Application of Operator Theory for the Representation of Continuous and Discrete Distributed Parameter Systems
Der Technischen Fakult¨at der Friedrich-Alexander-Universit¨at Erlangen-Nurnberg¨ zur Erlangung des Grades
Doktor-Ingenieur
vorgelegt von
Vitali Dymkou
Erlangen, 2006 Als Dissertation genehmigt von der Technischen Fakult¨at der Friedrich-Alexander-Universit¨at Erlangen-Nurnberg¨
Tag der Einreichung: 22. Dezember 2005 Tag der Promotion: 23. March 2006 Dekan: Prof. Dr.-Ing. Alfred Leipertz Berichterstatter: Apl. Prof. Dr.-Ing. habil. Peter Steffen Apl. Prof. Dr.-Ing. habil. Krzysztof Galkowski Acknowledgements
I would like to thank my supervisors, Apl. Prof. Dr.-Ing. habil. Peter Steffen and Priv. Doz. Dr.-Ing. habil. Rudolf Rabenstein, who introduced me to the subject of engineering problems and gave me the opportunity to work in their group. I would like to express them my gratitude for the excellent supervision and support and of course for the very warm atmosphere during my work and life in Erlangen. Also, I would like to thank Apl. Prof. Dr.-Ing. habil. Krzysztof Galkowski, who wisely advised me to start my scientific work in the Telecommunications Laboratory in Erlangen and for reviewing my thesis. I would like to thank ”Graduiertenkolleg Dreidimensionale Bildanalyse und -Synthese” and especially Prof. Dr. Gunther¨ Greiner for their financial support. I am deeply thankful to all my colleagues from the Telecommunications Laboratory for their patience and support over many months. I wish to thank Ursula Arnold for her help with all my administrative questions, Wolfgang Preiss for his software support and Manfred Lindner for his wonderful refrigerator. Especially I would like to thank Stefan Petrausch, who was always ready to translate, explain and answer all my private and scientific questions. I am deeply grateful to all my old friends in Russia and Belarus for their thousand calls and mails. They did not forget me. I would also like to thank my new friends in Germany Hamza Amasha, Juliane Gebhardt, Nael and Larissa Popova for always being there for me. Finally, and most importantly, I wish to thank my supportive family who accepted my time away from them. A special thanking word goes to my first teachers to my father Michael Dymkov and to my mother Raisa Dymkova. Also, of course, I want to thank my older brother Dymkou Siarhei, his wife Irina and their son Alexei. This work is dedicated to my family.
v
Contents
Abbreviations and Acronyms ix
List of mathematical symbols ix
Variables xi
1 Introduction 1
2 Basic Notions from Functional Analysis 5 2.1 Spaces ...... 5 2.2 Linearoperators...... 8 2.2.1 Unboundedoperators...... 8 2.2.2 Adjointoperators...... 9 2.2.3 Lineardifferentialforms ...... 10 2.2.3.1 Homogeneous boundary conditions ...... 10 2.2.3.2 Green’s formula and associated forms ...... 11 2.2.4 Lineardifferentialoperators ...... 12 2.2.4.1 Adjoint homogeneous boundary conditions and the adjoint operators ...... 13 2.2.4.2 Nonhomogeneous boundary conditions ...... 15 2.3 Chaptersummary...... 16
3 Spectral Theory of Operators 17 3.1 Theresolventoperator ...... 17 3.1.1 Canonical systems of the prime and the adjoint operator ...... 20 3.1.2 Expansion of the resolvent operator by the canonical system . . . . 24 3.1.3 Sectorialoperators ...... 25 3.1.3.1 Definitionofsectorialoperators ...... 26 3.1.3.2 Examplesofsectorialoperators ...... 26 3.1.4 Semigroups and canonical systems ...... 28 3.1.4.1 Definitions ...... 28 vi Contents
3.1.4.2 Connection between the semigroup of the operator, the resolvent and the Laplace transformation ...... 29 3.1.5 Expansion of the semigroup by the canonical system ...... 30 3.2 The C resolventoperator ...... 32 − 3.2.1 Canonical systems of the prime and the adjoint operator ...... 37 3.2.2 Expansion of the C resolvent operator by the canonical systems . . 39 − 3.3 Chaptersummary...... 40
4 Mathematical Modelling of Physical Processes 43 4.1 Ordinarydifferentialequations...... 43 4.1.1 ClassificationofODEs ...... 43 4.1.2 State-Spacemodel ...... 44 4.2 Partialdifferentialequations ...... 46 4.2.1 ClassificationofPDEs ...... 46 4.2.2 Initial-boundary-value problems ...... 48 4.2.3 Solution of initial-boundary-value problems in the Laplace domain . 49 4.3 Chaptersummary...... 51
5 Description of Multidimensional Systems 53 5.1 Multi-functional transformation (MFT) ...... 54 5.1.1 Definitionandproperties...... 55 5.1.2 InverseMFT ...... 56 5.2 Application of the MFT method to initial-boundary-value problems with homogeneousboundaryconditions...... 57 5.2.1 Initial-boundary-value problem with homogeneous boundary condi- tions...... 57 5.2.2 Laplacetransformation...... 58 5.2.3 Spatialtransformation(MFT) ...... 59 5.2.4 InverseMFT ...... 60 5.2.5 DiscretizationoftheMFTmodel ...... 61 5.2.5.1 Spatialdiscretization ...... 62 5.2.5.2 Timediscretization...... 62 5.2.6 Inverse z transformation ...... 63 − 5.2.7 Approximation ...... 64 5.3 Application of the MFT method to initial-boundary-value problems with nonhomogeneous boundary conditions ...... 64 5.3.1 Initial-boundary-value problems with nonhomogeneous boundary conditions ...... 64 5.3.2 Laplacetransformation...... 66 5.3.3 Spatialtransformation(MFT) ...... 66 Contents vii
5.3.4 InverseMFT ...... 69 5.3.5 DiscretizationoftheMFTmodel ...... 69 5.4 Application of the MFT method to general vector initial-boundary-value problems...... 70 5.4.1 Notation...... 70 5.4.2 General vector initial-boundary value problem ...... 70 5.4.3 Laplacetransformation...... 71 5.4.4 Spatialtransformation(MFT) ...... 72 5.4.5 InverseMFT ...... 74 5.4.6 DiscretizationoftheMFTmodel ...... 74 5.5 Chaptersummary...... 74
6 Examples of the MFT Simulations 77 6.1 Heatflowequation ...... 77 6.1.1 Primeandadjointoperators ...... 78 6.1.2 Eigenproblem ...... 78 6.1.3 Biorthogonality ...... 79 6.1.4 Thegeneralsolution ...... 79 6.1.5 MFTSimulation ...... 80 6.2 Heatflowthroughawall ...... 81 6.2.1 Adjointoperator ...... 82 6.2.2 Eigenvalueproblems ...... 82 6.2.3 Associatedvectors ...... 85 6.2.4 Biorthogonality ...... 87 6.2.5 Thegeneralsolution ...... 88 6.2.6 MFTsimulation...... 89 6.3 Telegraphequation ...... 89 6.3.1 Adjointoperator ...... 91 6.3.2 Eigenvalueproblems ...... 92 6.3.3 Associatedvectors ...... 95 6.3.4 Biorthogonality ...... 97 6.3.5 Thegeneralsolution ...... 98 6.4 Chaptersummary...... 98
7 Conclusions 99
A Appendix 101 A.1 ProofofTheorem2...... 101 A.2 ProofofExample5...... 109 A.3 ProofofTheorem5...... 112 viii Contents
B Titel, Inhaltsverzeichnis, Einleitung und Zusammenfassung 121 B.1 Titel ...... 121 B.2 Inhaltsverzeichnis ...... 121 B.3 Einleitung ...... 124 B.4 Zusammenfassung...... 128
Bibliography 130 Abbreviations and Acronyms ix
Abbreviations and Acronyms
FTM functional transformation method IIT impulse invariant transformation IBVP initial-boundary-value problem MD multi dimensional MFT multi-functional transformation ODE ordinary differential equation PDE partial differential equation SLT Sturm-Liouville transformation TFM transfer function model
Mathematical symbols
for all ∀ there exists ∃ subset of (or inclusion sign) ⊂ belongs to ∈ / does not belongs to ∈ 1 ( )− inverse operation of ( ) · · ( )H hermitian of ( ) · · ( )T transposed of ( ) · · ( )∗ conjugate complex of ( ) · · union ∪ intersection ∩ converges to → <, , >, inequality signs ≤ ≥ summation sign lim limit P + , , infinity ∞ −∞ ∞ Rank( ) rankof( ) · · Ran( ) imageorrangeof( ) · · Ker( ) kernel of ( ) · · dim dimension of ( ) {·} · Im imaginary part of ( ) {·} · Re real part of ( ) {·} · norm in H ||·||H ( , ) scalar product in H · · H span( ) linear span of ( ) · · x List of mathematical symbols inf( ) infimum of ( ) (greatest lower bound) · · sup( ) supremum of ( ) (least upper bound) · · min( ) minimum of ( ) · · max( ) maximum of ( ) · · det( ) determinant of ( ) · · k n Cn or binomial coefficient k ! π circle constant j imaginary unit Laplace transformation L{} multi-functional transformation T {} z transformation Z{} − δm,n Kronecker symbol N set of natural numbers Z set of integers R set of real numbers C set of complex numbers Cn(or Rn) set of complex (or real) n tuples − a sequence { n} (a, b) open interval [a, b] closed interval C([a, b]) set of continuous functions on [a, b] C(n)([a, b]) set of functions on [a, b] with continuous derivatives up to order n Ω subset of Rn ∂Ω boundary of Ω Ω closure of Ω
