Universitext Editorial Board (North America):
S. Axler K.A. Ribet
Universitext Editors (North America): S. Axler and ...

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

Universitext Editorial Board (North America):

S. Axler K.A. Ribet

Universitext Editors (North America): S. Axler and K.A. Ribet Aguilar/Gitler/Prieto: Algebraic Topology from a Homotopical Viewpoint Aksoy/Khamsi: Nonstandard Methods in Fixed Point Theory Andersson: Topics in Complex Analysis Aupetit: A Primer on Spectral Theory Bachman/Narici/Beckenstein: Fourier and Wavelet Analysis Badescu: Algebraic Surfaces Balakrishnan/Ranganathan: A Textbook of Graph Theory Balser: Formal Power Series and Linear Systems of Meromorphic Ordinary Differential Equations Bapat: Linear Algebra and Linear Models (2nd ed.) Berberian: Fundamentals of Real Analysis Blyth: Lattices and Ordered Algebraic Structures Boltyanskii/Efremovich: Intuitive Combinatorial Topology. (Shenitzer, trans.) Booss/Bleecker: Topology and Analysis Borkar: Probability Theory: An Advanced Course Böttcher/Silbermann: Introduction to Large Truncated Toeplitz Matrices Bridges/Vît¸a˘: Techniques of Constructive Analysis Carleson/Gamelin: Complex Dynamics Cecil: Lie Sphere Geometry: With Applications to Submanifolds Chae: Lebesgue Integration (2nd ed.) Charlap: Bieberbach Groups and Flat Manifolds Chern: Complex Manifolds Without Potential Theory Cohn: A Classical Invitation to Algebraic Numbers and Class Fields Curtis: Abstract Linear Algebra Curtis: Matrix Groups Debarre: Higher-Dimensional Algebraic Geometry Deitmar: A First Course in Harmonic Analysis (2nd ed.) DiBenedetto: Degenerate Parabolic Equations Dimca: Singularities and Topology of Hypersurfaces Edwards: A Formal Background to Mathematics I a/b Edwards: A Formal Background to Mathematics II a/b Engel/Nagel: A Short Course on Operator Semigroups Farenick: Algebras of Linear Transformations Foulds: Graph Theory Applications Friedman: Algebraic Surfaces and Holomorphic Vector Bundles Fuhrmann: A Polynomial Approach to Linear Algebra Gardiner: A First Course in Group Theory Gårding/Tambour: Algebra for Computer Science Goldblatt: Orthogonality and Spacetime Geometry Gustafson/Rao: Numerical Range: The Field of Values of Linear Operators and Matrices Hahn: Quadratic Algebras, Clifford Algebras, and Arithmetic Witt Groups Heinonen: Lectures on Analysis on Metric Spaces Holmgren: A First Course in Discrete Dynamical Systems Howe/Tan: Non-Abelian Harmonic Analysis: Applications of SL(2, R) Howes: Modern Analysis and Topology Hsieh/Sibuya: Basic Theory of Ordinary Differential Equations Humi/Miller: Second Course in Ordinary Differential Equations (continued after index)

Douglas S. Bridges and Luminit¸a Simona Vît¸a˘

Techniques of Constructive Analysis

Douglas S. Bridges Department of Mathematics/Statistics University of Canterbury Christchurch, New Zealand [email protected]

Luminit¸a Simona Vît¸a˘ Department of Mathematics/Statistics University of Canterbury Christchurch, New Zealand [email protected]

Editorial Board (North America): S. Axler Mathematics Department San Francisco State University San Francisco, CA 94132 USA [email protected]

K.A. Ribet Mathematics Department University of California at Berkeley Berkeley, CA 94720-3840 USA [email protected]

Mathematics Subject Classification (2000): 03F60, 26E40, 46S30, 47S30, 03F55, 03F65, 68Q99 Library of Congress Control Number: 2006926441 ISBN-10: 0-387-33646-X ISBN-13: 978-0387-33646-6

e-ISBN-10: 0-387-38147-3 e-ISBN-13: 978-0387-38147-3

Printed on acid-free paper. © 2006 Springer Science+Business Media, LLC All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. 9 8 7 6 5 4 3 2 1 springer.com

Ideal pierdut ˆın noaptea unei lumi ce nu mai este, Lume ce gˆandea ˆın basme ¸si vorbea ˆın poezii, O! te v˘ ad, te-aud, te cuget, tˆ an˘ ar˘ a ¸si dulce veste Dintr-un cer cu alte stele, cu-alte raiuri, cu alt¸i zei. —Mihai Eminescu, “Venere ¸si Madon˘ a” Oh, ideal lost in night-mists of a vanished universe: People who would think in legends—all a world who spoke in verse; I can see and think and hear you—youthful scout which gently nods From a sky with diﬀerent starlights, other Edens, other gods. —Mihai Eminescu, “Venus and Madonna” (translated by Andrei Banta¸s)

This image is a courtesy of The Times of London. Printed in the February 3, 2004 issue.

Preface

Rosencrantz: Shouldn’t we be doing something... constructive? Guildenstern: What did you have in mind? —Tom Stoppard, Rosencratz and Guildenstern Are Dead

We have written this book in order to provide an introduction to constructive analysis, emphasising techniques and results that have been obtained in the last twenty years. The intended readership comprises senior undergraduates, postgraduates, and professional researchers in mathematics and theoretical computer science. We hope that our work will help spread the message that doing mathematics constructively is interesting (it can even be fun!) and challenging, and produces new, deep computational information. An appreciation of the distinction between constructive and nonconstructive has become more widespread in this era of computers. Nevertheless, there are few books devoted to the development of mathematics in a rigorously constructive/computable fashion, although there are some, primarily concentrating on logic and foundations, in which the odd chapter deals with constructive mathematics proper as distinct from its underlying logic or set theory. It is now almost forty years since the publication of Errett Bishop’s seminal monograph Foundations of Constructive Analysis [9], which in our view is one of the most remarkable intellectual documents of the twentieth century, and more than twenty since the appearance of its outgrowth [12]. In the intervening years there has been considerable activity in constructive analysis, algebra, and topology; in related foundational areas such as type theory [69]; and in the relation between constructive mathematics and computer science (for example, program extraction from proofs [42, 70, 51]). Believing that a new introduction to the mathematical, as distinct from the foundational, side of the subject is overdue, we embarked upon this monograph. Our book is intended not to replace, but to supplement, Bishop’s original classic [9] and the later volume [12] based thereon. Both of those two monographs cover

viii

Preface

aspects of analysis, such as Haar measure and commutative Banach algebras, that we do not mention. We cover some topics that are found in [9] and [12] (it would be almost inconceivable to produce a book like ours, dealing with constructive mathematics for nonexperts, without proving, for example, basic results about locatedness and total boundedness); but we have tried to provide improved proofs whenever possible. However, much of the material we present was simply not around at the time of writing of [9] or [12]. Instead of systematically developing analysis, beginning with the real line and continuing through metric, normed, and Hilbert spaces to its higher reaches, we have chosen to write the chapters around certain themes or techniques (hence our title). For example, Chapter 3 is devoted to the λ-technique, which, since its ﬁrst use in the proof of Lemma 7 on page 177 of [9], has become a surprisingly powerful tool with applications in many areas of constructive analysis. A major inﬂuence in the application of the λ-technique was Ishihara’s remarkable paper [60], which showed that a subtle use of the technique could enable us to prove disjunctions whose proof, although trivial with classical logic, appears at ﬁrst sight to be constructively out of the question. This paper opened up many new pathways in constructive analysis. Chapter 1 introduces constructive mathematics and lays the foundations for the later chapters. In Chapter 2 we ﬁrst present a new construction of the real numbers, motivated by ideas in [2]. After deriving standard properties such as the completeness of R, we introduce metric spaces, with the major theme of locatedness, and normed linear spaces. When we discuss metric, normed, and Hilbert spaces, we assume some familiarity with the standard classical deﬁnitions of those concepts and with those elementary classical properties that pass over unchanged to the constructive setting. Chapter 3 we have already referred to. The main theme of Chapter 4 is ﬁnitedimensionality, but the chapter concludes with an introduction to Hilbert spaces. Chapter 5 deals with convexity in normed spaces. Starting with some elementary convex geometry in Rn , the chapter goes on to handle separation and Hahn–Banach theorems, locally convex spaces, and duality. Following Bishop, we describe those linear functionals that are weak∗ -uniformly continuous on the unit ball of the dual space. We then give a new application of the technique used to prove that result, thereby characterising certain continuous linear functionals on the space of bounded operators on a Hilbert space. In Chapter 6 we derive a range of results associated with the theme of locatedness and with the λ-technique introduced in Chapter 3. We pay particular attention to necessary and suﬃcient conditions for convex subsets of a normed space to be located, and to connections between properties of an operator on a Hilbert space and those of its adjoint—when that adjoint exists: it may not always do so constructively. The ﬁnal section of the book deals with a relatively recent version of Baire’s theorem and its applications, and culminates in constructive versions of three of the big guns in functional analysis: the open mapping, inverse mapping, and closed graph theorems.

Preface

ix

Which parts of the book deal with new material, compared with what appeared in [12]? We have already mentioned the new construction of the real numbers, in Chapter 2. Notable novelties in the later chapters include all but one result in Chapter 3 on the λ-technique; the section on convexity, Ishihara’s results on exact Hahn–Banach extensions, and our characterisation theorem for certain continuous linear functionals, all in Chapter 5; and virtually all of Chapter 6. Throughout the book there are what we hope will be seen as improvements and simpliﬁcations of proofs of many results that were given in [9] or [12]. What do we mean by “constructive analysis” in the title of this book? We do not mean analysis carried out with the usual “classical” logic within a framework, such as recursive function theory, designed to capture the concept of computability. In our view, such a notion of constructive has at least two drawbacks. First, by working within, say, the recursive setting, it can make the mathematics look less like normal mathematics and much harder to read. Secondly, the recursive constraint removes the possibility of other interpretations of the mathematics, such as Brouwer’s intuitionistic one [48]. Our approach, on the other hand, has neither of these features: the mathematics looks and reads just like the mathematics one is used to from undergraduate days, and all our proofs and results are valid in several models. They are valid in the recursive model, in intuitionistic mathematics, and, we believe, in any of the models for “computable mathematics” (including Weihrauch’s Type Two Eﬀectivity Theory [91], within which Andrej Bauer has recently found a realisability interpretation of constructive mathematics within Weihrauch’s theory [5]). They are also valid proofs in standard mathematics with classical logic. For example, our proof of the Hahn–Banach theorem (Theorem 5.3.3) is, as it stands, a valid algorithmic proof of the classical Hahn–Banach theorem. Moreover—and this is one advantage of a constructive proof in general—our proof embodies an algorithm for the construction of the functional whose existence is stated in the theorem. This algorithm can be extracted from the proof, and, as an undeserved bonus, the proof itself demonstrates that the algorithm is correct or, in computer science parlance, “meets its speciﬁcations”.1 So how do we achieve all this? Simply by changing the logic with which we do our mathematics! Instead of using classical logic, we systematically use intuitionistic logic, which was abstracted by Heyting [52] from the practice of Brouwer’s intuitionistic mathematics. The remarkable fact is that every proof carried out with intuitionistic logic is fully constructive/algorithmic. (Is this the “secret on the point of being blabbed” that appears in the epigraph to Bishop’s book?) Unfortunately, too few mathematicians outside the mathematical logic community are aware of this serendipity and dismiss both intuitionistic logic and constructive mathematics as at best a marginal curiosity. This contrasts sharply with the theoretical computer science community, in which there is considerable knowledge of, and interest in, the computational power of intuitionistic logic. 1 We do not carry out program-extraction from proofs in our book. For more on this topic see [42, 51, 70].

x

Preface

Reading constructive mathematics demands careful interpretation. A theorem in this book might look like a familiar one from classical analysis, but with more complicated hypotheses and proof. However, the statement of the theorem will be phrased so that the explicit algorithmic interpretation is left to the reader; and the additional hypotheses will be necessary for a constructive proof, which will contain algorithmic information that is excluded from the classical proof by the latter’s use of principles outside intuitionistic logic. Consider, for example, the following statement: (*) Let C be an open convex subset of a normed space X, let ξ ∈ C, and let z ∈ X be bounded away from C. Then the boundary of C intersects the segment [ξ, z] joining ξ and z. This is trivial to prove classically; but to ﬁnd/construct the (necessarily unique) point in which the boundary of C intersects [ξ, z] is a totally diﬀerent matter. The constructive theorem (Proposition 5.1.5 below) requires us to postulate that the union of C and its metric complement −C (the set of points bounded away from C) be dense in X, and that X itself be a complete normed space. The constructive proof, though elementary, requires some careful geometrical estimation that would be supererogatory in the natural classical proof by contradiction. The beneﬁt of that estimation and of the use of intuitionistic logic is that we could extract from the constructive proof an implementable algorithm for ﬁnding the point where the segment crosses the boundary. In turn, this would enable us to produce an algorithm for constructing separating hyperplanes and Hahn–Banach extensions of linear functionals, under appropriate hypotheses. We could have made the algorithmic interpretation of the constructive version of (*) explicit by stating the proposition in this way: There is a “boundary crossing algorithm” that, applied to the data consisting of (i) an open convex set C in a Banach space X such that C ∪−C is dense in X, (ii) a point ξ of C, and (iii) a point z of −C, constructs the point where the boundary of C intersects the segment [ξ, z] . Even this is not really explicit enough. A full description of the data to which the boundary crossing algorithm applies would require explicit information about the algorithms for such things as these: membership of C; the convergence of Cauchy sequences in X; the computation, for given x in X and ε > 0, of a point y of C ∪−C such that x − y < ε (and even the decision between the cases “y ∈ C” and “y ∈ −C”); and so on. Such explicit description of algorithmic hypotheses would become an ever greater burden on writer and reader alike as the book probed deeper and deeper into abstract analysis. It is a matter of sound sense, even sanity, to unburden ourselves from the outset, relying on the reader’s native wit in the interpretation of the statements of our constructive lemmas, propositions, and theorems. We should make it clear that we are not advocating the exclusive use of intuitionistic logic in mathematics. That logic is, we believe, the natural and right

Preface

xi

one to use when dealing with the constructive content of mathematics. To abandon classical logic in those ﬁelds (such as the higher reaches of set theory) where constructivity is of little or no signiﬁcance makes no sense whatsoever. Nevertheless, it is remarkable how much mathematics actually has what Bishop called “a deep underpinning of constructive truth”.

Christchurch, New Zealand January 2006

Douglas Bridges Luminit¸a Simona Vˆıt¸a ˘

Acknowledgments

It is never easy to apportion thanks properly among the many who have contributed to this book either directly or by their support and encouragement at various stages of our professional lives. We do, however, have special thanks for the following people: Cris Calude, who was responsible for bringing Vˆıt¸˘a to New Zealand to begin what has proved a very fruitful research partnership with Bridges, and who has been tireless in his encouragement of our work over many years. Hajime Ishihara, who, in conjunction with one or both of the authors, was largely responsible for much of the work in Chapter 6 and whose inﬂuence can be seen in several other places in the book. (Hajime’s contributions to constructive functional analysis have been remarkable and deserve to be recognised more widely.) In addition, he has hosted us many times at the Japan Advanced Institute of Science & Technology, each of our visits there being both memorable and highly productive. Peter Schuster, not only for his contributions to constructive mathematics since he joined our community ten years ago, but also for acting as organiser, fund-raiser, and host on our many research visits to Munich. His considerable eﬀorts in securing a DAAD Gastprofessorship for Bridges in the Mathematisches Institut der LudwigMaximilians-Universit¨ at (LMU), M¨ unchen, in 2003 provided us with the time and environment in which we could break the back of the writing of this book. Our early drafts of Chapters 1–5 formed the basis of graduate lectures by Bridges at LMU in 2003. We thank the students in that course for patiently receiving that material and for suggesting corrections and improvements to our presentation of it. We are also grateful to the DAAD for supporting Bridges as a Gastprofessor at LMU for that year; Otto Forster and Helmut Schwichtenberg, our hosts at LMU in 2003 and on several other occasions;

xiv

Acknowledgments

the New Zealand Foundation of Research, Science and Technology for Vˆıt¸˘a’s postdoctoral fellowship from 2002 to 2005 and for supporting her extended visit to LMU to enable us to work together on the book; our departments at Canterbury and Galat¸i; Josef Berger, Hannes Diener, Maarten Jordans, and Robin Havea, who have kindly assisted us with proofreading various chapters. We say a warm “thank you” to our friends Imola and Attila Zsigmond, and Helmut and Eva Pellinger, who were wonderful hosts during our time in Munich. Finally, we want to thank our families for their continuing love and support.

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vii

Introduction to Constructive Mathematics . . . . . . . . . . . . . . . . . . . . .

1

1.1 What Is Constructive Mathematics? . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.2 A Very Brief History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5

1.3 Intuitionistic Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.4 Informal Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

Techniques of Elementary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

2.1 The Real Number Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

2.2 Metric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

37

2.3 Normed Linear Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

47

The λ-Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

3.1 Introduction to the Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

3.2 Ishihara’s Tricks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

71

Finite-Dimensional and Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . .

81

4.1 Finite-Dimensional Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

4.2 Best Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

88

4.3 Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

1

2

3

4

5

Linearity and Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

xvi

Contents

5.1 Crossing Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 5.2 Separation Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 5.3 The Hahn–Banach Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 5.4 Locally Convex Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 6

Operators and Locatedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 6.1 Smooth and Uniformly Smooth Normed Spaces . . . . . . . . . . . . . . . . . 147 6.2 Locatedness of Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 6.3 Adjoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 6.4 Functions of Selfadjoint Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 6.5 Locating the Kernel and the Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 6.6 Baire’s Theorem, with Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

1 Introduction to Constructive Mathematics

My task which I am trying to achieve is, by the power of the written word, to make you hear, to make you feel—it is, before all, to make you see. That, and no more, and it is everything. —Joseph Conrad, The Nigger of the ‘Narcissus’ In this chapter we ﬁrst sketch the history and philosophy that motivated the early workers in the ﬁeld of constructive mathematics. We then describe informal intuitionistic logic and discuss a number of elementary classical theorems that do not carry over to the constructive setting. Finally, we introduce an informal constructive theory of sets and functions. All this will prepare us for the presentation of the constructive theory of the real line R in Chapter 2, and for the more abstract analysis that will be described in later chapters.

1.1 What Is Constructive Mathematics? Proposition 20 in Book IX of the thirteen volumes of Euclid’s Elements states that Prime numbers are more than any assigned multitude of prime numbers —in current terms, there are inﬁnitely many primes. The modernised version of Euclid’s proof is often presented as follows. Suppose that there are only ﬁnitely many primes, say p1 , . . . , pn , and consider the integer p = p1 × p2 × · · · × pn + 1. Being greater than 2, p has prime factors (it may even be prime itself). Since the numbers pk are not divisors of p, each prime factor of p is distinct from each pk . This is absurd, since {p1 , . . . , pn } is supposed to be the set of all primes. From this contradiction we conclude that the set of primes is inﬁnite.

2

1 Introduction to Constructive Mathematics

Although at one level there appears to be nothing untoward about this proof, it can be criticised on two counts. First, it uses a totally unnecessary contradiction argument. If you look carefully, you will see that the proof actually embodies an algorithm that, applied to any ﬁnite set {p1 , . . . , pn } of primes, enables you to compute a prime that is distinct from each element of that set. In other words, the use of a contradiction argument in the preceding paragraph has obscured the computational content of Euclid’s proof. The second criticism of the proof is a little more subtle, and deals with the notion of “inﬁnitely many”. The proof is based on the negative idea that a set is inﬁnite if and only if it is contradictory that it be ﬁnite. But an algorithmic recasting of Euclid’s proof, as suggested in the preceding paragraph, shows that the set S of primes is inﬁnite in a more positive, productive sense: namely, if we start with a ﬁnite subset F of S, then we can compute an element of S that is distinct from each element of F. From a traditional standpoint, the distinctions between the contradiction proof of Euclid’s theorem and the algorithmic one, and between the negative and positive notions of “inﬁnite”, are obscured if not invisible. For example, the two notions of “inﬁnite” are equivalent if we use traditional logic—or classical logic, as it is normally called—so the distinction has to be perceived at an aesthetic level rather than a mathematical one. The same applies, more generally, to a proof by contradiction of the existence of an object x with the property P (x). In such a proof one supposes that P (x) is false for all applicable objects x, deduces a contradiction, and then concludes that P (x) must, after all, hold for some x, even though the proof doesn’t tell us which x actually has the desired property. Classical logic draws no distinction between the “idealistic existence” demonstrated by such a proof and the “constructive existence” based on an algorithm that constructs x and shows that P (x) holds. In order to reveal such distinctions at a mathematical, rather than an aesthetic, level we shall adopt the radical expedient of changing our logic: throughout this book, we shall work with intuitionistic logic, an abstraction of the informal logic used in algorithmic thinking. How much analysis, as normally presented, is really nonconstructive —that is, essentially dependent on proofs by contradiction or other nonalgorithmic procedures? Consider, for example, the classical intermediate value theorem: If f : [a, b] −→ R is a continuous mapping such that f (a) < 0 and f (b) > 0, then there exists c ∈ (a, b) such that f (c) = 0. (Note that [a, b] and (a, b) respectively denote the closed and open intervals with endpoints a and b. We shall use standard notations, like (a, b] for the half-open interval, without further comment.) It might be thought that the common elementary proofs of the intermediate value theorem are constructive, enabling one to produce a zero c of the function f. For example, one proof uses interval-halving in the following way. Without loss of

1.1 What Is Constructive Mathematics?

3

generality, take a = 0 and b = 1. Consider f (1/2): if it is 0, then we take c = 0 and stop the process; if f (1/2) > 0, then f satisﬁes the hypotheses of the theorem with a = 0 and b = 1/2; if f (1/2) < 0, then f satisﬁes the hypotheses with a = 1/2 and b = 1. In each of the last two cases, we proceed with the interval-halving. This process either stops after a ﬁnite number of iterations and produces the required zero of f, or else it goes on ad inﬁnitum to produce a descending sequence of compact intervals whose unique point of intersection is the required zero. Isn’t this a fully algorithmic proof? Suppose we try to implement the algorithm embodied in this proof on a computer that works with 50-bit precision. What happens if we apply it to the cubic function f deﬁned on [0, 1] by 2 1 3 x− − 2−51 ? f (x) = x − 4 2 Here, 1 3 − 2−51 < 0, f (1) = − 2−51 > 0, 16 16 so we are well set to carry out the ﬁrst step of the interval-halving algorithm. Since our computer’s ﬂoating-point representation of f (1/2) is 0 (we have the phenomenon of underﬂow, in which the computer sets the small but nonzero number −2−51 equal to 0), the algorithm stops by outputting c = 1/2 as the place where f has a zero. But in this case the only zero of f in [0, 1] lies between 3/4 and 1, more than one quarter of the entire interval away from the output value 1/2. f (0) = −

Now, one could object that this example is misleading, in that the problem arises from the level of precision in the computer rather than any intrinsic failing in the algorithm itself. To deal with this point, for each positive integer n let G(n) signify that 2n + 2 is a sum of two primes. Construct a binary sequence (an )n1 such that for each n, an = 0 if and only if either G(k) for all k n or else there exists k < n such that ¬G(k). Deﬁne a =

∞ n=1

an 2−n , and note that a = 0 if and only if the Goldbach conjecture,

Every even integer greater than 2 is a sum of two primes, holds. Using classical logic, apply the classical interval-halving algorithm to the cubic function f deﬁned on [0, 1] by 2 1 3 x− − a. f (x) = x − 4 2 As long as the status of the Goldbach conjecture remains undecided (which it has done since the conjecture ﬁrst appeared in 1742), no matter what ﬁnite precision

4

1 Introduction to Constructive Mathematics

our computer has, the algorithm will output 1/2 as a zero of f ; but if the Goldbach conjecture is false, then f (1/2) = −a < 0 and the ﬁrst zero of f in [0, 1] occurs between 3/4 and 1. In fact, the classical algorithm will give the correct output if and only if the Goldbach conjecture is true. However, as the reader may verify, a constructive approximate interval-halving argument, such as that expected in the solution of Exercise 11, does not give the possibly false value 1/2 for a zero of f. It produces a value that approximates the zero of f lying between 3/4 and 1, as accurately as the precision of the computer permits. In this example, the classically (but not constructively) deﬁned function taking the parameter a to the smallest root r(a) of f is discontinuous at a = 0. The classical algorithm correctly outputs r(a) = 1/2 in the case a = 0; but if a > 0, then, by outputting the value 1/2, the algorithm has failed to spot that the value of r(a) jumps from 1/2 to more than 3/4 as the parameter a increases from 0. There is a general principle that constructive proofs will involve continuity in parameters. Thus we cannot expect to prove constructively that for each a the above cubic function f has a smallest zero. The problem with the classical interval-halving algorithm is that the ﬁnite precision of the computer prevents it from making correct comparisons between two very close, but distinct, real numbers. If we are to develop mathematics in a computational manner, we have to ensure that such comparisons are barred. This barring can be done in at least two ways. One way is to use classical logic and to preclude nonalgorithmic “decisions” (such as whether two given numbers are equal) by developing the mathematics using a standard programming language or a more abstract algorithmic framework like that of recursive function theory. Another way is to change from classical to intuitionistic logic. The advantages of this second way are, ﬁrst, that nonalgorithmic “decisions” are automatically barred by the logic, and, second, that the resulting mathematics looks like the mathematics we are used to from school and university, without any special logical notation such as is used in, for example, recursive function theory. In this book we explore mathematics with intuitionistic logic. We work throughout with notions, like that of “inﬁnitely many” discussed earlier, that have positive computational meaning; and we present only algorithmic proofs—ones that show how we can, at least in principle, construct the objects whose existence is asserted in the statement of a theorem. We hope to convince the reader that, contrary to a widely held belief, intuitionistic logic suﬃces for the development of deep, interesting mathematics and often opens up new vistas that are hidden by classical logic. In other words, we want to justify our belief in the power of positive (constructive) thinking in mathematics.

1.2 A Very Brief History

5

1.2 A Very Brief History Although luminaries such as Leopold Kronecker had advocated a constructive approach to mathematics in the nineteenth century, the story of modern constructivism really begins with the publication, in 1907, of the doctoral thesis “On the Foundations of Mathematics” [41], in which the Dutch mathematician L.E.J. Brouwer introduced his intuitionistic mathematics (INT) as an alternative to traditional classical mathematics (CLASS). According to Brouwer, mathematical objects are free creations of the human mind, independent of both logic and language, and a mathematical object comes into existence precisely when it is constructed. Such a belief naturally leads to a rejection of existence proofs by contradiction, and a consequent scepticism about the meaning of many of the theorems of CLASS. Not surprisingly, Brouwer’s views met with at best indiﬀerence, and at worst hostility, from the large majority of his peers, for whom the elimination of nonconstructive arguments, with all their apparent power and fruitfulness, was too great a price to pay for a clariﬁcation of the meaning of mathematics. If we adhere to the principle that “existence” should always be interpreted constructively, then we are forced to dispense with the unrestricted use of the logical law of excluded middle (or excluded third ), P or (not P ), which we shall abbreviate to LEM. Recognising this consequence of his philosophical views, Brouwer went as far as to claim, The belief in the universal validity of the principle of the excluded third in mathematics is considered by the intuitionists as a phenomenon of the history of civilization of the same kind as the former belief in the rationality of π , or in the rotation of the ﬁrmament about the earth [44].

Subsequently, he introduced into INT some principles that led to results apparently contradicting aspects of classical mathematics. For example, Brouwer was able to prove that any real-valued function on [0, 1] is uniformly continuous. But to regard Brouwer’s mathematics as inconsistent with its classical counterpart is a serious oversimpliﬁcation of the situation, since the two types of mathematics are in many respects incomparable. Nevertheless, there was, and remains, a commonly held belief that too much mathematics has to be given up in order to accommodate Brouwer’s ideas. For example, Hilbert expressed his disagreement with Brouwer in words both forceful and memorable: Forbidding a mathematician to make use of the principle of excluded middle is like forbidding an astronomer his telescope or a boxer the use of his ﬁsts [54].

Despite continuing opposition, intuitionism survived and new constructive approaches to mathematics arose. In 1948–1949 in the former Soviet Union, A.A.

6

1 Introduction to Constructive Mathematics

Markov initiated a programme of recursive constructive mathematics (RUSS)— mathematics using intuitionistic logic and based on the Church–Markov–Turing thesis that all computable partial functions from the set N of natural numbers to itself are recursive. This approach led to a number of technical successes [66, 67]. RUSS does not use any of Brouwer’s nonlogical intuitionistic principles; indeed, it could not, since it produces results that are false if interpreted directly within INT. For example, in RUSS there exists a continuous real-valued map on [0, 1] that is not uniformly continuous; more dramatically, there exists a uniformly continuous map f from [0, 1] onto (0, 1] that has inﬁmum equal to 0. Once again, one should not overreact to the apparent conﬂict with classical mathematics: the last of these results should really be interpreted as saying that there exists a recursively uniformly continuous recursive function f from the closed interval [0, 1] of the recursive real line onto the recursive interval (0, 1] that has inﬁmum equal to 0. Put this way, the result does not conﬂict with CLASS; indeed, it is a result of CLASS, since the proof within RUSS is actually a proof within CLASS that does not use such nonconstructive logical principles as LEM. By the mid-1960s, constructive mathematics was, when compared with its classical counterpart, virtually stagnant. The situation changed in 1967 with the publication of Errett Bishop’s monograph Foundations of Constructive Mathematics [9]. This book and its oﬀspring [12] represent the most far-reaching and systematic presentation of constructive analysis to date. In [9], Bishop revealed, by thoroughgoing constructive means but without resorting to either Brouwer’s principles or the formalism of recursive function theory, a vast panorama of constructive mathematics, covering elementary analysis, metric and normed spaces, abstract measure and integration, the spectral theory of selfadjoint operators on a Hilbert space, Haar measure, duality on locally compact groups, and Banach algebras. Bishop’s constructive mathematics (BISH) was founded on a primitive, unspeciﬁed notion of “algorithm”, or “ﬁnite routine”, and the Peano properties of natural numbers, and kept strictly to the interpretation of “existence ” as “computability”. His refusal to pin down the notion of algorithm led to criticism, particularly from philosophers of mathematics and from those committed to the Church–Markov–Turing thesis; but this very imprecision enabled Bishop’s work to have a variety of interpretations: his results are valid in CLASS, INT, RUSS, and all reasonable models of computable mathematics, such as the more recent one propounded by Weihrauch [91]. Indeed, from a purely formal viewpoint, each of INT, RUSS, and CLASS can be regarded as BISH plus some additional principles: INT can be regarded as BISH supplemented by Brouwer’s continuity principle and fan theorem; RUSS as BISH plus the Church–Markov–Turing thesis; and CLASS as BISH plus the law of excluded middle. One consequence of this multiplicity of interpretations is that we can often demonstrate that certain propositions P are independent of BISH; that is, neither P nor (not P ) can be proved within BISH. For example, since “every mapping from [0, 1] into R is uniformly continuous” is a theorem of INT, and “there exists a continuous map of [0, 1] into R that is not uniformly continuous” is a theorem of

1.3 Intuitionistic Logic

7

RUSS, and since both INT and RUSS are formally consistent with BISH, within BISH we cannot expect either to prove that every continuous map of [0, 1] into R is uniformly continuous or to construct an example of a real-valued function that is deﬁned, but not uniformly continuous, on [0, 1]. Over the years since the publication of Bishop’s book, it became clear to a number of researchers that, in essence, BISH is simply mathematics with intuitionistic logic together with some appropriate set-theoretic foundation. As we pointed out at the end of the preceding section, working with intuitionistic logic automatically bars noncomputational steps. As long as we keep strictly to intuitionistic logic, having made sure that our set-theoretic principles do not inadvertently imply LEM or some other nonconstructive proposition, the mathematics we develop turns out to be predictive, in the sense that every proof implicitly shows that if we perform certain calculations, we shall achieve certain results. Accordingly, when we speak of “constructive mathematics” or “BISH” in future, we shall mean “mathematics with intuitionistic logic”. It therefore behooves us to explain more clearly exactly what intuitionistic logic is.

1.3 Intuitionistic Logic

The meaning doesn’t matter if it’s only idle chatter of a transcendental kind. —W.S. Gilbert, Patience Everywhere one seeks to produce meaning, to make the world signify, to render it visible. —Jean Baudrillard, Seduction, or the Superﬁcial Abyss

For Brouwer, mathematics took precedence over logic. In order to describe the logic used by the (intuitionist) mathematician, it was necessary ﬁrst to analyse the mathematical processes of the mind, from which analysis the logic could be extracted. In 1930, Brouwer’s most famous pupil, Arend Heyting (1898–1980), published a set of formal axioms that so clearly characterise the logic used by the intuitionist that they have become universally known as the axioms for intuitionistic logic [52]. These axioms capture the so-called BHK interpretation of the connectives ∨ (or), ∧ (and), =⇒ (implies), ¬ (not) and quantiﬁers ∃ (there exists), which we now outline.

∀ (for all/each),

8

1 Introduction to Constructive Mathematics

P ∨ Q : either we have a proof of P or else we have a proof of Q. P ∧ Q : we have both a proof of P and a proof of Q. P =⇒ Q : by means of an algorithm we can convert any proof of P into a proof of Q. ¬P : assuming P, we can derive a contradiction (such as 0 = 1); equivalently, we can prove (P =⇒ (0 = 1)) . ∃x P (x) : we have (i) an algorithm that computes a certain object x, and (ii) an algorithm that, using the information supplied by the application of algorithm (i), demonstrates that P (x) holds. ∀x ∈ A P (x) : we have an algorithm that, applied to an object x and a proof that x ∈ A, demonstrates that P (x) holds. Note that in the interpretation of the statement ∀x ∈ A P (x), the proof of P (x) will normally use both the data describing the object x and the information supplied by a proof that x belongs to the set A. This is an important point, since upon it hinges a key argument against the use of the axiom of choice in constructive mathematics. We shall return to this matter later. A property P (x) is said to be decidable if for each x to which it might be applicable we have P (x) ∨ ¬P (x), where the disjunction and negation are given their BHK interpretations. Even for a decidable property P (n) of natural numbers n the property ∀n P (n) ∨ ¬∀n P (n), and hence a fortiori LEM, will not hold in general. As a result, many classical results cannot be proved constructively, since they would imply LEM or perhaps some other manifestly nonconstructive principle. To illustrate this point, consider the following simple statement, the limited principle of omniscience (LPO): ∀a ∈ {0, 1}

N+

(a = 0 ∨ a = 0) , N+

where a = (a1 , a2 , . . .) , N+ = {1, 2, . . .} is the set of positive integers, {0, 1} the set of all binary sequences, and

is

a = 0 ⇐⇒ ∀n (an = 0) , a = 0 ⇐⇒ ∃n (an = 1) . In words, LPO states that for each binary sequence (an )n1 , either an = 0 for all n or else there exists n such that an = 1. Of course, this is a triviality from the

1.3 Intuitionistic Logic

9

viewpoint of classical logic. But its BHK interpretation is not so simple: it says that there is an algorithm that, applied to any binary sequence a, either veriﬁes that all the terms of the sequence are 0 or else computes the index of a term equal to 1. Anyone familiar with computers ought to be highly sceptical about such an algorithm, since in the case a = 0 it would normally need to test each of the inﬁnitely many terms an in order to come up with the correct decision. For such reasons we feel justiﬁed in not accepting LPO, or any classical proposition that constructively implies LPO, as a valid principle of constructive mathematics. But we have another reason for not doing so: it can be shown that there are models of Heyting arithmetic—Peano arithmetic with intuitionistic logic—in which LPO is false; so LPO cannot be derived in Heyting arithmetic (see [34, 48]). Since LPO is a special case of the law of excluded middle, we are led, in turn, to renounce the latter when working constructively. Similar informal analyses lead us to exclude both the classical rule ¬¬P =⇒ P, which forms the basis of proof by contradiction, and the following lesser limited principle of omniscience (LLPO), which is easily seen to be a consequence of LPO: For each binary sequence a with at most one term equal to 1 (in the sense that am an = 0 for all distinct m and n), either a2n = 0 for all n or else a2n+1 = 0 for all n. The exclusion of such principles from constructive mathematics has serious consequences for mathematical practice. For example, we cannot hope to prove constructively the simple statement ∀x ∈ R (x = 0 ∨ x = 0) ,

(1.1)

where R denotes the set of real numbers and x = 0 means that we can compute a rational number strictly between 0 and x (which, as we shall see when we deal with the real numbers more formally in Chapter 2, is not the same, constructively, as proving that ¬ (x = 0)). To prove this, consider any binary sequence a, and use it to deﬁne the binary expansion of a real number x=

∞

an 2−n .

n=1

If x = 0, then a = 0. If x = 0, we can compute a positive integer N such that x > 2−N =

∞

2−n ;

n=N +1

it is then clear that, by testing the terms a1 , . . . , aN , we can ﬁnd n N such that an = 1. Thus statement (1.1) about real numbers implies LPO and is therefore essentially nonconstructive.

10

1 Introduction to Constructive Mathematics

If the binary sequence a has at most one term equal to 1, then we can use the real number ∞ n (−1) an 2−n n=1

to show that the statement ∀x ∈ R (x 0 ∨ x 0) implies LLPO. The following elementary classical statements also turn out to be nonconstructive. Each real number x is either rational or irrational (that is, x = r for each rational number r). To see this, consider x=

∞ 1 − an , n! n=1

where a is any increasing binary sequence (that is a binary sequence such that an an+1 for each n). Each real number x has a binary expansion. Note that the standard intervalhalving argument for “constructing” binary expansions does not work, since we cannot necessarily decide, for a given number x between 0 and 1, whether x 1/2 or x 1/2. In fact, the existence of binary expansions is equivalent to LLPO. The intermediate value theorem, which is equivalent to LLPO. For all x, y ∈ R, if xy = 0, then either x = 0 or y = 0. The constructive failure of this proposition clearly has implications for the theory of integral domains. We emphasise here that classically valid statements like “each real number is either rational or irrational” that imply omniscience principles are not false in constructive mathematics; they cannot be, since BISH is consistent with CLASS. One principle whose constructive status is controversial is Markov’s principle (MP): N+ (¬ (a = 0) =⇒ a = 0) ; ∀a ∈ {0, 1} in words, for any binary sequence a, if it is impossible for all the terms to equal 0, then there exists a term equal to 1. In order to accept this as a principle of constructive mathematics, you have to be convinced that the information conveyed by the antecedent ¬ (a = 0) is suﬃcient to enable us to compute an index n with an = 1. The argument in favour of MP says that we can carry out this computation by searching systematically through the terms an , since the hypothesis ¬ (a = 0)

1.4 Informal Set Theory

11

guarantees that we shall eventually stumble across a term equal to 1. The counterargument is that the antecedent provides us with no prior bound for such a search— it does not tell us how many terms we need to test before we arrive at one equal to 1—so the search might go on longer than the remaining life of the universe before it produced the desired result. Moreover, Markov’s principle, like LPO, cannot be proved within Heyting arithmetic. For these reasons, we shall follow the normal practice of excluding MP from the working principles of constructive mathematics. As a consequence we exclude the even stronger logical principle (∀x ∈ A (P (x) ∨ ¬P (x)) ∧ ¬∀x ∈ A ¬P (x)) =⇒ ∃x ∈ A P (x),

(1.2)

where A is a well-deﬁned set (the exact meaning of “well-deﬁned set” will become clear in the next section). In fact, even if we were to accept Markov’s principle on the grounds that an unbounded search through the natural numbers that cannot fail to terminate must eventually do so, we would balk at accepting (1.2), since for a general set A there will be no natural order allowing us to search systematically in the way we can with N. An example of the type dealt with earlier, in which a classically valid proposition P is shown constructively to entail an essentially nonconstructive principle like LEM, LPO, LLPO, or even MP, is called a Brouwerian counterexample to P (even though it is not a counterexample in the true sense of the word; it is merely an indication that P does not admit of constructive proof). There is another expression ∞ an 2−n that we may use in this context. For example, we refer to the number x = n=1

that we constructed from a given binary sequence (an )n1 and then used to show that (1.1) implies LPO as a Brouwerian example of a real number x for which we cannot decide whether x = 0 or x = 0.

1.4 Informal Set Theory Die ganzen Zahlen hat der liebe Gott gemacht, alles andere ist Menschenwerk. —L. Kronecker [90] The primary concern of mathematics is number, and this means the positive integers. We feel about number the way Kant felt about space. The positive integers and their arithmetic are presupposed by the very nature of our intelligence and, we are tempted to believe, by the very nature of intelligence in general. The development of the positive integers from the primitive concept of the unit, the concept of adjoining a unit, and the process of mathematical induction carries complete conviction. In the words of Kronecker, the positive integers were created by God. —Errett Bishop [9]

Building on the set of positive integers and using intuitionistic logic, we follow Bishop’s approach to developing constructive mathematics at higher and higher

12

1 Introduction to Constructive Mathematics

levels of abstraction. To do this, we need to clarify notions such as “set” and “function”. For us, a set X (other than our basic, primary set N+ of positive integers) is given by two pieces of data: a property that enables members of X to be constructed using objects that have already been constructed (note this last phrase, which rules out the possibility of impredicative deﬁnitions and therefore of Russell-type paradoxes), and an equivalence relation =X of equality between members of X. We write x ∈ A to signify that x is an element of the set A, and x ∈ / A instead of ¬ (x ∈ A) . The use of equivalence relations rather than intensional equality—that is, identity of description—is common, but often goes unnoticed, in classical mathematics. For example, we call the rational numbers 1/2 and 3/6 equal, even though, strictly speaking, they are equivalent and not intensionally identical. A subset S of a set X consists of a collection of elements drawn from X, together with the equality relation induced on S by the given equality on X; that is, for elements x, y of S, we deﬁne x =S y ⇐⇒ x =X y. We write S ⊂ T to signify that S is a subset of T. If P (x) is a property applicable to certain elements x of a set A, then we denote by {x : x ∈ A ∧ P (x)} or {x ∈ A : P (x)} the subset of A consisting of those elements x of A with the property P (x). The logical complement of a subset S of X is ¬S = {x ∈ X : x ∈ / S} . A particular example of this is the empty subset of X, deﬁned by ∅X = ¬X. We say that a subset S of X is inhabited if ∃x (x ∈ S) . We then write S = ∅X . Note that in order to show that S is inhabited, we cannot just prove that it is impossible for S to be empty; we must actually construct an element of S; see Exercise 2.

1.4 Informal Set Theory

13

Two sets X, Y are said to be equal sets if each is a subset of the other; in other words, if the sets have the same elements and the same equality relation. We need to be careful when constructing new sets from old. Since an equality is part of the data for a set, it does not make sense to talk of the union S ∪ T of two sets unless we can put together not only the sets as collections of objects but also, in some way, their given equality relations. In practice, this means that in order to construct their union S ∪ T, the sets S and T must be given as subsets of some set X. We then deﬁne S ∪ T = {x ∈ X : x ∈ S ∨ x ∈ T }, where the equality on S ∪ T is that induced by X. Likewise, the intersection of S and T is deﬁned only when S and T are subsets of some set X, and is then the subset S ∩ T = {x ∈ X : x ∈ S ∧ x ∈ T } of X. The (Cartesian) product of two sets X, Y is the set X × Y consisting of all ordered pairs (x, y) with x ∈ X and y ∈ Y, together with the equality given by ((x, y) =X×Y (x , y )) ⇐⇒ (x =X x ∧ y =Y y ) . In many situations—even, as we shall see, on the real line—we frequently need a set X to be equipped with an inequality relation =X describing what it means for two elements of X to be unequal, or distinct. Such a relation must satisfy the following two properties: x =X y =⇒ ¬ (x =X y) , x =X y =⇒ y =X x. If, in addition, ¬ (x =X y) =⇒ x =X y, we say that the inequality is tight. A set X with an inequality is discrete if, for any two elements x and y of X, either x =X y or x =X y; we then also describe the inequality itself as discrete. One inequality relation, the denial inequality, is deﬁned by setting x =X y if and only if ¬ (x =X y) . This inequality is normally too weak for practical purposes. For example, in the absence of Markov’s principle, on the real line R the property ¬ (x =R 0) is weaker than |x| > 0 (Exercise 3); for that reason we deﬁne the standard inequality on R to be not the denial inequality but the one given by x =R y ⇐⇒ |x − y| > 0. From now on, when the meaning is clear from the context, we write =, ∅, =, . . .

14

1 Introduction to Constructive Mathematics

rather than =X , ∅X , =X , . . . . A subset S of a set X with an inequality has a complement, deﬁned by ∼S = {x ∈ X : ∀s ∈ S (x = s)} . Then ∼S ⊂ ¬S; but unless the inequality on X is the denial inequality, the reverse inclusion will not hold. A subset S of a set X is said to be detachable (from, or in, X) if ∀x ∈ X (x ∈ S ∨ x ∈ / S) . Since statement (1.1) implies LPO, not even the singleton subset {0} is detachable from R. However, {0} is detachable in the set Q of rational numbers. When X comes with an inequality relation, we deﬁne the inequality on any subset S of X to be the one induced by that on X: x =S y ⇐⇒ x =X y. If also Y has an inequality relation, we deﬁne the inequality on the Cartesian product X × Y by ((x, y) =X×Y (x , y )) ⇐⇒ (x =X x ∨ y =Y y ) . It should be no surprise that we require functions to be given by algorithms and to respect equality. Thus a function f from a set X to a set Y —also called a map or mapping of X into Y , and written f : X −→ Y —is an algorithm that, applied to any element x of X, produces an element f (x) of Y such that f is extensional : ∀x ∈ X ∀x ∈ X (x =X x =⇒ f (x) =Y f (x )) . The element f (x) is called the value of f at x or the image of x under f . If X and Y have inequality relations, then we may require f to be strongly extensional : ∀x ∈ X ∀x ∈ X (f (x) =Y f (x ) =⇒ x =X x ) . Note that the statement “all functions from R to R are strongly extensional” is equivalent to Markov’s principle. Let f, g be two real-valued functions on a set X with an inequality. We say that an element x0 of X is the strongly unique element of X such that f (x) = g(x) if f (x0 ) = g(x0 ) and f (x) = g(x) whenever x ∈ X and x = x0 . Strong uniqueness will resurface in Chapter 4 in connection with best approximations. A partial function f : X −→ Y is a function from a subset of X into Y. The subset

1.4 Informal Set Theory

15

{x ∈ X : f (x) is deﬁned} is called the domain of f, denoted by dom (f ) ; and the set {y ∈ Y : ∃x ∈ X (y =Y f (x))} the range of f, denoted by ran (f ) . The image of a subset A of X under f is the set f (A) = {f (x) : x ∈ A}. The inverse image of a subset B of Y under f is the set f −1 (B) = {x ∈ dom (f ) : f (x) ∈ B} . A partial function f : X −→ Y is said to be a total partial function on X if dom(f ) = X. When the expression describing f (x) is given explicitly and the domain of the partial function f is clearly understood, we may denote the function by x f (x). For example, the partial function from R to itself whose value is deﬁned, for each real number x such that x =R 0, to be 1/x may be written x 1/x. An important type of total partial function is deﬁned as follows. Let X and I be sets. A family of elements of X with index set I (or indexed by I) is a mapping i xi of I into X; we commonly denote this family by (xi )i∈I . In particular, if I is N+ , the family is called a sequence in X and is usually written (xn )n1 . More general sequences of the form (xn )nN have the obvious analogous meaning when N is an integer. Let f and g be mappings from subsets of a set X into a set Y, where Y is equipped with a binary operation 3. We introduce the corresponding pointwise operation 3 on f and g by setting (f 3g)(x) = f (x)3g(x) whenever f (x) and g(x) are both deﬁned. Thus, taking Y = R, we see that the (pointwise) sum of f and g is given by (f + g)(x) = f (x) + g(x) if f (x) and g(x) are both deﬁned; and that the (pointwise) quotient of f and g is given by ( f / g)(x) = f (x)/ g(x) if f (x) and g(x) are deﬁned and g(x) = 0. If X = N+ , so that f = (xn )n1 and g = (yn )n1 are sequences, then we also speak of termwise operations; for example, the termwise product of f and g is the sequence (xn yn )n1 . Pointwise operations extend in the obvious ways to ﬁnitely many functions. In the case of a sequence (fn )n1 of functions with values in a normed space (see

16

1 Introduction to Constructive Mathematics

Chapter 2), once we have introduced the notion of a series in a normed space we ∞ shall interpret fn in the obvious pointwise way. n=1

A partial function f : X −→ Y can be identiﬁed with its graph, G (f ) = {(x, y) ∈ X × Y : x ∈ dom (f ) , y ∈ ran (f ) , y = f (x)}, a subset of the Cartesian product X×Y. We deﬁne two partial functions f : X −→ Y and g : X −→ Y to be equal if their graphs are equal as sets. Thus f, g are equal partial functions if and only if their domains are equal subsets of X and f (x) = g(x) for each x ∈ dom(f ). We say that a partial function f : X −→ Y is one-one if f (x) = f (x ) entails x = x ; injective if X and Y are equipped with inequality relations, and x = x entails f (x) = f (x ). If f is injective and the inequality on X is tight, then f is one-one: for in that case, if f (x) = f (x ), then ¬ (x = x ) and so, by tightness, x = x . The composition, or composite, of partial functions f : A −→ B and g : B −→ C is the partial function g ◦ f : A −→ C (sometimes written gf ) deﬁned by g ◦ f (x) = g(f (x)) wherever the right side exists. A partial mapping f : X −→ Y maps its domain onto Y if ∀y ∈ Y ∃x ∈ X (y = f (x)) . On the other hand, we say that f : X −→ Y is an epimorphism if there exists a mapping g : Y −→ X such that ∀y ∈ Y (f (g(y)) = y) . A one-one partial function f : X −→ Y has a one-one inverse f −1 : ran (f ) −→ dom (f ) deﬁned by f −1 (f (x)) = x. If f is injective, then its inverse is strongly extensional. A bijection between X and Y is a one-one mapping from X onto Y . The subtle distinction between mappings onto and epimorphisms is closely linked to the constructive status of the axiom of choice, which we shall discuss shortly. First, though, we introduce some notions of cardinality. Let S be a subset of a set X with an inequality relation. We say that S is ﬁnitely enumerable if there exist a natural number N and a mapping of the set

1.4 Informal Set Theory

17

{1, 2, . . . , N } = {n ∈ N+ : n N } onto S; ﬁnite if there exist a natural number N and an injective map of {1, 2, . . . , N } onto S. When we speak of ﬁnitely many objects, we mean that those objects constitute an inhabited, ﬁnitely enumerable, but not necessarily ﬁnite, set. Note that the case N = 0 of the deﬁnition of “ﬁnitely enumerable” shows that the empty subset of X is ﬁnitely enumerable. Clearly, ﬁnite implies ﬁnitely enumerable; the converse does not hold constructively. We say that S is countable if there exist a detachable subset D of N+ and a mapping φ of D onto S. Every ﬁnitely enumerable subset—in particular, the empty subset—of X is countable. If D = N+ , we call φ an enumeration of S, in which case we may denote S by {φ(1), φ(2), . . .} . If also φ is one-one, we say that S is countably inﬁnite. A set is countably inﬁnite if and only if it is the range of a one-one mapping whose domain is a countably inﬁnite, detachable subset of N+ . We now consider the axiom of choice (AC): If X, Y are inhabited sets, S is a subset of X × Y, and for each x ∈ X there exists y ∈ Y such that (x, y) ∈ S, then there exists a choice function f : X −→ Y such that (x, f (x)) ∈ S for each x ∈ X. Under the BHK interpretation, the hypothesis ∀x ∈ X ∃y ∈ Y ((x, y) ∈ S) of AC means that we have an algorithm that, applied to each element x of X and the data showing that x belongs to X, constructs an element y of Y and demonstrates that (x, y) ∈ S. This much is clear. However, there is no guarantee that the algorithm will respect the equality relation on X—in other words, that if x =X x , and the algorithm constructs y, y in Y such that (x, y) ∈ S and (x , y ) ∈ S, then y =Y y . Indeed, we should expect that the computation of y might use data that are associated with properties intrinsic to x that do not apply intrinsically to x . For example, anticipating our development of the real number set R, consider the case in which X is R and Y is N+ . A real number is (deﬁned as) a certain set of rational approximations. However, two equal real numbers x, x can have diﬀerent deﬁning sets of rational approximations. In that case, the algorithm that computes a positive integer n such that (x, n) ∈ S may, and in general will, compute a diﬀerent positive integer n such that (x , n ) ∈ S. These considerations throw real doubt over the possibility that there is a choice function implementing the algorithm.

18

1 Introduction to Constructive Mathematics

In fact, an argument of Diaconescu [46] and Goodman & Myhill [50], but preﬁgured by Bishop (see Problem 2 on page 58 of [9]), shows that AC cannot be allowed as a principle of constructive mathematics. Theorem 1.4.1. The axiom of choice implies the law of excluded middle. Proof. Let P be any constructively meaningful statement, and deﬁne the set X to consist of the two elements 0 and 1, together with the equality relation such that (0 =X 1) ⇐⇒ P. Let Y be the set {0, 1} with the standard equality, and let S be the subset {(0, 0), (1, 1)} of the Cartesian product X × Y, taken with the standard equality. Suppose there exists a function f : X −→ Y such that (x, f (x)) ∈ S for all x ∈ X. There are three cases to consider: (i) f (0) = 1, (ii) f (1) = 0, and (iii) both f (0) = 0 and f (1) = 1. In case (i) we have (0, 1) = (0, f (0)) ∈ S, so either (0, 1) =X×Y (0, 0) or (0, 1) =X×Y (1, 1). If the ﬁrst of these two alternatives holds, then, by deﬁnition of the equality on X × Y, we have 1 =Y 0, which is absurd. Hence, in fact, (0, 1) =X×Y (1, 1) . Thus, again by deﬁnition of the equality on X × Y, we have 0 =X 1 and therefore P holds. Case (ii) similarly leads to the conclusion that P holds. Finally, in case (iii) we have ¬ (f (0) =Y f (1)) ; therefore, since f is a func2 tion, ¬ (0 =X 1) and so ¬P holds. Thus we have derived P ∨ ¬P from AC.

The axiom of choice will hold constructively if the set X is one for which no computation is necessary to demonstrate that an element belongs to it; Bishop calls such sets basic sets. Following the practice of most constructive mathematicians, we consider N+ , the set N = {0, 1, 2, 3, . . .} of natural numbers, and the set Z = {0, ±1, ±2, . . .} of all integers to be basic sets. This practice is reﬂected in our acceptance of the principle of countable choice: If Y is an inhabited set, S is a subset of N+ × Y, and for each positive integer n there exists y ∈ Y such that (n, y) ∈ S, then there is a function f : N+ −→ Y such that (n, f (n)) ∈ S for each n ∈ N+ . In fact, many constructive proofs use the stronger principle of dependent choice: If X is a set, a ∈ X, S is a subset of X × X, and for each x ∈ X there exists y ∈ X such that (x, y) ∈ S, then there exists a sequence (xn )n1 in X such that x1 = a and (xn , xn+1 ) ∈ S for each n ∈ N+ .

1.4 Informal Set Theory

19

Another contentious matter in constructive mathematics is the status and role of the power set P (X) of a given set X: that is, the collection of all subsets of X, with equality of subsets as deﬁned earlier. The main objection to admitting P (X) into the constructive fold is that we thereby allow impredicativity, since there is then nothing to stop us constructing subsets of X whose deﬁning characteristics are self-referential. On the other hand, nobody has yet shown that adding the power set axiom ∀X∃Y ∀S (S ⊂ X ⇐⇒ S ∈ Y ) to constructive mathematics enables us to prove LEM or some other incontestably nonconstructive principle. There are ways of avoiding the power set. It often suﬃces to work with the set Y X of all mappings from X into a set Y. Since we have a clear idea of what mappings from X into Y are (something we do not have for subsets of X), the set Y X seems relatively innocent. Note that classically the set {0, 1}X can be identiﬁed with the power set of X, since it comprises the characteristic functions of subsets of X. This identiﬁcation is not possible constructively, since characteristic functions exist only for those subsets of X that are detachable. Another way to avoid the full generality of the power set is to work with a welldeﬁned but smaller set of subsets of X. For example, the set of compact subsets of a metric space X is well deﬁned (it is actually a metric space itself), and is often all we need for many parts of analysis. Let S (X) be a well-deﬁned set of subsets of X, with two elements taken as equal if and only if they are equal sets in the usual sense, and let I be some set. Then we can speak sensibly about a family (Si )i∈I of elements of S (X) . We can also deﬁne the union and intersection of such a family to be, respectively, the subsets Si = {x ∈ X : ∃i ∈ I (x ∈ Si )} i∈I

and

Si = {x ∈ X : ∀i ∈ I (x ∈ Si )}

i∈I

of X. set of positive integers, we denote the above union and intersection If I is the Sn and Sn , respectively. by n1

n1

The Cartesian product

i∈I

Si is the subset of X I consisting of those functions f

∈ I; if X comes with an inequality relation, then the such that f (i) ∈ Si for each i

Si is deﬁned by corresponding inequality on i∈I

f = g ⇐⇒ ∃i ∈ I (f (i) = g(i)) . If I

is the set of positive integers, then (Si )i∈I is a sequence of sets, and the elements Si are also sequences; in this case, the sets Si can be arbitrary and need not of i∈I

20

1 Introduction to Constructive Mathematics

be subsets

of a previously deﬁned set X. If I = {1, 2, . . . , n}, then we denote the Si by S1 × S2 × · · · × Sn , and refer to its elements (x1 , x2 , . . . , xn ) as product i∈I

ordered n-tuples, or, in the case n = 2, ordered pairs (something we used informally in our earlier deﬁnition of the Cartesian product of two sets). Other set-theoretic notions will be introduced later as they arise. It is now time to turn our attention away from foundational matters to analysis proper.

Exercises Although we shall not formally deﬁne the real number line R until Chapter 2, in these exercises we assume elementary properties of real numbers where necessary. 1. Justify informally Brouwer’s observation that ¬¬¬P implies ¬P. Using this, show that the proposition (¬¬P =⇒ P ) is equivalent to the law of excluded middle. 2. Prove that if the statement ¬ (S = ∅) =⇒ S = ∅ applies to all subsets S of {0} , then the law of excluded middle holds. 3. Prove that ∀x ∈ R (¬ (x = 0) =⇒ x = 0) is equivalent to Markov’s principle. 4. Fill in the details of the proof that the statement ∀x ∈ R (x 0 ∨ x 0) implies LLPO. 5. Give a Brouwerian counterexample to the proposition that every real number x is either rational or irrational (where “x is irrational” means that x = r for each rational number r). 6. Prove that every real number has a binary expansion if and only if LLPO holds. 7. Give a Brouwerian counterexample to the statement that if r1 , r2 , and r3 are real roots of a quadratic polynomial x2 + ax + b with a, b ∈ R, then there exist distinct i, j with ri = rj .

1.4 Informal Set Theory

21

8. Prove that the statement “for all real numbers x, y that have binary expansions, the sum x + y has a binary expansion” is equivalent to LLPO. 9. Prove that the statement ∀x, y ∈ R (xy = 0 =⇒ (x = 0 ∨ y = 0)) is equivalent to LLPO. 10. Prove that the intermediate value theorem is equivalent to LLPO. 11. Let f : [a, b] −→ R be sequentially continuous in the following sense: for each sequence (xn )n1 in [a, b] that converges to a limit x, the sequence (f (xn ))n1 converges to f (x). Suppose also that f (a)f (b) < 0, and that f is locally nonzero in the sense that for each x ∈ [a, b] and each r > 0 there exists y ∈ [a, b] with |x − y| < r and f (y) = 0. Prove that there exists c ∈ (a, b) such that f (c) = 0. (This version of the intermediate value theorem suﬃces for virtually all constructive purposes.) 12. Prove that the statement “all functions from R to R are strongly extensional” is equivalent to Markov’s principle. 13. Discuss the statement “every mapping from a set X onto a set Y is an epimorphism”. 14. Give a Brouwerian counterexample to the statement “every subset of a ﬁnite set is ﬁnitely enumerable”. 15. Prove that an inhabited set is countably inﬁnite if and only if it is the range of a function with domain N. 16. Prove that a subset S of N is countable if and only if there exists a sequence (Sn )n1 of ﬁnite subsets of N such that S1 ⊂ S2 ⊂ S3 ⊂ · · · . 17. Prove that the statement “every inhabited subset of N is countable” implies the principle of ﬁnite possibility (PFP): to each binary sequence (an )n1 there corresponds a binary sequence (bn )n1 such that an = 0 for each n if and only if there exists N such that bN = 1. 18. Prove that the intersection of two countable sets is countable. 19. Prove that the principle of dependent choice implies the principle of countable choice. Prove also that the principle of dependent choice can be derived from the following principle of internal choice: if X is an inhabited set, S is a subset of X × X, and for each x ∈ X there exists y ∈ X such that (x, y) ∈ S, then there exists a choice function f : X −→ X such that (x, f (x)) ∈ S for each x ∈ X.

22

1 Introduction to Constructive Mathematics

20. By a ﬁlter on a set X we mean a set F of inhabited subsets of X with the following properties: • X ∈ F. • If S and T belong to F, then S ∩ T ∈ F. •

If S ∈ F and S ⊂ T ⊂ X, then T ∈ F.

A ﬁlter U is called an ultraﬁlter if for all S ⊂ X, either S ∈ U or ¬S ∈ U. The classical ultraﬁlter principle states that every ﬁlter is contained in an ultraﬁlter. Prove that this principle implies the weak limited principle of omniscience (WLPO): ∀a ∈ 2N (∀n (an = 0) ∨ ¬∀n (an = 0)) .

Notes The use of the Goldbach conjecture in the example in Section 1.1 is in no sense a constraint on the argument; if the Goldbach conjecture were solved tomorrow, we could easily replace it by any one of many open problems of number theory. Indeed, until Wiles proved the Fermat conjecture in 1994, it was common to use that, rather than the Goldbach conjecture, in such Brouwerian examples of nonconstructivity. The reader interested in the history of foundations of mathematics should consult [44] and [84] for more information on the “Grundlagenstreit” that eventuated between Brouwer and Hilbert in the 1920s. For more on Hilbert’s programme for the secure foundation of mathematics see [93]. The designation “BHK interpretation” comes from Brouwer, Heyting, and Kolmogorov. The BHK interpretation of implication, while more natural than the classical one of material implication, in which (P =⇒ Q) is equivalent to (¬P ∨ Q), has not completely satisﬁed all researchers using constructive logic. Shortly before he died, Bishop communicated to Bridges his dissatisfaction with the standard constructive interpretation of implication. Unfortunately, he left nothing more than very rudimentary sketches of his ideas for its improvement. (Note, however, Bishop’s paper [10].) For a deeper analysis of the constructive interpretations of the connectives and quantiﬁers, we refer the reader to [48]. The classical invalidity of the recursive interpretation of LPO is not a matter of logic: it can be demonstrated, even with classical logic, that a recursive version of LPO would lead to a proof of the decidability of the halting problem, which is known to be impossible; see [34], pages 52–53. Andrej Bauer has recently shown that proofs and results in BISH can be translated into Weihrauch’s Type-2 Eﬀectivity framework by a realisability interpretation [5].

1.4 Informal Set Theory

23

The principle (¬¬P =⇒ P ) is equivalent to LEM even with intuitionistic logic; see Exercise 1. For the conﬁrmed intuitionist there is at least one other reason for rejecting MP: it contradicts Brouwer’s (admittedly controversial) theory of the creating subject, an add-on to his intuitionistic mathematics. See [34] (pages 116–117) and, for a rather diﬀerent view, [73]. Bishop used “subﬁnite” instead of “ﬁnitely enumerable”. Note that a subset of a ﬁnite set need not be ﬁnitely enumerable. In the proof of Theorem 1.4.1 we could have deﬁned X as a set of equivalence classes under the equivalence relation ∼ deﬁned by (0 ∼ 1) ⇐⇒ P, but it is more in keeping with Bishop’s approach to proceed by considering a special, if unusual, equality relation on X. Some constructive mathematicians argue against even the principle of countable choice; see, for example, [76, 82]. Myhill has shown that, under the Church–Markov–Turing thesis, the power set of a singleton is uncountable: in other words, there is no recursive mapping of N onto that power set [74] (page 364, Theorem 3). Within BISH we may not be able to prove the uncountability of the power set of {0} , but we certainly have an unending supply of subsets S of {0} for which we cannot decide that S = ∅ or S = {0}: given any constructively meaningful statement P, we can deﬁne a corresponding set S = {x : x = 0 ∧ P } such that (S = {0} ∨ S = ∅) if and only if P ∨ ¬P. Myhill’s formal theory—based on primitive notions of “set”, “function”, and “natural number”—is but one of several foundational theories advocated for constructive mathematics. Others include Aczel’s constructive set theory [3], MartinL¨ of’s type theory [69], and a largely unpublished constructive version of Morse’s set theory [16] in which membership of the universe appears to correspond to being constructively deﬁned. For more on constructive foundational theories see [6] and [88].

2 Techniques of Elementary Analysis

My deplorable mania for analysis exhausts me. —Gustave Flaubert, Letter (August 1846) We begin by using a form of interval arithmetic as a foundation for the construction of the real number line R. Having discussed the elementary algebraic and order-theoretic properties of real numbers, we prove that R is complete in two senses: Dedekind (order) complete and Cauchy (sequentially) complete. The next step is to generalise from the reals to metric and normed spaces. A particularly important property for us is total boundedness, which plays a big part in proving the existence of suprema and inﬁma; for that reason we need to prove that there are lots of totally bounded subsets in a totally bounded space. We also highlight the important property of locatedness for subsets of a metric space, a property that holds automatically in classical mathematics.

2.1 The Real Number Line Constructive analysis proper, as distinct from arithmetic, begins with the real numbers, which we shall construct, choice-free, by interval arithmetic. First, though, we observe that the purely algebraic constructions of the sets Z of integers and Q of rational numbers from N are carried out as in classical algebra, and that the standard inequality on Q —and hence a fortiori on Z regarded as a subset of Q —is the (in this case discrete) denial inequality. By a real number we mean a subset x of Q × Q such that for all (q, q ) in x, q q ; for all (q, q ) and (r, r ) in x, the closed intervals [q, q ] and [r, r ] in Q intersect in points of Q;

26

2 Techniques of Elementary Analysis

for each positive rational ε there exists (q, q ) in x such that q − q < ε. The last of these properties ensures that x is inhabited. It also legitimises our use of expressions like “pick an element of x”: to carry out such a selection, we simply take ε = 1 in the third of the deﬁning properties of the real number x, to produce a corresponding element (q, q ) of x. The intuition underlying our deﬁnition of “real number” is that the elements of x are the rational endpoints of closed intervals with one point—namely x —in common. Any rational number q gives rise to a canonical real number q = {(q, q)} with which the original rational q is identiﬁed. Two real numbers x and y are •

equal, written x = y, if for all (q, q ) ∈ x and all (r, r ) ∈ y, the intervals [q, q ] and [r, r ] in Q have a rational point in common;

•

unequal (or distinct), written x = y, if there exist (q, q ) ∈ x and (r, r ) ∈ y such that the intervals [q, q ] and [r, r ] in Q are disjoint.

It is almost immediate that = satisﬁes the deﬁning properties of an inequality relation. Let us check that equality is an equivalence relation. It is trivial that it is reﬂexive and symmetric, so only transitivity has to be handled. Let x = y and y = z, and suppose that for some (q, q ) ∈ x and (r, r ) ∈ z there are no rational points in [q, q ]∩[r, r ] . We may assume without loss of generality that q < r. Using the third of the deﬁning properties for a real number, choose (s, s ) ∈ y such that s − s < r − q . Now, the rational interval [s, s ] intersects [q, q ] in a rational point u, and [r, r ] in a rational point v, so r − q v − u s − s < r − q , a contradiction. Hence

¬ ([q, q ] ∩ [r, r ] = ∅) .

Since we are working with intervals in Q, with the aid of two simple lemmas we can turn this around to construct a point of [q, q ] ∩ [r, r ] , and therefore complete the proof that x = z, as follows. Lemma 2.1.1. Let a, b, c, d be rational numbers with a b and c d. Then there exists a rational number in [a, b] ∩ [c, d] if and only if a d and c b. Proof. If r is a rational number in the intersection of the intervals, then a r d and c r b, so a d and c b. If, conversely, these conditions hold, then either c < a and therefore a ∈ [a, b] ∩ [c, d] , or else a c and c ∈ [a, b] ∩ [c, d] . 2

2.1 The Real Number Line

27

Lemma 2.1.2. Let I, J be closed, bounded intervals in Q such that ¬ (I ∩ J = ∅) . Then there exists r ∈ Q such that r ∈ I ∩ J. Proof. Let I = [a, b] and J = [c, d] . If b < c, then I ∩ J = ∅, a contradiction. Hence c b. Likewise, a d. It remains to apply Lemma 2.1.1. 2

Taken with the equality and inequality we have deﬁned above, the collection of real numbers forms a set: the real line R. Let x, y be real numbers. We say that x is greater than y, and that y is less than x, if there exist (q, q ) ∈ x and (r, r ) ∈ y such that r < q; we then write x > y or, equivalently, y < x. On the other hand, we say that x is greater than or equal to y, and that y is less than or equal to x, if for all (q, q ) ∈ x and all (r, r ) ∈ y we have q r; we then write x y or, equivalently, y x. We write, for example, x > y to indicate that ¬ (x > y) , and x y to indicate that ¬ (x y) . Clearly, x x and x > x. Moreover, by Lemma 2.1.1, x = y if and only if both x y and y x; and x = y if and only if either x > y or else x < y. Lemma 2.1.3. If x > y, then y > x. Proof. Since x > y, there exist (q, q ) ∈ x and (r, r ) ∈ y such that r < q. If also y > x, then there exist (s, s ) ∈ x and (t, t ) ∈ y such that s < t. By the deﬁning conditions on real numbers, there exist rational numbers a, b such that a ∈ [q, q ] ∩ [s, s ] and b ∈ [r, r ] ∩ [t, t ] . But then a s < t b r < q a, a contradiction.

2

Lemma 2.1.4. x y if and only if y > x. Proof. By deﬁnition, x y if and only if for all (q, q ) ∈ x and all (r, r ) ∈ y we 2 have q r, which occurs precisely when ¬(y > x). Lemma 2.1.5. If x > y, then x y. Proof. By Lemma 2.1.3, y > x. The result follows from Lemma 2.1.4.

2

Lemma 2.1.6. If x > 0, then there exists a positive integer n such that x > 1/n.

28

2 Techniques of Elementary Analysis

Proof. First pick (q, q ) ∈ x such that 0 < q. Then choose a positive integer n such that q > 1/n. The deﬁnition of “greater than” ensures that x > 1/n. 2 From time to time we shall revisit some of our earlier Brouwerian examples in the light of our formal deﬁnitions of real numbers and their properties. Proposition 2.1.7. The statement ∀x, y ∈ R (¬ (x y) =⇒ y > x)

(2.1)

implies Markov’s principle. Proof. Assume (2.1). Let (an )n1 be an increasing binary sequence such that ¬∀n (an = 0) , and deﬁne a real number by 1 1 1 , : an = 0 ∪ : an = 1 − an−1 . x= 0, n n n

(2.2)

Then ¬ (0 x): for if 0 q for all (q, q ) ∈ x, then an = 0 for all n, a contradiction. It follows from (2.1) that

1 1x > 0 and hence that there exists (q, q ) ∈ x such that 2 q > 0. Then (q, q ) = n , n for a (unique) n such that an = 1 − an−1 . Proposition 2.1.8. The relation > is cotransitive: If a < b, then for all x ∈ R either a < x or x < b. Proof. There exist (q, q ) ∈ a and (r, r ) ∈ b such that q < r. Given a real number x, we can ﬁnd (s, s ) ∈ x such that s − s < r − q . If s < r, then x < b; if s r, 2 then q < s and so a < x. Lemma 2.1.9. If x > y z or x y > z, then x > z and x z. Proof. Assume, for example, that x > y z. By Proposition 2.1.8, either x > z or z > y. Since the latter is ruled out by Lemma 2.1.4, we have x > z and therefore, by Lemma 2.1.5, x z. 2 Lemma 2.1.10. If (q, q ) ∈ x, then q x q . Proof. By Lemma 2.1.1, for each (r, r ) ∈ x, since the rational intervals [r, r ] and [q, q ] intersect, r q and q r. The desired conclusion now follows from the deﬁnition of the relation . 2

2.1 The Real Number Line

29

Lemma 2.1.11. For each real number x there exist rational numbers q, q such that q < x < q . Proof. Let (r, r ) be any element of x. By Lemma 2.1.10, r x r . Choosing q, q in Q with q < r r < q , we see from the deﬁnition of the relation < that 2 q < x < q .

The following two propositions show that the classical law of trichotomy does not hold constructively, even in the weak form discussed in Proposition 2.1.13. Proposition 2.1.12. The statement ∀x ∈ R (x 0 =⇒ x > 0 ∨ x = 0) implies LPO. Proof. Given an increasing binary sequence (an )n1 , deﬁne the real number x as at (2.2). It is routine to check that x 0; that if x = 0, then an = 0 for all n; and 2 that if x > 0, then there exists n such that an = 1. Proposition 2.1.13. The statement ∀x ∈ R (x 0 ∨ x 0) implies LLPO. Proof. Let (an )n1 be a binary sequence with at most one term equal to 1, and deﬁne a real number by 1 1 n 1 n 1 , (−1) : ∀k n (ak = 0) ∪ : an = 1 . (−1) x= − , n n n n If x 0, then it is impossible that an = 1 for an odd n, so an = 0 for all odd n. 2 Likewise, if x 0, then an = 0 for all even n. Lemma 2.1.14. If x < y, then there exists s ∈ Q such that x < s < y. Proof. Since x < y, there exist (q, q ) ∈ x and (r, r ) ∈ y such that q < r. Then the rational number 1 1 (q + r) , (q + r) s= 2 2 has the desired property.

2

30

2 Techniques of Elementary Analysis

The maximum of two real numbers x and y is the set max {x, y} of all rational pairs of the form (max {q, r} , max {q , r }) where (q, q ) ∈ x, (r, r ) ∈ y, and, for example, max {q, r} is the maximum in Q of the two rational numbers q, r. The minimum of x and y is the set min {x, y} of all rational pairs of the form (min {q, r} , min {q , r }) where (q, q ) ∈ x, (r, r ) ∈ y, and min {q , r } is the minimum in Q of the two rational numbers q , r . The maximum, max {x1 , . . . , xn } , and minimum, min {x1 , . . . , xn } , of ﬁnitely many real numbers are deﬁned analogously. We deﬁne the negative of a real number x to be

− x = (q, q ) ∈ Q × Q : − q ,− q ∈ x , where, for a rational number q, the expression − q denotes the negative deﬁned in the usual way. The absolute value of the real number x is the set |x| = max − x, x . It is not hard to verify such properties as the following: −

x, and |x| are real numbers.

•

max{x, y}, min{x, y},

•

max {x, y} < z if and only if x < z and y < z.

• max {x, y} > z if and only if either x > z or y > z. • min {x, y} < z if and only if either x < z or y < z. •

x < y if and only if

• max {x, y} =

−

−

y 0; whence x = 1 and P holds. In the other case we must have ¬P. We deﬁne a set S of real numbers to be upper order located if for all rational numbers a, b with a < b, either x b for all x in S or else there exists x ∈ S with a < x. We can now state the constructive least-upper-bound principle, which is also known as the Dedekind (order) completeness of R. Theorem 2.1.18. Let S be an inhabited set of real numbers that is both bounded above and upper order located. Then the least upper bound of S exists. Proof. Let B be the set of upper bounds for S, and deﬁne ξ = {(q, q ) ∈ Q × Q : ∃s ∈ S ∃b ∈ B (q s b q )} . Taken with Lemmas 2.1.10 and 2.1.9, the hypotheses ensure that ξ is inhabited. If (q, q ) and (r, r ) belong to ξ, then there exist s1 , s2 ∈ S and b1 , b2 ∈ B such that q s1 b1 q and r s2 b2 r . Then s1 b2 , so q b2 r ; similarly, r q . By Lemma 2.1.1, there exists a rational number in [q, q ] ∩ [r, r ] . To complete the proof that ξ is a real number, we show that for each rational ε > 0 there exists (q, q ) ∈ ξ with q − q < ε. To this end, ﬁx (a, a ) in ξ. If a = a , then there is nothing to prove; so we may assume that a < a . Construct rational numbers a0 = a < a1 < a2 < · · · < an = a such that ai − ai−1 < ε/2 for 1 i n. Since S is upper order located, either a2 ∈ B or else a1 < s for some element s of S. In the ﬁrst case, (a0 , a2 ) ∈ ξ and a2 − a0 < ε. In the second, either a3 ∈ B

2.1 The Real Number Line

33

and therefore (a1 , a3 ) ∈ ξ and a3 − a1 < ε; or else a2 < s for some element s of X. Carrying on in this way, since an ∈ B we can be sure of ﬁnding k < n − 1 such that (ak , ak+2 ) ∈ ξ and ak+2 − ak < ε. Thus ξ is indeed a real number. To show that ξ is an upper bound for S, consider any (q, q ) in ξ and any s in S. There exists b ∈ B such that q b q . For any (r, r ) in s we have r s, by Lemma 2.1.10, and therefore r b; whence r q . It follows that s ξ. Finally, if x < ξ, then we can ﬁnd (q, q ) ∈ ξ and (r, r ) ∈ x such that q > r . It follows from Lemma 2.1.10 that x r < q. By deﬁnition of ξ, there exists s ∈ S with q s; whence x < s, by Lemma 2.1.9. This completes the proof that ξ is the least upper bound for S. 2

Let S be a set of real numbers. We say that a real number b is a lower bound of/for S if b s for all s ∈ S; and that b is the (perforce unique) greatest lower bound of S if it is a lower bound and for each x > b there exists s ∈ S such that x > s. In the latter event we also call b the inﬁmum of S and we denote it by inf S. Corollary 2.1.19. Let S be an inhabited set of real numbers that is bounded below and is lower order located in the following sense: for all rational numbers a, b with a < b, either a is a lower bound for S or else there exists s ∈ S with s < b. Then the greatest lower bound of S exists. Proof. Apply Theorem 2.1.18 to the set T = −s : s ∈ S , which is inhabited, bounded above, and upper order located, to construct its supre2 mum b. Then − b is the inﬁmum of S.

We now want to introduce the arithmetic operations on real numbers. Given real numbers x and y, we need to ﬁnd the appropriate ways of combining the rational pairs (representing intervals) that constitute the real number x with those that constitute y, in order to create the rational pairs that represent x ◦ y, where ◦ stands for any of the operations +, −, ×, ÷. We begin with the easy deﬁnitions of + and −, for the moment leaving aside the more complicated ones for × and ÷. We deﬁne the sum x + y and diﬀerence x − y of the real numbers x, y to be, respectively, x + y = {(s, s ) : ∃ (q, q ) ∈ x ∃ (r, r ) ∈ y (s = q + r ∧ s = q + r )} , x − y = {(s, s ) : ∃ (q, q ) ∈ x ∃ (r, r ) ∈ y (s = q − r ∧ s = q − r )} . Let us verify, for example, that x+y is a real number. Let (q1 , q1 ) ∈ x and (r1 , r1 ) ∈ y. Certainly, q1 + r1 q1 + r1 . Moreover, given a positive rational number ε, we can arrange that q1 −q1 < ε/2 and r1 −r1 < ε/2, so (q1 + r1 )−(q1 + r1 ) < ε. It remains

34

2 Techniques of Elementary Analysis

to show that if also (q2 , q2 ) ∈ x and (r2 , r2 ) ∈ y, then the intervals [qi + ri , qi + ri ] (i = 1, 2) in Q intersect. This is easy: there exist rational numbers ξ, η such that qi ξ qi and ri η ri (i = 1, 2) ; whence qi + ri ξ + η qi + ri

(i = 1, 2) .

Thus x + y is a real number. Note that if x < y, then x + z < y + z for all real z; and that x − y = x+ (− y) . In view of the latter, we adopt the normal convention of writing −x instead of − x. Lemma 2.1.20. For each real number x there exists a positive integer N such that max{|q| , |q |} < N whenever (q, q ) ∈ x and q − q < 1. Proof. With n as in Lemma 2.1.15, set N = n + 1. If (q, q ) ∈ x and q − q < 1, then, using Lemma 2.1.10, we obtain q x |x| < n < N and therefore

q < q + 1 < n + 1 = N.

Since (−q , −q) ∈ −x and |−x| = |x| , we likewise have −q < N and −q < N . 2 Hence |q| < N and |q | < N, so max {|q| , |q |} < N. We next deﬁne the product x × y, also written x · y or xy, to be the set of all rational pairs (s, s ) such that there exist (q, q ) ∈ x and (r, r ) ∈ y with s = min {qr, qr , q r, q r } , s = max {qr, qr , q r, q r } . Certainly, s s . Using Lemma 2.1.20, compute a positive integer N such that if (q, q ) ∈ x and q − q < 1, then max {|q| , |q |} < N, and such that if (r, r ) ∈ y and r − r < 1, then max {|r| , |r |} < N. Given a rational ε with 0 < ε < 1, choose (q, q ) ∈ x and (r, r ) ∈ y such that q − q < ε/2N and r − r < ε/2N. Routine rational arithmetic calculations show that s − s < (q − q) max {|r| , |r |} + (r − r) max {|q| , |q |} < ε. For i = 1, 2 let (qi , qi ) ∈ x, (ri , ri ) ∈ y, and si = min {qi ri , qi ri , qi ri , qi ri } , si = max {qi ri , qi ri , qi ri , qi ri } . Pick ξ in [q1 , q1 ] ∩ [q2 , q2 ] and η in [r1 , r1 ] ∩ [r2 , r2 ] ; it is easy to verify that ξη ∈ [s1 , s1 ] ∩ [s2 , s2 ] . This completes the proof that xy is a real number. When dealing with division, we consider two real numbers x, y with y = 0. In the case y > 0 we construct a rational pair (s, s ) in the quotient x/y (also written

2.1 The Real Number Line

35

x y)

as follows. We pick (q, q ) ∈ x and (r, r ) ∈ y such that r > 0. If q 0, we set s = q/r , s = q /r; if q < 0, we set s = q/r, s = q /r . In the case y < 0 we deﬁne x/y to be −((−x)/y). We leave as an exercise the details that x/y is indeed a real number, and that x(1/x) = 1. As classically, we say that a sequence (xn )n1 of real numbers converges to a real number x∞ , called the limit of the sequence, if ∀ε > 0 ∃N ∀n N (|x∞ − xn | < ε) . We then write xn −→ x∞ as n −→ ∞ or x∞ = lim xn . n→∞

The uniqueness of the limit, and the basic algebraic properties of sequences of real numbers, will be assumed without proof, since the proofs are virtually the same as their classical counterparts. However, we should note that the monotone convergence theorem for sequences, Every increasing sequence in R that is bounded above (that is, whose terms form a set that is bounded above) converges, implies LPO. By a Cauchy sequence of real numbers we mean a sequence (xn )n1 such that ∀ε > 0 ∃N ∀m, n N (|xm − xn | < ε) . Every convergent sequence of real numbers is a Cauchy sequence. We now counter the second major misconception about the constructive nature of R by establishing its so-called (Cauchy) completeness. Theorem 2.1.21. Every Cauchy sequence of real numbers converges to a real number. Proof. Let (xn )n1 be a Cauchy sequence of real numbers, and, using countable choice, compute a function k nk from N+ to N+ such that

∀k ∀m, n nk |xm − xn | < 2−k . Again using countable choice, construct a sequence ((qk , qk ))k1 such that

∀k (qk , qk ) ∈ xnk ∧ qk − qk < 2−k . Deﬁne rational numbers rk = qk − 2−k ,

rk = qk + 2−k .

36

2 Techniques of Elementary Analysis

Then for all n nk , rk xnk − 2−k < xn < xnk + 2−k rk .

(2.3)

It follows that for all j k, xnj ∈ rj , rj ∩ [rk , rk ] . Since rk − rk < 2−k+2 −→ 0 as k −→ ∞, we conclude that x∞ = {(rk , rk ) : k 1} is a real number. From (2.3) we have xn ∈ [rk , rk ] for all n nk . It follows that

∀k ∀n nk |xn − x∞ | rk − rk < 2−k+2 . Noting Lemma 2.1.6, we see that xn −→ x∞ as n −→ ∞.

2

We shall see in later chapters that completeness is often used in constructive mathematics to prove propositions that are immediate consequences of omniscience principles. For the remainder of this book we shall drop the use of boldface type to denote real numbers; it has served its purpose to signal a distinction between a rational number and a real number (a set of special pairs of rational numbers). We shall also assume, without further comment, basic properties of real numbers—for example, x2 0 for all real x—that can easily be deduced from the foregoing results. The complex plane C consists of all complex numbers—ordered pairs (x, y) of real numbers—with addition and multiplication deﬁned by (x, y) + (x , y ) = (x + x , y + y ) , (x, y) × (x , y ) = (xx − yy , xy + x y) . The equality and inequality on C are deﬁned by (x, y) = (x , y ) ⇐⇒ x = x ∧ y = y (x, y) = (x , y ) ⇐⇒ x = x ∨ y = y . We embed R as a subset of C in the usual way by identifying the real number x with the complex number (x, 0) . The pair i = (0, 1) then has the special property that i2 = −1; and every complex number z = (x, y) can be written in the form x + iy, with real part Re z = x and imaginary part Im z = y. The complex conjugate of z = x + iy is the number x − iy, denoted by z ∗ . We shall assume basic deﬁnitions and properties of C, such as its completeness, as they are needed. The same goes for the Euclidean spaces Rn and Cn , which are now deﬁned in the standard ways.

2.2 Metric Spaces

37

2.2 Metric Spaces Since we are assuming some familiarity with the classical theory of metric spaces, in the following we shall emphasise diﬀerences between the classical and the constructive theory, as well as those constructive properties, such as locatedness, that play no role in classical analysis. We normally denote the metric on a set X by ρ. The deﬁnitions of such notions as metric, (metric) subspace, ball, interior, and cluster point are exactly as in classical analysis; see, for example, [47]. The interior of a subset S of a metric space X is denoted by S ◦ and is a subset of S; if S = S ◦ , then S is said to be open. The set S of cluster points of S is called the closure of S (in X), and contains S; if S = S, then S is said to be closed in X. If S = X, then S is dense in in X. This happens if and only if for each x ∈ X and each ε > 0 there exists s ∈ S such that ρ (x, s) < ε. We denote the open and closed balls in X with centre a and radius r > 0 by B(a, r) and B(a, r) respectively. The inequality on a metric space is deﬁned by x = y ⇐⇒ ρ(x, y) > 0. If X1 , . . . , Xn are metric spaces, with metrics ρ1 , . . . , ρn respectively, then the Cartesian product X = X1 × · · · × Xn is a metric space relative to the product metric ρ = ρ 1 + · · · + ρn . We then call (X, ρ) the product of the metric spaces X1 , . . . , Xn . As in the classical theory, both X and ∅ are open, arbitrary unions of open sets are open, and ﬁnite intersections of open sets are open; in other words, the open sets form a topology on X. Likewise, both X and ∅ are closed, and arbitrary intersections of closed sets are closed. But we cannot prove that the union of two closed sets is closed, even in the case X = R and the two closed sets are closed intervals: both the intervals [−1, 0] and [0, 1] are closed in R, and their union is dense in [−1, 1]; but if that union is closed, then we have ∀x ∈ R (x 0 ∨ x 0) , a proposition equivalent to LLPO. Proposition 2.2.1. The logical complement of an open subset S of X is closed in X, and coincides with the complement of S. Proof. Given a cluster point a of ¬S and any point x of S, choose, in turn, r > 0 such that B(x, r) ⊂ S, and y ∈ ¬S such that ρ(a, y) < r/2. If ρ(a, x) < r/2, then ρ(x, y) ρ(a, x) + ρ(a, y) < r, so y ∈ S, which is absurd. Hence, by Proposition 2.1.8, ρ(a, x) > 0. Since x ∈ S is arbitrary, it follows that a ∈ ∼S ⊂ ¬S. Thus ¬S is closed, and ¬S ⊂ ∼S ⊂ ¬S;

38

2 Techniques of Elementary Analysis

whence ¬S = ∼S.

2

In contrast to the classical situation, we cannot prove constructively that the complement of a closed subset of R is open (Exercise 9). The classical deﬁnitions of convergent sequence, Cauchy sequence, and complete metric space carry over unchanged into the constructive setting, as they did in the special case of the metric space R. A complete subset of a metric space X is closed in X; and a closed subset of a complete space is complete. Moreover, the product of ﬁnitely many complete spaces is complete. We now introduce the ﬁrst of two major themes of metric space theory. Given ε > 0, by an ε-approximation to a subset S of a metric space X we mean an inhabited subset T of S such that for each s ∈ S there exists t ∈ T with ρ (s, t) < ε. If for each ε > 0 there exists a ﬁnite ε-approximation to S, then we say that S is totally bounded. The closure of a totally bounded subset of X is totally bounded. If a subset S of X contains a dense totally bounded set, then S itself is totally bounded. The product of ﬁnitely many totally bounded spaces is totally bounded. There are at least two reasons why total boundedness is so important in constructive analysis. The ﬁrst is that with its help we can compute suprema and inﬁma in many important situations. The second is that, coupled with completeness, total boundedness gives the only one of three classically equivalent notions of compactness that can be used in constructive analysis. The next lemma enables us to replace “ﬁnite” by the weaker “ﬁnitely enumerable” in the deﬁnition of total boundedness. Lemma 2.2.2. Let ε be a positive number, and X a metric space with a ﬁnitely enumerable ε-approximation. Then X has a ﬁnite η-approximation for each η > ε. Proof. Let {x1 , . . . , xn } be an ε-approximation to X, and let η > ε. We may assume that n > 1. Either ρ(xi , xj ) > 0 whenever i = j, or ρ(xi , xj ) < η − ε for some pair of distinct indices i, j. In the former case, {x1 , . . . , xn } is a ﬁnite η-approximation. In the latter case we can delete xj from {x1 , . . . , xn }, to obtain a ﬁnitely enumerable ε -approximation with ε = ε + ρ(xi , xj ) < η. Applying this argument at most n − 1 times, we arrive at a ﬁnite η-approximation to X. 2 Corollary 2.2.3. A metric space X is totally bounded if and only if for each ε > 0 there exists a ﬁnitely enumerable ε-approximation to X. Proof. This is an immediate consequence of Lemma 2.2.2.

2

A metric space X is separable if it has a countable dense subset S; an enumeration (xn )n1 of S is then called a dense sequence in X. The real line is separable,

2.2 Metric Spaces

39

since, by Lemma 2.1.10 and the deﬁnition of “real number”, it has Q as a countable dense subset. Proposition 2.2.4. A totally bounded metric space is separable. Proof. If, for each positive integer n, Fn is a ﬁnite 1/n-approximation to the metric Fn is a countable dense subset of X. 2 space X, then n1

We now have what is among the most useful of all results about totally bounded sets. Proposition 2.2.5. If S is a totally bounded subset of R, then sup S and inf S exist. Proof. We ﬁrst consider the case where S = {x1 , . . . , xn } is ﬁnitely enumerable. Given real numbers a, b with a < b, we apply Proposition 2.1.8 repeatedly, to prove that either xk < b for each k or else there exists j such that xj > a. It follows from the constructive least-upper-bound principle that sup S exists. Now consider the general case. Again let a, b be real numbers with a < b, but this time write ε = (b − a)/2 and construct a ﬁnite ε-approximation {x1 , . . . , xn } to S. By the ﬁrst part of the proof, σ = sup {x1 , . . . , xn } exists. By Proposition 2.1.8, either σ > a or σ < a + ε. In the ﬁrst case there exists j such that xj > a. In the second, consider any x ∈ S. Choosing j such that |x − xj | < ε, we have x xj + |x − xj | < σ + ε < a + 2ε = b. So in this case, b is an upper bound for S. It follows from Theorem 2.1.18 that sup S exists. Similar arguments show that inf S exists. 2 A mapping f : X −→ Y between metric spaces is •

sequentially continuous at x ∈ X if for each sequence (xn )n1 that converges in X to x, the sequence (f (xn ))n1 converges in Y to f (x);

• sequentially continuous if it is sequentially continuous at each point of X; • continuous at x ∈ X if for each ε > 0 there exists δ > 0 such that ρ (f (x), f (x )) < ε for all x ∈ X with ρ (x, x ) < δ; • continuous if it is continuous at each point of X.

40

2 Techniques of Elementary Analysis

Although sequential continuity is classically equivalent to continuity, the most we can say constructively is that the latter notion implies the former. For constructive practice it is frequently necessary to assume the following stronger continuity property. We say that f : X −→ Y is uniformly continuous (on X) if for each ε > 0 there exists δ > 0 such that for all x, x ∈ X, if ρ (x, x ) < δ, then ρ (f (x), f (x )) < ε. Clearly, uniform continuity implies continuity. Although in CLASS and in INT every continuous mapping from the bounded closed interval [0, 1] into R is uniformly continuous, since this is not the case in RUSS we cannot expect to prove it in BISH; see Chapter 3 of [34]. Proposition 2.2.6. If X is a totally bounded space, and f a uniformly continuous mapping of X into a metric space, then f (X) is totally bounded. Proof. Given ε > 0, compute δ > 0 such that if x, y ∈ X and ρ(x, y) < δ, then ρ(f (x), f (y)) < ε. There exists a ﬁnite δ-approximation {x1 , . . . , xn } to X. For each x ∈ X pick i such that ρ (x, xi ) < δ. Then ρ(f (x), f (xi )) < ε. Hence 2 {f (xi ), . . . , f (xn )} is a ﬁnitely enumerable ε-approximation to f (X). Corollary 2.2.7. Let f be a uniformly continuous mapping of a totally bounded space into R. Then ran(f ) is bounded, and the supremum sup f (x) = sup {y : y ∈ ran(f )}

x∈X

and inﬁmum inf f (x) = inf {y : y ∈ ran(f )}

x∈X

of f exist. Proof. This is an immediate consequence of Propositions 2.2.6 and 2.2.5.

2

Corollary 2.2.8. If X is a totally bounded metric space, then its diameter, diam(X) = sup {ρ(x, y) : x, y ∈ X} , exists. Proof. Since the mapping ρ on the totally bounded product space X × X is uniformly continuous, the desired conclusion follows from Corollary 2.2.7. 2

We now introduce a second major theme in constructive metric space theory. An inhabited subset S of a metric space X is located in X if for each x ∈ X the distance

2.2 Metric Spaces

41

ρ (x, S) = inf {ρ(x, s) : s ∈ S} exists. We cannot prove the proposition “every inhabited subset of R is located” constructively, since it implies LEM. To see this, let P be any syntactically correct proposition, and S = {0} ∪ {x ∈ R : (x = 1) ∧ P } . If S is located, then either ρ(1, S) > 0 or ρ(1, S) < 1. In the ﬁrst case we have ¬P. In the second, choosing s ∈ S such that ρ(1, s) < 1, we see that s ∈ / {0} , so s = 1 and P holds. It is hard to overemphasise the importance of locatedness in constructive analysis. To illustrate, we anticipate later sections and chapters by pointing out that the norm of a bounded linear functional on a normed space exists if and only if the kernel of the functional is located, and that locatedness plays a vital role in the separation and Hahn–Banach extension theorems. It is only when we work constructively that the signiﬁcance of locatedness (a property that holds automatically under classical logic) for the proofs of many existence theorems is revealed. A subset S of a metric space is located if and only if its closure is located; in fact, for each x ∈ X we have ρ(x, S) = ρ(x, S) if either side of this equation exists. Note that even if a subset S of a metric space X is located, we may not be able to prove that its metric complement −S = {x ∈ X : ρ(x, S) > 0} is located (Exercise 11). Proposition 2.2.9. A totally bounded subset of a metric space is located. Proof. Let S be totally bounded in the metric space X. For each x ∈ X the mapping s ρ(x, s) is uniformly continuous on S. It follows from Corollary 2.2.7 that 2 inf {ρ(x, s) : s ∈ S} exists.

The next two proofs use a technique that will reappear from time to time. We have a ﬁnitely enumerable set {x1 , . . . , xn } of a metric space X, a located subset S of X, and positive numbers a, b with a < b. We write {1, . . . , n} as a union of subsets P, Q such that i ∈ P =⇒ ρ(xi , S) < b, i ∈ Q =⇒ ρ(xi , S) > a. By concentrating on one of those sets, we are able to reach our desired conclusion.

42

2 Techniques of Elementary Analysis

Proposition 2.2.10. A located subset of a totally bounded metric space is totally bounded. Proof. Let S be a located subset of a totally bounded metric space X, and let ε > 0. There exists a ﬁnite ε/3-approximation {x1 , . . . , xn } to X. Using Proposition 2.1.8, write {1, . . . , n} as a union of two sets P and Q, where ρ(xi , S) < 2ε/3 if i ∈ P, and ρ(xi , S) > ε/3 if i ∈ Q. For each i ∈ P there exists si ∈ S such that ρ(xi , si ) < 2ε/3. / Q; Given s ∈ S, choose i such that ρ(s, xi ) < ε/3. Then ρ (xi , S) < ε/3, so i ∈ whence i ∈ P and therefore ρ(s, si ) ρ(s, xi ) + ρ(xi , si )

2−n+2 r

if i ∈ A, if i ∈ B.

Then Fn+1 = {xi : i ∈ A} clearly satisﬁes the appropriate instance of (b). Let x be any point of B(x0 , r). By the induction hypothesis, there exists y in Fn with ρ(x, y) < 2−n+1 r. Choosing i in {1, . . . , N } such that ρ(x, xi ) < 2−n r, we have ρ(xi , Fn ) ρ(xi , y) ρ(x, xi ) + ρ(x, y) < 2−n+2 r.

2.2 Metric Spaces

43

Thus i ∈ / B, so i ∈ A and therefore xi ∈ Fn+1 . Since ρ(x, xi ) < 2−(n+1)+1 r, the set Fn+1 therefore satisﬁes the appropriate instance of (a). This completes the inductive construction. Fn in X. We see from (a) that B(x0 , r) ⊂ K. On Let K be the closure of n1

the other hand, given x ∈ K and a positive integer n, we can ﬁnd m and y ∈ Fm such that ρ(x, y) < 2−n+4 r. If m n, then by (b), there exist points ym = y, ym−1 ∈ Fm−1 , . . . , yn ∈ Fn such that ρ(yi+1 , yi ) < 2−i+3 r for n i m − 1. Thus ρ(y, Fn ) ρ (y, yn )

m−1 i=n

ρ(yi+1 , yi )

0 there exist totally bounded sets K1 , . . . , Kn , each of diameter less than or equal to ε, such that n Ki . X= i=1

Proof. Given ε > 0, construct an ε/16-approximation {x1 , . . . , xn } to X. By Proposition 2.2.11, for each i in {1, . . . , n} there exists a closed, totally bounded set Ki n such that B (xi , ε/16) ⊂ Ki ⊂ B (xi , ε/2) . Clearly, X = Ki . Also, ρ(x, y) ε for all x, y in Ki , so diam (Ki ) ε.

i=1

2

A property P, applicable to certain elements of a set S, is said to hold for all but countably many x in S if there exists a sequence (xn )n1 in S such that P (x) holds whenever x ∈ S and x = xn for each n. The sequence (xn )n1 is then called the excluded sequence, and the elements x such that x = xn for each n are said to be admissible, for the property P. Theorem 2.2.13. Let f be a uniformly continuous mapping of a totally bounded metric space X into R. Then for all but countably many r ∈ R the set X(f, r) = {x ∈ X : f (x) r} is either totally bounded or empty.

44

2 Techniques of Elementary Analysis

Proof. By Corollary 2.2.12, for each positive integer k there exist a positive integer nk and totally bounded sets Xk,j (1 j nk ), each of diameter less than 1/k, whose union is X. Deﬁne the excluded sequence (rn )n1 to be an enumeration of the real numbers ck,j = inf {f (x) : x ∈ Xk,j }

(k 1; 1 j nk ).

(Note that the numbers ck,j exist in view of Corollary 2.2.7.) Let r = rn for each n. If ck,j > r for 1 j nk , then X (f, r) = ∅. So we may assume that there exists ν nk such that ck,j < r for 1 j ν, and ck,j > r for ν < j nk . For each j ν choose xk,j ∈ Xk,j such that f (xk,j ) < r. For all such k, j and all x ∈ Xk,j we have 1 (2.5) ρ(x, xk,j ) diam (Xk,j ) < . k Consider any x ∈ X(f, r) and any positive integer k. Choose j with 1 j nk and x ∈ Xk,j . Then ck,j f (x) r, so ck,j < r (since ck,j = r) and therefore j ν. By (2.5), ρ(x, xk,j ) < 1/k. Hence {xk,j : 1 j ν} is a ﬁnitely enumerable 1/k-approximation to X(f, r). Since k is arbitrary, it follows that X(f, r) is totally bounded. 2

A complete, totally bounded metric space X is said to be compact. The bounded closed intervals [a, b] in R, and the closed balls in C, are compact; the product of ﬁnitely many compact spaces is compact. A compact subset of a metric space is both closed and (by Proposition 2.2.9) located; and a closed, located subset of a compact space is compact (see Proposition 2.2.10). Corollary 2.2.14. Under the hypotheses of Theorem 2.2.13, if X is compact, then X(f, r) is either compact or empty for all but countably many r ∈ R. Proof. In view of Theorem 2.2.13, it suﬃces to note that, by the uniform continuity of f , the set X(f, r) is closed in X, and therefore complete, for each admissible r. 2

We say that two subsets S, T of a metric space X are apart from each other, and we write S T, if there exists r > 0 such that ρ(s, t) r for all s ∈ S and t ∈ T. A mapping f : X −→ Y between metric spaces is said to be strongly continuous if for all subsets S, T of X with f (S) f (T ) we have S T. It is simple to show that uniform continuity implies strong continuity. The following partial converse will be used shortly. Proposition 2.2.15. Let f be a strongly continuous mapping of a metric space X onto a totally bounded metric space Y. Then f is uniformly continuous.

2.2 Metric Spaces

45

Proof. Given ε > 0, construct an ε/8-approximation {f (x1 ), . . . , f (xn )} to Y , and deﬁne

Yi = B f (xi ), 8ε , Xi = f −1 (Yi ). Write {(i, j) : 1 i, j n} as the union of subsets P and Q such that 3ε , 4 ε (i, j) ∈ Q =⇒ ρ (f (xi ), f (xj )) > . 2 (i, j) ∈ P =⇒ ρ (f (xi ), f (xj ))

ε . 4

Hence Yi Yj . The strong continuity of f now yields Xi Xj ; whence there exists ri,j > 0 such that ρ (x, x ) ri,j for all x ∈ Xi and x ∈ Xj . Let δ = min {ri,j : (i, j) ∈ Q} > 0, and consider points x, x of X with ρ (x, x ) < δ. Choose i, j such that f (x) ∈ Yi and f (y) ∈ Yj . If (i, j) ∈ Q, then ρ (x, x ) ri,j δ, a contradiction. Hence (i, j) ∈ P ; 2 so ρ (f (xi ), f (xj )) < 3ε/4 and therefore ρ (f (x), f (x )) < ε. A mapping f : X −→ Y between metric spaces is said to be strongly injective, or a strong injection, if it is one-one and the inverse mapping is strongly continuous; in other words, if for all subsets S, T of X with S T we have f (S) f (T ). Every strongly injective mapping is injective. The following is a constructive substitute for the classical theorem that a continuous one-one mapping of a compact metric space into a metric space has compact range and a continuous inverse (a theorem that is false in RUSS). Proposition 2.2.16. Let f be a uniformly continuous strong injection of a compact metric space X into a metric space Y. Then f −1 is uniformly continuous and strongly injective on f (X), and f (X) is compact. Proof. The mapping f −1 : f (X) −→ X is strongly continuous, and X is totally bounded. Hence, by Proposition 2.2.15, f −1 is uniformly continuous on f (X). Since f is uniformly continuous and hence strongly continuous, f −1 is also strongly injective. By Proposition 2.2.6, f (X) is totally bounded; so it remains to prove that it is complete. Accordingly, let (f (xn ))n1 be a Cauchy sequence in f (X). Using the uniform continuity of f −1 , we see that (xn )n1 is a Cauchy sequence in X. Since X is complete, (xn )n1 converges to a limit x in X. Finally, since f is continuous, 2 (f (xn ))n1 converges to f (x) in f (X).

