canonical factorization (Section 16.4) is well-known; see e.g., [29] and the references given there. A surprising fact is that this property actually characterizes the circle or the line; for this see [109]. The state space results given in Sections 16.4 and 16.5 are based on [74].

Chapter 17

J-unitary rational matrix functions In this chapter realizations are used to study rational matrix functions of which the values on the imaginary axis are J-unitary matrices. Solutions to various factorization problems are given. Special attention is paid to factorization of J-unitary rational matrix functions into J-unitary factors. We also discuss the problem of embedding a contractive rational matrix function as the (1, 2) block in a unitary rational matrix function. The latter problem is related to the Darlington synthesis problem from network theory. This chapter consists of eight sections. Realization and minimal factorization of J-unitary rational matrix functions are the main topics of Sections 17.1 and 17.2. In Section 17.3 the factorization results are speciﬁed further for unitary rational matrix functions. The Redheﬀer transform, which allows one to relate J-unitary rational matrix functions to certain classes of unitary rational matrix functions, is introduced in Section 17.4. This transform is used in Section 17.5 in the study of Jinner rational matrix functions. A state space analysis of inner-outer factorization is the main topic of Section 17.6. The ﬁnal two sections deal with completion problems. Section 17.7 presents state space formulas for unitary completions of minimal degree, and Section 17.8 presents such formulas for bi-inner completions of non-square inner rational matrix functions.

17.1 Realizations of J-unitary rational matrix functions Throughout this section, J stands for an m × m signature matrix, that is, J is an invertible Hermitian matrix such that J = J −1 . An m × m matrix M is said to be J-unitary if M ∗ JM = J. Since all matrices in the latter identity are square and J is invertible, it follows that a J-unitary matrix M is invertible and M −1 = JM ∗ J.

314

Chapter 17. J-unitary rational matrix functions

If M is a J-unitary matrix, then M ∗ and M −1 are both J-unitary too. Indeed, M JM ∗ = (M ∗ J)−1 M ∗ = J −1 (M ∗ )−1 M ∗ = J −1 = J, M −1)∗ JM −1 = (JM ∗ J)∗ J(JM ∗ J) = J(M JM ∗ )J = J. In this chapter we deal with rational matrix functions of which the values on the imaginary axis are J-unitary matrices. A rational m × m matrix function W is called J-unitary on the imaginary axis if it takes J-unitary values on the imaginary axis. In other words, W is J-unitary with respect to the imaginary axis whenever W (λ)∗ JW (λ) = J, λ ∈ iR, λ not a pole of W . (17.1) Equivalently, W is J-unitary with respect to the imaginary axis if and only if ¯ ∗ JW (λ) = J, W (−λ)

¯ not a pole of W . λ, −λ

(17.2)

In the sequel we shall only consider matrix functions that are J-unitary with respect to the imaginary axis and not with respect to other contours. Therefore we shall feel free to omit the phrase “with respect to the imaginary axis.” ¯ ∗ Observe that if W is J-unitary, then both the functions W (λ)−1 and W (−λ) are J-unitary as well. Furthermore, if W1 and W2 are two J-unitary rational matrix functions, their product W1 W2 will also be J-unitary. First we shall characterize the property of being a J-unitary rational matrix function in terms of realizations. We shall assume throughout that the rational matrix functions are proper. Theorem 17.1. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a proper rational m × m matrix function. The following statements are equivalent: (i) W is J-unitary; (ii) D is J-unitary and there exists an n × n matrix H such that AH + HA∗ = BJB ∗ ,

CH = DJB ∗ ,

H = H ∗;

(17.3)

(iii) D is J-unitary and there exists an n × n matrix G such that GA + A∗ G = C ∗ JC,

GB = C ∗ JD,

G = G∗ .

(17.4)

In this case the matrices H and G are uniquely determined by the given realization, they are invertible and G = H −1 . Proof. Assume that W is J-unitary. Taking the limit in (17.1) for λ → ∞ we see that D∗ JD = J. Thus D is a J-unitary matrix. In particular, D is invertible, ¯ ∗ )−1 J = W (λ) for all λ for and hence W is biproper. By (17.2) we have J(W (−λ) ¯ which λ is not a pole of W and −λ is not a zero of W . Now one computes that ¯ −∗ J = JD−∗ J + JD−∗ B ∗ λIn − (−A∗ + C ∗ D−∗ B ∗ ) −1 C ∗ D−∗ J. JW (−λ)

17.1. Realizations of J-unitary rational matrix functions

315

The fact that the realization is minimal yields, by the state space similarity theorem, the existence of a unique (invertible) n × n matrix H such that AH = −HA∗ + HC ∗ D−∗ B ∗ ,

B = HC ∗ D−∗ J,

JD −∗ B ∗ = CH. (17.5)

Next, take adjoints and use D∗ JD = J to see that (17.5) also holds with H ∗ in place of H. By uniqueness it follows that H = H ∗ . Hence (17.3) holds, and so (i) implies (ii), even with the additional condition that H is invertible. Next assume D is J-unitary and there exists an n × n matrix H such that (17.3) holds. A straightforward computation gives ¯ ∗ = I + C(λ − A)−1 BD−1 DJD∗ I − D−∗ B ∗ (λ + A∗ )−1 C ∗ W (λ)JW (−λ) =

J + C(λ − A)−1 BJD ∗ − DJB ∗ (λ + A∗ )−1 C ∗ −C(λ − A)−1 BJB ∗ (λ + A∗ )−1 C ∗

=

J + C(λ − A)−1 HC ∗ − CH(λ + A∗ )−1 C ∗ −C(λ − A)−1 (H(λ + A∗ ) − (λ − A)H)(λ + A∗ )−1 C ∗

=

J + C(λ − A)−1 HC ∗ − CH(λ + A∗ )−1 C ∗ −C(λ − A)−1 HC ∗ + CH(λ + A∗ )−1 C ∗ = J.

Thus the function W (λ)∗ is J-unitary. But then so is W . We have now proved that (i) and (ii) are equivalent. The equivalence of (i) and (iii) can be established in the same way. Actually the implication (iii) ⇒ (i) can be obtained directly from (17.4) without having to take recourse to the function W (λ)∗ . As above, (i) implies the stronger version of (iii) with the extra requirement that G is invertible. The uniqueness and invertibility of H and G follow from the minimality. The invertibility can also be proved directly, and in fact from slightly weaker conditions. Assume (17.3) holds and that the pair (A, B) is controllable. Then H is invertible. Indeed, assume Hx = 0. Then DJB ∗ x = CHx = 0, so B ∗ x = 0. Hence (AH + HA∗ )x = 0 too. With Hx = 0, this gives HA∗ x = −AHx = 0. So Ker H ⊂ Ker B ∗ and A∗ [Ker H] ⊂ Ker H. Thus Ker H ⊂ Ker (B ∗ |A∗ ) = {0}. So H is invertible. Likewise, one shows that if (17.4) is satisﬁed and the pair (C, A) is observable, then G is invertible. Finally, let H be as in (17.3), then (17.4) holds with H −1 in place of G. By uniqueness it follows that G = H −1 . In the argument for the implication (ii) ⇒ (i) given above, the minimality of the given realization does not play a role. Similarly the minimality condition is irrelevant for the implication (iii) ⇒ (i). This is also reﬂected by the following proposition.

316

Chapter 17. J-unitary rational matrix functions

Proposition 17.2. Let W (λ) = D + C(λIn − A)−1 B be a realization of a rational m × m matrix function. Assume D is J-unitary, and let H and G be given Hermitian n × n matrices. Consider the following four statements: (i) AH + HA∗ = BJB ∗ ,

CH = DJB ∗ ;

(ii) AH + HA∗ = HC ∗ JCH,

CH = DJB ∗ ;

(iii) GA + A∗ G = C ∗ JC,

GB = C ∗ JD;

(iv) GA + A∗ G = GBJB ∗ G,

GB = C ∗ JD.

Then (i) and (ii) are equivalent, and so are (iii) and (iv). Each of (i)–(iv) implies that W is J-unitary. Moreover, if (A, B) is controllable and (i) holds, then all four statements are equivalent and the realization is minimal. Likewise, if (C, A) is observable and (iii) holds, then again all four statements are equivalent and the realization is minimal. Proof. To see the equivalence of (i) and (ii), use D ∗ JD = J to see that BJB ∗ = HC ∗ JCH. In an analogous manner one sees that (iii) and (iv) are equivalent. For the case when the realization is minimal the fact that (i) and (iii) imply that W is J-unitary is covered by Theorem 17.1. The general case is proved using the type of arguments occurring the proof of Theorem 17.1. Now suppose that (A, B) is controllable, and that (i) holds. In the proof of Theorem 17.1 we have already shown that this implies that H is invertible. Taking G = H −1 it follows that (iii) is satisﬁed, and hence also (iv). Next, we show that in this case (C, A) is observable. Indeed, by induction one shows that H −1 Ker (C|A) ⊂ Ker (B ∗ |A∗ ) = {0}. Hence the realization is minimal. The equivalence of all four statements now follows from Theorem 17.1. The reasoning for ﬁnal statement of the theorem is similar. The next proposition shows that under certain additional conditions the ﬁrst identity in (i) of Proposition 17.2 implies the second identity in (i), and analogously for (i) replaced by (iii). Proposition 17.3. Let W (λ) = D + C(λIn − A)−1 B be a realization of a J-unitary rational m × m matrix function, and let H and G be n × n Hermitian matrices. The following two statements are true: (i) If the pair (A, B) is controllable and GA + A∗ G = C ∗ JC, then GB = C ∗ JD. (ii) If the pair (C, A) is observable and AH +HA∗ = BJB ∗ , then CH = DJB ∗ . Proof. We only prove the ﬁrst part of the proposition, the second part can be ¯ ∗ JW (λ) established analogously. Assume that W is J-unitary. Computing W (−λ) one sees that this is equivalent to

−B ∗

D∗ JC

$ λ−

−A∗

C ∗ JC

0

A

%−1

C ∗ JD B

= 0.

(17.6)

17.1. Realizations of J-unitary rational matrix functions

317

Now assume that GA + A∗ G = C ∗ JC. Using I G S= 0 I as a similarity transformation in the realization (17.6), we see that (17.6) is equivalent to $ %−1 ∗ −A∗ 0 C JD − GB −B ∗ D ∗ JC − B ∗ G λ− = 0. 0 A B But this identity, in turn, is equivalent to (D ∗ JC−B ∗ G)(λ−A)−1 B = 0, λ ∈ ρ(A). The fact that (A, B) is controllable now implies that GB = C ∗ JD. The Hermitian matrix H in Theorem 17.1(ii), which is uniquely determined by the conditions stated there, will be called the Hermitian matrix associated with the minimal realization W (λ) = D + C(λIn − A)−1 B. Our next concern is how the associated Hermitian matrix behaves under similarity transformation on the realization. Proposition 17.4. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function. Write H for the Hermitian matrix associated with this realization, and let S be an invertible n × n matrix. Then the Hermitian matrix associated with the minimal realization W (λ) = D + CS −1 (λIn − SAS −1 )−1 SB

(17.7)

is given by SHS ∗ . Proof. For the (minimal) realization (17.7), the matrix SHS ∗ satisﬁes the requirements of condition (ii) in Theorem 17.1. As a consequence of the above proposition the number of positive and the number of negative eigenvalues of the matrix H do not depend on the particular choice of the minimal realization of the function W . The number of positive eigenvalues of H will be denoted by π+ (W ). At the end of this section, in Proposition 17.10, it will be seen how to express π+ (W ) completely in terms of W itself rather than in terms of the associated Hermitian matrix H. The next two propositions describe how the associated Hermitian matrix behaves under the operations of inversion, taking adjoints, and multiplication. Proposition 17.5. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function. Write H for the Hermitian matrix associated with this realization and, as usual, A× for the matrix A−BD −1 C. Then the Hermitian matrices associated with the minimal realizations W (λ)−1

=

D−1 − D−1 C(λIn − A× )−1 BD−1 ,

¯ ∗ W (−λ)

=

D∗ − B ∗ (λIn + A∗ )−1 C ∗ ,

318

Chapter 17. J-unitary rational matrix functions

are −H and −H −1 , respectively. Proof. For the ﬁrst realization, use (17.3) and observe that A× (−H) + (−H)(A× )∗

= −AH − HA∗ + BD−1 CH + HC ∗ D−∗ B ∗ = −BJB ∗ + BD−1 DJB ∗ + BJD∗ D−∗ B ∗ = BJB ∗ = (BD−1 )J(D −∗ B ∗ ),

and −D −1 C(−H) = D−1 CH = JB ∗ = D−1 J(BD−1 )∗ . The claim for the second realization is straightforward from the fact that in Theorem 17.1, the matrix G is the inverse of H. Proposition 17.6. For j = 1, 2, let Wj (λ) = Dj + Cj (λInj − Aj )−1 Bj be a minimal realization of a J-unitary rational m×m matrix function Wj having as the Hermitian matrix associated to it Hj . Suppose W = W1 , W2 is a minimal factorization. Then W is a J-unitary rational matrix function, W (λ) = D1 D2 +

D1 C2

C1

$

λIn1 +n2 −

A1

B1 C2

0

A2

%−1

B1 D 2

B2

is a minimal realization of W , and the associated Hermitian matrix is the block diagonal matrix diag (H1 , H2 ). Proof. Applying (17.3) to both realizations, using also D2 JD2∗ = J, one sees that ∗ H1 0 A1 A1 B1 C2 H1 0 0 + 0 A2 0 H2 C2∗ B1∗ A∗2 0 H2 =

B1 JB1∗

B1 C2 H2

H2 C2∗ B1∗

B2 JB2∗

B1 D2 = J D2∗ B1∗ B2

B2∗ .

So the ﬁrst equality in (17.3) is satisﬁed for the product realization. Also, H1 0 C1 D1 C2 = C1 H1 D1 C2 H2 0 H2 = D1 JB1∗

D1 D2 JB2∗ = (D1 D2 )J D2∗ B1∗

B2∗ ,

and this proves the second equality of (17.3) for the product realization.

Next, we present a few examples. As before, J stands for an m × m signature matrix.

17.1. Realizations of J-unitary rational matrix functions

319

Example 17.7. Let R be an m × m matrix such that R∗ JR = JR, and let ω ∈ / iR. Then the rational m × m matrix function W given by λ−ω R λ+ω ¯ is J-unitary. To be more speciﬁc, let u be a vector in Cm such that u∗ Ju = Ju, u = 0, and take for R the rank 1 matrix W (λ) = Im − R +

R=

1 Juu∗ . u∗ Ju

Then R∗ JR = JR = R∗ J = (u∗ Ju)−1 uu∗ . (Note here that uu∗ is a rank 1 matrix, while u∗ Ju is just a scalar.) A minimal realization for W for this particular choice of R may be obtained by setting (ω + ω ¯) . ∗ u Ju The associated Hermitian matrix satisﬁes AH + HA∗ = BJB ∗ , which in this case becomes −(ω + ω ¯ )H = u∗ Ju. So H = −(u∗ Ju)(2ω)−1 . Example 17.8. Let α ∈ iR, n ∈ N, and let x ∈ Cm be a J-neutral vector, i.e., x∗ Jx = 0. Then i W (λ) = Im + Jxx∗ (λ − α)2n is J-unitary. A minimal realization for W can be obtained by setting A = J2n (α), the Jordan block of size 2n with eigenvalue α, and ⎡ ⎤ 0 ⎢ .. ⎥ ⎢ ⎥ C = i Jx 0 · · · 0 , B = ⎢ . ⎥, ⎣0⎦ x∗ A = −¯ ω,

B = u∗ ,

C=−

where C is an m × 2n matrix and B is a 2n × m matrix. The associated Hermitian matrix can be computed to be the following matrix: 0 if p + q = 2n + 1, 2n H = [hp q ]p,q=1 , hp q = (−1)q i if p + q = 2n + 1. We conclude this section with a few remarks on matrix-valued kernel functions and their state space representations. Introduce the functions KW (λ, μ)

=

J − W (λ)JW (μ)∗ , λ+μ ¯

K∗,W (μ, λ)

=

J − W (μ)∗ JW (λ) . λ+μ ¯

Here W is a rational m × m matrix function. Furthermore, λ and μ are complex numbers, not poles of W , λ = −μ.

320

Chapter 17. J-unitary rational matrix functions

Lemma 17.9. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational matrix function having H as its associated Hermitian matrix. Then the following two identities hold: KW (λ, μ)

=

−C(λ − A)−1 H −1 (¯ μ − A∗ )−1 C ∗ ,

(17.8)

K∗,W (μ, λ)

=

−B ∗ (¯ μ − A∗ )−1 H −1 (λ − A)−1 B.

(17.9)

Proof. We shall only prove (17.9); identity (17.8) can be obtained in an analogous fashion. First note that W (μ)∗ JW (λ) = D∗ + B ∗ (¯ μ − A∗ )−1 C ∗ J D + C(λ − A)−1 B μ − A∗ )−1 C ∗ JD + D ∗ JC(λ − A)−1 B = D∗ JD + B ∗ (¯ μ − A∗ )−1 C ∗ JC(λ − A)−1 B. + B ∗ (¯ Now use the identities D∗ JD = J, C ∗ JD = HB and C ∗ JC = H −1 A + A∗ H −1 which hold by Theorem 17.1. Then one sees that ¯)B ∗ (¯ μ − A∗ )−1 H −1 (λ − A)−1 B. W (μ)∗ JW (λ) = J + (λ + μ From this (17.9) is immediate.

The kernel function KW (λ, μ) is said to have κ negative squares if for each r ∈ N and any collection of points ω1 , . . . , ωr in the complex plane, not poles of W , and any collection of vectors u1 , . . . , ur in Cm the r × r Hermitian matrix ∗ r (17.10) uj KW (ωj , ωi )ui i,j=1 has at most κ negative eigenvalues, and it has exactly κ negative eigenvalues for at least one choice of r, ω1 , . . . , ωr and u1 , . . . , ur . For K∗,W (μ, λ), the deﬁnition is of course similar. Proposition 17.10. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Then the number of negative squares of each of the functions KW and K∗,W is equal to π+ (W ), the number of positive eigenvalues of the matrix H. This result corroborates the already established fact that the integer π+ (W ) is independent of the particular minimal realization of W (cf., the paragraph after the proof of Proposition 17.4). Proof. It follows from the previous lemma that K∗,W has at most π+ (W ) negative squares. Indeed, if ω1 , . . . , ωr is a collection of points in the complex plane, not poles of W , and u1 , . . . , ur is a collection of vectors in Cm , then the r×r Hermitian matrix (17.10) can be written in the form −E ∗ H −1 E, where H is the Hermitian matrix associated with the given realization of W .

17.2. Factorization of J-unitary rational matrix functions

321

Next, consider M = span {(λ − A)−1 Bu | u ∈ Cm , λ ∈ C not an eigenvalue of A}. Clearly, for u ∈ Cm and λ not an eigenvalue of A, the vector λ(λ−A)−1 Bu belongs to M . Since M is closed in Cn , this implies that Bu = lim λ(λ − A)−1 Bu ∈ M, λ→∞

u ∈ Cm .

Thus Im B ⊂ M . Next, note that A(λ − A)−1 Bu = −Bu + λ(λ − A)−1 Bu ∈ M . Hence M is invariant under A. But then Im (A|B) ⊂ M . By hypothesis, the given realization of W is minimal. This implies that Im (A|B) = Cn . We conclude that M = Cn . The latter implies that Cn has a basis x1 , . . . , xn such that for each j the vector xj is of the form xj = (λi − A)−1 Buj for some vector uj ∈ Cm and some ωj ∈ C. Consider the n × n matrix X = [ x1 · · · xn ]. We obtain that for these uj and ωi we have n ∗ uj K∗,W (ωj , ωi )ui i,j=1 = −X ∗ H −1 X. As X is invertible, this matrix has exactly π+ (W ) negative eigenvalues. This settles the matter for K∗,W ; for KW the argument is similar.

17.2 Factorization of J-unitary rational matrix functions In this section minimal factorizations of J-unitary rational matrix functions into a product of two J-unitary rational matrix functions will be studied. Here, as in the previous section, J is an m × m signature matrix. To state the main theorem we need to recall a notion introduced in Section 11.1. Let H = H ∗ be an invertible n×n matrix. A subspace M ⊂ Cn is called H-nondegenerate if M ∩ [HM ]⊥ = {0}. ˙ [HM ]⊥ = Cn , as a simple dimension count shows. For such a subspace one has M + Also note that (HM )⊥ = H −1 [M ⊥ ]. Theorem 17.11. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Let M be an A-invariant H −1 -nondegenerate subspace, and denote by Π the projection of Cn onto H[M ⊥ ] along M . Let D = D1 D2 be a factorization of D into two J-unitary constant matrices, and put W1 (λ)

=

D1 + C(λIn − A)−1 (I − Π)BD2−1 ,

W2 (λ)

=

D2 + D1−1 CΠ(λIn − A)−1 B.

Then W = W1 W2 , this factorization is minimal, and the factors W1 and W2 are J-unitary. Conversely, any minimal factorization W = W1 W2 with J-unitary

322

Chapter 17. J-unitary rational matrix functions

factors W1 and W2 is obtained in this way. Moreover, given a ﬁxed factorization D = D1 D2 , the correspondence between minimal factorizations of W with two J-unitary factors and H-nondegenerate invariant subspaces of A is one-to-one. Proof. From (17.3) we know that A× = −HA∗ H −1 , where A× = A − BD−1 C. It follows that H[M ⊥ [ is A× -invariant because M is A-invariant. Since the subspace M is H −1 -nondegenerate, the projection Π is a supporting projection. Hence the factorization W = W1 W2 is a minimal one. To complete the proof of the ﬁrst part of the theorem it remains to show that the factors W1 and W2 are J-unitary rational matrix functions. In fact, it suﬃces to show that one of them is J-unitary, the J-unitarity of the other one then follows automatically. Since Π is a supporting projection we know that a minimal realization, W1 (λ) = D1 + C1 (λ − A1 )−1 B1 , of W1 is obtained by taking A1

=

τM ∗ AτM : M → M,

B1

=

τM ∗ (I − Π)BD2−1 : Cm → M,

C1

=

CτM : M → Cm .

Here τM is the canonical embedding of M into Cn , and hence τM ∗ τM is the orthogonal projection of Cn onto M . Put G1 = τM ∗ H −1 τM . Then G1 is invertible. Indeed, suppose G1 x = 0 for some x ∈ M . Then H −1 x ∈ Ker τM ∗ = M ⊥ , i.e., x ∈ H(M ⊥ ). So x ∈ M ∩ H(M ⊥ ) = {0}. Next, we shall show that the conditions of Theorem 17.1 (iii) are satisﬁed. First, note that (G1 A1 + A∗1 G1 ) =

τM ∗ H −1 τM τM ∗ AτM + τM ∗ AτM τM ∗ H −1 τM

=

τM ∗ (H −1 A + A∗ H −1 )τM

=

τM ∗ C ∗ JCτM ∗ = C1∗ JC1 .

Furthermore, we have G1 B1 = τM ∗ H −1 τM τM ∗ (I − Π)BD2−1 = τM ∗ H −1 (I − Π)BD2−1 . Now, as M is H −1 -nondegenerate, Im H −1 Π = M ⊥ and H −1 [M ]⊥ = H[M ⊥ ] = Im Π. This yields H −1 Πx, y = H −1 Πx, Πy = Πx, H −1 Πy = x, H −1 Πy. Hence Π∗ H −1 = H −1 Π, that is, the projection Π is H −1 -selfadjoint. Therefore H −1 (I − Π) = (I − Π∗ )H −1 . Moreover, as (I − Π)τM = τM we have the identity τM ∗ (I − Π∗ ) = τM ∗ . Thus G1 B1

=

τM ∗ H −1 (I − Π)BD2−1

=

τM ∗ (I − Π∗ )H −1 BD2−1

=

τM ∗ H −1 BD2−1 = τM ∗ C ∗ JD1 = C1∗ JD1 .

17.2. Factorization of J-unitary rational matrix functions

323

Hence the conditions of Theorem 17.1 (iii) are satisﬁed, and thus W1 is J-unitary. The converse statement is a direct consequence of Proposition 17.6 and Theorem 8.5. As a special case of the preceding theorem we state the following proposition concerning the case where one of the factors is of degree 1. Proposition 17.12. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Suppose x is an eigenvector of A corresponding to the eigenvalue ω of A, and assume H −1 x, x = 0. Then W admits a minimal factorization W = W1 W2 into two J-unitary factors where the factor W1 is given by 1 W1 (λ) = Im + Cxx∗ C ∗ J . (17.11) (λ − ω)H −1 x, x Furthermore, in case ω ∈ / iR the scalar x∗ C ∗ JCx is non-zero and λ+ω ¯ 1 1− Cxx∗ C ∗ J. W1 (λ) = Im − ∗ ∗ x C JCx λ−ω

(17.12)

Observe that the factor W1 is of the form as given in Example 17.7 Proof. As H −1 x, x = 0, the subspace M = span {x} is H −1 -nondegenerate. Therefore we can apply the previous theorem. The projection I − Π is given by (I − Π)v =

H −1 v, x x∗ H −1 v x = x. H −1 x, x x∗ H −1 x

Taking D1 = I and D2 = D one obtains W1 (λ)

= I + C(λ − A|M )−1 (I − Π)BD−1 = I+

Cxx∗ H −1 BD−1 (λ − ω)H −1 x, x

= I+

Cxx∗ C ∗ J . (λ − ω)H −1 x, x

This proves (17.11). Next we apply (17.4) in the present setting. Recall that G = H −1 . It follows that x∗ C ∗ JCx = (ω + ω ¯ )x∗ H −1 x. Thus, when ω ∈ / iR or, equivalently, ω + ω ¯ = 0, x∗ C ∗ JCx = 0,

H −1 x, x =

x∗ C ∗ JCx . ω+ω ¯

Employing this in (17.11) immediately yields (17.12).