Lp(Ω) Lebesgue space on Ω Wk p(Ω) Sobolev space on Ω ∂ Dt or ∂t first-order temporal derivative ∂ Dx or ∂x first-order spatial derivative y˙ first-order temporal derivative of y
y′ first-order spatial derivative of y X,Y Banach spaces H, Hilbert spaces H I identity operator L, A, C linear operators
L† adjoint of linear operator L rank(L) rankoftheoperator L ρ(L) resolvent set of the operator L Variables xi
σ(L) spectrumoftheoperator L R(s, L) resolvent of the operator L empty set ∅
Variables
λ eigenvalue x continuous space variable x vector of continuous spatial coordinates z complex frequency variable of the z transformation − s complex frequency variable of the Laplace transformation t continuous time variable i, m, p, indices µ, ν xii Variables 1 1 Introduction
The development of computer technologies, the availability of high-speed processors and various programming languages allow nowadays the researchers in different areas of sci- ence to investigate and design numerous algorithms to solve physical phenomena on the computer. However, to construct high precision models of a real process one has to begin with its mathematical description and analysis in order to obtain specific characteristics of the considered problem. This helps to design very efficient numerical methods which can be implemented directly on the computer. The past decades, in particular, have seen a continually growing interest in the applica- tion of functional analysis and especially general operator theory to engineering problems. This development is clearly related to the wide variety of applications of both practical and theoretical interests. Many physical and information processes in various fields pos- sess identical mathematical structures, that can be described in a common operator form. Such a generalization allows to construct general algorithms to solve a wide class of prob- lems. It is, however, not sufficient to analyze a given problem on a pure theoretical basis. The practical application may pose additional constraints like real-time performance of the system, low-delay requirements or restrictions on computing power or memory. Therefore, it is necessary to develop a discrete model for the solution of the problem which is suitable for computer implementation. The general steps in such a development are illustrated in Figure 1.1. The presented thesis is oriented towards the development of algorithms based on spec- tral operator theory and their application for solving practical problems arising in the discrete simulation of continuous systems governed by partial differential equations with unbounded and non-self-adjoint operators, in general. We deal with the theory of well- posed problems arising in systems simulation. The need to consider the case of unbounded operators has been stimulated by numerous engineering applications where system models of physical processes yield this class of operators. In spite of the large popularity of these processes, relatively few results of the rigorous mathematical theory have been applied to practice. The main problem is that a number of results obtained in the classical case of bounded and self-adjoint operators is no longer valid or is incomplete in the case of unboundedness and non-self-adjointness, in general. Moreover, as is well known, there is no general spectral theory for unbounded and non-self-adjoint operators in a Hilbert space. Spectral theory treats only some special classes of such operators (see [GK69]). 2 1. Introduction
Problem Mathematical Physical Description Set of PDEs
Discretization and the Model Transformations in Discrete Approximation Continuous Domain
Simulation
Figure 1.1: The general procedure between the formulation of the physical problems and their computer realization. Here the physical problem is a distributed parameter system, described by a set of partial differential equations (PDEs).
The most elementary case of finite dimensional spaces and matrices as operators, which we assume to be known to the reader, is well established (see [Gan59]). It has a set of powerful tools which also might be useful to have in the more general case of infinite dimensional spaces and general operators. Of particular interest is the problem of a canonical form for arbitrary linear transformations in a Hilbert space as a general- ization of the Jordan form for matrices (some texts refer to it as the Jordan canonical form or the Jordan normal form). The Jordan structure for matrices is obtainable by using the special basis, which in the most simple case is formed by the set of linear inde- pendent eigenvectors, and in the more general case the associated vectors (also known as generalized eigenfunctions) are to be added to form such a basis. The case of bounded operators in a general Banach space can be treated essentially in the same way. M.V. Keldysh was the first who generalized the notion of the Jor- dan chain of vectors [Gan59] to a wide class of non-self-adjoint bounded operators (see [Kel51],[Kel71]). For that reason it was called the Keldysh chain. Further results were obtained e.g. in the work of M.G. Krein [Kre59], I. Gohberg [IGK90], V.B. Lidskii [Lid59], A.S. Markus [Mar88], M.A. Naimark [Nai62], A.V. Fursikov [Fur01b],[Fur01a] respectively. For further references see [GK69]. In this work, methods of spectral analysis are applied to the class of unbounded and non-self-adjoint operators with compact resolvent. It summarizes, extends and generalizes recent research, performed by the author in [DRS04a],[DRS04b],[DRS05],[DRS]. The basic principles are based on the method of Keldysh. The intention is to obtain a spectral decomposition of the solution of initial-boundary-value problems, which is adapted to the structure of the spatial differential operator which leads to its representation with respect to its canonical system and its adjoint. It is a consequence of the spatial separation that the resulting structure is well adapted to the simulation of the complete system with 1. Introduction 3 distributed parameters, i.e. the initial-boundary-value problem. The desired discrete simulation can be finally done by standard approximation and simulation techniques. To this end, one needs to counteract the non-regularity of the system operators by exploiting properties of the bases in suitable functional spaces formed by the eigenfunc- tions and associated functions of certain operators. There are two representations of the solution in the case of bounded operators, a series expansion and an integral formula with a complex contour integral. However, in the case of unbounded operators, the contour integral admits a generalization to sectorial operators. If an operator is sectorial, the solution of the corresponding initial-boundary-value problem is given by an holomorphic semigroup. Sectorial operators and holomorphic semigroups are basic tools in the the- ory of abstract parabolic problems, and in the solution of systems of partial differential equations of parabolic type. They will be introduced and investigated in the following