Dominated operators

Contents

Foreword to the English Translation

ix

Preface

xi

Chapter 1. Boolean Algebras and Vector Lattices § 1.1. Boolean Algebras

............................................

§ 1.2. Representation of Boolean Algebras § 1.3. Vector Lattices

10

..............................................

17

............................

25

......................................

33

...................................................

40

§ 1.5. Normed Vector Lattices

Chapter 2. Lattice-Normed Spaces § 2.1. Preliminaries § 2.2. Completion § 2.3. Examples

2

..........................

§ 1.4. Representation of Vector Lattices

§ 1.6. Comments

1

44

................................................

45

..................................................

54

....................................................

62

§ 2.4. Continuous Banach Bundles

.................................

70

§ 2.5. Measurable Banach Bundles

.................................

79

...................................................

84

§ 2.6. Comments

vi

Contents

Chapter 3. Positive Operators

89

§ 3.1. Operators in Vector Lattices

.................................

§ 3.2. Fragments of a Positive Operator

............................

§ 3.3. Orthomorphisms and Lattice Homomorphisms § 3.4. Maharam Operators

99

...............

108

.........................................

118

§ 3.5. Maharam’s Extension of Positive Operators § 3.6. Comments

90

..................

126

...................................................

133

Chapter 4. Dominated Operators

141

§ 4.1. The Space of Dominated Operators

..........................

§ 4.2. Decomposability of the Space of Dominated Operators § 4.3. Order Continuous Operators

142

.......

150

.................................

156

§ 4.4. The Yosida–Hewitt-Type Theorems

..........................

163

§ 4.5. Extension of Dominated Operators

...........................

171

...................................................

179

§ 4.6. Comments

Chapter 5. Disjointness Preserving Operators § 5.1. Band Preserving Operators § 5.2. n-Disjoint Operators

..................................

188

.........................................

195

§ 5.3. Weight-Shift-Weight Factorization

...........................

206

................................

214

.....................................

221

...................................................

230

§ 5.4. Multiplicative Representation § 5.5. Decomposable Operators § 5.6. Comments

187

Chapter 6. Integral Operators § 6.1. Vector Integration

236

...........................................

237

§ 6.2. Integral Representation by Quasi-Radon Measures

............

248

§ 6.3. Functional Representation of Maharam’s Extension

...........

257

...........................................

266

§ 6.4. Integral Operators

§ 6.5. Pseudointegral Operators § 6.6. Comments

....................................

278

...................................................

286

Contents

vii

Chapter 7. Operators in Spaces with Mixed Norm § 7.1. Spaces with Mixed Norm § 7.2. Summing Operators

....................................

292

.........................................

299

§ 7.3. Isometric Classification

......................................

§ 7.4. Kaplansky–Hilbert Modules

316

...............................................

325

...................................................

334

Chapter 8. Applications of Boolean-Valued Analysis § 8.1. Real Numbers in Boolean-Valued Models

340

...................

352

..............................

362

..................................

371

§ 8.3. Boolean-Valued Banach Spaces § 8.4. Involutive Banach Algebras

338

.....................

§ 8.2. Boolean-Valued Analysis of Vector Lattices

§ 8.5. Cyclically Compact Operators § 8.6. Comments

307

..................................

§ 7.5. AW ∗ -Algebras § 7.4. Comments

291

...............................

379

...................................................

389

Appendix. Boolean-Valued Models

394

References

412

Symbol Index

434

Subject Index

437

Foreword to the English Translation

The topic of this book belongs to vector lattice theory. Dominated operators are remote relatives of the noble family of bounded linear operators and functionals residing in Banach space. The concept of a dominated operator was invented in the 1930s by Leonid Vital 0 evich Kantorovich, a renowned mathematician and a Nobel prize winner in economics. This concept utilizes the main sociological trick of functional analysis which rests on studying the population of some mathematical objects in order to reveal their individual features and acquired traits. The theory of dominated operators has ripened in the recent decades mainly due to the contribution by Professor A. G. Kusraev and his students in Vladikavkaz and Novosibirsk. A few years ago Professor A. G. Kusraev asked me to edit and introduce the English translation of his book. I undertook the task readily for two reasons: First, the topic of the book is attractive and close to my own area of research. Second, I feel proud of the achievements of Professor A. G. Kusraev, once a brilliant student of mine at Novosibirsk State University and now my inseparable fellow with whom I have been sharing many splendid days full of inspirational mathematics. Unfortunately, the Russian version of this book is still in preparation in view of the unstable situation of the Northern Caucasus which hindered us in working on the translation. I hope that the reader will take this regrettable circumstance into account and forgive the inadvertent shortcomings of the inadequate communication between the author and the editor. S. Kutateladze Novosibirsk Akademgorodok

Preface

The notion of a dominated or majorized operator rests on a simple idea that goes as far back as the Cauchy method of majorants. Loosely speaking, the idea can be expressed as follows. If an operator (equation) under study is dominated by another operator (equation), called a dominant or majorant, then the properties of the latter have a substantial influence on the properties of the former. Thus, operators or equations that have “nice” dominants must possess “nice” properties. In other words, an operator with a somehow qualified dominant must be qualified itself. Mathematical tools, putting the idea of domination into a natural and complete form, were suggested by L. V. Kantorovich in 1935–36. He introduced the fundamental notion of a vector space normed by elements of a vector lattice and that of a linear operator between such spaces which is dominated by a positive linear or monotone sublinear operator. He also applied these notions to solving functional equations. In the succeeding years many authors studied various particular cases of latticenormed spaces and different classes of dominated operators. However, research was performed within and in the spirit of the theory of vector and normed lattices. So, it is not an exaggeration to say that dominated operators, as independent objects of investigation, were beyond the reach of specialists for half a century. As a consequence, the most important structural properties and some interesting applications of dominated operators have become available since recently. By the early 1980s, certain qualitative changes took place in the theory of vector lattices. New methods were suggested, while the range of applications was considerably extended and enriched. Radically new ideas came from other branches of mathematics. All these facts provided necessary prerequisites for a deep study of dominated operators and led to formation of a novel theory of dominated operators. The aim of this book is to present the main results on dominated operators which were obtained in the last fifteen years, thus demonstrating a certain ripeness of the theory. The book consists of eight chapters. Chapter 1 contains definitions and prelim-

xii

Preface

inary information about Boolean algebras and vector lattices. Chapter 1 is included mainly for fixing terminology and notation. In Chapter 2 some structural properties are considered of a vector space endowed with a norm taking values in some vector lattice. Here, we address the questions of completing these spaces as well as representing them by means of continuous Banach bundles and measurable Banach bundles. Chapter 3 is devoted to positive operators. The content of Chapter 3 is traditional except for some new results concerning fragments and order-intervalpreserving extension of positive operators. In Chapter 4 we study the general properties of dominated operators. A dominated operator has the least (or exact) dominant under rather weak assumptions. Assigning to each dominant operator its exact dominant we come to some vector norm with values in the vector lattice of regular operators. The central result of this chapter is decomposability of this lattice-normed space of dominants. In Chapter 5 disjointness preserving and decomposable operators are considered. In particular, we give their analytic representations and decompositions into simpler parts. Chapter 6 deals with integrality and pseudointegrality for dominated operators. It turns out that a dominated operator inherits these properties from every dominant. From this fact we deduce criteria for weak integrality and pseudointegrality of dominated operators. Several assertions about the general form of dominated operators are also stated. Various classes of operators under study in functional analysis are often defined using the terms that mix norms and orders. This theme is developed in Chapter 7 in which we introduce some new classes of spaces and operators. The problem of isometric classification of spaces with mixed norm is also briefly discussed. Chapter 8 is devoted to the so-called Boolean-valued analysis of vector lattices and dominated operators. Our starting point is the assertion claiming that each field of the reals in a Boolean-valued model gives rise to a universally complete vector lattice. Therefore, a huge part of the general theory of vector lattices admits a straightforward derivation by translating and interpreting the well known properties of the conventional reals. This chapter also exposes the Boolean-valued approach to more advanced sections of analysis such as lattice-normed spaces, involutive Banach algebras, etc. Elementary exposition of the apparatus of Booleanvalued model theory resides in the Appendix. The diversity of topics and results we handle in the book determines the style of exposition. Information freely accessible to the reader is given without demonstration. All principal results are however furnished with complete proofs. Comments to all chapters contain some additional remarks and a guide to the literature. While writing the book, the author assumed the reader familiar with the standard courses

Preface

xiii

in the theory of vector lattices and positive operators. I seize the opportunity to express my gratitude to all those who helped me in preparation of the book. My pleasant debt is to acknowledge the financial support of the Sobolev Institute of Mathematics of the Siberian Division of the Russian Academy of Sciences, the North Ossetian Scientific Center, the North Ossetian State University, the Russian Foundation for Basic Research, the International Science Foundation, and the American Mathematical Society during the compilation of the monograph. July, 1999

A. G. Kusraev

Chapter 1 Boolean Algebras and Vector Lattices

The present chapter collects some basic facts from the theories of Boolean algebras and vector lattices. It lays the foundation for the theory of lattice-normed spaces and dominated operators to which the book is devoted. An elementary algebraic theory of Boolean algebras is briefly presented in Section 1.1. We expose the main constructions with Boolean algebras such as subalgebras, homomorphic images, ideals, factor algebras, Cartesian products, and partitions of unity (1.1.5). We state a very useful Exhaustion Principle for Boolean algebras and a pair of its corollaries (1.1.6). Some important examples of Boolean algebras are also provided: the algebras of regular open sets and clopen sets, the algebras of Borel sets modulo meager sets and measurable sets modulo negligible sets, etc. (1.1.7). Three representation theorems are proved in Section 1.2: the celebrated Stone Representation Theorem (1.2.3), the Loomis–Sikorski Theorem on representation of σ-complete Boolean algebras (1.2.6), and representation of multinormed Boolean algebras by measure spaces with direct sum property (1.2.10). Ogasawara’s characterization of order completeness of Boolean algebras (1.2.4) and the Sikorski representation of a Boolean homomorphism (1.2.5) via the Stone space are given. The concept of lifting and the corresponding immersion of a measure space into the Stone space of the respective complete Boolean algebra are also presented (1.2.7, 1.2.8). Section 1.3 begins with the notions of vector lattice and vector sublattice, ideal and factor lattice, order and relative uniform convergence, etc. (1.3.2, 1.3.4– 1.3.6). The disjointness relation in a vector lattice provides the three basic Boolean algebras of bands, band projections, and fragments of an order-unity (1.3.3). These are isomorphic to one another for an order complete vector lattice with order-unity (1.3.7). The spectral function of an element of vector lattice and a list of its useful properties are given (1.3.8). The Freudenthal Spectral Theorem is stated as well as a characterization of general vector lattices in which the weak and strong forms of this theorem hold (1.3.9).

2

Chapter 1

Three basic examples of universally complete vector lattices and some interconnections between them, as well as the representation of an arbitrary order complete vector lattice as an order-dense ideal in such spaces is the content of Section 1.4. The first example is the space of continuous functions on a quasiextremal compact space which assume possibly infinite values on a nowhere-dense set depending on a function (1.4.1, 1.4.2). This space is an order σ-complete vector lattice and even an order complete vector lattice provided that the underlying compact space is extremally disconnected (1.4.2). The second example is the space of all spectral functions with values in a σ-complete Boolean algebra (1.4.3). The Stone transform of a σ-complete Boolean algebra can be extended to a linear and lattice isomorphism of the corresponding space of spectral functions onto the space of extended valued continuous functions on the Stone space (1.4.4). Using the Freudenthal Spectral Theorem, we prove that an arbitrary order complete vector lattice is isomorphic to an order-dense ideal of the space of extended valued continuous functions on an extremal compact space (1.4.5, 1.4.6). The third example is the space of (equivalence classes of) measurable functions which is order complete provided that the underlying measure space has the direct sum property (1.4.7). In this case the space of essentially bounded measurable functions admits lifting (1.4.8); moreover, each lifting generates a linear and lattice isomorphism between the space of measurable functions and the space of extended valued continuous functions (1.4.9). Various conditions under which an abstract order complete vector space is isomorphic to an order-dense ideal in some space of measurable functions are also presented (1.4.10). Section 1.5 starts with basic definitions and elementary facts of the theory of normed vector lattices (1.5.1, 1.5.2). We then discuss various useful characterizations of Banach lattices with order continuous norm (1.5.3), as well as monotonically complete Banach lattices and Banach lattices with order semicontinuous norm (1.5.4). Abstract M - and Lp -spaces are introduced (1.5.5) and the corresponding representation theorems are also stated (1.5.6). In particular, every Banach lattice has some AM -structure locally. This observation opens a way to some functional calculus in an arbitrary Banach lattice (1.5.7). Two more applications of a local AM -structure are stated: a disjointly complete Banach lattice has the projection property (1.5.10) and a Banach lattice is order complete if and only if it is disjointly complete (1.5.11). To prove these results, some order properties (1.5.8, 1.5.9) of an AM -space with unity are needed. 1.1. Boolean Algebras In this section we sketch a minimum about Boolean algebras which we need in the sequel. A more explicit exposition may be found elsewhere; for instance, cf. [127, 283, 352, 384].

Boolean Algebras and Vector Lattices

3

1.1.1. To fix terminology and notation, we recall some well-known notions. An ordered set is a pair (M, ≤), where ≤ is an order on M (see A.1.10). An ordered set is also called a partially ordered set or, briefly, a poset. It is in common parlance to apply all names of (M, ≤) to the underlying set M of (M, ≤). We indulge in doing the same elsewhere without further circumlocution. An upper bound of a subset X of a poset M is an element a ∈ M such that x ≤ a for all x ∈ X. A least element of the set of upper bounds of X is called a least upper bound or supremum of X and denoted by sup(X) or sup X. In other words, a = sup(X) if and only if a is an upper bound of X and a ≤ b for every upper bound b of X. By reversal, i.e., by passing from the original order ≤ on a poset M to the reverse or opposite order ≤−1 (x ≤−1 ⇔ y ≤ x), define a lower bound of a subset X of M and a greatest lower bound, inf(X) of X, also called an infimum of X and denoted inf X. If a least upper or greatest lower bound of a set in M exists then it is unique and so deserves the definite article. It can be easily checked that the following commutativity and associativity laws hold in every poset on duly stipulating existence of the suprema and infima in question: (1) sup sup xα,β = sup sup xα,β ; α∈A β∈B

β∈B α∈A

(2) inf inf xα,β = inf inf xα,β ; α∈A β∈B β∈B α∈A
S (3) sup Xα = sup sup Xα ; α∈A α∈A
S Xα = inf inf Xα . (4) inf α∈A

α∈A

1.1.2. A lattice is an ordered set L in which each pair {x, y} has the join x ∨ y := sup{x, y} and meet x ∧ y := inf{x, y}. Given a subset X of a lattice L, we use the notation: _ ^ X := sup(X), X := inf(X), _ _ ^ ^ xα := {xα : α ∈ A}, xα := {xα : α ∈ A}, α∈A

α∈A n _ k=1 n ^

xk := x1 ∨ · · · ∨ xn := sup{x1 , . . . , xn }, xk := x1 ∧ · · · ∧ xn := inf{x1 , . . . , xn }.

k=1

Here (xα )α∈A is a family in L, and x1 , . . . , xn stand for some members of L. The binary operations join (x, y) 7→ x ∨ y and meet (x, y) 7→ x ∧ y act in every lattice L and possess the following properties:

4

Chapter 1 (1) commutativity: x ∨ y = y ∨ x,

x ∧ y = y ∧ x;

(2) associativity: x ∨ (y ∨ z) = (x ∨ y) ∨ z,

x ∧ (y ∧ z) = (x ∧ y) ∧ z.

By induction, from (2) we deduce that each nonempty finite set in a lattice has the meet and join. If each subset of a lattice L has the supremum and infimum then L is a complete lattice. A lattice L is distributive provided that the following distributive laws are valid: (3) x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z); (4) x ∨ (y ∧ z) = (x ∨ y) ∧ (x ∨ z). If a lattice L has the least or greatest element then the former is called the zero of L and the latter, the unity of L. The zero and unity of L are solemnly denoted by 0L and 1L . It is in common parlance to use the simpler symbols 0 and 1 and nicknames zero and unity provided that the context prompts the due details. Note also that 0 and 1 are neutral elements: (5) 0 ∨ x = x, 1 ∧ x = x. W V Specifying the general definitions, also note that ∅ = sup ∅ := 0 and ∅ = inf ∅ := 1. A complement x∗ of a member x of a lattice L with zero and unity is an element x∗ of L such that (6) x ∧ x∗ = 0, x ∨ x∗ = 1. Elements x and y in L are disjoint if x ∧ y = 0. So, every element x is disjoint from any complement x∗ . Note finally that if each element in L has at least one complement then we call L a complemented lattice. Recall by the way that a set U is disjoint whenever every two distinct members of U are disjoint. It is rather evident that an arbitrary lattice L may fail to have a complement to each element of L. 1.1.3. A Boolean algebra is a distributive complemented lattice with distinct zero and unity. Each element x of a Boolean algebra B has a unique complement denoted by x∗ . This gives rise to the mapping x 7→ x∗ (x ∈ B) which is idempotent (i.e., (∀ x ∈ B) (x∗∗ := (x∗ )∗ = x)) and presents a dual isomorphism or an antiisomorphism of B onto itself (i.e., it is an order isomorphism between (B, ≤) and (B, ≤−1 )). The three operations ∨, ∧, and ∗, living in every Boolean algebra B, are jointly referred to as Boolean operations. A Boolean algebra B is complete (σ-complete), if each subset (countable subset) of B has a supremum and an infimum. It is in common parlance to speak of σalgebras instead of σ-complete algebras.

Boolean Algebras and Vector Lattices

5

W V Associated with a Boolean algebra B, the mappings , : P(B) → B are available that ascribe to a set in B its supremum and infimum, respectively. These mappings are sometimes referred to as infinite operations. The infinite operations obey many important rules among which we for instance mention the De Morgan formulas  ∗  ∗ W V ∗ V W ∗ (1) xα = xα , xα = xα , α∈A

α∈A

α∈A

α∈A

and the infinite distributive laws V V (2) x ∨ xα = x ∨ xα , α∈A

(3) x ∧

W

α∈A

xα =

α∈A

W

x ∧ xα ,

α∈A

with xα ∈ B for all α ∈ A. 1.1.4. Let 2 := Z2 := P({∅}) := {0, 1} be the underlying set of the twoelement Boolean algebra now viewed as a field with the following operations: 0 + 0 := 0,

0 + 1 = 1 + 0 := 1,

0 · 1 = 1 · 0 := 0,

0 · 0 := 0,

1 + 1 := 0, 1 · 1 := 1.

Note that every member of 2 is idempotent. Consider an arbitrary set B with the structure of an associative ring whose every element is idempotent: (∀b ∈ B)(b2 = b). In this case B is called a Boolean ring. A Boolean ring is commutative and obeys the identity b = −b for b ∈ B. Each Boolean ring is obviously a vector space and, at the same time, a commutative algebra over 2. Recall that the unity of an algebra differs from its zero by definition. So, we may and will identify the field 2 with the subring of a Boolean ring comprising the zero and unity of the latter. We usually reflect the practice in symbols by letting 0 stand for the zero and 1, for the unity of whatever ring. This agreement leads clearly to a rather popular notational collision: the addition and multiplication of 2 may be redefined on making 0 play the role of 1 and vice versa. It is customary to endow a Boolean ring B with some order by the rule: b1 ≤ b2 ⇔ b1 b2 = b1

(b1 , b2 ∈ B).

The poset (B, ≤) obviously becomes a distributive lattice with the least element 0 and the greatest element 1. In this event, the lattice and ring operations are connected as follows: x ∨ y = x + y + xy,

x ∧ y = xy.

6

Chapter 1

Moreover, to each element b ∈ B there is a unique b∗ ∈ B, the complement of b, such that b∗ ∨ b = 1, b∗ ∧ b = 0. Obviously, b∗ = 1 + b. Hence, each Boolean ring is a Boolean algebra under the above order. In turn, we may transform a Boolean algebra B into a ring by putting x + y := x M y,

xy := x ∧ y

(x, y ∈ B),

where x M y := (x ∧ y ∗ ) ∨ (x∗ ∧ y) is a symmetric difference of x and y. In this case (B, +, · , 0, 1) becomes a unital Boolean ring whose natural order coincides with the initial order on B. Therefore, a Boolean algebra can be viewed as a unital algebra over 2 whose every element is idempotent. 1.1.5. Consider some methods of arranging new Boolean algebras. (1) A nonempty subset B0 of a Boolean algebra B is a subalgebra of B if B0 is closed under the Boolean operations ∨, ∧, and ∗; i.e., {x ∨ y, x ∧ y, x∗ } ⊂ B0 for all x, y ∈ B0 . Under the order induced from B, every subalgebra B0 is a Boolean algebra with the same zero and unity as those of B. In particular, B0 := {0B , 1B } is a subalgebra of B. A subalgebra B0 ⊂ B is regular or complete (σ-regularWor σ-complete) provided that for every set (countable set) A in B0 the elements A V and A, if exist in B, belong to B0 . The intersection of every family of subalgebras is itself a subalgebra. The same holds true for regular (σ-regular) subalgebras, which makes the definition to follow sound. The least subalgebra of B containing a nonempty subset M of B is the subalgebra generated by M . The regular (σ-regular) subalgebra generated by M is introduced in much the same manner. (2) An ideal of a Boolean algebra B is any nonempty set J in B obeying the conditions: x ∈ J, y ∈ J ⇒ x ∨ y ∈ J, x ∈ J, y ≤ x ⇒ y ∈ J. The set Ba := {x ∈ B : x ≤ a}, with a ∈ B, provides an example of an ideal of B. Such an ideal is called principal. If 0 6= e ∈ B then the principal ideal Be with the order induced from B is a Boolean algebra in its own right. The element e plays the role of unity in Be . The lattice operations of Be are inherited from B, and the complementation of Be has the form x 7→ e − x := e ∧ x∗ for all x ∈ B. An ideal J is proper provided that J 6= B. A regular ideal of B is often called a band or component of B.

Boolean Algebras and Vector Lattices

7

(3) Take Boolean algebras B and B 0 and a mapping h : B → B 0 . Say that h isotonic or monotone if (x ≤ y ⇒ h(x) ≤ h(y)). (Note by the way an isotonic mapping from B to B 0 with the opposite order is antitonic.) Say that h is a (Boolean) homomorphism, if for all x, y ∈ B the following equalities are fulfilled: h(x ∨ y) = h(x) ∨ h(y), h(x ∧ y) = h(x) ∧ h(y), h(x∗ ) = h(x)∗ . Every homomorphism h is monotone and the image h(B) of B is a subalgebra of B 0 . If h is bijective then we call h an isomorphism, and B and B 0 , isomorphic Boolean algebras. An injective homomorphism is a monomorphism. A homomorphism h (of B to a complete B 0 ) is complete if h preserves suprema and infima; i.e., h(sup(U )) = sup(h(U )) and h(inf(V )) = inf(h(V )) for all U ⊂ B and V ⊂ B for which there are sup(U ) and inf(V ). Observe that h : B → B 0 is a ring homomorphism (see 1.1.4) if and only if h is a Boolean homomorphism from B into the Boolean algebra Ba0 := [0, a] with a := h(1B ). Given a set C and a bijection h : B → C, we may furnish C with an order by putting h(x) ≤ h(y) whenever x ≤ y. In this event, C turns into a Boolean algebra and h becomes an isomorphism between B and C. (4) Let J be a proper ideal of a Boolean algebra B. Define the equivalence ∼ on B by the rule x∼y ⇔ xMy∈J

(x, y ∈ B).

Denote by ϕ the factor mapping of B onto the factor set B/J := B/∼. Recall that ϕ is also called canonical. Given cosets (equivalence classes) u and v, i.e., members of B/J; agree to write u ≤ v if and only if there are x ∈ u and y ∈ v satisfying x ≤ y. We have thus defined an order on B/J. In this event, B/J becomes a Boolean algebra which is called factor algebra or quotient algebra of B by J. The Boolean operations in B/J make ϕ a homomorphism. So, ϕ is referred to as the factor homomorphism of B onto B/J. If h : B → B 0 is a homomorphism then ker(h) := {x ∈ B : h(x) = 0} is an ideal of B and there is a unique monomorphism g : B/ ker(h) → B 0 satisfying g ◦ ϕ = h, where ϕ : B → B/ ker(h) is the factor homomorphism. Therefore, each homomorphic image of a Boolean algebra B is isomorphic to the factor algebra of B by a suitable ideal. B :=

Q

(5) Take a family of Boolean algebras (Bα )α∈A . Furnish the product α∈A Bα with the coordinatewise order or product order by putting x ≤ y for

8

Chapter 1

x, y ∈ B whenever x(α) ≤ y(α) for all α ∈ A. In this event B becomes a Boolean algebra. Each Boolean operation in B consists in implementing the respective operation in every coordinate Boolean algebra Bα , i.e., it is carried out coordinatewise. The zero 0B and unity 1B of B are as follows: 0B (α) := 0α and 1B (α) := 1α (α ∈ A), where 0α and 1α are the zero and unity in Bα . The Boolean algebra B is the Cartesian product or, simply, product of (Bα )α∈A . (6) A disjoint subset of a Boolean algebra is an antichain. In other words, a subset A of B is an antichain provided that a1 ∧ a2 = 0 for all distinct a1 , a2 ∈ A. If an antichain has the form A := {aξ : ξ ∈ „} then we presume that aξ ∧ aη = 0 whenever ξ 6= η. An antichain A in B is a partition of an element b ∈ B and a partition of W unity, when b is the unity of B, provided that b = A. Let (bξ )ξ∈„ be a partition of unity in B. According to 1.1.5 (2) Bξ := [0, bξ ] is a Boolean algebra with unity bξ . Q The complete Boolean algebra B is isomorphic to ξ∈„ Bξ . Some isomorphism is carried out by sending b ∈ B to the mapping eb by the rule eb(ξ) := b ∧ bξ (b ∈ B). 1.1.6. Recall a certain fundamental property of Boolean algebras. Let B be a Boolean algebra. A subset E of B minorizes a subset B0 of B if to each 0 < b ∈ B0 there is an x in E such that 0 < x ≤ b. It is also in common parlance to call E coinitial or minorant set to B0 . Denote by u.b.(M ) the set of all upper bounds of M . Exhaustion Principle. Let M be a nonempty subset of a Boolean algebra B. Assume given a subset E of B that minorizes the band B0 of B generated by M . Then some antichain E0 exists, E0 ⊂ E, such that u.b.(E0 ) = u.b.(M ) and to each x ∈ E0 there is an element y in M satisfying x ≤ y. C Consider the set A of all antichains A with the following properties: (a) A ⊂ E; (b) to each x ∈ A there is a y ∈ M satisfying x ≤ y. If 0 6= y ∈ M then, by the minorant condition, y ≥ x for some 0 6= x ∈ E. Hence, {x} ∈ A and A is nonempty. The inclusion ordered set A clearly obeys the hypotheses of the Kuratowski–Zorn Lemma. Therefore, there is a maximal element E0 ∈ A. The property (b) from the definition of A implies u.b.(M ) ⊂ u.b.(E0 ). In particular, we are done if u.b.(E0 ) = {1}. To prove the converse inclusion assume that b0 ∈ / u.b.(M ) for some b0 ∈ u.b.(E0 ), b0 6= 1. There exists an element x ∈ M , such that x0 := b∗0 ∧ x 6= 0. By the minorant condition, 0 < y ≤ x0 for some y ∈ E. The set E0 ∪ {y} belongs to A and has essentially more elements than E0 . This contradicts the fact that E0 is maximal, and so u.b.(E0 ) ⊂ u.b.(M ). B

Boolean Algebras and Vector Lattices

9

(1) To each nonempty set M ⊂ B havingWthe least W upper bound, there is an antichain A ⊂ B with the following properties: A = M and, given x ∈ A, we may find y in M such that x ≤ y. S C Choose the minorant E := y∈M [0, y] as a minorant for M and appeal to (1). B (2) A Boolean algebra is complete if and only if any antichain in it has the supremum. 1.1.7. Examples. (1) Recall that P(X) stands for the powerset of X comprising all subsets of X and also denoted by 2X . Given a nonempty set X, observe that the powerset P(X) of X ordered by inclusion is a complete Boolean algebra. This algebra is often called the boolean of X. The Boolean operations on every boolean are the set-theoretic operations of union, intersection, and complementation. (2) Let X be a topological space. Recall that a closed and open subset of X is called clopen. The collection of all clopen sets in X, ordered by inclusion, is a subalgebra of the boolean P(X). Denote this subalgebra by Clop(X). The Boolean operations in Clop(X) are inherited from P(X). Hence, they are settheoretic. However, Clop(X) is not a regular subalgebra of P(X); i.e., the infinite operations in P(X) and Clop(X) may differ essentially. (3) A closed subset F of a topological space X is called regular if F = cl(int(F )); i.e., if F coincides with the closure of the interior of F . By analogy, a regular open set G is defined by the formula G = int(cl(G). Let RC (X) and RO (X) stand for the collections of all regular closed subsets and all regular open subsets, respectively, of X. Furnished with the order by inclusion, RC (X) and RO (X) become complete Boolean algebras. The mapping F 7→ int(F ) (F ∈ RC (X)) is an isomorphism between them. Despite RC (X) and RO (X) are included in the boolean P(X), they are not subalgebras of the latter. For instance, the Boolean operations on RC (X) take the form E ∨ F = E ∪ F,

E ∧ F = cl(int(E ∩ F )),

F ∗ = cl(X − F ).

(4) Denote by Bor(X) the Borel σ-algebra of a topological space X (i.e., the σ-regular subalgebra of the boolean P(X) which is generated by the open sets of X). Consider the ideal N of Bor(X) comprising the meager subsets of X (also called the first category sets in X). The factor algebra Bor(X)/N is a complete Boolean algebra called the algebra of Borel sets modulo meager sets or briefly the Borel-by-meager algebra. We will arrive at an isomorphic algebra if instead of Bor(X) we take the σalgebra of sets with the Baire property. (A subset M of X has the Baire property if

10

Chapter 1

there is an open set G in X such that the symmetric difference M M G is a meager set.) If X is a Baire space, i.e., if X lacks nonempty open meager subsets, then the algebra in question is isomorphic to the algebra of regular closed sets RC (X). (5) Assume given a nonempty set Š, a σ-algebra B ⊂ P(Š) of subsets of Š, and a measure on B which is a positive countably additive function µ : B → R ∪ {+∞}, with R the reals. Countable additivity, as usual, means that [  X ∞ ∞ An = µ µ(An ) n=1

n=1

for every disjoint sequence (An ) of B. A triple (Š, B, µ) is said to be a measure space if the following conditions are met: (a) if A ⊂ Š and A ∩ K ∈ B for each K ∈ B with µ(K) < +∞, then A ∈ B; (b) if A ∈ B and µ(A) = +∞ then there is A0 ∈ B such that A0 ⊂ A and 0 < |A0 | < +∞; (c) if A ∈ B, µ(A) = 0 and A0 ⊂ A then A0 ∈ B. Let N := {A ∈ B : µ(A) = 0}. Then N is a σ-complete ideal. The factor algebra B(Š) := B(Š, B, µ) := B/N is also a σ-algebra called the associated algebra or the algebra of measurable sets by measure zero sets. The measure space (Š, B, µ) is said to possess the direct sum property if B contains a family (Aξ )ξ∈„ of pairwise disjoint sets of finite measure such that the following holds: for every measurable subset A ∈ B of finite measure there exists a countable set of indices ‚ ⊂ „ and a measure zero set A0 ∈ N such that [  A = A0 ∪ (A ∩ Aξ ) . ξ∈‚

If a measure space (Š, B, µ) possesses the direct sum property then the associate Boolean algebra B(Š, B, µ) is complete. 1.2. Representation of Boolean Algebras The Stone Representation Theorem opens up a distinct possibility of representing a Boolean algebra as the Boolean algebra of clopen subsets of a compact space. The basic goal of this section is to prove this theorem and to describe some opportunities it affords. 1.2.1. Let B be an arbitrary Boolean algebra. (1) A character of B is a Boolean homomorphism or, which is the same, a nonzero ring homomorphism χ : B → 2. Denote by X(B) the set of all

Boolean Algebras and Vector Lattices

11

characters of B and make X(B) into a topological space on furnishing it with the topology of pointwise convergence. To put it more explicitly, the topology on X(B) is induced by the product topology of 2B , where we consider 2 with the unique compact topology on this set, the discrete topology of 2. Observe that, unless stated otherwise, we presume all topological spaces to be Hausdorff. Recall that a topological space X is connected whenever the only clopen subsets of X are ∅ and X. A topological space X is totally disconnected provided that each connected subspace of X is at most a singleton. The topological space 2B , called sometimes a Cantor discontinuum, is compact and totally disconnected. A topological space with all these properties is a Boolean space. Evidently, X(B) is a closed subset of 2B . Therefore, X(B) is itself is a Boolean space. Say that the Boolean space X(B) is the character space of a Boolean algebra B. (2) Recall that a nonempty subset F of B is a filter on B provided that x ∈ F, y ∈ F ⇒ x ∨ y ∈ F, x ∈ F, x ≤ y ⇒ y ∈ F. A filter other than B is proper. A maximal element of the inclusion-ordered set of all proper filters on B is an ultrafilter on B. Let U (B) stand for the set of all ultrafilters on B, and denote by U (b) the set of ultrafilters containing b. We endow U (B) with the topology with base {U (b) : b ∈ B}. This definition is sound since it is easy to check that U (x ∧ y) = U (x) ∩ U (y) (x, y ∈ B); i.e., U (B) is closed under finite intersections. The topological space U (B) is often referred to as the Stone space of B and is denoted by S (B). (3) Denote by M (B) the set of all maximal (proper) ideals of a Boolean algebra B. An ideal here may be understood in accord with 1.1.5 (2) or in the conventional sense of ring theory. Clearly, a set J in B is an ideal of B if and only if J ∗ := {x∗ : x ∈ J} is a filter on B. Moreover, J ∈ M (B) ↔ J ∗ ∈ U (B). Therefore, the mapping J 7→ J ∗ presents a bijection between M (B) and U (B). The set M (B) is usually called the maximal ideal space of B and is always furnished with the inverse image topology translated from U (B) which makes the mapping J 7→ J ∗ a homeomorphism. 1.2.2. Recall the prerequisites we need for applying the Gelfand transform in the case of a Boolean algebra. (1) A Boolean ring B is a field if and only if B is the pair of 0 and 1. Hence, there is a unique Boolean field to within isomorphism; namely, 2. C Indeed, a nonzero element x ∈ B is invertible, and so the following implications are valid: xx−1 = 1 ⇒ xxx−1 = 1 ⇒ xx−1 = x ⇒ x = 1. B

12

Chapter 1

Given χ ∈ X(B), denote by χ∗ the mapping x 7→ χ(x)∗ (x ∈ B). Obviously, ker(χ) := {x ∈ B : χ(x) = 0} is an ideal, and ker(χ)∗ is a filter. (2) The mappings χ 7→ ker(χ) (χ ∈ X(B)) and χ 7→ ker(χ)∗ (χ ∈ X(B)) are homeomorphisms of X(B) onto M (B) and U (B), respectively. C The mapping χ 7→ ker(χ) is injective. If J ∈ M (B) then B/J is a field and, by (1), B/J is isomorphic to 2. Fix such an isomorphism λ : B/J → 2; and put χ := λ ◦ ϕ, where ϕ : B → B/J is the factor mapping. Obviously, ker(χ) = J and so the mapping under discussion is bijective. The remaining claims are obvious. B (3) For an x in B to equal zero it is necessary and sufficient that χ(b) = 0 for all χ ∈ X(B). C Assume that x 6= 0. Then the principal ideal {y ∈ B : y ≤ x∗ } is proper, and so can be extended to a maximal ideal J ∈ M (B). This claim, known as the Krull Theorem, is immediate from the Kuratowski–Zorn Lemma. By (2), J = ker(χ) for a certain χ ∈ X(B). Since x ∈ / J; therefore, χ(x) 6= 0. B 1.2.3. Stone Representation Theorem. Each Boolean algebra B is isomorphic to the Boolean algebra of clopen sets of a Boolean space unique up to homeomorphism, the Stone space of B. C Denote by C(X(B), 2) the algebra of continuous 2-valued functions on the character space X(B) of B which is a Boolean space. The Gelfand transform GB assigns to an element x ∈ B the 2-valued function x b : χ 7→ χ(x) (χ ∈ X(B)). Obviously, GB : B → C(X(B), 2) is a injective homomorphism, i.e., a monomorphism (cf. 1.2.2 (3)). Take f ∈ C(X(B), 2) and put Vf := {χ ∈ X(B) : f (χ) = 1}. The set Vf is clopen. By the definition of the topology of X(B), there are b1 , . . . , bk ∈ B and c1 , . . . , cl ∈ B such that Vf := {χ ∈ X(B) : χ(bn ) = 1 (n ≤ k),

χ(cm ) = 0 (m ≤ l)}.

Assign b0 := b1 ∧ · · · ∧ bk , c0 := c1 ∨ · · · ∨ cl and b := b0 ∧ c∗0 . The set Vf can be presented as follows: Vf = {χ ∈ X(B) : χ(b0 ) = 1, χ(c0 ) = 0} = {χ ∈ X(B) : χ(b) = 1} = {χ ∈ X(B) : bb(χ) = 1}. Therefore, f = bb, and so GB is an isomorphism. Assume now that Q1 and Q2 are Boolean spaces such that the mapping h : C(Q1 , 2) → C(Q2 , 2) is an isomorphism of these algebras. If χ is a character of

Boolean Algebras and Vector Lattices

13

C(Q2 , 2) then χ◦h is a character of C(Q2 , 2). Hence, χ 7→ χ◦h is a homeomorphism between the character spaces. On the other hand, the character space of C(Qk , 2) is homeomorphic to Qk . Therefore, the Boolean spaces Q1 and Q2 are homeomorphic. It suffices to note that the algebra C(X(B), 2) is isomorphic to the algebra of clopen sets of the space X(B) and so, of the space U (B) as well. B In view of this theorem, there is a mapping B 7→ Clop(S (B)) which is occasionally called the Stone transform of B. 1.2.4. In the sequel we are mostly interested in complete and σ-complete Boolean algebras. The notion of a complete Boolean algebra is closely tied with that of an extremally disconnected compact space. Recall that a topological space Q is called extremally (quasiextremally) disconnected or simply extremal (quasiextremal ) if the closure of an arbitrary open set (open Fσ -set) in it is open or, which is equivalent, the interior of an arbitrary closed set (closed Gδ -set) is closed. Clearly, an extremal (quasiextremal) space is totally disconnected. Ogasawara Theorem. A Boolean algebra is complete (σ-complete) if and only if its Stone space is extremal (quasiextremal). C We confine exposition to the case of complete Boolean algebra B. Assume further that h is an isomorphism of B on the algebra of clopen sets of the compact spaceSQ := U (B). Take an open set G ⊂ Q. Since Q is totally disconnected, G = U , where U stands for W the union of all clopen subsets of G. Put U 0 := −1 {h (U ) : U ∈ U } and b := U 0 . The clopen set h(b) is the closure of G. Indeed, cl(G) ⊂ h(b) and h(b)\ cl(G) is open. If the latter set is nonempty then h(c) ⊂ h(b)\ cl(G) for some 0 6= c ∈ B. This impliesWin turn that h(c) ∨ h(u) ≤ h(b) for all u ∈ U 0 , which contradicts the equality b = U . Therefore, cl(G) = h(b) is an open set. Assume now that the compact space Q is extremal. Let G stand for some S collection of clopen subsets of Q, and put G := G . The set G is open and the closure cl(G) of G must be open by the hypothesis about Q. Obviously, cl(G) is the least upper bound of G in the Boolean algebra of clopen sets Clop(Q). B 1.2.5. Sikorski Theorem. Assume that B and B 0 are Boolean algebras, and h : B → B 0 is a homomorphism between them. Denote by ı : B → Clop(S (B)) and ı0 : B 0 → Clop(S (B 0 )) the Stone transforms of B and B 0 . There is a unique continuous mapping θ : S (B 0 ) → S (B) such that h(x) = (ı0 )−1 θ−1 (ı(x)) (x ∈ B). The mapping h 7→ S (h) := θ carries out a bijection between the sets of all homomorphisms from B to B 0 and the set of all continuous mappings from S (B 0 ) to S (B). If B 00 is another Boolean algebra and g : B 0 → B 00 is a homomorphism, then S (g ◦ h) = S (h) ◦ S (g). Moreover, S (IB ) = IS (B) .

14

Chapter 1

C Denote Q := S (B) and Q0 := S (B 0 ). If q 0 is an ultrafilter in B 0 then obviously q := {b ∈ B : h(b) ∈ q 0 } is also an ultrafilter in B. Assigning θ(q 0 ) := q we arrive at the mapping θ : q 0 ∈ Q0 7→ q ∈ Q. Taking into consideration the equalities ı(b) = {q ∈ Q : b ∈ q} (b ∈ B) and ı0 (b0 ) = {q 0 ∈ Q0 : b0 ∈ q 0 } (b0 ∈ B 0 ) we deduce ı0 h(b) = {q 0 ∈ Q0 : h(b) ∈ q 0 } = {q 0 ∈ Q0 : b ∈ q} = {q 0 ∈ Q0 : θ(q 0 ) ∈ ı(b)} = θ−1 (ı(b)). In particular, θ(q 0 ) ∈ ı(b) if and only if q 0 ∈ ı0 (h(b)); therefore θ is continuous. The remaining properties of θ are obvious. B Two important particular cases of the situation under consideration are worthy of a special attention. (1) A Boolean algebra B0 is isomorphic to a subalgebra of a Boolean algebra B if and only if the Stone space S (B0 ) of B0 is a continuous image of the Stone space S (B) of B. (2) A Boolean algebra B 0 is a homomorphic image of a Boolean algebra B (or isomorphic with a factor algebra of B; see 1.1.6 (4)) if and only if the Stone space S (B 0 ) of B 0 is homeomorphic to a closed subset of the Stone space S (B) of B. 1.2.6. Loomis–Sikorski Theorem. Let Q be the Stone space of a Boolean σ-algebra B. Denote by Clopσ (Q) the σ-algebra of subsets of Q which is generated by the set Clop(Q) of all clopen subsets of Q. Let  stand for the σ-ideal of Clopσ (Q), comprising all meager sets. Then B is isomorphic with the factor algebra Clopσ (Q)/. If ı is an isomorphism of B onto Clop(Q) and ϕ is the factor mapping of Clopσ (Q) onto the factor algebra Clopσ (Q)/ then the mapping h := ϕ ◦ ı is an isomorphism of B onto Clopσ (Q)/. C Observe that h is a homomorphism as the composite of two homomorphisms. If h(b) = 0 then ı(b) ∈  and ı(b) = ∅, since no nonempty clopen set is meager. Thus h is injective. To prove that h is surjective put F := {A ∈ Clopσ (Q) : (∃b ∈ B) ϕ(A) = h(A)}. Since Clop(Q) ⊂ F ⊂ Clopσ (Q), it suffices to observe that F is a σ-algebra. If A ∈ F then ϕ(Q \ A) = h(b∗ ), so that Q \ A ∈ F . Now, consider a sequence (An ) of F and choose a sequence W∞ S∞(bn ) of B such that ϕ(An ) = h(bn ). According to 1.1.7 (2) ı ( n=1 bn ) = A0 ∪ n=1 ı(bn ) with a nowhere-dense subset A0 ⊂ Q.

Boolean Algebras and Vector Lattices

15

Using this equality we easily deduce ! ! ! ∞ ∞ ∞ [ [ [ ϕ An = ϕ A0 ∪ An = ϕ A0 ∪ ı(bn ) n=1

=ϕ ı

n=1 !! ∞ _

bn

n=1

n=1

=h

∞ _

! bn

n=1

and the result follows. B Birkhoff–Ulam Theorem. Let Q be a compact space. For every Borel set V ∈ Bor(Q) there exists a unique regular open set h(V ) such that the symmetric difference V 4h(V ) is meager. Let N stand for the σ-ideal of Bor(Q), comprising all meager sets. Then h is an order σ-continuous homomorphism from Bor(Q) onto RO (Q). Moreover, the kernel of h coincides with N and Bor(Q)/N is isomorphic to RO (Q). 1.2.7. Consider a measure space (Š, B, µ), and let ϕ : B → B(Š) be the factor homomorphism. A Boolean homomorphism ρ : B(Š) → B is called a lifting of the factor algebra B(Š) if ρ(A) ∈ A for each equivalence class A ∈ B(Š). The latter means that ϕ ◦ ρ is the identity mapping on B(Š); therefore, a lifting is a right-inverse of the homomorphism ϕ. (1) If a measure space (Š, B, µ) possesses the direct sum property then the factor algebra B(Š, B, µ) admits a lifting. C The proof can be found in [375]. B (2) Let ρ be a lifting of the factor algebra S B(Š). Then, for every family (Aξ )ξ∈„ T of elements of B(Š), the union As := ξ∈„ ρ(Aξ ) and the intersection Ai := ξ∈„ ρ(Aξ ) are measurable and, moreover, _ ^ ϕ(As ) = Aξ , ϕ(Ai ) = Aξ . ξ∈„

ξ∈„

(3) For every point ω ∈ Š, denote the ultrafilter {A ∈ B(Š) : ω ∈ ρ(A)} by τ (ω). The mapping τ : Š → Q thus constructed will be called the canonical immersion of Š in Q corresponding to the lifting ρ. 1.2.8. Theorem. Let ρ be a lifting of B(Š, B, µ), let τ be the corresponding canonical immersion of Š in the Stone space Q of the Boolean algebra B(Š, B, µ), and let ı be the Stone transform from B(Š) onto Clop(Q). The following are valid: (1) ρ(A) = τ −1 (ı(A)) for each class A ∈ B(Š); (2) ı−1 (U ) = ϕ(τ −1 (A)) for each clopen set U ∈ Clop(Q); (3) the mapping τ : Š → Q is Borel measurable and the image τ (Š) is dense in Q;

16

Chapter 1

(4) the inverse image τ −1 (N ) of every meager subset N ⊂ Q is measurable in Š and has zero measure. C Assertions (1) and (2) are straightforward. To prove (3) and (4), we consider an arbitrary open subset V ⊂ Q. Choose a family (Uξ )ξ∈„ of clopen subsets with S V = ξ∈„ Uξ . Then (1) implies measurability of the inverse images τ −1 (Uξ ) for all S −1 −1 ξ ∈ „ and 1.2.7 (2) implies measurability of the set τ (V ) = (Uξ ). The ξ∈„ τ W relation ξ∈„ Uξ = cl(V ) in the Boolean algebra Clop(Q), together with (2) and 1.2.7 (2), ensures the equality ϕ(τ −1 (V )) = ϕ(τ −1 (cl V ). B 1.2.9. A function µ : B → R ∪ {+∞} is called additive, countably additive, or completely additive if _  X µ xξ = µ(xξ ) ξ∈„

ξ∈„

for any finite, countable, or arbitrary antichain (xξ ) of B, respectively. To eliminate the trivial additive function µ ≡ +∞ it is always assumed that µ(0B ) = 0. An additive function on Boolean algebra is often referred to as a measure. We say that the function µ is positive, strictly positive, and finite if respectively µ(b) ≥ 0, µ(b) > 0, and µ(b) < +∞ for all 0 6= b ∈ B. Finally, a positive µ is said to be locally finite if for any b ∈ B with 0 < µ(b) there exists 0 < b0 ≤ b such that 0 < µ(b0 ) < +∞. A positive countably additive locally finite measure on the Boolean algebra Clopσ (Q) is called normal if it vanishes on the ideal of meager sets. An extremal space is said to be hyperstonian if there is a normal measure on Clopσ (Q) strictly positive on Clop(Q). Let M+ (B) denote the set of all finite completely additive positive measures on B. A complete Boolean algebra B is said to be multinormed if the set of all finite completely additive measures separates the points of B; in symbols, (∀0 6= b ∈ B) (∃µ ∈ M+ (B)) µ(b) > 0. In the case when there is a finite strictly positive completely additive measure µ on B, the pair (B, µ) is referred to as a normed Boolean algebra. A normed Boolean algebra (B, µ) can be endowed with the metric ρ(x, y) := µ(x M y) and it is not difficult to observe that the metric space (B, ρ) is complete. 1.2.10. Theorem. For a complete Boolean algebra B the following are equivalent: (1) B is a multinormed Boolean algebra; (2) B is isomorphic to the Cartesian product of a family of normed Boolean algebras; (3) there exists a strictly positive locally finite completely additive measure on B;

Boolean Algebras and Vector Lattices

17

(4) B is isomorphic to the associated algebra B(Š, B, µ) for some measure space (Š, B, µ) possessing the direct sum property; (5) the Stone space S (B) is hyperstonian. C (1) ⇒ (2): According to the Exhaustion Principle (see 1.1.6) we may choose a partition of unity (bξ )ξ∈„ in B and a family of positive completely additive measures (µξ )ξ∈„ such that µξ (bξ ) > 0 for all ξ ∈ „. If Bξ is the principal ideal generated by bξ and the restriction of µξ to Bξ is denoted by the same symbol then (Bξ , µξ ) is a normed Boolean algebra and B is isomorphic to the Cartesian product of the
family (Bξ , µξ ) ξ∈„ . (2) ⇒ (3): The measure on B with the required properties can be defined by µ(b) :=

X

µξ (b ∧ bξ )

(b ∈ B).

ξ∈„

(3) ⇒ (4): Observe first that if Š := S (B) then Clopσ (Š) consists of the sets U 4N where U ∈ Clop(Š) and N ⊂ Š is meager. Now, let ν be a strictly positive locally finite completely additive measure on Clop(Š). If B := Clopσ (Š) and the measure µ on B is defined by letting µ(U 4N ) := ν(U ) then (Š, B, µ) is a measure space possessing the direct sum property and Boolean algebras B(Š, B, µ) and Clop(Š) are isomorphic, see [162]. (4) ⇒ (5): According to 1.2.8 a normal measure µ ¯ on Clopσ (Q) which is strictly ¯(A) := µ(τ −1 (A)) (a ∈ Clopσ (Q)). positive on Clop(Q) can be obtain by putting µ (5) ⇒ (1): Assume that Q := S (B) is hyperstonian, and let µ be a normal measure on Clopσ (Q) strictly positive on Clop(Q). Take an arbitrary b ∈ B. Since µ is locally finite there is a clopen set V with µ(V ) > 0 and V ⊂ ı(b) where ı : B → Clop(Q) is the Stone transform. Putting µb (x) := µ(V ∩ ı(x)) (x ∈ B) we arrive at a finite positive completely additive measure µb on B with µb (b) = µ(V ) > 0. B 1.3. Vector Lattices In this section we give some preliminaries to the theory of vector lattices; a more explicit exposition may be found elsewhere [15, 23, 145, 162, 163, 262, 336, 341, 388, 409]. 1.3.1. Let F be a linearly ordered field. An ordered vector space over F is a pair (E, ≤), where E is a vector space over F and ≤ is an order in E satisfying the following conditions: (1) if x ≤ y and u ≤ v then x + u ≤ y + v whatever x, y, u, v ∈ E might be;

18

Chapter 1

(2) if x ≤ y then λx ≤ λy for all x, y ∈ E and 0 ≤ λ ∈ F. Informally speaking, we may “sum inequalities in E and multiply them by positive members of F.” This circumstance is worded as follows: ≤ is an order compatible with vector space structure or, briefly, ≤ is a vector order. Equipping a vector space E over F with some vector order is equivalent to indicating a set E+ ⊂ E, called the positive cone of E, with the following properties: E+ + E+ ⊂ E+ ,

λE+ ⊂ E+ (0 ≤ λ ∈ F),

E+ ∩ −E+ = {0}.

Moreover, the order ≤ and the cone E+ are connected by the relation x ≤ y ⇔ y − x ∈ E+

(x, y ∈ E).

The elements of E+ are called positive. An ordered vector space E is called Archimedean if for any pair of elements x, y ∈ E the relation (∀n ∈ N) nx ≤ y implies x ≤ 0. In the sequel, all ordered vector spaces are assumed Archimedean. 1.3.2. A vector lattice is an ordered vector space that is also a lattice. Thereby in a vector lattice there exist a least upper bound sup{x1 , . . . , xn } := x1 ∨ · · · ∨ xn and a greatest lower bound inf{x1 , . . . , xn } := x1 ∧ · · · ∧ xn for every finite set {x1 , . . . , xn } ⊂ E. In particular, every element x of a vector lattice has the positive part x+ := x ∨ 0, the negative part x− := (−x)+ := −x ∧ 0, and the modulus |x| := x ∨ (−x). Let E be a vector lattice. For all x, y, z ∈ E the following relations are valid: (1) x = x+ − x− , |x| = x+ + x− = x+ ∨ x− ; (2) x ≤ y ⇔ x+ ≤ y + & y − ≤ x− ; (3) x ∨ y = 21 (x + y + |x − y|), x ∧ y = 12 (x + y − |x − y|); (4) |x| ∨ |y| = 12 (|x + y| + |x − y|), |x| ∧ |y| = 21 (|x + y| − |x − y|); (5) x + y = x ∨ y + x ∧ y, |x − y| = x ∨ y − x ∧ y; (6) x + y ∨ z = (x + y) ∨ (x + z), x + y ∧ z = (x + y) ∧ (x + z); (7) x, y, z ∈ E+ ⇒ (x + y) ∧ z ≤ (x ∧ z) + (y ∧ z); (8) |x − y| = |x ∨ z − y ∨ z| + |x ∧ z − x ∧ z|. Let (xα ) and (yα ) be families in E for which sup(aα ) and inf(yα ) exist. Then for any z ∈ E the infinite distributive laws are valid: (9) z ∧ supα (xα ) = supα (z ∧ xα ), z ∨ inf α (yα ) = inf α (z ∨ yα ). For the same (xα ), (yα ), and z the following useful relations are also true;

Boolean Algebras and Vector Lattices

19

(10) z + supα (xα ) = supα (z + xα ); (11) z + inf α (yα ) = inf α (z + yα ); (12) supα (xα ) = − inf α (−xα ). An order interval in E is a set of the form [a, b] := {x ∈ E : a ≤ x ≤ b}, where a, b ∈ E. The following constantly used property of vector lattices is frequently called the Riesz Decomposition Property. (13) [0, x + y] = [0, x] + [0, y] (x, y ∈ E+ ); We indicate only two corollaries of (13). (14) (x1 + · · · + xn ) ∧ y ≤ x1 ∧ y + · · · + xn ∧ y (xk , y ∈ E+ ); Vn Pm P (15) k=1 l=1 xk,l ≤ j∈J x1,j(1) ∧ · · · ∧ xn,j(n) where xk,l ∈ E+ and J is the set of all functions j : {1, . . . , n} → {1, . . . , m}. 1.3.3. Two elements x and y are called disjoint if |x|∧|y| = 0. The disjointness of x and y is denoted by x ⊥ y. (1) The following properties of disjointness are easy from 1.3.2: x ⊥ y ⇔ |x + y| = |x − y| ⇔ |x| ∨ |y| = |x| + |y|; x+ ⊥ x− ; (x − x ∧ y) ⊥ (y − x ∧ y); x ⊥ y ⇒ |x + y| = |x| + |y|, (x + y)+ = x+ + y + , (x + y)− = x− + y − . Let u ∈ E+ and e ∧ (u − e) = 0 for some 0 ≤ e ∈ E. Then e is said to be a fragment, or a part, or a component of u, or a unit element with respect to u. (2) The set E(u) of all fragments of u with the order induced by E is a Boolean algebra. The lattice operations in E(u) are taken from E and the Boolean complement has the form e∗ := u − e (e ∈ E(u)). The disjoint complement M ⊥ of a nonempty set M ⊂ E is defined as M ⊥ := {x ∈ E : (∀y ∈ M ) x ⊥ y}. A nonempty set K in E meeting the identity K = K ⊥⊥ is called a band (a component in the Russian literature) of E. Every band of the form {x}⊥⊥ with x ∈ E is called principal. (3) The inclusion-ordered set of all bands of E is denoted by B(E) and presents a complete Boolean algebra. The Boolean operations of B(E) take the shape: L ∧ K = L ∩ K,

L ∨ K = (L ∪ K)⊥⊥ ,

L∗ = L⊥

(L, K ∈ B(E)).

20

Chapter 1

The Boolean algebra B(E) is the base of E. Let K be a band of the vector lattice E. If there is an element sup{u ∈ K : 0 ≤ u ≤ x} in E then it is called the band projection of x onto K and is denoted by [K]x (or πK x). Given an arbitrary x ∈ E, we put [K]x := [K]x+ − [K]x− . The band projection of an element x ∈ E onto K exists if and only if x is representable as x = y + z with y ∈ K and z ∈ K ⊥ . Furthermore, y = [K]x and z = [K ⊥ ]x. Assume that to each element x ∈ E there is a band projection onto K, then the operator x 7→ [K]x (x ∈ E) is a linear idempotent and 0 ≤ [K]x ≤ x for all 0 ≤ x ∈ E, called a band projection or an order projection. (4) The set P(E) of all band projections ordered by π ≤ ρ ⇔ π ◦ ρ = π is a Boolean algebra. The Boolean operations of P(E) take the shape π ∧ ρ = π ◦ ρ,

π ∨ ρ = π + ρ − π ◦ ρ,

π ∗ = IE − π

(π, ρ ∈ (E)).

The band projection onto a principal band is called principal. (5) The principal projection πu := [u] := [u⊥⊥ ], where 0 ≤ u ∈ E, can be calculated by the following rule simpler than that indicated above: πu x = sup{x ∧ (nu) : n ∈ N}. A vector lattice E is said to have the projection property (principal projection property) if every band (every principal band) in B(E) is a projection band. Clearly, every Kσ -space has the principal projection property. 1.3.4. The order relation in a vector lattice generates different types of convergence. Let (A, ≤) be an upward-directed set. A net (xα ) := (xα )α∈A in E is called increasing (decreasing) if xα ≤ xβ (xβ ≤ xα ) for α ≤ β (α, β ∈ A). We say that a net (xα ) in a vector lattice E o-converges to x ∈ E if there exists a decreasing net (eβ )β∈B in E such that inf{eβ : β ∈ B} = 0 and for each β ∈ B there is α(β) ∈ A with |xα − x| ≤ eβ (α ≥ α(β)). In this event, we call x the (o)

o-limit of the net (xα ) and write x = o-lim xα or xα → x. If a net (eβ ) in this definition is replaced by a sequence (λn e)n∈N , where 0 ≤ v ∈ E+ and (λn )n∈N is a numerical sequence with limn→∞ λn = 0, then we say that a net (xα )α∈A converges relatively uniformly or more precisely e-uniformly to x ∈ E. The elements e and x are called the regulator of convergence and the r-limit (r)

of (xα ), respectively. The notations x = r-limα∈A xα and xα → x are also frequent. A net (xα )α∈A is called o-fundamental (r-fundamental with regulator e) if the net (xα − xβ )(α,β)∈A×A o-converges (respectively, r-converges with regulator e) to zero. A vector lattice is said to be relatively uniformly complete if every rfundamental sequence is r-convergent.

Boolean Algebras and Vector Lattices

21

The presence of order convergence in a vector lattice allows us to determined the sum of an infinite family (xξ )ξ∈„ . Indeed, given θ := {ξ1 , . . . , ξn } ∈ Pfin („), put yθ := xξ1 + . . . + xξn . So, we arrive at the (yθ )θ∈‚ , where ‚ := Pfin („) is naturally ordered by inclusion. Assuming that there is some x satisfying x = o-limθ∈‚ yθ , we call the family (xξ ) summable in order or order summable or o-summable. The P element x is the o-sum of (xξ ) and we write x = o- ξ∈„ xξ . Obviously, if xξ ≥ 0 (ξ ∈ „) then for the o-sum of the family (xξ ) to exist it isPnecessary and sufficient that the net (yθ )θ∈‚ has the supremum, in which case o- ξ∈„ xξ = supθ∈‚ yθ . If (xξ ) is a disjoint family then X − oxξ = sup x+ ξ − sup xξ . ξ∈„

ξ∈„

ξ∈„

1.3.5. (1) A linear subspace J of a vector lattice is called an order ideal or o-ideal (or, finally, just an ideal, when it is clear from the context what is meant) if the inequality |x| ≤ |y| implies x ∈ J for arbitrary x ∈ E and y ∈ J. Every order ideal of a vector lattice is a vector lattice. If an ideal J possesses the additional property J ⊥⊥ = E (or, which is the same, J ⊥ = {0}) then J is referred to as an order-dense ideal of E (the term “foundation” is current in the Russian literature). (2) Let J be an ideal of a vector lattice E. Then the factor space e e is determined by the E := E/J is also a vector lattice, provided that the order on E e standing for the canonical factor mapping. positive cone ϕ(E+ ), with ϕ : E → E The factor lattice E/J is Archimedean if and only if N is closed under relative uniform convergence. If E is an f -algebra and J is a ring and order ideal then E/N is an f -algebra and ϕ is algebra homomorphism. If E is a Kσ -space and J is sequentially order-closed then E/J is a Kσ -space and ϕ is sequentially order continuous. (3) Denote by I (E) the set of all order ideals of E ordered by inclusion. Then I (E) is a complete lattice, with the lattice operations defined as I ∧J := I ∩J and I ∨J := I +J. Moreover, the lattice I (E) is distributive. The sublattice Ip (E) of principal ideals is also distributive. 1.3.6. (1) A vector sublattice is a vector subspace E0 ⊂ E such that x ∧ y, x ∨ y ∈ E0 for all x, y ∈ E0 . We say that a sublattice E0 is minorizing if, for every 0 6= x ∈ E+ , there exists an element x0 ∈ E0 satisfying the inequalities 0 < x0 ≤ x. We say that E0 is a majorizing or massive sublattice if, for every x ∈ E, there exists x0 ∈ E0 such that x ≤ x0 . Thus, E0 is a minorizing or majorizing sublattice if and only if E+ \ {0} = E+ + E0 + \ {0} and E = E+ + E0 , respectively. (2) A set in E is called (order ) bounded (or o-bounded ) if it is included in some S order interval. The o-ideal generated by the element 0 ≤ u ∈ E is the set ∞ E(u) := n=1 [−nu, nu]; clearly, E(u) is the smallest o-ideal in E containing u.

22

Chapter 1

If E(u) = E then we say that u is a strong unity or strong order-unity and E is a vector lattice of bounded elements. If E(u)⊥⊥ = E then we say that u is an order-unity or weak order-unity. It is evident that an element u ∈ E+ is an order-unity if {u}⊥⊥ = E; i.e., if E lacks nonzero elements disjoint from u. (3) An element x ≥ 0 of a vector lattice is called discrete if [0, x] = [0, 1]x; i.e., if 0 ≤ y ≤ x implies y = λx for some 0 ≤ λ ≤ 1. A vector lattice E is called discrete or atomic if, for every 0 6= y ∈ E+ , there exists a discrete element x ∈ E such that 0 < x ≤ y. If E lacks nonzero discrete elements then E is said to be continuous or diffuse. 1.3.7. A vector lattice is said to be (conditionally) order complete if each nonvoid order bounded set in it has least upper and greatest lower bounds. If, in a vector lattice, least upper and greatest lower bounds exist only for countable bounded sets, then it is called countably order complete. An order complete vector lattice and a countably order complete vector lattice are frequently referred to as a Dedekind complete vector lattice and a Dedekind σ-complete vector lattice or, in the Russian literature, K-space(= Kantorovich space) and a Kσ -space, respectively. We say that a K-space (Kσ -space) is universally complete or extended if its every subset (countable subset) of pairwise disjoint elements is bounded. (1) Theorem. Let E be an arbitrary K-space. Then E has the projection property and the operation of projecting onto bands determines the isomorphism K 7→ [K] of the Boolean algebras B(E) and P(E). If there is an order-unity 1 in E then the mappings π 7→ π1 from P(E) into E(E) and e 7→ {e}⊥⊥ from E(E) into B(E) are isomorphisms of Boolean algebras, too. A K-space is o-complete in the sense that every o-fundamental net in it is oconvergent. Each Kσ -space and, hence, a K-space is Archimedean. Henceforth all vector lattices are presumed to be Archimedean. Consider an order-bounded net (eα )α∈A in a K-space E , and let e ∈ E. (2) An order-boundednet (eα )α∈A o-converges to e if and only if the relation o-limα∈A [d] (|eα − e| − d)+ = 0 holds in the Boolean algebra P(E) for all positive d ∈ E. C It is easy to verify the necessity of the criterion in question. To prove its sufficiency, assign e0 := inf α∈A supβ>α |eβ − e|. If the net (eα )α∈A does not converge to e then e0 > 0 and, thus, there are π ∈ P(E), d ∈ D, and n ∈ N such that 0 < πd/n < e0 . Therefore, for each index α ∈ A, we have h i   sup [d] |eβ − e| > d/n = [d] sup |eβ − e| > d/n > π, β>α

β>α

  which contradicts the convergence of [d] |eα − e| > d/n to zero. B

Boolean Algebras and Vector Lattices

23

(3) Suppose that E is a K-space with order-unity 1; while (eα )α∈A is a bounded net in E, and e ∈ E. Then o-limα∈A eα = e if and only if the relation  o-limα∈A (|eα − e| − 1/n)+ = 0 holds in the Boolean algebra P(E) for all n ∈ N. 1.3.8. Let E be a Kσ -space with order-unity 1. We call the projection of the order-unity to the band {x}⊥⊥ the trace of x and denoted it by ex . Therefore, ex := sup{1 ∧ (n|x|) : n ∈ N}. The trace ex serves both as an order-unity of {x}⊥⊥ and a unit element of E. Given a real λ, denote the trace of the positive part of λ1 − x by exλ ; i.e., exλ := e(λ1−x)+ . The function λ 7→ exλ (λ ∈ R) arising in this case is called the spectral function or characteristic of x. Theorem. Let E be an arbitrary Kσ -space with order-unity 1 and P be a dense subfield of R. The spectral function λ 7→ exλ (λ ∈ R) of x ∈ E has the following properties: (1) (∀ λ, µ ∈ R) (λ ≤ µ ⇒ exλ ≤ exµ ); V W (2) ex+∞ := µ∈P exµ = 1, ex−∞ := µ∈P exµ = 0; W (3) exλ = {exµ : µ ∈ P, µ < λ} (λ ∈ R); (4) x ≤ y ⇔ (∀ λ ∈ P) (eyλ ≤ exλ ); W (5) ex+y = {exµ ∧ eyν : µ, ν ∈ P, µ + ν = λ}; λ W x (6) ex·y {eµ ∧ eyν : 0 ≤ µ, ν ∈ P, µν = λ} (x ≥ 0, y ≥ 0); λ = W (7) e−x {1 − ex−µ : µ ∈ P, µ < λ} = (1 − ex−λ ) · e(x+λ1) ; λ = (8) ex∧y = exλ ∨ eyλ ; ex∨y = exλ ∧ eyλ (λ ∈ R); λ λ W (9) x = inf(A) ⇔ (∀ λ ∈ P) (exλ = {eaλ : a ∈ A}); |x|

(10) eλ = exλ ∧ (1 − ex−λ ) ∧ ex+λ1

(λ ∈ R);

−x x αx (11) eαx λ = eλ/α (α > 0), eλ = e−λ/α (α < 0)

(λ ∈ R);

x ∗ cx x (12) ecx λ = c ∧ eλ + c (λ > 0), eλ = c ∧ eλ (λ ≤ 0)

(c ∈ E(E)).

A formula for x = sup(A) similar to (9) is generally not true. Nevertheless the following more complicated formula is valid:  x = sup(A) ⇔ (∀ λ ∈ P)

exλ

=

_^

{eaν

 : a ∈ A} .

ν f (t) there are λ, µ ∈ ƒ such that f (t) < µ < λ < ν, so that t ∈ Uµ ⊂ Wλ and h(t) < λ < ν. Letting ν tend to f (t), obtain h(t) ≤ f (t). The same inequality is immediate for f (t) = +∞. By analogy, Vµ ⊂ Uλ for µ < λ. Hence, f (t) ≤ g(t) for all t ∈ Q. Writing (b) as Wµ ⊂ Vλ (µ < λ), and arguing as above, conclude that g(t) ≤ h(t) for all t ∈ Q. Therefore, f = g = h. The fact that f is continuous follows from the equalities [ {Vµ : µ < λ, µ ∈ ƒ}, \ {f ≤ λ} = {h ≤ λ} = {Wµ : µ > λ, µ ∈ ƒ}, {f < λ} = {g < λ} =

since Vµ is open whereas Wµ is closed for all µ ∈ ƒ. B (2) Let Q be a quasiextremal compact space. Assume that Q0 is an open dense Fσ -set in Q and f : Q0 → R is a continuous function. Then there is a unique continuous function f¯ : Q0 → R such that f (t) = f¯(t) (t ∈ Q0 ). C Indeed, if Uµ := cl({f < µ}) then the mapping µ 7→ Uµ (µ ∈ R) increases and meets the condition (b) of (1). Therefore, there is a unique function f¯ : Q → R satisfying {f¯ < µ} ⊂ Uµ ⊂ {f¯ ≤ µ} (µ ∈ R). Obviously, in this case f¯  Q0 = f , i.e. the restriction of f¯ to Q0 coincides with f . B 1.4.2. Let Q be a quasiextremal compact space. Denote by C∞ (Q) the set of all continuous functions x : Q → R assuming the values ±∞ possibly on a nowheredense set. Order C∞ (Q) by assigning x ≤ y whenever x(t) ≤ y(t) for all t ∈ Q. Then, take x, y ∈ C∞ (Q) and put Q0 := {|x| < +∞} ∩ {|y| < +∞}. In this case Q0 is open and dense in Q. According to 1.4.1 (2), there is a unique continuous function z : Q → R such that z(t) = x(t) + y(t) for t ∈ Q0 . It is this function z that we declare the sum of x and y. In an analogous way we define the product of a pair of elements. Identifying the number λ with the identically λ function on Q, we obtain the product of x ∈ C∞ (Q) and λ ∈ R.

Boolean Algebras and Vector Lattices

27

Clearly, the space C∞ (Q) with the operations and order introduced above is a vector lattice and a faithful f -algebra. The identically one function 1 is a ring and order-unity. The order ideal generated by 1 is the space C(Q) of all continuous numeric functions on Q. (1) The space C∞ (Q) is a universally σ-complete Kσ -space. T∞ C Take a bounded increasing sequence (xn ) of elements of C∞ (Q). Put Vλ := n=1 {xn ≤ λ} and Uλ := int Vλ . Then Vλ is a closed Gδ -set and, by assumption, Uλ is a clopen set. According to 1.4.1 (2) there is a unique function x : Q → R such that {x < λ} ⊂ Uλ ⊂ {x ≤ λ} for all λ ∈ R. Now, it is not difficult to check that x = supn xn . Universally σ-completeness is obvious. B (2) The base of the vector lattice C∞ (Q) is isomorphic to the Boolean algebra of all regular open (closed) subsets of Q. C The same argument as used in (1) works. B (3) The space C∞ (Q) is an order complete vector lattice if and only if Q is extremal. C In the case when Q is extremal, the order completeness of C∞ (Q) may be proved as in (1). The converse follows from the Ogasawara Theorem 1.2.4, since Boolean algebras Clop(Q) and E(1) are isomorphic. B According to these arguments we may describe suprema and infima in C∞ (Q) as follows. If (xα ) is an order-bounded family in C∞ (Q) then x = supα xα if and only if there exists a comeager subset Q0 ⊂ Q such that x(t) = supα xα (t) for all t ∈ Q0 . 1.4.3. According to 1.3.8, to each element of a Kσ -space with order-unity there corresponds its spectral function; moreover, the operations transform in a rather definite way. This circumstance suggests that an arbitrary Kσ -space with unity can be realized as a space of “abstract spectral functions.” We will expatiate upon this. A resolution of unity or resolution of the identity in a Boolean algebra B is defined as a mapping e : R → B satisfying the conditions (1) s ≤ t → e(s) ≤ e(t) (s, t ∈ R); W V (2) t∈R e(t) = 1, t∈R e(t) = 0; W (3) s∈R,s 0 and 0(t) ¯ := 0 if P, s < t} and the zero element 0 ¯ := 1 if t > 1 and 1(t) ¯ := 0 if t ≤ 1. Finally, define the product of t ≤ 0. Set 1(t) an element e ∈ K(B) and a real α ∈ R by the rules (αe)(t) := e(t/α)

(α > 0, t ∈ R),

(αe)(t) := (−e)(−t/α)

(α < 0, t ∈ R).

To each element b ∈ B assign the resolution of unity ¯b defined as ¯b(t) := 1 if t > 1, ¯b(t) := b∗ := 1 − b if 0 ≤ t < 1, and ¯b(t) := 0 if t ≤ 0. Theorem. Let B be a complete (σ-complete) Boolean algebra. The set K(B) with the above-introduced operations and order represents a universally complete ¯ serves as an orderK-space (a universally σ-complete Kσ -space). The element 1 ¯ unity and the mapping b 7→ b (b ∈ B) is an isomorphism of Boolean algebras B and ¯ E(1). C The proof is elementary. Most effort is put into the routine calculation with resolutions of unity. B 1.4.4. Theorem. Let E be a Kσ -space with order-unity 1 and B := E(1). The mapping sending an element x ∈ E to the spectral function λ 7→ exλ (λ ∈ R) is an isomorphism of E onto an order-dense ideal in K(B). If E is universally σ-complete then E and K(B) are isomorphic. C Denote by h the mapping from E to K(B) we are interested in. By virtue of Theorem 1.3.8 h is linear and preserves order. Moreover, h is injective according to the Freudenthal Spectral Theorem 1.3.9. Prove that h(E) is an order ideal in K(B). Assume 0 ≤ s ≤ h(y) where s ∈ K(B) and y ∈ E. Let λ > µ, b := s(λ) − s(µ), and y0 := µb. It can be easily checked that µ¯b ≤ ¯bs, since  if t > µ,    1, b ∧ s(t) + b∗ , if t > 0, ∗ (µ¯b)(t) = b , if 0 < t ≤ µ, (¯bs)(t) =  0, if t ≤ 0.  0, if t ≤ 0; Therefore, h(y0 ) = µh(b) = µ¯b ≤ ¯bs ≤ s ≤ h(y) and we obtain y0 ≤ y. Finally, we deduce λ(s(λ) − s(µ)) = y0 + (λ − µ)b ≤ y + (λ − µ)1.

Boolean Algebras and Vector Lattices

29

Now, consider a partition of the real axis of the form βN := (tn )n∈Z where tn := n/N and N ∈ N. The disjoint sum ¯(βN ) := x

X

tn+1 (s(tn+1 ) − s(tn ))

n∈Z

exists in E, since tn+1 (s(tn+1 ) − s(tn )) ≤ y + (1/N )1 as was proven above. Denote ¯(βN ). Every element of the form by A the sequence of all elements x x(βN ) :=

X

tn (s(tn+1 ) − s(tn ))

n∈Z

x(β)}. Note that is a lower bound of A. Therefore, there exists x := inf(A) := inf{¯ x ¯(β)

eλ

=

W

{s(tn ) : tn < λ}.

Hence, by 1.3.8 (8), we infer exλ =

_ a∈A

eaλ =

_

s(t) = s(λ) (λ ∈ R).

t∈R,t 0 and A0 ∈ A , A0 ⊂ A, implies either µ(A0 ) = 0 or µ(A0 ) = µ(A). The discreteness of L0 (Š, A , µ) is equivalent to the fact that the measure µ is purely atomic, i.e., every set of nonzero measure contains an atom of µ. The equivalence class containing the identically unity function is an order and ring unity in L0 (Š, A , µ). 1.4.8. A mapping ρ : L∞ (Š) → L ∞ (Š) is said to be a lifting of L∞ (Š) if for all α, β ∈ R and f, g ∈ L∞ (Š) the following are true: (a) ρ(f ) ∈ f and dom(ρ(f )) = Š; (b) if f 6 g, then ρ(f ) 6 ρ(g) everywhere on Š; (c) ρ(αf +βg) = αρ(f )+βρ(g), ρ(f g) = ρ(f )ρ(g), ρ(f ∨g) = ρ(f )∨ρ(g), ρ(f ∧ g) = ρ(f ) ∧ ρ(g);

32

Chapter 1

(d) ρ(0) = 0 and ρ(1) = 1 everywhere on Š. Theorem. For a measure space (Š, A , µ), the following are equivalent: (1) (Š, A , µ) possesses the direct sum property; (2) L∞ (Š) admits a lifting; (3) L0 (Š, A , µ) is order complete and universally complete. The base of the K-space L0 (Š, A , µ) is isomorphic to the Boolean algebra B(Š, A , µ) of measurable sets modulo zero-measure sets. 1.4.9. Suppose that (Š, A , µ) is a measure space with a direct sum property. Let ρ be a lifting of L∞ (Š, A , µ). and τ : Š → Q be the corresponding canonical immersion of Š into the Stone space Q of the Boolean algebra B(Š, A , µ), see Denote by τ ∗ the mapping that sends each function f ∈ C∞ (Q) to the equivalence class of the measurable function f ◦ τ . Theorem. The mapping τ ∗ is a linear and order isomorphism from C∞ (Q) onto L0 (Š, A , µ). The image of C(Q) under the isomorphism τ ∗ coincides with L∞ (Š, A , µ). For u ∈ L0 (Š, A , µ) the function u b := τ ∗−1 (u) is called the Stone transform of u. 1.4.10. Now, we will answer the question: which vector lattices are representable as order-dense ideals in L0 (Š, A , µ)? An order complete vector lattice E is said to be a Kantorovich–Pinsker space if its base B(E) is a multinormed Boolean algebra (= there exists an essentially positive locally finite countably additive measure on B(E), see 1.2.9), or equivalently, if E contains an order-dense ideal with a total set of order continuous functionals. It can be proved that if (Š, A , µ) is a measure space possessing the direct sum property then L0 (Š, A , µ) is a Kantorovich–Pinsker space. Theorem. Each Kantorovich–Pinsker space E is linearly and order isomorphic to an order-dense ideal of L0 (Š) for a suitable measure space (Š, A , µ) with the direct sum property. If an order-unity 1 is fixed in E, then among such isomorphisms there is a unique isomorphism taking 1 to the equivalence class of the identically one function on Š. The space E is universally complete if and only if the image of each isomorphism from E onto an order-dense ideal of L0 (Š) coincides with L0 (Š). C The proof can be obtained by combining 1.2.10, 1.4.5, and 1.4.9. B 1.4.11. Let (Š, A , µ) be a measure space with the direct sum property. An arbitrary order ideal E in L0 (Š, †, µ) is called an ideal space on (Š, †, µ). We recall some basic properties of ideal function spaces. In this subsection fe := f ∼ stands for the coset of a measurable function f . (1) Every ideal space is an order complete vector lattice.

Boolean Algebras and Vector Lattices

33

(2) Let M be a subset in L0 (Š, A , µ) unbounded from above. Then M contains a countable subset that is also unbounded from above. (3) If µ is σ-finite then every nonempty order-bounded set M in ideal space contains a countable subset (fen ) with sup(M ) = supn fen . Moreover, fe := supn fen can be computed pointwise: f (t) = supn fn (t) (t ∈ Š). (4) A sequence (fen ) ⊂ E o-converges to fe if and only if it is orderbounded (in E) and fn (t) → f (t) for almost all t ∈ Š. For fe ∈ L0 (Š, A , µ) denote the support of fe by supp fe:= supp(f ) := {t ∈ Š : f (t) 6= 0}. Let ef stand for the coset of the characteristic function of supp(f ). The support of a nonempty subset M ⊂ L0 (Š, A , µ) is a measurable subset Š0 ⊂ Š such that sup{ef : f ∈ M } is the coset of the characteristic function of Š0 . Observe that supp(M ) is determined up to a set of measure zero. We write A ⊂ B mod (µ) if µ(B \ A) = 0. (5) Let M be a nonempty subset of E, Š0 := supp(M ) and Š1 := Š\Š0 . Then n o ⊥ e M = f ∈ E : supp(f ) ⊂ Š1 mod (µ) , n o M ⊥⊥ = fe ∈ E : supp(f ) ⊂ Š0 mod (µ) . (6) Let K be a band in E, and let χK be the characteristic function of supp(K). Then the band projection [K] has the form: [K](fe) = (χK f )∼ (fe ∈ E). (7) An ideal space E is order dense in L0 (Š, A , µ) if and only if supp(E) = Š mod (µ). If, in addition, µ is σ-finite then for every 0 ≤ f ∈ L0 (Š, A , µ) there exists an increasing sequence (fn ) in E such that f = o-limn fn . 1.5. Normed Vector Lattices In this section we expose various classes of normed vector lattices determined by the interplay of order and norm. Some fundamental properties are given. A more detailed presentation can be found in [23, 162, 231, 242, 336, 341, 388]. 1.5.1. Let E be a vector lattice. A norm k · k on E is called a lattice norm if |x| ≤ |y| implies kxk ≤ kyk for all x, y ∈ E. A lattice norm can be equivalently defined by following two relations: 0 ≤ x ≤ y ⇒ kxk ≤ kyk (x, y ∈ E) and k |x| k = kxk (x ∈ E). If k · k is a lattice norm on E, the pair (E, k · k) is called a normed (vector) lattice. A norm complete normed lattice is called a Banach lattice. It is immediate from 1.3.2 (1, 8) and monotonicity of a lattice norm that the mappings x 7→ x+ , x 7→ x− , x 7→ |x|, and the lattice operations (x, y) 7→ x ∨ y

34

Chapter 1

and (x, y) 7→ x ∧ y are uniformly continuous. In particular, the positive cone, the disjointness relation, and every band are closed. The closure of a vector sublattice (order ideal) of E is a vector sublattice (order ideal) of E. It is also clear that each of the lattice operations of a normed lattice E admits a unique continuous e so that E e becomes a Banach lattice. We state extension to the norm completion E also a useful criterion for norm completeness. For an arbitrary normed vector lattice E the following are equivalent: (1) E is a Banach lattice; (2) each increasing Cauchy sequence of positive elements in E has the least upper bound; (3) every absolutely convergent series of positive elements in E is order convergent. If, in addition, E has the principal projection property then each of the following two conditions is equivalent to (1): (4) each laterally increasing Cauchy sequence of positive elements in E has the least upper bound; (5) every absolutely convergent series of pairwise disjoint positive elements in E is order convergent. 1.5.2. Consider an arbitrary vector lattice E. A linear functional f on E is called order-bounded if the image of every order interval in E under f is a bounded set in R. Denote by E ∼ the set of all order-bounded functionals on E. A functional f is positive if f (x) ≥ 0 for all x ∈ E+ . The space E ∼ becomes an ordered vector ∼ of all positive functionals and space if a vector order in it is defined by the cone E+ is referred to as order dual space. A linear functional f ∈ E ∼ is called order continuous if limα f (xα ) = 0 for every decreasing net (xα ) in E with inf xα = 0. The set of all order continuous functionals is denoted by En∼ . In the sequel E 0 stands for the space of all norm continuous functionals on a normed lattice E. Each x ∈ E defines by x ^ : f 7→ hx, f i = f (x) (f ∈ E ∼ ) an order-bounded ∼ functional on E . The mapping assigning x ^ to each x ∈ E by hf, x ^i = hx, f i (f ∈ ∼ E ) is called the canonical embedding in the second order dual or the evaluation mapping This mapping preserves order and is injective provided that E ∼ is pointseparating. We now list a few basic properties of order-bounded functionals and order duals. (1) The order dual space is an order complete vector lattice. (2) The set En∼ is a band in E ∼ . (3) The evaluation mapping is an isomorphism from E onto a sublattice

Boolean Algebras and Vector Lattices

35

of E ∼∼ provided that E ∼ is point-separating. (4) Let E0 be a massive subspace of an ordered vector space E. Each positive linear functional on E0 has a positive linear extension to E. (5) If E is a normed lattice then E 0 is an order ideal in E ∼ . (6) If E is a normed lattice then E 0 is an order complete Banach lattice. (7) If E is a Banach lattice then E 0 = E ∼ and En∼ is a band in E 0 . (8) Let E0 be a vector sublattice of the normed vector lattice E. Each norm continuous positive linear functional on E0 has a norm preserving positive linear extension to E. 1.5.3. A Banach lattice is said to have an order continuous norm (σ-order continuous norm) if limα kxα k = 0 for every downward directed net (sequence) (xα ) with inf α xα = 0. In the Russian literature, order continuity is frequently called Property (A). We say that the norm in E is laterally σ-continuous if limn ken k = 0 for every decreasing sequence (en ) in E such that (en − en+1 ) ⊥ en+1 (n ∈ N) and inf n en = 0. Theorem. For every Banach lattice E, the following are equivalent: (1) The norm of E is order continuous; (2) E is order complete and the norm of E is order continuous; (3) E is order σ-complete and the norm of E is σ-order continuous; (4) E is order complete and the norm of E is laterally σ-continuous; (5) E is order σ-complete and each closed order ideal of E is a band; (6) every closed order ideal of E is a projection band; (7) every order interval of E is σ(E, E 0 )-compact; (8) every norm continuous linear functional on E is order continuous; (9) the canonical embedding E → E 00 maps E onto an ideal of the Banach lattice E 00 . C The proof can be found in [242, 263, 336, 341]. B 1.5.4. We say that a normed lattice E is monotonically complete or possesses Property (B) if for every increasing net (xα ) in E+ with supα kxα k < ∞ there exists the supremum x = supα xα . Every monotonically complete normed lattice is order complete. A vector lattice E is said to have the order semicontinuous norm or to possess Property (C) if o-limα xα = x implies limα kxα k = kxk for every increasing net (xα ) in E+ . The sequential variants of Properties (A), (B), and (C) are denoted by (A)σ , (B)σ , and (C)σ , respectively. A Banach lattice possessing (A)σ and (B)σ is called a KB-space.

36

Chapter 1

The significance of these properties for the theory of Banach lattices has various aspects. We emphasize only those related to the properties of canonical embedding into the second dual lattice. (1) A Banach lattice E with En∼ point-separating is monotonically complete if and only if the canonical embedding E → (En∼ )∼ n is an isometry. (2) An order complete Banach lattice E with En∼ point-separating has order semicontinuous norm if and only if the canonical embedding E and (En∼ )∼ n are isomorphic under the canonical embedding. (3) Let E be an order complete Banach lattice with En∼ point-separating. Then E is monotonically complete and has the order semicontinuous norm if and only if the image of E under the canonical embedding is the range of a positive contractive projection. (4) A Banach lattice E is a KB-space if and only if the canonical embedding κ maps E onto a band in E 00 . 1.5.5. (1) A Banach lattice E is called an abstract M -space or AM -space, for short, if kx ∨ yk = kxk ∨ kyk (x, y ∈ E+ ). If the unit ball of an AM -space E contains a largest element e, then e is a strong order-unity and the unit ball of E coincides with the symmetric order interval [−e, e]. In this case E is said to be an AM -space with unity. (2) Let E be an arbitrary vector lattice and u ∈ E. We may introduce the following seminorm in the principal ideal E(u): kxku := inf{λ ∈ R : |x| ≤ λu} (x ∈ E(u)). The seminorm k · ku is a norm if and only if the lattice E(u) is Archimedean. It can be easily seen that relative uniform convergence in E is the convergence in the norm of (E(v), k · kv ). Therefore, a vector lattice E is relatively uniformly complete if and only if the normed lattice (E(u), k · ku ) is complete for every u ∈ E+ . Let E be a Banach space and 0 6= u ∈ E. Then (E(u), k · ku ) is an AM -space with unity u and the identity embedding E(u) → E is continuous. This simple proposition asserts that every Banach lattice is locally arranged as an AM -space, which enables us to reduce the study of some properties of Banach lattices to that of AM -spaces. Such an approach is demonstrated below in 1.5.7, 1.5.11, and 1.5.12. (3) A Banach lattice E is called an ALp -space if kx + yk = kxkp + kykp


p1

Boolean Algebras and Vector Lattices

37

for all disjoint x, y ∈ E+ . Here 1 ≤ p ≤ ∞ and, in case p := ∞, we let (tp + sp )1/p := max{s, t} (0 < s, t ∈ R). The terms AM -space and AL-space are conventionally used instead of AL∞ -space and AL1 -space, respectively. This definition of AM space is equivalent to that given in (1) despite the fact that the required equality holds only for disjoint pairs of elements. 1.5.6. We recall three well-known facts from the theory of Banach lattices. Proofs can be found in [231, 242, 341]. (1) Theorem. Let E be an AM -space. Then there exist a compact space Q and a family of triples (tα , sα , λα )α∈A with tα , sα ∈ Q and 0 ≤ λα < 1 such that E is isometrically isomorphic to the closed sublattice F := {x ∈ C(Q) : (∀α ∈ A) x(tα ) = λα x(sα )}. (2) Brothers Kre˘ın–Kakutani Theorem. Every AM -space with unity is linearly isometric and order isomorphic to the space of continuous functions C(Q) on some compact space Q. (3) Theorem. If 1 ≤ p < +∞ then every ALp -space is linearly isometric and order isomorphic to Lp (Š, A , µ) for a suitable measure space (Š, A , µ) with the direct sum property. 1.5.7. The study of functions f : Rl → R, for which f (e1 , . . . , el ) can naturally be defined for e1 , . . . , el ∈ E, is called the functional calculus. To assign some values to expressions of the form f (e1 , . . . , el ) we will use the local AM -structure of a Banach lattice. Denote by H (Rl ) the space of all real continuous functions f on Rl which are positively homogeneous, i.e. f (λt) = λf (t) for all t ∈ Rn and λ ≥ 0. Clearly, H (Rn ) is a vector lattice under pointwise operations. Denote S := {(t1 , . . . , tl ) ∈ Rl : |t1 | + · · · + |tl | = 1}. Clearly, each function f ∈ H (Rl ) is completely defined by its values on S. Therefore, the restriction mapping f 7→ f |S is a linear and lattice isomorphism from H (Rl ) onto C(S). Thus, we may regard H (Rl ) as a Banach lattice with strong order-unity and norm kf k∞ := sup{|f (x)| : x ∈ S}. We define dxj ∈ H (Rn ) by dxj (t1 , . . . , tn ) = tj (j := 1, . . . , n). Theorem. Let E be a uniformly complete vector lattice, e1 , . . . , el ∈ E, and e := e1 + · · · + en . Then there is a unique lattice homomorphism h : H (Rl ) → E
l such that h(dxj ) = ej (j := 1, . . . , l). Moreover, h H (R ) is the e-uniform closure of the vector sublattice generated by {e1 , . . . , el }. C See [242, 341]. B 1.5.8. A normed (Banach) ideal space on (Š, †, µ) is defined as an ideal space E on (Š, †, µ) endowed with a lattice norm making E into a normed (Banach) space.

38

Chapter 1

If a sequence xn converges to x in the norm of a Banach ideal space E then xn → x(µ). If (xn ) is a Cauchy sequence in E then it converges in measure to some x ∈ L0 (T, †, µ). Now, we consider some properties of the vector lattice C(Q) of all continuous functions on a compact topological space Q. A vector lattice is called disjointly complete (disjointly σ-complete if every its order-bounded antichain (countable antichain) has supremum. 1.5.9. Theorem. For a compact space Q, the following are equivalent: (1) C(Q) is order complete (σ-complete); (2) C(Q) is disjointly complete (σ-complete); (3) Q is extremal (quasiextremal); (4) C(Q) possesses the projection property (principal projection property). C The implications (1) ⇒ (2) and (1) ⇒ (4) areSobvious. ∞ (2) ⇒ (3): Take an open G ⊂ Q such that G = n=1 Fn with Fn closed for all n ∈ N. We will prove that cl G is open. Since every compact topological space is normal, we may find an open G1 ⊂ Q with F1 ⊂ G1 ⊂ cl G1 ⊂ G. By the same argument there is an open G2 ⊂ Q with cl G1 ∪ F2 ⊂ G2 ⊂ cl G2 ⊂ G. By induction we may construct a sequence (Gn ) of open subsets such that cl Gn ∪Fn+1 ⊂ Gn+1 ⊂ cl Gn+1 ⊂ G. Clearly, ∞ ∞ [ [ G= Gn = cl Gn . n=1

n=1

Without loss of generality, we may assume that all Gn differ from one another. Now, we define a sequence of continuous functions (xn ) such that 0 ≤ xn ≤ 1 and  1, if t ∈ cl Gn−1 , xn (t) := 0, if t ∈ cl Gn−2 ∪ (Q \ Gn−1 ), where G−1 := G0 := ∅. It is easily seen that each of the sequences (x3n ), (x3n−1 ), and (x3n−2 ) is pairwise disjoint. By assumption there exist yk := supn∈N x3n−k (k := 0, 1, 2). Put y := y0 ∨ y1 ∨ y2 . Obviously, yk (Q \ cl G) = {0} (k := 0, 1, 2), and therefore y(Q \ cl G) = {0}. At the same time y(G) = {1}. Since y is continuous, we have y(cl G) = {1}. Thus, y = χcl G and cl G is clopen set. Now, suppose that C(Q) is disjointly complete. Then Q is quasiextremal as we have just proved. Moreover, the Boolean algebras Clop(Q) and E(1) are isomorphic; therefore, Q is the Stone space of the Boolean algebra of clopen sets Clop(Q). Whence the required assertion follows from 1.1.6 (2) and the Ogasawara Theorem (1.2.4).

Boolean Algebras and Vector Lattices

39

(3) ⇒ (1): follows from 1.4.2 (1, 3). (4) ⇒ (3): See [341]. B 1.5.10. Theorem. Let Q be a compact space and E be a vector sublattice in C(Q) containing constants and separating the points of Q. If E is disjointly complete then Q is extremal and E possesses the projection property. C First, we prove that Q is extremal. In view of 1.5.9 it suffices to establish disjoint completeness of C(Q). Take a pairwise disjoint family (xξ )ξ∈„ , 0 ≤ xξ ≤ 1, and put  xξ (t), if xξ (t) > 0, z(t) := 0, if t ∈ G := int{t ∈ Q : (∀ξ ∈ „)xξ (t) = 0}. Clearly, z is a continuous function defined on an open dense subset in Q. According to the Stone–Weierstrass Theorem, for each ξ ∈ „ there exists a sequence (yξ,n ∈ E+ ), such that 0 ≤ xξ − yξ,n ≤ (1/n)1, (n ∈ N). By assumption there exists yn := supξ∈„ yξ,n . This sequence is uniformly convergent since |yn+p − yn | ≤ (1/n)1 (n ∈ N). Put x := lim yn . Observe that 0 ≤ z(t)−yn (t) ≤ (1/n)1 for all t ∈ dom(z); thus x(t) = lim yn (t) = z(t) for all t ∈ dom(z). This means that x is a continuous extension of z onto Q; therefore x = supξ∈„ xξ . Now, take an arbitrary element e ∈ E+ and prove that eχQ0 ∈ E for any clopen set Q0 ⊂ Q. Since E contains constants and separates points of Q, for each q ∈ Q0 there exists eq ∈ E with eq (q) > e(q). Put Uq := {t ∈ Q : eq (t) > e(t)} and choose a finite subcover Uq1 , . . . , Uqm of Q0 . Then the following relations are consistent: e0 := eq1 ∨ · · · ∨ eqm ∈ E; e0 (t) = 0 (t ∈ Q \ Q0 );

e0 (t) ≥ e(t) (t ∈ Q0 ).

Now it is clear that e ∧ e0 = eχQ0 ∈ E. B 1.5.11. Theorem. Every Archimedean disjointly complete vector lattice has the projection property. Every Archimedean disjointly σ-complete vector lattice has the principal projection property. C We consider the claim only for disjointly complete vector lattices. The disjointly σ-complete case is handled by the same arguments. Disjoint completeness of E is inherited by order ideals. In particular, all principal ideals E(u) are disjointly ] of E(u) is an AM -space. Using complete. Clearly, the norm completion E(u) ] is isomorphic to C(Q) for some compact space Q. 1.5.6 (2) we conclude that E(u) The isomorphism brings E(u) in a vector sublattice containing constants and separating points of Q. According to 1.5.10 E possesses the projection property. B

40

Chapter 1

1.5.12. Theorem. Let E be a Banach lattice or relatively uniformly complete vector lattice. Then the following are equivalent: (1) E is order complete (order σ-complete); (2) E is disjointly complete (disjointly σ-complete); (3) E has the projection property (principal projection property). C The implication (1) ⇒ (2) is obvious and (2) ⇒ (3) follows from 1.5.11. To prove (3) ⇒ (1) use the local AM -structure of a Banach lattice (relatively uniformly complete vector lattice) and apply (1) ⇔ (4) from 1.5.9. B 1.5.13. Let A be a Boolean algebra. Denote by ba(A ) the set of all bounded finitely additive functions µ : A → R with a vector structure induced from RA . The set ba+ (A ) of all positive bounded finitely additive functions is a cone and defines some structure of a vector lattice in ba(A ). The lattice operations have the form: µ ∨ ν(a) = sup{µ(b) + µ(a ∧ b∗ ) : b ≤ a}, µ ∧ ν(a) = inf{µ(b) + µ(a ∧ b∗ ) : b ≤ a}, |µ|(a) = sup{|µ(b)| : b ≤ a}, where µ, ν ∈ ba(A ). The total variation of µ is defined as kµk := |µ|(1) where 1 is the unity of A . It is clear that the function µ 7→ kµk is a lattice norm. The vector lattice ba(A ) of bounded, finitely additive functions on a Boolean algebra is an AL-space. In particular, ba(A ) is order complete and the norm of ba(A ) is order continuous. 1.6. Comments 1.6.1. (1) The theory of Boolean algebras originated from the classical work by G. Boole “An Investigation of the Laws of Thought, on Which Are Founded the Mathematical Theories of Logic and Probabilities” [49, 50]. The author himself formulated his intentions as follows: “The design of the following treatise is to investigate the fundamental laws of those operations of the mind by which reasoning is performed; to give expression to them in the language of a Calculus, and upon this foundation to establish the science of Logic and construct its method to make that method itself the basis of a general method for the application of the mathematical doctrine of probabilities; and, finally, to collect from the various elements of truth brought to view in the course of these inquiries some probable intimations concerning the nature and constitution of the human mind.... ” Pursuing this end, G. Boole carried out, in fact, the algebraization of the logical system lying behind the classical mathematical reasoning. In a result, he become the author of the algebraic system omnipresent under the name of Boolean algebra.