46

2 Techniques of Elementary Analysis

Let K(X) denote the set of all compact subsets of a compact metric space X. For all A, B ∈ K(X), since the map x ρ (x, B) is deﬁned (B being located) and is uniformly continuous on A, the number mA,B = inf {ρ (x, B) : x ∈ A} exists, by Corollary 2.2.7. Likewise, mB,A exists. It is left as an exercise to show that the mapping (A, B) max {mA,B , mB,A } from K(X) × K(X) into the nonnegative real line R0+ = {x ∈ R : x 0} is a metric—we call it the Hausdorﬀ metric—with respect to which the space K(X) is compact. An inhabited metric space X is said to be locally totally bounded if each bounded subset of X is contained in a totally bounded subset; locally compact if it is both locally totally bounded and complete. Every compact space is locally compact. The spaces R and C, and the product spaces Rn and Cn , are locally compact. A metric space X is locally compact if and only if every bounded subset of X is contained in a compact set. The following lemma prepares the way for our next theorem, which deals with certain fundamental properties of a locally totally bounded space. Lemma 2.2.17. Let Y be a located subset of a metric space X, and T a totally bounded subset of X that intersects Y. Then there exists a totally bounded set S such that T ∩ Y ⊂ S ⊂ Y. Proof. Using Theorem 2.2.13, construct a sequence (αn )n1 such that for each positive integer n, 0 < αn+1 < αn < 1/n and Tn = {x ∈ T : ρ(x, Y ) αn } is totally bounded. Let Fn be a ﬁnite 1/n-approximation to Tn , and construct a mapping φn : Fn −→ Y such that ρ(x, φn (x)) < 1/n for each x ∈ Fn ; then set Sn = {φn (x) : x ∈ Fn } .

Sn in Y. We show that SN is a 3/N -approximation to Let S be the closure of n1 Sk . Let k N and y ∈ Sk . Then y = φk (x) ∈ Y for some x ∈ Fk . Since

kN

Fk ⊂ Tk ⊂ TN , there exists z ∈ FN such that ρ(x, z) < 1/N ; whence

2.3 Normed Linear Spaces

ρ(y, φN (z)) ρ(φk (x), x) + ρ(x, z) + ρ(z, φN (z)) < We now see that

N k=1

Sk is a 3/N -approximation to

n1

47

1 1 3 1 + + . k N N N

Sn . So

n1

Sn , and therefore

S, is totally bounded. If x ∈ T ∩ Y, then for each n we have x ∈ Tn ; so there exists z ∈ Fn such that ρ(x, z) < 1/n and therefore ρ(x, φn (z)) < 2/n. Thus ρ(x, S) < 2/n for each positive integer n. Since S is closed in Y , it follows that x belongs to S. 2

Proposition 2.2.18. Let Y be an inhabited subset of a metric space X. (a) If Y is locally totally bounded, then it is located. (b) If X is locally totally bounded and Y is located, then Y is locally totally bounded. Proof. Assume ﬁrst that Y is locally totally bounded. Let y0 ∈ Y and x ∈ X. The set B = {y ∈ Y : ρ(y, y0 ) 2ρ(x, y0 ) + 1} , being bounded in Y, is contained in a totally bounded subset K of Y. By Theorem 2.2.13, there exists r > 2ρ(x, y0 ) + 1 such that the set T = {y ∈ K : ρ(y, y0 ) r} is totally bounded and hence located. Then ρ(x, T ) ρ(x, y0 ) < ρ(x, y0 ) + 1, so for each y in Y, either ρ(x, y) > ρ(x, T ) or ρ(x, y) < ρ(x, y0 ) + 1. In the latter case, ρ(y, y0 ) ρ(x, y) + ρ(x, y0 ) 2ρ(x, y0 ) + 1 < r, so y ∈ T and therefore ρ(x, y) ρ(x, T ). It follows that ρ(x, Y ) exists and equals ρ(x, T ). Thus Y is located. Now assume that X is locally totally bounded and that Y is located in X. Let B be a bounded subset of Y. Then there exists a totally bounded subset T of X such that B ⊂ T. By Lemma 2.2.17, there exists a totally bounded subset S of Y such that T ∩Y ⊂ S and therefore B ⊂ S. Thus Y is locally totally bounded. 2

2.3 Normed Linear Spaces Let K stand for either R or C, and let X be a linear space over K. An inequality relation = on X is said to be compatible with the linear structure on X if for all x, y ∈ X and all t ∈ K,

48

2 Techniques of Elementary Analysis

x = y ⇐⇒ x − y = 0, x + y = 0 =⇒ x = 0 ∨ y = 0, and tx = 0 =⇒ t = 0 ∧ x = 0. It readily follows from the ﬁrst of these properties that x = y =⇒ ∀z ∈ X (x + z = y + z) . The requirement of compatibility between the inequality and the linear structure is a natural one and is automatically fulﬁlled by the denial inequality under classical logic. Constructively, it is certainly true that the denial inequality fulﬁls the ﬁrst of the three requirements for compatibility; but unless we accept Markov’s principle we cannot expect to prove that it satisﬁes the second or the third. In the case X = K the standard inequality is compatible with the linear structure. From now on, unless we state otherwise, when we refer to a linear space we mean one that is equipped with a compatible inequality. Let X be such a space. By a seminorm on X we mean a mapping : x x of X into R0+ such that for all x, y in X and all t in K, • x > 0 =⇒ x = 0, • tx = |t| x , and • x + y x + y . We call the pair (X, )—or, when no confusion is likely, just X itself—a seminormed (linear) space over K. We say that the seminormed space X is real or complex, depending on whether K is R or C. If x ∈ X and x > 0, then x is called a nonzero vector ; if x = 1, then x is called a unit vector . We call the seminormed space X nontrivial if it contains a nonzero vector. If the inequality on the seminormed space X satisﬁes x = 0 ⇐⇒ x > 0,

(2.6)

then we call a norm on X, x the norm of the vector x, and (X, )—or just X—a normed (linear) space over K. Note that every seminorm on a linear space X induces an inequality relation—namely, the one deﬁned by (2.6)—with respect to which becomes a norm. Let X be a normed space. Then the mapping (x, y) x − y of X × X into R is a metric on X, and is said to be associated with the norm on X. When we consider X as a metric space, it is understood that we are referring to the metric, usually denoted by ρ, associated with the given norm on X. Note that the inequality corresponding to the metric associated with the norm on X is just the original inequality on X. The unit ball of X is the closed ball with centre 0 and radius 1,

2.3 Normed Linear Spaces

49

BX = B X (0, 1) = B(0, 1) = {x ∈ X : x 1} , relative to that metric. This ball, like any open or closed ball in a normed space, is located. It is a simple consequence of the triangle inequality that | x − y| x − y for all vectors x, y in a normed space X. It follows from this that if a sequence (xn )n1 converges to a limit x in X, then the sequence (xn )n1 converges to x in R. Perhaps the simplest examples of a norm are the following ones on Kn : (x1 , . . . , xn ) max {|x1 | , . . . , |xn |} , 2 2 (x1 , . . . , xn ) |x1 | + · · · + |xn | , (x1 , . . . , xn ) |x1 | + · · · + |xn | . The second of these is called the Euclidean norm on Kn , which, when equipped with that norm, is called Euclidean n-space or simply Euclidean space. If S is a compact metric space, then for each uniformly continuous map f : S −→ K the sup norm f = sup {|f (x)| : x ∈ S} is well deﬁned, by Corollary 2.2.7. It is easy to see that the mapping f f is a norm on the space C(S, K) of uniformly continuous functions from S to K, taken with pointwise operations of addition and multiplication-by-scalars. We usually denote C(S, C) by C(S). Convergence and Cauchyness with respect to the sup norm on C(S, K), where S is a compact metric space, are called uniform convergence and uniform Cauchyness respectively. The space C(S, K) is an example of a Banach space: that is, a complete normed space. The standard proof of this in classical analysis is constructive, and is left as an exercise. The Euclidean space Kn is also a Banach space. Let X1 , X2 be normed spaces over K, and recall that the standard operations of addition and multiplication-by-scalars on the product vector space X = X1 × X2 are given by (x1 , x2 ) + (x1 , x2 ) = (x1 + x1 , x2 + x2 ) , t (x1 , x2 ) = (tx1 , tx2 ) , where xi ∈ Xi , xi ∈ Xi , and t ∈ K. It is easy to verify that the mapping (x1 , x2 ) max {x1 , x2 } is a norm on X, and that the metric associated with this norm is the product metric on X (considered as the product of the metric spaces X1 and X2 ). Taken with this

50

2 Techniques of Elementary Analysis

norm, which we call the product norm, X is known as the product of the normed spaces X1 and X2 . The product norm and the product space for a ﬁnite number of normed spaces are deﬁned analogously. The proof of the following is left as an exercise. Proposition 2.3.1. Let X be a normed space over K. Then (a) the mapping (x, y) x + y is uniformly continuous on X × X; (b) for each t ∈ K the mapping x tx is uniformly continuous on X; (c) for each x ∈ X the mapping t tx is uniformly continuous on K; (d) the mapping (t, x) tx is continuous on K × X. If X is a normed space and S is a linear subset of X, then the restriction to S of the norm on X is a norm on S; taken with this norm, S is called a normed linear subspace, or simply a subspace, of X. It is left as an exercise to show, from Proposition 2.3.1, that the closure of a subspace in X is also a subspace of X. Recall that a mapping u between vector spaces X, Y is linear if u(x + y) = u(x) + u(y) and u(tx) = tu(x) whenever x, y ∈ X and t ∈ K. For example, the mapping x Ax on Kn , where A is an n-by-n matrix over K, is linear. If X = Y, then we refer to a linear map u : X −→ Y as an operator on X. On the other hand, if Y = K, then u is called a linear functional on X. Continuous linear mappings between normed spaces are the backbone of functional analysis. In order to characterise these maps, we need a simple lemma, whose proof is left to the reader. Lemma 2.3.2. For each element x of a normed space X, 1 : t ∈ K, t = 0, tx 1 . x = inf |t| Proposition 2.3.3. The following are equivalent conditions on a linear mapping of a normed space X into a normed space Y : (a) u is continuous at 0. (b) u is continuous on X. (c) u is uniformly continuous on X. (d) u (BX ) is a bounded subset of Y .

2.3 Normed Linear Spaces

51

(e) u maps bounded subsets of X to bounded subsets of Y. (f) There exists a positive number c, called a bound for u, such that u(x) c x for all x ∈ X. Proof. Suppose that u is continuous at 0. Then there exists r > 0 such that u(x) = u(x) − u(0) 1 whenever x r. For each nonzero t ∈ K with tx 1 we have rtx r and therefore 1 1 u(rtx) . u(x) = r |t| r |t| It follows from Lemma 2.3.2 that u(x) (f).

1 r

x for all x ∈ X. Hence (a) implies

It is clear that (f) =⇒ (e) =⇒ (d). Next, suppose that there exists c > 0 such that u(x) c whenever x 1. Given x in X and ε > 0, we have either x = 0 or x < ε. In the ﬁrst case, 1 x c x . u(x) = x u x In the second, u ε−1 x c, so u(x) cε c (x + ε) . Thus u(x) c (x + ε) in each case. Letting ε −→ 0, we obtain u(x) c x. Thus (f) holds. We now have u(x − y) c x − y

(x, y ∈ X) ,

from which it follows that u is uniformly continuous on X. Thus (d) =⇒ (f) =⇒ (c). Finally, it is immediate that (c) =⇒ (b) =⇒ (a).

2

In view of property (e) of Proposition 2.3.3, we refer to a continuous linear mapping between normed spaces X, Y as a bounded linear mapping. In the cases Y = X and Y = K, we use the terms bounded operator and bounded linear functional, respectively. Bounded linear functionals are associated with linear subsets that, in the case of a general normed space, correspond to planes in three-dimensional geometry. A linear subset H of a normed space X is called a hyperplane if there exist an associated vector x0 ∈ X and a positive number c such that

52

2 Techniques of Elementary Analysis

x − x0 c for each x ∈ H, and each x ∈ X is represented (perforce uniquely) in the form x = tx0 + y with t ∈ K and y ∈ H. Recall that the kernel of a linear mapping u between linear spaces X, Y is the set

ker(u) = u−1 (0) .

We say that u is nonzero if there exists x ∈ X such that u(x) = 0. If u is a nonzero bounded linear functional, then ker(u) is a hyperplane. To see this, choose x0 ∈ X such that u(x0 ) = 1, and then (by Proposition 2.3.3) c > 0 such that c |u(x)| x for each x ∈ X. If x ∈ ker(u), then x − x0 c |u(x − x0 )| = c |u(x) − u(x0 )| = c. On the other hand, for each x ∈ X we have x = u(x)x0 + (x − u(x)x0 ) , where u(x) ∈ K and x − u(x)x0 ∈ ker(u). Proposition 2.3.4. Let X be a normed space, and H a hyperplane in X with associated vector x0 . Then there exists a unique bounded linear functional u on X such that ker(u) = H and u(x0 ) = 1. Proof. Compute c > 0 such that x − x0 c for each x ∈ H. For each x ∈ X there exist unique u(x) ∈ K and φ(x) ∈ H such that x = u(x)x0 + φ(x). By the uniqueness, u(x0 ) = 1 and H = ker(u). Moreover, for each ε > 0 we have either |u(x)|

1 x + ε c

(2.7)

or else u(x) = 0. In the latter case we have 1 φ(x) x = |u(x)| x0 − − |u(x)| c, u(x) from which we see that (2.7) holds. Since ε > 0 is arbitrary, we conclude that 1/c is a bound for u. 2 If u : X −→ Y is a bounded linear mapping between normed spaces and u = sup {u(x) : x ∈ X, x 1}

(2.8)

exists, we call this number the norm of u and we say that u is normed or normable. In that case,

2.3 Normed Linear Spaces

u(x) u x

53

(x ∈ X) ,

and if X is nontrivial, then u = sup {u(x) : x ∈ X, x = 1} . Although there is no general criterion for the existence of the norm of a bounded linear mapping between normed spaces, such a criterion—and a very useful one at that—exists for bounded linear functionals. Our proof depends on a preliminary lemma. Lemma 2.3.5. Let u be a bounded linear map of a normed space X into a normed space Y. Then ker(u) is located if and only if sx = inf {t : t > 0, u(x) ∈ tu(BX )}

(2.9)

exists for each x ∈ X. Moreover, if u(x) = 0, then sx > 0. Proof. For each x ∈ X we have {t > 0 : ∃y ∈ ker(u)(x − y < t)} = {t > 0 : ∃z ∈ X (z < 1 and u (x − tz) = 0)} = {t > 0 : u(x) ∈ tu(BX )}. The ﬁrst part of the lemma now follows, since ρ (x, ker(u)), if it exists, equals the inﬁmum of the ﬁrst set, whereas sx , if it exists, equals the inﬁmum of the third. Now let u(x) = 0, and choose r > 0 such that if y ∈ X and y < r, then |u(y)| < |u(x)| . Supposing that sx < r, we can ﬁnd a positive t < r and an element z of BX such that u(x) = tu(z) = u(tz); but tz < r, so |u(tz)| < |u(x)| , a con2 tradiction. Hence sx r.

Proposition 2.3.6. A nonzero bounded linear functional on a normed space is normed if and only if its kernel is located. Proof. Let u be a nonzero bounded linear functional on the normed space X. Supposing ﬁrst that u is normed—in which case u > 0—consider any a ∈ X. For each y ∈ ker(u) we have a − y

|u(a)| |u(a − y)| = . u u

On the other hand, if 0 < ε < u and we choose a unit vector x ∈ X such that u(x) > u − ε, then u(a) x ∈ ker(u) z =a− u(x)

54

2 Techniques of Elementary Analysis

and a − z =

|u(a)| |u(a)| < . u(x) u − ε

Hence ρ(a, ker(u)) exists and equals |u(a)| / u . Since a ∈ X is arbitrary, ker(u) is located. Conversely, if ker(u) is located, then by Lemma 2.3.5, s = inf {t > 0 : 1 ∈ tu(BX )} exists and is positive. We show that u = 1/s. For each x ∈ BX we have either u(x) 1/s or u(x) = 0; in the latter case, |u(x)| x u(x) 1 and 1=

1 u |u(x)|

|u(x)| x , u(x)

so 1/ |u(x)| s and therefore |u(x)| 1/s. Thus 1/s is a bound for u. On the other hand, given ε with 0 < ε < 1/s, choose t < s/ (1 − εs) and x ∈ BX such that 1 = tu(x). Then u(x) = 1/t > 1/s − ε. It now follows that u exists and equals 1/s. 2

We cannot expect the norm of a normed linear functional u to be attained at some vector in the unit ball of the domain space. For a certain important class of spaces, though, attainment of the norm does occur. A normed space X is said to be uniformly convex if for each exists ε > 0 there δ with 0 < δ < 1 such that if x, y are unit vectors in X with 12 (x + y) > 1 − δ, then x − y < ε. Every linear subspace of a uniformly convex normed space is itself uniformly convex. Proposition 2.3.7. If u is a nonzero normed linear functional on a uniformly convex Banach space X, then there exists a unique unit vector x ∈ X such that u(x) = u . Proof. Construct a sequence (xn )n1 of unit vectors in X such that u (xn ) −→ u as n −→ ∞. Given ε > 0, choose δ > 0 as in the deﬁnition of “uniformly convex”. Since 1 − 2δ < 1, there exists a positive integer N such that 1 − 2δ u < u(xn ) for all n N. For m, n N we then have

2.3 Normed Linear Spaces

55

1 1 u (xm + xn ) 2 |u(xm + xn )| 2 1 u(xm ) − |u(xm − xn )| 2 1 1 u(xm ) − (u − u(xm )) − (u − u(xn )) 2 2 δ δ δ u − u − u > 1− 2 4 4 = (1 − δ) u . Hence 12 (xm + xn ) > 1−δ and therefore xm − xn < ε. Since ε > 0 is arbitrary, we conclude that (xn )n1 is a Cauchy sequence. It therefore converges to a limit x in the complete space X. By the continuity of the norm and of u, we have x = 1 and u(x) = u . Now let y be any unit vector such that u(y) = u = u(x). For each positive integer n deﬁne z2n−1 = x and z2n = y. Then (zn )n1 is a sequence of unit vectors such that u(zn ) −→ u ; so, by the ﬁrst part of the proof, (zn )n1 converges to a unit vector, which clearly must equal both x and y. Hence x = y. 2

We next present an important construction of new normed spaces from old. Let Y be a located subspace of a normed space X, and deﬁne new equality and inequality relations on X by x =X/Y x ⇐⇒ ρ (x − x , Y ) = 0, x =X/Y x ⇐⇒ ρ (x − x , Y ) > 0. Taken with this equality relation, X becomes the quotient space of X by Y, and is usually redesignated X/Y. The identity (linear) mapping x x from the original normed space X into X/Y is then called the canonical injection iX/Y of X onto X/Y . If Y is also closed in X, then xX/Y = ρ (x, Y ) deﬁnes a norm—the quotient norm—on X/Y. Since xX/Y x − 0 = x , the canonical injection iX/Y is a bounded linear mapping. Note that the locatedness of Y is essential for the deﬁnition of the equality, inequality, and norm on X/Y. Proposition 2.3.8. If Y is a closed, located subspace of a Banach space X, then the quotient space X/Y is a Banach space. Proof. Given a Cauchy sequence (xn )n1 in X/Y, choose a strictly increasing sequence (nk )k1 of positive integers such that

56

2 Techniques of Elementary Analysis

xn

k+1

− xnk X/Y < 2−k

(k 1).

Setting y1 = 0, we construct inductively a sequence (yk )k1 in Y such that for each k, (xn − yk+1 ) − (xnk − yk ) < 2−k . (2.10) k+1 Indeed, having constructed elements y1 , . . . , yk of Y with the applicable properties, we have inf xnk+1 − (xnk − yk ) − y : y ∈ Y = xnk+1 − xnk X/Y < 2−k , so there exists yk+1 ∈ Y such that (2.10) holds. We now see from (2.10) that (xnk − yk )k1 is a Cauchy sequence in the Banach space X; whence it converges to a limit z in X. We then have xnk − zX/Y = xnk − z − yk X/Y (xnk − yk ) − z −→ 0 as k −→ ∞. Thus the Cauchy sequence (xn )n1 in X/Y has a subsequence that converges in 2 X/Y. It follows that the sequence (xn )n1 itself converges in X/Y. Normed spaces form the natural abstract context for inﬁnite series. Given a sequence (xn )n1 of elements of a normed space X, we deﬁne the corresponding ∞ n xn to be the sequence (sn )n1 , where sn = xk is the nth partial sum series n=1

of the series. The series

∞ n=1

k=1

xn is said to be

convergent if the sequence (sn )n1 converges to a limit s in X, called the sum of the series; absolutely convergent if the series In the ﬁrst case we write

∞ n=1

∞ n=1

xn is convergent in R.

xn = s.

If (an )n1 and (bn )n1 are sequences of real numbers such that 0 < an bn for ∞ ∞ each n, and if bn converges, then an converges (the comparison test). For, n=1

n=1

given ε > 0, since the partial sums of a convergent series form a Cauchy sequence we can ﬁnd N such that k bn < ε 0< n=j

whenever k > j N. For all such j, k we then have 0 < sums of

∞ n=1

k n=j

aj < ε. So the partial

an form a Cauchy sequence, which converges by the completeness of R.

2.3 Normed Linear Spaces

57

A particular case of this occurs when bn = rn for some ﬁxed r with |r| < 1: in that ∞ ∞ r case, bn is a geometric series, and an converges to a sum at most 1−r . n=1

n=1

We shall resume consideration of normed spaces in Chapter 4.

Exercises 1. Prove that two real numbers x and y are equal if and only if x ∪ y is a real number. (This is not a typographical error: remember, a real number is actually a set.) 2. Prove the following for real numbers x, y, z: x > y =⇒ x + z > y + z, (x > 0 ∧ y > 0) =⇒ xy > 0, xy = 0 =⇒ (x = 0 ∧ y = 0) . 3. Prove that for all x, y ∈ R the sets max {x, y} and min {x, y} are real numbers; that min {x, y} x max {x, y} ; that x = min {x, y} if and only if x y; and that x = max {x, y} if and only if x y. Prove also that max {x, y} < ε if and only if x < ε and y < ε; that max {x, y} ε if and only if x ε and y ε; and that max {x, y} > z if and only if either x > z or y > z. 4. Prove that |x| = x if and only if x 0, and that |x| = −x if and only if x 0. Prove also that |x| < y if and only if −y < x < y, and that |x| y if and only if −y x y. 5. Prove Archimedes’ axiom in the following form: If x > 0 and y 0, then there exists an integer n such that nx > y. 6. Let x, y be real numbers with y = 0. Prove that x/y is a real number and that y/y = 1. 7. Prove that the statement ∀x ∈ R (¬ (x = 0) =⇒ ∃y ∈ R (xy = 1)) implies Markov’s principle. 8. Prove that the mapping (x, y) x + y is a function on R × R, and that it is strongly extensional. Prove analogous results for the mappings x −x, (x, y) xy, (x, y) x/y (for y = 0), x |x| , (x, y) min {x, y} , and (x, y) max {x, y} .

58

2 Techniques of Elementary Analysis

9. Give Brouwerian examples to show that each of the following statements is essentially nonconstructive: (a) If S ⊂ R is closed, then ∼S is open in R. (b) If S ⊂ R and ∼S is open, then S is closed in R. 10. Let X be a separable metric space. Prove that (i) dense subsets and (ii) located subsets of X are separable. 11. Show that the statement “the metric complement of every located subset of R is located” implies the weak law of excluded middle (WLEM), ¬P ∨ ¬¬P. 12. Prove that the closure of a totally bounded subset of a metric space X is totally bounded, and that if a subset S of X contains a dense totally bounded set, then S itself is totally bounded. 13. Prove Corollary 2.2.8 directly, without using Proposition 2.2.5, Proposition 2.2.6, or Corollary 2.2.7. 14. Let h be a mapping of a metric space into a totally bounded space X such that f ◦ h is uniformly continuous for each uniformly continuous map f : X −→ R. Prove that h is uniformly continuous. 15. Let h be a mapping of a compact metric space X into a metric space Y such that f ◦ h is uniformly continuous on X for each uniformly continuous mapping f : Y −→ R. Prove that h is uniformly continuous if and only if its range is totally bounded. Prove also that if Y is locally compact, then h is uniformly continuous. 16. Let a < b, and let S, T be inhabited open subsets of R such that [a, b] ⊂ S ∪ T. Prove that S ∩ T is inhabited. 17. Let f : [a, b] −→ R be sequentially continuous, with f (a) f (b). Prove that the range of f is dense in [f (a), f (b)] . 18. Let I be a bounded interval, and f : I −→ R an increasing sequentially continuous function. Prove that f is uniformly continuous on I. 19. Let S be a dense subset of a metric space X, and f a uniformly continuous mapping of S into a complete metric space Y. Prove that f has a uniformly continuous extension to a mapping of X into Y ; that is, a uniformly continuous mapping F : X −→ Y such that F (x) = f (x) for all x ∈ S. 20. Prove that the Hausdorﬀ metric on the set K(X) of compact subsets of a compact metric space X is indeed a metric, and that K(X) is complete with respect to it. Why do we need the sets to be compact in order to be sure that ρ deﬁnes a metric?

2.3 Normed Linear Spaces

59

21. Let X be a compact metric space, K(X) the compact metric space of all compact subsets of X (see the preceding exercise), and E the set of all pairs (x, K) where K ∈ K(X) and x ∈ K. Deﬁne a metric d on E by d ((x, K) , (x , K )) = ρ (x, x ) + ρ (K, K ) , where the second ρ on the right-hand side denotes the Hausdorﬀ metric on K(X). Prove that E is compact with respect to this metric d. sequence of compact subsets 22. Let X be a metric space, and (Kn )n1 a decreasing Kn is compact. Does the of X whose diameters converge to 0. Prove that n1

conclusion hold without the hypothesis that the diameters converge to 0? 23. Under the hypotheses of Corollary 2.2.14, prove that if r ∈ R is admissible and ε > 0, then there exists δ > 0 such that for each admissible r with |r − r | < δ, ρ (X(f, r), X(f, r )) < ε, where ρ is the Hausdorﬀ metric on K(X). 24. Give a Brouwerian example to show that we cannot drop the condition “for all but countably many” from the conclusion of Theorem 2.2.13. 25. Prove that the product of ﬁnitely many (locally) compact metric spaces is (locally) compact. 26. Let f be a mapping from a metric space (X, ρ) into R with the following properties: (a) the set X(f, r) is compact for certain arbitrarily large real numbers r; (b) f is uniformly continuous on each of the sets X(f, r). Deﬁne a new metric ρ0 on X by ρ0 (x, x ) = ρ (x, x ) + |f (x) − f (x )|

(x, x ∈ X) .

Prove that (X, ρ0 ) is a locally compact metric space. 27. Let S be a located subset of a locally compact metric space (Y, ρ) such that X = −S is inhabited. Let the mapping h : Y −→ R be uniformly continuous on each bounded subset of Y, and such that the set Y (h, r) is bounded for −1 each r ∈ R. Prove that the mapping f : x h (x) + ρ (x, S) on X satisﬁes conditions (a) and (b) of the preceding exercise. 28. Let X be a uniformly convex Banach space, and u a normed linear functional on X. Prove that for any two distinct unit vectors x, y in X either |u(x)| < u or |u(y)| < u . (This is a strong form of the uniqueness result in Proposition 2.3.7.)

60

2 Techniques of Elementary Analysis

29. Prove Proposition 2.3.1. 30. Prove Lemma 2.3.2.

Notes In view of our earlier remarks on the status of the power set, it may seem strange, if not perverse, to deﬁne a real number as a set. We do not believe that our deﬁnition will give rise to any logical problems, since in practice a real number is likely to be given explicitly by interval approximations, rather than by any appeal to the existence of the full power set of Q × Q. In fact, many, if not most, real numbers will actually arise as limits of sequences of rational approximations, from which it is straightforward to construct a set of the type required by our deﬁnition of “real number”. Bishop deﬁned a real number to be a sequence (xn )n1 of rational numbers that is regular in the sense that |xm − xn | < 1/m + 1/n for all m and n, two such sequences (xn )n1 and (yn )n1 being called equal if |xn − yn | < 2/n for each n. For an axiomatic development of R see [19]. For us, = normally has a stronger meaning than the denial of equality. Some authors, notably the intuitionists, use # instead of = to denote an inequality relation, and = to denote the denial inequality. Bishop originally used the word “continuous” to describe mappings that are uniformly continuous on compact sets. While this gets round the independence of the uniform continuity theorem relative to BISH, it has the disadvantage that we cannot prove that the composition of continuous functions is continuous, since the image of a compact set, although totally bounded, cannot generally be proved complete. (To be more precise, in RUSS there is a uniformly continuous mapping of [0, 1] onto (0, 1].) In a later, unpublished manuscript, The neat category of stratiﬁed spaces, Bishop introduced the notion of a compact image: a subset of a metric space that is the image of a compact metric space under a uniformly continuous mapping. He then deﬁned a mapping f : X −→ Y between metric spaces to be continuous if it is uniformly continuous near each compact image, in the following sense: for each compact image S ⊂ X and each ε > 0 there exists δ > 0 such that ρ(f (x), f (x )) < ε whenever x ∈ X, x ∈ S, and ρ(x, x ) < δ. Compositions of functions that are continuous in this sense are continuous; but the interval (0, 1] is a compact image in RUSS, so we cannot prove that the mapping x 1/x on (0, 1] is uniformly continuous near compact images. It seems that without Brouwer’s fan theorem (see Chapter 5 of [34]), it is impossible to come up with a deﬁnition of “continuous” in BISH, other than the usual one of “pointwise continuous”, that will satisfy all the conditions that one might wish for. For more on this, see [81] and [89].

2.3 Normed Linear Spaces

61

In this connection, Bishop’s statement that “The concept of a pointwise continuous function is not relevant” ([9], pages ix–x) no longer seems appropriate: there are instances—for example, in the theory of operators between Banach spaces— where sequential convergence is both useful and the best we can hope to prove within BISH; see Chapter 6. Some authors remove the word “inhabited” from the deﬁnitions of “located” and “totally bounded”, thereby allowing the empty set to be both located and totally bounded. For us, located sets and totally bounded sets are inhabited, by deﬁnition. Why do we deﬁne “compact” as we have done? Classically, compactness is deﬁned in terms of the Heine–Borel–Lebesgue covering property: every open cover contains a ﬁnite subcover. In Brouwer’s intuitionistic mathematics, as classically, the interval [0, 1] has this open-cover compactness property; but Brouwer’s proof depends on his fan theorem, the addition of which to intuitionistic logic would lose us some of the ﬂexibility of interpretation of our constructive mathematics. On the other hand, if we add the Church–Markov–Turing thesis to intuitionistic logic, then we can prove that [0, 1] does not have the Heine–Borel–Lebesgue covering property ([34], page 60, Theorem (4.1)). So that property holds for [0, 1] in one model of our constructive mathematics but fails to hold in another. Therefore, without adding some principle to BISH, we cannot prove or disprove that it holds for [0, 1]. A second classical compactness property, equivalent to the Heine–Borel–Lebesgue one, is that of sequential compactness: every sequence in the space has a convergent subsequence. This fails even for the pair set {0, 1} in the constructive setting. Thus, in looking for a workable constructive notion of compactness for metric spaces, we are reduced to that of completeness plus total boundedness, which is classically equivalent to the other two considered above. Fortunately, completeness and total boundedness together make a combination suﬃciently powerful for most constructive purposes. Proposition 2.2.11 appears in Aspects of Constructivism, the unpublished notes of colloquium lectures given by Bishop at New Mexico State University in December 1972. It provides a neat proof of Corollary 2.2.12. The proof of Lemma 2.2.17 is new, although not unlike that of Lemma (4.10) on page 33 of [34]. Exercise 14 provides a constructive version of the classical theorem that the unique uniform structure U compatible with the given topology on a compact Hausdorﬀ space X is induced by the U-uniformly continuous mappings of X into R. Note, however, that we require X to be only totally bounded, not compact. The notion of strong continuity is best studied in the context of an apartness space: that is, an inhabited set X with an inequality and a binary relation of apartness between sets that satisﬁes certain natural axioms. For more on apartness spaces see [38, 40]. Exercise 16 describes one version of the connectedness of the interval [a, b]. For other, constructively inequivalent, types of connectedness see [18].

62

2 Techniques of Elementary Analysis

The result in Exercise 17 is perhaps the most general constructive intermediate value theorem. A linear map u : X −→ Y between normed spaces is said to be compact if u(BX ) is a totally bounded subset of Y. Proposition 2.3.6 can be generalised as follows: A bounded linear mapping of a normed linear space onto a ﬁnite-dimensional Banach space is compact if and only if its kernel is located ([34], page 36, Theorem (5.4)). Bishop required that a Banach space be separable. We prefer not to depend on separability unless it is absolutely necessary. We use the term “normed”, rather than the usual “normable”, for those linear mappings for which the norm exists.

3 The λ-Technique

Here is more matter for a hot brain. —Shakespeare, The Winter’s Tale, act 5, scene 2 In this chapter we discuss a peculiarly constructive technique that, normally under the hypothesis that one or more of the metric spaces under consideration is complete, enables us to prove results that otherwise would require some nonconstructive principle such as LPO, LLPO, or Markov’s principle.

3.1 Introduction to the Technique We begin with another elementary classical result that does not hold in constructive mathematics: For each complex number z there exists θ ∈ [0, 2π) such that z = |z| eiθ , and such that if θ = 0, then z = 0. We show that this proposition entails LPO. To do so, we consider an increasing binary sequence (λn )n1 with at most one term equal to 1. Deﬁne a sequence (zn )n1 of complex numbers such that λn = 0 =⇒ zn = 0, 1 λn = 1 − λn−1 =⇒ zk = eiπ/2 for all k n. n Then |zm − zn |< 1/n whenever m > n, so (zn )n1 is a Cauchy sequence and therefore converges to a limit z in the complete metric space C. Assume that z = |z| eiθ for some θ ∈ [0, 2π), and that if θ = 0, then z = 0. Either θ < π/2 or θ > 0. In the ﬁrst case we have λn = 0 for all n: for if we suppose that there exists n such

64

3 The λ-Technique

that λn = 1 − λn−1 , then z = eiπ/2 /n and therefore θ = π/2, a contradiction. In the case θ > 0 we have z = 0, so there exists N such that zN = 0; then λn = 1 for some n N. The constructive problem occurs when z is near 0 but we cannot decide whether z = 0 or z = 0. Indeed, if z = 0 or z = 0, then we can ﬁnd θ ∈ [0, 2π) such that z = |z| eiθ . Now, it might be thought that the failure of the modulus–argument decomposition of a general complex number would mean that there was no constructive proof of the existence of square roots in C; for in order to ﬁnd a square root of z, we normally would write z = |z| eiθ and then compute ± |z|eiθ/2 . Although this method of ﬁnding a square root certainly will not work unless we already can decide that z = 0 or z = 0—which, in general, we cannot—there is a constructive proof of the existence of square roots, one that uses the completeness of C. To see this, consider any complex number z, and note that for each positive integer n we have either |z| < 1/n2 or |z| > 1/(n + 1)2 . Thus we can successively construct the terms of an increasing binary sequence (λn )n1 such that λn = 0 =⇒ |z| < λn = 1 =⇒ |z| >

1 , n2 1 (n + 1)

2.

We may assume that λ1 = 0. If λn = 0, setζ n = 0; if λn = 1 − λn−1 , choose θ ∈ [0, 2π) such that z = |z| eiθ , and set ζ k = |z|eiθ/2 for all k n. Then (ζ n )n1 is a Cauchy sequence. To see this, let m n. If λm = 0 or λn = 1, then ζ m = ζ n . If λm = 1 and λn = 0, then there exists a unique k such that n < k m and 2 λk = 1 − λk−1 ; whence |z| < 1/ (k − 1) and |z|

0.

3.1 Introduction to the Technique

65

Proof. Let s0 ∈ S. If ρ(x, S) > 1/2, then we can take s = s0 . Hence we may assume that ρ(x, S) < 1. Now construct an increasing binary sequence (λn )n1 such that λ1 = 0 and 1 , n 1 . λn = 1 =⇒ ρ(x, S) > n+1

λn = 0 =⇒ ρ(x, S)

3/N. If λN = 0, then ρ(x, sN ) < 1/N and so ρ(s, sN ) ρ(x, s) − ρ(x, sN ) >

2 , N

a contradiction. Thus λN = 1 and therefore ρ(x, S) > 0.

2

Bishop’s lemma is simple to prove using classical logic. For if x ∈ S, then by taking s = x we render the hypothesis of the implication “if x = s, then ρ(x, S) > 0 ” false; whereas if x ∈ / S, then the conclusion of that implication holds since S is closed in X, so we may take s = s0 . This argument depends on the full law of excluded middle, but in fact only LPO is needed to establish Bishop’s lemma classically even when S is merely closed and not necessarily complete. To see this, ﬁrst construct the sequence (λn )n1 as in our constructive proof of the lemma. If λn = 0 for all n, then ρ(x, S) = 0, so x ∈ S and we can take s = x; whereas if λN = 1, then ρ(x, S) > 0 and we can take s = s0 . In order to obtain a constructive proof of Bishop’s lemma, we replace LPO by the completeness of S. This is a typical situation, in which a result proved classically using an omniscience principle is proved constructively under some completeness condition using the λ-technique. Now, it follows immediately from Bishop’s lemma that if S is a complete, located subset of a metric space, then ∼S = −S. The λ-technique provides us with a related result in a Banach space, in which locatedness is replaced by convexity. A subset C of a linear space X is said to be convex if tx + (1 − t) y ∈ C whenever x, y ∈ C and 0 t 1; absorbing if for each x ∈ X there exists t > 0 such that x ∈ tC.

66

3 The λ-Technique

Proposition 3.1.2. If C is a convex, absorbing subset of a Banach space X, then ∼C is dense in ¬C. Proof. Let x ∈ ¬C, let ε > 0, and choose δ > 0 such that δx < ε. Then x = (1 + δ)x ∈ (1 + δ)C. Given y ∈ C, we show that x = y. To that end, construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ x − y

n. Hence (zn )n1 is a Cauchy sequence and so converges to a limit z in the complete space X. Choose a positive integer N such that z ∈ N δC, and consider any integer n N . If λn = 1 − λn−1 , then z = n(x − y); whence x = y +

Nδ 1 z∈y+ C ⊂ C + δC = (1 + δ)C, n n

a contradiction. Thus λn = λn−1 for all n N . It follows that if λN = 0, then λn = 0 for all n, and therefore x = y ∈ C ⊂ (1 + δ)C. This contradiction ensures that λN = 1; whence x = y. Since y ∈ C is arbitrary, it follows that x ∈ ∼C. Since also x − x = δx < ε, and x ∈ ¬C and ε > 0 are arbitrary, we conclude that ∼C is dense in ¬C. 2

Classically, as the reader is invited to prove, the completeness of the space X can be removed from Proposition 3.1.2 since we are allowed to use LPO. For an example of the use of the λ-technique to avoid the application of Markov’s principle, we turn to an elementary result in metric topology. We observed in Chapter 2 that it cannot be proved constructively that the union of two closed subsets of a metric space is closed. Nevertheless, it seems reasonable to expect that if A, B are closed subsets of R, then the complement, in some appropriate sense, of A ∪ B will be the intersection of the complements of A and B. Since it is trivial that the ﬁrst of these complements is a subset of the intersection of the other two, it is enough to prove the reverse inclusion. With the help of Markov’s principle, we can prove that

3.1 Introduction to the Technique

67

(¬A ∩ ¬B) ∩ A ∪ B = ∅. Indeed, supposing that x ∈ (¬A ∩ ¬B) ∩ A ∪ B, choose a sequence (xn )n1 of elements of A ∪ B converging to x. Deﬁne an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (xk ∈ A) , λn = 1 − λn−1 =⇒ xn ∈ B. We may assume that λ1 = 0. If λn = 0 for all n, then x ∈ A = A, a contradiction. Hence, by Markov’s principle, there exists n > 1 such that λn = 1 − λn−1 , so there exists n1 > 1 with xn1 ∈ B. Repeated application of this argument enables us to construct a strictly increasing sequence (nk )k1 of positive integers such that xnk ∈ A for all even k, and xnk ∈ B for all odd k. Thus x ∈ A ∩ B, which is a ﬁnal contradiction. Since such an argument, depending on Markov’s principle, is not good enough for our constructive purposes, it is fortunate that if we work within a complete metric space and use the λ-technique (on more than one occasion), we can say something interesting about the intersection of the complements of two closed sets. This will require a preliminary result. Lemma 3.1.3. Let X be a complete metric space, A a closed subset of X, and B a subset of X. Let x ∈ ∼A, let y ∈ A ∪ B, and let (yn )n1 be a sequence in A ∪ B that converges to y. Then either x = y or there exists n such that yn ∈ B. Proof. We may assume that y1 ∈ A and, by passing to a subsequence if necessary, that ρ (yn , y) < 1/n for each n. Construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (yk ∈ A) , λn = 1 − λn−1 =⇒ yn ∈ B. If λn = 0, set ξ n = yn ; if λn = 1 − λn−1 , set ξ k = ξ n−1 for all k n. Then (ξ n )n1 is a Cauchy sequence in A; in fact, 2 (m n) . n Since X is complete and A is closed, (ξ n )n1 converges to a limit ξ ∈ A. Choose a positive integer N such that ρ (x, ξ) > 2/N. If λN = 1, then yn ∈ B for some n N. If λN = 0, then either x = y or, as we may suppose, ρ (x, y) < 1/N. If there exists m N such that λm+1 = 1 − λm , then ρ (ξ m , ξ n ) ρ (ξ m , y) + ρ (ξ n , y)

− , > N N m N a contradiction. Hence λn = 0 for all n N and therefore for all n; so yn ∈ A for all n. Since A is closed, y ∈ A and therefore x = y. 2

68

3 The λ-Technique

Proposition 3.1.4. If A, B are closed subsets of a complete metric space X, then ∼A ∩ ∼B = ∼A ∪ B. Proof. It is clear that ∼A ∪ B ⊂ ∼A ∩ ∼B. To prove the reverse inclusion, given x ∈ ∼A ∩ ∼B and y ∈ A ∪ B, we must prove that x = y. To this end, choose a sequence (yn )n1 in A ∪ B that converges to y. We may assume that y1 ∈ A and that ρ (yn , y) < 1/n for each n. Set λ1 = 0, n1 = 1, and ξ 1 = y1 . In view of Lemma 3.1.3, we may also assume that there exists n2 > 1 such that yn2 ∈ B; set λ2 = 0 and η 1 = yn2 . Now apply the same lemma, but with the roles of A and B interchanged. Either we have x = y, when for each k 3 we set λk = 1, nk = n2 , ξ k−1 = ξ 1 , and η k−1 = η 1 ; or else, as we may assume, there exists n3 > n2 such that yn3 ∈ A, in which case we set λ3 = 0 and ξ 2 = yn3 . Repeating such applications of Lemma 3.1.3, we construct an increasing binary sequence (λn )n1 , an increasing sequence (nk )k1 of positive integers, a sequence (ξ k )k1 of elements of A, and a sequence (η k )k1 of elements of B, such that for each k 1, • if λ2k = 0, then n2k > n2k−1 and η k = yn2k ∈ B; • if λ2k+1 = 0, then n2k+1 > n2k and ξ k+1 = yn2k+1 ∈ A; • if λ2k = 1−λ2k−1 , then x = y and for each j k, n2j+1 = n2j = n2j−1 , ξ j = ξ k , and η j = η k−1 ; • if λ2k+1 = 1−λ2k , then x = y and for each j k, n2j+2 = n2j+1 = n2k , ξ j = ξ k , and η j = η k . Then (ξ k )k1 is a Cauchy sequence in A, and (η k )k1 is a Cauchy sequence in B. Indeed, for m n we have 2 ρ (ξ m , ξ n ) ρ (ξ m , y) + ρ (ξ n , y) < , n and similarly ρ (η m , η n ) < 2/n. Since X is complete and the subsets A, B are closed in X, the sequences (ξ n )n1 , (η n )n1 converge to respective limits ξ ∈ A and η ∈ B. Choose an integer N > 4 such that ρ (x, ξ) > 3/N and ρ (x, η) > 3/N. If λN = 1, then x = y; so we may assume that λN = 0. Either x = y or, as we may further assume, ρ (x, y) < 1/N. Suppose there exists m > N such that λm = 1 − λm−1 . In the case where m is even, we have ρ(x, y) ρ (x, η) − ρ (η, y)

3 − ρ ynm−2 , y > N 1 3 − > N m−2 2 3 − (since m > 4) > N m 1 > , N

3.1 Introduction to the Technique

69

a contradiction. If m is odd, a similar argument again leads us to a contradiction. Hence λn = 0 for all n N and therefore for all n. It follows that y = ξ = η and therefore ρ (x, y) > 3/N, a contradiction. Hence the case ρ (x, y) < 1/N is ruled out, and so ρ (x, y) 1/N. 2

The λ-technique turns out to be very useful for clarifying the connections between various continuity properties of functions between metric spaces. It is trivial that a continuous mapping between metric spaces is strongly extensional. If the domain is complete, then we can weaken continuity to sequential continuity. Proposition 3.1.5. Let f be a sequentially continuous mapping of a complete metric space X into a metric space Y . Then f is strongly extensional. Proof. Given points x, y of X with f (x) = f (y), construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ρ(x, y)

0. Proof. Construct an increasing binary sequence (λn )n1 such that

70

3 The λ-Technique

λn = 0 =⇒ ρ ((x, y) , B)

0.

2,

We may assume that λ1 = 0. If λn = 0, set zn = 0; if λn = 1 − λn−1 , pick ξ ∈ X such that 1 (3.1) (ξ, T ξ) ∈ B and x − ξ + y − T ξ < 2 n and set zi = n (x − ξ) for all i n. If j > k and zj − zk > 0, then λk = 0 and λn = 1 − λn−1 for a unique value of n with k < n j; so there exists ξ ∈ X such that (3.1) holds and 1 1 < . n k Hence zj − zk < 1/k whenever j k. Therefore (zn )n1 is a Cauchy sequence in X and so converges to a limit z ∈ X. Now choose a positive integer N such that zj − zk = n (x − ξ)

N. If λn = 1 − λn−1 , then z = n (x − ξ) for some ξ satisfying (3.1), so 1 + T z = 1 + n T x − T ξ 1 + n T x − y − n y − T ξ 1 > 1 + N T x − y − n > N T x − y , a contradiction of our choice of N. Hence λn = λn−1 for all n N. It follows that if λN = 0, then λn = 0 for all n, and therefore ρ ((x, y) , B) = 0. Since B is closed in X × Y, we now have y = T x, which contradicts our original hypotheses. We 2 conclude that λN = 1. Corollary 3.1.7. A linear mapping of a Banach space into a normed space is strongly extensional. Proof. Let T be a linear mapping of a Banach space X into a normed space Y, and let x ∈ X satisfy T x = 0. Given z ∈ ker(T ), apply Proposition 3.1.6 with y = 0 and B = {(z, T z)} , to obtain (x, 0) = (z, T z); whence either x = z or else T z = 0. Since the latter is absurd, we conclude that x = z. 2 Markov’s principle is all that we need add to intuitionistic logic in order to prove Proposition 3.1.6 without the hypothesis that X is complete: for the argument towards the end of the proof of that proposition shows that it is impossible that ρ ((x, y) , B) = 0, so, by Markov’s principle, ρ ((x, y) , B) > 0. In fact (this is left to the exercises at the end of the chapter), if Proposition 3.1.6 holds without the completeness of X, then we can derive Markov’s principle.

3.2 Ishihara’s Tricks

71

3.2 Ishihara’s Tricks In [57], Ishihara introduced the following two lemmas, now called Ishihara’s tricks, in which we use completeness to make a decision which at ﬁrst sight would seem to be impossible with purely constructive techniques. Lemma 3.2.1. Let f be a strongly extensional mapping of a complete metric space X into a metric space Y , and let (xn )n1 be a sequence in X converging to a limit x. Then for all positive numbers α, β with α < β, either ρ(f (xn ), f (x)) > α for some n or ρ(f (xn ), f (x)) < β for all n. Proof. Let (nk )k1 be a strictly increasing sequence of integers such that ρ(x, xn ) < 1/ (k + 1) for all n nk . For convenience, set n0 = 1. For each k 1 set sk = max{ρ(f (xn ), f (x)) : nk−1 n < nk }. Construct an increasing binary sequence (λk )k1 such that λk = 0 =⇒ ∀j k (sj < β) , λk = 1 =⇒ ∃j k (sj > α) . We may assume that λ1 = 0. Deﬁne a sequence (yk )k1 in X as follows. If λk = 0, set yk = x; if λk = 1 − λk−1 , choose ν k with nk−1 ν k < nk and ρ(f (xν k ), f (x)) > α, and set yj = xν k for all j k. Let i j. If λi = 0 or λj = 1, then yi = yj . If λi = 1 − λj , then there exists k with j < k i and λk = 1 − λk−1 ; so yj = x, and yi = yk = xν k for some ν k such that nk−1 ν k < nk and ρ(f (xν k ), f (x)) > α; whence 1 1 < . ρ(yi , yj ) = ρ (xν k , x) νk + 1 j It follows that ρ(yi , yj ) < 1/j whenever i j; so (yk )k1 is a Cauchy sequence and therefore converges to a limit y in X. Either ρ(f (x), f (y)) < α or ρ(f (x), f (y)) > 0. In the ﬁrst case, if λk = 1 − λk−1 , then y = xν k with ρ(f (xν k ), f (x)) > α, a contradiction; whence λk = 0 for all k. Then ρ(f (xn ), f (x)) < β for all n. In the second case, since f is strongly extensional, x = y. Choose a positive integer κ such that x = yκ . If λκ = 0, then x = yκ = x, a contradiction; whence λκ = 1 and there 2 exists n such that ρ(f (xn ), f (x)) > α.

Let P (n) be a property of positive integers n. We say that P (n) holds • eventually, or for all suﬃciently large n, if there exists N such that P (n) holds for all n N ; •

inﬁnitely often if for each n there exists m > n such that P (m) holds.

72

3 The λ-Technique

The foregoing lemma leads to a technique allowing us to decide between alternatives that happen eventually and those that happen inﬁnitely often. Lemma 3.2.2. (Ishihara’s second trick) Let f be a strongly extensional mapping of a complete metric space X into a metric space Y , and let (xn )n1 be a sequence in X converging to a limit x. Then for all positive numbers α, β with α < β, either ρ(f (xn ), f (x)) > α for inﬁnitely many n or else ρ(f (xn ), f (x)) < β for all suﬃciently large n. Proof. Let (nk )k1 be a strictly increasing sequence of positive integers such that ρ(x, xn ) < 1/k for all n nk . Successively applying Lemma 3.2.1 to the subsequence (xn )nnk , construct an increasing binary sequence (λk )k1 such that λk = 0 =⇒ ∃n nk (ρ(f (xn ), f (x)) > α) , λk = 1 =⇒ ∀n nk (ρ(f (xn ), f (x)) < β) . We may assume that λ1 = 0. Deﬁne a sequence (yk )k1 in X as follows. If λk = 0, choose ν k nk such that ρ(f (xν k ), f (x)) > α, and set yk = xν k ; if λk = 1−λk−1 , set yi = yk−1 for all i k. Then (yk )k1 is a Cauchy sequence: in fact, ρ(yi , yj ) < 2/j whenever i j. Let y be the limit of (yk )k1 in the complete space X. Either 0 < ρ(f (x), f (y)) or ρ(f (x), f (y)) < α. In the ﬁrst case, since f is strongly extensional, x = y, so there exists a positive integer κ with ρ(x, yκ ) > 1/κ. If λκ = 0, then yκ = xν κ for some ν κ nκ ; whence ρ(x, yκ ) < 1/κ, a contradiction. Thus λκ = 1, and therefore ρ (f (xn ), f (x)) < β eventually. In the case ρ(f (x), f (y)) < α, if there exists k such that λk+1 = 1 − λk , then y = xν k and ρ(f (xν k ), f (x)) > α, a contradiction; whence λk = 0 for all k, and therefore ρ(f (xn ), f (x)) > α inﬁnitely often. 2 A mapping f : X −→ Y between metric spaces is said to be sequentially nondiscontinuous if it has the following property: if (xn )n1 converges to x ∈ X and ρ (f (xn ) , f (x)) δ for all n, then δ 0. Clearly, a sequentially continuous mapping is sequentially nondiscontinuous. Ishihara’s ﬁrst application of his tricks was to the converse. Proposition 3.2.3. Let f be a mapping of a complete metric space X into a metric space Y. Then f is sequentially continuous if and only if it is strongly extensional and sequentially nondiscontinuous. Proof. Suppose that f is strongly extensional and sequentially nondiscontinuous. Let (xn )n1 be a sequence converging to x ∈ X, and let ε > 0. By Lemma 3.2.2, either there exists a subsequence (xnk )k1 of (xn )n1 such that ρ (f (xnk ), f (x)) > ε/2 for all k, or else ρ (f (xn ), f (x)) < ε for all suﬃciently large n. In the former case, the sequential nondiscontinuity of f shows that ε 0, which is absurd. We conclude that the latter case obtains. Hence f is sequentially continuous.

3.2 Ishihara’s Tricks

73

The converse is an immediate consequence of Proposition 3.1.5 and the observation made before this proposition. 2 Corollary 3.2.4. For linear mappings of a Banach space into a normed space, sequential continuity and sequential nondiscontinuity are equivalent. Proof. This follows immediately from Corollary 3.1.7 and Proposition 3.2.3.

2

We now lift Ishihara’s tricks into a general setting. This both clariﬁes the ideas underlying those lemmas and raises the possibility that some other applications of their proof techniques in constructive analysis are, in fact, corollaries of our general results. We begin with a generalisation of Ishihara’s ﬁrst trick (Lemma 3.2.1). Proposition 3.2.5. Let X be a complete metric space, let P, Q be subsets of X such that X = P ∪ Q, and let x be an element of X such that for each y ∈ X, either x = y or y ∈ / Q. Then for each sequence (xn )n1 in X that converges to x, either xn ∈ P for all n or else there exists N such that xN ∈ Q. Proof. Construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (xk ∈ P ), λn = 1 − λn−1 =⇒ xn ∈ Q. We may assume that λ1 = 0. If λn = 0, set yn = x; if λn = 1−λn−1 , set yk = xn for all k n. To see that (yn )n1 is a Cauchy sequence in X, let ε > 0 and compute N such that ρ(x, xn ) < ε for all n N. Let m n N. If λm = 0, then ym = yn = x. So we may assume that λm = 1. If λn = 0, then ρ(ym , yn ) = ρ(xm , x) < ε. If λn = 1, then ρ(ym , yn ) = ρ(xm , xn ) ρ(x, xm ) + ρ(x, xn ) < 2ε. Thus in all cases, ρ(ym , yn ) < 2ε. Since X is complete, the Cauchy sequence (yn )n1 converges to a limit y ∈ X. Either x = y or y ∈ / Q. In the ﬁrst case, choosing N such that x = yN , we see that λN = 1. In the second case, if there exists m such that λm = 1 − λm−1 , then 2 y = xm ∈ Q, a contradiction; whence λn = 0, and therefore xn ∈ P, for all n. To derive Lemma 3.2.1, assume the hypotheses of that lemma, deﬁne P = {y ∈ X : ρ(f (y), f (x)) < β} , Q = {y ∈ X : ρ(f (y), f (x)) > α} , and apply Proposition 3.2.5, noting that for all y ∈ Y,

74

3 The λ-Technique

• either ρ(f (y), f (x)) > 0 , in which case, by the strong extensionality of f, x = y; • or else ρ(f (y), f (x)) < α and therefore y ∈ / Q. In order to tackle the generalisation of Ishihara’s second trick (Lemma 3.2.2), we prove two lemmas, the ﬁrst of which is a variant of Proposition 3.2.5. Lemma 3.2.6. Let X be a complete metric space, let P, Q be subsets of X such that X = P ∪ Q, and let x be a point of X such that for each y ∈ X, either x = y or y ∈ / P. Then for each sequence (xn )n1 in X that converges to x, either xn ∈ P for all n or else there exists N such that xN ∈ Q. Proof. Construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (xk ∈ P ) , λn = 1 − λn−1 =⇒ xn ∈ Q. We may assume that λ1 = 0. If λn = 0, set yn = x; if λn = 1 − λn−1 , set yk = xn−1 for each k n. Then (yn )n1 is a Cauchy sequence in X and so converges to a limit y ∈ X. Either x = y or else y ∈ / P. In the ﬁrst case choose N such that ρ (xn−1 , y) > 12 ρ (x, y) for all n N, and suppose that λN = 0. If λm = 1 − λm−1 for some m > N, then y = xm−1 and so 0 = ρ (xm−1 , y) > 12 ρ (x, y), a contradiction. Hence λn = 0 for all n N and therefore for all n; but this implies that y = x, another contradiction. Thus, in fact, λN = 1 and there exists n N such / P we must have λn = 0, and that xn ∈ Q. On the other hand, in the case y ∈ 2 therefore xn ∈ P, for all n. Lemma 3.2.7. Let X be a complete metric space, let P, Q be subsets of X such that X = P ∪ Q, and let x be an element of X. Suppose that for any sequence (xn )n1 converging to x in X, either there exists N such that x = xN or else xn ∈ / Q for all n. Then for each y ∈ X, either x = y or y ∈ / Q. Proof. Applying the hypotheses to the sequence (x, x, . . .) in X, we see that x ∈ / Q. Given y in X, construct a (perforce increasing) binary sequence (λn )n1 such that 1 , n 1 . λn = 1 =⇒ ρ (x, y) > n+1 λn = 0 =⇒ ρ (x, y)

α} , and, for convenience, P = {y ∈ X : ρ (f (x), f (y)) < α} , Q = {y ∈ X : ρ (f (x), f (y)) > α/2} . Then X = P ∪ Q and for each y ∈ X, either x = y or y ∈ / Q. Similarly, X = P ∪ Q and for each y ∈ X, either x = y or y ∈ / Q . It follows from the last statement and Proposition 3.2.5 that if (yn )n1 is any sequence converging to x in X, then either yn ∈ P for all n or else there exists N such that yN ∈ Q . Hence either / Q for all n or else there exists N such that x = yN . Thus P and Q satisfy the yn ∈

76

3 The λ-Technique

hypotheses of Proposition 3.2.8, from which we immediately deduce the conclusion of Ishihara’s second trick. We have now given a number of illustrations of the merit of the λ-technique as a replacement for omniscience principles. The technique will reappear throughout the book, where it will be used, for example, to construct best approximations from ﬁnite-dimensional spaces, to locate certain convex sets in a normed space, and to discuss the range of an operator with an adjoint on a Hilbert space.

Exercises 1. Using the λ-technique, show that if Ra = {ax : x ∈ R} is closed, then a = 0 or a = 0. 2. Let X be a complete metric space, and f : X −→ R a sequentially continuous mapping such that inf f exists. Suppose that to each ε > 0 there corresponds δ > 0 such that if x, y ∈ X and max {f (x), f (y)} < δ, then ρ(x, y) < ε. Prove that there exists a ∈ X such that if f (a) > 0, then inf f > 0. 3. A linear mapping T : X −→ Y between normed spaces is said to be well-behaved if T x = 0 whenever x ∈ ∼ ker(T ). Prove that the statement “every linear mapping between normed spaces is well-behaved” is equivalent to Markov’s principle. Prove also that every linear mapping from a Banach space into a normed space is well-behaved. 4. Prove that if Corollary 3.1.7 (and hence a fortiori Proposition 3.1.6) holds without the hypothesis that X is complete, then we can derive Markov’s principle. 5. Prove de Morgan’s rule for metric complements: If (S n )n1 is a sequence of Sn is complete and located subsets of a metric space X such that S = n1 located, then −S = −Sn . n1

6. A subset S of a metric space X is said to be uniformly almost located if there exists a strictly decreasing sequence (δ n )n1 of positive numbers converging to 0 such that the following holds: for each x ∈ X there exists y ∈ S such that for each n, if ρ (x, y) > δ n , then ρ(x, S) > δ n+1 . Prove that if S is an inhabited, uniformly almost located subset of a locally totally bounded space, then S is located. 7. Let X be a complete metric space, and a a point of X such that s = sup {ρ(a, x) : x ∈ X} exists. Prove that for each r > 1 there exists b ∈ X such that s rρ(a, b).

3.2 Ishihara’s Tricks

77

8. Let T be a sequentially continuous linear mapping of a Banach space X into a normed space Y, and let B be the unit ball of the range of T. Suppose that for each ε > 0, either there exists x ∈ ∼T −1 (B) with x < ε or else ∼T −1 (B) is bounded away from 0 (that is, there exists δ > 0 such that x > δ whenever T (x) ∈ B). Prove that T is a bounded linear mapping. 9. Let T : X −→ Y be a sequentially continuous linear map between normed spaces such that ker(T ) is located. Prove that if x0 ∈ X and T (x0 ) = 0, then ρ(x0 , ker(T )) > 0. 10. Let X be a metric space, and let S, T be subsets of X such that for each s ∈ S, each t ∈ T, and each ε > 0, either s = t or there exists y ∈ S ∩T with ρ(t, y) < ε. Prove that if S ∩ T is complete, then S and T intersect sharply in the following sense: if x ∈ ∼ (S ∩ T ), then for each s ∈ S and each t ∈ T, either x = s or x = t. 11. A sequence (xn )n1 in a metric space is said to be weakly discriminating if for all positive a, b with a < b, either ρ (xn , x1 ) < b for all n or else ρ(xn , x1 ) > a for some n. Prove that every totally bounded sequence in a metric space is weakly discriminating. Let f : X −→ Y be a function between metric spaces that maps Cauchy sequences to weakly discriminating sequences. Prove that f is strongly extensional. 12. A sequence (xn )n1 in a metric space is called an LEM-Cauchy sequence if ¬¬ (xn )n1 is a Cauchy sequence . Prove that the following are equivalent conditions on a mapping f between metric spaces: (a) f maps convergent sequences to weakly discriminating LEM-Cauchy sequences. (b) f is sequentially continuous. 13. Let f be a mapping of a complete metric space X into a metric space Y. Suppose that f is strongly extensional and that it is sequentially discontinuous at some point x ∈ X in the following sense: there exist ε > 0 and a sequence (xn )n1 converging to x such that ρ(f (x), f (xn )) > ε for each n. Prove that LPO holds. 14. Let f be a strongly extensional mapping of a complete metric space X into a compact metric space Y, and let (xn )n1 be a sequence converging to x in X. Prove that the sequence (f (xn ))n1 has a convergent subsequence. (Hint: First prove that if LPO holds, then every sequence in a compact metric space has a convergent subsequence. Then use Ishihara’s tricks and Exercise 13.)

78

3 The λ-Technique

15. Let (Un )n1 be a sequence of located open sets in a complete metric space X Un is inhabited. Prove that there such that the metric complement of U = n1

exist a point x∞ in −U and an increasing binary sequence (λn )n1 such that for each n, if λn = 0, then x∞ ∈ Un , and if λn = 1, then −Uk is inhabited for some k n. Use this result to prove that Markov’s principle is equivalent to the following statement (which is classically equivalent to Baire’s theorem): If (Un )n1 is a sequence of located open subsets of a complete metric space X such that Un is inhabited, then there exists n such that −Un is inhabited. − n1

Notes A variation of the argument at the beginning of the chapter shows that if for each z ∈ C, there exists θ ∈ [0, 2π) such that z = |z| eiθ , then LLPO holds. The classical proposition “sequential continuity implies pointwise continuity” is equivalent to the essentially nonconstructive principle BD-N that we discuss later, in Section 6.3. When taken with the Church–Markov–Turing thesis, Exercise 13 shows that it is impossible that there exist a sequentially discontinuous function from R to R, since LPO is false in RUSS. It is, however, a far cry from showing that the existence of a sequentially discontinuous mapping on R implies LPO to proving that every mapping f : R −→ R is pointwise continuous. The latter can be done with the aid of either Brouwer’s continuity principle or else both the Church–Markov–Turing thesis and Markov’s principle; see Chapters 3 and 5 of [34]. In connection with Exercise 1, Fred Richman has shown us the following choicefree proof that if Ra is complete, then either a = 0 or a = 0. Assume that Ra is complete. For each ε > 0 we have either |a| < ε or else |a| > 0; in the latter case,

a |a| = ± ∈ Ra. |a|

Since ε > 0 is arbitrary, |a| ∈ Ra = Ra and there exists r such that |a| = ra. Pick a positive integer N > r. Either |a| > 0 or |a| < 1/N 2 . In the latter case, if a = 0, then |a| |r| |a| = |ra| = |a| = > N |a| , |a| so |r| > N, a contradiction; whence a = 0. Exercise 3 and an extended version of Proposition 3.1.6 originated in [24].

3.2 Ishihara’s Tricks

79

Exercise 6 deals with a weak converse of Bishop’s lemma. There seems to be no obvious use for such converses, in spite of the extreme value of Bishop’s lemma itself. The notion of “weakly discriminating”, and Exercises 11 and 12, come from [37]. Exercise 14 produces a constructive substitute for the highly nonconstructive Bolzano–Weierstrass theorem; it ﬁrst appeared in [30]. For more constructive analyses of the Bolzano–Weierstrass theorem see [27] and [22]. For more on Baire’s theorem, see [34] (Chapter 2) and Chapter 6 below.

4 Finite-Dimensional and Hilbert Spaces

A mind that is stretched by a new experience can never go back to its old dimensions. —Oliver Wendell Holmes, attrib.

We ﬁrst examine ﬁnite-dimensional spaces, including an application of the λ-technique to the problem of ﬁnding best approximations by elements of a ﬁnite-dimensional subspace. We then introduce Hilbert spaces, which are natural generalisations of ﬁnite-dimensional Euclidean spaces.

4.1 Finite-Dimensional Spaces Let X be a linear space equipped with a compatible inequality. Finitely many n λi ei = 0 for all vectors e1 , . . . , en in X are said to be linearly independent if scalars λ1 , . . . , λn such that n i=1

n i=1

i=1

|λi | > 0. In that case, if the λi are scalars such that

λi ei = 0, then λi = 0 for each i. We say that the space X is ﬁnite-dimensional

if either X = {0} or else it contains ﬁnitely many linearly independent vectors e1 , . . . , en such that for each x ∈ X there exist scalars λ1 , . . . , λn for which x=

n

λi ei .

(4.1)

i=1

In the ﬁrst case we say that X has dimension 0 or is 0-dimensional. In the second case we say that X has dimension n or is n-dimensional, we call {e1 , . . . , en } a basis of/for X, and we say that the space X is spanned by, or is the (linear) span of, the set {e1 , . . . , en } . We denote the dimension of a ﬁnite-dimensional space X

82

4 Finite-Dimensional and Hilbert Spaces

by dim(X). Since any two bases of X have the same number of vectors, dim(X) is well deﬁned. If {e1 , . . . , en } is a basis for X, then for each x ∈ X, the coordinates λi in the representation (4.1) are uniquely deﬁned; so there are well-deﬁned, clearly linear, n ui (x)ei for each x in X. We coordinate functionals ui : X −→ K such that x = i=1

shall prove that each of the coordinate functionals is a bounded linear map relative to any norm on X. Lemma 4.1.1. Let e be a nonzero vector in a normed space X. Then the subspace Y = Ke is locally compact, and ∼Y = −Y. −1

Proof. The mapping λ λ e e is an isometric isomorphism of K onto Y, so Y is locally compact and therefore located. Since locally compact spaces are complete, it follows from Bishop’s lemma (Proposition 3.1.1) that ∼Y = −Y. 2

Lemma 4.1.2. If Y is a ﬁnite-dimensional subspace of a normed space X, then Y is located and ∼Y = −Y. Proof. The case dim(Y ) = 0 is trivial, and the case dim(Y ) = 1 is dealt with in Lemma 4.1.1. Assume that all subspaces of dimension at most n 1 in all normed spaces are located, and that their complements and metric complements coincide. Consider an (n + 1)-dimensional space Y with basis vectors e1 , . . . , en+1 . Let Z be the n-dimensional space spanned by {e1 , . . . , en } . First note that for all λ1 , . . . , λn n in K, since 1 + |λi | > 0, the linear independence of the vectors e1 , . . . , en+1 gives en+1 =

n

i=1

i=1

λi ei . Hence en+1 belongs to ∼Z and therefore, by our induction

hypothesis, to −Z; whence en+1 X/Z = ρ (en+1 , Z) > 0. It follows that Ken+1 is a 1-dimensional subspace of X/Z. Now, for each x ∈ X, ρ (x, Y ) = inf {x − ten+1 − z : t ∈ K, z ∈ Z} = inf {ρ (x − ten+1 , Z) : t ∈ K} = inf x − ten+1 X/Z : t ∈ K , which exists, since, by Lemma 4.1.1, the 1-dimensional subspace Ken+1 is located in X/Z. Hence Y is located in X. Now consider an element x of ∼Y. For all t ∈ K and z ∈ Z we have x = ten+1 +z and therefore x − ten+1 = z. Thus x − ten+1 ∈ ∼Z, and therefore, by our induction hypothesis,

4.1 Finite-Dimensional Spaces

83

x − ten+1 X/Z = ρ (x − ten+1 , Z) > 0. This shows that x =X/Z λen+1 ; whence x is in the complement of Ken+1 relative to the quotient norm on X/Z. Again applying our induction hypothesis, this time to the 1-dimensional subspace Ken+1 of X/Z, and denoting by ρX/Z the distance corresponding to the quotient norm on X/Z, we see that 0 < ρX/Z (x, Ken+1 ) = ρ (x, Y ) , so x ∈ −Y. Hence ∼Y ⊂ −Y and therefore ∼Y = −Y. This completes the induction step. 2 Proposition 4.1.3. Let {e1 , . . . , en } be a basis for an n-dimensional normed space X. Then the corresponding coordinate functionals are bounded linear functionals. Proof. In the case n = 1, the sole coordinate functional is the linear mapping λe1 λ, which, being an isometry (that is, distance preserving), is trivially continuous. n ui (x)ei . In order to prove Now consider the case n 2. For each x ∈ X write x = i=1

that the coordinate functional uk is bounded, we may relabel, if necessary, to take k = n. Let Z be the (n − 1)-dimensional subspace of X with basis {e1 , . . . , en−1 } . Since, as in the proof of Lemma 4.1.2, en ∈ ∼Z, it follows from that lemma that ρ (en , Z) > 0 and hence that Ken is a 1-dimensional subspace of X/Z. Moreover, for each x ∈ X we have n ui (x)ei , Z = ρ (un (x)en , Z) xX/Z = ρ i=1

= un (x)en X/Z = |un (x)| en X/Z and therefore |un (x)| =

1 1 xX/Z x . en X/Z en X/Z

Thus 1/ en X/Z is a bound for the linear functional un .

2

Corollary 4.1.4. Every linear mapping from a ﬁnite-dimensional normed space into a normed space is bounded. Proof. Let u be a linear mapping of a ﬁnite-dimensional normed space X into a normed space Y. We may assume that dim(X) > 0. Let {e1 , . . . , en } be a basis of X, and, using Proposition 4.1.3, compute a common bound c > 0 for the corresponding coordinate functionals ui (1 i n). For each x ∈ X we have

84

4 Finite-Dimensional and Hilbert Spaces

n n u(x) = u ui (x)ei = ui (x)u(ei )

n

i=1

ui (x)u(ei ) =

i=1

c

n

n

i=1

|ui (x)| u(ei )

i=1

u(ei ) x .

i=1

2

So u is a bounded linear map.

Corollary 4.1.5. Let {e1 , . . . , en } be a basis for an n-dimensional normed space X. Then the canonical mapping (λ1 , . . . , λn )

n

λi ei

(4.2)

i=1

is a bounded linear injection of the Euclidean space Kn onto X, and its inverse is a bounded linear injection. Proof. It is routine to verify the linearity of this map and its inverse, and that the maps are injective. The continuity of the mappings follows immediately from Corollary 4.1.4. 2 Proposition 4.1.6. A ﬁnite-dimensional normed space is locally compact. Proof. The 0-dimensional case is trivial. Consider a ﬁnite-dimensional normed space X with a basis {e1 , . . . , en }. Let u : Kn −→ X be the bounded linear mapping deﬁned at (4.2). By Corollary 4.1.5, u−1 is a bounded linear map. Hence for any bounded subset B of X, u−1 (B) is a bounded subset of the locally compact space Kn and is therefore contained in a compact subset K of Kn . Since u is injective and its inverse mapping is uniformly continuous, it is strongly injective. It follows from Proposition 2.2.16 that u(K), which clearly contains B, is a compact subset of X. 2 Corollary 4.1.7. The unit ball of a ﬁnite-dimensional normed space is compact. Proof. Let X be a ﬁnite-dimensional normed space, and B its unit ball. By Proposition 4.1.6, there exists a compact subset K of X that contains B. Now, for each x ∈ X, ρ(x, B) = max {0, 1 − x} exists, so B is totally bounded (by Proposition 2.2.10). Being also closed in K, it is complete and therefore compact. 2

4.1 Finite-Dimensional Spaces

85

Corollary 4.1.8. Every linear mapping from a ﬁnite-dimensional normed space into a normed space is normed. Proof. Let u be a linear mapping from a ﬁnite-dimensional normed space X into a normed space Y. We may assume that dim(X) 1. By Corollary 4.1.4, u is a bounded linear mapping and hence is uniformly continuous on X. In particular, this implies that u maps the compact (by Corollary 4.1.7) unit ball of X onto a totally bounded subset of Y ; so u = sup {u (x) : x ∈ X, x 1} 2

exists.

Two norms , on a vector space X are said to be equivalent if both the identity mapping from (X, ) onto (X, ) and its inverse are continuous; since those mappings are linear, it follows from Proposition 2.3.3 that and are equivalent norms on X if and only if there exist positive constants a, b such that a x x b x for all x ∈ X. Corollary 4.1.9. Any two norms on a ﬁnite-dimensional space are equivalent.

Proof. If and are two norms on a ﬁnite-dimensional space X, then, by Corollary 4.1.4, both the identity mapping from (X, ) to X, and its inverse are bounded, and hence continuous, linear mappings. 2 We want to prove the converse of Corollary 4.1.7. This requires three lemmas, the second of which will have several applications later in the book. Lemma 4.1.10. Let X be a normed space, Y an n-dimensional subspace of X with basis {e1 , . . . , en }, and e a vector such that ρ(e, Y ) > 0. Then the linear span of Y ∪ {e} is (n + 1)-dimensional, with basis {e1 , . . . , en , e}. Proof. We need only prove that the vectors e1 , . . . , en , e are linearly independent. n+1 |λi | > 0. Either To this end, consider elements λi (1 i n + 1) of K such that λn+1 = 0 or

n i=1

i=1

|λi | > 0. In the ﬁrst case,

n λi ei + λn+1 e |λn+1 | ρ (e, Y ) > 0. (4.3) i=1 n λ e In the second case, since e1 , . . . , en are linearly independent, i i > 0. Hence i=1 either λn+1 e > 0, so that λn+1 = 0 and we have (4.3); or else λn+1 e < n λi ei and therefore i=1

86

4 Finite-Dimensional and Hilbert Spaces

n n λi ei + λn+1 e λi ei − λn+1 e > 0. i=1

i=1

n

Thus in all cases we have

i=1

λi ei + λn+1 e = 0.

2

Lemma 4.1.11. Let S be the span of a ﬁnitely enumerable set {x1 , . . . , xn } in a normed space X, and let ε > 0. Then there exists a ﬁnite-dimensional subspace Y of S such that ρ (xi , Y ) < ε for each i. Proof. Setting X0 = {0} , suppose that for some k < n we have constructed ﬁnitedimensional subspaces X0 ⊂ X1 ⊂ · · · ⊂ Xk ⊂ S such that ρ (xi , Xk ) < ε for 1 i k. Then either ρ (xk+1 , Xk ) < ε, in which case we set Xk+1 = Xk , or else ρ (xk+1 , Xk ) > 0. In the latter case we take Xk+1 to be the span of Xk ∪ {xk+1 } , which is ﬁnite-dimensional by the preceding lemma. This completes the inductive construction of the ﬁnite-dimensional subspace Xk+1 . It remains to take Y = Xn . 2

Lemma 4.1.12. (Riesz’s lemma) Let Y be a closed located subspace with an inhabited metric complement in a normed space X, and let 0 < θ < 1. Then there exists a unit vector x ∈ X such that x − y > θ for each y ∈ Y. Proof. Fix x0 ∈ −Y. Then 0 < r = ρ(x0 , Y ) < θ−1 r. Choosing y0 ∈ Y such that r x0 − y0 < θ−1 r, let x=

1 (x0 − y0 ). x0 − y0

Then x = 1. Also, for each y ∈ Y, y0 + x0 − y0 y ∈ Y. Hence x0 − y0 x − y = x0 − (y0 + x0 − y0 y) ρ(x0 , Y ) = r and therefore x − y

r > θ. x0 − y0 2

4.1 Finite-Dimensional Spaces

87

Proposition 4.1.13. A locally totally bounded normed space is ﬁnite-dimensional. Proof. Let B be the unit ball of a locally totally bounded normed space X. Then B is located in X. By Proposition 2.2.18, B is locally totally bounded; being bounded, it is therefore totally bounded. Let {x1 , . . . , xn } be a 1/4-approximation to B, and, using Lemma 4.1.11, construct a ﬁnite-dimensional subspace Y of X such that ρ (xi , Y ) < 1/4 for each i. Let ξ be any point of X, and suppose that ρ (ξ, Y ) > 0. Then, by Riesz’s lemma, there exists a unit vector x ∈ X such that x − y > 1/2 for all y ∈ Y. But this is absurd: for since x ∈ B, there exists i such that x − xi < 1/4 and therefore ρ(x, Y ) < 1/2. We conclude that ρ (ξ, Y ) = 0; so ξ is in the closure of Y. But Y, being ﬁnite-dimensional and therefore locally compact (Proposition 4.1.6), is closed in X. Hence ξ ∈ Y. Since ξ ∈ X is arbitrary, 2 we see that X = Y. Corollary 4.1.14. A located subspace of a ﬁnite-dimensional normed space is ﬁnite-dimensional. Proof. Let Y be a located subspace of a ﬁnite-dimensional normed space X. Since, by Proposition 4.1.6, X is locally compact, we see from Proposition 2.2.18 that Y is locally totally bounded. Hence, by Proposition 4.1.13, Y is ﬁnite-dimensional. 2

By a convex combination of ﬁnitely many elements x1 , . . . , xn of X we mean n n λi xi where each λi 0 and λi = 1. The convex hull a point of the form i=1

i=1

of a subset S of X is the closure of the set of all points of X that are convex combinations of points of S. It is straightforward to show that the convex hull of S is the intersection of all convex subsets of X that contain S. According to Exercise 11 of Chapter 2, a located subset of R may not have its metric complement located. However, things are diﬀerent when the set is convex. Proposition 4.1.15. If S is a located convex subset of the product normed space Rn such that −S is inhabited, then −S is located. Proof. Consider any x in Rn , and for each r > 0 let B 1 (x, r) be the closed ball with centre x and radius r relative to the product metric ρ. Then B 1 (x, r) is an n-dimensional cube with centre x and sides of length 2r. Consider any two real numbers α, β such that α < β. Let v1 , . . . , v2n be the vertices of B 1 x, 12 (α + β) . It is left as an exercise to show that there exists δ > 0 such that for all points w1 , . . . , w2n with ρ(vi , wi ) < δ (1 i 2n ) , the convex hull of {w1 , . . . , w2n } contains B 1 (x, α). Either ρ(vi , S) < δ for all i, or ρ(vi , S) > 0 for some i. In the ﬁrst case, for each i in {1, . . . , 2n } choose wi in S such that ρ(vi , wi ) < δ; then S contains the convex hull of {wi , . . . , w2n } and

88

4 Finite-Dimensional and Hilbert Spaces

therefore contains B 1 (x, α). Thus ρ(x, y) α for all y in −S. On the other hand, if for some i we have ρ(vi , S) > 0, then vi ∈ −S and ρ(x, vi ) 12 (α + β) < β. Thus ρ(x, S) exists, by the constructive least-upper-bound principle. 2 Although we do not develop integration theory in this book, it is worth pointing out here that for convex subsets of Rn there is a close link between locatedness and Lebesgue measurability: a convex subset of Rn with inhabited interior is located if and only if it is Lebesgue measurable [17]. Informally, this result shows that a convex subset of Rn with inhabited interior can be located if and only if its size can be calculated.

4.2 Best Approximation Let V be an inhabited subspace of a metric space X, and let a, b be elements of X, V respectively. We say that b is a best approximation, or closest point, to a in V if ρ (a, b) ρ (a, v) for each v ∈ V. In that case, ρ (a, V ) exists and equals ρ (a, b). We call V proximinal in X if each element of X has a best approximation in V , in which case V is located. The fundamental theorem of classical approximation theory says that a ﬁnitedimensional subspace of a normed space is proximinal. Constructively, this theorem implies LLPO (see Exercise 1). However, by introducing the idea of at most one object even without knowing in advance that there exists one, we can produce a good constructive version of the fundamental theorem. Let X, V, and a be as in the ﬁrst paragraph of this section. We say that a has at most one best approximation in V if for all distinct points v, v in V, there exists x ∈ V such that max {ρ(a, v), ρ(a, v )} > ρ(a, x). We call V quasiproximinal if each point of X with at most one best approximation in V actually has a (perforce strongly unique) best approximation in V. Our destination in this section is the following constructive fundamental theorem of approximation theory. Theorem 4.2.1. Every ﬁnite-dimensional subspace of a real normed space is quasiproximinal. We defer the proof until we have prepared the pathway with some preliminary results, one of which has a proof that uses the λ-technique discussed in Chapter 3. Lemma 4.2.2. Let {e1 , . . . , en } be a basis for an n-dimensional subspace V of a normed space X, let 1 m < n, and let W be the subspace of X with basis {e1 , . . . , em } . Then the span of {em+1 , . . . , en } is an (n − m)-dimensional subspace of the quotient space X/W.

4.2 Best Approximation

Proof. The proof is relegated to Exercise 4.

89

2

Lemma 4.2.3. Let x, e be elements of a real normed space X with e = 0, and for each δ > ρ (x, Re) write Sδ = {t ∈ R : x − te δ} . If Sδ is compact, then it is a proper compact interval [m, M ] in R. Moreover, x − me = δ = x − M e .

(4.4)

Proof. Suppose that Sδ is compact, with inﬁmum m and supremum M. By Corollary 2.2.14, there exists δ such that ρ(x, Re) < δ < δ and Sδ is compact. Let m and M denote the inﬁmum and supremum, respectively, of Sδ . The uniform continuity of the map t x − te on R ensures that (4.4) and x − m e = δ = x − M e hold. Since Sδ ⊂ Sδ and δ < δ, it follows that m < m M < M and therefore m < M. Now consider any t ∈ [m, M ] . Writing τ=

t−m 0, M −m

we have t = (1 − τ ) m + τ M. Hence x − te (1 − τ ) x − me + τ x − M e = (1 − τ ) δ + τ δ = δ and so t ∈ Sδ . Thus Sδ = [m, M ] .

2

Lemma 4.2.4. Let x, e be elements of a real normed space X with e = 0, and let d 0. Suppose that max {x − te , x − t e} > d whenever t, t are distinct real numbers. Then there exists τ ∈ R such that x − τ e > d entails ρ (x, Re) > d. Proof. Without loss of generality, we may assume that ρ (x, Re) < d + 1. With Sδ as in Lemma 4.2.3, we construct an increasing binary sequence (λn )n1 , sequences (an )n1 , (bn )n1 of real numbers, and a decreasing sequence (δ n )n1 of positive numbers with the following properties: Sδn is the proper compact interval [an , bn ] ; if λn = 0, then ρ (x, Re) < δ n < d + 1/n, and if n 2, 0 < bn − an

2 (bn−1 − an−1 ) ; 3

(4.5)

90

4 Finite-Dimensional and Hilbert Spaces

if λn = 1, then ρ (x, Re) > d, an = an−1 , bn = bn−1 , and δ n = δ n−1 . To begin the construction, set λ1 = 0 and, using Corollary 2.2.14, choose δ 1 > 0 such that ρ (x, Re) < δ 1 < d+1 and Sδ1 is compact. By Lemma 4.2.3, Sδ1 is a proper compact interval [a1 , b1 ]. Now suppose we have constructed λn−1 , an−1 , bn−1 , and δ n−1 with the applicable properties. If λn−1 = 1, set λn = 1, an = an−1 , bn = bn−1 , and δ n = δ n−1 .

(4.6)

If λn−1 = 0, write 2 1 2 1 an−1 + bn−1 e , x − an−1 + bn−1 e r = max x − . 3 3 3 3 By our hypotheses, r > d; so either ρ (x, Re) > d or else ρ (x, Re) < min {r, d + 1/n} . In the ﬁrst case we deﬁne our numbers as at (4.6). In the second case we set λn = 0 and choose δ n such that 1 ρ (x, Re) < δ n < min δ n−1 , r, d + n and Sδn is compact; then by Lemma 4.2.3, Sδn is a proper compact interval [an , bn ] contained in Sδn−1 . Since r > δ n , either x − 1 an−1 + 2 bn−1 e > δ n 3 3 x − 2 an−1 + 1 bn−1 e > δ n . 3 3

or

Taking, for example, the ﬁrst case, and using the convexity of Sδn , we see that either 1 2 Sδn ⊂ an−1 , an−1 + bn−1 3 3 or 1 2 an−1 + bn−1 , bn−1 . Sδ n ⊂ 3 3 Hence (4.5) holds. This completes our inductive construction. We see from (4.5) that for n 2, if λn = 0, then n−1 2 (b1 − a1 ) 0 < bn − an 3 and therefore 0 an+1 − an

n−1 2 (b1 − a1 ) . 3

(4.7)

4.2 Best Approximation

91

Clearly, this last inequality also holds if λn = 1. It follows that (an )n1 is a Cauchy sequence and therefore converges to a limit τ in R. Supposing that x − τ e > d, compute a positive integer N such that x − aN e > d + 1/N. If λN = 0, then x − aN e = δ N < d +

1 , N

a contradiction. We conclude that λN = 1 and hence that ρ (x, Re) > d.

2

We now have the proof of Theorem 4.2.1. Proof. Let V be a ﬁnite-dimensional subspace of a normed space X over R, and let a be a point of X with at most one best approximation in V. If V has dimension 0, then there is nothing to prove. If dim(V ) = 1, then V = Re for some e = 0 in V, and we can apply Lemma 4.2.4 with x = a and d = ρ (a, V ) to construct τ ∈ R such that a − τ e = ρ (a, V ). Now let n be a positive integer, and suppose we have proved the desired result for all n-dimensional subspaces of real normed spaces. Consider the case where V has a basis {e1 , . . . , en+1 } , and let Y = Ren+1 . By Lemma 4.2.2, V /Y is an ndimensional subspace of X/Y with basis {e1 , . . . , en }; moreover, for each x ∈ X we have (4.8) ρ (x, V ) = inf x − vX/Y : v ∈ V . Next, note that (*) for each v ∈ V there exists α ∈ R such that if a − v − αen+1 > ρ (a, V ) , then a − vX/Y = ρ (a − v, Ren+1 ) > ρ (a, V ) . For if t, t are distinct real numbers, then (v + ten+1 ) − (v + t en+1 ) = |t − t | en+1 = 0 and therefore, by our hypotheses, max {a − v − ten+1 , a − v − t en+1 } > ρ (a, V ) . So we can apply Lemma 4.2.4 with x = a − v, e = en+1 , and d = ρ (a, V ) to compute the desired α. With v, α as above, now let v be a point of V distinct from v, and compute α ∈ R such that if a − v − α en+1 > ρ (a, V ) , then

a − v X/Y = ρ (a − v , Ren+1 ) > ρ (a, V ) . We have

92

4 Finite-Dimensional and Hilbert Spaces

(v + αen+1 ) − (v + α en+1 ) = (v − v ) + (α − α ) en+1 v − v X/Y > 0, so, by our hypotheses, max {a − v − αen+1 , a − v − α en+1 } > ρ (a, V ) . It follows from (4.8) and the deﬁning properties of α, α that max a − vX/Y , a − v X/Y > ρ(a, V ) = inf a − vX/Y : v ∈ V . We have now shown that a has at most one best approximation in the n-dimensional subspace V /Y of X/Y. By our induction hypothesis and (4.8), there exists v0 ∈ V such that ρ (a − v0 , Ren+1 ) = a − v0 X/Y = inf a − vX/Y : v ∈ V = ρ (a, V ) . Applying (*) once more, we compute t0 ∈ R such that a − v0 − t0 en+1 = ρ (a, V ) . Hence v0 + t0 en+1 is the best approximation to a with respect to the original norm on X, and our inductive proof is complete. 2

Theorem 4.2.1 is classically equivalent to its classical counterpart. For if V is a ﬁnite-dimensional subspace of the real normed space X and we suppose that some a ∈ X has no best approximation in V, then a clearly has at most one best approximation in X. Therefore, by Theorem 4.2.1, it has one, which is a contradiction. For applications of Theorem 4.2.1, see [15].

4.3 Hilbert Spaces As we did for metric and normed spaces, we assume that the reader is familiar with the deﬁnitions and elementary properties of an inner product and an inner product space. In particular, we are not going to prove that x = x, x deﬁnes the norm associated with a given inner product on a vector space X over K, and that the Cauchy–Schwarz inequality |x, y| x y and the parallelogram identity

4.3 Hilbert Spaces 2

2

2

93

2

x + y + x − y = 2 x + 2 y

then hold. The latter inequality enables us to prove that an inner product space X is uniformly convex, as follows: If 0 < δ < 1, and x, y are unit vectors in X with 1 2 x + y > 1 − δ, then 2

2

2

2

2

x − y = 2 x + 2 y − x + y < 4 − 4 (1 − δ) , which can be made as small as we please, independently of x and y, by a suitable initial choice of δ. The simplest example of an inner product space is, of course, the Euclidean space Kn with the usual inner product given by x, y =

n

xi yi∗ ,

i=1

where x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) . Another example is the space l2 (K), consisting of all sequences x = (xn )n1 in K that are square summable in the sense ∞ 2 that |xn | converges in R; in this case we work with termwise operations and n=1

with the inner product deﬁned by x, y =

∞

xn yn∗ .

n=1

It is left to the exercises to prove that this is indeed an inner product. An inner product space that is complete with respect to its norm is called a Hilbert space. Adding completeness to the inner product leads to a structure with powerful geometrical properties such as the following. Theorem 4.3.1. Let S be a closed, located subspace of a Hilbert space H. Then for each x ∈ H, there exists a strongly unique element P x of S such that x − P x = ρ(x, S). Moreover, P x is the strongly unique element y of S such that x − y, s = 0 for all s ∈ S. Proof. Fixing x ∈ H, let d = ρ (x, S) and choose a sequence (sn )n1 in S such that d = limn→∞ x − sn . Using the parallelogram identity and the convexity of S, we have 2

sm − sn = (sm − x) − (sn − x)

2

2 sm + sn 2 2 − x = 2 sm − x + 2 sn − x − 4 2 2

2

2 sm − x + 2 sn − x − 4d2 2 2 = 2 sm − x − d2 + 2 sn − x − d2 −→ 0 as m, n −→ ∞.

94

4 Finite-Dimensional and Hilbert Spaces

Hence (sn )n1 is a Cauchy sequence. Since H is complete and S is closed, this sequence converges to a limit P x ∈ S. The continuity of the norm on H now gives x − P x = d. Moreover, if y ∈ S and y = P x, then, again by the parallelogram identity and convexity, we have 2

2

0 < P x − y = (P x − x) − (y − x) 2 Px + y 2 2 − x = 2 P x − x + 2 y − x − 4 2 2 2 2 P x − x − d2 + 2 y − x − d2 2 = 2 y − x − d2 , so x − y > d. It follows that P x is the strongly unique closest point to x in S. Next note that for all y ∈ S and λ ∈ K, x − P x + λy, x − P x + λy d2 = x − P x, x − P x and therefore

2

2

|λ| y + 2 Re (λ∗ x − P x, y) 0.

(4.9)

Suppose that Re x − P x, y = 0. By choosing a suﬃciently small real number λ with λ Re x − P x, y < 0, we can contradict (4.9). It follows that Re x − P x, y = 0. A similar argument shows that Im x − P x, y = 0; whence x − P x, y = 0. Finally, if y ∈ S and y = P x, then since y − P x ∈ S, 2

0 < y − P x = y − P x, y − P x = x − P x, y − P x − x − y, y − P x = − x − y, y − P x , so x − y, y − P x = 0. It follows that P x is the strongly unique element y of S such that x − y, s = 0 for all s ∈ S. 2

The mapping P : H −→ S deﬁned in Theorem 4.3.1 is called the projection of H onto S, and for each x ∈ H, the vector P x is the projection of the vector x onto S. The mapping P is linear: for if λ ∈ K and x, x ∈ H, then since for all y ∈ S, λx + x − (λP x + P x ), y = λ x − P x, y + x − P x , y = 0, it follows from the uniqueness of the projection of λx + x onto S (Theorem 4.3.1) that P (λx + x ) = λP x + P x . For each y ∈ S, since y − y, s = 0 for all s ∈ S, we see from the uniqueness of the projection of y into S that P y = y; whence P maps H onto S. Since P x ∈ S, we have P 2 x = P (P x) = P x, so P is idempotent; that is, P 2 = P. Also,

4.3 Hilbert Spaces

95

2

x = x, x = x − P x + P x, x − P x + P x = x − P x, x − P x + x − P x, P x + P x, P x 2

2

= x − P x + P x , since x − P x, P x = 0 (by Theorem 4.3.1). Hence P x x , and so 1 is a bound for P. On the other hand, if S contains a nonzero vector x, then P x = x ; it follows that in this case, P is normed and P = 1. Two subsets S, T of an inner product space are said to be orthogonal if x, y = 0 for all x ∈ S and y ∈ T ; we then write S ⊥ T. The orthogonality relation is symmetric, S ⊥ T ⇐⇒ T ⊥ S, and any family (Ti )i∈I of subsets of X satisﬁes Ti ⇐⇒ ∀i ∈ I (S ⊥ Ti ) . S⊥ i∈I

A vector x is orthogonal to the subset S if {x} ⊥ S, in which case we write x ⊥ S; two vectors x, y are said to be orthogonal vectors if x ⊥ {y}, in which case we write x ⊥ y and we have the following generalisation of Pythagoras’s theorem: 2

2

2

x + y = x + y . We deﬁne the orthogonal complement of a subset S of X to be the set S ⊥ = {x ∈ X : x ⊥ S} , which is easily seen to be a closed linear subspace of X. If 0 ∈ S, then S ∩ S ⊥ = {0} . It readily follows that if S is a closed, located subspace of a Hilbert space H, with P the corresponding projection, and if x ∈ H, then the decomposition x = P x + (x − P x) is the unique expression of x as the sum of a vector in S and a vector orthogonal to S. Denoting by I the identity operator x x on H, we prove that I − P is the projection of H onto the subspace S ⊥ . For each y ∈ S ⊥ , the vector z = x − P x − y belongs to S ⊥ , so 2

2

2

2

2

2

x − y = P x + z = P x + z P x = x − (x − P x) . Since x − P x ∈ S ⊥ , we conclude from the uniqueness part of Theorem 4.3.1 that x − P x is the projection of x into S ⊥ . We need at this point to clarify what we mean by saying that the series λi ei i∈I

converges to the sum x and by writing x=

i∈I

λi ei ,

(4.10)

96

4 Finite-Dimensional and Hilbert Spaces

where the index set I on the right is not necessarily countable: we mean that for each ε > 0 there exists a ﬁnitely enumerable subset F of I with the property that if G is a ﬁnitely enumerable subset of I that contains F, then λi ei < ε. x − i∈G

In the case I = N+ , this condition is equivalent to the usual one for convergence of ∞ λn en to x. the series n=1

A family (ei )i∈I of vectors in a Hilbert space H is said to be orthonormal if ei ⊥ ej whenever i = j, and for each i, either ei = 1 or ei = 0. Such a family is called an orthonormal basis if each vector x ∈ H can be written uniquely in the form (4.10) with (λi )i∈I a family of elements of K such that λi = 0 whenever ei = 0. The scalar λi is then called the ith coordinate of x relative to the orthonormal basis. If (ei )i∈I is a ﬁnitely enumerable orthonormal family in H, then it spans a ﬁnite= 1 for each i. dimensional subspace of H. To see this, wemay assume that ei |λi | > 0, and let x = λi ei . Choose Let (λi )i∈I be a family of scalars such that i∈I

j ∈ I such that λj = 0. Then λi ei , ej = λj ej , ej = λj = 0, x, ej =

i∈I

i∈I

so x = 0. Thus the vectors ei (i ∈ I) are linearly independent and therefore span a ﬁnite-dimensional space. Classically, using (an equivalent of) the axiom of choice, we can prove that every Hilbert space has an orthonormal basis of unit vectors. Constructively we avoid the axiom of choice by adding separability to the hypotheses on H, by relaxing the requirements to allow basis vectors to be 0, and by using the Gram–Schmidt orthogonalisation process embodied in the proof of our next result. Proposition 4.3.2. Every separable Hilbert space has a countable orthonormal basis. If (en )n1 is such a basis, relative to which x has coordinates α1 , α2 , . . . and ∞ y has coordinates β 1 , β 2 , . . . , then αn = x, en for each n, x, y = αn β ∗n , and 2

x =

∞ n=1

n=1

2

|αn | .

Proof. Let (an )n1 be a dense sequence in H. The idea of the proof is to construct inductively the orthonormal sequence (en )n1 to ensure that at stage n there exists

4.3 Hilbert Spaces

97

zn in the subspace Sn of H spanned by {e1 , . . . , en } such that an − zn < 1/n. For convenience, set e0 = 0 and S0 = {0}. Assume that en has already been constructed. Being ﬁnite-dimensional, Sn is closed and located, so the projection Pn of H onto Sn exists. Either an+1 − Pn an+1 < 1/ (n + 1) or else an+1 = Pn an+1 . In the ﬁrst case, set en+1 = 0. In the second, deﬁne a unit vector by en+1 =

1 (an+1 − Pn an+1 ) . an+1 − Pn an+1

Note that en+1 is orthogonal to e0 , . . . , en since (by Theorem 4.3.1) an+1 − Pn an+1 is orthogonal to Sn , and that an+1 = Pn an+1 + an+1 − Pn an+1 en+1 is in the subspace of H spanned by {e1 , . . . , en+1 }. This completes the induction. Now consider the vectors x, y ∈ H. For each positive integer k we have unique representations k αn en Pk x = n=1

and Pk y =

k

β n en ,

n=1

with αn = β n = 0 whenever en = 0. For n k, αn =

k

αi ei , en = Pk x, en = x, en − x − Pk x, en = x, en .

i=1

Likewise, β n = y, en . Thus αn and β n do not depend on any k = n. Now, Pk x, Pk y =

k

αn β ∗n .

n=1

Since the sequence (an )n1 is dense in X and ρ (ak , Sk ) < 1/k, we see that both ρ (x, Sk ) and ρ (y, Sk ) approach 0 as k −→ ∞; whence Pk x −→ x and Pk y −→ y as k −→ ∞. Thus ∞ αn β ∗n . x, y = lim Pk x, Pk y = k→∞

n=1 2

Finally, if we take y = x, we obtain x =

∞ n=1

2

|αn | .

2

An elementary lemma will enable us to characterise dimensionality in terms of orthonormal bases.

98

4 Finite-Dimensional and Hilbert Spaces

Lemma 4.3.3. Let S be a ﬁnite-dimensional subspace of a separable Hilbert space H, and let (en )n1 be an orthonormal basis of H. Then there exists a positive integer N such that ρ (en , S) > 0 whenever n N and en = 1. Proof. By Corollary 4.1.7, there exists a ﬁnite 1/2-approximation {x1 , . . . , xm } to ∞ 2 |xk , en | the unit ball of H. Since, by the preceding proposition, the series n=1

converges, |xk , en | −→ 0 as n −→ ∞; so we can compute N such that |xk , en |

1 , 4

so ρ (en , S) = en − P en en − xk − P en − xk > 0, as we required.

2

Proposition 4.3.4. Let (en )n1 be an orthonormal basis of a separable Hilbert space H. Then H is ﬁnite-dimensional if and only if en = 0 for all suﬃciently large n. Proof. If H is ﬁnite-dimensional, then taking S = H in Lemma 4.3.3, we obtain N such that en = 0 for all n N. If, conversely, such N exists, then H is the ﬁnite-dimensional space of all linear combinations of those vectors e1 , . . . , eN that are nonzero. 2

Let X be a linear space with an inequality compatible with its linear structure. We say that X is inﬁnite-dimensional if the complement of each ﬁnite-dimensional subspace of X is inhabited. It follows from Bishop’s lemma (Proposition 3.1.1) that a normed space X is inﬁnite-dimensional if and only if for each ﬁnite-dimensional subspace S of X there exists x ∈ X such that ρ (x, S) > 0. Proposition 4.3.5. Let (en )n1 be an orthonormal basis of a separable Hilbert space H. Then H is inﬁnite-dimensional if and only if en = 0 for inﬁnitely many n. Proof. Suppose that H is inﬁnite-dimensional, and consider any positive integer N. Let S be the ﬁnite-dimensional subspace of H consisting of all linear combinations

4.3 Hilbert Spaces

99

of e1 , . . . , eN , and let P be the projection of H onto S. There exists x ∈ H such that ρ(x, S) > 0 and therefore, by Proposition 4.3.2, ∞

2

2

2

|x, en | = x − P x = ρ (x, S) > 0.

n=N +1

Hence there exists n > N such that x, en = 0 and therefore, by the Cauchy– Schwarz inequality, en = 0. Conversely, if there exists a strictly increasing sequence (nk )k1 of positive integers such that enk = 0 for each k, consider any ﬁnite-dimensional subspace S of H. Choose N as in Lemma 4.3.3. Then choose k such that nk > N. Since enk = 1, 2 we must have ρ (enk , S) > 0.

We now introduce a construction of new Hilbert spaces from old. Let H1 , H2 be Hilbert spaces, and for all x = (x1 , x2 ) and y = (y1 , y2 ) in the product vector space H1 × H2 deﬁne x, y = x1 , y1 + x2 , y2 . This deﬁnes an inner product with respect to which H1 × H2 is a Hilbert space, called the direct sum of H1 and H2 , and denoted by H1 ⊕ H2 . This construction helps us to prove the most general case of our next theorem. We say that a linear functional u on a Hilbert space H is represented by a vector a ∈ H if u(x) = x, a for all x ∈ H. The following Riesz representation theorem tells us that a functional is representable by a vector in this way if and only if it is normed. Theorem 4.3.6. A bounded linear functional u on a Hilbert space H is normed if and only if there exists a ∈ H such that u(x) = x, a

(x ∈ H) .

(4.11)

In that case, a is strongly unique: if y ∈ H and y = a, then there exists x ∈ H such that u(x) = x, y . Moreover, u = a . Proof. Suppose ﬁrst that there exists a vector a with property (4.11). Then, by the Cauchy–Schwarz inequality, |u(x)| a x for each x ∈ H. On the other hand, given ε > 0, we have either a > 0 or a < ε. In the ﬁrst case, 1 a = a > a − ε. u a In the second, u(0) = 0 > a − ε. Since ε > 0 is arbitrary, it follows that u exists and equals a .

100

4 Finite-Dimensional and Hilbert Spaces

Suppose, conversely, that u is normed. To begin with, take u > 0. Then ker(u) is located, by Proposition 2.3.6. Let P be the projection of H on ker(u), choose y ∈ H with u (y) > 0, and deﬁne x0 =

1 (y − P y) . u(y)

Then u(x0 ) = 1, so x − u(x)x0 ∈ ker(u). Since x0 ⊥ ker(u), we have 2

0 = x − u(x)x0 , x0 = x, x0 − u(x) x0 , from which it follows that u(x) = x, a with a=

1

2 x0 .

x0

It remains to remove the condition that u > 0. In doing so, we need to bear in mind that we have not ruled out the possibility that H = {0}. To deal with this, we consider the direct sum H ⊕ K, on which we deﬁne a bounded linear functional v by v(x, ζ) = u(x) + ζ. We ﬁrst observe that u is represented by the vector a if and only if v is represented by the vector (a, 1). Since v > 0, it follows from the ﬁrst part of the proof that it will suﬃce to prove that v is normed. To this end, let 0 < α < β and set ε = 12 (α + β). If u > 0, then, by the ﬁrst part of the proof, u is represented by a unique vector a ∈ H; whence v is represented by (a, 1) and is therefore normed. So we may assume that u < ε. Then either u + 1 > α + ε or u + 1 < β. In the ﬁrst case, v(0, 1) = 1 > α and (0, 1) = 1. In the second case, for each (x, ζ) with (x, ζ) 1 we have |v(x, ζ)| |u(x)| + |ζ| u + 1 < β. It now follows from the least-upper-bound principle that v exists.