chapters. The main goals of this work are: (i) to present the state of the art of the theory of unbounded and non-self-adjoint operators in the context of holomorphic semigroups, (ii) to expand the solution of initial-boundary-value problems with respect to the canonical systems of the sectorial operators, (iii) to design functional transformations for a gener- alized frequency domain representation, and for the subsequent discretization and (iv) to apply the presented methods to selected differential equations of technical interest. These four goals are reflected by Chapter 2 through Chapter 6. Chapter 2 contains the basic concepts of functional analysis such as spaces and opera- tors. This section is important in allowing us to be consistent with a rigorous mathemati- cal formulation of the problem. In the sense that for each problem one needs to define the corresponding space in which the problem is considered and the corresponding operators which reflect the physical behavior or relations. Moreover, the problem can possess some additional constraints such as boundary or initial conditions. Chapter 3 provides a summary of spectral theory of unbounded operators with compact resolvent. This detailed overview is included there because it is not easily accessible in the engineering literature and also some fundamental results are presented only in the Russian literature. The presentation is centered around general linear operators in Hilbert spaces and their canonical systems, i.e. their eigenvalues, eigenvectors and associated vectors. An important subclass form the so-called sectorial operators. They describe the evolution of dynamical systems through an appropriate holomorphic semigroup. This section closes with new results in the representation of a certain holomorphic semigroup by the canonical system of the underlying sectorial operator and the expansion of the generalized resolvent operator by its canonical system. Chapter 4 introduces initial-boundary-value problems which are defined by a sectorial differential operator with compact resolvent. It is shown how the methods compiled in Chapter 3 allow to represent the solution of initial-boundary-value problems by the 4 1. Introduction canonical systems of the corresponding spatial differential operators. In short, Chapter 4 presents a representation of the general solution of a wide class of problems frequently encountered in physics and engineering. Chapter 5 presents the main results of this contribution. It shifts the focus from mathematics to multidimensional systems theory. This shift is subtle, but important. Rather than emphasizing the solution of an initial-boundary-value problem, i.e. a func- tion of time and space, we consider the initial-boundary-value problem as the description of a multidimensional system. The representation of all functions under consideration by the canonical system of its spatial differential operator is then formalized through the introduction of a rather general functional transformation, the so-called multi-functional transformation (MFT). It describes the multidimensional system in the temporal and the generalized spatial frequency domain by an infinite set of one-dimensional systems. In turn, each one-dimensional system has a very simple structure. Its details are de- termined by the structure of the canonical system. The corresponding blocks are given in terms of the temporal and generalized spatial frequency variables. The uniform na- ture of these blocks makes it easy to apply standard discretization methods well-known for one-dimensional systems. This section concludes with a discrete-time, discrete-space approximation of multidimensional systems under consideration. The new results presented in this section may be summarized as follows: The structure of initial-boundary-value problems is mapped to the block structure of discrete-time, discrete-space multidimensional systems. These results provide a mathematically rigorous link between the theory of multidimensional systems (e.g. [Bos82],[DM84]) and their discrete computer implementation. The presented procedure is explained by examples of physical problems in Chapter 6. 5 2 Basic Notions from Functional Analysis
In this chapter we describe the appropriate context in which one can define and analyze the spectral properties of unbounded linear operators, particularly those which are closed and non-self-adjoint. The precise description of the operators will be the main focus of attention throughout this chapter. The reader is supposed to be familiar with the fundamental tools of applied functional analysis which will be used systematically throughout the text, otherwise we refer to [Bal76],[Yos80],[Gol66],[Has00].
2.1 Spaces
Before one can start to study the behaviour of operators one has to choose an appropriate space on which they act. It turns out that some properties of the operator can change depending upon the spaces on which it acts.
Definition 1 A set X is called a linear space (or vector space) over a field R (C) if the following conditions are satisfied:
1. an addition + is defined: for every elements x, y X there exists an associated ∈ element z of X, such that z = x + y;
2. x + y = y + x;
3. x +(y + z)=(x + y)+ z;
4. there exists a zero element of X, denoted by 0, such that x +0= x;
5. for every x X there exists an element x X such that x +( x)=0; ∈ − ∈ − 6. a scalar multiplication is defined: for every element x Xand each α R(C) there ∈ ∈ exists an associated element y of X, such that y = αx;
7. α(βx)=(αβ)x;
8. 1 x = x; · 6 2. Basic Notions from Functional Analysis
9. α(x + y)= αx + αy;
10. (α + β)x = αx + βx;
Definition 2 A linear space X is called a normed linear space, if for every x X, there ∈ is associated a real number x , the norm of the element x, such that: || ||X 1. x 0 and x =0 iff x = 0; || ||X ≥ || ||X 2. αx = α x , α R(C); || ||X | |·|| ||X ∈ 3. x + y x + y (triangle inequality); || ||X ≤ || ||X || ||X Definition 3 A sequence x in a normed linear space X converges to the element { n} x X if x x 0, n . ∈ || n − ||X → →∞ Definition 4 A sequence x in a normed linear space X is said to be a Cauchy sequence { n} if ε> 0 N n>N, p N : x x <ε. ∀ ∃ ∀ ∀ ∈ || n+p − n||X Definition 5 If every Cauchy sequence is convergent, a normed linear space is said to be complete. A Banach space is a complete normed linear space.
Definition 6 A subset S of a normed linear space X is dense in X if the closure of S is the entire space X (S¯ = X).
Definition 7 A subset S of a normed linear space X is called compact if every (infinite) sequence in X has a convergent subsequence.
Definition 8 An inner product (or scalar product) on a linear space X defined over the field R (C) is a map ( , ) : X X R(C) such that · · × −→ 1. (x, x) 0 and (x, x)=0 iff x = 0; ≥
2. (x, y)∗ =(y, x) for all x, y X; ∈ 3. (αx + βy,z)= α(x, z)+ β(y, z) for all x, y, z X, α,β R(C); ∈ ∈ A linear space X with an inner product ( , ) is called an inner product space. · · We can consider any inner product space as a normed linear space (X, ) by || · ||X defining the norm as = (x, x). ||·||X p Definition 9 A Hilbert space is an inner product space which is complete as a normed space under the induced norm. 2.1. Spaces 7
The spaces taken into consideration are the usual spaces of complex valued functions defined on RM or on an open set Ω of RM . The following two classes of Banach spaces are widely used in functional analysis and the theory of differential equations.
Definition 10 [Ada75] The Lebesgue space L (Ω), 1 p< + , is the set of all complex- p ≤ ∞ valued functions y(x)= y(x1, x2,...,xM ) defined in Ω and such that
1 p p y Lp(Ω) = y(x) dx < , (2.1) k k | | ∞ ZΩ where the integral is taken in the Lebesgue-sense.