Boolean Algebras and Vector Lattices

41

(2) Definition 1.1.2 looks somewhat strange at first sight. Indeed, it does not reveal the reasons for whatever distributive lattice to be called an algebra since the term “algebra” refers customarily to conventional objects (cf. Lie algebra, Banach algebra, C ∗ -algebra, etc.). The arising ambiguity is easily eliminated because a Boolean algebra is in fact an algebra over the two-element field. At the same time, it is perfectly natural to view Boolean algebras in different contexts at different angles. It is worth emphasizing that the particular Boolean algebras we deal with in functional analysis appear mostly as distributive complemented lattices. (3) Recall that a universal algebra is an algebraic system without predicates. This concept makes available another definition of Boolean algebra. Namely, a Boolean algebra B is a universal algebra (B, ∨, ∧, ∗, 0, 1) with two binary operations ∨ and ∧, one unary operation ∗, and two distinguished elements 0 and 1 obeying the conditions: (a) ∨ and ∧ are commutative and associative; (b) ∨ and ∧ are both distributive relative to one another; (c) x and x∗ complement one another; (d) 0 and 1 are neutral for ∨ and ∧, respectively. Conversely, given a universal algebra B of the above type, make B into a poset by letting x ≤ y whenever x ∧ y = x for x, y ∈ B. In this event, note that (B, ≤) is a distributive complemented lattice with join ∨, meet ∧, complementation ∗, zero 0, and unity 1. 1.6.2. (1) Theorem 1.2.3 shows that every Boolean algebra is completely determined from its Stone space. In more detail, each property of a Boolean algebra B translates into the topological language, becoming a property of the Stone space S (B) of B. This way of studying Boolean algebras is the Stone representation method. A detailed presentation can be found in [144, 283, 352]. (2) Denote by Bool the category of Boolean algebras and Boolean homomorphisms, and let Comp stand for the category of compact spaces and continuous mappings. Then 1.2.3 and 1.2.5 may be paraphrased as follows: Theorem. The mapping S is a contravariant functor from the category Bool to the category Comp. (3) About measure spaces with the direct sum property see [78, 162, 239, 375]. The problem of whether lifting exists was formulated by A. Haar for the Lebesgue measure on the real axis and was solved by J. von Neumann in 1931. It was D. Maharam who proved that every σ-finite measure admits lifting [270]. The general case is handled in [375]. For the results in 1.2.7 and 1.2.8 as well as other aspects of lifting theory see in [78, 239, 375]. 1.6.3. (1) In the history of functional analysis, the rise of the theory of ordered vector spaces is commonly attributed to the contribution of G. Birkhoff, L. V. Kantorovich, M. G. Kre˘ın, H. Nakano, F. Riesz, H. Freudenthal, et al. At present, the

42

Chapter 1

theory of ordered vector spaces and its applications constitute a noble branch of mathematics representing, in fact, one of the main sections of contemporary functional analysis. The theory with a vast field of applications is thoroughly covered in many monographs (see [15, 19, 23, 100, 139, 145, 162, 163, 170, 184, 185, 197, 230, 231, 242, 262, 282, 336, 341, 388, 399, 409, 411]). Observe also the surveys [60–62] each with a rich reference list. (2) The credit for finding the most important instance of ordered vector spaces, an order complete vector lattice or a K-space, is due to L. V. Kantorovich. This notion appeared in Kantorovich’s first fundamental article on this topic [153] where he wrote, “In this note, I define a new type of space that I call a semiordered linear space. The introduction of such a space allows us to study linear operations of one abstract class (those with values in such a space) as linear functionals.” Here L. V. Kantorovich stated an important methodological principle, the heuristic transfer principle for K-spaces. (3) The concept of universal completion for a K-space was introduced in another way by A. G. Pinsker (see [163]). He also proved existence of a universal completion unique to within isomorphism for an arbitrary K-space. The concepts of order-unity, fragment, and spectral function were introduced by H. Freudenthal. He also established Theorem 1.3.9 (1) (see [163, 388]). (4) Weak and strong Freudenthal properties for general vector lattices were introduced and studied B. Lavriˇc [234]. Using the principal ideal lattice he also introduced the notion of zero-dimensional vector lattice with strong unity and proved that a vector lattice has the weak Freudenthal property if and only if every principal ideal in it is zero-dimensional. M. Pannenberg [315] extended Lavriˇc’s notion of dimension zero to an arbitrary value of dimension by using the lattice Ip (E), see 1.3.5 (3). He also defined the topological stable rank tsr(E) for a Banach lattice E with quasi-interior positive elements and established that dim Ip (E) = tsr (E). (5) Sometimes the use of Freudenthal’s Spectral Theorem may be replaced by the following assertion: In a Banach lattice E with the principal projection property every order interval is the norm closed convex hull of its own extreme points. (The extreme points of the interval [0, e] are exactly the fragments of e.) The assertion remains valid if E is a Banach lattice with quasi-interior positive elements having the topological stable rank zero, see M. Pannenberg [315]. 1.6.4. (1) The fact that for a complete Boolean algebra B the set K(B) of resolutions of unity is a universally complete K-space with base isomorphic to B (see 1.4.3 and 1.4.4) is due to L. V. Kantorovich [163]. Theorem 1.4.4 was obtained by A. G. Pinsker (see [163]). The representation of an arbitrary K-space as an order-dense ideal in C∞ (Q) was established independently by B. Z. Vulikh and

Boolean Algebras and Vector Lattices

43

T. Ogasawara (see [163, 388]). (2) The theory of vector lattices and K-spaces arose historically before the theory of general ideal spaces which began developing in the 1950s with research by J. Diedonn`e, G. G. Lorentz, I. Galperin, H. W. Ellis, A. C. Zaanen, W. A. J. Luxemburg, et al. The synthesis of these theories occurred in the 1960s in the works of W. A. J. Luxemburg, A. C. Zaanen, and G. Ya. Lozanovski˘ı (see [61, 409]). After the papers of Diedonn`e, one of the most common terms for an ideal space is a K¨ othe space. The term “ideal space” was coined within the school of M. A. Krasnosel0 ski˘ı on integral operators and equations (P. P. Zabre˘ıko, P. E. Sobolevski˘ı, Ya. B. Rutitski˘ı et al.) due to the fact that an ideal space is actually an o-ideal in L0 (see [188, 189]). 1.6.5. (1) The material presented in 1.5.1–1.5.8 is traditional for the theory of normed lattices and can be found in [23, 162, 231, 242, 263, 336, 341, 409]. (2) Theorems 1.5.10, 1.5.11 and the equivalence (1) ⇔ (2) in 1.5.12 belong to A. I. Veksler and V. A. Ge˘ıler [381]. (2) In this book we never touch the theory of positive operators in a Banach space with cone which originated with M. G. Kre˘ın’s articles written in the 1940s and developed later by M. A. Krasnosel0 ski˘ı in Russia and by many research groups in other countries. The history, state of the art, and various applications are reflected in [35, 66, 139, 170, 184, 185, 187, 188, 285, 389, 390, 399].

Chapter 2 Lattice-Normed Spaces

In this chapter we consider structural properties of a vector space with some norm taking values in a vector lattice. Such a vector space is called a lattice-normed space; an LNS for short. The most important peculiarities of LNSs are connected with the decomposability property (2.1.1 (4)). The latter allows us, in particular, to indicate a complete Boolean algebra of linear projections in a lattice-normed space, which is isomorphic and closely related to the Boolean algebra of band projections of the norm lattice (2.1.3, 2.1.4). Moreover, a decomposable LNS admits a compatible module structure over a certain ring of orthomorphisms (2.1.8). These facts are closely related with the disjointness relation induced by the vector norm (2.1.2). If an LNS is simultaneously a vector lattice then there is another disjointness relation connected with its lattice structure. Some simple interrelation between them is reflected in the notions of norm-indecomposable and norm-n-decomposable elements (2.1.9). It turns out that every norm-n-decomposable elements is very often the sum of n norm-indecomposable elements (2.1.10). Partitions of unity in a Boolean algebra lead to the operation of mixing elements in a lattice-normed space (2.2.1). If there exists a mixing of every (vector) norm-bounded family in a lattice-normed space then the space is called disjointly complete (2.1.5). For instance, such are Banach–Kantorovich spaces (BKSs), i.e. decomposable and order complete latticenormed spaces (2.2.1). Moreover, a decomposable lattice-normed space is order complete if and only if it is disjointly (laterally) complete and complete with respect to relative uniform convergence (2.2.2). In addition, every LNS has a lattice norm completion that can be obtained by consecutively applying the operations of mixing and closure with respect to relative uniform convergence (2.2.8, 2.2.9). LNSs consisting of continuous and measurable vector-functions (Section 2.3) are most frequent in analysis. A more general example of an LNS is presented by a space of sections of a continuous Banach bundle (CBB) (2.4.7). It turns out that each BKS is linearly isometric to a space of almost global sections of some continuous Banach bundle (2.4.10). Thus, an LNS admits some functional representation.

Lattice-Normed Spaces

45

This representation is uniquely determined in the class of ample continuous Banach bundle (2.4.6). Every continuous Banach bundle over an extremal compact space has a unique ample hull to within isometry (2.4.11). The notion of ample continuous Banach bundle allows us, in particular, to introduce the CBB of operator spaces (2.4.13). The notion of measurable Banach bundle (MBB) is a measurable analog of a continuity structure. Measurable sections are defined as the limits of almost everywhere convergent sequences (on the subsets of finite measure) of elements of some set of sections given axiomatically and called a measurability structure (2.5.1). Under some assumptions the space of equivalence classes of measurable sections (2.5.2) of a measurable Banach bundle appears to be a bo-complete LNS (2.5.3). The notion of lifting in a factor space of measurable sections of an MBB is introduced (2.5.5). Measurable Banach bundles with lifting are called liftable. There is a constructive connection between liftable measurable Banach bundles and complete continuous Banach bundles. A method for constructing a liftable MBB given an arbitrary complete CBB over the corresponding extremal compact space is described in 2.5.8 and 2.5.9. Moreover, this method is universal: every liftable MBB is so obtainable (2.5.10). These facts enable us to represent Banach–Kantorovich spaces as the spaces of measurable functions associated with measurable Banach bundles (2.5.11–2.5.13). 2.1. Preliminaries In this section, we introduce basic notions and consider simple properties of vector norms. 2.1.1. Consider a vector space X and a real vector lattice E. (All vector lattices under consideration are assumed Archimedean.) A mapping · : X → E+ is a vector (E-valued ) norm if it satisfies the following axioms: (1) x = 0 ⇔ x = 0 (x ∈ X); (2) λx = |λ| x (λ ∈ R, x ∈ X); (3) x + y ≤ x + y (x, y ∈ X). A vector norm is called a decomposable or Kantorovich norm if (4) for all e1 , e2 ∈ E+ and x ∈ X, from x = e1 + e2 it follows that there exist x1 , x2 ∈ X such that x = x1 + x2 and xk = ek (k := 1, 2). In the case when condition (4) is valid only for disjoint e1 , e2 ∈ E+ , the norm is said to be disjointly-decomposable or, in short,

d-decomposable. A triple X, · , E in brief (X, E), X, · , or X with the default parameters
omitted is a lattice-normed space (over E) if · is an E-valued norm in the vector space X. The space E is called the norm lattice
of X. If the norm · is decomposable (d-decomposable) then the space X, · itself is called decomposable

46

Chapter 2

(d-decomposable). In the sequel we use the abbreviation LNS for “lattice-normed space.” 2.1.2. If x ∧ y = 0 then we call the elements x, y ∈ X disjoint and write x ⊥ y. As in the case of a vector lattice, a set of the form M ⊥ := {x ∈ X : (∀y ∈ M ) x ⊥ y}, with ∅ 6= M ⊂ X, is called a band or a component. The symbol B(X) denotes the set of all bands ordered by inclusion. We say that K ∈ B(X) is a projection band if K ⊕ K ⊥ = X. The projection h(π) onto the band K along the band K ⊥ is called a band projection. A lattice-normed space X is said to have the projection property whenever every band in X is a projection band. For uniformity, we often write B(X) instead of B(X) and take liberty of using the terminology from the theory of vector lattices. Such a treatment is admissible as long as X is considered as a LNS by itself. But, if X is simultaneously a vector lattice, we should be more accurate to avoid confusions, see  2.1.4 and 2.1.9 below. Given L ⊂ E and M ⊂ X, we let by definition h(L) := x ∈ X : x ∈ L and M := { x : x ∈ M }. It is clear that h(L) ⊂ L ∩ X . ⊥⊥

(1) Suppose that every band of the vector lattice E0 := X contains the norm of some nonzero element. Then B(X) is a complete Boolean algebra and ⊥⊥
and the mapping L 7→ h(L) is an isomorphism of the Boolean algebras B X B(X). C It is clear that the mapping h preserves the intersection of any nonempty family of bands. Therefore, h preserves infima, since in
the algebras under consider⊥⊥
ation they coincide with intersections. Moreover, h {0} = {0} and h X = X.
⊥⊥ Thus, it is sufficient to establish that h(L⊥ ) = h(L)⊥ for L ∈ B X . The inclusion h(L⊥ ) ⊂ h(L)⊥ is obvious. If 0 6= x ∈ h(L)⊥ then x is disjoint from all the elements of the form y in L. At the same time, x ∈ / h(L⊥ ) implies that ⊥⊥ 0 < e ≤ x for some e ∈ L+ . Therefore, in the band {e} there are no elements of the form y 6= 0, which contradicts our assumption. B (2) If elements x, y ∈ X are disjoint, then x + y = x + y . C Indeed, from the relations x ∧ y = 0 and x ≤ x + y + y we infer that x ≤

x+y + y




∧ x ≤ x+y ∧ x ≤ x+y .

Similarly, y ≤ x + y ; therefore, x + y = x ∨ y ≤ x+y . B (3) For every pair of disjoint elements e1 , e2 ∈ E the decomposition x = x1 + x2 with x1 = e1 and x2 = e2 is unique.

Lattice-Normed Spaces

47

C Assume x1 = y1 = e1 , x2 = y2 = e2 , and x = x1 + x2 = y1 + y2 . Then x1 − y1 ⊥ y2 − x2 , since x1 − y1 ≤ x1 + y1 = 2e1 and x2 − y2 ≤ 2e2 . By (2) 0 = (x1 − y1 ) + (x2 − y2 ) = x1 − y1 + x2 − y2 , whence x1 = y1 and x2 = y2 . B (4) Assume the hypotheses of Proposition (1) to be satisfied. Suppose that X is d-decomposable and there exists a band projection π onto L ∈ B(E0 ). Then the projection h(π) onto the band K := h(L) along the band K ⊥ exists and, moreover, π x = h(π)x for all x ∈ X. C In view of the d-decomposability condition, for every x ∈ X, there are x1 , x2 ∈ X such that x = x1 + x2 , x1 = π x , and x2 = π ⊥ x . This means that X is the direct sum of the bands K and K ⊥ . Let h(π) be the projection onto K along K ⊥ . By the definition of the isomorphism h, we have h(π)x ∈ K = h(πE0 ), i.e., h(π)x ∈ πE0 . Thus, π ⊥ h(π)x = 0, i.e., π h(π)x = h(π)x . Since h(π)x and h(π ⊥ )x are disjoint, in view of (2) we have
π x = π h(π)x + h(π ⊥ )x = π h(π)x . Consequently, π x = π h(π)x = h(π)x . B 2.1.3. In the sequel, by a Boolean algebra of projections in a vector space X we mean a set B of commuting idempotent linear operators that act in X. Moreover, the Boolean operations have the following form: π ∧ ρ := π ◦ ρ = ρ ◦ π,

π ∨ ρ = π + ρ − π ◦ ρ,

π ∗ = IX − π

(π, ρ ∈ B),

and the zero and identity operators in X serve as the zero and unity of the Boolean algebra B. ⊥⊥

Suppose that E0 := X is a vector lattice with the projection property and the space X is d-decomposable. Then X have the projection property. Moreover, there exists a complete Boolean algebra B of band projections in X and an isomorphism h from P(E0 ) onto B such that
b x = h(b)x b ∈ P(E0 ), x ∈ X . C A nonzero band L ∈ B(E0 ) cannot be disjoint from the set X . Hence, for some x ∈ X, we have x ∈ / L⊥ . If π is the projection onto L then the element π x differs from zero. Due to d-decomposability of X, we have x0 = π x ∈ L for some x0 ∈ X. Thus, we may apply 2.1.2 (1, 4). To every band K ∈ B(X) there corresponds the projection πK along K ⊥ . Assign B := {πK : K ∈ B(X)}. It is clear that B is a complete Boolean algebra of projections. With an order projection ρ ∈ P(E0 ), we associate the projection πK with K := h(ρE0 ). Denote the mapping thus obtained by the same letter h. Then h is an isomorphism of the Boolean algebras P(E0 ) and B. The required property of the isomorphism h follows from (3). B We identify the
Boolean algebras P(E0 ) and P(X) := B and write π x = πx x ∈ X, π ∈ P(E0 ) .

48

Chapter 2

2.1.4. Suppose that X is a vector lattice. The norm · is monotone if |x| ≤ |y| implies x ≤ y for x, y ∈ X. We say that · is order semicontinuous (osemicontinuous for short) if sup xα = sup xα for each increasing net (xα ) ⊂ X with a least upper bound x ∈ X and order continuous (o-continuous for short) if inf xα = 0 for any decreasing net (xα ) ⊂ X with inf α xα = 0. If X is a vector lattice with a monotone norm then P(X) consists of positive projections. If, in addition, X is order complete and the norm is order semicontinuous then P(X) is a complete subalgebra of the Boolean algebra P(X). C For a vector lattice with a monotone norm we have obviously that h(π) is an order ideal in X for any π ∈ P(E0 ). If 0 ≤ x ∈ h(π) and 0 ≤ y ∈ h(π ⊥ ) then x ∧ y ∈ π(E) ∩ π ⊥ (E) = {0} by monotonicity. Hence, x ∧ y = 0 and x and y are disjoint not only in the sense of Subsection 2.1.2 but also in the sense of the order relation in X. Therefore the projection operator corresponding to the decomposition X = h(π) ⊕ h(π ⊥ ) is positive. Now, assume that X is order complete and · is order semicontinuous. It suffices to prove that h(π) is order-closed in X. Let (xα ) ⊂ h(π)+ be a bounded increasing net. Since the net xα is bounded in E, by order semicontinuity of the norm we have sup xα = sup xα ∈ π(E). Therefore sup xα ∈ h(π) and h(π) is a band. B 2.1.5. We say that a net (xα )α∈A bo-converges to an element x ∈ X and write x = bo-lim xα if there exists a decreasing net (eγ )γ∈€ in E such that inf γ∈€ eγ = 0 and, for every γ ∈ € , there is an index α(γ) ∈ A such that x − xα ≤ eγ for all α ≥ α(γ). Given an element e ∈ E+ , let the following condition be satisfied: for every number ε > 0, there is an index α(ε) ∈ A such that x − xα ≤ εe for all α ≥ α(ε). Then we say that (xα ) br-converges to x and write x = br-lim xα . A net (xα ) is said to be bo-fundamental (br-fundamental ) if the net (xα − xβ )(α,β)∈A×A bo-converges (br-converges) to zero. A lattice-normed space is called bo-complete (br-complete) if every bo-fundamental (br-fundamental) net in it bo-converges (brconverges) to an element of the space. Take a family (xξ )ξ∈„ and associate with itP the net (yα )α∈A , where A := Pfin („) is the set of all finite subsets of „ and yα := ξ∈α xξ . If x := bo-lim yα exists then the family (x Pξ ) is said to be bo-summable and x is its sum. It is conventional to write x = bo- ξ∈„ xξ in this case. Sometimes, if X is not a vector lattice, we write o and r instead of bo and br in the terms like o-complete. A set M ⊂ X is called norm-bounded if there exists an e ∈ E+ such that x ≤ e for all x ∈ M . A space X is called disjointly complete or d-complete if every norm-bounded set in X of pairwise disjoint elements is bo-summable. 2.1.6. A subspace X0 ⊂ X is called a bo-ideal if, for x ∈ X and x0 ∈ X0 , from x ≤ x0 it follows that x ∈ X0 . In this subsection we assume that X is

Lattice-Normed Spaces

49

decomposable. (1) A subspace X0 ⊂ X is a bo-ideal if and only if X0 = h(L) for some o-ideal L ⊂ E. C Sufficiency is obvious. In order to prove necessity, we take an arbitrary oideal X0 ⊂ X. Let L be the bo-ideal generated by the set X0 . It is clear that X0 ⊂ h(L). If x ∈ h(L) then x ≤ u1 + · · · + un for suitable u1 , . . . , un ∈ X0 . Due to decomposability of X, we have the representation x = x1 + · · · + xn , where xk ≤ uk (k := 1, . . . , n). By the definition of bo-ideal, xk ∈ X0 ; hence, x ∈ X0 . Thus, X0 = h(L). B (2) The set h(L) minorizes X particular, 2.1.2 (1) holds.

⊥⊥

∩ L for every o-ideal L ⊂ E. In

⊥⊥

C Indeed, if 0 < e ∈ X ∩ L then there exists an element 0 < e0 ≤ e such that e0 ≤ x1 + · · · + xn for some x1 , . . . , xn ∈ X. Due to decomposability of X, the representation e0 = u1 + · · · + un holds, where u1 , . . . , un ∈ X. For at least one of the numbers k, we have uk 6= 0. Since e ∈ L and uk ≤ e, it follows that uk ∈ h(L). B (3) An order ideal K ⊂ X is a band if and only if it is bo-closed. C Necessity is obvious. Suppose that K is a closed bo-ideal. According to (1), ⊥⊥ we have K = h(L), where L is a bo-ideal in E. Demonstrate that K = h(L ).  ⊥⊥ If x ∈ h(L ) then x = sup E , where E := e ∈ L : 0 ≤ e ≤ x . By decomposability of X, we may choose an xe so that x − xe = x − e and xe = e. The set E is directed upward and the net (xe )e∈E o-converges to x. Consequently, x ∈ K and K ⊃ h(L⊥⊥ ). The reverse inclusion is obvious. B 2.1.7. Under the assumption that (X, E) is a decomposable LNS, we establish three auxiliary facts. (1) Suppose that an element x ∈ X and an increasing sequence (an ) ⊂ E+ are given and, moreover, (an ) ≤ x (n ∈ N). Then there exists a sequence (xn ) ⊂ X such that, for all n ∈ N and m > n, we have xn = an ,

x − xn = x − an ,

xm − xn = am − an .

C Denote bn := x − an (n ∈ N), b0 := x . By decomposability of X, we may choose sequences (un ) ⊂ X and (vn ) ⊂ X so that x = u1 + v 1 ,

u1 = a1 ,

v 1 = b1 ,

un+1 + vn+1 = vn , vn+1 = bn+1 , un+1 = bn − bn+1 . Pn Assign xn = k=1 uk . Then x = xn + vn and the following relations hold: xn ≤

n X k=1

uk =

n X k=1

(bk−1 − bk ) = b0 − bn = an .

50

Chapter 2

At the same time, x ≤ xn + vn ≤ an + bn = x ; consequently, xn = an . Thus, for m > n, we have xm − xn =

m X k=n+1

uk ≤

m X

uk = am − an ≤ xm − xn . B

k=n+1

(2) If (X, E) is d-complete then it has the projection property. In particular, every d-complete Archimedean vector lattice has the projection property. C Given y ∈ X, denote  by E(y) the o-ideal in E that is generated by the ele-
ment y . Assign X(y) := x ∈ X : x ∈ E(y) . It is easy to see that X(y), E(y) is a decomposable LNS. If K is a band in X and K(y) := K ∩ E(y) then projections of y onto K and K(y) exist simultaneously and they are equal. At the same time E(y) is disjointly complete. Indeed, for an arbitrary family (eξ ) that is bounded in E(y) and consists of disjoint elements, there is a family P P (xξ ) ⊂ X such that xξ = eξ . Therefore, for x = bo- xξ , we have x = o- xξ = sup eξ ∈ E(y). Thus, in view of 2.1.2 (3), it remains to apply 1.5.11. B ⊥⊥

(3) If (X, E) is d-complete and r-complete then E0 := X is a Kspace and X = E0+ . ^ be the Dedekind C According to (2) X has the projection property. Let E ^ +, completion of E. Since E(y) is d-complete, we deduce that, for every e ∈ E(y) there is an increasing sequence en ∈ E(y)+ that r-converges to e. Using (1), choose a sequence (xn ) ⊂ X such that xn = en and xm − xn ≤ em − en (m > n). It is clear that the sequence (xn ) is r-fundamental and, due to r-completeness of X, x = r-lim xn ∈ X exists. Moreover, x = r-lim en = e. Consequently, E(y) = ^ E(y). Next, let e be an arbitrary positive element in the Dedekind completion ⊥⊥ of the lattice X . Then, for every maximal disjoint family (eξ ) ⊂ X , we ^ so that π ξ e ≤ neξ , may choose a partition (πnξ ) of the identity projection in E n ξ ξ i.e., πn e ∈ X . Therefore, there is a family (xn ) ⊂ X and an x ∈ X such that ⊥⊥ xξn = πnξ e ≤ e and x = e. Consequently, E0 = X is a K-space and X = E0+ . B 2.1.8. Let A be a sublattice and subring in Orth(E) containing P(E). We say that a lattice-normed space X over E admits a compatible module structure over A if X can be endowed with a structure of a faithful unitary A-module such that: (a) the natural representation of A in X defines an isomorphism of the Boolean algebras P(E) and P(X) from 2.1.3; (b) ax = |a| x (a ∈ A, x ∈ X). In the case when X is a vector lattice we assume, in addition to (a) and (b), the following:

Lattice-Normed Spaces

51

(c) B(X) is a complete subalgebra of the Boolean algebra of bands B(X). Let X be a d-decomposable lattice-normed space over a vector lattice E with ⊥⊥ and let A be a sublattice and subring in Orth(E) containing P(E). E = X Each of the following assertions implies that X admits a compatible module structure over A: (1) A is the algebra of finite rank elements; (2) E is order σ-complete, A := Z (E), and X is br-complete; (3) E is order σ-complete, A = Orth(E), and X is sequentially bocomplete; (4) E is order complete, A = Orth(E), and X is a vector lattice with an order semicontinuous monotone norm. P C Let a ∈ A be a finite rank element, i.e. a = λk πk where λ1 , . . P . , λn ∈ R and π1 , . . . , πn is a finite partition of unity in P(E). Then we put ax := λk πk x. Taking into account 2.1.2 (2), 2.1.3, and the identification of the Boolean algebras P(E0 ) and P(X), ax =

X

λ k πk x =

X

|λk |πk x = a x .

Now an arbitrary a ∈ A is the order limit of an increasing sequence of finite rank elements (an ) ⊂ A. The sequence (an x) ⊂ X is bo-fundamental, since an x−am x = |an − am | x → 0. Therefore, we may assign ax := bo-lim an x. Moreover, ax = bo-lim an x = o-lim |an | x = a x . The remaining part of the proof is straightforward. B 2.1.9. Let X be a lattice-normed vector lattice over E. In this case we have two different disjointness relations in X, one induced by vector norm as defined in 2.1.2 and the other determined by the order relation in X, see 1.3.3. They will be referred to, if need be, as order disjointness and norm disjointness, respectively. Vn An elementPx ∈ X is said to be norm-n-decomposable if k=0 xk = 0 whenn ever |x| = k=0 xk and x0 , x1 , . . . , xn are pairwise order disjoint elements in X+ . A norm-1-decomposable element is called norm-indecomposable. If E is a vector lattice with the principal projection property then we may state the following equivalent definition: x is norm-n-decomposable if and only if for every collection of pairwise order disjoint elements x0 , x1 , . . . , xn ∈ X+ with x = x0 + x1 + · · · + xn there exists a partition of unity π0 . . . πn in P(F ) such that πk xk = 0 for all k = 0, 1, . . . , n. Indeed, e0 ∧ · · · ∧ en = 0 is equivalent to πk ek = 0 (∀k = 0, 1, . . . , n) for a suitable partition of unity π0 . . . πn in P(F ).

52

Chapter 2

In the case when E admits compatible module structure over Orth(E) we may give one more equivalent definition by replacing the partition of unity with the collection of positive orthomorphisms σ0 , σ1 , . . . , σn such that σ0 +σ1 +· · ·+σn = IF . (1) The sum of n norm-indecomposable elements is norm-n-decomposable element. C Let x = u1 + · · · + un where u1 , . . . , un are norm-indecomposable elements in X. Without loss of generality, we may assume that uk ≥ 0 (k := 1, . . . , n). Take pairwise order disjoint elements x0 , x1 , . . . , xn ∈ X with x0 + x1 + · · · + xn = u1 + · · · + un . Using the Riesz Decomposition Property 1.3.2 (13), choose uk,l ∈ X+ such that uk =

n X

uk,l (k := 1, . . . , n),

xl =

n X

uk,l (l := 0, 1, . . . , n).

k=0

l=1

From 1.3.2 (15) we deduce

0≤

n ^ l=0

≤

X

xl =

n X n ^ l=0 k=1

uk,l ≤

n X n ^

uk,l

l=0 k=1

uj(0),0 ∧ uj(1),1 ∧ · · · ∧ uj(n),n

j∈J

where J is the set of all functions j : {0, 1, . . . , m} → {1, . . . , m}. At least two indices in {j(0), . . . , j(n)} are equal, say j(r) = j(s) = m, 0 ≤ r, s ≤ n, r 6= s. Therefore, uj(0),0 ∧ uj(1),1 ∧ · · · ∧ uj(n),n ≤ um,r ∧ um,s = 0. The latter follows from the relations 0 ≤ um,r ≤ xr , 0 ≤ um,s ≤ xs , um = um,0 + · · · + um,n , since xr and xs are order disjoint and um is norm-indecomposable. B (2) Let E be a vector lattice with the principal projection property and X be a d-decomposable lattice-normed vector lattice over E. Let x ∈ X be a positive norm-n-decomposable element and x = x0 + x1 + · · · + xn−1 for some collection of pairwise disjoint x0 . . . xn−1 . If π is a band projection in E with  ⊥⊥ then y := π(x0 ) is norm-indecomposable and π(E) ⊂ x0 ∧ x1 ∧ · · · ∧ xn−1 z := π(x − x0 ) norm-(n − 1)-decomposable; moreover, y and z are disjoint fragments of x. C Let 0 ≤ y1 ≤ y, 0 ≤ y2 ≤ y, y1 ∧ y2 = 0, and y1 + y2 = y. Prove that y1 ∧ y2 = 0. Let πk be the band projection onto { xk }⊥⊥ and put π :=

Lattice-Normed Spaces

53

π0 π1 . . . πn−1 . Then n + 1 elements πx1 , . . . , πxn−1 , y1 , y2 are pairwise disjoint and their supremum is equal to πx1 + πx2 + · · · + πxn−1 + y1 + y2 = π(x − x0 ) + π(x0 ) = πx. Since any fragment of a norm-n-decomposable element is also norm-n-decomposable we obtain πx1 ∧
· · · ∧ πxn−1 ∧ y1 ∧ y2 = 0, whence πe = 0, e := π x1 ∧ · · · ∧ xn−1 ∧ y1 ∧ y2 . At the same time π ⊥ e = 0, since π ⊥ e ≤ π ⊥ y1 ≤ π ⊥ πx0 = 0, so that e = 0. It follows from this that π( y1 ∧ y2 ) = 0, because π is the band
⊥⊥ projection onto x0 ∧ · · · ∧ xn−1 . Thus, y1 ∧ y2 = π ⊥ ( y1 ∧ y2 ) ≤ π ⊥ πx0 = 0. Clearly y is a fragment of x. Suppose now z = z1 + · · · + zn and zk ⊥ xl (k 6= l). Then πx is the sum of n + 1 pairwise disjoint collection πx0 , z1 , , . . . , zn and πx0 ∧ z1 ∧· · ·∧ zn = 0, since πx is norm-n-decomposable. It follows from this that πe = 0, π( x0 ∧ z1 ∧· · ·∧ zn ) = 0. At the same time π ⊥ e = 0, so that e = 0. Consequently, π( z1 ∧ · · · ∧ zn ) = 0 and z1 ∧ · · · ∧ zn = 0. B 2.1.10. Theorem. Let E be a vector lattice with the principal projection property and X be a d-complete lattice-normed vector lattice over E. Then every norm-n-decomposable element x ∈ X there exist n norm-indecomposable elements x1 , . . . , xn such that x = x1 + · · · + xn . C The proof is by induction on n. The case n = 1 is a tautology. Suppose the claim is true for n − 1. By the Kuratowski–Zorn Lemma there is a maximal set P of pairwise disjoint band projections in E such that for each π ∈ P there Pn−1 exist n pairwise disjoint elements x0 , x1 , . . . , xn−1 ∈ X such that i=0 xi = x and π(E) ⊂ ( x0 ∧ x1 ∧ · · · ∧ xn−1 )⊥⊥ . Using 2.1.9 (2), we may construct a functions π 7→ xπ defined on P such that xπ is positive and norm-indecomposable, yπ := πx − xπ is positive norm-(n − 1)decomposable, xπ = πxπ , and xπ and yπ are disjoint fragments of W x. Since the family (xπ )π∈P is disjoint and X is d-complete we may define y := {xπ : π ∈ P} and z := x − y. Observe that y and z are disjoint fragments of x. By definition πy = xπ for every π ∈ P, so that, setting ρ := sup P, we obtain ρy = y and ρ⊥ y = 0. If y = y1 + y2 with y1 ∧ y2 = 0 then πy1 + πy2 = xπ for π ∈ P and π( y1 ∧ y2 ) = 0, since xπ is norm-indecomposable; thus ρ( y1 ∧ y2 ) = 0. But ρ⊥ ( y1 ∧ y2 ) ≤ ρ y = 0, whence y1 ∧ y2 = 0. Assume now that z = z0 +z1 +· · ·+ zn−1 for a disjoint collection z0 , z1 , . . . , zn−1 ∈ X+ and put e := z0 ∧ · · · ∧ zn−1 ). Then πe for every π ∈ P, because πz = yπ is n−1-decomposable. At the same time x = (z0 + y) + z1 + · · · + zn−1 and, taking into consideration the definition of P, we obtain π ⊥ ( z0 ∧ · · · ∧ zn−1 ) ≤ π ⊥ ( z0 + y ∧ z1 ∧ · · · ∧ zn−1 ) = 0. We conclude Vn−1 that i=0 zi = 0. Thus, we have proved that y is norm-indecomposable, z is normn − 1-decomposable and x = y + z and we are done in the case of positive x. For an

54

Chapter 2

arbitrary norm-n-indecomposable x the modulus |x| is also norm-n-indecomposable and by above proved fact |x| = y1 + · · · + yn with yk norm-indecomposable. In view of the Riesz Decomposition Property x+ = u1 + · · · + un , x− = v1 + · · · + vn , uk + vk = yk (uk , vk ∈ X; k := 1, . . . , n). The element xk := uk − vk is normindecomposable, since yk = |xk | and x = x1 + · · · + xn . B 2.2. Completion Here we briefly consider the questions of completing lattice-normed spaces. 2.2.1. Every decomposable o-complete lattice-normed space is called a Banach–Kantorovich space (a BKS for short). If a Banach–Kantorovich space is in addition a vector lattice and the norm is monotone then it is called a Banach– Kantorovich lattice. Let (X, E) be a Banach–Kantorovich space and, moreover, ⊥⊥ . According to 2.1.2 (2), the Boolean E= X
algebras P(E) and B(X) can be identified and π x = πx π ∈ P(E), x ∈ X . For every bounded family P (xξ )ξ∈„ in X and every partition of unity (πξ )ξ∈„ in P(X), the sum x := o- ξ∈„ πξ xξ exists. Moreover, x is a unique element in X satisfying the relations πξ x = πξ xξ (ξ ∈ „). C If e := sup xξ then, for α, β ∈ Pfin („), we have  X  X πξ xξ ≤ yα − yβ = πξ e ≤ e, ξ∈α4β

ξ∈α4β

P

where yγ := ξ∈γ πξ xξ and α4β is the symmetric difference of the sets α and β. Hence it is clear that the net (yα ) is bo-fundamental; hence, x := bo-lim yα exists. B This proposition, in particular, ensures d-completeness of all BKSs. Moreover, the definitions imply readily that every BKS is br-complete as well. Thus, an arbitrary BKS (X, E) is d-complete and br-complete; therefore, by 2.1.7 (3) we have ⊥⊥ and E+ = X . E= X 2.2.2. Let B be a complete Boolean algebra and let A be an arbitrary nonempty set. Assign n
B(A) := ν : A → B : (∀α, β ∈ A) α 6= β ⇒ ν(α) ∧ ν(β) = 0 o _ ∧ ν(α) = 1 . α∈A

Thus, B(A) is the set of all partitions of unity in B that are indexed by elements of the set A. If A is an ordered set then we may order the set B(A) as well:

ν ≤ µ ⇔ (∀α, β ∈ A) ν(α) ∧ µ(β) 6= 0 ⇒ α ≤ β ν, µ ∈ B(A) .

Lattice-Normed Spaces

55

It is easy to show that this relation is actually a partial order in B(A). If A is directed upward (downward) then so does B(A). Let Q be the Stone space of the algebra B. Identifying an element  ν(α) with a clopen subset of Q, we construct the mapping ν¯ : Qν → A, Qν := ∪ ν(α) : α ∈ A , by letting ν¯(q) = α whenever takes the value α on ν(α). Moreover, q ∈ ν(α). Thus, ν¯ is a step-function that
¯(q) . We shall use the set B(A) in the proof of ν ≤ µ ⇒ (∀q ∈ Qν ∩ Qµ ) ν¯(q) ≤ µ the following important completeness criterion. 2.2.3. Theorem. A decomposable lattice-normed space is order complete if and only if it is disjointly complete and complete with respect to relative uniform convergence. C Necessity was noted in 2.2.1. We will prove sufficiency. Suppose that a decomposable lattice-normed space X is d-complete and r-complete. According to ⊥⊥ 2.1.7 (3), we may assume without loss of generality that E := X is a K-space. Take an o-fundamental net P (xα )α∈A ⊂ X. Let B := P(X) be the base of X. Given ν ∈ B(A), assign xν := α∈A ν(α)xα . The element xν is well defined due to the d-completeness of X. A new net (xν )ν∈B(A) thus appeared. Show that the net is r-fundamental. Choose a decreasing net (eα )α∈A ⊂ E so that inf eα = 0 and xα − xβ ≤ eγ ≤ e whenever α, β ≥ γ (α, β, γ ∈ A), where e ∈ E. Take an arbitrary number ε > 0. There exist a partition of unity (ρξ )ξ∈„ in the algebra B and a mapping ϕ : „ → A such that ρξ eϕ(ξ) ≤ εe (ξ ∈ „). This is easily  deduced from the properties of o-convergence in a K-space. Assign πα := sup ρξ : ϕ(ξ) = α, ξ ∈ „ and πα = 0 in case α ∈ / im ϕ. Then (πα )α∈A is a partition of unity in B and πα eα ≤ εe (α ∈ A). Denote this partition of unity by ν(ε). Show that xν 0 − xν ≤ εe whenever ν, ν 0 ≥ ν(ε). Let ν = (ρα )α∈A and ν 0 = (τα )α∈A . If πα ρβ τγ 6= 0 then β, γ ≥ α; consequently, πα ρβ τγ xν − xν 0 = πα ρβ τγ xγ − ρβ τγ xβ = πα ρβ τγ xγ − xβ ≤ πα eα ≤ εe. Summing up over α, β, and γ, we find xν − xν 0 ≤ εe for ν, ν 0 ≥ ν(ε). Thus, the net (xν ) is r-fundamental and, due to the r-completeness, x := r-lim xν exists. It is clear that x − xν ≤ εe whenever ν ≥ ν(ε). For any fixed index γ ∈ A, construct a special partition of unity ν := (ρα )α∈A by letting ρα = πα in case α 6= γ and α 6= β, ργ = 0, and ρβ = πγ ∨ πβ , where β ∈ A and β ≥ γ. Then ν ≥ ν(ε) and πγ x − xβ = πγ ρβ x − xν = πγ x − xν ≤ εe.  Thus, πγ x − xβ ≤ εe for all β, γ ∈ A, β ≥ γ. Put cγ := sup x − xβ : β ≥ γ and c := inf{cγ : γ ∈ A}. Then πγ c ≤ πγ cγ ≤ εe. Since γ ∈ A and ε > 0 are arbitrary, we have c = 0; hence, o- lim x − xα = 0. B

56

Chapter 2

2.2.4. As before, let mE be a universal completion of a vector lattice E; moreover, let a unity 1 be fixed together with the respective multiplication structure ⊥⊥ = E. The operator-dual in mE. Take a lattice-normed space X with X ∗ ∗ space X is defined as follows. An operator x : X → mE belongs to X ∗ if and only if there exists an element 0 ≤ c ∈ mE such that hx, x∗ i := x∗ (x) ≤ c x

(x ∈ X).

The least element 0 ≤ c ∈ mE satisfying the indicated relation exists. This element is denoted by x∗ . It is easy to see that the mapping x∗ 7→ x∗ is an mE-valued norm in X ∗ and the following inequality holds: hx, x∗ i ≤ x x∗

(x ∈ X).

(1) If x∗ ∈ X ∗ then πx∗ = π x∗ for every projection π ∈ P(mE). C From the inequality hx, x∗ i ≤ c x it follows readily that hx, π
x∗ i ≤ πc x ; therefore, πx∗ ≤ π x∗ . On the other hand, hx, x∗ i ≤ πx∗ + π ⊥ x∗ x for all x ∈ X; hence, x∗ ≤ πx∗ + π ⊥ x∗ ≤ πx∗ + π ⊥ x∗ . Hence we deduce π x∗ ≤ πx∗ . B (2) The space X ∗ is d-complete and r-complete. C Take a partition of unity (πξ ) in P(mE), a family (x∗ξ ) ⊂ X ∗ ,P and an element ∗ x ∈ X. K-space there exist hx, x i := πξ hx, x∗ξ i and P In a∗ universally complete ∗ ∗ x i ≤ e x . Hence it is clear that x ∈ X ∗ . Since e := πξ xξ ; moreover, hx, P πξ x∗ = πξ x∗ξ , we have x∗ = πξ x∗ξ in view of (1). Thus, d-completeness is proven. Completeness with respect to relative uniform convergence is established below in a more general situation (see 4.2.1). B (3) The space X ∗ is a universally complete BKS. C Decomposability of X ∗ follows from (1) due to r-completeness, and it remains to refer to Theorem 2.2.3. B 2.2.5. According to 2.2.4 (3), the second operator-dual space, X ∗∗ , is a universally complete BKS. The canonical embedding κ : X → X ∗∗ is defined, as usual, by the formula hx∗ , κxi := κx(x∗ ) = hx, x∗ i (x∗ ∈ X ∗ ). The canonical embedding κ is a linear isometry. C We only need to show that κ preserves the norm, i.e., κx = x (x ∈ X). Observe first that κx ≤ x , since hx∗ , κxi ≤ x x∗ (x∗ ∈ X ∗ ). Next, since the operator · is sublinear, therefore, by the Hahn–Banach–Kantorovich Theorem, there exists an operator x∗ : X → mE such that x = hx, x∗ i and hy, x∗ i ≤ y (y ∈ X). Hence it is clear that x∗ ≤ 1; therefore, x = hx∗ , κxi ≤ κx x∗ ≤ κx . B

Lattice-Normed Spaces

57

2.2.6. By a universal completion (an order completion or, in short, a bocompletion) of a lattice-normed space
(X, E) we mean a universally complete BKS (Y, mE) respectively, a BKS (Y, oE) together with a linear isometry ı : X → Y such that each universally complete o-complete subspace of (Y, mE) respectively,
any decomposable o-complete subspace of (Y, oE) containing ıX coincides with Y . Here oE is a Dedekind completion of the vector lattice E, and mE is, as before, a universal completion of oE; moreover, we assume E ⊂ oE ⊂ mE, see 1.1.8. Given a set U in a lattice-normed space Y , we assign  r(U ) := y = br-lim yn : (yn )n∈N ⊂ U , n→∞  o(U ) := y = bo-lim yα : (yα )α∈A ⊂ U , n o X d(U ) := y = boπξ yξ : (yξ )ξ∈„ ⊂ U , ξ∈„

where A is an arbitrary directed set, (πξ ) is an arbitrary partition of unity in P(Y ), and the limits and the sums exist in Y . Let r0 (U ) be the part of r(U ) containing the limits of sequences in U converging with regulator 1, and let d0 (U ) be the part of d(U ) containing the finite sums. Observe the following simple relations: d0 (d0 (U )) = d0 (U ), d(d0 (U )) = d(U ). Now denote mX := rd(κX), where κ is the canonical embedding X → X ∗∗ of 2.2.5 and the operations d and r are calculated in the universally complete BKS (X ∗∗ , mE). 2.2.7. If Y is a decomposable BKS then, for any U ⊂ Y , the following hold: (1) dd(U ) = d(U ); (2) dr0 d(U ) = r0 d(U ); (3) rd(U ) = r0 d(U ); (4) rrd(U ) = rd(U ). C In each of the desired equalities, the inclusion ⊃ is obvious; therefore, the reasoning below concerns the reverse inclusion. (1): Take a family (yξ )ξ∈„ in d(U ) and a partitionPof unity (πξ )ξ∈„ in the algebra P(Y ). Represent an element yξ ∈ d(U ) as yξ = u∈U πu,ξ u, where (πu,ξ )u∈U P is a partition of unity in P(Y ) for each ξ ∈ „. Assign y := ξ∈„ πξ yξ and π(ξ,u) := πξ ◦ πu,ξ (ξ ∈ „, u ∈ U ). If (ξ, u) 6= (η, v) then π(ξ,u) ◦ π(η,v) = πξ ◦ πη ◦ πu,ξ ◦ πv,η = 0. At the same time, _ (ξ,u)∈„×U

π(ξ,u) =

_ ξ∈„

πξ ◦

_ n∈U

 πu,ξ = 1;

58

Chapter 2

consequently, (πλ )λ∈„×U is a partition of unity in P(Y ). Now observe that π(ξ,u) y = πu,ξ (πξ y) = πu,ξ (πξ yξ ) = πξ (πu,ξ u) = π(ξ,u) u. P Thus, y = (ξ,u) π(ξ,u) u and y ∈ d(U ). (2): Now let the family (yξ ) be included in r0 d(U ). Fix a number P ε > 0 and, for each ξ ∈ „, choose a uξ ∈ d(U ) so that yξ − uξ ≤ ε1. If u = πξ uξ and y is as above, then y − u ≤ ε1. However, due to (1), u ∈ d(U ); therefore, y ∈ r0 d(U ). (3): Suppose that a sequence (yn )n∈N in d(U ) converges to an element y ∈ Y with regulator e ∈ mE. Choose a partition of unity (πξ )ξ∈„ in P(mE) so that πξ e ≤ λξ 1 for suitable λξ > 0. Take an arbitrary ε > 0. For each P ξ ∈ „, there exists an index n(ξ) ∈ N such that πξ y − yn(ξ) ≤ ε1. Assign u = ξ∈„ πξ yn(ξ) and observe that u ∈ d(U ) due to (1). It remains to take account of the inequality y − u ≤ ε1, and we arrive at the conclusion y ∈ r0 d(U ). (4): First, note that the operation r0 is a topological closure and, therefore, r0 r0 (A) = r0 (A) for each A ⊂ Y . However, the latter can be easily proven directly. Next, applying (2) and (3), we may write rrd(U ) = rr0 d(U ) = rdr0 d(U ) = r0 dr0 d(U ) = r0 r0 d(U ) = r0 d(U ) = rd(U ). B 2.2.8. Theorem. Each lattice-normed space has a universal completion that is unique to within linear isometry. The space (mX, mE) serves as a universal completion for (X, E) and is referred to as the universal completion of (X, E) in the sequel. C The d-completeness of the space mX follows from 2.2.7 (2, 3), and its rcompleteness is clear from 2.2.7 (4). By Theorem 2.2.3, mX is a BKS. Moreover, mX is universally complete, since the operation d in the definition of mX is calculated in the universally complete space (X ∗∗ , mE), see 2.2.6. The embedding ı := κ : X → mX is a linear isometry (2.2.5). If Y is a decomposable BKS and ıX ⊂ Y ⊂ mX, then the values of the operations d and r are always in Y ; therefore, mX = rd(ıX) ⊂ rd(Y ) ⊂ Y ⊂ mX, i.e., Y = mX. Thus, mX is a universal completion for the space X. Assume that Y 0 is one more universal completion for X and let ı0 : X → Y 0 be the corresponding isometric embedding. The operator h := ı0 ◦ ı−1 : ıX → Y 0 is a linear isometry. We may extend h onto d(ıX) and next onto rd(ıX) preserving linearity and isometry. Moreover, h(mX) is a universally complete BKS and ı0 X ⊂ h(mX) ⊂ Y 0 ; consequently, h(mX) = Y 0 . B

Lattice-Normed Spaces

59

2.2.9. In this subsection X is a d-decomposable lattice-normed space over a vector lattice E possessing the principal projection property. (1) Let U ⊂ X and d0 (U ) = U . Then for every x ∈ X there exists
a net (uα )α∈A in U such that the net x − uα α∈A decreases and { x − uα : α ∈ A} = { x − u : u ∈ U }. In particular, o-limα∈A x − uα = inf u∈U x − u . C Take x ∈ X and introduce some equivalence relation and preorder in U by the following formulas: u∼v ⇔ x−u = x−v , u4v ⇔ x−u ≥ x−v . Since E is a lattice with the principal projection property we may choose a projection π ∈ P(X) such that π x − u + π ⊥ x − v = x − u ∧ x − v . The element w := πu + π ⊥ v lies in U since d0 (U ) = U . Moreover, x−w = x−u ∧ x−v , whence u 4 w and v 4 w. Thus, the preordered set (U, 4) is directed upward. Hence it is clear that the factor set A := U/∼, endowed with the factor order, is an upward-directed ordered set. Now, the desired net (uα )α∈A is defined by choosing uα ∈ α (α ∈ A). B We say that a set U ⊂ X approximates an element x ∈ X if inf u∈U x − u = 0 and that U approximates a set V ⊂ X if U approximates every element of V . In case if U approximates each x ∈ X we say that U is an approximating set. (2) If V approximates U and W approximates V , then W approximates U . C For an arbitrary u ∈ U , denote e := inf w∈W w − u and suppose that e 6= 0. Since inf v∈V v − u = 0, there exist v ∈ V and a band projection ρ such that ρ v − u < ρe/2. By the same arguments, inf w∈W w − v = 0 and there are w ∈ W and a band projection π with π w−v < πρe/2. Thus we arrive at the contradiction πρe 6 πρ w − u 6 πρ w − v + πρ v − u < πρe/2 + πρe/2 = πρe. B (3) Let U ⊂ and x ∈ X. Then U approximates x if and only if x is the bo-limit of some net of elements of d0 (U ).  C Indeed, if U approximates x then inf x − u : u ∈ d0 (U ) = 0. Thus, in view of (1) there is a net (uα )α∈A in d0 (U ) such that o-lim x − uα = 0. Conversely, if x is a bo-limit of a net containing in d0 (U ) then d0 (U ) approximates x. It remains to apply (2), since U approximates d0 (U ). B

60

Chapter 2

2.2.10. Assume that E is an order complete vector lattice, B is a complete Boolean algebra, and h : P(E) → B is a ring homomorphism. Say that a net (πα )α∈A in P(E) h-converges to zero and write h-limα∈A πα = 0 provided that o-limα∈A πα = 0 and o-limα∈A h(πα ) = 0 in the Boolean algebras P(E) and B respectively. We say that a net (eα )α∈A in E h-converges to e ∈ E and write h-limα∈A eα = e if the net (e  α )α∈A is bounded beginning with some member and h-limα∈A [d] (|eα − e| − d)+ = 0 for all positive d ∈ E. In this case, we call the element e the h-limit of the net (eα )α∈A . Finally, we say that a net (uα )α∈A in X h-converges to u ∈ X and write h-limα∈A uα = u if h-limα∈A uα −u = 0. In this case, we call the element u the h-limit of the net (uα )α∈A . For an arbitrary family  
+ (eξ )ξ∈„ in E+ , the notation h- inf ξ∈„ eξ = 0 means that inf ξ∈„ h [d] (eξ −d) =0 for all positive d ∈ E. (1) If h-limα∈A eα = e or h- inf ξ∈„ eξ = 0, then o-limα∈A eα = e or inf ξ∈„ eξ = 0, respectively. The converse is also true provided that the homomorphism h is o-continuous. Let U be a subset of an LNS X. We say that U h-approximates an element x ∈ X if h- inf u∈U x − u = 0. We say that U h-approximates a set V ⊂ X if U h-approximates every element of V . A subset of an LNS X is called happroximating if it h-approximates X. Every h-approximating set is approximating and, in case the homomorphism h is o-continuous, the notions of approximating and h-approximating set coincide. (2) Suppose that V h-approximates U and W h-approximates V . Then W h-approximates U . C Consider an arbitrary element u ∈ U , fix a
positive element d of the norm lattice, and assign b := inf w∈W h [d][( x − z − d)+ ] . By 2.2.9 (2), it is sufficient to establish the equality b = 0. For simplicity, we assume that h([d]) = 1. Suppose to the contrary that b 6= 0. Then, in view of inf v∈V h([( v − u − d/2)+ ]) = 0, there is an element v ∈ V such that b0 := b ∧ h([ v − u − d/2)+ ]) < b. Similarly, in view of the equality inf w∈W h([( w − v − d/2)+ ]) = 0, there is an element w ∈ W such that (b\b0 ) ∧ h([ w − x − d/2)+ ]) < (b\b0 ). It is easy to verify that w satisfies the inequality b ∧ h([( w − u − d)+ ]) < b, which contradicts the definition of b. B (3) A set U ⊂ X h-approximates an element x ∈ X if and only if x is the h-limit of some net in d0 (U ). C If U h-approximates x then, in
view of 2.2.9 (1), there exists a net (vα )α∈A in d0 (U ) such that the net x − vα α∈A decreases and { x − vα : α ∈ A} = { x − v : v ∈ d0 (U )}. It remains to observe that h-limα∈A x − vα = 0. Conversely, if x is the h-limit of a net in d0 (U ), then d0 (U ) h-approximates x. It remains to observe that U h-approximates d0 (U ) and to use (2). B

Lattice-Normed Spaces

61

2.2.11. (1) Theorem. For each lattice-normed space X, there exists a bocompletion which is unique to within linear isometry and is referred to as the bocompletion of X in the sequel. C Recall that E ⊂ oE ⊂ mE. Assign  ^ = x ∈ mX : x ∈ oE . X ^ is an order completion of the space X. B Then X We always assume that any lattice-normed space X is contained in its bo^ completion X. ^ of the space X, we have X ^ = rdX. If X (2) For the bo-completion X ⊥⊥ is decomposable and E0 := X is a vector lattice with the principal projection ^ property, then also X = oX. ^ ^ ∈ X C The first part of the assertion ensues from 2.2.7 and 2.2.8. Take x and choose a net (xα ) ⊂ X that bo-converges to x ^.
Apply Proposition 2.2.10 (1) ^ There exists a net xα in X such that the net with U := X and X := X. α∈A
x ^−xα α∈A decreases and e := inf x∈X x ^−x = inf α∈A x ^−xα , where the infimum ^ is calculated in oE. In view of the equality X = rdX, for an arbitrary number ε > 0,Pwe may find a family (xξ ) ⊂ X and a partition of unity (πξ ) ⊂ P(X) so that ^ − πξ xξ ≤ ε x . So, we may write x e = o-

X

X X ^ − xξ = x ^ − bo^. πξ e ≤ oπξ x πξ xξ ≤ ε x

^ = bo-lim xα . B Therefore, e = 0 and x 2.2.12. Let X be a vector lattice with a monotone norm. Then its norm ^ is a Banach–Kantorovich lattice and X is a sublattice in X. ^ completion X C Denote Y+ := rd(X+ ). Then Y+ is a cone. Indeed, Y+ is a wedge, since the operations r and d are linear. At the same time, if ±y ∈ Y+ then, by definition, for every 0 < ε ∈ R there are a partition of unity (πξ ) ⊂ P(E) and families (x0ξ ) and (x00ξ ) in X+ such that X y − boπξ x0ξ ≤ ε y ,

X y + boπξ x00ξ ≤ ε y .

From these inequalities we deduce X X 2y ≤ y − boπξ x0ξ + y + boπξ x00ξ ≤ 2ε y . Therefore, y = 0 and y = 0. Further details are obvious. B

62

Chapter 2

2.2.13. Let X be an arbitrary, not necessarily d-decomposable, lattice-normed space over an arbitrary vector lattice E. Suppose that a d-decomposable latticenormed space X over E includes X as a subspace with the induced norm. We say that X is a d-decomposable hull of X, if X = d0 (X), i.e., if every member of X is represented as a finite sum π1 x1 + · · · + πn xn , where x1 , . . . , xn ∈ X and pairwise disjoint π1 , . . . , πn ∈ P(E). (1) Suppose that a vector lattice E possesses the principal projection property. Then every (not necessarily d-decomposable) lattice-normed space over E has a d-decomposable hull that is unique to within an isometry. C Let mX be a universal completion of a lattice-normed space X. Define X Pn to be the set of all finite sums k=1 πk xk , where x1 , . . . , xn ∈ X and (πk )nk=1 is a partition of unity in the Boolean algebra P(E). It is easy to see that X is an LNS over E and is a d-decomposable hull of X. Uniqueness of this d-decomposable hull is obvious. B (2) We describe one more useful construction. Let E and F be K-spaces and let X be an LNS over
E. Suppose that a mapping S : E+ → F+ is subadditive
S(e1 + e2 ) 6 Se1 + Se2 , positively
homogeneous S(λe) = λSe (λ ∈ R+ ) , and increasing 0 6 e1 6 e2 ⇒ Se1 6 Se2 . Consider the vector subspace X0 := {x ∈ X : S( x ) = 0} and denote by SX x the coset in X/X0 containing an x ∈ X. It is easy to check that the space X/X0 is a lattice-normed space with respect to the norm SX x := S( x ). Observe that the LNS X/X0 need not be d-decomposable. We call a d-decomposable hull of the LNS X/X0 the norm transformation of X by means of S and denote it by SX. The linear operator SX : X → SX is called the operator of norm transformation of X by means of S. 2.3. Examples In this section we consider important examples of spaces of continuous, weakly continuous, measurable, and weakly measurable vector-functions that can be naturally normed by means of a vector lattice. 2.3.1. We begin with the simplest extreme cases, namely, vector lattices and normed spaces. If X = E then the modulus of an element can be taken as its vector norm: x := |x| = x ∨ (−x) (x ∈ E). Decomposability of this norm is easy from the Riesz Decomposition Property holding in every vector lattice. If E = R then X is a normed space. We shall use the conventional notation for the norm, k·k, and omit references to the order structure of the norm lattice. 2.3.2. Let Q be a topological space and let Y be a normed space. Let X := Cb (Q, Y ) be the space of bounded continuous vector-functions from Q into

Lattice-Normed Spaces

63

Y . Assign E := Cb (Q, R). Given f ∈ X, define its vector norm f by the relation f : t 7→ kf (t)k (t ∈ Q). Then · is a decomposable norm. Indeed, assume that f = h1 + h2 for some h1 , h2 ∈ E. Define a vector-function f1 : Q → Y so that f1 (t) = h1 (t)f (t)/kf (t)k whenever f (t) 6= 0 and f1 (t) = 0 whenever f (t) = 0. Then f1 ∈ X and f2 := f − f1 ∈ X; moreover, fk = hk (k := 1, 2). The space X is br-complete if and only if Y is a Banach space. 2.3.3. Suppose that E is an order-dense ideal of the universally complete vector lattice C∞ (Q), where Q is an extremal compact space. Let C∞ (Q, X) be the set of cosets of continuous vector-functions u that act from comeager subsets dom(u) ⊂ Q into some normed space X. Recall that a set is called comeager if its complement is meager. Vector-functions u and v are equivalent if u(t) = v(t) whenever t ∈ dom(u) ∩ dom(v). The set C∞ (Q, X) is endowed, in a natural way, with the structure of a module over C∞ (Q). Moreover, the continuous extension of the pointwise norm defines a decomposable norm on C∞ (Q, X) with values in C∞ (Q). Indeed, given any z ∈ C∞ (Q, X),
there exists a unique function xz ∈ C∞ (Q) such that ku(t)k = xz (t) t ∈ dom(u) , for every representative u of the coset z. Assign z := xz and  E(X) := z ∈ C∞ (Q, X) : z ∈ E . If X is a Banach space then E(X) presents a Banach–Kantorovich space with C∞ (Q, X) its universal completion. C The assertion can be easily deduced from 2.2.3. It also follows immediately from a more general fact which will be proved in the next section (see 2.4.6 and 2.4.7). B Denote by C# (Q, X) the part of C∞ (Q, Y ) consisting of cosets z with z ∈ C(Q). Thus, C# (Q, X) := E(X) where E := C(Q). Observe that C# (Q, X) is also a Banach–Kantorovich space, whereas the space C(Q, X) of everywhere defined continuous functions from Q to X, being a lattice-normed space over C(Q) (see 2.3.2) is not d-complete in general (see 2.4.9). Thus, C(Q, X) does not coincide with the space C# (Q, X), unless Q is finite or X is finite-dimensional. 2.3.4. For the same X and E define an embedding of the algebraic tensor Pn product E ⊗ X into the space E(X). Associate with an element z = k=1 ek ⊗ xk ∈ E ⊗ X the coset z¯ of the vector-function n n  n o X X q 7→ z¯(q) := ek (q)xk q ∈ Q(z) := s ∈ Q : |ek (s)| < +∞ . k=1

k=1

We also endow the space E ⊗ X with the norm n nX o z 0 := inf |ek |kxk k (z ∈ E ⊗ X), k=1

64

Chapter 2

Pn where the infimum is taken over all representations z = k=1 ek ⊗ xk . (1) The mapping ι : z 7→ z¯ from E ⊗ X into E(X) is a linear isometry. Pn C Clearly, ι is a linear embedding. If z = k=1 ek ⊗ xk then

n

n

X

X

z ≤ ek (·)xk ≤ |ek (·)|kxk k;

k=1

k=1

hence, z ≤ z 0 . For each P s ∈ Q(z) we may choose a function φs ∈ E such n 0 that φs (s) = 1. Put x := s k=1 ek (s)xk ∈ X and ek := ek − φs ek (s). Then Pn z = φs xs + k=1 e0k xk and for an arbitrary s ∈ Q(z) we have n   X z 0 (s) ≤ |φs |kxs k + |e0k |kxk k (s) k=1



= |φs (·)|kxs k +

n X

 |ek (·) − φs (·)ek (s)|kxk k (s)

k=1

X

n

= kz(s)k = z . B e (s)x = |φs (s)|kxs k = k k

k=1

Thus, E ⊗ X can be identified with a subspace of the LNS E(X). It turns out that E ⊗ X is, in a sense, dense in E(X).
(2) For every z ∈ E(X) and every ε > 0, there exist a family Q(ξ) ξ∈„ of pairwise disjoint clopen sets Q(ξ) ⊂ Q whose union is dense in Q and a family (zξ )ξ∈„ in E ⊗ X such that z|Q(ξ) − zξ |Q(ξ) ≤ ε z (ξ ∈ „).  ⊥⊥ C It is sufficient to consider the case in which z = E and the identical unity is contained in E. For every point q ∈ dom(z) ⊂ Q, choose a clopen neighborhood U (q) so that
kz(q) − z(q 0 )k ≤ εkz(q 0 )k q 0 ∈ U (q) . In the
Boolean algebra of clopen subsets of Q there exists a partition of unity Q(ξ) ξ∈„ which possesses the following property: for each ξ ∈ „, there is a point qξ ∈ Q such that Q(ξ) ⊂ U (qξ ). Now, if we assign zξ := χQ(ξ) z(qξ ) then
kz(q 0 ) − zξ (q 0 )k = kz(q 0 ) − z(qξ )k ≤ εkz(q 0 )k q 0 ∈ Q(ξ) . Thus, χQ(ξ) z − zξ ≤ ε z , which is equivalent to the desired inequality. B

Lattice-Normed Spaces

65

2.3.5. Now we turn to defining a space of weakly continuous vector-functions which is similar to E(X). We suppose that Z ⊂ X 0 is a norming subspace (for X), i.e., kxkX = sup{|hx, zi| : z ∈ Z, kzk ≤ 1} (x ∈ X). Here, as usual, X 0 is the dual space and h· , ·i is the canonical bilinear form of the duality X ↔ X 0 . Denote by M the set of σ(X, Z)-continuous vector-functions u : dom(u) → X such that dom(u) is a comeager set in Q. Consider the factor set C∞ (Q,
X|Z) := M /∼, where u ∼ v means that u(t) = v(t) t ∈ dom(u) ∩ dom(v) . The set C∞ (Q, X|Z) can be made into a vector space in the natural way: if u e stands for the cosets of u ∈ M then a sum λe u + µe v is interpreted as the coset of the pointwise sum λu(t) + µv(t), t ∈ dom(u) ∩ dom(v). For u ∈ M and z ∈ Z denote by hu, zi the unique continuous extension of the function t 7→ hu(t), zi (t ∈ dom(u)) onto the entire Q. If u ∼ v then clearly hu, zi = hv, zi; thus, for w ∈ C∞ (Q, X|Z) and an arbitrary u ∈ w we may put hw, zi := hu, zi. The set R(u) := hu, zi : z ∈ Z, kzk ≤ 1 is order-bounded in C∞ (Q), it is pointwise bounded on the comeager set dom(u). Therefore, for an arbitrary u ∈ w we may assign  w := u := sup hu, zi : z ∈ Z, kzk ≤ 1 , where supremum is taken in C∞ (Q). Observe that the function ku(·)k : t 7→ ku(t)k (t ∈ dom(u)) is pointwise the least upper bound of the same family R(u). Therefore, the functions u and ku(·)k coincide on a comeager subset of Q. Nevertheless, these function may differ on dom(u). It is easy to see that · is a decomposable norm with values in C∞ (Q). Moreover, C∞ (Q, X|Z) is naturally endowed with the structure of a faithful module over the ring C∞ (Q). Put  Ew (X, Z) := u ∈ C∞ (Q, X|Z) : u ∈ E . Indicate an important particular case Ew (X 0 ) := Ew (X 0 , X) that appears when X := X 0 and Z := X ⊂ X 00 . If X is a Banach space then for every order-dense ideal E ⊂ C∞ (Q) the set Ew (X, Z), endowed with the operations and E-valued norm · induced from C∞ (Q, X|Z), is a Banach–Kantorovich space over E and C∞ (Q, X|Z) is its universal completion. In particular, Ew (X 0 ) is a Banach–Kantorovich space over E. 2.3.6. Let X, Q, and E be the same as in 2.3.5. Let Y be one more normed space, with Z ⊂ Y 0 a norming space for Y and L (X, Y ) standing for the space of all bounded linear operators from X to Y . Denote by MQ (X, Y ) the set of all operator-functions K : dom(K) → L (X, Y ) that satisfy the following conditions:

66

Chapter 2

(a) Q0 := dom(K) is a comeager subset of Q; (b) the function s 7→ hK(s)x, zi (s ∈ Q0 ) is continuous for all x ∈ X and z ∈ Z. It follows from the definition that, given x ∈ X and z ∈ Z, there exists a unique element u ∈ C∞ (Q) for which u(t) = hK(t)x, zi (t ∈ Q0 ). Assign hKx, zi := u. Just as in 2.3.5 there exists a ϕ ∈ C∞ (Q) such that kK(s)k ≤ ϕ(s) (s ∈ Q0 ). Thus, it is clear that  K := sup hKx, zi : kxk ≤ 1, kzk ≤ 1 ≤ ϕ exists in the C∞ (Q). Introduce an equivalence in MQ (X, Y ) by letting K ∼ L e is the coset of an operatorwhenever K(s) = L(s) for all s ∈ dom(K)∩dom(L). If K e function K then we assign K := K . This definition is sound since K = L for equivalent K and L. Now introduce the space
 Ew L (X, Y ), Z := K ∈ MQ (X, Y )/∼ : K ∈ E .
It is easy to verify that Ew L (X, Y ), Z is a decomposable LNS. It coincides with Ew (X, Z) whenever X := L (X, Y ) and Z := X ⊗ Z. An important particular case is presented by the space


^ Y )0 , Ew L (X, Y 0 ) := Ew L (X, Y 0 ), Y := Ew (X ⊗ ^ Y is the projective tensor product of the spaces X and Y . where X ⊗ If Y is a Banach space then for every order-dense ideal E ⊂ C∞ (Q) the set
Ew L (X, Y ), Z is a Banach–Kantorovich space over E
and C∞ (Q, L (X, Y )|Z) 0 is its universal completion. In particular, Ew L (X, Y ) is a Banach–Kantorovich space over E. 2.3.7. Let (Š, †, µ) be a measure space with the direct sum property, let E be an order-dense ideal in L0 (Š, †, µ), and let X be a normed space. Let L0 (µ, X) := L0 (Š, †, µ, X) be the space of cosets of Bochner µ-measurable vectorfunctions acting from Š into X. As usual, vector-functions are equivalent if they assume equal values at almost all points of the set Š. If u e ∈ L0 (µ, X) is the coset of a measurable vector-function of u : Š → X then t 7→ ku(t)k (t ∈ Š) is a scalar measurable function whose coset is denoted by the symbol ze ∈ L0 (µ). Assign by definition E(X) := {u ∈ L0 (µ, X) : u ∈ E}.
Then E(X), E is a lattice-normed space with decomposable norm. Obviously, Lp (µ, X) (1 ≤ p ≤ ∞) coincides with E(X), where E = Lp (µ). If X is a Banach space then E(X) is a Banach–Kantorovich space and L0 (µ, X) is its universal completion. C The assertion can be easily deduced from 2.2.2. It also follows immediately from a more general fact which will be proved in Section 2.5 (see 2.5.4 (3)). B

Lattice-Normed Spaces

67

2.3.8. Suppose, as above, that (Š, †, µ) is a measure space with the direct sum property. Let ρ be a lifting of L∞ (Š, †, µ) and τ : Š → Q be the corresponding canonical immersion of Š into the Stone space Q of the Boolean algebra B(Š, †, µ), ∗ see 1.4.9. Denote by τ ∗ := τX the mapping that sends each function f ∈ C∞ (Q, X) to the equivalence class of the measurable vector-function f ◦ τ . (1) The mapping τ ∗ is a linear isometry (in the sense of vector norms) from C∞ (Q, X) onto L0 (Š, †, µ, X). The image of C# (Q, X) under τ ∗ coincides with the subspace L∞ (Š, †, µ, X). C A more general fact is proved in 2.5.9. B The algebraic tensor product E ⊗X is identified with the set of vector-functions of the form n X f (t) = ei (t)xi (ei ∈ E, xi ∈ X). i=1

It is clear that E ⊗ X is a linear subset in E(X). (2) Let E be an ideal space and let X be a Banach space. Then E ⊗ X is dense in E(X) with respect to the vector norm. C It is readily apparent from (1) and 2.3.4. B 2.3.9. Now we consider the lattice-normed space E[F ] of measurable functions of two variables defined on the product of some measure spaces (Š, †, µ) and (Š0 , †0 , µ0 ). To avoid technical inconveniences assume that (Š0 , †0 , µ0 ) is σ-finite; however, the definition and main properties remain valid in the general case. Let E be an ideal space on (Š, †, µ) and let F be a Banach ideal space on 0 (Š , †0 , µ0 ). Denote by E[F ] the space of all measurable functions K on Š0 × Š satisfying the following: (a) the coset of the function s 7→ K(s, t) belongs to F for almost all t ∈ Š; (b) the function t 7→ kK(·, t)kF is measurable and its coset K belongs to E. (1) Suppose that a norm on F is o-semicontinuous. Then E[F ] is a Banach–Kantorovich space over E and an ideal space on Š0 × Š. C There are two subtle reasonings in the proof. The first is to demonstrate that o-semicontinuity of the norm in F provides measurability for the function K (see [162, XI.1.4]). It follows from this that E[F ] is a linear space and an ideal space on Š0 × Š. Thus, E[F ] is also a decomposable d-complete lattice-normed space over E. The second is to prove bo-completeness of E[F ]. In virtue of Theorem 2.2.3 we need only to check that this space is br-complete. The latter is equivalent to br-completeness of all spaces Ee [F ] where e ∈ E+ and Ee := E(e). Define a scalar norm ||| · ||| on Ee [F ] by |||K||| := k K ke . In 7.1.2 we will prove that Ee [F ] is brcomplete if and only if (Ee [F ], ||| · |||) is norm complete. It can be also easily seen that Ee is an AM -space and (Ee [F ], ||| · |||) is a normed lattice. In view of 1.5.1 (5),

68

Chapter 2

it remains to prove that for every disjoint sequence P∞ (Kn ) of positive functions in Ee [F ] there exists K := supn Kn provided that n=1 |||Kn ||| < ∞. The sought least upper bound K exists in a universally complete vector space L0 (Š0 × Š, ν × µ). P∞ Show that K ∈ Ee [F ]. Observe that the series n=1 Kn is absolutely summable and its sum exists, since (Ee , kP· ke ) is norm complete. From this we deduce that ∞ for almost all t ∈ Š the series n=1 Kn (·, t) is absolutely summable in F . If Kt (·) stands for the sum of this series then Kt (s) = K(s, t) for almost allP s ∈ Š0 in view ∞ of 1.5.8. Thus K(·, t) ∈ F for almost all t ∈ Š. Moreover, K ≤ n=1 Kn and K ∈ Ee . B (2) Let F be a Banach ideal space on (Š0 , †0 , µ0 ) with an order semicontinuous norm. For every measurable vector-function f : Š → F , there is a measurable function K(s, t) on (Š0 × Š) such that, for almost all t ∈ Š the equality K(s, t) = (f (t))(s) holds for almost all s ∈ Š0 . The correspondence f 7→ K implements an isometric embedding of E(F ) onto a closed subspace and a sublattice of E[F ]. C Without loss of generality we may assume that (Š, †, µ) is a finite measure space. Approximate f by a sequence (fn ) of finite-valued functions in the sense of almost everywhere norm convergence in F : kfn (t) − f (t)kF → 0 for almost all t ∈ Š. Every function fn generates a measurable function Kn (·, ·) by formula Kn (s, t) = (fn (t))(s); moreover, Kn − Km (t) = kfn (t) − fm (t)kF → 0 almost everywhere. en − K e → 0 where L e ∈ E[F ] with K e Since E[F ] is norm o-complete there is K stands for the coset of measurable function L. Now the required identity follows from the relations: kf (t) − K(·, t)kF ≤ kf (t) − fn (t)kF + Kn − K (t) → 0. B (3) Let the measure µ be not purely atomic. Then the lattice-normed spaces E(F ) and E[F ] coincide (under embedding (2)) if and only if F is a Banach ideal space with order continuous norm. 2.3.10. Now we introduce the measurable version of the space Ew (X). Take the same E and X as in 2.3.7, and a norming subspace Z ⊂ X 0 , see 2.3.5. A vectorfunction u : Š → X is called σ(X, Z)-measurable or simply Z-measurable if, for each z ∈ Z, the function t 7→ hu(t), zi (t ∈ Š) is measurable. Denote the coset of the last function by the symbol hu, zi, so that hu, zi ∈ L0 (µ). Let M (Š, X|Z) be the set of Z-measurable vector-functions u : Š → X. We say that Z-measurable vector-functions u, v are Z-equivalent and write u ' v if measurable functions hu, zi and hv, zi are equal almost everywhere for each z ∈ Z. Consider the factor set

Lattice-Normed Spaces

69

L0 (µ, X|Z) := L0 (Š, †, µ, X|Z) := M (Š, X|Z)/ ' and define vector space structure in it by setting αe u + βe v := (αu + βv)∼ . For a coset u e ∈ L0 (µ, X|Z) with u ∈ M (Š, X|Z) put he u, zi := hu, zi.  Observe that the set R(e u) := he u, zi : z ∈ Z, kzk ≤ 1 is order-bounded in L0 (µ). Otherwise, according to 1.4.11 (2), one can choose a sequence (fn ) in R(e u) which is unbounded from above. But this is a contradiction, since the function f (t) := supn fn (t) is measurable and |f (t)| ≤ ku(t)kX < ∞ (t ∈ Š). Given u e∈ M / ', assign  u := sup hu, zi : z ∈ Z, kzk ≤ 1 , where the supremum is taken in the L0 (Š, †, µ). It is easy to verify that L0 (µ, X|Z) is a decomposable lattice-normed space over L0 (µ). Define now the set  Ew (X, Z) := u ∈ L0 (µ, X|Z) : u ∈ E . In the same manner as in 2.3.7, we indicate an important particular case, when X is a dual Banach space (X := X 0 ) and Z is its predual (Z := X ⊂ X 00 ). The notation Ew (X 0 ) := Ew (X 0 , X) is conventional. (1) For every order ideal E ⊂ L0 (µ) the space Ew (X 0 ) endowed with the operations and E-valued norm · induced from L0 (Š, †, µ, X 0 |X) is a Banach– Kantorovich space over L0 (Š, †, µ) and L0 (Š, †, µ, X 0 |X) is its universal completion. It is evident that every measurable vector-function is weakly measurable and u ∼ v implies u ' v for every pair of measurable vector-functions u and v. Therefore, there is a mapping (called the canonical embedding) assigning to each u e ∈ L0 (µ, X) the coset {v ∈ M (µ, X|Z) : v ' u} ∈ L0 (µ, X|Z). (2) The canonical embedding L0 (µ, X) → L0 (µ, X|Z) is a linear isometry. 2.3.11. Take the same E and X as in 2.3.10. Let Y be one more normed space and Z ⊂ Y 0 be a norming subspace. An operator-function K : Š → L (X, Y ) is called Z-weakly measurable if, for all x ∈ X and z ∈ Z, the function hz, Kxi : t 7→ hz, K(t)xi (t ∈ Š) is measurable. Denote by Mµ (X, Y ) the set of all Z-weakly measurable operator-functions K : Š → L (X, Y ). Introduce some equivalence relation ∼ in Mµ (X, Y ) by letting K ∼ L if and only if, for each pair (x, z) ∈ X × Z, the measurable functions hz, Kxi and hz, Lxi coincide almost everywhere. For a K ∈ Mµ (X, Y ), as well as for the corresponding e assign coset K,  e := sup hz, Kx i : kxk ≤ 1, kzk ≤ 1 , K = K

70

Chapter 2

where the supremum is taken in L0 (µ). This definition is sound, since the same reasoning as in 2.3.10 shows that K is the least upper bound of an order-bounded set in L0 (µ). Now put
 Ew L (X, Y ), Z := K ∈ Mµ (X, Y )/∼ : K ∈ E .
This notation agrees with 2.3.10, since it is easily seen that Ew L (X, Y ), Z = Ew (X, Z) whenever X := L (X, Y ) and Z := X ⊗ Z. We also note that Ew (X 0 ) =
Ew L (X, R) . Again, we distinguish an important particular case


^ Y )0 , Ew L (X, Y 0 ) := Ew L (X, Y 0 ), Y := Ew (X ⊗ ^ Y is the projective tensor product of the spaces X and Y . where X ⊗ The following assertions hold:
(1) Ew L (X, Y ), Z is a decomposable lattice-normed space;
(2) Ew L (X, Y 0 ) is a Banach–Kantorovich space. 2.3.12. An operator-function K : Š → L (X, Y ) is called simply measurable if, for each x ∈ X, the vector-function Kx : t 7→ K(t)x (t ∈ Š) is measurable. It is clear that any simply measurable operator-function is weakly measurable. Let Msµ (X, Y ) be the part of Mµ (X, Y ) consisting of simply measurable operator-functions with
values in L (X, Y ). As usual, we assume that Y ⊂ Z 0 and L (X, Y ) ⊂ L (X, Z 0 ). Assign
 Ews L (X, Y ) := K ∈ Msµ (X, Y )/≈ : K ∈ E , where K is the same as in 2.3.11 and the equivalence ≈ is defined as follows: L ≈ K if and only if, for each x ∈ X, the measurable vector-functions Kx and Lx coincide almost everywhere. If K ∈ Msµ (X, Y ) then we also have  K = sup kKxk : kxk ≤ 1 , where kKxk is
the measurable function t 7→ kK(t)xk (t ∈ Š). Observe that E(Y ) =
Ews L (R, Y ) and Ew (X 0 ) = Ews L (X, R) . 2.4. Continuous Banach Bundles The main subject of this section is the following assertion: Every Banach– Kantorovich space is linearly isomorphic to the space of almost global sections of some continuous Banach bundle.

Lattice-Normed Spaces

71

2.4.1. First of all, we give a general definition of a Banach bundle. Let Q be a topological space. A bundle over Q is defined to be a continuous surjection σ : X → Q of a topological space X onto Q. A nonempty set Xq = σ −1 (q) is called the stalk at a point q ∈ Q. A mapping s from a nonempty set dom(s) ⊂ Q into X is called a section over dom(s) if s(q) ∈ Xq for each q ∈ dom(s). A continuous section s is called local, almost global, or global, whenever its domain of definition dom(s) is respectively an open proper subset, a comeager subset, or the whole of Q. A bundle σ : X → Q is often identified with the mapping q 7→ Xq (q ∈ Q), and the notation X (q) is used instead of Xq . Moreover, the bundle is sometimes denoted by X , with the default parameters σ and Q omitted. A set of sections S is called stalkwise dense in X if the set {s(q) : s ∈ S} is dense in X (q) for every q ∈ Q. The product of two bundles σ : X → Q and σ 0 : X 0 → Q is the bundle τ : X ×Q X 0 → Q defined by the following formulas:  X ×Q X 0 := (x, x0 ) ∈ X × X 0 : σ(x) = σ 0 (x0 ) ,
τ : (x, x0 ) 7→ σ(x) = σ 0 (x0 ) (x, x0 ) ∈ X ×Q X 0 . 2.4.2. A continuous Banach bundle over Q is defined to be a bundle σ : X → Q that satisfies the following four conditions: (1) every stalk X (q) is a Banach space with the norm k·kq , and the topology defined by the norm coincides with that induced; (2) addition, X ×Q X 3 (x, y) 7→ x + y ∈ X , scalar multiplication, R × X 3 (α, x) 7→ αx ∈ X , and the zero section, Q 3 q 7→ 0q ∈ X (q), are continuous mappings; (3) the sets of the shape  U (s, ε) := x ∈ X : σ(x) ∈ dom(s), kx − s(σ(x))kσ(x) < ε , where ε > 0 and s is a local section, form a base for the topology on X ; (4) for each point x ∈ X and every number ε > 0, there exists a local section s such that σ(x) ∈ dom(s) and kx − s(σ(x))kσ(x) < ε. In the case of a paracompact space Q, the above definition can be simplified. Namely, instead of (3) and (4) we may require the following: (30 ) every neighborhood of zero 0q ∈ Xq contains a neighborhood of the form U (s, ε), where ε > 0 and s is the restriction of the zero section q 7→ 0q to some open set; (40 ) the mapping σ is open, and k·k : X → R, kxk := kxkσ(x) , is upper semicontinuous. In this case a continuous Banach bundle has a very useful property: (5)
For each x ∈ X , there exists a global continuous section s such that s σ(x) = x.

72

Chapter 2

2.4.3. It is sometimes convenient to define a topology for a Banach bundle with the help of a certain set of sections. Take a surjective mapping σ : X → Q and assume that Xq := σ −1 (q) is a Banach space for each q ∈ Q. The latter amounts to considering a mapping q 7→ Xq from Q into the class of Banach spaces. Introduce the structure of a vector space in the set of all global sections S(Q, X ) by letting (αu + βv)(q) = αu(q) + βv(q) (q ∈ Q) for α, β ∈ R and u, v ∈ S(Q, X ). For each section s ∈ S(Q, X ) we consider its pointwise norm |||s||| : q 7→ ks(q)kq (q ∈ Q). Now let C be a set of global sections that satisfies the following conditions: (1) C is a vector subspace of S(Q, X ); (2) for each s ∈ C , the function |||s||| is continuous;  (3) for each q ∈ Q, the set s(q) : s ∈ C is dense in the stalk Xq . With the help of C , called a continuity structure in X , we may introduce  a topology in X by taking the family U (s, ε) as a basis for the topology, where ε > 0 and s is the restriction of some section in C to an arbitrary open set. Then σ : X → Q is a continuous Banach bundle. Moreover, a section u ∈ S(Q, X ) is continuous if and only if the function q 7→ ku − skq (q ∈ Q) is continuous for every section s ∈ C . Let C(Q, X ) denote the space of all global continuous sections of X . 2.4.4. Consider an important example of a Banach bundle. Assume that
X, (k·kq )q∈Q is a multinormed space, where  Q is still a topological space. For each q ∈ Q, assign Xq := x ∈ X : kxkq = 0 and denote by Xq the completion of the normed factor space X/Xq . With each element x ∈ X associate a global section ı(x) ∈ S(Q, X ) by the formula ı(x)(q) = ϕq (x), where ϕq : Xq → X/Xq ⊂ Xq is the factor homomorphism. The set C := ı(X) satisfies conditions 2.4.3 (1–3) if and only if, for each x ∈ X, the function q 7→ kxkq (q ∈ Q) is continuous. If the last condition is satisfied then, due to 2.4.3, X becomes a continuous Banach bundle. In this case, the constructed Banach bundle X is said to be associated
with the multinormed space X, (k·kq )q∈Q , and the mapping ı : X → C(Q, X ) is called the canonical embedding. 2.4.5. We shall need the following auxiliary fact: Let (Qξ )ξ∈„ be a family of pairwise disjoint open sets whose union S is dense in Q. If Pξ is a comeager subset of Qξ for each ξ ∈ „, then the union P := ξ∈„ Pξ is again a comeager set in Q. C ForTeach ξ ∈ „, choose a sequence (Gn,ξ )n∈N of open dense subsets of Qξ such that n∈N Gn,ξ ⊂ Pξ (ξ ∈ „). Since Qξ are pairwise disjoint, we have [ \ \ [ Gn,ξ = Gn,ξ . ξ∈„ n∈N

n∈N ξ∈„

Lattice-Normed Spaces

73

S T Hence it is clear that if Gn := ξ∈„ Gn,ξ then P ⊃ n∈N Gn . Thus, P is a comeager set, since Gn is open and dense in Q for each n ∈ N. B 2.4.6. Now suppose that Q is an extremal compact space. Take a continuous Banach bundle X over Q. If u is an almost global section of the bundle X then the function q 7→ ku(q)kq is defined and continuous on a comeager set dom(u). Consequently, there
exists a unique function u ∈ C∞ (Q) such that u (q) = ku(q)kq q ∈ dom(u) . Introduce an equivalence relation in the set of almost global sections M(Q, X ) by letting u ∼ v if u(q) = v(q) whenever q ∈ dom(u) ∩ dom(v). ~ := u , where For equivalent u and v, we have u = v ; therefore, we may define u u ~ is the coset of the almost global section u. Denote the factor set M(Q, X )/∼ by C∞ (Q, X ). The set C∞ (Q, X ) can be naturally made a lattice-normed space. For instance, by the element u ~+~ v we mean the coset of the almost global section
q 7→ u(q) + v(q) q ∈ dom(u) ∩ dom(v) . If E is an order ideal in C∞ (Q) then we assign  E(X ) := u ∈ C∞ (Q, X ) : u ∈ E . In each coset u ~, there exists a unique section u ¯∈u ~ such that dom(v) ⊂ dom(¯ u) for all v ∈ u ~. The section u ¯ is called extended. The space C∞ (Q, X ) can thus be represented as the space of all extended almost global sections of the bundle X . The constant mapping X = Q × {X} that associates with each point of a topological space Q the same Banach space X is a simple example of a Banach bundle. If the totality C of constant functions c : Q → X is taken as a continuity structure C , then the continuous sections of the CBB (X , C ) are exactly continuous functions u : Q → X. Moreover, in this case we have C∞ (Q, X ) = C∞ (Q, X) and E(X ) = E(X). It will be shown in 2.4.8 that the space C(Q, X) is a Banach– Kantorovich space if and only if the trivial bundle Q × {X} is ample. 2.4.7. Theorem. Let X be a continuous Banach bundle over an extremal compact space Q and let E be an order-dense ideal in C∞ (Q). Then E(X ) is a Banach–Kantorovich space and C∞ (Q, X ) is its universal completion. C It is sufficient to establish that C∞ (Q, X ) is a universally complete Banach– Kantorovich space. Take a pairwise disjoint family (uξ )ξ∈„ in C∞ (Q, X ). Let Qξ be the least clopen set in Q on the complement of which uξ equals zero. Then Qξ ∩ Qη = ∅ for ξ 6= η. Without loss of generality we may assume that (Qξ )ξ∈„ is a partition of unity inSthe Boolean algebra of clopen subsets of Q. Assign Gξ := Qξ ∩ dom(uξ ) and G := ξ∈„ Gξ . Define a section u0 by the formulas dom(u) := G,

u0 |Gξ = uξ |Gξ (ξ ∈ „).

The set G is comeager due to 2.4.5, and the section u0 is continuous by definition. Consequently, u0 determines a unique element u ∈ C∞ (Q, X ) for which πξ uξ = πξ u

74

Chapter 2

(ξ ∈ „), where πξ is the projection induced by the operator of multiplication by P the characteristic function of the set Gξ . If θ is a finite set in „ and uθ = ξ∈θ uξ , then u − uθ ≤ πθ e, where πθ = sup{πξ : ξ ∈ θ}; therefore, u = bo-lim uθ . Now let a sequence (un )n∈N in C∞ (Q, X ) beTr-fundamental with regulator ∞ e ∈ C∞ (Q). On the comeager set G := dom(e) ∩ n=1 dom (un ), we may define a section u by the formula u(q) := lim un (q). Existence of the limits follows from the estimate kun (q) − um (q)k ≤ λk e(q) (n, m ≥ k), where λn → 0. After passing to the limit as m → ∞, it becomes clear from the last inequality that un converges to u locally uniformly. Consequently, the section u is continuous. It remains to observe that un − u ≤ λk e (n ≥ k), i.e., br-limn→∞ un = u. Thus, C∞ (Q, X ) is d-complete and r-complete and, therefore, the required assertion ensues from 2.2.3. B 2.4.8. A continuous Banach bundle X over an extremal compact space Q is called ample (or complete) if every bounded almost global continuous section of it can be extended to a global continuous section. Introduce the notation  C# (Q, X ) := u ∈ C∞ (Q, X ) : u ∈ C(Q) , i.e., C# (Q, X ) = E(X ) for E := C(Q). Let C(Q, X ) denote the set of all global continuous sections of X . The following assertions are equivalent: (1) the bundle X is ample; (2) C# (Q, X ) = C(Q, X ); (3) C(Q, X ) is a Banach–Kantorovich space. C The implications (1) ⇒ (2) ⇒ (3) are obvious. Suppose that (3) is satisfied and show that X is an ample bundle. Take a section u ∈ C# (Q, X ). There exist a partition of unity (Qξ )ξ∈„ in the Boolean algebra of clopen subsets of Q and a family of global continuous section (uξ )ξ∈„ such that ku(q) − uξ (q)kq ≤ 1/n for all q ∈ Qξ ∩ dom(u) and ξ ∈ „. This assertion can be easily deduced from 2.4.2 (5), since dom(u) is dense in Q. Let πξ be the projection in C∞ (Q, X ) that corresponds to the set Qξ . Due to bo-completeness of C(Q, X ), there exists a section vn ∈ C(Q, X ) with πξ vn = πξ uξ (ξ ∈ „). It follows from this that u − vn ≤ (1/n)1. Thus, the sequence (vn ) in C(Q, X ) converges relatively uniformly to u. Again we may apply bo-completeness of C(Q, X ) and conclude that u ∈ C(Q, X ), as required. B Let us show that assertions (1)–(3) are also equivalent to the following: (4) every bounded continuous section of the bundle X that is defined on a dense subset of Q can be extended to a global continuous section. It is not difficult to show that the trivial bundle Q × {X} is complete if and only if Q is a finite space or X is a finite-dimensional space.

