Generalized Inverse Operators and Fredholm Boundary-Value Problems 9783110944679

Table of contents :
Contents
NOTATION
PREFACE
1. PRELIMINARY INFORMATION
2. GENERALIZED INVERSE OPERATORS IN BANACH SPACES
3. PSEUDOINVERSE OPERATORS IN HILBERT SPACES
4. BOUNDARY-VALUE PROBLEMS FOR OPERATOR EQUATIONS
5. BOUNDARY-VALUE PROBLEMS FOR SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS
6. IMPULSIVE BOUNDARY-VALUE PROBLEMS FOR SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS
7. SOLUTIONS OF DIFFERENTIAL AND DIFFERENCE SYSTEMS BOUNDED ON THE ENTIRE REAL AXIS
References

Citation preview

Generalized Inverse Operators and Fredholm Boundary-Value Problems

GENERALIZED INVERSE OPERATORS AND FREDHOLM BOUNDARY-VALUE PROBLEMS

A.A. Boichuk and A.M. Samoilenko Translated by P.V. Malyshev and D.V. Malyshev

IIIVSPm Utrecht · Boston, 2004

VSP an imprint of Brill Academic Publishers P.O. Box 346 3700 AH Zeist The Netherlands

Tel: +31 30 692 5790 Fax: +31 30 693 2081 [email protected] www.brill.nl www.vsppub.com

© Koninklijke Brill NV 2004 First published in 2004 ISBN 90-6764-407-2

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior permission of the copyright owner.

A C.I.P. record for this book is available from the Library of Congress

Printed in The Netherlands by Ridderprint bv, Ridderkerk.

Contents

NOTATION

ix

PREFACE

xi

1. PRELIMINARY INFORMATION 1.1. 1.2. 1.3. 1.4. 1.5.

Metric and Normed Spaces Hilbert Spaces Banach Spaces Linear Operators Unilateral Inverse, Generalized Inverse, and Pseudo-Inverse Operators

1 1 6 10 14 25

2. GENERALIZED INVERSE OPERATORS IN BANACH SPACES 2.1. Finite-Dimensional Operators 2.2. An Analog of the Schmidt Lemma for Fredholm Operators . . . . 2.3. Generalized Inverse Operators for Bounded Linear Fredholm Operators 2.4. Generalized Inverse Matrices

29 29 35

3. PSEUDOINVERSE OPERATORS IN HILBERT SPACES 3.1. Orthoprojectors, Their Properties and Relation to Finite-Dimensional Operators 3.2. An Analog of the Schmidt Lemma for Fredholm Operators . . . . 3.3. Left and Right Pseudoinverse Operators for Bounded Linear Fredholm Operators 3.4. Pseudoinverse Operators for Bounded Linear Fredholm Operators 3.5. Inverse Operators for Fredholm Operators of Index Zero 3.6. A Criterion for Solvability and a Representation of Solutions of Fredholm Linear Operator Equations

47

ν

39 42

47 54 60 62 66 68

VI

Contents 3.7. Integral Fredholm Equations with Degenerate Kernels under Critical Conditions 3.8. Pseudoinverse Matrices

4. BOUNDARY-VALUE PROBLEMS FOR OPERATOR EQUATIONS 4.1. Linear Boundary-Value Problems for Fredholm Operator Equations 4.2. Generalized Green Operator 4.3. Examples 5. BOUNDARY-VALUE PROBLEMS FOR SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS 5.1. Linear Boundary-Value Problems. Criterion of Solvability . . . . 5.2. Weakly Nonlinear Boundary-Value Problems 5.3. Autonomous Boundary-Value Problems 5.4. General Scheme of Investigation of Boundary-Value Problems . . 5.5. Periodic Solutions of the Mathieu, Riccati, and Van der Pol Equations 5.6. Differential Systems with Delay 6. IMPULSIVE BOUNDARY-VALUE PROBLEMS FOR SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS 6.1. Linear Boundary-Value Problems. Criterion of Solvability . . . . 6.2. Generalized Green Operator for Semihomogeneous Boundary-Value Problem and Its Properties 6.3. Regularization of Linear Impulsive Boundary-Value Problems . . 6.4. Conditions for the Appearance of Solutions of Weakly Perturbed Linear Boundary-Value Problems 6.5. Weakly Nonlinear Boundary-Value Problems 6.6. Critical Case. Necessary Condition for the Existence of Solutions 6.7. Sufficient Condition for the Existence of Solutions. Iterative Algorithm for the Construction of Solutions 6.8. Critical Case of the Second Order

73 76

83 84 88 93

101 101 113 145 156 158 170

195 196 205 208 209 217 221 223 237

Contents 7. SOLUTIONS OF DIFFERENTIAL AND DIFFERENCE SYSTEMS BOUNDED ON THE ENTIRE REAL AXIS 7.1. Solutions of Linear Weakly Perturbed Systems Bounded on the Entire Real Axis 7.2. Nonlinear Systems 7.3. Solutions of Linear and Nonlinear Difference Equations Bounded on the Entire Real Axis REFERENCES

vii

257 257 277 286 305

NOTATION Β C[a, b]

Banach space of vector functions 2: [a, 6] —> Rn space of vector functions χ: [α, 6] —• Rn continuous on an interval [a, 6]

C1 [a, 6]

space of vector functions χ: [α, δ] —> Rn continuously differentiable on an interval [o, b} Banach space of vector functions ζ: [α,6] —> Rn absolutely continuous on an interval [a, &] with norm 11*11 D- = \ \Z\\L$ + ||«(ο)||λ» dimension of the image of an operator L dimension of the kernel of an operator L

Dp[a,b]

dim im L dim ker L (G-)(t)

generalized Green operator for a semihomogeneous boundary-value problem

G(t,r)

generalized Green matrix

Η im L (R(L)) ker L ( N ( L ) ) K(t, τ) I ΙΧ(·)

L~, L+ L^

Hilbert space of vector functions ζ: [α, b] —> Rn image of a linear operator L kernel (null space) of a linear operator L Cauchy matrix linear vector functional τη χ η constant matrix that is the result of the action of an Tridimensional linear vector functional I on the columns of an η χ η matrix X(t) operator generalized inverse or pseudoinverse to an operator L left (right) pseudoinverse of an operator L

l

L~

inverse of an operator L

Lz\

left (right) inverse of an operator L ix

χ

Notation L;[a,b}

VL •Pyl

PL =

PN(L)

Banach space of vector functions ζ : [α, b] —> Rn integ r a t e to pth power (1 < ρ < oo) equipped with norm

space of bounded linear operators L acting from a Banach space Bi into a Banach space B2 linear operator (projector) projecting a Banach space Β onto the null space of an operator L linear operator (projector) projecting a Banach space Β onto a subspace Y c Β isomorphic to the null space N(L*) of the operator adjoint to an operator L linear operator (orthoprojector) projecting a Hilbert space Η onto the null space of an operator L

Rn sign t öij Χ(α,6](ί)

Euclidean vector space of constant vectors sign function Kronecker symbol characteristic function of an interval [o, b}

PREFACE

The problems of development of constructive methods for the analysis of linear and weakly nonlinear boundary-value problems for a broad class of functional differential equations, including systems of ordinary differential and difference equations, systems of differential equations with delay, systems with pulse action, and integro-differential systems, traditionally occupy one of the central places in the qualitative theory of differential equations [5], [148], [51], [94], [118], [104]. This is explained, first of all, by the practical significance of the theory of boundary-value problems for various applications—theory of nonlinear oscillations [2], [7], [14], [15], [89], [97], [101], [139], [70], theory of stability of motion [50], [51], [98], [97], [108], control theory [158], [128], and numerous problems in radioengineering, mechanics, biology, etc. [121], [152], [81], [89], [103], As a specific feature of boundary-value problems, we can mention the fact that their linear part is, in most cases, an operator without inverse. This fact makes it impossible to use traditional methods based on the fixed-point principle for the investigation of boundary-value problems of this sort. The uninvertibility of the linear part of the operator is a consequence of the fact that the number m of boundary conditions does not coincide with the number η of unknown variables in the operator system. Problems of this kind for systems of functional differential equations are of Fredholm type (or with Fredholm linear parts). They include extremely complicated and insufficiently studied (both underdetermined and overdetermined) critical and noncritical boundary value-problems. The applicability of the well-known Schmidt lemma [148] to the investigation of boundary-value problems regarded as operator equations with bounded operators in the linear part with an aim to construct a generalized inverse operator resolving the original boundary-value problem is restricted by the requirement that the corresponding boundary-value problem must be of Fredholm type with index zero, i.e., that m — n. Therefore, a major part of the works dealing with problems of this sort were carried out under the assumption that these problems are of Fredholm type with index zero (Azbelev, Maksimov, and Rakhmatullina [8], Vejvoda [149], Wexler [153], xi

xii

Preface

Grebenikov, Lika, and Ryabov [65], [93], Malkin [101], Mitropol'skii and Martynyuk [107], Samoilenko, Perestyuk, and Ronto [139] and [140]). Moreover, a significant part of these results was, in fact, obtained under the assumption that the operator in the linear part of the original boundary-value problem has the inverse operator (noncritical case). We do not use this assumption. It is known (Atkinson [6], Vainberg and Trenogin [148], Pyt'yev [123], Turbin [147], Nashed [112]) that the classical Schmidt procedure [141] is applicable to the construction of generalized inverse operators only in the case of Fredholm operators of index zero. Thus, for boundary-value problems regarded as operator systems in abstract spaces [153], we suggest some methods for the construction of the generalized inverse (or pseudo-inverse) operators for the original linear Fredholm operators in Banach (or Hilbert) spaces. As a result of systematic application and development of the theory of generalized inverse operators [123], [112] and matrices [91], [147], [109], [117], new criteria of solvability were obtained and the structure of solutions was determined for linear Fredholm boundary-value problems for various classes of systems of functional differential operators. The methods used for the construction of the generalized Green's operators (and generalized matrices playing the role of kernels of their integral representations) for semihomogeneous boundary-value problems for systems of this sort are presented from the common viewpoint. We also study basic properties of the generalized Green's operator. In particular, it is shown how, using this operator, one can construct the generalized inverse of the operator of the original boundary-value problem. New efficient methods of perturbation theory were developed in analyzing weakly nonlinear boundary-value problems. These methods, including the Lyapunov-Poincare method of small parameter [97], [120], asymptotic methods of nonlinear mechanics developed by Krylov, Bogolyubov, Mitropol'skii, and Samoilenko [14], [15], [90], some methods proposed by Tikhonov [144], [145] and the Vishik-Lyusternik method [150], are extensively used for the solution of various problems encountered in different fields of science and engineering, such as radioengineering [121], [101], shipbuilding [89], celestial mechanics [65], [81], biology [152], [103], etc. These methods were developed and used in numerous works (Vainberg and Trenogin [148], Vejvoda [149], Grebenikov and Ryabov [65], Kato [80], Malkin [101], Mishchenko and Rozov [106], and Hayashi [74]). However, the application of the methods of perturbation theory to the analysis of weakly nonlinear boundary-value problems for various classes of differential systems was, for the most part, restricted to the case of ordinary periodic boundaryvalue problems in the theory of nonlinear oscillations (Grebenikov and Ryabov