Definition 11 [Ada75] Let k be a positive integer, and 1 p< + . The Sobolev space ≤ ∞ Wk p(Ω) is the set of all complex-valued functions y(x) = y(x1, x2,...,xM ) defined on Ω and such that
1 p a p y Wk(Ω) = D y(x) dx < , (2.2) k k p | | ∞ Z a k Ω |X|≤ where a = (α1,...,αM ) is a vector of nonnegative integers, a = α1 + + αM and a ∂α1 ∂αM | | ··· D = α . . . α is the weak or distributional partial derivative of order a . ∂x 1 ∂x M 1 M | | The following inequalities will be useful in the next chapters. Let 1
fgdx fg dx f L g L (2.4) ≤ | | ≤k k p(Ω) ·k k p′ (Ω) Z Z Ω Ω
holds for any functions f L (Ω) and g L ′ (Ω). ∈ p ∈ p Minkowski’s inequality
f + g L f L + g L (2.5) k k p(Ω) ≤k k p(Ω) k k p(Ω) holds for any f,g L (Ω). ∈ p Minkowski’s integral inequality
1 1 p p p p dx f(x1, x2)dx dx f(x1, x2) dx (2.6) 2 1 ≤ 1 | | 2 Z Z Z Z Ω2 Ω1 Ω1 Ω2 holds for any integrable function f(x , x ) defined on the set Ω Ω (x Ω RM1 , x 1 2 1× 2 1 ∈ 1 ⊂ 2 ∈ Ω RM2 ) and such that integral in the right part of (2.6) is finite. 2 ⊂ 8 2. Basic Notions from Functional Analysis
2.2 Linear operators
Let X,Y be Banach spaces over the complex field C and A be a map acting from X to Y. Denote by D(A) the set of elements from X where the mapping A is defined or its domain, by Ran(A) = y Y : Ax = y, x D(A) its image or range and by { ∈ ∈ } Ker(A)= x D(A) : Ax =0 its kernel. { ∈ } Definition 12 The mapping A : X Y is a linear operator if → i) the set D(A) is linear subspace of X; ii) the mapping A is linear on D(A), i. e.
A(α x + α x )= α Ax + α Ax , x , x D(A), α , α C. (2.7) 1 1 2 2 1 1 2 2 ∀ 1 1 ∈ 1 2 ∈ Note that D(A) is a linear subspace of X, while Ran(A) is a linear subspace of Y . The case D(A) = X and/or Ran(A) = Y may occur. 6 6
2.2.1 Unbounded operators
One of the basic instruments in functional analysis is a bounded linear operator. The need to consider the case of unbounded operators has been stimulated by numerous engineering applications where system models of the physical process yield this class of operators. Moreover, the most differential operators are unbounded when considered as acting on any usual Banach or Hilbert spaces.
Definition 13 The linear operator A : X Y is bounded if → i) D(A)= X; ii) there exists a constant c > 0 such that Ax c x for all x X. The norm k kY ≤ k kX ∈ A of the linear bounded operator A is defined as k k Ax A = sup k kY , (2.8) k k x=0 x X 6 k k Remark 1 The infimum of all such constants c is equal to the norm of the operator A, inf(c)= A . k k Thus, the linear operator A can be unbounded if the condition i) or/and the condition ii) of Definition 13 are violated. It is known that in the case where X and Y are normed linear spaces the boundedness of the linear operator A is equivalent to its continuity. It is said that the linear operator A : X Y is continuous at the point x D(A) if x x 0 leads to Ax → 0 ∈ k n − 0kX → k n − Ax 0. The linear operator A is continuous if it is continuous at an arbitrary point 0kY → x D(A). 0 ∈ 2.2. Linear operators 9
Example 1 [KF81] Unbounded operator
Denote by C([0, 2π]) the Banach space of all continuous complex valued functions defined on the compact interval [0, 2π] R with the norm x C = max x(t) ( x C = x ). 0 t 2π ∞ ∈ k k ≤ ≤ | | k k k k Define the differential operator A : C([0, 2π]) C([0, 2π]) by the formula Af = f ′, → with D(A) = C1([0, 2π]) C([0, 2π]), where C1([0, 2π]) is the set of all continuously ⊂ differentiable functions on [0, 2π]. It is obvious that C1([0, 2π]) is a dense subspace in C([0, 2π]). This operator is linear since
d dx dx (α x + α x )= α 1 + α 2 . dt 1 1 2 2 1 dt 2 dt
This operator is not bounded. To show this consider, for example, the sequences xn(t)= dxn dxn sin nt, xn C 1, C = max n cos nt = n. Hence sup C =+ , which proves dt 0 t 2π dt k k ≤ k k ≤ ≤ | | n k k ∞ the unboundedness of the operator A.
Example 2 [KF81] Bounded operator
In contrast to the case above, consider the differential operator A acting from the Banach 1 space C ([0, 2π]) with the norm x C1 = max x(t) + max x′(t) to the Banach space 0 t 2π 0 t 2π k k ≤ ≤ | | ≤ ≤ | | C([0, 2π]) by the same formula: Af = f ′. This operator is linear and bounded since
1 Ax C = x′ C = max x′(t) max x(t) + max x′(t) = x C1 , x C ([0, 2π]). 0 t 2π 0 t 2π 0 t 2π k k k k ≤ ≤ | | ≤ ≤ ≤ | | ≤ ≤ | | k k ∀ ∈
Remark 2 The given examples show that the operator given by the same formula can be bounded or unbounded with respect to different norms associated with the underlying Banach spaces.
2.2.2 Adjoint operators
Now we consider Hilbert spaces H1 and H2 over the field C with inner products denoted as ( , ) and ( , ) , respectively. L is assumed to be a linear operator acting from H · · H1 · · H2 1 into H with the domain D(L) H . 2 ⊂ 1
Definition 14 Let f ∞ be a sequence of elements f D(L). If from f f, n { n}n=1 n ∈ n → →∞ and Lf g, n it follows that f D(L) and Lf = g then the operator L is called n → →∞ ∈ a closed operator.
Definition 15 A bounded operator is called a compact operator if it maps each bounded
subset of H1 into a compact subset of H2. 10 2. Basic Notions from Functional Analysis
Definition 16 Let the operator L have a dense domain D(L) in H1. The adjoint operator
L† is that operator with domain D(L†) containing all those elements g H for which ∈ 2 there exists an element h H such that the following equality holds: ∈ 1 f D(L) (Lf,g) =(f, h) . (2.9) ∀ ∈ H2 H1
In this case, by definition, L†g = h.
Equivalently we can write:
f D(L) , g D(L†) (Lf,g) =(f, L†g) . (2.10) ∀ ∈ ∀ ∈ H2 H1
Lemma 1 [Yos80] Let the operator L have a dense domain D(L) in H1. Assume that 1 1 1 the operator L− exists and its domain D(L− ) is dense in H2. Then (L− )† exists and satisfies
1 1 (L− )† =(L†)− . (2.11)
In the following we assume that the Hilbert spaces Hi are identical, i.e H1 = H2 = H. The scalar product and its corresponding induced norm will now be simply denoted by ( , ) and , respectively. · · k·k
2.2.3 Linear differential forms In this section we introduce the basic concept of linear differential forms and their adjoints. We will consider the set C(n)([a, b]) of continuously differentiable functions up to the order n over finite interval [a, b]. An arbitrary linear differential form l applied to an arbitrary element y C(n)([a, b]) ∈ is defined as n (n) (n 1) (ν) l(y)(x) := pn(x)y (x)+ pn 1(x)y − (x)+ + p0(x)y(x)= pν(x)y (x), (2.12) − ··· ν=0 X where the functions p (x), ν =0, , n are continuously differentiable up to the order ν ν ··· over the given interval [a, b].