Lattice-Normed Spaces

75

2.4.9. Let X and Y be continuous Banach bundles over the same compact space Q. A mapping h : q 7→ hq := h(q) ∈ L (Xq , Yq ) (q ∈ Q) is called a homomorphism from X into Y if hq is a bounded linear operator from Xq into Yq for all q ∈ Q, and, for
each continuous section u ∈ C(Q, X ), the mapping hu := h ⊗ u : q 7→ hq u(q) (q ∈ Q) is continuous. (1) If h is a homomorphism from X into Y then the function |||h||| : q 7→ kh(q)k is bounded. C For every q ∈ Q define a linear operator Tq : C(Q, X ) → Y (q) by setting Tq u = h(q)u(q). Endow the space C(Q, X ) with the uniform norm kuk∞ := sup{ku(q)kX (q) ; q ∈ Q} and observe that kTq k ≤ kh(q)k. Moreover, sup kTq uk = sup kh ⊗ uk < ∞ q∈Q

q∈Q

for every u ∈ C(Q, X ). We may apply the Uniform Boundedness Principle and derive supq∈Q kTq k < ∞. It remains to note that kh(q)k = sup{kh(q)xk : x ∈ X (q), kxk 6 1} = sup{kh(q)u(q)k : u ∈ C(Q, X ), kuk∞ 6 1} = kTq k. B We say that a homomorphism h is an isometry if hq is an isometry for each q ∈ Q. The set of all homomorphism from X into Y denote by HomQ (X , Y
). If h is an isometry of X onto Y then the mapping u 7→ hu u ∈ C(Q, X ) is a linear isometry of the lattice-normed spaces C(Q, X ) and C(Q, Y ). The converse assertion is also valid; however, a stronger assertion holds in the case of ample bundles. (2) Suppose that X and Y are ample Banach bundles over an extremal compact space Q. Then X and Y are isometric if and only if the LNSs C∞ (Q, X ) and C∞ (Q, Y ) are linearly isometric. C Let J be a linear isometry of C∞ (Q, X ) onto C∞ (Q, Y ). Then u ∈ C(Q, X ) if and only if Ju ∈ C(Q, Y ), since C = C# for
any ample bundle (see 2.4.8 (2)). Given u ∈ C(Q, X ) and q ∈ Q, assign hq u(q) := (Ju)(q). It is easy to see that
khq u(q) kq = k(Ju)(q)kq = Ju (q) = u (q) = ku(q)kq , i.e., hq is a linear isometry of Xq into Yq . Moreover, it is obvious that hu ∈ C(Q, Y ) whenever u ∈ C(Q, X ). For each y ∈ Yq , we may choose a global continuous section
−1 v so that v(q) = y. If u := J (v) then hq u(q) = y; hence, the operator hq is surjective. Consequently, h is an isometry of X onto Y . B

76

Chapter 2

2.4.10. Theorem. Every Banach–Kantorovich space X over an order-dense ideal E ⊂ C∞ (Q) is linearly isometric to E(X ) for some ample continuous Banach bundle X over Q. Moreover, such a bundle X is unique to within linear isometry. e be the universal completion of the space X (see 2.2.8). Assign C Let X  e : x ∈ C(Q) and kxkq := x (q) (q ∈ Q, x ∈ X0 ). It is clear X0 := x ∈
X that k · kq q∈Q is a multinorm in X0 . Let X be the continuous Banach bundle associated with the multinorm and let ı : X0 → C(Q, X ) be the canonical embedding (see 2.4.4). From the definition of the topology in X (see 2.4.3) it is clear that an arbitrary section u ∈ C(Q, X ) goes through the ε-tube defined by some section of the form ı(x), x ∈ X0 , on a suitable clopen neighborhood of each point of Q. Consequently, for every ε > 0, we may choose a partition of unity (Qξ )ξ∈„ in the Boolean algebra of clopen subsets of Q and a family (xξ )ξ∈„ of elements in X0 such that ku(q)−ı(xξ )(q)kq ≤ ε (q ∈ Qξ , ξ ∈ „); i.e., πξ u−πξ ı(xξ ) ≤ ε1 (ξ ∈ „), where πξ is the projection induced by the operator of multiplication by the characteristic P function of the set Qξ . In the space X0 there exists an element xε = πξ xξ , for which ı(xε ) − u ≤ ε1. Due to br-completeness of X0 , the limit x := br-lim xε also exists as ε → 0; moreover, ı(x) = u and x = u . Thus, ı(X0 ) = C(Q, X ), and, hence, C(Q, X ) is a Banach–Kantorovich space. According to 2.4.8, the bundle X is ample. The uniqueness assertion follows from 2.4.9. B 2.4.11. A continuous Banach bundle X is called a subbundle of a continuous Banach bundle X if X (q) is a Banach subspace of X (q) for every q ∈ Q and, moreover, C(Q, X ) = C(Q, X ) ∩ S(Q, X ). A subbundle X is called dense in X if every section u ∈ C(Q, X ) takes the values u(q) ∈ X (q) on a dense subset of Q. An ample hull or completion of a continuous Banach bundle X is any ample continuous Banach bundle X containing X as a dense subbundle. (1) Theorem. Every continuous Banach bundle over an extremal compact space has an ample hull unique to within isometry. C Consider a continuous Banach
bundle X over Q and endow the space C# (Q, X ) with the multinorm k · kq q∈Q , by writing kukq := u (q) for all u ∈ C# (Q, X ) and q ∈ Q. According to 2.4.4, the multinormed space thus obtained generates a continuous Banach bundle over Q that we denote by X . Let ı : C# (Q, X ) → C(Q, X ) be the corresponding canonical embedding. At each point q ∈ Q it is possible to define an isometric monomorphism h(q) : X (q) → X (q), satisfying the equality h(q)u(q) = ı(u)(q) for all u ∈ C(Q, X ). By Definition 2.4.9 it follows that h is a homomorphism from X into X and, thus, performs an isometry from X onto a Banach subbundle of X . According to above definitions, this subbundle is dense in X . Uniqueness for the completion of a continuous Banach bundle to within isometry follows from 2.4.9 (2). B

Lattice-Normed Spaces

77

(2) Theorem. Let X and Y be CBBs over Q, let X and Y be respective ample hulls, and let D be a comeager subset of Q. Then every homomorphism h ∈ HomD (X , Y ) has a unique extension h ∈ HomD (X , Y ). Moreover, the pointwise norms h and h coincide on a comeager subset of Q. ¯ ∈ X (q0 ) define C Take h ∈ HomD (X , Y ). For arbitrary q0 ∈ D and x h(q0 )¯ x ∈ Y (q0 ). To this end choose a section u ¯ ∈ C(Q, X ) such that u ¯(q0 ) = x ¯ and denote by u the restriction of u ¯ to comeager subset {q ∈ Q : u ¯(q) ∈ X (q)}. Then h ⊗ u is a continuous almost global section of Y . Let h ⊗ u ∈ C∞ (Q, Y ) be a maximal extension of h ⊗ u in Y . According to 2.4.9 (1) |||h||| is a locally bounded function. Therefore, |||H ⊗u||| is bounded on a neighborhood of q0 and, consequently, q0 ∈ dom(h ⊗ u). Now, we put h(q0 )x := h ⊗ u(q0 ). Clearly, h ⊗ u(q0 ) does not depend on the choice of the section u ¯ ∈ C(Q, X ) satisfying u ¯(q) = x ¯ and, hence, the desired homomorphism h is
correctly defined. The uniqueness of h follows from the equality H(q) extX (u)(q) = extY (H ⊗u)(q) for all u ∈ C# (Q, X ) and q ∈ Q. Since C∞ (Q, X ) is a universal completion of C(Q, X ), for every section u ∈ C∞ (Q, X ), u 6 1, one can find a net (uα )α∈A in C(Q, X ) bo-converging to u and satisfying the inequality uα 6 1. Then h ⊗ uα − h ⊗ u 6 |||h||| uα − u holds on D ∩dom(u), so that h⊗uα bo-converges to h⊗u . Thus, h⊗u 6 supα∈A h⊗uα and, for q ∈ D, we deduce |||h|||(q) = sup{ h ⊗ u (q) : u ∈ C(Q, X ), u 6 1}, ≥ sup{ h ⊗ u (q) : u ∈ C∞ (Q, X ), u 6 1} = |||h|||(q). The reverse inequality is obvious. B (3) Let X be a continuous Banach bundle, with X an ample hull of X . Then the spaces C∞ (Q, X ) and C∞ (Q, X ) coincide. C This follows from the definition of dense subbundle, since it is not difficult to check that a Banach bundle X is a dense subbundle of a continuous Banach bundle X if and only if every section u ∈ C∞ (Q, X ) takes the values u(q) ∈ X (q) on a comeager subset of Q. B 2.4.12. Now we introduce the continuous Banach bundle B(X , Y ) whose continuous sections are homomorphisms from a continuous Banach bundle X into a continuous Banach bundle Y . Consider a nonempty extremal compact space Q, and continuous Banach bundles X and Y over Q. (1) Let D
be a dense subset of Q, and let a mapping h : q ∈ D 7→ h(q) ∈ L X (q), Y (q) be such that hu ∈ C(D, Y ) for all u ∈ C(Q, X ). Suppose that the bundle X is ample. Then the pointwise norm |||h||| is continuous. In view of 2.4.2 (5), the equality |||h|||(q) = sup{hu(q) : u ∈ C(Q, X ), u 6 1}

78

Chapter 2

is valid for all q ∈ D which implies that the function |||h||| is lower semicontinuous. In order to prove upper semicontinuity for |||h||| at an arbitrary point q ∈ D we assume that |||h|||(q) < λ and establish the inequality |||h|||(p) < λ for all elements p in a neighborhood of q within D. Assume the contrary. Then, taking a number λ0 so that |||h|||(q) < λ0 < λ and defining V = {p ∈ D : |||h|||(p) > λ0 } we conclude that q ∈ cl V . For every element p ∈ V let a section up ∈ C(Q, X ) satisfy up 6 1 and hup (p) > λ0 . In view of continuity of the function hup at each point p ∈ V there is a clopen neighborhood Vp ⊂ Q such that S hup > λ0 on Vp ∩ D. It is easy to verify that the following equality holds: cl p∈V Vp = cl V . The Exhaustion Principle implies existence of a family (Up )p∈V of pairwiseSdisjoint clopen subsets of Q, satisfying the conditions Up ⊂ Vp (p ∈ V ) and cl p∈V Up = cl V . Since S X is ample, S the continuous bounded section p∈V up Up ∪ 0 Q\ cl V of X over the dense subset p∈V Up ∪ (Q\ cl V ) can be extended to a global section u ∈ C(Q, X ). Obviously, u 6 1. Since the sections u and up coincide on Up , it follows that hu > λ0 on cl V . The last assertion contradicts the inequality |||h|||(q) < λ0 . B (2) Every homomorphism h ∈ HomQ (X , Y ) has a continuous pointwise norm |||h||| provided that the bundle X is ample. The following assertion is a version of the Stone–Weierstrass Theorem for a CBB over an extremal compact space. The section coinciding with u on a clopen set V and vanishing on its complement is denoted by [V ]u. (3) Theorem. Let a vector subspace U ⊂ C(Q, X ) be stalkwise dense in X and contain [V ]u for all elements u ∈ U and clopen subsets K ⊂ Q. Then U is uniformly dense in C(Q, X ). C Suppose that a subspace U ⊂ C(Q, X ) meets the hypotheses of the theorem. We will fix a section v ∈ C(Q, X ) and, given an arbitrary ε > 0, construct an element u ∈ U satisfying the inequality ku − vk∞ ≤ ε. Given a point q ∈ Q choose a section uq ∈ U such that kuq (q) − v(q)k < ε. In addition, denote by Vq a clopen neighborhood of q on which uq − v < ε. Let us refine a finite subcover Vq1 , . . . , Vqn from the open cover (Vq )q∈Q of the compact Q. According to the Exhaustion Principle (see 1.1.6) there is a partition W1 , . . . , Wn of Q into clopen Pnsubsets such that Wk ⊂ Vqk for all k = 1, . . . , n. It is clear that the section u = k=1 [Wk ]uqk is the desired one. B 2.4.13. Theorem. Suppose that X is an ample bundle. Then there exists a (unique) continuous Banach bundle B(X , Y ) over Q such that the
stalk B(X , Y )(q) over
each point q ∈ Q is a Banach subspace of B X (q), Y (q) and C Q, B(X , Y ) = HomQ (X , Y ). C Consider the discrete Banach bundle over Q with the stalk over each point q ∈ Q equal to
the closure of the subspace {h(q) : h ∈ HomQ (X , Y )} in the bundle B X (q), Y (q) . In view of 2.4.12 (2), the space HomQ (X , Y ) is a continuity

Lattice-Normed Spaces

79

structure in this bundle and thus makes it into the desired continuous Banach bundle B(X , Y ). The equality
C Q, B(X , Y ) = HomQ (X , Y ) ensues from 2.4.12 (3). B 2.4.14. Theorem. Suppose that X is an ample bundle and D is a comeager subset of Q and U ⊂ C(D, X ) is
stalkwise dense in X on D. Let a mapping h : q ∈ D 7→ h(q) ∈ B X (q), Y (q)
be such that h ⊗ u ∈ C(D, Y ) for all elements u ∈ U . Then h ∈ C D, B(X , Y ) . C Let Y be an ample hull of Y . Using 2.4.11 (2), prove that there exists a section
h ∈ C∞ Q, B(X , Y ) such that the operator h(q) is an extension of h(q) for all q ∈ D ∩ dom(h). Moreover, h and h coincide on D ∩ dom(h). Assume that there exists an element ¯ For every q ∈ D := dom( h ) choose a section uq ∈ C(Q, X ) q0 ∈ D\ dom(h). and a clopen set Uq ⊂ Q such that q ∈ Uq and the inequalities uq 6 1 and h ⊗ uq > h
− ε hold on Uq . According to the Exhaustion Principle we may find a family Vq q∈D of pairwise disjoint clopen sets in Q such that Vq ⊂ Uq (q ∈ D) ¯ ∈ C(Q, X ) be a section coinciding with and the union of all Vq is dense in Q. Let u ¯ > h − 1. Since U is stalkwise dense in X on D there is an uq on Vq . Then h ⊗ u ¯(q0 )k < 1/2; thus, u − u ¯ 6 1/2 in a neighborhood element u ∈ U with ku(q0 ) − u of q and we deduce

>

kh(q0 )u(q0 )k = h ⊗ u (q0 ) = h ⊗ u (q0 )

1 ¯ (q0 ) > h⊗u − h · u−u h − 1 (q0 ) = ∞. 2

The contradiction shows that D ⊂ dom h. B 2.5. Measurable Banach Bundles In this section, we introduce the notion and establish some elementary properties of measurable sections obtained by means of a measurability structure and prove that every Banach–Kantorovich space over an ideal space is linearly isometric to the space of measurable sections of a measurable Banach bundle. 2.5.1. Throughout the section, (Š, †, µ) is a nonzero measure space. Let X be a Banach bundle over Š. Denote by S∼ (Š, X ) the set of all sections of X defined almost everywhere on Q. We call a set of sections C ⊂ S∼ (Š, X ) a measurability structure on X , if it satisfies the following three conditions:

80

Chapter 2 (a) λ1 c1 + λ2 c2 ∈ C for all λ1 , λ2 ∈ R and c1 , c2 ∈ C ; (b) the pointwise norm |||c||| : Š → R of every element c ∈ C is measur-

able; (c) the set C is stalkwise dense in X . If C is a measurability structure in X then we call the pair (X , C ) a measurable Banach bundle over Š. We shall usually write simply X instead of (X , C ) and denote the measurability structure C by CX . Let (X , C ) be a measurable Banach bundle over Š. We say that s ∈ S∼ (Š, X ) is a C -step-section (or simply Pna step-section, if it is clear which measurability structure is meant), if s = k=1 [Ak ]ck for some n ∈ N, A1 , . . . , An ∈ † and c1 , . . . , cn ∈ C . A section u ∈ S∼ (Š, X ) is called C -measurable (or simply measurable) if, for every K ∈ † with µ(K) < +∞, there is a sequence (sn )n∈N of C -step-sections such that sn (ω) → u(ω) for almost all ω ∈ K. The set of all C measurable sections of X is denoted by M (Š, †, µ, X |C ) or M (Š, X ) for brevity. 2.5.2. Suppose that X is a measurable Banach bundle over Š. We consider the equivalence relation ∼ in the set M (Š, X ) which is the coincidence almost everywhere: u ∼ v means that u(ω) = v(ω) for almost all ω ∈ Š. The coset containing an element u ∈ M (Š, X ) is denoted by u∼ . The factor set M (Š, X )/∼ is made into a vector space in the natural way: we write λu∼ + µv ∼ = (λu + µv)∼ for λ, µ ∈ R and u, v ∈ M (Š, X ). In addition, for every element u∼ ∈ M (Š, X )/∼ we may define its (vector)
∼ ∼ norm u := |||u||| ∈ M (Š). It is clear that the pair M (Š, X )/∼, · is an LNS over L0 (Š, †, µ); we denote it by L0 (Š, †, µ, X ), or shortly L0 (Š, X ). Note that the space L0 (Š, X ) can be endowed with the natural structure of a module over the ring L0 (Š, †, µ); as follows: e∼ u∼ := (eu)∼ for all e ∈ M (Š) and u ∈ M (Š, X ). 2.5.3. Theorem. If a measure space (Š, †, µ) possesses the direct sum property and if X is a measurable Banach bundle over Š, then L0 (Š, X ) is a Banach– Kantorovich space over L0 (Š, †, µ). C Decomposability of the LNS L0 (Š, X ) is obvious. In view of 2.2.3 for proving bo-completeness of L0 (Š, X ), it suffices to establish its d- and br-completeness. 0 Let (u∼ ξ )ξ∈„ be a family of elements of L (Š, X ), with pairwise disjoint supports Aξ ∈ B(Š) (see 1.4.11). Denote by „ the enrichment „ ∪ {∞} of the set „
∗ with a new element ∞ and define u∞ := 0 and A∞ := supξ∈„ Aξ . Fix a lifting ρ of L∞ (Š)S(see 1.4.9) and denote Aξ := ρ(Aξ ) for each ξ ∈ „. In view of 1.2.7 (2), the union ξ∈„ Aξ is measurable and differs from Š by a set of measure zero. It S is easy to verify that the section ξ∈„ uξ Aξ is measurable and the corresponding P 0 class is the desired sum bo- ξ∈„ u∼ ξ ∈ L (Š, X ).

Lattice-Normed Spaces

81

0 Now, suppose that a sequence (u∼ n )n∈N
in L (Š, X ) is br-fundamental. Then, for almost all ω ∈ Š the sequence un (ω) n∈N is fundamental. Due to completeness of the stalks of X , there exists a section u ∈ S∼ (Š, X ) to which the sequence (un )n∈N converges almost everywhere. It is clear that the section u is measurable and the corresponding class is the desired br-limit of the sequence (u∼ n )n∈N . B

2.5.4. Some important corollaries of the above theorem are now available. (1) Theorem. If (Š, †, µ) is a measure space possessing the direct sum property and E is an ideal of L0 (Š) then the set E(X ) := {u ∈ L0 (Š, X ) : u ∈ E} endowed with the operations induced from L0 (Š, X ) is a Banach–Kantorovich over E.  ∞ (2) The symbol L (Š, X ) stands for the set u ∈ M (Š, X ) : |||u||| ∈ ∞ L (Š) and its elements are called (essentially) bounded measurable sections of X . The totality of equivalence classes constituted by essentially bounded sections is denoted by L∞ (Š, X ). Obviously, the space L∞ (Š, X ) coincides with E(X ), where E = L∞ (Š, †, µ). In particular, L∞ (Š, X ) is a Banach–Kantorovich space over L∞ (Š, †, µ), (3) Let X be a Banach space. Consider the trivial Banach bundle X = Š × {X} and let the totality of constant functions c : Š → X be taken as the measurability structure of X . Then the set M (Š, X ) consists exactly of all Bochner measurable X-valued functions defined almost everywhere in Š: M (Š, X) = M (Š, X ). Consequently, L0 (Š, X) = L0 (Š, X ). In particular, L0 (Š, X), L∞ (Š, X), and E(X), where E is an order ideal in L0 (Š, †, µ), are Banach–Kantorovich spaces over L0 (Š, †, µ). 2.5.5. Let X be a measurable Banach bundle over Š. Consider a lifting ρ : L (Š) → L ∞ (Š) (see 1.4.8). We call a mapping ρX : L∞ (Š, X ) → L ∞ (Š, X ) a lifting of L∞ (Š, X ) associated with ρ if, for all u, v ∈ L∞ (Š, X ) and e ∈ L∞ (Š), the following relations hold: ∞

(1) ρX (u) ∈ u∼ and dom(ρX (u)) = Š; (2) |||ρX (u)||| = ρ( u ); (3) ρX (u + v) = ρX (u) + ρX (v); (4) ρX (eu) = ρ(e)ρX (u); (5) the set {ρX (u) : u ∈ L∞ (Š, X )} is stalkwise dense in X . In case there exist a lifting of L∞ (Š) and a lifting of associated with it, we say that X is a liftable measurable Banach bundle. 2.5.6. Let ρ be a lifting of L∞ (Š) and let X and Y be liftable measurable Banach bundles over Š. We call the bundles X and Y ρ-isometric, if their liftings

82

Chapter 2

ρX and ρY are associated with ρ and there exists an isometry h from X onto Y such that ρY (hu) = hρX (u) for all u ∈ L∞ (Š, X ). Two points ω1 , ω2 ∈ Š are said to be ρ-indistinguishable if ρ(f )(ω1 ) = ρ(f )(ω2 ) for every class f ∈ L∞ (Š). We say that a measurable Banach bundle X is invariant with respect to ρ, or ρ-invariant, if X (ω1 ) = X (ω2 ) and ρ(u)(ω1 ) = ρ(u)(ω2 ) for all u ∈ L∞ (Š, X ) for arbitrary ρ-indistinguishable points ω1 , ω2 ∈ Š. 2.5.7. The following assertion enables us to assume without loss of generality that every measurable Banach bundle under consideration is invariant with respect to the corresponding lifting. For every measurable Banach bundle over Š having a lifting associated with ρ, there is a ρ-stationary liftable measurable Banach bundle ρ-isometric to it. C Suppose that a measurable Banach bundle X meets the hypothesis of the assertion. Fix an arbitrary pair of Eρ-indistinguishable points ω1 , ω2 ∈ Š. It suffices to
construct an isometry ı from the stalk X (ω1 ) onto X (ω2 ) such that ı ρ(u)(ω1 ) = ρ(u)(ω2 ) for all u ∈ L∞ (Š, X ). Indeed, in this case, the stalks over indistinguishable points can be identified by means of such isometries. For each ω ∈ Š, denote by X0 (ω) the subspace {ρ(u)(ω) : u ∈ L∞ (Š, X )} of the stalk X (ω). Let classes u, v ∈ L∞ (Š, X ) be such that ρ(u)(ω1 ) = ρ(v)(ω1 ). Then ρ(u)(ω2 ) = ρ(v)(ω2 ), because from the properties of lifting it follows that kρ(u)(ω2 )−ρ(v)(ω2 )k = ρ( u−v )(ω2 ) = ρ( u−v )(ω1 ) = kρ(u)(ω1 )−ρ(v)(ω1 )k = 0. This enables
us to consider a bijection ı0 : X0 (ω1 ) → X0 (ω2 ), defined by the rule ı0 ρ(u)(ω1 ) = ρ(u)(ω2 ) for every class u ∈ L∞ (Š, X ). Due to the fact that the subspaces X0 (ω) are dense in the corresponding stalks X (ω), the isometry ı0 can be extended to the desired isometry ı : X (ω1 ) → X (ω2 ). B 2.5.8. Suppose that Q is the Stone space of the Boolean algebra B(Š) and τ : Š → Q is the canonical immersion of Š into Q corresponding to the lifting ρ of L∞ (Š). Let Y be an ample continuous Banach bundle over Q and X = Y ◦ τ . If C is a continuous structure in Y , then the set C ◦ τ is a measurability structure in X , since |||c ◦ τ ||| = |||c||| ◦ τ and |||c||| ◦ τ is a measurable function. The bundle Y ◦ τ is always regarded as a measurable Banach bundle with respect to the measurability structure C ◦ τ . For every v ∈ C∞ (Q, Y ) the composite v ◦ τ is a measurable section of the bundle X . C This is evident. B For v ∈ C∞ (Q, Y ) we denote by (v ◦ τ )∼ the coset containing an element v ◦ τ .

Lattice-Normed Spaces

83

2.5.9. Theorem. The mapping v 7→ (v ◦ τ )∼ is an isometry from the BKS,
C∞ (Q, Y ) onto L0 (Š, X ) and is associated with the isomorphism e 7→ (e ◦ τ )∼ : C∞ (Q) → L0 (Š). The image of C(Q, Y ) under this isometry is L∞ (Š, X ). C We need only to prove that an almost everywhere defined section u of the bundle X is measurable if and only if u ∼ v ◦ τ for some element v ∈ C∞ (Q, Y ). In other words, it should be verified that U = L 0 (Š, X ), where
U := {u ∈ S∼ (Š, X ) : ∃v ∈ C∞ (Q, Y ) u ∼ v ◦ τ }. We confine ourself to the case of a σ-finite measure space. In view of 2.5.8 U ⊂ L 0 (Š, X ). Let us prove the reverse inclusion. ^ := A ^∼ be the clopen set in Q corresponding Let v ∈ C∞ (Q, Y ), A ∈ † and A to A under the Stone transform. Then 

^ (v ◦ τ ) = [A]v ^ ◦ τ, [A](v ◦ τ ) ∼ [ρ(A∼ )](v ◦ τ ) = τ −1 A which implies that the set U contains the fragments of all its elements. Thus, U contains also the set of all C -step-sections. Suppose that u ∈ L 0 (Š, X ). By the definition of measurability in 2.5.1, there exists a sequence of step-sections un ∼ vn ◦ τ that converges
to a section u ∈ S∼ (Š, X ) almost everywhere. Then we ∼ ∼ 0 have o-limn→∞ |||un − u||| = o-limu∼ n − u = 0 in the K-space L (Š, †, µ); thus, ∼ ∼ u∼ is an order isomorphism n − um → 0. Since the correspondence f 7→ (f ◦ τ ) 0 of vector lattices L (Š, †, µ) and C∞ (Q), it follows that o-limn→∞ vn − vm = 0 in C∞ (Q). Due to o-completeness of the LNS C∞ (Q, Y ), there exists the o-limit v ∈ C∞ (Q, Y ) of the sequence (vn ). Obviously, u ∼ v ◦ τ . B The inverse isometry from L0 (Š, X ) onto C∞ (Q, Y ) is defined by the rule ^, where u ^ is a unique section in C∞ (Q, Y ), representing u as (^ u ◦ τ )∼ . We u 7→ u ^ as the Stone transform of u. refer to the section u ^ ◦ τ is a lifting of L∞ (Š, X ) associated with ρ. Endowed The mapping u 7→ u with this lifting, the measurable Banach bundle X is ρ-stable. Theorem 2.5.9 describes a method of constructing a liftable measurable Banach bundle given a complete continuous Banach bundle over the corresponding Stone space. The following result shows that every liftable measurable Banach bundle can be obtained exactly in such a way. 2.5.10. Theorem. Let X be a ρ-invariant measurable Banach bundle over Š that has a lifting associated with ρ. Then there exists an ample continuous Banach bundle X^ over Q unique to within an isometry and such that X = X^ ◦ τ and ρ(u) = u ^ ◦ τ for all u ∈ L∞ (Š, X ). C Since X
is ρ-stationary, we may define a Banach bundle Y overτ (Š) by the formula Y τ (ω) := X (ω) and endow it with the continuity structure ρ(u) ◦ τ −1 : u ∈ L∞ (Š, X ) .

84

Chapter 2


Denote by U the vector space C b (D, Y ) with the multinorm k·kq q∈Q , kukq = u (q), where u : Q → R is a unique continuous extension of the bounded function q 7→ ku(q)k (q ∈ D). Let X^ be the CBB over Q generated by the multinormed space U and let ı : U → C(Q, X^) be corresponding canonical embedding (see 2.4.4). Since U is norm complete, stalkwise dense in X^ and invariant with respect to band projections, we may prove that ı(U ) = C(Q, X^). For each point q ∈ D we may define a linear isometry h(q) between the stalks X (q) and X^(q) satisfying the equality h(q)u(q) = ı(u)(q) for all u ∈ U . It is easy to see that h is an isometry ^ from X onto X D . Consequently, X^ is the desired continuous Banach bundle. Most of the necessary properties of the bundle X^ are easily verified. Prove that the bundle is ample. It follows from the construction of the bundle X^
that for each class u ∈ ∞ L (Š, X ), the bounded section ρ(u) ◦ τ ∈ C τ (Š), Y can be extended to an element of C(Q, X^) that will be denoted by u ^. It can be easily verified that the ^ is an isometry from the LNS L∞ (Š, X ) onto C(Q, X^) associated mapping u 7→ u e) from L∞ (Š) onto C(Q), see 1.4.9. It follows that with the isomorphism (e 7→ ^ the image of L∞ (Š, X ) under this isometry is o-dense in C(Q, X^). In view of the o-completeness of L∞ (Š, X ), this image coincides with C(Q, X^) and, hence, the LNS C(Q, X^) is o-complete as well. Uniqueness of the bundle X^ follows from 2.4.10. B We call the ample continuous Banach bundle X^ in the statement of the last theorem the Stone transform of the measurable Banach bundle X . Three facts presented below are immediate consequences of applying Theorem 2.5.9 and the representation result 2.4.9. 2.5.11. Theorem. Every Banach–Kantorovich space X over an order-dense ideal E ⊂ L0 (Š) is linearly isometric to E(X ) for some liftable measurable Banach bundle X over Š. Moreover, such a bundle X is unique to within a linear isometry. 2.5.12. Theorem. For every Banach–Kantorovich X over E and every isomorphism ı from the K-space E onto an order-dense ideal F ⊂ L0 (Š) there exist a liftable measurable Banach bundle X over Š (unique to within isometry) and some isometry from X onto F (X ) associated with ı. 2.5.13. Theorem. For every Banach–Kantorovich space X over a Kantorovich–Pinsker space, there exist a measurability structure Š, possessing the direct sum property, an order-dense ideal F ⊂ L0 (Š), and a liftable measurable Banach bundle X over Š such that the LNSs X and F (X ) are isometric. 2.6. Comments 2.6.1. (1) The concept of lattice-normed space was introduced for the first

Lattice-Normed Spaces

85

time by L. V. Kantorovich in 1936 [154]. These are vector spaces normed by elements of a vector lattice. Somewhat earlier, G. Kurepa [192] considered “espaces pseudodistanci´es,” i.e. a space with a metric that takes values in an ordered vector space. First applications of vector norms and metrics were related to the method of successive approximations in numerical analysis, see [154, 161, 163, 182, 340]. (2) It is worth stressing that [157] is the very paper in which the unusual decomposability axiom (see 2.1.1 (4)) for an abstract norm appeared for the first. Paradoxically, this axiom (4) was often omitted as inessential in the further research by other authors. The profound importance of (4) was rediscovered in connection with Boolean-valued analysis (see [196, 197]). (3) The connection between decomposability and existence of a Boolean algebra of projections in an LNS was discovered by A. G. Kusraev [196, 197]. Spaces with a fixed Boolean algebra of linear projections and a coordinated order (the so-called coordinated spaces) were studied by J. L. B. Cooper [69, 70]. Assertions 2.1.7 (1, 3) were obtained in [181]. (4) The notion of discrete element is important in the structure theory of vector lattices, see [163, 263, 388]. Discrete functionals and discrete (real-valued) measures are well studied and have a simple structure [139, 326]. As was shown by J. A. Crenshaw [71], under certain mild conditions a discrete element in the lattice of order-bounded operators is completely defined by a discrete functional on the domain vector lattice and a discrete element of the target vector lattice. Thus discrete operators comprise a poor class. At the same time there is a series of interesting results in which the concept of module discreteness plays the central role (see for instance [42, 197, 215, 220, 344, 393]). This motivates the study of concepts analogous to module discreteness, module atomicity, and module indecomposability in lattice-normed spaces. The notion of norm-n-indecomposable element from 2.1.9 and Theorem 2.1.10 are due to V. A. Radnaev [322, 323]. 2.6.2. (1) The completeness criterion 2.2.3 was stated by A. G. Kusraev in [196] under the condition that the norm lattice E is order complete. In [197], this was proven in a more general situation of spaces with decomposable vector multinorms. The assumption of order completeness for E was removed in [181]. For an Archimedean vector lattice (the case of X = E) the indicated fact was established by A. I. Veksler and V. A. Ge˘ıler [381]. (2) The concept of universal completion (maximal extension) for an arbitrary K-space was introduced and studied by A. G. Pinsker; see [163]. He established, in particular, that any K-space has a unique, to within an isomorphism, maximal extension. Assertion 2.2.8, which is a generalization of the Pinsker Theorem for LNSs, was essentially obtained in [197]. As regards Theorem 2.2.11 (1) on

86

Chapter 2

^ = oX in 2.2.11 (2) is order completion of an LNS, see [197, 218]. The assertion X due to A. E. Gutman. The properties of approximating sets and their applications to order-bounded operators were studied in [120]. (3) The operator-dual space X ∗ in 2.2.4, introduced here for constructing the maximal extension mX, is of independent interest; it was studied in [197]. In particular, an LNS X is operator-reflexive in the sense that X ∗∗ = κ(X) if and only if the set {x ∈ X : x ≤ e} is weakly cyclically compact for each element e ∈ E+ (see [197]). More about the conception of cyclical compactness will be said in Section 8.5. 2.6.3. (1) Information about measurable functions with values in a Banach space and, particularly, in the space of bounded linear operators, is presented in [74, 78, 87]. A general idea of constructing spaces in Section 1.2 can be expressed as follows: If X is a Banach (or locally convex) space and E is a function space then we may associate with them a class Z of vector-functions (measurable or continuous) by requiring that f ∈ Z if and only if l ◦f ∈ E for each l ∈ X 0 , etc. (see [54, 92, 238, 319]). This idea was elaborated while developing the theory of vector integration [51, 74, 78, 87, 91, 375]. (2) From 2.3.4 it is clear that the algebraic tensor product E ⊗ X is bo-dense in the space E(X). Denseness of E ⊗ X in E(X) with respect to the scalar norm is connected with order continuity of a Banach lattice E: see Phuong-Cac [319], A. V. Bukhvalov [55], and V. L. Levin [237, 238]. The vector norm in E ⊗ X, that is introduced in 1.2.10, is analogous to Levin’s cross-norm, see [237, 238]. Assertion 2.3.4 (1) is proven as in [238, Theorem 4.2]. (3) The tensor product of Banach lattice E and an arbitrary Banach space with cross-norm induced by the space of regular operators was introduced and learned by V. L. Levin [238]. V. T. Khudalov [169, 170] studied a similar crossnorm for a Banach space with regular cone. An important method for constructing tensor products of vector and Banach lattices was proposed by D. Fremlin [98, 99]. G. Shotaev [350, 351] constructed tensor products of lattice-normed spaces and of Banach spaces with mixed norm. (4) In the study of measurable vector-functions the following question is of interest: If X = F is a Banach ideal space of functions of a variable s then every vector-function ~f : T → X = F generates a function of two variables by the formula φ(s, t) = (~f (t))(s). However, φ(·, ·) may fail to be measurable as a function of two variables even in the simplest cases (for instance, for the measure space [0, 1] with the Lebesgue measure and X = L2 (0, 1)). Indeed, W. Sierpinski constructed an example of a subset of a square which is Lebesgue nonmeasurable and has at most two common points with every straight line. The characteristic function φ0 of this set is nonmeasurable as a function of two variables whereas the corresponding

Lattice-Normed Spaces

87

vector-function f is the zero function since φ0 (s, t) = 0 almost everywhere for all t. By adding φ0 to nonzero measurable vector-functions, we obtain similar examples with nonzero vector-functions. (5) The fact that Property (C) in F implies measurability for the function K (t) := kK(·, t)kF (see 2.3.9 (1)) was established by W. A. J. Luxemburg. It can be also deduced from one result of Yu. I. Gribanov [115]; see [162, XI.1.3, XI.1.4]. Proposition 2.3.9 (3) is essentially due to H. W. Ellis [90]. The case of an atomic measure is of no avail since then all functions are measurable. (6) A much more difficult question is the representation f (t)(·) = K(·, t) for weakly measurable functions. One of the possible solutions to this problem rests on the Bukhvalov integrality criterion 6.5.4 (see [59, 228]). If F is a Banach ideal space with order continuous norm then E(F 0 ) = Es (F 0 ). 2.6.4. (1) The invention of Banach bundles is customarily connected with the name of J. von Neumann who proposed in 1937 some ideas about varying Banach spaces. The corresponding formal descriptions appeared about 1950 in the papers of R. Godement, I. Kaplansky, and I. M. Gelfand and M. A. Na˘ımark. Presently, the theory of continuous Banach bundles is a rather wide area of research. An adequate description for the current state of this theory can be found in the surveys and references of [95]. The monograph [104] contains a detailed discussion of most of the necessary notions related to CBBs. Continuous Banach bundles are often used for representing various functional-analytical objects, see [95, 104, 131, 339]. (2) In the article [218] by A. G. Kusraev and V. Z. Strizhevski˘ı, the space E(X ) was introduced of almost global sections of a continuous Banach bundle X , and Theorem 2.4.7 was established; it was also shown that each BKS is linearly isometric to E(X ) for a suitable X; however, uniqueness of the bundle X was not established (cf. the existence result in Theorem 2.4.10). A. E. Gutman found a class of uniqueness for Theorem 1.4.7, the class of ample continuous Banach bundles, and proved Theorem 1.4.10 [123]. For a detailed presentation of these and other interesting results see [118–124]. The material of Section 2.4 shows that ample CBBs have other advantages. For instance, only after assuming the completeness of a CBB X it becomes possible to introduce the CBB L (X , Y ) and the dual CBB X 0 (2.4.13); this result is also due to A. E. Gutman [118]. 2.6.5. (1) The idea of a measurability structure has been proposed by N. Dinculeanu as early as in 1966, but has not been much studied since then. A different approach to defining measurability of sections has been prevalent so far in the papers on Banach bundles. Namely, let X be a continuous Banach bundle over a locally compact space Š with a fixed Radon measure. A section u ∈ S(Š, X ) is called measurable if, for every compact K ⊂ Š, there exists a sequence of continuous sections un ∈ S(Š, X ) converging to u almost everywhere on K.

88

Chapter 2

(2) In Section 2.5 we follow A. E. Gutman [119]. The way of introducing measurable sections in Section 2.5 is similar to Daniel’s construction and is formally more general than the traditional topological approach. As is seen from 1.5.10, liftable measurable Banach bundles in the class of all MBBs occupy in a sense the same place as complete continuous Banach bundles in the class of all CBBs. The connection between liftable measurable Banach bundles and complete CBBs, described in 2.5.8–2.5.10, enables us to transfer the facts of the theory of complete CBBs to the case of a measurable Banach bundle; see [119].

Chapter 3 Positive Operators

In this chapter we briefly present some basic results on positive operators in vector lattices. Two natural classes of operators are closely related to positive operators: the set of differences of positive operators called the space of regular operators, and the space of order-bounded operators (3.1.1). The fundamental Riesz–Kantorovich Theorem (3.1.2) claims that, under natural conditions, the space of regular operators is an order complete vector lattice and that a regular operator can be described as an order-bounded operator, i.e., an operator carrying orderbounded sets into order-bounded sets. Explicit expressions for finite and infinite lattice operations as well as for the moduli, positive and negative parts of orderbounded operators comprise the so-called order calculus. One of the approaches leans upon the concept of a generating set of projections (3.1.5). All these facts require order completeness of the target vector lattice. Order complete vector lattices appear by necessity also in the problem of dominated extension of linear operators: such an extension is possible in a mass setting if and only if the target vector lattice is order complete (the Hahn–Banach–Kantorovich Theorem (3.1.7) and the Bonnice–Silvermann–To Theorem (3.1.8)). Simplest representatives of the class of order-bounded operators are orthomorphisms (3.3.2) and lattice homomorphisms (3.3.1). Hahn–Banach–Kantorovichtype theorems are also true for lattice homomorphisms (3.3.10, 3.3.11). This fact is closely connected with the extreme structure of convex sets of operators (3.3.7– 3.3.9). Among order bounded operators are important classes of order continuous and order σ-continuous operators (3.2.1). Their antipodes, called singular and σsingular operators, are distinguish by the property of vanishing on an order-dense or σ-order-dense ideal, respectively (3.2.1). Order continuous and order σ-continuous operators comprise bands (3.2.3 (2)). Moreover, these bands are the disjoint complements of the sets of singular and σ-singular bands under the appropriate conditions (3.2.3 (1)). It is a long tradition stemming from measure theory to look for some

90

Chapter 3

explicit description of different fragments of an operator or measure. Several results in this direction are presented in 3.2.5–3.2.9 and 3.3.13. Next we consider the class of the so-called Maharam operators. A Maharam operator is a “full-valued” or order interval preserving order continuous operator. One of the main peculiarities of a Maharam operator is that the Boolean algebra of its fragments is isomorphic to the base of the domain vector lattice (3.4.5). Variants of the Hahn Decomposition Property and the Radon–Nikod´ ym Theorem are valid for Maharam operators (3.4.6 (1), 3.4.9). An important example of a Maharam operator is the transformation assigning to each order-bounded operator its restriction to a fixed massive sublattice of the domain vector lattice (3.5.2 (6)). This fact enables us to prove the existence of simultaneous extension from an arbitrary massive sublattice (3.4.11). Maharam operators may be useful in studying different classes of operators. Given a positive operator, we may always extend it to a Maharam operator by a construction very much resembling the Daniell construction of the Lebesgue integral (3.5.2, 5.5.3). The domain of the so-extended operator is then approximated in a way by the domain vector lattice of the initial operator (3.5.6–3.5.8). The band generated by the initial operator is linear and order isomorphic to that of the resulting Maharam operators (3.5.5). These results are used to obtain some description for the Boolean algebra of fragments of a positive operator by up-and-down procedures (3.5.9, 3.5.10). 3.1. Operators in Vector Lattices In this section we introduce spaces of order-bounded operators and present several fundamental results which are systematically used in the sequel. 3.1.1. Let E and F be vector lattices. A linear operator T : E → F is said to be positive provided that T (E+ ) ⊂ F+ ; regular provided that it can be represented as a difference of two positive operators; last, order-bounded or o-bounded provided that T carries every order-bounded subset in E onto an order-bounded subset in F . The set of all regular, order-bounded, and positive operators from E into F is denoted by Lr (E, F ), L∼ (E, F ), and L+ (E, F ) := L∼ (E, F )+ , respectively. The classes Lr (E, F ) and L∼ (E, F ) are vector subspaces in the vector space L(E, F ) of all linear operators from E to F , provided with the order relation: S ≥ T ⇔ S − T ≥ 0. Every positive operator is obviously order-bounded and the difference of orderbounded operators is order-bounded, too. Thus, every regular operator is orderbounded. The converse may fail but is also true in the case of order complete F . The latter follows immediately from the famous Riesz–Kantorovich Theorem to be proved in the next subsection. First we give a simple auxiliary fact.

Positive Operators

91

Let E be a vector lattice, let X be an arbitrary real vector space, and let U be an additive and positively homogeneous mapping from E+ to X: U (x + y) = U x + U y,

U (λx) = λU x (0 ≤ λ ∈ R; x, y ∈ E+ ).

Then U admits a unique linear extension T to the whole vector lattice E. If in addition F is a vector lattice and U (E+ ) ⊂ F+ then T is positive. C The uniqueness of extension is obvious from the formula x = x+ − x− ; therefore, it remains to prove existence. For every x ∈ E, we put T x = U x+ − U x− and observe that T is the sought extension. Indeed, for z = x − y, x, y ∈ E+ we have z + − z − = x − y or z + + y = x + z − . Therefore, U z + + U y = U x + U z − by hypotheses, whence T z = U z + − U z − = U x − U y = T x − T y. Now, we may conclude that T is additive on E. Additivity implies that T (−x) = −T x for all x ∈ E, which involves homogeneity of T by virtue of its positive homogeneity. B 3.1.2. Riesz–Kantorovich Theorem. If E is a vector lattice and F is some K-space then the set of order-bounded operators L∼ (E, F ) ordered by the cone of positive operators L∼ (E, F )+ is a K-space. Moreover, for every S, T ∈ L∼ (E, F ) and x ∈ E+ , the following formulas hold: (1) (S ∨ T )x = sup{Sx1 + T x2 : x1 , x2 > 0, x = x1 + x2 }; (2) (S ∧ T )x = inf{Sx1 + T x2 : x1 , x2 > 0, x = x1 + x2 }; (3) S + x = sup{Sy : 0 ≤ y ≤ x}; (4) S − x = − inf{Sy : 0 ≤ y ≤ x}; (5) |S|x = sup{|Sy| : |y| ≤ x}; Pn Pn (6) |S|x = sup { i=1 |Sxi | : x1 , . . . , xn > 0, x = i=1 xi , n ∈ N} ; (7) |Sx| ≤ |S|(|x|) (x ∈ E). C To establish that some order vector space L is a vector lattice, it suffices to check that, for every x ∈ L, there exists |x| = x ∨ (−x). Then it remains to use the formulas 1.3.2 (3). Denote by U x the right-hand side of the formula (5) and prove that the operator U : E+ → F is additive and positively homogeneous as in 2.1.1. Observe that for a fixed x ∈ E+ the set {y ∈ E : |y| ≤ x} is bounded and, since S ∈ L∼ (E, F ), its image {|Sy| : |y| ≤ x} is also bounded in F . Thus, the supremum in (5) exists and the operator U is correctly defined. Positive homogeneity of U is obvious and it remains to prove its additivity. Formula (5) can be slightly simplified. By virtue of associativity of bounds, we have U x = sup{|Sy| : |y| ≤ x} = sup{(Sy) ∨ (S(−y)) : |y| ≤ x} = sup{Sy : |y| ≤ x}.

92

Chapter 3

Take now x1 , x2 ∈ E+ and prove that U (x1 + x2 ) = U x1 + U x2 . If y1 and y2 are such that |y1 | ≤ x1 and |y2 | ≤ x2 then |y1 + y2 | ≤ x1 + x2 and Sy1 + Sy2 = S(y1 + y2 ) ≤ U (x1 + x2 ), whence U x1 + U x2 ≤ U (x1 + x2 ). Prove the reverse inequality. Let |y| ≤ x1 + x2 . Put y1 := x1 ∧ y + − x1 ∧ y − , y2 := y − y1 . It is easy to check that |y1 | ≤ x1 , |y2 | ≤ x2 , whence |Sy| = |S(y1 + y2 )| ≤ |Sy1 | + |Sy2 | ≤ U x1 + U x2 which proves additivity. In view of 3.1.1, the operator U admits a unique linear extension to E which we denote by |S|. Check that |S| = S∨(−S). From a simplified version of the formula for U it is obvious that [S] ≥ ±S and that V ≥ ±S yields V ≥ [S] by which the claim is verified. Thus, we established that L∼ (E, F ) is a vector lattice and the modulus of an operator in the vector lattice can be calculated by formula (5). Formulas (1)– (4) and (7) easily follow from (5) as was mentioned above. Prove that the vector lattice L∼ (E, F ) is order complete. To this end it suffices to prove existence of a least upper bound of an arbitrary bounded above set. Without loss of generality, we may assume that the set under study is directed upward. Thus, let M be an upward-directed set of operators and let V ∈ L∼ (E, F ) be an upper bound of it. Given x ∈ E+ , we put T x := sup{Sx : S ∈ M }. The supremum exists since Sx ≤ V x for all S ∈ M and all x ∈ E+ . Again we may apply the proposition from 3.1.1 to the operator U . Let T be an extension of U to E. Then T belongs to L∼ (E, F ), since S ≤ T ≤ V for every S ∈ M . It is now seen that T = sup M . Thus, we verified that L∼ (E, F ) is order complete. It remains to prove the formula (6). According to (7), the right-hand side in (6), say f , is less or equal to the left-hand one. To prove the reverse inequality fix x ∈ E+ and take y, |y| ≤ x. Then we have Sy = Sy + − Sy − ≤ |Sy + | + |Sy − | + |S(x − |y|)| ≤ f and the claim easily follows. B 3.1.3. Let E and F be vector lattices with F order complete. Assume that given a positive operator S : G → F on an order ideal G ⊂ E. If the set S([0, e]∩G) is order-bounded in F for every e ∈ E+ then we may define EG (S)e := sup{Sg : g ∈ G, 0 ≤ g ≤ e} := sup{S(g ∧ e) : g ∈ G} (e ∈ E+ ).

Positive Operators

93

An operator EG (S) : E+ → F is additive and positively homogeneous, so it can be extended to E by differences, see 3.1.1. The resulting operator, called the minimal extension of S, is denoted by the same symbol E (S) := EG (S). Specify some properties of minimal extensions. (1) The minimal extension E (S) agrees with S on G and vanishes on G⊥ . C It follows directly from the definition of minimal extension. B Let Lext (G, F ) denote the set of regular operator S ∈ L∼ (G, F ) such that S + and S − admit minimal extension. It is evident that S ∈ Lext (G, F ) if and only if there is a positive operator Se ∈ L∼ (E, F ), a dominant for S, such that e |S|(g) ≤ S(g) (g ∈ G+ ). Since the operator EG defined on Lext (G, F )+ is additive and positively homogeneous, it can be extended by differences as in 3.1.1 and we obtain a positive operator EG : Lext (G, F ) → L∼ (E, F ) which is called the minimal extension operator. Let RG : L∼ (E, F ) → L∼ (G, F ) be the restriction operator T 7→ T |G . The restriction operator is also positive. (2) The minimal extension operator EG preserves order continuity and sequential order continuity. C Ensure that EG preserves sequential order continuity. Take 0 < S ∈ Lext nσ (G, F ) and denote Se := EG S. For any increasing sequence (en )n∈N in E+ with supn en = e we have e = sup{S(g ∧ e) : g ∈ G} = sup{sup S(g ∧ en ) : g ∈ G} S(e) n

e n ). = sup sup{S(g ∧ en : g ∈ G} = sup S(e n

n

e n ) % S(e). e e 1 − en ) % S(e e 1 ), Thus, S(e Now, the relation en & 0 implies that S(e e n ) & 0. B whence S(e The null ideal of an operator T ∈ L∼ (E, F ) is the set N (U ) = { x ∈ E : U (|x|) = 0}. The disjoint complement N (U )⊥ is called the carrier or the band of essential positivity of the operator T and is denoted by CT . In case when N (U )⊥ = E we say that U is essentially positive. (3) The operator πG := EG ◦ RG is a band projection in L∼ (E, F ) and ⊥ ∼ πG L (E, F ) = {T ∈ L∼ (E, F ) : G ⊂ N (T )}.

C It is sufficient to prove that S := πG T is a fragment of T for every positive T ∈ L∼ (E, F ). Let V := S ∧ (T − S). If x ∈ E+ then Sxα % Sx for some increasing net (xα ) ⊂ G according to the definition of πG . Since V ≤ S, we have V (x − xα ) ≤ S(x − xα ) & 0. In virtue of (1) V vanishes on G; therefore, V x = o-lim V xα = 0. B

94

Chapter 3

(4) The subspace Lext (G, F ) is a band in L∼ (G, F ) linearly and latticially isomorphic to πG L∼ (E, F ) under the minimal extension operator. 3.1.4. Thus, we have constructed an isotonic mapping G 7→ πG from the lattice I (E) to the Boolean algebra P(L∼ (E, F )). Now, consider some properties of this mapping. In case G = e⊥⊥ we put π[e] := πG . (1) Let π be a band projection in E and G = π(E). Then πG T = T ◦ π for every T ∈ L∼ (E, F ). In particular, π[e] T = T ◦ [e], where [e] stands for the band projection onto {e}⊥⊥ . C According to the definition of minimal extension πG T x = sup{T (π(g ∧ x)) : g ∈ G} = sup{T (g ∧ π(x)) : g ∈ G} = T (πx). B (2) Suppose that C and D are order ideals in E and G = C ∩ D. Then πG = πC ∧ πD . In particular, πG and πD are disjoint if and only if C ∩ D = {0}. C Since C ∩ D = {c ∧ d : c ∈ C, d ∈ D}, for 0 ≤ x ∈ E and 0 ≤ T ∈ L∼ (E, F ) we deduce πG T (x) = sup{T (x ∧ c ∧ d) : c ∈ C, d ∈ D} = sup sup{T (x ∧ c ∧ d) : d ∈ D} = πC (πD T )(x). c∈C

Thus, πG = πC ◦ πD = πC ∧ πD . B (3) Suppose that C and D are order ideals in E and G = C + D. Then πG = πC ∨ πD . C Monotonicity of G 7→ πG implies πG ≥ πC ∨ πD . If u ∈ [0, x] ∩ (C + D) then u = c + d for some c ∈ C+ and d ∈ D+ and for x1 := c, x2 := x − c we have x = x1 + x2 , c ∈ [0, x1 ] ∩ C, and d ∈ [0, x2 ] ∩ D. Therefore, T u = T c + T d ≤ πC T (x1 ) + πD T (x2 ) ≤ (πD T ∨ πD T )x ≤ (πD ∨ πD )T x and πG T (x) = sup T u ≤ (πD ∨ πD )T x. B S (4) Let (Gα ) be an increasing family of order ideals in E and G := Gα . Then G is also an order ideal and πG = supα πGα . Moreover, the representation holds: ⊥ ∼ πG L (E, F ) = {T ∈ L∼ (E, F ) : (∃α) Gα ⊂ N (T )}. C This follows immediately from the definitions on using associativity of least upper bounds: (πG T )x = sup T (g ∧ x) = sup sup T (g ∧ x) = (sup πGα T )x. B g∈G

α g∈Gα

α

If G := E(e) is the ideal generated by e ∈ E+ then we write πe instead of πG .

Positive Operators

95

(5) The following representation for πe is valid: πe T x = sup T (ne ∧ x)

(x ∈ E + , T ∈ L+ (E, F )),

n

πe T x = πe T x+ − πe T x− πe T = πe T + − πe T −

(x ∈ E, T ∈ L+ (E, F )), (T ∈ L∼ (E, F )).

3.1.5. Very often it is useful to have formulas for the calculation of (T ∨S)x and (T ∧ S)x using the fragments of T and S. Different approaches to this problem can be unified by means of the following notion. A set of projections P ⊂ P(L∼ (E, F )) is said to be generating if for all T ∈ L+ (E, F ) and x ∈ E we have T x+ = sup{pT x : p ∈ P}. As an easy example we cite the following. To each band projection π ∈ P(E) assign π T 7→ T ◦ π acting in L∼ (E, F ) and denoted by P ◦ the set of the band projection ^ all such band projections. (1) Put P π := {πe : e ∈ E+ }. Then P π is a generating set of projec∼ tions in L (E, F ). C Indeed, if e := x+ then πe T x+ = T x+ and πe T x− = 0 by 3.1.3 (1); therefore πe T x = πe T x+ − πe T x− = T x+ . B (2) If a vector lattice E has the strong Freudenthal property then P ◦ is a generating set of projections in L∼ (E, F ). C According to 1.3.9 (2) the strong Freudenthal property implies that for every π T x for all x ∈ E there is a band projection π with x+ = πx. Therefore, T (x+ ) = ^ T ∈ L∼ (E, F ). B (3) Theorem. Let E and F be vector lattices with F order complete. A set P of band projections in L∼ (E, F ) is generating if and only if for any T, S ∈ L∼ (E, F ) and x ∈ E + the following formulas hold: (T ∧ S)x = inf{pT x + p⊥ Sx : p ∈ P}, (T ∨ S)x = sup{pT x + p⊥ Sx : p ∈ P}. C We confine exposition to sufficiency. Let P be a generating set of projections. In virtue of 1.3.2 (10) and 1.3.2 (1) it is sufficient to prove inf{pT x+p⊥ Sx} = 0 provided that T ∧ S = 0. First observe that V x+ − sup{pV x : p ∈ P} = V x+ + inf{−pV x+ + pV x− : p ∈ P} = inf{p⊥ V x+ + pV x− : p ∈ P}.

96

Chapter 3

Thus, the identities inf{pV x+ + p⊥ V x− : p ∈ P} = 0 and V x+ = sup{pV x : p ∈ P} are equivalent for all V ∈ L+ (E, F ) and x ∈ E. Note next that pT x = p(T (2y) + T (x − 2y)) ≤ 2T y + pT (x − 2y)− , p⊥ Sx = p⊥ (S(2x − 2y) + S(2y − x)) ≤ 2S(x − y) + p⊥ S(2y − x)+ for p ∈ P and 0 ≤ y ≤ x, x ∈ E + . Adding these inequalities and passing to the infimum we deduce inf {pT x + p⊥ Sx} ≤ 2 inf {T y + S(x − y)} p

y

+ inf {p(T + S)(2y − x)− + p⊥ (T + S)(2y − x)+ } = 0 p

and the result follows. B 3.1.6. Observe the following useful corollary to 3.1.2 and 3.1.5 (2). (1) If E has the strong Freudenthal property then it is possible to calculate the supremum in 3.1.2 (1), 3.1.2 (6), and the infimum in 3.1.2 (2) over partitions of x ∈ E+ into the disjoint sum: x1 ⊥ x2 in 3.1.2 (1, 2) and xk ⊥ xl (k 6= l) in 3.1.2 (6). C The assertions concerning 3.1.2 (1, 2) follow immediately from 3.1.5 (2, 3). Hence, we may write |T |x := (T ∨ (−T ))x = sup{T x1 − T x2 : x = x1 + x2 , x1 ⊥ x2 , x1 , x2 ∈ E+ } ≤ sup{|T x1 | + |T x2 | : x = x1 + x2 , x1 ⊥ x2 , x1 , x2 ∈ E+ }. This proves the inequality ≤ in 3.1.2 (5), while the reverse inequality is obvious. B (2) Let E and F be the same as in Theorem 3.1.2. Then operators T + and S in L (E, F ) are disjoint if and only if for every e ∈ E+ and 0 < ε ∈ R there exist a partition of unity πα ⊂ P(F ) and a family (eα ) ⊂ E, 0 ≤ eα ≤ e, such that for all α the inequalities hold: πα T eα ≤ εT e,

πα S(e − eα ) ≤ εSe.

If, in addition, E has the strong Freudenthal property then we may assume that (eα ) ⊂ E(e), i.e., eα is a fragment of e for every α. C Suppose that (S ∧ T )e := inf{T e + S(e − x) : 0 ≤ x ≤ e} = 0

Positive Operators

97

for some e ∈ E+ . Put u := Se∧T e+π(Se+T e), where π is the band projection onto {Se ∧ T e}⊥ . As is seen, Se + T e ∈ {u}⊥⊥ , [Se]u ≤ Se and [T e]u ≤ T e (as before, [a] stands for the band projection onto {a}⊥⊥ ). According to 1.4.6 (4) and (1) there are a partition of unity πα ⊂ P(F ) and a family (eα ) ⊂ [0, e] ⊂ E such that πα (T eα +S(e−eα )) ≤ εu for all α. Thus, πα T eα ≤ ε[T e]u and πα S(e−eα ) ≤ ε[Se]u. The converse is obvious. B (3) Let E be a vector lattice and let P be a generating set of band projections in L∼ (E, F ). Then operators T and S in L+ (E, F ) are disjoint if and only if for every e ∈ E+ and 0 < ε ∈ R there exist a partition of unity πα ⊂ P(F ) and a family (pα ) ⊂ P, such that for all α the inequalities hold: πα pα T e ≤ εT e,

πα p⊥ α Se ≤ εSe.

3.1.7. We will conclude this section with the problem of extension of positive operators and dominated extension of linear operators. Kantorovich Theorem. Let X be a preordered vector space, let X0 be a massive subspace in X, and let F be a K-space. Then each positive operator T0 : X0 → F has a positive extension T : X → F . C First, let X = X0 ⊕ X1 , where X1 is a one-dimensional subspace, X1 := {αx1 : α ∈ R}. Since X0 is massive and T0 is positive, the set U := {T0 x0 : x0 ∈ X0 , x0 ≥ x1 } is nonempty and bounded below. Assign T x := {T0 x0 + αf : x = x0 + αx1 , x0 ∈ X0 , α ∈ R} where f := inf U . It is clear that T is a linear operator from X to F and T |X0 = T0 . So, only the positivity of T needs checking. If x = x0 + αx1 and x ≥ 0, then the case of α equal to 0 is trivial. If α > 0 then x1 ≥ −x0 /α. This implies that −T0 x0 /α ≤ f , i.e., T x ∈ F+ . In a similar way for α < 0 observe that x1 ≤ −x0 /α. Thus, f ≤ −T0 x0 /α and, finally, T x = T0 x0 + αf ∈ F+ . Now let S be the collection of linear operators S : dom(S) → F such that S extends T0 and S(dom(S)+ ) ⊂ F+ . Clearly, S is inductive in order by inclusion and so, by the Kuratowski–Zorn Lemma, S has a maximal element T . If x1 ∈ X \ dom(T ), apply the above proved result with X := dom(T ) ⊕ X1 , X0 := dom(T ), T0 := T , and X1 := Rx1 to obtain an extension of T . But this contradicts the maximality of T ; thus, T is a sought operator. B 3.1.8. Let X be an arbitrary real vector space. An operator p : X → E is called sublinear if p(x + y) ≤ p(x) + p(y) and p(λx) = λp(x) for all x, y ∈ X and 0 ≤ λ ∈ R. The collection of all linear operators from X into E dominated by p is called the support set of p and denoted by ∂p; symbolically, ∂p := {T ∈ L(X, E) : (∀x ∈ X) T x ≤ p(x)},

98

Chapter 3

where L(X, E) is the space of all linear operators from X to E. A member of ∂p is a supporting operator of p. The epigraph → /− >> /(p) of p is defined by → /− >> /(p) := {(x, e) ∈ X × E : e ≥ p(x)}. The epigraph of any sublinear operator p is a cone and defines in X ×E a preorder relation, so that (x, e) ≤ (x0 , e0 ) if and only if e0 − e ≥ p(x − x0 ). (1) Let U ∈ L(E, F ) and V ∈ L(X, F ). Then the operator B : (x, e) 7→ V e−U x is positive on the space X ×E endowed with the above-mentioned preorder if and only if V ≥ 0 and U ∈ ∂(V ◦ P ). C Suppose B ≥ 0. Since (e, 0) ≥ 0 for every e ∈ E+ , we conclude V e = B(0, e) ≥ 0. Thus, for (x, e) ≥ 0 we have V e − U x ≥ 0 and U x ≤ V e ≤ V p(x). The converse is similar. B Assume that X0 is a subspace of X and T0 : X0 → E is a linear operator such that T0 x ≤ p(x) for all x ∈ X0 . If for every such X, X0 , T0 and p, there exists an operator T ∈ ∂p that is an extension of T0 from X0 to the whole of X, then we say that E admits dominated extension of linear operators or have the dominated extension property. A complete characterization of ordered vector spaces admitting dominated extension of linear operators resides in the following two theorems. (2) Hahn–Banach–Kantorovich Theorem. Every order complete vector lattice has the dominated extension property. C The claim follows from the Kantorovich Theorem on using (1) with V = IF . B (3) Bonnice–Silvermann–To Theorem. Every ordered vector space admitting dominated extension of linear operators is order complete. 3.1.9. In the following corollaries E is a K-space and p is a sublinear operator acting from a vector space X into E. (1) Each sublinear operator is the upper envelope of its support set, i.e. the next representation holds p(x) = sup{T x : T ∈ ∂p} (x ∈ X). Moreover, for an arbitrary point x0 ∈ X there exists a linear operator T from X into E supporting p at x0 , i.e. such that T x0 = p(x0 ) and T ∈ ∂p. C Put X0 = {λx0 : λ ∈ R} and define the linear operator T0 : X0 → E by T (λx0 ) := λp(x0 ). For λ ≥ 0 we have T (λx0 ) = λp(x0 ) = p(λx0 ). If λ < 0 then T0 (λx0 ) = λp(x0 ) = −|λ|p(x0 ) ≤ p(−|λ|x0 ) = p(λx0 ). Thus T0 supports the restriction p  X0 . By the Hahn–Banach–Kantorovich Theorem there exists an extension T of the operator T0 to the whole space X dominated by p on X. This ensures that T ∈ ∂p and T x0 = T0 x0 = p(x0 ). B

Positive Operators

99

(2) Let Y be another vector space and let T be a linear operator from Y into X. Then ∂(p ◦ T ) = ∂p ◦ T. C Consider an arbitrary element S from ∂(p ◦ T ). Clearly, −p(T (−y)) ≤ Sy ≤ p(T y). Therefore, T y = 0 implies Sy = 0. This means that ker(T ) ⊂ ker(S), where ker(R) := R−1 (0) is the kernel of the operator R. Consequently, the equation X ◦ T = S is solvable for an unknown linear operator X : T (Y ) → E. By assumption, every solution U0 to this equation satisfies the inequality U0 x0 ≤ p(x0 ) for all x0 ∈ X0 := T (Y ). Therefore, according to the Hahn–Banach–Kantorovich Theorem, there exists an extension U ∈ L(X, E) of the operator U0 supporting the sublinear operator p. Thus U ∈ ∂p and U ◦ T = S, i.e. S ∈ ∂p ◦ T . The reverse inclusion is obvious. B It should be stressed that if T is an identical embedding of a subspace X0 into the space X, then the proposition exactly expresses the dominated extension property. In this connection Proposition 1.4.14 (2) is often referred to as the Hahn– Banach formula. (3) Let X0 be a massive sublattice of a vector lattice X and let S0 : X0 → E be a positive operator with S0 x0 ≤ T x0 (x0 ∈ X0 ). Then there exists a positive extension S : X → E of S0 such that S ≤ T . C Put p(x) := T (x+ ) (x ∈ X). Then ∂p = [0, T ] and S0 x0 ≤ S0 (x+ 0) ≤ + T (x+ ) = p(x ) for all x ∈ X . By the Hahn–Banach–Kantorovich Theorem there 0 0 0 0 is an extension S of S0 with S ∈ ∂p. B 3.2. Fragments of a Positive Operator The aim of the section is to obtain some basic formulas for calculating specific fragments of a positive operator. Simple fragments associated with order ideals in the domain of the operator prove to be very useful for this purpose. 3.2.1. An operator T : E → F is called order continuous (order σ-continuous) if T xα order converges to T x for each net (xα )α∈A (each sequence (xα )α∈N ) in E with order limit x. The set of all order continuous regular operators (order σ-continuous regular operators) with the vector and order structure induced from ∼ ∼ L∼ (E, F ) is denoted by L∼ n (E, F ) (Lnσ (E, F )). If F = R then we shall write En rather than L∼ n (E, R), see 1.5.2. A positive operator T ∈ L∼ (E, F ) is order continuous (order σ-continuous) if o and only if T xα → 0 for every decreasing net (sequence) (xα ) in E with inf α xα = 0. C The proof can be found in [388] (Lemmas VIII.3.1 and VIII.4.3). B An ideal G in E is said to be a σ-order-dense ideal if for any e ∈ E+ there is an increasing sequence (gn ) ⊂ G+ such that e = sup gn . An operator T ∈ L∼ (E, F ) is

100

Chapter 3

called singular if it vanishes on some order-dense ideal G ⊂ E. If T vanishes on a σorder-dense ideal, then U is called σ-singular. The sets of singular and σ-singular ∼ operators are denoted by L∼ s (E, F ) and Lsσ (E, F ), respectively. It is easily seen ∼ ∼ from 3.1.2 (5, 7) that Ls (E, F ) and Lsσ (E, F ) are order ideals in L∼ (E, F ). We say that a vector lattice E is rich in σ-order-dense ideals if for any x ∈ E+ there exists an element e ∈ E+ such that x ∈ {e}⊥⊥ and {e}⊥ + {e}⊥⊥ is a σorder-dense ideal in E. Each of the following types of vector lattices is rich in σ-order-dense ideals: (a) vector lattice with a weak order-unity; (b) vector lattice with the countable sup property; (c) vector lattice with the principal projection property. (d) vector lattice E in which for any x ∈ E there exists a countable antichain (en ), ek ⊥ ej (j 6= k), such that {en : n ∈ N}⊥ = E and x ∈ {e1 }⊥⊥ . 3.2.2. Theorem. Let E and F be vector lattices with E rich in σ-order-dense ideals and F order complete. An operator T ∈ L∼ (E, F ) is σ-order continuous if and only if it is disjoint to all σ-singular operators: ∼ ⊥ L∼ σ (E, F ) = Lsσ (E, F ) . ⊥ ∼ is true without any additional asC The inclusion L∼ nσ (E, F ) ⊂ Lsσ (E, F ) ∼ sumption on E. Indeed, if 0 ≤ U ∈ Lnσ (E, F ) and 0 ≤ V ∈ L∼ sσ (E, F ), then T := U ∧ V ≥ 0 is σ-continuous and vanishes on a σ-order-dense ideal G ⊂ E; thus, ∼ T = 0. The sets L∼ nσ (E, F ) and Lsσ (E, F ) are thus seen to be disjoint, since they are order ideals. ⊥ Now, suppose that an operator 0 < U ∈ L∼ nσ (E, F ) is not σ-order continuous. Take a sequence xn & 0, with f := inf n U xn > 0. By assumption there exists an element e ∈ E+ such that g1 ∈ {e}⊥⊥ and G = {e}⊥ + {e}⊥⊥ is a σ-order-dense ideal. It is no restriction to assume that x1 = e, since we may put e := e ∨ x1 if necessary. We will find a nonzero fragment of U vanishing on some σ-order-dense ideal. Since U (e) ≥ f > 0, we may choose a number 0 < ε < 1 and a nonzero band projection ρ in F such that ρf > ερU (e) > 0. Let en := (xn − εe)+ , πn := πen , and V := inf n ρ(πn U ), where πen is the band projection defined in 3.1.4 (5). Clearly en & 0, the sequence (πn ) is decreasing, and V is a fragment of U . Moreover, this fragment is nonzero, since

V (e1 ) = inf sup ρU (men ∧ e1 ) ≥ inf ρU (en ) ≥ ρ(f − εU (e)) > 0, n

m

n

Prove that V vanishes on the σ-order-dense ideal G generated by the increasing sequence gn = (εe − xn )+ and the order ideal {e}⊥ , in symbols: [ G = {e}⊥ + {E(gn )}.

Positive Operators

101

For x := gn we have V x ≤ πn U (gn ) = sup U (men ∧ gn ) = 0, m

since en and gn are disjoint. If 0 ≤ x ∈ {e}⊥ = {x}⊥ 1 , then V x ≤ π1 U (x) = 0. Thus V |G = 0. It remains to check that G is a σ-order-dense ideal. For an arbitrary x ∈ E+ we may find increasing sequences (e00n ) ⊂ {e}⊥ and (e0n ) ⊂ {e}⊥⊥ such that x = supn (e00n + e0n ). Moreover, e = supk gk /ε and e0n = supm me ∧ e0n (k, n, m ∈ N), whence x = sup (e00n + (mgk /ε) ∧ e0n ) = sup(e00n + (ngn /ε) ∧ e0n ). B n

m,n,k

3.2.3. Let E and F be vector lattices with F order complete. (1) Theorem. An operator T ∈ L∼ (E, F ) is order continuous if and only if it is disjoint from all singular operators: ∼ ⊥ L∼ n (E, F ) = Ls (E, F ) .

C The proof can be obtained as a simple modification of the reasoning of 3.2.2, on replacing the sequences xn and en by nets xα and eα = (xα − εxα0 ), α ≥ α0 , and taking into consideration that the ideal {xα0 }⊥⊥ + {xα0 }⊥ is always order-dense. B ∼ ∼ (2) The spaces L∼ n (E, F ) and Lnσ (E, F ) are bands in L (E, F ). C The statement concerning L∼ n (E, F ) is immediate from (1). To prove the ∼ second part we denote by L[e] (E, F ) the set of all regular operators U such that the ⊥⊥ restriction of U to {e}⊥⊥ lies in L∼ , F ). According to 3.1.3 (4) and 3.2.2 nσ ({e} ∼ L[e] (E, F ) is a band. Now, it remains to observe that L∼ nσ (E, F ) =

\

L∼ [e] (E, F ). B

e∈E+

^ be an order completion of E. Then every positive order (3) Let E ^ → F which continuous operator T : E → F has a unique positive extension T^ : E ^ is also order continuous. The correspondence T 7→ T extends by differences to ∼ ^ an isomorphism from L∼ n (E, F ) onto Ln (E, F ). ^ According to (1) it C Let T^ be an arbitrary positive extension of T to E. ∼ ^ F ). Note that if 0 ≤ S^ ∈ L∼ (E, ^ F ) then S := S| ^E suffices to prove that T^ ⊥ Ls (E, s is also singular, since G ∩ E is an order-dense ideal in E for every order-dense ideal ^ Thus, T ⊥ S. Using 3.1.2 (2), it is easy to observe that S^ ⊥ T^. B G ⊂ E.

102

Chapter 3

3.2.4. Let F be a filter in the lattice I (E) of order ideals in E. Then the set of F -singular operators LF (E, F ) = {T ∈ L∼ (E, F ) : NT ∈ F } is an order ideal in L∼ (E, F ). (1) The operator πF = inf{πG : G ∈ F } is the band projection onto LF (E, F )⊥ . C By definition πF is a band projection in L∼ (E, F ). Denote by π the projection onto LF (E, F )⊥ . If T ∈ LF (E, F ) and G := NT , then πF T ≤ πG T = 0. Thus πF ≤ π. Conversely, suppose that T ∈ LF (E, F )⊥ and G ∈ F . Then T − πG T is contained in LF (E, F )⊥ . At the same time T − πG T ∈ LF (E, F ), since T − πG T vanishes on G. Thus, T − πG T ∈ LF (E, F ) ∩ LF (E, F )⊥ = 0. Thus, we conclude that T − πG T = 0, whence πF ≥ π. B Denote by Id (E) and Iσd (E) the sets of all order-dense ideals and σ-orderdense ideals in E. Assume that Tn and Tnσ stand for o-continuous and σ-order continuous parts of T . If we take in (1) F := Id (E) or F := Iσd (E) then, by virtue of 3.2.2 and 3.2.3, we arrive at the following: (2) For an order continuous part of a positive operator T we have Tn = inf{πG T : G ∈ Id (E)}. (3) If, in addition, a vector lattice E is rich in σ-order-dense ideals then the formula is also valid: Tnσ = inf{πG T : G ∈ Iσd (E)}. C According to 3.2.4 (1) the operators πn and πnσ , defined by πn = inf{πG : G ∈ Id (E)},

πnσ = inf{πG : G ∈ Iσd (E)},

⊥ ⊥ are the band projections onto L∼ and L∼ respectively. But we s (E, F ) sσ (E, F ) ∼ ∼ ⊥ know that Ln (E, F ) = Ls (E, F ) and if E is rich in σ-order-dense ideals then we ∼ ⊥ have L∼ σ (E, F )= Lsσ (E, F ) . Therefore πn T = Tn and πnσ T = Tnσ . B

Positive Operators

103

3.2.5. Using projections of type πG , we may localize assumptions in Theorem 3.2.2 and deduce further formulas for calculating σ-order continuous parts. ∼ Let L∼ sσ(G) (E, F ) be a subset of L (E, F ) consisting of the operators vanishing on σ-order-dense ideals in G, i.e., Lsσ(G) (E, F ) = {T : NT ∩ G ∈ Iσd (G)}. (1) Let G ∈ I (E) be an ideal rich in σ-order-dense ideals. For T ∈ L (E, F ) the operator πG T is σ-order continuous if and only if it is disjoint from all operators σ-singular on G: ∼

∼ ⊥ πG L ∼ nσ (E, F ) = πG Lsσ(G) (E, F ) . ⊥ ∼ C If G ∈ I (E) is rich in σ-order-dense ideals then L∼ nσ (G, F ) = Lsσ (G, F ) ext ⊥ by Theorem 3.2.2. Therefore, Lext nσ (G, F ) = Lsσ (G, F ) , where the disjoint comext plement is taken in L (G, F ). Indeed, the inclusion ⊂ is obvious. If 0 ≤ T ∈ ∼ Lext (G, F ), T ⊥ Lext sσ (G, F ), and an operator 0 ≤ S ≤ T is such that S ∈ Lsσ (G, F ), ext then actually S ∈ Lsσ (G, F ) and consequently S = 0. This proves the converse inclusion. Using 3.1.3 (2, 4), we may now deduce

ext ext ⊥ πG L∼ = πG Lsσ(G) (E, F )⊥ . B nσ (E, F ) = EG Lnσ (G, F ) = EG Lsσ (G, F )

For every 0 ≤ T ∈ L∼ (E, F ) the following representations hold: (2) πG Tnσ = inf{πD T : D ∈ Iσd (G)}; (3) Tnσ x = inf{πD T x : D ∈ Iσd (G)} (x ∈ G+ ); (4) Tnσ = supe∈E+ inf{πD T : D ∈ Iσd ({e}⊥⊥ )}. C By 3.2.4 (1) the operator π = inf{πD : D ∩ G ∈ Iσd (G)} is the band projection onto Lsσ(G) (E, F )⊥ . Therefore, πG π = inf{πG πD : D ∩ G ∈ Iσd (G)} = inf{πC : C ∈ Iσd (G)} is the band projection onto πG Lsσ(G) (E, F )⊥ . This proves (2). The formula (3) follows from 3.2.1 (a) and (4) is true, since {e}⊥⊥ is rich in σ-order-dense ideals for every e ∈ E+ . B 3.2.6. (1) For any operator 0 ≤ T ∈ L∼ (E, F ) the following representations hold: Tn x = inf{sup T xα : xα % x} (x ∈ E+ ), α

Tnσ x = inf{sup T xn : xn % x} (x ∈ E+ ). n

104

Chapter 3

C Denote by h(T, x) the right-hand side of the second formula. Choose a sequence xn % x and consider the set G (xn ) of all order ideals in {x}⊥⊥ containing (xn ). By definition supn T xn ≤ πG T x for each G ∈ G (xn ), whence h(T, x) ≤ inf{πG T x : G ∈ G (xn )}. At the same time each σ-order-dense ideal G contains a sequence x0n % x and, by definition, G ⊂ G (x0n ). This implies, in virtue of 3.2.4 (3), that h(T, x) ≤ Tnσ x, thereby 0 ≤ h(T − Tnσ , x) ≤ (T − Tnσ )nσ x = 0. Moreover, h(Tnσ , x) = Tnσ x, since Tnσ is sequentially order continuous. Using the additivity of h(·, x), we obtain h(T, x) = h(Tnσ , x) + h(T − Tnσ , x) = Tnσ x. B (2) If E is a vector lattice with the strong Freudenthal property then for any 0 ≤ T ∈ L∼ (E, F ) we have Tn x = inf{sup T πα x : πα % [x]} (x ∈ E+ ), α

Tnσ x = inf{sup T πn x : πn % [x]} (x ∈ E+ ), n

where πα , πn are band projections in E and [x] is the band projection onto {x}⊥⊥ . C The proof is in a similar spirit. B 3.2.7. Theorem. Let S be an upward-directed set in L∼ (E, F )+ and σ denote the band projection onto S ⊥⊥ . Then for every 0 ≤ T ∈ L∼ (E, F ) and e ∈ E+ the following representations hold: (σ ⊥ T )e = inf sup{πT x : π ∈ P(F ), 0 ≤ x ≤ e, πSx ≤ εSe}, 0 0 for all π1 6= 0, π1 ⊥ π0 . Finally let [x] be the band projection onto the band generated by the element x. Then π := π0⊥ ◦ [x] is a sought projection. Indeed, T ◦ πx = T ◦ π0⊥ x and T x = T ◦ π0⊥ x − (−T ◦ π0 x). It follows that T ◦ π0⊥ x ≥ (T x)+ > 0. On the other hand, if 0 ≤ y ∈ {x}⊥⊥ and (eyλ )λ∈R is the spectral function (or characteristic) of the element y relative to x then eyλ = 0 for λ < 0 and, for λ ≥ 0, we have T ◦ π(eyλ ) = T ◦ π0⊥ (eyλ ) = T ◦ π0⊥ ◦ [eyλ ]x ≥ 0. Appealing to the Freudenthal Spectral Theorem, we finally obtain: ∞  ∞ Z Z y  T ◦ π(y) = T ◦ π λ deλ λ d(T ◦ π(eyλ ) ) ≥ 0. B 0

0

3.4.5. Theorem. Let E and F be some K-spaces and let T : E → F be a Maharam operator. Then there exists an isomorphism ϕ from the Boolean algebra E (T ) of fragments of T onto P (ET ) such that T ◦ ϕ(S) = S for all S ∈ E (T ). C Again we may assume that E = ET . Let T0 stand for the unique o-continuous extension of the operator T to Dm (T ). Then, for each operator S ∈ E (T0 ), its restriction to E is an isomorphism of the Boolean algebras E (T0 ) and E (T ). Similarly, the Boolean algebras Pr (E) and Pr (Dm (T ) ) are isomorphic. Thus, without loss of generality, we may assume E = Dm (T ). Assign to each projection π ∈ Pr (E) the operator ψ (π) := T ◦ π. Then ψ is an increasing mapping from Pr (T ) into {T }⊥⊥ and ψ(0) = 0 and ψ(IE ) = T . Clearly, if the projections π and ρ are disjoint then the carriers of the operators ψ(π) and ψ(ρ) are disjoint too; therefore, ψ(π) ⊥ ψ(ρ). Moreover, for π ∈ Pr (E), the equalities ψ(IE − π) = T ◦ (IE − π) = T − T ◦ π = T − ψ(π) are valid. Consequently, ψ(π ⊥ ) = ψ(π)⊥ . Thus, ψ(π) ∈ E (T ) for all π ∈ Pr (E). Consider now two arbitrary projections π1 and π2 ∈ Pr (E). Since the projections ρl := πl − π1 ◦ π2 (l := 1, 2) are disjoint, the operators ψ(ρ1 ) and ψ(ρ2 ) are disjoint too. On the other hand, ψ(π1 ) ∧ ψ(π2 ) − ψ(π1 ∧ π2 ) = T ◦ π1 ∧ T ◦ π2 − T ◦ π1 ◦ π2 = (T ◦ π1 − T ◦ π1 ◦ π2 ) ∧ (T ◦ π2 − T ◦ π1 ◦ π2 ) = ψ(ρ1 ) ∧ ψ(ρ2 ) = 0.

122

Chapter 3

Hence, ψ(π1 ∧ π2 ) = ψ(π1 ) ∧ ψ(π2 ). Thus, ψ is a homomorphism from the Boolean algebra Pr (E) into the Boolean algebra E (T ). Essential positivity of the operator T implies that if ψ(π) = 0 for some π ∈ Pr (E) then π = 0. This means that ψ indeed is a monomorphism and it remains to establish surjectivity for it. Let S ∈ E (T ) and look at the set … := {π ∈ Orth (E)+ : T ◦ π ≤ S}. Using the Kuratowski–Zorn Lemma, we now demonstrate that … contains a maximal element. Indeed, … is nonempty and, for a linearly ordered set (πξ )ξ∈„ in … the set (T ◦ πξ )ξ∈„ is bounded since it is included [0, S]. But then the assumption E = Dm (T ) implies that (πξ )ξ∈„ is a bounded set. If π0 := sup {πξ : ξ ∈ „} then π0 = o- lim πξ and, by order continuity of the operator T , we have: T ◦ π0 = T ◦ (o- lim πξ ) = o- lim T ◦ πξ ≤ S, i.e. π0 ∈ …. Thus, there is a maximal element π ∈ … in the set …. Show that T ◦ π = S. To this end, suppose the contrary and let the operator S1 := S − T ◦ π take strictly positive value at some 0 < x0 ∈ E. Then, for a suitable 0 < ε < 1 and 0 6= ρ ∈ Pr (F ), we have ρ (S1 x0 − ερ ◦ T x0 ) > 0. The operator ρ ◦ |S1 − εT | is absolutely continuous with respect to T and, by Theorem 3.4.3, is a Maharam operator. According to 3.4.4, there exists a projection πε ∈ Pr (E) such that (S1 − εT ) ◦ πε x0 > 0 and (S1 − εT ) ◦ πε ≥ 0. The former relations implies that πε > 0 and, from the latter, we have T (π + επε ) ≤ S. Thus, π < π + επε ∈ …, which contradicts the maximality of π in …. This substantiates the relation S = T ◦ π. Further, by assumption, S ∧ (T − S) = 0; hence, 0 = (T ◦ π) ∧ (T ◦ (IX − π) ) ≥ T (π ∧ (IX − π) ) ≥ 0. In view of essential positivity of T , the last leads to the equality π ∧ (IE − π) = ∅ which is equivalent to the containment π ∈ Pr (E). The surjectivity of ψ is thus proven. It remains to observe that ϕ := ψ −1 is the isomorphism sought since T ◦ ϕ(S) = ψ ◦ ϕ (S) = S. B 3.4.6. Observe the following corollary to Theorems 3.4.4 and 3.4.5. (1) Let E, F , and T be the same as in Theorem 3.4.3. Then operators S1 and S2 of the band {T }⊥⊥ are disjoint if and only if their carriers CS1 and CS2 are disjoint. C Disjointness of the carriers CS1 and CS2 obviously implies that of the operators S1 and S2 (this fact does not depend on the Maharam property and is valid for arbitrary regular operators). To prove the converse, we first observe that if T1 and T2 are positive o-continuous operators and T1 ∈ {T2 }⊥⊥ then CT1 ⊂ CT2 . Indeed, assuming the contrary, we may find a projection π such that 0 < T1 ◦ π ≤ T1 and CT1 ◦π ⊥ CT2 , which contradicts the relation T1 ∈ {T2 }⊥⊥ by virtue of the previous remark.

Positive Operators

123

Let now S1 and S2 be disjoint. Then the band projections T1 and T2 of T onto {S1 }⊥⊥ and {S2 }⊥⊥ are disjoint too. On the other hand, CSl = CTl by the previous remarks. By Theorem 3.4.5, CT1 ⊥ CT2 has to be valid; therefore, CS1 ⊥ CS2 . B (2) Hahn Decomposition Theorem. Let E and F be some K-spaces and let T : E → F be an order-bounded operator with |T | a Maharam operator. Then there exists a band projection π ∈ P(E) such that T + = T ◦ π and T − = T ◦ π⊥ . C Apply Theorem 3.4.5 to the operator |T | and put π := ϕ(T + ). B 3.4.7. In the sequel, one more fact is needed on representations of order continuous operators. Let E and F be some K-spaces; and let m(E) be, as usual, the universal completion of the space E with a fixed ring and order-unity 1. Suppose that, on some order-dense ideal D (ˆ) ⊂ m(E), an essentially positive Maharam operator ˆ is defined which acts on F and D (ˆ) = Dm (ˆ). Let E0 := E ∩ D (ˆ), let ˆ0 be the restriction of the operator ˆ to the order-dense ideal E0 , and regard ˆ0 as an order-unity in the band {ˆ0 }⊥⊥ ⊂ L∼ (E0 , F ). Denote by the symbol Lˆ (E, F ) the set of all o-continuous regular operators from E into F whose restriction to E0 belongs to the band {ˆ0 }⊥⊥ , i.e. ⊥⊥ Lˆ (E, F ) := {S ∈ L∼ }. n (E, F ) : S  E0 ∈ {ˆ0 }

As is seen, an operator S belongs to Lˆ (E, F ) if and only if S results from the extension of some S0 ∈ {ˆ0 }⊥⊥ by o-continuity. This in particular implies that Lˆ (E, F ) is a band in L∼ n (E, F ). 0 Consider the set E ⊂ m(E) defined by the relation E 0 := {x0 ∈ m(E) : x0 · E ⊂ D (ˆ) }. 3.4.8. Theorem. The set E 0 is an order-dense ideal in m(E) linearly and latticially isomorphic to the space Lˆ (E, F ). The isomorphism may be implemented by assigning the operator Sx0 ∈ Lˆ (E, F ) to an element x0 ∈ E 0 by the formula Sx0 (x) = ˆ (x · x0 ) (x ∈ E). C The fact that E 0 is an order ideal in m(E) follows immediately from the definitions. On the other hand, the bases of the spaces Lˆ (E, F ) and m(E) are isomorphic by 3.4.5. Therefore, E 0 becomes an order-dense ideal in m(E) if we establish the sought isomorphism of the spaces m(E) and Lˆ (E, F ). Obviously, if x0 ∈ E 0 then Sx0 is an order continuous regular operator from E into F . Observe that ˆ0 is a Maharam operator. Consequently, if e ∈ E (1), i.e., e is a unit element of 1, then, by Theorem 3.4.5, the operator Se is a fragment of ˆ0

124

Chapter 3 0

and therefore Se ∈ {ˆ0 }⊥⊥ . Let (exλ )λ∈R be the spectral function of x0 . Then, by the Freudenthal Spectral Theorem, Z∞

0

x =

0

λ dexλ ,

−∞

where P∞ the integral on the right-hand side is the r-limit of integral sums of the form −∞ ln (eλn+1 − eλn ), ln ∈ (λn , λn+1 ), λn → +∞, and λ−n → −∞ for n → +∞. It follows that the operator Sx0 has the representation Z∞ Sx0 (x) =

0

λ d (ˆ (x · exλ ) ),

−∞ 0

i.e., the operator Sx0 is made from operators of the form Se , e := exλ by summation and passage to the o-limit. Since every band is closed under these operations, we have Sx0 ∈ {ˆ0 }⊥⊥ . Thus, S0 ∈ {ˆ0 }⊥⊥ and Sx0 ∈ Lˆ (E, F ). Also, it is clear that the correspondence x0 7→ Sx0 is an injective linear operator from E 0 into Lˆ (E, F ) and x0 ≥ 0 if and only if Sx0 ≥ 0. It remains to show that, for every S ∈ Lˆ (E, F ), there exists an x0 ∈ E 0 such that S = Sx0 . Indeed, let T be the restriction of S to E0 and consider the spectral function (eTλ )λ∈R of the operator T (with respect to the order-unity ˆ0 ). By virtue of 3.4.5, the family (h (eTλ ) )λ∈R is a resolution of unity in E (1); consequently, for 0 some x0 ∈ m(F ), we have exλ = h (eTλ ) for all λ ∈ R. Moreover, 0

eTλ (x) = ˆ (x · exλ ) for all λ ∈ R and x ∈ E0 . Appealing now to the Freudenthal Spectral Theorem and using some elementary properties of o-summable families, given x ∈ E+ we obtain the relations  ∞  Z Z∞ λ d (eTλ )  x = λ d (eTλ (x) ) Tx =  −∞

Z∞ = −∞

−∞

 0

λdˆ (x · exλ ) = ˆ x ·

Z∞

 0

λdexλ  = ˆ (x · x0 ).

−∞

Suppose now that x ∈ E+ , (xα ) ⊂ E 0 and sup (xα ) = x. Then ˆ (xα ·x0 ) ≤ S(x) and since D (ˆ) = Dm (ˆ), the family (xα ·x0 ) is bounded in D (ˆ). Hence, x·x0 ∈ D (ˆ) and Sx = ˆ (x · x0 ). Thus, x0 ∈ E 0 and the sought representation holds. B

Positive Operators

125

3.4.9. Luxemburg–Schep Theorem. Let E and F be order complete vector lattices; and let S and T be positive order continuous operators from E to F , with T possessing the Maharam property. Then the following are equivalent: (1) S ∈ {T }⊥⊥ ; (2) S  T ; (3) there exists an orthomorphism 0 ≤ ρ ∈ Orth∞ (E) such that Sx = T (ρx) for all x ∈ D(ρ); (4) there exists a sequence of orthomorphisms (ρn ) ⊂ Orth(E) such that Sx = supn T (ρn x) for all x ∈ E. C The implication (1) ⇒ (2) is trivial and the equivalences (1) ⇔ (3) and (1) ⇔ (4) can be easily deduced from 3.4.8. Prove (2) ⇒ (1). Suppose that S0 ∈ {T }⊥ and 0 ≤ S0 ≤ S. Then the operator S0 + T is absolutely continuous with respect to T and, by virtue of Theorem 3.4.3, possesses the Maharam property. According to 3.4.6 (1) CS0 ⊥ CT . Now, if 0 ≤ x ∈ CS0 then T x = 0 and S0 x ∈ {T x}⊥⊥ = {0}. Thus, S0 = 0 and this completes the proof. B 3.4.10. Theorem. Let E and F be K-spaces, and let T (E)⊥⊥ = F . Suppose that every increasing net (eα ) in E+ is order-bounded provided that (T eα ) is orderbounded in F . Then there exists an order continuous lattice homomorphism S : F → E such that T ◦ S = IF . C Observe that T (E) = F and CT = E by assumption. First suppose that F contains an order-unity 1. Then 1 = T e for some 0 ≤ e ∈ E. For every f ∈ F (1) there is a unique orthomorphism πf ∈ Z (F ) with πf 1 = f . Moreover, f 7→ πf is an isomorphism of the vector lattices F (1) and Z (F ). Let h be the lattice homomorphism of Theorem 3.4.3. Put S(f ) := h(πf )e (f ∈ F (1)). Then S is an order continuous lattice homomorphism from F (1) to E and (T ◦ S)f = T (h(πf )e) = πf T e = πf 1 = f . Observe further that a minimal extension of S to the whole of F , if such an extension exists, would be the sought operator. Take f ∈ F and put fn := f ∧ (n1) (n ∈ N). Then fn ∈ F (1) and for en := S(fn ) we have T (en ) = fn ≤ f . Thus (en ) is bounded in E and we may set Sf := supn Sen . In the case when F lacks order-unity we consider a family (fξ )ξ∈„ of positive pairwise disjoint elements with F = {fξ : ξ ∈ „}⊥⊥ . Then fξ = T (eξ ) for some eξ ∈ E and the family (eξ )ξ∈„ is pairwise disjoint, since T is essentially positive. Apply the above-proven fact to the operator πξ ◦ T , where πξ is the band projection onto {fξ }⊥⊥ . Thus, there is an operator Sξ : πξ F → h(πξ )E such that T ◦ Sξ is identity mapping on πξ F . Put X Sf := oSξ ([fξ ]f ) ξ∈„

(f ∈ F+ ).

126

Chapter 3

To ensure that the operator S is soundly defined, take a finite set θ := {ξ1 , . . . , ξn } in „ and put n n X X gθ := Sξk ([fξk ]f ), dθ := [fξk ]f. k=1

k=1

Then (dθ ) and (gθ ) are increasing nets and T gθ = dθ . By hypothesis (gθ ) is bounded, since (dθ ) is bounded above by f . Now it is an easy exercise to check that S is the desired operator. B 3.4.11. Theorem. Let E0 be a massive vector sublattice of a vector lattice E and let F be a K-space. Then there exists an order continuous lattice homomorphism ε from L∼ (E0 , F ) to L∼ (E, F ) such that % ◦ ε is the identical mapping on L∼ (E0 , F ). C We know already that % is an essentially positive Maharam operator from L∼ (E, F ) onto L∼ (E0 , F ), see 3.4.2 (6). Take an increasing net (Tα ) in L∼ (E, F ) and denote Sα := %(Tα ). Suppose that the net (Sα ) has an upper bound, say S ∈ L∼ (E0 , F ). For an arbitrary x ∈ E choose x0 ∈ E0 with x ≤ x0 . Then Tα x ≤ Tα x0 ≤ Sα x0 ≤ S0 x0 . Thus, (Tα ) is bounded above and we may apply Theorem 3.4.10. B 3.5. Maharam’s Extension of Positive Operators The main problem discussed in this section is the extension of an arbitrary positive operator to an order interval preserving order continuous operator, i.e., the Maharam extension. The structural properties of such extension allows us to deduce some results on approximation of the Boolean algebra of fragments of a positive operator by elementary fragments. 3.5.1. Consider vector lattices X and F , with F order complete, and an operator ˆ ∈ L+ (X, F ). Suppose that ˆ is essentially positive, i.e., N (ˆ) = {0}; and put x := ˆ(|x|) (x ∈ X). Then (X, · ) is a lattice-normed space. Denote by L1 (ˆ) the order norm completion (bo-completion) of X. According to Theorems 2.2.8 and 2.2.11 (1, 2) L1 (ˆ) is a Banach–Kantorovich lattice. This completion enjoys the following properties (see 2.2.7–2.2.12): (1) there is an isometric isomorphism ι from X onto the vector sublattice ι(X) ⊂ L1 (ˆ);
(2) there is a lattice and ring isomorphism h from Z (E) into Z L1 (ˆ) such that πˆ(x) = h(π)ι(x) (x ∈ X+ , π ∈ Z (E)+ ); (3) for every z ∈ L1 (ˆ) and 0 < ε ∈ R there are a partition (πξ ) of the projection [ z ] and a family (xξ ) ⊂ X such that X z − zε ≤ ε z , zε := πξ ι(xξ ).

Positive Operators

127

Let X be the order ideal in L1 (ˆ) generated by ι(X). 3.5.2. Theorem. For an arbitrary essentially positive operator ˆ ∈ L∼ (X, F ) the space L1 (ˆ) is an order complete Banach–Kantorovich lattice with an additive order continuous norm; moreover, ι(X)⊥⊥ = L1 (ˆ) and X is an order-dense ideal in L1 (ˆ). C It is evident that the above defined norm in X is additive on X+ . Take 0 00 x , x ∈ dι(X)+ and choose a partition of unity (πξ ) ⊂ F and two families (x0ξ ) ⊂ X+ and (x00ξ ) ⊂ X+ such that πξ x0 = πξ ι(x0ξ ) and πξ x00 = πξ ι(x00ξ ). Taking into account simple properties of a decomposable norm, we may deduce

πξ x0 + x00 = πξ ι(x0ξ ) + ι(x00ξ ) = πξ ι(x0ξ ) + ι(x00ξ ) = πξ x0 + x00 . Thus, the norm is additive on dι(X)+ . Additivity on a larger cone L1 (ˆ)+ is easily proven by passing to the limit in view of the representation L1 (ˆ) = rdι(X) from 2.2.11 (2). Suppose now that a net (zα ) ⊂ L1 (ˆ) decreases to zero and put e := inf zα . If α ≤ β then zβ − zα = zβ − zα = ( zβ − e) − ( zα − e) → 0. Thus, the net is bo-fundamental and there exists z := bo-lim zα . Since 0 ≤ z ≤ zα , we conclude that z = 0. At the same time e := o-lim zα = 0 and order continuity of the norm is proven. Finally, assume that a net (zα ) ⊂ L1 (ˆ) is increasing and bounded above. Then the net ( zα ) ⊂ F is also increasing and bounded above. In view of additivity of the norm the net (zα ) is again bo-fundamental and there exists z := bo-lim zα . It is easily seen that z = sup zα . Now we check the identity ι(X)⊥⊥ = L1 (ˆ). According to 2.2.9 the subspace ι(X) is dense in L1 (ˆ) in the sense of 3.5.1 (3). If y ∈ L1 (ˆ) and y ⊥ ι(X) then πξ y ⊥ πξ ι(xξ ); therefore, y ⊥ zε . Since bands are br-closed, we complete the proof by passage to the limit as ε tends to zero. B 3.5.3. Recall that the space L1 (ˆ) can be endowed with the structure of an Amodule for A := Orth(F ). Moreover, in view of 2.1.4 and 2.1.9, the natural representation of A in L1 (ˆ) defines an isomorphism of A onto a sublattice and subalgebra in Orth L1 (ˆ). Since the norm · is additive on the cone L1 (ˆ)+ , we may define the essentially ~ : L1 (ˆ) → F by putting positive operator ˆ ~ : z → z+ − z− ˆ

(z ∈ L1 (ˆ)).

~ is an order continuous operator. Clearly, order continuity of the norm implies that ˆ ~ is order interval preserving, and hence, a Maharam operator. Indeed, Moreover, ˆ

128

Chapter 3

~ if 0 ≤ e ≤ ˆ(z), z ∈ L1 (ˆ)+ , then there is an orthomorphism 0 ≤ α ≤ IF such that ~ ~ e = αˆ(z). By virtue of A+ -homogeneity of the norm e = α z = αz = ˆ(αz). ~ ~ At the same time αz ∈ [0, z] and ˆ([0, z]) = [0, ˆ(z)]. If A0 := Z (F ) then X is an A0 -submodule in L1 (ˆ). Moreover, the restriction ~ of ˆ ~ to X is also a Maharam operator. ˆ := ˆ| X 3.5.4. Theorem. For every operator S ∈ {ˆ}⊥⊥ there exists a unique operator S ∈ {ˆ}⊥⊥ such that S = S ◦ ι. The assignment S 7→ S implements an isomorphism of K-spaces {ˆ}⊥⊥ and {ˆ}⊥⊥ . C First of all observe that every positive (and thus every order-bounded) operator S : X → F admits a unique extension S0 to the vector sublattice dι(X) which commutes with all projections in P(F ). To demonstrate we need only to put   X X S0 z := πξ S(xξ ) z := πξ ι(xξ ) ∈ dι(X) . The definition is correct because |z| ≤ ιx for some x ∈ X and πξ S(xξ ) ≤ S(x) for all ξ. Let S lie in the order ideal generated by ˆ, i.e., |S| ≤ Cˆ. Then |S0 z| ≤ C z . Therefore, S0 is norm r-continuous and admits a unique extension S to the space X = rdι(X) by br-continuity. It is easy to check that S commutes with all projections in P(F ) and lies in the order ideal generated by ˆ. Take now an increasing net (Sα ) of positive operators in the order ideal generated by ˆ such that S := sup Sα ∈ {ˆ}⊥⊥ . If z ∈ X then |z| ≤ ι(x) for some x ∈ X; therefore we may estimate |S α z| ≤ S α (|z|) ≤ Sx. Thus it is correct to define some positive operator by putting Sz := sup S α z

(z ∈ X + ).