Preface

xiii

[65], Hale [70], Malkin [101], Proskuryakov [122], and Yakubovich and Starzhinskii [154] for systems of ordinary differential equations, Mitropol'skii and Martynyuk [107] and Shimanov [143] for systems with delay, and Samoilenko and Perestyuk [139] and Bainov and Simeonov [10] for systems with pulse action). We show that the principal results in the theory of weakly nonlinear periodic oscillations remain valid (with necessary refinements, changes, and supplements) for general weakly perturbed (with Fredholm-type linear parts) boundary-value problems for systems of functional differential equations. The boundary-value problems are specified by linear or weakly nonlinear vector fiinctionals such that the number of their components does not coincide with the dimension of the operator system. In our monograph, we develop a general theory of boundaryvalue problems of this kind, give a natural classification of critical and noncritical cases,1 establish efficient conditions for the coefficients guaranteeing the existence of solutions, and develop iterative algorithms for the construction of solutions of these problems. Numerous results presented in the monograph were originally obtained and approved in analyzing boundary-value problems for systems of ordinary differential equations (Boichuk [19]). Later, it was discovered that the proposed procedures of investigation and algorithms are applicable to the analysis of much more general objects, including boundary-value problems for ordinary systems with lumped delay [32], [37], [157], systems with pulse action [30], [29], [135], autonomous differential systems [25], [26], [36], and operator equations in functional spaces whose linear part is a normally resolvable operator but they are not everywhere solvable [33], [34], [35]. In the first chapters, to make our presentation more general, we give some results from the theory of generalized inversion of bounded linear operators in abstract spaces, which are then used for the investigation of boundary-value problems for systems of functional differential equations. Some of these results are of independent interest for the theory of linear operators, although our main aim was just to develop the tools required for the analysis of boundary-value problems for systems of functional differential equations. The methods used for the construction of generalized inverse operators in Banach and Hilbert spaces are presented separately because these spaces are characterized by absolutely different geometries. The construction of the generalized inverse operator for a linear Fredholm operator acting in Banach spaces is based on the Atkinson theorem [6] obtained as a generalization of the Nikol'skii theorem [113], which states that any bounded Fredholm operator can be represented in the form of a unilaterally invertible and 'in the literature, these cases are sometimes called resonance and nonresonance.

xiv

Preface

completely continuous (finite-dimensional) operator. By using this fact, we arrive at a construction of the generalized inverse of a Fredholm operator similar to the well-known Schmidt procedure [148] applicable only in the case of generalized inversion of Fredholm operators of index zero in Banach spaces. The theory of generalized inversion and pseudoinversion of linear Fredholm operators in Banach and Hilbert spaces enabled us to develop a unified procedure for the investigation of Fredholm boundary-value problems for operator equations solvable either everywhere or not everywhere (Chapter 4). The proposed approach is then improved for the analysis of boundary-value problems for standard operator systems, including systems of ordinary differential equations and equations with delay (Chapter 5) and systems with pulse action (Chapter 6). We obtain necessary and sufficient conditions for the existence of solutions of linear and nonlinear differential and difference systems bounded on the entire axis (Chapter 7). This enables us to take into account specific features of each analyzed differential system and present necessary examples. The readers interested primarily in the theory of boundary-value problems for specific differential systems may focus their attention on the corresponding chapters and omit the chapters containing preliminary information. The authors do not even try to present the complete bibliography on the subject, which is quite extensive, and mention only the works required for the completeness of presentation. In conclusion, the authors wish to express their deep gratitude to all participants of numerous seminars and conferences on the theory of differential equations and nonlinear oscillations, where all principal results included in the book were reported and discussed.

1. PRELIMINARY INFORMATION

In Chapter 1, we present some well-known definitions and results from functional analysis, the theory of linear operators in Hilbert and Banach spaces, and matrix theory, required for our subsequent presentation. The readers who are not familiar with the theory of linear operators in function spaces can find here an elementary presentation of the facts essentially used in what follows. The other readers can use this material for references. The theorems are presented without proofs, but the reader is referred to the sources for further details. In this chapter, we also introduce necessary notation.

1.1. Metric and Normed Spaces Definition 1.1. A metric space is defined as a set X equipped with a metric p(·, ·), i.e., a real function defined in the set X and such that (1) P(x,y)>

0

(p{x,y)

= Oiffx

= y);

(2) p(x, y) — p(y, x); (3) p(x,z)

< p(x,y) + p(y,z)

(triangle inequality).

Thus, an arbitrary set equipped with a metric is a metric space. Example 1. A set X whose points are collections of n-dimensional real vectors χ = ( χ ι , . . . , xn) turns into a metric space if we set

1

2

Preliminary

Chapter 1

Information

The same set X can also be equipped with other metrics, e.g., pi(x,y)

= max ja:» - yf|, 1 Y is called a linear operator if the following axioms are satisfied: (1) L(x + y) = Lx + Ly for any χ and y in X ; (2) L(ax) — aLx for any χ € X and a € R or C. Definition 1.10. A linear operator L is called bounded if there exists a number C > 0 such that, for any χ € X, \\Lx\\ < C||x||. The smallest of these numbers is denoted by ||L|| and is called the norm of the operator L. Definition 1.11. A linear operator Ζ»: X —> R is called a linear functional and an operator I = [l\,... ,lm]: X —> Rm defined by the equality lx — col {l\x,..., lmx} for all χ e X is called a linear vector functional. Thus, a linear functional maps a linear space into a field of constant elements, and a vector functional maps a linear space into a field of vectors with constant components. Definition 1.12. The set of all linear bounded operators mapping a linear space X into a space Y equipped with properly introduced operations of addition and multiplication by constant numbers is a linear space called the space of linear operators and denoted by C(X, Y).

1.2. Hilbert Spaces Hilbert spaces are fairly simple infinite-dimensional normed spaces. Their relative simplicity is explained by the fact that they possess an additional structure created by the operation of inner product (a generalization of the ordinary inner product in vector algebras). The presence of an additional algebraic structure substantially enriches the geometric properties of Hilbert spaces. It is worth noting that the analyzed structure allows one to introduce the notion of perpendicular vectors, which makes the geometry of Hilbert spaces much closer to the geometry of Euclidean spaces.

Section 1.2

Hilbert Spaces

7

Definition 1.13. Α Hilbert space is defined as a set Η of elements with the following properties:

f,g,h,...

(1) Η is a linear space, i.e., the operations of addition and multiplication by real or complex numbers are defined in Η (depending on the field of constants, the space Η is either real or complex). (2) Η is equipped with the operation of inner product, i.e., with a real function (/> 9) of two arguments satisfying the following axioms: (a) (af, g) = a(f, g) for any number a; (b) {f + h,g) = {f,g) + {h,g)\ (c) (/, g) — (g, /), where the overbar denotes the operation of complex conjugation; (d) { f , f ) > 0 if f φ 0, and (/, / ) = 0 whenever f = 0. In what follows, we consider only real Hilbert spaces for which axiom (2c) takes the form (f,g) = (g,f). (3) Η is a complete space with respect to the metric p(f,g) = ||/—g||, where the norm of an arbitrary element h € Η is defined as ||/i|| = (h, h)1/2. Example 1. Let Rn be the n-dimensional Euclidean space. The inner prod-

uct of two elements χ — (x\,..., defined as follows:

xn) and y = (yi,...,

yn) of the space Rn is

η

Thus, Rn is a Hilbert space. Example 2. The space L2 (a special case of the space Lp with ρ = 2) is a Hilbert space if the inner product of two functions f: [a, 6] —> Rn and g: [a, b} —> Rn is defined as

Preliminary

8

Information

Chapter 1

The norm of an element in this space is given by the formula 6

\\f\\=vüj)=([£\fw\2d*y/2· β

η

ί = ι

Example 3. If we consider the space lp with ρ = 2, then the inner product can be defined as follows: oo

x

( >y) = Σ Χ i=l

ί ν ί

'

where x\ and yi, ϊ = 1 , 2 , . . . , are real numbers and the norm of an element is given by the formula oo

imi =

(5>I i=l

2

).

The space h obtained as a result can be regarded as the simplest example of an infinite-dimensional Hilbert space. Example 4. The space D2 is a Hilbert space if the inner product is defined as follows: {X,V)D2

= (X,Y)L2

+

(x(a),y(a))Rn.

We now introduce some geometric notions valid in all spaces with inner product. Definition 1.14. Two vectors χ and y in a space Η are called orthogonal if(x,y)

= 0.

A collection of vectors {a?;} in a space Η is called orthonormal if (x;, xt) = 1 for all i and (xj, Xj) = 0 whenever i φ j. Definition 1.15. Let Hi and H2 be Hilbert spaces. Then the set {χ -I- y}, where χ e H\ and y e #2, is α Hilbert space called the direct sum of the spaces H\ and Η2 and denoted by H\ Θ Η2. Let Μ be a closed subspace of a given Hilbert space H. Then Μ is a Hilbert space with natural inner product inherited from H. Let M 1 - be the set of vectors from Η orthogonal to M. The set M L is called the orthogonal complement

Section 1.2

Hilbert Spaces

9

of Μ. The linearity and continuity of the inner product imply that Μ 1 is a closed linear subspace of the space H, and, hence, M 1 - is a Hilbert space. The only common element of the subspaces Μ and M 1 is zero [129]. The theorem presented below states that, for any closed proper subspace, one can find vectors orthogonal to it. In fact, we have Η — Μ @ Ml. This geometric property is one of the most important factors due to which Hilbert spaces are much easier for analysis than Banach spaces (see Section 1.3). Lemma 1.1 [1]. Let Η be α Hilbert space, let Μ be its closed subspace, and let χ € Η. Then there exists a unique element ζ £ Μ nearest to x. Theorem 1.1 (theorem on projection) [1]. Let Η be α Hilbert space and let Μ be its closed subspace. Then any element χ Ε Η admits a unique representation of the form x = z + h,

zeM,

heMx.

Definition 1.16. Two Hilbert spaces H\ and H2 are called isomorphic if there exists a continuous linear operator U mapping from H\ onto H2 and such that (Ux,Uy)H2 = (x,J/)hi for all x,y Ε H\. The operator U is called unitary and the mapping realized by this operator is called an isomorphism. An isomorphism is called isometric if it preserves the norm. For all g from a real Hilbert space H, we define an operator g*: Η —• R by the requirement that the equality g*(f) — ( f , g ) hold for all f £ H. By virtue of the Schwarz inequality [129], we have ||g*|| < ||g]| (for the norms in £(H, R) and H, respectively). Therefore, g* is a continuous linear functional on H, i.e., it is an element of the space C(H, R) — H*. It remains to clarify whether an arbitrary element of the dual space H* can be determined in a similar way using an element of Η. The Riesz theorem gives a positive answer to this question: Theorem 1.2 [129]. Any element g* of the space H* dual to α Hilbert space Η is associated with a single element g G Η such that g*{f) = ( f , g ) for all f G Η and, moreover, ||ff*||i/· = IMIi/·

10

Chapter 1

Preliminary Information

Thus, the space dual to a Hilbert space coincides with this space to within an isomorphism.

1.3. Banach Spaces Various concepts of finite-dimensional analysis admit generalizations to infinite-dimensional normed vector spaces. At the same time, the required analysis cannot be developed in all spaces of this sort. Serious difficulties are encountered in analyzing the problem of the convergence of sequences, which is of crucial importance for analysis. Indeed, a sequence expected to be convergent may appear to be divergent. In this case, it is necessary to restrict the class of spaces under consideration. To do this, one can, e.g., impose the condition of the completeness on the norm. The assumption of completeness substantially simplifies abstract analysis and, at the same time, a broad class of normed vector spaces satisfies this assumption [130], Definition 1.17. A complete normed linear space Β is called a Banach space. In the metric spaces Rn, C[a,b], Cn[a,b] (η > 1), lp, and Lp the norm of an element χ is defined as its distance to the zero element, i.e., ||x|| = p(x, 0), and, hence, these are Banach spaces. In Hilbert spaces, the norm is introduced with the help of the operation of inner product. In Banach spaces, it satisfies a weaker condition, namely, every element / G Β is associated with a number ||/|| > 0 called its norm and such that 11/ + ffll < 11/11 + \\g\\ and

||α/|| = | α | | | / | | .