324

Chapter 17. J-unitary rational matrix functions

17.3 Factorization of unitary rational matrix functions In this section we shall consider the special case of rational matrix functions that are unitary on the imaginary axis, that is, we continue the theme of the previous section with J = I. For simplicity, we call such functions unitary rational matrix functions and omit the additional qualiﬁer “on the imaginary axis.” Let W be a unitary rational matrix function. Then W is bounded by 1 on the imaginary axis, and hence W cannot have pure imaginary poles. Since W −1 is also a unitary rational matrix function, W cannot have pure imaginary zeros either. Replacing λ by λ−1 one also sees that W has to be biproper. Lemma 17.13. Let W (λ) = D + C(λ− A)−1 B be a minimal realization of a unitary rational m×m matrix function, and let H be the Hermitian matrix associated with this realization. Then A has no pure imaginary eigenvalues. Let P be the spectral projection of A corresponding to the part of σ(A) lying in the open right half plane. Then Im P is maximal H −1 -positive and Ker P is maximal H −1 -negative. Proof. Since the realization is minimal and W has no poles on the imaginary axis, the matrix A has no pure imaginary eigenvalues. By Theorem 17.1 with G = H −1 we have GA + A∗ G = C ∗ C. Because of the minimality of the realization we also know that the pair (C, A) is observable. Let us denote by ν(G) the number of negative eigenvalues of G, and by π(G) the number of positive eigenvalues of G. By a well-known inertia theorem (see Theorem 13.1.4 in [107]) we have ν(G) = dim Ker P and π(G) = dim Im P . Now put M = Im P , let τM be the canonical embedding of M into Cn , and introduce AM = τM ∗ AτM , GM = τM ∗ GτM and CM = CτM . Then GM is Hermitian, and (using the fact that M is invariant under A) we have GM AM + A∗M GM

=

τM ∗ GτM τM ∗ AτM + τM ∗ A∗ τM τM ∗ GτM

=

∗ τM ∗ (GA + A∗ G)τM = τM ∗ C ∗ CτM = CM CM .

The invariance of M under A also implies that Ker (CM |AM ) ⊂ Ker (C|A), and hence (CM , AM ) is an observable pair too. Moreover, AM has only eigenvalues in the open right half plane. The inertia theorem referred to above then gives that GM is positive deﬁnite. But this is equivalent to saying that Im P is H −1 -positive. As π(H −1 ) = dim Im P , it is actually maximal H −1 -positive. The other part of the proposition is proved in a similar way. Observe that an H −1 -positive subspace is in particular H −1 -nondegenerate . Likewise, an H −1 -negative subspace is H −1 -nondegenerate. So we are in a position to apply Theorem 17.11. This yields the following two results of which we shall only prove the second. Theorem 17.14. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a unitary rational m × m matrix function (so, in particular, D is invertible), and

17.3. Factorization of unitary rational matrix functions

325

let A× = A − BD−1 C be the associate main operator. Then W admits a minimal factorization W = W1 W2 having the following additional properties: (i) W1 has its poles in the left half plane and its zeros in the right half plane, (ii) W2 has its poles in the right half plane and its zeros in the left half plane, (iii) δ(W1 ) = n − π+ (W ) and δ(W2 ) = π+ (W ). Such a factorization can be obtained as follows. Let P denote the spectral projection corresponding to the part of σ(A) lying in the open left half plane, and write P × for the spectral projection of A× corresponding to the part of σ(A× ) lying in the ˙ Ker P × and the functions open right half plane. Then Cn = Im P + W1 (λ)

=

Im + C(λIn − A)−1 (In − Π)BD−1 ,

W2 (λ)

=

D + CΠ(λIn − A)−1 B,

(17.13)

meet the requirements. Here Π is the projection of Cn along Im P onto Ker P × . Theorem 17.15. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a unitary rational m × m matrix function (so, in particular, D is invertible), and let A× = A − BD−1 C be the associate main operator. Then W admits a minimal factorization W = W1 W2 having the following additional properties: (i) W1 has its poles in the right half plane and its zeros in the left half plane, (ii) W2 has its poles in the left half plane and its zeros in the right half plane, (iii) δ(W1 ) = π+ (W ) and δ(W2 ) = n − π+ (W ). Such a factorization can be obtained as follows. Let P denote the spectral projection corresponding to the part of σ(A) lying in the open right half plane, and write P × for the spectral projection of A× corresponding to the part of σ(A× ) lying in the ˙ Ker P × and the functions open left half plane. Then Cn = Im P + W1 (λ)

=

Im + C(λIn − A)−1 (In − Π)BD−1 ,

W2 (λ)

=

D + CΠ(λIn − A)−1 B,

(17.14)

meet the requirements. Here Π is the projection of Cn along Im P onto Ker P × . Proof. With Lemma 17.13 in mind, the idea is to apply Theorem 17.11 taking M = Im P . We need to ﬁnd H[Im P ]⊥ = Ker (P ∗ H −1 ). From (17.3) we know that A× = −HA∗ H −1 , hence P × = −HP ∗ H −1 . It follows that H[Im P ]⊥ = Im P × . Let Π be the projection along Im P onto Im P × . Then, by Theorem 17.11, the function W admits the factorization W = W1 W2 , where W1 and W2 are given by (17.14), and these factors are unitary. Moreover, the factorization is minimal. Finally, the poles of W1 are the eigenvalues of A|Im P (counting multiplicities), its zeros are the eigenvalues of A× |Im P × (counting multiplicities too). Similarly, the poles of W2 are the eigenvalues of A|Ker P , while the

326

Chapter 17. J-unitary rational matrix functions

zeros of W2 are the eigenvalues of A× |Ker P × . So the position of poles and zeros of W1 and W2 is as required. It also follows that δ(W1 ) = dim Ker Π = dim Im P = π+ (W ), and hence by minimality also δ(W2 ) = n − π+ (W ).

Our next theorem is on complete factorization of a unitary rational matrix function into unitary factors (cf., Part III in [20]). Theorem 17.16. Let W be a unitary rational m × m matrix function of McMillan degree n. Then W admits a minimal factorization into n factors of McMillan degree 1. Moreover, each of these factors can be taken to be unitary. In order to prove this theorem we ﬁrst show that a unitary rational matrix function allows for a realization with very special properties. Lemma 17.17. Let W be a unitary rational m × m matrix function with W (∞) = Im . Then W admits a minimal realization W (λ) = Im + C(λIn −A)−1 B such that A11 A12 A= , (17.15) 0 A22 where A11 and A22 are upper triangular, A11 has all its eigenvalues in the open right half plane, A22 has all its eigenvalues in the open left half plane, and the Hermitian matrix associated with the realization is given by In 0 H= . (17.16) 0 −In Proof. Take an arbitrary minimal realization W (λ) = I +C(λ−A)−1 B. By Schur’s theorem there is an orthogonal change of basis such that A is upper triangular. In fact, we may take the eigenvalues of A on the diagonal in any order we like. This is known as the ordered Schur form of A. We apply this to construct a similarity transformation such that A is of the form A11 A12 A= 0 A22 where A11 is upper triangular having all its eigenvalues in the open right half plane, and A22 is upper triangular having all its eigenvalues in the open left half plane. The spectral projection of A corresponding to its eigenvalues in the open right half plane is given by I 0 P = . 0 0

17.3. Factorization of unitary rational matrix functions

327

Let H be the Hermitian matrix associated with this realization, and let G be its inverse. Decompose G in the same way as A, and write G11 G12 . G= G∗12 G22 Because of Lemma 17.13 we have that Im P is maximal G-positive, and so G11 is positive deﬁnite. Likewise, since Ker P is maximal G-negative, G22 is negative deﬁnite. Next, we employ the Schur complement of G11 in G. So we factorize G as 0 I 0 G11 I G−1 11 G12 . G= G∗12 I 0 I 0 G22 − G∗12 G−1 11 G12 Since G11 is positive deﬁnite and G22 is negative deﬁnite, the Schur complement G22 − G∗12 G−1 11 G12 is negative deﬁnite too. ∗ Now take the Cholesky decomposition of G11 , that is, write G11 = C11 C11 with C11 upper triangular. Likewise, take the Cholesky decomposition of the Schur ∗ complement. Thus G22 − G∗12 G−1 11 G12 = −C22 C22 with C22 upper triangular. Put −1 −1 C11 −G−1 11 G12 C22 . S= −1 0 C22 Then, using Proposition 17.4, one checks that the realization W (λ) = I + CS(λ − S −1 AS)−1 S −1 B has all the desired properties.

Proof of Theorem 17.16. Without loss of generality we may assume that W has the value Im at inﬁnity. Let W (λ) = Im + C(λIn − A)−1 B be a minimal realization as in the previous lemma, and let H be the Hermitian matrix associated with this realization. In particular, A is upper triangular. For this realization we have by (17.3) that A× = −HA∗ H −1 . This is clearly a lower triangular matrix. Now let e1 , . . . , en be the standard basis of Cn . For k = 1, . . . , n, deﬁne Πk to be the orthogonal projection of Cn onto span {ek }. Then for j = 1, . . . , n − 1 the projection Πj+1 + · · · + Πn is a supporting projection for the minimal realization W (λ) = Im + C(λIn − A)−1 B. It then follows from Theorem 10.5 in [20] that W admits a factorization into n factors of degree 1. It remains to prove that each of the factors is unitary. Clearly, for each integer j = 1, . . . , n − 1 the image and kernel of Πj+1 + · · · + Πn are both H −1 -nondegenerate and are each other’s H-orthogonal complements. From Theorem 17.11 it then follows that for each j the products W1 · · · Wj and Wj+1 · · · Wn are unitary. From this one concludes that each Wj separately is unitary.

328

Chapter 17. J-unitary rational matrix functions

17.4 Intermezzo on the Redheﬀer transformation In this section we study the Redheﬀer transform of a J-unitary rational matrix function. This will allow us to relate J-unitary rational matrix functions to certain classes of unitary rational matrix functions. The results obtained will be used in the next section. All the time, J will be a signature matrix. The starting point of our considerations is a 2 × 2 block matrix M11 M12 M= , (17.17) M21 M22 with M11 a p×p matrix and M22 a q ×q matrix. When M22 is an invertible matrix, the Redheﬀer transform Λ of M is deﬁned as follows: −1 −1 M21 M12 M22 Λ11 Λ12 M11 − M12 M22 Λ= = . (17.18) −1 −1 Λ21 Λ22 −M22 M21 M22 We refer to the map M → Λ as the Redheﬀer transformation. Let J = diag (Ip , −Iq ). The matrix M in (17.17) is said to be J-contractive if M ∗ JM ≤ J. The next lemma shows that for such a matrix the requirement that M22 is invertible is automatically fulﬁlled. Hence the Redheﬀer transform of a J-contractive matrix M with J = diag (Ip , −Iq ) is well-deﬁned. Lemma 17.18. Let J = diag (Ip , −Iq ). If the matrix M in (17.17) is J-contractive, then M22 is invertible, the (well-deﬁned) Redheﬀer transform Λ of M is a con−1 traction, and M22 M21 < 1. Conversely, if M22 is invertible and the Redheﬀer transform Λ of M is a contraction, then M is J-contractive. Proof. Assume that the matrix M is J-contractive. By considering the (2, 2)-entry of M ∗ JM and using M ∗ JM ≤ J, we see that ∗ ∗ M22 M22 ≥ Iq + M12 M12 .

(17.19)

∗ M22 is positive deﬁnite, and hence, because M22 is square, the matrix Thus M22 −∗ M22 is invertible. Multiplying the inequality (17.19) from the left by M22 and from −1 −∗ −1 −∗ −1 −∗ −1 ∗ the right by M22 , we get Iq − M22 M12 M12 M22 ≥ M22 M22 . Since M22 M22 −∗ −1 ∗ is positive deﬁnite, we may conclude that so is Iq − M22 M12 M12 M22 . But this is −1 equivalent to M22 M21 < 1. Next assume that M22 is invertible and consider the equations x u M11 M12 = . (17.20) y v M21 M22

Then, as M22 is invertible, these equations are equivalent to Λ11 Λ12 x u = . Λ21 Λ22 v y

(17.21)

17.4. Intermezzo on the Redheﬀer transformation

329

Indeed, rewrite (17.20) as M11 x + M12 y = u and M21 x + M22 y = v. Solving for y in the second of these equations, one gets −1 −1 M21 x + M22 v. y = −M22

(17.22)

Inserting this in the ﬁrst of the two equations above, we obtain −1 −1 u = (M11 − M12 M22 M21 )x + M12 M22 v.

(17.23)

Together, (17.22) and (17.23) prove the desired equivalence between (17.20) and (17.21). Notice that the condition that the matrix M is J-contractive is equivalent to the inequality u2 − v2 ≤ x2 − y2 . Indeed, M ∗ JM ≤ J is equivalent to ( ) ( ) ( ) ( ) x x u u 2 2 u − v = J (17.24) ,M = JM , y y v v = M ∗ JM

( ) ( ) ( ) ( ) x x x x = x2 − y2 . , ≤ J , y y y y

Similarly, the condition that the Redheﬀer transform Λ is a contraction is equivalent to u2 + y2 ≤ x2 + v2 . But u2 − v2 ≤ x2 − y2 ⇐⇒ u2 + y2 ≤ x2 + v2 . Thus, as desired, M is J-contractive amounts to the same as M22 is invertible and Λ is a contraction. Corollary 17.19. Let J = diag (Ip , −Iq ), and assume that the matrix M in (17.17) is J-contractive. Then M ∗ is J-contractive too. Proof. By Lemma 17.18, the fact that M is J-contractive implies that M22 is invertible and the Redheﬀer transform Λ of M is a contraction. Since M22 is ∗ invertible, so is M22 . Thus the Redheﬀer transform of M ∗ is well-deﬁned. Moreover, the Redheﬀer transform of M ∗ is equal to Λ∗ . As Λ is a contraction, the same holds true for Λ∗ . But then the converse part of Lemma 17.18 shows that M ∗ is J-contractive too. Proposition 17.20. Let J = diag (Ip , −Iq ). The matrix M in (17.17) is J-unitary if and only if M22 is invertible and the Redheﬀer transform of M is unitary. Proof. Since a J-unitary matrix is J-contractive and a unitary matrix is a contraction, we see from Lemma 17.18 that without loss of generality we may assume that the matrix M22 is invertible. This allows us to use the equivalence of the equations (17.20) and (17.21). Next, using a calculation as in (17.24), one sees that M is J-contractive if and only if the equality x2 − y2 = u2 − v2 holds. Furthermore, the condition that Λ is unitary is equivalent to x2 + v2 = u2 + y2 . But x2 − y2 = u2 − v2 ⇐⇒ x2 + v2 = u2 + y2 .

330

Chapter 17. J-unitary rational matrix functions

Hence M is J-unitary if and only if Λ is unitary.

Next we pass from matrices to matrix functions. Consider a rational matrix function W, W11 (λ) W12 (λ) W (λ) = , (17.25) W21 (λ) W22 (λ) with W11 a p×p rational matrix function and W22 a q ×q rational matrix function. −1 Assume W22 to be regular, i.e., det W22 (λ) ≡ 0. Thus W22 is a well-deﬁned rational matrix function. Under these assumptions the Redheﬀer transform of W is deﬁned to be the rational matrix function Σ given by Σ11 (λ) Σ12 (λ) Σ(λ) = (17.26) Σ21 (λ) Σ22 (λ) =

W11 (λ) − W12 (λ)W22 (λ)−1 W21 (λ)

W12 (λ)W22 (λ)−1

−W22 (λ)−1 W21 (λ)

W22 (λ)−1

.

As before, let J = diag (Ip , −Iq ). If the rational matrix function W is J-unitary with respect to the imaginary axis, then we know from Proposition 17.20 that the Redheﬀer transform Σ is unitary. In particular, it has no pure imaginary poles and zeros (see the second paragraph of Section 17.3). The following theorem is the main result of this section. Theorem 17.21. Let W be a rational matrix function, and let C1 D1 0 + (λIn − A)−1 B1 B2 W (λ) = C2 0 D2

(17.27)

−1 be a realization of W . Assume D2 is invertible, and put A× 2 = A − B2 D2 C2 . Then the Redheﬀer transform Σ of W has the realization C1 0 D1 −1 + (λIn − A× Σ(λ) = B1 B2 D2−1 , (17.28) 2) −1 −1 −D2 C2 0 D2

and this realization is minimal if and only if so is the realization (17.28). Moreover, assuming both realizations (17.27) and (17.28) to be minimal, the following holds. Let J = diag (Ip , −Iq ) and suppose W is J-unitary on the imaginary axis. If HW and HΣ denote the Hermitian matrices associated with the realizations (17.27) and (17.28), respectively, then HW = HΣ . Proof. Write W in the form (17.25). From Theorem 2.4 we have −1 B2 D2−1 , W22 (λ)−1 = D2−1 − D2−1 C2 (λ − A× 2)

17.4. Intermezzo on the Redheﬀer transformation

331

and with the help of this expression one computes −1 B2 D2−1 W12 (λ)W22 (λ)−1 = C1 (λ − A)−1 B2 D2−1 − D2−1 C2 (λ − A× 2 ) = W22 (λ)−1 W21 (λ)

=

−1 C1 (λ − A× B2 D2−1 , 2)

−1 D2−1 − D2−1 C2 (λ − A× B2 D2−1 C2 (λ − A)−1 B1 2)

−1 D2−1 C2 (λ − A× B1 . 2) Now W12 (λ)W22 (λ)−1 W21 (λ) = C1 (λ − A)−1 B2 W22 (λ)−1 W21 (λ), and hence

=

W11 (λ) − W12 (λ)W22 (λ)−1 W21 (λ)

−1 = D1 + C1 (λ − A)−1 B1 − C1 (λ − A)−1 B2 D2−1 C2 (λ − A× B1 2)

× −1 = D1 + C1 (λ − A)−1 B1 − C1 (λ − A)−1 (A − A× B1 2 )(λ − A ) −1 = D1 + C1 (λ − A× B1 . 2)

This proves (17.28). Next we deal with minimality. Assume the realization (17.27) is minimal. To prove the minimality of the realization (17.28), assume the realization (17.28) is not observable. Then ( ) C1 × Ker , A = {0}. 2 −D2−1 C2 Observe that the subspace on the left-hand side is invariant under A× 2 . Hence there exists an eigenvalue λ0 of A× and there is a non-zero vector x such that 2 −1 × A× x = λ x, and C x = 0, −D C x = 0. By the deﬁnition of A this implies 0 1 2 2 2 2 −1 × that Ax = A× 2 x − B2 D2 C2 x = A2 x = λ0 x. So ( ) C1 , A = {0}. Ker C2 Hence the realization (17.27) is not observable, which is a contradiction. It follows that the realization (17.28) is observable. A similar argument proves that the realization (17.28) is controllable. The reverse implication, minimality of (17.28) implies minimality of (17.27), is proved in an analogous way. Now assume both realizations are minimal. It remains to prove the equality of the corresponding Hermitian matrices. This is seen as follows. According to Theorem 17.1 the matrix HW is uniquely determined by the four expressions D1∗ D1 = Ip , D2∗ D2 = Iq and ∗ B C D 1 1 1 HW = . AHW + HW A∗ = B1 B1∗ − B2 B2∗ , C2 −D2 B2∗

332

Chapter 17. J-unitary rational matrix functions

Next, using the same theorem with Ip+q as the signature matrix, we know that HΣ is uniquely determined by the identities D1∗ D1 = Ip , D2∗ D2 = Iq and × ∗ A× 2 HΣ + HΣ (A2 )

C1 −D2−1 C2

=

B1 B1∗ + B2 D2−∗ D2−1 B2∗ ,

HΣ

=

(17.29)

D1 B1∗

.

D2−1 D2−∗ B2∗

∗ −1 Since D2∗ D2 = Iq and A× 2 = A − B2 C2 = A + B2 B2 HW , we obtain that the formulas for HΣ are satisﬁed by HW . Uniqueness of the associated Hermitian matrix proves then that HW = HΣ .

We ﬁnish this section by returning to the examples of Section 17.1. Consider, ∗ for J = diag (Ip , −Iq ), the function W of Example 17.7. So, taking u = u∗1 u∗2 , W (λ) = Ip+q

( ) −u1 + (λ + ω ¯ )−1 u∗1 u2

u∗2

2ω . u∗ Ju

Using Theorem 17.21 one ﬁnds, for Redheﬀer transform Σ of W , Σ(λ) = Ip+q −

1 2ω uu∗ , λ − α u∗ Ju

where

−¯ ωu1 2 − ωu2 2 2ω 2 = u . 2 u∗ Ju u∗ Ju For the Example 17.8, things are somewhat ∗ more complicated. We use the realization presented there, writing x = x∗1 x∗2 . The Redheﬀer transform of α = −¯ ω−

W (λ) = Ip+q

x1 i x∗1 + (λ − α)2n −x2

x∗2

then becomes ⎡

Σ(λ) = Ip+q + i x

where

0 ···

⎡ ⎢ ⎢ A× 2 = J2n (α) + ⎢ ⎣

0 .. . 0 ix2 2

⎤ 0 ⎢.⎥ −1 ⎢ .. ⎥ 0 (λ − A× ) ⎢ ⎥, 2 ⎣0⎦ x∗ 0

···

0

···

⎤ 0 .. ⎥ .⎥ ⎥. 0⎦ 0

17.5. J-inner rational matrix functions

333

−1 Since Σ only involves the entry of (λ − A× in the upper right corner, this 2) can be computed further. The entry in question is just 1 over the characteristic polynomial of A× 2 , and so

Σ(λ) = Ip+q +

i (λ −

α)2n

− ix2 2

xx∗ .

17.5 J-inner rational matrix functions A matrix M is called a J-contraction if M ∗ JM ≤ J. A rational matrix function W is called J-inner if W is J-unitary on the imaginary axis and, in addition, W (λ) is a J-contraction for λ in the open right half plane, λ not a pole of W . Note that we restrict the attention here to functions that are J-inner relative to the imaginary axis. If W is J-inner with J = I, then W is called bi-inner or two-sided inner (cf., Section 17.6 below). Clearly, if W is bi-inner it cannot have poles in the right open half plane. Also, if a unitary rational matrix W is analytic on the right half plane, then by the maximum modulus theorem W (λ) ≤ 1 for λ > 0, i.e., W is bi-inner. Thus a unitary rational matrix function W is bi-inner if and only if it is analytic on the right half plane. Recall from the second paragraph of Section 17.3 that a unitary rational matrix function has no pure imaginary poles or zeros, and that it is biproper. The next theorem characterizes the property of being J-inner in terms of a minimal realization. Theorem 17.22. Let W (λ) = D + C(λIn − A)−1 B be a minimal realization of a J-unitary rational m × m matrix function, and let H be the Hermitian matrix associated with this realization. Then W is J-inner if and only if H is negative deﬁnite. First we state a result that is of independent interest, and which proves one direction of Theorem 17.22. Proposition 17.23. If W is a J-inner rational matrix function, where the signature matrix J has the form J = diag (Ip , −Iq ), then its Redheﬀer transform Σ is biinner. If, in addition, W is given by the minimal realization (17.27), then A× 2 = A − B2 D2−1 C2 has all its eigenvalues in the open left half plane, and the Hermitian matrix H associated with (17.27) is negative deﬁnite. Proof. The ﬁrst part of the proposition can be derived from Proposition 17.20 and Lemma 17.18. For the second part, consider a minimal realization of W written in the form (17.27) with the partitioning induced by J = diag (Ip , −Iq ). Then we also have a minimal realization (17.28) of Σ. Since Σ is bi-inner, it is analytic in the right half plane, and by minimality of the realization this shows that A× 2 has all its eigenvalues in the left half plane.

334

Chapter 17. J-unitary rational matrix functions

It follows from the fact that H satisﬁes the Lyapunov equation (17.29) and from minimality that H is negative deﬁnite (see Corollary 1 in Section 13.1 in [107]). Proof of Theorem 17.22. Assume H is negative deﬁnite. For λ > 0 we then have ¯ − A∗ )−1 C ∗ J D + C(λ − A)−1 B J − W (λ)∗ JW (λ) = J − D ∗ + B ∗ (λ ¯ − A∗ )−1 C ∗ JD − D ∗ JC(λ − A)−1 B = J − D∗ JD − B ∗ (λ ¯ − A∗ )−1 C ∗ JC(λ − A)−1 B. −B ∗ (λ Using the identities D∗ JD = J, C ∗ JD = HB and C ∗ JC = H −1 A + A∗ H −1 , which hold by Theorem 17.1, one sees that ¯ − A∗ )−1 H −1 (λ − A)−1 B ≥ 0. J − W (λ)∗ JW (λ) = −2(Re λ)B ∗ (λ Hence W is J-inner. Conversely, if W is J-inner, where J = diag (Ip , −Iq ), then H is negative deﬁnite by Proposition 17.23. So, it remains to show that we can reduce the general case to the situation where J is of the form J = diag (Ip , −Iq ). To this end, let T be an invertible matrix such that T ∗ JT = J1 = diag (Ip , −Iq ) for some nonnegative integers p and q. Such a T does exist. Observe that J = T −∗ J1 T −1 , and since J = J −1 , we obtain that J = T J1 T ∗ . Consider the matrix function W1 = T −1 W T . Then W1 is J1 -inner, and has a minimal realization W1 (λ) = T −1DT + T −1 C(λ − A)−1 BT. We claim that H is the Hermitian matrix associated with this minimal realization. Indeed, using J = T J1 T ∗ we have AH + HA∗ = BJB ∗ = BT J1 T ∗ B ∗ , T −1 CH = T −1 DJB ∗ = (T −1 DT )J1 T ∗ B ∗ . By Theorem 17.1, the matrix H is the Hermitian matrix associated with the given minimal realization of W1 . So we can apply Proposition 17.23 to W1 in order to conclude that H is negative deﬁnite. In the next theorem we analyze J-inner functions in terms of a realization which is not necessarily minimal. As always in this chapter, J stands for a signature matrix. Theorem 17.24. Let W (λ) = D + C(λIn − A)−1 B be a (possibly non-minimal) realization of a rational m × m matrix function. Suppose D ∗ JD = J, and assume there exists a Hermitian matrix X such that XA + A∗ X = C ∗ JC,

XB = C ∗ JD,

Ker (C|A) ⊂ Ker X.

Then W is J-unitary. In that case W is J-inner if and only if X is nonpositive.

17.5. J-inner rational matrix functions

335

Proof. With respect to the orthogonal decomposition Cn = Im X ⊕ Ker X write ) ( G 0 . X= 0 0 Note that G is invertible and Hermitian. Also, with respect to the decomposition Cn = Im X ⊕ Ker X, write B1 A11 A12 , B= , C = C1 C2 . A= A21 A22 B2 Then XB = C ∗ JD yields XB =

GB1 0

=

C1∗ C2∗

JD.