Obviously, S is a positive operator and S = sup S α ∈ {Sˆ}⊥⊥ . It is easily seen that the mapping S 7→ S is linear, positive, and one-to-one. It remains to show that an operator S ∈ {ˆ}⊥⊥ has only one extension lying in {ˆ}⊥⊥ . Observe that every operator T ∈ {ˆ}⊥⊥ commutes with all projections in P(F ), since ˆ is a Maharam operator, see 3.4.3 (3). Therefore, if T vanishes on ι(X) then X  X T πξ ι(xξ ) = πξ T (ι(xξ )) = 0. Since T is br-continuous and dι(X) is br-dense in X, we conclude that T is the zero operator. B 3.5.5. (1) Theorem. The space {ˆ}⊥⊥ is isomorphic to the ideal X 0 := {u ∈ mX : u·ι(X) ⊂ L1 (ˆ)}. The isomorphism is implemented by assigning the operator Su ∈␣L∼ (X, F ) to an element u ∈ X 0 by the formula Su (x) = ˆ (u · ιx) (x ∈ X). C The proof is immediate from 3.4.8 and 3.5.4. B

Positive Operators

129

(2) The space HomA (X, F ) of all order-bounded order continuous Alinear operators (A-module homomorphisms) from X to F coincides with the band {ˆ}⊥⊥ . C If S ∈ {ˆ}⊥⊥ then, according to 3.4.3, S is A-linear and S ∈ HomA (X, F ). Conversely, let 0 < S ∈ HomA (X, F ) and S ⊥ ˆ. Then T := S + ˆ is a Maharam operator and S, ˆ ∈ {T }⊥⊥ . From 3.4.6 (1) we conclude that CS ⊥ Cˆ . Since ˆ is essentially positive, Cˆ = X and CS = {0}. Thus S = 0 and this completes the proof. B 3.5.6. Introduce twoP sets M0 and M in X. The former consists of the elements representable as o- ρξ ι(xξ ), where (ρξ ) ⊂ P(F ) is an arbitrary partition of unity and (xξ ) ⊂ X is an order-bounded set. The latter comprises finite sums P n k=1 ρk · ι(xk ), where ρk ∈ P(F ), xk ∈ X (k = 1, . . . , n). Clearly, M and M0 are vector sublattices in X and ι(X) ⊂ M ⊂ M0 . Given a set M in X, denote by M ↓ the set of all elements z ∈ X, of the form z = inf zα , where (zα ) ⊂ M is a downward directed net. The set M ↑ is defined similarly on using upward-directed nets. If we take sequences instead of nets in these definitions then the corresponding sets are denoted by M  and M  . More precisely, M  is the set of all y = inf zn in X, where (zn ) ⊂ M is a decreasing sequence. Finally, we set M ↓↑ := (M ↓ )↑ and M  := (M  ) . An element y ∈ X belongs to M ↓↑ if and only if for arbitrary m ∈ X with m > |y| and n ∈ N there exists c ∈ M ↓↑ such that c ≤ y and ˆ(y − c) ≤ n1 ˆ(m). C If y ∈ M ↓↑ , then y = sup ϕy , where ϕy = {c ∈ M ↓ : c ≤ y}. Since M is a sublattice in X, the set M ↓ is also a sublattice; therefore, ϕy is directed upward. In view of order continuity of ˆ on X we conclude ˆ = sup ˆ(ϕy ) := sup{ˆ(c) : c ∈ ϕy }. The set ι(X) is a massive sublattice in X and we may choose m ∈ M with |y| ≤ m. Evidently ˆ(y) = sup ˆ(ϕy ∩ [−m, m]) = o-lim ˆ(ϕy ∩ [−m, m]). Now, by the properties of o-convergence in K-space, for each n ∈ N there exist a partition of unity (ρξ ) ⊂ P(F ) and a family (cξ ) ⊂ ϕy ∩ [−m, m] such that ρξ (ˆ(y) − ˆ(cξ )) ≤

1 ρξ ˆ(m) n

for every ξ. ↓ Denote c := inf{ρξ cξ + ρ⊥ ξ m} and observe that c ∈ M and c ≤ y. Since ˆ is A-linear and order continuous, we have ρξ ˆ(c) = ρξ ˆ(cξ ) for all ξ and ˆ(y − c) ≤ 1 n ˆ(m). Take a pair of elements y and m in X with |y| ≤ m. If for every n ∈ N there exists cn ∈ M ↓ such that c ≤ y and ˆ(y − cn ) < n1 ˆ(m), then y ∈ M ↓↑ . Indeed, if c = sup cn , then c ≤ y and ˆ(c) = sup ˆ(cn ) = ˆ(y); thus, y = c ∈ M ↓↑ , since ˆ is essentially positive. B

130

Chapter 3 3.5.7. Theorem. The following assertions hold: (1) X = M ↓↑ = M ↑↓ ;

(2) X = M0 = M0 . If F satisfy the countable chain condition then (3) X = M  = M  . C We confine exposition to (1). Show that M ↓↑ is a conditionally order complete sublattice in X. It is clear that M ↓↑ is conditionally complete upper semilattice, since (M ↓↑ )↑ = M ↓↑ . Prove the identity M ↓↑↓ = M ↓↑ . Take z ∈ M ↓↑↓ . There exists m ∈ M such that |z| ≤ m and for every 0 < ε ∈ R and n ∈ N one can choose a partition of unity (ρξ ) ⊂ P(F ) and a family (yξ ) ⊂ M ↓↑ satisfying the relations: |yξ | ≤ m, yξ ≥ z, ρξ ˆ(yξ − z) ≤ 2−n ερξ ˆ(m). ↓↑ Clearly, for yn := sup{ρξ yξ − ρ⊥ ξ m}, we have yn ∈ M , |yn | ≤ m, yn ≥ z, and ˆ(yn − z) ≤ 2−n εˆ(m). For each n ∈ N, according to the above-proved description of M ↓↑ , choose cn ∈ M ↓ with cn ≤ yn , |cn | ≤ m, and ˆ(yn − cn ) ≤ 2−n εˆ(m). Denote c = inf cn and observe that c ∈ M ↓ , |c| ≤ m, and c = inf cn ≤ inf yn = z. Prove the inequality ˆ(z − c) ≤ 2εˆ(m) which implies that z ∈ M ↓↑ and M ↓↑↓ = M ↓↑ . Consider a sequence c0n = inf{ck : k = 1, . . . , n}.PClearly, ˆ(z − c) = ˆ(|z − n c|) = o-limn ˆ(|z − c0n |). At the same time |z − c0n | ≤ k=1 |z − ck |, whence

ˆ(|z −

c0n |)

≤

n X k=1

ˆ(|z − ck |) ≤

n X


ˆ(|z − yk |) + ˆ(|yk − ck |) ≤ 2εˆ(m).

k=1

Thus, ˆ(z − c) ≤ 2εˆ(m). Now, using additivity of operations (·)↓ and (·)↑ , observe that M ↓↑ is a vector space: M ↓↑ + M ↓↑ = M ↓↑ , M ↓↑ − M ↓↑ = M ↓↑ + M ↑↓ = M ↓↑↓ = M ↓↑ . Finally we conclude that M ↓↑ is a K-space embedded in X as an order-closed sublattice. Observe that the restriction operator R : T 7→ T |M ↓↑ is an isomorphism of HomA (X, F ) onto HomA (M ↓↑ , F ). Indeed, if an operator T vanishes on M ↓↑ then T vanishes on X and T ◦ ı = 0; thus T = 0 by virtue of 3.5.4. This proves that R is injective. For an arbitrary S^ ∈ HomA (M ↓↑ , F ) put S := S^ ◦ ı and let S be a unique order continuous operator from X to F for which S ◦ ı = S (see

Positive Operators

131

Theorem 3.5.4). Then S and S^ agree on ı(X), and so they coincide on M ↓↑ . Thus, the vector lattices HomA (X, F ), HomA (M ↓↑ , F ), and {ˆ}⊥⊥ are isomorphic. It follows from 3.4.5 that the bases of the K-spaces M ↓↑ and X are also isomorphic, whence M ↓↑ = X. B 3.5.8. We observe the following simple consequences: (1) L1 (ˆ) = M ↓↑ . C The proof is immediate from 3.5.2 and 3.5.7 (1). B (2) The space X is the order completion of the vector sublattice W := M ↓ − M ↓ := {y1 − y2 : y1 , y2 ∈ M ↓ }. In particular, the restriction of each T ∈ {ˆ}⊥⊥ to W is an order continuous Alinear operator. C Obviously, M ↓ is a cone and hence M ↓ −M ↓ is a vector subspace. Moreover, W is a sublattice since (y1 − y2 ) ∨ 0 = y1 ∨ y2 − y2 ∈ W . Since X = M ↓↑ , for an arbitrary 0 < y ∈ X there is 0 < w ∈ W such that w ≤ y, i.e. W is a massive sublattice. It remains to observe that X = W ↑ = W ↓ . B 3.5.9. According to 3.5.4 and 3.5.5 (2), the vector lattices X, L1 (ˆ), {ˆ}⊥⊥ , and {ˆ}⊥⊥ have isomorphic bases. In the rest of the section we will give a detailed description for bases for X and {ˆ}⊥⊥ . As usual, we denote by [ιx] the band projection in X onto {ι(x)}⊥⊥ . (1) For x ∈ X+ the projection πx has the representation: πx (ˆ) = ˆ ◦ [ιx] ◦ ι. C Indeed, using 3.1.4 (5) and order continuity of ˆ, we deduce πx (ˆ)y = sup{ˆ(y ∧ nx) : n ∈ N} = sup{ˆ(ι(y) ∧ nι(x)) : n ∈ N} = ˆ(sup{ι(y) ∧ nι(x)) = ˆ ◦ [ιx](ι(y))). B Denote by S (X) and S (ˆ) the sets of all projections in X and the set of all fragments of ˆ representable in the form n _ k=1

ρk [ιxk ] and

n _

ρk πxk (ˆ),

k=1

where (xk ) ⊂ X+ , (ρk ) ⊂ P(F ), n ∈ N. Given a band K in X, denote by hKi
the ⊥⊥ ⊥⊥ band projection in X onto (ιK) , i.e. hKi := [ιK]. Put hxi := [ι {x} ] and πhxi := π{x}⊥⊥ (x ∈ X).

132

Chapter 3

(2) For every band K in X the projection πK has the representation: πK (ˆ) = (ˆ ◦ hKi) ◦ ι. In particular, πhxi (ˆ) = (ˆ ◦ hxi) ◦ ι. C The proof is similar to (1). B Wn Let C (X) denote theWset of projections k=1 ρk · hxk i, and let C (ˆ) be the set n of fragments of the form k=1 ρk · πhxk i where n ∈ N, (ρk ) ⊂ P(F ), and (xk ) ∈ X. In the case when X is a vector lattice with the principal projection property we may Wn consider one more set A (ˆ) consisting of the fragments of ˆ representable as k=1 ρk ◦ ˆ ◦ [xk ], where n ∈ N, (ρk ) ⊂ P(F ), and [xk ] is the band projection in X onto {xk }⊥⊥ . 3.5.10. The following are valid: (1) P(X) = S (X)↓↑ ; (2) E(ˆ) = S (ˆ)↓↑ . C By definition [ιy] ∈ S (X) for each y ∈ M . If 0 ≤ y ∈ M ↓ , we may choose x ∈ X+ and a family (mξ )ξ∈„ ⊂ M such that ιx ≥ mξ ≥ y (ξ ∈ „) and y = inf mξ . It is not difficult to verify that " + # _ ^  1 . [y] = mξ − ιx n n∈N ξ∈„

Since the element yn,ξ = (mξ − n1 ιx)+ is contained in M , it follows [yn,ξ ] ∈ S and [y] ∈ S ↓↑ . An arbitrary projection π ∈ P(X) has the representation π = sup{[y] : y ∈ X + , πy = y}. Thus, taking 3.5.8 into consideration we arrive at the desired
↑ inclusion π ∈ (S ↓↑ )↑ = S ↓↑ . B 3.5.11. The following are valid: (1) P(X) = C (X)↑↓↑ ; (2) E(ˆ) = C (ˆ)↑↓↑ . If X has the principal projection property then (3) E(ˆ) = A ↑↓↑ . C It is sufficient to show that [ιx] ∈ C ↑↓ for every x ∈ X+ . Then S (X) ⊂ C (X)↑↓ , so that  ↓↑ P(X) = S (X)↓↑ ⊂ C (X)↑↓ = C (X)↑↓↑ ⊂ P(X). Thus, what we need is only to justify the representation: ^ _ [ιx] = h(nx − t)+ i. t∈X+ n∈N

Positive Operators 133 V V Denote σt := n h(nx − t)+ i and σ = t σt . It is not difficult to observe that σt ≥ [ιx] for all t ∈ X+ . For an arbitrary projection ρ ∈ P(X) with ρ ∧ [ιx] = 0 put ρt := ρ ∧ [ιt] (t ∈ X+ ). Then ρt ≤ [ι(t − nx)+ ] ≤ h(t − nx)+ i for every + n ∈ N. Since h(nx − t)+ i = 0 it follows ρt ∧ h(nx − t)+ i = 0 and V h(t − nx) i ∧ ρt ∧ σt = n (ρt ∧ h(nx − t)+ i = 0. From this we obtain the following identities: ρt ∧ σ = 0,

ρ ∧ σ = sup ρt ∧ σ = 0.

Now, putting ρ = [ιx]⊥ , we arrive at the desired inequalities [ιx] ≤ σ ≤ [ιx]. (2): It follows from (1) and 3.5.9 (2). (3): It is sufficient to observe that (ˆ ◦ hxi) ◦ ι = ˆ ◦ πx . B 3.6. Comments 3.6.1. (1) In [157], L. V. Kantorovich laid grounds for the theory of regular operators in K-spaces. Also, the Riesz–Kantorovich Theorem (3.1.2) appeared in this article for the first time. F. Riesz [331] formulated an analogous assertion for the space of continuous linear functionals over the lattice C[a, b] in his famous talk at the International Mathematical Congress in Bologna in 1928 and thereby became enlisted in the cohort of the founders of the theory of ordered vector spaces. (2) The minimal extension operator and its properties are well known (see, for instance, [23]). The set of e ∈ mE for which S([0, |e|] ∩ G) is orderbounded in F is a maximal order-dense ideal in mE to which the operator S admits the minimal extension. Let p : G → F be an increasing (0 ≤ e1 ≤ e2 implies p(e1 ) ≤ p(e2 )) sublinear operator with p(g) = p(g + ) (g ∈ G). Denote by E the set of all e ∈ mG for which p([0, |e|] ∩ G) is order-bounded in F . The operator pm : E → F defined by pm (e) := sup{p(g) : g ∈ G, 0 ≤ g ≤ e+ } is also sublinear and ∂pm = {πG S : S ∈ L∼ (E, F ), S|G ∈ ∂p}. This construction is, in particular, a basis for Yu. A. Abramovich’s maximal extension of a vector lattice [1–3]. (3) The operators of the form πG S, where G ∈ I (E), have long been in use (see, for instance, [388]) but seem to be considered explicitly in A. R. Schep’s article [337], see also [23]. The idea to employ fragments of the form πG S and πe instead of the fragments S ◦ π belongs to E. V. Kolesnikov [14, 175, 176]. (4) Yu. A. Abramovich [3] developed a version of the calculus of 3.1.2 in which suprema and infima can be taken over partitions of the argument into disjoint parts, see 3.1.6 (2). For the modulus of a regular operator, this fact was independently established by W. A. J. Luxemburg and A. C. Zaanen [263], see also [23, 228, 409].

134

Chapter 3

(5) The concept of a generating set of projections as well as Theorem 3.1.5 (3) belongs to S. S. Kutateladze [227]. The main idea of [227] is as follows: The fragments of a positive operator U are the extreme points of the order interval [0, U ]. The latter set coincides with the supporting set ∂p of the sublinear operator p(x) := U x+ . Thereby, studying the fragments of a positive operator reduces to describing the extremal structure of a supporting set. Such a description for a general sublinear operator was obtained for the first time in the article [221] by S. S. Kutateladze (for a detailed exposition, see [209]). (6) The problem of dominated extension of linear operators originates with the Hahn–Banach Theorem; see [63] for its history. Theorem 3.1.8 (2) was discovered by L. V. Kantorovich in 1935 [153]. The equivalence between the extension and the least upper bound properties (Theorem 3.1.8 (3)) was first established by W. Bonnice and R. Silvermann [48] and T.-O. To [374]; an elegant proof with decisive simplifications is due to A. D. Ioffe [138]; see also [209]. Theorem 3.1.17 is due to L. V. Kantorovich [159]. (7) Theorem 3.1.8 (2) can be considered as an exemplar application of the heuristic transfer principle for K-spaces (see 1.6.3 (2)). It claims that the Kantorovich principle is valid in relation to the classical Dominated Extension Theorem; i.e., we may replace the reals in the standard Hahn–Banach Theorem by elements of an arbitrary K-space and a linear functional by a linear operator with values in this K-space. 3.6.2. (1) Theorem 3.2.2 was proved by E. V. Kolesnikov [178]; of course, it is well known at least when E is a Kσ -space. Theorems 3.2.3 (2) and 3.2.3 (3) are due to Ogasawara [302] and A. I. Veksler [376]. In [178] E. V. Kolesnikov suggested some simple localization method that enabled him to obtain the formulas of 3.2.5 for calculating the order continuous and σ-order continuous parts. The formulas 3.2.6 (1) took their final form (see [337] and [20]) gradually in the works of various authors (W. A. J. Luxemburg, A. C. Zaanen, C. D. Aliprantis, A. R. Schep, and P. van Eldik); a piece of this history can be learned from [23, 409]. Formulas 3.2.6 (2) are just variants of 3.2.6 (1) in the case when the vector lattice E has the strong Freudenthal property. (2) The projection formulas of 3.2.7 and 3.2.8 (1) were established by E. V. Kolesnikov [177]. Using nonstandard methods of analysis, S. S. Kutateladze independently found 3.2.8 (1) in [227] at E. V. Kolesnikov’s request. The particular cases given in 3.2.8. (3, 4) were earlier obtained by C. D. Aliprantis and O. Burkinshaw [22]. Kutateladze’s method of generating sets (see 3.6.1 (5)) works also for projection formulas: Theorem 3.2.9 (1) have been proven in [227]. Of course, this result enables us to find various projection formulas by taking concrete generating sets of projections. A particular case of Kutateladze’s formulas, presented in

Positive Operators

135

3.2.9 (2), was earlier obtained by A. G. Kusraev and V. Z. Strizhevski˘ı in [218]. Equivalences (1) ⇔ (3) and (1) ⇔ (4) in Theorem 3.2.10 were established in [218] and [227], respectively. A variant of the equivalence (1) ⇔ (3) in the case when Fn∼ separates the points of F was earlier obtained in [20]; see also [23]. (3) The shadow of an operator S : E → F is the mapping shdw(S) : P(E) → P(E) defined by shdw(π) = [Sπ(E)]. Thus, shdw(π) is the band projection onto (Sπ(X))⊥⊥ , see 5.2.2. Denote hshdw(S)i := {T ∈ L∼ (E, F ) : (∀σ ∈ P(E)) shdw(T )σ ≤ shdw(S)σ}. It is easy to observe that hshdw(S)i is a band containing {S}⊥⊥ . Let [shdw(S)] and [shdw(S)]n be the band projections onto hshdw(S)i and hshdw(S)i ∩ L∼ n (E, F ), respectively. Denote by  and n the sets of all finite and arbitrary partitions of unity in P(E) E. V. Kolesnikov proved that the following projection formulas hold: ( [shdw(S)]T = inf

n X

) shdw(σk )T σk : (σ1 , . . . , σn ) ∈ , n ∈ N ,

k=1

( [shdw(S)]n T = inf

) X oshdw(σα )T σα : (σα )α∈A ∈ n

.

α∈A

3.6.3. (1) The basic properties of lattice homomorphisms are presented in [23]. Theorem of 3.3.1 is due to M. Meyer [86, 282]. Theorem 3.3.3 was announced in [219] and proved in [220] by S. S. Kutateladze. He obtained it as a simple consequence of his powerful canonical sublinear operator method; for details see [209]. Another proof was found by W. A. J. Luxemburg and A. R. Schep [261]. Propositions 3.3.4 (1–5) are easy consequences of Kutateladze’s Theorem. V. A. Radnaev [324] demonstrated that the converse of 3.3.4 (4) is also true: a positive operator T ∈ L∼ (E, F ) is a lattice homomorphism if and only if for every 0 ≤ T1 , T2 ≤ T the relation T1 ⊥ T2 implies T1 (E) ⊥ T2 (E). Lattice homomorphisms in vector lattices are closely related to disjointness preserving operators to which Chapter 5 is devoted; see Section 5.6 for the relevant comments. (2) The theory of orthomorphisms stems from H. Nakano [287]. Orthomorphisms were studied by many authors under various names: dilatators (H. Nakano [287]), essentially positive operators (G. Birkhoff [46]), polar preserving endomorphisms (P. F. Conrad and J. E. Diem [68]), multiplication operators (R. C. Buck [53] and A. W. Wickstead [396]), and stabilisateurs (M. Meyer [281]). The main stages of this development are reflected in the books by A. Bigard, K. Keimel, and S. Wolfenstein [44], C. D. Aliprantis and O. Burkinshaw [23], A. C. Zaanen [409] etc.; see also the survey by A. V. Bukhvalov [60]. The results of this book are covered by [23, 409]. Available is an extensive bibliography

136

Chapter 3

on the theory of orthomorphisms; we indicated a portion of it connected with the subjects we discuss in the sequel: [4, 12, 13, 39, 45, 86, 122, 123, 135, 259, 261, 279, 310, 311, 321, 408]. (3) In 3.3.5–3.3.8 a portion of Kutateladze’s theory of extreme operators is exposed. The main results (Theorems 3.3.6, 3.3.7 and 3.3.8) were established in [221, 223]. In the same papers the following operator variant of the classical Kre˘ın– Milman Theorem was suggested: Take a K-space F and let 0 ≤ T ∈ L∼ (E, F ). Call an operator S ∈ ∂p a T extreme point of ∂p if T ◦ S ∈ Ch(T ◦ p). Say that S is an o-extreme point of ∂p or p if S is a T -extreme point of ∂p for every positive o-continuous operators T : E → F and every K-space F . Denote the set of all o-extreme points by E0 (P ). Kre˘ın–Milman Theorem for o-Extreme Points. Every sublinear operator p : X → E is the upper envelope of the set of its o-extreme points. Symbolically, p(x) = sup{T x : T ∈ E0 (p)} (x ∈ X). Moreover, the least upper bound on the right-hand side is attained for each x ∈ X. These and other facts from [221, 223, 226] give sound grounds for studying the extreme structure of convex sets of linear operators. The set E (T ) of all positive extensions of a positive operator T defined on a massive subspace G of a vector lattice E coincides with the supporting set ∂p whenever p : E → F is defined by p(e) := inf{T g : e ≤ g, g ∈ G} (e ∈ E). Now, it follows from the Kre˘ın–Milman Theorem that the convex set E (T ) has extreme points. This simple corollary was independently proved by Z. Lipecki [244] (see also [23]). About other extension results by Z. Lipecki see [243, 245, 248, 249]. (4) The Hahn–Banach formula for lattice homomorphisms (Theorem 3.3.10) was established by V. A. Radnaev [324]. Radnaev’s approach is based upon Kutateladze’s machinery for calculating supporting sets and their extreme boundaries. The auxiliary facts of 3.3.9 belong to V. A. Radnaev [324] (3.3.9 (1)) and G. J. H. M. Buskes and A. C. M. van Rooij [64] (3.3.9 (2, 3)). The important corollary 3.3.11 (1) was obtained by G. J. H. M. Buskes and A. C. M. van Rooij in [64]. Theorem 3.3.11 (2), conventionally referred to as the Lipecki–Luxemburg– Schep Theorem because of [260] and [244] (see, for instance, [23, 41, 63, 64]), was announced in [221] and proved in [223] by S. S. Kutateladze. (Actually S. S. Kutateladze proved Theorem 3.3.6 which contains 3.3.11 (2) as a particular case in view of 3.3.3.) Various approaches (mysteriously, except those by S. S. Kutateladze) to Hahn–Banach-type theorems for lattice homomorphisms are discussed in C. B. Bernau [41]; see also a nice survey by G. Buskes [63] in which the history, interconnections, and part of numerous generalizations of the Hahn–Banach Theorem are collected.

Positive Operators

137

(5) In [210] it was indicated that developing a calculus of extreme points is desirable and the following problem was formulated: For which sublinear operators p : X → E and q : E → F does the relation Ch(q ◦ p) ⊂ Ch(q) ◦ Ch(p) (or Ch(q ◦ p) = Ch(q) ◦ Ch(p)) hold? The same question relates to o-extreme points: For which p and q does the relation E0 (q ◦ p) ⊂ E0 (q) ◦ E0 (p), or E0 (q ◦ p) = E0 (q) ◦ E0 (p) hold? Of course, the problem is partially motivated by Kutateladze–Milman’s Theorem 3.3.8. V. A. Radnaev observed in [324] that Ch(πp) = π Ch(p) for a sublinear operator p : X → E and a positive orthomorphism π ∈ Orth(E). Radnaev’s formula of 3.3.10 can be also considered as a partial answer to the above-mentioned questions, since according to 3.3.9 (2, 3) it can be rewritten equivalently as 

[

π Ch(p ◦ T ) =

0≤π≤IE

[

 π Ch(p) ◦ T.

0≤π≤IE

(6) The results of 3.3.12 and 3.3.13 are from the article by C. B. Huijsmans and B. de Pagter [136]. This article also contains some interesting applications of 3.3.13 (1, 2). In the case when E is a vector lattice with the strong Freudenthal property and F is order complete E. V. Kolesnikov gave another projection formula: If 0 ≤ T ∈ L∼ (E, F ) and Td is the band projection of T onto L∼ d (E, F ) then  Td = inf sup{πT σ : πT σe ≤ εT e} : 0 < ε ∈ R, 0 ≤ e∈E . (The infimum and supremum are taken in L∼ (E, F ).) To obtain a pointwise formula we must make the set under the supremum on the right-hand side upward-directed by adding finite disjoint suprema: ( Td (x) = inf sup 0 0, there exists a partition of unity (ρα )α≥α P0 in P(E) such that ρα eα ≤ εe0 (α ≥ α0 ). Assign f := inf α∈A T eα and πθ := α∈θ ρα , where θ is a finite subset of {α ∈ A : α ≥ α0 }. Choose β ∈ A so that β ≥ α for all α ∈ θ. Then πα eβ ≤ πα eα ≤ εe0 (α ∈ θ); hence, f ≤ T eβ = T (πθ eβ ) + T (IE − πθ )eβ ≤ ε T e0 + T (IE − πθ )e0 . From complete additivity of the operator T it follows that o-lim T (IE − πθ )e0 = 0; θ

therefore, f ≤ ε T e0 . Since ε > 0 is arbitrary, f = 0. Consequently, the operator T is order continuous. B 4.3.8. Let E be a sublattice of a vector lattice G. Say that a net (xα ) ⊂ X is G-convergent to x ∈ X if there exists a decreasing net (eγ ) ⊂ E such that inf γ eγ = 0 in G and, for every γ, there is an index α(γ) such that x − xα ≤ eγ for all α ≥ α(γ). An operator T : X → Y is called G-continuous (sequentially Gcontinuous) if o-limα T xα = 0 (o-limn T xn = 0) for every net (xα ) ⊂ X (sequence (xn ) ⊂ X) G-converging to zero. The set of all dominated (regular) G-continuous operators is denoted by MG (X, Y ) (by L∼ G (E, F )). In the sequential case we use in this notation σG instead of G. Theorem. Let X be decomposable and let F be order complete. Then a dominated operator T : X → Y is G-continuous (sequentially G-continuous) if and only if so is its exact dominant; i.e., T ∈ MG (X, Y ) ⇔ T ∈ LG (E, F ), T ∈ MσG (X, Y ) ⇔ T ∈ LσG (E, F ). C The proof proceeds in much the same way as in 4.3.2. B

162

Chapter 4

4.3.9. Consider one more band of M (X, Y ). Let J(E, F ) be the band of L∼ (E, F ) generated by the set of finite-rank o-continuous operators: J(E, F ) := (En∼ ⊗ F )⊥⊥ . Define  MJ (X, Y ) := T ∈ M (X, Y ) : T ∈ J(E, F ) . Almost integral operators we call the elements of MJ (X, Y ). In order to obtain some internal description for the band MJ , we need the notion of ∗-convergence by P. S. Aleksandrov and P. S. Urysohn. We recall that a sequence (en )n∈N in E is (∗)

called ∗-convergent to e ∈ E (with respect to o-convergence), en −→ e in writing, if every subsequence (enk )k∈N of it contains a subsequence (enkl )l∈N o-convergent to x. We call an operator T : X → Y ∗-o-continuous if it is bo-continuous and takes (∗)

every bounded sequence ∗-convergent to zero (xn ∈ X, xn −→ 0) into a sequence (o)

o-convergent to zero (T xn −→ 0). Denote by M∗n (X, Y ) the set of all ∗-o-continuous dominated operators from X into Y and assign L∗n (E, F ) := M∗n (E, F ). It is easy to see that M∗n (X, Y ) is a band. Indeed, from the definitions it follows (∗)

(o)

immediately that M∗n (X, Y ) is an order ideal. Next, if xn −→ 0 and Tα −→ T , (o)

where xn ∈ X, xn ≤ e ∈ E, and Tα ∈ M∗n (X, Y ), then T xn −→ 0, since T xn ≤ T − Tα (e) + Tα xn . A K-space F is said to be regular if for every sequence (An ) of subsets An ⊂ F with o-lim inf An = 0 there exist finite subsets A0n such that o-lim inf A0n = 0. 4.3.10. Theorem. Let X and Y be BKSs and let the norm lattice F be a regular K-space. A dominated operator from X into Y is almost integral if and only if it is ∗-o-continuous: MJ (X, Y ) = M∗n (X, Y ). C Every functional e0 ∈ En∼ is ∗-o-continuous, therefore, En∼ ⊗ F ⊂ L∗n (E, F ). Since L∗n (E, F ) is a band, we have J(E, F ) ⊂ L∗n (E, F ); hence, MJ (X, Y ) ⊂ M∗n (X, Y ). Conversely, suppose that the operator T ∈ M (X, Y ) is ∗-o-continuous. By decomposability of the least dominant (see 4.2.6), there exists an operator U ∈ M (X, Y ) such that S := U ⊥ J(E, F ), T −U ∈ J(E, F ), and T = U + T −U . If we prove that S = 0 then U = 0 and T ∈ MJ (X, Y ). First of all, observe that the operator U is ∗-o-continuous, since T and T − U do. By Theorem 4.3.2, the operator S is o-continuous; consequently, without loss of generality, we may assume that there is a weak order-unity 1 ∈ F and an essentially positive functional f ∈ En∼ . Let e := x for some x ∈ X. The relation S ⊥ f ⊗ 1 implies  0 = (S ∧ f ⊗ 1)e = inf S(e − e0 ) + f (e0 ) · 1 . 0≤e0 ≤e

Dominated Operators 163  Assign An := S(e − e0 ) + f (e0 ) · 1 : 0 ≤ e0 ≤ e, f (e0 ) ≤ 1/n . It is clear that An+1 ⊂ An and inf(An ) = 0 for each n ∈ N. According to our assumption about F , there exist finite sets A0n ⊂ An
such that o-limn→∞ inf(A0n ) = 0. There exists a strictly increasing sequence k(n) n∈N of naturals and a sequence (em )m∈N  in E such that 0 ≤ em ≤ e and A0n = S(e−em )+f (em )·1 : k(n) ≤ m < k(n+1) . (∗)

Since the functional f ∈ En∼ is essentially positive and f (em ) → 0, we have em −→ 0. Consider a sequence (xm )m∈N in X, for which xm = em and x − xm = e − em (o)

(see 2.1.7 (1)). It is clear that (xm ) ∗-converges to zero; thus, U xm −→ 0. So, the following equalities hold: o- lim U xm = 0, m→∞

 o- lim inf S(e − em ) : k(n) ≤ m < k(n + 1) = 0. n→∞

For an arbitrary ε > 0, there exist partitions of unity (πm )m∈N and (ρn )n∈N in the Boolean algebra P(F ) such that πm U xk ≤ ε1 (k, m ∈ N, k ≥ m),  ρn inf S(e − ek ) : k(p) ≤ k < k(p + 1) ≤ ε1 (p ≥ n).  Whence, for k ≥ max k(p), m , p ≥ n, we infer πm U x ≤ S(e − ek ) + πm U xk ≤ S(e − ek ) + ε1,  ρn πm U x ≤ ρn inf S(e − ek ) : k (p) ≤ k < k(p + 1) + ε1 ≤ 2ε1. Summing up over n and m, we obtain U x ≤ 2ε1; consequently, U x = 0. Since the element x ∈ X is arbitrary, we have U = 0 and S = 0. B 4.3.11. Under the hypotheses of Theorem 4.3.10, for every dominated operator T : X → Y , there is a unique representation of T as the disjoint sum T = Ti + Tsi , where Ti is a ∗-o-continuous operator and Tsi is a dominated operator that has no ∗-o-continuous fragments. C The claim follows from 4.2.6 and 4.3.10. B 4.4. The Yosida–Hewitt-Type Theorems We will consider the question about decomposition of a dominated operator into the order continuous and order singular parts. These results are conventionally

164

Chapter 4

called theorems of Yosida–Hewitt-type because of the classical fact on decomposition of a finitely additive measure into countably additive and finitely additive parts. The decomposition in which singularity is understood as disjointness from the set of all order continuous operators is referred to as the weak form of the Yosida– Hewitt decomposition. Very often singularity means that an operator vanishes on a “huge” subset of its domain. In this case decomposition into the o-continuous and singular parts is called the strong form of the Yosida–Hewitt decomposition. 4.4.1. A dominated operator is called norm order singular (bo-singular ) or abnormal if it is disjoint to each bo-continuous dominated operator. Denote by Mos (X, Y ) (respectively, L∼ os (E, F ) the set of all dominated (regular) bo-singular operator from X into Y from E into F . Then the above definition can be rewritten as follows: T ∈ Mos (X, Y ) ⇐⇒ T ⊥ Mn (X, Y ), ∼ T ∈ L∼ os (E, F ) ⇐⇒ S ⊥ Ln (E, F ).

Let X be a decomposable lattice-normed space and let Y be a Banach–Kantorovich space. Then, for an operator T ∈ M (X, Y ), the following are equivalent: (1) T is norm order singular; (2) T is order singular; (3) T has no nonzero bo-continuous fragments. C The equivalence (2) ⇔ (1) follows from 4.3.2. The implication (1) ⇒ (3) is obvious and so it remains to observe that (3) ⇒ (2). If S is the projection of T onto the band L∼ n (E, F ) then, in view of decomposability of the dominant norm (Theorem 4.2.6), there is a fragment T0 of T such that T0 = S. Since T has no nonzero bo-continuous fragments, we have T = 0 and S = 0. B 4.4.2. Now the weak form of the Yosida–Hewitt decomposition can be easily deduced from the decomposability of dominant norm. Theorem. Let X be a decomposable lattice-normed space and let Y be a Banach–Kantorovich space. Then every operator T ∈ M (X, Y ) has a unique representation in the form T = Tn + Tos , where Tn ∈ Mn (X, Y ), and Tos ∈ Mos (X, Y ). Moreover, T = Tn + Tos , Tn = T n , Tos = T os . A decomposition of T with the properties indicated above is unique. C According to 3.2.3 (2) the positive operator T has a unique representation in ∼ the form T = T n + T os , where T n ∈ L∼ n and T os ∈ Los . Using Theorem 4.2.6

Dominated Operators

165

we may find T1 , T2 ∈ M (X, Y ) such that T1 = T n , T2 = T os , and T = T1 +T2 . It follows from 4.3.2 and 4.4.1 that T1 ∈ Mn (X, Y ) and T2 ∈ Mos (X, Y ). Thus, we may put Tn := T1 and Tos := T2 . Uniqueness follows from 4.2.7. B A similar result is valid for the σ-continuous fragments. 4.4.3. The following formulas for calculating the order continuous and order σ-continuous parts of a dominated operators can be easily deduced from 3.2.4 (2), 3.2.5 (4), 4.2.2, 4.2.4 (1): (1) Tn = bo- lim πA T A∈I

(I := Id (E)),

(2) π[e] Tnσ = bo- lim πA T A∈I

(I := Iσd ([e]), e ∈ E+ ).

The fragments can be also calculated as pointwise limits: (3) Tn x = bo- lim πA T x (I := Id (E), x ∈ X), A∈I

(4) Tnσ x = bo- lim πA T x (I := Iσd ({e}⊥⊥ ), e ∈ E+ , x ∈ {e}⊥⊥ ). A∈I

The limits are taken over the decreasing net of order-dense ideals A ⊂ E. Suppose now that E is a vector lattice with the projection property. Introduce two directed sets, € and . Let € be the set of all partitions of unity in the Boolean algebra P(X) which we order by “refinement.” More precisely, if γ and γ 0 are partitions of unity in P(X), then γ ≤ γ 0 means that, for every π 0 ∈ γ 0 , there exists a π ∈ γ such that π ≤ π 0 . Let the set  consist of all upward-directed subsets δ ⊂ P(X) such that sup δ = IX . The order in  is introduced as above: δ ≤ δ 0 ⇔ (∀π 0 ∈ δ 0 )(∃π ∈ δ) π ≤ π 0 . 4.4.4. Now, consider the following formulas: P (1) Tn x := bo-lim boT ◦ πx (x ∈ X); γ∈€

(2)

Tn e := inf

P

γ∈€ π∈γ

π∈γ

T ◦ πe (e ∈ E+ );

(3) Tn x := bo- lim bo-lim T ◦ πx (x ∈ X); δ∈

(4)

π∈δ

Tn e := inf sup T ◦ πe (e ∈ E+ ). δ∈ π∈δ

Theorem. Let (X, E) be a decomposable LNS, let (Y, F ) be a Banach–Kantorovich space, let E have the projection property. Then, for every T ∈ M (X, Y ), each of the formulas (1) and (3) well defines a dominated operator Tn : X → Y . Moreover, Tn ∈ Mn (X, Y ), T − Tn ∈ Mos (X, Y ), and, the exact dominant can be calculated by (2) or (4).

166

Chapter 4

C We confine exposition to considering only the first pair of formulas. The first formula is an easy consequence of the second. Indeed, if (2) is valid then for any partition of unity γ ∈ € we have X X T ◦ π = boT ◦ π → Tn , fγ := oπ∈γ

π∈γ

since the net (fγ )γ∈€ is decreasing and inf γ∈€ fγ = o-limγ∈€ fγ . Thus, P(1) follows from 4.2.4 (1) and 4.2.7. It remains to prove that Sn e = f := inf γ∈€ o- π∈γ S ◦ πe for a positive operator S : E → F and e ∈ E+ . Take an order-dense ideal A ⊂ E of the form [  X A := π(E) := lin π(E) , π∈γ

π∈γ

P where γ ∈ € . Then πA S = o- π∈γ S ◦ π in view of 4.2.3 (4) and 4.2.4 (2). By 3.2.4 (2) we may conclude Sn e ≤ f . To prove the reverse inequality, suppose that A is an order-dense ideal and choose a partition of unity γ ∈ € such that πe ∈ A for all π ∈ γ. According to 3.1.3 (1) πA Se ≥ πA Sπe = Sπe; therefore, πA Se ≥ f . Since A is arbitrary, we obtain Se ≥ f . B A similar result is valid for the σo-continuous fragment Tσ of an operator T ∈ M (X, Y ). Moreover, Tσ can be calculated by the formula Tσ = bo- lim

(πk )∈€σ

∞ X

T ◦ πk x (x ∈ X),

k=1

where €σ is the subset of € consisting of countable partitions of unity. The operator T − Tσ is σo-singular in the sense that it is disjoint from all sequentially bo-continuous dominated operators or, which is equivalent, has no sequentially bocontinuous fragments. The exact dominant satisfies Tσ = T σ and T − Tσ = T − Tσ . 4.4.5. We now present a strong form of the Yosida–Hewitt Theorem for dominated operators. An operator T : X → Y is called singular if there exists an orderdense ideal X0 ⊂ X on which T vanishes. From 4.1.5 it follows immediately that a dominated operator T is singular if and only if so is its exact dominant T . We denote the set of all dominated (regular) operators from X into Y (from E into F ) by Ms (X, Y ) (respectively, L∼ s (E, F )). The null ideal NT of the operator T is introduced by the formula  NT := x ∈ X : (∀u ∈ X)( u ≤ x ⇒ T u = 0)

Dominated Operators

167

Let E and F be vector lattices, with F order complete. For a positive operator S : E → F , the following hold: ∼ (1) S ∈ L∼ n (E, F ) ⇔ S ⊥ Ls (E, F );


(2) S ∈ L∼ n (E, F ) ⇔ (∀0 ≤ U ≤ S) NU ∈ B(E) . C We have only to verify (2), since (1) was already proven in 3.2.3 (1). If S is order continuous and 0 ≤ U ≤ S then U is also order continuous, so that NU is a band. Suppose that the right-hand side of the desired equivalence is valid. Let (gα ) be an increasing net in E with g = supα gα and let f := supα Sgα . Check that f = Sg. To this end, for 0 < ε < 1, denote eα := (εg − gα )+ and Sα := πeα S (see 3.1.3 (3) and 3.1.4 (5) for the definitions). Since (eα ) is decreasing, (Sα ) is also decreasing and U := inf α Sα ≤ S. Also U (εg − gα )− = 0 or (εg − gα )− ∈ NU , because Sα (εg − gα )− = 0. Moreover, o-lim(εg − gα )− = (1 − ε)g, and hence, since by our hypothesis NU is a band, g ∈ NU . Now observe that the following simple relations 0 ≤ Seα = Sα eα ≤ Sg, 0 ≤ g − gα = (1 − ε)g + (εg − gα ) ≤ (1 − ε)g + (εg − gα )+ imply 0 ≤ Sg − f ≤ S(g − gα ) ≤ (1 − ε)Sg + Seα ≤ (1 − ε)Sg + Sα g. From this, taking into consideration that o-lim Sα g = U g = 0, we deduce 0 ≤ Sg − f ≤ (1 − ε)Sg for all 0 < ε < 1. Hence, f = Sg and the result follows. B 4.4.6. Theorem. If the hypotheses of Theorem 4.2.6 are met then, for an operator T ∈ M (X, Y ), the following are equivalent: (1) T ∈ Mn (X, Y ); (2) T ⊥ Ms (X, Y ); (3) if U ∈ M (X, Y ) and U ≤ T then NU is a band. C (1) ⇒ (3): If T ∈ Mn (X, Y ) and U ≤ T then, in view of 4.3.2, U ∈ Ln (E, F ). According to 4.4.5 (2), N U is a band in E and it remains to take account of the obvious equality (see 2.1.2) NU = h(N

U

) := {x ∈ X : x ∈ N U }.

(3) ⇒ (2): Suppose that U ≤ T for some singular U ∈ M (X, Y ). Then NU is a band and, at the same time, X = NU⊥⊥ = NU . Thus, U = 0. (2) ⇒ (1): Suppose that 0 ≤ S ≤ T and the operator S is singular. According to 4.2.6, we may choose an operator U ∈ M (X, Y ), for which U = S. It is clear that U ⊥ T , since U is singular. But then S ⊥ T , i.e., S = 0. This means that T ⊥ Ls (E, F ) and, in view of 4.4.5 (1) and 4.3.2, we obtain T ∈ Mn (X, Y ). B

168

Chapter 4

4.4.7. To formulate our results for the strong form of the Yosida–Hewitt decomposition we need some new definitions and additional facts. A vector lattice E has the Egorov property if for every 0 ≤ e ∈ E and for every double sequence en,k ∈ E, we have that 0 ≤ en,k %k f (for all n ∈ N) implies the existence of a sequence 0 ≤ dm % f with the relation dm ≤ en,j(m,n) holding for m, n arbitrary and j(m, n) suitable. A vector lattice is said to be weakly σ-distributive, if for every e ∈ E and for every double sequence (fm,n )m,n∈N in E+ such that if (fm,n )n∈N increases and o-converges to e for each m ∈ N, there exists a net (eϕ )ϕ∈NN with the properties supϕ eϕ = e and eϕ ≤ em,ϕ(m) for all m ∈ N. The set NN is considered with the natural pointwise ordering: ϕ1 ≤ ϕ2 ⇔ (∀n ∈ N)ϕ1 (n) ≤ ϕ2 (n). It is obvious that a vector lattice with the Egorov property is weakly σ-distributive. The converse is false. We also recall that a vector lattice is said to have the countable sup property, if whenever an arbitrary subset D has a supremum, then there exists an at most countable subset C of D with the same supremum. 4.4.8. Theorem. The following are valid: (1) a countable intersection of σ-order-dense ideals in a vector lattice with the Egorov property is also a σ-order-dense ideal; (2) a countable intersection of σ-order-dense ideals in a weakly σdistributive vector lattice is an order-dense ideal; (3) a countable intersection of order-dense ideals in a vector lattice with both Egorov and countable sup property is also an order-dense ideal. C First we prove (1) and (2). Let (Gk ) be a countable family of σ-order-dense T∞ ideals in a vector lattice E and G = k=1 Gk . Take a nonzero e ∈ E+ . Each finite intersection of Gk is a σ-order-dense ideal. Thus, for any n ∈ N T we may choose n an increasing sequence (en,k ) such that e = supk en,k and en,k ∈ l=1 Gl . Using the Egorov property, we may find an increasing sequence (em ) ⊂ E+ such that e = sup em and for all m and n there is a suitable index l(m, n) with the property em ≤ dn,l(m,n) . This implies that em ∈ G for all m, and G is a σ-order-dense ideal. In the case when E is weakly σ-distributive we may find an increasing net (eϕ ) ⊂ E+ such that e = sup eϕ and eϕ ≤ em,ϕ(m) for all m ∈ N. This implies that eϕ ∈ G for all ϕ ∈ NN , and G is an order-dense ideal. Now, to prove (3) it is sufficient to observe that an order-dense ideal in a vector lattice with the countable sup property is σ-order-dense ideal. B 4.4.9. Let E and F be vector lattices with F order complete. (1) If F has the countable sup property then every essentially positive σ-order continuous operator T : E → F is order continuous.

Dominated Operators

169

C Assume 0 ≤ eτ % e ∈ E and set f = supτ T (eτ ). By the countable sup property of F , we find an increasing sequence dn := eτ (n) such that f = supn T (dn ). Assume that dn % e is false. Then we find h ∈ E and τ0 such that h ≥ dn for all n and g := eτ0 ∨ h − h > 0. Since eτ0 ∨ h − eτ0 ∨ dn ≤ h − dn , we deduce eτ0 ∨ dn − dn ≥ g > 0 and T (eτ0 ∨ dn ) ≥ T (dn ) + T (g). It follows that sup(T (dn ) + T (g)) = f + T (g) ≤ sup T (eτ0 ∨ dn ) ≤ sup T (eτ ) = f. n

n

τ

This is a contradiction since the last formula implies that T (g) = 0 for g > 0 despite of the strict positivity of T . Thus, we have proven that dn % e and, consequently, by σ-order continuity of T , we came to the sought relation f = T (e). B (2) If E has the Egorov property and F has the countable sup property then for every e ∈ E+ the infimum is attained in the formula (see 3.2.5 (1)): πA Tnσ e = inf{πG T e : G ∈ Iσd (A)}. C Let A be the ideal I (e) generated by e. Because of the countable sup property of F , in the formula it suffices to take a countable set of σ-order-dense ideals G(k) ⊂ I (e): Tnσ e = πA Tnσ e = inf πG(k) T e. k∈N

T∞

Put G := k=1 G(k). According to 4.4.8 (1) G is also a σ-order-dense ideal. Therefore, Tnσ e ≤ πG T e ≤ πG(k) T e. By passing to infimum we deduce that Tnσ e ≤ πG T e ≤ inf k πG(k) T e = Tnσ e, whence Tnσ e = πG T e. B (3) If E weakly σ-distributive and F has the countable sup property then for every e ∈ E+ there exists an order-dense ideal G ∈ Iσd (A) such that Tnσ e = πG T e. 4.4.10. Theorem. Let E be a vector lattice with the Egorov property and F be an order complete vector lattice with the countable sup property. Then ∼ (1) L∼ (E, F ) = L∼ nσ (E, F ) ⊕ Lsσ (E, F ); ∼ (2) L∼ (E, F ) = L∼ n (E, F ) ⊕ Ls (E, F ).

C (1): Of course, what we have to prove reads as follows: ∼ ⊥ L∼ nσ = (Lnσ ) .

It suffices to show that ⊥ ∼ (L∼ nσ ) ⊂ Lsσ ,

since the reverse inclusion is obvious.

170

Chapter 4

⊥ If 0 ≤ T ∈ (L∼ nσ ) , then Tnσ = 0. Thus, for an arbitrary e ∈ E+ , in view of 4.4.9 (2), there is some σ-order-dense ideal G ⊂ I (e) such that πG T e = Tn,σ e = 0. By definition of the projection πG T we have G ⊂ NT . Therefore, NT is a σ-orderdense ideal, whence T ∈ L∼ sσ . (2): Repeating the idea of the proof for (1) but using 4.4.9 (3) instead of ⊥ ∼ 4.4.9 (2), we may obtain (L∼ nσ ) ⊂ Ls . Using this inclusion and taking into account ∼ ∼ that Ln ⊂ Lnσ , we have ∼ ⊥ ∼ ∼ ⊥ ∼ ∼ ∼ ⊥ ∼ L∼ = L∼ n ⊕ (Ln ) = Lnσ ⊕ (Lnσ ) = Ln ⊕ [Lnσ ∩ (Ln ) ] ⊕ Lsσ . ∼ ⊥ ⊂ L∼ Therefore, it is sufficient to prove that L∼ s , since the inclusion nσ ∩ (Ln ) ∼ ∼ Lsσ ⊂ Ls is trivial. ∼ ⊥ ⊥ Take 0 ≤ T ∈ L∼ nσ ∩ (Ln ) denote by T0 its restriction to NT . So, we obtain ∼ ⊥ an essentially positive operator T0 ∈ Lnσ (NT , F ). By virtue of 4.4.9 (1) T0 is order continuous. Let S ∈ L∼ n (E, F ) be the smallest positive extension of T0 to E. ⊥ ⊥ and consequently S = 0. implies S ∈ (L∼ The inequality 0 ≤ S ≤ T ∈ (L∼ n) n) ⊥ ⊥ The latter says that NT = {0}. In other words NT is an order-dense ideal and T ∈ L∼ s . B

4.4.11. So, if the hypotheses of the Theorem 2.4.6 are met then Mn (X, Y ) = Ms (X, Y )⊥ , and Ms (X, Y ) is an order-dense ideal in Mn (X, Y )⊥ . In connection with the Yosida–Hewitt Theorem, of interest is the equality Ms (X, Y ) = Mn (X, Y )⊥ = Mos (X, Y ) or, which is the same, the representation M (X, Y ) = Mn (X, Y ) ⊕ Ms (X, Y ). Theorem. Let (X, E) be a decomposable LNS, let E possess the Egorov property, let (Y, F ) be a Banach–Kantorovich space, and let F be a K-space satisfying the countable sup condition. Then the above representation is valid. C The claim follows from 4.2.6 and 4.4.10 (2). B 4.4.12. A vector measure µ is countably additive (completely additive) if ! _ X µ aγ = µ(aγ ) γ∈€

γ∈€

for each countable (arbitrary) family (aγ )γ∈€ of pairwise disjoint elements of A . Denote by dca(A , Y ) and dao(A , Y ) the respective sets of countably additive vector measure and completely additive vector measure.

Dominated Operators

171

(1) It can be easily checked that µ ∈ dca(A , Y ) ⇐⇒ Iµ ∈ Mσ (S(A ), Y ), µ ∈ dao(A , Y ) ⇐⇒ Iµ ∈ Mn (S(A ), Y ). (2) The sets dca(A , Y ) and dao(A , Y ) are bands in da(A , Y ), so that da(A , Y ) = dca(A , Y ) ⊕ dca(A , Y )⊥ ,

da(A , Y ) = dao(A , Y ) ⊕ dao(A , Y )⊥ .

(3) Every µ ∈ da(A , Y ) has a unique representation in the form µ = µn + µos , where µn ∈ dca(A , Y ) and µos ∈ dca(A , Y )⊥ . Moreover, µ = µn + µos , µn = µ n , µos = µ

os .

(4) Let I be an ideal in A and ν ∈ ba+ (A , F ). Then there exists a unique measure πI ν ∈ ba+ (A , Y ) such that ν = πI ν on I and ν(a) = 0 for every a ∈ A satisfying a ∧ b = 0 for all b ∈ I . (5) Let I be an ideal in A and let Y be bo-complete. For every µ ∈ da(A , Y ) there exists a unique measure πI µ ∈ da(A , Y ) such that πI µ = πI µ and µ − πI µ = µ − πI µ . (6) If ν ∈ ba+ (A , F ) then  νn (a) = inf sup{ν(aγ ) : aγ % a} (a ∈ A ),  νc (a) = inf sup{ν(an ) : an % a} (a ∈ A ). 4.5. Extension of Dominated Operators In this section we present a method of extending sequentially order continuous dominated operators which essentially coincides with the classical Daniell–Stone construction of the Lebesgue integral. The method works if and only if the underlying target vector lattice possesses the weak σ-distributivity property. 4.5.1. A vector lattice F is said to be weakly (σ, ∞)-distributive, if whenever (fm,α )α∈A(m) (m ∈ N) is an order-bounded sequence of nets in F such that if (fm,α )α∈A(m) decreases to zero for each m ∈ N, then )

( inf

sup fm,ϕ(m) : ϕ ∈ m∈N

Y m∈N

A(m)

= 0.

172

Chapter 4

A vector lattice F is said to be weakly σ-distributive if whenever (fm,n )m,n∈N is an order-bounded double sequence in F such that if (fm,n )n∈N decreases and o-converges to zero for each m ∈ N, then   N inf sup fm,ϕ(m) : ϕ ∈ N = 0. m∈N

As usual, NN denotes the set of all functions ϕ : N → N. If the decreasing sequences (fm,n )n∈N in this definition are replaced by countable nets or countable directed sets then we obtain formally new formulations. But these two are equivalent to the above (presuming the countable choice axiom). Thus, any weakly (σ, ∞)-distributive lattice is weakly σ-distributive. Theorem. For an arbitrary Kσ -space (K-space) F the following are equivalent: (1) F is weakly σ-distributive (weakly (σ, ∞)-distributive); (2) for every 0 ≤ f ∈ F the Boolean algebra of fragments E(f ) is weakly σ-distributive (weakly (σ, ∞)-distributive); (3) for every 0 ≤ f ∈ F the Stone space of the Boolean algebra E(f ) have the property: the union of every sequence of closed nowhere-dense Gδ -sets is nowhere-dense (each meager set is nowhere-dense). 4.5.2. In order to apply the hypothesis of weak σ-distributivity to the operator extension problem, introduce the following concept. Take a double sequence (fn,m )n,m∈N in F and denote ˆ := NN . We say that the double sequence σ-depresses a nonempty set B ⊂ F towards zero if: (a) (fn,m )m∈N decreases to zero for every n ∈ N, and (b) for every ϕ ∈ ˆ, there is a b ∈ B such that b ≤ supn∈N fn,ϕ(n) . A nonempty set B ⊂ F is said to be σ-depressed towards zero if there is an orderbounded double sequence which depresses B towards zero. (1) If F is a weakly σ-distributive Kσ -space and B is a nonempty set which is σ-depressed towards zero, then inf{b+ : b ∈ B} = 0. C The proof is immediate from the definitions. B (2) If B and C are σ-depressed towards zero, then so is B +C := {b+c : b ∈ B, c ∈ C}. C If (fn,m )n,m∈N
and (gn,m )n,m∈N depress towards zero B and C respectively then fn,m + gn,m n,m∈N depresses B + C towards zero. B (3) If B is σ-depressed towards zero, and every b ∈ B dominates some c ∈ C then C is σ-depressed towards zero. If (fn )n∈N decreases to zero in F , then B := {fn : n ∈ N} is σ-depressed towards zero. C Follows immediately from the given definition. B

Dominated Operators

173

(4) Let (Bn ) be a sequence of nonempty sets in F , each σ-depressed towards zero. Then for every f ∈ F+ the set ( B :=

sup f ∧ n∈N

!

n X

bm

) : (∀m ∈ N) bm ∈ Bm

m=1

is σ-depressed towards zero. C For each n ∈ N choose a double sequence (fn,m,l )m,l∈N depressing Bn towards zero. Put wm,l := sup 2n fn,n−m,l ,

vm,l := f ∧ wm,l

(m, l ∈ N).

n≤m

Then the sequences (wm,l )l∈N and (vm,l )l∈N decrease to zero for each m ∈ N and the second net is order-bounded. Given ϕ ∈ ˆ, we may choose, for each k ∈ N, an element bk ∈ Bk such that bk ≤ sup{fk,m,ϕ(k+m) : m ∈ N}. Since wm+k,l ≥ 2k fk,m,l for k, m, l ∈ N, we may estimate

bk ∧ f ≤ inf f ∧ 2k wk+m,ϕ(k+m) ≤ sup f ∧ 2k wm,ϕ(m) =: uk . m∈N

m∈N

It is easy to check that uk = f ∧ 2n−k un (k ≤ n). Summarizing the indicated properties of uk , we deduce f∧

n X

bk ≤ f ∧

k=1

≤f∧ =f∧

n X k=1 n X k=1 n X

b+ k

=f∧

u+ k =f ∧

n X

(b+ k

k=1 n X

∧ f) = f ∧

n X

(bk ∧ f )+

k=1

uk = f ∧

k=1

n X

f ∧ 2n−k uk




k=1


2n−k un ≤ f ∧ 2n+1 un = f ∧ sup 2n+1 f ∧ 2wm,ϕ(m) m∈N

k=1

= sup vm,ϕ(m) . m∈N

Since n is arbitrary, we have sup vm,ϕ(m) ≥ sup m∈N

m∈N

f∧

n X

! bk

∈ B.

k=1

Thus, B is depressed towards zero by (vm,l )l∈N and the proof is complete. B

174

Chapter 4

4.5.3. Theorem. Let F and G be Kσ -spaces with F weakly σ-distributive. Let E be a massive sublattice in G and let Y be a sequentially bo-complete latticenormed space over F . Suppose that a linear operator T : E → Y has a sequentially G-continuous dominant S : E → F . Then there exist a sequentially o-closed sub^ ⊂ G and a unique pair of sequentially o-continuous operators T^ : E ^→Y lattice E ^ → F such that E ⊂ E, ^ T^ is an extension of T , S is an extension of S, and S^ : E and S^ is a dominant of T^. C The proof is given below in 4.5.4–4.5.9. B 4.5.4. Denote by E  the set of all x ∈ G such that x = supn xn for an increasing sequence (xn )n∈N ⊂ E. It is evident that the set E  is closed under addition and lattice operations. Given x ∈ E  , define S  x := sup{Sxn : n ∈ N}. The least upper bound on the left-hand side exists, since by the massiveness assumption x is dominated by some u ∈ E, and therefore, (Sxn )n∈N is bounded above by Su. If (x0n )n∈N is another sequence in E with sup x0n = x then o-limn,m (x0n −xm ) = 0 in G and, by G-continuity of S, we have o-lim Sx0n = o-lim Sxn . Thus, the operator S  : E  → F is defined correctly. (1) The operator S  is increasing, additive, agrees with S on E, and S  u = sup{S  un : n ∈ N} for an arbitrary sequence (un )n∈N in E  increasing to u ∈ E. If x and xn are as above then the sequence (T xn )n∈N is bo-fundamental, since (o)

T xn − T xm ≤ S(|xn − xm |) ≤ S  (|x − xα |) + S  (|x − xm |) −→ 0. If (x0m )m∈N is another sequence in E increasing to x ∈ E  , then (o)

T xn − T x0m ≤ S  (x − xn ) + S  (x − x0m ) −→ 0. Thus, the relation T  x := bo-lim T xn correctly defined an operator T  : E  → Y . (2) The operator T  is additive, positively homogeneous, and agrees with T on E. Replacing increasing sequences by decreasing sequences in above definitions we may introduce E  and operators S  : E  → F and T  : E  → Y . It can be easily verified that T  x ≤ S  (|x|)

(x ∈ E  ),

T  x ≤ S  (|x|)

(x ∈ E  ).

Moreover, E  = −E  , T  = −(−T ) , and the operators T  and T  coincide on E ∩ E.

Dominated Operators

175

^ ⊂ G to be the set of such x ∈ G that the set 4.5.5. Now define E B(x) := {Su − Sv : u ∈ E  , v ∈ E  , v ≤ x ≤ u} ^ = −E ^ and is σ-depressed towards zero in F . It can be easily observed that E ^ = E ^ for λ > 0, since B(x) = B(−x) and B(λx) = λB(x). Moreover, the λE following is true: ^ is a vector sublattice in G. (1) The set E ^ and x = x1 + x2 . Then the set B(xj ) is σ-depressed towards C Let x1 , x2 ∈ E zero for j := 1, 2. It follows easily from the definition of sets B(x), B(x1 ), B(x2 ) that every member of B(x1 ) + B(x2 ) dominates some member of B(x). Thus, in ^ view of 4.5.2 (2, 3), B(x) is σ-depressed towards zero and x ∈ E. ^ If u ∈ E  , v ∈ E  , and v ≤ x ≤ u. Then u+ ∈ E  , Now suppose that x ∈ E. +  + + + v ∈ E , v ≤ x ≤ u , and S(u+ − v + ) ≤ S(u − v). It follows that every member of B(x) dominates some member of B(x+ ). Thus, B(x+ ) is σ-depressed towards ^ B zero and x+ ∈ E. ^ If x ∈ E then inf B(x) = 0 according to 4.5.2 (1). Therefore, the expressions inf{Su : u ∈ E  , x ≤ u} and sup{Sv : v ∈ E  v ≤ x} have equal values, say w. ^ → F by Thus, we may define an operator S^ : E ^ := inf{Su : u ∈ E  , x ≤ u} = sup{Sv : v ∈ E  v ≤ x} (x ∈ E). ^ Sx It can be easily seen that B(x) is σ-depressed towards zero if and only if there exists w ∈ F such that the sets B + (x) := {Su − w : u ∈ E  u ≥ x} and B − (x) := {−Su + w : v ∈ E  v ≤ x} are both σ-depressed towards zero. Moreover, in this ^ = w. case Sx (2) The operator S^ is linear, positive and agrees with S on E. ^ is sequentially order-closed sublattice in G and S^ : E ^→F 4.5.6. The set E is sequentially order continuous. ^+ with x1 = 0 and let x := C Take an increasing sequence (xn )n∈N in E supn∈N xn exist in G. Fix on some e ∈ E such that x ≤ e, and put f := Se. Let w := supn∈N Sxn and note that w ≤ f . We will show that the set B(x) is σ-depressed towards zero. It is convenience to split the proof into two steps. (1) For each n ∈ N, denote ^ n+1 − xn ) : u ∈ E  , u ≥ xn+1 − xn }. Bn := {Su − S(x ^ n+1 − xn ) exists and Bn is It follows from 4.5.5 that for each n ∈ N the element S(x σ-depressed towards zero. Let B be defined as in 4.5.2 (4). Then B is σ-depressed

176 towards zero. By definition each b ∈ B is expressed as b = supn∈N ^ m+1 − xm ) with um ≥ xm+1 − xm . Denote where bm = Sum − S(x ! n X u := sup e ∧ um . n∈N

Chapter 4
Pn f ∧ m=1 bm

m=1

Observe that u ∈ E  and u ≥ x, since n n X X u≥e∧ um ≥ e ∧ (xm+1 − xm ) ≥ e ∧ xn+1 = xn+1 m=1

m=1

for every n ∈ N. Now taking 4.5.4 (1) into consideration, we deduce ! !! n n X X Su = sup S e ∧ um ≤ sup Se ∧ S um n∈N

= sup Se ∧ n∈N

= sup f ∧ n∈N

m=1 n X

n∈N

! Sum

m=1 n X

= sup f ∧ n∈N

m=1 n X

bm +

m=1

!! ^ n+1 bm + Sx

m=1

≤ sup f ∧ n∈N

n X

m=1 n X

!! ^ m+1 − xm ) S(x

bm

! + w = b + w.

m=1

Thus, every b ∈ B dominates some element of B + (x) := {Su − w : u ∈ E  u ≥ x} and, in view of 4.5.2 (2), B + (x) is σ-depressed to zero. (2) For each n ∈ N, denote ^ n : n ∈ N}, B1 := {w − Sx

^ n : u ∈ E  , u ≤ xn }. Bn+1 := {−Su + Sx

By 4.5.5 Bn is σ-depressed to zero for each n ∈ N. Again B, defined as in 4.5.2 (4), is σ-depressed to zero. But each b ∈ B can be represented as !! n X ^ k+ ^ m) b := sup f ∧ w − Sx (−Sum + Sx n∈N

m=1

^ m ≥ 0 we where k ∈ N and um ∈ E  , u ≤ xn for every n ∈ N. Since −Sum + Sx have ^ k + (−Sum + Sx ^ m )) = f ∧ (w − Sum ) b ≥ f ∧ (w − Sx = w ∧ (w − Sum ) = w + 0 ∧ S(−um ) ≥ w + S(0 ∧ (−um )) = w − S(0 ∨ um ). But 0 ∨ um ∈ E  and 0 ∨ um ≤ xm ≤ x. Thus, every b ∈ B dominates some element of the set B − (x) := {w − Sv : v ∈ E  , v ≤ x} and B − (x) is σ-depressed towards zero. According to 4.5.2 (2) B + (x) + B − (x) is also σ-depressed towards zero. Since ^ Moreover, Sx ^ = w. B B(x) = B + (x) + B − (x) we obtain that x ∈ E.

Dominated Operators

177

4.5.7. Let ˆ := NN and define an order relation in ˆ by ϕ1 ≤ ϕ2 ⇔ (∀n ∈ N) ϕ1 (n) ≤ ϕ2 (n). Clearly, ˆ is directed upward. We say that a net (yϕ )ϕ∈ˆ in Y is ˆ-fundamental (ˆ-converges to y ∈ Y ), if there is a double sequence (fm,n )m,n∈N in F such that (fm,n )n∈N decreases to zero for each m ∈ N and yϕ0 − yϕ0 ≤ sup fm,ϕ(m) m∈N

for all ϕ, ϕ0 , ϕ00 ∈ ˆ with ϕ0 ≥ ϕ and ϕ00 ≥ ϕ (respectively, y − yϕ ≤ sup fm,ϕ(m) m∈N

for all ϕ ∈ ˆ). Let F^ be an order completion of a σ-distributive vector lattice F . Then there exist a lattice-normed space Y^ over F^ and isometric embedding ı : Y → Y^ such that (1) if (yϕ )ϕ∈ˆ is a ˆ-fundamental net in Y , then the net (ıyϕ ) is boconvergent to some element in Y^ ; (2) every member of Y^ is the ˆ-limit of a net (yϕ )ϕ∈ˆ in Y . C Denote by ˆ(Y ) the set of all ˆ-fundamental nets in Y . Define some equivalence relation ∼ in ˆ(Y ) by putting (yϕ ) ∼ (zϕ ) if and only if there exists an orderbounded double sequence (fm,n ) in F such that the sequence (fm,n )n∈N decreases to zero for every m ∈ N and yϕ0 − zϕ00 ≤ sup fm,ϕ(m) m∈N

for all ϕ0 ≥ ϕ and ϕ00 ≥ ϕ. Let Y^ := ˆ(Y )/ ∼ be the factor set and ı be the mapping that sends each y ∈ Y to the coset of the constant net yϕ = y (ϕ ∈ ˆ). If an element y^ ∈ Y^ is the coset of a ˆ-fundamental net (yϕ ), then we set y^ := o-lim yϕ . It follows from the definitions and σ-distributivity assumption that o-lim y^ − yϕ0 = o-lim o-lim yϕ − yϕ0 = 0, 0 0 ϕ

ϕ

ϕ

which proves (1). Claim (2) paraphrases (1). B

178

Chapter 4

^ Take x ∈ E ^ and let a double sequence 4.5.8. Now we will extend T to E. (fm,n )m,n∈N σ-depresses B(x) to zero. Then for every ϕ ∈ ˆ there exist uϕ ∈ E  and v ∈ E  such that v ≤ x ≤ u and S  (uϕ − vϕ ) ≤ sup fm,ϕ(m) . m∈N

^ for u ∈ E , we have Since T u ≤ S(|u|) 



^ ϕ0 − uϕ00 |) T  uϕ0 − T  uϕ00 ≤ S(|u ^ ϕ0 − x) + S(u ^ ϕ00 − x) ≤ sup fm,ϕ(m) ≤ S(u m∈N

for all ϕ0 ≥ ϕ and ϕ00 ≥ ϕ. From this it follows that the net (T  uϕ )ϕ∈ˆ is ˆfundamental. In a similar manner the nets (T  uϕ )ϕ∈ˆ and (T  vϕ )ϕ∈ˆ are equivalent. According to 4.5.7 these nets have the same ˆ-limit in Y^ which is also their bo-limit because of the weak σ-distributivity assumption. Therefore, we may define T^x := lim T  uϕ = lim T  vϕ . ϕ∈ˆ

ϕ∈ˆ

^ ^ are obvious. So, ^ → Y^ and the inequality T^x ≤ S(|x|) (x ∈ E) Linearity of T^ : E it remains to prove that T^(σ(E)) ⊂ Y . Of course, we could use bo-completion from 2.2.11 instead of ˆ-completion from 4.5.7. But it was the sole use of the axiom of choice. Thus, the proof of Theorem 4.5.3 runs by the countable choice axiom. 4.5.9. A subset in G is called monotone if it contains order limits of all monotone sequences in this set. Let E0 be the intersection of all monotone subsets in G including E, so that E0 is the smallest monotone subset in G including E. Denote E1 := {x ∈ σ(E) : T~x ∈ Y }. If the sequence (xn ) ⊂ E1 increases to some x ∈ G then x ∈ σ(E), and (T^xn ) is ^ n − x|) and S^ is sequentially o-continuous. fundamental, since T^xn − T^x ≤ S(|x (see 4.5.7). But Y is sequentially bo-complete and T^x = lim T^xn lie in Y . Thus x ∈ E1 , so that E1 is a monotone subset including E. In particular, E0 ⊂ E1 ⊂ σ(E). It can be shown that the smallest monotone subset containing a vector sublattice is itself a vector sublattice. Therefore, E0 = σ(E) and E1 = σ(E). Thus, T^(σ(E)) ⊂ Y which completes the proof of Theorem 4.5.3. Thus, the Daniell scheme works and results in sequentially order continuous extension to the “Baire completion” if F is weakly σ-distributive. The converse is also true: weak σ-distributivity of F is not only sufficient but also necessary for Theorem 4.5.3 to be valid. There are similar interrelations between possibility of extension to the “Borel completion” and (σ, ∞)-distributivity. We state only one result in this direction.

Dominated Operators

179

4.5.10. Theorem. Let G and F be K-spaces with F weakly (σ, ∞)-distributive. Let E be a massive sublattice in G, and let Y be a bo-complete lattice-normed space over F . Suppose that a linear operator T : E → Y is G-continuous. Then e ⊂ G and an extension Te : E e → Y of T to E e such there exist a vector sublattice E that the following hold: e is a sequentially o-closed sublattice in G containing E ↑ ; (1) E (2) Te is a sequentially order continuous dominated operator; e → F is an extension of T ; (3) Te : E (4) if a net (xα ) ⊂ E ↑ increases to x ∈ E ↑ then Tex = limα Texα . C We confine exposition to a rough sketch. Denote S := T and observe that S : E → F is G-continuous (4.3.8). Given x ∈ E ↑ , we put by definition S ↑ x := sup{Sxα : α ∈ A}. The operator S ↑ : E ↑ → F is correctly defined because E is massive in G and S is G-continuous. Moreover, S ↑ y = sup{S ↑ yα : α ∈ A} for an arbitrary net (yα )α∈A in E ↑ increasing to y ∈ E ↑ . If x and xα are as above then we let T ↑ x := o-lim T xα . This relation correctly defined an additive and positively homogeneous operator T ↑ : E↑ → Y . We say that the sequence of nets (fn,α )α∈A(n) (σ, ∞)-depresses a nonempty set B ⊂ F towards zero if: (a) (fn,α ) decreases to zero for every n ∈ N, and (b) for every Q ϕ ∈ ˆ := n∈N A(n), there is a b ∈ B such that b ≤ supn∈N fn,ϕ(n) . A nonempty set B ⊂ F is said to be (σ, ∞)-depressed towards zero if there is an order-bounded sequence of nets which depresses B towards zero. e ⊂ G comprise such x ∈ G that the set Now let E B(x) := {Sy − Sz : y ∈ E ↑ , z ∈ E ↓ z ≤ x ≤ y} e is a sequentially o-closed is (σ, ∞)-depressed towards zero in F . Then the set E e is a sequentially o-closed vector sublattice vector sublattice in G. Then the set E e → Y with the required properties in G; moreover, there exists an extension Te : E (2)–(4). Further details may be carried out in a similar way. B 4.6. Comments 4.6.1. (1) The notion of dominated operator (4.1.1) appeared in the second half of the 1930s in the papers of L. V. Kantorovich [154, 157, 158, 161]. The notion was motivated by two reasons: a theoretical motif was the development of the general theory of operations in semiordered spaces (see [153, 155–157]) and an applied stimulus was tied with the approximate methods of analysis (see [154, 158, 161]).

180

Chapter 4

Speaking about the latter, it is worthwhile to use the words by L. V. Kantorovich himself [154]: “In proving existence for a solution to various classes of functional equations, the method of successive approximations is of rather frequent use in analysis; in this case, the proof for convergence of the approximations rests on the fact that a given equation may be dominated by some equation of a simple form. Proofs of this sort occur in the theory of integral and differential equations. Invoking semiordered spaces and operations in these spaces allows us to develop a rather easy but complete theory of functional equations of these type in abstract form.” (2) Without additional conditions like decomposability, we cannot guarantee existence of an exact dominant for a dominated operator (4.1.2). Simple examples demonstrate that a nonzero dominated operator can have even disjoint dominants. At the same time, decomposability conditions can be weakened to some extent. We point out three of such possibilities. (2.1) Suppose that X is r-decomposable in the following sense: for every x ∈ X and 0 ≤ e ≤ x , there exists an r-fundamental sequence (xn ) in X (r)

(r)

such that xn −→ e and x − xn −→ x − e. Then each dominated operator from X into Y has an exact dominant. (2.2) Let E be a vector lattice with the principal projection property and let X be d-decomposable. Then each dominated operator from X into Y has an exact dominant. (2.3) Suppose that X is o-decomposable in the following sense: for every x ∈ X and every 0 ≤ e ≤ x there exists an o-fundamental sequence (xn ) (o)

(o)

in X such that xn −→ e and x − xn −→ x − e. Then each order continuous (see 2.3.1) dominated operator from X into Y has an exact dominant. (3) Various classes of dominated operators were studied rather independently. Alongside bounded operators in normed spaces, much attention was paid to regular operators, see [23, 60, 162, 163, 336, 388, 409]. Operators with abstract norm (4.1.3 (3)) were introduced by L. V. Kantorovich [163], see [60, 228] and the bibliography therein. Dominated operators in the sense of 4.1.3 (4) were studied by many authors in generality of various degree and under various names, e.g., see [55, 78, 115-117, 188, 237, 238, 296, 336]. In particular, the formula of 4.1.3 (4) is contained in [55, 238]. The operators of 4.1.3 (5) possess a peculiarity that, after the Boolean-valued representation of the domain and target (see 8.3.3) these operators become bounded in the sense of the theory of normed spaces. This fact provides a possibility of studying generalized orthomorphisms by interpreting, in a suitable Boolean-valued model, results of the theory of bounded operators in normed spaces (see A. G. Kusraev and S. S. Kutateladze [212]). We also point out that operators

Dominated Operators

181

with abstract norm (4.1.3) can be represented as bounded functionals in a suitable Boolean-valued model (see E. I. Gordon [106]). (4) The general formula of 4.1.5, for calculating an exact dominant, and its corollary 4.1.6 were obtained by A. G. Kusraev and V. Z. Strizhevski˘ı [218]. The improvement 4.1.8, connected with the possibility of taking the supremum over disjoint partitions, is proven by S. A. Malyugin [181]. Subdominated operators appeared for the first time in the paper [157] by L. V. Kantorovich. Assertion 4.1.11 (1) is also presented there. (5) Consider a subdominated operator T : X → Y . Denote  Š := S : Y → F : |Ay| ≤ y (y ∈ Y ) (= ∂ · ). For each S ∈ Š, the operator S ◦ T is dominated and the following formula holds: [T ]e = sup{ S ◦ T e : S ∈ Š} (e ∈ E+ ). C By the Hahn–Banach–Kantorovich Theorem the following equality holds: y = sup{|Ay| : A ∈ Š} (y ∈ Y ). From 4.1.11 (1) it follows that the operator S ◦T is dominated and [A◦T ]e = A◦T e (e ∈ E+ ). It remains to apply the formula for [T ] (4.1.10) and [A ◦ T ] (4.1.5) and use associativity of suprema. B 4.6.2. (1) The fact that the space of dominated operators M (X, Y ), where X is decomposable and Y is bo-complete, is a bo-complete LNS is known since the second half of the 1930s, see [157, 158, 161, 163]; however, the question about decomposability has remained open. Decomposability of the space M (X, Y ) (Theorem 4.2.6) was established by A. G. Kusraev and V. Z. Strizhevski˘ı [218] in 1987, see also [181]. E. V. Kolesnikov observed (see [181]) that the decomposability hypothesis for X in Theorem 4.2.6 can be somewhat weakened. More precisely, the following assertions are valid:
(1.1) If X is r-decomposable see 4.6.1 (1) and Y is a BKS then M (X, Y ) is a BKS.
(1.2) If X is o-decomposable see 2.1.3 (3) and Y is a BKS then Mn (X, Y ) is a BKS. 4.6.3. (1) Theorems 4.3.2 and 4.3.3 are established by A. G. Kusraev and V. Z. Strizhevski˘ı [218] for the case of order complete norm lattices, see also [197, 181]. The result on extending an o-continuous regular operator from a vector lattice onto its Dedekind completion (which the proof of 4.3.3 is referred to) is established

182

Chapter 4

by A. I. Veksler [376]. Formula 4.3.5 (3) was mentioned in [199]. Theorem 4.3.7 is valid in a sequential version as well: A dominated operator is sequentially ocontinuous if and only if it is σ-additive (i.e., the condition of complete additivity in 4.3.6 is satisfied only for sequences). For a positive operator, this fact was established by A. G. Pinsker, see [163, Theorem VII.2.45]. (2) Almost integral operators were introduced by G. Ya. Lozanovski˘ı [253]. Theorem 2.5.10 is essentially due to A. G. Kusraev [199]. V. Z. Strizhevski˘ı [362] obtained the following result: Theorem. Let E and F be order complete vector lattices and let En∼ separate the points of E. If E is diffuse then every lattice homomorphism is disjoint from the band J(E, F ). If F is diffuse then every Maharam operator is disjoint from the band J(E, F ). 4.6.4. (1) The question of decomposing a measure, functional, or operator into order continuous and order singular parts was studied from various points of view and has many different versions of stating and solving, e.g., see [23, 213–215, 228, 326, 409] and the bibliography in [60]. The formulas 4.4.3 (1–4) were obtained by E. V. Kolesnikov [178]; and 4.4.3 (10 –40 ) are due to A. G. Kusraev and S. A. Malyugin [213]. Theorem 4.4.5 (2) belongs to C. D. Aliprantis and O. Burkinshaw [21], see also [23, 258, 388] and the survey [60]. Theorem 4.4.6 was proven in [199]; it is an easy consequence of Theorems 4.2.6 and 4.4.5 (2). (2) As another corollary of Theorem 4.2.6, we indicate a version of a result by C. D. Aliprantis and O. Burkinshaw [23, Theorem 4.10] for dominated operators [199]: Theorem. Let X be a decomposable LNS, and let Y be a BKS. Then the following assertions are equivalent: (i) every dominated operator from X into Y is order continuous; (ii) the null ideal of every dominated operator from X into Y is a band; (iii) every nonzero dominated operator from X into Y has nonzero carrier (= the disjoint complement of the null ideal). (3) It is well known that the strong form of the Yosida–Hewitt Theorem is not valid, even in the case of functionals, without some extra assumptions. For example, in the vector lattice E := C[0, 1] there are no nontrivial order continuous functionals (cf. [388]) but the Riemann integral gives an example of a nonsingular functional. The generalized Yosida–Hewitt Theorems 4.4.10 and 4.4.11 were obtained by A. V. Bukhvalov and M. Ya. Yakubson (see [33, 228]; as regards Theorem 2.5.8,

Dominated Operators

183

see [33]). The scalar case of Theorem 4.4.10 was established by W. A. J. Luxemburg [258]. For details about the Yosida–Hewitt Theorem and its applications, see [60, 239]. (4) We observe that Theorems 4.4.10 and 4.4.11 can be slightly improved. Namely, it suffices to require that E possesses an order-dense ideal E0 that admits decomposition into arbitrarily many bands with the Egorov property. An instance of such a vector lattice is provided by a vector lattice E possessing an order-dense ideal with sufficiently many o-continuous functionals. To prove this fact, we may take the order completion of such an ideal, which also has a total set of o-continuous functionals and is realizable as an ideal space (see 1.4.11). (5) The Egorov property was introduced by W. A. J. Luxemburg and A. C. Zaanen [263]. I. I. Shamaev [343, 347] gave an interesting characterization for the Egorov property. A Kσ -space E is said to be weakly σσ-distributive (in [343] this property is called weak σ-distributivity of countable type) if for every orderbounded double sequence (en,k ) ∈ E such that the sequence (en,k )k∈N is decreasing for each n ∈ N there is an increasing sequence (ϕn ) of mappings ϕ : N → N with _ ^ ^ _ en,k = en,ϕk (n) . n∈N k∈N

k∈N n∈N

A weakly σ-distributive Kσ -space is weakly σσ-distributive. The converse is true when E has the countable sup property but fails in general. If M is the set of all strictly increasing mappings N → N then the vector lattice l∞ (M ) is weakly σdistributive but not weakly σσ-distributive [343]. The following fact was established by I. I. Shamaev [343, 347]. Theorem. For an arbitrary Kσ -space E the following are equivalent: (i) E possesses the Egorov property; (ii) E is weakly σσ-distributive; (iii) for every e ∈ E+ the Stone space of the principal ideal E(e) possesses the property: every sequence of closed nowhere-dense sets of type Gδ is contained in a single nowhere-dense set of type Gδ . (6) A vector lattice E is said to have the diagonal property if, for any double sequence (en,k ) ∈ E with the order limits en := o-limk→+∞ en,k (n ∈ N) and e := o-limn→+∞ en , we may choose a strictly increasing sequence of indices (kn )n∈N with e = o-limn→+∞ en,kn . Every Kσ -space with the diagonal property is σσ-distributive; the space l∞ is a σσ-distributive Kσ -space without the diagonal property [343]. A universally complete K-space have the diagonal property if and only if it is σσ-distributive [343]. It is well known that a regular K-space have the diagonal property [163, 388]. D. A. Vladimirov [382] proved that the converse assertion is weaker than the continuum hypothesis.

184

Chapter 4

(7) The countable sup property of F is essential for the validity of Theorem 4.4.9. This is illustrated by the following example (see [33]). Take E = L∞ (0, 1) and, using the Brothers Kre˘ın–Kakutani Theorem, realize E as the space C(Q) for an extremal compact space Q. Let T be the identical embedding of ⊥ ∼ E := C(Q) into F := l∞ (Q). We claim that T ∈ (L∼ n (E, F )) \ Ls (E, F ). It is ∼ ⊥ clear that T does not belong to L∼ s (E, F ). To show that T ∈ (Ln (E, F )) , let Tn be the order continuous part of T and f0 the constant function 1 on Q. Proving Tn (f0 ) = 0 we get that Tn = 0. Let {Uα } be a base for neighborhoods of an arbitrary point q ∈ Q. Set Fα = Q \ Uα and α1 ≥ α2 if Uα1 ⊆ Uα2 . By complete regularity, given α, we find fα ∈ C(Q) with fα (q) = 0 and fα = 1 over Fα . As a consequence, we may affirm that fα ↑ f0 in C(Q) (Q has no isolated points since C(Q) is order isomorphic to L∞ (0, 1)) and by [23, Theorem 4.6] Tn (f0 ) ≤ supα T (fα ) (in `∞ (Q)), hence Tn (f0 )(q) = 0. Since q is arbitrary, we are done. 4.6.5. The extension method of Section 4.5 is essentially the classical Daniell– Stone method for constructing the Lebesgue integral. (1) For positive sequentially G-continuous operators with values in a regular K-space the scheme was implemented for the first time and the corresponding extension result was established in [160] by L. V. Kantorovich; see also [163]. Later an analogous result was obtained by E. J. McShane [280]. The result was extended to weakly (σ, ∞)-distributive Kσ -spaces by K. Matthes in [277] (the local ℵ-regularity condition from [277] is equivalent to (σ, ∞)-distributivity). It was J. D. M Wright [405] who proved that a σ-complete vector lattice admits σ-continuous extension of “preintegrals” (= positive G-continuous operators) if and only if it is weakly σ-distributive. Another proofs of this fact were given by D. H. Fremlin [101] and B. Rieˆcan [329, 328]. In Section 4.5 we follow D. H. Fremlin’s approach and use his concept of a set σ-depressed towards zero [101]. (2) It seems likely that weak σ-distributivity condition for Boolean algebras appeared for the first time in J. von Neumann’s Lectures on Continuous Geometry (see [294]). This work provides examples of a weakly σ-distributive Boolean algebra (the algebra of Lebesgue measurable subsets of the real line modulo zeromeasure sets) and a Boolean algebra which is not weakly σ-distributive (the same algebra modulo meager sets). In [265] weak σ-distributivity is one of the characteristic properties of Boolean algebras with measure, see also [167]. For the general concept of (n, m)-distributivity see in [352]. The equivalence relation (1) ⇔ (3) in 4.5.1 was established independently by J. Kelley and J. C. Oxtoby (see [167]) for Boolean algebras with the countable chain condition and by Z. T. Dikanova [77] for universally complete vector lattices with the countable chain condition. The case of a σ-complete (an order complete) vector lattice is a straightforward generalization. (3) Let A be a subalgebra of a complete Boolean algebra B. A countably additive measure µ : A → Y is called extendable if there is a countably

Dominated Operators

185

additive measure µ ^ : σ(A ) → Y with µ ^(a) = µ(a) for all a ∈ A . A lattice-normed space (or vector lattice) Y is said to have the measure extension property if each countable additive measure µ : A → Y is extendable. Analogously, we say that Y has the operator extension property if Theorem 4.5.3 is valid for all E, G, and T . J. D. M. Wright [402] also proved that a σ-complete vector lattice has the measure extension property if and only if it is weakly σ-distributive. This result remains valid for dominated measures with values in a sequentially bo-complete lattice-normed space. More precisely, the following result is true ([214]): Theorem. For an order σ-complete vector lattice F the following are equivalent: (i) F is weakly σ-distributive; (ii) F has the measure extension property; (iii) F has the operator extension property; (iv) every sequentially bo-complete lattice-normed space Y over F has the measure extension property; (v) every sequentially bo-complete lattice-normed space Y over F has the operator extension property. (4) In [406] J. D. M. Wright also proved that an order σ-complete vector lattice F is weakly (σ, ∞)-distributive if and only if each F -valued Baire measure on each compact space can be extended to a regular Fb-valued Borel measure (Fb is an order completion of F ). In this connection we recall one more result. Say that Y has the strong operator extension property if Theorem 4.5.10 is valid for all E, G, and T . Let A be a subalgebra of a complete Boolean algebra B. Say that a finitely additive measure µ : A → Y is smooth if µ(a) = bo-lim µ(aξ ) for every net (aξ ) ⊂ A increasing to a ∈ A in B. A countably additive measure µ : A → Y is called Borel extendable if there is a countably additive measure µ b : Ac→ Y such thatW(a) Ac is a σ-subalgebra of B containing A ↑ ; (b) µ b is an extension of µ; (c) µ b( aξ ) = bo-lim µ(aξ ) for every increasing net (aξ ) ⊂ A . A lattice-normed space (or vector lattice) Y is said to have the strong measure extension property if each smooth measure µ : A → Y is Borel extendable. The following result was established in [214]. Theorem. For an order complete vector lattice F the following are equivalent: (i) F is weakly (σ, ∞)-distributive; (ii) F has the strong measure extension property; (iii) F has the strong operator extension property; (iv) every bo-complete lattice-normed space Y over F has the strong measure extension property;

186

Chapter 4

(v) every bo-complete lattice-normed space Y over F has the strong operator extension property. (5) In [277] Matthes established that a σ-complete vector lattice E is (σ, ∞)-distributive if and only if the order completion of E is also weakly (σ, ∞)distributive. A similar assertion for weakly σ-distributive vector lattices is false. In [406] J. D. M. Wright (assuming the continuum hypothesis) constructed a σcomplete vector lattice which is weakly (σ, ∞)-distributive but whose order completion is not weakly σ-distributive. S. A. Malyugin informed the author that such a vector lattice can be constructed without employing the continuum hypothesis. Thus, weak (σ, ∞)-distributivity is a strictly stronger condition than weak σ-distributivity. (6) Theorems 4.5.3 and 4.5.10 are proved by a constructive method. There are several extension results for dominated operators that are “pure existence theorems” by nature. We state two such results. The first is an easy consequence of the Hahn–Banach–Kantorovich Theorem, while the second was proved (see A. G. Kusraev and S. A. Malyugin [213, 217]) by the Boolean-valued representation method using the Lipecki–Plachky–Thomsen characterization of extreme extensions [248] (compare the Comments on Chapter 3). Let E and F be vector lattices with F order complete and let E0 be a massive sublattice of E. Denote by ε+ (S0 ) ⊂ L∼ (E, F ) the set of all positive extensions to the whole E of a positive operator S0 : E0 → F . (6.1) Theorem. Let (X0 , E0 ) be a lattice-normed space. Suppose that T0 ∈ M (X0 , F ), while S0 : E0 → F is a dominant of T0 , and S ∈ ε+ (S0 ). Then there exists a linear extension T of T0 to the whole X such that T ≤ S. (6.2) Theorem. Let (Y, F ) be a Banach–Kantorovich space. Suppose that T0 ∈ M (E0 , Y ) and S is an extreme point of the set ε+ ( T0 ). Then there exists a unique dominated operator T : E → Y such that T is an extension of T0 and T = S.