In general, this norm is not induced by the inner product, and, hence, Banach spaces do not possess numerous important properties of the Hilbert spaces. Example 1 [73]. Let us show that the geometry of Banach spaces may exhibit some unexpected features. Let Β = R2. The norm of an element / = ( / i , / 2 ) is defined by setting 11/11 = l/il + l/ 2 |. A square with vertices at the points (1,0), ( 0 , - 1 ) , (—1,0), and (0,1) lying on the Ox- and Oy-axes plays the role of a unit ball in this space. We draw a line passing through the origin of coordinates at an angle of 45° to the

Section 1.3

Banach Spaces

11

Ox-axis and consider this line as a subspace Bo of the space B. Any vector g G Bo has coordinates (51,51), i.e., 5 = (51,51). Therefore, if we consider the vector /(1,0) and write ||/ - g\\ = |1 — 511 + |gi|, then the norm of the difference between the vectors / and g attains its minimum equal to one for any 9i, 0 < 51 < 1. Thus, the minimum distance from the vector / to the subspace Bo is attained on an infinite set of vectors g e Bolt is clear that, in the plane R2 with ordinary Euclidean distance, the indicated minimum is attained on a single vector. In some normed spaces, for infinite-dimensional subspaces Bq this minimum may not be attained at all. In Section 1.1, we have introduced the notion of bounded linear operators and denoted the set of all operators of this sort from X into Y by C(X,Y). If Y is complete, then C(X, Y) is a Banach space. In the case where Y consists of real numbers y G R, the space £(X,R) is denoted by X* and is called the space dual to X. The elements of X* are bounded linear functional defined in the space X. The dual space is a complete normed space. For some Banach spaces, the dual space can be described explicitly. This description is given by theorems on the general form of linear bounded functional in these spaces. Example 2. Consider the space Rn. Any bounded linear functional / in the η space Rn has the form f(x) = ^ XiVi, where y\, • • • ,yn are given constants. i=1 Thus, we obtain an isomorphism between (Rn)* and Rn. Example 3. Consider a Hilbert space. According to Theorem 1.2, any bounded linear functional admits a unique representation in the form f(x) = (χ, u) with 11/11 = IMI. Example 4. For 1 < ρ < +oo, the space (lp)* is isomorphic to the space lq with 1/p + 1 / 5 = 1. The isomorphism is specified by the formula 00

t=l where x = (xi,...,xn,...)

€lp,

y= (yi,...,y„,...)

e Z„

fe(lp)*.

Chapter 1

Preliminary Information

12

Example 5. For 1 < ρ < +oo, the space (Lp)* is isomorphic to the space Lq with 1/p + 1/q = 1. The corresponding isomorphism is established by the formula b f(x) = J x(t)g(t)dt,

x(t) € Lp,

g(t) G Lq,

f e (Lp)*.

a

Example 6. The space (C[a, b])* is isomorphic to the space V[a, b] of leftcontinuous functions with bounded variation on [a, 6] satisfying the condition b g(a) = 0. The norm in this space is introduced as follows: ||p|| = Vg. The a

required isomorphism is established by the formula b f{x) = J

x(t)dg(t),

a

where g(t)eV[a,b],

x(t) e C[a,b],

and

fe(C[a,b])*.

In working with Banach spaces, it is often necessary to solve problems of the construction of linear functionals with given properties. As a rule, this program is realized in two steps: first, the functional is defined on a subspace of a Banach space where the required properties can easily be verified, and then a general theorem is proved according to which any functional of this sort can be extended to the entire space with preservation of the required properties. Theorem 1.3 (Hahn-Banach) [79]. Let Μ be a linear subspace of a Banach space Β and let f* be a continuous linear functional defined on M. Then f* can be extended to a continuous linear functional on Β with the same norm. Corollary 1.1. Let Μ be a linear subspace of Β and let g be an arbitrary element of Β lying outside M. Then there exists a continuous linear functional f* on Β such that (a) /*(/)

= 0 for f € M;

Section 1.3

Banach Spaces

(b)

f*(g)

(c)

||/*|| = distT 1 (g,M), g to the set M.

13

= 1;

where dist (g, M) is the distance from the point

Corollary 1.2. For any nonzero f G B, one can find f* G B* such that \\r\\ = l a n d r ( f ) = ||/||. Corollary 1.3. If /*(/)

= 0 for all f* G B, then / = 0.

Corollary 1.4. For any f G B, we have 11/11= sup | Γ ( / ) | . II/* 11=1 Definition 1.18. Let B\ and B2 be two subspaces of a space Β intersecting only at zero. The set of all vectors of the form x\ + 2:2, where x\ G B\ and X2 G B2, is called the direct (algebraic) sum of the subspaces B\ and B2 and is denoted as follows: Β = Βχ+Β2. Moreover, if B\ and B2 are closed subspaces of the space Β, then we say that the space Β decomposes into a direct (topological) sum of closed subspaces. In this case, we write B = B1@B2. Definition 1.19. If B\ C Β is a closed subspace and there exists a subspace B2 C Β such that Β = B\ φ i?2> then B\ is called complemented in B, and B2 is called a direct complement of B\ with respect to B. Clearly, a direct complement is not unique. The direct sum is a linear manifold, but, generally speaking, it is not a subspace (because it can be nonclosed). The following theorem gives necessary and sufficient conditions for the direct sum to be closed: Theorem 1.4 [63]. Let Β be a Banach space and let B\ and B2 be two subspaces of Β intersecting only at zero. In order that the direct sum B\ 0 B2 be a subspace, it is necessary and sufficient that there exist a constant k > 0 such that ||xx + x 2 || > k{ ||χι|| + ||χ2||), xi € Bu G B2.

14

Preliminary Information

Chapter 1

Corollary 1.5. If one of the subspaces B\ and By, is finite-dimensional, then the direct sum B\ φ #2 is closed, i.e., any finite-dimensional subspace of a Banach space can always be complemented. If Β — Η is a Hilbert space, then any subspace H\ c Η possesses a direct complement, e.g., the orthogonal complement of H i . For Banach spaces, one can introduce the following analog of orthogonality based on the duality of the spaces Β and B*: Definition 1.20 [88]. For any given sets Μ C Β and M* C B*, the sets M± = {f*eB*-.r(f)

= 0, / e M } ,

M*1 = {f £ B: / ( / * ) = 0, f* e M*} are called the orthogonal complements of the sets Μ and M*, respectively. In this case, the analogy with Hilbert spaces is not so close as desired because M 1 - is a subspace of the dual space B*, but not of the original space B. This is an additional manifestation of the complex geometric structure of Banach spaces. Generally speaking, direct complements cannot be constructed for some subspaces of a Banach space. However, as already indicated, finite-dimensional subspaces can always be complemented. Theorem 1.5 (on complements) [4]. 1. If B\ is an n-dimensional subspace of a Banach space B, then there exists a closed complement of B\, which can be specified by η linearly independent functionals. 2. If i?2 is a closed subspace of a Banach space Β defined by a finite set of η linearly independent functionals B2 = {x: fi{x) = 0, i = 1 , . . . , n}, then B% possesses an n-dimensional complement with respect to the space Β.

1.4. Linear Operators Let B\ and B2 be Banach spaces and let £.{B\, B2) be the Banach space of bounded linear operators acting from B\ into B2. Definition 1.21. The set D(L) C B\ where an operator L G C.(B\, B2) acts is called the domain of this operator.

Section 1.4

Linear Operators

15

Definition 1.22. The set of all solutions of the homogeneous equation Lx = 0 forms a subspace ker L C B\ called the kernel of the operator L. The kernel of the operator L is often denoted by N(L) and is called the null space of the operator L. Definition 1.23. The set of all values of an operator L € £(B\, B%) forms a linear manifold I m i C ß 2 called the image of the operator L. The image of the operator L is also denoted by R(L). In connection with these definitions, we want to make the following remark: In the definition of an operator L acting from B\ into B2, it is possible that D(L) is a subset of B\. However, in what follows, the notation L: B\ —> B2 always means that D(L) — B\. Example 1. We now illustrate the notions introduced above by using the space C[a,b] as an example. First, we not necessarily consider the operation of ordinary differentiation. Since continuous functions are differentiable, we first restrict ourselves to the case of smooth functions, say, from C°° [a, 6]. For χ € C°° [a, 6], we set ( L x ) ( t ) = x'(t). This relation specifies an operator L mapping the space C[a, 6] into itself with the domain C°°[a,b}. Thus, we can write L : C°°{a,b} -» C[a,b}. However, the operator of differentiation is also meaningful in a broader space, e.g., in C x [a, b], and we can define another operator L by setting ( L x ) ( t ) = x'(t) for χ e C 1 [a, 6]. Despite the fact that Lx = Lx for χ G D(L), the operators L and L are regarded as different because D(L) φ D(L). Note that the operators can be regarded as identical only in the case where both their images and domains are identical. The operator L is called a continuation (extension) of the operator L. Definition 1.24. Two operators L,L e £(.Bi,£2) are called equal if D(L) = D(L) and Lx = Lx for all χ e D(L). The operator L is called an extension (continuation) of the operator L and the operator L is called a restriction of L if D(L) D D(L) and Lx = Lx for all χ e D(L). This relationship between the operators is denoted as follows: L C L. Definition 1.25. The dimensionality of an operator L £ £(βχ, B2) is defined as the dimensionality of its image, i.e.,

16

Preliminary Information

Chapter 1

dimL = dim Im L. An operator is called finite-dimensional if dimL < oo. Definition 1.26. An operator L € C{B\,B2) is called bounded if there exists a constant C > 0 independent of χ and such that \\Lx\\ < C ||x|| for χ € D{L). The smallest of these constants is called the norm of the operator L and is denoted by ||L||: II Τ II

\\LX\\B2

ill = sup -rr-r . ießi FllBx

In the collection of all linear operators acting in Banach spaces, we can select a class of operators (generally speaking, unbounded) whose properties are quite close to the properties of bounded operators. This is the class of closed operators. Definition 1.27. A linear operator L is called closed if the facts that xn —• χ and Lxn —> y imply that χ £ D(L) and Lx = y. Any bounded operator defined in the entire space Β is closed, but a closed operator can be unbounded. If a closed operator L has a nontrivial null space N(L), then this space is closed. Theorem 1.6 [84]. A closed operator L is bounded if and only if its domain D(L) is closed. Thus, a closed operator L defined in the entire space B\ is bounded. Definition 1.28. An operator L e C{B\,B2) space B\ if its domain D(L) is dense in B\.

is called densely defined in the

An operator defined in the entire space B\ is densely defined. The operator L*: i?2 ~^> B*i adjoint to a densely defined operator L : B\ —> B2, is uniquely determined: {Lmg)(x)=g(Lx)

Vxefii,

Vje^.

(1.1)

Section 1.4

Linear Operators

17

In the case of Hilbert spaces, the adjoint operator L* : H^ —> H{ = H\ is defined in terms of the inner product as follows: (L*y,x)Hl

-

(1.2)

(y,Lx)H2.