Since D∗ JD = J, we know that D is invertible. Hence JD is invertible, and so C2 = 0. Now XA + A∗ X = C ∗ JC gives ∗ ∗ C + A G GA JC 0 GA 11 12 1 11 1 = . XA + A∗ X = A∗12 G 0 0 0 As G is invertible, one obtains A12 = 0. Therefore W (λ) = D + C1 (λ − A11 )−1 B1 ,

(17.30)

and for this realization of W we have GA11 + A∗11 G = C1∗ JC1 and GB1 = C1∗ JD. It is now suﬃcient to show that (17.30) is minimal. Indeed, the proof can then be completed by applying Theorems 17.1 and 17.22. One checks that Ker CAj = Ker C1 Aj11 ⊕ Ker X,

j = 0, 1, 2, . . . .

As Ker (C|A) ⊂ Ker X by assumption, we obtain Ker (C1 |A11 ) = {0}. Thus (C1 , A11 ) is an observable pair. It remains to show that (A11 , B1 ) is controllable. For this it suﬃces to prove that (A× 11 , B1 ) is a controllable pair, where −1 −1 ∗ A× C1 . Now A× A11 G, while B1 = G−1 C1∗ JD. So it is 11 = A11 − B1 D 11 = −G enough to show that (−A∗11 , C1∗ JD) is a controllable pair. But this is equivalent to (D∗ JC1 , −A11 ) being an observable pair. Now D∗ J is invertible, and hence Ker (D∗ JC1 | − A11 ) = Ker (C1 |A11 ) = {0}, which completes the proof.

We ﬁnish this section with a theorem on the multiplicative structure of J-inner rational matrix functions. It states that a J-inner rational matrix function admits a complete factorization into J-inner factors of McMillan degree 1.

336

Chapter 17. J-unitary rational matrix functions

Theorem 17.25. Let W be a J-inner rational matrix function of McMillan degree n. Then there are J-inner rational matrix functions W1 , . . . , Wn of McMillan degree 1 such that W = W1 · · · Wn . Proof. Employing a similar argument as in the proof of Lemma 17.17, taking into account Theorem 17.22, one can prove that the J-inner rational matrix function W admits a realization with upper triangular main matrix and having −I as its associated Hermitian matrix. Following the line of argument of the proof of Theorem 17.16 one then proves that a J-inner rational matrix function admits a minimal factorization into n factors of degree 1, and that these factors can be taken to be J-unitary. It remains to show that the factors are actually J-inner. Let us consider for each of the factors a minimal realization of the form Wj (λ) = Dj +

1 Dj JBj∗ h−1 j Bj . λ − aj

The Hermitian matrix associated with this realization is denoted by hj ; it is just a real number in this case (compare Example 17.7). Consider the minimal realization for W resulting from taking the product realization of the above minimal realizations of the Wi ’s. According to Proposition 17.6, the Hermitian matrix H associated with this product realization is the diagonal matrix with the numbers h1 , . . . , hn on the diagonal. According to Proposition 17.4 and the state space similarity theorem, there is an invertible matrix S such that SHS ∗ is the Hermitian matrix associated with the minimal realization of W mentioned in the ﬁrst paragraph of this proof. That is, SHS ∗ = −I. But this is only possible if all numbers hi are negative. Then we can apply Theorem 17.22 to conclude that each of the factors is J-inner.

17.6 Inner-outer factorization In this section we consider inner-outer factorization of a possibly non-square p × q rational matrix function L. First we introduce the necessary terminology. A p×q rational matrix function V is called inner if V is analytic on the closed right half plane (including the imaginary axis and inﬁnity) and the values of V on the imaginary axis are isometries. The latter means that V (λ)∗ V (λ) = Ip for each λ ∈ iR. Since V is assumed to be proper, this identity also holds at inﬁnity. By the maximum modulus principle, an inner function V satisﬁes V (λ) ≤ 1,

λ ≥ 0.

Note that for V to be inner, we must have q ≤ p. If q = p , then V is inner if and only if V is bi-inner (cf., the ﬁrst two paragraphs of Section 17.5). A rational square matrix-valued function X is said to be an invertible outer function if X is analytic on the closed right half plane (inﬁnity included) and

17.6. Inner-outer factorization

337

det X(λ) = 0 for λ ≥ 0 (again with inﬁnity included). Finally, given a p × q rational matrix function L, we say that a factorization L(λ) = V (λ)X(λ), is an inner-outer factorization if V is a p × q inner rational function and X is a q × q invertible outer rational matrix function.1 Clearly, for such a factorization to exist L must be analytic in the closed right half plane (inﬁnity included) and the values of L on iR ∪ {∞} have to be left invertible matrices. As we shall see (Theorem 17.26 below), these two conditions are not only necessary for L to have an inner-outer factorization but also suﬃcient. ¯ ∗ L(λ). Obviously, if L has an inner-outer factorization Put Φ(λ) = L(−λ) L = V X (suppressing the variable λ), then, since V takes isometric values on the imaginary axis and at inﬁnity, we have ¯ ∗ X(λ), Φ(λ) = X(−λ) and this factorization is a left spectral factorization (with respect to iR) of the rational q × q matrix function Φ. This gives a hint about how to construct an inner-outer factorization. Indeed, assume L is analytic in the closed right half plane, inﬁnity included, ¯ ∗ X(λ) be a left spectral factorization of Φ with respect to and let Φ(λ) = X(−λ) iR. Put V (λ) = L(λ)X(λ)−1 . Then V is analytic in the closed right half plane (inﬁnity included) because both L and X −1 are analytic there. In addition, V takes isometric values on the imaginary axis. Hence L = V X is an inner-outer factorization. This leads to the following theorem. Theorem 17.26. Let L(λ) = D +C(λIn −A)−1 B be a realization of a p×q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L takes left invertible values on the imaginary axis, and D∗ D = Iq . Then L admits an inner-outer factorization L(λ) = V (λ)X(λ) with the inner factor V and the invertible outer factor X given by −1 V (λ) = D + (I − DD ∗ )C + DB ∗ P λIn − (A − BD ∗ C + BB ∗ P ) B, X(λ) = Iq + (D∗ C − B ∗ P )(λIn − A)−1 B. Here P is the (unique) Hermitian iR-stabilizing solution of P BB ∗ P + P (A − BD∗ C) + (A∗ − C ∗ DB ∗ )P − C ∗ (I − DD∗ )C = 0, that is, the solution P = P ∗ for which A − BD∗ C + BB ∗ P has all its eigenvalues in the open left half plane. 1 Note that in our deﬁnition of inner-outer factorization, the outer factor is required to be invertible outer. This restricted version of inner-outer factorization is used throughout the book.

338

Chapter 17. J-unitary rational matrix functions

¯ ∗ L(λ). Using D∗ D = I and the given realization for L Proof. Put Φ(λ) = L(−λ) − A) −1 B, where we compute that Φ is given by the realization Φ(λ) = I + C(λ ∗ C D −A∗ C ∗ C = −B ∗ D ∗ C . (17.31) = , B= , C A 0 A B Since L has left invertible values on the imaginary axis (that is, has full column rank there), Φ takes positive deﬁnite values on the imaginary axis. Thus we know from Section 9.2 that Φ admits a left spectral factorization with respect to iR. It follows that an inner-outer factorization does exist under the assumptions of the theorem. To ﬁnd the spectral factorization in concrete form, we proceed as in the proof of Theorem 13.1. In other words we apply Theorem 12.5 with the data given by (17.31). The same argument as in the proof of Theorem 13.1 gives that the Riccati equation featured in the theorem has a Hermitian stabilizing solution P . Now use P to deﬁne X(λ) by the expression given in the theorem which is the analogue of the expression for L− (λ) in Theorem 13.1. With the function X obtained this ¯ ∗ X(λ). way, we have the left spectral factorization Φ(λ) = X(−λ) −1 It remains to compute V (λ) = L(λ)X(λ) . Note that −1 X −1 (λ) = I − (D∗ C − B ∗ P ) λ − (A − BD∗ C + BB ∗ P ) B. From A − BD∗ C + BB ∗ P = A − B(D∗ C − B ∗ P ), we now obtain −1 (λ − A)−1 BX(λ)−1 = λ − (A − BD∗ C + BB ∗ P ) B. Using the latter identity it is straightforward to deduce the formula for V given in the theorem. The following corollary will be useful in the ﬁnal chapter of the book. Corollary 17.27. Let L(λ) = D +C(λIn −A)−1 B be a realization of a p×q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L takes left invertible values on the imaginary axis, and D ∗ D = Iq . Then there is a q × p rational matrix function L (λ) which has no poles on the imaginary line including inﬁnity, such that L (iω)L(iω) = Iq , ω ∈ R. Proof. Let L(λ) = V (λ)X(λ) be an inner-outer factorization of L and take L (λ) = X(λ)−1 V (−λ)∗ . Next we consider the dual problem of outer-co-inner factorization. A possibly non-square rational matrix function V is called co-inner if V is analytic on the closed right half plane (including inﬁnity), and takes co-isometric values on the ¯ ∗. imaginary axis. In other words, V is co-inner if V is inner, where V (λ) = V (λ) Note that for V to be co-inner, we must have p ≤ q.

17.7. Unitary completions of minimal degree

339

A factorization L(λ) = X(λ)V (λ), where X is invertible outer and V is co-inner is called an outer-co-inner factor¯ ∗ = X(λ)X(−λ) ¯ ∗ is a right ization.2 Obviously, in that case Φ(λ) = L(λ)L(−λ) spectral factorization with respect to iR, and conversely. Using a duality argument we obtain the following counterpart to Theorem 17.26. Theorem 17.28. Let L(λ) = D + C(λ − A)−1 B be a realization of a p × q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L(iω) is right invertible for each ω ∈ R, and DD ∗ = I. Then L admits an outer-co-inner factorization L(λ) = X(λ)V (λ), with the co-inner factor and the invertible outer factor being given by −1 V (λ) = D + C λIn − (A − BD∗ C + QC ∗ C) B(I − D ∗ D) + QC ∗ D , X(λ) = Ip + C(λIn − A)−1 (BD∗ − QC ∗ ). Here Q is the (unique) Hermitian iR-stabilizing solution QC ∗ CQ + (A − BD∗ C)Q + Q(A∗ − C ∗ DB ∗ ) − B(I − D∗ D)B ∗ = 0, that is, the solution Q = Q∗ for which A − BD∗ C + QC ∗ C has all its eigenvalues in the open left half plane. Note that Proof. Let L(λ) = D ∗ +B ∗ (λ−A∗ )−1 C ∗ , and apply Theorem 17.26 to L. ∗ A also has all its eigenvalues in the open left half plane. So, applying Theorem yields a factorization L(λ) is 17.26 to L = V (λ)X(λ), where V is inner and X ∗ ∗ ¯ is co-inner and X(λ) = X( ¯ is invertible λ) invertible outer. Then V (λ) = V (λ) outer. So L(λ) = X(λ)V (λ) is an outer-co-inner factorization of L. Theorem 17.26 Those for V and X are now obtained also gives formulas for the factors V and X. ∗ ¯ ¯ ∗. λ) from the expressions V (λ) = V (λ) and X(λ) = X(

17.7 Unitary completions of minimal degree In this section we deal with the following completion problem. Given a strictly proper rational m × p matrix function W , having contractive values on the imaginary axis, ﬁnd an (m + p) × (m + p) rational matrix function U having unitary values on the imaginary axis, such that U11 (λ) W (λ) . (17.32) U (λ) = U21 (λ) U22 (λ) 2 Note that in our deﬁnition of outer-co-inner factorization, the outer factor is required to be invertible outer (cf., footnote 1).

340

Chapter 17. J-unitary rational matrix functions

In other words, we want to ﬁnd a unitary rational matrix function U such that W is embedded as a (right upper) corner in U . Moreover, we wish to ﬁnd such a U which has the same McMillan degree as W . We shall normalize U so that U (∞) = Im+p . This problem can be treated for the more general case of a proper W (see [75]). However, for sake of simplicity we shall conﬁne ourselves to the strictly proper case. The following theorem describes all possible solutions. Theorem 17.29. Let W (λ) = C(λIn − A)−1 B be a minimal realization of an m × p strictly proper rational matrix function W which is contractive on the imaginary axis. Then the set of all unitary rational (m + p) × (m + p) matrix functions U of the form (17.32) with U (∞) = Im+p and δ(U ) = δ(W ) is in one-to-one correspondence with the set of Hermitian solutions of the algebraic Riccati equation XC ∗ CX − AX − XA∗ + BB ∗ = 0.

(17.33)

Moreover, these Hermitian solutions X are invertible, and the one-to-one correspondence referred to above is given by Im 0 C (17.34) U (λ) = + (λIn − A)−1 XC ∗ B . ∗ −1 B X 0 Ip Proof. Suppose U is a unitary rational matrix function with W as its right upper corner block entry, U (∞) = Im+p and δ(U ) = δ(W ). The McMillan degree of W is n, the size of the main matrix in the given minimal realization W (λ) = C(λIn − A)−1 B of W . Hence δ(U ) = n, and U has a realization of the type U (λ) =

Im

0

0

Ip

+

1 C 2 C

−1 B 1 (λIn − A)

. 2 . B

−1 B 1 (λIn − A) 2 is realization of W . Comparing this realization Clearly W (λ) = C with the given one, and using the state space similarity theorem for minimal real = S −1 AS, B 2 = izations, we see that there exists an invertible n×n matrix with A −1 −1 S B and C1 = CS. Introducing C2 = C2 S and B1 = SB1 , we get Im 0 C (17.35) U (λ) = + (λIn − A)−1 B1 B , 0 Ip C2 and this realization of U is a minimal one. Since U is unitary, there is a Hermitian X such that ∗ ∗ C B1 B 1 , X= . AX + XA∗ = B1 B B∗ B∗ C2

(17.36)

17.8. Bi-inner completions of inner functions

341

In particular, we have B1∗ = CX. Inserting B1∗ = CX into the ﬁrst part of (17.36) we obtain (17.33). Moreover, X is invertible by minimality of the realization of U (see Theorem 17.1), and so C2 = B ∗ X −1 , which yields (17.34). Conversely, suppose that X is a Hermitian solution of (17.33). By minimality of the realization of W we have that X is invertible. The argument is as follows. Suppose Xx = 0. Then (17.33) gives x∗ BB ∗ x = 0, hence B ∗ x = 0. Again using (17.33) we get XA∗ x = 0, and we see that Ker X ⊂ Ker (B ∗ |A∗ ) = {0}. Let U be given by (17.34). Then, by Theorem 17.1, the rational matrix function U is unitary. Obviously, W is the right upper corner block entry of U and δ(U ) = δ(W ), and U (∞) = Im+p . To show that the correspondence between Hermitian solutions X of (17.33) and the set of all unitary rational matrix functions U of the form (17.32) with U (∞) = I and δ(U ) = δ(W ) is one-to-one we argue as follows. We have seen in the previous part of the theorem that any such U is necessarily of the form (17.34) for some Hermitian solution of (17.33). Assume that for two solutions X1 and X2 the functions U1 and U2 given by (17.34) with these solutions in place of X coincide. Then, from (17.36) it is seen that C ∗ (X1 − X2 ) = 0. A(X1 − X2 ) + (X1 − X2 )A = 0, C2 Hence Im (X1 − X2 ) is A-invariant, and it is also contained in Ker C. This implies that Im (X1 − X2 ) ⊂ Ker (C|A) = {0}. Thus X1 = X2 .

17.8 Bi-inner completions of inner functions Our aim in this section is to complete a possibly non-square inner function to a (square) bi-inner one. It is convenient to begin with two propositions. With the notation used in the ﬁrst proposition we anticipate Theorem 17.32 below. + C(λIn − A)−1 B be a realization of a p × q Proposition 17.30. Let V (λ) = D rational matrix function, and assume ∗D = Iq , D

σ(A) ⊂ Cleft,

= C ∗ D, YB

(17.37)

where Y is the unique (Hermitian) solution of the Lyapunov equation Y A + A∗ Y = C ∗ C.

(17.38)

Then V is inner. Conversely, if V is inner, the given realization of V is minimal, and Y is the unique (Hermitian) solution of the Lyapunov equation (17.38), then (17.37) is satisﬁed. Since A has all its eigenvalues in the open left half plane, equation (17.38) has a unique solution Y , and this solution is given by ∞ ∗ etA C ∗ CetA dt. (17.39) Y =− 0

342

Chapter 17. J-unitary rational matrix functions

From this representation one sees that the matrix Y is generally negative semidefinite, and that it has the stronger property of being negative deﬁnite when the is minimal (or even just observable). Thus + C(λIn − A)−1 B realization V (λ) = D the above result can be viewed as a special case of Theorem 17.24. It is illustrative to give a direct proof. is Proof. Assume that (17.37) holds with Y as indicated in the theorem. Then D an isometry by the ﬁrst condition in (17.37). Thus p ≥ q. For pure imaginary λ, a straightforward computation, using (17.37) and (17.38), gives ∗ ¯ − A∗ )−1 C ∗ D +B ∗ (λ + C(λ − A)−1 B V (λ)∗ V (λ) = D ∗ (λ + A∗ )−1 Y B +B ∗ Y (λ − A−1 B = Iq − B ∗ (λ + A∗ )−1 (Y A + A∗ Y )(λ − A)−1 B −B ∗ (λ + A∗ )−1 Y B +B ∗ Y (λ − A)−1 B = Iq − B ∗ (λ + A∗ )−1 Y (A − λ) + (A∗ + λ)Y )(λ − A −1 B = Iq . −B Hence V has isometric values on iR. Since V is analytic in the open right half plane by the second condition in (17.37), we may conclude that V is inner. + C(λIn − A)−1 B be Next, let V be inner and let the realization V (λ) = D minimal. Clearly, since V is inner, the ﬁrst two conditions in (17.37) are satisﬁed. = C ∗ D. This Let Y be the unique solution of (17.38). It remains to show that Y B is done by using the same arguments as used in the proof of Proposition 17.3. Proposition 17.31. Let U (λ) = D + C(λIn − A)−1 B be a realization of a p × p rational matrix function. Assume D∗ D = Iq ,

σ(A) ⊂ Cleft,

Y B = C ∗ D,

(17.40)

where Y is the unique (Hermitian) solution of the Lyapunov equation Y A + A∗ Y = C ∗ C.

(17.41)

Then U is bi-inner and the McMillan degree of U is equal to the rank of Y which, in turn, is equal to dim Ker (C|A)⊥ . Proof. The fact that U is bi-inner follows from Proposition 17.30. Since σ(A) is contained in Cleft , the unique solution Y of (17.41) is given by the integral representation (17.39), from which we easily obtain Ker Y = Ker (C|A). Now consider the decomposition Cn = X1 ⊕ X2 , where X1 = Ker (C|A) and X2 is the orthogonal compliment of X1 in Cn . Thus X1 = Ker Y and X2 = Im Y . In particular rank Y = dim X2 . Write A, B, C and Y as block matrices according to the decomposition Cn = X1 ⊕ X2 . Then B1 A1 0 0 , B= , C = 0 C2 , , (17.42) A= Y = B2 0 A2 0 Y2

17.8. Bi-inner completions of inner functions

343

and U (λ) = D + C2 (λIn − A2 )−1 B2 . Since rank Y = rank Y2 = dim X2 , it suﬃces to prove that this second realization of U is minimal. From (17.41), the third identity in (17.40) and the partitioning of A, B, C and Y in (17.42), we see that Y2 A2 + A∗2 Y2 = C2∗ C2 ,

Y2 B2 = C2∗ D.

−1 C2 , the associate main matrix of the realization U (λ) = For A× 2 = A2 − B2 D D + C2 (λIn − A2 )−1 B2 , this gives −1 Y2 A× C2 = −A∗2 Y2 + C2∗ C2 − C2∗ DD−1 C2 = −A∗2 Y2 . 2 = Y2 A2 − Y2 B2 D ∗ Now Y2 is invertible. Thus A× 2 and −A2 are similar. From the second part of (17.40) and the partitioning of A in (17.42), we see that σ(A2 ) ⊂ Cleft . Taking × ∗ into account the similarity of A× 2 and −A2 , it follows that σ(A2 ) ⊂ Cright . In × particular, σ(A2 ) and σ(A2 ) are disjoint. But then, by a remark made after the proof of Theorem 7.6 in [20], the realization U (λ) = D + C2 (λIn − A2 )−1 B2 is minimal.

Let V be as in Proposition 17.30, so in particular V is inner. Returning to the aim of this section, we shall now complete V to a p × p bi-inner rational matrix function. Before turning to the theorem in question, we make some preparations. is an isometry. Thus According to the ﬁrst condition in (17.37) the matrix D p ≥ q. When p = q, there is nothing to do. Therefore in what follows we take D ∗ is an orthogonal is an isometry, implies that Ip − D p > q. The fact that D projection of rank p − q. Thus we can choose a p × (p − q) isometry E such that D ∗ = EE ∗ . Now note that there exists an n × (p − q) matrix B such that Ip − D Y B = C ∗ E.

(17.43)

Since Y is Hermitian, to prove that equation (17.43) has a solution of the desired form, it suﬃces to show that Ker Y ⊂ Ker E ∗ C. In fact, we have Ker Y ⊂ Ker C. Indeed, assume that Y x = 0, then we see from (17.38) that x∗ C ∗ Cx = 0, which is equivalent to Cx = 0. be a realization of a p × q + C(λIn − A)−1 B Theorem 17.32. Let V (λ) = D rational matrix function satisfying the conditions (17.37), where Y is the unique (Hermitian) solution of the Lyapunov equation (17.38). Let E be a p × (p − q) D ∗ = EE ∗ , and let B be an n × (p − q) matrix solution isometry such that Ip − D of (17.43). Put U (λ) = D + C(λIn − A)−1 B, where B and D are the p × p ] and D = [ E D ]. Then U is a p × p bi-inner matrices given by B = [ B B completion of V , that is, U is a bi-inner rational p × p matrix function of the form [ V (λ) V (λ) ], and the McMillan degree of U is equal to the rank of Y . The rational p × (p − q) matrix function V can be described explicitly; it is actually given by the realization V (λ) = E + C(λIn − A)−1 B .

344

Chapter 17. J-unitary rational matrix functions

Proof. To prove that U is bi-inner, apply Proposition 17.30 to U with its given realization. Since (17.38) holds, it suﬃces to show that Y B = C ∗ D and D∗ D = Ip . These facts follow from the third identity in (17.37) and the deﬁnitions of E and B . Indeed, we have = C∗E C ∗D = C ∗ D, Y B = Y B Y B ∗ E D ∗ = Ip , = EE ∗ + D DD∗ = E D ∗ D

and, since D is a square matrix, DD ∗ = Ip amounts to the same as D∗ D = Ip . The ﬁnal statement is an immediate corollary of Proposition 17.31. Next, we return to the inner-outer factorization discussed in Section 17.6. The point we focus on here is the completion of the inner factor to a bi-inner function. Let L(λ) = D + C(λIn − A)−1 B be a realization of a p × q rational matrix function. Assume A has all its eigenvalues in the open left half plane, L(iω) is left invertible for each ω ∈ R, and D ∗ D = Iq . Let L(λ) = V (λ)X(λ) be the inner-outer factorization constructed in Theorem 17.26, in particular, −1 V (λ) = D + (I − DD∗ )C + DB ∗ Y λIn − (A − BD ∗ C + BB ∗ Y ) B, where Y = Y ∗ satisﬁes the algebraic Riccati equation Y BB ∗ Y + Y (A − BD∗ C) + (A∗ − C ∗ DB ∗ )Y − C ∗ (I − DD∗ )C = 0, and A − BD∗ C + BB ∗ Y has all its eigenvalues in the open left half plane. Choose a p × (p − q) isometry E such that I − DD ∗ = EE ∗ , and let B be any n × (p − q) matrix such that Y B = C ∗ E. Corollary 17.33. In the situation described in the previous paragraph, introduce U (λ) = V (λ) V (λ) , where the rational p × (p − q) matrix function V is given by −1 V (λ) = E + (I − DD∗ )C + DB ∗ Y λIn − (A − BD ∗ C + BB ∗ Y ) B . Then U is bi-inner. Proof. All we need to show is that Theorem 17.32 may be applied with the matrices A − BD ∗ C + BB ∗ Y and (I − DD∗ )C + DB ∗ Y in place of A and ∗ C, respectively. For this we need to verify the identities (I − DD∗ )C + DB ∗ Y D = Y B and Y (A − BD∗ C + BB ∗ Y ) + (A − BD∗ C + BB ∗ Y )∗ Y ∗ = (I − DD ∗ )C + DB ∗ Y (I − DD∗ )C + DB ∗ Y . This involves nothing more than a routine computation using that D∗ D = I and that Y = Y ∗ is a solution of the Riccati equation featured in the paragraph preceding the corollary.

17.8. Bi-inner completions of inner functions

345

Notes The ﬁrst three sections are largely based on [3]. The Redheﬀer transformation of Section 17.4, which is a standard tool in the analysis of 2 × 2 block matrix functions, originates from [130]. Theorem 17.22 in Section 17.5 also implies that if W is a J-inner rational matrix function, then the function K∗,W (μ, λ) has no negative squares, that is, it is a positive deﬁnite kernel, see also Theorem 2.5 in [39]. Theorem 17.25 in Section 17.5 is a simple case of a more far-reaching theory concerning the multiplicative structure of general matrix-valued J-inner functions, which originates from [118]; see also Chapter 4 in [39]. Factorizations in degree 1 factors, of which Theorem 17.25 provides an example, are the main topic of Part III in [20]. Section 17.6 originates from Section 7.4 in [43]; for the corresponding state space formulas, see [146]. Section 17.7 is related to the problem of Darlington synthesis. The latter problem can be found in [4]. The presentation given here is based on [75]. For further results in this direction, including Darlington embedding for time-variant systems, see [36] and Chapter 6 in [117]. The result presented in Section 17.8 may be found in, e.g., Chapter 12 (page 249) in [149].

Part VII Applications of J-spectral factorizations In this part, the state space theory of J-spectral factorization, developed in the preceding two parts, is used to solve H∞ -problems. There are three chapters. The ﬁrst chapter (Chapter 18) presents the solution of the Nehari interpolation problem for rational matrix functions. The second chapter (Chapter 19) reviews elements from control and mathematical systems theory that play an essential role in the ﬁnal chapter. The third and ﬁnal chapter (Chapter 20) treats H∞ -control problems. Here we use the J-spectral factorization theory to obtain the solutions of some of the main problems in this area, namely the standard problem, the one-sided problem, and the full model matching problem.