2.2.3.1 Homogeneous boundary conditions
′ (n 1) ′ (n 1) Denote by y ,y , ,ya − and y ,y , ,y − the values of the function y(x) and its a a ··· b b ··· b derivatives at the points x = a and x = b, respectively. Let m linear independent forms be given by
µ µ ′ µ (n 1) Uµ(y) :=α0 ya + α1 ya + + αn 1ya − + ··· − (2.13) µ µ ′ µ (n 1) µ µ C +β0 yb + β1 yb + + βn 1yb − , αi , βi . ··· − ∈ 2.2. Linear operators 11
For an arbitrary number m 2n of such kind of forms we say that the equalities ≤
Uµ(y)=0, µ =1,...,m (2.14) represent homogeneous boundary conditions. If m = 2n, it is obvious, that from (2.14) ′ (n 1) follow the equalities y =0,y =0, ,ya − = 0. a a ···
2.2.3.2 Green’s formula and associated forms
Let us introduce the following vectors with n components
ya yb za zb (1) (1) (1) (1) ya yb za zb y = (2) , y = (2) , z = (2) , z = (2) . (2.15) a ya b yb a za b zb · · · · (n 1) (n 1) (n 1) (n 1) ya − y − za − z − b b b (ν) We consider the integral pν (x)y (x)z(x)∗dx for ν =0, , n. Repeated integration by a ··· parts shows that for any Ry, z C(n)([a, b]) and p (x) C(ν)([a, b]) the equality ∈ ν ∈ b (ν) (ν 1) ′ (ν 2) p (x)y (x)z(x)∗dx = p (x)z∗(x)y − (x) (p (x)z∗(x)) y − (x)+ ν ν − ν ··· Za x=b b ν 1 (ν 1) ν (ν) +( 1) − (p (x)z∗(x)) − y(x) +( 1) y(x) p (x)z(x)∗ dx ··· − ν − ν x=a Za (2.16) holds. The sum of all equalities (2.16) over the indices ν = 0, , n gives us that the ··· differential form l(y) satisfies the following identity
(l(y), z)= P (η,ζ)+(y, l†(z)). (2.17)
Here n n (n) n 1 (n 1) ν (ν) l†(z)=( 1) (pn∗ z) +( 1) − (pn∗ 1z) − + + p0∗z = ( 1) (pν∗z) (2.18) − − − ··· − ν=0 X represents an associated (formally adjoint) differential form to l(y) and
n ν x=b ν i (ν i) (i 1) P (η,ζ)= ( 1) − (p z∗) − y − (2.19) − ν ν=1 i=1 x=a X X is a bilinear form of the variables η = (y , y ) and ζ = (z , z ). The relation (2.17) is a b a b called the Green’s formula. 12 2. Basic Notions from Functional Analysis
If we introduce the matrix Q as follows,
Q =
(1) n 1 (n 1) n 2 n 1 p1 p2 + +( 1) − pn − p2 p3 p4 ( 1) − pn 1 ( 1) − pn − ··· − − − ··· − − − (1) n 2 (n 2) n 2 p p + +( 1) − pn − p p p ( 1) − p 0 2 − 3 ··· − − 3 4 − 5 ··· − n (1) n 3 (n 3) p3 p4 + +( 1) − pn − p4 p5 p6 0 0 − ··· − − − ··· (2.20) ··· ······ · · (1) (2) pn 2 pn 1 + pn pn 1 pn 0 0 0 − − − − − ··· (1) pn 1 pn pn 0 0 0 0 − − − ··· p 0 0 0 0 0 n ··· the bilinear form P (η,ζ) can be written in compact form as
P (η,ζ)= yT Q(b)z yT Q(a)z . (2.21) b b − a a
2.2.4 Linear differential operators Define the differential operator L : C(n)([a, b]) C([a, b]) as follows: → i) the domain D(L) of the operator is
D(L)= y C(n)([a, b]) : U (y)=0, µ =1,...,m ; (2.22a) ∈ µ ii) the operator is given by the formula
Ly = l(y), y D(L). (2.22b) ∈ Assigning different boundary conditions (2.14) to a fixed differential form l will generate different differential operators L with different domains, in general.
Remark 3 The precise specification of the domain of the operator is very important since it turns out that the same linear differential form (2.12) with different boundary conditions (2.14) produces different differential operators L, in general. For these reasons when using the term ”differential operator” we shall understand that we have already chosen the boundary conditions if we are thinking in more applied terms, or that we have already chosen the precise domain of definition of the operator if we are thinking in abstract terms.
It is obvious that D(L) is a linear subspace of C(n)([a, b]), and D(L) = C(n)([a, b]) if there are no restrictions of the form (2.22a). Therefore, the linear differential form l itself is a linear differential operator with domain D(l)= C(n)([a, b]). Also, we always have the inclusion D(L) D(l). ⊂ 2.2. Linear operators 13
2.2.4.1 Adjoint homogeneous boundary conditions and the adjoint operators
We consider m linear independent forms U1, , Um of the variables ′ (n 1) ′ (n 1) ··· y ,y , ,ya − ,y ,y , ,y − given by a a ··· b b ··· b
ya 1 1 1 1 1 1 (1) α0 α1 αn 1 β0 β1 βn 1 ya U ··· − ··· − 1 2 2 2 2 2 2 α0 α1 αn 1 β0 β1 βn 1 U ··· − ··· − 2 · (n 1) ya − · · ··· · · · ··· · = · . yb · · ··· · · · ··· · · (1) yb · · ··· · · · ··· · · m m m m m m α0 α1 αn 1 β0 β1 βn 1 Um ··· − ··· − m 1 m 2n · × × (n 1) y − b 2n 1 × (2.23) Using block-matrices A , B and the vector U of appropriate dimensions we rewrite 1 1 1 (2.23) in more compact form as
y a A B = U1 . (2.24) 1 1 " y # h i b h i
If m < 2n we can add 2n m arbitrary linear independent forms U , , U in − m+1 ··· 2n such a way that the resulting forms U , , U are linear independent as well. Then we 1 ··· 2n can write
1 1 1 1 1 1 α0 α1 αn 1 β0 β1 βn 1 ya U1 ··· − ··· − 2 2 2 2 2 2 (1) α0 α1 αn 1 β0 β1 βn 1 ya U2 ··· − ··· − · · ··· · · · ··· · · · (n 1) ya − · · ··· · · · ··· · = · , yb · · ··· · · · ··· · · m m m m m m (1) α0 α1 αn 1 β0 β1 βn 1 yb ··· − ··· − · · · ··· · · · ··· · · · 2n 2n 2n 2n 2n 2n (n 1) α0 α1 αn 1 β0 β1 βn 1 yb − U2n ··· − ··· − 2n 2n 2n 1 2n 1 × × (2.25)× or equivalently 14 2. Basic Notions from Functional Analysis
A B y U1 1 1 a = , (2.26) A B y U " 2 2 # " b # " 2 # where A and B are block-matrices of appropriate dimensions and U is a vector of 2 2 2 length 2n m. − In this case, by inverting the matrix in 2.26, each y and y can be represented as a a b linear combination of these forms
y C D U1 a = 1 1 . (2.27) y C D U " b # " 2 2 # " 2 # Thus y = C U + D U , (2.28a) a 1 1 1 2 y = C U + D U . (2.28b) b 2 1 2 2 Substitution of these representations into P (η,ζ) yields
P (η,ζ)=(C U + D U )T Q(b)z (C U + D U )T Q(a)z 2 1 2 2 b − 1 1 1 2 a = UT [CT Q(b)z CT Q(a)z ]+ UT [DT Q(b)z DT Q(a)z ] (2.29) 1 2 b − 1 a 2 2 b − 1 a T T = U1 V2 + U2 V1, where
T T T V2 = [V2n,V2n 1,...,V2n m] = C Q(b)zb C Q(a)za, (2.30a) − − 2 − 1 T T T V1 = [V2n m 1,V2n m 2,...,V1] = D Q(b)zb D Q(a)za, (2.30b) − − − − 2 − 1 and Vi, i =1,..., 2n are linear independent forms of the variables za and zb, too. Now we rewrite the bilinear form (2.29) in terms of Ui and Vi
P (η,ζ)= U1V2n + + UmV2n m+1 + Um+1V2n m + + U2nV1. (2.31) ··· − − ··· Substitution of the prime boundary conditions U =0, U =0, , U = 0 yields 1 2 ··· m
P (η,ζ)= Um+1V2n m + + U2nV1. (2.32) − ··· We say that the expressions
V1 =0, V2 =0, ,V2n m =0 (2.33) ··· − are the adjoint boundary conditions associated with the prime boundary conditions: U1 = 0, U =0, , U =0. The purpose of the special partitioning into two sets U and V 2 ··· m i i 2.2. Linear operators 15 is that under conditions (2.14) and (2.33) the expression P (η,ζ) will vanish, no matter how the remaining conditions are chosen.