Chapter 5 Disjointness Preserving Operators

In the current chapter, we study disjointness preserving operators in vector lattices and lattice-normed spaces. We will concentrate mainly on decomposition and analytical representation of disjointness preserving operators. Somewhat more general class comprises n-disjoint operators (5.2.1). A dominated operator is ndisjoint if and only if it is representable as a sum of n disjointness preserving dominated operators (5.2.7). Thus, the space of n-disjoint operators has a rather simple structure similar to that of disjointness preserving operators. The simplest representatives of the classes of disjointness preserving operators are band preserving operators. Simplicity of these operators notwithstanding, the question about their order-boundedness is far from triviality. All band preserving operators in a universally complete vector lattice are regular if and only if the vector lattice under study is locally one-dimensional (5.1.2) and a universally complete vector lattice is locally one-dimensional if and only if its base is σ-distributive (5.1.5). There exist nondiscrete locally one-dimensional vector latices (5.1.6, 5.1.7). A disjointness preserving operator defines a ring homomorphism between the Boolean algebras of bands, called the shadow of an operator (5.2.2). In turn, the shadow generates a so-called shift operator (5.3.2) which is a lattice homomorphism defined on a certain order-dense ideal of the universal completion of the domain vector lattice (5.3.1). Both are closely related with the initial disjointness preserving operator and concentrate, in a sense, its multiplicative properties. Using these simplest types of operators we may construct weighted shift operators, i.e. the composites W ◦ S ◦ w of two orthomorphisms w and W and a shift operator S (5.3.4, 5.3.9). An arbitrary disjointness preserving operator is representable as the strongly disjoint sum of weighted shift operators (5.3.6, 5.3.10). Every weighted shift dominated operator admits multiplicative representation, i.e., it can be represented as a composite of a generalized shift operator and the operator of multiplication by an operator-valued weight (5.4.4, 5.4.9). This fact allows us to construct one of

188

Chapter 5

the analytical representations of general disjointness preserving operators (5.4.5, 5.4.10). The notion of decomposable operator (5.5.5) is dual to that of disjointness preserving operator. An order continuous dominated operator is decomposable if and only if its exact dominant is a Maharam operator; i.e., if the dominant is order continuous and preserves order intervals (5.5.6). Decomposable operators admit a simple analytical description (5.5.9). In particular, an analog of the Radon– Nikod´ ym Theorem holds for them in which the role of the integral is played by the exact dominant of the operator (5.5.10). 5.1. Band Preserving Operators In this section we handle the problem: under what conditions all band preserving operators in a universally complete vector lattice are automatically orderbounded. Band preserving operators in lattice-normed space are also considered. 5.1.1. Let G be an arbitrary universally complete vector lattice with a fixed order-unity 1. We introduce some multiplication in G that makes G into a commutative ordered algebra with unity 1. A subset E ⊂ G is said to be locally linearly independent if whenever e1 , . . . , en ∈ E , λ1 , . . . , λn ∈ R, and π is a band projection in G with π(λ1 e1 + · · · + λn en ) = 0 we have πλk ek = 0 for all k := 1, . . . , n. A maximal locally linearly independent set in G is called a local Hamel basis for G. Using the Kuratowski–Zorn Lemma, we may easily deduce the existence of a local Hamel basis for G. A locally linearly independent set E in G is a local Hamel basis for G if and only if for every x ∈ G there exists a partition of unity (πξ )ξ∈„ in P(G) such that πξ x is a finite linear combination of elements of πξ E for each ξ ∈ „. Such representation of πξ x is unique in the band πξ (G). An element e ∈ G+ is called locally constant with respect to f ∈ G+ if e = supξ∈„ λξ πξ f for some numeric family (λξ )ξ∈„ and a family (πξ )ξ∈„ of pairwise disjoint band projections. For each universally complete vector lattice G the following are equivalent: (1) all elements of G+ are locally constant with respect to 1; (2) all elements of G+ are locally constant with respect to an arbitrary order-unity e ∈ G; (3) {1} is a local Hamel basis for G; (4) every local Hamel basis of G consists of pairwise disjoint members. C Obviously, (2) ⇒ (1). To prove the converse, note that for an arbitrary x ∈ G we may choose a partition of unity (πξ )ξ∈„ such that for each ξ ∈ „ both πξ x and

Disjointness Preserving Operators

189

πξ e are multiples of πξ 1. So, πξ x is a multiple of πξ e. A similar argument shows that {1} is a local Hamel basis if and only if so is {f } for every order-unity f ∈ G. Thus, if (4) holds and E is a local Hamel basis for G then f := sup{e : e ∈ E } exists and {f } is a local Hamel basis for G. It follows that (4) ⇒ (3); moreover, (3) ⇒ (1) is obvious. To complete the proof, we had to show (1) ⇒ (4). If (4) fails then we may choose a nonzero band projection π and a local Hamel basis with two members e1 and e2 such that both πe1 and πe2 are nonzero multiples of {π1}. Consequently, π(λ1 e1 + λ2 e2 ) = 0 for some λ1 , λ2 ∈ R and we arrive at the contradictory conclusion that {e1 , e2 } is not locally linearly independent. B A universally complete vector lattice G is called locally one-dimensional if G satisfies the equivalent conditions (1–4) of the above proposition. 5.1.2. Theorem. Let G be a universally complete vector lattice. Then the following are equivalent: (1) G is locally one-dimensional; (2) every band preserving operator T : G → G is order-bounded. C (1) ⇒ (2): First observe that a linear operator T : G → G is band preserving if and only if πT = T π for every band projection π in G. Assume that T is band preserving and put ρ := T 1. Since an arbitrary e ∈ G+ can be expressed as e = supξ∈„ λξ πξ 1, we deduce πξ T e = T (πξ e) = T (λξ πξ 1) = λξ πξ T (1) = πξ (e)T (1) = πξ eρ, whence T e = ρe. It follows that T is a multiplication operator in G which is obviously order-bounded. (2) ⇒ (1): Assume that (1) is false. Fix a local Hamel basis E in G. According to 5.1.1 (4) we may choose e1 , e2 ∈ E that are not disjoint. Then the band projection π := [e1 ] ∧ [e2 ] is nonzero. For an arbitrary x ∈ G there exists a partition of unity (πξ )ξ∈„ such that πξ x is a finite linear combination of elements of E . Assume the elements of E have been labelled so that πξ x = λ1 πξ e1 + λ2 πξ e2 + · · · . Define T x to be a unique element in G with πξ T x := λ1 ππξ e2 . It is easy to check that T is a well defined linear operator from G into itself. Take x, y ∈ G with x ⊥ y and let (πξ )ξ∈„ be a partition of unity such that both πξ x and πξ y are finite linear combination of elements from E . Refining the partition of unity if necessary we may also require that at least one of the elements πξ x and πξ y equals to zero for each ξ ∈ „. If πξ y 6= 0 then πξ x = 0, so the corresponding λ1 e1 = 0. If ππξ 6= 0 then λ1 = 0, and in any case πξ T x = 0. It follows that T x ⊥ y and T is band preserving. If T were order-bounded then T would be presentable as T x = ax (x ∈ G) for some a ∈ G. In particular, T e2 = ae2 and, since T e2 = 0 by definition, we have 0 = [e2 ]|a| ≥ π|a|. Thus πe2 = T (πe1 ) = aπe1 = 0, contradicting the definition of π. B

190

Chapter 5

5.1.3. A subset of a Boolean algebra whose supremum is unity is called a cover of the algebra. A σ-complete Boolean algebra B is called σ-distributive if for every double sequence (bn,m )n,m∈N in B the following conditions hold: _ ^ n∈N m∈N

^ _

bn,m =

ϕ∈NN

bn,ϕ(n) .

n∈N

Let B be an arbitrary Boolean algebra and let C be a cover of B. A subset C0 of the algebra B is said to be refined from C if, for each c0 ∈ C0 , there exists a c ∈ C such that c0 6 c. An element b ∈ B is called refined from C if the set {b} is refined from C, i.e., b 6 c for some element c ∈ C. If (Cn )n∈N is a sequence of covers of the algebra B and an element b ∈ B is refined from each of the covers Cn (n ∈ N), then we say that b is refined from the sequence (Cn )n∈N . We also refer to a cover, all elements of which are refined from the sequence (Cn )n∈N , as refined from the sequence. Let B be a complete Boolean algebra. The following are equivalent: (1) the algebra B is σ-distributive; (2) from every sequence of countable covers of B, it is possible to refine a (possibly, uncountable) cover; (3) from every sequence of finite covers of B, it is possible to refine a (possibly, infinite) cover; (4) from every sequence of countable partitions of unity in B, it is possible to refine a (possibly, uncountable) partition; (5) from every sequence of finite partitions of unity in B, it is possible to refine a (possibly, infinite) partition; C A proof of the equivalence (1) ⇔ (2) can be found in [352, Theorem 19.3]. The implication (2) ⇒ (3) is obvious. Moreover, (3) implies that a cover is refined from every sequence of two-element partitions of B. The latter is a paraphrase of the following equivalent definition of σ-distributivity: for every sequence (bn ) in B it holds _ ^ ε(n)bn = 1 ε∈{1,−1}N n∈N

where 1bn = bn and (−1)bn is the complement of bn . The equivalences (2) ⇔ (4) and (3) ⇔ (5) are easy consequences of the Exhaustion Principle (see 1.1.6). B 5.1.4. Let Q be the Stone space of the base of G, and Clop(Q) be the Boolean algebra of all clopen subsets of Q (see 1.1.7). Say that a function e ∈ C∞ (Q) is refined from a cover C of the Boolean algebra Clop(Q) if, for every two points q 0 , q 00 ∈ Q satisfying the equality e(q 0 ) = e(q 00 ), there exists an element U ∈ C

Disjointness Preserving Operators

191

such that q 0 , q 00 ∈ U . If (Cn )n∈N is a sequence of covers of the algebra Clop(Q) and a function e is refined from each of the covers Cn (n ∈ N), then we say that the function e is refined from the sequence (Cn )n∈N . To each sequence of finite covers of Clop(Q), there is a function in C(Q) refined from it. C Let (Cn )n∈N be a sequence of finite covers of the algebra Clop(Q). By induction, it is not difficult to construct a sequence of partitions Pm = {U1m , U2m , . . . , U2mm } of the algebra Clop(Q) possessing the following properties: (1) for every n ∈ N there is a number m ∈ N such that the partition Pm is refined from the cover Cn ; m+1 m+1 ∨ U2j for all m ∈ N and j ∈ {1, 2, . . . , 2m }. Given (2) Unm = U2j−1 a number m ∈ N, define the two-valued function χm ∈ C(Q) as follows: χm :=

m−1 2X

m χ(U2i ),

i=1

where ) is the characteristic function of a subset U ⊂ Q. Since the series P∞ χ(U 1 m=1 3m χm is uniformly convergent, its sum e belongs to C(Q). We will show that the function e is refined from (Cn )n∈N . By property (1) of the sequence (Pm )m∈N , it is sufficient for this to establish that the function e is refined from (Pm )m∈N . Assume the contrary and consider the smallest number m ∈ N for which the function e is not refined from the partition Pm . In this case, there are two points q 0 , q 00 ∈ Q that satisfy the equality e(q 0 ) = e(q 00 ) and belong to distinct elements of Pm . Since the function e is refined from the partition Pm−1 (for m > 1), from property (2) of the sequence (Pm )m∈N it follows that the points q 0 and q 00 m belong to adjacent elements of Pm , i.e. elements of the form Ujm and Uj+1 , where m 0 j ∈ {1, . . . , 2 − 1}. For definiteness, suppose that q belongs to an element with even subscript and q 00 with odd one, i.e., χm (q 0 ) = 1 and χm (q 00 ) = 0. Therefore, taking into account the fact that χk (q 0 ) = χk (q 00 ) for all k ∈ {1, . . . , m − 1}, we have: ∞ ∞ X X
1 1 1 1 1 0 00 e(q ) − e(q ) = m + χi (q ) − χk (q ) > m − = > 0, k k 3 3 3 3 2 · 3m 0

00

k=m+1

k=m+1

which contradicts the equality e(q 0 ) = e(q 00 ). B 5.1.5. Theorem. A universally complete vector lattice is locally one-dimensional if and only if its base is σ-distributive. C Let G be a universally complete K-space and let Q be the Stone space of its base. Suppose that G is locally one-dimensional and consider an arbitrary sequence

192

Chapter 5

(Pn )n∈N of finite partitions of the Boolean algebra Clop(Q). According to 5.1.3, in order to prove σ-distributivity of the base of G, it is sufficient to refine a cover of Clop(Q) from (Pn )n∈N . In view of Proposition 5.1.4, one can refine a function e ∈ C∞ (Q) from the sequence (Pn )n∈N . Since G is locally one-dimensional, there exists a partition (Uξ )ξ∈„ of the algebra Clop(Q) such that the function e is constant on each of the sets Uξ . Show that the partition (Uξ )ξ∈„ is refined from the sequence (Pn )n∈N . To this end, we fix arbitrary indices ξ ∈ „ and n ∈ N and establish that the set Uξ is refined from the partition Pn . We may assume that Uξ 6= ∅. Let q0 be an element of Uξ . Finiteness of the partition Pn allows us to find an element U of it such that q0 ∈ U . It remains to observe that Uξ ⊂ U . Indeed, if q ∈ Uξ then e(q) = e(q0 ) and, since the function e is refined from Pn , the points q and q0 belong to the same element of the partition Pn , i.e., q ∈ U . Now, assume that the base of G is σ-distributive and consider an arbitrary function e ∈ C∞ (Q). According to 5.1.1 (1), it is sufficient to construct a partition (Uξ )ξ∈„ of the algebra Clop(Q) such that the function e is constant on each of n the interior the sets Uξ . For every natural n and every integer m, denote by Um m+1 m of the  closure of the set of all points q ∈ Q for which n 6 e(q) < n and define n Pn := Um : m ∈ Z . Due to 5.1.3 (4), from the sequence (Pn )n∈N of countable partitions of the algebra Clop(Q), we may refine some partition (Uξ )ξ∈„ . It is not difficult to become convinced that the resultant partition was desired. B 5.1.6. Thus, the question about existence of a purely nonatomic locally onedimensional K-space is reduced to existence of a purely nonatomic σ-distributive complete Boolean algebra. In the next two subsections will be constructed such an algebra. A Boolean algebra B is called σ-inductive if every decreasing sequence of nonzero elements of B admits a nonzero lower bound. A subalgebra B0 of a Boolean algebra B is said to be dense if, for every nonzero element b ∈ B, there exists a nonzero element b0 ∈ B0 such that b0 6 b. If a σ-complete Boolean algebra contains a σ-inductive dense subalgebra then it is σ-distributive. C Let B be a σ-complete Boolean algebra and let B0 be a σ-inductive dense subalgebra of B. Consider an arbitrary sequence (Cn )n∈N of countable covers of B, denote by C the set of all elements in B that are refined from (Cn )n∈N , and assume by way of contradiction that C is not a cover of B. Then there exists a nonzero element b ∈ B that is disjoint with all elements of C. By induction, we construct sequences (bn )n∈N and (cn )n∈N as follows. Let c1 be an element of C1 such that b ∧ c1 6= 0. Since B0 is dense, there is an element b1 ∈ B0 such that 0 < b1 6 b ∧ c1 . Suppose that the elements bn and cn are already constructed. Let cn+1 be an element of Cn+1 such that bn ∧ cn+1 6= 0. As bn+1 we take an arbitrary element of B0 that satisfies the inequalities 0 < bn+1 6 bn ∧ cn+1 .

Disjointness Preserving Operators

193

Thus, we have constructed sequences (bn )n∈N and (cn )n∈N such that bn ∈ B0 , bn 6 cn ∈ Cn and 0 < bn+1 6 bn 6 b for all n ∈ N. Due to the fact that B0 is σinductive, it contains an element b0 which satisfies b0 6 bn for all n ∈ N. In view of the inequalities b0 6 cn , the element b0 is refined from (Cn )n∈N , i.e., belongs to C. On the other hand, b0 6 b, which contradicts disjointness of b with all elements of C. B 5.1.7. As is known, for every Boolean algebra B, there exists a complete Boolean algebra B that contains B as a dense subalgebra (see [352; Section 35]). Such an algebra B is unique to within isomorphism and called a completion of B. Obviously, a completion of a purely nonatomic Boolean algebra is purely nonatomic. In addition, due to 5.1.6, a completion of a σ-inductive algebra is σ-distributive. Therefore, in order to prove existence of a purely nonatomic σ-distributive complete Boolean algebra, it is sufficient to present an arbitrary purely nonatomic σ-inductive Boolean algebra. Examples of such algebras are readily available. For the sake of completeness, we present here one of the simplest constructions. Let B be the Boolean algebra of all subsets of N and let I be the ideal of B consisting of all finite subsets of N. Then the factor algebra B/I is purely nonatomic and σ-inductive. C Pure nonatomicity of the algebra B/I is obvious. In order to prove that the algebra is σ-inductive, it is sufficient to consider an arbitrary decreasing sequence (bn )n∈N of infinite subsets of N and construct an infinite subset b ⊂ N such that the difference b\bn is finite for each n ∈ N. We can easily obtain the desired set b = {mn : n ∈ N} with the help of induction by letting m1 := min b1 and mn+1 := min{m ∈ bn+1 : m > mn }. B 5.1.8. Now, introduce band preserving operators in lattice-normed spaces. Let E and F be vector sublattices of a vector lattice G, and let X and Y be latticenormed spaces over E and F , respectively. As in 2.1.2, we call elements x ∈ X and y ∈ Y disjoint and write x ⊥ y whenever x ∧ y = 0. The operator T is said to be band preserving if x ⊥ y implies T x ⊥ y for all x ∈ X and y ∈ Y . Clearly, if Y is a minorant subset in G+ then T is band preserving if it satisfies either of the following equivalent conditions: (a) x ⊥ g implies T x ⊥ g for all x ∈ X and g ∈ G; (b) { T x }⊥⊥ ⊂ { x }⊥⊥ for all x ∈ X, where {·}⊥⊥ is calculated in G. Obviously, the last definition agrees with the notion of band preserving operator in vector lattices (see 3.3.2). In the rest of this section we assume that E and F are order-dense ideals in a K-space G. Recall also that we identify the Boolean algebras P(G), P(E), and P(X) whenever X is decomposable.

194

Chapter 5

(1) Let X be decomposable. A dominated operator T : X → Y is band preserving if and only if T ∈ Orth(E, F ). In particular every dominated band preserving operator is bo-continuous. C Since a band preserving operator T preserves disjointness, by virtue of 4.1.11 (2) we have T e := sup{ T x : x ≤ e+ } (e ∈ E). From this it follows that e ⊥ f implies T e ⊥ f for every f ∈ F , because x ⊥ f for all x ∈ X with x ≤ e. Thus, T is band preserving. The converse is trivial. B (2) If X and Y are decomposable then a dominated operator T : X → Y is band preserving if and only if T π = πT for all π ∈ P(G). C If T is dominated and band preserving then, in view of (1) and 2.1.3 πT (π ⊥ x) = π T (π ⊥ x) ≤ π T (π ⊥ x) = ππ ⊥ T ( x ) = 0, whenever x ∈ X and π ∈ P(G). This implies πT π ⊥ = 0 or πT = πT π ⊥ . Substituting π ⊥ for π in the latter identity we obtain T π = πT π ⊥ , so that T π = πT . The converse is obvious. B (3) Let X be a bo-complete LNS, and T ∈ Orth(X, Y ). For every orthomorphism a ∈ Orth(E) the product aT (x) is defined in Y and the equality T (ax) = aT (x) holds. C It follows from (2) and the Freudenthal Spectral Theorem. B We will say that an operator T : X → Y is semibounded whenever the following condition holds: if a sequence (xn ) in X br-converges to zero then inf{ T (xn ) : n ∈ N} = 0. 5.1.10. Theorem. The following properties of a band preserving operator T from a BKS into an LNS are equivalent: (1) T is dominated; (2) T is order-bounded; (3) T is semibounded. C The implications (1) ⇒ (2) ⇒ (3) are obvious, (2) ⇒ (1) follows from 4.1.11. It remains to show that (3) ⇒ (1). Assume that a LNS X is order complete and an operator T : X → Y is band preserving and semibounded. Fix an arbitrary positive element e ∈ G and prove that the set { T x : u 6 e} is order-bounded in F . We split the proof into two steps. (a) Show first that the set {T u : u 6 e} is order-bounded in the universally complete K-space G. Without loss of generality, we may assume that G = C∞ (Q),

Disjointness Preserving Operators

195

where Q is an extremal compact space. Denote by D the totality of those points q ∈ Q, for which sup{T u(q) : u 6 e} = ∞. Assume that the set {T u : u 6 e} is not bounded in C∞ (Q). Then, according to [163; Chapter XIII, Theorem 2.32], the clopen set U := int cl D is nonempty. For each natural n and each point q ∈ U ∩ D, consider an element uqn ∈ U satisfying the conditions uqn 6 e and T uqn (q) > n. Denote by Unq a clopen subset of Q such that q ∈ Unq ⊂ U and T uqn (p) > n for all p ∈ Unq . It is clear that, for each n ∈ N the relation supq∈U ∩D Unq = U holds in the Boolean
algebra Clop(Q). In view of the Exhaustion Principle, there exists a family Vnq q∈U ∩D of pairwise disjoint elements of Clop(Q) such that Vnq ⊂ Unqfor P all q ∈ U ∩ D, and supq∈U ∩D Vnq = U . We know that the sum o- q∈U ∩D Vnq uqn exists in the BKS X. Denote this sum by un . For all n ∈ N and q ∈ U ∩ D, we have

q





 Vn T un = T Vnq un = T Vnq uqn = Vnq T uqn > nχVnq . After passing to the supremum over q ∈ U ∩ D, we obtain T un > nχU for all n ∈ N; which, together with the inequalities un 6 e, yields a contradiction with semiboundedness of T . (b) Denote by f the upper envelope of the set {T u : u 6 e} in the Kspace G and show that f ∈ F . Without loss of generality, we may assume that f > 0 on some comeager subset of Q. Then, according to 1.4.2 (4), the set of all points q ∈ Q, for which 0 < sup{T u(q) : u 6 e} = f (q) < ∞, is comeager in Q. For any such point q, consider an element uq ∈ U satisfying the conditions uq 6 e and T uq (q) > f (q)/2. By repeating the idea of step (a) and “mixing up” the elements uq in an appropriate way, we may construct an element u ∈ U such that T u > f /2; whence the containment f ∈ F follows directly. B 5.2. n-Disjoint Operators The main goal of the present section is to describe the order ideal that is generated by disjointness preserving operators (= d-homomorphisms) in the space of dominated operators. For this purpose a new class of n-disjoint operators is introduced. 5.2.1. Let X and Y be lattice-normed spaces over vector lattices E and F respectively. An operator T : X → Y is said to be n-disjoint  if, for all n+1 pairwise disjoint elements x0 , . . . , xn ∈ X, the infimum of the set T xk : k := 0, 1, . . . , n equals zero; symbolically: (∀x0 , x1 . . . , xn ∈ X) xk ⊥ xl (k 6= l) ⇒ T x0 ∧ · · · ∧ T xn = 0. Observe that an operator T : X → Y is disjointness preserving (or, which is the same, d-homomorphism) if it is 1-disjoint, i.e., x1 ⊥ x2 implies T x1 ⊥ T x2 for all x1 , x2 ∈ X.

196

Chapter 5

(1) Let X be decomposable and let F be order complete. A dominated operator T ∈ M (X, Y ) is n-disjoint if and only if its exact dominant T is an ndisjoint operator from E into F . C Sufficiency is obvious. Suppose that the operator T is n-disjoint. Take pairwise disjoint elements e0 , . . . , en ∈ E+ and assign fk := sup T u : u ≤ ek . If uk ≤ ek then uk ⊥ ul (k 6= l); therefore, T u0 ∧ · · · ∧ T un = 0. Passing to the supremum over u0 , . . . , P un in the last equality, we obtain f0 ∧ · · · ∧ fn = 0. m If x1 + · · · + xm ≤ ek then l=1 T xl ∈ {fk }⊥⊥ ; consequently, T ek ∈ {fk }⊥⊥ according to 4.1.5. Thus, T e0 ∧ · · · ∧ T en ∈ {f0 }⊥⊥ ∩ · · · ∩ {fk }⊥⊥ = {0}; hence, T e0 ∧ · · · ∧ T en = 0. B (2) A positive operator S : E → F is n-disjoint if and only if ! n n _ _ S xk = S(x0 ∨ · · · ∨ xk−1 ∨ xk+1 ∨ · · · ∨ xn ) k=0

k=0

for all x0 , . . . , xn ∈ X with x−1 := xn and xn+1 := x0 . C Let S be n-disjoint and take arbitrary x0 , . . . , xn ∈ X. If yk := x0 ∨ · · · ∨ xn − x0 ∨ · · · ∨ xk−1 ∨ xk+1 ∨ · · · ∨ xn then yk ≥ 0 and yk ⊥ yl (k 6= l). Thus, Sy0 ∧ · · · ∧ Syn = 0, which is equivalent to the required identity. Conversely, take a pairwise disjoint collection x0 , . . . , xn ∈ X. Without loss of generality we may assume that xk are positive. Then ! ! n n n X X _ Sxk = S xk = S xk k=0

=

n _

k=0

k=0

S(x0 ∨ · · · ∨ xk−1 ∨ xk+1 ∨ · · · ∨ xn ) =

k=0

n _

X

Sxl ,

k=0 l=0,...,n l6=k

whence we deduce  n ^ k=0

Sxk = −

n _ k=0

=−



n n _ X  X    (−Sxk ) = − Sx − Sx l l  

n _

k=0

X

k=0 l=0,...,n l6=k

Sxl +

l=0,...,n l6=k

n X l=0

Sxl = 0. B

l=0

Disjointness Preserving Operators

197

(3) Under the hypotheses of (1) a dominated operator T : X → Y is disjointness preserving if and only if its exact dominant T : E → F is a lattice homomorphism. In particular, every positive disjointness preserving operator from E to F is a lattice homomorphism. C In view of (1) we need only to prove the second part. For all x, y ∈ E the elements x − x ∧ y and y − x ∧ y are positive and disjoint. If 0 ≤ T ∈ L∼ (E, F ) is disjointness preserving then T (x − x ∧ y) and T (y − x ∧ y) are also positive and disjoint. Thus, 0 = (T x − z) ∧ (T y − z) = T (x) ∧ T (y) − z, where z = T (x ∧ y). B 5.2.2. Suppose that X and Y are decomposable lattice-normed spaces and YT := T (X)⊥⊥ . The shadow of an operator T : X → Y is the mapping shdw := shdw(T ) : P(X) → P(YT ) defined by the formula shdw(π) = [T π(X)]. In other words, shdw(π) is the band projection onto (T π(X))⊥⊥ . Clearly, shdw(π) = sup{[T πx] : x ∈ X}. (1) Let E and F have the projection property. A linear operator T : X → Y is disjointness preserving if and only if its shadow shdw(T ) is a Boolean homomorphism. C Sufficiency is trivial; prove necessity. Without loss of generality, we may assume that YT = Y . Assume that a linear operator T : X → Y is disjointness preserving. Let π1 , π2 ∈ P(Y ) and π1 ∧ π2 = π1 ◦ π2 = 0. Then _ _ _ shdw(π1 )∧shdw(π2 ) = [T π1 x1 ]∧ [T π2 x2 ] = [T π1 x1 ]∧[T π2 x2 ] = 0, x1 ∈X

x2 ∈X

x1 ,x2 ∈X

i.e., shdw(π1 ) ⊥ shdw(π2 ). Moreover, shdw(π1 ) ∨ shdw(π2 ) =

_

[T π1 u] ∨ [T π2 v]

u,v∈X

=

_  _  T (π1 u + π2 v) = [T (π1 + π2 )u] = shdw(π1 ∨ π2 ). u,v∈X

x∈X

It follows that (shdw(π ⊥ ) = shdw(π)⊥ for every π ∈ P(X) and shdw(π1 ) ∨ shdw(π2 ) = shdw(π1 ∨ π2 ) for disjoint π1 , π2 ∈ P(X). Thus, shdw is a Boolean homomorphism, since for arbitrary π1 , π2 ∈ P(X) we have
shdw(π1 ∨ π2 ) = shdw (π1 \π2 ) ∨ (π1 ∧ π2 ) ∨ (π2 \π1 ) = shdw(π1 \π2 ) ∨ shdw(π1 ∧ π2 ) ∨ shdw(π2 \π1 )
= shdw(π1 \π2 ) ∨ shdw(π1 ∧ π2 )
∨ shdw(π1 ∧ π2 ) ∨ shdw(π2 \π1 ) = shdw(π1 ) ∨ shdw(π2 ). B

198

Chapter 5

(2) Let F be order complete and T be dominated. Then the shadows of T and T coincide. C Denote h := shdw(T ) and h0 := shdw( T ). Of course, coincidence of the functions h : P(X) → P(Y ) and h0 : P(X) → P(Y ) is understood with the identifications P(X) = P(E) and P(Y ) = P(F ). The inequality h(π) 6 h0 (π)
π ∈ P(E) is obvious. To prove the reverse inequality, it is sufficient to observe, that for e ∈ E, π ∈ P(E), and u1 , . . . , un ∈ X the inequality u1 + · · · + un 6 πe implies 

T u1 + · · · + T un



=



T πu1 + · · · + T πun



= [T πu1 ] ∨ · · · ∨ [T πun ] 6 h(π),

and to use 4.1.5. B Let h : P(X) → P(Y ) be a ring homomorphism. We say that the mapping T : X → Y is h-o-continuous whenever h-limα∈A xα = x (see 2.2.10) implies o-limα∈A T xα = T x for every net (xα )α∈A in X and every x ∈ X. (3) Theorem. Let E and F be K-spaces. Every disjointness preserving operator T : E → F is h-o-continuous, where h is the shadow of T . C Since the shadow of |T | coincides with the shadow of T (see (2)), we may assume that the operator T is positive. To prove h-o-continuity of T , it is sufficient to consider a net (eα )α∈A in E with h-limα∈A eα = 0 and to show that o-limα∈A T eα = 0. According to 1.3.7 (2), o-convergence of T eα to zero will be established if we prove that o-limα∈A [T e][(T eα − T e/n])+ ] = 0 for all e ∈ E and n ∈ N. The latter relation can be obtained as follows: 
+  

 
(o) [T e] T (eα − e/n) = [T e] T (eα − e/n)+ 6 h [e] h (eα − e/n)+ → 0. B (4) Every disjointness preserving dominated operator from a BKS into an LNS is h-o-continuous, where h is its shadow. C The claim follows from (2) and (3). B 5.2.3. Theorem. Let X and Y be decomposable and let E and F be order complete. Suppose that an operator T ∈ M (X, Y ) is disjointness preserving. Then for every dominated operator S ∈ M (X, Y ) the following are equivalent: (1) S ∈ {T }⊥⊥ ; (2) Sx ∈ {T (x)}⊥⊥ (x ∈ X);
(3) Sπ = shdw(π)S π ∈ P(X) ; (4) S π = shdw(π) S


π ∈ P(X) .

Disjointness Preserving Operators

199

C (1) ⇒ (2): Obvious. (2) ⇒ (3): Fix arbitrary elements x ∈ X and π ∈ P(X). From (2) it follows that Sπx and Sπ ⊥ x are disjoint. Consequently, there exist a projection ρ ∈ P(Y ) such that Sπx = ρSx and Sπ ⊥ x = ρ⊥ Sx. Denote h := shdw(π). In order to ensure the equality ρSx = h(π)Sx, it is sufficient to show that ρ[Sx] = h(π)[Sx]. The relations ρ[Sx] = [Sπx] 6 h(π) imply the inequality ρ[Sx] 6 h(π)[Sx]. We may establish similarly that ρ⊥ [Sx] 6 h(π ⊥ )[Sx]. The two last inequalities directly imply the equality ρ[Sx] = h(π)[Sx]. (3) ⇒ (4): This follows from 5.2.2 (2). (4) ⇒ (1): Since S ∈ {T }⊥⊥ means that S ∈ { T }⊥⊥ , we may assume S := S and T := T . Suppose V ⊥ T for some V ≤ S and put
U := V + T . By the proven implication (1) ⇒ (3) U π = shdw(π)U π ∈ P(X) . In particular, U is a lattice homomorphism, as for disjoint e1 , e2 ∈ E+ we have U e1 ∧ U e2 = shdw([e1 ])U (e1 ) ∧ shdw([e2 ])U (e2 ) = 0. According to 3.3.3 V = ρU and T = ρ0 U for some orthomorphisms 0 ≤ ρ, ρ0 ≤ IF . Disjointness of T and V implies that 0 = inf{ρU e1 + ρ0 U e2 : e = e1 + e2 } ≥ inf{ρf1 + ρ0 f2 : f1 + f2 = U e} = (ρ ∧ ρ0 )U e ≥ 0 for every e ∈ E+ . Since (ρ ∧ ρ0 )f = (ρf ) ∧ (ρ0 f ) for f ∈ F+ we obtain V e ∧ T e = (ρU e) ∧ (ρ0 U e) = 0. But then for every e1 , e2 ∈ E+ we deduce V e1 ∧ T e2 = V e1 ∧ T (e1 + e2 ) and V e1 ∧ T e2 = V (e1 + e2 ) ∧ T e2 , so that V e1 ⊥ T e2 . Thus, if π := [e] then [V e] ⊥ shdw(π) and simultaneously V e = V πe = shdw(π)V e, i.e., [V e] ≤ shdw(π). It follows that V = 0. B 5.2.4. Our next goal is to prove that a dominated n-disjoint operator is representable as a sum of disjointness preserving operator. To this end, we need some auxiliary constructions. Let T and S be order-bounded operators from E to F . Define the set of order-bounded operators Z (S) and the mapping · : Z (S) → Orth(F ) as follows: Z (S) := {T ∈ L∼ (E, F ) : (∃ρ ∈ Orth(F )+ ) (|T | ≤ ρ ◦ S)}, T := inf{ρ ∈ Orth(F )+ : |T | ≤ ρ ◦ S} (T ∈ Z (S)). Observe that Z (S) coincide with Z (p) as defined in 3.3.9 if p(x) := S(|x|) (x ∈ E). (1) The triple (Z (S), · , Orth(F )) is a Banach–Kantorovich lattice with an order semicontinuous M -norm. C The fact that (Z (S), · , Orth(F )) is a Banach–Kantorovich space follows from 3.3.9 (2). Monotonicity of the norm · is seen from its definition. In particular,

200

Chapter 5

T1 ∨ T2 ≤ T1 ∨ T2 for positive T1 , T2 . But the reverse inequality is also true, as T1 ≤ T1 S and T2 ≤ T2 S, whence T1 ∨T2 ≤ ( T1 ∨ T2 )S. If (Tα ) is an increasing family of positive operators in Z (S) with T = supα Tα then the equivalences hold T ≤ γ ⇔ (∀α) Tα ≤ γ ⇔ sup T α ≤ γ. Thus, T = supα T α . B The space Z (S) has a natural module structure over Orth(F ) given by ρT := ρ ◦ T . At the same time, according to 2.1.8 Z (S) admits the structure of a module over Orth(Orth(F )). These two structures are identical in the following sense. (2) For every ρ ∈ Orth(F ) there exists a unique ρ¯ ∈ Orth(Orth(F )) such that ρ ◦ T = ρ¯ ∗ T for all T ∈ Z (S). C According to 3.3.2 (3) the f -algebras Orth(F ) and Orth(Orth(F )) are isomorphic. An isomorphism is given by assigning to ρ ∈ Orth(F ) the orthomorphism ρ¯ : σ 7→ ρ ◦ σ (σ ∈ Orth(F )). Given T ∈ Z (S) and ρ ∈ Orth(F ) the product ρ ∗ T is defined by ρ¯∗T = ρ¯∗ T = ρ◦ T and T −ρ¯∗T = T −ρ¯∗ T = ρ⊥ ◦ T . By virtue of 3.3.9 ρ ◦ T satisfies the same identities: T − ρ ◦ T = ρ⊥ ◦ T , ρ ◦ T = ρ ◦ T . Therefore, ρ¯ ∗ T = ρ ◦ T , since disjoint decomposition is unique by 2.1.2 (3). B (3) Let X be an arbitrary vector space and let p1 , . . . , pn : X → F be sublinear operators. Then the following holds: ∂(p1 ∨ · · · ∨ pn ) =

[

{∂(σ1 ◦ p1 ) + · · · + ∂(σn ◦ pn )} .

σ1 ,...,σn ∈Orth(F )+ σ1 +···+σn =IF

C The proof can be found in [209]. B (4) Let T ∈ L∼ (E, F ) and p(x1 , . . . , xn ) := T (x1 ∨ · · · ∨ xn ). Then p is a sublinear operator from E n to F and ( ) n X ∂p = (T1 , . . . , Tn ) : 0 ≤ Tk ∈ L∼ (E, F ), Tk = T , k=1

where (T1 , . . . , Tn ) denotes the linear operator (x1 , . . . , xn ) 7→ (T1 x1 + · · · + Tn xn ). C If U is the right-hand side of the desired equality then obviously ∂p ⊃ U . Conversely, suppose that 0 ≤ Tk ∈ L∼ (E, F ) and T1 x1 +· · ·+Tn xn ≤ T (x1 ∨· · ·∨xn ) for all xk ∈ E (k = 1, . . . , n). Substituting xl = 0 for l 6= k and xk ≤ 0 in this inequality we get Tk ≥ 0. Putting x1 = · · · = xn we arrive at T = T1 + · · · + Tn . B 5.2.5. Theorem. Let E and F be vector lattices with F order complete, and let T be an order-bounded operator from E to F . The following are equivalent:

Disjointness Preserving Operators

201

(1) T is an n-disjoint operator; (2) for every collection of n+1 positive operators T0 , . . . , Tn ∈ L∼ (E, F ) with |T | = T0 + · · · + Tn there exist collections of operators {Tk,l : k, l := 0, 1, . . . , n} and {σl : l = 0, 1, . . . , n} such that 0 ≤ σl ∈ Orth(F ), 0 ≤ Tk,l ∈ L∼ (E, F ), Tk,k = 0, n n n X X X σl = IF , σl Tk,l = |T |, σl Tk,l = Tk (k, l = 0, 1, . . . , n); l=1

k=1

l=1

(3) for a collection of n + 1 pairwise disjoint operators T0 , . . . , Tn ∈ L (E, F ) with |T | = T0 + · · · + Tn there exists a collection of orthomorphisms σ0 , . . . , σn ∈ Orth(F )+ such that σ0 +· · ·+σn = IF and σk ◦Tk = 0 (k = 0, 1, . . . , n); (4) for a collection of n + 1 pairwise disjoint operators T0 , . . . , Tn ∈ L∼ (E, F ) with |T | = T0 + · · · + Tn there exists a partition of unity π0 , . . . , πn in P(F ) such that πk ◦ Tk = 0 (k = 0, . . . , n); (5) T is a norm-n-decomposable element of the Banach–Kantorovich lattice Z (T ). C Without loss of generality, we may assume that T ≥ 0. (1) ⇔ (2): Define the operators p, p0 , p1 , . . . , pn : E n+1 → F by ∼

p(x0 , x1 , . . . , xn ) := T (x0 ∨ x1 ∨ · · · ∨ xn ), pk (x0 , x1 , . . . , xn ) := T (x0 ∨ · · · ∨ xk−1 ∨ xk+1 ∨ · · · ∨ xn ). It is easy to check that p, p0 , . . . , pn are increasing sublinear operators. In view of 5.2.1 (2) p = p0 ∨ p1 ∨ · · · ∨ pn . The latter is equivalent to ∂p = ∂(p0 ∨ p1 ∨ · · · ∨ pn ) by virtue of 3.1.9 (1). Using 5.2.4 (4), we deduce ( ∂p =

(T0 , T1 , . . . , Tn ) : 0 ≤ Tk ∈ L∼ (E, F ),

n X

) Tk = T

,

k=0

( ∂pl =

(T0,l , T1,l , . . . , Tn,l ) : 0 ≤ Tk,l ∈ L∼ (E, F ), Tk,k = 0,

n X

) Tk = T

.

k=0

It remains to apply 5.2.4 (3). (2) ⇒ (3): Assume that T = T0 + · · · + Tn with pairwise disjoint T0 , . . . , Tn ∈ L∼ (E, F )+ . Let Tk,l and σl be as in (2). Then σl Tk,l ⊥ Tl (k 6= l) and, by summing the left-hand side over k = 0, . . . , n, we arrive at σl T ⊥ Tl . But σk Tk ≤ Tk and σk Tk ≤ σk T , so that σk Tk = 0.

202

Chapter 5

(3) ⇔ (4) ⇔ (5): These equivalences follow from Definitions 2.1.9 and 5.2.4 (2). (5) ⇒ (1): Take a collection e0 , . . . , en of pairwise disjoint elements in E. Put πk := πek (k = 1, . . . , n) and π0 := (π1 + · · · + πn )⊥ where the band projections π1 , . . . , πn in L∼ (E, F ) are defined as in 3.1.4 (5). If Tk := πk T then Tk ⊥ Tl (k 6= l) according to 3.1.4 (2) and Tk (ek ) = T (ek ) in view of 3.1.3 (1). Moreover, T = T0 + T1 + · · · + Tn , so that T0 ∧ · · · ∧ Tn = 0 by Definition 2.1.9. But |T ek | = |Tk ek | ≤ Tk (|ek |) ≤ Tk T (|ek |) ≤ Tk f where f := T (|e0 |)∨· · ·∨T (|en |). B 5.2.6. Let X be a lattice-ordered module over a lattice-ordered ring A. An element x ∈ X is called A-discrete or module-discrete whenever for any y ∈ X with 0 ≤ y ≤ |x| there exists a ∈ A such that 0 ≤ a ≤ 1A and y = a|x|. In other words x ∈ X is A-discrete if and only if the equality of order intervals holds: [0, |x|] = [0, 1A ]|x|. Obviously, for the vector lattices E and F , with F order complete, L∼ (E, F ) is a lattice-ordered Orth(F )-module. Theorem 3.3.3 claims that an operator T ∈ L∼ (E, F ) is an Orth(F )-discrete if and only if T is disjointness preserving. (1) Let X be a decomposable br-complete lattice-normed lattice over a Kσ space E. Assume that X has the strong Freudenthal property. Then an arbitrary element x ∈ X is norm-indecomposable if and only if it is Orth(F )-discrete. C Assume that x ∈ X is Orth(F )-discrete and |x| = u + v for order disjoint elements u, v ∈ X+ . By assumption there exist positive orthomorphisms ρ, σ ∈ Orth(F ) such that u = ρ|x| and v = σ|x|. Since X admits a compatible structure of a module over Orth(F ) we may write u ∧ v = ρ|x| ∧ σ|x| = ρ x ∧ σ x = (ρ ∧ σ) x = (ρ ∧ σ)|x| = (ρ|x|) ∧ (σ|x|) = u ∧ v = 0. Let x be a positive norm-indecomposable element. If e is a fragment of x then by definition e ∧ x − e = 0 and x − e = x − e in view of 2.1.2 (2). Thus, e is a fragment of x if and only if e is a fragment of x . If π is the projection onto the band e then πx − e = π x − e =P π( x − e ) = 0, so that e = πx. It follows n that every finite-valued element x0 := k=1 λk ek inPX where ek is a fragment of n x is representable as x0 = σx with πk = [ek ], σ := k=1 λk πk ∈ Orth(F ). Using the Freudenthal Spectral Theorem and br-completeness of X, we conclude that x is discrete. B (2) Assume that the hypothesis of (1) are satisfied. Then every norm-ndecomposable element is a sum of n pairwise norm disjoint module discrete elements. C It follows from (1) and 2.1.10. B

Disjointness Preserving Operators

203

5.2.7. Theorem. Suppose that X is decomposable and F is order complete. A dominated operator T : X → Y is n-disjoint if and only if T is representable as a sum of n disjointness preserving dominated operators. C Apply 5.2.1 (1), 5.2.5 (5), 2.1.10, and decomposability of the exact dominant (see 4.2.6). B ∼ Denote by H := H(E, F ) the o-ideal by lattice ho generated in L (E, F ) momorphisms and assign MH (X, Y ) := T ∈ M (X, Y ) : T ∈ H . A positive operator preserves disjointness if and only if it is a lattice homomorphism. Therefore, from 3.3.3, 5.2.1 (1) and 5.2.7 it follows that if F is order complete then H consists of all order-bounded operators n-disjoint for some n ∈ N. In addition, if X is decomposable then MH (X, Y ) consists of all dominated operators n-disjoint for some n ∈ N. At the same time, these spaces consist of finite sums of disjointness preserving operators. Thus, the space MH has a rather simple structure modulo d-homomorphisms. We close this section with the boundedness criteria for disjointness preserving operators. As we have seen in 5.1.10 a band preserving operator is order-bounded if and only if it is semibounded. This is not true for general disjointness preserving operators but remains valid for a disjointness preserving operator defined on a vector lattice. 5.2.8. Theorem Let T be a disjointness preserving linear operator on a vector lattice E with values in a lattice-normed space Y . Then the following are equivalent: (1) T is dominated; (2) T is order-bounded; (3) T is semibounded. C The implications (1) ⇒ (2) ⇒ (3) are obvious, (2) ⇒ (1) follows from 4.1.11. It remains to show that (3) ⇒ (2). We will split the proof into several elementary steps. (1): Let us first observe that the claim is true if E possesses the strong Freudenthal property. This step is not necessary but in this case the proof becomes particularly simple and clear. Indeed, assume that an operator T enjoys condition (3). Take arbitrary elements x, y ∈ E with |x| 6 |y|, and denote by S the set ( n ) X πk λk |y| : πk ∈ P(E), |λk | 6 1 . k=1

It is not difficult to become convinced that T s 6 T y for all s ∈ S . Moreover, by the strong Freudenthal property, there exists a sequence (sn )n∈N in S that is relatively y-uniformly convergent to x. Condition (3) together with the relations T x 6 T x − T sn + T y (n ∈ N) now yields the desired inequality T x 6 T y .

204

Chapter 5

The general case is considered in (2)–(6). (2): If e1 , e2 ∈ E then T e1 ∧ T e2 ≤ T (e1 ∧ e2 ) . C Indeed, (e1 −e1 ∧e2 ) ⊥ (e2 −e1 ∧e2 ), so that T (e1 −e1 ∧e2 ) ⊥ T (e2 −e1 ∧e2 ). Hence T e1 ∧ T e2 − T (e1 ∧ e2 ) = ( T e1 − T (e1 ∧ e2 ) ) ∧ ( T e2 − T (e1 ∧ e2 ) ) ≤ T (e1 − e1 ∧ e2 ) ∧ T (e2 − e1 ∧ e2 ) = 0. B Take x, y ∈ E with 0 ≤ y ≤ x and prove that T y ≤ T x , so that T is certainly order-bounded. (3): If uλ := T ((y − λx)+ ) and vλ := T ((y − λx)− ) then for each λ ∈ [0, 1] uλ ∧ vλ ≥ T y − T x ;

0 ≤ µ ≤ λ ⇒ uλ ∧ vµ = 0.

C If λ ∈ [0, 1] then (y − λx)+ ⊥ (y − λx)− , so that T ((y − λx)+ ) ∨ T ((y − λx)− ) = T ((y − λx)+ ) + T ((y − λx)− ) ≥ T ((y − λx)+ ) − T ((y − λx)− ) = T (y − λx) ≥ Ty − λ Tx ≥ Ty − Tx . Note also that if 0 ≤ µ ≤ λ then (y − λx)+ ⊥ (y − µx)− , so that uλ ∧ vµ = 0. B (4): For a collection of reals 0 = λ0 < λ1 < λ2 < · · · < λn = 1 the inequality holds: n n _ _ (uλi ∧ uλi ). (uλi−1 ∧ vλi ) ≥ uλ1 ∨ i=1

i=2

C Put f k = vλ 1 ∨

k _

(uλi−1 ∧ vλi ) ∨ uλk

(k = 2, . . . , n),

i=2

and observe that fk ∧ vλk+1 = (vλ1 ∧ vλk+1 ) ∨

k _

(uλi−1 ∧ vλi ∧ vλk+1 ) ∨ (uλk ∧ vλk+1 )

i=2

≤ vλ1 ∨

k+1 _

(uλi−1 ∧ vλi ).

i=2

Moreover, using (3), we deduce fk ∧ uλk+1 = (vλ1 ∧ uλk+1 ) ∨

k _

(uλi−1 ∧ vλi ∧ uλk+1 ) ∨ (uλk ∧ uλk+1 )

i=2

= uλk ∧ uλk+1 ≤ uλk+1 .

Disjointness Preserving Operators

205

We see then that fk ∧ (uλk+1 ∧ vλk+1 ) = (fk ∧ uλk+1 ) ∧ (fk ∧ uλk+1 ) ≤ fk+1 . It follows from a simple induction argument that n ^

fn ≥

(uλi ∨ vλi ),

i=1

and as uλn = 0 we arrive at the desired relation. B (5): Prove that T z ≥ T y − T x if −

z := (y − λ1 x) ∨

n _


(y − λk−1 x)+ ∧ (y − λk x)− .

k=2

C If 2 ≤ k, l ≤ n and k 6= l, then (y − λ1 x)− , (y − λk−1 x)+ ∧ (y − λk x)− ⊥ (y − λl−1 x)+ ∧ (y − λl x)− . Applying successively of (2), (4) and (3) we deduce

T z = T (y − λ1 x)− ∨

n _

T (y − λi−1 x)+ ∧ (y − λi x)−




i=2

≥ vλ 1 ∨

n _

(uλi−1 ∧ vλi ) ≥ T y − T x . B

i=2

(6): Now note that if δ = max{λk − λk−1 : k = 1, . . . , n} then (y − λk−1 x)+ ∧ (λk x − y)+ ≤ (y − λk−1 x + λk x − y)+ ≤ δx, (y − λ1 x)+ ≤ λ1 x = (λ1 − λ0 )x ≤ δx, so that 0 ≤ z ≤ δx. If for each m ∈ N we choose a partition of [0, 1] with δ < 1/m and let zm V be the corresponding z, then zm → 0 relatively uniformly so that by ∞ hypothesis m=1 T zm = 0. Proposition (5) now shows that T y − T x ≤ 0 as required. B

206

Chapter 5 5.3. Weight-Shift-Weight Factorization

The main result of the present section is representation of an arbitrary disjointness preserving operator as a strongly disjoint sum of operators admitting weight-shift-weight factorization. 5.3.1. Throughout this section, E and F are order-dense ideals of some universally complete vector lattices E and F . In the spaces E and F , we fix orderunities 1E and 1F and consider the multiplication that makes the spaces f -algebras with unities 1E and 1F , respectively. We recall that orthomorphisms in the Kspaces under consideration are multiplication operators and we identify them with the corresponding multipliers. For every f ∈ E , there exists a unique element g ∈ E such that f g = [f ]1E . We denote such an element g by 1/f := 1E /f . The product e(1/f ) is denoted by e/f for brevity. The ideal of the K-space E which is generated by the element 1 := 1E is denoted by E (1) := E (1). Observe that some notions in this section depend on a specific choice of the unities 1E and 1F . Let h : P(E) → P(F ) be a Boolean homomorphism. Denote by E (1, h) the set of elements x ∈ E representable as x = o-

∞ X

πn xn

n=1

where (xn ) is an arbitrary sequence in E (1) and (πn ) is a countable partition of unity in P(E ) such that (h(πn )) is a partition of unity in P(F ). Of course, E = E (1, h) if and only if h is sequentially order continuous. (1) The set E (1, h) is an order ideal in E and E (1) is relatively uniformly dense in E (1, h). C It is clear that E (1, h) is a subspace in E . If x is as in the definition of E (1, h) and |y| P ≤ |x| then πn |y| ≤ πn |xn | ∈ E (1), so that yn := πn yn ∈ E (1). Moreover, ∞ y = o- n=1P πn yn whence y ∈ E (1, h). Furthermore, using the same notation, P∞ ∞ denote zn = k=1 πk xk and e = o- n=1 nπn yn . Then zn ∈ E (1), e ∈ E (1, h), and |x − zn | ≤ (1/n)e (n ∈ N). B Let D be a subset of E, and let Y be an LNS. We say that an operator T : E → Y is wide on the set D whenever T (D)⊥⊥ = T (E)⊥⊥ . If d is a positive element of E , then we say that an operator T is wide at the element d whenever it is wide on the set {e ∈ E : e is a fragment of d}. (2) Suppose that T : E → Y is a disjointness preserving operator, and h : P(E) → P(Y ) is its shadow. Then T is wide on the set D ⊂ E if and only if every element e ∈ E is the h-limit of some net in the order ideal ED generated by D.

Disjointness Preserving Operators

207

C First observe that the operator T is wide on the set D if and only if it is wide on the ideal ED generated by D. Assume T to be wide on ED , consider an arbitrary element e ∈ E, and show that h-limπ∈… πe = e, where … = {π ∈ Pr(E) : πe ∈ ED }. For every n ∈ N and d ∈ ED , assign πn := [(n|d| − |e|)+ ]. Obviously, πn ∈ …. Since |d − πn d| = πn


⊥

|d| 6 πn


⊥

|e|/n 6 |e|/n

for all n ∈ N, we have r-limn→∞ πn d = d.
Using r-continuity of the operator T
and taking account of the equality T πn d = h πn T d, we arrive at the relation
supn∈N h πn > [T d]. Since the element d ∈ ED was chosen arbitrarily, we conclude by (2) that supπ∈… h(π) = h(1) and, consequently, h-limπ∈… πe = e. Conversely, consider an arbitrary element e ∈ E and let e be the h-limit of some net (eα )α∈A of elements in ED . In view of 5.2.2 (4), we have o-limα∈A T (eα ) = T e. Thus, the set T (ED ) is o-dense in T (E) and T is wide on T (ED ). B (3) Let T ∈ L∼ (E, F ) be a disjointness preserving operator and h := shdw(T ). Then E ⊂ E (1, h) if and only if T is wide on 1. C Follows immediately from (2). B 5.3.2. For every ring homomorphism h : Pr(E ) → Pr(F ), there exists a unique regular operator S : E (1, h) → F such that the shadow of S is equal to h and S(1E ) = h(1)1F . Furthermore, the operator S is positive. C For the sake of convenience, assume that h(1) = 1. We split the construction of the operator S into three steps. (1): Define the operator S on the set of step-elements of E by letting S

n X i=1

! λi πi 1E

:=

n X

λi h(πi )1F

i=1

for arbitrary λ1 , . . . , λn ∈ R and π1 , . . . , πn ∈ Pr(E ). (2): Extend the operator S onto E (1). To this end, fix an arbitrary element e ∈ E (1) and choose a sequence (en )n∈N of step-elements in E so that it r-converges to e with regulator 1E . It is easy to verify that the sequence (Sen )n∈N is r-fundamental (with regulator 1F ). Assign Se := r-limn→∞ Sen . (3): Finally, extend S onto the P entire set E (1, h). Every element e ∈ E (1, h) can be represented as the mixing on∈N πn en of elements en ∈ E (1) by an h-partition P (πn )n∈N . Assign Se := o- n∈N h(πn )Sen . It is easy to verify that the definition of S is sound at each of the steps. Obvious positivity of S ensures its regularity. In order Pmto prove
uniqueness of S, it is sufficient to observe that, at step 3, the sequence n=1 πn en m∈N is r-convergent to e with P regulator o- n∈N nπn |en | ∈ E (1, h). B

208

Chapter 5

5.3.3. The operator S, whose existence is asserted in 5.3.2, is called the shift by h and denoted by Sh . Let E be an order-dense ideal of E and F be an order-dense ideal of F . We say that an operator S : E → F is a shift operator, if there exists a ring homomorphism h : P(E ) → P(F ) such that E ⊂ E (1, h) and S = Sh on E. It is clear that, in this case, the homomorphism h is the shadow of S. Observe that the notion of shift and that of a shift operator depend on the choice of unities 1E and 1F in the K-spaces E and F . Theorem. A linear operator S : E → F is a shift operator if and only if it satisfies the following conditions: (1) S is disjointness preserving; (2) S is regular; (3) S takes fragments of 1E into fragments of 1F ; (4) S is wide at 1E . C Necessity of (1)–(3) is obvious and necessity of (4) follows from 5.3.1 (3). Let us show sufficiency. Suppose that the operator S satisfies conditions (1)–(4), denote the shadow of S by h and assign  … := {π ∈ P(E ) : π1E ∈ E}. Proposition 5.3.1 (3) implies the equality S(π1E ) = h(π) for each π ∈ …, which, together with condition (3), yields S(π1E ) = Sh (π1E ). The same proposition ensures the inclusion E ⊂ E (1, h). In view of 5.2.2 (3) and 5.3.1 (1), we now conclude that S = Sh on E. B 5.3.4. Say that a linear operator T : E → F admits a weight-shift-weight factorization if there exist order-dense ideals E 0 ⊂ E and F 0 ⊂ F , orthomorphisms w : E → E 0 and W : F 0 → F , and a shift operator S : E 0 → F 0 such that T = W ◦ S ◦ w. The composite W ◦ S ◦ w is called a WSW -representation of T , and the operators W , S, and w are respectively called the outer weight, shift, and inner weight of the representation W ◦ S ◦ w. An operator can have different WSW -factorizations. However, the constituents of these factorizations for a given operator T are connected with one another. Two main aspects of this connection are presented in the following proposition. Let T : E → F admit a WSW -factorization T = W ◦S ◦w. Assign ρ := [im(T )]. (1) Denote the shift of T by ST . Then ST extends ρ◦S and the equality W ◦ S ◦ w = W ◦ ST ◦ w holds. C The claim follows readily from 5.3.2 and 3.3.3. B (2) Identify w and W with the corresponding elements of E and F and assign WT := o-limπ∈… T π(1E /w) ∈ F , where … = {π ∈ P(E ) : π(1E /w) ∈ E}. Then ρW = WT and W ◦ S ◦ w = WT ◦ S ◦ w.

Disjointness Preserving Operators

209

C By the obvious equality T ◦ [w]⊥ = 0, we do not restrict generality on assuming that [w] = [1]. Then
o- lim T π(1E /w) = o- lim W ST wπ(1E /w) = o- lim W ST π1E = sup h(π) W, π∈…

π∈…

π∈…

π∈…

where h is the shadow of T . Since ρ = h(1), it is sufficient to show the relation supπ∈… h(π) = h(1). From E ⊂ dom(ST ◦ w) it follows that w(E) ⊂ dom(ST ) = E (1, h) and, hence, E ⊂ E (1/w, h). It remains to employ Proposition 5.3.1 (1). B (3) The inner weight in a WSW -representation can be chosen positive. If W ◦ S ◦ w is a WSW -representation of an operator T with w positive, then the operators T + , T − , and |T | admit the following WSW -representations: T + = W + ◦ S ◦ w, T − = W − ◦ S ◦ w, and |T | = |W | ◦ S ◦ w. C Consider an arbitrary WSW -factorization W ◦ S ◦ w. Identifying the ortho+ morphism w  with an  element of E , denote the projection [w ] ∈ P(E) by π and assign ρ := S(π1E ) . Then
W ◦ S ◦ w = W ◦ S ◦ π|w| − π ⊥ |w|
= W ◦ ρ ◦ S ◦ |w| − ρ⊥ ◦ S ◦ |w| = (ρW − ρ⊥ W ) ◦ S ◦ |w|. The remaining is obvious. B 5.3.5. Theorem. Let w be an arbitrary positive element of E . A linear operator T : E → F admits a WSW -factorization with inner weight w if and only if T is disjointness preserving, regular, and wide at the element 1E /w. C Necessity ensues from 5.3.4 (2). Let us prove sufficiency. Suppose that a disjointness preserving operator T : E → F is wide at 1E /w. Without loss  of generality, we may assume that the operator T is positive. Assign … := π ∈ Pr(E ) : π(1E /w) ∈ E and denote by W the orthomorphism of multiplication by supπ∈… T π(1E /w) ∈ F . Consider the composite (1F /W ) ◦ T ◦ (1E /w) as an operator from w[E] into F and denote it by S. In accordance with Theorem 5.3.3, it is sufficient to show that the operator S satisfies conditions (1)–(4) presented in the statement of the theorem. Verification of the conditions causes no difficulties, so that S is a shift operator and we obtain the desired WSW -factorization W ◦S ◦w for T . B Operators S and T are called strongly disjoint if Su ⊥ T v for all u, v ∈ X. Let X and Y be LNSs and let (Tξ )ξ∈„ be a family of linear operators from X into Y . Say that an operator T :
X → Y decomposes into the strongly disjoint sum L of Tξ and write T = the operators Tξ are strongly disjoint ξ∈„ Tξ , whenever P and, for every x ∈ X, the relation T x = o- ξ∈„ Tξ x holds. It is easy to observe that the strongly disjoint sum of (Tξ ) is disjointness preserving if and only if so is each summand Tξ .

210

Chapter 5

5.3.6. Theorem. Let T : E → F be a disjointness preserving regular operator. Then there exists a partition of unity (ρξ )ξ∈„ in the algebra P(F ) such that, for each ξ ∈ „, the composite ρξ ◦ T admits a WSW -factorization with inner weight 1E /eξ , where eξ is a positive element of E. In this case, the operator T decomposes into the strongly disjoint sum T =

M

W ◦ ρξ S ◦ (1E /eξ ),

ξ∈„

wherePS is the shift of T and W : F → F is the orthomorphism of multiplication by o- ξ∈„ ρξ T eξ . C By applying the Exhaustion Principle (see 1.1.6) to the relation sup{[T e] : e ∈ E+ } = [im(T )], we obtain a disjoint family (ρξ )ξ∈„ in the algebra P(F ) and a family (eξ )ξ∈„ of positive elements in E such that supξ∈„ ρξ [T eξ ] = [im(T )]. After adding the band projection onto [im(T )]⊥ to the family (ρξ )ξ∈„ and the zero element to the family (eξ )ξ∈„ , we make (ρξ )ξ∈„ into a partition of unity while supξ∈„ ρξ [T eξ ] = [im(T )]. By Theorem 5.3.5, for each ξ ∈ „, the composite ρξ ◦T admits a WSW -factorization with inner weight 1E /eξ . If S is the shift of T then the shift of ρξ ◦ T is equal to ρξ S (see 3.3.3); thus, using Proposition 5.3.4 (2), we conclude that ρξ ◦ T = ρξ T eξ ◦ ρξ S ◦ (1E /eξ ). B 5.3.7. To obtain WSW -factorization representations for operators in Banach– Kantorovich spaces which are analogous to those for operators in K-spaces we need some auxiliary facts. Let X be an LNS over E, let X0 be a vector subspace of X, and let Y be an LNS over F . Consider a linear operator T0 : X0 → Y and a disjointness preserving positive operator S : E → F with the shadow h : P(E) → P(F ). Denote by hX0 the LNS of all elements of X that are h-approximated by X0 . Assume that T0 u0 6 S u0 ( T0 u0 = S u0 ) for all u0 ∈ X0 . Then there exists a unique linear extension
T : hX0 → Y of the operator T0 such that T u 6 S u respectively, T u = S u for all u ∈ hX0 . C First, we prove the assertion about extension with the inequality preserved. If π ∈ P(X) and u0 ∈ X0 are such that πu0 = 0, then h(π)T0 u0 = 0, since h(π) T0 u0 6 h(π)S u0 = Sπ u0 = 0. This fact implies that the following definition of an operator T 0 : d0 (X0 ) → Y is sound: T¯0

X n k=1

 πk uk :=

n X k=1

h(πk )T0 uk

Disjointness Preserving Operators

211

where uh ∈ X and πk ∈ P(X) are pairwise disjoint. Evidently, T¯0 and satisfies the inequality T¯0 u 6 S u for all u ∈ d0 (X0 ). In view of Proposition 2.2.10 (3), for every u ∈ hX0 , there exists a net (uα )α∈A in d0 (X0 ) that is h-convergent to u. From the inequality T¯0 uα − T¯0 uβ 6 S uα − uβ and h-o-continuity of S (see 5.2.2 (3)) it follows that the net (T¯0 uα )α∈A is o-fundamental. Since the LNS Y is bo-complete, it contains a bo-limit of the net. Obviously, the limit depends only on u and, therefore, can be denoted by T u. It is not difficult to become convinced that the operator T : hX0 → Y thus obtained is the desired one. Uniqueness of this extension is ensured by its h-o-continuity inherited from S. Assume now that T0 u0 = S u0 for all u0 ∈ X0 . In view of what was proven above, there exists an extension T : hX0 → Y of the operator T0 such that T u 6 S u for all u ∈ hX0 . For every u0 ∈ X0 and π ∈ Pr(X), the relations S u0 = T u0 = T πu0 + T π ⊥ u0 6 S πu0 + S π ⊥ u0 = S u0 and the inequalities T πu0 6 S πu0 and T π ⊥ u0 6 S π ⊥ u0 imply T πu0 = S πu0 . Since u0 ∈ X0 and π ∈ P(X) were chosen arbitrarily, we have T u = S u for all u ∈ d0 X0 . The equality T u = S u for all u ∈ hX0 is now deduced from the fact that d0 X0 is h-dense in X. B 5.3.8. An operator S : X → Y is called a shift operator if there exists a shift operator s : E → F such that Su = s u for all u ∈ X. Obviously, s = S , i.e., the operator s is the exact dominant of S. (1) An operator S : X → Y is a shift operator if and only if there exist a shift operator s : E → F and an F -isometric embedding ı : sX → Y such that S = ı ◦ sX , where sX : X → sX is the norm transformation of X by means of s (see 2.2.13 (2)). C Only necessity requires proving. An elementary verification shows that the formula X  X n n
ı ρk sX xk = ρk Sxk xk ∈ X, ρk ∈ P(Y ) k=1

k=1

soundly defines a function ı : sX → Y that is the desired isometry. B The following description of shift operators generalizes criterion 5.3.3 to LNSs. (2) Theorem. An operator S : X → Y is a shift operator if and only if S satisfies the following: (a) S is disjointness preserving; (b) S is bounded; (c) if u ∈ X and u is a fragment of 1E then Su is a fragment of 1F ; (d) S is wide at 1E .

212

Chapter 5

C The proof is omitted, since the claim is not used in the sequel. More details can be found in [121, 123]. B 5.3.9. Let X be a BKS over an order-dense ideal E ⊂ E and let Y be a BKS over an order-dense ideal F ⊂ F . We say that a linear operator T : X → Y admits a weight-shift-weight factorization if there exist a BKS X 0 over an order-dense ideal E 0 ⊂ E , a BKS Y 0 over an order-dense ideal F 0 ⊂ F , orthomorphisms w : X → X 0 and W : Y 0 → Y , and a shift operator S : X 0 → Y 0 such that T = W ◦ S ◦ w, i.e., the diagram T

X −−−−→   wy

Y x  W

S

X 0 −−−−→ Y 0 is commutative. As in the case of an operator in K-spaces, the composite W ◦ S ◦ w is called a WSW -representation of T and the operators W , S and w are respectively called the outer weight, the shift, and the inner weight of the representation W ◦ S ◦ w. In order to avoid confusion, we call a weighted shift representation scalar or vector, referring to Definitions 5.3.3 or 5.3.9, respectively. By analogous reasons, we speak about scalar or vector WSW -factorizations. A vector WSW -factorization W ◦ S ◦ w of an operator T : X → Y will be called semivector if w is a scalar orthomorphism, i.e., X and X 0 are order-dense ideals of the same BKS over E and the orthomorphism w acts by the rule u 7→ eu for some fixed orthomorphisms e ∈ Orth(E, E 0 ). ¯ ◦ S¯ ◦ w (1) If W ¯ is a vector WSW -representation of T then T admits ¯ . ¯ such that 0 6 W 6 W a scalar WSW -factorization W ◦ S¯ ◦ w ¯ ◦ S¯ ◦ w ¯ . According to 3.3.3 there exists π ∈ Orth(F ) C Clearly, T 6 W ¯ ◦ S¯ ◦ w ¯ .B ¯ , so that we may put W := π W with T =π W (2) Theorem. Assume that T : X → Y is a dominated operator and its exact dominant T : E → F admits a scalar WSW -factorization W ◦ S ◦ w with positive weights W and w. Then T admits a semivector WSW -factorization ¯ = W , S¯ = S, and w ¯ ◦S◦ ¯ w ¯ such that W ¯ is the orthomorphism of multiplication W by w. C Suppose that W ◦ S ◦ w is a scalar WSW -factorization of T , where w : E → E 0 , S : E 0 → F 0 and W : F 0 → F . Let mX be the universal completion ¯ : of X, let X 0 be the ideal {u ∈ mX : u ∈ E 0 } of the BKS mX, and let w 0 0 X → X be the orthomorphism of multiplication by w. Denote by Y the bocompletion of the norm transformation of X 0 by means of S (see 2.2.13 (2)) and consider the corresponding operator of norm transformation S¯ : X 0 → Y 0 . Now, we ¯ :Y0 →Y. are to construct an orthomorphism W

Disjointness Preserving Operators

213

¯ 0 : Y 0 → Y by letting ¯ and define a linear operator W Assign Y00 := (S¯ ◦ w)(X) 0 ¯ 0 (S¯wu) ¯ := T u. Such a definition is sound, since the equality S¯wu ¯ 1 = S¯wu ¯ 2 W implies T u1 − T u2 6 T u1 − u1 = W Sw u1 − u1 ¯ 1 − wu ¯ 2 = W S¯wu ¯ 1 − S¯wu ¯ 2 = 0. = W S wu    ¯ ¯ Denote ρ := [im(T )]. Since ρ 6 (S¯ ◦ w)(X) and w(X) = v 0 ∈ Y 0 : v 0 ∈ w(E) , the operator ρ◦ S¯ is wide on the ideal w(E) ⊂ E 0 . It follows from the definitions that then the shadow of the restriction of ρ◦S¯ onto
the set w(X) ¯ coincide with the shadow 0 0 ¯ ¯ ¯ ¯ approximates (ρ◦ S)(X ). The latter of ρ◦ S. Therefore, the set Y0 = (ρ◦ S) w(X) set, by the definition of the norm transformation 2.2.13 (2), approximates the set ρ(SX 0 ), which in turn approximates ρ(Y 0 ). Thus, by virtue of 2.2.10 (2), the set ¯ 0 v 0 6 W v 0 for all v 0 ∈ Y 0 . According to Y00 approximates ρ(Y 0 ). Obviously, W 0 0 0 0 ¯ 0 admits a (unique) linear extension W ¯ 1 : ρ[Y 0 ] → Y such 5.3.7, the operator W ¯ 1 ◦ ρ : Y 0 → Y satisfies ¯ 1 v 0 6 W v 0 for all v 0 ∈ Y 0 . Then the composite W that W ¯ 1 ◦ ρ 6 W and, consequently, it is an orthomorphism. Thus, we the inequality W ¯ 1 ◦ ρ) ◦ S¯ ◦ w ¯ of the operator T . However, have constructed a WSW -factorization (W ¯ 1 ◦ ρ = W is not be guaranteed and the operator W ¯ 1 ◦ ρ should be the equality W rectified. For all positive e ∈ E, we have ¯ 1 ◦ ρ Swe = sup W

¯ 1 ρv 0 : v 0 ∈ Y 0 , v 0 = Swe W  ¯ 0 v 0 : v 0 ∈ Y 0 , v 0 = Swe > sup ρ W 0 0 0 0  ¯ 0 S¯wu = sup ρ W ¯ : u ∈ X, S¯wu ¯ = Swe  = sup T u : S w u = Swe  > sup T u : u = e = T e = W Swe, 

¯ 1 ◦ ρ Swe = W Swe by the inequality W ¯ 1 ◦ ρ 6 W . Thus, W ◦ S ◦ w and whence W ¯ W1 ◦ ρ ◦ S ◦ w are two WSW -factorizations of the operator T . Hence, according ¯ 1 ◦ ρ = ρW holds. To ensure the equality to Proposition 5.3.4 (2), the equality W ¯ ¯ ¯ ¯ ¯ 2 : X → Y is an arbitrary W = W , we define W := W1 ◦ ρ + W2 ◦ ρ⊥ , where W ¯2 = W. B dominated operator with W (3) A linear operator T : X → Y is a vector weighted shift operator if and only if it is dominated and its exact dominant T : E → F is a scalar weighted shift operator. C Follows from (1) and (2). B

214

Chapter 5

5.3.10. Theorem. Suppose that X and Y are Banach–Kantorovich spaces over order-dense ideals E ⊂ E and F ⊂ F , and T : X → Y is a disjointness preserving bounded operator. Then there exists a partition of unity (ρξ )ξ∈„ in the algebra P(Y ) such that, for each ξ ∈ „, the composition ρξ ◦ T admits WSW factorization with inner weight of norm 1E /eξ , where eξ is a positive element of E. For each ξ ∈ „, assign Eξ := {e/eξ : e ∈ E} and Xξ := {u ∈ mX : u ∈ Eξ }, and denote by wξ : X → Xξ the scalar orthomorphism of multiplication by 1E /eξ . Then there exist a BKS Y 0 over F , strongly disjoint shift operators Sξ : Xξ → Y 0 (ξ ∈ „), and an orthomorphism W : Y 0 → mY such that the operators T and T decompose into the following strongly disjoint sums: T =

M ξ∈„

W ◦ Sξ ◦ wξ ,

T =

M

W ◦ Sξ ◦ wξ .

ξ∈„

C Consider an arbitrary disjointness preserving bounded operator T : X → Y . By Theorem 5.3.6, there exists a partition of unity (ρξ )ξ∈„ in the algebra P(F ) such that, for each ξ ∈ „, the composition ρξ ◦ T is a weighted shift operator and, moreover, admits a WSW -factorization with inner weight 1E /eξ , where eξ is a positive element of E. Define BKSs Xξ and orthomorphisms wξ : X → Xξ in the same way as in the statement of the theorem under proof. By Theorem 5.3.7 (2), for each ξ ∈ „, there exist a BKS Yξ over an order-dense ideal Fξ ⊂ ρξ [F ], a shift operator Sξ : Xξ → Yξ , and an orthomorphism Wξ : Yξ → ρξ (Y ) such that ρξ ◦ T = Wξ ◦ Sξ ◦ wξ and ρξ ◦ T = Wξ ◦ Sξ ◦ wξ . In order to complete the proof, it remains to construct the desired BKS Y 0 and “glue” the orthomorphisms Wξ together to obtainL a single orthomorphism W . 0 Assign Y0 := ξ∈„ Yξ and denote by Y 0 a universal completion of the BKS Y00 . 0 Naturally identifying Yξ and ρξ [Y00 ], we regard Sξ as an operator P from Xξ into Y . 0 0 0 For each element v0 = (vξ )ξ∈„ ∈ Y0 , assign W0 (v ) := o- ξ∈„ Wξ (vξ ) ∈ mV . Due to 5.3.7, the orthomorphism W0 : Y00 → mY admits a unique extension to an orthomorphism W : Y 0 → mY . B 5.4. Multiplicative Representation In this section we give some analytical representations of disjointness preserving operators constructed with the help of such operations as a continuous change of variable and the pointwise multiplication by a real-valued function. 5.4.1. Throughout the section, P and Q are extremal compact spaces, E and F are order-dense ideals in the universally complete vector lattices E := C∞ (P ) and F := C∞ (Q), respectively. The symbol 1M denotes the function on a set M which is identically equal to unity. Denote by C0 (Q, P ) the totality of all continuous functions s : Q0 → P defined on various clopen subsets Q0 ⊂ Q.

Disjointness Preserving Operators

215

(1) A mapping h : Clop(P ) → Clop(Q) is a ring homomorphism if and only if there exists a function s ∈ C0 (Q, P ) such that h(U ) = s−1 (U ) for all U ∈ Clop(P ). For every ring homomorphism h, such a function s is unique. C The claim follows directly from the Sikorski Theorem 1.1.5 (1). B The relation h(U ) = s−1 (U ) is called the representation of the ring homomorphism h by means of the function s. (2) Let E and E 0 be order-dense ideals of C∞ (Q). A mapping W : 0 E → E is an orthomorphism if and only if there exists a function w ∈ C∞ (Q) such that W (e) = we for all e ∈ E. For every orthomorphism W , such a function w is unique. C See 3.3.2 (5). B The relation W (e) = we is called the representation of W by means of w. 5.4.2. Given arbitrary s ∈ C0 (Q, P ) and e ∈ C∞ (P ), the function e•s : Q → R is defined as follows:
 e s(q) , if q ∈ dom(s), (e • s)(q) := 0, if q ∈ Q\ dom(s). Obviously, the function e•s is continuous but, in general, does not belong to C∞ (Q), since it can assume infinite values on a set with nonempty interior. The totality of all functions e ∈ C∞ (P ) for which e • s ∈ C∞ (Q) is denoted by Cs (P ).

(1) Let h : P C∞ (P ) → P C∞ (Q) be a ring homomorphism, and let Ch (P ) := E (1P , h) be the order-dense ideal of E := C∞ (P ) defined in 5.3.1. Then Ch (P ) = Cs (P ), where s ∈ C0 (Q, P ) represents h by means of a formula h(U ) = s−1 (U ). C The claim follows from the definition of E (1, h) from 5.3.1 and 5.4.1 (1). B (2) A mapping S : E → F is a shift operator if and only if there exists a function s ∈ C0 (Q, P ) such that Se = e • s for all e ∈ E. C Sufficiency can be easily established with the help of Theorem 5.3.3. Let us show necessity. Suppose that S : E → F is a shift operator and h : P(E) → P(F ) is its shadow which is identified with the corresponding homomorphism from Clop(P ) to Clop(Q). Consider the representation h(U ) = s−1 (U ) of h by means of a mapping s ∈ C0 (Q, P ). According to Proposition (1), the equality Ch (P ) = Cs (P ) holds. Since the operators e 7→ e • s acts from Cs (P ) to C∞ (Q) and Sh : Ch (P ) → C∞ (Q) have the same shadow h and satisfy the equalities 1P • s = Sh (1P ) = h(1)1Q , they coincide in view of Proposition 5.3.2. Therefore, Se = Sh e = e • s for all e ∈ E. B 5.4.3. The function s in the representation 5.4.2 (2) of the shift operator S is not unique in general. Indeed, assume that the compact space P contains two different nonisolated points p1 and p2 , assign E := {e ∈ C∞ (P ) : e(p1 ) = e(p2 ) = 0} and

216

Chapter 5

consider the functions s1 , s2 : Q → P identically equal to p1 and p2 , respectively. Then e • s1 = e • s2 = 0 for all e ∈ E. The following propositions clarifies the question about uniqueness S of a representation of a shift operator S : E → F . Assign Q0 := supp(im S) := cl e∈E supp(Se), where supp(f ) := cl{q ∈ Q : f (q) 6= 0}. (1) If functions s1 , s2 ∈ C0 (Q, P ) satisfy the equalities Se = e • s1 = e • s2 for all e ∈ E then Q0 ⊂ dom(s1 ) ∩ dom(s2 ) and s1 = s2 on Q0 . C Denote by D the set of all points in P , at which some functions in E are nonzero. Obviously, the set s−1 1 (D) is dense in Q0 ; therefore, it is sufficient to establish the equality s1 = s2 on this set. Take an arbitrary point q ∈ s−1 1 (D) and assume to the contrary that s1 (q) 6= s2 (q).
Since s1 (q) ∈ D,
there exists a function e ∈ E that satisfies the relations e s1 (q) 6= 0 and e s2 (q) = 0, which contradicts the equality e • s1 = e • s2 . B (2) There exists a unique function s ∈ C(Q0 , P ) such that Se = e • s for all e ∈ E. Furthermore, if s is such a function then h(U ) = s−1 (U ) is a representation of the shadow h of the operator S. C Such a function s exists according to Proposition 5.4.2 (2), and its uniqueness follows from (1). The fact that s represents the shadow of S ensues from the proof of Proposition 5.4.2 (2). B If a function s satisfies the conditions stated in assertion (2) then the relation Se = e • s is called the representation of the shift operator S by means of the function s. 5.4.4. (1) Theorem. A mapping T : E → F admits a WSW -representation if and only if there exist functions s ∈ C0 (Q, P ), w ∈ C∞ (P ), and W ∈ C∞ (Q) such that we • s ∈ C∞ (Q) and T e = W (we • s) for all e ∈ E. C The claim readily follows from Propositions 5.4.1 (2) and 5.4.2 (2). B Simple examples show that the constituents of the representation T e = W (we• s) of a weighted shift operator T are not unique. However, omitting certain details, we may say that the function s is unique and W is uniquely determined by the choice of w. Let T : E → F be a disjointness preserving regular operator. Assign Q0 := supp(im(T )). (2) Let functions s1 , s2 ∈ C0 (Q, P ), w1 , w2 ∈ C∞ (P ) and W1 , W2 ∈ C∞ (Q) be such that T e = W1 (w1 e • s1 ) = W2 (w2 e • s2 ) for all e ∈ E. Then Q0 ⊂ dom(s1 ) ∩ dom(s2 ) and s1 = s2 on Q0 . If, in addition, w1 = w2 then W1 = W2 on Q0 . C Follows immediately from Proposition 5.3.4 due to 5.4.1 (2) and 5.4.2 (2). B (3) Let a positive function w ∈ C∞ (P ) be such that the operator T is wide at 1/w (see 5.3.1). Then there exist unique functions s ∈ C(Q0 , P ) and W ∈ C∞ (Q) such that W = 0 outside Q0 and T e = W (we • s) for all e ∈ E.

Disjointness Preserving Operators

217

Furthermore, supp(W ) = s−1 (supp(w)) = Q0 , Se = e • s is a representation of the shift S of the operator T , and h(U ) = s−1 (U ) is a representation of its shadow h. C Such functions s and W exist by virtue of Theorem 5.3.5 and (1); their uniqueness follows from (1). Connection of the function s with the shift and shadow of the operator T follows from Propositions 5.3.4 (1) and 5.4.3 (2). B If s, w, and W satisfy the conditions stated in assertion (2), then the relation T e = W (we • s) is called the representation of the weighted shift operator T by means of the functions s, w, and W . (4) If T e = W (we • s) is a representation of T then the operators + − T , T , and |T | admit the following representations: T + e = W + (we • s), T − e = W − (we • s), and |T |e = |W |(we • s). 5.4.5. Theorem. Let E be an order-dense ideal of C∞ (P ), let F be an orderdense ideal of C∞ (Q), and let T : E → F be a disjointness preserving regular operator. Consider the representation h(U ) = s−1 (U ) of the shadow h of the operator T by means of a function s ∈ C0 (Q, P ). Then there exist a family (wξ )ξ∈„ of positive functions in C∞ (P ) and a family (Wξ )ξ∈„ of pairwise disjoint functions in C∞ (Q) such that 1/wξ ∈ E for all ξ ∈ „ and X T e = oWξ (wξ e • s)

(e ∈ E).

ξ∈„

C The assertion stated is a reformulation of Theorem 5.3.6 with account taken of Proposition 5.4.4 (3). B Observe that the functions wξ e•s in the above representation, being continuous functions from Q into R, need not belong to C∞ (Q), while the products Wξ (wξ e•s) do belong to CP ∞ (Q). If T e = o- ξ∈„ Wξ (wξ e • s) is a representation of the operator T then the operators T + , T − , and |T | admit the following representations: X T + e = oWξ+ (wξ e • s), ξ∈„

X T e = oWξ− (wξ e • s), −

ξ∈„

X |T |e = o|Wξ |(wξ e • s). ξ∈„

5.4.6. Now we proceed to constructing analytical representations for operators in Banach–Kantorovich spaces which are analogous to above results for operators in K-spaces. If X and Y are ample Banach bundles over Q, u ∈ C∞ (Q, X )

218

Chapter 5


and w ∈ C∞ Q, B(X , Y ) then the section w ⊗ u : q ∈ Q 7→ w(q)u(q) ∈ Yq is continuous and defines a unique element of C∞ (Q, Y ), see 2.4.6 and 2.4.13. This ¯ As was indicated in 2.4.6 w⊗u ¯ can be identified element will be denoted by w⊗u. with the maximal extension of w ⊗ u. Let X and Y be ample Banach bundles over Q, let E and E 0 be order-dense ideals in C∞ (Q). An operator W : E(X ) → E 0 (Y
) is an orthomorphism if and only if there exists a section w ∈ C∞ Q, B(X , Y ) such that the representation ¯ holds for all u ∈ E(X ). Moreover, W (e) = w · e (e ∈ E) and the W u = w⊗u function w with these properties is unique. C Let W : E(X ) → F (Y ) be such that W ∈ Orth(E, F ). Then the representation W (e) = ge (e ∈ E) holds with a suitable function g ∈ C∞ (Q). Denote by D the set of all q ∈ Q such that |g(q)| < ∞ and e(q) 6= 0 for some e ∈
E. Put E1 := E ∩ C(Q). Define an mapping w0 : q ∈ D 7→ w(q) ∈ B X (q), Y (q) by the following rule: for arbitrary q ∈ D and x ∈ X (q) take a section u ∈ E1 (X ) with u(q) = x, and put w0 (q)x := (W u)(q). This definition is sound and the operator w0 (q) : X (q) → Y (q) is bounded, since the identities

k(W u)(q)k = W u (q) 6 W u (q) = g u (q) = g(q)ku(q)k are valid for all q ∈
D and u ∈ E1 (X ). According to Theorem 2.4.13 we have
w0 ∈ C D, B(X , Y ) . If w := ext(w0 ) ∈ C∞ Q, B(X , Y ) . then it follows from ¯ for all u ∈ E1 (X ). The set E1 (X ) is a bo-dense the definitions that W u = w⊗u ¯ ideal in the Banach–Kantorovich space E(X ) and the operators W and u 7→ w⊗u ¯ coincide on E(X ). are bo-continuous. Thus, W and u 7→ w⊗u
¯ = Assume that the sections w1 , w2 ∈ C∞ Q, B(X , Y ) met the relation w1 ⊗u ¯ for all u ∈ E(X ). Denote by D0 the set of all q ∈ Q such that e(q) 6= 0 for w2 ⊗u some e ∈ E. Put D := D0 ∩ dom(w1 ) ∩ dom(w2 ) and take q ∈ D and x ∈ X (q). Since there exists a section u ∈ E(X ) with u(q) = x, we deduce ¯ ¯ w1 (q)x = (w1 ⊗u)(q) = (w2 ⊗u)(q) = w2 (q)x. It follows that w1 = w2 , because D is dense in Q. To prove the identity W (e) = w e observe that, by virtue of 2.4.2 (5), we ¯ : u ∈ C(Q, X ), u 6 1}. Therefore, for all e ∈ E+ we may have w = sup{ w⊗u deduce: W e = sup W u = sup W (eu) u 6e

u 61

¯ ¯ e = w e. B = sup w⊗(eu) = sup w⊗u u 61

u 61

¯ we will call the representation of W by means of the The relation W u = w⊗u section w.

Disjointness Preserving Operators

219

5.4.7. (1) Let X and Y be continuous Banach bundles over Q with Y ample. Assume that X and Y are decomposable subspaces of lattice-normed spaces E(X ) and E(Y ), respectively. If the set X is an approximating set in E(X ) then a mapping I : X → Y is an isometric embedding of lattice-normed space if and only if there exists an isometric imbedding ı of the bundle X into Y such that ¯ for all u ∈ X. I(u) = ı⊗u C Sufficiency is obvious. To prove necessity suppose that the operator I : X → Y is an isometric embedding. In view of 5.3.7 there exists an isometric embedding I¯ : C∞ (Q, X ) → C∞ (Q, Y ) extending I. Let X be an ample hull of X . Identify C∞ (Q, X ) and C∞ (Q, X ) according to 2.4.11 (3). By virtue of 5.4.6 we have ¯ the representation I(u)
= ı⊗u of the orthomorphism I by means of some section ı ∈ C∞ Q, B(X , Y ) . It is easy to check that ı is an isometric imbedding of X into Y . For q ∈ Q denote the restriction of ı(q) onto X (q) by the same symbol ı(q). By Definition 2.4.9 ı ∈ HomQ (X , Y ), and so ı is an isometric embedding of ¯ for all u ∈ X. B X into Y . Obviously I(u) = ı⊗u (2) Let X and Y be ample Banach bundles over Q and let E be an order-dense ideal in C∞ (Q). A linear operator I : E(X ) → E(Y ) is an isometric embedding (isometry) if and only if there exist an isometric embedding (isometry) ¯ for all u ∈ E(X ). ı of X into (onto) Y such that I(u) = ı⊗u 5.4.8. Let X be a Banach bundle over P and s be an arbitrary mapping from a subset dom(s) ⊂ Q to P . Then the composite X ◦ s is a Banach bundle over dom(s). Moreover, if u is a section of X over D ⊂ P then u ◦ s is a section of X ◦ s over s−1 (D). For an arbitrary set U of sections of X , denote by U ◦ s the set of sections {u ◦ s : u ∈ U } of a bundle X ◦ s. The bundle X ◦ s, extended by the zero stalks on Q\ dom(s), will be denoted by X • s. More precisely, the Banach bundle X • s over Q is defined by
 X s(q) , if q ∈ dom(s), (X • s)(q) := {0}, if q ∈ Q\ dom(s). If u is a section of X over D ⊂ P then we denote by u • s the section
 u s(q) , if q ∈ s−1 [D], (u • s)(q) := 0, if q ∈ Q\ dom(s) of the bundle X • s over s−1 (D) ∪ (Q\ dom(s)). Let CX be a continuity structure in the Banach bundle X . Then the set CX • s := {u • s : u ∈ CX } is a continuity structure in the Banach bundle X • s, so that X • s is considered as a continuous Banach bundle (2.4.3). (1) Let X be an ample Banach bundle over P , s ∈ C0 (Q, P ), and u ∈ C∞ (P, X ). If u ∈ Cs (P ) then u • s ∈ C∞ (Q, X • s).

220

Chapter 5

C Indeed, the domain of u • s coincides with dom( u • s) which is dense in Q due to u ∈ Cs (P ). At the same time, if the section u • s has a limit at q ∈ Q, then q ∈ dom u • s = dom( u • s) = dom(u • s). B (2) Let X be an ample Banach bundle over P , let S : E → F be a shift operator. Let U be the norm transformation of X by means of S and denote by S^ the operator of norm transformation (see 2.2.13 (2)). Let Se = e • s be the representation of S by means of some mapping s ∈ C0 (Q, P ). Then there exists ^ = u • s for all u ∈ E(X ). an isometric embedding ı : U → F (X • s) such that ıSu ^ ) → F (X • s) by putting ı0 Su ^ := u • s for all C Define an operator ı0 : S(U ^ (u ∈ U ), and u ∈ U . The definition is sound since u • s = u • s = S u = Su ^ ı0 (v) = v for all v ∈ S(U ). In view of 5.3.7 ı0 extends to a desired isometric embedding ı : U → F (X • s). B (3) Let X and Y be ample Banach bundles over P and Q, respectively. A linear operator S : E(X ) → F (Y ) is a shift operator if and only if there exist a mapping s ∈ C0 (Q, P ) and a isometric embedding ı of the bundle X • s into Y such that Su = ı ⊗ (u • s) for all u ∈ E(X ). Moreover, S e = e • s for all e ∈ E. C Sufficiency is easily verified by employing Theorem 5.3.8 (2). Necessity follows from (2), (3), and 5.4.7 (1). B 5.4.9. Theorem. Let X and Y be ample Banach bundles over P and Q, respectively. A linear operator T : E(X ) → F (Y ) admits WSW -representation if and only if there exist a positive function
w ∈ C∞ (P ), a mapping s ∈ C0 (Q, P ), and a section W ∈ C∞ Q, B(X • s, Y ) , where X • s is an ample hull of X • s, ¯ such that T u = W ⊗(wu • s) for all u ∈ E(X ) and T e = W (we•s) for all e ∈ E. C Sufficiency follows immediately from 5.4.6 and 5.4.8 (3). Let a linear operator T : E(X ) → F (Y ) admit WSW -factorization. Then by Theorem 5.3.9 (2) there exist a Banach–Kantorovich space V 0 over an order-dense ideal F 0 ⊂ C∞ (Q), a scalar orthomorphism w ¯ : E(X ) → C∞ (P, X ) generated by a positive orthomorphism w : E → C∞ (P ), a shift operator S¯ : (wE)(X ) → V 0 , and an ortho¯ : V 0 → F (Y ) such that T = W ¯ ◦ S¯ ◦ w ¯ ◦ S¯ ◦ w ¯. morphism W ¯ and T = W 0 0 Using Theorem 2.4.10, we may assume that V = F (Z ), where Z is an ample Banach bundle over Q. According to 5.4.8 (3) there are a mapping s ∈ C0 (Q, P ) ¯ = ı ⊗ (u • s) and an isometric imbedding ı of the bundle X • s into Z such that Su for all u ∈ (wE)(X ). By virtue of 2.4.11 (2) the homomorphism ı can be extended to an isometric embedding
¯ı of the bundle X • s into Z . In view of 5.4.6 the ¯ (¯ı ⊗ v) : F 0 (X • s) → F (Y ) has the representation v 7→ W ⊗v, ¯ operator v 7→ W
where W ∈ C∞ Q, B(X • s, Y ) . It is easy to verify that the functions w, s and W possess the desired properties. B 5.4.10. Theorem. Let X and Y be ample Banach bundles over P and Q, respectively. Let T : E(X ) → F (Y ) be a bounded disjointness preserving operator

Disjointness Preserving Operators

221

and let s ∈ C0 (Q, P ) be its shift function. Then there exist a family (wξ )ξ∈„ of positive functions in C∞ (P ), a pairwise
disjoint family (Qξ )ξ∈„ in S Clop(Q), and a section W ∈ C∞ Q, B(X • s, Y ) such that (supp(W )) = cl ξ∈„ Qξ = dom(s) = supp(im(T )), 1/wξ ∈ E for all ξ ∈ „ and M
Tu = W ⊗ (wξ u • s)|Qξ u ∈ E(X ) . ξ∈„

C The claim follows from Theorems 5.3.10 and 5.4.9. B 5.5. Decomposable Operators In this section we consider several results concerning an analytical description for the dual classes of disjointness preserving operators and decomposable operators acting in the spaces of continuous vector-functions. We give an independent representation though some results may be deduced from Section 5.3 or can be generalized to the case of Banach bundles. 5.5.1. Throughout this section X and Y are normed spaces with normed duals X and Y 0 ; furthermore, E and F are order-dense ideals in universally complete vector lattices E := C∞ (P ) and F := C∞ (Q), respectively. Recall (see 4.1.3 (3)) that LA (X, E) is the Banach–Kantorovich space of operators T : X → E with abstract norm T = sup{T x : kxk ≤ 1}. For the definitions of E(X), Es (X), and MQ (X, Y 0 ) see Section 2.3. (1) To each operator with abstract norm T : X → E there is a unique 0 uT ∈ Es (X ) satisfying T x = hx, uT i (x ∈ X). 0

The mapping T 7→ uT is a linear isometry between the Banach–Kantorovich spaces LA (X, E) and Es (X 0 ). C If e := T then, for every x ∈ X, the function T x ∈ C∞ (Q) takes a finite value at each point of Q0 := {t ∈ Q : e(t) < +∞} since |T x| ≤ ekxk. The last estimate also implies that, for t ∈ Q0 , the functional v(t) : x 7→ (T x)(t) (x ∈ X) is bounded and kv(t)k ≤ e(t). This gives rise to the mapping v : Q0 → X 0 continuous in the weak topology σ(X 0 , X). Let uT denote the coset of v. Then T x = hx, uT i for all x ∈ X. In particular, the following supremum exists: sup {hx, uT i| : kxk ≤ 1} = e. Hence, uT ∈ Es (X 0 ) and uT = T . We thus see that T 7→ uT is an isometry from LA (X, E) to Es (X 0 ). Clearly, this mapping is linear and surjective. B b E), where X ⊗Y b is the projective tensor product of X Consider T ∈ LA (X ⊗Y, and Y . It is an easy matter to show that the bilinear operator b := T ⊗ : X ×Y → E has the abstract norm b := sup{|b(x, y)| : kxk ≤ 1, kyk ≤ 1},

222

Chapter 5

with b = T . Denote by BA (X ×Y, E) the set of all bilinear operators b : X ×Y → E with abstract norm. We further let B(X × Y ) denote the set of all bilinear forms b )0 ' B(X × Y ) is available, from on X × Y . Since the isometric isomorphy (X ⊗Y (1) we derive the following proposition: (2) To every b ∈ BA (X × Y, E) there is a unique ub ∈ Es (B(X × Y )) such that b(x, y) = hx ⊗ y, ub i (x ∈ X, y ∈ Y ). The mapping b 7→ ub is a linear isometry between BA (X ×Y, E) and Es (B(X ×Y )). 5.5.2. Let ˆ : E → mF be a lattice homomorphism. Assume that F = ˆ(E)⊥⊥ is a K-space. In the universal completion mF , we fix the multiplicative structure  0 that determined by choice of unity. Assign F := f ∈ mF : f · ˆ(E) ⊂ is uniquely ˆ F . Let L (E, F ) be the set of all regular operators S : E → F such that S ∈ {ˆ}⊥⊥ , where S is regarded as an operator from E into mE. Assign

M ˆ E(X), Fs (Y 0 ) := T ∈ M E(X), Fs (Y 0 ) : T ∈ L ˆ (E, F ) . (1) Every disjointness preserving sequentially o-continuous regular operator T : E → F admits a WSW -representation with an arbitrary order-unity w ∈ C∞ (P ) as an inner weight. C If T is sequentially o-continuous then its shadow shdw(T ) is also sequentially o-continuous and, by Definition 5.1.3, E = E (1, h). According to 5.3.1 (3) T is wide at 1. Now, the claim follows from 5.3.5. B (2) Assume that ˆ : E → F be an order continuous lattice homomorphism. Then there exist a continuous open mapping ϕ from a closed subset of Q into P and a function α ∈ C∞ (Q) such that ˆe = α · e • ϕ (e ∈ E). C The representation ˆe = α · e • ϕ follows from (1) and 5.4.4 (1). If T is order continuous then shdw(T ) is also order continuous and then ϕ is an open mapping; see [352]. B (3) If the substitution operator ϕ∗ : e 7→ e • ϕ acts from E into F and is order continuous then for any Banach space X the mapping ϕ∗X : u 7→ u • ϕ is a unique dominated operator from E(X) to F (X) such that s∗X = s∗ and ϕ∗X (x ⊗ e) = x ⊗ ϕ∗ e (x ∈ X, e ∈ E). C The operator ϕ∗ is order continuous if and only if the mapping ϕ is open; moreover, the inverse image of every meager subset of Q under ϕ is a meager subset of P ; see [352]. Therefore, the mapping u 7→ u • ϕ acts from E(X) into C∞ (Q, X) and is uniquely defined on X ⊗ E by the formula X  X n n ∗ ϕX xk ⊗ ek := xk ⊗ ϕ∗ ek . k=1

k=1

Disjointness Preserving Operators

223

Straightforward verification shows that this operator is dominated. From Theorem 4.3.3 it is clear that ϕ∗X admits a unique extension onto E(X). The equality ϕ∗X = ϕ∗ follows from 4.3.5 (3). B 5.5.3. Theorem. Suppose that ˆ is an order continuous lattice
homomorphism from E into mF . For every operator T ∈ M ˆ E(X), Fs (Y 0 ) there exists a unique (to within equivalence) operator-function K ∈ MQ (X, Y 0 ) satisfying K ∈ F 0 and providing the representation

T u = K α · ϕ∗X (u) u ∈ E(X) . The correspondence T 7→ K defines a linear isometry

between the Banach–Kantoro0 ˆ 0 0 vich spaces M E(X), Fs (Y ) and Fs L (X, Y ) . C According to 5.5.2 (2), the operator ˆ admits the
multiplicative representa∗ ˆ 0 tion ˆe = αϕ e (e ∈ E). Take a T ∈ M E(X), F (Y ) and assign Se = Sx,y (e) := s

 y, T (x ⊗ e) , where x ∈ X, y ∈ Y , and e ∈ E. The element x ⊗ e ∈ E(X) is determined by the formula x ⊗ e : t 7→ e(t)x, |e(t)| < +∞. Observe that |Se| ≤ T (x ⊗ e) kyk ≤ T (|e|) · kxkkyk. Since T ∈ L ˆ , we also have S ∈ L ˆ , because L ˆ is a band. Denote by λ the isomorphism from L ˆ (E, F ) onto F 0 that was presented in 3.3.4 (2). Assign b(x, y) := λ(Sx,y ). From the definitions it is clear that the mapping (x, y) 7→ b(x, y) is a bilinear operator from X × Y into F 0 . Moreover,
|b(x, y)| = λ(|Sx,y |) ≤ λ T kxkkyk (x ∈ X, y ∈ Y ),
i.e., b is an operator with abstract norm and b ≤ λ T . By the definition of λ, we have

 y, T (x ⊗ e) = b(x, y)ˆe (x ∈ X, y ∈ Y, e ∈ E). Taking account of what was said, as well as of the formula 4.3.5 (3), we may write the chain nX o T e = sup T (x(k) ⊗ c(k) : (x, c) ∈ U n o X
= sup sup hy, T x(k) ⊗ c(k) i : (x, c) ∈ U kyk≤1

= sup

n

X

o
b x(k), y ˆ(c(k)) : (x, c) ∈ U

nX

b ˆ(c(k))kx(k)kkyk : (x, c) ∈ U

sup kyk≤1

≤ sup sup kyk≤1

 ≤ sup

X  bˆ c(k)

= b ˆ(e),

o

224

Chapter 5

where the set U is constituted by all pairs (x, c) in which x : {1, . . . , n} → X, c : {1, . . . , xnP } → E+ , and n ∈ N; moreover, kx(k)k ≤ 1 (k := 1, . . . , n), c(k)
⊥ c(l) n (k 6= l), and k=1 c(k) = e. From the estimates made it is clear that λ T ≤ b
. Taking account of the reverse inequality established above, we obtain b = λ T . Now apply Theorem 5.5.1 (2), according to which there exists an operator-function K ∈ MQ (X, Y 0 ) such that b = K and the following hold:




 y, T (x ⊗ e) = b(x, y)ˆ(e) = hy, Kxiαϕ∗ e = y, K αϕ∗X (x ⊗ e) . Since the choice of y ∈ Y is arbitrary,
T (x ⊗ e) = K αϕ∗X (x ⊗ e) (x ∈ X, e ∈ E). Using this fact, linearity of T and K, and the definition of ϕ∗X (see 5.5.2 (3)), we deduce that T u = K(αϕ∗X u) (u ∈ X ⊗ E). Since the operators T and ϕ∗X are bo-continuous and X ⊗ E is bo-dense in E(X), the above representation on X ⊗ E remains valid on the entire E(X). Uniqueness of K is clear from the following reasoning. If L ∈ MQ (X, Y 0 ) represents T too, then L(αϕ∗X u) = K(αϕ∗X u) for u ∈ E(X). In particular,

L(t)x − K(t)x α(t)e ϕ(t) = 0 (x ∈ X, e ∈ E). Since ˆ(E) is an order-dense ideal in C∞ (Q), we conclude that α is an order-unity in C∞ (Q) and dom(ϕ) is dense in Q; hence, L(t)x = K(t)x for all t ∈ Q with the possible exception of points in some meager set. It remains to observe that if K ∈ MQ (X, Y 0 ) and K ∈ F 0 then the operator Su = K(αϕ∗X u) belongs to M ˆ , because


Su ≤ K αϕ∗X (u) ≤ α K ϕ∗ u = K ˆ u u ∈ E(X) . B 5.5.4. Theorem. Let ˆ be an order continuous lattice homomorphism from E
into F and suppose that an operator T ∈ M E(X), Fs (Y, Z) satisfies the condition T ∈ {ˆ}⊥⊥ . Then there exist an operator-function K ∈ MQ (X, Y 0 ), a function α ∈ C∞ (Q), and a continuous mapping ϕ : dom(ϕ) → P , with dom(ϕ) a closed subset of Q, such that the following hold:
T u = K(αϕ∗X u) u ∈ E(X) , T e = K αϕ∗ e (e ∈ E), ˆe = αϕ∗ e

(e ∈ E).