Chapter 18

Application to the rational Nehari problem In this chapter the rational matrix version of the Nehari problem (relative to the imaginary axis) is solved using a J-spectral factorization approach. The data of the problem are given in realized form. This together with the state space results on J-spectral factorization derived in Chapter 14 allows us to solve the problem and to obtain an explicit linear fractional representation of all its solutions, again in realized form. The main attention is given to the so-called suboptimal case. The more general Nehari-Takagi problem is also solved using the J-spectral factorization method. This chapter consists of six sections. Section 18.1 presents the problem statement and the main theorem. Section 18.2 deals with the theory of linear fractional maps. Such maps will play an important role in this and the ﬁnal chapter. In Section 18.3 the rational matrix Nehari problem is reduced to a J-spectral factorization of a special kind, and all solutions are described in terms of the coeﬃcients of the J-spectral factor. This result is used in Section 18.4 to prove the main theorem of Section 18.1. Section 18.5 deals with the Nehari problem for the non-stable case, when the given function does not necessarily have all its poles in the open left half plane. Section 18.6, the ﬁnal section of the chapter, gives the solution of the rational matrix Nehari-Takagi problem.

18.1 Problem statement and main result Let R be a rational p × q matrix function which does not have a pole on the imaginary axis and at inﬁnity. In particular, R is proper. In this section we study the problem of ﬁnding all proper rational p × q matrix functions K such that K

350

Chapter 18. Application to the rational Nehari problem

has all its poles in the open right half plane and K − R∞ = sup K(s) − R(s) < γ, s∈iR

(18.1)

where γ is a pre-speciﬁed positive number. Note that both R and K are proper and have no pole on the imaginary axis, and hence the so-called inﬁnity norm K − R ∞ is well-deﬁned. We shall refer to this problem as the (suboptimal) rational Nehari problem for R relative to the imaginary axis with tolerance γ. The latter qualiﬁer will be omitted when γ = 1. The word “suboptimal” refers to the fact that we use in (18.1) a strict inequality. We ﬁrst deal with the case when R is stable. A rational matrix function is called iR-stable, or simply stable when no confusion is possible (as will be the case in this chapter), if all its poles are in the open left half plane. Note that such a function is proper and has no pole on iR. We shall assume additionally that R is strictly proper. To state the main result we start with a realization of R. Since R is stable and strictly proper, we can choose a realization of R of the form R(λ) = C(λIn − A)−1 B,

(18.2)

with the property that A has all its eigenvalues in the open left half plane. Let P and Q be the unique solutions of the Lyapunov equations AP + P A∗ = −BB ∗ ,

A∗ Q + QA = −C ∗ C,

respectively. Note that P and Q are given by ∞ τA ∗ τ A∗ e BB e dτ, Q= P = 0

∞

(18.3)

∗

eτ A C ∗ Ceτ A dτ.

0

Hence P and Q are nonnegative Hermitian matrices. One usually refers to P as the controllability gramian, and to Q as the observability gramian, corresponding to the realization (18.2). We shall prove the following theorem. Theorem 18.1. Let R(λ) = C(λIn − A)−1 B be a realization of the p × q rational matrix function R, and assume that A has all its eigenvalues in the open left half plane. Then the rational Nehari problem for R relative to the imaginary axis with tolerance γ is solvable if and only if the matrix γ 2 In − P 1/2 QP 1/2 is positive deﬁnite. In that case all solutions of the Nehari problem for R can be obtained in the following way. Introduce the rational matrix functions X11 (λ)

= Ip + CP (λIn + A∗ )−1 Z −1 C ∗ ,

(18.4)

X12 (λ)

= CP (λIn + A∗ )−1 Z −1 QB,

(18.5)

X21 (λ)

= −B ∗ (λIn + A∗ )−1 Z −1 C ∗ ,

(18.6)

X22 (λ)

= Iq − B ∗ (λIn + A∗ )−1 Z −1 QB,

(18.7)

18.1. Problem statement and main result

351

where Z = γ 2 In − QP . Then all solutions K of the rational Nehari problem for R relative to the imaginary axis are given by −1 K(λ) = − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) ,

(18.8)

where H is any rational p × q matrix function which has all its poles in the open right half plane and satisﬁes H∞ < γ. Moreover, there is a one-to-one correspondence between the solution K and the free parameter H. Before we prove the above theorem (in Section 18.4 below) it will be convenient ﬁrst to make some preparations. The following lemma restates the necessary and suﬃcient condition appearing in Theorem 18.1 in operator language. Lemma 18.2. Let R(λ) = C(λIn − A)−1 B be a realization of the p × q rational matrix function R, and assume that A has all its eigenvalues in the open left half plane. Consider the Hankel operator HR generated by R, that is the ﬁnite rank integral operator from Lq2 [0, ∞) into Lp2 [0, ∞) given by ∞ (HR f )(t) = CeA(t+τ ) Bf (τ ) dτ. 0

Then HR < γ if and only if the matrix γ 2 In − P 1/2 QP 1/2 is positive deﬁnite. Proof. We need the controllability operator Ξ and the observability operator Ω associated with the realization (18.2). Thus ∞ Ξf = eτ A Bf (τ ) dτ, Ξ : Lq2 [0, ∞) → Cq , 0

Ω : Cn → Lp2 [0, ∞),

(Ωx)(t) = CetA x,

t > 0.

Clearly P = ΞΞ∗ , Q = Ω∗ Ω and HR = ΩΞ. Now let λ1 (X) denote the largest eigenvalue of an operator X all of whose non-zero spectrum consists of positive eigenvalues. Then HR 2

=

∗ λ1 (HR HR ) = λ1 (Ξ∗ Ω∗ ΩΞ)

=

λ1 (ΞΞ∗ Ω∗ Ω) = λ1 (P Q) = λ1 (P 1/2 QP 1/2 ).

Hence HR < γ if and only if all the eigenvalues of P 1/2 QP 1/2 are strictly less than γ 2 . Thus HR < γ if and only if γ 2 I − P 1/2 QP 1/2 is positive deﬁnite. We close the section by showing that, without loss of generality, we may assume that in Theorem 18.1 the tolerance γ = 1. Indeed, consider for the original problem R(λ) = γ −1 R(λ), and K(λ) = γ −1 K(λ). Then we have R − K∞ < γ − K ∞ < 1. Moreover, if R is given by the realization (18.2), if and only if R = γ −1 C. One easily admits the realization R = C(λ − A)−1 B, where C then R

352

Chapter 18. Application to the rational Nehari problem

of the corresponding Lyapunov equations (18.3), sees that, for solutions P and Q −2 = γ Q. Hence Z = I − PQ = γ −2 Z. For the functions Xij (λ) one has P = P , Q appearing in Theorem 18.1 we have the following: 11 (λ) X

=

(λ + A∗ )−1 Z −1 C ∗ = X11 (λ), Ip + CP

12 (λ) X

=

(λ + A∗ )−1 Z −1 QB = γ −1 X21 (λ), CP

21 (λ) X

=

−1 C ∗ = γX21 (λ), −B ∗ (λ + A∗ )−1 Z

22 (λ) X

=

−1 QB = X22 (λ). Iq − B ∗ (λ + A∗ )−1 Z

Suppose that K(λ) is a solution to the problem with γ = 1, given by 11 (λ)H(λ) 12 (λ))(X 21 (λ)H(λ) 22 (λ))−1 , K(λ) = −(X +X +X satisfying H ∞ < 1. Now taking H(λ) = γ H(λ) for some H we have H∞ < γ, and with K(λ) = γ K(λ), we obtain that (18.8) holds.

18.2 Intermezzo about linear fractional maps The expression (18.8), which assigns to the rational matrix function H a rational matrix function K, is usually called a linear fractional map. Such maps will play an important role in this and the ﬁnal chapter. Therefore, we review some of the main properties of linear fractional maps in this section. It will be convenient ﬁrst to introduce some notation and terminology. Given a p × q rational matrix function F , we write F ∗ for the adjoint of F relative to the ¯ ∗ . (In engineering literature, including [76], imaginary axis, that is, F ∗ (λ) = F (−λ) [43]), this function is often denoted by F ∼ .) By Rat we shall denote the set of all rational matrix functions that are proper and have no pole on the imaginary axis iR, and Ratp×q will stand for the set of all F in Rat that are of size p × q. If F belongs to Ratp×q , then F ∗ belongs to Ratq×p . Note that Ratp×q is closed under the usual addition of matrix functions as well as under scalar multiplication. Also for F ∈ Ratp×q and G ∈ Ratq×r , we have F G ∈ Ratp×r . In particular Ratp×p is an algebra. The unit element in this algebra is Ep , the p × p matrix function which is identically equal to the p × p identity matrix Ip . A function F ∈ Ratp×p is said to be invertible in Ratp×p if F has an inverse G in Ratp×p , that is, G ∈ Ratp×p and F G = GF = Ep . For a rational p × p matrix function F such that det F (λ) ≡ 0, the pointwise inverse F −1 , deﬁned by F −1 (λ) = F (λ)−1 , is again a rational matrix function. If F ∈ Ratp×p and det F (λ) ≡ 0, then F −1 need not be an element of Ratp×p . Indeed, F −1 might have a pole on the imaginary axis or fail to be proper. In fact, F −1 ∈ Ratp×p if and only if F is biproper and det F (λ) has no zero on iR, and in that case F −1 is the inverse of F in the algebra Ratp×p .

18.2. Intermezzo about linear fractional maps

353

A function F in Ratp×q is analytic on the imaginary axis and at inﬁnity. Hence we can consider the norm F ∞ = sup F (s).

(18.9)

s ∈ iR

This is the usual L∞ -norm for bounded matrix functions on iR which we already p×q used in (18.1). We write F ∈ Ratp×q and its B , whenever F belongs to Rat p×q inﬁnity-norm F ∞ is strictly less than 1. Thus RatB is the open unit ball in Ratp×q with respect to the norm deﬁned by (18.9). Note that F ∞ < 1 is ¯ ∗ F (λ) being positive deﬁnite on iR ∪ {∞}. For the latter equivalent to Ip − F (−λ) property we use the notation Ep − F ∗ F > 0. Now let Θ ∈ Rat(p+q)×(p+q) , and let us partition Θ as a 2 × 2 block matrix function in the following way: Θ11 (λ) Θ12 (λ) (18.10) Θ(λ) = Θ21 (λ) Θ22 (λ) with Θ11 (λ) a p × p matrix and Θ22 (λ) a q × q matrix. With this partitioning of Θ we associate the linear fractional map −1 (FΘ H)(λ) = Θ11 (λ)H(λ) + Θ12 (λ) Θ21 (λ)H(λ) + Θ22 (λ) .

(18.11)

Here H is assumed to be in Ratp×q . In general, it is not clear for which H the map is well-deﬁned. However for a J-unitary Θ, with J = diag (Ip , −Iq ), we have the following result. Theorem 18.3. Let Θ ∈ Rat (p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ). Then Θ is invertible in Rat(p+q)×(p+q) , the maps FΘ and FΘ−1 are well-deﬁned on Ratp×q and map Ratp×q into itself. Moreover B B H ∈ Ratp×q B .

H = FΘ−1 FΘ H = FΘ FΘ−1 H,

(18.12)

Proof. We divide the proof into three parts. In the ﬁrst part it is shown that Θ−1 is in Rat(p+q)×(p+q) and is J-unitary, and also that the maps FΘ and FΘ−1 are p×q well-deﬁned on Ratp×q into B . In the second part we prove that FΘ maps RatB itself. In the ﬁnal part the identities in (18.12) will be established. Part 1. Since Θ is proper and has no pole on iR, the fact that Θ is J-unitary implies that for each λ ∈ iR ∪ {∞} the matrix Θ(λ) is J-unitary and hence invertible. It follows that Θ is invertible in Rat(p+q)×(p+q) and that Θ−1 is J-unitary. The fact that the matrix Θ(λ) is J-unitary for λ ∈ iR ∪ {∞} implies that Θ22 (λ) is invertible and Θ22 (λ)−1 Θ21 (λ) < 1 for λ ∈ iR ∪ {∞} . It follows that Θ22 is invertible in Ratq×q and that −1 Θ−1 Θ21 (λ) = 22 Θ21 ∞ = sup Θ22 (λ) λ∈R

max

λ ∈ iR ∪{∞}

Θ22 (λ)−1 Θ21 (λ) < 1.

354

Chapter 18. Application to the rational Nehari problem

p×q −1 −1 Next, take H ∈ Rat 21 ∞ H∞ < 1. Thus B−1 . Then Θ22 Θ21 H∞ ≤ Θ22 Θ Θ21 H + Θ22 = Θ22 Θ22 Θ21 H + Eq is invertible in Ratq×q . It follows that FΘ H −1 is also J-unitary, FΘ−1 is well-deﬁned is well-deﬁned for H ∈ Ratp×q B . Since Θ p×q on RatB too.

into itself. Take H in Ratp×q Part 2. In this part we show that FΘ maps Ratp×q B B , and write F = FΘ H. First note that F (Θ11 H + Θ12 )(Θ21 H + Θ22 )−1 H (18.13) = = Θ X −1 , −1 Eq (Θ21 H + Θ22 )(Θ21 H + Θ22 ) Eq

where X = Θ21 H + Θ22 . The fact that Θ is J-unitary, with J = diag (Ip , −Iq ) is equivalent to the identity 0 0 E E p p Θ∗ Θ = . (18.14) 0 −Eq 0 −Eq Hence, using (18.13), we obtain ∗

Eq − F F

=

−

F

∗

Eq

0

0

−Eq

−X

−∗

=

−X

−∗

=

X −∗ Eq − H ∗ H X −1 .

=

H

∗

H

∗

Eq

Eq

Ep

Θ

∗

F

Eq

Ep

0

0

−Eq

Ep

0

0

−Eq

Θ

H Eq

H

Eq

X −1

X −1

¯ ∗ F (λ) = X(−λ) ¯ −∗ Iq − H(−λ) ¯ ∗ H(λ) X(λ)−1 . Now It follows that Iq − F (−λ) ¯ ∗ H(λ) is positive deﬁnite on iR ∪ {∞}. H∞ < 1. This means that Ip − H(−λ) ∗ ¯ But then Iq − F (−λ) F (λ) is also positive deﬁnite on iR ∪ {∞}. The latter is equivalent to F ∞ < 1. Thus F ∈ Ratp×q B , as desired. From what has been proved so far, we conclude that the result of the previous steps also hold with Θ−1 instead of Θ. Thus FΘ−1 maps Ratp×q into itself. B Therefore, to complete the proof, it remains to prove the identities in (18.12). In fact, by interchanging the roles of Θ and Θ−1, it suﬃces to prove the ﬁrst identity in (18.12). This will be done in the next part. Part 3. Take H ∈ Ratp×q B , and put F = FΘ H, G = FΘ−1 F . From (18.14) we see that ∗ ∗ E Θ 0 0 −Θ E p p 11 21 Θ−1 = Θ∗ = . (18.15) −Θ∗12 Θ∗22 0 −Eq 0 −Eq

18.2. Intermezzo about linear fractional maps

355

By using (18.13) for Θ as well as for Θ−1 , we have

F

= Θ−1

Eq

(Θ21 H + Θ22 )−1 ,

Eq

G

H

=Θ

Eq

F Eq

(−Θ∗12 F + Θ∗22 )−1 .

Now observe that −Θ∗12 F

+

Θ∗22

=

=

=

0

Eq

0

Eq

0

Eq

Θ∗11 F − Θ∗21

=

−Θ∗12F + Θ∗22

Θ

−1

Θ

H

0

Eq

−1

Θ

F

Eq

(Θ21 H + Θ22 )−1

Eq

H Eq

(Θ21 H + Θ22 )−1 = (Θ21 H + Θ22 )−1 .

In particular, (Θ21 H + Θ22 )−1 (−Θ∗12 F + Θ∗22)−1 = Eq . But then G =

=

=

Ep

Ep

Ep

0

0

0

G

=

Eq

Θ

−1

Θ

H Eq

H

Ep

0

Θ

−1

F Eq

(−Θ∗12 F + Θ∗22 )−1

Eq

(Θ21 H + Θ22 )−1 (−Θ∗12 F + Θ∗22 )−1

= H,

which proves the ﬁrst identity in (18.12).

We are particulary interested in proper rational p × q matrix functions that are analytic on the closed left half plane with inﬁnity included. The class of these p×q functions will be denoted by Ratp×q have no pole + . Since the functions in Rat+ p×q p×q on iR and are proper, Rat+ is a linear subspace of Rat . We write Ratp×q +, B for the set of all F Ratp×q such that (18.9) holds. Thus + p×q ∩ Ratp×q Ratp×q +, B = Rat+ B .

356

Chapter 18. Application to the rational Nehari problem

Now, as in Theorem 18.3, let Θ ∈ Rat(p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ). Fix R ∈ Ratp×q , and consider Θ11 − RΘ21 Θ12 − RΘ22 Ep −R V11 V12 = = Θ. (18.16) V = 0 Eq V21 V22 Θ21 Θ22 Since Θ is invertible in Rat(p+q)×(p+q) by Theorem 18.3, it follows that the same holds true for V . Let FV be the linear fractional map deﬁned by V . Since V21 = Θ21 and V22 = Θ22 , we know from Theorem 18.3 that for each function H in Ratp×q the B function V21 H + V22 is invertible in Ratq×q . Thus FV is well-deﬁned on Ratp×q B . Moreover, since V11 H + V12

= (Θ11 − RΘ21 )H + (Θ12 − RΘ22 ) = (Θ11 H + Θ12 ) − R(Θ21 H + Θ22 ),

we see that

H ∈ Ratp×q B .

FV H = FΘ H − R,

The fact V22 = Θ22 implies that V22 is invertible in Rat is the second main result of this section.

q×q

(18.17)

. The following theorem

Theorem 18.4. Let Θ ∈ Rat(p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ), and let V be given by (18.16), where R ∈ Ratp×q . Then V is invertible in Rat (p+q)×(p+q) , and V22 is invertible in Rat q×q . Assume additionally that (p+q)×(p+q)

(a) V and V −1 belong to Rat+

,

−1 (b) V22 and V22 belong to Ratq×q + .

Then FV is well-deﬁned and one-to-one on Ratp×q +, B . Also p×q FV Ratp×q | R + K∞ < 1 . +, B = K ∈ Rat+

(18.18)

Note that conditions (a) and (b) in the above theorem are not independent. Indeed, the property that V22 belongs to Ratq×q follows from the fact that V + (p+q)×(p+q) . belongs to Rat+ Proof. The fact that V is invertible in Rat(p+q)×(p+q) and V22 in Ratq×q has already been proved in the two paragraphs preceding Theorem 18.4. From Theorem 18.3 we know that FΘ is well-deﬁned and one-to-one on Ratp×q B . But then we see p×q from (18.17) that the same holds true for FV . Now recall that Ratp×q +, B ⊂ RatB . p×q This allows us to conclude that FV is well-deﬁned and one-to-one on Rat+, B . It remains to show the identity (18.18). This will be done in two parts. The ﬁrst part covers the inclusion p×q | R + K∞ < 1 . (18.19) FV Ratp×q +, B ⊂ K ∈ Rat+

18.2. Intermezzo about linear fractional maps

357

The reverse inclusion is proved in the second step. p×q Part 1. Take H in Ratp×q +, B . We ﬁrst show that FV H belongs to Rat+ . From condition (a) we know that V is analytic on the closed left half plane. Hence the same holds true for the entries Vij , i, j = 1, 2. Now V22 = Θ22 is invertible in −1 Ratq×q + , and so V22 V21 H is analytic on the closed left half plane. Moreover, −1 −1 V21 H∞ ≤ V22 V21 ∞ H∞ ≤ Θ−1 V22 22 Θ21 ∞ H∞ < 1. −1 By the maximum modulus principle, this gives V22 (λ)V21 (λ)H(λ) < 1 for λ in −1 + Iq is invertible for each λ the closure of Cleft . It follows that V22 (λ)V21 (λ)H(λ) −1 in the closed left half plane, and that the function V22 (λ)V21 (λ)H(λ) + Iq )−1 is −1 again analytic on the closed left half plane. Thus V22 V21 H + Eq is invertible in q×q Ratq×q too. Combining these facts + . By assumption, V22 is invertible in Rat+ q×q we obtain that V21 H + V22 belongs to Rat+ and is invertible in Ratq×q + . But p×q then FV H belongs to Rat+ , as desired. Next, consider K = FV H. Using (18.17), we see that R + K = FΘ H. Since p×q Ratp×q and FΘ maps Ratp×q into itself (by Theorem 18.3), +, B is a subset of RatB B we know that R + K belongs to Ratp×q , that is, R + K∞ < 1. Thus (18.19) is B proved

Part 2. Take K ∈ Ratp×q + , and suppose R + K∞ < 1. Since R + K belongs p×q to RatB , we know from Theorem 18.3 that there exists a unique H in Ratp×q B such that FΘ H = R + K. In fact, by (18.12), the function in question is H = FΘ−1 (R + K). Furthermore, according to (18.17), the equality FΘ H = R + K yields FV H = K. Note that H has no poles on iR ∪ {∞}. The main diﬃculty is to show that H is analytic on the open left half plane Cleft. From (18.15) and H = FΘ−1 (R + K) we know that −1 H = FΘ−1 (R + K) = Θ∗11 (R + K) − Θ∗21 − Θ∗12 (R + K) + Θ∗22 . Put H1 = Θ∗11 (R + K) − Θ∗21 ,

H2 = −Θ∗12 (R + K) + Θ∗22 .

Then H2 is invertible in Ratq×q and H = H1 H2−1 . Moreover, R+K K H1 −1 −1 = Θ = V . H2 Eq Eq (p+q)×(p+q)

(18.20)

Since V −1 and K belong to Rat+ and Ratp×q + , respectively, we see and H2 belongs from the second equality in (18.20) that H1 belongs to Ratp×q + to Ratq×q . In other words, H and H are analytic in the open left half plane. 1 2 + Hence, in order to prove that H is analytic on the open left half plane Cleft , it remains to show H2−1 is analytic in Cleft .

358

Chapter 18. Application to the rational Nehari problem Multiplying (18.20) from the left by V we get V21 H1 + V22 H2 = Eq , hence

−1 −1 −1 −1 V22 = V22 (V21 H1 + V22 H2 ) = V22 (V21 H + V22 )H2 = (V22 V21 H + Eq )H2 .

Now introduce the scalar rational functions f (λ)

=

g(λ)

=

det V22 (λ)−1 , det V22 (λ)−1 V21 (λ)H(λ) + Iq ,

h(λ)

=

det H2 (λ).

Then f = gh. Also f, g and h have no poles or zeros on iR ∪{∞}. This allows us to use winding number arguments (see Section IV.5 in [32]; also [53], pages 143 and 152). For simplicity we write wn◦ (f ) for the winding number around the origin of f , and we use the analogous notation for g and h. Note that wn◦ (f ) is just equal to the diﬀerence of the number of zeros and number of poles (multiplicities taken into account) of f in Cleft , and similarly for wn◦ (g) and wn◦ (h). First observe −1 that, by condition (b) in our theorem, both V22 and V22 are analytic in the closed left half plane. Thus f has no zeros or poles in the closed left half plane, which implies that wn◦ (f ) = 0. Since −1 −1 V22 V21 H∞ ≤ V22 V21 ∞ H∞ < 1,

it follows that g is analytic on the closed left half plane and has no zeros in the closed left half plane. Thus wn◦ (g) is also zero. The fact that f = gh implies that wn◦ (f ) is the sum of wn◦ (g) and wn◦ (h). Hence wn◦ (h) = 0. We already know that h is analytic on the closed left half plane. Thus wn◦ (h) = 0 tells us that h has no zeros on the closed left half plane. This implies H2 is analytic on Cleft , and hence the same holds true for H. Next we present a more general version of Theorem 18.4. In this more general version K ∈ Ratp×q is not supposed to be analytic on the open left half plane Cleft but K is required to have a prescribed number of poles in Cleft . To state the result we need the following terminology. Let F ∈ Ratp×q . By the number of poles of F in the open left half plane, multiplicities taken into account, we mean the nonnegative integer

δ(F ; λ). (18.21) λ ∈ Cleft

Here δ(F ; λ) is the local degree of F at λ deﬁned in the one but last paragraph of Section 8.2. Since δ(F ; λ) is non-zero if and only if λ is a pole of F , the sum in (18.21) is ﬁnite. Theorem 18.5. Let Θ ∈ Rat(p+q)×(p+q) be J-unitary with J = diag (Ip , −Iq ), and let V be given by (18.16), where R ∈ Ratp×q . Then V is invertible in Rat(p+q)×(p+q) , and V22 is invertible in Ratq×q . Assume additionally that

18.3. The J-spectral factorization approach (p+q)×(p+q)

(α) V and V −1 belong to Rat+

359

,

−1 (β) V22 belongs to Ratq×q and V22 has precisely κ poles, multiplicities taken + into account, in Cleft.

Then FV is well-deﬁned and one-to-one on Ratp×q +, B . Also p×q | R + K∞ < 1 and K FV Ratp×q +, B = K ∈ Rat

(18.22) has κ poles in Cleft , multiplicities taken into account .

For κ = 0 the above theorem is just Theorem 18.4. To prove Theorem 18.5 one can use the same line of reasoning as in the proof of Theorem 18.4 above. However, the winding number argument employed in the ﬁnal paragraph of the proof of Theorem 18.4 has to be used in a more sophisticated way. For the details we refer to the literature; see, e.g., [86] and the references therein.

18.3 The J-spectral factorization approach In this section we shall exhibit the connection between the rational Nehari problem and J-spectral factorization. From the ﬁnal paragraph of Section 18.1 we know that without loss of generality the tolerance γ can be assumed to be equal to 1. Therefore, in what follows we take γ = 1. Let R be a stable rational p × q matrix function. With R we associate the (p + q) × (p + q) matrix function W given by ¯ ∗ JG(λ), W (λ) = G(−λ) (

where J=

Ip 0

0 −Iq

)

( ,

G(λ) =

Ip 0

(18.23) R(λ) Iq

) .