2

The second part of the foregoing proof contains an argument showing that the sum of two normed linear functionals on a Hilbert space is normed. Note that in general, the sum of two normed linear functionals on an arbitrary nontrivial normed space need not be normed. By an operator on a Hilbert space H we mean a linear mapping of H into itself. The set of bounded operators on H is denoted by B(H). Classically, the Riesz representation theorem enables us to prove, for a given element T of B(H), the existence of the adjoint T ∗ , which has the deﬁning property T x, y = x, T ∗ y

(x, y ∈ H) .

(4.12)

4.3 Hilbert Spaces

101

Indeed, given y ∈ H, we apply the Riesz representation theorem to the linear functional x T x, y to obtain a unique vector T ∗ y with the desired property. The constructive problem with this argument is that in order to apply Theorem 4.3.6, we require that the functional x T x, y be normed, which is something we cannot guarantee. In fact, as the following Brouwerian example shows, we cannot prove that a general normed operator on H has an adjoint. Let H be a complex Hilbert space with an orthonormal basis (en )n1 of unit vectors, and let (an )n1 be a binary sequence with at most one term equal to 1. ∞ Observe that for each x ∈ H the series an x, en converges absolutely. For, given ε > 0 and choosing N such that

n=1 ∞

n=N +1

have

k

|an x, en | =

n=N +1

n=1

an |x, en |

n=N +1

Hence the partial sums of ∞

k

∞ n=1

2

|x, en | < ε2 , for each k > N we

max

N +1nk

|x, en | < ε.

an |x, en | form a Cauchy sequence, and so the series

an x, en converges in C. Thus Tx =

∞

an x, en e1

n=1

deﬁnes a mapping—clearly an operator—from H to itself. It is left as an exercise to show that T is normed. Suppose that T ∗ exists. Then either T ∗ e1 = 0 or else T ∗ e1 < 1. In the ﬁrst case we can ﬁnd N such that T ∗ e1 , eN = 0. Suppose that aN = 0. If there exists n = N such that an = 1, then for all x, y ∈ H, T x, y = x, en e1 , y = x, y, e1 en and therefore

T ∗ y = y, e1 en .

(4.13)

∗

Hence T e1 , eN = 0, a contradiction. Thus an = 0 for all n = N and therefore for all n; whence T ∗ = 0, which is impossible since T ∗ e1 = 0. We conclude that aN = 1. In the case T ∗ e1 < 1, suppose that an = 1. Then (4.13) holds, so T ∗ e1 = en and therefore T ∗ e1 = 1, a contradiction. It follows that in this case we have an = 0 for all n. In view of this Brouwerian example, for any Hilbert space H and any not necessarily bounded operator T on H, we deﬁne the adjoint T ∗ , if it exists, by the equation (4.12); in which case we refer to T as jointed. It is then straightforward to show that T ∗ is an operator; that the adjoint of T ∗ is T ; and that any bound for T is one for T ∗ , and vice versa. Moreover, if S, T are jointed operators, then for each λ ∈ K, so are λS + T and ST, and

102

4 Finite-Dimensional and Hilbert Spaces ∗

(λS + T ) = λ∗ S ∗ + T ∗ , ∗ (ST ) = T ∗ S ∗ . For example, we have ST x, y = T x, S ∗ y = x, T ∗ (S ∗ y) = x, T ∗ S ∗ y . An operator T on H is called selfadjoint, or Hermitian, if T ∗ exists and equals T. The identity operator I is trivially selfadjoint. More generally, if P is the projection of H onto a closed, located subspace, then for all x, y ∈ H we have P x, y = P x, P y + P x, y − P y = P x, P y = P x, P y + x − P x, P y = x, P y , so P is selfadjoint. Conversely, if P is any bounded, idempotent, selfadjoint operator on H, then P is a projection. To see this, let V = {y ∈ H : P y = y} . It is clear that V is a linear subspace of H. If x ∈ H, then since P 2 x = P x, we have P x ∈ V. On the other hand, for each y ∈ V we have x − P x, y = x, y − P x, y = x, y − x, P y = 0, since P y = y. Hence x − P x is orthogonal to V, and therefore (since P x − y is in V) 2

2

x − y = x − P x + P x − y 2

2

2

= x − P x + P x − y x − P x . We now see that P x is a closest point to x in V. Hence V is located in H. Finally, the continuity of the bounded operator P ensures that V is closed; so P is the projection of H onto V. We end the chapter with an application of the Riesz representation theorem and a corollary, both due to Ishihara [58]. For this we need to know that a linear mapping T between normed spaces X, Y is deﬁned to be compact if T (B X (0, 1)) is a totally bounded subset of Y ; in that case, the norm of T exists, by Corollary 2.2.7. Every bounded linear mapping on a ﬁnite-dimensional normed space is compact. Proposition 4.3.7. Let T be a bounded linear mapping of a Hilbert space H into Cn , and for 1 i n let Pi : Cn −→ C be the ith projection mapping, deﬁned by Pi (z1 , z2 , . . . , zn ) = zi . Then T is compact if and only if Pi ◦ T is normed for each i.

4.3 Hilbert Spaces

103

Proof. If T is compact, then since the projection Pi is uniformly continuous, the set Pi ◦ T (B(0, 1)) is totally bounded, so Pi ◦ T exists. Suppose, conversely, that Pi ◦ T is normed for each i. By the Riesz representation theorem, for each i there exists ai ∈ H such that Pi ◦ T (x) = x, ai

(x ∈ H) .

Using Lemma 4.1.11, construct a ﬁnite-dimensional subspace X of H such that ρ(ai , X) < ε/2n for each i. Let P be the projection of H onto X. Since the restriction of T to X is a compact linear mapping, there exist x1 , x2 , . . . , xm in the unit ball BX of X such that {T x1 , T x2 , . . . , T xm } is an ε/2-approximation to T (BX ). Consider any x ∈ H with x 1. Working with the product norm on Cn , we have T (x − P x) = = =

n i=1 n i=1 n

|Pi ◦ T (x − P x)| x − P x, ai x, (I − P ) ai

i=1 n

n ε ε = . x ai − P ai < 2n 2 i=1 i=1

Choosing j such that T P x − T xj < ε/2, we now obtain T x − T xj T (x − P x) + T P x − T xj

0 is arbitrary, we conclude that T is a compact linear mapping on H. 2 Corollary 4.3.8. The sum of two compact operators on a Hilbert space is compact. Proof. Let S and T be compact operators on a Hilbert space H, and let ε > 0. Since S(BH ) and T (BH ) are totally bounded, it readily follows from Lemma 4.1.11 that there exists a ﬁnite-dimensional subspace X of H, with associated projection P, such that (4.14) (S + T )x − (P S + P T )x < ε (x ∈ BH ) . We may assume that X is nontrivial and so has a basis {e1 , . . . , en } . Writing P Sx =

n i=1

pi (x)ei , P T x =

n

qi (x)ei ,

i=1

we see from Proposition 4.3.7 that for each i, the linear functionals pi , qi on H are normed; whence pi + qi is normed (see the remark following the proof of the Riesz representation theorem). Since

104

4 Finite-Dimensional and Hilbert Spaces

(P S + P T ) (x) =

n

(pi + qi ) (x)ei ,

i=1

we see from Proposition 4.3.7 that P S + P T is compact. In view of (4.14), any ε-approximation to (P S + P T ) (BH ) is a 2ε-approximation to (S + T )(BH ). Hence S + T is compact. 2

Exercises 1. Let X be the space R2 with norm (x, y) = max {|x| , |y|} . Show that if the 1-dimensional subspace R (cos θ, sin θ) is proximinal in X for each θ ∈ R, then LLPO holds. 2. Let V be a nonzero linear subspace of a normed linear space X such that each x ∈ X has at most one closest point in V . Need V be located? 3. Prove that if every pair of nonzero vectors in R2 generates a ﬁnite-dimensional space, then LPO holds. 4. Prove Lemma 4.2.2. 5. Prove that a uniformly convex linear subspace of a normed space is proximinal. 6. A normed space X is said to be compactly generated if there exists a compact set K ⊂ X such that each point of X is a linear combination of ﬁnitely many points of K. Let X be a compactly generated Banach space, and Y a ﬁnite-dimensional subspace of X. Prove that either X = Y or else the metric complement of Y is inhabited. (Hint: Use the λ-technique.) 7. Let Y be a locally compact subspace of a complete metric space X such that each x ∈ X has a unique best approximation P x in Y. Prove that the mapping P is sequentially continuous on X. (Hint: First use Ishihara’s tricks.) 8. Let T be a bounded linear mapping of a normed space X onto a ﬁnitedimensional Banach space Y. Prove that there exists r > 0 such that BY (0, r) ⊂ T (BX (0, 1)). 9. Let T be a bounded linear mapping of a normed space X onto a ﬁnitedimensional Banach space Y. Prove that T is a compact linear mapping if and only if ker(T ) is located in X. 10. Let X be an inner product space. Prove that x = x, x deﬁnes a norm on X and that the Cauchy–Schwarz inequality |x, y| x y holds for all x, y ∈ X.

4.3 Hilbert Spaces

105

11. Let p 1, and let lp be the set of all sequences x = (xn )n1 in the ﬁeld K ∞ p that are p-summable, in the sense that |xn | converges. Equipping lp with n=1

termwise operations of addition and multiplication-by-scalars, prove that 1/p ∞ p |xn | xp = n=1

deﬁnes a norm on lp , and that lp is complete with respect to this lp -norm. In the case p = 2, show that the lp -norm arises from an inner product, as in Exercise 10. 12. Let S, T be orthogonal linear subspaces of a Hilbert space H such that S + T = {x + y : x ∈ S, y ∈ T } is dense in H. Prove that S and T are both located. 13. Let S be a linear subset of a Hilbert space H such that for each x ∈ H there exists y ∈ S with x − y orthogonal to S. Prove that S is closed and located. 14. Construct a Brouwerian counterexample to the statement that every selfadjoint operator on a Hilbert space H is normed. Does every normed operator on H have an adjoint? 15. Construct a Brouwerian counterexample to the statement that every selfadjoint operator on a Hilbert space that has located kernel also has located range. 16. Construct Brouwerian counterexamples to each of the following statements. (a) Every bounded linear functional on l2 (see Exercise 11) is compact. (b) If T is a bounded linear mapping of l1 into C2 such that Pi ◦ T is normed for i = 1, 2, then T is compact (where Pi is the ith projection of C2 onto C). 17. Construct a Brouwerian counterexample to the proposition that the sum of two compact linear mappings between normed spaces is compact. 18. Use the λ-technique as an alternative means of removing the restriction that u > 0 in the proof of the Riesz representation theorem. 19. Let R be an algebra of normed, jointed operators on a complex Hilbert space H, where the operation of multiplication on R is just the composition of operators. For each T ∈ R and each λ ∈ C deﬁne an operator T on the direct sum H ⊕ C by T (x, ζ) = (T x + λx, λζ) . Show that T is both selfadjoint and normed.

106

4 Finite-Dimensional and Hilbert Spaces

Notes Our deﬁnition of “linearly independent” is classically equivalent to the usual one: n λi ei = 0, then λi = 0 for each i. Constructively, the two deﬁnitions namely, if i=1

are equivalent if and only if MP holds. Bishop’s deﬁnition of “ﬁnite-dimensional” requires that the space have a norm from the outset. Our deﬁnition is more in the spirit of linear algebra. Theorem 4.2.1 can be applied to the case of Chebyshev approximation: best approximation of elements of C [0, 1], relative to the sup norm, by polynomials of degree at most n. However, in this special case a deeper analysis enables one to prove the existence of best approximations without the use of Theorem 4.2.1; moreover, that analysis reveals that the best approximation process is, as one would expect in a constructive context, continuous. See [15]. The induction step in the proof of Theorem 4.2.1 is a lot simpler if we are permitted to use Brouwer’s fan theorem; for we can then show that any uniformly continuous function from a compact metric space to the positive real numbers has a positive inﬁmum (see [34], Chapter 6). As Section 4.3 shows, the elements of Hilbert space theory require very few modiﬁcations to bring them into constructive line. However, we have to be careful about the requirement that the linear functionals be normed before we can apply the Riesz representation theorem, and about the possibility that a given bounded operator may not have an adjoint. For a diﬀerent proof of the Riesz representation theorem see [35]. Working with a direct sum, as was done in the proof of Theorem 4.3.6 in order to circumvent our inability to decide whether a space is trivial, is a useful technique in other applications; see [28]. In the proof that an idempotent, bounded, selfadjoint operator P on a Hilbert space is a projection, we used boundedness only to prove that the set V = {x ∈ H : P x = x} is closed. The Hellinger–Toeplitz theorem (see Chapter 6) shows that every selfadjoint operator P on H is sequentially continuous, a property strong enough to prove that the corresponding set S is closed. Thus, in fact, every idempotent selfadjoint operator on a Hilbert space is a (bounded linear) projection.

5 Linearity and Convexity

Every separation is a link. —Simone Weil, ‘Metaxu’, Gravity and Grace We begin the chapter by exploring some geometric aspects of convexity that are used later in the construction of one of the cornerstones of functional analysis: the separation theorem. In turn, this leads us to the Hahn–Banach extension theorem, which in its most general form for separable normed spaces produces only approximately norm-preserving extensions of normed linear functionals. We also give Ishihara’s version of the Hahn– Banach theorem, which provides a unique norm-preserving extension of a given normed linear functional in the case where the norm on the space is Gˆ ateaux diﬀerentiable. We then use the separation and Hahn–Banach theorems to explore the interplay between a normed space and its dual. In particular, we characterise certain linear functionals on spaces of bounded linear mappings. For our discussion of duality we develop the fundamentals of the theory of locally convex topological vector spaces.

5.1 Crossing Boundaries Suppose we start at a point ξ in the interior of a located subset C of a normed space X and move linearly towards a point z in the metric complement of C. Are we able to tell when we are crossing the boundary ∂C = C ∩ ∼C of C? In general, as is discussed in more detail in problems at the end of this chapter, the constructive answer is no. However, our geometric intuition suggests that when C is convex, we might succeed in pinpointing boundary crossing points. Our ﬁrst few lemmas are designed to lead us to a proof of the existence of the boundary crossing point, which depends continuously on the origin ξ and the terminus z of our path out of the convex set.

108

5 Linearity and Convexity

Lemma 5.1.1. Let C be a convex subset of a normed space X, let ξ ∈ C ◦ , and let r > 0 be such that B (ξ, r) ⊂ C. Let z = ξ, 0 < t < 1, and z = tξ + (1 − t) z. If the ball B(z, tr) intersects C, then B(z , t2 r) ⊂ C. Proof. Suppose that there exists y in B(z, tr) ∩ C. Let ζ ∈ B(z , t2 r) and 1 1 y + ζ . ξ = 1− t t Then

ξ − ξ = 1 − 1 z + 1 z − 1 − 1 y − 1 ζ t t t t 1 1 − 1 z − y + z − ζ t t 1 2 1 − 1 tr + t r < t t = r.

Hence ξ ∈ C. Since ζ = tξ + (1 − t) y, it follows by convexity that ζ ∈ C.

2

In the context of a vector space X over K we deﬁne intervals as follows: [x, y] = {tx + (1 − t) y : 0 t 1} , (x, y) = {tx + (1 − t) y : 0 < t < 1} , where x, y ∈ X. Lemma 5.1.2. Let C be an open convex subset of a normed space X such that C ∪ −C is dense in X, let ξ ∈ C, and let z ∈ −C. Then (C ∪ −C) ∩ [ξ, z] is dense in [ξ, z] . Proof. Fix ε > 0 and choose r such that 0

r/2 and y − z > r/2, so y = αξ + (1 − α) z for some α ∈ (0, 1) . Fix λ such that −α −ε , < λ < 0, (5.1) max ξ − z 1 − α and set

5.1 Crossing Boundaries

109

λ , λ−1 y1 = λξ + (1 − λ) y. t=

Then 0 < t < 1 and y = tξ + (1 − t) y1 . Also, y1 = [1 − (1 − λ)(1 − α)] ξ + (1 − λ) (1 − α) z, where, by (5.1), 0 < (1 − λ) (1 − α) < 1,

so y1 ∈ [ξ, z] . Now pick y ∈ C ∪ −C such that y − y < min r2 , t2 r . Take ﬁrst the case y ∈ C. By our choice of r, the point ξ + 1r (y − y ) belongs to C. Since C is convex, 1 1 r y + ξ + (y − y ) y = 1+r 1+r r also belongs to C. Moreover, y = and

1 r ξ+ y ∈ [ξ, y] ⊂ [ξ, z] , 1+r 1+r

y − y =

r r ξ − y ξ − z < ε. 1+r 1+r

We are left with the case y ∈ −C to dispose of. But then B(y, t2 r) intersects −C, so, by the preceding lemma, B(y1 , tr) ∩ C = ∅; whence y1 ∈ −C. Since y − y1 = |λ| ξ − y ε by (5.1), the proof is complete.

2

Lemma 5.1.3. Let X be a normed space, let x1 , x2 be distinct points of X, and let x3 = λx1 + (1 − λ) x2 with λ = 0, 1. For all α, β > 0, if x − x1 < α/|λ| and y − x2 < /β/|1 − λ|, then λx + (1 − λ) y − x3 < α + β. Proof. For such x and y we have λx + (1 − λ) y − x3 |λ| x − x1 + |1 − λ| y − x2 , from which the result follows almost immediately.

2

110

5 Linearity and Convexity

In the presence of convexity, the preceding, seemingly innocent, lemma turns out to be a powerful tool. Here is a ﬁrst example of its use. Lemma 5.1.4. If C is an inhabited, open, convex subset of a normed space X, then −C is dense in ∼C. Proof. Fixing ξ ∈ C, choose r > 0 such that B (ξ, r) ⊂ C. Consider any z ∈ ∼C and ε > 0. Setting ε t=1+ z − ξ and x2 = tz + (1 − t) ξ, we see that x2 − z = ε and that ξ=

1 −t z+ x2 . 1−t 1−t

It now suﬃces to show that x2 ∈ −C. Taking x1 = z, x3 = ξ, and λ = t/(t − 1) in Lemma 5.1.3, we see that if y − x2 < r (t − 1), then −t 1 0 such that B (ξ, r) ⊂ C and B (z, r) ⊂ −C, now observe that if r , t=1− ξ − z then 0 < t < 1 and ξ − (tξ + (1 − t) z) = r; whence tξ + (1 − t) z ∈ C. Thus m1−

r < 1. ξ − z

On the other hand, since z − (1 − t) ξ + tz = r and therefore zt ∈ −C, we see from (*) that m

r > 0. ξ − z

By the deﬁnition of m as an inﬁmum, the point h(z) = mξ + (1 − m) z belongs to the closure of C. Again applying Lemma 5.1.2, we can ﬁnd t ∈ (0, m) such that zt belongs to C ∪ −C and is arbitrarily close to zm . By the deﬁnition of / C, and so zt ∈ −C. Thus zm is in the closure of −C and hence m, we have zt ∈ in ∂C. The uniqueness part of (b) will follow immediately once we have proved (c) and (d).

112

5 Linearity and Convexity

Now observe that if m < t 1, then by the deﬁnition of m as an inﬁmum, there exists t such that m < t < t and zt ∈ C; since zt is in [ξ, zt ], it is in the convex set C. This proves (c). To dispose of (d), let 0 t < m and let η be the midpoint of the segment [ξ, zm ] . Then, by (c), η ∈ C, so there exists δ > 0 such that B (η, δ) ⊂ C. Also, η = λzm + (1 − λ) zt for some λ > 1. Consider any point y with y − zt < δ/(λ − 1). Applying Lemma 5.1.3 with x1 = zm , x2 = zt , and x3 = η, we see that λzm + (1 − λ) y − η < δ and therefore λzm + (1 − λ) y ∈ C. If also y ∈ C, then 1 1 y ∈ C, zm = (λzm + (1 − λ) y) + 1 − λ λ which is absurd since zm ∈ ∂C and C is open. Thus δ ⊂ ¬C B zt , λ−1 and therefore zt ∈ −C. This proves (d). To prove the continuity of the boundary crossing map on C × −C, ﬁx ξ ∈ C, z ∈ −C, and ε > 0. Using (c) and (d), choose a, b, s such that 0 < a < m < b < 1, s > 0, B (za , s) ⊂ −C ∩ B (zm , ε) , and B (zb , s) ⊂ C ∩ B (zm , ε) . Consider points z ∈ −C and ξ ∈ C with s max z − z , ξ − ξ < min 2

For each t ∈ [0, 1] set

1 1 1 1 , , , a b 1−a 1−b

.

zt = tξ + (1 − t) z .

Taking x1 = ξ, x2 = z, x = ξ , and y = z , and applying Lemma 5.1.3 with λ = a, we obtain za − za < s and therefore za ∈ −C ∩B(zm , ε). On the other hand, applying Lemma 5.1.3 with λ = b, we see that zb − zb < s and therefore zb ∈ C ∩ B(zm , ε). It follows that zγ(ξ ,x) is in the segment (za , zb ), which lies in B(zm , ε); whence 2 zγ(ξ ,z ) − zγ(ξ,z) < ε. For ﬁxed ξ ∈ C, we call the mapping z zγ(ξ,z) in the foregoing proposition the boundary crossing map of C relative to ξ. We shall use the existence of exact boundary crossings out of convex sets in the next section.

5.2 Separation Theorems

113

5.2 Separation Theorems Developing the theme of convexity, in this section we approach the fundamental theorems on the separation of points and convex sets by hyperplanes. In turn, this material will lead us in the next section to the Hahn–Banach theorem, one of the cornerstones of functional analysis. A subset C of a vector space X over K is called a cone if for all x, y ∈ C and all t > 0, both x + y and tx belong to C. In that case, C is convex. The closure of a cone is a cone, as is the intersection of two cones. If K is a convex subset of X, then the set c(K) = {tx : x ∈ K, t > 0} is a cone. For clearly, if x ∈ c(K), then tx ∈ c (K) for all t > 0; whereas if x1 , x2 ∈ c(K), then x1 = t1 y1 and x2 = t2 y2 for some y1 , y2 ∈ K and some t1 , t2 > 0, so x1 + x2 = (t1 + t2 ) z with z=

t2 t1 y1 + y2 ∈ K. t1 + t2 t1 + t2

We call c (K) the cone generated by the convex set K. If X is a normed space and K is open, then so is c(K). Lemma 5.2.1. Let K be a bounded located subset of a normed space X, a ∈ X, and let τ > 0. Then the set S = {tx + (1 − t) a : x ∈ K, 0 < t < τ } is located. Proof. Choose R > 0 such that x R for all x ∈ K. Fixing x0 ∈ X, note ﬁrst that for each t > 0, f (t) = ρ (x0 , {tx + (1 − t) a : x ∈ K}) exists and equals

tρ

t−1 1 x0 + a, K . t t

If also t > 0, then since x0 − tx − (1 − t) a x0 − t x − (1 − t ) a + (a + x) |t − t | for all x ∈ X, we see that f (t) f (t ) + (a + R) |t − t | . It follows that the mapping f is uniformly continuous on R+ ; whence ρ (x0 , S) exists as the inﬁmum of f on the totally bounded interval (0, τ ) . 2

114

5 Linearity and Convexity

Lemma 5.2.2. Let K be a bounded, located, convex subset of a normed space X such that ρ(0, K) > 0, and let a ∈ X be such that ρ(−a, c(K)) > 0. Then there exists r > 0 such that tx + (1 − t) a r for all x ∈ K and t > 0. Proof. Let δ=

ρ(0, K) . 3 (a + ρ(0, K))

Given x in K and t > 0, we have either |1 − t| > δ or |1 − t| < 2δ. In the ﬁrst case, t > δρ(−a, c(K)). x + (−a) tx + (1 − t) a = |1 − t| 1 − t In the case |1 − t| < 2δ, tx + (1 − t) a t x − |1 − t| a (1 − 2δ) ρ (0, K) − 2δ a 1 = ρ(0, K). 3 Setting

1 r = min δρ(−a, c(K)), ρ(0, K) > 0, 3

we see that tx + (1 − t) a r in either case.

2

Lemma 5.2.3. Let K be a bounded, located, convex subset of a normed space X such that ρ(0, K) > 0. Then c (K) is located. Proof. Fix x0 ∈ X. For each t > 0 and each x ∈ K we have x0 − tx |t| x − x0 . Hence ρ (x0 , tK) tρ (0, K) − x0 −→ ∞ as t −→ ∞. Compute τ > 1 such that ρ (x0 , tK) > ρ (x0 , K)

(t > τ − 1) .

(5.2)

Then d = inf {ρ (x0 , tK) : 0 < t < τ } exists, by the case a = 0 of Lemma 5.2.1. Since τ > 1 and therefore d ρ (x0 , K), 2 it follows from (5.2) that ρ (x0 , c (K)) exists and equals d.

5.2 Separation Theorems

115

Lemma 5.2.4. Let K and L be open cones in a normed space X such that K ∪ L is dense in X and K ⊂ ∼L. Then (a) K ⊂ −L and L ⊂ −K, (b) K ∪ −K and L ∪ −L are dense in X, and (c) K and L have a common boundary—namely, K ∩ L. If also L = {−x : x ∈ K}, then ∂K is a subspace of X. Proof. Since L ⊂ ∼∼L ⊂ ∼K, we see that K ∪ ∼K is dense; whence, by Lemma 5.1.4, K ∪−K is dense. On the other hand, K ⊂ ∼L and K is open, so K ⊂ −L and therefore K ⊂ −L. But −L is open and K ∪ L is dense, so −L ⊂ K and therefore, by Lemma 5.1.4, ∼L = −L ⊂ K. Hence K = −L = ∼L. Interchanging the roles of K and L, we obtain L ∪ −L dense and L = −K = ∼K. We now have ∂K = K ∩ ∼K = K ∩ L = ∼L ∩ L = ∂L. Since K, L, and therefore their closures are all cones, so is ∂K. Now suppose also that L = {−x : x ∈ K}, and consider x ∈ ∂K = K ∩ L. There exist points y ∈ K and z ∈ L arbitrarily close to x. Then −y ∈ L, −z ∈ K, and these two points are arbitrarily close to −x; whence −x belongs to ∂K. It follows that ∂K is a subspace of X. 2

Lemma 5.2.5. Let K be a located subset of a normed space X, and let r > 0. Then the set Kr = {x ∈ X : ρ(x, K) r} is located, and for each x0 ∈ X, ρ(x0 , Kr ) = max {0, ρ(x0 , K) − r} .

(5.3)

Proof. If x ∈ Kr and ε > 0, then there exists y ∈ K such that x − y < r + ε; so x0 − x x0 − y − x − y ρ(x0 , K) − r − ε. Since ε > 0 is arbitrary, we have x0 − x ρ(x0 , K) − r

(x ∈ Kr ) .

(5.4)

116

5 Linearity and Convexity

On the other hand, choosing z ∈ K such that x0 − z < ρ (x0 , K) + ε, we have either x0 − z < r, in which case (5.3) holds with each side equal to 0, or else x0 − z > 0. In the latter case, writing r (x0 − z) , ζ = min 1, x0 − z we have z + ζ ∈ Kr and

1 − min 1, x0 − (z + ζ) =

r x0 − z

(x0 − z)

= max {0, x0 − z − r} < max {0, ρ(x0 , K) − r} + ε. Since ε > 0 is arbitrary, it follows from this and (5.4) that (5.3) holds.

2

By a half-space of a normed space X we mean a convex subset K such that ∂K is a hyperplane and the set {x ∈ X : x ∈ K ∨ −x ∈ K} is dense in X. We are now ready for the basic separation theorem. Our proof illustrates an important observation about classical proofs using Zorn’s lemma (see [65]): for separable spaces it is often possible to replace such a proof by a constructive one that uses an induction argument. Theorem 5.2.6. Let X be a separable normed space, K0 a bounded, located, open, convex subset of X such that ρ(0, K0 ) > 0, and x0 a point of X such that −x0 ∈ K0 . Then there exists an open half-space K of X such that K0 ⊂ K, ρ(x0 , K) > 0, and the boundary of K is a located subspace of X that is a hyperplane with associated vector x0 . Proof. Let (xn )n1 be a dense sequence in X. The basic idea of the proof is this: we carry out a succession of located convex enlargements of K0 such that for n 1, the cone generated by the nth enlargement Kn is close to at least one of the points xn and −xn , and such that the union of the cones c(Kn ) is the desired open half-space. The idea may seem simple, but the details are rather complicated. To be more precise, we construct bounded, located, open, convex subsets K0 ⊂ K1 ⊂ K2 ⊂ · · · of X such that the following properties obtain for n 1: (a) ρ(x0 , c(Kn )) > (1 − 2−n ) ρ (x0 , c(Kn−1 )) , (b) ρ (0, Kn ) > 0,

5.2 Separation Theorems

(c) max {ρ (xn , c(Kn )) , ρ (−xn , c(Kn ))}

0 and ρ(−xn , c (Kn−1 )) > 1/2n, we see from Lemma 5.2.2 that ρ(0, Kn+ ) > 0; similarly, ρ (0, Kn− ) > 0. Next, given z + ∈ c(Kn+ ) and z − ∈ c(Kn− ), ﬁnd t+ , t− > 0 and x+ , x− ∈ c(Kn−1 ) such that z + = t+ xn + x+ and z − = −t− xn + x− . We have t− x0 − z + + t+ x0 − z −

= t− x0 − t+ xn + x+ + t+ x0 − −t− xn + x−

t+ + t− x0 − t− x+ + t+ x−

+ t− t+ − + − = t + t x0 − + x + + x − − t +t t +t

t+ + t− ρ (x0 , c(Kn−1 ))

> t+ + t− 1 − 2−n−1 ρ (x0 , c(Kn−1 )) . Hence either

t− x0 − z + − 1 − 2−n−1 ρ (x0 , c(Kn−1 )) > 0 or

t+ x0 − z − − 1 − 2−n−1 ρ (x0 , c(Kn−1 )) > 0,

from which we obtain max x0 − z + , x0 − z − > 1 − 2−n−1 ρ (x0 , c(Kn−1 )) .

118

5 Linearity and Convexity

Taking the inﬁmum as z + , z − run over c(Kn+ ), c(Kn− ) respectively, we obtain max ρ(x0 , c(Kn+ )), ρ(x0 , c(Kn− )) 1 − 2−n−1 ρ (x0 , c(Kn−1 ))

> 1 − 2−n ρ (x0 , c(Kn−1 )) . Thus either

ρ(x0 , c(Kn+ )) > 1 − 2−n ρ (x0 , c(Kn−1 )) ,

in which case, noting that xn ∈ Kn+ and therefore ρ (xn , c(Kn+ )) = 0 < 1/n, we take Kn = Kn+ ; or else

ρ(x0 , c(Kn− )) > 1 − 2−n ρ (x0 , c(Kn−1 )) , when we take Kn = Kn− . This completes the inductive construction of Kn . The set K=

∞

c(Kn )

n=0

is easily shown to be an open convex cone. Moreover, we see from (a) that (ρ(x0 , c(Kn )))n1 is a Cauchy sequence, which therefore converges in R. Since we have c(Kn−1 ) ⊂ c(Kn ) for each n, the limit of the sequence is ρ(x0 , K). Now, for n 1, n n

1 1 −k −k 1− 1−2 > , 2 2 4 k=1

k=2

so, by (a), ρ(x0 , K) = lim ρ (x0 , c (Kn )) n→∞

1 ρ (x0 , c (K0 )) . 4

But by the convexity of K0 , ρ (x0 , c(K0 )) = inf {tx − x0 : t > 0, x ∈ K0 } t 1 x + (−x ) : t > 0, x ∈ K = inf (t + 1) 0 0 t + 1 t+1 ρ(0, K0 ). Hence ρ(x0 , K) > 0. Let L = {x ∈ X : −x ∈ K} . Then L is an open convex cone containing x0 . By (c), for each n, either ρ (xn , K) < 1/n or ρ (xn , L) < 1/n. So K ∪ L is dense in X. On the other hand, K ⊂ ∼L: for if x ∈ K and y ∈ L, then there exists n such that x ∈ c (Kn ) and −y ∈ c (Kn ); so there exist t > 0 and z ∈ Kn such that x − y = tz and therefore x − y tρ (0, Kn ) > 0. It now follows from Lemma 5.2.4 that L ∪ −L is dense in X, and the common boundary of K and L is the set

5.2 Separation Theorems

119

N = K ∩ L, which is a linear subspace of X. To prove that N is a hyperplane, let γ denote the boundary crossing map of L relative to its interior point x0 . Note that, by Lemma 5.2.4, K ⊂ −L; so, by Proposition 5.1.5, for each x ∈ K there exists a unique t ∈ (0, 1) such that γ (x) = tx0 + (1 − t) x and therefore

t 1 (−x0 ) + γ (x) . 1−t 1−t Similar considerations using the boundary crossing map of K relative to −x0 complete a proof that each x ∈ K ∪ L can be expressed in the form αx0 + βz with α ∈ R and z ∈ N. In fact, such an expression obtains for all x ∈ X. For since K ∪ L is dense, there exist a sequence (αn )n1 in R and a sequence (zn )n1 in N such that αn x0 + zn −→ x as n −→ ∞. Given ε > 0 and m > n, we have either |αm − αn | < ε or |αm − αn | > 0; in the latter case, since zm − zn ∈ N ⊂ K, 1 (zm − zn ) (αm x0 + zm ) − (αn x0 + zn ) = |αm − αn | x0 − |αm − αn | |αm − αn | ρ(x0 , K), x=

so |αm − αn |

1 (αm x0 + zm ) − (αn x0 + zn ) . ρ(x0 , K)

It follows that (αn )n1 is a Cauchy sequence and therefore converges to a limit α in R. Writing z = x − αx0 , we have z − zn x − (αn x0 + zn ) + |α − αn | x0 −→ 0 as n −→ ∞, so z ∈ N = N. Thus N is a hyperplane with associated vector x0 , and therefore (since also K ∪ L is dense in X) K is a half-space. Now, for each x ∈ K we have γ(x) ∈ N ⊂ K and x0 − γ (x) x0 − x . It follows that ρ (x0 , N ) exists, equals ρ(x0 , K), and is positive. For each x ∈ X, choosing α ∈ R and z ∈ N such that x = αx0 + z, we easily see that ρ (x, N ) exists 2 and equals |α| ρ(x0 , K). Hence N is located. Corollary 5.2.7. Under the hypotheses of Theorem 5.2.6, there exists a normed linear functional u on X such that u = 1 and Re u is positive on K0 . Proof. First take the case K = R. Let K be as in the conclusion of Theorem 5.2.6, let

120

5 Linearity and Convexity

L = {−x : x ∈ K} , recall that K ⊂ −L, and let N be the common boundary of K and L. Since N is a located hyperplane, we see from Propositions 2.3.4 and 2.3.6 that there exists a normed linear functional v on X such that ker(v) = N and v(x0 ) = 1. Deﬁning −1 u = − v v, we obtain a normed linear functional on X with norm 1 such that ker(u) = N and u(x0 ) < 0. Let γ be the boundary crossing map of L relative to x0 . Given x ∈ K, choose λ ∈ (0, 1) such that γ(x) = λx + (1 − λ) x0 . Then u(γ(x)) = 0, so 1−λ u(x0 ) > 0. u(x) = − λ Hence u is positive on K and therefore on K0 . Now consider the case K = C. By the case just considered, there exists a real linear functional v : X −→ R on the real linear space X such that v has norm 1 and is positive on K0 . The following lemma then shows that u(x) = v(x) − iv(ix) deﬁnes a linear functional u : X −→ C on the complex linear space X such that u = 1 and Re u is positive on K0 . 2 Lemma 5.2.8. If X is a normed linear space over C, then there is a one-one correspondence between real linear functionals v on X and complex linear functionals u on X, given by (5.5) u(x) = v(x) − iv(ix). Moreover, if either u or v is normed, then both are and their norms are equal. Proof. If u is a complex linear functional, then u = v + iw for unique real linear functionals v and w. But iv(x) − w(x) = iu(x) = u(ix) = v(ix) + iw(ix), so w(x) = −v(ix). Conversely, if v is a real linear functional, then (5.5) deﬁnes a linear functional u that respects multiplication by real numbers; it is easily checked that u(ix) = iu(x). The ﬁnal conclusion, about the norms of u and v, is a straightforward consequence of the identity |u(x)| = sup{|v(λx)| : |λ| = 1}, (5.6) which we now establish. If |λ| = 1, then |u(x)| = |λu(x)| = |u(λx)| |Re u(λx)| = |v(λx)| . On the other hand, if ε > 0, then either |u(x)| − ε < |v(x)|, or else u(x) = 0. In the latter case, if we let λ = |u(x)|/u(x), then

5.2 Separation Theorems

121

u(λx) = λu(x) = |u(x)| ∈ R, so |u(x)| = Re u(λx) = v(λx). Notice that in this case the supremum in (5.6) is achieved. 2

The following result is the full form of the separation theorem. Theorem 5.2.9. Let A and B be bounded convex subsets of a separable normed space X such that the algebraic diﬀerence {y − x : x ∈ A, y ∈ B} is located and the mutual distance d = inf {y − x : x ∈ A, y ∈ B} is positive. Then for each ε > 0 there exists a normed linear functional u on X, with norm 1, such that Re u(y) > Re u(x) + d − ε

(x ∈ A, y ∈ B) .

Proof. We may assume that ε < d. Write ε K0 = y − x − z : x ∈ A, y ∈ B, z < d − , 2 which is bounded, open, and convex. It follows from Lemma 5.2.5 that K0 is located and ρ(0, K0 ) = ε/2. Hence, by Corollary 5.2.7, there exists a normed linear functional u on X such that u = 1 and Re u is positive on K0 . Choose z ∈ X such that z < d − ε/2 and u(z) > d − ε; this is possible since u = 1. For all x ∈ A and y ∈ B we have y − x − z ∈ K0 and so Re u(y − x − z) > 0; whence Re u(y) > Re (u(x) + u(z)) > Re u(x) + d − ε, as required.

2

An inhabited subset K of a normed space X is said to be balanced if αx ∈ K whenever x ∈ K and |α| 1; in that case, 0 ∈ K. Corollary 5.2.10. Let x0 be a vector in a separable normed space X, and K a located subset of X that is bounded, convex, and balanced, such that ρ(x0 , K) > 0. Then for each ε > 0 there exists a normed linear functional u on X with norm 1 such that u(x0 ) > |u(x)| + ρ(x0 , K) − ε for all x ∈ K.

122

5 Linearity and Convexity

Proof. We may assume that 0 < ε < ρ(x0 , K). Applying Theorem 5.2.9 with A = K and B = {x0 }, construct a normed linear functional v on X with norm 1 such that Re v(x0 ) > Re v(x) + ρ(x0 , K) − ε (x ∈ K) . Since 0 ∈ K, we have v(x0 ) = 0. Thus u=

|v(x0 )| v v(x0 )

is a normed linear functional on X with norm 1 such that u(x0 ) = |v(x0 )| Re v(x0 ) > Re v(x) + ρ(x0 , K) − ε (x ∈ K) . For each x ∈ K, either u(x0 ) > |u(x)| + ρ(x0 , K) − ε, or else u(x) = 0 and therefore v(x) = 0. In the latter case, since K is balanced, |v(x)| x ∈ K, v(x) and therefore

u(x0 ) > Re v

|v(x)| x + ρ(x0 , K) − ε v(x)

= |v(x)| + ρ(x0 , K) − ε = |u(x)| + ρ(x0 , K) − ε, as required.

2

5.3 The Hahn–Banach Theorem The Hahn–Banach theorem enables us to extend a normed linear functional, with an arbitrarily small increase in norm, from a subspace of a normed space to the entire space. This fundamental result has numerous applications throughout functional analysis. In the constructive context we deal only with the extension of linear functionals on subspaces of a separable normed space. The standard classical proofs extending the theorem to nonseparable normed spaces depend on Zorn’s lemma and are therefore nonconstructive. We begin with two more consequences of the separation theorem. Proposition 5.3.1. Let x be an element of a nontrivial separable normed space X, and let ε > 0. Then there exists a normed linear functional u on X such that u = 1 and u(x) > x − ε.

5.3 The Hahn–Banach Theorem

123

Proof. If x = 0, then we may apply the separation theorem (Theorem 5.2.9) with A = {0} and B = {x} . In the general case, choose a nonzero vector y such that x − y < ε/2, and then construct a normed linear functional u on X such that u = 1 and u(y) > y − ε/2. We have u(x) u(y) − |u(x) − u(y)| > y −

ε − x − y > x − ε, 2 2

as required.

Corollary 5.3.2. Let V be a located subspace of a separable normed space X such that the metric complement −V is inhabited. Then there exists a normed linear functional u on X such that u = 1 and u(V ) = {0} . Proof. Let e be a unit vector in −V. Applying Proposition 5.3.1 in the quotient space X/V, construct a normed linear functional u on X/V such that u = 1 and u(v) = 0. Clearly, V ⊂ ker(u). Since X/V , we see that u, regarded as a linear functional on the original normed space X, has bound 1. On the other hand, given ε > 0, we can ﬁnd x ∈ X with xX/V < 1 and u(x) > 1 − ε; choosing y ∈ V with x − y < 1, we then have u(x − y) = u(x) > 1 − ε. Since ε > 0 is arbitrary, it follows that, as a linear functional on X, u is normed and has norm 1. 2

Proposition 5.3.1 is crucial for our proof of the Hahn–Banach theorem. Theorem 5.3.3. Let v be a nonzero bounded linear functional on a linear subset Y of a separable normed linear space X such that ker(v) is located in X. Then for each ε > 0 there exists a normed linear functional u on X such that u v + ε and u(y) = v(y) for each y ∈ Y. Proof. First note that, by Proposition 2.3.6, v is normed as a linear functional on Y. Fix y0 in Y with v(y0 ) = 1. Then for each x ∈ ker(v), y0 − x

1 1 v(y0 − x) = , v v

from which it follows that y0 X/ ker(v) = ρ (y0 , ker(v))

1 . v

For each normed linear functional u on X/ ker(v) denote the norm by uX/ ker(v) . Using Proposition 5.3.1, construct a normed linear functional u0 on X/ ker(v) such that u0 X/ ker(v) = 1 and u0 (y0 ) > 1/(v + ε). Then u=

1 u0 u0 (y0 )

124

5 Linearity and Convexity

is a normed linear functional on X/ ker(v) such that uX/ ker(v) < v + ε and u(y0 ) = 1. Since X/ ker(v) on X, we see that u, regarded as a linear functional on the original normed space X, has bound uX/ ker(v) . To see that this is actually the norm of u on X, consider any δ > 0 and choose x ∈ X such that xX/ ker(v) < 1 and u(x) > uX/ ker(v) − δ. Finding y ∈ ker(v) such that x − y < 1, we have u(x − y) = u(x) > uX/ ker(v) − δ. Since δ > 0 is arbitrary, we conclude that u is normed on X, with u = uX/ ker(v) v + ε. Finally, for each y ∈ Y we have y − v(y)y0 ∈ ker(v), so 0 = u(y − v(y)y0 ) = u(y) − v(y), and therefore u(y) = v(y).

2

The Hahn–Banach theorem and some of its associates have surprising applications. As surprising as any is the following, whose classical proof is almost trivial. Proposition 5.3.4. Let x1 , . . . , xn be elements of an inﬁnite-dimensional normed space X, and let ε > 0. Then there exist linearly independent elements e1 , . . . , en of X such that xi − ei < ε for each i. Proof. Construct a ﬁnite-dimensional subspace V of span {x1 , . . . , xn } such that for each i there exists yi ∈ V with xi − yi < ε/2 (see Lemma 4.1.11). Since X is inﬁnite-dimensional and V is ﬁnite-dimensional, we can embed V in an ndimensional subspace W of X. We may assume that y1 = 0. Setting e1 = y1 , suppose that for some k < n, we have constructed linearly independent elements e1 , . . . , ek of W such that yi − ei < ε/2 for 1 i k. Let Vk be the k-dimensional subspace of W with basis {e1 , . . . , ek } . By Corollary 5.3.2, there exists a normed linear functional u on W such that u (Vk ) = {0} and u = 1. Construct a vector z ∈ W such that z = ε/2 and u (z) > ε/3. Either u (yk+1 ) = 0 or else |u (yk+1 )| < ε/3. In the ﬁrst case, ρ (yk+1 , Vk ) > 0 and we set ek+1 = yk+1 . In the second case, u(yk+1 − z) = 0, ρ (yk+1 − z, Vk ) > 0, and we set ek+1 = yk+1 − z. In each case, the vectors e1 , . . . , ek+1 are linearly independent, by Lemma 4.1.10, and 2 xk+1 − ek+1 < ε. This completes the inductive construction.

The classical Hahn–Banach theorem says that we can extend a bounded linear functional v from Y to a functional u on the whole space X with exact preservation of norm: that is, u = v . In general, as Exercise 5 shows, we cannot do this

5.3 The Hahn–Banach Theorem

125

constructively. But, as we shall see, if we impose extra conditions on the norm of X, then we can make the extension norm-preserving. A mapping u of a linear space X into R is said to be convex if u (λx + (1 − λ) y) λu(x) + (1 − λ) u(y) whenever x, y ∈ X and 0 λ 1. Lemma 5.3.5. Let u be a convex mapping of a linear space X into R such that u(x) = −u(−x) for each x ∈ X. Then u is linear. Proof. First note that u(0) = −u(−0) = −u(0), so u(0) = 0. For each x ∈ X and for 0 λ 1 we have u(λx) = u(λx + (1 − λ) 0) λu(x) + (1 − λ) u(0) = λu(x). Replacing x by −x, we obtain u(−λx) λu(−x); whence, by our hypotheses on u, −u(λx) −λu(x) and therefore λu(x) u(λx). Hence u(λx) = λu(x) whenever 0 λ 1. Now consider any real λ = 0. If λ > 0, then by the foregoing, 1 λ λ u(x) = u x = u(λx) λ+1 λ+1 λ+1 and therefore λu(x) = u(λx). If λ < 0, then u(λx) = u((−λ)(−x)) = −λu(−x) = λ(−u(−x)) = λu(x). Now consider any λ ∈ R, and assume that u(λx) = λu(x). The foregoing shows that ¬ (λ > 0 ∨ λ < 0) and therefore λ = 0. But this is absurd, since u(0x) = 0 = 0u(x). We conclude that u(λx) = λu(x) for all λ ∈ R. It remains to prove the additivity of u. By the convexity of u, 1 1 1 (x + y) u(x) + u(y). u 2 2 2 Applying the ﬁrst part of the proof yields u(x + y) u(x) + u(y). We now replace x, y by −x, −y respectively, to obtain u(−(x + y)) = u(−x − y) u(−x) + u(−y) and therefore, by the hypotheses on u, u(x + y) u(x) + u(y). Hence u(x + y) = u(x) + u(y). 2

126

5 Linearity and Convexity

The norm of a normed space X is said to be Gˆ ateaux diﬀerentiable at x ∈ X if for each y ∈ X the limit ux (y) = lim

t→0

x + ty − x t

exists. For each c > 0 we then have x + c−1 ty − x = ux (y). ucx (y) = lim t→0 c−1 t If the norm of X is Gˆateaux diﬀerentiable at each nonzero point of X, we say that X has Gˆ ateaux diﬀerentiable norm. In such a case we obtain a strengthening of Proposition 5.3.1. Proposition 5.3.6. Let x be a nonzero vector in a normed space X, and suppose that the norm of X is Gˆ ateaux diﬀerentiable at x. Then the mapping ux is the unique normed linear functional u on X with u = 1 and u(x) = x . Proof. It is enough to take the case K = R. It is straightforward to show that ux is convex, and that ux (y) = −ux (−y) for each y ∈ X; whence, by Lemma 5.3.5, ux is linear. Moreover, |ux (y)| y and ux (x) = x. So ux = 1, and it only remains to prove the uniqueness. To this end, let u be any normed linear functional on X such that u = 1 and u(x) = x . For each y ∈ X and each t > 0 we have u(x − ty) − u(x) u(−ty) x − ty − x = = u(y) −t −t −t and

x + ty − x u(x + ty) − u(x) . t t Keeping y ﬁxed and letting t −→ 0 in these two sets of displayed inequalities, we 2 obtain ux (y) u(y) ux (y); whence u(y) = ux (y). u(y) =

Proposition 5.3.7. Let Y be a subspace of a uniformly convex Banach space X with Gˆ ateaux diﬀerentiable norm, and let v be a nonzero normed linear functional on Y. Then there exists a unique normed linear functional u on X such that u = v and u(y) = v(y) for each y ∈ Y. Proof. We may assume that v = 1. Using Exercise 19 of Chapter 2 if necessary, we may further assume that Y is closed in X and is therefore a Banach space. Since Y is uniformly convex, it follows from Proposition 2.3.7 that there exists a unique a ∈ Y such that a = 1 = v(a). By Proposition 5.3.6, ua is a normed linear functional on X such that ua = 1 = ua (a). Applying the uniqueness part of Proposition 5.3.6 in the space Y, we conclude that ua (y) = v(y) for all y ∈ Y.

5.4 Locally Convex Spaces

127

Finally, if u is a normed linear functional on X such that u = v and u(y) = v(y) for all y ∈ Y, then u(a) = v(a) = 1 = v = u , so, again by the uniqueness part of Proposition 5.3.6, u = ua .

2

Not surprisingly, there is also a strong version of the separation theorem that applies in a uniformly convex separable Banach space; see Exercises 9 and 10.

5.4 Locally Convex Spaces Although Errett Bishop considered that in most cases of interest it seems to be unnecessary to make use of any deep facts from the general theory of locally convex spaces ([9], page 350), the development of constructive analysis (in particular, aspects of the theory of operators) in recent years has greatly beneﬁted from such a general theory, which we now present. A locally convex space consists of a linear space X over K, a family (pi )i∈I of seminorms on X, and the equality and (compatible, tight) inequality relations deﬁned by x = y ⇐⇒ ∀i ∈ I (pi (x − y) = 0) , x = y ⇐⇒ ∃i ∈ I (pi (x − y) > 0) . Following normal practice, we call X itself a locally convex space when it is clear which family of seminorms is under consideration. The family (pi )i∈I and the associated equality and inequality together form the locally convex structure on X. The corresponding locally convex topology on X is the family τ X of all subsets of X that are unions of sets of the form ! pi (x − a) < ε U (a, F, ε) = x ∈ X : i∈F

where a ∈ X, F is an inhabited ﬁnitely enumerable subset of I, and ε > 0. The seminorms pi (i ∈ I) are called the deﬁning seminorms of τ X . The members of τ X are called open subsets of X, and the sets U (a, F, ε) basic neighbourhoods of a. On the other hand, if S is a subset of X, then its closure S is the set of all elements x of X such that S ∩ U (x, F, ε) is inhabited for all ﬁnitely enumerable F ⊂ I and all ε > 0. We say that S is closed (in the locally convex topology τ X ) if S = S, and that S is dense in X if S = X. The unit ball of the locally convex space X is {x ∈ X : ∀i ∈ I (pi (x) 1)} ,

128

5 Linearity and Convexity

which is a closed subset of X. We regard a normed space (X, ) as a locally convex space in which the family of seminorms consists of the single norm . In that case, the locally convex topology is just the metric topology associated with the norm on X, and the various notions of convergence, total boundedness, locatedness, and so on arising from the locally convex structure on X are precisely those associated with the norm on X.

In the following, unless we state otherwise, X, (pi )i∈I and Y, (qj )j∈J are locally convex spaces over K. A mapping f of a subset S of X into Y is continuous at a ∈ S if for each ε > 0 and each ﬁnitely enumerable G ⊂ J, there exist δ > 0 and aﬁnitely enumerable F ⊂ I such that if x ∈ S and pi (x − a) < δ, then qj (f (x) − f (a)) < ε; i∈F

j∈G

continuous on S if it is continuous at each point of X; uniformly continuous on S if for each ε > 0 and each ﬁnitely enumerable subset G of J, there exist δ > 0 and a ﬁnitely enumerable subset F of I such that if pi (x − y) < δ, then qj (f (x) − f (y)) < ε. x, y ∈ S and i∈F

j∈G

Notice that each of the deﬁning seminorms pi on X is uniformly continuous on X. Proposition 5.4.1. The following are equivalent conditions on a linear mapping u between the locally convex spaces X and Y : (a) u is continuous at 0. (b) u is continuous on X. (c) u is uniformly continuous on X. (d) There exist a positive real number C and a ﬁnitely enumerable subset F of I such that |u(x)| C sup pi (x) i∈F

for each x ∈ X. Proof. It is routine to show that (d) =⇒ (c) =⇒ (b) =⇒ (a). Suppose that u is continuous at 0. There exist δ > 0 and a ﬁnitely enumerable subset F of I such pi (x) < δ, then |u(x)| < 1. It follows that for each x ∈ X and each ε > 0, that if i∈F

⎛ ⎞ δx u ⎝ ⎠ < 1 pi (x) + ε i∈F

5.4 Locally Convex Spaces

and therefore 1 |u(x)| < δ

129

pi (x) + ε .

i∈F

Since ε > 0 is arbitrary, it follows that |u(x)|

1 pi (x). δ i∈F

2

Thus (a) =⇒ (d).

Naturally, in view of the analogous situation in a normed space, we say that a mapping u satisfying one, and hence all, of conditions (a)–(d) of Proposition 5.4.1 is a bounded linear mapping.

We say that a sequence (xn )n1 in a locally convex space X, (pi )i∈I converges to a limit x in X if for each basic neighbourhood U = U (x, F, ε) of x in X there exists n0 such that xn ∈ U whenever n n0 ; is a Cauchy sequence if for each ε > 0 and each ﬁnitely enumerable subset F of pi (xm − xn ) < ε whenever m, n n0 . I, there exists n0 such that i∈F

We say that X is complete if every Cauchy sequence in X converges to a limit in X. (Strictly speaking, this deﬁnition applies only when X is separable—that is, has a countable dense subset. However, it will suﬃce for our purposes.) Let S be a subset of X, F a ﬁnitely enumerable subset of I, and ε > 0. By an ε-approximation to S relative to F we mean a subset T of S such that for each pi (x − y) < ε. We say that S is x ∈ X there exists y ∈ T with i∈F

totally bounded relative to F if for each ε > 0 there exists a ﬁnitely enumerable ε-approximation to S relative to F ; totally bounded if it is totally bounded relative to each ﬁnitely enumerable subset of I. The proofs of the next ﬁve results are similar to those of their counterparts in metric space theory (see Chapter 2, Section 2) and are left as exercises. Proposition 5.4.2. If S is a totally bounded subset of a locally convex space, and f is a uniformly continuous mapping of S into a locally convex space Y, then f (S) is totally bounded.

130

5 Linearity and Convexity

Proposition 5.4.3. If S is a totally bounded subset of a locally convex space, and f is a uniformly continuous mapping of S into R, then sup f (x) and inf f (x) exist. x∈S

x∈S

A subset S of the locally convex space X, (pi )i∈I is said to be located (in X) if ! inf pi (x − y) : y ∈ S i∈F

exists for each x ∈ X and each ﬁnitely enumerable subset F of I. As it does in metric spaces, locatedness plays an important role in locally convex spaces. Proposition 5.4.4. A totally bounded subset of a locally convex space is located. Proposition 5.4.5. A located subset of a totally bounded set in a locally convex space is totally bounded. Theorem 5.4.6. Let S be a totally bounded subset of a locally convex space X, and let f be a uniformly continuous mapping of S into R. Then for all but countably many t > inf f (x), the set x∈S

{x ∈ S : f (x) t}

is totally bounded. Now consider the set B(X, Y ) of all bounded linear mappings between the locally convex spaces X and Y. This set, which we often denote by B when it is clear which spaces X and Y are under consideration, becomes a locally convex space when taken with pointwise addition and multiplication-by-scalars and endowed with the seminorms px deﬁned by px (T ) = T x (x ∈ X, T ∈ B(X, Y )) . We denote the unit ball of B(X, Y ) by B1 (X, Y ) or just B1 . When X = Y, we usually write B(X) and B1 (X) rather than B(X, Y ) and B1 (X, Y ). In the special case where Y is the ground ﬁeld K, we obtain the space of all bounded linear functionals on X; this space is called the ∗ dual space of X, and is denoted by X ∗ or sometimes, for clarity, by X, (pi )i∈I ; its unit ball is denoted by X1∗ . The topology associated with the family of seminorms (px )x∈X on X ∗ is called the weak ∗ topology on X ∗ . When we are dealing with, for example, total boundedness relative to the locally convex structure on X ∗ , we speak of weak ∗ total boundedness. In classical analysis, when X and Y are normed spaces, a big role is played by the operator norm T = sup {T x : x ∈ X, x 1} of an element T of B(X, Y ). Constructively the operator norm exists only for those elements T that are normed; but Corollary 4.1.8 shows that it exists for all T ∈

5.4 Locally Convex Spaces

131

B(X, Y ) when X is ﬁnite-dimensional. In that case, X ∗ is algebraically isomorphic to X, every element of X ∗ is normed, the weak∗ topology on X ∗ coincides with the topology associated with the operator norm, and X ∗ is a Banach space relative to the operator norm; these facts are easily veriﬁed using results from Section 1 of Chapter 4. We now examine the unit ball of the dual space. Theorem 5.4.7. (Banach–Alaoglu theorem) If X is a separable normed space, then X1∗ is complete and totally bounded relative to the locally convex structure on X ∗. Proof. To prove the weak∗ -completeness, let (vn )n1 be a Cauchy sequence in X1∗ . Given ε > 0 and a ﬁnitely enumerable subset F of X, we can ﬁnd N such that |(vm − vn )(x)| < ε (m, n N ) . (5.7) x∈F

It readily follows that for each x ∈ X, (vn (x))n1 is a Cauchy sequence in K. Since K is complete, the limit v(x) of the sequence (vn (x))n1 exists in K for each x ∈ X. It is straightforward to show that the mapping v : X −→ K so deﬁned is linear and belongs to X1∗ . Finally, keeping n N ﬁxed and taking the limit with respect to m in (5.7), we obtain |(v − vn ) (x)| ε (n N ) . x∈F

Since ε, F are arbitrary, the sequence (vn )n1 converges to v in the weak∗ topology. To prove the weak∗ -total boundedness of X1∗ , let F = {x1 , . . . , xm } be a ﬁnitely enumerable subset of X, let M > 4 + max {xi : 1 i m} , and let 0 < ε < 1. By Lemma 4.1.11, there exists a ﬁnite-dimensional subspace X0 of X such that for 1 i m, ρ (xi , X0 ) < ε/m and therefore there exists yi ∈ X0 with xi − yi < ε/m. If X0 = {0}, then for each v ∈ X1∗ , m i=1

|v(xi ) − v(0)|

m

xi < ε,

i=1

so {0} is an ε-approximation to X1∗ relative to F. We may therefore assume that X0 has positive dimension. Then every element of X0∗ (the dual of X0 ) is normed, and X0∗ , taken with the operator norm, is a ﬁnite-dimensional Banach space. Hence its unit ball is compact relative to the operator norm (Corollary 4.1.7). Moreover, by Proposition 2.3.6, each nonzero element of X0∗ has its kernel located in X0 ; since X0 is locally compact, this kernel is locally compact and hence is located in the space X (Proposition 2.2.18). It follows that the Hahn–Banach theorem can be applied to extend nonzero bounded linear functionals from X0 to X.

132

5 Linearity and Convexity

∗ Let u01 , . . . , u0n be an ε/m-approximation to the unit ball of X0 in the operator 0 norm such that 0 < uk < 1 for each k. By the Hahn–Banach theorem, there exist normed linear functionals u1 , . . . , un in X1∗ such that uk (x) = u0k (x) for each x ∈ X0 and each k. Given u ∈ X1∗ , since the restriction of u to X0 is in the unit ball of X0∗ , we can ﬁnd k such that u(x) − u0k (x) < ε/m for all x ∈ X0 with x 1. For each i (1 i m) we have |u(xi ) − uk (xi )| |(u − uk ) (xi − yi )| + u − u0k (yi ) 1 0 yi 2 xi − yi + (1 + yi ) u − uk 1 + yi ε ε 2ε + 1 + xi + < m m m Mε . < m

Hence

m

|u(xi ) − uk (xi )| < M ε.

(5.8)

i=1

We now see that {u1 , . . . , un } is an M ε-approximation to X1∗ relative to F.

2

Corollary 5.4.8. If X is a separable normed space, then the normed linear functionals on X are weak∗ -dense in X ∗ . Proof. In the notation of the last paragraph of the foregoing proof, the function uk satisfying (5.8) is normed. Since ε > 0 and the ﬁnitely enumerable subset F of X are arbitrary, the result follows. 2 Let X be a normed space. It is easy to see that, for a ﬁxed vector x ∈ X, the linear functional u u(x) on X ∗ is weak∗ -uniformly continuous on X1∗ . We shall show that any element of X ∗∗ (the dual of X ∗ ) that is uniformly continuous on X1∗ can be approximated arbitrarily closely by functionals of this special form, and that if X is complete, then this approximation can be made exact. Lemma 5.4.9. Let V be a locally convex space whose unit ball V1 is totally bounded, and let φ be a nonzero linear functional on V that is uniformly continuous on V1 . Then the unit kernel, V1 ∩ ker(φ), of φ is totally bounded. Proof. Since φ is nonzero and uniformly continuous on the totally bounded set V1 , C = sup{|φ(y)| : y ∈ V1 } exists and is positive. Using the linearity of φ, choose y ∈ V1 with φ(y) > C/2. Then

5.4 Locally Convex Spaces

y0 =

133

C y 2φ(y)

belongs to V1 , and φ(y0 ) = C/2. Let ε be a positive number, and F = {p1 , . . . , pm } a ﬁnitely enumerable set of deﬁning seminorms on V . Using Theorem 5.4.6, compute a positive number Cε t< C + 4m such that St = {y ∈ V1 : |φ(y)| t} is totally bounded. Pick a t-approximation {s1 , . . . , sn } of St relative to F . Set C 2 sk − φ(sk )y0 (1 k n). yk = C + 2t C Then yk ∈ ker(φ). If pi is any deﬁning seminorm on V , then we have C 2 pi (sk ) + |φ(sk )|pi (y0 ) pi (yk ) C + 2t C 2t C 1+ C + 2t C = 1. So yk belongs to V1 ∩ ker(φ). Now consider any element y of V1 ∩ ker(φ). Since y ∈ St , there exists k such m that pi (y − sk ) < t and therefore i=1

m i=1

pi (y − yk )

m

pi (y − sk ) +

i=1

m

m

pi (sk − yk )

i=1

2 pi (tsk + φ(sk )y0 ) 0 there exist δ > 0 and a ﬁnite-dimensional subspace X0 of X such that if u ∈ X1∗ and |u(x)| < δ for all x in the unit ball of X0 , then |φ(u)| < ε. Proof. There exist ξ 1 , . . . , ξ n in X with ξ k < 1/2 for each k, and a positive number δ, such that if u ∈ X1∗ and |u(ξ k )| < 2δ for each k, then |φ(u)| < ε. By Lemma 4.1.11, there exist a ﬁnite-dimensional subspace X0 of X, and vectors x1 , . . . , xn of X0 , such that 1 ,δ (1 k n) . ξ k − xk < min 2 Suppose that u ∈ X1∗ and |u(x)| < δ for all x in the unit ball of X0 . Then for each k, |u(ξ k )| |u(ξ k − xk )| + |u(xk )| ξ k − xk + δ < 2δ. Hence |φ(u)| < ε.

2

Lemma 5.4.11. Let X be a separable normed space, X0 a ﬁnite-dimensional subspace of X, and φ a nonzero linear functional on X that is weak∗ -uniformly continuous on X1∗ . Denote the unit kernel of φ by N1 . Then the following hold: (a) φ = sup {|φ(u)| : u ∈ X1∗ } exists. (b) x0 = sup {|u(x)| : u ∈ N1 } deﬁnes a seminorm on X0 . (c) β = inf {x0 : x ∈ X0 , x = 1} is well deﬁned. Suppose that β > 0. Then (d) (X0 , 0 ) is a ﬁnite-dimensional Banach space, and 0

u = sup {|(u(x)| : x ∈ X0 , x0 1} deﬁnes the corresponding operator norm on the dual space X0∗ . Finally, for each u ∈ N1 let F (u) denote the restriction of u to X0∗ . Then 0

∗ (e) F is continuous, and ran(F ) is -dense in the unit ball S0∗ weak -uniformly 0 of X0∗ , .

Proof. Since φ is weak∗ -uniformly continuous on X1∗ and, by Theorem 5.4.7, the latter is weak∗ -totally bounded, we see that φ exists. Also, N1 is weak∗ -totally bounded, by Lemma 5.4.9, and the mapping u u(x) is weak∗ -uniformly continuous on X ∗ , so x0 exists. Clearly, 0 is a seminorm and 0 ; whence 0 is uniformly continuous on (X0 , ) . Since that space is ﬁnite-dimensional,

5.4 Locally Convex Spaces

135

{x ∈ X0 : x = 1} is compact, and so β exists. Now suppose that β > 0. Then β −1 0 , so 0 is a norm on X0 equivalent to , and (X0 , 0 ) is a ﬁnite-dimensional Banach space. The mapping 0 F : N1 −→ (X0∗ , ) is weak∗ -uniformly continuous on the weak∗ -totally bounded 0 set N1 , so its range is totally bounded, and hence located, in (X0∗ , ). To show 0 ∗ ∗ ∗ that ran(F ) is dense in the unit ball S0 of (X0 , ), ﬁx u0 in S0 and suppose that 0 0 < c = inf u0 − F (u) : u ∈ N1 . 0

By Corollary 5.2.10, there exists a normed linear functional Φ on (X0∗ , ) such that c (u ∈ N1 ) . Φ(u0 ) > |Φ(F (u))| + 2 Now X0∗ , being ﬁnite-dimensional, is equivalent to a Hilbert space. (Recall from Corollary 4.1.9 that all norms on a ﬁnite-dimensional space are equivalent to each other and hence to the Euclidean norm.) It follows from this and the Riesz representation theorem that there exists ζ ∈ X0 such that Φ(v) = v(ζ) for each v ∈ X0∗ . Therefore u0 (ζ) > sup {|F (u)(ζ)| : u ∈ N1 } = sup {|u(ζ)| : u ∈ N1 } = ζ0 . 0

This is absurd, since u0 ∈ S0∗ . Hence, in fact, c = 0 and ran(F ) is -dense in S0∗ . 2 Lemma 5.4.12. With the hypotheses and notation of Lemma 5.4.11 and its proof, suppose that φ > 1 and let α, δ be positive numbers. Suppose also that |φ(u)| < α/2 whenever (a) u ∈ X1∗ and (b) |u(x)| < δ for all x in the unit ball of X0 . Then β < α. Proof. Choose u0 ∈ X1∗ such that φ(u0 ) = 1. Since |u0 (βx)| = β |u0 (x)| β x x0 , the linear functional Ψ : x u0 (βx) belongs to S0∗ . Now, either β < α or β > 0. Assuming the latter inequality, we see from Lemma 5.4.11 that there exists u ∈ N1 such that 0

Ψ − F (u) < 2δ. So for all x in the unit ball of (X0 , 0 ) , |u0 (βx) − u(x)| < 2δ

136

5 Linearity and Convexity

and therefore

1 (βu0 − u) (x) < δ. 2

But β 1, and both u0 and −u belong to the convex set X1∗ , so whence φ 1 (βu0 − u) < α . 2 2

Therefore β = βφ(u0 ) − φ(u) = 2φ

1 (βu0 − u) 2

1 2

(βu0 − u) ∈ X1∗ ;

< α, 2

as we required.

Proposition 5.4.13. Let X be a separable normed space, and φ a linear functional on X ∗ that is weak∗ -uniformly continuous on X1∗ . Then for each ε > 0 there exists x ∈ X such that x < 3 φ and |φ(u) − u(x)| < ε

(u ∈ X1∗ ) .

Proof. If φ < ε, then we can take x = 0; so we may assume that φ > 0. Scaling if necessary, we may also assume that there exists u0 ∈ X1∗ with φ(u0 ) = 1. Pick a positive number α, which we shall specify further as the proof develops. Using Lemma 5.4.10, construct a ﬁnite-dimensional subspace X0 of X, and a positive number δ, such that if u ∈ X1∗ and |u(x)| < δ for all x in the unit ball of X0 , then |φ(u)| < α/2. Applying Lemma 5.4.12, we see that there exists z ∈ X0 such that z = 1 and |u(z)| < α for all u ∈ N1 . For each u ∈ X1∗ , since 1 (u − φ(u)u0 ) ∈ N1 , 1 + φ we therefore have

1 |(u − φ(u)u0 ) (z)| < α. 1 + φ

By Proposition 5.3.1, there exists u1 ∈ X1∗ such that u1 (z) > 1/2. Thus 1 < |(u1 − φ(u1 )u0 ) (z)| + |φ(u1 )| |u0 (z)| (1 + φ) α + φ |u0 (z)| , 2 and therefore

1 |u0 (z)| > φ

1 − (1 + φ) α 2

>

1 3 φ

provided we choose α small enough. In that case, writing x=

1 z, u0 (z)

5.4 Locally Convex Spaces

137

we see that x < 3 φ. Also, for each u ∈ X1∗ , |u(x) − φ(u)| =

1 |(u − φ(u)u0 ) (z)| < 3 φ (1 + φ) α, |u0 (z)|

which can be made less than ε by a suitably small choice of α.

2

Theorem 5.4.14. Let X be a separable Banach space, and φ a linear functional on X ∗ that is weak∗ -uniformly continuous on X1∗ . Then there exists x ∈ X such that φ(u) = u(x) for each u ∈ X ∗ . Proof. We may assume that φ < 1. Recursively applying Proposition 5.4.13, construct a sequence (xn )n1 of vectors in X such that for each n, n 1 u(xk ) < n φ(u) − 2

(u ∈ X1∗ )

(5.9)

k=1

and xn < 3/2n−1 . The series

∞ n=1

xn then converges to an element x of the com-

plete space X. Using the linearity and continuity of u, and letting n −→ ∞ in (5.9), we obtain the desired conclusion. 2

Let H be a nontrivial Hilbert space. One of the topologies on B(H) that is important in operator-algebra theory is the weak-operator topology τ w : that is, the locally convex topology deﬁned by the seminorms of the form T |T x, y| with x, y in H. Classically—but not constructively (see Exercise 17)—the sets of the type ⎫ ⎧ n ⎬ ⎨ |T ei , ej | < δ , T ∈ B(H) : ⎭ ⎩ i,j=1

with δ > 0 and {e1 , . . . , en } a set of pairwise orthogonal unit vectors in H, form a base of weak-operator neighbourhoods of 0 in B(H), so a linear functional φ on B(H) is τ w -continuous if and only if it has the following special continuity property:

SC

T here exist δ > 0 and a set {e1 , . . . , en } of pairwise orthogonal unit vectors n |T ei , ej | < δ, then |φ(T )| < 1. in H such that f or each T ∈ B(H), if i,j=1

We shall use the technique embodied in the proofs of Lemma 5.4.11, Lemma 5.4.12, and Proposition 5.4.13 to produce a characterisation of those linear functionals φ on B(H) with the property SC.

138

5 Linearity and Convexity

Before the next result, we mention that, in spite of what we wrote in the preceding paragraph, it is constructively provable that the sets of the form ⎫ ⎧ n ⎬ ⎨ |T ei , ej | < δ , T ∈ B1 (H) : ⎭ ⎩ i,j=1

with δ > 0 and {e1 , . . . , en } a set of pairwise orthogonal unit vectors in H, form a base of weak-operator neighbourhoods of 0 in the unit ball B1 (H). This is left as part of Exercise 17. Proposition 5.4.15. If H is a nontrivial Hilbert space, then B1 (H) is τ w -totally bounded. Proof. Let {e1 , . . . , en } be a ﬁnite set of pairwise orthogonal unit vectors generating a ﬁnite-dimensional subspace H0 of H. In view of the observation preceding this proposition, it will suﬃce to prove that B1 (H) is totally bounded with respect to the n seminorm pjk : T |T ej , ek | . Let P be the projection of H on H0 . Note that j,k=1

B(H0 ) is a ﬁnite-dimensional Banach space, and hence bounded unit has a totally 0 be an ε/n2 -approximation to ball, relative to the operator norm. Let T10 , . . . , Tm B1 (H0 ), and consider any T ∈ B1 (H). The restriction (P T )0 of P T to H0 belongs to B1 (H0 ), so there exists i such that (P T )0 − Ti0 < ε/n2 . Also, Ti0 P ∈ B1 (H). Thus if 1 j, k n, then , - , - , - T − Ti0 P ej , ek = T − Ti0 ej , P ek = P T − Ti0 ej , ek , - ε = (P T )0 − Ti0 ej , ek (P T )0 − Ti0 < 2 . n Hence

n , - 0 T − T 0 P ej , ek < ε. We now see that T10 P, . . . , Tm P is an εi

j,k=1

approximation to B1 (H) relative to the seminorm pjk .

2

Proposition 5.4.16. Let H be a nontrivial Hilbert space, and let φ be a linear functional on B(H) with the property SC. Then for each ε > 0 there exist a ﬁnite set {e1 , . . . , en } of pairwise orthogonal unit vectors in H, and elements cjk (1 j, k n) of K, such that n φ(T ) − < ε (T ∈ B1 (H)) . c T e , e jk j k j,k=1 Proof. With the proofs of Lemma 5.4.11, Lemma 5.4.12, and Proposition 5.4.13 at hand, we omit some of the grisly details of this one. We ﬁrst note that, in view of the

5.4 Locally Convex Spaces

139

observation immediately preceding Proposition 5.4.15, φ is τ w -uniformly continuous on B1 (H). It follows from this, Proposition 5.4.15, and Proposition 5.4.3 that φ = sup {|φ(T )| : T ∈ B1 (H)} exists. Scaling if necessary, we may assume that there exists T0 ∈ B1 (H) such that unit vectors in H, φ(T0 ) = 1. Choose a ﬁnite set {e1 , . . . , en } of pairwise orthogonal n and a positive number δ, such that if T ∈ B(H) and i,j=1 |T ei , ej | < 2δ, then |φ(T )| < 1. Let H0 = span {e1 , . . . , en } , and for each x = (x1 , . . . , xn ) in the n-fold product space H0n , deﬁne the product norm x =

n

xk .

k=1

Note that for each T ∈ B1 (H) and each x ∈ H0n , n n n T xk , ek |T xk , ek | xk = x . k=1

k=1

For such x, the mapping

(5.10)

k=1

n T T xk , ek k=1

is τ w -uniformly continuous on B1 (H). Since (by Proposition 5.4.15 and Lemma 5.4.9) the unit kernel N1 of φ is τ w -totally bounded, it follows from Proposition 5.4.3 that we can deﬁne a seminorm on H0n by n ! T xk , ek : T ∈ N1 . x0 = sup k=1

Then 0 , by (5.10), so the mapping 0 from (H0n , ) to R is uniformly continuous. Since H0n is ﬁnite-dimensional, the set {x ∈ H0n : x = 1} is -compact. Hence β = inf {x0 : x ∈ H0n , x = 1} is well deﬁned. This time (cf. Lemma 5.4.12) we prove that β = 0. Suppose that β > 0. Then (H0n , 0 ) is a ﬁnite-dimensional Banach space, and 0

u = sup {|u(x)| : x ∈ H0n , x0 1} ∗

deﬁnes the corresponding operator norm on the dual space (H0n ) . For each T ∈ N1 and each x ∈ H0n let

140

5 Linearity and Convexity

F (T )(x) =

n

T xk , ek .

k=1 ∗

0

The mapping F : N1 −→ ((H0n ) , ) is τ w -uniformly continuous, so F (N1 ) ∗ 0 is totally bounded, and therefore located, in ((H0n ) , ). Using the separation theorem as in the proof of Lemma 5.4.12, we can show that F (N1 ) is dense in the 0 unit ball S0∗ of (H0n )∗ , ). Since the linear functional Ψ :x

n

βT0 xk , ek

k=1 0

belongs to S0∗ , given t > 0 we can ﬁnd T ∈ N1 such that Ψ − F (T ) < n−2 δt. Write {(j, k) : 1 j, k n} = P ∪ Q, where P and Q are disjoint sets, (j, k) ∈ P =⇒ (βT0 − T ) ej , ek = 0, (j, k) ∈ Q =⇒ |(βT0 − T ) ej , ek | < n−2 δt. If (j, k) ∈ P, set rjk = (βT0 − T ) ej , ek

−1

|(βT0 − T ) ej , ek | ,

and if (j, k) ∈ Q, set rjk = 0. Now deﬁne ξ ∈ H0n by ξ k = n−2

n

(1 k n) ,

rjk ej

j=1

to obtain ξ k n−2

n

|rjk | n−1

j=1

and hence ξ0 ξ = Thus

ξ k 1.

k=1

n (βT0 − T ) ξ k , ek < n−2 δt. k=1

Moreover,

n

(5.11)

5.4 Locally Convex Spaces

141

n n n 2 n (βT0 − T ) ξ k , ek = rjk (βT0 − T ) ej , ek k=1 k=1 j=1 rjk (βT0 − T ) ej , ek = (j,k)∈P |(βT0 − T ) ej , ek | , = (j,k)∈P

so n

|(βT0 − T ) ej , ek |

j,k=1

|(βT0 − T ) ej , ek | +

(j,k)∈P

|(βT0 − T ) ej , ek |

(j,k)∈Q

n

(βT0 − T ) ξ k , ek + n2 n−2 δt 2δt, n2 k=1

the last step following from (5.11). Hence n −1 t (βT0 − T ) ej , ek < 2δ j,k=1

and therefore |φ(βT0 − T )| < t. We now have β = βφ(T0 ) − φ(T ) = φ(βT0 − T ) < t. Since t > 0 is arbitrary, it follows that β = 0. Thus for each α > 0 there exists z ∈ H0n such that z = 1 and n T zk , ek < (2n)−1 α (T ∈ N1 ) . k=1

We can arrange that each zk = 0, provided we replace (2n)−1 by n−1 in the last inequality; we assume that this has been done. Therefore n (T − φ(T )T0 ) zk , ek < n−1 (1 + φ) α (T ∈ B1 (H)) . k=1

We now introduce an operator T1 that will enable us to bound from 0. Setting T1 x = n−1

n k=1

we have

−1

zk

x, zk ek

(x ∈ H) ,

n k=1

T0 zk , ek away

142

5 Linearity and Convexity

T1 x n−1

n

−1

zk

x zk = x ,

k=1

so T1 ∈ B1 (H). Also, n

T1 zk , ek = n−1

k=1

Hence n−1

n

zk

−1

zk , zk = n−1

k=1

n

zk = n−1 .

k=1

n n (T1 − φ(T1 )T0 ) zk , ek + |φ(T1 )| T0 zk , ek k=1 k=1 n T0 zk , ek n−1 (1 + φ) α + φ k=1

and therefore n −1 −1 T0 zk , ek (n φ) (1 − (1 + φ) α) > (2n φ) k=1

provided α is small enough. Setting n −1 x= T0 zk , ek z, k=1

for each T ∈ B1 (H) we now obtain −1 n n n n T xk , ek = T0 zk , ek φ(T )T0 zk , ek − T zk , ek φ(T ) − k=1

k=1

k=1

k=1

< 2n φ n−1 (1 + φ)α = 2 φ (1 + φ) α, which can be made less than ε by choosing α suitably small. It remains to take 2 cjk = xk , ej for 1 j, k n.

Exercises 1. Let U, V be subsets of a Banach space X such that U ∪ V is dense. Prove the following: (a) If u ∈ U and v ∈ V, then for each ε > 0 there exist t ∈ [0, 1] and x ∈ U ∩ V such that x − tu − (1 − t)v < ε.

5.4 Locally Convex Spaces

143

(b) For each x ∈ X,

ρ x, U ∩ V = min {ρ(x, U ), ρ(x, V )} , in the sense that each side of this equation exists if and only if the other does. 2. We say that a subset S of a normed space X has the boundary crossing property if for each ξ ∈ S, each z ∈ ∼S, and each ε > 0 there exist t ∈ [0, 1] and x ∈ ∂S such that x − tξ − (1 − t) z < ε. Prove that if X is complete and S ∪ ∼S is dense, then S has the boundary crossing property. Do we necessarily obtain this conclusion if S ∪ ∼S is not known to be dense in X? 3. Let S be a subset of a Banach space such that S ∪∼S is dense, and let x0 ∈ ∼S. Prove that ρ (x0 , ∂S) exists if and only if ρ (x0 , S) exists, in which case these two distances are equal. 4. Let {e1 , . . . , en } be the canonical orthonormal basis of the Euclidean space Rn , and Σ the simplex whose vertices are ±ej (1 j n). Let ξ belong to the metric complement −Σ of Σ in Rn . Prove that the segment [0, ξ] contains points arbitrarily close to the union of the faces of Σ. 5. Show that if the Hahn–Banach theorem holds with exact preservation of the norm of the functional, then LLPO holds. 6. Let Y be a ﬁnite-dimensional subspace of a normed space X, and x0 ∈ X. Suppose that ρ(x0 , Y ) > 0 and that X is the span of Y ∪ {x0 } . Let v be a nonzero normed linear functional on Y, and ε > 0. Without using the separation theorem, the Hahn–Banach theorem, or any of their consequences, prove that there exists a normed linear functional u on X such that u v + ε and u(y) = v(y) for each y ∈ Y. Use this to give another proof of the Hahn–Banach theorem. 7. Use the Hahn–Banach theorem to prove Proposition 5.3.1. 8. Let p be an integer 2. Prove that the Banach space lp is uniformly convex and has Gˆ ateaux diﬀerentiable norm. (See Exercise 11 of Chapter 4.) 9. Let X be a normed space with Gˆ ateaux diﬀerentiable norm. Let C be a closed, located, convex subset of X, let x ∈ −C, and let y ∈ C. Prove that x − y = ρ (x, C) if and only if Re ux−y (z − y) 0 for all z ∈ C. 10. Let X be a uniformly convex Banach space with Gˆ ateaux diﬀerentiable norm. Let A and B be subsets of X whose algebraic diﬀerence is located and convex, and whose mutual distance is positive. Prove that there exists a normed linear functional u on X, with norm 1, such that Re u (y) Re u (x) + d for all x ∈ A

144

5 Linearity and Convexity

and y ∈ B. (The results embodied in Exercises 9 and 10 are due to Ishihara [55].) 11. Prove that if, under the other hypotheses of Lemma 5.4.9, the conclusion holds without the hypothesis that φ is either 0 or nonzero, then we can derive the essentially nonconstructive proposition ∀x ∈ R (x = 0 ∨ ¬ (x = 0)) . 12. Let f be a uniformly continuous linear functional on X, and S a balanced, totally bounded subset of X. Prove that the set St = {x ∈ S : |f (x)| t} is totally bounded for each t > 0. 13. Let (X, p) be a seminormed space, and S a balanced, totally bounded subset of X. Let the mapping f : X −→ K be both uniformly continuous on S and homogeneous—that is, f (λx) = λf (x) for all λ ∈ K and x ∈ X. Prove that the set St = {x ∈ S : |f (x)| t} is totally bounded for each t > 0. 14. Let X be a ﬁnite-dimensional locally convex space, and let S be the set of convex combinations of points in an inhabited ﬁnitely enumerable subset of X. Prove that S is totally bounded in X. 15. Fill in the missing details in the proof of Proposition 5.4.16. 16. Let (en )n1 be an orthonormal basis of a separable Hilbert space H, and let φ be a linear functional on B(H) that is weak-operator uniformly continuous on B1 (H). Prove that there exist complex numbers cjk (j, k 1) such that ∞ φ(T ) = cjk T ej , ek for all T ∈ B(H). j,k=1

17. Prove the following for a nontrivial Hilbert space H: (a) The sets of the form ⎫ ⎧ n ⎬ ⎨ |T ei , ej | < δ , T ∈ B1 (H) : ⎭ ⎩ i,j=1

with δ > 0 and {e1 , . . . , en } a set of pairwise orthogonal unit vectors in H, form a base of weak-operator neighbourhoods of 0 in the unit ball B1 (H).

5.4 Locally Convex Spaces

145

(b) If B1 (H) can be replaced by B(H) in (a), then LPO holds. 18. Let C be a convex subset of Cn such that sup {f (z) : z ∈ C} exists for each real (bounded) linear functional f on Cn . Prove that C is totally bounded. (Hint: Prove that C is bounded; that ∗ d(x) = sup inf Re f (x − y) : f ∈ S y∈C

exists for each x ∈ Cn , where S ∗ is the unit ball of the dual of Cn ; and that ρ(x, C) exists and equals d(x).) 19. Let X, Y be normed spaces. We say that φ is ultraweakly continuous if it is uniformly continuous on B1 (X, Y ) relative to the locally convex topology τ w deﬁned on B(X, Y ) by the seminorms T y ∗ (T x) with x ∈ X and y ∗ a normed linear functional on Y. Prove that the following conditions are equivalent: (a) B1 (X, Y ) is τ w -totally bounded. (b) Every ultraweakly continuous linear functional on B(X, Y ) is normed. (c) Every ultraweakly continuous linear functional on B (X, Y ) is compact.

Notes A classical method of establishing the continuity of the boundary crossing map in Proposition 5.1.5 is to use a contradiction argument to prove that the mapping is sequentially continuous on −C; see [87] (pages 271–272). Our argument, based on the simple Lemma 5.1.3, is much more transparent and shorter than either that classical argument or its constructive counterpart (which is given in [32] and proves only the sequential continuity, not the full continuity, of the boundary crossing map). More information about boundary crossing can be found in [36], from which Exercises 2–4 are taken. The Hahn–Banach theorem can be proved without recourse to the separation theorem (Exercise 6). Using the Church–Markov–Turing thesis, Metakides et al. [71] have produced an explicit example in which the hypotheses of the Hahn–Banach theorem hold but the linear functional cannot be extended with exact preservation of its norm. This makes Theorem 5.3.7 (which, together with Exercises 9 and 10, is due to Ishihara [55]) all the more signiﬁcant, especially as it applies to many important examples of normed spaces, such as the Lp spaces for p > 1. The locally convex topology on a ﬁnite-dimensional space X is unique and equivalent to the topology induced by any norm on X. The proof of this depends on some intricate geometric algorithms and is found in [39].

146

5 Linearity and Convexity

Finding necessary and suﬃcient conditions for the existence of the norm of a bounded linear mapping T : X −→ Y between normed spaces, other than the case Y = K, is an important unsolved problem, even in the case where T is an operator on a Hilbert space. The Banach–Alaoglu theorem (Theorem 5.4.7) without the hypothesis of separability is classically equivalent to the ultraﬁlter principle—every ﬁlter is contained in an ultraﬁlter (see [80], page 766)—and is therefore unlikely to be provable constructively; see Exercise 1.4 of Chapter 1. The introduction of the numbers rjk in the proof of Proposition 5.4.16 is occasioned by the problem with the modulus–argument decomposition discussed at the start of Chapter 2. Proposition 5.4.16 is a step in the direction of a constructive proof of the classical characterisation of weak-operator continuous linear functionals as those of the n T xk , yk with the vectors xk , yk in H. All attempts to generalise the form T k=1

technique used to prove Proposition 5.4.13 in order to characterise ultraweakly continuous linear functionals—those linear functionals on B(H) that are weak-operator uniformly continuous on B1 (H)—suﬀer from a curse of dimensionality, failing ben T xk , yk is usually more than 1. cause the n appearing in expressions like k=1

Although Exercise 16 requires only the weak-operator uniform continuity of the functional on the unit ball of B(H) and derives a stronger conclusion than that of Proposition 5.4.16, the solution that we have in mind for that exercise works only for a separable Hilbert space. In that case we can improve the conclusion to ﬁnd square-summable sequences (xn )n1 and (yn )n1 in H such that for each T ∈ B(H) ∞ we have φ(T ) = T xn , yn ; see [23]. n=1

If H is a Hilbert space, then B1 (H) is classically weak-operator compact; however, the completeness of B1 (H) relative to the uniform structure associated with the weak-operator topology is an essentially nonconstructive property; see [14]. The result in Exercise 18 ﬁrst appeared in [56].

6 Operators and Locatedness

Location! Location! Location! —Unknown Estate Agent

We begin by introducing normed spaces on which the norm is diﬀerentiable in some fashion. In Section 2, with substantial help from the λ-technique, we provide criteria for the locatedness of certain convex sets in a normed space. This work is applied in Section 3 to proving that a bounded operator on a Hilbert space H has an adjoint if and only if it maps the unit ball to a located set. In the next section we construct approximate eigenvectors of a selfadjoint operator on H, and then show that a bounded positive operator has a unique positive square root. This result is applied in Section 5, in which we make further good use of the λ-technique to show that, for a so-called weak-sequentially open operator T on H, the range of T is located if and only if the range of its adjoint is located. The section ends with a proof of the closed range theorem for operators with an adjoint. The ﬁnal section of the chapter presents a version of Baire’s theorem, which is then applied to the proofs of three of the pillars of functional analysis: the open mapping, inverse mapping, and closed graph theorems.

6.1 Smooth and Uniformly Smooth Normed Spaces As the reader will have observed, locatedness, classically a nugatory concept, plays an important role in many aspects of constructive analysis, such as the Hahn– Banach theorem and the separation theorem discussed in the preceding chapter. It is therefore ﬁtting that criteria for locatedness should be a primary theme of this ﬁnal chapter. Our ﬁrst aim is to ﬁnd necessary and suﬃcient conditions for the locatedness of convex sets in a normed space; this requires some preliminary results associated with the diﬀerentiability of the norm. Later in the chapter we shall apply our conditions to the locatedness of subsets associated with operators on a Hilbert space.

148

6 Operators and Locatedness

Throughout this section, unless we state otherwise, X will be a normed linear space. We say that X is smooth if its norm is Gˆ ateaux diﬀerentiable at each nonzero vector; this is the case if and only if the limit x + ty − x t→0 t

ux (y) = lim

(6.1)

exists for all unit vectors x, y ∈ X. We say X has Fr´echet diﬀerentiable norm if this limit is uniform in the unit vector y: that is, if for each unit vector x ∈ X and each ε > 0 there exists δ > 0 such that x + ty − x − ux (y) < ε (6.2) t whenever y ∈ X, y = 1, and 0 < |t| < δ. If, moreover, δ can be chosen independent of x, then we say that X has uniformly Fr´echet diﬀerentiable norm and is uniformly smooth. For this to be the case, for each ε > 0 there must exist δ > 0 such that (6.2) holds whenever x, y ∈ X, x = y = 1, and 0 < |t| < δ. With each vector x in a smooth normed space X we associate the mapping Jx : X −→ R deﬁned by Jx (y) = lim x t→0

x + ty − x . t

If x = 0, then Jx (y) = 0; if x = 0, then Jx (y) = x ux (y). In both these cases, |Jx (y)| x y . In order to show that Jx (y) is deﬁned for an arbitrary element x of X, construct an increasing binary sequence (λn )n1 such that 1 , n 1 . λn = 1 =⇒ x > n+1

λn = 0 =⇒ x

1; so (ζ n )n1 is a Cauchy sequence and therefore converges to a limit ζ ∈ R. Note that |ζ| x y . For each ε > 0, either x y < ε/2 or x y > 0. In the ﬁrst case we have x x + ty − x − ζ 2 x y < ε t for all t = 0. In the second case, ζ = Jx (y) and x x + ty − x − ζ < ε t for all suﬃciently small t = 0. Since ε > 0 is arbitrary, we conclude that Jx (y) exists and equals ζ.

6.1 Smooth and Uniformly Smooth Normed Spaces

149

Proposition 6.1.1. If X is smooth, then for each x in X, Jx is a normed linear 2 functional on X, Jx = x, and Jx (x) = x . Moreover, if c > 0 and x ∈ X, then Jcx = cJx . Proof. If x = 0, then the ﬁrst three conclusions follow from Proposition 5.3.6. On the other hand, if also c > 0, then Jcx = cx ucx = c x ux = cJx . 2

The extension to the general case is left to the reader. Lemma 6.1.2. If X is smooth, then for all x, y ∈ X, 2

2

y x − 2Jy (x − y) . Proof. Using the linearity of Jy and also Proposition 6.1.1, we compute 2

2

Jy (x − y) = Jy (x) − y y x − y . Hence 2

2

2

2

2

x − y − 2Jy (x − y) x + y − 2 x y = (x − y) 0. 2 Lemma 6.1.3. Let X be a smooth normed space, and x a unit vector in X. Let δ, ε be positive numbers such that if y = 1 and 0 < |t| < δ, then x + ty − x − Jx (y) < ε. t Then for such t and each unit vector y, x + ty + x − ty < 2 + 2tε. Proof. We may assume that 0 < t < δ. If y = 1, then x + ty − x − Jx (y) < ε t and

x − ty − x − Jx (y) > −ε. −t

Hence x + ty < 1 + tJx (y) + tε and x − ty < 1 − tJx (y) + tε, from which the result follows.

2

150

6 Operators and Locatedness

We store the following result for use in Section 2. Proposition 6.1.4. Let X be a smooth, uniformly convex Banach space, and let f be a normed linear functional on X. Then there exists a unique vector x ∈ X such that f = Jx . Proof. First assume that f is nonzero. By Proposition 2.3.7, there exists a unit −1 vector y ∈ X such that f (y) = f . Applying Proposition 5.3.6 to f f, we see that f = f uy . Now let x = f y. Then x = f and (since ucy = uy for all c > 0) f = f uy = x u f y = x ux = Jx . It remains to deal with the uniqueness. Let f = Jz for some z ∈ X. Then, by Proposition 6.1.1, z = Jz = f > 0, so f uy = f = f uz = f u z −1 z −1

and therefore u z −1 z = uy . Since z z is a unit vector, the uniqueness part of −1 Proposition 5.3.6 shows that z z = y; whence z = z y = f y = x. This completes the proof in the case where f is nonzero; the general case is left as an exercise. 2

The next lemma simpliﬁes the proof of the theorem following it. Lemma 6.1.5. Let f be a mapping of R − {0} into R such that (a) if 0 < t t, then f (−t) f (−t ) f (t ) f (t), and (b) for each ε > 0 there exists t > 0 such that f (t) − f (−t) < ε. Then limt→0 f (t) exists. Proof. Construct a sequence (tn )n1 of positive numbers decreasing strictly to 0 such that f (tn ) − f (−tn ) < 1/n for each n. In view of (a), for m n we have 0 < f (tm ) − f (tn ) f (tm ) − f (−tm )

0, choose t > 0 as in (b) and then N such that tN < t and 0 f (tN ) − l < ε. By (a), if 0 < |t| < tN , then 2 f (−tN ) f (t) f (tN ) and therefore |f (t) − l| f (tN ) − l < ε.

6.1 Smooth and Uniformly Smooth Normed Spaces

151

We have already deﬁned “uniformly convex” for normed spaces, in Chapter 2. To introduce this notion for the dual of a normed space X, we concentrate on those elements of X ∗ that have a norm. We say that X ∗ is uniformly convex if for each ε > 0 there exists δ with 0 < δ < 1 such that for all normed u, v ∈ X ∗ with u = v = 1, if there exists a unit vector ξ ∈ X such that 12 (u + v) (ξ) > 1 − δ, then (u − v) (x) < ε for all unit vectors x. Proposition 6.1.6. A separable normed space is uniformly smooth if and only if it has a uniformly convex dual. Proof. Let X be a separable normed space, and suppose to begin with that X ∗ is uniformly convex. Given unit vectors x, y ∈ X, for all nonzero t ∈ R deﬁne f (t) =

x + ty − x . t

Then f satisﬁes (a) of Lemma 6.1.5. It follows that in order to prove that X is uniformly smooth, we need only show that for each ε > 0 there exists t > 0, independent of the unit vectors x and y, such that f (t) − f (−t) < ε. To this end, let δ > 0 be as in the deﬁnition of “uniformly convex” above, ﬁx t > 0 such that 8t(2 + t) < δ, and consider any unit vectors x, y ∈ X. Note that x + ty > 1/2 and x − ty > 1/2. Deﬁning unit vectors ξ, η by ξ=

1 (x + ty), x + ty

η=

1 (x − ty), x − ty

we obtain 1 x − ty (x + ty) − x + ty (x − ty) x + ty x − ty 4 (x − ty − x + ty) x + t (x − ty + x + ty) y

ξ − η =

4 [|x − ty − x + ty| + t (x − ty + x + ty)] 4 (2t + 2t(1 + t)) = 8t(2 + t) < δ.

Now let γ = min

δ tε , 4(1 + t) 2

,

and use Proposition 5.3.1 to construct normed linear functionals u and v such that u = v = 1, u(ξ) > 1 − γ, and v(η) > 1 − γ. Then δ (u + v) (ξ) = u(ξ) + v(η) + v (ξ − η) > 2 − 2γ − ξ − η 2 − 2 − δ = 2 − 2δ 2 and therefore 12 (u + v)(ξ) > 1 − δ. It follows that (u − v) (z) < ε/2 for all unit vectors z ∈ X, and in particular that (u − v) (y) < ε/2. Now,

152

6 Operators and Locatedness

u (x + ty) = x + ty u(ξ) > x + ty (1 − γ) , so x + ty < u(x + ty) + x + ty γ u(x + ty) + (1 + t)γ u(x + ty) +

tε . 4

Likewise, x − ty < v(x − ty) +

tε . 4

Thus 1 x + ty − x x − ty − x − = (x + ty + x − ty − 2) t −t t tε 1 −2 u(x + ty) + v(x − ty) + t 2 tε 1 −2 u(x) + v(x) + t(u − v)(y) + = t 2 ε + (u − v) (y) < ε. 2 This completes the proof that X is uniformly smooth. Now suppose, conversely, that X is uniformly smooth. Then for each ε > 0 there exists δ such that ε ,1 0 < δ < min 4 and such that for all unit vectors x, y ∈ X and all t with 0 < |t| < 4δ/ε, ε x + ty − x − Jx (y) < . t 4 Letting t = 2δ/ε, we see from Lemma 6.1.3 that for all unit vectors x, y ∈ X, x + ty + x − ty < 2 +

tε . 2

Now consider normed u, v ∈ X ∗ with u = v = 1, and assume that 12 (u+v)(ξ) > 1 − δ for some unit vector ξ ∈ X. Then for all unit vectors x ∈ X we have 2−

tε = 2 − δ < (u + v)(ξ) 2 = u(ξ + tx) + v(ξ − tx) − t(u − v)(x) ξ + tx + ξ − tx − t(u − v)(x) tε < 2+ − t(u − v)(x) 2

and therefore (u − v)(x) < ε. Thus X ∗ is uniformly convex.

2

6.2 Locatedness of Convex Sets

153

If H is a Hilbert space, then, in view of the Riesz representation theorem, its dual is, like H itself, uniformly convex. It follows from Proposition 6.1.6 that a separable Hilbert space is uniformly smooth. In fact we can remove separability and completeness here. Proposition 6.1.7. An inner product space is uniformly smooth. Proof. Given unit vectors x, y in the inner product space X, for real t we have 2

2

x − ty − x = −2t Re x, y + t2 and therefore x − ty − x =

−2t Re x, y + t2 . x − ty + x

It follows that if 0 < |t| < 1/2, then x − ty − x −2 Re x, y + t = + Re x, y + Re x, y t x − ty + x 1 |(x − ty − 1) Re x, y + t| = x − ty + 1 4 |t| 2 |t| (Re x, y + 1) , 3 3 which tends to 0 with t, independently of x and y.

2

6.2 Locatedness of Convex Sets Our aim in this section is to give conditions ensuring the locatedness of a bounded convex subset of a normed space under certain strong hypotheses that apply, in particular, to the norm on a Hilbert space or an Lp -space1 for 1 < p < ∞. Speciﬁcally, we aim to prove the following two results and some of their consequences. Theorem 6.2.1. Let X be a uniformly smooth normed space over R, and C an inhabited, bounded, located convex subset of X. Then sup {Jx (y) : y ∈ C} exists for each x ∈ X.

1

We refer the reader to [9, 12] for the theory of Lp spaces.

154

6 Operators and Locatedness

Theorem 6.2.2. Let C be an inhabited, bounded, convex subset of a uniformly smooth normed space X over R, such that sup {Jx (y) : y ∈ C} exists for each x ∈ X. Then C is located. The path to Theorem 6.2.1 takes us through a tangle of technical lemmas, the ﬁrst of which is an expression of the continuity of Jx as a function of x. Lemma 6.2.3. Let X be a real normed space with a Fr´echet diﬀerentiable norm. For each unit vector x ∈ X and each ε > 0 there exists δ > 0 such that if y < δ, then |Jx (z) − Jx+y (z)| < ε for all unit vectors z ∈ X. Proof. Given a unit vector x ∈ X and ε > 0, compute γ ∈ (0, 1) such that if z = 1 and 0 < |t| < γ, then ε x + tz − x < . − J (6.3) (z) x 8 t

Let δ = min

γε 1 , 32 2

and consider a vector y ∈ X with y < δ. Noting that x + y > 1/2, set ξ=

1 (x + y). x + y

Then since |1 − x + y| = |x − x + y| y , we have x + y ξ − x = x + y − x + y x = (1 − x + y) x + y 2 y . 2

Also, by Proposition 6.1.1, Jξ (ξ) = ξ = 1, so 0 1 − Jξ (x) = Jξ (ξ − x) ξ − x

2 y < 4δ. x + y

(6.4)

On the other hand, setting t = γ/2 and applying Lemma 6.1.3, we see that for each unit vector z ∈ X, tε x + tz + x − tz < 2 + . 4 Therefore, by (6.4), the choice of t and δ, and Proposition 6.1.1,

6.2 Locatedness of Convex Sets

1−

155

tε 1 − 4δ 4 < Jξ (x) = Jx (x) + Jξ (x) − 1 = Jx (x + tz) + Jξ (x − tz) − 1 − t(Jx − Jξ )(z) x + tz + x − tz − 1 − t(Jx − Jξ )(z) tε < + 1 − t(Jx − Jξ )(z). 4

Hence (Jx − Jξ )(z) < ε/2. Replacing z by −z, we obtain −(Jx − Jξ )(z) < ε/2 and therefore |(Jx − Jξ )(z)| < ε/2. It follows from this and Proposition 6.1.1 that |(Jx − Jx+y )(z)| = |x + y (Jx − Jξ )(z) + (x − x + y) Jx (z)| x + y |(Jx − Jξ )(z)| + |x − x + y| |Jx (z)| 3 |(Jx − Jξ )(z)| + y x z 2 ε 3ε + < ε, < 4 32 2

as we wanted.

Lemma 6.2.4. Let X be a uniformly smooth normed space over R, and let C be an inhabited, bounded, located convex subset of X. Let x0 ∈ X and ε > 0. Then there exists τ > 0 such that for all y ∈ C, either there exists x ∈ C such that Jx0 (y) + τ ε/6 < Jx0 (x) or else Jx0 (z) < Jx0 (y) + ε for all z ∈ C. Proof. Choose M > 0 such that x − y < M for all x, y ∈ C. If M x0 < ε, then for all x, y ∈ C, Jx0 (x − y) x0 x − y M x0 < ε. Hence we may assume that M x0 > 0. From the deﬁnition of “uniformly smooth” and Lemma 6.2.3 we see that there exists δ > 0 such that for all unit vectors x, y ∈ X, • if 0 < |t| δ, then x + ty − y 2ε < − J , (y) x t 3 x0 M and

156

6 Operators and Locatedness

• if y δ, then for all unit vectors z ∈ X, |Jx (z) − Jx+y (z)|

0. Since (1 − τ ) y + τ x belongs to the convex set C, we see from (6.7) and Proposition 6.1.1 that −

2τ ε 2 2 < x0 + τ (y − x) − x0 3 2 = Jx0 +τ (y−x) (x0 + τ (y − x)) − x0 2

= Jx0 +τ (y−x) (x0 ) + τ Jx0 +τ (y−x) (y − x) − x0 2

x0 + τ (y − x) x0 − x0 + τ Jx0 +τ (y−x) (y − x). Hence 0 < x0

x0 + τ (y − x) − x0 2ε + Jx0 +τ (y−x) (y − x) + . τ 3

(6.8)

6.2 Locatedness of Convex Sets

157

The appearance here of the expression E=

x0 + τ (y − x) − x0 τ

in the presence of a uniformly Fr´echet diﬀerentiable norm suggests that we consider the unit vectors 1 1 x0 , v = (y − x). u= x0 y − x Writing t= we have

τ y − x , x0

u + tv − u 1 E= . y − x t

Since 0 < t δ, it follows that 2ε 1 > E − Ju (v) 3 x0 M y − x 1 1 E− Jx0 (y − x) = y − x x0 y − x 1 |x0 E − Jx0 (y − x)| = x0 y − x and therefore x0 x0 + τ (y − x) − x0 − Jx0 (y − x) < 2ε y − x < 2ε . τ 3M 3

(6.9)

In view of (6.8) and (6.9), it now makes sense to examine Jx0 (y − x) − Jx +τ (y−x) (y − x) , 0 which, by Proposition 6.1.1, equals x0 y − x |Ju (v) − Ju+tv (v)| . By our choice of δ, this last expression is less than x0 y − x

2ε , 3 x0 M

so

Jx (y − x) − Jx +τ (y−x) (y − x) < 2ε . 0 0 3 It follows from (6.8)–(6.10) that

(6.10)

158

6 Operators and Locatedness

x0 + τ (y − x) − x0 2ε + Jx0 +τ (y−x) (y − x) + 0 < x0 τ 3 2ε 2ε 2ε + Jx0 (y − x) + + , Jx0 (y − x) + 3 3 3 from which we obtain the inequality Jx0 (x) < Jx0 (y) + ε. Since x was an arbitrary point of C, the proof is complete. 2

Using Lemma 6.2.4 and the λ-technique, we can now prove Theorem 6.2.1. Proof. Under the hypotheses of Theorem 6.2.1, ﬁx M > 0 such that x − y M for all x, y ∈ C, and consider any element x of X. It is enough to prove that for each ε > 0 there exists y ∈ C such that Jx (z) < Jx (y) + ε

(z ∈ C) .