C Identify Y with a closed subspace of Z 0 . Then Fs (Y,
Z) is a bo-closed subspace of Fs (Z 0 ). It is clear that T ∈ M ˆ E(X), Fs (Z 0 ) , because T ∈ {ˆ}⊥⊥ By Theorem 5.5.3, there exists a K ∈ MQ (X, Z 0 ) for which the required representations hold. B

Disjointness Preserving Operators

225

5.5.5. Consider LNSs (X, E) and (Y, F ). A linear operator T : X → Y is called decomposable if, for all x ∈ X and disjoint y1 , y2 ∈ Y , the equality T x = y
1 + y2 implies existence of x1 , x2 ∈ X such that x = x1 + x2 and T BX ( xk ) ⊥ yl (k, l := 1, 2, k 6= l). The last relation means that if u ≤ xk then T u ⊥ yl (k 6= l). In particular, T xk ⊥ yl (k 6= l); therefore, the coinciding elements T x1 − y1 and y2 − T x2 are disjoint and, hence, T xk = yk (k = 1, 2). A positive operator S : E → F is called positively decomposable if, for all e ∈ E+ and disjoint f1 , f2 ∈ F+ , the equality Se = f1 + f2 implies existence of e1 , e2 ∈ E+ such that e = e1 + e2 and Sek = fk
(k = 1, 2). Recall that the operator S is said to be interval preserving if S [0, e] = [0, Se] for each e ∈ E+ . An order continuous, interval preserving operator is called a Maharam operator. 5.5.6. Theorem. Let X and Y be decomposable and let F be order complete. Given a dominated operator T : X → Y , the following hold: (1) T is decomposable if and only if T is positively decomposable on the o-ideal generated by X ; (2) T is decomposable and bo-continuous if and only if T is positively decomposable and o-continuous; (3) if E is order complete then T is decomposable and bo-continuous if and only if T is a Maharam operator. C Suppose that the operator T is positively decomposable on the set X . Assume that the representation T x = y1 + y2 holds for x ∈ X and y1 , y2 ∈ Y , ⊥⊥ y1 ⊥ y2 . Take
the projection π1 onto the
band y1 and assign π2 := π1⊥ and fk := πk T x (k := 1, 2). Then T x = f1 + f2 , f1 ⊥ f2 , and, hence, there are e1 , e2 ∈ E+ such that e1 + e2 = x and fk = T ek (k := 1, 2). Decomposability of X provides the representation x =
x1 + x2 with xk = ek (k := 1, 2). If u ∈ X and u ≤ xk then T u ≤ T u ≤ fk ; therefore, T u ⊥ fl (k 6= l). Since yl ≤ fl , we have T u ⊥ yl (k 6= l). Thus, T is decomposable. Conversely, let T be decomposable. Assume that T e = f1 + f2 for some e := x , x ∈ X, and disjoint f1 , f2 ∈ F+ . Assign yk := πk T x, where π is the projection onto the band {f1 }⊥⊥ and π2 := π1⊥ . In view of decomposability
of T , there are x1 , x2 ∈ X such that x = x1 + x2 , T xk = yk , and T BX ( xk ) ⊥ yl (k 6= l). It follows that e ≤ c1 + c2 , T ck ⊥ fl (k 6= l), and ck = xk . If ek ≤ ck are such that e = e1 + e2 then f1 + f2 = T e1 + T e2 and T ek ⊥ fl (k 6= l); consequently, T ek = fk (k = 1, 2). Suppose now that e ≤ x1 + · · · + xn , where x1 , . . . , xn ∈ X. Then, due to decomposability of X, we have e = c1 + · · · + cn with c1 , . . . , cn ∈ X . In view of what was Pnproven above, there exist Pn e1k , e2k ∈ E+ (k := 1, . . . , n) such that ck = e1k + e2k , k=1 T e1k = f1 , and k=1 T e2k = f2 . It remains to assign e1 := e11 +· · ·+e1n and e2 := e21 +· · ·+e2n . Thus, assertion (1) is completely proven. In order to prove (2), observe that an o-continuous positive

226

Chapter 5

operator S : E → F is decomposable if S is such on an order-dense ideal E0 (in our ⊥⊥ case, E0 is the ideal generated by X , thus being an order-dense ideal in X , ⊥ and S := T vanishes on X ). Taking account of (2), in (3) it is sufficient to prove the following: an ocontinuous positive operator S : E → F preserves intervals if and only if S is positively decomposable. If S preserves intervals then, obviously, S is positively decomposable. Suppose that S is positively decomposable. Without loss of generality, we may assume that S is essentially positive. Assign e S = S(|e|) (e ∈ E). Then · S is a d-decomposable F -valued norm in E, see 3.4.3. According to 2.1.3, there exist a complete Boolean algebra B of projections in E and an isomorphism
⊥⊥ h : P(F0 ) → B, with F0 = E S , such that π e S = h(π)e S e ∈ E, π ∈ P(F0 ) . Moreover, B is an order-closed subalgebra of P(E). Using 2.1.8, we may extend the isomorphism h onto ƒ := Orth(F ), thereby endowing E with the structure of a unital ƒ-module. Hence it easily follows that S preserves intervals. B 5.5.7. Theorem. Let X and Y be decomposable and let E and F be order complete. Suppose that an operator T ∈ M(X, Y ) is decomposable ⊥ and bocontinuous. Assign YT := T (X)⊥⊥ and XT := x ∈ X : T (|x|) = 0 . There exists a Boolean isomorphism h from P(YT ) onto P(XT ) such that, for each bocontinuous S ∈ M (X, Y ), the following are equivalent: (1) S ∈ {T }⊥⊥ ;
⊥⊥ (2) Sx ∈ T BX ( x ) (x ∈ X);
(3) πS = Sh(π) π ∈ P(YT ) . C Without loss of generality, we may assume that E = X (see 2.1.7 (3)). By Theorem 5.5.6 (3), ˆ := T is a Maharam operator. The general properties of Maharam operators (see 3.4.3) ensure existence of a Boolean isomorphism h from P(Fˆ ) onto P(Eˆ ) such that πˆ = ˆh(π) for all π ∈ P(Fˆ ). Denote by the same letter h the isomorphism from P(YT ) onto an order-closed subalgebra of P(XT ); such an isomorphism exists, since the Boolean algebras P(Fˆ ) and P(YT ), as well as P(Eˆ ) and P(XT ), are pairwise isomorphic. For completing the proof, it is sufficient to show that the required properties (1), (2), and (3) are respectively equivalent to the following: (10 ) ‰ ∈ {ˆ}⊥⊥ ; (20 ) ‰e ∈ {ˆe}⊥⊥ (e ∈ E);
(30 ) π‰ = ‰h(π) π ∈ P(Fˆ ) , where ‰ := S . (1) ⇔ (10 ): It is evident.
⊥⊥ (2) ⇔ (20 ): Suppose Sx ∈ T BX ( x ) (x ∈ X). If x1 , . . . , xn ∈ X are
⊥⊥
⊥⊥ Pn such that x = x1 + · · · + xn , then k=1 T BX ( xk ) ⊂ T BX ( x ) . At

Disjointness Preserving Operators

227


⊥⊥ Pn Pn Pn the same time k=1 S(xk ) ∈ k=1 T BX ( xk ) ; therefore, k=1 S(xk ) ∈ ⊥⊥ ˆ( x ) . Thus, we deduce that ( ‰( x ) = sup

n X

) ∈ ˆ( x )⊥⊥ .

S(xk )

k=1

Conversely, let ‰e ∈ {ˆe}⊥⊥ for all e ∈ E. Then Sx ∈ {‰e}⊥⊥ ⊂ {ˆe}⊥⊥ and it remains to observe that ( {ˆe}⊥⊥ = sup

n X

T xk :

_n

Ty

⊥⊥

)⊥⊥ xk = x , n ∈ N

k=1

k=1

=

n X

: y ≤ x

o


⊥⊥ . = T BX ( x )

(3) ⇔ (30 ): If (3) is valid then the following equalities hold: ( π‰e = sup

n X

k=1 n X

π Sxk :

( = sup

k=1 n X

πSxk :

n X k=1 n X

= sup

k=1 ( n X k=1

xk = e, n ∈ N ) xk = e, n ∈ N

k=1

( = sup

)

Sh(π)xk :

n X

) xk = e, n ∈ N

k=1

Syk :

n X

) yk = h(π)e, n ∈ N

k=1

=‰h(π)e.
Conversely, assume that π‰ = ‰h(π) for all π ∈ P(Fˆ ) . Then Sh(π)x ≤ π‰( x ) and π ⊥ Sh(π)x = π ⊥ Sh(π)x = 0. Thus, π ⊥ Sh(π) = 0 or πSh(π) = Sh(π) = 0. Replacing π by π ⊥ in the last equality we obtain also πSh(π) − πS = 0, whence πS = πSh(π). B 5.5.8. Considering the universal completion mE, fix the multiplicative structure that is uniquely determined by a choice of an order-unity 1. Suppose that an F -valued essentially positive Maharam operator ˆ is defined on some orderdense ideal D(ˆ) ⊂ mE. Denote by L1 (ˆ) the maximal order-dense ideal of mE onto which we may extend ˆ by o-continuity. We assume that D(ˆ) = L1 (ˆ). Let ˆ0 be the restriction of ˆ onto E0 := E ∩ L1 (ˆ). Denote by Lˆ (E, F ) the set of

228

Chapter 5

all o-continuous regular operators S : E → F such that the restriction of S onto E0 belongs to theband {ˆ0 }⊥⊥ . It is easily seen that Lˆ (E, F ) is a band of Ln (E, F ). Assign E 0 := e0 ∈ mE : e0 · E ⊂ L1 (ˆ) .
Now, given Banach spaces X and Y , consider the LNS Es0 L (X, Y 0 ) defined as a factor set of the class MP (X, Y 0 ) of operator-functions, see 2.3.6. Take K ∈ MP (X, Y 0 ), u ∈ E(X), and y ∈ Y . If s, s0 ∈ dom(K) ∩ dom(u) then



 | y, K(s)u(s) − y, K(s0 )u(s0 ) |


 ≤ K (s)ku(s) − u(s0 )k + | y, K(s) − K(s0 ) u(s0 ) |.

 Hence it is clear that the function t 7→ y, K(t)u(t) is continuous on dom(K) ∩ dom(u); therefore, it admits an R-valued continuous extension onto the entire P . Denote by hy, Kui the corresponding element of C∞ (P ). From the inequality |hy, Kxi| ≤ K kxk · kyk it follows that |hy, Kui| ≤ K ·
u . If K ∈ E 0 and u ∈ E then K · u ∈ L1 (ˆ); thus, the element ˆ hy, Kui ∈ F is defined. Now we formulate and prove the main result of the present section.
5.5.9. Theorem. For every dominated operator T ∈ Mˆ E(X), Fs (Y 0 ) , there exists a unique (to within equivalence) operator-function K ∈ MP (X, Y 0 ) such that K ∈ E 0 and the following holds:

hy, T ui = ˆ hy, Kui u ∈ E(X), y ∈ Y . The correspondence T 7→ K establishes a linear isometry of the Banach–Kantoro

0 0 0 vich spaces Mˆ E(X), Fs (Y ) and Es L (X, Y ) . C In the same way as in the proof of Theorem 5.5.3,for x ∈ X and y ∈ Y we

define an operator S := Sx,y : E → F by Se := y, T (x ⊗ e) . It turns out again that S ∈ Lˆ (E, F ). Denote by λ the isomorphism from Lˆ (E, F ) onto E 0 discussed in the theorem of 3.4.8 and
assign

b(x, y) := λ(Sx,y ). Then, due to the mentioned theorem, we have ˆ eb(x, y) = y, T (x ⊗ e) for all e ∈ E, x ∈ X, and y ∈ Y . It is easy to see that the mapping (x, y) 7→ b(x, y) is a bilinear operator from
X×Y 0 into E . Moreover, b is an operator with abstract norm and b ≤ λ T . At the same time, by formula 4.3.5 (3) we have nX o
T e = sup T x(k) ⊗ c(k) : (x, c) ∈ U n o X 
 = sup sup ˆ c(k)b x(k), y : (x, c) ∈ U kyk≤1

≤ sup sup

nX

o
ˆ c(k) b kx(k)kkyk : (x, c) ∈ U

kyk≤1

≤ˆ

X

c(k) b
=ˆ eb .



Disjointness Preserving Operators

229


Here the set U is the same as in 5.5.3. It is clear that λ T ≤ b ,
which, together with the above-mentioned reverse inequality, yields b
= λ T . According to Proposition 5.5.1 (2), there exists a K ∈ Es0 L (X, Y 0 ) such that K = b and b(x, y) = hy, Kxi (x ∈ X, y ∈ Y ). Taking the definition of b into account, we may write




y, T (x ⊗ e) = ˆ eb(x, y) = ˆ ehy, Kxi = ˆ hy, Kx ⊗ ei .
Thus, hy, T ui = ˆ hy, Kui for all u ∈
X ⊗ E. This means that the operators T1 u := hy, T ui and T2 u := ˆ hy, Kui from E(X) into F coincide on X ⊗ E. Moreover, these operators admit o-continuous dominants S1 and S2 , respectively:
S1 e := T (e)kyk, S2 e := ˆ e K kyk (e ∈ E). Since X ⊗ E is bo-dense in E(X), we have T1 = T2 on the entire E(X). It is also
clear that λ T = K .
0 If there is another element L ∈ Es0 L (X, Y ) that serves as a representing 
operator-function for T then ˆ y, (K − L)u = 0 for all u ∈ E(X). Assigning 


u :=x⊗e, we conclude that ˆ e y, (K −L)x = 0 for all e ∈ E. Therefore, y, (K − L)x = 0. Since x ∈ X and y ∈ Y are arbitrary, K ∼ L. Thus,
the correspondence
0 T 7→ K defines a linear isometry from Mˆ E(X), Fs (Y ) into Es0 L (X, Y 0 ) . We will prove that it is surjective.
Take K ∈ Es0 L (X, Y 0 ) ,
u ∈ E(X), and y ∈ Y . As was mentioned in 5.5.8,
we have the element ˆ hy, Kui ∈ F defined correctly. Assign S (y) := ˆ hy, Kui . u
The operator Su : Y → F is linear and |Su (y)| ≤ ˆ K u kyk. Thus, Su is an operator with abstract norm and, by Proposition 5.5.1 (2), there exists an element v ∈ Fs (Y 0 ) such that Su (y) = hy, vi (y ∈ Y ) and, moreover, v = Su . Assign
Tu := v. Then hy, T ui = Su (y) = ˆ hy, Kui . This defines a linear operator T : E(X) → Fs (Y 0 ) that satisfies the relations

T u = Su = sup ˆ hy, Kui ≤ ˆ K u . kyk≤1

Consequently, T admits an o-continuous dominant S : e 7→ ˆ e K
0 Lˆ (E, F ). Therefore, T ∈ Mˆ E(X), Fs (Y ) . B




and S ∈

5.5.10. Theorem. Let ˆ : E → F be a Maharam operator and let a dominated operator T : E(X) → Fs (Y, Z) be such that T ∈ {ˆ}⊥⊥ . Then there exists a unique (to within equivalence) operator-function K ∈ MP (X, Z 0 ) such that the following hold: hT u, zi = ˆ hz, Kui
T e=ˆ eK





u ∈ D(K), z ∈ Z ,
e ∈ D( K ) ,

230

Chapter 5
 where Z is a norming subspace of
Y 0 , D K := e ∈ E : e K ∈ E , and D(K) := u ∈ E(X) : u ∈ D K . C Arguing as in 5.5.4 and using Theorem 5.5.9, find an operator-function K ∈ MP (X, Z 0 ) satisfying

e hz, Kui hz, T ui = ˆ u ∈ E(X), z ∈ Z ,
e eK T e=ˆ (e ∈ E), e is the extension of ˆ onto L1 (ˆ) by o-continuity. In these representations, where ˆ
e in case u ∈ D(K) and e ∈ D K . B we may write ˆ instead of ˆ 5.6. Comments Disjointness preserving operators in vector lattices have attracted attention for a long time. B. Z. Vulikh [385–387] was one of the first who considered disjointness preserving operators. Later, disjointness preserving operators were studied by Yu. A. Abramovich, E. L. Arenson, W. Arendt, S. J. Bernau, A. E. Gutman, D. R. Hart, C. B. Huijsmans, A. K. Kitover, A. V. Koldunov, S. S. Kutateladze, P. T. N. McPolin, B. de Pagter, A. I. Veksler, A. W. Wickstead, A. C. Zaanen, and many others (see, for instance, [4, 7–12, 26, 27, 40–42, 121–123, 130, 136, 279, 310, 396, 408]). The theory of disjointness preserving operators is rich in results and covers such topics as boundedness, continuity, compactness, spectral properties, analytical representation, etc. Chapter 5 concentrates on boundedness, decomposition, and analytical representation. 5.6.1. (1) An orthomorphism is a band preserving operator that is orderbounded. In [396] A. W. Wickstead raised the question whether every band preserving operator must be order-bounded automatically. Existence of an unbounded band preserving operator was announced for the first time in [12: Theorem 1]. Later, it was clarified that the situation described in the paper is typical in a sense. Namely, Yu. A. Abramovich, A. I. Veksler, and A. V. Koldunov [11: Theorem 2.1] and P. T. N. McPolin and A. W. Wickstead [279: Theorem 3.2]) established that all band preserving operators in a universally complete K-space are automatically bounded if and only if this K-space is locally one-dimensional (Theorem 5.1.2). (The definitions of a locally one-dimensional K-space and a local Hamel basis, as well as the equivalence conditions (1)–(4) from 5.1.1, are presented in [279].) (2) Thus, A. W. Wickstead’s question about boundedness of band preserving operators was given an exhaustive answer modulo the structure of onedimensional K-space. It was conjectured implicitly in [11] and [279] that the notions of locally one-dimensional and discrete K-space coincide. In [13] A. W. Wickstead fixed the conjecture as an open question. The negative solution of 5.1.6

Disjointness Preserving Operators

231

and 5.1.7 was found by A. E. Gutman [122]: there exists a diffuse locally onedimensional K-space (see also [121, 123]). He established this result by describing a locally one-dimensional K-space in terms of its base (Theorem 5.1.5). (3) A. W. Wickstead’s question admits different answers depending on spaces in which the operator in question acts. There are many results that guarantee automatic boundedness for a band preserving operator acting in concrete classes of vector lattices. According to [11, 12] every band preserving operator from a Banach lattice to a normed vector lattice is bounded. This claim remains valid if the domain Banach lattice is replaced by a relatively uniformly complete vector lattice [11]. In [279] a similar result is obtained for band preserving operators acting in a relatively uniformly complete vector lattice which is endowed with a locally convex locally solid topology. (4) The problem of finding sufficient conditions for boundedness of disjointness preserving operators (see [199: 6.5, Problem 7]) remains actual for operators in lattice-normed spaces. Theorem 5.1.10 is due to A. E. Gutman [121, 123]. We call an operator semibounded (the term was introduced in [121]) (5.1.8) if it satisfy the McPolin–Wickstead condition of [279]. (5) In the case of a universally complete K-space, a band preserving order-unbounded operator can be constructed on using V(B) . Moreover, inside an appropriate V(B) this problem reduces to existence of a discontinuous authomorphism of the group (R, +), i.e., an additive but nonlinear function from R to R. Let E be a universally complete K-space and let B := B(E). Take a Boolean algebra B such that R∧ 6= R, see 8.6.1 (7). Then R is an infinite-dimensional space over R∧ inside V(B) . By the Kuratowski–Zorn Lemma, there exist an R∧ -linear but not Rlinear function u : R → R in the model V(B) . The operator U0 := u↓ : R↓ → R↓ is linear, band preserving, but order-unbounded. If ι is an isomorphism of E onto R↓ then U := ι−1 ◦ U0 ◦ ι is a band preserving o-unbounded operator. 5.6.2. (1) The notion of n-disjoint operator in vector lattices was introduced by S. J. Bernau, C. B. Huijsmans, and B. de Pagter [42] and was adapted to operators in lattice-normed spaces by A. G. Kusraev [204]. Different characterizations of n-disjoint operators presented in 5.2.1 (2) and 5.2.5 were obtained by V. A. Radnaev [323, 324] employing Kutateladze’s approach [219] to characterizing lattice homomorphisms. It should be emphasized that the equivalence (1) ⇔ (5) in 5.2.5 gives a purely algebraic characterization of n-disjoint operators as metric n-decomposable elements of the associated Banach–Kantorovich space (5.2.4). (2) Theorem 5.2.7 is due to S. J. Bernau, C. B. Huijsmans, and B. de Pagter [42] (see also [41]); an algebraic approach to the proof (2.1.10, 5.2.6, 5.2.7) was found by V. A. Radnaev [323, 324]. The method of proof in [42] makes it clear that the decomposition of an n-disjoint operator into the sum of n lattice

232

Chapter 5

homomorphisms is nonunique. V. A. Radnaev [323, 324] noticed that, first, the disjoint preserving operators T1 , . . . , Tn in the decomposition T = T1 + · · · + Tn can be chosen pairwise disjoint; second, if S1 , . . . , Sn is another disjoint collection of disjointness preserving operators with T = S1 + · · · + Sn then for every k = 1, . . . , n there exist a partition of unity πk1 , . . . , πkn in P(F ) such that Sk = πk1 T1 + · · · + πkn Sn for all k = 1, . . . , n. Decomposition of an operator into infinite series of lattice homomorphisms was considered in [344]. (3) A positive operator T : E → F is said to be a local homomorphism if there exists a partition of unity (πξ )ξ∈„ in P(E) such that T ◦ πξ is a lattice homomorphism for every ξ ∈ „. N. Kalton [147, 148] proved that each finite sum of lattice homomorphisms in an ideal space over a standard measure space is a local homomorphism. This result was generalized to operators in arbitrary order complete vector lattices by I. I. Shamaev [344]. (4) The shadow of an operator (5.2.2) as a Boolean homomorphism (without introducing the corresponding term) was first considered in [259] for lattices homomorphisms and in [199] for a disjointness preserving operators in lattice normed spaces. Theorem 5.2.3 is due to A. G. Kusraev [199]; for lattice homomorphisms the equivalence (1) ⇔ (2) was proved by W. A. J. Luxemburg and A. Schep [261]; while the equivalence (1) ⇔ (3) was proved by C. B. Huijsmans and B. de Pagter [136]. A. E. Gutman [123] found that some properties of disjointness preserving operators can be expressed in terms of its shadows; in particular, he proved 5.2.2 (3, 4). (5) Yu. A. Abramovich’s condition (R) [4: Theorem A] was the first paraphrase of boundedness for disjointness preserving operators weaker than sequential r-o-continuity. Later, this condition was also weakened: P. T. N. McPolin and A. W. Wickstead showed [279: Theorem 2.1] that, for a disjointness preserving operator in vector lattices to be bounded, it suffices that the operator under test be semibounded (see Theorem 5.2.8). The proof of the nontrivial implication (4) ⇒ (1) in Theorem 5.2.8 repeats P. T. N. McPolin and A. W. Wickstead arguments. In [279] the case Y = F is considered; however, the result remains valid for an operator with values in an arbitrary LNS. (6) Attempts are unsuccessful at generalizing the Abramovich–McPolin– Wickstead criterion to the case of operators in lattice-normed spaces. A. E. Gutman studied four types of boundedness for this class of operators but it turned out that they are pairwise different. Thus the problem of finding sufficient conditions for boundedness remains open for disjointness preserving operators in LNSs. A small step in this direction is made in the following proposition. Let X be a BKS over E and let Y be an LNS over F . A disjointness preserving operator T : X → Y is subdominated if and only if T is semibounded and, for every

Disjointness Preserving Operators

233

positive element e ∈ E, the set { T u : u ∈ X, u = e} is order-bounded in F . Note that every semibounded disjointness preserving operator defined on a vector lattice obviously meets the hypotheses of the last proposition. This allows us to consider this proposition as a generalization of Theorem 5.2.8. 5.6.3. (1) In Section 5.3 we follow A. E. Gutman [121, 123]. The main idea here is that, since every disjointness preserving regular or dominated operator T is h-o-continuous with respect to the shadow h of T (5.2.2 (3, 4)), T can be uniquely extended to the h-closure of the domain of T (5.7.3). From this it follows immediately that the operator admits a W SW -factorization on principal ideals of the domain which can be also extended to the corresponding h-closure. These h-closures are sufficiently representative if the operator under consideration is wide at some set (5.3.1 (2, 3)). Now, an easy application of the Exhaustion Principle gives description for disjointness preserving operators through W SW -representation (Theorems 5.3.6 and 5.3.19). (2) The domain E (1, h) of the shift Sh is maximally wide. More precisely, E (1, h) contains the domain of every regular operator S acting from an orderdense ideal of E into F , having shadow h, and satisfying the equality S(1E ) = h(1)1F . Shift operators are abstract analogs of the composite mappings f 7→ f ◦ s. As regards, composition in spaces of measurable functions, see the survey paper [60]. (3) A linear operator S : E → F defined on an order-dense ideal E ⊂ E is called multiplicative if Se1 Se2 = S(e1 e2 ) for any two elements e1 , e2 ∈ E whose product belongs to E. Observe that the notion of multiplicative operator depends on the choice of the unities 1E and 1F . The following fact was proved in [123]. Theorem. Let E be an order-dense ideal of E . A linear operator S : E → F is a shift operator if and only if S is multiplicative. The study of multiplicative operators in vector lattices was initiated by B. Z. Vulikh [385–387] who proved that o-continuous shift operators in K-spaces with unity are multiplicative (see also [163]). The above theorem generalizes this result to the case of arbitrary shift operators in arbitrary K-spaces. There are some results describing multiplicative operators as extreme points of certain sets of operators (see [71, 90, 93, 318]). (4) The idea of considering the shift of a disjointness preserving operator appears in different contexts. An analogous notion occurs, for instance in [9, 10, 130] and in many papers about isometries of Lp -spaces. The S-correspondence of a positive operator considered in [362] is also a functional analog of shift. If ı : P(E) → Clop(P ) and ı0 : P(F ) → Clop(Q) are the Stone transforms then the

234

Chapter 5

S-correspondence of T ∈ L∼ (E, F ) is defined by q 7→

o \ n   ı Cπ◦T : q ∈ ı0 (π), π ∈ P(F )

(q ∈ Q).

(5) The criterion for WSW-representability stated in 5.3.5 is close to [10]. Some other criteria are presented also in [8–10, 121, 123]). The conventional notion of weighted shift operator does not contain an inner weight (see [8, 10–12, 199]). Involving an inner weight allows us to decompose an arbitrary bounded disjointness preserving operator in lattice normed spaces into the strongly disjoint sum of weighted shift operators (Theorems 5.3.6 and 5.3.10). (6) Not every disjointness preserving regular operator admits a WSWfactorization. The corresponding example was given by Yu. A. Abramovich [4]. Let Q be an extremal compact space without isolated points. In this case, we may find an order-dense ideal E ⊂ C∞ (Q), a family (eξ )ξ∈„ in E, and a family (qξ )ξ∈„ in Q so that the following conditions be satisfied: the set {qξ  : ξ ∈ „} is dense in Q, eξ (qξ ) = ∞ for all ξ ∈ „, and, for each e ∈ E, the numeric set (e/eξ )(qξ ) : ξ ∈ „ is bounded. Then the operator T : E → l∞ („) acting by the rule (T e)(ξ) = (e/eξ )(qξ ) is disjointness preserving and regular (even positive), but T is not a weighted shift operator. Denoting by ρξ the operator of multiplication by the characteristic func
∞ tion χ{ξ} , we obtain a partition of unity (ρξ )ξ∈„ in the algebra P l („) such that all fragments of the form ρξ ◦ T are weighted shift operators. Theorem 5.3.6 says that all disjointness preserving regular operators has the same structure. 5.6.4. (1) The main results of Section 5.4 (Theorems 5.4.5 and 5.4.10) belong to A. E. Gutman [121, 123]. The facts presented in 5.4.1–5.1.4 just repeat Yu. A. Abramovich’s results [4]. Theorem 5.4.5 interprets the decomposition in 5.3.6 of a disjointness preserving operator into the sum of weighted shift operators in terms of their functional representations. As is seen from the proofs, Theorem 5.4.10 leans upon the theory of ample Banach bundles from Section 2.4. Further development of this approach and extension of the multiplicative representation to operators acting in lattice-normed spaces of continuous or measurable sections can be found in the papers of A. E. Gutman [118–123]. (2) The global representation of 5.4.5 and 5.4.10 for a disjointness preserving operator, as well as the notions of the shift of an operator and the corresponding shift function, allows us to interpret the abstract properties of the operator in terms of its concrete functional representation or in terms of the properties of its shift function. Examples of similar interpretations can be found, for instance, in [8–10, 27, 152, 173, 174]. (3) The isometries of function spaces prove very often to be disjointness preserving operators. This phenomenon seems to be discovered by J. Lamperti

Disjointness Preserving Operators

235

[232]; and so disjointness preserving operators (in spaces of measurable functions) are sometimes referred to as Lamperti operators. The literature on the theory of Lamperti operators is extensive; some aspects are reflected in [6, 9, 26, 36, 113, 114, 174, 358]; see also the survey [60]. 5.6.5. (1) Section 5.5 follows the articles [199, 204]. The multiplicative representation 5.5.2 (2) was established by Yu. A. Abramovich [4]. It should be noted that the recent progress in the multiplicative representation of disjointness preserving operators stems from this work of Yu. A. Abramovich. Theorems 5.5.3 and 5.5.4 were proved in [199] (see also [204]). As is seen from the proof, these results are obtained by combining the representation method of Yu. A. Abramovich and the technique of dominated operators. (2) The notion of a decomposable operator was introduced in [199]. The main results on decomposable operators, which are presented in 5.5.6, 5.5.7, 5.5.8, and 5.5.9, belong to A. G. Kusraev [199, 204]. The auxiliary result in 5.5.1 (2) belongs to G. N. Shotaev [349].

Chapter 6 Integral Operators

This chapter deals with different classes of dominated operators whose common feature is integrality. Speaking of integrality we mean the possibility of integral representation with respect to a vector-valued or scalar-valued measure. Given a σadditive measure taking values in a bo-complete lattice-normed space, a Lebesguetype integral of numerical functions or, more generally, of elements of a universally complete vector lattice can be constructed (6.1.1, 6.1.2, 6.1.3). This is fairly straightforward and all simple properties of the resultant integral as well as analog of the Lebesgue convergence results are easily deduced (6.1.4, 6.1.5, 6.1.6). This integral is used to obtain the Riesz–Markov-type representation theorem for dominated operators defined on a lattice of bounded continuous functions (6.2.6). The corresponding class of measures is quasi-Radon measures. A dominated measure and its least dominant are or are not quasi-Radon measures simultaneously (6.2.2). The space of all integrable functions (elements) is some lattice-normed space that is neither bo-complete nor decomposable in general. This circumstance gives rise to new classes of measures: the space of integrable elements is decomposable if and only if the measure is modular (6.1.9 (3)) and it is bo-complete if and only if the measure is ample (6.1.9 (4)). A Radon–Nikod´ ym-type theorem is also valid for ample measures (6.1.11). Ample measures are closely connected with the theory of Maharam operators. Maharam extension of a positive operator, described in Section 3.5, leads to a Maharam operator whose domain is too large in general. But for an order continuous operator the extended domain space admits functional representation, while the extended Maharam operator is represented as the integral with respect to some ample measure (6.3.6). Classical integral operators defined by measurable kernels admit the following intrinsic characterization: a linear operator between ideal spaces is an integral operator if and only if it takes order-bounded sequences converging to zero in measure into sequences converging to zero almost everywhere (6.4.5). The same is true for dominated operators acting in the spaces of measurable vector-functions (6.4.10,

Integral Operators

237

6.4.11) if integrality is suitably defined (6.4.6). But this fact relies upon the inheritance of integrality under domination. If some dominant of an operator acting in the spaces of measurable vector-functions is an integral operator then the operator itself admits weak integral representation (6.4.10 (2)). The converse is true too (6.4.7). A broad class of operators arises from integration with respect to a family of measures depending on some measurable parameter. Such operators are called pseudointegral (6.5.1). It turns out that a positive operator admits pseudointegral representation if and only if it is order continuous, i.e. takes order-bounded sequences converging to zero almost everywhere into sequences converging to zero almost everywhere (6.5.4). From this fact, using the properties of dominated operators, the following criterion for the weak pseudointegrality is deduced: a dominated operator admits weak pseudointegral representation whenever it is order continuous (6.5.8). The above-mentioned results imply assertions about the general form of various classes of dominated operators (6.1.7, 6.3.8, 6.3.11, 6.4.12, 6.5.9). 6.1. Vector Integration The main goal of this section is to present some Lebesgue-type integration for measures taking values in lattice-normed spaces. A Radon–Nikod´ ym-type theorem is also established for a specific class of ample measures. 6.1.1. Let G be a universally complete vector lattice with order-unity 1 and let (Y, F ) be a sequentially bo-complete lattice-normed space over an order complete vector lattice F . Fix a subalgebra A in the complete Boolean algebra G(1) of unit elements of G and a finitely additive measure µ : A → Y with the bounded vector variation µ : A → F . Denote by S(A ) the vector sublattice of G comprising all A -simpleP (finite-valued) elements, i.e. x ∈ S(A ) means that there is a representan tion x = k=1 αk ek , where {α1 , . . . , αn } ⊂ R and {e1 , . . . , en } ⊂ A are pairwise disjoint. Put Z n X
Iµ (x) := x dµ := αk µ(ek ) (x ∈ S A ) . k=1

It is clear that this formula correctly defines some dominated linear operator Iµ : S(A ) → Y and
R R (1) x dµ ≤ |x| d µ x ∈ S(A ) . Consider the principal ideal G(1) generated by 1 with the norm kxk := inf{λ : |x| ≤ λ1}, so that G(1) is an AM -space (see 1.5.5). Let C(A ) be the closure of S(A ) in the AM -space G(1).

238

Chapter 6

(2) The operator Iµ admits a unique dominated extension to C(A ) which is denoted by the same symbol. Moreover, Iµ = I µ . C From (1) we immediately deduce that Z
x dµ ≤ kxk · µ (1) x ∈ S(A ) , so that Iµ is bo-continuous. Now, to obtain a unique dominated extension of Iµ to S(A ) we apply Theorem 4.3.3, taking into consideration the sequential bocompleteness of Y . The inequality (1) preserves under passage to the limit. Thus, Iµ ≤ I µ and it remains to check that Iµ e = I µ e for all e ∈ A . Using Theorem 4.1.8 with E := X and 4.2.9 (1), we derive Iµ e =

( n _ X

Iµ ek : ek ∧ el = 0 (k 6= l)

n _

) ek = e

= µ (e) = I µ e. B

k=1

k=1

(3) For every dominated operator T : C(A ) → Y there is a unique dominated measure µ : A → Y such that Z
T x = x dµ x ∈ C(A ) . The correspondence T 7→ µ is a linear isometric isomorphism of the lattice-normed spaces M (C(A ), Y ) and da(A , Y ). Moreover, T ∈ MG (C(A ), Y ) (T ∈ MσG (C(A ), Y )) if and only if we have µ ∈ dao(A , Y ) (respectively µ ∈ dca(A , Y )). Now we assume that A is a σ-subalgebra in E(1). Consider a universally complete Kσ -space E ⊂ G comprising all A -valued resolutions of unity (spectral functions, see 1.4.3). The inclusion E ⊂ G is understood by virtue of Theorem 1.4.4 which enables us to identify the K-spaces G and K(E(1)). 6.1.2. For the sequel we need some technical facts. (1) If a sequence (xn )n∈N ⊂ E decreases to zero then (∀ 0 < ε ∈ R)

∞ ^

µ (1 − exε n ) = 0.

n=1

C The proof is immediate from the σ-additivity of µ (4.4.12) and the suitable properties of spectral functions (1.3.8 (9)): ∞ _ n=1

n xn exε n = einf = e0ε = 1. B ε

Integral Operators

239

(2) If a sequence (xn )n∈N ⊂ S(A ) decreases and inf n xn = 0 in G, then xn d µ = 0. C Take a positive number α with x1 ≤ α1. Then for every ε > 0 and n ≥ 1 we have xn ≤ ε1 + α(1 − exε n ). From this it is seen that Z xn d µ ≤ ε µ (1) + α µ (1 − exε n ). inf n

R

Using (1), we obtain ∞ Z ^

xn d µ ≤ ε µ (1).

n=1

Since ε > 0 is arbitrary, the result follows. B (3) For every positive element x ∈ E there is a sequence of positive finite-valued elements (xn )n∈N ⊂ S(A ) with xn % x. (4) Let E be a universally σ-complete vector lattice with order-unity 1 and let (xn )n∈N be an unbounded sequence in E+ . Then there exists a unit element e∞ ∈ E(1) such that the sequence ([1 − e∞ ]xn )n∈N is order-bounded and te∞ = supn (te∞ ∧ xn ) for every 0 < t ∈ R. C Put 0

σ (λ) :=

∞ ^

exλn ,

σ(λ) :=

n=1

_ ν t then τ (λ) = inf n eyλn ∨ e∞ = e∞ . According to what was said after Theorem 1.3.8 we conclude that τ is the spectral function of supn (te∞ ∧ yn ) and the result follows. B

240

Chapter 6

6.1.3. Now define the integral for elements which can be approximated by A -simple elements. We say that a positive element x ∈ E is integrable by µ, or µ-integrable if there is an increasing sequence (xn )n∈NR of positive elements in S(A ) o-converging in G to x and the supremum supn∈N xn d µ existing in F . For such a sequence (xn ) the sequence of the integrals (Iµ (xn ))n∈N is bo-fundamental. Indeed, by applying 6.1.1 (1) we have Z

Z xn dµ −

Z xm dµ ≤

|xn − xm | d µ ≤

∞ Z _

 xk d µ

Z −

xp d µ −−−→ 0,

k=1

p→∞

where p = min{m, n}. Now we may define the integral of x by putting Z Iµ (x) :=

Z x dµ := bo-lim

n→∞

xn dµ.

To check the soundness of this definition, take one more sequence (yn )n∈N ⊂ S(A ) increasing to x in G and assume that supn I µ (yn ) exists in F . Using 6.1.1 (1), 6.1.2 (2), and 1.3.2 (5) we deduce Z

Z

Z

xn dµ − ym dµ ≤ |xn − ym | d µ Z Z ≤ xn ∨ ym d µ − xn ∧ ym d µ Z _ Z = (xn ∨ ym ) ∧ (xk ∧ yl ) d µ − xn ∧ ym d µ k,l∈N

≤

_ Z

Z xk ∧ yl d µ −

xn ∧ ym d µ .

k,l∈N

Consequently, Z bo-lim

m,n→∞



Z xn dµ −

ym dµ

=0

and the soundness of our definition is established. An element x ∈ E is integrable (= µ-integrable) if its positive part x+ and negative part x− are both integrable. Denote by L 1 (µ) the set of all integrable elements and, given x ∈ L 1 (µ), put Z Iµ (x) :=

Z x dµ :=

+

x dµ −

Z

x− dµ.

Integral Operators

241

It can be easily checked, using 6.1.2 (2), that L 1 (µ) is an order-dense ideal in E and Iµ : L 1 (µ) → Y is a linear operator. Moreover, 6.1.1 (1) holds for all x ∈ L 1 (µ). Note also that the construction of the integral implies L 1 (µ) = L 1 ( µ ). Define in L 1 (µ) an F -valued seminorm Z
x 1 := |x| d µ x ∈ L 1 (µ) . We say that two elements x, y ∈ L 1 (µ) are µ-equivalent if there is a unit element e ∈ G(1) with µ (1 − e) = 0 and [e]x = [e]y. The set N (µ) of all elements that are µ-equivalent to zero is a sequentially o-closed order ideal in L 1 (µ). It follows from the definition of integral that N (µ) = {x ∈ L 1 (µ) : x 1 = 0}. Define the Kσ space L1 (µ) as the factor space of L 1 (µ) by σ-ideal N (µ). The coset of an element ~. An F -valued
norm in L 1 (µ) is introduced by x ∈ L 1 (µ) will be denoted by
x ~ 1 := x x ∈ L 1 (µ) . Thus, L1 (µ), · is a lattice-normed space. setting x 6.1.4. Monotone Convergence Theorem. Assume that (xn )n∈N is
an increasing sequence of µ-integrable elements and let the sequence I µ (xn ) n∈N be W ~ = nx ~n order-bounded in F . Then there is x ∈ L 1 (µ) such that the identity x holds in L1 (µ) and Z Z x dµ = bo-lim

n→∞

xn dµ.

C First we observe that if an increasing sequence (yn )n∈N of positive elements in L 1 (µ) has a least upper bound y := supn yn in G and the sequence I µ (yn ) has a least upper bound in F then y ∈ L 1 (µ) and Z ∞ Z _ yd µ = yn d µ . n=1

Indeed, for every n ∈ N take an increasing sequence of positive elements (zn,k )k∈N ⊂ S(A ) such that yn = supk zn,k (n ∈ N). Since y = supk supn≤k zn,k , employing the definition of integral with respect to the measure µ we deduce Z yd µ =

∞ Z _ _ k=1

zn,k d µ ≤

n≤k

∞ Z _ _ k=1

yn d µ =

∞ Z _

yn d µ .

n=1

n≤k

Since the reverse inequality is evident, the result follows. According to 6.1.2 (4) there is a unit element e ∈ A such that m(1 − e) = supn (m(1 − e) ∧ xn ) (m ∈ N). It follows from this and the above observation that ∞ Z ∞ Z _ _ m µ (1 − e) = m(1 − e) ∧ xn d µ ≤ xn d µ . n=1

n=1

242

Chapter 6

As m ∈ N is arbitrary and the integrals on the right-hand side are order-bounded, ~n and ~ = supn x we conclude that µ (1 − e) = 0. Put x := supn [e]xn . Then x Z ∞ Z ∞ Z _ _ xd µ = [e]xn d µ = xn d µ . n=1

n=1

Taking into consideration the fact that I Z

Z x dµ −

µ

is a dominant of Iµ we deduce

Z xn dµ ≤

Z |x − xn | d µ =

(o)

Z xd µ −

xn d µ −−−→ 0, n→∞

and the proof is complete. B 6.1.5. Dominated Convergence Theorem. Let (xn )n∈N be a sequence of µ-integrable elements and let o-lim xn = x in G. If y ∈ L 1 (µ) and |xn | ≤ y (n ∈ N) then x ∈ L 1 (µ) and Z Z x dµ = bo-lim

n→∞

xn dµ.

C Since L 1 (µ) is an order-dense ideal in E and E is sequentially order-closed in G, we have x ∈ L 1 (µ). The assumption that o-lim xn = x and the evident inequality |x − xn | ≤ 2y imply the existence of a decreasing sequence (yn )n∈N ⊂ L 1 (µ) such that |x − xn | ≤ yn and o-lim yn = 0. Thus, y1 − yn % y1 and, using the Monotone Convergence Theorem, we have Z Z y1 d µ = o-lim (y1 − yn ) d µ . n→∞

Consequently, Z

Z x dµ −

Z xn dµ ≤

Z |x − xn | d µ ≤

yn d µ −−−→ 0 n→∞

and the result follows. B 6.1.6. The above results can be summarized as follows: Theorem. Let µ : A → Y be a dominated countably additive measure. There exist an order-dense ideal L 1 (µ) ⊂ E and a sequentially bo-continuous dominated operator Iµ : L 1 (µ) → Y such that (1) L 1 (µ) ⊃ A ; (2) Iµ e = µ(e) (e ∈ A ); (3) if L ⊃ A and Ie = µ(e) (e ∈ A ) for some order-dense ideal L ⊂ E and a bo-continuous dominated operator I : E → Y then L ⊂ L 1 (µ) and Ix = Iµ x (x ∈ L);

Integral Operators

243

(4) Iµ = I µ . C The existence and sequential order continuity of Iµ defined on the orderdense ideal L 1 (µ) ⊂ E and satisfying (1) and (2) was justified in 6.1.1, 6.1.3 and 6.1.5. The condition (3) is deduced from 6.1.4 and 6.1.3 (3). Finally, (4) is an easy consequence of 6.1.1 (1), 6.1.2 (3), and sequential order continuity I µ . B 6.1.7. Theorem. Let E0 be an order-dense ideal in E containing order-unity and let T : E0 → Y be a sequentially bo-continuous dominated linear operator. Then there exists a unique dominated σ-additive measure µ : A → Y such that L 1 (µ) ⊃ E0 and Z Z T x = x dµ, T x = x d µ (x ∈ E0 ). C We have only to define µ(e) := T e (e ∈ A ) and observe that T and Iµ coincide on S(A ). Then T and Iµ coincide on L (µ) by bo-continuity. B 6.1.8. Now we proceed
to a natural question under what conditions the latticenormed space L1 (µ), · is bo-complete.
The space L1 (µ), · is br-complete. C The proof uses much the same arguments as the Riesz–Fisher Completeness Theorem for a scalar measure. We sketch the proof. If a sequence (xn )n∈N is brfundamental then xn − xm ≤ rk f (n, m ≥ k) for some f ∈ F and a numeric sequence (rn )n∈N converging to zero. Denote L(f ) := {x ∈ L1 (µ) : x ∈ F (f )}, where F (f ) is an order ideal generated by f . Then (xn ) ⊂ L(f ) and it suffices to show that P L(f ) is a Banach space under the norm kxk := k x kF (f ) . Suppose P∞ that ∞ the series n=1 xn is absolutely convergent in L(f ). Then the series n=1 xn converges in the AM -space F (f ) and also r-converges to the same limit. Put t :=

∞ X k=1

xk ,

σn :=

n X

xk ,

sn :=

k=1

n X

|xk |.

k=1

R

Clearly, (sn ) has positive entries, increases, and sn ≤ t. Hence, by the Monotone Convergence Theorem, there exists a limit y := o-lim sn , with the resultantP element ∞ 1 y a member of L (µ). The inequalities |σn | ≤ sn ≤ y imply that the series k=1 xk o-converges. For the sum x0 the estimate holds: |x0 | ≤ y, whence x0 ∈ L1 (µ). RAppealing to the Dominated Convergence Theorem, conclude that σn − f0 = |σn − x0 | d µ → 0. B To obtain further completeness results, some additional assumption are necessary. For example, sequential bo-completeness of L1 (µ) can be proved whenever F is regular. But the decomposability phenomenon involves essentially different properties of measures.

244

Chapter 6

6.1.9. In the rest of this section we assume that Y is a Banach–Kantorovich space over a K-space F and µ : A → Y is a countably additive measure. Define the null ideal of µ by N (µ) := {a ∈ A : (∀ a0 ∈ A ) a0 ≤ a ⇒ µ(a0 ) = 0}. Evidently N (µ) = N ( µ ) = {a ∈ A : µ (a) = 0}. Let A~ and φ denote the factor algebra A /N (µ) and the canonical factor mapping A /N (µ) → A~ respectively. ~ = f There is a unique measure µ ~ : A~ → Y such that µ ~ ◦ φ = µ. Furthermore, µ µ. Given a Boolean homomorphism h : B := P(F ) → A~, we say that µ is modular µ(φa) = µ ~(h(b) ∧ φ(a)) for all a ∈ A and b ∈ B. with respect to h, or h-modular if b~ Clearly, the modularity of µ means that bµ(a) = µ(b0 ∧ a0 ) for all a0 ∈ φ(a) and b0 ∈ h(b). W Let e := {b ∈ B : (∀ a ∈ A ) bµ(a) = 0}. Then eµ(A ) = {0} and µ(A ) ⊂ (1 − e)Y . Moreover, bµ(A ) = {0} if and only if h(b) ∈ N (µ). Thus, h is injective on [0, 1 − e]. In the sequel we agree that µ(A )⊥⊥ = Y and in this event h is an isomorphic embedding of B into A~. An h-modular measure µ is said to be ample (with respect to h) if for any partition of unity (bξ )ξ∈„ in B and an arbitrary family (aξ )ξ∈„ in A there exists a unique (to within equivalence) element a ∈ A such that bξ µ (a4aξ ) = 0 for all ξ ∈ „. This condition is equivalent to h(bξ ) ∧ φ(a) = h(bξ ) ∧ φ(aξ ) (ξ ∈ „), since µ is h-modular. (1) A measure µ is modular with respect to a Boolean isomorphism h if and only if so is its exact dominant µ .
C Suppose that µ is h-modular and prove the identity b f µ φ(a) = f µ h(b) ∧


f ~ φ(a) = µ ~ h(b) ∧ φ(a) , since µ = µ ~ . Now the φ(a) . It is equivalent to b µ desired identity is verified by means of the following simple calculations with the use of 4.2.9 (1): ( n ) _ X
~ φ(a) = b µ ~(~ ak ) : ~ a1 ∨ · · · ∨ ~ an = φ(a) bµ ( =

=

=

_

k=1 n X

)


µ φ(ak ) : a1 ∨ · · · ∨ an = a b~

(k=1 n _ X (k=1 n _ X

)
~ φ(b0 ) ∧ φ(ak ) : a1 ∨ · · · ∨ an = a µ )
~ φ(ck ) : c1 ∨ · · · ∨ cn = φ(b0 ) ∧ φ(a) µ

k=1


= µ h(b) ∧ φ(a) ,

Integral Operators

245

a1 , . . . , ~ an } ⊂ A~, {a1 , . . . , an } ⊂ A , and {c1 , . . . , cn } ⊂ A where the finite sets {~ are pairwise disjoint and h(b) = φ(b0 ) for some b0 ∈ A . The converse follows from Proposition 6.1.11 (1) below. B (2) Let F obey the countable chain condition. Then every countably additive h-modular measure µ defined on a σ-algebra is ample. C W For a countable partition of unity (bn ) ⊂ B and a sequence (an ) ⊂ A put a := {cn ∧ an : n ∈ N}, where (cn ) is a sequence of pairwise disjoint elements in A with h(bn ) = φ(cn ), n ∈ N. Using the modularity and countable additivity of µ we derive: bm µ(a) = =

∞ X n=1 ∞ X

bm µ(cn ∧ an ) =

∞ X


~ φ(cn ) ∧ φ(an ) bm µ

n=1



~ φ(cm ) ∧ φ(cn ) ∧ φ(an ) = µ µ ~ h(bm ) ∧ φ(am )

n=1


~ φ(am ) = bm µ(am ). B = bm µ In the sequel we follow the common practice of identifying µ and µ ~ when this leads to no confusion. (3) The lattice-normed space L1 (µ) is disjointly decomposable if and only if µ is a modular measure. C Suppose that µ is modular with respect to some Boolean isomorphism h. Prove that b x 1 = h(b)x 1 for all b ∈ B and x ∈ L1 (µ). We identify a unit element h(b) ∈ A with the band projection [h(b)] in E and write h(b)x instead of [h(b)]x. If x = τ1 a1 + · · · + τn an with τ1 , . . . , τn ∈ R and a1 , . . . , an ∈ A pairwise disjoint, then h(b)x = τ1 h(b) ∧ a1 + · · · + τn h(b) ∧ an and we may write Iµ (h(b)x) =

n X k=1

τk µ(h(b) ∧ ak ) =

n X

τk bµ(ak ) = bIµ (x).

k=1

Now take an increasing sequence (xn ) ⊂ S(A ) such that Iµ (x) = bo-limn Iµ (xn ). Then h(b)x = o-limn h(b)xn , since the band projection [h(b)] is order continuous. Moreover, h(b)xn ≤ x and also h(b)x ∈ L1 (µ), because L1 (µ) is an order ideal in E. Using the Dominated Convergence Theorem, we deduce
bIµ (x) = b bo-lim Iµ (xn ) = bo-lim Iµ (h(b)xn ) = Iµ (h(b)x). n→∞

n→∞

It follows that L1 (µ) is d-decomposable. The converse is evident. B (4) The lattice-normed space L1 (µ) is disjointly complete if and only if µ is an ample measure.

246

Chapter 6

C Suppose that µ is ample. Let (bξ )ξ∈„ be a partition of unity in P(F ) and let (xξ )ξ∈„ be a norm-bounded family in L1 (µ). Let eξ (·) : R → A be the spectral function of xξ and define e(λ) :=

_

h(bξ ) ∧ eξ (λ) (λ ∈ R).

ξ∈„

The function λ 7→ e(λ) is a resolution of unity, i.e. satisfies 1.4.3 (1–3). This is checked by straightforward calculation employing the associativity laws (1.1.2) and the infinite distributive laws (1.1.3). The only
claim that needs clarification is the V identity e := {e(λ) : λ ∈ R} = 0. Since h(bξ ) is a partition of unity in A /N (µ), it is sufficient to show that h(bξ ) ∧ e = 0 for all ξ ∈ „. The due calculations are as follows:

h(bη ) ∧ e =

∞ ^ n=1

=

∞ _ ^ n=1 ξ∈„

h(bη ) ∧ e(−n) =

∞ ^

 h(bη ) ∧

n=1

h(bη ) ∧ h(bξ )eξ (−n) =

 _

h(bξ )eξ (−n)

ξ∈„ ∞ ^

h(bη ) ∧ eη (−n) = 0.

n=1

By Theorem 1.4.4 there is an element x in the universal completion mL1 (µ) with e(λ) = exλ (λ ∈ R). If y ∈ L1 (µ) is an upper bound of the family (|xξ |) then |x| ≤ y, h(b )x h(b )x so that x ∈ L1 (µ). Applying 1.3.8 (12), obtain eλ ξ = eλ ξ ξ (λ ∈ R), whence h(bξ )x = h(bξ )xξ . Thus we have verified that L1 (µ) is disjointly complete. Since the reverse is evident, the proof is complete. B 6.1.10. Theorem. For a countably additive measure with values in a Banach– Kantorovich space the following are equivalent: (1) µ is an ample measure; (2) µ is an ample measure; (3) L1 (µ) is a Banach–Kantorovich space; (4) L1 (µ) is an order complete vector lattice and the operator T :
1 ~ = I µ (x) x ∈ L 1 (µ) is a Maharam operator. L (µ) → F defined by T x C (1) ⇒ (2): Follows from 6.1.9 (1). (2) ⇒ (3): Follows from 6.1.8, 6.1.9 (4), and 2.2.3. (3) ⇒ (4): We know that L1 (µ) is order σ-complete. The Dominated Convergence Theorem implies that the F -valued norm in L1 (µ) is sequentially order continuous, i.e. for every decreasing sequence (xn ) ⊂ L1 (µ) we have o-limn xn = 0 whenever o-limn xn = 0. Take an order-bounded set M ⊂ L1 (µ) and let u ∈ L1 (µ)

Integral Operators

247

be its upper bound. Without loss of generality P we may assume that M contains all elements of the form x1 ∨ · · · ∨ xn and bo- ξ∈„ h(bξ )xξ , where {x1 , . . . , xn } ⊂ M , (xξ ) ⊂ M , and (bξ ) is a partition of unity in P(F ), since supplementing M with these elements, we will not change the set of upper bounds of M . Put f := sup{ x : x ∈ M } ≤ u and choose a sequence (xn ) ⊂ M with f − xn ≤ (1/n)f . If yn := x1 ∨ · · · ∨ xn then f − yn ≤ (1/n)f (n ∈ N) and (yn ) is increasing. If y = supn yn then y = f . We now take an arbitrary z ∈ M and verify that z ≤ y. To this end, observe that o-limn yn ∨ z = y ∨ z ≥ y and o-limn yn ∨ z = y ∨ z ≤ f , since yn ∨ z ∈ M . Thus, f ≤ y ≤ y ∨ z ≤ f , so that f = y ∨ z . Since the norm on L1 (µ)+ is additive, we have f = (y ∨z −y)+y = y ∨z −y + y = y ∨z −y +f , whence y ∨ z − y = 0 and y ∨ z = y. So, we have proved that y = sup(M ) and L1 (µ) is order complete. Consider a downward directed set D with inf(D) = 0 and put f := inf{ x : x ∈ D}. Repeating the above arguments we may choose a decreasing sequence (yn ) ⊂ L1 (µ) (not in D in general) such that f = inf n yn and 0 = inf n yn . Since the norm is order σ-continuous, we obtain f = 0. Thus, T is order continuous, because T x = x+ − x− . According to 2.1.8 (4) L1 (µ) admits a compatible module structure over Orth(F ). Therefore, if 0 ≤ f ≤ T x for some 0 ≤ x ∈ L1 (µ), then there is π ∈ Orth(F ) such that f = πT x = π x = πx = T (πx), so that T possess the Maharam property. (4) ⇒ (1): This is an easy consequence of 3.4.3. B 6.1.11. Let ν : A → F and µ : A → Y be finitely additive measures. We say that µ is absolutely continuous with respect to ν and write µ  ν if µ(a) ∈ ν(A)⊥⊥ for all a ∈ A . If µ  ν then N (ν) ⊂ N (µ) and a natural Boolean homomorphism % : A /N (ν) → A /N (µ) can be defined by %◦φ0 = φ, where φ0 is a canonical factor mapping A → A /N (ν). Denote by χa the band projection in E corresponding to some unit element a ∈ A . (1) Suppose that ν is modular with respect to h and µ  ν. Then µ is modular with respect to h ◦ %. C Without loss of generality we may assume that ν is positive. By hypothesis,

⊥⊥ 0 µ h(b) ∧ φ(a) ∈ ν h(b) ∧ φ (a) , so that b⊥ µ h(b) for every b ∈ B we have

∧ φ(a) = 0 for
all b ∈ B and a ∈ A . From this we obtain µ h(b)
∧ φ(a) = ⊥ ⊥ bµ h(b) ∧ φ(a) . Substituting = 0, whence
b for b, we arrive at bµ h(b ) ∧ φ(a)
bµ(φ(a)) = bµ h(b) ∧ φ(a) . Hence, bµ(φ(a)) = µ h(b) ∧ φ(a) and the result follows. B (2) Radon–Nikod´ ym Theorem. Let µ, ν : A → F be countably additive measures with µ positive and ample. If ν is absolutely continuous with

248

Chapter 6

respect to µ then there exists y ∈ L 1 (µ) such that Z ν(a) = χa y dµ (a ∈ A ). C First assume that |ν| ≤ µ. Define an operator Sν : L1 (µ) → F by Z Sν (~ x) := x dµ (x ∈ L 1 (µ)). Clearly, Sν  T , so that by Theorem 3.4.9 there is ρ ∈ Orth(L1 (µ)) such that |ρ| ≤ I and Sν (u) = T (ρu) (u ∈ L(µ)). The orthomorphism ρ is representable as ρ(u) = y~u for some y~ ∈ L1 (µ), |y| ≤ 1, whence Z ν(a) = T (χa y~) = χa y dµ (a ∈ A ). To handle the general case, put νn := ν ∧(nµ) and Sn := Sνn (n ∈ N). Then νn % ν, Sn % Sν and by above proved there is an increasing sequence (yn ) ∈ L 1 (µ) with νn (a) = Iµ (χa yn ) (a ∈ A ). Since Iµ (yn ) = νn (1) ≤ ν(1), we may apply the Monotone Convergence Theorem. Thus, the element y := supn yn is contained in L 1 (µ) and ν(a) = Iµ (χa y) (a ∈ A ). B (3) Let µ be a positive ample measure. Then the mapping x → νx , where νx is defined by νx (a) := Iµ (χa x) (a ∈ A ), is a lattice isomorphism of vector lattices L1 (µ) and {µ}⊥⊥ . 6.2. Integral Representation by Quasi-Radon Measures In this section we establish an integral representation result for dominated operators on some lattices of continuous functions. Of vital importance in this connection is the class of quasi-Radon measures. 6.2.1. (1) We now specify the vector integral of the proceeding section for elements of some abstract Kσ -space. Take as a universally complete K-space G the vector lattice QR of all real-valued functions defined on a nonempty set Q. Let A be an algebra of subsets of Q, i.e. A ⊂ P(Q). This algebra we identify with the isomorphic algebra of the characteristic functions {1A := χA : A ∈ A } so that S(A Pn) is the space of all A -simple functions on Q, i.e., f ∈ S(A ) means that f = k=1 αk χAk for some α1 , . . . , αn ∈ R and disjoint A1 , . . . , An ∈ A . Let a measure µ be defined on A and take values in a bo-complete lattice-normed space Y over a K-space F . We suppose that µ ∈ da(A , Y ). If f ∈ S(A ) then we put by definition Z n X Iµ := f dµ = αk µ(Ak ). k=1

Integral Operators

249

As was described in Section 6.1, the integral Iµ can be extended to the spaces of µ-summable functions L 1 (µ) for which the more informative notations L 1 (Q, µ) and L 1 (Q, A , µ) are also used. On identifying equivalent functions, we obtain the Kσ -space L1 (µ) := L1 (Q, µ) := L1 (Q, A , µ). Observe that order convergence in G := QR coincides with pointwise convergence, while order convergence in L1 (µ) is defined by almost everywhere convergence. (2) In this section we fix the following notation: Q is a completely regular topological space; T , F , and K are respectively the collections of all open, closed and compact subsets of Q; Cb (Q) is the space of all bounded continuous functions on Q; M (Q) and Mb (Q) are respectively the spaces of all Borel and bounded Borel functions on Q. For a family D of subsets of Q denote by σ0 (D) (σ(D)) the smallest subalgebra (respectively, σ-algebra), containing D. In this event we say that σ0 (D) is generated and σ(D) is σ-generated by D. (3) Let C ∈ A . Consider the families of sets KC = {K ∈ K ∩A : K ⊂ C} and FC = {D ∈ F ∩ A : D ⊂ C} directed by inclusion. A measure µ : A → Y is said to be Radon (quasi-Radon) if for each C ∈ A (for every C ∈ T ∩ A ) the identity µ(C) = bo-lim{µ(K) : K ∈ KC } holds. A measure µ : A → Y is called regular (quasiregular) if for every C ∈ A (for every C ∈ T ∩ A ) the equality µ(C) = bo-lim{µ(D) : D ∈ FC } holds. In the case of a compact space Q these two definitions are equivalent. Moreover, we may prove that µ ∈ da(A , Y ) is a Radon measure (a regular measure) if and only if the vector variation of µ is a Radon measure (a regular measure). But in our considerations of particular interest is a similar result for quasi-Radon measure. To prove it, we need an auxiliary fact that can be easily deduced from the Birkhoff– Ulam Theorem. (4) Let Q be an extremal compact space. There exists an order σcontinuous lattice homomorphism β from the vector lattice of Borel functions M (Q, Bor(Q)) onto WC∞ (Q) such that h is the identity operator on C∞ (Q). Moreover, β(supα fα ) = α β(fα ) for every increasing order-bounded net W (fα ) in C∞ (Q) (sup on the left-hand side denotes the pointwise supremum, while in the righthand side means the least upper bound in C∞ (Q)). C We need only to put Z
β(f ) := f dϕ f ∈ M (Q, Bor(Q)) , Q

where ϕ : Bor(Q) → C∞ (Q) is the Birkhoff–Ulam homomorphism of 1.2.6. B 6.2.2. Theorem. Suppose that a measure µ ∈ da(A , Y ) satisfies one of the following conditions:

250

Chapter 6

(1) A is generated by its closed sets, i.e. A = σ0 (F ∩ A ); (2) µ is countably additive and A is σ-generated by its closed sets, i.e. A = σ(F ∩ A ). Then µ is quasi-Radon measure (quasiregular measure) if and only if its exact dominant µ is quasi-Radon measure (quasiregular measure). C We confine exposition to (2); the case (1) is settled along the same lines. Suppose that µ is a quasi-Radon measure but µ lacks this property. Then there is a set U ∈ T ∩ A such that _ f := µ (U ) − { µ (K) : K ∈ K ∩ A , K ⊂ U } > 0. If K ∈ K ∩ A and K ⊂ U then µ (U \ K) ≥ f . By Definition 4.2.9 (1) there are 0 < ε0 ∈ R, 0 < e0 ∈ G(e), e = µ (U ), and a finite collection (Ci )i∈N ⊂ A such that n n X [ Ci = U \K, Ck ∩ Ci = ∅ (k 6= i), µ(Ci ) ≥ ε0 e0 . i=1

i=1

Let ω1 be the least uncountable ordinal. For some countable ordinal α0 < ω1 all Ci (i = 1, . . . , n) are included to the Baire class Bα0 (σ0 (F ∩ A )) over the algebra σ0 (F ∩A ). The ordinal α0 may be thought nonlimit. Each set from the Baire class Bα = Bα (σ0 (F ∩A )) is either a countable union or a countable intersection of some sets from the preceding Baire classes. Therefore, there are α1 < α0 and sequences (Ci,k )k∈N (i = 1, . . . , n) in Bαi such that the sequence (Ci,k )k∈N is monotone and converges to Ci for all i. It can be also assumed that Ci,k ⊂ U \K. for i ≤ n and k ∈ N. Take an arbitrary δ > 0. By additivity of µ, we may choose e1 ∈ G(e), 0 ≤ e1 ≤ e0 , and k1 ∈ N, with     n X [ δ µ Ci,k1 \ Cj,k1 ≥ ε0 − e1 . 4 j α1 > · · · > αm−1 > αm , a sequence of elements n n (ek )m k=1 ⊂ G(e), sequences of sets (Ci )i=1 ⊂ Bαk (k = 0, 1, . . . , m) such that 0 < em ≤ em−1 ≤ · · · ≤ e1 and     n X [ δ δ k k µ Ci \ Cj ≥ ε0 − − · · · − k+1 ek (k = 0, 1, . . . , m). 4 2 i=0 j 0 there are a compact space K∅ ∈ K ∩ A , K∅ ⊂ H∅ , and a fragment g0 ∈ E(e), 0 < g0 ≤ g, such that for every K 0 ∈ K ∩ A with K∅ ⊂ K 0 ⊂ H∅ , the relations hold: g0 µ(K 0 \K∅ ) ≤ ε0 e,

gi µ(H∅ \K∅ ) ≤ ε0 e.

(Here we mean the multiplication of the order ideal F (e) with ring unity e.) For every i ∈ M there are a compact space K{i} ∈ K ∩ A and a fragment gi ∈ G(e) such that K{i} ⊂ H{i} \K∅ , 0 < gi ≤ g0 , and, for every K 0 ∈ K ∩ A , with K{i} ⊂ K 0 ⊂ H{i} \K∅ , the relations hold:
gi µ(K 0 \K{i} ) ≤ ε0 e, gi µ (H{i} \K∅ )\K{i} ≤ ε0 e. We may assume that gm ≤ gm−1 ≤ · · · ≤ g1 . If i 6= j then K{i} ∩ K{j} = H∅ ; therefore, g0 µ(K{i} ∩ K{j} ) ≤ ε0 e. Assume that for some k ≤ m the above construction is performed for each J ⊂ M , with card J < k. In particular, for J ⊂ M , card J < k, there is KJ ∈ K ∩ A . Let J ⊂ M and card J = k. Take KJ ∈ K ∩ A and gJ ∈ G(e) so that [ ^ KJ ⊂ HJ \ {KJ 0 : J 0 ⊂ J}, 0 < gJ ≤ {gJ 0 : card J 0 < k} S and for every K 0 ∈ K ∩ A with KJ ⊂ K 0 ⊂ HJ \ {KJ 0 : J 0 ⊂ J}, the relations hold [ gJ µ(K 0 \ KJ ) ≤ ε0 e, gJ µ (HJ \ {KJ 0 : J 0 ⊂ J})\KJ ) ≤ ε0 e. All gJ with card J ≤ k can be assumed linearly ordered. Moreover, for every two subsets J ⊂ M and J 0 ⊂ M , J 6= J 0 , we have KJ ∩ KJ 0 ⊂ HJ ∩ HJ 0 = HJ∩J 0 provided that either card J = k and card J 0 ≤ k, or card J ≤ k and card J 0 = k. If 00 J ⊂ J ∩ J 0 then by construction either KJ ∩ KJ 00 = ∅, or KJ 0 ∩ KJ 00 = ∅. This 0 0 0 0 amounts to V the fact that0 g¯ µ(K ) ≤ ε e for every K ∈ K ∩ A , K ⊂ KJ ∩ KJ 0 , with g¯ = {gJ 0 : card J ≤ k}. The construction is finished on the induction step k = m − 1. Thus we obtain the following: Sm σ(i) For any function σ : M → {0, 1} denote W σ = i=1 Wi , where Wi0 = (U \K)\Wi and Wi1 = Wi for i ∈ M . We have proven that for every ε0 > 0 there

252

Chapter 6

exist g¯ ∈ E(e), 0 < g¯ ≤ g, and Kσ ∈ K ∩ A (σ ∈ {0, 1}M ) such that Kσ ⊂ U \K and g¯ µ(W σ ) − µ(Kσ ) ≤ ε0 g¯,

g¯ µ(K 0 ) ≤ ε0 g¯

K 0 ∈ K ∩ A , K 0 ⊂ Kσ ∩ Kσ0 , σ, σ 0 ∈ {0, 1}M , σ 6= σ

(∗)

0

.

This amounts to the inequalities  [  X  [ g¯ µ (K1 ) ≥ g¯ µ Kσ \ Kσ0 + g¯ µ Kσ ∩ Kσ0 σ 0 6=σ

σ

σ6=σ 0

 X X [ ≥ g¯ µ(Kσ ) − g¯ µ Kσ ∪ Kσ0 , σ

σ

σ 0 6=σ

{Kσ : σ ∈ {0, 1}M }. For every σ ∈ {0, 1}M , evaluate the following S  σ σ0 0 term in the right-hand side of the above inequality: g¯ µ . To K ∩ K σ 6=σ

where K1 =

S

this end, denote Mσ := {0, 1}M \{σ} and Lσ0 = Kσ ∩ Kσ0 and consider the identity [  X  X  l
k+1 Lσ0 = µ (−1) µ Lσ1 ∩ · · · ∩ Lσk . σ0

σ1 0. If f := µ 0 (Q) then there are 0 < ε0 ∈ R and 0 < e0 ∈

254

Chapter 6

E(e) with µ 0 (An ) ≥ ε0 e0 (n ∈ N). By Theorem 6.2.2 µ 0 is a quasi-Radon measure, so that we may find K1 ∈ K ∩ A and e1 ∈ G(e0 ) such that K1 ⊂ A1 , 0 < e1 ≤ e0 , and e1 µ 0 (A1 \K1 ) ≤ (ε0 /4)e0 . By induction, we produce sequences (Kn )n∈N ⊂ K ∩ A and (en )n∈N ⊂ G(e) such that 0 ≤ en+1T≤ en , Kn ⊂ An , n and en µ 0 (An \Kn ) ≤ (ε0 /2n+1 )e0 (n ∈ N). Putting Kn0 = l=1 Kl , we arrive at en µ 0 (An \Kn0 ) ≤ (ε0 /2)e0 (n ∈ N). From this it follows that µ 0 (Kn0 ) = µ 0 (An ) − µ 0 (An \Kn0 ) ≥ (ε0 /2)en . But then Kn0 0 = ∅ for some n0 ∈ N, since Kn0 & ∅, This is a contradiction, and so the claim follows. B Similar fact is valid for Radon measures without any additional restrictions, i.e., if µ ∈ da(A , Y ) is a Radon measure then µ is σ-additive. 6.2.6. We now address the question of integral representation of dominated operators by quasi-Radon measures. Let Q be a completely regular topological space and let L be a sublattice of the vector lattice Cb (Q). Denote by T (L) the weakest topology in which all functions in L are continuous. If T (L) coincides with the initial topology T of Q, then we say that L generates T . Theorem. Assume that a vector sublattice L ⊂ Cb (Q) contains the identically one function 1 and generates the topology of Q. Let T : L → Y be a dominated operator. The following are equivalent: (1) there exists a unique quasi-Radon measure µ ∈ dca(Bor(Q), Y ) such that the representation holds Z T f = f dµ (f ∈ L); W V (2) T (1) = { { T g : g ∈ L, g ≥ χK } : K ∈ K }. C Suppose that (1) holds and take a net (fα )α∈A in L decreasing to zero pointwise. Fix some α0 ∈ A and ε > 0. Then for some 0 < M ∈ R we have 0 ≤ fα ≤ V M 1 for all α ≥ α0 . Given a compact set K ⊂ Q, denote aK := T (1) − { T f : f ∈ L, f ≥ χK }. Observe that _ _ aK = { T (1 − g) : g ≥ χK } = { T (h) : h ≤ 1 − χK }. Since the net (fα ) vanishes uniformly on the compact set K, we may choose α1 ≥ α0 so that fα − fα ∧ (ε1) vanishes on K whenever α ≥ α1 . Moreover, hα := M −1 (fα − fα ∧ (ε1)) ≤ 1, whence h ≤ χK for the same α. Thus, T (hα ) ≤ aK and we may evaluate
0 ≤ T (fα ) = T fα ∧ (ε1) + T (M hα ) ≤ ε T (1) + M aK . By Theorem 1.4.6 (1) the band { T (1)}⊥⊥ in F is isomorphic to an order-dense ideal F 0 in C∞ (P ), where P is an extremal compact space. We may assume that

Integral Operators

255

this isomorphism carries the element T (1) into the identically one function on P . By hypothesis the net (aK )K∈K decreases to zero. Therefore, there exists a comeager set P0 ⊂ P such that the numerical net aK (p) converges to zero for all p ∈ P0 . From the above evaluation it follows that if a net (fα )α∈A ⊂ L decreases to zero then for each p ∈ P0 the net ωp (fα ) := ( T fα )(p) converges to zero. The positive functional ωp then can be extended to an order σ-continuous functional ~q : Mb (Q) → R (see 4.5.3). Define an operator V : Mb (Q) → RP as follows: ω  ~p (f ), q ∈ P0 , w (V f )(p) := 0, p∈ / P0 . ~p Clearly, V f ∈ Mb (P ) for every f ∈ L. Moreover, from the order σ-continuity of ω it follows that if V fn ∈ Mb (P ) for a sequence fn ∈ Mb (P ) with f := o-limn fn then V f ∈ Mb (P ). Whence we conclude that V f ∈ Mb (P ) for every f ∈ Mb (Q). Put W = β ◦ V , where β : Mb (P ) → F is defined by 6.2.1 (4). Then W : Mb (P ) → F is an order σ-continuous extension of T . The operator T is extended as follows: Take f ∈ Mb (Q) and a bounded net (gα )α∈A in L that increases to f pointwise. From the estimate T gα − T gβ ≤ W ( gα −gβ ) (α, β ∈ A) it follows that T gα is bo-fundamental. Thus, we may define T0 f = bo-lim T gα . Let SCb denote the cone of all bounded lower semicontinuous functions on P . Then T can be extended to a dominated operator T0 : M0 → Y with M0 = SCb↑ − SCb↑ . Further extension goes on by transfinite induction up to the least uncountable ordinal ω1 . This extension preserves the inequality T0 f ≤ W (|f |) (f ∈ M0 ). Suppose that for every ordinal β < α < ω1 we have defined a vector sublattice Mβ ⊂ Mb (P ) and a linear operator Tβ : Mβ → Y , such that the following relations are valid: Tβ f ≤ W (|f |) (f ∈ Mβ );

Mβ ⊂ Mγ , Tγ |Mβ = Tβ , β < γ < α. S If α is a limit ordinal then we put Mα := {Mβ : β < α} and define a linear operator Tα : Mα → Y by the relation Tα |Mβ = Tβ (β < α). If α is a nonlimit ordinal then  we consider the set Mα−1 (of the least upper bounds of countable bounded subsets of Mα−1 ). Take an increasing sequence (fn )n∈N in Mα−1 with supn fn = f ∈  Mα−1 . Using the above reasoning we may easily check that (Tα−1 fn )n∈N is a bo f := bo-lim Tα−1 fn . By doing fundamental sequence. Hence we may define Tα−1    so we obtain an operator Tα−1 : Mα−1 → Y satisfying the inequality Tα−1 f ≤  W f (0 ≤ f ∈ Mα−1 ). Soundness of this definition follows from sequential order   continuity of W . Let Mα := Mα−1 −Mα−1 and let Tα : Mα → Y be the extension of  Tα−1 by differences. It is easy to see that Mb (Q) = Mω1 and the operator T1 := Tω1 is a sequentially order continuous extension of T to the space Mb (Q). Moreover, T1 f ≤ W (|f |) (f ∈ Mb (Q)).

256

Chapter 6

Now
it is clear that if we define a measure µ by letting µ(C) := T1 (χC ) C ∈ Bor(Q) then the integral representation 6.2.6 (1) holds. It should be proven that µ is a quasi-Radon measure. For every U ∈ T the characteristic function χU is lower semicontinuous and by the above construction _ µ (U ) = T (χU ) = { T f : f ∈ L, 0 ≤ f ≤ χU }. Fix 0 < ε ∈ R, f ≤ χU , f ∈ L, and put D := {x ∈ Q : f (x) ≥ ε}. Then f ≤ χD + ε1,

T f ≤ µ (D) + ε µ (U ).

This amounts to the following quasiregularity condition _ µ (U ) = { µ (D) : D ∈ F , D ⊂ U }. W In a similar way, using 6.2.6 (2), we may deduce the relation µ (Q) = { µ (K) : K ∈ K }. By virtue of 6.2.4 (3) µ and µ are quasi-Radon measures. Conversely, suppose that T admits the integral representation 6.2.6 (1) with a quasi-Radon µ ∈ dca(Bor(Q), Y ). If K ∈ K then _ µ (Q\K) = { µ (K 0 ) : K 0 ∈ K , K 0 ⊂ U \K}. Since for every K 0 ⊂ U \K there is a functionWf ∈ L such that f (K 0 ) = {0}, f (K) = {1}, and 0 ≤ f ≤ 1, we have µ (K) = { T f : f ∈ L, f ≥ χK } and the desired identity 6.2.6 (2) holds. The proof is complete. B 6.2.7. Observe two corollaries to Theorem 6.2.7. Denote by dqa(A , F ) the space of countably additive quasiregular F -valued measures on A . (1) Theorem. Let Q be an arbitrary compact space, and let F be a K-space. Then for every order-bounded linear operator T : C(Q) → F there is a unique order-bounded countably additive quasiregular measure µ : Bor(Q) → F such that Z T (f ) = f dµ (f ∈ C(Q)). Q

The correspondence T 7→ µ is a linear and lattice isomorphism between the vector lattices L∼ (C(Q), F ) and dqa(Bor(Q), F ). C Under the stated conditions, 6.2.2 (2) holds automatically; we may put K := Q, because Q is compact. B Now we give a corollary concerning the measure extension problem. Let C(A ) denote the uniform closure of S(A ), and let T0 stand for the functionally open sets. An algebra A ⊂ P(Q) is called tight, if the following conditions are met: (a) the vector lattice L = C(A ) ∩ Cb (Q) generates the topology T ; (b) for every V ∈ A ∩ T0 there is a function ϕ ∈ C(A ) ∩ Cb (Q) with V = {x ∈ Q : ϕ(x) > 0}.

Integral Operators

257

(2) Theorem. Assume that a quasi-Radon measure µ0 ∈ da(A , Y ) is defined on a tight algebra A with A = σ0 (F0 ∩ A ). Then there exists a unique quasi-Radon measure µ ∈ dca(Bor(Q), Y ) extending µ0 . C Consider the dominated R operator T : L → Y defined on the vector lattice L = C(A )∩Cb (Q) by T f = f dµ0 (f ∈ L). The operator T meets 6.2.2 (2), since L separates compact subsets in Q. Thus, there is a unique quasi-Radon measure µ ∈ dca(Bor(Q), Y ) such that the integral representation 6.2.2 (1) holds. To ensure that µ is an extension of µ0 it is sufficient to justify the identity µ0 (U ) = µ(U ) provided that U ∈ T0 ∩A . Since A is tight, there is a function ϕ ∈ L with U = {x ∈ Q : ϕ(x) > 0}. Put ϕn = (nϕ) ∧ 1 (n ∈ N) andR observe that ϕn % R χU . Using σadditivity of µ and µ0 , we obtain µ0 (U ) = bo-lim ϕn dµ0 = bo-lim ϕn dµ = µ(U ), which completes the proof. B (3) Let Q be a σ-compact topological space and let A ⊂ P(Q) be an algebra with A = σ0 (F0 ∩ A ). Then every measure µ0 ∈ dca(A , Y ) has a unique extension to a quasi-Radon measure µ ∈ dca(Bor(Q), Y ). 6.3. Functional Representation of Maharam’s Extension The Maharam extension of a positive operator presented in Section 3.5 result in a vector lattice L1 (ˆ) on which the operator is well behaved. But the space L1 (ˆ) has a complicated structure which is troublesome in applications. In this section we give an explicit description for L1 (ˆ) in terms of measurable functions. 6.3.1. Let P be a σ-compact topological space and let E0 := C0 (P ) be the vector lattice of compactly-supported continuous functions on P , i.e., C0 (P ) := {f ∈ C(P, R) : supp(f ) is a compact space}. In this section A is a nonempty set, A is a σ-algebra of its subsets, and N is a σ-ideal in A . Let M (A, A , N ) be the space of cosets of measurable functions on A as defined in 1.4.7 (1). We will suppose that the measurable space (A , N ) is of countable type; i.e., an arbitrary family (Aα ) ⊂ A \N with Aα ∩ Aβ ∈ N (α 6= β) is at most countable. In this event M (A, A , N ) is an order complete vector lattice. In this section F is an orderdense ideal in M (A, A , N ). A sequence (An ) ⊂ A of pairwise disjoint sets is called a partition of a measurable set A0 ∈ A if χA0 = sup χAn in F , where χC always stands for the characteristic function of C. Denote by V the vector of functions of two variables f : A × P → R Plattice ∞ representable as v(s, t) = o- n=1 χAn (s)en (t), where (An ) is a partition of A, (en ) is an order-bounded sequence in E0 , and the infinite sum is understood for every s ∈ P in the sense of order convergence in the vector lattice M (A, A , A0 ), i.e. almost everywhere on A. Extend the operator ˆ : E0 → F to the vector lattice P∞ V by letting ˆv = n=1 χAn ˆen . This definition is sound. Indeed, if v(s, t) =

258 P∞

Chapter 6

n=1 χBn (s)dn (t) for another P∞partition (Bn ) of A and sequence (dn ) ⊂ E0 , then the representation v(s, t) = n,k=1 χCn,k (s)gn,k (t) is also true with Cn,k := An ∩Bk and gn,k := en = dk , whenever Cn,k 6= ∅. Therefore,

o-

∞ X

χAn ˆen = o-

n=1 ∞ X

= o-

k=1

∞ X

n=1

o-

∞ X

n=1

o-

∞ X

χCn,k ˆgn,k

k=1

χCn,k ˆgn,k = o-

∞ X

χBk ˆek .

k=1

6.3.2. If a sequence (vn ) in the vector lattice V decreases and (vn (s, t)) converges to zero for all (s, t) ∈ A × P , then inf ˆvn = 0. C Suppose that a sequence (vn ) meets the hypotheses, and still inf ˆvn > εχA P∞ for some 0 < ε ∈ R and C ∈ A \N . Let vn = k=1 χAn,k en,k . Consider the e1,k ≤ e for all k ∈ N. sequence vn0 := χC vn and a function e ∈ (E0 )+ , withP ∞ 0 0 Evidently, each function vn can be represented as vn = k=1 χCn,k en,k + χAn ϕn , where for every n ∈ N the sequence (Cn,k )k∈N ⊂ A \N is a partition S∞ of C, the function ϕn : A×P → R is bounded above by χA ⊗e, and Cn := C\ k=1 Cn,k ∈ N . Moreover, the partitions of C can be chosen so that (Cn+1,k )∞ k=1 is a refinement of (Cn,k )∞ . Take any element s ∈ C\ ∪ C . There exist a decreasing sequence n k=1 T∞ Cn,kn (s) such that s ∈ n=1 An,kn (s) . Then for the sequence (en,kn (s) ) in C0 (P ) we have en,kn (s) (t) = vn (s, t) ≥ vn+1 (s, t) = en+1,kn+1 (s) (t) (t ∈ P ). Since limn vn (s, t) = 0, it follows that (en,kn (s) ) vanishes uniformly on the compact set supp(e). Thus, for every m ∈ N there is n ∈ N such that χCn,kn (s) en,kn (s) ≤ (ε/m) · e ∧ 1 ≤ (ε/m)g, where g is a positive compactly-supported function coinciding with unity on supp(e). From this we deduce εχCn,kn (s) = χCn,kn (s) (εχC ) ≤ εχCn,kn (s) ˆvn = εχCn,kn (s) ˆvn0 = εχCn,kn (s) ˆen,kn (s) ≤ (ε/m)ˆg, whence ε ≤ (ε/m)(‰g)(s) → 0. This is a contradiction, and the claim follows. B 6.3.3. A set D ⊂ A × P is said to be negligible or, more precisely, ˆ-negligible if for every 0 < ε ∈ R there exist an increasing sequence of positive functions (vn ) in V such that supn vn (s, t) ≥ 1 ((s, t) ∈ D) and ˆvn ≤ ε1A almost everywhere on A.

Integral Operators

259

(1) The countable union of negligible sets is negligible. C Suppose that D is the union of some sequence of negligible sets Dn ⊂ A × P . Take 0 < ε ∈ R, n ∈ N and let (vn,k )k∈N be an increasing sequence in V+ such that supk (vn,k (s, t) ≥ 1 for all (s, t) ∈ Dn and ˆ(vn,k ) ≤ ε/2n . If vk := v1,k ∨ · · · ∨ vk,k then supk∈N vk (s, t) ≥ 1 for all (s, t) ∈ D and ˆ(vk ) ≤

k X l=1

vl,k

k X ε = ε1A , ≤ 2l l=1

whence the claim is clear. B (2) If a decreasing sequence (vn ) ⊂ V vanishes almost everywhere on A × P then inf ˆvn = 0. C Since v1 is a bounded function, β := sup{v1 (s, t) : (s, t) ∈ A × P } < ∞. If β = 0 then there is nothing to prove, so that we assume β > 0. Let D ⊂ A × P be the negligible set on which the sequence (vn ) does not vanish. For a fixed ε > 0 choose an increasing sequence (wn ) in V such that supn wn (s, t) ≥ 1 for (s, t) ∈ D and ˆ(wn ) ≤ ε/β (n ∈ N). The sequence (un )n∈N with un := (vn − βwn )+ is decreasing and limn un = 0 everywhere on A × P . By 6.3.2 o-limn ˆ(un ) = 0, whence o-lim ˆ(vn ) − βo-lim ˆ(wn ) = o-lim ˆ(vn − βwn ) ≤ o-lim ˆ(un ) = 0, n→∞

n→∞

n→∞

so that 0 ≤ o-lim ˆ(vn ) ≤ βo-lim ˆ(wn ) ≤ β n→∞

n→∞

n→∞

ε = ε. B β

6.3.4. Now we will extend the “preintegral” ˆ : V → F to some larger function space. Unfortunately, we cannot apply the Extension Theorem 4.5.3 since F is not weakly σ-distributive. Nevertheless, the Daniell construction of the Lebesgue integral may be successfully carried out due to the specific properties of our “preintegral” ˆ. Consider the set V  comprising the extended real-valued functions f : A × P → R ∪ {±∞} that are representable as almost everywhere limits of increasing sequences (vn ) ⊂ V with supn ˆ(vn ) ∈ F . (1) Every function f ∈ V  takes finite values almost everywhere. C Indeed, take a sequence (vn ) ⊂ V+ with vn % f and sup ˆvn =: g ∈ F . Put D := {f = +∞} := {(s, t) ∈ A × P : f (s, t) = +∞}, C0 := {g = +∞}, and Cm := {m − 1 ≤ g < m} := {s ∈ A : m − 1 ≤ g(s) < m} (m ∈ N). Clearly, {f = +∞} ⊂ A × P , Cm ∈ A (m = 0, 1, . . . ) and (Cm )m∈N is a partition of A. It suffices to prove that Dk := D ∩ (Ck × P ) is a negligible set for every k ∈ N. Take an arbitrary ε > 0 and put wn = kε χCk vn . It is easy to see that

260

Chapter 6

wn ∈ V and ˆ(wn ) ≤ kε χCk g ≤ ε1A . Moreover, for every d = (s, t) ∈ Dk there is a number n ∈ N, such that vn (d) > kε (since vn (d) % +∞). From this we deduce wn (d) = kε χCk (t)vn (d) ≥ 1, so that sup wn (d) ≥ 1 for each d ∈ Dk . B For every f ∈ V  define ˆf := sup ˆvn , where (vn ) ⊂ V and vn % f . This soundly defines some mapping V  → F that we will denote by the same symbol. It is an easy matter to verify that V  is closed under lattice operations, addition, and multiplication by positive scalars. Moreover the following is true: (2) If the sequence (fn ) ⊂ V  increases almost everywhere to some f : A × P → R ∪ {±∞} and ˆ(fn ) is bounded above in F then f ∈ V  and limn ˆ(fn ) = ˆ(f ). C If for every n ∈ N the sequence (vn,k )k∈N in V increases and converges almost everywhere to fn , then under the stated conditions the sequence (vn )n∈N of the functions vn := v1,n ∨ · · · ∨ vn,n converges almost everywhere to f and ˆ(vn ) is bounded above, since vn ≤ fn . Thus, f ∈ V  . Since ˆ(vn ) ≤ ˆ(fn ) ≤ ˆ(f ), we have ˆ(f ) = supn ˆ(fn ). B 6.3.5. A function f = A × P → R ∪ {±∞} is said to be ˆ-summable or ˆ-integrable if ˆ∗ f := inf{ˆh : h ≥ f, h ∈ V  } = sup{−ˆg : −g ≤ f, g ∈ V  } =: ˆ∗ f. b := ˆ∗ f = ˆ∗ f . Denote by L 1 (A × P, ˆ) the set of all In this event we set ˆf ˆ-summable functions. Evidently, L 1 (A × P, ˆ) ⊃ V  ⊃ V . A function f = A × P → R ∪ {±∞} is ˆ-summable if and only if for every 0 < ε ∈ R there exist h ∈ V  and g ∈ −V  such that g ≤ f ≤ h and ˆ(h−g) ≤ ε1A . C Since F has the countable sup property, we may find a decreasing sequence (hn ) ⊂ V  and an increasing sequence (gn ) ⊂ −V  with gn ≤ f ≤ hn (n ∈ N) and ˆf = inf ˆhn = sup ˆgn . The last identities imply that there exist a partition (Cn ) ⊂ A of A such that χCn (ˆhn −ˆf ) ≤ (ε/2)1A and χCn (ˆf −ˆgn ) ≤ (ε/2)1A . Put ∞ ∞ X X h= χCn hn , g = χCn gn . n=1

n=1

Clearly, −g, h ∈ V  , g ≤ f ≤ h, and ˆh − ˆg ≤ ε1A . B Using this proposition, it is easy to verify that L 1 (ˆ) is an order σ-complete b : L 1 (A × P, ˆ) → F is an order σ-continuous operator. Morevector lattice and ˆ over, if (fn ) ⊂ L 1 (A × P, ˆ) is an increasing sequence with sup ˆfn ∈ F then sup fn ∈ L 1 (A × P, ˆ). A stronger assertion will be proved in the next subsection. b 1 − f2 ) = 0 in L 1 (A × P, ˆ) Introduce the equivalence relation f1 ∼ f2 ⇔ ˆ(f and denote by L1 (A × P, ˆ) the factor vector lattice L 1 (A × P, ˆ)/ ∼. Let  be

Integral Operators

261

the natural embedding of E0 into L 1 (A × P, ˆ); more precisely, e is the coset of the function 1A ⊗ e : (s, t) 7→ e(t) (s, t) ∈ A × P . Denote by φ the factor mapping which sends every ˆ-summable function f to its coset f~. There is a unique operator b = ‰ ◦ φ. We will also denote ˆ b := ‰. ‰ : L1 (A × P, ˆ) → F satisfying ˆ Recall that (see 3.5.2, 3.5.3) there exist an order complete Banach–Kantorovich e : L1 (ˆ) → F , and an lattice L1 (ˆ), an essentially positive Maharam operator ˆ e ◦ ı = ˆ. isomorphic embedding ı : E0 → L1 (ˆ) such that ι(X)⊥⊥ = L1 (ˆ) and ˆ 1 Moreover, L (ˆ) admits the structure of a lattice-ordered module over ƒ := Orth(F ) e is a module homomorphism, i.e. a ƒ-linear operator. such that ˆ 6.3.6. Theorem. The following hold: (1) L1 (A × P, ˆ) is an order complete vector lattice; b : L1 (A × P, ˆ) → F is an essentially positive Maharam operator; (2) ˆ (3) there exists a unique lattice isomorphism β from L1 (A × P, ˆ) onto e ◦ β = ˆ. b L1 (ˆ) such that β ◦  = ı and ˆ P∞ C It suffices to verify (3). Take an arbitrary v ∈ V , v = k=1 χCn en , where (Cn ) ⊂ A is a partition of A and (en ) ⊂ E0 is an order-bounded sequence. Let πn be the projection P∞ in F defined as multiplication by the characteristic function χCn . ~ Put β(v) := k=1 πk ı(ek ). If (vn ) ⊂ V , vn % h, and supn ˆ(vn ) ∈ F , then we set ~ := sup β(v ~ n ). Finally, if (hn ) ⊂ V  and hn & f , f ∈ L 1 (A × P, ˆ), then we β(h) n ~ ~ n ). Observe that β~ is a lattice homomorphism which maps define β(f ) := inf n β(h the lattice V onto M0 , the cone V  onto M0 , and the vector lattice L 1 (A × P, ˆ) e and ˆ b that onto the K-space L1 (ˆ) = M0 . It follows from the definitions of ˆ ~ 1 − f2 |); e ◦ β~ = ˆ. b Consequently, if f1 ∼ f2 then 0 = ˆ(|f b 1 − f2 |) = ˆ( e β|f ˆ ~ 1 − f2 |) = 0. Thus, we may define β on L1 (A × P, ˆ) by the equation therefore, β(|f ~ ), where f~ is the coset of a function f . Now, it is clear that ˆ e◦β = ˆ b β(f~) = β(f and β is the desired isomorphism. It remains to establish uniqueness of β. Suppose that there is an isomorphism 0 b = ˆ e ◦ β 0 . In order to verify β : L1 (A × P, ˆ) → L1 (ˆ) with β 0 ◦  = ı and ˆ 0 0 the identity β = β it suffices to check that β (χC e) = πC ı(e), where e ∈ (E0 )+ , C ∈ A , and πC ∈ B(F ) is the projection corresponding to χC . Observe that e ⊥ β 0 (χC e)) = π ⊥ ˆβ e 0 (χC e) = χA\C χC ˆe = 0. Hence, β 0 (χC e) = πC β 0 (χC e) ≤ ˆ(π C C e 0 (χC e) = χC ˆe = πC ˆe = ˆ(π e C ı(e). From πC β 0 (e) = πC ı(e). At the same time ˆβ e is essentially positive. B this we derive β 0 (χC e) = πC ı(e), since ˆ 6.3.7. A set D ⊂ A × P is called ˆ-summable if χD ∈ L1 (A × P, ˆ). A set is ˆ-measurable if it is representable as the union of a countable family of ˆ-summable sets.

262

Chapter 6

(1) If B ⊂ P is a Baire set and C ∈ A then the rectangle D := B × C is ˆ-measurable. C First we observe that every rectangle of the form D := A × {e >
1} with e ∈ (E0 )+ is ˆ-summable. Indeed, χD = χA ⊗supn e∧(1+1/n)1P −e∧1 /n ∈ V  . Now, the standard measure theoretic arguments show that A × B is measurable for every Baire set B ⊂ P . Since P is a σ-compact space, the rectangle C × P is also measurable for all C ∈ A . It remains to note that C × B = (C × P ) ∩ (A × B). B Denote by A ⊗ B the σ-algebra generated by the rectangles C × B where B ⊂ P is an arbitrary Baire set and C ∈ A . It follows from (1) that A ⊗ B consists of measurable sets. (2) For every ˆ-measurable set D ⊂ A×P there exists a set D0 ∈ A ⊗B such that the symmetric difference D4D0 is negligible.
e C Denote by D the subalgebra in the Boolean algebra of unit elements E L1 (ˆ) which is generated by cosets of the form φ(B × C). Take x ∈ C0 (P ) and let e is disjoint from (x) if and only if C := {p ∈ P : x(p) > 0}. A coset φ(v) ∈ L1 (ˆ) v may be chosen disjoint from the function (s, t) 7→ x(s) (t ∈ P ). Therefore, (x) and φ(v) are disjoint if and only if so are φ(χC ) and φ(v). This means that (x) e and φ(χC ) generates the same band in L1 (ˆ).
e = D ↓↑ . Since the vector lattice F Now, it follows from 3.5.10 (1) that E L1 (ˆ)
e = D  is also true. Taking has the countable chain property, the formula E L1 (ˆ)
e , we conclude that there exists into account the identity φ(A ⊗ B) = E L1 (ˆ) ^ ◦ φ(χD4D0 ) = 0, some D0 ∈ A ⊗ B with φ(D) = φ(D0 ), or φ(D4D0 ) = 0. Hence, ˆ 0 whence D4D is negligible. B 6.3.8. Denote by L 0 (A × P, ˆ) the space of almost everywhere finite measurable functions on A × P with respect to the σ-algebra A ⊗ B. Introduce the factor space L0 (A×P, ˆ) := L 0 (A×P, ˆ)/ ∼, where ∼ is the equivalence relation in 6.3.5. Clearly, L 0 (A×P, ˆ) is a Kσ -space. Moreover, L1 (A×P, ˆ) is an order-dense ideal in L0 (A × P, ˆ) and the latter is the universal completion of the former. Theorem. For any operator S ∈ {ˆ}⊥⊥ there exists a unique (up to equivab lence) ˆ-summable function KS ∈ L 1 (A × P, ˆ) such that Se = ˆ(φ(K S )e) (e ∈ E0 ). Moreover, the correspondence S 7→ KS induces a linear and order isomorphism from {ˆ}⊥⊥ onto L1 (A × P, ˆ). C Observe that for an arbitrary u ∈ L0 (A × P, ˆ) we have
u ·  C0 (P ) ⊂ L1 (A × P, ˆ) ⇔ u ∈ L1 (A × P, ˆ) Then the claim follows from 6.3.6, 3.4.5, and 3.5.5 (1). B

Integral Operators

263

6.3.9. Assume now that P is a compact space, so that a weak order-unity 1 := φ(1A×P ) belongs to L1 (A × P, ˆ). Then we are able to define an F -valued b measure ϕ : A ⊗ B → F by letting ϕ(D) := ˆ(D). According to 6.1.7 ϕ is a unique 1 1 σ-additive measure with L (A × P, ˆ) = L (ϕ) and Z b ˆ(φf )=

f dϕ (f ∈ L 1 (ϕ)).