The adjoint operator is always closed. Indeed, if gn —> g and L*gn —> / , then, for χ e D(L), we have gn(Lx) —» g(Lx) by virtue of the first relation, and gn(Lx) —> f(x) by virtue of the second relation, i.e., g(Lx) = f(x) for all χ € ( D ( L ) ) and / = L*g. The operator L* adjoint to a bounded linear operator is also a bounded linear operator. Definition 1.29. The kernel ker L* of the operator L* is called the cokernel of the operator L and is denoted by Coker L or N(L*) (the null space of the operator L*). If Bi and B^ are reflexive spaces, then L** — (L*)* = L, whence kerL = Coker L*. Example 2. Let Ηι = Rn , let Η-χ — Rm, and let a linear operator L: Hi —> H2 be specified by an τι χ m matrix A with entries al3 according to the formula η t=l The space dual to Rm is isomorphic to Rm, and the correspondence between these spaces is established by the formula m x

=

f( )= It follows from relation (1.2) that (L*f)(x)

m η = f(Lx) = Σ Σ ° ν j=l i=1

χ

& =

(t'Lx)·

The functional thus constructed can be represented in the form m

η

j=1 i=l

τιη j=1

Preliminary Information

18

Chapter 1

η where Tj = J2 a ^ . i=l

Thus, the action of the adjoint operator is determined by the mxn T A — (a.ji) of the matrix A — (aij).

transpose

Example 3. In the space Lp[0,1], we consider an integral operator ι (Lx)(t) = J K(t, s)x(s) ds, ο where K(t,s) is a bounded function measurable in the square [0,1] χ [0,1]. Any functional / G (Lp[0,1])* can be represented in the form 1 f(x)= J

x(t)g(t)dt,

where g £ ^ [ 0 , 1 ] (see Section 1.3). We fix g € £[0,1] and study, by using relation (1.1), the action of the operator L* on the corresponding functional: ι (L*g)(x)=g(Lx)

ι

= J [J ο

K(t,s)x(s)ds)g(t)dt.

ο

Changing the order of integration (by virtue of the Fubini theorem [84], this is possible), we obtain 1

1

(L*g)(x) = Jx(s)

1

JK(t,s)g(t)dt]ds

where

=

Jx{s)v(s)ds,

1 υ(β) = j

K(t,s)g(t)dt.

ο Thus, the adjoint operator L* transforms the functional corresponding to the function g into the functional corresponding to the function v, i.e., I (L*g)(s) = J

K(s,t)g(t)dt.

Section 1.4

Linear Operators

19

Thus, the operator adjoint to the integral operator with kernel K(t, s) is an integral operator with kernel K(s,t). Example 4. In the space £>2 [0,2] of absolutely continuous functions whose derivatives belong to the space 2], we consider the operator ι (Lx)(t) = x(t) — x ( l ) + x(0) =x(t)

- J x(s) ds,

t e [0,2],

ο acting into the space L2[0,2]. As in the previous example, we choose g e ^ [ 0 , 2 ] and consider the action of the operator L* on the corresponding functional: 2 (L*g)(x)=g(Lx)

= J(x(t) -

1

J x{s)ds}g{t)

0

dt

0

2

2

= JJX[o,i](s)i(s) 0 0 2

2

- J x{t)g(t)dt-

ds)g(t)dt

2

J ( J

0

0

Xm(s)x(s)ds)g(t)dt,

0

where X[o,i](i) is the characteristic function of the interval [0,1]. Changing the order of integration in the second integral, we get 2

(L*g)(x) = J ( g ( s ) -

2 X m

0

( s ) Jg(t)dt)x(s)

ds.

0

Thus, the adjoint operator has the form 2

(.L*g){t) = g(t)-X[0tl](t)

J g(s) ds. ο

Example 5. Consider an operator L: Lp[0,1] —> Lp{0,1], 1 < ρ < oo, of the form ( L x ) ( t ) = x(a(t)), where a: [0,1] —> [0,1] is a continuously

Chapter 1

Preliminary Information

20

differentiable invertible mapping of the interval [0,1] onto itself with a'(t) φ 0 [64]. The adjoint operator L* is constructed as follows: 1 g(Lx) = J

g(t)x(a(t))dt

r = a(t)

ät = ß>{r)dT^

=

J

9 { β { τ ) ) χ { τ ) β

,

{ τ ) ά τ

ο ι = J ß'(T)g(mT))x(T)dr

=

(L*g)(x),

ο where r = a(t),

/=/?(r),

and

dt =

ß'(r)dr.

Thus, the adjoint operator L* acts in the space Lq[0,1] formula (L*g){t) = ß'{t)g(ß{l)), where β = αΓ 1 .

according to the

In the set of closed operators, we distinguish the class of normally resolvable operators.

Definition 1.30. A densely defined operator L acting from B\ into Bi is called normally resolvable if its image is closed,, i.e., R(L) = R(L). Thus, an operator L : B\ > B2 for which R(L) — B2 is a simple example of a normally resolvable operator. We now consider the case of strict inclusion: R(L) C B2· Let us show that if L i s a normally resolvable operator and R(L) does not coincide with B2, then the null space N(L*) of the adjoint operator L* is nontrivial. Let / e B2 but / e i ? ( / / ) . Since, according to our assumption, R(L) is a closed linear manifold, we can use the Hahn-Banach theorem for the construction of a linear functional ψ G Β*λ such that d,

vL-.x kN{L*)

(2.7) _

ρ

VL'y

_

= ^ s ( y ) f s , p.

The remaining relations are proved by analogy.

2.2. An Analog of the Schmidt Lemma for Fredholm Operators As already indicated, the Schmidt lemma is not true for Fredholm operators. By using the Atkinson theorem [6] on the representation of Fredholm operators in the form of a sum of unilaterally invertible and finite-dimensional operators, we arrive at the following analog of the Schmidt lemma for Fredholm operators: Lemma 2.4. Let L : B\ —> B2 be a bounded linear Fredholm operator of any index. Then the operator L — L + Vl possesses the bounded inverse ({L + VL)Tx hr

le

fif°r

= \{L + VL)rl

right for

r > d.

Proof. Let r < d. It follows from Theorem 1.9 that the operator Lt 1 (left inverse to L) exists if and only if (a) ker L — {0} and (b) I m L is a subspace with direct complement in Bj) = 5kj, we obtain d

d

s,k=1 d

j=l d Y,ßik 1)6 kj()L = L

and L+LL+

= L;\IH2

- PL.)LL-\IH2

= 1-\IH2

- PL-) -

L+.

-

PL·)

62

Pseudoinverse Operators in Hilbert Spaces

Chapter 3

Finally, we get LL+ = LL;1(IH2

-PL.)

= I„2 - PL'

=

(LL+)

*

because = IH2 and = Pi-. Similarly, for r > d, the operator L+ = L~1(IH2 ~ PL*) is right pseudoinverse to L.

3.4. Pseudoinverse Operators for Bounded Linear Fredholm Operators For a bounded linear Fredholm operator L in a Banach space, the structure of the generalized inverse operator L~ specified by (2.14) does not determine the pseudoinverse operator in the case of Hilbert spaces. At the same time, the structure of the unilateral pseudoinverse operator L+r enables us to deduce a formula for the pseudoinverse operator. Theorem 3.5. The operator ' L+(IH2-Pl.)

L

+

for

r < d,

for

r > d

(3.20)

= K

(IHL

~ PL)L+

is the unique pseudoinverse operator for a bounded linear Fredholm operator L. Proof. Since L(I

H I

-P

L

) = L

and

(IH2 ~ PL')L

= L,

relations (i) and (ii) in (3.16) are true. It is necessary to check relations (iii) and (iv). Since L+ is the right pseudoinverse operator, we have LL+

= L(IHL

- PL)L+

= LL+ - IH2

- PL.. =

(LL+)*.

Further, L+L

= (IHL

- PL)LTL

= (IHL

- PL)L^{IH2

-

-

- PL)L;1L

= (IHL

(IHL

where VL is a projector to N(L). L-1LL~XL

PL.)L - PL){IHL

Indeed, since

= L~L{I

- PL.)L

=

L~LL,

- VL)

Section 3.4

Pseudoinverse Operators for Bounded Fredholm Operators

63

the operator LR1L is a projector to R(L*). Therefore, IHX - LR1L — VL is also a projector. By virtue of Lemma 3.2, we have PLVL = VL and, hence, (IH, - PL){JH\

- VL) = IHL - P L - V

L

+ PLVL

= IHL -

PL·

Therefore, L+L

= IHL -PL

=

(L+L)*

because I*HL = IHI and P£ = PL-

It is now necessary to check properties (iii) and (iv) for the second part of relation (3.20). We have L+L

= L+(/H2

- PL-)L

(L+L)*

= L+L = IHL - PL =

because L;+ is the left pseudoinverse operator. Further, LL+

= LL+(IH2

-P

L

' ) = L(IHI

- PL)LJ\IH2

= Llj\LH2

-

PL·)

— PL·) = ( I H 2 - VL*)(IH2

-

PL·),

where VL· is the projector of the operator L* to the null space N(L*). virtue of Lemma 3.2, we have VL*PL· =VL·, and, therefore, {IH2 - VL-){IH2

- PL·) — IH2 - VL·

- PL- + VL-PL·

= IH2 -

By

PL·,

whence LL+

= I„2 - PL·

=

(LL+)*.

This completes the proof. We can also deduce another formula for the operator pseudoinverse to a Fredholm operator in a Hilbert space. Since any Hilbert space coincides with its dual space to within an isomorphism [129], the following compositions of operators L: H\ —> H2 and L*: H2 —> H\ are meaningful: L*L:

HI - > # 1

and

LL* : H

2

^ H

2

.

Thus, we have introduced the operator S = L*L \ H\ —» H\. This operator is self-adjoint because S* = (L*L)*

= L*L** = L*L = S.

Finally, by using relations (3.4), we construct the orthoprojector Ps • H\ —> N{S).

64

Pseudoinverse Operators in Hilbert Spaces

Chapter 3

Lemma 3.8. The operators PL, Ps, and Ps* satisfy the relation PL = PS = PS

Proo/. As earlier, let {/i}[ = 1 be a basis in the null space N(L). Since Lfi = 0 for i = 1 , . . . , r and the operators L* and L are linear, we get SJi = VLfi

= 0,

i.e., N(L) C N(S). Let us show that the basis fc, i = 1 , . . . , r, of the kernel N(L) cannot be supplemented by adding a linearly independent element xo £ Hi such that xo G N(S), i.e., Sxο = 0. Assume the contrary, i.e., let there exist an element xo € H\ such that Lx ο = φ φ 0 and Sx ο = L*ijj = 0. In this case, we arrive at a contradiction: 0 = (Lx 0 ,0) = (x0,L*rJ>) = (Lx0,VO = {Φ,Φ) Φ 0Hence, the bases of the null spaces N(L) and N(S) coincide, and, therefore, PL = Ps- Since the operator S is self-adjoint, we have N(S) — N(S*) and Ps = Ps* • Theorem 3.6. The operator L+ = (L*L + Pl)~xL*

= L*(LL* + PL.)~X

(3.21)

is the bounded pseudoinverse operator for a bounded linear Fredholm operator L. Proof. We prove the relation L+ = (L*L + PL)~lL*. The relation L+ = L*(LL* + PL*) - 1 is proved by analogy. The operator 5 is a Fredholm operator of index zero. Hence, the Schmidt lemma [148] is true for this operator, and the operator S + Ps possesses the bounded inverse. The operator S+ - (S + Ps)-l-Ps is the generalized inverse of S and satisfies the relation S+S = IHX — Ps- Moreover, since Ps is an orthoprojector, the operator S+ is not only the generalized inverse but also the unique pseudoinverse operator satisfying (1.8). Indeed, we have (i) 5 5 + 5 = S(I -Ps) (ii) S+SS+

= S - SPs = S,

= {I - Ps)S+

= S+ - PsS+

because S+Ps = 0 and Ps(S + Ps)-1

= 5+

= Ps, and, moreover,

(3.22)

Section 3.4

Pseudoinverse Operators for Bounded Fredholm Operators

(iii) (5+5)* = (IHl - Ps)* = IHl -Ps (iv) {SS+Y = (IHl-Ps-)m

65

= S+S,

= lHl-Ps-=SS+

(3.23)

because PS = PS·, I*HL = IΗ ι, and PS = P5. By using relations (3.22) and (3.23), we can show that the operator L+ = S+L* satisfies relations (3.16) and, hence, is the pseudoinverse operator for the operator L. Since PL = PS, LPL = 0, S+S = IHL - PS, and PSS+ = 0, we can write (i) LL+L = LS+L*L

= L(IHl

(ii) L+LL+ - S+L*LL+

- Ps) = L(IHl

- PL) = L,

= (IHl - PS)L+ = L+-

PsS+L*

= L+.