(6.11)

For in that case, if 0 < α < β and we choose y ∈ C such that (6.11) holds with ε = (β − α) /2, then either Jx (y) > α, or else Jx (y) < 12 (α + β) and therefore Jx (z) < β for all z ∈ C; so the desired supremum exists by the least-upper-bound principle. Fixing x ∈ X, y0 ∈ C, and ε > 0, let τ be as in Lemma 6.2.4. We construct an increasing binary sequence (λn )n0 with λ0 = 0, and a sequence (yn )n0 in C, such that τε < Jx (yn ), 6 λn = 1 =⇒ Jx (z) < Jx (yn−1 ) + ε for all z ∈ C.

λn = 0 =⇒ Jx (yn−1 ) +

To do so, we assume that we have constructed λ0 , . . . , λn and y0 , . . . , yn . If λn = 1, set λn+1 = 1 and yn+1 = yn . If λn = 0, then by Lemma 6.2.4, either Jx (yn )+τ ε/6 < Jx (y ) for some y ∈ C, or else Jx (z) < Jx (yn ) + ε for all z ∈ C. In the ﬁrst case set λn+1 = 0 and yn+1 = y . In the second case set λn+1 = 1 and yn+1 = yn . This completes the inductive construction. Now choose N such that M x < N If λN = 0, then Jx (y0 ) + N

τε . 6

τε < Jx (yN ) 6

and therefore

τε < Jx (yN − y0 ) M x , 6 a contradiction. Hence, in fact, λN = 1. N

2

6.2 Locatedness of Convex Sets

159

Corollary 6.2.5. Let X be a uniformly convex, uniformly smooth Banach space over R, and let C be an inhabited, bounded, located convex subset of X. Then sup {f (y) : y ∈ C} exists for each normed linear functional f on X. Proof. For a nonzero normed linear functional the result follows from Proposition 6.1.4 and Theorem 6.2.1. The completion of the proof is left to the reader. 2

Turning now towards Theorem 6.2.2, we prove the following lemma. Lemma 6.2.6. Let C be an inhabited, bounded, convex subset of a uniformly smooth normed space X over R such that sup{Jx (z) : z ∈ C} exists for each x ∈ X, and let ε > 0. Then for each x ∈ X there exists σ with 0 < σ < 1 such that for each y ∈ C, either there exists y ∈ C with x − y < (1 − σ) x − y or else x − y < x − z + ε for all z ∈ C. Proof. We may assume that ε < 1. Fixing x ∈ X, choose a positive integer M > max {ε/4, 1} such that x − z < M for each z ∈ C. Since X is uniformly smooth, there exists δ with 0 < δ < 2 such that for all unit vectors u, v ∈ X, u − tv − u ε − Ju (v) < . (6.12) 0 < |t| < δ =⇒ −t 8M Deﬁne

εδ 8M and note that 0 < σ < 1. Let y ∈ C. If x − y < ε, then x − y < x − z + ε for all z ∈ C. Hence we may assume that ε/2 < x − y. Now, σ=

sup {Jx−y (z − y) : z ∈ C} = sup {Jx−y (z) : z ∈ C} − Jx−y (y) exists, by our hypotheses. So either Jx−y (z − y) < ε x − y for all z ∈ C or else there exists z0 ∈ C such that ε x − y < Jx−y (z0 − y) . 2 In the ﬁrst case, for all z ∈ C we have

(6.13)

160

6 Operators and Locatedness 2

x − y = Jx−y (x − y) = Jx−y (x − z) + Jx−y (z − y) < x − y x − z + ε x − y , and therefore x − y < x − z + ε. In the second case, setting τ=

δ x − y , 2M

y = y + τ (z0 − y),

we have 0 < τ < δ/2 < 1 and therefore y ∈ C. Deﬁne also u=

1 (x − y), x − y

v=

1 (z0 − y), z0 − y

t=

τ z0 − y . x − y

Note, for the deﬁnition of v, that z0 − y x − y Jx−y (z0 − y) >

ε x − y 2

and hence that z0 − y > 0. Let t=

δ z0 − y τ z0 − y = . x − y 2M

Then t < δ and, by Proposition 6.1.1 and (6.13), tJu (v) =

τ τ τε Ju (z0 − y) = 2 Jx−y (z0 − y) > 2 x − y . x − y x − y

Hence x − y = x − y − τ (z0 − y) = x − y u − tv tε < x − y u − tJu (v) + 8M τ ε τ ε z0 − y + . < x − y − 2 8M Since z0 − y z0 − x + x − y < 2M and (6.13) holds, this gives τε τε + 2 4 δε x − y = x − y − 8M = (1 − σ) x − y

x − y < x − y −

by (6.12)

6.2 Locatedness of Convex Sets

161

and completes the proof that, in this case, the ﬁrst conclusion of the lemma obtains. 2

We now prove Theorem 6.2.2. Proof. Under the hypotheses of Theorem 6.2.2, let x ∈ X and ε > 0. It suﬃces to show that there exists y ∈ C such that x − y < x − z + ε for all z ∈ C. Compute σ > 0 as in Lemma 6.2.6, and let y0 ∈ C. If x − y0 < ε, then x − y0 < x − z + ε for all z ∈ C. So we may assume that x − y0 > 0. Now construct an increasing binary sequence (λn )n0 with λ0 = 0, and a sequence (yn )n0 of elements of C, such that λn = 0 =⇒ x − yn < (1 − σ) x − yn−1 , λn = 1 =⇒ x − yn < x − z + ε for all z ∈ C. Suppose we have constructed λ0 , . . . , λn and y0 , . . . , yn . If λn = 1, set λn+1 = 1 and yn+1 = yn . If λn = 0, then by Lemma 6.2.6, either x − y < (1 − σ) x − yn for some y ∈ C, or else x − yn < x − z + ε for all z ∈ C. In the ﬁrst case set λn+1 = 0 and yn+1 = y ; in the second case set λn+1 = 1 and yn+1 = yn . Since 0 < σ < 1, we see that (1 − σ)n −→ 0 as n −→ ∞, so there exists N > 1 such that (1 − σ)n

ε/2. Our hypotheses ensure that M = sup {Re x − y, z − y : z ∈ C} exists. Either M < ε2 /2 or else ε2 /4 < M. In the ﬁrst case, for all z ∈ C we have 2

x − y = Re x − y, x − y = Re x − y, x − z + Re x − y, z − y ε2 < x − y x − z + 2 and therefore x − y < x − z +

ε2 < x − z + ε. 2 x − y

Hence (c) holds. We may therefore assume that ε2 /4 < M. Choosing z0 ∈ C such that

6.2 Locatedness of Convex Sets

τ = Re x − y, z0 − y >

163

ε2 , 4

we consider three exhaustive cases. 2

z0 − y < 3τ /2. We have

Case 1:

2

2

x − z0 = x − y − (z0 − y) 2

2

= x − y − 2τ + z0 − y 1 2 < x − y − τ 2 2 ε 2 < x − y − 8 ε4 2 , < x − y − 16n since 0 < ε < 1. Hence (a) holds with y0 = z0 . 2

n/2 < z0 − y . Using the parallelogram identity, we have

Case 2:

n 2 < (z0 − x) + (x − y) 2 2 2 2 = 2 x − z0 + 2 x − y − z0 + y − 2x 2

2

2 x − z0 + 2 x − y , so (b) holds. 2

2

Case 3: τ < z0 − y and z0 − y < n. Then y0 = y +

τ 2

z0 − y

(z0 − y)

belongs to C, and 2

2

x − y0 = x − y − 2

= x − y −

2τ Re x − y, z0 − y 2

z0 − y

+

τ2 2

z0 − y

τ2 2

z0 − y ε4 2 < x − y − 16n and so (a) holds. This completes the proof.

The classical uniform boundedness theorem says:

2

164

6 Operators and Locatedness

If (Ti )i∈I is a family of bounded linear mappings from a Banach space X into a normed space Y such that for each x ∈ X the family (Ti x)i∈I is bounded, then there exists M > 0 such that Ti M for each i ∈ I. The reader is invited to consider why this form of the theorem is unlikely to be provable constructively. The correct constructive approach is via the contrapositive; here is a pretty version due to Royden [78]. Theorem 6.2.11. (Royden’s uniform boundedness theorem) Let (Tn )n1 be a sequence of normed linear mappings from a Banach space X into a normed space Y such that Tn > n3n for each n. Then there exists x ∈ X such that Tn x > n for each n. Proof. For each n 1 choose xn ∈ X such that xn = 3−n and Tn xn > 3 4 Tn xn . Then ∞ xn x=4 n=1 ∞

is well deﬁned, since (by comparison with

3−n ) the series on the right converges

n=1

in the complete space X. For n 2, letting

yn = x1 + · · · + xn−1 , we have 3 Tn xn < 2 Tn xn = T (xn + yn ) + T (xn − yn ) 2 Tn (xn + yn ) + T (yn − xn ) , so at least one of the last two terms is greater than −xn , if necessary, we may assume that Tn (xn + yn ) > Hence

3 4

Tn xn . Replacing xn by

3 Tn xn . 4

n ∞ 1 Tn x = Tn xk + Tn xk 4 k=1 k=n+1 n−1 ∞ xk − Tn xk Tn x n + k=1

k=n+1

∞

3 Tn 3−n − Tn 3−k 4 k=n+1 1 3 1 − = Tn 3−n Tn 3−n 4 2 4

>

6.2 Locatedness of Convex Sets

and therefore Tn x > 3−n Tn > n.

165

2

Theorem 6.2.12. (Uniform boundedness theorem) Let (Tn )n1 be a sequence of normed linear mappings from a Banach space X into a normed space Y such that Tn −→ ∞ as n −→ ∞. Then there exists x ∈ X such that the sequence (Tn x)n1 is unbounded. Proof. Compute a strictly increasing sequence (nk )k1 of positive integers such that Tnk > k3k for each k, and then apply Theorem 6.2.11. 2

It is possible to remove from these two uniform boundedness theorems the condition that each Tn be normed; see Exercise 12. Theorem 6.2.13. Let C be an inhabited convex subset of a Hilbert space H such that sup {Re x, y : y ∈ C} exists for each x ∈ H. Then C is located. Proof. Fix ξ ∈ H and y0 ∈ C. In order to prove that ρ(ξ, C) exists, it is enough to show that if 0 < ε < 1, then there exists η ∈ C such that ξ − η < ξ − y + ε for all y ∈ C. If ξ − y0 < ε, then we may take η = y0 . Hence we may assume that ξ = y0 . Setting λ0 = 0, for the purposes of this proof we extend the λ-technique by constructing a sequence (λn )n0 with values in {−1, 0, 1}, and a sequence (yn )n0 in C, such that for each n 1, ξ − yn ξ − yn−1 and the following hold: 2

2

λn = −1 =⇒ ξ − yn < ξ − yn−1 − ε4 /16n, 2

2

λn = 0 =⇒ n/4 < ξ − yn + ξ − z for some z ∈ C, λn = 1 =⇒ ξ − yn < ξ − z + ε for all z ∈ C. Assume that we have already constructed λ0 , . . . , λn and y0 , . . . , yn . If λn = 1, set λn+1 = 1 and yn+1 = yn . If λn = 1, then we apply Lemma 6.2.10. If there 2 2 exists y ∈ C such that ξ − y < ξ − yn − ε4 /16n, we set λn+1 = −1 and 2 2 yn+1 = y; if n/4 < ξ − yn + ξ − z for some z ∈ C, we set λn+1 = 0 and yn+1 = yn ; if ξ − yn < ξ − z + ε for all z ∈ C, we set λn+1 = 1 and yn+1 = yn . This completes our inductive construction of the sequences (λn )n0 and (yn )n0 . It remains to compute n such that λn = 1. Let (Nk )k1 be a strictly increasing sequence of positive integers such that for each k,

166

6 Operators and Locatedness

2

Nk > 4 k + ξ − y0

2

and ξ − y0

k, 4 4

so wk −→ ∞ as k −→ ∞. We now apply Corollary 6.2.12 to the normed linear functionals x x, wk on H, to produce ξ 0 ∈ H such that the sequence {|ξ 0 , wk |}k1 is unbounded. By our hypotheses, there exists M > 0 such that |ξ 0 , ξ − z| < M for all z ∈ C. Choosing K such that M < |ξ 0 , wK | , suppose that λnK = 0. Then, by our construction of the vectors wk , there exists z ∈ C such that wK = ξ − z and therefore 2 |ξ 0 , wK | = |ξ 0 , ξ − z| < M, a contradiction. Hence λnK = 1.

We shall use some of these results on locatedness in the next section, when we look at adjoints of operators.

6.3 Adjoints Throughout the rest of the chapter, H will be a Hilbert space. By an operator on H we mean a linear mapping of H into itself. We observed in Section 3 of Chapter 5 that the constructive Riesz representation theorem, with its requirement that the linear functional be not just bounded but

6.3 Adjoints

167

normed, does not enable us to prove that every bounded operator on H has an adjoint. In fact, we showed that the proposition “every bounded operator on a Hilbert space has an adjoint” implies LPO. Are there general criteria for the existence of the adjoint? To answer this question aﬃrmatively, we ﬁrst prove a lemma. Lemma 6.3.1. Let T be an operator on H, and B a subset of H. Let a be a unit vector in H, and P the projection of H on the 1-dimensional subspace Ka. Then P T (B) is located if and only if the set {T x, a : x ∈ B} is located in K. Proof. For all x, y ∈ H, since (I − P ) y is orthogonal to P (H), we have 2

2

2

y − P T x = P y − P T x + (I − P ) y 2

= P y, a a − P T x, a a + (I − P ) y 2

2

2

= |y, a − T x, a| + (I − P ) y . Hence

2

inf y − P T x

x∈B

exists if and only if

2

inf |y, a − T x, a|

x∈B

exists, from which the desired conclusion follows.

2

Proposition 6.3.2. Let T be a jointed operator on H. Let B be the unit ball of H, and P the projection of H on a 1-dimensional subspace. Then P T (B) is located. Proof. Choose a unit vector a ∈ H such that P x = x, a a for all x ∈ H. By the preceding lemma, it suﬃces to prove that the set S = {T x, a : x ∈ B} = {x, T ∗ a : x ∈ B} is located in K. We do this by showing that S is dense in the (located) ball B K (0, T ∗ a). Given ε > 0, we have either T ∗ a < ε, in which case |ζ − 0| < ε for all ζ ∈ K with |ζ| T ∗ a, or else T ∗ a = 0. In the latter case, for each ζ ∈ K with ζ T ∗ a we have . / ζ ∗ ∗ ζ= 2 T a, T a T ∗ a and

ζ 2 T ∗ a

T ∗ a ∈ B.

Thus in either case there exists x ∈ B such that |ζ − x, T ∗ a| < ε. Since ε > 0 is 2 arbitrary, it follows that S is dense in B K (0, T ∗ a) .

168

6 Operators and Locatedness

Proposition 6.3.3. Let H be a Hilbert space with unit ball B, and T a bounded operator on H such that P T (B) is located for each 1-dimensional projection P. Then T has an adjoint. Proof. By Lemma 6.3.1, for each unit vector y ∈ H, the set C = {T x, y : x ∈ B} is located in K. Since T is bounded, C is a subset of a ball in K. It follows from Corollary 4.1.7 and Proposition 2.2.10 that C is totally bounded. Thus, by Propositions 2.2.6 and 2.2.5, the linear functional u deﬁned on H by u(x) = T x, y

(x ∈ H)

is normed. By the Riesz representation theorem, there exists a unique element T ∗ y 2 of H such that T x, y = x, T ∗ y for all x ∈ H. Theorem 6.3.4. Let T be a jointed operator on a Hilbert space H. Then the image under T of the unit ball is located. Proof. By Proposition 6.3.2 and Lemma 6.3.1, for each x ∈ H the set {T y, x : y ∈ B} is located in K. This set is also bounded, since |T y, x| = |y, T ∗ x| T ∗ x for all y ∈ B. Hence, as in the previous proof, it is totally bounded. Since the map ζ Re ζ is uniformly continuous on K, we now see that sup {Re x, T y : y ∈ B} = sup {Re T y, x : y ∈ B} exists for each x ∈ H. It follows from Theorem 6.2.13 applied to C = T (B) that T (B) is located in H. 2 Theorem 6.3.5. Let T be a bounded operator on a Hilbert space H that maps the unit ball to a located set. Then T has an adjoint. Proof. Given y ∈ H, we take C = T (B) in Corollary 6.2.9 to show that the linear functional x T y, x is normed. The result now follows from the Riesz representation theorem, as in the proof of Proposition 6.3.3. 2 Thus for a bounded operator on a Hilbert space, the existence of the adjoint is equivalent to the image of the unit ball being located. When is a jointed operator bounded? The classical answer is “always”; the constructive answer is less decisive.

6.3 Adjoints

169

Theorem 6.3.6. (Hellinger–Toeplitz theorem) Every jointed operator on a Hilbert space is sequentially continuous. Proof. Let T be a jointed operator on a Hilbert space H. It will suﬃce to prove that T ∗ is sequentially continuous, since we can then interchange T and T ∗ to obtain the desired result. Accordingly, let (xn )n1 be a sequence converging to 0 in H, and let ε > 0. By Ishihara’s second trick (Lemma 3.2.2), either T ∗ xn < ε eventually or else there exists a strictly increasing sequence (nk )k1 of positive integers such that T ∗ xnk > ε/2 for all k. In the latter case, passing to a subsequence if necessary, we may assume that ε xnk < 2 k 2k 3 for each k. Then, setting 2k3k xnk , yk = ε we have yk < 1/k and T ∗ yk > k3k . We now apply Royden’s uniform boundedness theorem (Theorem 6.2.11) to the normed linear functionals x x, T ∗ yk on H, to construct a vector x ∈ H such that |x, T ∗ yk | > k for each k. Since x, T ∗ yk = T x, yk T x yk −→ 0 as k −→ ∞, this is absurd. We conclude that T ∗ xn < ε for all suﬃciently large n.

2

A subset S of N is pseudobounded if limn→∞ n−1 sn = 0 for each sequence (sn )n1 in S. The following principle is trivially true in classical mathematics, holds in both INT and RUSS, and appears not to be provable in BISH (see [59]).

BD-N

Every inhabited, countable, pseudobounded set of positive integers is bounded.

We now prove that if “sequentially continuous” can be replaced by “bounded” in the conclusion of Theorem 6.3.6, then BD-N holds. Let A = {a1 , a2 , . . .} be a countable pseudobounded set of positive integers. Let H be an inﬁnitedimensional Hilbert space with an orthonormal basis (en )n1 of unit vectors. We ﬁrst prove that for each x ∈ H, Tx =

∞

an x, en en

n=1

is well deﬁned. Let (Nk )k1 be a strictly increasing sequence of positive integers such that

170

6 Operators and Locatedness ∞

2

|x, en | < 2−k−1 k −2 ,

n=Nk

and construct a binary sequence (λk )k1 such that Nk+1 −1

λk = 0 =⇒

2

a2n |x, en | < 2−k ,

n=Nk Nk+1 −1

λk = 1 =⇒

2

a2n |x, en | > 2−k−1 .

n=Nk

Deﬁne a sequence (bk )k1 in A as follows. If λk = 0, set bk = a1 . If λk = 1, then −k−1 −2

2

k

max

a2n

Nk+1 −1

: Nk n < Nk+1

2

a2n |x, en | > 2−k−1 ,

n=Nk

so an > k for some n with Nk n < Nk+1 ; in this case we set bk = an for this n. Since A is pseudobounded, there exists m such that bk /k < 1 for all k m. If λk = 1 for some k m, then 1 < bk /k < 1, a contradiction; hence λk = 0 for all k m, and therefore the series deﬁning T x converges. It is easily seen that T is a one-one selfadjoint linear mapping of H onto itself; so, by Theorem 6.3.6, it is sequentially continuous. But if T is bounded, then the pseudobounded set A is bounded (by any positive bound for T ).

6.4 Functions of Selfadjoint Operators We say that a sequence (Sn )n1 of operators on our Hilbert space H converges strongly to an operator S if Sx = lim Sn x n→∞

for each x ∈ H. Given an operator T on H, for each power series p(t) =

∞

cn tn

(6.15)

n=0

with complex coeﬃcients cn we can form the corresponding power series in T, p(T ) =

∞

cn T n ,

n=0

which makes sense provided the series on the right converges strongly. In particular, if p(t) is a polynomial, regarded as a power series whose coeﬃcients are eventually 0, then p(T ) is always deﬁned.

6.4 Functions of Selfadjoint Operators

171

If each coeﬃcient cn in (6.15) is real and T is selfadjoint, then p(T ) is also selfadjoint. Our ﬁrst aim is to prove that if, in that case, T has bound 1, then p(T ) is bounded by the sup norm of p on the interval [0, 1]. This will require us to prove some lemmas about approximate eigenvalues. Note that two operators S, T on H are said to commute if ST = T S. Also, T 0 = I and T n = T T n−1 for any n 1. Lemma 6.4.1. Let T be a selfadjoint operator on H with bound 1, and let x be a unit vector in H. Let −1 1 1 x+ v . v = T x − T x, x x, u = x + v 2 2 Then

1 2 v . 4 Moreover, if S is any operator that commutes with T, then Su 2 Sx . T u, u > T x, x +

⊥

⊥

Proof. Clearly, v ∈ {x} and T x − v ⊥ {x} , so (by Theorem 4.3.1) v is the ⊥ projection of T x onto {x} . Hence 2 x + 1 v = 1 + 1 v2 > 0, 2 4 and our deﬁnition of u makes sense. Also, 2

T x, v = T x, v − T x, x x, v = v

and therefore, since T = T ∗ , 1 0 1 1 1 T x + v , x + v = T x, x + Re T x, v + T v, v 2 2 4 1 2 T x, x + v − T v v 4 3 2 T x, x + v . 4 Hence −2 0 1 1 1 1 T x + v , x + v − T x, x T u, u − T x, x = x + v 2 2 2 −1 3 1 2 2 T x, x + v − T x, x 1 + v 4 4 −1 3 1 1 2 2 2 v − v T x, x = 1 + v 4 4 4 −1 1 1 2 2 v 1 + v 4 2 1 2 v , 4

172

6 Operators and Locatedness

since v 2. Finally, let S commute with T. Then 1 Su S x + v 2 1 1 = Sx − T x, x Sx + T Sx 2 2 1 1 Sx 2 Sx . 1 + |T x, x| + 2 2 2

We show how to construct approximate eigenvectors common to ﬁnitely many commuting selfadjoint operators. We deal ﬁrst with the case of a single selfadjoint operator. Lemma 6.4.2. Let T be a selfadjoint operator on H with bound 1, and let x1 be a unit vector in H. Let ε be a positive number, and N an integer greater than 32/ε2 . Deﬁne sequences (xn )n1 , (yn )n1 in H recursively by yn = T xn − T xn , xn xn ,

xn+1

−1 1 1 xn + y n . = xn + yn 2 2

Then T xn+1 , xn+1 T xn , xn for each n, and there exists n N such that yn < ε. Proof. By Lemma 6.4.1, T xn+1 , xn+1 − T xn , xn

1 2 yn . 4

On the other hand, either yn < ε for some n N, or else yn > ε/2 for all n N. In the latter case, Lemma 6.4.1 shows that T xn+1 , xn+1 − T xn , xn

ε2 16

for each n N, and therefore that T xN +1 , xN +1 T x1 , x1 +

N ε2 > −1 + 2 = 1, 16

which is absurd since xN +1 is a unit vector and T has bound 1.

2

6.4 Functions of Selfadjoint Operators

173

Lemma 6.4.3. Let T1 , . . . , Tn be commuting selfadjoint operators on H with common bound 1, and let x be a unit vector in H. Then for each ε > 0 there exists a unit vector u such that T1 u, u T1 x, x − ε and Tk u − Tk u, u u < ε

(1 k n) .

Proof. Noting that the case n = 1 has been disposed of in Lemma 6.4.2, let n > 1 and suppose that we have proved the desired result for n − 1 commuting selfadjoint operators. Fix ε > 0, and choose a positive integer N > 32/ε2 . By our induction hypothesis, there exists a unit vector x1 such that T1 x1 , x1 T1 x, x −

ε 2

and

ε (1 k n − 1) . 2N Taking T = Tn , deﬁne sequences (xk )k1 and (yk )k1 as in Lemma 6.4.2. By that lemma, there exists j (1 j N ) such that Tn xj − Tn xj , xj xj < ε. Setting u = xj , we see that if j = 1, then we are ﬁnished; so we may assume that j > 1. Then for 1 k n − 1, since the selfadjoint operator Tk − Tk x1 , x1 I commutes with Tn , we see from the ﬁnal part of Lemma 6.4.1 that Tk x1 − Tk x1 , x1 x1

T1 x, x − − 2 2 = T1 x, x − ε.

174

6 Operators and Locatedness

2

Our induction is now complete.

Proposition 6.4.4. Let T be a selfadjoint operator on H with bound 1. Then for each polynomial p with real coeﬃcients, the operator p(T ) has bound M = sup {|p(t)| : 0 t 1} . Proof. Write p(t) =

N

cn tn ,

n=0

where each cn ∈ R. Let x be a unit vector in H, and let ε > 0. Since p(T ) has only real coeﬃcients, it is selfadjoint. By Lemma 6.4.3, there exists a unit vector u ∈ H such that p(T )u, u p(T )x, x − ε and −1 N n |cn | ε. T u − T u, u u < 1 + n=1

Taking t = T u, u, for n > 1 we compute n−1 T n u − tn u T T n−1 u − tn−1 u + |t| T u − tu n−1 n−1 T u−t u + T u − tu ··· n T u − tu . Hence p(T )u − p(t)u

N

|cn | T n u − tn u

n=1

N

n |cn | T u − T u, u u < ε

n=1

and therefore p(T )x, x p(T )u, u + ε |p(t)| + |(p(T )u − p(t)u) , u| + ε M + p(T )u − p(t)u < M + 2ε. Since ε > 0 is arbitrary, it follows that p(T )x, x M. Replacing p by p2 , we now have , - , 2 p(T )x = p(T )2 x, x = p2 (T )x, x 2 sup |p(t)| : 0 t 1 = M 2 and therefore p(T )x M.

2

6.4 Functions of Selfadjoint Operators

175

An operator T on H is said to be positive if T x, x 0 for all x ∈ H. In that case, T is selfadjoint, since for all x, y ∈ H we have both T x, x = x, T x and the polarisation identity 4 T x, y = T (x + y) , x + y − T (x − y) , x − y + i T (x + iy) , x + iy − i T (x − iy) , x − iy (which is actually valid for any operator T on H). We introduce a partial order on the set of selfadjoint operators by writing S T, or equivalently T S, if and only if T − S is a positive operator. For our ﬁrst result about positive operators we use a weak constructive substitute for the classical result that a series of positive real numbers is convergent if its partial sums form a bounded sequence. Lemma 6.4.5. Let

∞ n=1

an be a series of nonnegative terms whose partial sums form

a bounded sequence. Then for each ε > 0 and each positive integer n there exists k n such that ak < ε. Proof. Let b > 0 be an upper bound for the partial sums of the series. Given ε > 0, choose a positive integer N such that N ε/2 > b. If ak > ε/2 for n < k n + N, then n+N n+N Nε > b, ak ak > 2 k=1

k=n+1

a contradiction. Hence ak < ε for some k with n < k n + N.

2

Proposition 6.4.6. Let S and T be commuting positive operators on H, with S bounded. Then ST is positive. Proof. The proof is similar to the classical one on pages 415–417 of [4], so we give only an outline, leaving the details as an exercise. Since S is bounded, we may assume that 0 S I. Deﬁne a sequence (Sn )n1 of bounded selfadjoint operators on H such that Sn+1 = Sn − Sn2 (n 1) . S1 = S, By induction, 0 Sn I for each n. We now have 0

n

Sk2 =

k=1

n

(Sk − Sk+1 ) = S − Sn+1 S.

k=1

It follows that for each x ∈ H, n k=1

2

Sk x =

n , k=1

Sk2 x, x = (S − Sn+1 ) x, x Sx, x .

176

6 Operators and Locatedness

By Lemma 6.4.5, there exists a strictly increasing sequence (ni )i1 of positive integers such that Sni x < 2−i for each i; whence ni

Sk2 x = Sx − Sni +1 x −→ Sx as i −→ ∞.

k=1

Since T commutes with S, it commutes with every Sn . Hence / . ni Sk2 x , x . ST x, x = T Sx, x = T lim i→∞

k=1

Now, T is positive and hence selfadjoint; so, by Theorem 6.3.6, it is sequentially continuous. Hence ST x, x = lim

i→∞

= lim

i→∞

= lim

i→∞

ni , k=1 ni k=1 ni

T Sk2 x, x

Sk T Sk x, x T Sk x, Sk x 0.

k=1

Since x is arbitrary, we conclude that ST 0.

2

Our next objective is to construct the positive square root of a bounded √ positive operator. We ﬁrst examine an iteration scheme for the function t t on [0, 1] . A standard classical proof of the convergence of that scheme uses Dini’s highly nonconstructive theorem on the uniform convergence of monotone sequences of continuous functions [47] (pages 131–132). Fortunately, with relatively little extra eﬀort, we can avoid using Dini’s theorem altogether. Lemma 6.4.7. Deﬁne a sequence (un )n1 of uniformly continuous mappings from [0, 1] into R iteratively by u1 (t) = 0, Then (un (t))n1 converges to

un+1 (t) = un (t) + √

1 t − u2n (t) . 2

t uniformly on [0, 1].

Proof. For each t ∈ [0, 1] and each n, √ √ 1 t − u2n (t) t − un+1 (t) = t − un (t) − 2 √ 1 √ t − un (t) 1 − t + un (t) . = 2

(6.17)

6.4 Functions of Selfadjoint Operators

177

√

Using this, it is simple to prove by induction that un+1 un and un (t) t. Let 0 < ε < 1, and compute a positive integer N such that n ε < ε (n N ) . 1− √ 2 2 √ Consider any t ∈ [0, 1] . If t < ε2 , then for all n we have 0 un (t) t < ε and therefore √ (6.18) 0 t − un (t) < ε. If t > ε2 /2, then for all n 1, 1√ ε 1 √ t + un (t) t> √ 2 2 2 2 and therefore 1−

ε 1 √ t + un (t) < 1 − √ . 2 2 2

It follows from this and (6.17) that for all n 2N, √ √ ε 0 t − un (t) 1 − √ t − un−1 (t) 2 2 2 √ ε t − un−2 (t) 1− √ 2 2 ··· n−N √ ε t − uN (t) 1− √ 2 2 N ε 1− √ 2 2 < ε. Thus (6.18) holds for all n 2N.

2

Our next lemma will enable us to transform the iteration in Lemma 6.4.7 into one for the square root of a bounded positive operator. Lemma 6.4.8. Let T be a selfadjoint operator on H that satisﬁes 0 T I, deﬁne the sequence (un )n1 as in Lemma 6.4.7, and write Un = un (T ). Then for each n we have 0 Un 12 (I + T ) , Un2 T, and

1 2 T − Un+1 = T − Un2 I − (Un+1 + Un ) . (6.19) 2

178

6 Operators and Locatedness

Proof. We have

2 2 T − Un+1 = T − Un2 + Un2 − Un+1 = T − Un2 − (Un+1 − Un ) (Un+1 + Un ) 1 T − Un2 (Un+1 + Un ) = T − Un2 − 2

1 2 = T − Un I − (Un+1 + Un ) . 2 Suppose that for some n we have 0 Un 12 (I + T ) and Un2 T. (These inequalities certainly hold for n = 1.) Then 0 Un I, so Un and I − Un are positive and 2 hence selfadjoint, and (I − Un ) 0. Thus 1 1 1 Un+1 = (I + T ) − I + Un − Un2 2 2 2 1 1 1 2 = (I + T ) − (I − Un ) (I + T ) . 2 2 2 On the other hand, Un+1 − Un =

1 T − Un2 0 2

and therefore Un+1 Un 0. Thus 0 and so 2 Hence T −Un+1

1 1 (Un+1 + Un ) Un+1 (I + T ) I, 2 2

1 I − (Un+1 + Un ) 0. 2 0, by (6.19) and Proposition 6.4.6. This completes the induction. 2

Proposition 6.4.9. Let T be a bounded positive operator on H. Then there exists a unique positive operator U on H such that U 2 = T. Moreover, U is bounded, U commutes with every operator that commutes with T, and the range of U is dense in the closure of the range of T. Proof. We may assume that 0 T I. Deﬁne (un )n1 and (Un )n1 as in Lemmas 6.4.7 and 6.4.8. By Lemma 6.4.8, 0 Un I. Since Un is a polynomial in T, it commutes with every operator that commutes with T. By Lemma 6.4.7, for each ε > 0 there exists Nε such that |um (t) − un (t)| < ε for all t ∈ [0, 1] and all m, n Nε . It follows from Proposition 6.4.4 that Um x − Un x ε x for all x ∈ H and all m, n Nε . Hence (Un )n1 converges strongly to an operator U on H. Clearly, 0 U I and U commutes with every operator that commutes with T. Moreover, for each n 1, since un (t) is a strict polynomial (one without constant term) over R, we have ran(Un ) ⊂ ran(T ), from which it follows that

6.4 Functions of Selfadjoint Operators

ran (U ) ⊂ ran (T ).

179

(6.20)

With ε and Nε as before, we see that √ √ 2 un (t) − t = un (t) + t un (t) − t √ 2 un (t) − t = 2 lim |un (t) − um (t)| m→∞

2ε

for all t ∈ [0, 1] and all n Nε . It follows from Proposition 6.4.4 that 2 Un − T x 2ε x for all x ∈ H and all n Nε . Hence 2 U − T x U 2 − Un2 x + Un2 − T x = (U + Un ) (U − Un ) x + Un2 − T x 2 (U − Un ) x + Un2 − T x −→ 0 as n −→ ∞ and therefore U 2 = T. Moreover,

Tx = U

lim Un x ,

n→∞

which, taken with (6.20), shows that ran(U ) is dense in the closure of ran(T ) . Now suppose that we have another positive operator S such that S 2 = T. Then ST = SS 2 = S 2 S = T S, so S commutes with T , and therefore U commutes with S. Given x ∈ H, write y = U x − Sx. Then U y, y + Sy, y = (U + S) (U − S) x, y =

,

U 2 − S 2 x, y = 0.

But U y, y 0 and Sy, y 0, so we must have U y, y = 0 = Sy, y . Now apply the ﬁrst part of the proof to U, to obtain a positive operator A such that A2 = U. Then , 2 Ay = Ay, Ay = A2 y, y = U y, y = 0, so Ay = 0. Hence U y = A2 y = 0, and similarly, Sy = 0. It now follows that 2

U x − Sx = (U − S) (U − S) x, x = (U − S) y, x = 0. Hence S = U.

by

2

The operator U in Proposition 6.4.9 is called the square root of T and is denoted √ T or T 1/2 .

180

6 Operators and Locatedness

6.5 Locating the Kernel and the Range Let T be a jointed operator on the Hilbert space H. In this section we deal with questions like the following: When is ran (T ) located? Is the locatedness of ran (T ) linked to the locatedness of ker (T ) or that of ran (T ∗ )? When is ran (T ) closed? As we shall see, the answers depend on some interesting applications of the λtechnique. We begin with two elementary lemmas. Lemma 6.5.1. Let X and Y be orthogonal subspaces of a Hilbert space H such that X + Y is dense in H. Then both X and Y are located. Proof. Given z ∈ H, choose a sequence (xn )n1 in X and a sequence (yn )n1 in Y such that xn + yn −→ z as n −→ ∞. For m n we have 2

2

= (xm − xn ) + (ym − yn )

2

= (xm + ym ) − (xn + yn ) −→ 0 as n −→ ∞.

2

xm − xn + ym − yn

Hence (xn )n1 , (yn )n1 are Cauchy sequences in X, Y respectively, and so converge to respective limits x∞ , y∞ in H such that z = x∞ + y∞ , x∞ ⊥ y∞ , and y∞ ⊥ X. Given x ∈ X, we have 2

2

2

2

2

z − x = y∞ + (x∞ − x) = y∞ + x∞ − x y∞ , with equality when x = x∞ . It follows that ρ (z, X) exists and equals y∞ . Hence X, and likewise Y, is located. 2 Lemma 6.5.2. Let T be a jointed operator on H. Then ran(T )⊥ = ker(T ∗ ). Also, ran (T ) is located if and only if ran(T ) + ker (T ∗ ) is dense in H, in which case ker (T ∗ ) is located. Proof. First observe that y ⊥ ran(T ) ⇐⇒ ∀x ∈ H (x, T ∗ y = 0) ⇐⇒ T ∗ y = 0, so ran(T )⊥ = ker (T ∗ ). It follows that if ran(T ) is located, then so is ker (T ∗ ); moreover, if P is the projection of H on the closure of ran(T ), then since x = P x + (I − P ) x for each ∈ H, we see that ran(T ) + ker (T ∗ ) is dense in H. If, conversely, ran(T ) + ker (T ∗ ) is dense in H, then we can apply the preceding 2 lemma to show that ran(T ) and ker (T ∗ ) are both located.

6.5 Locating the Kernel and the Range

181

Lemma 6.5.3. If T is a jointed operator on H, then ran (T T ∗ ) is dense in ran (T ) . Proof. Let B denote the closed unit ball of H. By Theorem 6.3.4, T T ∗ (B) is located, so T T ∗ (nB) is located for each positive integer n. Given x ∈ H and ε > 0, we need only show that ρ (T x, T T ∗ (Bn )) < ε for some n. To this end, construct an increasing binary sequence (λn )n1 such that ε , 2 ∗ λn = 1 =⇒ ρ (T x, T T (nB)) < ε.

λn = 0 =⇒ ρ (T x, T T ∗ (nB)) >

We may assume that λ1 = 0. If λn = 0, then, applying the separation and the Riesz representation theorems, we construct a unit vector yn ∈ H such that ε T T ∗ z, yn + < T x, yn (z ∈ nB) . 2 Then ε 2 n T ∗ yn < T T ∗ (nyn ), yn + < T x, yn T x 2 2 and so T ∗ yn < n−1 T x . If λn = 1, we set yn = 0. Clearly, the sequence ∗ (T yn )n1 converges to 0. Choose a positive integer N such that x, T ∗ yN < ε/2. If λN = 0, then ε ε ε + T T ∗ (N yN ) , yN < T x, yN = x, T ∗ yN < , 2 2 2 a contradiction. Hence λN = 1. 2 For a ﬁrst application of Lemmas 6.5.2 and 6.5.3, we call an operator T on H sequentially open if for each sequence (xn )n1 such that (T xn )n1 converges to 0, there exists a sequence (yn )n1 in ker(T ) such that xn + yn −→ 0. Proposition 6.5.4. Let T be a sequentially open operator on H with an adjoint. Then ran (T ) and ker (T ∗ ) are located. Proof. By Lemma 6.5.3, there exists a sequence (xn )n1 in H such that T (T ∗ xn − x) −→ 0 as n −→ ∞. Since T is sequentially open, there exists a sequence (yn )n1 in ker (T ) such that T ∗ xn + yn −→ x as n −→ ∞. We now see that ran (T ∗ ) + ker (T ) is dense in H; whence, by Lemma 6.5.2, ran (T ) is located. 2 If T is any bounded operator with an adjoint, then T ∗ T is a bounded positive operator, so, by Proposition 6.4.9, the absolute value of T, √ |T | = T ∗ T , exists as a bounded positive operator.

182

6 Operators and Locatedness

Proposition 6.5.5. If T is a bounded operator on H with an adjoint, then ran (|T |) is dense in the closure of ran (T ∗ ) . √ Proof. We see from Proposition 6.4.9 that the range of T ∗ T is dense in the closure 2 of the range of T ∗ T. Reference to Lemma 6.5.3 completes the proof.

We have already seen, in Lemma 6.5.2, that, for a jointed operator T, the kernel of T ∗ is the orthogonal complement of the range of T, and that if ran (T ) is located, then so is ker (T ∗ ) . If, conversely, ker (T ∗ ) is located, is ran (T ) located also? Classically it is, since the closure of ran (T ) is the orthogonal complement of ker (T ∗ ) . However, if the latter holds constructively, then we can prove Markov’s principle: for if a is any real number such that ¬ (a = 0) , then the operator T : z az on C is selfadjoint and has kernel {0} ; but if ran (T ) is located, then ¬ (ρ (1, ran (T )) > 0), so ρ(1, ran(T )) = 0, ran (T ) contains nonzero elements, and therefore a = 0. We say that a sequence (xn )n1 in H converges weakly to x ∈ H, and we write w

xn −→ x as n −→ ∞ w

(or just xn −→ x), if limn→∞ xn , y = x, y for all y ∈ H. We call an operator T on H weak-sequentially open if for any sequence (xn )n1 such that T xn −→ 0, w there exists a sequence (yn )n1 in ker(T ) such that xn + yn −→ 0. Proposition 6.5.6. If T is a jointed operator on H such that ran(T ∗ ) is located, then T is weak-sequentially open. Proof. Let P be the projection of H on the closure of ran(T ∗ ). Let (xn )n1 be a sequence in H such that T xn −→ 0, and for each n set yn = P xn − xn . Then yn ∈ ran(T ∗ )⊥ = ker(T ). For each z ∈ H we have xn + yn , T ∗ z = xn , T ∗ z − (I − P )xn , P T ∗ z = T xn , z , so

| xn + yn , T ∗ z | T xn z −→ 0 as n −→ ∞.

It follows that for each x ∈ H we have xn + yn , P x −→ 0 and therefore xn + yn , x = P (xn + yn ) , x = xn + yn , P x −→ 0. Hence T is weak-sequentially open.

2

Theorem 6.5.7. Let T be a weak-sequentially open, jointed operator on H with located kernel. Then T ∗ has located range.

6.5 Locating the Kernel and the Range

183

Proof. Let P be the projection of H on ker(T ). It suﬃces to show that for each x ∈ H, the vector x − P x is in the closure of ran(T ∗ ): for then x − y2 = (x − P x) + y2 + P x2 for all y ∈ ran(T ∗ ), so ρ (x, ran(T ∗ )) exists and equals P x. Accordingly, ﬁx x in H and ε > 0. Denote the closed unit ball in H by B. By Theorem 6.3.4, T ∗ (nB) is located in H for each positive integer n; so we can construct an increasing binary sequence (λn )n1 such that ε , 2 λn = 1 =⇒ ρ (x − P x, T ∗ (nB)) < ε.

λn = 0 =⇒ ρ (x − P x, T ∗ (nB)) >

Without loss of generality we may assume that λ1 = 0. If λn = 0, then by Corollary 5.2.10 and the Riesz representation theorem, there exists a unit vector yn such that for each z ∈ nB, x − P x, yn > |T ∗ z, yn | + Taking z = n T yn

−1

ε ε = |z, T yn | + . 2 2

T yn , we obtain ε + n T yn < x − P x, yn x 2

and therefore T yn < n−1 x . On the other hand, if λn = 1 − λn−1 , we set yk = 0 for all k n. Clearly, the sequence (T yn )n1 converges to 0. But T is weak-sequentially open, so there exists a sequence (zn )n1 in ker(T ) such that w

yn + zn −→ 0. Choose N such that |x − P x, yn | = |x − P x, yn + zn |

ε/2, which is absurd. Hence λN = 1. Since ε > 0 is arbitrary, we are through. 2 Corollary 6.5.8. If T is a jointed operator on H, then the following four statements are equivalent: (a) ran (T ) is located. (b) ran(T ∗ ) is located. (c) ker(T ) is located and T is weak-sequentially open. (d) ker (T ∗ ) is located and T ∗ is weak-sequentially open. Proof. This follows from Lemma 6.5.2, Theorem 6.5.7, and Proposition 6.5.6.

2

184

6 Operators and Locatedness

Turning in a slightly diﬀerent direction, we move towards a proof of the closed range theorem: Theorem 6.5.9. Let T be a jointed operator on H whose range is closed. Then ran (T ) and ker (T ) are located, and ran(T ∗ ) is closed. To set the scene, consider how the closed range theorem is proved classically. One simple proof uses the polar decomposition of the operator T ∗ ; but the existence of an exact polar decomposition in constructive analysis requires the range of the operator to be located from the outset (Exercise 16). In another classical proof of Theorem 6.5.9 the idea is to show that ran(T ∗ T ) is closed and then to apply Lemma 6.5.3. To that end, let (xn )n1 be a sequence such that (T ∗ T xn )n1 converges to a limit y ∈ H. Applying the classical uniform boundedness theorem (see page 163) to the bounded linear functionals fn deﬁned on the Hilbert space ran(T ) by fn (T x) = T x, T xn = x, T ∗ T xn , we obtain M > 0 such that fn M for each n. Hence the linear functional T x x, y on ran(T ) is bounded by M. By the Riesz representation theorem, there exists x∞ ∈ H such that x, y = T x, T x∞ = x, T ∗ T x∞ for all x. It follows that y = T ∗ T x∞ ∈ ran(T ∗ T ). This proof fails constructively in two places: ﬁrst, in its use of the classical version of the uniform boundedness theorem, and second, in its application of the Riesz representation theorem, which requires the linear functional to be not just bounded but normed. Fortunately, as the following sequence of results will show, these diﬃculties can be overcome. We begin with two lemmas that prepare us for a general result about sequentially continuous linear mappings between normed spaces. Lemma 6.5.10. Let T : X −→ Y be a sequentially continuous linear mapping between normed spaces, (xn )n1 a Cauchy sequence in X, and 0 < α < β. Then either T xn < β for all n or else there exists n such that T xn > α. Proof. In view of the linearity of T, we may assume that β − α > 1. Choosing a strictly increasing sequence (Nk )k1 of positive integers such that xm − xn < 2−3k for all m, n Nk , write sk = max {T xn : 1 n Nk } . Construct an increasing binary sequence (λk )k1 such that λk = 0 =⇒ ∀j k (sj < β − 2−2j ), λk = 1 =⇒ ∃j k (sj > β − 2−2j+1 ).

6.5 Locating the Kernel and the Range

185

We may assume that λ1 = λ2 = 0. Now construct a sequence (zk )k1 in X as follows. If λk+1 = 0 or λk = 1, set zk = 0. If λk+1 = 1 and λk = 0, then T xNk sk < β − 2−2k and sk+1 > β − 2−2k−1 , so we can choose j such that Nk < j Nk+1 and T xj > β − 2−2k−1 ; setting zk = 22k (xj − xNk ), we have zk < 2−k . Moreover, T zk = 22k T xj − T xNk 22k (T xj − T xNk )

1 > 22k β − 2−2k−1 − (β − 2−2k ) = . 2 This completes the construction of a sequence (zk )k1 converging to 0 in X. By the sequential continuity of T, limk→∞ T zk = 0. Choose K such that T zk < 1/2 for all k K. If λK = 1, then there exists n NK such that T xn > β − 2−2n+1 > α. On the other hand, if λK = 0 and there exists k K such that λk+1 = 1 − λk , then T zk > 1/2, a contradiction. Thus if λK = 0, then λk = 0 for all k K and 2 therefore for all k, so T xk < β for all k. Lemma 6.5.11. Let T : X −→ Y be a sequentially continuous linear mapping between normed spaces, and (xn )n1 a Cauchy sequence in X. Then supn1 T xn exists. Proof. In view of the previous lemma, it is enough to show that the sequence (T xn )n1 is bounded. To do so, choose R > 0 such that xn R for all n. Taking α = 1 and β = 2 in Lemma 6.5.10, we may assume that there exists n1 such that T xn1 > 1. Set λ1 = 0. Using Lemma 6.5.10 repeatedly, we construct an increasing binary sequence (λn )n1 , and an increasing sequence (nk )k1 of positive integers, such that λk = 0 =⇒ T xnk > k and nk > nk−1 , λk = 1 =⇒ (T xn )n1 is a bounded sequence and nk+1 = nk . Assume that we have constructed λk and nk . If λk = 1, we set λk+1 = λk and nk+1 = nk . If λk = 0, then T xnj > j for all j k. We then apply Lemma 6.5.10 to the Cauchy sequence (xj )j>nk . Either we obtain nk+1 > nk such that T xn > k + 1, or else T xj < k + 2 for all j > nk . In the ﬁrst case we set k+1 λk+1 = 0, and in the second, noting that (T xn )n1 is bounded, we set λk+1 = 1

186

6 Operators and Locatedness

and nk+1 = nk . This completes the inductive construction of the sequences (λk )k1 and (nk )k1 . If λk = 0, set zk = k −1 xnk ; if λk = 1, set zk = 0. Then zk R/k for each k, so zk −→ 0 and therefore, by the sequential continuity of T, T zk −→ 0. Choose K such that T zk < 1 for all k K. If λK = 0, then T zk =

1 T xnk > 1, k

a contradiction. Hence λK = 1 and so the sequence (T xn )n1 is bounded.

2

Proposition 6.5.12. A sequentially continuous linear mapping T : X −→ Y between normed spaces maps Cauchy sequences to Cauchy sequences. Proof. Given a Cauchy sequence (xn )n1 in X, choose a strictly increasing sequence (Nk )k1 of positive integers such that xm − xn < 2−k for all m, n Nk . For each k, the sequence (xn − xNk )nNk is a Cauchy sequence; so, by Lemma 6.5.11, sk = sup T xn − T xNk nNk

exists. Given ε > 0, we construct an increasing binary sequence (λk )k1 such that ε , 4 ε λk = 1 =⇒ sk < . 2

λk = 0 =⇒ sk >

We may assume that λ1 = 0. If λk = 0, choose j Nk such that T xj − T xNk > ε/4, and set zk = xj − xNk . If λk = 1, set zk = 0. Then zk < 2−k for each k, so zk −→ 0. Since T is sequentially continuous, T zk −→ 0 and we can choose K such that T zk < ε/4 for all k K. If λK = 0, then T zK > ε/4, which is absurd; so λK = 1 and therefore sK < ε/2. It follows that for all j, k NK , T xj − T xk T xj − T xNk + T xk − T xNk

α for some n or else T xn < β for all n.

6.5 Locating the Kernel and the Range

187

Proof. Let (nk )k1 be a strictly increasing sequence of positive integers such that T ∗ T xj

α. We may assume that λ1 = 0. Deﬁne a sequence (yk )k1 in ran(T ) as follows: if λk = 0, set yk = 0; if λk = 1 − λk−1 , choose i with nk−1 < i nk and T xi > α, and set 1 T xi yj = kT xi for all j k. Then (yk )k1 is a Cauchy sequence: in fact, yj − yk 1/(k + 1) whenever j k. Since ran (T ) is closed in H and therefore complete, there exists z ∈ H such that (yk )k1 converges to T z. Choosing a positive integer N such that z < N α, consider any integer k N . If λk = 1 − λk−1 , then Tz =

1 T xi kT xi

for some i with nk−1 < i nk and T xi > α, so 1 1 1 1 1 α < T xi = T xi , T z = T ∗ T xi , z 2 z < 2 N α α, k k k k k a contradiction. Hence λk = λk−1 for all k N . It follows that either λk = 0 for all k, or else λk = 1 − λk−1 for some k N. In the ﬁrst case, T xn < β for all n; 2 in the second, T xi > α for some i with nk−1 < i nk .

Lemma 6.5.14. Under the hypotheses of Lemma 6.5.13, for all positive numbers α, β with α < β, either T xn > α for inﬁnitely many n or else T xn < β for all suﬃciently large n. Proof. Let (nk )k1 be a strictly increasing sequence of positive integers such that T ∗ T xj

α and set yk =

1 T xi ; kT xi

if λk = 1 − λk−1 , set yj = yk−1 for all j k. Then yj − yk 2/k for all j k; so (yk )k1 is a Cauchy sequence in the complete space ran (T ) and therefore converges to T z for some z ∈ H. Choosing a positive integer N such that z < N α, consider any integer k > N . If λk = 1 − λk−1 , then Tz =

1 T xi (k − 1)T xi

for some i nk−1 with T xi > α, so 1 1 α< T xi = T xi , T z = T ∗ T xi , z k−1 k−1 1 1 1 2 z < 2 N α k − 1 α, (k − 1) (k − 1) a contradiction. Hence λk = λk−1 for all k > N . It follows that either λk = 1 for some k N or else λk = 0 for all k. In the ﬁrst case, T xi < β for all i nk . In 2 the second case, for each k there exists i nk such that T xi > α.

Lemma 6.5.15. Under the hypotheses of Lemma 6.5.13, (T xn )n1 converges to 0. Proof. By Lemma 6.5.14, for each ε > 0 either T xn > ε/2 for inﬁnitely many n or else T xn < ε for all suﬃciently large n. In the former case, passing to an appropriate subsequence, we may assume that T xn > ε/2 and T ∗ T xn < 1/n2 for all n. Applying the uniform boundedness theorem (Corollary 6.2.12) to the normed linear functionals T x T x, nT xn on the Hilbert space ran(T ), we can ﬁnd T z ∈ ran(T ) such that the sequence (|T z, nT xn |)n1 is unbounded. But |T z, nT xn | = |z, nT ∗ T xn |

1 z −→ 0 as n −→ ∞. n

This contradiction rules out the possibility that T xn > ε/2 for inﬁnitely many n. Since ε > 0 is arbitrary, the result follows. 2

Lemma 6.5.16. Let H be a Hilbert space, and T a jointed operator on H with closed range. Then ran(T ∗ T ) is complete.

6.6 Baire’s Theorem, with Applications

189

Proof. First observe that, by Theorem 6.3.6, both T and T ∗ are sequentially continuous. It is easily seen using Lemma 6.5.2 that T ∗ is one-one on ran (T ). So we can deﬁne a linear mapping S : ran (T ∗ T ) −→ ran (T ) by setting ST ∗ T x = T x for each x ∈ H. Lemma 6.5.15 shows that S is sequentially continuous. Consider any Cauchy sequence (T ∗ T xn )n1 in ran(T ∗ T ) . By Proposition 6.5.12, (T xn )n1 is a Cauchy sequence in the Hilbert space ran (T ). Hence there exists x ∈ H such that T xn −→ T x. The sequential continuity of T ∗ now 2 yields T ∗ T xn −→ T ∗ T x. Thus ran(T ∗ T ) is complete. We can now give the proof of the closed range theorem. Proof. By Lemmas 6.5.3 and 6.5.16, ran (T ∗ T ) is both dense in ran(T ∗ ) and complete. Hence ran(T ∗ ) is complete and therefore closed in H. Moreover, for each x ∈ H there exists y ∈ H such that T ∗ x = T ∗ T y; so x = T y + (x − T y) , where T y ∈ ran (T ) and (by Lemma 6.5.2) x − T y ∈ ker (T ∗ ) = ran(T )⊥ . It follows from Lemma 6.5.2 that both ran (T ) and ker (T ∗ ) are located. Applying 2 Corollary 6.5.8, we now see that ran(T ∗ ) and ker (T ) are located.

6.6 Baire’s Theorem, with Applications Baire’s theorem states that The intersection of a sequence of dense open subsets of a complete metric space is dense in that space. The standard classical proof of this theorem (see [79], page 97) passes over unchanged to the constructive setting. However, various classically equivalent versions of Baire’s theorem do not pass over unscathed; for example, the version that states that if a complete metric space is the union of a sequence of closed subsets, then one of those subsets is inhabited. In this section we present a constructive proof of a restricted form of this last version of Baire’s theorem in the context of a Banach space, and apply it to operator theory. The proof of our version of Baire’s theorem introduces yet another technique, in which we show that a certain property P holds by constructing an element of the set {x ∈ X : P }, which is empty if ¬P holds and inhabited if P holds.

190

6 Operators and Locatedness

Theorem 6.6.1. Let X be a Banach space, and C a closed, convex, balanced, lo-◦ nC and the distance ρ (0, −C) exists. Then C cated subset of X such that X = n1

is inhabited. Proof. For each positive integer n deﬁne the open set Un = −nC ∪ {x ∈ X : C ◦ is inhabited} , where −nC is the metric complement of nC in X. To prove that Un is dense in X, consider y ∈ X and ε > 0. Note that ρ(y, nC) exists and equals nρ n−1 y, C . Either ρ (y, nC) > 0 or ρ (y, nC) < ε. In the ﬁrst case, y ∈ −nC. In the second, choose z ∈ nC such that y − z < ε. Noting that ρ (0, −2nC) = 2nρ (0, −C) exists, we see that either ρ (0, −2nC) < 2ε or ρ (0, −2nC) > ε. In the former case, choose z ∈ −2nC such that z < 2ε. For each w ∈ nC we have −w ∈ nC (since C is balanced), so z − w ∈ nC + nC = 2nC, by the convexity of C. Hence (z − z ) − w = z − (z − w) ρ (z , 2nC) > 0. Thus z − z ∈ −nC. Since also y − (z − z ) y − z + z < 3ε, we see that ρ (y, Un ) < 3ε. Finally, in the case ρ (0, −2nC) > ε, for each x with x ε we have ρ (x, 2nC) = 0. Hence B (0, ε) ⊂ 2nC = 2nC, ◦

so (2nC) , and therefore C ◦ , is inhabited. In this case, Un = X. This completes the proof that Un is dense in X. Since X is complete, it follows from the standard version of Baire’s theorem Un is dense in X and therefore, in particular, contains a point ξ. Choose that n1

n such that ξ ∈ nC. Since also ξ ∈ Un , we must have ξ ∈ {x ∈ X : C ◦ is inhabited} . Hence C ◦ is indeed inhabited.

2

Recall the classical open mapping theorem for bounded linear mappings: a bounded linear mapping T of a Banach space X onto a Banach space Y is open, in

6.6 Baire’s Theorem, with Applications

191

the sense that it maps open subsets of X onto open subsets of Y (or, equivalently, it maps the open unit ball of X onto an open subset of Y ). This theorem does not hold constructively without some additional hypotheses: the inverse S of the mapping T in the example on page 169 is a bounded operator of a Hilbert space onto a Hilbert space, but if S is open, then T is bounded. In other words, if every bounded linear mapping of a Hilbert space onto a Hilbert space is open, then BD-N holds. Nonetheless, we can prove constructive versions of the open mapping theorem that are classically equivalent to the standard version. To that end, we begin with a lemma. Lemma 6.6.2. Let T be a sequentially continuous linear mapping of a Banach space X into a normed space Y such that T (BX (0, 1)) is located. Let r be a positive number, and y an element of BY (0, r). There exists x ∈ B X (0, 2) such that if y = T x, then ρ (z, T (BX (0, 1))) > 0 for some z ∈ BY (0, r). Proof. If ρ (y, T (BX (0, 1))) > 0, then we can take x to be any element of BX (0, 2). So we may assume that ρ (y, T (BX (0, 1))) < r/2. Choosing x1 ∈ BX (0, 2) such

that y − T 12 x1 < r/2 and therefore 2y − T x1 < r, we set λ1 = 0. This is the ﬁrst step in the inductive construction of an increasing binary sequence (λn )n0 and a sequence (xn )n0 of elements of BX (0, 2) such that for each n 1, if λn = 0, then ρ

2n−1 y −

n−1

2n−1−i T xi

, T (BX (0, 1))

0

i=1

and xi = 0 for all i n. Suppose that we have found λn−1 and xn−1 with the applicable properties. If λn−1 = 1, we set λn = 1 and xn = 0. If λn−1 = 0, we consider the two cases n−1 n−1 n−1−i ρ 2 y− 2 T xi , T (BX (0, 1)) > 0 i=1

and

192

6 Operators and Locatedness

ρ

n−1

2

y−

n−1

n−1−i

2

T xi

, T (BX (0, 1))

2−N r y −

and therefore

N N 2N −i T xi > r. 2 y −

i=1

i=1

We must therefore have λN = 1; so there exists n N such that λn = 1 − λn−1 . Setting n−1 2n−1−i T xi , z = 2n−1 y − i=1

we see that ρ (z, T (BX (0, 1))) > 0 (as λn = 1) and z < r (as λn−1 = 0). This completes the proof. 2 Lemma 6.6.3. Let C be a balanced convex subset of a normed space Y, and let y ∈ Y and r > 0 be such that B(y, r) ⊂ C. Then B(0, r) ⊂ C. Proof. If z ∈ Y and z < r, then y±z ∈ C and therefore z = C.

1 2

(y + z)− 12 (y − z) ∈ 2

Here is our version of the open mapping theorem. Theorem 6.6.4. Let X, Y be Banach spaces, and T a sequentially continuous linear mapping of X onto Y such that T (B(0, 1)) is located and ρ (0, −T (B (0, 1))) exists. Then T is an open mapping.

6.6 Baire’s Theorem, with Applications

Proof. Since Y =

193

n T (B (0, 1)) ,

n1

we can apply Theorem 6.6.1 to compute y0 ∈ Y, R > 0, and a positive integer N such that BY (y0 , R) ⊂ N T (B (0, 1)) . Writing y1 = N −1 y0 and r = R/N, we obtain BY (y1 , r) ⊂ T (B (0, 1)). It follows from Lemma 6.6.3 that BY (0, r) ⊂ T (B (0, 1)).

(6.21)

Now consider any y ∈ Y with y < r. Choose x ∈ B (0, 2) as in the conclusion of Lemma 6.6.2. If y = T x, then there exists z ∈ BY (0, r) such that ρ (z, T (B (0, 1))) > 0, which contradicts (6.21); hence (the inequality on a normed space being tight) y = T x. Thus BY (0, r) ⊂ T (B (0, 2)) , and therefore T is an open mapping. 2

A subset C of a linear space X is said to be a generating set for, or to generate, X if every element of X is a ﬁnite linear combination of elements of C. Our next lemma will enable us to prove that compactly generated Banach spaces are ﬁnitedimensional. Lemma 6.6.5. Let G be a compact generating set for a nontrivial Banach space X. Then there exists a compact generating set C for X that is both convex and balanced such that ρ (0, −C) exists. Proof. We may assume that G is both convex and balanced (the proof is left as an exercise). Now, 2G is compact, the mapping x ρ (x, G) is uniformly continuous on X, and X is nontrivial. Hence there exists δ > 0 such that both the sets C = {x ∈ 2G : ρ (x, G) δ} , D = {x ∈ 2G : ρ (x, G) δ} are compact. Note that C is convex and balanced, and, since it contains G, generates X. We show that −C is dense in D. To this end, consider any x ∈ D and any ε > 0. Choose t > 1 such that (t − 1) x < ε/2, and suppose that tx ∈ C. Then, since C is balanced, x ∈ C; whence x ∈ C ∩ D and therefore ρ (x, G) = δ. But then for each g ∈ G we have tx − g = t x − t−1 g tρ (x, G) = tδ; so ρ(tx, G) tδ > δ, which is absurd since tx ∈ C. We conclude that tx ∈ / C. It follows from Proposition 3.1.2 that there exists y ∈ ∼C such that tx − y < ε/2

194

6 Operators and Locatedness

and therefore x − y < ε. Applying Bishop’s lemma (Proposition 3.1.1), we see that y ∈ −C. This completes the proof that −C is dense in D. Since the norm function is uniformly continuous on the compact set D, it now follows that ρ (0, −C) = inf {x : x ∈ −C} = inf {x : x ∈ D} 2

exists.

A normed space with a compact generating set is said to be compactly generated. Theorem 6.6.6. A compactly generated Banach space is ﬁnite-dimensional. Proof. Let X be a compactly generated Banach space. We ﬁrst suppose that X contains a nonzero vector. By Lemma 6.6.5, X has a balanced, convex, compact nC, we can apply Thegenerating set C such that ρ (0, −C) exists. Since X = n1

orem 6.6.1 to show that C ◦ is inhabited; whence C contains a nontrivial ball. But every ball in a normed space is located, so the ball in question is totally bounded. It follows from Proposition 4.1.13 that X is ﬁnite-dimensional. It remains to remove the restriction that X be nontrivial. To do this, we work in the product Banach space X × K. This space is generated by the compact set G × {1}; so, by the foregoing, X × K is ﬁnite-dimensional. It follows that X, being isomorphic to the quotient space (X × K)/ K, is ﬁnite-dimensional. 2 Recall from page 102 that a linear mapping T : X −→ Y between normed spaces is said to be compact if T (BX (0, 1)) is a totally bounded subset of Y. Corollary 6.6.7. If T is a compact linear mapping of a normed space X onto a Banach space Y, then Y is ﬁnite-dimensional. Proof. The totally bounded set T (B(0, 1)) generates Y. Since Y is complete, it follows that T (B(0, 1)) is a compact generating set for Y ; whence, by Theorem 6.6.6, Y is ﬁnite-dimensional. 2

In the traditional development of functional analysis, the open mapping theorem is used to prove Banach’s inverse mapping theorem, the closed graph theorem, and the uniform boundedness theorem. The last of these three we have already discussed, in Section 2. To deal with the inverse mapping theorem we need another lemma about convex sets, and two further technical lemmas. Lemma 6.6.8. Let C be a convex, absorbing subset of a Banach space X. Then 0∈ / −C.

6.6 Baire’s Theorem, with Applications

195

Proof. Assuming that 0 ∈ −C, we ﬁrst show that X − nC is dense in X for each positive integer n. To do so, construct a sequence (xn )n1 in −C that converges to 0. Fixing a positive integer n, an element y of X, and ε > 0, compute positive numbers r, δ and a positive integer k such that −y ∈ rC, xk < ε/ (n + r), and xk − z δ for all z ∈ C. Let y1 = y + (n + r) xk . Then y − y1 = (n + r) xk < ε. On the other hand, since C is convex, for each z ∈ nC we have 1 1 (z − y) ∈ (nC + rC) = C, n+r n+r 1 (n + r) δ. (z − y) y1 − z = (n + r) − x k n+r

so

Hence y1 ∈ −nC. Since ε > 0 is arbitrary, it follows that −nC is dense in X. Moreover, being a metric complement, −nC is open. Applying the standard form −nC is inhabited, which is absurd since of Baire’s theorem, we now see that n1 nC. We conclude that 0 ∈ / −C. 2 X= n1

Sometimes when we want to prove that a certain proposition P is absurd, we ﬁrst prove that P implies LPO, and then (frequently by adapting a classical proof) show that the addition of LPO to our intuitionistic logic suﬃces for us to prove that P is false. The next lemma will enable us to rule out in this way an unwanted alternative in the proof of Banach’s inverse mapping theorem. Lemma 6.6.9. Let f be a strongly extensional mapping of a complete metric space X into a metric space Y. Suppose that there exist α > 0 and a sequence (xn )n1 converging to x in X such that ρ (f (x), f (xn )) > α for each n. Then LPO holds. Proof. Given an increasing binary sequence (λn )n1 with λ1 = 0, construct a sequence (zn )n1 in X such that •

if λk = 0 for all k n, then zn = x, and

•

if λn = 1 − λn−1 , then zk = xn for all k n.

Then (zn )n1 is a Cauchy sequence in X and so converges to a limit z ∈ X. Either f (x) = f (z) or else ρ (f (x), f (z)) < α. In the ﬁrst case, the strong extensionality of f shows that x = z; whence there exists n such that x = zn and therefore λn = 1. In the second case, if there exists n with λn = 1 − λn−1 , then we obtain the con2 tradiction ρ (f (x), f (z)) = ρ (f (x), f (xn )) > α; hence λn = 0 for all n.

196

6 Operators and Locatedness

Lemma 6.6.10. If LPO holds, then every separable subset of a metric space is located. Proof. Assuming LPO, consider a separable subset S of a metric space X. Let (sn )n1 be a dense sequence in S, let x ∈ X, and let 0 < α < β. Construct a binary sequence (λn )n1 such that

λn = 0 =⇒ ∀k n ρ(x, sk ) < 12 (α + β) , λn = 1 =⇒ ∃k n (ρ(x, sk ) > α) . By LPO, either λn = 0 for all n, in which case, since (sn )n1 is dense in S, ρ(x, s) 12 (α + β) < β for all s ∈ S; or else there exists n with λn = 1 and therefore ρ(x, sk ) > α for some k n. It follows from the constructive least-upperbound principle that ρ (x, S) exists. 2

This brings us to Ishihara’s version of Banach’s inverse mapping theorem. Theorem 6.6.11. Let T be a one-one, sequentially continuous linear mapping of a separable Banach space X onto a Banach space Y. Then T −1 is sequentially continuous. Proof. Let (xn )n1 be a sequence in X such that T xn −→ 0, and let ε > 0. By Corollary 3.1.7, the inverse linear mapping T −1 : Y −→ X is strongly extensional; whence, by Lemma 3.2.2, either xn < ε for all suﬃciently large n, or else xn > ε/2 for inﬁnitely many n. It suﬃces to rule out the latter case. To do so, we may assume that xn > ε/2 for all n. By Lemma 6.6.9, LPO holds; so, by Lemma 6.6.10, every separable subset of Y is located. Let (an )n1 be a dense sequence in

B X (0, 1). Since T is sequentially continuous, (T an )n1 is dense in T B X (0, 1) , which is therefore located. Writing xn = 2ε−1 xn , we see that xn > 1 and that T xn −→ 0 as n −→ ∞. Since T −1 is strongly extensional,

T xn ∈ ∼T B X (0, 1) . We now apply Lemma 6.6.2 with y = T (2xn ), to produce zn ∈ X such that

T zn < T xn + n−1 and ρ T (2zn ) , T B X (0, 1) > 0. Then Hence

ρ T zn , T B X (0, 12 ) > 0 and T zn −→ 0.

0 ∈ −T BX 0, 12 .

This contradicts Lemma 6.6.8, since T BX 0, 12 is both convex and absorbing. 2

6.6 Baire’s Theorem, with Applications

197

Recall that the graph of a mapping T : X −→ Y is the set G (T ) = {(x, T x) : x ∈ X} . As classically, Banach’s inverse mapping theorem leads to a version of the closed graph theorem: Corollary 6.6.12. Let T be a linear mapping of a Banach space X into a Banach space Y such that G (T ) is closed and separable. Then T is sequentially continuous. Proof. The mapping p : (x, T x) x of the Banach space G (T ) onto X is one-one and bounded linear. It follows from Theorem 6.6.11 that the inverse linear map is sequentially continuous, and hence that T is sequentially continuous. 2

In the theory of unbounded operators, the graph plays a signiﬁcant role. Particularly important properties for such a graph are closedness and locatedness. See [26, 83, 92] for more on such matters.

Exercises 1. Complete the proof of Proposition 6.1.1. 2. Complete the proof of Proposition 6.1.4. 3. Prove that a bounded linear mapping T of a normed space X into Cn is compact if and only if f ◦ T is normed for each linear functional f on Cn . Prove that if also the sum of any two normed linear functionals on X is normed, then T is compact if and only if pk ◦ T is normed for each k, where pk denotes the mapping (z1 , . . . , zn ) zk on Cn . 4. Let S be a compact operator on a Hilbert space H, and A a bounded operator on H. Prove that (a) λS is compact for each λ ∈ C; (b) S ∗ exists and is compact; (c) AS is compact. Prove also that if A∗ exists, then SA is compact. 5. A subset S of a normed space X is said the be weakly totally bounded if it is totally bounded relative to the locally convex structure deﬁned on X by the seminorms x |f (x)| , with f a normed linear functional on X. Prove the equivalence of the following conditions on X.

198

6 Operators and Locatedness

(a) The sum of any two normed linear functionals on X is normed. (b) The unit ball of X is weakly totally bounded. (Hint: To prove that (a) implies (b), note the second part of Exercise 3.) 6. An operator on a Hilbert space H is said to be weakly compact if it maps the unit ball of H to a weakly totally bounded subset of H. Prove that an operator T on H has an adjoint if and only if it is weakly compact. 7. Let (en )n1 be an orthonormal basis of a separable Hilbert space H, and T an operator on H. Prove that the following conditions are equivalent: (a) T has an adjoint. (b)

∞ n=1

(c)

∞ n=1

2

|T en , ek | converges for each k. 2

|T en , y| converges for each y ∈ H.

8. Let T be a weak-sequentially open operator on a Hilbert space such that ker (T ) is located. Prove that T is well-behaved (see Exercise 3 of Chapter 3). 9. Prove that every bounded linear mapping of a normed space onto a ﬁnitedimensional Banach space is an open mapping. 10. Show that the statement “every normed linear mapping T of a Hilbert space onto a ﬁnite-dimensional Banach space is compact” implies an omniscience principle. 11. Show that the statement “every one-one compact linear mapping T of a Hilbert space into a ﬁnite-dimensional Banach space has ﬁnite-dimensional range” implies an omniscience principle. 12. Show that the existence of the norms of the operators can be removed from the hypotheses of Royden’s version of the uniform boundedness theorem (Theorem 6.2.11). 13. Complete the details of the proof of Proposition 6.4.6. 14. A partial isometry is a jointed operator U on a Hilbert space H for which there exists a projection P, called the initial projection of U, such that U x = x ⊥ for all x ∈ ran (P ), and U x = 0 for all x ∈ ran (P ) . Prove that U ∗ U = P, that ∗ ∗ U U is a projection, and that ran (U U ) = ran (U ) . 15. Let T be a bounded operator such that T ∗ exists and has located range, and let P be the projection of H on ran (T ∗ ). Prove that there exists a partial isometry

6.6 Baire’s Theorem, with Applications

199

U on H whose initial projection is P such that T = U |T | and |T | = U ∗ T. Prove also that U U ∗ is the projection of H on ran (T ). (The expression of T as U |T | is called the polar decomposition of T, and is analogous to the modulus–argument form of a complex number.) 16. Prove that if a bounded jointed operator T on a Hilbert space H has a polar decomposition, then ran (T ) is located. 17. Prove the converse of Proposition 6.5.12: if a linear mapping T between normed spaces maps Cauchy sequences to Cauchy sequences, then T is sequentially continuous. 18. Let E be a dense linear subspace of a normed space X, and T a sequentially continuous linear mapping of E into a Banach space Y. Show that T extends to a sequentially continuous linear mapping of X into Y. 19. Prove that if G is a compact generating set for a Banach space X, then there exists a compact generating set G for X that is balanced and convex. 20. Let G be a balanced convex generating set for a normed space X. Prove that for each x ∈ X and each ε > 0 there exist t > 0 and g ∈ G such that x − tg ε. Can we replace ε by 0 in this result?

Notes Many of the results in this chapter come from papers by Ishihara and the authors. The work of Sections 1 and 2 is largely drawn from [61] and [63]. For related classical material see [86]. It is not known whether Theorems 6.2.1 and 6.2.2 can be extended to possibly unbounded convex subsets of a normed space. More general versions of Exercises 3–6 appeared in Ishihara’s thesis [56]. Since we cannot guarantee that a bounded operator on a Hilbert space has an adjoint, when we discuss such matters as the Gelfand representation theorem for an operator algebra A, we need to postulate that A is selfadjoint in the sense that each of its elements has an adjoint that also belongs to A; see [9, 12]. The principle BD-N, introduced by Ishihara in [59], has the unusual feature of being provable classically, intuitionistically, recursively, but not, apparently, within BISH. It is an interesting problem—a part of constructive reverse mathematics [7, 62]—to identify classical theorems that are equivalent to BD-N. The usual classical proof of Lemma 6.4.7 is based on the nonconstructive monotone convergence theorem for sequences; see [47] (7.3.1.1). The constructive proof is much more informative, in that it provides the rate of convergence of the sequence of functions. A full constructive analysis of Dini’s theorem is given in [8]; see also [21].

200

6 Operators and Locatedness

The existence of the square root of a selfadjoint operator is a special case of a more general result, the spectral theorem for sequences of commuting selfadjoint operators, which enables us to construct more general functions of an operator. Since that theorem requires measure theory, which we do not touch in this book, we refer the reader to the relevant chapters of [9] and [12]. It is easy to prove Lemma 6.5.3 classically, taking orthogonal complements in the identity ker(T ∗ T ) = ker(T ∗ ). For more on polar decompositions see [35]. Proposition 6.5.12 is trivial in CLASS, since in that context sequential continuity for linear maps implies uniform continuity. With classical logic, Theorem 6.6.1 is a simple consequence of the standard form of Baire’s theorem, the hypotheses of locatedness and the existence of ρ(0, −C) being redundant. For other, constructively inequivalent, versions of Baire’s theorem, see Chapter 2 of [34]. For more on open mapping theorems see [29] and [25]. It is interesting that sequential continuity, rather than boundedness, is the best we can get in the constructive versions of Banach’s inverse mapping theorem and the closed graph theorem. An extension of the former is given in [64].

What can be said at all can be said clearly; and whereof one cannot speak thereof one must be silent. —Ludwig Wittgenstein, Tractatus Logico-Philosophicus

References

1. O. Aberth, Computable Analysis, McGraw-Hill, New York, 1980. 2. O. Aberth, Computable Calculus, Academic Press, San Diego, CA, 2001. 3. P. Aczel and M. Rathjen, Notes on Constructive Set Theory, Report No. 40, Institut Mittag-Leﬄer, Royal Swedish Academy of Sciences, 2001. 4. G. Bachman and L. Narici, Functional Analysis, Academic Press, New York, 1966. 5. A. Bauer, Realizability as the connection between computable and constructive mathematics, Proceedings of CCA 2005, Kyoto, Japan, 25–29 August 2005; to appear. 6. M.J. Beeson, Foundations of Constructive Mathematics, Ergeb. Math. und ihrer Grenzgebiete, 3. Folge, 6, Springer-Verlag, Heidelberg, 1985. 7. J. Berger and D.S. Bridges, A bizarre property equivalent to the Π10 -fan theorem, Logic J. of the IGPL, to appear. 8. J. Berger and P.M. Schuster, Classifying Dinis theorem, to appear in Notre Dame J. Formal Logic. 9. E.A. Bishop, Foundations of Constructive Analysis, McGraw-Hill, New York, 1967. 10. E.A. Bishop, Mathematics as a numerical language, in: Intuitionism and Proof Theory (J. Myhill, A. Kino, and R.E. Vesley, eds.), 53–71, North-Holland, Amsterdam, 1970. 11. E.A. Bishop, Schizophrenia in contemporary mathematics, in: Errett Bishop: Reﬂections on Him and His Research (M. Rosenblatt, ed.), 1–32, Contemporary Math. 39, Amer. Math. Soc., Providence, RI, 1985. 12. E.A. Bishop and D.S. Bridges, Constructive Analysis, Grundlehren der math. Wissenschaften 279, Springer-Verlag, Heidelberg, 1985. 13. E.A. Bishop and H. Cheng, Constructive Measure Theory, Mem. Amer. Math. Soc. 116, 1972. 14. D.S. Bridges, On weak operator compactness of the unit ball of L(H), Zeit. math. Logik Grundlagen Math. 24, 493–494, 1978. 15. D.S. Bridges, Recent progress in constructive approximation theory, in The L.E.J. Brouwer Centenary Symposium (A.S. Troelstra and D. van Dalen, eds.), 41–50, North–Holland, Amsterdam, 1982. 16. D.S. Bridges, A constructive Morse theory of sets, in Mathematical Logic and Its Applications (D. Skordev, ed.), 61–79, Plenum Publishing Corp., New York, 1987. 17. D.S. Bridges, Locatedness, convexity, and Lebesgue measurability, Quart. J. Math. Oxford (2) 39, 411–421, 1988.

204

References

18. D.S. Bridges, A constructive look at the real number line, in: Real Numbers, Generalizations of the Reals, and Theories of Continua (P. Ehrlich, ed.), 29–92, Synthese Library 242, Kluwer Academic Publishers, Dordrecht, 1994. 19. D.S. Bridges, Constructive mathematics: a foundation for computable analysis, Theoretical Computer Science 219(1–2), 95–109, 1999. 20. D.S. Bridges, Constructive methods in mathematical economics, in Mathematical Utility Theory, J. Econ. (Zeitschrift f¨ ur National¨ okonomie), Suppl. 8, 1–21, 1999. 21. D.S. Bridges, Dinis Theorem: a constructive case study, in: Combinatorics, Computability and Logic (Proceedings of DMTCS01, Constant¸a, Romania, 2–6 July 2001; C.S. Calude, M.J. Dinneen, S. Sburlan, eds.), 69–80, DMTCS Series 17, SpringerVerlag, London, 2001. 22. D.S. Bridges, A weak constructive sequential compactness property and the fan theorem, Logic J. of the IGPL 13(2), 151–158, 2005. 23. D.S. Bridges and N.F. Dudley Ward, Constructing ultraweakly continuous linear functionals on B(H) , Proc. American Math. Soc. 126(11), 3347–3353, 1998. 24. D.S. Bridges and H. Ishihara, Linear mappings are fairly well-behaved, Arch. Math. 54, 558–562, 1990. 25. D.S. Bridges and H. Ishihara, A deﬁnitive constructive open mapping theorem?, Math. Logic Quarterly 44, 545–552, 1998. 26. D.S. Bridges and H. Ishihara, Constructive unbounded operators, in: Reuniting the Antipodes—Constructive and Nonstandard Views of the Continuum (Proceedings of the Symposion in San Servolo/Venice, Italy, May 17–22, 1999; U. Berger, H. Osswald, and P.M. Schuster, eds.), 25–37, Synthese Library, Kluwer Academic Publishers, Dordrecht, December 2001. 27. D.S. Bridges, H. Ishihara, and P.M. Schuster, Sequential Compactness in Construc¨ tive Analysis, Osterr. Akad. Wiss. Math.–Natur. Kl. Sitzungsber. II. 208, 159–163, 1999. 28. D.S. Bridges, H. Ishihara, and L.S. Vˆıt¸˘ a, Computing inﬁma on convex sets, with applications in Hilbert spaces, Proc. Amer. Math. Soc. 132(9)2723–2732, 2004. 29. D.S. Bridges, W.H. Julian, and R. Mines, A constructive treatment of open and unopen mapping theorems, Zeit. math. Logik Grundlagen Math. 35, 29–43, 1989. 30. D.S. Bridges and A. Mahalanobis, Bounded variation implies regulated: a constructive proof, J. Symb. Logic 66(4), 1695–1700, 2001. 31. D.S. Bridges and R. Mines, Sequentially continuous linear mappings in constructive analysis, J. Symb. Logic 63(2), 579–583, 1998. 32. D.S. Bridges and G. Popa, Exact, continuous boundary crossings out of convex sets in RN , Quart. J. Math. 54, 391–398, 2003. 33. D.S. Bridges and S. Reeves, Constructive mathematics, in theory and programming practice, Philosophia Mathematica 7(1), 65–104, 1999. 34. D.S. Bridges and F. Richman, Varieties of Constructive Mathematics, London Math. Soc, Lecture Notes 97, Cambridge Univ. Press, 1987. 35. D.S. Bridges, F. Richman, and P.M. Schuster, Adjoints, absolute values, and polar decompositions, J. Operator Theory 44, 243–254, 2000. 36. D.S. Bridges, F. Richman, and Y. Wang, Sets, complements and boundaries, Proc. Koninklijke Nederlandse Akad. Wetenschappen (Indag. Math. N.S.) 7(4), 425–445, 1996. 37. D.S. Bridges and L.S. Vˆıt¸a ˘, Weak continuity properties in constructive analysis, Logic J. of the IGPL 7(3), 277–281, 1999. 38. D.S. Bridges and L.S. Vˆıt¸˘ a, Apartness spaces as a framework for constructive topology, Ann. Pure Appl. Logic 119 (1–3), 61–83, 2003.

References

205

39. D.S. Bridges and L.S. Vˆıt¸˘ a, The constructive uniqueness of the locally convex topology on Rn , in: From Sets and Types to Topology and Analysis (L. Crosilla and P.M. Schuster, eds.), 304–315, Oxford Logic Guides 32, Clarendon Press, Oxford, 2005. 40. D.S. Bridges and L.S. Vˆıt¸a ˘, Apartness Spaces, book, in preparation. 41. L.E.J. Brouwer, Over de Grondslagen der Wiskunde, doctoral thesis, University of Amsterdam, 1907. Reprinted with additional material (D. van Dalen, ed.) by Matematisch Centrum, Amsterdam, 1981. 42. R.L. Constable et al., Implementing Mathematics with the Nuprl Proof Development System, Prentice-Hall, Englewood Cliﬀs, New Jersey, 1986. 43. L. Crosilla and P.M. Schuster (eds.), From Sets and Types to Topology and Analysis, Oxford Logic Guides 32, Clarendon Press, Oxford, 2005. 44. D. van Dalen (ed.), Brouwers Cambridge Lectures on Intuitionism, Cambridge University Press, Cambridge, 1981. 45. D. van Dalen, Mystic, Geometer, and Intuitionist, Clarendon Press, Oxford, 1999 (Vol. 1) and 2005 (Vol. 2). 46. R. Diaconescu, Axiom of choice and complementation, Proc. Amer. Math. Soc. 51, 176–178, 1975. 47. J. Dieudonn´e, Foundations of Modern Analysis, Academic Press, New York, 1960. 48. M.A.E. Dummett, Elements of Intuitionism (2nd edition), Oxford Logic Guides 39, Clarendon Press, Oxford, 2000. 49. H.M. Edwards, Essays in Constructive Mathematics, Springer Science + Business Media, Inc., New York, 2005. 50. N. D. Goodman and J. Myhill, Choice implies excluded middle, Zeit. math. Logik und Grundlagen Math. 24, 461. 51. S. Hayashi and H. Nakano, PX: A Computational Logic, MIT Press, Cambridge MA, 1988. 52. A. Heyting, Die formalen Regeln der intuitionistischen Logik, Sitzungsber. preuss. Akad. Wiss. Berlin, 42–56, 1930. ¨ 53. D. Hilbert, Uber das Unendliche, Mathematische Annalen 95 , 161–190, 1926; translated in Philosophy of Mathematics (P. Benacerraf and H. Putnam, eds.), 183–201, Cambridge University Press, Cambridge, 1964. 54. D. Hilbert, Die Grundlagen der Mathematik, Abhandlungen Math. Sem. Univ. Hamburg V, 65–85, 1927. 55. H. Ishihara,On the constructive Hahn–Banach theorem, Bull. London Math. Soc. 21, 79–81, 1989. 56. H. Ishihara, Boundedness, normability and compactness of constructive linear mappings, Ph.D. dissertation, Tokyo Institute of Technology, Tokyo, 1990. 57. H. Ishihara, Continuity and nondiscontinuity in constructive analysis, J. Symb. Logic 56(4), 1349–1354, 1991. 58. H. Ishihara, Constructive compact operators on a Hilbert space, Ann. Pure Appl. Logic 52, 31–37, 1991. 59. H. Ishihara, Continuity properties in constructive analysis, J. Symb. Logic 57, 557– 565, 1992. 60. H. Ishihara, A constructive version of Banachs inverse mapping theorem, New Zealand J. Math 23, 71–75, 1994. 61. H. Ishihara, Locating subsets of a Hilbert space, Proc. Amer. Math. Soc. 129(5), 1385–1390, 2001. 62. H. Ishihara, Constructive reverse mathematics: compactness properties, in: From Sets and Types to Topology and Analysis (L. Crosilla and P.M. Schuster, eds.), 245–267, Oxford Logic Guides 32, Clarendon Press, Oxford, 2005.

206

References

63. H. Ishihara and L.S. Vˆıt¸˘ a, Locating subsets of a normed space, Proc. Amer. Math. Soc. 131(10), 3231–3239, 2003. 64. H. Ishihara and L.S. Vˆıt¸˘ a, A constructive Banach inverse mapping theorem in Fspaces, New Zealand J. Math., to appear. 65. R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Algebras (Vol. 1), Academic Press, New York, 1988. 66. B.A. Kushner, Lectures on Constructive Mathematical Analysis, Amer. Math. Soc., Providence RI, 1985. 67. A.A. Markov, Theory of Algorithms (Russian), Trudy Mat. Istituta imeni V.A. Steklova 42 (Izdatelstvo Akademi Nauk SSSR, Moskva), 1954; English translation by J.J. Schoor-Kan and PST staﬀ, Israel Program for Scientiﬁc Translations, Jerusalem, 1961. 68. P. Martin-L¨ of, Notes on Constructive Mathematics, Almqvist and Wiksell, Stockholm, 1970. 69. P. Martin-L¨ of, An Intuitionistic Theory of Types: Predicative Part, in: Logic Colloquium 1973 (H.E. Rose and J.C. Shepherdson, eds.), 73–118, North-Holland, Amsterdam, 1975. 70. P. Martin-L¨ of, Constructive mathematics and computer programming, in Proc. 6th. Int. Congress for Logic, Methodology and Philosophy of Science (L. Jonathan Cohen, ed.), 153–179, North-Holland, Amsterdam, 1980. 71. G. Metakides, A. Nerode, and R. Shore, Recursive limits on the Hahn–Banach theorem, in: Errett Bishop: Reﬂections on Him and His Research (M. Rosenblatt, ed.), 85–91, Contemporary Math. 39, Amer. Math. Soc., Providence, R.I., 1985. 72. R. Mines, F. Richman, and W. Ruitenburg, A Course in Constructive Algebra, Universitext, Springer-Verlag, Heidelberg, 1988. 73. J.R. Moschovakis, The eﬀect of Markovs principle on the intuitionistic continuum, preprint, UCLA, April 2005. 74. J. Myhill, Constructive set theory, J. Symb. Logic 40, 347–382, 1975. 75. F. Richman (ed.), Constructive Mathematics (Proceedings of the Conference at Las Cruces, New Mexico, August 1980), Lecture Notes in Mathematics 873, SpringerVerlag, Heidelberg, 1981. 76. F. Richman, The fundamental theorem of algebra: a constructive development without choice, Paciﬁc J. Math. 196, 213–230, 2000. 77. F. Richman, Adjoints and the image of the unit ball, Proc. Amer. Math. Soc. 129, 1189–1193, 2001. 78. H.L. Royden, Aspects of constructive analysis, in: Errett Bishop: Reﬂections on Him and His Research (M. Rosenblatt, ed.), 57–64, Contemporary Mathematics 39, American Math. Soc., 1985. 79. W. Rudin, Real and Complex Analysis, McGraw-Hill, New York, 1970. 80. E. Schechter, Handbook of Analysis and Its Foundations, Academic Press, San Diego, 1997. 81. P.M. Schuster, What is continuity, constructively?, J.UCS 11(12), 2076–2085, 2005. 82. P.M Schuster, L.S. Vˆıt¸˘ a, and D.S. Bridges, Apartness as a relation between subsets, in: Combinatorics, Computability and Logic (Proceedings of DMTCS01, Constant¸a, Romania, 2–6 July 2001; C.S. Calude, M.J. Dinneen, S. Sburlan, eds.), 203–214, DMTCS Series 17, Springer-Verlag, London, 2001. 83. B. Spitters, Located operators, Math. Logic Quarterly 48(Suppl. 1), 107–122, 2002. 84. W.P. van Stigt, Brouwers Intuitionism, North–Holland, Amsterdam, 1990. 85. G. Stolzenberg, Review of [9], Bull. Amer. Math. Soc. 76, 301–323, 1970.

References

207

86. W. Takahashi, Nonlinear Functional Analysis, Yokohama Publishers, 2000. 87. A. Takayama, Mathematical Economics, The Dryden Press, Hinsdale, IL, 1973. 88. A.S. Troelstra and D. van Dalen, Constructivism in Mathematics: An Introduction (two volumes), North Holland, Amsterdam, 1988. 89. F. Waaldijk, On the foundations of constructive mathematics, Foundations of Science 10(3), 249–324, 2005. 90. H. Weber, Leopold Kronecker, Jahresber. der Deutschen Math. Verein 2, 5–31, 1893. 91. K. Weihrauch, Computable Analysis, Springer-Verlag, Heidelberg, 2000. 92. F. Ye, Towards a constructive theory of unbounded operators, J. Symb. Logic 65, 357–370, 2000. 93. R. Zach, Hilberts Program, in: The Stanford Encyclopedia of Philosophy (E.N. Zalta, ed.), URL http://plato.stanford.edu/archives/fall2003/ entries/hilbert-program/.

The above list contains only a fraction of the publications on constructive mathematics that have appeared in the last forty years, and does not include the sources of all results in our book. The reader should not fall into the trap of believing that an unascribed result was ﬁrst produced by the authors. We mention two websites that may interest the reader: http://www.math.canterbury.ac.nz/php/groups/cm/faq/ http://plato.stanford.edu/entries/mathematics-constructive/ In addition, many of the authors of items in the bibliography have websites that are worth a visit. The primary historical reference on constructive analysis is [9], the review of which [85] is interesting in its own right. Later references for Bishop-style constructivism are [12, 34], the latter of which gives comparisons between BISH, INT, and RUSS. Beeson [6] and Troelstra–van Dalen [88] contain a wealth of information about the logic, philosophy, and practice of constructive mathematics. For some applications of constructive mathematics, see [20, 33, 92]. The deﬁnitive reference for constructive algebra is [72], but [49] should be consulted for more recent work in the ﬁeld. The classic work on intuitionism is [48]. The life and works of Brouwer himself are discussed in [44, 45, 84]. Martin-L¨ ofs early work on constructive mathematics is found in [68], and his theory of types appears in [69]. Among the most recent varieties of computable analysis is that of Weihrauch [91]; the translation of BISH into Weihrauchs framework is described in [5].

Index

absolute value, 30 absolute value, operator, 181 absolutely convergent, 56 absorbing, 65 adjoint, 101 admissible, 43 algebraic diﬀerence, 121 algorithm, 6 apart, 44 approximation, 38, 129 approximation theory, fundamental theorem, 88 approximation, Chebyshev, 106 Archimedes, 57 arithmetic, Heyting, 9 arithmetic, Peano, 9 associated vector, 51 axiom of choice, 17 Baire’s theorem, 189 balanced, 121 ball, 37 Banach space, 49 Banach’s inverse mapping theorem, 196 Banach–Alaoglu theorem, 131 basic open neighbourhood, 127 basic separation theorem, 116 basic set, 18 basis, 81 best approximation, 88 best approximation, at most one, 88 BHK interpretation, 7 bijection, 16

BISH, 6 Bishop, 6 Bishop’s constructive mathematics, 6 Bishop’s lemma, 64 bound, 51 boundary, 107 boundary crossing map, 112 boundary crossing property, 143 bounded linear mapping, 51, 129 Brouwer, 5, 22 Brouwerian counterexample, 11 Brouwerian example, 11 canonical injection, 55 canonical mapping, 84 Cantor’s theorem, 31 Cartesian product, 13, 19 Cauchy, 38 Cauchy completeness, 35 Cauchy sequence, 35, 129 Cauchy–Schwarz inequality, 92 Church–Markov–Turing thesis, 6, 61 CLASS, 5 closed, 37 closed range theorem, 184 closed set, 127 closest point, 88 closure, 37, 127 cluster point, 37 commuting operators, 171 compact, 44 compact image, 60 compact linear mapping, 62, 102, 194

210

Index

compactly generated, 104, 194 comparison test, 56 complement, 14 complement, logical, 12 complete, 129 complete metric space, 38 complex number, 36 complex plane, 36 composite/composition, 16 cone, 113 cone, generated, 113 connectives, 7 constructive reverse mathematics, 199 continuous, 39, 61, 128 convergence, 35 convergent, 38, 56 converges, 129 converges strongly, 170 converges weakly, 182 converges, series, 95 convex, 65 convex combination, 87 convex hull, 87 convex mapping, 125 coordinate, 82, 96 coordinate functionals, 82 cotransitive, 28 countable, 17 countable choice, 18 countably inﬁnite, 17 countably many, for all but, 43 counterexample, Brouwerian, 11 de Morgan’s rule, 76 decidable, 8 Dedekind completeness, 32 denial inequality, 13 dense, 37, 127 dense sequence, 38 dependent choice, 18 detachable, 14 diameter, 40 diﬀerence, 33 dimension, 81 Dini’s theorem, 176 discrete, 13 distance, 40 distinct, 13 distinct real numbers, 26

division, 34 domain, 15 dual, 130 empty subset, 12 epimorphism, 16 equal functions, 16 equal real numbers, 26 equal sets, 13 equality, 12 equivalent norms, 85 Euclid, 1 Euclidean space, 49 eventually, 71 example, Brouwerian, 11 excluded middle, 5, 18 excluded sequence, 43 excluded third, 5 existence, constructive, 2 existence, idealistic, 2 extensional, 14 family, 15 fan theorem, 60, 106 ﬁnite, 17 ﬁnite possibility, 21 ﬁnite-dimensional, 81 ﬁnitely enumerable, 16 ﬁnitely many, 17 ﬂagging alternatives, 64 Fr´echet diﬀerentiable, 148 function, 14 Gˆ ateaux diﬀerentiable, 126 generating set, 193 Goldbach conjecture, 3 Gram–Schmidt process, 96 graph, 16 greater than, 27 greater than or equal to, 27 greatest lower bound, 33 Hahn–Banach theorem, 123 half-space, 116 Hausdorﬀ metric, 46 Heine–Borel–Lebesgue theorem, 61 Hellinger–Toeplitz theorem, 106, 169 Hermitian, 102 Heyting, 7, 22

Index Heyting arithmetic, 9 Hilbert, 5 Hilbert space, 93 hyperplane, 51 idempotent, 94 identity operator, 95 image of a subset, 15 image of an element, 14 imaginary part, 36 independent, 6 index set, 15 inequality relation, 13 inequality, compatible, 47 inequality, denial, 13 inequality, discrete, 13 inequality, metric space, 37 inequality, tight, 13 inﬁmum, 40 inﬁnite-dimensional, 98 inﬁnitely often, 71 inhabited, 12 initial projection, 198 injective, 16 INT, 5 integers, 25 interior, 37 intermediate value theorem, 2 internal choice, 21 intersect sharply, 77 intersection, 19 interval-halving, 3 intervals, in a vector space, 108 inverse function, 16 inverse image, 15 Ishihara’s tricks, 71 jointed, 101 kernel, 52 Kolmogorov, 22 Kronecker, 5 lambda technique, 64 least upper bound, 32 least-upper-bound principle, 32 LEM, 5 LEM-Cauchy, 77 less than, 27

less than or equal to, 27 limit, 35, 129 linear functional, 50 linear functional, bounded, 51 linear map, 50 linear space, 48 linear space, seminormed, 48 linearly independent, 81 LLPO, 9 locally compact, 46 locally convex space, 127 locally convex structure, 127 locally convex topology, 127 locally nonzero, 21 locally totally bounded, 46 located, 40, 130 logic, classical, 2 logic, intuitionistic, 2, 7 logical complement, 12 lower bound, 33 lower order located, 33 LPO, 8 map, 14 mapping, 14 Markov’s principle, 10 material implication, 22 mathematics, classical, 5 mathematics, intuitionistic, 5 maximum, 30 metric, 37 metric complement, 41 metric space, 37 metric, associated, 48 minimum, 30 modulus-argument decomposition, 64 monotone convergence, 35 MP, 10 mutual distance, 121 negative, 30 nonnegative real line, 46 nontrivial, 48 nonzero, 52 nonzero vector, 48 norm, 48 norm, Euclidean, 49 normable, 52 normed, 52

211

212

Index

normed space, 48 omniscience principles, 8 one-one, 16 onto, 16 open, 37, 127 open mapping theorem, 192 operator, 50, 100, 166 operator norm, 130 operator, bounded, 51 ordered, 20 ordered pair, 20 orthogonal complement, 95 orthogonal subsets, 95 orthogonal vector, 95 orthonormal basis, 96 orthonormal family, 96 parallelogram identity, 92 partial function, 14 partial function, total, 15 partial isometry, 198 partial sum, 56 Peano arithmetic, 9 pointwise, 15 polar decomposition, 199 polarisation identity, 175 positive operator, 175 power set, 19 product, 34 product metric, 37 product metric space, 37 product norm, 50 product normed space, 50 projection, 94 proximinal, 88 pseudobounded, 169 Pythagoras’s theorem, 95 quantiﬁers, 7 quasiproximinal, 88 quotient norm, 55 quotient space, 55 range, 15 rational numbers, 25 real line, 27 real number, 25 real part, 36

recursive constructive mathematics, 6 regular, 60 represented, 99 Riesz representation theorem, 99 Riesz’s lemma, 86 RUSS, 6 selfadjoint, 102 seminorm, 48 seminorm, deﬁning, 127 separable, 38, 129 separation theorem, 121 sequence, 15 sequentially continuous, 21, 39 sequentially discontinuous, 77 sequentially nondiscontinuous, 72 sequentially open, 181 series, 56 set, 12 set, basic, 18 smooth, 148 span, 81 spanned, 81 spectral theorem, 200 square root, 64 square root, operator, 179 square summable, 93 strong injection, 45 strongly continuous, 44 strongly extensional, 14 strongly injective, 45 strongly unique, 14 subset, 12 subspace, 50 suﬃciently large, 71 sum, 33 sup norm, 49 supremum, 32, 40 termwise, 15 tight, 13 topology, 37 totally bounded, 38, 129 trichotomy, 29 Type-2 Eﬀectivity, 22 ultraﬁlter principle, 146 ultraweakly continuous, 145 unequal, 13

Index unequal real numbers, 26 uniform boundedness theorem, 165 uniform Cauchyness, 49 uniform convergence, 49 uniformly almost located, 76 uniformly continuous, 40, 128 uniformly convex, 54, 126, 151 uniformly Fr´echet diﬀerentiable, 148 union, 19 unit ball, 48, 127 unit kernel, 132 unit vector, 48 upper bound, 32

upper order located, 32 value, 14 weak star topology, 130 weak-sequentially open, 182 weakly compact, 198 weakly discriminating, 77 weakly totally bounded, 197 Weihrauch, 6 well-behaved, 76 WLEM, 58 Zorn’s lemma, 116, 122

213

Universitext

(continued from p. ii)

Hurwitz/Kritikos: Lectures on Number Theory Jennings: Modern Geometry with Applications Jones/Morris/Pearson: Abstract Algebra and Famous Impossibilities Kac/Cheung: Quantum Calculus Kannan/Krueger: Advanced Analysis Kelly/Matthews: The Non-Euclidean Hyperbolic Plane Kostrikin: Introduction to Algebra Kuo: Introduction to Stochastic Integration Kurzweil/Stellmacher: The Theory of Finite Groups: An Introduction Lang: Introduction to Differentiable Manifolds Lorenz: Algebra: Volume I: Fields and Galois Theory Luecking/Rubel: Complex Analysis: A Functional Analysis Approach MacLane/Moerdijk: Sheaves in Geometry and Logic Marcus: Number Fields Martinez: An Introduction to Semiclassical and Microlocal Analysis Matsuki: Introduction to the Mori Program McCarthy: Introduction to Arithmetical Functions McCrimmon: A Taste of Jordan Algebras Meyer: Essential Mathematics for Applied Fields Mines/Richman/Ruitenburg: A Course in Constructive Algebra Moise: Introductory Problems Course in Analysis and Topology Morris: Introduction to Game Theory Poizat: A Course In Model Theory: An Introduction to Contemporary Mathematical Logic Polster: A Geometrical Picture Book Porter/Woods: Extensions and Absolutes of Hausdorff Spaces Procesi: Lie Groups Radjavi/Rosenthal: Simultaneous Triangularization Ramsay/Richtmyer: Introduction to Hyperbolic Geometry Rautenberg: A Concise Introduction to Mathematical Logic, 2nd ed. Reisel: Elementary Theory of Metric Spaces Ribenboim: Classical Theory of Algebraic Numbers Rickart: Natural Function Algebras Rotman: Galois Theory Rubel/Colliander: Entire and Meromorphic Functions Runde: A Taste of Topology Sagan: Space-Filling Curves Samelson: Notes on Lie Algebras Schiff: Normal Families Shapiro: Composition Operators and Classical Function Theory Simonnet: Measures and Probability Smith: Power Series From a Computational Point of View Smith/Kahanpää/Kekäläinen/Traves: An Invitation to Algebraic Geometry Smorynski: Self-Reference and Modal Logic Stillwell: Geometry of Surfaces Stroock: An Introduction to the Theory of Large Deviations Sunder: An Invitation to von Neumann Algebras Tondeur: Foliations on Riemannian Manifolds Toth: Finite Möbius Groups, Minimal Immersions of Spheres, and Moduli Van Brunt: The Calculus of Variations Weintraub: Galois Theory

Wong: Weyl Transforms Zhang: Matrix Theory: Basic Results and Techniques Zong: Sphere Packings Zong: Strange Phenomena in Convex and Discrete Geometry

S. Axler K.A. Ribet

Universitext Editors (North America): S. Axler and K.A. Ribet Aguilar/Gitler/Prieto: Algebraic Topology from a Homotopical Viewpoint Aksoy/Khamsi: Nonstandard Methods in Fixed Point Theory Andersson: Topics in Complex Analysis Aupetit: A Primer on Spectral Theory Bachman/Narici/Beckenstein: Fourier and Wavelet Analysis Badescu: Algebraic Surfaces Balakrishnan/Ranganathan: A Textbook of Graph Theory Balser: Formal Power Series and Linear Systems of Meromorphic Ordinary Differential Equations Bapat: Linear Algebra and Linear Models (2nd ed.) Berberian: Fundamentals of Real Analysis Blyth: Lattices and Ordered Algebraic Structures Boltyanskii/Efremovich: Intuitive Combinatorial Topology. (Shenitzer, trans.) Booss/Bleecker: Topology and Analysis Borkar: Probability Theory: An Advanced Course Böttcher/Silbermann: Introduction to Large Truncated Toeplitz Matrices Bridges/Vît¸a˘: Techniques of Constructive Analysis Carleson/Gamelin: Complex Dynamics Cecil: Lie Sphere Geometry: With Applications to Submanifolds Chae: Lebesgue Integration (2nd ed.) Charlap: Bieberbach Groups and Flat Manifolds Chern: Complex Manifolds Without Potential Theory Cohn: A Classical Invitation to Algebraic Numbers and Class Fields Curtis: Abstract Linear Algebra Curtis: Matrix Groups Debarre: Higher-Dimensional Algebraic Geometry Deitmar: A First Course in Harmonic Analysis (2nd ed.) DiBenedetto: Degenerate Parabolic Equations Dimca: Singularities and Topology of Hypersurfaces Edwards: A Formal Background to Mathematics I a/b Edwards: A Formal Background to Mathematics II a/b Engel/Nagel: A Short Course on Operator Semigroups Farenick: Algebras of Linear Transformations Foulds: Graph Theory Applications Friedman: Algebraic Surfaces and Holomorphic Vector Bundles Fuhrmann: A Polynomial Approach to Linear Algebra Gardiner: A First Course in Group Theory Gårding/Tambour: Algebra for Computer Science Goldblatt: Orthogonality and Spacetime Geometry Gustafson/Rao: Numerical Range: The Field of Values of Linear Operators and Matrices Hahn: Quadratic Algebras, Clifford Algebras, and Arithmetic Witt Groups Heinonen: Lectures on Analysis on Metric Spaces Holmgren: A First Course in Discrete Dynamical Systems Howe/Tan: Non-Abelian Harmonic Analysis: Applications of SL(2, R) Howes: Modern Analysis and Topology Hsieh/Sibuya: Basic Theory of Ordinary Differential Equations Humi/Miller: Second Course in Ordinary Differential Equations (continued after index)

Douglas S. Bridges and Luminit¸a Simona Vît¸a˘

Techniques of Constructive Analysis

Douglas S. Bridges Department of Mathematics/Statistics University of Canterbury Christchurch, New Zealand [email protected]

Luminit¸a Simona Vît¸a˘ Department of Mathematics/Statistics University of Canterbury Christchurch, New Zealand [email protected]

Editorial Board (North America): S. Axler Mathematics Department San Francisco State University San Francisco, CA 94132 USA [email protected]

K.A. Ribet Mathematics Department University of California at Berkeley Berkeley, CA 94720-3840 USA [email protected]

Mathematics Subject Classification (2000): 03F60, 26E40, 46S30, 47S30, 03F55, 03F65, 68Q99 Library of Congress Control Number: 2006926441 ISBN-10: 0-387-33646-X ISBN-13: 978-0387-33646-6

e-ISBN-10: 0-387-38147-3 e-ISBN-13: 978-0387-38147-3

Printed on acid-free paper. © 2006 Springer Science+Business Media, LLC All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. 9 8 7 6 5 4 3 2 1 springer.com

Ideal pierdut ˆın noaptea unei lumi ce nu mai este, Lume ce gˆandea ˆın basme ¸si vorbea ˆın poezii, O! te v˘ ad, te-aud, te cuget, tˆ an˘ ar˘ a ¸si dulce veste Dintr-un cer cu alte stele, cu-alte raiuri, cu alt¸i zei. —Mihai Eminescu, “Venere ¸si Madon˘ a” Oh, ideal lost in night-mists of a vanished universe: People who would think in legends—all a world who spoke in verse; I can see and think and hear you—youthful scout which gently nods From a sky with diﬀerent starlights, other Edens, other gods. —Mihai Eminescu, “Venus and Madonna” (translated by Andrei Banta¸s)

This image is a courtesy of The Times of London. Printed in the February 3, 2004 issue.