(18.24)

Note that J is a (p + q) × (p + q) signature matrix. The fact that R is stable implies that G and G−1 are analytic on the closed right half plane (inﬁnity included), and hence the right-hand side of (18.23) is a left J-spectral factorization of W relative to iR. In this section we shall show that the rational Nehari problem for R relative to the imaginary axis is solvable if and only if W admits a right J-spectral factorization of W relative to iR with an additional condition on the inverse of the spectral factor. The ﬁrst step is given by the next proposition. This proposition, which does not involve realizations and does not require R to be stable, will also provide one of the main steps in the proof of Theorem 18.1 which will be given in the next section. Proposition 18.6. Let R be a proper rational p × q matrix function, and consider ¯ ∗ JG(λ), where J and G are deﬁned by (18.24). the factorization W (λ) = G(−λ)

360

Chapter 18. Application to the rational Nehari problem

Assume that W admits a right J-spectral factorization with respect to the imagi¯ ∗ JL+ (λ), with the additional property that the rational nary axis, W (λ) = L+ (−λ) q ×q matrix function in the right lower corner of L−1 + (λ) is biproper and its inverse is analytic on the closed left half plane. Then the rational Nehari problem for R relative to the imaginary axis is solvable. Moreover, all solutions can be obtained in the following way. Partition L−1 + (λ) as a 2 × 2 block matrix function, L−1 + (λ)

=

Y11 (λ)

Y12 (λ)

Y21 (λ)

Y22 (λ)

,

(18.25)

where Y22 (λ) has size q × q. Then all solutions K of the rational Nehari problem for R relative to the imaginary axis are given by −1 K(λ) = − Y11 (λ)H(λ) + Y12 (λ) Y21 (λ)H(λ) + Y22 (λ) ,

(18.26)

where H is any rational p × q matrix function which has all its poles in the open right half plane and satisﬁes H∞ < 1. Finally, there is a one-to-one correspondence between the solution K and the free parameter H. Proof. We shall apply the results of the previous section. Put Ip R(λ) Θ(λ) = L(λ)−1 . 0 Iq Then Θ ∈ Rat(p+q)×(p+q) and Θ is J-unitary on the imaginary axis. Introduce V (λ) = L−1 + (λ). Then Ep −R V = Θ, 0 Eq −1 and thus (18.16) is satisﬁed. From V = L−1 + and the properties of L+ and L+ we see that V satisﬁes all conditions necessary to apply Theorem 18.4. Thus − K ∈ Ratp×q FV Ratp×q | R − K∞ < 1 . + +, B =

This proves that (18.26) indeed describes the set of all solutions of the rational Nehari problem for R relative to the imaginary axis. Since FV is one-to-one on Ratp×q +, B , by Theorem 18.3, we also obtain the one-to-one correspondence between the solutions K and the free parameter H. In Proposition 18.6 we have that W admits a J-spectral factorization W (λ) = ¯ ∗ JL+ (λ) with the additional property that the q × q matrix function in L+ (−λ) the right lower corner of L−1 + is biproper and has an analytic inverse on the closed left half plane. This property, which involves an inverse of a block of the inverse of L+ , can be replaced by the following more simple condition: the p × p matrix

18.4. Proof of the main result

361

function in the left upper corner of L+ is biproper and its inverse is analytic in the closed left half plane. To see this, write ⎤ ⎤ ⎡ ⎡ L11 (λ) L12 (λ) X11 (λ) X12 (λ) ⎦, ⎦. ⎣ L−1 L+ (λ) = ⎣ + (λ) = L21 (λ) L22 (λ) X21 (λ) X22 (λ) A straightforward Schur complement argument gives that L−1 11 is analytic in the −1 closed left half plane if and only if X22 is analytic in the closed left half plane. Indeed, from Section 2.2 in [20] we have that −1 (λ) X22

=

L22 (λ) − L21 (λ)L−1 11 (λ)L12 (λ),

L−1 11 (λ)

=

−1 X11 (λ) − X12 (λ)X22 (λ)X21 (λ).

This observation will be used in the ﬁnal chapter to smoothen the phrasing of several theorems.

18.4 Proof of the main result Proof of Theorem 18.1. We split the proof into ﬁve parts. Throughout this section we take γ = 1. As has been explained in the ﬁnal paragraph of Section 18.1, this can be done without loss of generality. Furthermore, in what follows R is the strictly proper p × q rational matrix function given by formula (18.2). Part 1. Let K be a solution of the rational Nehari problem for R relative to the imaginary axis. Deﬁne F to be the p × q rational matrix function on iR given by F (iλ) = K(iλ) − R(iλ). Note that F is continuous on the imaginary axis, limλ∈R, |λ|→∞ F (iλ) exists and is equal to a p × q matrix D, say. Furthermore, F ∞ = sup F (iλ) < 1. λ∈R

Now, since K is analytic, the Hankel operator generated by F is equal to the Hankel operator generated by −R, that is, HF = HK−R = HR and HF < 1 (see, e.g., Section XII.2 in [51]). So HR < 1, and hence, by Lemma 18.2, the matrix I − P 1/2 QP 1/2 is positive deﬁnite. In the remaining Parts 2–5 of the proof it is assumed that I − P 1/2 QP 1/2 is positive deﬁnite. We show that under this condition the Nehari problem is solvable and we derive all its solutions. The main work is done in Parts 3 and 4. Part 2 has a preliminary character, and in Part 5 we ﬁnish the proof by applying Proposition 18.6. Part 2. As a ﬁrst step we show that I − P 1/2 QP 1/2 is positive deﬁnite implies that I − Q1/2 P Q1/2 is positive deﬁnite too. To see this, we argue as follows. Introduce T = Q1/2 P 1/2 . Clearly I − T ∗ T is positive deﬁnite, and hence T is a

362

Chapter 18. Application to the rational Nehari problem

strict contraction (i.e., T < 1). But then so is T ∗ = P 1/2 Q1/2 . Thus, as desired, I − Q1/2 P Q1/2 is positive deﬁnite. Next, put K = Z −1 Q, where Z = I − QP while Q and P are the unique solutions to the Lyapunov equations (18.3). Note that Z is invertible, because the matrix I − P 1/2 QP 1/2 is positive deﬁnite. We claim that K is nonnegative and that the following identity holds: KA + A∗ K = KBB ∗ K − Z −1 C ∗ CZ −∗ .

(18.27)

To prove that K is nonnegative , we use ZQ1/2 = (I − QP )Q1/2 = Q1/2 (I − Q1/2 P Q1/2 ). This yields Z −1 Q1/2 = Q1/2 (I − Q1/2 P Q1/2 )−1 , and hence K = Z −1 Q = Q1/2 (I − Q1/2 P Q1/2 )−1 Q1/2 ≥ 0.

(18.28)

To prove (18.27) we ﬁrst multiply the second identity in (18.3) from the left by Z −1 and from the right by Z −∗ . Using K = Z −1 Q = QZ −∗ , this yields KAZ −∗ + Z −1 A∗ K = −Z −1 C ∗ CZ −∗ . Now observe that KAZ −∗

= KA(I − P Q)−1 = KA I + P (I − QP )−1 Q = KA + KAP Z −1 Q = KA + KAP K.

But then, taking advantage of the ﬁrst identity in (18.3) , we obtain KAZ −∗ + Z −1 A∗ K

= KA + A∗ K + K(AP + A∗ P )K = KA + A∗ K − KBB ∗ K.

Thus KA + A∗ K

=

KAZ −∗ + Z −1A∗ K + KBB ∗ K

=

KBB ∗ K − Z −1 C ∗ CZ −∗ ,

which proves (18.27). ¯ ∗ JG(λ), where J and G are deﬁned by (18.24). It Part 3. Put W (λ) = G(−λ) was already observed that this factorization is a left J-spectral factorization with respect to iR. In this part we prove that W also admits a right J-spectral factorization with respect to iR. To do this we use that I − P 1/2 QP 1/2 is positive deﬁnite and apply Theorem 14.14 with L− (λ) = G(λ).

18.4. Proof of the main result

363

Employing the realization (18.2) of R, one gets Ip R(λ) Ip 0 C L− (λ) = = + (λ − A)−1 0 B . 0 0 Iq 0 Iq So, with = A, A

= B

0 B

,

= C

C

,

0

and the associate main matrix of this − A) −1 B, we have L− (λ) = Ip+q + C(λ × = A −B C obviously coincides with the main matrix A = A. realization A the solutions of For the realization considered here we denote by P and Q are the the equations (14.53) and (14.52), respectively. In other words, P and Q unique solutions of AP + PA∗ = BB ∗ ,

+ QA = C ∗ C. A∗ Q

= −Q. It follows that I − PQ = I − P Q, and therefore (ii) So P = −P and Q 1/2 1/2 implies that I − P Q = (I −P Q) = P (I −P QP 1/2 )P −1/2 is invertible. Hence P is invertible too. I −Q Thus by Theorem 14.14 the rational (p + q) × (p + q) matrix function W ¯ ∗ JL+ (λ), with respect to admits a right J-spectral factorization, W (λ) = L+ (−λ) iR. In fact, for L+ one can take −CP Ip 0 (18.29) + Z −1 (λ + A∗ )−1 C ∗ QB , L+ (λ) = ∗ B 0 Iq where Z = I − QP . Theorem 14.14 also tells us that for this choice of the right J-spectral factor L+ we have −CP Ip 0 −1 − (λ + A∗ )−1 Z −1 C ∗ QB , L+ (λ) = (18.30) B∗ 0 Iq where, as before, Z = I − QP . Now partition L−1 + (λ) as V (λ) =

L−1 + (λ)

=

X11 (λ)

X12 (λ)

X21 (λ)

X22 (λ)

,

(18.31)

where the block in the right lower corner has size q × q. Comparing (18.30) and (18.31) we see that the rational matrix functions Xij , i, j = 1, 2, are precisely the functions given by (18.4)– (18.7). Part 4. In this part, again assuming I − P 1/2 QP 1/2 to be positive deﬁnite, we show that the q × q rational matrix function X22 (λ) in the right lower corner of

364

Chapter 18. Application to the rational Nehari problem

the block matrix in (18.31) has precisely the properties which will allow us to apply Proposition 18.6. Obviously, X22 is biproper. Since the eigenvalues of A are in the open left half plane, those of −A∗ are in the open right half plane as well, and hence X22 is −1 is also analytic analytic on the closed left half plane. It remains to show that X22 on the closed left half plane. From the expression for X22 (λ) we see that −1 X22 (λ) = I + B ∗ (λ − A0 )−1 Z −1 QB,

where A0 = −A∗ + Z −1 QBB ∗ = −A∗ + KBB ∗ , with K as in Part 2 of the present −1 proof. Thus, in order to show that X22 is analytic on the closed left half plane, it suﬃces to show that A0 has all its eigenvalues in the open right half plane. To determine the location of the eigenvalues of A0 we ﬁrst prove that A0 K + KA∗0 = KBB ∗ K + Z −1 C ∗ CZ −∗ .

(18.32)

This identity follows from (18.27). Indeed, using the deﬁnition of A0 , we have A0 K = (−A∗ + KBB ∗ )K = −A∗ K + KBB ∗ K. But then, using (18.27), we see that A0 K + KA∗0 = −A∗ K − KA + 2KBB ∗ K = KBB ∗ K + Z −1 C ∗ CZ −∗ , which proves (18.32). The identity (18.32) implies that A0 does not have pure imaginary eigenvalues. Indeed, suppose A0 has a pure imaginary eigenvalue. Then the same holds true for A∗0 , that is, there is a pure imaginary λ0 and a non-zero vector x such that A∗0 x = λ0 x. This implies x∗ A0 = −λ0 x∗ , and hence x∗ (A0 K + KA∗0 )x = 0. From (18.32) it then follows that x∗ KBB ∗ Kx = 0. In other words, x∗ KB = 0. Using the deﬁnition of A0 , we see that −λ0 x∗ = x∗ A0 = −x∗ A∗ + x∗ KBB ∗ = −x∗ A∗ . We conclude that A∗ has a pure imaginary eigenvalue which is impossible because by assumption A (and hence A∗ too) has all its eigenvalues in the open left half plane. Thus a contradiction has been obtained, and we conclude that A0 has no pure imaginary eigenvalue. It remains to show that A0 has no eigenvalues in the open left half plane. If K would be invertible, then K would be positive deﬁnite, and the statement that A0 has no eigenvalues in the open left half plane would now follow immediately from A0 K + KA∗0 ≥ 0 and the classical Carlson-Schneider inertia theorem (see Theorem 13.1.3 in [107]). However since K may not be invertible an additional argument is required, which will be presented in the next two paragraphs. Let n be the order of the square matrix A. Note that K, Q, and Z are also square matrices of order n. Put X1 = Im K and X2 = Ker K. Since K is selfadjoint,

18.4. Proof of the main result

365

we have the orthogonal direct sum decomposition Cn = X1 ⊕ X2 . The identity K = Z −1 Q implies that Ker Q = Ker K. Hence, by selfadjointness, Im Q = Im K. It follows that relative to the decomposition Cn = X1 ⊕ X2 the matrices K and Q admit the following 2 × 2 block matrix representation: Q1 0 K1 0 , Q= , K= 0 0 0 0 where both K1 and Q1 are positive deﬁnite. Next, we partition A, B, and C relative to the decomposition Cn = X1 ⊕ X2 . This yields B1 A11 0 , B= , C = [ C1 0 ]. A= B2 A21 A22 Here we used that X2 = Ker Q = Ker (C|A), which implies that X2 is A-invariant and that C is zero on X2 . From ZK = Q, we see that Z[Im K] = Im Q, and hence Z[X1 ] = X1 . Thus. relative to Cn = X1 ⊕ X2 , the matrix Z partitions as Z11 Z12 , Z= 0 Z22 where both Z11 and Z22 are invertible. Employing the block matrix representations for K, A and B we compute A0 . We have ∗ K1 0 B1 B1∗ B1 B2∗ A11 A∗21 + . A0 = − 0 A∗22 0 0 B2 B1∗ B2 B2∗ Thus A0 has the form

A0 =

A0,11

0

A0,22

,

where A0,11 = −A∗11 + K1 B1 B1∗ and A0,22 = A∗22 . Since A has all its eigenvalues in the open left half plane, the same holds true for A22 . Hence A0,22 has all its eigenvalues in the open right half plane. Thus, in order to prove that A0 has all its eigenvalues in the open right half plane, it suﬃces to show that A0,11 has this property. This will be done in the next paragraph. Since A0 has no pure imaginary eigenvalue, the same holds true for A0,11 . From (18.32), using the block matrices in the previous paragraph, we see that A0,11 K1 + K1 A∗0,11 ≥ 0. As K1 is positive deﬁnite we can now apply the CarlsonSchneider inertia theorem (i.e., Theorem 13.1.3 in [107]) to show that the inertia of A0,11 is equal to the inertia of K1 . Using again that K1 is positive deﬁnite, it follows that all the eigenvalues of A0,11 are in the open right half plane, as desired. Part 5. We are now ready to complete the proof. Assume I −P 1/2 QP 1/2 is positive deﬁnite. By the previous two parts of the proof, the rational matrix function

366

Chapter 18. Application to the rational Nehari problem

¯ ∗ JG(λ) admits a right J-spectral factorization with respect to the W (λ) = G(−λ) ¯ ∗ JL+ (λ), with the additional property imaginary axis, written W (λ) = L+ (−λ) that the q × q matrix function in the right lower corner of L−1 + (λ) is biproper and its inverse is analytic on the closed left half plane. It was also shown that L−1 + (λ) partitions as X11 (λ) X12 (λ) −1 L+ (λ) = , X21 (λ) X22 (λ) where the rational matrix functions Xij , i, j = 1, 2, are precisely the functions given by (18.4)– (18.7). But then we can apply Proposition 18.6 to get the desired description of all solutions. .

18.5 The case of a non-stable given function In this section we return to the general case, where the rational p × q matrix function R is not necessarily stable, i.e., does not necessarily have all its poles in the open left half plane. Throughout we assume R to be proper and to have no poles on the imaginary axis. Write R = R− + R+ , where R− is a stable rational p × q matrix function which is strictly proper, and R+ is a proper rational p × q matrix function which has all its poles in the open right half plane. The required location of the poles determines R− and R+ uniquely. Recall that we seek proper rational p × q matrix functions K such that K has all its poles in the open right half plane and (18.33) R − K∞ = R− − (K − R+ )∞ < γ. The second term in (18.33) gives us a hint of how to solve the Nehari problem for R. In fact, from (18.33) we see that K is a solution to the Nehari problem with tolerance γ for R if and only if K − R+ is a solution to the Nehari problem with tolerance γ for R− . This remark allows us to extend Theorem 18.1 to the case when the given function R is non-stable. To describe the resulting theorem, we shall assume that R− and R+ are given in the form R− (λ) = C− (λIn − A− )−1 B− ,

R+ (λ) = D + C+ (λIn − A+ )−1 B+ , (18.34)

where A− has all its eigenvalues in the open left half plane, and A+ has all its eigenvalues in the open right half plane. In the situation where the realizations in (18.34) are minimal, these conditions on the location of the spectra of A− and A+ are automatically fulﬁlled. Put ∞ ∞ ∗ τ A− ∗ τ A∗ ∗ − e B− B− e dτ, Q− = eτ A− C− C− eτ A− dτ. (18.35) P− = 0

0

Note that P− and Q− are well-deﬁned because all the eigenvalues of A− are in the open left half plane. The following theorem is the main result of this section.

18.5. The case of a non-stable given function

367

Theorem 18.7. Let R = R− + R+ with R− and R+ being given by (18.34). Assume A− and A+ have all their eigenvalues in the open left and open right half plane, respectively, and let P− and Q− be given by (18.35). Then the rational Nehari problem for R relative to the imaginary axis with tolerance γ is solvable if and 1/2 1/2 only if the matrix γ 2 In − P− Q− P− is positive deﬁnite. In this case the matrix 2 Z− = γ In − P− Q− is invertible and all solutions of the Nehari problem under consideration can be obtained in the following way. Introduce rational matrix functions Yij , i, j = 1, 2, by setting ∗ −C+ C− P− + DB− Ip −D Y11 (λ) Y12 (λ) + (18.36) = ∗ Y21 (λ) Y22 (λ) 0 Iq 0 −B− %−1 $ ∗ 0 B+ A+ −B+ B− . . λI2n − −1 ∗ −1 0 −A∗− Z− C− Z− Q− B− Then all solutions K to the rational Nehari problem for R relative to the imaginary axis with tolerance γ are given by −1 K(λ) = − Y11 (λ)H(λ) + Y12 (λ) Y21 (λ)H(λ) + Y22 (λ) , (18.37) where H is any rational p × q matrix function which has all its poles in the open right half plane and satisﬁes H∞ < γ. Moreover, there is a one-to-one correspondence between the solution K and the free parameter H. Proof. From Theorem 18.1 we know that the Nehari problem with tolerance γ for 1/2 1/2 R− is solvable if and only if the matrix γ 2 I − P− Q− P− is positive deﬁnite. On the other hand, we also know (see the second paragraph of this section) that the Nehari problem with tolerance γ for R is solvable if and only if the Nehari problem with tolerance γ for R+ is solvable. These two “if and only if” statements together yield the ﬁrst part of the theorem. 1/2 1/2 Next, assume that the matrix γ 2 I − P− Q− P− is positive deﬁnite. As we have already seen, K is a solution to the Nehari problem with tolerance γ for R + R+ , where K is an arbitrary solution to the if and only if K is of the form K Nehari problem with tolerance γ for R− . By Theorem 18.1, applied to R− in place of R, the latter solutions are given by −1 , K(λ) = − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) with the coeﬃcients in this linear fractional representation given by X11 (λ)

=

−1 ∗ Ip + C− P− (λ + A∗− )−1 Z− C− ,

X12 (λ)

=

−1 C− P− (λ + A∗− )−1 Z− Q− B− ,

X21 (λ)

=

−1 ∗ ∗ −B− (λ + A∗− )−1 Z− C− ,

X22 (λ)

=

−1 ∗ Iq − B− (λ + A∗− )−1 Z− Q− B − ,

368

Chapter 18. Application to the rational Nehari problem

where Z− = γ 2 I − P− Q− , which is invertible. It follows that K(λ)

= = =

=

=

R+ (λ) + K(λ) −1 R+ (λ) − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) −1 R+ (λ) X21 (λ)H(λ) + X22 (λ) X21 (λ)H(λ) + X22 (λ) −1 − X11 (λ)H(λ) + X12 (λ) X21 (λ)H(λ) + X22 (λ) − X11 (λ) − R+ (λ)X21 (λ) H(λ) + X12 (λ) − R+ (λ)X22 (λ) −1 . X21 (λ)H(λ) + X22 (λ) −1 − Y11 (λ)H(λ) + Y12 (λ) Y21 (λ)H(λ) + Y22 (λ) ,

where H is any rational p × q matrix function having all its poles in the open right half plane and satisﬁes H∞ < γ. Moreover, the coeﬃcient matrix Y11 (λ) Y12 (λ) Y (λ) = Y21 (λ) Y22 (λ) is given by

Y (λ) =

Ip

−R+ (λ)

0

Iq

X11 (λ)

X12 (λ)

X21 (λ)

X22 (λ)

.

Now, using the formulas for Xij , i, j = 1, 2, one gets Ip 0 C − P− X11 (λ) X12 (λ) −1 ∗ = + (λ + A∗− )−1 Z− C− ∗ −B− X21 (λ) X22 (λ) 0 Iq

−1 Z− Q− B − .

Furthermore, employing the realization of R+ , Ip −R+ (λ) Ip −D −C+ (λ − A+ )−1 0 B+ . = + 0 Iq 0 Iq 0 Taking the product of these realizations (see Theorem 2.5) we reach the conclusion that the coeﬃcient matrix Y (λ) admits the desired realization (18.36). The fact that there is one-to-one correspondence between the solution K and the free parameter H in (18.37) follows directly from the corresponding result in Theorem 18.1.

18.6 The Nehari-Takagi problem In the Nehari-Takagi problem the given function R is the same as in the Nehari problem. However the solutions K are allowed to come from a wider class. To

18.6. The Nehari-Takagi problem

369

be more more speciﬁc, let the rational p × q matrix function R be as in the ﬁrst paragraph of Section 18.1. Thus R is proper and does not have a pole on the imaginary axis. Let κ be a non-negative integer. Then the (rational) NehariTakagi problem (relative to the imaginary axis) is the problem of ﬁnding all proper rational p × q matrix functions K such that K has no pole on the imaginary axis and at most κ poles in the open left half plane (multiplicities taken into account), and K − R∞ = sup K(s) − R(s) < γ, (18.38) s∈iR

where γ is a pre-speciﬁed positive number. When κ = 0, the conditions on K reduce to the requirement that K has all its poles in the open right half plane. Thus with κ = 0 the Nehari-Takagi problem is just the Nehari problem considered in the preceding sections. In this section we take γ = 1, which can be done without loss of generality (cf., the last paragraph of Section 18.1), and we assume that R is strictly proper and stable. Thus R admits a realization R(λ) = C(λIn − A)−1 B where A has all its eigenvalues in the open left half plane. The following result is the analogue of Theorem 18.1 for the Nehari-Takagi problem. Theorem 18.8. Let (λ) = C(λIn − A)−1 B be a realization of the rational p × q matrix function R, assume A has all its eigenvalues in the open left half plane, and let ∞ ∞ ∗ sA ∗ sA∗ P = e BB e ds, Q= esA C ∗ CesA ds 0

0

(i.e., P and Q are the controllability and observability gramians corresponding to the given realization). Suppose In − P Q is invertible. Then the rational NehariTakagi problem for R relative to the imaginary axis with γ = 1 is solvable if and only if the matrix P Q has at most κ eigenvalues (multiplicities taken into account) larger than 1. Moreover, if κ0 is the number of eigenvalues of P Q larger than 1, then all solutions K of the Nehari-Takagi problem for R relative to the imaginary axis with γ = 1 such that K has precisely κ0 poles in the open left half plane are given by the linear fractional formula −1 . (18.39) K(λ) = − Θ11 (λ)G(λ) + Θ12 (λ) Θ21 (λ)G(λ) + Θ22 (λ) Here the free parameter G is an arbitrary rational p × q matrix function which has all its poles in the open right half plane and G∞ < 1. Furthermore, the coeﬃcients Θij , i, j = 1, 2, are given by Θ11 (λ)

= Ip + CP (λIn + A∗ )−1 (In − QP )−1 C ∗ ,

Θ12 (λ)

= CP (λIn + A∗ )−1 (In − QP )−1 QB,

Θ21 (λ)

= −B ∗ (λIn + A∗ )−1 (In − QP )−1 C ∗ ,

Θ22 (λ)

= Iq − B ∗ (λIn + A∗ )−1 (In − QP )−1 QB.

370

Chapter 18. Application to the rational Nehari problem

To prove the above theorem one can follow the same line of reasoning as used in this chapter to prove Theorem 18.1. The role of Theorem 18.4 has to be taken over by Theorem 18.5. For further details we refer to the literature; see for example [86] and the references therein.

Notes The Nehari problem has its roots in the classical papers of Nehari [114] and Adamjan-Arov-Krein [1], [2]. The rational matrix version played an important role in the early development of H-inﬁnity control theory; see, e.g., the lecture notes [43]. Here one already ﬁnds the J-spectral factorization approach. For an overview of the various methods to deal with the matrix Nehari problem we refer to the notes to Chapter 20 in [7]. The Takagi version of the Nehari problem has its roots in [142]. The result with a full proof can also be found in Section 20.5 of [7]. For an abstract approach to the Nehari-Takagi problem, covering applications to time-invariant inﬁnite-dimensional systems and time-varying ﬁnite-dimensional linear systems, we refer to [86].

Chapter 19

Review of some control theory for linear systems In this chapter a brief survey is given of a number of basic elements of control and mathematical systems theory. The main aim is to give the reader some understanding for the type of problems that will be treated in the ﬁnal chapter. The chapter consists of two sections. Section 19.1 introduces the concepts of stability of systems and the method of feedback to stabilize a system. Section 19.2 deals with the notion of internal stability of a closed loop system. In particular the Youla-Jabr-Bongiorno parametrization of all stabilizing compensators is presented.

19.1 Stability and feedback In this section we consider a causal input-output system Σ as in the ﬁgure below:

u

Σ

y

As usual (cf., Section 2.1) the symbol u denotes the input and y the output. Mathematically input and output are vector-valued functions of a (time) parameter t. Such an input-output system is called externally stable or bounded-input boundedoutput stable (BIBO-stable) if a bounded input u produces a bounded output y, that is, supt≥0 u(t) < ∞ implies supt≥0 y(t) < ∞.