Further, since [129] L" = L,

((S + Ps)'1)*

= (S +

Ps)'1

and, hence, (S+)* = ((S + Ps)-1 - PsY = ((s + P s ) - 1 ) * - P § = s+, we conclude that (iii) (L+L)* = (S + L*L)* = (S+S)* = S+S = S+L*L = L+L, (iv) (LL+Y = (LS+L*)* = L**{LS+)* = L(S+)*L*LS+L*

= LL+.

Finally, since PLL* = (LPL)* = 0 and PS = Pl = PI, we can write L+ - S+L* = [(5 + Ps)'1

- PS]L* = (L*L +

Pl^L*.

Remark 3.1. In the case of a singular square matrix L (a simple example of a Fredholm operator of index zero), relation (3.21) for the pseudoinverse matrix L+ was apparently deduced for the first time in [147]. In [19], this relation was generalized to the case of rectangular matrices. Remark 3.2. Relations (3.21) enable one to compute the operators pseudoinverse to n- and ci-normal operators in Hilbert spaces [88].

66

Pseudoinverse Operators in Hilbert Spaces

Chapter 3

Note that if an operator L is such that ker L — 0 or ker L* = 0, then relation (3.21) coincides with the formulas L+ = (L*L)~XL*

(PL = 0),

L+ = L*(LL*)~1

(PL.=0)

presented in [112], Remark 3.3. Relations (3.21) can be used for the determination of the pseudoinverse operator L+ for a closed densely defined operator L.

3.5. Inverse Operators for Fredholm Operators of Index Zero Relations (3.20) and (3.21) specifying the operators pseudoinverse to Fredholm operators remain valid for Fredholm operators of index zero. To construct the operator pseudoinverse to a Fredholm operator of index zero in a Hilbert space, one can use an approach based on the Schmidt lemma. Thus, let L be a Fredholm operator of index zero (ind L — 0, r = d = p) and let {/i}£ =1 and be bases in the spaces ker L and kerL*, respectively. Using relations (3.4) and (3.6), we construct the orthoprojectors PL'· HI —• N(L) and PL· : —> N(L*) and finite-dimensional operators PL - H\ —> N{L*) and PL.: H2 N{L). Lemma 3.9. The operators PL, PL*, PL, and PL· satisfy the relations (i) PL-PL = PLPL = PL; (ii)

PLPL-

= PL*PL> = PL-;

(3.24) (Hi) (iv)

PLPL'=PL'\ PL.PL

= PL.

Proof. Lemma 3.9 follows from Lemma 3.3. Relations (iii) and (iv) in (3.24) follow from relations (iii) and (iv) in (3.7) because, for r = d = p, the expres-

Section 3.5 sions

Inverse Operators for Fredholm Operators of Index Zero

ρ

67

ρ

Σ Ak^&icVWs s,k=1

and

are, in fact, the orthoprojectors Pl* Lemma 3.10. 77ie operator

Σ oij^ifj'^fi i,j=1

and P l , respectively.

L + Pl possesses

the bounded inverse

Proof. It is necessary and sufficient to show that ker (L + Pl) ker (L + Pl)* = { 0 } . Assume that there exists xo φ 0, xo e H\, such that

operator.

= { 0 } and

(L + PL)x ο = 0. This means that

Γ Lxo = - Σ i,j=1

ij (fj>xo) {L + Pl)*VO = 0

suc

^

68 or

Pseudoinverse Operators in Hilbert Spaces

τ L*yo = - Σ ij=1

Chapter 3

"ij^iViiyoifj·

Multiplying both sides of the last equality scalarly by an element fa, k — 1 , . . . ,r, we get Γ

0 = (L*y0,fk)

= (yo,Lfk)

- £

c^W^X/,·,/*)

i.J=l

r i,i=\

Since the basis vectors ψ{, i = 1,... ,r, are linearly independent, the equality ίυο,ΨΊ) = 0 holds only for y0 = 0. Thus, ker (L + PL)* = {0}. It follows from the proof that the operator (L + PL) establishes a one-to-one correspondence between the spaces H\ and H2· By virtue of the Banach inverse-operator theorem [146], the operator (L + PL)"1 exists and is bounded. Theorem 3.7. The operator L+ = (L + PL)-1 - PL-

(3.25)

is the bounded pseudoinverse operator for a bounded Fredholm operator L of index zero. Proof. Theorem 3.7 follows from Theorem 3.4. Indeed, since r = d = p, the operator (L + PL) possesses both left (L+PL)IL and right (L + PL)~1 inverse operators. This means that the inverse operator (L + PL)~1 exists. Therefore, by analogy with the proof of Theorem 3.4, we obtain the required assertion of Theorem 3.7.

3.6. A Criterion for Solvability and a Representation of Solutions of Fredholm Linear Operator Equations The relations deduced for the operators pseudoinverse to bounded linear Fredholm operators in Hilbert spaces enable us to obtain, as in the case of Banach spaces, explicit expressions for the solutions of linear operator equations Lx = y.

(3.26)

Section 3.6

A Criterion for Solvability of Fredholm Operator Equations

69

Unlike the case of Banach spaces, the expression χ = L+y gives a unique special solution of equation (3.26), provided that this solution exists. The relation PL*V = 0,

(3.27)

where Ρχ,» is the orthoprojector to the kernel ker L* of the operator L*, guarantees the existence of a solution of equation (3.26). The general solution of the homogeneous equation Lx = 0 corresponding to (3.26) can be represented in the form X = [fl,...,fr]Cr,

(3.28)

where {/i}[ = 1 are basis vectors of the kernel kerL. The expression (3.28) for the general solution of the homogeneous equation can be rewritten in the following equivalent form: Χ — PLX

Indeed, substituting the last equality in the equation Lx = 0, we arrive at the identity r

r

LPLX =

= Σ

i,j=1

^ H F I ^ W I = O-

i,j=l

Theorem 3.8. The Fredholm operator equation (3.26) is solvable if and only if y G i/2 satisfies condition (3.27). In this case, equation (3.26) possesses an r-parameter (r = dim kerL) family of solutions, which can be represented in the form of a direct orthogonal sum as follows: x = PLx®L+y=[f1,...,

fT]cr ® L+y.

(3.29)

The first term of this sum is the general solution of the corresponding homogeneous equation and the second term is the unique solution of the operator equation (3.26) orthogonal to any solution of the homogeneous equation. The assertion of the theorem follows from the equality Lx

=

L{PlX

+

L+y) =

LPlX

+ LL+y

= LPlX

provided that condition (3.27) is satisfied.

+ (IH2 - PL.)y

= IN2y = y,

70

Chapter 3

Pseudoinverse Operators in Hilbert Spaces

For Pl'D φ 0, the problem is ill posed [145]. In this case, y does not belong to the image R(L) of the operator L, and the operator equation (3.26) is unsolvable. However, this equation has a 50-called pseudosolution [145], which minimizes the residual \\Lx — y\\u2· In the set of pseudosolutions specified by (3.29), one can find a unique pseudosolution x+ G Ηγ c H\ orthogonal to the kernel ker L of the operator L: = L+y.

(3.30)

This pseudosolution can be regarded as the best possible approximate solution of equation (3.26) minimizing the norm of the residual IILx+ - y\\Ha = min \\Lx x€H l

y\\h2.

The vector x+ has the smallest norm ||®+||ffi = min ||x|| x€tf+

in the set of all vectors for which the norm of the residual attains its minimum. Furthermore, the norm of the residual is equal to the norm of the expression on the left-hand side of the criterion (3.27) for the solvability of equation (3.26): ||Lx+ - y\\H2 = \\LL+y - y\\Ha = ||(J - PL.)y - y\\H2 =

\\PL.y\\Ha.

Thus, in the case where relation (3.27) is true, the norm of the residual is equal to zero, and pseudosolution (3.30) turns into a solution of equation (3.26). Example 1. We now establish a condition for the solvability of the following operator equation [92]: 2π

(Lx){t) - x(t) + Ax(t) -

where x(t) = col [ ^ ( t ) , x2(t)}, A =

0 1

J Bx(s)ds

= f(t),

f(t) = col [ / i ( t ) , f 2 ( t ) } , -1 0

and

Β =

0 1

0 0

(3.31)

Section 3.6

A Criterion for Solvability of Fredholm Operator Equations

71

The operator L: D2 [0,2π] —> L2[0,2π] is a Fredholm operator. The homogeneous equation (Lx) (t) = 0 has three linearly independent solutions xi (t) =

cost sini ®2(t) = - sint > ®s(i(t)y(t)dt i=l a

= g(s),

(3.34)

where L: L2 [a, b] —• L2[a,b] and L*: Ζ/2 [α-, b] —> L,2[a,b}. To find the bases of the kernels of the operators L and L*, it is necessary to solve the homogeneous equations (Lx)(t)

= 0

and

(L*y)(s)

= 0.

(3.35)

74

Chapter 3

Pseudoinverse Operators in Hilbert Spaces

Solutions of these equations are sought in the form η

η

x(t) = X > W «

2/(s)

^

=

(3.36) i=l

i=1

Substituting solutions (3.36) in equations (3.35), we arrive at the following algebraic systems: = 0

(I-A)c where A =

and

(I-A

T

(3.37)

)d=0,

is an η χ η matrix with constant entries, c =

col [ c i , . . . , Cn], and d = col [ d j , . . . , dn]. The case where det(I — A) = det(7 — Ατ) ψ 0 (noncritical) was studied in [38]. In this case, systems (3.37) and, hence, the homogeneous equations (3.35) have only the trivial solutions. By virtue of the Fredholm alternative, the inhomogeneous equations (3.33) and (3.34) are solvable for any f(t) and g(s), and their solutions can be represented in explicit form. We now consider the critical case where det(7 - A) = 0

and

rank (I - A) = rank (I - AT) =

nx.

In this case, the solution of system (3.37) admits the representation c = Fcr,

d = F\dr,

cr,dr£Rr,

r — n — ni,

where F and F\ are η χ r matrices formed of basis vectors of the null spaces N(I-A) and N{I-AT)·. ^=[/(1),...,/(r)]

and

Fi^/W,..·,/?0].

As a result, solutions (3.36) take the form η

χ(ϊ)

η

= Σ,Μί)*,

y(s)

i=l

=

i=l

X^iOO^i,

where Mt) = [Mt),...Mt)}f{i)

and

=

VhWl/f1.

Thus, the complete systems of functions {V>i(i)}i=i and { < ä } [ = 1 form the bases of ker L and ker L*, respectively.

Section 3.7

Integral Fredholm Equations with Degenerate Kernels

75

Since L is a Fredholm operator of index zero, its pseudoinverse operator L+ is given by relation (3.25). We first find the orthoprojector PL* and the operators PL and PI·. We have 6

r {PL-y){t)

ß\jl) ( * ) I ^ ( r ) / ( r ) d r = 0. i=1 α In this case, the solution of equation (3.33) has the form

x(t) =

+

where the operator L + is defined earlier.