Preface

Rosencrantz: Shouldn’t we be doing something... constructive? Guildenstern: What did you have in mind? —Tom Stoppard, Rosencratz and Guildenstern Are Dead

We have written this book in order to provide an introduction to constructive analysis, emphasising techniques and results that have been obtained in the last twenty years. The intended readership comprises senior undergraduates, postgraduates, and professional researchers in mathematics and theoretical computer science. We hope that our work will help spread the message that doing mathematics constructively is interesting (it can even be fun!) and challenging, and produces new, deep computational information. An appreciation of the distinction between constructive and nonconstructive has become more widespread in this era of computers. Nevertheless, there are few books devoted to the development of mathematics in a rigorously constructive/computable fashion, although there are some, primarily concentrating on logic and foundations, in which the odd chapter deals with constructive mathematics proper as distinct from its underlying logic or set theory. It is now almost forty years since the publication of Errett Bishop’s seminal monograph Foundations of Constructive Analysis [9], which in our view is one of the most remarkable intellectual documents of the twentieth century, and more than twenty since the appearance of its outgrowth [12]. In the intervening years there has been considerable activity in constructive analysis, algebra, and topology; in related foundational areas such as type theory [69]; and in the relation between constructive mathematics and computer science (for example, program extraction from proofs [42, 70, 51]). Believing that a new introduction to the mathematical, as distinct from the foundational, side of the subject is overdue, we embarked upon this monograph. Our book is intended not to replace, but to supplement, Bishop’s original classic [9] and the later volume [12] based thereon. Both of those two monographs cover

viii

Preface

aspects of analysis, such as Haar measure and commutative Banach algebras, that we do not mention. We cover some topics that are found in [9] and [12] (it would be almost inconceivable to produce a book like ours, dealing with constructive mathematics for nonexperts, without proving, for example, basic results about locatedness and total boundedness); but we have tried to provide improved proofs whenever possible. However, much of the material we present was simply not around at the time of writing of [9] or [12]. Instead of systematically developing analysis, beginning with the real line and continuing through metric, normed, and Hilbert spaces to its higher reaches, we have chosen to write the chapters around certain themes or techniques (hence our title). For example, Chapter 3 is devoted to the λ-technique, which, since its ﬁrst use in the proof of Lemma 7 on page 177 of [9], has become a surprisingly powerful tool with applications in many areas of constructive analysis. A major inﬂuence in the application of the λ-technique was Ishihara’s remarkable paper [60], which showed that a subtle use of the technique could enable us to prove disjunctions whose proof, although trivial with classical logic, appears at ﬁrst sight to be constructively out of the question. This paper opened up many new pathways in constructive analysis. Chapter 1 introduces constructive mathematics and lays the foundations for the later chapters. In Chapter 2 we ﬁrst present a new construction of the real numbers, motivated by ideas in [2]. After deriving standard properties such as the completeness of R, we introduce metric spaces, with the major theme of locatedness, and normed linear spaces. When we discuss metric, normed, and Hilbert spaces, we assume some familiarity with the standard classical deﬁnitions of those concepts and with those elementary classical properties that pass over unchanged to the constructive setting. Chapter 3 we have already referred to. The main theme of Chapter 4 is ﬁnitedimensionality, but the chapter concludes with an introduction to Hilbert spaces. Chapter 5 deals with convexity in normed spaces. Starting with some elementary convex geometry in Rn , the chapter goes on to handle separation and Hahn–Banach theorems, locally convex spaces, and duality. Following Bishop, we describe those linear functionals that are weak∗ -uniformly continuous on the unit ball of the dual space. We then give a new application of the technique used to prove that result, thereby characterising certain continuous linear functionals on the space of bounded operators on a Hilbert space. In Chapter 6 we derive a range of results associated with the theme of locatedness and with the λ-technique introduced in Chapter 3. We pay particular attention to necessary and suﬃcient conditions for convex subsets of a normed space to be located, and to connections between properties of an operator on a Hilbert space and those of its adjoint—when that adjoint exists: it may not always do so constructively. The ﬁnal section of the book deals with a relatively recent version of Baire’s theorem and its applications, and culminates in constructive versions of three of the big guns in functional analysis: the open mapping, inverse mapping, and closed graph theorems.

Preface

ix

Which parts of the book deal with new material, compared with what appeared in [12]? We have already mentioned the new construction of the real numbers, in Chapter 2. Notable novelties in the later chapters include all but one result in Chapter 3 on the λ-technique; the section on convexity, Ishihara’s results on exact Hahn–Banach extensions, and our characterisation theorem for certain continuous linear functionals, all in Chapter 5; and virtually all of Chapter 6. Throughout the book there are what we hope will be seen as improvements and simpliﬁcations of proofs of many results that were given in [9] or [12]. What do we mean by “constructive analysis” in the title of this book? We do not mean analysis carried out with the usual “classical” logic within a framework, such as recursive function theory, designed to capture the concept of computability. In our view, such a notion of constructive has at least two drawbacks. First, by working within, say, the recursive setting, it can make the mathematics look less like normal mathematics and much harder to read. Secondly, the recursive constraint removes the possibility of other interpretations of the mathematics, such as Brouwer’s intuitionistic one [48]. Our approach, on the other hand, has neither of these features: the mathematics looks and reads just like the mathematics one is used to from undergraduate days, and all our proofs and results are valid in several models. They are valid in the recursive model, in intuitionistic mathematics, and, we believe, in any of the models for “computable mathematics” (including Weihrauch’s Type Two Eﬀectivity Theory [91], within which Andrej Bauer has recently found a realisability interpretation of constructive mathematics within Weihrauch’s theory [5]). They are also valid proofs in standard mathematics with classical logic. For example, our proof of the Hahn–Banach theorem (Theorem 5.3.3) is, as it stands, a valid algorithmic proof of the classical Hahn–Banach theorem. Moreover—and this is one advantage of a constructive proof in general—our proof embodies an algorithm for the construction of the functional whose existence is stated in the theorem. This algorithm can be extracted from the proof, and, as an undeserved bonus, the proof itself demonstrates that the algorithm is correct or, in computer science parlance, “meets its speciﬁcations”.1 So how do we achieve all this? Simply by changing the logic with which we do our mathematics! Instead of using classical logic, we systematically use intuitionistic logic, which was abstracted by Heyting [52] from the practice of Brouwer’s intuitionistic mathematics. The remarkable fact is that every proof carried out with intuitionistic logic is fully constructive/algorithmic. (Is this the “secret on the point of being blabbed” that appears in the epigraph to Bishop’s book?) Unfortunately, too few mathematicians outside the mathematical logic community are aware of this serendipity and dismiss both intuitionistic logic and constructive mathematics as at best a marginal curiosity. This contrasts sharply with the theoretical computer science community, in which there is considerable knowledge of, and interest in, the computational power of intuitionistic logic. 1 We do not carry out program-extraction from proofs in our book. For more on this topic see [42, 51, 70].

x

Preface

Reading constructive mathematics demands careful interpretation. A theorem in this book might look like a familiar one from classical analysis, but with more complicated hypotheses and proof. However, the statement of the theorem will be phrased so that the explicit algorithmic interpretation is left to the reader; and the additional hypotheses will be necessary for a constructive proof, which will contain algorithmic information that is excluded from the classical proof by the latter’s use of principles outside intuitionistic logic. Consider, for example, the following statement: (*) Let C be an open convex subset of a normed space X, let ξ ∈ C, and let z ∈ X be bounded away from C. Then the boundary of C intersects the segment [ξ, z] joining ξ and z. This is trivial to prove classically; but to ﬁnd/construct the (necessarily unique) point in which the boundary of C intersects [ξ, z] is a totally diﬀerent matter. The constructive theorem (Proposition 5.1.5 below) requires us to postulate that the union of C and its metric complement −C (the set of points bounded away from C) be dense in X, and that X itself be a complete normed space. The constructive proof, though elementary, requires some careful geometrical estimation that would be supererogatory in the natural classical proof by contradiction. The beneﬁt of that estimation and of the use of intuitionistic logic is that we could extract from the constructive proof an implementable algorithm for ﬁnding the point where the segment crosses the boundary. In turn, this would enable us to produce an algorithm for constructing separating hyperplanes and Hahn–Banach extensions of linear functionals, under appropriate hypotheses. We could have made the algorithmic interpretation of the constructive version of (*) explicit by stating the proposition in this way: There is a “boundary crossing algorithm” that, applied to the data consisting of (i) an open convex set C in a Banach space X such that C ∪−C is dense in X, (ii) a point ξ of C, and (iii) a point z of −C, constructs the point where the boundary of C intersects the segment [ξ, z] . Even this is not really explicit enough. A full description of the data to which the boundary crossing algorithm applies would require explicit information about the algorithms for such things as these: membership of C; the convergence of Cauchy sequences in X; the computation, for given x in X and ε > 0, of a point y of C ∪−C such that x − y < ε (and even the decision between the cases “y ∈ C” and “y ∈ −C”); and so on. Such explicit description of algorithmic hypotheses would become an ever greater burden on writer and reader alike as the book probed deeper and deeper into abstract analysis. It is a matter of sound sense, even sanity, to unburden ourselves from the outset, relying on the reader’s native wit in the interpretation of the statements of our constructive lemmas, propositions, and theorems. We should make it clear that we are not advocating the exclusive use of intuitionistic logic in mathematics. That logic is, we believe, the natural and right

Preface

xi

one to use when dealing with the constructive content of mathematics. To abandon classical logic in those ﬁelds (such as the higher reaches of set theory) where constructivity is of little or no signiﬁcance makes no sense whatsoever. Nevertheless, it is remarkable how much mathematics actually has what Bishop called “a deep underpinning of constructive truth”.

Christchurch, New Zealand January 2006

Douglas Bridges Luminit¸a Simona Vˆıt¸a ˘

Acknowledgments

It is never easy to apportion thanks properly among the many who have contributed to this book either directly or by their support and encouragement at various stages of our professional lives. We do, however, have special thanks for the following people: Cris Calude, who was responsible for bringing Vˆıt¸˘a to New Zealand to begin what has proved a very fruitful research partnership with Bridges, and who has been tireless in his encouragement of our work over many years. Hajime Ishihara, who, in conjunction with one or both of the authors, was largely responsible for much of the work in Chapter 6 and whose inﬂuence can be seen in several other places in the book. (Hajime’s contributions to constructive functional analysis have been remarkable and deserve to be recognised more widely.) In addition, he has hosted us many times at the Japan Advanced Institute of Science & Technology, each of our visits there being both memorable and highly productive. Peter Schuster, not only for his contributions to constructive mathematics since he joined our community ten years ago, but also for acting as organiser, fund-raiser, and host on our many research visits to Munich. His considerable eﬀorts in securing a DAAD Gastprofessorship for Bridges in the Mathematisches Institut der LudwigMaximilians-Universit¨ at (LMU), M¨ unchen, in 2003 provided us with the time and environment in which we could break the back of the writing of this book. Our early drafts of Chapters 1–5 formed the basis of graduate lectures by Bridges at LMU in 2003. We thank the students in that course for patiently receiving that material and for suggesting corrections and improvements to our presentation of it. We are also grateful to the DAAD for supporting Bridges as a Gastprofessor at LMU for that year; Otto Forster and Helmut Schwichtenberg, our hosts at LMU in 2003 and on several other occasions;

xiv

Acknowledgments

the New Zealand Foundation of Research, Science and Technology for Vˆıt¸˘a’s postdoctoral fellowship from 2002 to 2005 and for supporting her extended visit to LMU to enable us to work together on the book; our departments at Canterbury and Galat¸i; Josef Berger, Hannes Diener, Maarten Jordans, and Robin Havea, who have kindly assisted us with proofreading various chapters. We say a warm “thank you” to our friends Imola and Attila Zsigmond, and Helmut and Eva Pellinger, who were wonderful hosts during our time in Munich. Finally, we want to thank our families for their continuing love and support.

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vii

Introduction to Constructive Mathematics . . . . . . . . . . . . . . . . . . . . .

1

1.1 What Is Constructive Mathematics? . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.2 A Very Brief History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5

1.3 Intuitionistic Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.4 Informal Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

Techniques of Elementary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

2.1 The Real Number Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

2.2 Metric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

37

2.3 Normed Linear Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

47

The λ-Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

3.1 Introduction to the Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

3.2 Ishihara’s Tricks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

71

Finite-Dimensional and Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . .

81

4.1 Finite-Dimensional Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

4.2 Best Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

88

4.3 Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

1

2

3

4

5

Linearity and Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

xvi

Contents

5.1 Crossing Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 5.2 Separation Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 5.3 The Hahn–Banach Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 5.4 Locally Convex Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 6

Operators and Locatedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 6.1 Smooth and Uniformly Smooth Normed Spaces . . . . . . . . . . . . . . . . . 147 6.2 Locatedness of Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 6.3 Adjoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 6.4 Functions of Selfadjoint Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 6.5 Locating the Kernel and the Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 6.6 Baire’s Theorem, with Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

1 Introduction to Constructive Mathematics

My task which I am trying to achieve is, by the power of the written word, to make you hear, to make you feel—it is, before all, to make you see. That, and no more, and it is everything. —Joseph Conrad, The Nigger of the ‘Narcissus’ In this chapter we ﬁrst sketch the history and philosophy that motivated the early workers in the ﬁeld of constructive mathematics. We then describe informal intuitionistic logic and discuss a number of elementary classical theorems that do not carry over to the constructive setting. Finally, we introduce an informal constructive theory of sets and functions. All this will prepare us for the presentation of the constructive theory of the real line R in Chapter 2, and for the more abstract analysis that will be described in later chapters.

1.1 What Is Constructive Mathematics? Proposition 20 in Book IX of the thirteen volumes of Euclid’s Elements states that Prime numbers are more than any assigned multitude of prime numbers —in current terms, there are inﬁnitely many primes. The modernised version of Euclid’s proof is often presented as follows. Suppose that there are only ﬁnitely many primes, say p1 , . . . , pn , and consider the integer p = p1 × p2 × · · · × pn + 1. Being greater than 2, p has prime factors (it may even be prime itself). Since the numbers pk are not divisors of p, each prime factor of p is distinct from each pk . This is absurd, since {p1 , . . . , pn } is supposed to be the set of all primes. From this contradiction we conclude that the set of primes is inﬁnite.

2

1 Introduction to Constructive Mathematics

Although at one level there appears to be nothing untoward about this proof, it can be criticised on two counts. First, it uses a totally unnecessary contradiction argument. If you look carefully, you will see that the proof actually embodies an algorithm that, applied to any ﬁnite set {p1 , . . . , pn } of primes, enables you to compute a prime that is distinct from each element of that set. In other words, the use of a contradiction argument in the preceding paragraph has obscured the computational content of Euclid’s proof. The second criticism of the proof is a little more subtle, and deals with the notion of “inﬁnitely many”. The proof is based on the negative idea that a set is inﬁnite if and only if it is contradictory that it be ﬁnite. But an algorithmic recasting of Euclid’s proof, as suggested in the preceding paragraph, shows that the set S of primes is inﬁnite in a more positive, productive sense: namely, if we start with a ﬁnite subset F of S, then we can compute an element of S that is distinct from each element of F. From a traditional standpoint, the distinctions between the contradiction proof of Euclid’s theorem and the algorithmic one, and between the negative and positive notions of “inﬁnite”, are obscured if not invisible. For example, the two notions of “inﬁnite” are equivalent if we use traditional logic—or classical logic, as it is normally called—so the distinction has to be perceived at an aesthetic level rather than a mathematical one. The same applies, more generally, to a proof by contradiction of the existence of an object x with the property P (x). In such a proof one supposes that P (x) is false for all applicable objects x, deduces a contradiction, and then concludes that P (x) must, after all, hold for some x, even though the proof doesn’t tell us which x actually has the desired property. Classical logic draws no distinction between the “idealistic existence” demonstrated by such a proof and the “constructive existence” based on an algorithm that constructs x and shows that P (x) holds. In order to reveal such distinctions at a mathematical, rather than an aesthetic, level we shall adopt the radical expedient of changing our logic: throughout this book, we shall work with intuitionistic logic, an abstraction of the informal logic used in algorithmic thinking. How much analysis, as normally presented, is really nonconstructive —that is, essentially dependent on proofs by contradiction or other nonalgorithmic procedures? Consider, for example, the classical intermediate value theorem: If f : [a, b] −→ R is a continuous mapping such that f (a) < 0 and f (b) > 0, then there exists c ∈ (a, b) such that f (c) = 0. (Note that [a, b] and (a, b) respectively denote the closed and open intervals with endpoints a and b. We shall use standard notations, like (a, b] for the half-open interval, without further comment.) It might be thought that the common elementary proofs of the intermediate value theorem are constructive, enabling one to produce a zero c of the function f. For example, one proof uses interval-halving in the following way. Without loss of

1.1 What Is Constructive Mathematics?

3

generality, take a = 0 and b = 1. Consider f (1/2): if it is 0, then we take c = 0 and stop the process; if f (1/2) > 0, then f satisﬁes the hypotheses of the theorem with a = 0 and b = 1/2; if f (1/2) < 0, then f satisﬁes the hypotheses with a = 1/2 and b = 1. In each of the last two cases, we proceed with the interval-halving. This process either stops after a ﬁnite number of iterations and produces the required zero of f, or else it goes on ad inﬁnitum to produce a descending sequence of compact intervals whose unique point of intersection is the required zero. Isn’t this a fully algorithmic proof? Suppose we try to implement the algorithm embodied in this proof on a computer that works with 50-bit precision. What happens if we apply it to the cubic function f deﬁned on [0, 1] by 2 1 3 x− − 2−51 ? f (x) = x − 4 2 Here, 1 3 − 2−51 < 0, f (1) = − 2−51 > 0, 16 16 so we are well set to carry out the ﬁrst step of the interval-halving algorithm. Since our computer’s ﬂoating-point representation of f (1/2) is 0 (we have the phenomenon of underﬂow, in which the computer sets the small but nonzero number −2−51 equal to 0), the algorithm stops by outputting c = 1/2 as the place where f has a zero. But in this case the only zero of f in [0, 1] lies between 3/4 and 1, more than one quarter of the entire interval away from the output value 1/2. f (0) = −

Now, one could object that this example is misleading, in that the problem arises from the level of precision in the computer rather than any intrinsic failing in the algorithm itself. To deal with this point, for each positive integer n let G(n) signify that 2n + 2 is a sum of two primes. Construct a binary sequence (an )n1 such that for each n, an = 0 if and only if either G(k) for all k n or else there exists k < n such that ¬G(k). Deﬁne a =

∞ n=1

an 2−n , and note that a = 0 if and only if the Goldbach conjecture,

Every even integer greater than 2 is a sum of two primes, holds. Using classical logic, apply the classical interval-halving algorithm to the cubic function f deﬁned on [0, 1] by 2 1 3 x− − a. f (x) = x − 4 2 As long as the status of the Goldbach conjecture remains undecided (which it has done since the conjecture ﬁrst appeared in 1742), no matter what ﬁnite precision

4

1 Introduction to Constructive Mathematics

our computer has, the algorithm will output 1/2 as a zero of f ; but if the Goldbach conjecture is false, then f (1/2) = −a < 0 and the ﬁrst zero of f in [0, 1] occurs between 3/4 and 1. In fact, the classical algorithm will give the correct output if and only if the Goldbach conjecture is true. However, as the reader may verify, a constructive approximate interval-halving argument, such as that expected in the solution of Exercise 11, does not give the possibly false value 1/2 for a zero of f. It produces a value that approximates the zero of f lying between 3/4 and 1, as accurately as the precision of the computer permits. In this example, the classically (but not constructively) deﬁned function taking the parameter a to the smallest root r(a) of f is discontinuous at a = 0. The classical algorithm correctly outputs r(a) = 1/2 in the case a = 0; but if a > 0, then, by outputting the value 1/2, the algorithm has failed to spot that the value of r(a) jumps from 1/2 to more than 3/4 as the parameter a increases from 0. There is a general principle that constructive proofs will involve continuity in parameters. Thus we cannot expect to prove constructively that for each a the above cubic function f has a smallest zero. The problem with the classical interval-halving algorithm is that the ﬁnite precision of the computer prevents it from making correct comparisons between two very close, but distinct, real numbers. If we are to develop mathematics in a computational manner, we have to ensure that such comparisons are barred. This barring can be done in at least two ways. One way is to use classical logic and to preclude nonalgorithmic “decisions” (such as whether two given numbers are equal) by developing the mathematics using a standard programming language or a more abstract algorithmic framework like that of recursive function theory. Another way is to change from classical to intuitionistic logic. The advantages of this second way are, ﬁrst, that nonalgorithmic “decisions” are automatically barred by the logic, and, second, that the resulting mathematics looks like the mathematics we are used to from school and university, without any special logical notation such as is used in, for example, recursive function theory. In this book we explore mathematics with intuitionistic logic. We work throughout with notions, like that of “inﬁnitely many” discussed earlier, that have positive computational meaning; and we present only algorithmic proofs—ones that show how we can, at least in principle, construct the objects whose existence is asserted in the statement of a theorem. We hope to convince the reader that, contrary to a widely held belief, intuitionistic logic suﬃces for the development of deep, interesting mathematics and often opens up new vistas that are hidden by classical logic. In other words, we want to justify our belief in the power of positive (constructive) thinking in mathematics.

1.2 A Very Brief History

5

1.2 A Very Brief History Although luminaries such as Leopold Kronecker had advocated a constructive approach to mathematics in the nineteenth century, the story of modern constructivism really begins with the publication, in 1907, of the doctoral thesis “On the Foundations of Mathematics” [41], in which the Dutch mathematician L.E.J. Brouwer introduced his intuitionistic mathematics (INT) as an alternative to traditional classical mathematics (CLASS). According to Brouwer, mathematical objects are free creations of the human mind, independent of both logic and language, and a mathematical object comes into existence precisely when it is constructed. Such a belief naturally leads to a rejection of existence proofs by contradiction, and a consequent scepticism about the meaning of many of the theorems of CLASS. Not surprisingly, Brouwer’s views met with at best indiﬀerence, and at worst hostility, from the large majority of his peers, for whom the elimination of nonconstructive arguments, with all their apparent power and fruitfulness, was too great a price to pay for a clariﬁcation of the meaning of mathematics. If we adhere to the principle that “existence” should always be interpreted constructively, then we are forced to dispense with the unrestricted use of the logical law of excluded middle (or excluded third ), P or (not P ), which we shall abbreviate to LEM. Recognising this consequence of his philosophical views, Brouwer went as far as to claim, The belief in the universal validity of the principle of the excluded third in mathematics is considered by the intuitionists as a phenomenon of the history of civilization of the same kind as the former belief in the rationality of π , or in the rotation of the ﬁrmament about the earth [44].

Subsequently, he introduced into INT some principles that led to results apparently contradicting aspects of classical mathematics. For example, Brouwer was able to prove that any real-valued function on [0, 1] is uniformly continuous. But to regard Brouwer’s mathematics as inconsistent with its classical counterpart is a serious oversimpliﬁcation of the situation, since the two types of mathematics are in many respects incomparable. Nevertheless, there was, and remains, a commonly held belief that too much mathematics has to be given up in order to accommodate Brouwer’s ideas. For example, Hilbert expressed his disagreement with Brouwer in words both forceful and memorable: Forbidding a mathematician to make use of the principle of excluded middle is like forbidding an astronomer his telescope or a boxer the use of his ﬁsts [54].

Despite continuing opposition, intuitionism survived and new constructive approaches to mathematics arose. In 1948–1949 in the former Soviet Union, A.A.

6

1 Introduction to Constructive Mathematics

Markov initiated a programme of recursive constructive mathematics (RUSS)— mathematics using intuitionistic logic and based on the Church–Markov–Turing thesis that all computable partial functions from the set N of natural numbers to itself are recursive. This approach led to a number of technical successes [66, 67]. RUSS does not use any of Brouwer’s nonlogical intuitionistic principles; indeed, it could not, since it produces results that are false if interpreted directly within INT. For example, in RUSS there exists a continuous real-valued map on [0, 1] that is not uniformly continuous; more dramatically, there exists a uniformly continuous map f from [0, 1] onto (0, 1] that has inﬁmum equal to 0. Once again, one should not overreact to the apparent conﬂict with classical mathematics: the last of these results should really be interpreted as saying that there exists a recursively uniformly continuous recursive function f from the closed interval [0, 1] of the recursive real line onto the recursive interval (0, 1] that has inﬁmum equal to 0. Put this way, the result does not conﬂict with CLASS; indeed, it is a result of CLASS, since the proof within RUSS is actually a proof within CLASS that does not use such nonconstructive logical principles as LEM. By the mid-1960s, constructive mathematics was, when compared with its classical counterpart, virtually stagnant. The situation changed in 1967 with the publication of Errett Bishop’s monograph Foundations of Constructive Mathematics [9]. This book and its oﬀspring [12] represent the most far-reaching and systematic presentation of constructive analysis to date. In [9], Bishop revealed, by thoroughgoing constructive means but without resorting to either Brouwer’s principles or the formalism of recursive function theory, a vast panorama of constructive mathematics, covering elementary analysis, metric and normed spaces, abstract measure and integration, the spectral theory of selfadjoint operators on a Hilbert space, Haar measure, duality on locally compact groups, and Banach algebras. Bishop’s constructive mathematics (BISH) was founded on a primitive, unspeciﬁed notion of “algorithm”, or “ﬁnite routine”, and the Peano properties of natural numbers, and kept strictly to the interpretation of “existence ” as “computability”. His refusal to pin down the notion of algorithm led to criticism, particularly from philosophers of mathematics and from those committed to the Church–Markov–Turing thesis; but this very imprecision enabled Bishop’s work to have a variety of interpretations: his results are valid in CLASS, INT, RUSS, and all reasonable models of computable mathematics, such as the more recent one propounded by Weihrauch [91]. Indeed, from a purely formal viewpoint, each of INT, RUSS, and CLASS can be regarded as BISH plus some additional principles: INT can be regarded as BISH supplemented by Brouwer’s continuity principle and fan theorem; RUSS as BISH plus the Church–Markov–Turing thesis; and CLASS as BISH plus the law of excluded middle. One consequence of this multiplicity of interpretations is that we can often demonstrate that certain propositions P are independent of BISH; that is, neither P nor (not P ) can be proved within BISH. For example, since “every mapping from [0, 1] into R is uniformly continuous” is a theorem of INT, and “there exists a continuous map of [0, 1] into R that is not uniformly continuous” is a theorem of

1.3 Intuitionistic Logic

7

RUSS, and since both INT and RUSS are formally consistent with BISH, within BISH we cannot expect either to prove that every continuous map of [0, 1] into R is uniformly continuous or to construct an example of a real-valued function that is deﬁned, but not uniformly continuous, on [0, 1]. Over the years since the publication of Bishop’s book, it became clear to a number of researchers that, in essence, BISH is simply mathematics with intuitionistic logic together with some appropriate set-theoretic foundation. As we pointed out at the end of the preceding section, working with intuitionistic logic automatically bars noncomputational steps. As long as we keep strictly to intuitionistic logic, having made sure that our set-theoretic principles do not inadvertently imply LEM or some other nonconstructive proposition, the mathematics we develop turns out to be predictive, in the sense that every proof implicitly shows that if we perform certain calculations, we shall achieve certain results. Accordingly, when we speak of “constructive mathematics” or “BISH” in future, we shall mean “mathematics with intuitionistic logic”. It therefore behooves us to explain more clearly exactly what intuitionistic logic is.

1.3 Intuitionistic Logic

The meaning doesn’t matter if it’s only idle chatter of a transcendental kind. —W.S. Gilbert, Patience Everywhere one seeks to produce meaning, to make the world signify, to render it visible. —Jean Baudrillard, Seduction, or the Superﬁcial Abyss

For Brouwer, mathematics took precedence over logic. In order to describe the logic used by the (intuitionist) mathematician, it was necessary ﬁrst to analyse the mathematical processes of the mind, from which analysis the logic could be extracted. In 1930, Brouwer’s most famous pupil, Arend Heyting (1898–1980), published a set of formal axioms that so clearly characterise the logic used by the intuitionist that they have become universally known as the axioms for intuitionistic logic [52]. These axioms capture the so-called BHK interpretation of the connectives ∨ (or), ∧ (and), =⇒ (implies), ¬ (not) and quantiﬁers ∃ (there exists), which we now outline.

∀ (for all/each),

8

1 Introduction to Constructive Mathematics

P ∨ Q : either we have a proof of P or else we have a proof of Q. P ∧ Q : we have both a proof of P and a proof of Q. P =⇒ Q : by means of an algorithm we can convert any proof of P into a proof of Q. ¬P : assuming P, we can derive a contradiction (such as 0 = 1); equivalently, we can prove (P =⇒ (0 = 1)) . ∃x P (x) : we have (i) an algorithm that computes a certain object x, and (ii) an algorithm that, using the information supplied by the application of algorithm (i), demonstrates that P (x) holds. ∀x ∈ A P (x) : we have an algorithm that, applied to an object x and a proof that x ∈ A, demonstrates that P (x) holds. Note that in the interpretation of the statement ∀x ∈ A P (x), the proof of P (x) will normally use both the data describing the object x and the information supplied by a proof that x belongs to the set A. This is an important point, since upon it hinges a key argument against the use of the axiom of choice in constructive mathematics. We shall return to this matter later. A property P (x) is said to be decidable if for each x to which it might be applicable we have P (x) ∨ ¬P (x), where the disjunction and negation are given their BHK interpretations. Even for a decidable property P (n) of natural numbers n the property ∀n P (n) ∨ ¬∀n P (n), and hence a fortiori LEM, will not hold in general. As a result, many classical results cannot be proved constructively, since they would imply LEM or perhaps some other manifestly nonconstructive principle. To illustrate this point, consider the following simple statement, the limited principle of omniscience (LPO): ∀a ∈ {0, 1}

N+

(a = 0 ∨ a = 0) , N+

where a = (a1 , a2 , . . .) , N+ = {1, 2, . . .} is the set of positive integers, {0, 1} the set of all binary sequences, and

is

a = 0 ⇐⇒ ∀n (an = 0) , a = 0 ⇐⇒ ∃n (an = 1) . In words, LPO states that for each binary sequence (an )n1 , either an = 0 for all n or else there exists n such that an = 1. Of course, this is a triviality from the

1.3 Intuitionistic Logic

9

viewpoint of classical logic. But its BHK interpretation is not so simple: it says that there is an algorithm that, applied to any binary sequence a, either veriﬁes that all the terms of the sequence are 0 or else computes the index of a term equal to 1. Anyone familiar with computers ought to be highly sceptical about such an algorithm, since in the case a = 0 it would normally need to test each of the inﬁnitely many terms an in order to come up with the correct decision. For such reasons we feel justiﬁed in not accepting LPO, or any classical proposition that constructively implies LPO, as a valid principle of constructive mathematics. But we have another reason for not doing so: it can be shown that there are models of Heyting arithmetic—Peano arithmetic with intuitionistic logic—in which LPO is false; so LPO cannot be derived in Heyting arithmetic (see [34, 48]). Since LPO is a special case of the law of excluded middle, we are led, in turn, to renounce the latter when working constructively. Similar informal analyses lead us to exclude both the classical rule ¬¬P =⇒ P, which forms the basis of proof by contradiction, and the following lesser limited principle of omniscience (LLPO), which is easily seen to be a consequence of LPO: For each binary sequence a with at most one term equal to 1 (in the sense that am an = 0 for all distinct m and n), either a2n = 0 for all n or else a2n+1 = 0 for all n. The exclusion of such principles from constructive mathematics has serious consequences for mathematical practice. For example, we cannot hope to prove constructively the simple statement ∀x ∈ R (x = 0 ∨ x = 0) ,

(1.1)

where R denotes the set of real numbers and x = 0 means that we can compute a rational number strictly between 0 and x (which, as we shall see when we deal with the real numbers more formally in Chapter 2, is not the same, constructively, as proving that ¬ (x = 0)). To prove this, consider any binary sequence a, and use it to deﬁne the binary expansion of a real number x=

∞

an 2−n .

n=1

If x = 0, then a = 0. If x = 0, we can compute a positive integer N such that x > 2−N =

∞

2−n ;

n=N +1

it is then clear that, by testing the terms a1 , . . . , aN , we can ﬁnd n N such that an = 1. Thus statement (1.1) about real numbers implies LPO and is therefore essentially nonconstructive.

10

1 Introduction to Constructive Mathematics

If the binary sequence a has at most one term equal to 1, then we can use the real number ∞ n (−1) an 2−n n=1

to show that the statement ∀x ∈ R (x 0 ∨ x 0) implies LLPO. The following elementary classical statements also turn out to be nonconstructive. Each real number x is either rational or irrational (that is, x = r for each rational number r). To see this, consider x=

∞ 1 − an , n! n=1

where a is any increasing binary sequence (that is a binary sequence such that an an+1 for each n). Each real number x has a binary expansion. Note that the standard intervalhalving argument for “constructing” binary expansions does not work, since we cannot necessarily decide, for a given number x between 0 and 1, whether x 1/2 or x 1/2. In fact, the existence of binary expansions is equivalent to LLPO. The intermediate value theorem, which is equivalent to LLPO. For all x, y ∈ R, if xy = 0, then either x = 0 or y = 0. The constructive failure of this proposition clearly has implications for the theory of integral domains. We emphasise here that classically valid statements like “each real number is either rational or irrational” that imply omniscience principles are not false in constructive mathematics; they cannot be, since BISH is consistent with CLASS. One principle whose constructive status is controversial is Markov’s principle (MP): N+ (¬ (a = 0) =⇒ a = 0) ; ∀a ∈ {0, 1} in words, for any binary sequence a, if it is impossible for all the terms to equal 0, then there exists a term equal to 1. In order to accept this as a principle of constructive mathematics, you have to be convinced that the information conveyed by the antecedent ¬ (a = 0) is suﬃcient to enable us to compute an index n with an = 1. The argument in favour of MP says that we can carry out this computation by searching systematically through the terms an , since the hypothesis ¬ (a = 0)

1.4 Informal Set Theory

11

guarantees that we shall eventually stumble across a term equal to 1. The counterargument is that the antecedent provides us with no prior bound for such a search— it does not tell us how many terms we need to test before we arrive at one equal to 1—so the search might go on longer than the remaining life of the universe before it produced the desired result. Moreover, Markov’s principle, like LPO, cannot be proved within Heyting arithmetic. For these reasons, we shall follow the normal practice of excluding MP from the working principles of constructive mathematics. As a consequence we exclude the even stronger logical principle (∀x ∈ A (P (x) ∨ ¬P (x)) ∧ ¬∀x ∈ A ¬P (x)) =⇒ ∃x ∈ A P (x),

(1.2)

where A is a well-deﬁned set (the exact meaning of “well-deﬁned set” will become clear in the next section). In fact, even if we were to accept Markov’s principle on the grounds that an unbounded search through the natural numbers that cannot fail to terminate must eventually do so, we would balk at accepting (1.2), since for a general set A there will be no natural order allowing us to search systematically in the way we can with N. An example of the type dealt with earlier, in which a classically valid proposition P is shown constructively to entail an essentially nonconstructive principle like LEM, LPO, LLPO, or even MP, is called a Brouwerian counterexample to P (even though it is not a counterexample in the true sense of the word; it is merely an indication that P does not admit of constructive proof). There is another expression ∞ an 2−n that we may use in this context. For example, we refer to the number x = n=1

that we constructed from a given binary sequence (an )n1 and then used to show that (1.1) implies LPO as a Brouwerian example of a real number x for which we cannot decide whether x = 0 or x = 0.

1.4 Informal Set Theory Die ganzen Zahlen hat der liebe Gott gemacht, alles andere ist Menschenwerk. —L. Kronecker [90] The primary concern of mathematics is number, and this means the positive integers. We feel about number the way Kant felt about space. The positive integers and their arithmetic are presupposed by the very nature of our intelligence and, we are tempted to believe, by the very nature of intelligence in general. The development of the positive integers from the primitive concept of the unit, the concept of adjoining a unit, and the process of mathematical induction carries complete conviction. In the words of Kronecker, the positive integers were created by God. —Errett Bishop [9]

Building on the set of positive integers and using intuitionistic logic, we follow Bishop’s approach to developing constructive mathematics at higher and higher

12

1 Introduction to Constructive Mathematics

levels of abstraction. To do this, we need to clarify notions such as “set” and “function”. For us, a set X (other than our basic, primary set N+ of positive integers) is given by two pieces of data: a property that enables members of X to be constructed using objects that have already been constructed (note this last phrase, which rules out the possibility of impredicative deﬁnitions and therefore of Russell-type paradoxes), and an equivalence relation =X of equality between members of X. We write x ∈ A to signify that x is an element of the set A, and x ∈ / A instead of ¬ (x ∈ A) . The use of equivalence relations rather than intensional equality—that is, identity of description—is common, but often goes unnoticed, in classical mathematics. For example, we call the rational numbers 1/2 and 3/6 equal, even though, strictly speaking, they are equivalent and not intensionally identical. A subset S of a set X consists of a collection of elements drawn from X, together with the equality relation induced on S by the given equality on X; that is, for elements x, y of S, we deﬁne x =S y ⇐⇒ x =X y. We write S ⊂ T to signify that S is a subset of T. If P (x) is a property applicable to certain elements x of a set A, then we denote by {x : x ∈ A ∧ P (x)} or {x ∈ A : P (x)} the subset of A consisting of those elements x of A with the property P (x). The logical complement of a subset S of X is ¬S = {x ∈ X : x ∈ / S} . A particular example of this is the empty subset of X, deﬁned by ∅X = ¬X. We say that a subset S of X is inhabited if ∃x (x ∈ S) . We then write S = ∅X . Note that in order to show that S is inhabited, we cannot just prove that it is impossible for S to be empty; we must actually construct an element of S; see Exercise 2.

1.4 Informal Set Theory

13

Two sets X, Y are said to be equal sets if each is a subset of the other; in other words, if the sets have the same elements and the same equality relation. We need to be careful when constructing new sets from old. Since an equality is part of the data for a set, it does not make sense to talk of the union S ∪ T of two sets unless we can put together not only the sets as collections of objects but also, in some way, their given equality relations. In practice, this means that in order to construct their union S ∪ T, the sets S and T must be given as subsets of some set X. We then deﬁne S ∪ T = {x ∈ X : x ∈ S ∨ x ∈ T }, where the equality on S ∪ T is that induced by X. Likewise, the intersection of S and T is deﬁned only when S and T are subsets of some set X, and is then the subset S ∩ T = {x ∈ X : x ∈ S ∧ x ∈ T } of X. The (Cartesian) product of two sets X, Y is the set X × Y consisting of all ordered pairs (x, y) with x ∈ X and y ∈ Y, together with the equality given by ((x, y) =X×Y (x , y )) ⇐⇒ (x =X x ∧ y =Y y ) . In many situations—even, as we shall see, on the real line—we frequently need a set X to be equipped with an inequality relation =X describing what it means for two elements of X to be unequal, or distinct. Such a relation must satisfy the following two properties: x =X y =⇒ ¬ (x =X y) , x =X y =⇒ y =X x. If, in addition, ¬ (x =X y) =⇒ x =X y, we say that the inequality is tight. A set X with an inequality is discrete if, for any two elements x and y of X, either x =X y or x =X y; we then also describe the inequality itself as discrete. One inequality relation, the denial inequality, is deﬁned by setting x =X y if and only if ¬ (x =X y) . This inequality is normally too weak for practical purposes. For example, in the absence of Markov’s principle, on the real line R the property ¬ (x =R 0) is weaker than |x| > 0 (Exercise 3); for that reason we deﬁne the standard inequality on R to be not the denial inequality but the one given by x =R y ⇐⇒ |x − y| > 0. From now on, when the meaning is clear from the context, we write =, ∅, =, . . .

14

1 Introduction to Constructive Mathematics

rather than =X , ∅X , =X , . . . . A subset S of a set X with an inequality has a complement, deﬁned by ∼S = {x ∈ X : ∀s ∈ S (x = s)} . Then ∼S ⊂ ¬S; but unless the inequality on X is the denial inequality, the reverse inclusion will not hold. A subset S of a set X is said to be detachable (from, or in, X) if ∀x ∈ X (x ∈ S ∨ x ∈ / S) . Since statement (1.1) implies LPO, not even the singleton subset {0} is detachable from R. However, {0} is detachable in the set Q of rational numbers. When X comes with an inequality relation, we deﬁne the inequality on any subset S of X to be the one induced by that on X: x =S y ⇐⇒ x =X y. If also Y has an inequality relation, we deﬁne the inequality on the Cartesian product X × Y by ((x, y) =X×Y (x , y )) ⇐⇒ (x =X x ∨ y =Y y ) . It should be no surprise that we require functions to be given by algorithms and to respect equality. Thus a function f from a set X to a set Y —also called a map or mapping of X into Y , and written f : X −→ Y —is an algorithm that, applied to any element x of X, produces an element f (x) of Y such that f is extensional : ∀x ∈ X ∀x ∈ X (x =X x =⇒ f (x) =Y f (x )) . The element f (x) is called the value of f at x or the image of x under f . If X and Y have inequality relations, then we may require f to be strongly extensional : ∀x ∈ X ∀x ∈ X (f (x) =Y f (x ) =⇒ x =X x ) . Note that the statement “all functions from R to R are strongly extensional” is equivalent to Markov’s principle. Let f, g be two real-valued functions on a set X with an inequality. We say that an element x0 of X is the strongly unique element of X such that f (x) = g(x) if f (x0 ) = g(x0 ) and f (x) = g(x) whenever x ∈ X and x = x0 . Strong uniqueness will resurface in Chapter 4 in connection with best approximations. A partial function f : X −→ Y is a function from a subset of X into Y. The subset

1.4 Informal Set Theory

15

{x ∈ X : f (x) is deﬁned} is called the domain of f, denoted by dom (f ) ; and the set {y ∈ Y : ∃x ∈ X (y =Y f (x))} the range of f, denoted by ran (f ) . The image of a subset A of X under f is the set f (A) = {f (x) : x ∈ A}. The inverse image of a subset B of Y under f is the set f −1 (B) = {x ∈ dom (f ) : f (x) ∈ B} . A partial function f : X −→ Y is said to be a total partial function on X if dom(f ) = X. When the expression describing f (x) is given explicitly and the domain of the partial function f is clearly understood, we may denote the function by x f (x). For example, the partial function from R to itself whose value is deﬁned, for each real number x such that x =R 0, to be 1/x may be written x 1/x. An important type of total partial function is deﬁned as follows. Let X and I be sets. A family of elements of X with index set I (or indexed by I) is a mapping i xi of I into X; we commonly denote this family by (xi )i∈I . In particular, if I is N+ , the family is called a sequence in X and is usually written (xn )n1 . More general sequences of the form (xn )nN have the obvious analogous meaning when N is an integer. Let f and g be mappings from subsets of a set X into a set Y, where Y is equipped with a binary operation 3. We introduce the corresponding pointwise operation 3 on f and g by setting (f 3g)(x) = f (x)3g(x) whenever f (x) and g(x) are both deﬁned. Thus, taking Y = R, we see that the (pointwise) sum of f and g is given by (f + g)(x) = f (x) + g(x) if f (x) and g(x) are both deﬁned; and that the (pointwise) quotient of f and g is given by ( f / g)(x) = f (x)/ g(x) if f (x) and g(x) are deﬁned and g(x) = 0. If X = N+ , so that f = (xn )n1 and g = (yn )n1 are sequences, then we also speak of termwise operations; for example, the termwise product of f and g is the sequence (xn yn )n1 . Pointwise operations extend in the obvious ways to ﬁnitely many functions. In the case of a sequence (fn )n1 of functions with values in a normed space (see

16

1 Introduction to Constructive Mathematics

Chapter 2), once we have introduced the notion of a series in a normed space we ∞ shall interpret fn in the obvious pointwise way. n=1

A partial function f : X −→ Y can be identiﬁed with its graph, G (f ) = {(x, y) ∈ X × Y : x ∈ dom (f ) , y ∈ ran (f ) , y = f (x)}, a subset of the Cartesian product X×Y. We deﬁne two partial functions f : X −→ Y and g : X −→ Y to be equal if their graphs are equal as sets. Thus f, g are equal partial functions if and only if their domains are equal subsets of X and f (x) = g(x) for each x ∈ dom(f ). We say that a partial function f : X −→ Y is one-one if f (x) = f (x ) entails x = x ; injective if X and Y are equipped with inequality relations, and x = x entails f (x) = f (x ). If f is injective and the inequality on X is tight, then f is one-one: for in that case, if f (x) = f (x ), then ¬ (x = x ) and so, by tightness, x = x . The composition, or composite, of partial functions f : A −→ B and g : B −→ C is the partial function g ◦ f : A −→ C (sometimes written gf ) deﬁned by g ◦ f (x) = g(f (x)) wherever the right side exists. A partial mapping f : X −→ Y maps its domain onto Y if ∀y ∈ Y ∃x ∈ X (y = f (x)) . On the other hand, we say that f : X −→ Y is an epimorphism if there exists a mapping g : Y −→ X such that ∀y ∈ Y (f (g(y)) = y) . A one-one partial function f : X −→ Y has a one-one inverse f −1 : ran (f ) −→ dom (f ) deﬁned by f −1 (f (x)) = x. If f is injective, then its inverse is strongly extensional. A bijection between X and Y is a one-one mapping from X onto Y . The subtle distinction between mappings onto and epimorphisms is closely linked to the constructive status of the axiom of choice, which we shall discuss shortly. First, though, we introduce some notions of cardinality. Let S be a subset of a set X with an inequality relation. We say that S is ﬁnitely enumerable if there exist a natural number N and a mapping of the set

1.4 Informal Set Theory

17

{1, 2, . . . , N } = {n ∈ N+ : n N } onto S; ﬁnite if there exist a natural number N and an injective map of {1, 2, . . . , N } onto S. When we speak of ﬁnitely many objects, we mean that those objects constitute an inhabited, ﬁnitely enumerable, but not necessarily ﬁnite, set. Note that the case N = 0 of the deﬁnition of “ﬁnitely enumerable” shows that the empty subset of X is ﬁnitely enumerable. Clearly, ﬁnite implies ﬁnitely enumerable; the converse does not hold constructively. We say that S is countable if there exist a detachable subset D of N+ and a mapping φ of D onto S. Every ﬁnitely enumerable subset—in particular, the empty subset—of X is countable. If D = N+ , we call φ an enumeration of S, in which case we may denote S by {φ(1), φ(2), . . .} . If also φ is one-one, we say that S is countably inﬁnite. A set is countably inﬁnite if and only if it is the range of a one-one mapping whose domain is a countably inﬁnite, detachable subset of N+ . We now consider the axiom of choice (AC): If X, Y are inhabited sets, S is a subset of X × Y, and for each x ∈ X there exists y ∈ Y such that (x, y) ∈ S, then there exists a choice function f : X −→ Y such that (x, f (x)) ∈ S for each x ∈ X. Under the BHK interpretation, the hypothesis ∀x ∈ X ∃y ∈ Y ((x, y) ∈ S) of AC means that we have an algorithm that, applied to each element x of X and the data showing that x belongs to X, constructs an element y of Y and demonstrates that (x, y) ∈ S. This much is clear. However, there is no guarantee that the algorithm will respect the equality relation on X—in other words, that if x =X x , and the algorithm constructs y, y in Y such that (x, y) ∈ S and (x , y ) ∈ S, then y =Y y . Indeed, we should expect that the computation of y might use data that are associated with properties intrinsic to x that do not apply intrinsically to x . For example, anticipating our development of the real number set R, consider the case in which X is R and Y is N+ . A real number is (deﬁned as) a certain set of rational approximations. However, two equal real numbers x, x can have diﬀerent deﬁning sets of rational approximations. In that case, the algorithm that computes a positive integer n such that (x, n) ∈ S may, and in general will, compute a diﬀerent positive integer n such that (x , n ) ∈ S. These considerations throw real doubt over the possibility that there is a choice function implementing the algorithm.

18

1 Introduction to Constructive Mathematics

In fact, an argument of Diaconescu [46] and Goodman & Myhill [50], but preﬁgured by Bishop (see Problem 2 on page 58 of [9]), shows that AC cannot be allowed as a principle of constructive mathematics. Theorem 1.4.1. The axiom of choice implies the law of excluded middle. Proof. Let P be any constructively meaningful statement, and deﬁne the set X to consist of the two elements 0 and 1, together with the equality relation such that (0 =X 1) ⇐⇒ P. Let Y be the set {0, 1} with the standard equality, and let S be the subset {(0, 0), (1, 1)} of the Cartesian product X × Y, taken with the standard equality. Suppose there exists a function f : X −→ Y such that (x, f (x)) ∈ S for all x ∈ X. There are three cases to consider: (i) f (0) = 1, (ii) f (1) = 0, and (iii) both f (0) = 0 and f (1) = 1. In case (i) we have (0, 1) = (0, f (0)) ∈ S, so either (0, 1) =X×Y (0, 0) or (0, 1) =X×Y (1, 1). If the ﬁrst of these two alternatives holds, then, by deﬁnition of the equality on X × Y, we have 1 =Y 0, which is absurd. Hence, in fact, (0, 1) =X×Y (1, 1) . Thus, again by deﬁnition of the equality on X × Y, we have 0 =X 1 and therefore P holds. Case (ii) similarly leads to the conclusion that P holds. Finally, in case (iii) we have ¬ (f (0) =Y f (1)) ; therefore, since f is a func2 tion, ¬ (0 =X 1) and so ¬P holds. Thus we have derived P ∨ ¬P from AC.

The axiom of choice will hold constructively if the set X is one for which no computation is necessary to demonstrate that an element belongs to it; Bishop calls such sets basic sets. Following the practice of most constructive mathematicians, we consider N+ , the set N = {0, 1, 2, 3, . . .} of natural numbers, and the set Z = {0, ±1, ±2, . . .} of all integers to be basic sets. This practice is reﬂected in our acceptance of the principle of countable choice: If Y is an inhabited set, S is a subset of N+ × Y, and for each positive integer n there exists y ∈ Y such that (n, y) ∈ S, then there is a function f : N+ −→ Y such that (n, f (n)) ∈ S for each n ∈ N+ . In fact, many constructive proofs use the stronger principle of dependent choice: If X is a set, a ∈ X, S is a subset of X × X, and for each x ∈ X there exists y ∈ X such that (x, y) ∈ S, then there exists a sequence (xn )n1 in X such that x1 = a and (xn , xn+1 ) ∈ S for each n ∈ N+ .

1.4 Informal Set Theory

19

Another contentious matter in constructive mathematics is the status and role of the power set P (X) of a given set X: that is, the collection of all subsets of X, with equality of subsets as deﬁned earlier. The main objection to admitting P (X) into the constructive fold is that we thereby allow impredicativity, since there is then nothing to stop us constructing subsets of X whose deﬁning characteristics are self-referential. On the other hand, nobody has yet shown that adding the power set axiom ∀X∃Y ∀S (S ⊂ X ⇐⇒ S ∈ Y ) to constructive mathematics enables us to prove LEM or some other incontestably nonconstructive principle. There are ways of avoiding the power set. It often suﬃces to work with the set Y X of all mappings from X into a set Y. Since we have a clear idea of what mappings from X into Y are (something we do not have for subsets of X), the set Y X seems relatively innocent. Note that classically the set {0, 1}X can be identiﬁed with the power set of X, since it comprises the characteristic functions of subsets of X. This identiﬁcation is not possible constructively, since characteristic functions exist only for those subsets of X that are detachable. Another way to avoid the full generality of the power set is to work with a welldeﬁned but smaller set of subsets of X. For example, the set of compact subsets of a metric space X is well deﬁned (it is actually a metric space itself), and is often all we need for many parts of analysis. Let S (X) be a well-deﬁned set of subsets of X, with two elements taken as equal if and only if they are equal sets in the usual sense, and let I be some set. Then we can speak sensibly about a family (Si )i∈I of elements of S (X) . We can also deﬁne the union and intersection of such a family to be, respectively, the subsets Si = {x ∈ X : ∃i ∈ I (x ∈ Si )} i∈I

and

Si = {x ∈ X : ∀i ∈ I (x ∈ Si )}

i∈I

of X. set of positive integers, we denote the above union and intersection If I is the Sn and Sn , respectively. by n1

n1

The Cartesian product

i∈I

Si is the subset of X I consisting of those functions f

∈ I; if X comes with an inequality relation, then the such that f (i) ∈ Si for each i

Si is deﬁned by corresponding inequality on i∈I

f = g ⇐⇒ ∃i ∈ I (f (i) = g(i)) . If I

is the set of positive integers, then (Si )i∈I is a sequence of sets, and the elements Si are also sequences; in this case, the sets Si can be arbitrary and need not of i∈I

20

1 Introduction to Constructive Mathematics

be subsets

of a previously deﬁned set X. If I = {1, 2, . . . , n}, then we denote the Si by S1 × S2 × · · · × Sn , and refer to its elements (x1 , x2 , . . . , xn ) as product i∈I

ordered n-tuples, or, in the case n = 2, ordered pairs (something we used informally in our earlier deﬁnition of the Cartesian product of two sets). Other set-theoretic notions will be introduced later as they arise. It is now time to turn our attention away from foundational matters to analysis proper.

Exercises Although we shall not formally deﬁne the real number line R until Chapter 2, in these exercises we assume elementary properties of real numbers where necessary. 1. Justify informally Brouwer’s observation that ¬¬¬P implies ¬P. Using this, show that the proposition (¬¬P =⇒ P ) is equivalent to the law of excluded middle. 2. Prove that if the statement ¬ (S = ∅) =⇒ S = ∅ applies to all subsets S of {0} , then the law of excluded middle holds. 3. Prove that ∀x ∈ R (¬ (x = 0) =⇒ x = 0) is equivalent to Markov’s principle. 4. Fill in the details of the proof that the statement ∀x ∈ R (x 0 ∨ x 0) implies LLPO. 5. Give a Brouwerian counterexample to the proposition that every real number x is either rational or irrational (where “x is irrational” means that x = r for each rational number r). 6. Prove that every real number has a binary expansion if and only if LLPO holds. 7. Give a Brouwerian counterexample to the statement that if r1 , r2 , and r3 are real roots of a quadratic polynomial x2 + ax + b with a, b ∈ R, then there exist distinct i, j with ri = rj .

1.4 Informal Set Theory

21

8. Prove that the statement “for all real numbers x, y that have binary expansions, the sum x + y has a binary expansion” is equivalent to LLPO. 9. Prove that the statement ∀x, y ∈ R (xy = 0 =⇒ (x = 0 ∨ y = 0)) is equivalent to LLPO. 10. Prove that the intermediate value theorem is equivalent to LLPO. 11. Let f : [a, b] −→ R be sequentially continuous in the following sense: for each sequence (xn )n1 in [a, b] that converges to a limit x, the sequence (f (xn ))n1 converges to f (x). Suppose also that f (a)f (b) < 0, and that f is locally nonzero in the sense that for each x ∈ [a, b] and each r > 0 there exists y ∈ [a, b] with |x − y| < r and f (y) = 0. Prove that there exists c ∈ (a, b) such that f (c) = 0. (This version of the intermediate value theorem suﬃces for virtually all constructive purposes.) 12. Prove that the statement “all functions from R to R are strongly extensional” is equivalent to Markov’s principle. 13. Discuss the statement “every mapping from a set X onto a set Y is an epimorphism”. 14. Give a Brouwerian counterexample to the statement “every subset of a ﬁnite set is ﬁnitely enumerable”. 15. Prove that an inhabited set is countably inﬁnite if and only if it is the range of a function with domain N. 16. Prove that a subset S of N is countable if and only if there exists a sequence (Sn )n1 of ﬁnite subsets of N such that S1 ⊂ S2 ⊂ S3 ⊂ · · · . 17. Prove that the statement “every inhabited subset of N is countable” implies the principle of ﬁnite possibility (PFP): to each binary sequence (an )n1 there corresponds a binary sequence (bn )n1 such that an = 0 for each n if and only if there exists N such that bN = 1. 18. Prove that the intersection of two countable sets is countable. 19. Prove that the principle of dependent choice implies the principle of countable choice. Prove also that the principle of dependent choice can be derived from the following principle of internal choice: if X is an inhabited set, S is a subset of X × X, and for each x ∈ X there exists y ∈ X such that (x, y) ∈ S, then there exists a choice function f : X −→ X such that (x, f (x)) ∈ S for each x ∈ X.

22

1 Introduction to Constructive Mathematics

20. By a ﬁlter on a set X we mean a set F of inhabited subsets of X with the following properties: • X ∈ F. • If S and T belong to F, then S ∩ T ∈ F. •

If S ∈ F and S ⊂ T ⊂ X, then T ∈ F.

A ﬁlter U is called an ultraﬁlter if for all S ⊂ X, either S ∈ U or ¬S ∈ U. The classical ultraﬁlter principle states that every ﬁlter is contained in an ultraﬁlter. Prove that this principle implies the weak limited principle of omniscience (WLPO): ∀a ∈ 2N (∀n (an = 0) ∨ ¬∀n (an = 0)) .

Notes The use of the Goldbach conjecture in the example in Section 1.1 is in no sense a constraint on the argument; if the Goldbach conjecture were solved tomorrow, we could easily replace it by any one of many open problems of number theory. Indeed, until Wiles proved the Fermat conjecture in 1994, it was common to use that, rather than the Goldbach conjecture, in such Brouwerian examples of nonconstructivity. The reader interested in the history of foundations of mathematics should consult [44] and [84] for more information on the “Grundlagenstreit” that eventuated between Brouwer and Hilbert in the 1920s. For more on Hilbert’s programme for the secure foundation of mathematics see [93]. The designation “BHK interpretation” comes from Brouwer, Heyting, and Kolmogorov. The BHK interpretation of implication, while more natural than the classical one of material implication, in which (P =⇒ Q) is equivalent to (¬P ∨ Q), has not completely satisﬁed all researchers using constructive logic. Shortly before he died, Bishop communicated to Bridges his dissatisfaction with the standard constructive interpretation of implication. Unfortunately, he left nothing more than very rudimentary sketches of his ideas for its improvement. (Note, however, Bishop’s paper [10].) For a deeper analysis of the constructive interpretations of the connectives and quantiﬁers, we refer the reader to [48]. The classical invalidity of the recursive interpretation of LPO is not a matter of logic: it can be demonstrated, even with classical logic, that a recursive version of LPO would lead to a proof of the decidability of the halting problem, which is known to be impossible; see [34], pages 52–53. Andrej Bauer has recently shown that proofs and results in BISH can be translated into Weihrauch’s Type-2 Eﬀectivity framework by a realisability interpretation [5].

1.4 Informal Set Theory

23

The principle (¬¬P =⇒ P ) is equivalent to LEM even with intuitionistic logic; see Exercise 1. For the conﬁrmed intuitionist there is at least one other reason for rejecting MP: it contradicts Brouwer’s (admittedly controversial) theory of the creating subject, an add-on to his intuitionistic mathematics. See [34] (pages 116–117) and, for a rather diﬀerent view, [73]. Bishop used “subﬁnite” instead of “ﬁnitely enumerable”. Note that a subset of a ﬁnite set need not be ﬁnitely enumerable. In the proof of Theorem 1.4.1 we could have deﬁned X as a set of equivalence classes under the equivalence relation ∼ deﬁned by (0 ∼ 1) ⇐⇒ P, but it is more in keeping with Bishop’s approach to proceed by considering a special, if unusual, equality relation on X. Some constructive mathematicians argue against even the principle of countable choice; see, for example, [76, 82]. Myhill has shown that, under the Church–Markov–Turing thesis, the power set of a singleton is uncountable: in other words, there is no recursive mapping of N onto that power set [74] (page 364, Theorem 3). Within BISH we may not be able to prove the uncountability of the power set of {0} , but we certainly have an unending supply of subsets S of {0} for which we cannot decide that S = ∅ or S = {0}: given any constructively meaningful statement P, we can deﬁne a corresponding set S = {x : x = 0 ∧ P } such that (S = {0} ∨ S = ∅) if and only if P ∨ ¬P. Myhill’s formal theory—based on primitive notions of “set”, “function”, and “natural number”—is but one of several foundational theories advocated for constructive mathematics. Others include Aczel’s constructive set theory [3], MartinL¨ of’s type theory [69], and a largely unpublished constructive version of Morse’s set theory [16] in which membership of the universe appears to correspond to being constructively deﬁned. For more on constructive foundational theories see [6] and [88].

2 Techniques of Elementary Analysis

My deplorable mania for analysis exhausts me. —Gustave Flaubert, Letter (August 1846) We begin by using a form of interval arithmetic as a foundation for the construction of the real number line R. Having discussed the elementary algebraic and order-theoretic properties of real numbers, we prove that R is complete in two senses: Dedekind (order) complete and Cauchy (sequentially) complete. The next step is to generalise from the reals to metric and normed spaces. A particularly important property for us is total boundedness, which plays a big part in proving the existence of suprema and inﬁma; for that reason we need to prove that there are lots of totally bounded subsets in a totally bounded space. We also highlight the important property of locatedness for subsets of a metric space, a property that holds automatically in classical mathematics.

2.1 The Real Number Line Constructive analysis proper, as distinct from arithmetic, begins with the real numbers, which we shall construct, choice-free, by interval arithmetic. First, though, we observe that the purely algebraic constructions of the sets Z of integers and Q of rational numbers from N are carried out as in classical algebra, and that the standard inequality on Q —and hence a fortiori on Z regarded as a subset of Q —is the (in this case discrete) denial inequality. By a real number we mean a subset x of Q × Q such that for all (q, q ) in x, q q ; for all (q, q ) and (r, r ) in x, the closed intervals [q, q ] and [r, r ] in Q intersect in points of Q;

26

2 Techniques of Elementary Analysis

for each positive rational ε there exists (q, q ) in x such that q − q < ε. The last of these properties ensures that x is inhabited. It also legitimises our use of expressions like “pick an element of x”: to carry out such a selection, we simply take ε = 1 in the third of the deﬁning properties of the real number x, to produce a corresponding element (q, q ) of x. The intuition underlying our deﬁnition of “real number” is that the elements of x are the rational endpoints of closed intervals with one point—namely x —in common. Any rational number q gives rise to a canonical real number q = {(q, q)} with which the original rational q is identiﬁed. Two real numbers x and y are •

equal, written x = y, if for all (q, q ) ∈ x and all (r, r ) ∈ y, the intervals [q, q ] and [r, r ] in Q have a rational point in common;

•

unequal (or distinct), written x = y, if there exist (q, q ) ∈ x and (r, r ) ∈ y such that the intervals [q, q ] and [r, r ] in Q are disjoint.

It is almost immediate that = satisﬁes the deﬁning properties of an inequality relation. Let us check that equality is an equivalence relation. It is trivial that it is reﬂexive and symmetric, so only transitivity has to be handled. Let x = y and y = z, and suppose that for some (q, q ) ∈ x and (r, r ) ∈ z there are no rational points in [q, q ]∩[r, r ] . We may assume without loss of generality that q < r. Using the third of the deﬁning properties for a real number, choose (s, s ) ∈ y such that s − s < r − q . Now, the rational interval [s, s ] intersects [q, q ] in a rational point u, and [r, r ] in a rational point v, so r − q v − u s − s < r − q , a contradiction. Hence

¬ ([q, q ] ∩ [r, r ] = ∅) .

Since we are working with intervals in Q, with the aid of two simple lemmas we can turn this around to construct a point of [q, q ] ∩ [r, r ] , and therefore complete the proof that x = z, as follows. Lemma 2.1.1. Let a, b, c, d be rational numbers with a b and c d. Then there exists a rational number in [a, b] ∩ [c, d] if and only if a d and c b. Proof. If r is a rational number in the intersection of the intervals, then a r d and c r b, so a d and c b. If, conversely, these conditions hold, then either c < a and therefore a ∈ [a, b] ∩ [c, d] , or else a c and c ∈ [a, b] ∩ [c, d] . 2

2.1 The Real Number Line

27

Lemma 2.1.2. Let I, J be closed, bounded intervals in Q such that ¬ (I ∩ J = ∅) . Then there exists r ∈ Q such that r ∈ I ∩ J. Proof. Let I = [a, b] and J = [c, d] . If b < c, then I ∩ J = ∅, a contradiction. Hence c b. Likewise, a d. It remains to apply Lemma 2.1.1. 2

Taken with the equality and inequality we have deﬁned above, the collection of real numbers forms a set: the real line R. Let x, y be real numbers. We say that x is greater than y, and that y is less than x, if there exist (q, q ) ∈ x and (r, r ) ∈ y such that r < q; we then write x > y or, equivalently, y < x. On the other hand, we say that x is greater than or equal to y, and that y is less than or equal to x, if for all (q, q ) ∈ x and all (r, r ) ∈ y we have q r; we then write x y or, equivalently, y x. We write, for example, x > y to indicate that ¬ (x > y) , and x y to indicate that ¬ (x y) . Clearly, x x and x > x. Moreover, by Lemma 2.1.1, x = y if and only if both x y and y x; and x = y if and only if either x > y or else x < y. Lemma 2.1.3. If x > y, then y > x. Proof. Since x > y, there exist (q, q ) ∈ x and (r, r ) ∈ y such that r < q. If also y > x, then there exist (s, s ) ∈ x and (t, t ) ∈ y such that s < t. By the deﬁning conditions on real numbers, there exist rational numbers a, b such that a ∈ [q, q ] ∩ [s, s ] and b ∈ [r, r ] ∩ [t, t ] . But then a s < t b r < q a, a contradiction.

2

Lemma 2.1.4. x y if and only if y > x. Proof. By deﬁnition, x y if and only if for all (q, q ) ∈ x and all (r, r ) ∈ y we 2 have q r, which occurs precisely when ¬(y > x). Lemma 2.1.5. If x > y, then x y. Proof. By Lemma 2.1.3, y > x. The result follows from Lemma 2.1.4.

2

Lemma 2.1.6. If x > 0, then there exists a positive integer n such that x > 1/n.

28

2 Techniques of Elementary Analysis

Proof. First pick (q, q ) ∈ x such that 0 < q. Then choose a positive integer n such that q > 1/n. The deﬁnition of “greater than” ensures that x > 1/n. 2 From time to time we shall revisit some of our earlier Brouwerian examples in the light of our formal deﬁnitions of real numbers and their properties. Proposition 2.1.7. The statement ∀x, y ∈ R (¬ (x y) =⇒ y > x)

(2.1)

implies Markov’s principle. Proof. Assume (2.1). Let (an )n1 be an increasing binary sequence such that ¬∀n (an = 0) , and deﬁne a real number by 1 1 1 , : an = 0 ∪ : an = 1 − an−1 . x= 0, n n n

(2.2)

Then ¬ (0 x): for if 0 q for all (q, q ) ∈ x, then an = 0 for all n, a contradiction. It follows from (2.1) that

1 1x > 0 and hence that there exists (q, q ) ∈ x such that 2 q > 0. Then (q, q ) = n , n for a (unique) n such that an = 1 − an−1 . Proposition 2.1.8. The relation > is cotransitive: If a < b, then for all x ∈ R either a < x or x < b. Proof. There exist (q, q ) ∈ a and (r, r ) ∈ b such that q < r. Given a real number x, we can ﬁnd (s, s ) ∈ x such that s − s < r − q . If s < r, then x < b; if s r, 2 then q < s and so a < x. Lemma 2.1.9. If x > y z or x y > z, then x > z and x z. Proof. Assume, for example, that x > y z. By Proposition 2.1.8, either x > z or z > y. Since the latter is ruled out by Lemma 2.1.4, we have x > z and therefore, by Lemma 2.1.5, x z. 2 Lemma 2.1.10. If (q, q ) ∈ x, then q x q . Proof. By Lemma 2.1.1, for each (r, r ) ∈ x, since the rational intervals [r, r ] and [q, q ] intersect, r q and q r. The desired conclusion now follows from the deﬁnition of the relation . 2

2.1 The Real Number Line

29

Lemma 2.1.11. For each real number x there exist rational numbers q, q such that q < x < q . Proof. Let (r, r ) be any element of x. By Lemma 2.1.10, r x r . Choosing q, q in Q with q < r r < q , we see from the deﬁnition of the relation < that 2 q < x < q .

The following two propositions show that the classical law of trichotomy does not hold constructively, even in the weak form discussed in Proposition 2.1.13. Proposition 2.1.12. The statement ∀x ∈ R (x 0 =⇒ x > 0 ∨ x = 0) implies LPO. Proof. Given an increasing binary sequence (an )n1 , deﬁne the real number x as at (2.2). It is routine to check that x 0; that if x = 0, then an = 0 for all n; and 2 that if x > 0, then there exists n such that an = 1. Proposition 2.1.13. The statement ∀x ∈ R (x 0 ∨ x 0) implies LLPO. Proof. Let (an )n1 be a binary sequence with at most one term equal to 1, and deﬁne a real number by 1 1 n 1 n 1 , (−1) : ∀k n (ak = 0) ∪ : an = 1 . (−1) x= − , n n n n If x 0, then it is impossible that an = 1 for an odd n, so an = 0 for all odd n. 2 Likewise, if x 0, then an = 0 for all even n. Lemma 2.1.14. If x < y, then there exists s ∈ Q such that x < s < y. Proof. Since x < y, there exist (q, q ) ∈ x and (r, r ) ∈ y such that q < r. Then the rational number 1 1 (q + r) , (q + r) s= 2 2 has the desired property.

2

30

2 Techniques of Elementary Analysis

The maximum of two real numbers x and y is the set max {x, y} of all rational pairs of the form (max {q, r} , max {q , r }) where (q, q ) ∈ x, (r, r ) ∈ y, and, for example, max {q, r} is the maximum in Q of the two rational numbers q, r. The minimum of x and y is the set min {x, y} of all rational pairs of the form (min {q, r} , min {q , r }) where (q, q ) ∈ x, (r, r ) ∈ y, and min {q , r } is the minimum in Q of the two rational numbers q , r . The maximum, max {x1 , . . . , xn } , and minimum, min {x1 , . . . , xn } , of ﬁnitely many real numbers are deﬁned analogously. We deﬁne the negative of a real number x to be

− x = (q, q ) ∈ Q × Q : − q ,− q ∈ x , where, for a rational number q, the expression − q denotes the negative deﬁned in the usual way. The absolute value of the real number x is the set |x| = max − x, x . It is not hard to verify such properties as the following: −

x, and |x| are real numbers.

•

max{x, y}, min{x, y},

•

max {x, y} < z if and only if x < z and y < z.

• max {x, y} > z if and only if either x > z or y > z. • min {x, y} < z if and only if either x < z or y < z. •

x < y if and only if

• max {x, y} =

−

−

y 0; whence x = 1 and P holds. In the other case we must have ¬P. We deﬁne a set S of real numbers to be upper order located if for all rational numbers a, b with a < b, either x b for all x in S or else there exists x ∈ S with a < x. We can now state the constructive least-upper-bound principle, which is also known as the Dedekind (order) completeness of R. Theorem 2.1.18. Let S be an inhabited set of real numbers that is both bounded above and upper order located. Then the least upper bound of S exists. Proof. Let B be the set of upper bounds for S, and deﬁne ξ = {(q, q ) ∈ Q × Q : ∃s ∈ S ∃b ∈ B (q s b q )} . Taken with Lemmas 2.1.10 and 2.1.9, the hypotheses ensure that ξ is inhabited. If (q, q ) and (r, r ) belong to ξ, then there exist s1 , s2 ∈ S and b1 , b2 ∈ B such that q s1 b1 q and r s2 b2 r . Then s1 b2 , so q b2 r ; similarly, r q . By Lemma 2.1.1, there exists a rational number in [q, q ] ∩ [r, r ] . To complete the proof that ξ is a real number, we show that for each rational ε > 0 there exists (q, q ) ∈ ξ with q − q < ε. To this end, ﬁx (a, a ) in ξ. If a = a , then there is nothing to prove; so we may assume that a < a . Construct rational numbers a0 = a < a1 < a2 < · · · < an = a such that ai − ai−1 < ε/2 for 1 i n. Since S is upper order located, either a2 ∈ B or else a1 < s for some element s of S. In the ﬁrst case, (a0 , a2 ) ∈ ξ and a2 − a0 < ε. In the second, either a3 ∈ B

2.1 The Real Number Line

33

and therefore (a1 , a3 ) ∈ ξ and a3 − a1 < ε; or else a2 < s for some element s of X. Carrying on in this way, since an ∈ B we can be sure of ﬁnding k < n − 1 such that (ak , ak+2 ) ∈ ξ and ak+2 − ak < ε. Thus ξ is indeed a real number. To show that ξ is an upper bound for S, consider any (q, q ) in ξ and any s in S. There exists b ∈ B such that q b q . For any (r, r ) in s we have r s, by Lemma 2.1.10, and therefore r b; whence r q . It follows that s ξ. Finally, if x < ξ, then we can ﬁnd (q, q ) ∈ ξ and (r, r ) ∈ x such that q > r . It follows from Lemma 2.1.10 that x r < q. By deﬁnition of ξ, there exists s ∈ S with q s; whence x < s, by Lemma 2.1.9. This completes the proof that ξ is the least upper bound for S. 2

Let S be a set of real numbers. We say that a real number b is a lower bound of/for S if b s for all s ∈ S; and that b is the (perforce unique) greatest lower bound of S if it is a lower bound and for each x > b there exists s ∈ S such that x > s. In the latter event we also call b the inﬁmum of S and we denote it by inf S. Corollary 2.1.19. Let S be an inhabited set of real numbers that is bounded below and is lower order located in the following sense: for all rational numbers a, b with a < b, either a is a lower bound for S or else there exists s ∈ S with s < b. Then the greatest lower bound of S exists. Proof. Apply Theorem 2.1.18 to the set T = −s : s ∈ S , which is inhabited, bounded above, and upper order located, to construct its supre2 mum b. Then − b is the inﬁmum of S.

We now want to introduce the arithmetic operations on real numbers. Given real numbers x and y, we need to ﬁnd the appropriate ways of combining the rational pairs (representing intervals) that constitute the real number x with those that constitute y, in order to create the rational pairs that represent x ◦ y, where ◦ stands for any of the operations +, −, ×, ÷. We begin with the easy deﬁnitions of + and −, for the moment leaving aside the more complicated ones for × and ÷. We deﬁne the sum x + y and diﬀerence x − y of the real numbers x, y to be, respectively, x + y = {(s, s ) : ∃ (q, q ) ∈ x ∃ (r, r ) ∈ y (s = q + r ∧ s = q + r )} , x − y = {(s, s ) : ∃ (q, q ) ∈ x ∃ (r, r ) ∈ y (s = q − r ∧ s = q − r )} . Let us verify, for example, that x+y is a real number. Let (q1 , q1 ) ∈ x and (r1 , r1 ) ∈ y. Certainly, q1 + r1 q1 + r1 . Moreover, given a positive rational number ε, we can arrange that q1 −q1 < ε/2 and r1 −r1 < ε/2, so (q1 + r1 )−(q1 + r1 ) < ε. It remains

34

2 Techniques of Elementary Analysis

to show that if also (q2 , q2 ) ∈ x and (r2 , r2 ) ∈ y, then the intervals [qi + ri , qi + ri ] (i = 1, 2) in Q intersect. This is easy: there exist rational numbers ξ, η such that qi ξ qi and ri η ri (i = 1, 2) ; whence qi + ri ξ + η qi + ri

(i = 1, 2) .

Thus x + y is a real number. Note that if x < y, then x + z < y + z for all real z; and that x − y = x+ (− y) . In view of the latter, we adopt the normal convention of writing −x instead of − x. Lemma 2.1.20. For each real number x there exists a positive integer N such that max{|q| , |q |} < N whenever (q, q ) ∈ x and q − q < 1. Proof. With n as in Lemma 2.1.15, set N = n + 1. If (q, q ) ∈ x and q − q < 1, then, using Lemma 2.1.10, we obtain q x |x| < n < N and therefore

q < q + 1 < n + 1 = N.

Since (−q , −q) ∈ −x and |−x| = |x| , we likewise have −q < N and −q < N . 2 Hence |q| < N and |q | < N, so max {|q| , |q |} < N. We next deﬁne the product x × y, also written x · y or xy, to be the set of all rational pairs (s, s ) such that there exist (q, q ) ∈ x and (r, r ) ∈ y with s = min {qr, qr , q r, q r } , s = max {qr, qr , q r, q r } . Certainly, s s . Using Lemma 2.1.20, compute a positive integer N such that if (q, q ) ∈ x and q − q < 1, then max {|q| , |q |} < N, and such that if (r, r ) ∈ y and r − r < 1, then max {|r| , |r |} < N. Given a rational ε with 0 < ε < 1, choose (q, q ) ∈ x and (r, r ) ∈ y such that q − q < ε/2N and r − r < ε/2N. Routine rational arithmetic calculations show that s − s < (q − q) max {|r| , |r |} + (r − r) max {|q| , |q |} < ε. For i = 1, 2 let (qi , qi ) ∈ x, (ri , ri ) ∈ y, and si = min {qi ri , qi ri , qi ri , qi ri } , si = max {qi ri , qi ri , qi ri , qi ri } . Pick ξ in [q1 , q1 ] ∩ [q2 , q2 ] and η in [r1 , r1 ] ∩ [r2 , r2 ] ; it is easy to verify that ξη ∈ [s1 , s1 ] ∩ [s2 , s2 ] . This completes the proof that xy is a real number. When dealing with division, we consider two real numbers x, y with y = 0. In the case y > 0 we construct a rational pair (s, s ) in the quotient x/y (also written

2.1 The Real Number Line

35

x y)

as follows. We pick (q, q ) ∈ x and (r, r ) ∈ y such that r > 0. If q 0, we set s = q/r , s = q /r; if q < 0, we set s = q/r, s = q /r . In the case y < 0 we deﬁne x/y to be −((−x)/y). We leave as an exercise the details that x/y is indeed a real number, and that x(1/x) = 1. As classically, we say that a sequence (xn )n1 of real numbers converges to a real number x∞ , called the limit of the sequence, if ∀ε > 0 ∃N ∀n N (|x∞ − xn | < ε) . We then write xn −→ x∞ as n −→ ∞ or x∞ = lim xn . n→∞

The uniqueness of the limit, and the basic algebraic properties of sequences of real numbers, will be assumed without proof, since the proofs are virtually the same as their classical counterparts. However, we should note that the monotone convergence theorem for sequences, Every increasing sequence in R that is bounded above (that is, whose terms form a set that is bounded above) converges, implies LPO. By a Cauchy sequence of real numbers we mean a sequence (xn )n1 such that ∀ε > 0 ∃N ∀m, n N (|xm − xn | < ε) . Every convergent sequence of real numbers is a Cauchy sequence. We now counter the second major misconception about the constructive nature of R by establishing its so-called (Cauchy) completeness. Theorem 2.1.21. Every Cauchy sequence of real numbers converges to a real number. Proof. Let (xn )n1 be a Cauchy sequence of real numbers, and, using countable choice, compute a function k nk from N+ to N+ such that

∀k ∀m, n nk |xm − xn | < 2−k . Again using countable choice, construct a sequence ((qk , qk ))k1 such that

∀k (qk , qk ) ∈ xnk ∧ qk − qk < 2−k . Deﬁne rational numbers rk = qk − 2−k ,

rk = qk + 2−k .

36

2 Techniques of Elementary Analysis

Then for all n nk , rk xnk − 2−k < xn < xnk + 2−k rk .

(2.3)

It follows that for all j k, xnj ∈ rj , rj ∩ [rk , rk ] . Since rk − rk < 2−k+2 −→ 0 as k −→ ∞, we conclude that x∞ = {(rk , rk ) : k 1} is a real number. From (2.3) we have xn ∈ [rk , rk ] for all n nk . It follows that

∀k ∀n nk |xn − x∞ | rk − rk < 2−k+2 . Noting Lemma 2.1.6, we see that xn −→ x∞ as n −→ ∞.

2

We shall see in later chapters that completeness is often used in constructive mathematics to prove propositions that are immediate consequences of omniscience principles. For the remainder of this book we shall drop the use of boldface type to denote real numbers; it has served its purpose to signal a distinction between a rational number and a real number (a set of special pairs of rational numbers). We shall also assume, without further comment, basic properties of real numbers—for example, x2 0 for all real x—that can easily be deduced from the foregoing results. The complex plane C consists of all complex numbers—ordered pairs (x, y) of real numbers—with addition and multiplication deﬁned by (x, y) + (x , y ) = (x + x , y + y ) , (x, y) × (x , y ) = (xx − yy , xy + x y) . The equality and inequality on C are deﬁned by (x, y) = (x , y ) ⇐⇒ x = x ∧ y = y (x, y) = (x , y ) ⇐⇒ x = x ∨ y = y . We embed R as a subset of C in the usual way by identifying the real number x with the complex number (x, 0) . The pair i = (0, 1) then has the special property that i2 = −1; and every complex number z = (x, y) can be written in the form x + iy, with real part Re z = x and imaginary part Im z = y. The complex conjugate of z = x + iy is the number x − iy, denoted by z ∗ . We shall assume basic deﬁnitions and properties of C, such as its completeness, as they are needed. The same goes for the Euclidean spaces Rn and Cn , which are now deﬁned in the standard ways.

2.2 Metric Spaces

37

2.2 Metric Spaces Since we are assuming some familiarity with the classical theory of metric spaces, in the following we shall emphasise diﬀerences between the classical and the constructive theory, as well as those constructive properties, such as locatedness, that play no role in classical analysis. We normally denote the metric on a set X by ρ. The deﬁnitions of such notions as metric, (metric) subspace, ball, interior, and cluster point are exactly as in classical analysis; see, for example, [47]. The interior of a subset S of a metric space X is denoted by S ◦ and is a subset of S; if S = S ◦ , then S is said to be open. The set S of cluster points of S is called the closure of S (in X), and contains S; if S = S, then S is said to be closed in X. If S = X, then S is dense in in X. This happens if and only if for each x ∈ X and each ε > 0 there exists s ∈ S such that ρ (x, s) < ε. We denote the open and closed balls in X with centre a and radius r > 0 by B(a, r) and B(a, r) respectively. The inequality on a metric space is deﬁned by x = y ⇐⇒ ρ(x, y) > 0. If X1 , . . . , Xn are metric spaces, with metrics ρ1 , . . . , ρn respectively, then the Cartesian product X = X1 × · · · × Xn is a metric space relative to the product metric ρ = ρ 1 + · · · + ρn . We then call (X, ρ) the product of the metric spaces X1 , . . . , Xn . As in the classical theory, both X and ∅ are open, arbitrary unions of open sets are open, and ﬁnite intersections of open sets are open; in other words, the open sets form a topology on X. Likewise, both X and ∅ are closed, and arbitrary intersections of closed sets are closed. But we cannot prove that the union of two closed sets is closed, even in the case X = R and the two closed sets are closed intervals: both the intervals [−1, 0] and [0, 1] are closed in R, and their union is dense in [−1, 1]; but if that union is closed, then we have ∀x ∈ R (x 0 ∨ x 0) , a proposition equivalent to LLPO. Proposition 2.2.1. The logical complement of an open subset S of X is closed in X, and coincides with the complement of S. Proof. Given a cluster point a of ¬S and any point x of S, choose, in turn, r > 0 such that B(x, r) ⊂ S, and y ∈ ¬S such that ρ(a, y) < r/2. If ρ(a, x) < r/2, then ρ(x, y) ρ(a, x) + ρ(a, y) < r, so y ∈ S, which is absurd. Hence, by Proposition 2.1.8, ρ(a, x) > 0. Since x ∈ S is arbitrary, it follows that a ∈ ∼S ⊂ ¬S. Thus ¬S is closed, and ¬S ⊂ ∼S ⊂ ¬S;

38

2 Techniques of Elementary Analysis

whence ¬S = ∼S.

2

In contrast to the classical situation, we cannot prove constructively that the complement of a closed subset of R is open (Exercise 9). The classical deﬁnitions of convergent sequence, Cauchy sequence, and complete metric space carry over unchanged into the constructive setting, as they did in the special case of the metric space R. A complete subset of a metric space X is closed in X; and a closed subset of a complete space is complete. Moreover, the product of ﬁnitely many complete spaces is complete. We now introduce the ﬁrst of two major themes of metric space theory. Given ε > 0, by an ε-approximation to a subset S of a metric space X we mean an inhabited subset T of S such that for each s ∈ S there exists t ∈ T with ρ (s, t) < ε. If for each ε > 0 there exists a ﬁnite ε-approximation to S, then we say that S is totally bounded. The closure of a totally bounded subset of X is totally bounded. If a subset S of X contains a dense totally bounded set, then S itself is totally bounded. The product of ﬁnitely many totally bounded spaces is totally bounded. There are at least two reasons why total boundedness is so important in constructive analysis. The ﬁrst is that with its help we can compute suprema and inﬁma in many important situations. The second is that, coupled with completeness, total boundedness gives the only one of three classically equivalent notions of compactness that can be used in constructive analysis. The next lemma enables us to replace “ﬁnite” by the weaker “ﬁnitely enumerable” in the deﬁnition of total boundedness. Lemma 2.2.2. Let ε be a positive number, and X a metric space with a ﬁnitely enumerable ε-approximation. Then X has a ﬁnite η-approximation for each η > ε. Proof. Let {x1 , . . . , xn } be an ε-approximation to X, and let η > ε. We may assume that n > 1. Either ρ(xi , xj ) > 0 whenever i = j, or ρ(xi , xj ) < η − ε for some pair of distinct indices i, j. In the former case, {x1 , . . . , xn } is a ﬁnite η-approximation. In the latter case we can delete xj from {x1 , . . . , xn }, to obtain a ﬁnitely enumerable ε -approximation with ε = ε + ρ(xi , xj ) < η. Applying this argument at most n − 1 times, we arrive at a ﬁnite η-approximation to X. 2 Corollary 2.2.3. A metric space X is totally bounded if and only if for each ε > 0 there exists a ﬁnitely enumerable ε-approximation to X. Proof. This is an immediate consequence of Lemma 2.2.2.

2

A metric space X is separable if it has a countable dense subset S; an enumeration (xn )n1 of S is then called a dense sequence in X. The real line is separable,

2.2 Metric Spaces

39

since, by Lemma 2.1.10 and the deﬁnition of “real number”, it has Q as a countable dense subset. Proposition 2.2.4. A totally bounded metric space is separable. Proof. If, for each positive integer n, Fn is a ﬁnite 1/n-approximation to the metric Fn is a countable dense subset of X. 2 space X, then n1

We now have what is among the most useful of all results about totally bounded sets. Proposition 2.2.5. If S is a totally bounded subset of R, then sup S and inf S exist. Proof. We ﬁrst consider the case where S = {x1 , . . . , xn } is ﬁnitely enumerable. Given real numbers a, b with a < b, we apply Proposition 2.1.8 repeatedly, to prove that either xk < b for each k or else there exists j such that xj > a. It follows from the constructive least-upper-bound principle that sup S exists. Now consider the general case. Again let a, b be real numbers with a < b, but this time write ε = (b − a)/2 and construct a ﬁnite ε-approximation {x1 , . . . , xn } to S. By the ﬁrst part of the proof, σ = sup {x1 , . . . , xn } exists. By Proposition 2.1.8, either σ > a or σ < a + ε. In the ﬁrst case there exists j such that xj > a. In the second, consider any x ∈ S. Choosing j such that |x − xj | < ε, we have x xj + |x − xj | < σ + ε < a + 2ε = b. So in this case, b is an upper bound for S. It follows from Theorem 2.1.18 that sup S exists. Similar arguments show that inf S exists. 2 A mapping f : X −→ Y between metric spaces is •

sequentially continuous at x ∈ X if for each sequence (xn )n1 that converges in X to x, the sequence (f (xn ))n1 converges in Y to f (x);

• sequentially continuous if it is sequentially continuous at each point of X; • continuous at x ∈ X if for each ε > 0 there exists δ > 0 such that ρ (f (x), f (x )) < ε for all x ∈ X with ρ (x, x ) < δ; • continuous if it is continuous at each point of X.

40

2 Techniques of Elementary Analysis

Although sequential continuity is classically equivalent to continuity, the most we can say constructively is that the latter notion implies the former. For constructive practice it is frequently necessary to assume the following stronger continuity property. We say that f : X −→ Y is uniformly continuous (on X) if for each ε > 0 there exists δ > 0 such that for all x, x ∈ X, if ρ (x, x ) < δ, then ρ (f (x), f (x )) < ε. Clearly, uniform continuity implies continuity. Although in CLASS and in INT every continuous mapping from the bounded closed interval [0, 1] into R is uniformly continuous, since this is not the case in RUSS we cannot expect to prove it in BISH; see Chapter 3 of [34]. Proposition 2.2.6. If X is a totally bounded space, and f a uniformly continuous mapping of X into a metric space, then f (X) is totally bounded. Proof. Given ε > 0, compute δ > 0 such that if x, y ∈ X and ρ(x, y) < δ, then ρ(f (x), f (y)) < ε. There exists a ﬁnite δ-approximation {x1 , . . . , xn } to X. For each x ∈ X pick i such that ρ (x, xi ) < δ. Then ρ(f (x), f (xi )) < ε. Hence 2 {f (xi ), . . . , f (xn )} is a ﬁnitely enumerable ε-approximation to f (X). Corollary 2.2.7. Let f be a uniformly continuous mapping of a totally bounded space into R. Then ran(f ) is bounded, and the supremum sup f (x) = sup {y : y ∈ ran(f )}

x∈X

and inﬁmum inf f (x) = inf {y : y ∈ ran(f )}

x∈X

of f exist. Proof. This is an immediate consequence of Propositions 2.2.6 and 2.2.5.

2

Corollary 2.2.8. If X is a totally bounded metric space, then its diameter, diam(X) = sup {ρ(x, y) : x, y ∈ X} , exists. Proof. Since the mapping ρ on the totally bounded product space X × X is uniformly continuous, the desired conclusion follows from Corollary 2.2.7. 2

We now introduce a second major theme in constructive metric space theory. An inhabited subset S of a metric space X is located in X if for each x ∈ X the distance

2.2 Metric Spaces

41

ρ (x, S) = inf {ρ(x, s) : s ∈ S} exists. We cannot prove the proposition “every inhabited subset of R is located” constructively, since it implies LEM. To see this, let P be any syntactically correct proposition, and S = {0} ∪ {x ∈ R : (x = 1) ∧ P } . If S is located, then either ρ(1, S) > 0 or ρ(1, S) < 1. In the ﬁrst case we have ¬P. In the second, choosing s ∈ S such that ρ(1, s) < 1, we see that s ∈ / {0} , so s = 1 and P holds. It is hard to overemphasise the importance of locatedness in constructive analysis. To illustrate, we anticipate later sections and chapters by pointing out that the norm of a bounded linear functional on a normed space exists if and only if the kernel of the functional is located, and that locatedness plays a vital role in the separation and Hahn–Banach extension theorems. It is only when we work constructively that the signiﬁcance of locatedness (a property that holds automatically under classical logic) for the proofs of many existence theorems is revealed. A subset S of a metric space is located if and only if its closure is located; in fact, for each x ∈ X we have ρ(x, S) = ρ(x, S) if either side of this equation exists. Note that even if a subset S of a metric space X is located, we may not be able to prove that its metric complement −S = {x ∈ X : ρ(x, S) > 0} is located (Exercise 11). Proposition 2.2.9. A totally bounded subset of a metric space is located. Proof. Let S be totally bounded in the metric space X. For each x ∈ X the mapping s ρ(x, s) is uniformly continuous on S. It follows from Corollary 2.2.7 that 2 inf {ρ(x, s) : s ∈ S} exists.

The next two proofs use a technique that will reappear from time to time. We have a ﬁnitely enumerable set {x1 , . . . , xn } of a metric space X, a located subset S of X, and positive numbers a, b with a < b. We write {1, . . . , n} as a union of subsets P, Q such that i ∈ P =⇒ ρ(xi , S) < b, i ∈ Q =⇒ ρ(xi , S) > a. By concentrating on one of those sets, we are able to reach our desired conclusion.

42

2 Techniques of Elementary Analysis

Proposition 2.2.10. A located subset of a totally bounded metric space is totally bounded. Proof. Let S be a located subset of a totally bounded metric space X, and let ε > 0. There exists a ﬁnite ε/3-approximation {x1 , . . . , xn } to X. Using Proposition 2.1.8, write {1, . . . , n} as a union of two sets P and Q, where ρ(xi , S) < 2ε/3 if i ∈ P, and ρ(xi , S) > ε/3 if i ∈ Q. For each i ∈ P there exists si ∈ S such that ρ(xi , si ) < 2ε/3. / Q; Given s ∈ S, choose i such that ρ(s, xi ) < ε/3. Then ρ (xi , S) < ε/3, so i ∈ whence i ∈ P and therefore ρ(s, si ) ρ(s, xi ) + ρ(xi , si )

2−n+2 r

if i ∈ A, if i ∈ B.

Then Fn+1 = {xi : i ∈ A} clearly satisﬁes the appropriate instance of (b). Let x be any point of B(x0 , r). By the induction hypothesis, there exists y in Fn with ρ(x, y) < 2−n+1 r. Choosing i in {1, . . . , N } such that ρ(x, xi ) < 2−n r, we have ρ(xi , Fn ) ρ(xi , y) ρ(x, xi ) + ρ(x, y) < 2−n+2 r.

2.2 Metric Spaces

43

Thus i ∈ / B, so i ∈ A and therefore xi ∈ Fn+1 . Since ρ(x, xi ) < 2−(n+1)+1 r, the set Fn+1 therefore satisﬁes the appropriate instance of (a). This completes the inductive construction. Fn in X. We see from (a) that B(x0 , r) ⊂ K. On Let K be the closure of n1

the other hand, given x ∈ K and a positive integer n, we can ﬁnd m and y ∈ Fm such that ρ(x, y) < 2−n+4 r. If m n, then by (b), there exist points ym = y, ym−1 ∈ Fm−1 , . . . , yn ∈ Fn such that ρ(yi+1 , yi ) < 2−i+3 r for n i m − 1. Thus ρ(y, Fn ) ρ (y, yn )

m−1 i=n

ρ(yi+1 , yi )

0 there exist totally bounded sets K1 , . . . , Kn , each of diameter less than or equal to ε, such that n Ki . X= i=1

Proof. Given ε > 0, construct an ε/16-approximation {x1 , . . . , xn } to X. By Proposition 2.2.11, for each i in {1, . . . , n} there exists a closed, totally bounded set Ki n such that B (xi , ε/16) ⊂ Ki ⊂ B (xi , ε/2) . Clearly, X = Ki . Also, ρ(x, y) ε for all x, y in Ki , so diam (Ki ) ε.

i=1

2

A property P, applicable to certain elements of a set S, is said to hold for all but countably many x in S if there exists a sequence (xn )n1 in S such that P (x) holds whenever x ∈ S and x = xn for each n. The sequence (xn )n1 is then called the excluded sequence, and the elements x such that x = xn for each n are said to be admissible, for the property P. Theorem 2.2.13. Let f be a uniformly continuous mapping of a totally bounded metric space X into R. Then for all but countably many r ∈ R the set X(f, r) = {x ∈ X : f (x) r} is either totally bounded or empty.

44

2 Techniques of Elementary Analysis

Proof. By Corollary 2.2.12, for each positive integer k there exist a positive integer nk and totally bounded sets Xk,j (1 j nk ), each of diameter less than 1/k, whose union is X. Deﬁne the excluded sequence (rn )n1 to be an enumeration of the real numbers ck,j = inf {f (x) : x ∈ Xk,j }

(k 1; 1 j nk ).

(Note that the numbers ck,j exist in view of Corollary 2.2.7.) Let r = rn for each n. If ck,j > r for 1 j nk , then X (f, r) = ∅. So we may assume that there exists ν nk such that ck,j < r for 1 j ν, and ck,j > r for ν < j nk . For each j ν choose xk,j ∈ Xk,j such that f (xk,j ) < r. For all such k, j and all x ∈ Xk,j we have 1 (2.5) ρ(x, xk,j ) diam (Xk,j ) < . k Consider any x ∈ X(f, r) and any positive integer k. Choose j with 1 j nk and x ∈ Xk,j . Then ck,j f (x) r, so ck,j < r (since ck,j = r) and therefore j ν. By (2.5), ρ(x, xk,j ) < 1/k. Hence {xk,j : 1 j ν} is a ﬁnitely enumerable 1/k-approximation to X(f, r). Since k is arbitrary, it follows that X(f, r) is totally bounded. 2

A complete, totally bounded metric space X is said to be compact. The bounded closed intervals [a, b] in R, and the closed balls in C, are compact; the product of ﬁnitely many compact spaces is compact. A compact subset of a metric space is both closed and (by Proposition 2.2.9) located; and a closed, located subset of a compact space is compact (see Proposition 2.2.10). Corollary 2.2.14. Under the hypotheses of Theorem 2.2.13, if X is compact, then X(f, r) is either compact or empty for all but countably many r ∈ R. Proof. In view of Theorem 2.2.13, it suﬃces to note that, by the uniform continuity of f , the set X(f, r) is closed in X, and therefore complete, for each admissible r. 2

We say that two subsets S, T of a metric space X are apart from each other, and we write S T, if there exists r > 0 such that ρ(s, t) r for all s ∈ S and t ∈ T. A mapping f : X −→ Y between metric spaces is said to be strongly continuous if for all subsets S, T of X with f (S) f (T ) we have S T. It is simple to show that uniform continuity implies strong continuity. The following partial converse will be used shortly. Proposition 2.2.15. Let f be a strongly continuous mapping of a metric space X onto a totally bounded metric space Y. Then f is uniformly continuous.

2.2 Metric Spaces

45

Proof. Given ε > 0, construct an ε/8-approximation {f (x1 ), . . . , f (xn )} to Y , and deﬁne

Yi = B f (xi ), 8ε , Xi = f −1 (Yi ). Write {(i, j) : 1 i, j n} as the union of subsets P and Q such that 3ε , 4 ε (i, j) ∈ Q =⇒ ρ (f (xi ), f (xj )) > . 2 (i, j) ∈ P =⇒ ρ (f (xi ), f (xj ))

ε . 4

Hence Yi Yj . The strong continuity of f now yields Xi Xj ; whence there exists ri,j > 0 such that ρ (x, x ) ri,j for all x ∈ Xi and x ∈ Xj . Let δ = min {ri,j : (i, j) ∈ Q} > 0, and consider points x, x of X with ρ (x, x ) < δ. Choose i, j such that f (x) ∈ Yi and f (y) ∈ Yj . If (i, j) ∈ Q, then ρ (x, x ) ri,j δ, a contradiction. Hence (i, j) ∈ P ; 2 so ρ (f (xi ), f (xj )) < 3ε/4 and therefore ρ (f (x), f (x )) < ε. A mapping f : X −→ Y between metric spaces is said to be strongly injective, or a strong injection, if it is one-one and the inverse mapping is strongly continuous; in other words, if for all subsets S, T of X with S T we have f (S) f (T ). Every strongly injective mapping is injective. The following is a constructive substitute for the classical theorem that a continuous one-one mapping of a compact metric space into a metric space has compact range and a continuous inverse (a theorem that is false in RUSS). Proposition 2.2.16. Let f be a uniformly continuous strong injection of a compact metric space X into a metric space Y. Then f −1 is uniformly continuous and strongly injective on f (X), and f (X) is compact. Proof. The mapping f −1 : f (X) −→ X is strongly continuous, and X is totally bounded. Hence, by Proposition 2.2.15, f −1 is uniformly continuous on f (X). Since f is uniformly continuous and hence strongly continuous, f −1 is also strongly injective. By Proposition 2.2.6, f (X) is totally bounded; so it remains to prove that it is complete. Accordingly, let (f (xn ))n1 be a Cauchy sequence in f (X). Using the uniform continuity of f −1 , we see that (xn )n1 is a Cauchy sequence in X. Since X is complete, (xn )n1 converges to a limit x in X. Finally, since f is continuous, 2 (f (xn ))n1 converges to f (x) in f (X).