A×P

~ If π is a band projection in F then there is C ∈ A with π f~ = χg C f , f ∈ F . Define g is the coset some band projection h(π) in L1 (ϕ) as follows: if g ∈ L 1 (ϕ) then h(π)~ of the function (s, t) 7→ χC (s)g(s, t). It is clear that h is a Boolean homomorphism from P(F ) into P(L1 (ϕ)). (1) Under the above assumptions ϕ is ample with respect to h. C Follows from 6.3.6 and 6.1.10. B Summing up (1) and 6.3.6, we obtain the following: (2) Let ˆ : C(P ) → F be a positive operator. Then there exists a unique ample measure ϕ : A ⊗ B → F such that ZZ 1A ⊗ g dϕ (g ∈ C(P )).

ˆg = A×P

(3) For every order-bounded operator S ∈ {ˆ}⊥⊥ there exists a unique (up to ϕ-equivalence) ϕ-integrable function KS ∈ L 1 (ϕ) such that ZZ Sg =

KS 1A ⊗ g dϕ (g ∈ C(P )).

A×P

(4) The correspondence S 7→ φ(KS ) is a linear and lattice isomorphism from {ˆ}⊥⊥ onto L1 (ϕ). 6.3.10. Let µ be a regular Borel measure on a compact space P , and let L0 (P, µ) be the vector lattice of cosets of real µ-measurable functions on P . Assume that E is an order-dense ideal in L0 (P, µ) containing the identically one function 1P . (1) If ˆ admits an order continuous extension to E then every µnegligible set is ˆ-negligible. C Take a µ-negligible set Z R⊂ P . For every k ∈ N choose an increasing sequence (fn,k )n∈N in C(P )+ such that P fn,k dµ ≤ 1/k (n, k ∈ N) and gk (t) ≥ 1 (t ∈ Z),

264

Chapter 6

where gk (t) = supn fn,k (t) (t ∈ P ). Substituting, if necessary, fn,k ∧ fn,k+1 for fn,k+1 we may assume that fn,k ≥ fn,k+1 . By the Monotone Convergence Theorem Z Z 1 gk dµ = lim fn,k ≤ . n→∞ k P

P

R Thus (gk )k∈N is a decreasing sequence of µ-summable functions and P gk dµ → 0 as k → ∞, so that (gk ) converges to zero almost everywhere. Since ˆ is order continuous on E and almost everywhere convergence in E is order convergence, we have inf k∈N ˆ(gk ) = 0. From this we deduce that for every ε > 0 there is a countable partition (Ck )k∈N of the set A such that ˆ(gk )(s) ≤ 1 for all s ∈ Ck . Now define vn ∈ V by vn :=

∞ X

χCk ⊗ fn,k ,

χCk ⊗ fn,k : (s, t) 7→ fn,k (t)χCk (s)

(t ∈ P, s ∈ A).

k=1

Using the definition in 6.3.1 we infer ˆ(vn ) =

∞ X k=1

χCk ˆ(fn,k ) ≤

∞ X

χCk ˆ(gk ) ≤ ε1A .

k=1

If s ∈ Cm then supn vn (s, t) = supn fn,m (t) = gm (t) ≥ 1 for t ∈ Z. Since every s is contained in some Ck , we obtain that supn vn (s, t) ≥ 1 for all t ∈ Z and s ∈ A. B (2) If ˆ admits an order continuous extension to E then the function g¯ := 1A ⊗ g : (s, t) 7→ g(t)1P (s) = g(t) is ˆ-negligible for every µ-negligible function g. C If g is µ-negligible then there is a decreasing sequence (gn ) of lower semiR continuous functions with |g| ≤ gn ≤ 1P (n ∈ N) and P gn dµ −−−→ 0. Hence, n→∞

(gn ) converges to zero µ-almost everywhere. Moreover, (gn ) is contained in E and o-converges to zero there. By assumption, (ˆ(gn )) o-converges to zero. It follows b g¯n ). that ˆ∗ (¯ g ) = 0 provided that g¯n ∈ V  and ˆ(gn ) = ˆ( Take an increasing net of continuous functions (fα ) ⊂ C(P ) and put f (t) = supα fα (t) (t ∈ P ). Without loss of generality we may set f ≤ 1P . Then there is an increasing sequence (α(n))n∈N such that f (t) = supn fα(n) (t) µ-almost everywhere. In other words, the sequence (fα(n) ) o-converges to f . Again by order continuity,
ˆ(f ) = sup ˆ(fα(n) ) = ˆ sup fα(n) ≤ ˆ(1P ) ∈ F. b f¯n ) → ˆ(¯ b g ) by definition. Moreover, according to (1) At the same time ˆ(fn ) = ˆ( the identity f (t) = supn fα(n) is valid ˆ-almost everywhere. B

Integral Operators

265

6.3.11. Theorem. For an arbitrary order continuous positive operator T : E → F there exists a unique ample measure ϕ : A ⊗ B → F such that ZZ e= e ∈ E). T~ 1A ⊗ e(s, t) dϕ(s, t) (~ Moreover, for every order-bounded operator S ∈ {T }⊥⊥ there is a unique (up to ϕ-equivalence) ϕ-measurable function K := KS such that ZZ S~ e= K(s, t)1A ⊗ e(s, t) dϕ(s, t) (~ e ∈ E). A×P

The correspondence S 7→ KS defines a linear and lattice isomorphism from {T }⊥⊥ onto the order-dense ideal Lϕ (E) := {g ∈ L0 (A × P, ϕ) : g · (E) ⊂ L1 (A × P, ϕ)} in L0 (A × P, ϕ). C Let ˆ be the restriction of T to the lattice C(P ). Apply 6.3.9 to ˆ and find an ample measure ϕ : B ⊗ A → F with the representation 6.3.9 (2) valid. According to 6.3.10 (2) the integral on the right-hand side of 6.3.9 (2) is correctly defined for g ∈ L 0 (P, µ) with ~ g ∈ E. By o-continuity we thus have ZZ bg = ˆ~ 1A ⊗ g dµ (~ g ∈ E), A×P

g denotes the µ-equivalence class of g. where ~ For every K ∈ Lϕ (E) the formula ZZ g) = SK (~ K(s, t)1A ⊗ e(s, t) dϕ(s, t)

g ∈ E) (~

A×P

correctly defines some operator SK ∈ {T }⊥⊥ . Moreover, the mapping β : K 7→ SK is a linear and order isomorphism from Lϕ (E) into {T }⊥⊥ . Show that this mapping is surjective. By virtue of the Freudenthal Spectral Theorem it is sufficient to prove that β covers the Boolean algebra of unit elementsP E(T ). Let K denote the set n of all functions K ∈ Lϕ (E) representable as K = k=1 χCk ⊗ χBk for arbitrary Bn ∈ B and disjoint Cn ∈ A . Given such a K, we may check by simple calculations Pn that β(K) = k=1 [χCk ] ◦ T ◦ [χBk ]. Hence β(K ) covers A (T ) (see 3.5.9 for the definition) and E(T ) = (β(K ))↑↓↑ by 3.5.11 (3). Denote K1 := {g ∈ L 0 (P ⊗ A) : 0 ≤ g ≤ 1A×P }. It remains to prove that β(K1 ) ⊃ (β(K ))↑ and β(K1 ) ⊃ (β(K ))↓ , since in this event β(K1 ) covers E(T ). We confine exposition to the former relation; the latter is proved in a similar way.

266

Chapter 6

Take an increasing net (β(Kα )) with Kα ∈ K1 and S := supα β(Kα ). Denote ~α , where K ~α is the coset of Kα and the supremum is taken in L0 (A × L := supα K P, ϕ). Since the vector lattice L0 (A × P, ϕ) satisfies the countable chain condition, ~α(n) . Put K(s, t) := there is a sequence of indexes (α(n))n∈N such that L = supn K ~ and β(K) = supn Kα(n) (s, t) (s ∈ A, t ∈ P ). Observe that K ∈ K1 , L = K ~α ≤ K ~ supn β(Kα(n) ) by the Monotone Convergence Theorem. At the same time K implies β(Kα ) ≤ β(K), so that S ≤ β(K) ≤ S. Hence, S ∈ β(K1 ). B 6.3.12. Suppose now that P is a σ-compactSspace. Then there is an increasing sequence of compact sets Pn ⊂ P with P = n∈N Pn . Let Bn and πn denote respectively the Baire σ-algebra of Pn and the band projection in F corresponding to multiplication by χPn . For every n ∈ N, in view of 6.3.9, there is a unique ample measure ϕn : A ⊗ Bn → F such that ZZ b πn ˆ(φg) = o-lim g(s, t) dϕn (s, t) (g ∈ L 1 (ϕ)). n→∞ A×P

Thus, all that was said in 6.3.10 and 6.3.11 can be carried out in the general case of a σ-compact space P . Of course, we may define the measure ϕ(D) :=

∞ _

ϕn (D ∩ A × Pn )

(D ∈ A × B)

n=1

taking values in F ∪{+∞}. A little should be added to 6.1 for developing the theory of F ∪ {+∞}-valued measures along the same lines. We pass over these possibilities and proceed to integral representation by means of real-valued measures. 6.4. Integral Operators In the present section, we study the question of integral representation for a dominated operator between spaces of measurable vector-functions. Our approach grounds on the technique of dominated operators. 6.4.1. Throughout this section (A, A , λ) and (B, B, µ) are σ-finite measure spaces; (Š, †, ν) = (A×B, A ⊗B, λ⊗µ) is their product; and E and F denote some ideal spaces over (B, B, µ) and (A, A , λ), respectively. An operator S : E → F is ¯ such that for called integral if there exists a measurable function K : A × B → R every x ∈ E the value y = Sx is the function Z y(s) = K(s, t) x(t) dµ(t). B

The integral is understood to be the usual Lebesgue integral. The definition presumes the following two conditions to be satisfied:

Integral Operators

i.e.,

267

(1) For every x ∈ E the integrand is summable for almost all s ∈ A; Z |K(s, t) x(t)| dµ(t) < ∞

for almost all s ∈ A. (2) For every x ∈ E the function y(·) belongs to the space F . The function K(s, t) is referred to as the kernel of the integral operator S. In addition, we say that S admits integral representation or S is an integral operator with kernel K and write Z Sx(s) = K(s, t) x(t) dµ(t) (x ∈ E). B

The set of all integral operators from E to F is denoted by I (E, F ). (3) Let S : E → L0 be an integral operator. If xn → 0 in measure and |xn | ≤ x ∈ E (n ∈ N) then Sxn → 0 almost everywhere. C This is an obvious corollary to the Lebesgue Dominated Convergence Theorem by virtue of (1). B Alongside with S we consider the integral operator with kernel |K(·, ·)|: Z ¯ (Sx)(s) = |K(s, t)| x(t) dµ(t). B

¯ is a function defined for all x ∈ E and finite almost everywhere, i.e., By (1), Sx ¯ ∈ L0 (A, A , λ). Thus, the operator S¯ acts always from E into L0 . But S¯ may Sx fail to act from E into F . In this connection, we give the following definition: An operator S is called a regular integral operator from E into F if the operator S¯ with kernel |K(s, t)| acts from E into F . It is evident that if S¯ acts from E to F then S acts from E to F . The converse is true only for regular integral operators. Property (1) shows that every integral operator is a regular integral operator if considered as acting from E into L0 (A, λ). The set of all regular integral operators is denoted by I ∼ (E, F ). 6.4.2. Theorem. An integral operator S is a regular integral operator from E into F if and only if S is o-bounded from E into F . Moreover, if K(·, ·) is the kernel of S then the modulus |S| (in the sense of 3.1.2) is also an integral operator with kernel |K(·, ·)|, i.e., Z (|S|x)(s) = |K(s, t)| x(t) dµ(t), x ∈ E.

268

Chapter 6

C We confine exposition to the case of separable measure spaces. The general case is settled in [162]. It suffices to establish that if an integral operator S acts from E into L0 then the desired integral representation for |S| holds for all x ∈ E, x ≥ 0. Fix such a function x and consider the set M = {y : |y| ≤ x} involved in 3.1.2 (5) for calculating the modulus. Formula 3.1.2 (5) asserts that |S|x = sup{Sy : y ∈ M } =: f ; but the supremum on the right-hand side is that in L0 (A, A , λ). By virtue of the separability of L0 (A, A , λ), there exists a countable everywhere dense set (yn )∞ n=1 in M . For an arbitrary element y ∈ M there is a sequence ynk such that ynk → y in measure. Since the set M is o-bounded in E, we have Synk → Sy almost everywhere by 6.4.1 (3). Therefore, Sy ≤ sup Synk ≤ sup Syn , k∈N

which immediately implies sup{Sy : y ∈ M } = sup{Syn : n ∈ N}. Putting y := ys in the last identity with ys (t) := sign (K(s, t)) x(t), we obtain Z

Z

|K(s, t)| x(t) dµ(t) = K(s, t) ys (t) dµ(t) = (|Sx|)(s) Z  = sup K(s, t) yn (t) dµ(t) : n ∈ N almost everywhere, since for every s ∈ A the inequality |ys | ≤ x is satisfied. B 6.4.3. Define the dual space E 0 (see 3.4.7) as   Z 0 0 E := y ∈ L (B, µ) : |xy| dµ < ∞ (∀x ∈ E) . It is clear that the dual space is an ideal space. (If E = Lp (1 ≤ p ≤ ∞) then 0 p ), if 1 < p < ∞; p0 := ∞, if p = 1; p0 := 1, if p = ∞.) But it E 0 = Lp (p0 := p−1 may happen that E 0 = {0}: for instance, if E = L0 ([0, 1], dx) or E = Lp ([0, 1], dx), 0 < p < 1.) Given x0 ∈ E 0 , we may construct some linear functional ϕx0 on E by the formula Z ϕx0 (x) = x(t)x0 (t) dµ(t) (x ∈ E).

Integral Operators

269

By the Lebesgue Dominated Convergence Theorem it is obvious that ϕx0 ∈ En∼ . It is an easy consequence of 3.4.8 that the mapping x0 7→ ϕx0 is an order and linear isomorphism of the K-spaces E 0 and En∼ . Thus, En∼ is exactly the set of those functionals on E which admit integral representation. Fix x0 ∈ E 0 and y ∈ F . Denote by x0 ⊗ y the rank-one operator Z

0



0

(x ⊗ y)(x) =

x(t) x (t) dµ(t) y

(x ∈ E),

that obviously belongs to L∼ n (E, F ) and is a regular integral operator with ker0 nel K(s, t) = y(s) x (t). Recall that J(E, F ) denotes the band in the K-space 0 0 0 L∼ n (E, F ) which is generated by all operators of the form x ⊗ y (x ∈ E , y ∈ F ) (see 4.3.9). (1) Theorem. I ∼ (E, F ) = J(E, F ). C The proof is contained in 6.4.4 (1–3). B Before launching into proof, observe two corollaries to the above theorem. (2) The set of regular integral operators from E into F is a band. (3) An operator S : E → L0 is an integral operator if and only if there exists an integral operator V ≥ 0 such that |S| ≤ V . 6.4.4. (1) Every continuous linear operator S : L1 (B, µ) → L∞ (A, λ) is an integral operator. C Let G be the set of functions of the form L(s, t) =

n X

rk χAk (s) χBk (t)

k=1

where r1 , . . . , rn ∈ R, A1 , . . . , An ∈ A , B1 , . . . , Bn ∈ B with µ(Bk ), λ(Ak ) < ∞, and Bk ∩ Bl = ∅ (k 6= l). Put ϕ(L) =

n X k=1

Z rk

S(χBk ) dλ. Ak

It is easy to see that the linear functional ϕ is defined on the linear subspace G ⊂ L1 (Š, ν) correctly. Then |ϕ(L)| ≤ kSk

n X k=1

rk λ(Ak ) µ(Bk ) = kSk kLkL1 (Š,ν) ,

270

Chapter 6

whence ϕ is continuous on G. Since G is dense in L1 (Š, ν) in norm, we may extend ϕ by continuity to a bounded functional over L1 (Š, ν) for which we preserve the previous notation. Then there exists a function K ∈ L∞ (Š, ν) such that Z ϕ(L) =

L(s, t) K(s, t) dν (s, t)


L ∈ L1 (Š) .

Š

Taking the definition of ϕ into account, we conclude that Sx admits some integral representation from 6.4.1 with kernel K for finite functions x ∈ L1 (B, µ). Since K ∈ L∞ (Š, ν), we infer that the representation can be extended by continuity to the whole L1 (B, µ). B Without loss of generality we may assume that 1B ∈ E 0 and 1A ∈ F or, what is the same, L∞ (A, λ) ⊂ F and L∞ (B, µ) ⊂ E 0 . The general case reduces easily to this case by decomposition of measure spaces. (2) Let Vn be a (rank-one) integral operator with kernel n1Š = n1A×B and 0 ≤ S ≤ Vn . Then S is an integral operator. C It is obvious that Z kSxkL∞ ≤ kVn (|x|)kL∞ = n |x(t)| dµ(t) = nkxkL1 for all x ∈ L1 (B, µ); therefore, we may extend S to a continuous linear operator from L1 (B, µ) into L∞ (A, λ) and apply (1). B (3) For an arbitrary 0 ≤ S ∈ J(E, F ) the sequence (S∧Vn )n∈N increases and o-converges to S. 0 C First we prove the sought relation for S = x0 ⊗y, where x0 ∈ E+ and y ∈ F+ . It is clear that Z (S ∧ Vn )(x)(s) = x(t) (x0 (t) y(s) ∧ n1A×B (s, t)) dµ(t) (this follows, for example, from Theorem 6.4.2, but
can be checked straightforwardly). Since the sequence x0 (t) y(s) ∧ n1A×B (s, t) n∈N increases and converges pointwise to x0 (t) y(s), we need only to use Beppo Levy’s Theorem and the formula (supn S ∧ Vn )x = sup{(S ∧ Vn )x : n ∈ N}. We now show that {V1 }⊥⊥ = J(E, F ). Indeed, take an arbitrary operator 0 ≤ 0 U ∈ J(E, F ) with V1 ∧ U = 0 and put V := x0 ⊗ y, where x0 ∈ E+ and y ∈ F+ . Then sup(V ∧ Vn ) = V by the above. Hence, U ∧ V = U ∧ sup(V ∧ Vn ) = sup(U ∧ V ∧ Vn ) = 0,

Integral Operators

271

for 0 ≤ U ∧ Vn ≤ n(U ∧ V1 ) = 0. By the definition of J(E, F ) we have U = 0. By Theorem 3.1.2, there exists an operator V := sup(S ∧ Vn ) ≤ S in the K-space L∼ (E, F ). It is clear that sup(Vn ∧ V ) = sup[(Vn ∧ S) ∧ V ] = sup(Vn ∧ S) ∧ V = V. Put U := (S − V ) ∧ V1 ≥ 0 and observe that (Vn ∧ V ) + U = (Vn + U ) ∧ (V + U ) ≤ Vn+1 ∧ S whence, V = sup(S ∧ Vn ) = sup(S ∧ Vn+1 ) ≥ sup(V ∧ Vn ) + U = V + U. Consequently, U ≤ 0. Thus, U = 0, i.e., (S − V ) ∧ V1 = 0. But we have just proven that V1 is a weak order-unity in J(E, F ); therefore, S = V . B (4) Now we pass to proving Theorem 6.4.3 (1). C Take an operator S ∈ J(E, F ). Since S = S + − S − , we may assume that S ≥ 0. We have to show that S admits integral representation. By (2), for Sn := S ∧ Vn there exists a kernel Kn (s, t), i.e., Z (Sn x)(s) =

Kn (s, t) x(t) dµ(t)

(x ∈ E).

Since the sequence (Sn ) is increasing, (Kn ) is also an increasing sequence by Theorem 6.4.2. According to (3) supn Sn = S. Put K(s, t) = supn Kn (s, t). Since supn Sn x = Sx (x ∈ E+ ), by Beppo Levy’s Theorem, we infer that S is an integral operator with kernel K(s, t). Let S be a regular integral operator. Show that S ∈ J(E, F ). By the regularity of S, we have S ∈ L∼ n (E, F ). Without loss of generality we may assume that S ≥ 0. Assign Kn (s, t) := K(s, t) ∧n1Š (s, t) and let Sn be the integral operator with kernel Kn . Then for x ∈ E+ we have  Z  (Sn x)(s) ≤ n x(t) dµ(t) 1A (s) = (Vn x)(s). Since Vn ∈ J(E, F ) and 0 ≤ Sn ≤ Vn , we obtain Sn ∈ J(E, F ), for every band is an ideal. By Beppo Levy’s Theorem, supn Sn = S, whence S ∈ J(E, F ) by the definition of band. B

272

Chapter 6

6.4.5. Bukhvalov Theorem. Let S : E → F be a linear operator. The following are equivalent: (1) S is an integral operator; (2) if 0 ≤ xn ≤ x ∈ E (n ∈ N) and xn → 0 in measure then Sxn → 0 almost everywhere; (3) the operator S satisfies the following conditions: (a) if µ(Bn ) → 0 (Bn ∈ B) and χBn ≤ x ∈ E (n ∈ N) then S(χBn ) → 0 almost everywhere; (b) if 0 ≤ xn ≤ x ∈ E (n ∈ N) and xn → 0 almost everywhere then Sxn → 0 almost everywhere. C (1) ⇔ (2) follows from 4.3.10 and 6.4.3 (1); (2) ⇔ (3) is straightforward. B 6.4.6. Now we turn to the question of finding some conditions for a dominated linear operator in lattice-normed spaces of measurable vector-functions to admit integral representation. First of all we introduce the corresponding conceptions of integrality. (1) Take Banach spaces X and Y and let Z ⊂ Y 0 be a norming subspace. Consider a Z-weakly measurable operator-function K : Š → L (X, Z 0 ) and a measurable vector-function u : B → X. The vector-function (s, t) 7→ K(s, t)u(t) ((s, t) ∈ A × B) is Z-weakly measurable. Suppose that, for all z ∈ Z and almost all s ∈ B, the integral Z w(s, z) := hz, K(s, t)u(t)i dµ(t) B

is defined and, moreover, the linear functional z 7→ w(s, z) is continuous for almost all s ∈ A. Then the vector-function v : s 7→ w(s, ·) ∈ Z 0 is Z-weakly measurable. ¯ := v¯. If, for each u ¯ ∈ E(X), Denote the coset of this vector-function by v¯. Assign T u ¯ and if T u ¯ ∈ F , then a linear operator T : E(X) → Fs (Z 0 ) there exists a T u appears. In this case we say that T is the weakly integral operator with kernel K and take the liberty of writing Z
hz, T ui(s) = hz, K(s, t)u(t)i dµ(t) a.e. u ∈ E(X) . B

If the kernel K belongs to Mν (X, Z 0 ) (see 2.3.11) and the integral operator S with kernel K ∈ L0 (ν) acts from E into F , then the weakly integral operator T is dominated; moreover, S is its exact dominant (see 6.4.7 (1) below). (2) Using the same notation as in (1), take a simply measurable operator-function K : Š → L (X, Y ). If u : B → X is a measurable vector-function

Integral Operators

273

then the vector-function (s, t) 7→ K(s, t)u(t) is measurable too. Suppose that, for each u ¯ ∈ E(X), we have K(s, ·)u(·) ∈ L1 (µ, Y ) for almost all s ∈ B and that the measurable vector-function Z v(s) := K(s, t)u(t) dµ(t) B

satisfies v¯ ∈ F (Y ). Then a linear operator T : E(X) → F (Y ) can be defined by the formula T u ¯ = v¯. This operator is called a strongly integral operator with kernel K. Again, we take the liberty of writing Z
(T u)(s) = K(s, t)u(t) dµ(t) a.e. u ∈ E(X) . B

If the kernel K belongs to Msν (X, Y ) (see 2.3.12) and the integral operator S with kernel K acts from E into F , then T is a dominated operator and S is its exact dominant. 6.4.7. Consider a weakly (strongly) integral operator T acting from E(X)
into Fs (Y, Z) into F (Y ) with kernel K. We say that T is regular if K ∈ Mν (X, Z 0 ) and the measurable function K ∈ L0 (ν) is a kernel of an integral operator from E into F . (1) The weakly (strongly) integral operator T with kernel K is dominated if and only if it is regular. In this case, T is the integral operator with kernel K .

 C If K ∈ Mν (X, Z 0 ) then | z, K(s, t)x | ≤ K (s, t)kzkkxk; therefore,
|hz, T ui| ≤ kzkSK u , where SK is the integral operator with kernel K . Hence it is clear that T ≤ SK . Conversely, assume that T is a dominated operator. Then, for x ∈ X, z ∈ Z, and e ∈ E, the inequality Z

 z, K(s, t)x e(t) dµ(t) ≤ T (e)kxkkzk B

 holds. Therefore, for the integral operator Sx,z with kernel z, K(s, t)x we have |Sx,z | ≤ T whenever kzk ≤ 1 and kxk ≤ 1. If S := sup{|Sx,z | : kxk ≤ 1, kzk ≤ 1} then, according to 6.4.3 (2), S is an integral operator from E into F . It remains to observe that S ≤ T and S = SK . B

274

Chapter 6

(2) Let T be a dominated weakly (strongly) integral operator. If a sequence (un )n∈N in E(X) is such that un ≤ e ∈ E (n ∈ N) and un converges to zero in measure, then T un converges to zero almost everywhere. In particular, T is bo-continuous. C According to (1), the exact dominant T is an integral operator from
E to F . Hence, the required assertion follows from the estimate T un ≤ T un in view of 6.4.1 (3). B 6.4.8. Recall that an operator S : X → F is called an operator with abstract norm (in symbols S ∈ LA (X, F ) if there exists S := sup{|Sx| : kxk ≤ 1} in F . (1) To each operator with abstract norm S : X → F there is a unique 0 uS ∈ Fs (X ) satisfying T x = hx, uT i (x ∈ X). The mapping T 7→ uT is a linear isometry between the Banach–Kantorovich spaces LA (X, F ) and Fs (X 0 ). C If f := S then S(X) ⊂ {f }⊥⊥ and we may assume without loss of generality that 0 < f (t) < ∞ for every s ∈ A. Let Uf denote the operator in L0 (A, λ) defined by g 7→ f −1 g. Then Uf is an invertible order isomorphism and Uf−1 = Uf −1 . If
T := Uf ◦ S then obviously T ∈ LA X, L∞ (A, λ) and T = Uf (f ) = 1. Let ρ be a lifting of the space L∞ (A, λ), see 1.4.8. For each s ∈ A the functional ϕs : x 7→ (ρ ◦ T x)(s) is contained in X 0 , since |T x| ≤ kxk. Thus, the vectorfunction u : s 7→ ϕs is σ(X 0 , X)-measurable and hx , u(s)i = ϕs (x) = ρ(T x)(s). Moreover, ku(s)k = sup{kρ(T x)(s)k; kxk ≤ 1} = 1, so that u = 1. Denote by uS the coset of the vector-function f u. Then uS ∈ Fs (X), uS = f , and Sx = Uf−1 T x = f hx , u(·)i = hx , f (·)u(·)i = hx , f ui = hx , uS i. B Let E, ˆ, E 0 , and Lˆ (E, F ) be the same as in 5.5.8, but require in addition 0 that mE = L0 (µ). Take a Y -weakly measurable operator-function K

∈ Mν (X, Y ). If u ∈ E(X) then, for each y ∈ Y , the function hy, Kui : t 7→ y, K(t)u(t) is 0 1 measurable.
Moreover, hy, Kui ∈ L (ˆ) whenever K ∈ E , and the element ˆ hy, Kui ∈ F is defined.
(2) Theorem. For every dominated operator T ∈ Lˆ E(X), Fs (Y 0 ) , there exists a unique (to within equivalence) operator-function K ∈ Mν (X, Y 0 ) such that K ∈ E 0 and

hy, T ui = ˆ hy, Kui u ∈ E(X), y ∈ Y . The correspondence is a linear
isometry of the Banach–Kantorovich spaces
T 7→ K 0 0 Lˆ E(X), Fs (Y ) and Es L (X, Y 0 ) . C The proof repeats the reasoning of 5.5.10 with the only difference that we should use (1) instead of Theorem 5.5.1 (1). B

Integral Operators

275

(3) Every dominated operator T : L1 (µ, X) → L∞ (λ)s (Y 0 ) admits weak integral representation. C The operator T : L1 (µ) → L∞ (λ) is integral. Let L(·, ·) ∈ L∞ (ν) be its kernel. Consider the operator ˆ : L1,∞ (ν) → L∞ (λ) defined by the formula Z (ˆe)(s) =

L(s, t)e(s, t) dµ(t)


e ∈ L1,∞ (ν) ,

B

where L1,∞ (ν) is the space e(·, ·) such that R of measurable functions of two variables ∞ the function e (s) := B |e(s, t)| dµ(t) (s ∈ A) belongs to L (A, λ). Obviously, ˆ is a Maharam operator. By Theorem 4.3.3 there exists a unique dominated operator Te : L1,∞ (ν, X) → L∞ (λ)s (Y 0 ) such that Te = ˆ and gT f = Te(g ⊗ f ), where g ∈ L∞ (λ), f ∈ L1 (µ, X), and (g ⊗ f )(s, t) = g(s)f (t). According to (2), the following representation holds:
e hy, Teui = ˆ hy, Kui


u ∈ L1,∞ (ν, X), y ∈ Y ,

e ∈ Mν (X, Y 0 ) and K e = 1. Put K(s, t) := K(s, e t)L(s, t) (s ∈ A, t ∈ B). where K From the definitions of Te and ˆ it is clear that we obtain the desired integral representation for u := 1 ⊗ f . B 6.4.9. Let Y be a dual Banach space possessing the Radon–Nikod´ ym property. Every dominated operator T : L1 (µ, X) → L∞ (λ, Y ) admits strong integral representation. Pn C Denote by F the set of functions u : Š → R of the form u = k=1 ϕk ⊗ ψk , where ϕk ∈ L1 (B, µ), ψk ∈ L1 (A, λ), and ϕk ⊗ ψk : (s, t) 7→ ϕk (t)ψk (s) (k := 1, . . . , n). For any fixed x ∈ X, define the operator Gx : F → Y by the formula Gx (u) :=

n Z X

ψk (s)T (x ⊗ ϕk )(s) dλ(s).

k=1 A

The operator T : L1 (µ) → L∞ (λ) is integral and its kernel L belongs to L∞ (ν). Employing the Fubini Theorem, it is easy to derive the inequality Z kGx (u)k ≤ ˆ(|u|) := L(s, t)|u(s, t)| dν(s, t) (u ∈ F ). Š

By Theorem 4.3.3 the operator Gx admits a unique dominated extension onto L1 (ν); we denote this extension by the same symbol Gx . In our case, Gx ≤ ˆ.

276

Chapter 6

In particular, the operator Gx : L1 (ν) → Y is bounded. Without loss of generality, we may assume that the measure ν is finite. Since Y possesses the Radon–Nikod´ ym ∞ property, there exists a gx ∈ L (ν, Y ) such that Z Z
Gx e = gx e dν e ∈ L1 (µ), u ∈ L1 (µX) . Gx (u) = gx u dν, Š

Š

Assign U x := gx . We obtain a linear operator U : X → L∞ (ν, Y ) that satisfies the following conditions: Z

Z ψT (x ⊗ ϕ) dλ =

A

(U x)ϕ ⊗ ψ dν


ϕ ∈ L1 (λ), ψ ∈ L1 (µ) ,

Š

U := sup U x ≤ L. kxk≤1

Using 4.3.5 (3), as well as the formula  for calculating the supremum of a set of regular operators, we may show that sup Gx : kxk ≤ 1 = ˆ; therefore, U = L. Now let ρ be a lifting of the space L
∞ (ν, Y ) associated with that of L∞ (ν). Assign K(s, t)x := (ρU x)(s, t) (s, t) ∈ Š . As is seen, the operator-function K : Š → L (X, Y ) is simply measurable and K = L. The definitions of U and K imply the following representation:   Z Z Z ψT (x ⊗ ϕ) dλ = ψ(t)  K(s, t)(x ⊗ ϕ)(s) dµ(s) dλ(t). A

A

B

Hence we obtain Z (T u)(t) =

K(s, t)u(s) dµ(s)


u ∈ X ⊗ L1 (µ) .

B

The same equality remains valid for all u ∈ L1 (µ, X), since X ⊗ L1 (µ) is dense in L1 (µ, X), the operator T is o-continuous, and passage to the limit is possible under the Bochner integral sign. B 6.4.10. Theorem. For every dominated operator T from E(X) to Fs (Y, Z), the following are equivalent: (1) T admits weak integral representation; (2) some dominant of T admits integral representation; (3) if a sequence (un )n∈N in E(X) is such that un ≤ e ∈ E (n ∈ N) and un → 0 in measure, then T un → 0 almost everywhere;

Integral Operators

277

(4) T is sequentially o-continuous and if a sequence of measurable sets An ∈ A is such that λ(An ) → 0 then T (uχAn ) → 0 almost everywhere for each function u ∈ E(X). C The implication (1) ⇒ (2) follows from 6.4.7 (1), and (2) ⇒ (3) from 6.4.7 (2); it is obvious that (3) ⇒ (4). It remains to prove validity of (4) ⇒ (1). Taking it into account that ∗-convergence of an o-bounded sequence in L0 (µ) coincides with convergence in measure, from (4) we easily deduce ∗-o-continuity of the operator T . By Theorem 4.3.10 the operator T is almost integral. According to 6.4.3 (1), T is an integral operator; let L(·, ·) ∈ L0 (ν) be its kernel. There exists a countable partition (Šn ) of the set Š into disjoint ν-measurable subsets such that Ln := χŠn L ∈ L∞ (ν) for all n ∈ N. Denote by Sn the integral operator with kernel Ln . It is clear that (Sn )n∈N is a sequence of positive pairwise disjoint operators from E into F . Moreover, Sn can be also regarded as a positive operator from L1 (µ) into L∞ (λ). Since the exact dominant is decomposable (see 4.2.6), there exists a sequence of pairwisePdisjoint operators Tn
: E(X) → Fs (Z 0 ) such that Tn = Sn ∞ (n ∈ N) and T u = bo- k=1 Tn u u ∈ E(X) . The restriction of Tn onto L1 (µ, X) ∩ E(X) can be extended to a dominated operator from L1 (µ, X) into L∞ (λ)s (Z 0 ) with the exact dominant Sn preserved (see Theorem 4.3.3). We denote such an extension by the same symbol Tn . According to 6.4.8 (3), the operator Tn admits weak integral representation with kernel Kn : Š → L (X, Z 0 ), for which Kn = Ln . Now define an operator-function K : Š → L (X, Z 0 ) so that the restriction of K onto Šn coincide with Kn . Then K is a Z-weakly measurable function, K = L, and, moreover, for all z ∈ Z and u ∈ W := E(X) ∩ L1 (µ, X), the following equality holds almost everywhere: Z

 hz, T ui(s) = z, K(s, t)u(t) dµ(t). A

This representation is valid for all u ∈ E(X), since W is bo-dense in E(X), the operator T is bo-continuous, and passage to the limit is possible under the integral sign. B 6.4.11. Theorem. Let Y be a dual Banach space possessing the Radon– Nikod´ ym property. A dominated operator T : E(X) → F (Y ) admits strong integral representation if and only if one of the conditions (2)–(4) of Theorem 6.4.10 is satisfied. C The proof follows the same scheme as in 6.4.10, but we should use 6.4.9 instead of 6.4.8 (3). B 6.4.12. Theorems 6.4.10 and 6.4.11 allow us to obtain results about the general from of some classes of dominated operators. Recall that MJ (X, Y ) is the band of

278

Chapter 6

almost integral operators from X into Y (see 4.3.9). Denote by EF the o-ideal of L0 (ν) constituted by the kernels of regular integral operators from E into F . (1) Theorem. The mapping sending a weakly measurable operatorfunction to the corresponding weakly integral operator establishes a linear isometry

of the Banach–Kantorovich spaces EFs L (X, Y 0 ) and MJ E(X), Fs (Y 0 ) . (2) Theorem. Let Y be a dual Banach space possessing the Radon– Nikod´ ym property. The mapping, sending a simply measurable operator-function to the corresponding strongly integral operator,
establishes a linear
isometry of the Banach–Kantorovich spaces EFs L (X, Y ) and MJ E(X), F (Y ) . p,∞ (3) Theorem. Let Y be the
same as in (2) and G := L (ν ⊗ λ),
1 p 1 ≤ p < ∞. Then the BKSs Gs L (X, Y ) and M L (ν, X), L (λ, Y ) are linearly isometric in the sense of Theorem (2).

6.5. Pseudointegral Operators The main result of the present section is a criterion for pseudointegrality of a dominated operator in spaces of measurable vector-functions. 6.5.1. Let, as above, (A, A , λ) and (B, B, µ) be spaces with complete σ-finite measures; let E and F be ideal spaces on (B, B, µ) and (A, A , λ), respectively, and let (Š, †, ν) be the product of these measure spaces. By a representing measure or a measure kernel we mean a positive countably additive function m : † → R∪{+∞} satisfying the following conditions: (a) there exists a countable increasing set of measurable sets Bn ∈ B and, for every n ∈SN, a countableSset of pairwise disjoint ∞ ∞ measurable sets An,k ∈ A such that B = n=1 Bn , A = k=1 An,k (n ∈ N), and m(An,k × Bn ) < ∞ (n, k ∈ N); (b) m(A0 × B0 ) = 0 whenever A0 × B0 ∈ † and either λ(A0 ) = 0 or µ(B0 ) = 0. Take a representing measure m : † → R ∪ {+∞}. For e ∈ L0 (µ), denote  F (e) := F (m, e) := f ∈ L0 (λ) : e ⊗ f ∈ L1 (m) , where (e ⊗ f )(s, t) := e(t)f (s) (s ∈ A, t ∈ B). Let D(m) be constituted by e ∈ L0 (µ) such that F (e) is an orderdense ideal of L0 (λ). Then D(m) is an order-dense ideal of L0 (µ). (1) There exists a unique order continuous operator ˆm : D(m) → L0 (λ) such that Z

ZZ f (s)(ˆm e)(s) dλ(s) =

A

e(s)f (t) dm(s, t) Š

for all e ∈ D(m) and f ∈ F (m, e). The operator ˆm is positive if and only if so is m.

Integral Operators

279

C Indeed, for any fixed e ∈ E, the right-hand side of the indicated equality defines an o-continuous functional ϕe on F (m, e). By Theorem 3.4.8 there exists 0 0 0 0 1 a unique element g ∈ F (m, e) := f ∈ L (λ) : f · F (m, e) ⊂ L (λ) such that Z ϕe (f ) =

g(s)f (s) dλ(s). A

It remains to assign ˆm e := g and observe that if m ≥ 0 and e ≥ 0 then ϕe ≥ 0 and ˆm e ≥ 0. Order continuity of ˆm follows from the Dominated Convergence Theorem. B An operator S : E → F is called pseudointegral if there exists a representing measure m : † → R ∪ {+∞} such that E ⊂ D(m) and Se = ˆm e (e ∈ E). In view of the above proposition every pseudointegral operator is order continuous. (2) Suppose that (B, B) is a standard Borel space; i.e., (B, B) is Borel isomorphic to a Borel subset of a complete separable metric space. In this event the representing measure m admits disintegration: dm(s, t) = dms (t)dλ(s). More precisely, there exists a mapping s 7→ ms (s ∈ A), with each ms a Borel measure on B, such that the following conditions are satisfied: (1) if B0 ∈ B and µ(B0 ) = 0 then ms (B0 ) = 0 for λ-almost all s; (2) for every B0 ∈ B, the function s 7→ ms (B0 ) is Borel; (3) if A0 ∈ A , B0 ∈ B, and m(A0 × B0 ) < +∞ then Z m(A0 × B0 ) =

ms (B0 ) dλ(s). A0

A mapping s 7→ ms with this properties is called a disintegrated kernel. A pseudointegral representation under the stated additional assumption takes the form Z (T u)(s) =

e(t) dms (t). B


6.5.2. Let I (ˆm ) be the order ideal in L∼ D(m), L0 (λ) generated by the operator ˆm . There exists a linear and order isomorphism ƒ from I (ˆ) onto L∞ (m) such that, for each S ∈ I (ˆm ), the following representation holds: Z

ZZ (Se)(s)f (s) dλ(s) =

A

ƒ(S)(s, t)f (s)e(t) dm(s, t) Š


e ∈ D(m), f ∈ F (m, e) .

280

Chapter 6

C Suppose that e1 , . . . , en ∈ E+ and f1 , . . . , fn ∈ F (m, e)+ . If 0 ≤ S ≤ ˆm Pn and z = k=1 ek ⊗ fk then e := Sz

n Z X

ZZ S(ek )fk dλ ≤

k=1 A

z(s, t) dm(s, t). Š

This implies that there exists a unique bounded positive functional Se on L1 (m),
R e ⊗ f ) = (Se)f dµ e ∈ E, f ∈ F (m, e) . The mapping S 7→ Se is for which S(e A an algebraic and order isomorphism of I (ˆ) onto L1 (m)0 . It remains to observe that L1 (m)0 ' L∞ (m). B 6.5.3. Now suppose that B is a σ-compact topological space and µ is a regular Borel measure on B. Theorem. Let T : E → F be an order continuous positive operator. Then there exists a representing measure m such that Z

ZZ g(s)(T e)(s) dλ(s) =

g(s)e(t) dm(s, t)

A


e ∈ E, g ∈ F (m, e) .

Š

Moreover, for every operator S ∈ {T }⊥⊥ there is a m-measurable function K(s, t) such that Z ZZ
g(s)(Se)(s) dλ(s) = K(s, t)g(s)e(t) dm(s, t) e ∈ E, F (Km, e) . A

Š

S∞ C Let (Bn ) be a sequence of compact subspaces with B = n=1 Bn and let En be the band in E generated by Bn , i.e. En = [Bn ]E, where [Bn ] denotes the band e 7→ χ] projection ~ Bn e. Denote Tn := T ◦ [Bn ]. According to Theorem 6.3.11 there exists a unique modular measure ϕn : A ⊗ Bn → F such that ZZ e ∈ En ). e(t) dϕn (s, t) (~

Tn e =

Given n ∈ N, choose a partition (An,k )k∈N of A such that [An,k ]ϕn (D) ∈ L1 (λ) for all D ∈ A ⊗ Bn and n, k ∈ N. Define a real-valued measure Z mn (D) :=

ϕn (D) dλ(s)

(D ∈ A ⊗ Bn ).

Integral Operators

281

There is a unique positive countably additive numerical measure m on A ⊗ B that agrees with mn on A ⊗ Bn . Clearly m (An,k × Bn ) = mn (An,k × Bn ) < ∞. If B 0 ⊂ Bn and A0 ⊂ An,k then for D = A0 × B 0 , by modularity of ϕ, we have Z Z 0 0 m(D) = ϕ(A × B ) dλ = ϕ(A × B 0 ) dλ. A0

A

Now it is evident that λ(A0 ) = 0 implies m(D) = 0. The fact that m(B 0 ) = 0 implies m(D) = 0 was proved in 6.3.10 (1). Hence, m is a representing measure. Now, if e ∈ E and f ∈ F (m, e) then   Z Z Z f (s)e(t) dmn (s, t) = f (s)  e(t) dϕn (s, t) dλ(s). Š

A

Š

This identity is evident for e := 1C and f := 1D , needs some simple calculation if e and f are step-functions, and is deduced with the help of the Monotone Convergence Theorem for arbitrary e ∈ E and f ∈ F (m, e). For the operator we have the representation Z ZZ
g(s)[Bn ](T e)(s) dλ(s) = g(s)e(t) dmn (s, t) e ∈ E, g ∈ F (m, e), n ∈ N . Employing again the Monotone Convergence Theorem we obtain the desired representation for T . The assertion concerning S is deduced by similar arguments on using the second part of Theorem 6.3.11. B 6.5.4. Sourour Theorem. A positive operator from E into F admits pseudointegral representation if and only if it is order continuous. C See 6.5.1 (1) and 6.5.3. B 6.5.5. Given an order continuous operator ˆ : D(ˆ) → F , where D(ˆ) is an order-dense ideal in L0 (λ), we introduce the space Lˆ (E, F ) as in 3.2.4; the latter is a band of Ln (E, F ). Theorem. Let ˆ be an order continuous operator from D(ˆ) into F . There exists a linear and order isomorphism ƒ from Lˆ (E, F ) onto some order-dense ideal of the space L0 (m) such that, for each S ∈ Lˆ (E, F ), the following representation holds: Z ZZ (Se)(t)f (t) dλ(t) = ƒ(S)(s, t)f (s)e(t) dm(s, t) A

Š

whenever e ∈ E and f ∈ F (km, e) with k := ƒ(S).

282

Chapter 6

C According to 6.5.3, there exists a representing measure m, for which D(ˆ) = D(m) and ˆ = ˆm . Suppose that 0 ≤ S ∈ Lˆ (E, F ), assign E0 := E ∩ D(m), and denote by S0 and ˆ0 the restrictions of the respective operators S and ˆ onto E0 . Put Sn := S0 ∧ (nˆ0 ). By Proposition 6.5.2 the operator Sn is pseudointegral and admits kn m as a representing measure, where kn ∈ L∞ (m). The sequence (kn ) is increasing and bounded almost everywhere. Indeed, by Theorem 6.5.4 the operator S admits a pseudointegral representation with some representing measure m0 . Consequently, ZZ ZZ 0 f (s)e(t) dm (s, t) = lim kn (s, t)f (s)e(t) dm(s, t) n→∞

Š

Š

for all e ∈ E0 and f ∈ F (m0 , e). Therefore, the sequence (kn ) is bounded and m0 = km, where k = sup kn . Assign ƒ(S) := k. It is easy that ƒ defines a linear and order isomorphism. B 6.5.6. Now consider a linear operator T : E(X) → Fs (Y, Z), where X and Y are Banach spaces. Take a Z-weakly m-measurable operator-function K : Š → L (X, Z 0 ), a strongly µ-measurable vector-function u : B → X, and some orderdense ideal Fu in L0 (λ). Suppose that for all z ∈ Z and g ∈ Fu the following integral exists: Z w(z, g, u) := hz, K(s, t)u(t)ig(s) dm(s, t). Š

For a fixed z ∈ Z, the functional w(z, ·, u) on F0 is order continuous; therefore, there exists a function v(z, ·, u) ∈ L0 (λ) such that Z v(z, s)g(s) dλ(s) = w(z, g, u) (g ∈ F0 ). Next, observe that the operator U : z 7→ v(z, ·, u) (z ∈ Z) is linear. If it has the form Uz (s) = hz, T ui(s) (for the above operator T : E(X) → Fs (Y, Z)), then T is said to be the weakly pseudointegral operator with representing measure m and kernel K. In this case we take the liberty of writing ZZ Z hz, T ui(s)g(s) dλ(s) = hz, K(s, t)u(t)ig(s) dm(s, t). A

Š

If K ∈ Mm (X, Z 0 ) and the operator S : E → F acts by the formula Z ZZ (Se)(s) dλ(s) = K (s, t)e(t)g(s) dm(t, s), A

Š

Integral Operators

283

then S is a dominant for T (moreover, S is the exact dominant of T , see 6.5.7). A weakly pseudointegral operator T is called regular if K ∈ Mm (X, Z 0 ) and the pseudointegral operator with representing measure K m acts from E into F . If (A, A ) is a standard Borel space then the weakly pseudointegral operator has the form Z

 hz, T ui(s) = z, K(s, t)u(t) dms (t). A

6.5.7. Theorem. The weakly pseudointegral operator T : E(X) → Fs (Y, Z) with kernel K and representing measure m is dominated if and only if T is regular. In this case, T is the pseudointegral operator with representing measure K m. C Assume that a linear operator T admits weak pseudointegral representation 6.5.6. If K ∈ Mm (X, Z 0 ) and S : E → F is the pseudointegral operator with kernel K and representing measure m, then Z

Z hT u, zi(s)g(s) dλ(s) ≤ kzk

A


S u (s)g(s) dλ

A


for all g ∈ Fu := F ( K m, u ). Whence we deduce hT u, zi
≤ S u kzk, and passage to the supremum over all kzk ≤ 1 yields T u ≤ S u . Conversely, assuming that a weakly pseudointegral operator T is dominated, prove that its kernel K belongs to Mm (X, Z 0 ). Let ˆ be the pseudointegral operator from D(m) into L0 (λ) generated by a representing measure m. For x ∈ X and z ∈ Z, define the operator Sx,z from E into L0 (λ) by the formula Z

ZZ Sx,z (e)(s)g(s) dλ(s) =

A

 z, K(s, t)x e(t)g(s) dm(s, t),

Š

where e ∈ E and g ∈ F (e). It is clear that Sx,z ∈ Lˆ (E, F ). Since T is dominated, we have the estimate Z Z Sx,z (e)(s)g(s) dλ(s) ≤ kxkkzk · T (e)(s)g(s) dλ(s) A

A

for e ∈ E+ and g ∈ F (e). Hence it is clear that Sx,z ≤ T whenever kxk ≤ 1 and kzk ≤ 1. Let S0 be the supremum in L∼ (E, F ) of the bounded set {Sx,z : kxk ≤ 1, kzk ≤ 1}. Since Lˆ (E, F ) is a band, S0 ∈ Lˆ (E, F ). By Theorem 6.5.5  L := sup hz, Kxi : kxk ≤ 1, kzk ≤ 1

284

Chapter 6

exists in the space L0 (m) and the following representation holds: Z ZZ S(e)(s)g(s) dλ(s) = L(s, t)e(t)g(s) dm(s, t). A

Š

Thus, K ∈ Mm (X, Z 0 ), K = L, and S ≤ T . While proving sufficiency, we have seen that T ≤ S; hence, S = T , i.e., T is the pseudointegral operator with kernel K and representing measure m. B 6.5.8. Theorem. For every dominated operator T : E(X) → Fs (Y, Z), the following are equivalent: (1) the operator T is order continuous; (2) the operator T admits pseudointegral representation; (3) the operator T admits weak pseudointegral representation. C (1) ⇔ (2): This follows from Theorems 4.3.2 and 6.5.3. (2) ⇒ (3): Let m be a representing measure of the operator T . According  ⊥⊥ to 6.5.5, there exists an isomorphism ƒ from T onto an ideal subspace

 L⊂ 0 L (m). For x ∈ X and z ∈ Z, the operator S := Sx,z : e 7→ T (x ⊗ e), z acts  ⊥⊥ from E into F , is regular, and belongs to T . Assign b(x, z) := ƒ(S
x,z ). We know that b is a bilinear operator with abstract norm and b = ƒ T , see 6.4.8 (1) and 5.5.1 (2). By Theorem 6.4.8 (1) there exists a Z-weakly measurable operator-function K ∈ Mm (X, Z 0 ) such that b = K and b(x, z) = hz, Kxi (x ∈ X, z ∈ Z). For any operator S ∈



T

⊥⊥

, in view of what was said above, we may write:

Z

ZZ (Se)(s)g(s) dλ(s) =

B

ƒ(S)e(t)g(s) dν(s, t). Š

Hence, according to the definition of the operator b, we have: Z ZZ



 z, K(s, t)x e(t)g(s) dν(s, t). z, T (x ⊗ e) (s)g(s) dλ(s) = B

Š

Taking account of the fact that X ⊗ E is bo-dense in E(X), after appropriate passage to the limit under the integral sign, we obtain a required pseudointegral representation for T . (3) ⇒ (2): This is an immediate consequence of 6.5.7 and 6.5.1 (1). B

Integral Operators

285

6.5.9. Let G be the ideal of L0 (m) consisting of functions k such that the pseudointegral operator with representing measure km acts from E into F . Assign

Mˆ E(X), Fs (Y 0 ) := T ∈ M E(X), Fs (Y 0 ) : T ∈ Lˆ (E, F ) .
Theorem. To each K ∈ Gs L (X, Y
0 ) , there corresponds the weakly pseudointegral operator T ∈ M E(X), Fs (Y 0 ) with kernel K and representing measure m. The
mapping K 7→ T establishes a linear isometry between the BKSs
Gs L (X, Y 0 ) and Mˆ E(X), Fs (Y 0 ) . C Only the first part of the assertion requires proving, the rest of the claims
follows immediately from 6.5.3, 6.5.5, 6.5.7, and 6.5.8. Take a K ∈ Gs L (X, Y 0 ) . For any fixed u ∈ E(X) and y ∈ Y , the functional ZZ


ϕ : f 7→ y, K(s, t)u(s) f (t) dm(s, t) f ∈ F (m, u ) Š

is order continuous. Consequently, the following representation holds: Z
ϕ(f ) = f (s)g(s) dλ(s) f ∈ F (m, u ) , A

where g is a uniquely defined element of L0 (µ). Consider the operator S : Y → L0 (µ) defined by the formula Sy := g. If ‰ is the pseudointegral operator with representing measure K m then Z ZZ f (s)g(s) dλ(s) ≤ kyk K (s, t) u (s)f (t) dm(s, t) A

Š

Z = kyk




‰ u (s)f (s) dλ(s)

f ∈ F m, u





.

B


Hence it is clear that Sy := g ≤ ‰ u kyk. Therefore, S is an operator with abstract norm and S ∈ F . By the representation theorem for operators with abstract norm (see 6.4.8 (1)), there exists a v ∈ Fs (Y 0 ) such that Sy = hy, vi (y ∈ Y ). Now define the operator T by the equality T u:= v. From the definitions we have: Z Z hy, T ui(s)f (s) dλ(s) = g(s)f (s) dλ(s) A

A

ZZ = ϕ(g) = Š

and the result follows. B

 y, K(s, t)u(t) f (s) dm(s, t)

286

Chapter 6 6.6. Comments

6.6.1. (1) Some integration theory of scalar-valued function with respect to a σ-additive measure with values in a Stone algebra (say F ) has been developed by J. D. M. Wright [404]. This theory (including all convergence theorems) remains valid when F is an arbitrary Kσ -space, see [402]. The construction in 6.1.1–6.1.7 has been performed in [214] and essentially repeats J. D. M. Wright’s [404] considerations in a more general situation: integrable objects are elements of some Kσ -space as in [163, 388] and the vector lattice F is replaced by a lattice-normed space. (2) It can be easily observed that the Radon–Nikod´ ym Theorem for measures with values in an (order complete) vector lattices fails in general: if µ is a σ-additive (finite) measure then the R2 -valued measures µ1 := (1, 0)µ and µ2 := (0, 1)µ are absolutely continuous with respect to one another, but µ1 (A) = (1, 0)µ(A) 6= (0, 1)Iµ (f χA ) = Iµ2 (f χA ) provided that µ(A) 6= 0. J. D. M. Wright has shown in [400] that the Radon–Nikod´ ym Theorem is true for special class of ample measures. By definition [400] µ is ample if the space L2 (µ) is a Kaplansky– Hilbert module (see Section 7.4). J. D. M. Wright deduced his Radon–Nikod´ ym Theorem [400; Theorem 4.1] (that essentially coincides with 6.1.11 (2)) from the following auxiliary fact [400; Lemma 4.2] (which is immediate from Kaplansky’s Theorem [166; Theorem 5], see 7.5.7 (2)). Let µ be an ample C(Q)-valued measure and let T : L2 (µ) → C(Q) be a normbounded module homomorphism. Then there exists a unique g ∈ L2 (µ) such that Z
T f = f g dµ f ∈ L2 (µ) . (3) By [400] a measure µ : B → C(Q) (with Q an extremal compact space) is modular with respect to an algebra homomorphism π : C(Q) → L∞ (µ) if Z Z π(a)f dµ = a f dµ (a ∈ C(Q), f ∈ L1 (µ)). Equivalence of this definition to that in 6.1.9 follows from 6.1.8, 6.1.8 (3), and 2.1.8. According to 6.1.9 (4), 2.2.3, and 7.4.4 the definitions of ample measure in [400] and 6.1.9 are also equivalent. 6.6.2. (1) The main results of Section 6.2 (Theorems 6.2.2 and 6.2.6) as well as the concept of quasi-Radon measure are due to S. A. Malyugin [273] and stem from Wright’s theory of Stone-algebra-valued measures [400, 402, 404–406]. The key idea in the proof of Theorem 6.2.6 is to use the Birkhoff–Ulam homomorphism (in the form 6.1.1 (4) or like) was employed for the first time by J. D. M. Wright [404].

Integral Operators

287

Other Riesz-type representation theorems for positive and order-bounded operators are presented in [43, 103, 171, 172, 214, 247, 314]. In this connection we must mention the two classical works by A. A. Markov [276] and A. D. Alexandrov [16] which are inspirational sources for studying Riesz-type representations and related topics for more than half of a century (see also [34, 112]). (2) Theorem 6.2.7 (1) was proved by J. D. M. Wright [404; Theorem 4.1]. In this result µ cannot be chosen regular rather than quasiregular. Moreover, J. D. M. Wright [406; Theorem T] obtained the following elegant characterization of order complete vector lattices for which this choice is always possible. Theorem. Let F be an order complete vector lattice. Then the following are equivalent: (i) F is weakly (σ, ∞)-distributive; (ii) each F -valued Baire measure on every compact space can be extended to a regular F -valued Borel measure; (iii) every F -valued quasiregular Borel measure on every compact space is regular. An analogous result for measures with values in F ∪{∞} was obtained in [314]. (3) Another peculiarity of Theorem 6.2.7 (1) is that it cannot be proven by the Daniell extension method. The Daniell construction fails since the Baire measure may be irregular. The following fact is also due to J. D. M. Wright [402]. We say that F has the measure extension property if, for every set Š and each algebra A of subsets of Š each countably additive measure µ : A → F has some σ-additive extension to the σ-algebra σ(A ) generated by A . Theorem. Let F be an order σ-complete vector lattice. Then the following conditions are equivalent: (i) F is weakly σ-distributive; (ii) F has the measure extension property; (iii) every F -valued Baire measure on every compact space is regular. (4) Speaking of the measure extension problem, we should mention the so-called extreme extension of measures. This direction stems from the classical result by A. Horn and A. Tarski [133]: A finitely additive positive real-valued measure defined on a subalgebra of a Boolean algebra admits a finitely additive positive extension to the whole algebra. Let A be a Boolean algebra and let A0 be its subalgebra. Let Ch(ε+ F (µ0 )) denote the set of extreme extensions, i.e. the collection of all positive finitely additive extensions of a measure µ0 : A0 → F to the whole algebra A . The following characterization of an extreme extension of a real-valued measure was given by D. Plachky [320]: µ ∈ Ch ε+ R (µ0 )) ⇔ (∀ a ∈ A ) inf{µ(a4a0 ) : a0 ∈ A0 } = 0.

288

Chapter 6

This result was generalized to measures with values in an order complete vector lattice in [248]. Such results are somehow connected with the Buck–Phelps characterization of extreme points, see [132]. An operator version of the Buck–Phelps result was obtained by S. S. Kutateladze [223]; see also [209]. (5) Let A and A0 be the same as in (4), with F an order complete vector lattice and Y a Banach–Kantorovich space over F . Observe two more results of [217]. The first claims existence of an extreme extension and, in the case of measures with values in an order complete vector lattice, it was proved in [248]; the second guarantees existence of simultaneous extension and, for scalar-valued measures, it was obtained in [126].
(5.1) Theorem. Let µ0 ∈ da(A0 , Y ). Then Ch ε+ F ( µ0 ) 6= ∅ and for each extreme extension ν : A → F of µ0 there exists a unique measure µ ∈ da(A , Y ) such that µ extends µ0 and µ = ν. Denote by % and %0 the restriction mappings from ba(A , F ) onto ba(A0 , F ) and from da(A , F ) onto da(A0 , F ), respectively. (5.2) Theorem. There exist an order continuous lattice homomorphism ε0 from ba(A0 , F ) onto an order-closed sublattice of ba(A , F ) and an Orthlinear operator ε from da(A0 , Y ) into ba(A , Y ) such that εµ = ε0 µ , % ◦ ε is the identity mapping in da(A , F ), and %0 ◦ ε0 is the identity mapping onto da(A0 , F ). 6.6.3. (1) The material of Section 6.3 is mainly taken from the article by E. V. Kolesnikov and A. G. Kusraev [180]. After the first unusual step (extension of a given “preintegral” to the lattice V in 6.3.1) the construction follows the classical Daniell scheme, see [250, 348]. The main results remain valid if M (A, A , N ) satisfies the local countable chain condition (is of countable type locally), i.e. there is a partition of unity (πα ) in P(F ) such that πα satisfies the countable chain condition for all α. Evidently M (A, A , N ) satisfy the local countable chain condition if and only if so is (A , N ) The latter means that there is a partition (Aα ) of A satisfying: (a) (A ∩ Aα , N ∩ Aα ), where B ∩ Aα = {B ∩ Aα : B ∈ B}, obeys the countable chain condition for all α; (b) if C ⊂ A and C ∩ Aα ∈ A for all α then C ∈ A ; (c) if C ⊂ A and C ∩ Aα ∈ N for all α then C ∈ N . (2) In [398] A. W. Wickstead developed some integration theory for vector-valued functions by Stone-algebra-valued functions very close to that in Section 6.3. Let Q be a Stone space (= an extremal compact space) and let B be a σ-algebra of subsets of a nonempty set Š. Given a countablyPadditive positive ∞ measure µ : B → C(Q), the integral of a simple function f := o- n=1 χAn en with A1 , . . . , An ∈ B and e1 , . . . , en ∈ C∞ (Q) in [398] is defined by Z ∞ X Iµ := f dµ := oen µ(An ). Š

n=1

Integral Operators

289

Further, using Iµ as a preintegral, the Lebesgue integral is developed along the lines of the Daniell scheme. Thus, integrable functions are defined on Š, taking their values in C∞ (Q). On the contrary, our simple functions (the lattice V ) can be identified with the functions defined on Q and taking values in M (Š, B, N ) (see 6.5.1 and 3.5.6). This moderate difference actually turns out essential, leading to a simpler and more flexible construction. (3) As was mentioned in 3.6.4 the theory of Sections 3.4, 3.5 and 6.3 stems from the D. Maharam fundamental papers [264–269]. In particular, variants of Theorems 6.3.6, 6.3.8, 6.3.9, and 6.3.11 were obtained in [268, 269]. The main difference is that in Maharam’s representation is performed on the Stone spaces of vector lattices E and F , while our results are formulated in the initial terms of a locally σ-compact space Q and measurable structure (A, A , N ). (4) In [339] I. E. Schochetman studied integral operators in the spaces of measurable sections of Lebesgue spaces. The relevant definitions can be easily extended to the spaces E(X ). It would be interesting to obtain necessary and sufficient conditions for integrality of (dominated) linear operators in the spaces of type E(X ) as well as to study these integral operators in the spirit of [128, 183, 188]. 6.6.4. The main references to the topic of the classical integral operators, related to the theme of Section 6.4, are [128, 183, 188]; see also [162, 228]. (1) The question about integrality of a linear operator was posed by J. von Neumann. In the fundamental article [292] he solved the problem of finding all operators in L2 (0, 1) unitarily equivalent to some selfadjoint integral operator (see the monograph [183] by V. B. Korotkov) and posed the problem of finding necessary and sufficient conditions for a given operator in L2 (0, 1) to admit the integral representation of 6.4.1 [292, p. 4]. (2) Theorem 6.4.5 was established by A. V. Bukhvalov [56, 57]. The proof presented is very close to the original. The whole history, references, related ideas and results can be found in [59, 60], see also [228, 409, 410]. We only mention two preceding results by S. I. Zhdanov [412] and L. Lessner [235, 236] that are close to Bukhvalov’s criteria, the article by A. Schep [338] in which Theorem 6.4.5 was re-proved, and a new proof of the implication (1) ⇒ (2) in Theorem 6.4.5 found by L. Weis [392]. (3) The criteria for weak and strong integrality of a dominated operator in the spaces of measurable vector-functions (Theorems 6.4.10 and 6.4.11), as well as the general form of dominated operators (three theorems in 6.4.12), were obtained by A. G. Kusraev [199, 198]. For linear operators acting from the space of measurable vector-functions into an arbitrary Banach space, i.e., in case F = R

290

Chapter 6

similar results were obtained by V. G. Navodnov [290, 291]; for compact and weakly compact operators, see also the paper by Kevin [168]. (4) In spite of the fact that Theorem 6.4.2 is natural and easy for separable measure spaces, its proof is very involved in the general case. W. A. J. Luxemburg and A. C. Zaanen [263] gave a proof that grounds on approximating the kernel K(·, ·) with finite-rank kernels. Another proof presented in [162; Theorem XI.1.2] is based on a Yu. I. Gribanov’s result [115]. All subtlety of the theorem lies in the fact that, under some conditions on the set of functions in the domain of an integral operator, the supremum of the values of the operator on this set calculated in the K-space L0 coincides with the pointwise supremum. In the case when E and F are KB-spaces on [0, 1], Theorem 6.4.3 (1) was proven by G. Ya. Lozanovski˘ı in [253]. The general form was obtained in [56]. Theorem 6.4.8 (1) is due to A. V. Bukhvalov [55]; for 6.4.8 (2, 3) see [199, 204]. 6.6.5. (1) Pseudointegral operators (6.5.1) were introduced by W. Arveson [28] in connection with operator algebras in L2 . Pseudointegral operators were later considered by H. Fakhoury [94] (operators in L1 ) and N. J. Kalton [147] (operators in Lp with 0 < p ≤ 1). Different aspects of pseudointegral operators are reflected in [147–150, 344–347, 359–358, 393-395]. (2) In Section 6.5 we mainly follow [204]. Theorems 6.5.3 and 6.5.4 were proven by Sourour [361] but we pursue another our approach. The results 6.5.2 and 6.5.5 can be also deduced from [359, 361]. The main results 6.5.7, 6.5.8, and 6.5.9 on pseudointegrality of dominated operators in the spaces of measurable vector-functions are obtained by K. T. Tibilov [373]. (3) It follows from 4.6.3 (2) that in an ideal spaces over a diffuse measure space integral operators are disjoint from all lattice homomorphisms and all Maharam operators. In this connection an interesting example of a pseudointegral operator was constructed in [345]. Let € be a compact group of the unit disk and let ν be an arbitrary diffuse measure disjoint from the Haar measure µ (generated by the Lebesgue measure on [0, 1]). For every t R∈ € define the Borel measure νt on € by νt (B) := ν(t + B). The relation (T f )(t) = € f (s) dνt (s) defines a pseudointegral operator acting from L∞ (µ) to L∞ (µ). I. I. Shamaev [345] has proved that the operator T is disjoint from all integral operators, all lattice homomorphisms, and all Maharam operators in L∞ (µ).

Chapter 7 Operators in Spaces with Mixed Norm

In the present chapter, we study various classes of linear operators acting in spaces with mixed norm and defined in mixed terms of norm and order. If (X, E) is a lattice-normed space and E is a norm lattice of X then X can be endowed with a mixed norm so that X becomes a normed space, and even a Banach space in case the lattice-normed space is br-complete (7.1.1 and 7.1.2). The dual of a space with mixed norm is a space with mixed norm too (7.1.4); furthermore, the canonical embedding into the second dual preserves the vector norm (7.1.5). A more general result states that the space of dominated operators between spaces with mixed norm is itself a space with mixed norm if some natural conditions are met (7.1.9). Passage to the dual of an operator commutes rather often with the taking of the exact dominant of this operator (7.1.10). Various classes of operators under study in functional analysis are often defined in mixed terms that employ norm and order. Using the (positively homogeneous) functional calculus in Banach lattices, we introduce the class of (p, q)-summing operators in spaces with mixed norm (7.2.3). The set of (p, q)-summing operators acting in fixed spaces with mixed norm is a Banach space (7.2.4). A bounded operator is (p, q)-summing if and only if the dual operator is (q 0 , p0 )-summing (7.2.6). Note that (1, 1)-summing operators can be characterized in terms of convergent series (7.2.7) and, under some additional requirements, this class of operators coincides with the class of dominated operators (7.2.8). Particular cases of the notion of (p, q)-summing operator are presented by (p, q)-convex operators (7.2.11 (1)), (p, q)-concave operators (7.2.11 (2)), and (p, q)-regular operators (7.2.11 (3)). Lattice-normed spaces provide reasonable grounds for constructing some isometric classification of spaces with mixed norm. The key point in this respect is the presence of a complete Boolean algebra of projections in a Banach space as well as a special geometric property of the unit ball which is related to the algebra. The property is called B-cyclicity (7.3.3) or (B, p)-cyclicity with 1 ≤ p < +∞ (7.3.3). A Banach space is linearly isometric to a bo-complete space with mixed

292

Chapter 7

norm whose norm lattice is an order complete AM-space with unity (ALp -space,
1 6 p < ∞), if and only if it is (B, ∞)-cyclic (B, p)-cyclic with respect to some complete (Bad´e-complete) Boolean algebra B of projections (7.3.2 and 7.3.4). An interesting class of Banach–Kantorovich spaces arises as a generalization of Hilbert spaces by allowing the inner product to take values in a Stone algebra instead of the complex numbers (7.4.3). If the space is complete under the vectorvalued norm defined by the Stone-algebra-valued inner product then it is called a Kaplansky–Hilbert module (7.4.5). An orthonormal set and a basis are defined in a Kaplansky–Hilbert module just as it is done in a Hilbert space (7.4.6). Kaplansky– Hilbert modules with a basis are called homogeneous. Unlike Hilbert space, not every Kaplansky–Hilbert module has a basis, but every Kaplansky–Hilbert module splits into the direct sum of homogeneous bands (7.4.7). In spite of this essential dissimilarity, these two objects have common features (7.4.9, 7.4.10, 7.5.7 (2)). Moreover, every homogeneous Kaplansky–Hilbert module is unitarily equivalent to the space of continuous function defined on comeager sets of an extremal compact space and taking values in a Hilbert space (7.4.12). The space of all bounded endomorphisms of a Kaplansky–Hilbert module is a C ∗ -algebra; moreover, it is an AW ∗ -algebra (7.5.7), i.e. it meets additionally the axioms of a Baer ∗-ring (7.5.1). Conversely, every type I AW ∗ -algebra is ∗-isomorphic to such algebra (7.5.11). The endomorphism algebra is homogeneous if so is the underlying module. Any AW ∗ -algebra admits a Stone-algebravalued norm, so that the original norm is a mixed norm and the algebra itself is a Banach–Kantorovich space (7.5.5). Finally, every homogeneous AW ∗ -algebra is representable as the space of continuous functions defined on comeager sets of an extremal compact space and taking values in the algebra of bounded operators in a Hilbert space with the strong operator topology; an arbitrary type I AW ∗ -algebra splits into the direct sum of such algebras (7.4.11, 7.4.12). 7.1. Spaces with Mixed Norm In this section, we introduce spaces with mixed norm and study their simplest properties. We also consider the interrelation between the notions of dominated operator and mixed norm. 7.1.1. Recall that a normed (Banach) lattice is a vector lattice E that is simultaneously a normed (Banach) space whose norm is monotone in the following sense: if |x| 6 |y| then kxk 6 kyk (x, y ∈ E), see Section 1.5. If (X, E) is a latticenormed space with E a norm lattice of X then, by definition, x ∈ E for every x ∈ X, and we may introduce some mixed norm in X by the formula

|||x||| := x (x ∈ X).

Operators in Spaces with Mixed Norm

293

In this situation, the normed space (X, ||| · |||) is called a space with mixed norm. In view of the inequality x − y 6 x − y and monotonicity of the norm in E we have

x − y 6 |||x − y||| (x, y ∈ X), so that the vector norm · is a norm continuous operator from (X, ||| · |||) into E. All the notions introduced in Chapter 2 for a lattice-normed space make sense for a space with mixed norm, including decomposability, br-completeness, d-completeness, bo-completeness, etc. A Banach space with mixed norm is a pair (X, E) in which E is a Banach lattice and X is a br-complete lattice-normed space with E-valued norm. The following proposition justifies this definition. 7.1.2. Let E be a Banach lattice. Then (X, ||| · |||) is a Banach space if and only if the lattice-normed space (X, E) is complete with respect to relative uniform convergence. C ⇐ Take a fundamental sequence (xn ) ⊂ X. Without loss of generality, we may assume that |||xn+1 − xn ||| 6 1/n3 (n ∈ N). Assign n X en := x1 + k xk+1 − xk (n ∈ N). k=1

Then we may estimate

n+l

X

k xk+1 − xk ken+l − en k =

k=n+1

6

n+l X

k|||xk+1 − xk ||| 6

k=n+1

n+l X

k=n+1





1 −−−−→ 0. k 2 n,l→∞

Thus, the sequence (en ) is fundamental and hence it has a limit e := limn→∞ en . Since en+k > en (n, k ∈ N), we have e = sup en . If n > m then m xn+l − xn 6

n+l X

k xk+1 − xk 6 en+l − en 6 e;

k=n+1

consequently, xn+l − xn 6 (1/m)e. This means that the sequence (xn ) is brfundamental. By br-completeness, the limit x := br-limn→∞ xn exists. It is clear that limn→∞ |||x − xn ||| = 0. ⇒ Suppose now that a sequence (xn ) ⊂ X is br-fundamental; i.e., xn − xm 6 λk e (m, n, k ∈ N, m, n > k), where 0 6 e ∈ E and limk→∞ λk = 0. Then |||xn − xm ||| 6 λk kek → 0 as k → ∞; consequently, the limit x := limn→∞ xn exists. By continuity of the vector norm, we have x − xn 6 λk e (n ≥ k); therefore, x = br-lim xn . B

294

Chapter 7

7.1.3. Let E be a Banach lattice with an order continuous norm. Then the following assertions are valid: (1) (X, ||| · |||) is a Banach space if and only if (X, · ) is bo-complete; (2) if X is decomposable then (X, ||| · |||) is a Banach space if and only if (X, · ) is a Banach–Kantorovich space. C Assertion (2) is an obvious consequence of (1). Sufficiency in (1) follows from 7.1.2. It remains to show that if (X, ||| · |||) is a Banach space then (X, · ) is bo-complete. Let a net (xα ) ⊂ X be such that xα − xβ 6 eγ for α, β > α(γ) and suppose that eγ decreases and o-converges to zero. Since |||xα − xβ ||| 6 keγ k and
keγ k → 0 due to condition (A) , (xα ) is fundamental with respect to the mixed norm. Consequently, limα |||x − xα ||| = 0 for some x ∈ X. Taking continuity of the vector norm into account, we can pass to the limit over β in the mixed norm; therefore, x − xα 6 eγ for α > α(γ). This means that x = bo-limα xα . B 7.1.4. Theorem. If (X, E) is a decomposable space with mixed norm then (X , E 0 ) is a decomposable bo-complete space with mixed norm. The least dominant x0 ∈ E 0 of a functional x0 ∈ X 0 serves as its vector norm; in particular, 0

hx, x0 i 6



x , x0



(x ∈ X, x0 ∈ X 0 ).

 C A continuous functional x0 ∈ X 0 is bounded on the set x ∈ X : x 6 e for each e ∈ E+ . Hence, in view of 4.1.11 (1), it follows that x0 is dominated and

e, x0



 = sup hx, x0 i : x 6 e, x ∈ X .

(2)

0 By Theorem 4.2.6, M (X, R) is a Banach–Kantorovich space. Consequently, X is0  ∗ ∗ 0 0 a Banach–Kantorovich space too, since X := x ∈ M (X, R) : x ∈ E and E is an order ideal of Lr (E, R). It remains to observe that the norm in X 0 is mixed:

0 



x = sup e, x0 : e ∈ E+ , kek 6 1 n o  = sup sup hx, x0 i : x 6 e, x ∈ X : e ∈ E+ , kek 6 1

 = sup hx, x0 i : x ∈ X, x 6 1 = kx0 k. B 7.1.5. Let (X, E) be a decomposable space with mixed norm. The canonical embedding of X into the second dual X 00 preserves the vector norm. More precisely, if κ and λ are the canonical embeddings X → X 00 and E → E 00 , respectively, then κ(x) = λ( x ) (x ∈ X).

Operators in Spaces with Mixed Norm

295

C Observe first that, according to 7.1.4, we have hx0 , κ(x)i = hx, x0 i 6



x , x0



=



 x0 , λ( x )

for x ∈ X and x0 ∈ X. Consequently, κ(x) 6 λ( x ). Take an arbitrary functional e0 ∈ E 0 . By the Hahn–Banach Theorem, for a fixed x ∈ X there is a functional

 x0 ∈ X such that hx, x0 i = x , e0 and hu, x0 i 6 u , e0 for all u ∈ X. Using 7.1.4 again, we conclude that









 e0 , λ( x ) = x , e0 = hx, x0 i = x0 , κ(x) 6 x0 , x(x) 6 e0 , κ(x) .

Thus, λ( x ) 6 κ(x) . B 7.1.6. We say that the mixed norm in a lattice-normed space (X, E) is bo(bo)

continuous (with respect to the vector norm · ) if xα −−→ 0 implies |||xα ||| → 0. It is easy to see that, for a normed lattice X = E, bo-continuity of the mixed norm is equivalent to order continuity of the lattice norm as defined in 1.5.3. Let E0 denote the order ideal in E generated by X := { x : x ∈ X}. (1) Let E have the principal projection property and (X, E) be decomposable. Then the mixed norm is bo-continuous if and only if the norm in E0 is order continuous. C Sufficiency is evident. To prove the necessity take a decreasing net (eα ) with inf α eα = 0. According to 4.1.4 we may choose xα ∈ X such that xα = eα . Then the net (xα ) bo-converges to zero and, by hypothesis limα keα k = limα |||xα ||| = 0. B Suppose that the space X itself is a vector lattice. The vector norm in X is called order continuous if inf α xα = 0 implies inf α xα = 0 for every decreasing net (xα ) ⊂ X. We say that the mixed norm is mixed order continuous if the vector norm is order continuous and the norm of E0 is order continuous. (2) Suppose that (X, E) is a decomposable lattice-normed lattice and E0 is order complete. Then the mixed norm in X is order continuous if and only if it is mixed order continuous. C Sufficiency is again obvious. Assume that the mixed norm is order continuous. If e ≤ xα for some e ∈ E and decreasing net (xα ) ⊂ X with inf α xα = 0 then kek ≤ |||xα ||| → 0. Thus, inf α xα = 0 and the vector norm is order continuous. Next, take a decreasing sequence (en ) ⊂ E0 with inf n en = 0 and (en −en+1 ) ⊥ en+1 (n ∈ N). By 4.1.4 there is x1 ∈ X with x1 = e1 = e2 +(e1 −e2 ). By decomposability there exist x2 ∈ X such that x2 = e2 and 0 ≤ x2 ≤ x1 . By induction we may find a decreasing sequence (xn ) in X with xn = en . It follows that inf n xn = 0 and by hypothesis limα keα k = limα |||xα ||| = 0. In view of 1.5.3 (4) the norm of E0 is order continuous. B

296

Chapter 7

7.1.7. Theorem. For a decomposable bo-complete space with mixed norm (X, E), the following assertions are equivalent: (1) the mixed norm in X is bo-continuous; (2) if κ is the canonical embeddings X → X 00 then κ(X) is a decomposable subspace in (X 00 , E 00 ). C By virtue of 7.1.6 (1) and 1.5.3 (5) it sufficient to show that κ(X) is decomposable in X 00 if and only if λ(E) is an order ideal in E 00 , where E := E0 and λ is the canonical embedding E → E 00 . If λ(E) is an order ideal in E 00 and the representation λ( x ) = e1 + e2 holds for some x ∈ X and 0 ≤ e1 , e2 ∈ E then e1 , e2 ∈ λ(E). By the decomposability assumption there exists x1 , x2 ∈ X such that x = x1 + x2 and λ( xk ) = ek (k = 1, 2). Conversely, suppose that κ(X) is decomposable in X 00 . Take e ∈ E and e00 ∈ E 00 with 0 ≤ e00 ≤ e. From 2.1.7 (3) it follows that e00 = x for some x ∈ X. According to 7.1.8 κ(X) is a submodule of the Z (E 00 )-module X 00 . Therefore, there exists an orthomorphism 0 ≤ π ∈ Z (E 00 ) such that e00 = πe = πx . Thus, e00 ∈ λ(E), since πx ∈ κ(X). B 7.1.8. Theorem. For a decomposable bo-complete space with mixed norm (X, E), the following are equivalent: (1) the mixed norm in X is bo-continuous; (2) every closed bo-ideal in X is a bo-band; (3) every bounded functional on X is bo-continuous; (4) the norm in E0 is order continuous. C (1) ⇒ (2): If X ∈ (A) then every closed bo-ideal is bo-closed, and the claim follows from 2.1.6 (3). (2) ⇒ (3): The zero ideal of each bounded functional f ∈ X 0 is a closed bo-ideal, and it remains to apply 4.4.6. ⊥⊥ (3) ⇒ (4): Without loss of generality we may assume that E = X . Take a positive functional ϕ ∈ E 0 . By Theorem 4.4.2, we have the decomposition ϕ = ϕn + ϕs , where ϕn is bo-continuous and ϕs is bo-singular. Assume that ϕs 6= 0. Since the functional x 7→ ϕs ( x ) (x ∈ X) is sublinear, continuous, and nonzero, by the Hahn–Banach Theorem there exists a nonzero functional f ∈ X 0 such that f (x) 6 ϕs ( x ) (x ∈ X). Hence, it is clear that 0 6= f 6 ϕs . By condition (3), the functional f is bo-continuous; consequently, f is o-continuous according to 4.3.2. This contradicts bo-singularity of ϕs ; hence, we must have ϕs = 0 and 0 0 ϕ = ϕn . Thus, E 0 = E+ − E+ ⊂ En0 ; therefore, the norm in E is order continuous according to 1.5.3 (8). (4) ⇒ (1): This is obvious. B

Operators in Spaces with Mixed Norm

297

7.1.9. Now consider spaces with mixed norm which consist of dominated operators. Let (X, E) and (Y, F ) be spaces with mixed norm. Denote by M (X, Y ) the set of all linear operators from X into Y which admit bounded dominants. In other words, T ∈ M (X, Y ) means that T is linear and T x 6 S( x ) (x ∈ X) for some 0 6 S ∈ L (E, F ). The dominant norm µ(T ) of the operator T ∈ M (X, Y ) is defined by the formula  µ(T ) := inf kSk : S ∈ L (E, F ) ∩ maj(T ) . In the special case X
= E and Y = F ,
we evidently obtain the space of regular operators R(E, F ), ρ := M (X, Y ), µ . Observe that if E and F are Banach lattices then M (X, Y ) = M (X, Y ), since R(E, F ) = L∼ (E, F ) in this case. (1) If X is decomposable and F is order complete then the norm µ is mixed: µ(T ) = k T k. If,
moreover, E and F are Banach lattices and Y is a BKS then M (X, Y ), R(E, F ) is a decomposable bo-complete space with mixed norm. C The claim follows from 4.1.2 and 4.2.6. B A linear operator T : X → Y is called predominated if there exists an operator 0 6 S ∈ Lr (E, F 00 ), called a predominant, such that T x 6 S( x ) (x ∈ X). Assign  µp (T ) = inf kSk , where the infimum is taken over all S’s indicated. The space of all predominated operators is denoted by M p (X, Y ).

(2) If X is decomposable then the norm µp is mixed: µp (T ) = T , since the lattice F 00 is order complete and the least predominant of T exists in R(E, F 00 ). 7.1.10. We conclude this section by considering several simple properties of dominated operators. We start with the following question: When does passing to the dual of an operator commute with the taking of the exact dominant of this operator? Let (X, E) and (Y, F ) be decomposable spaces with mixed norm and take an operator T ∈ M (X, Y ). Then the dual T 0 ∈ L (Y 0 , X 0 ) is defined. In case F is order complete, the operator T exists. 0 (1) If T ∈ M (X, Y ) then T 0 ∈ M (Y 0 , X 0 ) and T 0 6 T . C For arbitrary y 0 ∈ Y 0 and e ∈ E+ ,

e, T 0 y 0





= e, y 0 ◦ T 6 h T e, y 0 i = e, T 0

0

 y0 .

It follows that T 0 y 0 6 T ( y 0 ) (y 0 ∈ Y 0 ); therefore, T ∈ M (Y 0 , X 0 ) and T 0 6 0 T .B The inequality between the exact dominants may be strict. However, this is impossible for an order continuous dominated operator. More precisely, the following proposition holds.

298

Chapter 7

(2) Suppose that E and F are Banach lattices with En0 and Fn0 pointseparating. Suppose further that F is order complete. Let (X, E) and (Y, F ) be decomposable Banach spaces with mixed norm. If T ∈ Mn (X, Y ) then T 0 (Yn0 ) ⊂ Xn0 and 0 T 0 f 0 = T f 0 (f 0 ∈ Fn0 ). C The desired inclusion is an obvious consequence of the bo-continuity of T . 0 ∗ By Theorem 4.3.2, T is o-continuous; therefore, T (Fn0 ) ⊂ En0 . Let T ∗ and T 0 Fn0 , respectively. be the restrictions of the operators T 0 and T onto Yn0 and Since Fn0 is an o-ideal in F 0 and Yn0 := y 0 ∈ Y : y 0 ∈ Fn0 ; therefore, T ∗ is the restriction of T 0 onto Fn0 . This follows readily from 2.2.2. ∗ ∗ Hence, in view of (1), we have T ∗ 6 T . It is clear that the operator T is o-continuous. Then T ∗ is bo-continuous too, since it has an o-continuous dominant.
0 0 0 0 0 0 ∗ 0 Now, for the operator S := T : Yn → Xn we have S (Xn )n ⊂ (Yn )n and, in ∗ ∗ view of what was already proven, S ∗ 6 S , i.e., T ∗∗ 6 T ∗ . It is clear that ∗∗ the restriction X coincides with T . Since E is an o-ideal in (En0 )0n  00 of T 00 onto and X = x ∈ X : x00 ∈ E , T is the restriction of T ∗∗ onto E. Thus, ∗ 0 then T 6 T ∗ E . Now if e ∈ E+ and f 0 ∈ Fn+



  

0 ∗ e, T f 0 = h T e, f 0 i 6 T ∗ e, f 0 = e, T ∗ f 0 . ∗

∗

This means that T 6 T ∗ , which yields T ∗ = T in view of what was obtained above. B (3) In the setting of Proposition (2), suppose further that X and Y 0 satisfy condition (A). If T ∈ Mn (X, Y ) then T 0 ∈ Mn (Y 0 , X 0 ) and T 0 = T . 7.1.11. Suppose that E and F are Banach lattices. The following assertions hold: (1) Every dominated operator T is bounded and kT k 6 µ(T ). (2) If T ∈ M (X, Y ) and U ∈ M (Y, Z) then U T ∈ M (X, Z) and µ(U ◦ T ) 6 µ(U ) · µ(T ). (3) If T ∈ M (X, Y ) then T 0 ∈ M (Y 0 , X 0 ) and µ(T 0 ) 6 µ(T ). (4) Let X and Y be decomposable. An operator T ∈ L (X, Y ) satisfies T 0 ∈ M (Y 0 , X 0 ) if and only if T ∈ M p (X, Y ); in this case, µp (T ) = µ(T 0 ). C Take a bounded operator T : X → Y
. Let T 0 ∈ M (Y 0 , X 0 ) and U := T 0 . Then U 0 is a dominant for T 00 see 7.1.9 (1) . Denote by S the restriction of U 0 onto E. It is clear that S ∈ L+ (E, F 00 ) and, in view of 7.1.5, T x 6 S( x ) (x ∈ X). This means that T ∈ M (X, Y ); furthermore, µp (T ) 6 kSk 6 kU 0 k = kU k = µ(T 0 ). Conversely, suppose that T ∈ M p (X, Y ) and an operator S ∈ L+ (E, F 00 ) serves as a predominant for T . As above, taking account of 7.1.4 and 7.1.5, we conclude that the restriction of S 0 onto F 0 is a dominant for T 0 . Therefore, T 0 ∈ M (Y 0 , X 0 ) and µ(T 0 ) ≤ kS 0 k = kSk. Hence, µ(T 0 ) 6 µp (T ) and, finally, µp (T ) = µ(T 0 ). B

Operators in Spaces with Mixed Norm

299

(5) If there is a positive projection with norm 1 from F 00 onto the image of F under the canonical embedding F → F 00 , then M (X, Y ) = M p (X, Y ) and µ = µp . If, furthermore, X and Y are decomposable then, for every bounded operator T : X → Y , the containments T ∈ M (X, Y ) and T 0 ∈ M (Y 0 , X 0 ) are equivalent and, moreover, µ(T ) = µ(T 0 ). C Let π : F 00 → F be a positive projection such that kπk = 1. Take a bounded operator T : X → Y . If an operator S ∈ L+ (E, F 00 ) is a predominant for T ∈ L(X, Y ) then π ◦ S is a dominant for T and kπ ◦ Sk 6 kSk. Therefore, T ∈ M p (X, Y ) implies T ∈ M (X, Y ) and µp (T ) > µ(T ). The reverse inequality is obvious. The second part of the assertion follows from the first and (4). B 7.2. Summing Operators Various classes of operators studied in functional analysis are defined in mixed terms of norm and order. A considerable part of the corresponding problems falls naturally within the theory of LNSs. Furthermore, new relations and possibilities arise. In this section, we consider a small fragment connected with the notion of summing operator. 7.2.1. Let E be a Banach lattice and let f : Rn → R be a continuous positively homogeneous function. Since E is uniformly complete, we may apply 1.5.7, so that there is a mapping f^ : E n → E continuous positively homogeneous and satisfying the condition f^(t1 e, . . . , tn e) = ef (t1 , . . . , tn )

(t1 , . . . , tn ∈ Rn , e ∈ E+ ).

The element f^(e1 , . . . , en ) belongs to the ideal generated by |e1 |, . . . , |en |. 1 Pn p p . For p = ∞, we define Take a function fp,n : (t , . . . , t ) → 7 ( |t | ) 1 n k k=1 
fp,n (t1 , . . . , tn ) = max |tk | : k := 1, . . . , n . The function is positively homogeneous and continuous. Let Qp,n := f^p,n be the corresponding mapping from En into E. We use a more intricate but expressive notation: !p1 n X p (e1 , . . . , en ∈ E). Qp,n (e1 , . . . , en ) =: |ek | k=1

Since h : H (Rl ) → E is a lattice homomorphism with h(dxj ) = ej (j := 1, . . . , l), it is easy to see that 1 !1 !∞ n n n n X X X _ 1 ∞ |ek | |ek |, |ek | = |ek |. = k=1

k=1

k=1

k=1

The following explicit description for this element can be taken as a definition:

300

Chapter 7

(1) For all e1 , . . . , en ∈ E and reals 1 ≤ p, q ≤ ∞ with 1/p + 1/q = 1 the following is valid:   !p1 !q1 n n n X  X X p q |ek | = sup ≤1 . λ k ek : |λk |   k=1

k=1

k=1

C Put u := |e1 | ∨ · · · ∨ |en | and observe that the set

M :=

 n X 

n X

λ k ek :

k=1

!q1 |λk |

q

≤1

  

k=1

is precompact in the AM -space E(u). According to the Arzel`a–Ascoli Theorem for a precompact set M ⊂ C(Q) the supremum exists pointwise and defines a continuous function (sup M )(t) = sup{f (t) : f ∈ M }. By virtue of the Brothers Kre˘ın–Kakutani Theorem 1.5.6 (2) the supremum e := sup M exists in E(u). MorePn p 1 over, e = ( k=1 |ek | ) p in E(u) and we are done, since supremum of M in E coincides with e. B (2) Let S : E → F be a positive operator and e1 , . . . , en ∈ E. Then n X

!p1 p

≤S

|Sek |

n X

!p1 |ek |

p

.

k=1

k=1

If, in addition, S preserves suprema (= S is order continuous lattice homomorphism) then we have equality instead of inequality. C Since S is linear and increasing, we use (1) and proceed as follows: n X k=1

!p1 p

|Sek |

 

!q1

  q = sup S λk ek : |λk | ≤1   k=1 k=1    !q1 n n  X X q  ≤ S sup λ k ek : |λk | ≤1    !

n X

k=1

=S

n X

n X

k=1

1 p

! p

|ek |

.

k=1

The second assertion follows immediately from (1). B

Operators in Spaces with Mixed Norm

301

7.2.2. We need two more properties of the mapping Qp,n . (1) Let E be a Banach lattice, e1 , . . . , en ∈ E, e01 , . . . , e0n ∈ E 0 , and 1 ≤ p ≤ ∞. Then the H¨older inequality holds: n X hek , e0k i ≤

*

n X

!p1 p

|ek |

n X

,

k=1

k=1

!q1 + |e0k |

q

.

k=1

C First we assume that E = C(Q) for a compact topological space Q. Then µk := |e0kP | (k = 1, . . . , n) are Radon measures absolutely continuous with respect n to µ := k=1 µk . By the Radon–Nikod´ ym Theorem we may find functions gk ∈ L1 (Q, µ) such that dµk = gk dµ (k = 1, . . . , n). Using the H¨older inequality we obtain n X

|hek , e0k i|

≤

h|ek |, |µk |i =

n Z X

n Z X

|ek | dµk

k=1

k=1

k=1

=

n X

n X

Z |ek |gk dµ ≤

k=1

!p1

n X

p

|ek |

k=1

!q1 gk q

dµ.

k=1

It remains to note that n X

!q1 gk q

µ=

k=1

n X

!q1 q

|e0k |

,

k=1

since L1 (Q, µ) is isomorphic to the band in C(Q)0 generated by µ. In the case of a general Banach lattice E we set u := |e1 | + · · · + |en | and note that ek ∈ E(u). Let J be an isomorphism from C(Q) onto E(u) and fk := J −1 ek . From the above we deduce n X

|hek , e0k i| =

k=1

n X

* |hfk , J 0 e0k i| ≤

k=1

n X

!p1 p

|fk |

n X

,

!q1 + q

|J 0 e0k |

=: t.

k=1

k=1

Since J preserves suprema and J 0 is positive, we may apply 7.2.1 (2) to derive: * t≤

n X k=1

!p1 p

|fk |

, J0

n X k=1

!q1 + q |e0k |

* =

n X k=1

!p1 p

|Jfk |

,

n X k=1

!q1 + q |e0k |

. B

302

Chapter 7 (2) Let E be a Banach lattice, e1 , . . . , en ∈ E and 1 ≤ p ≤ ∞. Then 

!p1

X

n X

n

p

|ek | = sup hek , e0k i :



k=1

k=1



!1

X



n 0 q q

|ek | ≤ 1 .



k=1

C By virtue of 3.1.2 (8) we obtain

*  !p1 !p1 +

n

n   X

X

p p 0 0 0 0

= sup |e | |e | , e : e ∈ E , ke k ≤ 1 k k +

 

k=1

k=1   * n + !q1 m m n X  X X X q 0 = sup sup λ0k ek , fl0 : fl0 ∈ E+ , e0 = fl0 , |λ0k | ≤1   e0 l=1 k=1 l=1 k=1 * + n m X X = sup sup ek , λ0k fl0 . k=1

It remains to observe that if e0k :=

Pm

l=1

l=1

λ0k fl0 then

!1

n

m

X 0 q q

X 0

fl = ke0 k ≤ 1. B |ek | ≤

k=1

l=1 7.2.3. Suppose that 1 ≤ q ≤ p ≤ +∞. A linear operator T : X → Y is called (p, q)-summing if there exists a number κ > 0 such that, for every finite set {x1 , . . . , xn } ⊂ X, the following inequality holds:

!p1 !q1

n

n X

X

p q

.