3.8. Pseudoinverse Matrices In the present section, to illustrate the procedures of the construction of pseudoinverse operators proposed in the previous sections, we give some auxiliary facts from linear algebra and matrix theory frequently used in what follows. Thus,

Section 3.8

Pseudoinverse Matrices

11

we consider the problem of the construction of pseudoinverse matrices for rectangular matrices, establish conditions for the solvability of linear algebraic systems in the general case where the number of unknowns is, generally speaking, not equal to the number of equations, and study representations of solutions of these systems. In what follows, we restrict ourselves to the case of real domains, and, thus, the conjugate matrix always coincides with the transpose. Let Q be an m χ η matrix with constant entries and let rank Q = n\ < min(n, m). Identifying matrices with operators in a fixed space Rn, we obtain Rn = R(Q*) Θ N(Q)

and

Rm = R(Q) Θ N(Q*).

The η χ η and τη χ τη matrices playing the role of orthoprojectors projecting the spaces Rn and Rm to the null spaces N(Q) and N(Q*) of the matrices Q and Q* are denoted by PQ and PQ- , respectively: PQ:Rn^N(Q),

N(Q) = PQRu

Pq.: Rm -> N(Q*),

N(Q*) =

Since the dimension of the null space N(Q) matrix Q and r a n k Q — ηχ, we have [151]

PQ.Rm.

is equal to the defect of the

dim N(Q) = n — rank Q — n — ni—r. In view of the fact that rank Q = rank Q*, for the dimension of the null space N{Q*) we get dim N(Q*) = m — r a n k Q = m — n\ = d. Therefore, rank PQ = r and rank PQ. = d. This means that the matrix PQ consists of r linearly independent columns, and the matrix PQ* consists of d linearly independent rows. To construct the matrices (orthoprojectors) PQ and PQ·, we can use the following procedure [147]: Denote the bases of the null spaces N(Q) and N(Q*) by {fi}ri=l and {v«}s=i! respectively. These vectors are used for the construction of nonsingular (τι χ τι)- and (τη χ m)-dimensional Gram matrices: " = - K · } = {(/», Λ ) }

and

β = {/?sfc} = {{φ.,

78

Chapter 3

Pseudoinverse Operators in Hilbert Spaces

where (·, ·) are the inner products in the corresponding Euclidean spaces. The orthoprojectors PQ and PQ. are given by the formulas P Q = E

^FIFL*

ij=1

PQ' = Σ Ä

1

s,k=1

(3.39)

W,

where a·" 1 ^ and β ^ 1 ^ are the entries of the matrices inverse to the symmetric Gram matrices a and β, respectively. The Moore-Penrose pseudoinverse nxm matrix Q+ is given (according to (3.21)) by the formula Q+ = (QTQ + PQ)-1Qt

= QT{QQT

+

PQ-Γ1·

In this case, we have PQ = In-Q+Q

PQ'=Im-QQ+.

and

(3.40)

We now establish a solvability criterion and construct solutions for a linear algebraic equation Qc = b, (3.41) where Q is a known m χ η matrix of rank n\, r a n k Q = n i < min(m, n), b is a known column vector from the space Rm, and c is an unknown column vector (from the space Rn). Since any rectangular matrix is a simple Fredholm operator, by using Theorem 3.9 we conclude that the algebraic system (3.41) is solvable if and only if its right-hand side b belongs to the orthogonal complement N1- (Q) = R(Q) of the subspace N(Q*), i.e., whenever PQ'b = 0.

(3.42)

In this case, the general solution of system (3.41) takes the form c = Q+b + c, where c is a vector from the null space N(Q), C = PqC

=

namely

PQC&N{Q).

Section 3.8

Pseudoinverse Matrices

79

Since rank Pq- = d, condition (3.42) contains d linearly independent relations, and the m χ m matrix Pq- in the condition analyzed can be replaced by the dxm matrix Pq· composed of a complete system of d linearly independent rows of the matrix Pq* . The column vector c = Pqc G N(Q) can be rewritten as c = Pqj.ct, where Pg r is the η χ r matrix composed of a complete system of r linearly independent columns of the matrix Pq. Thus, we have proved the following theorem: Theorem 3.9. If rank Q = ni < min(m,n), then the algebraic system (3.41) is solvable if and only if the column vector b G Rm satisfies the condition PQ*b = 0, d = m-nx. (3.43) In this case, the system has an r-parameter (r = η — n\) family of solutions of the form c = PQrCr + Q+b Vcv € Rr. (3.44) Corollary 3.2. If rankQ — n\ = n, then system (3.41) is solvable if and only if the column vector b G Rm satisfies the condition PQ.b = 0,

d — m — n.

In this case, the system has a unique solution of the form c = Q+b. Indeed, since r = 0, we have Pqt = 0 and Pqtct — 0. Corollary 3.3. If rankQ — ri\ = n, then system (3.41) is solvable for any b G Rm and its solution has the form c = PQrCr + Q+b,

r =

n-m.

Indeed, since rank Q = m, we have d = m — m = 0 and Pq» ξ 0, and, hence, condition (3.42) is always satisfied. If condition (3.43) is not satisfied, then we get a simple ill-posed problem [60], [145]. In this case, the algebraic system (3.41) is inconsistent and unsolvable. However, there exists a so-called pseudosolution that minimizes the norm of the residual \\Qc - 6|| in the space Rn.

Chapter 3

Pseudoinverse Operators in Hilbert Spaces

80 The set R+

of pseudosolutions is specified by relation (3.44). In this set, one

can find a single normal pseudosolution c+ e R+ C Rn space

orthogonal to the null

N(Q): c+ - Q+b.

(3.45)

This solution is an approximate solution of system (3.41) (which can be obtained by the least-squares method) that minimizes the norm of the residual, i.e., \\Qc+-b\\ =

mm\\Qc-b\\,

c£Rn

and has the smallest length in the set of all vectors for which this minimum is attained: I M

= min ||C||, c€R+

||c|| = ( £ i=1

|«| a ) 1 / 2 .

Note that the norm of the residual is equal to the norm of the expression on the left-hand side of the solvability criterion (3.42) for system (3.41): \\Qc+-b\\ = \\QQ+b-b\\ = \\PQ.b\\. Thus, in the case where condition (3.42) is satisfied, the residual is equal to zero, and pseudosolution (3.45) turns into a solution of system (3.41). Example

1. To illustrate the proposed algorithm for the investigation of linear

algebraic systems, we consider the problem of the existence and construction of solutions of the matrix equation

Qc = b,

Q =

Ί 0 1

"0"

0' 1 , 1

Cl

0

b =

>

c

=

1

,C2.

(3.46)

ceR2

First, we construct the matrices 2 -1 1 PQ = 0,

PQ. =

-

1 - 1

1

1

-1

-1

-1

1

PQI = g[l

1 -1]·

Further, we check the solvability criterion (3.43) for system (3.46): pQd*b = ±( 1 , 1 , —i)(o, 0,

=

Section 3.8

Pseudoinverse Matrices

Thus, equation (3.46) is unsolvable, but we can find its pseudosolution 2

- 1

1

-1

2

1

which minimizes the norm of the residual.

4. BOUNDARY-VALUE PROBLEMS FOR OPERATOR EQUATIONS The constructions of generalized inverse operators L~ in Banach spaces and pseudoinverse operators L+ in Hilbert spaces suggested in the previous chapters enable us to propose two approaches to the analysis of linear Fredholm boundaryvalue problems

(Lz)(t) = B2 is a bounded linear Fredholm operator (indL = dim ker L — dim ker L* = s — k < 00). Theorem 2.3 shows that the Fredholm operator equation (4.3) is solvable if and only if its right-hand tp(t) G B2 satisfies the condition V>L-YQ is the projector specified by relation (2.4). In this case, the general solution of equation (4.3) takes the form z(t) = X(t)c+(L~ N(Q*) be the m χ m matrix representing (2.22). The s χ m matrix inverse to Q in the generalized sense is denoted by Q~. By using the algebraic equation Qc = X — l{L~ip),

(4.9)

we find a constant c £ R" for which solution (4.8) of equation (4.3) [existing if condition (4.7) is satisfied] is a solution of the boundary-value problem (4.3), (4.4). It follows from Theorem 2.3 that equation (4.9) is solvable if and only if the following condition is satisfied: VQ*d{\ - l(L~)(·)) = 0

(d = m — rankQ).

In this case, the indicated equation possesses an r-parameter (r = s — rank Q) family of solutions of the form (3.43): c = VQrCr + Q-{\-l{L- 0 linearly independent solutions. If rank D = τι ι = m, then d\ — 0, PDtd = 0, and condition (4.29) is satisfied for all f(t). l This is a characteristic feature of Fredholm problems. We now consider the problem of the solvability of the boundary-value problem (4.25), (4.26) and study the structure of the set of its solutions. Assume that the solvability condition (4.29) is satisfied. Substituting solution (4.31) in the boundary conditions (4.26), we obtain Ix = l(F(·)) + 1(Μ·)Ρϋη

Cr, = a.

(4.33)

This yields the following algebraic system for the vector cri: Qcri = α — l(F(·)),

(4.34)

where Q = 1(^ο(·))Ρϋη is a ρ χ ri matrix. Using Theorem 4.5 together with a criterion for the solvability of system (4.34), we arrive at a condition necessary and sufficient for the boundary-value problem (4.25), (4.26) to be solvable. This condition improves the corresponding results established in Theorem 4.1.

98

Boundary-Value Problems for Operator Equations

Chapter 4

Theorem 4.6. Let rank Q = n2 < min(p, 7*1). The homogeneous problem (4.25), (4.26) ( / = 0, α = 0) possesses exactly Γ2 (Γ2 = r\ - «2) linearly independent solutions of the form X = * 0 (t)PD r i P Q r 2 Cr 2 ,

Cr2eRr2·

(4.35)

The inhomogeneous problem is solvable if and only if PDt d b = 0,

Pjyt (a - l(F(·))) = 0,

l

2

di = m — rank D,

(4.36)

d e ) > e ) - i J Κ(·,τ)ΙΙ(χ(τ,ε),τ,ε)

dr],

α e(G[/ 0 (r,cS) + Αι{τ)χ(τ,ε)

+

+ £X(t)Q+[Jo(zo(·,

R(x,T,e)\)(t) eg)) + hx(;

F ( 0 , t , 0 ) = ö,

ε) + ϋι(χ(·,ε),

ε)]},

=

and I n is the η χ η identity matrix. The upper-triangular block-matrix operator L W has zero blocks at the principal diagonal and below. Therefore, the operator system (5.53) admits the following transformation: y(t,e)

= LFy(t,e),

L = (Is -

s = 2n + r.

The system obtained as a result can be solved by the method of simple iterations. Moreover, by the method of Lyapunov majorants, one can show [19] that the iterative process converges in a sufficiently small neighborhood of the generating solution and establish required estimates for the error of approximation. In addition, the operator S = LF

of the analyzed system is a contraction operator in

Section 5.2

Weakly Nonlinear Boundary-Value Problems

125

a sufficiently small neighborhood of a point zo(t,c*), ε = 0, and, hence, one of the versions of the fixed-point principle [87] is applicable to this system for sufficiently small ε £ [0, ε»]. Thus, the method of simple iterations enables us to construct the solution of the boundary-value problem (5.48) as a result of the iterative process described below. Iterative Algorithm. x(t, ·) € C[e],

The iterative process used for finding the solution

x(t, 0) = 0, of the boundary-value problem (5.48) is based on

the operator system (5.52). As the first approximation χ ^ ί , ε) to the solution x^\t,e),

we take /ο (τ,eg) ε) = εΛ"

(t)

= e(G/ 0 (r, cg))(t) + eX{t)Q+J0{z0{-,

+ e G i ( t ) P B o c ( 1 ) + X{2)(T, ε)) + +

£

X ( t ) Q

+ ε01(·)ΡΒ

+

{ j

0

0

R(X(T,

( z o ( ; Cg)) + h[XT(-)(Ir

C ^ + X^(.,£)}

+

ε), Τ,

-

ε)])(ί) PBo)c

R1(X(.,£),£)}.