46

2 Techniques of Elementary Analysis

Let K(X) denote the set of all compact subsets of a compact metric space X. For all A, B ∈ K(X), since the map x ρ (x, B) is deﬁned (B being located) and is uniformly continuous on A, the number mA,B = inf {ρ (x, B) : x ∈ A} exists, by Corollary 2.2.7. Likewise, mB,A exists. It is left as an exercise to show that the mapping (A, B) max {mA,B , mB,A } from K(X) × K(X) into the nonnegative real line R0+ = {x ∈ R : x 0} is a metric—we call it the Hausdorﬀ metric—with respect to which the space K(X) is compact. An inhabited metric space X is said to be locally totally bounded if each bounded subset of X is contained in a totally bounded subset; locally compact if it is both locally totally bounded and complete. Every compact space is locally compact. The spaces R and C, and the product spaces Rn and Cn , are locally compact. A metric space X is locally compact if and only if every bounded subset of X is contained in a compact set. The following lemma prepares the way for our next theorem, which deals with certain fundamental properties of a locally totally bounded space. Lemma 2.2.17. Let Y be a located subset of a metric space X, and T a totally bounded subset of X that intersects Y. Then there exists a totally bounded set S such that T ∩ Y ⊂ S ⊂ Y. Proof. Using Theorem 2.2.13, construct a sequence (αn )n1 such that for each positive integer n, 0 < αn+1 < αn < 1/n and Tn = {x ∈ T : ρ(x, Y ) αn } is totally bounded. Let Fn be a ﬁnite 1/n-approximation to Tn , and construct a mapping φn : Fn −→ Y such that ρ(x, φn (x)) < 1/n for each x ∈ Fn ; then set Sn = {φn (x) : x ∈ Fn } .

Sn in Y. We show that SN is a 3/N -approximation to Let S be the closure of n1 Sk . Let k N and y ∈ Sk . Then y = φk (x) ∈ Y for some x ∈ Fk . Since

kN

Fk ⊂ Tk ⊂ TN , there exists z ∈ FN such that ρ(x, z) < 1/N ; whence

2.3 Normed Linear Spaces

ρ(y, φN (z)) ρ(φk (x), x) + ρ(x, z) + ρ(z, φN (z)) < We now see that

N k=1

Sk is a 3/N -approximation to

n1

47

1 1 3 1 + + . k N N N

Sn . So

n1

Sn , and therefore

S, is totally bounded. If x ∈ T ∩ Y, then for each n we have x ∈ Tn ; so there exists z ∈ Fn such that ρ(x, z) < 1/n and therefore ρ(x, φn (z)) < 2/n. Thus ρ(x, S) < 2/n for each positive integer n. Since S is closed in Y , it follows that x belongs to S. 2

Proposition 2.2.18. Let Y be an inhabited subset of a metric space X. (a) If Y is locally totally bounded, then it is located. (b) If X is locally totally bounded and Y is located, then Y is locally totally bounded. Proof. Assume ﬁrst that Y is locally totally bounded. Let y0 ∈ Y and x ∈ X. The set B = {y ∈ Y : ρ(y, y0 ) 2ρ(x, y0 ) + 1} , being bounded in Y, is contained in a totally bounded subset K of Y. By Theorem 2.2.13, there exists r > 2ρ(x, y0 ) + 1 such that the set T = {y ∈ K : ρ(y, y0 ) r} is totally bounded and hence located. Then ρ(x, T ) ρ(x, y0 ) < ρ(x, y0 ) + 1, so for each y in Y, either ρ(x, y) > ρ(x, T ) or ρ(x, y) < ρ(x, y0 ) + 1. In the latter case, ρ(y, y0 ) ρ(x, y) + ρ(x, y0 ) 2ρ(x, y0 ) + 1 < r, so y ∈ T and therefore ρ(x, y) ρ(x, T ). It follows that ρ(x, Y ) exists and equals ρ(x, T ). Thus Y is located. Now assume that X is locally totally bounded and that Y is located in X. Let B be a bounded subset of Y. Then there exists a totally bounded subset T of X such that B ⊂ T. By Lemma 2.2.17, there exists a totally bounded subset S of Y such that T ∩Y ⊂ S and therefore B ⊂ S. Thus Y is locally totally bounded. 2

2.3 Normed Linear Spaces Let K stand for either R or C, and let X be a linear space over K. An inequality relation = on X is said to be compatible with the linear structure on X if for all x, y ∈ X and all t ∈ K,

48

2 Techniques of Elementary Analysis

x = y ⇐⇒ x − y = 0, x + y = 0 =⇒ x = 0 ∨ y = 0, and tx = 0 =⇒ t = 0 ∧ x = 0. It readily follows from the ﬁrst of these properties that x = y =⇒ ∀z ∈ X (x + z = y + z) . The requirement of compatibility between the inequality and the linear structure is a natural one and is automatically fulﬁlled by the denial inequality under classical logic. Constructively, it is certainly true that the denial inequality fulﬁls the ﬁrst of the three requirements for compatibility; but unless we accept Markov’s principle we cannot expect to prove that it satisﬁes the second or the third. In the case X = K the standard inequality is compatible with the linear structure. From now on, unless we state otherwise, when we refer to a linear space we mean one that is equipped with a compatible inequality. Let X be such a space. By a seminorm on X we mean a mapping : x x of X into R0+ such that for all x, y in X and all t in K, • x > 0 =⇒ x = 0, • tx = |t| x , and • x + y x + y . We call the pair (X, )—or, when no confusion is likely, just X itself—a seminormed (linear) space over K. We say that the seminormed space X is real or complex, depending on whether K is R or C. If x ∈ X and x > 0, then x is called a nonzero vector ; if x = 1, then x is called a unit vector . We call the seminormed space X nontrivial if it contains a nonzero vector. If the inequality on the seminormed space X satisﬁes x = 0 ⇐⇒ x > 0,

(2.6)

then we call a norm on X, x the norm of the vector x, and (X, )—or just X—a normed (linear) space over K. Note that every seminorm on a linear space X induces an inequality relation—namely, the one deﬁned by (2.6)—with respect to which becomes a norm. Let X be a normed space. Then the mapping (x, y) x − y of X × X into R is a metric on X, and is said to be associated with the norm on X. When we consider X as a metric space, it is understood that we are referring to the metric, usually denoted by ρ, associated with the given norm on X. Note that the inequality corresponding to the metric associated with the norm on X is just the original inequality on X. The unit ball of X is the closed ball with centre 0 and radius 1,

2.3 Normed Linear Spaces

49

BX = B X (0, 1) = B(0, 1) = {x ∈ X : x 1} , relative to that metric. This ball, like any open or closed ball in a normed space, is located. It is a simple consequence of the triangle inequality that | x − y| x − y for all vectors x, y in a normed space X. It follows from this that if a sequence (xn )n1 converges to a limit x in X, then the sequence (xn )n1 converges to x in R. Perhaps the simplest examples of a norm are the following ones on Kn : (x1 , . . . , xn ) max {|x1 | , . . . , |xn |} , 2 2 (x1 , . . . , xn ) |x1 | + · · · + |xn | , (x1 , . . . , xn ) |x1 | + · · · + |xn | . The second of these is called the Euclidean norm on Kn , which, when equipped with that norm, is called Euclidean n-space or simply Euclidean space. If S is a compact metric space, then for each uniformly continuous map f : S −→ K the sup norm f = sup {|f (x)| : x ∈ S} is well deﬁned, by Corollary 2.2.7. It is easy to see that the mapping f f is a norm on the space C(S, K) of uniformly continuous functions from S to K, taken with pointwise operations of addition and multiplication-by-scalars. We usually denote C(S, C) by C(S). Convergence and Cauchyness with respect to the sup norm on C(S, K), where S is a compact metric space, are called uniform convergence and uniform Cauchyness respectively. The space C(S, K) is an example of a Banach space: that is, a complete normed space. The standard proof of this in classical analysis is constructive, and is left as an exercise. The Euclidean space Kn is also a Banach space. Let X1 , X2 be normed spaces over K, and recall that the standard operations of addition and multiplication-by-scalars on the product vector space X = X1 × X2 are given by (x1 , x2 ) + (x1 , x2 ) = (x1 + x1 , x2 + x2 ) , t (x1 , x2 ) = (tx1 , tx2 ) , where xi ∈ Xi , xi ∈ Xi , and t ∈ K. It is easy to verify that the mapping (x1 , x2 ) max {x1 , x2 } is a norm on X, and that the metric associated with this norm is the product metric on X (considered as the product of the metric spaces X1 and X2 ). Taken with this

50

2 Techniques of Elementary Analysis

norm, which we call the product norm, X is known as the product of the normed spaces X1 and X2 . The product norm and the product space for a ﬁnite number of normed spaces are deﬁned analogously. The proof of the following is left as an exercise. Proposition 2.3.1. Let X be a normed space over K. Then (a) the mapping (x, y) x + y is uniformly continuous on X × X; (b) for each t ∈ K the mapping x tx is uniformly continuous on X; (c) for each x ∈ X the mapping t tx is uniformly continuous on K; (d) the mapping (t, x) tx is continuous on K × X. If X is a normed space and S is a linear subset of X, then the restriction to S of the norm on X is a norm on S; taken with this norm, S is called a normed linear subspace, or simply a subspace, of X. It is left as an exercise to show, from Proposition 2.3.1, that the closure of a subspace in X is also a subspace of X. Recall that a mapping u between vector spaces X, Y is linear if u(x + y) = u(x) + u(y) and u(tx) = tu(x) whenever x, y ∈ X and t ∈ K. For example, the mapping x Ax on Kn , where A is an n-by-n matrix over K, is linear. If X = Y, then we refer to a linear map u : X −→ Y as an operator on X. On the other hand, if Y = K, then u is called a linear functional on X. Continuous linear mappings between normed spaces are the backbone of functional analysis. In order to characterise these maps, we need a simple lemma, whose proof is left to the reader. Lemma 2.3.2. For each element x of a normed space X, 1 : t ∈ K, t = 0, tx 1 . x = inf |t| Proposition 2.3.3. The following are equivalent conditions on a linear mapping of a normed space X into a normed space Y : (a) u is continuous at 0. (b) u is continuous on X. (c) u is uniformly continuous on X. (d) u (BX ) is a bounded subset of Y .

2.3 Normed Linear Spaces

51

(e) u maps bounded subsets of X to bounded subsets of Y. (f) There exists a positive number c, called a bound for u, such that u(x) c x for all x ∈ X. Proof. Suppose that u is continuous at 0. Then there exists r > 0 such that u(x) = u(x) − u(0) 1 whenever x r. For each nonzero t ∈ K with tx 1 we have rtx r and therefore 1 1 u(rtx) . u(x) = r |t| r |t| It follows from Lemma 2.3.2 that u(x) (f).

1 r

x for all x ∈ X. Hence (a) implies

It is clear that (f) =⇒ (e) =⇒ (d). Next, suppose that there exists c > 0 such that u(x) c whenever x 1. Given x in X and ε > 0, we have either x = 0 or x < ε. In the ﬁrst case, 1 x c x . u(x) = x u x In the second, u ε−1 x c, so u(x) cε c (x + ε) . Thus u(x) c (x + ε) in each case. Letting ε −→ 0, we obtain u(x) c x. Thus (f) holds. We now have u(x − y) c x − y

(x, y ∈ X) ,

from which it follows that u is uniformly continuous on X. Thus (d) =⇒ (f) =⇒ (c). Finally, it is immediate that (c) =⇒ (b) =⇒ (a).

2

In view of property (e) of Proposition 2.3.3, we refer to a continuous linear mapping between normed spaces X, Y as a bounded linear mapping. In the cases Y = X and Y = K, we use the terms bounded operator and bounded linear functional, respectively. Bounded linear functionals are associated with linear subsets that, in the case of a general normed space, correspond to planes in three-dimensional geometry. A linear subset H of a normed space X is called a hyperplane if there exist an associated vector x0 ∈ X and a positive number c such that

52

2 Techniques of Elementary Analysis

x − x0 c for each x ∈ H, and each x ∈ X is represented (perforce uniquely) in the form x = tx0 + y with t ∈ K and y ∈ H. Recall that the kernel of a linear mapping u between linear spaces X, Y is the set

ker(u) = u−1 (0) .

We say that u is nonzero if there exists x ∈ X such that u(x) = 0. If u is a nonzero bounded linear functional, then ker(u) is a hyperplane. To see this, choose x0 ∈ X such that u(x0 ) = 1, and then (by Proposition 2.3.3) c > 0 such that c |u(x)| x for each x ∈ X. If x ∈ ker(u), then x − x0 c |u(x − x0 )| = c |u(x) − u(x0 )| = c. On the other hand, for each x ∈ X we have x = u(x)x0 + (x − u(x)x0 ) , where u(x) ∈ K and x − u(x)x0 ∈ ker(u). Proposition 2.3.4. Let X be a normed space, and H a hyperplane in X with associated vector x0 . Then there exists a unique bounded linear functional u on X such that ker(u) = H and u(x0 ) = 1. Proof. Compute c > 0 such that x − x0 c for each x ∈ H. For each x ∈ X there exist unique u(x) ∈ K and φ(x) ∈ H such that x = u(x)x0 + φ(x). By the uniqueness, u(x0 ) = 1 and H = ker(u). Moreover, for each ε > 0 we have either |u(x)|

1 x + ε c

(2.7)

or else u(x) = 0. In the latter case we have 1 φ(x) x = |u(x)| x0 − − |u(x)| c, u(x) from which we see that (2.7) holds. Since ε > 0 is arbitrary, we conclude that 1/c is a bound for u. 2 If u : X −→ Y is a bounded linear mapping between normed spaces and u = sup {u(x) : x ∈ X, x 1}

(2.8)

exists, we call this number the norm of u and we say that u is normed or normable. In that case,

2.3 Normed Linear Spaces

u(x) u x

53

(x ∈ X) ,

and if X is nontrivial, then u = sup {u(x) : x ∈ X, x = 1} . Although there is no general criterion for the existence of the norm of a bounded linear mapping between normed spaces, such a criterion—and a very useful one at that—exists for bounded linear functionals. Our proof depends on a preliminary lemma. Lemma 2.3.5. Let u be a bounded linear map of a normed space X into a normed space Y. Then ker(u) is located if and only if sx = inf {t : t > 0, u(x) ∈ tu(BX )}

(2.9)

exists for each x ∈ X. Moreover, if u(x) = 0, then sx > 0. Proof. For each x ∈ X we have {t > 0 : ∃y ∈ ker(u)(x − y < t)} = {t > 0 : ∃z ∈ X (z < 1 and u (x − tz) = 0)} = {t > 0 : u(x) ∈ tu(BX )}. The ﬁrst part of the lemma now follows, since ρ (x, ker(u)), if it exists, equals the inﬁmum of the ﬁrst set, whereas sx , if it exists, equals the inﬁmum of the third. Now let u(x) = 0, and choose r > 0 such that if y ∈ X and y < r, then |u(y)| < |u(x)| . Supposing that sx < r, we can ﬁnd a positive t < r and an element z of BX such that u(x) = tu(z) = u(tz); but tz < r, so |u(tz)| < |u(x)| , a con2 tradiction. Hence sx r.

Proposition 2.3.6. A nonzero bounded linear functional on a normed space is normed if and only if its kernel is located. Proof. Let u be a nonzero bounded linear functional on the normed space X. Supposing ﬁrst that u is normed—in which case u > 0—consider any a ∈ X. For each y ∈ ker(u) we have a − y

|u(a)| |u(a − y)| = . u u

On the other hand, if 0 < ε < u and we choose a unit vector x ∈ X such that u(x) > u − ε, then u(a) x ∈ ker(u) z =a− u(x)

54

2 Techniques of Elementary Analysis

and a − z =

|u(a)| |u(a)| < . u(x) u − ε

Hence ρ(a, ker(u)) exists and equals |u(a)| / u . Since a ∈ X is arbitrary, ker(u) is located. Conversely, if ker(u) is located, then by Lemma 2.3.5, s = inf {t > 0 : 1 ∈ tu(BX )} exists and is positive. We show that u = 1/s. For each x ∈ BX we have either u(x) 1/s or u(x) = 0; in the latter case, |u(x)| x u(x) 1 and 1=

1 u |u(x)|

|u(x)| x , u(x)

so 1/ |u(x)| s and therefore |u(x)| 1/s. Thus 1/s is a bound for u. On the other hand, given ε with 0 < ε < 1/s, choose t < s/ (1 − εs) and x ∈ BX such that 1 = tu(x). Then u(x) = 1/t > 1/s − ε. It now follows that u exists and equals 1/s. 2

We cannot expect the norm of a normed linear functional u to be attained at some vector in the unit ball of the domain space. For a certain important class of spaces, though, attainment of the norm does occur. A normed space X is said to be uniformly convex if for each exists ε > 0 there δ with 0 < δ < 1 such that if x, y are unit vectors in X with 12 (x + y) > 1 − δ, then x − y < ε. Every linear subspace of a uniformly convex normed space is itself uniformly convex. Proposition 2.3.7. If u is a nonzero normed linear functional on a uniformly convex Banach space X, then there exists a unique unit vector x ∈ X such that u(x) = u . Proof. Construct a sequence (xn )n1 of unit vectors in X such that u (xn ) −→ u as n −→ ∞. Given ε > 0, choose δ > 0 as in the deﬁnition of “uniformly convex”. Since 1 − 2δ < 1, there exists a positive integer N such that 1 − 2δ u < u(xn ) for all n N. For m, n N we then have

2.3 Normed Linear Spaces

55

1 1 u (xm + xn ) 2 |u(xm + xn )| 2 1 u(xm ) − |u(xm − xn )| 2 1 1 u(xm ) − (u − u(xm )) − (u − u(xn )) 2 2 δ δ δ u − u − u > 1− 2 4 4 = (1 − δ) u . Hence 12 (xm + xn ) > 1−δ and therefore xm − xn < ε. Since ε > 0 is arbitrary, we conclude that (xn )n1 is a Cauchy sequence. It therefore converges to a limit x in the complete space X. By the continuity of the norm and of u, we have x = 1 and u(x) = u . Now let y be any unit vector such that u(y) = u = u(x). For each positive integer n deﬁne z2n−1 = x and z2n = y. Then (zn )n1 is a sequence of unit vectors such that u(zn ) −→ u ; so, by the ﬁrst part of the proof, (zn )n1 converges to a unit vector, which clearly must equal both x and y. Hence x = y. 2

We next present an important construction of new normed spaces from old. Let Y be a located subspace of a normed space X, and deﬁne new equality and inequality relations on X by x =X/Y x ⇐⇒ ρ (x − x , Y ) = 0, x =X/Y x ⇐⇒ ρ (x − x , Y ) > 0. Taken with this equality relation, X becomes the quotient space of X by Y, and is usually redesignated X/Y. The identity (linear) mapping x x from the original normed space X into X/Y is then called the canonical injection iX/Y of X onto X/Y . If Y is also closed in X, then xX/Y = ρ (x, Y ) deﬁnes a norm—the quotient norm—on X/Y. Since xX/Y x − 0 = x , the canonical injection iX/Y is a bounded linear mapping. Note that the locatedness of Y is essential for the deﬁnition of the equality, inequality, and norm on X/Y. Proposition 2.3.8. If Y is a closed, located subspace of a Banach space X, then the quotient space X/Y is a Banach space. Proof. Given a Cauchy sequence (xn )n1 in X/Y, choose a strictly increasing sequence (nk )k1 of positive integers such that

56

2 Techniques of Elementary Analysis

xn

k+1

− xnk X/Y < 2−k

(k 1).

Setting y1 = 0, we construct inductively a sequence (yk )k1 in Y such that for each k, (xn − yk+1 ) − (xnk − yk ) < 2−k . (2.10) k+1 Indeed, having constructed elements y1 , . . . , yk of Y with the applicable properties, we have inf xnk+1 − (xnk − yk ) − y : y ∈ Y = xnk+1 − xnk X/Y < 2−k , so there exists yk+1 ∈ Y such that (2.10) holds. We now see from (2.10) that (xnk − yk )k1 is a Cauchy sequence in the Banach space X; whence it converges to a limit z in X. We then have xnk − zX/Y = xnk − z − yk X/Y (xnk − yk ) − z −→ 0 as k −→ ∞. Thus the Cauchy sequence (xn )n1 in X/Y has a subsequence that converges in 2 X/Y. It follows that the sequence (xn )n1 itself converges in X/Y. Normed spaces form the natural abstract context for inﬁnite series. Given a sequence (xn )n1 of elements of a normed space X, we deﬁne the corresponding ∞ n xn to be the sequence (sn )n1 , where sn = xk is the nth partial sum series n=1

of the series. The series

∞ n=1

k=1

xn is said to be

convergent if the sequence (sn )n1 converges to a limit s in X, called the sum of the series; absolutely convergent if the series In the ﬁrst case we write

∞ n=1

∞ n=1

xn is convergent in R.

xn = s.

If (an )n1 and (bn )n1 are sequences of real numbers such that 0 < an bn for ∞ ∞ each n, and if bn converges, then an converges (the comparison test). For, n=1

n=1

given ε > 0, since the partial sums of a convergent series form a Cauchy sequence we can ﬁnd N such that k bn < ε 0< n=j

whenever k > j N. For all such j, k we then have 0 < sums of

∞ n=1

k n=j

aj < ε. So the partial

an form a Cauchy sequence, which converges by the completeness of R.

2.3 Normed Linear Spaces

57

A particular case of this occurs when bn = rn for some ﬁxed r with |r| < 1: in that ∞ ∞ r case, bn is a geometric series, and an converges to a sum at most 1−r . n=1

n=1

We shall resume consideration of normed spaces in Chapter 4.

Exercises 1. Prove that two real numbers x and y are equal if and only if x ∪ y is a real number. (This is not a typographical error: remember, a real number is actually a set.) 2. Prove the following for real numbers x, y, z: x > y =⇒ x + z > y + z, (x > 0 ∧ y > 0) =⇒ xy > 0, xy = 0 =⇒ (x = 0 ∧ y = 0) . 3. Prove that for all x, y ∈ R the sets max {x, y} and min {x, y} are real numbers; that min {x, y} x max {x, y} ; that x = min {x, y} if and only if x y; and that x = max {x, y} if and only if x y. Prove also that max {x, y} < ε if and only if x < ε and y < ε; that max {x, y} ε if and only if x ε and y ε; and that max {x, y} > z if and only if either x > z or y > z. 4. Prove that |x| = x if and only if x 0, and that |x| = −x if and only if x 0. Prove also that |x| < y if and only if −y < x < y, and that |x| y if and only if −y x y. 5. Prove Archimedes’ axiom in the following form: If x > 0 and y 0, then there exists an integer n such that nx > y. 6. Let x, y be real numbers with y = 0. Prove that x/y is a real number and that y/y = 1. 7. Prove that the statement ∀x ∈ R (¬ (x = 0) =⇒ ∃y ∈ R (xy = 1)) implies Markov’s principle. 8. Prove that the mapping (x, y) x + y is a function on R × R, and that it is strongly extensional. Prove analogous results for the mappings x −x, (x, y) xy, (x, y) x/y (for y = 0), x |x| , (x, y) min {x, y} , and (x, y) max {x, y} .

58

2 Techniques of Elementary Analysis

9. Give Brouwerian examples to show that each of the following statements is essentially nonconstructive: (a) If S ⊂ R is closed, then ∼S is open in R. (b) If S ⊂ R and ∼S is open, then S is closed in R. 10. Let X be a separable metric space. Prove that (i) dense subsets and (ii) located subsets of X are separable. 11. Show that the statement “the metric complement of every located subset of R is located” implies the weak law of excluded middle (WLEM), ¬P ∨ ¬¬P. 12. Prove that the closure of a totally bounded subset of a metric space X is totally bounded, and that if a subset S of X contains a dense totally bounded set, then S itself is totally bounded. 13. Prove Corollary 2.2.8 directly, without using Proposition 2.2.5, Proposition 2.2.6, or Corollary 2.2.7. 14. Let h be a mapping of a metric space into a totally bounded space X such that f ◦ h is uniformly continuous for each uniformly continuous map f : X −→ R. Prove that h is uniformly continuous. 15. Let h be a mapping of a compact metric space X into a metric space Y such that f ◦ h is uniformly continuous on X for each uniformly continuous mapping f : Y −→ R. Prove that h is uniformly continuous if and only if its range is totally bounded. Prove also that if Y is locally compact, then h is uniformly continuous. 16. Let a < b, and let S, T be inhabited open subsets of R such that [a, b] ⊂ S ∪ T. Prove that S ∩ T is inhabited. 17. Let f : [a, b] −→ R be sequentially continuous, with f (a) f (b). Prove that the range of f is dense in [f (a), f (b)] . 18. Let I be a bounded interval, and f : I −→ R an increasing sequentially continuous function. Prove that f is uniformly continuous on I. 19. Let S be a dense subset of a metric space X, and f a uniformly continuous mapping of S into a complete metric space Y. Prove that f has a uniformly continuous extension to a mapping of X into Y ; that is, a uniformly continuous mapping F : X −→ Y such that F (x) = f (x) for all x ∈ S. 20. Prove that the Hausdorﬀ metric on the set K(X) of compact subsets of a compact metric space X is indeed a metric, and that K(X) is complete with respect to it. Why do we need the sets to be compact in order to be sure that ρ deﬁnes a metric?

2.3 Normed Linear Spaces

59

21. Let X be a compact metric space, K(X) the compact metric space of all compact subsets of X (see the preceding exercise), and E the set of all pairs (x, K) where K ∈ K(X) and x ∈ K. Deﬁne a metric d on E by d ((x, K) , (x , K )) = ρ (x, x ) + ρ (K, K ) , where the second ρ on the right-hand side denotes the Hausdorﬀ metric on K(X). Prove that E is compact with respect to this metric d. sequence of compact subsets 22. Let X be a metric space, and (Kn )n1 a decreasing Kn is compact. Does the of X whose diameters converge to 0. Prove that n1

conclusion hold without the hypothesis that the diameters converge to 0? 23. Under the hypotheses of Corollary 2.2.14, prove that if r ∈ R is admissible and ε > 0, then there exists δ > 0 such that for each admissible r with |r − r | < δ, ρ (X(f, r), X(f, r )) < ε, where ρ is the Hausdorﬀ metric on K(X). 24. Give a Brouwerian example to show that we cannot drop the condition “for all but countably many” from the conclusion of Theorem 2.2.13. 25. Prove that the product of ﬁnitely many (locally) compact metric spaces is (locally) compact. 26. Let f be a mapping from a metric space (X, ρ) into R with the following properties: (a) the set X(f, r) is compact for certain arbitrarily large real numbers r; (b) f is uniformly continuous on each of the sets X(f, r). Deﬁne a new metric ρ0 on X by ρ0 (x, x ) = ρ (x, x ) + |f (x) − f (x )|

(x, x ∈ X) .

Prove that (X, ρ0 ) is a locally compact metric space. 27. Let S be a located subset of a locally compact metric space (Y, ρ) such that X = −S is inhabited. Let the mapping h : Y −→ R be uniformly continuous on each bounded subset of Y, and such that the set Y (h, r) is bounded for −1 each r ∈ R. Prove that the mapping f : x h (x) + ρ (x, S) on X satisﬁes conditions (a) and (b) of the preceding exercise. 28. Let X be a uniformly convex Banach space, and u a normed linear functional on X. Prove that for any two distinct unit vectors x, y in X either |u(x)| < u or |u(y)| < u . (This is a strong form of the uniqueness result in Proposition 2.3.7.)

60

2 Techniques of Elementary Analysis

29. Prove Proposition 2.3.1. 30. Prove Lemma 2.3.2.

Notes In view of our earlier remarks on the status of the power set, it may seem strange, if not perverse, to deﬁne a real number as a set. We do not believe that our deﬁnition will give rise to any logical problems, since in practice a real number is likely to be given explicitly by interval approximations, rather than by any appeal to the existence of the full power set of Q × Q. In fact, many, if not most, real numbers will actually arise as limits of sequences of rational approximations, from which it is straightforward to construct a set of the type required by our deﬁnition of “real number”. Bishop deﬁned a real number to be a sequence (xn )n1 of rational numbers that is regular in the sense that |xm − xn | < 1/m + 1/n for all m and n, two such sequences (xn )n1 and (yn )n1 being called equal if |xn − yn | < 2/n for each n. For an axiomatic development of R see [19]. For us, = normally has a stronger meaning than the denial of equality. Some authors, notably the intuitionists, use # instead of = to denote an inequality relation, and = to denote the denial inequality. Bishop originally used the word “continuous” to describe mappings that are uniformly continuous on compact sets. While this gets round the independence of the uniform continuity theorem relative to BISH, it has the disadvantage that we cannot prove that the composition of continuous functions is continuous, since the image of a compact set, although totally bounded, cannot generally be proved complete. (To be more precise, in RUSS there is a uniformly continuous mapping of [0, 1] onto (0, 1].) In a later, unpublished manuscript, The neat category of stratiﬁed spaces, Bishop introduced the notion of a compact image: a subset of a metric space that is the image of a compact metric space under a uniformly continuous mapping. He then deﬁned a mapping f : X −→ Y between metric spaces to be continuous if it is uniformly continuous near each compact image, in the following sense: for each compact image S ⊂ X and each ε > 0 there exists δ > 0 such that ρ(f (x), f (x )) < ε whenever x ∈ X, x ∈ S, and ρ(x, x ) < δ. Compositions of functions that are continuous in this sense are continuous; but the interval (0, 1] is a compact image in RUSS, so we cannot prove that the mapping x 1/x on (0, 1] is uniformly continuous near compact images. It seems that without Brouwer’s fan theorem (see Chapter 5 of [34]), it is impossible to come up with a deﬁnition of “continuous” in BISH, other than the usual one of “pointwise continuous”, that will satisfy all the conditions that one might wish for. For more on this, see [81] and [89].

2.3 Normed Linear Spaces

61

In this connection, Bishop’s statement that “The concept of a pointwise continuous function is not relevant” ([9], pages ix–x) no longer seems appropriate: there are instances—for example, in the theory of operators between Banach spaces— where sequential convergence is both useful and the best we can hope to prove within BISH; see Chapter 6. Some authors remove the word “inhabited” from the deﬁnitions of “located” and “totally bounded”, thereby allowing the empty set to be both located and totally bounded. For us, located sets and totally bounded sets are inhabited, by deﬁnition. Why do we deﬁne “compact” as we have done? Classically, compactness is deﬁned in terms of the Heine–Borel–Lebesgue covering property: every open cover contains a ﬁnite subcover. In Brouwer’s intuitionistic mathematics, as classically, the interval [0, 1] has this open-cover compactness property; but Brouwer’s proof depends on his fan theorem, the addition of which to intuitionistic logic would lose us some of the ﬂexibility of interpretation of our constructive mathematics. On the other hand, if we add the Church–Markov–Turing thesis to intuitionistic logic, then we can prove that [0, 1] does not have the Heine–Borel–Lebesgue covering property ([34], page 60, Theorem (4.1)). So that property holds for [0, 1] in one model of our constructive mathematics but fails to hold in another. Therefore, without adding some principle to BISH, we cannot prove or disprove that it holds for [0, 1]. A second classical compactness property, equivalent to the Heine–Borel–Lebesgue one, is that of sequential compactness: every sequence in the space has a convergent subsequence. This fails even for the pair set {0, 1} in the constructive setting. Thus, in looking for a workable constructive notion of compactness for metric spaces, we are reduced to that of completeness plus total boundedness, which is classically equivalent to the other two considered above. Fortunately, completeness and total boundedness together make a combination suﬃciently powerful for most constructive purposes. Proposition 2.2.11 appears in Aspects of Constructivism, the unpublished notes of colloquium lectures given by Bishop at New Mexico State University in December 1972. It provides a neat proof of Corollary 2.2.12. The proof of Lemma 2.2.17 is new, although not unlike that of Lemma (4.10) on page 33 of [34]. Exercise 14 provides a constructive version of the classical theorem that the unique uniform structure U compatible with the given topology on a compact Hausdorﬀ space X is induced by the U-uniformly continuous mappings of X into R. Note, however, that we require X to be only totally bounded, not compact. The notion of strong continuity is best studied in the context of an apartness space: that is, an inhabited set X with an inequality and a binary relation of apartness between sets that satisﬁes certain natural axioms. For more on apartness spaces see [38, 40]. Exercise 16 describes one version of the connectedness of the interval [a, b]. For other, constructively inequivalent, types of connectedness see [18].

62

2 Techniques of Elementary Analysis

The result in Exercise 17 is perhaps the most general constructive intermediate value theorem. A linear map u : X −→ Y between normed spaces is said to be compact if u(BX ) is a totally bounded subset of Y. Proposition 2.3.6 can be generalised as follows: A bounded linear mapping of a normed linear space onto a ﬁnite-dimensional Banach space is compact if and only if its kernel is located ([34], page 36, Theorem (5.4)). Bishop required that a Banach space be separable. We prefer not to depend on separability unless it is absolutely necessary. We use the term “normed”, rather than the usual “normable”, for those linear mappings for which the norm exists.

3 The λ-Technique

Here is more matter for a hot brain. —Shakespeare, The Winter’s Tale, act 5, scene 2 In this chapter we discuss a peculiarly constructive technique that, normally under the hypothesis that one or more of the metric spaces under consideration is complete, enables us to prove results that otherwise would require some nonconstructive principle such as LPO, LLPO, or Markov’s principle.

3.1 Introduction to the Technique We begin with another elementary classical result that does not hold in constructive mathematics: For each complex number z there exists θ ∈ [0, 2π) such that z = |z| eiθ , and such that if θ = 0, then z = 0. We show that this proposition entails LPO. To do so, we consider an increasing binary sequence (λn )n1 with at most one term equal to 1. Deﬁne a sequence (zn )n1 of complex numbers such that λn = 0 =⇒ zn = 0, 1 λn = 1 − λn−1 =⇒ zk = eiπ/2 for all k n. n Then |zm − zn |< 1/n whenever m > n, so (zn )n1 is a Cauchy sequence and therefore converges to a limit z in the complete metric space C. Assume that z = |z| eiθ for some θ ∈ [0, 2π), and that if θ = 0, then z = 0. Either θ < π/2 or θ > 0. In the ﬁrst case we have λn = 0 for all n: for if we suppose that there exists n such

64

3 The λ-Technique

that λn = 1 − λn−1 , then z = eiπ/2 /n and therefore θ = π/2, a contradiction. In the case θ > 0 we have z = 0, so there exists N such that zN = 0; then λn = 1 for some n N. The constructive problem occurs when z is near 0 but we cannot decide whether z = 0 or z = 0. Indeed, if z = 0 or z = 0, then we can ﬁnd θ ∈ [0, 2π) such that z = |z| eiθ . Now, it might be thought that the failure of the modulus–argument decomposition of a general complex number would mean that there was no constructive proof of the existence of square roots in C; for in order to ﬁnd a square root of z, we normally would write z = |z| eiθ and then compute ± |z|eiθ/2 . Although this method of ﬁnding a square root certainly will not work unless we already can decide that z = 0 or z = 0—which, in general, we cannot—there is a constructive proof of the existence of square roots, one that uses the completeness of C. To see this, consider any complex number z, and note that for each positive integer n we have either |z| < 1/n2 or |z| > 1/(n + 1)2 . Thus we can successively construct the terms of an increasing binary sequence (λn )n1 such that λn = 0 =⇒ |z| < λn = 1 =⇒ |z| >

1 , n2 1 (n + 1)

2.

We may assume that λ1 = 0. If λn = 0, setζ n = 0; if λn = 1 − λn−1 , choose θ ∈ [0, 2π) such that z = |z| eiθ , and set ζ k = |z|eiθ/2 for all k n. Then (ζ n )n1 is a Cauchy sequence. To see this, let m n. If λm = 0 or λn = 1, then ζ m = ζ n . If λm = 1 and λn = 0, then there exists a unique k such that n < k m and 2 λk = 1 − λk−1 ; whence |z| < 1/ (k − 1) and |z|

0.

3.1 Introduction to the Technique

65

Proof. Let s0 ∈ S. If ρ(x, S) > 1/2, then we can take s = s0 . Hence we may assume that ρ(x, S) < 1. Now construct an increasing binary sequence (λn )n1 such that λ1 = 0 and 1 , n 1 . λn = 1 =⇒ ρ(x, S) > n+1

λn = 0 =⇒ ρ(x, S)

3/N. If λN = 0, then ρ(x, sN ) < 1/N and so ρ(s, sN ) ρ(x, s) − ρ(x, sN ) >

2 , N

a contradiction. Thus λN = 1 and therefore ρ(x, S) > 0.

2

Bishop’s lemma is simple to prove using classical logic. For if x ∈ S, then by taking s = x we render the hypothesis of the implication “if x = s, then ρ(x, S) > 0 ” false; whereas if x ∈ / S, then the conclusion of that implication holds since S is closed in X, so we may take s = s0 . This argument depends on the full law of excluded middle, but in fact only LPO is needed to establish Bishop’s lemma classically even when S is merely closed and not necessarily complete. To see this, ﬁrst construct the sequence (λn )n1 as in our constructive proof of the lemma. If λn = 0 for all n, then ρ(x, S) = 0, so x ∈ S and we can take s = x; whereas if λN = 1, then ρ(x, S) > 0 and we can take s = s0 . In order to obtain a constructive proof of Bishop’s lemma, we replace LPO by the completeness of S. This is a typical situation, in which a result proved classically using an omniscience principle is proved constructively under some completeness condition using the λ-technique. Now, it follows immediately from Bishop’s lemma that if S is a complete, located subset of a metric space, then ∼S = −S. The λ-technique provides us with a related result in a Banach space, in which locatedness is replaced by convexity. A subset C of a linear space X is said to be convex if tx + (1 − t) y ∈ C whenever x, y ∈ C and 0 t 1; absorbing if for each x ∈ X there exists t > 0 such that x ∈ tC.

66

3 The λ-Technique

Proposition 3.1.2. If C is a convex, absorbing subset of a Banach space X, then ∼C is dense in ¬C. Proof. Let x ∈ ¬C, let ε > 0, and choose δ > 0 such that δx < ε. Then x = (1 + δ)x ∈ (1 + δ)C. Given y ∈ C, we show that x = y. To that end, construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ x − y

n. Hence (zn )n1 is a Cauchy sequence and so converges to a limit z in the complete space X. Choose a positive integer N such that z ∈ N δC, and consider any integer n N . If λn = 1 − λn−1 , then z = n(x − y); whence x = y +

Nδ 1 z∈y+ C ⊂ C + δC = (1 + δ)C, n n

a contradiction. Thus λn = λn−1 for all n N . It follows that if λN = 0, then λn = 0 for all n, and therefore x = y ∈ C ⊂ (1 + δ)C. This contradiction ensures that λN = 1; whence x = y. Since y ∈ C is arbitrary, it follows that x ∈ ∼C. Since also x − x = δx < ε, and x ∈ ¬C and ε > 0 are arbitrary, we conclude that ∼C is dense in ¬C. 2

Classically, as the reader is invited to prove, the completeness of the space X can be removed from Proposition 3.1.2 since we are allowed to use LPO. For an example of the use of the λ-technique to avoid the application of Markov’s principle, we turn to an elementary result in metric topology. We observed in Chapter 2 that it cannot be proved constructively that the union of two closed subsets of a metric space is closed. Nevertheless, it seems reasonable to expect that if A, B are closed subsets of R, then the complement, in some appropriate sense, of A ∪ B will be the intersection of the complements of A and B. Since it is trivial that the ﬁrst of these complements is a subset of the intersection of the other two, it is enough to prove the reverse inclusion. With the help of Markov’s principle, we can prove that

3.1 Introduction to the Technique

67

(¬A ∩ ¬B) ∩ A ∪ B = ∅. Indeed, supposing that x ∈ (¬A ∩ ¬B) ∩ A ∪ B, choose a sequence (xn )n1 of elements of A ∪ B converging to x. Deﬁne an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (xk ∈ A) , λn = 1 − λn−1 =⇒ xn ∈ B. We may assume that λ1 = 0. If λn = 0 for all n, then x ∈ A = A, a contradiction. Hence, by Markov’s principle, there exists n > 1 such that λn = 1 − λn−1 , so there exists n1 > 1 with xn1 ∈ B. Repeated application of this argument enables us to construct a strictly increasing sequence (nk )k1 of positive integers such that xnk ∈ A for all even k, and xnk ∈ B for all odd k. Thus x ∈ A ∩ B, which is a ﬁnal contradiction. Since such an argument, depending on Markov’s principle, is not good enough for our constructive purposes, it is fortunate that if we work within a complete metric space and use the λ-technique (on more than one occasion), we can say something interesting about the intersection of the complements of two closed sets. This will require a preliminary result. Lemma 3.1.3. Let X be a complete metric space, A a closed subset of X, and B a subset of X. Let x ∈ ∼A, let y ∈ A ∪ B, and let (yn )n1 be a sequence in A ∪ B that converges to y. Then either x = y or there exists n such that yn ∈ B. Proof. We may assume that y1 ∈ A and, by passing to a subsequence if necessary, that ρ (yn , y) < 1/n for each n. Construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (yk ∈ A) , λn = 1 − λn−1 =⇒ yn ∈ B. If λn = 0, set ξ n = yn ; if λn = 1 − λn−1 , set ξ k = ξ n−1 for all k n. Then (ξ n )n1 is a Cauchy sequence in A; in fact, 2 (m n) . n Since X is complete and A is closed, (ξ n )n1 converges to a limit ξ ∈ A. Choose a positive integer N such that ρ (x, ξ) > 2/N. If λN = 1, then yn ∈ B for some n N. If λN = 0, then either x = y or, as we may suppose, ρ (x, y) < 1/N. If there exists m N such that λm+1 = 1 − λm , then ρ (ξ m , ξ n ) ρ (ξ m , y) + ρ (ξ n , y)

− , > N N m N a contradiction. Hence λn = 0 for all n N and therefore for all n; so yn ∈ A for all n. Since A is closed, y ∈ A and therefore x = y. 2

68

3 The λ-Technique

Proposition 3.1.4. If A, B are closed subsets of a complete metric space X, then ∼A ∩ ∼B = ∼A ∪ B. Proof. It is clear that ∼A ∪ B ⊂ ∼A ∩ ∼B. To prove the reverse inclusion, given x ∈ ∼A ∩ ∼B and y ∈ A ∪ B, we must prove that x = y. To this end, choose a sequence (yn )n1 in A ∪ B that converges to y. We may assume that y1 ∈ A and that ρ (yn , y) < 1/n for each n. Set λ1 = 0, n1 = 1, and ξ 1 = y1 . In view of Lemma 3.1.3, we may also assume that there exists n2 > 1 such that yn2 ∈ B; set λ2 = 0 and η 1 = yn2 . Now apply the same lemma, but with the roles of A and B interchanged. Either we have x = y, when for each k 3 we set λk = 1, nk = n2 , ξ k−1 = ξ 1 , and η k−1 = η 1 ; or else, as we may assume, there exists n3 > n2 such that yn3 ∈ A, in which case we set λ3 = 0 and ξ 2 = yn3 . Repeating such applications of Lemma 3.1.3, we construct an increasing binary sequence (λn )n1 , an increasing sequence (nk )k1 of positive integers, a sequence (ξ k )k1 of elements of A, and a sequence (η k )k1 of elements of B, such that for each k 1, • if λ2k = 0, then n2k > n2k−1 and η k = yn2k ∈ B; • if λ2k+1 = 0, then n2k+1 > n2k and ξ k+1 = yn2k+1 ∈ A; • if λ2k = 1−λ2k−1 , then x = y and for each j k, n2j+1 = n2j = n2j−1 , ξ j = ξ k , and η j = η k−1 ; • if λ2k+1 = 1−λ2k , then x = y and for each j k, n2j+2 = n2j+1 = n2k , ξ j = ξ k , and η j = η k . Then (ξ k )k1 is a Cauchy sequence in A, and (η k )k1 is a Cauchy sequence in B. Indeed, for m n we have 2 ρ (ξ m , ξ n ) ρ (ξ m , y) + ρ (ξ n , y) < , n and similarly ρ (η m , η n ) < 2/n. Since X is complete and the subsets A, B are closed in X, the sequences (ξ n )n1 , (η n )n1 converge to respective limits ξ ∈ A and η ∈ B. Choose an integer N > 4 such that ρ (x, ξ) > 3/N and ρ (x, η) > 3/N. If λN = 1, then x = y; so we may assume that λN = 0. Either x = y or, as we may further assume, ρ (x, y) < 1/N. Suppose there exists m > N such that λm = 1 − λm−1 . In the case where m is even, we have ρ(x, y) ρ (x, η) − ρ (η, y)

3 − ρ ynm−2 , y > N 1 3 − > N m−2 2 3 − (since m > 4) > N m 1 > , N

3.1 Introduction to the Technique

69

a contradiction. If m is odd, a similar argument again leads us to a contradiction. Hence λn = 0 for all n N and therefore for all n. It follows that y = ξ = η and therefore ρ (x, y) > 3/N, a contradiction. Hence the case ρ (x, y) < 1/N is ruled out, and so ρ (x, y) 1/N. 2

The λ-technique turns out to be very useful for clarifying the connections between various continuity properties of functions between metric spaces. It is trivial that a continuous mapping between metric spaces is strongly extensional. If the domain is complete, then we can weaken continuity to sequential continuity. Proposition 3.1.5. Let f be a sequentially continuous mapping of a complete metric space X into a metric space Y . Then f is strongly extensional. Proof. Given points x, y of X with f (x) = f (y), construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ρ(x, y)

0. Proof. Construct an increasing binary sequence (λn )n1 such that

70

3 The λ-Technique

λn = 0 =⇒ ρ ((x, y) , B)

0.

2,

We may assume that λ1 = 0. If λn = 0, set zn = 0; if λn = 1 − λn−1 , pick ξ ∈ X such that 1 (3.1) (ξ, T ξ) ∈ B and x − ξ + y − T ξ < 2 n and set zi = n (x − ξ) for all i n. If j > k and zj − zk > 0, then λk = 0 and λn = 1 − λn−1 for a unique value of n with k < n j; so there exists ξ ∈ X such that (3.1) holds and 1 1 < . n k Hence zj − zk < 1/k whenever j k. Therefore (zn )n1 is a Cauchy sequence in X and so converges to a limit z ∈ X. Now choose a positive integer N such that zj − zk = n (x − ξ)

N. If λn = 1 − λn−1 , then z = n (x − ξ) for some ξ satisfying (3.1), so 1 + T z = 1 + n T x − T ξ 1 + n T x − y − n y − T ξ 1 > 1 + N T x − y − n > N T x − y , a contradiction of our choice of N. Hence λn = λn−1 for all n N. It follows that if λN = 0, then λn = 0 for all n, and therefore ρ ((x, y) , B) = 0. Since B is closed in X × Y, we now have y = T x, which contradicts our original hypotheses. We 2 conclude that λN = 1. Corollary 3.1.7. A linear mapping of a Banach space into a normed space is strongly extensional. Proof. Let T be a linear mapping of a Banach space X into a normed space Y, and let x ∈ X satisfy T x = 0. Given z ∈ ker(T ), apply Proposition 3.1.6 with y = 0 and B = {(z, T z)} , to obtain (x, 0) = (z, T z); whence either x = z or else T z = 0. Since the latter is absurd, we conclude that x = z. 2 Markov’s principle is all that we need add to intuitionistic logic in order to prove Proposition 3.1.6 without the hypothesis that X is complete: for the argument towards the end of the proof of that proposition shows that it is impossible that ρ ((x, y) , B) = 0, so, by Markov’s principle, ρ ((x, y) , B) > 0. In fact (this is left to the exercises at the end of the chapter), if Proposition 3.1.6 holds without the completeness of X, then we can derive Markov’s principle.

3.2 Ishihara’s Tricks

71

3.2 Ishihara’s Tricks In [57], Ishihara introduced the following two lemmas, now called Ishihara’s tricks, in which we use completeness to make a decision which at ﬁrst sight would seem to be impossible with purely constructive techniques. Lemma 3.2.1. Let f be a strongly extensional mapping of a complete metric space X into a metric space Y , and let (xn )n1 be a sequence in X converging to a limit x. Then for all positive numbers α, β with α < β, either ρ(f (xn ), f (x)) > α for some n or ρ(f (xn ), f (x)) < β for all n. Proof. Let (nk )k1 be a strictly increasing sequence of integers such that ρ(x, xn ) < 1/ (k + 1) for all n nk . For convenience, set n0 = 1. For each k 1 set sk = max{ρ(f (xn ), f (x)) : nk−1 n < nk }. Construct an increasing binary sequence (λk )k1 such that λk = 0 =⇒ ∀j k (sj < β) , λk = 1 =⇒ ∃j k (sj > α) . We may assume that λ1 = 0. Deﬁne a sequence (yk )k1 in X as follows. If λk = 0, set yk = x; if λk = 1 − λk−1 , choose ν k with nk−1 ν k < nk and ρ(f (xν k ), f (x)) > α, and set yj = xν k for all j k. Let i j. If λi = 0 or λj = 1, then yi = yj . If λi = 1 − λj , then there exists k with j < k i and λk = 1 − λk−1 ; so yj = x, and yi = yk = xν k for some ν k such that nk−1 ν k < nk and ρ(f (xν k ), f (x)) > α; whence 1 1 < . ρ(yi , yj ) = ρ (xν k , x) νk + 1 j It follows that ρ(yi , yj ) < 1/j whenever i j; so (yk )k1 is a Cauchy sequence and therefore converges to a limit y in X. Either ρ(f (x), f (y)) < α or ρ(f (x), f (y)) > 0. In the ﬁrst case, if λk = 1 − λk−1 , then y = xν k with ρ(f (xν k ), f (x)) > α, a contradiction; whence λk = 0 for all k. Then ρ(f (xn ), f (x)) < β for all n. In the second case, since f is strongly extensional, x = y. Choose a positive integer κ such that x = yκ . If λκ = 0, then x = yκ = x, a contradiction; whence λκ = 1 and there 2 exists n such that ρ(f (xn ), f (x)) > α.

Let P (n) be a property of positive integers n. We say that P (n) holds • eventually, or for all suﬃciently large n, if there exists N such that P (n) holds for all n N ; •

inﬁnitely often if for each n there exists m > n such that P (m) holds.

72

3 The λ-Technique

The foregoing lemma leads to a technique allowing us to decide between alternatives that happen eventually and those that happen inﬁnitely often. Lemma 3.2.2. (Ishihara’s second trick) Let f be a strongly extensional mapping of a complete metric space X into a metric space Y , and let (xn )n1 be a sequence in X converging to a limit x. Then for all positive numbers α, β with α < β, either ρ(f (xn ), f (x)) > α for inﬁnitely many n or else ρ(f (xn ), f (x)) < β for all suﬃciently large n. Proof. Let (nk )k1 be a strictly increasing sequence of positive integers such that ρ(x, xn ) < 1/k for all n nk . Successively applying Lemma 3.2.1 to the subsequence (xn )nnk , construct an increasing binary sequence (λk )k1 such that λk = 0 =⇒ ∃n nk (ρ(f (xn ), f (x)) > α) , λk = 1 =⇒ ∀n nk (ρ(f (xn ), f (x)) < β) . We may assume that λ1 = 0. Deﬁne a sequence (yk )k1 in X as follows. If λk = 0, choose ν k nk such that ρ(f (xν k ), f (x)) > α, and set yk = xν k ; if λk = 1−λk−1 , set yi = yk−1 for all i k. Then (yk )k1 is a Cauchy sequence: in fact, ρ(yi , yj ) < 2/j whenever i j. Let y be the limit of (yk )k1 in the complete space X. Either 0 < ρ(f (x), f (y)) or ρ(f (x), f (y)) < α. In the ﬁrst case, since f is strongly extensional, x = y, so there exists a positive integer κ with ρ(x, yκ ) > 1/κ. If λκ = 0, then yκ = xν κ for some ν κ nκ ; whence ρ(x, yκ ) < 1/κ, a contradiction. Thus λκ = 1, and therefore ρ (f (xn ), f (x)) < β eventually. In the case ρ(f (x), f (y)) < α, if there exists k such that λk+1 = 1 − λk , then y = xν k and ρ(f (xν k ), f (x)) > α, a contradiction; whence λk = 0 for all k, and therefore ρ(f (xn ), f (x)) > α inﬁnitely often. 2 A mapping f : X −→ Y between metric spaces is said to be sequentially nondiscontinuous if it has the following property: if (xn )n1 converges to x ∈ X and ρ (f (xn ) , f (x)) δ for all n, then δ 0. Clearly, a sequentially continuous mapping is sequentially nondiscontinuous. Ishihara’s ﬁrst application of his tricks was to the converse. Proposition 3.2.3. Let f be a mapping of a complete metric space X into a metric space Y. Then f is sequentially continuous if and only if it is strongly extensional and sequentially nondiscontinuous. Proof. Suppose that f is strongly extensional and sequentially nondiscontinuous. Let (xn )n1 be a sequence converging to x ∈ X, and let ε > 0. By Lemma 3.2.2, either there exists a subsequence (xnk )k1 of (xn )n1 such that ρ (f (xnk ), f (x)) > ε/2 for all k, or else ρ (f (xn ), f (x)) < ε for all suﬃciently large n. In the former case, the sequential nondiscontinuity of f shows that ε 0, which is absurd. We conclude that the latter case obtains. Hence f is sequentially continuous.

3.2 Ishihara’s Tricks

73

The converse is an immediate consequence of Proposition 3.1.5 and the observation made before this proposition. 2 Corollary 3.2.4. For linear mappings of a Banach space into a normed space, sequential continuity and sequential nondiscontinuity are equivalent. Proof. This follows immediately from Corollary 3.1.7 and Proposition 3.2.3.

2

We now lift Ishihara’s tricks into a general setting. This both clariﬁes the ideas underlying those lemmas and raises the possibility that some other applications of their proof techniques in constructive analysis are, in fact, corollaries of our general results. We begin with a generalisation of Ishihara’s ﬁrst trick (Lemma 3.2.1). Proposition 3.2.5. Let X be a complete metric space, let P, Q be subsets of X such that X = P ∪ Q, and let x be an element of X such that for each y ∈ X, either x = y or y ∈ / Q. Then for each sequence (xn )n1 in X that converges to x, either xn ∈ P for all n or else there exists N such that xN ∈ Q. Proof. Construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (xk ∈ P ), λn = 1 − λn−1 =⇒ xn ∈ Q. We may assume that λ1 = 0. If λn = 0, set yn = x; if λn = 1−λn−1 , set yk = xn for all k n. To see that (yn )n1 is a Cauchy sequence in X, let ε > 0 and compute N such that ρ(x, xn ) < ε for all n N. Let m n N. If λm = 0, then ym = yn = x. So we may assume that λm = 1. If λn = 0, then ρ(ym , yn ) = ρ(xm , x) < ε. If λn = 1, then ρ(ym , yn ) = ρ(xm , xn ) ρ(x, xm ) + ρ(x, xn ) < 2ε. Thus in all cases, ρ(ym , yn ) < 2ε. Since X is complete, the Cauchy sequence (yn )n1 converges to a limit y ∈ X. Either x = y or y ∈ / Q. In the ﬁrst case, choosing N such that x = yN , we see that λN = 1. In the second case, if there exists m such that λm = 1 − λm−1 , then 2 y = xm ∈ Q, a contradiction; whence λn = 0, and therefore xn ∈ P, for all n. To derive Lemma 3.2.1, assume the hypotheses of that lemma, deﬁne P = {y ∈ X : ρ(f (y), f (x)) < β} , Q = {y ∈ X : ρ(f (y), f (x)) > α} , and apply Proposition 3.2.5, noting that for all y ∈ Y,

74

3 The λ-Technique

• either ρ(f (y), f (x)) > 0 , in which case, by the strong extensionality of f, x = y; • or else ρ(f (y), f (x)) < α and therefore y ∈ / Q. In order to tackle the generalisation of Ishihara’s second trick (Lemma 3.2.2), we prove two lemmas, the ﬁrst of which is a variant of Proposition 3.2.5. Lemma 3.2.6. Let X be a complete metric space, let P, Q be subsets of X such that X = P ∪ Q, and let x be a point of X such that for each y ∈ X, either x = y or y ∈ / P. Then for each sequence (xn )n1 in X that converges to x, either xn ∈ P for all n or else there exists N such that xN ∈ Q. Proof. Construct an increasing binary sequence (λn )n1 such that λn = 0 =⇒ ∀k n (xk ∈ P ) , λn = 1 − λn−1 =⇒ xn ∈ Q. We may assume that λ1 = 0. If λn = 0, set yn = x; if λn = 1 − λn−1 , set yk = xn−1 for each k n. Then (yn )n1 is a Cauchy sequence in X and so converges to a limit y ∈ X. Either x = y or else y ∈ / P. In the ﬁrst case choose N such that ρ (xn−1 , y) > 12 ρ (x, y) for all n N, and suppose that λN = 0. If λm = 1 − λm−1 for some m > N, then y = xm−1 and so 0 = ρ (xm−1 , y) > 12 ρ (x, y), a contradiction. Hence λn = 0 for all n N and therefore for all n; but this implies that y = x, another contradiction. Thus, in fact, λN = 1 and there exists n N such / P we must have λn = 0, and that xn ∈ Q. On the other hand, in the case y ∈ 2 therefore xn ∈ P, for all n. Lemma 3.2.7. Let X be a complete metric space, let P, Q be subsets of X such that X = P ∪ Q, and let x be an element of X. Suppose that for any sequence (xn )n1 converging to x in X, either there exists N such that x = xN or else xn ∈ / Q for all n. Then for each y ∈ X, either x = y or y ∈ / Q. Proof. Applying the hypotheses to the sequence (x, x, . . .) in X, we see that x ∈ / Q. Given y in X, construct a (perforce increasing) binary sequence (λn )n1 such that 1 , n 1 . λn = 1 =⇒ ρ (x, y) > n+1 λn = 0 =⇒ ρ (x, y)

α} , and, for convenience, P = {y ∈ X : ρ (f (x), f (y)) < α} , Q = {y ∈ X : ρ (f (x), f (y)) > α/2} . Then X = P ∪ Q and for each y ∈ X, either x = y or y ∈ / Q. Similarly, X = P ∪ Q and for each y ∈ X, either x = y or y ∈ / Q . It follows from the last statement and Proposition 3.2.5 that if (yn )n1 is any sequence converging to x in X, then either yn ∈ P for all n or else there exists N such that yN ∈ Q . Hence either / Q for all n or else there exists N such that x = yN . Thus P and Q satisfy the yn ∈

76

3 The λ-Technique

hypotheses of Proposition 3.2.8, from which we immediately deduce the conclusion of Ishihara’s second trick. We have now given a number of illustrations of the merit of the λ-technique as a replacement for omniscience principles. The technique will reappear throughout the book, where it will be used, for example, to construct best approximations from ﬁnite-dimensional spaces, to locate certain convex sets in a normed space, and to discuss the range of an operator with an adjoint on a Hilbert space.

Exercises 1. Using the λ-technique, show that if Ra = {ax : x ∈ R} is closed, then a = 0 or a = 0. 2. Let X be a complete metric space, and f : X −→ R a sequentially continuous mapping such that inf f exists. Suppose that to each ε > 0 there corresponds δ > 0 such that if x, y ∈ X and max {f (x), f (y)} < δ, then ρ(x, y) < ε. Prove that there exists a ∈ X such that if f (a) > 0, then inf f > 0. 3. A linear mapping T : X −→ Y between normed spaces is said to be well-behaved if T x = 0 whenever x ∈ ∼ ker(T ). Prove that the statement “every linear mapping between normed spaces is well-behaved” is equivalent to Markov’s principle. Prove also that every linear mapping from a Banach space into a normed space is well-behaved. 4. Prove that if Corollary 3.1.7 (and hence a fortiori Proposition 3.1.6) holds without the hypothesis that X is complete, then we can derive Markov’s principle. 5. Prove de Morgan’s rule for metric complements: If (S n )n1 is a sequence of Sn is complete and located subsets of a metric space X such that S = n1 located, then −S = −Sn . n1

6. A subset S of a metric space X is said to be uniformly almost located if there exists a strictly decreasing sequence (δ n )n1 of positive numbers converging to 0 such that the following holds: for each x ∈ X there exists y ∈ S such that for each n, if ρ (x, y) > δ n , then ρ(x, S) > δ n+1 . Prove that if S is an inhabited, uniformly almost located subset of a locally totally bounded space, then S is located. 7. Let X be a complete metric space, and a a point of X such that s = sup {ρ(a, x) : x ∈ X} exists. Prove that for each r > 1 there exists b ∈ X such that s rρ(a, b).

3.2 Ishihara’s Tricks

77

8. Let T be a sequentially continuous linear mapping of a Banach space X into a normed space Y, and let B be the unit ball of the range of T. Suppose that for each ε > 0, either there exists x ∈ ∼T −1 (B) with x < ε or else ∼T −1 (B) is bounded away from 0 (that is, there exists δ > 0 such that x > δ whenever T (x) ∈ B). Prove that T is a bounded linear mapping. 9. Let T : X −→ Y be a sequentially continuous linear map between normed spaces such that ker(T ) is located. Prove that if x0 ∈ X and T (x0 ) = 0, then ρ(x0 , ker(T )) > 0. 10. Let X be a metric space, and let S, T be subsets of X such that for each s ∈ S, each t ∈ T, and each ε > 0, either s = t or there exists y ∈ S ∩T with ρ(t, y) < ε. Prove that if S ∩ T is complete, then S and T intersect sharply in the following sense: if x ∈ ∼ (S ∩ T ), then for each s ∈ S and each t ∈ T, either x = s or x = t. 11. A sequence (xn )n1 in a metric space is said to be weakly discriminating if for all positive a, b with a < b, either ρ (xn , x1 ) < b for all n or else ρ(xn , x1 ) > a for some n. Prove that every totally bounded sequence in a metric space is weakly discriminating. Let f : X −→ Y be a function between metric spaces that maps Cauchy sequences to weakly discriminating sequences. Prove that f is strongly extensional. 12. A sequence (xn )n1 in a metric space is called an LEM-Cauchy sequence if ¬¬ (xn )n1 is a Cauchy sequence . Prove that the following are equivalent conditions on a mapping f between metric spaces: (a) f maps convergent sequences to weakly discriminating LEM-Cauchy sequences. (b) f is sequentially continuous. 13. Let f be a mapping of a complete metric space X into a metric space Y. Suppose that f is strongly extensional and that it is sequentially discontinuous at some point x ∈ X in the following sense: there exist ε > 0 and a sequence (xn )n1 converging to x such that ρ(f (x), f (xn )) > ε for each n. Prove that LPO holds. 14. Let f be a strongly extensional mapping of a complete metric space X into a compact metric space Y, and let (xn )n1 be a sequence converging to x in X. Prove that the sequence (f (xn ))n1 has a convergent subsequence. (Hint: First prove that if LPO holds, then every sequence in a compact metric space has a convergent subsequence. Then use Ishihara’s tricks and Exercise 13.)

78

3 The λ-Technique

15. Let (Un )n1 be a sequence of located open sets in a complete metric space X Un is inhabited. Prove that there such that the metric complement of U = n1

exist a point x∞ in −U and an increasing binary sequence (λn )n1 such that for each n, if λn = 0, then x∞ ∈ Un , and if λn = 1, then −Uk is inhabited for some k n. Use this result to prove that Markov’s principle is equivalent to the following statement (which is classically equivalent to Baire’s theorem): If (Un )n1 is a sequence of located open subsets of a complete metric space X such that Un is inhabited, then there exists n such that −Un is inhabited. − n1

Notes A variation of the argument at the beginning of the chapter shows that if for each z ∈ C, there exists θ ∈ [0, 2π) such that z = |z| eiθ , then LLPO holds. The classical proposition “sequential continuity implies pointwise continuity” is equivalent to the essentially nonconstructive principle BD-N that we discuss later, in Section 6.3. When taken with the Church–Markov–Turing thesis, Exercise 13 shows that it is impossible that there exist a sequentially discontinuous function from R to R, since LPO is false in RUSS. It is, however, a far cry from showing that the existence of a sequentially discontinuous mapping on R implies LPO to proving that every mapping f : R −→ R is pointwise continuous. The latter can be done with the aid of either Brouwer’s continuity principle or else both the Church–Markov–Turing thesis and Markov’s principle; see Chapters 3 and 5 of [34]. In connection with Exercise 1, Fred Richman has shown us the following choicefree proof that if Ra is complete, then either a = 0 or a = 0. Assume that Ra is complete. For each ε > 0 we have either |a| < ε or else |a| > 0; in the latter case,

a |a| = ± ∈ Ra. |a|

Since ε > 0 is arbitrary, |a| ∈ Ra = Ra and there exists r such that |a| = ra. Pick a positive integer N > r. Either |a| > 0 or |a| < 1/N 2 . In the latter case, if a = 0, then |a| |r| |a| = |ra| = |a| = > N |a| , |a| so |r| > N, a contradiction; whence a = 0. Exercise 3 and an extended version of Proposition 3.1.6 originated in [24].

3.2 Ishihara’s Tricks

79

Exercise 6 deals with a weak converse of Bishop’s lemma. There seems to be no obvious use for such converses, in spite of the extreme value of Bishop’s lemma itself. The notion of “weakly discriminating”, and Exercises 11 and 12, come from [37]. Exercise 14 produces a constructive substitute for the highly nonconstructive Bolzano–Weierstrass theorem; it ﬁrst appeared in [30]. For more constructive analyses of the Bolzano–Weierstrass theorem see [27] and [22]. For more on Baire’s theorem, see [34] (Chapter 2) and Chapter 6 below.

4 Finite-Dimensional and Hilbert Spaces

A mind that is stretched by a new experience can never go back to its old dimensions. —Oliver Wendell Holmes, attrib.

We ﬁrst examine ﬁnite-dimensional spaces, including an application of the λ-technique to the problem of ﬁnding best approximations by elements of a ﬁnite-dimensional subspace. We then introduce Hilbert spaces, which are natural generalisations of ﬁnite-dimensional Euclidean spaces.

4.1 Finite-Dimensional Spaces Let X be a linear space equipped with a compatible inequality. Finitely many n λi ei = 0 for all vectors e1 , . . . , en in X are said to be linearly independent if scalars λ1 , . . . , λn such that n i=1

n i=1

i=1

|λi | > 0. In that case, if the λi are scalars such that

λi ei = 0, then λi = 0 for each i. We say that the space X is ﬁnite-dimensional

if either X = {0} or else it contains ﬁnitely many linearly independent vectors e1 , . . . , en such that for each x ∈ X there exist scalars λ1 , . . . , λn for which x=

n

λi ei .

(4.1)

i=1

In the ﬁrst case we say that X has dimension 0 or is 0-dimensional. In the second case we say that X has dimension n or is n-dimensional, we call {e1 , . . . , en } a basis of/for X, and we say that the space X is spanned by, or is the (linear) span of, the set {e1 , . . . , en } . We denote the dimension of a ﬁnite-dimensional space X

82

4 Finite-Dimensional and Hilbert Spaces

by dim(X). Since any two bases of X have the same number of vectors, dim(X) is well deﬁned. If {e1 , . . . , en } is a basis for X, then for each x ∈ X, the coordinates λi in the representation (4.1) are uniquely deﬁned; so there are well-deﬁned, clearly linear, n ui (x)ei for each x in X. We coordinate functionals ui : X −→ K such that x = i=1

shall prove that each of the coordinate functionals is a bounded linear map relative to any norm on X. Lemma 4.1.1. Let e be a nonzero vector in a normed space X. Then the subspace Y = Ke is locally compact, and ∼Y = −Y. −1

Proof. The mapping λ λ e e is an isometric isomorphism of K onto Y, so Y is locally compact and therefore located. Since locally compact spaces are complete, it follows from Bishop’s lemma (Proposition 3.1.1) that ∼Y = −Y. 2

Lemma 4.1.2. If Y is a ﬁnite-dimensional subspace of a normed space X, then Y is located and ∼Y = −Y. Proof. The case dim(Y ) = 0 is trivial, and the case dim(Y ) = 1 is dealt with in Lemma 4.1.1. Assume that all subspaces of dimension at most n 1 in all normed spaces are located, and that their complements and metric complements coincide. Consider an (n + 1)-dimensional space Y with basis vectors e1 , . . . , en+1 . Let Z be the n-dimensional space spanned by {e1 , . . . , en } . First note that for all λ1 , . . . , λn n in K, since 1 + |λi | > 0, the linear independence of the vectors e1 , . . . , en+1 gives en+1 =

n

i=1

i=1

λi ei . Hence en+1 belongs to ∼Z and therefore, by our induction

hypothesis, to −Z; whence en+1 X/Z = ρ (en+1 , Z) > 0. It follows that Ken+1 is a 1-dimensional subspace of X/Z. Now, for each x ∈ X, ρ (x, Y ) = inf {x − ten+1 − z : t ∈ K, z ∈ Z} = inf {ρ (x − ten+1 , Z) : t ∈ K} = inf x − ten+1 X/Z : t ∈ K , which exists, since, by Lemma 4.1.1, the 1-dimensional subspace Ken+1 is located in X/Z. Hence Y is located in X. Now consider an element x of ∼Y. For all t ∈ K and z ∈ Z we have x = ten+1 +z and therefore x − ten+1 = z. Thus x − ten+1 ∈ ∼Z, and therefore, by our induction hypothesis,

4.1 Finite-Dimensional Spaces

83

x − ten+1 X/Z = ρ (x − ten+1 , Z) > 0. This shows that x =X/Z λen+1 ; whence x is in the complement of Ken+1 relative to the quotient norm on X/Z. Again applying our induction hypothesis, this time to the 1-dimensional subspace Ken+1 of X/Z, and denoting by ρX/Z the distance corresponding to the quotient norm on X/Z, we see that 0 < ρX/Z (x, Ken+1 ) = ρ (x, Y ) , so x ∈ −Y. Hence ∼Y ⊂ −Y and therefore ∼Y = −Y. This completes the induction step. 2 Proposition 4.1.3. Let {e1 , . . . , en } be a basis for an n-dimensional normed space X. Then the corresponding coordinate functionals are bounded linear functionals. Proof. In the case n = 1, the sole coordinate functional is the linear mapping λe1 λ, which, being an isometry (that is, distance preserving), is trivially continuous. n ui (x)ei . In order to prove Now consider the case n 2. For each x ∈ X write x = i=1

that the coordinate functional uk is bounded, we may relabel, if necessary, to take k = n. Let Z be the (n − 1)-dimensional subspace of X with basis {e1 , . . . , en−1 } . Since, as in the proof of Lemma 4.1.2, en ∈ ∼Z, it follows from that lemma that ρ (en , Z) > 0 and hence that Ken is a 1-dimensional subspace of X/Z. Moreover, for each x ∈ X we have n ui (x)ei , Z = ρ (un (x)en , Z) xX/Z = ρ i=1

= un (x)en X/Z = |un (x)| en X/Z and therefore |un (x)| =

1 1 xX/Z x . en X/Z en X/Z

Thus 1/ en X/Z is a bound for the linear functional un .

2

Corollary 4.1.4. Every linear mapping from a ﬁnite-dimensional normed space into a normed space is bounded. Proof. Let u be a linear mapping of a ﬁnite-dimensional normed space X into a normed space Y. We may assume that dim(X) > 0. Let {e1 , . . . , en } be a basis of X, and, using Proposition 4.1.3, compute a common bound c > 0 for the corresponding coordinate functionals ui (1 i n). For each x ∈ X we have

84

4 Finite-Dimensional and Hilbert Spaces

n n u(x) = u ui (x)ei = ui (x)u(ei )

n

i=1

ui (x)u(ei ) =

i=1

c

n

n

i=1

|ui (x)| u(ei )

i=1

u(ei ) x .

i=1

2

So u is a bounded linear map.

Corollary 4.1.5. Let {e1 , . . . , en } be a basis for an n-dimensional normed space X. Then the canonical mapping (λ1 , . . . , λn )

n

λi ei

(4.2)

i=1

is a bounded linear injection of the Euclidean space Kn onto X, and its inverse is a bounded linear injection. Proof. It is routine to verify the linearity of this map and its inverse, and that the maps are injective. The continuity of the mappings follows immediately from Corollary 4.1.4. 2 Proposition 4.1.6. A ﬁnite-dimensional normed space is locally compact. Proof. The 0-dimensional case is trivial. Consider a ﬁnite-dimensional normed space X with a basis {e1 , . . . , en }. Let u : Kn −→ X be the bounded linear mapping deﬁned at (4.2). By Corollary 4.1.5, u−1 is a bounded linear map. Hence for any bounded subset B of X, u−1 (B) is a bounded subset of the locally compact space Kn and is therefore contained in a compact subset K of Kn . Since u is injective and its inverse mapping is uniformly continuous, it is strongly injective. It follows from Proposition 2.2.16 that u(K), which clearly contains B, is a compact subset of X. 2 Corollary 4.1.7. The unit ball of a ﬁnite-dimensional normed space is compact. Proof. Let X be a ﬁnite-dimensional normed space, and B its unit ball. By Proposition 4.1.6, there exists a compact subset K of X that contains B. Now, for each x ∈ X, ρ(x, B) = max {0, 1 − x} exists, so B is totally bounded (by Proposition 2.2.10). Being also closed in K, it is complete and therefore compact. 2

4.1 Finite-Dimensional Spaces

85

Corollary 4.1.8. Every linear mapping from a ﬁnite-dimensional normed space into a normed space is normed. Proof. Let u be a linear mapping from a ﬁnite-dimensional normed space X into a normed space Y. We may assume that dim(X) 1. By Corollary 4.1.4, u is a bounded linear mapping and hence is uniformly continuous on X. In particular, this implies that u maps the compact (by Corollary 4.1.7) unit ball of X onto a totally bounded subset of Y ; so u = sup {u (x) : x ∈ X, x 1} 2

exists.

Two norms , on a vector space X are said to be equivalent if both the identity mapping from (X, ) onto (X, ) and its inverse are continuous; since those mappings are linear, it follows from Proposition 2.3.3 that and are equivalent norms on X if and only if there exist positive constants a, b such that a x x b x for all x ∈ X. Corollary 4.1.9. Any two norms on a ﬁnite-dimensional space are equivalent.

Proof. If and are two norms on a ﬁnite-dimensional space X, then, by Corollary 4.1.4, both the identity mapping from (X, ) to X, and its inverse are bounded, and hence continuous, linear mappings. 2 We want to prove the converse of Corollary 4.1.7. This requires three lemmas, the second of which will have several applications later in the book. Lemma 4.1.10. Let X be a normed space, Y an n-dimensional subspace of X with basis {e1 , . . . , en }, and e a vector such that ρ(e, Y ) > 0. Then the linear span of Y ∪ {e} is (n + 1)-dimensional, with basis {e1 , . . . , en , e}. Proof. We need only prove that the vectors e1 , . . . , en , e are linearly independent. n+1 |λi | > 0. Either To this end, consider elements λi (1 i n + 1) of K such that λn+1 = 0 or

n i=1

i=1

|λi | > 0. In the ﬁrst case,

n λi ei + λn+1 e |λn+1 | ρ (e, Y ) > 0. (4.3) i=1 n λ e In the second case, since e1 , . . . , en are linearly independent, i i > 0. Hence i=1 either λn+1 e > 0, so that λn+1 = 0 and we have (4.3); or else λn+1 e < n λi ei and therefore i=1

86

4 Finite-Dimensional and Hilbert Spaces

n n λi ei + λn+1 e λi ei − λn+1 e > 0. i=1

i=1

n

Thus in all cases we have

i=1

λi ei + λn+1 e = 0.

2

Lemma 4.1.11. Let S be the span of a ﬁnitely enumerable set {x1 , . . . , xn } in a normed space X, and let ε > 0. Then there exists a ﬁnite-dimensional subspace Y of S such that ρ (xi , Y ) < ε for each i. Proof. Setting X0 = {0} , suppose that for some k < n we have constructed ﬁnitedimensional subspaces X0 ⊂ X1 ⊂ · · · ⊂ Xk ⊂ S such that ρ (xi , Xk ) < ε for 1 i k. Then either ρ (xk+1 , Xk ) < ε, in which case we set Xk+1 = Xk , or else ρ (xk+1 , Xk ) > 0. In the latter case we take Xk+1 to be the span of Xk ∪ {xk+1 } , which is ﬁnite-dimensional by the preceding lemma. This completes the inductive construction of the ﬁnite-dimensional subspace Xk+1 . It remains to take Y = Xn . 2

Lemma 4.1.12. (Riesz’s lemma) Let Y be a closed located subspace with an inhabited metric complement in a normed space X, and let 0 < θ < 1. Then there exists a unit vector x ∈ X such that x − y > θ for each y ∈ Y. Proof. Fix x0 ∈ −Y. Then 0 < r = ρ(x0 , Y ) < θ−1 r. Choosing y0 ∈ Y such that r x0 − y0 < θ−1 r, let x=

1 (x0 − y0 ). x0 − y0

Then x = 1. Also, for each y ∈ Y, y0 + x0 − y0 y ∈ Y. Hence x0 − y0 x − y = x0 − (y0 + x0 − y0 y) ρ(x0 , Y ) = r and therefore x − y

r > θ. x0 − y0 2

4.1 Finite-Dimensional Spaces

87

Proposition 4.1.13. A locally totally bounded normed space is ﬁnite-dimensional. Proof. Let B be the unit ball of a locally totally bounded normed space X. Then B is located in X. By Proposition 2.2.18, B is locally totally bounded; being bounded, it is therefore totally bounded. Let {x1 , . . . , xn } be a 1/4-approximation to B, and, using Lemma 4.1.11, construct a ﬁnite-dimensional subspace Y of X such that ρ (xi , Y ) < 1/4 for each i. Let ξ be any point of X, and suppose that ρ (ξ, Y ) > 0. Then, by Riesz’s lemma, there exists a unit vector x ∈ X such that x − y > 1/2 for all y ∈ Y. But this is absurd: for since x ∈ B, there exists i such that x − xi < 1/4 and therefore ρ(x, Y ) < 1/2. We conclude that ρ (ξ, Y ) = 0; so ξ is in the closure of Y. But Y, being ﬁnite-dimensional and therefore locally compact (Proposition 4.1.6), is closed in X. Hence ξ ∈ Y. Since ξ ∈ X is arbitrary, 2 we see that X = Y. Corollary 4.1.14. A located subspace of a ﬁnite-dimensional normed space is ﬁnite-dimensional. Proof. Let Y be a located subspace of a ﬁnite-dimensional normed space X. Since, by Proposition 4.1.6, X is locally compact, we see from Proposition 2.2.18 that Y is locally totally bounded. Hence, by Proposition 4.1.13, Y is ﬁnite-dimensional. 2

By a convex combination of ﬁnitely many elements x1 , . . . , xn of X we mean n n λi xi where each λi 0 and λi = 1. The convex hull a point of the form i=1

i=1

of a subset S of X is the closure of the set of all points of X that are convex combinations of points of S. It is straightforward to show that the convex hull of S is the intersection of all convex subsets of X that contain S. According to Exercise 11 of Chapter 2, a located subset of R may not have its metric complement located. However, things are diﬀerent when the set is convex. Proposition 4.1.15. If S is a located convex subset of the product normed space Rn such that −S is inhabited, then −S is located. Proof. Consider any x in Rn , and for each r > 0 let B 1 (x, r) be the closed ball with centre x and radius r relative to the product metric ρ. Then B 1 (x, r) is an n-dimensional cube with centre x and sides of length 2r. Consider any two real numbers α, β such that α < β. Let v1 , . . . , v2n be the vertices of B 1 x, 12 (α + β) . It is left as an exercise to show that there exists δ > 0 such that for all points w1 , . . . , w2n with ρ(vi , wi ) < δ (1 i 2n ) , the convex hull of {w1 , . . . , w2n } contains B 1 (x, α). Either ρ(vi , S) < δ for all i, or ρ(vi , S) > 0 for some i. In the ﬁrst case, for each i in {1, . . . , 2n } choose wi in S such that ρ(vi , wi ) < δ; then S contains the convex hull of {wi , . . . , w2n } and

88

4 Finite-Dimensional and Hilbert Spaces

therefore contains B 1 (x, α). Thus ρ(x, y) α for all y in −S. On the other hand, if for some i we have ρ(vi , S) > 0, then vi ∈ −S and ρ(x, vi ) 12 (α + β) < β. Thus ρ(x, S) exists, by the constructive least-upper-bound principle. 2 Although we do not develop integration theory in this book, it is worth pointing out here that for convex subsets of Rn there is a close link between locatedness and Lebesgue measurability: a convex subset of Rn with inhabited interior is located if and only if it is Lebesgue measurable [17]. Informally, this result shows that a convex subset of Rn with inhabited interior can be located if and only if its size can be calculated.

4.2 Best Approximation Let V be an inhabited subspace of a metric space X, and let a, b be elements of X, V respectively. We say that b is a best approximation, or closest point, to a in V if ρ (a, b) ρ (a, v) for each v ∈ V. In that case, ρ (a, V ) exists and equals ρ (a, b). We call V proximinal in X if each element of X has a best approximation in V , in which case V is located. The fundamental theorem of classical approximation theory says that a ﬁnitedimensional subspace of a normed space is proximinal. Constructively, this theorem implies LLPO (see Exercise 1). However, by introducing the idea of at most one object even without knowing in advance that there exists one, we can produce a good constructive version of the fundamental theorem. Let X, V, and a be as in the ﬁrst paragraph of this section. We say that a has at most one best approximation in V if for all distinct points v, v in V, there exists x ∈ V such that max {ρ(a, v), ρ(a, v )} > ρ(a, x). We call V quasiproximinal if each point of X with at most one best approximation in V actually has a (perforce strongly unique) best approximation in V. Our destination in this section is the following constructive fundamental theorem of approximation theory. Theorem 4.2.1. Every ﬁnite-dimensional subspace of a real normed space is quasiproximinal. We defer the proof until we have prepared the pathway with some preliminary results, one of which has a proof that uses the λ-technique discussed in Chapter 3. Lemma 4.2.2. Let {e1 , . . . , en } be a basis for an n-dimensional subspace V of a normed space X, let 1 m < n, and let W be the subspace of X with basis {e1 , . . . , em } . Then the span of {em+1 , . . . , en } is an (n − m)-dimensional subspace of the quotient space X/W.

4.2 Best Approximation

Proof. The proof is relegated to Exercise 4.

89

2

Lemma 4.2.3. Let x, e be elements of a real normed space X with e = 0, and for each δ > ρ (x, Re) write Sδ = {t ∈ R : x − te δ} . If Sδ is compact, then it is a proper compact interval [m, M ] in R. Moreover, x − me = δ = x − M e .

(4.4)

Proof. Suppose that Sδ is compact, with inﬁmum m and supremum M. By Corollary 2.2.14, there exists δ such that ρ(x, Re) < δ < δ and Sδ is compact. Let m and M denote the inﬁmum and supremum, respectively, of Sδ . The uniform continuity of the map t x − te on R ensures that (4.4) and x − m e = δ = x − M e hold. Since Sδ ⊂ Sδ and δ < δ, it follows that m < m M < M and therefore m < M. Now consider any t ∈ [m, M ] . Writing τ=

t−m 0, M −m

we have t = (1 − τ ) m + τ M. Hence x − te (1 − τ ) x − me + τ x − M e = (1 − τ ) δ + τ δ = δ and so t ∈ Sδ . Thus Sδ = [m, M ] .

2

Lemma 4.2.4. Let x, e be elements of a real normed space X with e = 0, and let d 0. Suppose that max {x − te , x − t e} > d whenever t, t are distinct real numbers. Then there exists τ ∈ R such that x − τ e > d entails ρ (x, Re) > d. Proof. Without loss of generality, we may assume that ρ (x, Re) < d + 1. With Sδ as in Lemma 4.2.3, we construct an increasing binary sequence (λn )n1 , sequences (an )n1 , (bn )n1 of real numbers, and a decreasing sequence (δ n )n1 of positive numbers with the following properties: Sδn is the proper compact interval [an , bn ] ; if λn = 0, then ρ (x, Re) < δ n < d + 1/n, and if n 2, 0 < bn − an

2 (bn−1 − an−1 ) ; 3

(4.5)

90

4 Finite-Dimensional and Hilbert Spaces

if λn = 1, then ρ (x, Re) > d, an = an−1 , bn = bn−1 , and δ n = δ n−1 . To begin the construction, set λ1 = 0 and, using Corollary 2.2.14, choose δ 1 > 0 such that ρ (x, Re) < δ 1 < d+1 and Sδ1 is compact. By Lemma 4.2.3, Sδ1 is a proper compact interval [a1 , b1 ]. Now suppose we have constructed λn−1 , an−1 , bn−1 , and δ n−1 with the applicable properties. If λn−1 = 1, set λn = 1, an = an−1 , bn = bn−1 , and δ n = δ n−1 .

(4.6)

If λn−1 = 0, write 2 1 2 1 an−1 + bn−1 e , x − an−1 + bn−1 e r = max x − . 3 3 3 3 By our hypotheses, r > d; so either ρ (x, Re) > d or else ρ (x, Re) < min {r, d + 1/n} . In the ﬁrst case we deﬁne our numbers as at (4.6). In the second case we set λn = 0 and choose δ n such that 1 ρ (x, Re) < δ n < min δ n−1 , r, d + n and Sδn is compact; then by Lemma 4.2.3, Sδn is a proper compact interval [an , bn ] contained in Sδn−1 . Since r > δ n , either x − 1 an−1 + 2 bn−1 e > δ n 3 3 x − 2 an−1 + 1 bn−1 e > δ n . 3 3

or

Taking, for example, the ﬁrst case, and using the convexity of Sδn , we see that either 1 2 Sδn ⊂ an−1 , an−1 + bn−1 3 3 or 1 2 an−1 + bn−1 , bn−1 . Sδ n ⊂ 3 3 Hence (4.5) holds. This completes our inductive construction. We see from (4.5) that for n 2, if λn = 0, then n−1 2 (b1 − a1 ) 0 < bn − an 3 and therefore 0 an+1 − an

n−1 2 (b1 − a1 ) . 3

(4.7)

4.2 Best Approximation

91

Clearly, this last inequality also holds if λn = 1. It follows that (an )n1 is a Cauchy sequence and therefore converges to a limit τ in R. Supposing that x − τ e > d, compute a positive integer N such that x − aN e > d + 1/N. If λN = 0, then x − aN e = δ N < d +

1 , N

a contradiction. We conclude that λN = 1 and hence that ρ (x, Re) > d.

2

We now have the proof of Theorem 4.2.1. Proof. Let V be a ﬁnite-dimensional subspace of a normed space X over R, and let a be a point of X with at most one best approximation in V. If V has dimension 0, then there is nothing to prove. If dim(V ) = 1, then V = Re for some e = 0 in V, and we can apply Lemma 4.2.4 with x = a and d = ρ (a, V ) to construct τ ∈ R such that a − τ e = ρ (a, V ). Now let n be a positive integer, and suppose we have proved the desired result for all n-dimensional subspaces of real normed spaces. Consider the case where V has a basis {e1 , . . . , en+1 } , and let Y = Ren+1 . By Lemma 4.2.2, V /Y is an ndimensional subspace of X/Y with basis {e1 , . . . , en }; moreover, for each x ∈ X we have (4.8) ρ (x, V ) = inf x − vX/Y : v ∈ V . Next, note that (*) for each v ∈ V there exists α ∈ R such that if a − v − αen+1 > ρ (a, V ) , then a − vX/Y = ρ (a − v, Ren+1 ) > ρ (a, V ) . For if t, t are distinct real numbers, then (v + ten+1 ) − (v + t en+1 ) = |t − t | en+1 = 0 and therefore, by our hypotheses, max {a − v − ten+1 , a − v − t en+1 } > ρ (a, V ) . So we can apply Lemma 4.2.4 with x = a − v, e = en+1 , and d = ρ (a, V ) to compute the desired α. With v, α as above, now let v be a point of V distinct from v, and compute α ∈ R such that if a − v − α en+1 > ρ (a, V ) , then

a − v X/Y = ρ (a − v , Ren+1 ) > ρ (a, V ) . We have

92

4 Finite-Dimensional and Hilbert Spaces

(v + αen+1 ) − (v + α en+1 ) = (v − v ) + (α − α ) en+1 v − v X/Y > 0, so, by our hypotheses, max {a − v − αen+1 , a − v − α en+1 } > ρ (a, V ) . It follows from (4.8) and the deﬁning properties of α, α that max a − vX/Y , a − v X/Y > ρ(a, V ) = inf a − vX/Y : v ∈ V . We have now shown that a has at most one best approximation in the n-dimensional subspace V /Y of X/Y. By our induction hypothesis and (4.8), there exists v0 ∈ V such that ρ (a − v0 , Ren+1 ) = a − v0 X/Y = inf a − vX/Y : v ∈ V = ρ (a, V ) . Applying (*) once more, we compute t0 ∈ R such that a − v0 − t0 en+1 = ρ (a, V ) . Hence v0 + t0 en+1 is the best approximation to a with respect to the original norm on X, and our inductive proof is complete. 2

Theorem 4.2.1 is classically equivalent to its classical counterpart. For if V is a ﬁnite-dimensional subspace of the real normed space X and we suppose that some a ∈ X has no best approximation in V, then a clearly has at most one best approximation in X. Therefore, by Theorem 4.2.1, it has one, which is a contradiction. For applications of Theorem 4.2.1, see [15].

4.3 Hilbert Spaces As we did for metric and normed spaces, we assume that the reader is familiar with the deﬁnitions and elementary properties of an inner product and an inner product space. In particular, we are not going to prove that x = x, x deﬁnes the norm associated with a given inner product on a vector space X over K, and that the Cauchy–Schwarz inequality |x, y| x y and the parallelogram identity

4.3 Hilbert Spaces 2

2

2

93

2

x + y + x − y = 2 x + 2 y

then hold. The latter inequality enables us to prove that an inner product space X is uniformly convex, as follows: If 0 < δ < 1, and x, y are unit vectors in X with 1 2 x + y > 1 − δ, then 2

2

2

2

2

x − y = 2 x + 2 y − x + y < 4 − 4 (1 − δ) , which can be made as small as we please, independently of x and y, by a suitable initial choice of δ. The simplest example of an inner product space is, of course, the Euclidean space Kn with the usual inner product given by x, y =

n

xi yi∗ ,

i=1

where x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) . Another example is the space l2 (K), consisting of all sequences x = (xn )n1 in K that are square summable in the sense ∞ 2 that |xn | converges in R; in this case we work with termwise operations and n=1

with the inner product deﬁned by x, y =

∞

xn yn∗ .

n=1

It is left to the exercises to prove that this is indeed an inner product. An inner product space that is complete with respect to its norm is called a Hilbert space. Adding completeness to the inner product leads to a structure with powerful geometrical properties such as the following. Theorem 4.3.1. Let S be a closed, located subspace of a Hilbert space H. Then for each x ∈ H, there exists a strongly unique element P x of S such that x − P x = ρ(x, S). Moreover, P x is the strongly unique element y of S such that x − y, s = 0 for all s ∈ S. Proof. Fixing x ∈ H, let d = ρ (x, S) and choose a sequence (sn )n1 in S such that d = limn→∞ x − sn . Using the parallelogram identity and the convexity of S, we have 2

sm − sn = (sm − x) − (sn − x)

2

2 sm + sn 2 2 − x = 2 sm − x + 2 sn − x − 4 2 2

2

2 sm − x + 2 sn − x − 4d2 2 2 = 2 sm − x − d2 + 2 sn − x − d2 −→ 0 as m, n −→ ∞.

94

4 Finite-Dimensional and Hilbert Spaces

Hence (sn )n1 is a Cauchy sequence. Since H is complete and S is closed, this sequence converges to a limit P x ∈ S. The continuity of the norm on H now gives x − P x = d. Moreover, if y ∈ S and y = P x, then, again by the parallelogram identity and convexity, we have 2

2

0 < P x − y = (P x − x) − (y − x) 2 Px + y 2 2 − x = 2 P x − x + 2 y − x − 4 2 2 2 2 P x − x − d2 + 2 y − x − d2 2 = 2 y − x − d2 , so x − y > d. It follows that P x is the strongly unique closest point to x in S. Next note that for all y ∈ S and λ ∈ K, x − P x + λy, x − P x + λy d2 = x − P x, x − P x and therefore

2

2

|λ| y + 2 Re (λ∗ x − P x, y) 0.

(4.9)

Suppose that Re x − P x, y = 0. By choosing a suﬃciently small real number λ with λ Re x − P x, y < 0, we can contradict (4.9). It follows that Re x − P x, y = 0. A similar argument shows that Im x − P x, y = 0; whence x − P x, y = 0. Finally, if y ∈ S and y = P x, then since y − P x ∈ S, 2

0 < y − P x = y − P x, y − P x = x − P x, y − P x − x − y, y − P x = − x − y, y − P x , so x − y, y − P x = 0. It follows that P x is the strongly unique element y of S such that x − y, s = 0 for all s ∈ S. 2

The mapping P : H −→ S deﬁned in Theorem 4.3.1 is called the projection of H onto S, and for each x ∈ H, the vector P x is the projection of the vector x onto S. The mapping P is linear: for if λ ∈ K and x, x ∈ H, then since for all y ∈ S, λx + x − (λP x + P x ), y = λ x − P x, y + x − P x , y = 0, it follows from the uniqueness of the projection of λx + x onto S (Theorem 4.3.1) that P (λx + x ) = λP x + P x . For each y ∈ S, since y − y, s = 0 for all s ∈ S, we see from the uniqueness of the projection of y into S that P y = y; whence P maps H onto S. Since P x ∈ S, we have P 2 x = P (P x) = P x, so P is idempotent; that is, P 2 = P. Also,

4.3 Hilbert Spaces

95

2

x = x, x = x − P x + P x, x − P x + P x = x − P x, x − P x + x − P x, P x + P x, P x 2

2

= x − P x + P x , since x − P x, P x = 0 (by Theorem 4.3.1). Hence P x x , and so 1 is a bound for P. On the other hand, if S contains a nonzero vector x, then P x = x ; it follows that in this case, P is normed and P = 1. Two subsets S, T of an inner product space are said to be orthogonal if x, y = 0 for all x ∈ S and y ∈ T ; we then write S ⊥ T. The orthogonality relation is symmetric, S ⊥ T ⇐⇒ T ⊥ S, and any family (Ti )i∈I of subsets of X satisﬁes Ti ⇐⇒ ∀i ∈ I (S ⊥ Ti ) . S⊥ i∈I

A vector x is orthogonal to the subset S if {x} ⊥ S, in which case we write x ⊥ S; two vectors x, y are said to be orthogonal vectors if x ⊥ {y}, in which case we write x ⊥ y and we have the following generalisation of Pythagoras’s theorem: 2

2

2

x + y = x + y . We deﬁne the orthogonal complement of a subset S of X to be the set S ⊥ = {x ∈ X : x ⊥ S} , which is easily seen to be a closed linear subspace of X. If 0 ∈ S, then S ∩ S ⊥ = {0} . It readily follows that if S is a closed, located subspace of a Hilbert space H, with P the corresponding projection, and if x ∈ H, then the decomposition x = P x + (x − P x) is the unique expression of x as the sum of a vector in S and a vector orthogonal to S. Denoting by I the identity operator x x on H, we prove that I − P is the projection of H onto the subspace S ⊥ . For each y ∈ S ⊥ , the vector z = x − P x − y belongs to S ⊥ , so 2

2

2

2

2

2

x − y = P x + z = P x + z P x = x − (x − P x) . Since x − P x ∈ S ⊥ , we conclude from the uniqueness part of Theorem 4.3.1 that x − P x is the projection of x into S ⊥ . We need at this point to clarify what we mean by saying that the series λi ei i∈I

converges to the sum x and by writing x=

i∈I

λi ei ,

(4.10)

96

4 Finite-Dimensional and Hilbert Spaces

where the index set I on the right is not necessarily countable: we mean that for each ε > 0 there exists a ﬁnitely enumerable subset F of I with the property that if G is a ﬁnitely enumerable subset of I that contains F, then λi ei < ε. x − i∈G

In the case I = N+ , this condition is equivalent to the usual one for convergence of ∞ λn en to x. the series n=1

A family (ei )i∈I of vectors in a Hilbert space H is said to be orthonormal if ei ⊥ ej whenever i = j, and for each i, either ei = 1 or ei = 0. Such a family is called an orthonormal basis if each vector x ∈ H can be written uniquely in the form (4.10) with (λi )i∈I a family of elements of K such that λi = 0 whenever ei = 0. The scalar λi is then called the ith coordinate of x relative to the orthonormal basis. If (ei )i∈I is a ﬁnitely enumerable orthonormal family in H, then it spans a ﬁnite= 1 for each i. dimensional subspace of H. To see this, wemay assume that ei |λi | > 0, and let x = λi ei . Choose Let (λi )i∈I be a family of scalars such that i∈I

j ∈ I such that λj = 0. Then λi ei , ej = λj ej , ej = λj = 0, x, ej =

i∈I

i∈I

so x = 0. Thus the vectors ei (i ∈ I) are linearly independent and therefore span a ﬁnite-dimensional space. Classically, using (an equivalent of) the axiom of choice, we can prove that every Hilbert space has an orthonormal basis of unit vectors. Constructively we avoid the axiom of choice by adding separability to the hypotheses on H, by relaxing the requirements to allow basis vectors to be 0, and by using the Gram–Schmidt orthogonalisation process embodied in the proof of our next result. Proposition 4.3.2. Every separable Hilbert space has a countable orthonormal basis. If (en )n1 is such a basis, relative to which x has coordinates α1 , α2 , . . . and ∞ y has coordinates β 1 , β 2 , . . . , then αn = x, en for each n, x, y = αn β ∗n , and 2

x =

∞ n=1

n=1

2

|αn | .

Proof. Let (an )n1 be a dense sequence in H. The idea of the proof is to construct inductively the orthonormal sequence (en )n1 to ensure that at stage n there exists

4.3 Hilbert Spaces

97

zn in the subspace Sn of H spanned by {e1 , . . . , en } such that an − zn < 1/n. For convenience, set e0 = 0 and S0 = {0}. Assume that en has already been constructed. Being ﬁnite-dimensional, Sn is closed and located, so the projection Pn of H onto Sn exists. Either an+1 − Pn an+1 < 1/ (n + 1) or else an+1 = Pn an+1 . In the ﬁrst case, set en+1 = 0. In the second, deﬁne a unit vector by en+1 =

1 (an+1 − Pn an+1 ) . an+1 − Pn an+1

Note that en+1 is orthogonal to e0 , . . . , en since (by Theorem 4.3.1) an+1 − Pn an+1 is orthogonal to Sn , and that an+1 = Pn an+1 + an+1 − Pn an+1 en+1 is in the subspace of H spanned by {e1 , . . . , en+1 }. This completes the induction. Now consider the vectors x, y ∈ H. For each positive integer k we have unique representations k αn en Pk x = n=1

and Pk y =

k

β n en ,

n=1

with αn = β n = 0 whenever en = 0. For n k, αn =

k

αi ei , en = Pk x, en = x, en − x − Pk x, en = x, en .

i=1

Likewise, β n = y, en . Thus αn and β n do not depend on any k = n. Now, Pk x, Pk y =

k

αn β ∗n .

n=1

Since the sequence (an )n1 is dense in X and ρ (ak , Sk ) < 1/k, we see that both ρ (x, Sk ) and ρ (y, Sk ) approach 0 as k −→ ∞; whence Pk x −→ x and Pk y −→ y as k −→ ∞. Thus ∞ αn β ∗n . x, y = lim Pk x, Pk y = k→∞

n=1 2

Finally, if we take y = x, we obtain x =

∞ n=1

2

|αn | .

2

An elementary lemma will enable us to characterise dimensionality in terms of orthonormal bases.

98

4 Finite-Dimensional and Hilbert Spaces

Lemma 4.3.3. Let S be a ﬁnite-dimensional subspace of a separable Hilbert space H, and let (en )n1 be an orthonormal basis of H. Then there exists a positive integer N such that ρ (en , S) > 0 whenever n N and en = 1. Proof. By Corollary 4.1.7, there exists a ﬁnite 1/2-approximation {x1 , . . . , xm } to ∞ 2 |xk , en | the unit ball of H. Since, by the preceding proposition, the series n=1

converges, |xk , en | −→ 0 as n −→ ∞; so we can compute N such that |xk , en |

1 , 4

so ρ (en , S) = en − P en en − xk − P en − xk > 0, as we required.

2

Proposition 4.3.4. Let (en )n1 be an orthonormal basis of a separable Hilbert space H. Then H is ﬁnite-dimensional if and only if en = 0 for all suﬃciently large n. Proof. If H is ﬁnite-dimensional, then taking S = H in Lemma 4.3.3, we obtain N such that en = 0 for all n N. If, conversely, such N exists, then H is the ﬁnite-dimensional space of all linear combinations of those vectors e1 , . . . , eN that are nonzero. 2

Let X be a linear space with an inequality compatible with its linear structure. We say that X is inﬁnite-dimensional if the complement of each ﬁnite-dimensional subspace of X is inhabited. It follows from Bishop’s lemma (Proposition 3.1.1) that a normed space X is inﬁnite-dimensional if and only if for each ﬁnite-dimensional subspace S of X there exists x ∈ X such that ρ (x, S) > 0. Proposition 4.3.5. Let (en )n1 be an orthonormal basis of a separable Hilbert space H. Then H is inﬁnite-dimensional if and only if en = 0 for inﬁnitely many n. Proof. Suppose that H is inﬁnite-dimensional, and consider any positive integer N. Let S be the ﬁnite-dimensional subspace of H consisting of all linear combinations

4.3 Hilbert Spaces

99

of e1 , . . . , eN , and let P be the projection of H onto S. There exists x ∈ H such that ρ(x, S) > 0 and therefore, by Proposition 4.3.2, ∞

2

2

2

|x, en | = x − P x = ρ (x, S) > 0.

n=N +1

Hence there exists n > N such that x, en = 0 and therefore, by the Cauchy– Schwarz inequality, en = 0. Conversely, if there exists a strictly increasing sequence (nk )k1 of positive integers such that enk = 0 for each k, consider any ﬁnite-dimensional subspace S of H. Choose N as in Lemma 4.3.3. Then choose k such that nk > N. Since enk = 1, 2 we must have ρ (enk , S) > 0.

We now introduce a construction of new Hilbert spaces from old. Let H1 , H2 be Hilbert spaces, and for all x = (x1 , x2 ) and y = (y1 , y2 ) in the product vector space H1 × H2 deﬁne x, y = x1 , y1 + x2 , y2 . This deﬁnes an inner product with respect to which H1 × H2 is a Hilbert space, called the direct sum of H1 and H2 , and denoted by H1 ⊕ H2 . This construction helps us to prove the most general case of our next theorem. We say that a linear functional u on a Hilbert space H is represented by a vector a ∈ H if u(x) = x, a for all x ∈ H. The following Riesz representation theorem tells us that a functional is representable by a vector in this way if and only if it is normed. Theorem 4.3.6. A bounded linear functional u on a Hilbert space H is normed if and only if there exists a ∈ H such that u(x) = x, a

(x ∈ H) .

(4.11)

In that case, a is strongly unique: if y ∈ H and y = a, then there exists x ∈ H such that u(x) = x, y . Moreover, u = a . Proof. Suppose ﬁrst that there exists a vector a with property (4.11). Then, by the Cauchy–Schwarz inequality, |u(x)| a x for each x ∈ H. On the other hand, given ε > 0, we have either a > 0 or a < ε. In the ﬁrst case, 1 a = a > a − ε. u a In the second, u(0) = 0 > a − ε. Since ε > 0 is arbitrary, it follows that u exists and equals a .

100

4 Finite-Dimensional and Hilbert Spaces

Suppose, conversely, that u is normed. To begin with, take u > 0. Then ker(u) is located, by Proposition 2.3.6. Let P be the projection of H on ker(u), choose y ∈ H with u (y) > 0, and deﬁne x0 =

1 (y − P y) . u(y)

Then u(x0 ) = 1, so x − u(x)x0 ∈ ker(u). Since x0 ⊥ ker(u), we have 2

0 = x − u(x)x0 , x0 = x, x0 − u(x) x0 , from which it follows that u(x) = x, a with a=

1

2 x0 .

x0

It remains to remove the condition that u > 0. In doing so, we need to bear in mind that we have not ruled out the possibility that H = {0}. To deal with this, we consider the direct sum H ⊕ K, on which we deﬁne a bounded linear functional v by v(x, ζ) = u(x) + ζ. We ﬁrst observe that u is represented by the vector a if and only if v is represented by the vector (a, 1). Since v > 0, it follows from the ﬁrst part of the proof that it will suﬃce to prove that v is normed. To this end, let 0 < α < β and set ε = 12 (α + β). If u > 0, then, by the ﬁrst part of the proof, u is represented by a unique vector a ∈ H; whence v is represented by (a, 1) and is therefore normed. So we may assume that u < ε. Then either u + 1 > α + ε or u + 1 < β. In the ﬁrst case, v(0, 1) = 1 > α and (0, 1) = 1. In the second case, for each (x, ζ) with (x, ζ) 1 we have |v(x, ζ)| |u(x)| + |ζ| u + 1 < β. It now follows from the least-upper-bound principle that v exists.