T x x ≤ κ k k





k=1

k=1 Denote by σpq (T ) the infimum of the set of all κ’s indicated. Let Gp,q (X, Y ) be the space of all (p, q)-summing operators from X into Y . We can see from the definition that every summing operator T is bounded and kT k ≤ σpq (T ). It is also clear that σpq is a norm in Gp,q (X, Y ). The triangle inequality for σpq follows Pn p 1 from sublinearity of the mapping Qp : (f1 , . . . , fn ) 7→ ( k=1 |fk | ) p acting from Fn

Operators in Spaces with Mixed Norm

303

into F . The norm σpq can be calculated by the following formulas:

  !p1 !q1

n

n   X X



p q

: x ∈ X, n ∈ N, ≤ 1 , T x x σpq (T ) = sup k k k

 

k=1

k=1   p1

 P

n  

p     T x

k   n  k=1 

X



σpq (T ) = sup . : x ∈ X, n ∈ N, x = 6 0 k k q1

P   n  

q k=1     xk

 

k=1

In case p = q = 1, we use the term a summing operator and write σ instead of σ11 and G instead of G11 . 7.2.4. Theorem. Let X and Y be spaces with mixed norm and suppose that Y
is a Banach space. Then Gp,q (X, Y ), σpq is a Banach space. C Let (Tn ) be a fundamental sequence in Gpq (X, Y ). Then (Tn ) is also fundamental in the weaker norm of the space L (X, Y ). The space is complete; consequently, there exists an operator T ∈ L (X, Y ) such that limn→∞ kT − Tn k = 0. For an ε > 0, choose a number k0 ∈ N so that σpq (Tn − Tm ) ≤ ε for all n, m ≥ k0 . Then, for each finite set x1 , . . . , xl ∈ X we have

!q1 !p1

X

X

l l



q p

≤ 1 ⇒ αnm :=

≤ ε. xk Tn xk − Tm xk



k=1

k=1

Observe that the operator Qp,l is continuous, since the following inequality holds due to sublinearity: Qp,l (f1 , . . . , fl ) − Qp,l (g1 , . . . , gl ) ≤ Qp,l (f1 − g1 , . . . , fl − gl ). Taking it into account that a vector norm is a continuous operator from Y into F , we conclude !p1 !p1 l l X X p p lim Tn xk − Tm xk = Tn xk − T xk . m→∞

k=1

k=1

Thus, lim αnm

m→∞

!p1

X

l

p

≤ ε. = T x − T x n k k

k=1

Arbitrariness of the choice of x1 , . . . , xn implies σpq (Tn − T ) ≤ ε for n ≥ k0 . Hence, it is clear that T ∈ Gp,q (X, Y ) and σpq (Tn − T ) → 0 as n → ∞. B

304

Chapter 7

7.2.5. If T ∈ G(X, Y ), S ∈ M (X, Y1 ), and U ∈ M (X1 , X), then ST U ∈ Gpq (X1 , Y1 ) and σpq (ST U ) ≤ µ(S)σpq (T )µ(U ). C Let S0 and U0 be arbitrary dominants of the operators S and U respectively. Then, by applying 7.2.1 (2) twice for arbitrary x1 , . . . , xn ∈ X1 , we obtain:

!p1 !p1

X

X

n n



p p

≤ (S0 T V xk ) ST V xk



k=1

k=1



!p1 !p1

X

X

n n



p p

≤ kS0 kσpq (T )kU0 k

. ≤ kS0 kσpq (T ) V xk xk



k=1

k=1

Thus, ST U ∈ Gpq (X1 , Y1 ) and σpq (ST U ) ≤ kS0 kσpq (T )kU0 k. Taking the infimum over S0 and U0 , we come to the desired inequality. B 7.2.6. Theorem. Let X and Y be Banach spaces with decomposable mixed norms. A bounded operator T : X → Y is (p, q)-summing if and only if the dual T 0 of T is (q 0 , p0 )-summing. Furthermore, σpq (T ) = σq0 p0 (T 0 ). C Suppose that T ∈ Gp,q (X, Y ) and take a finite n-tuple y10 , . . . , yn0 ∈ Y 0 . Consecutively applying formulas 7.2.2 (2), 7.1.4, and 7.2.2 (1), we may write down the following chain:



 !10 !q1

X

X

n n q X  0

n



 q 0 0 q 0 0



ek , T yk : ek ∈ E+ , T yk ek

= sup 

≤ 1

k=1

k=1 k=1 

 !q1

X n n X 

q 0 0

= sup sup {hx, T yk i : x ≤ ek } : ek ∈ E+ , ek

≤ 1 

k=1

k=1



 !q1

n n X  X



 q

≤1 ≤ sup T xk , yk0 : xk ∈ X, x k

 

k=1

k=1 *  !p1 !10 + !q1

n

n n p   X X X 0

p p q

≤1 T xk , yk0 : x ≤ sup k

 

k=1

k=1 k=1 

 !p1 !10 !q1

X

X

n n n p   X 0



p q 0 p



T x y x ≤ 1 ≤ sup : k k k



 

k=1

k=1

k=1

!10

X

n p 0

0 p

. = σpq (T ) yk

k=1

Operators in Spaces with Mixed Norm

305

Thus, T 0 ∈ Gq0 ,p0 (Y 0 , X 0 ) and σq0 p0 (T 0 ) ≤ σpq (T ). Conversely, assume that T 0 is a (q 0 , p0 )-summing operator. Then, in view of what was proven, T 00 is (p, q)-summing and σpq (T 00 ) ≤ σq0 p0 (T 0 ). Since the canonical embeddings X → X 00 and Y → Y 00 preserve vector norms (see 7.1.5); therefore, T ∈ Gp,q (X, Y ), and from the definition of σpq (see 7.2.3) it is clear that σpq (T ) ≤ σpq (T 00 ). Thus, σpq (T ) = σq0 p0 (T 0 ). B 7.2.7. Theorem. A linear operator T : X → Y is summing P∞if and only if, for every sequence (vn ) ⊂ X, convergence in norm of the series k=1 vk implies P∞ convergence in norm of the series k=1 T vk . C Necessity is obvious. In order to prove sufficiency, assume that T ∈ / G(X, Y ). Then, for each n ∈ N, we may choose a finite set {vn1 , . . . , vnl(n) } ⊂ X such that

X

l(n)

vnj

j=1



≤ 1,



X

l(n)

T vnj

j=1



≥ 2n .

n Construct a new sequence (uk ) ⊂ X
by letting uk := (1/2 P∞ )vnj for k := l(1) + · · · + l(n − 1) + j n ∈ N, j := 1, . . . , l(n) . Then the series k=1 uk converges in norm, since



l(m)+···+l(m+s)

m+s

X

m+s l(n) X X X 1



1



. u ≤ v ≤ k nj

n n

k=l(1)+···+l(m−1)+1

n=m 2 j=1

n=m 2

At the same time, for the element l(1)+···+l(n)

sn :=

X

T uk

k=l(1)+···+l(n−1)+1

we have ksn k ≥ 1. Consequently, the series

P∞

l(n) 1 X = n T vnj , 2 j=1

k=1

T uk diverges. A contradiction. B

7.2.8. Theorem. Let X be decomposable, and let one of the following conditions be satisfied: (1) E = X and there is a projection with norm 1 from F 00 onto the image 00 of F under the i.e., F enjoys Property (B) and
canonical embedding F → F Property (C) ; (2) F is an order complete AM -space with unity. Then M (V, W ) = G(V, W ) and µ = σ. of sums of the form a(θ) := Pn C Take a T ∈ G(V, W+) and denote by U (e) the set P n T v , where e ∈ E , θ := {v , . . . , v } ⊂ X, and k 1 n k=1 k=1 vk ≤ e. It is clear that ka(θ)k ≤ σ(T )kek, i.e., U (e) is norm-bounded subset in F . If (1) is true, the set U (e) is directed upward and the norm in F is order semicontinuous.

306

Chapter 7

If (2) is true then norm-boundedness is equivalent to order-boundedness. In both cases, T ∈ M (X, Y ). In view of 4.1.5, µ(T ) = sup k T ek = sup kek=1 e≥0

sup kak = σ(T )

kek=1 a∈U (e) e≥0

and the result follows. B 7.2.9. Corollary. Assume that one of the conditions (1) or (2) in 4.2.8 is satisfied. Then, for a linear operator T : X → Y , the following are equivalent: (1) T is a dominated operator; (2) T is a summing operator; P∞ (3) for every sequence (xnP ) ⊂ X, convergence of the series k=1 xk ∞ in E implies convergence of the series k=1 T xk in F . 7.2.10. If the hypotheses of Theorem 4.2.8 are met then the following formula holds:

n

( n )

X

X



µ(T ) = sup T xk : xk ∈ X, n ∈ N, xk ≤ 1 .



k=1

k=1

7.2.11. We indicate some particular cases of the class Gpq (X, Y ) which may be encountered in the literature on functional analysis. (1) (p, q)-convex operators. This term is conventionally used for operators of the class Gp,q (X, Y ) in case Y = F and E = R (i.e., X is a Banach space). An operator T : X → F is called (p, q)-convex if there exists a number κ > 0 such that, for all x1 , . . . , xn ∈ X, the following inequality holds:

!p1 !q1

X

n n X

p q

|T xk | kxk k .

≤κ

k=1

k=1 The terms a p-superadditive or dominated operator are used instead of a (p, 1)convex or (∞, ∞)-convex operator, respectively. (2) (p, q)-concave operators. The class of these operators coincides with Gp,q (X, Y ) in case E = X and F = R (i.e., Y is a Banach space). Thus, an operator T : E → Y is called (p, q)-concave if there exists a number κ > 0 such that, for all e1 , . . . , en ∈ X, the following holds:

!q1 !p1

X

n n X

p q

≤ κ kT ek k |ek |

.

k=1 k=1 For the particular cases of (∞, q)-concave and (1, 1)-concave operators, the respective terms q-superadditive and summing operators are used.

Operators in Spaces with Mixed Norm

307

(3) (p, q)-regular operators. This class of operators arises when two Banach lattices are considered, i.e., when X = E and Y = F . Hence, an operator T : E → F is called (p, q)-regular if there exists a number κ > 0 such that, for all e1 , . . . , en ∈ E, the following inequality holds:

!q1 !p1

X

X n



n q p

≤ κ

.

|e | |T e | k k



k=1

k=1 Theorems 7.2.6 and 7.2.9 imply the following corollaries. (4) An operator T : X → F (with X a Banach space) is (p, q)-convex if and only if T 0 is a (q 0 , p0 )-concave operator. An operator T : E → Y (with Y a Banach space) is (p, q)-concave if and only if T 0 is a (p0 , q 0 )-convex operator. (5) If there is a positive projection with norm 1 from F 00 onto the image of F under the canonical embedding then an operator T : E → F is (1, 1)-regular if and only if T is regular. 7.3. Isometric Classification In this section, we are going to show that lattice-normed spaces provide reasonable grounds for constructing isometric classification of spaces with mixed norm. To this end, we chose a small number of natural questions connected with norms take values AM - and ALp -spaces. 7.3.1. We now turn to the following natural question: Which Banach spaces are linearly isometric to Banach spaces with mixed norm? We confine our study to the case in which the norm lattice is an AM - or an ALp -space. Below, we present several results in this direction. We start with the necessary definitions. Let X be a normed space. Suppose that L (X) has a complete Boolean algebra of norm-one projections B which is isomorphic to B. In this event we will identify the Boolean algebras B and B, writing B ⊂ L (X). Say that X is a normed Bspace if B ⊂ L (X) and for every partition of unity (bξ )ξ∈„ in B the two conditions hold: (1) If bξ x = 0 (ξ ∈ „) for some x ∈ X then x = 0; (2) If bξ x = bξ xξ (ξ ∈ „) for x ∈ X and a family (xξ )ξ∈„ in X then kxk ≤ sup{kbξ xξ k : ξ ∈ „}. Conditions (1) and (2) amount to the respective conditions (10 ) and (20 ): (10 ) To each x ∈ X there corresponds the greatest projection b ∈ B such that bx = 0;

308

Chapter 7 (20 ) If x, (xξ ), and (bξ ) are the same as in (2) then kxk = sup{kbξ xξ k :

ξ ∈ „}. From (20 ) it follows in particular that

n

X

bk x = max kbk xk

k:=1,...,n k=1

for x ∈ X and pairwise disjoint projections b1 , . . . , bn in B. Given a partition of unity (bξ ), we refer to x ∈ X satisfying the condition (∀ ξ ∈ „) bξ x = bξ xξ as a mixing of (xξ ) by (bξ ). If (1) holds then there is a unique mixing x of (xξ ) by (bξ ). In these circumstances we naturally call x the mixing of (xξ ) by (bξ ). Condition (2) maybe paraphrased as follows: The unit ball UX of X is closed under mixing or is mix-complete. 7.3.2. Theorem. For a Banach space X the following are equivalent: (1) X is a decomposable space with mixed norm whose norm lattice is an AM -space with unity; (2) X is a Banach B-space. C (1) ⇒ (2): Appeal to the appropriate definitions and 2.1.3. (2) ⇒ (1): Suppose that X is a Banach B-space and J : B → B is a corresponding isomorphism of B onto the Boolean algebra of projections B. Denote ¯ in the universally complete K-space of all B-valued by E the ideal generated by 1 Pn resolutions of unity (cf. 1.4.3). Take the finite-valued element d := i=1 λi bi ∈ E, where λ1 , . . . , λn ∈ R, the family {b1 , . . . , bn } is a partition of unity in B, and λb stands for the spectral function e : µ 7→ e(µ) ∈ B equalPto the zero of B for µ ≤ λ n and equal to the unity of B for µ > λ. Put J(α) := i=1 λi J(bi ) and note that J(α) is a bounded linear operator in X. Calculating the norm of J(α), obtain kJ(α)k = sup kJ(α)xk = sup kxk≤1

=

sup {kπl xk|λl |}

kxk≤1 l=1,...,n

sup sup{kπl xk |λl | : kxk ≤ 1} = max{|λ1 |, . . . , |λn |}. l=1,...,n

On the other hand, the norm kαk∞ of a member α of the AM -space E coincides with max{|λ1 |, . . . , |λn |} too. Hence, J is a linear isometry of the subspace E0 of finite-valued members of E to the algebra of bounded operators L (X). It is also clear that J(αβ) = J(α) ◦ J(β) for all α, β ∈ E0 . Since E0 is norm dense in E and L (X) is a Banach algebra; therefore, we may extend J by continuity to an isometric isomorphism of E onto a closed subalgebra of L (X). Assigning xα := αx := J(α)x for x ∈ X and α ∈ E, we make X into a unital E-module so that kαxk ≤ kxk kαk∞

(α ∈ E, x ∈ X).

Operators in Spaces with Mixed Norm

309

Moreover, αUX + βUX ⊂ UX for |α| + |β| ≤ 1. Define the mapping p : X → E+ by the formula p(x) := inf{α ∈ E+ : x ∈ αUX } (x ∈ X), with the infimum taken in the K-space E. If p(x) = 0 then to ε > 0 there are a partition of unity (πξ ) ⊂ B and a family (αξ ) ⊂ E+ such that πξ αξ ≤ ε1 and x ∈ αξ UX for all ξ. But then πξ x ∈ πξ αξ UX ⊂ εUX . Since the unit ball UX is closed under mixing; therefore, x = mix(πξ xξ ) ∈ εUX . The arbitrary choice of ε > 0 implies x = 0. If x ∈ αUX and y ∈ βUX for some α, β ∈ E+ , then, putting γ := α + β + ε1, we may write down x + y = γ (γ −1 x + γ −1 y) ∈ γ (γ −1 αUX + γ −1 βUX ) ⊂ γUX . Consequently, p(x + y) ≤ α + β + ε1; and taking the infimum over α, β, and ε yields p(x + y) ≤ p(x) + p(y). Furthermore, granted π ∈ B and x ∈ X, observe the equalities πp(x) = inf{πα : 0 ≤ α ∈ E, x ∈ αUX } = inf{α ∈ E+ : πx ∈ αUX } = p(πx). But then, for α =

P

λi πi , with {π1 , . . . , πn } a partition of unity in B, we see that

p(αx) =

X

πi p(λi x) =

n X

πi |λi |p(x) = |α|p(x).

ı=1

Hence, p(αx) = |α|p(x) for all α ∈ E. Therefore, (X, p, E) is a decomposable lattice-normed space. Show now that the norm of X is a mixed norm; i.e., kxk = kp(x)k∞ (x ∈ X). Take 0 6= x ∈ X and put y = x/kxk. Then y ∈ UX and p(y) ≤ 1. Consequently, p(x) ≤ kxk1 or kp(x)k∞ ≤ kxkk1k∞ = kxk. Conversely, given ε > 0, we may find a partition of unity (πξ )ξ∈„ in Pr(E) and a family (αξ )ξ ⊂ E+ such that πξ αξ ≤ p(x) + ε1 ≤ (kp(x)k∞ + ε)1 and x ∈ αξ UX (ξ ∈ „). Whence πξ xξ ∈ πξ αξ UX ⊂ (kp(x)k∞ + ε)πξ 1UX ⊂ (kp(x)k∞ + ε)UX . Consequently, kπξ xξ k ≤ kp(x)k∞ + ε. Considering the arbitrary choice of ε > 0 together with 7.3.1 (2), we deduce kxk ≤ kp(x)k∞ . B 7.3.3. A normed B-space X is B-cyclic if we may find in X a mixing of each norm-bounded family by any partition of unity in B. Taking 7.3.2 into consideration, we may assert that X is a B-cyclic normed space if and only if, given a partition of unity (bξ ) ⊂ B and a family (xξ ) ⊂ UX , we may find a unique element x ∈ UX such that bξ x = bξ xξ for all ξ.

310

Chapter 7

(1) Theorem. A Banach space is linearly isometric to a bo-complete space with mixed norm, whose norm lattice is an order complete AM -space with unity, if and only if it is B-cyclic with respect to some complete Boolean algebra B of projections. C In view of 7.3.2 it is sufficient to observe that a Banach B-space is B-cyclic if and only if it is disjointly complete as a lattice-normed space. B A Boolean algebra B is called Bad´e-complete if it is order complete and, for every increasing net of projections (bα ) ⊂ B, from b = sup bα it follows that hbx, x0 i = limα hbα x, x0 i whenever x ∈ X and x0 ∈ X 0 . We say that a set C ⊂ X is B-bounded if bounded is the following set of reals:

( n )

X

bk xk : x1 , . . . , xn ∈ C, b1 , . . . , bn ∈ B, n ∈ N, bk ◦ bl = 0 (k 6= l) .

k=1

A Banach space X is called (B, p)-cyclic (1 ≤ p < ∞) if the following conditions are satisfied: (a) there is a complete Boolean algebra B of norm-one projections in X; (b) kbx + b⊥ ykp = kbxkp + kb⊥ ykp for all b ∈ B and x, y ∈ X; (c) for each B-bounded family in X, there exists a mixing by an arbitrary partition of unity (with the same index set) in B; (2) If a Banach space X is (B, p)-cyclic and B is Bad´e-complete then, for every decreasing net (bα ) ⊂ B, from inf α bα = 0 it follows that limn kbα xk = 0 for all x ∈ X. C Let (bα ) be a decreasing net of projections; suppose that inf bα = 0 but kbα xk does not converge to zero. Without loss of generality, we may assume that kbα xk > ε > 0 for all α. Since the Boolean algebra is Bad´e-complete, the net bα x converges to zero in the weak topology σ(X, X 0 ). By the P Mazur Theorem, there Pn n exists a convex combination y = k=1 λk bα(k) x, λk ∈ R+ , k=1 λk = 1, such that kyk < (ε/2). If bα ≤ bα(k) , k = 1, . . . , n, then bα y = bα x and, by condition (b) in the definition of (B, p)-cyclicity, we have:

p

= kykp < (ε/2)p . εp < kbα xkp ≤ kbα ykp + b⊥ y α A contradiction is obtained. Hence, lim kbα xk = 0. B (3) Under the hypotheses P of Proposition (1), there is a unique mixing x = mixξ∈„ (bξ xξ ). Moreover, x = ξ∈„ bξ xξ and kxkp =

X ξ∈„

kbξ xξ kp .

Operators in Spaces with Mixed Norm

311

7.3.4. Theorem. A Banach space is linearly isometric to a bo-complete space with mixed norm whose norm lattice is an ALp -space with 1 ≤ p < +∞ if and only if the space is (B, p)-cyclic with respect to some Bad´e-complete Boolean algebra of projections. C We only need to prove sufficiency. Suppose that X is a (B, p)-cyclic Banach space with 1 ≤ p < ∞ and let B be a Bad´e-complete Boolean algebra of projections in X. If b ∈ B is a nonzero projection then hbx, x0 i 6= 0 for suitable x ∈ X and x0 ∈ X 0 . At the same time, the function b 7→ hbx, x0 i (b ∈ B) is additive and, due to Bad´e-completeness, it is o-continuous as well. Thus, there exists a point-separating set of o-continuous measures on B. Let Z be a universally complete K-space of all resolutions of unity in the algebra B. Then the base of Z is isomorphic to B; consequently, there exists an order-dense ideal in Z, on which an essentially positive o-continuous functional ˆ is defined, see 1.4.10. Denote by L1 (ˆ) the greatest order-dense ideal onto which ˆ can be extended by S∞ ∞ o-continuity. Assign L (ˆ) := n=1 [−n1, n1]; i.e., L∞ (ˆ) is the o-ideal in Z generated by the unity 1 ∈ Z. The mapping that associates with an element z ∈ L1 (ˆ)
the functional α 7→ ˆ(αz) α ∈ L∞ (ˆ) , is a linear and order isomorphism of L1 (ˆ) onto L∞ (ˆ)0n , see 3.4.8. In the same way as in 7.3.2, we may establish that X can be endowed with the structure of a faithful unital
module over the ring L∞ (ˆ); furthermore, kαxk ≤ kαk · kxk α ∈ L∞ (ˆ), x ∈ X . Now, take an arbitrary element x ∈ X and define a function ϕx : b 7→ kbxkp (b ∈ B). From the (B, p)-cyclicity condition (see 7.3.3 (b)) it is clear that ϕx is additive and o-continuous. Given α ∈ L∞ (ˆ), assign kαk Z

ˆx (α) :=

λdϕx (eα λ ),

−kαk

(eα λ)

where is the spectral function of the element α, and the integral is defined as the r-limit of the integral sums k−1 X

α ln ϕx (eα λn+1 − eλn ) ln ∈ [λn , λn+1 ),

n=−k

over refining partitions of the interval [−kαk, kαk] of the real line: −kαk =: λ−k < λ−k+1 < · · · < λk−1 < λk = kαk. Then ˆx is a positive o-continuous functional on L∞ (ˆ); consequently, there exists a unique positive element z ∈ L1 (ˆ) such that ˆx (α) = ˆ(αz) (α ∈ L∞ (ˆ)). Define √ p x := z. Introduce an ALp -space Lp (ˆ) by the following formulas: 
1
Lp (ˆ) := z ∈ Z : |z|p ∈ L1 (ˆ) , kzkp := ˆ(|z|p ) p z ∈ Lp (ˆ) .

312

Chapter 7

Thus, a mapping · : X → Lp (ˆ) is defined so that p

kbxkp = ˆ(b x )

(b ∈ B, x ∈ X).

Using the last relation, show that k·k is a decomposable norm and X is a space

with mixed norm. First of all, observe that kxk = x p (x ∈ X). In particular, if x = 0 then x = 0. Next, for every partition of unity {b1 , . . . , bn } ⊂ B, every set {x1 , . . . , xn } ⊂ X, and every element b ∈ B, we have (see 7.3.3 (b)): ˆ b

n X

p

!

bk x k

k=1

p ! n n n

X

X X

p = bbk xk = . kbbk xk kp = ˆ b bk x k

k=1

k=1

k=1

Since the projection b ∈ B is arbitrary and the elements bk xk are pairwise disjoint in Lp (ˆ), we have n X

bk x k =

k=1

n X k=1

!p1 bk xk

p

=

n X

bk x k .

k=1

Hence, by letting α := λ1 b1 + · · · + λn bn and xk := λk x (k := 1, . . . , n), we obtain αx = |α| · x . The latter equality is true for all α ∈ L∞ (ˆ). Indeed, if αn → α in L∞ (ˆ) then we have p

p

ˆ(b αn x ) = kbαn xkp → kbαxkp = ˆ(b αx ) for all x ∈ X and b ∈ B. Therefore,

p ˆ(b αx ) = ˆ b(|α| x )p = lim ˆ b( αn x )p . Since b ∈ B was chosen arbitrarily, αx = |α| x . Now, take x, y ∈ X and numbers 0 ≤ α, β ∈ R such that α + β ≤ 1. Employing the triangle inequality for the norm k · kp and the H¨older inequality for finite sums, we obtain 1 1 p p
p
ˆ(b αx + βy ) ≤ ˆ(b αx ) p + ˆ(b βy ) p  1 1 p 1 1 p
p p
p 1− p 1− p = α αˆ(b x ) + β βˆ(b y )
0 0
1 1 p p ≤ αˆ(b x ) + βˆ(b y ) α(1− p )p + β (1− p )p p p
≤ ˆ b(α x + β y ) .

Operators in Spaces with Mixed Norm

313 p

p

p

Using again arbitrariness of b ∈ B, we infer that αx + βy ≤ α x + β y . If α := α1 b1 + · · · + αn bn and β := β1 b1 + · · · + βn bn , with αk , βk ∈ R+ , αk + βk = 1 (k := 1, . . . , n), then αx + βy

p

=

n X

p

bk (αk x + βk y)

=

k=1

≤

n X

bk αk x

n X

bk αk x + βk x

k=1 p

+ βk y

p


=αx

p

p

+β y .

k=1

As above, using uniform approximation by finite-valued elements of Lp (ˆ), we may say that the inequality is true for α, β ∈ L∞ (ˆ)+ whenever α + β ≤ 1. Now, assign γ := x + y + 2ε1, α := γ −1 ( x + ε1), and β := γ −1 ( y + ε1), ε > 0. It is clear that α ≥ 0, β ≥ 0, and α + β ≤ 1. Moreover, if x1 := ( x + ε1)−1 x and y1 := ( y + ε1)−1 y then x1 ≤ 1 and y1 ≤ 1; consequently, γ −1 (x + y) = αx1 + βy1 ≤ (α x

p

p

1

+ β y ) p ≤ 1.

Thus, x + y ≤ γ. Passing to the limit as ε → 0, we obtain the triangle inequality for x .
p Thus, X, · , L (ˆ) is an LNS with decomposable norm and, furthermore,



x = kxk (x ∈ X). In view of 7.1.2, X is br-complete. In order to prove p disjoint completeness, take a family (xξ )ξ∈„ that is vector norm-bounded in X. If ‚ ⊂ „ is a finite set and (bν )ν∈‚ is a partition of unity in B, then

p

X

X X

p
bν xν = kbν xν kp = ˆ bν x ν ≤ ˆ(ep ),

ν∈‚

ν∈‚

ν∈‚

where e := supξ∈„ xξ . Hence, it is clear that the family (xξ ) is B-bounded; therefore, there exists
a mixing mix(πξ xξ ) by an arbitrary partition of unity (πξ ). Finally, X, · , Lp (ˆ) is a BKS and the original norm in X is a mixed norm. B 7.3.5. Theorem. A Banach space Y is linearly isometric to␣Lp (µ, X), 1 6 p < ∞, for some finite measure µ and Banach space X, if and only if there is a closed subspace Z in Y and a Bad´e-complete Boolean algebra B of projections in Y with the following properties: (1) Y is (B, p)-cyclic; (2) X and Z are linearly isometric; Pn (3) the set of all sums of the form k=1 πk zk , with π1 , . . . , πn ∈ B, z1 , . . . , zn ∈ Z, and n ∈ N, is dense in Y ;

314

Chapter 7

(4)for every 0 6= π ∈ B, the function z 7→ kπzk is a positive constant on the sphere z ∈ Z : kzk = 1 . C Necessity of these properties is checked directly. To proof sufficiency, suppose that Y, Z and B satisfy (1)–(4). By virtue of Theorem 7.3.4 we may assume that Y is a space with mixed norm, its norm lattice is Lp (µ) for some measure µ, and the relation  p1 Z p (y ∈ Y, π ∈ B) ||πy|| = π y dµ holds. Put e := z /||z||, z ∈ Z. By hypothesis e does not depend on z, since Z

Z

p

π( z /||z||)p dµ = (||πz||/||z||)p = const

πe dµ =

for each 0 6= π ∈ B. From this we may also see that πe 6= 0 for every nonzero π; thus, ep is an order-unity in L1 (µ). Put ν := ep µ and y 0 := y /e, y ∈ Y . Clearly, ν is a finite measure and (Y, Lp (ν)) is a bo-complete lattice-normed space, since y 7→ y 0 is an Lp (ν)-valued decomposable norm in Y . Moreover, Z

p

||πy|| = and in particular 

||πz|| ||z||

p

(y ∈ Y, π ∈ B)

π y 0 dν

p

1 = ||z||

Z

p

π z 0 dν = ν(π).

Let J denote a linear isometry from Z onto X. Take y = ∈ Z and π1 , . . . , πn ∈ B pairwise disjoint and define J(y) :=

n X

Pn

k=1

πk zk with z1 , . . . , zn

πi J(zi ) ∈ Lp (ν),

i=1

where πi x denotes the vector-function with constant value x on πi and vanishing in πi⊥ . Then, using the fact that Y is Bp -cyclic, we may derive Z kJ(y)k =

p 0

J(y) dν

 p1 =

X n

ν(πi )kzi k

i=1

=

X n i=1

kπi zi kp

 p1

n

X

= πi zi = kyk.

i=1

p

 p1

Operators in Spaces with Mixed Norm

315

It follows that J is a linear isometry of dense subspaces, which can be extended by continuity up to an isometry of spaces Y and Lp (ν, X). It is also clear that J preserves the vector norm, i.e., J is an isometry also in the sense of vector norms. B Observe a corollary to Theorem 7.3.5. Let µ and ν be finite measures. Denote by Lp,q (µ ⊗ ν) the space of measurable functions u of two variables having finite mixed norm  p1 Z Z  pq kukpq := |u(s, t)|q dµ(s) dν(t) . Equivalent functions are conventionally identified. 7.3.6. Theorem. A Banach space Y is linearly isometric to the space Lp,q (µ⊗ ν) for some finite measures µ and ν if and only if there are Bad´e-complete Boolean algebras A and B of projections in Y and an element e ∈ Y such that the following are satisfied: (1) Y is (B, p)-cyclic; (2) for all π ∈ B, disjoint projections ρ, σ ∈ A , and numbers α, β, the equality kαπρe + βπσekq = kαπρekq + kβπσekq holds; (3) the linear span of the set {πρe : π ∈ B, ρ ∈ A } is dense in Y ; (4) for every 0 6= π ∈ B, the function ρ 7→ kπρek/kρek, 0 6= ρ ∈ A , is a positive constant. C Denote by Z the closure of the linear span of the set {ρe : ρ ∈ A }. It follows from (3) that the set X n

πi zi : π1 , . . . , πn ∈ B, z1 , . . . , zn ∈ Z, n ∈ N



i=1

Pn is dense in Y . Let 0 6= π ∈ B and z = i=1 αi ρi e, where ρ1 , . . . , ρn ∈ A are pairwise disjoint and α1 , . . . , αn ∈ R. Using (2) and (4), we may deduce n P

kαi πρi ekq kα1 πρ1 ekq kπzkq kαn πρn ekq i=1 = · · · = = = . n P kα1 ρ1 ekq kαn ρn ekq kzkq q kαi ρi ek i=1

Now, we see that the function z 7→ kπzk is a positive constant on the unit sphere of Z whatever the projections 0 6= π ∈ B may be. Let Z+ be the closure of the conic hull of {ρe : ρ ∈ A }. Then the space Z, ordered by the cone Z+ , is a Banach Pn lattice. q Moreover, Z is an AL -space with weak order-unity e. Indeed, if z1 = i=1 αi ρi e

316

Chapter 7

Pn and z2 = j=1 β j σj e, with 0 ≤ α1 , . . . , β m ∈ R, and ρ1 , . . . , ρn , σ1 , . . . , σm ∈ A pairwise disjoint then, in view of (2), q

kz1 + z2 k =

n X

i

q

kα ρi ek +

i=1

n X

kβ j σj ekq = kz1 kq + kz2 kq .

j=1

It should be also noted that elements of the form z1 and z2 constitute a dense subset in the cone Z+ and arbitrary u1 , u2 ∈ Z are disjoint if and only if there exists a projection ρ ∈ A such that ρu1 = u1 and ρu2 = 0. In view of Theorem 1.5.6 (3) Z is isometrically and latticially isomorphic to the Banach lattice Lq (µ) for a suitable finite measure µ. It remains to apply Theorem 7.3.5 with X := Lq (µ) and observe that the Banach lattices Lp (ν, Lq (µ)) and Lp,q (µ × ν) are isometrically isomorphic. B 7.4. Kaplansky–Hilbert Modules In this section we introduce the class of AW ∗ -modules as bo-complete Banach modules with mixed norm, consider some important structural properties, and establish that every such module may be constructed, up to isomorphism, by forming direct sums and “smearing” of a Hilbert space over an extremal compact space. 7.4.1. We recall some preliminaries concerning complex algebras. Note also that by an algebra we always mean a unital associative algebra. An involutive algebra or ∗-algebra A is a complex algebra with involution, i.e. a mapping x 7→ x∗ (x ∈ A) satisfying the conditions: (1) x∗∗ = x (x ∈ A); (2) (x + y)∗ = x∗ + y ∗

(x, y ∈ A);

(3) (λx)∗ = λ∗ x∗

(λ ∈ C, x ∈ A);

(4) (xy)∗ = y ∗ x∗

(x, y ∈ A).

An element x of an involutive algebra A is called hermitian if x∗ = x. An element x of A is called normal if x∗ x = xx∗ . A hermitian element p is a projection whenever p is an idempotent, i.e. p2 = p. The symbol P(A) stands for the set of all projections of an involutive algebra A. Two projections p, q ∈ P(A) are called orthogonal if pq = 0. A projection p is central if px = xp for all x ∈ A. Denote the set of all central projections by Pc (A). A scalar λ ∈ C is a spectral value of x, if λ − x is not invertible in A. The set of all spectral values of x is called the spectrum of x and denoted by Sp(x). An element x of a ∗-algebra A is called positive if x is hermitian and Sp(x) ⊂ R+ . The set of all positive elements of A is denoted by A+ .

Operators in Spaces with Mixed Norm

317

If (A, ∗) and (B, ∗) are involutive algebras and R : A → B is a multiplicative linear operator, then R is called a ∗-representation of A in B whenever R(x∗ ) = R(x)∗ for all x ∈ A. If R is also an isomorphism then R is a ∗-isomorphism of A and B. In the presence of norms in the algebras, the naturally understood terms “isometric ∗-representation” and “isometric ∗-isomorphism” are in common parlance. 7.4.2. A norm k · k on an algebra A is submultiplicative if kxyk ≤ kxk kyk (x, y ∈ A). A Banach algebra A is an algebra furnished with a submultiplicative norm making A into a Banach space. An involutive Banach algebra is a Banach algebra which is also an involutive algebra and its involution satisfies the condition kx∗ k = kxk (x ∈ X). If A is an involutive Banach algebra satisfying kxx∗ k = kxk2

(x ∈ A)

then A is called a C ∗ -algebra. The spectrum of an element of a C ∗ -algebra is a nonempty compact subset of C. Let C(Sp(x), C) denote the C ∗ -algebra of complex continuous functions on Sp(x). (1) Spectral Theorem. Let x be a normal element of a C ∗ -algebra A, with Sp(x) the spectrum of x. There is a unique isometric ∗-representation Rx : C(Sp(x), C) → A such that x = Rx (ı), where ı is the identity mapping on Sp(x). The representation Rx : C(Sp(x), C) → A is called the continuous functional calculus (for a normal element x of A). The element Rx (f ) with f ∈ C(Sp(x), C) is usually denoted by f (x). In particular, for every positive x ∈ A the square root √ x is defined, since √ Sp(x) ⊂ R+ , and for each normal x ∈ A the modulus can be defined as |x| := x∗ x. (2) Let x be a normal element of a C ∗ -algebra A and f ∈ C(Sp(x), C). Then (g ◦ f )(x) = g(f (x)) for each g ∈ C(Sp f (x), C). (3) An element x of a C ∗ -algebra A is positive if and only if x = y ∗ y for some y ∈ A. The set A+ of all positive elements is an ordering cone and so (A, A+ ) is an ordered vector space. Treating a ∗-algebra A as an ordered vector space, we always imply the order that is conventionally induced by A+ .

318

Chapter 7

7.4.3. Suppose that ƒ is an order complete complex AM -space (see 1.3.11 and 1.5.5) with strong order-unity 1. According to the Brothers Kre˘ın–Kakutani Theorem 1.5.6 (2) and Theorem 1.5.9 ƒ is linearly isometric and order isomorphic to the space of continuous functions C(Q) on some extremal compact space Q. Therefore, A can be endowed with some multiplication and involution so that ƒ becomes a commutative C ∗ -algebra. Such C ∗ -algebra is often called a Stone algebra. Thus, a Stone algebra is a commutative C ∗ -algebra (with unity) which is order complete vector lattice with respect to the ordering 7.4.2 (3). An element e ∈ ƒ is a projection if and only if it is a fragment of 1. Moreover, the isomorphism ƒ → C(Q) defines a bijection between the set of fragments of 1 and the set of characteristic functions of clopen sets in Q, so that the Boolean algebras E(1) := E(A) coincides with the set of all projections P(ƒ) and is isomorphic to Clop(Q). A hermitian element p ∈ ƒ is a projection if and only if the multiplicative operator x 7→ px is a band projection. Given a complete Boolean B there exists a unique (up to ∗-isomorphism) Stone algebra ƒ such that B and P(ƒ) are isomorphic. Each of these algebras will be denoted by S (B). (The same symbol S (B) we used in 1.2.1 (2) to denote the Stone space of B. We hope that this liberty will not lead to confusions, since the meaning is always clear from the context.) Let ƒ be a Stone algebra and consider a unitary ƒ-module X. The mapping h· | ·i : X × X → ƒ is a ƒ-valued inner product, if for all x, y, z ∈ X and a ∈ ƒ the following are satisfied: (1) hx | xi ≥ 0; hx | xi = 0 ⇔ x = 0; (2) hx | yi = hy | xi∗ ; (3) hax | yi = ahx | yi; (4) hx + y | zi = hx | zi + hy | zi. Using a ƒ-valued inner product, we may introduce the norm in X by the formula p (5) |||x||| := khx|xik (x ∈ X), and the vector norm (6) x :=

p hx|xi (x ∈ X).

Employing the continuous functional calculus 7.4.2 (1, 2) we may deduce from the properties (2) and (3) that λx = |λ| x for all λ ∈ ƒ and x ∈ X. The fact that · satisfies the triangle inequality results as usual from the Cauchy–Bunyakovski˘ı– Schwarz inequality

Operators in Spaces with Mixed Norm

319

(7) hx | yi ≤ x y . On taking norms in (7) and using submultiplicativity and monotonicity of the norm in ƒ, we further obtain the numerical version of the Cauchy–Bunyakovski˘ı– Schwarz inequality (8) khx | yik ≤ |||x||| |||y|||. It follows from (5) and (6) that

(9) |||x||| = x (x ∈ X), √ √ since kak = k( a)2 k = k ak2 for every positive a ∈ ƒ. Therefore, the formula (5) defines a mixed norm on X (cf. 7.1.1). 7.4.4. Let X be a ƒ-module with an inner product h· | ·i : X × X → ƒ. If X is complete with respect to the mixed norm ||| · |||, it is called a C ∗ -module over ƒ. Theorem. Let X be a C ∗ -module. The pair (X, ||| · |||) is a B-cyclic Banach space if and only if (X, · ) is a Banach–Kantorovich space over ƒ := S (B). C Note that 7.4.4 (6) gives a decomposable norm since bx = b x (x ∈ X, b ∈ B). By Theorem 7.1.2, the normed space (X, ||| · |||) is complete if and only if (X, · ) is br-complete. Furthermore, it is clear that the B-cyclicity of (X, ||| · |||) amount to the disjoint completeness of (X, · ). The above remarks justify 2.2.3, so completing the proof. B 7.4.5. A Kaplansky–Hilbert module or an AW ∗ -module over ƒ is a unitary C ∗ module satisfying each of the equivalent conditions of Theorem 7.4.4. According to 2.2.3 X is a Kaplansky–Hilbert module over ƒ if and only if it is an C ∗ -module over ƒ and enjoys the following two properties: (1) let x be an arbitrary element in X, and let (eξ )ξ∈„ be a partition of unity in P(ƒ) with eξ x = 0 for all ξ ∈ „; then x = 0; (2) let (xξ )ξ∈„ be a norm-bounded family in X, and let (eξ )ξ∈„ be a partition of unity in P(ƒ); then there exists an element x ∈ X such that eξ x = eξ xξ for all ξ ∈ „. The element of (2) is the bo-sum of the family (eξ xξ )ξ∈„ , see 2.2.1. According to 7.4.3 (7) the inner product is bo-continuous in each variable. In particular, D P E

 P (3) bo- ξ∈„ eξ xξ | y = bo- ξ∈„ eξ xξ | y for every bounded family (xξ )ξ∈„ in X and partition of unity (eξ )ξ∈„ in P(ƒ). Let X be a Kaplansky–Hilbert module over ƒ. A Kaplansky–Hilbert submodule of X is a bo-closed submodule X0 ⊂ X. Theorems 7.1.2 and 2.2.3 imply that X0

320

Chapter 7

is a Kaplansky–Hilbert submodule if and only if X0 is a submodule in the conventional P algebraic sense closed in norm topology and containing all sums of the form bo- ξ∈„ eξ xξ , where (xξ )ξ∈„ is a bounded family in X0 and (eξ )ξ∈„ is a partition of unity in P(ƒ). The intersection of any number of Kaplansky–Hilbert submodules is again a Kaplansky–Hilbert submodule. Thus, for each nonempty subset M ⊂ X there exists a smallest Kaplansky–Hilbert submodule containing M ; it is called the Kaplansky–Hilbert submodule generated by M . (4) The orthogonal complement M ⊥ := {x ∈ X : (∀y ∈ M ) hx | yi = 0} of any nonempty subset M ⊂ X is a Kaplansky–Hilbert submodule of X. A Kaplansky–Hilbert module over ƒ is called faithful if for every a ∈ ƒ the condition (∀x ∈ X) ax = 0 implies that a = 0. It is clear that the faithfulness of ⊥⊥ = ƒ. In the sequel we restrict our attention to X amounts to the condition X faithful Kaplansky–Hilbert modules over ƒ. 7.4.6. Suppose that X is a unitary Kaplansky–Hilbert module over a Stone algebra ƒ. A subset E of X is said to be orthonormal if (1) hx | yi = 0 for all distinct x, y ∈ E ; (2) hx | xi = 1 for every x ∈ E . An orthonormal set E ⊂ X is a basis for X provided that (3) the condition (∀e ∈ E )hx | ei = 0 implies x = 0. Say that a Kaplansky–Hilbert module X is λ-homogeneous, if λ is a cardinal and X has a basis of cardinality λ. Granted 0 6= b ∈ B, denote by κ(b) the least cardinal γ such that a Kaplansky–Hilbert module bX over bƒ is γ-homogeneous. If X is homogeneous then κ(b) is defined for all 0 6= b ∈ B. It is convenient to assume that κ(0) = 0. We shall say that a Kaplansky–Hilbert module X is strictly γ-homogeneous if X is homogeneous and γ = κ(b) for all nonzero b ∈ B. A Kaplansky–Hilbert module is said to be (strictly homogeneous) if it is λhomogeneous (strictly λ-homogeneous) for some cardinal λ. If γ is a finite cardinal then the property of γ-homogeneity and strict γhomogeneity of a AW ∗ -module are equivalent. Denote by |M | the cardinality of M ; i.e., a cardinal number bijective with M . 7.4.7. (1) Suppose that X is a Kaplansky–Hilbert module over ƒ. The mapping κ preserves suprema of nonempty sets, i.e. κ(sup(D)) = sup(κ(D)) for every D ⊂ B. C Put ¯b := sup D. By definition κ is increasing: b1 ≤ b2 ⇒ κ(b1 ) ≤ κ(b2 ). Therefore, supb∈D κ(b) ≤ κ(¯b). Prove the reverse inequality. For an arbitrary nonzero b ∈ B the set of cardinals {κ(b0 ) : 0 6= b0 ≤ b} has the least element, say γ := κ(b0 ). Obviously, b0 6= 0 and κ(b0 ) = κ(b0 ) for all nonzero b0 ≤ b0 .

Operators in Spaces with Mixed Norm

321

Thus the set D0 of all b ∈ B, with bX strictly homogeneous, minorizes D. By the Exhaustion Principle (see 1.1.6) there exists a partition (bξ )ξ∈„ of the element ¯b such that bξ X is a strictly κ(bξ )-homogeneous Kaplansky–Hilbert module over bξ ƒ. Let Eξ := (eγ,ξ )γ 0, i.e., [[ y 6= 0 ]] = [[ x < λ∧ ]]. B 8.1.6. Theorem. Let E be an Archimedean vector lattice, let R be the reals in the model V(B) , and let  be an isomorphism of B onto the base B(E). Then there exists an element E ∈ V(B) satisfying the following conditions: (1) V(B) |= “E is a vector sublattice of R considered as a vector lattice over R∧ ”;

Applications of Boolean-Valued Analysis

347

(2) E 0 := E ↓ is a vector sublattice of R↓ invariant under each band projection χ(b) (b ∈ B) and such that every set of positive pairwise disjoint elements in it has a supremum; (3) there is an o-continuous lattice isomorphism ι : E → E 0 such that ι(E) is a massive sublattice in R↓; (4) for every b ∈ B the band projection in R↓ onto {ι((b))}⊥⊥ coincides with χ(b). C Assign d(x, y) := −1 ({|x − y|}⊥⊥ ). Let E be the Boolean-valued realization of the B-set (E, d) and E 0 := E ↓ (see A.12 (1, 2)). By A.12 (2), without loss of generality we may assume that E ⊂ E 0 , d(x, y) = [[ x 6= y ]] (x, y ∈ E), and E 0 = mix E. Further, furnish E 0 with a vector lattice structure. To this end, take a number λ ∈ R and elements x, y ∈ E 0 of the form x := mix(bξ xξ ) and y := mix(bξ yξ ), where (xξ ) ⊂ E, (yξ ) ⊂ E, and (bξ ) is a partition of unity in B, and define x + y := mix(bξ (xξ + yξ )), λx := mix(bξ (λxξ )), x ≤ y ⇔ x = mix(bξ (xξ ∧ yξ )). Inside V(B) , define the addition ⊕, multiplication , and order v in the set E as the ascents of the corresponding objects in E 0 . More precisely, the operations ⊕ : E × E → E and : E × R∧ → E and the predicate v ⊂ E × E are defined by the relations [[ x ⊕ y = x + y ]] = 1, [[ λ x = λx ]] = 1 (x, y ∈ E 0 , λ ∈ R), W [[ x v y ]] = {[[ x = x0 ]] ∧ [[ y = y 0 ]] : x0 , y 0 ∈ E 0 , x0 ≤ y 0 }. Thus, we may claim that E is a vector lattice over the field R∧ and, in particular, a lattice-ordered group inside V(B) . Also, it is clear that the Archimedean axiom is valid on E , since E 0 is an Archimedean lattice. Note that if x ∈ E+ then {x}⊥⊥ = d(x, 0) = [[ x 6= 0 ]], i.e., {x}⊥ = [[ x = 0 ]]. Consequently, we have [[ x = 0 ]] ∨ [[ y = 0 ]] = {x}⊥ ∨ {y}⊥ = 1B for disjoint x, y ∈ E. Hence, we easily derive that [[ E is linearly ordered ]] = 1, for [[ (∀x ∈ E )(∀y ∈ E ) (|x| ∧ |y| = 0 → x = 0 ∨ y = 0) ]] = 1.

348

Chapter 8

It is well known that an Archimedean linearly ordered group is isomorphic to an additive subgroup of the reals. Applying this assertion to E inside V(B) , without loss of generality we may assume that E is an additive subgroup of the field R. Furthermore, we suppose that 1 ∈ E , since otherwise E could be replaced by the isomorphic group e−1 E with 0 < e ∈ E . The multiplication represents a continuous R∧ -bilinear mapping from R∧ × E to E . Let β : R × R → R be its extension by continuity. Then β is R-bilinear and β(1, 1) = 1∧ 1 = 1. Consequently, β coincides with the usual multiplication in R; i.e., E is a vector sublattice of the field R considered as a vector lattice over R∧ . Thereby E 0 ⊂ R↓. The fact that E 0 is massive in R↓ obviously ensues from the fact that [[ E is dense in R ]] = 1. Prove that E is minorizing in E 0 . It follows from the properties of the isomorphism χ (see 8.1.2) that χ(b)ιx = 0 ⇔ (b) ≤ {x}⊥ ⇔ x ∈ (b⊥ ), whatever b ∈ B and x ∈ E+ might be. Hence, χ(b) is the band projection onto the band in R↓ generated by the set ι((b)). Moreover, if χ(b)x = 0 for all x ∈ E+ then b = {0}. Thus, for every b ∈ B we may find a positive element P y ∈ E for which y = χ(b)y. Now, take 0 < z ∈ E 0 . The representation z = o- ξ∈„ χ(bξ )xξ is valid, where (bξ ) is a partition of unity in B and (xξ ) ⊂ E+ . We see that χ(bξ )xξ 6= 0 at least for one index ξ. Let π := χ(bξ ) ◦ χ([[ xξ 6= 0 ]]) and y be a strictly positive element in E such that y = πy. Then for x0 := y ∧ xξ we have 0 < x0 ≤ πxξ ≤ χ(bξ )xξ ≤ z and x0 ∈ E. Thereby E is minorizing in E 0 . B 8.1.7. The element E ∈ V(B) arising in Theorem 8.1.6 is called the Booleanvalued realization of the vector lattice E. Thus, the Boolean-valued realizations of Archimedean vector lattices are vector sublattices of the reals R considered as a vector lattice over the field R∧ . Now, we indicated some corollaries to 8.1.2 and 8.1.6, with the same notations B, E, E 0 , E , ι, and R. (1) For every x0 ∈ E 0 there exist a family (xξ ) ⊂ E and a partition of unity (πξ ) in P(R↓) such that X x0 = oπξ ιxξ . ξ∈„

(2) For arbitrary x ∈ R↓ and ε > 0 there is xε ∈ E 0 such that |x − xε | ≤ ε1.

Applications of Boolean-Valued Analysis

349

C This is a consequence of the fact that [[ E is dense in R ]] = 1. B (3) If h : E → R↓ is a lattice isomorphism and for every b ∈ B the band projection onto the band in R↓ generated by the set h((b)) coincides with χ(b) then there exists a ∈ R↓ such that hx = a · ι(x) (x ∈ E). C Indeed, if E0 := im ι and h0 := h ◦ ι−1 then the isomorphism h0 : E0 → R↓ is extensional; therefore, for τ := h0 ↑ we have [[ the mapping τ : E → R is isotonic, injective, and additive ]] = 1. Consequently, h0 is continuous and has the form τ (α) = a · α (α ∈ R), where a is a fixed element in R↓. Hence, we derive that h0 (y) = a · y (y ∈ E0 ) or h(x) = a · ι(x) (x ∈ E). B (4) If there exists an order-unity 1 in E then the isomorphism ι is uniquely determined by the extra requirement that ι1 = 1. (5) If E is a K-space then E = R, E 0 = R↓, and ι(E) is an order-dense ideal of the K-space R↓. Moreover, ι−1 ◦ χ(b) ◦ ι is the band projection onto (b) for every b ∈ B. C If E is order complete then so is the lattice E 0 . From 8.1.4 (2) we see that the order completeness of E 0 is equivalent to the axiom of existence of exact bounds for bounded sets in E . By 8.1.1, E = R and E 0 = R↓. Let e ∈ E+ , y ∈ R↓, and |y| ≤ ιe. Since ι(E) is a massive sublattice in R↓, we have y + = sup ι(A), where A := {x ∈ E+ : ιx ≤ y + }. But the set A is bounded in E by the element e; therefore, sup A ∈ E and y + = ι(sup A) ∈ ιE. Similarly, y − ∈ ι(E) and, finally, y ∈ ι(E). B (6) The image ι(E) coincides with the whole R↓ if and only if E is a universally complete K-space. C If E is a K-space then E = R by (5) and, hence, R↓ = E ↓ = mix ι(E). However, for the universally complete K-space E we have mix ι(E) = ι(E). The converse is obvious. B (7) Universally complete K-spaces are isomorphic if and only if their bases are isomorphic. C If E and F are universally complete K-spaces and the Boolean algebras B(E) and B(F ) are isomorphic then E and F are isomorphic to the same K-space R↓ by (6). On the other hand, if h is an isomorphism from E onto F then the mapping K 7→ h(K) (K ∈ B(E)) is an isomorphism between the bases. B (8) Let E be a universally complete K-space with unity 1. Then we may uniquely define the multiplication in E so as to make E into an f -algebra and 1, into a ring unity. C By (6) and (4), we may assume that E = R↓ and 1 = 1. The existence of the required multiplication in E follows from 8.1.3. Assume that there is another multiplication : E × E → E in E and (E, +, , ≤) is a faithful f -algebra with

350

Chapter 8

unity 1. The faithfulness of the f -algebra implies that is an extensional mapping. But then the ascent × := ↑ is a multiplication in R. By virtue of uniqueness of the multiplicative structure in R, we conclude that × = · . Hence, we derive that coincides with the original multiplication in E (see 8.1.3). B 8.1.8. Interpreting the concept of a convergent numeric net inside V(B) and employing 8.1.4 (3) and 8.1.7 (5), we obtain useful tests for o-convergence in a Kspace E with unity 1. Theorem. Let (xα )α∈A be an order-bounded net in E and x ∈ E. The following are equivalent: (1) the net (xα ) o-converges to the element x; y(α)

(2) for every number ε > 0 the net (eε )α∈A of unit elements, where y(α) := |x − xα |, o-converges to 1; (3) for every number ε > 0 there exists a partition of unity (πα )α∈A in the Boolean-valued algebra P(E) such that πα |x − xβ | ≤ ε1 (α, β ∈ A, β ≥ α); (4) for every number ε > 0 there exists an increasing net (ρα )α∈A ⊂ P(E) of projections such that ρα |x − xβ | ≤ ε1 (α, β ∈ A, β ≥ α). C Without loss of generality we may assume that E is an order-dense ideal of the universally complete K-space R↓ (see 8.1.7 (5)). (1) ⇔ (2): It suffices to consider the case yα = xα (α ∈ A), i.e., (xα ) ⊂ E+ (o)

and xα → 0. Let σ be the modified ascent of the mapping s : α → xα . Then [[σ is a net in R+ ]] = 1. By 8.1.4 (3), o-lim s = 0 if and only if [[lim σ = 0]] = 1. We may rewrite the last equality in equivalent form: 1 = [[(∀ε ∈ R∧ )(ε > 0 → (∃α ∈ A∧ )(∀β ∈ A∧ ) (β ≥ α → xβ < ε))]]. Calculating the Boolean truth-values for the quantifiers, we find another equivalent form _  (∀ε > 0)(∃(bα )α∈A ⊂ B) bα = 1 ∧ (∀β ∈ A) (β ≥ α → [[xβ < ε∧ ]] ≥ bα ) α∈A

which in turn amounts to the following: _ ^  (∀ε > 0) [[xβ < ε∧ ]] = 1 . α∈A β∈A β≥α

Applications of Boolean-Valued Analysis

351 (o)

x

Since χ([[xβ < ε∧ ]]) = eε β (see 8.1.5), we see from the above that xα → x if and only if _ ^ x lim inf exε α = eε β = 1 α∈A

α∈A β∈A β≥α

(o)

for every ε > 0, i.e., exε α → 1 for every ε > 0. (1) ⇔ (3): Arguing as in (1) ⇒ (2), we find that the relation o-lim xα = x is equivalent to the following: _  (∀ε > 0)(∃(cα )α∈A ⊂ B) cα = 1 ∧ (∀β ∈ A)(β ≥ α ⇒ cα ≤ [[|xα − x| ≤ ε∧ ]]) . α∈A

By virtue of the Exhaustion Principle for Boolean algebras, there exist a partition of unity W (dξ )ξ∈„ in B and a mapping δ : „ → A such that dξ ≤ cδ(ξ) (ξ ∈ „). Put bα := {dξ : α = δ(ξ)} if α ∈ δ(„) and bα = 0 if α ∈ / δ(„). We see that (bα )α∈A is a partition of unity and bα ≤ cα (α ∈ A). Thus, if xα → x then for every ε > 0 there is a partition of unity (bα ) such that bα ≤ [[|x − xβ | ≤ ε∧ ]]

(α, β ∈ A, β ≥ α).

As follows from 8.1.2, the latter means that πα |x − xβ | ≤ ε1 (α, β ∈ A, β ≥ α), where πα := χ(bα ). Since (πα ) is a partition of unity in P(E), necessity is proven. To prove sufficiency, observe that if the indicated conditions are satisfied and a := lim sup |xα − x| then πα a ≤

_

|xβ − x| ≤ επα 1

β≥α

for all α ∈ A. Consequently, 0≤a=

X

πα a ≤ ε

X

πα 1 = ε1.

Since ε > 0 is arbitrary, we have a = 0 and o-lim xα = x. W (3) ⇔ (4): We only have to put ρα := {πβ : β ∈ A, α ≤ β} in (3). B

352

Chapter 8

8.1.9. Let C be the field of complex numbers in the model V(B) . Then the algebraic system C ↓ represents the complexification of the K-space R↓. In particular, C ↓ is a universally complete complex K-space and a complex algebra. C Since C = R ⊕ iR is equivalent to a bounded formula, we have [[ C∧ = R∧ ⊕ R∧ ]] = 1 (see A.8 (4)), where i is the imaginary unity and the element i∧ is denoted by the same letter i. From 8.1.1 we see that [[ C∧ is a dense subfield of the field C ]] = 1 and, in particular, [[ i is the imaginary unity of the field C ]] = 1. If z ∈ C ↓ then z is a complex number inside V(B) ; therefore, [[ (∃!x ∈ R)(∃!y ∈ R) z = x + iy ]] = 1. The Maximum Principle implies that there is a unique pair of elements x, y ∈ V(B) such that [[ x, y ∈ R ]] = [[ z = x + iy ]] = 1. Hence, we obtain x, y ∈ R↓, z = x + iy, and thereby C ↓ = R↓ ⊕ iR↓. Appealing to 8.1.2 and 8.1.4, complete the proof. B 8.2. Boolean-Valued Analysis of Vector Lattices In this section, we show that the most important structure properties of vector lattices such as representability by means of function spaces, the spectral theorem, the functional calculus, etc. are the images of properties of the reals in an appropriate Boolean-valued model. 8.2.1. We start with several useful remarks to be used below without further specification. Take a Kσ -space E. By Theorem 8.1.6, we may assume that E is a sublattice of the universally complete K-space R↓, where, as usual, R is the field b := I(E) of the reals in the model V(B) and B := B(E). Moreover, the ideal E generated by the set E in R↓ is an order-dense ideal of R↓ and an o-completion of E. The unity 1 of the lattice E is also a unity in R↓. The exact bounds of countable sets in E are inherited from R↓. In more detail, if the least upper (greatest lower) bound x of a sequence (xn ) ⊂ E exists in R↓ then x is also the least upper (greatest lower) bound in E, provided that x ∈ E. Thus, it does not matter whether the o-limit (o-sum) of a sequence in E is calculated in E or R↓, provided the result belongs to E. The same is true for the r-limit and r-sums. In particular, we may claim that if x ∈ E then the trace ex and the spectral function (characteristic) exλ of an element x calculated in R↓ are an element of B := E(E) and a mapping from R to B respectively. 8.2.2. As a first easy application we prove the properties of every spectral function of 1.3.8 (1–12). According to the remarks of 8.2.1, without loss of generality we may assume that the Kσ -space under consideration coincides with R↓. But

Applications of Boolean-Valued Analysis

353

then the required relations can be easily derived from the elementary properties of numbers with the help of 8.1.5. We confine exposition to (2), (6), and (8). C First of all observe that P∧ is a dense subfield of the field R inside V(B) Take x ∈ R↓ and consider the two formulas ϕ(x) := (∃t ∈ P∧ ) (x < t) and ψ(x) := (∀t ∈ P∧ ) (x < t). For a real number x the formula ϕ(x) is true and ψ(x) is false. Consequently, the Transfer Principle implies [[ϕ(x)]] = 1 and [[ψ(x)]] = 0. Calculating the Boolean truth-values for the quantifiers by the rules of A.8 (1) yields _ ^ [[x < t∧ ]] = 1, [[x < t∧ ]] = 0 t∈P

t∈P

which is equivalent to (2) by 8.1.5. Take positive elements x, y ∈ R↓ and a number 0 < t ∈ P. Then x, y, and t∧ are reals in the model V(B) . Make use of the following property of numbers: x ≥ 0 ∧ y ≥ 0 → (xy < t∧ ↔ (∃r, s ∈ P∧+ )(x < r ∧ y < s ∧ rs = t)). Employing again the Transfer Principle and the rules of A.8 (1) for calculating the Boolean truth-values, we arrive at the relation [[xy < t∧ ]] =

_

V [[x < r∧ ]] [[y < s∧ ]].

0≤r,s∈P rs=t

Hence, the required equality (6) ensues if we apply χ to both sides of the preceding equality (see 8.1.5). Now, let A be a set in the considered Kσ -space. Then A↑ is some set of reals inside V(B) and the formula inf(A) < t ⇔ (∃a ∈ A↑)(a < t) holds. Employing 8.1.4 (2) and A.10 (1), we may write down the following chain of equivalences: x = inf(A) ⇔ [[x = inf(A↑)]] = 1 ⇔ [[(∀t ∈ P∧ ) (x < t ↔ inf(A↑) < t)]] = 1 ⇔ (∀t ∈ P)[[x < t∧ ]] _ =[[(∃a ∈ A↑)(a < t∧ )]] ⇔ (∀t ∈ P)[[x < t∧ ]] = [[a < t∧ ]]. a∈A

Appealing to 8.1.5, complete the proof of (8). B 8.2.3. Thus, to each element of a Kσ -space with unity there corresponds the spectral function; moreover, the operations transform in a rather definite way. This circumstance suggests that an arbitrary Kσ -space with unity can be realized as some space of “abstract spectral functions.” This was done in 1.4.3 and 1.4.4. Now, we will pursue the Boolean-valued approach to the same problem.

354

Chapter 8

Theorem. Let B be a complete Boolean algebra. The set K(B) with introduced operations and order represents a universally complete K-space. The mapping sending an element x ∈ R↓ to the resolution of unity t 7→ [[x < t∧ ]] (t ∈ R) is an isomorphism between the K-spaces R↓ and K(B). C Denote the indicated mapping from R↓ to K(B) by the letter h. By Theorem 1.3.8, h preserves the operations and order. Moreover, h is one-to-one, since the equality h(x) = h(y) means [[x < t∧ ]] = [[y < t∧ ]]

(t ∈ R)

or (see A.8 (1)) [[(∀t ∈ R∧ ) (x < t ↔ y < t)]] = 1 and thereby is equivalent to coincidence of two numbers x and y inside V(B) . By virtue of Gordon’s Theorem 8.1.2, it remains to establish that h is surjective. Take an arbitrary resolution of unity e in the Boolean algebra B. Let β := (tn )n∈Z be a partition of the real axis; i.e., tn < tn+1 (n ∈ Z), limn→∞ tn = ∞, and limn→−∞ tn = −∞. The disjoint sum ¯(β) := x

X

tn+1 (χ(e(tn+1 )) − χ(e(tn )))

n∈Z

exists in the universally complete K-space R↓; here χ is the isomorphism of B ¯(β). onto E(R↓) (see 8.1.2 and 8.1.3). Denote by the letter A the set of all elements x Every element of the form x(β) :=

X

tn (χ(e(tn+1 )) − χ(e(tn )))

n∈Z

x(β)}. It is easy to is a lower bound of A. Therefore, there exists x := inf A := inf{¯ observe that W x ¯(β) eλ = {χ(e(tn )) : tn < λ}. Hence, by 1.3.8 (9), we infer exλ =

_ a∈A

_

eaλ =

t∈R,t λ; therefore, bξ µ(Ak ) = 0. Thereby ! ∞ ∞ _ ^ [ b= bξ ≤ µ(Ak )∗ = µ Š − Ak = µ({f < λ}). ξ∈„

k=1

k=1

On the other hand, b∗ = [[Iµ (f ) ≥ λ∧ ]] and, by 8.1.2, we again infer that λb ≤ b∗ Iµ (f ) ≤ b∗ σ(f, β) or ∗

λb∗ µ(Ak ) ≤ b∗ λk µ(Ak )

(k ∈ Z).

For k < 0 we have λk < λ; therefore, b∗ µ(Ak ) = 0. Consequently, ! −∞ −∞ ^ [ ∗ ∗ b ≤ µ(Ak ) = µ Š − Ak = µ({f ≥ λ}). k=−1

k=−1

This implies b ≥ µ({f < λ}) and we finally obtain b = µ({f < λ}). Assume that [[x < λ∧ ]] = µ({f < λ}) (λ ∈ R) for some x ∈ R↓. Then by what was established above we have [[x < λ∧ ]] = [[Iµ (f ) < λ∧ ]] for all λ ∈ R. This is equivalent to the relation [[(∀λ ∈ R∧ ) (x < λ ↔ Iµ (f ) < λ]] = 1. Hence, recalling that R∧ is dense in R, we obtain the equality [[x = Iµ (f )]] = 1 or x = Iµ (f ). B

Applications of Boolean-Valued Analysis

359

8.2.9. Take a measurable function f : Š → R and a spectral measure µ : † → B := E(E), where E is some K-space. If the integral Iµ (f ) ∈ E exists then λ 7→ µ({f < λ}) (λ ∈ R) coincides with the spectral function of the element Iµ (f ). C We have only to compare 8.2.8 with 8.1.5. B 8.2.10. Theorem. Let E be a universally complete Kσ -space, and let µ : † → B0 := E(E) be some spectral measure. The spectral integral Iµ (·) represents a sequential o-continuous (linear, multiplicative, and latticial) homomorphism from the f -algebra M (Š, †) of measurable functions into E. C Without loss of generality we may assume that E ⊂ R↓ and R↓ is a universal completion of E (see 8.1.7). Here R is the field of the reals in V(B) , where B is a completion of the algebra B0 . It is obvious that the operator Iµ is linear and positive. Prove its sequential o-continuity. Take a decreasing sequence (fn )n∈N of measurable functions such that limn→∞ fn (t) = 0 for all t ∈ Š, and let xn S := Iµ (fn ) ∞ (n ∈ N) and 0 < ε ∈ R. If we assign An := {t ∈ Š : fn (t) < ε} then Š = n=1 An . By Proposition 8.2.9, we may write down o-lim

n→∞

exε n

= o-lim µ(An ) = n→∞

∞ _

µ(An ) = 1.

n=1

Appealing to the test for o-convergence 8.1.9 (2), we obtain o-limn→∞ xn = 0. Further, given arbitrary measurable functions f, g : Š → R, we derive from 1.3.8 (8) and 8.2.9 that I(f ∨g)

eλ

I(f )

= µ({f ∨ g < λ}) = µ({f < λ}) ∧ µ({g < λ}) = eλ

I(g)

∧ eλ

I(f )∨I(g)

= eλ

(with I := Iµ ); consequently, I(f ∨ g) = I(f ) ∨ I(g). It means that Iµ is a lattice homomorphism. In a similar way, for f ≥ 0 and g ≥ 0 it follows from 1.3.8 (6) and 8.2.9 that  [  I(f ·g) eλ = µ({f · g < λ}) = µ {f < r} ∩ {g < s} r,s∈Q+ rs=λ

=

_

µ({f < r}) ∧ µ({g < s}) =

r,s∈Q+ rs=λ

_

I(f )·I(g)

) eI(f ∧ eI(g) = eλ r s

r,s∈Q+ rs=λ

for λ > 0, λ ∈ Q, with Q the rationals. Thus, I(f · g) = I(f ) · I(g). The validity of the latter equality for arbitrary functions f and g ensues from the above-established properties of the spectral integral: I(f · g) = I(f + g + ) + I(f − g − ) − I(f + g − ) − I(f − g + ) = I(f )+ I(g)+ + I(f )− I(g)− − I(f )+ I(g)− − I(f )− I(g)+ = I(f ) · I(g). B

360

Chapter 8

8.2.11. Let e1 , . . . , en : R → B be a finite collection of spectral functions with values in a σ-algebra B. Then there exists a unique B-valued spectral measure µ defined on the Borel σ-algebra Bor(Rn ) of the space Rn such that ! n n Y ^ µ (−∞, λk ) = ek (λk ) k=1

k=1

for all λ1 , . . . , λn ∈ R. C Without loss of generality we may assume that B = Clop(Q), where Q is the Stone space of B. According to 1.4.1 (1), there are continuous functions xk : Q → R such that ek (λ) = cl{xk < λ} for all λ ∈ R and k := 1, . . . , n. Put f (t) = (x1 (t), . . . , xn (t)) if all xk (t) are finite and f (t) = ∞ if xk (t) = ±∞ at least for one index k. Thereby we have defined a continuous mapping f : Q → Rn ∪ {∞} (the neighborhood filterbase of the point ∞ is composed of the complements to various balls with center the origin). It is clear that f is measurable with respect to the Borel algebras Bor(Q) and Bor(Rn ). Let Clopσ (Q) and ϕ be the same as in 1.2.6. Define the mapping µ : Bor(Rn ) → B by the formula
µ(A) := ϕ f −1 (A) (A ∈ Bor(Rn )). Qn It is obvious that µ is a spectral measure. If A := k=1 (−∞, λk ) then f

−1

(A) =

n \

{xk < λk },

k=1

and hence µ(A) = e1 (λ1 ) ∧ · · · ∧ en (λn ). If ν is another spectral measure with the same properties as µ then the set B := {A ∈ Bor(Rn ) : ν(A) = µ(A)} is a σ-algebra containing all sets of the form n Y

(−∞, λk )

(λ1 , . . . , λn ∈ R).

k=1

Hence, B = Bor(Rn ). B 8.2.12. Now, take an ordered collection of elements x1 , . . . , xn in a Kσ -space E with unity 1. Let exk : R → B := E(1) denote the spectral function of the element xk . According to the above-proven assertion, there exists a spectral measure µ : Bor(Rn ) → B such that ! n n Y ^ µ (−∞, λk ) = exk (λk ). k=1

k=1

Applications of Boolean-Valued Analysis

361

We may see that the measure µ is uniquely determined by the ordered collection x := (x1 , . . . , xn ) ∈ E n . For this reason, we write µx := µ and say that µx is the spectral measure of the collection x. The following notations are accepted for the integral of a measurable function f : Rn → R with respect to the spectral measure µx : ^x(f ) := f (x) := f (x1 , . . . , xn ) := Iµ (f ). If x = (x) then we also write x ^(f ) := f (x) := Iµ (f ) and call µx := µ the spectral measure of x. Recall that the space B(Rn , R) of all Borel functions in Rn is a universally complete Kσ -space and a faithful f -algebra. 8.2.13. Theorem. The spectral measures of a collection x := (x1 , . . . , xn ) and the element f (x) maintain the relation µf (x) = µx ◦ f ← , where f ← : Bor(R) → Bor(Rn ) is the homomorphism acting by the rule A 7→ f −1 (A). In particular, (f ◦ g)(x) = g(f (x)) for measurable functions f ∈ B(Rn , R) and g ∈ B(R, R) whenever f (x) and g(f (x)) exist. C By 8.2.9, we have f (x)

µf (x) (−∞, t) = et

= [[f (x) < t]] = µx ◦ f −1 (−∞, t)

for every t ∈ R. Hence, the spectral measures µf (x) and µx ◦ f ← defined on Bor(R) coincide on the intervals of the form (−∞, t). Reasoning in a standard manner, we then conclude that the measures coincide everywhere. To prove the second part, it suffices to observe that (g ◦ f )← = f ← ◦ g ← and apply what was established above twice. B 8.2.14. Theorem. For every ordered collection x := (x1 , . . . , xn ) of a universally complete Kσ -space E, the mapping ^x : f 7→ ^x(f )

(f ∈ B(Rn , R))

is a unique sequentially o-continuous homomorphism of the f -algebra B(Rn , R) into E satisfying the conditions ^x(dtk ) = xk

(k := 1, . . . , n),

where dtk : (t1 , . . . , tn ) 7→ tk stands for the kth coordinate function on Rn .

362

Chapter 8

C As was established in 8.2.10, the mapping f 7→ ^x(f ) is a sequentially o-continuous homomorphism of f -algebras. Theorem 8.2.13 yields the equalities µdtk (x) = µx ◦ (dtk )← = µxk . Consequently, the elements ^x(dtk ) = dtk (x) and xk coincide, for they have the same spectral function. If h : B(Rn , R) → E is another homomorphism of f -algebras with the same properties as ^x(·) then h and ^x(·) coincide on all polynomials. Afterwards, we infer that h and ^x(·) coincide on the whole B(Rn , R) due to o-continuity. B 8.2.15. Theorem. An element x ∈ E has the form x = f (x) with some x ∈ E n and f ∈ B(Rn , R) if and only if im(µx ) ⊂ im(µx ). C Necessity follows from 8.2.13. Sufficiency is left to the reader as an exercise. B 8.3. Boolean-Valued Banach Spaces In this section we discuss the Boolean-valued transfer principle for latticenormed spaces. The interpretation of a Banach space inside an arbitrary Booleanvalued model is a Banach–Kantorovich space. Conversely, the maximal extension of any lattice-normed space, having been embedded into a suitable Boolean-valued model, becomes a Banach space. A possibility thus appears of transferring theorems on Banach spaces to analogous results on lattice-normed spaces by means of Boolean-valued methods. As in Section 8.1, B is a fixed complete Boolean algebra and V(B) is the Boolean-valued universe constructed over B. Let R and C be the fields of real and complex numbers inside V(B) . Denote by ⊕ and the addition and multiplication in the fields R and C . 8.3.1. Theorem. Let (X , ρ) be a Banach space in the model V(B) . Assign X := X ↓ and · := ρ↓(·). The following hold:
(1) X, · , R↓ is a universally complete Banach–Kantorovich space; (2) the space X can be endowed with the structure of a faithful unital module over the ring ƒ = C ↓ so that (a) (λ1)x = λx (λ ∈ C , x ∈ X); (b) ax = |a| x (a ∈ C ↓, x ∈ X); (c) b ≤ [[ x = 0 ]] ⇔ χ(b)x = 0 (b ∈ B , x ∈ X), where χ is an isomorphism of B onto P(X). C We denote the addition in X , C , and R by the same symbol ⊕. Let denote the external composition law C × X → X of the complex vector space X , as well as the multiplication in R and in C . Assign + := ⊕↓ and · := ↓. This means that x + y = z ⇔ [[ x ⊕ y = z ]] = 1 (x, y, z ∈ X); a · x = y ⇔ [[ a x = y ]] = 1 (a ∈ ƒ, x, y ∈ X).

Applications of Boolean-Valued Analysis

363

From simple properties of the descent, it follows that (X, +) is an abelian group. For instance, commutativity of the + operation is deduced as follows. Inside the model, we have [[ ⊕ ◦ ı = ⊕ ]] = 1, where ı : X × X → X × X is a permutation of coordinates. Therefore,  = ı↓ is a permutation of coordinates in X × X, and + ◦  = ⊕↓ ◦ ı↓ = (⊕ ◦ ı)↓ = ⊕↓ = +. Given arbitrary b ∈ B and x ∈ X, assign χ(b)x := mix {bx, b∗ 0}, where 0 is the neutral element of the group (X, +). In other words, χ(b)x is a unique element of X for which [[ χ(b)x = x ]] ≥ b and [[ χ(b)x = 0 ]] ≥ b∗ . A mapping χ(b) : X → X is thus defined; moreover, χ(b) is additive and idempotent. Let P := {χ(b) : b ∈ B}. Then P is a complete Boolean algebra and χ is a Boolean isomorphism. Taking account of the fact that, inside the model V(B) , the axioms of a vector space are valid for X , we may write a · (x + y) = a (x + y) = a x + a y = a · x + a · y, (a + b) · x = (a + b) x = a x + b x = a · x + b · x, (ab) · x = (ab) x = a (b x) = a · (b · x), 1 · x = 1 x = x (a, b ∈ ƒ; x, y ∈ X). Since V(B) is separated, from these relations it follows that the + and · operations determine the structure of a unital ƒ-module in X. Letting λx := (λ1) · x (λ ∈ C, x ∈ X), we obtain the structure of a complex vector space in X; moreover, equality (a) is valid. Since, in the model V(B) , the following hold: χ(b) = 1 ⇒ χ(b) x = x, χ(b) = 0 ⇒ χ(b) x = 0, for b ≤ [[ x = 0 ]] we have by definition b ≤ [[ χ(b) x = x ]] ∧ [[ x = 0 ]] ≤ [[ χ(b) · x = 0 ]], b∗ ≤ [[ χ(b) x = 0 ]] = [[ χ(b) · x = 0 ]]. Hence, [[ χ(b) · x = 0 ]] = 1, i.e., χ(b)x = 0, which implies (c). Now turn to Banach properties of the space (X , ρ). Subadditivity and homogeneity of the norm ρ can be written as follows: ρ ◦ ⊕ ≤ ⊕ ◦ (ρ × ρ),


ρ ◦ = ◦ |·| × ρ ,

364

Chapter 8



where ρ × ρ : (x, y) 7→ ρ(x), ρ(y) and |·| × ρ : (a, x) 7→ |a|, ρ(x) . Taking account of the rules of descending composition, for p := · we have p ◦ + ≤ + ◦ (p × p),


p ◦ · = · ◦ |·| × p .

This means that the operator · : X → Re ƒ is a vector seminorm and condition (b) is satisfied. If x = 0 for some x ∈ X, from [[ ρ(x) = x ]] = 1 we have [[ ρ(x) = 0 ]] = 1; hence, [[ x = 0 ]] = 1, i.e., x = 0. Thus, · is a vector norm. Decomposability ensues from property (b). Indeed, assume that c := p(x) = c1 + c2 (x ∈ X; c1 , c2 ∈ ƒ+ ). There exist a1 , a2 ∈ ƒ+ such that ak c = ck (k := 1, 2) and a1 +a2 = 1.
−1
Assign ak := ck c+(1−ec ) , where ec is the trace of the element c. If xk := ak ·x (k := 1, 2) then x = x1 + x2 and xk = ak x = ak x = ck . It remains to prove bo-completeness for X. Take a bo-fundamental net s : A → X. If s¯(α, β) := s(α) − s(β) (α, β ∈ A), then lim · ◦ s¯(α, β) = 0. Let σ : A∧ → X be the modified ascent of s, and σ ¯ (α, β) := σ(α) − σ(β) (α, β ∈ A∧ ). ¯ is the modified ascent of s¯, and ρ ◦ σ ¯ is the modified ascent of · ◦ s. Then σ ¯ = 0 ]] = 1, i.e., V(B) |=“σ is a fundamental net in Then, due to 8.1.4 (3), [[ lim ρ ◦ σ X .” Since X is a Banach space inside V(B) , by the Maximum Principle there is an element x ∈ X such that [[ lim ρ ◦ σ0 = 0 ]] = 1, where σ0 : A∧ → X is defined by the formula σ0 (α) := σ(α) − x (α ∈ A∧ ). The modified descent of σ0 is presented by the net s0 : α 7→ s(α) − x (α ∈ A). Consequently, according to 8.1.4 (3), we have o-lim |·| ◦ s0 = 0, i.e., o-lim s(α) − x = 0. B A universally complete Banach–Kantorovich space X ↓ := (X , ρ)↓ := (X ↓, ρ↓) is called the descent of the Banach space (X , ρ). 8.3.2. Theorem. For every lattice-normed space (X, p), there exists a unique ⊥⊥
(to within a linear isometry) Banach space X inside V(B) , with B ' B X , for which the descent X ↓ is the universal completion of X. C Without loss of generality, we assume that E = p(X)⊥⊥ ⊂ mE = R↓ and B = B(E). Assign d(x, y) := p(x − y)⊥⊥ (x, y ∈ X). It is easy to verify that d is a B-metric in the set X. If we endow the field C with the discrete B-metric d0 , then the operations of addition, + : X × X → X, and multiplication, · : C × X → X, are nonexpanding mappings. The vector norm p is nonexpanding too. All these assertions are almost evident. For instance, as regards multiplication, whenever α, β ∈ C and x, y ∈ X, we have
⊥⊥
⊥⊥ d(αx, βy) = p(αx − βy)⊥⊥ ≤ |α|p(x − y) ∨ |α − β|p(y) ≤ d(x, y) ∨ d0 (α, β).

Applications of Boolean-Valued Analysis

365

Let X0 be the Boolean-valued representation of the B-set (X, d) (see A.12). Assign ρ0 := F ∼ (p), ⊕ := F ∼ (+), and := F ∼ (·), where F ∼ is the immersion defined in A.12 (3). The mappings ⊕ and define the structure of a vector space over the field C∧ in the set X0 , and the function ρ0 : X0 × X0 → R is a norm. In view of the Maximum Principle, there exist elements X , ρ ∈ V(B) , for which [[ (X , ρ) is a complex Banach space that is the completion of the normed space (X0 , ρ0 ) ]] = 1. Moreover, we may assume that [[ X0 is a dense C∧ -subspace of X ]] = 1. Let ı : X → X0 := X0 ↓ be the canonical injection (see A.12). Since + is a nonexpanding mapping from X × X into X, the addition in X0 , i.e. + := ⊕↓, is uniquely determined by the relation ı ◦ + = + ◦ (ı × ı), where ı × ı : (x, y) 7→ (ıx, ıy) is the canonical injection of the B-set X × X. However, it is equivalent to additivity of ı. Similarly, for the operation · := ↓, we have ı ◦ · = · ◦ (κ × ı), where κ × ı : (λ, x) 7→ (λ∧ , ıx) (λ ∈ C, x ∈ X). Thus, ı is a linear operator. By repeating the same reasoning for p0 := ρ0 ↓, we obtain ıE ◦p0 = p0 ◦ı, where ıE is the canonical injection of E. This means that ı is an isometry, i.e., ı preserves the vector norm. Consider a subspace Y ⊂ X ↓, ıX ⊂ Y , that is a universally complete Banach– Kantorovich space with the norm q(y) := ρ↓(y) (y ∈ Y ). From decomposability of the norm q and disjoint completeness of Y , it follows that X0 ⊂ Y . Indeed, X0 = mix (ıX), and, due to P (c) of 8.3.1 (2), given an x ∈ X ↓, we have x = mix (bξ ıxξ ) if and only if x = bo- χ(bξ )ıxξ . On the other hand, Y is decomposable and dcomplete; hence, according to 2.1.3 and 2.2.1, Y is invariant under every projection x 7→ χ(b)x and contains all sums of the indicated type. Arguing analogously, we have Y = mix Y . If Y := Y ↑ then [[ X0 ⊂ Y ⊂ X ]] = 1; moreover, Y ↓ = Y . Let σ : ω ∧ → Y be a Cauchy sequence and let s be its modified descent. Then s is a bo-fundamental sequence in Y ; consequently, y = lim s exists. From 8.1.4 (3) it is clear that [[ y = lim σ ]] = 1. Completeness of the set Y is thus established; therefore, X = Y and X = Y . Let Z be a Banach space inside V(B) ; moreover, let Z ↓ be the universal extension of the lattice-normed space X. If ı0 is the corresponding isometric embedding of X into Z ↓, then ı0 ◦ ı uniquely extends to a linear isometry of X0 onto a disjointly complete subspace Z0 ⊂ Z ↓. The spaces X0 and Z0 := Z0 ↑ are isometric. Therefore, their completions X and Y ⊂ Z are isometric too. Since Y ↓ is a Banach–Kantorovich space and ıX ⊂ Y ↓ ⊂ Z ↓, it follows that Y ↓ = Z ↓. Therefore, Y = Z and, thus, X and Z are linearly isometric. B 8.3.3. Theorem. Let X and Y be Boolean-valued realizations for Banach– Kantorovich spaces X and Y normed by some universally complete K-space E. Let L B (X , Y ) be the space of bounded linear operators from X into Y inside V(B) , where B := B(E). The immersion mapping T 7→ T ∼ of the operators implements a linear isometry between the lattice-normed spaces LE (X, Y ) and L B (X , Y )↓. C By Theorem 8.3.2, without loss of generality we may assume that E = R↓,

366

Chapter 8

X = X ↓, and Y ↓ = Y (see 4.1.3 (5) for definition of LE (X, Y )). Take a mapping T : X → Y inside V(B) and put T := T ↓. Let ρ and θ be the norms of the Banach spaces X and Y , let p := ρ↓ and q := θ↓, and let + stand for all addition in X , Y , X, and Y . The linearity and boundedness of T imply validity for the relations T ◦ + = + ◦ (T × T ), θ ◦ T ≤ kρ, where 0 ≤ k ∈ R↓. The descent and ascent rules for composition allow us to write down the relations in the following equivalent form: T ◦ + = + ◦ (T × T ),

q ◦ T ≤ kp.

But this means that T is linear and bounded. Let K be the set constituted of 0 ≤ k ∈ R↓ such that q(T x) ≤ kp(x) (x ∈ X). Then K↑ = {k ∈ R+ : θ ◦ T ≤ kρ} inside V(B) . Appealing to 8.1.4 (2), we derive V(B) |= T = inf K = inf(K↑) = kT k. Hence, the mapping T 7→ T ↓ preserves the vector norm. To justify the linearity of the mapping, it suffices to check its additivity. Given T1 , T2 ∈ L B (X , Y )↓, we have (T1 + T2 )↓(x) = (T1 + T2 )(x) = T1 x + T2 x = T1 ↓x + T2 ↓x = (T1 ↓ + T2 ↓)x inside V(B) for every x ∈ X. Consequently, (T1 + T2 )↓ = T1 ↓ + T2 ↓. So, the descent implement a linear isometry of L B (X , Y )↓ onto the space of all extensional bounded linear operators from X into Y . It remains to observe that every bounded linear operator from X into Y is nonexpanding, or which is the same, satisfies the inequality [[x = 0]] ≤ [[T x = 0]]. Indeed, if b := [[x = 0]] then χ(b)x = 0 by 8.3.1 (2); therefore, χ(b)q(T x) ≤ χ(b)kp(x) = kp(χ(b)x) = 0. Hence, q(χ(b)T x) = 0 or χ(b)T x = 0 and, employing 8.3.1 (2) again, we conclude that b ≤ [[T x = 0]]. B e is the completion of X. Let 8.3.4. Assume that X is a normed space and X ∧ ∧ X be the completion of the R -normed space X inside V(B) . (1) Theorem. The universally complete Banach–Kantorovich space e with Q the Stone space of B. X ↓ is linearly isometric to the space C∞ (Q, X),

Applications of Boolean-Valued Analysis

367

C Identify the K-spaces R ↓ and C∞ (Q), and apply Theorem 8.3.2 to the lattice-normed space (X, p, R ↓), with p(x) = kxk · 1. Using the notation of the proof of 8.3.2, note that X0 = X ∧ . Hence, X ↓ := (X ↓, q, R↓) is the universal completion of (X, p, R↓). For simplicity, assume that X ⊂ X ↓. Since e and ε > 0 [[ X ∧ is norm-dense in X ]] = 1, we deduce that to u ∈ C∞ (Q, X) there are a family (xξ ) ⊂ X and a partition of unity (Qξ ) ⊂ Clop(Q) for which the step-function uε , equal to xξ on the set Qξ , obeys the estimate u − uε ≤ ε1. Put T (uε ) = mix(bξ xξ ) where bξ stands for the member of B corresponding to the clopen set Qξ . Now, T (uε ) = uε . Hence, T is a linear isometric embedding of (r)

the subspace of all
vector-functions of the shape uε . If ε → 0 then uε − u −→ 0, and so T (u1/n ) is an r-fundamental sequence. Since X ↓ is complete, X ↓ contains the limit v := r-lim T (u1/n ). Assigning T (u) := v, obtain a linear isometric e → X ↓. If Z := im(T ) then Z is a decomposable boembedding T : C∞ (Q, X) complete subspace of X ↓ and X ⊂ Z. By Theorem 8.3.2 and the definition of 2.2.6, Z = X ↓. B (2) Assume that X 0 is the dual of X inside V(B) . Then the spaces 0 X ↓ and Es (X 0 ), with E = C∞ (Q), are linearly isometric. C Apply Theorem 8.3.3 to Y := E and X := (X, p, E), with p(x) = kxk1. Deduce so that the spaces X 0↓ := L (B) (X , R)↓ and LA (X, E) are linearly isometric. To complete the proof, refer to 5.5.1 (1). B 8.3.5. An isometry ı between normed B-spaces is B-isometry if ı is linear and commutes with every projection in B. Say that Y is a B-cyclic completion of a B-space X if Y is B-cyclic and there is a B-isometry ı : X → Y such that every B-cyclic subspace of Y containing ı(X) coincides with Y . (1) Each Banach B-space possesses a B-cyclic completion unique up to B-isometry. C The claim follows from 2.2.8, 7.1.3, and 7.3.3 (1). B Let ƒ be the bounded part of the universally complete K-space C ↓, i.e. ƒ is the order-dense ideal in C ↓ generated by the order-unity 1 := 1∧ ∈ C ↓. Take a Banach space X inside V(B) . Denote X ↓∞:= {x ∈ X ↓: x ∈ ƒ}. Then X ↓∞ is a Banach–Kantorovich space called the bounded descent of X . Since ƒ is an order complete AM -space with unity, X ↓∞ is a Banach space with mixed norm over ƒ, and hence, B-cyclic Banach space, see 7.3.4. If Y is another Banach space and T : X → Y is a bounded linear operator inside V(B) with T ↓ ∈ ƒ then the bounded descent of T is the restriction of T ↓ to X ↓∞ . Clearly, the bounded descent of T is a bounded linear operator from X ↓∞ to Y ↓∞ .

368

Chapter 8

(2) Theorem. A Banach space X is linearly isometric to the bounded descent of some Banach space inside V(B) if and only if X is B-cyclic. C Cf. 8.3.1, 8.3.2, 7.1.3, and 7.3.3 (1). B e the norm completion (3) Take a normed B-space X. Denote by X e is a Banach B-space for every projection b ∈ B admits a unique of X. Then X e which preserve the norm of b. By (1), X e possesses a extension to the whole of X cyclic B-completion which we denoted by X. Applying Theorem (2), we now take a Banach space X inside V(B) whose bounded descent is B-isometric with X. The element X ∈ V(B) is called the Boolean-valued representation of X. (4) Let X and Y be normed space such that B ⊂ L (X) and B ⊂ L (Y ). An operator T : X → Y is B-linear, if T commutes with every projection in B; i.e. b ◦ T = T ◦ b for all b ∈ B. Denote by LB (X, Y ) the set of all bounded B-linear operators from X to Y . In this event W := LB (X, Y ) is a Banach space and B ⊂ W . If Y is B-cyclic then so is W . A projection b ∈ B acts in W by the rule T 7→ b ◦ T (T ∈ W ). We call X # := LB (X, ƒ) the B-dual of X. If X # and Y are B-isometric to each other then we say that Y is a B-dual space and X is a B-predual of Y . In symbols, X = Y# . 8.3.6. Theorem. Assume that X is a normed B-space and Y is a B-cyclic Banach space. Let X and Y stand for the Boolean-valued representation of X and Y . The space LB (X, Y ) is B-isometric to the bounded descent of the space L B (X , Y ) of all bounded linear operators from X to Y inside V(B) . Moreover, to T ∈ LB (X, Y ) there corresponds the member T := T↑ of V(B) determined from the formulas [[ T : X → Y ]] = 1, [[ T ıx = ıT x ]] = 1 (x ∈ X), where ı stands for a mapping that embeds X into X ↓ and Y into Y ↓. C Without loss of generality, assume that X and Y the bounded descents of some Banach spaces X and Y (cf. 8.3.5 (1) and 8.3.5 (2)). Put X0 := X ↓ and Y0 := Y ↓. By 8.3.3, the spaces L B (X , Y ) ↓ and LR (X0 , Y0 ) are linearly isometric. Moreover, the restriction of LR (X0 , Y0 ) relative to S (B) coincides with the bounded descent of L (X , Y ). It suffices to note that each member T of LR (X, Y ) admits a unique extension which preserves the norm of T . B 8.3.7. Let X ∗ be the dual of X inside V(B) . Denote by ' and 'B the relations of isometric isomorphy and isometric B-isomorphy between Banach spaces. Suppose also that X, Y , X , and Y are the same as in 8.3.6. (1) The following equivalence holds: X # 'B Y ⇔ [[ X ∗ ' Y ]] = 1. #

(2) If X is the B-cyclic completion of X then X # = X . 8.3.8. Theorem. The bounded descent of an arbitrary Hilbert space in V(B) is a Kaplansky–Hilbert module over the Stone algebra S (B). Conversely, if X is

Applications of Boolean-Valued Analysis

369

a Kaplansky–Hilbert module over S (B), then there is a Hilbert space X in V(B) whose bounded descent is unitarily equivalent with X. This space is unique to within unitary equivalence inside V(B) . C Without loss of generality, we may assume that S (B) ⊂ C ↓. Suppose that X is a Hilbert space inside V(B) and X is the bounded descent of X . Then the pair (X, · ), with · the descent of the norm of X is a Banach–Kantorovich space and the pair (X, k| · k|), with k|xk| = k x k (x ∈ X), is a B-cyclic Banach space (cf. 8.3.5 (2)). In particular, X is a unitary module over S (B). Suppose that (· | ·) ∈ V(B) is the inner product in X and h· | ·i is the descent of (· | ·). It is easy to check that h· | ·i satisfies 7.4.3 (1–4) for all x, y, z ∈ X ↓ and a ∈ C ↓. If x, y ∈ X then [[ |(x | y)| ≤ kxk · kyk ]] = 1. Hence, |hx | yi| ≤ x · y . Since x , y ∈ S (B); therefore, hx | yi ∈ S (B). Thus, the restriction of h· | ·i to X × X, denoted by the same inner product on X. It suffices p symbol, is a S (B)-valued p to note that x = hx | xi, since [[ kxk = (x | x) ]] = 1 and the descent of the √ function : R + → R + depicts the square root in S (B). Now, consider a Kaplansky–Hilbert module X over S (B). By Theorem 8.3.2, the Boolean-valued representation X ∈ V(B) of the Banach–Kantorovich space (X, · , S (B)) is a Banach space inside V(B) . We may thus assume that X ⊂ X ↓. Let (· | ·) stand for the ascent of the S (B)-valued inner product h· | ·i in X. Then (· V(B) . Arguing as above, we see that [[kxk = p| ·) is an inner product on X insidep (x | x) (x ∈ X )]] = 1, since x = hx | xi (x ∈ X). Suppose that Y is another Hilbert space inside V(B) and the bounded descent Y of Y is unitarily equivalent with X. If U : X → Y is a unitary isomorphism then u := U ↑ is a linear bijection from X to Y . Since U enjoys the property h· | ·i ◦ (U × U ) = h· | ·i, we note inside V(B) that (· | ·) ◦ (u × u) = h· | ·i ↑ ◦ (U↑ ×U↑) = (h· | ·i ◦ (U × U ))↑ = h· | ·i↑ = (· | ·). Hence, u is a unitary equivalence between X and Y . This end the proof. B As usual, we call X the Boolean-valued representation of a Kaplansky–Hilbert module X. Suppose that L B (X , Y ) the space of bounded linear operators from X to Y inside V(B) . Let Hom(X, Y ) stand for the space of all bounded ƒ-linear operators from X to Y where X and Y are Kaplansky–Hilbert modules over the commutative AW ∗ -algebra S (B) = ƒ. As before, we let S (B) stand for the bounded descent of the field C. It is easy to see that Hom(X, Y ) = Lƒ (X, Y ) (cf. 4.1.3 (5), 8.3.3). 8.3.9. Theorem. Suppose that X and Y are Hilbert spaces inside V(B) . Let X and Y stand for the bounded descents of X and Y . For every bounded ƒ-linear operator ˆ : X → Y the element ϕ := ˆ↑ is a bounded linear operator from X to Y inside V(B) .

370

Chapter 8

Moreover, [[ kϕk ≤ c∧ ]] = 1 for some c ∈ R. The mapping ˆ 7→ ϕ is a B-linear isometry between the B-cyclic Banach spaces Hom(X, Y ) and L B (X , Y )↓∞ . C Appealing to 8.3.5 (2) and 8.3.6, we complete the proof. B 8.3.10. Theorem. For a Kaplansky–Hilbert module X to be λ-homogeneous it is necessary and sufficient that [[ dim(X ) = |λ∧ | ]] = 1. C By Theorem 8.3.2 we may assume, that X ⊂ X ↓. The mapping h· ·i and the descent of the form (· | ·) agree on X × X. Therefore, for all x, y ∈ X and a ∈ ƒ the following are equivalent hx|yi = a and [[ (x|y) = a ]] = 1. We thus see that the orthogonality relation on X is the restriction to X of the descent of the orthogonality relation on X . From these observations it follows that a subset E of X is orthonormal if and only if [[ E ↑ is an orthonormal set in X ]] = 1. Applying the descent rules for polars to orthogonal complements in X and X , we obtain (E ↑)⊥ ↓ = (E ↑↓)⊥ . Observe also that E ⊥ = (E ↑↓)⊥ . Hence, E ⊥ ↑ = (E ↑)⊥ . In particular, E ⊥ = 0 if and only if [[ (E ↑)⊥ = {0} ]] = 1. Thus, E is a basis for X only on condition that [[ E is a basis for X ]] = 1. If |E | = λ and ϕ : λ → E are bijections then the modified ascent ϕ↑ is a bijection of λ∧ to E ↑. Conversely, suppose that D is a basis for X and [[ ψ : λ∧ → D is a bijection ]] = 1 for some cardinal λ. In this case the modified descent ϕ := ψ↓ : λ → D↓ is injective. Consequently, the set E := im(ϕ) has cardinality λ. Moreover, as shown above, it is orthonormal. We are left with observing that D ↓ = mix(E ) = E ↑↓, i.e., [[ E ↑ = D ]] = 1. Finally, E is a basis for X, which completes the proof. B 8.3.11. Theorem. For a Kaplansky–Hilbert module X to be strictly λ-homogeneous it is necessary and sufficient that [[ dim(X ) = λ∧ ]] = 1. C Suppose that X is strictly λ-homogeneous module. By Theorem 8.3.10 [[ dim(X ) = |λ∧ | ]] = 1. On the other hand, there is a partition of unity (bα )α