Thus, we have the adjoint linear differential form l†(z) and the adjoint boundary conditions V1, ,V2n m. ··· − Since the adjoint operator L† to the operator L can be defined only in some Hilbert space (see Definition 16), let C(n)([a, b]) H, where H is the Hilbert space with the scalar b ⊂ product (y, z)= y(x)z∗(x)dx (for example we can set H = L2([a, b])). a Therefore, theR adjoint linear differential form l†(z) and the adjoint boundary conditions
V1, ,V2n m generate the adjoint operator L† defined as: ··· − i’) the domain D(L†) is
(n) D(L†)= z C ([a, b]) : V (z)=0, µ =1,..., 2n m ; (2.34a) { ∈ µ − } ii’) the operator acts by the formula
L†z = l†(z), z D(L†). (2.34b) ∈
2.2.4.2 Nonhomogeneous boundary conditions
In the case of nonhomogeneous boundary conditions m
Uµ(y)= φµ, φµφµ∗ > 0 (2.35) µ=1 X the differential operator L is defined by the following conditions i) the domain D(L) of the operator is
D(L)= y C(n)([a, b]) : U (y)= φ , µ =1,...,m ; (2.36a) ∈ µ µ ii) the operator is given by the formula
Ly = l(y), y D(L). (2.36b) ∈ When no confusion may arise, we introduce some new notations. In particular, denote by
l the differential form of the operator under consideration, by L0 the operator with ho-
mogeneous boundary conditions and by Lφ the operator with nonhomogeneous boundary conditions. Thus
L : D(L )= y C(n)([a, b]) : U (y)=0, µ =1,...,m , (2.37) 0 0 { ∈ µ } and L : D(L )= y C(n)([a, b]) : U (y)= φ , µ =1,...,m . (2.38) φ φ { ∈ µ µ }
Remark 4 Note that due to the nonhomogeneous boundary condition the operator Lφ is no longer linear. 16 2. Basic Notions from Functional Analysis
2.3 Chapter summary
This chapter described briefly the basic notions and facts from functional analysis and accompanied by the appropriate author’s interpretation. More exactly, the main focus was on linear operators in Hilbert spaces. It has been shown that to define an operator in rigorous mathematical terms, in addition to its formal definition it is necessary to define the space where operator acts and to specify its domain. The corresponding space contains those elements which satisfy the natural constraints such as continuity, differentiability or integrability. In turn, the domain of the operator is a subspace of this abstract or general space, and consists of such a set of elements which satisfy the desired properties or more specific restrictions generated in many cases by the nature of physical processes. Also, since we have defined the Hilbert space, it is possible to consider some relations between the elements such as scalar product and norm. And of course, as we will see in the next chapters, the notion of the adjoint operator is a very powerful tool in the investigation of the prime operator. In this chapter we have restricted our detailed consideration to the case of ordinary differential operators of order n which act on the subspace C(n)([a, b]) of scalar functions of the Hilbert space L2([a, b]). But in a similar way the case of spaces with functions f : R CM and ordinary differential operators on it can be treated (see [Naj68]). → Moreover, the case of partial differential operators can be also considered analogously. However, since partial differential operators include more complicated boundary con- ditions (due to the multidimensional argument), the generalization to a such compact presentation described above involves the corresponding generalization of the 1D Green’s formula (2.17) to its MD analogy (see [Tay], [Sho77]). Also, the strong requirements to have a continuous partial derivative for functions in a usual (classical) sense (C(n)([a, b])), can be substantially weakened by introducing the corresponding Sobolev spaces and weak or distributional partial derivatives. For further details about Sobolev spaces for use in some application we refer to [Ada75], [Eva02]. On the one hand, the main attempt has been made in this chapter to present all necessary basic material of operator theory in sufficient generality, in order to be able to manipulate in abstract terms. On the other hand, it was intended to cover in detail the procedure of defining the operators in special cases, in order to understand the applied aspects of the presented theory. 17 3 Spectral Theory of Operators
In the theory of linear and time-invariant systems and network theory one of the funda- mental concepts is the system function or transfer function. It helps to describe the sys- tem, its behaviour and its transfer properties from the input to the output (see [GRS01]). Most applications lead to models with meromorphic transfer functions. This means such functions are completely determined by all their poles and zeroes up to a multiplicative constant. We remark that the number of poles and zeroes might be infinite. In any case, the class of rational functions is included in the more general class of meromorphic functions. Since prescribed poles and zeroes are easy to implement in realizing structures, frequency-domain models are often used. In principle, there is a similar situation in the general operator theory, as will become obvious in this chapter. The application of spectral methods in the presence of general operators provides similar advantages as the introduction of transfer function above. In this chapter we will presuppose a closed linear operator L : H H, the domain D(L) of 7→ which is dense in the Hilbert space H. In addition we will consider the operator
L =(sI L) : H H, s C, s − 7→ ∈
where I is the identity operator. The investigation of the operator Ls is called the spectral theory for the operator L. It includes the characterization of the distribution of the values
of s for which Ls has an inverse and the properties of this inverse when it exists. Similar
to the former case, the general theory of the inverse of Ls leads to spectral decomposition
based on a certain set of poles. As we shall see later, the inverse of the operator Ls plays a similar role in obtaining the solution of initial-boundary-value problems as the system function in common systems.
3.1 The resolvent operator
Definition 17 If s0 is such that the image Ran(Ls0 ) is dense in H and Ls0 has a bounded 1 inverse (s0I L)− , we say that s0 is a regular point of L and we denote this inverse 1 − (s I L)− by R(s , L) and call it the resolvent operator of L at s . 0 − 0 0 Definition 18 The set of all regular points of the operator L is called the resolvent set, and is denoted by ρ(L). The set C ρ(L) is called the spectrum of the operator L and is \ denoted by σ(L). 18 3. Spectral Theory of Operators
In the following we assume that neither the spectrum nor the resolvent set is empty. Thus, for any s ρ(L) the resolvent operator is defined as ∈
1 R(s, L)=(sI L)− , s ρ(L). (3.1) − ∈
Let L† : H H be the adjoint operator of L. Since the operator L is assumed to be 7→ closed with dense domain D(L) in H, it follows that its adjoint L† is also closed, and its
domain D(L†) is dense in H, too (see e.g. [Bal76],[Yos80],[Gol66]). From the equality
1 1 R(s, L†)=(sI L†)− = [(s∗I L)− ]† = R(s∗, L)†, s ρ(L†) (3.2) − − ∀ ∈ it follows immediately that
ρ(L†)= ρ(L)∗, and σ(L†)= σ(L)∗. (3.3)
In addition, any resolvent satisfies the Hilbert identity (see e.g. [Bal76],[Yos80])
R(s, L) R(w, L)=(w s)R(s, L)R(w, L), s,w ρ(L), (3.4) − − ∈ which can easily be verified by use of (3.1).