2

The second part of the foregoing proof contains an argument showing that the sum of two normed linear functionals on a Hilbert space is normed. Note that in general, the sum of two normed linear functionals on an arbitrary nontrivial normed space need not be normed. By an operator on a Hilbert space H we mean a linear mapping of H into itself. The set of bounded operators on H is denoted by B(H). Classically, the Riesz representation theorem enables us to prove, for a given element T of B(H), the existence of the adjoint T ∗ , which has the deﬁning property T x, y = x, T ∗ y

(x, y ∈ H) .

(4.12)

4.3 Hilbert Spaces

101

Indeed, given y ∈ H, we apply the Riesz representation theorem to the linear functional x T x, y to obtain a unique vector T ∗ y with the desired property. The constructive problem with this argument is that in order to apply Theorem 4.3.6, we require that the functional x T x, y be normed, which is something we cannot guarantee. In fact, as the following Brouwerian example shows, we cannot prove that a general normed operator on H has an adjoint. Let H be a complex Hilbert space with an orthonormal basis (en )n1 of unit vectors, and let (an )n1 be a binary sequence with at most one term equal to 1. ∞ Observe that for each x ∈ H the series an x, en converges absolutely. For, given ε > 0 and choosing N such that

n=1 ∞

n=N +1

have

k

|an x, en | =

n=N +1

n=1

an |x, en |

n=N +1

Hence the partial sums of ∞

k

∞ n=1

2

|x, en | < ε2 , for each k > N we

max

N +1nk

|x, en | < ε.

an |x, en | form a Cauchy sequence, and so the series

an x, en converges in C. Thus Tx =

∞

an x, en e1

n=1

deﬁnes a mapping—clearly an operator—from H to itself. It is left as an exercise to show that T is normed. Suppose that T ∗ exists. Then either T ∗ e1 = 0 or else T ∗ e1 < 1. In the ﬁrst case we can ﬁnd N such that T ∗ e1 , eN = 0. Suppose that aN = 0. If there exists n = N such that an = 1, then for all x, y ∈ H, T x, y = x, en e1 , y = x, y, e1 en and therefore

T ∗ y = y, e1 en .

(4.13)

∗

Hence T e1 , eN = 0, a contradiction. Thus an = 0 for all n = N and therefore for all n; whence T ∗ = 0, which is impossible since T ∗ e1 = 0. We conclude that aN = 1. In the case T ∗ e1 < 1, suppose that an = 1. Then (4.13) holds, so T ∗ e1 = en and therefore T ∗ e1 = 1, a contradiction. It follows that in this case we have an = 0 for all n. In view of this Brouwerian example, for any Hilbert space H and any not necessarily bounded operator T on H, we deﬁne the adjoint T ∗ , if it exists, by the equation (4.12); in which case we refer to T as jointed. It is then straightforward to show that T ∗ is an operator; that the adjoint of T ∗ is T ; and that any bound for T is one for T ∗ , and vice versa. Moreover, if S, T are jointed operators, then for each λ ∈ K, so are λS + T and ST, and

102

4 Finite-Dimensional and Hilbert Spaces ∗

(λS + T ) = λ∗ S ∗ + T ∗ , ∗ (ST ) = T ∗ S ∗ . For example, we have ST x, y = T x, S ∗ y = x, T ∗ (S ∗ y) = x, T ∗ S ∗ y . An operator T on H is called selfadjoint, or Hermitian, if T ∗ exists and equals T. The identity operator I is trivially selfadjoint. More generally, if P is the projection of H onto a closed, located subspace, then for all x, y ∈ H we have P x, y = P x, P y + P x, y − P y = P x, P y = P x, P y + x − P x, P y = x, P y , so P is selfadjoint. Conversely, if P is any bounded, idempotent, selfadjoint operator on H, then P is a projection. To see this, let V = {y ∈ H : P y = y} . It is clear that V is a linear subspace of H. If x ∈ H, then since P 2 x = P x, we have P x ∈ V. On the other hand, for each y ∈ V we have x − P x, y = x, y − P x, y = x, y − x, P y = 0, since P y = y. Hence x − P x is orthogonal to V, and therefore (since P x − y is in V) 2

2

x − y = x − P x + P x − y 2

2

2

= x − P x + P x − y x − P x . We now see that P x is a closest point to x in V. Hence V is located in H. Finally, the continuity of the bounded operator P ensures that V is closed; so P is the projection of H onto V. We end the chapter with an application of the Riesz representation theorem and a corollary, both due to Ishihara [58]. For this we need to know that a linear mapping T between normed spaces X, Y is deﬁned to be compact if T (B X (0, 1)) is a totally bounded subset of Y ; in that case, the norm of T exists, by Corollary 2.2.7. Every bounded linear mapping on a ﬁnite-dimensional normed space is compact. Proposition 4.3.7. Let T be a bounded linear mapping of a Hilbert space H into Cn , and for 1 i n let Pi : Cn −→ C be the ith projection mapping, deﬁned by Pi (z1 , z2 , . . . , zn ) = zi . Then T is compact if and only if Pi ◦ T is normed for each i.

4.3 Hilbert Spaces

103

Proof. If T is compact, then since the projection Pi is uniformly continuous, the set Pi ◦ T (B(0, 1)) is totally bounded, so Pi ◦ T exists. Suppose, conversely, that Pi ◦ T is normed for each i. By the Riesz representation theorem, for each i there exists ai ∈ H such that Pi ◦ T (x) = x, ai

(x ∈ H) .

Using Lemma 4.1.11, construct a ﬁnite-dimensional subspace X of H such that ρ(ai , X) < ε/2n for each i. Let P be the projection of H onto X. Since the restriction of T to X is a compact linear mapping, there exist x1 , x2 , . . . , xm in the unit ball BX of X such that {T x1 , T x2 , . . . , T xm } is an ε/2-approximation to T (BX ). Consider any x ∈ H with x 1. Working with the product norm on Cn , we have T (x − P x) = = =

n i=1 n i=1 n

|Pi ◦ T (x − P x)| x − P x, ai x, (I − P ) ai

i=1 n

n ε ε = . x ai − P ai < 2n 2 i=1 i=1

Choosing j such that T P x − T xj < ε/2, we now obtain T x − T xj T (x − P x) + T P x − T xj

0 is arbitrary, we conclude that T is a compact linear mapping on H. 2 Corollary 4.3.8. The sum of two compact operators on a Hilbert space is compact. Proof. Let S and T be compact operators on a Hilbert space H, and let ε > 0. Since S(BH ) and T (BH ) are totally bounded, it readily follows from Lemma 4.1.11 that there exists a ﬁnite-dimensional subspace X of H, with associated projection P, such that (4.14) (S + T )x − (P S + P T )x < ε (x ∈ BH ) . We may assume that X is nontrivial and so has a basis {e1 , . . . , en } . Writing P Sx =

n i=1

pi (x)ei , P T x =

n

qi (x)ei ,

i=1

we see from Proposition 4.3.7 that for each i, the linear functionals pi , qi on H are normed; whence pi + qi is normed (see the remark following the proof of the Riesz representation theorem). Since

104

4 Finite-Dimensional and Hilbert Spaces

(P S + P T ) (x) =

n

(pi + qi ) (x)ei ,

i=1

we see from Proposition 4.3.7 that P S + P T is compact. In view of (4.14), any ε-approximation to (P S + P T ) (BH ) is a 2ε-approximation to (S + T )(BH ). Hence S + T is compact. 2

Exercises 1. Let X be the space R2 with norm (x, y) = max {|x| , |y|} . Show that if the 1-dimensional subspace R (cos θ, sin θ) is proximinal in X for each θ ∈ R, then LLPO holds. 2. Let V be a nonzero linear subspace of a normed linear space X such that each x ∈ X has at most one closest point in V . Need V be located? 3. Prove that if every pair of nonzero vectors in R2 generates a ﬁnite-dimensional space, then LPO holds. 4. Prove Lemma 4.2.2. 5. Prove that a uniformly convex linear subspace of a normed space is proximinal. 6. A normed space X is said to be compactly generated if there exists a compact set K ⊂ X such that each point of X is a linear combination of ﬁnitely many points of K. Let X be a compactly generated Banach space, and Y a ﬁnite-dimensional subspace of X. Prove that either X = Y or else the metric complement of Y is inhabited. (Hint: Use the λ-technique.) 7. Let Y be a locally compact subspace of a complete metric space X such that each x ∈ X has a unique best approximation P x in Y. Prove that the mapping P is sequentially continuous on X. (Hint: First use Ishihara’s tricks.) 8. Let T be a bounded linear mapping of a normed space X onto a ﬁnitedimensional Banach space Y. Prove that there exists r > 0 such that BY (0, r) ⊂ T (BX (0, 1)). 9. Let T be a bounded linear mapping of a normed space X onto a ﬁnitedimensional Banach space Y. Prove that T is a compact linear mapping if and only if ker(T ) is located in X. 10. Let X be an inner product space. Prove that x = x, x deﬁnes a norm on X and that the Cauchy–Schwarz inequality |x, y| x y holds for all x, y ∈ X.

4.3 Hilbert Spaces

105

11. Let p 1, and let lp be the set of all sequences x = (xn )n1 in the ﬁeld K ∞ p that are p-summable, in the sense that |xn | converges. Equipping lp with n=1

termwise operations of addition and multiplication-by-scalars, prove that 1/p ∞ p |xn | xp = n=1

deﬁnes a norm on lp , and that lp is complete with respect to this lp -norm. In the case p = 2, show that the lp -norm arises from an inner product, as in Exercise 10. 12. Let S, T be orthogonal linear subspaces of a Hilbert space H such that S + T = {x + y : x ∈ S, y ∈ T } is dense in H. Prove that S and T are both located. 13. Let S be a linear subset of a Hilbert space H such that for each x ∈ H there exists y ∈ S with x − y orthogonal to S. Prove that S is closed and located. 14. Construct a Brouwerian counterexample to the statement that every selfadjoint operator on a Hilbert space H is normed. Does every normed operator on H have an adjoint? 15. Construct a Brouwerian counterexample to the statement that every selfadjoint operator on a Hilbert space that has located kernel also has located range. 16. Construct Brouwerian counterexamples to each of the following statements. (a) Every bounded linear functional on l2 (see Exercise 11) is compact. (b) If T is a bounded linear mapping of l1 into C2 such that Pi ◦ T is normed for i = 1, 2, then T is compact (where Pi is the ith projection of C2 onto C). 17. Construct a Brouwerian counterexample to the proposition that the sum of two compact linear mappings between normed spaces is compact. 18. Use the λ-technique as an alternative means of removing the restriction that u > 0 in the proof of the Riesz representation theorem. 19. Let R be an algebra of normed, jointed operators on a complex Hilbert space H, where the operation of multiplication on R is just the composition of operators. For each T ∈ R and each λ ∈ C deﬁne an operator T on the direct sum H ⊕ C by T (x, ζ) = (T x + λx, λζ) . Show that T is both selfadjoint and normed.

106

4 Finite-Dimensional and Hilbert Spaces

Notes Our deﬁnition of “linearly independent” is classically equivalent to the usual one: n λi ei = 0, then λi = 0 for each i. Constructively, the two deﬁnitions namely, if i=1

are equivalent if and only if MP holds. Bishop’s deﬁnition of “ﬁnite-dimensional” requires that the space have a norm from the outset. Our deﬁnition is more in the spirit of linear algebra. Theorem 4.2.1 can be applied to the case of Chebyshev approximation: best approximation of elements of C [0, 1], relative to the sup norm, by polynomials of degree at most n. However, in this special case a deeper analysis enables one to prove the existence of best approximations without the use of Theorem 4.2.1; moreover, that analysis reveals that the best approximation process is, as one would expect in a constructive context, continuous. See [15]. The induction step in the proof of Theorem 4.2.1 is a lot simpler if we are permitted to use Brouwer’s fan theorem; for we can then show that any uniformly continuous function from a compact metric space to the positive real numbers has a positive inﬁmum (see [34], Chapter 6). As Section 4.3 shows, the elements of Hilbert space theory require very few modiﬁcations to bring them into constructive line. However, we have to be careful about the requirement that the linear functionals be normed before we can apply the Riesz representation theorem, and about the possibility that a given bounded operator may not have an adjoint. For a diﬀerent proof of the Riesz representation theorem see [35]. Working with a direct sum, as was done in the proof of Theorem 4.3.6 in order to circumvent our inability to decide whether a space is trivial, is a useful technique in other applications; see [28]. In the proof that an idempotent, bounded, selfadjoint operator P on a Hilbert space is a projection, we used boundedness only to prove that the set V = {x ∈ H : P x = x} is closed. The Hellinger–Toeplitz theorem (see Chapter 6) shows that every selfadjoint operator P on H is sequentially continuous, a property strong enough to prove that the corresponding set S is closed. Thus, in fact, every idempotent selfadjoint operator on a Hilbert space is a (bounded linear) projection.

5 Linearity and Convexity

Every separation is a link. —Simone Weil, ‘Metaxu’, Gravity and Grace We begin the chapter by exploring some geometric aspects of convexity that are used later in the construction of one of the cornerstones of functional analysis: the separation theorem. In turn, this leads us to the Hahn–Banach extension theorem, which in its most general form for separable normed spaces produces only approximately norm-preserving extensions of normed linear functionals. We also give Ishihara’s version of the Hahn– Banach theorem, which provides a unique norm-preserving extension of a given normed linear functional in the case where the norm on the space is Gˆ ateaux diﬀerentiable. We then use the separation and Hahn–Banach theorems to explore the interplay between a normed space and its dual. In particular, we characterise certain linear functionals on spaces of bounded linear mappings. For our discussion of duality we develop the fundamentals of the theory of locally convex topological vector spaces.

5.1 Crossing Boundaries Suppose we start at a point ξ in the interior of a located subset C of a normed space X and move linearly towards a point z in the metric complement of C. Are we able to tell when we are crossing the boundary ∂C = C ∩ ∼C of C? In general, as is discussed in more detail in problems at the end of this chapter, the constructive answer is no. However, our geometric intuition suggests that when C is convex, we might succeed in pinpointing boundary crossing points. Our ﬁrst few lemmas are designed to lead us to a proof of the existence of the boundary crossing point, which depends continuously on the origin ξ and the terminus z of our path out of the convex set.

108

5 Linearity and Convexity

Lemma 5.1.1. Let C be a convex subset of a normed space X, let ξ ∈ C ◦ , and let r > 0 be such that B (ξ, r) ⊂ C. Let z = ξ, 0 < t < 1, and z = tξ + (1 − t) z. If the ball B(z, tr) intersects C, then B(z , t2 r) ⊂ C. Proof. Suppose that there exists y in B(z, tr) ∩ C. Let ζ ∈ B(z , t2 r) and 1 1 y + ζ . ξ = 1− t t Then

ξ − ξ = 1 − 1 z + 1 z − 1 − 1 y − 1 ζ t t t t 1 1 − 1 z − y + z − ζ t t 1 2 1 − 1 tr + t r < t t = r.

Hence ξ ∈ C. Since ζ = tξ + (1 − t) y, it follows by convexity that ζ ∈ C.

2

In the context of a vector space X over K we deﬁne intervals as follows: [x, y] = {tx + (1 − t) y : 0 t 1} , (x, y) = {tx + (1 − t) y : 0 < t < 1} , where x, y ∈ X. Lemma 5.1.2. Let C be an open convex subset of a normed space X such that C ∪ −C is dense in X, let ξ ∈ C, and let z ∈ −C. Then (C ∪ −C) ∩ [ξ, z] is dense in [ξ, z] . Proof. Fix ε > 0 and choose r such that 0

r/2 and y − z > r/2, so y = αξ + (1 − α) z for some α ∈ (0, 1) . Fix λ such that −α −ε , < λ < 0, (5.1) max ξ − z 1 − α and set

5.1 Crossing Boundaries

109

λ , λ−1 y1 = λξ + (1 − λ) y. t=

Then 0 < t < 1 and y = tξ + (1 − t) y1 . Also, y1 = [1 − (1 − λ)(1 − α)] ξ + (1 − λ) (1 − α) z, where, by (5.1), 0 < (1 − λ) (1 − α) < 1,

so y1 ∈ [ξ, z] . Now pick y ∈ C ∪ −C such that y − y < min r2 , t2 r . Take ﬁrst the case y ∈ C. By our choice of r, the point ξ + 1r (y − y ) belongs to C. Since C is convex, 1 1 r y + ξ + (y − y ) y = 1+r 1+r r also belongs to C. Moreover, y = and

1 r ξ+ y ∈ [ξ, y] ⊂ [ξ, z] , 1+r 1+r

y − y =

r r ξ − y ξ − z < ε. 1+r 1+r

We are left with the case y ∈ −C to dispose of. But then B(y, t2 r) intersects −C, so, by the preceding lemma, B(y1 , tr) ∩ C = ∅; whence y1 ∈ −C. Since y − y1 = |λ| ξ − y ε by (5.1), the proof is complete.

2

Lemma 5.1.3. Let X be a normed space, let x1 , x2 be distinct points of X, and let x3 = λx1 + (1 − λ) x2 with λ = 0, 1. For all α, β > 0, if x − x1 < α/|λ| and y − x2 < /β/|1 − λ|, then λx + (1 − λ) y − x3 < α + β. Proof. For such x and y we have λx + (1 − λ) y − x3 |λ| x − x1 + |1 − λ| y − x2 , from which the result follows almost immediately.

2

110

5 Linearity and Convexity

In the presence of convexity, the preceding, seemingly innocent, lemma turns out to be a powerful tool. Here is a ﬁrst example of its use. Lemma 5.1.4. If C is an inhabited, open, convex subset of a normed space X, then −C is dense in ∼C. Proof. Fixing ξ ∈ C, choose r > 0 such that B (ξ, r) ⊂ C. Consider any z ∈ ∼C and ε > 0. Setting ε t=1+ z − ξ and x2 = tz + (1 − t) ξ, we see that x2 − z = ε and that ξ=

1 −t z+ x2 . 1−t 1−t

It now suﬃces to show that x2 ∈ −C. Taking x1 = z, x3 = ξ, and λ = t/(t − 1) in Lemma 5.1.3, we see that if y − x2 < r (t − 1), then −t 1 0 such that B (ξ, r) ⊂ C and B (z, r) ⊂ −C, now observe that if r , t=1− ξ − z then 0 < t < 1 and ξ − (tξ + (1 − t) z) = r; whence tξ + (1 − t) z ∈ C. Thus m1−

r < 1. ξ − z

On the other hand, since z − (1 − t) ξ + tz = r and therefore zt ∈ −C, we see from (*) that m

r > 0. ξ − z

By the deﬁnition of m as an inﬁmum, the point h(z) = mξ + (1 − m) z belongs to the closure of C. Again applying Lemma 5.1.2, we can ﬁnd t ∈ (0, m) such that zt belongs to C ∪ −C and is arbitrarily close to zm . By the deﬁnition of / C, and so zt ∈ −C. Thus zm is in the closure of −C and hence m, we have zt ∈ in ∂C. The uniqueness part of (b) will follow immediately once we have proved (c) and (d).

112

5 Linearity and Convexity

Now observe that if m < t 1, then by the deﬁnition of m as an inﬁmum, there exists t such that m < t < t and zt ∈ C; since zt is in [ξ, zt ], it is in the convex set C. This proves (c). To dispose of (d), let 0 t < m and let η be the midpoint of the segment [ξ, zm ] . Then, by (c), η ∈ C, so there exists δ > 0 such that B (η, δ) ⊂ C. Also, η = λzm + (1 − λ) zt for some λ > 1. Consider any point y with y − zt < δ/(λ − 1). Applying Lemma 5.1.3 with x1 = zm , x2 = zt , and x3 = η, we see that λzm + (1 − λ) y − η < δ and therefore λzm + (1 − λ) y ∈ C. If also y ∈ C, then 1 1 y ∈ C, zm = (λzm + (1 − λ) y) + 1 − λ λ which is absurd since zm ∈ ∂C and C is open. Thus δ ⊂ ¬C B zt , λ−1 and therefore zt ∈ −C. This proves (d). To prove the continuity of the boundary crossing map on C × −C, ﬁx ξ ∈ C, z ∈ −C, and ε > 0. Using (c) and (d), choose a, b, s such that 0 < a < m < b < 1, s > 0, B (za , s) ⊂ −C ∩ B (zm , ε) , and B (zb , s) ⊂ C ∩ B (zm , ε) . Consider points z ∈ −C and ξ ∈ C with s max z − z , ξ − ξ < min 2

For each t ∈ [0, 1] set

1 1 1 1 , , , a b 1−a 1−b

.

zt = tξ + (1 − t) z .

Taking x1 = ξ, x2 = z, x = ξ , and y = z , and applying Lemma 5.1.3 with λ = a, we obtain za − za < s and therefore za ∈ −C ∩B(zm , ε). On the other hand, applying Lemma 5.1.3 with λ = b, we see that zb − zb < s and therefore zb ∈ C ∩ B(zm , ε). It follows that zγ(ξ ,x) is in the segment (za , zb ), which lies in B(zm , ε); whence 2 zγ(ξ ,z ) − zγ(ξ,z) < ε. For ﬁxed ξ ∈ C, we call the mapping z zγ(ξ,z) in the foregoing proposition the boundary crossing map of C relative to ξ. We shall use the existence of exact boundary crossings out of convex sets in the next section.

5.2 Separation Theorems

113

5.2 Separation Theorems Developing the theme of convexity, in this section we approach the fundamental theorems on the separation of points and convex sets by hyperplanes. In turn, this material will lead us in the next section to the Hahn–Banach theorem, one of the cornerstones of functional analysis. A subset C of a vector space X over K is called a cone if for all x, y ∈ C and all t > 0, both x + y and tx belong to C. In that case, C is convex. The closure of a cone is a cone, as is the intersection of two cones. If K is a convex subset of X, then the set c(K) = {tx : x ∈ K, t > 0} is a cone. For clearly, if x ∈ c(K), then tx ∈ c (K) for all t > 0; whereas if x1 , x2 ∈ c(K), then x1 = t1 y1 and x2 = t2 y2 for some y1 , y2 ∈ K and some t1 , t2 > 0, so x1 + x2 = (t1 + t2 ) z with z=

t2 t1 y1 + y2 ∈ K. t1 + t2 t1 + t2

We call c (K) the cone generated by the convex set K. If X is a normed space and K is open, then so is c(K). Lemma 5.2.1. Let K be a bounded located subset of a normed space X, a ∈ X, and let τ > 0. Then the set S = {tx + (1 − t) a : x ∈ K, 0 < t < τ } is located. Proof. Choose R > 0 such that x R for all x ∈ K. Fixing x0 ∈ X, note ﬁrst that for each t > 0, f (t) = ρ (x0 , {tx + (1 − t) a : x ∈ K}) exists and equals

tρ

t−1 1 x0 + a, K . t t

If also t > 0, then since x0 − tx − (1 − t) a x0 − t x − (1 − t ) a + (a + x) |t − t | for all x ∈ X, we see that f (t) f (t ) + (a + R) |t − t | . It follows that the mapping f is uniformly continuous on R+ ; whence ρ (x0 , S) exists as the inﬁmum of f on the totally bounded interval (0, τ ) . 2

114

5 Linearity and Convexity

Lemma 5.2.2. Let K be a bounded, located, convex subset of a normed space X such that ρ(0, K) > 0, and let a ∈ X be such that ρ(−a, c(K)) > 0. Then there exists r > 0 such that tx + (1 − t) a r for all x ∈ K and t > 0. Proof. Let δ=

ρ(0, K) . 3 (a + ρ(0, K))

Given x in K and t > 0, we have either |1 − t| > δ or |1 − t| < 2δ. In the ﬁrst case, t > δρ(−a, c(K)). x + (−a) tx + (1 − t) a = |1 − t| 1 − t In the case |1 − t| < 2δ, tx + (1 − t) a t x − |1 − t| a (1 − 2δ) ρ (0, K) − 2δ a 1 = ρ(0, K). 3 Setting

1 r = min δρ(−a, c(K)), ρ(0, K) > 0, 3

we see that tx + (1 − t) a r in either case.

2

Lemma 5.2.3. Let K be a bounded, located, convex subset of a normed space X such that ρ(0, K) > 0. Then c (K) is located. Proof. Fix x0 ∈ X. For each t > 0 and each x ∈ K we have x0 − tx |t| x − x0 . Hence ρ (x0 , tK) tρ (0, K) − x0 −→ ∞ as t −→ ∞. Compute τ > 1 such that ρ (x0 , tK) > ρ (x0 , K)

(t > τ − 1) .

(5.2)

Then d = inf {ρ (x0 , tK) : 0 < t < τ } exists, by the case a = 0 of Lemma 5.2.1. Since τ > 1 and therefore d ρ (x0 , K), 2 it follows from (5.2) that ρ (x0 , c (K)) exists and equals d.

5.2 Separation Theorems

115

Lemma 5.2.4. Let K and L be open cones in a normed space X such that K ∪ L is dense in X and K ⊂ ∼L. Then (a) K ⊂ −L and L ⊂ −K, (b) K ∪ −K and L ∪ −L are dense in X, and (c) K and L have a common boundary—namely, K ∩ L. If also L = {−x : x ∈ K}, then ∂K is a subspace of X. Proof. Since L ⊂ ∼∼L ⊂ ∼K, we see that K ∪ ∼K is dense; whence, by Lemma 5.1.4, K ∪−K is dense. On the other hand, K ⊂ ∼L and K is open, so K ⊂ −L and therefore K ⊂ −L. But −L is open and K ∪ L is dense, so −L ⊂ K and therefore, by Lemma 5.1.4, ∼L = −L ⊂ K. Hence K = −L = ∼L. Interchanging the roles of K and L, we obtain L ∪ −L dense and L = −K = ∼K. We now have ∂K = K ∩ ∼K = K ∩ L = ∼L ∩ L = ∂L. Since K, L, and therefore their closures are all cones, so is ∂K. Now suppose also that L = {−x : x ∈ K}, and consider x ∈ ∂K = K ∩ L. There exist points y ∈ K and z ∈ L arbitrarily close to x. Then −y ∈ L, −z ∈ K, and these two points are arbitrarily close to −x; whence −x belongs to ∂K. It follows that ∂K is a subspace of X. 2

Lemma 5.2.5. Let K be a located subset of a normed space X, and let r > 0. Then the set Kr = {x ∈ X : ρ(x, K) r} is located, and for each x0 ∈ X, ρ(x0 , Kr ) = max {0, ρ(x0 , K) − r} .

(5.3)

Proof. If x ∈ Kr and ε > 0, then there exists y ∈ K such that x − y < r + ε; so x0 − x x0 − y − x − y ρ(x0 , K) − r − ε. Since ε > 0 is arbitrary, we have x0 − x ρ(x0 , K) − r

(x ∈ Kr ) .

(5.4)

116

5 Linearity and Convexity

On the other hand, choosing z ∈ K such that x0 − z < ρ (x0 , K) + ε, we have either x0 − z < r, in which case (5.3) holds with each side equal to 0, or else x0 − z > 0. In the latter case, writing r (x0 − z) , ζ = min 1, x0 − z we have z + ζ ∈ Kr and

1 − min 1, x0 − (z + ζ) =

r x0 − z

(x0 − z)

= max {0, x0 − z − r} < max {0, ρ(x0 , K) − r} + ε. Since ε > 0 is arbitrary, it follows from this and (5.4) that (5.3) holds.

2

By a half-space of a normed space X we mean a convex subset K such that ∂K is a hyperplane and the set {x ∈ X : x ∈ K ∨ −x ∈ K} is dense in X. We are now ready for the basic separation theorem. Our proof illustrates an important observation about classical proofs using Zorn’s lemma (see [65]): for separable spaces it is often possible to replace such a proof by a constructive one that uses an induction argument. Theorem 5.2.6. Let X be a separable normed space, K0 a bounded, located, open, convex subset of X such that ρ(0, K0 ) > 0, and x0 a point of X such that −x0 ∈ K0 . Then there exists an open half-space K of X such that K0 ⊂ K, ρ(x0 , K) > 0, and the boundary of K is a located subspace of X that is a hyperplane with associated vector x0 . Proof. Let (xn )n1 be a dense sequence in X. The basic idea of the proof is this: we carry out a succession of located convex enlargements of K0 such that for n 1, the cone generated by the nth enlargement Kn is close to at least one of the points xn and −xn , and such that the union of the cones c(Kn ) is the desired open half-space. The idea may seem simple, but the details are rather complicated. To be more precise, we construct bounded, located, open, convex subsets K0 ⊂ K1 ⊂ K2 ⊂ · · · of X such that the following properties obtain for n 1: (a) ρ(x0 , c(Kn )) > (1 − 2−n ) ρ (x0 , c(Kn−1 )) , (b) ρ (0, Kn ) > 0,

5.2 Separation Theorems

(c) max {ρ (xn , c(Kn )) , ρ (−xn , c(Kn ))}

0 and ρ(−xn , c (Kn−1 )) > 1/2n, we see from Lemma 5.2.2 that ρ(0, Kn+ ) > 0; similarly, ρ (0, Kn− ) > 0. Next, given z + ∈ c(Kn+ ) and z − ∈ c(Kn− ), ﬁnd t+ , t− > 0 and x+ , x− ∈ c(Kn−1 ) such that z + = t+ xn + x+ and z − = −t− xn + x− . We have t− x0 − z + + t+ x0 − z −

= t− x0 − t+ xn + x+ + t+ x0 − −t− xn + x−

t+ + t− x0 − t− x+ + t+ x−

+ t− t+ − + − = t + t x0 − + x + + x − − t +t t +t

t+ + t− ρ (x0 , c(Kn−1 ))

> t+ + t− 1 − 2−n−1 ρ (x0 , c(Kn−1 )) . Hence either

t− x0 − z + − 1 − 2−n−1 ρ (x0 , c(Kn−1 )) > 0 or

t+ x0 − z − − 1 − 2−n−1 ρ (x0 , c(Kn−1 )) > 0,

from which we obtain max x0 − z + , x0 − z − > 1 − 2−n−1 ρ (x0 , c(Kn−1 )) .

118

5 Linearity and Convexity

Taking the inﬁmum as z + , z − run over c(Kn+ ), c(Kn− ) respectively, we obtain max ρ(x0 , c(Kn+ )), ρ(x0 , c(Kn− )) 1 − 2−n−1 ρ (x0 , c(Kn−1 ))

> 1 − 2−n ρ (x0 , c(Kn−1 )) . Thus either

ρ(x0 , c(Kn+ )) > 1 − 2−n ρ (x0 , c(Kn−1 )) ,

in which case, noting that xn ∈ Kn+ and therefore ρ (xn , c(Kn+ )) = 0 < 1/n, we take Kn = Kn+ ; or else

ρ(x0 , c(Kn− )) > 1 − 2−n ρ (x0 , c(Kn−1 )) , when we take Kn = Kn− . This completes the inductive construction of Kn . The set K=

∞

c(Kn )

n=0

is easily shown to be an open convex cone. Moreover, we see from (a) that (ρ(x0 , c(Kn )))n1 is a Cauchy sequence, which therefore converges in R. Since we have c(Kn−1 ) ⊂ c(Kn ) for each n, the limit of the sequence is ρ(x0 , K). Now, for n 1, n n

1 1 −k −k 1− 1−2 > , 2 2 4 k=1

k=2

so, by (a), ρ(x0 , K) = lim ρ (x0 , c (Kn )) n→∞

1 ρ (x0 , c (K0 )) . 4

But by the convexity of K0 , ρ (x0 , c(K0 )) = inf {tx − x0 : t > 0, x ∈ K0 } t 1 x + (−x ) : t > 0, x ∈ K = inf (t + 1) 0 0 t + 1 t+1 ρ(0, K0 ). Hence ρ(x0 , K) > 0. Let L = {x ∈ X : −x ∈ K} . Then L is an open convex cone containing x0 . By (c), for each n, either ρ (xn , K) < 1/n or ρ (xn , L) < 1/n. So K ∪ L is dense in X. On the other hand, K ⊂ ∼L: for if x ∈ K and y ∈ L, then there exists n such that x ∈ c (Kn ) and −y ∈ c (Kn ); so there exist t > 0 and z ∈ Kn such that x − y = tz and therefore x − y tρ (0, Kn ) > 0. It now follows from Lemma 5.2.4 that L ∪ −L is dense in X, and the common boundary of K and L is the set

5.2 Separation Theorems

119

N = K ∩ L, which is a linear subspace of X. To prove that N is a hyperplane, let γ denote the boundary crossing map of L relative to its interior point x0 . Note that, by Lemma 5.2.4, K ⊂ −L; so, by Proposition 5.1.5, for each x ∈ K there exists a unique t ∈ (0, 1) such that γ (x) = tx0 + (1 − t) x and therefore

t 1 (−x0 ) + γ (x) . 1−t 1−t Similar considerations using the boundary crossing map of K relative to −x0 complete a proof that each x ∈ K ∪ L can be expressed in the form αx0 + βz with α ∈ R and z ∈ N. In fact, such an expression obtains for all x ∈ X. For since K ∪ L is dense, there exist a sequence (αn )n1 in R and a sequence (zn )n1 in N such that αn x0 + zn −→ x as n −→ ∞. Given ε > 0 and m > n, we have either |αm − αn | < ε or |αm − αn | > 0; in the latter case, since zm − zn ∈ N ⊂ K, 1 (zm − zn ) (αm x0 + zm ) − (αn x0 + zn ) = |αm − αn | x0 − |αm − αn | |αm − αn | ρ(x0 , K), x=

so |αm − αn |

1 (αm x0 + zm ) − (αn x0 + zn ) . ρ(x0 , K)

It follows that (αn )n1 is a Cauchy sequence and therefore converges to a limit α in R. Writing z = x − αx0 , we have z − zn x − (αn x0 + zn ) + |α − αn | x0 −→ 0 as n −→ ∞, so z ∈ N = N. Thus N is a hyperplane with associated vector x0 , and therefore (since also K ∪ L is dense in X) K is a half-space. Now, for each x ∈ K we have γ(x) ∈ N ⊂ K and x0 − γ (x) x0 − x . It follows that ρ (x0 , N ) exists, equals ρ(x0 , K), and is positive. For each x ∈ X, choosing α ∈ R and z ∈ N such that x = αx0 + z, we easily see that ρ (x, N ) exists 2 and equals |α| ρ(x0 , K). Hence N is located. Corollary 5.2.7. Under the hypotheses of Theorem 5.2.6, there exists a normed linear functional u on X such that u = 1 and Re u is positive on K0 . Proof. First take the case K = R. Let K be as in the conclusion of Theorem 5.2.6, let

120

5 Linearity and Convexity

L = {−x : x ∈ K} , recall that K ⊂ −L, and let N be the common boundary of K and L. Since N is a located hyperplane, we see from Propositions 2.3.4 and 2.3.6 that there exists a normed linear functional v on X such that ker(v) = N and v(x0 ) = 1. Deﬁning −1 u = − v v, we obtain a normed linear functional on X with norm 1 such that ker(u) = N and u(x0 ) < 0. Let γ be the boundary crossing map of L relative to x0 . Given x ∈ K, choose λ ∈ (0, 1) such that γ(x) = λx + (1 − λ) x0 . Then u(γ(x)) = 0, so 1−λ u(x0 ) > 0. u(x) = − λ Hence u is positive on K and therefore on K0 . Now consider the case K = C. By the case just considered, there exists a real linear functional v : X −→ R on the real linear space X such that v has norm 1 and is positive on K0 . The following lemma then shows that u(x) = v(x) − iv(ix) deﬁnes a linear functional u : X −→ C on the complex linear space X such that u = 1 and Re u is positive on K0 . 2 Lemma 5.2.8. If X is a normed linear space over C, then there is a one-one correspondence between real linear functionals v on X and complex linear functionals u on X, given by (5.5) u(x) = v(x) − iv(ix). Moreover, if either u or v is normed, then both are and their norms are equal. Proof. If u is a complex linear functional, then u = v + iw for unique real linear functionals v and w. But iv(x) − w(x) = iu(x) = u(ix) = v(ix) + iw(ix), so w(x) = −v(ix). Conversely, if v is a real linear functional, then (5.5) deﬁnes a linear functional u that respects multiplication by real numbers; it is easily checked that u(ix) = iu(x). The ﬁnal conclusion, about the norms of u and v, is a straightforward consequence of the identity |u(x)| = sup{|v(λx)| : |λ| = 1}, (5.6) which we now establish. If |λ| = 1, then |u(x)| = |λu(x)| = |u(λx)| |Re u(λx)| = |v(λx)| . On the other hand, if ε > 0, then either |u(x)| − ε < |v(x)|, or else u(x) = 0. In the latter case, if we let λ = |u(x)|/u(x), then

5.2 Separation Theorems

121

u(λx) = λu(x) = |u(x)| ∈ R, so |u(x)| = Re u(λx) = v(λx). Notice that in this case the supremum in (5.6) is achieved. 2

The following result is the full form of the separation theorem. Theorem 5.2.9. Let A and B be bounded convex subsets of a separable normed space X such that the algebraic diﬀerence {y − x : x ∈ A, y ∈ B} is located and the mutual distance d = inf {y − x : x ∈ A, y ∈ B} is positive. Then for each ε > 0 there exists a normed linear functional u on X, with norm 1, such that Re u(y) > Re u(x) + d − ε

(x ∈ A, y ∈ B) .

Proof. We may assume that ε < d. Write ε K0 = y − x − z : x ∈ A, y ∈ B, z < d − , 2 which is bounded, open, and convex. It follows from Lemma 5.2.5 that K0 is located and ρ(0, K0 ) = ε/2. Hence, by Corollary 5.2.7, there exists a normed linear functional u on X such that u = 1 and Re u is positive on K0 . Choose z ∈ X such that z < d − ε/2 and u(z) > d − ε; this is possible since u = 1. For all x ∈ A and y ∈ B we have y − x − z ∈ K0 and so Re u(y − x − z) > 0; whence Re u(y) > Re (u(x) + u(z)) > Re u(x) + d − ε, as required.

2

An inhabited subset K of a normed space X is said to be balanced if αx ∈ K whenever x ∈ K and |α| 1; in that case, 0 ∈ K. Corollary 5.2.10. Let x0 be a vector in a separable normed space X, and K a located subset of X that is bounded, convex, and balanced, such that ρ(x0 , K) > 0. Then for each ε > 0 there exists a normed linear functional u on X with norm 1 such that u(x0 ) > |u(x)| + ρ(x0 , K) − ε for all x ∈ K.

122

5 Linearity and Convexity

Proof. We may assume that 0 < ε < ρ(x0 , K). Applying Theorem 5.2.9 with A = K and B = {x0 }, construct a normed linear functional v on X with norm 1 such that Re v(x0 ) > Re v(x) + ρ(x0 , K) − ε (x ∈ K) . Since 0 ∈ K, we have v(x0 ) = 0. Thus u=

|v(x0 )| v v(x0 )

is a normed linear functional on X with norm 1 such that u(x0 ) = |v(x0 )| Re v(x0 ) > Re v(x) + ρ(x0 , K) − ε (x ∈ K) . For each x ∈ K, either u(x0 ) > |u(x)| + ρ(x0 , K) − ε, or else u(x) = 0 and therefore v(x) = 0. In the latter case, since K is balanced, |v(x)| x ∈ K, v(x) and therefore

u(x0 ) > Re v

|v(x)| x + ρ(x0 , K) − ε v(x)

= |v(x)| + ρ(x0 , K) − ε = |u(x)| + ρ(x0 , K) − ε, as required.

2

5.3 The Hahn–Banach Theorem The Hahn–Banach theorem enables us to extend a normed linear functional, with an arbitrarily small increase in norm, from a subspace of a normed space to the entire space. This fundamental result has numerous applications throughout functional analysis. In the constructive context we deal only with the extension of linear functionals on subspaces of a separable normed space. The standard classical proofs extending the theorem to nonseparable normed spaces depend on Zorn’s lemma and are therefore nonconstructive. We begin with two more consequences of the separation theorem. Proposition 5.3.1. Let x be an element of a nontrivial separable normed space X, and let ε > 0. Then there exists a normed linear functional u on X such that u = 1 and u(x) > x − ε.

5.3 The Hahn–Banach Theorem

123

Proof. If x = 0, then we may apply the separation theorem (Theorem 5.2.9) with A = {0} and B = {x} . In the general case, choose a nonzero vector y such that x − y < ε/2, and then construct a normed linear functional u on X such that u = 1 and u(y) > y − ε/2. We have u(x) u(y) − |u(x) − u(y)| > y −

ε − x − y > x − ε, 2 2

as required.

Corollary 5.3.2. Let V be a located subspace of a separable normed space X such that the metric complement −V is inhabited. Then there exists a normed linear functional u on X such that u = 1 and u(V ) = {0} . Proof. Let e be a unit vector in −V. Applying Proposition 5.3.1 in the quotient space X/V, construct a normed linear functional u on X/V such that u = 1 and u(v) = 0. Clearly, V ⊂ ker(u). Since X/V , we see that u, regarded as a linear functional on the original normed space X, has bound 1. On the other hand, given ε > 0, we can ﬁnd x ∈ X with xX/V < 1 and u(x) > 1 − ε; choosing y ∈ V with x − y < 1, we then have u(x − y) = u(x) > 1 − ε. Since ε > 0 is arbitrary, it follows that, as a linear functional on X, u is normed and has norm 1. 2

Proposition 5.3.1 is crucial for our proof of the Hahn–Banach theorem. Theorem 5.3.3. Let v be a nonzero bounded linear functional on a linear subset Y of a separable normed linear space X such that ker(v) is located in X. Then for each ε > 0 there exists a normed linear functional u on X such that u v + ε and u(y) = v(y) for each y ∈ Y. Proof. First note that, by Proposition 2.3.6, v is normed as a linear functional on Y. Fix y0 in Y with v(y0 ) = 1. Then for each x ∈ ker(v), y0 − x

1 1 v(y0 − x) = , v v

from which it follows that y0 X/ ker(v) = ρ (y0 , ker(v))

1 . v

For each normed linear functional u on X/ ker(v) denote the norm by uX/ ker(v) . Using Proposition 5.3.1, construct a normed linear functional u0 on X/ ker(v) such that u0 X/ ker(v) = 1 and u0 (y0 ) > 1/(v + ε). Then u=

1 u0 u0 (y0 )

124

5 Linearity and Convexity

is a normed linear functional on X/ ker(v) such that uX/ ker(v) < v + ε and u(y0 ) = 1. Since X/ ker(v) on X, we see that u, regarded as a linear functional on the original normed space X, has bound uX/ ker(v) . To see that this is actually the norm of u on X, consider any δ > 0 and choose x ∈ X such that xX/ ker(v) < 1 and u(x) > uX/ ker(v) − δ. Finding y ∈ ker(v) such that x − y < 1, we have u(x − y) = u(x) > uX/ ker(v) − δ. Since δ > 0 is arbitrary, we conclude that u is normed on X, with u = uX/ ker(v) v + ε. Finally, for each y ∈ Y we have y − v(y)y0 ∈ ker(v), so 0 = u(y − v(y)y0 ) = u(y) − v(y), and therefore u(y) = v(y).

2

The Hahn–Banach theorem and some of its associates have surprising applications. As surprising as any is the following, whose classical proof is almost trivial. Proposition 5.3.4. Let x1 , . . . , xn be elements of an inﬁnite-dimensional normed space X, and let ε > 0. Then there exist linearly independent elements e1 , . . . , en of X such that xi − ei < ε for each i. Proof. Construct a ﬁnite-dimensional subspace V of span {x1 , . . . , xn } such that for each i there exists yi ∈ V with xi − yi < ε/2 (see Lemma 4.1.11). Since X is inﬁnite-dimensional and V is ﬁnite-dimensional, we can embed V in an ndimensional subspace W of X. We may assume that y1 = 0. Setting e1 = y1 , suppose that for some k < n, we have constructed linearly independent elements e1 , . . . , ek of W such that yi − ei < ε/2 for 1 i k. Let Vk be the k-dimensional subspace of W with basis {e1 , . . . , ek } . By Corollary 5.3.2, there exists a normed linear functional u on W such that u (Vk ) = {0} and u = 1. Construct a vector z ∈ W such that z = ε/2 and u (z) > ε/3. Either u (yk+1 ) = 0 or else |u (yk+1 )| < ε/3. In the ﬁrst case, ρ (yk+1 , Vk ) > 0 and we set ek+1 = yk+1 . In the second case, u(yk+1 − z) = 0, ρ (yk+1 − z, Vk ) > 0, and we set ek+1 = yk+1 − z. In each case, the vectors e1 , . . . , ek+1 are linearly independent, by Lemma 4.1.10, and 2 xk+1 − ek+1 < ε. This completes the inductive construction.

The classical Hahn–Banach theorem says that we can extend a bounded linear functional v from Y to a functional u on the whole space X with exact preservation of norm: that is, u = v . In general, as Exercise 5 shows, we cannot do this

5.3 The Hahn–Banach Theorem

125

constructively. But, as we shall see, if we impose extra conditions on the norm of X, then we can make the extension norm-preserving. A mapping u of a linear space X into R is said to be convex if u (λx + (1 − λ) y) λu(x) + (1 − λ) u(y) whenever x, y ∈ X and 0 λ 1. Lemma 5.3.5. Let u be a convex mapping of a linear space X into R such that u(x) = −u(−x) for each x ∈ X. Then u is linear. Proof. First note that u(0) = −u(−0) = −u(0), so u(0) = 0. For each x ∈ X and for 0 λ 1 we have u(λx) = u(λx + (1 − λ) 0) λu(x) + (1 − λ) u(0) = λu(x). Replacing x by −x, we obtain u(−λx) λu(−x); whence, by our hypotheses on u, −u(λx) −λu(x) and therefore λu(x) u(λx). Hence u(λx) = λu(x) whenever 0 λ 1. Now consider any real λ = 0. If λ > 0, then by the foregoing, 1 λ λ u(x) = u x = u(λx) λ+1 λ+1 λ+1 and therefore λu(x) = u(λx). If λ < 0, then u(λx) = u((−λ)(−x)) = −λu(−x) = λ(−u(−x)) = λu(x). Now consider any λ ∈ R, and assume that u(λx) = λu(x). The foregoing shows that ¬ (λ > 0 ∨ λ < 0) and therefore λ = 0. But this is absurd, since u(0x) = 0 = 0u(x). We conclude that u(λx) = λu(x) for all λ ∈ R. It remains to prove the additivity of u. By the convexity of u, 1 1 1 (x + y) u(x) + u(y). u 2 2 2 Applying the ﬁrst part of the proof yields u(x + y) u(x) + u(y). We now replace x, y by −x, −y respectively, to obtain u(−(x + y)) = u(−x − y) u(−x) + u(−y) and therefore, by the hypotheses on u, u(x + y) u(x) + u(y). Hence u(x + y) = u(x) + u(y). 2

126

5 Linearity and Convexity

The norm of a normed space X is said to be Gˆ ateaux diﬀerentiable at x ∈ X if for each y ∈ X the limit ux (y) = lim

t→0

x + ty − x t

exists. For each c > 0 we then have x + c−1 ty − x = ux (y). ucx (y) = lim t→0 c−1 t If the norm of X is Gˆateaux diﬀerentiable at each nonzero point of X, we say that X has Gˆ ateaux diﬀerentiable norm. In such a case we obtain a strengthening of Proposition 5.3.1. Proposition 5.3.6. Let x be a nonzero vector in a normed space X, and suppose that the norm of X is Gˆ ateaux diﬀerentiable at x. Then the mapping ux is the unique normed linear functional u on X with u = 1 and u(x) = x . Proof. It is enough to take the case K = R. It is straightforward to show that ux is convex, and that ux (y) = −ux (−y) for each y ∈ X; whence, by Lemma 5.3.5, ux is linear. Moreover, |ux (y)| y and ux (x) = x. So ux = 1, and it only remains to prove the uniqueness. To this end, let u be any normed linear functional on X such that u = 1 and u(x) = x . For each y ∈ X and each t > 0 we have u(x − ty) − u(x) u(−ty) x − ty − x = = u(y) −t −t −t and

x + ty − x u(x + ty) − u(x) . t t Keeping y ﬁxed and letting t −→ 0 in these two sets of displayed inequalities, we 2 obtain ux (y) u(y) ux (y); whence u(y) = ux (y). u(y) =

Proposition 5.3.7. Let Y be a subspace of a uniformly convex Banach space X with Gˆ ateaux diﬀerentiable norm, and let v be a nonzero normed linear functional on Y. Then there exists a unique normed linear functional u on X such that u = v and u(y) = v(y) for each y ∈ Y. Proof. We may assume that v = 1. Using Exercise 19 of Chapter 2 if necessary, we may further assume that Y is closed in X and is therefore a Banach space. Since Y is uniformly convex, it follows from Proposition 2.3.7 that there exists a unique a ∈ Y such that a = 1 = v(a). By Proposition 5.3.6, ua is a normed linear functional on X such that ua = 1 = ua (a). Applying the uniqueness part of Proposition 5.3.6 in the space Y, we conclude that ua (y) = v(y) for all y ∈ Y.

5.4 Locally Convex Spaces

127

Finally, if u is a normed linear functional on X such that u = v and u(y) = v(y) for all y ∈ Y, then u(a) = v(a) = 1 = v = u , so, again by the uniqueness part of Proposition 5.3.6, u = ua .

2

Not surprisingly, there is also a strong version of the separation theorem that applies in a uniformly convex separable Banach space; see Exercises 9 and 10.

5.4 Locally Convex Spaces Although Errett Bishop considered that in most cases of interest it seems to be unnecessary to make use of any deep facts from the general theory of locally convex spaces ([9], page 350), the development of constructive analysis (in particular, aspects of the theory of operators) in recent years has greatly beneﬁted from such a general theory, which we now present. A locally convex space consists of a linear space X over K, a family (pi )i∈I of seminorms on X, and the equality and (compatible, tight) inequality relations deﬁned by x = y ⇐⇒ ∀i ∈ I (pi (x − y) = 0) , x = y ⇐⇒ ∃i ∈ I (pi (x − y) > 0) . Following normal practice, we call X itself a locally convex space when it is clear which family of seminorms is under consideration. The family (pi )i∈I and the associated equality and inequality together form the locally convex structure on X. The corresponding locally convex topology on X is the family τ X of all subsets of X that are unions of sets of the form ! pi (x − a) < ε U (a, F, ε) = x ∈ X : i∈F

where a ∈ X, F is an inhabited ﬁnitely enumerable subset of I, and ε > 0. The seminorms pi (i ∈ I) are called the deﬁning seminorms of τ X . The members of τ X are called open subsets of X, and the sets U (a, F, ε) basic neighbourhoods of a. On the other hand, if S is a subset of X, then its closure S is the set of all elements x of X such that S ∩ U (x, F, ε) is inhabited for all ﬁnitely enumerable F ⊂ I and all ε > 0. We say that S is closed (in the locally convex topology τ X ) if S = S, and that S is dense in X if S = X. The unit ball of the locally convex space X is {x ∈ X : ∀i ∈ I (pi (x) 1)} ,

128

5 Linearity and Convexity

which is a closed subset of X. We regard a normed space (X, ) as a locally convex space in which the family of seminorms consists of the single norm . In that case, the locally convex topology is just the metric topology associated with the norm on X, and the various notions of convergence, total boundedness, locatedness, and so on arising from the locally convex structure on X are precisely those associated with the norm on X.

In the following, unless we state otherwise, X, (pi )i∈I and Y, (qj )j∈J are locally convex spaces over K. A mapping f of a subset S of X into Y is continuous at a ∈ S if for each ε > 0 and each ﬁnitely enumerable G ⊂ J, there exist δ > 0 and aﬁnitely enumerable F ⊂ I such that if x ∈ S and pi (x − a) < δ, then qj (f (x) − f (a)) < ε; i∈F

j∈G

continuous on S if it is continuous at each point of X; uniformly continuous on S if for each ε > 0 and each ﬁnitely enumerable subset G of J, there exist δ > 0 and a ﬁnitely enumerable subset F of I such that if pi (x − y) < δ, then qj (f (x) − f (y)) < ε. x, y ∈ S and i∈F

j∈G

Notice that each of the deﬁning seminorms pi on X is uniformly continuous on X. Proposition 5.4.1. The following are equivalent conditions on a linear mapping u between the locally convex spaces X and Y : (a) u is continuous at 0. (b) u is continuous on X. (c) u is uniformly continuous on X. (d) There exist a positive real number C and a ﬁnitely enumerable subset F of I such that |u(x)| C sup pi (x) i∈F

for each x ∈ X. Proof. It is routine to show that (d) =⇒ (c) =⇒ (b) =⇒ (a). Suppose that u is continuous at 0. There exist δ > 0 and a ﬁnitely enumerable subset F of I such pi (x) < δ, then |u(x)| < 1. It follows that for each x ∈ X and each ε > 0, that if i∈F

⎛ ⎞ δx u ⎝ ⎠ < 1 pi (x) + ε i∈F

5.4 Locally Convex Spaces

and therefore 1 |u(x)| < δ

129

pi (x) + ε .

i∈F

Since ε > 0 is arbitrary, it follows that |u(x)|

1 pi (x). δ i∈F

2

Thus (a) =⇒ (d).

Naturally, in view of the analogous situation in a normed space, we say that a mapping u satisfying one, and hence all, of conditions (a)–(d) of Proposition 5.4.1 is a bounded linear mapping.

We say that a sequence (xn )n1 in a locally convex space X, (pi )i∈I converges to a limit x in X if for each basic neighbourhood U = U (x, F, ε) of x in X there exists n0 such that xn ∈ U whenever n n0 ; is a Cauchy sequence if for each ε > 0 and each ﬁnitely enumerable subset F of pi (xm − xn ) < ε whenever m, n n0 . I, there exists n0 such that i∈F

We say that X is complete if every Cauchy sequence in X converges to a limit in X. (Strictly speaking, this deﬁnition applies only when X is separable—that is, has a countable dense subset. However, it will suﬃce for our purposes.) Let S be a subset of X, F a ﬁnitely enumerable subset of I, and ε > 0. By an ε-approximation to S relative to F we mean a subset T of S such that for each pi (x − y) < ε. We say that S is x ∈ X there exists y ∈ T with i∈F

totally bounded relative to F if for each ε > 0 there exists a ﬁnitely enumerable ε-approximation to S relative to F ; totally bounded if it is totally bounded relative to each ﬁnitely enumerable subset of I. The proofs of the next ﬁve results are similar to those of their counterparts in metric space theory (see Chapter 2, Section 2) and are left as exercises. Proposition 5.4.2. If S is a totally bounded subset of a locally convex space, and f is a uniformly continuous mapping of S into a locally convex space Y, then f (S) is totally bounded.

130

5 Linearity and Convexity

Proposition 5.4.3. If S is a totally bounded subset of a locally convex space, and f is a uniformly continuous mapping of S into R, then sup f (x) and inf f (x) exist. x∈S

x∈S

A subset S of the locally convex space X, (pi )i∈I is said to be located (in X) if ! inf pi (x − y) : y ∈ S i∈F

exists for each x ∈ X and each ﬁnitely enumerable subset F of I. As it does in metric spaces, locatedness plays an important role in locally convex spaces. Proposition 5.4.4. A totally bounded subset of a locally convex space is located. Proposition 5.4.5. A located subset of a totally bounded set in a locally convex space is totally bounded. Theorem 5.4.6. Let S be a totally bounded subset of a locally convex space X, and let f be a uniformly continuous mapping of S into R. Then for all but countably many t > inf f (x), the set x∈S

{x ∈ S : f (x) t}

is totally bounded. Now consider the set B(X, Y ) of all bounded linear mappings between the locally convex spaces X and Y. This set, which we often denote by B when it is clear which spaces X and Y are under consideration, becomes a locally convex space when taken with pointwise addition and multiplication-by-scalars and endowed with the seminorms px deﬁned by px (T ) = T x (x ∈ X, T ∈ B(X, Y )) . We denote the unit ball of B(X, Y ) by B1 (X, Y ) or just B1 . When X = Y, we usually write B(X) and B1 (X) rather than B(X, Y ) and B1 (X, Y ). In the special case where Y is the ground ﬁeld K, we obtain the space of all bounded linear functionals on X; this space is called the ∗ dual space of X, and is denoted by X ∗ or sometimes, for clarity, by X, (pi )i∈I ; its unit ball is denoted by X1∗ . The topology associated with the family of seminorms (px )x∈X on X ∗ is called the weak ∗ topology on X ∗ . When we are dealing with, for example, total boundedness relative to the locally convex structure on X ∗ , we speak of weak ∗ total boundedness. In classical analysis, when X and Y are normed spaces, a big role is played by the operator norm T = sup {T x : x ∈ X, x 1} of an element T of B(X, Y ). Constructively the operator norm exists only for those elements T that are normed; but Corollary 4.1.8 shows that it exists for all T ∈

5.4 Locally Convex Spaces

131

B(X, Y ) when X is ﬁnite-dimensional. In that case, X ∗ is algebraically isomorphic to X, every element of X ∗ is normed, the weak∗ topology on X ∗ coincides with the topology associated with the operator norm, and X ∗ is a Banach space relative to the operator norm; these facts are easily veriﬁed using results from Section 1 of Chapter 4. We now examine the unit ball of the dual space. Theorem 5.4.7. (Banach–Alaoglu theorem) If X is a separable normed space, then X1∗ is complete and totally bounded relative to the locally convex structure on X ∗. Proof. To prove the weak∗ -completeness, let (vn )n1 be a Cauchy sequence in X1∗ . Given ε > 0 and a ﬁnitely enumerable subset F of X, we can ﬁnd N such that |(vm − vn )(x)| < ε (m, n N ) . (5.7) x∈F

It readily follows that for each x ∈ X, (vn (x))n1 is a Cauchy sequence in K. Since K is complete, the limit v(x) of the sequence (vn (x))n1 exists in K for each x ∈ X. It is straightforward to show that the mapping v : X −→ K so deﬁned is linear and belongs to X1∗ . Finally, keeping n N ﬁxed and taking the limit with respect to m in (5.7), we obtain |(v − vn ) (x)| ε (n N ) . x∈F

Since ε, F are arbitrary, the sequence (vn )n1 converges to v in the weak∗ topology. To prove the weak∗ -total boundedness of X1∗ , let F = {x1 , . . . , xm } be a ﬁnitely enumerable subset of X, let M > 4 + max {xi : 1 i m} , and let 0 < ε < 1. By Lemma 4.1.11, there exists a ﬁnite-dimensional subspace X0 of X such that for 1 i m, ρ (xi , X0 ) < ε/m and therefore there exists yi ∈ X0 with xi − yi < ε/m. If X0 = {0}, then for each v ∈ X1∗ , m i=1

|v(xi ) − v(0)|

m

xi < ε,

i=1

so {0} is an ε-approximation to X1∗ relative to F. We may therefore assume that X0 has positive dimension. Then every element of X0∗ (the dual of X0 ) is normed, and X0∗ , taken with the operator norm, is a ﬁnite-dimensional Banach space. Hence its unit ball is compact relative to the operator norm (Corollary 4.1.7). Moreover, by Proposition 2.3.6, each nonzero element of X0∗ has its kernel located in X0 ; since X0 is locally compact, this kernel is locally compact and hence is located in the space X (Proposition 2.2.18). It follows that the Hahn–Banach theorem can be applied to extend nonzero bounded linear functionals from X0 to X.

132

5 Linearity and Convexity

∗ Let u01 , . . . , u0n be an ε/m-approximation to the unit ball of X0 in the operator 0 norm such that 0 < uk < 1 for each k. By the Hahn–Banach theorem, there exist normed linear functionals u1 , . . . , un in X1∗ such that uk (x) = u0k (x) for each x ∈ X0 and each k. Given u ∈ X1∗ , since the restriction of u to X0 is in the unit ball of X0∗ , we can ﬁnd k such that u(x) − u0k (x) < ε/m for all x ∈ X0 with x 1. For each i (1 i m) we have |u(xi ) − uk (xi )| |(u − uk ) (xi − yi )| + u − u0k (yi ) 1 0 yi 2 xi − yi + (1 + yi ) u − uk 1 + yi ε ε 2ε + 1 + xi + < m m m Mε . < m

Hence

m

|u(xi ) − uk (xi )| < M ε.

(5.8)

i=1

We now see that {u1 , . . . , un } is an M ε-approximation to X1∗ relative to F.

2

Corollary 5.4.8. If X is a separable normed space, then the normed linear functionals on X are weak∗ -dense in X ∗ . Proof. In the notation of the last paragraph of the foregoing proof, the function uk satisfying (5.8) is normed. Since ε > 0 and the ﬁnitely enumerable subset F of X are arbitrary, the result follows. 2 Let X be a normed space. It is easy to see that, for a ﬁxed vector x ∈ X, the linear functional u u(x) on X ∗ is weak∗ -uniformly continuous on X1∗ . We shall show that any element of X ∗∗ (the dual of X ∗ ) that is uniformly continuous on X1∗ can be approximated arbitrarily closely by functionals of this special form, and that if X is complete, then this approximation can be made exact. Lemma 5.4.9. Let V be a locally convex space whose unit ball V1 is totally bounded, and let φ be a nonzero linear functional on V that is uniformly continuous on V1 . Then the unit kernel, V1 ∩ ker(φ), of φ is totally bounded. Proof. Since φ is nonzero and uniformly continuous on the totally bounded set V1 , C = sup{|φ(y)| : y ∈ V1 } exists and is positive. Using the linearity of φ, choose y ∈ V1 with φ(y) > C/2. Then

5.4 Locally Convex Spaces

y0 =

133

C y 2φ(y)

belongs to V1 , and φ(y0 ) = C/2. Let ε be a positive number, and F = {p1 , . . . , pm } a ﬁnitely enumerable set of deﬁning seminorms on V . Using Theorem 5.4.6, compute a positive number Cε t< C + 4m such that St = {y ∈ V1 : |φ(y)| t} is totally bounded. Pick a t-approximation {s1 , . . . , sn } of St relative to F . Set C 2 sk − φ(sk )y0 (1 k n). yk = C + 2t C Then yk ∈ ker(φ). If pi is any deﬁning seminorm on V , then we have C 2 pi (sk ) + |φ(sk )|pi (y0 ) pi (yk ) C + 2t C 2t C 1+ C + 2t C = 1. So yk belongs to V1 ∩ ker(φ). Now consider any element y of V1 ∩ ker(φ). Since y ∈ St , there exists k such m that pi (y − sk ) < t and therefore i=1

m i=1

pi (y − yk )

m

pi (y − sk ) +

i=1

m

m

pi (sk − yk )

i=1

2 pi (tsk + φ(sk )y0 ) 0 there exist δ > 0 and a ﬁnite-dimensional subspace X0 of X such that if u ∈ X1∗ and |u(x)| < δ for all x in the unit ball of X0 , then |φ(u)| < ε. Proof. There exist ξ 1 , . . . , ξ n in X with ξ k < 1/2 for each k, and a positive number δ, such that if u ∈ X1∗ and |u(ξ k )| < 2δ for each k, then |φ(u)| < ε. By Lemma 4.1.11, there exist a ﬁnite-dimensional subspace X0 of X, and vectors x1 , . . . , xn of X0 , such that 1 ,δ (1 k n) . ξ k − xk < min 2 Suppose that u ∈ X1∗ and |u(x)| < δ for all x in the unit ball of X0 . Then for each k, |u(ξ k )| |u(ξ k − xk )| + |u(xk )| ξ k − xk + δ < 2δ. Hence |φ(u)| < ε.

2

Lemma 5.4.11. Let X be a separable normed space, X0 a ﬁnite-dimensional subspace of X, and φ a nonzero linear functional on X that is weak∗ -uniformly continuous on X1∗ . Denote the unit kernel of φ by N1 . Then the following hold: (a) φ = sup {|φ(u)| : u ∈ X1∗ } exists. (b) x0 = sup {|u(x)| : u ∈ N1 } deﬁnes a seminorm on X0 . (c) β = inf {x0 : x ∈ X0 , x = 1} is well deﬁned. Suppose that β > 0. Then (d) (X0 , 0 ) is a ﬁnite-dimensional Banach space, and 0

u = sup {|(u(x)| : x ∈ X0 , x0 1} deﬁnes the corresponding operator norm on the dual space X0∗ . Finally, for each u ∈ N1 let F (u) denote the restriction of u to X0∗ . Then 0

∗ (e) F is continuous, and ran(F ) is -dense in the unit ball S0∗ weak -uniformly 0 of X0∗ , .

Proof. Since φ is weak∗ -uniformly continuous on X1∗ and, by Theorem 5.4.7, the latter is weak∗ -totally bounded, we see that φ exists. Also, N1 is weak∗ -totally bounded, by Lemma 5.4.9, and the mapping u u(x) is weak∗ -uniformly continuous on X ∗ , so x0 exists. Clearly, 0 is a seminorm and 0 ; whence 0 is uniformly continuous on (X0 , ) . Since that space is ﬁnite-dimensional,

5.4 Locally Convex Spaces

135

{x ∈ X0 : x = 1} is compact, and so β exists. Now suppose that β > 0. Then β −1 0 , so 0 is a norm on X0 equivalent to , and (X0 , 0 ) is a ﬁnite-dimensional Banach space. The mapping 0 F : N1 −→ (X0∗ , ) is weak∗ -uniformly continuous on the weak∗ -totally bounded 0 set N1 , so its range is totally bounded, and hence located, in (X0∗ , ). To show 0 ∗ ∗ ∗ that ran(F ) is dense in the unit ball S0 of (X0 , ), ﬁx u0 in S0 and suppose that 0 0 < c = inf u0 − F (u) : u ∈ N1 . 0

By Corollary 5.2.10, there exists a normed linear functional Φ on (X0∗ , ) such that c (u ∈ N1 ) . Φ(u0 ) > |Φ(F (u))| + 2 Now X0∗ , being ﬁnite-dimensional, is equivalent to a Hilbert space. (Recall from Corollary 4.1.9 that all norms on a ﬁnite-dimensional space are equivalent to each other and hence to the Euclidean norm.) It follows from this and the Riesz representation theorem that there exists ζ ∈ X0 such that Φ(v) = v(ζ) for each v ∈ X0∗ . Therefore u0 (ζ) > sup {|F (u)(ζ)| : u ∈ N1 } = sup {|u(ζ)| : u ∈ N1 } = ζ0 . 0

This is absurd, since u0 ∈ S0∗ . Hence, in fact, c = 0 and ran(F ) is -dense in S0∗ . 2 Lemma 5.4.12. With the hypotheses and notation of Lemma 5.4.11 and its proof, suppose that φ > 1 and let α, δ be positive numbers. Suppose also that |φ(u)| < α/2 whenever (a) u ∈ X1∗ and (b) |u(x)| < δ for all x in the unit ball of X0 . Then β < α. Proof. Choose u0 ∈ X1∗ such that φ(u0 ) = 1. Since |u0 (βx)| = β |u0 (x)| β x x0 , the linear functional Ψ : x u0 (βx) belongs to S0∗ . Now, either β < α or β > 0. Assuming the latter inequality, we see from Lemma 5.4.11 that there exists u ∈ N1 such that 0

Ψ − F (u) < 2δ. So for all x in the unit ball of (X0 , 0 ) , |u0 (βx) − u(x)| < 2δ

136

5 Linearity and Convexity

and therefore

1 (βu0 − u) (x) < δ. 2

But β 1, and both u0 and −u belong to the convex set X1∗ , so whence φ 1 (βu0 − u) < α . 2 2

Therefore β = βφ(u0 ) − φ(u) = 2φ

1 (βu0 − u) 2

1 2

(βu0 − u) ∈ X1∗ ;

< α, 2

as we required.

Proposition 5.4.13. Let X be a separable normed space, and φ a linear functional on X ∗ that is weak∗ -uniformly continuous on X1∗ . Then for each ε > 0 there exists x ∈ X such that x < 3 φ and |φ(u) − u(x)| < ε

(u ∈ X1∗ ) .

Proof. If φ < ε, then we can take x = 0; so we may assume that φ > 0. Scaling if necessary, we may also assume that there exists u0 ∈ X1∗ with φ(u0 ) = 1. Pick a positive number α, which we shall specify further as the proof develops. Using Lemma 5.4.10, construct a ﬁnite-dimensional subspace X0 of X, and a positive number δ, such that if u ∈ X1∗ and |u(x)| < δ for all x in the unit ball of X0 , then |φ(u)| < α/2. Applying Lemma 5.4.12, we see that there exists z ∈ X0 such that z = 1 and |u(z)| < α for all u ∈ N1 . For each u ∈ X1∗ , since 1 (u − φ(u)u0 ) ∈ N1 , 1 + φ we therefore have

1 |(u − φ(u)u0 ) (z)| < α. 1 + φ

By Proposition 5.3.1, there exists u1 ∈ X1∗ such that u1 (z) > 1/2. Thus 1 < |(u1 − φ(u1 )u0 ) (z)| + |φ(u1 )| |u0 (z)| (1 + φ) α + φ |u0 (z)| , 2 and therefore

1 |u0 (z)| > φ

1 − (1 + φ) α 2

>

1 3 φ

provided we choose α small enough. In that case, writing x=

1 z, u0 (z)

5.4 Locally Convex Spaces

137

we see that x < 3 φ. Also, for each u ∈ X1∗ , |u(x) − φ(u)| =

1 |(u − φ(u)u0 ) (z)| < 3 φ (1 + φ) α, |u0 (z)|

which can be made less than ε by a suitably small choice of α.

2

Theorem 5.4.14. Let X be a separable Banach space, and φ a linear functional on X ∗ that is weak∗ -uniformly continuous on X1∗ . Then there exists x ∈ X such that φ(u) = u(x) for each u ∈ X ∗ . Proof. We may assume that φ < 1. Recursively applying Proposition 5.4.13, construct a sequence (xn )n1 of vectors in X such that for each n, n 1 u(xk ) < n φ(u) − 2

(u ∈ X1∗ )

(5.9)

k=1

and xn < 3/2n−1 . The series

∞ n=1

xn then converges to an element x of the com-

plete space X. Using the linearity and continuity of u, and letting n −→ ∞ in (5.9), we obtain the desired conclusion. 2

Let H be a nontrivial Hilbert space. One of the topologies on B(H) that is important in operator-algebra theory is the weak-operator topology τ w : that is, the locally convex topology deﬁned by the seminorms of the form T |T x, y| with x, y in H. Classically—but not constructively (see Exercise 17)—the sets of the type ⎫ ⎧ n ⎬ ⎨ |T ei , ej | < δ , T ∈ B(H) : ⎭ ⎩ i,j=1

with δ > 0 and {e1 , . . . , en } a set of pairwise orthogonal unit vectors in H, form a base of weak-operator neighbourhoods of 0 in B(H), so a linear functional φ on B(H) is τ w -continuous if and only if it has the following special continuity property:

SC

T here exist δ > 0 and a set {e1 , . . . , en } of pairwise orthogonal unit vectors n |T ei , ej | < δ, then |φ(T )| < 1. in H such that f or each T ∈ B(H), if i,j=1

We shall use the technique embodied in the proofs of Lemma 5.4.11, Lemma 5.4.12, and Proposition 5.4.13 to produce a characterisation of those linear functionals φ on B(H) with the property SC.

138

5 Linearity and Convexity

Before the next result, we mention that, in spite of what we wrote in the preceding paragraph, it is constructively provable that the sets of the form ⎫ ⎧ n ⎬ ⎨ |T ei , ej | < δ , T ∈ B1 (H) : ⎭ ⎩ i,j=1

with δ > 0 and {e1 , . . . , en } a set of pairwise orthogonal unit vectors in H, form a base of weak-operator neighbourhoods of 0 in the unit ball B1 (H). This is left as part of Exercise 17. Proposition 5.4.15. If H is a nontrivial Hilbert space, then B1 (H) is τ w -totally bounded. Proof. Let {e1 , . . . , en } be a ﬁnite set of pairwise orthogonal unit vectors generating a ﬁnite-dimensional subspace H0 of H. In view of the observation preceding this proposition, it will suﬃce to prove that B1 (H) is totally bounded with respect to the n seminorm pjk : T |T ej , ek | . Let P be the projection of H on H0 . Note that j,k=1

B(H0 ) is a ﬁnite-dimensional Banach space, and hence bounded unit has a totally 0 be an ε/n2 -approximation to ball, relative to the operator norm. Let T10 , . . . , Tm B1 (H0 ), and consider any T ∈ B1 (H). The restriction (P T )0 of P T to H0 belongs to B1 (H0 ), so there exists i such that (P T )0 − Ti0 < ε/n2 . Also, Ti0 P ∈ B1 (H). Thus if 1 j, k n, then , - , - , - T − Ti0 P ej , ek = T − Ti0 ej , P ek = P T − Ti0 ej , ek , - ε = (P T )0 − Ti0 ej , ek (P T )0 − Ti0 < 2 . n Hence

n , - 0 T − T 0 P ej , ek < ε. We now see that T10 P, . . . , Tm P is an εi

j,k=1

approximation to B1 (H) relative to the seminorm pjk .

2

Proposition 5.4.16. Let H be a nontrivial Hilbert space, and let φ be a linear functional on B(H) with the property SC. Then for each ε > 0 there exist a ﬁnite set {e1 , . . . , en } of pairwise orthogonal unit vectors in H, and elements cjk (1 j, k n) of K, such that n φ(T ) − < ε (T ∈ B1 (H)) . c T e , e jk j k j,k=1 Proof. With the proofs of Lemma 5.4.11, Lemma 5.4.12, and Proposition 5.4.13 at hand, we omit some of the grisly details of this one. We ﬁrst note that, in view of the

5.4 Locally Convex Spaces

139

observation immediately preceding Proposition 5.4.15, φ is τ w -uniformly continuous on B1 (H). It follows from this, Proposition 5.4.15, and Proposition 5.4.3 that φ = sup {|φ(T )| : T ∈ B1 (H)} exists. Scaling if necessary, we may assume that there exists T0 ∈ B1 (H) such that unit vectors in H, φ(T0 ) = 1. Choose a ﬁnite set {e1 , . . . , en } of pairwise orthogonal n and a positive number δ, such that if T ∈ B(H) and i,j=1 |T ei , ej | < 2δ, then |φ(T )| < 1. Let H0 = span {e1 , . . . , en } , and for each x = (x1 , . . . , xn ) in the n-fold product space H0n , deﬁne the product norm x =

n

xk .

k=1

Note that for each T ∈ B1 (H) and each x ∈ H0n , n n n T xk , ek |T xk , ek | xk = x . k=1

k=1

For such x, the mapping

(5.10)

k=1

n T T xk , ek k=1

is τ w -uniformly continuous on B1 (H). Since (by Proposition 5.4.15 and Lemma 5.4.9) the unit kernel N1 of φ is τ w -totally bounded, it follows from Proposition 5.4.3 that we can deﬁne a seminorm on H0n by n ! T xk , ek : T ∈ N1 . x0 = sup k=1

Then 0 , by (5.10), so the mapping 0 from (H0n , ) to R is uniformly continuous. Since H0n is ﬁnite-dimensional, the set {x ∈ H0n : x = 1} is -compact. Hence β = inf {x0 : x ∈ H0n , x = 1} is well deﬁned. This time (cf. Lemma 5.4.12) we prove that β = 0. Suppose that β > 0. Then (H0n , 0 ) is a ﬁnite-dimensional Banach space, and 0

u = sup {|u(x)| : x ∈ H0n , x0 1} ∗

deﬁnes the corresponding operator norm on the dual space (H0n ) . For each T ∈ N1 and each x ∈ H0n let

140

5 Linearity and Convexity

F (T )(x) =

n

T xk , ek .

k=1 ∗

0

The mapping F : N1 −→ ((H0n ) , ) is τ w -uniformly continuous, so F (N1 ) ∗ 0 is totally bounded, and therefore located, in ((H0n ) , ). Using the separation theorem as in the proof of Lemma 5.4.12, we can show that F (N1 ) is dense in the 0 unit ball S0∗ of (H0n )∗ , ). Since the linear functional Ψ :x

n

βT0 xk , ek

k=1 0

belongs to S0∗ , given t > 0 we can ﬁnd T ∈ N1 such that Ψ − F (T ) < n−2 δt. Write {(j, k) : 1 j, k n} = P ∪ Q, where P and Q are disjoint sets, (j, k) ∈ P =⇒ (βT0 − T ) ej , ek = 0, (j, k) ∈ Q =⇒ |(βT0 − T ) ej , ek | < n−2 δt. If (j, k) ∈ P, set rjk = (βT0 − T ) ej , ek

−1

|(βT0 − T ) ej , ek | ,

and if (j, k) ∈ Q, set rjk = 0. Now deﬁne ξ ∈ H0n by ξ k = n−2

n

(1 k n) ,

rjk ej

j=1

to obtain ξ k n−2

n

|rjk | n−1

j=1

and hence ξ0 ξ = Thus

ξ k 1.

k=1

n (βT0 − T ) ξ k , ek < n−2 δt. k=1

Moreover,

n

(5.11)

5.4 Locally Convex Spaces

141

n n n 2 n (βT0 − T ) ξ k , ek = rjk (βT0 − T ) ej , ek k=1 k=1 j=1 rjk (βT0 − T ) ej , ek = (j,k)∈P |(βT0 − T ) ej , ek | , = (j,k)∈P

so n

|(βT0 − T ) ej , ek |

j,k=1

|(βT0 − T ) ej , ek | +

(j,k)∈P

|(βT0 − T ) ej , ek |

(j,k)∈Q

n

(βT0 − T ) ξ k , ek + n2 n−2 δt 2δt, n2 k=1

the last step following from (5.11). Hence n −1 t (βT0 − T ) ej , ek < 2δ j,k=1

and therefore |φ(βT0 − T )| < t. We now have β = βφ(T0 ) − φ(T ) = φ(βT0 − T ) < t. Since t > 0 is arbitrary, it follows that β = 0. Thus for each α > 0 there exists z ∈ H0n such that z = 1 and n T zk , ek < (2n)−1 α (T ∈ N1 ) . k=1

We can arrange that each zk = 0, provided we replace (2n)−1 by n−1 in the last inequality; we assume that this has been done. Therefore n (T − φ(T )T0 ) zk , ek < n−1 (1 + φ) α (T ∈ B1 (H)) . k=1

We now introduce an operator T1 that will enable us to bound from 0. Setting T1 x = n−1

n k=1

we have

−1

zk

x, zk ek

(x ∈ H) ,

n k=1

T0 zk , ek away

142

5 Linearity and Convexity

T1 x n−1

n

−1

zk

x zk = x ,

k=1

so T1 ∈ B1 (H). Also, n

T1 zk , ek = n−1

k=1

Hence n−1

n

zk

−1

zk , zk = n−1

k=1

n

zk = n−1 .

k=1

n n (T1 − φ(T1 )T0 ) zk , ek + |φ(T1 )| T0 zk , ek k=1 k=1 n T0 zk , ek n−1 (1 + φ) α + φ k=1

and therefore n −1 −1 T0 zk , ek (n φ) (1 − (1 + φ) α) > (2n φ) k=1

provided α is small enough. Setting n −1 x= T0 zk , ek z, k=1

for each T ∈ B1 (H) we now obtain −1 n n n n T xk , ek = T0 zk , ek φ(T )T0 zk , ek − T zk , ek φ(T ) − k=1

k=1

k=1

k=1

< 2n φ n−1 (1 + φ)α = 2 φ (1 + φ) α, which can be made less than ε by choosing α suitably small. It remains to take 2 cjk = xk , ej for 1 j, k n.

Exercises 1. Let U, V be subsets of a Banach space X such that U ∪ V is dense. Prove the following: (a) If u ∈ U and v ∈ V, then for each ε > 0 there exist t ∈ [0, 1] and x ∈ U ∩ V such that x − tu − (1 − t)v < ε.

5.4 Locally Convex Spaces

143

(b) For each x ∈ X,

ρ x, U ∩ V = min {ρ(x, U ), ρ(x, V )} , in the sense that each side of this equation exists if and only if the other does. 2. We say that a subset S of a normed space X has the boundary crossing property if for each ξ ∈ S, each z ∈ ∼S, and each ε > 0 there exist t ∈ [0, 1] and x ∈ ∂S such that x − tξ − (1 − t) z < ε. Prove that if X is complete and S ∪ ∼S is dense, then S has the boundary crossing property. Do we necessarily obtain this conclusion if S ∪ ∼S is not known to be dense in X? 3. Let S be a subset of a Banach space such that S ∪∼S is dense, and let x0 ∈ ∼S. Prove that ρ (x0 , ∂S) exists if and only if ρ (x0 , S) exists, in which case these two distances are equal. 4. Let {e1 , . . . , en } be the canonical orthonormal basis of the Euclidean space Rn , and Σ the simplex whose vertices are ±ej (1 j n). Let ξ belong to the metric complement −Σ of Σ in Rn . Prove that the segment [0, ξ] contains points arbitrarily close to the union of the faces of Σ. 5. Show that if the Hahn–Banach theorem holds with exact preservation of the norm of the functional, then LLPO holds. 6. Let Y be a ﬁnite-dimensional subspace of a normed space X, and x0 ∈ X. Suppose that ρ(x0 , Y ) > 0 and that X is the span of Y ∪ {x0 } . Let v be a nonzero normed linear functional on Y, and ε > 0. Without using the separation theorem, the Hahn–Banach theorem, or any of their consequences, prove that there exists a normed linear functional u on X such that u v + ε and u(y) = v(y) for each y ∈ Y. Use this to give another proof of the Hahn–Banach theorem. 7. Use the Hahn–Banach theorem to prove Proposition 5.3.1. 8. Let p be an integer 2. Prove that the Banach space lp is uniformly convex and has Gˆ ateaux diﬀerentiable norm. (See Exercise 11 of Chapter 4.) 9. Let X be a normed space with Gˆ ateaux diﬀerentiable norm. Let C be a closed, located, convex subset of X, let x ∈ −C, and let y ∈ C. Prove that x − y = ρ (x, C) if and only if Re ux−y (z − y) 0 for all z ∈ C. 10. Let X be a uniformly convex Banach space with Gˆ ateaux diﬀerentiable norm. Let A and B be subsets of X whose algebraic diﬀerence is located and convex, and whose mutual distance is positive. Prove that there exists a normed linear functional u on X, with norm 1, such that Re u (y) Re u (x) + d for all x ∈ A

144

5 Linearity and Convexity

and y ∈ B. (The results embodied in Exercises 9 and 10 are due to Ishihara [55].) 11. Prove that if, under the other hypotheses of Lemma 5.4.9, the conclusion holds without the hypothesis that φ is either 0 or nonzero, then we can derive the essentially nonconstructive proposition ∀x ∈ R (x = 0 ∨ ¬ (x = 0)) . 12. Let f be a uniformly continuous linear functional on X, and S a balanced, totally bounded subset of X. Prove that the set St = {x ∈ S : |f (x)| t} is totally bounded for each t > 0. 13. Let (X, p) be a seminormed space, and S a balanced, totally bounded subset of X. Let the mapping f : X −→ K be both uniformly continuous on S and homogeneous—that is, f (λx) = λf (x) for all λ ∈ K and x ∈ X. Prove that the set St = {x ∈ S : |f (x)| t} is totally bounded for each t > 0. 14. Let X be a ﬁnite-dimensional locally convex space, and let S be the set of convex combinations of points in an inhabited ﬁnitely enumerable subset of X. Prove that S is totally bounded in X. 15. Fill in the missing details in the proof of Proposition 5.4.16. 16. Let (en )n1 be an orthonormal basis of a separable Hilbert space H, and let φ be a linear functional on B(H) that is weak-operator uniformly continuous on B1 (H). Prove that there exist complex numbers cjk (j, k 1) such that ∞ φ(T ) = cjk T ej , ek for all T ∈ B(H). j,k=1

17. Prove the following for a nontrivial Hilbert space H: (a) The sets of the form ⎫ ⎧ n ⎬ ⎨ |T ei , ej | < δ , T ∈ B1 (H) : ⎭ ⎩ i,j=1

with δ > 0 and {e1 , . . . , en } a set of pairwise orthogonal unit vectors in H, form a base of weak-operator neighbourhoods of 0 in the unit ball B1 (H).

5.4 Locally Convex Spaces

145

(b) If B1 (H) can be replaced by B(H) in (a), then LPO holds. 18. Let C be a convex subset of Cn such that sup {f (z) : z ∈ C} exists for each real (bounded) linear functional f on Cn . Prove that C is totally bounded. (Hint: Prove that C is bounded; that ∗ d(x) = sup inf Re f (x − y) : f ∈ S y∈C

exists for each x ∈ Cn , where S ∗ is the unit ball of the dual of Cn ; and that ρ(x, C) exists and equals d(x).) 19. Let X, Y be normed spaces. We say that φ is ultraweakly continuous if it is uniformly continuous on B1 (X, Y ) relative to the locally convex topology τ w deﬁned on B(X, Y ) by the seminorms T y ∗ (T x) with x ∈ X and y ∗ a normed linear functional on Y. Prove that the following conditions are equivalent: (a) B1 (X, Y ) is τ w -totally bounded. (b) Every ultraweakly continuous linear functional on B(X, Y ) is normed. (c) Every ultraweakly continuous linear functional on B (X, Y ) is compact.

Notes A classical method of establishing the continuity of the boundary crossing map in Proposition 5.1.5 is to use a contradiction argument to prove that the mapping is sequentially continuous on −C; see [87] (pages 271–272). Our argument, based on the simple Lemma 5.1.3, is much more transparent and shorter than either that classical argument or its constructive counterpart (which is given in [32] and proves only the sequential continuity, not the full continuity, of the boundary crossing map). More information about boundary crossing can be found in [36], from which Exercises 2–4 are taken. The Hahn–Banach theorem can be proved without recourse to the separation theorem (Exercise 6). Using the Church–Markov–Turing thesis, Metakides et al. [71] have produced an explicit example in which the hypotheses of the Hahn–Banach theorem hold but the linear functional cannot be extended with exact preservation of its norm. This makes Theorem 5.3.7 (which, together with Exercises 9 and 10, is due to Ishihara [55]) all the more signiﬁcant, especially as it applies to many important examples of normed spaces, such as the Lp spaces for p > 1. The locally convex topology on a ﬁnite-dimensional space X is unique and equivalent to the topology induced by any norm on X. The proof of this depends on some intricate geometric algorithms and is found in [39].

146

5 Linearity and Convexity

Finding necessary and suﬃcient conditions for the existence of the norm of a bounded linear mapping T : X −→ Y between normed spaces, other than the case Y = K, is an important unsolved problem, even in the case where T is an operator on a Hilbert space. The Banach–Alaoglu theorem (Theorem 5.4.7) without the hypothesis of separability is classically equivalent to the ultraﬁlter principle—every ﬁlter is contained in an ultraﬁlter (see [80], page 766)—and is therefore unlikely to be provable constructively; see Exercise 1.4 of Chapter 1. The introduction of the numbers rjk in the proof of Proposition 5.4.16 is occasioned by the problem with the modulus–argument decomposition discussed at the start of Chapter 2. Proposition 5.4.16 is a step in the direction of a constructive proof of the classical characterisation of weak-operator continuous linear functionals as those of the n T xk , yk with the vectors xk , yk in H. All attempts to generalise the form T k=1

technique used to prove Proposition 5.4.13 in order to characterise ultraweakly continuous linear functionals—those linear functionals on B(H) that are weak-operator uniformly continuous on B1 (H)—suﬀer from a curse of dimensionality, failing ben T xk , yk is usually more than 1. cause the n appearing in expressions like k=1

Although Exercise 16 requires only the weak-operator uniform continuity of the functional on the unit ball of B(H) and derives a stronger conclusion than that of Proposition 5.4.16, the solution that we have in mind for that exercise works only for a separable Hilbert space. In that case we can improve the conclusion to ﬁnd square-summable sequences (xn )n1 and (yn )n1 in H such that for each T ∈ B(H) ∞ we have φ(T ) = T xn , yn ; see [23]. n=1

If H is a Hilbert space, then B1 (H) is classically weak-operator compact; however, the completeness of B1 (H) relative to the uniform structure associated with the weak-operator topology is an essentially nonconstructive property; see [14]. The result in Exercise 18 ﬁrst appeared in [56].

6 Operators and Locatedness

Location! Location! Location! —Unknown Estate Agent

We begin by introducing normed spaces on which the norm is diﬀerentiable in some fashion. In Section 2, with substantial help from the λ-technique, we provide criteria for the locatedness of certain convex sets in a normed space. This work is applied in Section 3 to proving that a bounded operator on a Hilbert space H has an adjoint if and only if it maps the unit ball to a located set. In the next section we construct approximate eigenvectors of a selfadjoint operator on H, and then show that a bounded positive operator has a unique positive square root. This result is applied in Section 5, in which we make further good use of the λ-technique to show that, for a so-called weak-sequentially open operator T on H, the range of T is located if and only if the range of its adjoint is located. The section ends with a proof of the closed range theorem for operators with an adjoint. The ﬁnal section of the chapter presents a version of Baire’s theorem, which is then applied to the proofs of three of the pillars of functional analysis: the open mapping, inverse mapping, and closed graph theorems.

6.1 Smooth and Uniformly Smooth Normed Spaces As the reader will have observed, locatedness, classically a nugatory concept, plays an important role in many aspects of constructive analysis, such as the Hahn– Banach theorem and the separation theorem discussed in the preceding chapter. It is therefore ﬁtting that criteria for locatedness should be a primary theme of this ﬁnal chapter. Our ﬁrst aim is to ﬁnd necessary and suﬃcient conditions for the locatedness of convex sets in a normed space; this requires some preliminary results associated with the diﬀerentiability of the norm. Later in the chapter we shall apply our conditions to the locatedness of subsets associated with operators on a Hilbert space.

148

6 Operators and Locatedness

Throughout this section, unless we state otherwise, X will be a normed linear space. We say that X is smooth if its norm is Gˆ ateaux diﬀerentiable at each nonzero vector; this is the case if and only if the limit x + ty − x t→0 t

ux (y) = lim

(6.1)

exists for all unit vectors x, y ∈ X. We say X has Fr´echet diﬀerentiable norm if this limit is uniform in the unit vector y: that is, if for each unit vector x ∈ X and each ε > 0 there exists δ > 0 such that x + ty − x − ux (y) < ε (6.2) t whenever y ∈ X, y = 1, and 0 < |t| < δ. If, moreover, δ can be chosen independent of x, then we say that X has uniformly Fr´echet diﬀerentiable norm and is uniformly smooth. For this to be the case, for each ε > 0 there must exist δ > 0 such that (6.2) holds whenever x, y ∈ X, x = y = 1, and 0 < |t| < δ. With each vector x in a smooth normed space X we associate the mapping Jx : X −→ R deﬁned by Jx (y) = lim x t→0

x + ty − x . t

If x = 0, then Jx (y) = 0; if x = 0, then Jx (y) = x ux (y). In both these cases, |Jx (y)| x y . In order to show that Jx (y) is deﬁned for an arbitrary element x of X, construct an increasing binary sequence (λn )n1 such that 1 , n 1 . λn = 1 =⇒ x > n+1

λn = 0 =⇒ x

1; so (ζ n )n1 is a Cauchy sequence and therefore converges to a limit ζ ∈ R. Note that |ζ| x y . For each ε > 0, either x y < ε/2 or x y > 0. In the ﬁrst case we have x x + ty − x − ζ 2 x y < ε t for all t = 0. In the second case, ζ = Jx (y) and x x + ty − x − ζ < ε t for all suﬃciently small t = 0. Since ε > 0 is arbitrary, we conclude that Jx (y) exists and equals ζ.

6.1 Smooth and Uniformly Smooth Normed Spaces

149

Proposition 6.1.1. If X is smooth, then for each x in X, Jx is a normed linear 2 functional on X, Jx = x, and Jx (x) = x . Moreover, if c > 0 and x ∈ X, then Jcx = cJx . Proof. If x = 0, then the ﬁrst three conclusions follow from Proposition 5.3.6. On the other hand, if also c > 0, then Jcx = cx ucx = c x ux = cJx . 2

The extension to the general case is left to the reader. Lemma 6.1.2. If X is smooth, then for all x, y ∈ X, 2

2

y x − 2Jy (x − y) . Proof. Using the linearity of Jy and also Proposition 6.1.1, we compute 2

2

Jy (x − y) = Jy (x) − y y x − y . Hence 2

2

2

2

2

x − y − 2Jy (x − y) x + y − 2 x y = (x − y) 0. 2 Lemma 6.1.3. Let X be a smooth normed space, and x a unit vector in X. Let δ, ε be positive numbers such that if y = 1 and 0 < |t| < δ, then x + ty − x − Jx (y) < ε. t Then for such t and each unit vector y, x + ty + x − ty < 2 + 2tε. Proof. We may assume that 0 < t < δ. If y = 1, then x + ty − x − Jx (y) < ε t and

x − ty − x − Jx (y) > −ε. −t

Hence x + ty < 1 + tJx (y) + tε and x − ty < 1 − tJx (y) + tε, from which the result follows.

2

150

6 Operators and Locatedness

We store the following result for use in Section 2. Proposition 6.1.4. Let X be a smooth, uniformly convex Banach space, and let f be a normed linear functional on X. Then there exists a unique vector x ∈ X such that f = Jx . Proof. First assume that f is nonzero. By Proposition 2.3.7, there exists a unit −1 vector y ∈ X such that f (y) = f . Applying Proposition 5.3.6 to f f, we see that f = f uy . Now let x = f y. Then x = f and (since ucy = uy for all c > 0) f = f uy = x u f y = x ux = Jx . It remains to deal with the uniqueness. Let f = Jz for some z ∈ X. Then, by Proposition 6.1.1, z = Jz = f > 0, so f uy = f = f uz = f u z −1 z −1

and therefore u z −1 z = uy . Since z z is a unit vector, the uniqueness part of −1 Proposition 5.3.6 shows that z z = y; whence z = z y = f y = x. This completes the proof in the case where f is nonzero; the general case is left as an exercise. 2

The next lemma simpliﬁes the proof of the theorem following it. Lemma 6.1.5. Let f be a mapping of R − {0} into R such that (a) if 0 < t t, then f (−t) f (−t ) f (t ) f (t), and (b) for each ε > 0 there exists t > 0 such that f (t) − f (−t) < ε. Then limt→0 f (t) exists. Proof. Construct a sequence (tn )n1 of positive numbers decreasing strictly to 0 such that f (tn ) − f (−tn ) < 1/n for each n. In view of (a), for m n we have 0 < f (tm ) − f (tn ) f (tm ) − f (−tm )

0, choose t > 0 as in (b) and then N such that tN < t and 0 f (tN ) − l < ε. By (a), if 0 < |t| < tN , then 2 f (−tN ) f (t) f (tN ) and therefore |f (t) − l| f (tN ) − l < ε.

6.1 Smooth and Uniformly Smooth Normed Spaces

151

We have already deﬁned “uniformly convex” for normed spaces, in Chapter 2. To introduce this notion for the dual of a normed space X, we concentrate on those elements of X ∗ that have a norm. We say that X ∗ is uniformly convex if for each ε > 0 there exists δ with 0 < δ < 1 such that for all normed u, v ∈ X ∗ with u = v = 1, if there exists a unit vector ξ ∈ X such that 12 (u + v) (ξ) > 1 − δ, then (u − v) (x) < ε for all unit vectors x. Proposition 6.1.6. A separable normed space is uniformly smooth if and only if it has a uniformly convex dual. Proof. Let X be a separable normed space, and suppose to begin with that X ∗ is uniformly convex. Given unit vectors x, y ∈ X, for all nonzero t ∈ R deﬁne f (t) =

x + ty − x . t

Then f satisﬁes (a) of Lemma 6.1.5. It follows that in order to prove that X is uniformly smooth, we need only show that for each ε > 0 there exists t > 0, independent of the unit vectors x and y, such that f (t) − f (−t) < ε. To this end, let δ > 0 be as in the deﬁnition of “uniformly convex” above, ﬁx t > 0 such that 8t(2 + t) < δ, and consider any unit vectors x, y ∈ X. Note that x + ty > 1/2 and x − ty > 1/2. Deﬁning unit vectors ξ, η by ξ=

1 (x + ty), x + ty

η=

1 (x − ty), x − ty

we obtain 1 x − ty (x + ty) − x + ty (x − ty) x + ty x − ty 4 (x − ty − x + ty) x + t (x − ty + x + ty) y

ξ − η =

4 [|x − ty − x + ty| + t (x − ty + x + ty)] 4 (2t + 2t(1 + t)) = 8t(2 + t) < δ.

Now let γ = min

δ tε , 4(1 + t) 2

,

and use Proposition 5.3.1 to construct normed linear functionals u and v such that u = v = 1, u(ξ) > 1 − γ, and v(η) > 1 − γ. Then δ (u + v) (ξ) = u(ξ) + v(η) + v (ξ − η) > 2 − 2γ − ξ − η 2 − 2 − δ = 2 − 2δ 2 and therefore 12 (u + v)(ξ) > 1 − δ. It follows that (u − v) (z) < ε/2 for all unit vectors z ∈ X, and in particular that (u − v) (y) < ε/2. Now,

152

6 Operators and Locatedness

u (x + ty) = x + ty u(ξ) > x + ty (1 − γ) , so x + ty < u(x + ty) + x + ty γ u(x + ty) + (1 + t)γ u(x + ty) +

tε . 4

Likewise, x − ty < v(x − ty) +

tε . 4

Thus 1 x + ty − x x − ty − x − = (x + ty + x − ty − 2) t −t t tε 1 −2 u(x + ty) + v(x − ty) + t 2 tε 1 −2 u(x) + v(x) + t(u − v)(y) + = t 2 ε + (u − v) (y) < ε. 2 This completes the proof that X is uniformly smooth. Now suppose, conversely, that X is uniformly smooth. Then for each ε > 0 there exists δ such that ε ,1 0 < δ < min 4 and such that for all unit vectors x, y ∈ X and all t with 0 < |t| < 4δ/ε, ε x + ty − x − Jx (y) < . t 4 Letting t = 2δ/ε, we see from Lemma 6.1.3 that for all unit vectors x, y ∈ X, x + ty + x − ty < 2 +

tε . 2

Now consider normed u, v ∈ X ∗ with u = v = 1, and assume that 12 (u+v)(ξ) > 1 − δ for some unit vector ξ ∈ X. Then for all unit vectors x ∈ X we have 2−

tε = 2 − δ < (u + v)(ξ) 2 = u(ξ + tx) + v(ξ − tx) − t(u − v)(x) ξ + tx + ξ − tx − t(u − v)(x) tε < 2+ − t(u − v)(x) 2

and therefore (u − v)(x) < ε. Thus X ∗ is uniformly convex.

2

6.2 Locatedness of Convex Sets

153

If H is a Hilbert space, then, in view of the Riesz representation theorem, its dual is, like H itself, uniformly convex. It follows from Proposition 6.1.6 that a separable Hilbert space is uniformly smooth. In fact we can remove separability and completeness here. Proposition 6.1.7. An inner product space is uniformly smooth. Proof. Given unit vectors x, y in the inner product space X, for real t we have 2

2

x − ty − x = −2t Re x, y + t2 and therefore x − ty − x =

−2t Re x, y + t2 . x − ty + x

It follows that if 0 < |t| < 1/2, then x − ty − x −2 Re x, y + t = + Re x, y + Re x, y t x − ty + x 1 |(x − ty − 1) Re x, y + t| = x − ty + 1 4 |t| 2 |t| (Re x, y + 1) , 3 3 which tends to 0 with t, independently of x and y.

2

6.2 Locatedness of Convex Sets Our aim in this section is to give conditions ensuring the locatedness of a bounded convex subset of a normed space under certain strong hypotheses that apply, in particular, to the norm on a Hilbert space or an Lp -space1 for 1 < p < ∞. Speciﬁcally, we aim to prove the following two results and some of their consequences. Theorem 6.2.1. Let X be a uniformly smooth normed space over R, and C an inhabited, bounded, located convex subset of X. Then sup {Jx (y) : y ∈ C} exists for each x ∈ X.

1

We refer the reader to [9, 12] for the theory of Lp spaces.

154

6 Operators and Locatedness

Theorem 6.2.2. Let C be an inhabited, bounded, convex subset of a uniformly smooth normed space X over R, such that sup {Jx (y) : y ∈ C} exists for each x ∈ X. Then C is located. The path to Theorem 6.2.1 takes us through a tangle of technical lemmas, the ﬁrst of which is an expression of the continuity of Jx as a function of x. Lemma 6.2.3. Let X be a real normed space with a Fr´echet diﬀerentiable norm. For each unit vector x ∈ X and each ε > 0 there exists δ > 0 such that if y < δ, then |Jx (z) − Jx+y (z)| < ε for all unit vectors z ∈ X. Proof. Given a unit vector x ∈ X and ε > 0, compute γ ∈ (0, 1) such that if z = 1 and 0 < |t| < γ, then ε x + tz − x < . − J (6.3) (z) x 8 t

Let δ = min

γε 1 , 32 2

and consider a vector y ∈ X with y < δ. Noting that x + y > 1/2, set ξ=

1 (x + y). x + y

Then since |1 − x + y| = |x − x + y| y , we have x + y ξ − x = x + y − x + y x = (1 − x + y) x + y 2 y . 2

Also, by Proposition 6.1.1, Jξ (ξ) = ξ = 1, so 0 1 − Jξ (x) = Jξ (ξ − x) ξ − x

2 y < 4δ. x + y

(6.4)

On the other hand, setting t = γ/2 and applying Lemma 6.1.3, we see that for each unit vector z ∈ X, tε x + tz + x − tz < 2 + . 4 Therefore, by (6.4), the choice of t and δ, and Proposition 6.1.1,

6.2 Locatedness of Convex Sets

1−

155

tε 1 − 4δ 4 < Jξ (x) = Jx (x) + Jξ (x) − 1 = Jx (x + tz) + Jξ (x − tz) − 1 − t(Jx − Jξ )(z) x + tz + x − tz − 1 − t(Jx − Jξ )(z) tε < + 1 − t(Jx − Jξ )(z). 4

Hence (Jx − Jξ )(z) < ε/2. Replacing z by −z, we obtain −(Jx − Jξ )(z) < ε/2 and therefore |(Jx − Jξ )(z)| < ε/2. It follows from this and Proposition 6.1.1 that |(Jx − Jx+y )(z)| = |x + y (Jx − Jξ )(z) + (x − x + y) Jx (z)| x + y |(Jx − Jξ )(z)| + |x − x + y| |Jx (z)| 3 |(Jx − Jξ )(z)| + y x z 2 ε 3ε + < ε, < 4 32 2

as we wanted.

Lemma 6.2.4. Let X be a uniformly smooth normed space over R, and let C be an inhabited, bounded, located convex subset of X. Let x0 ∈ X and ε > 0. Then there exists τ > 0 such that for all y ∈ C, either there exists x ∈ C such that Jx0 (y) + τ ε/6 < Jx0 (x) or else Jx0 (z) < Jx0 (y) + ε for all z ∈ C. Proof. Choose M > 0 such that x − y < M for all x, y ∈ C. If M x0 < ε, then for all x, y ∈ C, Jx0 (x − y) x0 x − y M x0 < ε. Hence we may assume that M x0 > 0. From the deﬁnition of “uniformly smooth” and Lemma 6.2.3 we see that there exists δ > 0 such that for all unit vectors x, y ∈ X, • if 0 < |t| δ, then x + ty − y 2ε < − J , (y) x t 3 x0 M and

156

6 Operators and Locatedness

• if y δ, then for all unit vectors z ∈ X, |Jx (z) − Jx+y (z)|

0. Since (1 − τ ) y + τ x belongs to the convex set C, we see from (6.7) and Proposition 6.1.1 that −

2τ ε 2 2 < x0 + τ (y − x) − x0 3 2 = Jx0 +τ (y−x) (x0 + τ (y − x)) − x0 2

= Jx0 +τ (y−x) (x0 ) + τ Jx0 +τ (y−x) (y − x) − x0 2

x0 + τ (y − x) x0 − x0 + τ Jx0 +τ (y−x) (y − x). Hence 0 < x0

x0 + τ (y − x) − x0 2ε + Jx0 +τ (y−x) (y − x) + . τ 3

(6.8)

6.2 Locatedness of Convex Sets

157

The appearance here of the expression E=

x0 + τ (y − x) − x0 τ

in the presence of a uniformly Fr´echet diﬀerentiable norm suggests that we consider the unit vectors 1 1 x0 , v = (y − x). u= x0 y − x Writing t= we have

τ y − x , x0

u + tv − u 1 E= . y − x t

Since 0 < t δ, it follows that 2ε 1 > E − Ju (v) 3 x0 M y − x 1 1 E− Jx0 (y − x) = y − x x0 y − x 1 |x0 E − Jx0 (y − x)| = x0 y − x and therefore x0 x0 + τ (y − x) − x0 − Jx0 (y − x) < 2ε y − x < 2ε . τ 3M 3

(6.9)

In view of (6.8) and (6.9), it now makes sense to examine Jx0 (y − x) − Jx +τ (y−x) (y − x) , 0 which, by Proposition 6.1.1, equals x0 y − x |Ju (v) − Ju+tv (v)| . By our choice of δ, this last expression is less than x0 y − x

2ε , 3 x0 M

so

Jx (y − x) − Jx +τ (y−x) (y − x) < 2ε . 0 0 3 It follows from (6.8)–(6.10) that

(6.10)

158

6 Operators and Locatedness

x0 + τ (y − x) − x0 2ε + Jx0 +τ (y−x) (y − x) + 0 < x0 τ 3 2ε 2ε 2ε + Jx0 (y − x) + + , Jx0 (y − x) + 3 3 3 from which we obtain the inequality Jx0 (x) < Jx0 (y) + ε. Since x was an arbitrary point of C, the proof is complete. 2

Using Lemma 6.2.4 and the λ-technique, we can now prove Theorem 6.2.1. Proof. Under the hypotheses of Theorem 6.2.1, ﬁx M > 0 such that x − y M for all x, y ∈ C, and consider any element x of X. It is enough to prove that for each ε > 0 there exists y ∈ C such that Jx (z) < Jx (y) + ε

(z ∈ C) .