Section 5.2

Weakly Nonlinear Boundary-Value

131

Problems

In view of representation (5.61) and the condition of solvability (5.59) of the second equation in (5.50), we arrive at the following algebraic system: ε Β ^ 1 ) + PB'0P N(B\) and let P g · be a d χ d matrix orthoprojector RD —> N(B{). Then, in order that equation (5.62) be solvable, it is necessary and sufficient that PB'lPB'0PQ'd{hx{2)(;s)

+

Ri(x(;E),e)

b -I J Κ(·,τ)[Α1(τ)χ(2)(τ,ε)

+ Ρ(χ(τ,ε),τ,ε)]άτ}

= 0.

(5.63)

a

If the null spaces N(BQ) and N(B{) have empty intersection, i.e., PB'PBI — 0, then equality (5.63) always holds and system (5.62) is solvable with respect to sc^ e N(B\) to within an arbitrary vector constant Pb0PbiC (Vc G RR): ecU = - B+PB'Pcrd{hxP\;e)

+

Ri(x(-,e),e)

b

-I j

ίΤ(·,τ)[Αι(τ)χ( 2 )(τ,ε) + ϋ(χ(τ,ε),τ,ε)] (·,^)] +

Rl{*kM,e)

b -I I ^ . . r J I A i W ^ M ^

+ ifir.e))

a + β(^(τ,ε),τ,ε)]ίίτ},

(5.71)

/o(r, c5) + Αι(τ)(ΧΓ{τ)(Ιτ

-

Pb0)cP

+ £G1(T)PBoC^1+X^(T,S)) +

R(xk(T,e),r,e)

® g 1 ( i > e ) = eA"

(t), J0(z0(.,ct))

+

h(Xr(.)(Ir-PBo)c{V

+ ε01{·)ΡΒο$1ι \L xk+1(t,e)

+

χ«\.,ε))

+Ri(xk{;e)e)

\)

= Xr(t)(/r - Pßo)40) + εσιίΟΡΒοίΛ + 4 + ι ( Μ ) · A: = 0 , 1 , 2 , . . . ,

x0(t, ε) = x(02) (t, ε) = 0.

As above, to show that the iterative process converges, one can use the method of majorizing Lyapunov equations described, e.g., in [19], [65], or [77]. Thus, we have proved the following assertion:

Theorem 5.6. Assume that the boundary-value problem (5.43) satisfies the conditions imposed above for the critical case (rank Q = n\ N(Q*), and PQ r is an η χ r matrix composed of a complete system of r linearly independent columns of the orthoprojector PQ : Rr —• N(Q). Necessary Condition. Both the autonomous periodic problem [101] and the weakly nonlinear boundary-value problem (5.85), (5.86) are quite different from similar nonautonomous boundary-value problems. Unlike the nonautonomous problems, the right endpoint of the interval where the solution of problem (5.85), (5.86) is sought is unknown and should be found in the process of construction of the required solution. We seek this endpoint in the form b(e) = b* + £(b*-a)ß(e),

β(0)=β*.

(5.91)

By using this relation and the change of variables t = a + (r-a)(

1+εβ(ε))

(5.92)

in problem (5.85), (5.86), we arrive at the problem of finding a solution ζ(τ, ε): ζ(·,ε) e Cl\r], Z(T,·) e C[e], ζ(Τ,0) = ZO(T,CT), of the boundary-value problem dz — =Az

+ f + ε{β(Αζ

+ f ) + ( 1 + εβ)Ζ{ζ, ε)},

lz(-,e) = a + eJ(z(-,e),e).

(5.93) (5.94)

n

Thus, the required solution ζ — ζ(·,ε): [a, 6*] —> R of the boundary-value problem (5.93), (5.94) is defined on the interval of fixed length [a, b*]. From the geometric point of view, the change of variables (5.91), (5.92) is a linear mapping of the interval [a, 6(e)] whose length depends on ε onto the interval [a,b*\ of constant length. Moreover, for β > 0, the interval [a, 6(e)]

Section 5.3

Autonomous Boundary-Value Problems

147

suffers linear contraction and, for β < 0, it is expanded. It is easy to see that, for a periodic boundary-value problem with a = 0 and b* = T, the change of variables (5.91), (5.92) coincides with the change of variables used in [65], [101], and [149]. In problem (5.93), (5.94), we now set ζ(τ,ε) =

z0(T,CR)

+

χ(τ,ε).

This yields the boundary-value problem dx — = Ax + ε{β[Α(ζ0 + X) + f] + (1 + εβ)Ζ(ζ0 ατ

+ χ, ε)},

Ιχ{·,ε) = ε.7(ζο(·, 0).

This solution satisfies the boundary condition ζ(0)-ζ(1

+ εβ(ε))

= 1η

tanc tan(e(l + εβ(ε))/2

+ c)

= 0

for tanc - tan(e(l + εβ{ε))/2

+ c) = 0,

and, therefore, sin(e(l+e/3(e))) = 0. This equality does not hold for sufficiently small ε

(0 < ε < εο) and, consequently, the boundary-value problem (5.98)

does not have the required solution.

Sufficient Condition. Assume that the necessary condition for the existence of solutions of the analyzed problem is satisfied. In this case, the solution of problem (5.95), (5.96) admits the following representation: χ ( τ , ε ) = Xt{T)CT

-(-

χ^(τ,Ε),

where XW(T,£)

=

£

X(T)Q

+

J{Z

0

+

X,£)

+ G{ß[A(zo The vector function Ζ(ζ,ε)

+ χ) + f ] + (1 + εβ)Ζ{ζ0

+ ®,ε)}(τ).

and vector functional J(z(-,a),e)

are contin-

uously differentiable with respect to 2 in the vicinity of the generating solution ZQ(T, C*) and continuous in ε in the vicinity of the point ε = 0. Hence, we can select terms linear in χ and terms of order zero with respect to ε in the vector function Z(ZQ + χ,ε) Z(Z0(T,

and vector functional J(ZQ + χ,ε).

c*r) + χ(τ, ε),ε)

= Ζ(ζο(Τ, c*T), 0) + ΑΧ{Τ)χ

As a result, we obtain +

Ψι{χ,ε),

(5.99)

150 Boundary-Value Problems for Ordinary Differential Equations

Chapter 5

where »>1(0,0) = 0,

3 & M - 0 ,

and J(z0(-,c*r)

+

x(-,e),£) = J{z0(-,c*r),0)

where 1\χ(·,Ε)

+ hx(-,s)

+ Μχ(-,ε),ε)

£ Rm,

is a linear (as a function of x) part of the functional

J(ZQ(·,

(5.100) C*)

J i ( x ( - , e ) , e ) is its nonlinear part, Ji(0,0) = 0, and dJ\(0,0)/dx

χ(-,ε),ε),

+

=

0. Expansion (5.99) is used to transform the inhomogeneity of system (5.95). We have ß[A(z0

+ x) + f ] + ( 1 + εβ)Ζ(ζ0 = ßA[XT{r)c*r

+ χ, ε)

+ ζο{τ, /)] + β Αχ + ßf + + Αι (τ)χ

= fo (τ, c*) + βΑζ0(τ, + Αί{τ)χ

+ φ1(χ,ε)

/O(T,C*) +

+ εβΖ(ζ,

ε)

c*) + ßf + β* Αχ + β Αχ + φι(χ,ε)

= /ο(τ, c*) + [β*Α + Αι(τ)]χ =

Ζ(ζ0,0)

+

εβΖ(ζ,ε)

+ [Αζο(τ, c*r) + f}ß + R(x, ε)

[β* A + Λ ι ( τ ) ] Χ Γ ( τ ) £ Ν + [β* A + Α 1 ( τ ) ] χ ( 1 )

+ [Az0(T,c*r)

+ f]ß+

= /ο (τ, c*) + Α(τ)ο+[β*Α

R(x,£)

+ Α1(τ)}χ^(τ,ε)

+

ϋ(χ,ε),

where Αχ{τ)

= {\β*Α + Αχ{τ)}Χτ(τ),Αζο{τ,*

+ l1x(;£)

PQ.{j(z0(;C*r),0)

+ J1(x(;£),£)-l

J Κ (•, s){f0(s,

C*)

a

+ Ä i ( s ) c + [β* A + A ! ( s ) ] x ( 1 ) ( s , ε) + Λ ( χ , ε ) } ^ } = 0. By using the equation for generating amplitudes (5.97), we obtain f>*

+ hxW(;£)

PQ'd{hXr(-)cr

+ J1(x(;£),£)-l

J Κ(·,

β){^ι(β)θ

a

+ [β*A + A i ( s ) ] x W ( s , e ) + R(x,£)}ds

} = 0,

whence b· PQ>{hXr(-)h

~ IJ

tfO^Ä^sjc

a

= -PQ.{hx{1\;£)

+

J1(x(;£),£)

b' - I J K{;

s){[ß*A

+ Ai(s)]xW +

ϋ(χ,ε)}ά8γ

152 Boundary-Value Problems for Ordinary Differential Equations

Chapter 5

If we now introduce a d χ (r + 1) matrix with constant entries 6*

B0 = PQ.d[hXT{-)h-l

J

Κ(.,s)Äi(e)de},

α

then we get the following equation for the vector c € ß r + 1 , c = c( ) e C[e]: B0c = - Ρ ο - ^ ι χ ^ - , ε ) + Ji(s(.,e),e) b' -I J Ki^s^A

+ Aiis)]!^+

R(x,e)}dsy

(5.103)

α In order that this equation be solvable, it is necessary and sufficient that PB*0PQ*{hx{1)(;e)

+

Ji(x(;£),e)

b· -I J A:(-,s){[/?*A + Ai(s)]x( 1 ) + i ? ( x , e ) } i i s } = 0 .

(5.104)

a

Assume that Ρβζ = 0, which means that the matrix Bo is of full rank, i.e., rank Bo = d. Then condition (5.104) is satisfied. In this case, system (5.103) is solvable and possesses the (n — m + 1)-parameter family of solutions c = -B+PQ'd{hx{1)(;e) + Λ(*(·,ε),ε) b' - I J AT(-,s){[/3*A + A 1 (s)]x( 1 ) + R(x,e)}ds}

+ PBoc

Vc €

Rr+1

a

(rank Ρβ0 = dim N(Bo) = r + l — d = n — η\ — m + n\ + l = n — m + 1 ) , where Pb* is a d χ d matrix orthoprojector {Pb'0 '• Rd —> N(Bq)), Pb0 is an (r + 1) χ (r + 1) matrix orthoprojector (Pß 0 : Rr+1 —> Ν (Bo)), and I\ is an r χ (r + 1) matrix with constant entries:

r

h =

"1 0

0 1

··· •··

0 0

0 ' 0

0

0

···

1

0

Section 5.3

Autonomous Boundary-Value Problems

153

Thus, under the assumption that Pb· = 0, one of the solutions χ(τ,ε): 1 χ(·,ε) e C [τ], χ (τ, ·) e C[e], x(r, 0) = 0 of the boundary-value problem (5.101), (5.102) can be found from the following equivalent operator system: χ(τ,ε)

= Xr(r)Iic

+χ^\τ,ε),

c=-B+PQi{l1x^M

+

J1{x(.,£),e)

b· -I J K(-,s){[ß*A

+ Ai(s)]xW + Λ ( ® , ε ) } ώ } ,

(5.105)

a

χΜ{τ,ε)

= £ X ( t ) Q + J { z o + χ,ε) +

£G{f0(s,c*)

+ Ä{s)c + [β* A + Λι(β)]ζω(β,ε) + Ä(s,e)}(r). It is known [19] that the operator system (5.105) belongs to the class of systems whose solution can be found by the method of simple iterations convergent for all ε e [0,ε*]. The lower bound of the quantity ε* can be established by using the Lyapunov majorizing equations. Thus, assume that the matrix Bo is of full rank, i.e., rank Bo = d. It is natural to seek the first approximation χι(τ, ε) to the solution χ (τ, ε) of system (5.105) in the form of a solution of the boundary-value problem

Ιχι(Ίε)

= eJ(zö(-,cf),0),

T+1

where c* = col (c*,ß*) 6 R satisfies equation (5.97). The required solution has the form χ ι ( τ , ε ) = ΧΓ(τ)Ιιοο

+ χ^(τ,ε),

x?(T,£)=£X(T)Q+J(zo(;c*r),0)

Co = 0, +

£Gfo(s,c*)(T).