Definition 19 The complex number λ σ(L) is called an eigenvalue of the operator L, ∈ if Ker(λI L) = 0 . (3.5) − 6 { } Any vector e = e(λ) Ker(λI L) 0 , is called an eigenvector of the operator L ∈ − \ { } corresponding to the eigenvalue λ.
It will turn out for all relevant cases that the eigenspaces Ker(λI L) have finite dimension, − i.e. P = P (λ) = dim Ker(λI L) < . Hence, we can assume that a basis B of this − ∞ 0 eigenspace is given by the set B = B (λ)= e , e , , e (3.6) 0 0 1,0 2,0 ··· P,0 of the linear independent eigenvectors ep,0 corresponding to the eigenvalue λ. For later purposes we have introduced the subscript 0. Using the simple property (3.4) of the resolvent operator of L we are already able to obtain a useful characterization of the spectrum of L.
Lemma 2 [Bal76] Let L be a closed operator with dense domain in H. If there exists s ρ(L) such that the resolvent operator R(s , L) is compact, then 0 ∈ 0 R(s, L) is compact in any point of the resolvent set ρ(L); moreover, the spectrum σ(L) consists of a discrete set of points and hence is denumerable. 3.1. The resolvent operator 19
The important statement of this lemma is the fact that it guaranties that the spectrum of L consists of a countable set of points. Consequently, all those values λ are isolated points. They are the poles of the resolvent operator R(s, L). It can therefore be represented in a certain neighborhood of such a singular point λ by the following Laurent series expansion
+ ∞ R(s, L)= (s λ)νR , (3.7) − ν ν= X−∞ where the coefficients Rν are operators and can be represented by the formula
1 ν 1 R = (s λ)− − R(s, L)ds (3.8) ν 2πj − s Iλ =r | − | for sufficiently small r> 0. Here j denotes the imaginary unit.
The next two lemmas contain some properties of the operators Rν in the series expan- sion (3.7).
Lemma 3 [Yos80]
a) the operators Rν in (3.7) commute with each other and the operator L;
2 b) the operator R 1 is a projection operator, i.e. R 1 = R 1; − − − c) the following formula is valid
Rν 1 +(λI L)Rν = δ0,νI, ν Z. (3.9) − − ∈
Here δµ,ν is the standard Kronecker symbol defined as
1, µ = ν δ = . (3.10) µ,ν 0, µ = ν ( 6 Lemma 4 [Fur01a][Fur01b] Let L be a closed operator with dense domain and compact resolvent operator and λ σ(L), then ∈ d) the Laurent series (3.7) in the neighborhood of the eigenvalue λ has a finite number M of terms with negative powers of (s λ) − + ∞ R(s, L)= (s λ)νR , (3.11) − ν ν= M X− i.e.
R =0, ν< M; (3.12) ν − 20 3. Spectral Theory of Operators
e) the operator R 1 can be characterized by the condition −
M Ran(R 1) =Ker([λI L] ); (3.13a) − −
the operators R ν , ν> 0 have finite dimensional images −
dim Ran(R ν ) < , ν> 0 (3.13b) { − } ∞ and satisfy the relations
ν ν R ν 1 =( 1) (λI L) R 1, ν =1,...,M 1. (3.13c) − − − − − −
It is obvious that the Laurent series expansion of the adjoint resolvent operator R(s, L†)
in the neighborhood of λ∗ σ(L†) is ∈ + ∞ ν R(s, L†)= (s λ∗) R† . (3.14) − ν ν= M X−
In particular, it has the same order M of the pole λ∗ and the corresponding operators Rν† are adjoint to the operators Rν. Thus, by analogy with Lemma 3 and Lemma 4 we have
Z Rν† 1 +(λ∗I L†)Rν† = δ0,νI, ν , − − ∈
dim Ran(R† ν ) = dim Ran(R ν) < , ν> 0, { − } { − } ∞ (3.15) M Ran(R† 1) = Ker([λ∗I L†] ), − − ν ν R† ν 1 =( 1) (λ∗I L†) R† 1, ν =1,...,M 1. − − − − − −
3.1.1 Canonical systems of the prime and the adjoint operator
An important part of matrix theory is the transformation which leads to Jordan form of a matrix and the corresponding Jordan chains of vectors (the sets of eigenvectors and generalized eigenvectors) (see [Gan59]). There is a similar situation in the general spectral theory of operators. M.V. Keldysh was the first who generalized the notions of the Jordan chains of vectors to a more general classes of operators (see [Kel51],[Kel71]). He has considered the canonical systems for the operator pencil of the form L(λ) = n 1 n 1 n n A0 + λBA1 + + λ − B − An 1 + λ B in the Hilbert space H, where A0, A1,...,An 1 ··· − − are arbitrary compact operators, B is a self-adjoint operator and Bf = 0, if f = 0. 6 6 Here we reformulate the definitions introduced by Keldysh in a context which is suit- able for our purpose. We consider a closed linear operator L : H H, the domain D(L) 7→ of which is dense in the Hilbert space H. 3.1. The resolvent operator 21
Definition 20 For a fixed eigenvalue λ and a fixed eigenvector e , where 1 p P (λ) p,0 ≤ ≤ (P (λ) is the dimension of the eigenvector space belonging to the eigenvalue λ, see (3.6)), we say that the vector e = 0 is an associated vector (or generalized eigenvector) of p,m 6 order m of the eigenvector ep,0, if it satisfies the following equations
(λI L)e =0, − p,0 ep,0 +(λI L)ep,1 =0, − (3.16) , ········· ep,m 1 +(λI L)ep,m =0. − −
The maximum possible number m of associated vectors is denoted by Mp = Mp(λ) and
(Mp + 1) is called the multiplicity of the eigenvector ep,0. The set
E = E (λ)= e , e , , e (3.17) p p p,0 p,1 ··· p,Mp is called the complete chain (or Keldysh chain) of associated vectors that belongs to the eigenvector ep,0 of λ (note that the eigenvector itself is also included).
Remark 5 When no confusion may arise, we will write P instead of P (λ).
It will be shown later that all multiplicities Mp + 1 are less or equal to the order M of the corresponding pole in the Laurent series expansion (3.7) for the considered eigenvalue λ, M +1 M. p ≤ Definition 21 For each fixed eigenvalue λ we consider the union
P (λ)
E(λ)= Ep(λ) (3.18) p=1 [ over the set of linear independent eigenvectors ep,0, p = 1,...,P (λ). E(λ) is called a canonical system of the operator L corresponding to the fixed eigenvalue λ. P (λ) The number N(λ)= M1 +1+ M2 +1+ + MP (λ) +1= P (λ)+ Mp is called the ··· p=1 multiplicity of the considered eigenvalue λ. P
In the following we assume without loss of generality that the sequence of eigenvectors is ordered with respect to their multiplicities (M 1) = M M M 0. − 1 ≥ 2 ≥···≥ P (λ) ≥ To illustrate the notion used in (3.16)-(3.18), the canonical system in (3.18) can be arranged as 22 3. Spectral Theory of Operators
E(λ)= E (λ), E (λ), , E (λ), , E (λ) 1 2 ··· p ··· P (λ) e e e e 1,0 2,0 ··· p,0 ··· P (λ),0 e1,1 e2,1 ep,1 eP (λ),1 ··· ··· · · ···· ···· e1,m e2,m ep,m eP (λ),m (3.19) ··· ··· = . · · ···· ···· e · · ···· ··· P (λ),MP (λ) e p,Mp · · ··· e2,M2 · e 1,M1 The first row contains the set of P (λ) eigenvectors, and the p th column shows the − complete chain of associated vectors corresponding to the eigenvector ep,0 as defined in (3.16).