(6.11)

For in that case, if 0 < α < β and we choose y ∈ C such that (6.11) holds with ε = (β − α) /2, then either Jx (y) > α, or else Jx (y) < 12 (α + β) and therefore Jx (z) < β for all z ∈ C; so the desired supremum exists by the least-upper-bound principle. Fixing x ∈ X, y0 ∈ C, and ε > 0, let τ be as in Lemma 6.2.4. We construct an increasing binary sequence (λn )n0 with λ0 = 0, and a sequence (yn )n0 in C, such that τε < Jx (yn ), 6 λn = 1 =⇒ Jx (z) < Jx (yn−1 ) + ε for all z ∈ C.

λn = 0 =⇒ Jx (yn−1 ) +

To do so, we assume that we have constructed λ0 , . . . , λn and y0 , . . . , yn . If λn = 1, set λn+1 = 1 and yn+1 = yn . If λn = 0, then by Lemma 6.2.4, either Jx (yn )+τ ε/6 < Jx (y ) for some y ∈ C, or else Jx (z) < Jx (yn ) + ε for all z ∈ C. In the ﬁrst case set λn+1 = 0 and yn+1 = y . In the second case set λn+1 = 1 and yn+1 = yn . This completes the inductive construction. Now choose N such that M x < N If λN = 0, then Jx (y0 ) + N

τε . 6

τε < Jx (yN ) 6

and therefore

τε < Jx (yN − y0 ) M x , 6 a contradiction. Hence, in fact, λN = 1. N

2

6.2 Locatedness of Convex Sets

159

Corollary 6.2.5. Let X be a uniformly convex, uniformly smooth Banach space over R, and let C be an inhabited, bounded, located convex subset of X. Then sup {f (y) : y ∈ C} exists for each normed linear functional f on X. Proof. For a nonzero normed linear functional the result follows from Proposition 6.1.4 and Theorem 6.2.1. The completion of the proof is left to the reader. 2

Turning now towards Theorem 6.2.2, we prove the following lemma. Lemma 6.2.6. Let C be an inhabited, bounded, convex subset of a uniformly smooth normed space X over R such that sup{Jx (z) : z ∈ C} exists for each x ∈ X, and let ε > 0. Then for each x ∈ X there exists σ with 0 < σ < 1 such that for each y ∈ C, either there exists y ∈ C with x − y < (1 − σ) x − y or else x − y < x − z + ε for all z ∈ C. Proof. We may assume that ε < 1. Fixing x ∈ X, choose a positive integer M > max {ε/4, 1} such that x − z < M for each z ∈ C. Since X is uniformly smooth, there exists δ with 0 < δ < 2 such that for all unit vectors u, v ∈ X, u − tv − u ε − Ju (v) < . (6.12) 0 < |t| < δ =⇒ −t 8M Deﬁne

εδ 8M and note that 0 < σ < 1. Let y ∈ C. If x − y < ε, then x − y < x − z + ε for all z ∈ C. Hence we may assume that ε/2 < x − y. Now, σ=

sup {Jx−y (z − y) : z ∈ C} = sup {Jx−y (z) : z ∈ C} − Jx−y (y) exists, by our hypotheses. So either Jx−y (z − y) < ε x − y for all z ∈ C or else there exists z0 ∈ C such that ε x − y < Jx−y (z0 − y) . 2 In the ﬁrst case, for all z ∈ C we have

(6.13)

160

6 Operators and Locatedness 2

x − y = Jx−y (x − y) = Jx−y (x − z) + Jx−y (z − y) < x − y x − z + ε x − y , and therefore x − y < x − z + ε. In the second case, setting τ=

δ x − y , 2M

y = y + τ (z0 − y),

we have 0 < τ < δ/2 < 1 and therefore y ∈ C. Deﬁne also u=

1 (x − y), x − y

v=

1 (z0 − y), z0 − y

t=

τ z0 − y . x − y

Note, for the deﬁnition of v, that z0 − y x − y Jx−y (z0 − y) >

ε x − y 2

and hence that z0 − y > 0. Let t=

δ z0 − y τ z0 − y = . x − y 2M

Then t < δ and, by Proposition 6.1.1 and (6.13), tJu (v) =

τ τ τε Ju (z0 − y) = 2 Jx−y (z0 − y) > 2 x − y . x − y x − y

Hence x − y = x − y − τ (z0 − y) = x − y u − tv tε < x − y u − tJu (v) + 8M τ ε τ ε z0 − y + . < x − y − 2 8M Since z0 − y z0 − x + x − y < 2M and (6.13) holds, this gives τε τε + 2 4 δε x − y = x − y − 8M = (1 − σ) x − y

x − y < x − y −

by (6.12)

6.2 Locatedness of Convex Sets

161

and completes the proof that, in this case, the ﬁrst conclusion of the lemma obtains. 2

We now prove Theorem 6.2.2. Proof. Under the hypotheses of Theorem 6.2.2, let x ∈ X and ε > 0. It suﬃces to show that there exists y ∈ C such that x − y < x − z + ε for all z ∈ C. Compute σ > 0 as in Lemma 6.2.6, and let y0 ∈ C. If x − y0 < ε, then x − y0 < x − z + ε for all z ∈ C. So we may assume that x − y0 > 0. Now construct an increasing binary sequence (λn )n0 with λ0 = 0, and a sequence (yn )n0 of elements of C, such that λn = 0 =⇒ x − yn < (1 − σ) x − yn−1 , λn = 1 =⇒ x − yn < x − z + ε for all z ∈ C. Suppose we have constructed λ0 , . . . , λn and y0 , . . . , yn . If λn = 1, set λn+1 = 1 and yn+1 = yn . If λn = 0, then by Lemma 6.2.6, either x − y < (1 − σ) x − yn for some y ∈ C, or else x − yn < x − z + ε for all z ∈ C. In the ﬁrst case set λn+1 = 0 and yn+1 = y ; in the second case set λn+1 = 1 and yn+1 = yn . Since 0 < σ < 1, we see that (1 − σ)n −→ 0 as n −→ ∞, so there exists N > 1 such that (1 − σ)n

ε/2. Our hypotheses ensure that M = sup {Re x − y, z − y : z ∈ C} exists. Either M < ε2 /2 or else ε2 /4 < M. In the ﬁrst case, for all z ∈ C we have 2

x − y = Re x − y, x − y = Re x − y, x − z + Re x − y, z − y ε2 < x − y x − z + 2 and therefore x − y < x − z +

ε2 < x − z + ε. 2 x − y

Hence (c) holds. We may therefore assume that ε2 /4 < M. Choosing z0 ∈ C such that

6.2 Locatedness of Convex Sets

τ = Re x − y, z0 − y >

163

ε2 , 4

we consider three exhaustive cases. 2

z0 − y < 3τ /2. We have

Case 1:

2

2

x − z0 = x − y − (z0 − y) 2

2

= x − y − 2τ + z0 − y 1 2 < x − y − τ 2 2 ε 2 < x − y − 8 ε4 2 , < x − y − 16n since 0 < ε < 1. Hence (a) holds with y0 = z0 . 2

n/2 < z0 − y . Using the parallelogram identity, we have

Case 2:

n 2 < (z0 − x) + (x − y) 2 2 2 2 = 2 x − z0 + 2 x − y − z0 + y − 2x 2

2

2 x − z0 + 2 x − y , so (b) holds. 2

2

Case 3: τ < z0 − y and z0 − y < n. Then y0 = y +

τ 2

z0 − y

(z0 − y)

belongs to C, and 2

2

x − y0 = x − y − 2

= x − y −

2τ Re x − y, z0 − y 2

z0 − y

+

τ2 2

z0 − y

τ2 2

z0 − y ε4 2 < x − y − 16n and so (a) holds. This completes the proof.

The classical uniform boundedness theorem says:

2

164

6 Operators and Locatedness

If (Ti )i∈I is a family of bounded linear mappings from a Banach space X into a normed space Y such that for each x ∈ X the family (Ti x)i∈I is bounded, then there exists M > 0 such that Ti M for each i ∈ I. The reader is invited to consider why this form of the theorem is unlikely to be provable constructively. The correct constructive approach is via the contrapositive; here is a pretty version due to Royden [78]. Theorem 6.2.11. (Royden’s uniform boundedness theorem) Let (Tn )n1 be a sequence of normed linear mappings from a Banach space X into a normed space Y such that Tn > n3n for each n. Then there exists x ∈ X such that Tn x > n for each n. Proof. For each n 1 choose xn ∈ X such that xn = 3−n and Tn xn > 3 4 Tn xn . Then ∞ xn x=4 n=1 ∞

is well deﬁned, since (by comparison with

3−n ) the series on the right converges

n=1

in the complete space X. For n 2, letting

yn = x1 + · · · + xn−1 , we have 3 Tn xn < 2 Tn xn = T (xn + yn ) + T (xn − yn ) 2 Tn (xn + yn ) + T (yn − xn ) , so at least one of the last two terms is greater than −xn , if necessary, we may assume that Tn (xn + yn ) > Hence

3 4

Tn xn . Replacing xn by

3 Tn xn . 4

n ∞ 1 Tn x = Tn xk + Tn xk 4 k=1 k=n+1 n−1 ∞ xk − Tn xk Tn x n + k=1

k=n+1

∞

3 Tn 3−n − Tn 3−k 4 k=n+1 1 3 1 − = Tn 3−n Tn 3−n 4 2 4

>

6.2 Locatedness of Convex Sets

and therefore Tn x > 3−n Tn > n.

165

2

Theorem 6.2.12. (Uniform boundedness theorem) Let (Tn )n1 be a sequence of normed linear mappings from a Banach space X into a normed space Y such that Tn −→ ∞ as n −→ ∞. Then there exists x ∈ X such that the sequence (Tn x)n1 is unbounded. Proof. Compute a strictly increasing sequence (nk )k1 of positive integers such that Tnk > k3k for each k, and then apply Theorem 6.2.11. 2

It is possible to remove from these two uniform boundedness theorems the condition that each Tn be normed; see Exercise 12. Theorem 6.2.13. Let C be an inhabited convex subset of a Hilbert space H such that sup {Re x, y : y ∈ C} exists for each x ∈ H. Then C is located. Proof. Fix ξ ∈ H and y0 ∈ C. In order to prove that ρ(ξ, C) exists, it is enough to show that if 0 < ε < 1, then there exists η ∈ C such that ξ − η < ξ − y + ε for all y ∈ C. If ξ − y0 < ε, then we may take η = y0 . Hence we may assume that ξ = y0 . Setting λ0 = 0, for the purposes of this proof we extend the λ-technique by constructing a sequence (λn )n0 with values in {−1, 0, 1}, and a sequence (yn )n0 in C, such that for each n 1, ξ − yn ξ − yn−1 and the following hold: 2

2

λn = −1 =⇒ ξ − yn < ξ − yn−1 − ε4 /16n, 2

2

λn = 0 =⇒ n/4 < ξ − yn + ξ − z for some z ∈ C, λn = 1 =⇒ ξ − yn < ξ − z + ε for all z ∈ C. Assume that we have already constructed λ0 , . . . , λn and y0 , . . . , yn . If λn = 1, set λn+1 = 1 and yn+1 = yn . If λn = 1, then we apply Lemma 6.2.10. If there 2 2 exists y ∈ C such that ξ − y < ξ − yn − ε4 /16n, we set λn+1 = −1 and 2 2 yn+1 = y; if n/4 < ξ − yn + ξ − z for some z ∈ C, we set λn+1 = 0 and yn+1 = yn ; if ξ − yn < ξ − z + ε for all z ∈ C, we set λn+1 = 1 and yn+1 = yn . This completes our inductive construction of the sequences (λn )n0 and (yn )n0 . It remains to compute n such that λn = 1. Let (Nk )k1 be a strictly increasing sequence of positive integers such that for each k,

166

6 Operators and Locatedness

2

Nk > 4 k + ξ − y0

2

and ξ − y0

k, 4 4

so wk −→ ∞ as k −→ ∞. We now apply Corollary 6.2.12 to the normed linear functionals x x, wk on H, to produce ξ 0 ∈ H such that the sequence {|ξ 0 , wk |}k1 is unbounded. By our hypotheses, there exists M > 0 such that |ξ 0 , ξ − z| < M for all z ∈ C. Choosing K such that M < |ξ 0 , wK | , suppose that λnK = 0. Then, by our construction of the vectors wk , there exists z ∈ C such that wK = ξ − z and therefore 2 |ξ 0 , wK | = |ξ 0 , ξ − z| < M, a contradiction. Hence λnK = 1.

We shall use some of these results on locatedness in the next section, when we look at adjoints of operators.

6.3 Adjoints Throughout the rest of the chapter, H will be a Hilbert space. By an operator on H we mean a linear mapping of H into itself. We observed in Section 3 of Chapter 5 that the constructive Riesz representation theorem, with its requirement that the linear functional be not just bounded but

6.3 Adjoints

167

normed, does not enable us to prove that every bounded operator on H has an adjoint. In fact, we showed that the proposition “every bounded operator on a Hilbert space has an adjoint” implies LPO. Are there general criteria for the existence of the adjoint? To answer this question aﬃrmatively, we ﬁrst prove a lemma. Lemma 6.3.1. Let T be an operator on H, and B a subset of H. Let a be a unit vector in H, and P the projection of H on the 1-dimensional subspace Ka. Then P T (B) is located if and only if the set {T x, a : x ∈ B} is located in K. Proof. For all x, y ∈ H, since (I − P ) y is orthogonal to P (H), we have 2

2

2

y − P T x = P y − P T x + (I − P ) y 2

= P y, a a − P T x, a a + (I − P ) y 2

2

2

= |y, a − T x, a| + (I − P ) y . Hence

2

inf y − P T x

x∈B

exists if and only if

2

inf |y, a − T x, a|

x∈B

exists, from which the desired conclusion follows.

2

Proposition 6.3.2. Let T be a jointed operator on H. Let B be the unit ball of H, and P the projection of H on a 1-dimensional subspace. Then P T (B) is located. Proof. Choose a unit vector a ∈ H such that P x = x, a a for all x ∈ H. By the preceding lemma, it suﬃces to prove that the set S = {T x, a : x ∈ B} = {x, T ∗ a : x ∈ B} is located in K. We do this by showing that S is dense in the (located) ball B K (0, T ∗ a). Given ε > 0, we have either T ∗ a < ε, in which case |ζ − 0| < ε for all ζ ∈ K with |ζ| T ∗ a, or else T ∗ a = 0. In the latter case, for each ζ ∈ K with ζ T ∗ a we have . / ζ ∗ ∗ ζ= 2 T a, T a T ∗ a and

ζ 2 T ∗ a

T ∗ a ∈ B.

Thus in either case there exists x ∈ B such that |ζ − x, T ∗ a| < ε. Since ε > 0 is 2 arbitrary, it follows that S is dense in B K (0, T ∗ a) .

168

6 Operators and Locatedness

Proposition 6.3.3. Let H be a Hilbert space with unit ball B, and T a bounded operator on H such that P T (B) is located for each 1-dimensional projection P. Then T has an adjoint. Proof. By Lemma 6.3.1, for each unit vector y ∈ H, the set C = {T x, y : x ∈ B} is located in K. Since T is bounded, C is a subset of a ball in K. It follows from Corollary 4.1.7 and Proposition 2.2.10 that C is totally bounded. Thus, by Propositions 2.2.6 and 2.2.5, the linear functional u deﬁned on H by u(x) = T x, y

(x ∈ H)

is normed. By the Riesz representation theorem, there exists a unique element T ∗ y 2 of H such that T x, y = x, T ∗ y for all x ∈ H. Theorem 6.3.4. Let T be a jointed operator on a Hilbert space H. Then the image under T of the unit ball is located. Proof. By Proposition 6.3.2 and Lemma 6.3.1, for each x ∈ H the set {T y, x : y ∈ B} is located in K. This set is also bounded, since |T y, x| = |y, T ∗ x| T ∗ x for all y ∈ B. Hence, as in the previous proof, it is totally bounded. Since the map ζ Re ζ is uniformly continuous on K, we now see that sup {Re x, T y : y ∈ B} = sup {Re T y, x : y ∈ B} exists for each x ∈ H. It follows from Theorem 6.2.13 applied to C = T (B) that T (B) is located in H. 2 Theorem 6.3.5. Let T be a bounded operator on a Hilbert space H that maps the unit ball to a located set. Then T has an adjoint. Proof. Given y ∈ H, we take C = T (B) in Corollary 6.2.9 to show that the linear functional x T y, x is normed. The result now follows from the Riesz representation theorem, as in the proof of Proposition 6.3.3. 2 Thus for a bounded operator on a Hilbert space, the existence of the adjoint is equivalent to the image of the unit ball being located. When is a jointed operator bounded? The classical answer is “always”; the constructive answer is less decisive.

6.3 Adjoints

169

Theorem 6.3.6. (Hellinger–Toeplitz theorem) Every jointed operator on a Hilbert space is sequentially continuous. Proof. Let T be a jointed operator on a Hilbert space H. It will suﬃce to prove that T ∗ is sequentially continuous, since we can then interchange T and T ∗ to obtain the desired result. Accordingly, let (xn )n1 be a sequence converging to 0 in H, and let ε > 0. By Ishihara’s second trick (Lemma 3.2.2), either T ∗ xn < ε eventually or else there exists a strictly increasing sequence (nk )k1 of positive integers such that T ∗ xnk > ε/2 for all k. In the latter case, passing to a subsequence if necessary, we may assume that ε xnk < 2 k 2k 3 for each k. Then, setting 2k3k xnk , yk = ε we have yk < 1/k and T ∗ yk > k3k . We now apply Royden’s uniform boundedness theorem (Theorem 6.2.11) to the normed linear functionals x x, T ∗ yk on H, to construct a vector x ∈ H such that |x, T ∗ yk | > k for each k. Since x, T ∗ yk = T x, yk T x yk −→ 0 as k −→ ∞, this is absurd. We conclude that T ∗ xn < ε for all suﬃciently large n.

2

A subset S of N is pseudobounded if limn→∞ n−1 sn = 0 for each sequence (sn )n1 in S. The following principle is trivially true in classical mathematics, holds in both INT and RUSS, and appears not to be provable in BISH (see [59]).

BD-N

Every inhabited, countable, pseudobounded set of positive integers is bounded.

We now prove that if “sequentially continuous” can be replaced by “bounded” in the conclusion of Theorem 6.3.6, then BD-N holds. Let A = {a1 , a2 , . . .} be a countable pseudobounded set of positive integers. Let H be an inﬁnitedimensional Hilbert space with an orthonormal basis (en )n1 of unit vectors. We ﬁrst prove that for each x ∈ H, Tx =

∞

an x, en en

n=1

is well deﬁned. Let (Nk )k1 be a strictly increasing sequence of positive integers such that

170

6 Operators and Locatedness ∞

2

|x, en | < 2−k−1 k −2 ,

n=Nk

and construct a binary sequence (λk )k1 such that Nk+1 −1

λk = 0 =⇒

2

a2n |x, en | < 2−k ,

n=Nk Nk+1 −1

λk = 1 =⇒

2

a2n |x, en | > 2−k−1 .

n=Nk

Deﬁne a sequence (bk )k1 in A as follows. If λk = 0, set bk = a1 . If λk = 1, then −k−1 −2

2

k

max

a2n

Nk+1 −1

: Nk n < Nk+1

2

a2n |x, en | > 2−k−1 ,

n=Nk

so an > k for some n with Nk n < Nk+1 ; in this case we set bk = an for this n. Since A is pseudobounded, there exists m such that bk /k < 1 for all k m. If λk = 1 for some k m, then 1 < bk /k < 1, a contradiction; hence λk = 0 for all k m, and therefore the series deﬁning T x converges. It is easily seen that T is a one-one selfadjoint linear mapping of H onto itself; so, by Theorem 6.3.6, it is sequentially continuous. But if T is bounded, then the pseudobounded set A is bounded (by any positive bound for T ).

6.4 Functions of Selfadjoint Operators We say that a sequence (Sn )n1 of operators on our Hilbert space H converges strongly to an operator S if Sx = lim Sn x n→∞

for each x ∈ H. Given an operator T on H, for each power series p(t) =

∞

cn tn

(6.15)

n=0

with complex coeﬃcients cn we can form the corresponding power series in T, p(T ) =

∞

cn T n ,

n=0

which makes sense provided the series on the right converges strongly. In particular, if p(t) is a polynomial, regarded as a power series whose coeﬃcients are eventually 0, then p(T ) is always deﬁned.

6.4 Functions of Selfadjoint Operators

171

If each coeﬃcient cn in (6.15) is real and T is selfadjoint, then p(T ) is also selfadjoint. Our ﬁrst aim is to prove that if, in that case, T has bound 1, then p(T ) is bounded by the sup norm of p on the interval [0, 1]. This will require us to prove some lemmas about approximate eigenvalues. Note that two operators S, T on H are said to commute if ST = T S. Also, T 0 = I and T n = T T n−1 for any n 1. Lemma 6.4.1. Let T be a selfadjoint operator on H with bound 1, and let x be a unit vector in H. Let −1 1 1 x+ v . v = T x − T x, x x, u = x + v 2 2 Then

1 2 v . 4 Moreover, if S is any operator that commutes with T, then Su 2 Sx . T u, u > T x, x +

⊥

⊥

Proof. Clearly, v ∈ {x} and T x − v ⊥ {x} , so (by Theorem 4.3.1) v is the ⊥ projection of T x onto {x} . Hence 2 x + 1 v = 1 + 1 v2 > 0, 2 4 and our deﬁnition of u makes sense. Also, 2

T x, v = T x, v − T x, x x, v = v

and therefore, since T = T ∗ , 1 0 1 1 1 T x + v , x + v = T x, x + Re T x, v + T v, v 2 2 4 1 2 T x, x + v − T v v 4 3 2 T x, x + v . 4 Hence −2 0 1 1 1 1 T x + v , x + v − T x, x T u, u − T x, x = x + v 2 2 2 −1 3 1 2 2 T x, x + v − T x, x 1 + v 4 4 −1 3 1 1 2 2 2 v − v T x, x = 1 + v 4 4 4 −1 1 1 2 2 v 1 + v 4 2 1 2 v , 4

172

6 Operators and Locatedness

since v 2. Finally, let S commute with T. Then 1 Su S x + v 2 1 1 = Sx − T x, x Sx + T Sx 2 2 1 1 Sx 2 Sx . 1 + |T x, x| + 2 2 2

We show how to construct approximate eigenvectors common to ﬁnitely many commuting selfadjoint operators. We deal ﬁrst with the case of a single selfadjoint operator. Lemma 6.4.2. Let T be a selfadjoint operator on H with bound 1, and let x1 be a unit vector in H. Let ε be a positive number, and N an integer greater than 32/ε2 . Deﬁne sequences (xn )n1 , (yn )n1 in H recursively by yn = T xn − T xn , xn xn ,

xn+1

−1 1 1 xn + y n . = xn + yn 2 2

Then T xn+1 , xn+1 T xn , xn for each n, and there exists n N such that yn < ε. Proof. By Lemma 6.4.1, T xn+1 , xn+1 − T xn , xn

1 2 yn . 4

On the other hand, either yn < ε for some n N, or else yn > ε/2 for all n N. In the latter case, Lemma 6.4.1 shows that T xn+1 , xn+1 − T xn , xn

ε2 16

for each n N, and therefore that T xN +1 , xN +1 T x1 , x1 +

N ε2 > −1 + 2 = 1, 16

which is absurd since xN +1 is a unit vector and T has bound 1.

2

6.4 Functions of Selfadjoint Operators

173

Lemma 6.4.3. Let T1 , . . . , Tn be commuting selfadjoint operators on H with common bound 1, and let x be a unit vector in H. Then for each ε > 0 there exists a unit vector u such that T1 u, u T1 x, x − ε and Tk u − Tk u, u u < ε

(1 k n) .

Proof. Noting that the case n = 1 has been disposed of in Lemma 6.4.2, let n > 1 and suppose that we have proved the desired result for n − 1 commuting selfadjoint operators. Fix ε > 0, and choose a positive integer N > 32/ε2 . By our induction hypothesis, there exists a unit vector x1 such that T1 x1 , x1 T1 x, x −

ε 2

and

ε (1 k n − 1) . 2N Taking T = Tn , deﬁne sequences (xk )k1 and (yk )k1 as in Lemma 6.4.2. By that lemma, there exists j (1 j N ) such that Tn xj − Tn xj , xj xj < ε. Setting u = xj , we see that if j = 1, then we are ﬁnished; so we may assume that j > 1. Then for 1 k n − 1, since the selfadjoint operator Tk − Tk x1 , x1 I commutes with Tn , we see from the ﬁnal part of Lemma 6.4.1 that Tk x1 − Tk x1 , x1 x1

T1 x, x − − 2 2 = T1 x, x − ε.

174

6 Operators and Locatedness

2

Our induction is now complete.

Proposition 6.4.4. Let T be a selfadjoint operator on H with bound 1. Then for each polynomial p with real coeﬃcients, the operator p(T ) has bound M = sup {|p(t)| : 0 t 1} . Proof. Write p(t) =

N

cn tn ,

n=0

where each cn ∈ R. Let x be a unit vector in H, and let ε > 0. Since p(T ) has only real coeﬃcients, it is selfadjoint. By Lemma 6.4.3, there exists a unit vector u ∈ H such that p(T )u, u p(T )x, x − ε and −1 N n |cn | ε. T u − T u, u u < 1 + n=1

Taking t = T u, u, for n > 1 we compute n−1 T n u − tn u T T n−1 u − tn−1 u + |t| T u − tu n−1 n−1 T u−t u + T u − tu ··· n T u − tu . Hence p(T )u − p(t)u

N

|cn | T n u − tn u

n=1

N

n |cn | T u − T u, u u < ε

n=1

and therefore p(T )x, x p(T )u, u + ε |p(t)| + |(p(T )u − p(t)u) , u| + ε M + p(T )u − p(t)u < M + 2ε. Since ε > 0 is arbitrary, it follows that p(T )x, x M. Replacing p by p2 , we now have , - , 2 p(T )x = p(T )2 x, x = p2 (T )x, x 2 sup |p(t)| : 0 t 1 = M 2 and therefore p(T )x M.

2

6.4 Functions of Selfadjoint Operators

175

An operator T on H is said to be positive if T x, x 0 for all x ∈ H. In that case, T is selfadjoint, since for all x, y ∈ H we have both T x, x = x, T x and the polarisation identity 4 T x, y = T (x + y) , x + y − T (x − y) , x − y + i T (x + iy) , x + iy − i T (x − iy) , x − iy (which is actually valid for any operator T on H). We introduce a partial order on the set of selfadjoint operators by writing S T, or equivalently T S, if and only if T − S is a positive operator. For our ﬁrst result about positive operators we use a weak constructive substitute for the classical result that a series of positive real numbers is convergent if its partial sums form a bounded sequence. Lemma 6.4.5. Let

∞ n=1

an be a series of nonnegative terms whose partial sums form

a bounded sequence. Then for each ε > 0 and each positive integer n there exists k n such that ak < ε. Proof. Let b > 0 be an upper bound for the partial sums of the series. Given ε > 0, choose a positive integer N such that N ε/2 > b. If ak > ε/2 for n < k n + N, then n+N n+N Nε > b, ak ak > 2 k=1

k=n+1

a contradiction. Hence ak < ε for some k with n < k n + N.

2

Proposition 6.4.6. Let S and T be commuting positive operators on H, with S bounded. Then ST is positive. Proof. The proof is similar to the classical one on pages 415–417 of [4], so we give only an outline, leaving the details as an exercise. Since S is bounded, we may assume that 0 S I. Deﬁne a sequence (Sn )n1 of bounded selfadjoint operators on H such that Sn+1 = Sn − Sn2 (n 1) . S1 = S, By induction, 0 Sn I for each n. We now have 0

n

Sk2 =

k=1

n

(Sk − Sk+1 ) = S − Sn+1 S.

k=1

It follows that for each x ∈ H, n k=1

2

Sk x =

n , k=1

Sk2 x, x = (S − Sn+1 ) x, x Sx, x .

176

6 Operators and Locatedness

By Lemma 6.4.5, there exists a strictly increasing sequence (ni )i1 of positive integers such that Sni x < 2−i for each i; whence ni

Sk2 x = Sx − Sni +1 x −→ Sx as i −→ ∞.

k=1

Since T commutes with S, it commutes with every Sn . Hence / . ni Sk2 x , x . ST x, x = T Sx, x = T lim i→∞

k=1

Now, T is positive and hence selfadjoint; so, by Theorem 6.3.6, it is sequentially continuous. Hence ST x, x = lim

i→∞

= lim

i→∞

= lim

i→∞

ni , k=1 ni k=1 ni

T Sk2 x, x

Sk T Sk x, x T Sk x, Sk x 0.

k=1

Since x is arbitrary, we conclude that ST 0.

2

Our next objective is to construct the positive square root of a bounded √ positive operator. We ﬁrst examine an iteration scheme for the function t t on [0, 1] . A standard classical proof of the convergence of that scheme uses Dini’s highly nonconstructive theorem on the uniform convergence of monotone sequences of continuous functions [47] (pages 131–132). Fortunately, with relatively little extra eﬀort, we can avoid using Dini’s theorem altogether. Lemma 6.4.7. Deﬁne a sequence (un )n1 of uniformly continuous mappings from [0, 1] into R iteratively by u1 (t) = 0, Then (un (t))n1 converges to

un+1 (t) = un (t) + √

1 t − u2n (t) . 2

t uniformly on [0, 1].

Proof. For each t ∈ [0, 1] and each n, √ √ 1 t − u2n (t) t − un+1 (t) = t − un (t) − 2 √ 1 √ t − un (t) 1 − t + un (t) . = 2

(6.17)

6.4 Functions of Selfadjoint Operators

177

√

Using this, it is simple to prove by induction that un+1 un and un (t) t. Let 0 < ε < 1, and compute a positive integer N such that n ε < ε (n N ) . 1− √ 2 2 √ Consider any t ∈ [0, 1] . If t < ε2 , then for all n we have 0 un (t) t < ε and therefore √ (6.18) 0 t − un (t) < ε. If t > ε2 /2, then for all n 1, 1√ ε 1 √ t + un (t) t> √ 2 2 2 2 and therefore 1−

ε 1 √ t + un (t) < 1 − √ . 2 2 2

It follows from this and (6.17) that for all n 2N, √ √ ε 0 t − un (t) 1 − √ t − un−1 (t) 2 2 2 √ ε t − un−2 (t) 1− √ 2 2 ··· n−N √ ε t − uN (t) 1− √ 2 2 N ε 1− √ 2 2 < ε. Thus (6.18) holds for all n 2N.

2

Our next lemma will enable us to transform the iteration in Lemma 6.4.7 into one for the square root of a bounded positive operator. Lemma 6.4.8. Let T be a selfadjoint operator on H that satisﬁes 0 T I, deﬁne the sequence (un )n1 as in Lemma 6.4.7, and write Un = un (T ). Then for each n we have 0 Un 12 (I + T ) , Un2 T, and

1 2 T − Un+1 = T − Un2 I − (Un+1 + Un ) . (6.19) 2

178

6 Operators and Locatedness

Proof. We have

2 2 T − Un+1 = T − Un2 + Un2 − Un+1 = T − Un2 − (Un+1 − Un ) (Un+1 + Un ) 1 T − Un2 (Un+1 + Un ) = T − Un2 − 2

1 2 = T − Un I − (Un+1 + Un ) . 2 Suppose that for some n we have 0 Un 12 (I + T ) and Un2 T. (These inequalities certainly hold for n = 1.) Then 0 Un I, so Un and I − Un are positive and 2 hence selfadjoint, and (I − Un ) 0. Thus 1 1 1 Un+1 = (I + T ) − I + Un − Un2 2 2 2 1 1 1 2 = (I + T ) − (I − Un ) (I + T ) . 2 2 2 On the other hand, Un+1 − Un =

1 T − Un2 0 2

and therefore Un+1 Un 0. Thus 0 and so 2 Hence T −Un+1

1 1 (Un+1 + Un ) Un+1 (I + T ) I, 2 2

1 I − (Un+1 + Un ) 0. 2 0, by (6.19) and Proposition 6.4.6. This completes the induction. 2

Proposition 6.4.9. Let T be a bounded positive operator on H. Then there exists a unique positive operator U on H such that U 2 = T. Moreover, U is bounded, U commutes with every operator that commutes with T, and the range of U is dense in the closure of the range of T. Proof. We may assume that 0 T I. Deﬁne (un )n1 and (Un )n1 as in Lemmas 6.4.7 and 6.4.8. By Lemma 6.4.8, 0 Un I. Since Un is a polynomial in T, it commutes with every operator that commutes with T. By Lemma 6.4.7, for each ε > 0 there exists Nε such that |um (t) − un (t)| < ε for all t ∈ [0, 1] and all m, n Nε . It follows from Proposition 6.4.4 that Um x − Un x ε x for all x ∈ H and all m, n Nε . Hence (Un )n1 converges strongly to an operator U on H. Clearly, 0 U I and U commutes with every operator that commutes with T. Moreover, for each n 1, since un (t) is a strict polynomial (one without constant term) over R, we have ran(Un ) ⊂ ran(T ), from which it follows that

6.4 Functions of Selfadjoint Operators

ran (U ) ⊂ ran (T ).

179

(6.20)

With ε and Nε as before, we see that √ √ 2 un (t) − t = un (t) + t un (t) − t √ 2 un (t) − t = 2 lim |un (t) − um (t)| m→∞

2ε

for all t ∈ [0, 1] and all n Nε . It follows from Proposition 6.4.4 that 2 Un − T x 2ε x for all x ∈ H and all n Nε . Hence 2 U − T x U 2 − Un2 x + Un2 − T x = (U + Un ) (U − Un ) x + Un2 − T x 2 (U − Un ) x + Un2 − T x −→ 0 as n −→ ∞ and therefore U 2 = T. Moreover,

Tx = U

lim Un x ,

n→∞

which, taken with (6.20), shows that ran(U ) is dense in the closure of ran(T ) . Now suppose that we have another positive operator S such that S 2 = T. Then ST = SS 2 = S 2 S = T S, so S commutes with T , and therefore U commutes with S. Given x ∈ H, write y = U x − Sx. Then U y, y + Sy, y = (U + S) (U − S) x, y =

,

U 2 − S 2 x, y = 0.

But U y, y 0 and Sy, y 0, so we must have U y, y = 0 = Sy, y . Now apply the ﬁrst part of the proof to U, to obtain a positive operator A such that A2 = U. Then , 2 Ay = Ay, Ay = A2 y, y = U y, y = 0, so Ay = 0. Hence U y = A2 y = 0, and similarly, Sy = 0. It now follows that 2

U x − Sx = (U − S) (U − S) x, x = (U − S) y, x = 0. Hence S = U.

by

2

The operator U in Proposition 6.4.9 is called the square root of T and is denoted √ T or T 1/2 .

180

6 Operators and Locatedness

6.5 Locating the Kernel and the Range Let T be a jointed operator on the Hilbert space H. In this section we deal with questions like the following: When is ran (T ) located? Is the locatedness of ran (T ) linked to the locatedness of ker (T ) or that of ran (T ∗ )? When is ran (T ) closed? As we shall see, the answers depend on some interesting applications of the λtechnique. We begin with two elementary lemmas. Lemma 6.5.1. Let X and Y be orthogonal subspaces of a Hilbert space H such that X + Y is dense in H. Then both X and Y are located. Proof. Given z ∈ H, choose a sequence (xn )n1 in X and a sequence (yn )n1 in Y such that xn + yn −→ z as n −→ ∞. For m n we have 2

2

= (xm − xn ) + (ym − yn )

2

= (xm + ym ) − (xn + yn ) −→ 0 as n −→ ∞.

2

xm − xn + ym − yn

Hence (xn )n1 , (yn )n1 are Cauchy sequences in X, Y respectively, and so converge to respective limits x∞ , y∞ in H such that z = x∞ + y∞ , x∞ ⊥ y∞ , and y∞ ⊥ X. Given x ∈ X, we have 2

2

2

2

2

z − x = y∞ + (x∞ − x) = y∞ + x∞ − x y∞ , with equality when x = x∞ . It follows that ρ (z, X) exists and equals y∞ . Hence X, and likewise Y, is located. 2 Lemma 6.5.2. Let T be a jointed operator on H. Then ran(T )⊥ = ker(T ∗ ). Also, ran (T ) is located if and only if ran(T ) + ker (T ∗ ) is dense in H, in which case ker (T ∗ ) is located. Proof. First observe that y ⊥ ran(T ) ⇐⇒ ∀x ∈ H (x, T ∗ y = 0) ⇐⇒ T ∗ y = 0, so ran(T )⊥ = ker (T ∗ ). It follows that if ran(T ) is located, then so is ker (T ∗ ); moreover, if P is the projection of H on the closure of ran(T ), then since x = P x + (I − P ) x for each ∈ H, we see that ran(T ) + ker (T ∗ ) is dense in H. If, conversely, ran(T ) + ker (T ∗ ) is dense in H, then we can apply the preceding 2 lemma to show that ran(T ) and ker (T ∗ ) are both located.

6.5 Locating the Kernel and the Range

181

Lemma 6.5.3. If T is a jointed operator on H, then ran (T T ∗ ) is dense in ran (T ) . Proof. Let B denote the closed unit ball of H. By Theorem 6.3.4, T T ∗ (B) is located, so T T ∗ (nB) is located for each positive integer n. Given x ∈ H and ε > 0, we need only show that ρ (T x, T T ∗ (Bn )) < ε for some n. To this end, construct an increasing binary sequence (λn )n1 such that ε , 2 ∗ λn = 1 =⇒ ρ (T x, T T (nB)) < ε.

λn = 0 =⇒ ρ (T x, T T ∗ (nB)) >

We may assume that λ1 = 0. If λn = 0, then, applying the separation and the Riesz representation theorems, we construct a unit vector yn ∈ H such that ε T T ∗ z, yn + < T x, yn (z ∈ nB) . 2 Then ε 2 n T ∗ yn < T T ∗ (nyn ), yn + < T x, yn T x 2 2 and so T ∗ yn < n−1 T x . If λn = 1, we set yn = 0. Clearly, the sequence ∗ (T yn )n1 converges to 0. Choose a positive integer N such that x, T ∗ yN < ε/2. If λN = 0, then ε ε ε + T T ∗ (N yN ) , yN < T x, yN = x, T ∗ yN < , 2 2 2 a contradiction. Hence λN = 1. 2 For a ﬁrst application of Lemmas 6.5.2 and 6.5.3, we call an operator T on H sequentially open if for each sequence (xn )n1 such that (T xn )n1 converges to 0, there exists a sequence (yn )n1 in ker(T ) such that xn + yn −→ 0. Proposition 6.5.4. Let T be a sequentially open operator on H with an adjoint. Then ran (T ) and ker (T ∗ ) are located. Proof. By Lemma 6.5.3, there exists a sequence (xn )n1 in H such that T (T ∗ xn − x) −→ 0 as n −→ ∞. Since T is sequentially open, there exists a sequence (yn )n1 in ker (T ) such that T ∗ xn + yn −→ x as n −→ ∞. We now see that ran (T ∗ ) + ker (T ) is dense in H; whence, by Lemma 6.5.2, ran (T ) is located. 2 If T is any bounded operator with an adjoint, then T ∗ T is a bounded positive operator, so, by Proposition 6.4.9, the absolute value of T, √ |T | = T ∗ T , exists as a bounded positive operator.

182

6 Operators and Locatedness

Proposition 6.5.5. If T is a bounded operator on H with an adjoint, then ran (|T |) is dense in the closure of ran (T ∗ ) . √ Proof. We see from Proposition 6.4.9 that the range of T ∗ T is dense in the closure 2 of the range of T ∗ T. Reference to Lemma 6.5.3 completes the proof.

We have already seen, in Lemma 6.5.2, that, for a jointed operator T, the kernel of T ∗ is the orthogonal complement of the range of T, and that if ran (T ) is located, then so is ker (T ∗ ) . If, conversely, ker (T ∗ ) is located, is ran (T ) located also? Classically it is, since the closure of ran (T ) is the orthogonal complement of ker (T ∗ ) . However, if the latter holds constructively, then we can prove Markov’s principle: for if a is any real number such that ¬ (a = 0) , then the operator T : z az on C is selfadjoint and has kernel {0} ; but if ran (T ) is located, then ¬ (ρ (1, ran (T )) > 0), so ρ(1, ran(T )) = 0, ran (T ) contains nonzero elements, and therefore a = 0. We say that a sequence (xn )n1 in H converges weakly to x ∈ H, and we write w

xn −→ x as n −→ ∞ w

(or just xn −→ x), if limn→∞ xn , y = x, y for all y ∈ H. We call an operator T on H weak-sequentially open if for any sequence (xn )n1 such that T xn −→ 0, w there exists a sequence (yn )n1 in ker(T ) such that xn + yn −→ 0. Proposition 6.5.6. If T is a jointed operator on H such that ran(T ∗ ) is located, then T is weak-sequentially open. Proof. Let P be the projection of H on the closure of ran(T ∗ ). Let (xn )n1 be a sequence in H such that T xn −→ 0, and for each n set yn = P xn − xn . Then yn ∈ ran(T ∗ )⊥ = ker(T ). For each z ∈ H we have xn + yn , T ∗ z = xn , T ∗ z − (I − P )xn , P T ∗ z = T xn , z , so

| xn + yn , T ∗ z | T xn z −→ 0 as n −→ ∞.

It follows that for each x ∈ H we have xn + yn , P x −→ 0 and therefore xn + yn , x = P (xn + yn ) , x = xn + yn , P x −→ 0. Hence T is weak-sequentially open.

2

Theorem 6.5.7. Let T be a weak-sequentially open, jointed operator on H with located kernel. Then T ∗ has located range.

6.5 Locating the Kernel and the Range

183

Proof. Let P be the projection of H on ker(T ). It suﬃces to show that for each x ∈ H, the vector x − P x is in the closure of ran(T ∗ ): for then x − y2 = (x − P x) + y2 + P x2 for all y ∈ ran(T ∗ ), so ρ (x, ran(T ∗ )) exists and equals P x. Accordingly, ﬁx x in H and ε > 0. Denote the closed unit ball in H by B. By Theorem 6.3.4, T ∗ (nB) is located in H for each positive integer n; so we can construct an increasing binary sequence (λn )n1 such that ε , 2 λn = 1 =⇒ ρ (x − P x, T ∗ (nB)) < ε.

λn = 0 =⇒ ρ (x − P x, T ∗ (nB)) >

Without loss of generality we may assume that λ1 = 0. If λn = 0, then by Corollary 5.2.10 and the Riesz representation theorem, there exists a unit vector yn such that for each z ∈ nB, x − P x, yn > |T ∗ z, yn | + Taking z = n T yn

−1

ε ε = |z, T yn | + . 2 2

T yn , we obtain ε + n T yn < x − P x, yn x 2

and therefore T yn < n−1 x . On the other hand, if λn = 1 − λn−1 , we set yk = 0 for all k n. Clearly, the sequence (T yn )n1 converges to 0. But T is weak-sequentially open, so there exists a sequence (zn )n1 in ker(T ) such that w

yn + zn −→ 0. Choose N such that |x − P x, yn | = |x − P x, yn + zn |

ε/2, which is absurd. Hence λN = 1. Since ε > 0 is arbitrary, we are through. 2 Corollary 6.5.8. If T is a jointed operator on H, then the following four statements are equivalent: (a) ran (T ) is located. (b) ran(T ∗ ) is located. (c) ker(T ) is located and T is weak-sequentially open. (d) ker (T ∗ ) is located and T ∗ is weak-sequentially open. Proof. This follows from Lemma 6.5.2, Theorem 6.5.7, and Proposition 6.5.6.

2

184

6 Operators and Locatedness

Turning in a slightly diﬀerent direction, we move towards a proof of the closed range theorem: Theorem 6.5.9. Let T be a jointed operator on H whose range is closed. Then ran (T ) and ker (T ) are located, and ran(T ∗ ) is closed. To set the scene, consider how the closed range theorem is proved classically. One simple proof uses the polar decomposition of the operator T ∗ ; but the existence of an exact polar decomposition in constructive analysis requires the range of the operator to be located from the outset (Exercise 16). In another classical proof of Theorem 6.5.9 the idea is to show that ran(T ∗ T ) is closed and then to apply Lemma 6.5.3. To that end, let (xn )n1 be a sequence such that (T ∗ T xn )n1 converges to a limit y ∈ H. Applying the classical uniform boundedness theorem (see page 163) to the bounded linear functionals fn deﬁned on the Hilbert space ran(T ) by fn (T x) = T x, T xn = x, T ∗ T xn , we obtain M > 0 such that fn M for each n. Hence the linear functional T x x, y on ran(T ) is bounded by M. By the Riesz representation theorem, there exists x∞ ∈ H such that x, y = T x, T x∞ = x, T ∗ T x∞ for all x. It follows that y = T ∗ T x∞ ∈ ran(T ∗ T ). This proof fails constructively in two places: ﬁrst, in its use of the classical version of the uniform boundedness theorem, and second, in its application of the Riesz representation theorem, which requires the linear functional to be not just bounded but normed. Fortunately, as the following sequence of results will show, these diﬃculties can be overcome. We begin with two lemmas that prepare us for a general result about sequentially continuous linear mappings between normed spaces. Lemma 6.5.10. Let T : X −→ Y be a sequentially continuous linear mapping between normed spaces, (xn )n1 a Cauchy sequence in X, and 0 < α < β. Then either T xn < β for all n or else there exists n such that T xn > α. Proof. In view of the linearity of T, we may assume that β − α > 1. Choosing a strictly increasing sequence (Nk )k1 of positive integers such that xm − xn < 2−3k for all m, n Nk , write sk = max {T xn : 1 n Nk } . Construct an increasing binary sequence (λk )k1 such that λk = 0 =⇒ ∀j k (sj < β − 2−2j ), λk = 1 =⇒ ∃j k (sj > β − 2−2j+1 ).

6.5 Locating the Kernel and the Range

185

We may assume that λ1 = λ2 = 0. Now construct a sequence (zk )k1 in X as follows. If λk+1 = 0 or λk = 1, set zk = 0. If λk+1 = 1 and λk = 0, then T xNk sk < β − 2−2k and sk+1 > β − 2−2k−1 , so we can choose j such that Nk < j Nk+1 and T xj > β − 2−2k−1 ; setting zk = 22k (xj − xNk ), we have zk < 2−k . Moreover, T zk = 22k T xj − T xNk 22k (T xj − T xNk )

1 > 22k β − 2−2k−1 − (β − 2−2k ) = . 2 This completes the construction of a sequence (zk )k1 converging to 0 in X. By the sequential continuity of T, limk→∞ T zk = 0. Choose K such that T zk < 1/2 for all k K. If λK = 1, then there exists n NK such that T xn > β − 2−2n+1 > α. On the other hand, if λK = 0 and there exists k K such that λk+1 = 1 − λk , then T zk > 1/2, a contradiction. Thus if λK = 0, then λk = 0 for all k K and 2 therefore for all k, so T xk < β for all k. Lemma 6.5.11. Let T : X −→ Y be a sequentially continuous linear mapping between normed spaces, and (xn )n1 a Cauchy sequence in X. Then supn1 T xn exists. Proof. In view of the previous lemma, it is enough to show that the sequence (T xn )n1 is bounded. To do so, choose R > 0 such that xn R for all n. Taking α = 1 and β = 2 in Lemma 6.5.10, we may assume that there exists n1 such that T xn1 > 1. Set λ1 = 0. Using Lemma 6.5.10 repeatedly, we construct an increasing binary sequence (λn )n1 , and an increasing sequence (nk )k1 of positive integers, such that λk = 0 =⇒ T xnk > k and nk > nk−1 , λk = 1 =⇒ (T xn )n1 is a bounded sequence and nk+1 = nk . Assume that we have constructed λk and nk . If λk = 1, we set λk+1 = λk and nk+1 = nk . If λk = 0, then T xnj > j for all j k. We then apply Lemma 6.5.10 to the Cauchy sequence (xj )j>nk . Either we obtain nk+1 > nk such that T xn > k + 1, or else T xj < k + 2 for all j > nk . In the ﬁrst case we set k+1 λk+1 = 0, and in the second, noting that (T xn )n1 is bounded, we set λk+1 = 1

186

6 Operators and Locatedness

and nk+1 = nk . This completes the inductive construction of the sequences (λk )k1 and (nk )k1 . If λk = 0, set zk = k −1 xnk ; if λk = 1, set zk = 0. Then zk R/k for each k, so zk −→ 0 and therefore, by the sequential continuity of T, T zk −→ 0. Choose K such that T zk < 1 for all k K. If λK = 0, then T zk =

1 T xnk > 1, k

a contradiction. Hence λK = 1 and so the sequence (T xn )n1 is bounded.

2

Proposition 6.5.12. A sequentially continuous linear mapping T : X −→ Y between normed spaces maps Cauchy sequences to Cauchy sequences. Proof. Given a Cauchy sequence (xn )n1 in X, choose a strictly increasing sequence (Nk )k1 of positive integers such that xm − xn < 2−k for all m, n Nk . For each k, the sequence (xn − xNk )nNk is a Cauchy sequence; so, by Lemma 6.5.11, sk = sup T xn − T xNk nNk

exists. Given ε > 0, we construct an increasing binary sequence (λk )k1 such that ε , 4 ε λk = 1 =⇒ sk < . 2

λk = 0 =⇒ sk >

We may assume that λ1 = 0. If λk = 0, choose j Nk such that T xj − T xNk > ε/4, and set zk = xj − xNk . If λk = 1, set zk = 0. Then zk < 2−k for each k, so zk −→ 0. Since T is sequentially continuous, T zk −→ 0 and we can choose K such that T zk < ε/4 for all k K. If λK = 0, then T zK > ε/4, which is absurd; so λK = 1 and therefore sK < ε/2. It follows that for all j, k NK , T xj − T xk T xj − T xNk + T xk − T xNk

α for some n or else T xn < β for all n.

6.5 Locating the Kernel and the Range

187

Proof. Let (nk )k1 be a strictly increasing sequence of positive integers such that T ∗ T xj

α. We may assume that λ1 = 0. Deﬁne a sequence (yk )k1 in ran(T ) as follows: if λk = 0, set yk = 0; if λk = 1 − λk−1 , choose i with nk−1 < i nk and T xi > α, and set 1 T xi yj = kT xi for all j k. Then (yk )k1 is a Cauchy sequence: in fact, yj − yk 1/(k + 1) whenever j k. Since ran (T ) is closed in H and therefore complete, there exists z ∈ H such that (yk )k1 converges to T z. Choosing a positive integer N such that z < N α, consider any integer k N . If λk = 1 − λk−1 , then Tz =

1 T xi kT xi

for some i with nk−1 < i nk and T xi > α, so 1 1 1 1 1 α < T xi = T xi , T z = T ∗ T xi , z 2 z < 2 N α α, k k k k k a contradiction. Hence λk = λk−1 for all k N . It follows that either λk = 0 for all k, or else λk = 1 − λk−1 for some k N. In the ﬁrst case, T xn < β for all n; 2 in the second, T xi > α for some i with nk−1 < i nk .

Lemma 6.5.14. Under the hypotheses of Lemma 6.5.13, for all positive numbers α, β with α < β, either T xn > α for inﬁnitely many n or else T xn < β for all suﬃciently large n. Proof. Let (nk )k1 be a strictly increasing sequence of positive integers such that T ∗ T xj

α and set yk =

1 T xi ; kT xi

if λk = 1 − λk−1 , set yj = yk−1 for all j k. Then yj − yk 2/k for all j k; so (yk )k1 is a Cauchy sequence in the complete space ran (T ) and therefore converges to T z for some z ∈ H. Choosing a positive integer N such that z < N α, consider any integer k > N . If λk = 1 − λk−1 , then Tz =

1 T xi (k − 1)T xi

for some i nk−1 with T xi > α, so 1 1 α< T xi = T xi , T z = T ∗ T xi , z k−1 k−1 1 1 1 2 z < 2 N α k − 1 α, (k − 1) (k − 1) a contradiction. Hence λk = λk−1 for all k > N . It follows that either λk = 1 for some k N or else λk = 0 for all k. In the ﬁrst case, T xi < β for all i nk . In 2 the second case, for each k there exists i nk such that T xi > α.

Lemma 6.5.15. Under the hypotheses of Lemma 6.5.13, (T xn )n1 converges to 0. Proof. By Lemma 6.5.14, for each ε > 0 either T xn > ε/2 for inﬁnitely many n or else T xn < ε for all suﬃciently large n. In the former case, passing to an appropriate subsequence, we may assume that T xn > ε/2 and T ∗ T xn < 1/n2 for all n. Applying the uniform boundedness theorem (Corollary 6.2.12) to the normed linear functionals T x T x, nT xn on the Hilbert space ran(T ), we can ﬁnd T z ∈ ran(T ) such that the sequence (|T z, nT xn |)n1 is unbounded. But |T z, nT xn | = |z, nT ∗ T xn |

1 z −→ 0 as n −→ ∞. n

This contradiction rules out the possibility that T xn > ε/2 for inﬁnitely many n. Since ε > 0 is arbitrary, the result follows. 2

Lemma 6.5.16. Let H be a Hilbert space, and T a jointed operator on H with closed range. Then ran(T ∗ T ) is complete.

6.6 Baire’s Theorem, with Applications

189

Proof. First observe that, by Theorem 6.3.6, both T and T ∗ are sequentially continuous. It is easily seen using Lemma 6.5.2 that T ∗ is one-one on ran (T ). So we can deﬁne a linear mapping S : ran (T ∗ T ) −→ ran (T ) by setting ST ∗ T x = T x for each x ∈ H. Lemma 6.5.15 shows that S is sequentially continuous. Consider any Cauchy sequence (T ∗ T xn )n1 in ran(T ∗ T ) . By Proposition 6.5.12, (T xn )n1 is a Cauchy sequence in the Hilbert space ran (T ). Hence there exists x ∈ H such that T xn −→ T x. The sequential continuity of T ∗ now 2 yields T ∗ T xn −→ T ∗ T x. Thus ran(T ∗ T ) is complete. We can now give the proof of the closed range theorem. Proof. By Lemmas 6.5.3 and 6.5.16, ran (T ∗ T ) is both dense in ran(T ∗ ) and complete. Hence ran(T ∗ ) is complete and therefore closed in H. Moreover, for each x ∈ H there exists y ∈ H such that T ∗ x = T ∗ T y; so x = T y + (x − T y) , where T y ∈ ran (T ) and (by Lemma 6.5.2) x − T y ∈ ker (T ∗ ) = ran(T )⊥ . It follows from Lemma 6.5.2 that both ran (T ) and ker (T ∗ ) are located. Applying 2 Corollary 6.5.8, we now see that ran(T ∗ ) and ker (T ) are located.

6.6 Baire’s Theorem, with Applications Baire’s theorem states that The intersection of a sequence of dense open subsets of a complete metric space is dense in that space. The standard classical proof of this theorem (see [79], page 97) passes over unchanged to the constructive setting. However, various classically equivalent versions of Baire’s theorem do not pass over unscathed; for example, the version that states that if a complete metric space is the union of a sequence of closed subsets, then one of those subsets is inhabited. In this section we present a constructive proof of a restricted form of this last version of Baire’s theorem in the context of a Banach space, and apply it to operator theory. The proof of our version of Baire’s theorem introduces yet another technique, in which we show that a certain property P holds by constructing an element of the set {x ∈ X : P }, which is empty if ¬P holds and inhabited if P holds.

190

6 Operators and Locatedness

Theorem 6.6.1. Let X be a Banach space, and C a closed, convex, balanced, lo-◦ nC and the distance ρ (0, −C) exists. Then C cated subset of X such that X = n1

is inhabited. Proof. For each positive integer n deﬁne the open set Un = −nC ∪ {x ∈ X : C ◦ is inhabited} , where −nC is the metric complement of nC in X. To prove that Un is dense in X, consider y ∈ X and ε > 0. Note that ρ(y, nC) exists and equals nρ n−1 y, C . Either ρ (y, nC) > 0 or ρ (y, nC) < ε. In the ﬁrst case, y ∈ −nC. In the second, choose z ∈ nC such that y − z < ε. Noting that ρ (0, −2nC) = 2nρ (0, −C) exists, we see that either ρ (0, −2nC) < 2ε or ρ (0, −2nC) > ε. In the former case, choose z ∈ −2nC such that z < 2ε. For each w ∈ nC we have −w ∈ nC (since C is balanced), so z − w ∈ nC + nC = 2nC, by the convexity of C. Hence (z − z ) − w = z − (z − w) ρ (z , 2nC) > 0. Thus z − z ∈ −nC. Since also y − (z − z ) y − z + z < 3ε, we see that ρ (y, Un ) < 3ε. Finally, in the case ρ (0, −2nC) > ε, for each x with x ε we have ρ (x, 2nC) = 0. Hence B (0, ε) ⊂ 2nC = 2nC, ◦

so (2nC) , and therefore C ◦ , is inhabited. In this case, Un = X. This completes the proof that Un is dense in X. Since X is complete, it follows from the standard version of Baire’s theorem Un is dense in X and therefore, in particular, contains a point ξ. Choose that n1

n such that ξ ∈ nC. Since also ξ ∈ Un , we must have ξ ∈ {x ∈ X : C ◦ is inhabited} . Hence C ◦ is indeed inhabited.

2

Recall the classical open mapping theorem for bounded linear mappings: a bounded linear mapping T of a Banach space X onto a Banach space Y is open, in

6.6 Baire’s Theorem, with Applications

191

the sense that it maps open subsets of X onto open subsets of Y (or, equivalently, it maps the open unit ball of X onto an open subset of Y ). This theorem does not hold constructively without some additional hypotheses: the inverse S of the mapping T in the example on page 169 is a bounded operator of a Hilbert space onto a Hilbert space, but if S is open, then T is bounded. In other words, if every bounded linear mapping of a Hilbert space onto a Hilbert space is open, then BD-N holds. Nonetheless, we can prove constructive versions of the open mapping theorem that are classically equivalent to the standard version. To that end, we begin with a lemma. Lemma 6.6.2. Let T be a sequentially continuous linear mapping of a Banach space X into a normed space Y such that T (BX (0, 1)) is located. Let r be a positive number, and y an element of BY (0, r). There exists x ∈ B X (0, 2) such that if y = T x, then ρ (z, T (BX (0, 1))) > 0 for some z ∈ BY (0, r). Proof. If ρ (y, T (BX (0, 1))) > 0, then we can take x to be any element of BX (0, 2). So we may assume that ρ (y, T (BX (0, 1))) < r/2. Choosing x1 ∈ BX (0, 2) such

that y − T 12 x1 < r/2 and therefore 2y − T x1 < r, we set λ1 = 0. This is the ﬁrst step in the inductive construction of an increasing binary sequence (λn )n0 and a sequence (xn )n0 of elements of BX (0, 2) such that for each n 1, if λn = 0, then ρ

2n−1 y −

n−1

2n−1−i T xi

, T (BX (0, 1))

0

i=1

and xi = 0 for all i n. Suppose that we have found λn−1 and xn−1 with the applicable properties. If λn−1 = 1, we set λn = 1 and xn = 0. If λn−1 = 0, we consider the two cases n−1 n−1 n−1−i ρ 2 y− 2 T xi , T (BX (0, 1)) > 0 i=1

and

192

6 Operators and Locatedness

ρ

n−1

2

y−

n−1

n−1−i

2

T xi

, T (BX (0, 1))

2−N r y −

and therefore

N N 2N −i T xi > r. 2 y −

i=1

i=1

We must therefore have λN = 1; so there exists n N such that λn = 1 − λn−1 . Setting n−1 2n−1−i T xi , z = 2n−1 y − i=1

we see that ρ (z, T (BX (0, 1))) > 0 (as λn = 1) and z < r (as λn−1 = 0). This completes the proof. 2 Lemma 6.6.3. Let C be a balanced convex subset of a normed space Y, and let y ∈ Y and r > 0 be such that B(y, r) ⊂ C. Then B(0, r) ⊂ C. Proof. If z ∈ Y and z < r, then y±z ∈ C and therefore z = C.

1 2

(y + z)− 12 (y − z) ∈ 2

Here is our version of the open mapping theorem. Theorem 6.6.4. Let X, Y be Banach spaces, and T a sequentially continuous linear mapping of X onto Y such that T (B(0, 1)) is located and ρ (0, −T (B (0, 1))) exists. Then T is an open mapping.

6.6 Baire’s Theorem, with Applications

Proof. Since Y =

193

n T (B (0, 1)) ,

n1

we can apply Theorem 6.6.1 to compute y0 ∈ Y, R > 0, and a positive integer N such that BY (y0 , R) ⊂ N T (B (0, 1)) . Writing y1 = N −1 y0 and r = R/N, we obtain BY (y1 , r) ⊂ T (B (0, 1)). It follows from Lemma 6.6.3 that BY (0, r) ⊂ T (B (0, 1)).

(6.21)

Now consider any y ∈ Y with y < r. Choose x ∈ B (0, 2) as in the conclusion of Lemma 6.6.2. If y = T x, then there exists z ∈ BY (0, r) such that ρ (z, T (B (0, 1))) > 0, which contradicts (6.21); hence (the inequality on a normed space being tight) y = T x. Thus BY (0, r) ⊂ T (B (0, 2)) , and therefore T is an open mapping. 2

A subset C of a linear space X is said to be a generating set for, or to generate, X if every element of X is a ﬁnite linear combination of elements of C. Our next lemma will enable us to prove that compactly generated Banach spaces are ﬁnitedimensional. Lemma 6.6.5. Let G be a compact generating set for a nontrivial Banach space X. Then there exists a compact generating set C for X that is both convex and balanced such that ρ (0, −C) exists. Proof. We may assume that G is both convex and balanced (the proof is left as an exercise). Now, 2G is compact, the mapping x ρ (x, G) is uniformly continuous on X, and X is nontrivial. Hence there exists δ > 0 such that both the sets C = {x ∈ 2G : ρ (x, G) δ} , D = {x ∈ 2G : ρ (x, G) δ} are compact. Note that C is convex and balanced, and, since it contains G, generates X. We show that −C is dense in D. To this end, consider any x ∈ D and any ε > 0. Choose t > 1 such that (t − 1) x < ε/2, and suppose that tx ∈ C. Then, since C is balanced, x ∈ C; whence x ∈ C ∩ D and therefore ρ (x, G) = δ. But then for each g ∈ G we have tx − g = t x − t−1 g tρ (x, G) = tδ; so ρ(tx, G) tδ > δ, which is absurd since tx ∈ C. We conclude that tx ∈ / C. It follows from Proposition 3.1.2 that there exists y ∈ ∼C such that tx − y < ε/2

194

6 Operators and Locatedness

and therefore x − y < ε. Applying Bishop’s lemma (Proposition 3.1.1), we see that y ∈ −C. This completes the proof that −C is dense in D. Since the norm function is uniformly continuous on the compact set D, it now follows that ρ (0, −C) = inf {x : x ∈ −C} = inf {x : x ∈ D} 2

exists.

A normed space with a compact generating set is said to be compactly generated. Theorem 6.6.6. A compactly generated Banach space is ﬁnite-dimensional. Proof. Let X be a compactly generated Banach space. We ﬁrst suppose that X contains a nonzero vector. By Lemma 6.6.5, X has a balanced, convex, compact nC, we can apply Thegenerating set C such that ρ (0, −C) exists. Since X = n1

orem 6.6.1 to show that C ◦ is inhabited; whence C contains a nontrivial ball. But every ball in a normed space is located, so the ball in question is totally bounded. It follows from Proposition 4.1.13 that X is ﬁnite-dimensional. It remains to remove the restriction that X be nontrivial. To do this, we work in the product Banach space X × K. This space is generated by the compact set G × {1}; so, by the foregoing, X × K is ﬁnite-dimensional. It follows that X, being isomorphic to the quotient space (X × K)/ K, is ﬁnite-dimensional. 2 Recall from page 102 that a linear mapping T : X −→ Y between normed spaces is said to be compact if T (BX (0, 1)) is a totally bounded subset of Y. Corollary 6.6.7. If T is a compact linear mapping of a normed space X onto a Banach space Y, then Y is ﬁnite-dimensional. Proof. The totally bounded set T (B(0, 1)) generates Y. Since Y is complete, it follows that T (B(0, 1)) is a compact generating set for Y ; whence, by Theorem 6.6.6, Y is ﬁnite-dimensional. 2

In the traditional development of functional analysis, the open mapping theorem is used to prove Banach’s inverse mapping theorem, the closed graph theorem, and the uniform boundedness theorem. The last of these three we have already discussed, in Section 2. To deal with the inverse mapping theorem we need another lemma about convex sets, and two further technical lemmas. Lemma 6.6.8. Let C be a convex, absorbing subset of a Banach space X. Then 0∈ / −C.

6.6 Baire’s Theorem, with Applications

195

Proof. Assuming that 0 ∈ −C, we ﬁrst show that X − nC is dense in X for each positive integer n. To do so, construct a sequence (xn )n1 in −C that converges to 0. Fixing a positive integer n, an element y of X, and ε > 0, compute positive numbers r, δ and a positive integer k such that −y ∈ rC, xk < ε/ (n + r), and xk − z δ for all z ∈ C. Let y1 = y + (n + r) xk . Then y − y1 = (n + r) xk < ε. On the other hand, since C is convex, for each z ∈ nC we have 1 1 (z − y) ∈ (nC + rC) = C, n+r n+r 1 (n + r) δ. (z − y) y1 − z = (n + r) − x k n+r

so

Hence y1 ∈ −nC. Since ε > 0 is arbitrary, it follows that −nC is dense in X. Moreover, being a metric complement, −nC is open. Applying the standard form −nC is inhabited, which is absurd since of Baire’s theorem, we now see that n1 nC. We conclude that 0 ∈ / −C. 2 X= n1

Sometimes when we want to prove that a certain proposition P is absurd, we ﬁrst prove that P implies LPO, and then (frequently by adapting a classical proof) show that the addition of LPO to our intuitionistic logic suﬃces for us to prove that P is false. The next lemma will enable us to rule out in this way an unwanted alternative in the proof of Banach’s inverse mapping theorem. Lemma 6.6.9. Let f be a strongly extensional mapping of a complete metric space X into a metric space Y. Suppose that there exist α > 0 and a sequence (xn )n1 converging to x in X such that ρ (f (x), f (xn )) > α for each n. Then LPO holds. Proof. Given an increasing binary sequence (λn )n1 with λ1 = 0, construct a sequence (zn )n1 in X such that •

if λk = 0 for all k n, then zn = x, and

•

if λn = 1 − λn−1 , then zk = xn for all k n.

Then (zn )n1 is a Cauchy sequence in X and so converges to a limit z ∈ X. Either f (x) = f (z) or else ρ (f (x), f (z)) < α. In the ﬁrst case, the strong extensionality of f shows that x = z; whence there exists n such that x = zn and therefore λn = 1. In the second case, if there exists n with λn = 1 − λn−1 , then we obtain the con2 tradiction ρ (f (x), f (z)) = ρ (f (x), f (xn )) > α; hence λn = 0 for all n.

196

6 Operators and Locatedness

Lemma 6.6.10. If LPO holds, then every separable subset of a metric space is located. Proof. Assuming LPO, consider a separable subset S of a metric space X. Let (sn )n1 be a dense sequence in S, let x ∈ X, and let 0 < α < β. Construct a binary sequence (λn )n1 such that

λn = 0 =⇒ ∀k n ρ(x, sk ) < 12 (α + β) , λn = 1 =⇒ ∃k n (ρ(x, sk ) > α) . By LPO, either λn = 0 for all n, in which case, since (sn )n1 is dense in S, ρ(x, s) 12 (α + β) < β for all s ∈ S; or else there exists n with λn = 1 and therefore ρ(x, sk ) > α for some k n. It follows from the constructive least-upperbound principle that ρ (x, S) exists. 2

This brings us to Ishihara’s version of Banach’s inverse mapping theorem. Theorem 6.6.11. Let T be a one-one, sequentially continuous linear mapping of a separable Banach space X onto a Banach space Y. Then T −1 is sequentially continuous. Proof. Let (xn )n1 be a sequence in X such that T xn −→ 0, and let ε > 0. By Corollary 3.1.7, the inverse linear mapping T −1 : Y −→ X is strongly extensional; whence, by Lemma 3.2.2, either xn < ε for all suﬃciently large n, or else xn > ε/2 for inﬁnitely many n. It suﬃces to rule out the latter case. To do so, we may assume that xn > ε/2 for all n. By Lemma 6.6.9, LPO holds; so, by Lemma 6.6.10, every separable subset of Y is located. Let (an )n1 be a dense sequence in

B X (0, 1). Since T is sequentially continuous, (T an )n1 is dense in T B X (0, 1) , which is therefore located. Writing xn = 2ε−1 xn , we see that xn > 1 and that T xn −→ 0 as n −→ ∞. Since T −1 is strongly extensional,

T xn ∈ ∼T B X (0, 1) . We now apply Lemma 6.6.2 with y = T (2xn ), to produce zn ∈ X such that

T zn < T xn + n−1 and ρ T (2zn ) , T B X (0, 1) > 0. Then Hence

ρ T zn , T B X (0, 12 ) > 0 and T zn −→ 0.

0 ∈ −T BX 0, 12 .

This contradicts Lemma 6.6.8, since T BX 0, 12 is both convex and absorbing. 2

6.6 Baire’s Theorem, with Applications

197

Recall that the graph of a mapping T : X −→ Y is the set G (T ) = {(x, T x) : x ∈ X} . As classically, Banach’s inverse mapping theorem leads to a version of the closed graph theorem: Corollary 6.6.12. Let T be a linear mapping of a Banach space X into a Banach space Y such that G (T ) is closed and separable. Then T is sequentially continuous. Proof. The mapping p : (x, T x) x of the Banach space G (T ) onto X is one-one and bounded linear. It follows from Theorem 6.6.11 that the inverse linear map is sequentially continuous, and hence that T is sequentially continuous. 2

In the theory of unbounded operators, the graph plays a signiﬁcant role. Particularly important properties for such a graph are closedness and locatedness. See [26, 83, 92] for more on such matters.

Exercises 1. Complete the proof of Proposition 6.1.1. 2. Complete the proof of Proposition 6.1.4. 3. Prove that a bounded linear mapping T of a normed space X into Cn is compact if and only if f ◦ T is normed for each linear functional f on Cn . Prove that if also the sum of any two normed linear functionals on X is normed, then T is compact if and only if pk ◦ T is normed for each k, where pk denotes the mapping (z1 , . . . , zn ) zk on Cn . 4. Let S be a compact operator on a Hilbert space H, and A a bounded operator on H. Prove that (a) λS is compact for each λ ∈ C; (b) S ∗ exists and is compact; (c) AS is compact. Prove also that if A∗ exists, then SA is compact. 5. A subset S of a normed space X is said the be weakly totally bounded if it is totally bounded relative to the locally convex structure deﬁned on X by the seminorms x |f (x)| , with f a normed linear functional on X. Prove the equivalence of the following conditions on X.

198

6 Operators and Locatedness

(a) The sum of any two normed linear functionals on X is normed. (b) The unit ball of X is weakly totally bounded. (Hint: To prove that (a) implies (b), note the second part of Exercise 3.) 6. An operator on a Hilbert space H is said to be weakly compact if it maps the unit ball of H to a weakly totally bounded subset of H. Prove that an operator T on H has an adjoint if and only if it is weakly compact. 7. Let (en )n1 be an orthonormal basis of a separable Hilbert space H, and T an operator on H. Prove that the following conditions are equivalent: (a) T has an adjoint. (b)

∞ n=1

(c)

∞ n=1

2

|T en , ek | converges for each k. 2

|T en , y| converges for each y ∈ H.

8. Let T be a weak-sequentially open operator on a Hilbert space such that ker (T ) is located. Prove that T is well-behaved (see Exercise 3 of Chapter 3). 9. Prove that every bounded linear mapping of a normed space onto a ﬁnitedimensional Banach space is an open mapping. 10. Show that the statement “every normed linear mapping T of a Hilbert space onto a ﬁnite-dimensional Banach space is compact” implies an omniscience principle. 11. Show that the statement “every one-one compact linear mapping T of a Hilbert space into a ﬁnite-dimensional Banach space has ﬁnite-dimensional range” implies an omniscience principle. 12. Show that the existence of the norms of the operators can be removed from the hypotheses of Royden’s version of the uniform boundedness theorem (Theorem 6.2.11). 13. Complete the details of the proof of Proposition 6.4.6. 14. A partial isometry is a jointed operator U on a Hilbert space H for which there exists a projection P, called the initial projection of U, such that U x = x ⊥ for all x ∈ ran (P ), and U x = 0 for all x ∈ ran (P ) . Prove that U ∗ U = P, that ∗ ∗ U U is a projection, and that ran (U U ) = ran (U ) . 15. Let T be a bounded operator such that T ∗ exists and has located range, and let P be the projection of H on ran (T ∗ ). Prove that there exists a partial isometry

6.6 Baire’s Theorem, with Applications

199

U on H whose initial projection is P such that T = U |T | and |T | = U ∗ T. Prove also that U U ∗ is the projection of H on ran (T ). (The expression of T as U |T | is called the polar decomposition of T, and is analogous to the modulus–argument form of a complex number.) 16. Prove that if a bounded jointed operator T on a Hilbert space H has a polar decomposition, then ran (T ) is located. 17. Prove the converse of Proposition 6.5.12: if a linear mapping T between normed spaces maps Cauchy sequences to Cauchy sequences, then T is sequentially continuous. 18. Let E be a dense linear subspace of a normed space X, and T a sequentially continuous linear mapping of E into a Banach space Y. Show that T extends to a sequentially continuous linear mapping of X into Y. 19. Prove that if G is a compact generating set for a Banach space X, then there exists a compact generating set G for X that is balanced and convex. 20. Let G be a balanced convex generating set for a normed space X. Prove that for each x ∈ X and each ε > 0 there exist t > 0 and g ∈ G such that x − tg ε. Can we replace ε by 0 in this result?

Notes Many of the results in this chapter come from papers by Ishihara and the authors. The work of Sections 1 and 2 is largely drawn from [61] and [63]. For related classical material see [86]. It is not known whether Theorems 6.2.1 and 6.2.2 can be extended to possibly unbounded convex subsets of a normed space. More general versions of Exercises 3–6 appeared in Ishihara’s thesis [56]. Since we cannot guarantee that a bounded operator on a Hilbert space has an adjoint, when we discuss such matters as the Gelfand representation theorem for an operator algebra A, we need to postulate that A is selfadjoint in the sense that each of its elements has an adjoint that also belongs to A; see [9, 12]. The principle BD-N, introduced by Ishihara in [59], has the unusual feature of being provable classically, intuitionistically, recursively, but not, apparently, within BISH. It is an interesting problem—a part of constructive reverse mathematics [7, 62]—to identify classical theorems that are equivalent to BD-N. The usual classical proof of Lemma 6.4.7 is based on the nonconstructive monotone convergence theorem for sequences; see [47] (7.3.1.1). The constructive proof is much more informative, in that it provides the rate of convergence of the sequence of functions. A full constructive analysis of Dini’s theorem is given in [8]; see also [21].

200

6 Operators and Locatedness

The existence of the square root of a selfadjoint operator is a special case of a more general result, the spectral theorem for sequences of commuting selfadjoint operators, which enables us to construct more general functions of an operator. Since that theorem requires measure theory, which we do not touch in this book, we refer the reader to the relevant chapters of [9] and [12]. It is easy to prove Lemma 6.5.3 classically, taking orthogonal complements in the identity ker(T ∗ T ) = ker(T ∗ ). For more on polar decompositions see [35]. Proposition 6.5.12 is trivial in CLASS, since in that context sequential continuity for linear maps implies uniform continuity. With classical logic, Theorem 6.6.1 is a simple consequence of the standard form of Baire’s theorem, the hypotheses of locatedness and the existence of ρ(0, −C) being redundant. For other, constructively inequivalent, versions of Baire’s theorem, see Chapter 2 of [34]. For more on open mapping theorems see [29] and [25]. It is interesting that sequential continuity, rather than boundedness, is the best we can get in the constructive versions of Banach’s inverse mapping theorem and the closed graph theorem. An extension of the former is given in [64].

What can be said at all can be said clearly; and whereof one cannot speak thereof one must be silent. —Ludwig Wittgenstein, Tractatus Logico-Philosophicus

References

1. O. Aberth, Computable Analysis, McGraw-Hill, New York, 1980. 2. O. Aberth, Computable Calculus, Academic Press, San Diego, CA, 2001. 3. P. Aczel and M. Rathjen, Notes on Constructive Set Theory, Report No. 40, Institut Mittag-Leﬄer, Royal Swedish Academy of Sciences, 2001. 4. G. Bachman and L. Narici, Functional Analysis, Academic Press, New York, 1966. 5. A. Bauer, Realizability as the connection between computable and constructive mathematics, Proceedings of CCA 2005, Kyoto, Japan, 25–29 August 2005; to appear. 6. M.J. Beeson, Foundations of Constructive Mathematics, Ergeb. Math. und ihrer Grenzgebiete, 3. Folge, 6, Springer-Verlag, Heidelberg, 1985. 7. J. Berger and D.S. Bridges, A bizarre property equivalent to the Π10 -fan theorem, Logic J. of the IGPL, to appear. 8. J. Berger and P.M. Schuster, Classifying Dinis theorem, to appear in Notre Dame J. Formal Logic. 9. E.A. Bishop, Foundations of Constructive Analysis, McGraw-Hill, New York, 1967. 10. E.A. Bishop, Mathematics as a numerical language, in: Intuitionism and Proof Theory (J. Myhill, A. Kino, and R.E. Vesley, eds.), 53–71, North-Holland, Amsterdam, 1970. 11. E.A. Bishop, Schizophrenia in contemporary mathematics, in: Errett Bishop: Reﬂections on Him and His Research (M. Rosenblatt, ed.), 1–32, Contemporary Math. 39, Amer. Math. Soc., Providence, RI, 1985. 12. E.A. Bishop and D.S. Bridges, Constructive Analysis, Grundlehren der math. Wissenschaften 279, Springer-Verlag, Heidelberg, 1985. 13. E.A. Bishop and H. Cheng, Constructive Measure Theory, Mem. Amer. Math. Soc. 116, 1972. 14. D.S. Bridges, On weak operator compactness of the unit ball of L(H), Zeit. math. Logik Grundlagen Math. 24, 493–494, 1978. 15. D.S. Bridges, Recent progress in constructive approximation theory, in The L.E.J. Brouwer Centenary Symposium (A.S. Troelstra and D. van Dalen, eds.), 41–50, North–Holland, Amsterdam, 1982. 16. D.S. Bridges, A constructive Morse theory of sets, in Mathematical Logic and Its Applications (D. Skordev, ed.), 61–79, Plenum Publishing Corp., New York, 1987. 17. D.S. Bridges, Locatedness, convexity, and Lebesgue measurability, Quart. J. Math. Oxford (2) 39, 411–421, 1988.

204

References

18. D.S. Bridges, A constructive look at the real number line, in: Real Numbers, Generalizations of the Reals, and Theories of Continua (P. Ehrlich, ed.), 29–92, Synthese Library 242, Kluwer Academic Publishers, Dordrecht, 1994. 19. D.S. Bridges, Constructive mathematics: a foundation for computable analysis, Theoretical Computer Science 219(1–2), 95–109, 1999. 20. D.S. Bridges, Constructive methods in mathematical economics, in Mathematical Utility Theory, J. Econ. (Zeitschrift f¨ ur National¨ okonomie), Suppl. 8, 1–21, 1999. 21. D.S. Bridges, Dinis Theorem: a constructive case study, in: Combinatorics, Computability and Logic (Proceedings of DMTCS01, Constant¸a, Romania, 2–6 July 2001; C.S. Calude, M.J. Dinneen, S. Sburlan, eds.), 69–80, DMTCS Series 17, SpringerVerlag, London, 2001. 22. D.S. Bridges, A weak constructive sequential compactness property and the fan theorem, Logic J. of the IGPL 13(2), 151–158, 2005. 23. D.S. Bridges and N.F. Dudley Ward, Constructing ultraweakly continuous linear functionals on B(H) , Proc. American Math. Soc. 126(11), 3347–3353, 1998. 24. D.S. Bridges and H. Ishihara, Linear mappings are fairly well-behaved, Arch. Math. 54, 558–562, 1990. 25. D.S. Bridges and H. Ishihara, A deﬁnitive constructive open mapping theorem?, Math. Logic Quarterly 44, 545–552, 1998. 26. D.S. Bridges and H. Ishihara, Constructive unbounded operators, in: Reuniting the Antipodes—Constructive and Nonstandard Views of the Continuum (Proceedings of the Symposion in San Servolo/Venice, Italy, May 17–22, 1999; U. Berger, H. Osswald, and P.M. Schuster, eds.), 25–37, Synthese Library, Kluwer Academic Publishers, Dordrecht, December 2001. 27. D.S. Bridges, H. Ishihara, and P.M. Schuster, Sequential Compactness in Construc¨ tive Analysis, Osterr. Akad. Wiss. Math.–Natur. Kl. Sitzungsber. II. 208, 159–163, 1999. 28. D.S. Bridges, H. Ishihara, and L.S. Vˆıt¸˘ a, Computing inﬁma on convex sets, with applications in Hilbert spaces, Proc. Amer. Math. Soc. 132(9)2723–2732, 2004. 29. D.S. Bridges, W.H. Julian, and R. Mines, A constructive treatment of open and unopen mapping theorems, Zeit. math. Logik Grundlagen Math. 35, 29–43, 1989. 30. D.S. Bridges and A. Mahalanobis, Bounded variation implies regulated: a constructive proof, J. Symb. Logic 66(4), 1695–1700, 2001. 31. D.S. Bridges and R. Mines, Sequentially continuous linear mappings in constructive analysis, J. Symb. Logic 63(2), 579–583, 1998. 32. D.S. Bridges and G. Popa, Exact, continuous boundary crossings out of convex sets in RN , Quart. J. Math. 54, 391–398, 2003. 33. D.S. Bridges and S. Reeves, Constructive mathematics, in theory and programming practice, Philosophia Mathematica 7(1), 65–104, 1999. 34. D.S. Bridges and F. Richman, Varieties of Constructive Mathematics, London Math. Soc, Lecture Notes 97, Cambridge Univ. Press, 1987. 35. D.S. Bridges, F. Richman, and P.M. Schuster, Adjoints, absolute values, and polar decompositions, J. Operator Theory 44, 243–254, 2000. 36. D.S. Bridges, F. Richman, and Y. Wang, Sets, complements and boundaries, Proc. Koninklijke Nederlandse Akad. Wetenschappen (Indag. Math. N.S.) 7(4), 425–445, 1996. 37. D.S. Bridges and L.S. Vˆıt¸a ˘, Weak continuity properties in constructive analysis, Logic J. of the IGPL 7(3), 277–281, 1999. 38. D.S. Bridges and L.S. Vˆıt¸˘ a, Apartness spaces as a framework for constructive topology, Ann. Pure Appl. Logic 119 (1–3), 61–83, 2003.

References

205

39. D.S. Bridges and L.S. Vˆıt¸˘ a, The constructive uniqueness of the locally convex topology on Rn , in: From Sets and Types to Topology and Analysis (L. Crosilla and P.M. Schuster, eds.), 304–315, Oxford Logic Guides 32, Clarendon Press, Oxford, 2005. 40. D.S. Bridges and L.S. Vˆıt¸a ˘, Apartness Spaces, book, in preparation. 41. L.E.J. Brouwer, Over de Grondslagen der Wiskunde, doctoral thesis, University of Amsterdam, 1907. Reprinted with additional material (D. van Dalen, ed.) by Matematisch Centrum, Amsterdam, 1981. 42. R.L. Constable et al., Implementing Mathematics with the Nuprl Proof Development System, Prentice-Hall, Englewood Cliﬀs, New Jersey, 1986. 43. L. Crosilla and P.M. Schuster (eds.), From Sets and Types to Topology and Analysis, Oxford Logic Guides 32, Clarendon Press, Oxford, 2005. 44. D. van Dalen (ed.), Brouwers Cambridge Lectures on Intuitionism, Cambridge University Press, Cambridge, 1981. 45. D. van Dalen, Mystic, Geometer, and Intuitionist, Clarendon Press, Oxford, 1999 (Vol. 1) and 2005 (Vol. 2). 46. R. Diaconescu, Axiom of choice and complementation, Proc. Amer. Math. Soc. 51, 176–178, 1975. 47. J. Dieudonn´e, Foundations of Modern Analysis, Academic Press, New York, 1960. 48. M.A.E. Dummett, Elements of Intuitionism (2nd edition), Oxford Logic Guides 39, Clarendon Press, Oxford, 2000. 49. H.M. Edwards, Essays in Constructive Mathematics, Springer Science + Business Media, Inc., New York, 2005. 50. N. D. Goodman and J. Myhill, Choice implies excluded middle, Zeit. math. Logik und Grundlagen Math. 24, 461. 51. S. Hayashi and H. Nakano, PX: A Computational Logic, MIT Press, Cambridge MA, 1988. 52. A. Heyting, Die formalen Regeln der intuitionistischen Logik, Sitzungsber. preuss. Akad. Wiss. Berlin, 42–56, 1930. ¨ 53. D. Hilbert, Uber das Unendliche, Mathematische Annalen 95 , 161–190, 1926; translated in Philosophy of Mathematics (P. Benacerraf and H. Putnam, eds.), 183–201, Cambridge University Press, Cambridge, 1964. 54. D. Hilbert, Die Grundlagen der Mathematik, Abhandlungen Math. Sem. Univ. Hamburg V, 65–85, 1927. 55. H. Ishihara,On the constructive Hahn–Banach theorem, Bull. London Math. Soc. 21, 79–81, 1989. 56. H. Ishihara, Boundedness, normability and compactness of constructive linear mappings, Ph.D. dissertation, Tokyo Institute of Technology, Tokyo, 1990. 57. H. Ishihara, Continuity and nondiscontinuity in constructive analysis, J. Symb. Logic 56(4), 1349–1354, 1991. 58. H. Ishihara, Constructive compact operators on a Hilbert space, Ann. Pure Appl. Logic 52, 31–37, 1991. 59. H. Ishihara, Continuity properties in constructive analysis, J. Symb. Logic 57, 557– 565, 1992. 60. H. Ishihara, A constructive version of Banachs inverse mapping theorem, New Zealand J. Math 23, 71–75, 1994. 61. H. Ishihara, Locating subsets of a Hilbert space, Proc. Amer. Math. Soc. 129(5), 1385–1390, 2001. 62. H. Ishihara, Constructive reverse mathematics: compactness properties, in: From Sets and Types to Topology and Analysis (L. Crosilla and P.M. Schuster, eds.), 245–267, Oxford Logic Guides 32, Clarendon Press, Oxford, 2005.

206

References

63. H. Ishihara and L.S. Vˆıt¸˘ a, Locating subsets of a normed space, Proc. Amer. Math. Soc. 131(10), 3231–3239, 2003. 64. H. Ishihara and L.S. Vˆıt¸˘ a, A constructive Banach inverse mapping theorem in Fspaces, New Zealand J. Math., to appear. 65. R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Algebras (Vol. 1), Academic Press, New York, 1988. 66. B.A. Kushner, Lectures on Constructive Mathematical Analysis, Amer. Math. Soc., Providence RI, 1985. 67. A.A. Markov, Theory of Algorithms (Russian), Trudy Mat. Istituta imeni V.A. Steklova 42 (Izdatelstvo Akademi Nauk SSSR, Moskva), 1954; English translation by J.J. Schoor-Kan and PST staﬀ, Israel Program for Scientiﬁc Translations, Jerusalem, 1961. 68. P. Martin-L¨ of, Notes on Constructive Mathematics, Almqvist and Wiksell, Stockholm, 1970. 69. P. Martin-L¨ of, An Intuitionistic Theory of Types: Predicative Part, in: Logic Colloquium 1973 (H.E. Rose and J.C. Shepherdson, eds.), 73–118, North-Holland, Amsterdam, 1975. 70. P. Martin-L¨ of, Constructive mathematics and computer programming, in Proc. 6th. Int. Congress for Logic, Methodology and Philosophy of Science (L. Jonathan Cohen, ed.), 153–179, North-Holland, Amsterdam, 1980. 71. G. Metakides, A. Nerode, and R. Shore, Recursive limits on the Hahn–Banach theorem, in: Errett Bishop: Reﬂections on Him and His Research (M. Rosenblatt, ed.), 85–91, Contemporary Math. 39, Amer. Math. Soc., Providence, R.I., 1985. 72. R. Mines, F. Richman, and W. Ruitenburg, A Course in Constructive Algebra, Universitext, Springer-Verlag, Heidelberg, 1988. 73. J.R. Moschovakis, The eﬀect of Markovs principle on the intuitionistic continuum, preprint, UCLA, April 2005. 74. J. Myhill, Constructive set theory, J. Symb. Logic 40, 347–382, 1975. 75. F. Richman (ed.), Constructive Mathematics (Proceedings of the Conference at Las Cruces, New Mexico, August 1980), Lecture Notes in Mathematics 873, SpringerVerlag, Heidelberg, 1981. 76. F. Richman, The fundamental theorem of algebra: a constructive development without choice, Paciﬁc J. Math. 196, 213–230, 2000. 77. F. Richman, Adjoints and the image of the unit ball, Proc. Amer. Math. Soc. 129, 1189–1193, 2001. 78. H.L. Royden, Aspects of constructive analysis, in: Errett Bishop: Reﬂections on Him and His Research (M. Rosenblatt, ed.), 57–64, Contemporary Mathematics 39, American Math. Soc., 1985. 79. W. Rudin, Real and Complex Analysis, McGraw-Hill, New York, 1970. 80. E. Schechter, Handbook of Analysis and Its Foundations, Academic Press, San Diego, 1997. 81. P.M. Schuster, What is continuity, constructively?, J.UCS 11(12), 2076–2085, 2005. 82. P.M Schuster, L.S. Vˆıt¸˘ a, and D.S. Bridges, Apartness as a relation between subsets, in: Combinatorics, Computability and Logic (Proceedings of DMTCS01, Constant¸a, Romania, 2–6 July 2001; C.S. Calude, M.J. Dinneen, S. Sburlan, eds.), 203–214, DMTCS Series 17, Springer-Verlag, London, 2001. 83. B. Spitters, Located operators, Math. Logic Quarterly 48(Suppl. 1), 107–122, 2002. 84. W.P. van Stigt, Brouwers Intuitionism, North–Holland, Amsterdam, 1990. 85. G. Stolzenberg, Review of [9], Bull. Amer. Math. Soc. 76, 301–323, 1970.

References

207

86. W. Takahashi, Nonlinear Functional Analysis, Yokohama Publishers, 2000. 87. A. Takayama, Mathematical Economics, The Dryden Press, Hinsdale, IL, 1973. 88. A.S. Troelstra and D. van Dalen, Constructivism in Mathematics: An Introduction (two volumes), North Holland, Amsterdam, 1988. 89. F. Waaldijk, On the foundations of constructive mathematics, Foundations of Science 10(3), 249–324, 2005. 90. H. Weber, Leopold Kronecker, Jahresber. der Deutschen Math. Verein 2, 5–31, 1893. 91. K. Weihrauch, Computable Analysis, Springer-Verlag, Heidelberg, 2000. 92. F. Ye, Towards a constructive theory of unbounded operators, J. Symb. Logic 65, 357–370, 2000. 93. R. Zach, Hilberts Program, in: The Stanford Encyclopedia of Philosophy (E.N. Zalta, ed.), URL http://plato.stanford.edu/archives/fall2003/ entries/hilbert-program/.

The above list contains only a fraction of the publications on constructive mathematics that have appeared in the last forty years, and does not include the sources of all results in our book. The reader should not fall into the trap of believing that an unascribed result was ﬁrst produced by the authors. We mention two websites that may interest the reader: http://www.math.canterbury.ac.nz/php/groups/cm/faq/ http://plato.stanford.edu/entries/mathematics-constructive/ In addition, many of the authors of items in the bibliography have websites that are worth a visit. The primary historical reference on constructive analysis is [9], the review of which [85] is interesting in its own right. Later references for Bishop-style constructivism are [12, 34], the latter of which gives comparisons between BISH, INT, and RUSS. Beeson [6] and Troelstra–van Dalen [88] contain a wealth of information about the logic, philosophy, and practice of constructive mathematics. For some applications of constructive mathematics, see [20, 33, 92]. The deﬁnitive reference for constructive algebra is [72], but [49] should be consulted for more recent work in the ﬁeld. The classic work on intuitionism is [48]. The life and works of Brouwer himself are discussed in [44, 45, 84]. Martin-L¨ ofs early work on constructive mathematics is found in [68], and his theory of types appears in [69]. Among the most recent varieties of computable analysis is that of Weihrauch [91]; the translation of BISH into Weihrauchs framework is described in [5].

Index

absolute value, 30 absolute value, operator, 181 absolutely convergent, 56 absorbing, 65 adjoint, 101 admissible, 43 algebraic diﬀerence, 121 algorithm, 6 apart, 44 approximation, 38, 129 approximation theory, fundamental theorem, 88 approximation, Chebyshev, 106 Archimedes, 57 arithmetic, Heyting, 9 arithmetic, Peano, 9 associated vector, 51 axiom of choice, 17 Baire’s theorem, 189 balanced, 121 ball, 37 Banach space, 49 Banach’s inverse mapping theorem, 196 Banach–Alaoglu theorem, 131 basic open neighbourhood, 127 basic separation theorem, 116 basic set, 18 basis, 81 best approximation, 88 best approximation, at most one, 88 BHK interpretation, 7 bijection, 16

BISH, 6 Bishop, 6 Bishop’s constructive mathematics, 6 Bishop’s lemma, 64 bound, 51 boundary, 107 boundary crossing map, 112 boundary crossing property, 143 bounded linear mapping, 51, 129 Brouwer, 5, 22 Brouwerian counterexample, 11 Brouwerian example, 11 canonical injection, 55 canonical mapping, 84 Cantor’s theorem, 31 Cartesian product, 13, 19 Cauchy, 38 Cauchy completeness, 35 Cauchy sequence, 35, 129 Cauchy–Schwarz inequality, 92 Church–Markov–Turing thesis, 6, 61 CLASS, 5 closed, 37 closed range theorem, 184 closed set, 127 closest point, 88 closure, 37, 127 cluster point, 37 commuting operators, 171 compact, 44 compact image, 60 compact linear mapping, 62, 102, 194

210

Index

compactly generated, 104, 194 comparison test, 56 complement, 14 complement, logical, 12 complete, 129 complete metric space, 38 complex number, 36 complex plane, 36 composite/composition, 16 cone, 113 cone, generated, 113 connectives, 7 constructive reverse mathematics, 199 continuous, 39, 61, 128 convergence, 35 convergent, 38, 56 converges, 129 converges strongly, 170 converges weakly, 182 converges, series, 95 convex, 65 convex combination, 87 convex hull, 87 convex mapping, 125 coordinate, 82, 96 coordinate functionals, 82 cotransitive, 28 countable, 17 countable choice, 18 countably inﬁnite, 17 countably many, for all but, 43 counterexample, Brouwerian, 11 de Morgan’s rule, 76 decidable, 8 Dedekind completeness, 32 denial inequality, 13 dense, 37, 127 dense sequence, 38 dependent choice, 18 detachable, 14 diameter, 40 diﬀerence, 33 dimension, 81 Dini’s theorem, 176 discrete, 13 distance, 40 distinct, 13 distinct real numbers, 26

division, 34 domain, 15 dual, 130 empty subset, 12 epimorphism, 16 equal functions, 16 equal real numbers, 26 equal sets, 13 equality, 12 equivalent norms, 85 Euclid, 1 Euclidean space, 49 eventually, 71 example, Brouwerian, 11 excluded middle, 5, 18 excluded sequence, 43 excluded third, 5 existence, constructive, 2 existence, idealistic, 2 extensional, 14 family, 15 fan theorem, 60, 106 ﬁnite, 17 ﬁnite possibility, 21 ﬁnite-dimensional, 81 ﬁnitely enumerable, 16 ﬁnitely many, 17 ﬂagging alternatives, 64 Fr´echet diﬀerentiable, 148 function, 14 Gˆ ateaux diﬀerentiable, 126 generating set, 193 Goldbach conjecture, 3 Gram–Schmidt process, 96 graph, 16 greater than, 27 greater than or equal to, 27 greatest lower bound, 33 Hahn–Banach theorem, 123 half-space, 116 Hausdorﬀ metric, 46 Heine–Borel–Lebesgue theorem, 61 Hellinger–Toeplitz theorem, 106, 169 Hermitian, 102 Heyting, 7, 22

Index Heyting arithmetic, 9 Hilbert, 5 Hilbert space, 93 hyperplane, 51 idempotent, 94 identity operator, 95 image of a subset, 15 image of an element, 14 imaginary part, 36 independent, 6 index set, 15 inequality relation, 13 inequality, compatible, 47 inequality, denial, 13 inequality, discrete, 13 inequality, metric space, 37 inequality, tight, 13 inﬁmum, 40 inﬁnite-dimensional, 98 inﬁnitely often, 71 inhabited, 12 initial projection, 198 injective, 16 INT, 5 integers, 25 interior, 37 intermediate value theorem, 2 internal choice, 21 intersect sharply, 77 intersection, 19 interval-halving, 3 intervals, in a vector space, 108 inverse function, 16 inverse image, 15 Ishihara’s tricks, 71 jointed, 101 kernel, 52 Kolmogorov, 22 Kronecker, 5 lambda technique, 64 least upper bound, 32 least-upper-bound principle, 32 LEM, 5 LEM-Cauchy, 77 less than, 27

less than or equal to, 27 limit, 35, 129 linear functional, 50 linear functional, bounded, 51 linear map, 50 linear space, 48 linear space, seminormed, 48 linearly independent, 81 LLPO, 9 locally compact, 46 locally convex space, 127 locally convex structure, 127 locally convex topology, 127 locally nonzero, 21 locally totally bounded, 46 located, 40, 130 logic, classical, 2 logic, intuitionistic, 2, 7 logical complement, 12 lower bound, 33 lower order located, 33 LPO, 8 map, 14 mapping, 14 Markov’s principle, 10 material implication, 22 mathematics, classical, 5 mathematics, intuitionistic, 5 maximum, 30 metric, 37 metric complement, 41 metric space, 37 metric, associated, 48 minimum, 30 modulus-argument decomposition, 64 monotone convergence, 35 MP, 10 mutual distance, 121 negative, 30 nonnegative real line, 46 nontrivial, 48 nonzero, 52 nonzero vector, 48 norm, 48 norm, Euclidean, 49 normable, 52 normed, 52

211

212

Index

normed space, 48 omniscience principles, 8 one-one, 16 onto, 16 open, 37, 127 open mapping theorem, 192 operator, 50, 100, 166 operator norm, 130 operator, bounded, 51 ordered, 20 ordered pair, 20 orthogonal complement, 95 orthogonal subsets, 95 orthogonal vector, 95 orthonormal basis, 96 orthonormal family, 96 parallelogram identity, 92 partial function, 14 partial function, total, 15 partial isometry, 198 partial sum, 56 Peano arithmetic, 9 pointwise, 15 polar decomposition, 199 polarisation identity, 175 positive operator, 175 power set, 19 product, 34 product metric, 37 product metric space, 37 product norm, 50 product normed space, 50 projection, 94 proximinal, 88 pseudobounded, 169 Pythagoras’s theorem, 95 quantiﬁers, 7 quasiproximinal, 88 quotient norm, 55 quotient space, 55 range, 15 rational numbers, 25 real line, 27 real number, 25 real part, 36

recursive constructive mathematics, 6 regular, 60 represented, 99 Riesz representation theorem, 99 Riesz’s lemma, 86 RUSS, 6 selfadjoint, 102 seminorm, 48 seminorm, deﬁning, 127 separable, 38, 129 separation theorem, 121 sequence, 15 sequentially continuous, 21, 39 sequentially discontinuous, 77 sequentially nondiscontinuous, 72 sequentially open, 181 series, 56 set, 12 set, basic, 18 smooth, 148 span, 81 spanned, 81 spectral theorem, 200 square root, 64 square root, operator, 179 square summable, 93 strong injection, 45 strongly continuous, 44 strongly extensional, 14 strongly injective, 45 strongly unique, 14 subset, 12 subspace, 50 suﬃciently large, 71 sum, 33 sup norm, 49 supremum, 32, 40 termwise, 15 tight, 13 topology, 37 totally bounded, 38, 129 trichotomy, 29 Type-2 Eﬀectivity, 22 ultraﬁlter principle, 146 ultraweakly continuous, 145 unequal, 13

Index unequal real numbers, 26 uniform boundedness theorem, 165 uniform Cauchyness, 49 uniform convergence, 49 uniformly almost located, 76 uniformly continuous, 40, 128 uniformly convex, 54, 126, 151 uniformly Fr´echet diﬀerentiable, 148 union, 19 unit ball, 48, 127 unit kernel, 132 unit vector, 48 upper bound, 32

upper order located, 32 value, 14 weak star topology, 130 weak-sequentially open, 182 weakly compact, 198 weakly discriminating, 77 weakly totally bounded, 197 Weihrauch, 6 well-behaved, 76 WLEM, 58 Zorn’s lemma, 116, 122

213

Universitext

(continued from p. ii)

Hurwitz/Kritikos: Lectures on Number Theory Jennings: Modern Geometry with Applications Jones/Morris/Pearson: Abstract Algebra and Famous Impossibilities Kac/Cheung: Quantum Calculus Kannan/Krueger: Advanced Analysis Kelly/Matthews: The Non-Euclidean Hyperbolic Plane Kostrikin: Introduction to Algebra Kuo: Introduction to Stochastic Integration Kurzweil/Stellmacher: The Theory of Finite Groups: An Introduction Lang: Introduction to Differentiable Manifolds Lorenz: Algebra: Volume I: Fields and Galois Theory Luecking/Rubel: Complex Analysis: A Functional Analysis Approach MacLane/Moerdijk: Sheaves in Geometry and Logic Marcus: Number Fields Martinez: An Introduction to Semiclassical and Microlocal Analysis Matsuki: Introduction to the Mori Program McCarthy: Introduction to Arithmetical Functions McCrimmon: A Taste of Jordan Algebras Meyer: Essential Mathematics for Applied Fields Mines/Richman/Ruitenburg: A Course in Constructive Algebra Moise: Introductory Problems Course in Analysis and Topology Morris: Introduction to Game Theory Poizat: A Course In Model Theory: An Introduction to Contemporary Mathematical Logic Polster: A Geometrical Picture Book Porter/Woods: Extensions and Absolutes of Hausdorff Spaces Procesi: Lie Groups Radjavi/Rosenthal: Simultaneous Triangularization Ramsay/Richtmyer: Introduction to Hyperbolic Geometry Rautenberg: A Concise Introduction to Mathematical Logic, 2nd ed. Reisel: Elementary Theory of Metric Spaces Ribenboim: Classical Theory of Algebraic Numbers Rickart: Natural Function Algebras Rotman: Galois Theory Rubel/Colliander: Entire and Meromorphic Functions Runde: A Taste of Topology Sagan: Space-Filling Curves Samelson: Notes on Lie Algebras Schiff: Normal Families Shapiro: Composition Operators and Classical Function Theory Simonnet: Measures and Probability Smith: Power Series From a Computational Point of View Smith/Kahanpää/Kekäläinen/Traves: An Invitation to Algebraic Geometry Smorynski: Self-Reference and Modal Logic Stillwell: Geometry of Surfaces Stroock: An Introduction to the Theory of Large Deviations Sunder: An Invitation to von Neumann Algebras Tondeur: Foliations on Riemannian Manifolds Toth: Finite Möbius Groups, Minimal Immersions of Spheres, and Moduli Van Brunt: The Calculus of Variations Weintraub: Galois Theory

Wong: Weyl Transforms Zhang: Matrix Theory: Basic Results and Techniques Zong: Sphere Packings Zong: Strange Phenomena in Convex and Discrete Geometry

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close