The second approximation Χ2 (τ, ε) is sought as a solution of the problem j = Ax2 + ε { / ο ( τ , c*) + Ä1 (T)CI

+ [/3Μ + Α 1 (τ)]χί 1 ) (τ ) ε) + Λ ( 4 1 ) , ε ) } , Ιχ2(·,ε)

=^εJ{zo(•,c*) +

Xl{•,ε),ε)

154 Boundary-Value Problems for Ordinary Differential Equations

Chapter 5

in the form x\(τ,ε)

— Xt{t)I\c\

4 1 } ( τ , ε ) = sX(t)Q+J{zo

+41)(t>£).

+ χ ι , ε ) + eG{f0(s,

c*)

+ Äi(e)ci + [β* A + Α ^ χ ^ ^ , ε ) + ^ ( x ^ . e J X r ) . The condition of solvability of the system deduced for the second approximation yields the following equation for the vector c\: Boci = - P Q ' d { h x ? ( ; £ ) + ^ ( χ ^ Ο , ε Χ ε ) b' -I J K^sHIFA

+A ^ x P +

RixM^ds}.

a

The solution of this equation has the form CI = - B + P ^ h x ^ (·, ε) + Jx

(·, ε), ε)

b* -I I K(-,s){[ß*A

+ A ^ x ? + R i x ? ,e)}ds}+Ppcp,

a

where Pp is an (r + 1) χ ρ matrix whose columns form a complete system of ρ linearly independent columns of the orthoprojector Pb0, ρ = rank Pb 0 = η — m + 1, and cp = Cp(-) G C[e], cp £ Rp, is a vector such that cp(0) = 0. We continue the process of calculations and, as a result, conclude that one of the solutions of the operator system (5.105) can be obtained with the help of the following iterative procedure: ck = - B + P Q ^ h x ^ M + Ji(x f c (.,e),e) b'

Section 5.3 ®£ii(T> Ε)

Autonomous Boundary-Value Problems c*) + xfc(·, ε), ε) + eG{f0(s, c*)

= SX{T)Q+J(ZO{·,

+ Ä^ck

155

+ [β* A + Α ^ χ ^ ^ , ε ) +

R(xk{s,£),€)}(r),

x f c + i (τ,ε) = X r (r)/ic f c + χ ^ τ , ε ) ,

(5.106)

{

χο(τ,ε) = χ ο \ τ , ε ) = 0, fc = 0 , 1 , 2 , 3 , . . . . Assume that the homogeneous generating problem (5.87), (5.88) has an rparameter (r = η — η ι, rank Q = ηχ) family of nontrivial solutions and condition (5.89) is satisfied. Then the inhomogeneous generating problem is solvable and possesses a solution of the form (5.90). The sufficient conditions for the existence of solutions of the original weakly nonlinear problem (5.85), (5.86) are established by the following assertion: Theorem 5.8. Assume that the matrix Bo is offull rank (Pß* = 0). Then, for each root c* = col (c*,ß*) € RT+l of the equation for generating amplitudes (5.97), the boundary-value problem (5.95), (5.96) possesses at least one solution χ(τ,ε) suchthat χ(·,ε) € C^a,^*], χ(τ, ·) € C[0,e*], and χ(τ,0) = 0. This solution can be obtained from the operator system (5.105) as a result of the iterative procedure (5.106) convergent for sufficiently small ε £ [0,ε,]. In this case, the boundary-value problem (5.93), (5.94) has at least one solution ζ(τ,ε) suchthat ζ(·,ε) G C^O.b*], ζ(τ, ·) € C[0,e o ], and ζ(τ, 0) = ZQ(T, c*). This solution can be found with the help of the iterative process (5.106) and the following formula: zjk(T,e)

= ZO(T,C*)

+ xk(T,E),

k = 0,1,2,... .

In view of the change of the independent variable (5.92), Theorem 5.8 specifies the solutions ζ(·,ε) e (71[α,6(ε)], z(t,·) e C[0,eo], z(t,0) = z(t,c*), of the boundary-value problem (5.85), (5.86) depending on an arbitrary constant cp e Rp (p = rank Ρβ0 — η — τη + 1). Furthermore, the ( r + l ) t h component β = β(ε) = β (ε) — β* of the vector constant c = c(s) e Rr+1 is a correction to the length of the interval used for the construction of the required solution z(t, ε) of the original problem. In the case τη — τι (most frequently encountered in the theory of oscillations), the equality Ρ β · = 0 means that det BQ Φ 0. In view of the fact that the last component of the vector cv can be made zero by the proper choice of the reference point for the independent variable t, the

156

Boundary-Value Problems for Ordinary Differential Equations

Chapter 5

periodic boundary-value problems (m = η, ·7(ζ(·,ε),ε) =0, a — 0, / = 0, Ιζ{·, ε) = z(0, ε) — z(Ti (ε), ε)) can be studied with the help of a theorem which follows from Theorem 5.8 and can be formulated as follows [36]: Theorem 5.9. Assume that the boundary-value problem (5.85), (5.86) satisfies the conditions presented above. Then, for each simple (det BQ φ 0) root c* = col (c*_l5 /3*) € Rr of the equation for generating amplitudes τ F(C*) = PQ4 J K(·, s)fo(s, c*)ds = 0,

ο

the boundary-value problem possesses a unique T\ (ε)-periodic solution z(t,s) suchthat ζ(·,ε) e (^[Ο,Τ^ε)] and z(t,·) e C[e] which turns into the generating solution z(t, c*_ x ) for ε = 0. The indicated solution can be obtained as a result of the iterative process τ

Cfc =

—BQ^PQ'J, J

ο

X(.,S){[/3*A + J4i(s)]xl1)(S)e) + JR(ifc(S,e),e)}dS,

τ

4 + ι ( ^ ) = ε / G(T>s){/o(e,C*) 0 + Äi(s)cfc +

[β* A + Ai^xj^s.O

+ R(xk(s,

e),e)}ds,

Xfe+i (τ,ε) = Xr_i(r)/icfc + ζ ^ τ , ε ) , zk+i{r,s) = zo(r, c*_j) + xIC+I{T,£),

χο(τ,ε) = χ{ο\τ,ε) ξ 0, fc = 0,1,2,... , convergent for ε

5.4.

£ [0, ε»] C [Ο,εο].

General Scheme of Investigation of Boundary-Value Problems

Let us now outline the general scheme used for the analysis of weakly perturbed linear and nonlinear boundary-value problems.

Section 5.4 General Scheme of Investigation of Boundary-Value Problems 157 1. For a given weakly nonlinear boundary-value problem, we consider the corresponding generating (ε = 0) boundary-value problem. To do this, for the known normal fundamental matrix X(t) of the homogeneous differential system (5.3), we construct the m χ η matrix Q = IX(-) and the orthoprojectors PQ and PQ*. Further, it is necessary to determine whether the problem is critical (rank Q < m ~ PQ- Φ 0) or noncritical (rank Q — M ~ PQ· = 0 ) . Then we check the criterion of solvability (5.9) of the generating boundary-value problem and find the family of generating solutions zo(t,Cr) of the form (5.11). 2. In the noncritical case (PQ* = 0 ) , the criterion of solvability (5.15) is satisfied for the generating boundary-value problem and it has at least one generating solution z0(t, Cr) given by (5.11) (r = η — m ) . Then, for any nonlinearity satisfying (5.36), the original weakly nonlinear noncritical boundary-value problem (5.32) possesses at least one solution z(t, ·) e C[e] which turns into the generating solution for ε = 0. This solution can be obtained from the n-dimensional operator system (5.34), (5.37) as a result of the iterative process (5.39), (5.40) convergent for ε £ [0, ε»]. The range of convergence of the iterative process and the accuracy of the approximate solutions can be estimated by the method of finite majorizing Lyapunov equations [19], [65]. 3. In the critical case (rank Q = RI\ < m), the investigation of boundaryvalue problems of the form (5.43) becomes much more complicated. Assume that the criterion of solvability (5.9) is satisfied for the generating boundary-value problem and it has an r-parameter (r = η — η χ) family of generating solutions zo(t,Cr) of the form (5.11). Then, by using the equation for generating amplitudes (5.46), we determine constants cv = CQ e RR specifying the generating solution associated with the solution z(t, ·) e C[s], z(t, 0) = zo(t, CQ) of the original boundary-value problem (5.43). Further, we construct the d χ r matrix BQ. If rank BQ = d ~ Ρβ* — 0, then the critical boundary-value problem of the first order (5.43) possesses at least one (unique for η = m) solution z(t, •) £ C[e\ which turns into the generating solution zo(i, cj) for ε = 0. This solution can be obtained from the (2η + r)-dimensional operator system (5.47), (5.52) as a result of the iterative process (5.57) convergent for ε e [0,ε„]. If rank BQ < d ~ ΡΒ* φ 0, then, in order to establish conditions required for the existence of solutions of the original problem, it is necessary to construct a d χ r matrix B\ given by relation (5.62) and its orthoprojectors PBx and Ρβ*. If Ρ β · Ρ β · = 0, then, for any nonlinearity satisfying relation (5.73), the

158 Boundary-Value Problems for Ordinary Differential Equations

Chapter 5

critical boundary-value problem of the second order (5.43) possesses at least one (unique for Pbo-Pbi = 0) solution z(t, •) G C[e] which turns into the generating solution z(t, 0) = zo(t, Cq) for ε = 0. The indicated solution can be found from the 2(n + r)-dimensional operator system (5.66) by using the iterative process (5.71) convergent for ε £ [Ο,ε*]. 4. If the generating (ε = 0) inhomogeneous linear boundary-value problem is unsolvable (condition (5.9) is not satisfied), then, as indicated above, one can either construct a quasisolution of the problem or make it solvable by adding small linear perturbations [17]. In this case, the requirements guaranteeing the appearance of solutions of weakly perturbed linear boundary-value problems are specified by introducing dx r matrices Bo, B\, etc. [19], [98]. The solutions of perturbed boundary-value problems can be constructed by using the VishikLyusternik algorithm [150] in the form of convergent Laurent series in powers of a small parameter (see Sections 5.6 and 6.4). Thus, the data required to establish the conditions of solvability of weakly perturbed linear and nonlinear boundary-value problems are contained in the m χ η matrix Q and the chain of dx r matrices Bo, B\,..., constructed according to the coefficients of the original differential system.

5.5. Periodic Solutions of the Mathieu, Riccati, and Van der Pol Equations The results obtained in the previous sections remain true and can be applied [16], [17], [18], [99] to the investigation of periodic boundary-value problems frequently encountered in applications and fairly well studied in [65], [93], [94], [97], [122], [143]. The examples presented below illustrate some possibilities of the proposed procedures. Example 1. Consider the periodic boundary-value problem for the scalar Riccati equation i = ao(t)z +