In addition to the canonical system E(λ) with the elements ep,m in (3.18) we define
another canonical system E† consisting of elements ǫp,m, that corresponds to the eigen-
value λ∗ of the adjoint operator L†. This collection of eigenvectors together with the corresponding set of associated vectors satisfies the following equations
(λ∗I L†)ǫ =0, − p,0 ǫ +(λ∗I L†)ǫ =0, p,0 − p,1 , ········· (3.20) ǫp,m 1 +(λ∗I L†)ǫp,m =0, − − , ········· ǫ +(λ∗I L†)ǫ =0, p,Mp−1 − p,Mp for each p =1, ,P (λ∗). ··· By analogy with (3.16)-(3.18), we introduce the complete chains
E† = E†(λ∗)= ǫ , ǫ , , ǫ (3.21) p p p,0 p,1 ··· p,Mp and the canonical system of the operator L† for a fixed eigenvalue λ∗ as
P (λ∗)
E†(λ∗)= Ep†(λ∗). (3.22) p=1 [ In addition, we will use the following notations :
E = E(λi) (3.23) λi ∪σ(L) ∈ 3.1. The resolvent operator 23 is called the canonical system of the operator L;
E† = E†(λi∗) (3.24) λ∗ ∪σ(L†) i ∈
is called the canonical system of the adjoint operator L†.
We have to mention here that in the general case neither E nor E† will be an orthogonal set. However, the following theorem is true
Theorem 1 After proper normalization of ǫp,Mp for each eigenvalue the introduced canon- ical systems E and E† are biorthogonal
(ep,m(λi), ǫl,n(λw∗ )) = δi,wδp,lδm,Mp n. (3.25) −
Proof. see Remark 18 in Appendix. To illustrate the biorthogonality relations of the canonical systems we assume that it has for example only 4 (Mp = 3) elements in the chain Ep(λ). In Figure 3.1 solid lines mean nonorthogonal relations, and dotted lines show orthogonality.
ep,0 ǫp,3
ep,1 ǫp,2
ep,2 ǫp,1
ep,3 ǫp,0
Figure 3.1: Biorthogonality relations of a canonical systems
Definition 22 A system A = ai i Z, ai H is called minimal if and only if none of { } ∈ ∈ the elements a A is contained in the closed linear span of the remaining elements i ∈ a / span(A a ). (3.26) i ∈ \{ i}
From the biorthogonality-relation we can deduce that the canonical systems E and E† are both minimal systems. The eigenvectors and associated vectors of the canonical system are linear independent.
Remark 6 Note, that for an arbitrary linear independent system it does not necessary follow the existence of the corresponding biorthogonal system. For example, the sequence i a = x ∞ is linear independent system in L ([0, 1]) and has no biorthogonal system. { i }i=0 2 24 3. Spectral Theory of Operators
For each fixed eigenvalue the corresponding elements of the canonical system form a subspace and its dimension is equal to the order of the eigenvalue. Therefore, one can set the natural question about the basis property of such a system.
Remark 7 Currently, the development of so-called completeness and basis properties of the introduced canonical systems for the considered case of operators is a subject of future work. But we note that the fundamental theorem concerning completeness of the canonical system was presented by Keldysh (see [Kel71]). That the canonical system constitutes a
basis was established by Ilyin for the case of differential operators of order n in L2 (see [Il’76]).
3.1.2 Expansion of the resolvent operator by the canonical sys- tem We consider closed operators L with dense domain and compact resolvent. Lemma 2 states that in this case the spectrum σ(L) consists only of a discrete set of eigenvalues and the resolvent operator can be represented by the expansion (3.11). In the following we will formulate some of these important properties of the resolvent operator which are needed below.
Agreement 1 We introduce the following notation: for u H and v H the symbol ∈ ∈ [u, v] represents the operator S : H H defined on functions f H by the following → ∈ formula:
[u, v]f = Sf =(f, v)u, (3.27) where ( , ) is the inner product in H. · · Obviously, the operator S in (3.27) projects the whole space H onto the onedimensional space spanned by the vector u. Consequently the operator S has rank 1. Since for f,g H the following equality holds ∈
(Sf,g)=((f, v)u,g)=(f, v)(u,g)=(f, (u,g)∗v)=(f, (g,u)v), (3.28) the adjoint operator S† acts on element g H as ∈
[u, v]†g = S†g =(g,u)v. (3.29)
The expansion of the resolvent operator by the canonical system was established first by M. Keldysh [Kel51] for the case of a compact operator pencil. But for our purpose we extend his results to the class of unbounded and non-self-adjoint operators L with compact resolvent. We use the fact that the class of compact operators excludes the class of operators with compact resolvent. Lemma 5 states this fact. 3.1. The resolvent operator 25
Lemma 5 [Bal76] A compact operator cannot have a compact resolvent.
Using the presented results of lemmas 2-4 we are now able to give an exact represen- tation of the resolvent comprised in theorem 2. Theorem 2 [Fur01b][Fur01b] Let L be a closed operator with dense domain, compact resolvent operator R(s, L), and non-empty spectrum σ(L). For an arbitrary pole s = λ of the resolvent R(s, L) of order M the following statements are valid:
i) the pole λ is an eigenvalue of L, and λ∗ is an eigenvalue of L†. Moreover, the
eigenvectors corresponding to these eigenvalues λ and λ∗ have identical multiplicities; ii) in a sufficiently small neighborhood of λ, the main part of the Laurent series ex- pansion for the resolvent R(s, L) can be represented as follows
1 P (λ) − ν [ep,0, ǫp,0] [ep,0, ǫp,1] + [ep,1, ǫp,0] (s λ) Rν = + + . . . − (s λ)Mp+1 (s λ)Mp ν= M p=1 X− X − − (3.30)
[ep,0, ǫp,Mp ] + [ep,1, ǫp,Mp 1]+ + [ep,Mp , ǫp,0] + − ··· , s λ −
and hence
P (λ) Mp+1 ν −
R ν = [ep,µ, ǫp,Mp+1 ν µ], ν =1, , M, (3.31) − − − ··· p=1 µ=0 X X where ep,m(λ) are the elements from an arbitrary canonical system E(λ) corresponding to
the eigenvalue λ of the operator L, and ǫp,n(λ∗) are elements from the canonical system
E†(λ∗) corresponding to the eigenvalue λ∗ of the operator L†, which uniquely follows from E(λ). Proof. See Section A.1 in Appendix. This theorem states that each pole of the resolvent is an eigenvalue of L. Moreover, the expression (3.30) shows how the main part of the resolvent R(s, L) at a pole λ of order M can be determined by means of the corresponding canonical systems E(λ) and
E†(λ∗).
3.1.3 Sectorial operators Now we will introduce the subclass of those operators termed as sectorial operators [Hen81, LLMP05]. This class is suitable for the mathematical description of many en- gineering problems. An essential feature of these operators is that the corresponding initial-boundary-value problem admits an analytical representation of the solution by means of holomorphic semigroups. 26 3. Spectral Theory of Operators
3.1.3.1 Definition of sectorial operators
Definition 23 [Hen81] A linear operator L on a Hilbert space H is called a sectorial operator, if i) it is closed and has a dense domain D(L) in H; ii) for some ϕ (0,π/2) and a R the sector (see Figure 3.2) ∈ ∈ S = s C ϕ arg(s a) π, s = a (3.32) a,ϕ { ∈ | ≤| − | ≤ 6 } belongs to the resolvent set ρ(L) and iii) for some C > 0 the following estimate
1 (sI L)− C/ s a , s S (3.33) k − k ≤ | − | ∀ ∈ a,ϕ holds.
Im(s)