372

Chapter 19. Review of some control theory for linear systems

Now let us assume that Σ is a causal linear time invariant system given by the following ﬁnite dimensional state space representation: x (t) = Ax(t) + Bu(t), (19.1) y(t) = Cx(t) + Du(t), t ≥ 0. Here A, B, C, D are matrices of appropriate sizes, and A is a square matrix. We refer to (19.1) as a realization of the system. The realization (19.1) is called stable if for any initial value x(0), with zero input u, the state x(t) will go to zero if t → ∞. It is easily seen that stability of the realization (19.1) is equivalent to the requirement that the matrix A has all its eigenvalues in the open left half plane. If the latter holds, A is said to be a stable matrix . Given (19.1) the eﬀect of inputs on outputs can be described in the time domain by a lower triangular integral operator y(t) = CetA x(0) +

t 0

k(t − s)u(s) ds + Du(t),

(19.2)

where k(t) is the so-called impulse response function. As we have already seen in Section 2.1, in the frequency domain with x(0) = 0 the connection between input and output is given by y(λ) = W (λ) u(λ), where W is the transfer function of the system, and u and y denote the Laplace transforms of the input u and the output y, respectively. In terms of (19.1) we have k(t) = CetA B,

W (λ) = D + C(λ − A)−1 B.

(19.3)

From (19.2) and the ﬁrst identity in (19.3) it is clear that stability of the realization (19.1) implies external stability of the corresponding system. The converse is also true when the realization is minimal, that is, when the pair (A, B) is controllable and the pair (C, A) is observable. We summarize this and related results in the following theorems. Theorem 19.1. Let (19.1) be a minimal realization, then the corresponding system is externally stable if and only if the realization is stable. Theorem 19.2. Let k be the impulse response function and let W be the transfer function of the linear time invariant system given by (19.1). The following statements are equivalent: 1. The system given by (19.1) is externally stable; ∞ 2. k(t) dt < ∞; 0

3. The rational matrix function W is iR-stable, that is, W has all its poles in the open left half plane.

19.1. Stability and feedback

373

An important issue is stabilizing an unstable system. The simplest method is that of static state feedback. To explain this method consider the system given by the state space representation: x (t) = Ax(t) + Bu(t), y(t)

=

t ≥ 0.

x(t),

Note that the output is equal to the state. This case is sometimes referred to as the full information case. The problem is to ﬁnd a static feedback control law u(t) = F x(t) + v(t) that will make the system sending v to x stable. That is, to ﬁnd a matrix F of appropriate size such that x (t) = (A + BF )x(t) + Bv(t) is stable. This amounts to requiring that the matrix A + BF is stable, i.e., all its eigenvalues are in the open left half plane. For such a matrix F to exist the pair (A, B) should be stabilizable in the sense of Section 13.2. Two questions appear: ﬁrst, when is a pair of matrices (A, B) stabilizable, and second, how to construct a stabilizing matrix F ? We start with an observation concerning the so-called single input case. In that situation, the matrix B is an n × 1 vector, and one may assume without loss of generality that A and B have the form ⎡ ⎤ ⎤ ⎡ 0 0 1 ⎢ ⎥ ⎥ ⎢ .. ⎢ ... ⎥ ⎥ ⎢ . ⎢ ⎥ ⎢ B=⎢ ⎥ A=⎢ ⎥, ⎥. ⎣0⎦ ⎣ 1 ⎦ −an

···

Consider F = [ fn · · · f1 ]. Then ⎡ ⎢ ⎢ A + BF = ⎢ ⎢ ⎣

···

0

−a1

1

1

⎤ ..

fn − an

···

.

···

1 f1 − a1

⎥ ⎥ ⎥. ⎥ ⎦

So, in this case, any polynomial can be obtained as the characteristic polynomial of A + BF by an appropriate choice of F . Next we make a second observation. Let A be an n × n matrix, let B be an n × m matrix, and write Cn = Im (A|B) X0 . With respect to this direct sum decomposition, the matrices A and B can be written as A11 A10 B1 , A= , B= 0 A00 0

374

Chapter 19. Review of some control theory for linear systems

with (A11 , B1 ) controllable. Thus, for any m × n matrix F = F1 A11 + B1 F1 A10 + B1 F0 , A + BF = 0 A00

F0 , one has

and hence σ(A+BF ) = σ(A11 +B1 F1 ) ∪ σ(A00 ). Note that σ(A00 ), the second part in the right-hand side of the preceding identity, is independent of the particular choice of X0 and also of the choice of F . Therefore the eigenvalues of A00 are called the uncontrollable eigenvalues of A relative to the matrix B. Clearly, A has no uncontrollable eigenvalues relative to B if and only if the pair (A, B) is controllable. From the discussion in the previous paragraph we conclude that, in order for (A, B) to be stabilizable, it is necessary that the uncontrollable eigenvalues of A relative to B are in the open left half plane. The converse of this observation would follow if any controllable pair is stabilizable. This is the case for single input as we have already seen. That it is true in general appears from the next result which is actually quite a bit stronger, and is known as the pole placement theorem. Theorem 19.3. Let A be an n × n matrix, and let B be an n × m matrix. The following two statements are equivalent: (i) The pair (A, B) is controllable; (ii) For any scalar polynomial p(λ) = λn + p1 λn−1 + · · · + pn−1 λ + pn , there is an m × n matrix F such that the characteristic polynomial of A + BF coincides with p. Corollary 19.4. Let A be an n × n matrix and let B be an n × m matrix. The pair (A, B) is stabilizable if and only if the uncontrollable eigenvalues of A relative to the matrix B are in the open left half plane. Let A be an n × n matrix and let C be an m × n matrix. The pair (C, A) is called detectable when there exists an n × m matrix R such that A − RC is stable. In other words the pair (C, A) is detectable if and only if the pair (A∗ , C ∗ ) is stabilizable. By deﬁnition the unobservable eigenvalues of A relative to C are the uncontrollable eigenvalues of A∗ relative to C ∗ . It is also possible to give a direct deﬁnition of the latter notion, involving a decomposition of the type Cn = Ker (C|A) X0 . From the above deﬁnitions and Corollary 19.4 it is clear that the pair (C, A) is detectable if and only if the unobservable eigenvalues of A relative to the matrix C are in the open left half plane.

19.2 Parametrization of internally stabilizing compensators In this section G is the transfer function of a system Σ with two inputs u and w, and two outputs y and z. Here u is the control input, w a disturbance, y is the

19.2. Parametrization of internally stabilizing compensators

375

output which can be measured and z is the output to be controlled. Throughout, we shall assume that the system Σ is given in state space form as follows: ⎧ x (t) = Ax(t) + B1 w(t) + B2 u(t), ⎪ ⎪ ⎨ z(t) = C1 x(t) + D1 u(t), (19.4) ⎪ ⎪ ⎩ t ≥ 0. y(t) = C2 x(t) + D2 w(t), It will be convenient to rewrite the realization (19.4) in the form ⎧ ⎪ w(t) ⎪ ⎪ , x (t) = Ax(t) + B1 B2 ⎪ ⎪ ⎪ u(t) ⎨ ⎪ ⎪ z(t) ⎪ ⎪ ⎪ = ⎪ ⎩ y(t)

C1 C2

x(t) +

0

D1

D2

0

w(t) u(t)

.

From the latter representation we see that the transfer function of (19.4) is given by G11 (λ) G12 (λ) 0 D1 C1 G(λ) = = + (λ − A)−1 B1 B2 . D2 0 C2 G21 (λ) G22 (λ) In particular, the transfer function G22 is strictly proper. Let C be a causal ﬁnite dimensional linear time invariant system of the type considered in the previous section, and let K be its transfer function. Thus K is a proper rational matrix function To deﬁne what it means that C is an internally stabilizing compensator for Σ we introduce two additional inputs v1 and v2 as in the following ﬁgure:

These two additional inputs are regarded as disturbances: v1 is a disturbance on the control input u, while v2 is a disturbance on the measured output. Then the system

376

Chapter 19. Review of some control theory for linear systems

C with transfer function K is said to be an internally stabilizing compensator for the system Σ if the nine transfer functions from the disturbances w, v1 , v2 to z, u and y are all stable rational matrix functions. In this case, by slight abuse of terminology, we shall also say that K is an internally stabilizing compensator for the transfer function G of Σ. After Laplace transform, the nine transfer functions from the disturbances w, v1 , v2 to z, u and y are given by ⎡

zˆ

⎤

⎡

I

⎢ ⎥ ⎢ ⎢ u ⎥ ⎢ ⎣ ˆ ⎦=⎣ 0 0 yˆ

−G12 I −G22

0

⎤−1 ⎡

⎥ −K ⎥ ⎦ I

⎢ ⎢ ⎣

G11

0

0

I

G21

0

0

⎤⎡

w ˆ

⎤

⎥ ⎥⎢ ⎥ ⎢ 0 ⎥ ⎦ ⎣ vˆ1 ⎦ . I vˆ2

(19.5)

Now G22 is strictly proper and K is proper. Hence the rational matrix functions I − G22 (λ)K(λ) and I − K(λ)G22 (λ) are biproper with the value I at inﬁnity. It follows that the inverses I − G22 K and I − KG22 are well-deﬁned. Using these facts, the product of the ﬁrst two matrices in the right-hand side of the identity (19.5)can be computed as ⎡ ⎢ ⎢ ⎣

G11 + G12 K(I − G22 K)−1 G21

G12 (I − KG22 )−1

G12 K(I − G22 K)−1

K(I − G22 K)−1 G21

(I − KG22 )−1

K(I − G22 K)−1

(I − G22 K)−1 G21

G22 (I − KG22 )−1

(I − G22 K)−1

⎤ ⎥ ⎥. ⎦

Theorem 19.5. Let G be the transfer function of the system Σ given by (19.4), and let C be a causal ﬁnite dimensional linear time invariant system whose transfer function K is a proper rational matrix function. Then C is an internally stabilizing compensator for Σ if and only if K stabilizes G22 in the sense that the transfer functions from v1 and v2 to u and y are stable rational matrix functions, that is, the four functions (I − KG22 )−1 ,

K(I − G22 K)−1 ,

G22 (I − KG22 )−1 ,

(I − G22 K)−1 ,

are stable. There is a beautiful parametrization of all internally stabilizing compensators, known as the Youla-Jabr-Bongiorno parametrization. In order to state the parametrization we need a doubly coprime factorization of G22 , that is, a factorization !(λ)−1 N (λ), G22 (λ) = N (λ)M (λ)−1 = M (19.6) and M ! are iR-stable rational matrix functions of appropriate sizes, where N, M, N with the additional property that there exist iR-stable rational matrix functions

19.2. Parametrization of internally stabilizing compensators and Y such that X, Y, X X(λ) −Y (λ) M (λ) (λ) −N

!(λ) M =

N (λ) M (λ)

Y (λ)

N (λ)

X(λ)

Y (λ)

377

X(λ)

X(λ)

−Y (λ)

(λ) −N

!(λ) M

(19.7)

= I.

Such a factorization always exists, in fact we can readily give formulas for all matrix functions involved in terms of the realization of G22 . To do this we assume that the realization G22 (λ) = C2 (λI − A)−1 B2 , has two additional properties, namely (C2 , A) is detectable, and (A, B2 ) is stabilizable. That is, there exist matrices F and H such that the matrices AF = A + B2 F and AH = A + HC2 are both stable. Then, one choice of a doubly coprime factorization is given by the functions ⎧ −1 N (λ) = C2 (λ − AF )−1 B2 , ⎪ ⎪M (λ) = I + F (λ − AF ) B2 , ⎪ ⎪ ⎪ ⎪ !(λ) = I + C2 (λ − AH )−1 H, (λ) = C2 (λ − AH )−1 B2 , ⎨M N (19.8) ⎪ ⎪ Y (λ) = −F (λ − AF )−1 H, X(λ) = I − C2 (λ − AF )−1 H, ⎪ ⎪ ⎪ ⎪ ⎩ Y (λ) = −F (λ − AH )−1 H. X(λ) = I − F (λ − AH )−1 B2 , Next, we give the Youla-Jabr-Bongiorno parametrization, which describes all internally stabilizing compensators of Σ in terms of iR- stable, proper rational matrix functions in a one-to-one way. Theorem 19.6. Let G be the transfer function of the system Σ given by (19.4), and let M , N , X, Y be the iR-stable rational matrix functions related to the doubly coprime factorization of G22 . Let C be a causal ﬁnite dimensional linear time invariant system whose transfer function K is a proper rational matrix function. Then C is an internally stabilizing compensator of Σ if and only if K has the form K(λ)

=

−1 Y (λ) − M (λ)Q(λ) X(λ) − N (λ)Q(λ) ,

(19.9)

where Q is an iR-stable rational matrix function. Moreover, the map from Q to K is one-to-one. !, N , X, Y we have the following alternative Replacing M , N , X, Y by M expression for the transfer function K of the compensator: !(λ) . (λ) −1 Y (λ) − Q(λ)M K(λ) = X(λ) − Q(λ)N

378

Chapter 19. Review of some control theory for linear systems

Notes The results of the ﬁrst section are standard results in mathematical systems theory, see, e.g., [94] or the more recent [33], [84]. For analogous results in the discrete time case we refer to [94], Chapter 21 of [150], and to [85]. A proof of Theorem 19.5 can be found in Chapter 4 of [43]. The formulas (19.8) giving the doubly coprime factorization in state space terms were derived in [115], see also Section 4.5 in [43]. Theorem 19.6 presents a result of [148].

Chapter 20

H-inﬁnity control applications The focus of the chapter is on a part of control theory called H-inﬁnity control. The problem involved is the general H-inﬁnity control problem, the so-called standard problem. It concerns the construction of a stabilizing controller with additional constraints on the maximum of the norm of the closed loop transfer function, taken over the values of the argument on the imaginary line. In its simplest form the problem is equivalent to the rational matrix Nehari problem considered in Chapter 18. The label H-inﬁnity is related to the fact that a proper rational matrix function is stable if and only if it is analytic and uniformly bounded in the open right half plane. A function with the latter properties is usually referred to as an H∞ -function (on the right half plane). The chapter consists of four sections. Section 20.1 introduces the standard problem mentioned above, and shows how this problem can be reduced to a model matching problem. In the next two sections we discuss a one-sided model matching problem (Section 20.2) and the two-sided model matching problem (Section 20.3). In particular, it will be shown how these two problems reduce to J-spectral factorization problems involving certain rational matrix functions. All of this will be done in general terms, without any state space formulas as yet. In the ﬁnal section (Section 20.4) we use results from Chapter 14 and present the solution to the model matching problem in state space terms. This leads to the solution of the standard problem in these terms too. In this chapter, as in Section 18.2, we use the following notation: if R is a rational matrix function, then R∗ denotes the rational matrix function given by R∗ (λ) = ¯ ∗ . (In engineering literature, including [76] and [43], this function is often R(−λ) denoted by R∼ .) Recall also from Section 18.2 that Rat denotes the set of all proper rational matrix functions that are analytic on the imaginary axis. Furthermore, Ratp×q stands for the set of all F in Rat that are of size p × q, and Ratp×q + denotes the set of all F in Ratp×q that are analytic on the closed left half plane. In the present chapter we shall also use the notation Ratp×q (Rat− ) which will −

380

Chapter 20. H-inﬁnity control applications

denote the set of all F in Ratp×q (in Rat) that are analytic in the closed right half plane. In other words, F belongs to Ratp×q if and only if F is an iR stable p × q − if and only if F ∗ ∈ Ratq×p rational matrix function. Note also that F ∈ Ratp×q − + .

20.1 The standard problem and model matching Throughout this chapter G is the transfer function of a system Σ with two inputs u and w, and two outputs y and z. The input u is the control input, w is a disturbance, y is the output we can measure, and z is the output to be controlled. As in Section 19.2 we assume that the system is given by the state space representation ⎧ ⎪ ⎪ x (t) = Ax(t) + B1 w(t) + B2 u(t), ⎨ z(t) = C1 x(t) + D1 u(t), (20.1) ⎪ ⎪ ⎩ y(t) = C2 x(t) + D2 w(t), t ≥ 0. In particular, the function G is of the form G(λ)

G11 (λ)

G12 (λ)

G21 (λ)

G22 (λ)

= =

0

D1

D2

0

+

C1 C2

(λ − A)−1 B1

B2 .

(20.2)

Taking Laplace transforms and assuming the system to be at rest at t = 0 we have

z(λ) y(λ)

G11 (λ)

G12 (λ)

G21 (λ)

G22 (λ)

=

w(λ) u (λ)

.

(20.3)

Our goal is to ﬁnd a proper rational matrix function K such that: (1) K is the transfer function of an internally stabilizing compensator C of Σ (see Section 19.2), and (2) the inﬂuence of w on z is kept small in a sense we shall explain presently. Inserting u (λ) = K(λ) y (λ) into (20.3), one sees that ⎧ ⎨z(λ) = G11 (λ)w(λ) + G12 (λ)K(λ) y (λ), (20.4) ⎩y(λ) = G (λ)w(λ) + G22 (λ)K(λ) y (λ). 21 Since G22 is strictly proper, so is G22 K, and hence the determinant of the matrix I − G22 (λ)K(λ) does not vanish identically. By the second equation in (20.4) we −1 have y(λ) = I − G22 (λ)K(λ) G21 (λ)w(λ). Inserting this into the ﬁrst equation

20.1. The standard problem and model matching

381

of (20.4), we obtain that the closed loop transfer function from w to z is given by the Redheﬀer representation (20.5) z(λ) = RG (K)(λ) w(λ) =

−1 G11 (λ) + G12 (λ)K(λ) I − G22 (λ)K(λ) G21 (λ) w(λ).

The second requirement on K is that, given a tolerance γ, we want RG (K) to be in Rat and to satisfy the bound RG (K)∞ = max RG (K)(λ) < γ. λ∈iR

(20.6)

This problem is known in control theory as the standard problem of H-inﬁnity control . The approach to solving this problem using J-spectral factorization techniques starts from the Youla parametrization of internally stabilizing compensators which we reviewed in Section 19.2. This leads, as we shall see in the ﬁnal two paragraphs of this section, to an equivalent and easier to handle problem. Indeed, from the given rational matrix function G one constructs three rational matrix functions, T1 , T2 and T3 such that internally stabilizing compensators for which (20.6) holds are in one-to-one correspondence with iR-stable rational matrix functions Q for which T1 − T2 QT3 ∞ < γ. (20.7) The latter problem is called the model matching problem . It turns out that under mild assumptions (see Section 20.4 below) the rational matrix functions T1 , T2 and T3 are iR stable. In particular, these functions have no poles on the imaginary axis and at inﬁnity, and hence they are all in Rat. Furthermore, we shall see that T2 has a left inverse in Rat and T3 has a right inverse in Rat. A particular case (see the next section) of the model matching problem, when T2 is square and T3 = I, is a variation on the Nehari problem as discussed in Chapter 18. Next, we present the reduction of the standard problem to a model matching problem. All necessary calculations take place in Rat, i.e., in the set of rational matrix functions that are analytic on iR and at inﬁnity. As before, we partition the transfer function G as in the ﬁrst part of (20.3). Also we shall employ the same notation as in Section 19.2 insofar as it concerns the doubly coprime factorization of G22 in (19.6) and the parametrization of the transfer functions of the internally stabilizing compensators of the system Σ in Theorem 19.6. We can then introduce three new functions, namely T1 (λ)

=

G11 (λ) + G12 (λ)M (λ)Y (λ)G21 (λ),

(20.8)

T2 (λ)

=

G12 (λ)M (λ),

(20.9)

T3 (λ)

=

!(λ)G21 (λ). M

(20.10)

382

Chapter 20. H-inﬁnity control applications

Recall that the problem we wish to solve is to ﬁnd, if possible, internally stabilizing compensators C of the system Σ with a proper transfer function K such that RG (K) belongs to Rat and (20.6) is satisﬁed, i.e., RG (K)∞ = max RG (K)(λ) < γ. λ∈iR

Here RG (K) is given by −1 G21 (λ); RG (K)(λ) = G11 (λ) + G12 (λ)K(λ) I − G22 (λ)K(λ)

(20.11)

see (20.5). In case K is given by (19.9) involving the function Q featured there, we can rewrite RG (K) as follows. Theorem 20.1. With K as in (19.9), the closed loop transfer function is given by RG (K)(λ) = T1 (λ) − T2 (λ)Q(λ)T3 (λ), where T1 , T2 and T3 are given by (20.8), (20.9) and (20.10), respectively !(λ)−1 N (λ) and (19.9) into I − G22 (λ)K(λ) −1 , and Proof. Inserting G22 (λ) = M suppressing the variable λ for notational convenience, we get !(X − N Q) − N (Y − M Q) −1 M ! (I − G22 K)−1 = (X − N Q) M =

!. (X − N Q)M

In the actual derivation of these identities, the doubly coprime factorization in (19.6) and the deﬁning properties given by (19.7) are employed. Again using (19.9), !. Substituting this in the formula for we arrive at K(I − G22 K)−1 = (Y − M Q)M the closed loop transfer function (20.11) yields RG (K) = =

!G21 ) − G12 M QM !G21 (G11 + G12 Y M !G21 ) − T2 QT3 . (G11 + G12 Y M

Now from the deﬁning properties of a doubly coprime factorization (19.6) one !. Inserting this in the formula above we obtain that T1 = sees that M Y = Y M ! G11 + G12 Y M G21 . This completes the proof.

20.2 The one-sided model matching problem In this section we consider the model matching problem (20.7) with T1 ∈ Ratl×p , T2 ∈ Ratl×q and T3 = Ip . Furthermore, we assume that T2 has a left inverse in − Ratq×l . In particular, T1 is analytic on the imaginary axis (with inﬁnity included) and T2 is iR-stable. Note that the left invertibility of T2 implies that l ≥ q, that is, T2 is a “tall” matrix.

20.2. The one-sided model matching problem

383

Given T1 and T2 as in the previous paragraph, the problem is to ﬁnd necessary and suﬃcient conditions for the existence of an iR-stable rational q × p matrix function Q, i.e., Q ∈ Ratq×p − , such that T1 − T2 Q∞ < γ, and to give a full parametrization of all such Q. We refer to this problem as the one-sided model matching problem corresponding to T1 and T2 . We shall explain how this problem reduces to the Nehari problem, and we shall present a necessary and suﬃcient condition for its solution in terms of a J-spectral factorization. The following theorem is the main result of this section. Theorem 20.2. Let T1 ∈ Ratl×p and T2 ∈ Ratl×q be given, and assume T2 has a − left inverse in Rat. Let γ > 0, and put ∗ Il T2 (λ) T1 (λ) T2 (λ) 0 0 , Υ(λ) = T1∗ (λ) Ip 0 −γ 2 Ip 0 Ip J

=

Iq

0

0

−Ip

.

such the norm constraint T1 − T2 Q∞ < γ is Then there exists Q ∈ Ratq×p − satisﬁed if and only if Υ admits a left J-spectral factorization Υ(λ) = W ∗ (λ)JW (λ),

(20.12)

with respect to the imaginary axis having the additional property that the q × q block in the left upper corner of W (λ) has an inverse in Ratq×q − . Moreover, writing 2 −1 W (λ) = ωij (λ) i,j=1 , where ω11 (λ) and ω22 (λ) are of sizes q × q and p × p, respectively, all solutions Q of the one-sided model matching problem corresponding to T1 and T2 are given by −1 , Q(λ) = − ω11 (λ)U (λ) + ω12 (λ) ω21 (λ)U (λ) + ω22 (λ)

(20.13)

with U ∞ < 1. where U is a rational matrix function in Ratq×p − and has a left inverse in Rat, we know from Proof. Since T2 belongs to Ratl×q − Theorem 17.26 that T2 admits an inner-outer factorization with an invertible outer factor. Thus T2 = V X, where V is inner, and both X and X −1 are analytic in the closed right half plane. If T2 happened to be square, the reduction to the Nehari problem would now be easy. Indeed, in that case V is bi-inner, and hence ∞, T1 − T2 Q∞ = T1 − V XQ∞ = V ∗ T1 − XQ∞ = R − Q = XQ. Actually, since both X and Q are in Rat− , also Q where R = V ∗ T1 and Q is in Rat− . Thus, this is not quite the Nehari problem as presented in Chapter 18, ∗ . Also note that R∗ is not but applying the results of Section 18.3 to R∗ yields Q

384

Chapter 20. H-inﬁnity control applications

stable, but it is just in Rat. At this point we use the fact that Proposition 18.6, when applied to R∗ , does not require R∗ to be stable. Recall that this point was made explicitly in the paragraph preceding the statement of Proposition 18.6. However, in general, T2 is only left invertible and not square, in which case V is only inner and not bi-inner. To deal with this more general case, we proceed = V V is bi-inner. We as follows (see Section 17.8): take V such that U has the same McMillan degree as V , that is, in the way choose V such that U outlined in Section 17.8. Then V∗ XQ T1 − T2 Q∞ = ∞ . T1 − (V )∗ 0 It follows that T1 − T2 Q∞ < γ if and only if for each for λ ∈ iR ∪ {∞} the following two conditions hold: (a) Φ(λ) = γ 2 Ip − T1∗ (λ)V (λ)(V )∗ (λ)T1 (λ) > 0, (b) γ 2 Ip − T1∗ (λ)V (λ)(V )∗ (λ)T1 (λ) −(V ∗ (λ)T1 (λ) − X(λ)Q(λ))∗ (V ∗ (λ)T1 (λ) − X(λ)Q(λ)) > 0. Using (a), the inequality (b) can be reduced to Φ−(V ∗ T1 −XQ)∗ (V ∗ T1 −XQ) > 0 where, for notational convenience, the variable λ being suppressed. Now, let Φ(λ) = N ∗ (λ)N (λ) be a left canonical factorization of Φ relative to the imaginary axis. Then condition (b) above is equivalent to Ip − N −∗ (V ∗ T1 − XQ)∗ (V ∗ T1 − XQ)N −1 > 0, i.e., to V ∗ T1 N −1 − XQN −1∞ < 1. Observe that this, in turn, is precisely an = XQN −1 . instance of Nehari’s problem, with R = V ∗ T1 N −1 and Q We apply the Nehari problem to R. Applying the result of Section 18.3, in particular Proposition 18.6 (which we apply with left half plane and right half plane interchanged) one sees that this Nehari problem is solvable if and only if the function Ψ(λ), deﬁned by Iq Iq V ∗ (λ)T1 (λ)N −1 (λ) Iq 0 0 , Ψ(λ) = N −∗ (λ)T1∗ (λ)V (λ) Ip 0 −Ip 0 Ip has a left J-spectral factorization of the form ) ( 0 Iq ∗ L− (λ), Ψ(λ) = L− (λ) 0 −Ip

(20.14)

with the additional property that the p × p block entry in the right lower corner p×p −1 of L−1 − has an inverse in Rat− . Moreover, in that case, if we partition L− (λ)

20.2. The one-sided model matching problem

385

2 Lij (λ) i,j=1 , with L11 a q × q rational matrix function, then all as L−1 − (λ) = solutions to this Nehari problem are given by Q(λ) = −(L11 (λ)U (λ) + L12 (λ))(L21 (λ)U (λ) + L22 (λ))−1 , where U runs over all functions in Ratq×p for which U ∞ < 1. Finally, recall − (see the ﬁnal paragraph of Section 18.3) that the additional property of the p × p block entry in the right lower corner of L−1 − is equivalent to the q × q block entry in the left upper corner of L− having an inverse in Ratp×p − . Put Q(λ) = X −1 (λ)Q(λ)N (λ). From the results of the previous paragraph, we get that all solutions to the one-sided model matching problem are given by Q(λ) = − X −1 (λ)L11 (λ)U (λ) + X −1(λ)L12 (λ) −1 , · N −1 (λ)L21 (λ)U (λ) + N −1 (λ)L22 (λ) where U runs over all functions in Ratq×p for which U ∞ < 1. − Next, introduce W (λ) = L− (λ)

X(λ)

0

0

N (λ)

.

(20.15)

Note that the q × q block entry in the left upper corner of W has an inverse in Ratq×q − . Furthermore, W

−1

(λ) =

X −1 (λ)L11 (λ)

X −1 (λ)L12 (λ)

N −1 (λ)L21 (λ)

N −1 (λ)L22 (λ)

.

So all solutions are parametrized by the function W −1 . It remains to establish the identity (20.12), that is, once more suppressing the variable λ, Υ = W ∗ JW. Let us denote the right side of the previous identity by Ξ. Thus Ξ = W ∗ JW . Using the deﬁnition of W in (20.15) together with formula (20.14), we see that Ξ=

X∗

0

0

N∗

Ψ

X

0

0

N

.

386

Chapter 20. H-inﬁnity control applications

It follows that Ξ= = = = = = = =

X∗

0

T1∗ V

N∗

X∗

0

T1∗ V

N∗

T2∗

0

T1∗ V V ∗

I

T2∗

0

T1∗ V V ∗

I

I

0

0

−I

V ∗V

0

0

−I

X

V ∗ T1

0

N

X

V ∗ T1

0

N

I

0

0

−N ∗ N

T2

V V ∗ T1

0

I

I

0

0

−γ 2 I + T1∗ V (V )∗ T1 T2∗ V V ∗ T1

T1∗ V V ∗ T2

−γ 2 I + T1∗ (V V ∗ + V (V )∗ )T1 X ∗ V ∗ T1

T1∗ V X

−γ 2 I + T1∗ T1 T2∗ T1

T1∗ T2

−γ 2 I + T1∗ T1

T1∗

I

I

0

0

−γ 2 I

0

I

T2∗ T2

V V ∗ T1

T2∗ T2

0

T2

T2∗ T2

T2∗

T2

T1

0

I

= Υ.

Thus we conclude that we may obtain W from a J-spectral factorization of a function that is easily described in terms of T1 and T2 , as desired. Note also that the positivity of γ 2 − T1∗ V (V )∗ T1 on iR ∪ {∞} is implied by the J-spectral factorization.

20.3 The two-sided model matching problem In this section we extend the analysis of the previous section to the two-sided model matching problem. It will turn out that in this case we need two J-spectral factorizations. l×q Theorem 20.3. Let T1 ∈ Ratl×p and T3 ∈ Ratm×p . Assume that − , T2 ∈ Rat− − T2 has a left inverse in Rat, and T3 has a right inverse in Rat. Let γ > 0, and

20.3. The two-sided model matching problem put Ω(λ) =

T3 (λ)

0

T1 (λ)

Il

387

Ip

0

0

−γ 2 Il

T3∗ (λ)

T1∗ (λ)

0

Il

.

(20.16)

Then there exists Q ∈ Ratq×m such that T1 − T2 QT3 < γ if and only if two − conditions (i) and (ii) hold. The ﬁrst condition (i) is as follows: (i) With respect to the imaginary axis, Ω admits a right J-spectral factorization 0 Im ∗ ¯ , where J = Ω(λ) = V (λ)JV (−λ) , (20.17) 0 −Il having the additional property that the m × m block in the upper left-hand . corner of V has an inverse in Ratm×m − With V as in (20.17), deﬁne 0 −T2∗(λ) −Im Ω(λ) = V −∗ (λ) I 0 0

0 Il

V −1 (λ)

0

I

−T2 (λ)

0

.

(20.18)

Then the second condition (ii) is: admits a left J-spectral factorization (ii) With respect to the imaginary axis, Ω of the form 0 I q ¯ ∗ JW (λ), where J = Ω(λ) = W (−λ) , (20.19) 0 −Im having the additional property that the q × q block in the upper left-hand corner of W has an inverse in Ratq×q − . Moreover, when (i) and (ii) are satisﬁed, (all ) the solutions Q to the two-sided model matching problem corresponding to T1 , T2 and T3 can be obtained as follows. 2 Partition W −1 = Xij i,j=1 , with X11 a q × q rational matrix function. Then Q = −(X11 U + X12 )(X21 U + X22 )−1 ,

(20.20)

where U is an iR-stable rational q × m matrix function with U ∞ < 1. Proof. The idea of the proof is to reduce the two-sided model matching problem to the one-sided model matching problem discussed in the previous section. The proof is divided into several steps. Part 1. We ﬁrst show that condition (i) in the theorem is a necessary condition. ¯ ∗ and T (λ) = T3 (λ) ¯ ∗ . Note the crucial To this end, introduce T1 (λ) = T1 (λ) 3 diﬀerence with the functions T1∗ and T3∗ : the functions T1 and T3 are analytic in the closed right half plane, inﬁnity included. With the help of these functions,

388

Chapter 20. H-inﬁnity control applications

∞ < γ, where rewrite T1 − T2 QT3 ∞ < γ in the following way: T1 − T3 Q = Q T2 (with the obvious interpretations for these functions). Taking into Q account Theorem 20.2, this gives that the ﬁrst condition is necessary. Indeed, with T3 T1 L= 0 I and V = W , we obtain L

∗

I

0

0

−γ 2 I

L= W

∗

I

0

0

−γ 2 I

W.

Part 2. The next step is to rewrite the two-sided model matching problem in an equivalent way. Use Theorem 17.28 to write T3 (λ) = Y (λ)V1 (λ) where Y is an m × m invertible outer function and V1 is an m × p co-inner function. Let V1 be such that V = V1∗ (V1 )∗ is bi-inner (see Corollary 17.33). Write R = T1 V = T1 V1∗ T1 (V1 )∗ = R1 R2 , where R1 is an l × m and R2 is an l × (p − m) matrix function. As V is bi-inner, T3 V = Y 0 . Thus we have T1 − T2 QT3 ∞ < γ if and only if

R1

R2 − T2 QY

0 ∞ < γ.

In turn, this can be rewritten as ∗ γ 2 Il > R1 (λ) − T2 (λ)Q(λ)Y (λ) R1 (λ) − T2 (λ)Q(λ)Y (λ) + R2 (λ)R2∗ (λ), for all λ ∈ iR ∪ {∞}, or equivalently, suppressing the variable λ again, as γ 2 Il − R2 R2∗ > (R1 − T2 QY )(R1 − T2 QY )∗ . This implies that γ 2 Il − R2 R2∗ > 0, and if we write γ 2 Il − R2 R2∗ = M M ∗ with M and M −1 in Ratl×l − , then we can rewrite the inequality above as Il > M −1 (R1 − T2 QY )(R1 − T2 QY )∗ M −∗ . Thus T1 − T2 QT3 ∞ < γ if and only if the following two conditions hold: γ 2 Il − R2 R2∗ > 0,

M −1 R1 − M −1 T2 QY ∞ < 1.

(20.21)

Note that the last of these two conditions is a one-sided model matching prob. Observe also that M −1 R1 = lem for QY , as both Y and Y −1 are in Ratm×m − l×m −1 ∗ ∗ M T1 V1 is in Rat , because V1 is inner and hence analytic in the closed left half plane, inﬁnity included. Also M −1 T2 is in Ratl×q − . Although we do not know

20.3. The two-sided model matching problem

389

that M −1 R1 is in Ratl×m (that is, we do not know that it is analytic in the closed − right half plane), still all conditions of Theorem 20.2 are met. Thus we may apply Theorem 20.2, to see that solvability of the one-sided model matching problem, which is the second condition in (20.21), is equivalent to a J-spectral factorization problem in the following way. Put −1 M T2 M −1 R1 K= . 0 Im Then, by Theorem 20.2, solvability of the one-sided model matching problem, which (as just noted) is the second part of (20.21), is equivalent to existence of a (m+q)×(m+q) , matrix function P such that P and P −1 are in Rat− ( ) ( ) 0 0 I I K∗ l K = P∗ q P, 0 −Im 0 −Im and, in addition, the q × q-block of P in the upper left corner has an inverse in . Recall that the last condition is equivalent to the requirement that the Ratm×m − m × m-block in the right lower corner of P −1 is in Ratm×m . Moreover, (all) the − solutions Q to the one-sided model matching problem corresponding to M −1 T2 2 and M −1 R1 are generated by P −1 as follows: if P −1 = Pij i,j=1 , with P11 of size q × q, then Q

= −(P11 U + P12 )(P21 U + P22 )−1 Y −1 = −(P11 U + P12 )(Y P21 U + Y P22 )−1 .

Introduce W = P

Iq 0

0 Y −1

.

Then W and W −1 are analytic in the right half plane and the m × m block in the right lower corner of W −1 is equal to Y P22 , which is also in Ratm×m . Finally, W − generates all solutions Q. Let 0 Iq K=K . 0 Y −1 We conclude that solvability of the one-sided model matching problem, which is the second part of (20.21), is equivalent to existence of a J-spectral factorization of the form Iq Il 0 0 ∗ ∗ W =K K, (20.22) W 0 −Im 0 −Im with the additional property that the m × m block in the right lower corner of W −1 is in Ratm×m . −

390

Chapter 20. H-inﬁnity control applications

Part 3. Continuing with the considerations above, we compute −1 M T2 M −1 R1 Iq 0 = K 0 Im 0 Y −1 −M −1

=

M −1R1 Y −1 Y −1

0 −M

=

=

R1

0

Y

0

Y

−M

R1

−T2

∗ K

0

−T2

0

0

Im

0

Im

−T2

0

−1

It follows that

−1

0

Im

.

Il

0

0

−Im

Il

0

0

−Im

K

is equal to

0

−T2∗

Im

0

0

−M ∗

Y∗

R1∗

−1

0

Y

−M

R1

−1

0

Im

−T2

0

,

which, in turn, can be written as,

0

−T2∗

Im

0

$

0

Y

−M

R1

Il 0

0 −Im

0

−M ∗

Y∗

R1∗

%−1

0

Im

−T2

0

.

Now the product of the middle three terms is easily seen to be equal to Y R1∗ YY∗ . R1 Y R1 R1∗ − M M ∗ Observe also that Y Y ∗ = T3 T3∗ and Y R1∗ = Y V1 T1∗ = T3 T1∗ . Furthermore, R1 R1∗ − M M ∗

= R1 R1∗ − γ 2 Il + R2 R2∗ 2

= −γ Il + R1

R2

R1∗ R2∗

= −γ 2 Il + T1 V ∗ V T1∗ = −γ 2 Il + T1 T1∗ .

20.3. The two-sided model matching problem Hence

YY∗

Y R1∗

R1 Y

R1 R1∗ − M M ∗

= =

391

T3 T3∗

T3 T1∗

T1 T3∗

−γ 2 Il + T1 T1∗

T3

0

Ip

T1

I

0

∗ T3

0 −γ 2 IL

T1∗

0

I

= Ω.

Part 4. After these preliminaries we can now complete the proof in one direction. Indeed, to show that both the conditions (i) and (ii) need to be satisﬁed, note that we already saw at the beginning of the proof that (i) is necessary. Assuming that (i) holds, we continue the computation above, with V as in (20.17), and see that ∗

K

Il

0

0

−Im

K

= =

0

−T2∗

Im

0

0

−T2∗

Im

0

$ V V

−∗

Im

0

0

−Il

Im 0

0 −Il

%−1 V

0

Im

−T2

0

∗

V

−1

0

Im

−T2

0

.

Thus, by (20.22), the second condition (ii) is necessary as well. Part 5. For the converse, assume that both (i) and (ii) are satisﬁed. As in the proof of Theorem 20.2, applied to T1 and T3 , in place of T1 and T2 , we see that (i) implies that the ﬁrst condition in (20.21) holds. Now follow the arguments in Parts 3 and 4 backwards to see that also the second condition in (20.21) is met. As we have already seen that these two conditions taken together are equivalent to the two-sided model matching problem, the proof is complete. Note that for the factorization (20.17) we need the analogue of Theorem 14.7 for right J-spectral factorization, applied to the function Ω given by (20.16). This analogue can be obtained by applying the left factorization result of Theorem 14.7 to the function Ω(−λ); cf., the paragraphs immediately following Theorem 14.8. In addition, the analogue of Theorem 14.7 for right J-spectral factorization provides us with a formula for the right J-spectral factor V , satisfying I 0 m ¯ ∗ V (λ). Ω(λ) = V (−λ) 0 −Il ¯ ∗ . We state the result of carrying The function we need will then be V (λ) = V (−λ) out all this in state space form as a lemma, which will be useful in the next section. Lemma 20.4. Let H(λ) = D+C(λIn −A)−1 B be a realization of an (m+l)×(p+l) rational matrix function H. Write J = diag (Ip , −Il ), J = diag (Im , −Il ), and

392

Chapter 20. H-inﬁnity control applications

assume that DJ D∗ = J. Also assume that A has all its eigenvalues in the open ¯ ∗ . Then Ω admits a right J-spectral left half plane. Put Ω(λ) = H(λ)J H(−λ) factorization with respect to the imaginary axis if and only if the algebraic Riccati equation XC ∗ JCX + X(A∗ − C ∗ J −1 DJ B ∗ ) + (A − BJ D ∗ J −1 C)X +BJ D∗ JDJ B ∗ − BJB ∗ = 0 has a Hermitian solution X such that A∗ − C ∗ J −1 (DJ B ∗ − CX) has its eigenvalues in the open left half plane. If X is such a solution (necessarily unique), and V (λ) = Im+l + C(λIn − A)−1 (BJ D∗ − XC ∗ )J −1 , then Ω(λ) = V (λ)JV (−λ)∗ is a right J-spectral factorization of Ω with respect to the imaginary axis.

20.4 State space solution of the standard problem In this section we return to the standard problem. We recall the basic facts about the problem. The starting point is a system in state space form ⎧ x (t) = Ax(t) + B1 w(t) + B2 u(t), ⎪ ⎪ ⎨ z(t) = C1 x(t) + D1 u(t), (20.23) ⎪ ⎪ ⎩ t ≥ 0. y(t) = C2 x(t) + D2 w(t), The input vector u(t) belongs to Cq , the noise vector w(t) belongs to Cp , the state vector x(t) belongs to Cn , the measured output y(t) belongs to Cm , and ﬁnally, the output z(t) to be controlled belongs to Cl . Thus the sizes of the matrices featured in (20.23) are as follows: A is n × n, B1 is n × p, B2 is n × q, C1 is l × n, C2 is m × n, D1 is l × q, and D2 is m × p. Throughout the section we assume that the following simplifying assumptions hold: A1. (A, B1 ) is controllable and (C1 , A) is observable, A2. (A, B2 ) is stabilizable and (C2 , A) is detectable, that is, there are matrices F and H so that both A + B2 F and A + HC2 have all their eigenvalues in the open left half plane. A3. D1∗ C1 = 0, D1∗ D1 = Iq , D2 B1∗ = 0, D2 D2∗ = Im . Given is also γ > 0. The problem we consider is to ﬁnd an internally stabilizing compensator K from y to u such that (20.6) holds. As we have explained in Section 20.1 this problem can be transformed into a model matching problem, using the rational matrix functions T1 , T2 , and T3

20.4. State space solution of the standard problem

393

appearing in (20.8)–(20.10). First we shall use (20.23) to derive state space realizations for T1 , T2 , and T3 . For this purpose we ﬁx matrices H and F such that AF = A+B2 F and AH = A+HC2 are stable matrices. Recall that assumption A2 guarantees the existence of matrices H and F with these properties. It is a matter of straightforward calculations to check that the following proposition holds. Proposition 20.5. Write G22 (λ) = C2 (λIn − A)−1 B2 , and assume assumption A2 is satisﬁed. Let F and H be matrices such that AF = A+ B2 F and AH = A+ HC2 are stable matrices. Suppose a doubly coprime factorization of G22 (λ) is given by the functions in (19.8). Then T1 (λ) =

C1 + D1 F

−D1 F

$ λI2n −

AF

−B2 F

0

AH

%−1

B1 B1 + HD2

,

T2 (λ) = D1 + (C1 + D1 F )(λIn − AF )−1 B2 , T3 (λ) = D2 + C2 (λIn − AH )−1 (B1 + HD2 ). Observe that T1 , T2 and T3 are in Rat− . Next, we show that T2 has a left inverse, while T3 has a right inverse, both in Rat. Lemma 20.6. Under the assumptions A1, A2, A3, the matrix function T2 has a left inverse in Rat and T3 has a right inverse in Rat. Proof. By Corollary 17.27, it suﬃces to show that T2 (λ) is left invertible for all λ ∈ iR and that T3 (λ) is right invertible for all λ ∈ iR. First we show that A − λIn B2 (20.24) C1 D1 is left invertible for all λ ∈ iR if and only if T2 (λ) is left invertible for all λ ∈ iR. To see that this is the case, we ﬁrst establish that T2 (λ) is left invertible for all λ ∈ iR if and only if AF − λIn B2 (20.25) C1 + D1 F D1 is left invertible for all λ ∈ iR. Indeed, assume that T2 (λ) is left invertible for all pure imaginary λ, and that for some λ0 ∈ iR and some vectors u and x we have 0 AF − λ0 In B2 x = . (20.26) 0 C1 + D1 F D1 u Then, since λ0 is not an eigenvalue of AF , it follows that x = (λ0 − AF )−1 B2 u. Inserting this in (C1 + D1 F )x + D1 u = 0, gives T2 (λ0 )u = 0. Since T2 (λ0 ) is left invertible u = 0, and hence also x = 0.

394

Chapter 20. H-inﬁnity control applications

Conversely, assume T2 (λ0 )u = 0 for some u and some pure imaginary λ0 . Suppose that (20.25) is left invertible for all λ ∈ iR. Put x = (λ0 − AF )−1 B2 u, then (20.26) holds, hence x = 0 and u = 0. Now (20.25) can be written as

AF − λ0 In

B2

C1 + D1 F

D1

=

A − λIn

B2

In

0

C1

D1

F

Iq

.

Thus we see that (20.25) is left invertible if and only if (20.24) is left invertible. Next we show that A − λIn B2 C1

D1

is left invertible for all λ ∈ iR. Indeed, assume that for some λ0 ∈ iR and some vectors u and x we have A − λ0 In B2 x 0 = . C1 D1 u 0 Then, in particular, C1 x + D1 u = 0. Using D1∗ D1 = Iq and D1∗ C1 = 0, this implies that u = 0. But then (A − λ0 In )x = 0 and C1 x = 0. Since the pair (C1 , A1 ) is observable by assumption, it follows that x = 0. For sake of convenience, and without loss of generality, we shall assume from now on that γ = 1. The ﬁrst main result in this section is the following theorem. Theorem 20.7. Suppose the system (20.23) satisﬁes the assumptions A1, A2 and A3, and let γ = 1. Then there is an internally stabilizing compensator K for the system (20.23) satisfying (20.6) if and only if the following two conditions hold: (i) there is a Hermitian solution Y of the Riccati equation Y (C1∗ C1 − C2∗ C2 )Y + AY + Y A∗ + B1 B1∗ = 0

(20.27)

with the additional properties that A∗ +(C1∗ C1 −C2∗ C2 )Y is stable and Y > 0, (ii) with the unique Y from (i) there is a Hermitian solution Z of the Riccati equation Z(Y C2∗ C2 Y − B2 B2∗ )Z + Z(A + Y C1∗ C1 ) + (A∗ + C1∗ C1 Y )Z + C1∗ C1 = 0 (20.28) with the additional properties that A+Y C1∗ C1 −B2 B2∗ Z +Y C2∗ C2 Y Z is stable and Z > 0.

20.4. State space solution of the standard problem

395

Moreover, when (i) and (ii) are satisﬁed, (all ) the internally stabilizing compensators can be obtained as follows. Introduce Ψ11 (λ) Ψ12 (λ) Ψ(λ) = Ψ21 (λ) Ψ22 (λ) Iq 0 −B2∗ Z −1 B2 Y C2∗ , (20.29) = (λIn − A) + 0 Im C2 (In + Y Z) = A+Y C1∗ C1 −B2 B2∗ Z +Y C2∗ C2 Y Z. Then (all ) the internally stabilizing where A compensators satisfying (20.6) are given by −1 K(λ) = Ψ11 (λ)U (λ) + Ψ12 (λ) Ψ21 (λ)U (λ) + Ψ22 (λ) , where U is an iR-stable rational q × m matrix function with U ∞ < 1. Note that condition (i) requires the Riccati equation (20.27) to have a positive deﬁnite iR-stabilizing solution. From Theorem 13.3 we know that the iR-stabilizing solution is unique. Similarly, condition (ii) requires (20.28) to have a positive deﬁnite iR-stabilizing solution, which is unique for the same reason. It will be convenient to split the proof in a number of lemmas. Lemma 20.8. The existence of a right J-spectral factorization (20.17) in condition (i) of Theorem 20.3 is equivalent to the existence of an iR-stabilizing Hermitian solution Y to the Riccati equation (20.27). Moreover, the additional property that the m × m block in the left upper corner of V has an inverse in Ratm×m is − equivalent to Y > 0. Proof. We split the proof in two parts. Part 1. Starting from the formulas for T1 and T3 given in Proposition 20.5 we form T3 0 . L= T1 Il −1 B, where n − A) This matrix function has the realization L(λ) = D + C(λI 0 B1 AF −B2 F = = , B , A 0 AH B1 + HD2 0 = C

0

C2

C1 + D1 F

−D1 F

,

D=

D2

0

0

Il

.

It will be more convenient however to work with a similar realization. Put In In . (20.30) S= 0 In

396

Chapter 20. H-inﬁnity control applications

Note that =S A

AF 0

= C

−HC2 AH

S

0

C2

C1 + D1 F

−C1

−1

,

=S B

−HD2

0

B1 + HD2

0

,

S −1 .

Also put J = diag (Ip , −Il ) and J = diag (Im , −Il ). Using the factorization principle from Section 2.6 one sees that L can be factored as L(λ) = L1 (λ)L2 (λ), where Im 0 0 L1 (λ) = (λ − AF )−1 −H 0 , + C1 + D1 F 0 Il

L2 (λ) =

D2 0

0

+

Il

C2 C1

Because L1 is of the form

(λ − AH )−1 B1 + HD2 (

Im L1 (λ) = Ξ(λ) where

0 .

) 0 , Il

Ξ(λ) = −(C1 + D1 F )(λIn − AH )−1 H,

we have that L1 and its inverse are in Rat− . Thus Ω admits a right J-spectral factorization if and only if Ω2 = L2 J L∗2 admits a right J-spectral factorization. Moreover, Ω = V JV ∗ with V and its inverse in Rat− if and only if Ω2 = V2 JV2∗ , where V2 = L−1 1 V , and V2 and its inverse are in Rat− . Now applying Lemma 20.4 to Ω2 , and using that D2 D2∗ = Im and D2 B1∗ = 0, we obtain that a right J-spectral factorization of Ω2 exists if and only if the algebraic Riccati equation X(C2∗ C2 − C1∗ C1 )X + XA∗ + AX − B1 B1∗ = 0 has a Hermitian solution X for which A∗ + (C2∗ C2 − C1∗ C1 )X has all its eigenvalues in the open left half plane. Comparing with (20.27) we see that this is equivalent to taking X = −Y . Observe also that this solution Y is unique since X is unique. In addition V (λ) = L1 (λ)−1 V2 (λ), where C2 Im 0 + (λIn − AH )−1 H − XC2∗ XC1∗ V2 (λ) = C1 0 Il

=

Im

0

0

Il

+

C2 C1

(λIn − AH )−1 H + Y C2∗

−Y C1∗ .

20.4. State space solution of the standard problem

397

Part 2. Next, we show that the property that the m × m block in the left upper corner of V has an inverse in Rat− , is equivalent to Y being positive deﬁnite. Because of the special form of H1 , we have that the m × m block in the left upper corner of V is equal to the m × m block in the left upper corner of V2 . Let us denote this block by V11 . Then −1 V11 (λ)−1 = Im − C2 λIn − (A − Y C2∗ C2 ) (H + Y C2∗ ). Now using (20.27) we have that (A − Y C2∗ C2 )Y + Y (A∗ − C2∗ C2 Y ) = −B1 B1∗ − Y (C1∗ C1 + C2∗ C2 )Y ≤ −B1 B1∗ ≤ 0.

(20.31)

Since the pair (A, B1 ) is controllable it follows from standard arguments concerning Lyapunov equations (see, e.g., Theorem 4 in Section 13.1 in [107]) that A−Y C2∗ C2 has its spectrum in the open left half plane if and only if Y is positive deﬁnite. This ﬁnishes the ﬁrst part of the proof of Theorem 20.7. Next we consider the second condition in Theorem 20.3 and its equivalence to the remaining parts of Theorem 20.7. Lemma 20.9. The existence of a left J-spectral factorization as in (20.19) in condition (ii) of Theorem 20.3 is equivalent to the existence of an iR-stabilizing solution Z of (20.28). Moreover, the additional property that the q × q block in the upper left corner of W is in Ratq×q is equivalent to Z being positive deﬁnite. − Proof. Again we shall split the argument into several parts. Part 1. For the ﬁrst step we start by computing the function from condition (ii) of Theorem 20.3 as follows. Using the notation of the proof of Lemma 20.8, deﬁne 0 Im 0 Im −1 −1 −1 L(λ) = V (λ) = V2 (λ) L1 (λ) . −T2 (λ) 0 −T2 (λ) 0 in condition (ii) of Theorem 20.3 is given by Observe that the function Ω −Im 0 ∗ Ω(λ) = L (λ)J L(λ), where J = . 0 Il First we show that the existence of a left J-spectral factorization 0 Iq ∗ = W JW, where J = Ω 0 −Il amounts to the existence of a left J-spectral factorization of the matrix function 1 arises from a certain factorization of L. In fact, the argument 1 , where L ∗J L L 1 will be similar to the one used in the proof of the previous lemma.

398

Chapter 20. H-inﬁnity control applications Using the product rule and then simplifying, we get $ 0 I 0 m = Im+l + (λIn − AF )−1 H L1 (λ)−1 −T2 (λ) 0 C1 + D1 F $ · $ =

0

Im

−D1

0

0

Im

−D1

0

+

+

0

C1 + D1 F

−1

(λIn − AF )

0

−1

C1 + D1 F

(λIn − AF )

−B2

H

0

−B2

%

%

0

%

.

Thus, again applying the multiplication rule, we obtain a formula for L(λ), by −1 ∗ pre-multiplying the above expression with V2 (λ) . Using also C1 D1 = 0, this + C(λ − A) −1 B, where yields L(λ) =D A − Y C2∗ C2 + Y C1∗ C1 −Y C1∗ C1 0 H + Y C2∗ A= , B= , 0 AF −B2 H −C2 0 0 Im = = , C . D −D1 0 −C1 C1 + D1 F It is convenient to consider another realization. With S as in (20.30) and writing AY = A − Y C2∗ C2 + Y C1∗ C1 , we have AY −B2 F − Y C2∗ C2 −1 B2 Y C2∗ =S =S A S , B , 0 AF −B2 H = C

−C2

−C2

−C1

D1 F

S −1 .

=L 1 L 2 , where It is now easily checked that L 0 Im −C2 L1 (λ) = + (λIn − AY )−1 B2 −D1 0 −C1 2 (λ) L

=

Iq

0

0

Im

+

−F −C2

(λIn − AF )−1 −B2

Y C2∗ , H .

2 is in Rat− , and as Since AF is stable, L −F AF − −B2 H = AF − B2 F + HC2 = AH −C2

20.4. State space solution of the standard problem

399

−1 is in Rat− too. has all its eigenvalues in the open left half plane, L 2 From the considerations in the previous paragraph it follows that the rational admits a left J-spectral factorization if and only if the =L ∗J L matrix function Ω function L1 J L1 admits a left J-spectral factorization.. In that case, if W1 is a left ∗ J L. 1 , then W = W1 L 2 is a J-spectral factor of L ∗1 J L J-spectral factor of L Part 2 . In this part we continue to use the notation of the previous part. We 1 . This yields that there exists a left J-spectral now apply Theorem 14.7 to L ∗ factorization of Ω = L J L if and only if there is a Hermitian solution X of the algebraic Riccati equation X(B2 B2∗ − Y C2∗ C2 Y )X + X(A + Y C1∗ C1 ) + (A∗ + C1∗ C1 Y )X − C1∗ C1 = 0 having the additional property σ(A + Y C1∗ C1 + B2 B2∗ X − Y C2∗ C2 Y X) ⊂ Cleft . This solution X is unique. Taking Z = −X we see that Z satisﬁes the algebraic Riccati equation (20.28) and is the iR-stabilizing solution of that equation. Thus the left J-spectral factor ∗ J L 1 is given by W1 of L 1 B2∗ Z (λIn − AY )−1 B2 Y C2∗ , (20.32) W1 (λ) = Iq+m + −C2 − C2 Y Z 2 (λ) becomes and the product W (λ) = W1 (λ)L %−1 ( )$ B2 AY −B2 F − Y C2∗ C2 B2∗ Z −F Iq+m + λ− −C2 − C2 Y Z −C2 0 AF −B2

Y C2∗ H

.

Part 3. We now consider the additional property that the q × q block in the upper left corner of W has an inverse in Ratq×q − , and prove that this is equivalent to Z being positive deﬁnite. Let us denote the q × q block in the upper left corner of W by W11 . Then $ %−1 AY −B2 F − Y C2∗ C2 B2 ∗ λI2n − . W11 (λ) = Iq + B2 Z −F 0 AF −B2 −1 is Thus the main operator in the realization of W11 AY −B2 F − Y C2∗ C2 B2 = B2∗ Z A − 0 AF −B2

=

AY − B2 B2∗ Z

−Y C2∗ C2

B2 B2∗ Z

A

−F

.

We have to show that this matrix has all its eigenvalues in the open left half plane if and only if Z is positive deﬁnite.

400

Chapter 20. H-inﬁnity control applications In order to do this, it is helpful to consider a similar matrix. Take In 0 S= , −In In

and put = = S −1 AS A

A + Y C1 C1∗ − B2 B2∗ Z

−Y C2∗ C2

Y C1∗ C1

A − Y C2∗ C2

.

has all its eigenvalues in the left half We shall show that Z > 0 ifand onlyif A Z 0 Z 0 ∗ +A A plane. To this end, consider 0 Y −1 0 Y −1 =

−C1∗ C1 − ZB2 B2∗ − ZY C2∗ C2 Y Z

C1∗ C1 − ZY C2∗ C2

C1∗ C1 − C2∗ C2 Y Z

Λ

,

where, because of (20.31), Λ = =

Y −1 (A − Y C2∗ C2 ) + (A∗ − C2∗ C2 Y )Y −1 −Y −1 B1 B1∗ Y −1 − C1∗ C1 − C2∗ C2 .

Substituting the latter expression for Λ in the right lower corner of the matrix above, we obtain Z 0 Z 0 ∗ +A A 0 Y −1 0 Y −1 C1∗ C1 − ZY C2∗ C2 −C1∗ C1 − ZB2 B2∗ − ZY C2∗ C2 Y Z = C1∗ C1 − C2∗ C2 Y Z −Y −1 B1 B1∗ Y −1 − C1∗ C1 − C2∗ C2 ⎤ ⎡ −C1 C1 ⎥ ⎢ ∗ ∗ ZB2 0 ZY C2∗ ⎢ 0 ⎥ C1 ⎥ ⎢ B2 Z =− ⎥. ⎢ ∗ −1 ∗ ∗ −1 ⎥ ⎢ B1 Y ⎦ −C1 0 Y B1 C2 ⎣ 0 C2 Y Z

C2

as shorthand for the latter factor, this reduces to With the notation C Z 0 Z 0 +A ∗ ∗ C. A = −C 0 Y −1 0 Y −1

(20.33)

20.4. State space solution of the standard problem

401

A is observable. Suppose Next, we show that the pair C, A

x y

= λ0

x y

,

C

x y

= 0,

or, which comes down to the same, (A + Y C1∗ C1 − B2 B2∗ Z)x − Y C2 C2∗ y

= λ0 x,

Y C1∗ C1 x + (A − Y C2∗ C2 )y

= λ0 y,

and C1 x = C1 y,

B2∗ Zx = 0,

B1∗ Y −1 y = 0,

C2 Y Zx = −C2 y.

Using C1 x = C1 y, it follows that (A − Y C2∗ C2 + Y C1∗ C1 )y = λ0 y. Combining this with B1∗ Y −1 y = 0, and putting w = Y −1 y, we obtain (AY − Y C2∗ C2 Y + Y C1 C1∗ Y + B1 B1∗ )w = λ0 Y w,

B1∗ w = 0.

Now use (20.27) to see that this implies Y A∗ w = λ0 Y w. As Y is invertible we have A∗ w = λ0 w and B1∗ w = 0. Since (A, B1 ) is controllable, it follows that w = 0. Hence y = 0 too. From (A + Y C1∗ C1 − B2 B2∗ Z)x − Y C2 C2∗ y = λ0 x, combined with y = 0, C1 x = C1 y = 0 and B2∗ Zx = 0 we then have Ax = λ0 x. The observability of the pair (C1 , A) ﬁnally gives x = 0. We ﬁnish by applying the result of Theorem 4 in Section 13.1 in [107] to the equation (20.33). Combined with the fact that Y > 0, this gives that Z > 0 if and has all its eigenvalues in the open left half plane. only if A This concludes the proof of the equivalence of (i) and (ii) in Theorem 20.7. We bring the argument to a close as follows. Proof of Theorem 20.7. In view of the two preceding lemmas, it remains to prove the formulas for the parametrization of the internally stabilizing compensators satisfying (20.6). Recall from Theorem 19.6, in particular from formula (19.9), that K = (Y − M Q)(X − N Q)−1 . Also we have formula (20.20), that is the expression Q = −(X11 U + X12 )(X21 U + X22 )−1 , where W −1 =

X11

X12

X21

X22

.

402

Chapter 20. H-inﬁnity control applications

Here W is obtained from Part 1 of the proof of Lemma 20.9. Combining the expressions, we see that K = (Y X21 + M X11 )U + (Y X22 + M X12 ) −1 · (XX21 + N X11 )U + (XX22 + N X12) = (Ψ11 U + Ψ12 )(Ψ21 U + Ψ22 )−1 , with Ψ given by Ψ=

Ψ11

Ψ12

Ψ21

Ψ22

=

M

Y

X11

X12

N

X

X21

X22

.

The formulas in (19.8) now give F M (λ) Y (λ) Iq 0 + (λIn − AF )−1 B2 = C2 0 Im N (λ) X(λ)

−H .

2 (λ) from Part 1 of the proof of the Fortuitously, this is equal to the function L 2 W −1 = W −1 , where W1 is previous lemma. Since W = W1 L2 , we get Ψ = L 1 given by (20.32). Hence Ψ is given by (20.29), as desired. We conclude with the second main result of this chapter. Theorem 20.10. Suppose the system (20.23) satisﬁes the assumptions A1, A2 and A3, and let γ be an arbitrary positive number. Then there is an internally stabilizing compensator K for the system (20.23) satisfying (20.6) if and only if the following three conditions hold: (i) there is a positive deﬁnite iR-stabilizing solution X of the Riccati equation X(γ −2 B1 B1∗ − B2 B2∗ )X + A∗ X + XA + C1∗ C1 = 0,

(20.34)

(ii) there is a positive deﬁnite iR-stabilizing solution Y of the Riccati equation Y (γ −2 C1∗ C1 − C2∗ C2 )Y + AY + Y A∗ + B1 B1∗ = 0,

(20.35)

(iii) X < γ −2 Y −1 or, equivalently, all eigenvalues of XY are in the open disc {z | |z| < γ −2}. In that case (all ) the internally stabilizing compensators K can be obtained as follows. Introduce Φ11 (λ) Φ12 (λ) Φ(λ) = Φ21 (λ) Φ22 (λ) ) ( ∗ ) ( B2 X 0 Iq −1 Y C2∗ B2 , − (I − γ −2 Y X)−1 (λ − A) = C2 Im 0

20.4. State space solution of the standard problem

403

= A − Y (C2∗ C2 − γ −2 C1∗ C1 ) − B2 B2∗ X(In − γ −2 Y X)−1 . Then (all ) the where A internally stabilizing compensators satisfying (20.6) are given by K(λ) = Φ11 (λ) + Φ12 (λ)U (λ)(Im − Φ22 (λ)U (λ))−1 Φ21 (λ), where U is an iR-stable rational q × m matrix function satisfying U ∞ < γ. Proof. The theorem may be derived from the previous one upon giving the connections between X and Z. Again we assume γ = 1 without loss of generality. Under this assumption, condition (i) in Theorem 20.7 is exactly the same as the second condition in the present theorem. Henceforth we suppose it is satisﬁed. Thus, throughout the proof, Y will be a positive deﬁnite iR-stabilizing solution of (20.27), or, equivalently, of (20.35) with γ = 1. The argument below is divided into four parts. Part 1. Introduce the block matrices −A∗ −C1∗ C1 −C1∗ C1 −A∗ − C1∗ C1 Y = H= . , H B1 B1∗ − B2 B2∗ A Y C2∗ C2 Y − B2 B2∗ A + Y C1∗ C1 In the terminology of Section 12.1 the matrix H is the Hamiltonian of the Riccati is the Hamiltonian of the Riccati equation equation (20.34) with γ = 1, while H (20.28). Introduce also ) ( In 0 . S= Y In Since Y is a solution of the Riccati equation (20.35) and γ = 1, a direct computa tion gives S −1 HS = H. Part 2. Here we assume that Z is the (unique) Hermitian iR-stabilizing solution of equation (20.28), and in addition that Z is positive deﬁnite. That is, it is assumed that condition (ii) in Theorem 20.7 is met. Since Z is iR-stabilizing, the space ∗ corresponding to the open left half Im Z ∗ In is the spectral subspace of H plane. It follows that Z Z = Im SIm In + Y Z In is the spectral subspace of H corresponding to the open left half plane. Our next concern is the invertibility of In + Y Z. Since Z is positive deﬁnite, In + Y Z = Z −1/2 (In + Z 1/2 Y Z 1/2 )Z 1/2 is similar to a positive deﬁnite matrix. Consequently, In + Y Z is invertible. Next, put X = Z(In + Y Z)−1 . We shall show that X is positive deﬁnite, X is the iR-stabilizing solution of (20.34) (with γ = 1), and that X < Y −1 . For this, note that X = Z(In + Y Z)−1 = (Z −1 + Y )−1 , so that X is positive deﬁnite. Furthermore, Z X Im = Im . In + Y Z In

404

Chapter 20. H-inﬁnity control applications

Hence X is the Hermitian iR-stabilizing solution of (20.34). In addition, since Z is positive deﬁnite also X −1 > Y , and as both X and Y are positive deﬁnite this yields X < Y −1 . We conclude that all conditions of Theorem 20.10 are satisﬁed. Part 3. This part deals with the reverse implication. So, we start with the positive deﬁnite iR-stabilizing solution X of (20.34) with γ = 1 such that X < Y −1 . We show that Z = (In − Y X)−1 is well-deﬁned and positive deﬁnite, and that Z is the iR-stabilizing solution of (20.28). Since X < Y −1 , the matrix I − Y X is invertible, hence Z is well-deﬁned. In addition, Z = X(In −Y X)−1 = (X −1 −Y )−1 is positive deﬁnite because X 0 A1/2 A−1 A−∗ λ−A ρ(A) σ(A) P (A; Γ) AM A[M ] A−1 [M ] A|M A(X1 → X2 ) D(A)

quotient space of M over N symbol for orthogonality in Hilbert space orthogonal complement of subspace M in Hilbert space orthogonality of sets V and W orthogonal direct sum (of subspaces) of Hilbert spaces algebraic (possibly non-orthogonal) direct sum of linear manifolds or (sub)spaces

List of symbols L(Y ) L(U, Y ) C(Γ, U )

Lp (Ω) Lm p (Ω) Lm×r (Ω) p Lm 1,ω (R) Dm 1 (R) Dm 1 [0, ∞) L2 (R+ , H) · , · [· , ·] A[ ] D + C(λI − A)−1 B A× Ker (C|A) Im (A|B) E(·; A) etS etS PΘ prΠ (Θ) W −1 δ(W ) δ(W ; λ0 )

417 Banach algebra of all bounded linear operators on Banach space Y Banach space of all bounded linear operators from Banach space U into Banach space Y Banach space of all U -valued continuous functions on Γ endowed with the supremum norm Lebesgue space of p-integrable functions on a measurable set Ω space of Cm -valued functions of which the entries are in Lp (Ω) space of m × r matrix functions of which the columns are in Lm p (Ω) a weighted Lm 1 -space; see Section 5.3 a certain linear submanifold of Lm 1 (R); see Section 5.3 linear manifold of all functions f ∈ Dm 1 (R) with f (t) = 0 for t < 0 the space of all square integrable functions on [0, ∞) with values in Hilbert space H standard inner product in Cm or L2 [−1, 1] alternative inner product in Cm or L2 [−1, 1] adjoint of an operator with respect to alternative inner product (in L2 [−1, 1]) realization associate state space operator (or matrix), associate main operator (or matrix) corresponding to a realization stands for Ker C ∩ Ker CA ∩ Ker CA2 ∩ · · · stands for Im B + Im AB + Im A2 B + · · · bisemigroup generated by exponentially dichotomous operator A value at t(< 0) of the left semigroup generated by S value at t(> 0) of the right semigroup generated by S separating projection for −iA where A is the main operator of the spectral triple Θ projection of realization triple Θ = (A, B, C) associated with a projection Π pointwise inverse of rational matrix function W , deﬁned by W −1 (λ) = W (λ)−1 ) McMillan degree of a rational matrix function W local degree of W at λ0

418

List of symbols

π+ (W )

number of positive eigenvalues of the Hermitian matrix associated with a minimal realization of J-unitary rational matrix function W

F∗

adjoint of the rational matrix function F relative to the ¯ ∗ imaginary axis, deﬁned by F ∗ (λ) = F (−λ) the set of all rational matrix functions that are proper and have no pole at the imaginary axis the set of all p × q matrix functions in Rat the set of all F in Ratp×q such that sups∈iR F (s) ≤ 1 the set of all matrix functions in Ratp×q that are analytic on the closed left half plane, inﬁnity included the set Ratp×q ∩ Ratp×q + B the set of all rational matrix functions that are analytic on the closed right half plane, inﬁnity included, that is, the set of all iR-stable rational matrix functions the set of all p × q matrix functions in Rat− the unit element in the algebra Ratp×p

Rat Ratp×q Ratp×q B Ratp×q + Ratp×q +, B Rat− Ratp×q − Ep

Index H-adjoint of matrix, 212 H-Lagrangian subspace, 212 H-dissipative, 215 H-negative vector, 211 H-neutral subspace, 211 H-neutral vector, 211 H-nondegenerate, 212 H-nonnegative subspace, 211 H-nonpositive subspace, 211 H-orthogonal vectors, 212 H-orthogonality, 212 H-positive vector, 211 H-selfadjoint matrix, 184, 213 J-contraction, 333 J-contractive, 328 J-inner rational matrix function, 333 J-unitary matrix, 313 J-unitary rational matrix function on the imaginary axis, 314 R-stabilizing solution of Riccati equation, 273 T-stabilizing solution of discrete Riccati equation, 264, 270 Ω-regular (linear matrix) pencil, 57 iR-stabilizing solution of Riccati equation, 238 iR-stable rational matrix function, 350 angular operator, 219 angular subspace, 219 associate main matrix of matrix realization, 21

associate main operator of realization, 21 realization triple, 93 associate state space matrix of matrix realization, 21 associate state space operator of realization, 21 bi-inner rational matrix function, 333 BIBO-stable, 371 bigenerator of bisemigroup, 80 biproper rational matrix function, 175 biproper realization, 21 bisemigroup generated by exponentially dichotomous operator, 80 block Toeplitz equation, 13 bounded-input bounded-output stable, 371 Cauchy contour, 5 Cauchy domain, 5 co-inner rational matrix, 339 co-pole function, 173 corresponding to Jordan chain, 174 complete factorization, 175 contractive rational matrix function on imaginary axis, 301 on real line, 310 controllability gramian, 350 controllable pair of matrices, 171 controllable realization, 172

420 coupling relation (between operators), 106 derivative of f ∈ Dm 1 (R), 87 detectable pair of matrices, 374 doubly comprime factorization, 376 exponential type of exponentially dichotomous operator, 79 of realization triple, 87 of semigroup, 78 exponentially decaying semigroup, 78 exponentially dichotomous operator, 79 exterior domain of Cauchy contour, 5 external matrix of matrix realization, 21 external operator of realization, 21 externally stable system, 371 ﬁnite dimensional realization, 20 gramian, 350 half range problem, 116 Hamiltonian, 220 Hermitian matrix associated with minimal realization, 317 Hermitian rational matrix function on imaginary axis, 181 on real line, 181 on unit circle, 189 indeﬁnite inner product given by Hermitian matrix, 211 inner rational matrix function, 336 inner-outer factorization (with invertible outer factor), 337 input matrix of matrix realization, 21 input operator of realization, 21 input space of realization triple, 88 interior domain of Cauchy contour, 5 internal stability, 376

Index internally stabilizing compensator for system, 376 invertible in Ratp×p , 352 invertible outer rational matrix function, 337 kernel function associated with realization triple, 90 of Wiener-Hopf equation, 9 left J-spectral factorization with respect to the imaginary axis, 250 with respect to the real line, 272 with respect to the unit circle, 262 left (C0 -)semigroup, 78 left canonical factorization of operator function (with respect to Cauchy contour), 144 of rational matrix function, 39 left Hankel operator, 100 left pseudo-canonical factorization, 176 left pseudo-spectral factorization with respect to imaginary axis, 199 with respect to real line, 198 with respect to unit circle, 204 left spectral factor, 185 left spectral factorization with respect to the imaginary axis, 185 with respect to the real line, 185 with respect to the unit circle, 192 left Wiener-Hopf factorization (with respect to Cauchy contour), 158 linear fractional map, 352 linear manifold, 30 linear matrix pencil, 57

Index local minimality at a given point, 172 main matrix of matrix realization, 21 main operator of realization, 21 realization triple, 88 manifold, 30 matching condition, 31 matricially coupled operators, 105 matrix-valued realization of rational matrix function, 19 maximal H-nonnegative subspace, 212 maximal H-nonpositive subspace, 212 McMillan degree, 171 minimal factorization (involving arbitrary number of factors), 175 minimal realization, 20 model matching problem, 381 one-sided, 383 two-sided, 381, 386 negative squares, 320 Nehari problem, 350 Nehari-Takagi problem (relative to the imaginary axis), 369 nonnegative rational matrix function on imaginary axis, 197 on real line, 197 on unit circle, 197 nonnegative real part on the real line, 291 number of poles of function in Ratp×q in the open left half plane, multiplicities taken into account, 358 observability gramian, 350 observable pair of matrices, 172 observable realization, 172 one-sided model matching problem , 383

421 outer rational matrix function (invertible), 337 outer-co-inner factorization (with invertible outer factor), 339 output matrix of matrix realization, 21 output operator of realization, 21 output space of realization triple, 88 pair of matching subspaces, 31 partial indices, 159 pencil, 57 pole placement theorem, 374 positive deﬁnite rational matrix function on imaginary axis, 185 on real line, 185 on unit circle, 192 positive deﬁnite real part on the real line, 291 product of realizations, 30 projection of realization triple associated with a projection, 110 propagator function, 117 proper rational matrix function, 19 rational Nehari problem (relative to the imaginary axis with given tolerance), 350 rational Nehari-Takagi problem (relative to the imaginary axis), 369 realization of a system, 372 of operator function on given set, 20 of rational matrix function, 19 realization triple, 88 of given exponential type, 87 Redheﬀer transform of 2 × 2 block matrix, 328 of rational matrix function, 330 Redheﬀer transformation, 328 regular (linear matrix) pencil, 57

422 resolvent set of operator, 20 Riccati equation algebraic, 220 discrete algebraic, 264, 270 symmetric algebraic, 238 Riemann-Hilbert boundary value problem, 52 Riesz projection, 38 right J-spectral factorization with respect to the imaginary axis, 250 with respect to the real line, 272 with respect to the unit circle, 262 right (C0 -)semigroup, 78 right (Wiener-Hopf) factorization indices (with respect to Cauchy contour), 159 right canonical factorization (of symbol) with respect to real line, 10 (of symbol) with respect to the unit circle, 13 of operator function (with respect to Cauchy contour), 144 of rational matrix function, 39 of Wiener-Hopf equation with integrable operator-valued kernel function, 122 with respect to Cauchy contour, 16 right Hankel operator, 100 right pseudo-canonical factorization, 176 right pseudo-spectral factorization with respect to imaginary axis, 199 with respect to real line, 198 with respect to unit circle, 204 right spectral factor, 185 right spectral factorization with respect to the imaginary axis, 185

Index with respect to the real line, 185 with respect to the unit circle, 192 right Wiener-Hopf factorization (with respect to Cauchy contour), 157 scattering function, 116 selfadjoint rational matrix function on imaginary axis, 181 on real line, 181 on the unit circle, 189 separating projection for exponentially dichotomous operator, 79 sign characteristic of pair of matrices, 213 signature matrix, 250 similarity between realizations, 172 singular integral equation, 15 skew-Hermitian matrix, 182 skew-Hermitian rational matrix function on real line, 298 solution of Riemann-Hilbert boundary problem, 71 spectral projection, 38 spectral subspace, 38 splitting of spectrum, 37 stabilizable pair of matrices, 238 stabilizing solution of discrete Riccati equation, 264, 270 of Riccati equation, 238, 273 stable matrix, 372 stable rational matrix function, 350 stable realization of system, 372 standard problem of H-inﬁnity control, 381 state space matrix of matrix realization, 21 state space of realization, 21 triple, 88 state space operator of realization, 21

Index state space similarity between realizations, 172 theorem, 172 strictly H-dissipative matrix, 216 strictly contractive at inﬁnity, 302 strictly contractive rational matrix function on imaginary axis, 304 on the real line, 307 strictly proper rational matrix function, 305 strictly proper realization, 21 suboptimal rational Nehari problem relative to the imaginary axis with given tolerance, 350 supporting pair of subspaces, 31 supporting projection for realization, 31 realization triple, 110 symbol of (block) Toeplitz equation, 13 (block) Toeplitz operator, 14 Wiener-Hopf equation with integrable operator-valued kernel, 122 Wiener-Hopf integral equation, 10 Wiener-Hopf integral operator, 12 transfer function of realization triple, 90 of system, 21 two-sided inner rational matrix function, 333 two-sided model matching problem, 381, 386 uncontrollable eigenvalues, 374 unital realization, 21 unitary rational matrix functions, 324 unobservable eigenvalues of, 374

423 Wiener-Hopf equation, 9 Wiener-Hopf integral operator, 11 Wiener-Hopf operator with kernel function k, 99 zero of rational matrix function, 172

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close